From 87744d2abc27be9f65bb538375e3468e7aad1cbb Mon Sep 17 00:00:00 2001
From: dianjixz <18637716021@163.com>
Date: Mon, 12 May 2025 10:20:54 +0800
Subject: [PATCH 01/79] [update] README.md add model list

---
 README_zh.md | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/README_zh.md b/README_zh.md
index dbeff436..4216b0aa 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -14,6 +14,7 @@
 
 * [特性](#特性)
 * [Demo](#demo)
+* [模型列表](#模型列表)
 * [环境要求](#环境要求)
 * [编译](#编译)
 * [安装](#安装)
@@ -54,6 +55,11 @@ StackFlow 语音助手的主要工作模式：
 - [StackFlow yolo 视觉检测](https://github.com/Abandon-ht/ModuleLLM_Development_Guide/tree/dev/ESP32/cpp)
 - [StackFlow VLM 图片描述](https://github.com/Abandon-ht/ModuleLLM_Development_Guide/tree/dev/ESP32/cpp)
 
+## 模型列表
+| 模型名 | 模型类型 | 模型大小 | 模型能力 | 模型配置文件 | 计算单元 |
+| :----: | :----: | :----: | :----: | :----: | :----: |
+| [sherpa-onnx-kws-zipformer-gigaspeech-3.3M-2024-01-01](https://github.com/k2-fsa/sherpa-onnx/releases/download/kws-models/sherpa-onnx-kws-zipformer-gigaspeech-3.3M-2024-01-01.tar.bz2) | KWS | 6.4M | 关键词识别 | [mode_sherpa-onnx-kws-zipformer-gigaspeech-3.3M-2024-01-01.json](projects/llm_framework/main_kws/mode_sherpa-onnx-kws-zipformer-gigaspeech-3.3M-2024-01-01.json) | CPU |
+
 ## 环境要求 ##
 当前 StackFlow 的 AI 单元是建立在 AXERA 加速平台之上的，主要的芯片平台为 ax630c、ax650n。系统要求为 ubuntu。
 

From 10e4bdf828912595765ef4031b95435073054ea7 Mon Sep 17 00:00:00 2001
From: yuyun2000 <15515722313yxw@gmail.com>
Date: Thu, 15 May 2025 15:17:04 +0800
Subject: [PATCH 02/79] Refactor SOLA component code

Streamline and simplify code in the SOLA module for improved readability and maintenance
---
 .../llm_framework/main_melotts/src/main.cpp   | 119 ++----------------
 1 file changed, 12 insertions(+), 107 deletions(-)

diff --git a/projects/llm_framework/main_melotts/src/main.cpp b/projects/llm_framework/main_melotts/src/main.cpp
index 4fab699f..5c371d97 100644
--- a/projects/llm_framework/main_melotts/src/main.cpp
+++ b/projects/llm_framework/main_melotts/src/main.cpp
@@ -243,7 +243,6 @@ class llm_task {
         try {
             std::vector<int16_t> wav_pcm_data;
             if (msg_str.empty()) {
-                SLOGI("empty");
                 if (out_callback_) {
                     std::string output = wav_pcm_data.empty() ? std::string()
                                                               : std::string((char *)wav_pcm_data.data(),
@@ -252,9 +251,7 @@ class llm_task {
                 }
                 return false;
             }
-            SLOGI("Processing text: %s", msg_str.c_str());
 
-            // Convert text to phonemes and tones
             std::vector<int> phones_bef, tones_bef;
             lexicon_->convert(msg_str, phones_bef, tones_bef);
             auto phones   = intersperse(phones_bef, 0);
@@ -262,9 +259,6 @@ class llm_task {
             int phone_len = phones.size();
             std::vector<int> langids(phone_len, 3);
 
-            SLOGI("Phoneme conversion completed, length: %d", phone_len);
-
-            // Run the encoder to generate hidden representations
             auto encoder_output =
                 encoder_->Run(phones, tones, langids, g_matrix, mode_config_.noise_scale, mode_config_.noise_scale_w,
                               mode_config_.get_length_scale(), mode_config_.sdp_ratio);
@@ -273,33 +267,22 @@ class llm_task {
             auto zp_info   = encoder_output.at(0).GetTensorTypeAndShapeInfo();
             auto zp_shape  = zp_info.GetShape();
 
-            SLOGI("Encoder output completed, shape: [%ld, %ld, %ld], expected audio length: %d", zp_shape[0],
-                  zp_shape[1], zp_shape[2], audio_len);
-
-            // Calculate decoder parameters
             int zp_size         = decoder_->GetInputSize(0) / sizeof(float);
             int dec_len         = zp_size / zp_shape[1];
             int audio_slice_len = decoder_->GetOutputSize(0) / sizeof(float);
 
-            const int pad_frames        = 16;
+            const int pad_frames        = 24;
             const int samples_per_frame = 512;
 
-            SLOGI("Decoder configuration: frame length=%d, audio slice length=%d, pad length=%d, samples per frame=%d",
-                  dec_len, audio_slice_len, pad_frames, samples_per_frame);
-
             const int effective_frames = dec_len - 2 * pad_frames;
 
             int dec_slice_num =
                 static_cast<int>(std::ceil(static_cast<double>(zp_shape[2]) / static_cast<double>(effective_frames)));
 
-            SLOGI("Will perform %d inferences, each with effective frames: %d", dec_slice_num, effective_frames);
+            const int sola_buffer_frame = pad_frames * samples_per_frame;
+            const int sola_search_frame = pad_frames * samples_per_frame;
+            const int block_frame       = (dec_len - 2 * pad_frames) * samples_per_frame;
 
-            // SOLA parameters setup
-            const int sola_buffer_frame = pad_frames * samples_per_frame;                  // Overlap buffer length
-            const int sola_search_frame = pad_frames * samples_per_frame;                  // Search window length
-            const int block_frame       = (dec_len - 2 * pad_frames) * samples_per_frame;  // Effective block length
-
-            // Create fade-in/fade-out windows for smooth transitions
             std::vector<float> fade_in_window(sola_buffer_frame);
             std::vector<float> fade_out_window(sola_buffer_frame);
 
@@ -308,50 +291,35 @@ class llm_task {
                 fade_out_window[i] = 1.0f - fade_in_window[i];
             }
 
-            // Initialize SOLA buffer
             std::vector<float> sola_buffer(sola_buffer_frame, 0.0f);
             bool first_frame = true;
 
             std::vector<float> pcmlist;
 
-            // Main decoding loop - process each slice
             for (int i = 0; i < dec_slice_num; i++) {
-                // Calculate start position for current batch input
                 int input_start = i * effective_frames;
-                // Consider forward padding, but ensure non-negative
                 if (i > 0) {
                     input_start -= pad_frames;
                 }
                 input_start = std::max(0, input_start);
 
-                // Actual input length
                 int actual_len = std::min(dec_len, static_cast<int>(zp_shape[2] - input_start));
 
-                // Calculate effective output range (frame level)
                 int output_start_frame, output_end_frame;
 
                 if (i == 0) {
-                    // First frame: skip padding at beginning
                     output_start_frame = 0;
                     output_end_frame   = effective_frames - 1;
                 } else if (i == dec_slice_num - 1) {
-                    // Last frame: calculate from current segment start
                     output_start_frame = i * effective_frames;
-                    // Last frame extends to encoder's maximum output length
-                    output_end_frame = static_cast<int>(zp_shape[2]) - 1;
+                    output_end_frame   = static_cast<int>(zp_shape[2]) - 1;
                 } else {
-                    // Middle frames: standard calculation
                     output_start_frame = i * effective_frames;
                     output_end_frame   = (i + 1) * effective_frames - 1;
                 }
 
-                SLOGI("Inference #%d: input frame range=[%d-%d], actual length=%d, output frame range=[%d-%d]", i + 1,
-                      input_start, input_start + actual_len - 1, actual_len, output_start_frame, output_end_frame);
-
-                // Prepare decoder input, initialize all to zero
                 std::vector<float> zp(zp_size, 0);
 
-                // Copy data to decoder input
                 for (int n = 0; n < zp_shape[1]; n++) {
                     int copy_size = std::min(actual_len, static_cast<int>(zp_shape[2] - input_start));
                     if (copy_size > 0) {
@@ -360,76 +328,49 @@ class llm_task {
                     }
                 }
 
-                // Run decoder
                 std::vector<float> decoder_output(audio_slice_len);
                 decoder_->SetInput(zp.data(), 0);
                 decoder_->SetInput(g_matrix.data(), 1);
 
-                SLOGI("Inference #%d: starting decoding...", i + 1);
-
                 if (0 != decoder_->Run()) {
-                    SLOGI("Inference #%d: decoding failed", i + 1);
                     throw std::string("decoder_ RunSync error");
                 }
 
                 decoder_->GetOutput(decoder_output.data(), 0);
 
-                // === SOLA Processing Logic ===
                 if (first_frame) {
-                    // Special handling for first frame - should not skip initial content
-                    // First frame starts directly from decoder output without skipping
-                    int audio_start = 0;  // Start from beginning, don't skip pad_frames
+                    int audio_start = 0;
+                    int audio_len   = decoder_output.size() - sola_buffer_frame;
+                    audio_len       = std::max(0, audio_len);
 
-                    // Calculate data length for first frame
-                    // First frame should preserve complete decoder output, only reserving sola_buffer_frame at the end
-                    // for next frame alignment
-                    int audio_len = decoder_output.size() - sola_buffer_frame;
-
-                    // Boundary check
-                    audio_len = std::max(0, audio_len);  // Ensure non-negative
-
-                    // Add first frame data
                     if (audio_len > 0) {
                         pcmlist.insert(pcmlist.end(), decoder_output.begin() + audio_start,
                                        decoder_output.begin() + audio_start + audio_len);
                     }
 
-                    // Save sola_buffer_frame length from the end to SOLA buffer for next frame alignment
                     int buffer_start = audio_len;
 
-                    // Ensure sufficient data is available for copying
                     if (buffer_start + sola_buffer_frame <= decoder_output.size()) {
                         std::copy(decoder_output.begin() + buffer_start,
                                   decoder_output.begin() + buffer_start + sola_buffer_frame, sola_buffer.begin());
                     } else {
-                        // Possible case: first frame data is shorter than sola_buffer_frame
                         int available = static_cast<int>(decoder_output.size() - buffer_start);
                         if (available > 0) {
                             std::copy(decoder_output.begin() + buffer_start, decoder_output.end(), sola_buffer.begin());
-                            // Fill with zeros
                             std::fill(sola_buffer.begin() + available, sola_buffer.end(), 0.0f);
                         } else {
-                            // Completely insufficient data, fill all with zeros
                             std::fill(sola_buffer.begin(), sola_buffer.end(), 0.0f);
                         }
                     }
 
                     first_frame = false;
-
-                    SLOGI(
-                        "Inference #%d: First frame processing, added %d samples from position %d to output, saved %d "
-                        "samples to SOLA buffer",
-                        i + 1, audio_len, audio_start, sola_buffer_frame);
                 } else {
-                    // Non-first frame: SOLA alignment required
                     int audio_start = pad_frames * samples_per_frame;
 
-                    // 1. Prepare search window - beginning portion of current frame
                     std::vector<float> search_window(sola_buffer_frame + sola_search_frame);
                     std::copy(decoder_output.begin() + audio_start,
                               decoder_output.begin() + audio_start + search_window.size(), search_window.begin());
 
-                    // 2. Find best alignment point (calculate cross-correlation)
                     int best_offset        = 0;
                     float best_correlation = -1.0;
 
@@ -442,7 +383,6 @@ class llm_task {
                             energy += search_window[j + offset] * search_window[j + offset];
                         }
 
-                        // Normalize correlation (avoid division by zero)
                         float normalized_correlation = (energy > 1e-8) ? correlation / std::sqrt(energy) : 0.0f;
 
                         if (normalized_correlation > best_correlation) {
@@ -451,40 +391,28 @@ class llm_task {
                         }
                     }
 
-                    SLOGI("Inference #%d: SOLA found best alignment offset %d with correlation coefficient %f", i + 1,
-                          best_offset, best_correlation);
-
-                    // 3. Apply alignment offset
                     int aligned_start = audio_start + best_offset;
 
-                    // 4. Smooth transition processing (crossfade in alignment region)
                     std::vector<float> crossfade_region(sola_buffer_frame);
 
                     for (int j = 0; j < sola_buffer_frame; j++) {
-                        // Apply fade-in/fade-out window functions
                         crossfade_region[j] =
                             decoder_output[aligned_start + j] * fade_in_window[j] + sola_buffer[j] * fade_out_window[j];
                     }
 
-                    // 5. Add crossfade region to output
                     pcmlist.insert(pcmlist.end(), crossfade_region.begin(), crossfade_region.end());
 
                     int remaining_start = aligned_start + sola_buffer_frame;
 
                     if (i == dec_slice_num - 1) {
                         int total_expected_samples = audio_len * samples_per_frame / 512;
-
-                        int processed_samples = static_cast<int>(pcmlist.size());
-
-                        int remaining_needed = total_expected_samples - processed_samples;
-                        remaining_needed     = std::max(0, remaining_needed);
+                        int processed_samples      = static_cast<int>(pcmlist.size());
+                        int remaining_needed       = total_expected_samples - processed_samples;
+                        remaining_needed           = std::max(0, remaining_needed);
 
                         int remaining_len =
                             std::min(remaining_needed, static_cast<int>(decoder_output.size() - remaining_start));
 
-                        SLOGI("Inference #%d (final): Expected total=%d, processed=%d, needed=%d, available=%d", i + 1,
-                              total_expected_samples, processed_samples, remaining_needed, remaining_len);
-
                         if (remaining_len > 0) {
                             pcmlist.insert(pcmlist.end(), decoder_output.begin() + remaining_start,
                                            decoder_output.begin() + remaining_start + remaining_len);
@@ -492,7 +420,6 @@ class llm_task {
 
                     } else {
                         int remaining_len = (dec_len - 2 * pad_frames) * samples_per_frame - sola_buffer_frame;
-
                         remaining_len =
                             std::min(remaining_len, static_cast<int>(decoder_output.size() - remaining_start));
 
@@ -514,55 +441,33 @@ class llm_task {
                             }
                             std::fill(sola_buffer.begin() + avail, sola_buffer.end(), 0.0f);
                         }
-
-                        SLOGI("Inference #%d: Added %d + %d samples to output, cumulative length: %zu", i + 1,
-                              sola_buffer_frame, remaining_len, pcmlist.size());
                     }
                 }
             }
 
-            SLOGI("All inference completed, raw generated PCM length: %zu", pcmlist.size());
-
             if (pcmlist.size() > audio_len) {
-                SLOGI("Truncating output from %zu to %d samples as per encoder prediction", pcmlist.size(), audio_len);
                 pcmlist.resize(audio_len);
             }
 
-            SLOGI("Final PCM length after truncation: %zu", pcmlist.size());
-
-            // Post-processing: resample and convert to int16
             double src_ratio =
                 static_cast<double>(mode_config_.audio_rate) / static_cast<double>(mode_config_.mode_rate);
             std::vector<float> tmp_pcm((pcmlist.size() * src_ratio + 1));
             int len;
 
-            SLOGI("Starting audio resampling, source rate: %f, target rate: %f, ratio: %f",
-                  static_cast<float>(mode_config_.mode_rate), static_cast<float>(mode_config_.audio_rate), src_ratio);
-
             resample_audio(pcmlist.data(), pcmlist.size(), tmp_pcm.data(), &len, src_ratio);
 
-            SLOGI("Resampling completed, length after resampling: %d", len);
-
-            // Convert to 16-bit PCM
             wav_pcm_data.reserve(len);
             std::transform(tmp_pcm.begin(), tmp_pcm.begin() + len, std::back_inserter(wav_pcm_data),
                            [](const auto val) { return static_cast<int16_t>(val * INT16_MAX); });
 
-            SLOGI("Final audio length: %zu samples", wav_pcm_data.size());
-
-            // Call the output callback function with the result
             if (out_callback_) {
                 out_callback_(
                     std::string(reinterpret_cast<char *>(wav_pcm_data.data()), wav_pcm_data.size() * sizeof(int16_t)),
                     finish);
             }
-
-            SLOGI("TTS processing completed, output callback invoked");
         } catch (const std::exception &e) {
-            SLOGI("TTS processing exception: %s", e.what());
             return true;
         } catch (...) {
-            SLOGI("TTS processing encountered an unknown exception");
             return true;
         }
         return false;
@@ -975,4 +880,4 @@ int main(int argc, char *argv[])
     }
     llm.llm_firework_exit();
     return 0;
-}
\ No newline at end of file
+}

From 74c41a3e4479d883c527fffef052a1ce4f70fb10 Mon Sep 17 00:00:00 2001
From: yuyun2000 <15515722313yxw@gmail.com>
Date: Fri, 16 May 2025 11:10:58 +0800
Subject: [PATCH 03/79] Add text normalization for Chinese, Japanese, and
 English

Implement regex-based text normalization functionality to support trilingual (CJE) content processing
---
 .../llm_framework/include/fst/accumulator.h   |  903 +++++++++
 projects/llm_framework/include/fst/add-on.h   |  248 +++
 .../llm_framework/include/fst/arc-arena.h     |  232 +++
 projects/llm_framework/include/fst/arc-map.h  | 1285 +++++++++++++
 projects/llm_framework/include/fst/arc.h      |  317 ++++
 .../llm_framework/include/fst/arcfilter.h     |   93 +
 projects/llm_framework/include/fst/arcsort.h  |  211 +++
 projects/llm_framework/include/fst/bi-table.h |  480 +++++
 projects/llm_framework/include/fst/cache.h    | 1327 +++++++++++++
 projects/llm_framework/include/fst/closure.h  |  134 ++
 .../llm_framework/include/fst/compact-fst.h   | 1564 ++++++++++++++++
 projects/llm_framework/include/fst/compat.h   |  130 ++
 .../llm_framework/include/fst/complement.h    |  277 +++
 .../include/fst/compose-filter.h              |  571 ++++++
 projects/llm_framework/include/fst/compose.h  | 1035 ++++++++++
 projects/llm_framework/include/fst/concat.h   |  220 +++
 projects/llm_framework/include/fst/config.h   |    3 +
 .../llm_framework/include/fst/config.h.in     |   11 +
 projects/llm_framework/include/fst/connect.h  |  323 ++++
 .../llm_framework/include/fst/const-fst.h     |  485 +++++
 .../llm_framework/include/fst/determinize.h   | 1093 +++++++++++
 .../llm_framework/include/fst/dfs-visit.h     |  202 ++
 .../llm_framework/include/fst/difference.h    |  205 ++
 .../llm_framework/include/fst/disambiguate.h  |  564 ++++++
 projects/llm_framework/include/fst/edit-fst.h |  702 +++++++
 projects/llm_framework/include/fst/encode.h   |  556 ++++++
 .../llm_framework/include/fst/epsnormalize.h  |   61 +
 projects/llm_framework/include/fst/equal.h    |  169 ++
 .../llm_framework/include/fst/equivalent.h    |  230 +++
 .../llm_framework/include/fst/expanded-fst.h  |  179 ++
 .../include/fst/expectation-weight.h          |  134 ++
 .../fst/extensions/compress/compress-script.h |   53 +
 .../fst/extensions/compress/compress.h        |  906 +++++++++
 .../include/fst/extensions/compress/elias.h   |   97 +
 .../include/fst/extensions/compress/gzfile.h  |  127 ++
 .../include/fst/extensions/compress/randmod.h |  102 +
 .../fst/extensions/far/compile-strings.h      |  260 +++
 .../include/fst/extensions/far/create.h       |   46 +
 .../include/fst/extensions/far/equal.h        |   69 +
 .../include/fst/extensions/far/extract.h      |  118 ++
 .../include/fst/extensions/far/far-class.h    |  258 +++
 .../include/fst/extensions/far/far.h          |  481 +++++
 .../include/fst/extensions/far/farlib.h       |   19 +
 .../include/fst/extensions/far/farscript.h    |  269 +++
 .../include/fst/extensions/far/getters.h      |   30 +
 .../include/fst/extensions/far/info.h         |  147 ++
 .../include/fst/extensions/far/isomorphic.h   |   69 +
 .../fst/extensions/far/print-strings.h        |  105 ++
 .../include/fst/extensions/far/script-impl.h  |   23 +
 .../include/fst/extensions/far/stlist.h       |  273 +++
 .../include/fst/extensions/far/sttable.h      |  353 ++++
 .../linear/linear-fst-data-builder.h          | 1074 +++++++++++
 .../fst/extensions/linear/linear-fst-data.h   |  526 ++++++
 .../fst/extensions/linear/linear-fst.h        | 1173 ++++++++++++
 .../fst/extensions/linear/linearscript.h      |  391 ++++
 .../fst/extensions/linear/loglinear-apply.h   |   77 +
 .../include/fst/extensions/linear/trie.h      |  444 +++++
 .../include/fst/extensions/mpdt/compose.h     |  267 +++
 .../include/fst/extensions/mpdt/expand.h      |  335 ++++
 .../include/fst/extensions/mpdt/info.h        |  190 ++
 .../include/fst/extensions/mpdt/mpdt.h        |  357 ++++
 .../include/fst/extensions/mpdt/mpdtlib.h     |   18 +
 .../include/fst/extensions/mpdt/mpdtscript.h  |  156 ++
 .../fst/extensions/mpdt/read_write_utils.h    |   86 +
 .../include/fst/extensions/mpdt/reverse.h     |   54 +
 .../fst/extensions/ngram/bitmap-index.h       |  168 ++
 .../include/fst/extensions/ngram/ngram-fst.h  | 1027 ++++++++++
 .../include/fst/extensions/ngram/nthbit.h     |   49 +
 .../include/fst/extensions/pdt/collection.h   |  107 ++
 .../include/fst/extensions/pdt/compose.h      |  493 +++++
 .../include/fst/extensions/pdt/expand.h       |  933 +++++++++
 .../include/fst/extensions/pdt/getters.h      |   22 +
 .../include/fst/extensions/pdt/info.h         |  152 ++
 .../include/fst/extensions/pdt/paren.h        |  440 +++++
 .../include/fst/extensions/pdt/pdt.h          |  165 ++
 .../include/fst/extensions/pdt/pdtlib.h       |   19 +
 .../include/fst/extensions/pdt/pdtscript.h    |  244 +++
 .../include/fst/extensions/pdt/replace.h      |  827 ++++++++
 .../include/fst/extensions/pdt/reverse.h      |   38 +
 .../fst/extensions/pdt/shortest-path.h        |  715 +++++++
 .../include/fst/extensions/special/phi-fst.h  |  183 ++
 .../include/fst/extensions/special/rho-fst.h  |  172 ++
 .../fst/extensions/special/sigma-fst.h        |  176 ++
 .../llm_framework/include/fst/factor-weight.h |  496 +++++
 .../llm_framework/include/fst/filter-state.h  |  199 ++
 projects/llm_framework/include/fst/flags.h    |  219 +++
 .../llm_framework/include/fst/float-weight.h  |  858 +++++++++
 projects/llm_framework/include/fst/fst-decl.h |  254 +++
 projects/llm_framework/include/fst/fst.h      | 1007 ++++++++++
 projects/llm_framework/include/fst/fstlib.h   |  130 ++
 .../include/fst/generic-register.h            |  126 ++
 projects/llm_framework/include/fst/heap.h     |  168 ++
 projects/llm_framework/include/fst/icu.h      |  129 ++
 .../llm_framework/include/fst/intersect.h     |  181 ++
 .../llm_framework/include/fst/interval-set.h  |  398 ++++
 projects/llm_framework/include/fst/invert.h   |  139 ++
 .../llm_framework/include/fst/isomorphic.h    |  183 ++
 .../include/fst/label-reachable.h             |  511 +++++
 .../include/fst/lexicographic-weight.h        |  173 ++
 projects/llm_framework/include/fst/lock.h     |   62 +
 projects/llm_framework/include/fst/log.h      |   78 +
 .../include/fst/lookahead-filter.h            |  623 ++++++
 .../include/fst/lookahead-matcher.h           |  841 +++++++++
 projects/llm_framework/include/fst/map.h      |  110 ++
 .../llm_framework/include/fst/mapped-file.h   |   81 +
 .../llm_framework/include/fst/matcher-fst.h   |  347 ++++
 projects/llm_framework/include/fst/matcher.h  | 1575 ++++++++++++++++
 projects/llm_framework/include/fst/memory.h   |  443 +++++
 projects/llm_framework/include/fst/minimize.h |  568 ++++++
 .../llm_framework/include/fst/mutable-fst.h   |  398 ++++
 .../llm_framework/include/fst/pair-weight.h   |  155 ++
 .../llm_framework/include/fst/partition.h     |  305 +++
 .../llm_framework/include/fst/power-weight.h  |  168 ++
 .../include/fst/product-weight.h              |  107 ++
 projects/llm_framework/include/fst/project.h  |  159 ++
 .../llm_framework/include/fst/properties.h    |  468 +++++
 projects/llm_framework/include/fst/prune.h    |  341 ++++
 projects/llm_framework/include/fst/push.h     |  155 ++
 projects/llm_framework/include/fst/queue.h    |  948 ++++++++++
 .../include/fst/randequivalent.h              |  114 ++
 projects/llm_framework/include/fst/randgen.h  |  756 ++++++++
 projects/llm_framework/include/fst/rational.h |  307 +++
 projects/llm_framework/include/fst/register.h |  115 ++
 projects/llm_framework/include/fst/relabel.h  |  472 +++++
 .../llm_framework/include/fst/replace-util.h  |  629 +++++++
 projects/llm_framework/include/fst/replace.h  | 1492 +++++++++++++++
 projects/llm_framework/include/fst/reverse.h  |  116 ++
 projects/llm_framework/include/fst/reweight.h |  127 ++
 .../llm_framework/include/fst/rmepsilon.h     |  548 ++++++
 .../include/fst/rmfinalepsilon.h              |   80 +
 .../include/fst/script/arc-class.h            |   40 +
 .../include/fst/script/arciterator-class.h    |  212 +++
 .../include/fst/script/arcsort.h              |   44 +
 .../include/fst/script/arg-packs.h            |   37 +
 .../include/fst/script/closure.h              |   28 +
 .../include/fst/script/compile-impl.h         |  217 +++
 .../include/fst/script/compile.h              |   98 +
 .../include/fst/script/compose.h              |   34 +
 .../llm_framework/include/fst/script/concat.h |   40 +
 .../include/fst/script/connect.h              |   23 +
 .../include/fst/script/convert.h              |   35 +
 .../llm_framework/include/fst/script/decode.h |   49 +
 .../include/fst/script/determinize.h          |   59 +
 .../include/fst/script/difference.h           |   35 +
 .../include/fst/script/disambiguate.h         |   54 +
 .../include/fst/script/draw-impl.h            |  227 +++
 .../llm_framework/include/fst/script/draw.h   |   85 +
 .../llm_framework/include/fst/script/encode.h |   51 +
 .../include/fst/script/encodemapper-class.h   |  169 ++
 .../include/fst/script/epsnormalize.h         |   31 +
 .../llm_framework/include/fst/script/equal.h  |   32 +
 .../include/fst/script/equivalent.h           |   34 +
 .../include/fst/script/fst-class.h            |  530 ++++++
 .../include/fst/script/fstscript-decl.h       |   32 +
 .../include/fst/script/fstscript.h            |  155 ++
 .../include/fst/script/getters.h              |   76 +
 .../include/fst/script/info-impl.h            |  314 ++++
 .../llm_framework/include/fst/script/info.h   |   50 +
 .../include/fst/script/intersect.h            |   35 +
 .../llm_framework/include/fst/script/invert.h |   23 +
 .../include/fst/script/isomorphic.h           |   34 +
 .../llm_framework/include/fst/script/map.h    |  158 ++
 .../include/fst/script/minimize.h             |   33 +
 .../include/fst/script/print-impl.h           |  132 ++
 .../llm_framework/include/fst/script/print.h  |   79 +
 .../include/fst/script/project.h              |   28 +
 .../llm_framework/include/fst/script/prune.h  |   51 +
 .../llm_framework/include/fst/script/push.h   |   53 +
 .../include/fst/script/randequivalent.h       |   67 +
 .../include/fst/script/randgen.h              |   63 +
 .../include/fst/script/register.h             |   99 +
 .../include/fst/script/relabel.h              |   64 +
 .../include/fst/script/replace.h              |   72 +
 .../include/fst/script/reverse.h              |   30 +
 .../include/fst/script/reweight.h             |   37 +
 .../include/fst/script/rmepsilon.h            |  109 ++
 .../include/fst/script/script-impl.h          |  211 +++
 .../include/fst/script/shortest-distance.h    |  214 +++
 .../include/fst/script/shortest-path.h        |  116 ++
 .../include/fst/script/stateiterator-class.h  |   85 +
 .../include/fst/script/synchronize.h          |   29 +
 .../include/fst/script/text-io.h              |   28 +
 .../include/fst/script/topsort.h              |   26 +
 .../llm_framework/include/fst/script/union.h  |   29 +
 .../llm_framework/include/fst/script/verify.h |   27 +
 .../include/fst/script/weight-class.h         |  235 +++
 .../llm_framework/include/fst/set-weight.h    |  618 ++++++
 .../include/fst/shortest-distance.h           |  351 ++++
 .../llm_framework/include/fst/shortest-path.h |  549 ++++++
 .../include/fst/signed-log-weight.h           |  440 +++++
 .../include/fst/sparse-power-weight.h         |  209 +++
 .../include/fst/sparse-tuple-weight.h         |  422 +++++
 .../llm_framework/include/fst/state-map.h     |  613 ++++++
 .../include/fst/state-reachable.h             |  224 +++
 .../llm_framework/include/fst/state-table.h   |  494 +++++
 .../llm_framework/include/fst/statesort.h     |   74 +
 .../llm_framework/include/fst/string-weight.h |  807 ++++++++
 projects/llm_framework/include/fst/string.h   |  286 +++
 .../include/fst/symbol-table-ops.h            |   76 +
 .../llm_framework/include/fst/symbol-table.h  |  445 +++++
 .../llm_framework/include/fst/synchronize.h   |  414 ++++
 .../include/fst/test-properties.h             |  246 +++
 .../include/fst/test/algo_test.h              | 1414 ++++++++++++++
 .../llm_framework/include/fst/test/fst_test.h |  318 ++++
 .../llm_framework/include/fst/test/rand-fst.h |   90 +
 .../include/fst/test/weight-tester.h          |  207 ++
 projects/llm_framework/include/fst/topsort.h  |   95 +
 .../llm_framework/include/fst/tuple-weight.h  |  163 ++
 projects/llm_framework/include/fst/types.h    |   41 +
 .../llm_framework/include/fst/union-find.h    |   84 +
 .../llm_framework/include/fst/union-weight.h  |  505 +++++
 projects/llm_framework/include/fst/union.h    |  157 ++
 projects/llm_framework/include/fst/util.h     |  400 ++++
 .../llm_framework/include/fst/vector-fst.h    |  796 ++++++++
 projects/llm_framework/include/fst/verify.h   |  100 +
 projects/llm_framework/include/fst/visit.h    |  321 ++++
 projects/llm_framework/include/fst/weight.h   |  389 ++++
 .../llm_framework/include/gflags/defines.h    |   48 +
 .../llm_framework/include/gflags/gflags.h     |  626 +++++++
 .../include/gflags/gflags_completions.h       |  121 ++
 .../include/gflags/gflags_declare.h           |  156 ++
 .../llm_framework/include/glog/log_severity.h |   92 +
 projects/llm_framework/include/glog/logging.h | 1662 +++++++++++++++++
 .../llm_framework/include/glog/raw_logging.h  |  180 ++
 .../llm_framework/include/glog/stl_logging.h  |  220 +++
 .../llm_framework/include/glog/vlog_is_on.h   |  129 ++
 .../llm_framework/main_melotts/SConstruct     |    7 +-
 .../main_melotts/mode_melotts-en-default.json |    2 +
 .../main_melotts/mode_melotts-en-us.json      |    8 +-
 .../main_melotts/mode_melotts-ja-jp.json      |    2 +
 .../main_melotts/mode_melotts-zh-cn.json      |    2 +
 .../llm_framework/main_melotts/src/main.cpp   |   20 +-
 .../main_melotts/src/runner/Lexicon.hpp       |   25 +-
 .../main_melotts/src/runner/base64.cpp        |  100 +-
 .../main_melotts/src/runner/base64.h          |   10 +-
 .../src/runner/processor/CMakeLists.txt       |   13 +
 .../src/runner/processor/wetext_processor.cc  |   86 +
 .../src/runner/processor/wetext_processor.h   |   51 +
 .../runner/processor/wetext_token_parser.cc   |  161 ++
 .../runner/processor/wetext_token_parser.h    |   94 +
 .../src/runner/utils/CMakeLists.txt           |    3 +
 .../src/runner/utils/wetext_flags.h           |   23 +
 .../src/runner/utils/wetext_log.h             |   23 +
 .../src/runner/utils/wetext_string.cc         |   89 +
 .../src/runner/utils/wetext_string.h          |   42 +
 245 files changed, 66914 insertions(+), 73 deletions(-)
 create mode 100644 projects/llm_framework/include/fst/accumulator.h
 create mode 100644 projects/llm_framework/include/fst/add-on.h
 create mode 100644 projects/llm_framework/include/fst/arc-arena.h
 create mode 100644 projects/llm_framework/include/fst/arc-map.h
 create mode 100644 projects/llm_framework/include/fst/arc.h
 create mode 100644 projects/llm_framework/include/fst/arcfilter.h
 create mode 100644 projects/llm_framework/include/fst/arcsort.h
 create mode 100644 projects/llm_framework/include/fst/bi-table.h
 create mode 100644 projects/llm_framework/include/fst/cache.h
 create mode 100644 projects/llm_framework/include/fst/closure.h
 create mode 100644 projects/llm_framework/include/fst/compact-fst.h
 create mode 100644 projects/llm_framework/include/fst/compat.h
 create mode 100644 projects/llm_framework/include/fst/complement.h
 create mode 100644 projects/llm_framework/include/fst/compose-filter.h
 create mode 100644 projects/llm_framework/include/fst/compose.h
 create mode 100644 projects/llm_framework/include/fst/concat.h
 create mode 100644 projects/llm_framework/include/fst/config.h
 create mode 100644 projects/llm_framework/include/fst/config.h.in
 create mode 100644 projects/llm_framework/include/fst/connect.h
 create mode 100644 projects/llm_framework/include/fst/const-fst.h
 create mode 100644 projects/llm_framework/include/fst/determinize.h
 create mode 100644 projects/llm_framework/include/fst/dfs-visit.h
 create mode 100644 projects/llm_framework/include/fst/difference.h
 create mode 100644 projects/llm_framework/include/fst/disambiguate.h
 create mode 100644 projects/llm_framework/include/fst/edit-fst.h
 create mode 100644 projects/llm_framework/include/fst/encode.h
 create mode 100644 projects/llm_framework/include/fst/epsnormalize.h
 create mode 100644 projects/llm_framework/include/fst/equal.h
 create mode 100644 projects/llm_framework/include/fst/equivalent.h
 create mode 100644 projects/llm_framework/include/fst/expanded-fst.h
 create mode 100644 projects/llm_framework/include/fst/expectation-weight.h
 create mode 100644 projects/llm_framework/include/fst/extensions/compress/compress-script.h
 create mode 100644 projects/llm_framework/include/fst/extensions/compress/compress.h
 create mode 100644 projects/llm_framework/include/fst/extensions/compress/elias.h
 create mode 100644 projects/llm_framework/include/fst/extensions/compress/gzfile.h
 create mode 100644 projects/llm_framework/include/fst/extensions/compress/randmod.h
 create mode 100644 projects/llm_framework/include/fst/extensions/far/compile-strings.h
 create mode 100644 projects/llm_framework/include/fst/extensions/far/create.h
 create mode 100644 projects/llm_framework/include/fst/extensions/far/equal.h
 create mode 100644 projects/llm_framework/include/fst/extensions/far/extract.h
 create mode 100644 projects/llm_framework/include/fst/extensions/far/far-class.h
 create mode 100644 projects/llm_framework/include/fst/extensions/far/far.h
 create mode 100644 projects/llm_framework/include/fst/extensions/far/farlib.h
 create mode 100644 projects/llm_framework/include/fst/extensions/far/farscript.h
 create mode 100644 projects/llm_framework/include/fst/extensions/far/getters.h
 create mode 100644 projects/llm_framework/include/fst/extensions/far/info.h
 create mode 100644 projects/llm_framework/include/fst/extensions/far/isomorphic.h
 create mode 100644 projects/llm_framework/include/fst/extensions/far/print-strings.h
 create mode 100644 projects/llm_framework/include/fst/extensions/far/script-impl.h
 create mode 100644 projects/llm_framework/include/fst/extensions/far/stlist.h
 create mode 100644 projects/llm_framework/include/fst/extensions/far/sttable.h
 create mode 100644 projects/llm_framework/include/fst/extensions/linear/linear-fst-data-builder.h
 create mode 100644 projects/llm_framework/include/fst/extensions/linear/linear-fst-data.h
 create mode 100644 projects/llm_framework/include/fst/extensions/linear/linear-fst.h
 create mode 100644 projects/llm_framework/include/fst/extensions/linear/linearscript.h
 create mode 100644 projects/llm_framework/include/fst/extensions/linear/loglinear-apply.h
 create mode 100644 projects/llm_framework/include/fst/extensions/linear/trie.h
 create mode 100644 projects/llm_framework/include/fst/extensions/mpdt/compose.h
 create mode 100644 projects/llm_framework/include/fst/extensions/mpdt/expand.h
 create mode 100644 projects/llm_framework/include/fst/extensions/mpdt/info.h
 create mode 100644 projects/llm_framework/include/fst/extensions/mpdt/mpdt.h
 create mode 100644 projects/llm_framework/include/fst/extensions/mpdt/mpdtlib.h
 create mode 100644 projects/llm_framework/include/fst/extensions/mpdt/mpdtscript.h
 create mode 100644 projects/llm_framework/include/fst/extensions/mpdt/read_write_utils.h
 create mode 100644 projects/llm_framework/include/fst/extensions/mpdt/reverse.h
 create mode 100644 projects/llm_framework/include/fst/extensions/ngram/bitmap-index.h
 create mode 100644 projects/llm_framework/include/fst/extensions/ngram/ngram-fst.h
 create mode 100644 projects/llm_framework/include/fst/extensions/ngram/nthbit.h
 create mode 100644 projects/llm_framework/include/fst/extensions/pdt/collection.h
 create mode 100644 projects/llm_framework/include/fst/extensions/pdt/compose.h
 create mode 100644 projects/llm_framework/include/fst/extensions/pdt/expand.h
 create mode 100644 projects/llm_framework/include/fst/extensions/pdt/getters.h
 create mode 100644 projects/llm_framework/include/fst/extensions/pdt/info.h
 create mode 100644 projects/llm_framework/include/fst/extensions/pdt/paren.h
 create mode 100644 projects/llm_framework/include/fst/extensions/pdt/pdt.h
 create mode 100644 projects/llm_framework/include/fst/extensions/pdt/pdtlib.h
 create mode 100644 projects/llm_framework/include/fst/extensions/pdt/pdtscript.h
 create mode 100644 projects/llm_framework/include/fst/extensions/pdt/replace.h
 create mode 100644 projects/llm_framework/include/fst/extensions/pdt/reverse.h
 create mode 100644 projects/llm_framework/include/fst/extensions/pdt/shortest-path.h
 create mode 100644 projects/llm_framework/include/fst/extensions/special/phi-fst.h
 create mode 100644 projects/llm_framework/include/fst/extensions/special/rho-fst.h
 create mode 100644 projects/llm_framework/include/fst/extensions/special/sigma-fst.h
 create mode 100644 projects/llm_framework/include/fst/factor-weight.h
 create mode 100644 projects/llm_framework/include/fst/filter-state.h
 create mode 100644 projects/llm_framework/include/fst/flags.h
 create mode 100644 projects/llm_framework/include/fst/float-weight.h
 create mode 100644 projects/llm_framework/include/fst/fst-decl.h
 create mode 100644 projects/llm_framework/include/fst/fst.h
 create mode 100644 projects/llm_framework/include/fst/fstlib.h
 create mode 100644 projects/llm_framework/include/fst/generic-register.h
 create mode 100644 projects/llm_framework/include/fst/heap.h
 create mode 100644 projects/llm_framework/include/fst/icu.h
 create mode 100644 projects/llm_framework/include/fst/intersect.h
 create mode 100644 projects/llm_framework/include/fst/interval-set.h
 create mode 100644 projects/llm_framework/include/fst/invert.h
 create mode 100644 projects/llm_framework/include/fst/isomorphic.h
 create mode 100644 projects/llm_framework/include/fst/label-reachable.h
 create mode 100644 projects/llm_framework/include/fst/lexicographic-weight.h
 create mode 100644 projects/llm_framework/include/fst/lock.h
 create mode 100644 projects/llm_framework/include/fst/log.h
 create mode 100644 projects/llm_framework/include/fst/lookahead-filter.h
 create mode 100644 projects/llm_framework/include/fst/lookahead-matcher.h
 create mode 100644 projects/llm_framework/include/fst/map.h
 create mode 100644 projects/llm_framework/include/fst/mapped-file.h
 create mode 100644 projects/llm_framework/include/fst/matcher-fst.h
 create mode 100644 projects/llm_framework/include/fst/matcher.h
 create mode 100644 projects/llm_framework/include/fst/memory.h
 create mode 100644 projects/llm_framework/include/fst/minimize.h
 create mode 100644 projects/llm_framework/include/fst/mutable-fst.h
 create mode 100644 projects/llm_framework/include/fst/pair-weight.h
 create mode 100644 projects/llm_framework/include/fst/partition.h
 create mode 100644 projects/llm_framework/include/fst/power-weight.h
 create mode 100644 projects/llm_framework/include/fst/product-weight.h
 create mode 100644 projects/llm_framework/include/fst/project.h
 create mode 100644 projects/llm_framework/include/fst/properties.h
 create mode 100644 projects/llm_framework/include/fst/prune.h
 create mode 100644 projects/llm_framework/include/fst/push.h
 create mode 100644 projects/llm_framework/include/fst/queue.h
 create mode 100644 projects/llm_framework/include/fst/randequivalent.h
 create mode 100644 projects/llm_framework/include/fst/randgen.h
 create mode 100644 projects/llm_framework/include/fst/rational.h
 create mode 100644 projects/llm_framework/include/fst/register.h
 create mode 100644 projects/llm_framework/include/fst/relabel.h
 create mode 100644 projects/llm_framework/include/fst/replace-util.h
 create mode 100644 projects/llm_framework/include/fst/replace.h
 create mode 100644 projects/llm_framework/include/fst/reverse.h
 create mode 100644 projects/llm_framework/include/fst/reweight.h
 create mode 100644 projects/llm_framework/include/fst/rmepsilon.h
 create mode 100644 projects/llm_framework/include/fst/rmfinalepsilon.h
 create mode 100644 projects/llm_framework/include/fst/script/arc-class.h
 create mode 100644 projects/llm_framework/include/fst/script/arciterator-class.h
 create mode 100644 projects/llm_framework/include/fst/script/arcsort.h
 create mode 100644 projects/llm_framework/include/fst/script/arg-packs.h
 create mode 100644 projects/llm_framework/include/fst/script/closure.h
 create mode 100644 projects/llm_framework/include/fst/script/compile-impl.h
 create mode 100644 projects/llm_framework/include/fst/script/compile.h
 create mode 100644 projects/llm_framework/include/fst/script/compose.h
 create mode 100644 projects/llm_framework/include/fst/script/concat.h
 create mode 100644 projects/llm_framework/include/fst/script/connect.h
 create mode 100644 projects/llm_framework/include/fst/script/convert.h
 create mode 100644 projects/llm_framework/include/fst/script/decode.h
 create mode 100644 projects/llm_framework/include/fst/script/determinize.h
 create mode 100644 projects/llm_framework/include/fst/script/difference.h
 create mode 100644 projects/llm_framework/include/fst/script/disambiguate.h
 create mode 100644 projects/llm_framework/include/fst/script/draw-impl.h
 create mode 100644 projects/llm_framework/include/fst/script/draw.h
 create mode 100644 projects/llm_framework/include/fst/script/encode.h
 create mode 100644 projects/llm_framework/include/fst/script/encodemapper-class.h
 create mode 100644 projects/llm_framework/include/fst/script/epsnormalize.h
 create mode 100644 projects/llm_framework/include/fst/script/equal.h
 create mode 100644 projects/llm_framework/include/fst/script/equivalent.h
 create mode 100644 projects/llm_framework/include/fst/script/fst-class.h
 create mode 100644 projects/llm_framework/include/fst/script/fstscript-decl.h
 create mode 100644 projects/llm_framework/include/fst/script/fstscript.h
 create mode 100644 projects/llm_framework/include/fst/script/getters.h
 create mode 100644 projects/llm_framework/include/fst/script/info-impl.h
 create mode 100644 projects/llm_framework/include/fst/script/info.h
 create mode 100644 projects/llm_framework/include/fst/script/intersect.h
 create mode 100644 projects/llm_framework/include/fst/script/invert.h
 create mode 100644 projects/llm_framework/include/fst/script/isomorphic.h
 create mode 100644 projects/llm_framework/include/fst/script/map.h
 create mode 100644 projects/llm_framework/include/fst/script/minimize.h
 create mode 100644 projects/llm_framework/include/fst/script/print-impl.h
 create mode 100644 projects/llm_framework/include/fst/script/print.h
 create mode 100644 projects/llm_framework/include/fst/script/project.h
 create mode 100644 projects/llm_framework/include/fst/script/prune.h
 create mode 100644 projects/llm_framework/include/fst/script/push.h
 create mode 100644 projects/llm_framework/include/fst/script/randequivalent.h
 create mode 100644 projects/llm_framework/include/fst/script/randgen.h
 create mode 100644 projects/llm_framework/include/fst/script/register.h
 create mode 100644 projects/llm_framework/include/fst/script/relabel.h
 create mode 100644 projects/llm_framework/include/fst/script/replace.h
 create mode 100644 projects/llm_framework/include/fst/script/reverse.h
 create mode 100644 projects/llm_framework/include/fst/script/reweight.h
 create mode 100644 projects/llm_framework/include/fst/script/rmepsilon.h
 create mode 100644 projects/llm_framework/include/fst/script/script-impl.h
 create mode 100644 projects/llm_framework/include/fst/script/shortest-distance.h
 create mode 100644 projects/llm_framework/include/fst/script/shortest-path.h
 create mode 100644 projects/llm_framework/include/fst/script/stateiterator-class.h
 create mode 100644 projects/llm_framework/include/fst/script/synchronize.h
 create mode 100644 projects/llm_framework/include/fst/script/text-io.h
 create mode 100644 projects/llm_framework/include/fst/script/topsort.h
 create mode 100644 projects/llm_framework/include/fst/script/union.h
 create mode 100644 projects/llm_framework/include/fst/script/verify.h
 create mode 100644 projects/llm_framework/include/fst/script/weight-class.h
 create mode 100644 projects/llm_framework/include/fst/set-weight.h
 create mode 100644 projects/llm_framework/include/fst/shortest-distance.h
 create mode 100644 projects/llm_framework/include/fst/shortest-path.h
 create mode 100644 projects/llm_framework/include/fst/signed-log-weight.h
 create mode 100644 projects/llm_framework/include/fst/sparse-power-weight.h
 create mode 100644 projects/llm_framework/include/fst/sparse-tuple-weight.h
 create mode 100644 projects/llm_framework/include/fst/state-map.h
 create mode 100644 projects/llm_framework/include/fst/state-reachable.h
 create mode 100644 projects/llm_framework/include/fst/state-table.h
 create mode 100644 projects/llm_framework/include/fst/statesort.h
 create mode 100644 projects/llm_framework/include/fst/string-weight.h
 create mode 100644 projects/llm_framework/include/fst/string.h
 create mode 100644 projects/llm_framework/include/fst/symbol-table-ops.h
 create mode 100644 projects/llm_framework/include/fst/symbol-table.h
 create mode 100644 projects/llm_framework/include/fst/synchronize.h
 create mode 100644 projects/llm_framework/include/fst/test-properties.h
 create mode 100644 projects/llm_framework/include/fst/test/algo_test.h
 create mode 100644 projects/llm_framework/include/fst/test/fst_test.h
 create mode 100644 projects/llm_framework/include/fst/test/rand-fst.h
 create mode 100644 projects/llm_framework/include/fst/test/weight-tester.h
 create mode 100644 projects/llm_framework/include/fst/topsort.h
 create mode 100644 projects/llm_framework/include/fst/tuple-weight.h
 create mode 100644 projects/llm_framework/include/fst/types.h
 create mode 100644 projects/llm_framework/include/fst/union-find.h
 create mode 100644 projects/llm_framework/include/fst/union-weight.h
 create mode 100644 projects/llm_framework/include/fst/union.h
 create mode 100644 projects/llm_framework/include/fst/util.h
 create mode 100644 projects/llm_framework/include/fst/vector-fst.h
 create mode 100644 projects/llm_framework/include/fst/verify.h
 create mode 100644 projects/llm_framework/include/fst/visit.h
 create mode 100644 projects/llm_framework/include/fst/weight.h
 create mode 100644 projects/llm_framework/include/gflags/defines.h
 create mode 100644 projects/llm_framework/include/gflags/gflags.h
 create mode 100644 projects/llm_framework/include/gflags/gflags_completions.h
 create mode 100644 projects/llm_framework/include/gflags/gflags_declare.h
 create mode 100644 projects/llm_framework/include/glog/log_severity.h
 create mode 100644 projects/llm_framework/include/glog/logging.h
 create mode 100644 projects/llm_framework/include/glog/raw_logging.h
 create mode 100644 projects/llm_framework/include/glog/stl_logging.h
 create mode 100644 projects/llm_framework/include/glog/vlog_is_on.h
 create mode 100644 projects/llm_framework/main_melotts/src/runner/processor/CMakeLists.txt
 create mode 100644 projects/llm_framework/main_melotts/src/runner/processor/wetext_processor.cc
 create mode 100644 projects/llm_framework/main_melotts/src/runner/processor/wetext_processor.h
 create mode 100644 projects/llm_framework/main_melotts/src/runner/processor/wetext_token_parser.cc
 create mode 100644 projects/llm_framework/main_melotts/src/runner/processor/wetext_token_parser.h
 create mode 100644 projects/llm_framework/main_melotts/src/runner/utils/CMakeLists.txt
 create mode 100644 projects/llm_framework/main_melotts/src/runner/utils/wetext_flags.h
 create mode 100644 projects/llm_framework/main_melotts/src/runner/utils/wetext_log.h
 create mode 100644 projects/llm_framework/main_melotts/src/runner/utils/wetext_string.cc
 create mode 100644 projects/llm_framework/main_melotts/src/runner/utils/wetext_string.h

diff --git a/projects/llm_framework/include/fst/accumulator.h b/projects/llm_framework/include/fst/accumulator.h
new file mode 100644
index 00000000..5ae19247
--- /dev/null
+++ b/projects/llm_framework/include/fst/accumulator.h
@@ -0,0 +1,903 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Classes to accumulate arc weights. Useful for weight lookahead.
+
+#ifndef FST_ACCUMULATOR_H_
+#define FST_ACCUMULATOR_H_
+
+#include <algorithm>
+#include <functional>
+#include <unordered_map>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/arcfilter.h>
+#include <fst/arcsort.h>
+#include <fst/dfs-visit.h>
+#include <fst/expanded-fst.h>
+#include <fst/replace.h>
+
+namespace fst {
+
+// This class accumulates arc weights using the semiring Plus().
+template <class A>
+class DefaultAccumulator {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  DefaultAccumulator() {}
+
+  DefaultAccumulator(const DefaultAccumulator &acc, bool safe = false) {}
+
+  void Init(const Fst<Arc> &fst, bool copy = false) {}
+
+  void SetState(StateId state) {}
+
+  Weight Sum(Weight w, Weight v) { return Plus(w, v); }
+
+  template <class ArcIter>
+  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) {
+    Adder<Weight> adder(w);  // maintains cumulative sum accurately
+    aiter->Seek(begin);
+    for (auto pos = begin; pos < end; aiter->Next(), ++pos)
+      adder.Add(aiter->Value().weight);
+    return adder.Sum();
+  }
+
+  constexpr bool Error() const { return false; }
+
+ private:
+  DefaultAccumulator &operator=(const DefaultAccumulator &) = delete;
+};
+
+// This class accumulates arc weights using the log semiring Plus() assuming an
+// arc weight has a WeightConvert specialization to and from log64 weights.
+template <class A>
+class LogAccumulator {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  LogAccumulator() {}
+
+  LogAccumulator(const LogAccumulator &acc, bool safe = false) {}
+
+  void Init(const Fst<Arc> &fst, bool copy = false) {}
+
+  void SetState(StateId s) {}
+
+  Weight Sum(Weight w, Weight v) { return LogPlus(w, v); }
+
+  template <class ArcIter>
+  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) {
+    auto sum = w;
+    aiter->Seek(begin);
+    for (auto pos = begin; pos < end; aiter->Next(), ++pos) {
+      sum = LogPlus(sum, aiter->Value().weight);
+    }
+    return sum;
+  }
+
+  constexpr bool Error() const { return false; }
+
+ private:
+  Weight LogPlus(Weight w, Weight v) {
+    if (w == Weight::Zero()) {
+      return v;
+    }
+    const auto f1 = to_log_weight_(w).Value();
+    const auto f2 = to_log_weight_(v).Value();
+    if (f1 > f2) {
+      return to_weight_(Log64Weight(f2 - internal::LogPosExp(f1 - f2)));
+    } else {
+      return to_weight_(Log64Weight(f1 - internal::LogPosExp(f2 - f1)));
+    }
+  }
+
+  const WeightConvert<Weight, Log64Weight> to_log_weight_{};
+  const WeightConvert<Log64Weight, Weight> to_weight_{};
+
+  LogAccumulator &operator=(const LogAccumulator &) = delete;
+};
+
+// Interface for shareable data for fast log accumulator copies. Holds pointers
+// to data only, storage is provided by derived classes.
+class FastLogAccumulatorData {
+ public:
+  FastLogAccumulatorData(int arc_limit, int arc_period)
+      : arc_limit_(arc_limit),
+        arc_period_(arc_period),
+        weights_ptr_(nullptr),
+        num_weights_(0),
+        weight_positions_ptr_(nullptr),
+        num_positions_(0) {}
+
+  virtual ~FastLogAccumulatorData() {}
+
+  // Cummulative weight per state for all states s.t. # of arcs > arc_limit_
+  // with arcs in order. The first element per state is Log64Weight::Zero().
+  const double *Weights() const { return weights_ptr_; }
+
+  int NumWeights() const { return num_weights_; }
+
+  // Maps from state to corresponding beginning weight position in weights_.
+  // osition -1 means no pre-computed weights for that state.
+  const int *WeightPositions() const { return weight_positions_ptr_; }
+
+  int NumPositions() const { return num_positions_; }
+
+  int ArcLimit() const { return arc_limit_; }
+
+  int ArcPeriod() const { return arc_period_; }
+
+  // Returns true if the data object is mutable and supports SetData().
+  virtual bool IsMutable() const = 0;
+
+  // Does not take ownership but may invalidate the contents of weights and
+  // weight_positions.
+  virtual void SetData(std::vector<double> *weights,
+                       std::vector<int> *weight_positions) = 0;
+
+ protected:
+  void Init(int num_weights, const double *weights, int num_positions,
+            const int *weight_positions) {
+    weights_ptr_ = weights;
+    num_weights_ = num_weights;
+    weight_positions_ptr_ = weight_positions;
+    num_positions_ = num_positions;
+  }
+
+ private:
+  const int arc_limit_;
+  const int arc_period_;
+  const double *weights_ptr_;
+  int num_weights_;
+  const int *weight_positions_ptr_;
+  int num_positions_;
+
+  FastLogAccumulatorData(const FastLogAccumulatorData &) = delete;
+  FastLogAccumulatorData &operator=(const FastLogAccumulatorData &) = delete;
+};
+
+// FastLogAccumulatorData with mutable storage; filled by
+// FastLogAccumulator::Init.
+class MutableFastLogAccumulatorData : public FastLogAccumulatorData {
+ public:
+  MutableFastLogAccumulatorData(int arc_limit, int arc_period)
+      : FastLogAccumulatorData(arc_limit, arc_period) {}
+
+  bool IsMutable() const override { return true; }
+
+  void SetData(std::vector<double> *weights,
+               std::vector<int> *weight_positions) override {
+    weights_.swap(*weights);
+    weight_positions_.swap(*weight_positions);
+    Init(weights_.size(), weights_.data(), weight_positions_.size(),
+         weight_positions_.data());
+  }
+
+ private:
+  std::vector<double> weights_;
+  std::vector<int> weight_positions_;
+
+  MutableFastLogAccumulatorData(const MutableFastLogAccumulatorData &) = delete;
+  MutableFastLogAccumulatorData &operator=(
+      const MutableFastLogAccumulatorData &) = delete;
+};
+
+// This class accumulates arc weights using the log semiring Plus() assuming an
+// arc weight has a WeightConvert specialization to and from log64 weights. The
+// member function Init(fst) has to be called to setup pre-computed weight
+// information.
+template <class A>
+class FastLogAccumulator {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  explicit FastLogAccumulator(ssize_t arc_limit = 20, ssize_t arc_period = 10)
+      : to_log_weight_(),
+        to_weight_(),
+        arc_limit_(arc_limit),
+        arc_period_(arc_period),
+        data_(std::make_shared<MutableFastLogAccumulatorData>(arc_limit,
+                                                              arc_period)),
+        state_weights_(nullptr),
+        error_(false) {}
+
+  explicit FastLogAccumulator(std::shared_ptr<FastLogAccumulatorData> data)
+      : to_log_weight_(),
+        to_weight_(),
+        arc_limit_(data->ArcLimit()),
+        arc_period_(data->ArcPeriod()),
+        data_(data),
+        state_weights_(nullptr),
+        error_(false) {}
+
+  FastLogAccumulator(const FastLogAccumulator<Arc> &acc, bool safe = false)
+      : to_log_weight_(),
+        to_weight_(),
+        arc_limit_(acc.arc_limit_),
+        arc_period_(acc.arc_period_),
+        data_(acc.data_),
+        state_weights_(nullptr),
+        error_(acc.error_) {}
+
+  void SetState(StateId s) {
+    const auto *weights = data_->Weights();
+    const auto *weight_positions = data_->WeightPositions();
+    state_weights_ = nullptr;
+    if (s < data_->NumPositions()) {
+      const auto pos = weight_positions[s];
+      if (pos >= 0) state_weights_ = &(weights[pos]);
+    }
+  }
+
+  Weight Sum(Weight w, Weight v) const { return LogPlus(w, v); }
+
+  template <class ArcIter>
+  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) const {
+    if (error_) return Weight::NoWeight();
+    auto sum = w;
+    // Finds begin and end of pre-stored weights.
+    ssize_t index_begin = -1;
+    ssize_t index_end = -1;
+    ssize_t stored_begin = end;
+    ssize_t stored_end = end;
+    if (state_weights_) {
+      index_begin = begin > 0 ? (begin - 1) / arc_period_ + 1 : 0;
+      index_end = end / arc_period_;
+      stored_begin = index_begin * arc_period_;
+      stored_end = index_end * arc_period_;
+    }
+    // Computes sum before pre-stored weights.
+    if (begin < stored_begin) {
+      const auto pos_end = std::min(stored_begin, end);
+      aiter->Seek(begin);
+      for (auto pos = begin; pos < pos_end; aiter->Next(), ++pos) {
+        sum = LogPlus(sum, aiter->Value().weight);
+      }
+    }
+    // Computes sum between pre-stored weights.
+    if (stored_begin < stored_end) {
+      const auto f1 = state_weights_[index_end];
+      const auto f2 = state_weights_[index_begin];
+      if (f1 < f2) sum = LogPlus(sum, LogMinus(f1, f2));
+      // Commented out for efficiency; adds Zero().
+      /*
+      else {
+        // explicitly computes if cumulative sum lacks precision
+        aiter->Seek(stored_begin);
+        for (auto pos = stored_begin; pos < stored_end; aiter->Next(), ++pos)
+          sum = LogPlus(sum, aiter->Value().weight);
+      }
+      */
+    }
+    // Computes sum after pre-stored weights.
+    if (stored_end < end) {
+      const auto pos_start = std::max(stored_begin, stored_end);
+      aiter->Seek(pos_start);
+      for (auto pos = pos_start; pos < end; aiter->Next(), ++pos) {
+        sum = LogPlus(sum, aiter->Value().weight);
+      }
+    }
+    return sum;
+  }
+
+  template <class FST>
+  void Init(const FST &fst, bool copy = false) {
+    if (copy || !data_->IsMutable()) return;
+    if (data_->NumPositions() != 0 || arc_limit_ < arc_period_) {
+      FSTERROR() << "FastLogAccumulator: Initialization error";
+      error_ = true;
+      return;
+    }
+    std::vector<double> weights;
+    std::vector<int> weight_positions;
+    weight_positions.reserve(CountStates(fst));
+    for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
+      const auto s = siter.Value();
+      if (fst.NumArcs(s) >= arc_limit_) {
+        auto sum = FloatLimits<double>::PosInfinity();
+        if (weight_positions.size() <= s) weight_positions.resize(s + 1, -1);
+        weight_positions[s] = weights.size();
+        weights.push_back(sum);
+        size_t narcs = 0;
+        ArcIterator<FST> aiter(fst, s);
+        aiter.SetFlags(kArcWeightValue | kArcNoCache, kArcFlags);
+        for (; !aiter.Done(); aiter.Next()) {
+          const auto &arc = aiter.Value();
+          sum = LogPlus(sum, arc.weight);
+          // Stores cumulative weight distribution per arc_period_.
+          if (++narcs % arc_period_ == 0) weights.push_back(sum);
+        }
+      }
+    }
+    data_->SetData(&weights, &weight_positions);
+  }
+
+  bool Error() const { return error_; }
+
+  std::shared_ptr<FastLogAccumulatorData> GetData() const { return data_; }
+
+ private:
+  static double LogPosExp(double x) {
+    return x == FloatLimits<double>::PosInfinity() ? 0.0
+                                                   : log(1.0F + exp(-x));
+  }
+
+  static double LogMinusExp(double x) {
+    return x == FloatLimits<double>::PosInfinity() ? 0.0
+                                                   : log(1.0F - exp(-x));
+  }
+
+  Weight LogPlus(Weight w, Weight v) const {
+    if (w == Weight::Zero()) {
+      return v;
+    }
+    const auto f1 = to_log_weight_(w).Value();
+    const auto f2 = to_log_weight_(v).Value();
+    if (f1 > f2) {
+      return to_weight_(Log64Weight(f2 - LogPosExp(f1 - f2)));
+    } else {
+      return to_weight_(Log64Weight(f1 - LogPosExp(f2 - f1)));
+    }
+  }
+
+  double LogPlus(double f1, Weight v) const {
+    const auto f2 = to_log_weight_(v).Value();
+    if (f1 == FloatLimits<double>::PosInfinity()) {
+      return f2;
+    } else if (f1 > f2) {
+      return f2 - LogPosExp(f1 - f2);
+    } else {
+      return f1 - LogPosExp(f2 - f1);
+    }
+  }
+
+  // Assumes f1 < f2.
+  Weight LogMinus(double f1, double f2) const {
+    if (f2 == FloatLimits<double>::PosInfinity()) {
+      return to_weight_(Log64Weight(f1));
+    } else {
+      return to_weight_(Log64Weight(f1 - LogMinusExp(f2 - f1)));
+    }
+  }
+
+  const WeightConvert<Weight, Log64Weight> to_log_weight_{};
+  const WeightConvert<Log64Weight, Weight> to_weight_{};
+  const ssize_t arc_limit_;   // Minimum number of arcs to pre-compute state.
+  const ssize_t arc_period_;  // Saves cumulative weights per arc_period_.
+  std::shared_ptr<FastLogAccumulatorData> data_;
+  const double *state_weights_;
+  bool error_;
+
+  FastLogAccumulator &operator=(const FastLogAccumulator &) = delete;
+};
+
+// Stores shareable data for cache log accumulator copies. All copies share the
+// same cache.
+template <class Arc>
+class CacheLogAccumulatorData {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  CacheLogAccumulatorData(bool gc, size_t gc_limit)
+      : cache_gc_(gc), cache_limit_(gc_limit), cache_size_(0) {}
+
+  CacheLogAccumulatorData(const CacheLogAccumulatorData<Arc> &data)
+      : cache_gc_(data.cache_gc_),
+        cache_limit_(data.cache_limit_),
+        cache_size_(0) {}
+
+  bool CacheDisabled() const { return cache_gc_ && cache_limit_ == 0; }
+
+  std::vector<double> *GetWeights(StateId s) {
+    auto it = cache_.find(s);
+    if (it != cache_.end()) {
+      it->second.recent = true;
+      return it->second.weights.get();
+    } else {
+      return nullptr;
+    }
+  }
+
+  void AddWeights(StateId s, std::vector<double> *weights) {
+    if (cache_gc_ && cache_size_ >= cache_limit_) GC(false);
+    cache_.insert(std::make_pair(s, CacheState(weights, true)));
+    if (cache_gc_) cache_size_ += weights->capacity() * sizeof(double);
+  }
+
+ private:
+  // Cached information for a given state.
+  struct CacheState {
+    std::unique_ptr<std::vector<double>> weights;  // Accumulated weights.
+    bool recent;  // Has this state been accessed since last GC?
+
+    CacheState(std::vector<double> *weights, bool recent)
+        : weights(weights), recent(recent) {}
+  };
+
+  // Garbage collect: Deletes from cache states that have not been accessed
+  // since the last GC ('free_recent = false') until 'cache_size_' is 2/3 of
+  // 'cache_limit_'. If it does not free enough memory, start deleting
+  // recently accessed states.
+  void GC(bool free_recent) {
+    auto cache_target = (2 * cache_limit_) / 3 + 1;
+    auto it = cache_.begin();
+    while (it != cache_.end() && cache_size_ > cache_target) {
+      auto &cs = it->second;
+      if (free_recent || !cs.recent) {
+        cache_size_ -= cs.weights->capacity() * sizeof(double);
+        cache_.erase(it++);
+      } else {
+        cs.recent = false;
+        ++it;
+      }
+    }
+    if (!free_recent && cache_size_ > cache_target) GC(true);
+  }
+
+  std::unordered_map<StateId, CacheState> cache_;  // Cache.
+  bool cache_gc_;       // Enables garbage collection.
+  size_t cache_limit_;  // # of bytes cached.
+  size_t cache_size_;   // # of bytes allowed before GC.
+
+  CacheLogAccumulatorData &operator=(const CacheLogAccumulatorData &) = delete;
+};
+
+// This class accumulates arc weights using the log semiring Plus() has a
+// WeightConvert specialization to and from log64 weights. It is similar to the
+// FastLogAccumator. However here, the accumulated weights are pre-computed and
+// stored only for the states that are visited. The member function Init(fst)
+// has to be called to setup this accumulator.
+template <class Arc>
+class CacheLogAccumulator {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  explicit CacheLogAccumulator(ssize_t arc_limit = 10, bool gc = false,
+                               size_t gc_limit = 10 * 1024 * 1024)
+      : arc_limit_(arc_limit),
+        data_(std::make_shared<CacheLogAccumulatorData<Arc>>(gc, gc_limit)),
+        s_(kNoStateId),
+        error_(false) {}
+
+  CacheLogAccumulator(const CacheLogAccumulator<Arc> &acc, bool safe = false)
+      : arc_limit_(acc.arc_limit_),
+        fst_(acc.fst_ ? acc.fst_->Copy() : nullptr),
+        data_(safe ? std::make_shared<CacheLogAccumulatorData<Arc>>(*acc.data_)
+                   : acc.data_),
+        s_(kNoStateId),
+        error_(acc.error_) {}
+
+  // Argument arc_limit specifies the minimum number of arcs to pre-compute.
+  void Init(const Fst<Arc> &fst, bool copy = false) {
+    if (!copy && fst_) {
+      FSTERROR() << "CacheLogAccumulator: Initialization error";
+      error_ = true;
+      return;
+    }
+    fst_.reset(fst.Copy());
+  }
+
+  void SetState(StateId s, int depth = 0) {
+    if (s == s_) return;
+    s_ = s;
+    if (data_->CacheDisabled() || error_) {
+      weights_ = nullptr;
+      return;
+    }
+    if (!fst_) {
+      FSTERROR() << "CacheLogAccumulator::SetState: Incorrectly initialized";
+      error_ = true;
+      weights_ = nullptr;
+      return;
+    }
+    weights_ = data_->GetWeights(s);
+    if ((weights_ == nullptr) && (fst_->NumArcs(s) >= arc_limit_)) {
+      weights_ = new std::vector<double>;
+      weights_->reserve(fst_->NumArcs(s) + 1);
+      weights_->push_back(FloatLimits<double>::PosInfinity());
+      data_->AddWeights(s, weights_);
+    }
+  }
+
+  Weight Sum(Weight w, Weight v) { return LogPlus(w, v); }
+
+  template <class ArcIter>
+  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) {
+    if (weights_ == nullptr) {
+      auto sum = w;
+      aiter->Seek(begin);
+      for (auto pos = begin; pos < end; aiter->Next(), ++pos) {
+        sum = LogPlus(sum, aiter->Value().weight);
+      }
+      return sum;
+    } else {
+      Extend(end, aiter);
+      const auto &f1 = (*weights_)[end];
+      const auto &f2 = (*weights_)[begin];
+      if (f1 < f2) {
+        return LogPlus(w, LogMinus(f1, f2));
+      } else {
+        // Commented out for efficiency; adds Zero().
+        /*
+        auto sum = w;
+        // Explicitly computes if cumulative sum lacks precision.
+        aiter->Seek(begin);
+        for (auto pos = begin; pos < end; aiter->Next(), ++pos) {
+          sum = LogPlus(sum, aiter->Value().weight);
+        }
+        return sum;
+        */
+        return w;
+      }
+    }
+  }
+
+  // Returns first position from aiter->Position() whose accumulated
+  // value is greater or equal to w (w.r.t. Zero() < One()). The
+  // iterator may be repositioned.
+  template <class ArcIter>
+  size_t LowerBound(Weight w, ArcIter *aiter) {
+    const auto f = to_log_weight_(w).Value();
+    auto pos = aiter->Position();
+    if (weights_) {
+      Extend(fst_->NumArcs(s_), aiter);
+      return std::lower_bound(weights_->begin() + pos + 1, weights_->end(),
+                              f, std::greater<double>()) -
+          weights_->begin() - 1;
+    } else {
+      size_t n = 0;
+      auto x = FloatLimits<double>::PosInfinity();
+      for (aiter->Reset(); !aiter->Done(); aiter->Next(), ++n) {
+        x = LogPlus(x, aiter->Value().weight);
+        if (n >= pos && x <= f) break;
+      }
+      return n;
+    }
+  }
+
+  bool Error() const { return error_; }
+
+ private:
+  double LogPosExp(double x) {
+    return x == FloatLimits<double>::PosInfinity() ? 0.0
+                                                   : log(1.0F + exp(-x));
+  }
+
+  double LogMinusExp(double x) {
+    return x == FloatLimits<double>::PosInfinity() ? 0.0
+                                                   : log(1.0F - exp(-x));
+  }
+
+  Weight LogPlus(Weight w, Weight v) {
+    if (w == Weight::Zero()) {
+      return v;
+    }
+    const auto f1 = to_log_weight_(w).Value();
+    const auto f2 = to_log_weight_(v).Value();
+    if (f1 > f2) {
+      return to_weight_(Log64Weight(f2 - LogPosExp(f1 - f2)));
+    } else {
+      return to_weight_(Log64Weight(f1 - LogPosExp(f2 - f1)));
+    }
+  }
+
+  double LogPlus(double f1, Weight v) {
+    const auto f2 = to_log_weight_(v).Value();
+    if (f1 == FloatLimits<double>::PosInfinity()) {
+      return f2;
+    } else if (f1 > f2) {
+      return f2 - LogPosExp(f1 - f2);
+    } else {
+      return f1 - LogPosExp(f2 - f1);
+    }
+  }
+
+  // Assumes f1 < f2.
+  Weight LogMinus(double f1, double f2) {
+    if (f2 == FloatLimits<double>::PosInfinity()) {
+      return to_weight_(Log64Weight(f1));
+    } else {
+      return to_weight_(Log64Weight(f1 - LogMinusExp(f2 - f1)));
+    }
+  }
+
+  // Extends weights up to index 'end'.
+  template <class ArcIter>
+  void Extend(ssize_t end, ArcIter *aiter) {
+    if (weights_->size() <= end) {
+      for (aiter->Seek(weights_->size() - 1); weights_->size() <= end;
+           aiter->Next()) {
+        weights_->push_back(LogPlus(weights_->back(), aiter->Value().weight));
+      }
+    }
+  }
+
+
+  const WeightConvert<Weight, Log64Weight> to_log_weight_{};
+  const WeightConvert<Log64Weight, Weight> to_weight_{};
+  ssize_t arc_limit_;                    // Minimum # of arcs to cache a state.
+  std::vector<double> *weights_;         // Accumulated weights for cur. state.
+  std::unique_ptr<const Fst<Arc>> fst_;  // Input FST.
+  std::shared_ptr<CacheLogAccumulatorData<Arc>> data_;  // Cache data.
+  StateId s_;                                           // Current state.
+  bool error_;
+};
+
+// Stores shareable data for replace accumulator copies.
+template <class Accumulator, class T>
+class ReplaceAccumulatorData {
+ public:
+  using Arc = typename Accumulator::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using StateTable = T;
+  using StateTuple = typename StateTable::StateTuple;
+
+  ReplaceAccumulatorData() : state_table_(nullptr) {}
+
+  explicit ReplaceAccumulatorData(
+      const std::vector<Accumulator *> &accumulators)
+      : state_table_(nullptr) {
+    accumulators_.reserve(accumulators.size());
+    for (const auto accumulator : accumulators) {
+      accumulators_.emplace_back(accumulator);
+    }
+  }
+
+  void Init(const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_tuples,
+            const StateTable *state_table) {
+    state_table_ = state_table;
+    accumulators_.resize(fst_tuples.size());
+    for (Label i = 0; i < accumulators_.size(); ++i) {
+      if (!accumulators_[i]) {
+        accumulators_[i].reset(new Accumulator());
+        accumulators_[i]->Init(*(fst_tuples[i].second));
+      }
+      fst_array_.emplace_back(fst_tuples[i].second->Copy());
+    }
+  }
+
+  const StateTuple &GetTuple(StateId s) const { return state_table_->Tuple(s); }
+
+  Accumulator *GetAccumulator(size_t i) { return accumulators_[i].get(); }
+
+  const Fst<Arc> *GetFst(size_t i) const { return fst_array_[i].get(); }
+
+ private:
+  const StateTable *state_table_;
+  std::vector<std::unique_ptr<Accumulator>> accumulators_;
+  std::vector<std::unique_ptr<const Fst<Arc>>> fst_array_;
+};
+
+// This class accumulates weights in a ReplaceFst.  The 'Init' method takes as
+// input the argument used to build the ReplaceFst and the ReplaceFst state
+// table. It uses accumulators of type 'Accumulator' in the underlying FSTs.
+template <class Accumulator,
+          class T = DefaultReplaceStateTable<typename Accumulator::Arc>>
+class ReplaceAccumulator {
+ public:
+  using Arc = typename Accumulator::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using StateTable = T;
+  using StateTuple = typename StateTable::StateTuple;
+  using Weight = typename Arc::Weight;
+
+  ReplaceAccumulator()
+      : init_(false),
+        data_(std::make_shared<
+              ReplaceAccumulatorData<Accumulator, StateTable>>()),
+        error_(false) {}
+
+  explicit ReplaceAccumulator(const std::vector<Accumulator *> &accumulators)
+      : init_(false),
+        data_(std::make_shared<ReplaceAccumulatorData<Accumulator, StateTable>>(
+            accumulators)),
+        error_(false) {}
+
+  ReplaceAccumulator(const ReplaceAccumulator<Accumulator, StateTable> &acc,
+                     bool safe = false)
+      : init_(acc.init_), data_(acc.data_), error_(acc.error_) {
+    if (!init_) {
+      FSTERROR() << "ReplaceAccumulator: Can't copy unintialized accumulator";
+    }
+    if (safe) FSTERROR() << "ReplaceAccumulator: Safe copy not supported";
+  }
+
+  // Does not take ownership of the state table, the state table is owned by
+  // the ReplaceFst.
+  void Init(const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_tuples,
+            const StateTable *state_table) {
+    init_ = true;
+    data_->Init(fst_tuples, state_table);
+  }
+
+  // Method required by LookAheadMatcher. However, ReplaceAccumulator needs to
+  // be initialized by calling the Init method above before being passed to
+  // LookAheadMatcher.
+  //
+  // TODO(allauzen): Revisit this. Consider creating a method
+  // Init(const ReplaceFst<A, T, C>&, bool) and using friendship to get access
+  // to the innards of ReplaceFst.
+  void Init(const Fst<Arc> &fst, bool copy = false) {
+    if (!init_) {
+      FSTERROR() << "ReplaceAccumulator::Init: Accumulator needs to be"
+                 << " initialized before being passed to LookAheadMatcher";
+      error_ = true;
+    }
+  }
+
+  void SetState(StateId s) {
+    if (!init_) {
+      FSTERROR() << "ReplaceAccumulator::SetState: Incorrectly initialized";
+      error_ = true;
+      return;
+    }
+    auto tuple = data_->GetTuple(s);
+    fst_id_ = tuple.fst_id - 1;  // Replace FST ID is 1-based.
+    data_->GetAccumulator(fst_id_)->SetState(tuple.fst_state);
+    if ((tuple.prefix_id != 0) &&
+        (data_->GetFst(fst_id_)->Final(tuple.fst_state) != Weight::Zero())) {
+      offset_ = 1;
+      offset_weight_ = data_->GetFst(fst_id_)->Final(tuple.fst_state);
+    } else {
+      offset_ = 0;
+      offset_weight_ = Weight::Zero();
+    }
+    aiter_.reset(
+        new ArcIterator<Fst<Arc>>(*data_->GetFst(fst_id_), tuple.fst_state));
+  }
+
+  Weight Sum(Weight w, Weight v) {
+    if (error_) return Weight::NoWeight();
+    return data_->GetAccumulator(fst_id_)->Sum(w, v);
+  }
+
+  template <class ArcIter>
+  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) {
+    if (error_) return Weight::NoWeight();
+    auto sum = begin == end ? Weight::Zero()
+                            : data_->GetAccumulator(fst_id_)->Sum(
+                                  w, aiter_.get(), begin ? begin - offset_ : 0,
+                                  end - offset_);
+    if (begin == 0 && end != 0 && offset_ > 0) sum = Sum(offset_weight_, sum);
+    return sum;
+  }
+
+  bool Error() const { return error_; }
+
+ private:
+  bool init_;
+  std::shared_ptr<ReplaceAccumulatorData<Accumulator, StateTable>> data_;
+  Label fst_id_;
+  size_t offset_;
+  Weight offset_weight_;
+  std::unique_ptr<ArcIterator<Fst<Arc>>> aiter_;
+  bool error_;
+};
+
+// SafeReplaceAccumulator accumulates weights in a ReplaceFst and copies of it
+// are always thread-safe copies.
+template <class Accumulator, class T>
+class SafeReplaceAccumulator {
+ public:
+  using Arc = typename Accumulator::Arc;
+  using StateId = typename Arc::StateId;
+  using Label = typename Arc::Label;
+  using Weight = typename Arc::Weight;
+  using StateTable = T;
+  using StateTuple = typename StateTable::StateTuple;
+
+  SafeReplaceAccumulator() {}
+
+  SafeReplaceAccumulator(const SafeReplaceAccumulator &copy, bool safe)
+      : SafeReplaceAccumulator(copy) {}
+
+  explicit SafeReplaceAccumulator(
+      const std::vector<Accumulator> &accumulators) {
+    for (const auto &accumulator : accumulators) {
+      accumulators_.emplace_back(accumulator, true);
+    }
+  }
+
+  void Init(const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_tuples,
+            const StateTable *state_table) {
+    state_table_ = state_table;
+    for (Label i = 0; i < fst_tuples.size(); ++i) {
+      if (i == accumulators_.size()) {
+        accumulators_.resize(accumulators_.size() + 1);
+        accumulators_[i].Init(*(fst_tuples[i].second));
+      }
+      fst_array_.emplace_back(fst_tuples[i].second->Copy(true));
+    }
+    init_ = true;
+  }
+
+  void Init(const Fst<Arc> &fst, bool copy = false) {
+    if (!init_) {
+      FSTERROR() << "SafeReplaceAccumulator::Init: Accumulator needs to be"
+                 << " initialized before being passed to LookAheadMatcher";
+      error_ = true;
+    }
+  }
+
+  void SetState(StateId s) {
+    auto tuple = state_table_->Tuple(s);
+    fst_id_ = tuple.fst_id - 1;  // Replace FST ID is 1-based
+    GetAccumulator(fst_id_)->SetState(tuple.fst_state);
+    offset_ = 0;
+    offset_weight_ = Weight::Zero();
+    const auto final_weight = GetFst(fst_id_)->Final(tuple.fst_state);
+    if ((tuple.prefix_id != 0) && (final_weight != Weight::Zero())) {
+      offset_ = 1;
+      offset_weight_ = final_weight;
+    }
+    aiter_.Set(*GetFst(fst_id_), tuple.fst_state);
+  }
+
+  Weight Sum(Weight w, Weight v) {
+    if (error_) return Weight::NoWeight();
+    return GetAccumulator(fst_id_)->Sum(w, v);
+  }
+
+  template <class ArcIter>
+  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) {
+    if (error_) return Weight::NoWeight();
+    if (begin == end) return Weight::Zero();
+    auto sum = GetAccumulator(fst_id_)->Sum(
+        w, aiter_.get(), begin ? begin - offset_ : 0, end - offset_);
+    if (begin == 0 && end != 0 && offset_ > 0) {
+      sum = Sum(offset_weight_, sum);
+    }
+    return sum;
+  }
+
+  bool Error() const { return error_; }
+
+ private:
+  class ArcIteratorPtr {
+   public:
+    ArcIteratorPtr() {}
+
+    ArcIteratorPtr(const ArcIteratorPtr &copy) {}
+
+    void Set(const Fst<Arc> &fst, StateId state_id) {
+      ptr_.reset(new ArcIterator<Fst<Arc>>(fst, state_id));
+    }
+
+    ArcIterator<Fst<Arc>> *get() { return ptr_.get(); }
+
+   private:
+    std::unique_ptr<ArcIterator<Fst<Arc>>> ptr_;
+  };
+
+  Accumulator *GetAccumulator(size_t i) { return &accumulators_[i]; }
+
+  const Fst<Arc> *GetFst(size_t i) const { return fst_array_[i].get(); }
+
+  const StateTable *state_table_;
+  std::vector<Accumulator> accumulators_;
+  std::vector<std::shared_ptr<Fst<Arc>>> fst_array_;
+  ArcIteratorPtr aiter_;
+  bool init_ = false;
+  bool error_ = false;
+  Label fst_id_;
+  size_t offset_;
+  Weight offset_weight_;
+};
+
+}  // namespace fst
+
+#endif  // FST_ACCUMULATOR_H_
diff --git a/projects/llm_framework/include/fst/add-on.h b/projects/llm_framework/include/fst/add-on.h
new file mode 100644
index 00000000..4a95111f
--- /dev/null
+++ b/projects/llm_framework/include/fst/add-on.h
@@ -0,0 +1,248 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// FST implementation class to attach an arbitrary object with a read/write
+// method to an FST and its file representation. The FST is given a new type
+// name.
+
+#ifndef FST_ADD_ON_H_
+#define FST_ADD_ON_H_
+
+#include <stddef.h>
+#include <memory>
+#include <string>
+#include <utility>
+
+#include <fst/log.h>
+
+#include <fst/fst.h>
+
+
+namespace fst {
+
+// Identifies stream data as an add-on FST.
+static constexpr int32 kAddOnMagicNumber = 446681434;
+
+// Nothing to save.
+class NullAddOn {
+ public:
+  NullAddOn() {}
+
+  static NullAddOn *Read(std::istream &strm, const FstReadOptions &opts) {
+    return new NullAddOn();
+  }
+
+  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
+    return true;
+  }
+};
+
+// Create a new add-on from a pair of add-ons.
+template <class A1, class A2>
+class AddOnPair {
+ public:
+  // Argument reference count incremented.
+  AddOnPair(std::shared_ptr<A1> a1, std::shared_ptr<A2> a2)
+      : a1_(std::move(a1)), a2_(std::move(a2)) {}
+
+  const A1 *First() const { return a1_.get(); }
+
+  const A2 *Second() const { return a2_.get(); }
+
+  std::shared_ptr<A1> SharedFirst() const { return a1_; }
+
+  std::shared_ptr<A2> SharedSecond() const { return a2_; }
+
+  static AddOnPair<A1, A2> *Read(std::istream &istrm,
+                                 const FstReadOptions &opts) {
+    A1 *a1 = nullptr;
+    bool have_addon1 = false;
+    ReadType(istrm, &have_addon1);
+    if (have_addon1) a1 = A1::Read(istrm, opts);
+
+    A2 *a2 = nullptr;
+    bool have_addon2 = false;
+    ReadType(istrm, &have_addon2);
+    if (have_addon2) a2 = A2::Read(istrm, opts);
+
+    return new AddOnPair<A1, A2>(std::shared_ptr<A1>(a1),
+                                 std::shared_ptr<A2>(a2));
+  }
+
+  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
+    bool have_addon1 = a1_ != nullptr;
+    WriteType(ostrm, have_addon1);
+    if (have_addon1) a1_->Write(ostrm, opts);
+    bool have_addon2 = a2_ != nullptr;
+    WriteType(ostrm, have_addon2);
+    if (have_addon2) a2_->Write(ostrm, opts);
+    return true;
+  }
+
+ private:
+  std::shared_ptr<A1> a1_;
+  std::shared_ptr<A2> a2_;
+};
+
+namespace internal {
+
+// Adds an object of type T to an FST. T must support:
+//
+//     T* Read(std::istream &);
+//     bool Write(std::ostream &);
+//
+// The resulting type is a new FST implementation.
+template <class FST, class T>
+class AddOnImpl : public FstImpl<typename FST::Arc> {
+ public:
+  using Arc = typename FST::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FstImpl<Arc>::SetType;
+  using FstImpl<Arc>::SetInputSymbols;
+  using FstImpl<Arc>::SetOutputSymbols;
+  using FstImpl<Arc>::SetProperties;
+  using FstImpl<Arc>::WriteHeader;
+
+  // We make a thread-safe copy of the FST by default since an FST
+  // implementation is expected to not share mutable data between objects.
+  AddOnImpl(const FST &fst, const string &type,
+            std::shared_ptr<T> t = std::shared_ptr<T>())
+      : fst_(fst, true), t_(std::move(t)) {
+    SetType(type);
+    SetProperties(fst_.Properties(kFstProperties, false));
+    SetInputSymbols(fst_.InputSymbols());
+    SetOutputSymbols(fst_.OutputSymbols());
+  }
+
+  // Conversion from const Fst<Arc> & to F always copies the underlying
+  // implementation.
+  AddOnImpl(const Fst<Arc> &fst, const string &type,
+            std::shared_ptr<T> t = std::shared_ptr<T>())
+      : fst_(fst), t_(std::move(t)) {
+    SetType(type);
+    SetProperties(fst_.Properties(kFstProperties, false));
+    SetInputSymbols(fst_.InputSymbols());
+    SetOutputSymbols(fst_.OutputSymbols());
+  }
+
+  // We make a thread-safe copy of the FST by default since an FST
+  // implementation is expected to not share mutable data between objects.
+  AddOnImpl(const AddOnImpl<FST, T> &impl)
+      : fst_(impl.fst_, true), t_(impl.t_) {
+    SetType(impl.Type());
+    SetProperties(fst_.Properties(kCopyProperties, false));
+    SetInputSymbols(fst_.InputSymbols());
+    SetOutputSymbols(fst_.OutputSymbols());
+  }
+
+  StateId Start() const { return fst_.Start(); }
+
+  Weight Final(StateId s) const { return fst_.Final(s); }
+
+  size_t NumArcs(StateId s) const { return fst_.NumArcs(s); }
+
+  size_t NumInputEpsilons(StateId s) const { return fst_.NumInputEpsilons(s); }
+
+  size_t NumOutputEpsilons(StateId s) const {
+    return fst_.NumOutputEpsilons(s);
+  }
+
+  size_t NumStates() const { return fst_.NumStates(); }
+
+  static AddOnImpl<FST, T> *Read(std::istream &strm,
+                                 const FstReadOptions &opts) {
+    FstReadOptions nopts(opts);
+    FstHeader hdr;
+    if (!nopts.header) {
+      hdr.Read(strm, nopts.source);
+      nopts.header = &hdr;
+    }
+    std::unique_ptr<AddOnImpl<FST, T>> impl(
+        new AddOnImpl<FST, T>(nopts.header->FstType()));
+    if (!impl->ReadHeader(strm, nopts, kMinFileVersion, &hdr)) return nullptr;
+    impl.reset();
+    int32 magic_number = 0;
+    ReadType(strm, &magic_number);  // Ensures this is an add-on FST.
+    if (magic_number != kAddOnMagicNumber) {
+      LOG(ERROR) << "AddOnImpl::Read: Bad add-on header: " << nopts.source;
+      return nullptr;
+    }
+    FstReadOptions fopts(opts);
+    fopts.header = nullptr;  // Contained header was written out.
+    std::unique_ptr<FST> fst(FST::Read(strm, fopts));
+    if (!fst) return nullptr;
+    std::shared_ptr<T> t;
+    bool have_addon = false;
+    ReadType(strm, &have_addon);
+    if (have_addon) {  // Reads add-on object if present.
+      t = std::shared_ptr<T>(T::Read(strm, fopts));
+      if (!t) return nullptr;
+    }
+    return new AddOnImpl<FST, T>(*fst, nopts.header->FstType(), t);
+  }
+
+  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
+    FstHeader hdr;
+    FstWriteOptions nopts(opts);
+    nopts.write_isymbols = false;  // Allows contained FST to hold any symbols.
+    nopts.write_osymbols = false;
+    WriteHeader(strm, nopts, kFileVersion, &hdr);
+    WriteType(strm, kAddOnMagicNumber);  // Ensures this is an add-on FST.
+    FstWriteOptions fopts(opts);
+    fopts.write_header = true;  // Forces writing contained header.
+    if (!fst_.Write(strm, fopts)) return false;
+    bool have_addon = !!t_;
+    WriteType(strm, have_addon);
+    // Writes add-on object if present.
+    if (have_addon) t_->Write(strm, opts);
+    return true;
+  }
+
+  void InitStateIterator(StateIteratorData<Arc> *data) const {
+    fst_.InitStateIterator(data);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const {
+    fst_.InitArcIterator(s, data);
+  }
+
+  FST &GetFst() { return fst_; }
+
+  const FST &GetFst() const { return fst_; }
+
+  const T *GetAddOn() const { return t_.get(); }
+
+  std::shared_ptr<T> GetSharedAddOn() const { return t_; }
+
+  void SetAddOn(std::shared_ptr<T> t) { t_ = t; }
+
+ private:
+  explicit AddOnImpl(const string &type) : t_() {
+    SetType(type);
+    SetProperties(kExpanded);
+  }
+
+  // Current file format version.
+  static constexpr int kFileVersion = 1;
+  // Minimum file format version supported.
+  static constexpr int kMinFileVersion = 1;
+
+  FST fst_;
+  std::shared_ptr<T> t_;
+
+  AddOnImpl &operator=(const AddOnImpl &) = delete;
+};
+
+template <class FST, class T>
+constexpr int AddOnImpl<FST, T>::kFileVersion;
+
+template <class FST, class T>
+constexpr int AddOnImpl<FST, T>::kMinFileVersion;
+
+}  // namespace internal
+}  // namespace fst
+
+#endif  // FST_ADD_ON_H_
diff --git a/projects/llm_framework/include/fst/arc-arena.h b/projects/llm_framework/include/fst/arc-arena.h
new file mode 100644
index 00000000..13fe918a
--- /dev/null
+++ b/projects/llm_framework/include/fst/arc-arena.h
@@ -0,0 +1,232 @@
+#ifndef FST_ARC_ARENA_H_
+#define FST_ARC_ARENA_H_
+
+#include <deque>
+#include <memory>
+#include <utility>
+#include <fst/fst.h>
+#include <fst/memory.h>
+#include <unordered_map>
+
+namespace fst {
+
+// ArcArena is used for fast allocation of contiguous arrays of arcs.
+//
+// To create an arc array:
+//   for each state:
+//     for each arc:
+//       arena.PushArc();
+//     // Commits these arcs and returns pointer to them.
+//     Arc *arcs = arena.GetArcs();
+//
+//     OR
+//
+//     arena.DropArcs();  // Throws away current arcs, reuse the space.
+//
+// The arcs returned are guaranteed to be contiguous and the pointer returned
+// will never be invalidated until the arena is cleared for reuse.
+//
+// The contents of the arena can be released with a call to arena.Clear() after
+// which the arena will restart with an initial allocation capable of holding at
+// least all of the arcs requested in the last usage before Clear() making
+// subsequent uses of the Arena more efficient.
+//
+// The max_retained_size option can limit the amount of arc space requested on
+// Clear() to avoid excess growth from intermittent high usage.
+template <typename Arc>
+class ArcArena {
+ public:
+  explicit ArcArena(size_t block_size = 256,
+                    size_t max_retained_size = 1e6)
+      : block_size_(block_size),
+        max_retained_size_(max_retained_size) {
+    blocks_.emplace_back(MakeSharedBlock(block_size_));
+    first_block_size_ = block_size_;
+    total_size_ = block_size_;
+    arcs_ = blocks_.back().get();
+    end_ = arcs_ + block_size_;
+    next_ = arcs_;
+  }
+
+  ArcArena(const ArcArena& copy)
+      : arcs_(copy.arcs_), next_(copy.next_), end_(copy.end_),
+        block_size_(copy.block_size_),
+        first_block_size_(copy.first_block_size_),
+        total_size_(copy.total_size_),
+        max_retained_size_(copy.max_retained_size_),
+        blocks_(copy.blocks_) {
+    NewBlock(block_size_);
+  }
+
+  void ReserveArcs(size_t n) {
+    if (next_ + n < end_) return;
+    NewBlock(n);
+  }
+
+  void PushArc(const Arc& arc) {
+    if (next_ == end_) {
+      size_t length = next_ - arcs_;
+      NewBlock(length * 2);
+    }
+    *next_ = arc;
+    ++next_;
+  }
+
+  const Arc* GetArcs() {
+    const auto *arcs = arcs_;
+    arcs_ = next_;
+    return arcs;
+  }
+
+  void DropArcs() { next_ = arcs_; }
+
+  size_t Size() { return total_size_; }
+
+  void Clear() {
+    blocks_.resize(1);
+    if (total_size_ > first_block_size_) {
+      first_block_size_ = std::min(max_retained_size_, total_size_);
+      blocks_.back() = MakeSharedBlock(first_block_size_);
+    }
+    total_size_ = first_block_size_;
+    arcs_ = blocks_.back().get();
+    end_ = arcs_ + first_block_size_;
+    next_ = arcs_;
+  }
+
+ private:
+  // Allocates a new block with capacity of at least n or block_size,
+  // copying incomplete arc sequence from old block to new block.
+  void NewBlock(size_t n) {
+    const auto length = next_ - arcs_;
+    const auto new_block_size = std::max(n, block_size_);
+    total_size_ += new_block_size;
+    blocks_.emplace_back(MakeSharedBlock(new_block_size));
+    std::copy(arcs_, next_, blocks_.back().get());
+    arcs_ = blocks_.back().get();
+    next_ = arcs_ + length;
+    end_ = arcs_ + new_block_size;
+  }
+
+  std::shared_ptr<Arc> MakeSharedBlock(size_t size) {
+    return std::shared_ptr<Arc>(new Arc[size], std::default_delete<Arc[]>());
+  }
+
+  Arc *arcs_;
+  Arc *next_;
+  const Arc *end_;
+  size_t block_size_;
+  size_t first_block_size_;
+  size_t total_size_;
+  size_t max_retained_size_;
+  std::list<std::shared_ptr<Arc>> blocks_;
+};
+
+// ArcArenaStateStore uses a resusable ArcArena to store arc arrays and does not
+// require that the Expander call ReserveArcs first.
+//
+// TODO(tombagby): Make cache type configurable.
+// TODO(tombagby): Provide ThreadLocal/Concurrent configuration.
+template <class A>
+class ArcArenaStateStore {
+ public:
+  using Arc = A;
+  using Weight = typename Arc::Weight;
+  using StateId = typename Arc::StateId;
+
+  ArcArenaStateStore() : arena_(64 * 1024) {
+  }
+
+  class State {
+   public:
+    Weight Final() const { return final_; }
+
+    size_t NumInputEpsilons() const { return niepsilons_; }
+
+    size_t NumOutputEpsilons() const { return noepsilons_; }
+
+    size_t NumArcs() const { return narcs_; }
+
+    const Arc &GetArc(size_t n) const { return arcs_[n]; }
+
+    const Arc *Arcs() const { return arcs_; }
+
+    int* MutableRefCount() const { return nullptr; }
+
+   private:
+    State(Weight weight, int32 niepsilons, int32 noepsilons, int32 narcs,
+          const Arc *arcs)
+        : final_(std::move(weight)),
+          niepsilons_(niepsilons),
+          noepsilons_(noepsilons),
+          narcs_(narcs),
+          arcs_(arcs) {}
+
+    Weight final_;
+    size_t niepsilons_;
+    size_t noepsilons_;
+    size_t narcs_;
+    const Arc *arcs_;
+
+    friend class ArcArenaStateStore<Arc>;
+  };
+
+  template <class Expander>
+  State *FindOrExpand(Expander &expander, StateId state_id) {  // NOLINT
+    auto it = cache_.insert(std::pair<StateId, State*>(state_id, nullptr));
+    if (!it.second) return it.first->second;
+    // Needs a new state.
+    StateBuilder builder(&arena_);
+    expander.Expand(state_id, &builder);
+    const auto arcs = arena_.GetArcs();
+    size_t narcs = builder.narcs_;
+    size_t niepsilons = 0;
+    size_t noepsilons = 0;
+    for (size_t i = 0; i < narcs; ++i) {
+      if (arcs[i].ilabel == 0) ++niepsilons;
+      if (arcs[i].olabel == 0) ++noepsilons;
+    }
+    states_.emplace_back(
+        State(builder.final_, niepsilons, noepsilons, narcs, arcs));
+    // Places it in the cache.
+    auto state = &states_.back();
+    it.first->second = state;
+    return state;
+  }
+
+  State *Find(StateId state_id) const {
+    auto it = cache_.find(state_id);
+    return (it == cache_.end()) ? nullptr : it->second;
+  }
+
+ private:
+  class StateBuilder {
+   public:
+    explicit StateBuilder(ArcArena<Arc>* arena)
+       : arena_(arena), final_(Weight::Zero()), narcs_(0) {}
+
+    void SetFinal(Weight weight) { final_ = std::move(weight); }
+
+    void ReserveArcs(size_t n) { arena_->ReserveArcs(n); }
+
+    void AddArc(const Arc &arc) {
+      ++narcs_;
+      arena_->PushArc(arc);
+    }
+
+   private:
+    friend class ArcArenaStateStore<Arc>;
+
+    ArcArena<Arc> *arena_;
+    Weight final_;
+    size_t narcs_;
+  };
+
+  std::unordered_map<StateId, State *> cache_;
+  std::deque<State> states_;
+  ArcArena<Arc> arena_;
+};
+
+}  // namespace fst
+
+#endif  // FST_ARC_ARENA_H_
diff --git a/projects/llm_framework/include/fst/arc-map.h b/projects/llm_framework/include/fst/arc-map.h
new file mode 100644
index 00000000..24db4911
--- /dev/null
+++ b/projects/llm_framework/include/fst/arc-map.h
@@ -0,0 +1,1285 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to map over/transform arcs e.g., change semirings or
+// implement project/invert. Consider using when operation does
+// not change the number of arcs (except possibly superfinal arcs).
+
+#ifndef FST_ARC_MAP_H_
+#define FST_ARC_MAP_H_
+
+#include <string>
+#include <unordered_map>
+#include <utility>
+
+#include <fst/log.h>
+
+#include <fst/cache.h>
+#include <fst/mutable-fst.h>
+
+
+namespace fst {
+
+// Determines how final weights are mapped.
+enum MapFinalAction {
+  // A final weight is mapped into a final weight. An error is raised if this
+  // is not possible.
+  MAP_NO_SUPERFINAL,
+  // A final weight is mapped to an arc to the superfinal state when the result
+  // cannot be represented as a final weight. The superfinal state will be
+  // added only if it is needed.
+  MAP_ALLOW_SUPERFINAL,
+  // A final weight is mapped to an arc to the superfinal state unless the
+  // result can be represented as a final weight of weight Zero(). The
+  // superfinal state is always added (if the input is not the empty FST).
+  MAP_REQUIRE_SUPERFINAL
+};
+
+// Determines how symbol tables are mapped.
+enum MapSymbolsAction {
+  // Symbols should be cleared in the result by the map.
+  MAP_CLEAR_SYMBOLS,
+  // Symbols should be copied from the input FST by the map.
+  MAP_COPY_SYMBOLS,
+  // Symbols should not be modified in the result by the map itself.
+  // (They may set by the mapper).
+  MAP_NOOP_SYMBOLS
+};
+
+// The ArcMapper interfaces defines how arcs and final weights are mapped.
+// This is useful for implementing operations that do not change the number of
+// arcs (except possibly superfinal arcs).
+//
+// template <class A, class B>
+// class ArcMapper {
+//  public:
+//   using FromArc = A;
+//   using ToArc = B;
+//
+//   // Maps an arc type FromArc to arc type ToArc.
+//   ToArc operator()(const FromArc &arc);
+//
+//   // Specifies final action the mapper requires (see above).
+//   // The mapper will be passed final weights as arcs of the form
+//   // Arc(0, 0, weight, kNoStateId).
+//   MapFinalAction FinalAction() const;
+//
+//   // Specifies input symbol table action the mapper requires (see above).
+//   MapSymbolsAction InputSymbolsAction() const;
+//
+//   // Specifies output symbol table action the mapper requires (see above).
+//   MapSymbolsAction OutputSymbolsAction() const;
+//
+//   // This specifies the known properties of an FST mapped by this mapper. It
+//   takes as argument the input FSTs's known properties.
+//   uint64 Properties(uint64 props) const;
+// };
+//
+// The ArcMap functions and classes below will use the FinalAction()
+// method of the mapper to determine how to treat final weights, e.g., whether
+// to add a superfinal state. They will use the Properties() method to set the
+// result FST properties.
+//
+// We include a various map versions below. One dimension of variation is
+// whether the mapping mutates its input, writes to a new result FST, or is an
+// on-the-fly FST. Another dimension is how we pass the mapper. We allow passing
+// the mapper by pointer for cases that we need to change the state of the
+// user's mapper.  This is the case with the EncodeMapper, which is reused
+// during decoding. We also include map versions that pass the mapper by value
+// or const reference when this suffices.
+
+// Maps an arc type A using a mapper function object C, passed
+// by pointer.  This version modifies its Fst input.
+template <class A, class C>
+void ArcMap(MutableFst<A> *fst, C *mapper) {
+  using FromArc = A;
+  using ToArc = A;
+  using Weight = typename FromArc::Weight;
+  if (mapper->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
+    fst->SetInputSymbols(nullptr);
+  }
+  if (mapper->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
+    fst->SetOutputSymbols(nullptr);
+  }
+  if (fst->Start() == kNoStateId) return;
+  const auto props = fst->Properties(kFstProperties, false);
+  const auto final_action = mapper->FinalAction();
+  auto superfinal = kNoStateId;
+  if (final_action == MAP_REQUIRE_SUPERFINAL) {
+    superfinal = fst->AddState();
+    fst->SetFinal(superfinal, Weight::One());
+  }
+  for (StateIterator<MutableFst<FromArc>> siter(*fst); !siter.Done();
+       siter.Next()) {
+    const auto state = siter.Value();
+    for (MutableArcIterator<MutableFst<FromArc>> aiter(fst, state);
+         !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      aiter.SetValue((*mapper)(arc));
+    }
+    switch (final_action) {
+      case MAP_NO_SUPERFINAL:
+      default: {
+        const FromArc arc(0, 0, fst->Final(state), kNoStateId);
+        const auto final_arc = (*mapper)(arc);
+        if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
+          FSTERROR() << "ArcMap: Non-zero arc labels for superfinal arc";
+          fst->SetProperties(kError, kError);
+        }
+        fst->SetFinal(state, final_arc.weight);
+        break;
+      }
+      case MAP_ALLOW_SUPERFINAL: {
+        if (state != superfinal) {
+          const FromArc arc(0, 0, fst->Final(state), kNoStateId);
+          auto final_arc = (*mapper)(arc);
+          if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
+            // Add a superfinal state if not already done.
+            if (superfinal == kNoStateId) {
+              superfinal = fst->AddState();
+              fst->SetFinal(superfinal, Weight::One());
+            }
+            final_arc.nextstate = superfinal;
+            fst->AddArc(state, std::move(final_arc));
+            fst->SetFinal(state, Weight::Zero());
+          } else {
+            fst->SetFinal(state, final_arc.weight);
+          }
+        }
+        break;
+      }
+      case MAP_REQUIRE_SUPERFINAL: {
+        if (state != superfinal) {
+          const FromArc arc(0, 0, fst->Final(state), kNoStateId);
+          const auto final_arc = (*mapper)(arc);
+          if (final_arc.ilabel != 0 || final_arc.olabel != 0 ||
+              final_arc.weight != Weight::Zero()) {
+            fst->AddArc(state, ToArc(final_arc.ilabel, final_arc.olabel,
+                                     final_arc.weight, superfinal));
+          }
+          fst->SetFinal(state, Weight::Zero());
+        }
+        break;
+      }
+    }
+  }
+  fst->SetProperties(mapper->Properties(props), kFstProperties);
+}
+
+// Maps an arc type A using a mapper function object C, passed by value. This
+// version modifies its FST input.
+template <class A, class C>
+void ArcMap(MutableFst<A> *fst, C mapper) {
+  ArcMap(fst, &mapper);
+}
+
+// Maps an arc type A to an arc type B using mapper function object C,
+// passed by pointer. This version writes the mapped input FST to an
+// output MutableFst.
+template <class A, class B, class C>
+void ArcMap(const Fst<A> &ifst, MutableFst<B> *ofst, C *mapper) {
+  using FromArc = A;
+  using StateId = typename FromArc::StateId;
+  ofst->DeleteStates();
+  if (mapper->InputSymbolsAction() == MAP_COPY_SYMBOLS) {
+    ofst->SetInputSymbols(ifst.InputSymbols());
+  } else if (mapper->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
+    ofst->SetInputSymbols(nullptr);
+  }
+  if (mapper->OutputSymbolsAction() == MAP_COPY_SYMBOLS) {
+    ofst->SetOutputSymbols(ifst.OutputSymbols());
+  } else if (mapper->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
+    ofst->SetOutputSymbols(nullptr);
+  }
+  const auto iprops = ifst.Properties(kCopyProperties, false);
+  if (ifst.Start() == kNoStateId) {
+    if (iprops & kError) ofst->SetProperties(kError, kError);
+    return;
+  }
+  const auto final_action = mapper->FinalAction();
+  if (ifst.Properties(kExpanded, false)) {
+    ofst->ReserveStates(
+        CountStates(ifst) + (final_action == MAP_NO_SUPERFINAL ? 0 : 1));
+  }
+  // Adds all states.
+  for (StateIterator<Fst<A>> siter(ifst); !siter.Done(); siter.Next()) {
+    ofst->AddState();
+  }
+  StateId superfinal = kNoStateId;
+  if (final_action == MAP_REQUIRE_SUPERFINAL) {
+    superfinal = ofst->AddState();
+    ofst->SetFinal(superfinal, B::Weight::One());
+  }
+  for (StateIterator<Fst<A>> siter(ifst); !siter.Done(); siter.Next()) {
+    StateId s = siter.Value();
+    if (s == ifst.Start()) ofst->SetStart(s);
+    ofst->ReserveArcs(
+        s, ifst.NumArcs(s) + (final_action != MAP_NO_SUPERFINAL ? 1 : 0));
+    for (ArcIterator<Fst<A>> aiter(ifst, s); !aiter.Done(); aiter.Next()) {
+      ofst->AddArc(s, (*mapper)(aiter.Value()));
+    }
+    switch (final_action) {
+      case MAP_NO_SUPERFINAL:
+      default: {
+        B final_arc = (*mapper)(A(0, 0, ifst.Final(s), kNoStateId));
+        if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
+          FSTERROR() << "ArcMap: Non-zero arc labels for superfinal arc";
+          ofst->SetProperties(kError, kError);
+        }
+        ofst->SetFinal(s, final_arc.weight);
+        break;
+      }
+      case MAP_ALLOW_SUPERFINAL: {
+        B final_arc = (*mapper)(A(0, 0, ifst.Final(s), kNoStateId));
+        if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
+          // Add a superfinal state if not already done.
+          if (superfinal == kNoStateId) {
+            superfinal = ofst->AddState();
+            ofst->SetFinal(superfinal, B::Weight::One());
+          }
+          final_arc.nextstate = superfinal;
+          ofst->AddArc(s, std::move(final_arc));
+          ofst->SetFinal(s, B::Weight::Zero());
+        } else {
+          ofst->SetFinal(s, final_arc.weight);
+        }
+        break;
+      }
+      case MAP_REQUIRE_SUPERFINAL: {
+        B final_arc = (*mapper)(A(0, 0, ifst.Final(s), kNoStateId));
+        if (final_arc.ilabel != 0 || final_arc.olabel != 0 ||
+            final_arc.weight != B::Weight::Zero()) {
+          ofst->AddArc(s, B(final_arc.ilabel, final_arc.olabel,
+                            final_arc.weight, superfinal));
+        }
+        ofst->SetFinal(s, B::Weight::Zero());
+        break;
+      }
+    }
+  }
+  const auto oprops = ofst->Properties(kFstProperties, false);
+  ofst->SetProperties(mapper->Properties(iprops) | oprops, kFstProperties);
+}
+
+// Maps an arc type A to an arc type B using mapper function
+// object C, passed by value. This version writes the mapped input
+// Fst to an output MutableFst.
+template <class A, class B, class C>
+void ArcMap(const Fst<A> &ifst, MutableFst<B> *ofst, C mapper) {
+  ArcMap(ifst, ofst, &mapper);
+}
+
+struct ArcMapFstOptions : public CacheOptions {
+  // ArcMapFst default caching behaviour is to do no caching. Most mappers are
+  // cheap and therefore we save memory by not doing caching.
+  ArcMapFstOptions() : CacheOptions(true, 0) {}
+
+  explicit ArcMapFstOptions(const CacheOptions &opts) : CacheOptions(opts) {}
+};
+
+template <class A, class B, class C>
+class ArcMapFst;
+
+namespace internal {
+
+// Implementation of delayed ArcMapFst.
+template <class A, class B, class C>
+class ArcMapFstImpl : public CacheImpl<B> {
+ public:
+  using Arc = B;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FstImpl<B>::SetType;
+  using FstImpl<B>::SetProperties;
+  using FstImpl<B>::SetInputSymbols;
+  using FstImpl<B>::SetOutputSymbols;
+
+  using CacheImpl<B>::EmplaceArc;
+  using CacheImpl<B>::HasArcs;
+  using CacheImpl<B>::HasFinal;
+  using CacheImpl<B>::HasStart;
+  using CacheImpl<B>::PushArc;
+  using CacheImpl<B>::SetArcs;
+  using CacheImpl<B>::SetFinal;
+  using CacheImpl<B>::SetStart;
+
+  friend class StateIterator<ArcMapFst<A, B, C>>;
+
+  ArcMapFstImpl(const Fst<A> &fst, const C &mapper,
+                const ArcMapFstOptions &opts)
+      : CacheImpl<B>(opts),
+        fst_(fst.Copy()),
+        mapper_(new C(mapper)),
+        own_mapper_(true),
+        superfinal_(kNoStateId),
+        nstates_(0) {
+    Init();
+  }
+
+  ArcMapFstImpl(const Fst<A> &fst, C *mapper, const ArcMapFstOptions &opts)
+      : CacheImpl<B>(opts),
+        fst_(fst.Copy()),
+        mapper_(mapper),
+        own_mapper_(false),
+        superfinal_(kNoStateId),
+        nstates_(0) {
+    Init();
+  }
+
+  ArcMapFstImpl(const ArcMapFstImpl<A, B, C> &impl)
+      : CacheImpl<B>(impl),
+        fst_(impl.fst_->Copy(true)),
+        mapper_(new C(*impl.mapper_)),
+        own_mapper_(true),
+        superfinal_(kNoStateId),
+        nstates_(0) {
+    Init();
+  }
+
+  ~ArcMapFstImpl() override {
+    if (own_mapper_) delete mapper_;
+  }
+
+  StateId Start() {
+    if (!HasStart()) SetStart(FindOState(fst_->Start()));
+    return CacheImpl<B>::Start();
+  }
+
+  Weight Final(StateId s) {
+    if (!HasFinal(s)) {
+      switch (final_action_) {
+        case MAP_NO_SUPERFINAL:
+        default: {
+          const auto final_arc =
+              (*mapper_)(A(0, 0, fst_->Final(FindIState(s)), kNoStateId));
+          if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
+            FSTERROR() << "ArcMapFst: Non-zero arc labels for superfinal arc";
+            SetProperties(kError, kError);
+          }
+          SetFinal(s, final_arc.weight);
+          break;
+        }
+        case MAP_ALLOW_SUPERFINAL: {
+          if (s == superfinal_) {
+            SetFinal(s, Weight::One());
+          } else {
+            const auto final_arc =
+                (*mapper_)(A(0, 0, fst_->Final(FindIState(s)), kNoStateId));
+            if (final_arc.ilabel == 0 && final_arc.olabel == 0) {
+              SetFinal(s, final_arc.weight);
+            } else {
+              SetFinal(s, Weight::Zero());
+            }
+          }
+          break;
+        }
+        case MAP_REQUIRE_SUPERFINAL: {
+          SetFinal(s, s == superfinal_ ? Weight::One() : Weight::Zero());
+          break;
+        }
+      }
+    }
+    return CacheImpl<B>::Final(s);
+  }
+
+  size_t NumArcs(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<B>::NumArcs(s);
+  }
+
+  size_t NumInputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<B>::NumInputEpsilons(s);
+  }
+
+  size_t NumOutputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<B>::NumOutputEpsilons(s);
+  }
+
+  uint64 Properties() const override { return Properties(kFstProperties); }
+
+  // Sets error if found, and returns other FST impl properties.
+  uint64 Properties(uint64 mask) const override {
+    if ((mask & kError) && (fst_->Properties(kError, false) ||
+                            (mapper_->Properties(0) & kError))) {
+      SetProperties(kError, kError);
+    }
+    return FstImpl<Arc>::Properties(mask);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<B> *data) {
+    if (!HasArcs(s)) Expand(s);
+    CacheImpl<B>::InitArcIterator(s, data);
+  }
+
+  void Expand(StateId s) {
+    // Add exiting arcs.
+    if (s == superfinal_) {
+      SetArcs(s);
+      return;
+    }
+    for (ArcIterator<Fst<A>> aiter(*fst_, FindIState(s)); !aiter.Done();
+         aiter.Next()) {
+      auto aarc = aiter.Value();
+      aarc.nextstate = FindOState(aarc.nextstate);
+      PushArc(s, (*mapper_)(aarc));
+    }
+
+    // Check for superfinal arcs.
+    if (!HasFinal(s) || Final(s) == Weight::Zero()) {
+      switch (final_action_) {
+        case MAP_NO_SUPERFINAL:
+        default:
+          break;
+        case MAP_ALLOW_SUPERFINAL: {
+          auto final_arc =
+              (*mapper_)(A(0, 0, fst_->Final(FindIState(s)), kNoStateId));
+          if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
+            if (superfinal_ == kNoStateId) superfinal_ = nstates_++;
+            final_arc.nextstate = superfinal_;
+            PushArc(s, std::move(final_arc));
+          }
+          break;
+        }
+        case MAP_REQUIRE_SUPERFINAL: {
+          const auto final_arc =
+              (*mapper_)(A(0, 0, fst_->Final(FindIState(s)), kNoStateId));
+          if (final_arc.ilabel != 0 || final_arc.olabel != 0 ||
+              final_arc.weight != B::Weight::Zero()) {
+            EmplaceArc(s, final_arc.ilabel, final_arc.olabel, final_arc.weight,
+                       superfinal_);
+          }
+          break;
+        }
+      }
+    }
+    SetArcs(s);
+  }
+
+ private:
+  void Init() {
+    SetType("map");
+    if (mapper_->InputSymbolsAction() == MAP_COPY_SYMBOLS) {
+      SetInputSymbols(fst_->InputSymbols());
+    } else if (mapper_->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
+      SetInputSymbols(nullptr);
+    }
+    if (mapper_->OutputSymbolsAction() == MAP_COPY_SYMBOLS) {
+      SetOutputSymbols(fst_->OutputSymbols());
+    } else if (mapper_->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
+      SetOutputSymbols(nullptr);
+    }
+    if (fst_->Start() == kNoStateId) {
+      final_action_ = MAP_NO_SUPERFINAL;
+      SetProperties(kNullProperties);
+    } else {
+      final_action_ = mapper_->FinalAction();
+      uint64 props = fst_->Properties(kCopyProperties, false);
+      SetProperties(mapper_->Properties(props));
+      if (final_action_ == MAP_REQUIRE_SUPERFINAL) superfinal_ = 0;
+    }
+  }
+
+  // Maps from output state to input state.
+  StateId FindIState(StateId s) {
+    if (superfinal_ == kNoStateId || s < superfinal_) {
+      return s;
+    } else {
+      return s - 1;
+    }
+  }
+
+  // Maps from input state to output state.
+  StateId FindOState(StateId is) {
+    auto os = is;
+    if (!(superfinal_ == kNoStateId || is < superfinal_)) ++os;
+    if (os >= nstates_) nstates_ = os + 1;
+    return os;
+  }
+
+  std::unique_ptr<const Fst<A>> fst_;
+  C *mapper_;
+  const bool own_mapper_;
+  MapFinalAction final_action_;
+  StateId superfinal_;
+  StateId nstates_;
+};
+
+}  // namespace internal
+
+// Maps an arc type A to an arc type B using Mapper function object
+// C. This version is a delayed FST.
+template <class A, class B, class C>
+class ArcMapFst : public ImplToFst<internal::ArcMapFstImpl<A, B, C>> {
+ public:
+  using Arc = B;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Store = DefaultCacheStore<B>;
+  using State = typename Store::State;
+  using Impl = internal::ArcMapFstImpl<A, B, C>;
+
+  friend class ArcIterator<ArcMapFst<A, B, C>>;
+  friend class StateIterator<ArcMapFst<A, B, C>>;
+
+  ArcMapFst(const Fst<A> &fst, const C &mapper, const ArcMapFstOptions &opts)
+      : ImplToFst<Impl>(std::make_shared<Impl>(fst, mapper, opts)) {}
+
+  ArcMapFst(const Fst<A> &fst, C *mapper, const ArcMapFstOptions &opts)
+      : ImplToFst<Impl>(std::make_shared<Impl>(fst, mapper, opts)) {}
+
+  ArcMapFst(const Fst<A> &fst, const C &mapper)
+      : ImplToFst<Impl>(
+            std::make_shared<Impl>(fst, mapper, ArcMapFstOptions())) {}
+
+  ArcMapFst(const Fst<A> &fst, C *mapper)
+      : ImplToFst<Impl>(
+            std::make_shared<Impl>(fst, mapper, ArcMapFstOptions())) {}
+
+  // See Fst<>::Copy() for doc.
+  ArcMapFst(const ArcMapFst<A, B, C> &fst, bool safe = false)
+      : ImplToFst<Impl>(fst, safe) {}
+
+  // Get a copy of this ArcMapFst. See Fst<>::Copy() for further doc.
+  ArcMapFst<A, B, C> *Copy(bool safe = false) const override {
+    return new ArcMapFst<A, B, C>(*this, safe);
+  }
+
+  inline void InitStateIterator(StateIteratorData<B> *data) const override;
+
+  void InitArcIterator(StateId s, ArcIteratorData<B> *data) const override {
+    GetMutableImpl()->InitArcIterator(s, data);
+  }
+
+ protected:
+  using ImplToFst<Impl>::GetImpl;
+  using ImplToFst<Impl>::GetMutableImpl;
+
+ private:
+  ArcMapFst &operator=(const ArcMapFst &) = delete;
+};
+
+// Specialization for ArcMapFst.
+//
+// This may be derived from.
+template <class A, class B, class C>
+class StateIterator<ArcMapFst<A, B, C>> : public StateIteratorBase<B> {
+ public:
+  using StateId = typename B::StateId;
+
+  explicit StateIterator(const ArcMapFst<A, B, C> &fst)
+      : impl_(fst.GetImpl()),
+        siter_(*impl_->fst_),
+        s_(0),
+        superfinal_(impl_->final_action_ == MAP_REQUIRE_SUPERFINAL) {
+    CheckSuperfinal();
+  }
+
+  bool Done() const final { return siter_.Done() && !superfinal_; }
+
+  StateId Value() const final { return s_; }
+
+  void Next() final {
+    ++s_;
+    if (!siter_.Done()) {
+      siter_.Next();
+      CheckSuperfinal();
+    } else if (superfinal_) {
+      superfinal_ = false;
+    }
+  }
+
+  void Reset() final {
+    s_ = 0;
+    siter_.Reset();
+    superfinal_ = impl_->final_action_ == MAP_REQUIRE_SUPERFINAL;
+    CheckSuperfinal();
+  }
+
+ private:
+  void CheckSuperfinal() {
+    if (impl_->final_action_ != MAP_ALLOW_SUPERFINAL || superfinal_) return;
+    if (!siter_.Done()) {
+      const auto final_arc =
+          (*impl_->mapper_)(A(0, 0, impl_->fst_->Final(s_), kNoStateId));
+      if (final_arc.ilabel != 0 || final_arc.olabel != 0) superfinal_ = true;
+    }
+  }
+
+  const internal::ArcMapFstImpl<A, B, C> *impl_;
+  StateIterator<Fst<A>> siter_;
+  StateId s_;
+  bool superfinal_;  // True if there is a superfinal state and not done.
+};
+
+// Specialization for ArcMapFst.
+template <class A, class B, class C>
+class ArcIterator<ArcMapFst<A, B, C>>
+    : public CacheArcIterator<ArcMapFst<A, B, C>> {
+ public:
+  using StateId = typename A::StateId;
+
+  ArcIterator(const ArcMapFst<A, B, C> &fst, StateId s)
+      : CacheArcIterator<ArcMapFst<A, B, C>>(fst.GetMutableImpl(), s) {
+    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
+  }
+};
+
+template <class A, class B, class C>
+inline void ArcMapFst<A, B, C>::InitStateIterator(
+    StateIteratorData<B> *data) const {
+  data->base = new StateIterator<ArcMapFst<A, B, C>>(*this);
+}
+
+// Utility Mappers.
+
+// Mapper that returns its input.
+template <class A>
+class IdentityArcMapper {
+ public:
+  using FromArc = A;
+  using ToArc = A;
+
+  constexpr ToArc operator()(const FromArc &arc) const { return arc; }
+
+  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr uint64 Properties(uint64 props) const { return props; }
+};
+
+// Mapper that converts all input symbols to epsilon.
+template <class A>
+class InputEpsilonMapper {
+ public:
+  using FromArc = A;
+  using ToArc = A;
+
+  constexpr ToArc operator()(const FromArc &arc) const {
+    return ToArc(0, arc.olabel, arc.weight, arc.nextstate);
+  }
+
+  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_CLEAR_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr uint64 Properties(uint64 props) const {
+    return (props & kSetArcProperties) | kIEpsilons | kILabelSorted;
+  }
+};
+
+// Mapper that converts all output symbols to epsilon.
+template <class A>
+class OutputEpsilonMapper {
+ public:
+  using FromArc = A;
+  using ToArc = A;
+
+  constexpr ToArc operator()(const FromArc &arc) const {
+    return ToArc(arc.ilabel, 0, arc.weight, arc.nextstate);
+  }
+
+  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_CLEAR_SYMBOLS;
+  }
+
+  constexpr uint64 Properties(uint64 props) const {
+    return (props & kSetArcProperties) | kOEpsilons | kOLabelSorted;
+  }
+};
+
+// Mapper that returns its input with final states redirected to a single
+// super-final state.
+template <class A>
+class SuperFinalMapper {
+ public:
+  using FromArc = A;
+  using ToArc = A;
+  using Label = typename FromArc::Label;
+  using Weight = typename FromArc::Weight;;
+
+  // Arg allows setting super-final label.
+  explicit SuperFinalMapper(Label final_label = 0)
+      : final_label_(final_label) {}
+
+  ToArc operator()(const FromArc &arc) const {
+    // Super-final arc.
+    if (arc.nextstate == kNoStateId && arc.weight != Weight::Zero()) {
+      return ToArc(final_label_, final_label_, arc.weight, kNoStateId);
+    } else {
+      return arc;
+    }
+  }
+
+  constexpr MapFinalAction FinalAction() const {
+    return MAP_REQUIRE_SUPERFINAL;
+  }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  uint64 Properties(uint64 props) const {
+    if (final_label_ == 0) {
+      return props & kAddSuperFinalProperties;
+    } else {
+      return props & kAddSuperFinalProperties &
+          kILabelInvariantProperties & kOLabelInvariantProperties;
+    }
+  }
+
+ private:
+  Label final_label_;
+};
+
+// Mapper that leaves labels and nextstate unchanged and constructs a new weight
+// from the underlying value of the arc weight. If no weight converter is
+// explictly specified, requires that there is a WeightConvert class
+// specialization that converts the weights.
+template <class A, class B,
+          class C = WeightConvert<typename A::Weight, typename B::Weight>>
+class WeightConvertMapper {
+ public:
+  using FromArc = A;
+  using ToArc = B;
+  using Converter = C;
+  using FromWeight = typename FromArc::Weight;
+  using ToWeight = typename ToArc::Weight;
+
+  constexpr explicit WeightConvertMapper(const Converter &c = Converter())
+      : convert_weight_(c) {}
+
+  constexpr ToArc operator()(const FromArc &arc) const {
+    return ToArc(arc.ilabel, arc.olabel, convert_weight_(arc.weight),
+                 arc.nextstate);
+  }
+
+  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr uint64 Properties(uint64 props) const { return props; }
+
+ private:
+  const Converter convert_weight_;
+};
+
+// Non-precision-changing weight conversions; consider using more efficient
+// Cast method instead.
+
+using StdToLogMapper = WeightConvertMapper<StdArc, LogArc>;
+
+using LogToStdMapper = WeightConvertMapper<LogArc, StdArc>;
+
+// Precision-changing weight conversions.
+
+using StdToLog64Mapper = WeightConvertMapper<StdArc, Log64Arc>;
+
+using LogToLog64Mapper = WeightConvertMapper<LogArc, Log64Arc>;
+
+using Log64ToStdMapper = WeightConvertMapper<Log64Arc, StdArc>;
+
+using Log64ToLogMapper = WeightConvertMapper<Log64Arc, LogArc>;
+
+// Mapper from A to GallicArc<A>.
+template <class A, GallicType G = GALLIC_LEFT>
+class ToGallicMapper {
+ public:
+  using FromArc = A;
+  using ToArc = GallicArc<A, G>;
+
+  using SW = StringWeight<typename A::Label, GallicStringType(G)>;
+  using AW = typename FromArc::Weight;
+  using GW = typename ToArc::Weight;
+
+  ToArc operator()(const FromArc &arc) const {
+    // Super-final arc.
+    if (arc.nextstate == kNoStateId && arc.weight != AW::Zero()) {
+      return ToArc(0, 0, GW(SW::One(), arc.weight), kNoStateId);
+    // Super-non-final arc.
+    } else if (arc.nextstate == kNoStateId) {
+      return ToArc(0, 0, GW::Zero(), kNoStateId);
+    // Epsilon label.
+    } else if (arc.olabel == 0) {
+      return ToArc(arc.ilabel, arc.ilabel, GW(SW::One(), arc.weight),
+                   arc.nextstate);
+    // Regular label.
+    } else {
+      return ToArc(arc.ilabel, arc.ilabel, GW(SW(arc.olabel), arc.weight),
+                   arc.nextstate);
+    }
+  }
+
+  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_CLEAR_SYMBOLS;
+  }
+
+  uint64 Properties(uint64 props) const {
+    return ProjectProperties(props, true) & kWeightInvariantProperties;
+  }
+};
+
+// Mapper from GallicArc<A> to A.
+template <class A, GallicType G = GALLIC_LEFT>
+class FromGallicMapper {
+ public:
+  using FromArc = GallicArc<A, G>;
+  using ToArc = A;
+
+  using Label = typename ToArc::Label;
+  using AW = typename ToArc::Weight;
+  using GW = typename FromArc::Weight;
+
+  explicit FromGallicMapper(Label superfinal_label = 0)
+      : superfinal_label_(superfinal_label), error_(false) {}
+
+  ToArc operator()(const FromArc &arc) const {
+    // 'Super-non-final' arc.
+    if (arc.nextstate == kNoStateId && arc.weight == GW::Zero()) {
+      return A(arc.ilabel, 0, AW::Zero(), kNoStateId);
+    }
+    Label l = kNoLabel;
+    AW weight;
+    if (!Extract(arc.weight, &weight, &l) || arc.ilabel != arc.olabel) {
+      FSTERROR() << "FromGallicMapper: Unrepresentable weight: " << arc.weight
+                 << " for arc with ilabel = " << arc.ilabel
+                 << ", olabel = " << arc.olabel
+                 << ", nextstate = " << arc.nextstate;
+      error_ = true;
+    }
+    if (arc.ilabel == 0 && l != 0 && arc.nextstate == kNoStateId) {
+      return ToArc(superfinal_label_, l, weight, arc.nextstate);
+    } else {
+      return ToArc(arc.ilabel, l, weight, arc.nextstate);
+    }
+  }
+
+  constexpr MapFinalAction FinalAction() const { return MAP_ALLOW_SUPERFINAL; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_CLEAR_SYMBOLS;
+  }
+
+  uint64 Properties(uint64 inprops) const {
+    uint64 outprops = inprops & kOLabelInvariantProperties &
+                      kWeightInvariantProperties & kAddSuperFinalProperties;
+    if (error_) outprops |= kError;
+    return outprops;
+  }
+
+ private:
+  template <GallicType GT>
+  static bool Extract(const GallicWeight<Label, AW, GT> &gallic_weight,
+                      typename A::Weight *weight, typename A::Label *label) {
+    using GWT = StringWeight<Label, GallicStringType(GT)>;
+    const GWT &w1 = gallic_weight.Value1();
+    const AW &w2 = gallic_weight.Value2();
+    typename GWT::Iterator iter1(w1);
+    const Label l = w1.Size() == 1 ? iter1.Value() : 0;
+    if (l == kStringInfinity || l == kStringBad || w1.Size() > 1) return false;
+    *label = l;
+    *weight = w2;
+    return true;
+  }
+
+  static bool Extract(const GallicWeight<Label, AW, GALLIC> &gallic_weight,
+                      typename A::Weight *weight, typename A::Label *label) {
+    if (gallic_weight.Size() > 1) return false;
+    if (gallic_weight.Size() == 0) {
+      *label = 0;
+      *weight = A::Weight::Zero();
+      return true;
+    }
+    return Extract<GALLIC_RESTRICT>(gallic_weight.Back(), weight, label);
+  }
+
+  const Label superfinal_label_;
+  mutable bool error_;
+};
+
+// Mapper from GallicArc<A> to A.
+template <class A, GallicType G = GALLIC_LEFT>
+class GallicToNewSymbolsMapper {
+ public:
+  using FromArc = GallicArc<A, G>;
+  using ToArc = A;
+
+  using Label = typename ToArc::Label;
+  using StateId = typename ToArc::StateId;
+  using AW = typename ToArc::Weight;
+  using GW = typename FromArc::Weight;
+  using SW = StringWeight<Label, GallicStringType(G)>;
+
+  explicit GallicToNewSymbolsMapper(MutableFst<ToArc> *fst)
+      : fst_(fst),
+        lmax_(0),
+        osymbols_(fst->OutputSymbols()),
+        isymbols_(nullptr),
+        error_(false) {
+    fst_->DeleteStates();
+    state_ = fst_->AddState();
+    fst_->SetStart(state_);
+    fst_->SetFinal(state_, AW::One());
+    if (osymbols_) {
+      string name = osymbols_->Name() + "_from_gallic";
+      fst_->SetInputSymbols(new SymbolTable(name));
+      isymbols_ = fst_->MutableInputSymbols();
+      const int64 zero = 0;
+      isymbols_->AddSymbol(osymbols_->Find(zero), 0);
+    } else {
+      fst_->SetInputSymbols(nullptr);
+    }
+  }
+
+  ToArc operator()(const FromArc &arc) {
+    // Super-non-final arc.
+    if (arc.nextstate == kNoStateId && arc.weight == GW::Zero()) {
+      return ToArc(arc.ilabel, 0, AW::Zero(), kNoStateId);
+    }
+    SW w1 = arc.weight.Value1();
+    AW w2 = arc.weight.Value2();
+    Label l;
+    if (w1.Size() == 0) {
+      l = 0;
+    } else {
+      auto insert_result = map_.insert(std::make_pair(w1, kNoLabel));
+      if (!insert_result.second) {
+        l = insert_result.first->second;
+      } else {
+        l = ++lmax_;
+        insert_result.first->second = l;
+        StringWeightIterator<SW> iter1(w1);
+        StateId n;
+        string s;
+        for (size_t i = 0, p = state_; i < w1.Size();
+             ++i, iter1.Next(), p = n) {
+          n = i == w1.Size() - 1 ? state_ : fst_->AddState();
+          fst_->AddArc(p, ToArc(i ? 0 : l, iter1.Value(), AW::One(), n));
+          if (isymbols_) {
+            if (i) s = s + "_";
+            s = s + osymbols_->Find(iter1.Value());
+          }
+        }
+        if (isymbols_) isymbols_->AddSymbol(s, l);
+      }
+    }
+    if (l == kStringInfinity || l == kStringBad || arc.ilabel != arc.olabel) {
+      FSTERROR() << "GallicToNewSymbolMapper: Unrepresentable weight: " << l;
+      error_ = true;
+    }
+    return ToArc(arc.ilabel, l, w2, arc.nextstate);
+  }
+
+  constexpr MapFinalAction FinalAction() const { return MAP_ALLOW_SUPERFINAL; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_CLEAR_SYMBOLS;
+  }
+
+  uint64 Properties(uint64 inprops) const {
+    uint64 outprops = inprops & kOLabelInvariantProperties &
+                      kWeightInvariantProperties & kAddSuperFinalProperties;
+    if (error_) outprops |= kError;
+    return outprops;
+  }
+
+ private:
+  class StringKey {
+   public:
+    size_t operator()(const SW &x) const { return x.Hash(); }
+  };
+
+  using Map = std::unordered_map<SW, Label, StringKey>;
+
+  MutableFst<ToArc> *fst_;
+  Map map_;
+  Label lmax_;
+  StateId state_;
+  const SymbolTable *osymbols_;
+  SymbolTable *isymbols_;
+  mutable bool error_;
+};
+
+// TODO(kbg): Add common base class for those mappers which do nothing except
+// mutate their weights.
+
+// Mapper to add a constant to all weights.
+template <class A>
+class PlusMapper {
+ public:
+  using FromArc = A;
+  using ToArc = A;
+  using Weight = typename FromArc::Weight;
+
+  constexpr explicit PlusMapper(Weight weight) : weight_(std::move(weight)) {}
+
+  ToArc operator()(const FromArc &arc) const {
+    if (arc.weight == Weight::Zero()) return arc;
+    return ToArc(arc.ilabel, arc.olabel, Plus(arc.weight, weight_),
+                 arc.nextstate);
+  }
+
+  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr uint64 Properties(uint64 props) const {
+    return props & kWeightInvariantProperties;
+  }
+
+ private:
+  const Weight weight_;
+};
+
+// Mapper to (right) multiply a constant to all weights.
+template <class A>
+class TimesMapper {
+ public:
+  using FromArc = A;
+  using ToArc = A;
+  using Weight = typename FromArc::Weight;
+
+  constexpr explicit TimesMapper(Weight weight) : weight_(std::move(weight)) {}
+
+  ToArc operator()(const FromArc &arc) const {
+    if (arc.weight == Weight::Zero()) return arc;
+    return ToArc(arc.ilabel, arc.olabel, Times(arc.weight, weight_),
+                 arc.nextstate);
+  }
+
+  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr uint64 Properties(uint64 props) const {
+    return props & kWeightInvariantProperties;
+  }
+
+ private:
+  const Weight weight_;
+};
+
+// Mapper to take all weights to a constant power. The power argument is stored
+// as a double, so if there is a floating-point power implementation for this
+// weight type, it will take precedence. Otherwise, the power argument's 53 bits
+// of integer precision will be implicitly converted to a size_t and the default
+// power implementation (iterated multiplication) will be used instead.
+template <class A>
+class PowerMapper {
+ public:
+  using FromArc = A;
+  using ToArc = A;
+  using Weight = typename FromArc::Weight;
+
+  explicit PowerMapper(double power) : power_(power) {}
+
+  ToArc operator()(const FromArc &arc) const {
+    return ToArc(arc.ilabel, arc.olabel, Power(arc.weight, power_),
+                 arc.nextstate);
+  }
+
+  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr uint64 Properties(uint64 props) const {
+    return props & kWeightInvariantProperties;
+  }
+
+ private:
+  const double power_;
+};
+
+// Mapper to reciprocate all non-Zero() weights.
+template <class A>
+class InvertWeightMapper {
+ public:
+  using FromArc = A;
+  using ToArc = A;
+  using Weight = typename FromArc::Weight;
+
+  ToArc operator()(const FromArc &arc) const {
+    if (arc.weight == Weight::Zero()) return arc;
+    return ToArc(arc.ilabel, arc.olabel, Divide(Weight::One(), arc.weight),
+                 arc.nextstate);
+  }
+
+  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr uint64 Properties(uint64 props) const {
+    return props & kWeightInvariantProperties;
+  }
+};
+
+// Mapper to map all non-Zero() weights to One().
+template <class A, class B = A>
+class RmWeightMapper {
+ public:
+  using FromArc = A;
+  using ToArc = B;
+  using FromWeight = typename FromArc::Weight;
+  using ToWeight = typename ToArc::Weight;
+
+  ToArc operator()(const FromArc &arc) const {
+    return ToArc(arc.ilabel, arc.olabel,
+                 arc.weight != FromWeight::Zero() ?
+                 ToWeight::One() : ToWeight::Zero(),
+                 arc.nextstate);
+  }
+
+  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr uint64 Properties(uint64 props) const {
+    return (props & kWeightInvariantProperties) | kUnweighted;
+  }
+};
+
+// Mapper to quantize all weights.
+template <class A, class B = A>
+class QuantizeMapper {
+ public:
+  using FromArc = A;
+  using ToArc = B;
+  using FromWeight = typename FromArc::Weight;
+  using ToWeight = typename ToArc::Weight;
+
+  QuantizeMapper() : delta_(kDelta) {}
+
+  explicit QuantizeMapper(float d) : delta_(d) {}
+
+  ToArc operator()(const FromArc &arc) const {
+    return ToArc(arc.ilabel, arc.olabel, arc.weight.Quantize(delta_),
+                 arc.nextstate);
+  }
+
+  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr uint64 Properties(uint64 props) const {
+    return props & kWeightInvariantProperties;
+  }
+
+ private:
+  const float delta_;
+};
+
+// Mapper from A to B under the assumption:
+//
+//    B::Weight = A::Weight::ReverseWeight
+//    B::Label == A::Label
+//    B::StateId == A::StateId
+//
+// The weight is reversed, while the label and nextstate are preserved.
+template <class A, class B>
+class ReverseWeightMapper {
+ public:
+  using FromArc = A;
+  using ToArc = B;
+
+  constexpr ToArc operator()(const FromArc &arc) const {
+    return ToArc(arc.ilabel, arc.olabel, arc.weight.Reverse(), arc.nextstate);
+  }
+
+  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr uint64 Properties(uint64 props) const { return props; }
+};
+
+}  // namespace fst
+
+#endif  // FST_ARC_MAP_H_
diff --git a/projects/llm_framework/include/fst/arc.h b/projects/llm_framework/include/fst/arc.h
new file mode 100644
index 00000000..651b11df
--- /dev/null
+++ b/projects/llm_framework/include/fst/arc.h
@@ -0,0 +1,317 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Commonly used FST arc types.
+
+#ifndef FST_ARC_H_
+#define FST_ARC_H_
+
+#include <climits>
+#include <string>
+#include <type_traits>
+#include <utility>
+
+
+#include <fst/expectation-weight.h>
+#include <fst/float-weight.h>
+#include <fst/lexicographic-weight.h>
+#include <fst/power-weight.h>
+#include <fst/product-weight.h>
+#include <fst/signed-log-weight.h>
+#include <fst/sparse-power-weight.h>
+#include <fst/string-weight.h>
+
+
+namespace fst {
+
+template <class W>
+struct ArcTpl {
+ public:
+  using Weight = W;
+  using Label = int;
+  using StateId = int;
+
+  Label ilabel;
+  Label olabel;
+  Weight weight;
+  StateId nextstate;
+
+  ArcTpl() noexcept(std::is_nothrow_default_constructible<Weight>::value) {}
+
+  template <class T>
+  ArcTpl(Label ilabel, Label olabel, T &&weight, StateId nextstate)
+      : ilabel(ilabel),
+        olabel(olabel),
+        weight(std::forward<T>(weight)),
+        nextstate(nextstate) {}
+
+  static const string &Type() {
+    static const auto *const type =
+        new string(Weight::Type() == "tropical" ? "standard" : Weight::Type());
+    return *type;
+  }
+};
+
+using StdArc = ArcTpl<TropicalWeight>;
+using LogArc = ArcTpl<LogWeight>;
+using Log64Arc = ArcTpl<Log64Weight>;
+using SignedLogArc = ArcTpl<SignedLogWeight>;
+using SignedLog64Arc = ArcTpl<SignedLog64Weight>;
+using MinMaxArc = ArcTpl<MinMaxWeight>;
+
+// Arc with integer labels and state IDs and string weights.
+template <StringType S = STRING_LEFT>
+struct StringArc {
+ public:
+  using Label = int;
+  using Weight = StringWeight<int, S>;
+  using StateId = int;
+
+  Label ilabel;
+  Label olabel;
+  Weight weight;
+  StateId nextstate;
+
+  StringArc() = default;
+
+  template <class W>
+  StringArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
+      : ilabel(ilabel),
+        olabel(olabel),
+        weight(std::forward<W>(weight)),
+        nextstate(nextstate) {}
+
+  static const string &Type() {
+    static const auto *const type = new string(
+        S == STRING_LEFT ? "left_standard_string"
+                         : (S == STRING_RIGHT ? "right_standard_string"
+                                              : "restricted_standard_string"));
+    return *type;
+  }
+};
+
+// Arc with label and state Id type the same as template arg and with
+// weights over the Gallic semiring w.r.t the output labels and weights of A.
+template <class A, GallicType G = GALLIC_LEFT>
+struct GallicArc {
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = GallicWeight<Label, typename Arc::Weight, G>;
+
+  Label ilabel;
+  Label olabel;
+  Weight weight;
+  StateId nextstate;
+
+  GallicArc() = default;
+
+  template <class W>
+  GallicArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
+      : ilabel(ilabel),
+        olabel(olabel),
+        weight(std::forward<W>(weight)),
+        nextstate(nextstate) {}
+
+  explicit GallicArc(const Arc &arc)
+      : ilabel(arc.ilabel), olabel(arc.ilabel), weight(arc.olabel, arc.weight),
+        nextstate(arc.nextstate) {}
+
+  static const string &Type() {
+    static const auto *const type = new string(
+        (G == GALLIC_LEFT
+             ? "left_gallic_"
+             : (G == GALLIC_RIGHT
+                    ? "right_gallic_"
+                    : (G == GALLIC_RESTRICT
+                           ? "restricted_gallic_"
+                           : (G == GALLIC_MIN ? "min_gallic_" : "gallic_")))) +
+        Arc::Type());
+    return *type;
+  }
+};
+
+// Arc with the reverse of the weight found in its template arg.
+template <class A>
+struct ReverseArc {
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using AWeight = typename Arc::Weight;
+  using Weight = typename AWeight::ReverseWeight;
+
+  Label ilabel;
+  Label olabel;
+  Weight weight;
+  StateId nextstate;
+
+  ReverseArc() = default;
+
+  template <class W>
+  ReverseArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
+      : ilabel(ilabel),
+        olabel(olabel),
+        weight(std::forward<W>(weight)),
+        nextstate(nextstate) {}
+
+  static const string &Type() {
+    static const auto *const type = new string("reverse_" + Arc::Type());
+    return *type;
+  }
+};
+
+// Arc with integer labels and state IDs and lexicographic weights.
+template <class Weight1, class Weight2>
+struct LexicographicArc {
+  using Label = int;
+  using StateId = int;
+  using Weight = LexicographicWeight<Weight1, Weight2>;
+
+  Label ilabel;
+  Label olabel;
+  Weight weight;
+  StateId nextstate;
+
+  LexicographicArc() = default;
+
+  template <class W>
+  LexicographicArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
+      : ilabel(ilabel),
+        olabel(olabel),
+        weight(std::forward<W>(weight)),
+        nextstate(nextstate) {}
+
+  static const string &Type() {
+    static const string *const type = new string(Weight::Type());
+    return *type;
+  }
+};
+
+// Arc with integer labels and state IDs and product weights.
+template <class Weight1, class Weight2>
+struct ProductArc {
+  using Label = int;
+  using StateId = int;
+  using Weight = ProductWeight<Weight1, Weight2>;
+
+  Label ilabel;
+  Label olabel;
+  Weight weight;
+  StateId nextstate;
+
+  ProductArc() = default;
+
+  template <class W>
+  ProductArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
+      : ilabel(ilabel),
+        olabel(olabel),
+        weight(std::forward<W>(weight)),
+        nextstate(nextstate) {}
+
+  static const string &Type() {
+    static const auto *const type = new string(Weight::Type());
+    return *type;
+  }
+};
+
+// Arc with label and state ID type the same as first template argument and with
+// weights over the n-th Cartesian power of the weight type of the template
+// argument.
+template <class A, size_t n>
+struct PowerArc {
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = PowerWeight<typename Arc::Weight, n>;
+
+  Label ilabel;
+  Label olabel;
+  Weight weight;
+  StateId nextstate;
+
+  PowerArc() = default;
+
+  template <class W>
+  PowerArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
+      : ilabel(ilabel),
+        olabel(olabel),
+        weight(std::forward<W>(weight)),
+        nextstate(nextstate) {}
+
+  static const string &Type() {
+    static const auto *const type =
+        new string(Arc::Type() + "_^" + std::to_string(n));
+    return *type;
+  }
+};
+
+// Arc with label and state ID type the same as first template argument and with
+// weights over the arbitrary Cartesian power of the weight type.
+template <class A, class K = int>
+struct SparsePowerArc {
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::Label;
+  using Weight = SparsePowerWeight<typename Arc::Weight, K>;
+
+  Label ilabel;
+  Label olabel;
+  Weight weight;
+  StateId nextstate;
+
+  SparsePowerArc() = default;
+
+  template <class W>
+  SparsePowerArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
+      : ilabel(ilabel),
+        olabel(olabel),
+        weight(std::forward<W>(weight)),
+        nextstate(nextstate) {}
+
+  static const string &Type() {
+    static const string *const type = [] {
+      string type = Arc::Type() + "_^n";
+      if (sizeof(K) != sizeof(uint32)) {
+        type += "_" + std::to_string(CHAR_BIT * sizeof(K));
+      }
+      return new string(type);
+    }();
+    return *type;
+  }
+};
+
+// Arc with label and state ID type the same as first template argument and with
+// expectation weight over the first template argument's weight type and the
+// second template argument.
+template <class A, class X2>
+struct ExpectationArc {
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using X1 = typename Arc::Weight;
+  using Weight = ExpectationWeight<X1, X2>;
+
+  Label ilabel;
+  Label olabel;
+  Weight weight;
+  StateId nextstate;
+
+  ExpectationArc() = default;
+
+  template <class W>
+  ExpectationArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
+      : ilabel(ilabel),
+        olabel(olabel),
+        weight(std::forward<W>(weight)),
+        nextstate(nextstate) {}
+
+  static const string &Type() {
+    static const auto *const type =
+        new string("expectation_" + Arc::Type() + "_" + X2::Type());
+    return *type;
+  }
+};
+
+}  // namespace fst
+
+#endif  // FST_ARC_H_
diff --git a/projects/llm_framework/include/fst/arcfilter.h b/projects/llm_framework/include/fst/arcfilter.h
new file mode 100644
index 00000000..598e543f
--- /dev/null
+++ b/projects/llm_framework/include/fst/arcfilter.h
@@ -0,0 +1,93 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Function objects to restrict which arcs are traversed in an FST.
+
+#ifndef FST_ARCFILTER_H_
+#define FST_ARCFILTER_H_
+
+
+#include <fst/fst.h>
+#include <fst/util.h>
+
+
+namespace fst {
+
+// True for all arcs.
+template <class Arc>
+class AnyArcFilter {
+ public:
+  bool operator()(const Arc &arc) const { return true; }
+};
+
+// True for (input/output) epsilon arcs.
+template <class Arc>
+class EpsilonArcFilter {
+ public:
+  bool operator()(const Arc &arc) const {
+    return arc.ilabel == 0 && arc.olabel == 0;
+  }
+};
+
+// True for input epsilon arcs.
+template <class Arc>
+class InputEpsilonArcFilter {
+ public:
+  bool operator()(const Arc &arc) const { return arc.ilabel == 0; }
+};
+
+// True for output epsilon arcs.
+template <class Arc>
+class OutputEpsilonArcFilter {
+ public:
+  bool operator()(const Arc &arc) const { return arc.olabel == 0; }
+};
+
+// True if specified label matches (doesn't match) when keep_match is
+// true (false).
+template <class Arc>
+class LabelArcFilter {
+ public:
+  using Label = typename Arc::Label;
+
+  explicit LabelArcFilter(Label label, bool match_input = true,
+                          bool keep_match = true)
+      : label_(label), match_input_(match_input), keep_match_(keep_match) {}
+
+  bool operator()(const Arc &arc) const {
+    const bool match = (match_input_ ? arc.ilabel : arc.olabel) == label_;
+    return keep_match_ ? match : !match;
+  }
+
+ private:
+  const Label label_;
+  const bool match_input_;
+  const bool keep_match_;
+};
+
+// True if specified labels match (don't match) when keep_match is true (false).
+template <class Arc>
+class MultiLabelArcFilter {
+ public:
+  using Label = typename Arc::Label;
+
+  explicit MultiLabelArcFilter(bool match_input = true, bool keep_match = true)
+      : match_input_(match_input), keep_match_(keep_match) {}
+
+  bool operator()(const Arc &arc) const {
+    const Label label = match_input_ ? arc.ilabel : arc.olabel;
+    const bool match = labels_.Find(label) != labels_.End();
+    return keep_match_ ? match : !match;
+  }
+
+  void AddLabel(Label label) { labels_.Insert(label); }
+
+ private:
+  CompactSet<Label, kNoLabel> labels_;
+  const bool match_input_;
+  const bool keep_match_;
+};
+
+}  // namespace fst
+
+#endif  // FST_ARCFILTER_H_
diff --git a/projects/llm_framework/include/fst/arcsort.h b/projects/llm_framework/include/fst/arcsort.h
new file mode 100644
index 00000000..b5ab50e0
--- /dev/null
+++ b/projects/llm_framework/include/fst/arcsort.h
@@ -0,0 +1,211 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes to sort arcs in an FST.
+
+#ifndef FST_ARCSORT_H_
+#define FST_ARCSORT_H_
+
+#include <algorithm>
+#include <string>
+#include <vector>
+
+#include <fst/cache.h>
+#include <fst/state-map.h>
+#include <fst/test-properties.h>
+
+
+namespace fst {
+
+template <class Arc, class Compare>
+class ArcSortMapper {
+ public:
+  using FromArc = Arc;
+  using ToArc = Arc;
+
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  constexpr ArcSortMapper(const Fst<Arc> &fst, const Compare &comp)
+      : fst_(fst), comp_(comp), i_(0) {}
+
+  // Allows updating Fst argument; pass only if changed.
+  ArcSortMapper(const ArcSortMapper<Arc, Compare> &mapper,
+                const Fst<Arc> *fst = nullptr)
+      : fst_(fst ? *fst : mapper.fst_), comp_(mapper.comp_), i_(0) {}
+
+  StateId Start() { return fst_.Start(); }
+
+  Weight Final(StateId s) const { return fst_.Final(s); }
+
+  void SetState(StateId s) {
+    i_ = 0;
+    arcs_.clear();
+    arcs_.reserve(fst_.NumArcs(s));
+    for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
+      arcs_.push_back(aiter.Value());
+    }
+    std::sort(arcs_.begin(), arcs_.end(), comp_);
+  }
+
+  bool Done() const { return i_ >= arcs_.size(); }
+
+  const Arc &Value() const { return arcs_[i_]; }
+
+  void Next() { ++i_; }
+
+  MapSymbolsAction InputSymbolsAction() const { return MAP_COPY_SYMBOLS; }
+
+  MapSymbolsAction OutputSymbolsAction() const { return MAP_COPY_SYMBOLS; }
+
+  uint64 Properties(uint64 props) const { return comp_.Properties(props); }
+
+ private:
+  const Fst<Arc> &fst_;
+  const Compare &comp_;
+  std::vector<Arc> arcs_;
+  ssize_t i_;  // current arc position
+
+  ArcSortMapper &operator=(const ArcSortMapper &) = delete;
+};
+
+// Sorts the arcs in an FST according to function object 'comp' of type Compare.
+// This version modifies its input. Comparison function objects ILabelCompare
+// and OLabelCompare are provided by the library. In general, Compare must meet
+// the requirements for a  comparison function object (e.g., similar to those
+// used by std::sort). It must also have a member Properties(uint64) that
+// specifies the known properties of the sorted FST; it takes as argument the
+// input FST's known properties before the sort.
+//
+// Complexity:
+//
+// - Time: O(v d log d)
+// - Space: O(d)
+//
+// where v = # of states and d = maximum out-degree.
+template <class Arc, class Compare>
+void ArcSort(MutableFst<Arc> *fst, Compare comp) {
+  ArcSortMapper<Arc, Compare> mapper(*fst, comp);
+  StateMap(fst, mapper);
+}
+
+using ArcSortFstOptions = CacheOptions;
+
+// Sorts the arcs in an FST according to function object 'comp' of type Compare.
+// This version is a delayed FST. Comparsion function objects ILabelCompare and
+// OLabelCompare are provided by the library. In general, Compare must meet the
+// requirements for a comparision function object (e.g., similar to those
+// used by std::sort). It must also have a member Properties(uint64) that
+// specifies the known properties of the sorted FST; it takes as argument the
+// input FST's known properties.
+//
+// Complexity:
+//
+// - Time: O(v d log d)
+// - Space: O(d)
+//
+// where v = # of states visited, d = maximum out-degree of states visited.
+// Constant time and space to visit an input state is assumed and exclusive of
+// caching.
+template <class Arc, class Compare>
+class ArcSortFst : public StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>> {
+  using StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>>::GetImpl;
+
+ public:
+  using StateId = typename Arc::StateId;
+  using Mapper = ArcSortMapper<Arc, Compare>;
+
+  ArcSortFst(const Fst<Arc> &fst, const Compare &comp)
+      : StateMapFst<Arc, Arc, Mapper>(fst,
+                                      ArcSortMapper<Arc, Compare>(fst, comp)) {}
+
+  ArcSortFst(const Fst<Arc> &fst, const Compare &comp,
+             const ArcSortFstOptions &opts)
+      : StateMapFst<Arc, Arc, Mapper>(fst, Mapper(fst, comp), opts) {}
+
+  // See Fst<>::Copy() for doc.
+  ArcSortFst(const ArcSortFst<Arc, Compare> &fst, bool safe = false)
+      : StateMapFst<Arc, Arc, Mapper>(fst, safe) {}
+
+  // Gets a copy of this ArcSortFst. See Fst<>::Copy() for further doc.
+  ArcSortFst<Arc, Compare> *Copy(bool safe = false) const override {
+    return new ArcSortFst(*this, safe);
+  }
+
+  size_t NumArcs(StateId s) const override {
+    return GetImpl()->GetFst()->NumArcs(s);
+  }
+
+  size_t NumInputEpsilons(StateId s) const override {
+    return GetImpl()->GetFst()->NumInputEpsilons(s);
+  }
+
+  size_t NumOutputEpsilons(StateId s) const override {
+    return GetImpl()->GetFst()->NumOutputEpsilons(s);
+  }
+};
+
+// Specialization for ArcSortFst.
+template <class Arc, class Compare>
+class StateIterator<ArcSortFst<Arc, Compare>>
+    : public StateIterator<StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>>> {
+ public:
+  explicit StateIterator(const ArcSortFst<Arc, Compare> &fst)
+      : StateIterator<StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>>>(fst) {
+  }
+};
+
+// Specialization for ArcSortFst.
+template <class Arc, class Compare>
+class ArcIterator<ArcSortFst<Arc, Compare>>
+    : public ArcIterator<StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>>> {
+ public:
+  ArcIterator(const ArcSortFst<Arc, Compare> &fst, typename Arc::StateId s)
+      : ArcIterator<StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>>>(fst,
+                                                                        s) {}
+};
+
+// Compare class for comparing input labels of arcs.
+template <class Arc>
+class ILabelCompare {
+ public:
+  constexpr ILabelCompare() {}
+
+  constexpr bool operator()(const Arc &arc1, const Arc &arc2) const {
+    return arc1.ilabel < arc2.ilabel;
+  }
+
+  constexpr uint64 Properties(uint64 props) const {
+    return (props & kArcSortProperties) | kILabelSorted |
+           (props & kAcceptor ? kOLabelSorted : 0);
+  }
+};
+
+// Compare class for comparing output labels of arcs.
+template <class Arc>
+class OLabelCompare {
+ public:
+  constexpr OLabelCompare() {}
+
+  constexpr bool operator()(const Arc &arc1, const Arc &arc2) const {
+    return arc1.olabel < arc2.olabel;
+  }
+
+  constexpr uint64 Properties(uint64 props) const {
+    return (props & kArcSortProperties) | kOLabelSorted |
+           (props & kAcceptor ? kILabelSorted : 0);
+  }
+};
+
+// Useful aliases when using StdArc.
+
+template <class Compare>
+using StdArcSortFst = ArcSortFst<StdArc, Compare>;
+
+using StdILabelCompare = ILabelCompare<StdArc>;
+
+using StdOLabelCompare = OLabelCompare<StdArc>;
+
+}  // namespace fst
+
+#endif  // FST_ARCSORT_H_
diff --git a/projects/llm_framework/include/fst/bi-table.h b/projects/llm_framework/include/fst/bi-table.h
new file mode 100644
index 00000000..9651cfe8
--- /dev/null
+++ b/projects/llm_framework/include/fst/bi-table.h
@@ -0,0 +1,480 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Classes for representing a bijective mapping between an arbitrary entry
+// of type T and a signed integral ID.
+
+#ifndef FST_BI_TABLE_H_
+#define FST_BI_TABLE_H_
+
+#include <deque>
+#include <functional>
+#include <memory>
+#include <type_traits>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+#include <fst/log.h>
+#include <fst/memory.h>
+#include <unordered_set>
+
+namespace fst {
+
+// Bitables model bijective mappings between entries of an arbitrary type T and
+// an signed integral ID of type I. The IDs are allocated starting from 0 in
+// order.
+//
+// template <class I, class T>
+// class BiTable {
+//  public:
+//
+//   // Required constructors.
+//   BiTable();
+//
+//   // Looks up integer ID from entry. If it doesn't exist and insert
+//   / is true, adds it; otherwise, returns -1.
+//   I FindId(const T &entry, bool insert = true);
+//
+//   // Looks up entry from integer ID.
+//   const T &FindEntry(I) const;
+//
+//   // Returns number of stored entries.
+//   I Size() const;
+// };
+
+// An implementation using a hash map for the entry to ID mapping. H is the
+// hash function and E is the equality function. If passed to the constructor,
+// ownership is given to this class.
+template <class I, class T, class H, class E = std::equal_to<T>>
+class HashBiTable {
+ public:
+  // Reserves space for table_size elements. If passing H and E to the
+  // constructor, this class owns them.
+  explicit HashBiTable(size_t table_size = 0, H *h = nullptr, E *e = nullptr) :
+      hash_func_(h ? h : new H()), hash_equal_(e ? e : new E()),
+      entry2id_(table_size, *hash_func_, *hash_equal_) {
+    if (table_size) id2entry_.reserve(table_size);
+  }
+
+  HashBiTable(const HashBiTable<I, T, H, E> &table)
+      : hash_func_(new H(*table.hash_func_)),
+        hash_equal_(new E(*table.hash_equal_)),
+        entry2id_(table.entry2id_.begin(), table.entry2id_.end(),
+                  table.entry2id_.size(), *hash_func_, *hash_equal_),
+        id2entry_(table.id2entry_) {}
+
+  I FindId(const T &entry, bool insert = true) {
+    if (!insert) {
+      const auto it = entry2id_.find(entry);
+      return it == entry2id_.end() ? -1 : it->second - 1;
+    }
+    I &id_ref = entry2id_[entry];
+    if (id_ref == 0) {  // T not found; stores and assigns a new ID.
+      id2entry_.push_back(entry);
+      id_ref = id2entry_.size();
+    }
+    return id_ref - 1;  // NB: id_ref = ID + 1.
+  }
+
+  const T &FindEntry(I s) const { return id2entry_[s]; }
+
+  I Size() const { return id2entry_.size(); }
+
+  // TODO(riley): Add fancy clear-to-size, as in CompactHashBiTable.
+  void Clear() {
+    entry2id_.clear();
+    id2entry_.clear();
+  }
+
+ private:
+  std::unique_ptr<H> hash_func_;
+  std::unique_ptr<E> hash_equal_;
+  std::unordered_map<T, I, H, E> entry2id_;
+  std::vector<T> id2entry_;
+};
+
+// Enables alternative hash set representations below.
+enum HSType { HS_STL = 0, HS_DENSE = 1, HS_SPARSE = 2, HS_FLAT = 3 };
+
+// Default hash set is STL hash_set.
+template <class K, class H, class E, HSType HS>
+struct HashSet : public std::unordered_set<K, H, E, PoolAllocator<K>> {
+  explicit HashSet(size_t n = 0, const H &h = H(), const E &e = E())
+      : std::unordered_set<K, H, E, PoolAllocator<K>>(n, h, e) {}
+
+  void rehash(size_t n) {}
+};
+
+// An implementation using a hash set for the entry to ID mapping. The hash set
+// holds keys which are either the ID or kCurrentKey. These keys can be mapped
+// to entries either by looking up in the entry vector or, if kCurrentKey, in
+// current_entry_. The hash and key equality functions map to entries first. H
+// is the hash function and E is the equality function. If passed to the
+// constructor, ownership is given to this class.
+// TODO(rybach): remove support for (deprecated and unused) HS_DENSE, HS_SPARSE.
+template <class I, class T, class H, class E = std::equal_to<T>,
+          HSType HS = HS_FLAT>
+class CompactHashBiTable {
+ public:
+  friend class HashFunc;
+  friend class HashEqual;
+
+  // Reserves space for table_size elements. If passing H and E to the
+  // constructor, this class owns them.
+  explicit CompactHashBiTable(size_t table_size = 0, H *h = nullptr,
+                              E *e = nullptr) :
+        hash_func_(h ? h : new H()), hash_equal_(e ? e : new E()),
+        compact_hash_func_(*this), compact_hash_equal_(*this),
+        keys_(table_size, compact_hash_func_, compact_hash_equal_) {
+    if (table_size) id2entry_.reserve(table_size);
+  }
+
+  CompactHashBiTable(const CompactHashBiTable<I, T, H, E, HS> &table)
+      : hash_func_(new H(*table.hash_func_)),
+        hash_equal_(new E(*table.hash_equal_)),
+        compact_hash_func_(*this), compact_hash_equal_(*this),
+        keys_(table.keys_.size(), compact_hash_func_, compact_hash_equal_),
+        id2entry_(table.id2entry_) {
+    keys_.insert(table.keys_.begin(), table.keys_.end());
+  }
+
+  I FindId(const T &entry, bool insert = true) {
+    current_entry_ = &entry;
+    if (insert) {
+      auto result = keys_.insert(kCurrentKey);
+      if (!result.second) return *result.first;  // Already exists.
+      // Overwrites kCurrentKey with a new key value; this is safe because it
+      // doesn't affect hashing or equality testing.
+      I key = id2entry_.size();
+      const_cast<I &>(*result.first) = key;
+      id2entry_.push_back(entry);
+      return key;
+    }
+    const auto it = keys_.find(kCurrentKey);
+    return it == keys_.end() ? -1 : *it;
+  }
+
+  const T &FindEntry(I s) const { return id2entry_[s]; }
+
+  I Size() const { return id2entry_.size(); }
+
+  // Clears content; with argument, erases last n IDs.
+  void Clear(ssize_t n = -1) {
+    if (n < 0 || n >= id2entry_.size()) {  // Clears completely.
+      keys_.clear();
+      id2entry_.clear();
+    } else if (n == id2entry_.size() - 1) {  // Leaves only key 0.
+      const T entry = FindEntry(0);
+      keys_.clear();
+      id2entry_.clear();
+      FindId(entry, true);
+    } else {
+      while (n-- > 0) {
+        I key = id2entry_.size() - 1;
+        keys_.erase(key);
+        id2entry_.pop_back();
+      }
+      keys_.rehash(0);
+    }
+  }
+
+ private:
+  static_assert(std::is_signed<I>::value, "I must be a signed type");
+  // ... otherwise >= kCurrentKey comparisons as used below don't work.
+  // TODO(rybach): (1) remove kEmptyKey, kDeletedKey, (2) don't use >= for key
+  // comparison, (3) allow unsigned key types.
+  static constexpr I kCurrentKey = -1;
+  static constexpr I kEmptyKey = -2;
+  static constexpr I kDeletedKey = -3;
+
+  class HashFunc {
+   public:
+    explicit HashFunc(const CompactHashBiTable &ht) : ht_(&ht) {}
+
+    size_t operator()(I k) const {
+      if (k >= kCurrentKey) {
+        return (*ht_->hash_func_)(ht_->Key2Entry(k));
+      } else {
+        return 0;
+      }
+    }
+
+   private:
+    const CompactHashBiTable *ht_;
+  };
+
+  class HashEqual {
+   public:
+    explicit HashEqual(const CompactHashBiTable &ht) : ht_(&ht) {}
+
+    bool operator()(I k1, I k2) const {
+      if (k1 == k2) {
+        return true;
+      } else if (k1 >= kCurrentKey && k2 >= kCurrentKey) {
+        return (*ht_->hash_equal_)(ht_->Key2Entry(k1), ht_->Key2Entry(k2));
+      } else {
+        return false;
+      }
+    }
+
+   private:
+    const CompactHashBiTable *ht_;
+  };
+
+  using KeyHashSet = HashSet<I, HashFunc, HashEqual, HS>;
+
+  const T &Key2Entry(I k) const {
+    if (k == kCurrentKey) {
+      return *current_entry_;
+    } else {
+      return id2entry_[k];
+    }
+  }
+
+  std::unique_ptr<H> hash_func_;
+  std::unique_ptr<E> hash_equal_;
+  HashFunc compact_hash_func_;
+  HashEqual compact_hash_equal_;
+  KeyHashSet keys_;
+  std::vector<T> id2entry_;
+  const T *current_entry_;
+};
+
+template <class I, class T, class H, class E, HSType HS>
+constexpr I CompactHashBiTable<I, T, H, E, HS>::kCurrentKey;
+
+template <class I, class T, class H, class E, HSType HS>
+constexpr I CompactHashBiTable<I, T, H, E, HS>::kEmptyKey;
+
+template <class I, class T, class H, class E, HSType HS>
+constexpr I CompactHashBiTable<I, T, H, E, HS>::kDeletedKey;
+
+// An implementation using a vector for the entry to ID mapping. It is passed a
+// function object FP that should fingerprint entries uniquely to an integer
+// that can used as a vector index. Normally, VectorBiTable constructs the FP
+// object. The user can instead pass in this object; in that case, VectorBiTable
+// takes its ownership.
+template <class I, class T, class FP>
+class VectorBiTable {
+ public:
+  // Reserves table_size cells of space. If passing FP argument to the
+  // constructor, this class owns it.
+  explicit VectorBiTable(FP *fp = nullptr, size_t table_size = 0) :
+      fp_(fp ? fp : new FP()) {
+    if (table_size) id2entry_.reserve(table_size);
+  }
+
+  VectorBiTable(const VectorBiTable<I, T, FP> &table)
+      : fp_(new FP(*table.fp_)), fp2id_(table.fp2id_),
+        id2entry_(table.id2entry_) {}
+
+  I FindId(const T &entry, bool insert = true) {
+    ssize_t fp = (*fp_)(entry);
+    if (fp >= fp2id_.size()) fp2id_.resize(fp + 1);
+    I &id_ref = fp2id_[fp];
+    if (id_ref == 0) {  // T not found.
+      if (insert) {     // Stores and assigns a new ID.
+        id2entry_.push_back(entry);
+        id_ref = id2entry_.size();
+      } else {
+        return -1;
+      }
+    }
+    return id_ref - 1;  // NB: id_ref = ID + 1.
+  }
+
+  const T &FindEntry(I s) const { return id2entry_[s]; }
+
+  I Size() const { return id2entry_.size(); }
+
+  const FP &Fingerprint() const { return *fp_; }
+
+ private:
+  std::unique_ptr<FP> fp_;
+  std::vector<I> fp2id_;
+  std::vector<T> id2entry_;
+};
+
+// An implementation using a vector and a compact hash table. The selecting
+// functor S returns true for entries to be hashed in the vector. The
+// fingerprinting functor FP returns a unique fingerprint for each entry to be
+// hashed in the vector (these need to be suitable for indexing in a vector).
+// The hash functor H is used when hashing entry into the compact hash table.
+// If passed to the constructor, ownership is given to this class.
+template <class I, class T, class S, class FP, class H, HSType HS = HS_DENSE>
+class VectorHashBiTable {
+ public:
+  friend class HashFunc;
+  friend class HashEqual;
+
+  explicit VectorHashBiTable(S *s, FP *fp, H *h, size_t vector_size = 0,
+                             size_t entry_size = 0)
+      : selector_(s), fp_(fp), h_(h), hash_func_(*this), hash_equal_(*this),
+        keys_(0, hash_func_, hash_equal_) {
+    if (vector_size) fp2id_.reserve(vector_size);
+    if (entry_size) id2entry_.reserve(entry_size);
+  }
+
+  VectorHashBiTable(const VectorHashBiTable<I, T, S, FP, H, HS> &table)
+      : selector_(new S(table.s_)), fp_(new FP(*table.fp_)),
+        h_(new H(*table.h_)), id2entry_(table.id2entry_),
+        fp2id_(table.fp2id_), hash_func_(*this), hash_equal_(*this),
+        keys_(table.keys_.size(), hash_func_, hash_equal_) {
+    keys_.insert(table.keys_.begin(), table.keys_.end());
+  }
+
+  I FindId(const T &entry, bool insert = true) {
+    if ((*selector_)(entry)) {  // Uses the vector if selector_(entry) == true.
+      uint64 fp = (*fp_)(entry);
+      if (fp2id_.size() <= fp) fp2id_.resize(fp + 1, 0);
+      if (fp2id_[fp] == 0) {  // T not found.
+        if (insert) {         // Stores and assigns a new ID.
+          id2entry_.push_back(entry);
+          fp2id_[fp] = id2entry_.size();
+        } else {
+          return -1;
+        }
+      }
+      return fp2id_[fp] - 1;  // NB: assoc_value = ID + 1.
+    } else {                  // Uses the hash table otherwise.
+      current_entry_ = &entry;
+      const auto it = keys_.find(kCurrentKey);
+      if (it == keys_.end()) {
+        if (insert) {
+          I key = id2entry_.size();
+          id2entry_.push_back(entry);
+          keys_.insert(key);
+          return key;
+        } else {
+          return -1;
+        }
+      } else {
+        return *it;
+      }
+    }
+  }
+
+  const T &FindEntry(I s) const { return id2entry_[s]; }
+
+  I Size() const { return id2entry_.size(); }
+
+  const S &Selector() const { return *selector_; }
+
+  const FP &Fingerprint() const { return *fp_; }
+
+  const H &Hash() const { return *h_; }
+
+ private:
+  static constexpr I kCurrentKey = -1;
+  static constexpr I kEmptyKey = -2;
+
+  class HashFunc {
+   public:
+    explicit HashFunc(const VectorHashBiTable &ht) : ht_(&ht) {}
+
+    size_t operator()(I k) const {
+      if (k >= kCurrentKey) {
+        return (*(ht_->h_))(ht_->Key2Entry(k));
+      } else {
+        return 0;
+      }
+    }
+
+   private:
+    const VectorHashBiTable *ht_;
+  };
+
+  class HashEqual {
+   public:
+    explicit HashEqual(const VectorHashBiTable &ht) : ht_(&ht) {}
+
+    bool operator()(I k1, I k2) const {
+      if (k1 >= kCurrentKey && k2 >= kCurrentKey) {
+        return ht_->Key2Entry(k1) == ht_->Key2Entry(k2);
+      } else {
+        return k1 == k2;
+      }
+    }
+
+   private:
+    const VectorHashBiTable *ht_;
+  };
+
+  using KeyHashSet = HashSet<I, HashFunc, HashEqual, HS>;
+
+  const T &Key2Entry(I k) const {
+    if (k == kCurrentKey) {
+      return *current_entry_;
+    } else {
+      return id2entry_[k];
+    }
+  }
+
+  std::unique_ptr<S> selector_;  // True if entry hashed into vector.
+  std::unique_ptr<FP> fp_;       // Fingerprint used for hashing into vector.
+  std::unique_ptr<H> h_;         // Hash funcion used for hashing into hash_set.
+
+  std::vector<T> id2entry_;  // Maps state IDs to entry.
+  std::vector<I> fp2id_;     // Maps entry fingerprints to IDs.
+
+  // Compact implementation of the hash table mapping entries to state IDs
+  // using the hash function h_.
+  HashFunc hash_func_;
+  HashEqual hash_equal_;
+  KeyHashSet keys_;
+  const T *current_entry_;
+};
+
+template <class I, class T, class S, class FP, class H, HSType HS>
+constexpr I VectorHashBiTable<I, T, S, FP, H, HS>::kCurrentKey;
+
+template <class I, class T, class S, class FP, class H, HSType HS>
+constexpr I VectorHashBiTable<I, T, S, FP, H, HS>::kEmptyKey;
+
+// An implementation using a hash map for the entry to ID mapping. This version
+// permits erasing of arbitrary states. The entry T must have == defined and
+// its default constructor must produce a entry that will never be seen. F is
+// the hash function.
+template <class I, class T, class F>
+class ErasableBiTable {
+ public:
+  ErasableBiTable() : first_(0) {}
+
+  I FindId(const T &entry, bool insert = true) {
+    I &id_ref = entry2id_[entry];
+    if (id_ref == 0) {  // T not found.
+      if (insert) {     // Stores and assigns a new ID.
+        id2entry_.push_back(entry);
+        id_ref = id2entry_.size() + first_;
+      } else {
+        return -1;
+      }
+    }
+    return id_ref - 1;  // NB: id_ref = ID + 1.
+  }
+
+  const T &FindEntry(I s) const { return id2entry_[s - first_]; }
+
+  I Size() const { return id2entry_.size(); }
+
+  void Erase(I s) {
+    auto &ref = id2entry_[s - first_];
+    entry2id_.erase(ref);
+    ref = empty_entry_;
+    while (!id2entry_.empty() && id2entry_.front() == empty_entry_) {
+      id2entry_.pop_front();
+      ++first_;
+    }
+  }
+
+ private:
+  std::unordered_map<T, I, F> entry2id_;
+  std::deque<T> id2entry_;
+  const T empty_entry_;
+  I first_;  // I of first element in the deque.
+};
+
+}  // namespace fst
+
+#endif  // FST_BI_TABLE_H_
diff --git a/projects/llm_framework/include/fst/cache.h b/projects/llm_framework/include/fst/cache.h
new file mode 100644
index 00000000..13b7cf81
--- /dev/null
+++ b/projects/llm_framework/include/fst/cache.h
@@ -0,0 +1,1327 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// An FST implementation that caches FST elements of a delayed computation.
+
+#ifndef FST_CACHE_H_
+#define FST_CACHE_H_
+
+#include <algorithm>
+#include <functional>
+#include <list>
+#include <vector>
+
+#include <fst/flags.h>
+#include <fst/log.h>
+
+#include <fst/vector-fst.h>
+
+#include <unordered_map>
+
+DECLARE_bool(fst_default_cache_gc);
+DECLARE_int64(fst_default_cache_gc_limit);
+
+namespace fst {
+
+// Options for controlling caching behavior; higher level than CacheImplOptions.
+struct CacheOptions {
+  bool gc;          // Enables GC.
+  size_t gc_limit;  // Number of bytes allowed before GC.
+
+  explicit CacheOptions(bool gc = FLAGS_fst_default_cache_gc,
+                        size_t gc_limit = FLAGS_fst_default_cache_gc_limit)
+      : gc(gc), gc_limit(gc_limit) {}
+};
+
+// Options for controlling caching behavior, at a lower level than
+// CacheOptions; templated on the cache store and allows passing the store.
+template <class CacheStore>
+struct CacheImplOptions {
+  bool gc;            // Enables GC.
+  size_t gc_limit;    // Number of bytes allowed before GC.
+  CacheStore *store;  // Cache store.
+  bool own_store;     // Should CacheImpl takes ownership of the store?
+
+  explicit CacheImplOptions(bool gc = FLAGS_fst_default_cache_gc,
+                            size_t gc_limit = FLAGS_fst_default_cache_gc_limit,
+                            CacheStore *store = nullptr)
+      : gc(gc), gc_limit(gc_limit), store(store), own_store(true) {}
+
+  explicit CacheImplOptions(const CacheOptions &opts)
+      : gc(opts.gc), gc_limit(opts.gc_limit), store(nullptr), own_store(true) {}
+};
+
+// Cache flags.
+constexpr uint32 kCacheFinal = 0x0001;   // Final weight has been cached.
+constexpr uint32 kCacheArcs = 0x0002;    // Arcs have been cached.
+constexpr uint32 kCacheInit = 0x0004;    // Initialized by GC.
+constexpr uint32 kCacheRecent = 0x0008;  // Visited since GC.
+constexpr uint32 kCacheFlags =
+    kCacheFinal | kCacheArcs | kCacheInit | kCacheRecent;
+
+// Cache state, with arcs stored in a per-state std::vector.
+template <class A, class M = PoolAllocator<A>>
+class CacheState {
+ public:
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using ArcAllocator = M;
+  using StateAllocator =
+      typename ArcAllocator::template rebind<CacheState<A, M>>::other;
+
+  // Provides STL allocator for arcs.
+  explicit CacheState(const ArcAllocator &alloc)
+      : final_(Weight::Zero()),
+        niepsilons_(0),
+        noepsilons_(0),
+        arcs_(alloc),
+        flags_(0),
+        ref_count_(0) {}
+
+  CacheState(const CacheState<A> &state, const ArcAllocator &alloc)
+      : final_(state.Final()),
+        niepsilons_(state.NumInputEpsilons()),
+        noepsilons_(state.NumOutputEpsilons()),
+        arcs_(state.arcs_.begin(), state.arcs_.end(), alloc),
+        flags_(state.Flags()),
+        ref_count_(0) {}
+
+  void Reset() {
+    final_ = Weight::Zero();
+    niepsilons_ = 0;
+    noepsilons_ = 0;
+    ref_count_ = 0;
+    flags_ = 0;
+    arcs_.clear();
+  }
+
+  Weight Final() const { return final_; }
+
+  size_t NumInputEpsilons() const { return niepsilons_; }
+
+  size_t NumOutputEpsilons() const { return noepsilons_; }
+
+  size_t NumArcs() const { return arcs_.size(); }
+
+  const Arc &GetArc(size_t n) const { return arcs_[n]; }
+
+  // Used by the ArcIterator<Fst<Arc>> efficient implementation.
+  const Arc *Arcs() const { return !arcs_.empty() ? &arcs_[0] : nullptr; }
+
+  // Accesses flags; used by the caller.
+  uint32 Flags() const { return flags_; }
+
+  // Accesses ref count; used by the caller.
+  int RefCount() const { return ref_count_; }
+
+  void SetFinal(Weight weight) { final_ = std::move(weight); }
+
+  void ReserveArcs(size_t n) { arcs_.reserve(n); }
+
+  // Adds one arc at a time with all needed book-keeping; use PushArc and
+  // SetArcs for a more efficient alternative.
+  void AddArc(const Arc &arc) {
+    IncrementNumEpsilons(arc);
+    arcs_.push_back(arc);
+  }
+
+  void AddArc(Arc &&arc) {
+    IncrementNumEpsilons(arc);
+    arcs_.push_back(std::move(arc));
+  }
+
+  // Adds one arc at a time with delayed book-keeping; finalize with SetArcs().
+  void PushArc(const Arc &arc) { arcs_.push_back(arc); }
+
+  void PushArc(Arc &&arc) { arcs_.push_back(std::move(arc)); }
+
+  // Adds one arc at a time with delayed book-keeping; finalize with SetArcs().
+  template <class... T>
+  void EmplaceArc(T &&... ctor_args) {
+    arcs_.emplace_back(std::forward<T>(ctor_args)...);
+  }
+
+  // Finalizes arcs book-keeping; call only once.
+  void SetArcs() {
+    for (const auto &arc : arcs_) {
+      IncrementNumEpsilons(arc);
+    }
+  }
+
+  // Modifies nth arc.
+  void SetArc(const Arc &arc, size_t n) {
+    if (arcs_[n].ilabel == 0) --niepsilons_;
+    if (arcs_[n].olabel == 0) --noepsilons_;
+    IncrementNumEpsilons(arc);
+    arcs_[n] = arc;
+  }
+
+  // Deletes all arcs.
+  void DeleteArcs() {
+    niepsilons_ = 0;
+    noepsilons_ = 0;
+    arcs_.clear();
+  }
+
+  void DeleteArcs(size_t n) {
+    for (size_t i = 0; i < n; ++i) {
+      if (arcs_.back().ilabel == 0) --niepsilons_;
+      if (arcs_.back().olabel == 0) --noepsilons_;
+      arcs_.pop_back();
+    }
+  }
+
+  // Sets status flags; used by the caller.
+  void SetFlags(uint32 flags, uint32 mask) const {
+    flags_ &= ~mask;
+    flags_ |= flags;
+  }
+
+  // Mutates reference counts; used by the caller.
+
+  int IncrRefCount() const { return ++ref_count_; }
+
+  int DecrRefCount() const { return --ref_count_; }
+
+  // Used by the ArcIterator<Fst<Arc>> efficient implementation.
+  int *MutableRefCount() const { return &ref_count_; }
+
+  // Used for state class allocation.
+  void *operator new(size_t size, StateAllocator *alloc) {
+    return alloc->allocate(1);
+  }
+
+  // For state destruction and memory freeing.
+  static void Destroy(CacheState<Arc> *state, StateAllocator *alloc) {
+    if (state) {
+      state->~CacheState<Arc>();
+      alloc->deallocate(state, 1);
+    }
+  }
+
+ private:
+  // Update the number of epsilons as a result of having added an arc.
+  void IncrementNumEpsilons(const Arc &arc) {
+    if (arc.ilabel == 0) ++niepsilons_;
+    if (arc.olabel == 0) ++noepsilons_;
+  }
+
+  Weight final_;                         // Final weight.
+  size_t niepsilons_;                    // # of input epsilons.
+  size_t noepsilons_;                    // # of output epsilons.
+  std::vector<Arc, ArcAllocator> arcs_;  // Arcs representation.
+  mutable uint32 flags_;
+  mutable int ref_count_;  // If 0, available for GC.
+};
+
+// Cache store, allocating and storing states, providing a mapping from state
+// IDs to cached states, and an iterator over these states. The state template
+// argument must implement the CacheState interface. The state for a StateId s
+// is constructed when requested by GetMutableState(s) if it is not yet stored.
+// Initially, a state has a reference count of zero, but the user may increment
+// or decrement this to control the time of destruction. In particular, a state
+// is destroyed when:
+//
+// 1. This instance is destroyed, or
+// 2. Clear() or Delete() is called, or
+// 3. Possibly (implementation-dependently) when:
+//    - Garbage collection is enabled (as defined by opts.gc),
+//    - The cache store size exceeds the limits (as defined by opts.gc_limits),
+//    - The state's reference count is zero, and
+//    - The state is not the most recently requested state.
+//
+// template <class S>
+// class CacheStore {
+//  public:
+//   using State = S;
+//   using Arc = typename State::Arc;
+//   using StateId = typename Arc::StateId;
+//
+//   // Required constructors/assignment operators.
+//   explicit CacheStore(const CacheOptions &opts);
+//
+//   // Returns nullptr if state is not stored.
+//   const State *GetState(StateId s);
+//
+//   // Creates state if state is not stored.
+//   State *GetMutableState(StateId s);
+//
+//   // Similar to State::AddArc() but updates cache store book-keeping.
+//   void AddArc(State *state, const Arc &arc);
+//
+//   // Similar to State::SetArcs() but updates cache store book-keeping; call
+//   // only once.
+//   void SetArcs(State *state);
+//
+//   // Similar to State::DeleteArcs() but updates cache store book-keeping.
+//
+//   void DeleteArcs(State *state);
+//
+//   void DeleteArcs(State *state, size_t n);
+//
+//   // Deletes all cached states.
+//   void Clear();
+//
+//   // Number of cached states.
+//   StateId CountStates();
+//
+//   // Iterates over cached states (in an arbitrary order); only needed if
+//   // opts.gc is true.
+//   bool Done() const;      // End of iteration.
+//   StateId Value() const;  // Current state.
+//   void Next();            // Advances to next state (when !Done).
+//   void Reset();           // Returns to initial condition.
+//   void Delete();          // Deletes current state and advances to next.
+// };
+
+// Container cache stores.
+
+// This class uses a vector of pointers to states to store cached states.
+template <class S>
+class VectorCacheStore {
+ public:
+  using State = S;
+  using Arc = typename State::Arc;
+  using StateId = typename Arc::StateId;
+  using StateList = std::list<StateId, PoolAllocator<StateId>>;
+
+  // Required constructors/assignment operators.
+  explicit VectorCacheStore(const CacheOptions &opts) : cache_gc_(opts.gc) {
+    Clear();
+    Reset();
+  }
+
+  VectorCacheStore(const VectorCacheStore<S> &store)
+      : cache_gc_(store.cache_gc_) {
+    CopyStates(store);
+    Reset();
+  }
+
+  ~VectorCacheStore() { Clear(); }
+
+  VectorCacheStore<State> &operator=(const VectorCacheStore<State> &store) {
+    if (this != &store) {
+      CopyStates(store);
+      Reset();
+    }
+    return *this;
+  }
+
+  bool InBounds(StateId s) const {
+    return s < static_cast<StateId>(state_vec_.size());
+  }
+
+  // Returns nullptr if state is not stored.
+  const State *GetState(StateId s) const {
+    return InBounds(s) ? state_vec_[s] : nullptr;
+  }
+
+  // Creates state if state is not stored.
+  State *GetMutableState(StateId s) {
+    State *state = nullptr;
+    if (InBounds(s)) {
+      state = state_vec_[s];
+    } else {
+      state_vec_.resize(s + 1, nullptr);
+    }
+    if (!state) {
+      state = new (&state_alloc_) State(arc_alloc_);
+      state_vec_[s] = state;
+      if (cache_gc_) state_list_.push_back(s);
+    }
+    return state;
+  }
+
+  // Similar to State::AddArc() but updates cache store book-keeping
+  void AddArc(State *state, const Arc &arc) { state->AddArc(arc); }
+
+  // Similar to State::SetArcs() but updates cache store book-keeping; call
+  // only once.
+  void SetArcs(State *state) { state->SetArcs(); }
+
+  // Deletes all arcs.
+  void DeleteArcs(State *state) { state->DeleteArcs(); }
+
+  // Deletes some arcs.
+  void DeleteArcs(State *state, size_t n) { state->DeleteArcs(n); }
+
+  // Deletes all cached states.
+  void Clear() {
+    for (State *s : state_vec_) {
+      State::Destroy(s, &state_alloc_);
+    }
+    state_vec_.clear();
+    state_list_.clear();
+  }
+
+  StateId CountStates() const {
+    return std::count_if(state_vec_.begin(), state_vec_.end(),
+                         [](const State *s) { return s != nullptr; });
+  }
+
+  // Iterates over cached states (in an arbitrary order); only works if GC is
+  // enabled (o.w. avoiding state_list_ overhead).
+  bool Done() const { return iter_ == state_list_.end(); }
+
+  StateId Value() const { return *iter_; }
+
+  void Next() { ++iter_; }
+
+  void Reset() { iter_ = state_list_.begin(); }
+
+  // Deletes current state and advances to next.
+  void Delete() {
+    State::Destroy(state_vec_[*iter_], &state_alloc_);
+    state_vec_[*iter_] = nullptr;
+    state_list_.erase(iter_++);
+  }
+
+ private:
+  void CopyStates(const VectorCacheStore<State> &store) {
+    Clear();
+    state_vec_.reserve(store.state_vec_.size());
+    for (size_t s = 0; s < store.state_vec_.size(); ++s) {
+      State *state = nullptr;
+      const auto *store_state = store.state_vec_[s];
+      if (store_state) {
+        state = new (&state_alloc_) State(*store_state, arc_alloc_);
+        if (cache_gc_) state_list_.push_back(s);
+      }
+      state_vec_.push_back(state);
+    }
+  }
+
+  bool cache_gc_;                               // Supports iteration when true.
+  std::vector<State *> state_vec_;              // Vector of states (or null).
+  StateList state_list_;                        // List of states.
+  typename StateList::iterator iter_;           // State list iterator.
+  typename State::StateAllocator state_alloc_;  // For state allocation.
+  typename State::ArcAllocator arc_alloc_;      // For arc allocation.
+};
+
+// This class uses a hash map from state IDs to pointers to cached states.
+template <class S>
+class HashCacheStore {
+ public:
+  using State = S;
+  using Arc = typename State::Arc;
+  using StateId = typename Arc::StateId;
+
+  using StateMap =
+      std::unordered_map<StateId, State *, std::hash<StateId>,
+                          std::equal_to<StateId>,
+                          PoolAllocator<std::pair<const StateId, State *>>>;
+
+  // Required constructors/assignment operators.
+  explicit HashCacheStore(const CacheOptions &opts) {
+    Clear();
+    Reset();
+  }
+
+  HashCacheStore(const HashCacheStore<S> &store) {
+    CopyStates(store);
+    Reset();
+  }
+
+  ~HashCacheStore() { Clear(); }
+
+  HashCacheStore<State> &operator=(const HashCacheStore<State> &store) {
+    if (this != &store) {
+      CopyStates(store);
+      Reset();
+    }
+    return *this;
+  }
+
+  // Returns nullptr if state is not stored.
+  const State *GetState(StateId s) const {
+    const auto it = state_map_.find(s);
+    return it != state_map_.end() ? it->second : nullptr;
+  }
+
+  // Creates state if state is not stored.
+  State *GetMutableState(StateId s) {
+    auto *&state = state_map_[s];
+    if (!state) state = new (&state_alloc_) State(arc_alloc_);
+    return state;
+  }
+
+  // Similar to State::AddArc() but updates cache store book-keeping.
+  void AddArc(State *state, const Arc &arc) { state->AddArc(arc); }
+
+  // Similar to State::SetArcs() but updates internal cache size; call only
+  // once.
+  void SetArcs(State *state) { state->SetArcs(); }
+
+  // Deletes all arcs.
+  void DeleteArcs(State *state) { state->DeleteArcs(); }
+
+  // Deletes some arcs.
+  void DeleteArcs(State *state, size_t n) { state->DeleteArcs(n); }
+
+  // Deletes all cached states.
+  void Clear() {
+    for (auto it = state_map_.begin(); it != state_map_.end(); ++it) {
+      State::Destroy(it->second, &state_alloc_);
+    }
+    state_map_.clear();
+  }
+
+  StateId CountStates() const { return state_map_.size(); }
+
+  // Iterates over cached states (in an arbitrary order).
+  bool Done() const { return iter_ == state_map_.end(); }
+
+  StateId Value() const { return iter_->first; }
+
+  void Next() { ++iter_; }
+
+  void Reset() { iter_ = state_map_.begin(); }
+
+  // Deletes current state and advances to next.
+  void Delete() {
+    State::Destroy(iter_->second, &state_alloc_);
+    state_map_.erase(iter_++);
+  }
+
+ private:
+  void CopyStates(const HashCacheStore<State> &store) {
+    Clear();
+    for (auto it = store.state_map_.begin(); it != store.state_map_.end();
+         ++it) {
+      state_map_[it->first] =
+          new (&state_alloc_) State(*it->second, arc_alloc_);
+    }
+  }
+
+  StateMap state_map_;                          // Map from state ID to state.
+  typename StateMap::iterator iter_;            // State map iterator.
+  typename State::StateAllocator state_alloc_;  // For state allocation.
+  typename State::ArcAllocator arc_alloc_;      // For arc allocation.
+};
+
+// Garbage-colllection cache stores.
+
+// This class implements a simple garbage collection scheme when
+// 'opts.gc_limit = 0'. In particular, the first cached state is reused for each
+// new state so long as the reference count is zero on the to-be-reused state.
+// Otherwise, the full underlying store is used. The caller can increment the
+// reference count to inhibit the GC of in-use states (e.g., in an ArcIterator).
+//
+// The typical use case for this optimization is when a single pass over a
+// cached
+// FST is performed with only one-state expanded at a time.
+template <class CacheStore>
+class FirstCacheStore {
+ public:
+  using State = typename CacheStore::State;
+  using Arc = typename State::Arc;
+  using StateId = typename Arc::StateId;
+
+  // Required constructors/assignment operators.
+  explicit FirstCacheStore(const CacheOptions &opts)
+      : store_(opts),
+        cache_gc_(opts.gc_limit == 0),  // opts.gc ignored historically.
+        cache_first_state_id_(kNoStateId),
+        cache_first_state_(nullptr) {}
+
+  FirstCacheStore(const FirstCacheStore<CacheStore> &store)
+      : store_(store.store_),
+        cache_gc_(store.cache_gc_),
+        cache_first_state_id_(store.cache_first_state_id_),
+        cache_first_state_(store.cache_first_state_id_ != kNoStateId
+                               ? store_.GetMutableState(0)
+                               : nullptr) {}
+
+  FirstCacheStore<CacheStore> &operator=(
+      const FirstCacheStore<CacheStore> &store) {
+    if (this != &store) {
+      store_ = store.store_;
+      cache_gc_ = store.cache_gc_;
+      cache_first_state_id_ = store.cache_first_state_id_;
+      cache_first_state_ = store.cache_first_state_id_ != kNoStateId
+                               ? store_.GetMutableState(0)
+                               : nullptr;
+    }
+    return *this;
+  }
+
+  // Returns nullptr if state is not stored.
+  const State *GetState(StateId s) const {
+    // store_ state 0 may hold first cached state; the rest are shifted by 1.
+    return s == cache_first_state_id_ ? cache_first_state_
+                                      : store_.GetState(s + 1);
+  }
+
+  // Creates state if state is not stored.
+  State *GetMutableState(StateId s) {
+    // store_ state 0 used to hold first cached state; the rest are shifted by
+    // 1.
+    if (cache_first_state_id_ == s) {
+      return cache_first_state_;  // Request for first cached state.
+    }
+    if (cache_gc_) {
+      if (cache_first_state_id_ == kNoStateId) {
+        cache_first_state_id_ = s;  // Sets first cached state.
+        cache_first_state_ = store_.GetMutableState(0);
+        cache_first_state_->SetFlags(kCacheInit, kCacheInit);
+        cache_first_state_->ReserveArcs(2 * kAllocSize);
+        return cache_first_state_;
+      } else if (cache_first_state_->RefCount() == 0) {
+        cache_first_state_id_ = s;  // Updates first cached state.
+        cache_first_state_->Reset();
+        cache_first_state_->SetFlags(kCacheInit, kCacheInit);
+        return cache_first_state_;
+      } else {  // Keeps first cached state.
+        cache_first_state_->SetFlags(0, kCacheInit);  // Clears initialized bit.
+        cache_gc_ = false;                            // Disables GC.
+      }
+    }
+    auto *state = store_.GetMutableState(s + 1);
+    return state;
+  }
+
+  // Similar to State::AddArc() but updates cache store book-keeping.
+  void AddArc(State *state, const Arc &arc) { store_.AddArc(state, arc); }
+
+  // Similar to State::SetArcs() but updates internal cache size; call only
+  // once.
+  void SetArcs(State *state) { store_.SetArcs(state); }
+
+  // Deletes all arcs
+  void DeleteArcs(State *state) { store_.DeleteArcs(state); }
+
+  // Deletes some arcs
+  void DeleteArcs(State *state, size_t n) { store_.DeleteArcs(state, n); }
+
+  // Deletes all cached states
+  void Clear() {
+    store_.Clear();
+    cache_first_state_id_ = kNoStateId;
+    cache_first_state_ = nullptr;
+  }
+
+  StateId CountStates() const { return store_.CountStates(); }
+
+  // Iterates over cached states (in an arbitrary order). Only needed if GC is
+  // enabled.
+  bool Done() const { return store_.Done(); }
+
+  StateId Value() const {
+    // store_ state 0 may hold first cached state; rest shifted + 1.
+    const auto s = store_.Value();
+    return s ? s - 1 : cache_first_state_id_;
+  }
+
+  void Next() { store_.Next(); }
+
+  void Reset() { store_.Reset(); }
+
+  // Deletes current state and advances to next.
+  void Delete() {
+    if (Value() == cache_first_state_id_) {
+      cache_first_state_id_ = kNoStateId;
+      cache_first_state_ = nullptr;
+    }
+    store_.Delete();
+  }
+
+ private:
+  CacheStore store_;              // Underlying store.
+  bool cache_gc_;                 // GC enabled.
+  StateId cache_first_state_id_;  // First cached state ID.
+  State *cache_first_state_;      // First cached state.
+};
+
+// This class implements mark-sweep garbage collection on an underlying cache
+// store. If GC is enabled, garbage collection of states is performed in a
+// rough approximation of LRU order once when 'gc_limit' bytes is reached. The
+// caller can increment the reference count to inhibit the GC of in-use state
+// (e.g., in an ArcIterator). With GC enabled, the 'gc_limit' parameter allows
+// the caller to trade-off time vs. space.
+template <class CacheStore>
+class GCCacheStore {
+ public:
+  using State = typename CacheStore::State;
+  using Arc = typename State::Arc;
+  using StateId = typename Arc::StateId;
+
+  // Required constructors/assignment operators.
+  explicit GCCacheStore(const CacheOptions &opts)
+      : store_(opts),
+        cache_gc_request_(opts.gc),
+        cache_limit_(opts.gc_limit > kMinCacheLimit ? opts.gc_limit
+                                                    : kMinCacheLimit),
+        cache_gc_(false),
+        cache_size_(0) {}
+
+  // Returns 0 if state is not stored.
+  const State *GetState(StateId s) const { return store_.GetState(s); }
+
+  // Creates state if state is not stored
+  State *GetMutableState(StateId s) {
+    auto *state = store_.GetMutableState(s);
+    if (cache_gc_request_ && !(state->Flags() & kCacheInit)) {
+      state->SetFlags(kCacheInit, kCacheInit);
+      cache_size_ += sizeof(State) + state->NumArcs() * sizeof(Arc);
+      // GC is enabled once an uninited state (from underlying store) is seen.
+      cache_gc_ = true;
+      if (cache_size_ > cache_limit_) GC(state, false);
+    }
+    return state;
+  }
+
+  // Similar to State::AddArc() but updates cache store book-keeping.
+  void AddArc(State *state, const Arc &arc) {
+    store_.AddArc(state, arc);
+    if (cache_gc_ && (state->Flags() & kCacheInit)) {
+      cache_size_ += sizeof(Arc);
+      if (cache_size_ > cache_limit_) GC(state, false);
+    }
+  }
+
+  // Similar to State::SetArcs() but updates internal cache size; call only
+  // once.
+  void SetArcs(State *state) {
+    store_.SetArcs(state);
+    if (cache_gc_ && (state->Flags() & kCacheInit)) {
+      cache_size_ += state->NumArcs() * sizeof(Arc);
+      if (cache_size_ > cache_limit_) GC(state, false);
+    }
+  }
+
+  // Deletes all arcs.
+  void DeleteArcs(State *state) {
+    if (cache_gc_ && (state->Flags() & kCacheInit)) {
+      cache_size_ -= state->NumArcs() * sizeof(Arc);
+    }
+    store_.DeleteArcs(state);
+  }
+
+  // Deletes some arcs.
+  void DeleteArcs(State *state, size_t n) {
+    if (cache_gc_ && (state->Flags() & kCacheInit)) {
+      cache_size_ -= n * sizeof(Arc);
+    }
+    store_.DeleteArcs(state, n);
+  }
+
+  // Deletes all cached states.
+  void Clear() {
+    store_.Clear();
+    cache_size_ = 0;
+  }
+
+  StateId CountStates() const { return store_.CountStates(); }
+
+  // Iterates over cached states (in an arbitrary order); only needed if GC is
+  // enabled.
+  bool Done() const { return store_.Done(); }
+
+  StateId Value() const { return store_.Value(); }
+
+  void Next() { store_.Next(); }
+
+  void Reset() { store_.Reset(); }
+
+  // Deletes current state and advances to next.
+  void Delete() {
+    if (cache_gc_) {
+      const auto *state = store_.GetState(Value());
+      if (state->Flags() & kCacheInit) {
+        cache_size_ -= sizeof(State) + state->NumArcs() * sizeof(Arc);
+      }
+    }
+    store_.Delete();
+  }
+
+  // Removes from the cache store (not referenced-counted and not the current)
+  // states that have not been accessed since the last GC until at most
+  // cache_fraction * cache_limit_ bytes are cached. If that fails to free
+  // enough, attempts to uncaching recently visited states as well. If still
+  // unable to free enough memory, then widens cache_limit_.
+  void GC(const State *current, bool free_recent, float cache_fraction = 0.666);
+
+  // Returns the current cache size in bytes or 0 if GC is disabled.
+  size_t CacheSize() const { return cache_size_; }
+
+  // Returns the cache limit in bytes.
+  size_t CacheLimit() const { return cache_limit_; }
+
+ private:
+  static constexpr size_t kMinCacheLimit = 8096;  // Minimum cache limit.
+
+  CacheStore store_;       // Underlying store.
+  bool cache_gc_request_;  // GC requested but possibly not yet enabled.
+  size_t cache_limit_;     // Number of bytes allowed before GC.
+  bool cache_gc_;          // GC enabled
+  size_t cache_size_;      // Number of bytes cached.
+};
+
+template <class CacheStore>
+void GCCacheStore<CacheStore>::GC(const State *current, bool free_recent,
+                                  float cache_fraction) {
+  if (!cache_gc_) return;
+  VLOG(2) << "GCCacheStore: Enter GC: object = "
+          << "(" << this << "), free recently cached = " << free_recent
+          << ", cache size = " << cache_size_
+          << ", cache frac = " << cache_fraction
+          << ", cache limit = " << cache_limit_ << "\n";
+  size_t cache_target = cache_fraction * cache_limit_;
+  store_.Reset();
+  while (!store_.Done()) {
+    auto *state = store_.GetMutableState(store_.Value());
+    if (cache_size_ > cache_target && state->RefCount() == 0 &&
+        (free_recent || !(state->Flags() & kCacheRecent)) && state != current) {
+      if (state->Flags() & kCacheInit) {
+        size_t size = sizeof(State) + state->NumArcs() * sizeof(Arc);
+        if (size < cache_size_) {
+          cache_size_ -= size;
+        }
+      }
+      store_.Delete();
+    } else {
+      state->SetFlags(0, kCacheRecent);
+      store_.Next();
+    }
+  }
+  if (!free_recent && cache_size_ > cache_target) {  // Recurses on recent.
+    GC(current, true, cache_fraction);
+  } else if (cache_target > 0) {  // Widens cache limit.
+    while (cache_size_ > cache_target) {
+      cache_limit_ *= 2;
+      cache_target *= 2;
+    }
+  } else if (cache_size_ > 0) {
+    FSTERROR() << "GCCacheStore:GC: Unable to free all cached states";
+  }
+  VLOG(2) << "GCCacheStore: Exit GC: object = "
+          << "(" << this << "), free recently cached = " << free_recent
+          << ", cache size = " << cache_size_
+          << ", cache frac = " << cache_fraction
+          << ", cache limit = " << cache_limit_ << "\n";
+}
+
+template <class CacheStore>
+constexpr size_t GCCacheStore<CacheStore>::kMinCacheLimit;
+
+// This class is the default cache state and store used by CacheBaseImpl.
+// It uses VectorCacheStore for storage decorated by FirstCacheStore
+// and GCCacheStore to do (optional) garbage collection.
+template <class Arc>
+class DefaultCacheStore
+    : public GCCacheStore<FirstCacheStore<VectorCacheStore<CacheState<Arc>>>> {
+ public:
+  explicit DefaultCacheStore(const CacheOptions &opts)
+      : GCCacheStore<FirstCacheStore<VectorCacheStore<CacheState<Arc>>>>(opts) {
+  }
+};
+
+namespace internal {
+
+// This class is used to cache FST elements stored in states of type State
+// (see CacheState) with the flags used to indicate what has been cached. Use
+// HasStart(), HasFinal(), and HasArcs() to determine if cached and SetStart(),
+// SetFinal(), AddArc(), (or PushArc() and SetArcs()) to cache. Note that you
+// must set the final weight even if the state is non-final to mark it as
+// cached. The state storage method and any garbage collection policy are
+// determined by the cache store. If the store is passed in with the options,
+// CacheBaseImpl takes ownership.
+template <class State,
+          class CacheStore = DefaultCacheStore<typename State::Arc>>
+class CacheBaseImpl : public FstImpl<typename State::Arc> {
+ public:
+  using Arc = typename State::Arc;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Store = CacheStore;
+
+  using FstImpl<Arc>::Type;
+  using FstImpl<Arc>::Properties;
+
+  explicit CacheBaseImpl(const CacheOptions &opts = CacheOptions())
+      : has_start_(false),
+        cache_start_(kNoStateId),
+        nknown_states_(0),
+        min_unexpanded_state_id_(0),
+        max_expanded_state_id_(-1),
+        cache_gc_(opts.gc),
+        cache_limit_(opts.gc_limit),
+        cache_store_(new CacheStore(opts)),
+        new_cache_store_(true),
+        own_cache_store_(true) {}
+
+  explicit CacheBaseImpl(const CacheImplOptions<CacheStore> &opts)
+      : has_start_(false),
+        cache_start_(kNoStateId),
+        nknown_states_(0),
+        min_unexpanded_state_id_(0),
+        max_expanded_state_id_(-1),
+        cache_gc_(opts.gc),
+        cache_limit_(opts.gc_limit),
+        cache_store_(opts.store ? opts.store : new CacheStore(CacheOptions(
+                                                   opts.gc, opts.gc_limit))),
+        new_cache_store_(!opts.store),
+        own_cache_store_(opts.store ? opts.own_store : true) {}
+
+  // Preserve gc parameters. If preserve_cache is true, also preserves
+  // cache data.
+  CacheBaseImpl(const CacheBaseImpl<State, CacheStore> &impl,
+                bool preserve_cache = false)
+      : FstImpl<Arc>(),
+        has_start_(false),
+        cache_start_(kNoStateId),
+        nknown_states_(0),
+        min_unexpanded_state_id_(0),
+        max_expanded_state_id_(-1),
+        cache_gc_(impl.cache_gc_),
+        cache_limit_(impl.cache_limit_),
+        cache_store_(new CacheStore(CacheOptions(cache_gc_, cache_limit_))),
+        new_cache_store_(impl.new_cache_store_ || !preserve_cache),
+        own_cache_store_(true) {
+    if (preserve_cache) {
+      *cache_store_ = *impl.cache_store_;
+      has_start_ = impl.has_start_;
+      cache_start_ = impl.cache_start_;
+      nknown_states_ = impl.nknown_states_;
+      expanded_states_ = impl.expanded_states_;
+      min_unexpanded_state_id_ = impl.min_unexpanded_state_id_;
+      max_expanded_state_id_ = impl.max_expanded_state_id_;
+    }
+  }
+
+  ~CacheBaseImpl() override { if (own_cache_store_) delete cache_store_; }
+
+  void SetStart(StateId s) {
+    cache_start_ = s;
+    has_start_ = true;
+    if (s >= nknown_states_) nknown_states_ = s + 1;
+  }
+
+  void SetFinal(StateId s, Weight weight) {
+    auto *state = cache_store_->GetMutableState(s);
+    state->SetFinal(std::move(weight));
+    static constexpr auto flags = kCacheFinal | kCacheRecent;
+    state->SetFlags(flags, flags);
+  }
+
+// Disabled to ensure PushArc not AddArc is used in existing code
+// TODO(sorenj): re-enable for backing store
+#if 0
+  // AddArc adds a single arc to a state and does incremental cache
+  // book-keeping. For efficiency, prefer PushArc and SetArcs below
+  // when possible.
+  void AddArc(StateId s, const Arc &arc) {
+    auto *state = cache_store_->GetMutableState(s);
+    cache_store_->AddArc(state, arc);
+    if (arc.nextstate >= nknown_states_)
+      nknown_states_ = arc.nextstate + 1;
+    SetExpandedState(s);
+    static constexpr auto flags = kCacheArcs | kCacheRecent;
+    state->SetFlags(flags, flags);
+  }
+#endif
+
+  // Adds a single arc to a state but delays cache book-keeping. SetArcs must
+  // be called when all PushArc and EmplaceArc calls at a state are complete.
+  // Do not mix with calls to AddArc.
+  void PushArc(StateId s, const Arc &arc) {
+    auto *state = cache_store_->GetMutableState(s);
+    state->PushArc(arc);
+  }
+
+  void PushArc(StateId s, Arc &&arc) {
+    auto *state = cache_store_->GetMutableState(s);
+    state->PushArc(std::move(arc));
+  }
+
+  // Adds a single arc to a state but delays cache book-keeping. SetArcs must
+  // be called when all PushArc and EmplaceArc calls at a state are complete.
+  // Do not mix with calls to AddArc.
+  template <class... T>
+  void EmplaceArc(StateId s, T &&... ctor_args) {
+    auto *state = cache_store_->GetMutableState(s);
+    state->EmplaceArc(std::forward<T>(ctor_args)...);
+  }
+
+  // Marks arcs of a state as cached and does cache book-keeping after all
+  // calls to PushArc have been completed. Do not mix with calls to AddArc.
+  void SetArcs(StateId s) {
+    auto *state = cache_store_->GetMutableState(s);
+    cache_store_->SetArcs(state);
+    const auto narcs = state->NumArcs();
+    for (size_t a = 0; a < narcs; ++a) {
+      const auto &arc = state->GetArc(a);
+      if (arc.nextstate >= nknown_states_) nknown_states_ = arc.nextstate + 1;
+    }
+    SetExpandedState(s);
+    static constexpr auto flags = kCacheArcs | kCacheRecent;
+    state->SetFlags(flags, flags);
+  }
+
+  void ReserveArcs(StateId s, size_t n) {
+    auto *state = cache_store_->GetMutableState(s);
+    state->ReserveArcs(n);
+  }
+
+  void DeleteArcs(StateId s) {
+    auto *state = cache_store_->GetMutableState(s);
+    cache_store_->DeleteArcs(state);
+  }
+
+  void DeleteArcs(StateId s, size_t n) {
+    auto *state = cache_store_->GetMutableState(s);
+    cache_store_->DeleteArcs(state, n);
+  }
+
+  void Clear() {
+    nknown_states_ = 0;
+    min_unexpanded_state_id_ = 0;
+    max_expanded_state_id_ = -1;
+    has_start_ = false;
+    cache_start_ = kNoStateId;
+    cache_store_->Clear();
+  }
+
+  // Is the start state cached?
+  bool HasStart() const {
+    if (!has_start_ && Properties(kError)) has_start_ = true;
+    return has_start_;
+  }
+
+  // Is the final weight of the state cached?
+  bool HasFinal(StateId s) const {
+    const auto *state = cache_store_->GetState(s);
+    if (state && state->Flags() & kCacheFinal) {
+      state->SetFlags(kCacheRecent, kCacheRecent);
+      return true;
+    } else {
+      return false;
+    }
+  }
+
+  // Are arcs of the state cached?
+  bool HasArcs(StateId s) const {
+    const auto *state = cache_store_->GetState(s);
+    if (state && state->Flags() & kCacheArcs) {
+      state->SetFlags(kCacheRecent, kCacheRecent);
+      return true;
+    } else {
+      return false;
+    }
+  }
+
+  StateId Start() const { return cache_start_; }
+
+  Weight Final(StateId s) const {
+    const auto *state = cache_store_->GetState(s);
+    return state->Final();
+  }
+
+  size_t NumArcs(StateId s) const {
+    const auto *state = cache_store_->GetState(s);
+    return state->NumArcs();
+  }
+
+  size_t NumInputEpsilons(StateId s) const {
+    const auto *state = cache_store_->GetState(s);
+    return state->NumInputEpsilons();
+  }
+
+  size_t NumOutputEpsilons(StateId s) const {
+    const auto *state = cache_store_->GetState(s);
+    return state->NumOutputEpsilons();
+  }
+
+  // Provides information needed for generic arc iterator.
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const {
+    const auto *state = cache_store_->GetState(s);
+    data->base = nullptr;
+    data->narcs = state->NumArcs();
+    data->arcs = state->Arcs();
+    data->ref_count = state->MutableRefCount();
+    state->IncrRefCount();
+  }
+
+  // Number of known states.
+  StateId NumKnownStates() const { return nknown_states_; }
+
+  // Updates number of known states, taking into account the passed state ID.
+  void UpdateNumKnownStates(StateId s) {
+    if (s >= nknown_states_) nknown_states_ = s + 1;
+  }
+
+  // Finds the mininum never-expanded state ID.
+  StateId MinUnexpandedState() const {
+    while (min_unexpanded_state_id_ <= max_expanded_state_id_ &&
+           ExpandedState(min_unexpanded_state_id_)) {
+      ++min_unexpanded_state_id_;
+    }
+    return min_unexpanded_state_id_;
+  }
+
+  // Returns maximum ever-expanded state ID.
+  StateId MaxExpandedState() const { return max_expanded_state_id_; }
+
+  void SetExpandedState(StateId s) {
+    if (s > max_expanded_state_id_) max_expanded_state_id_ = s;
+    if (s < min_unexpanded_state_id_) return;
+    if (s == min_unexpanded_state_id_) ++min_unexpanded_state_id_;
+    if (cache_gc_ || cache_limit_ == 0) {
+      if (expanded_states_.size() <= static_cast<size_t>(s))
+        expanded_states_.resize(s + 1, false);
+      expanded_states_[s] = true;
+    }
+  }
+
+  bool ExpandedState(StateId s) const {
+    if (cache_gc_ || cache_limit_ == 0) {
+      return expanded_states_[s];
+    } else if (new_cache_store_) {
+      return cache_store_->GetState(s) != nullptr;
+    } else {
+      // If the cache was not created by this class, then the cached state needs
+      // to be inspected to update nknown_states_.
+      return false;
+    }
+  }
+
+  const CacheStore *GetCacheStore() const { return cache_store_; }
+
+  CacheStore *GetCacheStore() { return cache_store_; }
+
+  // Caching on/off switch, limit and size accessors.
+
+  bool GetCacheGc() const { return cache_gc_; }
+
+  size_t GetCacheLimit() const { return cache_limit_; }
+
+ private:
+  mutable bool has_start_;                   // Is the start state cached?
+  StateId cache_start_;                      // ID of start state.
+  StateId nknown_states_;                    // Number of known states.
+  std::vector<bool> expanded_states_;        // States that have been expanded.
+  mutable StateId min_unexpanded_state_id_;  // Minimum never-expanded state ID
+  mutable StateId max_expanded_state_id_;    // Maximum ever-expanded state ID
+  bool cache_gc_;                            // GC enabled.
+  size_t cache_limit_;       // Number of bytes allowed before GC.
+  CacheStore *cache_store_;  // The store of cached states.
+  bool new_cache_store_;     // Was the store was created by class?
+  bool own_cache_store_;     // Is the store owned by class?
+
+  CacheBaseImpl &operator=(const CacheBaseImpl &impl) = delete;
+};
+
+// A CacheBaseImpl with the default cache state type.
+template <class Arc>
+class CacheImpl : public CacheBaseImpl<CacheState<Arc>> {
+ public:
+  using State = CacheState<Arc>;
+
+  CacheImpl() {}
+
+  explicit CacheImpl(const CacheOptions &opts)
+      : CacheBaseImpl<CacheState<Arc>>(opts) {}
+
+  CacheImpl(const CacheImpl<Arc> &impl, bool preserve_cache = false)
+      : CacheBaseImpl<State>(impl, preserve_cache) {}
+
+ private:
+  CacheImpl &operator=(const CacheImpl &impl) = delete;
+};
+
+}  // namespace internal
+
+// Use this to make a state iterator for a CacheBaseImpl-derived FST, which must
+// have Arc and Store types defined. Note this iterator only returns those
+// states reachable from the initial state, so consider implementing a
+// class-specific one.
+//
+// This class may be derived from.
+template <class FST>
+class CacheStateIterator : public StateIteratorBase<typename FST::Arc> {
+ public:
+  using Arc = typename FST::Arc;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Store = typename FST::Store;
+  using State = typename Store::State;
+  using Impl = internal::CacheBaseImpl<State, Store>;
+
+  CacheStateIterator(const FST &fst, Impl *impl)
+      : fst_(fst), impl_(impl), s_(0) {
+    fst_.Start();  // Forces start state.
+  }
+
+  bool Done() const final {
+    if (s_ < impl_->NumKnownStates()) return false;
+    for (StateId u = impl_->MinUnexpandedState(); u < impl_->NumKnownStates();
+         u = impl_->MinUnexpandedState()) {
+      // Forces state expansion.
+      ArcIterator<FST> aiter(fst_, u);
+      aiter.SetFlags(kArcValueFlags, kArcValueFlags | kArcNoCache);
+      for (; !aiter.Done(); aiter.Next()) {
+        impl_->UpdateNumKnownStates(aiter.Value().nextstate);
+      }
+      impl_->SetExpandedState(u);
+      if (s_ < impl_->NumKnownStates()) return false;
+    }
+    return true;
+  }
+
+  StateId Value() const final { return s_; }
+
+  void Next() final { ++s_; }
+
+  void Reset() final { s_ = 0; }
+
+ private:
+  const FST &fst_;
+  Impl *impl_;
+  StateId s_;
+};
+
+// Used to make an arc iterator for a CacheBaseImpl-derived FST, which must
+// have Arc and State types defined.
+template <class FST>
+class CacheArcIterator {
+ public:
+  using Arc = typename FST::Arc;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Store = typename FST::Store;
+  using State = typename Store::State;
+  using Impl = internal::CacheBaseImpl<State, Store>;
+
+  CacheArcIterator(Impl *impl, StateId s) : i_(0) {
+    state_ = impl->GetCacheStore()->GetMutableState(s);
+    state_->IncrRefCount();
+  }
+
+  ~CacheArcIterator() { state_->DecrRefCount(); }
+
+  bool Done() const { return i_ >= state_->NumArcs(); }
+
+  const Arc &Value() const { return state_->GetArc(i_); }
+
+  void Next() { ++i_; }
+
+  size_t Position() const { return i_; }
+
+  void Reset() { i_ = 0; }
+
+  void Seek(size_t a) { i_ = a; }
+
+  constexpr uint32 Flags() const { return kArcValueFlags; }
+
+  void SetFlags(uint32 flags, uint32 mask) {}
+
+ private:
+  const State *state_;
+  size_t i_;
+
+  CacheArcIterator(const CacheArcIterator &) = delete;
+  CacheArcIterator &operator=(const CacheArcIterator &) = delete;
+};
+
+// Use this to make a mutable arc iterator for a CacheBaseImpl-derived FST,
+// which must have types Arc and Store defined.
+template <class FST>
+class CacheMutableArcIterator
+    : public MutableArcIteratorBase<typename FST::Arc> {
+ public:
+  using Arc = typename FST::Arc;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Store = typename FST::Store;
+  using State = typename Store::State;
+  using Impl = internal::CacheBaseImpl<State, Store>;
+
+  // User must call MutateCheck() in the constructor.
+  CacheMutableArcIterator(Impl *impl, StateId s) : i_(0), s_(s), impl_(impl) {
+    state_ = impl_->GetCacheStore()->GetMutableState(s_);
+    state_->IncrRefCount();
+  }
+
+  ~CacheMutableArcIterator() override { state_->DecrRefCount(); }
+
+  bool Done() const final { return i_ >= state_->NumArcs(); }
+
+  const Arc &Value() const final { return state_->GetArc(i_); }
+
+  void Next() final { ++i_; }
+
+  size_t Position() const final { return i_; }
+
+  void Reset() final { i_ = 0; }
+
+  void Seek(size_t a) final { i_ = a; }
+
+  void SetValue(const Arc &arc) final { state_->SetArc(arc, i_); }
+
+  uint32 Flags() const final { return kArcValueFlags; }
+
+  void SetFlags(uint32, uint32) final {}
+
+ private:
+  size_t i_;
+  StateId s_;
+  Impl *impl_;
+  State *state_;
+
+  CacheMutableArcIterator(const CacheMutableArcIterator &) = delete;
+  CacheMutableArcIterator &operator=(const CacheMutableArcIterator &) = delete;
+};
+
+// Wrap existing CacheStore implementation to use with ExpanderFst.
+template <class CacheStore>
+class ExpanderCacheStore {
+ public:
+  using State = typename CacheStore::State;
+  using Arc = typename CacheStore::Arc;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  explicit ExpanderCacheStore(const CacheOptions &opts = CacheOptions())
+      : store_(opts) {}
+
+  template <class Expander>
+  State *FindOrExpand(Expander &expander, StateId s) {  // NOLINT
+    auto *state = store_.GetMutableState(s);
+    if (state->Flags()) {
+      state->SetFlags(kCacheRecent, kCacheRecent);
+    } else {
+      StateBuilder builder(state);
+      expander.Expand(s, &builder);
+      state->SetFlags(kCacheFlags, kCacheFlags);
+      store_.SetArcs(state);
+    }
+    return state;
+  }
+
+ private:
+  CacheStore store_;
+
+  struct StateBuilder {
+    State *state;
+
+    explicit StateBuilder(State *state_) : state(state_) {}
+
+    void AddArc(const Arc &arc) { state->PushArc(arc); }
+
+    void AddArc(Arc &&arc) { state->PushArc(std::move(arc)); }
+
+    void SetFinal(Weight weight) { state->SetFinal(std::move(weight)); }
+  };
+};
+
+}  // namespace fst
+
+#endif  // FST_CACHE_H_
diff --git a/projects/llm_framework/include/fst/closure.h b/projects/llm_framework/include/fst/closure.h
new file mode 100644
index 00000000..13beea9c
--- /dev/null
+++ b/projects/llm_framework/include/fst/closure.h
@@ -0,0 +1,134 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes to compute the concatenative closure of an FST.
+
+#ifndef FST_CLOSURE_H_
+#define FST_CLOSURE_H_
+
+#include <algorithm>
+#include <vector>
+
+#include <fst/mutable-fst.h>
+#include <fst/rational.h>
+
+
+namespace fst {
+
+// Computes the concatenative closure. This version modifies its
+// MutableFst input. If an FST transduces string x to y with weight a,
+// then its closure transduces x to y with weight a, xx to yy with
+// weight Times(a, a), xxx to yyy with with Times(Times(a, a), a),
+// etc. If closure_type == CLOSURE_STAR, then the empty string is
+// transduced to itself with weight Weight::One() as well.
+//
+// Complexity:
+//
+//   Time: O(V)
+//   Space: O(V)
+//
+// where V is the number of states.
+template <class Arc>
+void Closure(MutableFst<Arc> *fst, ClosureType closure_type) {
+  using Weight = typename Arc::Weight;
+  const auto props = fst->Properties(kFstProperties, false);
+  const auto start = fst->Start();
+  for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
+       siter.Next()) {
+    const auto s = siter.Value();
+    const auto weight = fst->Final(s);
+    if (weight != Weight::Zero()) fst->AddArc(s, Arc(0, 0, weight, start));
+  }
+  if (closure_type == CLOSURE_STAR) {
+    fst->ReserveStates(fst->NumStates() + 1);
+    const auto nstart = fst->AddState();
+    fst->SetStart(nstart);
+    fst->SetFinal(nstart, Weight::One());
+    if (start != kNoLabel) fst->AddArc(nstart, Arc(0, 0, Weight::One(), start));
+  }
+  fst->SetProperties(ClosureProperties(props, closure_type == CLOSURE_STAR),
+                     kFstProperties);
+}
+
+// Computes the concatenative closure. This version modifies its
+// RationalFst input.
+template <class Arc>
+void Closure(RationalFst<Arc> *fst, ClosureType closure_type) {
+  fst->GetMutableImpl()->AddClosure(closure_type);
+}
+
+struct ClosureFstOptions : RationalFstOptions {
+  ClosureType type;
+
+  ClosureFstOptions(const RationalFstOptions &opts,
+                    ClosureType type = CLOSURE_STAR)
+      : RationalFstOptions(opts), type(type) {}
+
+  explicit ClosureFstOptions(ClosureType type = CLOSURE_STAR) : type(type) {}
+};
+
+// Computes the concatenative closure. This version is a delayed FST. If an FST
+// transduces string x to y with weight a, then its closure transduces x to y
+// with weight a, xx to yy with weight Times(a, a), xxx to yyy with weight
+// Times(Times(a, a), a), etc. If closure_type == CLOSURE_STAR, then the empty
+// string is transduced to itself with weight Weight::One() as well.
+//
+// Complexity:
+//
+//   Time: O(v)
+//   Space: O(v)
+//
+// where v is the number of states visited. Constant time and space to visit an
+// input state or arc is assumed and exclusive of caching.
+template <class A>
+class ClosureFst : public RationalFst<A> {
+ public:
+  using Arc = A;
+
+  ClosureFst(const Fst<Arc> &fst, ClosureType closure_type) {
+    GetMutableImpl()->InitClosure(fst, closure_type);
+  }
+
+  ClosureFst(const Fst<Arc> &fst, const ClosureFstOptions &opts)
+      : RationalFst<A>(opts) {
+    GetMutableImpl()->InitClosure(fst, opts.type);
+  }
+
+  // See Fst<>::Copy() for doc.
+  ClosureFst(const ClosureFst<Arc> &fst, bool safe = false)
+      : RationalFst<A>(fst, safe) {}
+
+  // Gets a copy of this ClosureFst. See Fst<>::Copy() for further doc.
+  ClosureFst<A> *Copy(bool safe = false) const override {
+    return new ClosureFst<A>(*this, safe);
+  }
+
+ private:
+  using ImplToFst<internal::RationalFstImpl<Arc>>::GetImpl;
+  using ImplToFst<internal::RationalFstImpl<Arc>>::GetMutableImpl;
+};
+
+// Specialization for ClosureFst.
+template <class Arc>
+class StateIterator<ClosureFst<Arc>> : public StateIterator<RationalFst<Arc>> {
+ public:
+  explicit StateIterator(const ClosureFst<Arc> &fst)
+      : StateIterator<RationalFst<Arc>>(fst) {}
+};
+
+// Specialization for ClosureFst.
+template <class Arc>
+class ArcIterator<ClosureFst<Arc>> : public ArcIterator<RationalFst<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const ClosureFst<Arc> &fst, StateId s)
+      : ArcIterator<RationalFst<Arc>>(fst, s) {}
+};
+
+// Useful alias when using StdArc.
+using StdClosureFst = ClosureFst<StdArc>;
+
+}  // namespace fst
+
+#endif  // FST_CLOSURE_H_
diff --git a/projects/llm_framework/include/fst/compact-fst.h b/projects/llm_framework/include/fst/compact-fst.h
new file mode 100644
index 00000000..402c87b7
--- /dev/null
+++ b/projects/llm_framework/include/fst/compact-fst.h
@@ -0,0 +1,1564 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// FST Class for memory-efficient representation of common types of
+// FSTs: linear automata, acceptors, unweighted FSTs, ...
+
+#ifndef FST_COMPACT_FST_H_
+#define FST_COMPACT_FST_H_
+
+#include <climits>
+#include <iterator>
+#include <memory>
+#include <tuple>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/cache.h>
+#include <fst/expanded-fst.h>
+#include <fst/fst-decl.h>  // For optional argument declarations
+#include <fst/mapped-file.h>
+#include <fst/matcher.h>
+#include <fst/test-properties.h>
+#include <fst/util.h>
+
+
+namespace fst {
+
+struct CompactFstOptions : public CacheOptions {
+  // The default caching behaviour is to do no caching. Most compactors are
+  // cheap and therefore we save memory by not doing caching.
+  CompactFstOptions() : CacheOptions(true, 0) {}
+
+  explicit CompactFstOptions(const CacheOptions &opts) : CacheOptions(opts) {}
+};
+
+// New upcoming (Fst) Compactor interface - currently used internally
+// by CompactFstImpl.
+//
+// class Compactor {
+//  public:
+//   // Constructor from the Fst to be compacted.
+//   Compactor(const Fst<Arc> &fst, ...);
+//   // Copy constructor
+//   Compactor(const Compactor &compactor, bool safe = false)
+//   // Default constructor (optional, see comment below).
+//   Compactor();
+//
+//   // Returns the start state, number of states, and total number of arcs
+//   // of the compacted Fst
+//   StateId Start() const;
+//   StateId NumStates() const;
+//   size_t NumArcs() const;
+//
+//   // Accessor class for state attributes.
+//   class State {
+//    public:
+//     State();  // Required, corresponds to kNoStateId.
+//     State(const Compactor *c, StateId);  // Accessor for StateId 's'.
+//     StateId GetStateId() const;
+//     Weight Final() const;
+//     size_t NumArcs() const;
+//     Arc GetArc(size_t i, uint32 f) const;
+//   };
+//
+//   // Modifies 'state' accessor to provide access to state id 's'.
+//   void SetState(StateId s, State *state);
+//   // Tests whether 'fst' can be compacted by this compactor.
+//   bool IsCompatible(const Fst<A> &fst) const;
+//   // Return the properties that are always true for an fst
+//   // compacted using this compactor
+//   uint64 Properties() const;
+//   // Return a string identifying the type of compactor.
+//   static const string &Type();
+//   // Return true if an error has occured.
+//   bool Error() const;
+//   // Writes a compactor to a file.
+//   bool Write(std::ostream &strm, const FstWriteOptions &opts) const;
+//   // Reads a compactor from a file.
+//   static Compactor*Read(std::istream &strm, const FstReadOptions &opts,
+//                         const FstHeader &hdr);
+// };
+//
+
+// Old (Arc) Compactor Interface:
+//
+// The ArcCompactor class determines how arcs and final weights are compacted
+// and expanded.
+//
+// Final weights are treated as transitions to the superfinal state, i.e.,
+// ilabel = olabel = kNoLabel and nextstate = kNoStateId.
+//
+// There are two types of compactors:
+//
+// * Fixed out-degree compactors: 'compactor.Size()' returns a positive integer
+//   's'. An FST can be compacted by this compactor only if each state has
+//   exactly 's' outgoing transitions (counting a non-Zero() final weight as a
+//   transition). A typical example is a compactor for string FSTs, i.e.,
+//   's == 1'.
+//
+// * Variable out-degree compactors: 'compactor.Size() == -1'. There are no
+//   out-degree restrictions for these compactors.
+//
+// Interface:
+//
+// class ArcCompactor {
+//  public:
+//   // Element is the type of the compacted transitions.
+//   using Element = ...
+//
+//   // Returns the compacted representation of a transition 'arc'
+//   // at a state 's'.
+//   Element Compact(StateId s, const Arc &arc);
+//
+//   // Returns the transition at state 's' represented by the compacted
+//   // transition 'e'.
+//   Arc Expand(StateId s, const Element &e) const;
+//
+//   // Returns -1 for variable out-degree compactors, and the mandatory
+//   // out-degree otherwise.
+//   ssize_t Size() const;
+//
+//   // Tests whether an FST can be compacted by this compactor.
+//   bool Compatible(const Fst<A> &fst) const;
+//
+//   // Returns the properties that are always true for an FST compacted using
+//   // this compactor
+//   uint64 Properties() const;
+//
+//   // Returns a string identifying the type of compactor.
+//   static const string &Type();
+//
+//   // Writes a compactor to a file.
+//   bool Write(std::ostream &strm) const;
+//
+//   // Reads a compactor from a file.
+//   static ArcCompactor *Read(std::istream &strm);
+//
+//   // Default constructor (optional, see comment below).
+//   ArcCompactor();
+// };
+//
+// The default constructor is only required for FST_REGISTER to work (i.e.,
+// enabling Convert() and the command-line utilities to work with this new
+// compactor). However, a default constructor always needs to be specified for
+// this code to compile, but one can have it simply raise an error when called,
+// like so:
+//
+// Compactor::Compactor() {
+//   FSTERROR() << "Compactor: No default constructor";
+// }
+
+// Default implementation data for CompactFst, which can shared between
+// otherwise independent copies.
+//
+// The implementation contains two arrays: 'states_' and 'compacts_'.
+//
+// For fixed out-degree compactors, the 'states_' array is unallocated. The
+// 'compacts_' contains the compacted transitions. Its size is 'ncompacts_'.
+// The outgoing transitions at a given state are stored consecutively. For a
+// given state 's', its 'compactor.Size()' outgoing transitions (including
+// superfinal transition when 's' is final), are stored in position
+// ['s*compactor.Size()', '(s+1)*compactor.Size()').
+//
+// For variable out-degree compactors, the states_ array has size
+// 'nstates_ + 1' and contains pointers to positions into 'compacts_'. For a
+// given state 's', the compacted transitions of 's' are stored in positions
+// ['states_[s]', 'states_[s + 1]') in 'compacts_'. By convention,
+// 'states_[nstates_] == ncompacts_'.
+//
+// In both cases, the superfinal transitions (when 's' is final, i.e.,
+// 'Final(s) != Weight::Zero()') are stored first.
+//
+// The unsigned type U is used to represent indices into the compacts_ array.
+template <class Element, class Unsigned>
+class DefaultCompactStore {
+ public:
+  DefaultCompactStore()
+      : states_(nullptr),
+        compacts_(nullptr),
+        nstates_(0),
+        ncompacts_(0),
+        narcs_(0),
+        start_(kNoStateId),
+        error_(false) {}
+
+  template <class Arc, class Compactor>
+  DefaultCompactStore(const Fst<Arc> &fst, const Compactor &compactor);
+
+  template <class Iterator, class Compactor>
+  DefaultCompactStore(const Iterator &begin, const Iterator &end,
+                      const Compactor &compactor);
+
+  ~DefaultCompactStore() {
+    if (!states_region_) delete[] states_;
+    if (!compacts_region_) delete[] compacts_;
+  }
+
+  template <class Compactor>
+  static DefaultCompactStore<Element, Unsigned> *Read(
+      std::istream &strm, const FstReadOptions &opts, const FstHeader &hdr,
+      const Compactor &compactor);
+
+  bool Write(std::ostream &strm, const FstWriteOptions &opts) const;
+
+  Unsigned States(ssize_t i) const { return states_[i]; }
+
+  const Element &Compacts(size_t i) const { return compacts_[i]; }
+
+  size_t NumStates() const { return nstates_; }
+
+  size_t NumCompacts() const { return ncompacts_; }
+
+  size_t NumArcs() const { return narcs_; }
+
+  ssize_t Start() const { return start_; }
+
+  bool Error() const { return error_; }
+
+  // Returns a string identifying the type of data storage container.
+  static const string &Type();
+
+ private:
+  std::unique_ptr<MappedFile> states_region_;
+  std::unique_ptr<MappedFile> compacts_region_;
+  Unsigned *states_;
+  Element *compacts_;
+  size_t nstates_;
+  size_t ncompacts_;
+  size_t narcs_;
+  ssize_t start_;
+  bool error_;
+};
+
+template <class Element, class Unsigned>
+template <class Arc, class Compactor>
+DefaultCompactStore<Element, Unsigned>::DefaultCompactStore(
+    const Fst<Arc> &fst, const Compactor &compactor)
+    : states_(nullptr),
+      compacts_(nullptr),
+      nstates_(0),
+      ncompacts_(0),
+      narcs_(0),
+      start_(kNoStateId),
+      error_(false) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  start_ = fst.Start();
+  // Counts # of states and arcs.
+  StateId nfinals = 0;
+  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
+    ++nstates_;
+    const auto s = siter.Value();
+    narcs_ += fst.NumArcs(s);
+    if (fst.Final(s) != Weight::Zero()) ++nfinals;
+  }
+  if (compactor.Size() == -1) {
+    states_ = new Unsigned[nstates_ + 1];
+    ncompacts_ = narcs_ + nfinals;
+    compacts_ = new Element[ncompacts_];
+    states_[nstates_] = ncompacts_;
+  } else {
+    states_ = nullptr;
+    ncompacts_ = nstates_ * compactor.Size();
+    if ((narcs_ + nfinals) != ncompacts_) {
+      FSTERROR() << "DefaultCompactStore: Compactor incompatible with FST";
+      error_ = true;
+      return;
+    }
+    compacts_ = new Element[ncompacts_];
+  }
+  size_t pos = 0;
+  size_t fpos = 0;
+  for (size_t s = 0; s < nstates_; ++s) {
+    fpos = pos;
+    if (compactor.Size() == -1) states_[s] = pos;
+    if (fst.Final(s) != Weight::Zero()) {
+      compacts_[pos++] = compactor.Compact(
+          s, Arc(kNoLabel, kNoLabel, fst.Final(s), kNoStateId));
+    }
+    for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
+      compacts_[pos++] = compactor.Compact(s, aiter.Value());
+    }
+    if ((compactor.Size() != -1) && (pos != fpos + compactor.Size())) {
+      FSTERROR() << "DefaultCompactStore: Compactor incompatible with FST";
+      error_ = true;
+      return;
+    }
+  }
+  if (pos != ncompacts_) {
+    FSTERROR() << "DefaultCompactStore: Compactor incompatible with FST";
+    error_ = true;
+    return;
+  }
+}
+
+template <class Element, class Unsigned>
+template <class Iterator, class Compactor>
+DefaultCompactStore<Element, Unsigned>::DefaultCompactStore(
+    const Iterator &begin, const Iterator &end, const Compactor &compactor)
+    : states_(nullptr),
+      compacts_(nullptr),
+      nstates_(0),
+      ncompacts_(0),
+      narcs_(0),
+      start_(kNoStateId),
+      error_(false) {
+  using Arc = typename Compactor::Arc;
+  using Weight = typename Arc::Weight;
+  if (compactor.Size() != -1) {
+    ncompacts_ = std::distance(begin, end);
+    if (compactor.Size() == 1) {
+      // For strings, allows implicit final weight. Empty input is the empty
+      // string.
+      if (ncompacts_ == 0) {
+        ++ncompacts_;
+      } else {
+        const auto arc =
+            compactor.Expand(ncompacts_ - 1, *(begin + (ncompacts_ - 1)));
+        if (arc.ilabel != kNoLabel) ++ncompacts_;
+      }
+    }
+    if (ncompacts_ % compactor.Size()) {
+      FSTERROR() << "DefaultCompactStore: Size of input container incompatible"
+                 << " with compactor";
+      error_ = true;
+      return;
+    }
+    if (ncompacts_ == 0) return;
+    start_ = 0;
+    nstates_ = ncompacts_ / compactor.Size();
+    compacts_ = new Element[ncompacts_];
+    size_t i = 0;
+    Iterator it = begin;
+    for (; it != end; ++it, ++i) {
+      compacts_[i] = *it;
+      if (compactor.Expand(i, *it).ilabel != kNoLabel) ++narcs_;
+    }
+    if (i < ncompacts_) {
+      compacts_[i] = compactor.Compact(
+          i, Arc(kNoLabel, kNoLabel, Weight::One(), kNoStateId));
+    }
+  } else {
+    if (std::distance(begin, end) == 0) return;
+    // Count # of states, arcs and compacts.
+    auto it = begin;
+    for (size_t i = 0; it != end; ++it, ++i) {
+      const auto arc = compactor.Expand(i, *it);
+      if (arc.ilabel != kNoLabel) {
+        ++narcs_;
+        ++ncompacts_;
+      } else {
+        ++nstates_;
+        if (arc.weight != Weight::Zero()) ++ncompacts_;
+      }
+    }
+    start_ = 0;
+    compacts_ = new Element[ncompacts_];
+    states_ = new Unsigned[nstates_ + 1];
+    states_[nstates_] = ncompacts_;
+    size_t i = 0;
+    size_t s = 0;
+    for (it = begin; it != end; ++it) {
+      const auto arc = compactor.Expand(i, *it);
+      if (arc.ilabel != kNoLabel) {
+        compacts_[i++] = *it;
+      } else {
+        states_[s++] = i;
+        if (arc.weight != Weight::Zero()) compacts_[i++] = *it;
+      }
+    }
+    if ((s != nstates_) || (i != ncompacts_)) {
+      FSTERROR() << "DefaultCompactStore: Ill-formed input container";
+      error_ = true;
+      return;
+    }
+  }
+}
+
+template <class Element, class Unsigned>
+template <class Compactor>
+DefaultCompactStore<Element, Unsigned>
+    *DefaultCompactStore<Element, Unsigned>::Read(std::istream &strm,
+                                                  const FstReadOptions &opts,
+                                                  const FstHeader &hdr,
+                                                  const Compactor &compactor) {
+  std::unique_ptr<DefaultCompactStore<Element, Unsigned>> data(
+      new DefaultCompactStore<Element, Unsigned>());
+  data->start_ = hdr.Start();
+  data->nstates_ = hdr.NumStates();
+  data->narcs_ = hdr.NumArcs();
+  if (compactor.Size() == -1) {
+    if ((hdr.GetFlags() & FstHeader::IS_ALIGNED) && !AlignInput(strm)) {
+      LOG(ERROR) << "DefaultCompactStore::Read: Alignment failed: "
+                 << opts.source;
+      return nullptr;
+    }
+    auto b = (data->nstates_ + 1) * sizeof(Unsigned);
+    data->states_region_.reset(MappedFile::Map(
+        &strm, opts.mode == FstReadOptions::MAP, opts.source, b));
+    if (!strm || !data->states_region_) {
+      LOG(ERROR) << "DefaultCompactStore::Read: Read failed: " << opts.source;
+      return nullptr;
+    }
+    data->states_ =
+        static_cast<Unsigned *>(data->states_region_->mutable_data());
+  } else {
+    data->states_ = nullptr;
+  }
+  data->ncompacts_ = compactor.Size() == -1 ? data->states_[data->nstates_]
+                                            : data->nstates_ * compactor.Size();
+  if ((hdr.GetFlags() & FstHeader::IS_ALIGNED) && !AlignInput(strm)) {
+    LOG(ERROR) << "DefaultCompactStore::Read: Alignment failed: "
+               << opts.source;
+    return nullptr;
+  }
+  size_t b = data->ncompacts_ * sizeof(Element);
+  data->compacts_region_.reset(
+      MappedFile::Map(&strm, opts.mode == FstReadOptions::MAP, opts.source, b));
+  if (!strm || !data->compacts_region_) {
+    LOG(ERROR) << "DefaultCompactStore::Read: Read failed: " << opts.source;
+    return nullptr;
+  }
+  data->compacts_ =
+      static_cast<Element *>(data->compacts_region_->mutable_data());
+  return data.release();
+}
+
+template <class Element, class Unsigned>
+bool DefaultCompactStore<Element, Unsigned>::Write(
+    std::ostream &strm, const FstWriteOptions &opts) const {
+  if (states_) {
+    if (opts.align && !AlignOutput(strm)) {
+      LOG(ERROR) << "DefaultCompactStore::Write: Alignment failed: "
+                 << opts.source;
+      return false;
+    }
+    strm.write(reinterpret_cast<char *>(states_),
+               (nstates_ + 1) * sizeof(Unsigned));
+  }
+  if (opts.align && !AlignOutput(strm)) {
+    LOG(ERROR) << "DefaultCompactStore::Write: Alignment failed: "
+               << opts.source;
+    return false;
+  }
+  strm.write(reinterpret_cast<char *>(compacts_), ncompacts_ * sizeof(Element));
+  strm.flush();
+  if (!strm) {
+    LOG(ERROR) << "DefaultCompactStore::Write: Write failed: " << opts.source;
+    return false;
+  }
+  return true;
+}
+
+template <class Element, class Unsigned>
+const string &DefaultCompactStore<Element, Unsigned>::Type() {
+  static const string *const type = new string("compact");
+  return *type;
+}
+
+template <class C, class U, class S> class DefaultCompactState;
+
+// Wraps an arc compactor and a compact store as a new Fst compactor.
+template <class C, class U,
+          class S = DefaultCompactStore<typename C::Element, U>>
+class DefaultCompactor {
+ public:
+  using ArcCompactor = C;
+  using Unsigned = U;
+  using CompactStore = S;
+  using Element = typename C::Element;
+  using Arc = typename C::Arc;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using State = DefaultCompactState<C, U, S>;
+  friend State;
+
+  DefaultCompactor()
+      : arc_compactor_(nullptr), compact_store_(nullptr) {}
+
+  // Constructs from Fst.
+  DefaultCompactor(const Fst<Arc> &fst,
+                   std::shared_ptr<ArcCompactor> arc_compactor)
+      : arc_compactor_(std::move(arc_compactor)),
+        compact_store_(std::make_shared<S>(fst, *arc_compactor_)) {}
+
+  DefaultCompactor(const Fst<Arc> &fst,
+                   std::shared_ptr<DefaultCompactor<C, U, S>> compactor)
+      : arc_compactor_(compactor->arc_compactor_),
+        compact_store_(compactor->compact_store_ == nullptr ?
+                       std::make_shared<S>(fst, *arc_compactor_) :
+                       compactor->compact_store_) {}
+
+  // Constructs from CompactStore.
+  DefaultCompactor(std::shared_ptr<ArcCompactor> arc_compactor,
+                   std::shared_ptr<CompactStore> compact_store)
+      : arc_compactor_(std::move(arc_compactor)),
+        compact_store_(std::move(compact_store)) {}
+
+  // Constructs from set of compact elements (when arc_compactor.Size() != -1).
+  template <class Iterator>
+  DefaultCompactor(const Iterator &b, const Iterator &e,
+                   std::shared_ptr<C> arc_compactor)
+      : arc_compactor_(std::move(arc_compactor)),
+        compact_store_(std::make_shared<S>(b, e, *arc_compactor_)) {}
+
+  // Copy constructor.
+  DefaultCompactor(const DefaultCompactor<C, U, S> &compactor)
+      : arc_compactor_(std::make_shared<C>(*compactor.GetArcCompactor())),
+        compact_store_(compactor.SharedCompactStore()) {}
+
+  template <class OtherC>
+  explicit DefaultCompactor(const DefaultCompactor<OtherC, U, S> &compactor)
+      : arc_compactor_(std::make_shared<C>(*compactor.GetArcCompactor())),
+        compact_store_(compactor.SharedCompactStore()) {}
+
+  StateId Start() const { return compact_store_->Start(); }
+  StateId NumStates() const { return compact_store_->NumStates(); }
+  size_t NumArcs() const { return compact_store_->NumArcs(); }
+
+  void SetState(StateId s, State *state) const {
+    if (state->GetStateId() != s) state->Set(this, s);
+  }
+
+  static DefaultCompactor<C, U, S> *Read(std::istream &strm,
+                                         const FstReadOptions &opts,
+                                         const FstHeader &hdr) {
+    std::shared_ptr<C> arc_compactor(C::Read(strm));
+    if (arc_compactor == nullptr) return nullptr;
+    std::shared_ptr<S> compact_store(S::Read(strm, opts, hdr, *arc_compactor));
+    if (compact_store == nullptr) return nullptr;
+    return new DefaultCompactor<C, U, S>(arc_compactor, compact_store);
+  }
+
+  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
+    return arc_compactor_->Write(strm) && compact_store_->Write(strm, opts);
+  }
+
+  uint64 Properties() const { return arc_compactor_->Properties(); }
+
+  bool IsCompatible(const Fst<Arc> &fst) const {
+    return arc_compactor_->Compatible(fst);
+  }
+
+  bool Error() const { return compact_store_->Error(); }
+
+  bool HasFixedOutdegree() const { return arc_compactor_->Size() != -1; }
+
+  static const string &Type() {
+    static const string *const type = [] {
+      string type = "compact";
+      if (sizeof(U) != sizeof(uint32)) type += std::to_string(8 * sizeof(U));
+      type += "_";
+      type += C::Type();
+      if (CompactStore::Type() != "compact") {
+        type += "_";
+        type += CompactStore::Type();
+      }
+      return new string(type);
+    }();
+    return *type;
+  }
+
+  const ArcCompactor *GetArcCompactor() const { return arc_compactor_.get(); }
+  CompactStore *GetCompactStore() const { return compact_store_.get(); }
+
+  std::shared_ptr<ArcCompactor> SharedArcCompactor() const {
+    return arc_compactor_;
+  }
+
+  std::shared_ptr<CompactStore> SharedCompactStore() const {
+    return compact_store_;
+  }
+
+  // TODO(allauzen): remove dependencies on this method and make private.
+  Arc ComputeArc(StateId s, Unsigned i, uint32 f) const {
+    return arc_compactor_->Expand(s, compact_store_->Compacts(i), f);
+  }
+
+ private:
+  std::pair<Unsigned, Unsigned> CompactsRange(StateId s) const {
+    std::pair<size_t, size_t> range;
+    if (HasFixedOutdegree()) {
+      range.first = s * arc_compactor_->Size();
+      range.second = arc_compactor_->Size();
+    } else {
+      range.first = compact_store_->States(s);
+      range.second = compact_store_->States(s + 1) - range.first;
+    }
+    return range;
+  }
+
+ private:
+  std::shared_ptr<ArcCompactor> arc_compactor_;
+  std::shared_ptr<CompactStore> compact_store_;
+};
+
+// Default implementation of state attributes accessor class for
+// DefaultCompactor. Use of efficient specialization strongly encouraged.
+template <class C, class U, class S>
+class DefaultCompactState {
+ public:
+  using Arc = typename C::Arc;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  DefaultCompactState() = default;
+
+  DefaultCompactState(const DefaultCompactor<C, U, S> *compactor, StateId s)
+      : compactor_(compactor),
+        s_(s),
+        range_(compactor->CompactsRange(s)),
+        has_final_(
+            range_.second != 0 &&
+            compactor->ComputeArc(s, range_.first,
+                                 kArcILabelValue).ilabel == kNoLabel) {
+    if (has_final_) {
+      ++range_.first;
+      --range_.second;
+    }
+  }
+
+  void Set(const DefaultCompactor<C, U, S> *compactor, StateId s) {
+    compactor_ = compactor;
+    s_ = s;
+    range_ = compactor->CompactsRange(s);
+    if (range_.second != 0 &&
+        compactor->ComputeArc(s, range_.first, kArcILabelValue).ilabel
+        == kNoLabel) {
+      has_final_ = true;
+      ++range_.first;
+      --range_.second;
+    } else {
+      has_final_ = false;
+    }
+  }
+
+  StateId GetStateId() const { return s_; }
+
+  Weight Final() const {
+    if (!has_final_) return Weight::Zero();
+    return compactor_->ComputeArc(s_, range_.first - 1, kArcWeightValue).weight;
+  }
+
+  size_t NumArcs() const { return range_.second; }
+
+  Arc GetArc(size_t i, uint32 f) const {
+    return compactor_->ComputeArc(s_, range_.first + i, f);
+  }
+
+ private:
+  const DefaultCompactor<C, U, S> *compactor_ = nullptr;  // borrowed ref.
+  StateId s_ = kNoStateId;
+  std::pair<U, U> range_ = {0, 0};
+  bool has_final_ = false;
+};
+
+// Specialization for DefaultCompactStore.
+template <class C, class U>
+class DefaultCompactState<C, U, DefaultCompactStore<typename C::Element, U>> {
+ public:
+  using Arc = typename C::Arc;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using CompactStore = DefaultCompactStore<typename C::Element, U>;
+
+  DefaultCompactState() = default;
+
+  DefaultCompactState(
+      const DefaultCompactor<C, U, CompactStore> *compactor, StateId s)
+      : arc_compactor_(compactor->GetArcCompactor()), s_(s) {
+    Init(compactor);
+  }
+
+  void Set(const DefaultCompactor<C, U, CompactStore> *compactor, StateId s) {
+    arc_compactor_ = compactor->GetArcCompactor();
+    s_ = s;
+    has_final_ = false;
+    Init(compactor);
+  }
+
+  StateId GetStateId() const { return s_; }
+
+  Weight Final() const {
+    if (!has_final_) return Weight::Zero();
+    return arc_compactor_->Expand(s_, *(compacts_ - 1), kArcWeightValue).weight;
+  }
+
+  size_t NumArcs() const { return num_arcs_; }
+
+  Arc GetArc(size_t i, uint32 f) const {
+    return arc_compactor_->Expand(s_, compacts_[i], f);
+  }
+
+ private:
+  void Init(const DefaultCompactor<C, U, CompactStore> *compactor) {
+    const auto *store = compactor->GetCompactStore();
+    U offset;
+    if (!compactor->HasFixedOutdegree()) {  // Variable out-degree compactor.
+      offset = store->States(s_);
+      num_arcs_ = store->States(s_ + 1) - offset;
+    } else {  // Fixed out-degree compactor.
+      offset = s_ * arc_compactor_->Size();
+      num_arcs_ = arc_compactor_->Size();
+    }
+    if (num_arcs_ > 0) {
+      compacts_ = &(store->Compacts(offset));
+      if (arc_compactor_->Expand(s_, *compacts_, kArcILabelValue).ilabel
+          == kNoStateId) {
+        ++compacts_;
+        --num_arcs_;
+        has_final_ = true;
+      }
+    }
+  }
+
+ private:
+  const C *arc_compactor_ = nullptr;               // Borrowed reference.
+  const typename C::Element *compacts_ = nullptr;  // Borrowed reference.
+  StateId s_ = kNoStateId;
+  U num_arcs_ = 0;
+  bool has_final_ = false;
+};
+
+template <class Arc, class ArcCompactor, class Unsigned, class CompactStore,
+          class CacheStore>
+class CompactFst;
+
+template <class F, class G>
+void Cast(const F &, G *);
+
+namespace internal {
+
+// Implementation class for CompactFst, which contains parametrizeable
+// Fst data storage (DefaultCompactStore by default) and Fst cache.
+template <class Arc, class C, class CacheStore = DefaultCacheStore<Arc>>
+class CompactFstImpl
+    : public CacheBaseImpl<typename CacheStore::State, CacheStore> {
+ public:
+  using Weight = typename Arc::Weight;
+  using StateId = typename Arc::StateId;
+  using Compactor = C;
+
+  using FstImpl<Arc>::SetType;
+  using FstImpl<Arc>::SetProperties;
+  using FstImpl<Arc>::Properties;
+  using FstImpl<Arc>::SetInputSymbols;
+  using FstImpl<Arc>::SetOutputSymbols;
+  using FstImpl<Arc>::WriteHeader;
+
+  using ImplBase = CacheBaseImpl<typename CacheStore::State, CacheStore>;
+  using ImplBase::PushArc;
+  using ImplBase::HasArcs;
+  using ImplBase::HasFinal;
+  using ImplBase::HasStart;
+  using ImplBase::SetArcs;
+  using ImplBase::SetFinal;
+  using ImplBase::SetStart;
+
+  CompactFstImpl()
+      : ImplBase(CompactFstOptions()),
+        compactor_() {
+    SetType(Compactor::Type());
+    SetProperties(kNullProperties | kStaticProperties);
+  }
+
+  CompactFstImpl(const Fst<Arc> &fst, std::shared_ptr<Compactor> compactor,
+                 const CompactFstOptions &opts)
+      : ImplBase(opts),
+        compactor_(std::make_shared<Compactor>(fst, compactor)) {
+    SetType(Compactor::Type());
+    SetInputSymbols(fst.InputSymbols());
+    SetOutputSymbols(fst.OutputSymbols());
+    if (compactor_->Error()) SetProperties(kError, kError);
+    uint64 copy_properties = fst.Properties(kMutable, false) ?
+        fst.Properties(kCopyProperties, true):
+        CheckProperties(fst,
+                        kCopyProperties & ~kWeightedCycles & ~kUnweightedCycles,
+                        kCopyProperties);
+    if ((copy_properties & kError) || !compactor_->IsCompatible(fst)) {
+      FSTERROR() << "CompactFstImpl: Input Fst incompatible with compactor";
+      SetProperties(kError, kError);
+      return;
+    }
+    SetProperties(copy_properties | kStaticProperties);
+  }
+
+  CompactFstImpl(std::shared_ptr<Compactor> compactor,
+                 const CompactFstOptions &opts)
+      : ImplBase(opts),
+        compactor_(compactor) {
+    SetType(Compactor::Type());
+    SetProperties(kStaticProperties | compactor_->Properties());
+    if (compactor_->Error()) SetProperties(kError, kError);
+  }
+
+  CompactFstImpl(const CompactFstImpl<Arc, Compactor, CacheStore> &impl)
+      : ImplBase(impl),
+        compactor_(impl.compactor_ == nullptr ?
+                   std::make_shared<Compactor>() :
+                   std::make_shared<Compactor>(*impl.compactor_)) {
+    SetType(impl.Type());
+    SetProperties(impl.Properties());
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+  }
+
+  // Allows to change the cache store from OtherI to I.
+  template <class OtherCacheStore>
+  CompactFstImpl(const CompactFstImpl<Arc, Compactor, OtherCacheStore> &impl)
+      : ImplBase(CacheOptions(impl.GetCacheGc(), impl.GetCacheLimit())),
+        compactor_(impl.compactor_ == nullptr ?
+                   std::make_shared<Compactor>() :
+                   std::make_shared<Compactor>(*impl.compactor_)) {
+    SetType(impl.Type());
+    SetProperties(impl.Properties());
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+  }
+
+  StateId Start() {
+    if (!HasStart()) SetStart(compactor_->Start());
+    return ImplBase::Start();
+  }
+
+  Weight Final(StateId s) {
+    if (HasFinal(s)) return ImplBase::Final(s);
+    compactor_->SetState(s, &state_);
+    return state_.Final();
+  }
+
+  StateId NumStates() const {
+    if (Properties(kError)) return 0;
+    return compactor_->NumStates();
+  }
+
+  size_t NumArcs(StateId s) {
+    if (HasArcs(s)) return ImplBase::NumArcs(s);
+    compactor_->SetState(s, &state_);
+    return state_.NumArcs();
+  }
+
+  size_t NumInputEpsilons(StateId s) {
+    if (!HasArcs(s) && !Properties(kILabelSorted)) Expand(s);
+    if (HasArcs(s)) return ImplBase::NumInputEpsilons(s);
+    return CountEpsilons(s, false);
+  }
+
+  size_t NumOutputEpsilons(StateId s) {
+    if (!HasArcs(s) && !Properties(kOLabelSorted)) Expand(s);
+    if (HasArcs(s)) return ImplBase::NumOutputEpsilons(s);
+    return CountEpsilons(s, true);
+  }
+
+  size_t CountEpsilons(StateId s, bool output_epsilons) {
+    compactor_->SetState(s, &state_);
+    const uint32 f = output_epsilons ? kArcOLabelValue : kArcILabelValue;
+    size_t num_eps = 0;
+    for (size_t i = 0; i < state_.NumArcs(); ++i) {
+      const auto& arc = state_.GetArc(i, f);
+      const auto label = output_epsilons ? arc.olabel : arc.ilabel;
+      if (label == 0)
+        ++num_eps;
+      else if (label > 0)
+        break;
+    }
+    return num_eps;
+  }
+
+  static CompactFstImpl<Arc, Compactor, CacheStore> *Read(
+      std::istream &strm, const FstReadOptions &opts) {
+    std::unique_ptr<CompactFstImpl<Arc, Compactor, CacheStore>> impl(
+      new CompactFstImpl<Arc, Compactor, CacheStore>());
+    FstHeader hdr;
+    if (!impl->ReadHeader(strm, opts, kMinFileVersion, &hdr)) {
+      return nullptr;
+    }
+    // Ensures compatibility.
+    if (hdr.Version() == kAlignedFileVersion) {
+      hdr.SetFlags(hdr.GetFlags() | FstHeader::IS_ALIGNED);
+    }
+    impl->compactor_ = std::shared_ptr<Compactor>(
+        Compactor::Read(strm, opts, hdr));
+    if (!impl->compactor_) {
+      return nullptr;
+    }
+    return impl.release();
+  }
+
+  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
+    FstHeader hdr;
+    hdr.SetStart(compactor_->Start());
+    hdr.SetNumStates(compactor_->NumStates());
+    hdr.SetNumArcs(compactor_->NumArcs());
+    // Ensures compatibility.
+    const auto file_version = opts.align ? kAlignedFileVersion : kFileVersion;
+    WriteHeader(strm, opts, file_version, &hdr);
+    return compactor_->Write(strm, opts);
+  }
+
+  // Provides information needed for generic state iterator.
+  void InitStateIterator(StateIteratorData<Arc> *data) const {
+    data->base = nullptr;
+    data->nstates = compactor_->NumStates();
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
+    if (!HasArcs(s)) Expand(s);
+    ImplBase::InitArcIterator(s, data);
+  }
+
+  void Expand(StateId s) {
+    compactor_->SetState(s, &state_);
+    for (size_t i = 0; i < state_.NumArcs(); ++i)
+      PushArc(s, state_.GetArc(i, kArcValueFlags));
+    SetArcs(s);
+    if (!HasFinal(s)) SetFinal(s, state_.Final());
+  }
+
+  const Compactor *GetCompactor() const { return compactor_.get(); }
+  std::shared_ptr<Compactor> SharedCompactor() const { return compactor_; }
+  void SetCompactor(std::shared_ptr<Compactor> compactor) {
+    // TODO(allauzen): is this correct? is this needed?
+    // TODO(allauzen): consider removing and forcing this through direct calls
+    // to compactor.
+    compactor_ = compactor;
+  }
+
+  // Properties always true of this FST class.
+  static constexpr uint64 kStaticProperties = kExpanded;
+
+ protected:
+  template <class OtherArc, class OtherCompactor, class OtherCacheStore>
+  explicit CompactFstImpl(
+    const CompactFstImpl<OtherArc, OtherCompactor, OtherCacheStore> &impl)
+    : compactor_(std::make_shared<Compactor>(*impl.GetCompactor())) {
+    SetType(impl.Type());
+    SetProperties(impl.Properties());
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+  }
+
+ private:
+  // Allows access during write.
+  template <class AnyArc, class ArcCompactor, class Unsigned,
+            class CompactStore, class AnyCacheStore>
+  friend class ::fst::CompactFst;  // allow access during write.
+
+  // Current unaligned file format version.
+  static constexpr int kFileVersion = 2;
+  // Current aligned file format version.
+  static constexpr int kAlignedFileVersion = 1;
+  // Minimum file format version supported.
+  static constexpr int kMinFileVersion = 1;
+
+  std::shared_ptr<Compactor> compactor_;
+  typename Compactor::State state_;
+};
+
+template <class Arc, class Compactor, class CacheStore>
+constexpr uint64 CompactFstImpl<Arc, Compactor, CacheStore>::kStaticProperties;
+
+template <class Arc, class Compactor, class CacheStore>
+constexpr int CompactFstImpl<Arc, Compactor, CacheStore>::kFileVersion;
+
+template <class Arc, class Compactor, class CacheStore>
+constexpr int CompactFstImpl<Arc, Compactor, CacheStore>::kAlignedFileVersion;
+
+template <class Arc, class Compactor, class CacheStore>
+constexpr int CompactFstImpl<Arc, Compactor, CacheStore>::kMinFileVersion;
+
+}  // namespace internal
+
+// This class attaches interface to implementation and handles reference
+// counting, delegating most methods to ImplToExpandedFst. The Unsigned type
+// is used to represent indices into the compact arc array. (Template
+// argument defaults are declared in fst-decl.h.)
+template <class A, class ArcCompactor, class Unsigned, class CompactStore,
+          class CacheStore>
+class CompactFst
+    : public ImplToExpandedFst<internal::CompactFstImpl<
+          A,
+          DefaultCompactor<ArcCompactor, Unsigned, CompactStore>,
+          CacheStore>> {
+ public:
+  template <class F, class G>
+  void friend Cast(const F &, G *);
+
+  using Arc = A;
+  using StateId = typename A::StateId;
+  using Compactor = DefaultCompactor<ArcCompactor, Unsigned, CompactStore>;
+  using Impl = internal::CompactFstImpl<A, Compactor, CacheStore>;
+  using Store = CacheStore;  // for CacheArcIterator
+
+  friend class StateIterator<
+      CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>>;
+  friend class ArcIterator<
+      CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>>;
+
+  CompactFst() : ImplToExpandedFst<Impl>(std::make_shared<Impl>()) {}
+
+  // If data is not nullptr, it is assumed to be already initialized.
+  explicit CompactFst(
+      const Fst<A> &fst,
+      const ArcCompactor &compactor = ArcCompactor(),
+      const CompactFstOptions &opts = CompactFstOptions(),
+      std::shared_ptr<CompactStore> data = std::shared_ptr<CompactStore>())
+      : ImplToExpandedFst<Impl>(
+            std::make_shared<Impl>(
+                fst,
+                std::make_shared<Compactor>(
+                    std::make_shared<ArcCompactor>(compactor), data),
+                opts)) {}
+
+  // If data is not nullptr, it is assumed to be already initialized.
+  CompactFst(
+      const Fst<Arc> &fst,
+      std::shared_ptr<ArcCompactor> compactor,
+      const CompactFstOptions &opts = CompactFstOptions(),
+      std::shared_ptr<CompactStore> data = std::shared_ptr<CompactStore>())
+      : ImplToExpandedFst<Impl>(
+            std::make_shared<Impl>(fst,
+                                   std::make_shared<Compactor>(compactor, data),
+                                   opts)) {}
+
+  // The following 2 constructors take as input two iterators delimiting a set
+  // of (already) compacted transitions, starting with the transitions out of
+  // the initial state. The format of the input differs for fixed out-degree
+  // and variable out-degree compactors.
+  //
+  // - For fixed out-degree compactors, the final weight (encoded as a
+  // compacted transition) needs to be given only for final states. All strings
+  // (compactor of size 1) will be assume to be terminated by a final state
+  // even when the final state is not implicitely given.
+  //
+  // - For variable out-degree compactors, the final weight (encoded as a
+  // compacted transition) needs to be given for all states and must appeared
+  // first in the list (for state s, final weight of s, followed by outgoing
+  // transitons in s).
+  //
+  // These 2 constructors allows the direct construction of a CompactFst
+  // without first creating a more memory-hungry regular FST. This is useful
+  // when memory usage is severely constrained.
+  template <class Iterator>
+  explicit CompactFst(const Iterator &begin, const Iterator &end,
+                      const ArcCompactor &compactor = ArcCompactor(),
+                      const CompactFstOptions &opts = CompactFstOptions())
+      : ImplToExpandedFst<Impl>(
+            std::make_shared<Impl>(
+                std::make_shared<Compactor>(
+                    begin, end, std::make_shared<ArcCompactor>(compactor)),
+                opts)) {}
+
+  template <class Iterator>
+  CompactFst(const Iterator &begin, const Iterator &end,
+             std::shared_ptr<ArcCompactor> compactor,
+             const CompactFstOptions &opts = CompactFstOptions())
+      : ImplToExpandedFst<Impl>(
+            std::make_shared<Impl>(
+                std::make_shared<Compactor>(begin, end, compactor), opts)) {}
+
+  // See Fst<>::Copy() for doc.
+  CompactFst(
+      const CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>
+      &fst,
+      bool safe = false)
+      : ImplToExpandedFst<Impl>(fst, safe) {}
+
+  // Get a copy of this CompactFst. See Fst<>::Copy() for further doc.
+  CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore> *Copy(
+      bool safe = false) const override {
+    return new CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>(
+        *this, safe);
+  }
+
+  // Read a CompactFst from an input stream; return nullptr on error
+  static CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore> *Read(
+      std::istream &strm, const FstReadOptions &opts) {
+    auto *impl = Impl::Read(strm, opts);
+    return impl ? new CompactFst<A, ArcCompactor, Unsigned, CompactStore,
+                                 CacheStore>(std::shared_ptr<Impl>(impl))
+                : nullptr;
+  }
+
+  // Read a CompactFst from a file; return nullptr on error
+  // Empty filename reads from standard input
+  static CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore> *Read(
+      const string &filename) {
+    auto *impl = ImplToExpandedFst<Impl>::Read(filename);
+    return impl ? new CompactFst<A, ArcCompactor, Unsigned, CompactStore,
+                                 CacheStore>(std::shared_ptr<Impl>(impl))
+                : nullptr;
+  }
+
+  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
+    return GetImpl()->Write(strm, opts);
+  }
+
+  bool Write(const string &filename) const override {
+    return Fst<Arc>::WriteFile(filename);
+  }
+
+  template <class FST>
+  static bool WriteFst(const FST &fst, const ArcCompactor &compactor,
+                       std::ostream &strm, const FstWriteOptions &opts);
+
+  void InitStateIterator(StateIteratorData<Arc> *data) const override {
+    GetImpl()->InitStateIterator(data);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
+    GetMutableImpl()->InitArcIterator(s, data);
+  }
+
+  MatcherBase<Arc> *InitMatcher(MatchType match_type) const override {
+    return new SortedMatcher<
+        CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>>(
+        *this, match_type);
+  }
+
+  template <class Iterator>
+  void SetCompactElements(const Iterator &b, const Iterator &e) {
+    GetMutableImpl()->SetCompactor(std::make_shared<Compactor>(
+        b, e, std::make_shared<ArcCompactor>()));
+  }
+
+ private:
+  using ImplToFst<Impl, ExpandedFst<Arc>>::GetImpl;
+  using ImplToFst<Impl, ExpandedFst<Arc>>::GetMutableImpl;
+
+  explicit CompactFst(std::shared_ptr<Impl> impl)
+      : ImplToExpandedFst<Impl>(impl) {}
+
+  // Use overloading to extract the type of the argument.
+  static Impl *GetImplIfCompactFst(
+      const CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>
+          &compact_fst) {
+    return compact_fst.GetImpl();
+  }
+
+  // This does not give privileged treatment to subclasses of CompactFst.
+  template <typename NonCompactFst>
+  static Impl *GetImplIfCompactFst(const NonCompactFst &fst) {
+    return nullptr;
+  }
+
+  CompactFst &operator=(const CompactFst &fst) = delete;
+};
+
+// Writes FST in Compact format, with a possible pass over the machine before
+// writing to compute the number of states and arcs.
+template <class A, class ArcCompactor, class Unsigned, class CompactStore,
+          class CacheStore>
+template <class FST>
+bool CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>::WriteFst(
+    const FST &fst, const ArcCompactor &compactor, std::ostream &strm,
+    const FstWriteOptions &opts) {
+  using Arc = A;
+  using Weight = typename A::Weight;
+  using Element = typename ArcCompactor::Element;
+  const auto file_version =
+      opts.align ? Impl::kAlignedFileVersion : Impl::kFileVersion;
+  size_t num_arcs = -1;
+  size_t num_states = -1;
+  auto first_pass_compactor = compactor;
+  if (auto *impl = GetImplIfCompactFst(fst)) {
+    num_arcs = impl->GetCompactor()->GetCompactStore()->NumArcs();
+    num_states = impl->GetCompactor()->GetCompactStore()->NumStates();
+    first_pass_compactor = *impl->GetCompactor()->GetArcCompactor();
+  } else {
+    // A first pass is needed to compute the state of the compactor, which
+    // is saved ahead of the rest of the data structures. This unfortunately
+    // means forcing a complete double compaction when writing in this format.
+    // TODO(allauzen): eliminate mutable state from compactors.
+    num_arcs = 0;
+    num_states = 0;
+    for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
+      const auto s = siter.Value();
+      ++num_states;
+      if (fst.Final(s) != Weight::Zero()) {
+        first_pass_compactor.Compact(
+            s, Arc(kNoLabel, kNoLabel, fst.Final(s), kNoStateId));
+      }
+      for (ArcIterator<FST> aiter(fst, s); !aiter.Done(); aiter.Next()) {
+        ++num_arcs;
+        first_pass_compactor.Compact(s, aiter.Value());
+      }
+    }
+  }
+  FstHeader hdr;
+  hdr.SetStart(fst.Start());
+  hdr.SetNumStates(num_states);
+  hdr.SetNumArcs(num_arcs);
+  string type = "compact";
+  if (sizeof(Unsigned) != sizeof(uint32)) {
+    type += std::to_string(CHAR_BIT * sizeof(Unsigned));
+  }
+  type += "_";
+  type += ArcCompactor::Type();
+  if (CompactStore::Type() != "compact") {
+    type += "_";
+    type += CompactStore::Type();
+  }
+  const auto copy_properties = fst.Properties(kCopyProperties, true);
+  if ((copy_properties & kError) || !compactor.Compatible(fst)) {
+    FSTERROR() << "Fst incompatible with compactor";
+    return false;
+  }
+  uint64 properties = copy_properties | Impl::kStaticProperties;
+  internal::FstImpl<Arc>::WriteFstHeader(fst, strm, opts, file_version, type,
+                                         properties, &hdr);
+  first_pass_compactor.Write(strm);
+  if (first_pass_compactor.Size() == -1) {
+    if (opts.align && !AlignOutput(strm)) {
+      LOG(ERROR) << "CompactFst::Write: Alignment failed: " << opts.source;
+      return false;
+    }
+    Unsigned compacts = 0;
+    for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
+      const auto s = siter.Value();
+      strm.write(reinterpret_cast<const char *>(&compacts), sizeof(compacts));
+      if (fst.Final(s) != Weight::Zero()) {
+        ++compacts;
+      }
+      compacts += fst.NumArcs(s);
+    }
+    strm.write(reinterpret_cast<const char *>(&compacts), sizeof(compacts));
+  }
+  if (opts.align && !AlignOutput(strm)) {
+    LOG(ERROR) << "Could not align file during write after writing states";
+  }
+  const auto &second_pass_compactor = compactor;
+  Element element;
+  for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
+    const auto s = siter.Value();
+    if (fst.Final(s) != Weight::Zero()) {
+      element = second_pass_compactor.Compact(
+          s, A(kNoLabel, kNoLabel, fst.Final(s), kNoStateId));
+      strm.write(reinterpret_cast<const char *>(&element), sizeof(element));
+    }
+    for (ArcIterator<FST> aiter(fst, s); !aiter.Done(); aiter.Next()) {
+      element = second_pass_compactor.Compact(s, aiter.Value());
+      strm.write(reinterpret_cast<const char *>(&element), sizeof(element));
+    }
+  }
+  strm.flush();
+  if (!strm) {
+    LOG(ERROR) << "CompactFst write failed: " << opts.source;
+    return false;
+  }
+  return true;
+}
+
+// Specialization for CompactFst; see generic version in fst.h for sample
+// usage (but use the CompactFst type!). This version should inline.
+template <class Arc, class ArcCompactor, class Unsigned, class CompactStore,
+          class CacheStore>
+class StateIterator<
+    CompactFst<Arc, ArcCompactor, Unsigned, CompactStore, CacheStore>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  explicit StateIterator(
+      const CompactFst<Arc, ArcCompactor, Unsigned, CompactStore,
+                       CacheStore> &fst)
+      : nstates_(fst.GetImpl()->NumStates()), s_(0) {}
+
+  bool Done() const { return s_ >= nstates_; }
+
+  StateId Value() const { return s_; }
+
+  void Next() { ++s_; }
+
+  void Reset() { s_ = 0; }
+
+ private:
+  StateId nstates_;
+  StateId s_;
+};
+
+// Specialization for CompactFst. Never caches,
+// always iterates over the underlying compact elements.
+template <class Arc, class ArcCompactor, class Unsigned,
+          class CompactStore, class CacheStore>
+class ArcIterator<CompactFst<
+    Arc, ArcCompactor, Unsigned, CompactStore, CacheStore>> {
+ public:
+  using StateId = typename Arc::StateId;
+  using Element = typename ArcCompactor::Element;
+  using Compactor = DefaultCompactor<ArcCompactor, Unsigned, CompactStore>;
+  using State = typename Compactor::State;
+
+  ArcIterator(const CompactFst<Arc, ArcCompactor, Unsigned, CompactStore,
+                               CacheStore> &fst,
+              StateId s)
+      : state_(fst.GetImpl()->GetCompactor(), s),
+        pos_(0),
+        flags_(kArcValueFlags) {}
+
+  bool Done() const { return pos_ >= state_.NumArcs(); }
+
+  const Arc &Value() const {
+    arc_ = state_.GetArc(pos_, flags_);
+    return arc_;
+  }
+
+  void Next() { ++pos_; }
+
+  size_t Position() const { return pos_; }
+
+  void Reset() { pos_ = 0; }
+
+  void Seek(size_t pos) { pos_ = pos; }
+
+  uint32 Flags() const { return flags_; }
+
+  void SetFlags(uint32 f, uint32 m) {
+    flags_ &= ~m;
+    flags_ |= (f & kArcValueFlags);
+  }
+
+ private:
+  State state_;
+  size_t pos_;
+  mutable Arc arc_;
+  uint32 flags_;
+};
+
+// ArcCompactor for unweighted string FSTs.
+template <class A>
+class StringCompactor {
+ public:
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Element = Label;
+
+  Element Compact(StateId s, const Arc &arc) const { return arc.ilabel; }
+
+  Arc Expand(StateId s, const Element &p, uint32 f = kArcValueFlags) const {
+    return Arc(p, p, Weight::One(), p != kNoLabel ? s + 1 : kNoStateId);
+  }
+
+  constexpr ssize_t Size() const { return 1; }
+
+  constexpr uint64 Properties() const {
+    return kString | kAcceptor | kUnweighted;
+  }
+
+  bool Compatible(const Fst<Arc> &fst) const {
+    const auto props = Properties();
+    return fst.Properties(props, true) == props;
+  }
+
+  static const string &Type() {
+    static const string *const type = new string("string");
+    return *type;
+  }
+
+  bool Write(std::ostream &strm) const { return true; }
+
+  static StringCompactor *Read(std::istream &strm) {
+    return new StringCompactor;
+  }
+};
+
+// ArcCompactor for weighted string FSTs.
+template <class A>
+class WeightedStringCompactor {
+ public:
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Element = std::pair<Label, Weight>;
+
+  Element Compact(StateId s, const Arc &arc) const {
+    return std::make_pair(arc.ilabel, arc.weight);
+  }
+
+  Arc Expand(StateId s, const Element &p, uint32 f = kArcValueFlags) const {
+    return Arc(p.first, p.first, p.second,
+               p.first != kNoLabel ? s + 1 : kNoStateId);
+  }
+
+  constexpr ssize_t Size() const { return 1; }
+
+  constexpr uint64 Properties() const { return kString | kAcceptor; }
+
+  bool Compatible(const Fst<Arc> &fst) const {
+    const auto props = Properties();
+    return fst.Properties(props, true) == props;
+  }
+
+  static const string &Type() {
+    static const string *const type = new string("weighted_string");
+    return *type;
+  }
+
+  bool Write(std::ostream &strm) const { return true; }
+
+  static WeightedStringCompactor *Read(std::istream &strm) {
+    return new WeightedStringCompactor;
+  }
+};
+
+// ArcCompactor for unweighted acceptor FSTs.
+template <class A>
+class UnweightedAcceptorCompactor {
+ public:
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Element = std::pair<Label, StateId>;
+
+  Element Compact(StateId s, const Arc &arc) const {
+    return std::make_pair(arc.ilabel, arc.nextstate);
+  }
+
+  Arc Expand(StateId s, const Element &p, uint32 f = kArcValueFlags) const {
+    return Arc(p.first, p.first, Weight::One(), p.second);
+  }
+
+  constexpr ssize_t Size() const { return -1; }
+
+  constexpr uint64 Properties() const { return kAcceptor | kUnweighted; }
+
+  bool Compatible(const Fst<Arc> &fst) const {
+    const auto props = Properties();
+    return fst.Properties(props, true) == props;
+  }
+
+  static const string &Type() {
+    static const string *const type = new string("unweighted_acceptor");
+    return *type;
+  }
+
+  bool Write(std::ostream &strm) const { return true; }
+
+  static UnweightedAcceptorCompactor *Read(std::istream &istrm) {
+    return new UnweightedAcceptorCompactor;
+  }
+};
+
+// ArcCompactor for weighted acceptor FSTs.
+template <class A>
+class AcceptorCompactor {
+ public:
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Element = std::pair<std::pair<Label, Weight>, StateId>;
+
+  Element Compact(StateId s, const Arc &arc) const {
+    return std::make_pair(std::make_pair(arc.ilabel, arc.weight),
+                          arc.nextstate);
+  }
+
+  Arc Expand(StateId s, const Element &p, uint32 f = kArcValueFlags) const {
+    return Arc(p.first.first, p.first.first, p.first.second, p.second);
+  }
+
+  constexpr ssize_t Size() const { return -1; }
+
+  constexpr uint64 Properties() const { return kAcceptor; }
+
+  bool Compatible(const Fst<Arc> &fst) const {
+    const auto props = Properties();
+    return fst.Properties(props, true) == props;
+  }
+
+  static const string &Type() {
+    static const string *const type = new string("acceptor");
+    return *type;
+  }
+
+  bool Write(std::ostream &strm) const { return true; }
+
+  static AcceptorCompactor *Read(std::istream &strm) {
+    return new AcceptorCompactor;
+  }
+};
+
+// ArcCompactor for unweighted FSTs.
+template <class A>
+class UnweightedCompactor {
+ public:
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Element = std::pair<std::pair<Label, Label>, StateId>;
+
+  Element Compact(StateId s, const Arc &arc) const {
+    return std::make_pair(std::make_pair(arc.ilabel, arc.olabel),
+                          arc.nextstate);
+  }
+
+  Arc Expand(StateId s, const Element &p, uint32 f = kArcValueFlags) const {
+    return Arc(p.first.first, p.first.second, Weight::One(), p.second);
+  }
+
+  constexpr ssize_t Size() const { return -1; }
+
+  constexpr uint64 Properties() const { return kUnweighted; }
+
+  bool Compatible(const Fst<Arc> &fst) const {
+    const auto props = Properties();
+    return fst.Properties(props, true) == props;
+  }
+
+  static const string &Type() {
+    static const string *const type = new string("unweighted");
+    return *type;
+  }
+
+  bool Write(std::ostream &strm) const { return true; }
+
+  static UnweightedCompactor *Read(std::istream &strm) {
+    return new UnweightedCompactor;
+  }
+};
+
+template <class Arc, class Unsigned /* = uint32 */>
+using CompactStringFst = CompactFst<Arc, StringCompactor<Arc>, Unsigned>;
+
+template <class Arc, class Unsigned /* = uint32 */>
+using CompactWeightedStringFst =
+    CompactFst<Arc, WeightedStringCompactor<Arc>, Unsigned>;
+
+template <class Arc, class Unsigned /* = uint32 */>
+using CompactAcceptorFst = CompactFst<Arc, AcceptorCompactor<Arc>, Unsigned>;
+
+template <class Arc, class Unsigned /* = uint32 */>
+using CompactUnweightedFst =
+    CompactFst<Arc, UnweightedCompactor<Arc>, Unsigned>;
+
+template <class Arc, class Unsigned /* = uint32 */>
+using CompactUnweightedAcceptorFst =
+    CompactFst<Arc, UnweightedAcceptorCompactor<Arc>, Unsigned>;
+
+using StdCompactStringFst = CompactStringFst<StdArc, uint32>;
+
+using StdCompactWeightedStringFst = CompactWeightedStringFst<StdArc, uint32>;
+
+using StdCompactAcceptorFst = CompactAcceptorFst<StdArc, uint32>;
+
+using StdCompactUnweightedFst = CompactUnweightedFst<StdArc, uint32>;
+
+using StdCompactUnweightedAcceptorFst =
+    CompactUnweightedAcceptorFst<StdArc, uint32>;
+
+}  // namespace fst
+
+#endif  // FST_COMPACT_FST_H_
diff --git a/projects/llm_framework/include/fst/compat.h b/projects/llm_framework/include/fst/compat.h
new file mode 100644
index 00000000..73ed5737
--- /dev/null
+++ b/projects/llm_framework/include/fst/compat.h
@@ -0,0 +1,130 @@
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_LIB_COMPAT_H_
+#define FST_LIB_COMPAT_H_
+
+#include <climits>
+#include <cstdlib>
+#include <cstring>
+#include <iostream>
+#include <string>
+#include <vector>
+
+// Makes copy constructor and operator= private
+// Deprecated: now just use =delete.
+#define DISALLOW_COPY_AND_ASSIGN(type)    \
+  type(const type&);                      \
+  void operator=(const type&)
+
+#if defined(__GNUC__) || defined(__clang__)
+#define OPENFST_DEPRECATED(message) __attribute__((deprecated(message)))
+#elif defined(_MSC_VER)
+#define OPENFST_DEPRECATED(message) __declspec(deprecated(message))
+#else
+#define OPENFST_DEPRECATED(message)
+#endif
+
+#include <fst/config.h>
+#include <fst/types.h>
+#include <fst/lock.h>
+#include <fst/flags.h>
+#include <fst/log.h>
+#include <fst/icu.h>
+
+using std::string;
+
+void FailedNewHandler();
+
+#ifdef _MSC_VER
+#include <intrin.h>
+const char* basename(const char* path);
+#define __builtin_popcount __popcnt
+
+#ifdef _M_X64
+// Using 64-bit MSVC intrinsics.
+#define __builtin_popcountll __popcnt64
+inline unsigned int __builtin_ctzll(std::uint64_t w) {
+  unsigned long v;
+  return _BitScanForward64(&v, std::uint32_t(w)) ? v : 0;
+}
+#else
+// Using 32-bit MSVC intrinsics.
+inline unsigned int __builtin_popcountll(std::uint64_t w) {
+  return __popcnt(std::uint32_t(w)) + __popcnt(std::uint32_t(w >> 32));
+}
+inline unsigned int __builtin_ctzll(std::uint64_t w) {
+  unsigned long v;
+  return (_BitScanForward(&v, std::uint32_t(w)) ? v :
+          _BitScanForward(&v, std::uint32_t(w >> 32)) ? v + 32 : 0);
+}
+#endif  // _M_X64
+#endif  // _MSC_VER
+
+namespace fst {
+
+// Downcasting.
+template<typename To, typename From>
+inline To down_cast(From* f) { return static_cast<To>(f); }
+
+// Bitcasting.
+template <class Dest, class Source>
+inline Dest bit_cast(const Source &source) {
+  static_assert(sizeof(Dest) == sizeof(Source),
+                "Bitcasting unsafe for specified types");
+  Dest dest;
+  memcpy(&dest, &source, sizeof(dest));
+  return dest;
+}
+
+// Check sums
+class CheckSummer {
+ public:
+  CheckSummer() : count_(0) {
+    check_sum_.resize(kCheckSumLength, '\0');
+  }
+
+  void Reset() {
+    count_ = 0;
+    for (int i = 0; i < kCheckSumLength; ++i) check_sum_[i] = '\0';
+  }
+
+  void Update(void const *data, int size) {
+    const char *p = reinterpret_cast<const char *>(data);
+    for (int i = 0; i < size; ++i) {
+      check_sum_[(count_++) % kCheckSumLength] ^= p[i];
+    }
+  }
+
+  void Update(string const &data) {
+    for (int i = 0; i < data.size(); ++i) {
+      check_sum_[(count_++) % kCheckSumLength] ^= data[i];
+    }
+  }
+
+  string Digest() { return check_sum_; }
+
+ private:
+  static const int kCheckSumLength = 32;
+  int count_;
+  string check_sum_;
+
+  CheckSummer(const CheckSummer &) = delete;
+  CheckSummer &operator=(const CheckSummer &) = delete;
+};
+
+}  // namespace fst
+
+#endif  // FST_LIB_COMPAT_H_
diff --git a/projects/llm_framework/include/fst/complement.h b/projects/llm_framework/include/fst/complement.h
new file mode 100644
index 00000000..64eebc03
--- /dev/null
+++ b/projects/llm_framework/include/fst/complement.h
@@ -0,0 +1,277 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to complement an FST.
+
+#ifndef FST_COMPLEMENT_H_
+#define FST_COMPLEMENT_H_
+
+#include <algorithm>
+#include <string>
+#include <vector>
+#include <fst/log.h>
+
+#include <fst/fst.h>
+#include <fst/test-properties.h>
+
+
+namespace fst {
+
+template <class Arc>
+class ComplementFst;
+
+namespace internal {
+
+// Implementation of delayed ComplementFst. The algorithm used completes the
+// (deterministic) FSA and then exchanges final and non-final states.
+// Completion, i.e. ensuring that all labels can be read from every state, is
+// accomplished by using ρ-labels, which match all labels that are otherwise
+// not found leaving a state. The first state in the output is reserved to be a
+// new state that is the destination of all ρ-labels. Each remaining output
+// state s corresponds to input state s - 1. The first arc in the output at
+// these states is the ρ-label, the remaining arcs correspond to the input
+// arcs.
+template <class A>
+class ComplementFstImpl : public FstImpl<A> {
+ public:
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FstImpl<A>::SetType;
+  using FstImpl<A>::SetProperties;
+  using FstImpl<A>::SetInputSymbols;
+  using FstImpl<A>::SetOutputSymbols;
+
+  friend class StateIterator<ComplementFst<Arc>>;
+  friend class ArcIterator<ComplementFst<Arc>>;
+
+  explicit ComplementFstImpl(const Fst<Arc> &fst) : fst_(fst.Copy()) {
+    SetType("complement");
+    uint64 props = fst.Properties(kILabelSorted, false);
+    SetProperties(ComplementProperties(props), kCopyProperties);
+    SetInputSymbols(fst.InputSymbols());
+    SetOutputSymbols(fst.OutputSymbols());
+  }
+
+  ComplementFstImpl(const ComplementFstImpl<Arc> &impl)
+      : fst_(impl.fst_->Copy()) {
+    SetType("complement");
+    SetProperties(impl.Properties(), kCopyProperties);
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+  }
+
+  StateId Start() const {
+    if (Properties(kError)) return kNoStateId;
+    auto start = fst_->Start();
+    return start != kNoStateId ? start + 1 : 0;
+  }
+
+  // Exchange final and non-final states; makes ρ-destination state final.
+  Weight Final(StateId s) const {
+    if (s == 0 || fst_->Final(s - 1) == Weight::Zero()) {
+      return Weight::One();
+    } else {
+      return Weight::Zero();
+    }
+  }
+
+  size_t NumArcs(StateId s) const {
+    return s == 0 ? 1 : fst_->NumArcs(s - 1) + 1;
+  }
+
+  size_t NumInputEpsilons(StateId s) const {
+    return s == 0 ? 0 : fst_->NumInputEpsilons(s - 1);
+  }
+
+  size_t NumOutputEpsilons(StateId s) const {
+    return s == 0 ? 0 : fst_->NumOutputEpsilons(s - 1);
+  }
+
+  uint64 Properties() const override { return Properties(kFstProperties); }
+
+  // Sets error if found, and returns other FST impl properties.
+  uint64 Properties(uint64 mask) const override {
+    if ((mask & kError) && fst_->Properties(kError, false)) {
+      SetProperties(kError, kError);
+    }
+    return FstImpl<Arc>::Properties(mask);
+  }
+
+ private:
+  std::unique_ptr<const Fst<Arc>> fst_;
+};
+
+}  // namespace internal
+
+// Complements an automaton. This is a library-internal operation that
+// introduces a (negative) ρ-label; use Difference/DifferenceFst in user code,
+// which will not see this label. This version is a delayed FST.
+//
+// This class attaches interface to implementation and handles
+// reference counting, delegating most methods to ImplToFst.
+template <class A>
+class ComplementFst : public ImplToFst<internal::ComplementFstImpl<A>> {
+ public:
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Impl = internal::ComplementFstImpl<Arc>;
+
+  friend class StateIterator<ComplementFst<Arc>>;
+  friend class ArcIterator<ComplementFst<Arc>>;
+
+  explicit ComplementFst(const Fst<Arc> &fst)
+      : ImplToFst<Impl>(std::make_shared<Impl>(fst)) {
+    static constexpr auto props =
+        kUnweighted | kNoEpsilons | kIDeterministic | kAcceptor;
+    if (fst.Properties(props, true) != props) {
+      FSTERROR() << "ComplementFst: Argument not an unweighted "
+                 << "epsilon-free deterministic acceptor";
+      GetImpl()->SetProperties(kError, kError);
+    }
+  }
+
+  // See Fst<>::Copy() for doc.
+  ComplementFst(const ComplementFst<Arc> &fst, bool safe = false)
+      : ImplToFst<Impl>(fst, safe) {}
+
+  // Gets a copy of this FST. See Fst<>::Copy() for further doc.
+  ComplementFst<Arc> *Copy(bool safe = false) const override {
+    return new ComplementFst<Arc>(*this, safe);
+  }
+
+  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
+
+  inline void InitArcIterator(StateId s,
+                              ArcIteratorData<Arc> *data) const override;
+
+  // Label that represents the ρ-transition; we use a negative value private to
+  // the library and which will preserve FST label sort order.
+  static const Label kRhoLabel = -2;
+
+ private:
+  using ImplToFst<Impl>::GetImpl;
+
+  ComplementFst &operator=(const ComplementFst &) = delete;
+};
+
+template <class Arc>
+const typename Arc::Label ComplementFst<Arc>::kRhoLabel;
+
+// Specialization for ComplementFst.
+template <class Arc>
+class StateIterator<ComplementFst<Arc>> : public StateIteratorBase<Arc> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  explicit StateIterator(const ComplementFst<Arc> &fst)
+      : siter_(*fst.GetImpl()->fst_), s_(0) {}
+
+  bool Done() const final { return s_ > 0 && siter_.Done(); }
+
+  StateId Value() const final { return s_; }
+
+  void Next() final {
+    if (s_ != 0) siter_.Next();
+    ++s_;
+  }
+
+  void Reset() final {
+    siter_.Reset();
+    s_ = 0;
+  }
+
+ private:
+  StateIterator<Fst<Arc>> siter_;
+  StateId s_;
+};
+
+// Specialization for ComplementFst.
+template <class Arc>
+class ArcIterator<ComplementFst<Arc>> : public ArcIteratorBase<Arc> {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  ArcIterator(const ComplementFst<Arc> &fst, StateId s) : s_(s), pos_(0) {
+    if (s_ != 0) {
+      aiter_.reset(new ArcIterator<Fst<Arc>>(*fst.GetImpl()->fst_, s - 1));
+    }
+  }
+
+  bool Done() const final {
+    if (s_ != 0) {
+      return pos_ > 0 && aiter_->Done();
+    } else {
+      return pos_ > 0;
+    }
+  }
+
+  // Adds the ρ-label to the ρ destination state.
+  const Arc &Value() const final {
+    if (pos_ == 0) {
+      arc_.ilabel = arc_.olabel = ComplementFst<Arc>::kRhoLabel;
+      arc_.weight = Weight::One();
+      arc_.nextstate = 0;
+    } else {
+      arc_ = aiter_->Value();
+      ++arc_.nextstate;
+    }
+    return arc_;
+  }
+
+  void Next() final {
+    if (s_ != 0 && pos_ > 0) aiter_->Next();
+    ++pos_;
+  }
+
+  size_t Position() const final { return pos_; }
+
+  void Reset() final {
+    if (s_ != 0) aiter_->Reset();
+    pos_ = 0;
+  }
+
+  void Seek(size_t a) final {
+    if (s_ != 0) {
+      if (a == 0) {
+        aiter_->Reset();
+      } else {
+        aiter_->Seek(a - 1);
+      }
+    }
+    pos_ = a;
+  }
+
+  uint32 Flags() const final { return kArcValueFlags; }
+
+  void SetFlags(uint32, uint32) final {}
+
+ private:
+  std::unique_ptr<ArcIterator<Fst<Arc>>> aiter_;
+  StateId s_;
+  size_t pos_;
+  mutable Arc arc_;
+};
+
+template <class Arc>
+inline void ComplementFst<Arc>::InitStateIterator(
+    StateIteratorData<Arc> *data) const {
+  data->base = new StateIterator<ComplementFst<Arc>>(*this);
+}
+
+template <class Arc>
+inline void ComplementFst<Arc>::InitArcIterator(StateId s,
+    ArcIteratorData<Arc> *data) const {
+  data->base = new ArcIterator<ComplementFst<Arc>>(*this, s);
+}
+
+// Useful alias when using StdArc.
+using StdComplementFst = ComplementFst<StdArc>;
+
+}  // namespace fst
+
+#endif  // FST_COMPLEMENT_H_
diff --git a/projects/llm_framework/include/fst/compose-filter.h b/projects/llm_framework/include/fst/compose-filter.h
new file mode 100644
index 00000000..7251e273
--- /dev/null
+++ b/projects/llm_framework/include/fst/compose-filter.h
@@ -0,0 +1,571 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Classes for filtering the composition matches, e.g. for correct epsilon
+// handling.
+
+#ifndef FST_COMPOSE_FILTER_H_
+#define FST_COMPOSE_FILTER_H_
+
+#include <fst/filter-state.h>
+#include <fst/fst-decl.h>  // For optional argument declarations
+#include <fst/fst.h>
+#include <fst/matcher.h>
+
+
+namespace fst {
+
+// Composition filters determine which matches are allowed to proceed. The
+// filter's state is represeted by the type ComposeFilter::FilterState.
+// The basic filters handle correct epsilon matching. Their interface is:
+//
+// template <class M1, class M2>
+// class ComposeFilter {
+//  public:
+//   using Matcher1 = ...;
+//   using Matcher2 = ...;
+//   using FST1 = typename M1::FST;
+//   using FST2 = typename M2::FST;
+//   using FilterState = ...;
+//
+//   using Arc = typename FST1::Arc;
+//   using StateId = typename Arc::StateId;
+//   using Weight = typename Arc::Weight;
+//
+//   // Required constructor.
+//   ComposeFilter(const FST1 &fst1, const FST2 &fst2,
+//                 M1 *matcher1 = nullptr, M2 *matcher2 = nullptr);
+//
+//   // If safe=true, the copy is thread-safe. See Fst<>::Copy()
+//   // for further doc.
+//   ComposeFilter(const ComposeFilter<M1, M2> &filter,
+//                 bool safe = false);
+//
+//   // Return start state of filter.
+//   FilterState Start() const;
+//
+//   // Specifies current composition state.
+//   void SetState(StateId s1, StateId s2, const FilterState &fs);
+//
+//   // Apply filter at current composition state to these transitions. If an
+//   // arc label to be matched is kNolabel, then that side does not consume a
+//   // symbol. Returns the new filter state or, if disallowed,
+//   // FilterState::NoState(). The filter is permitted to modify its inputs
+//   // (e.g. for optimization reasons).
+//   FilterState FilterArc(Arc *arc1, Arc *arc2) const;
+
+//   // Apply filter at current composition state to these final weights
+//   // (cf. superfinal transitions). The filter may modify its inputs
+//   // (e.g. for optimization reasons).
+//   void FilterFinal(Weight *w1, Weight *w2) const;
+//
+//   // Return the respective matchers. Ownership stays with filter. These
+//   // methods allow the filter to access and possibly modify the compositio
+//   // matchers (useful, e.g., with lookahead).
+//
+//   Matcher1 *GetMatcher1();
+//
+//   Matcher2 *GetMatcher2();
+//
+//   // This specifies how the filter affects the composition result properties.
+//   It takes as argument the properties that would apply with a trivial
+//   // composition filter.
+//   uint64 Properties(uint64 props) const;
+// };
+//
+// This filter allows only exact matching of symbols from FST1 with on FST2;
+// e.g., no special interpretation of epsilons.
+template <class M1, class M2 /* = M1 */>
+class NullComposeFilter {
+ public:
+  using Matcher1 = M1;
+  using Matcher2 = M2;
+  using FST1 = typename M1::FST;
+  using FST2 = typename M2::FST;
+  using FilterState = TrivialFilterState;
+
+  using Arc = typename FST1::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  NullComposeFilter(const FST1 &fst1, const FST2 &fst2,
+                    Matcher1 *matcher1 = nullptr, Matcher2 *matcher2 = nullptr)
+      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
+        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
+        fst1_(matcher1_->GetFst()),
+        fst2_(matcher2_->GetFst()) {}
+
+  NullComposeFilter(const NullComposeFilter<M1, M2> &filter, bool safe = false)
+      : matcher1_(filter.matcher1_->Copy(safe)),
+        matcher2_(filter.matcher2_->Copy(safe)),
+        fst1_(matcher1_->GetFst()),
+        fst2_(matcher2_->GetFst()) {}
+
+  FilterState Start() const { return FilterState(true); }
+
+  void SetState(StateId, StateId, const FilterState &) {}
+
+  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
+    return (arc1->olabel == kNoLabel || arc2->ilabel == kNoLabel)
+               ? FilterState::NoState()
+               : FilterState(true);
+  }
+
+  void FilterFinal(Weight *, Weight *) const {}
+
+  Matcher1 *GetMatcher1() { return matcher1_.get(); }
+
+  Matcher2 *GetMatcher2() { return matcher2_.get(); }
+
+  uint64 Properties(uint64 props) const { return props; }
+
+ private:
+  std::unique_ptr<Matcher1> matcher1_;
+  std::unique_ptr<Matcher2> matcher2_;
+  const FST1 &fst1_;
+  const FST2 &fst2_;
+};
+
+// This filter allows all epsilon matches, potentially resulting in redundant
+// epsilon paths. The use of this filter gives correct results iff one of the
+// following conditions hold:
+//
+//  (1) The semiring is idempotent,
+//  (2) the first FST is output-epsilon free, or
+//  (3) the second FST is input-epsilon free.
+//
+// For (1), redundant epsilon paths may be created but won't hurt correctness.
+// For (2) and (3), no redundant paths are created.
+template <class M1, class M2 /* = M1 */>
+class TrivialComposeFilter {
+ public:
+  using Matcher1 = M1;
+  using Matcher2 = M2;
+  using FST1 = typename M1::FST;
+  using FST2 = typename M2::FST;
+  using FilterState = TrivialFilterState;
+
+  using Arc = typename FST1::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  TrivialComposeFilter(const FST1 &fst1, const FST2 &fst2,
+                       Matcher1 *matcher1 = nullptr,
+                       Matcher2 *matcher2 = nullptr)
+      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
+        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
+        fst1_(matcher1_->GetFst()),
+        fst2_(matcher2_->GetFst()) {}
+
+  TrivialComposeFilter(const TrivialComposeFilter<Matcher1, Matcher2> &filter,
+                       bool safe = false)
+      : matcher1_(filter.matcher1_->Copy(safe)),
+        matcher2_(filter.matcher2_->Copy(safe)),
+        fst1_(matcher1_->GetFst()),
+        fst2_(matcher2_->GetFst()) {}
+
+  FilterState Start() const { return FilterState(true); }
+
+  void SetState(StateId, StateId, const FilterState &) {}
+
+  FilterState FilterArc(Arc *, Arc *) const { return FilterState(true); }
+
+  void FilterFinal(Weight *, Weight *) const {}
+
+  Matcher1 *GetMatcher1() { return matcher1_.get(); }
+
+  Matcher2 *GetMatcher2() { return matcher2_.get(); }
+
+  uint64 Properties(uint64 props) const { return props; }
+
+ private:
+  std::unique_ptr<Matcher1> matcher1_;
+  std::unique_ptr<Matcher2> matcher2_;
+  const FST1 &fst1_;
+  const FST2 &fst2_;
+};
+
+// This filter requires epsilons on FST1 to be read before epsilons on FST2.
+template <class M1, class M2 /* = M1 */>
+class SequenceComposeFilter {
+ public:
+  using Matcher1 = M1;
+  using Matcher2 = M2;
+  using FST1 = typename M1::FST;
+  using FST2 = typename M2::FST;
+  using FilterState = CharFilterState;
+
+  using Arc = typename FST1::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  SequenceComposeFilter(const FST1 &fst1, const FST2 &fst2,
+                        Matcher1 *matcher1 = nullptr,
+                        Matcher2 *matcher2 = nullptr)
+      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
+        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
+        fst1_(matcher1_->GetFst()),
+        s1_(kNoStateId),
+        s2_(kNoStateId),
+        fs_(kNoStateId) {}
+
+  SequenceComposeFilter(const SequenceComposeFilter<Matcher1, Matcher2> &filter,
+                        bool safe = false)
+      : matcher1_(filter.matcher1_->Copy(safe)),
+        matcher2_(filter.matcher2_->Copy(safe)),
+        fst1_(matcher1_->GetFst()),
+        s1_(kNoStateId),
+        s2_(kNoStateId),
+        fs_(kNoStateId) {}
+
+  FilterState Start() const { return FilterState(0); }
+
+  void SetState(StateId s1, StateId s2, const FilterState &fs) {
+    if (s1_ == s1 && s2_ == s2 && fs == fs_) return;
+    s1_ = s1;
+    s2_ = s2;
+    fs_ = fs;
+    const auto na1 = internal::NumArcs(fst1_, s1);
+    const auto ne1 = internal::NumOutputEpsilons(fst1_, s1);
+    const bool fin1 = internal::Final(fst1_, s1) != Weight::Zero();
+    alleps1_ = na1 == ne1 && !fin1;
+    noeps1_ = ne1 == 0;
+  }
+
+  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
+    if (arc1->olabel == kNoLabel) {
+      return alleps1_ ? FilterState::NoState() : noeps1_ ? FilterState(0)
+                                                         : FilterState(1);
+    } else if (arc2->ilabel == kNoLabel) {
+      return fs_ != FilterState(0) ? FilterState::NoState() : FilterState(0);
+    } else {
+      return arc1->olabel == 0 ? FilterState::NoState() : FilterState(0);
+    }
+  }
+
+  void FilterFinal(Weight *, Weight *) const {}
+
+  Matcher1 *GetMatcher1() { return matcher1_.get(); }
+
+  Matcher2 *GetMatcher2() { return matcher2_.get(); }
+
+  uint64 Properties(uint64 props) const { return props; }
+
+ private:
+  std::unique_ptr<Matcher1> matcher1_;
+  std::unique_ptr<Matcher2> matcher2_;
+  const FST1 &fst1_;
+  StateId s1_;      // Current fst1_ state.
+  StateId s2_;      // Current fst2_ state.
+  FilterState fs_;  // Current filter state.
+  bool alleps1_;   // Only epsilons (and non-final) leaving s1_?
+  bool noeps1_;    // No epsilons leaving s1_?
+};
+
+// This filter requires epsilons on FST2 to be read before epsilons on FST1.
+template <class M1, class M2 /* = M1 */>
+class AltSequenceComposeFilter {
+ public:
+  using Matcher1 = M1;
+  using Matcher2 = M2;
+  using FST1 = typename M1::FST;
+  using FST2 = typename M2::FST;
+  using FilterState = CharFilterState;
+
+  using Arc = typename FST1::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  AltSequenceComposeFilter(const FST1 &fst1, const FST2 &fst2,
+                           Matcher1 *matcher1 = nullptr,
+                           Matcher2 *matcher2 = nullptr)
+      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
+        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
+        fst2_(matcher2_->GetFst()),
+        s1_(kNoStateId),
+        s2_(kNoStateId),
+        fs_(kNoStateId) {}
+
+  AltSequenceComposeFilter(
+      const AltSequenceComposeFilter<Matcher1, Matcher2> &filter,
+      bool safe = false)
+      : matcher1_(filter.matcher1_->Copy(safe)),
+        matcher2_(filter.matcher2_->Copy(safe)),
+        fst2_(matcher2_->GetFst()),
+        s1_(kNoStateId),
+        s2_(kNoStateId),
+        fs_(kNoStateId) {}
+
+  FilterState Start() const { return FilterState(0); }
+
+  void SetState(StateId s1, StateId s2, const FilterState &fs) {
+    if (s1_ == s1 && s2_ == s2 && fs == fs_) return;
+    s1_ = s1;
+    s2_ = s2;
+    fs_ = fs;
+    const auto na2 = internal::NumArcs(fst2_, s2);
+    const auto ne2 = internal::NumInputEpsilons(fst2_, s2);
+    const bool fin2 = internal::Final(fst2_, s2) != Weight::Zero();
+    alleps2_ = na2 == ne2 && !fin2;
+    noeps2_ = ne2 == 0;
+  }
+
+  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
+    if (arc2->ilabel == kNoLabel) {
+      return alleps2_ ? FilterState::NoState() : noeps2_ ? FilterState(0)
+                                                         : FilterState(1);
+    } else if (arc1->olabel == kNoLabel) {
+      return fs_ == FilterState(1) ? FilterState::NoState() : FilterState(0);
+    } else {
+      return arc1->olabel == 0 ? FilterState::NoState() : FilterState(0);
+    }
+  }
+
+  void FilterFinal(Weight *, Weight *) const {}
+
+  Matcher1 *GetMatcher1() { return matcher1_.get(); }
+
+  Matcher2 *GetMatcher2() { return matcher2_.get(); }
+
+  uint64 Properties(uint64 props) const { return props; }
+
+ private:
+  std::unique_ptr<Matcher1> matcher1_;
+  std::unique_ptr<Matcher2> matcher2_;
+  const FST2 &fst2_;
+  StateId s1_;      // Current fst1_ state.
+  StateId s2_;      // Current fst2_ state.
+  FilterState fs_;  // Current filter state.
+  bool alleps2_;    // Only epsilons (and non-final) leaving s2_?
+  bool noeps2_;     // No epsilons leaving s2_?
+};
+
+// This filter requires epsilons on FST1 to be matched with epsilons on FST2
+// whenever possible. (Template arg default declared in fst-decl.h.)
+template <class M1, class M2 /* = M1 */>
+class MatchComposeFilter {
+ public:
+  using Matcher1 = M1;
+  using Matcher2 = M2;
+  using FST1 = typename M1::FST;
+  using FST2 = typename M2::FST;
+  using FilterState = CharFilterState;
+
+  using Arc = typename FST1::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  MatchComposeFilter(const FST1 &fst1, const FST2 &fst2,
+                     Matcher1 *matcher1 = nullptr, Matcher2 *matcher2 = nullptr)
+      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
+        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
+        fst1_(matcher1_->GetFst()),
+        fst2_(matcher2_->GetFst()),
+        s1_(kNoStateId),
+        s2_(kNoStateId),
+        fs_(kNoStateId) {}
+
+  MatchComposeFilter(const MatchComposeFilter<Matcher1, Matcher2> &filter,
+                     bool safe = false)
+      : matcher1_(filter.matcher1_->Copy(safe)),
+        matcher2_(filter.matcher2_->Copy(safe)),
+        fst1_(matcher1_->GetFst()),
+        fst2_(matcher2_->GetFst()),
+        s1_(kNoStateId),
+        s2_(kNoStateId),
+        fs_(kNoStateId) {}
+
+  FilterState Start() const { return FilterState(0); }
+
+  void SetState(StateId s1, StateId s2, const FilterState &fs) {
+    if (s1_ == s1 && s2_ == s2 && fs == fs_) return;
+    s1_ = s1;
+    s2_ = s2;
+    fs_ = fs;
+    size_t na1 = internal::NumArcs(fst1_, s1);
+    size_t ne1 = internal::NumOutputEpsilons(fst1_, s1);
+    bool f1 = internal::Final(fst1_, s1) != Weight::Zero();
+    alleps1_ = na1 == ne1 && !f1;
+    noeps1_ = ne1 == 0;
+    size_t na2 = internal::NumArcs(fst2_, s2);
+    size_t ne2 = internal::NumInputEpsilons(fst2_, s2);
+    bool f2 = internal::Final(fst2_, s2) != Weight::Zero();
+    alleps2_ = na2 == ne2 && !f2;
+    noeps2_ = ne2 == 0;
+  }
+
+  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
+    if (arc2->ilabel == kNoLabel) {  // Epsilon in FST1.
+      return fs_ == FilterState(0)
+                 ? (noeps2_
+                        ? FilterState(0)
+                        : (alleps2_ ? FilterState::NoState() : FilterState(1)))
+                 : (fs_ == FilterState(1) ? FilterState(1)
+                                          : FilterState::NoState());
+    } else if (arc1->olabel == kNoLabel) {  // Epsilon in FST2.
+      return fs_ == FilterState(0)
+                 ? (noeps1_
+                        ? FilterState(0)
+                        : (alleps1_ ? FilterState::NoState() : FilterState(2)))
+                 : (fs_ == FilterState(2) ? FilterState(2)
+                                          : FilterState::NoState());
+    } else if (arc1->olabel == 0) {  // Epsilon in both.
+      return fs_ == FilterState(0) ? FilterState(0) : FilterState::NoState();
+    } else {  // Both are non-epsilons.
+      return FilterState(0);
+    }
+  }
+
+  void FilterFinal(Weight *, Weight *) const {}
+
+  Matcher1 *GetMatcher1() { return matcher1_.get(); }
+
+  Matcher2 *GetMatcher2() { return matcher2_.get(); }
+
+  uint64 Properties(uint64 props) const { return props; }
+
+ private:
+  std::unique_ptr<Matcher1> matcher1_;
+  std::unique_ptr<Matcher2> matcher2_;
+  const FST1 &fst1_;
+  const FST2 &fst2_;
+  StateId s1_;      // Current fst1_ state.
+  StateId s2_;      // Current fst2_ state.
+  FilterState fs_;  // Current filter state ID.
+  bool alleps1_;    // Only epsilson (and non-final) leaving s1?
+  bool alleps2_;    // Only epsilons (and non-final) leaving s2?
+  bool noeps1_;     // No epsilons leaving s1?
+  bool noeps2_;     // No epsilons leaving s2?
+};
+
+// This filter disallows matching epsilons on FST1 with epsilons on FST2,
+// but allows all other matches, potentially resulting in redundant
+// epsilon paths. The use of this filter gives correct results iff one of the
+// following conditions hold:
+//
+//  (1) The semiring is idempotent,
+//  (2) the first FST is output-epsilon free, or
+//  (3) the second FST is input-epsilon free.
+//
+// For (1), redundant epsilon paths may be created but won't hurt correctness.
+// For (2) and (3), no redundant paths are created.
+template <class M1, class M2 /* = M1 */>
+class NoMatchComposeFilter {
+ public:
+  using Matcher1 = M1;
+  using Matcher2 = M2;
+  using FST1 = typename M1::FST;
+  using FST2 = typename M2::FST;
+  using FilterState = TrivialFilterState;
+
+  using Arc = typename FST1::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  NoMatchComposeFilter(const FST1 &fst1, const FST2 &fst2,
+                       Matcher1 *matcher1 = nullptr,
+                       Matcher2 *matcher2 = nullptr)
+      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
+        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
+        fst1_(matcher1_->GetFst()),
+        fst2_(matcher2_->GetFst()) {}
+
+  NoMatchComposeFilter(const NoMatchComposeFilter<Matcher1, Matcher2> &filter,
+                       bool safe = false)
+      : matcher1_(filter.matcher1_->Copy(safe)),
+        matcher2_(filter.matcher2_->Copy(safe)),
+        fst1_(matcher1_->GetFst()),
+        fst2_(matcher2_->GetFst()) {}
+
+  FilterState Start() const { return FilterState(true); }
+
+  void SetState(StateId, StateId, const FilterState &) {}
+
+  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
+    return FilterState(arc1->olabel != 0 || arc2->ilabel != 0);
+  }
+
+  void FilterFinal(Weight *, Weight *) const {}
+
+  Matcher1 *GetMatcher1() { return matcher1_.get(); }
+
+  Matcher2 *GetMatcher2() { return matcher2_.get(); }
+
+  uint64 Properties(uint64 props) const { return props; }
+
+ private:
+  std::unique_ptr<Matcher1> matcher1_;
+  std::unique_ptr<Matcher2> matcher2_;
+  const FST1 &fst1_;
+  const FST2 &fst2_;
+};
+
+// This filter works with the MultiEpsMatcher to determine if multi-epsilons are
+// preserved in the composition output (rather than rewritten as 0) and
+// ensures correct properties.
+template <class Filter>
+class MultiEpsFilter {
+ public:
+  using Matcher1 = typename Filter::Matcher1;
+  using Matcher2 = typename Filter::Matcher2;
+  using FST1 = typename Filter::FST1;
+  using FST2 = typename Filter::FST2;
+  using FilterState = typename Filter::FilterState;
+
+  using Arc = typename Filter::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  MultiEpsFilter(const FST1 &fst1, const FST2 &fst2,
+                 Matcher1 *matcher1 = nullptr, Matcher2 *matcher2 = nullptr,
+                 bool keep_multi_eps = false)
+      : filter_(fst1, fst2, matcher1, matcher2),
+        keep_multi_eps_(keep_multi_eps) {}
+
+  MultiEpsFilter(const MultiEpsFilter<Filter> &filter, bool safe = false)
+      : filter_(filter.filter_, safe),
+        keep_multi_eps_(filter.keep_multi_eps_) {}
+
+  FilterState Start() const { return filter_.Start(); }
+
+  void SetState(StateId s1, StateId s2, const FilterState &fs) {
+    return filter_.SetState(s1, s2, fs);
+  }
+
+  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
+    const auto fs = filter_.FilterArc(arc1, arc2);
+    if (keep_multi_eps_) {
+      if (arc1->olabel == kNoLabel) arc1->ilabel = arc2->ilabel;
+      if (arc2->ilabel == kNoLabel) arc2->olabel = arc1->olabel;
+    }
+    return fs;
+  }
+
+  void FilterFinal(Weight *w1, Weight *w2) const {
+    return filter_.FilterFinal(w1, w2);
+  }
+
+  Matcher1 *GetMatcher1() { return filter_.GetMatcher1(); }
+
+  Matcher2 *GetMatcher2() { return filter_.GetMatcher2(); }
+
+  uint64 Properties(uint64 iprops) const {
+    const auto oprops = filter_.Properties(iprops);
+    return oprops & kILabelInvariantProperties & kOLabelInvariantProperties;
+  }
+
+ private:
+  Filter filter_;
+  bool keep_multi_eps_;
+};
+
+}  // namespace fst
+
+#endif  // FST_COMPOSE_FILTER_H_
diff --git a/projects/llm_framework/include/fst/compose.h b/projects/llm_framework/include/fst/compose.h
new file mode 100644
index 00000000..1066d097
--- /dev/null
+++ b/projects/llm_framework/include/fst/compose.h
@@ -0,0 +1,1035 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to compute the composition of two FSTs.
+
+#ifndef FST_COMPOSE_H_
+#define FST_COMPOSE_H_
+
+#include <algorithm>
+#include <memory>
+#include <utility>
+
+#include <fst/log.h>
+
+#include <fst/cache.h>
+#include <fst/compose-filter.h>
+#include <fst/fst-decl.h>  // For optional argument declarations
+#include <fst/lookahead-filter.h>
+#include <fst/matcher.h>
+#include <fst/state-table.h>
+#include <fst/test-properties.h>
+
+
+namespace fst {
+
+// Delayed composition options templated on the arc type, the matcher,
+// the composition filter, and the composition state table.  By
+// default, the matchers, filter, and state table are constructed by
+// composition. If set below, the user can instead pass in these
+// objects; in that case, ComposeFst takes their ownership. This
+// version controls composition implemented between generic Fst<Arc>
+// types and a shared matcher type M for Fst<Arc>. This should be
+// adequate for most applications, giving a reasonable tradeoff
+// between efficiency and code sharing (but see ComposeFstImplOptions).
+template <class Arc, class M = Matcher<Fst<Arc>>,
+          class Filter = SequenceComposeFilter<M>,
+          class StateTable =
+              GenericComposeStateTable<Arc, typename Filter::FilterState>>
+struct ComposeFstOptions : public CacheOptions {
+  M *matcher1;              // FST1 matcher.
+  M *matcher2;              // FST2 matcher.
+  Filter *filter;           // Composition filter.
+  StateTable *state_table;  // Composition state table.
+
+  explicit ComposeFstOptions(const CacheOptions &opts = CacheOptions(),
+                             M *matcher1 = nullptr, M *matcher2 = nullptr,
+                             Filter *filter = nullptr,
+                             StateTable *state_table = nullptr)
+      : CacheOptions(opts),
+        matcher1(matcher1),
+        matcher2(matcher2),
+        filter(filter),
+        state_table(state_table) {}
+};
+
+// Forward declaration of ComposeFstMatcher.
+template <class C, class F, class T>
+class ComposeFstMatcher;
+
+// Delayed composition options templated on the two matcher types, the
+// composition filter, the composition state table and the cache store. By
+// default, the matchers, filter, state table and cache store are constructed
+// by composition. If set below, the user can instead pass in these objects; in
+// that case, ComposeFst takes their ownership. This version controls
+// composition implemented using arbitrary matchers (of the same arc type but
+// otherwise arbitrary FST type). The user must ensure the matchers are
+// compatible. These options permit the most efficient use, but shares the
+// least code. This is for advanced use only in the most demanding or
+// specialized applications that can benefit from it; otherwise, prefer
+// ComposeFstOptions).
+template <class M1, class M2, class Filter = SequenceComposeFilter<M1, M2>,
+          class StateTable = GenericComposeStateTable<
+              typename M1::Arc, typename Filter::FilterState>,
+          class CacheStore = DefaultCacheStore<typename M1::Arc>>
+struct ComposeFstImplOptions : public CacheImplOptions<CacheStore> {
+  M1 *matcher1;    // FST1 matcher (see matcher.h)....
+  M2 *matcher2;    // FST2 matcher.
+  Filter *filter;  // Composition filter (see compose-filter.h).
+  StateTable
+    *state_table;        // Composition state table (see compose-state-table.h).
+  bool own_state_table;   // ComposeFstImpl takes ownership of 'state_table'?
+  bool allow_noncommute;  // Allow non-commutative weights
+
+  explicit ComposeFstImplOptions(const CacheOptions &opts,
+                                 M1 *matcher1 = nullptr, M2 *matcher2 = nullptr,
+                                 Filter *filter = nullptr,
+                                 StateTable *state_table = nullptr)
+      : CacheImplOptions<CacheStore>(opts),
+        matcher1(matcher1),
+        matcher2(matcher2),
+        filter(filter),
+        state_table(state_table),
+        own_state_table(true),
+        allow_noncommute(false) {}
+
+  explicit ComposeFstImplOptions(const CacheImplOptions<CacheStore> &opts,
+                                 M1 *matcher1 = nullptr, M2 *matcher2 = nullptr,
+                                 Filter *filter = nullptr,
+                                 StateTable *state_table = nullptr)
+      : CacheImplOptions<CacheStore>(opts),
+        matcher1(matcher1),
+        matcher2(matcher2),
+        filter(filter),
+        state_table(state_table),
+        own_state_table(true),
+        allow_noncommute(false) {}
+
+  ComposeFstImplOptions()
+      : matcher1(nullptr),
+        matcher2(nullptr),
+        filter(nullptr),
+        state_table(nullptr),
+        own_state_table(true),
+        allow_noncommute(false) {}
+};
+
+namespace internal {
+
+// Implementation of delayed composition. This base class is common to the
+// variants with different matchers, composition filters and state tables.
+template <class Arc, class CacheStore = DefaultCacheStore<Arc>,
+          class F = ComposeFst<Arc, CacheStore>>
+class ComposeFstImplBase
+    : public CacheBaseImpl<typename CacheStore::State, CacheStore> {
+ public:
+  using FST = F;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using State = typename CacheStore::State;
+  using CacheImpl = CacheBaseImpl<State, CacheStore>;
+
+  using FstImpl<Arc>::SetType;
+  using FstImpl<Arc>::SetProperties;
+  using FstImpl<Arc>::Properties;
+  using FstImpl<Arc>::SetInputSymbols;
+  using FstImpl<Arc>::SetOutputSymbols;
+
+  using CacheImpl::HasStart;
+  using CacheImpl::HasFinal;
+  using CacheImpl::HasArcs;
+  using CacheImpl::SetFinal;
+  using CacheImpl::SetStart;
+
+  ComposeFstImplBase(const CacheImplOptions<CacheStore> &opts)
+      : CacheImpl(opts) {}
+
+  ComposeFstImplBase(const CacheOptions &opts) : CacheImpl(opts) {}
+
+  ComposeFstImplBase(const ComposeFstImplBase &impl) : CacheImpl(impl, true) {
+    SetType(impl.Type());
+    SetProperties(impl.Properties(), kCopyProperties);
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+  }
+
+  virtual ComposeFstImplBase *Copy() const = 0;
+
+  ~ComposeFstImplBase() override {}
+
+  StateId Start() {
+    if (!HasStart()) {
+      const auto start = ComputeStart();
+      if (start != kNoStateId) SetStart(start);
+    }
+    return CacheImpl::Start();
+  }
+
+  Weight Final(StateId s) {
+    if (!HasFinal(s)) SetFinal(s, ComputeFinal(s));
+    return CacheImpl::Final(s);
+  }
+
+  virtual void Expand(StateId s) = 0;
+
+  size_t NumArcs(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl::NumArcs(s);
+  }
+
+  size_t NumInputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl::NumInputEpsilons(s);
+  }
+
+  size_t NumOutputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl::NumOutputEpsilons(s);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
+    if (!HasArcs(s)) Expand(s);
+    CacheImpl::InitArcIterator(s, data);
+  }
+
+  virtual MatcherBase<Arc> *InitMatcher(const F &fst,
+                                        MatchType match_type) const {
+    // Use the default matcher if no override is provided.
+    return nullptr;
+  }
+
+ protected:
+  virtual StateId ComputeStart() = 0;
+  virtual Weight ComputeFinal(StateId s) = 0;
+};
+
+// Implementation of delayed composition templated on the matchers (see
+// matcher.h), composition filter (see compose-filter.h) and the composition
+// state table (see compose-state-table.h).
+template <class CacheStore, class Filter, class StateTable>
+class ComposeFstImpl
+    : public ComposeFstImplBase<typename CacheStore::Arc, CacheStore> {
+ public:
+  using Matcher1 = typename Filter::Matcher1;
+  using Matcher2 = typename Filter::Matcher2;
+
+  using FST1 = typename Matcher1::FST;
+  using FST2 = typename Matcher2::FST;
+
+  using Arc = typename CacheStore::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FilterState = typename Filter::FilterState;
+  using State = typename CacheStore::State;
+
+  using CacheImpl = CacheBaseImpl<State, CacheStore>;
+
+  using StateTuple = typename StateTable::StateTuple;
+
+  friend class ComposeFstMatcher<CacheStore, Filter, StateTable>;
+
+  using FstImpl<Arc>::SetInputSymbols;
+  using FstImpl<Arc>::SetOutputSymbols;
+  using FstImpl<Arc>::SetType;
+  using FstImpl<Arc>::SetProperties;
+
+  template <class M1, class M2>
+  ComposeFstImpl(const FST1 &fst1, const FST2 &fst2,
+                 const ComposeFstImplOptions<M1, M2, Filter, StateTable,
+                                             CacheStore> &opts);
+
+  ComposeFstImpl(const ComposeFstImpl &impl)
+      : ComposeFstImplBase<Arc, CacheStore>(impl),
+        filter_(new Filter(*impl.filter_, true)),
+        matcher1_(filter_->GetMatcher1()),
+        matcher2_(filter_->GetMatcher2()),
+        fst1_(matcher1_->GetFst()),
+        fst2_(matcher2_->GetFst()),
+        state_table_(new StateTable(*impl.state_table_)),
+        own_state_table_(true),
+        match_type_(impl.match_type_) {}
+
+  ~ComposeFstImpl() override {
+    if (own_state_table_) delete state_table_;
+  }
+
+  ComposeFstImpl *Copy() const override { return new ComposeFstImpl(*this); }
+
+  uint64 Properties() const override { return Properties(kFstProperties); }
+
+  // Sets error if found, and returns other FST impl properties.
+  uint64 Properties(uint64 mask) const override {
+    if ((mask & kError) &&
+        (fst1_.Properties(kError, false) || fst2_.Properties(kError, false) ||
+         (matcher1_->Properties(0) & kError) ||
+         (matcher2_->Properties(0) & kError) |
+             (filter_->Properties(0) & kError) ||
+         state_table_->Error())) {
+      SetProperties(kError, kError);
+    }
+    return FstImpl<Arc>::Properties(mask);
+  }
+
+  // Arranges it so that the first arg to OrderedExpand is the Fst
+  // that will be matched on.
+  void Expand(StateId s) override {
+    const auto &tuple = state_table_->Tuple(s);
+    const auto s1 = tuple.StateId1();
+    const auto s2 = tuple.StateId2();
+    filter_->SetState(s1, s2, tuple.GetFilterState());
+    if (MatchInput(s1, s2)) {
+      OrderedExpand(s, fst2_, s2, fst1_, s1, matcher2_, true);
+    } else {
+      OrderedExpand(s, fst1_, s1, fst2_, s2, matcher1_, false);
+    }
+  }
+
+  const FST1 &GetFst1() const { return fst1_; }
+
+  const FST2 &GetFst2() const { return fst2_; }
+
+  const Matcher1 *GetMatcher1() const { return matcher1_; }
+
+  Matcher1 *GetMatcher1() { return matcher1_; }
+
+  const Matcher2 *GetMatcher2() const { return matcher2_; }
+
+  Matcher2 *GetMatcher2() { return matcher2_; }
+
+  const Filter *GetFilter() const { return filter_.get(); }
+
+  Filter *GetFilter() { return filter_.get(); }
+
+  const StateTable *GetStateTable() const { return state_table_; }
+
+  StateTable *GetStateTable() { return state_table_; }
+
+  MatcherBase<Arc> *InitMatcher(const ComposeFst<Arc, CacheStore> &fst,
+                                MatchType match_type) const override {
+    const auto test_props = match_type == MATCH_INPUT
+                                ? kFstProperties & ~kILabelInvariantProperties
+                                : kFstProperties & ~kOLabelInvariantProperties;
+    // If both matchers support 'match_type' and we have a guarantee that a
+    // call to 'filter_->FilterArc(arc1, arc2)' will not modify the ilabel of
+    // arc1 when MATCH_INPUT or the olabel or arc2 when MATCH_OUTPUT, then
+    // ComposeFstMatcher can be used.
+    if ((matcher1_->Type(false) == match_type) &&
+        (matcher2_->Type(false) == match_type) &&
+        (filter_->Properties(test_props) == test_props)) {
+      return new ComposeFstMatcher<
+        CacheStore, Filter, StateTable>(&fst, match_type);
+    }
+    return nullptr;
+  }
+
+ private:
+  // This does that actual matching of labels in the composition. The
+  // arguments are ordered so matching is called on state 'sa' of
+  // 'fsta' for each arc leaving state 'sb' of 'fstb'. The 'match_input' arg
+  // determines whether the input or output label of arcs at 'sb' is
+  // the one to match on.
+  template <class FST, class Matcher>
+  void OrderedExpand(StateId s, const Fst<Arc> &, StateId sa, const FST &fstb,
+                     StateId sb, Matcher *matchera, bool match_input) {
+    matchera->SetState(sa);
+    // First processes non-consuming symbols (e.g., epsilons) on FSTA.
+    const Arc loop(match_input ? 0 : kNoLabel, match_input ? kNoLabel : 0,
+                   Weight::One(), sb);
+    MatchArc(s, matchera, loop, match_input);
+    // Then processes matches on FSTB.
+    for (ArcIterator<FST> iterb(fstb, sb); !iterb.Done(); iterb.Next()) {
+      MatchArc(s, matchera, iterb.Value(), match_input);
+    }
+    CacheImpl::SetArcs(s);
+  }
+
+  // Matches a single transition from 'fstb' against 'fata' at 's'.
+  template <class Matcher>
+  void MatchArc(StateId s, Matcher *matchera, const Arc &arc,
+                bool match_input) {
+    if (matchera->Find(match_input ? arc.olabel : arc.ilabel)) {
+      for (; !matchera->Done(); matchera->Next()) {
+        auto arca = matchera->Value();
+        auto arcb = arc;
+        if (match_input) {
+          const auto &fs = filter_->FilterArc(&arcb, &arca);
+          if (fs != FilterState::NoState()) AddArc(s, arcb, arca, fs);
+        } else {
+          const auto &fs = filter_->FilterArc(&arca, &arcb);
+          if (fs != FilterState::NoState()) AddArc(s, arca, arcb, fs);
+        }
+      }
+    }
+  }
+
+  // Add a matching transition at 's'.
+  void AddArc(StateId s, const Arc &arc1, const Arc &arc2,
+              const FilterState &f) {
+    const StateTuple tuple(arc1.nextstate, arc2.nextstate, f);
+    CacheImpl::EmplaceArc(
+        s, arc1.ilabel, arc2.olabel, Times(arc1.weight, arc2.weight),
+        state_table_->FindState(tuple));
+  }
+
+  StateId ComputeStart() override {
+    const auto s1 = fst1_.Start();
+    if (s1 == kNoStateId) return kNoStateId;
+    const auto s2 = fst2_.Start();
+    if (s2 == kNoStateId) return kNoStateId;
+    const auto &fs = filter_->Start();
+    const StateTuple tuple(s1, s2, fs);
+    return state_table_->FindState(tuple);
+  }
+
+  Weight ComputeFinal(StateId s) override {
+    const auto &tuple = state_table_->Tuple(s);
+    const auto s1 = tuple.StateId1();
+    auto final1 = matcher1_->Final(s1);
+    if (final1 == Weight::Zero()) return final1;
+    const auto s2 = tuple.StateId2();
+    auto final2 = matcher2_->Final(s2);
+    if (final2 == Weight::Zero()) return final2;
+    filter_->SetState(s1, s2, tuple.GetFilterState());
+    filter_->FilterFinal(&final1, &final2);
+    return Times(final1, final2);
+  }
+
+  // Determines which side to match on per composition state.
+  bool MatchInput(StateId s1, StateId s2) {
+    switch (match_type_) {
+      case MATCH_INPUT:
+        return true;
+      case MATCH_OUTPUT:
+        return false;
+      default:  // MATCH_BOTH
+        const auto priority1 = matcher1_->Priority(s1);
+        const auto priority2 = matcher2_->Priority(s2);
+        if (priority1 == kRequirePriority && priority2 == kRequirePriority) {
+          FSTERROR() << "ComposeFst: Both sides can't require match";
+          SetProperties(kError, kError);
+          return true;
+        }
+        if (priority1 == kRequirePriority) return false;
+        if (priority2 == kRequirePriority) {
+          return true;
+        }
+        return priority1 <= priority2;
+    }
+  }
+
+  // Identifies and verifies the capabilities of the matcher to be used for
+  // composition.
+  void SetMatchType();
+
+  std::unique_ptr<Filter> filter_;
+  Matcher1 *matcher1_;  // Borrowed reference.
+  Matcher2 *matcher2_;  // Borrowed reference.
+  const FST1 &fst1_;
+  const FST2 &fst2_;
+  StateTable *state_table_;
+  bool own_state_table_;
+
+  MatchType match_type_;
+};
+
+template <class CacheStore, class Filter, class StateTable>
+template <class M1, class M2>
+ComposeFstImpl<CacheStore, Filter, StateTable>::ComposeFstImpl(
+    const FST1 &fst1, const FST2 &fst2,
+    const ComposeFstImplOptions<M1, M2, Filter, StateTable, CacheStore> &opts)
+    : ComposeFstImplBase<Arc, CacheStore>(opts),
+      filter_(opts.filter
+                  ? opts.filter
+                  : new Filter(fst1, fst2, opts.matcher1, opts.matcher2)),
+      matcher1_(filter_->GetMatcher1()),
+      matcher2_(filter_->GetMatcher2()),
+      fst1_(matcher1_->GetFst()),
+      fst2_(matcher2_->GetFst()),
+      state_table_(opts.state_table ? opts.state_table
+                                    : new StateTable(fst1_, fst2_)),
+      own_state_table_(opts.state_table ? opts.own_state_table : true) {
+  SetType("compose");
+  if (!CompatSymbols(fst2.InputSymbols(), fst1.OutputSymbols())) {
+    FSTERROR() << "ComposeFst: Output symbol table of 1st argument "
+               << "does not match input symbol table of 2nd argument";
+    SetProperties(kError, kError);
+  }
+  SetInputSymbols(fst1_.InputSymbols());
+  SetOutputSymbols(fst2_.OutputSymbols());
+  SetMatchType();
+  VLOG(2) << "ComposeFstImpl: Match type: " << match_type_;
+  if (match_type_ == MATCH_NONE) SetProperties(kError, kError);
+  const auto fprops1 = fst1.Properties(kFstProperties, false);
+  const auto fprops2 = fst2.Properties(kFstProperties, false);
+  const auto mprops1 = matcher1_->Properties(fprops1);
+  const auto mprops2 = matcher2_->Properties(fprops2);
+  const auto cprops = ComposeProperties(mprops1, mprops2);
+  SetProperties(filter_->Properties(cprops), kCopyProperties);
+  if (state_table_->Error()) SetProperties(kError, kError);
+}
+
+template <class CacheStore, class Filter, class StateTable>
+void ComposeFstImpl<CacheStore, Filter, StateTable>::SetMatchType() {
+  // Ensures any required matching is possible and known.
+  if ((matcher1_->Flags() & kRequireMatch) &&
+      matcher1_->Type(true) != MATCH_OUTPUT) {
+    FSTERROR() << "ComposeFst: 1st argument cannot perform required matching "
+               << "(sort?).";
+    match_type_ = MATCH_NONE;
+    return;
+  }
+  if ((matcher2_->Flags() & kRequireMatch) &&
+      matcher2_->Type(true) != MATCH_INPUT) {
+    FSTERROR() << "ComposeFst: 2nd argument cannot perform required matching "
+               << "(sort?).";
+    match_type_ = MATCH_NONE;
+    return;
+  }
+  // Finds which sides to match on (favoring minimal testing of capabilities).
+  const auto type1 = matcher1_->Type(false);
+  const auto type2 = matcher2_->Type(false);
+  if (type1 == MATCH_OUTPUT && type2 == MATCH_INPUT) {
+    match_type_ = MATCH_BOTH;
+  } else if (type1 == MATCH_OUTPUT) {
+    match_type_ = MATCH_OUTPUT;
+  } else if (type2 == MATCH_INPUT) {
+    match_type_ = MATCH_INPUT;
+  } else if (matcher1_->Type(true) == MATCH_OUTPUT) {
+    match_type_ = MATCH_OUTPUT;
+  } else if (matcher2_->Type(true) == MATCH_INPUT) {
+    match_type_ = MATCH_INPUT;
+  } else {
+    FSTERROR() << "ComposeFst: 1st argument cannot match on output labels "
+               << "and 2nd argument cannot match on input labels (sort?).";
+    match_type_ = MATCH_NONE;
+  }
+}
+
+}  // namespace internal
+
+// Computes the composition of two transducers. This version is a delayed FST.
+// If FST1 transduces string x to y with weight a and FST2 transduces y to z
+// with weight b, then their composition transduces string x to z with weight
+// Times(x, z).
+//
+// The output labels of the first transducer or the input labels of the second
+// transducer must be sorted (with the default matcher). The weights need to
+// form a commutative semiring (valid for TropicalWeight and LogWeight).
+//
+// Complexity:
+//
+// Assuming the first FST is unsorted and the second is sorted,
+//
+//   Time: O(v1 v2 d1 (log d2 + m2)),
+//   Space: O(v1 v2)
+//
+// where vi = # of states visited, di = maximum out-degree, and mi the
+// maximum multiplicity of the states visited, for the ith FST. Constant time
+// and space to visit an input state or arc is assumed and exclusive of caching.
+//
+// Caveats:
+// - ComposeFst does not trim its output (since it is a delayed operation).
+// - The efficiency of composition can be strongly affected by several factors:
+//   - the choice of which transducer is sorted - prefer sorting the FST
+//     that has the greater average out-degree.
+//   - the amount of non-determinism
+//   - the presence and location of epsilon transitions - avoid epsilon
+//     transitions on the output side of the first transducer or
+//     the input side of the second transducer or prefer placing
+//     them later in a path since they delay matching and can
+//     introduce non-coaccessible states and transitions.
+//
+// This class attaches interface to implementation and handles reference
+// counting, delegating most methods to ImplToFst. The CacheStore specifies the
+// cache store (default declared in fst-decl.h).
+template <class A, class CacheStore /* = DefaultCacheStore<A> */>
+class ComposeFst
+    : public ImplToFst<internal::ComposeFstImplBase<A, CacheStore>> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Store = CacheStore;
+  using State = typename CacheStore::State;
+
+  using Impl = internal::ComposeFstImplBase<A, CacheStore>;
+
+  friend class ArcIterator<ComposeFst<Arc, CacheStore>>;
+  friend class StateIterator<ComposeFst<Arc, CacheStore>>;
+  template <class, class, class> friend class ComposeFstMatcher;
+
+  // Compose specifying only caching options.
+  ComposeFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+             const CacheOptions &opts = CacheOptions())
+      : ImplToFst<Impl>(CreateBase(fst1, fst2, opts)) {}
+
+  // Compose specifying one shared matcher type M. Requires that the input FSTs
+  // and matcher FST types be Fst<Arc>. Recommended for best code-sharing and
+  // matcher compatiblity.
+  template <class Matcher, class Filter, class StateTuple>
+  ComposeFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+             const ComposeFstOptions<Arc, Matcher, Filter, StateTuple> &opts)
+      : ImplToFst<Impl>(CreateBase1(fst1, fst2, opts)) {}
+
+  // Compose specifying two matcher types Matcher1 and Matcher2. Requires input
+  // FST (of the same Arc type, but o.w. arbitrary) match the corresponding
+  // matcher FST types). Recommended only for advanced use in demanding or
+  // specialized applications due to potential code bloat and matcher
+  // incompatibilities.
+  template <class Matcher1, class Matcher2, class Filter, class StateTuple>
+  ComposeFst(const typename Matcher1::FST &fst1,
+             const typename Matcher2::FST &fst2,
+             const ComposeFstImplOptions<Matcher1, Matcher2, Filter, StateTuple,
+                                         CacheStore> &opts)
+      : ImplToFst<Impl>(CreateBase2(fst1, fst2, opts)) {}
+
+  // See Fst<>::Copy() for doc.
+  ComposeFst(const ComposeFst<A, CacheStore> &fst, bool safe = false)
+      : ImplToFst<Impl>(safe ? std::shared_ptr<Impl>(fst.GetImpl()->Copy())
+                             : fst.GetSharedImpl()) {}
+
+  // Get a copy of this ComposeFst. See Fst<>::Copy() for further doc.
+  ComposeFst<A, CacheStore> *Copy(bool safe = false) const override {
+    return new ComposeFst<A, CacheStore>(*this, safe);
+  }
+
+  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
+    GetMutableImpl()->InitArcIterator(s, data);
+  }
+
+  MatcherBase<Arc> *InitMatcher(MatchType match_type) const override {
+    return GetImpl()->InitMatcher(*this, match_type);
+  }
+
+ protected:
+  using ImplToFst<Impl>::GetImpl;
+  using ImplToFst<Impl>::GetMutableImpl;
+
+  explicit ComposeFst(std::shared_ptr<Impl> impl) : ImplToFst<Impl>(impl) {}
+
+  // Create compose implementation specifying two matcher types.
+  template <class Matcher1, class Matcher2, class Filter, class StateTuple>
+  static std::shared_ptr<Impl> CreateBase2(
+      const typename Matcher1::FST &fst1, const typename Matcher2::FST &fst2,
+      const ComposeFstImplOptions<Matcher1, Matcher2, Filter, StateTuple,
+                                  CacheStore> &opts) {
+    auto impl = std::make_shared<
+        internal::ComposeFstImpl<CacheStore, Filter, StateTuple>>(fst1, fst2,
+                                                                  opts);
+    if (!(Weight::Properties() & kCommutative) && !opts.allow_noncommute) {
+      const auto props1 = fst1.Properties(kUnweighted, true);
+      const auto props2 = fst2.Properties(kUnweighted, true);
+      if (!(props1 & kUnweighted) && !(props2 & kUnweighted)) {
+        FSTERROR() << "ComposeFst: Weights must be a commutative semiring: "
+                   << Weight::Type();
+        impl->SetProperties(kError, kError);
+      }
+    }
+    return impl;
+  }
+
+  // Create compose implementation specifying one matcher type; requires that
+  // input and matcher FST types be Fst<Arc>.
+  template <class Matcher, class Filter, class StateTuple>
+  static std::shared_ptr<Impl> CreateBase1(
+      const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+      const ComposeFstOptions<Arc, Matcher, Filter, StateTuple> &opts) {
+    ComposeFstImplOptions<Matcher, Matcher, Filter, StateTuple, CacheStore>
+        nopts(opts, opts.matcher1, opts.matcher2, opts.filter,
+              opts.state_table);
+    return CreateBase2(fst1, fst2, nopts);
+  }
+
+  // Create compose implementation specifying no matcher type.
+  static std::shared_ptr<Impl> CreateBase(const Fst<Arc> &fst1,
+                                          const Fst<Arc> &fst2,
+                                          const CacheOptions &opts) {
+    switch (LookAheadMatchType(fst1, fst2)) {  // Check for lookahead matchers
+      default:
+      case MATCH_NONE: {  // Default composition (no look-ahead).
+        ComposeFstOptions<Arc> nopts(opts);
+        return CreateBase1(fst1, fst2, nopts);
+      }
+      case MATCH_OUTPUT: {  // Lookahead on fst1.
+        using M = typename DefaultLookAhead<Arc, MATCH_OUTPUT>::FstMatcher;
+        using F = typename DefaultLookAhead<Arc, MATCH_OUTPUT>::ComposeFilter;
+        ComposeFstOptions<Arc, M, F> nopts(opts);
+        return CreateBase1(fst1, fst2, nopts);
+      }
+      case MATCH_INPUT: {  // Lookahead on fst2
+        using M = typename DefaultLookAhead<Arc, MATCH_INPUT>::FstMatcher;
+        using F = typename DefaultLookAhead<Arc, MATCH_INPUT>::ComposeFilter;
+        ComposeFstOptions<Arc, M, F> nopts(opts);
+        return CreateBase1(fst1, fst2, nopts);
+      }
+    }
+  }
+
+ private:
+  ComposeFst &operator=(const ComposeFst &fst) = delete;
+};
+
+// Specialization for ComposeFst.
+template <class Arc, class CacheStore>
+class StateIterator<ComposeFst<Arc, CacheStore>>
+    : public CacheStateIterator<ComposeFst<Arc, CacheStore>> {
+ public:
+  explicit StateIterator(const ComposeFst<Arc, CacheStore> &fst)
+      : CacheStateIterator<ComposeFst<Arc, CacheStore>>(fst,
+                                                        fst.GetMutableImpl()) {}
+};
+
+// Specialization for ComposeFst.
+template <class Arc, class CacheStore>
+class ArcIterator<ComposeFst<Arc, CacheStore>>
+    : public CacheArcIterator<ComposeFst<Arc, CacheStore>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const ComposeFst<Arc, CacheStore> &fst, StateId s)
+      : CacheArcIterator<ComposeFst<Arc, CacheStore>>(fst.GetMutableImpl(), s) {
+    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
+  }
+};
+
+template <class Arc, class CacheStore>
+inline void ComposeFst<Arc, CacheStore>::InitStateIterator(
+    StateIteratorData<Arc> *data) const {
+  data->base = new StateIterator<ComposeFst<Arc, CacheStore>>(*this);
+}
+
+// Specialized matcher for ComposeFst. Supports MATCH_INPUT or MATCH_OUTPUT,
+// iff the underlying matchers for the two FSTS being composed support
+// MATCH_INPUT or MATCH_OUTPUT, respectively.
+template <class CacheStore, class Filter, class StateTable>
+class ComposeFstMatcher : public MatcherBase<typename CacheStore::Arc> {
+ public:
+  using Arc = typename CacheStore::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Matcher1 = typename Filter::Matcher1;
+  using Matcher2 = typename Filter::Matcher2;
+  using FilterState = typename Filter::FilterState;
+
+  using StateTuple = typename StateTable::StateTuple;
+  using Impl = internal::ComposeFstImpl<CacheStore, Filter, StateTable>;
+
+  // The compose FST arg must match the filter and state table types.
+  // This makes a copy of the FST.
+  ComposeFstMatcher(const ComposeFst<Arc, CacheStore> &fst,
+                    MatchType match_type)
+      : owned_fst_(fst.Copy()),
+        fst_(*owned_fst_),
+        impl_(static_cast<const Impl *>(fst_.GetImpl())),
+        s_(kNoStateId),
+        match_type_(match_type),
+        matcher1_(impl_->matcher1_->Copy()),
+        matcher2_(impl_->matcher2_->Copy()),
+        current_loop_(false),
+        loop_(kNoLabel, 0, Weight::One(), kNoStateId) {
+    if (match_type_ == MATCH_OUTPUT) std::swap(loop_.ilabel, loop_.olabel);
+  }
+
+  // The compose FST arg must match the filter and state table types.
+  // This doesn't copy the FST (although it may copy components).
+  ComposeFstMatcher(const ComposeFst<Arc, CacheStore> *fst,
+                    MatchType match_type)
+      : fst_(*fst),
+        impl_(static_cast<const Impl *>(fst_.GetImpl())),
+        s_(kNoStateId),
+        match_type_(match_type),
+        matcher1_(impl_->matcher1_->Copy()),
+        matcher2_(impl_->matcher2_->Copy()),
+        current_loop_(false),
+        loop_(kNoLabel, 0, Weight::One(), kNoStateId) {
+    if (match_type_ == MATCH_OUTPUT) std::swap(loop_.ilabel, loop_.olabel);
+  }
+
+  // This makes a copy of the FST.
+  ComposeFstMatcher(
+      const ComposeFstMatcher<CacheStore, Filter, StateTable> &matcher,
+      bool safe = false)
+      : owned_fst_(matcher.fst_.Copy(safe)),
+        fst_(*owned_fst_),
+        impl_(static_cast<const Impl *>(fst_.GetImpl())),
+        s_(kNoStateId),
+        match_type_(matcher.match_type_),
+        matcher1_(matcher.matcher1_->Copy(safe)),
+        matcher2_(matcher.matcher2_->Copy(safe)),
+        current_loop_(false),
+        loop_(kNoLabel, 0, Weight::One(), kNoStateId) {
+    if (match_type_ == MATCH_OUTPUT) std::swap(loop_.ilabel, loop_.olabel);
+  }
+
+  ComposeFstMatcher<CacheStore, Filter, StateTable> *Copy(
+      bool safe = false) const override {
+    return new ComposeFstMatcher<CacheStore, Filter, StateTable>(*this, safe);
+  }
+
+  MatchType Type(bool test) const override {
+    if ((matcher1_->Type(test) == MATCH_NONE) ||
+        (matcher2_->Type(test) == MATCH_NONE)) {
+      return MATCH_NONE;
+    }
+    if (((matcher1_->Type(test) == MATCH_UNKNOWN) &&
+         (matcher2_->Type(test) == MATCH_UNKNOWN)) ||
+        ((matcher1_->Type(test) == MATCH_UNKNOWN) &&
+         (matcher2_->Type(test) == match_type_)) ||
+        ((matcher1_->Type(test) == match_type_) &&
+         (matcher2_->Type(test) == MATCH_UNKNOWN))) {
+      return MATCH_UNKNOWN;
+    }
+    if ((matcher1_->Type(test) == match_type_) &&
+        (matcher2_->Type(test) == match_type_)) {
+      return match_type_;
+    }
+    return MATCH_NONE;
+  }
+
+  const Fst<Arc> &GetFst() const override { return fst_; }
+
+  uint64 Properties(uint64 inprops) const override {
+    return inprops;
+  }
+
+  void SetState(StateId s) final {
+    if (s_ == s) return;
+    s_ = s;
+    const auto &tuple = impl_->state_table_->Tuple(s);
+    matcher1_->SetState(tuple.StateId1());
+    matcher2_->SetState(tuple.StateId2());
+    loop_.nextstate = s_;
+  }
+
+  bool Find(Label label) final {
+    bool found = false;
+    current_loop_ = false;
+    if (label == 0) {
+      current_loop_ = true;
+      found = true;
+    }
+    if (match_type_ == MATCH_INPUT) {
+      found = found || FindLabel(label, matcher1_.get(), matcher2_.get());
+    } else {  // match_type_ == MATCH_OUTPUT
+      found = found || FindLabel(label, matcher2_.get(), matcher1_.get());
+    }
+    return found;
+  }
+
+  bool Done() const final {
+    return !current_loop_ && matcher1_->Done() && matcher2_->Done();
+  }
+
+  const Arc &Value() const final { return current_loop_ ? loop_ : arc_; }
+
+  void Next() final {
+    if (current_loop_) {
+      current_loop_ = false;
+    } else if (match_type_ == MATCH_INPUT) {
+      FindNext(matcher1_.get(), matcher2_.get());
+    } else {  // match_type_ == MATCH_OUTPUT
+      FindNext(matcher2_.get(), matcher1_.get());
+    }
+  }
+
+  ssize_t Priority(StateId s) final { return fst_.NumArcs(s); }
+
+ private:
+  // Processes a match with the filter and creates resulting arc.
+  bool MatchArc(StateId s, Arc arc1,
+                Arc arc2) {  // FIXME(kbg): copy but not assignment.
+    const auto &fs = impl_->filter_->FilterArc(&arc1, &arc2);
+    if (fs == FilterState::NoState()) return false;
+    const StateTuple tuple(arc1.nextstate, arc2.nextstate, fs);
+    arc_.ilabel = arc1.ilabel;
+    arc_.olabel = arc2.olabel;
+    arc_.weight = Times(arc1.weight, arc2.weight);
+    arc_.nextstate = impl_->state_table_->FindState(tuple);
+    return true;
+  }
+
+  // Finds the first match allowed by the filter.
+  template <class MatcherA, class MatcherB>
+  bool FindLabel(Label label, MatcherA *matchera, MatcherB *matcherb) {
+    if (matchera->Find(label)) {
+      matcherb->Find(match_type_ == MATCH_INPUT ? matchera->Value().olabel
+                                                : matchera->Value().ilabel);
+      return FindNext(matchera, matcherb);
+    }
+    return false;
+  }
+
+  // Finds the next match allowed by the filter, returning true iff such a
+  // match is found.
+  template <class MatcherA, class MatcherB>
+  bool FindNext(MatcherA *matchera, MatcherB *matcherb) {
+    // State when entering this function:
+    // 'matchera' is pointed to a match x, y for label x, and a match for y was
+    // requested on 'matcherb'.
+    while (!matchera->Done() || !matcherb->Done()) {
+      if (matcherb->Done()) {
+        // If no more matches for y on 'matcherb', moves forward on 'matchera'
+        // until a match x, y' is found such that there is a match for y' on
+        // 'matcherb'.
+        matchera->Next();
+        while (!matchera->Done() &&
+               !matcherb->Find(match_type_ == MATCH_INPUT
+                                   ? matchera->Value().olabel
+                                   : matchera->Value().ilabel)) {
+          matchera->Next();
+        }
+      }
+      while (!matcherb->Done()) {
+        // 'matchera' is pointing to a match x, y' ('arca') and 'matcherb' is
+        // pointing to a match y', z' ('arcb'). If combining these two arcs is
+        // allowed by the filter (hence resulting in an arc x, z') return true.
+        // Position 'matcherb' on the next potential match for y' before
+        // returning.
+        const auto &arca = matchera->Value();
+        const auto &arcb = matcherb->Value();
+        // Position 'matcherb' on the next potential match for y'.
+        matcherb->Next();
+        // Returns true If combining these two arcs is allowed by the filter
+        // (hence resulting in an arc x, z'); otherwise consider next match
+        // for y' on 'matcherb'.
+        if (MatchArc(s_, match_type_ == MATCH_INPUT ? arca : arcb,
+                     match_type_ == MATCH_INPUT ? arcb : arca)) {
+          return true;
+        }
+      }
+    }
+    // Both 'matchera' and 'matcherb' are done, no more match to analyse.
+    return false;
+  }
+
+  std::unique_ptr<const ComposeFst<Arc, CacheStore>> owned_fst_;
+  const ComposeFst<Arc, CacheStore> &fst_;
+  const Impl *impl_;
+  StateId s_;
+  MatchType match_type_;
+  std::unique_ptr<Matcher1> matcher1_;
+  std::unique_ptr<Matcher2> matcher2_;
+  bool current_loop_;
+  Arc loop_;
+  Arc arc_;
+};
+
+// Useful alias when using StdArc.
+using StdComposeFst = ComposeFst<StdArc>;
+
+enum ComposeFilter {
+  AUTO_FILTER,
+  NULL_FILTER,
+  TRIVIAL_FILTER,
+  SEQUENCE_FILTER,
+  ALT_SEQUENCE_FILTER,
+  MATCH_FILTER,
+  NO_MATCH_FILTER
+};
+
+struct ComposeOptions {
+  bool connect;               // Connect output?
+  ComposeFilter filter_type;  // Pre-defined filter to use.
+
+  explicit ComposeOptions(bool connect = true,
+                          ComposeFilter filter_type = AUTO_FILTER)
+      : connect(connect), filter_type(filter_type) {}
+};
+
+// Computes the composition of two transducers. This version writes
+// the composed FST into a MutableFst. If FST1 transduces string x to
+// y with weight a and FST2 transduces y to z with weight b, then
+// their composition transduces string x to z with weight
+// Times(x, z).
+//
+// The output labels of the first transducer or the input labels of
+// the second transducer must be sorted.  The weights need to form a
+// commutative semiring (valid for TropicalWeight and LogWeight).
+//
+// Complexity:
+//
+// Assuming the first FST is unsorted and the second is sorted:
+//
+//   Time: O(V1 V2 D1 (log D2 + M2)),
+//   Space: O(V1 V2 D1 M2)
+//
+// where Vi = # of states, Di = maximum out-degree, and Mi is the maximum
+// multiplicity, for the ith FST.
+//
+// Caveats:
+//
+// - Compose trims its output.
+// - The efficiency of composition can be strongly affected by several factors:
+//   - the choice of which transducer is sorted - prefer sorting the FST
+//     that has the greater average out-degree.
+//   - the amount of non-determinism
+//   - the presence and location of epsilon transitions - avoid epsilon
+//     transitions on the output side of the first transducer or
+//     the input side of the second transducer or prefer placing
+//     them later in a path since they delay matching and can
+//     introduce non-coaccessible states and transitions.
+template <class Arc>
+void Compose(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
+             MutableFst<Arc> *ofst,
+             const ComposeOptions &opts = ComposeOptions()) {
+  using M = Matcher<Fst<Arc>>;
+  // In each case, we cache only the last state for fastest copy.
+  switch (opts.filter_type) {
+    case AUTO_FILTER: {
+      CacheOptions nopts;
+      nopts.gc_limit = 0;
+      *ofst = ComposeFst<Arc>(ifst1, ifst2, nopts);
+      break;
+    }
+    case NULL_FILTER: {
+      ComposeFstOptions<Arc, M, NullComposeFilter<M>> copts;
+      copts.gc_limit = 0;
+      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
+      break;
+    }
+    case SEQUENCE_FILTER: {
+      ComposeFstOptions<Arc, M, SequenceComposeFilter<M>> copts;
+      copts.gc_limit = 0;
+      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
+      break;
+    }
+    case ALT_SEQUENCE_FILTER: {
+      ComposeFstOptions<Arc, M, AltSequenceComposeFilter<M>> copts;
+      copts.gc_limit = 0;
+      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
+      break;
+    }
+    case MATCH_FILTER: {
+      ComposeFstOptions<Arc, M, MatchComposeFilter<M>> copts;
+      copts.gc_limit = 0;
+      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
+      break;
+    }
+    case NO_MATCH_FILTER: {
+      ComposeFstOptions<Arc, M, NoMatchComposeFilter<M>> copts;
+      copts.gc_limit = 0;
+      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
+      break;
+    }
+    case TRIVIAL_FILTER: {
+      ComposeFstOptions<Arc, M, TrivialComposeFilter<M>> copts;
+      copts.gc_limit = 0;
+      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
+      break;
+    }
+  }
+  if (opts.connect) Connect(ofst);
+}
+
+}  // namespace fst
+
+#endif  // FST_COMPOSE_H_
diff --git a/projects/llm_framework/include/fst/concat.h b/projects/llm_framework/include/fst/concat.h
new file mode 100644
index 00000000..74d22c22
--- /dev/null
+++ b/projects/llm_framework/include/fst/concat.h
@@ -0,0 +1,220 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes to compute the concatenation of two FSTs.
+
+#ifndef FST_CONCAT_H_
+#define FST_CONCAT_H_
+
+#include <algorithm>
+#include <vector>
+
+#include <fst/mutable-fst.h>
+#include <fst/rational.h>
+
+
+namespace fst {
+
+// Computes the concatenation (product) of two FSTs. If FST1 transduces string
+// x to y with weight a and FST2 transduces string w to v with weight b, then
+// their concatenation transduces string xw to yv with weight Times(a, b).
+//
+// This version modifies its MutableFst argument (in first position).
+//
+// Complexity:
+//
+//   Time: O(V1 + V2 + E2)
+//   Space: O(V1 + V2 + E2)
+//
+// where Vi is the number of states, and Ei is the number of arcs, of the ith
+// FST.
+template <class Arc>
+void Concat(MutableFst<Arc> *fst1, const Fst<Arc> &fst2) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  // Checks that the symbol table are compatible.
+  if (!CompatSymbols(fst1->InputSymbols(), fst2.InputSymbols()) ||
+      !CompatSymbols(fst1->OutputSymbols(), fst2.OutputSymbols())) {
+    FSTERROR() << "Concat: Input/output symbol tables of 1st argument "
+               << "does not match input/output symbol tables of 2nd argument";
+    fst1->SetProperties(kError, kError);
+    return;
+  }
+  const auto props1 = fst1->Properties(kFstProperties, false);
+  const auto props2 = fst2.Properties(kFstProperties, false);
+  const auto start1 = fst1->Start();
+  if (start1 == kNoStateId) {
+    if (props2 & kError) fst1->SetProperties(kError, kError);
+    return;
+  }
+  const auto numstates1 = fst1->NumStates();
+  if (fst2.Properties(kExpanded, false)) {
+    fst1->ReserveStates(numstates1 + CountStates(fst2));
+  }
+  for (StateIterator<Fst<Arc>> siter2(fst2); !siter2.Done(); siter2.Next()) {
+    const auto s1 = fst1->AddState();
+    const auto s2 = siter2.Value();
+    fst1->SetFinal(s1, fst2.Final(s2));
+    fst1->ReserveArcs(s1, fst2.NumArcs(s2));
+    for (ArcIterator<Fst<Arc>> aiter(fst2, s2); !aiter.Done(); aiter.Next()) {
+      auto arc = aiter.Value();
+      arc.nextstate += numstates1;
+      fst1->AddArc(s1, arc);
+    }
+  }
+  const auto start2 = fst2.Start();
+  for (StateId s1 = 0; s1 < numstates1; ++s1) {
+    const auto weight = fst1->Final(s1);
+    if (weight != Weight::Zero()) {
+      fst1->SetFinal(s1, Weight::Zero());
+      if (start2 != kNoStateId) {
+        fst1->AddArc(s1, Arc(0, 0, weight, start2 + numstates1));
+      }
+    }
+  }
+  if (start2 != kNoStateId) {
+    fst1->SetProperties(ConcatProperties(props1, props2), kFstProperties);
+  }
+}
+
+// Computes the concatentation of two FSTs.  This version modifies its
+// MutableFst argument (in second position).
+//
+// Complexity:
+//
+//   Time: O(V1 + E1)
+//   Space: O(V1 + E1)
+//
+// where Vi is the number of states, and Ei is the number of arcs, of the ith
+// FST.
+template <class Arc>
+void Concat(const Fst<Arc> &fst1, MutableFst<Arc> *fst2) {
+  using Weight = typename Arc::Weight;
+  // Checks that the symbol table are compatible.
+  if (!CompatSymbols(fst1.InputSymbols(), fst2->InputSymbols()) ||
+      !CompatSymbols(fst1.OutputSymbols(), fst2->OutputSymbols())) {
+    FSTERROR() << "Concat: Input/output symbol tables of 1st argument "
+               << "does not match input/output symbol tables of 2nd argument";
+    fst2->SetProperties(kError, kError);
+    return;
+  }
+  const auto props1 = fst1.Properties(kFstProperties, false);
+  const auto props2 = fst2->Properties(kFstProperties, false);
+  const auto start2 = fst2->Start();
+  if (start2 == kNoStateId) {
+    if (props1 & kError) fst2->SetProperties(kError, kError);
+    return;
+  }
+  const auto numstates2 = fst2->NumStates();
+  if (fst1.Properties(kExpanded, false)) {
+    fst2->ReserveStates(numstates2 + CountStates(fst1));
+  }
+  for (StateIterator<Fst<Arc>> siter(fst1); !siter.Done(); siter.Next()) {
+    const auto s1 = siter.Value();
+    const auto s2 = fst2->AddState();
+    const auto weight = fst1.Final(s1);
+    if (weight != Weight::Zero()) {
+      fst2->ReserveArcs(s2, fst1.NumArcs(s1) + 1);
+      fst2->AddArc(s2, Arc(0, 0, weight, start2));
+    } else {
+      fst2->ReserveArcs(s2, fst1.NumArcs(s1));
+    }
+    for (ArcIterator<Fst<Arc>> aiter(fst1, s1); !aiter.Done(); aiter.Next()) {
+      auto arc = aiter.Value();
+      arc.nextstate += numstates2;
+      fst2->AddArc(s2, arc);
+    }
+  }
+  const auto start1 = fst1.Start();
+  if (start1 != kNoStateId) {
+    fst2->SetStart(start1 + numstates2);
+    fst2->SetProperties(ConcatProperties(props1, props2), kFstProperties);
+  } else {
+    fst2->SetStart(fst2->AddState());
+  }
+}
+
+// Computes the concatentation of two FSTs. This version modifies its
+// RationalFst input (in first position).
+template <class Arc>
+void Concat(RationalFst<Arc> *fst1, const Fst<Arc> &fst2) {
+  fst1->GetMutableImpl()->AddConcat(fst2, true);
+}
+
+// Computes the concatentation of two FSTs. This version modifies its
+// RationalFst input (in second position).
+template <class Arc>
+void Concat(const Fst<Arc> &fst1, RationalFst<Arc> *fst2) {
+  fst2->GetMutableImpl()->AddConcat(fst1, false);
+}
+
+using ConcatFstOptions = RationalFstOptions;
+
+// Computes the concatenation (product) of two FSTs; this version is a delayed
+// FST. If FST1 transduces string x to y with weight a and FST2 transduces
+// string w to v with weight b, then their concatenation transduces string xw
+// to yv with Times(a, b).
+//
+// Complexity:
+//
+//   Time: O(v1 + e1 + v2 + e2),
+//   Space: O(v1 + v2)
+//
+// where vi is the number of states visited, and ei is the number of arcs
+// visited, of the ith FST. Constant time and space to visit an input state or
+// arc is assumed and exclusive of caching.
+template <class A>
+class ConcatFst : public RationalFst<A> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  ConcatFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {
+    GetMutableImpl()->InitConcat(fst1, fst2);
+  }
+
+  ConcatFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+            const ConcatFstOptions &opts)
+      : RationalFst<Arc>(opts) {
+    GetMutableImpl()->InitConcat(fst1, fst2);
+  }
+
+  // See Fst<>::Copy() for doc.
+  ConcatFst(const ConcatFst<Arc> &fst, bool safe = false)
+      : RationalFst<Arc>(fst, safe) {}
+
+  // Get a copy of this ConcatFst. See Fst<>::Copy() for further doc.
+  ConcatFst<Arc> *Copy(bool safe = false) const override {
+    return new ConcatFst<Arc>(*this, safe);
+  }
+
+ private:
+  using ImplToFst<internal::RationalFstImpl<Arc>>::GetImpl;
+  using ImplToFst<internal::RationalFstImpl<Arc>>::GetMutableImpl;
+};
+
+// Specialization for ConcatFst.
+template <class Arc>
+class StateIterator<ConcatFst<Arc>> : public StateIterator<RationalFst<Arc>> {
+ public:
+  explicit StateIterator(const ConcatFst<Arc> &fst)
+      : StateIterator<RationalFst<Arc>>(fst) {}
+};
+
+// Specialization for ConcatFst.
+template <class Arc>
+class ArcIterator<ConcatFst<Arc>> : public ArcIterator<RationalFst<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const ConcatFst<Arc> &fst, StateId s)
+      : ArcIterator<RationalFst<Arc>>(fst, s) {}
+};
+
+// Useful alias when using StdArc.
+using StdConcatFst = ConcatFst<StdArc>;
+
+}  // namespace fst
+
+#endif  // FST_CONCAT_H_
diff --git a/projects/llm_framework/include/fst/config.h b/projects/llm_framework/include/fst/config.h
new file mode 100644
index 00000000..32f2a653
--- /dev/null
+++ b/projects/llm_framework/include/fst/config.h
@@ -0,0 +1,3 @@
+// Windows-specific OpenFst config file
+// No dynamic registration.
+#define FST_NO_DYNAMIC_LINKING 1
diff --git a/projects/llm_framework/include/fst/config.h.in b/projects/llm_framework/include/fst/config.h.in
new file mode 100644
index 00000000..7815dfcd
--- /dev/null
+++ b/projects/llm_framework/include/fst/config.h.in
@@ -0,0 +1,11 @@
+// OpenFst config file 
+
+/* Define to 1 if you have the ICU library. */
+#undef HAVE_ICU
+
+/* Define to 1 if the system has the type `std::tr1::hash<long long
+   unsigned>'. */
+#define HAVE_STD__TR1__HASH_LONG_LONG_UNSIGNED_ 1
+
+/* Define to 1 if the system has the type `__gnu_cxx::slist<int>'. */
+#define HAVE___GNU_CXX__SLIST_INT_ 1
diff --git a/projects/llm_framework/include/fst/connect.h b/projects/llm_framework/include/fst/connect.h
new file mode 100644
index 00000000..4c989292
--- /dev/null
+++ b/projects/llm_framework/include/fst/connect.h
@@ -0,0 +1,323 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Classes and functions to remove unsuccessful paths from an FST.
+
+#ifndef FST_CONNECT_H_
+#define FST_CONNECT_H_
+
+#include <algorithm>
+#include <memory>
+#include <vector>
+
+#include <fst/dfs-visit.h>
+#include <fst/mutable-fst.h>
+#include <fst/union-find.h>
+
+
+namespace fst {
+
+// Finds and returns connected components. Use with Visit().
+template <class Arc>
+class CcVisitor {
+ public:
+  using Weight = typename Arc::Weight;
+  using StateId = typename Arc::StateId;
+
+  // cc[i]: connected component number for state i.
+  explicit CcVisitor(std::vector<StateId> *cc)
+      : comps_(new UnionFind<StateId>(0, kNoStateId)), cc_(cc), nstates_(0) {}
+
+  // comps: connected components equiv classes.
+  explicit CcVisitor(UnionFind<StateId> *comps)
+      : comps_(comps), cc_(nullptr), nstates_(0) {}
+
+  ~CcVisitor() {
+    if (cc_) delete comps_;
+  }
+
+  void InitVisit(const Fst<Arc> &fst) {}
+
+  bool InitState(StateId s, StateId root) {
+    ++nstates_;
+    if (comps_->FindSet(s) == kNoStateId) comps_->MakeSet(s);
+    return true;
+  }
+
+  bool WhiteArc(StateId s, const Arc &arc) {
+    comps_->MakeSet(arc.nextstate);
+    comps_->Union(s, arc.nextstate);
+    return true;
+  }
+
+  bool GreyArc(StateId s, const Arc &arc) {
+    comps_->Union(s, arc.nextstate);
+    return true;
+  }
+
+  bool BlackArc(StateId s, const Arc &arc) {
+    comps_->Union(s, arc.nextstate);
+    return true;
+  }
+
+  void FinishState(StateId s) {}
+
+  void FinishVisit() {
+    if (cc_) GetCcVector(cc_);
+  }
+
+  // Returns number of components.
+  // cc[i]: connected component number for state i.
+  int GetCcVector(std::vector<StateId> *cc) {
+    cc->clear();
+    cc->resize(nstates_, kNoStateId);
+    StateId ncomp = 0;
+    for (StateId s = 0; s < nstates_; ++s) {
+      const auto rep = comps_->FindSet(s);
+      auto &comp = (*cc)[rep];
+      if (comp == kNoStateId) {
+        comp = ncomp;
+        ++ncomp;
+      }
+      (*cc)[s] = comp;
+    }
+    return ncomp;
+  }
+
+ private:
+  UnionFind<StateId> *comps_;  // Components.
+  std::vector<StateId> *cc_;   // State's cc number.
+  StateId nstates_;            // State count.
+};
+
+// Finds and returns strongly-connected components, accessible and
+// coaccessible states and related properties. Uses Tarjan's single
+// DFS SCC algorithm (see Aho, et al, "Design and Analysis of Computer
+// Algorithms", 189pp). Use with DfsVisit();
+template <class Arc>
+class SccVisitor {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  // scc[i]: strongly-connected component number for state i.
+  //   SCC numbers will be in topological order for acyclic input.
+  // access[i]: accessibility of state i.
+  // coaccess[i]: coaccessibility of state i.
+  // Any of above can be NULL.
+  // props: related property bits (cyclicity, initial cyclicity,
+  //   accessibility, coaccessibility) set/cleared (o.w. unchanged).
+  SccVisitor(std::vector<StateId> *scc, std::vector<bool> *access,
+             std::vector<bool> *coaccess, uint64 *props)
+      : scc_(scc), access_(access), coaccess_(coaccess), props_(props) {}
+  explicit SccVisitor(uint64 *props)
+      : scc_(nullptr), access_(nullptr), coaccess_(nullptr), props_(props) {}
+
+  void InitVisit(const Fst<Arc> &fst);
+
+  bool InitState(StateId s, StateId root);
+
+  bool TreeArc(StateId s, const Arc &arc) { return true; }
+
+  bool BackArc(StateId s, const Arc &arc) {
+    const auto t = arc.nextstate;
+    if ((*dfnumber_)[t] < (*lowlink_)[s]) (*lowlink_)[s] = (*dfnumber_)[t];
+    if ((*coaccess_)[t]) (*coaccess_)[s] = true;
+    *props_ |= kCyclic;
+    *props_ &= ~kAcyclic;
+    if (t == start_) {
+      *props_ |= kInitialCyclic;
+      *props_ &= ~kInitialAcyclic;
+    }
+    return true;
+  }
+
+  bool ForwardOrCrossArc(StateId s, const Arc &arc) {
+    const auto t = arc.nextstate;
+    if ((*dfnumber_)[t] < (*dfnumber_)[s] /* cross edge */ && (*onstack_)[t] &&
+        (*dfnumber_)[t] < (*lowlink_)[s]) {
+      (*lowlink_)[s] = (*dfnumber_)[t];
+    }
+    if ((*coaccess_)[t]) (*coaccess_)[s] = true;
+    return true;
+  }
+
+  // Last argument always ignored, but required by the interface.
+  void FinishState(StateId state, StateId p, const Arc *);
+
+  void FinishVisit() {
+    // Numbers SCCs in topological order when acyclic.
+    if (scc_) {
+      for (size_t s = 0; s < scc_->size(); ++s) {
+        (*scc_)[s] = nscc_ - 1 - (*scc_)[s];
+      }
+    }
+    if (coaccess_internal_) delete coaccess_;
+    dfnumber_.reset();
+    lowlink_.reset();
+    onstack_.reset();
+    scc_stack_.reset();
+  }
+
+ private:
+  std::vector<StateId> *scc_;    // State's scc number.
+  std::vector<bool> *access_;    // State's accessibility.
+  std::vector<bool> *coaccess_;  // State's coaccessibility.
+  uint64 *props_;
+  const Fst<Arc> *fst_;
+  StateId start_;
+  StateId nstates_;  // State count.
+  StateId nscc_;     // SCC count.
+  bool coaccess_internal_;
+  std::unique_ptr<std::vector<StateId>> dfnumber_;  // State discovery times.
+  std::unique_ptr<std::vector<StateId>>
+      lowlink_;  // lowlink[state] == dfnumber[state] => SCC root
+  std::unique_ptr<std::vector<bool>> onstack_;  // Is a state on the SCC stack?
+  std::unique_ptr<std::vector<StateId>>
+      scc_stack_;  // SCC stack, with random access.
+};
+
+template <class Arc>
+inline void SccVisitor<Arc>::InitVisit(const Fst<Arc> &fst) {
+  if (scc_) scc_->clear();
+  if (access_) access_->clear();
+  if (coaccess_) {
+    coaccess_->clear();
+    coaccess_internal_ = false;
+  } else {
+    coaccess_ = new std::vector<bool>;
+    coaccess_internal_ = true;
+  }
+  *props_ |= kAcyclic | kInitialAcyclic | kAccessible | kCoAccessible;
+  *props_ &= ~(kCyclic | kInitialCyclic | kNotAccessible | kNotCoAccessible);
+  fst_ = &fst;
+  start_ = fst.Start();
+  nstates_ = 0;
+  nscc_ = 0;
+  dfnumber_.reset(new std::vector<StateId>());
+  lowlink_.reset(new std::vector<StateId>());
+  onstack_.reset(new std::vector<bool>());
+  scc_stack_.reset(new std::vector<StateId>());
+}
+
+template <class Arc>
+inline bool SccVisitor<Arc>::InitState(StateId s, StateId root) {
+  scc_stack_->push_back(s);
+  if (static_cast<StateId>(dfnumber_->size()) <= s) {
+    if (scc_) scc_->resize(s + 1, -1);
+    if (access_) access_->resize(s + 1, false);
+    coaccess_->resize(s + 1, false);
+    dfnumber_->resize(s + 1, -1);
+    lowlink_->resize(s + 1, -1);
+    onstack_->resize(s + 1, false);
+  }
+  (*dfnumber_)[s] = nstates_;
+  (*lowlink_)[s] = nstates_;
+  (*onstack_)[s] = true;
+  if (root == start_) {
+    if (access_) (*access_)[s] = true;
+  } else {
+    if (access_) (*access_)[s] = false;
+    *props_ |= kNotAccessible;
+    *props_ &= ~kAccessible;
+  }
+  ++nstates_;
+  return true;
+}
+
+template <class Arc>
+inline void SccVisitor<Arc>::FinishState(StateId s, StateId p, const Arc *) {
+  if (fst_->Final(s) != Weight::Zero()) (*coaccess_)[s] = true;
+  if ((*dfnumber_)[s] == (*lowlink_)[s]) {  // Root of new SCC.
+    bool scc_coaccess = false;
+    auto i = scc_stack_->size();
+    StateId t;
+    do {
+      t = (*scc_stack_)[--i];
+      if ((*coaccess_)[t]) scc_coaccess = true;
+    } while (s != t);
+    do {
+      t = scc_stack_->back();
+      if (scc_) (*scc_)[t] = nscc_;
+      if (scc_coaccess) (*coaccess_)[t] = true;
+      (*onstack_)[t] = false;
+      scc_stack_->pop_back();
+    } while (s != t);
+    if (!scc_coaccess) {
+      *props_ |= kNotCoAccessible;
+      *props_ &= ~kCoAccessible;
+    }
+    ++nscc_;
+  }
+  if (p != kNoStateId) {
+    if ((*coaccess_)[s]) (*coaccess_)[p] = true;
+    if ((*lowlink_)[s] < (*lowlink_)[p]) (*lowlink_)[p] = (*lowlink_)[s];
+  }
+}
+
+// Trims an FST, removing states and arcs that are not on successful paths.
+// This version modifies its input.
+//
+// Complexity:
+//
+//   Time:  O(V + E)
+//   Space: O(V + E)
+//
+// where V = # of states and E = # of arcs.
+template <class Arc>
+void Connect(MutableFst<Arc> *fst) {
+  using StateId = typename Arc::StateId;
+  std::vector<bool> access;
+  std::vector<bool> coaccess;
+  uint64 props = 0;
+  SccVisitor<Arc> scc_visitor(nullptr, &access, &coaccess, &props);
+  DfsVisit(*fst, &scc_visitor);
+  std::vector<StateId> dstates;
+  dstates.reserve(access.size());
+  for (StateId s = 0; s < access.size(); ++s) {
+    if (!access[s] || !coaccess[s]) dstates.push_back(s);
+  }
+  fst->DeleteStates(dstates);
+  fst->SetProperties(kAccessible | kCoAccessible, kAccessible | kCoAccessible);
+}
+
+// Returns an acyclic FST where each SCC in the input FST has been condensed to
+// a single state with transitions between SCCs retained and within SCCs
+// dropped. Also populates 'scc' with a mapping from input to output states.
+template <class Arc>
+void Condense(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
+              std::vector<typename Arc::StateId> *scc) {
+  using StateId = typename Arc::StateId;
+  ofst->DeleteStates();
+  uint64 props = 0;
+  SccVisitor<Arc> scc_visitor(scc, nullptr, nullptr, &props);
+  DfsVisit(ifst, &scc_visitor);
+  const auto iter = std::max_element(scc->cbegin(), scc->cend());
+  if (iter == scc->cend()) return;
+  const StateId num_condensed_states = 1 + *iter;
+  ofst->ReserveStates(num_condensed_states);
+  for (StateId c = 0; c < num_condensed_states; ++c) {
+    ofst->AddState();
+  }
+  for (StateId s = 0; s < scc->size(); ++s) {
+    const auto c = (*scc)[s];
+    if (s == ifst.Start()) ofst->SetStart(c);
+    const auto weight = ifst.Final(s);
+    if (weight != Arc::Weight::Zero())
+      ofst->SetFinal(c, Plus(ofst->Final(c), weight));
+    for (ArcIterator<Fst<Arc>> aiter(ifst, s); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      const auto nextc = (*scc)[arc.nextstate];
+      if (nextc != c) {
+        Arc condensed_arc = arc;
+        condensed_arc.nextstate = nextc;
+        ofst->AddArc(c, std::move(condensed_arc));
+      }
+    }
+  }
+  ofst->SetProperties(kAcyclic | kInitialAcyclic, kAcyclic | kInitialAcyclic);
+}
+
+}  // namespace fst
+
+#endif  // FST_CONNECT_H_
diff --git a/projects/llm_framework/include/fst/const-fst.h b/projects/llm_framework/include/fst/const-fst.h
new file mode 100644
index 00000000..09c81c7b
--- /dev/null
+++ b/projects/llm_framework/include/fst/const-fst.h
@@ -0,0 +1,485 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Simple concrete immutable FST whose states and arcs are each stored in
+// single arrays.
+
+#ifndef FST_CONST_FST_H_
+#define FST_CONST_FST_H_
+
+#include <climits>
+#include <string>
+#include <vector>
+
+// Google-only...
+// ...Google-only
+#include <fst/log.h>
+
+#include <fst/expanded-fst.h>
+#include <fst/fst-decl.h>
+#include <fst/mapped-file.h>
+#include <fst/test-properties.h>
+#include <fst/util.h>
+
+
+namespace fst {
+
+template <class A, class Unsigned>
+class ConstFst;
+
+template <class F, class G>
+void Cast(const F &, G *);
+
+namespace internal {
+
+// States and arcs each implemented by single arrays, templated on the
+// Arc definition. Unsigned is used to represent indices into the arc array.
+template <class A, class Unsigned>
+class ConstFstImpl : public FstImpl<A> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FstImpl<A>::SetInputSymbols;
+  using FstImpl<A>::SetOutputSymbols;
+  using FstImpl<A>::SetType;
+  using FstImpl<A>::SetProperties;
+  using FstImpl<A>::Properties;
+
+  ConstFstImpl()
+      : states_(nullptr),
+        arcs_(nullptr),
+        narcs_(0),
+        nstates_(0),
+        start_(kNoStateId) {
+    string type = "const";
+    if (sizeof(Unsigned) != sizeof(uint32)) {
+      type += std::to_string(CHAR_BIT * sizeof(Unsigned));
+    }
+    SetType(type);
+    SetProperties(kNullProperties | kStaticProperties);
+  }
+
+  explicit ConstFstImpl(const Fst<Arc> &fst);
+
+  StateId Start() const { return start_; }
+
+  Weight Final(StateId s) const { return states_[s].weight; }
+
+  StateId NumStates() const { return nstates_; }
+
+  size_t NumArcs(StateId s) const { return states_[s].narcs; }
+
+  size_t NumInputEpsilons(StateId s) const { return states_[s].niepsilons; }
+
+  size_t NumOutputEpsilons(StateId s) const { return states_[s].noepsilons; }
+
+  static ConstFstImpl<Arc, Unsigned> *Read(std::istream &strm,
+                                           const FstReadOptions &opts);
+
+  const Arc *Arcs(StateId s) const { return arcs_ + states_[s].pos; }
+
+  // Provide information needed for generic state iterator.
+  void InitStateIterator(StateIteratorData<Arc> *data) const {
+    data->base = nullptr;
+    data->nstates = nstates_;
+  }
+
+  // Provide information needed for the generic arc iterator.
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const {
+    data->base = nullptr;
+    data->arcs = arcs_ + states_[s].pos;
+    data->narcs = states_[s].narcs;
+    data->ref_count = nullptr;
+  }
+
+ private:
+  // Used to find narcs_ and nstates_ in Write.
+  friend class ConstFst<Arc, Unsigned>;
+
+  // States implemented by array *states_ below, arcs by (single) *arcs_.
+  struct ConstState {
+    Weight weight;        // Final weight.
+    Unsigned pos;         // Start of state's arcs in *arcs_.
+    Unsigned narcs;       // Number of arcs (per state).
+    Unsigned niepsilons;  // Number of input epsilons.
+    Unsigned noepsilons;  // Number of output epsilons.
+
+    ConstState() : weight(Weight::Zero()) {}
+  };
+
+  // Properties always true of this FST class.
+  static constexpr uint64 kStaticProperties = kExpanded;
+  // Current unaligned file format version. The unaligned version was added and
+  // made the default since the aligned version does not work on pipes.
+  static constexpr int kFileVersion = 2;
+  // Current aligned file format version.
+  static constexpr int kAlignedFileVersion = 1;
+  // Minimum file format version supported.
+  static constexpr int kMinFileVersion = 1;
+
+  std::unique_ptr<MappedFile> states_region_;  // Mapped file for states.
+  std::unique_ptr<MappedFile> arcs_region_;    // Mapped file for arcs.
+  ConstState *states_;                         // States representation.
+  Arc *arcs_;                                  // Arcs representation.
+  size_t narcs_;                               // Number of arcs.
+  StateId nstates_;                            // Number of states.
+  StateId start_;                              // Initial state.
+
+  ConstFstImpl(const ConstFstImpl &) = delete;
+  ConstFstImpl &operator=(const ConstFstImpl &) = delete;
+};
+
+template <class Arc, class Unsigned>
+constexpr uint64 ConstFstImpl<Arc, Unsigned>::kStaticProperties;
+
+template <class Arc, class Unsigned>
+constexpr int ConstFstImpl<Arc, Unsigned>::kFileVersion;
+
+template <class Arc, class Unsigned>
+constexpr int ConstFstImpl<Arc, Unsigned>::kAlignedFileVersion;
+
+template <class Arc, class Unsigned>
+constexpr int ConstFstImpl<Arc, Unsigned>::kMinFileVersion;
+
+template <class Arc, class Unsigned>
+ConstFstImpl<Arc, Unsigned>::ConstFstImpl(const Fst<Arc> &fst)
+    : narcs_(0), nstates_(0) {
+  string type = "const";
+  if (sizeof(Unsigned) != sizeof(uint32)) {
+    type += std::to_string(CHAR_BIT * sizeof(Unsigned));
+  }
+  SetType(type);
+  SetInputSymbols(fst.InputSymbols());
+  SetOutputSymbols(fst.OutputSymbols());
+  start_ = fst.Start();
+  // Counts states and arcs.
+  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
+    ++nstates_;
+    narcs_ += fst.NumArcs(siter.Value());
+  }
+  states_region_.reset(MappedFile::Allocate(nstates_ * sizeof(*states_)));
+  arcs_region_.reset(MappedFile::Allocate(narcs_ * sizeof(*arcs_)));
+  states_ = reinterpret_cast<ConstState *>(states_region_->mutable_data());
+  arcs_ = reinterpret_cast<Arc *>(arcs_region_->mutable_data());
+  size_t pos = 0;
+  for (StateId s = 0; s < nstates_; ++s) {
+    states_[s].weight = fst.Final(s);
+    states_[s].pos = pos;
+    states_[s].narcs = 0;
+    states_[s].niepsilons = 0;
+    states_[s].noepsilons = 0;
+    for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      ++states_[s].narcs;
+      if (arc.ilabel == 0) ++states_[s].niepsilons;
+      if (arc.olabel == 0) ++states_[s].noepsilons;
+      arcs_[pos] = arc;
+      ++pos;
+    }
+  }
+  const auto props =
+      fst.Properties(kMutable, false)
+          ? fst.Properties(kCopyProperties, true)
+          : CheckProperties(
+                fst, kCopyProperties & ~kWeightedCycles & ~kUnweightedCycles,
+                kCopyProperties);
+  SetProperties(props | kStaticProperties);
+}
+
+template <class Arc, class Unsigned>
+ConstFstImpl<Arc, Unsigned> *ConstFstImpl<Arc, Unsigned>::Read(
+    std::istream &strm, const FstReadOptions &opts) {
+  using ConstState = typename ConstFstImpl<Arc, Unsigned>::ConstState;
+  std::unique_ptr<ConstFstImpl<Arc, Unsigned>> impl(
+      new ConstFstImpl<Arc, Unsigned>());
+  FstHeader hdr;
+  if (!impl->ReadHeader(strm, opts, kMinFileVersion, &hdr)) return nullptr;
+  impl->start_ = hdr.Start();
+  impl->nstates_ = hdr.NumStates();
+  impl->narcs_ = hdr.NumArcs();
+  // Ensures compatibility.
+  if (hdr.Version() == kAlignedFileVersion) {
+    hdr.SetFlags(hdr.GetFlags() | FstHeader::IS_ALIGNED);
+  }
+  if ((hdr.GetFlags() & FstHeader::IS_ALIGNED) && !AlignInput(strm)) {
+    LOG(ERROR) << "ConstFst::Read: Alignment failed: " << opts.source;
+    return nullptr;
+  }
+  size_t b = impl->nstates_ * sizeof(ConstState);
+  impl->states_region_.reset(
+      MappedFile::Map(&strm, opts.mode == FstReadOptions::MAP, opts.source, b));
+  if (!strm || !impl->states_region_) {
+    LOG(ERROR) << "ConstFst::Read: Read failed: " << opts.source;
+    return nullptr;
+  }
+  impl->states_ =
+      reinterpret_cast<ConstState *>(impl->states_region_->mutable_data());
+  if ((hdr.GetFlags() & FstHeader::IS_ALIGNED) && !AlignInput(strm)) {
+    LOG(ERROR) << "ConstFst::Read: Alignment failed: " << opts.source;
+    return nullptr;
+  }
+  b = impl->narcs_ * sizeof(Arc);
+  impl->arcs_region_.reset(
+      MappedFile::Map(&strm, opts.mode == FstReadOptions::MAP, opts.source, b));
+  if (!strm || !impl->arcs_region_) {
+    LOG(ERROR) << "ConstFst::Read: Read failed: " << opts.source;
+    return nullptr;
+  }
+  impl->arcs_ = reinterpret_cast<Arc *>(impl->arcs_region_->mutable_data());
+  return impl.release();
+}
+
+}  // namespace internal
+
+// Simple concrete immutable FST. This class attaches interface to
+// implementation and handles reference counting, delegating most methods to
+// ImplToExpandedFst. The unsigned type U is used to represent indices into the
+// arc array (default declared in fst-decl.h).
+template <class A, class Unsigned>
+class ConstFst : public ImplToExpandedFst<internal::ConstFstImpl<A, Unsigned>> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+
+  using Impl = internal::ConstFstImpl<A, Unsigned>;
+  using ConstState = typename Impl::ConstState;
+
+  friend class StateIterator<ConstFst<Arc, Unsigned>>;
+  friend class ArcIterator<ConstFst<Arc, Unsigned>>;
+
+  template <class F, class G>
+  void friend Cast(const F &, G *);
+
+  ConstFst() : ImplToExpandedFst<Impl>(std::make_shared<Impl>()) {}
+
+  explicit ConstFst(const Fst<Arc> &fst)
+      : ImplToExpandedFst<Impl>(std::make_shared<Impl>(fst)) {}
+
+  ConstFst(const ConstFst<A, Unsigned> &fst, bool safe = false)
+      : ImplToExpandedFst<Impl>(fst) {}
+
+  // Gets a copy of this ConstFst. See Fst<>::Copy() for further doc.
+  ConstFst<A, Unsigned> *Copy(bool safe = false) const override {
+    return new ConstFst<A, Unsigned>(*this, safe);
+  }
+
+  // Reads a ConstFst from an input stream, returning nullptr on error.
+  static ConstFst<A, Unsigned> *Read(std::istream &strm,
+                                     const FstReadOptions &opts) {
+    auto *impl = Impl::Read(strm, opts);
+    return impl ? new ConstFst<A, Unsigned>(std::shared_ptr<Impl>(impl))
+                : nullptr;
+  }
+
+  // Read a ConstFst from a file; return nullptr on error; empty filename reads
+  // from standard input.
+  static ConstFst<A, Unsigned> *Read(const string &filename) {
+    auto *impl = ImplToExpandedFst<Impl>::Read(filename);
+    return impl ? new ConstFst<A, Unsigned>(std::shared_ptr<Impl>(impl))
+                : nullptr;
+  }
+
+  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
+    return WriteFst(*this, strm, opts);
+  }
+
+  bool Write(const string &filename) const override {
+    return Fst<Arc>::WriteFile(filename);
+  }
+
+  template <class FST>
+  static bool WriteFst(const FST &fst, std::ostream &strm,
+                       const FstWriteOptions &opts);
+
+  void InitStateIterator(StateIteratorData<Arc> *data) const override {
+    GetImpl()->InitStateIterator(data);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
+    GetImpl()->InitArcIterator(s, data);
+  }
+
+ private:
+  explicit ConstFst(std::shared_ptr<Impl> impl)
+      : ImplToExpandedFst<Impl>(impl) {}
+
+  using ImplToFst<Impl, ExpandedFst<Arc>>::GetImpl;
+
+  // Uses overloading to extract the type of the argument.
+  static const Impl *GetImplIfConstFst(const ConstFst &const_fst) {
+    return const_fst.GetImpl();
+  }
+
+  // NB: this does not give privileged treatment to subtypes of ConstFst.
+  template <typename FST>
+  static Impl *GetImplIfConstFst(const FST &fst) {
+    return nullptr;
+  }
+
+  ConstFst &operator=(const ConstFst &) = delete;
+};
+
+// Writes FST in Const format, potentially with a pass over the machine before
+// writing to compute number of states and arcs.
+template <class Arc, class Unsigned>
+template <class FST>
+bool ConstFst<Arc, Unsigned>::WriteFst(const FST &fst, std::ostream &strm,
+                                       const FstWriteOptions &opts) {
+  const auto file_version =
+      opts.align ? internal::ConstFstImpl<Arc, Unsigned>::kAlignedFileVersion
+                 : internal::ConstFstImpl<Arc, Unsigned>::kFileVersion;
+  size_t num_arcs = 0;    // To silence -Wsometimes-uninitialized warnings.
+  size_t num_states = 0;  // Ditto.
+  std::streamoff start_offset = 0;
+  bool update_header = true;
+  if (const auto *impl = GetImplIfConstFst(fst)) {
+    num_arcs = impl->narcs_;
+    num_states = impl->nstates_;
+    update_header = false;
+  } else if (opts.stream_write || (start_offset = strm.tellp()) == -1) {
+    // precompute values needed for header when we cannot seek to rewrite it.
+    num_arcs = 0;
+    num_states = 0;
+    for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
+      num_arcs += fst.NumArcs(siter.Value());
+      ++num_states;
+    }
+    update_header = false;
+  }
+  FstHeader hdr;
+  hdr.SetStart(fst.Start());
+  hdr.SetNumStates(num_states);
+  hdr.SetNumArcs(num_arcs);
+  string type = "const";
+  if (sizeof(Unsigned) != sizeof(uint32)) {
+    type += std::to_string(CHAR_BIT * sizeof(Unsigned));
+  }
+  const auto properties =
+      fst.Properties(kCopyProperties, true) |
+      internal::ConstFstImpl<Arc, Unsigned>::kStaticProperties;
+  internal::FstImpl<Arc>::WriteFstHeader(fst, strm, opts, file_version, type,
+                                         properties, &hdr);
+  if (opts.align && !AlignOutput(strm)) {
+    LOG(ERROR) << "Could not align file during write after header";
+    return false;
+  }
+  size_t pos = 0;
+  size_t states = 0;
+  typename ConstFst<Arc, Unsigned>::ConstState state;
+  for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
+    const auto s = siter.Value();
+    state.weight = fst.Final(s);
+    state.pos = pos;
+    state.narcs = fst.NumArcs(s);
+    state.niepsilons = fst.NumInputEpsilons(s);
+    state.noepsilons = fst.NumOutputEpsilons(s);
+    strm.write(reinterpret_cast<const char *>(&state), sizeof(state));
+    pos += state.narcs;
+    ++states;
+  }
+  hdr.SetNumStates(states);
+  hdr.SetNumArcs(pos);
+  if (opts.align && !AlignOutput(strm)) {
+    LOG(ERROR) << "Could not align file during write after writing states";
+  }
+  for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
+    for (ArcIterator<FST> aiter(fst, siter.Value()); !aiter.Done();
+         aiter.Next()) {
+      const auto &arc = aiter.Value();
+// Google-only...
+#ifdef MEMORY_SANITIZER
+      // arc may contain padding which has unspecified contents. Tell MSAN to
+      // not complain about it when writing it to a file.
+      ANNOTATE_MEMORY_IS_INITIALIZED(reinterpret_cast<const char *>(&arc),
+                                     sizeof(arc));
+#endif
+      // ...Google-only
+      strm.write(reinterpret_cast<const char *>(&arc), sizeof(arc));
+    }
+  }
+  strm.flush();
+  if (!strm) {
+    LOG(ERROR) << "ConstFst::WriteFst: write failed: " << opts.source;
+    return false;
+  }
+  if (update_header) {
+    return internal::FstImpl<Arc>::UpdateFstHeader(
+        fst, strm, opts, file_version, type, properties, &hdr, start_offset);
+  } else {
+    if (hdr.NumStates() != num_states) {
+      LOG(ERROR) << "Inconsistent number of states observed during write";
+      return false;
+    }
+    if (hdr.NumArcs() != num_arcs) {
+      LOG(ERROR) << "Inconsistent number of arcs observed during write";
+      return false;
+    }
+  }
+  return true;
+}
+
+// Specialization for ConstFst; see generic version in fst.h for sample usage
+// (but use the ConstFst type instead). This version should inline.
+template <class Arc, class Unsigned>
+class StateIterator<ConstFst<Arc, Unsigned>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  explicit StateIterator(const ConstFst<Arc, Unsigned> &fst)
+      : nstates_(fst.GetImpl()->NumStates()), s_(0) {}
+
+  bool Done() const { return s_ >= nstates_; }
+
+  StateId Value() const { return s_; }
+
+  void Next() { ++s_; }
+
+  void Reset() { s_ = 0; }
+
+ private:
+  const StateId nstates_;
+  StateId s_;
+};
+
+// Specialization for ConstFst; see generic version in fst.h for sample usage
+// (but use the ConstFst type instead). This version should inline.
+template <class Arc, class Unsigned>
+class ArcIterator<ConstFst<Arc, Unsigned>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const ConstFst<Arc, Unsigned> &fst, StateId s)
+      : arcs_(fst.GetImpl()->Arcs(s)),
+        narcs_(fst.GetImpl()->NumArcs(s)),
+        i_(0) {}
+
+  bool Done() const { return i_ >= narcs_; }
+
+  const Arc &Value() const { return arcs_[i_]; }
+
+  void Next() { ++i_; }
+
+  size_t Position() const { return i_; }
+
+  void Reset() { i_ = 0; }
+
+  void Seek(size_t a) { i_ = a; }
+
+  constexpr uint32 Flags() const { return kArcValueFlags; }
+
+  void SetFlags(uint32, uint32) {}
+
+ private:
+  const Arc *arcs_;
+  size_t narcs_;
+  size_t i_;
+};
+
+// A useful alias when using StdArc.
+using StdConstFst = ConstFst<StdArc>;
+
+}  // namespace fst
+
+#endif  // FST_CONST_FST_H_
diff --git a/projects/llm_framework/include/fst/determinize.h b/projects/llm_framework/include/fst/determinize.h
new file mode 100644
index 00000000..736a140e
--- /dev/null
+++ b/projects/llm_framework/include/fst/determinize.h
@@ -0,0 +1,1093 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes to determinize an FST.
+
+#ifndef FST_DETERMINIZE_H_
+#define FST_DETERMINIZE_H_
+
+#include <algorithm>
+#include <climits>
+#include <forward_list>
+#include <map>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/arc-map.h>
+#include <fst/bi-table.h>
+#include <fst/cache.h>
+#include <fst/factor-weight.h>
+#include <fst/filter-state.h>
+#include <fst/prune.h>
+#include <fst/test-properties.h>
+
+
+namespace fst {
+
+// Common divisors are used in determinization to compute transition weights.
+// In the simplest case, it is the same as semiring Plus, but other choices
+// permit more efficient determinization when the output contains strings.
+
+// The default common divisor uses the semiring Plus.
+template <class W>
+struct DefaultCommonDivisor {
+ public:
+  using Weight = W;
+
+  Weight operator()(const Weight &w1, const Weight &w2) const {
+    return Plus(w1, w2);
+  }
+};
+
+// The label common divisor for a (left) string semiring selects a single
+// letter common prefix or the empty string. This is used in the
+// determinization of output strings so that at most a single letter will
+// appear in the output of a transtion.
+template <typename Label, StringType S>
+struct LabelCommonDivisor {
+ public:
+  using Weight = StringWeight<Label, S>;
+
+  Weight operator()(const Weight &w1, const Weight &w2) const {
+    typename Weight::Iterator iter1(w1);
+    typename Weight::Iterator iter2(w2);
+    if (!(StringWeight<Label, S>::Properties() & kLeftSemiring)) {
+      FSTERROR() << "LabelCommonDivisor: Weight needs to be left semiring";
+      return Weight::NoWeight();
+    } else if (w1.Size() == 0 || w2.Size() == 0) {
+      return Weight::One();
+    } else if (w1 == Weight::Zero()) {
+      return Weight(iter2.Value());
+    } else if (w2 == Weight::Zero()) {
+      return Weight(iter1.Value());
+    } else if (iter1.Value() == iter2.Value()) {
+      return Weight(iter1.Value());
+    } else {
+      return Weight::One();
+    }
+  }
+};
+
+// The gallic common divisor uses the label common divisor on the string
+// component and the common divisor on the weight component, which defaults to
+// the default common divisor.
+template <class Label, class W, GallicType G,
+          class CommonDivisor = DefaultCommonDivisor<W>>
+class GallicCommonDivisor {
+ public:
+  using Weight = GallicWeight<Label, W, G>;
+
+  Weight operator()(const Weight &w1, const Weight &w2) const {
+    return Weight(label_common_divisor_(w1.Value1(), w2.Value1()),
+                  weight_common_divisor_(w1.Value2(), w2.Value2()));
+  }
+
+ private:
+  LabelCommonDivisor<Label, GallicStringType(G)> label_common_divisor_;
+  CommonDivisor weight_common_divisor_;
+};
+
+// Specialization for general GALLIC weight.
+template <class Label, class W, class CommonDivisor>
+class GallicCommonDivisor<Label, W, GALLIC, CommonDivisor> {
+ public:
+  using Weight = GallicWeight<Label, W, GALLIC>;
+  using GRWeight = GallicWeight<Label, W, GALLIC_RESTRICT>;
+  using Iterator =
+      UnionWeightIterator<GRWeight, GallicUnionWeightOptions<Label, W>>;
+
+  Weight operator()(const Weight &w1, const Weight &w2) const {
+    auto weight = GRWeight::Zero();
+    for (Iterator iter(w1); !iter.Done(); iter.Next()) {
+      weight = common_divisor_(weight, iter.Value());
+    }
+    for (Iterator iter(w2); !iter.Done(); iter.Next()) {
+      weight = common_divisor_(weight, iter.Value());
+    }
+    return weight == GRWeight::Zero() ? Weight::Zero() : Weight(weight);
+  }
+
+ private:
+  GallicCommonDivisor<Label, W, GALLIC_RESTRICT, CommonDivisor> common_divisor_;
+};
+
+namespace internal {
+
+// Represents an element in a subset
+template <class Arc>
+struct DeterminizeElement {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  DeterminizeElement(StateId s, Weight weight)
+      : state_id(s), weight(std::move(weight)) {}
+
+  inline bool operator==(const DeterminizeElement<Arc> &element) const {
+    return state_id == element.state_id && weight == element.weight;
+  }
+
+  inline bool operator!=(const DeterminizeElement<Arc> &element) const {
+    return !(*this == element);
+  }
+
+  inline bool operator<(const DeterminizeElement<Arc> &element) const {
+    return state_id < element.state_id;
+  }
+
+  StateId state_id;  // Input state ID.
+  Weight weight;     // Residual weight.
+};
+
+// Represents a weighted subset and determinization filter state
+template <typename A, typename FilterState>
+struct DeterminizeStateTuple {
+  using Arc = A;
+  using Element = DeterminizeElement<Arc>;
+  using Subset = std::forward_list<Element>;
+
+  DeterminizeStateTuple() : filter_state(FilterState::NoState()) {}
+
+  inline bool operator==(
+      const DeterminizeStateTuple<Arc, FilterState> &tuple) const {
+    return (tuple.filter_state == filter_state) && (tuple.subset == subset);
+  }
+
+  inline bool operator!=(
+      const DeterminizeStateTuple<Arc, FilterState> &tuple) const {
+    return (tuple.filter_state != filter_state) || (tuple.subset != subset);
+  }
+
+  Subset subset;
+  FilterState filter_state;
+};
+
+// Proto-transition for determinization.
+template <class StateTuple>
+struct DeterminizeArc {
+  using Arc = typename StateTuple::Arc;
+  using Label = typename Arc::Label;
+  using Weight = typename Arc::Weight;
+
+  DeterminizeArc()
+      : label(kNoLabel), weight(Weight::Zero()), dest_tuple(nullptr) {}
+
+  explicit DeterminizeArc(const Arc &arc)
+      : label(arc.ilabel), weight(Weight::Zero()), dest_tuple(new StateTuple) {}
+
+  Label label;             // Arc label.
+  Weight weight;           // Arc weight.
+  StateTuple *dest_tuple;  // Destination subset and filter state.
+};
+
+}  // namespace internal
+
+// Determinization filters are used to compute destination state tuples based
+// on the source tuple, transition, and destination element or on similar
+// super-final transition information. The filter operates on a map between a
+// label and the corresponding destination state tuples. It must define the map
+// type LabelMap. The default filter is used for weighted determinization.
+// A determinize filter for implementing weighted determinization.
+template <class Arc>
+class DefaultDeterminizeFilter {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FilterState = CharFilterState;
+  using Element = internal::DeterminizeElement<Arc>;
+  using StateTuple = internal::DeterminizeStateTuple<Arc, FilterState>;
+  using LabelMap = std::map<Label, internal::DeterminizeArc<StateTuple>>;
+
+  // This is needed e.g. to go into the gallic domain for transducers.
+  template <class A>
+  struct rebind {
+    using Other = DefaultDeterminizeFilter<A>;
+  };
+
+  explicit DefaultDeterminizeFilter(const Fst<Arc> &fst) : fst_(fst.Copy()) {}
+
+  // This is needed (e.g.) to go into the gallic domain for transducers.
+  // Ownership of the templated filter argument is given to this class.
+  template <class Filter>
+  DefaultDeterminizeFilter(const Fst<Arc> &fst, Filter *filter)
+      : fst_(fst.Copy()) {
+    delete filter;
+  }
+
+  // Copy constructor; the FST can be passed if it has been deep-copied.
+  DefaultDeterminizeFilter(const DefaultDeterminizeFilter<Arc> &filter,
+                           const Fst<Arc> *fst = nullptr)
+      : fst_(fst ? fst->Copy() : filter.fst_->Copy()) {}
+
+  FilterState Start() const { return FilterState(0); }
+
+  // Does no work.
+  void SetState(StateId s, const StateTuple &tuple) {}
+
+  // Filters transition, possibly modifying label map. Returns true if arc is
+  // added to the label map.
+  bool FilterArc(const Arc &arc, const Element &src_element,
+                 Element &&dest_element, LabelMap *label_map) const {
+    // Adds element to unique state tuple for arc label.
+    auto &det_arc = (*label_map)[arc.ilabel];
+    if (det_arc.label == kNoLabel) {
+      det_arc = internal::DeterminizeArc<StateTuple>(arc);
+      det_arc.dest_tuple->filter_state = FilterState(0);
+    }
+    det_arc.dest_tuple->subset.push_front(std::move(dest_element));
+    return true;
+  }
+
+  // Filters super-final transition, returning new final weight.
+  Weight FilterFinal(Weight weight, const Element &element) { return weight; }
+
+  static uint64 Properties(uint64 props) { return props; }
+
+ private:
+  std::unique_ptr<Fst<Arc>> fst_;
+};
+
+// Determinization state table interface:
+//
+// template <class Arc, class FilterState>
+// class DeterminizeStateTable {
+//  public:
+//   using StateId = typename Arc::StateId;
+//   using StateTuple = internal::DeterminizeStateTuple<Arc, FilterState>;
+//
+//   // Required sub-class. This is needed (e.g.) to go into the gallic domain.
+//   template <class B, class G>
+//   struct rebind {
+//     using Other = DeterminizeStateTable<B, G>;
+//   }
+//
+//   // Required constuctor.
+//   DeterminizeStateTable();
+//
+//   // Required copy constructor that does not copy state.
+//   DeterminizeStateTable(const DeterminizeStateTable<Arc, FilterState>
+//   &table);
+//
+//   // Looks up state ID by state tuple; if it doesn't exist, then adds it.
+//   // FindState takes ownership of the state tuple argument so that it
+//   // doesn't have to copy it if it creates a new state.
+//   StateId FindState(StateTuple *tuple);
+//
+//   // Looks up state tuple by ID.
+//   const StateTuple *Tuple(StateId id) const;
+// };
+
+// The default determinization state table based on the compact hash bi-table.
+template <class Arc, class FilterState>
+class DefaultDeterminizeStateTable {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using StateTuple = internal::DeterminizeStateTuple<Arc, FilterState>;
+  using Element = typename StateTuple::Element;
+  using Subset = typename StateTuple::Subset;
+
+  template <class B, class G>
+  struct rebind {
+    using Other = DefaultDeterminizeStateTable<B, G>;
+  };
+
+  explicit DefaultDeterminizeStateTable(size_t table_size = 0)
+      : table_size_(table_size), tuples_(table_size_) {}
+
+  DefaultDeterminizeStateTable(
+      const DefaultDeterminizeStateTable<Arc, FilterState> &table)
+      : table_size_(table.table_size_), tuples_(table_size_) {}
+
+  ~DefaultDeterminizeStateTable() {
+    for (StateId s = 0; s < tuples_.Size(); ++s) delete tuples_.FindEntry(s);
+  }
+
+  // Finds the state corresponding to a state tuple. Only creates a new state if
+  // the tuple is not found. FindState takes ownership of the tuple argument so
+  // that it doesn't have to copy it if it creates a new state.
+  StateId FindState(StateTuple *tuple) {
+    const StateId ns = tuples_.Size();
+    const auto s = tuples_.FindId(tuple);
+    if (s != ns) delete tuple;  // Tuple found.
+    return s;
+  }
+
+  const StateTuple *Tuple(StateId s) { return tuples_.FindEntry(s); }
+
+ private:
+  // Comparison object for StateTuples.
+  class StateTupleEqual {
+   public:
+    bool operator()(const StateTuple *tuple1, const StateTuple *tuple2) const {
+      return *tuple1 == *tuple2;
+    }
+  };
+
+  // Hash function for StateTuples.
+  class StateTupleKey {
+   public:
+    size_t operator()(const StateTuple *tuple) const {
+      size_t h = tuple->filter_state.Hash();
+      for (auto it = tuple->subset.begin(); it != tuple->subset.end(); ++it) {
+        const size_t h1 = it->state_id;
+        static constexpr auto lshift = 5;
+        static constexpr auto rshift = CHAR_BIT * sizeof(size_t) - 5;
+        h ^= h << 1 ^ h1 << lshift ^ h1 >> rshift ^ it->weight.Hash();
+      }
+      return h;
+    }
+  };
+
+  size_t table_size_;
+  CompactHashBiTable<StateId, StateTuple *, StateTupleKey, StateTupleEqual,
+                     HS_STL>
+      tuples_;
+
+  DefaultDeterminizeStateTable &operator=(
+      const DefaultDeterminizeStateTable &) = delete;
+};
+
+// Determinization type.
+enum DeterminizeType {
+  // Input transducer is known to be functional (or error).
+  DETERMINIZE_FUNCTIONAL,  // Input transducer is functional (error if not).
+  // Input transducer is not known to be functional.
+  DETERMINIZE_NONFUNCTIONAL,
+  // Input transducer is not known to be functional but only keep the min of
+  // of ambiguous outputs.
+  DETERMINIZE_DISAMBIGUATE
+};
+
+// Options for finite-state transducer determinization templated on the arc
+// type, common divisor, the determinization filter and the state table.
+// DeterminizeFst takes ownership of the determinization filter and state table,
+// if provided.
+template <class Arc,
+          class CommonDivisor = DefaultCommonDivisor<typename Arc::Weight>,
+          class Filter = DefaultDeterminizeFilter<Arc>,
+          class StateTable =
+              DefaultDeterminizeStateTable<Arc, typename Filter::FilterState>>
+struct DeterminizeFstOptions : public CacheOptions {
+  using Label = typename Arc::Label;
+
+  float delta;                // Quantization delta for subset weights.
+  Label subsequential_label;  // Label used for residual final output
+                              // when producing subsequential transducers.
+  DeterminizeType type;       // Determinization type.
+  bool increment_subsequential_label;  // When creating several subsequential
+                                       // arcs at a given state, make their
+                                       // label distinct by incrementing.
+  Filter *filter;                      // Determinization filter;
+                                       // DeterminizeFst takes ownership.
+  StateTable *state_table;             // Determinization state table;
+                                       // DeterminizeFst takes ownership.
+
+  explicit DeterminizeFstOptions(const CacheOptions &opts, float delta = kDelta,
+                                 Label subsequential_label = 0,
+                                 DeterminizeType type = DETERMINIZE_FUNCTIONAL,
+                                 bool increment_subsequential_label = false,
+                                 Filter *filter = nullptr,
+                                 StateTable *state_table = nullptr)
+      : CacheOptions(opts),
+        delta(delta),
+        subsequential_label(subsequential_label),
+        type(type),
+        increment_subsequential_label(increment_subsequential_label),
+        filter(filter),
+        state_table(state_table) {}
+
+  explicit DeterminizeFstOptions(float delta = kDelta,
+                                 Label subsequential_label = 0,
+                                 DeterminizeType type = DETERMINIZE_FUNCTIONAL,
+                                 bool increment_subsequential_label = false,
+                                 Filter *filter = nullptr,
+                                 StateTable *state_table = nullptr)
+      : delta(delta),
+        subsequential_label(subsequential_label),
+        type(type),
+        increment_subsequential_label(increment_subsequential_label),
+        filter(filter),
+        state_table(state_table) {}
+};
+
+namespace internal {
+
+// Implementation of delayed DeterminizeFst. This base class is
+// common to the variants that implement acceptor and transducer
+// determinization.
+template <class Arc>
+class DeterminizeFstImplBase : public CacheImpl<Arc> {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Store = DefaultCacheStore<Arc>;
+  using State = typename Store::State;
+
+  using FstImpl<Arc>::SetType;
+  using FstImpl<Arc>::SetProperties;
+  using FstImpl<Arc>::Properties;
+  using FstImpl<Arc>::SetInputSymbols;
+  using FstImpl<Arc>::SetOutputSymbols;
+
+  using CacheBaseImpl<CacheState<Arc>>::HasStart;
+  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
+  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
+  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
+  using CacheBaseImpl<CacheState<Arc>>::SetStart;
+
+  template <class CommonDivisor, class Filter, class StateTable>
+  DeterminizeFstImplBase(
+      const Fst<Arc> &fst,
+      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable> &opts)
+      : CacheImpl<Arc>(opts), fst_(fst.Copy()) {
+    SetType("determinize");
+    const auto iprops = fst.Properties(kFstProperties, false);
+    const auto dprops =
+        DeterminizeProperties(iprops, opts.subsequential_label != 0,
+                              opts.type == DETERMINIZE_NONFUNCTIONAL
+                                  ? opts.increment_subsequential_label
+                                  : true);
+    SetProperties(Filter::Properties(dprops), kCopyProperties);
+    SetInputSymbols(fst.InputSymbols());
+    SetOutputSymbols(fst.OutputSymbols());
+  }
+
+  DeterminizeFstImplBase(const DeterminizeFstImplBase<Arc> &impl)
+      : CacheImpl<Arc>(impl), fst_(impl.fst_->Copy(true)) {
+    SetType("determinize");
+    SetProperties(impl.Properties(), kCopyProperties);
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+  }
+
+  virtual DeterminizeFstImplBase<Arc> *Copy() const = 0;
+
+  StateId Start() {
+    if (!HasStart()) {
+      const auto start = ComputeStart();
+      if (start != kNoStateId) SetStart(start);
+    }
+    return CacheImpl<Arc>::Start();
+  }
+
+  Weight Final(StateId s) {
+    if (!HasFinal(s)) SetFinal(s, ComputeFinal(s));
+    return CacheImpl<Arc>::Final(s);
+  }
+
+  virtual void Expand(StateId s) = 0;
+
+  size_t NumArcs(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<Arc>::NumArcs(s);
+  }
+
+  size_t NumInputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<Arc>::NumInputEpsilons(s);
+  }
+
+  size_t NumOutputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<Arc>::NumOutputEpsilons(s);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
+    if (!HasArcs(s)) Expand(s);
+    CacheImpl<Arc>::InitArcIterator(s, data);
+  }
+
+  virtual StateId ComputeStart() = 0;
+
+  virtual Weight ComputeFinal(StateId s) = 0;
+
+  const Fst<Arc> &GetFst() const { return *fst_; }
+
+ private:
+  std::unique_ptr<const Fst<Arc>> fst_;  // Input FST.
+};
+
+// Implementation of delayed determinization for weighted acceptors.
+template <class Arc, class CommonDivisor, class Filter, class StateTable>
+class DeterminizeFsaImpl : public DeterminizeFstImplBase<Arc> {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FilterState = typename Filter::FilterState;
+  using StateTuple = internal::DeterminizeStateTuple<Arc, FilterState>;
+  using Element = typename StateTuple::Element;
+  using Subset = typename StateTuple::Subset;
+  using LabelMap = typename Filter::LabelMap;
+
+  using FstImpl<Arc>::SetProperties;
+  using DeterminizeFstImplBase<Arc>::GetFst;
+  using DeterminizeFstImplBase<Arc>::SetArcs;
+
+  DeterminizeFsaImpl(
+      const Fst<Arc> &fst, const std::vector<Weight> *in_dist,
+      std::vector<Weight> *out_dist,
+      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable> &opts)
+      : DeterminizeFstImplBase<Arc>(fst, opts),
+        delta_(opts.delta),
+        in_dist_(in_dist),
+        out_dist_(out_dist),
+        filter_(opts.filter ? opts.filter : new Filter(fst)),
+        state_table_(opts.state_table ? opts.state_table : new StateTable()) {
+    if (!fst.Properties(kAcceptor, true)) {
+      FSTERROR() << "DeterminizeFst: Argument not an acceptor";
+      SetProperties(kError, kError);
+    }
+    if (!(Weight::Properties() & kLeftSemiring)) {
+      FSTERROR() << "DeterminizeFst: Weight must be left distributive: "
+                 << Weight::Type();
+      SetProperties(kError, kError);
+    }
+    if (out_dist_) out_dist_->clear();
+  }
+
+  DeterminizeFsaImpl(
+      const DeterminizeFsaImpl<Arc, CommonDivisor, Filter, StateTable> &impl)
+      : DeterminizeFstImplBase<Arc>(impl),
+        delta_(impl.delta_),
+        in_dist_(nullptr),
+        out_dist_(nullptr),
+        filter_(new Filter(*impl.filter_, &GetFst())),
+        state_table_(new StateTable(*impl.state_table_)) {
+    if (impl.out_dist_) {
+      FSTERROR() << "DeterminizeFsaImpl: Cannot copy with out_dist vector";
+      SetProperties(kError, kError);
+    }
+  }
+
+  DeterminizeFsaImpl<Arc, CommonDivisor, Filter, StateTable> *Copy()
+      const override {
+    return new DeterminizeFsaImpl<Arc, CommonDivisor, Filter, StateTable>(
+        *this);
+  }
+
+  uint64 Properties() const override { return Properties(kFstProperties); }
+
+  // Sets error if found, and returns other FST impl properties.
+  uint64 Properties(uint64 mask) const override {
+    if ((mask & kError) && (GetFst().Properties(kError, false))) {
+      SetProperties(kError, kError);
+    }
+    return FstImpl<Arc>::Properties(mask);
+  }
+
+  StateId ComputeStart() override {
+    const auto s = GetFst().Start();
+    if (s == kNoStateId) return kNoStateId;
+    auto *tuple = new StateTuple;
+    tuple->subset.emplace_front(s, Weight::One());
+    tuple->filter_state = filter_->Start();
+    return FindState(tuple);
+  }
+
+  Weight ComputeFinal(StateId s) override {
+    const auto *tuple = state_table_->Tuple(s);
+    filter_->SetState(s, *tuple);
+    auto final_weight = Weight::Zero();
+    for (auto it = tuple->subset.begin(); it != tuple->subset.end(); ++it) {
+      const auto &element = *it;
+      final_weight =
+          Plus(final_weight,
+               Times(element.weight, GetFst().Final(element.state_id)));
+      final_weight = filter_->FilterFinal(final_weight, element);
+      if (!final_weight.Member()) SetProperties(kError, kError);
+    }
+    return final_weight;
+  }
+
+  StateId FindState(StateTuple *tuple) {
+    const auto s = state_table_->FindState(tuple);
+    if (in_dist_ && out_dist_->size() <= s) {
+      out_dist_->push_back(ComputeDistance(tuple->subset));
+    }
+    return s;
+  }
+
+  // Computes distance from a state to the final states in the DFA given the
+  // distances in the NFA.
+  Weight ComputeDistance(const Subset &subset) {
+    auto outd = Weight::Zero();
+    for (auto it = subset.begin(); it != subset.end(); ++it) {
+      const auto &element = *it;
+      const auto ind =
+          (element.state_id < in_dist_->size() ? (*in_dist_)[element.state_id]
+                                               : Weight::Zero());
+      outd = Plus(outd, Times(element.weight, ind));
+    }
+    return outd;
+  }
+
+  // Computes the outgoing transitions from a state, creating new destination
+  // states as needed.
+  void Expand(StateId s) override {
+    LabelMap label_map;
+    GetLabelMap(s, &label_map);
+    for (auto it = label_map.begin(); it != label_map.end(); ++it) {
+      AddArc(s, std::move(it->second));
+    }
+    SetArcs(s);
+  }
+
+ private:
+  using DetArc = internal::DeterminizeArc<StateTuple>;
+
+  // Constructs proto-determinization transition, including destination subset,
+  // per label.
+  void GetLabelMap(StateId s, LabelMap *label_map) {
+    const auto *src_tuple = state_table_->Tuple(s);
+    filter_->SetState(s, *src_tuple);
+    for (auto it = src_tuple->subset.begin(); it != src_tuple->subset.end();
+         ++it) {
+      const auto &src_element = *it;
+      for (ArcIterator<Fst<Arc>> aiter(GetFst(), src_element.state_id);
+           !aiter.Done(); aiter.Next()) {
+        const auto &arc = aiter.Value();
+        Element dest_element(arc.nextstate,
+                             Times(src_element.weight, arc.weight));
+        filter_->FilterArc(arc, src_element, std::move(dest_element),
+                           label_map);
+      }
+    }
+    for (auto it = label_map->begin(); it != label_map->end(); ++it) {
+      NormArc(&it->second);
+    }
+  }
+
+  // Sorts subsets and removes duplicate elements, normalizing transition and
+  // subset weights.
+  void NormArc(DetArc *det_arc) {
+    auto *dest_tuple = det_arc->dest_tuple;
+    dest_tuple->subset.sort();
+    auto piter = dest_tuple->subset.begin();
+    for (auto diter = dest_tuple->subset.begin();
+         diter != dest_tuple->subset.end();) {
+      auto &dest_element = *diter;
+      auto &prev_element = *piter;
+      // Computes arc weight.
+      det_arc->weight = common_divisor_(det_arc->weight, dest_element.weight);
+      if (piter != diter && dest_element.state_id == prev_element.state_id) {
+        // Found duplicate state: sums state weight and deletes duplicate.
+        prev_element.weight = Plus(prev_element.weight, dest_element.weight);
+        if (!prev_element.weight.Member()) SetProperties(kError, kError);
+        ++diter;
+        dest_tuple->subset.erase_after(piter);
+      } else {
+        piter = diter;
+        ++diter;
+      }
+    }
+    // Divides out label weight from destination subset elements, quantizing to
+    // ensure comparisons are effective.
+    for (auto diter = dest_tuple->subset.begin();
+         diter != dest_tuple->subset.end(); ++diter) {
+      auto &dest_element = *diter;
+      dest_element.weight =
+          Divide(dest_element.weight, det_arc->weight, DIVIDE_LEFT);
+      dest_element.weight = dest_element.weight.Quantize(delta_);
+    }
+  }
+
+  // Adds an arc from state S to the destination state associated with state
+  // tuple in det_arc as created by GetLabelMap.
+  void AddArc(StateId s, DetArc &&det_arc) {
+    CacheImpl<Arc>::EmplaceArc(
+        s, det_arc.label, det_arc.label, std::move(det_arc.weight),
+        FindState(det_arc.dest_tuple));
+  }
+
+  float delta_;                         // Quantization delta for weights.
+  const std::vector<Weight> *in_dist_;  // Distance to final NFA states.
+  std::vector<Weight> *out_dist_;       // Distance to final DFA states.
+
+  // FIXME(kbg): Ought to be static const?
+  CommonDivisor common_divisor_;
+  std::unique_ptr<Filter> filter_;
+  std::unique_ptr<StateTable> state_table_;
+};
+
+// Implementation of delayed determinization for transducers. Transducer
+// determinization is implemented by mapping the input to the Gallic semiring as
+// an acceptor whose weights contain the output strings and using acceptor
+// determinization above to determinize that acceptor.
+template <class Arc, GallicType G, class CommonDivisor, class Filter,
+          class StateTable>
+class DeterminizeFstImpl : public DeterminizeFstImplBase<Arc> {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using ToMapper = ToGallicMapper<Arc, G>;
+  using ToArc = typename ToMapper::ToArc;
+  using ToFst = ArcMapFst<Arc, ToArc, ToMapper>;
+  using FromMapper = FromGallicMapper<Arc, G>;
+  using FromFst = ArcMapFst<ToArc, Arc, FromMapper>;
+
+  using ToCommonDivisor = GallicCommonDivisor<Label, Weight, G, CommonDivisor>;
+  using ToFilter = typename Filter::template rebind<ToArc>::Other;
+  using ToFilterState = typename ToFilter::FilterState;
+  using ToStateTable =
+      typename StateTable::template rebind<ToArc, ToFilterState>::Other;
+  using FactorIterator = GallicFactor<Label, Weight, G>;
+
+  using FstImpl<Arc>::SetProperties;
+  using DeterminizeFstImplBase<Arc>::GetFst;
+  using CacheBaseImpl<CacheState<Arc>>::GetCacheGc;
+  using CacheBaseImpl<CacheState<Arc>>::GetCacheLimit;
+
+  DeterminizeFstImpl(
+      const Fst<Arc> &fst,
+      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable> &opts)
+      : DeterminizeFstImplBase<Arc>(fst, opts),
+        delta_(opts.delta),
+        subsequential_label_(opts.subsequential_label),
+        increment_subsequential_label_(opts.increment_subsequential_label) {
+    if (opts.state_table) {
+      FSTERROR() << "DeterminizeFst: "
+                 << "A state table can not be passed with transducer input";
+      SetProperties(kError, kError);
+      return;
+    }
+    Init(GetFst(), opts.filter);
+  }
+
+  DeterminizeFstImpl(
+      const DeterminizeFstImpl<Arc, G, CommonDivisor, Filter, StateTable> &impl)
+      : DeterminizeFstImplBase<Arc>(impl),
+        delta_(impl.delta_),
+        subsequential_label_(impl.subsequential_label_),
+        increment_subsequential_label_(impl.increment_subsequential_label_) {
+    Init(GetFst(), nullptr);
+  }
+
+  DeterminizeFstImpl<Arc, G, CommonDivisor, Filter, StateTable> *Copy()
+      const override {
+    return new DeterminizeFstImpl<Arc, G, CommonDivisor, Filter, StateTable>(
+        *this);
+  }
+
+  uint64 Properties() const override { return Properties(kFstProperties); }
+
+  // Sets error if found, and returns other FST impl properties.
+  uint64 Properties(uint64 mask) const override {
+    if ((mask & kError) && (GetFst().Properties(kError, false) ||
+                            from_fst_->Properties(kError, false))) {
+      SetProperties(kError, kError);
+    }
+    return FstImpl<Arc>::Properties(mask);
+  }
+
+  StateId ComputeStart() override { return from_fst_->Start(); }
+
+  Weight ComputeFinal(StateId s) override { return from_fst_->Final(s); }
+
+  void Expand(StateId s) override {
+    for (ArcIterator<FromFst> aiter(*from_fst_, s); !aiter.Done();
+         aiter.Next()) {
+      CacheImpl<Arc>::PushArc(s, aiter.Value());
+    }
+    CacheImpl<Arc>::SetArcs(s);
+  }
+
+ private:
+  // Initialization of transducer determinization implementation, which is
+  // defined after DeterminizeFst since it calls it.
+  void Init(const Fst<Arc> &fst, Filter *filter);
+
+  float delta_;
+  Label subsequential_label_;
+  bool increment_subsequential_label_;
+  std::unique_ptr<FromFst> from_fst_;
+};
+
+}  // namespace internal
+
+// Determinizes a weighted transducer. This version is a delayed
+// FST. The result will be an equivalent FST that has the property
+// that no state has two transitions with the same input label.
+// For this algorithm, epsilon transitions are treated as regular
+// symbols (cf. RmEpsilon).
+//
+// The transducer must be functional. The weights must be (weakly) left
+// divisible (valid for TropicalWeight and LogWeight for instance) and be
+// zero-sum-free if for all a, b: (Plus(a, b) == 0) => a = b = 0.
+//
+// Complexity:
+//
+//   Determinizable: exponential (polynomial in the size of the output).
+//   Non-determinizable: does not terminate.
+//
+// The determinizable automata include all unweighted and all acyclic input.
+//
+// For more information, see:
+//
+// Mohri, M. 1997. Finite-state transducers in language and speech processing.
+// Computational Linguistics 23(2): 269-311.
+//
+// This class attaches interface to implementation and handles reference
+// counting, delegating most methods to ImplToFst.
+template <class A>
+class DeterminizeFst : public ImplToFst<internal::DeterminizeFstImplBase<A>> {
+ public:
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Store = DefaultCacheStore<Arc>;
+  using State = typename Store::State;
+  using Impl = internal::DeterminizeFstImplBase<Arc>;
+
+  friend class ArcIterator<DeterminizeFst<Arc>>;
+  friend class StateIterator<DeterminizeFst<Arc>>;
+
+  template <class B, GallicType G, class CommonDivisor, class Filter,
+            class StateTable>
+  friend class DeterminizeFstImpl;
+
+  explicit DeterminizeFst(const Fst<A> &fst)
+      : ImplToFst<Impl>(CreateImpl(fst)) {}
+
+  template <class CommonDivisor, class Filter, class StateTable>
+  DeterminizeFst(
+      const Fst<Arc> &fst,
+      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable>
+          &opts =
+              DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable>())
+      : ImplToFst<Impl>(CreateImpl(fst, opts)) {}
+
+  // This acceptor-only version additionally computes the distance to final
+  // states in the output if provided with those distances for the input; this
+  // is useful for e.g., computing the k-shortest unique paths.
+  template <class CommonDivisor, class Filter, class StateTable>
+  DeterminizeFst(
+      const Fst<Arc> &fst, const std::vector<Weight> *in_dist,
+      std::vector<Weight> *out_dist,
+      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable>
+          &opts =
+              DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable>())
+      : ImplToFst<Impl>(
+            std::make_shared<internal::DeterminizeFsaImpl<Arc, CommonDivisor,
+                                                          Filter, StateTable>>(
+                fst, in_dist, out_dist, opts)) {
+    if (!fst.Properties(kAcceptor, true)) {
+      FSTERROR() << "DeterminizeFst: "
+                 << "Distance to final states computed for acceptors only";
+      GetMutableImpl()->SetProperties(kError, kError);
+    }
+  }
+
+  // See Fst<>::Copy() for doc.
+  DeterminizeFst(const DeterminizeFst<Arc> &fst, bool safe = false)
+      : ImplToFst<Impl>(safe ? std::shared_ptr<Impl>(fst.GetImpl()->Copy())
+                             : fst.GetSharedImpl()) {}
+
+  // Get a copy of this DeterminizeFst. See Fst<>::Copy() for further doc.
+  DeterminizeFst<Arc> *Copy(bool safe = false) const override {
+    return new DeterminizeFst<Arc>(*this, safe);
+  }
+
+  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
+    GetMutableImpl()->InitArcIterator(s, data);
+  }
+
+ private:
+  using ImplToFst<Impl>::GetImpl;
+  using ImplToFst<Impl>::GetMutableImpl;
+
+  static std::shared_ptr<Impl> CreateImpl(const Fst<Arc> &fst) {
+    using D = DefaultCommonDivisor<Weight>;
+    using F = DefaultDeterminizeFilter<Arc>;
+    using T = DefaultDeterminizeStateTable<Arc, typename F::FilterState>;
+    const DeterminizeFstOptions<Arc, D, F, T> opts;
+    return CreateImpl(fst, opts);
+  }
+
+  template <class CommonDivisor, class Filter, class StateTable>
+  static std::shared_ptr<Impl> CreateImpl(
+      const Fst<Arc> &fst,
+      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable>
+          &opts) {
+    if (fst.Properties(kAcceptor, true)) {
+      // Calls implementation for acceptors.
+      return std::make_shared<
+          internal::DeterminizeFsaImpl<Arc, CommonDivisor, Filter, StateTable>>(
+          fst, nullptr, nullptr, opts);
+    } else if (opts.type == DETERMINIZE_DISAMBIGUATE) {
+      auto rv = std::make_shared<internal::DeterminizeFstImpl<
+          Arc, GALLIC_MIN, CommonDivisor, Filter, StateTable>>(fst, opts);
+      if (!(Weight::Properties() & kPath)) {
+        FSTERROR() << "DeterminizeFst: Weight needs to have the "
+                   << "path property to disambiguate output: "
+                   << Weight::Type();
+        rv->SetProperties(kError, kError);
+      }
+      // Calls disambiguating implementation for non-functional transducers.
+      return rv;
+    } else if (opts.type == DETERMINIZE_FUNCTIONAL) {
+      // Calls implementation for functional transducers.
+      return std::make_shared<internal::DeterminizeFstImpl<
+          Arc, GALLIC_RESTRICT, CommonDivisor, Filter, StateTable>>(fst, opts);
+    } else {  // opts.type == DETERMINIZE_NONFUNCTIONAL
+      // Calls implementation for non functional transducers;
+      return std::make_shared<internal::DeterminizeFstImpl<
+          Arc, GALLIC, CommonDivisor, Filter, StateTable>>(fst, opts);
+    }
+  }
+
+  DeterminizeFst &operator=(const DeterminizeFst &) = delete;
+};
+
+namespace internal {
+
+// Initialization of transducer determinization implementation, which is defined
+// after DeterminizeFst since it calls it.
+template <class Arc, GallicType G, class D, class Filter, class T>
+void DeterminizeFstImpl<Arc, G, D, Filter, T>::Init(const Fst<Arc> &fst, Filter *filter) {
+  // Mapper to an acceptor.
+  const ToFst to_fst(fst, ToMapper());
+  auto *to_filter = filter ? new ToFilter(to_fst, filter) : nullptr;
+  // This recursive call terminates since it is to a (non-recursive)
+  // different constructor.
+  const CacheOptions copts(GetCacheGc(), GetCacheLimit());
+  const DeterminizeFstOptions<ToArc, ToCommonDivisor, ToFilter, ToStateTable>
+      dopts(copts, delta_, 0, DETERMINIZE_FUNCTIONAL, false, to_filter);
+  // Uses acceptor-only constructor to avoid template recursion.
+  const DeterminizeFst<ToArc> det_fsa(to_fst, nullptr, nullptr, dopts);
+  // Mapper back to transducer.
+  const FactorWeightOptions<ToArc> fopts(
+      CacheOptions(true, 0), delta_, kFactorFinalWeights, subsequential_label_,
+      subsequential_label_, increment_subsequential_label_,
+      increment_subsequential_label_);
+  const FactorWeightFst<ToArc, FactorIterator> factored_fst(det_fsa, fopts);
+  from_fst_.reset(new FromFst(factored_fst, FromMapper(subsequential_label_)));
+}
+
+}  // namespace internal
+
+// Specialization for DeterminizeFst.
+template <class Arc>
+class StateIterator<DeterminizeFst<Arc>>
+    : public CacheStateIterator<DeterminizeFst<Arc>> {
+ public:
+  explicit StateIterator(const DeterminizeFst<Arc> &fst)
+      : CacheStateIterator<DeterminizeFst<Arc>>(fst, fst.GetMutableImpl()) {}
+};
+
+// Specialization for DeterminizeFst.
+template <class Arc>
+class ArcIterator<DeterminizeFst<Arc>>
+    : public CacheArcIterator<DeterminizeFst<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const DeterminizeFst<Arc> &fst, StateId s)
+      : CacheArcIterator<DeterminizeFst<Arc>>(fst.GetMutableImpl(), s) {
+    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
+  }
+};
+
+template <class Arc>
+inline void DeterminizeFst<Arc>::InitStateIterator(
+    StateIteratorData<Arc> *data) const {
+  data->base = new StateIterator<DeterminizeFst<Arc>>(*this);
+}
+
+// Useful aliases when using StdArc.
+using StdDeterminizeFst = DeterminizeFst<StdArc>;
+
+template <class Arc>
+struct DeterminizeOptions {
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  float delta;                // Quantization delta for subset weights.
+  Weight weight_threshold;    // Pruning weight threshold.
+  StateId state_threshold;    // Pruning state threshold.
+  Label subsequential_label;  // Label used for residual final output.
+  DeterminizeType type;
+  bool increment_subsequential_label;  // When creating several subsequential
+                                       // arcs at a given state, make their
+                                       // label distinct by incrementation?
+
+  explicit DeterminizeOptions(float delta = kDelta,
+                              Weight weight_threshold = Weight::Zero(),
+                              StateId state_threshold = kNoStateId,
+                              Label subsequential_label = 0,
+                              DeterminizeType type = DETERMINIZE_FUNCTIONAL,
+                              bool increment_subsequential_label = false)
+      : delta(delta),
+        weight_threshold(std::move(weight_threshold)),
+        state_threshold(state_threshold),
+        subsequential_label(subsequential_label),
+        type(type),
+        increment_subsequential_label(increment_subsequential_label) {}
+};
+
+// Determinizes a weighted transducer. This version writes the
+// determinized Fst to an output MutableFst. The result will be an
+// equivalent FST that has the property that no state has two
+// transitions with the same input label. For this algorithm, epsilon
+// transitions are treated as regular symbols (cf. RmEpsilon).
+//
+// The transducer must be functional. The weights must be (weakly)
+// left divisible (valid for TropicalWeight and LogWeight).
+//
+// Complexity:
+//
+//   Determinizable: exponential (polynomial in the size of the output)
+//   Non-determinizable: does not terminate
+//
+// The determinizable automata include all unweighted and all acyclic input.
+template <class Arc>
+void Determinize(
+    const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
+    const DeterminizeOptions<Arc> &opts = DeterminizeOptions<Arc>()) {
+  using Weight = typename Arc::Weight;
+  DeterminizeFstOptions<Arc> nopts;
+  nopts.delta = opts.delta;
+  nopts.subsequential_label = opts.subsequential_label;
+  nopts.type = opts.type;
+  nopts.increment_subsequential_label = opts.increment_subsequential_label;
+  nopts.gc_limit = 0;  // Caches only the last state for fastest copy.
+  if (opts.weight_threshold != Weight::Zero() ||
+      opts.state_threshold != kNoStateId) {
+    if (ifst.Properties(kAcceptor, false)) {
+      std::vector<Weight> idistance;
+      std::vector<Weight> odistance;
+      ShortestDistance(ifst, &idistance, true);
+      DeterminizeFst<Arc> dfst(ifst, &idistance, &odistance, nopts);
+      PruneOptions<Arc, AnyArcFilter<Arc>> popts(
+          opts.weight_threshold, opts.state_threshold, AnyArcFilter<Arc>(),
+          &odistance);
+      Prune(dfst, ofst, popts);
+    } else {
+      *ofst = DeterminizeFst<Arc>(ifst, nopts);
+      Prune(ofst, opts.weight_threshold, opts.state_threshold);
+    }
+  } else {
+    *ofst = DeterminizeFst<Arc>(ifst, nopts);
+  }
+}
+
+}  // namespace fst
+
+#endif  // FST_DETERMINIZE_H_
diff --git a/projects/llm_framework/include/fst/dfs-visit.h b/projects/llm_framework/include/fst/dfs-visit.h
new file mode 100644
index 00000000..a7b18a6c
--- /dev/null
+++ b/projects/llm_framework/include/fst/dfs-visit.h
@@ -0,0 +1,202 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Depth-first search visitation. See visit.h for more general search queue
+// disciplines.
+
+#ifndef FST_DFS_VISIT_H_
+#define FST_DFS_VISIT_H_
+
+#include <stack>
+#include <vector>
+
+#include <fst/arcfilter.h>
+#include <fst/fst.h>
+
+
+namespace fst {
+
+// Visitor Interface: class determining actions taken during a depth-first
+// search-style visit. If any of the boolean member functions return false, the
+// DFS is aborted by first calling FinishState() on all currently grey states
+// and then calling FinishVisit().
+//
+// This is similar to the more general visitor interface in visit.h, except
+// that FinishState returns additional information appropriate only for a DFS
+// and some methods names here are better suited to a DFS.
+//
+// template <class Arc>
+// class Visitor {
+//  public:
+//   using StateId = typename Arc::StateId;
+//
+//   Visitor(T *return_data);
+//
+//   // Invoked before DFS visit.
+//   void InitVisit(const Fst<Arc> &fst);
+//
+//   // Invoked when state discovered (2nd arg is DFS tree root).
+//   bool InitState(StateId s, StateId root);
+//
+//   // Invoked when tree arc to white/undiscovered state examined.
+//   bool TreeArc(StateId s, const Arc &arc);
+//
+//   // Invoked when back arc to grey/unfinished state examined.
+//   bool BackArc(StateId s, const Arc &arc);
+//
+//   // Invoked when forward or cross arc to black/finished state examined.
+//   bool ForwardOrCrossArc(StateId s, const Arc &arc);
+//
+//   // Invoked when state finished ('s' is tree root, 'parent' is kNoStateId,
+//   // and 'arc' is nullptr).
+//   void FinishState(StateId s, StateId parent, const Arc *arc);
+//
+//   // Invoked after DFS visit.
+//   void FinishVisit();
+// };
+
+namespace internal {
+
+// An FST state's DFS stack state.
+template <class FST>
+struct DfsState {
+  using Arc = typename FST::Arc;
+  using StateId = typename Arc::StateId;
+
+  DfsState(const FST &fst, StateId s) : state_id(s), arc_iter(fst, s) {}
+
+  void *operator new(size_t size, MemoryPool<DfsState<FST>> *pool) {
+    return pool->Allocate();
+  }
+
+  static void Destroy(DfsState<FST> *dfs_state,
+                      MemoryPool<DfsState<FST>> *pool) {
+    if (dfs_state) {
+      dfs_state->~DfsState<FST>();
+      pool->Free(dfs_state);
+    }
+  }
+
+  StateId state_id;           // FST state.
+  ArcIterator<FST> arc_iter;  // The corresponding arcs.
+};
+
+}  // namespace internal
+
+// Performs depth-first visitation. Visitor class argument determines actions
+// and contains any return data. ArcFilter determines arcs that are considered.
+// If 'access_only' is true, performs visitation only to states accessible from
+// the initial state.
+//
+// Note this is similar to Visit() in visit.h called with a LIFO queue, except
+// this version has a Visitor class specialized and augmented for a DFS.
+template <class FST, class Visitor, class ArcFilter>
+void DfsVisit(const FST &fst, Visitor *visitor, ArcFilter filter,
+              bool access_only = false) {
+  visitor->InitVisit(fst);
+  const auto start = fst.Start();
+  if (start == kNoStateId) {
+    visitor->FinishVisit();
+    return;
+  }
+  // An FST state's DFS status
+  static constexpr uint8 kDfsWhite = 0;  // Undiscovered.
+  static constexpr uint8 kDfsGrey = 1;   // Discovered but unfinished.
+  static constexpr uint8 kDfsBlack = 2;  // Finished.
+  std::vector<uint8> state_color;
+  std::stack<internal::DfsState<FST> *> state_stack;  // DFS execution stack.
+  MemoryPool<internal::DfsState<FST>> state_pool;     // Pool for DFSStates.
+  auto nstates = start + 1;  // Number of known states in general case.
+  bool expanded = false;
+  if (fst.Properties(kExpanded, false)) {  // Tests if expanded case, then
+    nstates = CountStates(fst);            // uses ExpandedFst::NumStates().
+    expanded = true;
+  }
+  state_color.resize(nstates, kDfsWhite);
+  StateIterator<FST> siter(fst);
+  // Continue DFS while true.
+  bool dfs = true;
+  // Iterate over trees in DFS forest.
+  for (auto root = start; dfs && root < nstates;) {
+    state_color[root] = kDfsGrey;
+    state_stack.push(new (&state_pool) internal::DfsState<FST>(fst, root));
+    dfs = visitor->InitState(root, root);
+    while (!state_stack.empty()) {
+      auto *dfs_state = state_stack.top();
+      const auto s = dfs_state->state_id;
+      if (s >= static_cast<typename FST::Arc::StateId>(state_color.size())) {
+        nstates = s + 1;
+        state_color.resize(nstates, kDfsWhite);
+      }
+      ArcIterator<FST> &aiter = dfs_state->arc_iter;
+      if (!dfs || aiter.Done()) {
+        state_color[s] = kDfsBlack;
+        internal::DfsState<FST>::Destroy(dfs_state, &state_pool);
+        state_stack.pop();
+        if (!state_stack.empty()) {
+          auto *parent_state = state_stack.top();
+          auto &piter = parent_state->arc_iter;
+          visitor->FinishState(s, parent_state->state_id, &piter.Value());
+          piter.Next();
+        } else {
+          visitor->FinishState(s, kNoStateId, nullptr);
+        }
+        continue;
+      }
+      const auto &arc = aiter.Value();
+      if (arc.nextstate >= state_color.size()) {
+        nstates = arc.nextstate + 1;
+        state_color.resize(nstates, kDfsWhite);
+      }
+      if (!filter(arc)) {
+        aiter.Next();
+        continue;
+      }
+      const auto next_color = state_color[arc.nextstate];
+      switch (next_color) {
+        default:
+        case kDfsWhite:
+          dfs = visitor->TreeArc(s, arc);
+          if (!dfs) break;
+          state_color[arc.nextstate] = kDfsGrey;
+          state_stack.push(new (&state_pool)
+                               internal::DfsState<FST>(fst, arc.nextstate));
+          dfs = visitor->InitState(arc.nextstate, root);
+          break;
+        case kDfsGrey:
+          dfs = visitor->BackArc(s, arc);
+          aiter.Next();
+          break;
+        case kDfsBlack:
+          dfs = visitor->ForwardOrCrossArc(s, arc);
+          aiter.Next();
+          break;
+      }
+    }
+    if (access_only) break;
+    // Finds next tree root.
+    for (root = root == start ? 0 : root + 1;
+         root < nstates && state_color[root] != kDfsWhite; ++root) {
+    }
+    // Checks for a state beyond the largest known state.
+    if (!expanded && root == nstates) {
+      for (; !siter.Done(); siter.Next()) {
+        if (siter.Value() == nstates) {
+          ++nstates;
+          state_color.push_back(kDfsWhite);
+          break;
+        }
+      }
+    }
+  }
+  visitor->FinishVisit();
+}
+
+template <class Arc, class Visitor>
+void DfsVisit(const Fst<Arc> &fst, Visitor *visitor) {
+  DfsVisit(fst, visitor, AnyArcFilter<Arc>());
+}
+
+}  // namespace fst
+
+#endif  // FST_DFS_VISIT_H_
diff --git a/projects/llm_framework/include/fst/difference.h b/projects/llm_framework/include/fst/difference.h
new file mode 100644
index 00000000..f073b0e1
--- /dev/null
+++ b/projects/llm_framework/include/fst/difference.h
@@ -0,0 +1,205 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to compute the difference between two FSAs.
+
+#ifndef FST_DIFFERENCE_H_
+#define FST_DIFFERENCE_H_
+
+#include <memory>
+
+
+#include <fst/cache.h>
+#include <fst/complement.h>
+#include <fst/compose.h>
+
+
+namespace fst {
+
+template <class Arc, class M = Matcher<Fst<Arc>>,
+          class Filter = SequenceComposeFilter<M>,
+          class StateTable =
+              GenericComposeStateTable<Arc, typename Filter::FilterState>>
+struct DifferenceFstOptions
+    : public ComposeFstOptions<Arc, M, Filter, StateTable> {
+  explicit DifferenceFstOptions(const CacheOptions &opts = CacheOptions(),
+                                M *matcher1 = nullptr, M *matcher2 = nullptr,
+                                Filter *filter = nullptr,
+                                StateTable *state_table = nullptr)
+      : ComposeFstOptions<Arc, M, Filter, StateTable>(opts, matcher1, matcher2,
+                                                      filter, state_table) {}
+};
+
+// Computes the difference between two FSAs. This version is a delayed FST.
+// Only strings that are in the first automaton but not in second are retained
+// in the result.
+//
+// The first argument must be an acceptor; the second argument must be an
+// unweighted, epsilon-free, deterministic acceptor. One of the arguments must
+// be label-sorted.
+//
+// Complexity: same as ComposeFst.
+//
+// Caveats: same as ComposeFst.
+template <class A>
+class DifferenceFst : public ComposeFst<A> {
+ public:
+  using Arc = A;
+  using Weight = typename Arc::Weight;
+  using StateId = typename Arc::StateId;
+
+  using ComposeFst<Arc>::CreateBase1;
+
+  // A - B = A ^ B'.
+  DifferenceFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+                const CacheOptions &opts = CacheOptions())
+      : ComposeFst<Arc>(CreateDifferenceImplWithCacheOpts(fst1, fst2, opts)) {
+    if (!fst1.Properties(kAcceptor, true)) {
+      FSTERROR() << "DifferenceFst: 1st argument not an acceptor";
+      GetImpl()->SetProperties(kError, kError);
+    }
+  }
+
+  template <class Matcher, class Filter, class StateTable>
+  DifferenceFst(
+      const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+      const DifferenceFstOptions<Arc, Matcher, Filter, StateTable> &opts)
+      : ComposeFst<Arc>(
+            CreateDifferenceImplWithDifferenceOpts(fst1, fst2, opts)) {
+    if (!fst1.Properties(kAcceptor, true)) {
+      FSTERROR() << "DifferenceFst: 1st argument not an acceptor";
+      GetImpl()->SetProperties(kError, kError);
+    }
+  }
+
+  // See Fst<>::Copy() for doc.
+  DifferenceFst(const DifferenceFst<Arc> &fst, bool safe = false)
+      : ComposeFst<Arc>(fst, safe) {}
+
+  // Get a copy of this DifferenceFst. See Fst<>::Copy() for further doc.
+  DifferenceFst<Arc> *Copy(bool safe = false) const override {
+    return new DifferenceFst<Arc>(*this, safe);
+  }
+
+ private:
+  using Impl = internal::ComposeFstImplBase<Arc>;
+  using ImplToFst<Impl>::GetImpl;
+
+  static std::shared_ptr<Impl> CreateDifferenceImplWithCacheOpts(
+      const Fst<Arc> &fst1, const Fst<Arc> &fst2, const CacheOptions &opts) {
+    using RM = RhoMatcher<Matcher<Fst<A>>>;
+    ComplementFst<Arc> cfst(fst2);
+    ComposeFstOptions<A, RM> copts(
+        CacheOptions(), new RM(fst1, MATCH_NONE),
+        new RM(cfst, MATCH_INPUT, ComplementFst<Arc>::kRhoLabel));
+    return CreateBase1(fst1, cfst, copts);
+  }
+
+  template <class Matcher, class Filter, class StateTable>
+  static std::shared_ptr<Impl> CreateDifferenceImplWithDifferenceOpts(
+      const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+      const DifferenceFstOptions<Arc, Matcher, Filter, StateTable> &opts) {
+    using RM = RhoMatcher<Matcher>;
+    ComplementFst<Arc> cfst(fst2);
+    ComposeFstOptions<Arc, RM> copts(opts);
+    copts.matcher1 = new RM(fst1, MATCH_NONE, kNoLabel, MATCHER_REWRITE_ALWAYS,
+                            opts.matcher1);
+    copts.matcher2 = new RM(cfst, MATCH_INPUT, ComplementFst<Arc>::kRhoLabel,
+                            MATCHER_REWRITE_ALWAYS, opts.matcher2);
+    return CreateBase1(fst1, cfst, copts);
+  }
+};
+
+// Specialization for DifferenceFst.
+template <class Arc>
+class StateIterator<DifferenceFst<Arc>>
+    : public StateIterator<ComposeFst<Arc>> {
+ public:
+  explicit StateIterator(const DifferenceFst<Arc> &fst)
+      : StateIterator<ComposeFst<Arc>>(fst) {}
+};
+
+// Specialization for DifferenceFst.
+template <class Arc>
+class ArcIterator<DifferenceFst<Arc>> : public ArcIterator<ComposeFst<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const DifferenceFst<Arc> &fst, StateId s)
+      : ArcIterator<ComposeFst<Arc>>(fst, s) {}
+};
+
+using DifferenceOptions = ComposeOptions;
+
+// Useful alias when using StdArc.
+using StdDifferenceFst = DifferenceFst<StdArc>;
+
+using DifferenceOptions = ComposeOptions;
+
+// Computes the difference between two FSAs. This version writes the difference
+// to an output MutableFst. Only strings that are in the first automaton but not
+// in the second are retained in the result.
+//
+// The first argument must be an acceptor; the second argument must be an
+// unweighted, epsilon-free, deterministic acceptor. One of the arguments must
+// be label-sorted.
+//
+// Complexity: same as Compose.
+//
+// Caveats: same as Compose.
+template <class Arc>
+void Difference(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
+                MutableFst<Arc> *ofst,
+                const DifferenceOptions &opts = DifferenceOptions()) {
+  using M = Matcher<Fst<Arc>>;
+  // In each case, we cache only the last state for fastest copy.
+  switch (opts.filter_type) {
+    case AUTO_FILTER: {
+      CacheOptions nopts;
+      nopts.gc_limit = 0;
+      *ofst = DifferenceFst<Arc>(ifst1, ifst2, nopts);
+      break;
+    }
+    case SEQUENCE_FILTER: {
+      DifferenceFstOptions<Arc> dopts;
+      dopts.gc_limit = 0;
+      *ofst = DifferenceFst<Arc>(ifst1, ifst2, dopts);
+      break;
+    }
+    case ALT_SEQUENCE_FILTER: {
+      DifferenceFstOptions<Arc, M, AltSequenceComposeFilter<M>> dopts;
+      dopts.gc_limit = 0;
+      *ofst = DifferenceFst<Arc>(ifst1, ifst2, dopts);
+      break;
+    }
+    case MATCH_FILTER: {
+      DifferenceFstOptions<Arc, M, MatchComposeFilter<M>> dopts;
+      dopts.gc_limit = 0;
+      *ofst = DifferenceFst<Arc>(ifst1, ifst2, dopts);
+      break;
+    }
+    case NO_MATCH_FILTER: {
+      DifferenceFstOptions<Arc, M, NoMatchComposeFilter<M>> dopts;
+      dopts.gc_limit = 0;
+      *ofst = DifferenceFst<Arc>(ifst1, ifst2, dopts);
+      break;
+    }
+    case NULL_FILTER: {
+      DifferenceFstOptions<Arc, M, NullComposeFilter<M>> dopts;
+      dopts.gc_limit = 0;
+      *ofst = DifferenceFst<Arc>(ifst1, ifst2, dopts);
+      break;
+    }
+    case TRIVIAL_FILTER: {
+      DifferenceFstOptions<Arc, M, TrivialComposeFilter<M>> dopts;
+      dopts.gc_limit = 0;
+      *ofst = DifferenceFst<Arc>(ifst1, ifst2, dopts);
+      break;
+    }
+  }
+  if (opts.connect) Connect(ofst);
+}
+
+}  // namespace fst
+
+#endif  // FST_DIFFERENCE_H_
diff --git a/projects/llm_framework/include/fst/disambiguate.h b/projects/llm_framework/include/fst/disambiguate.h
new file mode 100644
index 00000000..7e3ba37c
--- /dev/null
+++ b/projects/llm_framework/include/fst/disambiguate.h
@@ -0,0 +1,564 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes to disambiguate an FST.
+
+#ifndef FST_DISAMBIGUATE_H_
+#define FST_DISAMBIGUATE_H_
+
+#include <list>
+#include <map>
+#include <set>
+#include <utility>
+#include <vector>
+
+#include <fst/arcsort.h>
+#include <fst/compose.h>
+#include <fst/connect.h>
+#include <fst/determinize.h>
+#include <fst/dfs-visit.h>
+#include <fst/project.h>
+#include <fst/prune.h>
+#include <fst/state-map.h>
+#include <fst/state-table.h>
+#include <fst/union-find.h>
+#include <fst/verify.h>
+
+
+namespace fst {
+
+template <class Arc>
+struct DisambiguateOptions : public DeterminizeOptions<Arc> {
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  explicit DisambiguateOptions(float delta = kDelta,
+                               Weight weight = Weight::Zero(),
+                               StateId n = kNoStateId, Label label = 0)
+      : DeterminizeOptions<Arc>(delta, std::move(weight), n, label,
+                                DETERMINIZE_FUNCTIONAL) {}
+};
+
+namespace internal {
+
+// A determinization filter based on a subset element relation. The relation is
+// assumed to be reflexive and symmetric.
+template <class Arc, class Relation>
+class RelationDeterminizeFilter {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FilterState = IntegerFilterState<StateId>;
+  using StateTuple = DeterminizeStateTuple<Arc, FilterState>;
+  using Subset = typename StateTuple::Subset;
+  using Element = typename StateTuple::Element;
+  using LabelMap = std::multimap<Label, DeterminizeArc<StateTuple>>;
+
+  // This is needed (e.g.) to go into the gallic domain for transducers; there
+  // is no need to rebind the relation since its use here only depends on the
+  // state IDs.
+  template <class A>
+  struct rebind {
+    using Other = RelationDeterminizeFilter<A, Relation>;
+  };
+
+  explicit RelationDeterminizeFilter(const Fst<Arc> &fst)
+      : fst_(fst.Copy()), r_(new Relation()), s_(kNoStateId), head_(nullptr) {}
+
+  // Ownership of the relation is given to this class.
+  RelationDeterminizeFilter(const Fst<Arc> &fst, Relation *r)
+      : fst_(fst.Copy()), r_(r), s_(kNoStateId), head_(0) {}
+
+  // Ownership of the relation is given to this class.
+  RelationDeterminizeFilter(const Fst<Arc> &fst, Relation *r,
+                            std::vector<StateId> *head)
+      : fst_(fst.Copy()), r_(r), s_(kNoStateId), head_(head) {}
+
+  // This is needed, e.g., to go into the gallic domain for transducers.
+  // Ownership of the templated filter argument is given to this class.
+  template <class Filter>
+  RelationDeterminizeFilter(const Fst<Arc> &fst, Filter *filter)
+      : fst_(fst.Copy()),
+        r_(new Relation(filter->GetRelation())),
+        s_(kNoStateId),
+        head_(filter->GetHeadStates()) {
+    delete filter;
+  }
+
+  // Copy constructor; the FST can be passed if it has been deep-copied.
+  RelationDeterminizeFilter(const RelationDeterminizeFilter &filter,
+                            const Fst<Arc> *fst = nullptr)
+      : fst_(fst ? fst->Copy() : filter.fst_->Copy()),
+        r_(new Relation(*filter.r_)),
+        s_(kNoStateId),
+        head_() {}
+
+  FilterState Start() const { return FilterState(fst_->Start()); }
+
+  void SetState(StateId s, const StateTuple &tuple) {
+    if (s_ != s) {
+      s_ = s;
+      tuple_ = &tuple;
+      const auto head = tuple.filter_state.GetState();
+      is_final_ = fst_->Final(head) != Weight::Zero();
+      if (head_) {
+        if (head_->size() <= s) head_->resize(s + 1, kNoStateId);
+        (*head_)[s] = head;
+      }
+    }
+  }
+
+  // Filters transition, possibly modifying label map. Returns true if arc is
+  // added to label map.
+  bool FilterArc(const Arc &arc, const Element &src_element,
+                 const Element &dest_element, LabelMap *label_map) const;
+
+  // Filters super-final transition, returning new final weight.
+  Weight FilterFinal(const Weight final_weight, const Element &element) const {
+    return is_final_ ? final_weight : Weight::Zero();
+  }
+
+  static uint64 Properties(uint64 props) {
+    return props & ~(kIDeterministic | kODeterministic);
+  }
+
+  const Relation &GetRelation() { return *r_; }
+
+  std::vector<StateId> *GetHeadStates() { return head_; }
+
+ private:
+  // Pairs arc labels with state tuples with possible heads and empty subsets.
+  void InitLabelMap(LabelMap *label_map) const;
+
+  std::unique_ptr<Fst<Arc>> fst_;  // Input FST.
+  std::unique_ptr<Relation> r_;    // Relation compatible with inv. trans. fnc.
+  StateId s_;                      // Current state.
+  const StateTuple *tuple_;        // Current tuple.
+  bool is_final_;                  // Is the current head state final?
+  std::vector<StateId> *head_;     // Head state for a given state,
+                                   // owned by the Disambiguator.
+};
+
+template <class Arc, class Relation>
+bool RelationDeterminizeFilter<Arc, Relation>::FilterArc(
+    const Arc &arc, const Element &src_element, const Element &dest_element,
+    LabelMap *label_map) const {
+  bool added = false;
+  if (label_map->empty()) InitLabelMap(label_map);
+  // Adds element to state tuple if element state is related to tuple head.
+  for (auto liter = label_map->lower_bound(arc.ilabel);
+       liter != label_map->end() && liter->first == arc.ilabel; ++liter) {
+    auto *dest_tuple = liter->second.dest_tuple;
+    const auto dest_head = dest_tuple->filter_state.GetState();
+    if ((*r_)(dest_element.state_id, dest_head)) {
+      dest_tuple->subset.push_front(dest_element);
+      added = true;
+    }
+  }
+  return added;
+}
+
+template <class Arc, class Relation>
+void RelationDeterminizeFilter<Arc, Relation>::InitLabelMap(
+    LabelMap *label_map) const {
+  const auto src_head = tuple_->filter_state.GetState();
+  Label label = kNoLabel;
+  StateId nextstate = kNoStateId;
+  for (ArcIterator<Fst<Arc>> aiter(*fst_, src_head); !aiter.Done();
+       aiter.Next()) {
+    const auto &arc = aiter.Value();
+    // Continues if multiarc.
+    if (arc.ilabel == label && arc.nextstate == nextstate) continue;
+    DeterminizeArc<StateTuple> det_arc(arc);
+    det_arc.dest_tuple->filter_state = FilterState(arc.nextstate);
+    label_map->insert(std::make_pair(arc.ilabel, det_arc));
+    label = arc.ilabel;
+    nextstate = arc.nextstate;
+  }
+}
+
+// Helper class to disambiguate an FST via Disambiguate().
+template <class Arc>
+class Disambiguator {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  // IDs arcs with state ID and arc position. Arc position -1 indicates final
+  // (super-final transition).
+  using ArcId = std::pair<StateId, ssize_t>;
+
+  Disambiguator() : error_(false) {}
+
+  void Disambiguate(
+      const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
+      const DisambiguateOptions<Arc> &opts = DisambiguateOptions<Arc>()) {
+    VectorFst<Arc> sfst(ifst);
+    Connect(&sfst);
+    ArcSort(&sfst, ArcCompare());
+    PreDisambiguate(sfst, ofst, opts);
+    ArcSort(ofst, ArcCompare());
+    FindAmbiguities(*ofst);
+    RemoveSplits(ofst);
+    MarkAmbiguities();
+    RemoveAmbiguities(ofst);
+    if (error_) ofst->SetProperties(kError, kError);
+  }
+
+ private:
+  // Comparison functor for comparing input labels and next states of arcs. This
+  // sort order facilitates the predisambiguation.
+  class ArcCompare {
+   public:
+    bool operator()(const Arc &arc1, const Arc &arc2) const {
+      return arc1.ilabel < arc2.ilabel ||
+             (arc1.ilabel == arc2.ilabel && arc1.nextstate < arc2.nextstate);
+    }
+
+    uint64 Properties(uint64 props) const {
+      return (props & kArcSortProperties) | kILabelSorted |
+             (props & kAcceptor ? kOLabelSorted : 0);
+    }
+  };
+
+  // Comparison functor for comparing transitions represented by their arc ID.
+  // This sort order facilitates ambiguity detection.
+  class ArcIdCompare {
+   public:
+    explicit ArcIdCompare(const std::vector<StateId> &head) : head_(head) {}
+
+    bool operator()(const ArcId &a1, const ArcId &a2) const {
+      // Sort first by source head state...
+      const auto src1 = a1.first;
+      const auto src2 = a2.first;
+      const auto head1 = head_[src1];
+      const auto head2 = head_[src2];
+      if (head1 < head2) return true;
+      if (head2 < head1) return false;
+      // ...then by source state...
+      if (src1 < src2) return true;
+      if (src2 < src1) return false;
+      // ...then by position.
+      return a1.second < a2.second;
+    }
+
+   private:
+    const std::vector<StateId> &head_;
+  };
+
+  // A relation that determines if two states share a common future.
+  class CommonFuture {
+   public:
+    using StateTable = GenericComposeStateTable<Arc, TrivialFilterState>;
+    using StateTuple = typename StateTable::StateTuple;
+
+    // Needed for compilation with DeterminizeRelationFilter.
+    CommonFuture() {
+      FSTERROR() << "Disambiguate::CommonFuture: FST not provided";
+    }
+
+    explicit CommonFuture(const Fst<Arc> &ifst) {
+      using M = Matcher<Fst<Arc>>;
+      ComposeFstOptions<Arc, M, NullComposeFilter<M>> opts;
+      // Ensures composition is between acceptors.
+      const bool trans = ifst.Properties(kNotAcceptor, true);
+      const auto *fsa =
+          trans ? new ProjectFst<Arc>(ifst, PROJECT_INPUT) : &ifst;
+      opts.state_table = new StateTable(*fsa, *fsa);
+      const ComposeFst<Arc> cfst(*fsa, *fsa, opts);
+      std::vector<bool> coaccess;
+      uint64 props = 0;
+      SccVisitor<Arc> scc_visitor(nullptr, nullptr, &coaccess, &props);
+      DfsVisit(cfst, &scc_visitor);
+      for (StateId s = 0; s < coaccess.size(); ++s) {
+        if (coaccess[s]) {
+          related_.insert(opts.state_table->Tuple(s).StatePair());
+        }
+      }
+      if (trans) delete fsa;
+    }
+
+    bool operator()(const StateId s1, StateId s2) const {
+      return related_.count(std::make_pair(s1, s2)) > 0;
+    }
+
+   private:
+    // States s1 and s2 resp. are in this relation iff they there is a
+    // path from s1 to a final state that has the same label as some
+    // path from s2 to a final state.
+    std::set<std::pair<StateId, StateId>> related_;
+  };
+
+  using ArcIdMap = std::multimap<ArcId, ArcId, ArcIdCompare>;
+
+  // Inserts candidate into the arc ID map.
+  inline void InsertCandidate(StateId s1, StateId s2, const ArcId &a1,
+                              const ArcId &a2) {
+    candidates_->insert(head_[s1] > head_[s2] ? std::make_pair(a1, a2)
+                                              : std::make_pair(a2, a1));
+  }
+
+  // Returns the arc corresponding to ArcId a.
+  static Arc GetArc(const Fst<Arc> &fst, ArcId aid) {
+    if (aid.second == -1) {  // Returns super-final transition.
+      return Arc(kNoLabel, kNoLabel, fst.Final(aid.first), kNoStateId);
+    } else {
+      ArcIterator<Fst<Arc>> aiter(fst, aid.first);
+      aiter.Seek(aid.second);
+      return aiter.Value();
+    }
+  }
+
+  // Outputs an equivalent FST whose states are subsets of states that have a
+  // future path in common.
+  void PreDisambiguate(const ExpandedFst<Arc> &ifst, MutableFst<Arc> *ofst,
+                       const DisambiguateOptions<Arc> &opts);
+
+  // Finds transitions that are ambiguous candidates in the result of
+  // PreDisambiguate.
+  void FindAmbiguities(const ExpandedFst<Arc> &fst);
+
+  // Finds transition pairs that are ambiguous candidates from two specified
+  // source states.
+  void FindAmbiguousPairs(const ExpandedFst<Arc> &fst, StateId s1, StateId s2);
+
+  // Marks ambiguous transitions to be removed.
+  void MarkAmbiguities();
+
+  // Deletes spurious ambiguous transitions (due to quantization).
+  void RemoveSplits(MutableFst<Arc> *ofst);
+
+  // Deletes actual ambiguous transitions.
+  void RemoveAmbiguities(MutableFst<Arc> *ofst);
+
+  // States s1 and s2 are in this relation iff there is a path from the initial
+  // state to s1 that has the same label as some path from the initial state to
+  // s2. We store only state pairs s1, s2 such that s1 <= s2.
+  std::set<std::pair<StateId, StateId>> coreachable_;
+
+  // Queue of disambiguation-related states to be processed. We store only
+  // state pairs s1, s2 such that s1 <= s2.
+  std::list<std::pair<StateId, StateId>> queue_;
+
+  // Head state in the pre-disambiguation for a given state.
+  std::vector<StateId> head_;
+
+  // Maps from a candidate ambiguous arc A to each ambiguous candidate arc B
+  // with the same label and destination state as A, whose source state s' is
+  // coreachable with the source state s of A, and for which head(s') < head(s).
+  std::unique_ptr<ArcIdMap> candidates_;
+
+  // Set of ambiguous transitions to be removed.
+  std::set<ArcId> ambiguous_;
+
+  // States to merge due to quantization issues.
+  std::unique_ptr<UnionFind<StateId>> merge_;
+  // Marks error condition.
+  bool error_;
+
+  Disambiguator(const Disambiguator &) = delete;
+  Disambiguator &operator=(const Disambiguator &) = delete;
+};
+
+template <class Arc>
+void Disambiguator<Arc>::PreDisambiguate(const ExpandedFst<Arc> &ifst,
+                                         MutableFst<Arc> *ofst,
+                                         const DisambiguateOptions<Arc> &opts) {
+  using CommonDivisor = DefaultCommonDivisor<Weight>;
+  using Filter = RelationDeterminizeFilter<Arc, CommonFuture>;
+  // Subset elements with states s1 and s2 (resp.) are in this relation iff they
+  // there is a path from s1 to a final state that has the same label as some
+  // path from s2 to a final state.
+  auto *common_future = new CommonFuture(ifst);
+  DeterminizeFstOptions<Arc, CommonDivisor, Filter> nopts;
+  nopts.delta = opts.delta;
+  nopts.subsequential_label = opts.subsequential_label;
+  nopts.filter = new Filter(ifst, common_future, &head_);
+  // The filter takes ownership of 'common_future', and determinization takes
+  // ownership of the filter itself.
+  nopts.gc_limit = 0;  // Cache only the last state for fastest copy.
+  if (opts.weight_threshold != Weight::Zero() ||
+      opts.state_threshold != kNoStateId) {
+    /* TODO(riley): fails regression test; understand why
+    if (ifst.Properties(kAcceptor, true)) {
+      std::vector<Weight> idistance, odistance;
+      ShortestDistance(ifst, &idistance, true);
+      DeterminizeFst<Arc> dfst(ifst, &idistance, &odistance, nopts);
+      PruneOptions< Arc, AnyArcFilter<Arc>> popts(opts.weight_threshold,
+                                                   opts.state_threshold,
+                                                   AnyArcFilter<Arc>(),
+                                                   &odistance);
+      Prune(dfst, ofst, popts);
+      } else */ {
+      *ofst = DeterminizeFst<Arc>(ifst, nopts);
+      Prune(ofst, opts.weight_threshold, opts.state_threshold);
+    }
+  } else {
+    *ofst = DeterminizeFst<Arc>(ifst, nopts);
+  }
+  head_.resize(ofst->NumStates(), kNoStateId);
+}
+
+template <class Arc>
+void Disambiguator<Arc>::FindAmbiguities(const ExpandedFst<Arc> &fst) {
+  if (fst.Start() == kNoStateId) return;
+  candidates_.reset(new ArcIdMap(ArcIdCompare(head_)));
+  const auto start_pr = std::make_pair(fst.Start(), fst.Start());
+  coreachable_.insert(start_pr);
+  queue_.push_back(start_pr);
+  while (!queue_.empty()) {
+    const auto &pr = queue_.front();
+    const auto s1 = pr.first;
+    const auto s2 = pr.second;
+    queue_.pop_front();
+    FindAmbiguousPairs(fst, s1, s2);
+  }
+}
+
+template <class Arc>
+void Disambiguator<Arc>::FindAmbiguousPairs(const ExpandedFst<Arc> &fst,
+                                            StateId s1, StateId s2) {
+  if (fst.NumArcs(s2) > fst.NumArcs(s1)) FindAmbiguousPairs(fst, s2, s1);
+  SortedMatcher<Fst<Arc>> matcher(fst, MATCH_INPUT);
+  matcher.SetState(s2);
+  for (ArcIterator<Fst<Arc>> aiter(fst, s1); !aiter.Done(); aiter.Next()) {
+    const auto &arc1 = aiter.Value();
+    const ArcId a1(s1, aiter.Position());
+    if (matcher.Find(arc1.ilabel)) {
+      for (; !matcher.Done(); matcher.Next()) {
+        const auto &arc2 = matcher.Value();
+        // Continues on implicit epsilon match.
+        if (arc2.ilabel == kNoLabel) continue;
+        const ArcId a2(s2, matcher.Position());
+        // Actual transition is ambiguous.
+        if (s1 != s2 && arc1.nextstate == arc2.nextstate) {
+          InsertCandidate(s1, s2, a1, a2);
+        }
+        const auto spr = arc1.nextstate <= arc2.nextstate
+                             ? std::make_pair(arc1.nextstate, arc2.nextstate)
+                             : std::make_pair(arc2.nextstate, arc1.nextstate);
+        // Not already marked as coreachable?
+        if (coreachable_.insert(spr).second) {
+          // Only possible if state split by quantization issues.
+          if (spr.first != spr.second &&
+              head_[spr.first] == head_[spr.second]) {
+            if (!merge_) {
+              merge_.reset(new UnionFind<StateId>(fst.NumStates(), kNoStateId));
+              merge_->MakeAllSet(fst.NumStates());
+            }
+            merge_->Union(spr.first, spr.second);
+          } else {
+            queue_.push_back(spr);
+          }
+        }
+      }
+    }
+  }
+  // Super-final transition is ambiguous.
+  if (s1 != s2 && fst.Final(s1) != Weight::Zero() &&
+      fst.Final(s2) != Weight::Zero()) {
+    const ArcId a1(s1, -1);
+    const ArcId a2(s2, -1);
+    InsertCandidate(s1, s2, a1, a2);
+  }
+}
+
+template <class Arc>
+void Disambiguator<Arc>::MarkAmbiguities() {
+  if (!candidates_) return;
+  for (auto it = candidates_->begin(); it != candidates_->end(); ++it) {
+    const auto a = it->first;
+    const auto b = it->second;
+    // If b is not to be removed, then a is.
+    if (ambiguous_.count(b) == 0) ambiguous_.insert(a);
+  }
+  coreachable_.clear();
+  candidates_.reset();
+}
+
+template <class Arc>
+void Disambiguator<Arc>::RemoveSplits(MutableFst<Arc> *ofst) {
+  if (!merge_) return;
+  // Merges split states to remove spurious ambiguities.
+  for (StateIterator<MutableFst<Arc>> siter(*ofst); !siter.Done();
+       siter.Next()) {
+    for (MutableArcIterator<MutableFst<Arc>> aiter(ofst, siter.Value());
+         !aiter.Done(); aiter.Next()) {
+      auto arc = aiter.Value();
+      const auto nextstate = merge_->FindSet(arc.nextstate);
+      if (nextstate != arc.nextstate) {
+        arc.nextstate = nextstate;
+        aiter.SetValue(arc);
+      }
+    }
+  }
+  // Repeats search for actual ambiguities on modified FST.
+  coreachable_.clear();
+  merge_.reset();
+  candidates_.reset();
+  FindAmbiguities(*ofst);
+  if (merge_) {  // Shouldn't get here; sanity test.
+    FSTERROR() << "Disambiguate: Unable to remove spurious ambiguities";
+    error_ = true;
+    return;
+  }
+}
+
+template <class Arc>
+void Disambiguator<Arc>::RemoveAmbiguities(MutableFst<Arc> *ofst) {
+  if (ambiguous_.empty()) return;
+  // Adds dead state to redirect ambiguous transitions to be removed.
+  const auto dead = ofst->AddState();
+  for (auto it = ambiguous_.begin(); it != ambiguous_.end(); ++it) {
+    const auto pos = it->second;
+    if (pos >= 0) {  // Actual transition.
+      MutableArcIterator<MutableFst<Arc>> aiter(ofst, it->first);
+      aiter.Seek(pos);
+      auto arc = aiter.Value();
+      arc.nextstate = dead;
+      aiter.SetValue(arc);
+    } else {  // Super-final transition.
+      ofst->SetFinal(it->first, Weight::Zero());
+    }
+  }
+  Connect(ofst);
+  ambiguous_.clear();
+}
+
+}  // namespace internal
+
+// Disambiguates a weighted FST. This version writes the disambiguated FST to an
+// output MutableFst. The result will be an equivalent FST that has the
+// property that there are not two distinct paths from the initial state to a
+// final state with the same input labeling.
+//
+// The weights must be (weakly) left divisible (valid for Tropical and
+// LogWeight).
+//
+// Complexity:
+//
+//   Disambiguable: exponential (polynomial in the size of the output).
+//   Non-disambiguable: does not terminate.
+//
+// The disambiguable transducers include all automata and functional transducers
+// that are unweighted or that are acyclic or that are unambiguous.
+//
+// For more information, see:
+//
+// Mohri, M. and Riley, M. 2015. On the disambiguation of weighted automata.
+// In CIAA, pages 263-278.
+template <class Arc>
+void Disambiguate(
+    const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
+    const DisambiguateOptions<Arc> &opts = DisambiguateOptions<Arc>()) {
+  internal::Disambiguator<Arc> disambiguator;
+  disambiguator.Disambiguate(ifst, ofst, opts);
+}
+
+}  // namespace fst
+
+#endif  // FST_DISAMBIGUATE_H_
diff --git a/projects/llm_framework/include/fst/edit-fst.h b/projects/llm_framework/include/fst/edit-fst.h
new file mode 100644
index 00000000..e7bdfd61
--- /dev/null
+++ b/projects/llm_framework/include/fst/edit-fst.h
@@ -0,0 +1,702 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// An FST implementation that allows non-destructive edit operations on an
+// existing FST.
+//
+// The EditFst class enables non-destructive edit operations on a wrapped
+// ExpandedFst. The implementation uses copy-on-write semantics at the node
+// level: if a user has an underlying fst on which he or she wants to perform a
+// relatively small number of edits (read: mutations), then this implementation
+// will copy the edited node to an internal MutableFst and perform any edits in
+// situ on that copied node. This class supports all the methods of MutableFst
+// except for DeleteStates(const std::vector<StateId> &); thus, new nodes may
+// also be
+// added, and one may add transitions from existing nodes of the wrapped fst to
+// new nodes.
+//
+// N.B.: The documentation for Fst::Copy(true) says that its behavior is
+// undefined if invoked on an fst that has already been accessed.  This class
+// requires that the Fst implementation it wraps provides consistent, reliable
+// behavior when its Copy(true) method is invoked, where consistent means
+// the graph structure, graph properties and state numbering and do not change.
+// VectorFst and CompactFst, for example, are both well-behaved in this regard.
+
+#ifndef FST_EDIT_FST_H_
+#define FST_EDIT_FST_H_
+
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/cache.h>
+
+
+namespace fst {
+namespace internal {
+
+// The EditFstData class is a container for all mutable data for EditFstImpl;
+// also, this class provides most of the actual implementation of what EditFst
+// does (that is, most of EditFstImpl's methods delegate to methods in this, the
+// EditFstData class).  Instances of this class are reference-counted and can be
+// shared between otherwise independent EditFstImpl instances. This scheme
+// allows EditFstImpl to implement the thread-safe, copy-on-write semantics
+// required by Fst::Copy(true).
+//
+// template parameters:
+//   A the type of arc to use
+//   WrappedFstT the type of fst wrapped by the EditFst instance that
+//     this EditFstData instance is backing
+//   MutableFstT the type of mutable fst to use internally for edited states;
+//     crucially, MutableFstT::Copy(false) *must* yield an fst that is
+//     thread-safe for reading (VectorFst, for example, has this property)
+template <typename Arc, typename WrappedFstT = ExpandedFst<Arc>,
+          typename MutableFstT = VectorFst<Arc>>
+class EditFstData {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  EditFstData() : num_new_states_(0) {}
+
+  EditFstData(const EditFstData &other)
+      : edits_(other.edits_),
+        external_to_internal_ids_(other.external_to_internal_ids_),
+        edited_final_weights_(other.edited_final_weights_),
+        num_new_states_(other.num_new_states_) {}
+
+  ~EditFstData() {}
+
+  static EditFstData<Arc, WrappedFstT, MutableFstT> *Read(
+      std::istream &strm, const FstReadOptions &opts);
+
+  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
+    // Serialize all private data members of this class.
+    FstWriteOptions edits_opts(opts);
+    edits_opts.write_header = true;  // Force writing contained header.
+    edits_.Write(strm, edits_opts);
+    WriteType(strm, external_to_internal_ids_);
+    WriteType(strm, edited_final_weights_);
+    WriteType(strm, num_new_states_);
+    if (!strm) {
+      LOG(ERROR) << "EditFstData::Write: Write failed: " << opts.source;
+      return false;
+    }
+    return true;
+  }
+
+  StateId NumNewStates() const { return num_new_states_; }
+
+  // accessor methods for the fst holding edited states
+  StateId EditedStart() const { return edits_.Start(); }
+
+  Weight Final(StateId s, const WrappedFstT *wrapped) const {
+    auto final_weight_it = GetFinalWeightIterator(s);
+    if (final_weight_it == NotInFinalWeightMap()) {
+      auto it = GetEditedIdMapIterator(s);
+      return it == NotInEditedMap() ? wrapped->Final(s)
+                                    : edits_.Final(it->second);
+    } else {
+      return final_weight_it->second;
+    }
+  }
+
+  size_t NumArcs(StateId s, const WrappedFstT *wrapped) const {
+    auto it = GetEditedIdMapIterator(s);
+    return it == NotInEditedMap() ? wrapped->NumArcs(s)
+                                  : edits_.NumArcs(it->second);
+  }
+
+  size_t NumInputEpsilons(StateId s, const WrappedFstT *wrapped) const {
+    auto it = GetEditedIdMapIterator(s);
+    return it == NotInEditedMap() ? wrapped->NumInputEpsilons(s)
+                                  : edits_.NumInputEpsilons(it->second);
+  }
+
+  size_t NumOutputEpsilons(StateId s, const WrappedFstT *wrapped) const {
+    auto it = GetEditedIdMapIterator(s);
+    return it == NotInEditedMap() ? wrapped->NumOutputEpsilons(s)
+                                  : edits_.NumOutputEpsilons(it->second);
+  }
+
+  void SetEditedProperties(uint64 props, uint64 mask) {
+    edits_.SetProperties(props, mask);
+  }
+
+  // Non-const MutableFst operations.
+
+  // Sets the start state for this FST.
+  void SetStart(StateId s) { edits_.SetStart(s); }
+
+  // Sets the final state for this FST.
+  Weight SetFinal(StateId s, Weight w, const WrappedFstT *wrapped) {
+    Weight old_weight = Final(s, wrapped);
+    auto it = GetEditedIdMapIterator(s);
+    // If we haven't already edited state s, don't add it to edited_ (which can
+    // be expensive if s has many transitions); just use the
+    // edited_final_weights_ map.
+    if (it == NotInEditedMap()) {
+      edited_final_weights_[s] = w;
+    } else {
+      edits_.SetFinal(GetEditableInternalId(s, wrapped), w);
+    }
+    return old_weight;
+  }
+
+  // Adds a new state to this FST, initially with no arcs.
+  StateId AddState(StateId curr_num_states) {
+    StateId internal_state_id = edits_.AddState();
+    StateId external_state_id = curr_num_states;
+    external_to_internal_ids_[external_state_id] = internal_state_id;
+    num_new_states_++;
+    return external_state_id;
+  }
+
+  // Adds the specified arc to the specified state of this FST.
+  const Arc *AddArc(StateId s, const Arc &arc, const WrappedFstT *wrapped) {
+    const auto internal_id = GetEditableInternalId(s, wrapped);
+    const auto num_arcs = edits_.NumArcs(internal_id);
+    ArcIterator<MutableFstT> arc_it(edits_, internal_id);
+    const Arc *prev_arc = nullptr;
+    if (num_arcs > 0) {
+      // grab the final arc associated with this state in edits_
+      arc_it.Seek(num_arcs - 1);
+      prev_arc = &(arc_it.Value());
+    }
+    edits_.AddArc(internal_id, arc);
+    return prev_arc;
+  }
+
+  void DeleteStates() {
+    edits_.DeleteStates();
+    num_new_states_ = 0;
+    external_to_internal_ids_.clear();
+    edited_final_weights_.clear();
+  }
+
+  // Removes all but the first n outgoing arcs of the specified state.
+  void DeleteArcs(StateId s, size_t n, const WrappedFstT *wrapped) {
+    edits_.DeleteArcs(GetEditableInternalId(s, wrapped), n);
+  }
+
+  // Removes all outgoing arcs from the specified state.
+  void DeleteArcs(StateId s, const WrappedFstT *wrapped) {
+    edits_.DeleteArcs(GetEditableInternalId(s, wrapped));
+  }
+
+  // End methods for non-const MutableFst operations.
+
+  // Provides information for the generic arc iterator.
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data,
+                       const WrappedFstT *wrapped) const {
+    auto id_map_it = GetEditedIdMapIterator(s);
+    if (id_map_it == NotInEditedMap()) {
+      VLOG(3) << "EditFstData::InitArcIterator: iterating on state " << s
+              << " of original fst";
+      wrapped->InitArcIterator(s, data);
+    } else {
+      VLOG(2) << "EditFstData::InitArcIterator: iterating on edited state " << s
+              << " (internal state id: " << id_map_it->second << ")";
+      edits_.InitArcIterator(id_map_it->second, data);
+    }
+  }
+
+  // Provides information for the generic mutable arc iterator.
+  void InitMutableArcIterator(StateId s, MutableArcIteratorData<Arc> *data,
+                              const WrappedFstT *wrapped) {
+    data->base = new MutableArcIterator<MutableFstT>(
+        &edits_, GetEditableInternalId(s, wrapped));
+  }
+
+  // Prints out the map from external to internal state id's (for debugging
+  // purposes).
+  void PrintMap() {
+    for (auto map_it = external_to_internal_ids_.begin();
+         map_it != NotInEditedMap(); ++map_it) {
+      LOG(INFO) << "(external,internal)=(" << map_it->first << ","
+                << map_it->second << ")";
+    }
+  }
+
+ private:
+  // Returns the iterator of the map from external to internal state id's
+  // of edits_ for the specified external state id.
+  typename std::unordered_map<StateId, StateId>::const_iterator
+      GetEditedIdMapIterator(StateId s) const {
+    return external_to_internal_ids_.find(s);
+  }
+
+  typename std::unordered_map<StateId, StateId>::const_iterator
+      NotInEditedMap() const {
+    return external_to_internal_ids_.end();
+  }
+
+  typename std::unordered_map<StateId, Weight>::const_iterator
+      GetFinalWeightIterator(StateId s) const {
+    return edited_final_weights_.find(s);
+  }
+
+  typename std::unordered_map<StateId, Weight>::const_iterator
+      NotInFinalWeightMap() const {
+    return edited_final_weights_.end();
+  }
+
+  // Returns the internal state ID of the specified external ID if the state has
+  // already been made editable, or else copies the state from wrapped_ to
+  // edits_ and returns the state id of the newly editable state in edits_.
+  StateId GetEditableInternalId(StateId s, const WrappedFstT *wrapped) {
+    auto id_map_it = GetEditedIdMapIterator(s);
+    if (id_map_it == NotInEditedMap()) {
+      StateId new_internal_id = edits_.AddState();
+      VLOG(2) << "EditFstData::GetEditableInternalId: editing state " << s
+              << " of original fst; new internal state id:" << new_internal_id;
+      external_to_internal_ids_[s] = new_internal_id;
+      for (ArcIterator<Fst<Arc>> arc_iterator(*wrapped, s);
+           !arc_iterator.Done(); arc_iterator.Next()) {
+        edits_.AddArc(new_internal_id, arc_iterator.Value());
+      }
+      // Copies the final weight.
+      auto final_weight_it = GetFinalWeightIterator(s);
+      if (final_weight_it == NotInFinalWeightMap()) {
+        edits_.SetFinal(new_internal_id, wrapped->Final(s));
+      } else {
+        edits_.SetFinal(new_internal_id, final_weight_it->second);
+        edited_final_weights_.erase(s);
+      }
+      return new_internal_id;
+    } else {
+      return id_map_it->second;
+    }
+  }
+
+  // A mutable FST (by default, a VectorFst) to contain new states, and/or
+  // copies of states from a wrapped ExpandedFst that have been modified in
+  // some way.
+  MutableFstT edits_;
+  // A mapping from external state IDs to the internal IDs of states that
+  // appear in edits_.
+  std::unordered_map<StateId, StateId> external_to_internal_ids_;
+  // A mapping from external state IDs to final state weights assigned to
+  // those states.  The states in this map are *only* those whose final weight
+  // has been modified; if any other part of the state has been modified,
+  // the entire state is copied to edits_, and all modifications reside there.
+  std::unordered_map<StateId, Weight> edited_final_weights_;
+  // The number of new states added to this mutable fst impl, which is <= the
+  // number of states in edits_ (since edits_ contains both edited *and* new
+  // states).
+  StateId num_new_states_;
+};
+
+// EditFstData method implementations: just the Read method.
+template <typename A, typename WrappedFstT, typename MutableFstT>
+EditFstData<A, WrappedFstT, MutableFstT> *
+EditFstData<A, WrappedFstT, MutableFstT>::Read(std::istream &strm,
+                                               const FstReadOptions &opts) {
+  auto *data = new EditFstData<A, WrappedFstT, MutableFstT>();
+  // next read in MutabelFstT machine that stores edits
+  FstReadOptions edits_opts(opts);
+  // Contained header was written out, so read it in.
+  edits_opts.header = nullptr;
+
+  // Because our internal representation of edited states is a solid object
+  // of type MutableFstT (defaults to VectorFst<A>) and not a pointer,
+  // and because the static Read method allocates a new object on the heap,
+  // we need to call Read, check if there was a failure, use
+  // MutableFstT::operator= to assign the object (not the pointer) to the
+  // edits_ data member (which will increase the ref count by 1 on the impl)
+  // and, finally, delete the heap-allocated object.
+  std::unique_ptr<MutableFstT> edits(MutableFstT::Read(strm, edits_opts));
+  if (!edits) return nullptr;
+  data->edits_ = *edits;
+  edits.reset();
+  // Finally, reads in rest of private data members.
+  ReadType(strm, &data->external_to_internal_ids_);
+  ReadType(strm, &data->edited_final_weights_);
+  ReadType(strm, &data->num_new_states_);
+  if (!strm) {
+    LOG(ERROR) << "EditFst::Read: read failed: " << opts.source;
+    return nullptr;
+  }
+  return data;
+}
+
+// This class enables non-destructive edit operations on a wrapped ExpandedFst.
+// The implementation uses copy-on-write semantics at the node level: if a user
+// has an underlying fst on which he or she wants to perform a relatively small
+// number of edits (read: mutations), then this implementation will copy the
+// edited node to an internal MutableFst and perform any edits in situ on that
+// copied node. This class supports all the methods of MutableFst except for
+// DeleteStates(const std::vector<StateId> &); thus, new nodes may also be
+// added, and
+// one may add transitions from existing nodes of the wrapped fst to new nodes.
+//
+// template parameters:
+//   A the type of arc to use
+//   WrappedFstT the type of fst wrapped by the EditFst instance that
+//     this EditFstImpl instance is backing
+//   MutableFstT the type of mutable fst to use internally for edited states;
+//     crucially, MutableFstT::Copy(false) *must* yield an fst that is
+//     thread-safe for reading (VectorFst, for example, has this property)
+template <typename A, typename WrappedFstT = ExpandedFst<A>,
+          typename MutableFstT = VectorFst<A>>
+class EditFstImpl : public FstImpl<A> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FstImpl<Arc>::SetProperties;
+  using FstImpl<Arc>::SetInputSymbols;
+  using FstImpl<Arc>::SetOutputSymbols;
+  using FstImpl<Arc>::WriteHeader;
+
+  // Constructs an editable FST implementation with no states. Effectively, this
+  // initially-empty fst will in every way mimic the behavior of a
+  // VectorFst---more precisely, a VectorFstImpl instance---but with slightly
+  // slower performance (by a constant factor), due to the fact that
+  // this class maintains a mapping between external state id's and
+  // their internal equivalents.
+  EditFstImpl() : wrapped_(new MutableFstT()) {
+    FstImpl<Arc>::SetType("edit");
+    InheritPropertiesFromWrapped();
+    data_ = std::make_shared<EditFstData<Arc, WrappedFstT, MutableFstT>>();
+  }
+
+  // Wraps the specified ExpandedFst. This constructor requires that the
+  // specified Fst is an ExpandedFst instance. This requirement is only enforced
+  // at runtime. (See below for the reason.)
+  //
+  // This library uses the pointer-to-implementation or "PIMPL" design pattern.
+  // In particular, to make it convenient to bind an implementation class to its
+  // interface, there are a pair of template "binder" classes, one for immutable
+  // and one for mutable fst's (ImplToFst and ImplToMutableFst, respectively).
+  // As it happens, the API for the ImplToMutableFst<I,F> class requires that
+  // the implementation class--the template parameter "I"--have a constructor
+  // taking a const Fst<A> reference.  Accordingly, the constructor here must
+  // perform a static_cast to the WrappedFstT type required by EditFst and
+  // therefore EditFstImpl.
+  explicit EditFstImpl(const Fst<Arc> &wrapped)
+      : wrapped_(static_cast<WrappedFstT *>(wrapped.Copy())) {
+    FstImpl<Arc>::SetType("edit");
+    data_ = std::make_shared<EditFstData<Arc, WrappedFstT, MutableFstT>>();
+    // have edits_ inherit all properties from wrapped_
+    data_->SetEditedProperties(wrapped_->Properties(kFstProperties, false),
+                               kFstProperties);
+    InheritPropertiesFromWrapped();
+  }
+
+  // A copy constructor for this implementation class, used to implement
+  // the Copy() method of the Fst interface.
+  EditFstImpl(const EditFstImpl &impl)
+      : FstImpl<Arc>(),
+        wrapped_(static_cast<WrappedFstT *>(impl.wrapped_->Copy(true))),
+        data_(impl.data_) {
+    SetProperties(impl.Properties());
+  }
+
+  // const Fst/ExpandedFst operations, declared in the Fst and ExpandedFst
+  // interfaces
+  StateId Start() const {
+    const auto edited_start = data_->EditedStart();
+    return edited_start == kNoStateId ? wrapped_->Start() : edited_start;
+  }
+
+  Weight Final(StateId s) const { return data_->Final(s, wrapped_.get()); }
+
+  size_t NumArcs(StateId s) const { return data_->NumArcs(s, wrapped_.get()); }
+
+  size_t NumInputEpsilons(StateId s) const {
+    return data_->NumInputEpsilons(s, wrapped_.get());
+  }
+
+  size_t NumOutputEpsilons(StateId s) const {
+    return data_->NumOutputEpsilons(s, wrapped_.get());
+  }
+
+  StateId NumStates() const {
+    return wrapped_->NumStates() + data_->NumNewStates();
+  }
+
+  static EditFstImpl<Arc, WrappedFstT, MutableFstT> *Read(
+      std::istream &strm, const FstReadOptions &opts);
+
+  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
+    FstHeader hdr;
+    hdr.SetStart(Start());
+    hdr.SetNumStates(NumStates());
+    FstWriteOptions header_opts(opts);
+    // Allows the contained FST to hold any symbols.
+    header_opts.write_isymbols = false;
+    header_opts.write_osymbols = false;
+    WriteHeader(strm, header_opts, kFileVersion, &hdr);
+    // First, serializes the wrapped FST to stream.
+    FstWriteOptions wrapped_opts(opts);
+    // Forcse writing the contained header.
+    wrapped_opts.write_header = true;
+    wrapped_->Write(strm, wrapped_opts);
+    data_->Write(strm, opts);
+    strm.flush();
+    if (!strm) {
+      LOG(ERROR) << "EditFst::Write: Write failed: " << opts.source;
+      return false;
+    }
+    return true;
+  }
+
+  // Sets the start state for this FST.
+  void SetStart(StateId s) {
+    MutateCheck();
+    data_->SetStart(s);
+    SetProperties(SetStartProperties(FstImpl<Arc>::Properties()));
+  }
+
+  // Sets the final state for this fst.
+  void SetFinal(StateId s, Weight weight) {
+    MutateCheck();
+    Weight old_weight = data_->SetFinal(s, weight, wrapped_.get());
+    SetProperties(
+        SetFinalProperties(FstImpl<Arc>::Properties(), old_weight, weight));
+  }
+
+  // Adds a new state to this fst, initially with no arcs.
+  StateId AddState() {
+    MutateCheck();
+    SetProperties(AddStateProperties(FstImpl<Arc>::Properties()));
+    return data_->AddState(NumStates());
+  }
+
+  // Adds the specified arc to the specified state of this fst.
+  void AddArc(StateId s, const Arc &arc) {
+    MutateCheck();
+    const auto *prev_arc = data_->AddArc(s, arc, wrapped_.get());
+    SetProperties(
+        AddArcProperties(FstImpl<Arc>::Properties(), s, arc, prev_arc));
+  }
+
+  void DeleteStates(const std::vector<StateId> &dstates) {
+    FSTERROR() << ": EditFstImpl::DeleteStates(const std::vector<StateId>&): "
+               << " not implemented";
+    SetProperties(kError, kError);
+  }
+
+  // Deletes all states in this fst.
+  void DeleteStates();
+
+  // Removes all but the first n outgoing arcs of the specified state.
+  void DeleteArcs(StateId s, size_t n) {
+    MutateCheck();
+    data_->DeleteArcs(s, n, wrapped_.get());
+    SetProperties(DeleteArcsProperties(FstImpl<Arc>::Properties()));
+  }
+
+  // Removes all outgoing arcs from the specified state.
+  void DeleteArcs(StateId s) {
+    MutateCheck();
+    data_->DeleteArcs(s, wrapped_.get());
+    SetProperties(DeleteArcsProperties(FstImpl<Arc>::Properties()));
+  }
+
+  void ReserveStates(StateId s) {}
+
+  void ReserveArcs(StateId s, size_t n) {}
+
+  // Ends non-const MutableFst operations.
+
+  // Provides information for the generic state iterator.
+  void InitStateIterator(StateIteratorData<Arc> *data) const {
+    data->base = nullptr;
+    data->nstates = NumStates();
+  }
+
+  // Provides information for the generic arc iterator.
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const {
+    data_->InitArcIterator(s, data, wrapped_.get());
+  }
+
+  // Provides information for the generic mutable arc iterator.
+  void InitMutableArcIterator(StateId s, MutableArcIteratorData<Arc> *data) {
+    MutateCheck();
+    data_->InitMutableArcIterator(s, data, wrapped_.get());
+  }
+
+ private:
+  // Properties always true of this FST class.
+  static constexpr uint64 kStaticProperties = kExpanded | kMutable;
+  // Current file format version.
+  static constexpr int kFileVersion = 2;
+  // Minimum file format version supported
+  static constexpr int kMinFileVersion = 2;
+
+  // Causes this FST to inherit all the properties from its wrapped FST, except
+  // for the two properties that always apply to EditFst instances: kExpanded
+  // and kMutable.
+  void InheritPropertiesFromWrapped() {
+    SetProperties(wrapped_->Properties(kCopyProperties, false) |
+                  kStaticProperties);
+    SetInputSymbols(wrapped_->InputSymbols());
+    SetOutputSymbols(wrapped_->OutputSymbols());
+  }
+
+  // This method ensures that any operations that alter the mutable data
+  // portion of this EditFstImpl cause the data_ member to be copied when its
+  // reference count is greater than 1.  Note that this method is distinct from
+  // MutableFst::Mutate, which gets invoked whenever one of the basic mutation
+  // methods defined in MutableFst is invoked, such as SetInputSymbols.
+  // The MutateCheck here in EditFstImpl is invoked whenever one of the
+  // mutating methods specifically related to the types of edits provided
+  // by EditFst is performed, such as changing an arc of an existing state
+  // of the wrapped fst via a MutableArcIterator, or adding a new state via
+  // AddState().
+  void MutateCheck() {
+    if (!data_.unique()) {
+      data_ =
+          std::make_shared<EditFstData<Arc, WrappedFstT, MutableFstT>>(*data_);
+    }
+  }
+
+  // The FST that this FST wraps. The purpose of this class is to enable
+  // non-destructive edits on this wrapped FST.
+  std::unique_ptr<const WrappedFstT> wrapped_;
+  // The mutable data for this EditFst instance, with delegates for all the
+  // methods that can mutate data.
+  std::shared_ptr<EditFstData<Arc, WrappedFstT, MutableFstT>> data_;
+};
+
+template <typename Arc, typename WrappedFstT, typename MutableFstT>
+constexpr uint64 EditFstImpl<Arc, WrappedFstT, MutableFstT>::kStaticProperties;
+
+template <typename Arc, typename WrappedFstT, typename MutableFstT>
+constexpr int EditFstImpl<Arc, WrappedFstT, MutableFstT>::kFileVersion;
+
+template <typename Arc, typename WrappedFstT, typename MutableFstT>
+constexpr int EditFstImpl<Arc, WrappedFstT, MutableFstT>::kMinFileVersion;
+
+template <typename Arc, typename WrappedFstT, typename MutableFstT>
+inline void EditFstImpl<Arc, WrappedFstT, MutableFstT>::DeleteStates() {
+  data_->DeleteStates();
+  // we are deleting all states, so just forget about pointer to wrapped_
+  // and do what default constructor does: set wrapped_ to a new VectorFst
+  wrapped_.reset(new MutableFstT());
+  const auto new_props =
+      DeleteAllStatesProperties(FstImpl<Arc>::Properties(), kStaticProperties);
+  FstImpl<Arc>::SetProperties(new_props);
+}
+
+template <typename Arc, typename WrappedFstT, typename MutableFstT>
+EditFstImpl<Arc, WrappedFstT, MutableFstT> *
+EditFstImpl<Arc, WrappedFstT, MutableFstT>::Read(std::istream &strm,
+                                                 const FstReadOptions &opts) {
+  auto *impl = new EditFstImpl();
+  FstHeader hdr;
+  if (!impl->ReadHeader(strm, opts, kMinFileVersion, &hdr)) return nullptr;
+  impl->SetStart(hdr.Start());
+  // Reads in wrapped FST.
+  FstReadOptions wrapped_opts(opts);
+  // Contained header was written out, so reads it in too.
+  wrapped_opts.header = nullptr;
+  std::unique_ptr<Fst<Arc>> wrapped_fst(Fst<Arc>::Read(strm, wrapped_opts));
+  if (!wrapped_fst) return nullptr;
+  impl->wrapped_.reset(static_cast<WrappedFstT *>(wrapped_fst.release()));
+  impl->data_ = std::shared_ptr<EditFstData<Arc, WrappedFstT, MutableFstT>>(
+      EditFstData<Arc, WrappedFstT, MutableFstT>::Read(strm, opts));
+  if (!impl->data_) return nullptr;
+  return impl;
+}
+
+}  // namespace internal
+
+// Concrete, editable FST.  This class attaches interface to implementation.
+template <typename A, typename WrappedFstT = ExpandedFst<A>,
+          typename MutableFstT = VectorFst<A>>
+class EditFst : public ImplToMutableFst<
+                    internal::EditFstImpl<A, WrappedFstT, MutableFstT>> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+
+  using Impl = internal::EditFstImpl<Arc, WrappedFstT, MutableFstT>;
+
+  friend class MutableArcIterator<EditFst<Arc, WrappedFstT, MutableFstT>>;
+
+  EditFst() : ImplToMutableFst<Impl>(std::make_shared<Impl>()) {}
+
+  explicit EditFst(const Fst<Arc> &fst)
+      : ImplToMutableFst<Impl>(std::make_shared<Impl>(fst)) {}
+
+  explicit EditFst(const WrappedFstT &fst)
+      : ImplToMutableFst<Impl>(std::make_shared<Impl>(fst)) {}
+
+  // See Fst<>::Copy() for doc.
+  EditFst(const EditFst<Arc, WrappedFstT, MutableFstT> &fst, bool safe = false)
+      : ImplToMutableFst<Impl>(fst, safe) {}
+
+  ~EditFst() override {}
+
+  // Gets a copy of this EditFst. See Fst<>::Copy() for further doc.
+  EditFst<Arc, WrappedFstT, MutableFstT> *Copy(
+      bool safe = false) const override {
+    return new EditFst<Arc, WrappedFstT, MutableFstT>(*this, safe);
+  }
+
+  EditFst<Arc, WrappedFstT, MutableFstT> &operator=(
+      const EditFst<Arc, WrappedFstT, MutableFstT> &fst) {
+    SetImpl(fst.GetSharedImpl());
+    return *this;
+  }
+
+  EditFst<Arc, WrappedFstT, MutableFstT> &operator=(
+      const Fst<Arc> &fst) override {
+    SetImpl(std::make_shared<Impl>(fst));
+    return *this;
+  }
+
+  // Reads an EditFst from an input stream, returning nullptr on error.
+  static EditFst<Arc, WrappedFstT, MutableFstT> *Read(
+      std::istream &strm, const FstReadOptions &opts) {
+    auto *impl = Impl::Read(strm, opts);
+    return impl ? new EditFst<Arc>(std::shared_ptr<Impl>(impl)) : nullptr;
+  }
+
+  // Reads an EditFst from a file, returning nullptr on error. If the filename
+  // argument is an empty string, it reads from standard input.
+  static EditFst<Arc, WrappedFstT, MutableFstT> *Read(const string &filename) {
+    auto *impl = ImplToExpandedFst<Impl, MutableFst<Arc>>::Read(filename);
+    return impl ? new EditFst<Arc, WrappedFstT, MutableFstT>(
+                      std::shared_ptr<Impl>(impl))
+                : nullptr;
+  }
+
+  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
+    return GetImpl()->Write(strm, opts);
+  }
+
+  bool Write(const string &filename) const override {
+    return Fst<Arc>::WriteFile(filename);
+  }
+
+  void InitStateIterator(StateIteratorData<Arc> *data) const override {
+    GetImpl()->InitStateIterator(data);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
+    GetImpl()->InitArcIterator(s, data);
+  }
+
+  void InitMutableArcIterator(StateId s,
+                              MutableArcIteratorData<A> *data) override {
+    GetMutableImpl()->InitMutableArcIterator(s, data);
+  }
+
+ private:
+  explicit EditFst(std::shared_ptr<Impl> impl) : ImplToMutableFst<Impl>(impl) {}
+
+  using ImplToFst<Impl, MutableFst<Arc>>::GetImpl;
+  using ImplToFst<Impl, MutableFst<Arc>>::GetMutableImpl;
+  using ImplToFst<Impl, MutableFst<Arc>>::SetImpl;
+};
+
+}  // namespace fst
+
+#endif  // FST_EDIT_FST_H_
diff --git a/projects/llm_framework/include/fst/encode.h b/projects/llm_framework/include/fst/encode.h
new file mode 100644
index 00000000..f251bbfc
--- /dev/null
+++ b/projects/llm_framework/include/fst/encode.h
@@ -0,0 +1,556 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to encode and decode an FST.
+
+#ifndef FST_ENCODE_H_
+#define FST_ENCODE_H_
+
+#include <iostream>
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+#include <fstream>
+
+#include <fst/arc-map.h>
+#include <fst/rmfinalepsilon.h>
+
+
+namespace fst {
+
+enum EncodeType { ENCODE = 1, DECODE = 2 };
+
+static constexpr uint32 kEncodeLabels = 0x0001;
+static constexpr uint32 kEncodeWeights = 0x0002;
+static constexpr uint32 kEncodeFlags = 0x0003;
+
+namespace internal {
+
+static constexpr uint32 kEncodeHasISymbols = 0x0004;
+static constexpr uint32 kEncodeHasOSymbols = 0x0008;
+
+// Identifies stream data as an encode table (and its endianity)
+static const int32 kEncodeMagicNumber = 2129983209;
+
+// The following class encapsulates implementation details for the encoding and
+// decoding of label/weight tuples used for encoding and decoding of FSTs. The
+// EncodeTable is bidirectional. I.e, it stores both the Tuple of encode labels
+// and weights to a unique label, and the reverse.
+template <class Arc>
+class EncodeTable {
+ public:
+  using Label = typename Arc::Label;
+  using Weight = typename Arc::Weight;
+
+  // Encoded data consists of arc input/output labels and arc weight.
+  struct Tuple {
+    Tuple() {}
+
+    Tuple(Label ilabel_, Label olabel_, Weight weight_)
+        : ilabel(ilabel_), olabel(olabel_), weight(std::move(weight_)) {}
+
+    Tuple(const Tuple &tuple)
+        : ilabel(tuple.ilabel),
+          olabel(tuple.olabel),
+          weight(std::move(tuple.weight)) {}
+
+    Label ilabel;
+    Label olabel;
+    Weight weight;
+  };
+
+  // Comparison object for hashing EncodeTable Tuple(s).
+  class TupleEqual {
+   public:
+    bool operator()(const Tuple *x, const Tuple *y) const {
+      return (x->ilabel == y->ilabel && x->olabel == y->olabel &&
+              x->weight == y->weight);
+    }
+  };
+
+  // Hash function for EncodeTabe Tuples. Based on the encode flags
+  // we either hash the labels, weights or combination of them.
+  class TupleKey {
+   public:
+    TupleKey() : encode_flags_(kEncodeLabels | kEncodeWeights) {}
+
+    TupleKey(const TupleKey &key) : encode_flags_(key.encode_flags_) {}
+
+    explicit TupleKey(uint32 encode_flags) : encode_flags_(encode_flags) {}
+
+    size_t operator()(const Tuple *x) const {
+      size_t hash = x->ilabel;
+      static constexpr int lshift = 5;
+      static constexpr int rshift = CHAR_BIT * sizeof(size_t) - 5;
+      if (encode_flags_ & kEncodeLabels) {
+        hash = hash << lshift ^ hash >> rshift ^ x->olabel;
+      }
+      if (encode_flags_ & kEncodeWeights) {
+        hash = hash << lshift ^ hash >> rshift ^ x->weight.Hash();
+      }
+      return hash;
+    }
+
+   private:
+    int32 encode_flags_;
+  };
+
+  explicit EncodeTable(uint32 encode_flags)
+      : flags_(encode_flags), encode_hash_(1024, TupleKey(encode_flags)) {}
+
+  using EncodeHash = std::unordered_map<const Tuple *, Label, TupleKey,
+                                        TupleEqual>;
+
+  // Given an arc, encodes either input/output labels or input/costs or both.
+  Label Encode(const Arc &arc) {
+    std::unique_ptr<Tuple> tuple(
+        new Tuple(arc.ilabel, flags_ & kEncodeLabels ? arc.olabel : 0,
+                  flags_ & kEncodeWeights ? arc.weight : Weight::One()));
+    auto insert_result = encode_hash_.insert(
+        std::make_pair(tuple.get(), encode_tuples_.size() + 1));
+    if (insert_result.second) encode_tuples_.push_back(std::move(tuple));
+    return insert_result.first->second;
+  }
+
+  // Given an arc, looks up its encoded label or returns kNoLabel if not found.
+  Label GetLabel(const Arc &arc) const {
+    const Tuple tuple(arc.ilabel, flags_ & kEncodeLabels ? arc.olabel : 0,
+                      flags_ & kEncodeWeights ? arc.weight : Weight::One());
+    auto it = encode_hash_.find(&tuple);
+    return (it == encode_hash_.end()) ?  kNoLabel : it->second;
+  }
+
+  // Given an encoded arc label, decodes back to input/output labels and costs.
+  const Tuple *Decode(Label key) const {
+    if (key < 1 || key > encode_tuples_.size()) {
+      LOG(ERROR) << "EncodeTable::Decode: Unknown decode key: " << key;
+      return nullptr;
+    }
+    return encode_tuples_[key - 1].get();
+  }
+
+  size_t Size() const { return encode_tuples_.size(); }
+
+  bool Write(std::ostream &strm, const string &source) const;
+
+  static EncodeTable<Arc> *Read(std::istream &strm, const string &source);
+
+  uint32 Flags() const { return flags_ & kEncodeFlags; }
+
+  const SymbolTable *InputSymbols() const { return isymbols_.get(); }
+
+  const SymbolTable *OutputSymbols() const { return osymbols_.get(); }
+
+  void SetInputSymbols(const SymbolTable *syms) {
+    if (syms) {
+      isymbols_.reset(syms->Copy());
+      flags_ |= kEncodeHasISymbols;
+    } else {
+      isymbols_.reset();
+      flags_ &= ~kEncodeHasISymbols;
+    }
+  }
+
+  void SetOutputSymbols(const SymbolTable *syms) {
+    if (syms) {
+      osymbols_.reset(syms->Copy());
+      flags_ |= kEncodeHasOSymbols;
+    } else {
+      osymbols_.reset();
+      flags_ &= ~kEncodeHasOSymbols;
+    }
+  }
+
+ private:
+  uint32 flags_;
+  std::vector<std::unique_ptr<Tuple>> encode_tuples_;
+  EncodeHash encode_hash_;
+  std::unique_ptr<SymbolTable> isymbols_;  // Pre-encoded input symbol table.
+  std::unique_ptr<SymbolTable> osymbols_;  // Pre-encoded output symbol table.
+
+  EncodeTable(const EncodeTable &) = delete;
+  EncodeTable &operator=(const EncodeTable &) = delete;
+};
+
+template <class Arc>
+bool EncodeTable<Arc>::Write(std::ostream &strm,
+                                  const string &source) const {
+  WriteType(strm, kEncodeMagicNumber);
+  WriteType(strm, flags_);
+  const int64 size = encode_tuples_.size();
+  WriteType(strm, size);
+  for (const auto &tuple : encode_tuples_) {
+    WriteType(strm, tuple->ilabel);
+    WriteType(strm, tuple->olabel);
+    tuple->weight.Write(strm);
+  }
+  if (flags_ & kEncodeHasISymbols) isymbols_->Write(strm);
+  if (flags_ & kEncodeHasOSymbols) osymbols_->Write(strm);
+  strm.flush();
+  if (!strm) {
+    LOG(ERROR) << "EncodeTable::Write: Write failed: " << source;
+    return false;
+  }
+  return true;
+}
+
+template <class Arc>
+EncodeTable<Arc> *EncodeTable<Arc>::Read(std::istream &strm,
+                                         const string &source) {
+  int32 magic_number = 0;
+  ReadType(strm, &magic_number);
+  if (magic_number != kEncodeMagicNumber) {
+    LOG(ERROR) << "EncodeTable::Read: Bad encode table header: " << source;
+    return nullptr;
+  }
+  uint32 flags;
+  ReadType(strm, &flags);
+  int64 size;
+  ReadType(strm, &size);
+  if (!strm) {
+    LOG(ERROR) << "EncodeTable::Read: Read failed: " << source;
+    return nullptr;
+  }
+  std::unique_ptr<EncodeTable<Arc>> table(new EncodeTable<Arc>(flags));
+  for (int64 i = 0; i < size; ++i) {
+    std::unique_ptr<Tuple> tuple(new Tuple());
+    ReadType(strm, &tuple->ilabel);
+    ReadType(strm, &tuple->olabel);
+    tuple->weight.Read(strm);
+    if (!strm) {
+      LOG(ERROR) << "EncodeTable::Read: Read failed: " << source;
+      return nullptr;
+    }
+    table->encode_tuples_.push_back(std::move(tuple));
+    table->encode_hash_[table->encode_tuples_.back().get()] =
+        table->encode_tuples_.size();
+  }
+  if (flags & kEncodeHasISymbols) {
+    table->isymbols_.reset(SymbolTable::Read(strm, source));
+  }
+  if (flags & kEncodeHasOSymbols) {
+    table->osymbols_.reset(SymbolTable::Read(strm, source));
+  }
+  return table.release();
+}
+
+}  // namespace internal
+
+// A mapper to encode/decode weighted transducers. Encoding of an FST is used
+// for performing classical determinization or minimization on a weighted
+// transducer viewing it as an unweighted acceptor over encoded labels.
+//
+// The mapper stores the encoding in a local hash table (EncodeTable). This
+// table is shared (and reference-counted) between the encoder and decoder.
+// A decoder has read-only access to the EncodeTable.
+//
+// The EncodeMapper allows on the fly encoding of the machine. As the
+// EncodeTable is generated the same table may by used to decode the machine
+// on the fly. For example in the following sequence of operations
+//
+//  Encode -> Determinize -> Decode
+//
+// we will use the encoding table generated during the encode step in the
+// decode, even though the encoding is not complete.
+template <class Arc>
+class EncodeMapper {
+  using Label = typename Arc::Label;
+  using Weight = typename Arc::Weight;
+
+ public:
+  EncodeMapper(uint32 flags, EncodeType type)
+      : flags_(flags),
+        type_(type),
+        table_(std::make_shared<internal::EncodeTable<Arc>>(flags)),
+        error_(false) {}
+
+  EncodeMapper(const EncodeMapper &mapper)
+      : flags_(mapper.flags_),
+        type_(mapper.type_),
+        table_(mapper.table_),
+        error_(false) {}
+
+  // Copy constructor but setting the type, typically to DECODE.
+  EncodeMapper(const EncodeMapper &mapper, EncodeType type)
+      : flags_(mapper.flags_),
+        type_(type),
+        table_(mapper.table_),
+        error_(mapper.error_) {}
+
+  Arc operator()(const Arc &arc);
+
+  MapFinalAction FinalAction() const {
+    return (type_ == ENCODE && (flags_ & kEncodeWeights))
+               ? MAP_REQUIRE_SUPERFINAL
+               : MAP_NO_SUPERFINAL;
+  }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_CLEAR_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_CLEAR_SYMBOLS;
+  }
+
+  uint64 Properties(uint64 inprops) {
+    uint64 outprops = inprops;
+    if (error_) outprops |= kError;
+    uint64 mask = kFstProperties;
+    if (flags_ & kEncodeLabels) {
+      mask &= kILabelInvariantProperties & kOLabelInvariantProperties;
+    }
+    if (flags_ & kEncodeWeights) {
+      mask &= kILabelInvariantProperties & kWeightInvariantProperties &
+              (type_ == ENCODE ? kAddSuperFinalProperties
+                               : kRmSuperFinalProperties);
+    }
+    return outprops & mask;
+  }
+
+  uint32 Flags() const { return flags_; }
+
+  EncodeType Type() const { return type_; }
+
+  bool Write(std::ostream &strm, const string &source) const {
+    return table_->Write(strm, source);
+  }
+
+  bool Write(const string &filename) const {
+    std::ofstream strm(filename,
+                             std::ios_base::out | std::ios_base::binary);
+    if (!strm) {
+      LOG(ERROR) << "EncodeMap: Can't open file: " << filename;
+      return false;
+    }
+    return Write(strm, filename);
+  }
+
+  static EncodeMapper<Arc> *Read(std::istream &strm, const string &source,
+                               EncodeType type = ENCODE) {
+    auto *table = internal::EncodeTable<Arc>::Read(strm, source);
+    return table ? new EncodeMapper(table->Flags(), type, table) : nullptr;
+  }
+
+  static EncodeMapper<Arc> *Read(const string &filename,
+                                 EncodeType type = ENCODE) {
+    std::ifstream strm(filename,
+                            std::ios_base::in | std::ios_base::binary);
+    if (!strm) {
+      LOG(ERROR) << "EncodeMap: Can't open file: " << filename;
+      return nullptr;
+    }
+    return Read(strm, filename, type);
+  }
+
+  const SymbolTable *InputSymbols() const { return table_->InputSymbols(); }
+
+  const SymbolTable *OutputSymbols() const { return table_->OutputSymbols(); }
+
+  void SetInputSymbols(const SymbolTable *syms) {
+    table_->SetInputSymbols(syms);
+  }
+
+  void SetOutputSymbols(const SymbolTable *syms) {
+    table_->SetOutputSymbols(syms);
+  }
+
+ private:
+  uint32 flags_;
+  EncodeType type_;
+  std::shared_ptr<internal::EncodeTable<Arc>> table_;
+  bool error_;
+
+  explicit EncodeMapper(uint32 flags, EncodeType type,
+                        internal::EncodeTable<Arc> *table)
+      : flags_(flags), type_(type), table_(table), error_(false) {}
+
+  EncodeMapper &operator=(const EncodeMapper &) = delete;
+};
+
+template <class Arc>
+Arc EncodeMapper<Arc>::operator()(const Arc &arc) {
+  if (type_ == ENCODE) {
+    if ((arc.nextstate == kNoStateId && !(flags_ & kEncodeWeights)) ||
+        (arc.nextstate == kNoStateId && (flags_ & kEncodeWeights) &&
+         arc.weight == Weight::Zero())) {
+      return arc;
+    } else {
+      const auto label = table_->Encode(arc);
+      return Arc(label, flags_ & kEncodeLabels ? label : arc.olabel,
+                 flags_ & kEncodeWeights ? Weight::One() : arc.weight,
+                 arc.nextstate);
+    }
+  } else {  // type_ == DECODE
+    if (arc.nextstate == kNoStateId) {
+      return arc;
+    } else {
+      if (arc.ilabel == 0) return arc;
+      if (flags_ & kEncodeLabels && arc.ilabel != arc.olabel) {
+        FSTERROR() << "EncodeMapper: Label-encoded arc has different "
+                      "input and output labels";
+        error_ = true;
+      }
+      if (flags_ & kEncodeWeights && arc.weight != Weight::One()) {
+        FSTERROR() << "EncodeMapper: Weight-encoded arc has non-trivial weight";
+        error_ = true;
+      }
+      const auto tuple = table_->Decode(arc.ilabel);
+      if (!tuple) {
+        FSTERROR() << "EncodeMapper: Decode failed";
+        error_ = true;
+        return Arc(kNoLabel, kNoLabel, Weight::NoWeight(), arc.nextstate);
+      } else {
+        return Arc(tuple->ilabel,
+                   flags_ & kEncodeLabels ? tuple->olabel : arc.olabel,
+                   flags_ & kEncodeWeights ? tuple->weight : arc.weight,
+                   arc.nextstate);
+      }
+    }
+  }
+}
+
+// Complexity: O(E + V).
+template <class Arc>
+inline void Encode(MutableFst<Arc> *fst, EncodeMapper<Arc> *mapper) {
+  mapper->SetInputSymbols(fst->InputSymbols());
+  mapper->SetOutputSymbols(fst->OutputSymbols());
+  ArcMap(fst, mapper);
+}
+
+template <class Arc>
+inline void Decode(MutableFst<Arc> *fst, const EncodeMapper<Arc> &mapper) {
+  ArcMap(fst, EncodeMapper<Arc>(mapper, DECODE));
+  RmFinalEpsilon(fst);
+  fst->SetInputSymbols(mapper.InputSymbols());
+  fst->SetOutputSymbols(mapper.OutputSymbols());
+}
+
+// On-the-fly encoding of an input FST.
+//
+// Complexity:
+//
+//   Construction: O(1)
+//   Traversal: O(e + v)
+//
+// where e is the number of arcs visited and v is the number of states visited.
+// Constant time and space to visit an input state or arc is assumed and
+// exclusive of caching.
+template <class Arc>
+class EncodeFst : public ArcMapFst<Arc, Arc, EncodeMapper<Arc>> {
+ public:
+  using Mapper = EncodeMapper<Arc>;
+  using Impl = internal::ArcMapFstImpl<Arc, Arc, Mapper>;
+
+  EncodeFst(const Fst<Arc> &fst, Mapper *encoder)
+      : ArcMapFst<Arc, Arc, Mapper>(fst, encoder, ArcMapFstOptions()) {
+    encoder->SetInputSymbols(fst.InputSymbols());
+    encoder->SetOutputSymbols(fst.OutputSymbols());
+  }
+
+  EncodeFst(const Fst<Arc> &fst, const Mapper &encoder)
+      : ArcMapFst<Arc, Arc, Mapper>(fst, encoder, ArcMapFstOptions()) {}
+
+  // See Fst<>::Copy() for doc.
+  EncodeFst(const EncodeFst<Arc> &fst, bool copy = false)
+      : ArcMapFst<Arc, Arc, Mapper>(fst, copy) {}
+
+  // Makes a copy of this EncodeFst. See Fst<>::Copy() for further doc.
+  EncodeFst<Arc> *Copy(bool safe = false) const override {
+    if (safe) {
+      FSTERROR() << "EncodeFst::Copy(true): Not allowed";
+      GetImpl()->SetProperties(kError, kError);
+    }
+    return new EncodeFst(*this);
+  }
+
+ private:
+  using ImplToFst<Impl>::GetImpl;
+  using ImplToFst<Impl>::GetMutableImpl;
+};
+
+// On-the-fly decoding of an input FST.
+//
+// Complexity:
+//
+//   Construction: O(1).
+//   Traversal: O(e + v)
+//
+// Constant time and space to visit an input state or arc is assumed and
+// exclusive of caching.
+template <class Arc>
+class DecodeFst : public ArcMapFst<Arc, Arc, EncodeMapper<Arc>> {
+ public:
+  using Mapper = EncodeMapper<Arc>;
+  using Impl = internal::ArcMapFstImpl<Arc, Arc, Mapper>;
+  using ImplToFst<Impl>::GetImpl;
+
+  DecodeFst(const Fst<Arc> &fst, const Mapper &encoder)
+      : ArcMapFst<Arc, Arc, Mapper>(fst, Mapper(encoder, DECODE),
+                                    ArcMapFstOptions()) {
+    GetMutableImpl()->SetInputSymbols(encoder.InputSymbols());
+    GetMutableImpl()->SetOutputSymbols(encoder.OutputSymbols());
+  }
+
+  // See Fst<>::Copy() for doc.
+  DecodeFst(const DecodeFst<Arc> &fst, bool safe = false)
+      : ArcMapFst<Arc, Arc, Mapper>(fst, safe) {}
+
+  // Makes a copy of this DecodeFst. See Fst<>::Copy() for further doc.
+  DecodeFst<Arc> *Copy(bool safe = false) const override {
+    return new DecodeFst(*this, safe);
+  }
+
+ private:
+  using ImplToFst<Impl>::GetMutableImpl;
+};
+
+// Specialization for EncodeFst.
+template <class Arc>
+class StateIterator<EncodeFst<Arc>>
+    : public StateIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>> {
+ public:
+  explicit StateIterator(const EncodeFst<Arc> &fst)
+      : StateIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>>(fst) {}
+};
+
+// Specialization for EncodeFst.
+template <class Arc>
+class ArcIterator<EncodeFst<Arc>>
+    : public ArcIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>> {
+ public:
+  ArcIterator(const EncodeFst<Arc> &fst, typename Arc::StateId s)
+      : ArcIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>>(fst, s) {}
+};
+
+// Specialization for DecodeFst.
+template <class Arc>
+class StateIterator<DecodeFst<Arc>>
+    : public StateIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>> {
+ public:
+  explicit StateIterator(const DecodeFst<Arc> &fst)
+      : StateIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>>(fst) {}
+};
+
+// Specialization for DecodeFst.
+template <class Arc>
+class ArcIterator<DecodeFst<Arc>>
+    : public ArcIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>> {
+ public:
+  ArcIterator(const DecodeFst<Arc> &fst, typename Arc::StateId s)
+      : ArcIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>>(fst, s) {}
+};
+
+// Useful aliases when using StdArc.
+
+using StdEncodeFst = EncodeFst<StdArc>;
+
+using StdDecodeFst = DecodeFst<StdArc>;
+
+}  // namespace fst
+
+#endif  // FST_ENCODE_H_
diff --git a/projects/llm_framework/include/fst/epsnormalize.h b/projects/llm_framework/include/fst/epsnormalize.h
new file mode 100644
index 00000000..18105fb1
--- /dev/null
+++ b/projects/llm_framework/include/fst/epsnormalize.h
@@ -0,0 +1,61 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Function that implements epsilon-normalization.
+
+#ifndef FST_EPSNORMALIZE_H_
+#define FST_EPSNORMALIZE_H_
+
+
+#include <fst/arc-map.h>
+#include <fst/factor-weight.h>
+#include <fst/invert.h>
+#include <fst/rmepsilon.h>
+
+
+namespace fst {
+
+enum EpsNormalizeType { EPS_NORM_INPUT, EPS_NORM_OUTPUT };
+
+// Returns an equivalent FST that is epsilon-normalized. An acceptor is
+// epsilon-normalized if it is epsilon-removed. A transducer is input
+// epsilon-normalized if additionally if on each path any epsilon input
+// label follows all non-epsilon input labels. Output epsilon-normalized
+// is defined similarly.
+//
+// For more information, see:
+//
+// Mohri, M. 2002. Generic epsilon-removal and input epsilon-normalization
+// algorithms for weighted transducers. International Journal of Computer
+// Science, 13(1): 129-143, 2002.
+template <class Arc>
+void EpsNormalize(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
+                  EpsNormalizeType type = EPS_NORM_INPUT) {
+  EpsNormalize<Arc, GALLIC>(ifst, ofst, type);
+}
+
+// Same as above, except allows specifying explicitly the gallic weight type.
+template <class Arc, GallicType G>
+void EpsNormalize(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
+                  EpsNormalizeType type) {
+  VectorFst<GallicArc<Arc, G>> gfst;
+  std::unique_ptr<SymbolTable> symbols;
+  if (type == EPS_NORM_INPUT) {
+    ArcMap(ifst, &gfst, ToGallicMapper<Arc, G>());
+    if (ifst.OutputSymbols()) symbols.reset(ifst.OutputSymbols()->Copy());
+  } else {  // type == EPS_NORM_OUTPUT
+    ArcMap(InvertFst<Arc>(ifst), &gfst, ToGallicMapper<Arc, G>());
+    if (ifst.InputSymbols()) symbols.reset(ifst.InputSymbols()->Copy());
+  }
+  RmEpsilon(&gfst);
+  FactorWeightFst<GallicArc<Arc, G>,
+                  GallicFactor<typename Arc::Label, typename Arc::Weight, G>>
+      fwfst(gfst);
+  ArcMap(fwfst, ofst, FromGallicMapper<Arc, G>());
+  ofst->SetOutputSymbols(symbols.get());
+  if (type == EPS_NORM_OUTPUT) Invert(ofst);
+}
+
+}  // namespace fst
+
+#endif  // FST_EPSNORMALIZE_H_
diff --git a/projects/llm_framework/include/fst/equal.h b/projects/llm_framework/include/fst/equal.h
new file mode 100644
index 00000000..ed89c6ce
--- /dev/null
+++ b/projects/llm_framework/include/fst/equal.h
@@ -0,0 +1,169 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Function to test equality of two FSTs.
+
+#ifndef FST_EQUAL_H_
+#define FST_EQUAL_H_
+
+#include <fst/log.h>
+
+#include <fst/fst.h>
+#include <fst/test-properties.h>
+
+
+namespace fst {
+
+constexpr uint32 kEqualFsts = 0x0001;
+constexpr uint32 kEqualFstTypes = 0x0002;
+constexpr uint32 kEqualCompatProperties = 0x0004;
+constexpr uint32 kEqualCompatSymbols = 0x0008;
+constexpr uint32 kEqualAll =
+    kEqualFsts | kEqualFstTypes | kEqualCompatProperties | kEqualCompatSymbols;
+
+class WeightApproxEqual {
+ public:
+  explicit WeightApproxEqual(float delta) : delta_(delta) {}
+
+  template <class Weight>
+  bool operator()(const Weight &w1, const Weight &w2) const {
+    return ApproxEqual(w1, w2, delta_);
+  }
+
+ private:
+  float delta_;
+};
+
+// Tests if two Fsts have the same states and arcs in the same order (when
+// etype & kEqualFst).
+// Also optional checks equality of Fst types (etype & kEqualFstTypes) and
+// compatibility of stored properties (etype & kEqualCompatProperties) and
+// of symbol tables (etype & kEqualCompatSymbols).
+template <class Arc, class WeightEqual>
+bool Equal(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+           WeightEqual weight_equal, uint32 etype = kEqualFsts) {
+  if ((etype & kEqualFstTypes) && (fst1.Type() != fst2.Type())) {
+    VLOG(1) << "Equal: Mismatched FST types (" << fst1.Type() << " != "
+            << fst2.Type() << ")";
+    return false;
+  }
+  if ((etype & kEqualCompatProperties) &&
+      !CompatProperties(fst1.Properties(kCopyProperties, false),
+                        fst2.Properties(kCopyProperties, false))) {
+    VLOG(1) << "Equal: Properties not compatible";
+    return false;
+  }
+  if (etype & kEqualCompatSymbols) {
+    if (!CompatSymbols(fst1.InputSymbols(), fst2.InputSymbols(), false)) {
+      VLOG(1) << "Equal: Input symbols not compatible";
+      return false;
+    }
+    if (!CompatSymbols(fst1.OutputSymbols(), fst2.OutputSymbols(), false)) {
+      VLOG(1) << "Equal: Output symbols not compatible";
+      return false;
+    }
+  }
+  if (!(etype & kEqualFsts)) return true;
+  if (fst1.Start() != fst2.Start()) {
+    VLOG(1) << "Equal: Mismatched start states (" << fst1.Start() << " != "
+            << fst2.Start() << ")";
+    return false;
+  }
+  StateIterator<Fst<Arc>> siter1(fst1);
+  StateIterator<Fst<Arc>> siter2(fst2);
+  while (!siter1.Done() || !siter2.Done()) {
+    if (siter1.Done() || siter2.Done()) {
+      VLOG(1) << "Equal: Mismatched number of states";
+      return false;
+    }
+    const auto s1 = siter1.Value();
+    const auto s2 = siter2.Value();
+    if (s1 != s2) {
+      VLOG(1) << "Equal: Mismatched states (" << s1 << "!= "
+              << s2 << ")";
+      return false;
+    }
+    const auto &final1 = fst1.Final(s1);
+    const auto &final2 = fst2.Final(s2);
+    if (!weight_equal(final1, final2)) {
+      VLOG(1) << "Equal: Mismatched final weights at state " << s1
+              << " (" << final1 << " != " << final2 << ")";
+      return false;
+    }
+    ArcIterator<Fst<Arc>> aiter1(fst1, s1);
+    ArcIterator<Fst<Arc>> aiter2(fst2, s2);
+    for (auto a = 0; !aiter1.Done() || !aiter2.Done(); ++a) {
+      if (aiter1.Done() || aiter2.Done()) {
+        VLOG(1) << "Equal: Mismatched number of arcs at state " << s1;
+        return false;
+      }
+      const auto &arc1 = aiter1.Value();
+      const auto &arc2 = aiter2.Value();
+      if (arc1.ilabel != arc2.ilabel) {
+        VLOG(1) << "Equal: Mismatched arc input labels at state " << s1
+                << ", arc " << a << " (" << arc1.ilabel << " != "
+                << arc2.ilabel << ")";
+        return false;
+      } else if (arc1.olabel != arc2.olabel) {
+        VLOG(1) << "Equal: Mismatched arc output labels at state " << s1
+                << ", arc " << a << " (" << arc1.olabel << " != "
+                << arc2.olabel << ")";
+        return false;
+      } else if (!weight_equal(arc1.weight, arc2.weight)) {
+        VLOG(1) << "Equal: Mismatched arc weights at state " << s1
+                << ", arc " << a << " (" << arc1.weight << " != "
+                << arc2.weight << ")";
+        return false;
+      } else if (arc1.nextstate != arc2.nextstate) {
+        VLOG(1) << "Equal: Mismatched next state at state " << s1
+                << ", arc " << a << " (" << arc1.nextstate << " != "
+                << arc2.nextstate << ")";
+        return false;
+      }
+      aiter1.Next();
+      aiter2.Next();
+    }
+    // Sanity checks: should never fail.
+    if (fst1.NumArcs(s1) != fst2.NumArcs(s2)) {
+      FSTERROR() << "Equal: Inconsistent arc counts at state " << s1
+                 << " (" << fst1.NumArcs(s1) << " != "
+                 << fst2.NumArcs(s2) << ")";
+      return false;
+    }
+    if (fst1.NumInputEpsilons(s1) != fst2.NumInputEpsilons(s2)) {
+      FSTERROR() << "Equal: Inconsistent input epsilon counts at state " << s1
+                 << " (" << fst1.NumInputEpsilons(s1) << " != "
+                 << fst2.NumInputEpsilons(s2) << ")";
+      return false;
+    }
+    if (fst1.NumOutputEpsilons(s1) != fst2.NumOutputEpsilons(s2)) {
+      FSTERROR() << "Equal: Inconsistent output epsilon counts at state " << s1
+                 << " (" << fst1.NumOutputEpsilons(s1) << " != "
+                 << fst2.NumOutputEpsilons(s2) << ")";
+    }
+    siter1.Next();
+    siter2.Next();
+  }
+  return true;
+}
+
+template <class Arc>
+bool Equal(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+           float delta = kDelta, uint32 etype = kEqualFsts) {
+  return Equal(fst1, fst2, WeightApproxEqual(delta), etype);
+}
+
+// Support double deltas without forcing all clients to cast to float.
+// Without this overload, Equal<Arc, WeightEqual=double> will be chosen,
+// since it is a better match than double -> float narrowing, but
+// the instantiation will fail.
+template <class Arc>
+bool Equal(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+           double delta, uint32 etype = kEqualFsts) {
+  return Equal(fst1, fst2, WeightApproxEqual(static_cast<float>(delta)), etype);
+}
+
+
+}  // namespace fst
+
+#endif  // FST_EQUAL_H_
diff --git a/projects/llm_framework/include/fst/equivalent.h b/projects/llm_framework/include/fst/equivalent.h
new file mode 100644
index 00000000..cf3fdb61
--- /dev/null
+++ b/projects/llm_framework/include/fst/equivalent.h
@@ -0,0 +1,230 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes to determine the equivalence of two FSTs.
+
+#ifndef FST_EQUIVALENT_H_
+#define FST_EQUIVALENT_H_
+
+#include <algorithm>
+#include <deque>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+#include <fst/log.h>
+
+#include <fst/encode.h>
+#include <fst/push.h>
+#include <fst/union-find.h>
+#include <fst/vector-fst.h>
+
+
+namespace fst {
+namespace internal {
+
+// Traits-like struct holding utility functions/typedefs/constants for
+// the equivalence algorithm.
+//
+// Encoding device: in order to make the statesets of the two acceptors
+// disjoint, we map Arc::StateId on the type MappedId. The states of
+// the first acceptor are mapped on odd numbers (s -> 2s + 1), and
+// those of the second one on even numbers (s -> 2s + 2). The number 0
+// is reserved for an implicit (non-final) dead state (required for
+// the correct treatment of non-coaccessible states; kNoStateId is mapped to
+// kDeadState for both acceptors). The union-find algorithm operates on the
+// mapped IDs.
+template <class Arc>
+struct EquivalenceUtil {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using MappedId = StateId;  // ID for an equivalence class.
+
+  // MappedId for an implicit dead state.
+  static constexpr MappedId kDeadState = 0;
+
+  // MappedId for lookup failure.
+  static constexpr MappedId kInvalidId = -1;
+
+  // Maps state ID to the representative of the corresponding
+  // equivalence class. The parameter 'which_fst' takes the values 1
+  // and 2, identifying the input FST.
+  static MappedId MapState(StateId s, int32 which_fst) {
+    return (kNoStateId == s) ? kDeadState
+                             : (static_cast<MappedId>(s) << 1) + which_fst;
+  }
+
+  // Maps set ID to State ID.
+  static StateId UnMapState(MappedId id) {
+    return static_cast<StateId>((--id) >> 1);
+  }
+
+  // Convenience function: checks if state with MappedId s is final in
+  // acceptor fa.
+  static bool IsFinal(const Fst<Arc> &fa, MappedId s) {
+    return (kDeadState == s) ? false
+                             : (fa.Final(UnMapState(s)) != Weight::Zero());
+  }
+  // Convenience function: returns the representative of ID in sets,
+  // creating a new set if needed.
+  static MappedId FindSet(UnionFind<MappedId> *sets, MappedId id) {
+    const auto repr = sets->FindSet(id);
+    if (repr != kInvalidId) {
+      return repr;
+    } else {
+      sets->MakeSet(id);
+      return id;
+    }
+  }
+};
+
+template <class Arc>
+constexpr
+    typename EquivalenceUtil<Arc>::MappedId EquivalenceUtil<Arc>::kDeadState;
+
+template <class Arc>
+constexpr
+    typename EquivalenceUtil<Arc>::MappedId EquivalenceUtil<Arc>::kInvalidId;
+
+}  // namespace internal
+
+// Equivalence checking algorithm: determines if the two FSTs fst1 and fst2
+// are equivalent. The input FSTs must be deterministic input-side epsilon-free
+// acceptors, unweighted or with weights over a left semiring. Two acceptors are
+// considered equivalent if they accept exactly the same set of strings (with
+// the same weights).
+//
+// The algorithm (cf. Aho, Hopcroft and Ullman, "The Design and Analysis of
+// Computer Programs") successively constructs sets of states that can be
+// reached by the same prefixes, starting with a set containing the start states
+// of both acceptors. A disjoint tree forest (the union-find algorithm) is used
+// to represent the sets of states. The algorithm returns false if one of the
+// constructed sets contains both final and non-final states. Returns an
+// optional error value (useful when FLAGS_error_fatal = false).
+//
+// Complexity:
+//
+// Quasi-linear, i.e., O(n G(n)), where
+//
+//   n = |S1| + |S2| is the number of states in both acceptors
+//
+//   G(n) is a very slowly growing function that can be approximated
+//        by 4 by all practical purposes.
+template <class Arc>
+bool Equivalent(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+                float delta = kDelta, bool *error = nullptr) {
+  using Weight = typename Arc::Weight;
+  if (error) *error = false;
+  // Check that the symbol table are compatible.
+  if (!CompatSymbols(fst1.InputSymbols(), fst2.InputSymbols()) ||
+      !CompatSymbols(fst1.OutputSymbols(), fst2.OutputSymbols())) {
+    FSTERROR() << "Equivalent: Input/output symbol tables of 1st argument "
+               << "do not match input/output symbol tables of 2nd argument";
+    if (error) *error = true;
+    return false;
+  }
+  // Check properties first.
+  static constexpr auto props = kNoEpsilons | kIDeterministic | kAcceptor;
+  if (fst1.Properties(props, true) != props) {
+    FSTERROR() << "Equivalent: 1st argument not an"
+               << " epsilon-free deterministic acceptor";
+    if (error) *error = true;
+    return false;
+  }
+  if (fst2.Properties(props, true) != props) {
+    FSTERROR() << "Equivalent: 2nd argument not an"
+               << " epsilon-free deterministic acceptor";
+    if (error) *error = true;
+    return false;
+  }
+  if ((fst1.Properties(kUnweighted, true) != kUnweighted) ||
+      (fst2.Properties(kUnweighted, true) != kUnweighted)) {
+    VectorFst<Arc> efst1(fst1);
+    VectorFst<Arc> efst2(fst2);
+    Push(&efst1, REWEIGHT_TO_INITIAL, delta);
+    Push(&efst2, REWEIGHT_TO_INITIAL, delta);
+    ArcMap(&efst1, QuantizeMapper<Arc>(delta));
+    ArcMap(&efst2, QuantizeMapper<Arc>(delta));
+    EncodeMapper<Arc> mapper(kEncodeWeights | kEncodeLabels, ENCODE);
+    ArcMap(&efst1, &mapper);
+    ArcMap(&efst2, &mapper);
+    return Equivalent(efst1, efst2);
+  }
+  using Util = internal::EquivalenceUtil<Arc>;
+  using MappedId = typename Util::MappedId;
+  enum { FST1 = 1, FST2 = 2 };  // Required by Util::MapState(...)
+  auto s1 = Util::MapState(fst1.Start(), FST1);
+  auto s2 = Util::MapState(fst2.Start(), FST2);
+  // The union-find structure.
+  UnionFind<MappedId> eq_classes(1000, Util::kInvalidId);
+  // Initializes the union-find structure.
+  eq_classes.MakeSet(s1);
+  eq_classes.MakeSet(s2);
+  // Data structure for the (partial) acceptor transition function of fst1 and
+  // fst2: input labels mapped to pairs of MappedIds representing destination
+  // states of the corresponding arcs in fst1 and fst2, respectively.
+  using Label2StatePairMap =
+      std::unordered_map<typename Arc::Label, std::pair<MappedId, MappedId>>;
+  Label2StatePairMap arc_pairs;
+  // Pairs of MappedId's to be processed, organized in a queue.
+  std::deque<std::pair<MappedId, MappedId>> q;
+  bool ret = true;
+  // Returns early if the start states differ w.r.t. finality.
+  if (Util::IsFinal(fst1, s1) != Util::IsFinal(fst2, s2)) ret = false;
+  // Main loop: explores the two acceptors in a breadth-first manner, updating
+  // the equivalence relation on the statesets. Loop invariant: each block of
+  // the states contains either final states only or non-final states only.
+  for (q.push_back(std::make_pair(s1, s2)); ret && !q.empty(); q.pop_front()) {
+    s1 = q.front().first;
+    s2 = q.front().second;
+    // Representatives of the equivalence classes of s1/s2.
+    const auto rep1 = Util::FindSet(&eq_classes, s1);
+    const auto rep2 = Util::FindSet(&eq_classes, s2);
+    if (rep1 != rep2) {
+      eq_classes.Union(rep1, rep2);
+      arc_pairs.clear();
+      // Copies outgoing arcs starting at s1 into the hash-table.
+      if (Util::kDeadState != s1) {
+        ArcIterator<Fst<Arc>> arc_iter(fst1, Util::UnMapState(s1));
+        for (; !arc_iter.Done(); arc_iter.Next()) {
+          const auto &arc = arc_iter.Value();
+          // Zero-weight arcs are treated as if they did not exist.
+          if (arc.weight != Weight::Zero()) {
+            arc_pairs[arc.ilabel].first = Util::MapState(arc.nextstate, FST1);
+          }
+        }
+      }
+      // Copies outgoing arcs starting at s2 into the hashtable.
+      if (Util::kDeadState != s2) {
+        ArcIterator<Fst<Arc>> arc_iter(fst2, Util::UnMapState(s2));
+        for (; !arc_iter.Done(); arc_iter.Next()) {
+          const auto &arc = arc_iter.Value();
+          // Zero-weight arcs are treated as if they did not exist.
+          if (arc.weight != Weight::Zero()) {
+            arc_pairs[arc.ilabel].second = Util::MapState(arc.nextstate, FST2);
+          }
+        }
+      }
+      // Iterates through the hashtable and process pairs of target states.
+      for (const auto &arc_iter : arc_pairs) {
+        const auto &pair = arc_iter.second;
+        if (Util::IsFinal(fst1, pair.first) !=
+            Util::IsFinal(fst2, pair.second)) {
+          // Detected inconsistency: return false.
+          ret = false;
+          break;
+        }
+        q.push_back(pair);
+      }
+    }
+  }
+  if (fst1.Properties(kError, false) || fst2.Properties(kError, false)) {
+    if (error) *error = true;
+    return false;
+  }
+  return ret;
+}
+
+}  // namespace fst
+
+#endif  // FST_EQUIVALENT_H_
diff --git a/projects/llm_framework/include/fst/expanded-fst.h b/projects/llm_framework/include/fst/expanded-fst.h
new file mode 100644
index 00000000..2c7d514c
--- /dev/null
+++ b/projects/llm_framework/include/fst/expanded-fst.h
@@ -0,0 +1,179 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Generic FST augmented with state count-interface class definition.
+
+#ifndef FST_EXPANDED_FST_H_
+#define FST_EXPANDED_FST_H_
+
+#include <sys/types.h>
+#include <istream>
+#include <memory>
+#include <string>
+
+#include <fst/log.h>
+#include <fstream>
+
+#include <fst/fst.h>
+
+
+namespace fst {
+
+// A generic FST plus state count.
+template <class A>
+class ExpandedFst : public Fst<A> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+
+  virtual StateId NumStates() const = 0;  // State count
+
+  // Get a copy of this ExpandedFst. See Fst<>::Copy() for further doc.
+  ExpandedFst<Arc> *Copy(bool safe = false) const override = 0;
+
+  // Read an ExpandedFst from an input stream; return NULL on error.
+  static ExpandedFst<Arc> *Read(std::istream &strm,
+                                const FstReadOptions &opts) {
+    FstReadOptions ropts(opts);
+    FstHeader hdr;
+    if (ropts.header) {
+      hdr = *opts.header;
+    } else {
+      if (!hdr.Read(strm, opts.source)) return nullptr;
+      ropts.header = &hdr;
+    }
+    if (!(hdr.Properties() & kExpanded)) {
+      LOG(ERROR) << "ExpandedFst::Read: Not an ExpandedFst: " << ropts.source;
+      return nullptr;
+    }
+    const auto reader =
+        FstRegister<Arc>::GetRegister()->GetReader(hdr.FstType());
+    if (!reader) {
+      LOG(ERROR) << "ExpandedFst::Read: Unknown FST type \"" << hdr.FstType()
+                 << "\" (arc type = \"" << A::Type() << "\"): " << ropts.source;
+      return nullptr;
+    }
+    auto *fst = reader(strm, ropts);
+    if (!fst) return nullptr;
+    return static_cast<ExpandedFst<Arc> *>(fst);
+  }
+
+  // Read an ExpandedFst from a file; return NULL on error.
+  // Empty filename reads from standard input.
+  static ExpandedFst<Arc> *Read(const string &filename) {
+    if (!filename.empty()) {
+      std::ifstream strm(filename,
+                              std::ios_base::in | std::ios_base::binary);
+      if (!strm) {
+        LOG(ERROR) << "ExpandedFst::Read: Can't open file: " << filename;
+        return nullptr;
+      }
+      return Read(strm, FstReadOptions(filename));
+    } else {
+      return Read(std::cin, FstReadOptions("standard input"));
+    }
+  }
+};
+
+namespace internal {
+
+//  ExpandedFst<A> case - abstract methods.
+template <class Arc>
+inline typename Arc::Weight Final(const ExpandedFst<Arc> &fst,
+                                  typename Arc::StateId s) {
+  return fst.Final(s);
+}
+
+template <class Arc>
+inline ssize_t NumArcs(const ExpandedFst<Arc> &fst, typename Arc::StateId s) {
+  return fst.NumArcs(s);
+}
+
+template <class Arc>
+inline ssize_t NumInputEpsilons(const ExpandedFst<Arc> &fst,
+                                typename Arc::StateId s) {
+  return fst.NumInputEpsilons(s);
+}
+
+template <class Arc>
+inline ssize_t NumOutputEpsilons(const ExpandedFst<Arc> &fst,
+                                 typename Arc::StateId s) {
+  return fst.NumOutputEpsilons(s);
+}
+
+}  // namespace internal
+
+// A useful alias when using StdArc.
+using StdExpandedFst = ExpandedFst<StdArc>;
+
+// This is a helper class template useful for attaching an ExpandedFst
+// interface to its implementation, handling reference counting. It
+// delegates to ImplToFst the handling of the Fst interface methods.
+template <class Impl, class FST = ExpandedFst<typename Impl::Arc>>
+class ImplToExpandedFst : public ImplToFst<Impl, FST> {
+ public:
+  using Arc = typename FST::Arc;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  StateId NumStates() const override { return GetImpl()->NumStates(); }
+
+ protected:
+  using ImplToFst<Impl, FST>::GetImpl;
+
+  explicit ImplToExpandedFst(std::shared_ptr<Impl> impl)
+      : ImplToFst<Impl, FST>(impl) {}
+
+  ImplToExpandedFst(const ImplToExpandedFst<Impl, FST> &fst, bool safe)
+      : ImplToFst<Impl, FST>(fst, safe) {}
+
+  static Impl *Read(std::istream &strm, const FstReadOptions &opts) {
+    return Impl::Read(strm, opts);
+  }
+
+  // Read FST implementation from a file; return NULL on error.
+  // Empty filename reads from standard input.
+  static Impl *Read(const string &filename) {
+    if (!filename.empty()) {
+      std::ifstream strm(filename,
+                              std::ios_base::in | std::ios_base::binary);
+      if (!strm) {
+        LOG(ERROR) << "ExpandedFst::Read: Can't open file: " << filename;
+        return nullptr;
+      }
+      return Impl::Read(strm, FstReadOptions(filename));
+    } else {
+      return Impl::Read(std::cin, FstReadOptions("standard input"));
+    }
+  }
+};
+
+// Function to return the number of states in an FST, counting them
+// if necessary.
+template <class Arc>
+typename Arc::StateId CountStates(const Fst<Arc> &fst) {
+  if (fst.Properties(kExpanded, false)) {
+    const auto *efst = static_cast<const ExpandedFst<Arc> *>(&fst);
+    return efst->NumStates();
+  } else {
+    typename Arc::StateId nstates = 0;
+    for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
+      ++nstates;
+    }
+    return nstates;
+  }
+}
+
+// Function to return the number of arcs in an FST.
+template <class Arc>
+typename Arc::StateId CountArcs(const Fst<Arc> &fst) {
+  size_t narcs = 0;
+  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
+    narcs += fst.NumArcs(siter.Value());
+  }
+  return narcs;
+}
+
+}  // namespace fst
+
+#endif  // FST_EXPANDED_FST_H_
diff --git a/projects/llm_framework/include/fst/expectation-weight.h b/projects/llm_framework/include/fst/expectation-weight.h
new file mode 100644
index 00000000..f996cbc6
--- /dev/null
+++ b/projects/llm_framework/include/fst/expectation-weight.h
@@ -0,0 +1,134 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Expectation semiring as described by Jason Eisner:
+// See: doi=10.1.1.22.9398
+// Multiplex semiring operations and identities:
+//    One: <One, Zero>
+//    Zero: <Zero, Zero>
+//    Plus: <a1, b1> + <a2, b2> = < (a1 + a2) , (b1 + b2) >
+//    Times: <a1, b1> * <a2, b2> = < (a1 * a2) , [(a1 * b2) + (a2 * b1)] >
+//    Division: Undefined (currently)
+//
+// Usually used to store the pair <probability, random_variable> so that
+// ShortestDistance[Fst<ArcTpl<ExpectationWeight<P, V>>>]
+//    == < PosteriorProbability, Expected_Value[V] >
+
+#ifndef FST_EXPECTATION_WEIGHT_H_
+#define FST_EXPECTATION_WEIGHT_H_
+
+#include <string>
+
+#include <fst/log.h>
+
+#include <fst/pair-weight.h>
+#include <fst/product-weight.h>
+
+
+namespace fst {
+
+// X1 is usually a probability weight like LogWeight.
+// X2 is usually a random variable or vector (see SignedLogWeight or
+// SparsePowerWeight).
+//
+// If X1 is distinct from X2, it is required that there is an external product
+// between X1 and X2 and if both semriring are commutative, or left or right
+// semirings, then result must have those properties.
+template <class X1, class X2>
+class ExpectationWeight : public PairWeight<X1, X2> {
+ public:
+  using PairWeight<X1, X2>::Value1;
+  using PairWeight<X1, X2>::Value2;
+
+  using PairWeight<X1, X2>::Reverse;
+  using PairWeight<X1, X2>::Quantize;
+  using PairWeight<X1, X2>::Member;
+
+  using ReverseWeight =
+      ExpectationWeight<typename X1::ReverseWeight, typename X2::ReverseWeight>;
+
+  ExpectationWeight() : PairWeight<X1, X2>(Zero()) {}
+
+  explicit ExpectationWeight(const PairWeight<X1, X2> &weight)
+      : PairWeight<X1, X2>(weight) {}
+
+  ExpectationWeight(const X1 &x1, const X2 &x2) : PairWeight<X1, X2>(x1, x2) {}
+
+  static const ExpectationWeight &Zero() {
+    static const ExpectationWeight zero(X1::Zero(), X2::Zero());
+    return zero;
+  }
+
+  static const ExpectationWeight &One() {
+    static const ExpectationWeight one(X1::One(), X2::Zero());
+    return one;
+  }
+
+  static const ExpectationWeight &NoWeight() {
+    static const ExpectationWeight no_weight(X1::NoWeight(), X2::NoWeight());
+    return no_weight;
+  }
+
+  static const string &Type() {
+    static const string *const type =
+        new string("expectation_" + X1::Type() + "_" + X2::Type());
+    return *type;
+  }
+
+  PairWeight<X1, X2> Quantize(float delta = kDelta) const {
+    return ExpectationWeight(PairWeight<X1, X2>::Quantize());
+  }
+
+  ReverseWeight Reverse() const {
+    return ReverseWeight(PairWeight<X1, X2>::Reverse());
+  }
+
+  bool Member() const { return PairWeight<X1, X2>::Member(); }
+
+  static constexpr uint64 Properties() {
+    return X1::Properties() & X2::Properties() &
+           (kLeftSemiring | kRightSemiring | kCommutative | kIdempotent);
+  }
+};
+
+template <class X1, class X2>
+inline ExpectationWeight<X1, X2> Plus(const ExpectationWeight<X1, X2> &w1,
+                                      const ExpectationWeight<X1, X2> &w2) {
+  return ExpectationWeight<X1, X2>(Plus(w1.Value1(), w2.Value1()),
+                                   Plus(w1.Value2(), w2.Value2()));
+}
+
+template <class X1, class X2>
+inline ExpectationWeight<X1, X2> Times(const ExpectationWeight<X1, X2> &w1,
+                                       const ExpectationWeight<X1, X2> &w2) {
+  return ExpectationWeight<X1, X2>(
+      Times(w1.Value1(), w2.Value1()),
+      Plus(Times(w1.Value1(), w2.Value2()), Times(w1.Value2(), w2.Value1())));
+}
+
+template <class X1, class X2>
+inline ExpectationWeight<X1, X2> Divide(const ExpectationWeight<X1, X2> &w1,
+                                        const ExpectationWeight<X1, X2> &w2,
+                                        DivideType typ = DIVIDE_ANY) {
+  FSTERROR() << "ExpectationWeight::Divide: Not implemented";
+  return ExpectationWeight<X1, X2>::NoWeight();
+}
+
+// This function object generates weights by calling the underlying generators
+// for the template weight types, like all other pair weight types. This is
+// intended primarily for testing.
+template <class X1, class X2>
+class WeightGenerate<ExpectationWeight<X1, X2>>
+    : public WeightGenerate<PairWeight<X1, X2>> {
+ public:
+  using Weight = ExpectationWeight<X1, X2>;
+  using Generate = WeightGenerate<PairWeight<X1, X2>>;
+
+  explicit WeightGenerate(bool allow_zero = true) : Generate(allow_zero) {}
+
+  Weight operator()() const { return Weight(Generate::operator()()); }
+};
+
+}  // namespace fst
+
+#endif  // FST_EXPECTATION_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/extensions/compress/compress-script.h b/projects/llm_framework/include/fst/extensions/compress/compress-script.h
new file mode 100644
index 00000000..bad238aa
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/compress/compress-script.h
@@ -0,0 +1,53 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Declarations of 'scriptable' versions of compression operations, that is,
+// those that can be called with FstClass-type arguments.
+
+#ifndef FST_EXTENSIONS_COMPRESS_COMPRESS_SCRIPT_H_
+#define FST_EXTENSIONS_COMPRESS_COMPRESS_SCRIPT_H_
+
+#include <string>
+#include <tuple>
+
+#include <fst/log.h>
+#include <fst/extensions/compress/compress.h>
+#include <fst/mutable-fst.h>
+#include <fst/util.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+typedef std::tuple<const FstClass &, const string &, const bool> CompressArgs;
+
+template <class Arc>
+void Compress(CompressArgs *args) {
+  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
+  const string &filename = std::get<1>(*args);
+  const bool gzip = std::get<2>(*args);
+
+  if (!fst::Compress(fst, filename, gzip)) FSTERROR() << "Compress: failed";
+}
+
+void Compress(const FstClass &fst, const string &filename, const bool gzip);
+
+typedef std::tuple<const string &, MutableFstClass *, const bool>
+    DecompressArgs;
+
+template <class Arc>
+void Decompress(DecompressArgs *args) {
+  const string &filename = std::get<0>(*args);
+  MutableFst<Arc> *fst = std::get<1>(*args)->GetMutableFst<Arc>();
+  const bool gzip = std::get<2>(*args);
+
+  if (!fst::Decompress(filename, fst, gzip))
+    FSTERROR() << "Decompress: failed";
+}
+
+void Decompress(const string &filename, MutableFstClass *fst, const bool gzip);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_COMPRESS_COMPRESS_SCRIPT_H_
diff --git a/projects/llm_framework/include/fst/extensions/compress/compress.h b/projects/llm_framework/include/fst/extensions/compress/compress.h
new file mode 100644
index 00000000..aa94848f
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/compress/compress.h
@@ -0,0 +1,906 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Compresses and decompresses unweighted FSTs.
+
+#ifndef FST_EXTENSIONS_COMPRESS_COMPRESS_H_
+#define FST_EXTENSIONS_COMPRESS_COMPRESS_H_
+
+#include <algorithm>
+#include <cstdio>
+#include <iostream>
+#include <map>
+#include <memory>
+#include <queue>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include <fst/compat.h>
+#include <fst/log.h>
+#include <fst/extensions/compress/elias.h>
+#include <fst/extensions/compress/gzfile.h>
+#include <fst/encode.h>
+#include <fst/fst.h>
+#include <fst/mutable-fst.h>
+#include <fst/statesort.h>
+
+namespace fst {
+
+// Identifies stream data as a vanilla compressed FST.
+static const int32 kCompressMagicNumber = 1858869554;
+// Identifies stream data as (probably) a Gzip file accidentally read from
+// a vanilla stream, without gzip support.
+static const int32 kGzipMagicNumber = 0x8b1f;
+// Selects the two most significant bytes.
+constexpr uint32 kGzipMask = 0xffffffff >> 16;
+
+namespace internal {
+
+// Expands a Lempel Ziv code and returns the set of code words. expanded_code[i]
+// is the i^th Lempel Ziv codeword.
+template <class Var, class Edge>
+bool ExpandLZCode(const std::vector<std::pair<Var, Edge>> &code,
+                  std::vector<std::vector<Edge>> *expanded_code) {
+  expanded_code->resize(code.size());
+  for (int i = 0; i < code.size(); ++i) {
+    if (code[i].first > i) {
+      LOG(ERROR) << "ExpandLZCode: Not a valid code";
+      return false;
+    }
+    if (code[i].first == 0) {
+      (*expanded_code)[i].resize(1, code[i].second);
+    } else {
+      (*expanded_code)[i].resize((*expanded_code)[code[i].first - 1].size() +
+                                 1);
+      std::copy((*expanded_code)[code[i].first - 1].begin(),
+                (*expanded_code)[code[i].first - 1].end(),
+                (*expanded_code)[i].begin());
+      (*expanded_code)[i][(*expanded_code)[code[i].first - 1].size()] =
+          code[i].second;
+    }
+  }
+  return true;
+}
+
+}  // namespace internal
+
+// Lempel Ziv on data structure Edge, with a less than operator
+// EdgeLessThan and an equals operator  EdgeEquals.
+// Edge has a value defaultedge which it never takes and
+// Edge is defined, it is initialized to defaultedge
+template <class Var, class Edge, class EdgeLessThan, class EdgeEquals>
+class LempelZiv {
+ public:
+  LempelZiv() : dict_number_(0), default_edge_() {
+    root_.current_number = dict_number_++;
+    root_.current_edge = default_edge_;
+    decode_vector_.push_back(std::make_pair(0, default_edge_));
+  }
+  // Encodes a vector input into output
+  void BatchEncode(const std::vector<Edge> &input,
+                   std::vector<std::pair<Var, Edge>> *output);
+
+  // Decodes codedvector to output. Returns false if
+  // the index exceeds the size.
+  bool BatchDecode(const std::vector<std::pair<Var, Edge>> &input,
+                   std::vector<Edge> *output);
+
+  // Decodes a single dictionary element. Returns false
+  // if the index exceeds the size.
+  bool SingleDecode(const Var &index, Edge *output) {
+    if (index >= decode_vector_.size()) {
+      LOG(ERROR) << "LempelZiv::SingleDecode: "
+                 << "Index exceeded the dictionary size";
+      return false;
+    } else {
+      *output = decode_vector_[index].second;
+      return true;
+    }
+  }
+
+  ~LempelZiv() {
+    for (auto it = (root_.next_number).begin(); it != (root_.next_number).end();
+         ++it) {
+      CleanUp(it->second);
+    }
+  }
+  // Adds a single dictionary element while decoding
+  //  void AddDictElement(const std::pair<Var, Edge> &newdict) {
+  //    EdgeEquals InstEdgeEquals;
+  //  if (InstEdgeEquals(newdict.second, default_edge_) != 1)
+  //     decode_vector_.push_back(newdict);
+  //  }
+
+ private:
+  // Node datastructure is used for encoding
+
+  struct Node {
+    Var current_number;
+    Edge current_edge;
+    std::map<Edge, Node *, EdgeLessThan> next_number;
+  };
+
+  void CleanUp(Node *temp) {
+    for (auto it = (temp->next_number).begin(); it != (temp->next_number).end();
+         ++it) {
+      CleanUp(it->second);
+    }
+    delete temp;
+  }
+  Node root_;
+  Var dict_number_;
+  // decode_vector_ is used for decoding
+  std::vector<std::pair<Var, Edge>> decode_vector_;
+  Edge default_edge_;
+};
+
+template <class Var, class Edge, class EdgeLessThan, class EdgeEquals>
+void LempelZiv<Var, Edge, EdgeLessThan, EdgeEquals>::BatchEncode(
+    const std::vector<Edge> &input, std::vector<std::pair<Var, Edge>> *output) {
+  for (typename std::vector<Edge>::const_iterator it = input.begin();
+       it != input.end(); ++it) {
+    Node *temp_node = &root_;
+    while (it != input.end()) {
+      auto next = (temp_node->next_number).find(*it);
+      if (next != (temp_node->next_number).end()) {
+        temp_node = next->second;
+        ++it;
+      } else {
+        break;
+      }
+    }
+    if (it == input.end() && temp_node->current_number != 0) {
+      output->push_back(
+          std::make_pair(temp_node->current_number, default_edge_));
+    } else if (it != input.end()) {
+      output->push_back(std::make_pair(temp_node->current_number, *it));
+      Node *new_node = new (Node);
+      new_node->current_number = dict_number_++;
+      new_node->current_edge = *it;
+      (temp_node->next_number)[*it] = new_node;
+    }
+    if (it == input.end()) break;
+  }
+}
+
+template <class Var, class Edge, class EdgeLessThan, class EdgeEquals>
+bool LempelZiv<Var, Edge, EdgeLessThan, EdgeEquals>::BatchDecode(
+    const std::vector<std::pair<Var, Edge>> &input, std::vector<Edge> *output) {
+  for (typename std::vector<std::pair<Var, Edge>>::const_iterator it =
+           input.begin();
+       it != input.end(); ++it) {
+    std::vector<Edge> temp_output;
+    EdgeEquals InstEdgeEquals;
+    if (InstEdgeEquals(it->second, default_edge_) != 1) {
+      decode_vector_.push_back(*it);
+      temp_output.push_back(it->second);
+    }
+    Var temp_integer = it->first;
+    if (temp_integer >= decode_vector_.size()) {
+      LOG(ERROR) << "LempelZiv::BatchDecode: "
+                 << "Index exceeded the dictionary size";
+      return false;
+    } else {
+      while (temp_integer != 0) {
+        temp_output.push_back(decode_vector_[temp_integer].second);
+        temp_integer = decode_vector_[temp_integer].first;
+      }
+      std::reverse(temp_output.begin(), temp_output.end());
+      output->insert(output->end(), temp_output.begin(), temp_output.end());
+    }
+  }
+  return true;
+}
+
+// The main Compressor class
+template <class Arc>
+class Compressor {
+ public:
+  typedef typename Arc::StateId StateId;
+  typedef typename Arc::Label Label;
+  typedef typename Arc::Weight Weight;
+
+  Compressor() {}
+
+  // Compresses fst into a boolean vector code. Returns true on sucesss.
+  bool Compress(const Fst<Arc> &fst, std::ostream &strm);
+
+  // Decompresses the boolean vector into Fst. Returns true on sucesss.
+  bool Decompress(std::istream &strm, const string &source,
+                  MutableFst<Arc> *fst);
+
+  // Finds the BFS order of a fst
+  void BfsOrder(const ExpandedFst<Arc> &fst, std::vector<StateId> *order);
+
+  // Preprocessing step to convert fst to a isomorphic fst
+  // Returns a preproccess fst and a dictionary
+  void Preprocess(const Fst<Arc> &fst, MutableFst<Arc> *preprocessedfst,
+                  EncodeMapper<Arc> *encoder);
+
+  // Performs Lempel Ziv and outputs a stream of integers
+  // and sends it to a stream
+  void EncodeProcessedFst(const ExpandedFst<Arc> &fst, std::ostream &strm);
+
+  // Decodes fst from the stream
+  void DecodeProcessedFst(const std::vector<StateId> &input,
+                          MutableFst<Arc> *fst, bool unweighted);
+
+  // Converts buffer_code_ to uint8 and writes to a stream.
+
+  // Writes the boolean file to the stream
+  void WriteToStream(std::ostream &strm);
+
+  // Writes the weights to the stream
+  void WriteWeight(const std::vector<Weight> &input, std::ostream &strm);
+
+  void ReadWeight(std::istream &strm, std::vector<Weight> *output);
+
+  // Same as fst::Decode without the line RmFinalEpsilon(fst)
+  void DecodeForCompress(MutableFst<Arc> *fst, const EncodeMapper<Arc> &mapper);
+
+  // Updates the buffer_code_
+  template <class CVar>
+  void WriteToBuffer(CVar input) {
+    std::vector<bool> current_code;
+    Elias<CVar>::DeltaEncode(input, &current_code);
+    if (!buffer_code_.empty()) {
+      buffer_code_.insert(buffer_code_.end(), current_code.begin(),
+                          current_code.end());
+    } else {
+      buffer_code_.assign(current_code.begin(), current_code.end());
+    }
+  }
+
+ private:
+  struct LZLabel {
+    LZLabel() : label(0) {}
+    Label label;
+  };
+
+  struct LabelLessThan {
+    bool operator()(const LZLabel &labelone, const LZLabel &labeltwo) const {
+      return labelone.label < labeltwo.label;
+    }
+  };
+
+  struct LabelEquals {
+    bool operator()(const LZLabel &labelone, const LZLabel &labeltwo) const {
+      return labelone.label == labeltwo.label;
+    }
+  };
+
+  struct Transition {
+    Transition() : nextstate(0), label(0), weight(Weight::Zero()) {}
+
+    StateId nextstate;
+    Label label;
+    Weight weight;
+  };
+
+  struct TransitionLessThan {
+    bool operator()(const Transition &transition_one,
+                    const Transition &transition_two) const {
+      if (transition_one.nextstate == transition_two.nextstate)
+        return transition_one.label < transition_two.label;
+      else
+        return transition_one.nextstate < transition_two.nextstate;
+    }
+  } transition_less_than;
+
+  struct TransitionEquals {
+    bool operator()(const Transition &transition_one,
+                    const Transition &transition_two) const {
+      return transition_one.nextstate == transition_two.nextstate &&
+             transition_one.label == transition_two.label;
+    }
+  } transition_equals;
+
+  struct OldDictCompare {
+    bool operator()(const std::pair<StateId, Transition> &pair_one,
+                    const std::pair<StateId, Transition> &pair_two) const {
+      if ((pair_one.second).nextstate == (pair_two.second).nextstate)
+        return (pair_one.second).label < (pair_two.second).label;
+      else
+        return (pair_one.second).nextstate < (pair_two.second).nextstate;
+    }
+  } old_dict_compare;
+
+  std::vector<bool> buffer_code_;
+  std::vector<Weight> arc_weight_;
+  std::vector<Weight> final_weight_;
+};
+
+template <class Arc>
+inline void Compressor<Arc>::DecodeForCompress(
+    MutableFst<Arc> *fst, const EncodeMapper<Arc> &mapper) {
+  ArcMap(fst, EncodeMapper<Arc>(mapper, DECODE));
+  fst->SetInputSymbols(mapper.InputSymbols());
+  fst->SetOutputSymbols(mapper.OutputSymbols());
+}
+
+// Compressor::BfsOrder
+template <class Arc>
+void Compressor<Arc>::BfsOrder(const ExpandedFst<Arc> &fst,
+                               std::vector<StateId> *order) {
+  Arc arc;
+  StateId bfs_visit_number = 0;
+  std::queue<StateId> states_queue;
+  order->assign(fst.NumStates(), kNoStateId);
+  states_queue.push(fst.Start());
+  (*order)[fst.Start()] = bfs_visit_number++;
+  while (!states_queue.empty()) {
+    for (ArcIterator<Fst<Arc>> aiter(fst, states_queue.front()); !aiter.Done();
+         aiter.Next()) {
+      arc = aiter.Value();
+      StateId nextstate = arc.nextstate;
+      if ((*order)[nextstate] == kNoStateId) {
+        (*order)[nextstate] = bfs_visit_number++;
+        states_queue.push(nextstate);
+      }
+    }
+    states_queue.pop();
+  }
+
+  // If the FST is unconnected, then the following
+  // code finds them
+  while (bfs_visit_number < fst.NumStates()) {
+    int unseen_state = 0;
+    for (unseen_state = 0; unseen_state < fst.NumStates(); ++unseen_state) {
+      if ((*order)[unseen_state] == kNoStateId) break;
+    }
+    states_queue.push(unseen_state);
+    (*order)[unseen_state] = bfs_visit_number++;
+    while (!states_queue.empty()) {
+      for (ArcIterator<Fst<Arc>> aiter(fst, states_queue.front());
+           !aiter.Done(); aiter.Next()) {
+        arc = aiter.Value();
+        StateId nextstate = arc.nextstate;
+        if ((*order)[nextstate] == kNoStateId) {
+          (*order)[nextstate] = bfs_visit_number++;
+          states_queue.push(nextstate);
+        }
+      }
+      states_queue.pop();
+    }
+  }
+}
+
+template <class Arc>
+void Compressor<Arc>::Preprocess(const Fst<Arc> &fst,
+                                 MutableFst<Arc> *preprocessedfst,
+                                 EncodeMapper<Arc> *encoder) {
+  *preprocessedfst = fst;
+  if (!preprocessedfst->NumStates()) {
+    return;
+  }
+  // Relabels the edges and develops a dictionary
+  Encode(preprocessedfst, encoder);
+  std::vector<StateId> order;
+  // Finds the BFS sorting order of the fst
+  BfsOrder(*preprocessedfst, &order);
+  // Reorders the states according to the BFS order
+  StateSort(preprocessedfst, order);
+}
+
+template <class Arc>
+void Compressor<Arc>::EncodeProcessedFst(const ExpandedFst<Arc> &fst,
+                                         std::ostream &strm) {
+  std::vector<StateId> output;
+  LempelZiv<StateId, LZLabel, LabelLessThan, LabelEquals> dict_new;
+  LempelZiv<StateId, Transition, TransitionLessThan, TransitionEquals> dict_old;
+  std::vector<LZLabel> current_new_input;
+  std::vector<Transition> current_old_input;
+  std::vector<std::pair<StateId, LZLabel>> current_new_output;
+  std::vector<std::pair<StateId, Transition>> current_old_output;
+  std::vector<StateId> final_states;
+
+  StateId number_of_states = fst.NumStates();
+
+  StateId seen_states = 0;
+  // Adding the number of states
+  WriteToBuffer<StateId>(number_of_states);
+
+  for (StateId state = 0; state < number_of_states; ++state) {
+    current_new_input.clear();
+    current_old_input.clear();
+    current_new_output.clear();
+    current_old_output.clear();
+    if (state > seen_states) ++seen_states;
+
+    // Collecting the final states
+    if (fst.Final(state) != Weight::Zero()) {
+      final_states.push_back(state);
+      final_weight_.push_back(fst.Final(state));
+    }
+
+    // Reading the states
+    for (ArcIterator<Fst<Arc>> aiter(fst, state); !aiter.Done(); aiter.Next()) {
+      Arc arc = aiter.Value();
+      if (arc.nextstate > seen_states) {  // RILEY: > or >= ?
+        ++seen_states;
+        LZLabel temp_label;
+        temp_label.label = arc.ilabel;
+        arc_weight_.push_back(arc.weight);
+        current_new_input.push_back(temp_label);
+      } else {
+        Transition temp_transition;
+        temp_transition.nextstate = arc.nextstate;
+        temp_transition.label = arc.ilabel;
+        temp_transition.weight = arc.weight;
+        current_old_input.push_back(temp_transition);
+      }
+    }
+    // Adding new states
+    dict_new.BatchEncode(current_new_input, &current_new_output);
+    WriteToBuffer<StateId>(current_new_output.size());
+
+    for (auto it = current_new_output.begin(); it != current_new_output.end();
+         ++it) {
+      WriteToBuffer<StateId>(it->first);
+      WriteToBuffer<Label>((it->second).label);
+    }
+    // Adding old states by sorting and using difference coding
+    std::sort(current_old_input.begin(), current_old_input.end(),
+              transition_less_than);
+    for (auto it = current_old_input.begin(); it != current_old_input.end();
+         ++it) {
+      arc_weight_.push_back(it->weight);
+    }
+    dict_old.BatchEncode(current_old_input, &current_old_output);
+    std::vector<StateId> dict_old_temp;
+    std::vector<Transition> transition_old_temp;
+    for (auto it = current_old_output.begin(); it != current_old_output.end();
+         ++it) {
+      dict_old_temp.push_back(it->first);
+      transition_old_temp.push_back(it->second);
+    }
+    if (!transition_old_temp.empty()) {
+      if ((transition_old_temp.back()).nextstate == 0 &&
+          (transition_old_temp.back()).label == 0) {
+        transition_old_temp.pop_back();
+      }
+    }
+    std::sort(dict_old_temp.begin(), dict_old_temp.end());
+    std::sort(transition_old_temp.begin(), transition_old_temp.end(),
+              transition_less_than);
+
+    WriteToBuffer<StateId>(dict_old_temp.size());
+    if (dict_old_temp.size() != transition_old_temp.size())
+      WriteToBuffer<int>(1);
+    else
+      WriteToBuffer<int>(0);
+
+    StateId previous;
+    if (!dict_old_temp.empty()) {
+      WriteToBuffer<StateId>(dict_old_temp.front());
+      previous = dict_old_temp.front();
+    }
+    if (dict_old_temp.size() > 1) {
+      for (auto it = dict_old_temp.begin() + 1; it != dict_old_temp.end();
+           ++it) {
+        WriteToBuffer<StateId>(*it - previous);
+        previous = *it;
+      }
+    }
+    if (!transition_old_temp.empty()) {
+      WriteToBuffer<StateId>((transition_old_temp.front()).nextstate);
+      previous = (transition_old_temp.front()).nextstate;
+      WriteToBuffer<Label>((transition_old_temp.front()).label);
+    }
+    if (transition_old_temp.size() > 1) {
+      for (auto it = transition_old_temp.begin() + 1;
+           it != transition_old_temp.end(); ++it) {
+        WriteToBuffer<StateId>(it->nextstate - previous);
+        previous = it->nextstate;
+        WriteToBuffer<StateId>(it->label);
+      }
+    }
+  }
+  // Adding final states
+  WriteToBuffer<StateId>(final_states.size());
+  if (!final_states.empty()) {
+    for (auto it = final_states.begin(); it != final_states.end(); ++it) {
+      WriteToBuffer<StateId>(*it);
+    }
+  }
+  WriteToStream(strm);
+  uint8 unweighted = (fst.Properties(kUnweighted, true) == kUnweighted);
+  WriteType(strm, unweighted);
+  if (unweighted == 0) {
+    WriteWeight(arc_weight_, strm);
+    WriteWeight(final_weight_, strm);
+  }
+}
+
+template <class Arc>
+void Compressor<Arc>::DecodeProcessedFst(const std::vector<StateId> &input,
+                                         MutableFst<Arc> *fst,
+                                         bool unweighted) {
+  LempelZiv<StateId, LZLabel, LabelLessThan, LabelEquals> dict_new;
+  LempelZiv<StateId, Transition, TransitionLessThan, TransitionEquals> dict_old;
+  std::vector<std::pair<StateId, LZLabel>> current_new_input;
+  std::vector<std::pair<StateId, Transition>> current_old_input;
+  std::vector<LZLabel> current_new_output;
+  std::vector<Transition> current_old_output;
+  std::vector<std::pair<StateId, Transition>> actual_old_dict_numbers;
+  std::vector<Transition> actual_old_dict_transitions;
+  auto arc_weight_it = arc_weight_.begin();
+  Transition default_transition;
+  StateId seen_states = 1;
+
+  // Adding states.
+  const StateId num_states = input.front();
+  if (num_states > 0) {
+    const StateId start_state = fst->AddState();
+    fst->SetStart(start_state);
+    for (StateId state = 1; state < num_states; ++state) {
+      fst->AddState();
+    }
+  }
+
+  typename std::vector<StateId>::const_iterator main_it = input.begin();
+  ++main_it;
+
+  for (StateId current_state = 0; current_state < num_states; ++current_state) {
+    if (current_state >= seen_states) ++seen_states;
+    current_new_input.clear();
+    current_new_output.clear();
+    current_old_input.clear();
+    current_old_output.clear();
+    // New states
+    StateId current_number_new_elements = *main_it;
+    ++main_it;
+    for (StateId new_integer = 0; new_integer < current_number_new_elements;
+         ++new_integer) {
+      std::pair<StateId, LZLabel> temp_new_dict_element;
+      temp_new_dict_element.first = *main_it;
+      ++main_it;
+      LZLabel temp_label;
+      temp_label.label = *main_it;
+      ++main_it;
+      temp_new_dict_element.second = temp_label;
+      current_new_input.push_back(temp_new_dict_element);
+    }
+    dict_new.BatchDecode(current_new_input, &current_new_output);
+    for (auto it = current_new_output.begin(); it != current_new_output.end();
+         ++it) {
+      if (!unweighted) {
+        fst->AddArc(current_state,
+                    Arc(it->label, it->label, *arc_weight_it, seen_states++));
+        ++arc_weight_it;
+      } else {
+        fst->AddArc(current_state,
+                    Arc(it->label, it->label, Weight::One(), seen_states++));
+      }
+    }
+
+    // Old states dictionary
+    StateId current_number_old_elements = *main_it;
+    ++main_it;
+    StateId is_zero_removed = *main_it;
+    ++main_it;
+    StateId previous = 0;
+    actual_old_dict_numbers.clear();
+    for (StateId new_integer = 0; new_integer < current_number_old_elements;
+         ++new_integer) {
+      std::pair<StateId, Transition> pair_temp_transition;
+      if (new_integer == 0) {
+        pair_temp_transition.first = *main_it;
+        previous = *main_it;
+      } else {
+        pair_temp_transition.first = *main_it + previous;
+        previous = pair_temp_transition.first;
+      }
+      ++main_it;
+      Transition temp_test;
+      if (!dict_old.SingleDecode(pair_temp_transition.first, &temp_test)) {
+        FSTERROR() << "Compressor::Decode: failed";
+        fst->DeleteStates();
+        fst->SetProperties(kError, kError);
+        return;
+      }
+      pair_temp_transition.second = temp_test;
+      actual_old_dict_numbers.push_back(pair_temp_transition);
+    }
+
+    // Reordering the dictionary elements
+    std::sort(actual_old_dict_numbers.begin(), actual_old_dict_numbers.end(),
+              old_dict_compare);
+
+    // Transitions
+    previous = 0;
+    actual_old_dict_transitions.clear();
+
+    for (StateId new_integer = 0;
+         new_integer < current_number_old_elements - is_zero_removed;
+         ++new_integer) {
+      Transition temp_transition;
+      if (new_integer == 0) {
+        temp_transition.nextstate = *main_it;
+        previous = *main_it;
+      } else {
+        temp_transition.nextstate = *main_it + previous;
+        previous = temp_transition.nextstate;
+      }
+      ++main_it;
+      temp_transition.label = *main_it;
+      ++main_it;
+      actual_old_dict_transitions.push_back(temp_transition);
+    }
+
+    if (is_zero_removed == 1) {
+      actual_old_dict_transitions.push_back(default_transition);
+    }
+
+    auto trans_it = actual_old_dict_transitions.begin();
+    auto dict_it = actual_old_dict_numbers.begin();
+
+    while (trans_it != actual_old_dict_transitions.end() &&
+           dict_it != actual_old_dict_numbers.end()) {
+      if (dict_it->first == 0) {
+        ++dict_it;
+      } else {
+        std::pair<StateId, Transition> temp_pair;
+        if (transition_equals(*trans_it, default_transition) == 1) {
+          temp_pair.first = dict_it->first;
+          temp_pair.second = default_transition;
+          ++dict_it;
+        } else if (transition_less_than(dict_it->second, *trans_it) == 1) {
+          temp_pair.first = dict_it->first;
+          temp_pair.second = *trans_it;
+          ++dict_it;
+        } else {
+          temp_pair.first = 0;
+          temp_pair.second = *trans_it;
+        }
+        ++trans_it;
+        current_old_input.push_back(temp_pair);
+      }
+    }
+    while (trans_it != actual_old_dict_transitions.end()) {
+      std::pair<StateId, Transition> temp_pair;
+      temp_pair.first = 0;
+      temp_pair.second = *trans_it;
+      ++trans_it;
+      current_old_input.push_back(temp_pair);
+    }
+
+    // Adding old elements in the dictionary
+    if (!dict_old.BatchDecode(current_old_input, &current_old_output)) {
+      FSTERROR() << "Compressor::Decode: Failed";
+      fst->DeleteStates();
+      fst->SetProperties(kError, kError);
+      return;
+    }
+
+    for (auto it = current_old_output.begin(); it != current_old_output.end();
+         ++it) {
+      if (!unweighted) {
+        fst->AddArc(current_state,
+                    Arc(it->label, it->label, *arc_weight_it, it->nextstate));
+        ++arc_weight_it;
+      } else {
+        fst->AddArc(current_state,
+                    Arc(it->label, it->label, Weight::One(), it->nextstate));
+      }
+    }
+  }
+  // Adding the final states
+  StateId number_of_final_states = *main_it;
+  if (number_of_final_states > 0) {
+    ++main_it;
+    for (StateId temp_int = 0; temp_int < number_of_final_states; ++temp_int) {
+      if (!unweighted) {
+        fst->SetFinal(*main_it, final_weight_[temp_int]);
+      } else {
+        fst->SetFinal(*main_it, Weight(0));
+      }
+      ++main_it;
+    }
+  }
+}
+
+template <class Arc>
+void Compressor<Arc>::ReadWeight(std::istream &strm,
+                                 std::vector<Weight> *output) {
+  int64 size;
+  Weight weight;
+  ReadType(strm, &size);
+  for (int64 i = 0; i < size; ++i) {
+    weight.Read(strm);
+    output->push_back(weight);
+  }
+}
+
+template <class Arc>
+bool Compressor<Arc>::Decompress(std::istream &strm, const string &source,
+                                 MutableFst<Arc> *fst) {
+  fst->DeleteStates();
+  int32 magic_number = 0;
+  ReadType(strm, &magic_number);
+  if (magic_number != kCompressMagicNumber) {
+    LOG(ERROR) << "Decompress: Bad compressed Fst: " << source;
+    // If the most significant two bytes of the magic number match the
+    // gzip magic number, then we are probably reading a gzip file as an
+    // ordinary stream.
+    if ((magic_number & kGzipMask) == kGzipMagicNumber) {
+      LOG(ERROR) << "Decompress: Fst appears to be compressed with Gzip, but "
+                    "gzip decompression was not requested. Try with "
+                    "the --gzip flag"
+                    ".";
+    }
+    return false;
+  }
+  std::unique_ptr<EncodeMapper<Arc>> encoder(
+      EncodeMapper<Arc>::Read(strm, "Decoding", DECODE));
+  std::vector<bool> bool_code;
+  uint8 block;
+  uint8 msb = 128;
+  int64 data_size;
+  ReadType(strm, &data_size);
+  for (int64 i = 0; i < data_size; ++i) {
+    ReadType(strm, &block);
+    for (int j = 0; j < 8; ++j) {
+      uint8 temp = msb & block;
+      if (temp == 128)
+        bool_code.push_back(1);
+      else
+        bool_code.push_back(0);
+      block = block << 1;
+    }
+  }
+  std::vector<StateId> int_code;
+  Elias<StateId>::BatchDecode(bool_code, &int_code);
+  bool_code.clear();
+  uint8 unweighted;
+  ReadType(strm, &unweighted);
+  if (unweighted == 0) {
+    ReadWeight(strm, &arc_weight_);
+    ReadWeight(strm, &final_weight_);
+  }
+  DecodeProcessedFst(int_code, fst, unweighted);
+  DecodeForCompress(fst, *encoder);
+  return !fst->Properties(kError, false);
+}
+
+template <class Arc>
+void Compressor<Arc>::WriteWeight(const std::vector<Weight> &input,
+                                  std::ostream &strm) {
+  int64 size = input.size();
+  WriteType(strm, size);
+  for (typename std::vector<Weight>::const_iterator it = input.begin();
+       it != input.end(); ++it) {
+    it->Write(strm);
+  }
+}
+
+template <class Arc>
+void Compressor<Arc>::WriteToStream(std::ostream &strm) {
+  while (buffer_code_.size() % 8 != 0) buffer_code_.push_back(1);
+  int64 data_size = buffer_code_.size() / 8;
+  WriteType(strm, data_size);
+  std::vector<bool>::const_iterator it;
+  int64 i;
+  uint8 block;
+  for (it = buffer_code_.begin(), i = 0; it != buffer_code_.end(); ++it, ++i) {
+    if (i % 8 == 0) {
+      if (i > 0) WriteType(strm, block);
+      block = 0;
+    } else {
+      block = block << 1;
+    }
+    block |= *it;
+  }
+  WriteType(strm, block);
+}
+
+template <class Arc>
+bool Compressor<Arc>::Compress(const Fst<Arc> &fst, std::ostream &strm) {
+  VectorFst<Arc> processedfst;
+  EncodeMapper<Arc> encoder(kEncodeLabels, ENCODE);
+  Preprocess(fst, &processedfst, &encoder);
+  WriteType(strm, kCompressMagicNumber);
+  encoder.Write(strm, "encoder stream");
+  EncodeProcessedFst(processedfst, strm);
+  return true;
+}
+
+// Convenience functions that call the compressor and decompressor.
+
+template <class Arc>
+void Compress(const Fst<Arc> &fst, std::ostream &strm) {
+  Compressor<Arc> comp;
+  comp.Compress(fst, strm);
+}
+
+// Returns true on success.
+template <class Arc>
+bool Compress(const Fst<Arc> &fst, const string &file_name,
+              const bool gzip = false) {
+  if (gzip) {
+    if (file_name.empty()) {
+      std::stringstream strm;
+      Compress(fst, strm);
+      OGzFile gzfile(fileno(stdout));
+      gzfile.write(strm);
+      if (!gzfile) {
+        LOG(ERROR) << "Compress: Can't write to file: stdout";
+        return false;
+      }
+    } else {
+      std::stringstream strm;
+      Compress(fst, strm);
+      OGzFile gzfile(file_name);
+      if (!gzfile) {
+        LOG(ERROR) << "Compress: Can't open file: " << file_name;
+        return false;
+      }
+      gzfile.write(strm);
+      if (!gzfile) {
+        LOG(ERROR) << "Compress: Can't write to file: " << file_name;
+        return false;
+      }
+    }
+  } else if (file_name.empty()) {
+    Compress(fst, std::cout);
+  } else {
+    std::ofstream strm(file_name,
+                             std::ios_base::out | std::ios_base::binary);
+    if (!strm) {
+      LOG(ERROR) << "Compress: Can't open file: " << file_name;
+      return false;
+    }
+    Compress(fst, strm);
+  }
+  return true;
+}
+
+template <class Arc>
+void Decompress(std::istream &strm, const string &source,
+                MutableFst<Arc> *fst) {
+  Compressor<Arc> comp;
+  comp.Decompress(strm, source, fst);
+}
+
+// Returns true on success.
+template <class Arc>
+bool Decompress(const string &file_name, MutableFst<Arc> *fst,
+                const bool gzip = false) {
+  if (gzip) {
+    if (file_name.empty()) {
+      IGzFile gzfile(fileno(stdin));
+      Decompress(*gzfile.read(), "stdin", fst);
+      if (!gzfile) {
+        LOG(ERROR) << "Decompress: Can't read from file: stdin";
+        return false;
+      }
+    } else {
+      IGzFile gzfile(file_name);
+      if (!gzfile) {
+        LOG(ERROR) << "Decompress: Can't open file: " << file_name;
+        return false;
+      }
+      Decompress(*gzfile.read(), file_name, fst);
+      if (!gzfile) {
+        LOG(ERROR) << "Decompress: Can't read from file: " << file_name;
+        return false;
+      }
+    }
+  } else if (file_name.empty()) {
+    Decompress(std::cin, "stdin", fst);
+  } else {
+    std::ifstream strm(file_name,
+                            std::ios_base::in | std::ios_base::binary);
+    if (!strm) {
+      LOG(ERROR) << "Decompress: Can't open file: " << file_name;
+      return false;
+    }
+    Decompress(strm, file_name, fst);
+  }
+  return true;
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_COMPRESS_COMPRESS_H_
diff --git a/projects/llm_framework/include/fst/extensions/compress/elias.h b/projects/llm_framework/include/fst/extensions/compress/elias.h
new file mode 100644
index 00000000..6c576d33
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/compress/elias.h
@@ -0,0 +1,97 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Compresses and decompresses unweighted FSTs.
+
+#ifndef FST_EXTENSIONS_COMPRESS_ELIAS_H_
+#define FST_EXTENSIONS_COMPRESS_ELIAS_H_
+
+#include <stack>
+#include <vector>
+
+#include <fst/compat.h>
+namespace fst {
+
+template <class Var>
+class Elias {
+ public:
+  // Gamma encoding is a subroutine for Delta encoding
+  static void GammaEncode(const Var &input, std::vector<bool> *code);
+
+  // Elias Delta encoding for a single integer
+  static void DeltaEncode(const Var &input, std::vector<bool> *code);
+
+  // Batch decoding of a set of integers
+  static void BatchDecode(const std::vector<bool> &input,
+                          std::vector<Var> *output);
+};
+
+template <class Var>
+void Elias<Var>::GammaEncode(const Var &input, std::vector<bool> *code) {
+  Var input_copy = input;
+  std::stack<bool> reverse_code;
+  while (input_copy > 0) {
+    reverse_code.push(input_copy % 2);
+    input_copy = input_copy / 2;
+  }
+  for (Var auxvar = 0; auxvar < reverse_code.size() - 1; auxvar++)
+    code->push_back(0);
+  while (reverse_code.empty() != 1) {
+    code->push_back(reverse_code.top());
+    reverse_code.pop();
+  }
+}
+template <class Var>
+void Elias<Var>::DeltaEncode(const Var &input, std::vector<bool> *code) {
+  Var input_copy = input + 1;
+  std::stack<bool> reverse_remainder;
+  Var auxvar = 0;
+  while (input_copy != 0) {
+    reverse_remainder.push(input_copy % 2);
+    input_copy = input_copy / 2;
+    auxvar = auxvar + 1;
+  }
+  GammaEncode(auxvar, code);
+  reverse_remainder.pop();
+  while (reverse_remainder.empty() != 1) {
+    code->push_back(reverse_remainder.top());
+    reverse_remainder.pop();
+  }
+}
+
+template <class Var>
+void Elias<Var>::BatchDecode(const std::vector<bool> &input,
+                             std::vector<Var> *output) {
+  Var lead_zeros = 0;
+  Var remainder_bits = 0;
+  Var current_word = 1;
+  Var value = 1;
+  std::vector<bool>::const_iterator it = input.begin();
+  while (it != input.end()) {
+    lead_zeros = 0;
+    remainder_bits = 0;
+    current_word = 1;
+    value = 1;
+    while (*it != 1) {
+      it++;
+      lead_zeros++;
+    }
+    it++;
+    while (lead_zeros > 0) {
+      lead_zeros--;
+      current_word = 2 * current_word + *it;
+      it++;
+    }
+    current_word--;
+    while (current_word > 0) {
+      value = 2 * value + *it;
+      current_word--;
+      it++;
+    }
+    output->push_back(value - 1);
+  }
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_COMPRESS_ELIAS_H_
diff --git a/projects/llm_framework/include/fst/extensions/compress/gzfile.h b/projects/llm_framework/include/fst/extensions/compress/gzfile.h
new file mode 100644
index 00000000..2d2e6243
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/compress/gzfile.h
@@ -0,0 +1,127 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Resource handles for gzip files written to or read from stringstreams. These
+// are necessary to provide the compression routines with streams reading from
+// or writing to compressed files (or the UNIX standard streams), and are not
+// intended for general use.
+
+#ifndef FST_EXTENSIONS_COMPRESS_GZFILE_H_
+#define FST_EXTENSIONS_COMPRESS_GZFILE_H_
+
+#include <cstddef>
+#include <memory>
+#include <sstream>
+#include <string>
+
+#include <fst/compat.h>
+#include <fst/log.h>
+#include <fst/fst.h>
+#include <zlib.h>
+
+using std::stringstream;
+using std::unique_ptr;
+
+namespace fst {
+
+// Gives the zlib gzFile type an OO-like interface. String inputs are all
+// C-style strings. The caller is responsible to get the file modes appropriate
+// for the IO methods being called, and for opening the file descriptors
+// if that constructor is used. The ! operator can be used to check for errors
+// after construction or read/writing.
+class GzFile {
+ public:
+  GzFile(const char *filename, const char *mode)
+      : gzfile_(gzopen(filename, mode)), error_(check_handle()) {
+  }
+
+  // The caller is responsible to ensure the corresponding FD is open and has
+  // the needed modes ("r" for reading, "w" or "a" for writing).
+  explicit GzFile(const int fd, const char *mode)
+      : gzfile_(gzdopen(fd, mode)), error_(check_handle()), close_me_(false) {
+  }
+
+  // If the instance was constructed from an FD, flush the buffer; otherwise,
+  // close the file, which flushes the buffer as a side-effect.
+  ~GzFile() { close_me_ ? gzclose(gzfile_) : gzflush(gzfile_, Z_FINISH); }
+
+  inline bool operator!() const { return error_; }
+
+  // Returns false on EOF and sets error if short read does not reach an EOF.
+  int read(void *buf, unsigned int size) {
+    int bytes_read = gzread(gzfile_, buf, size);
+    if ((bytes_read < size) && !gzeof(gzfile_)) error_ = true;
+    return bytes_read;
+  }
+
+  // Sets error on short writes.
+  void write(const char *buf, unsigned int size) {
+    if (gzwrite(gzfile_, buf, size) != size) error_ = true;
+  }
+
+ private:
+  // gzopen and gzdopen signal failure by returning null.
+  bool check_handle() { return gzfile_ == nullptr; }
+
+  gzFile gzfile_ = nullptr;
+  bool error_ = false;
+  bool close_me_ = false;
+};
+
+// Resource handle for writing stringstream to GzFile.
+class OGzFile {
+ public:
+  explicit OGzFile(const string &filename) : OGzFile(filename.c_str()) {}
+
+  explicit OGzFile(const char *filename) : gz_(GzFile(filename, mode_)) {}
+
+  explicit OGzFile(const int fd) : gz_(GzFile(fd, mode_)) {}
+
+  inline bool operator!() const { return !gz_; }
+
+  void write(const std::stringstream &ssbuf) {
+    string sbuf = ssbuf.str();
+    gz_.write(sbuf.data(), sbuf.size());
+  }
+
+ private:
+  GzFile gz_;
+  static constexpr auto &mode_ = "wb";
+};
+
+// Resource handle for reading stringstream from GzFile.
+class IGzFile {
+ public:
+  explicit IGzFile(const string &filename) : IGzFile(filename.c_str()) {}
+
+  explicit IGzFile(const char *filename) : gz_(GzFile(filename, mode_)) {}
+
+  explicit IGzFile(const int fd) : gz_(GzFile(fd, mode_)) {}
+
+  inline bool operator!() const { return !gz_; }
+
+  // This is a great case for "move", but GCC 4 is missing the C+11 standard
+  // move constructor for stringstream, so a unique_ptr is the next best thing.
+  std::unique_ptr<std::stringstream> read() {
+    char buf[bufsize_];
+    std::unique_ptr<std::stringstream> sstrm(new std::stringstream);
+    // We always read at least once, and the result of the last read is always
+    // pushed onto the stringstream. We use the "write" member because << onto
+    // a stringstream stops at the null byte, which might be data!
+    int bytes_read;
+    while ((bytes_read = gz_.read(buf, bufsize_)) == bufsize_)
+      sstrm->write(buf, bufsize_);
+    sstrm->write(buf, bytes_read);
+    return sstrm;
+  }
+
+ private:
+  GzFile gz_;
+  static constexpr auto &mode_ = "rb";
+  // This is the same size as the default internal buffer for zlib.
+  static const size_t bufsize_ = 8192;
+};
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_COMPRESS_GZFILE_H_
diff --git a/projects/llm_framework/include/fst/extensions/compress/randmod.h b/projects/llm_framework/include/fst/extensions/compress/randmod.h
new file mode 100644
index 00000000..bfe2de5b
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/compress/randmod.h
@@ -0,0 +1,102 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Generates a random FST according to a class-specific transition model.
+
+#ifndef FST_EXTENSIONS_COMPRESS_RANDMOD_H_
+#define FST_EXTENSIONS_COMPRESS_RANDMOD_H_
+
+#include <cstdlib>
+#include <vector>
+
+#include <fst/compat.h>
+#include <fst/mutable-fst.h>
+
+namespace fst {
+
+template <class Arc, class G>
+class RandMod {
+ public:
+  typedef typename Arc::StateId StateId;
+  typedef typename Arc::Label Label;
+  typedef typename Arc::Weight Weight;
+
+  // Generates random FST with 'nstates' with 'nclasses' in the probability
+  // generation model, and 'nlabels' in the alphabet. If 'trans' = true, then
+  // a transducer is generated; iff 'generate_' is non-null, the output is
+  // randomly weighted.
+  RandMod(StateId nstates, StateId nclasses, Label nlabels, bool trans,
+          const G *generate)
+      : nstates_(nstates),
+        nclasses_(nclasses),
+        nlabels_(nlabels),
+        trans_(trans),
+        generate_(generate) {
+    for (StateId s = 0; s < nstates; ++s) {
+      classes_.push_back(rand() % nclasses);  // NOLINT
+    }
+  }
+
+  // Generates a random FST according to a class-specific transition model
+  void Generate(StdMutableFst *fst) {
+    StateId start = rand() % nstates_;  // NOLINT
+    fst->DeleteStates();
+    for (StateId s = 0; s < nstates_; ++s) {
+      fst->AddState();
+      if (s == start) fst->SetStart(start);
+      for (StateId n = 0; n <= nstates_; ++n) {
+        Arc arc;
+        StateId d = n == nstates_ ? kNoStateId : n;
+        if (!RandArc(s, d, &arc)) continue;
+        if (d == kNoStateId) {  // A super-final transition?
+          fst->SetFinal(s, arc.weight);
+        } else {
+          fst->AddArc(s, arc);
+        }
+      }
+    }
+  }
+
+ private:
+  // Generates a transition from s to d. If d == kNoStateId, a superfinal
+  // transition is generated. Returns false if no transition generated.
+  bool RandArc(StateId s, StateId d, Arc *arc) {
+    StateId sclass = classes_[s];
+    StateId dclass = d != kNoStateId ? classes_[d] : 0;
+
+    int r = sclass + dclass + 2;
+    if ((rand() % r) != 0)  // NOLINT
+      return false;
+
+    arc->nextstate = d;
+
+    Label ilabel = kNoLabel;
+    Label olabel = kNoLabel;
+    if (d != kNoStateId) {
+      ilabel = (dclass % nlabels_) + 1;
+      if (trans_)
+        olabel = (sclass % nlabels_) + 1;
+      else
+        olabel = ilabel;
+    }
+
+    Weight weight = Weight::One();
+    if (generate_) weight = (*generate_)();
+
+    arc->ilabel = ilabel;
+    arc->olabel = olabel;
+    arc->weight = weight;
+    return true;
+  }
+
+  StateId nstates_;
+  StateId nclasses_;
+  Label nlabels_;
+  bool trans_;
+  const G *generate_;
+  std::vector<StateId> classes_;
+};
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_COMPRESS_RANDMOD_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/compile-strings.h b/projects/llm_framework/include/fst/extensions/far/compile-strings.h
new file mode 100644
index 00000000..96271c17
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/far/compile-strings.h
@@ -0,0 +1,260 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_EXTENSIONS_FAR_COMPILE_STRINGS_H_
+#define FST_EXTENSIONS_FAR_COMPILE_STRINGS_H_
+
+#ifndef _MSC_VER
+#include <libgen.h>
+#else
+#include <fst/compat.h>
+#endif
+
+#include <fstream>
+#include <istream>
+#include <string>
+#include <vector>
+
+#include <fst/extensions/far/far.h>
+#include <fstream>
+#include <fst/string.h>
+
+namespace fst {
+
+// Constructs a reader that provides FSTs from a file (stream) either on a
+// line-by-line basis or on a per-stream basis. Note that the freshly
+// constructed reader is already set to the first input.
+//
+// Sample usage:
+//
+//   for (StringReader<Arc> reader(...); !reader.Done(); reader.Next()) {
+//     auto *fst = reader.GetVectorFst();
+//   }
+template <class Arc>
+class StringReader {
+ public:
+  using Label = typename Arc::Label;
+  using Weight = typename Arc::Weight;
+
+  enum EntryType { LINE = 1, FILE = 2 };
+
+  StringReader(std::istream &istrm, const string &source, EntryType entry_type,
+               StringTokenType token_type, bool allow_negative_labels,
+               const SymbolTable *syms = nullptr,
+               Label unknown_label = kNoStateId)
+      : nline_(0),
+        istrm_(istrm),
+        source_(source),
+        entry_type_(entry_type),
+        token_type_(token_type),
+        symbols_(syms),
+        done_(false),
+        compiler_(token_type, syms, unknown_label, allow_negative_labels) {
+    Next();  // Initialize the reader to the first input.
+  }
+
+  bool Done() { return done_; }
+
+  void Next() {
+    VLOG(1) << "Processing source " << source_ << " at line " << nline_;
+    if (!istrm_) {  // We're done if we have no more input.
+      done_ = true;
+      return;
+    }
+    if (entry_type_ == LINE) {
+      getline(istrm_, content_);
+      ++nline_;
+    } else {
+      content_.clear();
+      string line;
+      while (getline(istrm_, line)) {
+        ++nline_;
+        content_.append(line);
+        content_.append("\n");
+      }
+    }
+    if (!istrm_ && content_.empty())  // We're also done if we read off all the
+      done_ = true;                   // whitespace at the end of a file.
+  }
+
+  VectorFst<Arc> *GetVectorFst(bool keep_symbols = false) {
+    std::unique_ptr<VectorFst<Arc>> fst(new VectorFst<Arc>());
+    if (keep_symbols) {
+      fst->SetInputSymbols(symbols_);
+      fst->SetOutputSymbols(symbols_);
+    }
+    if (compiler_(content_, fst.get())) {
+      return fst.release();
+    } else {
+      return nullptr;
+    }
+  }
+
+  CompactStringFst<Arc> *GetCompactFst(bool keep_symbols = false) {
+    std::unique_ptr<CompactStringFst<Arc>> fst;
+    if (keep_symbols) {
+      VectorFst<Arc> tmp;
+      tmp.SetInputSymbols(symbols_);
+      tmp.SetOutputSymbols(symbols_);
+      fst.reset(new CompactStringFst<Arc>(tmp));
+    } else {
+      fst.reset(new CompactStringFst<Arc>());
+    }
+    if (compiler_(content_, fst.get())) {
+      return fst.release();
+    } else {
+      return nullptr;
+    }
+  }
+
+ private:
+  size_t nline_;
+  std::istream &istrm_;
+  string source_;
+  EntryType entry_type_;
+  StringTokenType token_type_;
+  const SymbolTable *symbols_;
+  bool done_;
+  StringCompiler<Arc> compiler_;
+  string content_;  // The actual content of the input stream's next FST.
+
+  StringReader(const StringReader &) = delete;
+  StringReader &operator=(const StringReader &) = delete;
+};
+
+// Computes the minimal length required to encode each line number as a decimal
+// number.
+int KeySize(const char *filename);
+
+template <class Arc>
+void FarCompileStrings(const std::vector<string> &in_fnames,
+                       const string &out_fname, const string &fst_type,
+                       const FarType &far_type, int32 generate_keys,
+                       FarEntryType fet, FarTokenType tt,
+                       const string &symbols_fname,
+                       const string &unknown_symbol, bool keep_symbols,
+                       bool initial_symbols, bool allow_negative_labels,
+                       const string &key_prefix, const string &key_suffix) {
+  typename StringReader<Arc>::EntryType entry_type;
+  if (fet == FET_LINE) {
+    entry_type = StringReader<Arc>::LINE;
+  } else if (fet == FET_FILE) {
+    entry_type = StringReader<Arc>::FILE;
+  } else {
+    FSTERROR() << "FarCompileStrings: Unknown entry type";
+    return;
+  }
+  StringTokenType token_type;
+  if (tt == FTT_SYMBOL) {
+    token_type = StringTokenType::SYMBOL;
+  } else if (tt == FTT_BYTE) {
+    token_type = StringTokenType::BYTE;
+  } else if (tt == FTT_UTF8) {
+    token_type = StringTokenType::UTF8;
+  } else {
+    FSTERROR() << "FarCompileStrings: Unknown token type";
+    return;
+  }
+  bool compact;
+  if (fst_type.empty() || (fst_type == "vector")) {
+    compact = false;
+  } else if (fst_type == "compact") {
+    compact = true;
+  } else {
+    FSTERROR() << "FarCompileStrings: Unknown FST type: " << fst_type;
+    return;
+  }
+  std::unique_ptr<const SymbolTable> syms;
+  typename Arc::Label unknown_label = kNoLabel;
+  if (!symbols_fname.empty()) {
+    const SymbolTableTextOptions opts(allow_negative_labels);
+    syms.reset(SymbolTable::ReadText(symbols_fname, opts));
+    if (!syms) {
+      LOG(ERROR) << "FarCompileStrings: Error reading symbol table: "
+                 << symbols_fname;
+      return;
+    }
+    if (!unknown_symbol.empty()) {
+      unknown_label = syms->Find(unknown_symbol);
+      if (unknown_label == kNoLabel) {
+        FSTERROR() << "FarCompileStrings: Label \"" << unknown_label
+                   << "\" missing from symbol table: " << symbols_fname;
+        return;
+      }
+    }
+  }
+  std::unique_ptr<FarWriter<Arc>> far_writer(
+      FarWriter<Arc>::Create(out_fname, far_type));
+  if (!far_writer) return;
+  int n = 0;
+  for (const auto &in_fname : in_fnames) {
+    if (generate_keys == 0 && in_fname.empty()) {
+      FSTERROR() << "FarCompileStrings: Read from a file instead of stdin or"
+                 << " set the --generate_keys flags.";
+      return;
+    }
+    int key_size =
+        generate_keys ? generate_keys : (entry_type == StringReader<Arc>::FILE
+                                             ? 1 : KeySize(in_fname.c_str()));
+    std::ifstream fstrm;
+    if (!in_fname.empty()) {
+      fstrm.open(in_fname);
+      if (!fstrm) {
+        FSTERROR() << "FarCompileStrings: Can't open file: " << in_fname;
+        return;
+      }
+    }
+    std::istream &istrm = fstrm.is_open() ? fstrm : std::cin;
+    bool keep_syms = keep_symbols;
+    for (StringReader<Arc> reader(
+             istrm, in_fname.empty() ? "stdin" : in_fname, entry_type,
+             token_type, allow_negative_labels, syms.get(), unknown_label);
+         !reader.Done(); reader.Next()) {
+      ++n;
+      std::unique_ptr<const Fst<Arc>> fst;
+      if (compact) {
+        fst.reset(reader.GetCompactFst(keep_syms));
+      } else {
+        fst.reset(reader.GetVectorFst(keep_syms));
+      }
+      if (initial_symbols) keep_syms = false;
+      if (!fst) {
+        FSTERROR() << "FarCompileStrings: Compiling string number " << n
+                   << " in file " << in_fname << " failed with token_type = "
+                   << (tt == FTT_BYTE
+                           ? "byte"
+                           : (tt == FTT_UTF8
+                                  ? "utf8"
+                                  : (tt == FTT_SYMBOL ? "symbol" : "unknown")))
+                   << " and entry_type = "
+                   << (fet == FET_LINE
+                           ? "line"
+                           : (fet == FET_FILE ? "file" : "unknown"));
+        return;
+      }
+      std::ostringstream keybuf;
+      keybuf.width(key_size);
+      keybuf.fill('0');
+      keybuf << n;
+      string key;
+      if (generate_keys > 0) {
+        key = keybuf.str();
+      } else {
+        auto *filename = new char[in_fname.size() + 1];
+        strcpy(filename, in_fname.c_str());
+        key = basename(filename);
+        if (entry_type != StringReader<Arc>::FILE) {
+          key += "-";
+          key += keybuf.str();
+        }
+        delete[] filename;
+      }
+      far_writer->Add(key_prefix + key + key_suffix, *fst);
+    }
+    if (generate_keys == 0) n = 0;
+  }
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_FAR_COMPILE_STRINGS_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/create.h b/projects/llm_framework/include/fst/extensions/far/create.h
new file mode 100644
index 00000000..31c95896
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/far/create.h
@@ -0,0 +1,46 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Creates a finite-state archive from component FSTs.
+
+#ifndef FST_EXTENSIONS_FAR_CREATE_H_
+#define FST_EXTENSIONS_FAR_CREATE_H_
+
+#include <sstream>
+#include <string>
+#include <vector>
+
+#include <fst/extensions/far/far.h>
+
+namespace fst {
+
+template <class Arc>
+void FarCreate(const std::vector<string> &in_fnames, const string &out_fname,
+               const int32 generate_keys, const FarType &far_type,
+               const string &key_prefix, const string &key_suffix) {
+  std::unique_ptr<FarWriter<Arc>> far_writer(
+      FarWriter<Arc>::Create(out_fname, far_type));
+  if (!far_writer) return;
+  for (size_t i = 0; i < in_fnames.size(); ++i) {
+    std::unique_ptr<Fst<Arc>> ifst(Fst<Arc>::Read(in_fnames[i]));
+    if (!ifst) return;
+    string key;
+    if (generate_keys > 0) {
+      std::ostringstream keybuf;
+      keybuf.width(generate_keys);
+      keybuf.fill('0');
+      keybuf << i + 1;
+      key = keybuf.str();
+    } else {
+      auto *filename = new char[in_fnames[i].size() + 1];
+      strcpy(filename, in_fnames[i].c_str());
+      key = basename(filename);
+      delete[] filename;
+    }
+    far_writer->Add(key_prefix + key + key_suffix, *ifst);
+  }
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_FAR_CREATE_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/equal.h b/projects/llm_framework/include/fst/extensions/far/equal.h
new file mode 100644
index 00000000..04511d21
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/far/equal.h
@@ -0,0 +1,69 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_EXTENSIONS_FAR_EQUAL_H_
+#define FST_EXTENSIONS_FAR_EQUAL_H_
+
+#include <memory>
+#include <string>
+
+#include <fst/extensions/far/far.h>
+#include <fst/equal.h>
+
+namespace fst {
+
+template <class Arc>
+bool FarEqual(const string &filename1, const string &filename2,
+              float delta = kDelta, const string &begin_key = string(),
+              const string &end_key = string()) {
+  std::unique_ptr<FarReader<Arc>> reader1(FarReader<Arc>::Open(filename1));
+  if (!reader1) {
+    LOG(ERROR) << "FarEqual: Could not open FAR file " << filename1;
+    return false;
+  }
+  std::unique_ptr<FarReader<Arc>> reader2(FarReader<Arc>::Open(filename2));
+  if (!reader2) {
+    LOG(ERROR) << "FarEqual: Could not open FAR file " << filename2;
+    return false;
+  }
+  if (!begin_key.empty()) {
+    bool find_begin1 = reader1->Find(begin_key);
+    bool find_begin2 = reader2->Find(begin_key);
+    if (!find_begin1 || !find_begin2) {
+      bool ret = !find_begin1 && !find_begin2;
+      if (!ret) {
+        LOG(ERROR) << "FarEqual: Key " << begin_key << " missing from "
+                   << (find_begin1 ? "second" : "first") << " archive";
+      }
+      return ret;
+    }
+  }
+  for (; !reader1->Done() && !reader2->Done();
+       reader1->Next(), reader2->Next()) {
+    const auto &key1 = reader1->GetKey();
+    const auto &key2 = reader2->GetKey();
+    if (!end_key.empty() && end_key < key1 && end_key < key2) {
+      return true;
+    }
+    if (key1 != key2) {
+      LOG(ERROR) << "FarEqual: Mismatched keys " << key1 << " and " << key2;
+      return false;
+    }
+    if (!Equal(*(reader1->GetFst()), *(reader2->GetFst()), delta)) {
+      LOG(ERROR) << "FarEqual: FSTs for key " << key1 << " are not equal";
+      return false;
+    }
+  }
+  if (!reader1->Done() || !reader2->Done()) {
+    LOG(ERROR) << "FarEqual: Key "
+               << (reader1->Done() ? reader2->GetKey() : reader1->GetKey())
+               << " missing from " << (reader2->Done() ? "first" : "second")
+               << " archive";
+    return false;
+  }
+  return true;
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_FAR_EQUAL_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/extract.h b/projects/llm_framework/include/fst/extensions/far/extract.h
new file mode 100644
index 00000000..e707835f
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/far/extract.h
@@ -0,0 +1,118 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Extracts component FSTs from an finite-state archive.
+
+#ifndef FST_EXTENSIONS_FAR_EXTRACT_H_
+#define FST_EXTENSIONS_FAR_EXTRACT_H_
+
+#include <memory>
+#include <string>
+#include <vector>
+
+#include <fst/extensions/far/far.h>
+#include <fst/util.h>
+
+namespace fst {
+
+template <class Arc>
+inline void FarWriteFst(const Fst<Arc> *fst, string key, string *okey,
+                        int *nrep, int32 generate_filenames, int i,
+                        const string &filename_prefix,
+                        const string &filename_suffix) {
+  if (key == *okey) {
+    ++*nrep;
+  } else {
+    *nrep = 0;
+  }
+  *okey = key;
+  string ofilename;
+  if (generate_filenames) {
+    std::ostringstream tmp;
+    tmp.width(generate_filenames);
+    tmp.fill('0');
+    tmp << i;
+    ofilename = tmp.str();
+  } else {
+    if (*nrep > 0) {
+      std::ostringstream tmp;
+      tmp << '.' << nrep;
+      key.append(tmp.str().data(), tmp.str().size());
+    }
+    ofilename = key;
+  }
+  fst->Write(filename_prefix + ofilename + filename_suffix);
+}
+
+template <class Arc>
+void FarExtract(const std::vector<string> &ifilenames, int32 generate_filenames,
+                const string &keys, const string &key_separator,
+                const string &range_delimiter, const string &filename_prefix,
+                const string &filename_suffix) {
+  std::unique_ptr<FarReader<Arc>> far_reader(
+      FarReader<Arc>::Open(ifilenames));
+  if (!far_reader) return;
+  string okey;
+  int nrep = 0;
+  std::vector<char *> key_vector;
+  // User has specified a set of FSTs to extract, where some of these may in
+  // fact be ranges.
+  if (!keys.empty()) {
+    auto *keys_cstr = new char[keys.size() + 1];
+    strcpy(keys_cstr, keys.c_str());
+    SplitString(keys_cstr, key_separator.c_str(), &key_vector, true);
+    int i = 0;
+    for (size_t k = 0; k < key_vector.size(); ++k, ++i) {
+      string key = key_vector[k];
+      auto *key_cstr = new char[key.size() + 1];
+      strcpy(key_cstr, key.c_str());
+      std::vector<char *> range_vector;
+      SplitString(key_cstr, range_delimiter.c_str(), &range_vector, false);
+      if (range_vector.size() == 1) {  // Not a range
+        if (!far_reader->Find(key)) {
+          LOG(ERROR) << "FarExtract: Cannot find key " << key;
+          return;
+        }
+        const auto *fst = far_reader->GetFst();
+        FarWriteFst(fst, key, &okey, &nrep, generate_filenames, i,
+                    filename_prefix, filename_suffix);
+      } else if (range_vector.size() == 2) {  // A legal range
+        string begin_key = range_vector[0];
+        string end_key = range_vector[1];
+        if (begin_key.empty() || end_key.empty()) {
+          LOG(ERROR) << "FarExtract: Illegal range specification " << key;
+          return;
+        }
+        if (!far_reader->Find(begin_key)) {
+          LOG(ERROR) << "FarExtract: Cannot find key " << begin_key;
+          return;
+        }
+        for (; !far_reader->Done(); far_reader->Next(), ++i) {
+          const auto &ikey = far_reader->GetKey();
+          if (end_key < ikey) break;
+          const auto *fst = far_reader->GetFst();
+          FarWriteFst(fst, ikey, &okey, &nrep, generate_filenames, i,
+                      filename_prefix, filename_suffix);
+        }
+      } else {
+        LOG(ERROR) << "FarExtract: Illegal range specification " << key;
+        return;
+      }
+      delete[] key_cstr;
+    }
+    delete[] keys_cstr;
+    return;
+  }
+  // Nothing specified, so just extracts everything.
+  for (size_t i = 1; !far_reader->Done(); far_reader->Next(), ++i) {
+    const auto &key = far_reader->GetKey();
+    const auto *fst = far_reader->GetFst();
+    FarWriteFst(fst, key, &okey, &nrep, generate_filenames, i, filename_prefix,
+                filename_suffix);
+  }
+  return;
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_FAR_EXTRACT_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/far-class.h b/projects/llm_framework/include/fst/extensions/far/far-class.h
new file mode 100644
index 00000000..50d7be9b
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/far/far-class.h
@@ -0,0 +1,258 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Scripting API support for FarReader and FarWriter.
+
+#ifndef FST_EXTENSIONS_FAR_FAR_CLASS_H_
+#define FST_EXTENSIONS_FAR_FAR_CLASS_H_
+
+#include <memory>
+#include <string>
+#include <vector>
+
+#include <fst/extensions/far/far.h>
+#include <fst/script/arg-packs.h>
+#include <fst/script/fstscript.h>
+
+namespace fst {
+namespace script {
+
+
+// FarReader API.
+
+// Virtual interface implemented by each concrete FarReaderImpl<A>.
+// See the FarReader interface in far.h for the exact semantics.
+class FarReaderImplBase {
+ public:
+  virtual const string &ArcType() const = 0;
+  virtual bool Done() const = 0;
+  virtual bool Error() const = 0;
+  virtual const string &GetKey() const = 0;
+  virtual const FstClass *GetFstClass() const = 0;
+  virtual bool Find(const string &key) = 0;
+  virtual void Next() = 0;
+  virtual void Reset() = 0;
+  virtual FarType Type() const = 0;
+  virtual ~FarReaderImplBase() {}
+};
+
+// Templated implementation.
+template <class Arc>
+class FarReaderClassImpl : public FarReaderImplBase {
+ public:
+  explicit FarReaderClassImpl(const string &filename)
+      : impl_(FarReader<Arc>::Open(filename)) {}
+
+  explicit FarReaderClassImpl(const std::vector<string> &filenames)
+      : impl_(FarReader<Arc>::Open(filenames)) {}
+
+  const string &ArcType() const final { return Arc::Type(); }
+
+  bool Done() const final { return impl_->Done(); }
+
+  bool Error() const final { return impl_->Error(); }
+
+  bool Find(const string &key) final { return impl_->Find(key); }
+
+  const FstClass *GetFstClass() const final {
+    fstc_.reset(new FstClass(*impl_->GetFst()));
+    return fstc_.get();
+  }
+
+  const string &GetKey() const final { return impl_->GetKey(); }
+
+  void Next() final { return impl_->Next(); }
+
+  void Reset() final { impl_->Reset(); }
+
+  FarType Type() const final { return impl_->Type(); }
+
+  const FarReader<Arc> *GetImpl() const { return impl_.get(); }
+
+  FarReader<Arc> *GetImpl() { return impl_.get(); }
+
+ private:
+  std::unique_ptr<FarReader<Arc>> impl_;
+  mutable std::unique_ptr<FstClass> fstc_;
+};
+
+
+class FarReaderClass;
+
+using OpenFarReaderClassArgs =
+    WithReturnValue<FarReaderClass *, const std::vector<string> &>;
+
+// Untemplated user-facing class holding a templated pimpl.
+class FarReaderClass {
+ public:
+  const string &ArcType() const { return impl_->ArcType(); }
+
+  bool Done() const { return impl_->Done(); }
+
+  // Returns True if the impl is null (i.e., due to read failure).
+  // Attempting to call any other function will result in null dereference.
+  bool Error() const { return (impl_) ? impl_->Error() : true; }
+
+  bool Find(const string &key) { return impl_->Find(key); }
+
+  const FstClass *GetFstClass() const { return impl_->GetFstClass(); }
+
+  const string &GetKey() const { return impl_->GetKey(); }
+
+  void Next() { impl_->Next(); }
+
+  void Reset() { impl_->Reset(); }
+
+  FarType Type() const { return impl_->Type(); }
+
+  template <class Arc>
+  const FarReader<Arc> *GetFarReader() const {
+    if (Arc::Type() != ArcType()) return nullptr;
+    const FarReaderClassImpl<Arc> *typed_impl =
+        static_cast<FarReaderClassImpl<Arc> *>(impl_.get());
+    return typed_impl->GetImpl();
+  }
+
+  template <class Arc>
+  FarReader<Arc> *GetFarReader() {
+    if (Arc::Type() != ArcType()) return nullptr;
+    FarReaderClassImpl<Arc> *typed_impl =
+        static_cast<FarReaderClassImpl<Arc> *>(impl_.get());
+    return typed_impl->GetImpl();
+  }
+
+  template <class Arc>
+  friend void OpenFarReaderClass(OpenFarReaderClassArgs *args);
+
+  // Defined in the CC.
+
+  static FarReaderClass *Open(const string &filename);
+
+  static FarReaderClass *Open(const std::vector<string> &filenames);
+
+ private:
+  template <class Arc>
+  explicit FarReaderClass(FarReaderClassImpl<Arc> *impl) : impl_(impl) {}
+
+  std::unique_ptr<FarReaderImplBase> impl_;
+};
+
+// These exist solely for registration purposes; users should call the
+// static method FarReaderClass::Open instead.
+
+template <class Arc>
+void OpenFarReaderClass(OpenFarReaderClassArgs *args) {
+  args->retval = new FarReaderClass(new FarReaderClassImpl<Arc>(args->args));
+}
+
+// FarWriter API.
+
+// Virtual interface implemented by each concrete FarWriterImpl<A>.
+class FarWriterImplBase {
+ public:
+  // Unlike the lower-level library, this returns a boolean to signal failure
+  // due to non-conformant arc types.
+  virtual bool Add(const string &key, const FstClass &fst) = 0;
+  virtual const string &ArcType() const = 0;
+  virtual bool Error() const = 0;
+  virtual FarType Type() const = 0;
+  virtual ~FarWriterImplBase() {}
+};
+
+
+// Templated implementation.
+template <class Arc>
+class FarWriterClassImpl : public FarWriterImplBase {
+ public:
+  explicit FarWriterClassImpl(const string &filename,
+                              FarType type = FAR_DEFAULT)
+      : impl_(FarWriter<Arc>::Create(filename, type)) {}
+
+  bool Add(const string &key, const FstClass &fst) final {
+    if (ArcType() != fst.ArcType()) {
+      FSTERROR() << "Cannot write FST with " << fst.ArcType() << " arcs to "
+                 << "FAR with " << ArcType() << " arcs";
+      return false;
+    }
+    impl_->Add(key, *(fst.GetFst<Arc>()));
+    return true;
+  }
+
+  const string &ArcType() const final { return Arc::Type(); }
+
+  bool Error() const final { return impl_->Error(); }
+
+  FarType Type() const final { return impl_->Type(); }
+
+  const FarWriter<Arc> *GetImpl() const { return impl_.get(); }
+
+  FarWriter<Arc> *GetImpl() { return impl_.get(); }
+
+ private:
+  std::unique_ptr<FarWriter<Arc>> impl_;
+};
+
+
+class FarWriterClass;
+
+using CreateFarWriterClassInnerArgs = std::pair<const string &, FarType>;
+
+using CreateFarWriterClassArgs =
+    WithReturnValue<FarWriterClass *, CreateFarWriterClassInnerArgs>;
+
+// Untemplated user-facing class holding a templated pimpl.
+class FarWriterClass {
+ public:
+  static FarWriterClass *Create(const string &filename, const string &arc_type,
+                                FarType type = FAR_DEFAULT);
+
+  bool Add(const string &key, const FstClass &fst) {
+    return impl_->Add(key, fst);
+  }
+
+  // Returns True if the impl is null (i.e., due to construction failure).
+  // Attempting to call any other function will result in null dereference.
+  bool Error() const { return (impl_) ? impl_->Error() : true; }
+
+  const string &ArcType() const { return impl_->ArcType(); }
+
+  FarType Type() const { return impl_->Type(); }
+
+  template <class Arc>
+  const FarWriter<Arc> *GetFarWriter() const {
+    if (Arc::Type() != ArcType()) return nullptr;
+    const FarWriterClassImpl<Arc> *typed_impl =
+        static_cast<FarWriterClassImpl<Arc> *>(impl_.get());
+    return typed_impl->GetImpl();
+  }
+
+  template <class Arc>
+  FarWriter<Arc> *GetFarWriter() {
+    if (Arc::Type() != ArcType()) return nullptr;
+    FarWriterClassImpl<Arc> *typed_impl =
+        static_cast<FarWriterClassImpl<Arc> *>(impl_.get());
+    return typed_impl->GetImpl();
+  }
+
+  template <class Arc>
+  friend void CreateFarWriterClass(CreateFarWriterClassArgs *args);
+
+ private:
+  template <class Arc>
+  explicit FarWriterClass(FarWriterClassImpl<Arc> *impl) : impl_(impl) {}
+
+  std::unique_ptr<FarWriterImplBase> impl_;
+};
+
+// This exists solely for registration purposes; users should call the
+// static method FarWriterClass::Create instead.
+template <class Arc>
+void CreateFarWriterClass(CreateFarWriterClassArgs *args) {
+  args->retval = new FarWriterClass(new FarWriterClassImpl<Arc>(
+      std::get<0>(args->args), std::get<1>(args->args)));
+}
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_FAR_FAR_CLASS_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/far.h b/projects/llm_framework/include/fst/extensions/far/far.h
new file mode 100644
index 00000000..c24c7dab
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/far/far.h
@@ -0,0 +1,481 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Finite-State Transducer (FST) archive classes.
+
+#ifndef FST_EXTENSIONS_FAR_FAR_H_
+#define FST_EXTENSIONS_FAR_FAR_H_
+
+#include <iostream>
+#include <sstream>
+
+#include <fst/log.h>
+#include <fst/extensions/far/stlist.h>
+#include <fst/extensions/far/sttable.h>
+#include <fst/fst.h>
+#include <fst/vector-fst.h>
+#include <fstream>
+
+namespace fst {
+
+enum FarEntryType { FET_LINE, FET_FILE };
+
+enum FarTokenType { FTT_SYMBOL, FTT_BYTE, FTT_UTF8 };
+
+inline bool IsFst(const string &filename) {
+  std::ifstream strm(filename, std::ios_base::in | std::ios_base::binary);
+  if (!strm) return false;
+  return IsFstHeader(strm, filename);
+}
+
+// FST archive header class
+class FarHeader {
+ public:
+  const string &ArcType() const { return arctype_; }
+
+  const string &FarType() const { return fartype_; }
+
+  bool Read(const string &filename) {
+    FstHeader fsthdr;
+    if (filename.empty()) {
+      // Header reading unsupported on stdin. Assumes STList and StdArc.
+      fartype_ = "stlist";
+      arctype_ = "standard";
+      return true;
+    } else if (IsSTTable(filename)) {  // Checks if STTable.
+      ReadSTTableHeader(filename, &fsthdr);
+      fartype_ = "sttable";
+      arctype_ = fsthdr.ArcType().empty() ? "unknown" : fsthdr.ArcType();
+      return true;
+    } else if (IsSTList(filename)) {  // Checks if STList.
+      ReadSTListHeader(filename, &fsthdr);
+      fartype_ = "stlist";
+      arctype_ = fsthdr.ArcType().empty() ? "unknown" : fsthdr.ArcType();
+      return true;
+    } else if (IsFst(filename)) {  // Checks if FST.
+      std::ifstream istrm(filename,
+                               std::ios_base::in | std::ios_base::binary);
+      fsthdr.Read(istrm, filename);
+      fartype_ = "fst";
+      arctype_ = fsthdr.ArcType().empty() ? "unknown" : fsthdr.ArcType();
+      return true;
+    }
+    return false;
+  }
+
+ private:
+  string fartype_;
+  string arctype_;
+};
+
+enum FarType {
+  FAR_DEFAULT = 0,
+  FAR_STTABLE = 1,
+  FAR_STLIST = 2,
+  FAR_FST = 3,
+};
+
+// This class creates an archive of FSTs.
+template <class A>
+class FarWriter {
+ public:
+  using Arc = A;
+
+  // Creates a new (empty) FST archive; returns null on error.
+  static FarWriter *Create(const string &filename, FarType type = FAR_DEFAULT);
+
+  // Adds an FST to the end of an archive. Keys must be non-empty and
+  // in lexicographic order. FSTs must have a suitable write method.
+  virtual void Add(const string &key, const Fst<Arc> &fst) = 0;
+
+  virtual FarType Type() const = 0;
+
+  virtual bool Error() const = 0;
+
+  virtual ~FarWriter() {}
+
+ protected:
+  FarWriter() {}
+};
+
+// This class iterates through an existing archive of FSTs.
+template <class A>
+class FarReader {
+ public:
+  using Arc = A;
+
+  // Opens an existing FST archive in a single file; returns null on error.
+  // Sets current position to the beginning of the achive.
+  static FarReader *Open(const string &filename);
+
+  // Opens an existing FST archive in multiple files; returns null on error.
+  // Sets current position to the beginning of the achive.
+  static FarReader *Open(const std::vector<string> &filenames);
+
+  // Resets current position to beginning of archive.
+  virtual void Reset() = 0;
+
+  // Sets current position to first entry >= key.  Returns true if a match.
+  virtual bool Find(const string &key) = 0;
+
+  // Current position at end of archive?
+  virtual bool Done() const = 0;
+
+  // Move current position to next FST.
+  virtual void Next() = 0;
+
+  // Returns key at the current position. This reference is invalidated if
+  // the current position in the archive is changed.
+  virtual const string &GetKey() const = 0;
+
+  // Returns pointer to FST at the current position. This is invalidated if
+  // the current position in the archive is changed.
+  virtual const Fst<Arc> *GetFst() const = 0;
+
+  virtual FarType Type() const = 0;
+
+  virtual bool Error() const = 0;
+
+  virtual ~FarReader() {}
+
+ protected:
+  FarReader() {}
+};
+
+template <class Arc>
+class FstWriter {
+ public:
+  void operator()(std::ostream &strm, const Fst<Arc> &fst) const {
+    fst.Write(strm, FstWriteOptions());
+  }
+};
+
+template <class A>
+class STTableFarWriter : public FarWriter<A> {
+ public:
+  using Arc = A;
+
+  static STTableFarWriter *Create(const string &filename) {
+    auto *writer = STTableWriter<Fst<Arc>, FstWriter<Arc>>::Create(filename);
+    return new STTableFarWriter(writer);
+  }
+
+  void Add(const string &key, const Fst<Arc> &fst) final {
+    writer_->Add(key, fst);
+  }
+
+  FarType Type() const final { return FAR_STTABLE; }
+
+  bool Error() const final { return writer_->Error(); }
+
+ private:
+  explicit STTableFarWriter(STTableWriter<Fst<Arc>, FstWriter<Arc>> *writer)
+      : writer_(writer) {}
+
+  std::unique_ptr<STTableWriter<Fst<Arc>, FstWriter<Arc>>> writer_;
+};
+
+template <class A>
+class STListFarWriter : public FarWriter<A> {
+ public:
+  using Arc = A;
+
+  static STListFarWriter *Create(const string &filename) {
+    auto *writer = STListWriter<Fst<Arc>, FstWriter<Arc>>::Create(filename);
+    return new STListFarWriter(writer);
+  }
+
+  void Add(const string &key, const Fst<Arc> &fst) final {
+    writer_->Add(key, fst);
+  }
+
+  constexpr FarType Type() const final { return FAR_STLIST; }
+
+  bool Error() const final { return writer_->Error(); }
+
+ private:
+  explicit STListFarWriter(STListWriter<Fst<Arc>, FstWriter<Arc>> *writer)
+      : writer_(writer) {}
+
+  std::unique_ptr<STListWriter<Fst<Arc>, FstWriter<Arc>>> writer_;
+};
+
+template <class A>
+class FstFarWriter : public FarWriter<A> {
+ public:
+  using Arc = A;
+
+  explicit FstFarWriter(const string &filename)
+      : filename_(filename), error_(false), written_(false) {}
+
+  static FstFarWriter *Create(const string &filename) {
+    return new FstFarWriter(filename);
+  }
+
+  void Add(const string &key, const Fst<A> &fst) final {
+    if (written_) {
+      LOG(WARNING) << "FstFarWriter::Add: only one FST supported,"
+                   << " subsequent entries discarded.";
+    } else {
+      error_ = !fst.Write(filename_);
+      written_ = true;
+    }
+  }
+
+  constexpr FarType Type() const final { return FAR_FST; }
+
+  bool Error() const final { return error_; }
+
+  ~FstFarWriter() final {}
+
+ private:
+  string filename_;
+  bool error_;
+  bool written_;
+};
+
+template <class Arc>
+FarWriter<Arc> *FarWriter<Arc>::Create(const string &filename, FarType type) {
+  switch (type) {
+    case FAR_DEFAULT:
+      if (filename.empty()) return STListFarWriter<Arc>::Create(filename);
+    case FAR_STTABLE:
+      return STTableFarWriter<Arc>::Create(filename);
+    case FAR_STLIST:
+      return STListFarWriter<Arc>::Create(filename);
+    case FAR_FST:
+      return FstFarWriter<Arc>::Create(filename);
+    default:
+      LOG(ERROR) << "FarWriter::Create: Unknown FAR type";
+      return nullptr;
+  }
+}
+
+template <class Arc>
+class FstReader {
+ public:
+  Fst<Arc> *operator()(std::istream &strm) const {
+    return Fst<Arc>::Read(strm, FstReadOptions());
+  }
+};
+
+template <class A>
+class STTableFarReader : public FarReader<A> {
+ public:
+  using Arc = A;
+
+  static STTableFarReader *Open(const string &filename) {
+    auto *reader = STTableReader<Fst<Arc>, FstReader<Arc>>::Open(filename);
+    if (!reader || reader->Error()) return nullptr;
+    return new STTableFarReader(reader);
+  }
+
+  static STTableFarReader *Open(const std::vector<string> &filenames) {
+    auto *reader = STTableReader<Fst<Arc>, FstReader<Arc>>::Open(filenames);
+    if (!reader || reader->Error()) return nullptr;
+    return new STTableFarReader(reader);
+  }
+
+  void Reset() final { reader_->Reset(); }
+
+  bool Find(const string &key) final { return reader_->Find(key); }
+
+  bool Done() const final { return reader_->Done(); }
+
+  void Next() final { return reader_->Next(); }
+
+  const string &GetKey() const final { return reader_->GetKey(); }
+
+  const Fst<Arc> *GetFst() const final { return reader_->GetEntry(); }
+
+  constexpr FarType Type() const final { return FAR_STTABLE; }
+
+  bool Error() const final { return reader_->Error(); }
+
+ private:
+  explicit STTableFarReader(STTableReader<Fst<Arc>, FstReader<Arc>> *reader)
+      : reader_(reader) {}
+
+  std::unique_ptr<STTableReader<Fst<Arc>, FstReader<Arc>>> reader_;
+};
+
+template <class A>
+class STListFarReader : public FarReader<A> {
+ public:
+  using Arc = A;
+
+  static STListFarReader *Open(const string &filename) {
+    auto *reader = STListReader<Fst<Arc>, FstReader<Arc>>::Open(filename);
+    if (!reader || reader->Error()) return nullptr;
+    return new STListFarReader(reader);
+  }
+
+  static STListFarReader *Open(const std::vector<string> &filenames) {
+    auto *reader = STListReader<Fst<Arc>, FstReader<Arc>>::Open(filenames);
+    if (!reader || reader->Error()) return nullptr;
+    return new STListFarReader(reader);
+  }
+
+  void Reset() final { reader_->Reset(); }
+
+  bool Find(const string &key) final { return reader_->Find(key); }
+
+  bool Done() const final { return reader_->Done(); }
+
+  void Next() final { return reader_->Next(); }
+
+  const string &GetKey() const final { return reader_->GetKey(); }
+
+  const Fst<Arc> *GetFst() const final { return reader_->GetEntry(); }
+
+  constexpr FarType Type() const final { return FAR_STLIST; }
+
+  bool Error() const final { return reader_->Error(); }
+
+ private:
+  explicit STListFarReader(STListReader<Fst<Arc>, FstReader<Arc>> *reader)
+      : reader_(reader) {}
+
+  std::unique_ptr<STListReader<Fst<Arc>, FstReader<Arc>>> reader_;
+};
+
+template <class A>
+class FstFarReader : public FarReader<A> {
+ public:
+  using Arc = A;
+
+  static FstFarReader *Open(const string &filename) {
+    std::vector<string> filenames;
+    filenames.push_back(filename);
+    return new FstFarReader<Arc>(filenames);
+  }
+
+  static FstFarReader *Open(const std::vector<string> &filenames) {
+    return new FstFarReader<Arc>(filenames);
+  }
+
+  explicit FstFarReader(const std::vector<string> &filenames)
+      : keys_(filenames), has_stdin_(false), pos_(0), error_(false) {
+    std::sort(keys_.begin(), keys_.end());
+    streams_.resize(keys_.size(), 0);
+    for (size_t i = 0; i < keys_.size(); ++i) {
+      if (keys_[i].empty()) {
+        if (!has_stdin_) {
+          streams_[i] = &std::cin;
+          has_stdin_ = true;
+        } else {
+          FSTERROR() << "FstFarReader::FstFarReader: standard input should "
+                        "only appear once in the input file list";
+          error_ = true;
+          return;
+        }
+      } else {
+        streams_[i] = new std::ifstream(
+            keys_[i], std::ios_base::in | std::ios_base::binary);
+        if (streams_[i]->fail()) {
+          FSTERROR() << "FstFarReader::FstFarReader: Error reading file: "
+                     << filenames[i];
+          error_ = true;
+          return;
+        }
+      }
+    }
+    if (pos_ >= keys_.size()) return;
+    ReadFst();
+  }
+
+  void Reset() final {
+    if (has_stdin_) {
+      FSTERROR()
+          << "FstFarReader::Reset: Operation not supported on standard input";
+      error_ = true;
+      return;
+    }
+    pos_ = 0;
+    ReadFst();
+  }
+
+  bool Find(const string &key) final {
+    if (has_stdin_) {
+      FSTERROR()
+          << "FstFarReader::Find: Operation not supported on standard input";
+      error_ = true;
+      return false;
+    }
+    pos_ = 0;  // TODO
+    ReadFst();
+    return true;
+  }
+
+  bool Done() const final { return error_ || pos_ >= keys_.size(); }
+
+  void Next() final {
+    ++pos_;
+    ReadFst();
+  }
+
+  const string &GetKey() const final { return keys_[pos_]; }
+
+  const Fst<Arc> *GetFst() const final { return fst_.get(); }
+
+  constexpr FarType Type() const final { return FAR_FST; }
+
+  bool Error() const final { return error_; }
+
+  ~FstFarReader() final {
+    for (size_t i = 0; i < keys_.size(); ++i) {
+      if (streams_[i] != &std::cin) {
+        delete streams_[i];
+      }
+    }
+  }
+
+ private:
+  void ReadFst() {
+    fst_.reset();
+    if (pos_ >= keys_.size()) return;
+    streams_[pos_]->seekg(0);
+    fst_.reset(Fst<Arc>::Read(*streams_[pos_], FstReadOptions()));
+    if (!fst_) {
+      FSTERROR() << "FstFarReader: Error reading Fst from: " << keys_[pos_];
+      error_ = true;
+    }
+  }
+
+  std::vector<string> keys_;
+  std::vector<std::istream *> streams_;
+  bool has_stdin_;
+  size_t pos_;
+  mutable std::unique_ptr<Fst<Arc>> fst_;
+  mutable bool error_;
+};
+
+template <class Arc>
+FarReader<Arc> *FarReader<Arc>::Open(const string &filename) {
+  if (filename.empty())
+    return STListFarReader<Arc>::Open(filename);
+  else if (IsSTTable(filename))
+    return STTableFarReader<Arc>::Open(filename);
+  else if (IsSTList(filename))
+    return STListFarReader<Arc>::Open(filename);
+  else if (IsFst(filename))
+    return FstFarReader<Arc>::Open(filename);
+  return nullptr;
+}
+
+template <class Arc>
+FarReader<Arc> *FarReader<Arc>::Open(const std::vector<string> &filenames) {
+  if (!filenames.empty() && filenames[0].empty())
+    return STListFarReader<Arc>::Open(filenames);
+  else if (!filenames.empty() && IsSTTable(filenames[0]))
+    return STTableFarReader<Arc>::Open(filenames);
+  else if (!filenames.empty() && IsSTList(filenames[0]))
+    return STListFarReader<Arc>::Open(filenames);
+  else if (!filenames.empty() && IsFst(filenames[0]))
+    return FstFarReader<Arc>::Open(filenames);
+  return nullptr;
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_FAR_FAR_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/farlib.h b/projects/llm_framework/include/fst/extensions/far/farlib.h
new file mode 100644
index 00000000..c9bb1710
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/far/farlib.h
@@ -0,0 +1,19 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// A finite-state archive (FAR) is used to store an indexable collection of
+// FSTs in a single file. Utilities are provided to create FARs from FSTs,
+// to iterate over FARs, and to extract specific FSTs from FARs.
+
+#ifndef FST_EXTENSIONS_FAR_FARLIB_H_
+#define FST_EXTENSIONS_FAR_FARLIB_H_
+
+#include <fst/extensions/far/compile-strings.h>
+#include <fst/extensions/far/create.h>
+#include <fst/extensions/far/extract.h>
+#include <fst/extensions/far/far.h>
+#include <fst/extensions/far/getters.h>
+#include <fst/extensions/far/info.h>
+#include <fst/extensions/far/print-strings.h>
+
+#endif  // FST_EXTENSIONS_FAR_FARLIB_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/farscript.h b/projects/llm_framework/include/fst/extensions/far/farscript.h
new file mode 100644
index 00000000..4bd11a94
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/far/farscript.h
@@ -0,0 +1,269 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Convenience file for including all of the FAR operations, or registering
+// them for new arc types.
+
+#ifndef FST_EXTENSIONS_FAR_FARSCRIPT_H_
+#define FST_EXTENSIONS_FAR_FARSCRIPT_H_
+
+#include <string>
+#include <vector>
+
+#include <fst/types.h>
+#include <fst/extensions/far/compile-strings.h>
+#include <fst/extensions/far/create.h>
+#include <fst/extensions/far/equal.h>
+#include <fst/extensions/far/extract.h>
+#include <fst/extensions/far/far.h>
+#include <fst/extensions/far/far-class.h>
+#include <fst/extensions/far/info.h>
+#include <fst/extensions/far/isomorphic.h>
+#include <fst/extensions/far/print-strings.h>
+#include <fst/extensions/far/script-impl.h>
+#include <fst/script/arg-packs.h>
+
+namespace fst {
+namespace script {
+
+// Note: it is safe to pass these strings as references because this struct is
+// only used to pass them deeper in the call graph. Be sure you understand why
+// this is so before using this struct for anything else!
+struct FarCompileStringsArgs {
+  const std::vector<string> &in_fnames;
+  const string &out_fname;
+  const string &fst_type;
+  const FarType &far_type;
+  const int32 generate_keys;
+  const FarEntryType fet;
+  const FarTokenType tt;
+  const string &symbols_fname;
+  const string &unknown_symbol;
+  const bool keep_symbols;
+  const bool initial_symbols;
+  const bool allow_negative_labels;
+  const string &key_prefix;
+  const string &key_suffix;
+
+  FarCompileStringsArgs(const std::vector<string> &in_fnames,
+                        const string &out_fname, const string &fst_type,
+                        const FarType &far_type, int32 generate_keys,
+                        FarEntryType fet, FarTokenType tt,
+                        const string &symbols_fname,
+                        const string &unknown_symbol, bool keep_symbols,
+                        bool initial_symbols, bool allow_negative_labels,
+                        const string &key_prefix, const string &key_suffix)
+      : in_fnames(in_fnames),
+        out_fname(out_fname),
+        fst_type(fst_type),
+        far_type(far_type),
+        generate_keys(generate_keys),
+        fet(fet),
+        tt(tt),
+        symbols_fname(symbols_fname),
+        unknown_symbol(unknown_symbol),
+        keep_symbols(keep_symbols),
+        initial_symbols(initial_symbols),
+        allow_negative_labels(allow_negative_labels),
+        key_prefix(key_prefix),
+        key_suffix(key_suffix) {}
+};
+
+template <class Arc>
+void FarCompileStrings(FarCompileStringsArgs *args) {
+  FarCompileStrings<Arc>(
+      args->in_fnames, args->out_fname, args->fst_type, args->far_type,
+      args->generate_keys, args->fet, args->tt, args->symbols_fname,
+      args->unknown_symbol, args->keep_symbols, args->initial_symbols,
+      args->allow_negative_labels, args->key_prefix, args->key_suffix);
+}
+
+void FarCompileStrings(const std::vector<string> &in_fnames,
+                       const string &out_fname, const string &arc_type,
+                       const string &fst_type, const FarType &far_type,
+                       int32 generate_keys, FarEntryType fet, FarTokenType tt,
+                       const string &symbols_fname,
+                       const string &unknown_symbol, bool keep_symbols,
+                       bool initial_symbols, bool allow_negative_labels,
+                       const string &key_prefix, const string &key_suffix);
+
+// Note: it is safe to pass these strings as references because this struct is
+// only used to pass them deeper in the call graph. Be sure you understand why
+// this is so before using this struct for anything else!
+struct FarCreateArgs {
+  const std::vector<string> &in_fnames;
+  const string &out_fname;
+  const int32 generate_keys;
+  const FarType &far_type;
+  const string &key_prefix;
+  const string &key_suffix;
+
+  FarCreateArgs(const std::vector<string> &in_fnames, const string &out_fname,
+                const int32 generate_keys, const FarType &far_type,
+                const string &key_prefix, const string &key_suffix)
+      : in_fnames(in_fnames),
+        out_fname(out_fname),
+        generate_keys(generate_keys),
+        far_type(far_type),
+        key_prefix(key_prefix),
+        key_suffix(key_suffix) {}
+};
+
+template <class Arc>
+void FarCreate(FarCreateArgs *args) {
+  FarCreate<Arc>(args->in_fnames, args->out_fname, args->generate_keys,
+                 args->far_type, args->key_prefix, args->key_suffix);
+}
+
+void FarCreate(const std::vector<string> &in_fnames, const string &out_fname,
+               const string &arc_type, const int32 generate_keys,
+               const FarType &far_type, const string &key_prefix,
+               const string &key_suffix);
+
+using FarEqualInnerArgs = std::tuple<const string &, const string &, float,
+                                     const string &, const string &>;
+
+using FarEqualArgs = WithReturnValue<bool, FarEqualInnerArgs>;
+
+template <class Arc>
+void FarEqual(FarEqualArgs *args) {
+  args->retval = fst::FarEqual<Arc>(
+      std::get<0>(args->args), std::get<1>(args->args), std::get<2>(args->args),
+      std::get<3>(args->args), std::get<4>(args->args));
+}
+
+bool FarEqual(const string &filename1, const string &filename2,
+              const string &arc_type, float delta = kDelta,
+              const string &begin_key = string(),
+              const string &end_key = string());
+
+using FarExtractArgs =
+    std::tuple<const std::vector<string> &, int32, const string &,
+               const string &, const string &, const string &, const string &>;
+
+template <class Arc>
+void FarExtract(FarExtractArgs *args) {
+  fst::FarExtract<Arc>(std::get<0>(*args), std::get<1>(*args),
+                           std::get<2>(*args), std::get<3>(*args),
+                           std::get<4>(*args), std::get<5>(*args),
+                           std::get<6>(*args));
+}
+
+void FarExtract(const std::vector<string> &ifilenames, const string &arc_type,
+                int32 generate_filenames, const string &keys,
+                const string &key_separator, const string &range_delimiter,
+                const string &filename_prefix, const string &filename_suffix);
+
+using FarInfoArgs = std::tuple<const std::vector<string> &, const string &,
+                               const string &, const bool>;
+
+template <class Arc>
+void FarInfo(FarInfoArgs *args) {
+  fst::FarInfo<Arc>(std::get<0>(*args), std::get<1>(*args),
+                        std::get<2>(*args), std::get<3>(*args));
+}
+
+void FarInfo(const std::vector<string> &filenames, const string &arc_type,
+             const string &begin_key, const string &end_key,
+             const bool list_fsts);
+
+using GetFarInfoArgs = std::tuple<const std::vector<string> &, const string &,
+                                  const string &, const bool, FarInfoData *>;
+
+template <class Arc>
+void GetFarInfo(GetFarInfoArgs *args) {
+  fst::GetFarInfo<Arc>(std::get<0>(*args), std::get<1>(*args),
+                           std::get<2>(*args), std::get<3>(*args),
+                           std::get<4>(*args));
+}
+
+void GetFarInfo(const std::vector<string> &filenames, const string &arc_type,
+                const string &begin_key, const string &end_key,
+                const bool list_fsts, FarInfoData *);
+
+using FarIsomorphicInnerArgs = std::tuple<const string &, const string &, float,
+                                          const string &, const string &>;
+
+using FarIsomorphicArgs = WithReturnValue<bool, FarIsomorphicInnerArgs>;
+
+template <class Arc>
+void FarIsomorphic(FarIsomorphicArgs *args) {
+  args->retval = fst::FarIsomorphic<Arc>(
+      std::get<0>(args->args), std::get<1>(args->args), std::get<2>(args->args),
+      std::get<3>(args->args), std::get<4>(args->args));
+}
+
+bool FarIsomorphic(const string &filename1, const string &filename2,
+                   const string &arc_type, float delta = kDelta,
+                   const string &begin_key = string(),
+                   const string &end_key = string());
+
+struct FarPrintStringsArgs {
+  const std::vector<string> &ifilenames;
+  const FarEntryType entry_type;
+  const FarTokenType token_type;
+  const string &begin_key;
+  const string &end_key;
+  const bool print_key;
+  const bool print_weight;
+  const string &symbols_fname;
+  const bool initial_symbols;
+  const int32 generate_filenames;
+  const string &filename_prefix;
+  const string &filename_suffix;
+
+  FarPrintStringsArgs(const std::vector<string> &ifilenames,
+                      const FarEntryType entry_type,
+                      const FarTokenType token_type, const string &begin_key,
+                      const string &end_key, const bool print_key,
+                      const bool print_weight, const string &symbols_fname,
+                      const bool initial_symbols,
+                      const int32 generate_filenames,
+                      const string &filename_prefix,
+                      const string &filename_suffix)
+      : ifilenames(ifilenames),
+        entry_type(entry_type),
+        token_type(token_type),
+        begin_key(begin_key),
+        end_key(end_key),
+        print_key(print_key),
+        print_weight(print_weight),
+        symbols_fname(symbols_fname),
+        initial_symbols(initial_symbols),
+        generate_filenames(generate_filenames),
+        filename_prefix(filename_prefix),
+        filename_suffix(filename_suffix) {}
+};
+
+template <class Arc>
+void FarPrintStrings(FarPrintStringsArgs *args) {
+  fst::FarPrintStrings<Arc>(
+      args->ifilenames, args->entry_type, args->token_type, args->begin_key,
+      args->end_key, args->print_key, args->print_weight, args->symbols_fname,
+      args->initial_symbols, args->generate_filenames, args->filename_prefix,
+      args->filename_suffix);
+}
+
+void FarPrintStrings(const std::vector<string> &ifilenames,
+                     const string &arc_type, const FarEntryType entry_type,
+                     const FarTokenType token_type, const string &begin_key,
+                     const string &end_key, const bool print_key,
+                     const bool print_weight, const string &symbols_fname,
+                     const bool initial_symbols, const int32 generate_filenames,
+                     const string &filename_prefix,
+                     const string &filename_suffix);
+
+}  // namespace script
+}  // namespace fst
+
+#define REGISTER_FST_FAR_OPERATIONS(ArcType)                                 \
+  REGISTER_FST_OPERATION(FarCompileStrings, ArcType, FarCompileStringsArgs); \
+  REGISTER_FST_OPERATION(FarCreate, ArcType, FarCreateArgs);                 \
+  REGISTER_FST_OPERATION(FarEqual, ArcType, FarEqualArgs);                   \
+  REGISTER_FST_OPERATION(FarExtract, ArcType, FarExtractArgs);               \
+  REGISTER_FST_OPERATION(FarInfo, ArcType, FarInfoArgs);                     \
+  REGISTER_FST_OPERATION(FarIsomorphic, ArcType, FarIsomorphicArgs);         \
+  REGISTER_FST_OPERATION(FarPrintStrings, ArcType, FarPrintStringsArgs);     \
+  REGISTER_FST_OPERATION(GetFarInfo, ArcType, GetFarInfoArgs)
+
+#endif  // FST_EXTENSIONS_FAR_FARSCRIPT_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/getters.h b/projects/llm_framework/include/fst/extensions/far/getters.h
new file mode 100644
index 00000000..3dde4194
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/far/getters.h
@@ -0,0 +1,30 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Classes and functions for registering and invoking FAR main
+// functions that support multiple and extensible arc types.
+
+#ifndef FST_EXTENSIONS_FAR_GETTERS_H_
+#define FST_EXTENSIONS_FAR_GETTERS_H_
+
+#include <fst/flags.h>
+#include <fst/extensions/far/far.h>
+
+namespace fst {
+namespace script {
+
+FarType GetFarType(const string &str);
+
+bool GetFarEntryType(const string &str, FarEntryType *entry_type);
+
+bool GetFarTokenType(const string &str, FarTokenType *token_type);
+
+void ExpandArgs(int argc, char **argv, int *argcp, char ***argvp);
+
+}  // namespace script
+
+string GetFarTypeString(FarType type);
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_FAR_GETTERS_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/info.h b/projects/llm_framework/include/fst/extensions/far/info.h
new file mode 100644
index 00000000..0391c1f4
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/far/info.h
@@ -0,0 +1,147 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_EXTENSIONS_FAR_INFO_H_
+#define FST_EXTENSIONS_FAR_INFO_H_
+
+#include <iomanip>
+#include <memory>
+#include <set>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include <fst/extensions/far/far.h>
+#include <fst/extensions/far/getters.h>
+
+namespace fst {
+
+template <class Arc>
+void AccumulateStatesAndArcs(const Fst<Arc> &fst, size_t *nstate, size_t *narc,
+                             size_t *nfinal) {
+  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done();
+       siter.Next(), ++(*nstate)) {
+    ArcIterator<Fst<Arc>> aiter(fst, siter.Value());
+    for (; !aiter.Done(); aiter.Next(), ++(*narc)) {
+    }
+    if (fst.Final(siter.Value()) != Arc::Weight::Zero()) ++(*nfinal);
+  }
+}
+
+struct KeyInfo {
+  string key;
+  string type;
+  size_t nstate = 0;
+  size_t narc = 0;
+  size_t nfinal = 0;
+};
+
+struct FarInfoData {
+  std::vector<KeyInfo> key_infos;
+  string far_type;
+  string arc_type;
+  size_t nfst = 0;
+  size_t nstate = 0;
+  size_t narc = 0;
+  size_t nfinal = 0;
+  std::set<string> fst_types;
+};
+
+template <class Arc>
+void GetFarInfo(const std::vector<string> &filenames, const string &begin_key,
+                const string &end_key, const bool list_fsts,
+                FarInfoData *far_info) {
+  *far_info = FarInfoData();
+  std::unique_ptr<FarReader<Arc>> reader(FarReader<Arc>::Open(filenames));
+  if (!reader) {
+    LOG(ERROR) << "GetFarInfo: failed to create far reader.";
+    return;
+  }
+  if (!begin_key.empty()) reader->Find(begin_key);
+
+  for (; !reader->Done(); reader->Next()) {
+    const auto &key = reader->GetKey();
+    if (!end_key.empty() && end_key < key) break;
+    ++far_info->nfst;
+    const auto *fst = reader->GetFst();
+    far_info->fst_types.insert(fst->Type());
+    if (list_fsts) {
+      KeyInfo info;
+      info.key = key;
+      info.type = fst->Type();
+      AccumulateStatesAndArcs(*fst, &info.nstate, &info.narc, &info.nfinal);
+      far_info->nstate += info.nstate;
+      far_info->narc += info.narc;
+      far_info->nfinal += info.nfinal;
+      far_info->key_infos.push_back(info);
+    } else {
+      AccumulateStatesAndArcs(*fst, &far_info->nstate, &far_info->narc,
+                              &far_info->nfinal);
+    }
+  }
+  far_info->far_type = GetFarTypeString(reader->Type());
+  far_info->arc_type = Arc::Type();
+}
+
+template <class Arc>
+void FarInfo(const std::vector<string> &filenames, const string &begin_key,
+             const string &end_key, const bool list_fsts) {
+  FarInfoData info;
+  GetFarInfo<Arc>(filenames, begin_key, end_key, list_fsts, &info);
+  if (!list_fsts) {
+    std::cout << std::left << std::setw(50) << "far type" << info.far_type
+              << std::endl;
+    std::cout << std::left << std::setw(50) << "arc type" << Arc::Type()
+              << std::endl;
+    std::cout << std::left << std::setw(50) << "fst type";
+    for (auto iter = info.fst_types.begin(); iter != info.fst_types.end();
+         ++iter) {
+      if (iter != info.fst_types.begin()) std::cout << ",";
+      std::cout << *iter;
+    }
+    std::cout << std::endl;
+    std::cout << std::left << std::setw(50) << "# of FSTs" << info.nfst
+              << std::endl;
+    std::cout << std::left << std::setw(50) << "total # of states"
+              << info.nstate << std::endl;
+    std::cout << std::left << std::setw(50) << "total # of arcs" << info.narc
+              << std::endl;
+    std::cout << std::left << std::setw(50) << "total # of final states"
+              << info.nfinal << std::endl;
+  } else {
+    // FIXME(kbg): Grok, then document this.
+    int wkey = 10;
+    int wtype = 10;
+    int wnstate = 14;
+    int wnarc = 12;
+    int wnfinal = 20;
+    for (const auto &key_info : info.key_infos) {
+      if (key_info.key.size() + 2 > wkey) wkey = key_info.key.size() + 2;
+      if (key_info.type.size() + 2 > wtype) wtype = key_info.type.size() + 2;
+      if (ceil(log10(key_info.nstate)) + 2 > wnstate) {
+        wnstate = ceil(log10(key_info.nstate)) + 2;
+      }
+      if (ceil(log10(key_info.narc)) + 2 > wnarc) {
+        wnarc = ceil(log10(key_info.narc)) + 2;
+      }
+      if (ceil(log10(key_info.nfinal)) + 2 > wnfinal) {
+        wnfinal = ceil(log10(key_info.nfinal)) + 2;
+      }
+    }
+    std::cout << std::left << std::setw(wkey) << "key" << std::setw(wtype)
+              << "type" << std::right << std::setw(wnstate) << "# of states"
+              << std::setw(wnarc) << "# of arcs" << std::setw(wnfinal)
+              << "# of final states" << std::endl;
+    for (const auto &key_info : info.key_infos) {
+      std::cout << std::left << std::setw(wkey) << key_info.key
+                << std::setw(wtype) << key_info.type << std::right
+                << std::setw(wnstate) << key_info.nstate << std::setw(wnarc)
+                << key_info.narc << std::setw(wnfinal) << key_info.nfinal
+                << std::endl;
+    }
+  }
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_FAR_INFO_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/isomorphic.h b/projects/llm_framework/include/fst/extensions/far/isomorphic.h
new file mode 100644
index 00000000..1e6e9cb3
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/far/isomorphic.h
@@ -0,0 +1,69 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_EXTENSIONS_FAR_ISOMORPHIC_H_
+#define FST_EXTENSIONS_FAR_ISOMORPHIC_H_
+
+#include <memory>
+#include <string>
+
+#include <fst/extensions/far/far.h>
+#include <fst/isomorphic.h>
+
+namespace fst {
+
+template <class Arc>
+bool FarIsomorphic(const string &filename1, const string &filename2,
+                   float delta = kDelta, const string &begin_key = string(),
+                   const string &end_key = string()) {
+  std::unique_ptr<FarReader<Arc>> reader1(FarReader<Arc>::Open(filename1));
+  if (!reader1) {
+    LOG(ERROR) << "FarIsomorphic: Cannot open FAR file " << filename1;
+    return false;
+  }
+  std::unique_ptr<FarReader<Arc>> reader2(FarReader<Arc>::Open(filename2));
+  if (!reader2) {
+    LOG(ERROR) << "FarIsomorphic: Cannot open FAR file " << filename2;
+    return false;
+  }
+  if (!begin_key.empty()) {
+    bool find_begin1 = reader1->Find(begin_key);
+    bool find_begin2 = reader2->Find(begin_key);
+    if (!find_begin1 || !find_begin2) {
+      bool ret = !find_begin1 && !find_begin2;
+      if (!ret) {
+        VLOG(1) << "FarIsomorphic: Key " << begin_key << " missing from "
+                << (find_begin1 ? "second" : "first") << " archive.";
+      }
+      return ret;
+    }
+  }
+  for (; !reader1->Done() && !reader2->Done();
+       reader1->Next(), reader2->Next()) {
+    const auto &key1 = reader1->GetKey();
+    const auto &key2 = reader2->GetKey();
+    if (!end_key.empty() && end_key < key1 && end_key < key2) return true;
+    if (key1 != key2) {
+      LOG(ERROR) << "FarIsomorphic: Mismatched keys " << key1 << " and "
+                 << key2;
+      return false;
+    }
+    if (!Isomorphic(*(reader1->GetFst()), *(reader2->GetFst()), delta)) {
+      LOG(ERROR) << "FarIsomorphic: FSTs for key " << key1
+                 << " are not isomorphic";
+      return false;
+    }
+  }
+  if (!reader1->Done() || !reader2->Done()) {
+    LOG(ERROR) << "FarIsomorphic: Key "
+               << (reader1->Done() ? reader2->GetKey() : reader1->GetKey())
+               << " missing form " << (reader2->Done() ? "first" : "second")
+               << " archive";
+    return false;
+  }
+  return true;
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_FAR_ISOMORPHIC_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/print-strings.h b/projects/llm_framework/include/fst/extensions/far/print-strings.h
new file mode 100644
index 00000000..dc428401
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/far/print-strings.h
@@ -0,0 +1,105 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Outputs as strings the string FSTs in a finite-state archive.
+
+#ifndef FST_EXTENSIONS_FAR_PRINT_STRINGS_H_
+#define FST_EXTENSIONS_FAR_PRINT_STRINGS_H_
+
+#include <iomanip>
+#include <string>
+#include <vector>
+
+#include <fst/flags.h>
+#include <fst/extensions/far/far.h>
+#include <fstream>
+#include <fst/shortest-distance.h>
+#include <fst/string.h>
+
+DECLARE_string(far_field_separator);
+
+namespace fst {
+
+template <class Arc>
+void FarPrintStrings(const std::vector<string> &ifilenames,
+                     FarEntryType entry_type, FarTokenType far_token_type,
+                     const string &begin_key, const string &end_key,
+                     bool print_key, bool print_weight,
+                     const string &symbols_fname, bool initial_symbols,
+                     int32 generate_filenames, const string &filename_prefix,
+                     const string &filename_suffix) {
+  StringTokenType token_type;
+  if (far_token_type == FTT_SYMBOL) {
+    token_type = StringTokenType::SYMBOL;
+  } else if (far_token_type == FTT_BYTE) {
+    token_type = StringTokenType::BYTE;
+  } else if (far_token_type == FTT_UTF8) {
+    token_type = StringTokenType::UTF8;
+  } else {
+    FSTERROR() << "FarPrintStrings: Unknown token type";
+    return;
+  }
+  std::unique_ptr<const SymbolTable> syms;
+  if (!symbols_fname.empty()) {
+    // TODO(kbg): Allow negative flag?
+    const SymbolTableTextOptions opts(true);
+    syms.reset(SymbolTable::ReadText(symbols_fname, opts));
+    if (!syms) {
+      LOG(ERROR) << "FarPrintStrings: Error reading symbol table "
+                 << symbols_fname;
+      return;
+    }
+  }
+  std::unique_ptr<FarReader<Arc>> far_reader(FarReader<Arc>::Open(ifilenames));
+  if (!far_reader) return;
+  if (!begin_key.empty()) far_reader->Find(begin_key);
+  string okey;
+  int nrep = 0;
+  for (int i = 1; !far_reader->Done(); far_reader->Next(), ++i) {
+    const auto &key = far_reader->GetKey();
+    if (!end_key.empty() && end_key < key) break;
+    if (okey == key) {
+      ++nrep;
+    } else {
+      nrep = 0;
+    }
+    okey = key;
+    const auto *fst = far_reader->GetFst();
+    if (i == 1 && initial_symbols && !syms && fst->InputSymbols())
+      syms.reset(fst->InputSymbols()->Copy());
+    string str;
+    VLOG(2) << "Handling key: " << key;
+    StringPrinter<Arc> string_printer(token_type,
+                                      syms ? syms.get() : fst->InputSymbols());
+    string_printer(*fst, &str);
+    if (entry_type == FET_LINE) {
+      if (print_key) std::cout << key << FLAGS_far_field_separator[0];
+      std::cout << str;
+      if (print_weight)
+        std::cout << FLAGS_far_field_separator[0] << ShortestDistance(*fst);
+      std::cout << std::endl;
+    } else if (entry_type == FET_FILE) {
+      std::stringstream sstrm;
+      if (generate_filenames) {
+        sstrm.fill('0');
+        sstrm << std::right << std::setw(generate_filenames) << i;
+      } else {
+        sstrm << key;
+        if (nrep > 0) sstrm << "." << nrep;
+      }
+      string filename;
+      filename = filename_prefix + sstrm.str() + filename_suffix;
+      std::ofstream ostrm(filename);
+      if (!ostrm) {
+        LOG(ERROR) << "FarPrintStrings: Can't open file: " << filename;
+        return;
+      }
+      ostrm << str;
+      if (token_type == StringTokenType::SYMBOL) ostrm << "\n";
+    }
+  }
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_FAR_PRINT_STRINGS_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/script-impl.h b/projects/llm_framework/include/fst/extensions/far/script-impl.h
new file mode 100644
index 00000000..a0586cc3
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/far/script-impl.h
@@ -0,0 +1,23 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Classes and functions for registering and invoking Far main
+// functions that support multiple and extensible arc types.
+
+#ifndef FST_EXTENSIONS_FAR_SCRIPT_IMPL_H_
+#define FST_EXTENSIONS_FAR_SCRIPT_IMPL_H_
+
+#include <string>
+
+#include <fst/compat.h>
+namespace fst {
+namespace script {
+
+string LoadArcTypeFromFar(const string &far_fname);
+
+string LoadArcTypeFromFst(const string &fst_fname);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_FAR_SCRIPT_IMPL_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/stlist.h b/projects/llm_framework/include/fst/extensions/far/stlist.h
new file mode 100644
index 00000000..b155e17e
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/far/stlist.h
@@ -0,0 +1,273 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// A generic (string,type) list file format.
+//
+// This is a stripped-down version of STTable that does not support the Find()
+// operation but that does support reading/writting from standard in/out.
+
+#ifndef FST_EXTENSIONS_FAR_STLIST_H_
+#define FST_EXTENSIONS_FAR_STLIST_H_
+
+#include <algorithm>
+#include <functional>
+#include <iostream>
+#include <memory>
+#include <queue>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include <fstream>
+#include <fst/util.h>
+
+namespace fst {
+
+static constexpr int32 kSTListMagicNumber = 5656924;
+static constexpr int32 kSTListFileVersion = 1;
+
+// String-type list writing class for object of type T using a functor Writer.
+// The Writer functor must provide at least the following interface:
+//
+//   struct Writer {
+//     void operator()(std::ostream &, const T &) const;
+//   };
+template <class T, class Writer>
+class STListWriter {
+ public:
+  explicit STListWriter(const string &filename)
+      : stream_(filename.empty() ? &std::cout : new std::ofstream(
+                                                    filename,
+                                                    std::ios_base::out |
+                                                        std::ios_base::binary)),
+        error_(false) {
+    WriteType(*stream_, kSTListMagicNumber);
+    WriteType(*stream_, kSTListFileVersion);
+    if (!stream_) {
+      FSTERROR() << "STListWriter::STListWriter: Error writing to file: "
+                 << filename;
+      error_ = true;
+    }
+  }
+
+  static STListWriter<T, Writer> *Create(const string &filename) {
+    return new STListWriter<T, Writer>(filename);
+  }
+
+  void Add(const string &key, const T &t) {
+    if (key == "") {
+      FSTERROR() << "STListWriter::Add: Key empty: " << key;
+      error_ = true;
+    } else if (key < last_key_) {
+      FSTERROR() << "STListWriter::Add: Key out of order: " << key;
+      error_ = true;
+    }
+    if (error_) return;
+    last_key_ = key;
+    WriteType(*stream_, key);
+    entry_writer_(*stream_, t);
+  }
+
+  bool Error() const { return error_; }
+
+  ~STListWriter() {
+    WriteType(*stream_, string());
+    if (stream_ != &std::cout) delete stream_;
+  }
+
+ private:
+  Writer entry_writer_;
+  std::ostream *stream_;  // Output stream.
+  string last_key_;       // Last key.
+  bool error_;
+
+  STListWriter(const STListWriter &) = delete;
+  STListWriter &operator=(const STListWriter &) = delete;
+};
+
+// String-type list reading class for object of type T using a functor Reader.
+// Reader must provide at least the following interface:
+//
+//   struct Reader {
+//     T *operator()(std::istream &) const;
+//   };
+template <class T, class Reader>
+class STListReader {
+ public:
+  explicit STListReader(const std::vector<string> &filenames)
+      : sources_(filenames), error_(false) {
+    streams_.resize(filenames.size(), 0);
+    bool has_stdin = false;
+    for (size_t i = 0; i < filenames.size(); ++i) {
+      if (filenames[i].empty()) {
+        if (!has_stdin) {
+          streams_[i] = &std::cin;
+          sources_[i] = "stdin";
+          has_stdin = true;
+        } else {
+          FSTERROR() << "STListReader::STListReader: Cannot read multiple "
+                     << "inputs from standard input";
+          error_ = true;
+          return;
+        }
+      } else {
+        streams_[i] = new std::ifstream(
+            filenames[i], std::ios_base::in | std::ios_base::binary);
+        if (streams_[i]->fail()) {
+          FSTERROR() << "STListReader::STListReader: Error reading file: "
+                     << filenames[i];
+          error_ = true;
+          return;
+        }
+      }
+      int32 magic_number = 0;
+      ReadType(*streams_[i], &magic_number);
+      int32 file_version = 0;
+      ReadType(*streams_[i], &file_version);
+      if (magic_number != kSTListMagicNumber) {
+        FSTERROR() << "STListReader::STListReader: Wrong file type: "
+                   << filenames[i];
+        error_ = true;
+        return;
+      }
+      if (file_version != kSTListFileVersion) {
+        FSTERROR() << "STListReader::STListReader: Wrong file version: "
+                   << filenames[i];
+        error_ = true;
+        return;
+      }
+      string key;
+      ReadType(*streams_[i], &key);
+      if (!key.empty()) heap_.push(std::make_pair(key, i));
+      if (!*streams_[i]) {
+        FSTERROR() << "STListReader: Error reading file: " << sources_[i];
+        error_ = true;
+        return;
+      }
+    }
+    if (heap_.empty()) return;
+    const auto current = heap_.top().second;
+    entry_.reset(entry_reader_(*streams_[current]));
+    if (!entry_ || !*streams_[current]) {
+      FSTERROR() << "STListReader: Error reading entry for key "
+                 << heap_.top().first << ", file " << sources_[current];
+      error_ = true;
+    }
+  }
+
+  ~STListReader() {
+    for (auto &stream : streams_) {
+      if (stream != &std::cin) delete stream;
+    }
+  }
+
+  static STListReader<T, Reader> *Open(const string &filename) {
+    std::vector<string> filenames;
+    filenames.push_back(filename);
+    return new STListReader<T, Reader>(filenames);
+  }
+
+  static STListReader<T, Reader> *Open(const std::vector<string> &filenames) {
+    return new STListReader<T, Reader>(filenames);
+  }
+
+  void Reset() {
+    FSTERROR() << "STListReader::Reset: Operation not supported";
+    error_ = true;
+  }
+
+  bool Find(const string &key) {
+    FSTERROR() << "STListReader::Find: Operation not supported";
+    error_ = true;
+    return false;
+  }
+
+  bool Done() const { return error_ || heap_.empty(); }
+
+  void Next() {
+    if (error_) return;
+    auto current = heap_.top().second;
+    string key;
+    heap_.pop();
+    ReadType(*(streams_[current]), &key);
+    if (!*streams_[current]) {
+      FSTERROR() << "STListReader: Error reading file: " << sources_[current];
+      error_ = true;
+      return;
+    }
+    if (!key.empty()) heap_.push(std::make_pair(key, current));
+    if (!heap_.empty()) {
+      current = heap_.top().second;
+      entry_.reset(entry_reader_(*streams_[current]));
+      if (!entry_ || !*streams_[current]) {
+        FSTERROR() << "STListReader: Error reading entry for key: "
+                   << heap_.top().first << ", file: " << sources_[current];
+        error_ = true;
+      }
+    }
+  }
+
+  const string &GetKey() const { return heap_.top().first; }
+
+  const T *GetEntry() const { return entry_.get(); }
+
+  bool Error() const { return error_; }
+
+ private:
+  Reader entry_reader_;                  // Read functor.
+  std::vector<std::istream *> streams_;  // Input streams.
+  std::vector<string> sources_;          // Corresponding filenames.
+  std::priority_queue<
+      std::pair<string, size_t>, std::vector<std::pair<string, size_t>>,
+      std::greater<std::pair<string, size_t>>> heap_;  // (Key, stream id) heap
+  mutable std::unique_ptr<T> entry_;  // The currently read entry.
+  bool error_;
+
+  STListReader(const STListReader &) = delete;
+  STListReader &operator=(const STListReader &) = delete;
+};
+
+// String-type list header reading function, templated on the entry header type.
+// The Header type must provide at least the following interface:
+//
+//  struct Header {
+//    void Read(std::istream &strm, const string &filename);
+//  };
+template <class Header>
+bool ReadSTListHeader(const string &filename, Header *header) {
+  if (filename.empty()) {
+    LOG(ERROR) << "ReadSTListHeader: Can't read header from standard input";
+    return false;
+  }
+  std::ifstream strm(filename, std::ios_base::in | std::ios_base::binary);
+  if (!strm) {
+    LOG(ERROR) << "ReadSTListHeader: Could not open file: " << filename;
+    return false;
+  }
+  int32 magic_number = 0;
+  ReadType(strm, &magic_number);
+  int32 file_version = 0;
+  ReadType(strm, &file_version);
+  if (magic_number != kSTListMagicNumber) {
+    LOG(ERROR) << "ReadSTListHeader: Wrong file type: " << filename;
+    return false;
+  }
+  if (file_version != kSTListFileVersion) {
+    LOG(ERROR) << "ReadSTListHeader: Wrong file version: " << filename;
+    return false;
+  }
+  string key;
+  ReadType(strm, &key);
+  header->Read(strm, filename + ":" + key);
+  if (!strm) {
+    LOG(ERROR) << "ReadSTListHeader: Error reading file: " << filename;
+    return false;
+  }
+  return true;
+}
+
+bool IsSTList(const string &filename);
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_FAR_STLIST_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/sttable.h b/projects/llm_framework/include/fst/extensions/far/sttable.h
new file mode 100644
index 00000000..2a01bb16
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/far/sttable.h
@@ -0,0 +1,353 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// A generic string-to-type table file format.
+//
+// This is not meant as a generalization of SSTable. This is more of a simple
+// replacement for SSTable in order to provide an open-source implementation
+// of the FAR format for the external version of the FST library.
+
+#ifndef FST_EXTENSIONS_FAR_STTABLE_H_
+#define FST_EXTENSIONS_FAR_STTABLE_H_
+
+#include <algorithm>
+#include <istream>
+#include <memory>
+
+#include <fstream>
+#include <fst/util.h>
+
+namespace fst {
+
+static constexpr int32 kSTTableMagicNumber = 2125656924;
+static constexpr int32 kSTTableFileVersion = 1;
+
+// String-type table writing class for an object of type T using a functor
+// Writer. The Writer functor must provide at least the following interface:
+//
+//   struct Writer {
+//     void operator()(std::ostream &, const T &) const;
+//   };
+template <class T, class Writer>
+class STTableWriter {
+ public:
+  explicit STTableWriter(const string &filename)
+      : stream_(filename, std::ios_base::out | std::ios_base::binary),
+        error_(false) {
+    WriteType(stream_, kSTTableMagicNumber);
+    WriteType(stream_, kSTTableFileVersion);
+    if (stream_.fail()) {
+      FSTERROR() << "STTableWriter::STTableWriter: Error writing to file: "
+                 << filename;
+      error_ = true;
+    }
+  }
+
+  static STTableWriter<T, Writer> *Create(const string &filename) {
+    if (filename.empty()) {
+      LOG(ERROR) << "STTableWriter: Writing to standard out unsupported.";
+      return nullptr;
+    }
+    return new STTableWriter<T, Writer>(filename);
+  }
+
+  void Add(const string &key, const T &t) {
+    if (key == "") {
+      FSTERROR() << "STTableWriter::Add: Key empty: " << key;
+      error_ = true;
+    } else if (key < last_key_) {
+      FSTERROR() << "STTableWriter::Add: Key out of order: " << key;
+      error_ = true;
+    }
+    if (error_) return;
+    last_key_ = key;
+    positions_.push_back(stream_.tellp());
+    WriteType(stream_, key);
+    entry_writer_(stream_, t);
+  }
+
+  bool Error() const { return error_; }
+
+  ~STTableWriter() {
+    WriteType(stream_, positions_);
+    WriteType(stream_, static_cast<int64>(positions_.size()));
+  }
+
+ private:
+  Writer entry_writer_;
+  std::ofstream stream_;
+  std::vector<int64> positions_;  // Position in file of each key-entry pair.
+  string last_key_;               // Last key.
+  bool error_;
+
+  STTableWriter(const STTableWriter &) = delete;
+  STTableWriter &operator=(const STTableWriter &) = delete;
+};
+
+// String-type table reading class for object of type T using a functor Reader.
+// Reader must provide at least the following interface:
+//
+//   struct Reader {
+//     T *operator()(std::istream &) const;
+//   };
+//
+template <class T, class Reader>
+class STTableReader {
+ public:
+  explicit STTableReader(const std::vector<string> &filenames)
+      : sources_(filenames), error_(false) {
+    compare_.reset(new Compare(&keys_));
+    keys_.resize(filenames.size());
+    streams_.resize(filenames.size(), 0);
+    positions_.resize(filenames.size());
+    for (size_t i = 0; i < filenames.size(); ++i) {
+      streams_[i] = new std::ifstream(
+          filenames[i], std::ios_base::in | std::ios_base::binary);
+      if (streams_[i]->fail()) {
+        FSTERROR() << "STTableReader::STTableReader: Error reading file: "
+                 << filenames[i];
+        error_ = true;
+        return;
+      }
+      int32 magic_number = 0;
+      ReadType(*streams_[i], &magic_number);
+      int32 file_version = 0;
+      ReadType(*streams_[i], &file_version);
+      if (magic_number != kSTTableMagicNumber) {
+        FSTERROR() << "STTableReader::STTableReader: Wrong file type: "
+                   << filenames[i];
+        error_ = true;
+        return;
+      }
+      if (file_version != kSTTableFileVersion) {
+        FSTERROR() << "STTableReader::STTableReader: Wrong file version: "
+                   << filenames[i];
+        error_ = true;
+        return;
+      }
+      int64 num_entries;
+      streams_[i]->seekg(-static_cast<int>(sizeof(int64)), std::ios_base::end);
+      ReadType(*streams_[i], &num_entries);
+      if (num_entries > 0) {
+        streams_[i]->seekg(-static_cast<int>(sizeof(int64)) * (num_entries + 1),
+                           std::ios_base::end);
+        positions_[i].resize(num_entries);
+        for (size_t j = 0; (j < num_entries) && (!streams_[i]->fail()); ++j) {
+          ReadType(*streams_[i], &(positions_[i][j]));
+        }
+        streams_[i]->seekg(positions_[i][0]);
+        if (streams_[i]->fail()) {
+          FSTERROR() << "STTableReader::STTableReader: Error reading file: "
+                     << filenames[i];
+          error_ = true;
+          return;
+        }
+      }
+    }
+    MakeHeap();
+  }
+
+  ~STTableReader() {
+    for (auto &stream : streams_) delete stream;
+  }
+
+  static STTableReader<T, Reader> *Open(const string &filename) {
+    if (filename.empty()) {
+      LOG(ERROR) << "STTableReader: Operation not supported on standard input";
+      return nullptr;
+    }
+    std::vector<string> filenames;
+    filenames.push_back(filename);
+    return new STTableReader<T, Reader>(filenames);
+  }
+
+  static STTableReader<T, Reader> *Open(const std::vector<string> &filenames) {
+    return new STTableReader<T, Reader>(filenames);
+  }
+
+  void Reset() {
+    if (error_) return;
+    for (size_t i = 0; i < streams_.size(); ++i)
+      streams_[i]->seekg(positions_[i].front());
+    MakeHeap();
+  }
+
+  bool Find(const string &key) {
+    if (error_) return false;
+    for (size_t i = 0; i < streams_.size(); ++i) LowerBound(i, key);
+    MakeHeap();
+    if (heap_.empty()) return false;
+    return keys_[current_] == key;
+  }
+
+  bool Done() const { return error_ || heap_.empty(); }
+
+  void Next() {
+    if (error_) return;
+    if (streams_[current_]->tellg() <= positions_[current_].back()) {
+      ReadType(*(streams_[current_]), &(keys_[current_]));
+      if (streams_[current_]->fail()) {
+        FSTERROR() << "STTableReader: Error reading file: "
+                   << sources_[current_];
+        error_ = true;
+        return;
+      }
+      std::push_heap(heap_.begin(), heap_.end(), *compare_);
+    } else {
+      heap_.pop_back();
+    }
+    if (!heap_.empty()) PopHeap();
+  }
+
+  const string &GetKey() const { return keys_[current_]; }
+
+  const T *GetEntry() const { return entry_.get(); }
+
+  bool Error() const { return error_; }
+
+ private:
+  // Comparison functor used to compare stream IDs in the heap.
+  struct Compare {
+    explicit Compare(const std::vector<string> *keys) : keys(keys) {}
+
+    bool operator()(size_t i, size_t j) const {
+      return (*keys)[i] > (*keys)[j];
+    };
+
+   private:
+    const std::vector<string> *keys;
+  };
+
+  // Positions the stream at the position corresponding to the lower bound for
+  // the specified key.
+  void LowerBound(size_t id, const string &find_key) {
+    auto *strm = streams_[id];
+    const auto &positions = positions_[id];
+    if (positions.empty()) return;
+    size_t low = 0;
+    size_t high = positions.size() - 1;
+    while (low < high) {
+      size_t mid = (low + high) / 2;
+      strm->seekg(positions[mid]);
+      string key;
+      ReadType(*strm, &key);
+      if (key > find_key) {
+        high = mid;
+      } else if (key < find_key) {
+        low = mid + 1;
+      } else {
+        for (size_t i = mid; i > low; --i) {
+          strm->seekg(positions[i - 1]);
+          ReadType(*strm, &key);
+          if (key != find_key) {
+            strm->seekg(positions[i]);
+            return;
+          }
+        }
+        strm->seekg(positions[low]);
+        return;
+      }
+    }
+    strm->seekg(positions[low]);
+  }
+
+  // Adds all streams to the heap.
+  void MakeHeap() {
+    heap_.clear();
+    for (size_t i = 0; i < streams_.size(); ++i) {
+      if (positions_[i].empty()) continue;
+      ReadType(*streams_[i], &(keys_[i]));
+      if (streams_[i]->fail()) {
+        FSTERROR() << "STTableReader: Error reading file: " << sources_[i];
+        error_ = true;
+        return;
+      }
+      heap_.push_back(i);
+    }
+    if (heap_.empty()) return;
+    std::make_heap(heap_.begin(), heap_.end(), *compare_);
+    PopHeap();
+  }
+
+  // Positions the stream with the lowest key at the top of the heap, sets
+  // current_ to the ID of that stream, and reads the current entry from that
+  // stream.
+  void PopHeap() {
+    std::pop_heap(heap_.begin(), heap_.end(), *compare_);
+    current_ = heap_.back();
+    entry_.reset(entry_reader_(*streams_[current_]));
+    if (!entry_) error_ = true;
+    if (streams_[current_]->fail()) {
+      FSTERROR() << "STTableReader: Error reading entry for key: "
+                 << keys_[current_] << ", file: " << sources_[current_];
+      error_ = true;
+    }
+  }
+
+  Reader entry_reader_;
+  std::vector<std::istream *> streams_;        // Input streams.
+  std::vector<string> sources_;                // Corresponding file names.
+  std::vector<std::vector<int64>> positions_;  // Index of positions.
+  std::vector<string> keys_;  // Lowest unread key for each stream.
+  std::vector<int64> heap_;   // Heap containing ID of streams with unread keys.
+  int64 current_;             // ID of current stream to be read.
+  std::unique_ptr<Compare> compare_;  // Functor comparing stream IDs.
+  mutable std::unique_ptr<T> entry_;  // The currently read entry.
+  bool error_;
+};
+
+// String-type table header reading function template on the entry header type.
+// The Header type must provide at least the following interface:
+//
+//   struct Header {
+//     void Read(std::istream &istrm, const string &filename);
+//   };
+template <class Header>
+bool ReadSTTableHeader(const string &filename, Header *header) {
+  if (filename.empty()) {
+    LOG(ERROR) << "ReadSTTable: Can't read header from standard input";
+    return false;
+  }
+  std::ifstream strm(filename, std::ios_base::in | std::ios_base::binary);
+  if (!strm) {
+    LOG(ERROR) << "ReadSTTableHeader: Could not open file: " << filename;
+    return false;
+  }
+  int32 magic_number = 0;
+  ReadType(strm, &magic_number);
+  int32 file_version = 0;
+  ReadType(strm, &file_version);
+  if (magic_number != kSTTableMagicNumber) {
+    LOG(ERROR) << "ReadSTTableHeader: Wrong file type: " << filename;
+    return false;
+  }
+  if (file_version != kSTTableFileVersion) {
+    LOG(ERROR) << "ReadSTTableHeader: Wrong file version: " << filename;
+    return false;
+  }
+  int64 i = -1;
+  strm.seekg(-static_cast<int>(sizeof(int64)), std::ios_base::end);
+  ReadType(strm, &i);  // Reads number of entries
+  if (strm.fail()) {
+    LOG(ERROR) << "ReadSTTableHeader: Error reading file: " << filename;
+    return false;
+  }
+  if (i == 0) return true;  // No entry header to read.
+  strm.seekg(-2 * static_cast<int>(sizeof(int64)), std::ios_base::end);
+  ReadType(strm, &i);  // Reads position for last entry in file.
+  strm.seekg(i);
+  string key;
+  ReadType(strm, &key);
+  header->Read(strm, filename + ":" + key);
+  if (strm.fail()) {
+    LOG(ERROR) << "ReadSTTableHeader: Error reading file: " << filename;
+    return false;
+  }
+  return true;
+}
+
+bool IsSTTable(const string &filename);
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_FAR_STTABLE_H_
diff --git a/projects/llm_framework/include/fst/extensions/linear/linear-fst-data-builder.h b/projects/llm_framework/include/fst/extensions/linear/linear-fst-data-builder.h
new file mode 100644
index 00000000..a6ac7279
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/linear/linear-fst-data-builder.h
@@ -0,0 +1,1074 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_BUILDER_H_
+#define FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_BUILDER_H_
+
+#include <map>
+#include <queue>
+#include <set>
+#include <sstream>
+#include <stack>
+#include <string>
+#include <vector>
+
+#include <fst/compat.h>
+#include <fst/log.h>
+#include <fst/fst.h>
+#include <fst/symbol-table.h>
+#include <fst/util.h>
+
+#include <fst/extensions/linear/linear-fst-data.h>
+
+namespace fst {
+
+// Forward declaration
+template <class A>
+class FeatureGroupBuilder;
+
+// For logging purposes
+inline string TranslateLabel(int64 label, const SymbolTable *syms);
+template <class Iterator>
+string JoinLabels(Iterator begin, Iterator end, const SymbolTable *syms);
+template <class Label>
+string JoinLabels(const std::vector<Label> &labels, const SymbolTable *syms);
+
+// Guesses the appropriate boundary label (start- or end-of-sentence)
+// for all labels equal to `boundary` and modifies the `sequence`
+// in-place. Returns the number of positions that are still uncertain.
+template <class A>
+typename A::Label GuessStartOrEnd(std::vector<typename A::Label> *sequence,
+                                  typename A::Label boundary);
+
+// Builds a `LinearFstData` object by adding words and feature
+// weights. A few conventions:
+//
+// - Input labels forms a dense non-empty range from 1 to `MaxInputLabel()`.
+// - Feature labels, output labels are > 0.
+// - Being a discriminative linear model, it only makes sense to use tropical
+// semirings.
+template <class A>
+class LinearFstDataBuilder {
+ public:
+  typedef typename A::Label Label;
+  typedef typename A::Weight Weight;
+
+  // Constructs a builder with associated symbol tables for diagonstic
+  // output. Each of these symbol tables may also be nullptr.
+  explicit LinearFstDataBuilder(const SymbolTable *isyms = nullptr,
+                                const SymbolTable *fsyms = nullptr,
+                                const SymbolTable *osyms = nullptr)
+      : error_(false),
+        max_future_size_(0),
+        max_input_label_(1),
+        isyms_(isyms),
+        fsyms_(fsyms),
+        osyms_(osyms) {}
+
+  // Tests whether the builder has encountered any error. No operation
+  // is valid if the builder is already at error state. All other
+  // public methods should check this before any actual operations.
+  bool Error() const { return error_; }
+
+  // Adds a word and its feature labels to the vocabulary; this
+  // version allows the word to have any output label. Returns true
+  // iff the word is added.
+  //
+  // This may fail if the word is added twice or if the feature labels
+  // are non-positive.
+  bool AddWord(Label word, const std::vector<Label> &features);
+
+  // Adds a word and its feature labels to the vocabulary; this
+  // version puts constraint on possible output labels the word can
+  // have. `possible_output` must not be empty. Returns true iff the
+  // word is added.
+  //
+  // In addition to the reasons above in the two-parameter version,
+  // this may also fail if `possible_output` is empty or any output
+  // label in it is non-positive.
+  bool AddWord(Label word, const std::vector<Label> &word_features,
+               const std::vector<Label> &possible_output);
+
+  // Creates a new feature group with specified future size (size of
+  // the look-ahead window), returns the group id to be used for
+  // adding actual feature weights or a negative number when called at
+  // error state.
+  //
+  // This does not fail unless called at error state.
+  int AddGroup(size_t future_size);
+
+  // Adds an instance of feature weight to the specified feature
+  // group. If some weight has already been added with the same
+  // feature, the product of the old and new weights are
+  // stored. Returns true iff the weight is added. A weight is not
+  // added when the context has ill-formed context involving start-,
+  // end-of-sentence marks.
+  //
+  // For two features to be within the same group, it must satisfy
+  // that (1) they have the same future size; (2) the two either have
+  // disjoint context or one is the back-off context of the
+  // other. Furthermore, for all features in a single group, there
+  // must be one and only one other context (not necessarily an active
+  // feature) that the feature immediately backs off to (i.e. there is
+  // no other context that is the back-off of the first and backs off
+  // to the second).
+  //
+  // Consider for example features with zero look-ahead of the form
+  // (input, OUTPUT).
+  //
+  // - The following two features can be put in the same group because
+  // their context is disjoint: (a a a, A A), (b, B B);
+  //
+  // - The following two features can be put in the same group because
+  // one is the back-off context of the other: (a a a, A A), (a a, A
+  // A);
+  //
+  // - The following two features can NOT be put in the same group
+  // because there is overlap but neither is the other's back-off: (a
+  // a a, A), (a a, A A);
+  //
+  // - Finally, the following three features cannot be in a same group
+  // because the first one can immediately back off to either of the
+  // rest: (a a a, A A), (a a, A A), (a a a, A).
+  //
+  // The easiest way to satisfy the constraints is to create a feature
+  // group for each feature template. However, better feature grouping
+  // may help improve speed.
+  //
+  // This may fail if any of input or output labels are non-positive,
+  // or if any call to `FeatureGroupBuilder<>::AddWeight()` fails.
+  bool AddWeight(size_t group, const std::vector<Label> &input,
+                 const std::vector<Label> &output, Weight weight);
+
+  // Returns a newly created `LinearFstData` object or nullptr in case
+  // of failure. The caller takes the ownership of the memory. No
+  // other methods shall be called after this --- this is enforced by
+  // putting the builder at error state, even when a
+  // `LinearFstData<>` object is successfully built.
+  //
+  // This may fail if the call to any `FeatureGroupBuilder<>::Dump()`
+  // fails.
+  LinearFstData<A> *Dump();
+
+ private:
+  bool error_;
+  CompactSet<Label, kNoLabel> all_output_labels_;
+  std::map<Label, std::set<Label>> word_output_map_, word_feat_map_;
+  std::map<Label, std::set<size_t>> feat_groups_;
+  std::vector<std::unique_ptr<FeatureGroupBuilder<A>>> groups_;
+  size_t max_future_size_;
+  Label max_input_label_;
+  const SymbolTable *isyms_, *fsyms_, *osyms_;
+
+  LinearFstDataBuilder(const LinearFstDataBuilder &) = delete;
+  LinearFstDataBuilder &operator=(const LinearFstDataBuilder &) = delete;
+};
+
+// Builds a LinearFstData tailored for a LinearClassifierFst. The
+// major difference between an ordinary LinearFstData that works on
+// taggers and a LinearFstData that works on classifiers is that
+// feature groups are divided into sections by the prediction class
+// label. For a prediction label `pred` and a logical group id
+// `group`, the actual group id is `group * num_classes + pred -
+// 1`.
+//
+// This layout saves us from recording output labels in each single
+// FeatureGroup. Because there is no need for any delaying, stripping
+// the output allows features with different shapes but using the same
+// set of feature label mapping to reside in a single FeatureGroup.
+template <class A>
+class LinearClassifierFstDataBuilder {
+ public:
+  typedef typename A::Label Label;
+  typedef typename A::Weight Weight;
+
+  // Constructs a builder for a `num_classes`-class classifier,
+  // optinally with associated symbol tables for diagnostic
+  // output. The output labels (i.e. prediction) must be in the range
+  // of [1, num_classes].
+  explicit LinearClassifierFstDataBuilder(size_t num_classes,
+                                          const SymbolTable *isyms = nullptr,
+                                          const SymbolTable *fsyms = nullptr,
+                                          const SymbolTable *osyms = nullptr)
+      : error_(false),
+        num_classes_(num_classes),
+        num_groups_(0),
+        builder_(isyms, fsyms, osyms) {}
+
+  // Tests whether the builder has encountered any error. Similar to
+  // LinearFstDataBuilder<>::Error().
+  bool Error() const { return error_; }
+
+  // Same as LinearFstDataBuilder<>::AddWord().
+  bool AddWord(Label word, const std::vector<Label> &features);
+
+  // Adds a logical feature group. Similar to
+  // LinearFstDataBuilder<>::AddGroup(), with the exception that the
+  // returned group id is the logical group id. Also there is no need
+  // for "future" in a classifier.
+  int AddGroup();
+
+  // Adds an instance of feature weight to the specified logical
+  // feature group. Instead of a vector of output, only a single
+  // prediction is needed as the output.
+  //
+  // This may fail if `pred` is not in the range of [1, num_classes_].
+  bool AddWeight(size_t group, const std::vector<Label> &input, Label pred,
+                 Weight weight);
+
+  // Returns a newly created `LinearFstData` object or nullptr in case of
+  // failure.
+  LinearFstData<A> *Dump();
+
+ private:
+  std::vector<Label> empty_;
+  bool error_;
+  size_t num_classes_, num_groups_;
+  LinearFstDataBuilder<A> builder_;
+};
+
+// Builds a single feature group. Usually used in
+// `LinearFstDataBuilder::AddWeight()`. See that method for the
+// constraints on grouping features.
+template <class A>
+class FeatureGroupBuilder {
+ public:
+  typedef typename A::Label Label;
+  typedef typename A::Weight Weight;
+
+  // Constructs a builder with the given future size. All features
+  // added to the group will have look-ahead windows of this size.
+  FeatureGroupBuilder(size_t future_size, const SymbolTable *fsyms,
+                      const SymbolTable *osyms)
+      : error_(false), future_size_(future_size), fsyms_(fsyms), osyms_(osyms) {
+    // This edge is special; see doc of class `FeatureGroup` on the
+    // details.
+    start_ = trie_.Insert(trie_.Root(), InputOutputLabel(kNoLabel, kNoLabel));
+  }
+
+  // Tests whether the builder has encountered any error. No operation
+  // is valid if the builder is already at error state. All other
+  // public methods should check this before any actual operations.
+  bool Error() const { return error_; }
+
+  // Adds a feature weight with the given context. Returns true iff
+  // the weight is added. A weight is not added if it has ill-formed
+  // context involving start-, end-of-sentence marks.
+  //
+  // Note: `input` is the sequence of input
+  // features, instead of input labels themselves. `input` must be at
+  // least as long as `future_size`; `output` may be empty, but
+  // usually should be non-empty because an empty output context is
+  // useless in discriminative modelling. All labels in both `input`
+  // and `output` must be > 0 (this is checked in
+  // `LinearFstDataBuilder::AddWeight()`). See
+  // LinearFstDataBuilder<>::AddWeight for more details.
+  //
+  // This may fail if the input is smaller than the look-ahead window.
+  bool AddWeight(const std::vector<Label> &input,
+                 const std::vector<Label> &output, Weight weight);
+
+  // Creates an actual FeatureGroup<> object. Connects back-off links;
+  // pre-accumulates weights from back-off features. Returns nullptr if
+  // there is any violation in unique immediate back-off
+  // constraints.
+  //
+  // Regardless of whether the call succeeds or not, the error flag is
+  // always set before this returns, to prevent repeated dumping.
+  //
+  // TODO(wuke): check overlapping top-level contexts (see
+  // `DumpOverlappingContext()` in tests).
+  FeatureGroup<A> *Dump(size_t max_future_size);
+
+ private:
+  typedef typename FeatureGroup<A>::InputOutputLabel InputOutputLabel;
+  typedef typename FeatureGroup<A>::InputOutputLabelHash InputOutputLabelHash;
+  typedef typename FeatureGroup<A>::WeightBackLink WeightBackLink;
+  // Nested trie topology uses more memory but we can traverse a
+  // node's children easily, which is required in `BuildBackLinks()`.
+  typedef NestedTrieTopology<InputOutputLabel, InputOutputLabelHash> Topology;
+  typedef MutableTrie<InputOutputLabel, WeightBackLink, Topology> Trie;
+
+  // Finds the first node with an arc with `label` following the
+  // back-off chain of `parent`. Returns the node index or
+  // `kNoTrieNodeId` when not found. The number of hops is stored in
+  // `hop` when it is not `nullptr`.
+  //
+  // This does not fail.
+  int FindFirstMatch(InputOutputLabel label, int parent, int *hop) const;
+
+  // Links each node to its immediate back-off. root is linked to -1.
+  //
+  // This may fail when the unique immediate back-off constraint is
+  // violated.
+  void BuildBackLinks();
+
+  // Traces back on the back-chain for each node to multiply the
+  // weights from back-offs to the node itself.
+  //
+  // This does not fail.
+  void PreAccumulateWeights();
+
+  // Reconstruct the path from trie root to given node for logging.
+  bool TrieDfs(const Topology &topology, int cur, int target,
+               std::vector<InputOutputLabel> *path) const;
+  string TriePath(int node, const Topology &topology) const;
+
+  bool error_;
+  size_t future_size_;
+  Trie trie_;
+  int start_;
+  const SymbolTable *fsyms_, *osyms_;
+
+  FeatureGroupBuilder(const FeatureGroupBuilder &) = delete;
+  FeatureGroupBuilder &operator=(const FeatureGroupBuilder &) = delete;
+};
+
+//
+// Implementation of methods in `LinearFstDataBuilder`
+//
+template <class A>
+bool LinearFstDataBuilder<A>::AddWord(Label word,
+                                      const std::vector<Label> &features) {
+  if (error_) {
+    FSTERROR() << "Calling LinearFstDataBuilder<>::AddWord() at error state";
+    return false;
+  }
+  if (word == LinearFstData<A>::kStartOfSentence ||
+      word == LinearFstData<A>::kEndOfSentence) {
+    LOG(WARNING) << "Ignored: adding boundary label: "
+                 << TranslateLabel(word, isyms_)
+                 << "(start-of-sentence=" << LinearFstData<A>::kStartOfSentence
+                 << ", end-of-sentence=" << LinearFstData<A>::kEndOfSentence
+                 << ")";
+    return false;
+  }
+  if (word <= 0) {
+    error_ = true;
+    FSTERROR() << "Word label must be > 0; got " << word;
+    return false;
+  }
+  if (word > max_input_label_) max_input_label_ = word;
+  // Make sure the word hasn't been added before
+  if (word_feat_map_.find(word) != word_feat_map_.end()) {
+    error_ = true;
+    FSTERROR() << "Input word " << TranslateLabel(word, isyms_)
+               << " is added twice";
+    return false;
+  }
+  // Store features
+  std::set<Label> *feats = &word_feat_map_[word];
+  for (size_t i = 0; i < features.size(); ++i) {
+    Label feat = features[i];
+    if (feat <= 0) {
+      error_ = true;
+      FSTERROR() << "Feature label must be > 0; got " << feat;
+      return false;
+    }
+    feats->insert(feat);
+  }
+  return true;
+}
+
+template <class A>
+bool LinearFstDataBuilder<A>::AddWord(
+    Label word, const std::vector<Label> &word_features,
+    const std::vector<Label> &possible_output) {
+  if (error_) {
+    FSTERROR() << "Calling LinearFstDataBuilder<>::AddWord() at error state";
+    return false;
+  }
+  if (!AddWord(word, word_features)) return false;
+  // Store possible output constraint
+  if (possible_output.empty()) {
+    error_ = true;
+    FSTERROR() << "Empty possible output constraint; "
+               << "use the two-parameter version if no constraint is need.";
+    return false;
+  }
+  std::set<Label> *outputs = &word_output_map_[word];
+  for (size_t i = 0; i < possible_output.size(); ++i) {
+    Label output = possible_output[i];
+    if (output == LinearFstData<A>::kStartOfSentence ||
+        output == LinearFstData<A>::kEndOfSentence) {
+      LOG(WARNING) << "Ignored: word = " << TranslateLabel(word, isyms_)
+                   << ": adding boundary label as possible output: " << output
+                   << "(start-of-sentence="
+                   << LinearFstData<A>::kStartOfSentence
+                   << ", end-of-sentence=" << LinearFstData<A>::kEndOfSentence
+                   << ")";
+      continue;
+    }
+    if (output <= 0) {
+      error_ = true;
+      FSTERROR() << "Output label must be > 0; got " << output;
+      return false;
+    }
+    outputs->insert(output);
+    all_output_labels_.Insert(output);
+  }
+  return true;
+}
+
+template <class A>
+inline int LinearFstDataBuilder<A>::AddGroup(size_t future_size) {
+  if (error_) {
+    FSTERROR() << "Calling LinearFstDataBuilder<>::AddGroup() at error state";
+    return -1;
+  }
+  size_t ret = groups_.size();
+  groups_.emplace_back(new FeatureGroupBuilder<A>(future_size, fsyms_, osyms_));
+  if (future_size > max_future_size_) max_future_size_ = future_size;
+  return ret;
+}
+
+template <class A>
+bool LinearFstDataBuilder<A>::AddWeight(size_t group,
+                                        const std::vector<Label> &input,
+                                        const std::vector<Label> &output,
+                                        Weight weight) {
+  if (error_) {
+    FSTERROR() << "Calling LinearFstDataBuilder<>::AddWeight() at error state";
+    return false;
+  }
+  // Check well-formedness of boundary marks on the input.
+  {
+    bool start_in_middle = false, end_in_middle = false;
+    for (int i = 1; i < input.size(); ++i) {
+      if (input[i] == LinearFstData<A>::kStartOfSentence &&
+          input[i - 1] != LinearFstData<A>::kStartOfSentence)
+        start_in_middle = true;
+      if (input[i - 1] == LinearFstData<A>::kEndOfSentence &&
+          input[i] != LinearFstData<A>::kEndOfSentence)
+        end_in_middle = true;
+    }
+    if (start_in_middle) {
+      LOG(WARNING) << "Ignored: start-of-sentence in the middle of the input!";
+      LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
+      LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
+      return false;
+    }
+    if (end_in_middle) {
+      LOG(WARNING) << "Ignored: end-of-sentence in the middle of the input!";
+      LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
+      LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
+      return false;
+    }
+  }
+  // Check well-formedness of boundary marks on the output.
+  {
+    bool non_first_start = false, non_last_end = false;
+    for (int i = 1; i < output.size(); ++i) {
+      if (output[i] == LinearFstData<A>::kStartOfSentence)
+        non_first_start = true;
+      if (output[i - 1] == LinearFstData<A>::kEndOfSentence)
+        non_last_end = true;
+    }
+    if (non_first_start) {
+      LOG(WARNING) << "Ignored: start-of-sentence not appearing "
+                   << "as the first label in the output!";
+      LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
+      LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
+      return false;
+    }
+    if (non_last_end) {
+      LOG(WARNING) << "Ignored: end-of-sentence not appearing "
+                   << "as the last label in the output!";
+      LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
+      LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
+      return false;
+    }
+  }
+
+  for (size_t i = 0; i < input.size(); ++i) {
+    Label feat = input[i];
+    if (feat != LinearFstData<A>::kStartOfSentence &&
+        feat != LinearFstData<A>::kEndOfSentence && feat <= 0) {
+      error_ = true;
+      FSTERROR() << "Feature label must be > 0; got " << feat;
+      return false;
+    }
+    feat_groups_[feat].insert(group);
+  }
+  for (size_t i = 0; i < output.size(); ++i) {
+    Label label = output[i];
+    if (label != LinearFstData<A>::kStartOfSentence &&
+        label != LinearFstData<A>::kEndOfSentence && label <= 0) {
+      error_ = true;
+      FSTERROR() << "Output label must be > 0; got " << label;
+      return false;
+    }
+    if (label != LinearFstData<A>::kStartOfSentence &&
+        label != LinearFstData<A>::kEndOfSentence)
+      all_output_labels_.Insert(label);
+  }
+
+  // Everything looks good at this point (more checks on the way in
+  // the feature group). Add this feature weight.
+  bool added = groups_[group]->AddWeight(input, output, weight);
+  if (groups_[group]->Error()) {
+    error_ = true;
+    FSTERROR() << "FeatureGroupBuilder<>::AddWeight() failed";
+    return false;
+  }
+  return added;
+}
+
+template <class A>
+LinearFstData<A> *LinearFstDataBuilder<A>::Dump() {
+  if (error_) {
+    FSTERROR() << "Calling LinearFstDataBuilder<>::Dump() at error state";
+    return nullptr;
+  }
+
+  std::unique_ptr<LinearFstData<A>> data(new LinearFstData<A>());
+  data->max_future_size_ = max_future_size_;
+  data->max_input_label_ = max_input_label_;
+
+  // Feature groups; free builders after it's dumped.
+  data->groups_.resize(groups_.size());
+  for (int group = 0; group != groups_.size(); ++group) {
+    FeatureGroup<A> *new_group = groups_[group]->Dump(max_future_size_);
+    if (new_group == nullptr) {
+      error_ = true;
+      FSTERROR() << "Error in dumping group " << group;
+      return nullptr;
+    }
+    data->groups_[group].reset(new_group);
+    groups_[group].reset();
+    VLOG(1) << "Group " << group << ": " << new_group->Stats();
+  }
+
+  // Per-group feature mapping
+  data->group_feat_map_.Init(data->NumGroups(), max_input_label_ + 1);
+  for (Label word = 1; word <= max_input_label_; ++word) {
+    typename std::map<Label, std::set<Label>>::const_iterator it =
+        word_feat_map_.find(word);
+    if (it == word_feat_map_.end()) continue;
+    for (typename std::set<Label>::const_iterator oit = it->second.begin();
+         oit != it->second.end(); ++oit) {
+      Label feat = *oit;
+      typename std::map<Label, std::set<size_t>>::const_iterator jt =
+          feat_groups_.find(feat);
+      if (jt == feat_groups_.end()) continue;
+      for (std::set<size_t>::const_iterator git = jt->second.begin();
+           git != jt->second.end(); ++git) {
+        size_t group_id = *git;
+        if (!data->group_feat_map_.Set(group_id, word, feat)) {
+          error_ = true;
+          return nullptr;
+        }
+      }
+    }
+  }
+
+  // Possible output labels
+  {
+    std::vector<typename LinearFstData<A>::InputAttribute> *input_attribs =
+        &data->input_attribs_;
+    std::vector<Label> *output_pool = &data->output_pool_;
+    input_attribs->resize(max_input_label_ + 1);
+    for (Label word = 0; word <= max_input_label_; ++word) {
+      typename std::map<Label, std::set<Label>>::const_iterator it =
+          word_output_map_.find(word);
+      if (it == word_output_map_.end()) {
+        (*input_attribs)[word].output_begin = 0;
+        (*input_attribs)[word].output_length = 0;
+      } else {
+        (*input_attribs)[word].output_begin = output_pool->size();
+        (*input_attribs)[word].output_length = it->second.size();
+        for (typename std::set<Label>::const_iterator oit = it->second.begin();
+             oit != it->second.end(); ++oit) {
+          Label olabel = *oit;
+          output_pool->push_back(olabel);
+        }
+      }
+    }
+  }
+
+  for (typename CompactSet<Label, kNoLabel>::const_iterator it =
+           all_output_labels_.Begin();
+       it != all_output_labels_.End(); ++it)
+    data->output_set_.push_back(*it);
+
+  error_ = true;  // prevent future calls on this object
+  return data.release();
+}
+
+//
+// Implementation of methods in `LinearClassifierFstDataBuilder`
+//
+template <class A>
+inline bool LinearClassifierFstDataBuilder<A>::AddWord(
+    Label word, const std::vector<Label> &features) {
+  if (error_) {
+    FSTERROR() << "Calling LinearClassifierFstDataBuilder<>::AddWord() at "
+                  "error state";
+    return false;
+  }
+  bool added = builder_.AddWord(word, features);
+  if (builder_.Error()) error_ = true;
+  return added;
+}
+
+template <class A>
+inline int LinearClassifierFstDataBuilder<A>::AddGroup() {
+  if (error_) {
+    FSTERROR() << "Calling LinearClassifierFstDataBuilder<>::AddGroup() at "
+                  "error state";
+    return -1;
+  }
+  for (int i = 0; i < num_classes_; ++i) builder_.AddGroup(0);
+  if (builder_.Error()) {
+    error_ = true;
+    return -1;
+  }
+  return num_groups_++;
+}
+
+template <class A>
+inline bool LinearClassifierFstDataBuilder<A>::AddWeight(
+    size_t group, const std::vector<Label> &input, Label pred, Weight weight) {
+  if (error_) {
+    FSTERROR() << "Calling LinearClassifierFstDataBuilder<>::AddWeight() at "
+                  "error state";
+    return false;
+  }
+  if (pred <= 0 || pred > num_classes_) {
+    FSTERROR() << "Out-of-range prediction label: " << pred
+               << " (num classes = " << num_classes_ << ")";
+    error_ = true;
+    return false;
+  }
+  size_t real_group = group * num_classes_ + pred - 1;
+  bool added = builder_.AddWeight(real_group, input, empty_, weight);
+  if (builder_.Error()) error_ = true;
+  return added;
+}
+
+template <class A>
+inline LinearFstData<A> *LinearClassifierFstDataBuilder<A>::Dump() {
+  if (error_) {
+    FSTERROR()
+        << "Calling LinearClassifierFstDataBuilder<>::Dump() at error state";
+    return nullptr;
+  }
+  LinearFstData<A> *data = builder_.Dump();
+  error_ = true;
+  return data;
+}
+
+//
+// Implementation of methods in `FeatureGroupBuilder`
+//
+template <class A>
+bool FeatureGroupBuilder<A>::AddWeight(const std::vector<Label> &input,
+                                       const std::vector<Label> &output,
+                                       Weight weight) {
+  if (error_) {
+    FSTERROR() << "Calling FeatureGroupBuilder<>::AddWeight() at error state";
+    return false;
+  }
+
+  // `LinearFstDataBuilder<>::AddWeight()` ensures prefix/suffix
+  // properties for us. We can directly count.
+  int num_input_start = 0;
+  while (num_input_start < input.size() &&
+         input[num_input_start] == LinearFstData<A>::kStartOfSentence)
+    ++num_input_start;
+  int num_output_start = 0;
+  while (num_output_start < output.size() &&
+         output[num_output_start] == LinearFstData<A>::kStartOfSentence)
+    ++num_output_start;
+  int num_input_end = 0;
+  for (int i = input.size() - 1;
+       i >= 0 && input[i] == LinearFstData<A>::kEndOfSentence; --i)
+    ++num_input_end;
+  int num_output_end = 0;
+  for (int i = output.size() - 1;
+       i >= 0 && output[i] == LinearFstData<A>::kEndOfSentence; --i)
+    ++num_output_end;
+
+  DCHECK_LE(num_output_end, 1);
+
+  if (input.size() - num_input_start < future_size_) {
+    LOG(WARNING) << "Ignored: start-of-sentence in the future!";
+    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
+    LOG(WARNING) << "\tOutput: " << JoinLabels(output, fsyms_);
+    return false;
+  }
+  if (num_input_start > 0 &&
+      input.size() - future_size_ - num_input_start <
+          output.size() - num_output_start) {
+    LOG(WARNING) << "Ignored: matching start-of-sentence with actual output!";
+    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
+    LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
+    return false;
+  }
+  if (num_output_start > 0 &&
+      input.size() - future_size_ - num_input_start >
+          output.size() - num_output_start) {
+    LOG(WARNING) << "Ignored: matching start-of-sentence with actual input!";
+    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
+    LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
+    return false;
+  }
+  // The following two require `num_output_end` <= 1.
+  if (num_input_end > future_size_ && num_input_end - future_size_ != 1) {
+    LOG(WARNING) << "Ignored: matching end-of-sentence with actual output!";
+    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
+    LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
+    return false;
+  }
+  if (num_output_end > 0 &&
+      ((input.size() == future_size_ && future_size_ != num_input_end) ||
+       (input.size() > future_size_ &&
+        num_input_end != future_size_ + num_output_end))) {
+    LOG(WARNING) << "Ignored: matching end-of-sentence with actual input!";
+    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
+    LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
+    return false;
+  }
+  // Check if the context has no other labels than boundary marks
+  // (such features are useless).
+  if (num_input_start + num_input_end == input.size() &&
+      num_output_start + num_output_end == output.size()) {
+    LOG(WARNING)
+        << "Ignored: feature context consisting of only boundary marks!";
+    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
+    LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
+    return false;
+  }
+
+  // Start point for insertion in the trie. Insert at `start_` iff the
+  // beginning of the context is non-consumed start-of-sentence.
+  int cur = (num_input_start == 0 && num_output_start <= future_size_)
+                ? trie_.Root()
+                : start_;
+  // Skip all input start-of-sentence marks
+  size_t ipos = num_input_start;
+  // Skip to keep at most `future_size_` start-of-sentence marks
+  size_t opos =
+      num_output_start <= future_size_ ? 0 : num_output_start - future_size_;
+  // Skip `num_output_end` end-of-sentence marks on both input and output
+  size_t iend = !input.empty() ? input.size() - num_output_end : 0,
+         oend = output.size() - num_output_end;
+  // Further, when output is empty, keep at most `future_size_`
+  // end-of-sentence marks on input.
+  if (output.empty() && num_input_end > future_size_)
+    iend = input.size() - num_input_end + future_size_;
+
+  // Actual feature context is (input[ipos:iend], output[opos:oend]).
+
+  // Pad `kNoLabel` as don't cares on the shorter of actual `input`
+  // and `output`.
+  const size_t effective_input_size = iend - ipos,
+               effective_output_size = oend - opos;
+  if (effective_input_size > effective_output_size) {
+    for (size_t pad = effective_input_size - effective_output_size; pad != 0;
+         --pad, ++ipos)
+      cur = trie_.Insert(cur, InputOutputLabel(input[ipos], kNoLabel));
+  } else if (effective_input_size < effective_output_size) {
+    for (size_t pad = effective_output_size - effective_input_size; pad != 0;
+         --pad, ++opos)
+      cur = trie_.Insert(cur, InputOutputLabel(kNoLabel, output[opos]));
+  }
+  CHECK_EQ(iend - ipos, oend - opos);
+  for (; ipos != iend; ++ipos, ++opos)
+    cur = trie_.Insert(cur, InputOutputLabel(input[ipos], output[opos]));
+  // We only need to attach final weight when there is an output
+  // end-of-sentence. When there is only end-of-sentence on the input,
+  // they are all consumed as the end-of-sentence paddings from
+  // `LinearFstImpl<>::ShiftBuffer()`. `LinearFstImpl<>::Expand()`
+  // and `LinearFstImpl<>::MatchInput()` ensures no other
+  // transition takes place after consuming the padding.
+  if (num_output_end > 0 || (output.empty() && num_input_end > future_size_))
+    trie_[cur].final_weight = Times(trie_[cur].final_weight, weight);
+  else
+    trie_[cur].weight = Times(trie_[cur].weight, weight);
+
+  return true;
+}
+
+template <class A>
+FeatureGroup<A> *FeatureGroupBuilder<A>::Dump(size_t max_future_size) {
+  if (error_) {
+    FSTERROR() << "Calling FeatureGroupBuilder<>::PreAccumulateWeights() "
+               << "at error state";
+    return nullptr;
+  }
+
+  if (max_future_size < future_size_) {
+    error_ = true;
+    FSTERROR() << "max_future_size (= " << max_future_size
+               << ") is smaller the builder's future_size (= " << future_size_
+               << ")";
+    return nullptr;
+  }
+
+  BuildBackLinks();
+  if (error_) return nullptr;
+  PreAccumulateWeights();  // does not fail
+
+  FeatureGroup<A> *ret =
+      new FeatureGroup<A>(max_future_size - future_size_, start_);
+
+  // Walk around the trie to compute next states
+  ret->next_state_.resize(trie_.NumNodes());
+  const Topology &topology = trie_.TrieTopology();
+  for (int i = 0; i < topology.NumNodes(); ++i) {
+    int next = i;
+    while (next != topology.Root() && topology.ChildrenOf(next).empty() &&
+           trie_[next].final_weight ==
+               trie_[trie_[next].back_link].final_weight)
+      next = trie_[next].back_link;
+    ret->next_state_[i] = next;
+  }
+
+  // Copy the trie
+  typename FeatureGroup<A>::Trie store_trie(trie_);
+  ret->trie_.swap(store_trie);
+
+  // Put the builder at error state to prevent repeated call of `Dump()`.
+  error_ = true;
+  return ret;
+}
+
+template <class A>
+int FeatureGroupBuilder<A>::FindFirstMatch(InputOutputLabel label, int parent,
+                                           int *hop) const {
+  int hop_count = 0;
+  int ret = kNoTrieNodeId;
+  for (; parent >= 0; parent = trie_[parent].back_link, ++hop_count) {
+    int next = trie_.Find(parent, label);
+    if (next != kNoTrieNodeId) {
+      ret = next;
+      break;
+    }
+  }
+  if (hop != nullptr) *hop = hop_count;
+  return ret;
+}
+
+template <class A>
+void FeatureGroupBuilder<A>::BuildBackLinks() {
+  // Breadth first search from the root. In the case where we only
+  // have the input label, the immedate back-off is simply the longest
+  // suffix of the current node that is also in the trie. For a node
+  // reached from its parent with label L, we can simply walk through
+  // the parent's back-off chain to find the first state with an arc
+  // of the same label L. The uniqueness is always
+  // guanranteed. However, in the case with both input and output
+  // labels, it is possible to back off by removing first labels from
+  // either side, which in general causes non-uniqueness.
+
+  const Topology &topology = trie_.TrieTopology();
+  std::queue<int> q;  // all enqueued or visited nodes have known links
+
+  // Note: nodes have back link initialized to -1 in their
+  // constructor.
+  q.push(trie_.Root());
+  while (!error_ && !q.empty()) {
+    int parent = q.front();
+    q.pop();
+    // Find links for every child
+    const typename Topology::NextMap &children = topology.ChildrenOf(parent);
+    for (typename Topology::NextMap::const_iterator eit = children.begin();
+         eit != children.end(); ++eit) {
+      const std::pair<InputOutputLabel, int> &edge = *eit;
+      InputOutputLabel label = edge.first;
+      int child = edge.second;
+      if (label.input == kNoLabel || label.output == kNoLabel) {
+        // Label pairs from root to here all have one and only one
+        // `kNoLabel` on the same side; equivalent to the
+        // "longest-suffix" case.
+        trie_[child].back_link =
+            FindFirstMatch(label, trie_[parent].back_link, nullptr);
+      } else {
+        // Neither side is `kNoLabel` at this point, there are
+        // three possible ways to back-off: if the parent backs
+        // off to some context with only one side non-empty, the
+        // empty side may remain empty; or else an exact match of
+        // both sides is needed. Try to find all three possible
+        // backs and look for the closest one (in terms of hops
+        // along the parent's back-off chain).
+        int only_input_hop, only_output_hop, full_hop;
+        int only_input_link =
+                FindFirstMatch(InputOutputLabel(label.input, kNoLabel), parent,
+                               &only_input_hop),
+            only_output_link =
+                FindFirstMatch(InputOutputLabel(kNoLabel, label.output), parent,
+                               &only_output_hop),
+            full_link =
+                FindFirstMatch(label, trie_[parent].back_link, &full_hop);
+        if (only_input_link != -1 && only_output_link != -1) {
+          error_ = true;
+          FSTERROR() << "Branching back-off chain:\n"
+                     << "\tnode " << child << ": " << TriePath(child, topology)
+                     << "\n"
+                     << "\tcan back-off to node " << only_input_link << ": "
+                     << TriePath(only_input_link, topology) << "\n"
+                     << "\tcan back-off to node " << only_output_link << ": "
+                     << TriePath(only_output_link, topology);
+          return;
+        } else if (full_link != -1) {
+          ++full_hop;
+          if (full_hop <= only_input_hop && full_hop <= only_output_hop) {
+            trie_[child].back_link = full_link;
+          } else {
+            error_ = true;
+            int problem_link = only_input_link != kNoTrieNodeId
+                                   ? only_input_link
+                                   : only_output_link;
+            CHECK_NE(problem_link, kNoTrieNodeId);
+            FSTERROR() << "Branching back-off chain:\n"
+                       << "\tnode " << child << ": "
+                       << TriePath(child, topology) << "\n"
+                       << "\tcan back-off to node " << full_link << ": "
+                       << TriePath(full_link, topology) << "\n"
+                       << "tcan back-off to node " << problem_link << ": "
+                       << TriePath(problem_link, topology);
+            return;
+          }
+        } else {
+          trie_[child].back_link =
+              only_input_link != -1 ? only_input_link : only_output_link;
+        }
+      }
+      if (error_) break;
+      // Point to empty context (root) when no back-off can be found
+      if (trie_[child].back_link == -1) trie_[child].back_link = 0;
+      q.push(child);
+    }
+  }
+}
+
+template <class A>
+void FeatureGroupBuilder<A>::PreAccumulateWeights() {
+  std::vector<bool> visited(trie_.NumNodes(), false);
+  visited[trie_.Root()] = true;
+
+  for (size_t i = 0; i != trie_.NumNodes(); ++i) {
+    std::stack<int> back_offs;
+    for (int j = i; !visited[j]; j = trie_[j].back_link) back_offs.push(j);
+    while (!back_offs.empty()) {
+      int j = back_offs.top();
+      back_offs.pop();
+      WeightBackLink &node = trie_[j];
+      node.weight = Times(node.weight, trie_[node.back_link].weight);
+      node.final_weight =
+          Times(node.final_weight, trie_[node.back_link].final_weight);
+      visited[j] = true;
+    }
+  }
+}
+
+template <class A>
+bool FeatureGroupBuilder<A>::TrieDfs(
+    const Topology &topology, int cur, int target,
+    std::vector<InputOutputLabel> *path) const {
+  if (cur == target) return true;
+  const typename Topology::NextMap &children = topology.ChildrenOf(cur);
+  for (typename Topology::NextMap::const_iterator eit = children.begin();
+       eit != children.end(); ++eit) {
+    const std::pair<InputOutputLabel, int> &edge = *eit;
+    path->push_back(edge.first);
+    if (TrieDfs(topology, edge.second, target, path)) return true;
+    path->pop_back();
+  }
+  return false;
+}
+
+template <class A>
+string FeatureGroupBuilder<A>::TriePath(int node,
+                                        const Topology &topology) const {
+  std::vector<InputOutputLabel> labels;
+  TrieDfs(topology, topology.Root(), node, &labels);
+  bool first = true;
+  std::ostringstream strm;
+  for (typename std::vector<InputOutputLabel>::const_iterator it =
+           labels.begin();
+       it != labels.end(); ++it) {
+    InputOutputLabel i = *it;
+    if (first)
+      first = false;
+    else
+      strm << ", ";
+    strm << "(" << TranslateLabel(i.input, fsyms_) << ", "
+         << TranslateLabel(i.output, osyms_) << ")";
+  }
+  return strm.str();
+}
+
+inline string TranslateLabel(int64 label, const SymbolTable *syms) {
+  string ret;
+  if (syms != nullptr) ret += syms->Find(label);
+  if (ret.empty()) {
+    std::ostringstream strm;
+    strm << '<' << label << '>';
+    ret = strm.str();
+  }
+  return ret;
+}
+
+template <class Iterator>
+string JoinLabels(Iterator begin, Iterator end, const SymbolTable *syms) {
+  if (begin == end) return "<empty>";
+  std::ostringstream strm;
+  bool first = true;
+  for (Iterator it = begin; it != end; ++it) {
+    if (first)
+      first = false;
+    else
+      strm << '|';
+    strm << TranslateLabel(*it, syms);
+  }
+  return strm.str();
+}
+
+template <class Label>
+string JoinLabels(const std::vector<Label> &labels, const SymbolTable *syms) {
+  return JoinLabels(labels.begin(), labels.end(), syms);
+}
+
+template <class A>
+typename A::Label GuessStartOrEnd(std::vector<typename A::Label> *sequence,
+                                  typename A::Label boundary) {
+  const size_t length = sequence->size();
+  std::vector<bool> non_boundary_on_left(length, false),
+      non_boundary_on_right(length, false);
+  for (size_t i = 1; i < length; ++i) {
+    non_boundary_on_left[i] =
+        non_boundary_on_left[i - 1] || (*sequence)[i - 1] != boundary;
+    non_boundary_on_right[length - 1 - i] = non_boundary_on_right[length - i] ||
+                                            (*sequence)[length - i] != boundary;
+  }
+  int unresolved = 0;
+  for (size_t i = 0; i < length; ++i) {
+    if ((*sequence)[i] != boundary) continue;
+    const bool left = non_boundary_on_left[i], right = non_boundary_on_right[i];
+    if (left && right) {
+      // Boundary in the middle
+      LOG(WARNING) << "Boundary label in the middle of the sequence! position: "
+                   << i << "; boundary: " << boundary
+                   << "; sequence: " << JoinLabels(*sequence, nullptr);
+      LOG(WARNING)
+          << "This is an invalid sequence anyway so I will set it to start.";
+      (*sequence)[i] = LinearFstData<A>::kStartOfSentence;
+    } else if (left && !right) {
+      // Can only be end
+      (*sequence)[i] = LinearFstData<A>::kEndOfSentence;
+    } else if (!left && right) {
+      // Can only be start
+      (*sequence)[i] = LinearFstData<A>::kStartOfSentence;
+    } else {
+      // !left && !right; can't really tell
+      ++unresolved;
+    }
+  }
+  return unresolved;
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_BUILDER_H_
diff --git a/projects/llm_framework/include/fst/extensions/linear/linear-fst-data.h b/projects/llm_framework/include/fst/extensions/linear/linear-fst-data.h
new file mode 100644
index 00000000..3b39c29d
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/linear/linear-fst-data.h
@@ -0,0 +1,526 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Data structures for storing and looking up the actual feature weights.
+
+#ifndef FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_H_
+#define FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_H_
+
+#include <memory>
+#include <numeric>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include <fst/compat.h>
+#include <fst/fst.h>
+
+#include <fst/extensions/linear/trie.h>
+
+namespace fst {
+
+// Forward declarations
+template <class A>
+class LinearFstDataBuilder;
+template <class A>
+class FeatureGroup;
+
+// Immutable data storage of the feature weights in a linear
+// model. Produces state tuples that represent internal states of a
+// LinearTaggerFst. Object of this class can only be constructed via
+// either `LinearFstDataBuilder::Dump()` or `LinearFstData::Read()`
+// and usually used as refcount'd object shared across mutiple
+// `LinearTaggerFst` copies.
+//
+// TODO(wuke): more efficient trie implementation
+template <class A>
+class LinearFstData {
+ public:
+  friend class LinearFstDataBuilder<A>;  // For builder access
+
+  typedef typename A::Label Label;
+  typedef typename A::Weight Weight;
+
+  // Sentence boundary labels. Both of them are negative labels other
+  // than `kNoLabel`.
+  static const Label kStartOfSentence;
+  static const Label kEndOfSentence;
+
+  // Constructs empty data; for non-trivial ways of construction see
+  // `Read()` and `LinearFstDataBuilder`.
+  LinearFstData()
+      : max_future_size_(0), max_input_label_(1), input_attribs_(1) {}
+
+  // Appends the state tuple of the start state to `output`, where
+  // each tuple holds the node ids of a trie for each feature group.
+  void EncodeStartState(std::vector<Label> *output) const {
+    for (int i = 0; i < NumGroups(); ++i) output->push_back(GroupStartState(i));
+  }
+
+  // Takes a transition from the trie states stored in
+  // `(trie_state_begin, trie_state_end)` with input label `ilabel`
+  // and output label `olabel`; appends the destination state tuple to
+  // `next` and multiplies the weight of the transition onto
+  // `weight`. `next` should be the shifted input buffer of the caller
+  // in `LinearTaggerFstImpl` (i.e. of size `LinearTaggerFstImpl::delay_`;
+  // the last element is `ilabel`).
+  template <class Iterator>
+  void TakeTransition(Iterator buffer_end, Iterator trie_state_begin,
+                      Iterator trie_state_end, Label ilabel, Label olabel,
+                      std::vector<Label> *next, Weight *weight) const;
+
+  // Returns the final weight of the given trie state sequence.
+  template <class Iterator>
+  Weight FinalWeight(Iterator trie_state_begin, Iterator trie_state_end) const;
+
+  // Returns the start trie state of the given group.
+  Label GroupStartState(int group_id) const {
+    return groups_[group_id]->Start();
+  }
+
+  // Takes a transition only within the given group. Returns the
+  // destination trie state and multiplies the weight onto `weight`.
+  Label GroupTransition(int group_id, int trie_state, Label ilabel,
+                        Label olabel, Weight *weight) const;
+
+  // Returns the final weight of the given trie state in the given group.
+  Weight GroupFinalWeight(int group_id, int trie_state) const {
+    return groups_[group_id]->FinalWeight(trie_state);
+  }
+
+  Label MinInputLabel() const { return 1; }
+
+  Label MaxInputLabel() const { return max_input_label_; }
+
+  // Returns the maximum future size of all feature groups. Future is
+  // the look-ahead window of a feature, e.g. if a feature looks at
+  // the next 2 words after the current input, then the future size is
+  // 2. There is no look-ahead for output. Features inside a single
+  // `FeatureGroup` must have equal future size.
+  size_t MaxFutureSize() const { return max_future_size_; }
+
+  // Returns the number of feature groups
+  size_t NumGroups() const { return groups_.size(); }
+
+  // Returns the range of possible output labels for an input label.
+  std::pair<typename std::vector<Label>::const_iterator,
+            typename std::vector<Label>::const_iterator>
+  PossibleOutputLabels(Label word) const;
+
+  static LinearFstData<A> *Read(std::istream &strm);  // NOLINT
+  std::ostream &Write(std::ostream &strm) const;      // NOLINT
+
+ private:
+  // Offsets in `output_pool_`
+  struct InputAttribute {
+    size_t output_begin, output_length;
+
+    std::istream &Read(std::istream &strm);         // NOLINT
+    std::ostream &Write(std::ostream &strm) const;  // NOLINT
+  };
+
+  // Mapping from input label to per-group feature label
+  class GroupFeatureMap;
+
+  // Translates the input label into input feature label of group
+  // `group`; returns `kNoLabel` when there is no feature for that
+  // group.
+  Label FindFeature(size_t group, Label word) const;
+
+  size_t max_future_size_;
+  Label max_input_label_;
+  std::vector<std::unique_ptr<const FeatureGroup<A>>> groups_;
+  std::vector<InputAttribute> input_attribs_;
+  std::vector<Label> output_pool_, output_set_;
+  GroupFeatureMap group_feat_map_;
+
+  LinearFstData(const LinearFstData &) = delete;
+  LinearFstData &operator=(const LinearFstData &) = delete;
+};
+
+template <class A>
+const typename A::Label LinearFstData<A>::kStartOfSentence = -3;
+template <class A>
+const typename A::Label LinearFstData<A>::kEndOfSentence = -2;
+
+template <class A>
+template <class Iterator>
+void LinearFstData<A>::TakeTransition(Iterator buffer_end,
+                                      Iterator trie_state_begin,
+                                      Iterator trie_state_end, Label ilabel,
+                                      Label olabel, std::vector<Label> *next,
+                                      Weight *weight) const {
+  DCHECK_EQ(trie_state_end - trie_state_begin, groups_.size());
+  DCHECK(ilabel > 0 || ilabel == kEndOfSentence);
+  DCHECK(olabel > 0 || olabel == kStartOfSentence);
+  size_t group_id = 0;
+  for (Iterator it = trie_state_begin; it != trie_state_end; ++it, ++group_id) {
+    size_t delay = groups_[group_id]->Delay();
+    // On the buffer, there may also be `kStartOfSentence` from the
+    // initial empty buffer.
+    Label real_ilabel = delay == 0 ? ilabel : *(buffer_end - delay);
+    next->push_back(
+        GroupTransition(group_id, *it, real_ilabel, olabel, weight));
+  }
+}
+
+template <class A>
+typename A::Label LinearFstData<A>::GroupTransition(int group_id,
+                                                    int trie_state,
+                                                    Label ilabel, Label olabel,
+                                                    Weight *weight) const {
+  Label group_ilabel = FindFeature(group_id, ilabel);
+  return groups_[group_id]->Walk(trie_state, group_ilabel, olabel, weight);
+}
+
+template <class A>
+template <class Iterator>
+inline typename A::Weight LinearFstData<A>::FinalWeight(
+    Iterator trie_state_begin, Iterator trie_state_end) const {
+  DCHECK_EQ(trie_state_end - trie_state_begin, groups_.size());
+  size_t group_id = 0;
+  Weight accum = Weight::One();
+  for (Iterator it = trie_state_begin; it != trie_state_end; ++it, ++group_id)
+    accum = Times(accum, GroupFinalWeight(group_id, *it));
+  return accum;
+}
+
+template <class A>
+inline std::pair<typename std::vector<typename A::Label>::const_iterator,
+                 typename std::vector<typename A::Label>::const_iterator>
+LinearFstData<A>::PossibleOutputLabels(Label word) const {
+  const InputAttribute &attrib = input_attribs_[word];
+  if (attrib.output_length == 0)
+    return std::make_pair(output_set_.begin(), output_set_.end());
+  else
+    return std::make_pair(
+        output_pool_.begin() + attrib.output_begin,
+        output_pool_.begin() + attrib.output_begin + attrib.output_length);
+}
+
+template <class A>
+inline LinearFstData<A> *LinearFstData<A>::Read(std::istream &strm) {  // NOLINT
+  std::unique_ptr<LinearFstData<A>> data(new LinearFstData<A>());
+  ReadType(strm, &(data->max_future_size_));
+  ReadType(strm, &(data->max_input_label_));
+  // Feature groups
+  size_t num_groups = 0;
+  ReadType(strm, &num_groups);
+  data->groups_.resize(num_groups);
+  for (size_t i = 0; i < num_groups; ++i)
+    data->groups_[i].reset(FeatureGroup<A>::Read(strm));
+  // Other data
+  ReadType(strm, &(data->input_attribs_));
+  ReadType(strm, &(data->output_pool_));
+  ReadType(strm, &(data->output_set_));
+  ReadType(strm, &(data->group_feat_map_));
+  if (strm) {
+    return data.release();
+  } else {
+    return nullptr;
+  }
+}
+
+template <class A>
+inline std::ostream &LinearFstData<A>::Write(
+    std::ostream &strm) const {  // NOLINT
+  WriteType(strm, max_future_size_);
+  WriteType(strm, max_input_label_);
+  // Feature groups
+  WriteType(strm, groups_.size());
+  for (size_t i = 0; i < groups_.size(); ++i) {
+    groups_[i]->Write(strm);
+  }
+  // Other data
+  WriteType(strm, input_attribs_);
+  WriteType(strm, output_pool_);
+  WriteType(strm, output_set_);
+  WriteType(strm, group_feat_map_);
+  return strm;
+}
+
+template <class A>
+typename A::Label LinearFstData<A>::FindFeature(size_t group,
+                                                Label word) const {
+  DCHECK(word > 0 || word == kStartOfSentence || word == kEndOfSentence);
+  if (word == kStartOfSentence || word == kEndOfSentence)
+    return word;
+  else
+    return group_feat_map_.Find(group, word);
+}
+
+template <class A>
+inline std::istream &LinearFstData<A>::InputAttribute::Read(
+    std::istream &strm) {  // NOLINT
+  ReadType(strm, &output_begin);
+  ReadType(strm, &output_length);
+  return strm;
+}
+
+template <class A>
+inline std::ostream &LinearFstData<A>::InputAttribute::Write(
+    std::ostream &strm) const {  // NOLINT
+  WriteType(strm, output_begin);
+  WriteType(strm, output_length);
+  return strm;
+}
+
+// Forward declaration
+template <class A>
+class FeatureGroupBuilder;
+
+// An immutable grouping of features with similar context shape. Like
+// `LinearFstData`, this can only be constructed via `Read()` or
+// via its builder.
+//
+// Internally it uses a trie to store all feature n-grams and their
+// weights. The label of a trie edge is a pair (feat, olabel) of
+// labels. They can be either positive (ordinary label), `kNoLabel`,
+// `kStartOfSentence`, or `kEndOfSentence`. `kNoLabel` usually means
+// matching anything, with one exception: from the root of the trie,
+// there is a special (kNoLabel, kNoLabel) that leads to the implicit
+// start-of-sentence state. This edge is never actually matched
+// (`FindFirstMatch()` ensures this).
+template <class A>
+class FeatureGroup {
+ public:
+  friend class FeatureGroupBuilder<A>;  // for builder access
+
+  typedef typename A::Label Label;
+  typedef typename A::Weight Weight;
+
+  int Start() const { return start_; }
+
+  // Finds destination node from `cur` by consuming `ilabel` and
+  // `olabel`. The transition weight is multiplied onto `weight`.
+  int Walk(int cur, Label ilabel, Label olabel, Weight *weight) const;
+
+  // Returns the final weight of the current trie state. Only valid if
+  // the state is already known to be part of a final state (see
+  // `LinearFstData<>::CanBeFinal()`).
+  Weight FinalWeight(int trie_state) const {
+    return trie_[trie_state].final_weight;
+  }
+
+  static FeatureGroup<A> *Read(std::istream &strm) {  // NOLINT
+    size_t delay;
+    ReadType(strm, &delay);
+    int start;
+    ReadType(strm, &start);
+    Trie trie;
+    ReadType(strm, &trie);
+    std::unique_ptr<FeatureGroup<A>> ret(new FeatureGroup<A>(delay, start));
+    ret->trie_.swap(trie);
+    ReadType(strm, &ret->next_state_);
+    if (strm) {
+      return ret.release();
+    } else {
+      return nullptr;
+    }
+  }
+
+  std::ostream &Write(std::ostream &strm) const {  // NOLINT
+    WriteType(strm, delay_);
+    WriteType(strm, start_);
+    WriteType(strm, trie_);
+    WriteType(strm, next_state_);
+    return strm;
+  }
+
+  size_t Delay() const { return delay_; }
+
+  string Stats() const;
+
+ private:
+  // Label along the arcs on the trie. `kNoLabel` means anything
+  // (non-negative label) can match; both sides holding `kNoLabel`
+  // is not allow; otherwise the label is > 0 (enforced by
+  // `LinearFstDataBuilder::AddWeight()`).
+  struct InputOutputLabel;
+  struct InputOutputLabelHash;
+
+  // Data to be stored on the trie
+  struct WeightBackLink {
+    int back_link;
+    Weight weight, final_weight;
+
+    WeightBackLink()
+        : back_link(kNoTrieNodeId),
+          weight(Weight::One()),
+          final_weight(Weight::One()) {}
+
+    std::istream &Read(std::istream &strm) {  // NOLINT
+      ReadType(strm, &back_link);
+      ReadType(strm, &weight);
+      ReadType(strm, &final_weight);
+      return strm;
+    }
+
+    std::ostream &Write(std::ostream &strm) const {  // NOLINT
+      WriteType(strm, back_link);
+      WriteType(strm, weight);
+      WriteType(strm, final_weight);
+      return strm;
+    }
+  };
+
+  typedef FlatTrieTopology<InputOutputLabel, InputOutputLabelHash> Topology;
+  typedef MutableTrie<InputOutputLabel, WeightBackLink, Topology> Trie;
+
+  explicit FeatureGroup(size_t delay, int start)
+      : delay_(delay), start_(start) {}
+
+  // Finds the first node with an arc with `label` following the
+  // back-off chain of `parent`. Returns the node index or
+  // `kNoTrieNodeId` when not found.
+  int FindFirstMatch(InputOutputLabel label, int parent) const;
+
+  size_t delay_;
+  int start_;
+  Trie trie_;
+  // Where to go after hitting this state. When we reach a state with
+  // no child and with no additional final weight (i.e. its final
+  // weight is the same as its back-off), we can immediately go to its
+  // back-off state.
+  std::vector<int> next_state_;
+
+  FeatureGroup(const FeatureGroup &) = delete;
+  FeatureGroup &operator=(const FeatureGroup &) = delete;
+};
+
+template <class A>
+struct FeatureGroup<A>::InputOutputLabel {
+  Label input, output;
+
+  InputOutputLabel(Label i = kNoLabel, Label o = kNoLabel)
+      : input(i), output(o) {}
+
+  bool operator==(InputOutputLabel that) const {
+    return input == that.input && output == that.output;
+  }
+
+  std::istream &Read(std::istream &strm) {  // NOLINT
+    ReadType(strm, &input);
+    ReadType(strm, &output);
+    return strm;
+  }
+
+  std::ostream &Write(std::ostream &strm) const {  // NOLINT
+    WriteType(strm, input);
+    WriteType(strm, output);
+    return strm;
+  }
+};
+
+template <class A>
+struct FeatureGroup<A>::InputOutputLabelHash {
+  size_t operator()(InputOutputLabel label) const {
+    return static_cast<size_t>(label.input * 7853 + label.output);
+  }
+};
+
+template <class A>
+int FeatureGroup<A>::Walk(int cur, Label ilabel, Label olabel,
+                          Weight *weight) const {
+  // Note: user of this method need to ensure `ilabel` and `olabel`
+  // are valid (e.g. see DCHECKs in
+  // `LinearFstData<>::TakeTransition()` and
+  // `LinearFstData<>::FindFeature()`).
+  int next;
+  if (ilabel == LinearFstData<A>::kStartOfSentence) {
+    // An observed start-of-sentence only occurs in the beginning of
+    // the input, when this feature group is delayed (i.e. there is
+    // another feature group with a larger future size). The actual
+    // input hasn't arrived so stay at the start state.
+    DCHECK_EQ(cur, start_);
+    next = start_;
+  } else {
+    // First, try exact match
+    next = FindFirstMatch(InputOutputLabel(ilabel, olabel), cur);
+    // Then try with don't cares
+    if (next == kNoTrieNodeId)
+      next = FindFirstMatch(InputOutputLabel(ilabel, kNoLabel), cur);
+    if (next == kNoTrieNodeId)
+      next = FindFirstMatch(InputOutputLabel(kNoLabel, olabel), cur);
+    // All failed, go to empty context
+    if (next == kNoTrieNodeId) next = trie_.Root();
+    *weight = Times(*weight, trie_[next].weight);
+    next = next_state_[next];
+  }
+  return next;
+}
+
+template <class A>
+inline int FeatureGroup<A>::FindFirstMatch(InputOutputLabel label,
+                                           int parent) const {
+  if (label.input == kNoLabel && label.output == kNoLabel)
+    return kNoTrieNodeId;  // very important; see class doc.
+  for (; parent != kNoTrieNodeId; parent = trie_[parent].back_link) {
+    int next = trie_.Find(parent, label);
+    if (next != kNoTrieNodeId) return next;
+  }
+  return kNoTrieNodeId;
+}
+
+template <class A>
+inline string FeatureGroup<A>::Stats() const {
+  std::ostringstream strm;
+  int num_states = 2;
+  for (int i = 2; i < next_state_.size(); ++i)
+    num_states += i == next_state_[i];
+  strm << trie_.NumNodes() << " node(s); " << num_states << " state(s)";
+  return strm.str();
+}
+
+template <class A>
+class LinearFstData<A>::GroupFeatureMap {
+ public:
+  GroupFeatureMap() {}
+
+  void Init(size_t num_groups, size_t num_words) {
+    num_groups_ = num_groups;
+    pool_.clear();
+    pool_.resize(num_groups * num_words, kNoLabel);
+  }
+
+  Label Find(size_t group_id, Label ilabel) const {
+    return pool_[IndexOf(group_id, ilabel)];
+  }
+
+  bool Set(size_t group_id, Label ilabel, Label feat) {
+    size_t i = IndexOf(group_id, ilabel);
+    if (pool_[i] != kNoLabel && pool_[i] != feat) {
+      FSTERROR() << "Feature group " << group_id
+                 << " already has feature for word " << ilabel;
+      return false;
+    }
+    pool_[i] = feat;
+    return true;
+  }
+
+  std::istream &Read(std::istream &strm) {  // NOLINT
+    ReadType(strm, &num_groups_);
+    ReadType(strm, &pool_);
+    return strm;
+  }
+
+  std::ostream &Write(std::ostream &strm) const {  // NOLINT
+    WriteType(strm, num_groups_);
+    WriteType(strm, pool_);
+    return strm;
+  }
+
+ private:
+  size_t IndexOf(size_t group_id, Label ilabel) const {
+    return ilabel * num_groups_ + group_id;
+  }
+
+  size_t num_groups_;
+  // `pool_[ilabel * num_groups_ + group_id]` is the feature active
+  // for group `group_id` with input `ilabel`
+  std::vector<Label> pool_;
+};
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_H_
diff --git a/projects/llm_framework/include/fst/extensions/linear/linear-fst.h b/projects/llm_framework/include/fst/extensions/linear/linear-fst.h
new file mode 100644
index 00000000..e4239a87
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/linear/linear-fst.h
@@ -0,0 +1,1173 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Classes for building, storing and representing log-linear models as FSTs.
+
+#ifndef FST_EXTENSIONS_LINEAR_LINEAR_FST_H_
+#define FST_EXTENSIONS_LINEAR_LINEAR_FST_H_
+
+#include <algorithm>
+#include <iostream>
+#include <memory>
+#include <vector>
+
+#include <fst/compat.h>
+#include <fst/log.h>
+#include <fst/extensions/pdt/collection.h>
+#include <fst/bi-table.h>
+#include <fst/cache.h>
+#include <fstream>
+#include <fst/fst.h>
+#include <fst/matcher.h>
+#include <fst/symbol-table.h>
+
+#include <fst/extensions/linear/linear-fst-data.h>
+
+namespace fst {
+
+// Forward declaration of the specialized matcher for both
+// LinearTaggerFst and LinearClassifierFst.
+template <class F>
+class LinearFstMatcherTpl;
+
+namespace internal {
+
+// Implementation class for on-the-fly generated LinearTaggerFst with
+// special optimization in matching.
+template <class A>
+class LinearTaggerFstImpl : public CacheImpl<A> {
+ public:
+  using FstImpl<A>::SetType;
+  using FstImpl<A>::SetProperties;
+  using FstImpl<A>::SetInputSymbols;
+  using FstImpl<A>::SetOutputSymbols;
+  using FstImpl<A>::WriteHeader;
+
+  using CacheBaseImpl<CacheState<A>>::PushArc;
+  using CacheBaseImpl<CacheState<A>>::HasArcs;
+  using CacheBaseImpl<CacheState<A>>::HasFinal;
+  using CacheBaseImpl<CacheState<A>>::HasStart;
+  using CacheBaseImpl<CacheState<A>>::SetArcs;
+  using CacheBaseImpl<CacheState<A>>::SetFinal;
+  using CacheBaseImpl<CacheState<A>>::SetStart;
+
+  typedef A Arc;
+  typedef typename A::Label Label;
+  typedef typename A::Weight Weight;
+  typedef typename A::StateId StateId;
+  typedef typename Collection<StateId, Label>::SetIterator NGramIterator;
+
+  // Constructs an empty FST by default.
+  LinearTaggerFstImpl()
+      : CacheImpl<A>(CacheOptions()),
+        data_(std::make_shared<LinearFstData<A>>()),
+        delay_(0) {
+    SetType("linear-tagger");
+  }
+
+  // Constructs the FST with given data storage and symbol
+  // tables.
+  //
+  // TODO(wuke): when there is no constraint on output we can delay
+  // less than `data->MaxFutureSize` positions.
+  LinearTaggerFstImpl(const LinearFstData<Arc> *data, const SymbolTable *isyms,
+                      const SymbolTable *osyms, CacheOptions opts)
+      : CacheImpl<A>(opts), data_(data), delay_(data->MaxFutureSize()) {
+    SetType("linear-tagger");
+    SetProperties(kILabelSorted, kFstProperties);
+    SetInputSymbols(isyms);
+    SetOutputSymbols(osyms);
+    ReserveStubSpace();
+  }
+
+  // Copy by sharing the underlying data storage.
+  LinearTaggerFstImpl(const LinearTaggerFstImpl &impl)
+      : CacheImpl<A>(impl), data_(impl.data_), delay_(impl.delay_) {
+    SetType("linear-tagger");
+    SetProperties(impl.Properties(), kCopyProperties);
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+    ReserveStubSpace();
+  }
+
+  StateId Start() {
+    if (!HasStart()) {
+      StateId start = FindStartState();
+      SetStart(start);
+    }
+    return CacheImpl<A>::Start();
+  }
+
+  Weight Final(StateId s) {
+    if (!HasFinal(s)) {
+      state_stub_.clear();
+      FillState(s, &state_stub_);
+      if (CanBeFinal(state_stub_))
+        SetFinal(s, data_->FinalWeight(InternalBegin(state_stub_),
+                                       InternalEnd(state_stub_)));
+      else
+        SetFinal(s, Weight::Zero());
+    }
+    return CacheImpl<A>::Final(s);
+  }
+
+  size_t NumArcs(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<A>::NumArcs(s);
+  }
+
+  size_t NumInputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<A>::NumInputEpsilons(s);
+  }
+
+  size_t NumOutputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<A>::NumOutputEpsilons(s);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<A> *data) {
+    if (!HasArcs(s)) Expand(s);
+    CacheImpl<A>::InitArcIterator(s, data);
+  }
+
+  // Computes the outgoing transitions from a state, creating new
+  // destination states as needed.
+  void Expand(StateId s);
+
+  // Appends to `arcs` all out-going arcs from state `s` that matches `label` as
+  // the input label.
+  void MatchInput(StateId s, Label ilabel, std::vector<Arc> *arcs);
+
+  static LinearTaggerFstImpl *Read(std::istream &strm,
+                                   const FstReadOptions &opts);
+
+  bool Write(std::ostream &strm,  // NOLINT
+             const FstWriteOptions &opts) const {
+    FstHeader header;
+    header.SetStart(kNoStateId);
+    WriteHeader(strm, opts, kFileVersion, &header);
+    data_->Write(strm);
+    if (!strm) {
+      LOG(ERROR) << "LinearTaggerFst::Write: Write failed: " << opts.source;
+      return false;
+    }
+    return true;
+  }
+
+ private:
+  static const int kMinFileVersion;
+  static const int kFileVersion;
+
+  // A collection of functions to access parts of the state tuple. A
+  // state tuple is a vector of `Label`s with two parts:
+  //   [buffer] [internal].
+  //
+  // - [buffer] is a buffer of observed input labels with length
+  // `delay_`. `LinearFstData<A>::kStartOfSentence`
+  // (resp. `LinearFstData<A>::kEndOfSentence`) are used as
+  // paddings when the buffer has fewer than `delay_` elements, which
+  // can only appear as the prefix (resp. suffix) of the buffer.
+  //
+  // - [internal] is the internal state tuple for `LinearFstData`
+  typename std::vector<Label>::const_iterator BufferBegin(
+      const std::vector<Label> &state) const {
+    return state.begin();
+  }
+
+  typename std::vector<Label>::const_iterator BufferEnd(
+      const std::vector<Label> &state) const {
+    return state.begin() + delay_;
+  }
+
+  typename std::vector<Label>::const_iterator InternalBegin(
+      const std::vector<Label> &state) const {
+    return state.begin() + delay_;
+  }
+
+  typename std::vector<Label>::const_iterator InternalEnd(
+      const std::vector<Label> &state) const {
+    return state.end();
+  }
+
+  // The size of state tuples are fixed, reserve them in stubs
+  void ReserveStubSpace() {
+    state_stub_.reserve(delay_ + data_->NumGroups());
+    next_stub_.reserve(delay_ + data_->NumGroups());
+  }
+
+  // Computes the start state tuple and maps it to the start state id.
+  StateId FindStartState() {
+    // Empty buffer with start-of-sentence paddings
+    state_stub_.clear();
+    state_stub_.resize(delay_, LinearFstData<A>::kStartOfSentence);
+    // Append internal states
+    data_->EncodeStartState(&state_stub_);
+    return FindState(state_stub_);
+  }
+
+  // Tests whether the buffer in `(begin, end)` is empty.
+  bool IsEmptyBuffer(typename std::vector<Label>::const_iterator begin,
+                     typename std::vector<Label>::const_iterator end) const {
+    // The following is guanranteed by `ShiftBuffer()`:
+    // - buffer[i] == LinearFstData<A>::kEndOfSentence =>
+    //       buffer[i+x] == LinearFstData<A>::kEndOfSentence
+    // - buffer[i] == LinearFstData<A>::kStartOfSentence =>
+    //       buffer[i-x] == LinearFstData<A>::kStartOfSentence
+    return delay_ == 0 || *(end - 1) == LinearFstData<A>::kStartOfSentence ||
+           *begin == LinearFstData<A>::kEndOfSentence;
+  }
+
+  // Tests whether the given state tuple can be a final state. A state
+  // is final iff there is no observed input in the buffer.
+  bool CanBeFinal(const std::vector<Label> &state) {
+    return IsEmptyBuffer(BufferBegin(state), BufferEnd(state));
+  }
+
+  // Finds state corresponding to an n-gram. Creates new state if n-gram not
+  // found.
+  StateId FindState(const std::vector<Label> &ngram) {
+    StateId sparse = ngrams_.FindId(ngram, true);
+    StateId dense = condensed_.FindId(sparse, true);
+    return dense;
+  }
+
+  // Appends after `output` the state tuple corresponding to the state id. The
+  // state id must exist.
+  void FillState(StateId s, std::vector<Label> *output) {
+    s = condensed_.FindEntry(s);
+    for (NGramIterator it = ngrams_.FindSet(s); !it.Done(); it.Next()) {
+      Label label = it.Element();
+      output->push_back(label);
+    }
+  }
+
+  // Shifts the buffer in `state` by appending `ilabel` and popping
+  // the one in the front as the return value. `next_stub_` is a
+  // shifted buffer of size `delay_` where the first `delay_ - 1`
+  // elements are the last `delay_ - 1` elements in the buffer of
+  // `state`. The last (if any) element in `next_stub_` will be
+  // `ilabel` after the call returns.
+  Label ShiftBuffer(const std::vector<Label> &state, Label ilabel,
+                    std::vector<Label> *next_stub_);
+
+  // Builds an arc from state tuple `state` consuming `ilabel` and
+  // `olabel`. `next_stub_` is the buffer filled in `ShiftBuffer`.
+  Arc MakeArc(const std::vector<Label> &state, Label ilabel, Label olabel,
+              std::vector<Label> *next_stub_);
+
+  // Expands arcs from state `s`, equivalent to state tuple `state`,
+  // with input `ilabel`. `next_stub_` is the buffer filled in
+  // `ShiftBuffer`.
+  void ExpandArcs(StateId s, const std::vector<Label> &state, Label ilabel,
+                  std::vector<Label> *next_stub_);
+
+  // Appends arcs from state `s`, equivalent to state tuple `state`,
+  // with input `ilabel` to `arcs`. `next_stub_` is the buffer filled
+  // in `ShiftBuffer`.
+  void AppendArcs(StateId s, const std::vector<Label> &state, Label ilabel,
+                  std::vector<Label> *next_stub_, std::vector<Arc> *arcs);
+
+  std::shared_ptr<const LinearFstData<A>> data_;
+  size_t delay_;
+  // Mapping from internal state tuple to *non-consecutive* ids
+  Collection<StateId, Label> ngrams_;
+  // Mapping from non-consecutive id to actual state id
+  CompactHashBiTable<StateId, StateId, std::hash<StateId>> condensed_;
+  // Two frequently used vectors, reuse to avoid repeated heap
+  // allocation
+  std::vector<Label> state_stub_, next_stub_;
+
+  LinearTaggerFstImpl &operator=(const LinearTaggerFstImpl &) = delete;
+};
+
+template <class A>
+const int LinearTaggerFstImpl<A>::kMinFileVersion = 1;
+
+template <class A>
+const int LinearTaggerFstImpl<A>::kFileVersion = 1;
+
+template <class A>
+inline typename A::Label LinearTaggerFstImpl<A>::ShiftBuffer(
+    const std::vector<Label> &state, Label ilabel,
+    std::vector<Label> *next_stub_) {
+  DCHECK(ilabel > 0 || ilabel == LinearFstData<A>::kEndOfSentence);
+  if (delay_ == 0) {
+    DCHECK_GT(ilabel, 0);
+    return ilabel;
+  } else {
+    (*next_stub_)[BufferEnd(*next_stub_) - next_stub_->begin() - 1] = ilabel;
+    return *BufferBegin(state);
+  }
+}
+
+template <class A>
+inline A LinearTaggerFstImpl<A>::MakeArc(const std::vector<Label> &state,
+                                         Label ilabel, Label olabel,
+                                         std::vector<Label> *next_stub_) {
+  DCHECK(ilabel > 0 || ilabel == LinearFstData<A>::kEndOfSentence);
+  DCHECK(olabel > 0 || olabel == LinearFstData<A>::kStartOfSentence);
+  Weight weight(Weight::One());
+  data_->TakeTransition(BufferEnd(state), InternalBegin(state),
+                        InternalEnd(state), ilabel, olabel, next_stub_,
+                        &weight);
+  StateId nextstate = FindState(*next_stub_);
+  // Restore `next_stub_` to its size before the call
+  next_stub_->resize(delay_);
+  // In the actual arc, we use epsilons instead of boundaries.
+  return A(ilabel == LinearFstData<A>::kEndOfSentence ? 0 : ilabel,
+           olabel == LinearFstData<A>::kStartOfSentence ? 0 : olabel, weight,
+           nextstate);
+}
+
+template <class A>
+inline void LinearTaggerFstImpl<A>::ExpandArcs(StateId s,
+                                               const std::vector<Label> &state,
+                                               Label ilabel,
+                                               std::vector<Label> *next_stub_) {
+  // Input label to constrain the output with, observed `delay_` steps
+  // back. `ilabel` is the input label to be put on the arc, which
+  // fires features.
+  Label obs_ilabel = ShiftBuffer(state, ilabel, next_stub_);
+  if (obs_ilabel == LinearFstData<A>::kStartOfSentence) {
+    // This happens when input is shorter than `delay_`.
+    PushArc(s, MakeArc(state, ilabel, LinearFstData<A>::kStartOfSentence,
+                       next_stub_));
+  } else {
+    std::pair<typename std::vector<typename A::Label>::const_iterator,
+              typename std::vector<typename A::Label>::const_iterator> range =
+        data_->PossibleOutputLabels(obs_ilabel);
+    for (typename std::vector<typename A::Label>::const_iterator it =
+             range.first;
+         it != range.second; ++it)
+      PushArc(s, MakeArc(state, ilabel, *it, next_stub_));
+  }
+}
+
+// TODO(wuke): this has much in duplicate with `ExpandArcs()`
+template <class A>
+inline void LinearTaggerFstImpl<A>::AppendArcs(StateId /*s*/,
+                                               const std::vector<Label> &state,
+                                               Label ilabel,
+                                               std::vector<Label> *next_stub_,
+                                               std::vector<Arc> *arcs) {
+  // Input label to constrain the output with, observed `delay_` steps
+  // back. `ilabel` is the input label to be put on the arc, which
+  // fires features.
+  Label obs_ilabel = ShiftBuffer(state, ilabel, next_stub_);
+  if (obs_ilabel == LinearFstData<A>::kStartOfSentence) {
+    // This happens when input is shorter than `delay_`.
+    arcs->push_back(
+        MakeArc(state, ilabel, LinearFstData<A>::kStartOfSentence, next_stub_));
+  } else {
+    std::pair<typename std::vector<typename A::Label>::const_iterator,
+              typename std::vector<typename A::Label>::const_iterator> range =
+        data_->PossibleOutputLabels(obs_ilabel);
+    for (typename std::vector<typename A::Label>::const_iterator it =
+             range.first;
+         it != range.second; ++it)
+      arcs->push_back(MakeArc(state, ilabel, *it, next_stub_));
+  }
+}
+
+template <class A>
+void LinearTaggerFstImpl<A>::Expand(StateId s) {
+  VLOG(3) << "Expand " << s;
+  state_stub_.clear();
+  FillState(s, &state_stub_);
+
+  // Precompute the first `delay_ - 1` elements in the buffer of
+  // next states, which are identical for different input/output.
+  next_stub_.clear();
+  next_stub_.resize(delay_);
+  if (delay_ > 0)
+    std::copy(BufferBegin(state_stub_) + 1, BufferEnd(state_stub_),
+              next_stub_.begin());
+
+  // Epsilon transition for flushing out the next observed input
+  if (!IsEmptyBuffer(BufferBegin(state_stub_), BufferEnd(state_stub_)))
+    ExpandArcs(s, state_stub_, LinearFstData<A>::kEndOfSentence, &next_stub_);
+
+  // Non-epsilon input when we haven't flushed
+  if (delay_ == 0 ||
+      *(BufferEnd(state_stub_) - 1) != LinearFstData<A>::kEndOfSentence)
+    for (Label ilabel = data_->MinInputLabel();
+         ilabel <= data_->MaxInputLabel(); ++ilabel)
+      ExpandArcs(s, state_stub_, ilabel, &next_stub_);
+
+  SetArcs(s);
+}
+
+template <class A>
+void LinearTaggerFstImpl<A>::MatchInput(StateId s, Label ilabel,
+                                        std::vector<Arc> *arcs) {
+  state_stub_.clear();
+  FillState(s, &state_stub_);
+
+  // Precompute the first `delay_ - 1` elements in the buffer of
+  // next states, which are identical for different input/output.
+  next_stub_.clear();
+  next_stub_.resize(delay_);
+  if (delay_ > 0)
+    std::copy(BufferBegin(state_stub_) + 1, BufferEnd(state_stub_),
+              next_stub_.begin());
+
+  if (ilabel == 0) {
+    // Epsilon transition for flushing out the next observed input
+    if (!IsEmptyBuffer(BufferBegin(state_stub_), BufferEnd(state_stub_)))
+      AppendArcs(s, state_stub_, LinearFstData<A>::kEndOfSentence, &next_stub_,
+                 arcs);
+  } else {
+    // Non-epsilon input when we haven't flushed
+    if (delay_ == 0 ||
+        *(BufferEnd(state_stub_) - 1) != LinearFstData<A>::kEndOfSentence)
+      AppendArcs(s, state_stub_, ilabel, &next_stub_, arcs);
+  }
+}
+
+template <class A>
+inline LinearTaggerFstImpl<A> *LinearTaggerFstImpl<A>::Read(
+    std::istream &strm, const FstReadOptions &opts) {  // NOLINT
+  std::unique_ptr<LinearTaggerFstImpl<A>> impl(new LinearTaggerFstImpl<A>());
+  FstHeader header;
+  if (!impl->ReadHeader(strm, opts, kMinFileVersion, &header)) {
+    return nullptr;
+  }
+  impl->data_ = std::shared_ptr<LinearFstData<A>>(LinearFstData<A>::Read(strm));
+  if (!impl->data_) {
+    return nullptr;
+  }
+  impl->delay_ = impl->data_->MaxFutureSize();
+  impl->ReserveStubSpace();
+  return impl.release();
+}
+
+}  // namespace internal
+
+// This class attaches interface to implementation and handles
+// reference counting, delegating most methods to ImplToFst.
+template <class A>
+class LinearTaggerFst : public ImplToFst<internal::LinearTaggerFstImpl<A>> {
+ public:
+  friend class ArcIterator<LinearTaggerFst<A>>;
+  friend class StateIterator<LinearTaggerFst<A>>;
+  friend class LinearFstMatcherTpl<LinearTaggerFst<A>>;
+
+  typedef A Arc;
+  typedef typename A::Label Label;
+  typedef typename A::Weight Weight;
+  typedef typename A::StateId StateId;
+  typedef DefaultCacheStore<A> Store;
+  typedef typename Store::State State;
+  using Impl = internal::LinearTaggerFstImpl<A>;
+
+  LinearTaggerFst() : ImplToFst<Impl>(std::make_shared<Impl>()) {}
+
+  explicit LinearTaggerFst(LinearFstData<A> *data,
+                           const SymbolTable *isyms = nullptr,
+                           const SymbolTable *osyms = nullptr,
+                           CacheOptions opts = CacheOptions())
+      : ImplToFst<Impl>(std::make_shared<Impl>(data, isyms, osyms, opts)) {}
+
+  explicit LinearTaggerFst(const Fst<A> &fst)
+      : ImplToFst<Impl>(std::make_shared<Impl>()) {
+    LOG(FATAL) << "LinearTaggerFst: no constructor from arbitrary FST.";
+  }
+
+  // See Fst<>::Copy() for doc.
+  LinearTaggerFst(const LinearTaggerFst<A> &fst, bool safe = false)
+      : ImplToFst<Impl>(fst, safe) {}
+
+  // Get a copy of this LinearTaggerFst. See Fst<>::Copy() for further doc.
+  LinearTaggerFst<A> *Copy(bool safe = false) const override {
+    return new LinearTaggerFst<A>(*this, safe);
+  }
+
+  inline void InitStateIterator(StateIteratorData<A> *data) const override;
+
+  void InitArcIterator(StateId s, ArcIteratorData<A> *data) const override {
+    GetMutableImpl()->InitArcIterator(s, data);
+  }
+
+  MatcherBase<A> *InitMatcher(MatchType match_type) const override {
+    return new LinearFstMatcherTpl<LinearTaggerFst<A>>(this, match_type);
+  }
+
+  static LinearTaggerFst<A> *Read(const string &filename) {
+    if (!filename.empty()) {
+      std::ifstream strm(filename,
+                              std::ios_base::in | std::ios_base::binary);
+      if (!strm) {
+        LOG(ERROR) << "LinearTaggerFst::Read: Can't open file: " << filename;
+        return nullptr;
+      }
+      return Read(strm, FstReadOptions(filename));
+    } else {
+      return Read(std::cin, FstReadOptions("standard input"));
+    }
+  }
+
+  static LinearTaggerFst<A> *Read(std::istream &in,  // NOLINT
+                                  const FstReadOptions &opts) {
+    auto *impl = Impl::Read(in, opts);
+    return impl ? new LinearTaggerFst<A>(std::shared_ptr<Impl>(impl)) : nullptr;
+  }
+
+  bool Write(const string &filename) const override {
+    if (!filename.empty()) {
+      std::ofstream strm(filename,
+                               std::ios_base::out | std::ios_base::binary);
+      if (!strm) {
+        LOG(ERROR) << "LinearTaggerFst::Write: Can't open file: " << filename;
+        return false;
+      }
+      return Write(strm, FstWriteOptions(filename));
+    } else {
+      return Write(std::cout, FstWriteOptions("standard output"));
+    }
+  }
+
+  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
+    return GetImpl()->Write(strm, opts);
+  }
+
+ private:
+  using ImplToFst<Impl>::GetImpl;
+  using ImplToFst<Impl>::GetMutableImpl;
+
+  explicit LinearTaggerFst(std::shared_ptr<Impl> impl)
+      : ImplToFst<Impl>(impl) {}
+
+  void operator=(const LinearTaggerFst<A> &fst) = delete;
+};
+
+// Specialization for LinearTaggerFst.
+template <class Arc>
+class StateIterator<LinearTaggerFst<Arc>>
+    : public CacheStateIterator<LinearTaggerFst<Arc>> {
+ public:
+  explicit StateIterator(const LinearTaggerFst<Arc> &fst)
+      : CacheStateIterator<LinearTaggerFst<Arc>>(fst, fst.GetMutableImpl()) {}
+};
+
+// Specialization for LinearTaggerFst.
+template <class Arc>
+class ArcIterator<LinearTaggerFst<Arc>>
+    : public CacheArcIterator<LinearTaggerFst<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const LinearTaggerFst<Arc> &fst, StateId s)
+      : CacheArcIterator<LinearTaggerFst<Arc>>(fst.GetMutableImpl(), s) {
+    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
+  }
+};
+
+template <class Arc>
+inline void LinearTaggerFst<Arc>::InitStateIterator(
+    StateIteratorData<Arc> *data) const {
+  data->base = new StateIterator<LinearTaggerFst<Arc>>(*this);
+}
+
+namespace internal {
+
+// Implementation class for on-the-fly generated LinearClassifierFst with
+// special optimization in matching.
+template <class A>
+class LinearClassifierFstImpl : public CacheImpl<A> {
+ public:
+  using FstImpl<A>::SetType;
+  using FstImpl<A>::SetProperties;
+  using FstImpl<A>::SetInputSymbols;
+  using FstImpl<A>::SetOutputSymbols;
+  using FstImpl<A>::WriteHeader;
+
+  using CacheBaseImpl<CacheState<A>>::PushArc;
+  using CacheBaseImpl<CacheState<A>>::HasArcs;
+  using CacheBaseImpl<CacheState<A>>::HasFinal;
+  using CacheBaseImpl<CacheState<A>>::HasStart;
+  using CacheBaseImpl<CacheState<A>>::SetArcs;
+  using CacheBaseImpl<CacheState<A>>::SetFinal;
+  using CacheBaseImpl<CacheState<A>>::SetStart;
+
+  typedef A Arc;
+  typedef typename A::Label Label;
+  typedef typename A::Weight Weight;
+  typedef typename A::StateId StateId;
+  typedef typename Collection<StateId, Label>::SetIterator NGramIterator;
+
+  // Constructs an empty FST by default.
+  LinearClassifierFstImpl()
+      : CacheImpl<A>(CacheOptions()),
+        data_(std::make_shared<LinearFstData<A>>()) {
+    SetType("linear-classifier");
+    num_classes_ = 0;
+    num_groups_ = 0;
+  }
+
+  // Constructs the FST with given data storage, number of classes and
+  // symbol tables.
+  LinearClassifierFstImpl(const LinearFstData<Arc> *data, size_t num_classes,
+                          const SymbolTable *isyms, const SymbolTable *osyms,
+                          CacheOptions opts)
+      : CacheImpl<A>(opts),
+        data_(data),
+        num_classes_(num_classes),
+        num_groups_(data_->NumGroups() / num_classes_) {
+    SetType("linear-classifier");
+    SetProperties(kILabelSorted, kFstProperties);
+    SetInputSymbols(isyms);
+    SetOutputSymbols(osyms);
+    ReserveStubSpace();
+  }
+
+  // Copy by sharing the underlying data storage.
+  LinearClassifierFstImpl(const LinearClassifierFstImpl &impl)
+      : CacheImpl<A>(impl),
+        data_(impl.data_),
+        num_classes_(impl.num_classes_),
+        num_groups_(impl.num_groups_) {
+    SetType("linear-classifier");
+    SetProperties(impl.Properties(), kCopyProperties);
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+    ReserveStubSpace();
+  }
+
+  StateId Start() {
+    if (!HasStart()) {
+      StateId start = FindStartState();
+      SetStart(start);
+    }
+    return CacheImpl<A>::Start();
+  }
+
+  Weight Final(StateId s) {
+    if (!HasFinal(s)) {
+      state_stub_.clear();
+      FillState(s, &state_stub_);
+      SetFinal(s, FinalWeight(state_stub_));
+    }
+    return CacheImpl<A>::Final(s);
+  }
+
+  size_t NumArcs(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<A>::NumArcs(s);
+  }
+
+  size_t NumInputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<A>::NumInputEpsilons(s);
+  }
+
+  size_t NumOutputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<A>::NumOutputEpsilons(s);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<A> *data) {
+    if (!HasArcs(s)) Expand(s);
+    CacheImpl<A>::InitArcIterator(s, data);
+  }
+
+  // Computes the outgoing transitions from a state, creating new
+  // destination states as needed.
+  void Expand(StateId s);
+
+  // Appends to `arcs` all out-going arcs from state `s` that matches
+  // `label` as the input label.
+  void MatchInput(StateId s, Label ilabel, std::vector<Arc> *arcs);
+
+  static LinearClassifierFstImpl<A> *Read(std::istream &strm,
+                                          const FstReadOptions &opts);
+
+  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
+    FstHeader header;
+    header.SetStart(kNoStateId);
+    WriteHeader(strm, opts, kFileVersion, &header);
+    data_->Write(strm);
+    WriteType(strm, num_classes_);
+    if (!strm) {
+      LOG(ERROR) << "LinearClassifierFst::Write: Write failed: " << opts.source;
+      return false;
+    }
+    return true;
+  }
+
+ private:
+  static const int kMinFileVersion;
+  static const int kFileVersion;
+
+  // A collection of functions to access parts of the state tuple. A
+  // state tuple is a vector of `Label`s with two parts:
+  //   [prediction] [internal].
+  //
+  // - [prediction] is a single label of the predicted class. A state
+  //   must have a positive class label, unless it is the start state.
+  //
+  // - [internal] is the internal state tuple for `LinearFstData` of
+  //   the given class; or kNoTrieNodeId's if in start state.
+  Label &Prediction(std::vector<Label> &state) { return state[0]; }  // NOLINT
+  Label Prediction(const std::vector<Label> &state) const { return state[0]; }
+
+  Label &InternalAt(std::vector<Label> &state, int index) {  // NOLINT
+    return state[index + 1];
+  }
+  Label InternalAt(const std::vector<Label> &state, int index) const {
+    return state[index + 1];
+  }
+
+  // The size of state tuples are fixed, reserve them in stubs
+  void ReserveStubSpace() {
+    size_t size = 1 + num_groups_;
+    state_stub_.reserve(size);
+    next_stub_.reserve(size);
+  }
+
+  // Computes the start state tuple and maps it to the start state id.
+  StateId FindStartState() {
+    // A start state tuple has no prediction
+    state_stub_.clear();
+    state_stub_.push_back(kNoLabel);
+    // For a start state, we don't yet know where we are in the tries.
+    for (size_t i = 0; i < num_groups_; ++i)
+      state_stub_.push_back(kNoTrieNodeId);
+    return FindState(state_stub_);
+  }
+
+  // Tests if the state tuple represents the start state.
+  bool IsStartState(const std::vector<Label> &state) const {
+    return state[0] == kNoLabel;
+  }
+
+  // Computes the actual group id in the data storage.
+  int GroupId(Label pred, int group) const {
+    return group * num_classes_ + pred - 1;
+  }
+
+  // Finds out the final weight of the given state. A state is final
+  // iff it is not the start.
+  Weight FinalWeight(const std::vector<Label> &state) const {
+    if (IsStartState(state)) {
+      return Weight::Zero();
+    }
+    Label pred = Prediction(state);
+    DCHECK_GT(pred, 0);
+    DCHECK_LE(pred, num_classes_);
+    Weight final_weight = Weight::One();
+    for (size_t group = 0; group < num_groups_; ++group) {
+      int group_id = GroupId(pred, group);
+      int trie_state = InternalAt(state, group);
+      final_weight =
+          Times(final_weight, data_->GroupFinalWeight(group_id, trie_state));
+    }
+    return final_weight;
+  }
+
+  // Finds state corresponding to an n-gram. Creates new state if n-gram not
+  // found.
+  StateId FindState(const std::vector<Label> &ngram) {
+    StateId sparse = ngrams_.FindId(ngram, true);
+    StateId dense = condensed_.FindId(sparse, true);
+    return dense;
+  }
+
+  // Appends after `output` the state tuple corresponding to the state id. The
+  // state id must exist.
+  void FillState(StateId s, std::vector<Label> *output) {
+    s = condensed_.FindEntry(s);
+    for (NGramIterator it = ngrams_.FindSet(s); !it.Done(); it.Next()) {
+      Label label = it.Element();
+      output->push_back(label);
+    }
+  }
+
+  std::shared_ptr<const LinearFstData<A>> data_;
+  // Division of groups in `data_`; num_classes_ * num_groups_ ==
+  // data_->NumGroups().
+  size_t num_classes_, num_groups_;
+  // Mapping from internal state tuple to *non-consecutive* ids
+  Collection<StateId, Label> ngrams_;
+  // Mapping from non-consecutive id to actual state id
+  CompactHashBiTable<StateId, StateId, std::hash<StateId>> condensed_;
+  // Two frequently used vectors, reuse to avoid repeated heap
+  // allocation
+  std::vector<Label> state_stub_, next_stub_;
+
+  void operator=(const LinearClassifierFstImpl<A> &) = delete;
+};
+
+template <class A>
+const int LinearClassifierFstImpl<A>::kMinFileVersion = 0;
+
+template <class A>
+const int LinearClassifierFstImpl<A>::kFileVersion = 0;
+
+template <class A>
+void LinearClassifierFstImpl<A>::Expand(StateId s) {
+  VLOG(3) << "Expand " << s;
+  state_stub_.clear();
+  FillState(s, &state_stub_);
+  next_stub_.clear();
+  next_stub_.resize(1 + num_groups_);
+
+  if (IsStartState(state_stub_)) {
+    // Make prediction
+    for (Label pred = 1; pred <= num_classes_; ++pred) {
+      Prediction(next_stub_) = pred;
+      for (int i = 0; i < num_groups_; ++i)
+        InternalAt(next_stub_, i) = data_->GroupStartState(GroupId(pred, i));
+      PushArc(s, A(0, pred, Weight::One(), FindState(next_stub_)));
+    }
+  } else {
+    Label pred = Prediction(state_stub_);
+    DCHECK_GT(pred, 0);
+    DCHECK_LE(pred, num_classes_);
+    for (Label ilabel = data_->MinInputLabel();
+         ilabel <= data_->MaxInputLabel(); ++ilabel) {
+      Prediction(next_stub_) = pred;
+      Weight weight = Weight::One();
+      for (int i = 0; i < num_groups_; ++i)
+        InternalAt(next_stub_, i) =
+            data_->GroupTransition(GroupId(pred, i), InternalAt(state_stub_, i),
+                                   ilabel, pred, &weight);
+      PushArc(s, A(ilabel, 0, weight, FindState(next_stub_)));
+    }
+  }
+
+  SetArcs(s);
+}
+
+template <class A>
+void LinearClassifierFstImpl<A>::MatchInput(StateId s, Label ilabel,
+                                            std::vector<Arc> *arcs) {
+  state_stub_.clear();
+  FillState(s, &state_stub_);
+  next_stub_.clear();
+  next_stub_.resize(1 + num_groups_);
+
+  if (IsStartState(state_stub_)) {
+    // Make prediction if `ilabel` is epsilon.
+    if (ilabel == 0) {
+      for (Label pred = 1; pred <= num_classes_; ++pred) {
+        Prediction(next_stub_) = pred;
+        for (int i = 0; i < num_groups_; ++i)
+          InternalAt(next_stub_, i) = data_->GroupStartState(GroupId(pred, i));
+        arcs->push_back(A(0, pred, Weight::One(), FindState(next_stub_)));
+      }
+    }
+  } else if (ilabel != 0) {
+    Label pred = Prediction(state_stub_);
+    Weight weight = Weight::One();
+    Prediction(next_stub_) = pred;
+    for (int i = 0; i < num_groups_; ++i)
+      InternalAt(next_stub_, i) = data_->GroupTransition(
+          GroupId(pred, i), InternalAt(state_stub_, i), ilabel, pred, &weight);
+    arcs->push_back(A(ilabel, 0, weight, FindState(next_stub_)));
+  }
+}
+
+template <class A>
+inline LinearClassifierFstImpl<A> *LinearClassifierFstImpl<A>::Read(
+    std::istream &strm, const FstReadOptions &opts) {
+  std::unique_ptr<LinearClassifierFstImpl<A>> impl(
+      new LinearClassifierFstImpl<A>());
+  FstHeader header;
+  if (!impl->ReadHeader(strm, opts, kMinFileVersion, &header)) {
+    return nullptr;
+  }
+  impl->data_ = std::shared_ptr<LinearFstData<A>>(LinearFstData<A>::Read(strm));
+  if (!impl->data_) {
+    return nullptr;
+  }
+  ReadType(strm, &impl->num_classes_);
+  if (!strm) {
+    return nullptr;
+  }
+  impl->num_groups_ = impl->data_->NumGroups() / impl->num_classes_;
+  if (impl->num_groups_ * impl->num_classes_ != impl->data_->NumGroups()) {
+    FSTERROR() << "Total number of feature groups is not a multiple of the "
+                  "number of classes: num groups = "
+               << impl->data_->NumGroups()
+               << ", num classes = " << impl->num_classes_;
+    return nullptr;
+  }
+  impl->ReserveStubSpace();
+  return impl.release();
+}
+
+}  // namespace internal
+
+// This class attaches interface to implementation and handles
+// reference counting, delegating most methods to ImplToFst.
+template <class A>
+class LinearClassifierFst
+    : public ImplToFst<internal::LinearClassifierFstImpl<A>> {
+ public:
+  friend class ArcIterator<LinearClassifierFst<A>>;
+  friend class StateIterator<LinearClassifierFst<A>>;
+  friend class LinearFstMatcherTpl<LinearClassifierFst<A>>;
+
+  typedef A Arc;
+  typedef typename A::Label Label;
+  typedef typename A::Weight Weight;
+  typedef typename A::StateId StateId;
+  typedef DefaultCacheStore<A> Store;
+  typedef typename Store::State State;
+  using Impl = internal::LinearClassifierFstImpl<A>;
+
+  LinearClassifierFst() : ImplToFst<Impl>(std::make_shared<Impl>()) {}
+
+  explicit LinearClassifierFst(LinearFstData<A> *data, size_t num_classes,
+                               const SymbolTable *isyms = nullptr,
+                               const SymbolTable *osyms = nullptr,
+                               CacheOptions opts = CacheOptions())
+      : ImplToFst<Impl>(
+            std::make_shared<Impl>(data, num_classes, isyms, osyms, opts)) {}
+
+  explicit LinearClassifierFst(const Fst<A> &fst)
+      : ImplToFst<Impl>(std::make_shared<Impl>()) {
+    LOG(FATAL) << "LinearClassifierFst: no constructor from arbitrary FST.";
+  }
+
+  // See Fst<>::Copy() for doc.
+  LinearClassifierFst(const LinearClassifierFst<A> &fst, bool safe = false)
+      : ImplToFst<Impl>(fst, safe) {}
+
+  // Get a copy of this LinearClassifierFst. See Fst<>::Copy() for further doc.
+  LinearClassifierFst<A> *Copy(bool safe = false) const override {
+    return new LinearClassifierFst<A>(*this, safe);
+  }
+
+  inline void InitStateIterator(StateIteratorData<A> *data) const override;
+
+  void InitArcIterator(StateId s, ArcIteratorData<A> *data) const override {
+    GetMutableImpl()->InitArcIterator(s, data);
+  }
+
+  MatcherBase<A> *InitMatcher(MatchType match_type) const override {
+    return new LinearFstMatcherTpl<LinearClassifierFst<A>>(this, match_type);
+  }
+
+  static LinearClassifierFst<A> *Read(const string &filename) {
+    if (!filename.empty()) {
+      std::ifstream strm(filename,
+                              std::ios_base::in | std::ios_base::binary);
+      if (!strm) {
+        LOG(ERROR) << "LinearClassifierFst::Read: Can't open file: "
+                   << filename;
+        return nullptr;
+      }
+      return Read(strm, FstReadOptions(filename));
+    } else {
+      return Read(std::cin, FstReadOptions("standard input"));
+    }
+  }
+
+  static LinearClassifierFst<A> *Read(std::istream &in,
+                                      const FstReadOptions &opts) {
+    auto *impl = Impl::Read(in, opts);
+    return impl ? new LinearClassifierFst<A>(std::shared_ptr<Impl>(impl))
+                : nullptr;
+  }
+
+  bool Write(const string &filename) const override {
+    if (!filename.empty()) {
+      std::ofstream strm(filename,
+                               std::ios_base::out | std::ios_base::binary);
+      if (!strm) {
+        LOG(ERROR) << "ProdLmFst::Write: Can't open file: " << filename;
+        return false;
+      }
+      return Write(strm, FstWriteOptions(filename));
+    } else {
+      return Write(std::cout, FstWriteOptions("standard output"));
+    }
+  }
+
+  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
+    return GetImpl()->Write(strm, opts);
+  }
+
+ private:
+  using ImplToFst<Impl>::GetImpl;
+  using ImplToFst<Impl>::GetMutableImpl;
+
+  explicit LinearClassifierFst(std::shared_ptr<Impl> impl)
+      : ImplToFst<Impl>(impl) {}
+
+  void operator=(const LinearClassifierFst<A> &fst) = delete;
+};
+
+// Specialization for LinearClassifierFst.
+template <class Arc>
+class StateIterator<LinearClassifierFst<Arc>>
+    : public CacheStateIterator<LinearClassifierFst<Arc>> {
+ public:
+  explicit StateIterator(const LinearClassifierFst<Arc> &fst)
+      : CacheStateIterator<LinearClassifierFst<Arc>>(fst,
+                                                     fst.GetMutableImpl()) {}
+};
+
+// Specialization for LinearClassifierFst.
+template <class Arc>
+class ArcIterator<LinearClassifierFst<Arc>>
+    : public CacheArcIterator<LinearClassifierFst<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const LinearClassifierFst<Arc> &fst, StateId s)
+      : CacheArcIterator<LinearClassifierFst<Arc>>(fst.GetMutableImpl(), s) {
+    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
+  }
+};
+
+template <class Arc>
+inline void LinearClassifierFst<Arc>::InitStateIterator(
+    StateIteratorData<Arc> *data) const {
+  data->base = new StateIterator<LinearClassifierFst<Arc>>(*this);
+}
+
+// Specialized Matcher for LinearFsts. This matcher only supports
+// matching from the input side. This is intentional because comparing
+// the scores of different input sequences with the same output
+// sequence is meaningless in a discriminative model.
+template <class F>
+class LinearFstMatcherTpl : public MatcherBase<typename F::Arc> {
+ public:
+  typedef typename F::Arc Arc;
+  typedef typename Arc::Label Label;
+  typedef typename Arc::Weight Weight;
+  typedef typename Arc::StateId StateId;
+  typedef F FST;
+
+  // This makes a copy of the FST.
+  LinearFstMatcherTpl(const FST &fst, MatchType match_type)
+      : owned_fst_(fst.Copy()),
+        fst_(*owned_fst_),
+        match_type_(match_type),
+        s_(kNoStateId),
+        current_loop_(false),
+        loop_(kNoLabel, 0, Weight::One(), kNoStateId),
+        cur_arc_(0),
+        error_(false) {
+    switch (match_type_) {
+      case MATCH_INPUT:
+      case MATCH_OUTPUT:
+      case MATCH_NONE:
+        break;
+      default:
+        FSTERROR() << "LinearFstMatcherTpl: Bad match type";
+        match_type_ = MATCH_NONE;
+        error_ = true;
+    }
+  }
+
+  // This doesn't copy the FST.
+  LinearFstMatcherTpl(const FST *fst, MatchType match_type)
+      : fst_(*fst),
+        match_type_(match_type),
+        s_(kNoStateId),
+        current_loop_(false),
+        loop_(kNoLabel, 0, Weight::One(), kNoStateId),
+        cur_arc_(0),
+        error_(false) {
+    switch (match_type_) {
+      case MATCH_INPUT:
+      case MATCH_OUTPUT:
+      case MATCH_NONE:
+        break;
+      default:
+        FSTERROR() << "LinearFstMatcherTpl: Bad match type";
+        match_type_ = MATCH_NONE;
+        error_ = true;
+    }
+  }
+
+  // This makes a copy of the FST.
+  LinearFstMatcherTpl(const LinearFstMatcherTpl<F> &matcher, bool safe = false)
+      : owned_fst_(matcher.fst_.Copy(safe)),
+        fst_(*owned_fst_),
+        match_type_(matcher.match_type_),
+        s_(kNoStateId),
+        current_loop_(false),
+        loop_(matcher.loop_),
+        cur_arc_(0),
+        error_(matcher.error_) {}
+
+  LinearFstMatcherTpl<F> *Copy(bool safe = false) const override {
+    return new LinearFstMatcherTpl<F>(*this, safe);
+  }
+
+  MatchType Type(bool /*test*/) const override {
+    // `MATCH_INPUT` is the only valid type
+    return match_type_ == MATCH_INPUT ? match_type_ : MATCH_NONE;
+  }
+
+  void SetState(StateId s) final {
+    if (s_ == s) return;
+    s_ = s;
+    // `MATCH_INPUT` is the only valid type
+    if (match_type_ != MATCH_INPUT) {
+      FSTERROR() << "LinearFstMatcherTpl: Bad match type";
+      error_ = true;
+    }
+    loop_.nextstate = s;
+  }
+
+  bool Find(Label label) final {
+    if (error_) {
+      current_loop_ = false;
+      return false;
+    }
+    current_loop_ = label == 0;
+    if (label == kNoLabel) label = 0;
+    arcs_.clear();
+    cur_arc_ = 0;
+    fst_.GetMutableImpl()->MatchInput(s_, label, &arcs_);
+    return current_loop_ || !arcs_.empty();
+  }
+
+  bool Done() const final {
+    return !(current_loop_ || cur_arc_ < arcs_.size());
+  }
+
+  const Arc &Value() const final {
+    return current_loop_ ? loop_ : arcs_[cur_arc_];
+  }
+
+  void Next() final {
+    if (current_loop_)
+      current_loop_ = false;
+    else
+      ++cur_arc_;
+  }
+
+  ssize_t Priority(StateId s) final { return kRequirePriority; }
+
+  const FST &GetFst() const override { return fst_; }
+
+  uint64 Properties(uint64 props) const override {
+    if (error_) props |= kError;
+    return props;
+  }
+
+  uint32 Flags() const override { return kRequireMatch; }
+
+ private:
+  std::unique_ptr<const FST> owned_fst_;
+  const FST &fst_;
+  MatchType match_type_;  // Type of match to perform.
+  StateId s_;             // Current state.
+  bool current_loop_;     // Current arc is the implicit loop.
+  Arc loop_;              // For non-consuming symbols.
+  // All out-going arcs matching the label in last Find() call.
+  std::vector<Arc> arcs_;
+  size_t cur_arc_;  // Index to the arc that `Value()` should return.
+  bool error_;      // Error encountered.
+};
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_LINEAR_LINEAR_FST_H_
diff --git a/projects/llm_framework/include/fst/extensions/linear/linearscript.h b/projects/llm_framework/include/fst/extensions/linear/linearscript.h
new file mode 100644
index 00000000..54106d20
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/linear/linearscript.h
@@ -0,0 +1,391 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_EXTENSIONS_LINEAR_LINEARSCRIPT_H_
+#define FST_EXTENSIONS_LINEAR_LINEARSCRIPT_H_
+
+#include <istream>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#include <fst/compat.h>
+#include <fst/extensions/linear/linear-fst-data-builder.h>
+#include <fst/extensions/linear/linear-fst.h>
+#include <fstream>
+#include <fst/symbol-table.h>
+#include <fst/script/arg-packs.h>
+#include <fst/script/script-impl.h>
+
+DECLARE_string(delimiter);
+DECLARE_string(empty_symbol);
+DECLARE_string(start_symbol);
+DECLARE_string(end_symbol);
+DECLARE_bool(classifier);
+
+namespace fst {
+namespace script {
+typedef std::tuple<const string &, const string &, const string &, char **, int,
+                   const string &, const string &, const string &,
+                   const string &>
+    LinearCompileArgs;
+
+bool ValidateDelimiter();
+bool ValidateEmptySymbol();
+
+// Returns the proper label given the symbol. For symbols other than
+// `FLAGS_start_symbol` or `FLAGS_end_symbol`, looks up the symbol
+// table to decide the label. Depending on whether
+// `FLAGS_start_symbol` and `FLAGS_end_symbol` are identical, it
+// either returns `kNoLabel` for later processing or decides the label
+// right away.
+template <class Arc>
+inline typename Arc::Label LookUp(const string &str, SymbolTable *syms) {
+  if (str == FLAGS_start_symbol)
+    return str == FLAGS_end_symbol ? kNoLabel
+                                   : LinearFstData<Arc>::kStartOfSentence;
+  else if (str == FLAGS_end_symbol)
+    return LinearFstData<Arc>::kEndOfSentence;
+  else
+    return syms->AddSymbol(str);
+}
+
+// Splits `str` with `delim` as the delimiter and stores the labels in
+// `output`.
+template <class Arc>
+void SplitAndPush(const string &str, const char delim, SymbolTable *syms,
+                  std::vector<typename Arc::Label> *output) {
+  if (str == FLAGS_empty_symbol) return;
+  std::istringstream strm(str);
+  string buf;
+  while (std::getline(strm, buf, delim))
+    output->push_back(LookUp<Arc>(buf, syms));
+}
+
+// Like `std::replace_copy` but returns the number of modifications
+template <class InputIterator, class OutputIterator, class T>
+size_t ReplaceCopy(InputIterator first, InputIterator last,
+                   OutputIterator result, const T &old_value,
+                   const T &new_value) {
+  size_t changes = 0;
+  while (first != last) {
+    if (*first == old_value) {
+      *result = new_value;
+      ++changes;
+    } else {
+      *result = *first;
+    }
+    ++first;
+    ++result;
+  }
+  return changes;
+}
+
+template <class Arc>
+bool GetVocabRecord(const string &vocab, std::istream &strm,  // NOLINT
+                    SymbolTable *isyms, SymbolTable *fsyms, SymbolTable *osyms,
+                    typename Arc::Label *word,
+                    std::vector<typename Arc::Label> *feature_labels,
+                    std::vector<typename Arc::Label> *possible_labels,
+                    size_t *num_line);
+
+template <class Arc>
+bool GetModelRecord(const string &model, std::istream &strm,  // NOLINT
+                    SymbolTable *fsyms, SymbolTable *osyms,
+                    std::vector<typename Arc::Label> *input_labels,
+                    std::vector<typename Arc::Label> *output_labels,
+                    typename Arc::Weight *weight, size_t *num_line);
+
+// Reads in vocabulary file. Each line is in the following format
+//
+//   word <whitespace> features [ <whitespace> possible output ]
+//
+// where features and possible output are `FLAGS_delimiter`-delimited lists of
+// tokens
+template <class Arc>
+void AddVocab(const string &vocab, SymbolTable *isyms, SymbolTable *fsyms,
+              SymbolTable *osyms, LinearFstDataBuilder<Arc> *builder) {
+  std::ifstream in(vocab);
+  if (!in) LOG(FATAL) << "Can't open file: " << vocab;
+  size_t num_line = 0, num_added = 0;
+  std::vector<string> fields;
+  std::vector<typename Arc::Label> feature_labels, possible_labels;
+  typename Arc::Label word;
+  while (GetVocabRecord<Arc>(vocab, in, isyms, fsyms, osyms, &word,
+                             &feature_labels, &possible_labels, &num_line)) {
+    if (word == kNoLabel) {
+      LOG(WARNING) << "Ignored: boundary word: " << fields[0];
+      continue;
+    }
+    if (possible_labels.empty())
+      num_added += builder->AddWord(word, feature_labels);
+    else
+      num_added += builder->AddWord(word, feature_labels, possible_labels);
+  }
+  VLOG(1) << "Read " << num_added << " words in " << num_line << " lines from "
+          << vocab;
+}
+
+template <class Arc>
+void AddVocab(const string &vocab, SymbolTable *isyms, SymbolTable *fsyms,
+              SymbolTable *osyms,
+              LinearClassifierFstDataBuilder<Arc> *builder) {
+  std::ifstream in(vocab);
+  if (!in) LOG(FATAL) << "Can't open file: " << vocab;
+  size_t num_line = 0, num_added = 0;
+  std::vector<string> fields;
+  std::vector<typename Arc::Label> feature_labels, possible_labels;
+  typename Arc::Label word;
+  while (GetVocabRecord<Arc>(vocab, in, isyms, fsyms, osyms, &word,
+                             &feature_labels, &possible_labels, &num_line)) {
+    if (!possible_labels.empty())
+      LOG(FATAL)
+          << "Classifier vocabulary should not have possible output constraint";
+    if (word == kNoLabel) {
+      LOG(WARNING) << "Ignored: boundary word: " << fields[0];
+      continue;
+    }
+    num_added += builder->AddWord(word, feature_labels);
+  }
+  VLOG(1) << "Read " << num_added << " words in " << num_line << " lines from "
+          << vocab;
+}
+
+// Reads in model file. The first line is an integer designating the
+// size of future window in the input sequences. After this, each line
+// is in the following format
+//
+//   input sequence <whitespace> output sequence <whitespace> weight
+//
+// input sequence is a `FLAGS_delimiter`-delimited sequence of feature
+// labels (see `AddVocab()`) . output sequence is a
+// `FLAGS_delimiter`-delimited sequence of output labels where the
+// last label is the output of the feature position before the history
+// boundary.
+template <class Arc>
+void AddModel(const string &model, SymbolTable *fsyms, SymbolTable *osyms,
+              LinearFstDataBuilder<Arc> *builder) {
+  std::ifstream in(model);
+  if (!in) LOG(FATAL) << "Can't open file: " << model;
+  string line;
+  std::getline(in, line);
+  if (!in) LOG(FATAL) << "Empty file: " << model;
+  size_t future_size;
+  {
+    std::istringstream strm(line);
+    strm >> future_size;
+    if (!strm) LOG(FATAL) << "Can't read future size: " << model;
+  }
+  size_t num_line = 1, num_added = 0;
+  const int group = builder->AddGroup(future_size);
+  VLOG(1) << "Group " << group << ": from " << model << "; future size is "
+          << future_size << ".";
+  // Add the rest of lines as a single feature group
+  std::vector<string> fields;
+  std::vector<typename Arc::Label> input_labels, output_labels;
+  typename Arc::Weight weight;
+  while (GetModelRecord<Arc>(model, in, fsyms, osyms, &input_labels,
+                             &output_labels, &weight, &num_line)) {
+    if (output_labels.empty())
+      LOG(FATAL) << "Empty output sequence in source " << model << ", line "
+                 << num_line;
+
+    const typename Arc::Label marks[] = {LinearFstData<Arc>::kStartOfSentence,
+                                         LinearFstData<Arc>::kEndOfSentence};
+
+    std::vector<typename Arc::Label> copy_input(input_labels.size()),
+        copy_output(output_labels.size());
+    for (int i = 0; i < 2; ++i) {
+      for (int j = 0; j < 2; ++j) {
+        size_t num_input_changes =
+            ReplaceCopy(input_labels.begin(), input_labels.end(),
+                        copy_input.begin(), kNoLabel, marks[i]);
+        size_t num_output_changes =
+            ReplaceCopy(output_labels.begin(), output_labels.end(),
+                        copy_output.begin(), kNoLabel, marks[j]);
+        if ((num_input_changes > 0 || i == 0) &&
+            (num_output_changes > 0 || j == 0))
+          num_added +=
+              builder->AddWeight(group, copy_input, copy_output, weight);
+      }
+    }
+  }
+  VLOG(1) << "Group " << group << ": read " << num_added << " weight(s) in "
+          << num_line << " lines.";
+}
+
+template <class Arc>
+void AddModel(const string &model, SymbolTable *fsyms, SymbolTable *osyms,
+              LinearClassifierFstDataBuilder<Arc> *builder) {
+  std::ifstream in(model);
+  if (!in) LOG(FATAL) << "Can't open file: " << model;
+  string line;
+  std::getline(in, line);
+  if (!in) LOG(FATAL) << "Empty file: " << model;
+  size_t future_size;
+  {
+    std::istringstream strm(line);
+    strm >> future_size;
+    if (!strm) LOG(FATAL) << "Can't read future size: " << model;
+  }
+  if (future_size != 0)
+    LOG(FATAL) << "Classifier model must have future size = 0; got "
+               << future_size << " from " << model;
+  size_t num_line = 1, num_added = 0;
+  const int group = builder->AddGroup();
+  VLOG(1) << "Group " << group << ": from " << model << "; future size is "
+          << future_size << ".";
+  // Add the rest of lines as a single feature group
+  std::vector<string> fields;
+  std::vector<typename Arc::Label> input_labels, output_labels;
+  typename Arc::Weight weight;
+  while (GetModelRecord<Arc>(model, in, fsyms, osyms, &input_labels,
+                             &output_labels, &weight, &num_line)) {
+    if (output_labels.size() != 1)
+      LOG(FATAL) << "Output not a single label in source " << model << ", line "
+                 << num_line;
+
+    const typename Arc::Label marks[] = {LinearFstData<Arc>::kStartOfSentence,
+                                         LinearFstData<Arc>::kEndOfSentence};
+
+    typename Arc::Label pred = output_labels[0];
+
+    std::vector<typename Arc::Label> copy_input(input_labels.size());
+    for (int i = 0; i < 2; ++i) {
+      size_t num_input_changes =
+          ReplaceCopy(input_labels.begin(), input_labels.end(),
+                      copy_input.begin(), kNoLabel, marks[i]);
+      if (num_input_changes > 0 || i == 0)
+        num_added += builder->AddWeight(group, copy_input, pred, weight);
+    }
+  }
+  VLOG(1) << "Group " << group << ": read " << num_added << " weight(s) in "
+          << num_line << " lines.";
+}
+
+void SplitByWhitespace(const string &str, std::vector<string> *out);
+int ScanNumClasses(char **models, int models_length);
+
+template <class Arc>
+void LinearCompileTpl(LinearCompileArgs *args) {
+  const string &epsilon_symbol = std::get<0>(*args);
+  const string &unknown_symbol = std::get<1>(*args);
+  const string &vocab = std::get<2>(*args);
+  char **models = std::get<3>(*args);
+  const int models_length = std::get<4>(*args);
+  const string &out = std::get<5>(*args);
+  const string &save_isymbols = std::get<6>(*args);
+  const string &save_fsymbols = std::get<7>(*args);
+  const string &save_osymbols = std::get<8>(*args);
+
+  SymbolTable isyms,  // input (e.g. word tokens)
+      osyms,          // output (e.g. tags)
+      fsyms;          // feature (e.g. word identity, suffix, etc.)
+  isyms.AddSymbol(epsilon_symbol);
+  osyms.AddSymbol(epsilon_symbol);
+  fsyms.AddSymbol(epsilon_symbol);
+  isyms.AddSymbol(unknown_symbol);
+
+  VLOG(1) << "start-of-sentence label is "
+          << LinearFstData<Arc>::kStartOfSentence;
+  VLOG(1) << "end-of-sentence label is " << LinearFstData<Arc>::kEndOfSentence;
+
+  if (FLAGS_classifier) {
+    int num_classes = ScanNumClasses(models, models_length);
+    LinearClassifierFstDataBuilder<Arc> builder(num_classes, &isyms, &fsyms,
+                                                &osyms);
+
+    AddVocab(vocab, &isyms, &fsyms, &osyms, &builder);
+    for (int i = 0; i < models_length; ++i)
+      AddModel(models[i], &fsyms, &osyms, &builder);
+
+    LinearClassifierFst<Arc> fst(builder.Dump(), num_classes, &isyms, &osyms);
+    fst.Write(out);
+  } else {
+    LinearFstDataBuilder<Arc> builder(&isyms, &fsyms, &osyms);
+
+    AddVocab(vocab, &isyms, &fsyms, &osyms, &builder);
+    for (int i = 0; i < models_length; ++i)
+      AddModel(models[i], &fsyms, &osyms, &builder);
+
+    LinearTaggerFst<Arc> fst(builder.Dump(), &isyms, &osyms);
+    fst.Write(out);
+  }
+
+  if (!save_isymbols.empty()) isyms.WriteText(save_isymbols);
+  if (!save_fsymbols.empty()) fsyms.WriteText(save_fsymbols);
+  if (!save_osymbols.empty()) osyms.WriteText(save_osymbols);
+}
+
+void LinearCompile(const string &arc_type, const string &epsilon_symbol,
+                   const string &unknown_symbol, const string &vocab,
+                   char **models, int models_len, const string &out,
+                   const string &save_isymbols, const string &save_fsymbols,
+                   const string &save_osymbols);
+
+template <class Arc>
+bool GetVocabRecord(const string &vocab, std::istream &strm,  // NOLINT
+                    SymbolTable *isyms, SymbolTable *fsyms, SymbolTable *osyms,
+                    typename Arc::Label *word,
+                    std::vector<typename Arc::Label> *feature_labels,
+                    std::vector<typename Arc::Label> *possible_labels,
+                    size_t *num_line) {
+  string line;
+  if (!std::getline(strm, line)) return false;
+  ++(*num_line);
+
+  std::vector<string> fields;
+  SplitByWhitespace(line, &fields);
+  if (fields.size() != 3)
+    LOG(FATAL) << "Wrong number of fields in source " << vocab << ", line "
+               << num_line;
+
+  feature_labels->clear();
+  possible_labels->clear();
+
+  *word = LookUp<Arc>(fields[0], isyms);
+
+  const char delim = FLAGS_delimiter[0];
+  SplitAndPush<Arc>(fields[1], delim, fsyms, feature_labels);
+  SplitAndPush<Arc>(fields[2], delim, osyms, possible_labels);
+
+  return true;
+}
+
+template <class Arc>
+bool GetModelRecord(const string &model, std::istream &strm,  // NOLINT
+                    SymbolTable *fsyms, SymbolTable *osyms,
+                    std::vector<typename Arc::Label> *input_labels,
+                    std::vector<typename Arc::Label> *output_labels,
+                    typename Arc::Weight *weight, size_t *num_line) {
+  string line;
+  if (!std::getline(strm, line)) return false;
+  ++(*num_line);
+
+  std::vector<string> fields;
+  SplitByWhitespace(line, &fields);
+  if (fields.size() != 3)
+    LOG(FATAL) << "Wrong number of fields in source " << model << ", line "
+               << num_line;
+
+  input_labels->clear();
+  output_labels->clear();
+
+  const char delim = FLAGS_delimiter[0];
+  SplitAndPush<Arc>(fields[0], delim, fsyms, input_labels);
+  SplitAndPush<Arc>(fields[1], delim, osyms, output_labels);
+
+  *weight = StrToWeight<typename Arc::Weight>(fields[2], model, *num_line);
+
+  GuessStartOrEnd<Arc>(input_labels, kNoLabel);
+  GuessStartOrEnd<Arc>(output_labels, kNoLabel);
+
+  return true;
+}
+}  // namespace script
+}  // namespace fst
+
+#define REGISTER_FST_LINEAR_OPERATIONS(Arc) \
+  REGISTER_FST_OPERATION(LinearCompileTpl, Arc, LinearCompileArgs);
+
+#endif  // FST_EXTENSIONS_LINEAR_LINEARSCRIPT_H_
diff --git a/projects/llm_framework/include/fst/extensions/linear/loglinear-apply.h b/projects/llm_framework/include/fst/extensions/linear/loglinear-apply.h
new file mode 100644
index 00000000..1b5d2eaf
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/linear/loglinear-apply.h
@@ -0,0 +1,77 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_EXTENSIONS_LINEAR_LOGLINEAR_APPLY_H_
+#define FST_EXTENSIONS_LINEAR_LOGLINEAR_APPLY_H_
+
+#include <fst/compat.h>
+#include <fst/arc.h>
+#include <fst/arc-map.h>
+#include <fst/compose.h>
+#include <fst/determinize.h>
+#include <fst/float-weight.h>
+#include <fst/fst.h>
+#include <fst/minimize.h>
+#include <fst/mutable-fst.h>
+#include <fst/project.h>
+#include <fst/rmepsilon.h>
+#include <fst/vector-fst.h>
+
+namespace fst {
+
+// Applies a FST model as a discriminative model to weighted input
+// `ifst`. `A` is an arc type with tropical weight of all the
+// input/output FSTs.
+//
+// In general, consider `ifst` an unnormalized probability
+// distribution between its input X and output Y, P(X, Y); and `lfst`
+// a group of unnormalized probability distributions of all its output
+// Z for every input Y, Q(Z|Y). `normalize` controls whether Q is
+// normalized for every Y before chaining with P(X, Y). I.e., for a
+// path (X, Y, Z) in `ofst` (where Y is hidden),
+//
+// - When `normalize` is true, its weight is P(X, Y) Q(Z|Y) / sum_z Q(z|Y);
+// - When `normalize` is false, its weight is P(X, Y) Q(Z|Y).
+template <class A>
+void LogLinearApply(const Fst<A> &ifst, const Fst<A> &lfst, MutableFst<A> *ofst,
+                    bool normalize = true) {
+  LogLinearApply<A, LogArc>(ifst, lfst, ofst, normalize);
+}
+
+// This version gives finer control over the arc type (`B`) to be used
+// in normalization. `B` is an arc type with log weight (e.g. `LogArc`
+// or `Log64Arc`).
+template <class A, class B>
+void LogLinearApply(const Fst<A> &ifst, const Fst<A> &lfst, MutableFst<A> *ofst,
+                    bool normalize = true) {
+  if (normalize) {
+    VectorFst<A> unnormalized_ofst, rescored_ifsa;
+    Compose(ifst, lfst, &unnormalized_ofst);
+    {
+      VectorFst<A> tropical_ifsa(unnormalized_ofst);
+      Project(&tropical_ifsa, PROJECT_INPUT);
+      {
+        VectorFst<B> minimal_log_ifsa;
+        {
+          VectorFst<B> log_ifsa;
+          ArcMap(tropical_ifsa, &log_ifsa, WeightConvertMapper<A, B>());
+          RmEpsilon(&log_ifsa);
+          Determinize(log_ifsa, &minimal_log_ifsa);
+        }
+        Minimize(&minimal_log_ifsa);
+        ArcMap(&minimal_log_ifsa, InvertWeightMapper<B>());
+        ArcMap(minimal_log_ifsa, &tropical_ifsa, WeightConvertMapper<B, A>());
+      }
+      ArcSort(&tropical_ifsa, OLabelCompare<A>());
+      Compose(tropical_ifsa, ifst, &rescored_ifsa);
+    }
+    ArcSort(&rescored_ifsa, OLabelCompare<A>());
+    Compose(rescored_ifsa, unnormalized_ofst, ofst);
+  } else {
+    Compose(ifst, lfst, ofst);
+  }
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_LINEAR_LOGLINEAR_APPLY_H_
diff --git a/projects/llm_framework/include/fst/extensions/linear/trie.h b/projects/llm_framework/include/fst/extensions/linear/trie.h
new file mode 100644
index 00000000..b5ddb3ad
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/linear/trie.h
@@ -0,0 +1,444 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_EXTENSIONS_LINEAR_TRIE_H_
+#define FST_EXTENSIONS_LINEAR_TRIE_H_
+
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include <fst/compat.h>
+#include <fst/util.h>
+
+namespace fst {
+
+const int kNoTrieNodeId = -1;
+
+// Forward declarations of all available trie topologies.
+template <class L, class H>
+class NestedTrieTopology;
+template <class L, class H>
+class FlatTrieTopology;
+
+// A pair of parent node id and label, part of a trie edge
+template <class L>
+struct ParentLabel {
+  int parent;
+  L label;
+
+  ParentLabel() {}
+  ParentLabel(int p, L l) : parent(p), label(l) {}
+
+  bool operator==(const ParentLabel &that) const {
+    return parent == that.parent && label == that.label;
+  }
+
+  std::istream &Read(std::istream &strm) {  // NOLINT
+    ReadType(strm, &parent);
+    ReadType(strm, &label);
+    return strm;
+  }
+
+  std::ostream &Write(std::ostream &strm) const {  // NOLINT
+    WriteType(strm, parent);
+    WriteType(strm, label);
+    return strm;
+  }
+};
+
+template <class L, class H>
+struct ParentLabelHash {
+  size_t operator()(const ParentLabel<L> &pl) const {
+    return static_cast<size_t>(pl.parent * 7853 + H()(pl.label));
+  }
+};
+
+// The trie topology in a nested tree of hash maps; allows efficient
+// iteration over children of a specific node.
+template <class L, class H>
+class NestedTrieTopology {
+ public:
+  typedef L Label;
+  typedef H Hash;
+  typedef std::unordered_map<L, int, H> NextMap;
+
+  class const_iterator {
+   public:
+    typedef std::forward_iterator_tag iterator_category;
+    typedef std::pair<ParentLabel<L>, int> value_type;
+    typedef std::ptrdiff_t difference_type;
+    typedef const value_type *pointer;
+    typedef const value_type &reference;
+
+    friend class NestedTrieTopology<L, H>;
+
+    const_iterator() : ptr_(nullptr), cur_node_(kNoTrieNodeId), cur_edge_() {}
+
+    reference operator*() {
+      UpdateStub();
+      return stub_;
+    }
+    pointer operator->() {
+      UpdateStub();
+      return &stub_;
+    }
+
+    const_iterator &operator++();
+    const_iterator &operator++(int);  // NOLINT
+
+    bool operator==(const const_iterator &that) const {
+      return ptr_ == that.ptr_ && cur_node_ == that.cur_node_ &&
+             cur_edge_ == that.cur_edge_;
+    }
+    bool operator!=(const const_iterator &that) const {
+      return !(*this == that);
+    }
+
+   private:
+    const_iterator(const NestedTrieTopology *ptr, int cur_node)
+        : ptr_(ptr), cur_node_(cur_node) {
+      SetProperCurEdge();
+    }
+
+    void SetProperCurEdge() {
+      if (cur_node_ < ptr_->NumNodes())
+        cur_edge_ = ptr_->nodes_[cur_node_]->begin();
+      else
+        cur_edge_ = ptr_->nodes_[0]->begin();
+    }
+
+    void UpdateStub() {
+      stub_.first = ParentLabel<L>(cur_node_, cur_edge_->first);
+      stub_.second = cur_edge_->second;
+    }
+
+    const NestedTrieTopology *ptr_;
+    int cur_node_;
+    typename NextMap::const_iterator cur_edge_;
+    value_type stub_;
+  };
+
+  NestedTrieTopology();
+  NestedTrieTopology(const NestedTrieTopology &that);
+  ~NestedTrieTopology();
+  void swap(NestedTrieTopology &that);
+  NestedTrieTopology &operator=(const NestedTrieTopology &that);
+  bool operator==(const NestedTrieTopology &that) const;
+  bool operator!=(const NestedTrieTopology &that) const;
+
+  int Root() const { return 0; }
+  size_t NumNodes() const { return nodes_.size(); }
+  int Insert(int parent, const L &label);
+  int Find(int parent, const L &label) const;
+  const NextMap &ChildrenOf(int parent) const { return *nodes_[parent]; }
+
+  std::istream &Read(std::istream &strm);         // NOLINT
+  std::ostream &Write(std::ostream &strm) const;  // NOLINT
+
+  const_iterator begin() const { return const_iterator(this, 0); }
+  const_iterator end() const { return const_iterator(this, NumNodes()); }
+
+ private:
+  std::vector<NextMap *>
+      nodes_;  // Use pointers to avoid copying the maps when the
+               // vector grows
+};
+
+template <class L, class H>
+NestedTrieTopology<L, H>::NestedTrieTopology() {
+  nodes_.push_back(new NextMap);
+}
+
+template <class L, class H>
+NestedTrieTopology<L, H>::NestedTrieTopology(const NestedTrieTopology &that) {
+  nodes_.reserve(that.nodes_.size());
+  for (size_t i = 0; i < that.nodes_.size(); ++i) {
+    NextMap *node = that.nodes_[i];
+    nodes_.push_back(new NextMap(*node));
+  }
+}
+
+template <class L, class H>
+NestedTrieTopology<L, H>::~NestedTrieTopology() {
+  for (size_t i = 0; i < nodes_.size(); ++i) {
+    NextMap *node = nodes_[i];
+    delete node;
+  }
+}
+
+// TODO(wuke): std::swap compatibility
+template <class L, class H>
+inline void NestedTrieTopology<L, H>::swap(NestedTrieTopology &that) {
+  nodes_.swap(that.nodes_);
+}
+
+template <class L, class H>
+inline NestedTrieTopology<L, H> &NestedTrieTopology<L, H>::operator=(
+    const NestedTrieTopology &that) {
+  NestedTrieTopology copy(that);
+  swap(copy);
+  return *this;
+}
+
+template <class L, class H>
+inline bool NestedTrieTopology<L, H>::operator==(
+    const NestedTrieTopology &that) const {
+  if (NumNodes() != that.NumNodes()) return false;
+  for (int i = 0; i < NumNodes(); ++i)
+    if (ChildrenOf(i) != that.ChildrenOf(i)) return false;
+  return true;
+}
+
+template <class L, class H>
+inline bool NestedTrieTopology<L, H>::operator!=(
+    const NestedTrieTopology &that) const {
+  return !(*this == that);
+}
+
+template <class L, class H>
+inline int NestedTrieTopology<L, H>::Insert(int parent, const L &label) {
+  int ret = Find(parent, label);
+  if (ret == kNoTrieNodeId) {
+    ret = NumNodes();
+    (*nodes_[parent])[label] = ret;
+    nodes_.push_back(new NextMap);
+  }
+  return ret;
+}
+
+template <class L, class H>
+inline int NestedTrieTopology<L, H>::Find(int parent, const L &label) const {
+  typename NextMap::const_iterator it = nodes_[parent]->find(label);
+  return it == nodes_[parent]->end() ? kNoTrieNodeId : it->second;
+}
+
+template <class L, class H>
+inline std::istream &NestedTrieTopology<L, H>::Read(
+    std::istream &strm) {  // NOLINT
+  NestedTrieTopology new_trie;
+  size_t num_nodes;
+  if (!ReadType(strm, &num_nodes)) return strm;
+  for (size_t i = 1; i < num_nodes; ++i) new_trie.nodes_.push_back(new NextMap);
+  for (size_t i = 0; i < num_nodes; ++i) ReadType(strm, new_trie.nodes_[i]);
+  if (strm) swap(new_trie);
+  return strm;
+}
+
+template <class L, class H>
+inline std::ostream &NestedTrieTopology<L, H>::Write(
+    std::ostream &strm) const {  // NOLINT
+  WriteType(strm, NumNodes());
+  for (size_t i = 0; i < NumNodes(); ++i) WriteType(strm, *nodes_[i]);
+  return strm;
+}
+
+template <class L, class H>
+inline typename NestedTrieTopology<L, H>::const_iterator
+    &NestedTrieTopology<L, H>::const_iterator::operator++() {
+  ++cur_edge_;
+  if (cur_edge_ == ptr_->nodes_[cur_node_]->end()) {
+    ++cur_node_;
+    while (cur_node_ < ptr_->NumNodes() && ptr_->nodes_[cur_node_]->empty())
+      ++cur_node_;
+    SetProperCurEdge();
+  }
+  return *this;
+}
+
+template <class L, class H>
+inline typename NestedTrieTopology<L, H>::const_iterator
+    &NestedTrieTopology<L, H>::const_iterator::operator++(int) {  // NOLINT
+  const_iterator save(*this);
+  ++(*this);
+  return save;
+}
+
+// The trie topology in a single hash map; only allows iteration over
+// all the edges in arbitrary order.
+template <class L, class H>
+class FlatTrieTopology {
+ private:
+  typedef std::unordered_map<ParentLabel<L>, int, ParentLabelHash<L, H>>
+      NextMap;
+
+ public:
+  // Iterator over edges as std::pair<ParentLabel<L>, int>
+  typedef typename NextMap::const_iterator const_iterator;
+  typedef L Label;
+  typedef H Hash;
+
+  FlatTrieTopology() {}
+  FlatTrieTopology(const FlatTrieTopology &that) : next_(that.next_) {}
+  template <class T>
+  explicit FlatTrieTopology(const T &that);
+
+  // TODO(wuke): std::swap compatibility
+  void swap(FlatTrieTopology &that) { next_.swap(that.next_); }
+
+  bool operator==(const FlatTrieTopology &that) const {
+    return next_ == that.next_;
+  }
+  bool operator!=(const FlatTrieTopology &that) const {
+    return !(*this == that);
+  }
+
+  int Root() const { return 0; }
+  size_t NumNodes() const { return next_.size() + 1; }
+  int Insert(int parent, const L &label);
+  int Find(int parent, const L &label) const;
+
+  std::istream &Read(std::istream &strm) {  // NOLINT
+    return ReadType(strm, &next_);
+  }
+  std::ostream &Write(std::ostream &strm) const {  // NOLINT
+    return WriteType(strm, next_);
+  }
+
+  const_iterator begin() const { return next_.begin(); }
+  const_iterator end() const { return next_.end(); }
+
+ private:
+  NextMap next_;
+};
+
+template <class L, class H>
+template <class T>
+FlatTrieTopology<L, H>::FlatTrieTopology(const T &that)
+    : next_(that.begin(), that.end()) {}
+
+template <class L, class H>
+inline int FlatTrieTopology<L, H>::Insert(int parent, const L &label) {
+  int ret = Find(parent, label);
+  if (ret == kNoTrieNodeId) {
+    ret = NumNodes();
+    next_[ParentLabel<L>(parent, label)] = ret;
+  }
+  return ret;
+}
+
+template <class L, class H>
+inline int FlatTrieTopology<L, H>::Find(int parent, const L &label) const {
+  typename NextMap::const_iterator it =
+      next_.find(ParentLabel<L>(parent, label));
+  return it == next_.end() ? kNoTrieNodeId : it->second;
+}
+
+// A collection of implementations of the trie data structure. The key
+// is a sequence of type `L` which must be hashable. The value is of
+// `V` which must be default constructible and copyable. In addition,
+// a value object is stored for each node in the trie therefore
+// copying `V` should be cheap.
+//
+// One can access the store values with an integer node id, using the
+// [] operator. A valid node id can be obtained by the following ways:
+//
+// 1. Using the `Root()` method to get the node id of the root.
+//
+// 2. Iterating through 0 to `NumNodes() - 1`. The node ids are dense
+// so every integer in this range is a valid node id.
+//
+// 3. Using the node id returned from a successful `Insert()` or
+// `Find()` call.
+//
+// 4. Iterating over the trie edges with an `EdgeIterator` and using
+// the node ids returned from its `Parent()` and `Child()` methods.
+//
+// Below is an example of inserting keys into the trie:
+//
+//   const string words[] = {"hello", "health", "jello"};
+//   Trie<char, bool> dict;
+//   for (auto word : words) {
+//     int cur = dict.Root();
+//     for (char c : word) {
+//       cur = dict.Insert(cur, c);
+//     }
+//     dict[cur] = true;
+//   }
+//
+// And the following is an example of looking up the longest prefix of
+// a string using the trie constructed above:
+//
+//   string query = "healed";
+//   size_t prefix_length = 0;
+//   int cur = dict.Find(dict.Root(), query[prefix_length]);
+//   while (prefix_length < query.size() &&
+//     cur != Trie<char, bool>::kNoNodeId) {
+//     ++prefix_length;
+//     cur = dict.Find(cur, query[prefix_length]);
+//   }
+template <class L, class V, class T>
+class MutableTrie {
+ public:
+  template <class LL, class VV, class TT>
+  friend class MutableTrie;
+
+  typedef L Label;
+  typedef V Value;
+  typedef T Topology;
+
+  // Constructs a trie with only the root node.
+  MutableTrie() {}
+
+  // Conversion from another trie of a possiblly different
+  // topology. The underlying topology must supported conversion.
+  template <class S>
+  explicit MutableTrie(const MutableTrie<L, V, S> &that)
+      : topology_(that.topology_), values_(that.values_) {}
+
+  // TODO(wuke): std::swap compatibility
+  void swap(MutableTrie &that) {
+    topology_.swap(that.topology_);
+    values_.swap(that.values_);
+  }
+
+  int Root() const { return topology_.Root(); }
+  size_t NumNodes() const { return topology_.NumNodes(); }
+
+  // Inserts an edge with given `label` at node `parent`. Returns the
+  // child node id. If the node already exists, returns the node id
+  // right away.
+  int Insert(int parent, const L &label) {
+    int ret = topology_.Insert(parent, label);
+    values_.resize(NumNodes());
+    return ret;
+  }
+
+  // Finds the node id of the node from `parent` via `label`. Returns
+  // `kNoTrieNodeId` when such a node does not exist.
+  int Find(int parent, const L &label) const {
+    return topology_.Find(parent, label);
+  }
+
+  const T &TrieTopology() const { return topology_; }
+
+  // Accesses the value stored for the given node.
+  V &operator[](int node_id) { return values_[node_id]; }
+  const V &operator[](int node_id) const { return values_[node_id]; }
+
+  // Comparison by content
+  bool operator==(const MutableTrie &that) const {
+    return topology_ == that.topology_ && values_ == that.values_;
+  }
+
+  bool operator!=(const MutableTrie &that) const { return !(*this == that); }
+
+  std::istream &Read(std::istream &strm) {  // NOLINT
+    ReadType(strm, &topology_);
+    ReadType(strm, &values_);
+    return strm;
+  }
+  std::ostream &Write(std::ostream &strm) const {  // NOLINT
+    WriteType(strm, topology_);
+    WriteType(strm, values_);
+    return strm;
+  }
+
+ private:
+  T topology_;
+  std::vector<V> values_;
+};
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_LINEAR_TRIE_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/compose.h b/projects/llm_framework/include/fst/extensions/mpdt/compose.h
new file mode 100644
index 00000000..47714e37
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/mpdt/compose.h
@@ -0,0 +1,267 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Compose an MPDT and an FST.
+
+#ifndef FST_EXTENSIONS_MPDT_COMPOSE_H_
+#define FST_EXTENSIONS_MPDT_COMPOSE_H_
+
+#include <list>
+
+#include <fst/extensions/mpdt/mpdt.h>
+#include <fst/extensions/pdt/compose.h>
+#include <fst/compose.h>
+
+namespace fst {
+
+template <class Filter>
+class MPdtParenFilter {
+ public:
+  using FST1 = typename Filter::FST1;
+  using FST2 = typename Filter::FST2;
+  using Arc = typename Filter::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Matcher1 = typename Filter::Matcher1;
+  using Matcher2 = typename Filter::Matcher2;
+
+  using StackId = StateId;
+  using ParenStack = internal::MPdtStack<StackId, Label>;
+  using FilterState1 = typename Filter::FilterState;
+  using FilterState2 = IntegerFilterState<StackId>;
+  using FilterState = PairFilterState<FilterState1, FilterState2>;
+
+  MPdtParenFilter(const FST1 &fst1, const FST2 &fst2,
+                  Matcher1 *matcher1 = nullptr, Matcher2 *matcher2 = nullptr,
+                  const std::vector<std::pair<Label, Label>> *parens = nullptr,
+                  const std::vector<Label> *assignments = nullptr,
+                  bool expand = false, bool keep_parens = true)
+      : filter_(fst1, fst2, matcher1, matcher2),
+        parens_(parens ? *parens : std::vector<std::pair<Label, Label>>()),
+        assignments_(assignments ? *assignments : std::vector<Label>()),
+        expand_(expand),
+        keep_parens_(keep_parens),
+        fs_(FilterState::NoState()),
+        stack_(parens_, assignments_),
+        paren_id_(-1) {
+    if (parens) {
+      for (const auto &pair : *parens) {
+        parens_.push_back(pair);
+        GetMatcher1()->AddOpenParen(pair.first);
+        GetMatcher2()->AddOpenParen(pair.first);
+        if (!expand_) {
+          GetMatcher1()->AddCloseParen(pair.second);
+          GetMatcher2()->AddCloseParen(pair.second);
+        }
+      }
+    }
+  }
+
+  MPdtParenFilter(const MPdtParenFilter &filter, bool safe = false)
+      : filter_(filter.filter_, safe),
+        parens_(filter.parens_),
+        expand_(filter.expand_),
+        keep_parens_(filter.keep_parens_),
+        fs_(FilterState::NoState()),
+        stack_(filter.parens_, filter.assignments_),
+        paren_id_(-1) {}
+
+  FilterState Start() const {
+    return FilterState(filter_.Start(), FilterState2(0));
+  }
+
+  void SetState(StateId s1, StateId s2, const FilterState &fs) {
+    fs_ = fs;
+    filter_.SetState(s1, s2, fs_.GetState1());
+    if (!expand_) return;
+    const auto paren_id = stack_.Top(fs.GetState2().GetState());
+    if (paren_id != paren_id_) {
+      if (paren_id_ != -1) {
+        GetMatcher1()->RemoveCloseParen(parens_[paren_id_].second);
+        GetMatcher2()->RemoveCloseParen(parens_[paren_id_].second);
+      }
+      paren_id_ = paren_id;
+      if (paren_id_ != -1) {
+        GetMatcher1()->AddCloseParen(parens_[paren_id_].second);
+        GetMatcher2()->AddCloseParen(parens_[paren_id_].second);
+      }
+    }
+  }
+
+  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
+    const auto fs1 = filter_.FilterArc(arc1, arc2);
+    const auto &fs2 = fs_.GetState2();
+    if (fs1 == FilterState1::NoState()) return FilterState::NoState();
+    if (arc1->olabel == kNoLabel && arc2->ilabel) {  // arc2 parentheses.
+      if (keep_parens_) {
+        arc1->ilabel = arc2->ilabel;
+      } else if (arc2->ilabel) {
+        arc2->olabel = arc1->ilabel;
+      }
+      return FilterParen(arc2->ilabel, fs1, fs2);
+    } else if (arc2->ilabel == kNoLabel && arc1->olabel) {  // arc1 parentheses
+      if (keep_parens_) {
+        arc2->olabel = arc1->olabel;
+      } else {
+        arc1->ilabel = arc2->olabel;
+      }
+      return FilterParen(arc1->olabel, fs1, fs2);
+    } else {
+      return FilterState(fs1, fs2);
+    }
+  }
+
+  void FilterFinal(Weight *w1, Weight *w2) const {
+    if (fs_.GetState2().GetState() != 0) *w1 = Weight::Zero();
+    filter_.FilterFinal(w1, w2);
+  }
+
+  // Returns respective matchers; ownership stays with filter.
+
+  Matcher1 *GetMatcher1() { return filter_.GetMatcher1(); }
+
+  Matcher2 *GetMatcher2() { return filter_.GetMatcher2(); }
+
+  uint64 Properties(uint64 iprops) const {
+    const auto oprops = filter_.Properties(iprops);
+    return oprops & kILabelInvariantProperties & kOLabelInvariantProperties;
+  }
+
+ private:
+  const FilterState FilterParen(Label label, const FilterState1 &fs1,
+                                const FilterState2 &fs2) const {
+    if (!expand_) return FilterState(fs1, fs2);
+    const auto stack_id = stack_.Find(fs2.GetState(), label);
+    if (stack_id < 0) {
+      return FilterState::NoState();
+    } else {
+      return FilterState(fs1, FilterState2(stack_id));
+    }
+  }
+
+  Filter filter_;
+  std::vector<std::pair<Label, Label>> parens_;
+  std::vector<Label> assignments_;
+  bool expand_;       // Expands to FST?
+  bool keep_parens_;  // Retains parentheses in output?
+  FilterState fs_;    // Current filter state.
+  mutable ParenStack stack_;
+  ssize_t paren_id_;
+};
+
+// Class to setup composition options for MPDT composition. Default is to take
+// the MPDT as the first composition argument.
+template <class Arc, bool left_pdt = true>
+class MPdtComposeFstOptions
+    : public ComposeFstOptions<Arc, ParenMatcher<Fst<Arc>>,
+                               MPdtParenFilter<AltSequenceComposeFilter<
+                                   ParenMatcher<Fst<Arc>> >> > {
+ public:
+  using Label = typename Arc::Label;
+  using MPdtMatcher = ParenMatcher<Fst<Arc>>;
+  using MPdtFilter = MPdtParenFilter<AltSequenceComposeFilter<MPdtMatcher>>;
+
+  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::matcher1;
+  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::matcher2;
+  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::filter;
+
+  MPdtComposeFstOptions(const Fst<Arc> &ifst1,
+                        const std::vector<std::pair<Label, Label>> &parens,
+                        const std::vector<typename Arc::Label> &assignments,
+                        const Fst<Arc> &ifst2, bool expand = false,
+                        bool keep_parens = true) {
+    matcher1 = new MPdtMatcher(ifst1, MATCH_OUTPUT, kParenList);
+    matcher2 = new MPdtMatcher(ifst2, MATCH_INPUT, kParenLoop);
+    filter = new MPdtFilter(ifst1, ifst2, matcher1, matcher2, &parens,
+                            &assignments, expand, keep_parens);
+  }
+};
+
+// Class to setup composition options for PDT with FST composition.
+// Specialization is for the FST as the first composition argument.
+template <class Arc>
+class MPdtComposeFstOptions<Arc, false>
+    : public ComposeFstOptions<
+          Arc, ParenMatcher<Fst<Arc>>,
+          MPdtParenFilter<SequenceComposeFilter<ParenMatcher<Fst<Arc>> >> > {
+ public:
+  using Label = typename Arc::Label;
+  using MPdtMatcher = ParenMatcher<Fst<Arc>>;
+  using MPdtFilter = MPdtParenFilter<SequenceComposeFilter<MPdtMatcher>>;
+
+  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::matcher1;
+  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::matcher2;
+  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::filter;
+
+  MPdtComposeFstOptions(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
+                        const std::vector<std::pair<Label, Label>> &parens,
+                        const std::vector<typename Arc::Label> &assignments,
+                        bool expand = false, bool keep_parens = true) {
+    matcher1 = new MPdtMatcher(ifst1, MATCH_OUTPUT, kParenLoop);
+    matcher2 = new MPdtMatcher(ifst2, MATCH_INPUT, kParenList);
+    filter = new MPdtFilter(ifst1, ifst2, matcher1, matcher2, &parens,
+                            &assignments, expand, keep_parens);
+  }
+};
+
+struct MPdtComposeOptions {
+  bool connect;                  // Connect output?
+  PdtComposeFilter filter_type;  // Which pre-defined filter to use.
+
+  explicit MPdtComposeOptions(bool connect = true,
+                              PdtComposeFilter filter_type = PAREN_FILTER)
+      : connect(connect), filter_type(filter_type) {}
+};
+
+// Composes multi-pushdown transducer (MPDT) encoded as an FST (1st arg) and an
+// FST (2nd arg) with the result also an MPDT encoded as an FST (3rd arg). In
+// theMPDTs, some transitions are labeled with open or close parentheses (and
+// associated with a stack). To be interpreted as an MPDT, the parents on each
+// stack must balance on a path (see MPdtExpand()). The open-close parenthesis
+// label pairs are passed using the parens arguments, and the stack assignments
+// are passed using the assignments argument.
+template <class Arc>
+void Compose(
+    const Fst<Arc> &ifst1,
+    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
+        &parens,
+    const std::vector<typename Arc::Label> &assignments, const Fst<Arc> &ifst2,
+    MutableFst<Arc> *ofst,
+    const MPdtComposeOptions &opts = MPdtComposeOptions()) {
+  bool expand = opts.filter_type != PAREN_FILTER;
+  bool keep_parens = opts.filter_type != EXPAND_FILTER;
+  MPdtComposeFstOptions<Arc, true> copts(ifst1, parens, assignments, ifst2,
+                                         expand, keep_parens);
+  copts.gc_limit = 0;
+  *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
+  if (opts.connect) Connect(ofst);
+}
+
+// Composes an FST (1st arg) and a multi-pushdown transducer (MPDT) encoded as
+// an FST (2nd arg) with the result also an MPDT encoded as an FST (3rd arg).
+// In the MPDTs, some transitions are labeled with open or close parentheses
+// (and associated with a stack). To be interpreted as an MPDT, the parents on
+// each stack must balance on a path (see MPdtExpand()). The open-close
+// parenthesis label pairs are passed using the parens arguments, and the stack
+// assignments are passed using the assignments argument.
+template <class Arc>
+void Compose(
+    const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
+    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
+        &parens,
+    const std::vector<typename Arc::Label> &assignments, MutableFst<Arc> *ofst,
+    const MPdtComposeOptions &opts = MPdtComposeOptions()) {
+  bool expand = opts.filter_type != PAREN_FILTER;
+  bool keep_parens = opts.filter_type != EXPAND_FILTER;
+  MPdtComposeFstOptions<Arc, false> copts(ifst1, ifst2, parens, assignments,
+                                          expand, keep_parens);
+  copts.gc_limit = 0;
+  *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
+  if (opts.connect) Connect(ofst);
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_MPDT_COMPOSE_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/expand.h b/projects/llm_framework/include/fst/extensions/mpdt/expand.h
new file mode 100644
index 00000000..0e06cdce
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/mpdt/expand.h
@@ -0,0 +1,335 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Expands an MPDT to an FST.
+
+#ifndef FST_EXTENSIONS_MPDT_EXPAND_H_
+#define FST_EXTENSIONS_MPDT_EXPAND_H_
+
+#include <vector>
+
+#include <fst/extensions/mpdt/mpdt.h>
+#include <fst/extensions/pdt/paren.h>
+#include <fst/cache.h>
+#include <fst/mutable-fst.h>
+#include <fst/queue.h>
+#include <fst/state-table.h>
+#include <fst/test-properties.h>
+
+namespace fst {
+
+template <class Arc>
+struct MPdtExpandFstOptions : public CacheOptions {
+  bool keep_parentheses;
+  internal::MPdtStack<typename Arc::StateId, typename Arc::Label> *stack;
+  PdtStateTable<typename Arc::StateId, typename Arc::StateId> *state_table;
+
+  MPdtExpandFstOptions(
+      const CacheOptions &opts = CacheOptions(), bool kp = false,
+      internal::MPdtStack<typename Arc::StateId, typename Arc::Label> *s =
+          nullptr,
+      PdtStateTable<typename Arc::StateId, typename Arc::StateId> *st = nullptr)
+      : CacheOptions(opts), keep_parentheses(kp), stack(s), state_table(st) {}
+};
+
+// Properties for an expanded PDT.
+inline uint64 MPdtExpandProperties(uint64 inprops) {
+  return inprops & (kAcceptor | kAcyclic | kInitialAcyclic | kUnweighted);
+}
+
+namespace internal {
+
+// Implementation class for ExpandFst
+template <class Arc>
+class MPdtExpandFstImpl : public CacheImpl<Arc> {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using StackId = StateId;
+  using StateTuple = PdtStateTuple<StateId, StackId>;
+  using ParenStack = internal::MPdtStack<StateId, Label>;
+
+  using FstImpl<Arc>::SetType;
+  using FstImpl<Arc>::SetProperties;
+  using FstImpl<Arc>::Properties;
+  using FstImpl<Arc>::SetInputSymbols;
+  using FstImpl<Arc>::SetOutputSymbols;
+
+  using CacheBaseImpl<CacheState<Arc>>::PushArc;
+  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
+  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
+  using CacheBaseImpl<CacheState<Arc>>::HasStart;
+  using CacheBaseImpl<CacheState<Arc>>::SetArcs;
+  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
+  using CacheBaseImpl<CacheState<Arc>>::SetStart;
+
+  MPdtExpandFstImpl(const Fst<Arc> &fst,
+                    const std::vector<std::pair<Label, Label>> &parens,
+                    const std::vector<Label> &assignments,
+                    const MPdtExpandFstOptions<Arc> &opts)
+      : CacheImpl<Arc>(opts),
+        fst_(fst.Copy()),
+        stack_(opts.stack ? opts.stack : new ParenStack(parens, assignments)),
+        state_table_(opts.state_table ? opts.state_table
+                                      : new PdtStateTable<StateId, StackId>()),
+        own_stack_(!opts.stack),
+        own_state_table_(!opts.state_table),
+        keep_parentheses_(opts.keep_parentheses) {
+    SetType("expand");
+    const auto props = fst.Properties(kFstProperties, false);
+    SetProperties(MPdtExpandProperties(props), kCopyProperties);
+    SetInputSymbols(fst.InputSymbols());
+    SetOutputSymbols(fst.OutputSymbols());
+  }
+
+  MPdtExpandFstImpl(const MPdtExpandFstImpl &impl)
+      : CacheImpl<Arc>(impl),
+        fst_(impl.fst_->Copy(true)),
+        stack_(new ParenStack(*impl.stack_)),
+        state_table_(new PdtStateTable<StateId, StackId>()),
+        own_stack_(true),
+        own_state_table_(true),
+        keep_parentheses_(impl.keep_parentheses_) {
+    SetType("expand");
+    SetProperties(impl.Properties(), kCopyProperties);
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+  }
+
+  ~MPdtExpandFstImpl() override {
+    if (own_stack_) delete stack_;
+    if (own_state_table_) delete state_table_;
+  }
+
+  StateId Start() {
+    if (!HasStart()) {
+      const auto s = fst_->Start();
+      if (s == kNoStateId) return kNoStateId;
+      const StateTuple tuple(s, 0);
+      const auto start = state_table_->FindState(tuple);
+      SetStart(start);
+    }
+    return CacheImpl<Arc>::Start();
+  }
+
+  Weight Final(StateId s) {
+    if (!HasFinal(s)) {
+      const auto &tuple = state_table_->Tuple(s);
+      const auto weight = fst_->Final(tuple.state_id);
+      SetFinal(s,
+               (weight != Weight::Zero() && tuple.stack_id == 0)
+                   ? weight
+                   : Weight::Zero());
+    }
+    return CacheImpl<Arc>::Final(s);
+  }
+
+  size_t NumArcs(StateId s) {
+    if (!HasArcs(s)) ExpandState(s);
+    return CacheImpl<Arc>::NumArcs(s);
+  }
+
+  size_t NumInputEpsilons(StateId s) {
+    if (!HasArcs(s)) ExpandState(s);
+    return CacheImpl<Arc>::NumInputEpsilons(s);
+  }
+
+  size_t NumOutputEpsilons(StateId s) {
+    if (!HasArcs(s)) ExpandState(s);
+    return CacheImpl<Arc>::NumOutputEpsilons(s);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
+    if (!HasArcs(s)) ExpandState(s);
+    CacheImpl<Arc>::InitArcIterator(s, data);
+  }
+
+  // Computes the outgoing transitions from a state, creating new destination
+  // states as needed.
+  void ExpandState(StateId s) {
+    const auto tuple = state_table_->Tuple(s);
+    for (ArcIterator<Fst<Arc>> aiter(*fst_, tuple.state_id); !aiter.Done();
+         aiter.Next()) {
+      auto arc = aiter.Value();
+      const auto stack_id = stack_->Find(tuple.stack_id, arc.ilabel);
+      if (stack_id == -1) {
+        continue;  // Non-matching close parenthesis.
+      } else if ((stack_id != tuple.stack_id) && !keep_parentheses_) {
+        arc.ilabel = arc.olabel = 0;  // Stack push/pop.
+      }
+      const StateTuple ntuple(arc.nextstate, stack_id);
+      arc.nextstate = state_table_->FindState(ntuple);
+      PushArc(s, arc);
+    }
+    SetArcs(s);
+  }
+
+  const ParenStack &GetStack() const { return *stack_; }
+
+  const PdtStateTable<StateId, StackId> &GetStateTable() const {
+    return *state_table_;
+  }
+
+ private:
+  std::unique_ptr<const Fst<Arc>> fst_;
+  ParenStack *stack_;
+  PdtStateTable<StateId, StackId> *state_table_;
+  const bool own_stack_;
+  const bool own_state_table_;
+  const bool keep_parentheses_;
+
+  MPdtExpandFstImpl &operator=(const MPdtExpandFstImpl &) = delete;
+};
+
+}  // namespace internal
+
+// Expands a multi-pushdown transducer (MPDT) encoded as an FST into an FST.
+// This version is a delayed FST. In the MPDT, some transitions are labeled with
+// open or close parentheses. To be interpreted as an MPDT, the parens for each
+// stack must balance on a path. The open-close parenthesis label
+// pairs are passed using the parens argument, and the assignment of those pairs
+// to stacks is passed using the assignments argument. Expansion enforces the
+// parenthesis constraints. The MPDT must be
+// expandable as an FST.
+//
+// This class attaches interface to implementation and handles
+// reference counting, delegating most methods to ImplToFst.
+template <class A>
+class MPdtExpandFst : public ImplToFst<internal::MPdtExpandFstImpl<A>> {
+ public:
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using StackId = StateId;
+  using ParenStack = internal::MPdtStack<StackId, Label>;
+  using Store = DefaultCacheStore<Arc>;
+  using State = typename Store::State;
+  using Impl = internal::MPdtExpandFstImpl<Arc>;
+
+  friend class ArcIterator<MPdtExpandFst<Arc>>;
+  friend class StateIterator<MPdtExpandFst<Arc>>;
+
+  MPdtExpandFst(const Fst<Arc> &fst,
+                const std::vector<std::pair<Label, Label>> &parens,
+                const std::vector<Label> &assignments)
+      : ImplToFst<Impl>(std::make_shared<Impl>(fst, parens, assignments,
+                                               MPdtExpandFstOptions<Arc>())) {}
+
+  MPdtExpandFst(const Fst<Arc> &fst,
+                const std::vector<std::pair<Label, Label>> &parens,
+                const std::vector<Label> &assignments,
+                const MPdtExpandFstOptions<Arc> &opts)
+      : ImplToFst<Impl>(
+            std::make_shared<Impl>(fst, parens, assignments, opts)) {}
+
+  // See Fst<>::Copy() for doc.
+  MPdtExpandFst(const MPdtExpandFst<Arc> &fst, bool safe = false)
+      : ImplToFst<Impl>(fst, safe) {}
+
+  // Get a copy of this ExpandFst. See Fst<>::Copy() for further doc.
+  MPdtExpandFst<Arc> *Copy(bool safe = false) const override {
+    return new MPdtExpandFst<A>(*this, safe);
+  }
+
+  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
+    GetMutableImpl()->InitArcIterator(s, data);
+  }
+
+  const ParenStack &GetStack() const { return GetImpl()->GetStack(); }
+
+  const PdtStateTable<StateId, StackId> &GetStateTable() const {
+    return GetImpl()->GetStateTable();
+  }
+
+ private:
+  using ImplToFst<Impl>::GetImpl;
+  using ImplToFst<Impl>::GetMutableImpl;
+
+  void operator=(const MPdtExpandFst &) = delete;
+};
+
+// Specialization for MPdtExpandFst.
+template <class Arc>
+class StateIterator<MPdtExpandFst<Arc>>
+    : public CacheStateIterator<MPdtExpandFst<Arc>> {
+ public:
+  explicit StateIterator(const MPdtExpandFst<Arc> &fst)
+      : CacheStateIterator<MPdtExpandFst<Arc>>(fst, fst.GetMutableImpl()) {}
+};
+
+// Specialization for MPdtExpandFst.
+template <class Arc>
+class ArcIterator<MPdtExpandFst<Arc>>
+    : public CacheArcIterator<MPdtExpandFst<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const MPdtExpandFst<Arc> &fst, StateId s)
+      : CacheArcIterator<MPdtExpandFst<Arc>>(fst.GetMutableImpl(), s) {
+    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->ExpandState(s);
+  }
+};
+
+template <class Arc>
+inline void MPdtExpandFst<Arc>::InitStateIterator(
+    StateIteratorData<Arc> *data) const {
+  data->base = new StateIterator<MPdtExpandFst<Arc>>(*this);
+}
+
+struct MPdtExpandOptions {
+  bool connect;
+  bool keep_parentheses;
+
+  explicit MPdtExpandOptions(bool connect = true, bool keep_parentheses = false)
+      : connect(connect), keep_parentheses(keep_parentheses) {}
+};
+
+// Expands a multi-pushdown transducer (MPDT) encoded as an FST into an FST.
+// This version writes the expanded PDT to a mutable FST. In the MPDT, some
+// transitions are labeled with open or close parentheses. To be interpreted as
+// an MPDT, the parens for each stack must balance on a path. The open-close
+// parenthesis label pair sets are passed using the parens argument, and the
+// assignment of those pairs to stacks is passed using the assignments argument.
+// The expansion enforces the parenthesis constraints. The MPDT must be
+// expandable as an FST.
+template <class Arc>
+void Expand(const Fst<Arc> &ifst,
+            const std::vector<
+            std::pair<typename Arc::Label, typename Arc::Label>> &parens,
+            const std::vector<typename Arc::Label> &assignments,
+            MutableFst<Arc> *ofst, const MPdtExpandOptions &opts) {
+  MPdtExpandFstOptions<Arc> eopts;
+  eopts.gc_limit = 0;
+  eopts.keep_parentheses = opts.keep_parentheses;
+  *ofst = MPdtExpandFst<Arc>(ifst, parens, assignments, eopts);
+  if (opts.connect) Connect(ofst);
+}
+
+// Expands a multi-pushdown transducer (MPDT) encoded as an FST into an FST.
+// This version writes the expanded PDT to a mutable FST. In the MPDT, some
+// transitions are labeled with open or close parentheses. To be interpreted as
+// an MPDT, the parens for each stack must balance on a path. The open-close
+// parenthesis label pair sets are passed using the parens argument, and the
+// assignment of those pairs to stacks is passed using the assignments argument.
+// The expansion enforces the parenthesis constraints. The MPDT must be
+// expandable as an FST.
+template <class Arc>
+void Expand(const Fst<Arc> &ifst,
+            const std::vector<std::pair<typename Arc::Label,
+            typename Arc::Label>> &parens,
+            const std::vector<typename Arc::Label> &assignments,
+            MutableFst<Arc> *ofst, bool connect = true,
+            bool keep_parentheses = false) {
+  const MPdtExpandOptions opts(connect, keep_parentheses);
+  Expand(ifst, parens, assignments, ofst, opts);
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_MPDT_EXPAND_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/info.h b/projects/llm_framework/include/fst/extensions/mpdt/info.h
new file mode 100644
index 00000000..512fcfa8
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/mpdt/info.h
@@ -0,0 +1,190 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Prints information about an MPDT.
+
+#ifndef FST_EXTENSIONS_MPDT_INFO_H_
+#define FST_EXTENSIONS_MPDT_INFO_H_
+
+#include <unordered_map>
+#include <vector>
+
+#include <fst/extensions/mpdt/mpdt.h>
+#include <fst/fst.h>
+
+namespace fst {
+
+// Compute various information about MPDTs, helper class for mpdtinfo.cc.
+template <class Arc, typename Arc::Label nlevels = 2>
+class MPdtInfo {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  MPdtInfo(const Fst<Arc> &fst,
+           const std::vector<std::pair<Label, Label>> &parens,
+           const std::vector<Label> &assignments);
+
+  const string &FstType() const { return fst_type_; }
+
+  const string &ArcType() const { return Arc::Type(); }
+
+  int64 NumStates() const { return nstates_; }
+
+  int64 NumArcs() const { return narcs_; }
+
+  int64 NumLevels() const { return nlevels; }
+
+  int64 NumOpenParens(Label level) const { return nopen_parens_[level]; }
+
+  int64 NumCloseParens(Label level) const { return nclose_parens_[level]; }
+
+  int64 NumUniqueOpenParens(Label level) const {
+    return nuniq_open_parens_[level];
+  }
+
+  int64 NumUniqueCloseParens(Label level) const {
+    return nuniq_close_parens_[level];
+  }
+  int64 NumOpenParenStates(Label level) const {
+    return nopen_paren_states_[level];
+  }
+
+  int64 NumCloseParenStates(Label level) const {
+    return nclose_paren_states_[level];
+  }
+
+  void Print();
+
+ private:
+  string fst_type_;
+  int64 nstates_;
+  int64 narcs_;
+  int64 nopen_parens_[nlevels];
+  int64 nclose_parens_[nlevels];
+  int64 nuniq_open_parens_[nlevels];
+  int64 nuniq_close_parens_[nlevels];
+  int64 nopen_paren_states_[nlevels];
+  int64 nclose_paren_states_[nlevels];
+  bool error_;
+};
+
+template <class Arc, typename Arc::Label nlevels>
+MPdtInfo<Arc, nlevels>::MPdtInfo(
+    const Fst<Arc> &fst,
+    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
+        &parens,
+    const std::vector<typename Arc::Label> &assignments)
+    : fst_type_(fst.Type()), nstates_(0), narcs_(0), error_(false) {
+  std::unordered_map<Label, size_t> paren_map;
+  std::unordered_set<Label> paren_set;
+  std::unordered_map<Label, int> paren_levels;
+  std::unordered_set<StateId> open_paren_state_set;
+  std::unordered_set<StateId> close_paren_state_set;
+  if (parens.size() != assignments.size()) {
+    FSTERROR() << "MPdtInfo: Parens of different size from assignments";
+    error_ = true;
+    return;
+  }
+  for (Label i = 0; i < assignments.size(); ++i) {
+    // Assignments here start at 0, so assuming the human-readable version has
+    // them starting at 1, we should subtract 1 here.
+    Label level = assignments[i] - 1;
+    if (level < 0 || level >= nlevels) {
+      FSTERROR() << "MPdtInfo: Specified level " << level << " out of bounds";
+      error_ = true;
+      return;
+    }
+    const auto &pair = parens[i];
+    paren_levels[pair.first] = level;
+    paren_levels[pair.second] = level;
+    paren_map[pair.first] = i;
+    paren_map[pair.second] = i;
+  }
+  for (Label i = 0; i < nlevels; ++i) {
+    nopen_parens_[i] = 0;
+    nclose_parens_[i] = 0;
+    nuniq_open_parens_[i] = 0;
+    nuniq_close_parens_[i] = 0;
+    nopen_paren_states_[i] = 0;
+    nclose_paren_states_[i] = 0;
+  }
+  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
+    ++nstates_;
+    const auto s = siter.Value();
+    for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      ++narcs_;
+      const auto it = paren_map.find(arc.ilabel);
+      if (it != paren_map.end()) {
+        const auto open_paren = parens[it->second].first;
+        const auto close_paren = parens[it->second].second;
+        const auto level = paren_levels[arc.ilabel];
+        if (arc.ilabel == open_paren) {
+          ++nopen_parens_[level];
+          if (!paren_set.count(open_paren)) {
+            ++nuniq_open_parens_[level];
+            paren_set.insert(open_paren);
+          }
+          if (!open_paren_state_set.count(arc.nextstate)) {
+            ++nopen_paren_states_[level];
+            open_paren_state_set.insert(arc.nextstate);
+          }
+        } else {
+          ++nclose_parens_[level];
+          if (!paren_set.count(close_paren)) {
+            ++nuniq_close_parens_[level];
+            paren_set.insert(close_paren);
+          }
+          if (!close_paren_state_set.count(s)) {
+            ++nclose_paren_states_[level];
+            close_paren_state_set.insert(s);
+          }
+        }
+      }
+    }
+  }
+}
+
+template <class Arc, typename Arc::Label nlevels>
+void MPdtInfo<Arc, nlevels>::Print() {
+  const auto old = std::cout.setf(std::ios::left);
+  std::cout.width(50);
+  std::cout << "fst type" << FstType() << std::endl;
+  std::cout.width(50);
+  std::cout << "arc type" << ArcType() << std::endl;
+  std::cout.width(50);
+  std::cout << "# of states" << NumStates() << std::endl;
+  std::cout.width(50);
+  std::cout << "# of arcs" << NumArcs() << std::endl;
+  std::cout.width(50);
+  std::cout << "# of levels" << NumLevels() << std::endl;
+  std::cout.width(50);
+  for (typename Arc::Label i = 0; i < nlevels; ++i) {
+    int level = i + 1;
+    std::cout << "# of open parentheses at levelel " << level << "\t"
+              << NumOpenParens(i) << std::endl;
+    std::cout.width(50);
+    std::cout << "# of close parentheses at levelel " << level << "\t"
+              << NumCloseParens(i) << std::endl;
+    std::cout.width(50);
+    std::cout << "# of unique open parentheses at levelel " << level << "\t"
+              << NumUniqueOpenParens(i) << std::endl;
+    std::cout.width(50);
+    std::cout << "# of unique close parentheses at levelel " << level << "\t"
+              << NumUniqueCloseParens(i) << std::endl;
+    std::cout.width(50);
+    std::cout << "# of open parenthesis dest. states at levelel " << level
+              << "\t" << NumOpenParenStates(i) << std::endl;
+    std::cout.width(50);
+    std::cout << "# of close parenthesis source states at levelel " << level
+              << "\t" << NumCloseParenStates(i) << std::endl;
+    std::cout.width(50);
+  }
+  std::cout.setf(old);
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_MPDT_INFO_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/mpdt.h b/projects/llm_framework/include/fst/extensions/mpdt/mpdt.h
new file mode 100644
index 00000000..77ff7574
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/mpdt/mpdt.h
@@ -0,0 +1,357 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Common classes for Multi Pushdown Transducer (MPDT) expansion/traversal.
+
+#ifndef FST_EXTENSIONS_MPDT_MPDT_H_
+#define FST_EXTENSIONS_MPDT_MPDT_H_
+
+#include <array>
+#include <functional>
+#include <map>
+#include <vector>
+
+#include <fst/compat.h>
+#include <fst/extensions/pdt/pdt.h>
+
+namespace fst {
+
+enum MPdtType {
+  MPDT_READ_RESTRICT,   // Can only read from first empty stack
+  MPDT_WRITE_RESTRICT,  // Can only write to first empty stack
+  MPDT_NO_RESTRICT,     // No read-write restrictions
+};
+
+namespace internal {
+
+// PLEASE READ THIS CAREFULLY:
+//
+// When USEVECTOR is set, the stack configurations --- the statewise
+// representation of the StackId's for each substack --- is stored in a vector.
+// I would like to do this using an array for efficiency reasons, thus the
+// definition of StackConfig below. However, while this *works* in that tests
+// pass, etc. It causes a memory leak in the compose and expand tests, evidently
+// in the map[] that is being used to store the mapping between these
+// StackConfigs and the external StackId that the caller sees. There are no
+// memory leaks when I use a vector, only with this StackId. Why there should be
+// memory leaks given that I am not mallocing anything is a mystery. In case you
+// were wondering, clearing the map at the end does not help.
+
+template <typename StackId, typename Level, Level nlevels>
+struct StackConfig {
+  StackConfig() : array_() {}
+
+  StackConfig(const StackConfig<StackId, Level, nlevels> &config) {
+    array_ = config.array_;
+  }
+
+  StackId &operator[](const int index) { return array_[index]; }
+
+  const StackId &operator[](const int index) const { return array_[index]; }
+
+  StackConfig &operator=(const StackConfig<StackId, Level, nlevels> &config) {
+    if (this == &config) return *this;
+    array_ = config.array_;
+    return *this;
+  }
+
+  std::array<StackId, nlevels> array_;
+};
+
+template <typename StackId, typename Level, Level nlevels>
+class CompConfig {
+ public:
+  using Config = StackConfig<StackId, Level, nlevels>;
+
+  bool operator()(const Config &x, const Config &y) const {
+    for (Level level = 0; level < nlevels; ++level) {
+      if (x.array_[level] < y.array_[level]) {
+        return true;
+      } else if (x.array_[level] > y.array_[level]) {
+        return false;
+      }
+    }
+    return false;
+  }
+};
+
+// Defines the KeyPair type used as the key to MPdtStack.paren_id_map_. The hash
+// function is provided as a separate struct to match templating syntax.
+template <typename Level>
+struct KeyPair {
+  Level level;
+  size_t underlying_id;
+
+  KeyPair(Level level, size_t id) : level(level), underlying_id(id) {}
+
+  inline bool operator==(const KeyPair<Level> &rhs) const {
+    return level == rhs.level && underlying_id == rhs.underlying_id;
+  }
+};
+
+template <typename Level>
+struct KeyPairHasher {
+  inline size_t operator()(const KeyPair<Level> &keypair) const {
+    return std::hash<Level>()(keypair.level) ^
+           (std::hash<size_t>()(keypair.underlying_id) << 1);
+  }
+};
+
+template <typename StackId, typename Level, Level nlevels = 2,
+          MPdtType restrict = MPDT_READ_RESTRICT>
+class MPdtStack {
+ public:
+  using Label = Level;
+  using Config = StackConfig<StackId, Level, nlevels>;
+  using ConfigToStackId =
+      std::map<Config, StackId, CompConfig<StackId, Level, nlevels>>;
+
+  MPdtStack(const std::vector<std::pair<Label, Label>> &parens,
+            const std::vector<Level> &assignments);
+
+  MPdtStack(const MPdtStack &mstack);
+
+  ~MPdtStack() {
+    for (Level level = 0; level < nlevels; ++level) delete stacks_[level];
+  }
+
+  StackId Find(StackId stack_id, Label label);
+
+  // For now we do not implement Pop since this is needed only for
+  // ShortestPath().
+
+  // For Top we find the first non-empty config, and find the paren ID of that
+  // (or -1) if there is none, then map that to the external stack_id to return.
+  ssize_t Top(StackId stack_id) const {
+    if (stack_id == -1) return -1;
+    const auto config = InternalStackIds(stack_id);
+    Level level = 0;
+    StackId underlying_id = -1;
+    for (; level < nlevels; ++level) {
+      if (!Empty(config, level)) {
+        underlying_id = stacks_[level]->Top(config[level]);
+        break;
+      }
+    }
+    if (underlying_id == -1) return -1;
+    const auto it = paren_id_map_.find(KeyPair<Level>(level, underlying_id));
+    if (it == paren_id_map_.end()) return -1;  // NB: shouldn't happen.
+    return it->second;
+  }
+
+  ssize_t ParenId(Label label) const {
+    const auto it = paren_map_.find(label);
+    return it != paren_map_.end() ? it->second : -1;
+  }
+
+  // TODO(rws): For debugging purposes only: remove later.
+  string PrintConfig(const Config &config) const {
+    string result = "[";
+    for (Level i = 0; i < nlevels; ++i) {
+      char s[128];
+      snprintf(s, sizeof(s), "%d", config[i]);
+      result += string(s);
+      if (i < nlevels - 1) result += ", ";
+    }
+    result += "]";
+    return result;
+  }
+
+  bool Error() { return error_; }
+
+  // Each component stack has an internal stack ID for a given configuration and
+  // label.
+  // This function maps a configuration of those to the stack ID the caller
+  // sees.
+  inline StackId ExternalStackId(const Config &config) {
+    const auto it = config_to_stack_id_map_.find(config);
+    StackId result;
+    if (it == config_to_stack_id_map_.end()) {
+      result = next_stack_id_++;
+      config_to_stack_id_map_.insert(
+          std::pair<Config, StackId>(config, result));
+      stack_id_to_config_map_[result] = config;
+    } else {
+      result = it->second;
+    }
+    return result;
+  }
+
+  // Retrieves the internal stack ID for a corresponding external stack ID.
+  inline const Config InternalStackIds(StackId stack_id) const {
+    auto it = stack_id_to_config_map_.find(stack_id);
+    if (it == stack_id_to_config_map_.end()) {
+      it = stack_id_to_config_map_.find(-1);
+    }
+    return it->second;
+  }
+
+  inline bool Empty(const Config &config, Level level) const {
+    return config[level] <= 0;
+  }
+
+  inline bool AllEmpty(const Config &config) {
+    for (Level level = 0; level < nlevels; ++level) {
+      if (!Empty(config, level)) return false;
+    }
+    return true;
+  }
+
+  bool error_;
+  Label min_paren_;
+  Label max_paren_;
+  // Stores level of each paren.
+  std::unordered_map<Label, Label> paren_levels_;
+  std::vector<std::pair<Label, Label>> parens_;  // As in pdt.h.
+  std::unordered_map<Label, size_t> paren_map_;  // As in pdt.h.
+  // Maps between internal paren_id and external paren_id.
+  std::unordered_map<KeyPair<Level>, size_t, KeyPairHasher<Level>>
+      paren_id_map_;
+  // Maps between internal stack ids and external stack id.
+  ConfigToStackId config_to_stack_id_map_;
+  std::unordered_map<StackId, Config> stack_id_to_config_map_;
+  StackId next_stack_id_;
+  // Array of stacks.
+  PdtStack<StackId, Label> *stacks_[nlevels];
+};
+
+template <typename StackId, typename Level, Level nlevels, MPdtType restrict>
+MPdtStack<StackId, Level, nlevels, restrict>::MPdtStack(
+    const std::vector<std::pair<Level, Level>> &parens,  // NB: Label = Level.
+    const std::vector<Level> &assignments)
+    : error_(false),
+      min_paren_(kNoLabel),
+      max_paren_(kNoLabel),
+      parens_(parens),
+      next_stack_id_(1) {
+  using Label = Level;
+  if (parens.size() != assignments.size()) {
+    FSTERROR() << "MPdtStack: Parens of different size from assignments";
+    error_ = true;
+    return;
+  }
+  std::vector<std::pair<Label, Label>> vectors[nlevels];
+  for (Level i = 0; i < assignments.size(); ++i) {
+    // Assignments here start at 0, so assuming the human-readable version has
+    // them starting at 1, we should subtract 1 here
+    const auto level = assignments[i] - 1;
+    if (level < 0 || level >= nlevels) {
+      FSTERROR() << "MPdtStack: Specified level " << level << " out of bounds";
+      error_ = true;
+      return;
+    }
+    const auto &pair = parens[i];
+    vectors[level].push_back(pair);
+    paren_levels_[pair.first] = level;
+    paren_levels_[pair.second] = level;
+    paren_map_[pair.first] = i;
+    paren_map_[pair.second] = i;
+    const KeyPair<Level> key(level, vectors[level].size() - 1);
+    paren_id_map_[key] = i;
+    if (min_paren_ == kNoLabel || pair.first < min_paren_) {
+      min_paren_ = pair.first;
+    }
+    if (pair.second < min_paren_) min_paren_ = pair.second;
+    if (max_paren_ == kNoLabel || pair.first > max_paren_) {
+      max_paren_ = pair.first;
+    }
+    if (pair.second > max_paren_) max_paren_ = pair.second;
+  }
+  using Config = StackConfig<StackId, Level, nlevels>;
+  Config neg_one;
+  Config zero;
+  for (Level level = 0; level < nlevels; ++level) {
+    stacks_[level] = new PdtStack<StackId, Label>(vectors[level]);
+    neg_one[level] = -1;
+    zero[level] = 0;
+  }
+  config_to_stack_id_map_[neg_one] = -1;
+  config_to_stack_id_map_[zero] = 0;
+  stack_id_to_config_map_[-1] = neg_one;
+  stack_id_to_config_map_[0] = zero;
+}
+
+template <typename StackId, typename Level, Level nlevels, MPdtType restrict>
+MPdtStack<StackId, Level, nlevels, restrict>::MPdtStack(
+    const MPdtStack<StackId, Level, nlevels, restrict> &mstack)
+    : error_(mstack.error_),
+      min_paren_(mstack.min_paren_),
+      max_paren_(mstack.max_paren_),
+      parens_(mstack.parens_),
+      next_stack_id_(mstack.next_stack_id_) {
+  for (const auto &kv : mstack.paren_levels_) {
+    paren_levels_[kv.first] = kv.second;
+  }
+  for (const auto &paren : mstack.parens_) parens_.push_back(paren);
+  for (const auto &kv : mstack.paren_map_) {
+    paren_map_[kv.first] = kv.second;
+  }
+  for (const auto &kv : mstack.paren_id_map_) {
+    paren_id_map_[kv.first] = kv.second;
+  }
+  for (auto it = mstack.config_to_stack_id_map_.begin();
+       it != mstack.config_to_stack_id_map_.end(); ++it) {
+    config_to_stack_id_map_[it->first] = it->second;
+  }
+  for (const auto &kv : mstack.stack_id_to_config_map_) {
+    using Config = StackConfig<StackId, Level, nlevels>;
+    const Config config(kv.second);
+    stack_id_to_config_map_[kv.first] = config;
+  }
+  for (Level level = 0; level < nlevels; ++level)
+    stacks_[level] = mstack.stacks_[level];
+}
+
+template <typename StackId, typename Level, Level nlevels, MPdtType restrict>
+StackId MPdtStack<StackId, Level, nlevels, restrict>::Find(StackId stack_id,
+                                                           Level label) {
+  // Non-paren.
+  if (min_paren_ == kNoLabel || label < min_paren_ || label > max_paren_) {
+    return stack_id;
+  }
+  const auto it = paren_map_.find(label);
+  // Non-paren.
+  if (it == paren_map_.end()) return stack_id;
+  ssize_t paren_id = it->second;
+  // Gets the configuration associated with this stack_id.
+  const auto config = InternalStackIds(stack_id);
+  // Gets the level.
+  const auto level = paren_levels_.find(label)->second;
+  // If the label is an open paren we push:
+  //
+  // 1) if the restrict type is not MPDT_WRITE_RESTRICT, or
+  // 2) the restrict type is MPDT_WRITE_RESTRICT, and all the stacks above the
+  // level are empty.
+  if (label == parens_[paren_id].first) {  // Open paren.
+    if (restrict == MPDT_WRITE_RESTRICT) {
+      for (Level upper_level = 0; upper_level < level; ++upper_level) {
+        if (!Empty(config, upper_level)) return -1;  // Non-empty stack blocks.
+      }
+    }
+    // If the label is an close paren we pop:
+    //
+    // 1) if the restrict type is not MPDT_READ_RESTRICT, or
+    // 2) the restrict type is MPDT_READ_RESTRICT, and all the stacks above the
+    // level are empty.
+  } else if (restrict == MPDT_READ_RESTRICT) {
+    for (Level lower_level = 0; lower_level < level; ++lower_level) {
+      if (!Empty(config, lower_level)) return -1;  // Non-empty stack blocks.
+    }
+  }
+  const auto nid = stacks_[level]->Find(config[level], label);
+  // If the new ID is -1, that means that there is no valid transition at the
+  // level we want.
+  if (nid == -1) {
+    return -1;
+  } else {
+    using Config = StackConfig<StackId, Level, nlevels>;
+    Config nconfig(config);
+    nconfig[level] = nid;
+    return ExternalStackId(nconfig);
+  }
+}
+
+}  // namespace internal
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_MPDT_MPDT_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/mpdtlib.h b/projects/llm_framework/include/fst/extensions/mpdt/mpdtlib.h
new file mode 100644
index 00000000..bbf206f1
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/mpdt/mpdtlib.h
@@ -0,0 +1,18 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// This is an experimental multipush-down transducer (MPDT) library. An MPDT is
+// encoded as an FST, where some transitions are labeled with open or close
+// parentheses, each mated pair of which is associated to one stack. To be
+// interpreted as an MPDT, the parentheses within a stack must balance on a
+// path.
+
+#ifndef FST_EXTENSIONS_MPDT_MPDTLIB_H_
+#define FST_EXTENSIONS_MPDT_MPDTLIB_H_
+
+#include <fst/extensions/mpdt/compose.h>
+#include <fst/extensions/mpdt/expand.h>
+#include <fst/extensions/mpdt/mpdt.h>
+#include <fst/extensions/mpdt/reverse.h>
+
+#endif  // FST_EXTENSIONS_MPDT_MPDTLIB_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/mpdtscript.h b/projects/llm_framework/include/fst/extensions/mpdt/mpdtscript.h
new file mode 100644
index 00000000..66340992
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/mpdt/mpdtscript.h
@@ -0,0 +1,156 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Convenience file for including all MPDT operations at once, and/or
+// registering them for new arc types.
+
+#ifndef FST_EXTENSIONS_MPDT_MPDTSCRIPT_H_
+#define FST_EXTENSIONS_MPDT_MPDTSCRIPT_H_
+
+#include <algorithm>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+#include <fst/compose.h>  // for ComposeOptions
+#include <fst/util.h>
+
+#include <fst/script/arg-packs.h>
+#include <fst/script/fst-class.h>
+#include <fst/script/shortest-path.h>
+
+#include <fst/extensions/mpdt/compose.h>
+#include <fst/extensions/mpdt/expand.h>
+#include <fst/extensions/mpdt/info.h>
+#include <fst/extensions/mpdt/reverse.h>
+
+#include <fst/extensions/pdt/pdtscript.h>  // For LabelClassPair,
+                                               // FstClassPair, and to detect
+                                               // any collisions.
+
+namespace fst {
+namespace script {
+
+using MPdtComposeArgs =
+    std::tuple<const FstClass &, const FstClass &,
+               const std::vector<LabelPair> &, const std::vector<int64> &,
+               MutableFstClass *, const MPdtComposeOptions &, bool>;
+
+template <class Arc>
+void MPdtCompose(MPdtComposeArgs *args) {
+  const Fst<Arc> &ifst1 = *(std::get<0>(*args).GetFst<Arc>());
+  const Fst<Arc> &ifst2 = *(std::get<1>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<4>(*args)->GetMutableFst<Arc>();
+  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
+      std::get<2>(*args).size());
+  std::copy(std::get<2>(*args).begin(), std::get<2>(*args).end(),
+            typed_parens.begin());
+  using Level = typename Arc::Label;
+  std::vector<Level> typed_assignments(std::get<3>(*args).size());
+  std::copy(std::get<3>(*args).begin(), std::get<3>(*args).end(),
+            typed_assignments.begin());
+  if (std::get<6>(*args)) {
+    Compose(ifst1, typed_parens, typed_assignments, ifst2, ofst,
+            std::get<5>(*args));
+  } else {
+    Compose(ifst1, ifst2, typed_parens, typed_assignments, ofst,
+            std::get<5>(*args));
+  }
+}
+
+void MPdtCompose(const FstClass &ifst1, const FstClass &ifst2,
+                 const std::vector<LabelPair> &parens,
+                 const std::vector<int64> &assignments, MutableFstClass *ofst,
+                 const MPdtComposeOptions &copts, bool left_pdt);
+
+using MPdtExpandArgs =
+    std::tuple<const FstClass &, const std::vector<LabelPair> &,
+               const std::vector<int64> &, MutableFstClass *,
+               const MPdtExpandOptions &>;
+
+template <class Arc>
+void MPdtExpand(MPdtExpandArgs *args) {
+  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<3>(*args)->GetMutableFst<Arc>();
+  // In case Arc::Label is not the same as FstClass::Label, we make copies.
+  // Truncation may occur if FstClass::Label has more precision than
+  // Arc::Label.
+  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
+      std::get<1>(*args).size());
+  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
+            typed_parens.begin());
+  using Level = typename Arc::Label;
+  std::vector<Level> typed_assignments(std::get<2>(*args).size());
+  std::copy(std::get<2>(*args).begin(), std::get<2>(*args).end(),
+            typed_assignments.begin());
+  Expand(fst, typed_parens, typed_assignments, ofst,
+         MPdtExpandOptions(std::get<4>(*args).connect,
+                           std::get<4>(*args).keep_parentheses));
+}
+
+void MPdtExpand(const FstClass &ifst, const std::vector<LabelPair> &parens,
+                const std::vector<int64> &assignments, MutableFstClass *ofst,
+                const MPdtExpandOptions &opts);
+
+using MPdtReverseArgs =
+    std::tuple<const FstClass &, const std::vector<LabelPair> &,
+               std::vector<int64> *, MutableFstClass *>;
+
+template <class Arc>
+void MPdtReverse(MPdtReverseArgs *args) {
+  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<3>(*args)->GetMutableFst<Arc>();
+  // In case Arc::Label is not the same as FstClass::Label, we make copies.
+  // Truncation may occur if FstClass::Label has more precision than
+  // Arc::Label.
+  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
+      std::get<1>(*args).size());
+  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
+            typed_parens.begin());
+  using Level = typename Arc::Label;
+  std::vector<Level> typed_assignments(std::get<2>(*args)->size());
+  std::copy(std::get<2>(*args)->begin(), std::get<2>(*args)->end(),
+            typed_assignments.begin());
+  Reverse(fst, typed_parens, &typed_assignments, ofst);
+  // Reassign stack assignments to input assignment vector.
+  std::copy(typed_assignments.begin(), typed_assignments.end(),
+            std::get<2>(*args)->begin());
+}
+
+void MPdtReverse(const FstClass &ifst, const std::vector<LabelPair> &parens,
+                 std::vector<int64> *assignments, MutableFstClass *ofst);
+
+using PrintMPdtInfoArgs =
+    std::tuple<const FstClass &, const std::vector<LabelPair> &,
+               const std::vector<int64> &>;
+
+template <class Arc>
+void PrintMPdtInfo(PrintMPdtInfoArgs *args) {
+  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
+  // In case Arc::Label is not the same as FstClass::Label, we make copies.
+  // Truncation may occur if FstClass::Label has more precision than
+  // Arc::Label.
+  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
+      std::get<1>(*args).size());
+  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
+            typed_parens.begin());
+  using Level = typename Arc::Label;
+  std::vector<Level> typed_assignments(std::get<2>(*args).size());
+  std::copy(std::get<2>(*args).begin(), std::get<2>(*args).end(),
+            typed_assignments.begin());
+  MPdtInfo<Arc> mpdtinfo(fst, typed_parens, typed_assignments);
+  mpdtinfo.Print();
+}
+
+void PrintMPdtInfo(const FstClass &ifst, const std::vector<LabelPair> &parens,
+                   const std::vector<int64> &assignments);
+
+}  // namespace script
+}  // namespace fst
+
+#define REGISTER_FST_MPDT_OPERATIONS(ArcType)                    \
+  REGISTER_FST_OPERATION(MPdtCompose, ArcType, MPdtComposeArgs); \
+  REGISTER_FST_OPERATION(MPdtExpand, ArcType, MPdtExpandArgs);   \
+  REGISTER_FST_OPERATION(MPdtReverse, ArcType, MPdtReverseArgs); \
+  REGISTER_FST_OPERATION(PrintMPdtInfo, ArcType, PrintMPdtInfoArgs)
+#endif  // FST_EXTENSIONS_MPDT_MPDTSCRIPT_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/read_write_utils.h b/projects/llm_framework/include/fst/extensions/mpdt/read_write_utils.h
new file mode 100644
index 00000000..38f9f963
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/mpdt/read_write_utils.h
@@ -0,0 +1,86 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Definition of ReadLabelTriples based on ReadLabelPairs, like that in
+// nlp/fst/lib/util.h for pairs, and similarly for WriteLabelTriples.
+
+#ifndef FST_EXTENSIONS_MPDT_READ_WRITE_UTILS_H_
+#define FST_EXTENSIONS_MPDT_READ_WRITE_UTILS_H_
+
+#include <string>
+#include <utility>
+#include <vector>
+
+#include <fstream>
+#include <fst/test-properties.h>
+
+namespace fst {
+
+// Returns true on success.
+template <typename Label>
+bool ReadLabelTriples(const string &filename,
+                      std::vector<std::pair<Label, Label>> *pairs,
+                      std::vector<Label> *assignments,
+                      bool allow_negative = false) {
+  std::ifstream fstrm(filename);
+  if (!fstrm) {
+    LOG(ERROR) << "ReadIntTriples: Can't open file: " << filename;
+    return false;
+  }
+  static constexpr auto kLineLen = 8096;
+  char line[kLineLen];
+  size_t nline = 0;
+  pairs->clear();
+  while (fstrm.getline(line, kLineLen)) {
+    ++nline;
+    std::vector<char *> col;
+    SplitString(line, "\n\t ", &col, true);
+    // Empty line or comment?
+    if (col.empty() || col[0][0] == '\0' || col[0][0] == '#') continue;
+    if (col.size() != 3) {
+      LOG(ERROR) << "ReadLabelTriples: Bad number of columns, "
+                 << "file = " << filename << ", line = " << nline;
+      return false;
+    }
+    bool err;
+    const Label i1 = StrToInt64(col[0], filename, nline, allow_negative, &err);
+    if (err) return false;
+    const Label i2 = StrToInt64(col[1], filename, nline, allow_negative, &err);
+    if (err) return false;
+    using Level = Label;
+    const Level i3 = StrToInt64(col[2], filename, nline, allow_negative, &err);
+    if (err) return false;
+    pairs->push_back(std::make_pair(i1, i2));
+    assignments->push_back(i3);
+  }
+  return true;
+}
+
+// Returns true on success.
+template <typename Label>
+bool WriteLabelTriples(const string &filename,
+                       const std::vector<std::pair<Label, Label>> &pairs,
+                       const std::vector<Label> &assignments) {
+  if (pairs.size() != assignments.size()) {
+    LOG(ERROR) << "WriteLabelTriples: Pairs and assignments of different sizes";
+    return false;
+  }
+  std::ofstream fstrm(filename);
+  if (!fstrm) {
+    LOG(ERROR) << "WriteLabelTriples: Can't open file: " << filename;
+    return false;
+  }
+  for (size_t n = 0; n < pairs.size(); ++n)
+    fstrm << pairs[n].first << "\t" << pairs[n].second << "\t" << assignments[n]
+          << "\n";
+  if (!fstrm) {
+    LOG(ERROR) << "WriteLabelTriples: Write failed: "
+               << (filename.empty() ? "standard output" : filename);
+    return false;
+  }
+  return true;
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_MPDT_READ_WRITE_UTILS_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/reverse.h b/projects/llm_framework/include/fst/extensions/mpdt/reverse.h
new file mode 100644
index 00000000..413dd3f0
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/mpdt/reverse.h
@@ -0,0 +1,54 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Reverses an MPDT.
+
+#ifndef FST_EXTENSIONS_MPDT_REVERSE_H_
+#define FST_EXTENSIONS_MPDT_REVERSE_H_
+
+#include <limits>
+#include <vector>
+
+#include <fst/mutable-fst.h>
+#include <fst/relabel.h>
+#include <fst/reverse.h>
+
+namespace fst {
+
+// Reverses a multi-stack pushdown transducer (MPDT) encoded as an FST.
+template <class Arc, class RevArc>
+void Reverse(
+    const Fst<Arc> &ifst,
+    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
+        &parens,
+    std::vector<typename Arc::Label> *assignments, MutableFst<RevArc> *ofst) {
+  using Label = typename Arc::Label;
+  // Reverses FST component.
+  Reverse(ifst, ofst);
+  // Exchanges open and close parenthesis pairs.
+  std::vector<std::pair<Label, Label>> relabel_pairs;
+  relabel_pairs.reserve(2 * parens.size());
+  for (const auto &pair : parens) {
+    relabel_pairs.emplace_back(pair.first, pair.second);
+    relabel_pairs.emplace_back(pair.second, pair.first);
+  }
+  Relabel(ofst, relabel_pairs, relabel_pairs);
+  // Computes new bounds for the stack assignments.
+  Label max_level = -1;
+  Label min_level = std::numeric_limits<Label>::max();
+  for (const auto assignment : *assignments) {
+    if (assignment < min_level) {
+      min_level = assignment;
+    } else if (assignment > max_level) {
+      max_level = assignment;
+    }
+  }
+  // Actually reverses stack assignments.
+  for (auto &assignment : *assignments) {
+    assignment = (max_level - assignment) + min_level;
+  }
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_MPDT_REVERSE_H_
diff --git a/projects/llm_framework/include/fst/extensions/ngram/bitmap-index.h b/projects/llm_framework/include/fst/extensions/ngram/bitmap-index.h
new file mode 100644
index 00000000..09ad6d7d
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/ngram/bitmap-index.h
@@ -0,0 +1,168 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_EXTENSIONS_NGRAM_BITMAP_INDEX_H_
+#define FST_EXTENSIONS_NGRAM_BITMAP_INDEX_H_
+
+#include <utility>
+#include <vector>
+
+#include <fst/compat.h>
+
+// This class is a bitstring storage class with an index that allows
+// seeking to the Nth set or clear bit in time O(Log(N)) where N is
+// the length of the bit vector.  In addition, it allows counting set or
+// clear bits over ranges in constant time.
+//
+// This is accomplished by maintaining an "secondary" index of limited
+// size in bits that maintains a running count of the number of bits set
+// in each block of bitmap data.  A block is defined as the number of
+// uint64 values that can fit in the secondary index before an overflow
+// occurs.
+//
+// To handle overflows, a "primary" index containing a running count of
+// bits set in each block is created using the type uint64.
+
+namespace fst {
+
+class BitmapIndex {
+ public:
+  static size_t StorageSize(size_t size) {
+    return ((size + kStorageBlockMask) >> kStorageLogBitSize);
+  }
+
+  BitmapIndex() : bits_(nullptr), size_(0) {}
+
+  bool Get(size_t index) const {
+    return (bits_[index >> kStorageLogBitSize] &
+            (kOne << (index & kStorageBlockMask))) != 0;
+  }
+
+  static void Set(uint64* bits, size_t index) {
+    bits[index >> kStorageLogBitSize] |= (kOne << (index & kStorageBlockMask));
+  }
+
+  static void Clear(uint64* bits, size_t index) {
+    bits[index >> kStorageLogBitSize] &= ~(kOne << (index & kStorageBlockMask));
+  }
+
+  size_t Bits() const { return size_; }
+
+  size_t ArraySize() const { return StorageSize(size_); }
+
+  // Returns the number of one bits in the bitmap
+  size_t GetOnesCount() const {
+    return primary_index_[primary_index_size() - 1];
+  }
+
+  // Returns the number of one bits in positions 0 to limit - 1.
+  // REQUIRES: limit <= Bits()
+  size_t Rank1(size_t end) const;
+
+  // Returns the number of one bits in the range start to end - 1.
+  // REQUIRES: limit <= Bits()
+  size_t GetOnesCountInRange(size_t start, size_t end) const {
+    return Rank1(end) - Rank1(start);
+  }
+
+  // Returns the number of zero bits in positions 0 to limit - 1.
+  // REQUIRES: limit <= Bits()
+  size_t Rank0(size_t end) const { return end - Rank1(end); }
+
+  // Returns the number of zero bits in the range start to end - 1.
+  // REQUIRES: limit <= Bits()
+  size_t GetZeroesCountInRange(size_t start, size_t end) const {
+    return end - start - GetOnesCountInRange(start, end);
+  }
+
+  // Return true if any bit between begin inclusive and end exclusive
+  // is set.  0 <= begin <= end <= Bits() is required.
+  //
+  bool TestRange(size_t start, size_t end) const {
+    return Rank1(end) > Rank1(start);
+  }
+
+  // Returns the offset to the nth set bit (zero based)
+  // or Bits() if index >= number of ones
+  size_t Select1(size_t bit_index) const;
+
+  // Returns the offset to the nth clear bit (zero based)
+  // or Bits() if index > number of
+  size_t Select0(size_t bit_index) const;
+
+  // Returns the offset of the nth and nth+1 clear bit (zero based),
+  // equivalent to two calls to Select0, but more efficient.
+  std::pair<size_t, size_t> Select0s(size_t bit_index) const;
+
+  // Rebuilds from index for the associated Bitmap, should be called
+  // whenever changes have been made to the Bitmap or else behavior
+  // of the indexed bitmap methods will be undefined.
+  void BuildIndex(const uint64* bits, size_t size);
+
+  // the secondary index accumulates counts until it can possibly overflow
+  // this constant computes the number of uint64 units that can fit into
+  // units the size of uint16.
+  static const uint64 kOne = 1;
+  static const uint32 kStorageBitSize = 64;
+  static const uint32 kStorageLogBitSize = 6;
+  static const uint32 kSecondaryBlockSize =
+      ((1 << 16) - 1) >> kStorageLogBitSize;
+
+ private:
+  static const uint32 kStorageBlockMask = kStorageBitSize - 1;
+
+  // returns, from the index, the count of ones up to array_index
+  size_t get_index_ones_count(size_t array_index) const;
+
+  // because the indexes, both primary and secondary, contain a running
+  // count of the population of one bits contained in [0,i), there is
+  // no reason to have an element in the zeroth position as this value would
+  // necessarily be zero.  (The bits are indexed in a zero based way.)  Thus
+  // we don't store the 0th element in either index.  Both of the following
+  // functions, if greater than 0, must be decremented by one before retreiving
+  // the value from the corresponding array.
+  // returns the 1 + the block that contains the bitindex in question
+  // the inverted version works the same but looks for zeros using an inverted
+  // view of the index
+  size_t find_primary_block(size_t bit_index) const;
+
+  size_t find_inverted_primary_block(size_t bit_index) const;
+
+  // similarly, the secondary index (which resets its count to zero at
+  // the end of every kSecondaryBlockSize entries) does not store the element
+  // at 0.  Note that the rem_bit_index parameter is the number of bits
+  // within the secondary block, after the bits accounted for by the primary
+  // block have been removed (i.e. the remaining bits)  And, because we
+  // reset to zero with each new block, there is no need to store those
+  // actual zeros.
+  // returns 1 + the secondary block that contains the bitindex in question
+  size_t find_secondary_block(size_t block, size_t rem_bit_index) const;
+
+  size_t find_inverted_secondary_block(size_t block,
+                                       size_t rem_bit_index) const;
+
+  // We create a primary index based upon the number of secondary index
+  // blocks.  The primary index uses fields wide enough to accomodate any
+  // index of the bitarray so cannot overflow
+  // The primary index is the actual running
+  // count of one bits set for all blocks (and, thus, all uint64s).
+  size_t primary_index_size() const {
+    return (ArraySize() + kSecondaryBlockSize - 1) / kSecondaryBlockSize;
+  }
+
+  const uint64* bits_;
+  size_t size_;
+
+  // The primary index contains the running popcount of all blocks
+  // which means the nth value contains the popcounts of
+  // [0,n*kSecondaryBlockSize], however, the 0th element is omitted.
+  std::vector<uint32> primary_index_;
+  // The secondary index contains the running popcount of the associated
+  // bitmap.  It is the same length (in units of uint16) as the
+  // bitmap's map is in units of uint64s.
+  std::vector<uint16> secondary_index_;
+};
+
+}  // end namespace fst
+
+#endif  // FST_EXTENSIONS_NGRAM_BITMAP_INDEX_H_
diff --git a/projects/llm_framework/include/fst/extensions/ngram/ngram-fst.h b/projects/llm_framework/include/fst/extensions/ngram/ngram-fst.h
new file mode 100644
index 00000000..8183a8c0
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/ngram/ngram-fst.h
@@ -0,0 +1,1027 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// NgramFst implements a n-gram language model based upon the LOUDS data
+// structure.  Please refer to "Unary Data Structures for Language Models"
+// http://research.google.com/pubs/archive/37218.pdf
+
+#ifndef FST_EXTENSIONS_NGRAM_NGRAM_FST_H_
+#define FST_EXTENSIONS_NGRAM_NGRAM_FST_H_
+
+#include <stddef.h>
+#include <string.h>
+#include <algorithm>
+#include <iostream>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include <fst/compat.h>
+#include <fst/log.h>
+#include <fstream>
+#include <fst/extensions/ngram/bitmap-index.h>
+#include <fst/fstlib.h>
+#include <fst/mapped-file.h>
+
+namespace fst {
+template <class A>
+class NGramFst;
+template <class A>
+class NGramFstMatcher;
+
+// Instance data containing mutable state for bookkeeping repeated access to
+// the same state.
+template <class A>
+struct NGramFstInst {
+  typedef typename A::Label Label;
+  typedef typename A::StateId StateId;
+  typedef typename A::Weight Weight;
+  StateId state_;
+  size_t num_futures_;
+  size_t offset_;
+  size_t node_;
+  StateId node_state_;
+  std::vector<Label> context_;
+  StateId context_state_;
+  NGramFstInst()
+      : state_(kNoStateId),
+        node_state_(kNoStateId),
+        context_state_(kNoStateId) {}
+};
+
+namespace internal {
+
+// Implementation class for LOUDS based NgramFst interface.
+template <class A>
+class NGramFstImpl : public FstImpl<A> {
+  using FstImpl<A>::SetInputSymbols;
+  using FstImpl<A>::SetOutputSymbols;
+  using FstImpl<A>::SetType;
+  using FstImpl<A>::WriteHeader;
+
+  friend class ArcIterator<NGramFst<A>>;
+  friend class NGramFstMatcher<A>;
+
+ public:
+  using FstImpl<A>::InputSymbols;
+  using FstImpl<A>::SetProperties;
+  using FstImpl<A>::Properties;
+
+  typedef A Arc;
+  typedef typename A::Label Label;
+  typedef typename A::StateId StateId;
+  typedef typename A::Weight Weight;
+
+  NGramFstImpl() {
+    SetType("ngram");
+    SetInputSymbols(nullptr);
+    SetOutputSymbols(nullptr);
+    SetProperties(kStaticProperties);
+  }
+
+  NGramFstImpl(const Fst<A> &fst, std::vector<StateId> *order_out);
+
+  explicit NGramFstImpl(const Fst<A> &fst) : NGramFstImpl(fst, nullptr) {}
+
+  NGramFstImpl(const NGramFstImpl &other) {
+    FSTERROR() << "Copying NGramFst Impls is not supported, use safe = false.";
+    SetProperties(kError, kError);
+  }
+
+  ~NGramFstImpl() override {
+    if (owned_) {
+      delete[] data_;
+    }
+  }
+
+  static NGramFstImpl<A> *Read(std::istream &strm,  // NOLINT
+                               const FstReadOptions &opts) {
+    NGramFstImpl<A> *impl = new NGramFstImpl();
+    FstHeader hdr;
+    if (!impl->ReadHeader(strm, opts, kMinFileVersion, &hdr)) return 0;
+    uint64 num_states, num_futures, num_final;
+    const size_t offset =
+        sizeof(num_states) + sizeof(num_futures) + sizeof(num_final);
+    // Peek at num_states and num_futures to see how much more needs to be read.
+    strm.read(reinterpret_cast<char *>(&num_states), sizeof(num_states));
+    strm.read(reinterpret_cast<char *>(&num_futures), sizeof(num_futures));
+    strm.read(reinterpret_cast<char *>(&num_final), sizeof(num_final));
+    size_t size = Storage(num_states, num_futures, num_final);
+    MappedFile *data_region = MappedFile::Allocate(size);
+    char *data = reinterpret_cast<char *>(data_region->mutable_data());
+    // Copy num_states, num_futures and num_final back into data.
+    memcpy(data, reinterpret_cast<char *>(&num_states), sizeof(num_states));
+    memcpy(data + sizeof(num_states), reinterpret_cast<char *>(&num_futures),
+           sizeof(num_futures));
+    memcpy(data + sizeof(num_states) + sizeof(num_futures),
+           reinterpret_cast<char *>(&num_final), sizeof(num_final));
+    strm.read(data + offset, size - offset);
+    if (strm.fail()) {
+      delete impl;
+      return nullptr;
+    }
+    impl->Init(data, false, data_region);
+    return impl;
+  }
+
+  bool Write(std::ostream &strm,  // NOLINT
+             const FstWriteOptions &opts) const {
+    FstHeader hdr;
+    hdr.SetStart(Start());
+    hdr.SetNumStates(num_states_);
+    WriteHeader(strm, opts, kFileVersion, &hdr);
+    strm.write(data_, StorageSize());
+    return !strm.fail();
+  }
+
+  StateId Start() const { return start_; }
+
+  Weight Final(StateId state) const {
+    if (final_index_.Get(state)) {
+      return final_probs_[final_index_.Rank1(state)];
+    } else {
+      return Weight::Zero();
+    }
+  }
+
+  size_t NumArcs(StateId state, NGramFstInst<A> *inst = nullptr) const {
+    if (inst == nullptr) {
+      const std::pair<size_t, size_t> zeros =
+          (state == 0) ? select_root_ : future_index_.Select0s(state);
+      return zeros.second - zeros.first - 1;
+    }
+    SetInstFuture(state, inst);
+    return inst->num_futures_ + ((state == 0) ? 0 : 1);
+  }
+
+  size_t NumInputEpsilons(StateId state) const {
+    // State 0 has no parent, thus no backoff.
+    if (state == 0) return 0;
+    return 1;
+  }
+
+  size_t NumOutputEpsilons(StateId state) const {
+    return NumInputEpsilons(state);
+  }
+
+  StateId NumStates() const { return num_states_; }
+
+  void InitStateIterator(StateIteratorData<A> *data) const {
+    data->base = 0;
+    data->nstates = num_states_;
+  }
+
+  static size_t Storage(uint64 num_states, uint64 num_futures,
+                        uint64 num_final) {
+    uint64 b64;
+    Weight weight;
+    Label label;
+    size_t offset =
+        sizeof(num_states) + sizeof(num_futures) + sizeof(num_final);
+    offset +=
+        sizeof(b64) * (BitmapIndex::StorageSize(num_states * 2 + 1) +
+                       BitmapIndex::StorageSize(num_futures + num_states + 1) +
+                       BitmapIndex::StorageSize(num_states));
+    offset += (num_states + 1) * sizeof(label) + num_futures * sizeof(label);
+    // Pad for alignemnt, see
+    // http://en.wikipedia.org/wiki/Data_structure_alignment#Computing_padding
+    offset = (offset + sizeof(weight) - 1) & ~(sizeof(weight) - 1);
+    offset += (num_states + 1) * sizeof(weight) + num_final * sizeof(weight) +
+              (num_futures + 1) * sizeof(weight);
+    return offset;
+  }
+
+  void SetInstFuture(StateId state, NGramFstInst<A> *inst) const {
+    if (inst->state_ != state) {
+      inst->state_ = state;
+      const std::pair<size_t, size_t> zeros = future_index_.Select0s(state);
+      inst->num_futures_ = zeros.second - zeros.first - 1;
+      inst->offset_ = future_index_.Rank1(zeros.first + 1);
+    }
+  }
+
+  void SetInstNode(NGramFstInst<A> *inst) const {
+    if (inst->node_state_ != inst->state_) {
+      inst->node_state_ = inst->state_;
+      inst->node_ = context_index_.Select1(inst->state_);
+    }
+  }
+
+  void SetInstContext(NGramFstInst<A> *inst) const {
+    SetInstNode(inst);
+    if (inst->context_state_ != inst->state_) {
+      inst->context_state_ = inst->state_;
+      inst->context_.clear();
+      size_t node = inst->node_;
+      while (node != 0) {
+        inst->context_.push_back(context_words_[context_index_.Rank1(node)]);
+        node = context_index_.Select1(context_index_.Rank0(node) - 1);
+      }
+    }
+  }
+
+  // Access to the underlying representation
+  const char *GetData(size_t *data_size) const {
+    *data_size = StorageSize();
+    return data_;
+  }
+
+  void Init(const char *data, bool owned, MappedFile *file = nullptr);
+
+  const std::vector<Label> &GetContext(StateId s, NGramFstInst<A> *inst) const {
+    SetInstFuture(s, inst);
+    SetInstContext(inst);
+    return inst->context_;
+  }
+
+  size_t StorageSize() const {
+    return Storage(num_states_, num_futures_, num_final_);
+  }
+
+  void GetStates(const std::vector<Label> &context,
+                 std::vector<StateId> *states) const;
+
+ private:
+  StateId Transition(const std::vector<Label> &context, Label future) const;
+
+  // Properties always true for this Fst class.
+  static const uint64 kStaticProperties =
+      kAcceptor | kIDeterministic | kODeterministic | kEpsilons | kIEpsilons |
+      kOEpsilons | kILabelSorted | kOLabelSorted | kWeighted | kCyclic |
+      kInitialAcyclic | kNotTopSorted | kAccessible | kCoAccessible |
+      kNotString | kExpanded;
+  // Current file format version.
+  static const int kFileVersion = 4;
+  // Minimum file format version supported.
+  static const int kMinFileVersion = 4;
+
+  std::unique_ptr<MappedFile> data_region_;
+  const char *data_ = nullptr;
+  bool owned_ = false;  // True if we own data_
+  StateId start_ = fst::kNoStateId;
+  uint64 num_states_ = 0;
+  uint64 num_futures_ = 0;
+  uint64 num_final_ = 0;
+  std::pair<size_t, size_t> select_root_;
+  const Label *root_children_ = nullptr;
+  // borrowed references
+  const uint64 *context_ = nullptr;
+  const uint64 *future_ = nullptr;
+  const uint64 *final_ = nullptr;
+  const Label *context_words_ = nullptr;
+  const Label *future_words_ = nullptr;
+  const Weight *backoff_ = nullptr;
+  const Weight *final_probs_ = nullptr;
+  const Weight *future_probs_ = nullptr;
+  BitmapIndex context_index_;
+  BitmapIndex future_index_;
+  BitmapIndex final_index_;
+};
+
+template <typename A>
+inline void NGramFstImpl<A>::GetStates(
+    const std::vector<Label> &context,
+    std::vector<typename A::StateId> *states) const {
+  states->clear();
+  states->push_back(0);
+  typename std::vector<Label>::const_reverse_iterator cit = context.rbegin();
+  const Label *children = root_children_;
+  size_t num_children = select_root_.second - 2;
+  const Label *loc = std::lower_bound(children, children + num_children, *cit);
+  if (loc == children + num_children || *loc != *cit) return;
+  size_t node = 2 + loc - children;
+  states->push_back(context_index_.Rank1(node));
+  if (context.size() == 1) return;
+  size_t node_rank = context_index_.Rank1(node);
+  std::pair<size_t, size_t> zeros =
+      node_rank == 0 ? select_root_ : context_index_.Select0s(node_rank);
+  size_t first_child = zeros.first + 1;
+  ++cit;
+  if (context_index_.Get(first_child) != false) {
+    size_t last_child = zeros.second - 1;
+    while (cit != context.rend()) {
+      children = context_words_ + context_index_.Rank1(first_child);
+      loc = std::lower_bound(children, children + last_child - first_child + 1,
+                             *cit);
+      if (loc == children + last_child - first_child + 1 || *loc != *cit) {
+        break;
+      }
+      ++cit;
+      node = first_child + loc - children;
+      states->push_back(context_index_.Rank1(node));
+      node_rank = context_index_.Rank1(node);
+      zeros =
+          node_rank == 0 ? select_root_ : context_index_.Select0s(node_rank);
+      first_child = zeros.first + 1;
+      if (context_index_.Get(first_child) == false) break;
+      last_child = zeros.second - 1;
+    }
+  }
+}
+
+}  // namespace internal
+
+/*****************************************************************************/
+template <class A>
+class NGramFst : public ImplToExpandedFst<internal::NGramFstImpl<A>> {
+  friend class ArcIterator<NGramFst<A>>;
+  friend class NGramFstMatcher<A>;
+
+ public:
+  typedef A Arc;
+  typedef typename A::StateId StateId;
+  typedef typename A::Label Label;
+  typedef typename A::Weight Weight;
+  typedef internal::NGramFstImpl<A> Impl;
+
+  explicit NGramFst(const Fst<A> &dst)
+      : ImplToExpandedFst<Impl>(std::make_shared<Impl>(dst, nullptr)) {}
+
+  NGramFst(const Fst<A> &fst, std::vector<StateId> *order_out)
+      : ImplToExpandedFst<Impl>(std::make_shared<Impl>(fst, order_out)) {}
+
+  // Because the NGramFstImpl is a const stateless data structure, there
+  // is never a need to do anything beside copy the reference.
+  NGramFst(const NGramFst<A> &fst, bool safe = false)
+      : ImplToExpandedFst<Impl>(fst, false) {}
+
+  NGramFst() : ImplToExpandedFst<Impl>(std::make_shared<Impl>()) {}
+
+  // Non-standard constructor to initialize NGramFst directly from data.
+  NGramFst(const char *data, bool owned)
+      : ImplToExpandedFst<Impl>(std::make_shared<Impl>()) {
+    GetMutableImpl()->Init(data, owned, nullptr);
+  }
+
+  // Get method that gets the data associated with Init().
+  const char *GetData(size_t *data_size) const {
+    return GetImpl()->GetData(data_size);
+  }
+
+  const std::vector<Label> GetContext(StateId s) const {
+    return GetImpl()->GetContext(s, &inst_);
+  }
+
+  // Consumes as much as possible of context from right to left, returns the
+  // the states corresponding to the increasingly conditioned input sequence.
+  void GetStates(const std::vector<Label> &context,
+                 std::vector<StateId> *state) const {
+    return GetImpl()->GetStates(context, state);
+  }
+
+  size_t NumArcs(StateId s) const override {
+    return GetImpl()->NumArcs(s, &inst_);
+  }
+
+  NGramFst<A> *Copy(bool safe = false) const override {
+    return new NGramFst(*this, safe);
+  }
+
+  static NGramFst<A> *Read(std::istream &strm, const FstReadOptions &opts) {
+    Impl *impl = Impl::Read(strm, opts);
+    return impl ? new NGramFst<A>(std::shared_ptr<Impl>(impl)) : nullptr;
+  }
+
+  static NGramFst<A> *Read(const string &filename) {
+    if (!filename.empty()) {
+      std::ifstream strm(filename,
+                              std::ios_base::in | std::ios_base::binary);
+      if (!strm.good()) {
+        LOG(ERROR) << "NGramFst::Read: Can't open file: " << filename;
+        return nullptr;
+      }
+      return Read(strm, FstReadOptions(filename));
+    } else {
+      return Read(std::cin, FstReadOptions("standard input"));
+    }
+  }
+
+  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
+    return GetImpl()->Write(strm, opts);
+  }
+
+  bool Write(const string &filename) const override {
+    return Fst<A>::WriteFile(filename);
+  }
+
+  inline void InitStateIterator(StateIteratorData<A> *data) const override {
+    GetImpl()->InitStateIterator(data);
+  }
+
+  inline void InitArcIterator(StateId s,
+                              ArcIteratorData<A> *data) const override;
+
+  MatcherBase<A> *InitMatcher(MatchType match_type) const override {
+    return new NGramFstMatcher<A>(this, match_type);
+  }
+
+  size_t StorageSize() const { return GetImpl()->StorageSize(); }
+
+  static bool HasRequiredProps(const Fst<A> &fst) {
+    static const auto props =
+        kAcceptor | kIDeterministic | kILabelSorted | kIEpsilons | kAccessible;
+    return fst.Properties(props, true) == props;
+  }
+
+  static bool HasRequiredStructure(const Fst<A> &fst) {
+    if (!HasRequiredProps(fst)) {
+      return false;
+    }
+    typename A::StateId unigram = fst.Start();
+    while (true) {  // Follows epsilon arc chain to find unigram state.
+      if (unigram == fst::kNoStateId) return false;  // No unigram state.
+      typename fst::ArcIterator<Fst<A>> aiter(fst, unigram);
+      if (aiter.Done() || aiter.Value().ilabel != 0) break;
+      unigram = aiter.Value().nextstate;
+      aiter.Next();
+    }
+    // Other requirement: all states other than unigram an epsilon arc.
+    for (fst::StateIterator<Fst<A>> siter(fst); !siter.Done();
+         siter.Next()) {
+      const typename A::StateId &state = siter.Value();
+      fst::ArcIterator<Fst<A>> aiter(fst, state);
+      if (state != unigram) {
+        if (aiter.Done()) return false;
+        if (aiter.Value().ilabel != 0) return false;
+        aiter.Next();
+        if (!aiter.Done() && aiter.Value().ilabel == 0) return false;
+      }
+    }
+    return true;
+  }
+
+ private:
+  using ImplToExpandedFst<Impl, ExpandedFst<A>>::GetImpl;
+  using ImplToExpandedFst<Impl, ExpandedFst<A>>::GetMutableImpl;
+
+  explicit NGramFst(std::shared_ptr<Impl> impl)
+      : ImplToExpandedFst<Impl>(impl) {}
+
+  mutable NGramFstInst<A> inst_;
+};
+
+template <class A>
+inline void NGramFst<A>::InitArcIterator(StateId s,
+                                         ArcIteratorData<A> *data) const {
+  GetImpl()->SetInstFuture(s, &inst_);
+  GetImpl()->SetInstNode(&inst_);
+  data->base = new ArcIterator<NGramFst<A>>(*this, s);
+}
+
+namespace internal {
+
+template <typename A>
+NGramFstImpl<A>::NGramFstImpl(const Fst<A> &fst,
+                              std::vector<StateId> *order_out) {
+  typedef A Arc;
+  typedef typename Arc::Label Label;
+  typedef typename Arc::Weight Weight;
+  typedef typename Arc::StateId StateId;
+  SetType("ngram");
+  SetInputSymbols(fst.InputSymbols());
+  SetOutputSymbols(fst.OutputSymbols());
+  SetProperties(kStaticProperties);
+
+  // Check basic requirements for an OpenGrm language model Fst.
+  if (!NGramFst<A>::HasRequiredProps(fst)) {
+    FSTERROR() << "NGramFst only accepts OpenGrm language models as input";
+    SetProperties(kError, kError);
+    return;
+  }
+
+  int64 num_states = CountStates(fst);
+  Label *context = new Label[num_states];
+
+  // Find the unigram state by starting from the start state, following
+  // epsilons.
+  StateId unigram = fst.Start();
+  while (1) {
+    if (unigram == kNoStateId) {
+      FSTERROR() << "Could not identify unigram state";
+      SetProperties(kError, kError);
+      return;
+    }
+    ArcIterator<Fst<A>> aiter(fst, unigram);
+    if (aiter.Done()) {
+      LOG(WARNING) << "Unigram state " << unigram << " has no arcs.";
+      break;
+    }
+    if (aiter.Value().ilabel != 0) break;
+    unigram = aiter.Value().nextstate;
+  }
+
+  // Each state's context is determined by the subtree it is under from the
+  // unigram state.
+  std::queue<std::pair<StateId, Label>> label_queue;
+  std::vector<bool> visited(num_states);
+  // Force an epsilon link to the start state.
+  label_queue.push(std::make_pair(fst.Start(), 0));
+  for (ArcIterator<Fst<A>> aiter(fst, unigram); !aiter.Done(); aiter.Next()) {
+    label_queue.push(
+        std::make_pair(aiter.Value().nextstate, aiter.Value().ilabel));
+  }
+  // investigate states in breadth first fashion to assign context words.
+  while (!label_queue.empty()) {
+    std::pair<StateId, Label> &now = label_queue.front();
+    if (!visited[now.first]) {
+      context[now.first] = now.second;
+      visited[now.first] = true;
+      for (ArcIterator<Fst<A>> aiter(fst, now.first); !aiter.Done();
+           aiter.Next()) {
+        const Arc &arc = aiter.Value();
+        if (arc.ilabel != 0) {
+          label_queue.push(std::make_pair(arc.nextstate, now.second));
+        }
+      }
+    }
+    label_queue.pop();
+  }
+  visited.clear();
+
+  // The arc from the start state should be assigned an epsilon to put it
+  // in front of the all other labels (which makes Start state 1 after
+  // unigram which is state 0).
+  context[fst.Start()] = 0;
+
+  // Build the tree of contexts fst by reversing the epsilon arcs from fst.
+  VectorFst<Arc> context_fst;
+  uint64 num_final = 0;
+  for (int i = 0; i < num_states; ++i) {
+    if (fst.Final(i) != Weight::Zero()) {
+      ++num_final;
+    }
+    context_fst.SetFinal(context_fst.AddState(), fst.Final(i));
+  }
+  context_fst.SetStart(unigram);
+  context_fst.SetInputSymbols(fst.InputSymbols());
+  context_fst.SetOutputSymbols(fst.OutputSymbols());
+  int64 num_context_arcs = 0;
+  int64 num_futures = 0;
+  for (StateIterator<Fst<A>> siter(fst); !siter.Done(); siter.Next()) {
+    const StateId &state = siter.Value();
+    num_futures += fst.NumArcs(state) - fst.NumInputEpsilons(state);
+    ArcIterator<Fst<A>> aiter(fst, state);
+    if (!aiter.Done()) {
+      const Arc &arc = aiter.Value();
+      // this arc goes from state to arc.nextstate, so create an arc from
+      // arc.nextstate to state to reverse it.
+      if (arc.ilabel == 0) {
+        context_fst.AddArc(arc.nextstate, Arc(context[state], context[state],
+                                              arc.weight, state));
+        num_context_arcs++;
+      }
+    }
+  }
+  if (num_context_arcs != context_fst.NumStates() - 1) {
+    FSTERROR() << "Number of contexts arcs != number of states - 1";
+    SetProperties(kError, kError);
+    return;
+  }
+  if (context_fst.NumStates() != num_states) {
+    FSTERROR() << "Number of contexts != number of states";
+    SetProperties(kError, kError);
+    return;
+  }
+  int64 context_props =
+      context_fst.Properties(kIDeterministic | kILabelSorted, true);
+  if (!(context_props & kIDeterministic)) {
+    FSTERROR() << "Input Fst is not structured properly";
+    SetProperties(kError, kError);
+    return;
+  }
+  if (!(context_props & kILabelSorted)) {
+    ArcSort(&context_fst, ILabelCompare<Arc>());
+  }
+
+  delete[] context;
+
+  uint64 b64;
+  Weight weight;
+  Label label = kNoLabel;
+  const size_t storage = Storage(num_states, num_futures, num_final);
+  MappedFile *data_region = MappedFile::Allocate(storage);
+  char *data = reinterpret_cast<char *>(data_region->mutable_data());
+  memset(data, 0, storage);
+  size_t offset = 0;
+  memcpy(data + offset, reinterpret_cast<char *>(&num_states),
+         sizeof(num_states));
+  offset += sizeof(num_states);
+  memcpy(data + offset, reinterpret_cast<char *>(&num_futures),
+         sizeof(num_futures));
+  offset += sizeof(num_futures);
+  memcpy(data + offset, reinterpret_cast<char *>(&num_final),
+         sizeof(num_final));
+  offset += sizeof(num_final);
+  uint64 *context_bits = reinterpret_cast<uint64 *>(data + offset);
+  offset += BitmapIndex::StorageSize(num_states * 2 + 1) * sizeof(b64);
+  uint64 *future_bits = reinterpret_cast<uint64 *>(data + offset);
+  offset +=
+      BitmapIndex::StorageSize(num_futures + num_states + 1) * sizeof(b64);
+  uint64 *final_bits = reinterpret_cast<uint64 *>(data + offset);
+  offset += BitmapIndex::StorageSize(num_states) * sizeof(b64);
+  Label *context_words = reinterpret_cast<Label *>(data + offset);
+  offset += (num_states + 1) * sizeof(label);
+  Label *future_words = reinterpret_cast<Label *>(data + offset);
+  offset += num_futures * sizeof(label);
+  offset = (offset + sizeof(weight) - 1) & ~(sizeof(weight) - 1);
+  Weight *backoff = reinterpret_cast<Weight *>(data + offset);
+  offset += (num_states + 1) * sizeof(weight);
+  Weight *final_probs = reinterpret_cast<Weight *>(data + offset);
+  offset += num_final * sizeof(weight);
+  Weight *future_probs = reinterpret_cast<Weight *>(data + offset);
+  int64 context_arc = 0, future_arc = 0, context_bit = 0, future_bit = 0,
+        final_bit = 0;
+
+  // pseudo-root bits
+  BitmapIndex::Set(context_bits, context_bit++);
+  ++context_bit;
+  context_words[context_arc] = label;
+  backoff[context_arc] = Weight::Zero();
+  context_arc++;
+
+  ++future_bit;
+  if (order_out) {
+    order_out->clear();
+    order_out->resize(num_states);
+  }
+
+  std::queue<StateId> context_q;
+  context_q.push(context_fst.Start());
+  StateId state_number = 0;
+  while (!context_q.empty()) {
+    const StateId &state = context_q.front();
+    if (order_out) {
+      (*order_out)[state] = state_number;
+    }
+
+    const Weight final_weight = context_fst.Final(state);
+    if (final_weight != Weight::Zero()) {
+      BitmapIndex::Set(final_bits, state_number);
+      final_probs[final_bit] = final_weight;
+      ++final_bit;
+    }
+
+    for (ArcIterator<VectorFst<A>> aiter(context_fst, state); !aiter.Done();
+         aiter.Next()) {
+      const Arc &arc = aiter.Value();
+      context_words[context_arc] = arc.ilabel;
+      backoff[context_arc] = arc.weight;
+      ++context_arc;
+      BitmapIndex::Set(context_bits, context_bit++);
+      context_q.push(arc.nextstate);
+    }
+    ++context_bit;
+
+    for (ArcIterator<Fst<A>> aiter(fst, state); !aiter.Done(); aiter.Next()) {
+      const Arc &arc = aiter.Value();
+      if (arc.ilabel != 0) {
+        future_words[future_arc] = arc.ilabel;
+        future_probs[future_arc] = arc.weight;
+        ++future_arc;
+        BitmapIndex::Set(future_bits, future_bit++);
+      }
+    }
+    ++future_bit;
+    ++state_number;
+    context_q.pop();
+  }
+
+  if ((state_number != num_states) || (context_bit != num_states * 2 + 1) ||
+      (context_arc != num_states) || (future_arc != num_futures) ||
+      (future_bit != num_futures + num_states + 1) ||
+      (final_bit != num_final)) {
+    FSTERROR() << "Structure problems detected during construction";
+    SetProperties(kError, kError);
+    return;
+  }
+
+  Init(data, false, data_region);
+}
+
+template <typename A>
+inline void NGramFstImpl<A>::Init(const char *data, bool owned,
+                                  MappedFile *data_region) {
+  if (owned_) {
+    delete[] data_;
+  }
+  data_region_.reset(data_region);
+  owned_ = owned;
+  data_ = data;
+  size_t offset = 0;
+  num_states_ = *(reinterpret_cast<const uint64 *>(data_ + offset));
+  offset += sizeof(num_states_);
+  num_futures_ = *(reinterpret_cast<const uint64 *>(data_ + offset));
+  offset += sizeof(num_futures_);
+  num_final_ = *(reinterpret_cast<const uint64 *>(data_ + offset));
+  offset += sizeof(num_final_);
+  uint64 bits;
+  size_t context_bits = num_states_ * 2 + 1;
+  size_t future_bits = num_futures_ + num_states_ + 1;
+  context_ = reinterpret_cast<const uint64 *>(data_ + offset);
+  offset += BitmapIndex::StorageSize(context_bits) * sizeof(bits);
+  future_ = reinterpret_cast<const uint64 *>(data_ + offset);
+  offset += BitmapIndex::StorageSize(future_bits) * sizeof(bits);
+  final_ = reinterpret_cast<const uint64 *>(data_ + offset);
+  offset += BitmapIndex::StorageSize(num_states_) * sizeof(bits);
+  context_words_ = reinterpret_cast<const Label *>(data_ + offset);
+  offset += (num_states_ + 1) * sizeof(*context_words_);
+  future_words_ = reinterpret_cast<const Label *>(data_ + offset);
+  offset += num_futures_ * sizeof(*future_words_);
+  offset = (offset + sizeof(*backoff_) - 1) & ~(sizeof(*backoff_) - 1);
+  backoff_ = reinterpret_cast<const Weight *>(data_ + offset);
+  offset += (num_states_ + 1) * sizeof(*backoff_);
+  final_probs_ = reinterpret_cast<const Weight *>(data_ + offset);
+  offset += num_final_ * sizeof(*final_probs_);
+  future_probs_ = reinterpret_cast<const Weight *>(data_ + offset);
+
+  context_index_.BuildIndex(context_, context_bits);
+  future_index_.BuildIndex(future_, future_bits);
+  final_index_.BuildIndex(final_, num_states_);
+
+  select_root_ = context_index_.Select0s(0);
+  if (context_index_.Rank1(0) != 0 || select_root_.first != 1 ||
+      context_index_.Get(2) == false) {
+    FSTERROR() << "Malformed file";
+    SetProperties(kError, kError);
+    return;
+  }
+  root_children_ = context_words_ + context_index_.Rank1(2);
+  start_ = 1;
+}
+
+template <typename A>
+inline typename A::StateId NGramFstImpl<A>::Transition(
+    const std::vector<Label> &context, Label future) const {
+  const Label *children = root_children_;
+  size_t num_children = select_root_.second - 2;
+  const Label *loc =
+      std::lower_bound(children, children + num_children, future);
+  if (loc == children + num_children || *loc != future) {
+    return context_index_.Rank1(0);
+  }
+  size_t node = 2 + loc - children;
+  size_t node_rank = context_index_.Rank1(node);
+  std::pair<size_t, size_t> zeros =
+      (node_rank == 0) ? select_root_ : context_index_.Select0s(node_rank);
+  size_t first_child = zeros.first + 1;
+  if (context_index_.Get(first_child) == false) {
+    return context_index_.Rank1(node);
+  }
+  size_t last_child = zeros.second - 1;
+  for (int word = context.size() - 1; word >= 0; --word) {
+    children = context_words_ + context_index_.Rank1(first_child);
+    loc = std::lower_bound(children, children + last_child - first_child + 1,
+                           context[word]);
+    if (loc == children + last_child - first_child + 1 ||
+        *loc != context[word]) {
+      break;
+    }
+    node = first_child + loc - children;
+    node_rank = context_index_.Rank1(node);
+    zeros =
+        (node_rank == 0) ? select_root_ : context_index_.Select0s(node_rank);
+    first_child = zeros.first + 1;
+    if (context_index_.Get(first_child) == false) break;
+    last_child = zeros.second - 1;
+  }
+  return context_index_.Rank1(node);
+}
+
+}  // namespace internal
+
+/*****************************************************************************/
+template <class A>
+class NGramFstMatcher : public MatcherBase<A> {
+ public:
+  typedef A Arc;
+  typedef typename A::Label Label;
+  typedef typename A::StateId StateId;
+  typedef typename A::Weight Weight;
+
+  // This makes a copy of the FST.
+  NGramFstMatcher(const NGramFst<A> &fst, MatchType match_type)
+      : owned_fst_(fst.Copy()),
+        fst_(*owned_fst_),
+        inst_(fst_.inst_),
+        match_type_(match_type),
+        current_loop_(false),
+        loop_(kNoLabel, 0, A::Weight::One(), kNoStateId) {
+    if (match_type_ == MATCH_OUTPUT) {
+      std::swap(loop_.ilabel, loop_.olabel);
+    }
+  }
+
+  // This doesn't copy the FST.
+  NGramFstMatcher(const NGramFst<A> *fst, MatchType match_type)
+      : fst_(*fst),
+        inst_(fst_.inst_),
+        match_type_(match_type),
+        current_loop_(false),
+        loop_(kNoLabel, 0, A::Weight::One(), kNoStateId) {
+    if (match_type_ == MATCH_OUTPUT) {
+      std::swap(loop_.ilabel, loop_.olabel);
+    }
+  }
+
+  // This makes a copy of the FST.
+  NGramFstMatcher(const NGramFstMatcher<A> &matcher, bool safe = false)
+      : owned_fst_(matcher.fst_.Copy(safe)),
+        fst_(*owned_fst_),
+        inst_(matcher.inst_),
+        match_type_(matcher.match_type_),
+        current_loop_(false),
+        loop_(kNoLabel, 0, A::Weight::One(), kNoStateId) {
+    if (match_type_ == MATCH_OUTPUT) {
+      std::swap(loop_.ilabel, loop_.olabel);
+    }
+  }
+
+  NGramFstMatcher<A> *Copy(bool safe = false) const override {
+    return new NGramFstMatcher<A>(*this, safe);
+  }
+
+  MatchType Type(bool test) const override { return match_type_; }
+
+  const Fst<A> &GetFst() const override { return fst_; }
+
+  uint64 Properties(uint64 props) const override { return props; }
+
+  void SetState(StateId s) final {
+    fst_.GetImpl()->SetInstFuture(s, &inst_);
+    current_loop_ = false;
+  }
+
+  bool Find(Label label) final {
+    const Label nolabel = kNoLabel;
+    done_ = true;
+    if (label == 0 || label == nolabel) {
+      if (label == 0) {
+        current_loop_ = true;
+        loop_.nextstate = inst_.state_;
+      }
+      // The unigram state has no epsilon arc.
+      if (inst_.state_ != 0) {
+        arc_.ilabel = arc_.olabel = 0;
+        fst_.GetImpl()->SetInstNode(&inst_);
+        arc_.nextstate = fst_.GetImpl()->context_index_.Rank1(
+            fst_.GetImpl()->context_index_.Select1(
+                fst_.GetImpl()->context_index_.Rank0(inst_.node_) - 1));
+        arc_.weight = fst_.GetImpl()->backoff_[inst_.state_];
+        done_ = false;
+      }
+    } else {
+      current_loop_ = false;
+      const Label *start = fst_.GetImpl()->future_words_ + inst_.offset_;
+      const Label *end = start + inst_.num_futures_;
+      const Label *search = std::lower_bound(start, end, label);
+      if (search != end && *search == label) {
+        size_t state = search - start;
+        arc_.ilabel = arc_.olabel = label;
+        arc_.weight = fst_.GetImpl()->future_probs_[inst_.offset_ + state];
+        fst_.GetImpl()->SetInstContext(&inst_);
+        arc_.nextstate = fst_.GetImpl()->Transition(inst_.context_, label);
+        done_ = false;
+      }
+    }
+    return !Done();
+  }
+
+  bool Done() const final { return !current_loop_ && done_; }
+
+  const Arc &Value() const final { return (current_loop_) ? loop_ : arc_; }
+
+  void Next() final {
+    if (current_loop_) {
+      current_loop_ = false;
+    } else {
+      done_ = true;
+    }
+  }
+
+  ssize_t Priority(StateId s) final { return fst_.NumArcs(s); }
+
+ private:
+  std::unique_ptr<NGramFst<A>> owned_fst_;
+  const NGramFst<A> &fst_;
+  NGramFstInst<A> inst_;
+  MatchType match_type_;  // Supplied by caller
+  bool done_;
+  Arc arc_;
+  bool current_loop_;  // Current arc is the implicit loop
+  Arc loop_;
+};
+
+/*****************************************************************************/
+// Specialization for NGramFst; see generic version in fst.h
+// for sample usage (but use the ProdLmFst type!). This version
+// should inline.
+template <class A>
+class StateIterator<NGramFst<A>> : public StateIteratorBase<A> {
+ public:
+  typedef typename A::StateId StateId;
+
+  explicit StateIterator(const NGramFst<A> &fst)
+      : s_(0), num_states_(fst.NumStates()) {}
+
+  bool Done() const final { return s_ >= num_states_; }
+
+  StateId Value() const final { return s_; }
+
+  void Next() final { ++s_; }
+
+  void Reset() final { s_ = 0; }
+
+ private:
+  StateId s_;
+  StateId num_states_;
+};
+
+/*****************************************************************************/
+template <class A>
+class ArcIterator<NGramFst<A>> : public ArcIteratorBase<A> {
+ public:
+  typedef A Arc;
+  typedef typename A::Label Label;
+  typedef typename A::StateId StateId;
+  typedef typename A::Weight Weight;
+
+  ArcIterator(const NGramFst<A> &fst, StateId state)
+      : lazy_(~0), impl_(fst.GetImpl()), i_(0), flags_(kArcValueFlags) {
+    inst_ = fst.inst_;
+    impl_->SetInstFuture(state, &inst_);
+    impl_->SetInstNode(&inst_);
+  }
+
+  bool Done() const final {
+    return i_ >=
+           ((inst_.node_ == 0) ? inst_.num_futures_ : inst_.num_futures_ + 1);
+  }
+
+  const Arc &Value() const final {
+    bool eps = (inst_.node_ != 0 && i_ == 0);
+    StateId state = (inst_.node_ == 0) ? i_ : i_ - 1;
+    if (flags_ & lazy_ & (kArcILabelValue | kArcOLabelValue)) {
+      arc_.ilabel = arc_.olabel =
+          eps ? 0 : impl_->future_words_[inst_.offset_ + state];
+      lazy_ &= ~(kArcILabelValue | kArcOLabelValue);
+    }
+    if (flags_ & lazy_ & kArcNextStateValue) {
+      if (eps) {
+        arc_.nextstate =
+            impl_->context_index_.Rank1(impl_->context_index_.Select1(
+                impl_->context_index_.Rank0(inst_.node_) - 1));
+      } else {
+        if (lazy_ & kArcNextStateValue) {
+          impl_->SetInstContext(&inst_);  // first time only.
+        }
+        arc_.nextstate = impl_->Transition(
+            inst_.context_, impl_->future_words_[inst_.offset_ + state]);
+      }
+      lazy_ &= ~kArcNextStateValue;
+    }
+    if (flags_ & lazy_ & kArcWeightValue) {
+      arc_.weight = eps ? impl_->backoff_[inst_.state_]
+                        : impl_->future_probs_[inst_.offset_ + state];
+      lazy_ &= ~kArcWeightValue;
+    }
+    return arc_;
+  }
+
+  void Next() final {
+    ++i_;
+    lazy_ = ~0;
+  }
+
+  size_t Position() const final { return i_; }
+
+  void Reset() final {
+    i_ = 0;
+    lazy_ = ~0;
+  }
+
+  void Seek(size_t a) final {
+    if (i_ != a) {
+      i_ = a;
+      lazy_ = ~0;
+    }
+  }
+
+  uint32 Flags() const final { return flags_; }
+
+  void SetFlags(uint32 flags, uint32 mask) final {
+    flags_ &= ~mask;
+    flags_ |= (flags & kArcValueFlags);
+  }
+
+ private:
+  mutable Arc arc_;
+  mutable uint32 lazy_;
+  const internal::NGramFstImpl<A> *impl_;  // Borrowed reference.
+  mutable NGramFstInst<A> inst_;
+
+  size_t i_;
+  uint32 flags_;
+};
+
+}  // namespace fst
+#endif  // FST_EXTENSIONS_NGRAM_NGRAM_FST_H_
diff --git a/projects/llm_framework/include/fst/extensions/ngram/nthbit.h b/projects/llm_framework/include/fst/extensions/ngram/nthbit.h
new file mode 100644
index 00000000..1e6ec635
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/ngram/nthbit.h
@@ -0,0 +1,49 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_EXTENSIONS_NGRAM_NTHBIT_H_
+#define FST_EXTENSIONS_NGRAM_NTHBIT_H_
+
+#include <fst/types.h>
+#include <fst/compat.h>
+
+#ifdef __BMI2__
+// PDEP requires BMI2.
+
+// Returns the position (0-63) of the r-th 1 bit in v.
+// 1 <= r <= CountOnes(v) <= 64.  Therefore, v must not be 0.
+inline uint32 nth_bit(uint64 v, uint32 r) {
+  // PDEP example from https://stackoverflow.com/a/27453505
+  return __builtin_ctzll(_pdep_u64(uint64{1} << (r - 1), v));
+}
+
+#else  // !defined(__BMI2__)
+
+extern const uint32 nth_bit_bit_offset[];
+
+// Returns the position (0-63) of the r-th 1 bit in v.
+// 1 <= r <= CountOnes(v) <= 64.  Therefore, v must not be 0.
+inline uint32 nth_bit(uint64 v, uint32 r) {
+  uint32 shift = 0;
+  uint32 c = __builtin_popcount(v & 0xffffffff);
+  uint32 mask = -(r > c);
+  r -= c & mask;
+  shift += (32 & mask);
+
+  c = __builtin_popcount((v >> shift) & 0xffff);
+  mask = -(r > c);
+  r -= c & mask;
+  shift += (16 & mask);
+
+  c = __builtin_popcount((v >> shift) & 0xff);
+  mask = -(r > c);
+  r -= c & mask;
+  shift += (8 & mask);
+
+  return shift +
+         ((nth_bit_bit_offset[(v >> shift) & 0xff] >> ((r - 1) << 2)) & 0xf);
+}
+
+#endif  // !defined(__BMI2__)
+
+#endif  // FST_EXTENSIONS_NGRAM_NTHBIT_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/collection.h b/projects/llm_framework/include/fst/extensions/pdt/collection.h
new file mode 100644
index 00000000..ae34aba7
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/pdt/collection.h
@@ -0,0 +1,107 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to store a collection of ordered (multi-)sets with elements of type T.
+
+#ifndef FST_EXTENSIONS_PDT_COLLECTION_H_
+#define FST_EXTENSIONS_PDT_COLLECTION_H_
+
+#include <functional>
+#include <vector>
+
+#include <fst/log.h>
+#include <fst/bi-table.h>
+
+namespace fst {
+
+// Stores a collection of non-empty, ordered (multi-)sets with elements of type
+// T. A default constructor, operator==, and an STL-style hash functor must be
+// defined on the elements. Provides signed integer ID (of type I) for each
+// unique set. The IDs are allocated starting from 0 in order.
+template <class I, class T>
+class Collection {
+ public:
+  struct Node {  // Trie node.
+    I node_id;   // Root is kNoNodeId;
+    T element;
+
+    Node() : node_id(kNoNodeId), element(T()) {}
+
+    Node(I i, const T &t) : node_id(i), element(t) {}
+
+    bool operator==(const Node &n) const {
+      return n.node_id == node_id && n.element == element;
+    }
+  };
+
+  struct NodeHash {
+    size_t operator()(const Node &n) const {
+      static constexpr auto kPrime = 7853;
+      return n.node_id + hash_(n.element) * kPrime;
+    }
+  };
+
+  using NodeTable = CompactHashBiTable<I, Node, NodeHash>;
+
+  class SetIterator {
+   public:
+    SetIterator(I id, Node node, NodeTable *node_table)
+        : id_(id), node_(node), node_table_(node_table) {}
+
+    bool Done() const { return id_ == kNoNodeId; }
+
+    const T &Element() const { return node_.element; }
+
+    void Next() {
+      id_ = node_.node_id;
+      if (id_ != kNoNodeId) node_ = node_table_->FindEntry(id_);
+    }
+
+   private:
+    I id_;       // Iterator set node ID.
+    Node node_;  // Iterator set node.
+    NodeTable *node_table_;
+  };
+
+  Collection() {}
+
+  // Looks up integer ID from ordered multi-se, and if it doesn't exist and
+  // insert is true, then adds it. Otherwise returns -1.
+  I FindId(const std::vector<T> &set, bool insert = true) {
+    I node_id = kNoNodeId;
+    for (ssize_t i = set.size() - 1; i >= 0; --i) {
+      Node node(node_id, set[i]);
+      node_id = node_table_.FindId(node, insert);
+      if (node_id == -1) break;
+    }
+    return node_id;
+  }
+
+  // Finds ordered (multi-)set given integer ID. Returns set iterator to
+  // traverse result.
+  SetIterator FindSet(I id) {
+    if (id < 0 || id >= node_table_.Size()) {
+      return SetIterator(kNoNodeId, Node(kNoNodeId, T()), &node_table_);
+    } else {
+      return SetIterator(id, node_table_.FindEntry(id), &node_table_);
+    }
+  }
+
+  I Size() const { return node_table_.Size(); }
+
+ private:
+  static constexpr I kNoNodeId = -1;
+  static const std::hash<T> hash_;
+
+  NodeTable node_table_;
+};
+
+template <class I, class T>
+constexpr I Collection<I, T>::kNoNodeId;
+
+template <class I, class T>
+const std::hash<T> Collection<I, T>::hash_ = {};
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_PDT_COLLECTION_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/compose.h b/projects/llm_framework/include/fst/extensions/pdt/compose.h
new file mode 100644
index 00000000..525d613a
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/pdt/compose.h
@@ -0,0 +1,493 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Composes a PDT and an FST.
+
+#ifndef FST_EXTENSIONS_PDT_COMPOSE_H_
+#define FST_EXTENSIONS_PDT_COMPOSE_H_
+
+#include <list>
+
+#include <fst/extensions/pdt/pdt.h>
+#include <fst/compose.h>
+
+namespace fst {
+
+// Returns paren arcs for Find(kNoLabel).
+constexpr uint32 kParenList = 0x00000001;
+
+// Returns a kNolabel loop for Find(paren).
+constexpr uint32 kParenLoop = 0x00000002;
+
+// This class is a matcher that treats parens as multi-epsilon labels.
+// It is most efficient if the parens are in a range non-overlapping with
+// the non-paren labels.
+template <class F>
+class ParenMatcher {
+ public:
+  using FST = F;
+  using M = SortedMatcher<FST>;
+  using Arc = typename FST::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  // This makes a copy of the FST.
+  ParenMatcher(const FST &fst, MatchType match_type,
+               uint32 flags = (kParenLoop | kParenList))
+      : matcher_(fst, match_type), match_type_(match_type), flags_(flags) {
+    if (match_type == MATCH_INPUT) {
+      loop_.ilabel = kNoLabel;
+      loop_.olabel = 0;
+    } else {
+      loop_.ilabel = 0;
+      loop_.olabel = kNoLabel;
+    }
+    loop_.weight = Weight::One();
+    loop_.nextstate = kNoStateId;
+  }
+
+  // This doesn't copy the FST.
+  ParenMatcher(const FST *fst, MatchType match_type,
+               uint32 flags = (kParenLoop | kParenList))
+      : matcher_(fst, match_type), match_type_(match_type), flags_(flags) {
+    if (match_type == MATCH_INPUT) {
+      loop_.ilabel = kNoLabel;
+      loop_.olabel = 0;
+    } else {
+      loop_.ilabel = 0;
+      loop_.olabel = kNoLabel;
+    }
+    loop_.weight = Weight::One();
+    loop_.nextstate = kNoStateId;
+  }
+
+  // This makes a copy of the FST.
+  ParenMatcher(const ParenMatcher<FST> &matcher, bool safe = false)
+      : matcher_(matcher.matcher_, safe),
+        match_type_(matcher.match_type_),
+        flags_(matcher.flags_),
+        open_parens_(matcher.open_parens_),
+        close_parens_(matcher.close_parens_),
+        loop_(matcher.loop_) {
+    loop_.nextstate = kNoStateId;
+  }
+
+  ParenMatcher<FST> *Copy(bool safe = false) const {
+    return new ParenMatcher<FST>(*this, safe);
+  }
+
+  MatchType Type(bool test) const { return matcher_.Type(test); }
+
+  void SetState(StateId s) {
+    matcher_.SetState(s);
+    loop_.nextstate = s;
+  }
+
+  bool Find(Label match_label);
+
+  bool Done() const { return done_; }
+
+  const Arc &Value() const { return paren_loop_ ? loop_ : matcher_.Value(); }
+
+  void Next();
+
+  Weight Final(StateId s) { return matcher_.Final(s); }
+
+  ssize_t Priority(StateId s) { return matcher_.Priority(s); }
+
+  const FST &GetFst() const { return matcher_.GetFst(); }
+
+  uint64 Properties(uint64 props) const { return matcher_.Properties(props); }
+
+  uint32 Flags() const { return matcher_.Flags(); }
+
+  void AddOpenParen(Label label) {
+    if (label == 0) {
+      FSTERROR() << "ParenMatcher: Bad open paren label: 0";
+    } else {
+      open_parens_.Insert(label);
+    }
+  }
+
+  void AddCloseParen(Label label) {
+    if (label == 0) {
+      FSTERROR() << "ParenMatcher: Bad close paren label: 0";
+    } else {
+      close_parens_.Insert(label);
+    }
+  }
+
+  void RemoveOpenParen(Label label) {
+    if (label == 0) {
+      FSTERROR() << "ParenMatcher: Bad open paren label: 0";
+    } else {
+      open_parens_.Erase(label);
+    }
+  }
+
+  void RemoveCloseParen(Label label) {
+    if (label == 0) {
+      FSTERROR() << "ParenMatcher: Bad close paren label: 0";
+    } else {
+      close_parens_.Erase(label);
+    }
+  }
+
+  void ClearOpenParens() { open_parens_.Clear(); }
+
+  void ClearCloseParens() { close_parens_.Clear(); }
+
+  bool IsOpenParen(Label label) const { return open_parens_.Member(label); }
+
+  bool IsCloseParen(Label label) const { return close_parens_.Member(label); }
+
+ private:
+  // Advances matcher to next open paren, returning true if it exists.
+  bool NextOpenParen();
+
+  // Advances matcher to next close paren, returning true if it exists.
+  bool NextCloseParen();
+
+  M matcher_;
+  MatchType match_type_;  // Type of match to perform.
+  uint32 flags_;
+  // Open paren label set.
+  CompactSet<Label, kNoLabel> open_parens_;
+  // Close paren label set.
+  CompactSet<Label, kNoLabel> close_parens_;
+  bool open_paren_list_;   // Matching open paren list?
+  bool close_paren_list_;  // Matching close paren list?
+  bool paren_loop_;        // Current arc is the implicit paren loop?
+  mutable Arc loop_;       // For non-consuming symbols.
+  bool done_;              // Matching done?
+
+  ParenMatcher &operator=(const ParenMatcher &) = delete;
+};
+
+template <class FST>
+inline bool ParenMatcher<FST>::Find(Label match_label) {
+  open_paren_list_ = false;
+  close_paren_list_ = false;
+  paren_loop_ = false;
+  done_ = false;
+  // Returns all parenthesis arcs.
+  if (match_label == kNoLabel && (flags_ & kParenList)) {
+    if (open_parens_.LowerBound() != kNoLabel) {
+      matcher_.LowerBound(open_parens_.LowerBound());
+      open_paren_list_ = NextOpenParen();
+      if (open_paren_list_) return true;
+    }
+    if (close_parens_.LowerBound() != kNoLabel) {
+      matcher_.LowerBound(close_parens_.LowerBound());
+      close_paren_list_ = NextCloseParen();
+      if (close_paren_list_) return true;
+    }
+  }
+  // Returns the implicit paren loop.
+  if (match_label > 0 && (flags_ & kParenLoop) &&
+      (IsOpenParen(match_label) || IsCloseParen(match_label))) {
+    paren_loop_ = true;
+    return true;
+  }
+  // Returns all other labels.
+  if (matcher_.Find(match_label)) return true;
+  done_ = true;
+  return false;
+}
+
+template <class FST>
+inline void ParenMatcher<FST>::Next() {
+  if (paren_loop_) {
+    paren_loop_ = false;
+    done_ = true;
+  } else if (open_paren_list_) {
+    matcher_.Next();
+    open_paren_list_ = NextOpenParen();
+    if (open_paren_list_) return;
+    if (close_parens_.LowerBound() != kNoLabel) {
+      matcher_.LowerBound(close_parens_.LowerBound());
+      close_paren_list_ = NextCloseParen();
+      if (close_paren_list_) return;
+    }
+    done_ = !matcher_.Find(kNoLabel);
+  } else if (close_paren_list_) {
+    matcher_.Next();
+    close_paren_list_ = NextCloseParen();
+    if (close_paren_list_) return;
+    done_ = !matcher_.Find(kNoLabel);
+  } else {
+    matcher_.Next();
+    done_ = matcher_.Done();
+  }
+}
+
+// Advances matcher to next open paren, returning true if it exists.
+template <class FST>
+inline bool ParenMatcher<FST>::NextOpenParen() {
+  for (; !matcher_.Done(); matcher_.Next()) {
+    Label label = match_type_ == MATCH_INPUT ? matcher_.Value().ilabel
+                                             : matcher_.Value().olabel;
+    if (label > open_parens_.UpperBound()) return false;
+    if (IsOpenParen(label)) return true;
+  }
+  return false;
+}
+
+// Advances matcher to next close paren, returning true if it exists.
+template <class FST>
+inline bool ParenMatcher<FST>::NextCloseParen() {
+  for (; !matcher_.Done(); matcher_.Next()) {
+    Label label = match_type_ == MATCH_INPUT ? matcher_.Value().ilabel
+                                             : matcher_.Value().olabel;
+    if (label > close_parens_.UpperBound()) return false;
+    if (IsCloseParen(label)) return true;
+  }
+  return false;
+}
+
+template <class Filter>
+class ParenFilter {
+ public:
+  using FST1 = typename Filter::FST1;
+  using FST2 = typename Filter::FST2;
+  using Arc = typename Filter::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Matcher1 = typename Filter::Matcher1;
+  using Matcher2 = typename Filter::Matcher2;
+
+  using StackId = StateId;
+  using ParenStack = PdtStack<StackId, Label>;
+  using FilterState1 = typename Filter::FilterState;
+  using FilterState2 = IntegerFilterState<StackId>;
+  using FilterState = PairFilterState<FilterState1, FilterState2>;
+
+  ParenFilter(const FST1 &fst1, const FST2 &fst2, Matcher1 *matcher1 = nullptr,
+              Matcher2 *matcher2 = nullptr,
+              const std::vector<std::pair<Label, Label>> *parens = nullptr,
+              bool expand = false, bool keep_parens = true)
+      : filter_(fst1, fst2, matcher1, matcher2),
+        parens_(parens ? *parens : std::vector<std::pair<Label, Label>>()),
+        expand_(expand),
+        keep_parens_(keep_parens),
+        fs_(FilterState::NoState()),
+        stack_(parens_),
+        paren_id_(-1) {
+    if (parens) {
+      for (const auto &pair : *parens) {
+        parens_.push_back(pair);
+        GetMatcher1()->AddOpenParen(pair.first);
+        GetMatcher2()->AddOpenParen(pair.first);
+        if (!expand_) {
+          GetMatcher1()->AddCloseParen(pair.second);
+          GetMatcher2()->AddCloseParen(pair.second);
+        }
+      }
+    }
+  }
+
+  ParenFilter(const ParenFilter &filter, bool safe = false)
+      : filter_(filter.filter_, safe),
+        parens_(filter.parens_),
+        expand_(filter.expand_),
+        keep_parens_(filter.keep_parens_),
+        fs_(FilterState::NoState()),
+        stack_(filter.parens_),
+        paren_id_(-1) {}
+
+  FilterState Start() const {
+    return FilterState(filter_.Start(), FilterState2(0));
+  }
+
+  void SetState(StateId s1, StateId s2, const FilterState &fs) {
+    fs_ = fs;
+    filter_.SetState(s1, s2, fs_.GetState1());
+    if (!expand_) return;
+    ssize_t paren_id = stack_.Top(fs.GetState2().GetState());
+    if (paren_id != paren_id_) {
+      if (paren_id_ != -1) {
+        GetMatcher1()->RemoveCloseParen(parens_[paren_id_].second);
+        GetMatcher2()->RemoveCloseParen(parens_[paren_id_].second);
+      }
+      paren_id_ = paren_id;
+      if (paren_id_ != -1) {
+        GetMatcher1()->AddCloseParen(parens_[paren_id_].second);
+        GetMatcher2()->AddCloseParen(parens_[paren_id_].second);
+      }
+    }
+  }
+
+  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
+    const auto fs1 = filter_.FilterArc(arc1, arc2);
+    const auto &fs2 = fs_.GetState2();
+    if (fs1 == FilterState1::NoState()) return FilterState::NoState();
+    if (arc1->olabel == kNoLabel && arc2->ilabel) {  // arc2 parentheses.
+      if (keep_parens_) {
+        arc1->ilabel = arc2->ilabel;
+      } else if (arc2->ilabel) {
+        arc2->olabel = arc1->ilabel;
+      }
+      return FilterParen(arc2->ilabel, fs1, fs2);
+    } else if (arc2->ilabel == kNoLabel && arc1->olabel) {  // arc1 parentheses.
+      if (keep_parens_) {
+        arc2->olabel = arc1->olabel;
+      } else {
+        arc1->ilabel = arc2->olabel;
+      }
+      return FilterParen(arc1->olabel, fs1, fs2);
+    } else {
+      return FilterState(fs1, fs2);
+    }
+  }
+
+  void FilterFinal(Weight *w1, Weight *w2) const {
+    if (fs_.GetState2().GetState() != 0) *w1 = Weight::Zero();
+    filter_.FilterFinal(w1, w2);
+  }
+
+  // Returns respective matchers; ownership stays with filter.
+
+  Matcher1 *GetMatcher1() { return filter_.GetMatcher1(); }
+
+  Matcher2 *GetMatcher2() { return filter_.GetMatcher2(); }
+
+  uint64 Properties(uint64 iprops) const {
+    return filter_.Properties(iprops) & kILabelInvariantProperties &
+           kOLabelInvariantProperties;
+  }
+
+ private:
+  const FilterState FilterParen(Label label, const FilterState1 &fs1,
+                                const FilterState2 &fs2) const {
+    if (!expand_) return FilterState(fs1, fs2);
+    const auto stack_id = stack_.Find(fs2.GetState(), label);
+    if (stack_id < 0) {
+      return FilterState::NoState();
+    } else {
+      return FilterState(fs1, FilterState2(stack_id));
+    }
+  }
+
+  Filter filter_;
+  std::vector<std::pair<Label, Label>> parens_;
+  bool expand_;       // Expands to FST?
+  bool keep_parens_;  // Retains parentheses in output?
+  FilterState fs_;    // Current filter state.
+  mutable ParenStack stack_;
+  ssize_t paren_id_;
+};
+
+// Class to setup composition options for PDT composition. Default is to take
+// the PDT as the first composition argument.
+template <class Arc, bool left_pdt = true>
+class PdtComposeFstOptions
+    : public ComposeFstOptions<
+          Arc, ParenMatcher<Fst<Arc>>,
+          ParenFilter<AltSequenceComposeFilter<ParenMatcher<Fst<Arc>>>>> {
+ public:
+  using Label = typename Arc::Label;
+  using PdtMatcher = ParenMatcher<Fst<Arc>>;
+  using PdtFilter = ParenFilter<AltSequenceComposeFilter<PdtMatcher>>;
+
+  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::matcher1;
+  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::matcher2;
+  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::filter;
+
+  PdtComposeFstOptions(const Fst<Arc> &ifst1,
+                       const std::vector<std::pair<Label, Label>> &parens,
+                       const Fst<Arc> &ifst2, bool expand = false,
+                       bool keep_parens = true) {
+    matcher1 = new PdtMatcher(ifst1, MATCH_OUTPUT, kParenList);
+    matcher2 = new PdtMatcher(ifst2, MATCH_INPUT, kParenLoop);
+    filter = new PdtFilter(ifst1, ifst2, matcher1, matcher2, &parens, expand,
+                           keep_parens);
+  }
+};
+
+// Class to setup composition options for PDT with FST composition.
+// Specialization is for the FST as the first composition argument.
+template <class Arc>
+class PdtComposeFstOptions<Arc, false>
+    : public ComposeFstOptions<
+          Arc, ParenMatcher<Fst<Arc>>,
+          ParenFilter<SequenceComposeFilter<ParenMatcher<Fst<Arc>>>>> {
+ public:
+  using Label = typename Arc::Label;
+  using PdtMatcher = ParenMatcher<Fst<Arc>>;
+  using PdtFilter = ParenFilter<SequenceComposeFilter<PdtMatcher>>;
+
+  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::matcher1;
+  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::matcher2;
+  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::filter;
+
+  PdtComposeFstOptions(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
+                       const std::vector<std::pair<Label, Label>> &parens,
+                       bool expand = false, bool keep_parens = true) {
+    matcher1 = new PdtMatcher(ifst1, MATCH_OUTPUT, kParenLoop);
+    matcher2 = new PdtMatcher(ifst2, MATCH_INPUT, kParenList);
+    filter = new PdtFilter(ifst1, ifst2, matcher1, matcher2, &parens, expand,
+                           keep_parens);
+  }
+};
+
+enum PdtComposeFilter {
+  PAREN_FILTER,         // Bar-Hillel construction; keeps parentheses.
+  EXPAND_FILTER,        // Bar-Hillel + expansion; removes parentheses.
+  EXPAND_PAREN_FILTER,  // Bar-Hillel + expansion; keeps parentheses.
+};
+
+struct PdtComposeOptions {
+  bool connect;                  // Connect output?
+  PdtComposeFilter filter_type;  // Pre-defined filter to use.
+
+  explicit PdtComposeOptions(bool connect = true,
+                             PdtComposeFilter filter_type = PAREN_FILTER)
+      : connect(connect), filter_type(filter_type) {}
+};
+
+// Composes pushdown transducer (PDT) encoded as an FST (1st arg) and an FST
+// (2nd arg) with the result also a PDT encoded as an FST (3rd arg). In the
+// PDTs, some transitions are labeled with open or close parentheses. To be
+// interpreted as a PDT, the parens must balance on a path (see PdtExpand()).
+// The open-close parenthesis label pairs are passed using the parens argument.
+template <class Arc>
+void Compose(const Fst<Arc> &ifst1,
+             const std::vector<
+                 std::pair<typename Arc::Label, typename Arc::Label>> &parens,
+             const Fst<Arc> &ifst2, MutableFst<Arc> *ofst,
+             const PdtComposeOptions &opts = PdtComposeOptions()) {
+  bool expand = opts.filter_type != PAREN_FILTER;
+  bool keep_parens = opts.filter_type != EXPAND_FILTER;
+  PdtComposeFstOptions<Arc, true> copts(ifst1, parens, ifst2, expand,
+                                        keep_parens);
+  copts.gc_limit = 0;
+  *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
+  if (opts.connect) Connect(ofst);
+}
+
+// Composes an FST (1st arg) and pushdown transducer (PDT) encoded as an FST
+// (2nd arg) with the result also a PDT encoded as an FST (3rd arg). In the
+// PDTs, some transitions are labeled with open or close parentheses. To be
+// interpreted as a PDT, the parens must balance on a path (see ExpandFst()).
+// The open-close parenthesis label pairs are passed using the parens argument.
+template <class Arc>
+void Compose(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
+             const std::vector<
+                 std::pair<typename Arc::Label, typename Arc::Label>> &parens,
+             MutableFst<Arc> *ofst,
+             const PdtComposeOptions &opts = PdtComposeOptions()) {
+  bool expand = opts.filter_type != PAREN_FILTER;
+  bool keep_parens = opts.filter_type != EXPAND_FILTER;
+  PdtComposeFstOptions<Arc, false> copts(ifst1, ifst2, parens, expand,
+                                         keep_parens);
+  copts.gc_limit = 0;
+  *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
+  if (opts.connect) Connect(ofst);
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_PDT_COMPOSE_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/expand.h b/projects/llm_framework/include/fst/extensions/pdt/expand.h
new file mode 100644
index 00000000..eeee781e
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/pdt/expand.h
@@ -0,0 +1,933 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Expands a PDT to an FST.
+
+#ifndef FST_EXTENSIONS_PDT_EXPAND_H_
+#define FST_EXTENSIONS_PDT_EXPAND_H_
+
+#include <forward_list>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/extensions/pdt/paren.h>
+#include <fst/extensions/pdt/pdt.h>
+#include <fst/extensions/pdt/reverse.h>
+#include <fst/extensions/pdt/shortest-path.h>
+#include <fst/cache.h>
+#include <fst/mutable-fst.h>
+#include <fst/queue.h>
+#include <fst/state-table.h>
+#include <fst/test-properties.h>
+
+namespace fst {
+
+template <class Arc>
+struct PdtExpandFstOptions : public CacheOptions {
+  bool keep_parentheses;
+  PdtStack<typename Arc::StateId, typename Arc::Label> *stack;
+  PdtStateTable<typename Arc::StateId, typename Arc::StateId> *state_table;
+
+  explicit PdtExpandFstOptions(
+      const CacheOptions &opts = CacheOptions(), bool keep_parentheses = false,
+      PdtStack<typename Arc::StateId, typename Arc::Label> *stack = nullptr,
+      PdtStateTable<typename Arc::StateId, typename Arc::StateId> *state_table =
+          nullptr)
+      : CacheOptions(opts),
+        keep_parentheses(keep_parentheses),
+        stack(stack),
+        state_table(state_table) {}
+};
+
+namespace internal {
+
+// Implementation class for PdtExpandFst.
+template <class Arc>
+class PdtExpandFstImpl : public CacheImpl<Arc> {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using StackId = StateId;
+  using StateTuple = PdtStateTuple<StateId, StackId>;
+
+  using FstImpl<Arc>::SetType;
+  using FstImpl<Arc>::SetProperties;
+  using FstImpl<Arc>::Properties;
+  using FstImpl<Arc>::SetInputSymbols;
+  using FstImpl<Arc>::SetOutputSymbols;
+
+  using CacheBaseImpl<CacheState<Arc>>::PushArc;
+  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
+  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
+  using CacheBaseImpl<CacheState<Arc>>::HasStart;
+  using CacheBaseImpl<CacheState<Arc>>::SetArcs;
+  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
+  using CacheBaseImpl<CacheState<Arc>>::SetStart;
+
+  PdtExpandFstImpl(const Fst<Arc> &fst,
+                   const std::vector<std::pair<Label, Label>> &parens,
+                   const PdtExpandFstOptions<Arc> &opts)
+      : CacheImpl<Arc>(opts),
+        fst_(fst.Copy()),
+        stack_(opts.stack ? opts.stack : new PdtStack<StateId, Label>(parens)),
+        state_table_(opts.state_table ? opts.state_table
+                                      : new PdtStateTable<StateId, StackId>()),
+        own_stack_(opts.stack == 0),
+        own_state_table_(opts.state_table == 0),
+        keep_parentheses_(opts.keep_parentheses) {
+    SetType("expand");
+    const auto props = fst.Properties(kFstProperties, false);
+    SetProperties(PdtExpandProperties(props), kCopyProperties);
+    SetInputSymbols(fst.InputSymbols());
+    SetOutputSymbols(fst.OutputSymbols());
+  }
+
+  PdtExpandFstImpl(const PdtExpandFstImpl &impl)
+      : CacheImpl<Arc>(impl),
+        fst_(impl.fst_->Copy(true)),
+        stack_(new PdtStack<StateId, Label>(*impl.stack_)),
+        state_table_(new PdtStateTable<StateId, StackId>()),
+        own_stack_(true),
+        own_state_table_(true),
+        keep_parentheses_(impl.keep_parentheses_) {
+    SetType("expand");
+    SetProperties(impl.Properties(), kCopyProperties);
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+  }
+
+  ~PdtExpandFstImpl() override {
+    if (own_stack_) delete stack_;
+    if (own_state_table_) delete state_table_;
+  }
+
+  StateId Start() {
+    if (!HasStart()) {
+      const auto s = fst_->Start();
+      if (s == kNoStateId) return kNoStateId;
+      StateTuple tuple(s, 0);
+      const auto start = state_table_->FindState(tuple);
+      SetStart(start);
+    }
+    return CacheImpl<Arc>::Start();
+  }
+
+  Weight Final(StateId s) {
+    if (!HasFinal(s)) {
+      const auto &tuple = state_table_->Tuple(s);
+      const auto weight = fst_->Final(tuple.state_id);
+      if (weight != Weight::Zero() && tuple.stack_id == 0)
+        SetFinal(s, weight);
+      else
+        SetFinal(s, Weight::Zero());
+    }
+    return CacheImpl<Arc>::Final(s);
+  }
+
+  size_t NumArcs(StateId s) {
+    if (!HasArcs(s)) ExpandState(s);
+    return CacheImpl<Arc>::NumArcs(s);
+  }
+
+  size_t NumInputEpsilons(StateId s) {
+    if (!HasArcs(s)) ExpandState(s);
+    return CacheImpl<Arc>::NumInputEpsilons(s);
+  }
+
+  size_t NumOutputEpsilons(StateId s) {
+    if (!HasArcs(s)) ExpandState(s);
+    return CacheImpl<Arc>::NumOutputEpsilons(s);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
+    if (!HasArcs(s)) ExpandState(s);
+    CacheImpl<Arc>::InitArcIterator(s, data);
+  }
+
+  // Computes the outgoing transitions from a state, creating new destination
+  // states as needed.
+  void ExpandState(StateId s) {
+    StateTuple tuple = state_table_->Tuple(s);
+    for (ArcIterator<Fst<Arc>> aiter(*fst_, tuple.state_id); !aiter.Done();
+         aiter.Next()) {
+      auto arc = aiter.Value();
+      const auto stack_id = stack_->Find(tuple.stack_id, arc.ilabel);
+      if (stack_id == -1) {  // Non-matching close parenthesis.
+        continue;
+      } else if ((stack_id != tuple.stack_id) && !keep_parentheses_) {
+        // Stack push/pop.
+        arc.ilabel = 0;
+        arc.olabel = 0;
+      }
+      StateTuple ntuple(arc.nextstate, stack_id);
+      arc.nextstate = state_table_->FindState(ntuple);
+      PushArc(s, arc);
+    }
+    SetArcs(s);
+  }
+
+  const PdtStack<StackId, Label> &GetStack() const { return *stack_; }
+
+  const PdtStateTable<StateId, StackId> &GetStateTable() const {
+    return *state_table_;
+  }
+
+ private:
+  // Properties for an expanded PDT.
+  inline uint64 PdtExpandProperties(uint64 inprops) {
+    return inprops & (kAcceptor | kAcyclic | kInitialAcyclic | kUnweighted);
+  }
+
+  std::unique_ptr<const Fst<Arc>> fst_;
+  PdtStack<StackId, Label> *stack_;
+  PdtStateTable<StateId, StackId> *state_table_;
+  bool own_stack_;
+  bool own_state_table_;
+  bool keep_parentheses_;
+};
+
+}  // namespace internal
+
+// Expands a pushdown transducer (PDT) encoded as an FST into an FST. This
+// version is a delayed FST. In the PDT, some transitions are labeled with open
+// or close parentheses. To be interpreted as a PDT, the parens must balance on
+// a path. The open-close parenthesis label pairs are passed using the parens
+// argument. The expansion enforces the parenthesis constraints. The PDT must be
+// expandable as an FST.
+//
+// This class attaches interface to implementation and handles reference
+// counting, delegating most methods to ImplToFst.
+template <class A>
+class PdtExpandFst : public ImplToFst<internal::PdtExpandFstImpl<A>> {
+ public:
+  using Arc = A;
+
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using StackId = StateId;
+  using Store = DefaultCacheStore<Arc>;
+  using State = typename Store::State;
+  using Impl = internal::PdtExpandFstImpl<Arc>;
+
+  friend class ArcIterator<PdtExpandFst<Arc>>;
+  friend class StateIterator<PdtExpandFst<Arc>>;
+
+  PdtExpandFst(const Fst<Arc> &fst,
+               const std::vector<std::pair<Label, Label>> &parens)
+      : ImplToFst<Impl>(
+            std::make_shared<Impl>(fst, parens, PdtExpandFstOptions<A>())) {}
+
+  PdtExpandFst(const Fst<Arc> &fst,
+               const std::vector<std::pair<Label, Label>> &parens,
+               const PdtExpandFstOptions<Arc> &opts)
+      : ImplToFst<Impl>(std::make_shared<Impl>(fst, parens, opts)) {}
+
+  // See Fst<>::Copy() for doc.
+  PdtExpandFst(const PdtExpandFst<Arc> &fst, bool safe = false)
+      : ImplToFst<Impl>(fst, safe) {}
+
+  // Gets a copy of this ExpandFst. See Fst<>::Copy() for further doc.
+  PdtExpandFst<Arc> *Copy(bool safe = false) const override {
+    return new PdtExpandFst<Arc>(*this, safe);
+  }
+
+  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
+    GetMutableImpl()->InitArcIterator(s, data);
+  }
+
+  const PdtStack<StackId, Label> &GetStack() const {
+    return GetImpl()->GetStack();
+  }
+
+  const PdtStateTable<StateId, StackId> &GetStateTable() const {
+    return GetImpl()->GetStateTable();
+  }
+
+ private:
+  using ImplToFst<Impl>::GetImpl;
+  using ImplToFst<Impl>::GetMutableImpl;
+
+  void operator=(const PdtExpandFst &) = delete;
+};
+
+// Specialization for PdtExpandFst.
+template <class Arc>
+class StateIterator<PdtExpandFst<Arc>>
+    : public CacheStateIterator<PdtExpandFst<Arc>> {
+ public:
+  explicit StateIterator(const PdtExpandFst<Arc> &fst)
+      : CacheStateIterator<PdtExpandFst<Arc>>(fst, fst.GetMutableImpl()) {}
+};
+
+// Specialization for PdtExpandFst.
+template <class Arc>
+class ArcIterator<PdtExpandFst<Arc>>
+    : public CacheArcIterator<PdtExpandFst<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const PdtExpandFst<Arc> &fst, StateId s)
+      : CacheArcIterator<PdtExpandFst<Arc>>(fst.GetMutableImpl(), s) {
+    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->ExpandState(s);
+  }
+};
+
+template <class Arc>
+inline void PdtExpandFst<Arc>::InitStateIterator(
+    StateIteratorData<Arc> *data) const {
+  data->base = new StateIterator<PdtExpandFst<Arc>>(*this);
+}
+
+// PrunedExpand prunes the delayed expansion of a pushdown transducer (PDT)
+// encoded as an FST into an FST. In the PDT, some transitions are labeled with
+// open or close parentheses. To be interpreted as a PDT, the parens must
+// balance on a path. The open-close parenthesis label pairs are passed
+// using the parens argument. The expansion enforces the parenthesis
+// constraints.
+//
+// The algorithm works by visiting the delayed ExpandFst using a shortest-stack
+// first queue discipline and relies on the shortest-distance information
+// computed using a reverse shortest-path call to perform the pruning.
+//
+// The algorithm maintains the same state ordering between the ExpandFst being
+// visited (efst_) and the result of pruning written into the MutableFst (ofst_)
+// to improve readability.
+template <class Arc>
+class PdtPrunedExpand {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using StackId = StateId;
+  using Stack = PdtStack<StackId, Label>;
+  using StateTable = PdtStateTable<StateId, StackId>;
+  using SetIterator = typename internal::PdtBalanceData<Arc>::SetIterator;
+
+  // Constructor taking as input a PDT specified by by an input FST and a vector
+  // of parentheses. The keep_parentheses argument specifies whether parentheses
+  // are replaced by epsilons or not during the expansion. The cache options are
+  // passed to the underlying ExpandFst.
+  PdtPrunedExpand(const Fst<Arc> &ifst,
+                  const std::vector<std::pair<Label, Label>> &parens,
+                  bool keep_parentheses = false,
+                  const CacheOptions &opts = CacheOptions())
+      : ifst_(ifst.Copy()),
+        keep_parentheses_(keep_parentheses),
+        stack_(parens),
+        efst_(ifst, parens,
+              PdtExpandFstOptions<Arc>(opts, true, &stack_, &state_table_)),
+        queue_(state_table_, stack_, stack_length_, distance_, fdistance_),
+        error_(false) {
+    Reverse(*ifst_, parens, &rfst_);
+    VectorFst<Arc> path;
+    reverse_shortest_path_.reset(new PdtShortestPath<Arc, FifoQueue<StateId>>(
+        rfst_, parens,
+        PdtShortestPathOptions<Arc, FifoQueue<StateId>>(true, false)));
+    reverse_shortest_path_->ShortestPath(&path);
+    error_ = (path.Properties(kError, true) == kError);
+    balance_data_.reset(reverse_shortest_path_->GetBalanceData()->Reverse(
+        rfst_.NumStates(), 10, -1));
+    InitCloseParenMultimap(parens);
+  }
+
+  bool Error() const { return error_; }
+
+  // Expands and prunes the input PDT according to the provided weight
+  // threshold, wirting the result into an output mutable FST.
+  void Expand(MutableFst<Arc> *ofst, const Weight &threshold);
+
+ private:
+  static constexpr uint8 kEnqueued = 0x01;
+  static constexpr uint8 kExpanded = 0x02;
+  static constexpr uint8 kSourceState = 0x04;
+
+  // Comparison functor used by the queue:
+  //
+  // 1. States corresponding to shortest stack first, and
+  // 2. for stacks of matching length, reverse lexicographic order is used, and
+  // 3. for states with the same stack, shortest-first order is used.
+  class StackCompare {
+   public:
+    StackCompare(const StateTable &state_table, const Stack &stack,
+                 const std::vector<StackId> &stack_length,
+                 const std::vector<Weight> &distance,
+                 const std::vector<Weight> &fdistance)
+        : state_table_(state_table),
+          stack_(stack),
+          stack_length_(stack_length),
+          distance_(distance),
+          fdistance_(fdistance) {}
+
+    bool operator()(StateId s1, StateId s2) const {
+      auto si1 = state_table_.Tuple(s1).stack_id;
+      auto si2 = state_table_.Tuple(s2).stack_id;
+      if (stack_length_[si1] < stack_length_[si2]) return true;
+      if (stack_length_[si1] > stack_length_[si2]) return false;
+      // If stack IDs are equal, use A*.
+      if (si1 == si2) {
+        return less_(Distance(s1), Distance(s2));
+      }
+      // If lengths are equal, uses reverse lexicographic order.
+      for (; si1 != si2; si1 = stack_.Pop(si1), si2 = stack_.Pop(si2)) {
+        if (stack_.Top(si1) < stack_.Top(si2)) return true;
+        if (stack_.Top(si1) > stack_.Top(si2)) return false;
+      }
+      return false;
+    }
+
+   private:
+    Weight Distance(StateId s) const {
+      return (s < distance_.size()) && (s < fdistance_.size())
+                 ? Times(distance_[s], fdistance_[s])
+                 : Weight::Zero();
+    }
+
+    const StateTable &state_table_;
+    const Stack &stack_;
+    const std::vector<StackId> &stack_length_;
+    const std::vector<Weight> &distance_;
+    const std::vector<Weight> &fdistance_;
+    const NaturalLess<Weight> less_;
+  };
+
+  class ShortestStackFirstQueue
+      : public ShortestFirstQueue<StateId, StackCompare> {
+   public:
+    ShortestStackFirstQueue(const PdtStateTable<StateId, StackId> &state_table,
+                            const Stack &stack,
+                            const std::vector<StackId> &stack_length,
+                            const std::vector<Weight> &distance,
+                            const std::vector<Weight> &fdistance)
+        : ShortestFirstQueue<StateId, StackCompare>(StackCompare(
+              state_table, stack, stack_length, distance, fdistance)) {}
+  };
+
+  void InitCloseParenMultimap(
+      const std::vector<std::pair<Label, Label>> &parens);
+
+  Weight DistanceToDest(StateId source, StateId dest) const;
+
+  uint8 Flags(StateId s) const;
+
+  void SetFlags(StateId s, uint8 flags, uint8 mask);
+
+  Weight Distance(StateId s) const;
+
+  void SetDistance(StateId s, Weight weight);
+
+  Weight FinalDistance(StateId s) const;
+
+  void SetFinalDistance(StateId s, Weight weight);
+
+  StateId SourceState(StateId s) const;
+
+  void SetSourceState(StateId s, StateId p);
+
+  void AddStateAndEnqueue(StateId s);
+
+  void Relax(StateId s, const Arc &arc, Weight weight);
+
+  bool PruneArc(StateId s, const Arc &arc);
+
+  void ProcStart();
+
+  void ProcFinal(StateId s);
+
+  bool ProcNonParen(StateId s, const Arc &arc, bool add_arc);
+
+  bool ProcOpenParen(StateId s, const Arc &arc, StackId si, StackId nsi);
+
+  bool ProcCloseParen(StateId s, const Arc &arc);
+
+  void ProcDestStates(StateId s, StackId si);
+
+  // Input PDT.
+  std::unique_ptr<Fst<Arc>> ifst_;
+  // Reversed PDT.
+  VectorFst<Arc> rfst_;
+  // Keep parentheses in ofst?
+  const bool keep_parentheses_;
+  // State table for efst_.
+  StateTable state_table_;
+  // Stack trie.
+  Stack stack_;
+  // Expanded PDT.
+  PdtExpandFst<Arc> efst_;
+  // Length of stack for given stack ID.
+  std::vector<StackId> stack_length_;
+  // Distance from initial state in efst_/ofst.
+  std::vector<Weight> distance_;
+  // Distance to final states in efst_/ofst.
+  std::vector<Weight> fdistance_;
+  // Queue used to visit efst_.
+  ShortestStackFirstQueue queue_;
+  // Construction time failure?
+  bool error_;
+  // Status flags for states in efst_/ofst.
+  std::vector<uint8> flags_;
+  // PDT source state for each expanded state.
+  std::vector<StateId> sources_;
+  // Shortest path for rfst_.
+  std::unique_ptr<PdtShortestPath<Arc, FifoQueue<StateId>>>
+      reverse_shortest_path_;
+  std::unique_ptr<internal::PdtBalanceData<Arc>> balance_data_;
+  // Maps open paren arcs to balancing close paren arcs.
+  typename PdtShortestPath<Arc, FifoQueue<StateId>>::CloseParenMultimap
+      close_paren_multimap_;
+  MutableFst<Arc> *ofst_;  // Output FST.
+  Weight limit_;           // Weight limit.
+
+  // Maps a state s in ifst (i.e., the source of a closed paranthesis matching
+  // the top of current_stack_id_ to final states in efst_.
+  std::unordered_map<StateId, Weight> dest_map_;
+  // Stack ID of the states currently at the top of the queue, i.e., the states
+  // currently being popped and processed.
+  StackId current_stack_id_;
+  ssize_t current_paren_id_;  // Paren ID at top of current stack.
+  ssize_t cached_stack_id_;
+  StateId cached_source_;
+  // The set of pairs of destination states and weights to final states for the
+  // source state cached_source_ and the paren ID cached_paren_id_; i.e., the
+  // set of source states of a closed parenthesis with paren ID cached_paren_id
+  // balancing an incoming open parenthesis with paren ID cached_paren_id_ in
+  // state cached_source_.
+  std::forward_list<std::pair<StateId, Weight>> cached_dest_list_;
+  NaturalLess<Weight> less_;
+};
+
+// Initializes close paren multimap, mapping pairs (s, paren_id) to all the arcs
+// out of s labeled with close parenthese for paren_id.
+template <class Arc>
+void PdtPrunedExpand<Arc>::InitCloseParenMultimap(
+    const std::vector<std::pair<Label, Label>> &parens) {
+  std::unordered_map<Label, Label> paren_map;
+  for (size_t i = 0; i < parens.size(); ++i) {
+    const auto &pair = parens[i];
+    paren_map[pair.first] = i;
+    paren_map[pair.second] = i;
+  }
+  for (StateIterator<Fst<Arc>> siter(*ifst_); !siter.Done(); siter.Next()) {
+    const auto s = siter.Value();
+    for (ArcIterator<Fst<Arc>> aiter(*ifst_, s); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      const auto it = paren_map.find(arc.ilabel);
+      if (it == paren_map.end()) continue;
+      if (arc.ilabel == parens[it->second].second) {  // Close paren.
+        const internal::ParenState<Arc> key(it->second, s);
+        close_paren_multimap_.emplace(key, arc);
+      }
+    }
+  }
+}
+
+// Returns the weight of the shortest balanced path from source to dest
+// in ifst_; dest must be the source state of a close paren arc.
+template <class Arc>
+typename Arc::Weight PdtPrunedExpand<Arc>::DistanceToDest(StateId source,
+                                                          StateId dest) const {
+  using SearchState =
+      typename PdtShortestPath<Arc, FifoQueue<StateId>>::SearchState;
+  const SearchState ss(source + 1, dest + 1);
+  const auto distance =
+      reverse_shortest_path_->GetShortestPathData().Distance(ss);
+  VLOG(2) << "D(" << source << ", " << dest << ") =" << distance;
+  return distance;
+}
+
+// Returns the flags for state s in ofst_.
+template <class Arc>
+uint8 PdtPrunedExpand<Arc>::Flags(StateId s) const {
+  return s < flags_.size() ? flags_[s] : 0;
+}
+
+// Modifies the flags for state s in ofst_.
+template <class Arc>
+void PdtPrunedExpand<Arc>::SetFlags(StateId s, uint8 flags, uint8 mask) {
+  while (flags_.size() <= s) flags_.push_back(0);
+  flags_[s] &= ~mask;
+  flags_[s] |= flags & mask;
+}
+
+// Returns the shortest distance from the initial state to s in ofst_.
+template <class Arc>
+typename Arc::Weight PdtPrunedExpand<Arc>::Distance(StateId s) const {
+  return s < distance_.size() ? distance_[s] : Weight::Zero();
+}
+
+// Sets the shortest distance from the initial state to s in ofst_.
+template <class Arc>
+void PdtPrunedExpand<Arc>::SetDistance(StateId s, Weight weight) {
+  while (distance_.size() <= s) distance_.push_back(Weight::Zero());
+  distance_[s] = std::move(weight);
+}
+
+// Returns the shortest distance from s to the final states in ofst_.
+template <class Arc>
+typename Arc::Weight PdtPrunedExpand<Arc>::FinalDistance(StateId s) const {
+  return s < fdistance_.size() ? fdistance_[s] : Weight::Zero();
+}
+
+// Sets the shortest distance from s to the final states in ofst_.
+template <class Arc>
+void PdtPrunedExpand<Arc>::SetFinalDistance(StateId s, Weight weight) {
+  while (fdistance_.size() <= s) fdistance_.push_back(Weight::Zero());
+  fdistance_[s] = std::move(weight);
+}
+
+// Returns the PDT source state of state s in ofst_.
+template <class Arc>
+typename Arc::StateId PdtPrunedExpand<Arc>::SourceState(StateId s) const {
+  return s < sources_.size() ? sources_[s] : kNoStateId;
+}
+
+// Sets the PDT source state of state s in ofst_ to state p'in ifst_.
+template <class Arc>
+void PdtPrunedExpand<Arc>::SetSourceState(StateId s, StateId p) {
+  while (sources_.size() <= s) sources_.push_back(kNoStateId);
+  sources_[s] = p;
+}
+
+// Adds state s of efst_ to ofst_ and inserts it in the queue, modifying the
+// flags for s accordingly.
+template <class Arc>
+void PdtPrunedExpand<Arc>::AddStateAndEnqueue(StateId s) {
+  if (!(Flags(s) & (kEnqueued | kExpanded))) {
+    while (ofst_->NumStates() <= s) ofst_->AddState();
+    queue_.Enqueue(s);
+    SetFlags(s, kEnqueued, kEnqueued);
+  } else if (Flags(s) & kEnqueued) {
+    queue_.Update(s);
+  }
+  // TODO(allauzen): Check everything is fine when kExpanded?
+}
+
+// Relaxes arc out of state s in ofst_ as follows:
+//
+// 1. If the distance to s times the weight of arc is smaller than
+//   the currently stored distance for arc.nextstate, updates
+//   Distance(arc.nextstate) with a new estimate
+// 2. If fd is less than the currently stored distance from arc.nextstate to the
+// final state, updates with new estimate.
+template <class Arc>
+void PdtPrunedExpand<Arc>::Relax(StateId s, const Arc &arc, Weight fd) {
+  const auto nd = Times(Distance(s), arc.weight);
+  if (less_(nd, Distance(arc.nextstate))) {
+    SetDistance(arc.nextstate, nd);
+    SetSourceState(arc.nextstate, SourceState(s));
+  }
+  if (less_(fd, FinalDistance(arc.nextstate))) {
+    SetFinalDistance(arc.nextstate, fd);
+  }
+  VLOG(2) << "Relax: " << s << ", d[s] = " << Distance(s) << ", to "
+          << arc.nextstate << ", d[ns] = " << Distance(arc.nextstate)
+          << ", nd = " << nd;
+}
+
+// Returns whether the arc out of state s in efst needs pruned.
+template <class Arc>
+bool PdtPrunedExpand<Arc>::PruneArc(StateId s, const Arc &arc) {
+  VLOG(2) << "Prune ?";
+  auto fd = Weight::Zero();
+  if ((cached_source_ != SourceState(s)) ||
+      (cached_stack_id_ != current_stack_id_)) {
+    cached_source_ = SourceState(s);
+    cached_stack_id_ = current_stack_id_;
+    cached_dest_list_.clear();
+    if (cached_source_ != ifst_->Start()) {
+      for (auto set_iter =
+               balance_data_->Find(current_paren_id_, cached_source_);
+           !set_iter.Done(); set_iter.Next()) {
+        auto dest = set_iter.Element();
+        const auto it = dest_map_.find(dest);
+        cached_dest_list_.push_front(*it);
+      }
+    } else {
+      // TODO(allauzen): queue discipline should prevent this from ever
+      // happening.
+      // Replace by a check.
+      cached_dest_list_.push_front(
+          std::make_pair(rfst_.Start() - 1, Weight::One()));
+    }
+  }
+  for (auto it = cached_dest_list_.begin(); it != cached_dest_list_.end();
+       ++it) {
+    const auto d =
+        DistanceToDest(state_table_.Tuple(arc.nextstate).state_id, it->first);
+    fd = Plus(fd, Times(d, it->second));
+  }
+  Relax(s, arc, fd);
+  return less_(limit_, Times(Distance(s), Times(arc.weight, fd)));
+}
+
+// Adds start state of efst_ to ofst_, enqueues it, and initializes the distance
+// data structures.
+template <class Arc>
+void PdtPrunedExpand<Arc>::ProcStart() {
+  const auto s = efst_.Start();
+  AddStateAndEnqueue(s);
+  ofst_->SetStart(s);
+  SetSourceState(s, ifst_->Start());
+  current_stack_id_ = 0;
+  current_paren_id_ = -1;
+  stack_length_.push_back(0);
+  const auto r = rfst_.Start() - 1;
+  cached_source_ = ifst_->Start();
+  cached_stack_id_ = 0;
+  cached_dest_list_.push_front(std::make_pair(r, Weight::One()));
+  const PdtStateTuple<StateId, StackId> tuple(r, 0);
+  SetFinalDistance(state_table_.FindState(tuple), Weight::One());
+  SetDistance(s, Weight::One());
+  const auto d = DistanceToDest(ifst_->Start(), r);
+  SetFinalDistance(s, d);
+  VLOG(2) << d;
+}
+
+// Makes s final in ofst_ if shortest accepting path ending in s is below
+// threshold.
+template <class Arc>
+void PdtPrunedExpand<Arc>::ProcFinal(StateId s) {
+  const auto weight = efst_.Final(s);
+  if (weight == Weight::Zero()) return;
+  if (less_(limit_, Times(Distance(s), weight))) return;
+  ofst_->SetFinal(s, weight);
+}
+
+// Returns true when an arc (or meta-arc) leaving state s in efst_ is below the
+// threshold. When add_arc is true, arc is added to ofst_.
+template <class Arc>
+bool PdtPrunedExpand<Arc>::ProcNonParen(StateId s, const Arc &arc,
+                                        bool add_arc) {
+  VLOG(2) << "ProcNonParen: " << s << " to " << arc.nextstate << ", "
+          << arc.ilabel << ":" << arc.olabel << " / " << arc.weight
+          << ", add_arc = " << (add_arc ? "true" : "false");
+  if (PruneArc(s, arc)) return false;
+  if (add_arc) ofst_->AddArc(s, arc);
+  AddStateAndEnqueue(arc.nextstate);
+  return true;
+}
+
+// Processes an open paren arc leaving state s in ofst_. When the arc is labeled
+// with an open paren,
+//
+// 1. Considers each (shortest) balanced path starting in s by taking the arc
+// and ending by a close paren balancing the open paren of as a meta-arc,
+// processing and pruning each meta-arc as a non-paren arc, inserting its
+// destination to the queue;
+// 2. if at least one of these meta-arcs has not been pruned, adds the
+// destination of arc to ofst_ as a new source state for the stack ID nsi, and
+// inserts it in the queue.
+template <class Arc>
+bool PdtPrunedExpand<Arc>::ProcOpenParen(StateId s, const Arc &arc, StackId si,
+                                         StackId nsi) {
+  // Updates the stack length when needed.
+  while (stack_length_.size() <= nsi) stack_length_.push_back(-1);
+  if (stack_length_[nsi] == -1) stack_length_[nsi] = stack_length_[si] + 1;
+  const auto ns = arc.nextstate;
+  VLOG(2) << "Open paren: " << s << "(" << state_table_.Tuple(s).state_id
+          << ") to " << ns << "(" << state_table_.Tuple(ns).state_id << ")";
+  bool proc_arc = false;
+  auto fd = Weight::Zero();
+  const auto paren_id = stack_.ParenId(arc.ilabel);
+  std::forward_list<StateId> sources;
+  for (auto set_iter =
+           balance_data_->Find(paren_id, state_table_.Tuple(ns).state_id);
+       !set_iter.Done(); set_iter.Next()) {
+    sources.push_front(set_iter.Element());
+  }
+  for (const auto source : sources) {
+    VLOG(2) << "Close paren source: " << source;
+    const internal::ParenState<Arc> paren_state(paren_id, source);
+    for (auto it = close_paren_multimap_.find(paren_state);
+         it != close_paren_multimap_.end() && paren_state == it->first; ++it) {
+      auto meta_arc = it->second;
+      const PdtStateTuple<StateId, StackId> tuple(meta_arc.nextstate, si);
+      meta_arc.nextstate = state_table_.FindState(tuple);
+      const auto state_id = state_table_.Tuple(ns).state_id;
+      const auto d = DistanceToDest(state_id, source);
+      VLOG(2) << state_id << ", " << source;
+      VLOG(2) << "Meta arc weight = " << arc.weight << " Times " << d
+              << " Times " << meta_arc.weight;
+      meta_arc.weight = Times(arc.weight, Times(d, meta_arc.weight));
+      proc_arc |= ProcNonParen(s, meta_arc, false);
+      fd = Plus(
+          fd,
+          Times(Times(DistanceToDest(state_table_.Tuple(ns).state_id, source),
+                      it->second.weight),
+                FinalDistance(meta_arc.nextstate)));
+    }
+  }
+  if (proc_arc) {
+    VLOG(2) << "Proc open paren " << s << " to " << arc.nextstate;
+    ofst_->AddArc(
+        s, keep_parentheses_ ? arc : Arc(0, 0, arc.weight, arc.nextstate));
+    AddStateAndEnqueue(arc.nextstate);
+    const auto nd = Times(Distance(s), arc.weight);
+    if (less_(nd, Distance(arc.nextstate))) SetDistance(arc.nextstate, nd);
+    // FinalDistance not necessary for source state since pruning decided using
+    // meta-arcs above.  But this is a problem with A*, hence the following.
+    if (less_(fd, FinalDistance(arc.nextstate)))
+      SetFinalDistance(arc.nextstate, fd);
+    SetFlags(arc.nextstate, kSourceState, kSourceState);
+  }
+  return proc_arc;
+}
+
+// Checks that shortest path through close paren arc in efst_ is below
+// threshold, and if so, adds it to ofst_.
+template <class Arc>
+bool PdtPrunedExpand<Arc>::ProcCloseParen(StateId s, const Arc &arc) {
+  const auto weight =
+      Times(Distance(s), Times(arc.weight, FinalDistance(arc.nextstate)));
+  if (less_(limit_, weight)) return false;
+  ofst_->AddArc(s,
+                keep_parentheses_ ? arc : Arc(0, 0, arc.weight, arc.nextstate));
+  return true;
+}
+
+// When state s in ofst_ is a source state for stack ID si, identifies all the
+// corresponding possible destination states, that is, all the states in ifst_
+// that have an outgoing close paren arc balancing the incoming open paren taken
+// to get to s. For each such state t, computes the shortest distance from (t,
+// si) to the final states in ofst_. Stores this information in dest_map_.
+template <class Arc>
+void PdtPrunedExpand<Arc>::ProcDestStates(StateId s, StackId si) {
+  if (!(Flags(s) & kSourceState)) return;
+  if (si != current_stack_id_) {
+    dest_map_.clear();
+    current_stack_id_ = si;
+    current_paren_id_ = stack_.Top(current_stack_id_);
+    VLOG(2) << "StackID " << si << " dequeued for first time";
+  }
+  // TODO(allauzen): clean up source state business; rename current function to
+  // ProcSourceState.
+  SetSourceState(s, state_table_.Tuple(s).state_id);
+  const auto paren_id = stack_.Top(si);
+  for (auto set_iter =
+           balance_data_->Find(paren_id, state_table_.Tuple(s).state_id);
+       !set_iter.Done(); set_iter.Next()) {
+    const auto dest_state = set_iter.Element();
+    if (dest_map_.find(dest_state) != dest_map_.end()) continue;
+    auto dest_weight = Weight::Zero();
+    internal::ParenState<Arc> paren_state(paren_id, dest_state);
+    for (auto it = close_paren_multimap_.find(paren_state);
+         it != close_paren_multimap_.end() && paren_state == it->first; ++it) {
+      const auto &arc = it->second;
+      const PdtStateTuple<StateId, StackId> tuple(arc.nextstate,
+                                                  stack_.Pop(si));
+      dest_weight =
+          Plus(dest_weight,
+               Times(arc.weight, FinalDistance(state_table_.FindState(tuple))));
+    }
+    dest_map_[dest_state] = dest_weight;
+    VLOG(2) << "State " << dest_state << " is a dest state for stack ID " << si
+            << " with weight " << dest_weight;
+  }
+}
+
+// Expands and prunes the input PDT, writing the result in ofst.
+template <class Arc>
+void PdtPrunedExpand<Arc>::Expand(MutableFst<Arc> *ofst,
+                                  const typename Arc::Weight &threshold) {
+  ofst_ = ofst;
+  if (error_) {
+    ofst_->SetProperties(kError, kError);
+    return;
+  }
+  ofst_->DeleteStates();
+  ofst_->SetInputSymbols(ifst_->InputSymbols());
+  ofst_->SetOutputSymbols(ifst_->OutputSymbols());
+  limit_ = Times(DistanceToDest(ifst_->Start(), rfst_.Start() - 1), threshold);
+  flags_.clear();
+  ProcStart();
+  while (!queue_.Empty()) {
+    const auto s = queue_.Head();
+    queue_.Dequeue();
+    SetFlags(s, kExpanded, kExpanded | kEnqueued);
+    VLOG(2) << s << " dequeued!";
+    ProcFinal(s);
+    StackId stack_id = state_table_.Tuple(s).stack_id;
+    ProcDestStates(s, stack_id);
+    for (ArcIterator<PdtExpandFst<Arc>> aiter(efst_, s); !aiter.Done();
+         aiter.Next()) {
+      const auto &arc = aiter.Value();
+      const auto nextstack_id = state_table_.Tuple(arc.nextstate).stack_id;
+      if (stack_id == nextstack_id) {
+        ProcNonParen(s, arc, true);
+      } else if (stack_id == stack_.Pop(nextstack_id)) {
+        ProcOpenParen(s, arc, stack_id, nextstack_id);
+      } else {
+        ProcCloseParen(s, arc);
+      }
+    }
+    VLOG(2) << "d[" << s << "] = " << Distance(s) << ", fd[" << s
+            << "] = " << FinalDistance(s);
+  }
+}
+
+// Expand functions.
+
+template <class Arc>
+struct PdtExpandOptions {
+  using Weight = typename Arc::Weight;
+
+  bool connect;
+  bool keep_parentheses;
+  Weight weight_threshold;
+
+  PdtExpandOptions(bool connect = true, bool keep_parentheses = false,
+                   Weight weight_threshold = Weight::Zero())
+      : connect(connect),
+        keep_parentheses(keep_parentheses),
+        weight_threshold(std::move(weight_threshold)) {}
+};
+
+// Expands a pushdown transducer (PDT) encoded as an FST into an FST. This
+// version writes the expanded PDT to a mutable FST. In the PDT, some
+// transitions are labeled with open or close parentheses. To be interpreted as
+// a PDT, the parens must balance on a path. The open-close parenthesis label
+// pairs are passed using the parens argument. Expansion enforces the
+// parenthesis constraints. The PDT must be expandable as an FST.
+template <class Arc>
+void Expand(
+    const Fst<Arc> &ifst,
+    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
+        &parens,
+    MutableFst<Arc> *ofst, const PdtExpandOptions<Arc> &opts) {
+  PdtExpandFstOptions<Arc> eopts;
+  eopts.gc_limit = 0;
+  if (opts.weight_threshold == Arc::Weight::Zero()) {
+    eopts.keep_parentheses = opts.keep_parentheses;
+    *ofst = PdtExpandFst<Arc>(ifst, parens, eopts);
+  } else {
+    PdtPrunedExpand<Arc> pruned_expand(ifst, parens, opts.keep_parentheses);
+    pruned_expand.Expand(ofst, opts.weight_threshold);
+  }
+  if (opts.connect) Connect(ofst);
+}
+
+// Expands a pushdown transducer (PDT) encoded as an FST into an FST. This
+// version writes the expanded PDT result to a mutable FST. In the PDT, some
+// transitions are labeled with open or close parentheses. To be interpreted as
+// a PDT, the parens must balance on a path. The open-close parenthesis label
+// pairs are passed using the parents argument. Expansion enforces the
+// parenthesis constraints. The PDT must be expandable as an FST.
+template <class Arc>
+void Expand(const Fst<Arc> &ifst,
+    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
+    &parens, MutableFst<Arc> *ofst, bool connect = true,
+    bool keep_parentheses = false) {
+  const PdtExpandOptions<Arc> opts(connect, keep_parentheses);
+  Expand(ifst, parens, ofst, opts);
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_PDT_EXPAND_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/getters.h b/projects/llm_framework/include/fst/extensions/pdt/getters.h
new file mode 100644
index 00000000..69dd150d
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/pdt/getters.h
@@ -0,0 +1,22 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_EXTENSIONS_PDT_GETTERS_H_
+#define FST_EXTENSIONS_PDT_GETTERS_H_
+
+#include <string>
+
+#include <fst/extensions/pdt/compose.h>
+#include <fst/extensions/pdt/replace.h>
+
+namespace fst {
+namespace script {
+
+bool GetPdtComposeFilter(const string &str, PdtComposeFilter *cf);
+
+bool GetPdtParserType(const string &str, PdtParserType *pt);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_PDT_GETTERS_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/info.h b/projects/llm_framework/include/fst/extensions/pdt/info.h
new file mode 100644
index 00000000..3de54772
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/pdt/info.h
@@ -0,0 +1,152 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Prints information about a PDT.
+
+#ifndef FST_EXTENSIONS_PDT_INFO_H_
+#define FST_EXTENSIONS_PDT_INFO_H_
+
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+#include <fst/extensions/pdt/pdt.h>
+#include <fst/fst.h>
+
+namespace fst {
+
+// Compute various information about PDTs.
+template <class Arc>
+class PdtInfo {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  PdtInfo(const Fst<Arc> &fst,
+          const std::vector<std::pair<Label, Label>> &parents);
+
+  const string &FstType() const { return fst_type_; }
+
+  const string &ArcType() const { return Arc::Type(); }
+
+  int64 NumStates() const { return nstates_; }
+
+  int64 NumArcs() const { return narcs_; }
+
+  int64 NumOpenParens() const { return nopen_parens_; }
+
+  int64 NumCloseParens() const { return nclose_parens_; }
+
+  int64 NumUniqueOpenParens() const { return nuniq_open_parens_; }
+
+  int64 NumUniqueCloseParens() const { return nuniq_close_parens_; }
+
+  int64 NumOpenParenStates() const { return nopen_paren_states_; }
+
+  int64 NumCloseParenStates() const { return nclose_paren_states_; }
+
+ private:
+  string fst_type_;
+  int64 nstates_;
+  int64 narcs_;
+  int64 nopen_parens_;
+  int64 nclose_parens_;
+  int64 nuniq_open_parens_;
+  int64 nuniq_close_parens_;
+  int64 nopen_paren_states_;
+  int64 nclose_paren_states_;
+};
+
+template <class Arc>
+PdtInfo<Arc>::PdtInfo(
+    const Fst<Arc> &fst,
+    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
+        &parens)
+    : fst_type_(fst.Type()),
+      nstates_(0),
+      narcs_(0),
+      nopen_parens_(0),
+      nclose_parens_(0),
+      nuniq_open_parens_(0),
+      nuniq_close_parens_(0),
+      nopen_paren_states_(0),
+      nclose_paren_states_(0) {
+  std::unordered_map<Label, size_t> paren_map;
+  std::unordered_set<Label> paren_set;
+  std::unordered_set<StateId> open_paren_state_set;
+  std::unordered_set<StateId> close_paren_state_set;
+  for (size_t i = 0; i < parens.size(); ++i) {
+    const auto &pair = parens[i];
+    paren_map[pair.first] = i;
+    paren_map[pair.second] = i;
+  }
+  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
+    ++nstates_;
+    const auto s = siter.Value();
+    for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      ++narcs_;
+      const auto it = paren_map.find(arc.ilabel);
+      if (it != paren_map.end()) {
+        const auto open_paren = parens[it->second].first;
+        const auto close_paren = parens[it->second].second;
+        if (arc.ilabel == open_paren) {
+          ++nopen_parens_;
+          if (!paren_set.count(open_paren)) {
+            ++nuniq_open_parens_;
+            paren_set.insert(open_paren);
+          }
+          if (!open_paren_state_set.count(arc.nextstate)) {
+            ++nopen_paren_states_;
+            open_paren_state_set.insert(arc.nextstate);
+          }
+        } else {
+          ++nclose_parens_;
+          if (!paren_set.count(close_paren)) {
+            ++nuniq_close_parens_;
+            paren_set.insert(close_paren);
+          }
+          if (!close_paren_state_set.count(s)) {
+            ++nclose_paren_states_;
+            close_paren_state_set.insert(s);
+          }
+        }
+      }
+    }
+  }
+}
+
+template <class Arc>
+void PrintPdtInfo(const PdtInfo<Arc> &info) {
+  const auto old = std::cout.setf(std::ios::left);
+  std::cout.width(50);
+  std::cout << "fst type" << info.FstType() << std::endl;
+  std::cout.width(50);
+  std::cout << "arc type" << info.ArcType() << std::endl;
+  std::cout.width(50);
+  std::cout << "# of states" << info.NumStates() << std::endl;
+  std::cout.width(50);
+  std::cout << "# of arcs" << info.NumArcs() << std::endl;
+  std::cout.width(50);
+  std::cout << "# of open parentheses" << info.NumOpenParens() << std::endl;
+  std::cout.width(50);
+  std::cout << "# of close parentheses" << info.NumCloseParens() << std::endl;
+  std::cout.width(50);
+  std::cout << "# of unique open parentheses" << info.NumUniqueOpenParens()
+            << std::endl;
+  std::cout.width(50);
+  std::cout << "# of unique close parentheses" << info.NumUniqueCloseParens()
+            << std::endl;
+  std::cout.width(50);
+  std::cout << "# of open parenthesis dest. states" << info.NumOpenParenStates()
+            << std::endl;
+  std::cout.width(50);
+  std::cout << "# of close parenthesis source states"
+            << info.NumCloseParenStates() << std::endl;
+  std::cout.setf(old);
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_PDT_INFO_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/paren.h b/projects/llm_framework/include/fst/extensions/pdt/paren.h
new file mode 100644
index 00000000..0347848d
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/pdt/paren.h
@@ -0,0 +1,440 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Common classes for PDT parentheses.
+
+#ifndef FST_EXTENSIONS_PDT_PAREN_H_
+#define FST_EXTENSIONS_PDT_PAREN_H_
+
+#include <algorithm>
+#include <set>
+#include <unordered_map>
+#include <unordered_set>
+
+#include <fst/log.h>
+
+#include <fst/extensions/pdt/collection.h>
+#include <fst/extensions/pdt/pdt.h>
+#include <fst/dfs-visit.h>
+#include <fst/fst.h>
+
+
+namespace fst {
+namespace internal {
+
+// ParenState: Pair of an open (close) parenthesis and its destination (source)
+// state.
+
+template <class Arc>
+struct ParenState {
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+
+  Label paren_id;    // ID of open (close) paren.
+  StateId state_id;  // Destination (source) state of open (close) paren.
+
+  explicit ParenState(Label paren_id = kNoLabel, StateId state_id = kNoStateId)
+      : paren_id(paren_id), state_id(state_id) {}
+
+  bool operator==(const ParenState<Arc> &other) const {
+    if (&other == this) return true;
+    return other.paren_id == paren_id && other.state_id == state_id;
+  }
+
+  bool operator!=(const ParenState<Arc> &other) const {
+    return !(other == *this);
+  }
+
+  struct Hash {
+    size_t operator()(const ParenState<Arc> &pstate) const {
+      static constexpr auto prime = 7853;
+      return pstate.paren_id + pstate.state_id * prime;
+    }
+  };
+};
+
+// Creates an FST-style const iterator from an STL-style map.
+template <class Map>
+class MapIterator {
+ public:
+  using StlIterator = typename Map::const_iterator;
+  using ValueType = typename Map::mapped_type;
+
+  MapIterator(const Map &map, StlIterator it)
+      : begin_(it), end_(map.end()), it_(it) {}
+
+  bool Done() const { return it_ == end_ || it_->first != begin_->first; }
+
+  ValueType Value() const { return it_->second; }
+
+  void Next() { ++it_; }
+
+  void Reset() { it_ = begin_; }
+
+ private:
+  const StlIterator begin_;
+  const StlIterator end_;
+  StlIterator it_;
+};
+
+// PdtParenReachable: Provides various parenthesis reachability information.
+
+template <class Arc>
+class PdtParenReachable {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+
+  using State = ParenState<Arc>;
+  using StateHash = typename State::Hash;
+
+  // Maps from state ID to reachable paren IDs from (to) that state.
+  using ParenMultimap = std::unordered_multimap<StateId, Label>;
+
+  // Maps from paren ID and state ID to reachable state set ID.
+  using StateSetMap = std::unordered_map<State, ssize_t, StateHash>;
+
+  // Maps from paren ID and state ID to arcs exiting that state with that
+  // Label.
+  using ParenArcMultimap = std::unordered_map<State, Arc, StateHash>;
+
+  using ParenIterator = MapIterator<ParenMultimap>;
+
+  using ParenArcIterator = MapIterator<ParenArcMultimap>;
+
+  using SetIterator = typename Collection<ssize_t, StateId>::SetIterator;
+
+  // Computes close (open) parenthesis reachability information for a PDT with
+  // bounded stack.
+  PdtParenReachable(const Fst<Arc> &fst,
+                    const std::vector<std::pair<Label, Label>> &parens,
+                    bool close)
+      : fst_(fst), parens_(parens), close_(close), error_(false) {
+    paren_map_.reserve(2 * parens.size());
+    for (size_t i = 0; i < parens.size(); ++i) {
+      const auto &pair = parens[i];
+      paren_map_[pair.first] = i;
+      paren_map_[pair.second] = i;
+    }
+    if (close_) {
+      const auto start = fst.Start();
+      if (start == kNoStateId) return;
+      if (!DFSearch(start)) {
+        FSTERROR() << "PdtReachable: Underlying cyclicity not supported";
+        error_ = true;
+      }
+    } else {
+      FSTERROR() << "PdtParenReachable: Open paren info not implemented";
+      error_ = true;
+    }
+  }
+
+  bool Error() const { return error_; }
+
+  // Given a state ID, returns an iterator over paren IDs for close (open)
+  // parens reachable from that state along balanced paths.
+  ParenIterator FindParens(StateId s) const {
+    return ParenIterator(paren_multimap_, paren_multimap_.find(s));
+  }
+
+  // Given a paren ID and a state ID s, returns an iterator over states that can
+  // be reached along balanced paths from (to) s that have have close (open)
+  // parentheses matching the paren ID exiting (entering) those states.
+  SetIterator FindStates(Label paren_id, StateId s) const {
+    const State paren_state(paren_id, s);
+    const auto it = set_map_.find(paren_state);
+    if (it == set_map_.end()) {
+      return state_sets_.FindSet(-1);
+    } else {
+      return state_sets_.FindSet(it->second);
+    }
+  }
+
+  // Given a paren ID and a state ID s, return an iterator over arcs that exit
+  // (enter) s and are labeled with a close (open) parenthesis matching the
+  // paren ID.
+  ParenArcIterator FindParenArcs(Label paren_id, StateId s) const {
+    const State paren_state(paren_id, s);
+    return ParenArcIterator(paren_arc_multimap_,
+                            paren_arc_multimap_.find(paren_state));
+  }
+
+ private:
+  // Returns false when cycle detected during DFS gathering paren and state set
+  // information.
+  bool DFSearch(StateId s);
+
+  // Unions state sets together gathered by the DFS.
+  void ComputeStateSet(StateId s);
+
+  // Gathers state set(s) from state.
+  void UpdateStateSet(StateId nextstate, std::set<Label> *paren_set,
+                      std::vector<std::set<StateId>> *state_sets) const;
+
+  const Fst<Arc> &fst_;
+  // Paren IDs to labels.
+  const std::vector<std::pair<Label, Label>> &parens_;
+  // Close/open paren info?
+  const bool close_;
+  // Labels to paren IDs.
+  std::unordered_map<Label, Label> paren_map_;
+  // Paren reachability.
+  ParenMultimap paren_multimap_;
+  // Paren arcs.
+  ParenArcMultimap paren_arc_multimap_;
+  // DFS states.
+  std::vector<uint8> state_color_;
+  // Reachable states to IDs.
+  mutable Collection<ssize_t, StateId> state_sets_;
+  // IDs to reachable states.
+  StateSetMap set_map_;
+  bool error_;
+
+  PdtParenReachable(const PdtParenReachable &) = delete;
+  PdtParenReachable &operator=(const PdtParenReachable &) = delete;
+};
+
+// Gathers paren and state set information.
+template <class Arc>
+bool PdtParenReachable<Arc>::DFSearch(StateId s) {
+  static constexpr uint8 kWhiteState = 0x01;  // Undiscovered.
+  static constexpr uint8 kGreyState = 0x02;   // Discovered & unfinished.
+  static constexpr uint8 kBlackState = 0x04;  // Finished.
+  if (s >= state_color_.size()) state_color_.resize(s + 1, kWhiteState);
+  if (state_color_[s] == kBlackState) return true;
+  if (state_color_[s] == kGreyState) return false;
+  state_color_[s] = kGreyState;
+  for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
+    const auto &arc = aiter.Value();
+    const auto it = paren_map_.find(arc.ilabel);
+    if (it != paren_map_.end()) {  // Paren?
+      const auto paren_id = it->second;
+      if (arc.ilabel == parens_[paren_id].first) {  // Open paren?
+        if (!DFSearch(arc.nextstate)) return false;
+        for (auto set_iter = FindStates(paren_id, arc.nextstate);
+             !set_iter.Done(); set_iter.Next()) {
+          for (auto paren_arc_iter =
+                   FindParenArcs(paren_id, set_iter.Element());
+               !paren_arc_iter.Done(); paren_arc_iter.Next()) {
+            const auto &cparc = paren_arc_iter.Value();
+            if (!DFSearch(cparc.nextstate)) return false;
+          }
+        }
+      }
+    } else if (!DFSearch(arc.nextstate)) {  // Non-paren.
+      return false;
+    }
+  }
+  ComputeStateSet(s);
+  state_color_[s] = kBlackState;
+  return true;
+}
+
+// Unions state sets.
+template <class Arc>
+void PdtParenReachable<Arc>::ComputeStateSet(StateId s) {
+  std::set<Label> paren_set;
+  std::vector<std::set<StateId>> state_sets(parens_.size());
+  for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
+    const auto &arc = aiter.Value();
+    const auto it = paren_map_.find(arc.ilabel);
+    if (it != paren_map_.end()) {  // Paren?
+      const auto paren_id = it->second;
+      if (arc.ilabel == parens_[paren_id].first) {  // Open paren?
+        for (auto set_iter = FindStates(paren_id, arc.nextstate);
+             !set_iter.Done(); set_iter.Next()) {
+          for (auto paren_arc_iter =
+                   FindParenArcs(paren_id, set_iter.Element());
+               !paren_arc_iter.Done(); paren_arc_iter.Next()) {
+            const auto &cparc = paren_arc_iter.Value();
+            UpdateStateSet(cparc.nextstate, &paren_set, &state_sets);
+          }
+        }
+      } else {  // Close paren.
+        paren_set.insert(paren_id);
+        state_sets[paren_id].insert(s);
+        const State paren_state(paren_id, s);
+        paren_arc_multimap_.insert(std::make_pair(paren_state, arc));
+      }
+    } else {  // Non-paren.
+      UpdateStateSet(arc.nextstate, &paren_set, &state_sets);
+    }
+  }
+  std::vector<StateId> state_set;
+  for (auto paren_iter = paren_set.begin(); paren_iter != paren_set.end();
+       ++paren_iter) {
+    state_set.clear();
+    const auto paren_id = *paren_iter;
+    paren_multimap_.insert(std::make_pair(s, paren_id));
+    for (auto state_iter = state_sets[paren_id].begin();
+         state_iter != state_sets[paren_id].end(); ++state_iter) {
+      state_set.push_back(*state_iter);
+    }
+    const State paren_state(paren_id, s);
+    set_map_[paren_state] = state_sets_.FindId(state_set);
+  }
+}
+
+// Gathers state sets.
+template <class Arc>
+void PdtParenReachable<Arc>::UpdateStateSet(
+    StateId nextstate, std::set<Label> *paren_set,
+    std::vector<std::set<StateId>> *state_sets) const {
+  for (auto paren_iter = FindParens(nextstate); !paren_iter.Done();
+       paren_iter.Next()) {
+    const auto paren_id = paren_iter.Value();
+    paren_set->insert(paren_id);
+    for (auto set_iter = FindStates(paren_id, nextstate); !set_iter.Done();
+         set_iter.Next()) {
+      (*state_sets)[paren_id].insert(set_iter.Element());
+    }
+  }
+}
+
+// Stores balancing parenthesis data for a PDT. Unlike PdtParenReachable above
+// this allows on-the-fly construction (e.g., in PdtShortestPath).
+template <class Arc>
+class PdtBalanceData {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+
+  using State = ParenState<Arc>;
+  using StateHash = typename State::Hash;
+
+  // Set for open parens.
+  using OpenParenSet = std::unordered_set<State, StateHash>;
+
+  // Maps from open paren destination state to parenthesis ID.
+  using OpenParenMap = std::unordered_multimap<StateId, Label>;
+
+  // Maps from open paren state to source states of matching close parens
+  using CloseParenMap = std::unordered_multimap<State, StateId, StateHash>;
+
+  // Maps from open paren state to close source set ID.
+  using CloseSourceMap = std::unordered_map<State, ssize_t, StateHash>;
+
+  using SetIterator = typename Collection<ssize_t, StateId>::SetIterator;
+
+  PdtBalanceData() {}
+
+  void Clear() {
+    open_paren_map_.clear();
+    close_paren_map_.clear();
+  }
+
+  // Adds an open parenthesis with destination state open_dest.
+  void OpenInsert(Label paren_id, StateId open_dest) {
+    const State key(paren_id, open_dest);
+    if (!open_paren_set_.count(key)) {
+      open_paren_set_.insert(key);
+      open_paren_map_.emplace(open_dest, paren_id);
+    }
+  }
+
+  // Adds a matching closing parenthesis with source state close_source
+  // balancing an open_parenthesis with destination state open_dest if
+  // OpenInsert() previously called.
+  void CloseInsert(Label paren_id, StateId open_dest, StateId close_source) {
+    const State key(paren_id, open_dest);
+    if (open_paren_set_.count(key)) {
+      close_paren_map_.emplace(key, close_source);
+    }
+  }
+
+  // Finds close paren source states matching an open parenthesis. The following
+  // methods are then used to iterate through those matching states. Should be
+  // called only after FinishInsert(open_dest).
+  SetIterator Find(Label paren_id, StateId open_dest) {
+    const State key(paren_id, open_dest);
+    const auto it = close_source_map_.find(key);
+    if (it == close_source_map_.end()) {
+      return close_source_sets_.FindSet(-1);
+    } else {
+      return close_source_sets_.FindSet(it->second);
+    }
+  }
+
+  // Called when all open and close parenthesis insertions (w.r.t. open
+  // parentheses entering state open_dest) are finished. Must be called before
+  // Find(open_dest).
+  void FinishInsert(StateId open_dest) {
+    std::vector<StateId> close_sources;
+    for (auto oit = open_paren_map_.find(open_dest);
+         oit != open_paren_map_.end() && oit->first == open_dest;) {
+      const auto paren_id = oit->second;
+      close_sources.clear();
+      const State key(paren_id, open_dest);
+      open_paren_set_.erase(open_paren_set_.find(key));
+      for (auto cit = close_paren_map_.find(key);
+           cit != close_paren_map_.end() && cit->first == key;) {
+        close_sources.push_back(cit->second);
+        close_paren_map_.erase(cit++);
+      }
+      std::sort(close_sources.begin(), close_sources.end());
+      auto unique_end = std::unique(close_sources.begin(), close_sources.end());
+      close_sources.resize(unique_end - close_sources.begin());
+      if (!close_sources.empty()) {
+        close_source_map_[key] = close_source_sets_.FindId(close_sources);
+      }
+      open_paren_map_.erase(oit++);
+    }
+  }
+
+  // Returns a new balance data object representing the reversed balance
+  // information.
+  PdtBalanceData<Arc> *Reverse(StateId num_states, StateId num_split,
+                               StateId state_id_shift) const;
+
+ private:
+  // Open paren at destintation state?
+  OpenParenSet open_paren_set_;
+  // Open parens per state.
+  OpenParenMap open_paren_map_;
+  // Current open destination state.
+  State open_dest_;
+  // Current open paren/state.
+  typename OpenParenMap::const_iterator open_iter_;
+  // Close states to (open paren, state).
+  CloseParenMap close_paren_map_;
+  // (Paren, state) to set ID.
+  CloseSourceMap close_source_map_;
+  mutable Collection<ssize_t, StateId> close_source_sets_;
+};
+
+// Return a new balance data object representing the reversed balance
+// information.
+template <class Arc>
+PdtBalanceData<Arc> *PdtBalanceData<Arc>::Reverse(
+    StateId num_states, StateId num_split, StateId state_id_shift) const {
+  auto *bd = new PdtBalanceData<Arc>;
+  std::unordered_set<StateId> close_sources;
+  const auto split_size = num_states / num_split;
+  for (StateId i = 0; i < num_states; i += split_size) {
+    close_sources.clear();
+    for (auto it = close_source_map_.begin(); it != close_source_map_.end();
+         ++it) {
+      const auto &okey = it->first;
+      const auto open_dest = okey.state_id;
+      const auto paren_id = okey.paren_id;
+      for (auto set_iter = close_source_sets_.FindSet(it->second);
+           !set_iter.Done(); set_iter.Next()) {
+        const auto close_source = set_iter.Element();
+        if ((close_source < i) || (close_source >= i + split_size)) continue;
+        close_sources.insert(close_source + state_id_shift);
+        bd->OpenInsert(paren_id, close_source + state_id_shift);
+        bd->CloseInsert(paren_id, close_source + state_id_shift,
+                        open_dest + state_id_shift);
+      }
+    }
+    for (auto it = close_sources.begin(); it != close_sources.end(); ++it) {
+      bd->FinishInsert(*it);
+    }
+  }
+  return bd;
+}
+
+}  // namespace internal
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_PDT_PAREN_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/pdt.h b/projects/llm_framework/include/fst/extensions/pdt/pdt.h
new file mode 100644
index 00000000..d6c5946f
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/pdt/pdt.h
@@ -0,0 +1,165 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Common classes for PDT expansion/traversal.
+
+#ifndef FST_EXTENSIONS_PDT_PDT_H_
+#define FST_EXTENSIONS_PDT_PDT_H_
+
+#include <map>
+#include <set>
+#include <unordered_map>
+
+#include <fst/compat.h>
+#include <fst/log.h>
+#include <fst/fst.h>
+#include <fst/state-table.h>
+
+namespace fst {
+
+// Provides bijection between parenthesis stacks and signed integral stack IDs.
+// Each stack ID is unique to each distinct stack. The open-close parenthesis
+// label pairs are passed using the parens argument.
+template <typename StackId, typename Label>
+class PdtStack {
+ public:
+  // The stacks are stored in a tree. The nodes are stored in a vector. Each
+  // node represents the top of some stack and is identified by its position in
+  // the vector. Its' parent node represents the stack with the top popped and
+  // its children are stored in child_map_ and accessed by stack_id and label.
+  // The paren_id is
+  // the position in parens of the parenthesis for that node.
+  struct StackNode {
+    StackId parent_id;
+    size_t paren_id;
+
+    StackNode(StackId p, size_t i) : parent_id(p), paren_id(i) {}
+  };
+
+  explicit PdtStack(const std::vector<std::pair<Label, Label>> &parens)
+      : parens_(parens), min_paren_(kNoLabel), max_paren_(kNoLabel) {
+    for (size_t i = 0; i < parens.size(); ++i) {
+      const auto &pair = parens[i];
+      paren_map_[pair.first] = i;
+      paren_map_[pair.second] = i;
+      if (min_paren_ == kNoLabel || pair.first < min_paren_) {
+        min_paren_ = pair.first;
+      }
+      if (pair.second < min_paren_) min_paren_ = pair.second;
+      if (max_paren_ == kNoLabel || pair.first > max_paren_) {
+        max_paren_ = pair.first;
+      }
+      if (pair.second > max_paren_) max_paren_ = pair.second;
+    }
+    nodes_.push_back(StackNode(-1, -1));  // Tree root.
+  }
+
+  // Returns stack ID given the current stack ID (0 if empty) and label read.
+  // Pushes onto the stack if the label is an open parenthesis, returning the
+  // new stack ID. Pops the stack if the label is a close parenthesis that
+  // matches the top of the stack, returning the parent stack ID. Returns -1 if
+  // label is an unmatched close parenthesis. Otherwise, returns the current
+  // stack ID.
+  StackId Find(StackId stack_id, Label label) {
+    if (min_paren_ == kNoLabel || label < min_paren_ || label > max_paren_) {
+      return stack_id;  // Non-paren.
+    }
+    const auto it = paren_map_.find(label);
+    // Non-paren.
+    if (it == paren_map_.end()) return stack_id;
+    const auto paren_id = it->second;
+    // Open paren.
+    if (label == parens_[paren_id].first) {
+      auto &child_id = child_map_[std::make_pair(stack_id, label)];
+      if (child_id == 0) {  // Child not found; pushes label.
+        child_id = nodes_.size();
+        nodes_.push_back(StackNode(stack_id, paren_id));
+      }
+      return child_id;
+    }
+    const auto &node = nodes_[stack_id];
+    // Matching close paren.
+    if (paren_id == node.paren_id) return node.parent_id;
+    // Non-matching close paren.
+    return -1;
+  }
+
+  // Returns the stack ID obtained by popping the label at the top of the
+  // current stack ID.
+  StackId Pop(StackId stack_id) const { return nodes_[stack_id].parent_id; }
+
+  // Returns the paren ID at the top of the stack.
+  ssize_t Top(StackId stack_id) const { return nodes_[stack_id].paren_id; }
+
+  ssize_t ParenId(Label label) const {
+    const auto it = paren_map_.find(label);
+    if (it == paren_map_.end()) return -1;  // Non-paren.
+    return it->second;
+  }
+
+ private:
+  struct ChildHash {
+    size_t operator()(const std::pair<StackId, Label> &pair) const {
+      static constexpr size_t prime = 7853;
+      return static_cast<size_t>(pair.first) +
+             static_cast<size_t>(pair.second) * prime;
+    }
+  };
+
+  std::vector<std::pair<Label, Label>> parens_;
+  std::vector<StackNode> nodes_;
+  std::unordered_map<Label, size_t> paren_map_;
+  // Child of stack node w.r.t label
+  std::unordered_map<std::pair<StackId, Label>, StackId, ChildHash> child_map_;
+  Label min_paren_;
+  Label max_paren_;
+};
+
+// State tuple for PDT expansion.
+template <typename S, typename K>
+struct PdtStateTuple {
+  using StateId = S;
+  using StackId = K;
+
+  StateId state_id;
+  StackId stack_id;
+
+  PdtStateTuple(StateId state_id = kNoStateId, StackId stack_id = -1)
+      : state_id(state_id), stack_id(stack_id) {}
+};
+
+// Equality of PDT state tuples.
+template <typename S, typename K>
+inline bool operator==(const PdtStateTuple<S, K> &x,
+                       const PdtStateTuple<S, K> &y) {
+  if (&x == &y) return true;
+  return x.state_id == y.state_id && x.stack_id == y.stack_id;
+}
+
+// Hash function object for PDT state tuples
+template <class T>
+class PdtStateHash {
+ public:
+  size_t operator()(const T &tuple) const {
+    static constexpr auto prime = 7853;
+    return tuple.state_id + tuple.stack_id * prime;
+  }
+};
+
+// Tuple to PDT state bijection.
+template <class StateId, class StackId>
+class PdtStateTable : public CompactHashStateTable<
+                          PdtStateTuple<StateId, StackId>,
+                          PdtStateHash<PdtStateTuple<StateId, StackId>>> {
+ public:
+  PdtStateTable() {}
+
+  PdtStateTable(const PdtStateTable &other) {}
+
+ private:
+  PdtStateTable &operator=(const PdtStateTable &) = delete;
+};
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_PDT_PDT_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/pdtlib.h b/projects/llm_framework/include/fst/extensions/pdt/pdtlib.h
new file mode 100644
index 00000000..41d98986
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/pdt/pdtlib.h
@@ -0,0 +1,19 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// This is an experimental push-down transducer (PDT) library. A PDT is
+// encoded as an FST, where some transitions are labeled with open or close
+// parentheses. To be interpreted as a PDT, the parentheses must balance on a
+// path.
+
+#ifndef FST_EXTENSIONS_PDT_PDTLIB_H_
+#define FST_EXTENSIONS_PDT_PDTLIB_H_
+
+#include <fst/extensions/pdt/compose.h>
+#include <fst/extensions/pdt/expand.h>
+#include <fst/extensions/pdt/pdt.h>
+#include <fst/extensions/pdt/replace.h>
+#include <fst/extensions/pdt/reverse.h>
+#include <fst/extensions/pdt/shortest-path.h>
+
+#endif  // FST_EXTENSIONS_PDT_PDTLIB_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/pdtscript.h b/projects/llm_framework/include/fst/extensions/pdt/pdtscript.h
new file mode 100644
index 00000000..bdf5b4ea
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/pdt/pdtscript.h
@@ -0,0 +1,244 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Convenience file for including all PDT operations at once, and/or
+// registering them for new arc types.
+
+#ifndef FST_EXTENSIONS_PDT_PDTSCRIPT_H_
+#define FST_EXTENSIONS_PDT_PDTSCRIPT_H_
+
+#include <algorithm>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+#include <fst/compose.h>  // for ComposeOptions
+#include <fst/util.h>
+
+#include <fst/script/arg-packs.h>
+#include <fst/script/fstscript.h>
+#include <fst/script/shortest-path.h>
+
+#include <fst/extensions/pdt/compose.h>
+#include <fst/extensions/pdt/expand.h>
+#include <fst/extensions/pdt/info.h>
+#include <fst/extensions/pdt/replace.h>
+#include <fst/extensions/pdt/reverse.h>
+#include <fst/extensions/pdt/shortest-path.h>
+
+namespace fst {
+namespace script {
+
+using PdtComposeArgs =
+    std::tuple<const FstClass &, const FstClass &,
+               const std::vector<LabelPair> &, MutableFstClass *,
+               const PdtComposeOptions &, bool>;
+
+template <class Arc>
+void PdtCompose(PdtComposeArgs *args) {
+  const Fst<Arc> &ifst1 = *(std::get<0>(*args).GetFst<Arc>());
+  const Fst<Arc> &ifst2 = *(std::get<1>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<3>(*args)->GetMutableFst<Arc>();
+  // In case Arc::Label is not the same as FstClass::Label, we make a
+  // copy. Truncation may occur if FstClass::Label has more precision than
+  // Arc::Label.
+  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
+      std::get<2>(*args).size());
+  std::copy(std::get<2>(*args).begin(), std::get<2>(*args).end(),
+            typed_parens.begin());
+  if (std::get<5>(*args)) {
+    Compose(ifst1, typed_parens, ifst2, ofst, std::get<4>(*args));
+  } else {
+    Compose(ifst1, ifst2, typed_parens, ofst, std::get<4>(*args));
+  }
+}
+
+void PdtCompose(const FstClass &ifst1, const FstClass &ifst2,
+                const std::vector<LabelPair> &parens,
+                MutableFstClass *ofst, const PdtComposeOptions &opts,
+                bool left_pdt);
+
+struct PdtExpandOptions {
+  bool connect;
+  bool keep_parentheses;
+  const WeightClass &weight_threshold;
+
+  PdtExpandOptions(bool c, bool k, const WeightClass &w)
+      : connect(c), keep_parentheses(k), weight_threshold(w) {}
+};
+
+using PdtExpandArgs =
+    std::tuple<const FstClass &, const std::vector<LabelPair> &,
+               MutableFstClass *, const PdtExpandOptions &>;
+
+template <class Arc>
+void PdtExpand(PdtExpandArgs *args) {
+  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
+  // In case Arc::Label is not the same as FstClass::Label, we make a
+  // copy. Truncation may occur if FstClass::Label has more precision than
+  // Arc::Label.
+  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
+      std::get<1>(*args).size());
+  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
+            typed_parens.begin());
+  Expand(fst, typed_parens, ofst,
+         fst::PdtExpandOptions<Arc>(
+             std::get<3>(*args).connect, std::get<3>(*args).keep_parentheses,
+             *(std::get<3>(*args)
+                   .weight_threshold.GetWeight<typename Arc::Weight>())));
+}
+
+void PdtExpand(const FstClass &ifst, const std::vector<LabelPair> &parens,
+               MutableFstClass *ofst, const PdtExpandOptions &opts);
+
+void PdtExpand(const FstClass &ifst, const std::vector<LabelPair> &parens,
+               MutableFstClass *ofst, bool connect, bool keep_parentheses,
+               const WeightClass &weight_threshold);
+
+using PdtReplaceArgs =
+    std::tuple<const std::vector<LabelFstClassPair> &, MutableFstClass *,
+               std::vector<LabelPair> *, int64, PdtParserType, int64,
+               const string &, const string &>;
+
+template <class Arc>
+void PdtReplace(PdtReplaceArgs *args) {
+  const auto &untyped_pairs = std::get<0>(*args);
+  auto size = untyped_pairs.size();
+  std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>> typed_pairs(
+      size);
+  for (size_t i = 0; i < size; ++i) {
+    typed_pairs[i].first = untyped_pairs[i].first;
+    typed_pairs[i].second = untyped_pairs[i].second->GetFst<Arc>();
+  }
+  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
+  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens;
+  const PdtReplaceOptions<Arc> opts(std::get<3>(*args), std::get<4>(*args),
+                                    std::get<5>(*args), std::get<6>(*args),
+                                    std::get<7>(*args));
+  Replace(typed_pairs, ofst, &typed_parens, opts);
+  // Copies typed parens into arg3.
+  std::get<2>(*args)->resize(typed_parens.size());
+  std::copy(typed_parens.begin(), typed_parens.end(),
+            std::get<2>(*args)->begin());
+}
+
+void PdtReplace(const std::vector<LabelFstClassPair> &pairs,
+                MutableFstClass *ofst, std::vector<LabelPair> *parens,
+                int64 root, PdtParserType parser_type = PDT_LEFT_PARSER,
+                int64 start_paren_labels = kNoLabel,
+                const string &left_paren_prefix = "(_",
+                const string &right_paren_prefix = "_)");
+
+using PdtReverseArgs =
+    std::tuple<const FstClass &, const std::vector<LabelPair> &,
+               MutableFstClass *>;
+
+template <class Arc>
+void PdtReverse(PdtReverseArgs *args) {
+  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
+  // In case Arc::Label is not the same as FstClass::Label, we make a
+  // copy. Truncation may occur if FstClass::Label has more precision than
+  // Arc::Label.
+  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
+      std::get<1>(*args).size());
+  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
+            typed_parens.begin());
+  Reverse(fst, typed_parens, ofst);
+}
+
+void PdtReverse(const FstClass &ifst, const std::vector<LabelPair> &,
+                MutableFstClass *ofst);
+
+// PDT SHORTESTPATH
+
+struct PdtShortestPathOptions {
+  QueueType queue_type;
+  bool keep_parentheses;
+  bool path_gc;
+
+  PdtShortestPathOptions(QueueType qt = FIFO_QUEUE, bool kp = false,
+                         bool gc = true)
+      : queue_type(qt), keep_parentheses(kp), path_gc(gc) {}
+};
+
+using PdtShortestPathArgs =
+    std::tuple<const FstClass &, const std::vector<LabelPair> &,
+               MutableFstClass *, const PdtShortestPathOptions &>;
+
+template <class Arc>
+void PdtShortestPath(PdtShortestPathArgs *args) {
+  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
+  const PdtShortestPathOptions &opts = std::get<3>(*args);
+  // In case Arc::Label is not the same as FstClass::Label, we make a
+  // copy. Truncation may occur if FstClass::Label has more precision than
+  // Arc::Label.
+  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
+      std::get<1>(*args).size());
+  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
+            typed_parens.begin());
+  switch (opts.queue_type) {
+    default:
+      FSTERROR() << "Unknown queue type: " << opts.queue_type;
+    case FIFO_QUEUE: {
+      using Queue = FifoQueue<typename Arc::StateId>;
+      fst::PdtShortestPathOptions<Arc, Queue> spopts(opts.keep_parentheses,
+                                                         opts.path_gc);
+      ShortestPath(fst, typed_parens, ofst, spopts);
+      return;
+    }
+    case LIFO_QUEUE: {
+      using Queue = LifoQueue<typename Arc::StateId>;
+      fst::PdtShortestPathOptions<Arc, Queue> spopts(opts.keep_parentheses,
+                                                         opts.path_gc);
+      ShortestPath(fst, typed_parens, ofst, spopts);
+      return;
+    }
+    case STATE_ORDER_QUEUE: {
+      using Queue = StateOrderQueue<typename Arc::StateId>;
+      fst::PdtShortestPathOptions<Arc, Queue> spopts(opts.keep_parentheses,
+                                                         opts.path_gc);
+      ShortestPath(fst, typed_parens, ofst, spopts);
+      return;
+    }
+  }
+}
+
+void PdtShortestPath(const FstClass &ifst,
+    const std::vector<LabelPair> &parens, MutableFstClass *ofst,
+    const PdtShortestPathOptions &opts = PdtShortestPathOptions());
+
+// PRINT INFO
+
+using PrintPdtInfoArgs =
+    std::pair<const FstClass &, const std::vector<LabelPair> &>;
+
+template <class Arc>
+void PrintPdtInfo(PrintPdtInfoArgs *args) {
+  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
+  // In case Arc::Label is not the same as FstClass::Label, we make a
+  // copy. Truncation may occur if FstClass::Label has more precision than
+  // Arc::Label.
+  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
+      std::get<1>(*args).size());
+  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
+            typed_parens.begin());
+  PdtInfo<Arc> pdtinfo(fst, typed_parens);
+  PrintPdtInfo(pdtinfo);
+}
+
+void PrintPdtInfo(const FstClass &ifst, const std::vector<LabelPair> &parens);
+
+}  // namespace script
+}  // namespace fst
+
+#define REGISTER_FST_PDT_OPERATIONS(ArcType)                             \
+  REGISTER_FST_OPERATION(PdtCompose, ArcType, PdtComposeArgs);           \
+  REGISTER_FST_OPERATION(PdtExpand, ArcType, PdtExpandArgs);             \
+  REGISTER_FST_OPERATION(PdtReplace, ArcType, PdtReplaceArgs);           \
+  REGISTER_FST_OPERATION(PdtReverse, ArcType, PdtReverseArgs);           \
+  REGISTER_FST_OPERATION(PdtShortestPath, ArcType, PdtShortestPathArgs); \
+  REGISTER_FST_OPERATION(PrintPdtInfo, ArcType, PrintPdtInfoArgs)
+#endif  // FST_EXTENSIONS_PDT_PDTSCRIPT_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/replace.h b/projects/llm_framework/include/fst/extensions/pdt/replace.h
new file mode 100644
index 00000000..76341e8d
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/pdt/replace.h
@@ -0,0 +1,827 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Recursively replaces FST arcs with other FSTs, returning a PDT.
+
+#ifndef FST_EXTENSIONS_PDT_REPLACE_H_
+#define FST_EXTENSIONS_PDT_REPLACE_H_
+
+#include <map>
+#include <memory>
+#include <set>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include <fst/replace.h>
+#include <fst/replace-util.h>
+#include <fst/symbol-table-ops.h>
+
+namespace fst {
+namespace internal {
+
+// Hash to paren IDs
+template <typename S>
+struct ReplaceParenHash {
+  size_t operator()(const std::pair<size_t, S> &paren) const {
+    static constexpr auto prime = 7853;
+    return paren.first + paren.second * prime;
+  }
+};
+
+}  // namespace internal
+
+// Parser types characterize the PDT construction method. When applied to a CFG,
+// each non-terminal is encoded as a DFA that accepts precisely the RHS's of
+// productions of that non-terminal. For parsing (rather than just recognition),
+// production numbers can used as outputs (placed as early as possible) in the
+// DFAs promoted to DFTs. For more information on the strongly regular
+// construction, see:
+//
+// Mohri, M., and Pereira, F. 1998. Dynamic compilation of weighted context-free
+// grammars. In Proc. ACL, pages 891-897.
+enum PdtParserType {
+  // Top-down construction. Applied to a simple LL(1) grammar (among others),
+  // gives a DPDA. If promoted to a DPDT, with outputs being production
+  // numbers, gives a leftmost derivation. Left recursive grammars are
+  // problematic in use.
+  PDT_LEFT_PARSER,
+
+  // Top-down construction. Similar to PDT_LEFT_PARSE except bounded-stack
+  // (expandable as an FST) result with regular or, more generally, strongly
+  // regular grammars. Epsilons may replace some parentheses, which may
+  // introduce some non-determinism.
+  PDT_LEFT_SR_PARSER,
+
+  /* TODO(riley):
+  // Bottom-up construction. Applied to a LR(0) grammar, gives a DPDA.
+  // If promoted to a DPDT, with outputs being the production nubmers,
+  // gives the reverse of a rightmost derivation.
+  PDT_RIGHT_PARSER,
+  */
+};
+
+template <class Arc>
+struct PdtReplaceOptions {
+  using Label = typename Arc::Label;
+
+  explicit PdtReplaceOptions(Label root,
+                             PdtParserType type = PDT_LEFT_PARSER,
+                             Label start_paren_labels = kNoLabel,
+                             string left_paren_prefix = "(_",
+                             string right_paren_prefix = ")_") :
+      root(root), type(type), start_paren_labels(start_paren_labels),
+      left_paren_prefix(std::move(left_paren_prefix)),
+      right_paren_prefix(std::move(right_paren_prefix)) {}
+
+  Label root;
+  PdtParserType type;
+  Label start_paren_labels;
+  const string left_paren_prefix;
+  const string right_paren_prefix;
+};
+
+// PdtParser: Base PDT parser class common to specific parsers.
+
+template <class Arc>
+class PdtParser {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using LabelFstPair = std::pair<Label, const Fst<Arc> *>;
+  using LabelPair = std::pair<Label, Label>;
+  using LabelStatePair = std::pair<Label, StateId>;
+  using StateWeightPair = std::pair<StateId, Weight>;
+  using ParenKey = std::pair<size_t, StateId>;
+  using ParenMap =
+      std::unordered_map<ParenKey, size_t, internal::ReplaceParenHash<StateId>>;
+
+  PdtParser(const std::vector<LabelFstPair> &fst_array,
+            const PdtReplaceOptions<Arc> &opts) :
+      root_(opts.root), start_paren_labels_(opts.start_paren_labels),
+      left_paren_prefix_(std::move(opts.left_paren_prefix)),
+      right_paren_prefix_(std::move(opts.right_paren_prefix)),
+      error_(false) {
+    for (size_t i = 0; i < fst_array.size(); ++i) {
+      if (!CompatSymbols(fst_array[0].second->InputSymbols(),
+                         fst_array[i].second->InputSymbols())) {
+        FSTERROR() << "PdtParser: Input symbol table of input FST " << i
+                   << " does not match input symbol table of 0th input FST";
+        error_ = true;
+      }
+      if (!CompatSymbols(fst_array[0].second->OutputSymbols(),
+                         fst_array[i].second->OutputSymbols())) {
+        FSTERROR() << "PdtParser: Output symbol table of input FST " << i
+                   << " does not match input symbol table of 0th input FST";
+        error_ = true;
+      }
+      fst_array_.emplace_back(fst_array[i].first, fst_array[i].second->Copy());
+      // Builds map from non-terminal label to FST ID.
+      label2id_[fst_array[i].first] = i;
+    }
+  }
+
+  virtual ~PdtParser() {
+    for (auto &pair : fst_array_) delete pair.second;
+  }
+
+  // Constructs the output PDT, dependent on the derived parser type.
+  virtual void GetParser(MutableFst<Arc> *ofst,
+                         std::vector<LabelPair> *parens) = 0;
+
+ protected:
+  const std::vector<LabelFstPair> &FstArray() const { return fst_array_; }
+
+  Label Root() const { return root_; }
+
+  // Maps from non-terminal label to corresponding FST ID, or returns
+  // kNoStateId to signal lookup failure.
+  StateId Label2Id(Label l) const {
+    auto it = label2id_.find(l);
+    return it == label2id_.end() ? kNoStateId : it->second;
+  }
+
+  // Maps from output state to input FST label, state pair, or returns a
+  // (kNoLabel, kNoStateId) pair to signal lookup failure.
+  LabelStatePair GetLabelStatePair(StateId os) const {
+    if (os >= label_state_pairs_.size()) {
+      static const LabelStatePair no_pair(kNoLabel, kNoLabel);
+      return no_pair;
+    } else {
+      return label_state_pairs_[os];
+    }
+  }
+
+  // Maps to output state from input FST (label, state) pair, or returns
+  // kNoStateId to signal lookup failure.
+  StateId GetState(const LabelStatePair &lsp) const {
+    auto it = state_map_.find(lsp);
+    if (it == state_map_.end()) {
+      return kNoStateId;
+    } else {
+      return it->second;
+    }
+  }
+
+  // Builds single FST combining all referenced input FSTs, leaving in the
+  // non-termnals for now; also tabulates the PDT states that correspond to the
+  // start and final states of the input FSTs.
+  void CreateFst(MutableFst<Arc> *ofst, std::vector<StateId> *open_dest,
+                 std::vector<std::vector<StateWeightPair>> *close_src);
+
+  // Assigns parenthesis labels from total allocated paren IDs.
+  void AssignParenLabels(size_t total_nparens, std::vector<LabelPair> *parens) {
+    parens->clear();
+    for (size_t paren_id = 0; paren_id < total_nparens; ++paren_id) {
+      const auto open_paren = start_paren_labels_ + paren_id;
+      const auto close_paren = open_paren + total_nparens;
+      parens->emplace_back(open_paren, close_paren);
+    }
+  }
+
+  // Determines how non-terminal instances are assigned parentheses IDs.
+  virtual size_t AssignParenIds(const Fst<Arc> &ofst,
+                                ParenMap *paren_map) const = 0;
+
+  // Changes a non-terminal transition to an open parenthesis transition
+  // redirected to the PDT state specified in the open_dest argument, when
+  // indexed by the input FST ID for the non-terminal. Adds close parenthesis
+  // transitions (with specified weights) from the PDT states specified in the
+  // close_src argument, when indexed by the input FST ID for the non-terminal,
+  // to the former destination state of the non-terminal transition. The
+  // paren_map argument gives the parenthesis ID for a given non-terminal FST ID
+  // and destination state pair. The close_non_term_weight vector specifies
+  // non-terminals for which the non-terminal arc weight should be applied on
+  // the close parenthesis (multiplying the close_src weight above) rather than
+  // on the open parenthesis. If no paren ID is found, then an epsilon replaces
+  // the parenthesis that would carry the non-terminal arc weight and the other
+  // parenthesis is omitted (appropriate for the strongly-regular case).
+  void AddParensToFst(
+      const std::vector<LabelPair> &parens,
+      const ParenMap &paren_map,
+      const std::vector<StateId> &open_dest,
+      const std::vector<std::vector<StateWeightPair>> &close_src,
+      const std::vector<bool> &close_non_term_weight,
+      MutableFst<Arc> *ofst);
+
+  // Ensures that parentheses arcs are added to the symbol table.
+  void AddParensToSymbolTables(const std::vector<LabelPair> &parens,
+                               MutableFst<Arc> *ofst);
+
+ private:
+  std::vector<LabelFstPair> fst_array_;
+  Label root_;
+  // Index to use for the first parenthesis.
+  Label start_paren_labels_;
+  const string left_paren_prefix_;
+  const string right_paren_prefix_;
+  // Maps from non-terminal label to FST ID.
+  std::unordered_map<Label, StateId> label2id_;
+  // Given an output state, specifies the input FST (label, state) pair.
+  std::vector<LabelStatePair> label_state_pairs_;
+  // Given an FST (label, state) pair, specifies the output FST state ID.
+  std::map<LabelStatePair, StateId> state_map_;
+  bool error_;
+};
+
+template <class Arc>
+void PdtParser<Arc>::CreateFst(
+    MutableFst<Arc> *ofst, std::vector<StateId> *open_dest,
+    std::vector<std::vector<StateWeightPair>> *close_src) {
+  ofst->DeleteStates();
+  if (error_) {
+    ofst->SetProperties(kError, kError);
+    return;
+  }
+  open_dest->resize(fst_array_.size(), kNoStateId);
+  close_src->resize(fst_array_.size());
+  // Queue of non-terminals to replace.
+  std::deque<Label> non_term_queue;
+  non_term_queue.push_back(root_);
+  // Has a non-terminal been enqueued?
+  std::vector<bool> enqueued(fst_array_.size(), false);
+  enqueued[label2id_[root_]] = true;
+  Label max_label = kNoLabel;
+  for (StateId soff = 0; !non_term_queue.empty(); soff = ofst->NumStates()) {
+    const auto label = non_term_queue.front();
+    non_term_queue.pop_front();
+    StateId fst_id = Label2Id(label);
+    const auto *ifst = fst_array_[fst_id].second;
+    for (StateIterator<Fst<Arc>> siter(*ifst); !siter.Done(); siter.Next()) {
+      const auto is = siter.Value();
+      const auto os = ofst->AddState();
+      const LabelStatePair lsp(label, is);
+      label_state_pairs_.push_back(lsp);
+      state_map_[lsp] = os;
+      if (is == ifst->Start()) {
+        (*open_dest)[fst_id] = os;
+        if (label == root_) ofst->SetStart(os);
+      }
+      if (ifst->Final(is) != Weight::Zero()) {
+        if (label == root_) ofst->SetFinal(os, ifst->Final(is));
+        (*close_src)[fst_id].emplace_back(os, ifst->Final(is));
+      }
+      for (ArcIterator<Fst<Arc>> aiter(*ifst, is); !aiter.Done();
+           aiter.Next()) {
+        auto arc = aiter.Value();
+        arc.nextstate += soff;
+        if (max_label == kNoLabel || arc.olabel > max_label)
+          max_label = arc.olabel;
+        const auto nfst_id = Label2Id(arc.olabel);
+        if (nfst_id != kNoStateId) {
+          if (fst_array_[nfst_id].second->Start() == kNoStateId) continue;
+          if (!enqueued[nfst_id]) {
+            non_term_queue.push_back(arc.olabel);
+            enqueued[nfst_id] = true;
+          }
+        }
+        ofst->AddArc(os, arc);
+      }
+    }
+  }
+  if (start_paren_labels_ == kNoLabel)
+    start_paren_labels_ = max_label + 1;
+}
+
+template <class Arc>
+void PdtParser<Arc>::AddParensToFst(
+    const std::vector<LabelPair> &parens,
+    const ParenMap &paren_map,
+    const std::vector<StateId> &open_dest,
+    const std::vector<std::vector<StateWeightPair>> &close_src,
+    const std::vector<bool> &close_non_term_weight,
+    MutableFst<Arc> *ofst) {
+  StateId dead_state = kNoStateId;
+  using MIter = MutableArcIterator<MutableFst<Arc>>;
+  for (StateIterator<Fst<Arc>> siter(*ofst); !siter.Done(); siter.Next()) {
+    StateId os = siter.Value();
+    std::unique_ptr<MIter> aiter(new MIter(ofst, os));
+    for (auto n = 0; !aiter->Done(); aiter->Next(), ++n) {
+      const auto arc = aiter->Value();  // A reference here may go stale.
+      StateId nfst_id = Label2Id(arc.olabel);
+      if (nfst_id != kNoStateId) {
+        // Gets parentheses.
+        const ParenKey paren_key(nfst_id, arc.nextstate);
+        auto it = paren_map.find(paren_key);
+        Label open_paren = 0;
+        Label close_paren = 0;
+        if (it != paren_map.end()) {
+          const auto paren_id = it->second;
+          open_paren = parens[paren_id].first;
+          close_paren = parens[paren_id].second;
+        }
+        // Sets open parenthesis.
+        if (open_paren != 0 || !close_non_term_weight[nfst_id]) {
+          const auto open_weight =
+              close_non_term_weight[nfst_id] ? Weight::One() : arc.weight;
+          const Arc sarc(open_paren, open_paren, open_weight,
+                         open_dest[nfst_id]);
+          aiter->SetValue(sarc);
+        } else {
+          if (dead_state == kNoStateId) {
+            dead_state = ofst->AddState();
+          }
+          const Arc sarc(0, 0, Weight::One(), dead_state);
+          aiter->SetValue(sarc);
+        }
+        // Adds close parentheses.
+        if (close_paren != 0 || close_non_term_weight[nfst_id]) {
+          for (size_t i = 0; i < close_src[nfst_id].size(); ++i) {
+            const auto &pair = close_src[nfst_id][i];
+            const auto close_weight = close_non_term_weight[nfst_id]
+                                          ? Times(arc.weight, pair.second)
+                                          : pair.second;
+            const Arc farc(close_paren, close_paren, close_weight,
+                           arc.nextstate);
+
+            ofst->AddArc(pair.first, farc);
+            if (os == pair.first) {  // Invalidated iterator.
+              aiter.reset(new MIter(ofst, os));
+              aiter->Seek(n);
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+template <class Arc>
+void PdtParser<Arc>::AddParensToSymbolTables(
+    const std::vector<LabelPair> &parens, MutableFst<Arc> *ofst) {
+  auto size = parens.size();
+  if (ofst->InputSymbols()) {
+    if (!AddAuxiliarySymbols(left_paren_prefix_, start_paren_labels_, size,
+                             ofst->MutableInputSymbols())) {
+      ofst->SetProperties(kError, kError);
+      return;
+    }
+    if (!AddAuxiliarySymbols(right_paren_prefix_, start_paren_labels_ + size,
+                             size, ofst->MutableInputSymbols())) {
+      ofst->SetProperties(kError, kError);
+      return;
+    }
+  }
+  if (ofst->OutputSymbols()) {
+    if (!AddAuxiliarySymbols(left_paren_prefix_, start_paren_labels_, size,
+                             ofst->MutableOutputSymbols())) {
+      ofst->SetProperties(kError, kError);
+      return;
+    }
+    if (!AddAuxiliarySymbols(right_paren_prefix_, start_paren_labels_ + size,
+                             size, ofst->MutableOutputSymbols())) {
+      ofst->SetProperties(kError, kError);
+      return;
+    }
+  }
+}
+
+// Builds a PDT by recursive replacement top-down, where the call and return are
+// encoded in the parentheses.
+template <class Arc>
+class PdtLeftParser final : public PdtParser<Arc> {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using LabelFstPair = typename PdtParser<Arc>::LabelFstPair;
+  using LabelPair = typename PdtParser<Arc>::LabelPair;
+  using LabelStatePair = typename PdtParser<Arc>::LabelStatePair;
+  using StateWeightPair = typename PdtParser<Arc>::StateWeightPair;
+  using ParenKey = typename PdtParser<Arc>::ParenKey;
+  using ParenMap = typename PdtParser<Arc>::ParenMap;
+
+  using PdtParser<Arc>::AddParensToFst;
+  using PdtParser<Arc>::AddParensToSymbolTables;
+  using PdtParser<Arc>::AssignParenLabels;
+  using PdtParser<Arc>::CreateFst;
+  using PdtParser<Arc>::FstArray;
+  using PdtParser<Arc>::GetLabelStatePair;
+  using PdtParser<Arc>::GetState;
+  using PdtParser<Arc>::Label2Id;
+  using PdtParser<Arc>::Root;
+
+  PdtLeftParser(const std::vector<LabelFstPair> &fst_array,
+                const PdtReplaceOptions<Arc> &opts) :
+      PdtParser<Arc>(fst_array, opts) { }
+
+  void GetParser(MutableFst<Arc> *ofst,
+                 std::vector<LabelPair> *parens) override;
+
+ protected:
+  // Assigns a unique parenthesis ID for each non-terminal, destination
+  // state pair.
+  size_t AssignParenIds(const Fst<Arc> &ofst,
+                        ParenMap *paren_map) const override;
+};
+
+template <class Arc>
+void PdtLeftParser<Arc>::GetParser(
+    MutableFst<Arc> *ofst,
+    std::vector<LabelPair> *parens) {
+  ofst->DeleteStates();
+  parens->clear();
+  const auto &fst_array = FstArray();
+  // Map that gives the paren ID for a (non-terminal, dest. state) pair
+  // (which can be unique).
+  ParenMap paren_map;
+  // Specifies the open parenthesis destination state for a given non-terminal.
+  // The source is the non-terminal instance source state.
+  std::vector<StateId> open_dest(fst_array.size(), kNoStateId);
+  // Specifies close parenthesis source states and weights for a given
+  // non-terminal. The destination is the non-terminal instance destination
+  // state.
+  std::vector<std::vector<StateWeightPair>> close_src(fst_array.size());
+  // Specifies non-terminals for which the non-terminal arc weight
+  // should be applied on the close parenthesis (multiplying the
+  // 'close_src' weight above) rather than on the open parenthesis.
+  std::vector<bool> close_non_term_weight(fst_array.size(), false);
+  CreateFst(ofst, &open_dest, &close_src);
+  auto total_nparens = AssignParenIds(*ofst, &paren_map);
+  AssignParenLabels(total_nparens, parens);
+  AddParensToFst(*parens, paren_map, open_dest, close_src,
+                 close_non_term_weight, ofst);
+  if (!fst_array.empty()) {
+    ofst->SetInputSymbols(fst_array[0].second->InputSymbols());
+    ofst->SetOutputSymbols(fst_array[0].second->OutputSymbols());
+  }
+  AddParensToSymbolTables(*parens, ofst);
+}
+
+template <class Arc>
+size_t PdtLeftParser<Arc>::AssignParenIds(
+    const Fst<Arc> &ofst,
+    ParenMap *paren_map) const {
+  // Number of distinct parenthesis pairs per FST.
+  std::vector<size_t> nparens(FstArray().size(), 0);
+  // Number of distinct parenthesis pairs overall.
+  size_t total_nparens = 0;
+  for (StateIterator<Fst<Arc>> siter(ofst); !siter.Done(); siter.Next()) {
+    const auto os = siter.Value();
+    for (ArcIterator<Fst<Arc>> aiter(ofst, os); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      const auto nfst_id = Label2Id(arc.olabel);
+      if (nfst_id != kNoStateId) {
+        const ParenKey paren_key(nfst_id, arc.nextstate);
+        auto it = paren_map->find(paren_key);
+        if (it == paren_map->end()) {
+          // Assigns new paren ID for this (FST, dest state) pair.
+          (*paren_map)[paren_key] = nparens[nfst_id]++;
+          if (nparens[nfst_id] > total_nparens)
+            total_nparens = nparens[nfst_id];
+        }
+      }
+    }
+  }
+  return total_nparens;
+}
+
+// Similar to PdtLeftParser but:
+//
+// 1. Uses epsilons rather than parentheses labels for any non-terminal
+//    instances within a left- (right-) linear dependency SCC,
+// 2. Allocates a paren ID uniquely for each such dependency SCC (rather than
+//    non-terminal = dependency state) and destination state.
+template <class Arc>
+class PdtLeftSRParser final : public PdtParser<Arc> {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using LabelFstPair = typename PdtParser<Arc>::LabelFstPair;
+  using LabelPair = typename PdtParser<Arc>::LabelPair;
+  using LabelStatePair = typename PdtParser<Arc>::LabelStatePair;
+  using StateWeightPair = typename PdtParser<Arc>::StateWeightPair;
+  using ParenKey = typename PdtParser<Arc>::ParenKey;
+  using ParenMap = typename PdtParser<Arc>::ParenMap;
+
+  using PdtParser<Arc>::AddParensToFst;
+  using PdtParser<Arc>::AddParensToSymbolTables;
+  using PdtParser<Arc>::AssignParenLabels;
+  using PdtParser<Arc>::CreateFst;
+  using PdtParser<Arc>::FstArray;
+  using PdtParser<Arc>::GetLabelStatePair;
+  using PdtParser<Arc>::GetState;
+  using PdtParser<Arc>::Label2Id;
+  using PdtParser<Arc>::Root;
+
+  PdtLeftSRParser(const std::vector<LabelFstPair> &fst_array,
+                  const PdtReplaceOptions<Arc> &opts) :
+      PdtParser<Arc>(fst_array, opts),
+      replace_util_(fst_array, ReplaceUtilOptions(opts.root)) { }
+
+  void GetParser(MutableFst<Arc> *ofst,
+                 std::vector<LabelPair> *parens) override;
+
+ protected:
+  // Assigns a unique parenthesis ID for each non-terminal, destination state
+  // pair when the non-terminal refers to a non-linear FST. Otherwise, assigns
+  // a unique parenthesis ID for each dependency SCC, destination state pair if
+  // the non-terminal instance is between
+  // SCCs. Otherwise does nothing.
+  size_t AssignParenIds(const Fst<Arc> &ofst,
+                        ParenMap *paren_map) const override;
+
+  // Returns dependency SCC for given label.
+  size_t SCC(Label label) const { return replace_util_.SCC(label); }
+
+  // Is a given dependency SCC left-linear?
+  bool SCCLeftLinear(size_t scc_id) const {
+    const auto ll_props = kReplaceSCCLeftLinear | kReplaceSCCNonTrivial;
+    const auto scc_props = replace_util_.SCCProperties(scc_id);
+    return (scc_props & ll_props) == ll_props;
+  }
+
+  // Is a given dependency SCC right-linear?
+  bool SCCRightLinear(size_t scc_id) const {
+    const auto lr_props = kReplaceSCCRightLinear | kReplaceSCCNonTrivial;
+    const auto scc_props = replace_util_.SCCProperties(scc_id);
+    return (scc_props & lr_props) == lr_props;
+  }
+
+  // Components of left- (right-) linear dependency SCC; empty o.w.
+  const std::vector<size_t> &SCCComps(size_t scc_id) const {
+    if (scc_comps_.empty()) GetSCCComps();
+    return scc_comps_[scc_id];
+  }
+
+  // Returns the representative state of an SCC. For left-linear grammars, it
+  // is one of the initial states. For right-linear grammars, it is one of the
+  // non-terminal destination states; otherwise, it is kNoStateId.
+  StateId RepState(size_t scc_id) const {
+    if (SCCComps(scc_id).empty()) return kNoStateId;
+    const auto fst_id = SCCComps(scc_id).front();
+    const auto &fst_array = FstArray();
+    const auto label = fst_array[fst_id].first;
+    const auto *ifst = fst_array[fst_id].second;
+    if (SCCLeftLinear(scc_id)) {
+      const LabelStatePair lsp(label, ifst->Start());
+      return GetState(lsp);
+    } else {  // Right-linear.
+      const LabelStatePair lsp(label, *NonTermDests(fst_id).begin());
+      return GetState(lsp);
+    }
+    return kNoStateId;
+  }
+
+ private:
+  // Merges initial (final) states of in a left- (right-) linear dependency SCC
+  // after dealing with the non-terminal arc and final weights.
+  void ProcSCCs(MutableFst<Arc> *ofst,
+                std::vector<StateId> *open_dest,
+                std::vector<std::vector<StateWeightPair>> *close_src,
+                std::vector<bool> *close_non_term_weight) const;
+
+  // Computes components of left- (right-) linear dependency SCC.
+  void GetSCCComps() const {
+    const std::vector<LabelFstPair> &fst_array = FstArray();
+    for (size_t i = 0; i < fst_array.size(); ++i) {
+      const auto label = fst_array[i].first;
+      const auto scc_id = SCC(label);
+      if (scc_comps_.size() <= scc_id) scc_comps_.resize(scc_id + 1);
+      if (SCCLeftLinear(scc_id) || SCCRightLinear(scc_id)) {
+        scc_comps_[scc_id].push_back(i);
+      }
+    }
+  }
+
+  const std::set<StateId> &NonTermDests(StateId fst_id) const {
+    if (non_term_dests_.empty()) GetNonTermDests();
+    return non_term_dests_[fst_id];
+  }
+
+  // Finds non-terminal destination states for right-linear FSTS, or does
+  // nothing if not found.
+  void GetNonTermDests() const;
+
+  // Dependency SCC info.
+  mutable ReplaceUtil<Arc> replace_util_;
+  // Components of left- (right-) linear dependency SCCs, or empty otherwise.
+  mutable std::vector<std::vector<size_t>> scc_comps_;
+  // States that have non-terminals entering them for each (right-linear) FST.
+  mutable std::vector<std::set<StateId>> non_term_dests_;
+};
+
+template <class Arc>
+void PdtLeftSRParser<Arc>::GetParser(
+    MutableFst<Arc> *ofst,
+    std::vector<LabelPair> *parens) {
+  ofst->DeleteStates();
+  parens->clear();
+  const auto &fst_array = FstArray();
+  // Map that gives the paren ID for a (non-terminal, dest. state) pair.
+  ParenMap paren_map;
+  // Specifies the open parenthesis destination state for a given non-terminal.
+  // The source is the non-terminal instance source state.
+  std::vector<StateId> open_dest(fst_array.size(), kNoStateId);
+  // Specifies close parenthesis source states and weights for a given
+  // non-terminal. The destination is the non-terminal instance destination
+  // state.
+  std::vector<std::vector<StateWeightPair>> close_src(fst_array.size());
+  // Specifies non-terminals for which the non-terminal arc weight should be
+  // applied on the close parenthesis (multiplying the close_src weight above)
+  // rather than on the open parenthesis.
+  std::vector<bool> close_non_term_weight(fst_array.size(), false);
+  CreateFst(ofst, &open_dest, &close_src);
+  ProcSCCs(ofst, &open_dest, &close_src, &close_non_term_weight);
+  const auto total_nparens = AssignParenIds(*ofst, &paren_map);
+  AssignParenLabels(total_nparens, parens);
+  AddParensToFst(*parens, paren_map, open_dest, close_src,
+                 close_non_term_weight, ofst);
+  if (!fst_array.empty()) {
+    ofst->SetInputSymbols(fst_array[0].second->InputSymbols());
+    ofst->SetOutputSymbols(fst_array[0].second->OutputSymbols());
+  }
+  AddParensToSymbolTables(*parens, ofst);
+  Connect(ofst);
+}
+
+template <class Arc>
+void PdtLeftSRParser<Arc>::ProcSCCs(
+    MutableFst<Arc> *ofst,
+    std::vector<StateId> *open_dest,
+    std::vector<std::vector<StateWeightPair>> *close_src,
+    std::vector<bool> *close_non_term_weight) const {
+  const auto &fst_array = FstArray();
+  for (StateIterator<Fst<Arc>> siter(*ofst); !siter.Done(); siter.Next()) {
+    const auto os = siter.Value();
+    const auto label = GetLabelStatePair(os).first;
+    const auto is = GetLabelStatePair(os).second;
+    const auto fst_id = Label2Id(label);
+    const auto scc_id = SCC(label);
+    const auto rs = RepState(scc_id);
+    const auto *ifst = fst_array[fst_id].second;
+    // SCC LEFT-LINEAR: puts non-terminal weights on close parentheses. Merges
+    // initial states into SCC representative state and updates open_dest.
+    if (SCCLeftLinear(scc_id)) {
+      (*close_non_term_weight)[fst_id] = true;
+      if (is == ifst->Start() && os != rs) {
+        for (ArcIterator<Fst<Arc>> aiter(*ofst, os); !aiter.Done();
+             aiter.Next()) {
+          const auto &arc = aiter.Value();
+          ofst->AddArc(rs, arc);
+        }
+        ofst->DeleteArcs(os);
+        if (os == ofst->Start())
+          ofst->SetStart(rs);
+        (*open_dest)[fst_id] = rs;
+      }
+    }
+    // SCC RIGHT-LINEAR: pushes back final weights onto non-terminals, if
+    // possible, or adds weighted epsilons to the SCC representative state.
+    // Merges final states into SCC representative state and updates close_src.
+    if (SCCRightLinear(scc_id)) {
+      for (MutableArcIterator<MutableFst<Arc>> aiter(ofst, os); !aiter.Done();
+           aiter.Next()) {
+        auto arc = aiter.Value();
+        const auto idest = GetLabelStatePair(arc.nextstate).second;
+        if (NonTermDests(fst_id).count(idest) > 0) {
+          if (ofst->Final(arc.nextstate) != Weight::Zero()) {
+            ofst->SetFinal(arc.nextstate, Weight::Zero());
+            ofst->SetFinal(rs, Weight::One());
+          }
+          arc.weight = Times(arc.weight, ifst->Final(idest));
+          arc.nextstate = rs;
+          aiter.SetValue(arc);
+        }
+      }
+      const auto final_weight = ifst->Final(is);
+      if (final_weight != Weight::Zero() &&
+          NonTermDests(fst_id).count(is) == 0) {
+        ofst->AddArc(os, Arc(0, 0, final_weight, rs));
+        if (ofst->Final(os) != Weight::Zero()) {
+          ofst->SetFinal(os, Weight::Zero());
+          ofst->SetFinal(rs, Weight::One());
+        }
+      }
+      if (is == ifst->Start()) {
+        (*close_src)[fst_id].clear();
+        (*close_src)[fst_id].emplace_back(rs, Weight::One());
+      }
+    }
+  }
+}
+
+template <class Arc>
+void PdtLeftSRParser<Arc>::GetNonTermDests() const {
+  const auto &fst_array = FstArray();
+  non_term_dests_.resize(fst_array.size());
+  for (size_t fst_id = 0; fst_id < fst_array.size(); ++fst_id) {
+    const auto label = fst_array[fst_id].first;
+    const auto scc_id = SCC(label);
+    if (SCCRightLinear(scc_id)) {
+      const auto *ifst = fst_array[fst_id].second;
+      for (StateIterator<Fst<Arc>> siter(*ifst); !siter.Done(); siter.Next()) {
+        const auto is = siter.Value();
+        for (ArcIterator<Fst<Arc>> aiter(*ifst, is); !aiter.Done();
+             aiter.Next()) {
+          const auto &arc = aiter.Value();
+          if (Label2Id(arc.olabel) != kNoStateId) {
+            non_term_dests_[fst_id].insert(arc.nextstate);
+          }
+        }
+      }
+    }
+  }
+}
+
+template <class Arc>
+size_t PdtLeftSRParser<Arc>::AssignParenIds(
+    const Fst<Arc> &ofst,
+    ParenMap *paren_map) const {
+  const auto &fst_array = FstArray();
+  // Number of distinct parenthesis pairs per FST.
+  std::vector<size_t> nparens(fst_array.size(), 0);
+  // Number of distinct parenthesis pairs overall.
+  size_t total_nparens = 0;
+  for (StateIterator<Fst<Arc>> siter(ofst); !siter.Done(); siter.Next()) {
+    const auto os = siter.Value();
+    const auto label = GetLabelStatePair(os).first;
+    const auto scc_id = SCC(label);
+    for (ArcIterator<Fst<Arc>> aiter(ofst, os); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      const auto nfst_id = Label2Id(arc.olabel);
+      if (nfst_id != kNoStateId) {
+        size_t nscc_id = SCC(arc.olabel);
+        bool nscc_linear = !SCCComps(nscc_id).empty();
+        // Assigns a parenthesis ID for the non-terminal transition
+        // if the non-terminal belongs to a (left-/right-) linear dependency
+        // SCC or if the transition is in an FST from a different SCC
+        if (!nscc_linear || scc_id != nscc_id) {
+          // For (left-/right-) linear SCCs instead of using nfst_id, we
+          // will use its SCC prototype pfst_id for assigning distinct
+          // parenthesis IDs.
+          const auto pfst_id =
+              nscc_linear ? SCCComps(nscc_id).front() : nfst_id;
+          ParenKey paren_key(pfst_id, arc.nextstate);
+          const auto it = paren_map->find(paren_key);
+          if (it == paren_map->end()) {
+            // Assigns new paren ID for this (FST/SCC, dest. state) pair.
+            if (nscc_linear) {
+              // This is mapping we'll need, but we also store (harmlessly)
+              // for the prototype below so we can easily keep count per SCC.
+              const ParenKey nparen_key(nfst_id, arc.nextstate);
+              (*paren_map)[nparen_key] = nparens[pfst_id];
+            }
+            (*paren_map)[paren_key] = nparens[pfst_id]++;
+            if (nparens[pfst_id] > total_nparens) {
+              total_nparens = nparens[pfst_id];
+            }
+          }
+        }
+      }
+    }
+  }
+  return total_nparens;
+}
+
+// Builds a pushdown transducer (PDT) from an RTN specification. The result is
+// a PDT written to a mutable FST where some transitions are labeled with
+// open or close parentheses. To be interpreted as a PDT, the parens must
+// balance on a path (see PdtExpand()). The open/close parenthesis label pairs
+// are returned in the parens argument.
+template <class Arc>
+void Replace(
+    const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
+        &ifst_array,
+    MutableFst<Arc> *ofst,
+    std::vector<std::pair<typename Arc::Label, typename Arc::Label>> *parens,
+    const PdtReplaceOptions<Arc> &opts) {
+  switch (opts.type) {
+    case PDT_LEFT_PARSER:
+      {
+        PdtLeftParser<Arc> pr(ifst_array, opts);
+        pr.GetParser(ofst, parens);
+        return;
+      }
+    case PDT_LEFT_SR_PARSER:
+      {
+        PdtLeftSRParser<Arc> pr(ifst_array, opts);
+        pr.GetParser(ofst, parens);
+        return;
+      }
+    default:
+      FSTERROR() << "Replace: Unknown PDT parser type: " << opts.type;
+      ofst->DeleteStates();
+      ofst->SetProperties(kError, kError);
+      parens->clear();
+      return;
+  }
+}
+
+// Variant where the only user-controlled arguments is the root ID.
+template <class Arc>
+void Replace(
+    const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
+        &ifst_array,
+    MutableFst<Arc> *ofst,
+    std::vector<std::pair<typename Arc::Label, typename Arc::Label>> *parens,
+    typename Arc::Label root) {
+  PdtReplaceOptions<Arc> opts(root);
+  Replace(ifst_array, ofst, parens, opts);
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_PDT_REPLACE_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/reverse.h b/projects/llm_framework/include/fst/extensions/pdt/reverse.h
new file mode 100644
index 00000000..3bd44e21
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/pdt/reverse.h
@@ -0,0 +1,38 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Expands a PDT to an FST.
+
+#ifndef FST_EXTENSIONS_PDT_REVERSE_H_
+#define FST_EXTENSIONS_PDT_REVERSE_H_
+
+#include <vector>
+
+#include <fst/mutable-fst.h>
+#include <fst/relabel.h>
+#include <fst/reverse.h>
+
+namespace fst {
+
+// Reverses a pushdown transducer (PDT) encoded as an FST.
+template <class Arc, class RevArc>
+void Reverse(const Fst<Arc> &ifst,
+             const std::vector<
+                 std::pair<typename Arc::Label, typename Arc::Label>> &parens,
+             MutableFst<RevArc> *ofst) {
+  using Label = typename Arc::Label;
+  // Reverses FST component.
+  Reverse(ifst, ofst);
+  // Exchanges open and close parenthesis pairs.
+  std::vector<std::pair<Label, Label>> relabel_pairs;
+  relabel_pairs.reserve(2 * parens.size());
+  for (const auto &pair : parens) {
+    relabel_pairs.emplace_back(pair.first, pair.second);
+    relabel_pairs.emplace_back(pair.second, pair.first);
+  }
+  Relabel(ofst, relabel_pairs, relabel_pairs);
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_PDT_REVERSE_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/shortest-path.h b/projects/llm_framework/include/fst/extensions/pdt/shortest-path.h
new file mode 100644
index 00000000..591c1ea4
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/pdt/shortest-path.h
@@ -0,0 +1,715 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions to find shortest paths in a PDT.
+
+#ifndef FST_EXTENSIONS_PDT_SHORTEST_PATH_H_
+#define FST_EXTENSIONS_PDT_SHORTEST_PATH_H_
+
+#include <stack>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/extensions/pdt/paren.h>
+#include <fst/extensions/pdt/pdt.h>
+#include <fst/shortest-path.h>
+
+namespace fst {
+
+template <class Arc, class Queue>
+struct PdtShortestPathOptions {
+  bool keep_parentheses;
+  bool path_gc;
+
+  PdtShortestPathOptions(bool keep_parentheses = false, bool path_gc = true)
+      : keep_parentheses(keep_parentheses), path_gc(path_gc) {}
+};
+
+namespace internal {
+
+// Flags for shortest path data.
+
+constexpr uint8 kPdtInited = 0x01;
+constexpr uint8 kPdtFinal = 0x02;
+constexpr uint8 kPdtMarked = 0x04;
+
+// Stores shortest path tree info Distance(), Parent(), and ArcParent()
+// information keyed on two types:
+//
+// 1. SearchState: This is a usual node in a shortest path tree but:
+//    a. is w.r.t a PDT search state (a pair of a PDT state and a "start" state,
+//    either the PDT start state or the destination state of an open
+//    parenthesis).
+//    b. the Distance() is from this "start" state to the search state.
+//    c. Parent().state is kNoLabel for the "start" state.
+//
+// 2. ParenSpec: This connects shortest path trees depending on the the
+// parenthesis taken. Given the parenthesis spec:
+//    a. the Distance() is from the Parent() "start" state to the parenthesis
+//    destination state.
+//    b. The ArcParent() is the parenthesis arc.
+template <class Arc>
+class PdtShortestPathData {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  struct SearchState {
+    StateId state;  // PDT state.
+    StateId start;  // PDT paren "start" state.
+
+    SearchState(StateId s = kNoStateId, StateId t = kNoStateId)
+        : state(s), start(t) {}
+
+    bool operator==(const SearchState &other) const {
+      if (&other == this) return true;
+      return other.state == state && other.start == start;
+    }
+  };
+
+  // Specifies paren ID, source and dest "start" states of a paren. These are
+  // the "start" states of the respective sub-graphs.
+  struct ParenSpec {
+    ParenSpec(Label paren_id = kNoLabel, StateId src_start = kNoStateId,
+              StateId dest_start = kNoStateId)
+        : paren_id(paren_id), src_start(src_start), dest_start(dest_start) {}
+
+    Label paren_id;
+    StateId src_start;   // Sub-graph "start" state for paren source.
+    StateId dest_start;  // Sub-graph "start" state for paren dest.
+
+    bool operator==(const ParenSpec &other) const {
+      if (&other == this) return true;
+      return (other.paren_id == paren_id &&
+              other.src_start == other.src_start &&
+              other.dest_start == dest_start);
+    }
+  };
+
+  struct SearchData {
+    SearchData()
+        : distance(Weight::Zero()),
+          parent(kNoStateId, kNoStateId),
+          paren_id(kNoLabel),
+          flags(0) {}
+
+    Weight distance;     // Distance to this state from PDT "start" state.
+    SearchState parent;  // Parent state in shortest path tree.
+    int16 paren_id;      // If parent arc has paren, paren ID (or kNoLabel).
+    uint8 flags;         // First byte reserved for PdtShortestPathData use.
+  };
+
+  PdtShortestPathData(bool gc)
+      : gc_(gc), nstates_(0), ngc_(0), finished_(false) {}
+
+  ~PdtShortestPathData() {
+    VLOG(1) << "opm size: " << paren_map_.size();
+    VLOG(1) << "# of search states: " << nstates_;
+    if (gc_) VLOG(1) << "# of GC'd search states: " << ngc_;
+  }
+
+  void Clear() {
+    search_map_.clear();
+    search_multimap_.clear();
+    paren_map_.clear();
+    state_ = SearchState(kNoStateId, kNoStateId);
+    nstates_ = 0;
+    ngc_ = 0;
+  }
+
+  // TODO(kbg): Currently copying SearchState and passing a const reference to
+  // ParenSpec. Benchmark to confirm this is the right thing to do.
+
+  Weight Distance(SearchState s) const { return GetSearchData(s)->distance; }
+
+  Weight Distance(const ParenSpec &paren) const {
+    return GetSearchData(paren)->distance;
+  }
+
+  SearchState Parent(SearchState s) const { return GetSearchData(s)->parent; }
+
+  SearchState Parent(const ParenSpec &paren) const {
+    return GetSearchData(paren)->parent;
+  }
+
+  Label ParenId(SearchState s) const { return GetSearchData(s)->paren_id; }
+
+  uint8 Flags(SearchState s) const { return GetSearchData(s)->flags; }
+
+  void SetDistance(SearchState s, Weight weight) {
+    GetSearchData(s)->distance = std::move(weight);
+  }
+
+  void SetDistance(const ParenSpec &paren, Weight weight) {
+    GetSearchData(paren)->distance = std::move(weight);
+  }
+
+  void SetParent(SearchState s, SearchState p) { GetSearchData(s)->parent = p; }
+
+  void SetParent(const ParenSpec &paren, SearchState p) {
+    GetSearchData(paren)->parent = p;
+  }
+
+  void SetParenId(SearchState s, Label p) {
+    if (p >= 32768) {
+      FSTERROR() << "PdtShortestPathData: Paren ID does not fit in an int16";
+    }
+    GetSearchData(s)->paren_id = p;
+  }
+
+  void SetFlags(SearchState s, uint8 f, uint8 mask) {
+    auto *data = GetSearchData(s);
+    data->flags &= ~mask;
+    data->flags |= f & mask;
+  }
+
+  void GC(StateId s);
+
+  void Finish() { finished_ = true; }
+
+ private:
+  // Hash for search state.
+  struct SearchStateHash {
+    size_t operator()(const SearchState &s) const {
+      static constexpr auto prime = 7853;
+      return s.state + s.start * prime;
+    }
+  };
+
+  // Hash for paren map.
+  struct ParenHash {
+    size_t operator()(const ParenSpec &paren) const {
+      static constexpr auto prime0 = 7853;
+      static constexpr auto prime1 = 7867;
+      return paren.paren_id + paren.src_start * prime0 +
+             paren.dest_start * prime1;
+    }
+  };
+
+  using SearchMap =
+      std::unordered_map<SearchState, SearchData, SearchStateHash>;
+
+  using SearchMultimap = std::unordered_multimap<StateId, StateId>;
+
+  // Hash map from paren spec to open paren data.
+  using ParenMap = std::unordered_map<ParenSpec, SearchData, ParenHash>;
+
+  SearchData *GetSearchData(SearchState s) const {
+    if (s == state_) return state_data_;
+    if (finished_) {
+      auto it = search_map_.find(s);
+      if (it == search_map_.end()) return &null_search_data_;
+      state_ = s;
+      return state_data_ = &(it->second);
+    } else {
+      state_ = s;
+      state_data_ = &search_map_[s];
+      if (!(state_data_->flags & kPdtInited)) {
+        ++nstates_;
+        if (gc_) search_multimap_.insert(std::make_pair(s.start, s.state));
+        state_data_->flags = kPdtInited;
+      }
+      return state_data_;
+    }
+  }
+
+  SearchData *GetSearchData(ParenSpec paren) const {
+    if (paren == paren_) return paren_data_;
+    if (finished_) {
+      auto it = paren_map_.find(paren);
+      if (it == paren_map_.end()) return &null_search_data_;
+      paren_ = paren;
+      return state_data_ = &(it->second);
+    } else {
+      paren_ = paren;
+      return paren_data_ = &paren_map_[paren];
+    }
+  }
+
+  mutable SearchMap search_map_;            // Maps from search state to data.
+  mutable SearchMultimap search_multimap_;  // Maps from "start" to subgraph.
+  mutable ParenMap paren_map_;              // Maps paren spec to search data.
+  mutable SearchState state_;               // Last state accessed.
+  mutable SearchData *state_data_;          // Last state data accessed.
+  mutable ParenSpec paren_;                 // Last paren spec accessed.
+  mutable SearchData *paren_data_;          // Last paren data accessed.
+  bool gc_;                                 // Allow GC?
+  mutable size_t nstates_;                  // Total number of search states.
+  size_t ngc_;                              // Number of GC'd search states.
+  mutable SearchData null_search_data_;     // Null search data.
+  bool finished_;                           // Read-only access when true.
+
+  PdtShortestPathData(const PdtShortestPathData &) = delete;
+  PdtShortestPathData &operator=(const PdtShortestPathData &) = delete;
+};
+
+// Deletes inaccessible search data from a given "start" (open paren dest)
+// state. Assumes "final" (close paren source or PDT final) states have
+// been flagged kPdtFinal.
+template <class Arc>
+void PdtShortestPathData<Arc>::GC(StateId start) {
+  if (!gc_) return;
+  std::vector<StateId> finals;
+  for (auto it = search_multimap_.find(start);
+       it != search_multimap_.end() && it->first == start; ++it) {
+    const SearchState s(it->second, start);
+    if (search_map_[s].flags & kPdtFinal) finals.push_back(s.state);
+  }
+  // Mark phase.
+  for (const auto state : finals) {
+    SearchState ss(state, start);
+    while (ss.state != kNoLabel) {
+      auto &sdata = search_map_[ss];
+      if (sdata.flags & kPdtMarked) break;
+      sdata.flags |= kPdtMarked;
+      const auto p = sdata.parent;
+      if (p.start != start && p.start != kNoLabel) {  // Entering sub-subgraph.
+        const ParenSpec paren(sdata.paren_id, ss.start, p.start);
+        ss = paren_map_[paren].parent;
+      } else {
+        ss = p;
+      }
+    }
+  }
+  // Sweep phase.
+  auto it = search_multimap_.find(start);
+  while (it != search_multimap_.end() && it->first == start) {
+    const SearchState s(it->second, start);
+    auto mit = search_map_.find(s);
+    const SearchData &data = mit->second;
+    if (!(data.flags & kPdtMarked)) {
+      search_map_.erase(mit);
+      ++ngc_;
+    }
+    search_multimap_.erase(it++);
+  }
+}
+
+}  // namespace internal
+
+// This computes the single source shortest (balanced) path (SSSP) through a
+// weighted PDT that has a bounded stack (i.e., is expandable as an FST). It is
+// a generalization of the classic SSSP graph algorithm that removes a state s
+// from a queue (defined by a user-provided queue type) and relaxes the
+// destination states of transitions leaving s. In this PDT version, states that
+// have entering open parentheses are treated as source states for a sub-graph
+// SSSP problem with the shortest path up to the open parenthesis being first
+// saved. When a close parenthesis is then encountered any balancing open
+// parenthesis is examined for this saved information and multiplied back. In
+// this way, each sub-graph is entered only once rather than repeatedly. If
+// every state in the input PDT has the property that there is a unique "start"
+// state for it with entering open parentheses, then this algorithm is quite
+// straightforward. In general, this will not be the case, so the algorithm
+// (implicitly) creates a new graph where each state is a pair of an original
+// state and a possible parenthesis "start" state for that state.
+template <class Arc, class Queue>
+class PdtShortestPath {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using SpData = internal::PdtShortestPathData<Arc>;
+  using SearchState = typename SpData::SearchState;
+  using ParenSpec = typename SpData::ParenSpec;
+  using CloseSourceIterator =
+      typename internal::PdtBalanceData<Arc>::SetIterator;
+
+  PdtShortestPath(const Fst<Arc> &ifst,
+                  const std::vector<std::pair<Label, Label>> &parens,
+                  const PdtShortestPathOptions<Arc, Queue> &opts)
+      : ifst_(ifst.Copy()),
+        parens_(parens),
+        keep_parens_(opts.keep_parentheses),
+        start_(ifst.Start()),
+        sp_data_(opts.path_gc),
+        error_(false) {
+    // TODO(kbg): Make this a compile-time static_assert once:
+    // 1) All weight properties are made constexpr for all weight types.
+    // 2) We have a pleasant way to "deregister" this oepration for non-path
+    //    semirings so an informative error message is produced. The best
+    //    solution will probably involve some kind of SFINAE magic.
+    if ((Weight::Properties() & (kPath | kRightSemiring)) !=
+        (kPath | kRightSemiring)) {
+      FSTERROR() << "PdtShortestPath: Weight needs to have the path"
+                 << " property and be right distributive: " << Weight::Type();
+      error_ = true;
+    }
+    for (Label i = 0; i < parens.size(); ++i) {
+      const auto &pair = parens[i];
+      paren_map_[pair.first] = i;
+      paren_map_[pair.second] = i;
+    }
+  }
+
+  ~PdtShortestPath() {
+    VLOG(1) << "# of input states: " << CountStates(*ifst_);
+    VLOG(1) << "# of enqueued: " << nenqueued_;
+    VLOG(1) << "cpmm size: " << close_paren_multimap_.size();
+  }
+
+  void ShortestPath(MutableFst<Arc> *ofst) {
+    Init(ofst);
+    GetDistance(start_);
+    GetPath();
+    sp_data_.Finish();
+    if (error_) ofst->SetProperties(kError, kError);
+  }
+
+  const internal::PdtShortestPathData<Arc> &GetShortestPathData() const {
+    return sp_data_;
+  }
+
+  internal::PdtBalanceData<Arc> *GetBalanceData() { return &balance_data_; }
+
+ public:
+  // Hash multimap from close paren label to an paren arc.
+  using CloseParenMultimap =
+      std::unordered_multimap<internal::ParenState<Arc>, Arc,
+                              typename internal::ParenState<Arc>::Hash>;
+
+  const CloseParenMultimap &GetCloseParenMultimap() const {
+    return close_paren_multimap_;
+  }
+
+ private:
+  void Init(MutableFst<Arc> *ofst);
+
+  void GetDistance(StateId start);
+
+  void ProcFinal(SearchState s);
+
+  void ProcArcs(SearchState s);
+
+  void ProcOpenParen(Label paren_id, SearchState s, StateId nexstate,
+                     const Weight &weight);
+
+  void ProcCloseParen(Label paren_id, SearchState s, const Weight &weight);
+
+  void ProcNonParen(SearchState s, StateId nextstate, const Weight &weight);
+
+  void Relax(SearchState s, SearchState t, StateId nextstate,
+             const Weight &weight, Label paren_id);
+
+  void Enqueue(SearchState d);
+
+  void GetPath();
+
+  Arc GetPathArc(SearchState s, SearchState p, Label paren_id, bool open);
+
+  std::unique_ptr<Fst<Arc>> ifst_;
+  MutableFst<Arc> *ofst_;
+  const std::vector<std::pair<Label, Label>> &parens_;
+  bool keep_parens_;
+  Queue *state_queue_;
+  StateId start_;
+  Weight fdistance_;
+  SearchState f_parent_;
+  SpData sp_data_;
+  std::unordered_map<Label, Label> paren_map_;
+  CloseParenMultimap close_paren_multimap_;
+  internal::PdtBalanceData<Arc> balance_data_;
+  ssize_t nenqueued_;
+  bool error_;
+
+  static constexpr uint8 kEnqueued = 0x10;
+  static constexpr uint8 kExpanded = 0x20;
+  static constexpr uint8 kFinished = 0x40;
+
+  static const Arc kNoArc;
+};
+
+template <class Arc, class Queue>
+void PdtShortestPath<Arc, Queue>::Init(MutableFst<Arc> *ofst) {
+  ofst_ = ofst;
+  ofst->DeleteStates();
+  ofst->SetInputSymbols(ifst_->InputSymbols());
+  ofst->SetOutputSymbols(ifst_->OutputSymbols());
+  if (ifst_->Start() == kNoStateId) return;
+  fdistance_ = Weight::Zero();
+  f_parent_ = SearchState(kNoStateId, kNoStateId);
+  sp_data_.Clear();
+  close_paren_multimap_.clear();
+  balance_data_.Clear();
+  nenqueued_ = 0;
+  // Finds open parens per destination state and close parens per source state.
+  for (StateIterator<Fst<Arc>> siter(*ifst_); !siter.Done(); siter.Next()) {
+    const auto s = siter.Value();
+    for (ArcIterator<Fst<Arc>> aiter(*ifst_, s); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      const auto it = paren_map_.find(arc.ilabel);
+      if (it != paren_map_.end()) {  // Is a paren?
+        const auto paren_id = it->second;
+        if (arc.ilabel == parens_[paren_id].first) {  // Open paren.
+          balance_data_.OpenInsert(paren_id, arc.nextstate);
+        } else {  // Close paren.
+          const internal::ParenState<Arc> paren_state(paren_id, s);
+          close_paren_multimap_.emplace(paren_state, arc);
+        }
+      }
+    }
+  }
+}
+
+// Computes the shortest distance stored in a recursive way. Each sub-graph
+// (i.e., different paren "start" state) begins with weight One().
+template <class Arc, class Queue>
+void PdtShortestPath<Arc, Queue>::GetDistance(StateId start) {
+  if (start == kNoStateId) return;
+  Queue state_queue;
+  state_queue_ = &state_queue;
+  const SearchState q(start, start);
+  Enqueue(q);
+  sp_data_.SetDistance(q, Weight::One());
+  while (!state_queue_->Empty()) {
+    const auto state = state_queue_->Head();
+    state_queue_->Dequeue();
+    const SearchState s(state, start);
+    sp_data_.SetFlags(s, 0, kEnqueued);
+    ProcFinal(s);
+    ProcArcs(s);
+    sp_data_.SetFlags(s, kExpanded, kExpanded);
+  }
+  sp_data_.SetFlags(q, kFinished, kFinished);
+  balance_data_.FinishInsert(start);
+  sp_data_.GC(start);
+}
+
+// Updates best complete path.
+template <class Arc, class Queue>
+void PdtShortestPath<Arc, Queue>::ProcFinal(SearchState s) {
+  if (ifst_->Final(s.state) != Weight::Zero() && s.start == start_) {
+    const auto weight = Times(sp_data_.Distance(s), ifst_->Final(s.state));
+    if (fdistance_ != Plus(fdistance_, weight)) {
+      if (f_parent_.state != kNoStateId) {
+        sp_data_.SetFlags(f_parent_, 0, internal::kPdtFinal);
+      }
+      sp_data_.SetFlags(s, internal::kPdtFinal, internal::kPdtFinal);
+      fdistance_ = Plus(fdistance_, weight);
+      f_parent_ = s;
+    }
+  }
+}
+
+// Processes all arcs leaving the state s.
+template <class Arc, class Queue>
+void PdtShortestPath<Arc, Queue>::ProcArcs(SearchState s) {
+  for (ArcIterator<Fst<Arc>> aiter(*ifst_, s.state); !aiter.Done();
+       aiter.Next()) {
+    const auto &arc = aiter.Value();
+    const auto weight = Times(sp_data_.Distance(s), arc.weight);
+    const auto it = paren_map_.find(arc.ilabel);
+    if (it != paren_map_.end()) {  // Is a paren?
+      const auto paren_id = it->second;
+      if (arc.ilabel == parens_[paren_id].first) {
+        ProcOpenParen(paren_id, s, arc.nextstate, weight);
+      } else {
+        ProcCloseParen(paren_id, s, weight);
+      }
+    } else {
+      ProcNonParen(s, arc.nextstate, weight);
+    }
+  }
+}
+
+// Saves the shortest path info for reaching this parenthesis and starts a new
+// SSSP in the sub-graph pointed to by the parenthesis if previously unvisited.
+// Otherwise it finds any previously encountered closing parentheses and relaxes
+// them using the recursively stored shortest distance to them.
+template <class Arc, class Queue>
+inline void PdtShortestPath<Arc, Queue>::ProcOpenParen(Label paren_id,
+                                                       SearchState s,
+                                                       StateId nextstate,
+                                                       const Weight &weight) {
+  const SearchState d(nextstate, nextstate);
+  const ParenSpec paren(paren_id, s.start, d.start);
+  const auto pdist = sp_data_.Distance(paren);
+  if (pdist != Plus(pdist, weight)) {
+    sp_data_.SetDistance(paren, weight);
+    sp_data_.SetParent(paren, s);
+    const auto dist = sp_data_.Distance(d);
+    if (dist == Weight::Zero()) {
+      auto *state_queue = state_queue_;
+      GetDistance(d.start);
+      state_queue_ = state_queue;
+    } else if (!(sp_data_.Flags(d) & kFinished)) {
+      FSTERROR()
+          << "PdtShortestPath: open parenthesis recursion: not bounded stack";
+      error_ = true;
+    }
+    for (auto set_iter = balance_data_.Find(paren_id, nextstate);
+         !set_iter.Done(); set_iter.Next()) {
+      const SearchState cpstate(set_iter.Element(), d.start);
+      const internal::ParenState<Arc> paren_state(paren_id, cpstate.state);
+      for (auto cpit = close_paren_multimap_.find(paren_state);
+           cpit != close_paren_multimap_.end() && paren_state == cpit->first;
+           ++cpit) {
+        const auto &cparc = cpit->second;
+        const auto cpw =
+            Times(weight, Times(sp_data_.Distance(cpstate), cparc.weight));
+        Relax(cpstate, s, cparc.nextstate, cpw, paren_id);
+      }
+    }
+  }
+}
+
+// Saves the correspondence between each closing parenthesis and its balancing
+// open parenthesis info. Relaxes any close parenthesis destination state that
+// has a balancing previously encountered open parenthesis.
+template <class Arc, class Queue>
+inline void PdtShortestPath<Arc, Queue>::ProcCloseParen(Label paren_id,
+                                                        SearchState s,
+                                                        const Weight &weight) {
+  const internal::ParenState<Arc> paren_state(paren_id, s.start);
+  if (!(sp_data_.Flags(s) & kExpanded)) {
+    balance_data_.CloseInsert(paren_id, s.start, s.state);
+    sp_data_.SetFlags(s, internal::kPdtFinal, internal::kPdtFinal);
+  }
+}
+
+// Classical relaxation for non-parentheses.
+template <class Arc, class Queue>
+inline void PdtShortestPath<Arc, Queue>::ProcNonParen(SearchState s,
+                                                      StateId nextstate,
+                                                      const Weight &weight) {
+  Relax(s, s, nextstate, weight, kNoLabel);
+}
+
+// Classical relaxation on the search graph for an arc with destination state
+// nexstate from state s. State t is in the same sub-graph as nextstate (i.e.,
+// has the same paren "start").
+template <class Arc, class Queue>
+inline void PdtShortestPath<Arc, Queue>::Relax(SearchState s, SearchState t,
+                                               StateId nextstate,
+                                               const Weight &weight,
+                                               Label paren_id) {
+  const SearchState d(nextstate, t.start);
+  Weight dist = sp_data_.Distance(d);
+  if (dist != Plus(dist, weight)) {
+    sp_data_.SetParent(d, s);
+    sp_data_.SetParenId(d, paren_id);
+    sp_data_.SetDistance(d, Plus(dist, weight));
+    Enqueue(d);
+  }
+}
+
+template <class Arc, class Queue>
+inline void PdtShortestPath<Arc, Queue>::Enqueue(SearchState s) {
+  if (!(sp_data_.Flags(s) & kEnqueued)) {
+    state_queue_->Enqueue(s.state);
+    sp_data_.SetFlags(s, kEnqueued, kEnqueued);
+    ++nenqueued_;
+  } else {
+    state_queue_->Update(s.state);
+  }
+}
+
+// Follows parent pointers to find the shortest path. A stack is used since the
+// shortest distance is stored recursively.
+template <class Arc, class Queue>
+void PdtShortestPath<Arc, Queue>::GetPath() {
+  SearchState s = f_parent_;
+  SearchState d = SearchState(kNoStateId, kNoStateId);
+  StateId s_p = kNoStateId;
+  StateId d_p = kNoStateId;
+  auto arc = kNoArc;
+  Label paren_id = kNoLabel;
+  std::stack<ParenSpec> paren_stack;
+  while (s.state != kNoStateId) {
+    d_p = s_p;
+    s_p = ofst_->AddState();
+    if (d.state == kNoStateId) {
+      ofst_->SetFinal(s_p, ifst_->Final(f_parent_.state));
+    } else {
+      if (paren_id != kNoLabel) {                     // Paren?
+        if (arc.ilabel == parens_[paren_id].first) {  // Open paren?
+          paren_stack.pop();
+        } else {  // Close paren?
+          const ParenSpec paren(paren_id, d.start, s.start);
+          paren_stack.push(paren);
+        }
+        if (!keep_parens_) arc.ilabel = arc.olabel = 0;
+      }
+      arc.nextstate = d_p;
+      ofst_->AddArc(s_p, arc);
+    }
+    d = s;
+    s = sp_data_.Parent(d);
+    paren_id = sp_data_.ParenId(d);
+    if (s.state != kNoStateId) {
+      arc = GetPathArc(s, d, paren_id, false);
+    } else if (!paren_stack.empty()) {
+      const ParenSpec paren = paren_stack.top();
+      s = sp_data_.Parent(paren);
+      paren_id = paren.paren_id;
+      arc = GetPathArc(s, d, paren_id, true);
+    }
+  }
+  ofst_->SetStart(s_p);
+  ofst_->SetProperties(
+      ShortestPathProperties(ofst_->Properties(kFstProperties, false)),
+      kFstProperties);
+}
+
+// Finds transition with least weight between two states with label matching
+// paren_id and open/close paren type or a non-paren if kNoLabel.
+template <class Arc, class Queue>
+Arc PdtShortestPath<Arc, Queue>::GetPathArc(SearchState s, SearchState d,
+                                            Label paren_id, bool open_paren) {
+  auto path_arc = kNoArc;
+  for (ArcIterator<Fst<Arc>> aiter(*ifst_, s.state); !aiter.Done();
+       aiter.Next()) {
+    const auto &arc = aiter.Value();
+    if (arc.nextstate != d.state) continue;
+    Label arc_paren_id = kNoLabel;
+    const auto it = paren_map_.find(arc.ilabel);
+    if (it != paren_map_.end()) {
+      arc_paren_id = it->second;
+      bool arc_open_paren = (arc.ilabel == parens_[arc_paren_id].first);
+      if (arc_open_paren != open_paren) continue;
+    }
+    if (arc_paren_id != paren_id) continue;
+    if (arc.weight == Plus(arc.weight, path_arc.weight)) path_arc = arc;
+  }
+  if (path_arc.nextstate == kNoStateId) {
+    FSTERROR() << "PdtShortestPath::GetPathArc: Failed to find arc";
+    error_ = true;
+  }
+  return path_arc;
+}
+
+template <class Arc, class Queue>
+const Arc PdtShortestPath<Arc, Queue>::kNoArc = Arc(kNoLabel, kNoLabel,
+                                                    Weight::Zero(), kNoStateId);
+
+// Functional variants.
+
+template <class Arc, class Queue>
+void ShortestPath(
+    const Fst<Arc> &ifst,
+    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
+        &parens,
+    MutableFst<Arc> *ofst, const PdtShortestPathOptions<Arc, Queue> &opts) {
+  PdtShortestPath<Arc, Queue> psp(ifst, parens, opts);
+  psp.ShortestPath(ofst);
+}
+
+template <class Arc>
+void ShortestPath(
+    const Fst<Arc> &ifst,
+    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
+        &parens,
+    MutableFst<Arc> *ofst) {
+  using Q = FifoQueue<typename Arc::StateId>;
+  const PdtShortestPathOptions<Arc, Q> opts;
+  PdtShortestPath<Arc, Q> psp(ifst, parens, opts);
+  psp.ShortestPath(ofst);
+}
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_PDT_SHORTEST_PATH_H_
diff --git a/projects/llm_framework/include/fst/extensions/special/phi-fst.h b/projects/llm_framework/include/fst/extensions/special/phi-fst.h
new file mode 100644
index 00000000..db40d611
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/special/phi-fst.h
@@ -0,0 +1,183 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_EXTENSIONS_SPECIAL_PHI_FST_H_
+#define FST_EXTENSIONS_SPECIAL_PHI_FST_H_
+
+#include <memory>
+#include <string>
+
+#include <fst/const-fst.h>
+#include <fst/matcher-fst.h>
+#include <fst/matcher.h>
+
+DECLARE_int64(phi_fst_phi_label);
+DECLARE_bool(phi_fst_phi_loop);
+DECLARE_string(phi_fst_rewrite_mode);
+
+namespace fst {
+namespace internal {
+
+template <class Label>
+class PhiFstMatcherData {
+ public:
+  PhiFstMatcherData(
+      Label phi_label = FLAGS_phi_fst_phi_label,
+      bool phi_loop = FLAGS_phi_fst_phi_loop,
+      MatcherRewriteMode rewrite_mode = RewriteMode(FLAGS_phi_fst_rewrite_mode))
+      : phi_label_(phi_label),
+        phi_loop_(phi_loop),
+        rewrite_mode_(rewrite_mode) {}
+
+  PhiFstMatcherData(const PhiFstMatcherData &data)
+      : phi_label_(data.phi_label_),
+        phi_loop_(data.phi_loop_),
+        rewrite_mode_(data.rewrite_mode_) {}
+
+  static PhiFstMatcherData<Label> *Read(std::istream &istrm,
+                                        const FstReadOptions &read) {
+    auto *data = new PhiFstMatcherData<Label>();
+    ReadType(istrm, &data->phi_label_);
+    ReadType(istrm, &data->phi_loop_);
+    int32 rewrite_mode;
+    ReadType(istrm, &rewrite_mode);
+    data->rewrite_mode_ = static_cast<MatcherRewriteMode>(rewrite_mode);
+    return data;
+  }
+
+  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
+    WriteType(ostrm, phi_label_);
+    WriteType(ostrm, phi_loop_);
+    WriteType(ostrm, static_cast<int32>(rewrite_mode_));
+    return !ostrm ? false : true;
+  }
+
+  Label PhiLabel() const { return phi_label_; }
+
+  bool PhiLoop() const { return phi_loop_; }
+
+  MatcherRewriteMode RewriteMode() const { return rewrite_mode_; }
+
+ private:
+  static MatcherRewriteMode RewriteMode(const string &mode) {
+    if (mode == "auto") return MATCHER_REWRITE_AUTO;
+    if (mode == "always") return MATCHER_REWRITE_ALWAYS;
+    if (mode == "never") return MATCHER_REWRITE_NEVER;
+    LOG(WARNING) << "PhiFst: Unknown rewrite mode: " << mode << ". "
+                 << "Defaulting to auto.";
+    return MATCHER_REWRITE_AUTO;
+  }
+
+  Label phi_label_;
+  bool phi_loop_;
+  MatcherRewriteMode rewrite_mode_;
+};
+
+}  // namespace internal
+
+constexpr uint8 kPhiFstMatchInput = 0x01;   // Input matcher is PhiMatcher.
+constexpr uint8 kPhiFstMatchOutput = 0x02;  // Output matcher is PhiMatcher.
+
+template <class M, uint8 flags = kPhiFstMatchInput | kPhiFstMatchOutput>
+class PhiFstMatcher : public PhiMatcher<M> {
+ public:
+  using FST = typename M::FST;
+  using Arc = typename M::Arc;
+  using StateId = typename Arc::StateId;
+  using Label = typename Arc::Label;
+  using Weight = typename Arc::Weight;
+  using MatcherData = internal::PhiFstMatcherData<Label>;
+
+  enum : uint8 { kFlags = flags };
+
+  // This makes a copy of the FST.
+  PhiFstMatcher(const FST &fst, MatchType match_type,
+      std::shared_ptr<MatcherData> data = std::make_shared<MatcherData>())
+      : PhiMatcher<M>(fst, match_type,
+                      PhiLabel(match_type, data ? data->PhiLabel()
+                                                : MatcherData().PhiLabel()),
+                      data ? data->PhiLoop() : MatcherData().PhiLoop(),
+                      data ? data->RewriteMode() : MatcherData().RewriteMode()),
+        data_(data) {}
+
+  // This doesn't copy the FST.
+  PhiFstMatcher(const FST *fst, MatchType match_type,
+      std::shared_ptr<MatcherData> data = std::make_shared<MatcherData>())
+      : PhiMatcher<M>(fst, match_type,
+                      PhiLabel(match_type, data ? data->PhiLabel()
+                                                : MatcherData().PhiLabel()),
+                      data ? data->PhiLoop() : MatcherData().PhiLoop(),
+                      data ? data->RewriteMode() : MatcherData().RewriteMode()),
+        data_(data) {}
+
+  // This makes a copy of the FST.
+  PhiFstMatcher(const PhiFstMatcher<M, flags> &matcher, bool safe = false)
+      : PhiMatcher<M>(matcher, safe), data_(matcher.data_) {}
+
+  PhiFstMatcher<M, flags> *Copy(bool safe = false) const override {
+    return new PhiFstMatcher<M, flags>(*this, safe);
+  }
+
+  const MatcherData *GetData() const { return data_.get(); }
+
+  std::shared_ptr<MatcherData> GetSharedData() const { return data_; }
+
+ private:
+  static Label PhiLabel(MatchType match_type, Label label) {
+    if (match_type == MATCH_INPUT && flags & kPhiFstMatchInput) return label;
+    if (match_type == MATCH_OUTPUT && flags & kPhiFstMatchOutput) return label;
+    return kNoLabel;
+  }
+
+  std::shared_ptr<MatcherData> data_;
+};
+
+extern const char phi_fst_type[];
+extern const char input_phi_fst_type[];
+extern const char output_phi_fst_type[];
+
+using StdPhiFst =
+    MatcherFst<ConstFst<StdArc>, PhiFstMatcher<SortedMatcher<ConstFst<StdArc>>>,
+               phi_fst_type>;
+
+using LogPhiFst =
+    MatcherFst<ConstFst<LogArc>, PhiFstMatcher<SortedMatcher<ConstFst<LogArc>>>,
+               phi_fst_type>;
+
+using Log64PhiFst = MatcherFst<ConstFst<Log64Arc>,
+                               PhiFstMatcher<SortedMatcher<ConstFst<Log64Arc>>>,
+                               input_phi_fst_type>;
+
+using StdInputPhiFst =
+    MatcherFst<ConstFst<StdArc>, PhiFstMatcher<SortedMatcher<ConstFst<StdArc>>,
+                                               kPhiFstMatchInput>,
+               input_phi_fst_type>;
+
+using LogInputPhiFst =
+    MatcherFst<ConstFst<LogArc>, PhiFstMatcher<SortedMatcher<ConstFst<LogArc>>,
+                                               kPhiFstMatchInput>,
+               input_phi_fst_type>;
+
+using Log64InputPhiFst = MatcherFst<
+    ConstFst<Log64Arc>,
+    PhiFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kPhiFstMatchInput>,
+    input_phi_fst_type>;
+
+using StdOutputPhiFst =
+    MatcherFst<ConstFst<StdArc>, PhiFstMatcher<SortedMatcher<ConstFst<StdArc>>,
+                                               kPhiFstMatchOutput>,
+               output_phi_fst_type>;
+
+using LogOutputPhiFst =
+    MatcherFst<ConstFst<LogArc>, PhiFstMatcher<SortedMatcher<ConstFst<LogArc>>,
+                                               kPhiFstMatchOutput>,
+               output_phi_fst_type>;
+
+using Log64OutputPhiFst = MatcherFst<
+    ConstFst<Log64Arc>,
+    PhiFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kPhiFstMatchOutput>,
+    output_phi_fst_type>;
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_SPECIAL_PHI_FST_H_
diff --git a/projects/llm_framework/include/fst/extensions/special/rho-fst.h b/projects/llm_framework/include/fst/extensions/special/rho-fst.h
new file mode 100644
index 00000000..ababecc0
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/special/rho-fst.h
@@ -0,0 +1,172 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_EXTENSIONS_SPECIAL_RHO_FST_H_
+#define FST_EXTENSIONS_SPECIAL_RHO_FST_H_
+
+#include <memory>
+#include <string>
+
+#include <fst/const-fst.h>
+#include <fst/matcher-fst.h>
+#include <fst/matcher.h>
+
+DECLARE_int64(rho_fst_rho_label);
+DECLARE_string(rho_fst_rewrite_mode);
+
+namespace fst {
+namespace internal {
+
+template <class Label>
+class RhoFstMatcherData {
+ public:
+  explicit RhoFstMatcherData(
+      Label rho_label = FLAGS_rho_fst_rho_label,
+      MatcherRewriteMode rewrite_mode = RewriteMode(FLAGS_rho_fst_rewrite_mode))
+      : rho_label_(rho_label), rewrite_mode_(rewrite_mode) {}
+
+  RhoFstMatcherData(const RhoFstMatcherData &data)
+      : rho_label_(data.rho_label_), rewrite_mode_(data.rewrite_mode_) {}
+
+  static RhoFstMatcherData<Label> *Read(std::istream &istrm,
+                                    const FstReadOptions &read) {
+    auto *data = new RhoFstMatcherData<Label>();
+    ReadType(istrm, &data->rho_label_);
+    int32 rewrite_mode;
+    ReadType(istrm, &rewrite_mode);
+    data->rewrite_mode_ = static_cast<MatcherRewriteMode>(rewrite_mode);
+    return data;
+  }
+
+  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
+    WriteType(ostrm, rho_label_);
+    WriteType(ostrm, static_cast<int32>(rewrite_mode_));
+    return !ostrm ? false : true;
+  }
+
+  Label RhoLabel() const { return rho_label_; }
+
+  MatcherRewriteMode RewriteMode() const { return rewrite_mode_; }
+
+ private:
+  static MatcherRewriteMode RewriteMode(const string &mode) {
+    if (mode == "auto") return MATCHER_REWRITE_AUTO;
+    if (mode == "always") return MATCHER_REWRITE_ALWAYS;
+    if (mode == "never") return MATCHER_REWRITE_NEVER;
+    LOG(WARNING) << "RhoFst: Unknown rewrite mode: " << mode << ". "
+                 << "Defaulting to auto.";
+    return MATCHER_REWRITE_AUTO;
+  }
+
+  Label rho_label_;
+  MatcherRewriteMode rewrite_mode_;
+};
+
+}  // namespace internal
+
+constexpr uint8 kRhoFstMatchInput = 0x01;   // Input matcher is RhoMatcher.
+constexpr uint8 kRhoFstMatchOutput = 0x02;  // Output matcher is RhoMatcher.
+
+template <class M, uint8 flags = kRhoFstMatchInput | kRhoFstMatchOutput>
+class RhoFstMatcher : public RhoMatcher<M> {
+ public:
+  using FST = typename M::FST;
+  using Arc = typename M::Arc;
+  using StateId = typename Arc::StateId;
+  using Label = typename Arc::Label;
+  using Weight = typename Arc::Weight;
+  using MatcherData = internal::RhoFstMatcherData<Label>;
+
+  enum : uint8 { kFlags = flags };
+
+  // This makes a copy of the FST.
+  RhoFstMatcher(
+      const FST &fst, MatchType match_type,
+      std::shared_ptr<MatcherData> data = std::make_shared<MatcherData>())
+      : RhoMatcher<M>(fst, match_type,
+                      RhoLabel(match_type, data ? data->RhoLabel()
+                                                : MatcherData().RhoLabel()),
+                      data ? data->RewriteMode() : MatcherData().RewriteMode()),
+        data_(data) {}
+
+  // This doesn't copy the FST.
+  RhoFstMatcher(
+      const FST *fst, MatchType match_type,
+      std::shared_ptr<MatcherData> data = std::make_shared<MatcherData>())
+      : RhoMatcher<M>(fst, match_type,
+                      RhoLabel(match_type, data ? data->RhoLabel()
+                                                : MatcherData().RhoLabel()),
+                      data ? data->RewriteMode() : MatcherData().RewriteMode()),
+        data_(data) {}
+
+  // This makes a copy of the FST.
+  RhoFstMatcher(const RhoFstMatcher<M, flags> &matcher, bool safe = false)
+      : RhoMatcher<M>(matcher, safe), data_(matcher.data_) {}
+
+  RhoFstMatcher<M, flags> *Copy(bool safe = false) const override {
+    return new RhoFstMatcher<M, flags>(*this, safe);
+  }
+
+  const MatcherData *GetData() const { return data_.get(); }
+
+  std::shared_ptr<MatcherData> GetSharedData() const { return data_; }
+
+ private:
+  static Label RhoLabel(MatchType match_type, Label label) {
+    if (match_type == MATCH_INPUT && flags & kRhoFstMatchInput) return label;
+    if (match_type == MATCH_OUTPUT && flags & kRhoFstMatchOutput) return label;
+    return kNoLabel;
+  }
+
+  std::shared_ptr<MatcherData> data_;
+};
+
+extern const char rho_fst_type[];
+extern const char input_rho_fst_type[];
+extern const char output_rho_fst_type[];
+
+using StdRhoFst =
+    MatcherFst<ConstFst<StdArc>, RhoFstMatcher<SortedMatcher<ConstFst<StdArc>>>,
+               rho_fst_type>;
+
+using LogRhoFst =
+    MatcherFst<ConstFst<LogArc>, RhoFstMatcher<SortedMatcher<ConstFst<LogArc>>>,
+               rho_fst_type>;
+
+using Log64RhoFst = MatcherFst<ConstFst<Log64Arc>,
+                               RhoFstMatcher<SortedMatcher<ConstFst<Log64Arc>>>,
+                               input_rho_fst_type>;
+
+using StdInputRhoFst =
+    MatcherFst<ConstFst<StdArc>, RhoFstMatcher<SortedMatcher<ConstFst<StdArc>>,
+                                               kRhoFstMatchInput>,
+               input_rho_fst_type>;
+
+using LogInputRhoFst =
+    MatcherFst<ConstFst<LogArc>, RhoFstMatcher<SortedMatcher<ConstFst<LogArc>>,
+                                               kRhoFstMatchInput>,
+               input_rho_fst_type>;
+
+using Log64InputRhoFst = MatcherFst<
+    ConstFst<Log64Arc>,
+    RhoFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kRhoFstMatchInput>,
+    input_rho_fst_type>;
+
+using StdOutputRhoFst =
+    MatcherFst<ConstFst<StdArc>, RhoFstMatcher<SortedMatcher<ConstFst<StdArc>>,
+                                               kRhoFstMatchOutput>,
+               output_rho_fst_type>;
+
+using LogOutputRhoFst =
+    MatcherFst<ConstFst<LogArc>, RhoFstMatcher<SortedMatcher<ConstFst<LogArc>>,
+                                               kRhoFstMatchOutput>,
+               output_rho_fst_type>;
+
+using Log64OutputRhoFst = MatcherFst<
+    ConstFst<Log64Arc>,
+    RhoFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kRhoFstMatchOutput>,
+    output_rho_fst_type>;
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_SPECIAL_RHO_FST_H_
diff --git a/projects/llm_framework/include/fst/extensions/special/sigma-fst.h b/projects/llm_framework/include/fst/extensions/special/sigma-fst.h
new file mode 100644
index 00000000..0d6cfd03
--- /dev/null
+++ b/projects/llm_framework/include/fst/extensions/special/sigma-fst.h
@@ -0,0 +1,176 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_EXTENSIONS_SPECIAL_SIGMA_FST_H_
+#define FST_EXTENSIONS_SPECIAL_SIGMA_FST_H_
+
+#include <memory>
+#include <string>
+
+#include <fst/const-fst.h>
+#include <fst/matcher-fst.h>
+#include <fst/matcher.h>
+
+DECLARE_int64(sigma_fst_sigma_label);
+DECLARE_string(sigma_fst_rewrite_mode);
+
+namespace fst {
+namespace internal {
+
+template <class Label>
+class SigmaFstMatcherData {
+ public:
+  explicit SigmaFstMatcherData(Label sigma_label = FLAGS_sigma_fst_sigma_label,
+      MatcherRewriteMode rewrite_mode =
+                          RewriteMode(FLAGS_sigma_fst_rewrite_mode))
+      : sigma_label_(sigma_label), rewrite_mode_(rewrite_mode) {}
+
+  SigmaFstMatcherData(const SigmaFstMatcherData &data)
+      : sigma_label_(data.sigma_label_), rewrite_mode_(data.rewrite_mode_) {}
+
+  static SigmaFstMatcherData<Label> *Read(std::istream &istrm,
+                                      const FstReadOptions &read) {
+    auto *data = new SigmaFstMatcherData<Label>();
+    ReadType(istrm, &data->sigma_label_);
+    int32 rewrite_mode;
+    ReadType(istrm, &rewrite_mode);
+    data->rewrite_mode_ = static_cast<MatcherRewriteMode>(rewrite_mode);
+    return data;
+  }
+
+  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
+    WriteType(ostrm, sigma_label_);
+    WriteType(ostrm, static_cast<int32>(rewrite_mode_));
+    return !ostrm ? false : true;
+  }
+
+  Label SigmaLabel() const { return sigma_label_; }
+
+  MatcherRewriteMode RewriteMode() const { return rewrite_mode_; }
+
+ private:
+  static MatcherRewriteMode RewriteMode(const string &mode) {
+    if (mode == "auto") return MATCHER_REWRITE_AUTO;
+    if (mode == "always") return MATCHER_REWRITE_ALWAYS;
+    if (mode == "never") return MATCHER_REWRITE_NEVER;
+    LOG(WARNING) << "SigmaFst: Unknown rewrite mode: " << mode << ". "
+                 << "Defaulting to auto.";
+    return MATCHER_REWRITE_AUTO;
+  }
+
+  Label sigma_label_;
+  MatcherRewriteMode rewrite_mode_;
+};
+
+}  // namespace internal
+
+constexpr uint8 kSigmaFstMatchInput = 0x01;   // Input matcher is SigmaMatcher.
+constexpr uint8 kSigmaFstMatchOutput = 0x02;  // Output matcher is SigmaMatcher.
+
+template <class M, uint8 flags = kSigmaFstMatchInput | kSigmaFstMatchOutput>
+class SigmaFstMatcher : public SigmaMatcher<M> {
+ public:
+  using FST = typename M::FST;
+  using Arc = typename M::Arc;
+  using StateId = typename Arc::StateId;
+  using Label = typename Arc::Label;
+  using Weight = typename Arc::Weight;
+  using MatcherData = internal::SigmaFstMatcherData<Label>;
+
+  enum : uint8 { kFlags = flags };
+
+  // This makes a copy of the FST.
+  SigmaFstMatcher(
+      const FST &fst, MatchType match_type,
+      std::shared_ptr<MatcherData> data = std::make_shared<MatcherData>())
+      : SigmaMatcher<M>(
+            fst, match_type,
+            SigmaLabel(match_type,
+                       data ? data->SigmaLabel() : MatcherData().SigmaLabel()),
+            data ? data->RewriteMode() : MatcherData().RewriteMode()),
+        data_(data) {}
+
+  // This doesn't copy the FST.
+  SigmaFstMatcher(
+      const FST *fst, MatchType match_type,
+      std::shared_ptr<MatcherData> data = std::make_shared<MatcherData>())
+      : SigmaMatcher<M>(
+            fst, match_type,
+            SigmaLabel(match_type,
+                       data ? data->SigmaLabel() : MatcherData().SigmaLabel()),
+            data ? data->RewriteMode() : MatcherData().RewriteMode()),
+        data_(data) {}
+
+  // This makes a copy of the FST.
+  SigmaFstMatcher(const SigmaFstMatcher<M, flags> &matcher, bool safe = false)
+      : SigmaMatcher<M>(matcher, safe), data_(matcher.data_) {}
+
+  SigmaFstMatcher<M, flags> *Copy(bool safe = false) const override {
+    return new SigmaFstMatcher<M, flags>(*this, safe);
+  }
+
+  const MatcherData *GetData() const { return data_.get(); }
+
+  std::shared_ptr<MatcherData> GetSharedData() const { return data_; }
+
+ private:
+  static Label SigmaLabel(MatchType match_type, Label label) {
+    if (match_type == MATCH_INPUT && flags & kSigmaFstMatchInput) return label;
+    if (match_type == MATCH_OUTPUT && flags & kSigmaFstMatchOutput)
+      return label;
+    return kNoLabel;
+  }
+
+  std::shared_ptr<MatcherData> data_;
+};
+
+extern const char sigma_fst_type[];
+extern const char input_sigma_fst_type[];
+extern const char output_sigma_fst_type[];
+
+using StdSigmaFst = MatcherFst<ConstFst<StdArc>,
+                               SigmaFstMatcher<SortedMatcher<ConstFst<StdArc>>>,
+                               sigma_fst_type>;
+
+using LogSigmaFst = MatcherFst<ConstFst<LogArc>,
+                               SigmaFstMatcher<SortedMatcher<ConstFst<LogArc>>>,
+                               sigma_fst_type>;
+
+using Log64SigmaFst =
+    MatcherFst<ConstFst<Log64Arc>,
+               SigmaFstMatcher<SortedMatcher<ConstFst<Log64Arc>>>,
+               input_sigma_fst_type>;
+
+using StdInputSigmaFst = MatcherFst<
+    ConstFst<StdArc>,
+    SigmaFstMatcher<SortedMatcher<ConstFst<StdArc>>, kSigmaFstMatchInput>,
+    input_sigma_fst_type>;
+
+using LogInputSigmaFst = MatcherFst<
+    ConstFst<LogArc>,
+    SigmaFstMatcher<SortedMatcher<ConstFst<LogArc>>, kSigmaFstMatchInput>,
+    input_sigma_fst_type>;
+
+using Log64InputSigmaFst = MatcherFst<
+    ConstFst<Log64Arc>,
+    SigmaFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kSigmaFstMatchInput>,
+    input_sigma_fst_type>;
+
+using StdOutputSigmaFst = MatcherFst<
+    ConstFst<StdArc>,
+    SigmaFstMatcher<SortedMatcher<ConstFst<StdArc>>, kSigmaFstMatchOutput>,
+    output_sigma_fst_type>;
+
+using LogOutputSigmaFst = MatcherFst<
+    ConstFst<LogArc>,
+    SigmaFstMatcher<SortedMatcher<ConstFst<LogArc>>, kSigmaFstMatchOutput>,
+    output_sigma_fst_type>;
+
+using Log64OutputSigmaFst = MatcherFst<
+    ConstFst<Log64Arc>,
+    SigmaFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kSigmaFstMatchOutput>,
+    output_sigma_fst_type>;
+
+}  // namespace fst
+
+#endif  // FST_EXTENSIONS_SPECIAL_SIGMA_FST_H_
diff --git a/projects/llm_framework/include/fst/factor-weight.h b/projects/llm_framework/include/fst/factor-weight.h
new file mode 100644
index 00000000..30e02320
--- /dev/null
+++ b/projects/llm_framework/include/fst/factor-weight.h
@@ -0,0 +1,496 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Classes to factor weights in an FST.
+
+#ifndef FST_FACTOR_WEIGHT_H_
+#define FST_FACTOR_WEIGHT_H_
+
+#include <algorithm>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/cache.h>
+#include <fst/test-properties.h>
+
+
+namespace fst {
+
+constexpr uint32 kFactorFinalWeights = 0x00000001;
+constexpr uint32 kFactorArcWeights = 0x00000002;
+
+template <class Arc>
+struct FactorWeightOptions : CacheOptions {
+  using Label = typename Arc::Label;
+
+  float delta;
+  uint32 mode;         // Factor arc weights and/or final weights.
+  Label final_ilabel;  // Input label of arc when factoring final weights.
+  Label final_olabel;  // Output label of arc when factoring final weights.
+  bool increment_final_ilabel;  // When factoring final w' results in > 1 arcs
+  bool increment_final_olabel;  // at state, increment labels to make distinct?
+
+  explicit FactorWeightOptions(const CacheOptions &opts, float delta = kDelta,
+                               uint32 mode = kFactorArcWeights |
+                                             kFactorFinalWeights,
+                               Label final_ilabel = 0, Label final_olabel = 0,
+                               bool increment_final_ilabel = false,
+                               bool increment_final_olabel = false)
+      : CacheOptions(opts),
+        delta(delta),
+        mode(mode),
+        final_ilabel(final_ilabel),
+        final_olabel(final_olabel),
+        increment_final_ilabel(increment_final_ilabel),
+        increment_final_olabel(increment_final_olabel) {}
+
+  explicit FactorWeightOptions(float delta = kDelta,
+                               uint32 mode = kFactorArcWeights |
+                                             kFactorFinalWeights,
+                               Label final_ilabel = 0, Label final_olabel = 0,
+                               bool increment_final_ilabel = false,
+                               bool increment_final_olabel = false)
+      : delta(delta),
+        mode(mode),
+        final_ilabel(final_ilabel),
+        final_olabel(final_olabel),
+        increment_final_ilabel(increment_final_ilabel),
+        increment_final_olabel(increment_final_olabel) {}
+};
+
+// A factor iterator takes as argument a weight w and returns a sequence of
+// pairs of weights (xi, yi) such that the sum of the products xi times yi is
+// equal to w. If w is fully factored, the iterator should return nothing.
+//
+// template <class W>
+// class FactorIterator {
+//  public:
+//   explicit FactorIterator(W w);
+//
+//   bool Done() const;
+//
+//   void Next();
+//
+//   std::pair<W, W> Value() const;
+//
+//   void Reset();
+// }
+
+// Factors trivially.
+template <class W>
+class IdentityFactor {
+ public:
+  explicit IdentityFactor(const W &weight) {}
+
+  bool Done() const { return true; }
+
+  void Next() {}
+
+  std::pair<W, W> Value() const { return std::make_pair(W::One(), W::One()); }
+
+  void Reset() {}
+};
+
+// Factors a StringWeight w as 'ab' where 'a' is a label.
+template <typename Label, StringType S = STRING_LEFT>
+class StringFactor {
+ public:
+  explicit StringFactor(const StringWeight<Label, S> &weight)
+      : weight_(weight), done_(weight.Size() <= 1) {}
+
+  bool Done() const { return done_; }
+
+  void Next() { done_ = true; }
+
+  std::pair<StringWeight<Label, S>, StringWeight<Label, S>> Value() const {
+    using Weight = StringWeight<Label, S>;
+    typename Weight::Iterator siter(weight_);
+    Weight w1(siter.Value());
+    Weight w2;
+    for (siter.Next(); !siter.Done(); siter.Next()) w2.PushBack(siter.Value());
+    return std::make_pair(w1, w2);
+  }
+
+  void Reset() { done_ = weight_.Size() <= 1; }
+
+ private:
+  const StringWeight<Label, S> weight_;
+  bool done_;
+};
+
+// Factor a GallicWeight using StringFactor.
+template <class Label, class W, GallicType G = GALLIC_LEFT>
+class GallicFactor {
+ public:
+  using GW = GallicWeight<Label, W, G>;
+
+  explicit GallicFactor(const GW &weight)
+      : weight_(weight), done_(weight.Value1().Size() <= 1) {}
+
+  bool Done() const { return done_; }
+
+  void Next() { done_ = true; }
+
+  std::pair<GW, GW> Value() const {
+    StringFactor<Label, GallicStringType(G)> siter(weight_.Value1());
+    GW w1(siter.Value().first, weight_.Value2());
+    GW w2(siter.Value().second, W::One());
+    return std::make_pair(w1, w2);
+  }
+
+  void Reset() { done_ = weight_.Value1().Size() <= 1; }
+
+ private:
+  const GW weight_;
+  bool done_;
+};
+
+// Specialization for the (general) GALLIC type GallicWeight.
+template <class Label, class W>
+class GallicFactor<Label, W, GALLIC> {
+ public:
+  using GW = GallicWeight<Label, W, GALLIC>;
+  using GRW = GallicWeight<Label, W, GALLIC_RESTRICT>;
+
+  explicit GallicFactor(const GW &weight)
+      : iter_(weight),
+        done_(weight.Size() == 0 ||
+              (weight.Size() == 1 && weight.Back().Value1().Size() <= 1)) {}
+
+  bool Done() const { return done_ || iter_.Done(); }
+
+  void Next() { iter_.Next(); }
+
+  void Reset() { iter_.Reset(); }
+
+  std::pair<GW, GW> Value() const {
+    const auto weight = iter_.Value();
+    StringFactor<Label, GallicStringType(GALLIC_RESTRICT)> siter(
+        weight.Value1());
+    GRW w1(siter.Value().first, weight.Value2());
+    GRW w2(siter.Value().second, W::One());
+    return std::make_pair(GW(w1), GW(w2));
+  }
+
+ private:
+  UnionWeightIterator<GRW, GallicUnionWeightOptions<Label, W>> iter_;
+  bool done_;
+};
+
+namespace internal {
+
+// Implementation class for FactorWeight
+template <class Arc, class FactorIterator>
+class FactorWeightFstImpl : public CacheImpl<Arc> {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FstImpl<Arc>::SetType;
+  using FstImpl<Arc>::SetProperties;
+  using FstImpl<Arc>::SetInputSymbols;
+  using FstImpl<Arc>::SetOutputSymbols;
+
+  using CacheBaseImpl<CacheState<Arc>>::EmplaceArc;
+  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
+  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
+  using CacheBaseImpl<CacheState<Arc>>::HasStart;
+  using CacheBaseImpl<CacheState<Arc>>::SetArcs;
+  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
+  using CacheBaseImpl<CacheState<Arc>>::SetStart;
+
+  struct Element {
+    Element() {}
+
+    Element(StateId s, Weight weight_) : state(s), weight(std::move(weight_)) {}
+
+    StateId state;  // Input state ID.
+    Weight weight;  // Residual weight.
+  };
+
+  FactorWeightFstImpl(const Fst<Arc> &fst, const FactorWeightOptions<Arc> &opts)
+      : CacheImpl<Arc>(opts),
+        fst_(fst.Copy()),
+        delta_(opts.delta),
+        mode_(opts.mode),
+        final_ilabel_(opts.final_ilabel),
+        final_olabel_(opts.final_olabel),
+        increment_final_ilabel_(opts.increment_final_ilabel),
+        increment_final_olabel_(opts.increment_final_olabel) {
+    SetType("factor_weight");
+    const auto props = fst.Properties(kFstProperties, false);
+    SetProperties(FactorWeightProperties(props), kCopyProperties);
+    SetInputSymbols(fst.InputSymbols());
+    SetOutputSymbols(fst.OutputSymbols());
+    if (mode_ == 0) {
+      LOG(WARNING) << "FactorWeightFst: Factor mode is set to 0; "
+                   << "factoring neither arc weights nor final weights";
+    }
+  }
+
+  FactorWeightFstImpl(const FactorWeightFstImpl<Arc, FactorIterator> &impl)
+      : CacheImpl<Arc>(impl),
+        fst_(impl.fst_->Copy(true)),
+        delta_(impl.delta_),
+        mode_(impl.mode_),
+        final_ilabel_(impl.final_ilabel_),
+        final_olabel_(impl.final_olabel_),
+        increment_final_ilabel_(impl.increment_final_ilabel_),
+        increment_final_olabel_(impl.increment_final_olabel_) {
+    SetType("factor_weight");
+    SetProperties(impl.Properties(), kCopyProperties);
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+  }
+
+  StateId Start() {
+    if (!HasStart()) {
+      const auto s = fst_->Start();
+      if (s == kNoStateId) return kNoStateId;
+      SetStart(FindState(Element(fst_->Start(), Weight::One())));
+    }
+    return CacheImpl<Arc>::Start();
+  }
+
+  Weight Final(StateId s) {
+    if (!HasFinal(s)) {
+      const auto &element = elements_[s];
+      // TODO(sorenj): fix so cast is unnecessary
+      const auto weight =
+          element.state == kNoStateId
+              ? element.weight
+              : (Weight)Times(element.weight, fst_->Final(element.state));
+      FactorIterator siter(weight);
+      if (!(mode_ & kFactorFinalWeights) || siter.Done()) {
+        SetFinal(s, weight);
+      } else {
+        SetFinal(s, Weight::Zero());
+      }
+    }
+    return CacheImpl<Arc>::Final(s);
+  }
+
+  size_t NumArcs(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<Arc>::NumArcs(s);
+  }
+
+  size_t NumInputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<Arc>::NumInputEpsilons(s);
+  }
+
+  size_t NumOutputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<Arc>::NumOutputEpsilons(s);
+  }
+
+  uint64 Properties() const override { return Properties(kFstProperties); }
+
+  // Sets error if found, and returns other FST impl properties.
+  uint64 Properties(uint64 mask) const override {
+    if ((mask & kError) && fst_->Properties(kError, false)) {
+      SetProperties(kError, kError);
+    }
+    return FstImpl<Arc>::Properties(mask);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
+    if (!HasArcs(s)) Expand(s);
+    CacheImpl<Arc>::InitArcIterator(s, data);
+  }
+
+  // Finds state corresponding to an element, creating new state if element not
+  // found.
+  StateId FindState(const Element &element) {
+    if (!(mode_ & kFactorArcWeights) && element.weight == Weight::One() &&
+        element.state != kNoStateId) {
+      while (unfactored_.size() <= element.state)
+        unfactored_.push_back(kNoStateId);
+      if (unfactored_[element.state] == kNoStateId) {
+        unfactored_[element.state] = elements_.size();
+        elements_.push_back(element);
+      }
+      return unfactored_[element.state];
+    } else {
+      const auto insert_result =
+          element_map_.insert(std::make_pair(element, elements_.size()));
+      if (insert_result.second) {
+        elements_.push_back(element);
+      }
+      return insert_result.first->second;
+    }
+  }
+
+  // Computes the outgoing transitions from a state, creating new destination
+  // states as needed.
+  void Expand(StateId s) {
+    const auto element = elements_[s];
+    if (element.state != kNoStateId) {
+      for (ArcIterator<Fst<Arc>> ait(*fst_, element.state); !ait.Done();
+           ait.Next()) {
+        const auto &arc = ait.Value();
+        auto weight = Times(element.weight, arc.weight);
+        FactorIterator fiter(weight);
+        if (!(mode_ & kFactorArcWeights) || fiter.Done()) {
+          const auto dest = FindState(Element(arc.nextstate, Weight::One()));
+          EmplaceArc(s, arc.ilabel, arc.olabel, std::move(weight), dest);
+        } else {
+          for (; !fiter.Done(); fiter.Next()) {
+            auto pair = fiter.Value();
+            const auto dest =
+                FindState(Element(arc.nextstate, pair.second.Quantize(delta_)));
+            EmplaceArc(s, arc.ilabel, arc.olabel, std::move(pair.first), dest);
+          }
+        }
+      }
+    }
+    if ((mode_ & kFactorFinalWeights) &&
+        ((element.state == kNoStateId) ||
+         (fst_->Final(element.state) != Weight::Zero()))) {
+      const auto weight =
+          element.state == kNoStateId
+              ? element.weight
+              : Times(element.weight, fst_->Final(element.state));
+      auto ilabel = final_ilabel_;
+      auto olabel = final_olabel_;
+      for (FactorIterator fiter(weight); !fiter.Done(); fiter.Next()) {
+        auto pair = fiter.Value();
+        const auto dest =
+            FindState(Element(kNoStateId, pair.second.Quantize(delta_)));
+        EmplaceArc(s, ilabel, olabel, std::move(pair.first), dest);
+        if (increment_final_ilabel_) ++ilabel;
+        if (increment_final_olabel_) ++olabel;
+      }
+    }
+    SetArcs(s);
+  }
+
+ private:
+  // Equality function for Elements, assume weights have been quantized.
+  class ElementEqual {
+   public:
+    bool operator()(const Element &x, const Element &y) const {
+      return x.state == y.state && x.weight == y.weight;
+    }
+  };
+
+  // Hash function for Elements to Fst states.
+  class ElementKey {
+   public:
+    size_t operator()(const Element &x) const {
+      static constexpr auto prime = 7853;
+      return static_cast<size_t>(x.state * prime + x.weight.Hash());
+    }
+  };
+
+  using ElementMap =
+      std::unordered_map<Element, StateId, ElementKey, ElementEqual>;
+
+  std::unique_ptr<const Fst<Arc>> fst_;
+  float delta_;
+  uint32 mode_;         // Factoring arc and/or final weights.
+  Label final_ilabel_;  // ilabel of arc created when factoring final weights.
+  Label final_olabel_;  // olabel of arc created when factoring final weights.
+  bool increment_final_ilabel_;    // When factoring final weights results in
+  bool increment_final_olabel_;    // mutiple arcs, increment labels?
+  std::vector<Element> elements_;  // mapping from FST state to Element.
+  ElementMap element_map_;         // mapping from Element to FST state.
+  // Mapping between old/new StateId for states that do not need to be factored
+  // when mode_ is 0 or kFactorFinalWeights.
+  std::vector<StateId> unfactored_;
+};
+
+}  // namespace internal
+
+// FactorWeightFst takes as template parameter a FactorIterator as defined
+// above. The result of weight factoring is a transducer equivalent to the
+// input whose path weights have been factored according to the FactorIterator.
+// States and transitions will be added as necessary. The algorithm is a
+// generalization to arbitrary weights of the second step of the input
+// epsilon-normalization algorithm.
+//
+// This class attaches interface to implementation and handles reference
+// counting, delegating most methods to ImplToFst.
+template <class A, class FactorIterator>
+class FactorWeightFst
+    : public ImplToFst<internal::FactorWeightFstImpl<A, FactorIterator>> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Store = DefaultCacheStore<Arc>;
+  using State = typename Store::State;
+  using Impl = internal::FactorWeightFstImpl<Arc, FactorIterator>;
+
+  friend class ArcIterator<FactorWeightFst<Arc, FactorIterator>>;
+  friend class StateIterator<FactorWeightFst<Arc, FactorIterator>>;
+
+  explicit FactorWeightFst(const Fst<Arc> &fst)
+      : ImplToFst<Impl>(
+            std::make_shared<Impl>(fst, FactorWeightOptions<Arc>())) {}
+
+  FactorWeightFst(const Fst<Arc> &fst, const FactorWeightOptions<Arc> &opts)
+      : ImplToFst<Impl>(std::make_shared<Impl>(fst, opts)) {}
+
+  // See Fst<>::Copy() for doc.
+  FactorWeightFst(const FactorWeightFst<Arc, FactorIterator> &fst, bool copy)
+      : ImplToFst<Impl>(fst, copy) {}
+
+  // Get a copy of this FactorWeightFst. See Fst<>::Copy() for further doc.
+  FactorWeightFst<Arc, FactorIterator> *Copy(bool copy = false) const override {
+    return new FactorWeightFst<Arc, FactorIterator>(*this, copy);
+  }
+
+  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
+    GetMutableImpl()->InitArcIterator(s, data);
+  }
+
+ private:
+  using ImplToFst<Impl>::GetImpl;
+  using ImplToFst<Impl>::GetMutableImpl;
+
+  FactorWeightFst &operator=(const FactorWeightFst &) = delete;
+};
+
+// Specialization for FactorWeightFst.
+template <class Arc, class FactorIterator>
+class StateIterator<FactorWeightFst<Arc, FactorIterator>>
+    : public CacheStateIterator<FactorWeightFst<Arc, FactorIterator>> {
+ public:
+  explicit StateIterator(const FactorWeightFst<Arc, FactorIterator> &fst)
+      : CacheStateIterator<FactorWeightFst<Arc, FactorIterator>>(
+            fst, fst.GetMutableImpl()) {}
+};
+
+// Specialization for FactorWeightFst.
+template <class Arc, class FactorIterator>
+class ArcIterator<FactorWeightFst<Arc, FactorIterator>>
+    : public CacheArcIterator<FactorWeightFst<Arc, FactorIterator>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const FactorWeightFst<Arc, FactorIterator> &fst, StateId s)
+      : CacheArcIterator<FactorWeightFst<Arc, FactorIterator>>(
+            fst.GetMutableImpl(), s) {
+    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
+  }
+};
+
+template <class Arc, class FactorIterator>
+inline void FactorWeightFst<Arc, FactorIterator>::InitStateIterator(
+    StateIteratorData<Arc> *data) const {
+  data->base = new StateIterator<FactorWeightFst<Arc, FactorIterator>>(*this);
+}
+
+}  // namespace fst
+
+#endif  // FST_FACTOR_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/filter-state.h b/projects/llm_framework/include/fst/filter-state.h
new file mode 100644
index 00000000..289b628e
--- /dev/null
+++ b/projects/llm_framework/include/fst/filter-state.h
@@ -0,0 +1,199 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Classes for storing filter state in various algorithms like composition.
+
+#ifndef FST_FILTER_STATE_H_
+#define FST_FILTER_STATE_H_
+
+#include <forward_list>
+#include <utility>
+
+#include <fst/fst-decl.h>  // For optional argument declarations
+#include <fst/fst.h>
+#include <fst/matcher.h>
+
+
+namespace fst {
+
+// The filter state interface represents the state of a (e.g., composition)
+// filter.
+//
+// class FilterState {
+//  public:
+//   // Required constructors.
+//
+//   FilterState();
+//
+//   FilterState(const FilterState &fs);
+//
+//   // An invalid filter state.
+//   static const FilterState NoState();
+//
+//   // Maps state to integer for hashing.
+//   size_t Hash() const;
+//
+//   // Equality of filter states.
+//   bool operator==(const FilterState &fs) const;
+//
+//   // Inequality of filter states.
+//   bool operator!=(const FilterState &fs) const;
+//
+//   // Assignment to filter states.
+//   FilterState &operator=(const FilterState& fs);
+// };
+
+// Filter state that is a signed integral type.
+template <typename T>
+class IntegerFilterState {
+ public:
+  IntegerFilterState() : state_(kNoStateId) {}
+
+  explicit IntegerFilterState(T s) : state_(s) {}
+
+  static const IntegerFilterState NoState() { return IntegerFilterState(); }
+
+  size_t Hash() const { return static_cast<size_t>(state_); }
+
+  bool operator==(const IntegerFilterState &fs) const {
+    return state_ == fs.state_;
+  }
+
+  bool operator!=(const IntegerFilterState &fs) const {
+    return state_ != fs.state_;
+  }
+
+  T GetState() const { return state_; }
+
+  void SetState(T state) { state_ = state; }
+
+ private:
+  T state_;
+};
+
+using CharFilterState = IntegerFilterState<signed char>;
+using ShortFilterState = IntegerFilterState<short>;  // NOLINT
+using IntFilterState = IntegerFilterState<int>;
+
+// Filter state that is a weight (class).
+template <class W>
+class WeightFilterState {
+ public:
+  WeightFilterState() : weight_(W::Zero()) {}
+
+  explicit WeightFilterState(W weight) : weight_(std::move(weight)) {}
+
+  static const WeightFilterState NoState() { return WeightFilterState(); }
+
+  size_t Hash() const { return weight_.Hash(); }
+
+  bool operator==(const WeightFilterState &fs) const {
+    return weight_ == fs.weight_;
+  }
+
+  bool operator!=(const WeightFilterState &fs) const {
+    return weight_ != fs.weight_;
+  }
+
+  W GetWeight() const { return weight_; }
+
+  void SetWeight(W weight) { weight_ = std::move(weight); }
+
+ private:
+  W weight_;
+};
+
+// Filter state is a list of signed integer types T. Order matters
+// for equality.
+template <typename T>
+class ListFilterState {
+ public:
+  ListFilterState() {}
+
+  explicit ListFilterState(T s) { list_.push_front(s); }
+
+  static const ListFilterState NoState() { return ListFilterState(kNoStateId); }
+
+  size_t Hash() const {
+    size_t h = 0;
+    for (const auto &elem : list_) h ^= h << 1 ^ elem;
+    return h;
+  }
+
+  bool operator==(const ListFilterState &fs) const { return list_ == fs.list_; }
+
+  bool operator!=(const ListFilterState &fs) const { return list_ != fs.list_; }
+
+  const std::forward_list<T> &GetState() const { return list_; }
+
+  std::forward_list<T> *GetMutableState() { return &list_; }
+
+  void SetState(const std::forward_list<T> &state) { list_ = state; }
+
+ private:
+  std::forward_list<T> list_;
+};
+
+// Filter state that is the combination of two filter states.
+template <class FS1, class FS2>
+class PairFilterState {
+ public:
+  PairFilterState() : fs1_(FS1::NoState()), fs2_(FS2::NoState()) {}
+
+  PairFilterState(const FS1 &fs1, const FS2 &fs2) : fs1_(fs1), fs2_(fs2) {}
+
+  static const PairFilterState NoState() { return PairFilterState(); }
+
+  size_t Hash() const {
+    const auto h1 = fs1_.Hash();
+    static constexpr auto lshift = 5;
+    static constexpr auto rshift = CHAR_BIT * sizeof(size_t) - 5;
+    return h1 << lshift ^ h1 >> rshift ^ fs2_.Hash();
+  }
+
+  bool operator==(const PairFilterState &fs) const {
+    return fs1_ == fs.fs1_ && fs2_ == fs.fs2_;
+  }
+
+  bool operator!=(const PairFilterState &fs) const {
+    return fs1_ != fs.fs1_ || fs2_ != fs.fs2_;
+  }
+
+  const FS1 &GetState1() const { return fs1_; }
+
+  const FS2 &GetState2() const { return fs2_; }
+
+  void SetState(const FS1 &fs1, const FS2 &fs2) {
+    fs1_ = fs1;
+    fs2_ = fs2;
+  }
+
+ private:
+  FS1 fs1_;
+  FS2 fs2_;
+};
+
+// Single non-blocking filter state.
+class TrivialFilterState {
+ public:
+  explicit TrivialFilterState(bool state = false) : state_(state) {}
+
+  static const TrivialFilterState NoState() { return TrivialFilterState(); }
+
+  size_t Hash() const { return 0; }
+
+  bool operator==(const TrivialFilterState &fs) const {
+    return state_ == fs.state_;
+  }
+
+  bool operator!=(const TrivialFilterState &fs) const {
+    return state_ != fs.state_;
+  }
+
+ private:
+  bool state_;
+};
+
+}  // namespace fst
+
+#endif  // FST_FILTER_STATE_H_
diff --git a/projects/llm_framework/include/fst/flags.h b/projects/llm_framework/include/fst/flags.h
new file mode 100644
index 00000000..6ef1257f
--- /dev/null
+++ b/projects/llm_framework/include/fst/flags.h
@@ -0,0 +1,219 @@
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Google-style flag handling declarations and inline definitions.
+
+#ifndef FST_LIB_FLAGS_H_
+#define FST_LIB_FLAGS_H_
+
+#include <cstdlib>
+
+#include <iostream>
+#include <map>
+#include <set>
+#include <sstream>
+#include <string>
+
+#include <fst/lock.h>
+#include <fst/types.h>
+
+#include "gflags/gflags.h"
+#include "glog/logging.h"
+
+using std::string;
+
+// FLAGS USAGE:
+//
+// Definition example:
+//
+//    DEFINE_int32(length, 0, "length");
+//
+// This defines variable FLAGS_length, initialized to 0.
+//
+// Declaration example:
+//
+//    DECLARE_int32(length);
+//
+// SET_FLAGS() can be used to set flags from the command line
+// using, for example, '--length=2'.
+//
+// ShowUsage() can be used to print out command and flag usage.
+
+// #define DECLARE_bool(name) extern bool FLAGS_ ## name
+// #define DECLARE_string(name) extern string FLAGS_ ## name
+// #define DECLARE_int32(name) extern int32 FLAGS_ ## name
+// #define DECLARE_int64(name) extern int64 FLAGS_ ## name
+// #define DECLARE_double(name) extern double FLAGS_ ## name
+
+template <typename T>
+struct FlagDescription {
+  FlagDescription(T* addr, const char* doc, const char* type, const char* file,
+                  const T val)
+      : address(addr),
+        doc_string(doc),
+        type_name(type),
+        file_name(file),
+        default_value(val) {}
+
+  T* address;
+  const char* doc_string;
+  const char* type_name;
+  const char* file_name;
+  const T default_value;
+};
+
+template <typename T>
+class FlagRegister {
+ public:
+  static FlagRegister<T>* GetRegister() {
+    static auto reg = new FlagRegister<T>;
+    return reg;
+  }
+
+  const FlagDescription<T>& GetFlagDescription(const string& name) const {
+    fst::MutexLock l(&flag_lock_);
+    auto it = flag_table_.find(name);
+    return it != flag_table_.end() ? it->second : 0;
+  }
+
+  void SetDescription(const string& name, const FlagDescription<T>& desc) {
+    fst::MutexLock l(&flag_lock_);
+    flag_table_.insert(make_pair(name, desc));
+  }
+
+  bool SetFlag(const string& val, bool* address) const {
+    if (val == "true" || val == "1" || val.empty()) {
+      *address = true;
+      return true;
+    } else if (val == "false" || val == "0") {
+      *address = false;
+      return true;
+    } else {
+      return false;
+    }
+  }
+
+  bool SetFlag(const string& val, string* address) const {
+    *address = val;
+    return true;
+  }
+
+  bool SetFlag(const string& val, int32* address) const {
+    char* p = 0;
+    *address = strtol(val.c_str(), &p, 0);
+    return !val.empty() && *p == '\0';
+  }
+
+  bool SetFlag(const string& val, int64* address) const {
+    char* p = 0;
+    *address = strtoll(val.c_str(), &p, 0);
+    return !val.empty() && *p == '\0';
+  }
+
+  bool SetFlag(const string& val, double* address) const {
+    char* p = 0;
+    *address = strtod(val.c_str(), &p);
+    return !val.empty() && *p == '\0';
+  }
+
+  bool SetFlag(const string& arg, const string& val) const {
+    for (typename std::map<string, FlagDescription<T>>::const_iterator it =
+             flag_table_.begin();
+         it != flag_table_.end(); ++it) {
+      const string& name = it->first;
+      const FlagDescription<T>& desc = it->second;
+      if (arg == name) return SetFlag(val, desc.address);
+    }
+    return false;
+  }
+
+  void GetUsage(std::set<std::pair<string, string>>* usage_set) const {
+    for (auto it = flag_table_.begin(); it != flag_table_.end(); ++it) {
+      const string& name = it->first;
+      const FlagDescription<T>& desc = it->second;
+      string usage = "  --" + name;
+      usage += ": type = ";
+      usage += desc.type_name;
+      usage += ", default = ";
+      usage += GetDefault(desc.default_value) + "\n  ";
+      usage += desc.doc_string;
+      usage_set->insert(make_pair(desc.file_name, usage));
+    }
+  }
+
+ private:
+  string GetDefault(bool default_value) const {
+    return default_value ? "true" : "false";
+  }
+
+  string GetDefault(const string& default_value) const {
+    return "\"" + default_value + "\"";
+  }
+
+  template <class V>
+  string GetDefault(const V& default_value) const {
+    std::ostringstream strm;
+    strm << default_value;
+    return strm.str();
+  }
+
+  mutable fst::Mutex flag_lock_;  // Multithreading lock.
+  std::map<string, FlagDescription<T>> flag_table_;
+};
+
+template <typename T>
+class FlagRegisterer {
+ public:
+  FlagRegisterer(const string& name, const FlagDescription<T>& desc) {
+    auto registr = FlagRegister<T>::GetRegister();
+    registr->SetDescription(name, desc);
+  }
+
+ private:
+  FlagRegisterer(const FlagRegisterer&) = delete;
+  FlagRegisterer& operator=(const FlagRegisterer&) = delete;
+};
+
+#define DEFINE_VAR(type, name, value, doc)             \
+  type FLAGS_##name = value;                           \
+  static FlagRegisterer<type> name##_flags_registerer( \
+      #name,                                           \
+      FlagDescription<type>(&FLAGS_##name, doc, #type, __FILE__, value))
+
+// #define DEFINE_bool(name, value, doc) DEFINE_VAR(bool, name, value, doc)
+// #define DEFINE_string(name, value, doc) \
+//   DEFINE_VAR(string, name, value, doc)
+// #define DEFINE_int32(name, value, doc) DEFINE_VAR(int32, name, value, doc)
+// #define DEFINE_int64(name, value, doc) DEFINE_VAR(int64, name, value, doc)
+// #define DEFINE_double(name, value, doc) DEFINE_VAR(double, name, value, doc)
+
+// Temporary directory.
+DECLARE_string(tmpdir);
+
+void SetFlags(const char* usage, int* argc, char*** argv, bool remove_flags,
+              const char* src = "");
+
+#define SET_FLAGS(usage, argc, argv, rmflags) \
+  gflags::ParseCommandLineFlags(argc, argv, true)
+// SetFlags(usage, argc, argv, rmflags, __FILE__)
+
+// Deprecated; for backward compatibility.
+inline void InitFst(const char* usage, int* argc, char*** argv, bool rmflags) {
+  return SetFlags(usage, argc, argv, rmflags);
+}
+
+void ShowUsage(bool long_usage = true);
+
+#endif  // FST_LIB_FLAGS_H_
diff --git a/projects/llm_framework/include/fst/float-weight.h b/projects/llm_framework/include/fst/float-weight.h
new file mode 100644
index 00000000..8fade684
--- /dev/null
+++ b/projects/llm_framework/include/fst/float-weight.h
@@ -0,0 +1,858 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Float weight set and associated semiring operation definitions.
+
+#ifndef FST_FLOAT_WEIGHT_H_
+#define FST_FLOAT_WEIGHT_H_
+
+#include <algorithm>
+#include <climits>
+#include <cmath>
+#include <cstdlib>
+#include <cstring>
+#include <limits>
+#include <sstream>
+#include <string>
+#include <type_traits>
+
+#include <fst/util.h>
+#include <fst/weight.h>
+
+
+namespace fst {
+
+// Numeric limits class.
+template <class T>
+class FloatLimits {
+ public:
+  static constexpr T PosInfinity() {
+    return std::numeric_limits<T>::infinity();
+  }
+
+  static constexpr T NegInfinity() { return -PosInfinity(); }
+
+  static constexpr T NumberBad() { return std::numeric_limits<T>::quiet_NaN(); }
+};
+
+// Weight class to be templated on floating-points types.
+template <class T = float>
+class FloatWeightTpl {
+ public:
+  using ValueType = T;
+
+  FloatWeightTpl() noexcept {}
+
+  constexpr FloatWeightTpl(T f) : value_(f) {}  // NOLINT
+
+  // TODO(mjansche): Leave implicit once Android NDK r18 is the default.
+  FloatWeightTpl(const FloatWeightTpl<T> &) = default;
+  FloatWeightTpl(FloatWeightTpl<T> &&) noexcept = default;
+
+  FloatWeightTpl<T> &operator=(const FloatWeightTpl<T> &) = default;
+  FloatWeightTpl<T> &operator=(FloatWeightTpl<T> &&) noexcept = default;
+
+  std::istream &Read(std::istream &strm) { return ReadType(strm, &value_); }
+
+  std::ostream &Write(std::ostream &strm) const {
+    return WriteType(strm, value_);
+  }
+
+  size_t Hash() const {
+    size_t hash = 0;
+    // Avoid using union, which would be undefined behavior.
+    // Use memcpy, similar to bit_cast, but sizes may be different.
+    // This should be optimized into a single move instruction by
+    // any reasonable compiler.
+    std::memcpy(&hash, &value_, std::min(sizeof(hash), sizeof(value_)));
+    return hash;
+  }
+
+  constexpr const T &Value() const { return value_; }
+
+ protected:
+  void SetValue(const T &f) { value_ = f; }
+
+  static constexpr const char *GetPrecisionString() {
+    return sizeof(T) == 4 ? ""
+        : sizeof(T) == 1 ? "8"
+        : sizeof(T) == 2 ? "16"
+        : sizeof(T) == 8 ? "64"
+        : "unknown";
+  }
+
+ private:
+  T value_;
+};
+
+// Single-precision float weight.
+using FloatWeight = FloatWeightTpl<float>;
+
+template <class T>
+constexpr bool operator==(const FloatWeightTpl<T> &w1,
+                          const FloatWeightTpl<T> &w2) {
+#if (defined(__i386__) || defined(__x86_64__)) && !defined(__SSE2_MATH__)
+// With i387 instructions, excess precision on a weight in an 80-bit
+// register may cause it to compare unequal to that same weight when
+// stored to memory.  This breaks =='s reflexivity, in turn breaking
+// NaturalLess.
+#error "Please compile with -msse -mfpmath=sse, or equivalent."
+#endif
+  return w1.Value() == w2.Value();
+}
+
+// These seemingly unnecessary overloads are actually needed to make
+// comparisons like FloatWeightTpl<float> == float compile.  If only the
+// templated version exists, the FloatWeightTpl<float>(float) conversion
+// won't be found.
+constexpr bool operator==(const FloatWeightTpl<float> &w1,
+                          const FloatWeightTpl<float> &w2) {
+  return operator==<float>(w1, w2);
+}
+
+constexpr bool operator==(const FloatWeightTpl<double> &w1,
+                          const FloatWeightTpl<double> &w2) {
+  return operator==<double>(w1, w2);
+}
+
+template <class T>
+constexpr bool operator!=(const FloatWeightTpl<T> &w1,
+                          const FloatWeightTpl<T> &w2) {
+  return !(w1 == w2);
+}
+
+constexpr bool operator!=(const FloatWeightTpl<float> &w1,
+                          const FloatWeightTpl<float> &w2) {
+  return operator!=<float>(w1, w2);
+}
+
+constexpr bool operator!=(const FloatWeightTpl<double> &w1,
+                          const FloatWeightTpl<double> &w2) {
+  return operator!=<double>(w1, w2);
+}
+
+template <class T>
+constexpr bool FloatApproxEqual(T w1, T w2, float delta = kDelta) {
+  return w1 <= w2 + delta && w2 <= w1 + delta;
+}
+
+template <class T>
+constexpr bool ApproxEqual(const FloatWeightTpl<T> &w1,
+                           const FloatWeightTpl<T> &w2, float delta = kDelta) {
+  return FloatApproxEqual(w1.Value(), w2.Value(), delta);
+}
+
+template <class T>
+inline std::ostream &operator<<(std::ostream &strm,
+                                const FloatWeightTpl<T> &w) {
+  if (w.Value() == FloatLimits<T>::PosInfinity()) {
+    return strm << "Infinity";
+  } else if (w.Value() == FloatLimits<T>::NegInfinity()) {
+    return strm << "-Infinity";
+  } else if (w.Value() != w.Value()) {  // Fails for IEEE NaN.
+    return strm << "BadNumber";
+  } else {
+    return strm << w.Value();
+  }
+}
+
+template <class T>
+inline std::istream &operator>>(std::istream &strm, FloatWeightTpl<T> &w) {
+  string s;
+  strm >> s;
+  if (s == "Infinity") {
+    w = FloatWeightTpl<T>(FloatLimits<T>::PosInfinity());
+  } else if (s == "-Infinity") {
+    w = FloatWeightTpl<T>(FloatLimits<T>::NegInfinity());
+  } else {
+    char *p;
+    T f = strtod(s.c_str(), &p);
+    if (p < s.c_str() + s.size()) {
+      strm.clear(std::ios::badbit);
+    } else {
+      w = FloatWeightTpl<T>(f);
+    }
+  }
+  return strm;
+}
+
+// Tropical semiring: (min, +, inf, 0).
+template <class T>
+class TropicalWeightTpl : public FloatWeightTpl<T> {
+ public:
+  using typename FloatWeightTpl<T>::ValueType;
+  using FloatWeightTpl<T>::Value;
+  using ReverseWeight = TropicalWeightTpl<T>;
+  using Limits = FloatLimits<T>;
+
+  TropicalWeightTpl() noexcept : FloatWeightTpl<T>() {}
+
+  constexpr TropicalWeightTpl(T f) : FloatWeightTpl<T>(f) {}
+
+  static constexpr TropicalWeightTpl<T> Zero() { return Limits::PosInfinity(); }
+
+  static constexpr TropicalWeightTpl<T> One() { return 0; }
+
+  static constexpr TropicalWeightTpl<T> NoWeight() {
+    return Limits::NumberBad();
+  }
+
+  static const string &Type() {
+    static const string *const type =
+        new string(string("tropical") +
+                   FloatWeightTpl<T>::GetPrecisionString());
+    return *type;
+  }
+
+  constexpr bool Member() const {
+    // All floating point values except for NaNs and negative infinity are valid
+    // tropical weights.
+    //
+    // Testing membership of a given value can be done by simply checking that
+    // it is strictly greater than negative infinity, which fails for negative
+    // infinity itself but also for NaNs. This can usually be accomplished in a
+    // single instruction (such as *UCOMI* on x86) without branching logic.
+    //
+    // An additional wrinkle involves constexpr correctness of floating point
+    // comparisons against NaN. GCC is uneven when it comes to which expressions
+    // it considers compile-time constants. In particular, current versions of
+    // GCC do not always consider (nan < inf) to be a constant expression, but
+    // do consider (inf < nan) to be a constant expression. (See
+    // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88173 and
+    // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88683 for details.) In order
+    // to allow Member() to be a constexpr function accepted by GCC, we write
+    // the comparison here as (-inf < v).
+    return Limits::NegInfinity() < Value();
+  }
+
+  TropicalWeightTpl<T> Quantize(float delta = kDelta) const {
+    if (!Member() || Value() == Limits::PosInfinity()) {
+      return *this;
+    } else {
+      return TropicalWeightTpl<T>(std::floor(Value() / delta + 0.5F) * delta);
+    }
+  }
+
+  constexpr TropicalWeightTpl<T> Reverse() const { return *this; }
+
+  static constexpr uint64 Properties() {
+    return kLeftSemiring | kRightSemiring | kCommutative | kPath | kIdempotent;
+  }
+};
+
+// Single precision tropical weight.
+using TropicalWeight = TropicalWeightTpl<float>;
+
+template <class T>
+constexpr TropicalWeightTpl<T> Plus(const TropicalWeightTpl<T> &w1,
+                                    const TropicalWeightTpl<T> &w2) {
+  return (!w1.Member() || !w2.Member()) ? TropicalWeightTpl<T>::NoWeight()
+      : w1.Value() < w2.Value() ? w1 : w2;
+}
+
+// See comment at operator==(FloatWeightTpl<float>, FloatWeightTpl<float>)
+// for why these overloads are present.
+constexpr TropicalWeightTpl<float> Plus(const TropicalWeightTpl<float> &w1,
+                                        const TropicalWeightTpl<float> &w2) {
+  return Plus<float>(w1, w2);
+}
+
+constexpr TropicalWeightTpl<double> Plus(const TropicalWeightTpl<double> &w1,
+                                         const TropicalWeightTpl<double> &w2) {
+  return Plus<double>(w1, w2);
+}
+
+template <class T>
+constexpr TropicalWeightTpl<T> Times(const TropicalWeightTpl<T> &w1,
+                                     const TropicalWeightTpl<T> &w2) {
+  // The following is safe in the context of the Tropical (and Log) semiring
+  // for all IEEE floating point values, including infinities and NaNs,
+  // because:
+  //
+  // * If one or both of the floating point Values is NaN and hence not a
+  //   Member, the result of addition below is NaN, so the result is not a
+  //   Member. This supersedes all other cases, so we only consider non-NaN
+  //   values next.
+  //
+  // * If both Values are finite, there is no issue.
+  //
+  // * If one of the Values is infinite, or if both are infinities with the
+  //   same sign, the result of floating point addition is the same infinity,
+  //   so there is no issue.
+  //
+  // * If both of the Values are infinities with opposite signs, the result of
+  //   adding IEEE floating point -inf + inf is NaN and hence not a Member. But
+  //   since -inf was not a Member to begin with, returning a non-Member result
+  //   is fine as well.
+  return TropicalWeightTpl<T>(w1.Value() + w2.Value());
+}
+
+constexpr TropicalWeightTpl<float> Times(const TropicalWeightTpl<float> &w1,
+                                         const TropicalWeightTpl<float> &w2) {
+  return Times<float>(w1, w2);
+}
+
+constexpr TropicalWeightTpl<double> Times(const TropicalWeightTpl<double> &w1,
+                                          const TropicalWeightTpl<double> &w2) {
+  return Times<double>(w1, w2);
+}
+
+template <class T>
+constexpr TropicalWeightTpl<T> Divide(const TropicalWeightTpl<T> &w1,
+                                      const TropicalWeightTpl<T> &w2,
+                                      DivideType typ = DIVIDE_ANY) {
+  // The following is safe in the context of the Tropical (and Log) semiring
+  // for all IEEE floating point values, including infinities and NaNs,
+  // because:
+  //
+  // * If one or both of the floating point Values is NaN and hence not a
+  //   Member, the result of subtraction below is NaN, so the result is not a
+  //   Member. This supersedes all other cases, so we only consider non-NaN
+  //   values next.
+  //
+  // * If both Values are finite, there is no issue.
+  //
+  // * If w2.Value() is -inf (and hence w2 is not a Member), the result of ?:
+  //   below is NoWeight, which is not a Member.
+  //
+  //   Whereas in IEEE floating point semantics 0/inf == 0, this does not carry
+  //   over to this semiring (since TropicalWeight(-inf) would be the analogue
+  //   of floating point inf) and instead Divide(Zero(), TropicalWeight(-inf))
+  //   is NoWeight().
+  //
+  // * If w2.Value() is inf (and hence w2 is Zero), the resulting floating
+  //   point value is either NaN (if w1 is Zero or if w1.Value() is NaN) and
+  //   hence not a Member, or it is -inf and hence not a Member; either way,
+  //   division by Zero results in a non-Member result.
+  using Weight = TropicalWeightTpl<T>;
+  return w2.Member() ? Weight(w1.Value() - w2.Value()) : Weight::NoWeight();
+}
+
+constexpr TropicalWeightTpl<float> Divide(const TropicalWeightTpl<float> &w1,
+                                          const TropicalWeightTpl<float> &w2,
+                                          DivideType typ = DIVIDE_ANY) {
+  return Divide<float>(w1, w2, typ);
+}
+
+constexpr TropicalWeightTpl<double> Divide(const TropicalWeightTpl<double> &w1,
+                                           const TropicalWeightTpl<double> &w2,
+                                           DivideType typ = DIVIDE_ANY) {
+  return Divide<double>(w1, w2, typ);
+}
+
+// Power(w, n) calculates the n-th power of w with respect to semiring Times.
+//
+// In the case of the Tropical (and Log) semiring, the exponent n is not
+// restricted to be an integer. It can be a floating point value, for example.
+//
+// In weight.h, a narrower and hence more broadly applicable version of
+// Power(w, n) is defined for arbitrary weight types and non-negative integer
+// exponents n (of type size_t) and implemented in terms of repeated
+// multiplication using Times.
+//
+// Without further provisions this means that, when an expression such as
+//
+//   Power(TropicalWeightTpl<float>::One(), static_cast<size_t>(2))
+//
+// is specified, the overload of Power() is ambiguous. The template function
+// below could be instantiated as
+//
+//   Power<float, size_t>(const TropicalWeightTpl<float> &, size_t)
+//
+// and the template function defined in weight.h (further specialized below)
+// could be instantiated as
+//
+//   Power<TropicalWeightTpl<float>>(const TropicalWeightTpl<float> &, size_t)
+//
+// That would lead to two definitions with identical signatures, which results
+// in a compilation error. To avoid that, we hide the definition of Power<T, V>
+// when V is size_t, so only Power<W> is visible. Power<W> is further
+// specialized to Power<TropicalWeightTpl<...>>, and the overloaded definition
+// of Power<T, V> is made conditionally available only to that template
+// specialization.
+
+template <class T, class V,
+          bool Enable = !std::is_same<V, size_t>::value,
+          typename std::enable_if<Enable>::type * = nullptr>
+constexpr TropicalWeightTpl<T> Power(const TropicalWeightTpl<T> &w, V n) {
+  using Weight = TropicalWeightTpl<T>;
+  return (!w.Member() || n != n) ? Weight::NoWeight()
+      : (n == 0 || w == Weight::One()) ? Weight::One()
+      : Weight(w.Value() * n);
+}
+
+// Specializes the library-wide template to use the above implementation; rules
+// of function template instantiation require this be a full instantiation.
+
+template <>
+constexpr TropicalWeightTpl<float> Power<TropicalWeightTpl<float>>(
+    const TropicalWeightTpl<float> &weight, size_t n) {
+  return Power<float, size_t, true>(weight, n);
+}
+
+template <>
+constexpr TropicalWeightTpl<double> Power<TropicalWeightTpl<double>>(
+    const TropicalWeightTpl<double> &weight, size_t n) {
+  return Power<double, size_t, true>(weight, n);
+}
+
+
+// Log semiring: (log(e^-x + e^-y), +, inf, 0).
+template <class T>
+class LogWeightTpl : public FloatWeightTpl<T> {
+ public:
+  using typename FloatWeightTpl<T>::ValueType;
+  using FloatWeightTpl<T>::Value;
+  using ReverseWeight = LogWeightTpl;
+  using Limits = FloatLimits<T>;
+
+  LogWeightTpl() noexcept : FloatWeightTpl<T>() {}
+
+  constexpr LogWeightTpl(T f) : FloatWeightTpl<T>(f) {}
+
+  static constexpr LogWeightTpl Zero() { return Limits::PosInfinity(); }
+
+  static constexpr LogWeightTpl One() { return 0; }
+
+  static constexpr LogWeightTpl NoWeight() { return Limits::NumberBad(); }
+
+  static const string &Type() {
+    static const string *const type =
+        new string(string("log") + FloatWeightTpl<T>::GetPrecisionString());
+    return *type;
+  }
+
+  constexpr bool Member() const {
+    // The comments for TropicalWeightTpl<>::Member() apply here unchanged.
+    return Limits::NegInfinity() < Value();
+  }
+
+  LogWeightTpl<T> Quantize(float delta = kDelta) const {
+    if (!Member() || Value() == Limits::PosInfinity()) {
+      return *this;
+    } else {
+      return LogWeightTpl<T>(std::floor(Value() / delta + 0.5F) * delta);
+    }
+  }
+
+  constexpr LogWeightTpl<T> Reverse() const { return *this; }
+
+  static constexpr uint64 Properties() {
+    return kLeftSemiring | kRightSemiring | kCommutative;
+  }
+};
+
+// Single-precision log weight.
+using LogWeight = LogWeightTpl<float>;
+
+// Double-precision log weight.
+using Log64Weight = LogWeightTpl<double>;
+
+namespace internal {
+
+// -log(e^-x + e^-y) = x - LogPosExp(y - x), assuming x >= 0.0.
+inline double LogPosExp(double x) {
+  DCHECK(!(x < 0));  // NB: NaN values are allowed.
+  return log1p(exp(-x));
+}
+
+// -log(e^-x - e^-y) = x - LogNegExp(y - x), assuming x > 0.0.
+inline double LogNegExp(double x) {
+  DCHECK_GT(x, 0);
+  return log1p(-exp(-x));
+}
+
+// a +_log b = -log(e^-a + e^-b) = KahanLogSum(a, b, ...).
+// Kahan compensated summation provides an error bound that is
+// independent of the number of addends. Assumes b >= a;
+// c is the compensation.
+inline double KahanLogSum(double a, double b, double *c) {
+  DCHECK_GE(b, a);
+  double y = -LogPosExp(b - a) - *c;
+  double t = a + y;
+  *c = (t - a) - y;
+  return t;
+}
+
+// a -_log b = -log(e^-a - e^-b) = KahanLogDiff(a, b, ...).
+// Kahan compensated summation provides an error bound that is
+// independent of the number of addends. Assumes b > a;
+// c is the compensation.
+inline double KahanLogDiff(double a, double b, double *c) {
+  DCHECK_GT(b, a);
+  double y = -LogNegExp(b - a) - *c;
+  double t = a + y;
+  *c = (t - a) - y;
+  return t;
+}
+
+}  // namespace internal
+
+template <class T>
+inline LogWeightTpl<T> Plus(const LogWeightTpl<T> &w1,
+                            const LogWeightTpl<T> &w2) {
+  using Limits = FloatLimits<T>;
+  const T f1 = w1.Value();
+  const T f2 = w2.Value();
+  if (f1 == Limits::PosInfinity()) {
+    return w2;
+  } else if (f2 == Limits::PosInfinity()) {
+    return w1;
+  } else if (f1 > f2) {
+    return LogWeightTpl<T>(f2 - internal::LogPosExp(f1 - f2));
+  } else {
+    return LogWeightTpl<T>(f1 - internal::LogPosExp(f2 - f1));
+  }
+}
+
+inline LogWeightTpl<float> Plus(const LogWeightTpl<float> &w1,
+                                const LogWeightTpl<float> &w2) {
+  return Plus<float>(w1, w2);
+}
+
+inline LogWeightTpl<double> Plus(const LogWeightTpl<double> &w1,
+                                 const LogWeightTpl<double> &w2) {
+  return Plus<double>(w1, w2);
+}
+
+template <class T>
+constexpr LogWeightTpl<T> Times(const LogWeightTpl<T> &w1,
+                                const LogWeightTpl<T> &w2) {
+  // The comments for Times(Tropical...) above apply here unchanged.
+  return LogWeightTpl<T>(w1.Value() + w2.Value());
+}
+
+constexpr LogWeightTpl<float> Times(const LogWeightTpl<float> &w1,
+                                    const LogWeightTpl<float> &w2) {
+  return Times<float>(w1, w2);
+}
+
+constexpr LogWeightTpl<double> Times(const LogWeightTpl<double> &w1,
+                                     const LogWeightTpl<double> &w2) {
+  return Times<double>(w1, w2);
+}
+
+template <class T>
+constexpr LogWeightTpl<T> Divide(const LogWeightTpl<T> &w1,
+                                 const LogWeightTpl<T> &w2,
+                                 DivideType typ = DIVIDE_ANY) {
+  // The comments for Divide(Tropical...) above apply here unchanged.
+  using Weight = LogWeightTpl<T>;
+  return w2.Member() ? Weight(w1.Value() - w2.Value()) : Weight::NoWeight();
+}
+
+constexpr LogWeightTpl<float> Divide(const LogWeightTpl<float> &w1,
+                                     const LogWeightTpl<float> &w2,
+                                     DivideType typ = DIVIDE_ANY) {
+  return Divide<float>(w1, w2, typ);
+}
+
+constexpr LogWeightTpl<double> Divide(const LogWeightTpl<double> &w1,
+                                      const LogWeightTpl<double> &w2,
+                                      DivideType typ = DIVIDE_ANY) {
+  return Divide<double>(w1, w2, typ);
+}
+
+// The comments for Power<>(Tropical...) above apply here unchanged.
+
+template <class T, class V,
+          bool Enable = !std::is_same<V, size_t>::value,
+          typename std::enable_if<Enable>::type * = nullptr>
+constexpr LogWeightTpl<T> Power(const LogWeightTpl<T> &w, V n) {
+  using Weight = LogWeightTpl<T>;
+  return (!w.Member() || n != n) ? Weight::NoWeight()
+      : (n == 0 || w == Weight::One()) ? Weight::One()
+      : Weight(w.Value() * n);
+}
+
+// Specializes the library-wide template to use the above implementation; rules
+// of function template instantiation require this be a full instantiation.
+
+template <>
+constexpr LogWeightTpl<float> Power<LogWeightTpl<float>>(
+    const LogWeightTpl<float> &weight, size_t n) {
+  return Power<float, size_t, true>(weight, n);
+}
+
+template <>
+constexpr LogWeightTpl<double> Power<LogWeightTpl<double>>(
+    const LogWeightTpl<double> &weight, size_t n) {
+  return Power<double, size_t, true>(weight, n);
+}
+
+// Specialization using the Kahan compensated summation.
+template <class T>
+class Adder<LogWeightTpl<T>> {
+ public:
+  using Weight = LogWeightTpl<T>;
+
+  explicit Adder(Weight w = Weight::Zero())
+      : sum_(w.Value()),
+        c_(0.0) { }
+
+  Weight Add(const Weight &w) {
+    using Limits = FloatLimits<T>;
+    const T f = w.Value();
+    if (f == Limits::PosInfinity()) {
+      return Sum();
+    } else if (sum_ == Limits::PosInfinity()) {
+      sum_ = f;
+      c_ = 0.0;
+    } else if (f > sum_) {
+      sum_ = internal::KahanLogSum(sum_, f, &c_);
+    } else {
+      sum_ = internal::KahanLogSum(f, sum_, &c_);
+    }
+    return Sum();
+  }
+
+  Weight Sum() { return Weight(sum_); }
+
+  void Reset(Weight w = Weight::Zero()) {
+    sum_ = w.Value();
+    c_ = 0.0;
+  }
+
+ private:
+  double sum_;
+  double c_;   // Kahan compensation.
+};
+
+// MinMax semiring: (min, max, inf, -inf).
+template <class T>
+class MinMaxWeightTpl : public FloatWeightTpl<T> {
+ public:
+  using typename FloatWeightTpl<T>::ValueType;
+  using FloatWeightTpl<T>::Value;
+  using ReverseWeight = MinMaxWeightTpl<T>;
+  using Limits = FloatLimits<T>;
+
+  MinMaxWeightTpl() noexcept : FloatWeightTpl<T>() {}
+
+  constexpr MinMaxWeightTpl(T f) : FloatWeightTpl<T>(f) {}  // NOLINT
+
+  static constexpr MinMaxWeightTpl Zero() { return Limits::PosInfinity(); }
+
+  static constexpr MinMaxWeightTpl One() { return Limits::NegInfinity(); }
+
+  static constexpr MinMaxWeightTpl NoWeight() { return Limits::NumberBad(); }
+
+  static const string &Type() {
+    static const string *const type =
+        new string(string("minmax") + FloatWeightTpl<T>::GetPrecisionString());
+    return *type;
+  }
+
+  // Fails for IEEE NaN.
+  constexpr bool Member() const { return Value() == Value(); }
+
+  MinMaxWeightTpl<T> Quantize(float delta = kDelta) const {
+    // If one of infinities, or a NaN.
+    if (!Member() ||
+        Value() == Limits::NegInfinity() || Value() == Limits::PosInfinity()) {
+      return *this;
+    } else {
+      return MinMaxWeightTpl<T>(std::floor(Value() / delta + 0.5F) * delta);
+    }
+  }
+
+  constexpr MinMaxWeightTpl<T> Reverse() const { return *this; }
+
+  static constexpr uint64 Properties() {
+    return kLeftSemiring | kRightSemiring | kCommutative | kIdempotent | kPath;
+  }
+};
+
+// Single-precision min-max weight.
+using MinMaxWeight = MinMaxWeightTpl<float>;
+
+// Min.
+template <class T>
+constexpr MinMaxWeightTpl<T> Plus(const MinMaxWeightTpl<T> &w1,
+                               const MinMaxWeightTpl<T> &w2) {
+  return (!w1.Member() || !w2.Member()) ? MinMaxWeightTpl<T>::NoWeight()
+      : w1.Value() < w2.Value() ? w1 : w2;
+}
+
+constexpr MinMaxWeightTpl<float> Plus(const MinMaxWeightTpl<float> &w1,
+                                      const MinMaxWeightTpl<float> &w2) {
+  return Plus<float>(w1, w2);
+}
+
+constexpr MinMaxWeightTpl<double> Plus(const MinMaxWeightTpl<double> &w1,
+                                       const MinMaxWeightTpl<double> &w2) {
+  return Plus<double>(w1, w2);
+}
+
+// Max.
+template <class T>
+constexpr MinMaxWeightTpl<T> Times(const MinMaxWeightTpl<T> &w1,
+                                   const MinMaxWeightTpl<T> &w2) {
+  return (!w1.Member() || !w2.Member()) ? MinMaxWeightTpl<T>::NoWeight()
+      : w1.Value() >= w2.Value() ? w1 : w2;
+}
+
+constexpr MinMaxWeightTpl<float> Times(const MinMaxWeightTpl<float> &w1,
+                                       const MinMaxWeightTpl<float> &w2) {
+  return Times<float>(w1, w2);
+}
+
+constexpr MinMaxWeightTpl<double> Times(const MinMaxWeightTpl<double> &w1,
+                                        const MinMaxWeightTpl<double> &w2) {
+  return Times<double>(w1, w2);
+}
+
+// Defined only for special cases.
+template <class T>
+constexpr MinMaxWeightTpl<T> Divide(const MinMaxWeightTpl<T> &w1,
+                                    const MinMaxWeightTpl<T> &w2,
+                                    DivideType typ = DIVIDE_ANY) {
+  return w1.Value() >= w2.Value() ? w1 : MinMaxWeightTpl<T>::NoWeight();
+}
+
+constexpr MinMaxWeightTpl<float> Divide(const MinMaxWeightTpl<float> &w1,
+                                        const MinMaxWeightTpl<float> &w2,
+                                        DivideType typ = DIVIDE_ANY) {
+  return Divide<float>(w1, w2, typ);
+}
+
+constexpr MinMaxWeightTpl<double> Divide(const MinMaxWeightTpl<double> &w1,
+                                         const MinMaxWeightTpl<double> &w2,
+                                         DivideType typ = DIVIDE_ANY) {
+  return Divide<double>(w1, w2, typ);
+}
+
+// Converts to tropical.
+template <>
+struct WeightConvert<LogWeight, TropicalWeight> {
+  constexpr TropicalWeight operator()(const LogWeight &w) const {
+    return w.Value();
+  }
+};
+
+template <>
+struct WeightConvert<Log64Weight, TropicalWeight> {
+  constexpr TropicalWeight operator()(const Log64Weight &w) const {
+    return w.Value();
+  }
+};
+
+// Converts to log.
+template <>
+struct WeightConvert<TropicalWeight, LogWeight> {
+  constexpr LogWeight operator()(const TropicalWeight &w) const {
+    return w.Value();
+  }
+};
+
+template <>
+struct WeightConvert<Log64Weight, LogWeight> {
+  constexpr LogWeight operator()(const Log64Weight &w) const {
+    return w.Value();
+  }
+};
+
+// Converts to log64.
+template <>
+struct WeightConvert<TropicalWeight, Log64Weight> {
+  constexpr Log64Weight operator()(const TropicalWeight &w) const {
+    return w.Value();
+  }
+};
+
+template <>
+struct WeightConvert<LogWeight, Log64Weight> {
+  constexpr Log64Weight operator()(const LogWeight &w) const {
+    return w.Value();
+  }
+};
+
+// This function object returns random integers chosen from [0,
+// num_random_weights). The boolean 'allow_zero' determines whether Zero() and
+// zero divisors should be returned in the random weight generation. This is
+// intended primary for testing.
+template <class Weight>
+class FloatWeightGenerate {
+ public:
+  explicit FloatWeightGenerate(
+      bool allow_zero = true,
+      const size_t num_random_weights = kNumRandomWeights)
+      : allow_zero_(allow_zero), num_random_weights_(num_random_weights) {}
+
+  Weight operator()() const {
+    const int n = rand() % (num_random_weights_ + allow_zero_);  // NOLINT
+    if (allow_zero_ && n == num_random_weights_) return Weight::Zero();
+    return Weight(n);
+  }
+
+ private:
+  // Permits Zero() and zero divisors.
+  const bool allow_zero_;
+  // Number of alternative random weights.
+  const size_t num_random_weights_;
+};
+
+template <class T>
+class WeightGenerate<TropicalWeightTpl<T>>
+    : public FloatWeightGenerate<TropicalWeightTpl<T>> {
+ public:
+  using Weight = TropicalWeightTpl<T>;
+  using Generate = FloatWeightGenerate<Weight>;
+
+  explicit WeightGenerate(bool allow_zero = true,
+                          size_t num_random_weights = kNumRandomWeights)
+      : Generate(allow_zero, num_random_weights) {}
+
+  Weight operator()() const { return Weight(Generate::operator()()); }
+};
+
+template <class T>
+class WeightGenerate<LogWeightTpl<T>>
+    : public FloatWeightGenerate<LogWeightTpl<T>> {
+ public:
+  using Weight = LogWeightTpl<T>;
+  using Generate = FloatWeightGenerate<Weight>;
+
+  explicit WeightGenerate(bool allow_zero = true,
+                          size_t num_random_weights = kNumRandomWeights)
+      : Generate(allow_zero, num_random_weights) {}
+
+  Weight operator()() const { return Weight(Generate::operator()()); }
+};
+
+// This function object returns random integers chosen from [0,
+// num_random_weights). The boolean 'allow_zero' determines whether Zero() and
+// zero divisors should be returned in the random weight generation. This is
+// intended primary for testing.
+template <class T>
+class WeightGenerate<MinMaxWeightTpl<T>> {
+ public:
+  using Weight = MinMaxWeightTpl<T>;
+
+  explicit WeightGenerate(bool allow_zero = true,
+                          size_t num_random_weights = kNumRandomWeights)
+      : allow_zero_(allow_zero), num_random_weights_(num_random_weights) {}
+
+  Weight operator()() const {
+    const int n = (rand() %  // NOLINT
+                   (2 * num_random_weights_ + allow_zero_)) -
+                  num_random_weights_;
+    if (allow_zero_ && n == num_random_weights_) {
+      return Weight::Zero();
+    } else if (n == -num_random_weights_) {
+      return Weight::One();
+    } else {
+      return Weight(n);
+    }
+  }
+
+ private:
+  // Permits Zero() and zero divisors.
+  const bool allow_zero_;
+  // Number of alternative random weights.
+  const size_t num_random_weights_;
+};
+
+}  // namespace fst
+
+#endif  // FST_FLOAT_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/fst-decl.h b/projects/llm_framework/include/fst/fst-decl.h
new file mode 100644
index 00000000..9319dfc9
--- /dev/null
+++ b/projects/llm_framework/include/fst/fst-decl.h
@@ -0,0 +1,254 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// This file contains declarations of classes in the Fst template library.
+
+#ifndef FST_FST_DECL_H_
+#define FST_FST_DECL_H_
+
+#include <sys/types.h>
+#include <memory>  // for allocator<>
+
+#include <fst/types.h>
+
+namespace fst {
+
+// Symbol table and iterator.
+
+class SymbolTable;
+
+class SymbolTableIterator;
+
+// Weight templates and weights.
+
+template <class T>
+class FloatWeightTpl;
+
+template <class T>
+class TropicalWeightTpl;
+
+template <class T>
+class LogWeightTpl;
+
+template <class T>
+class MinMaxWeightTpl;
+
+using FloatWeight = FloatWeightTpl<float>;
+
+using TropicalWeight = TropicalWeightTpl<float>;
+
+using LogWeight = LogWeightTpl<float>;
+
+using MinMaxWeight = MinMaxWeightTpl<float>;
+
+// Arc templates and arcs.
+
+template <class Weight>
+struct ArcTpl;
+
+using StdArc = ArcTpl<TropicalWeight>;
+
+using LogArc = ArcTpl<LogWeight>;
+
+// Stores.
+
+template <class Element, class U>
+class DefaultCompactStore;
+
+template <class Arc>
+class DefaultCacheStore;
+
+// FST templates.
+
+template <class A, class ArcCompactor, class Unsigned = uint32,
+    class CompactStore = DefaultCompactStore<typename ArcCompactor::Element, Unsigned>,
+    class CacheStore = DefaultCacheStore<A>>
+class CompactFst;
+
+template <class Arc, class U = uint32>
+class ConstFst;
+
+template <class Arc, class Weight, class Matcher>
+class EditFst;
+
+template <class Arc>
+class ExpandedFst;
+
+template <class Arc>
+class Fst;
+
+template <class Arc>
+class MutableFst;
+
+template <class A, class Allocator = std::allocator<A>>
+class VectorState;
+
+template <class A, class State = VectorState<A>>
+class VectorFst;
+
+template <class Arc, class U = ssize_t>
+class DefaultReplaceStateTable;
+
+// On-the-fly operations.
+
+template <class Arc, class Compare>
+class ArcSortFst;
+
+template <class Arc>
+class ClosureFst;
+
+template <class A, class Store = DefaultCacheStore<A>>
+class ComposeFst;
+
+template <class Arc>
+class ConcatFst;
+
+template <class Arc>
+class DeterminizeFst;
+
+template <class Arc>
+class DifferenceFst;
+
+template <class Arc>
+class IntersectFst;
+
+template <class Arc>
+class InvertFst;
+
+template <class AArc, class BArc, class Mapper>
+class ArcMapFst;
+
+template <class Arc>
+class ProjectFst;
+
+template <class AArc, class BArc, class Selector>
+class RandGenFst;
+
+template <class Arc>
+class RelabelFst;
+
+template <class A, class StateTable = DefaultReplaceStateTable<A>,
+          class Store = DefaultCacheStore<A>>
+class ReplaceFst;
+
+template <class Arc>
+class RmEpsilonFst;
+
+template <class Arc>
+class UnionFst;
+
+// Heap.
+
+template <class T, class Compare>
+class Heap;
+
+// Compactors.
+
+template <class Arc>
+class AcceptorCompactor;
+
+template <class Arc>
+class StringCompactor;
+
+template <class Arc>
+class UnweightedAcceptorCompactor;
+
+template <class Arc>
+class UnweightedCompactor;
+
+template <class Arc>
+class WeightedStringCompactor;
+
+// Compact FSTs.
+
+template <class Arc, class U = uint32>
+using CompactStringFst = CompactFst<Arc, StringCompactor<Arc>, U>;
+
+template <class Arc, class U = uint32>
+using CompactWeightedStringFst =
+    CompactFst<Arc, WeightedStringCompactor<Arc>, U>;
+
+template <class Arc, class U = uint32>
+using CompactAcceptorFst = CompactFst<Arc, AcceptorCompactor<Arc>, U>;
+
+template <class Arc, class U = uint32>
+using CompactUnweightedFst = CompactFst<Arc, UnweightedCompactor<Arc>, U>;
+
+template <class Arc, class U = uint32>
+using CompactUnweightedAcceptorFst =
+    CompactFst<Arc, UnweightedAcceptorCompactor<Arc>, U>;
+
+// StdArc aliases for FSTs.
+
+using StdConstFst = ConstFst<StdArc>;
+using StdExpandedFst = ExpandedFst<StdArc>;
+using StdFst = Fst<StdArc>;
+using StdMutableFst = MutableFst<StdArc>;
+using StdVectorFst = VectorFst<StdArc>;
+
+// StdArc aliases for on-the-fly operations.
+
+template <class Compare>
+using StdArcSortFst = ArcSortFst<StdArc, Compare>;
+
+using StdClosureFst = ClosureFst<StdArc>;
+
+using StdComposeFst = ComposeFst<StdArc>;
+
+using StdConcatFst = ConcatFst<StdArc>;
+
+using StdDeterminizeFst = DeterminizeFst<StdArc>;
+
+using StdDifferenceFst = DifferenceFst<StdArc>;
+
+using StdIntersectFst = IntersectFst<StdArc>;
+
+using StdInvertFst = InvertFst<StdArc>;
+
+using StdProjectFst = ProjectFst<StdArc>;
+
+using StdRelabelFst = RelabelFst<StdArc>;
+
+using StdReplaceFst = ReplaceFst<StdArc>;
+
+using StdRmEpsilonFst = RmEpsilonFst<StdArc>;
+
+using StdUnionFst = UnionFst<StdArc>;
+
+// Filter states.
+
+template <class T>
+class IntegerFilterState;
+
+using CharFilterState = IntegerFilterState<signed char>;
+
+using ShortFilterState = IntegerFilterState<short>;  // NOLINT
+
+using IntFilterState = IntegerFilterState<int>;
+
+// Matchers and filters.
+
+template <class FST>
+class Matcher;
+
+template <class M1, class M2 = M1>
+class NullComposeFilter;
+
+template <class M1, class M2 = M1>
+class TrivialComposeFilter;
+
+template <class M1, class M2 = M1>
+class SequenceComposeFilter;
+
+template <class M1, class M2 = M1>
+class AltSequenceComposeFilter;
+
+template <class M1, class M2 = M1>
+class MatchComposeFilter;
+
+template <class M1, class M2 = M1>
+class NoMatchComposeFilter;
+
+}  // namespace fst
+
+#endif  // FST_FST_DECL_H_
diff --git a/projects/llm_framework/include/fst/fst.h b/projects/llm_framework/include/fst/fst.h
new file mode 100644
index 00000000..20e6bb3c
--- /dev/null
+++ b/projects/llm_framework/include/fst/fst.h
@@ -0,0 +1,1007 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// FST abstract base class definition, state and arc iterator interface, and
+// suggested base implementation.
+
+#ifndef FST_FST_H_
+#define FST_FST_H_
+
+#include <sys/types.h>
+
+#include <cmath>
+#include <cstddef>
+
+#include <iostream>
+#include <memory>
+#include <sstream>
+#include <string>
+#include <utility>
+
+#include <fst/compat.h>
+#include <fst/flags.h>
+#include <fst/log.h>
+#include <fstream>
+
+#include <fst/arc.h>
+#include <fst/memory.h>
+#include <fst/properties.h>
+#include <fst/register.h>
+#include <fst/symbol-table.h>
+#include <fst/util.h>
+
+
+DECLARE_bool(fst_align);
+
+namespace fst {
+
+bool IsFstHeader(std::istream &, const string &);
+
+class FstHeader;
+
+template <class Arc>
+struct StateIteratorData;
+
+template <class Arc>
+struct ArcIteratorData;
+
+template <class Arc>
+class MatcherBase;
+
+struct FstReadOptions {
+  // FileReadMode(s) are advisory, there are many conditions than prevent a
+  // file from being mapped, READ mode will be selected in these cases with
+  // a warning indicating why it was chosen.
+  enum FileReadMode { READ, MAP };
+
+  string source;                // Where you're reading from.
+  const FstHeader *header;      // Pointer to FST header; if non-zero, use
+                                // this info (don't read a stream header).
+  const SymbolTable *isymbols;  // Pointer to input symbols; if non-zero, use
+                                // this info (read and skip stream isymbols)
+  const SymbolTable *osymbols;  // Pointer to output symbols; if non-zero, use
+                                // this info (read and skip stream osymbols)
+  FileReadMode mode;            // Read or map files (advisory, if possible)
+  bool read_isymbols;           // Read isymbols, if any (default: true).
+  bool read_osymbols;           // Read osymbols, if any (default: true).
+
+  explicit FstReadOptions(const string &source = "<unspecified>",
+                          const FstHeader *header = nullptr,
+                          const SymbolTable *isymbols = nullptr,
+                          const SymbolTable *osymbols = nullptr);
+
+  explicit FstReadOptions(const string &source, const SymbolTable *isymbols,
+                          const SymbolTable *osymbols = nullptr);
+
+  // Helper function to convert strings FileReadModes into their enum value.
+  static FileReadMode ReadMode(const string &mode);
+
+  // Outputs a debug string for the FstReadOptions object.
+  string DebugString() const;
+};
+
+struct FstWriteOptions {
+  string source;        // Where you're writing to.
+  bool write_header;    // Write the header?
+  bool write_isymbols;  // Write input symbols?
+  bool write_osymbols;  // Write output symbols?
+  bool align;           // Write data aligned (may fail on pipes)?
+  bool stream_write;    // Avoid seek operations in writing.
+
+  explicit FstWriteOptions(const string &source = "<unspecifed>",
+                           bool write_header = true, bool write_isymbols = true,
+                           bool write_osymbols = true,
+                           bool align = FLAGS_fst_align,
+                           bool stream_write = false)
+      : source(source),
+        write_header(write_header),
+        write_isymbols(write_isymbols),
+        write_osymbols(write_osymbols),
+        align(align),
+        stream_write(stream_write) {}
+};
+
+// Header class.
+//
+// This is the recommended file header representation.
+
+class FstHeader {
+ public:
+  enum {
+    HAS_ISYMBOLS = 0x1,  // Has input symbol table.
+    HAS_OSYMBOLS = 0x2,  // Has output symbol table.
+    IS_ALIGNED = 0x4,    // Memory-aligned (where appropriate).
+  } Flags;
+
+  FstHeader() : version_(0), flags_(0), properties_(0), start_(-1),
+      numstates_(0), numarcs_(0) {}
+
+  const string &FstType() const { return fsttype_; }
+
+  const string &ArcType() const { return arctype_; }
+
+  int32 Version() const { return version_; }
+
+  int32 GetFlags() const { return flags_; }
+
+  uint64 Properties() const { return properties_; }
+
+  int64 Start() const { return start_; }
+
+  int64 NumStates() const { return numstates_; }
+
+  int64 NumArcs() const { return numarcs_; }
+
+  void SetFstType(const string &type) { fsttype_ = type; }
+
+  void SetArcType(const string &type) { arctype_ = type; }
+
+  void SetVersion(int32 version) { version_ = version; }
+
+  void SetFlags(int32 flags) { flags_ = flags; }
+
+  void SetProperties(uint64 properties) { properties_ = properties; }
+
+  void SetStart(int64 start) { start_ = start; }
+
+  void SetNumStates(int64 numstates) { numstates_ = numstates; }
+
+  void SetNumArcs(int64 numarcs) { numarcs_ = numarcs; }
+
+  bool Read(std::istream &strm, const string &source,
+            bool rewind = false);
+
+  bool Write(std::ostream &strm, const string &source) const;
+
+  // Outputs a debug string for the FstHeader object.
+  string DebugString() const;
+
+ private:
+  string fsttype_;     // E.g. "vector".
+  string arctype_;     // E.g. "standard".
+  int32 version_;      // Type version number.
+  int32 flags_;        // File format bits.
+  uint64 properties_;  // FST property bits.
+  int64 start_;        // Start state.
+  int64 numstates_;    // # of states.
+  int64 numarcs_;      // # of arcs.
+};
+
+// Specifies matcher action.
+enum MatchType {
+  MATCH_INPUT = 1,   // Match input label.
+  MATCH_OUTPUT = 2,  // Match output label.
+  MATCH_BOTH = 3,    // Match input or output label.
+  MATCH_NONE = 4,    // Match nothing.
+  MATCH_UNKNOWN = 5
+};  // Otherwise, match type unknown.
+
+constexpr int kNoLabel = -1;    // Not a valid label.
+constexpr int kNoStateId = -1;  // Not a valid state ID.
+
+// A generic FST, templated on the arc definition, with common-demoninator
+// methods (use StateIterator and ArcIterator to iterate over its states and
+// arcs).
+template <class A>
+class Fst {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  virtual ~Fst() {}
+
+  // Initial state.
+  virtual StateId Start() const = 0;
+
+  // State's final weight.
+  virtual Weight Final(StateId) const = 0;
+
+  // State's arc count.
+  virtual size_t NumArcs(StateId) const = 0;
+
+  // State's input epsilon count.
+  virtual size_t NumInputEpsilons(StateId) const = 0;
+
+  // State's output epsilon count.
+  virtual size_t NumOutputEpsilons(StateId) const = 0;
+
+  // Property bits. If test = false, return stored properties bits for mask
+  // (some possibly unknown); if test = true, return property bits for mask
+  // (computing o.w. unknown).
+  virtual uint64 Properties(uint64 mask, bool test) const = 0;
+
+  // FST type name.
+  virtual const string &Type() const = 0;
+
+  // Gets a copy of this Fst. The copying behaves as follows:
+  //
+  // (1) The copying is constant time if safe = false or if safe = true
+  // and is on an otherwise unaccessed FST.
+  //
+  // (2) If safe = true, the copy is thread-safe in that the original
+  // and copy can be safely accessed (but not necessarily mutated) by
+  // separate threads. For some FST types, 'Copy(true)' should only be
+  // called on an FST that has not otherwise been accessed. Behavior is
+  // otherwise undefined.
+  //
+  // (3) If a MutableFst is copied and then mutated, then the original is
+  // unmodified and vice versa (often by a copy-on-write on the initial
+  // mutation, which may not be constant time).
+  virtual Fst<Arc> *Copy(bool safe = false) const = 0;
+
+  // Reads an FST from an input stream; returns nullptr on error.
+  static Fst<Arc> *Read(std::istream &strm, const FstReadOptions &opts) {
+    FstReadOptions ropts(opts);
+    FstHeader hdr;
+    if (ropts.header) {
+      hdr = *opts.header;
+    } else {
+      if (!hdr.Read(strm, opts.source)) return nullptr;
+      ropts.header = &hdr;
+    }
+    const auto &fst_type = hdr.FstType();
+    const auto reader = FstRegister<Arc>::GetRegister()->GetReader(fst_type);
+    if (!reader) {
+      LOG(ERROR) << "Fst::Read: Unknown FST type " << fst_type
+                 << " (arc type = " << Arc::Type() << "): " << ropts.source;
+      return nullptr;
+    }
+    return reader(strm, ropts);
+  }
+
+  // Reads an FST from a file; returns nullptr on error. An empty filename
+  // results in reading from standard input.
+  static Fst<Arc> *Read(const string &filename) {
+    if (!filename.empty()) {
+      std::ifstream strm(filename,
+                              std::ios_base::in | std::ios_base::binary);
+      if (!strm) {
+        LOG(ERROR) << "Fst::Read: Can't open file: " << filename;
+        return nullptr;
+      }
+      return Read(strm, FstReadOptions(filename));
+    } else {
+      return Read(std::cin, FstReadOptions("standard input"));
+    }
+  }
+
+  // Writes an FST to an output stream; returns false on error.
+  virtual bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
+    LOG(ERROR) << "Fst::Write: No write stream method for " << Type()
+               << " FST type";
+    return false;
+  }
+
+  // Writes an FST to a file; returns false on error; an empty filename
+  // results in writing to standard output.
+  virtual bool Write(const string &filename) const {
+    LOG(ERROR) << "Fst::Write: No write filename method for " << Type()
+               << " FST type";
+    return false;
+  }
+
+  // Returns input label symbol table; return nullptr if not specified.
+  virtual const SymbolTable *InputSymbols() const = 0;
+
+  // Return output label symbol table; return nullptr if not specified.
+  virtual const SymbolTable *OutputSymbols() const = 0;
+
+  // For generic state iterator construction (not normally called directly by
+  // users). Does not copy the FST.
+  virtual void InitStateIterator(StateIteratorData<Arc> *data) const = 0;
+
+  // For generic arc iterator construction (not normally called directly by
+  // users). Does not copy the FST.
+  virtual void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const = 0;
+
+  // For generic matcher construction (not normally called directly by users).
+  // Does not copy the FST.
+  virtual MatcherBase<Arc> *InitMatcher(MatchType match_type) const;
+
+ protected:
+  bool WriteFile(const string &filename) const {
+    if (!filename.empty()) {
+      std::ofstream strm(filename,
+                               std::ios_base::out | std::ios_base::binary);
+      if (!strm) {
+        LOG(ERROR) << "Fst::Write: Can't open file: " << filename;
+        return false;
+      }
+      bool val = Write(strm, FstWriteOptions(filename));
+      if (!val) LOG(ERROR) << "Fst::Write failed: " << filename;
+      return val;
+    } else {
+      return Write(std::cout, FstWriteOptions("standard output"));
+    }
+  }
+};
+
+// A useful alias when using StdArc.
+using StdFst = Fst<StdArc>;
+
+// State and arc iterator definitions.
+//
+// State iterator interface templated on the Arc definition; used for
+// StateIterator specializations returned by the InitStateIterator FST method.
+template <class Arc>
+class StateIteratorBase {
+ public:
+  using StateId = typename Arc::StateId;
+
+  virtual ~StateIteratorBase() {}
+
+  // End of iterator?
+  virtual bool Done() const = 0;
+  // Returns current state (when !Done()).
+  virtual StateId Value() const = 0;
+  // Advances to next state (when !Done()).
+  virtual void Next() = 0;
+  // Resets to initial condition.
+  virtual void Reset() = 0;
+};
+
+// StateIterator initialization data.
+
+template <class Arc>
+struct StateIteratorData {
+  using StateId = typename Arc::StateId;
+
+  // Specialized iterator if non-zero.
+  StateIteratorBase<Arc> *base;
+  // Otherwise, the total number of states.
+  StateId nstates;
+
+  StateIteratorData() : base(nullptr), nstates(0) {}
+
+  StateIteratorData(const StateIteratorData &) = delete;
+  StateIteratorData &operator=(const StateIteratorData &) = delete;
+};
+
+// Generic state iterator, templated on the FST definition (a wrapper
+// around a pointer to a specific one). Here is a typical use:
+//
+//   for (StateIterator<StdFst> siter(fst);
+//        !siter.Done();
+//        siter.Next()) {
+//     StateId s = siter.Value();
+//     ...
+//   }
+// There is no copying of the FST.
+template <class FST>
+class StateIterator {
+ public:
+  using Arc = typename FST::Arc;
+  using StateId = typename Arc::StateId;
+
+  explicit StateIterator(const FST &fst) : s_(0) {
+    fst.InitStateIterator(&data_);
+  }
+
+  ~StateIterator() { delete data_.base; }
+
+  bool Done() const {
+    return data_.base ? data_.base->Done() : s_ >= data_.nstates;
+  }
+
+  StateId Value() const { return data_.base ? data_.base->Value() : s_; }
+
+  void Next() {
+    if (data_.base) {
+      data_.base->Next();
+    } else {
+      ++s_;
+    }
+  }
+
+  void Reset() {
+    if (data_.base) {
+      data_.base->Reset();
+    } else {
+      s_ = 0;
+    }
+  }
+
+ private:
+  StateIteratorData<Arc> data_;
+  StateId s_;
+};
+
+// Flags to control the behavior on an arc iterator.
+static constexpr uint32 kArcILabelValue =
+    0x0001;  // Value() gives valid ilabel.
+static constexpr uint32 kArcOLabelValue = 0x0002;  //  "       "     " olabel.
+static constexpr uint32 kArcWeightValue = 0x0004;  //  "       "     " weight.
+static constexpr uint32 kArcNextStateValue =
+    0x0008;                                    //  "       "     " nextstate.
+static constexpr uint32 kArcNoCache = 0x0010;  // No need to cache arcs.
+
+static constexpr uint32 kArcValueFlags =
+    kArcILabelValue | kArcOLabelValue | kArcWeightValue | kArcNextStateValue;
+
+static constexpr uint32 kArcFlags = kArcValueFlags | kArcNoCache;
+
+// Arc iterator interface, templated on the arc definition; used for arc
+// iterator specializations that are returned by the InitArcIterator FST method.
+template <class Arc>
+class ArcIteratorBase {
+ public:
+  using StateId = typename Arc::StateId;
+
+  virtual ~ArcIteratorBase() {}
+
+  // End of iterator?
+  virtual bool Done() const = 0;
+  // Returns current arc (when !Done()).
+  virtual const Arc &Value() const = 0;
+  // Advances to next arc (when !Done()).
+  virtual void Next() = 0;
+  // Returns current position.
+  virtual size_t Position() const = 0;
+  // Returns to initial condition.
+  virtual void Reset() = 0;
+  // Advances to arbitrary arc by position.
+  virtual void Seek(size_t) = 0;
+  // Returns current behavorial flags
+  virtual uint32 Flags() const = 0;
+  // Sets behavorial flags.
+  virtual void SetFlags(uint32, uint32) = 0;
+};
+
+// ArcIterator initialization data.
+template <class Arc>
+struct ArcIteratorData {
+  ArcIteratorData()
+      : base(nullptr), arcs(nullptr), narcs(0), ref_count(nullptr) {}
+
+  ArcIteratorData(const ArcIteratorData &) = delete;
+
+  ArcIteratorData &operator=(const ArcIteratorData &) = delete;
+
+  ArcIteratorBase<Arc> *base;  // Specialized iterator if non-zero.
+  const Arc *arcs;             // O.w. arcs pointer
+  size_t narcs;                // ... and arc count.
+  int *ref_count;              // ... and reference count if non-zero.
+};
+
+// Generic arc iterator, templated on the FST definition (a wrapper around a
+// pointer to a specific one). Here is a typical use:
+//
+//   for (ArcIterator<StdFst> aiter(fst, s);
+//        !aiter.Done();
+//         aiter.Next()) {
+//     StdArc &arc = aiter.Value();
+//     ...
+//   }
+// There is no copying of the FST.
+template <class FST>
+class ArcIterator {
+ public:
+  using Arc = typename FST::Arc;
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const FST &fst, StateId s) : i_(0) {
+    fst.InitArcIterator(s, &data_);
+  }
+
+  explicit ArcIterator(const ArcIteratorData<Arc> &data) : data_(data), i_(0) {
+    if (data_.ref_count) ++(*data_.ref_count);
+  }
+
+  ~ArcIterator() {
+    if (data_.base) {
+      delete data_.base;
+    } else if (data_.ref_count) {
+      --(*data_.ref_count);
+    }
+  }
+
+  bool Done() const {
+    return data_.base ? data_.base->Done() : i_ >= data_.narcs;
+  }
+
+  const Arc &Value() const {
+    return data_.base ? data_.base->Value() : data_.arcs[i_];
+  }
+
+  void Next() {
+    if (data_.base) {
+      data_.base->Next();
+    } else {
+      ++i_;
+    }
+  }
+
+  void Reset() {
+    if (data_.base) {
+      data_.base->Reset();
+    } else {
+      i_ = 0;
+    }
+  }
+
+  void Seek(size_t a) {
+    if (data_.base) {
+      data_.base->Seek(a);
+    } else {
+      i_ = a;
+    }
+  }
+
+  size_t Position() const { return data_.base ? data_.base->Position() : i_; }
+
+  uint32 Flags() const {
+    if (data_.base) {
+      return data_.base->Flags();
+    } else {
+      return kArcValueFlags;
+    }
+  }
+
+  void SetFlags(uint32 flags, uint32 mask) {
+    if (data_.base) data_.base->SetFlags(flags, mask);
+  }
+
+ private:
+  ArcIteratorData<Arc> data_;
+  size_t i_;
+};
+
+}  // namespace fst
+
+// ArcIterator placement operator new and destroy function; new needs to be in
+// the global namespace.
+
+template <class FST>
+void *operator new(size_t size,
+                   fst::MemoryPool<fst::ArcIterator<FST>> *pool) {
+  return pool->Allocate();
+}
+
+namespace fst {
+
+template <class FST>
+void Destroy(ArcIterator<FST> *aiter, MemoryPool<ArcIterator<FST>> *pool) {
+  if (aiter) {
+    aiter->~ArcIterator<FST>();
+    pool->Free(aiter);
+  }
+}
+
+// Matcher definitions.
+
+template <class Arc>
+MatcherBase<Arc> *Fst<Arc>::InitMatcher(MatchType match_type) const {
+  return nullptr;  // One should just use the default matcher.
+}
+
+// FST accessors, useful in high-performance applications.
+
+namespace internal {
+
+// General case, requires non-abstract, 'final' methods. Use for inlining.
+
+template <class F>
+inline typename F::Arc::Weight Final(const F &fst, typename F::Arc::StateId s) {
+  return fst.F::Final(s);
+}
+
+template <class F>
+inline ssize_t NumArcs(const F &fst, typename F::Arc::StateId s) {
+  return fst.F::NumArcs(s);
+}
+
+template <class F>
+inline ssize_t NumInputEpsilons(const F &fst, typename F::Arc::StateId s) {
+  return fst.F::NumInputEpsilons(s);
+}
+
+template <class F>
+inline ssize_t NumOutputEpsilons(const F &fst, typename F::Arc::StateId s) {
+  return fst.F::NumOutputEpsilons(s);
+}
+
+// Fst<Arc> case, abstract methods.
+
+template <class Arc>
+inline typename Arc::Weight Final(const Fst<Arc> &fst,
+                                  typename Arc::StateId s) {
+  return fst.Final(s);
+}
+
+template <class Arc>
+inline size_t NumArcs(const Fst<Arc> &fst, typename Arc::StateId s) {
+  return fst.NumArcs(s);
+}
+
+template <class Arc>
+inline size_t NumInputEpsilons(const Fst<Arc> &fst, typename Arc::StateId s) {
+  return fst.NumInputEpsilons(s);
+}
+
+template <class Arc>
+inline size_t NumOutputEpsilons(const Fst<Arc> &fst, typename Arc::StateId s) {
+  return fst.NumOutputEpsilons(s);
+}
+
+// FST implementation base.
+//
+// This is the recommended FST implementation base class. It will handle
+// reference counts, property bits, type information and symbols.
+//
+// Users are discouraged, but not prohibited, from subclassing this outside the
+// FST library.
+template <class Arc>
+class FstImpl {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  FstImpl() : properties_(0), type_("null") {}
+
+  FstImpl(const FstImpl<Arc> &impl)
+      : properties_(impl.properties_),
+        type_(impl.type_),
+        isymbols_(impl.isymbols_ ? impl.isymbols_->Copy() : nullptr),
+        osymbols_(impl.osymbols_ ? impl.osymbols_->Copy() : nullptr) {}
+
+  FstImpl(FstImpl<Arc> &&impl) noexcept;
+
+  virtual ~FstImpl() {}
+
+  FstImpl &operator=(const FstImpl<Arc> &impl) {
+    properties_ = impl.properties_;
+    type_ = impl.type_;
+    isymbols_ = impl.isymbols_ ? impl.isymbols_->Copy() : nullptr;
+    osymbols_ = impl.osymbols_ ? impl.osymbols_->Copy() : nullptr;
+    return *this;
+  }
+
+  FstImpl &operator=(FstImpl<Arc> &&impl) noexcept;
+
+  const string &Type() const { return type_; }
+
+  void SetType(const string &type) { type_ = type; }
+
+  virtual uint64 Properties() const { return properties_; }
+
+  virtual uint64 Properties(uint64 mask) const { return properties_ & mask; }
+
+  void SetProperties(uint64 props) {
+    properties_ &= kError;  // kError can't be cleared.
+    properties_ |= props;
+  }
+
+  void SetProperties(uint64 props, uint64 mask) {
+    properties_ &= ~mask | kError;  // kError can't be cleared.
+    properties_ |= props & mask;
+  }
+
+  // Allows (only) setting error bit on const FST implementations.
+  void SetProperties(uint64 props, uint64 mask) const {
+    if (mask != kError) {
+      FSTERROR() << "FstImpl::SetProperties() const: Can only set kError";
+    }
+    properties_ |= kError;
+  }
+
+  const SymbolTable *InputSymbols() const { return isymbols_.get(); }
+
+  const SymbolTable *OutputSymbols() const { return osymbols_.get(); }
+
+  SymbolTable *InputSymbols() { return isymbols_.get(); }
+
+  SymbolTable *OutputSymbols() { return osymbols_.get(); }
+
+  void SetInputSymbols(const SymbolTable *isyms) {
+    isymbols_.reset(isyms ? isyms->Copy() : nullptr);
+  }
+
+  void SetOutputSymbols(const SymbolTable *osyms) {
+    osymbols_.reset(osyms ? osyms->Copy() : nullptr);
+  }
+
+  // Reads header and symbols from input stream, initializes FST, and returns
+  // the header. If opts.header is non-null, skips reading and uses the option
+  // value instead. If opts.[io]symbols is non-null, reads in (if present), but
+  // uses the option value.
+  bool ReadHeader(std::istream &strm, const FstReadOptions &opts,
+                  int min_version, FstHeader *hdr);
+
+  // Writes header and symbols to output stream. If opts.header is false, skips
+  // writing header. If opts.[io]symbols is false, skips writing those symbols.
+  // This method is needed for implementations that implement Write methods.
+  void WriteHeader(std::ostream &strm, const FstWriteOptions &opts,
+                   int version, FstHeader *hdr) const {
+    if (opts.write_header) {
+      hdr->SetFstType(type_);
+      hdr->SetArcType(Arc::Type());
+      hdr->SetVersion(version);
+      hdr->SetProperties(properties_);
+      int32 file_flags = 0;
+      if (isymbols_ && opts.write_isymbols) {
+        file_flags |= FstHeader::HAS_ISYMBOLS;
+      }
+      if (osymbols_ && opts.write_osymbols) {
+        file_flags |= FstHeader::HAS_OSYMBOLS;
+      }
+      if (opts.align) file_flags |= FstHeader::IS_ALIGNED;
+      hdr->SetFlags(file_flags);
+      hdr->Write(strm, opts.source);
+    }
+    if (isymbols_ && opts.write_isymbols) isymbols_->Write(strm);
+    if (osymbols_ && opts.write_osymbols) osymbols_->Write(strm);
+  }
+
+  // Writes out header and symbols to output stream. If opts.header is false,
+  // skips writing header. If opts.[io]symbols is false, skips writing those
+  // symbols. `type` is the FST type being written. This method is used in the
+  // cross-type serialization methods Fst::WriteFst.
+  static void WriteFstHeader(const Fst<Arc> &fst, std::ostream &strm,
+                             const FstWriteOptions &opts, int version,
+                             const string &type, uint64 properties,
+                             FstHeader *hdr) {
+    if (opts.write_header) {
+      hdr->SetFstType(type);
+      hdr->SetArcType(Arc::Type());
+      hdr->SetVersion(version);
+      hdr->SetProperties(properties);
+      int32 file_flags = 0;
+      if (fst.InputSymbols() && opts.write_isymbols) {
+        file_flags |= FstHeader::HAS_ISYMBOLS;
+      }
+      if (fst.OutputSymbols() && opts.write_osymbols) {
+        file_flags |= FstHeader::HAS_OSYMBOLS;
+      }
+      if (opts.align) file_flags |= FstHeader::IS_ALIGNED;
+      hdr->SetFlags(file_flags);
+      hdr->Write(strm, opts.source);
+    }
+    if (fst.InputSymbols() && opts.write_isymbols) {
+      fst.InputSymbols()->Write(strm);
+    }
+    if (fst.OutputSymbols() && opts.write_osymbols) {
+      fst.OutputSymbols()->Write(strm);
+    }
+  }
+
+  // In serialization routines where the header cannot be written until after
+  // the machine has been serialized, this routine can be called to seek to the
+  // beginning of the file an rewrite the header with updated fields. It
+  // repositions the file pointer back at the end of the file. Returns true on
+  // success, false on failure.
+  static bool UpdateFstHeader(const Fst<Arc> &fst, std::ostream &strm,
+                              const FstWriteOptions &opts, int version,
+                              const string &type, uint64 properties,
+                              FstHeader *hdr, size_t header_offset) {
+    strm.seekp(header_offset);
+    if (!strm) {
+      LOG(ERROR) << "Fst::UpdateFstHeader: Write failed: " << opts.source;
+      return false;
+    }
+    WriteFstHeader(fst, strm, opts, version, type, properties, hdr);
+    if (!strm) {
+      LOG(ERROR) << "Fst::UpdateFstHeader: Write failed: " << opts.source;
+      return false;
+    }
+    strm.seekp(0, std::ios_base::end);
+    if (!strm) {
+      LOG(ERROR) << "Fst::UpdateFstHeader: Write failed: " << opts.source;
+      return false;
+    }
+    return true;
+  }
+
+ protected:
+  mutable uint64 properties_;  // Property bits.
+
+ private:
+  string type_;  // Unique name of FST class.
+  std::unique_ptr<SymbolTable> isymbols_;
+  std::unique_ptr<SymbolTable> osymbols_;
+};
+
+template <class Arc>
+inline FstImpl<Arc>::FstImpl(FstImpl<Arc> &&) noexcept = default;
+
+template <class Arc>
+inline FstImpl<Arc> &FstImpl<Arc>::operator=(
+    FstImpl<Arc> &&) noexcept = default;
+
+template <class Arc>
+bool FstImpl<Arc>::ReadHeader(std::istream &strm, const FstReadOptions &opts,
+                              int min_version, FstHeader *hdr) {
+  if (opts.header) {
+    *hdr = *opts.header;
+  } else if (!hdr->Read(strm, opts.source)) {
+    return false;
+  }
+  if (FLAGS_v >= 2) {
+    LOG(INFO) << "FstImpl::ReadHeader: source: " << opts.source
+              << ", fst_type: " << hdr->FstType()
+              << ", arc_type: " << Arc::Type()
+              << ", version: " << hdr->Version()
+              << ", flags: " << hdr->GetFlags();
+  }
+  if (hdr->FstType() != type_) {
+    LOG(ERROR) << "FstImpl::ReadHeader: FST not of type " << type_
+               << ": " << opts.source;
+    return false;
+  }
+  if (hdr->ArcType() != Arc::Type()) {
+    LOG(ERROR) << "FstImpl::ReadHeader: Arc not of type " << Arc::Type()
+               << ": " << opts.source;
+    return false;
+  }
+  if (hdr->Version() < min_version) {
+    LOG(ERROR) << "FstImpl::ReadHeader: Obsolete " << type_
+               << " FST version: " << opts.source;
+    return false;
+  }
+  properties_ = hdr->Properties();
+  if (hdr->GetFlags() & FstHeader::HAS_ISYMBOLS) {
+    isymbols_.reset(SymbolTable::Read(strm, opts.source));
+  }
+  // Deletes input symbol table.
+  if (!opts.read_isymbols) SetInputSymbols(nullptr);
+  if (hdr->GetFlags() & FstHeader::HAS_OSYMBOLS) {
+    osymbols_.reset(SymbolTable::Read(strm, opts.source));
+  }
+  // Deletes output symbol table.
+  if (!opts.read_osymbols) SetOutputSymbols(nullptr);
+  if (opts.isymbols) {
+    isymbols_.reset(opts.isymbols->Copy());
+  }
+  if (opts.osymbols) {
+    osymbols_.reset(opts.osymbols->Copy());
+  }
+  return true;
+}
+
+}  // namespace internal
+
+template <class Arc>
+uint64 TestProperties(const Fst<Arc> &fst, uint64 mask, uint64 *known);
+
+// This is a helper class template useful for attaching an FST interface to
+// its implementation, handling reference counting.
+template <class Impl, class FST = Fst<typename Impl::Arc>>
+class ImplToFst : public FST {
+ public:
+  using Arc = typename Impl::Arc;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  StateId Start() const override { return impl_->Start(); }
+
+  Weight Final(StateId s) const override { return impl_->Final(s); }
+
+  size_t NumArcs(StateId s) const override { return impl_->NumArcs(s); }
+
+  size_t NumInputEpsilons(StateId s) const override {
+    return impl_->NumInputEpsilons(s);
+  }
+
+  size_t NumOutputEpsilons(StateId s) const override {
+    return impl_->NumOutputEpsilons(s);
+  }
+
+  uint64 Properties(uint64 mask, bool test) const override {
+    if (test) {
+      uint64 knownprops, testprops = TestProperties(*this, mask, &knownprops);
+      impl_->SetProperties(testprops, knownprops);
+      return testprops & mask;
+    } else {
+      return impl_->Properties(mask);
+    }
+  }
+
+  const string &Type() const override { return impl_->Type(); }
+
+  const SymbolTable *InputSymbols() const override {
+    return impl_->InputSymbols();
+  }
+
+  const SymbolTable *OutputSymbols() const override {
+    return impl_->OutputSymbols();
+  }
+
+ protected:
+  explicit ImplToFst(std::shared_ptr<Impl> impl) : impl_(std::move(impl)) {}
+
+  // This constructor presumes there is a copy constructor for the
+  // implementation.
+  ImplToFst(const ImplToFst<Impl, FST> &fst, bool safe) {
+    if (safe) {
+      impl_ = std::make_shared<Impl>(*(fst.impl_));
+    } else {
+      impl_ = fst.impl_;
+    }
+  }
+
+  ImplToFst() = delete;
+
+  ImplToFst(const ImplToFst<Impl, FST> &fst) : impl_(fst.impl_) {}
+
+  ImplToFst(ImplToFst<Impl, FST> &&fst) noexcept
+      : impl_(std::move(fst.impl_)) {
+    fst.impl_ = std::make_shared<Impl>();
+  }
+
+  ImplToFst<Impl, FST> &operator=(const ImplToFst<Impl, FST> &fst) {
+    impl_ = fst.impl_;
+    return *this;
+  }
+
+  ImplToFst<Impl, FST> &operator=(ImplToFst<Impl, FST> &&fst) noexcept {
+    if (this != &fst) {
+      impl_ = std::move(fst.impl_);
+      fst.impl_ = std::make_shared<Impl>();
+    }
+    return *this;
+  }
+
+  // Returns raw pointers to the shared object.
+  const Impl *GetImpl() const { return impl_.get(); }
+
+  Impl *GetMutableImpl() const { return impl_.get(); }
+
+  // Returns a ref-counted smart poiner to the implementation.
+  std::shared_ptr<Impl> GetSharedImpl() const { return impl_; }
+
+  bool Unique() const { return impl_.unique(); }
+
+  void SetImpl(std::shared_ptr<Impl> impl) { impl_ = std::move(impl); }
+
+ private:
+  template <class IFST, class OFST>
+  friend void Cast(const IFST &ifst, OFST *ofst);
+
+  std::shared_ptr<Impl> impl_;
+};
+
+// Converts FSTs by casting their implementations, where this makes sense
+// (which excludes implementations with weight-dependent virtual methods).
+// Must be a friend of the FST classes involved (currently the concrete FSTs:
+// ConstFst, CompactFst, and VectorFst). This can only be safely used for arc
+// types that have identical storage characteristics. As with an FST
+// copy constructor and Copy() method, this is a constant time operation
+// (but subject to copy-on-write if it is a MutableFst and modified).
+template <class IFST, class OFST>
+void Cast(const IFST &ifst, OFST *ofst) {
+  using OImpl = typename OFST::Impl;
+  ofst->impl_ = std::shared_ptr<OImpl>(ifst.impl_,
+      reinterpret_cast<OImpl *>(ifst.impl_.get()));
+}
+
+// FST serialization.
+
+template <class Arc>
+string FstToString(const Fst<Arc> &fst,
+                   const FstWriteOptions &options =
+                       FstWriteOptions("FstToString")) {
+  std::ostringstream ostrm;
+  fst.Write(ostrm, options);
+  return ostrm.str();
+}
+
+template <class Arc>
+void FstToString(const Fst<Arc> &fst, string *result) {
+  *result = FstToString(fst);
+}
+
+template <class Arc>
+void FstToString(const Fst<Arc> &fst, string *result,
+                 const FstWriteOptions &options) {
+  *result = FstToString(fst, options);
+}
+
+template <class Arc>
+Fst<Arc> *StringToFst(const string &s) {
+  std::istringstream istrm(s);
+  return Fst<Arc>::Read(istrm, FstReadOptions("StringToFst"));
+}
+
+}  // namespace fst
+
+#endif  // FST_FST_H_
diff --git a/projects/llm_framework/include/fst/fstlib.h b/projects/llm_framework/include/fst/fstlib.h
new file mode 100644
index 00000000..e8b1c3a1
--- /dev/null
+++ b/projects/llm_framework/include/fst/fstlib.h
@@ -0,0 +1,130 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// This is a library for constructing, combining, optimizing, and searching
+// "weighted finite-state transducers" (FSTs). Weighted finite-state transducers
+// are automata where each transition has an input label, an output label, and a
+// weight. The more familiar finite-state acceptor is represented as a
+// transducer with each transition's input and output the same. Finite-state
+// acceptors are used to represent sets of strings (specifically, "regular" or
+// "rational sets"); finite-state transducers are used to represent binary
+// relations between pairs of strings (specifically, "rational transductions").
+// The weights can be used to represent the cost of taking a particular
+// transition.
+//
+// In this library, transducers are templated on the Arc (transition)
+// definition, which allows changing the label, weight, and state ID sets.
+// Labels and state IDs are restricted to signed integral types but the weight
+// can be an arbitrary type whose members satisfy certain algebraic ("semiring")
+// properties.
+//
+// This convenience file includes all other FST header files.
+
+#ifndef FST_FSTLIB_H_
+#define FST_FSTLIB_H_
+
+
+// Abstract FST classes.
+#include <fst/expanded-fst.h>
+#include <fst/fst.h>
+#include <fst/mutable-fst.h>
+
+// Concrete FST classes.
+#include <fst/compact-fst.h>
+#include <fst/const-fst.h>
+#include <fst/edit-fst.h>
+#include <fst/vector-fst.h>
+
+// FST algorithms and delayed FST classes.
+#include <fst/arc-map.h>
+#include <fst/arcsort.h>
+#include <fst/closure.h>
+#include <fst/compose.h>
+#include <fst/concat.h>
+#include <fst/connect.h>
+#include <fst/determinize.h>
+#include <fst/difference.h>
+#include <fst/disambiguate.h>
+#include <fst/encode.h>
+#include <fst/epsnormalize.h>
+#include <fst/equal.h>
+#include <fst/equivalent.h>
+#include <fst/factor-weight.h>
+#include <fst/intersect.h>
+#include <fst/invert.h>
+#include <fst/isomorphic.h>
+#include <fst/map.h>
+#include <fst/minimize.h>
+#include <fst/project.h>
+#include <fst/prune.h>
+#include <fst/push.h>
+#include <fst/randequivalent.h>
+#include <fst/randgen.h>
+#include <fst/rational.h>
+#include <fst/relabel.h>
+#include <fst/replace-util.h>
+#include <fst/replace.h>
+#include <fst/reverse.h>
+#include <fst/reweight.h>
+#include <fst/rmepsilon.h>
+#include <fst/rmfinalepsilon.h>
+#include <fst/shortest-distance.h>
+#include <fst/shortest-path.h>
+#include <fst/state-map.h>
+#include <fst/statesort.h>
+#include <fst/synchronize.h>
+#include <fst/topsort.h>
+#include <fst/union.h>
+#include <fst/verify.h>
+#include <fst/visit.h>
+
+// Weights.
+#include <fst/expectation-weight.h>
+#include <fst/float-weight.h>
+#include <fst/lexicographic-weight.h>
+#include <fst/pair-weight.h>
+#include <fst/power-weight.h>
+#include <fst/product-weight.h>
+#include <fst/signed-log-weight.h>
+#include <fst/sparse-power-weight.h>
+#include <fst/sparse-tuple-weight.h>
+#include <fst/string-weight.h>
+#include <fst/tuple-weight.h>
+#include <fst/weight.h>
+
+// Auxiliary classes for composition.
+#include <fst/compose-filter.h>
+#include <fst/lookahead-filter.h>
+#include <fst/lookahead-matcher.h>
+#include <fst/matcher-fst.h>
+#include <fst/matcher.h>
+#include <fst/state-table.h>
+
+// Data structures.
+#include <fst/heap.h>
+#include <fst/interval-set.h>
+#include <fst/queue.h>
+#include <fst/union-find.h>
+
+// Miscellaneous.
+#include <fst/accumulator.h>
+#include <fst/add-on.h>
+#include <fst/arc.h>
+#include <fst/arcfilter.h>
+#include <fst/cache.h>
+#include <fst/complement.h>
+#include <fst/dfs-visit.h>
+#include <fst/generic-register.h>
+#include <fst/label-reachable.h>
+#include <fst/partition.h>
+#include <fst/properties.h>
+#include <fst/register.h>
+#include <fst/state-reachable.h>
+#include <fst/string.h>
+#include <fst/symbol-table-ops.h>
+#include <fst/symbol-table.h>
+#include <fst/test-properties.h>
+#include <fst/util.h>
+
+
+#endif  // FST_FSTLIB_H_
diff --git a/projects/llm_framework/include/fst/generic-register.h b/projects/llm_framework/include/fst/generic-register.h
new file mode 100644
index 00000000..ea6b8fe1
--- /dev/null
+++ b/projects/llm_framework/include/fst/generic-register.h
@@ -0,0 +1,126 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_GENERIC_REGISTER_H_
+#define FST_GENERIC_REGISTER_H_
+
+#include <fst/compat.h>
+#ifndef FST_NO_DYNAMIC_LINKING
+#include <dlfcn.h>
+#endif
+#include <map>
+#include <string>
+
+#include <fst/types.h>
+#include <fst/log.h>
+
+// Generic class representing a globally-stored correspondence between
+// objects of KeyType and EntryType.
+//
+// KeyType must:
+//
+// * be such as can be stored as a key in a std::map<>.
+// * be concatenable with a const char* with the + operator
+//   (or you must subclass and redefine LoadEntryFromSharedObject)
+//
+// EntryType must be default constructible.
+//
+// The third template parameter should be the type of a subclass of this class
+// (think CRTP). This is to allow GetRegister() to instantiate and return an
+// object of the appropriate type.
+
+namespace fst {
+
+template <class KeyType, class EntryType, class RegisterType>
+class GenericRegister {
+ public:
+  using Key = KeyType;
+  using Entry = EntryType;
+
+  static RegisterType *GetRegister() {
+    static auto reg = new RegisterType;
+    return reg;
+  }
+
+  void SetEntry(const KeyType &key, const EntryType &entry) {
+    MutexLock l(&register_lock_);
+    register_table_.insert(std::make_pair(key, entry));
+  }
+
+  EntryType GetEntry(const KeyType &key) const {
+    const auto *entry = LookupEntry(key);
+    if (entry) {
+      return *entry;
+    } else {
+      return LoadEntryFromSharedObject(key);
+    }
+  }
+
+  virtual ~GenericRegister() {}
+
+ protected:
+  // Override this if you want to be able to load missing definitions from
+  // shared object files.
+  virtual EntryType LoadEntryFromSharedObject(const KeyType &key) const {
+#ifdef FST_NO_DYNAMIC_LINKING
+    return EntryType();
+#else
+    const auto so_filename = ConvertKeyToSoFilename(key);
+    void *handle = dlopen(so_filename.c_str(), RTLD_LAZY);
+    if (handle == nullptr) {
+      LOG(ERROR) << "GenericRegister::GetEntry: " << dlerror();
+      return EntryType();
+    }
+#ifdef RUN_MODULE_INITIALIZERS
+    RUN_MODULE_INITIALIZERS();
+#endif
+    // We assume that the DSO constructs a static object in its global scope
+    // that does the registration. Thus we need only load it, not call any
+    // methods.
+    const auto *entry = this->LookupEntry(key);
+    if (entry == nullptr) {
+      LOG(ERROR) << "GenericRegister::GetEntry: "
+                 << "lookup failed in shared object: " << so_filename;
+      return EntryType();
+    }
+    return *entry;
+#endif  // FST_NO_DYNAMIC_LINKING
+  }
+
+  // Override this to define how to turn a key into an SO filename.
+  virtual string ConvertKeyToSoFilename(const KeyType &key) const = 0;
+
+  virtual const EntryType *LookupEntry(const KeyType &key) const {
+    MutexLock l(&register_lock_);
+    const auto it = register_table_.find(key);
+    if (it != register_table_.end()) {
+      return &it->second;
+    } else {
+      return nullptr;
+    }
+  }
+
+ private:
+  mutable Mutex register_lock_;
+  std::map<KeyType, EntryType> register_table_;
+};
+
+// Generic register-er class capable of creating new register entries in the
+// given RegisterType template parameter. This type must define types Key and
+// Entry, and have appropriate static GetRegister() and instance SetEntry()
+// functions. An easy way to accomplish this is to have RegisterType be the
+// type of a subclass of GenericRegister.
+template <class RegisterType>
+class GenericRegisterer {
+ public:
+  using Key = typename RegisterType::Key;
+  using Entry = typename RegisterType::Entry;
+
+  GenericRegisterer(Key key, Entry entry) {
+    RegisterType::GetRegister()->SetEntry(key, entry);
+  }
+};
+
+}  // namespace fst
+
+#endif  // FST_GENERIC_REGISTER_H_
diff --git a/projects/llm_framework/include/fst/heap.h b/projects/llm_framework/include/fst/heap.h
new file mode 100644
index 00000000..041a4bb9
--- /dev/null
+++ b/projects/llm_framework/include/fst/heap.h
@@ -0,0 +1,168 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Implementation of a heap as in STL, but allows tracking positions in heap
+// using a key. The key can be used to do an in-place update of values in the
+// heap.
+
+#ifndef FST_HEAP_H_
+#define FST_HEAP_H_
+
+#include <utility>
+#include <vector>
+
+#include <fst/compat.h>
+namespace fst {
+
+// A templated heap implementation that supports in-place update of values.
+//
+// The templated heap implementation is a little different from the STL
+// priority_queue and the *_heap operations in STL. This heap supports
+// indexing of values in the heap via an associated key.
+//
+// Each value is internally associated with a key which is returned to the
+// calling functions on heap insert. This key can be used to later update
+// the specific value in the heap.
+//
+// T: the element type of the hash. It can be POD, Data or a pointer to Data.
+// Compare: comparison functor for determining min-heapness.
+template <class T, class Compare>
+class Heap {
+ public:
+  using Value = T;
+
+  static constexpr int kNoKey = -1;
+
+  // Initializes with a specific comparator.
+  explicit Heap(Compare comp = Compare()) : comp_(comp), size_(0) {}
+
+  // Inserts a value into the heap.
+  int Insert(const Value &value) {
+    if (size_ < values_.size()) {
+      values_[size_] = value;
+      pos_[key_[size_]] = size_;
+    } else {
+      values_.push_back(value);
+      pos_.push_back(size_);
+      key_.push_back(size_);
+    }
+    ++size_;
+    return Insert(value, size_ - 1);
+  }
+
+  // Updates a value at position given by the key. The pos_ array is first
+  // indexed by the key. The position gives the position in the heap array.
+  // Once we have the position we can then use the standard heap operations
+  // to calculate the parent and child positions.
+  void Update(int key, const Value &value) {
+    const auto i = pos_[key];
+    const bool is_better = comp_(value, values_[Parent(i)]);
+    values_[i] = value;
+    if (is_better) {
+      Insert(value, i);
+    } else {
+      Heapify(i);
+    }
+  }
+
+  // Returns the least value.
+  Value Pop() {
+    Value top = values_.front();
+    Swap(0, size_-1);
+    size_--;
+    Heapify(0);
+    return top;
+  }
+
+  // Returns the least value w.r.t.  the comparison function from the
+  // heap.
+  const Value &Top() const { return values_.front(); }
+
+  // Returns the element for the given key.
+  const Value &Get(int key) const { return values_[pos_[key]]; }
+
+  // Checks if the heap is empty.
+  bool Empty() const { return size_ == 0; }
+
+  void Clear() { size_ = 0; }
+
+  int Size() const { return size_; }
+
+  void Reserve(int size) {
+    values_.reserve(size);
+    pos_.reserve(size);
+    key_.reserve(size);
+  }
+
+  const Compare &GetCompare() const { return comp_; }
+
+ private:
+  // The following private routines are used in a supportive role
+  // for managing the heap and keeping the heap properties.
+
+  // Computes left child of parent.
+  static int Left(int i) {
+    return 2 * (i + 1) - 1;  // 0 -> 1, 1 -> 3
+  }
+
+  // Computes right child of parent.
+  static int Right(int i) {
+    return 2 * (i + 1);  // 0 -> 2, 1 -> 4
+  }
+
+  // Given a child computes parent.
+  static int Parent(int i) {
+    return (i - 1) / 2;  // 0 -> 0, 1 -> 0, 2 -> 0,  3 -> 1,  4 -> 1, ...
+  }
+
+  // Swaps a child and parent. Use to move element up/down tree. Note the use of
+  // a little trick here. When we swap we need to swap:
+  //
+  // - the value
+  // - the associated keys
+  // - the position of the value in the heap
+  void Swap(int j, int k) {
+    const auto tkey = key_[j];
+    pos_[key_[j] = key_[k]] = j;
+    pos_[key_[k] = tkey] = k;
+    using std::swap;
+    swap(values_[j], values_[k]);
+  }
+
+  // Heapifies the subtree rooted at index i.
+  void Heapify(int i) {
+    const auto l = Left(i);
+    const auto r = Right(i);
+    auto largest = (l < size_ && comp_(values_[l], values_[i])) ? l : i;
+    if (r < size_ && comp_(values_[r], values_[largest])) largest = r;
+    if (largest != i) {
+      Swap(i, largest);
+      Heapify(largest);
+    }
+  }
+
+  // Inserts (updates) element at subtree rooted at index i.
+  int Insert(const Value &value, int i) {
+    int p;
+    while (i > 0 && !comp_(values_[p = Parent(i)], value)) {
+      Swap(i, p);
+      i = p;
+    }
+    return key_[i];
+  }
+
+ private:
+  const Compare comp_;
+
+  std::vector<int> pos_;
+  std::vector<int> key_;
+  std::vector<Value> values_;
+  int size_;
+};
+
+template <class T, class Compare>
+constexpr int Heap<T, Compare>::kNoKey;
+
+}  // namespace fst
+
+#endif  // FST_HEAP_H_
diff --git a/projects/llm_framework/include/fst/icu.h b/projects/llm_framework/include/fst/icu.h
new file mode 100644
index 00000000..5459da24
--- /dev/null
+++ b/projects/llm_framework/include/fst/icu.h
@@ -0,0 +1,129 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// This library implements an unrestricted Thompson/Pike UTF-8 parser and
+// serializer. UTF-8 is a restricted subset of this byte stream encoding. For
+// a description of the encoding details, see:
+//
+//     http://en.wikipedia.org/wiki/UTF-8
+
+#ifndef FST_ICU_H_
+#define FST_ICU_H_
+
+#include <sstream>
+#include <vector>
+
+#include <fst/log.h>
+
+namespace fst {
+
+// Trivial function to copy bytestrings into vectors of labels, truncating
+// if necessary. It is possible to use this sensibly with as little as 8 bits
+// of Label precision. This returns `true` deterministically for compatibility.
+template <class Label>
+bool ByteStringToLabels(const string &str, std::vector<Label> *labels) {
+  for (const unsigned char ch : str) labels->push_back(ch);
+  return true;
+}
+
+// This function writes UTF-8 strings into a vector of Labels, truncating if
+// necessary. It is possible to use this sensibly with as little as 16 bits of
+// Label precision (i.e., when all characters are within the Basic Multilingual
+// Plane). With 21 bits, one can label all UTF-8 labelpoints, including those
+// from the various Astral Planes. Naturally, it is safe to use this with larger
+// Labels (e.g., 64 bits).
+template <class Label>
+bool UTF8StringToLabels(const string &str, std::vector<Label> *labels) {
+  for (auto it = str.begin(); it != str.end();) {
+    int c = *it & 0xff;
+    ++it;
+    if ((c & 0x80) == 0) {
+      labels->push_back(c);
+    } else {
+      if ((c & 0xc0) == 0x80) {
+        LOG(ERROR) << "UTF8StringToLabels: Continuation byte as lead byte";
+        return false;
+      }
+      int count =
+          (c >= 0xc0) + (c >= 0xe0) + (c >= 0xf0) + (c >= 0xf8) + (c >= 0xfc);
+      int32 label = c & ((1 << (6 - count)) - 1);
+      while (count != 0) {
+        if (it == str.end()) {
+          LOG(ERROR) << "UTF8StringToLabels: Truncated UTF-8 byte sequence";
+          return false;
+        }
+        char cb = *it;
+        ++it;
+        if ((cb & 0xc0) != 0x80) {
+          LOG(ERROR) << "UTF8StringToLabels: Missing/invalid continuation byte";
+          return false;
+        }
+        label = (label << 6) | (cb & 0x3f);
+        --count;
+      }
+      if (label < 0) {
+        // Should be unreachable.
+        LOG(ERROR) << "UTF8StringToLabels: Invalid character found: " << c;
+        return false;
+      }
+      labels->push_back(label);
+    }
+  }
+  return true;
+}
+
+template <class Label>
+bool LabelsToByteString(const std::vector<Label> &labels, string *str) {
+  std::ostringstream ostrm;
+  for (const char label : labels) {
+    if (label != 0) ostrm << label;
+  }
+  *str = ostrm.str();
+  return !!ostrm;
+}
+
+template <class Label>
+bool LabelsToUTF8String(const std::vector<Label> &labels, string *str) {
+  std::ostringstream ostrm;
+  for (const int32 label : labels) {
+    if (label < 0) {
+      LOG(ERROR) << "LabelsToUTF8String: Invalid character found: " << label;
+      return false;
+    } else if (label == 0) {
+      continue;
+    } else if (label < 0x80) {
+      ostrm << static_cast<char>(label);
+    } else if (label < 0x800) {
+      ostrm << static_cast<char>((label >> 6) | 0xc0);
+      ostrm << static_cast<char>((label & 0x3f) | 0x80);
+    } else if (label < 0x10000) {
+      ostrm << static_cast<char>((label >> 12) | 0xe0);
+      ostrm << static_cast<char>(((label >> 6) & 0x3f) | 0x80);
+      ostrm << static_cast<char>((label & 0x3f) | 0x80);
+    } else if (label < 0x200000) {
+      ostrm << static_cast<char>((label >> 18) | 0xf0);
+      ostrm << static_cast<char>(((label >> 12) & 0x3f) | 0x80);
+      ostrm << static_cast<char>(((label >> 6) & 0x3f) | 0x80);
+      ostrm << static_cast<char>((label & 0x3f) | 0x80);
+    } else if (label < 0x4000000) {
+      ostrm << static_cast<char>((label >> 24) | 0xf8);
+      ostrm << static_cast<char>(((label >> 18) & 0x3f) | 0x80);
+      ostrm << static_cast<char>(((label >> 12) & 0x3f) | 0x80);
+      ostrm << static_cast<char>(((label >> 6) & 0x3f) | 0x80);
+      ostrm << static_cast<char>((label & 0x3f) | 0x80);
+    } else {
+      ostrm << static_cast<char>((label >> 30) | 0xfc);
+      ostrm << static_cast<char>(((label >> 24) & 0x3f) | 0x80);
+      ostrm << static_cast<char>(((label >> 18) & 0x3f) | 0x80);
+      ostrm << static_cast<char>(((label >> 12) & 0x3f) | 0x80);
+      ostrm << static_cast<char>(((label >> 6) & 0x3f) | 0x80);
+      ostrm << static_cast<char>((label & 0x3f) | 0x80);
+    }
+  }
+  *str = ostrm.str();
+  return !!ostrm;
+}
+
+}  // namespace fst
+
+#endif  // FST_ICU_H_
diff --git a/projects/llm_framework/include/fst/intersect.h b/projects/llm_framework/include/fst/intersect.h
new file mode 100644
index 00000000..fa66206d
--- /dev/null
+++ b/projects/llm_framework/include/fst/intersect.h
@@ -0,0 +1,181 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to compute the intersection of two FSAs.
+
+#ifndef FST_INTERSECT_H_
+#define FST_INTERSECT_H_
+
+#include <algorithm>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/cache.h>
+#include <fst/compose.h>
+
+
+namespace fst {
+
+using IntersectOptions = ComposeOptions;
+
+template <class Arc, class M = Matcher<Fst<Arc>>,
+          class Filter = SequenceComposeFilter<M>,
+          class StateTable =
+              GenericComposeStateTable<Arc, typename Filter::FilterState>>
+struct IntersectFstOptions
+    : public ComposeFstOptions<Arc, M, Filter, StateTable> {
+  IntersectFstOptions() {}
+
+  explicit IntersectFstOptions(const CacheOptions &opts, M *matcher1 = nullptr,
+                               M *matcher2 = nullptr, Filter *filter = nullptr,
+                               StateTable *state_table = nullptr)
+      : ComposeFstOptions<Arc, M, Filter, StateTable>(opts, matcher1, matcher2,
+                                                      filter, state_table) {}
+};
+
+// Computes the intersection (Hadamard product) of two FSAs. This version is a
+// delayed FST. Only strings that are in both automata are retained in the
+// result.
+//
+// The two arguments must be acceptors. One of the arguments must be
+// label-sorted.
+//
+// Complexity: same as ComposeFst.
+//
+// Caveats: same as ComposeFst.
+template <class A>
+class IntersectFst : public ComposeFst<A> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using ComposeFst<A>::CreateBase;
+  using ComposeFst<A>::CreateBase1;
+  using ComposeFst<A>::Properties;
+
+  IntersectFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+               const CacheOptions &opts = CacheOptions())
+      : ComposeFst<Arc>(CreateBase(fst1, fst2, opts)) {
+    const bool acceptors =
+        fst1.Properties(kAcceptor, true) && fst2.Properties(kAcceptor, true);
+    if (!acceptors) {
+      FSTERROR() << "IntersectFst: Input FSTs are not acceptors";
+      GetMutableImpl()->SetProperties(kError);
+    }
+  }
+
+  template <class M, class Filter, class StateTable>
+  IntersectFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+               const IntersectFstOptions<Arc, M, Filter, StateTable> &opts)
+      : ComposeFst<Arc>(CreateBase1(fst1, fst2, opts)) {
+    const bool acceptors =
+        fst1.Properties(kAcceptor, true) && fst2.Properties(kAcceptor, true);
+    if (!acceptors) {
+      FSTERROR() << "IntersectFst: input FSTs are not acceptors";
+      GetMutableImpl()->SetProperties(kError);
+    }
+  }
+
+  // See Fst<>::Copy() for doc.
+  IntersectFst(const IntersectFst<Arc> &fst, bool safe = false)
+      : ComposeFst<Arc>(fst, safe) {}
+
+  // Get a copy of this IntersectFst. See Fst<>::Copy() for further doc.
+  IntersectFst<Arc> *Copy(bool safe = false) const override {
+    return new IntersectFst<Arc>(*this, safe);
+  }
+
+ private:
+  using ImplToFst<internal::ComposeFstImplBase<A>>::GetImpl;
+  using ImplToFst<internal::ComposeFstImplBase<A>>::GetMutableImpl;
+};
+
+// Specialization for IntersectFst.
+template <class Arc>
+class StateIterator<IntersectFst<Arc>> : public StateIterator<ComposeFst<Arc>> {
+ public:
+  explicit StateIterator(const IntersectFst<Arc> &fst)
+      : StateIterator<ComposeFst<Arc>>(fst) {}
+};
+
+// Specialization for IntersectFst.
+template <class Arc>
+class ArcIterator<IntersectFst<Arc>> : public ArcIterator<ComposeFst<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const IntersectFst<Arc> &fst, StateId s)
+      : ArcIterator<ComposeFst<Arc>>(fst, s) {}
+};
+
+// Useful alias when using StdArc.
+using StdIntersectFst = IntersectFst<StdArc>;
+
+// Computes the intersection (Hadamard product) of two FSAs. This version
+// writes the intersection to an output MurableFst. Only strings that are in
+// both automata are retained in the result.
+//
+// The two arguments must be acceptors. One of the arguments must be
+// label-sorted.
+//
+// Complexity: same as Compose.
+//
+// Caveats: same as Compose.
+template <class Arc>
+void Intersect(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
+               MutableFst<Arc> *ofst,
+               const IntersectOptions &opts = IntersectOptions()) {
+  using M = Matcher<Fst<Arc>>;
+  // In each case, we cache only the last state for fastest copy.
+  switch (opts.filter_type) {
+    case AUTO_FILTER: {
+      CacheOptions nopts;
+      nopts.gc_limit = 0;
+      *ofst = IntersectFst<Arc>(ifst1, ifst2, nopts);
+      break;
+    }
+    case SEQUENCE_FILTER: {
+      IntersectFstOptions<Arc> iopts;
+      iopts.gc_limit = 0;
+      *ofst = IntersectFst<Arc>(ifst1, ifst2, iopts);
+      break;
+    }
+    case ALT_SEQUENCE_FILTER: {
+      IntersectFstOptions<Arc, M, AltSequenceComposeFilter<M>> iopts;
+      iopts.gc_limit = 0;
+      *ofst = IntersectFst<Arc>(ifst1, ifst2, iopts);
+      break;
+    }
+    case MATCH_FILTER: {
+      IntersectFstOptions<Arc, M, MatchComposeFilter<M>> iopts;
+      iopts.gc_limit = 0;
+      *ofst = IntersectFst<Arc>(ifst1, ifst2, iopts);
+      break;
+    }
+    case NO_MATCH_FILTER: {
+      IntersectFstOptions<Arc, M, NoMatchComposeFilter<M>> iopts;
+      iopts.gc_limit = 0;
+      *ofst = IntersectFst<Arc>(ifst1, ifst2, iopts);
+      break;
+    }
+    case NULL_FILTER: {
+      IntersectFstOptions<Arc, M, NullComposeFilter<M>> iopts;
+      iopts.gc_limit = 0;
+      *ofst = IntersectFst<Arc>(ifst1, ifst2, iopts);
+      break;
+    }
+    case TRIVIAL_FILTER: {
+      IntersectFstOptions<Arc, M, TrivialComposeFilter<M>> iopts;
+      iopts.gc_limit = 0;
+      *ofst = IntersectFst<Arc>(ifst1, ifst2, iopts);
+      break;
+    }
+  }
+  if (opts.connect) Connect(ofst);
+}
+
+}  // namespace fst
+
+#endif  // FST_INTERSECT_H_
diff --git a/projects/llm_framework/include/fst/interval-set.h b/projects/llm_framework/include/fst/interval-set.h
new file mode 100644
index 00000000..0942ea00
--- /dev/null
+++ b/projects/llm_framework/include/fst/interval-set.h
@@ -0,0 +1,398 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to represent and operate on sets of intervals.
+
+#ifndef FST_INTERVAL_SET_H_
+#define FST_INTERVAL_SET_H_
+
+#include <algorithm>
+#include <iostream>
+#include <vector>
+
+
+#include <fst/util.h>
+
+
+namespace fst {
+
+// Half-open integral interval [a, b) of signed integers of type T.
+template <class T>
+struct IntInterval {
+  T begin;
+  T end;
+
+  IntInterval() : begin(-1), end(-1) {}
+
+  IntInterval(T begin, T end) : begin(begin), end(end) {}
+
+  bool operator<(const IntInterval<T> &i) const {
+    return begin < i.begin || (begin == i.begin && end > i.end);
+  }
+
+  bool operator==(const IntInterval<T> &i) const {
+    return begin == i.begin && end == i.end;
+  }
+
+  bool operator!=(const IntInterval<T> &i) const {
+    return begin != i.begin || end != i.end;
+  }
+
+  std::istream &Read(std::istream &strm) {
+    T n;
+    ReadType(strm, &n);
+    begin = n;
+    ReadType(strm, &n);
+    end = n;
+    return strm;
+  }
+
+  std::ostream &Write(std::ostream &strm) const {
+    T n = begin;
+    WriteType(strm, n);
+    n = end;
+    WriteType(strm, n);
+    return strm;
+  }
+};
+
+// Stores IntIntervals<T> in a vector. In addition, keeps the count of points in
+// all intervals.
+template <class T>
+class VectorIntervalStore {
+ public:
+  using Interval = IntInterval<T>;
+  using Iterator = typename std::vector<Interval>::const_iterator;
+
+  VectorIntervalStore() : count_(-1) {}
+
+  std::vector<Interval> *MutableIntervals() { return &intervals_; }
+
+  const Interval *Intervals() const { return intervals_.data(); }
+
+  T Size() const { return intervals_.size(); }
+
+  T Count() const { return count_; }
+
+  void SetCount(T count) { count_ = count; }
+
+  void Clear() {
+    intervals_.clear();
+    count_ = 0;
+  }
+
+  Iterator begin() const { return intervals_.begin(); }
+
+  Iterator end() const { return intervals_.end(); }
+
+  std::istream &Read(std::istream &strm) {
+    ReadType(strm, &intervals_);
+    return ReadType(strm, &count_);
+  }
+
+  std::ostream &Write(std::ostream &strm) const {
+    WriteType(strm, intervals_);
+    return WriteType(strm, count_);
+  }
+
+ private:
+  std::vector<Interval> intervals_;
+  T count_;
+};
+
+// Stores and operates on a set of half-open integral intervals [a, b)
+// of signed integers of type T.
+template <class T, class Store = VectorIntervalStore<T>>
+class IntervalSet {
+ public:
+  using Interval = IntInterval<T>;
+
+  template <class... A>
+  explicit IntervalSet(A... args) : intervals_(args...) {}
+
+  // Returns the interval set as a vector.
+  std::vector<Interval> *MutableIntervals() {
+    return intervals_.MutableIntervals();
+  }
+
+  // Returns a pointer to an array of Size() elements.
+  const Interval *Intervals() const { return intervals_.Intervals(); }
+
+  bool Empty() const { return Size() == 0; }
+
+  T Size() const { return intervals_.Size(); }
+
+  // Number of points in the intervals (undefined if not normalized).
+  T Count() const { return intervals_.Count(); }
+
+  void Clear() { intervals_.Clear(); }
+
+  // Adds an interval set to the set. The result may not be normalized.
+  void Union(const IntervalSet<T, Store> &iset) {
+    intervals_.MutableIntervals()->insert(intervals_.MutableIntervals()->end(),
+                                          iset.intervals_.begin(),
+                                          iset.intervals_.end());
+  }
+
+  // Requires intervals be normalized.
+  bool Member(T value) const {
+    const Interval interval(value, value);
+    auto lb = std::lower_bound(intervals_.begin(), intervals_.end(), interval);
+    if (lb == intervals_.begin()) return false;
+    return (--lb)->end > value;
+  }
+
+  // Requires intervals be normalized.
+  bool operator==(const IntervalSet<T, Store> &iset) const {
+    return Size() == iset.Size() &&
+           std::equal(intervals_.begin(), intervals_.end(),
+                      iset.intervals_.begin());
+  }
+
+  // Requires intervals be normalized.
+  bool operator!=(const IntervalSet<T, Store> &iset) const {
+    return Size() != iset.Size() ||
+           !std::equal(intervals_.begin(), intervals_.end(),
+                       iset.intervals_.begin());
+  }
+
+  bool Singleton() const {
+    return Size() == 1 &&
+           intervals_.begin()->begin + 1 == intervals_.begin()->end;
+  }
+
+  // Sorts, collapses overlapping and adjacent interals, and sets count.
+  void Normalize();
+
+  // Intersects an interval set with the set. Requires intervals be normalized.
+  // The result is normalized.
+  void Intersect(const IntervalSet<T, Store> &iset,
+                 IntervalSet<T, Store> *oset) const;
+
+  // Complements the set w.r.t [0, maxval). Requires intervals be normalized.
+  // The result is normalized.
+  void Complement(T maxval, IntervalSet<T, Store> *oset) const;
+
+  // Subtract an interval set from the set. Requires intervals be normalized.
+  // The result is normalized.
+  void Difference(const IntervalSet<T, Store> &iset,
+                  IntervalSet<T, Store> *oset) const;
+
+  // Determines if an interval set overlaps with the set. Requires intervals be
+  // normalized.
+  bool Overlaps(const IntervalSet<T, Store> &iset) const;
+
+  // Determines if an interval set overlaps with the set but neither is
+  // contained in the other. Requires intervals be normalized.
+  bool StrictlyOverlaps(const IntervalSet<T, Store> &iset) const;
+
+  // Determines if an interval set is contained within the set. Requires
+  // intervals be normalized.
+  bool Contains(const IntervalSet<T, Store> &iset) const;
+
+  std::istream &Read(std::istream &strm) { return intervals_.Read(strm); }
+
+  std::ostream &Write(std::ostream &strm) const {
+    return intervals_.Write(strm);
+  }
+
+  typename Store::Iterator begin() const { return intervals_.begin(); }
+
+  typename Store::Iterator end() const { return intervals_.end(); }
+
+ private:
+  Store intervals_;
+};
+
+// Sorts, collapses overlapping and adjacent intervals, and sets count.
+template <typename T, class Store>
+void IntervalSet<T, Store>::Normalize() {
+  auto &intervals = *intervals_.MutableIntervals();
+  std::sort(intervals.begin(), intervals.end());
+  T count = 0;
+  T size = 0;
+  for (T i = 0; i < intervals.size(); ++i) {
+    auto &inti = intervals[i];
+    if (inti.begin == inti.end) continue;
+    for (T j = i + 1; j < intervals.size(); ++j) {
+      auto &intj = intervals[j];
+      if (intj.begin > inti.end) break;
+      if (intj.end > inti.end) inti.end = intj.end;
+      ++i;
+    }
+    count += inti.end - inti.begin;
+    intervals[size++] = inti;
+  }
+  intervals.resize(size);
+  intervals_.SetCount(count);
+}
+
+// Intersects an interval set with the set. Requires intervals be normalized.
+// The result is normalized.
+template <typename T, class Store>
+void IntervalSet<T, Store>::Intersect(const IntervalSet<T, Store> &iset,
+                                      IntervalSet<T, Store> *oset) const {
+  auto *ointervals = oset->MutableIntervals();
+  auto it1 = intervals_.begin();
+  auto it2 = iset.intervals_.begin();
+  ointervals->clear();
+  T count = 0;
+  while (it1 != intervals_.end() && it2 != iset.intervals_.end()) {
+    if (it1->end <= it2->begin) {
+      ++it1;
+    } else if (it2->end <= it1->begin) {
+      ++it2;
+    } else {
+      ointervals->emplace_back(std::max(it1->begin, it2->begin),
+                               std::min(it1->end, it2->end));
+      count += ointervals->back().end - ointervals->back().begin;
+      if (it1->end < it2->end) {
+        ++it1;
+      } else {
+        ++it2;
+      }
+    }
+  }
+  oset->intervals_.SetCount(count);
+}
+
+// Complements the set w.r.t [0, maxval). Requires intervals be normalized.
+// The result is normalized.
+template <typename T, class Store>
+void IntervalSet<T, Store>::Complement(T maxval,
+                                       IntervalSet<T, Store> *oset) const {
+  auto *ointervals = oset->MutableIntervals();
+  ointervals->clear();
+  T count = 0;
+  Interval interval;
+  interval.begin = 0;
+  for (auto it = intervals_.begin(); it != intervals_.end(); ++it) {
+    interval.end = std::min(it->begin, maxval);
+    if ((interval.begin) < (interval.end)) {
+      ointervals->push_back(interval);
+      count += interval.end - interval.begin;
+    }
+    interval.begin = it->end;
+  }
+  interval.end = maxval;
+  if ((interval.begin) < (interval.end)) {
+    ointervals->push_back(interval);
+    count += interval.end - interval.begin;
+  }
+  oset->intervals_.SetCount(count);
+}
+
+// Subtract an interval set from the set. Requires intervals be normalized.
+// The result is normalized.
+template <typename T, class Store>
+void IntervalSet<T, Store>::Difference(const IntervalSet<T, Store> &iset,
+                                       IntervalSet<T, Store> *oset) const {
+  if (Empty()) {
+    oset->MutableIntervals()->clear();
+    oset->intervals_.SetCount(0);
+  } else {
+    IntervalSet<T, Store> cset;
+    iset.Complement(intervals_.Intervals()[intervals_.Size() - 1].end, &cset);
+    Intersect(cset, oset);
+  }
+}
+
+// Determines if an interval set overlaps with the set. Requires intervals be
+// normalized.
+template <typename T, class Store>
+bool IntervalSet<T, Store>::Overlaps(const IntervalSet<T, Store> &iset) const {
+  auto it1 = intervals_.begin();
+  auto it2 = iset.intervals_.begin();
+  while (it1 != intervals_.end() && it2 != iset.intervals_.end()) {
+    if (it1->end <= it2->begin) {
+      ++it1;
+    } else if (it2->end <= it1->begin) {
+      ++it2;
+    } else {
+      return true;
+    }
+  }
+  return false;
+}
+
+// Determines if an interval set overlaps with the set but neither is contained
+// in the other. Requires intervals be normalized.
+template <typename T, class Store>
+bool IntervalSet<T, Store>::StrictlyOverlaps(
+    const IntervalSet<T, Store> &iset) const {
+  auto it1 = intervals_.begin();
+  auto it2 = iset.intervals_.begin();
+  bool only1 = false;    // Point in intervals_ but not intervals.
+  bool only2 = false;    // Point in intervals but not intervals_.
+  bool overlap = false;  // Point in both intervals_ and intervals.
+  while (it1 != intervals_.end() && it2 != iset.intervals_.end()) {
+    if (it1->end <= it2->begin) {  // no overlap - it1 first
+      only1 = true;
+      ++it1;
+    } else if (it2->end <= it1->begin) {  // no overlap - it2 first
+      only2 = true;
+      ++it2;
+    } else if (it2->begin == it1->begin && it2->end == it1->end) {  // equals
+      overlap = true;
+      ++it1;
+      ++it2;
+    } else if (it2->begin <= it1->begin && it2->end >= it1->end) {  // 1 c 2
+      only2 = true;
+      overlap = true;
+      ++it1;
+    } else if (it1->begin <= it2->begin && it1->end >= it2->end) {  // 2 c 1
+      only1 = true;
+      overlap = true;
+      ++it2;
+    } else {  // Strict overlap.
+      only1 = true;
+      only2 = true;
+      overlap = true;
+    }
+    if (only1 == true && only2 == true && overlap == true) return true;
+  }
+  if (it1 != intervals_.end()) only1 = true;
+  if (it2 != iset.intervals_.end()) only2 = true;
+  return only1 == true && only2 == true && overlap == true;
+}
+
+// Determines if an interval set is contained within the set. Requires intervals
+// be normalized.
+template <typename T, class Store>
+bool IntervalSet<T, Store>::Contains(const IntervalSet<T, Store> &iset) const {
+  if (iset.Count() > Count()) return false;
+  auto it1 = intervals_.begin();
+  auto it2 = iset.intervals_.begin();
+  while (it1 != intervals_.end() && it2 != iset.intervals_.end()) {
+    if ((it1->end) <= (it2->begin)) {  // No overlap; it1 first.
+      ++it1;
+    } else if ((it2->begin) < (it1->begin) ||
+               (it2->end) > (it1->end)) {  // No C.
+      return false;
+    } else if (it2->end == it1->end) {
+      ++it1;
+      ++it2;
+    } else {
+      ++it2;
+    }
+  }
+  return it2 == iset.intervals_.end();
+}
+
+template <typename T, class Store>
+std::ostream &operator<<(std::ostream &strm, const IntervalSet<T, Store> &s) {
+  strm << "{";
+  for (T i = 0; i < s.Size(); ++i) {
+    if (i > 0) {
+      strm << ",";
+    }
+    const auto &interval = s.Intervals()[i];
+    strm << "[" << interval.begin << "," << interval.end << ")";
+  }
+  strm << "}";
+  return strm;
+}
+
+}  // namespace fst
+
+#endif  // FST_INTERVAL_SET_H_
diff --git a/projects/llm_framework/include/fst/invert.h b/projects/llm_framework/include/fst/invert.h
new file mode 100644
index 00000000..bd243c62
--- /dev/null
+++ b/projects/llm_framework/include/fst/invert.h
@@ -0,0 +1,139 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes to invert an FST.
+
+#ifndef FST_INVERT_H_
+#define FST_INVERT_H_
+
+#include <fst/arc-map.h>
+#include <fst/mutable-fst.h>
+
+
+namespace fst {
+
+// Mapper to implement inversion of an arc.
+template <class A>
+struct InvertMapper {
+  using FromArc = A;
+  using ToArc = A;
+
+  InvertMapper() {}
+
+  ToArc operator()(const FromArc &arc) const {
+    return ToArc(arc.olabel, arc.ilabel, arc.weight, arc.nextstate);
+  }
+
+  constexpr MapFinalAction FinalAction() const {
+     return MAP_NO_SUPERFINAL;
+  }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_CLEAR_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_CLEAR_SYMBOLS;
+  }
+
+  uint64 Properties(uint64 props) const {
+    return InvertProperties(props);
+  }
+};
+
+// Inverts the transduction corresponding to an FST by exchanging the
+// FST's input and output labels.
+//
+// Complexity:
+//
+//   Time: O(V + E)
+//   Space: O(1)
+//
+// where V is the number of states and E is the number of arcs.
+template <class Arc>
+inline void Invert(const Fst<Arc> &ifst, MutableFst<Arc> *ofst) {
+  std::unique_ptr<SymbolTable> input(
+      ifst.InputSymbols() ? ifst.InputSymbols()->Copy() : nullptr);
+  std::unique_ptr<SymbolTable> output(
+      ifst.OutputSymbols() ? ifst.OutputSymbols()->Copy() : nullptr);
+  ArcMap(ifst, ofst, InvertMapper<Arc>());
+  ofst->SetInputSymbols(output.get());
+  ofst->SetOutputSymbols(input.get());
+}
+
+// Destructive variant of the above.
+template <class Arc>
+inline void Invert(MutableFst<Arc> *fst) {
+  std::unique_ptr<SymbolTable> input(
+      fst->InputSymbols() ? fst->InputSymbols()->Copy() : nullptr);
+  std::unique_ptr<SymbolTable> output(
+      fst->OutputSymbols() ? fst->OutputSymbols()->Copy() : nullptr);
+  ArcMap(fst, InvertMapper<Arc>());
+  fst->SetInputSymbols(output.get());
+  fst->SetOutputSymbols(input.get());
+}
+
+// Inverts the transduction corresponding to an FST by exchanging the
+// FST's input and output labels. This version is a delayed FST.
+//
+// Complexity:
+//
+//   Time: O(v + e)
+//   Space: O(1)
+//
+// where v is the number of states visited and e is the number of arcs visited.
+// Constant time and to visit an input state or arc is assumed and exclusive of
+// caching.
+template <class A>
+class InvertFst : public ArcMapFst<A, A, InvertMapper<A>> {
+ public:
+  using Arc = A;
+
+  using Mapper = InvertMapper<Arc>;
+  using Impl = internal::ArcMapFstImpl<A, A, InvertMapper<A>>;
+
+  explicit InvertFst(const Fst<Arc> &fst)
+      : ArcMapFst<Arc, Arc, Mapper>(fst, Mapper()) {
+    GetMutableImpl()->SetOutputSymbols(fst.InputSymbols());
+    GetMutableImpl()->SetInputSymbols(fst.OutputSymbols());
+  }
+
+  // See Fst<>::Copy() for doc.
+  InvertFst(const InvertFst<Arc> &fst, bool safe = false)
+      : ArcMapFst<Arc, Arc, Mapper>(fst, safe) {}
+
+  // Get a copy of this InvertFst. See Fst<>::Copy() for further doc.
+  InvertFst<Arc> *Copy(bool safe = false) const override {
+    return new InvertFst(*this, safe);
+  }
+
+ private:
+  using ImplToFst<Impl>::GetMutableImpl;
+};
+
+// Specialization for InvertFst.
+template <class Arc>
+class StateIterator<InvertFst<Arc>>
+    : public StateIterator<ArcMapFst<Arc, Arc, InvertMapper<Arc>>> {
+ public:
+  explicit StateIterator(const InvertFst<Arc> &fst)
+      : StateIterator<ArcMapFst<Arc, Arc, InvertMapper<Arc>>>(fst) {}
+};
+
+// Specialization for InvertFst.
+template <class Arc>
+class ArcIterator<InvertFst<Arc>>
+    : public ArcIterator<ArcMapFst<Arc, Arc, InvertMapper<Arc>>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const InvertFst<Arc> &fst, StateId s)
+      : ArcIterator<ArcMapFst<Arc, Arc, InvertMapper<Arc>>>(fst, s) {}
+};
+
+// Useful alias when using StdArc.
+using StdInvertFst = InvertFst<StdArc>;
+
+}  // namespace fst
+
+#endif  // FST_INVERT_H_
diff --git a/projects/llm_framework/include/fst/isomorphic.h b/projects/llm_framework/include/fst/isomorphic.h
new file mode 100644
index 00000000..b100b0a6
--- /dev/null
+++ b/projects/llm_framework/include/fst/isomorphic.h
@@ -0,0 +1,183 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Function to test two FSTs are isomorphic, i.e., they are equal up to a state
+// and arc re-ordering. FSTs should be deterministic when viewed as
+// unweighted automata.
+
+#ifndef FST_ISOMORPHIC_H_
+#define FST_ISOMORPHIC_H_
+
+#include <algorithm>
+#include <list>
+#include <type_traits>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/fst.h>
+
+
+namespace fst {
+namespace internal {
+
+// Orders weights for equality checking.
+template <class Weight, typename std::enable_if<
+                            IsIdempotent<Weight>::value>::type * = nullptr>
+bool WeightCompare(const Weight &w1, const Weight &w2, float delta,
+                   bool *error) {
+  return NaturalLess<Weight>()(w1, w2);
+}
+
+template <class Weight, typename std::enable_if<
+                            !IsIdempotent<Weight>::value>::type * = nullptr>
+bool WeightCompare(const Weight &w1, const Weight &w2, float delta,
+                   bool *error) {
+  // No natural order; use hash.
+  const auto q1 = w1.Quantize(delta);
+  const auto q2 = w2.Quantize(delta);
+  auto n1 = q1.Hash();
+  auto n2 = q2.Hash();
+  // Hash not unique; very unlikely to happen.
+  if (n1 == n2 && q1 != q2) {
+    VLOG(1) << "Isomorphic: Weight hash collision";
+    *error = true;
+  }
+  return n1 < n2;
+}
+
+template <class Arc>
+class Isomorphism {
+  using StateId = typename Arc::StateId;
+
+ public:
+  Isomorphism(const Fst<Arc> &fst1, const Fst<Arc> &fst2, float delta)
+      : fst1_(fst1.Copy()),
+        fst2_(fst2.Copy()),
+        delta_(delta),
+        error_(false),
+        comp_(delta, &error_) {}
+
+  // Checks if input FSTs are isomorphic.
+  bool IsIsomorphic() {
+    if (fst1_->Start() == kNoStateId && fst2_->Start() == kNoStateId) {
+      return true;
+    }
+    if (fst1_->Start() == kNoStateId || fst2_->Start() == kNoStateId) {
+      return false;
+    }
+    PairState(fst1_->Start(), fst2_->Start());
+    while (!queue_.empty()) {
+      const auto &pr = queue_.front();
+      if (!IsIsomorphicState(pr.first, pr.second)) return false;
+      queue_.pop_front();
+    }
+    return true;
+  }
+
+  bool Error() const { return error_; }
+
+ private:
+  // Orders arcs for equality checking.
+  class ArcCompare {
+   public:
+    ArcCompare(float delta, bool *error) : delta_(delta), error_(error) {}
+
+    bool operator()(const Arc &arc1, const Arc &arc2) const {
+      if (arc1.ilabel < arc2.ilabel) return true;
+      if (arc1.ilabel > arc2.ilabel) return false;
+      if (arc1.olabel < arc2.olabel) return true;
+      if (arc1.olabel > arc2.olabel) return false;
+      return WeightCompare(arc1.weight, arc2.weight, delta_, error_);
+    }
+
+   private:
+    float delta_;
+    bool *error_;
+  };
+
+  // Maintains state correspondences and queue.
+  bool PairState(StateId s1, StateId s2) {
+    if (state_pairs_.size() <= s1) state_pairs_.resize(s1 + 1, kNoStateId);
+    if (state_pairs_[s1] == s2) {
+      return true;  // already seen this pair
+    } else if (state_pairs_[s1] != kNoStateId) {
+      return false;  // s1 already paired with another s2
+    }
+    state_pairs_[s1] = s2;
+    queue_.push_back(std::make_pair(s1, s2));
+    return true;
+  }
+
+  // Checks if state pair is isomorphic
+  bool IsIsomorphicState(StateId s1, StateId s2);
+
+  std::unique_ptr<Fst<Arc>> fst1_;
+  std::unique_ptr<Fst<Arc>> fst2_;
+  float delta_;                          // Weight equality delta.
+  std::vector<Arc> arcs1_;               // For sorting arcs on FST1.
+  std::vector<Arc> arcs2_;               // For sorting arcs on FST2.
+  std::vector<StateId> state_pairs_;     // Maintains state correspondences.
+  std::list<std::pair<StateId, StateId>> queue_;  // Queue of state pairs.
+  bool error_;                           // Error flag.
+  ArcCompare comp_;
+};
+
+template <class Arc>
+bool Isomorphism<Arc>::IsIsomorphicState(StateId s1, StateId s2) {
+  if (!ApproxEqual(fst1_->Final(s1), fst2_->Final(s2), delta_)) return false;
+  auto narcs1 = fst1_->NumArcs(s1);
+  auto narcs2 = fst2_->NumArcs(s2);
+  if (narcs1 != narcs2) return false;
+  ArcIterator<Fst<Arc>> aiter1(*fst1_, s1);
+  ArcIterator<Fst<Arc>> aiter2(*fst2_, s2);
+  arcs1_.clear();
+  arcs1_.reserve(narcs1);
+  arcs2_.clear();
+  arcs2_.reserve(narcs2);
+  for (; !aiter1.Done(); aiter1.Next(), aiter2.Next()) {
+    arcs1_.push_back(aiter1.Value());
+    arcs2_.push_back(aiter2.Value());
+  }
+  std::sort(arcs1_.begin(), arcs1_.end(), comp_);
+  std::sort(arcs2_.begin(), arcs2_.end(), comp_);
+  for (size_t i = 0; i < arcs1_.size(); ++i) {
+    const auto &arc1 = arcs1_[i];
+    const auto &arc2 = arcs2_[i];
+    if (arc1.ilabel != arc2.ilabel) return false;
+    if (arc1.olabel != arc2.olabel) return false;
+    if (!ApproxEqual(arc1.weight, arc2.weight, delta_)) return false;
+    if (!PairState(arc1.nextstate, arc2.nextstate)) return false;
+    if (i > 0) {  // Checks for non-determinism.
+      const auto &arc0 = arcs1_[i - 1];
+      if (arc1.ilabel == arc0.ilabel && arc1.olabel == arc0.olabel &&
+          ApproxEqual(arc1.weight, arc0.weight, delta_)) {
+        VLOG(1) << "Isomorphic: Non-determinism as an unweighted automaton";
+        error_ = true;
+        return false;
+      }
+    }
+  }
+  return true;
+}
+
+}  // namespace internal
+
+// Tests if two FSTs have the same states and arcs up to a reordering.
+// Inputs should be non-deterministic when viewed as unweighted automata.
+template <class Arc>
+bool Isomorphic(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+                float delta = kDelta) {
+  internal::Isomorphism<Arc> iso(fst1, fst2, delta);
+  bool result = iso.IsIsomorphic();
+  if (iso.Error()) {
+    FSTERROR() << "Isomorphic: Cannot determine if inputs are isomorphic";
+    return false;
+  } else {
+    return result;
+  }
+}
+
+}  // namespace fst
+
+#endif  // FST_ISOMORPHIC_H_
diff --git a/projects/llm_framework/include/fst/label-reachable.h b/projects/llm_framework/include/fst/label-reachable.h
new file mode 100644
index 00000000..f3d7f2bc
--- /dev/null
+++ b/projects/llm_framework/include/fst/label-reachable.h
@@ -0,0 +1,511 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to determine if a non-epsilon label can be read as the first
+// non-epsilon symbol along some path from a given state.
+
+#ifndef FST_LABEL_REACHABLE_H_
+#define FST_LABEL_REACHABLE_H_
+
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/accumulator.h>
+#include <fst/arcsort.h>
+#include <fst/interval-set.h>
+#include <fst/state-reachable.h>
+#include <fst/util.h>
+#include <fst/vector-fst.h>
+
+
+namespace fst {
+
+// Stores shareable data for label reachable class copies.
+template <typename Label>
+class LabelReachableData {
+ public:
+  using LabelIntervalSet = IntervalSet<Label>;
+  using Interval = typename LabelIntervalSet::Interval;
+
+  explicit LabelReachableData(bool reach_input, bool keep_relabel_data = true)
+      : reach_input_(reach_input),
+        keep_relabel_data_(keep_relabel_data),
+        have_relabel_data_(true),
+        final_label_(kNoLabel) {}
+
+  ~LabelReachableData() {}
+
+  bool ReachInput() const { return reach_input_; }
+
+  std::vector<LabelIntervalSet> *MutableIntervalSets() {
+    return &interval_sets_;
+  }
+
+  const LabelIntervalSet &GetIntervalSet(int s) const {
+    return interval_sets_[s];
+  }
+
+  int NumIntervalSets() const { return interval_sets_.size(); }
+
+  std::unordered_map<Label, Label> *Label2Index() {
+    if (!have_relabel_data_) {
+      FSTERROR() << "LabelReachableData: No relabeling data";
+    }
+    return &label2index_;
+  }
+
+  void SetFinalLabel(Label final_label) { final_label_ = final_label; }
+
+  Label FinalLabel() const { return final_label_; }
+
+  static LabelReachableData<Label> *Read(std::istream &istrm,
+                                         const FstReadOptions &opts) {
+    auto *data = new LabelReachableData<Label>();
+    ReadType(istrm, &data->reach_input_);
+    ReadType(istrm, &data->keep_relabel_data_);
+    data->have_relabel_data_ = data->keep_relabel_data_;
+    if (data->keep_relabel_data_) ReadType(istrm, &data->label2index_);
+    ReadType(istrm, &data->final_label_);
+    ReadType(istrm, &data->interval_sets_);
+    return data;
+  }
+
+  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
+    WriteType(ostrm, reach_input_);
+    WriteType(ostrm, keep_relabel_data_);
+    if (keep_relabel_data_) WriteType(ostrm, label2index_);
+    WriteType(ostrm, FinalLabel());
+    WriteType(ostrm, interval_sets_);
+    return true;
+  }
+
+ private:
+  LabelReachableData() {}
+
+  bool reach_input_;                              // Input labels considered?
+  bool keep_relabel_data_;                        // Save label2index_ to file?
+  bool have_relabel_data_;                        // Using label2index_?
+  Label final_label_;                             // Final label.
+  std::unordered_map<Label, Label> label2index_;  // Finds index for a label.
+  std::vector<LabelIntervalSet> interval_sets_;   // Interval sets per state.
+};
+
+// Tests reachability of labels from a given state. If reach_input is true, then
+// input labels are considered, o.w. output labels are considered. To test for
+// reachability from a state s, first do SetState(s), then a label l can be
+// reached from state s of FST f iff Reach(r) is true where r = Relabel(l). The
+// relabeling is required to ensure a compact representation of the reachable
+// labels.
+
+// The whole FST can be relabeled instead with Relabel(&f, reach_input) so that
+// the test Reach(r) applies directly to the labels of the transformed FST f.
+// The relabeled FST will also be sorted appropriately for composition.
+//
+// Reachablity of a final state from state s (via an epsilon path) can be
+// tested with ReachFinal().
+//
+// Reachability can also be tested on the set of labels specified by an arc
+// iterator, useful for FST composition. In particular, Reach(aiter, ...) is
+// true if labels on the input (output) side of the transitions of the arc
+// iterator, when iter_input is true (false), can be reached from the state s.
+// The iterator labels must have already been relabeled.
+//
+// With the arc iterator test of reachability, the begin position, end position
+// and accumulated arc weight of the matches can be returned. The optional
+// template argument controls how reachable arc weights are accumulated. The
+// default uses semiring Plus(). Alternative ones can be used to distribute the
+// weights in composition in various ways.
+template <class Arc, class Accumulator = DefaultAccumulator<Arc>,
+          class D = LabelReachableData<typename Arc::Label>>
+class LabelReachable {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using Data = D;
+
+  using LabelIntervalSet = typename Data::LabelIntervalSet;
+
+  using Interval = typename LabelIntervalSet::Interval;
+
+  LabelReachable(const Fst<Arc> &fst, bool reach_input,
+                 Accumulator *accumulator = nullptr,
+                 bool keep_relabel_data = true)
+      : fst_(new VectorFst<Arc>(fst)),
+        s_(kNoStateId),
+        data_(std::make_shared<Data>(reach_input, keep_relabel_data)),
+        accumulator_(accumulator ? accumulator : new Accumulator()),
+        ncalls_(0),
+        nintervals_(0),
+        reach_fst_input_(false),
+        error_(false) {
+    const auto ins = fst_->NumStates();
+    TransformFst();
+    FindIntervals(ins);
+    fst_.reset();
+  }
+
+  explicit LabelReachable(std::shared_ptr<Data> data,
+                          Accumulator *accumulator = nullptr)
+      : s_(kNoStateId),
+        data_(std::move(data)),
+        accumulator_(accumulator ? accumulator : new Accumulator()),
+        ncalls_(0),
+        nintervals_(0),
+        reach_fst_input_(false),
+        error_(false) {}
+
+  LabelReachable(const LabelReachable<Arc, Accumulator, Data> &reachable,
+                 bool safe = false)
+      : s_(kNoStateId),
+        data_(reachable.data_),
+        accumulator_(new Accumulator(*reachable.accumulator_, safe)),
+        ncalls_(0),
+        nintervals_(0),
+        reach_fst_input_(reachable.reach_fst_input_),
+        error_(reachable.error_) {}
+
+  ~LabelReachable() {
+    if (ncalls_ > 0) {
+      VLOG(2) << "# of calls: " << ncalls_;
+      VLOG(2) << "# of intervals/call: " << (nintervals_ / ncalls_);
+    }
+  }
+
+  // Relabels w.r.t labels that give compact label sets.
+  Label Relabel(Label label) {
+    if (label == 0 || error_) return label;
+    auto &label2index = *data_->Label2Index();
+    auto &relabel = label2index[label];
+    if (!relabel) relabel = label2index.size() + 1;  // Adds new label.
+    return relabel;
+  }
+
+  // Relabels FST w.r.t to labels that give compact label sets.
+  void Relabel(MutableFst<Arc> *fst, bool relabel_input) {
+    for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
+         siter.Next()) {
+      for (MutableArcIterator<MutableFst<Arc>> aiter(fst, siter.Value());
+           !aiter.Done(); aiter.Next()) {
+        auto arc = aiter.Value();
+        if (relabel_input) {
+          arc.ilabel = Relabel(arc.ilabel);
+        } else {
+          arc.olabel = Relabel(arc.olabel);
+        }
+        aiter.SetValue(arc);
+      }
+    }
+    if (relabel_input) {
+      ArcSort(fst, ILabelCompare<Arc>());
+      fst->SetInputSymbols(nullptr);
+    } else {
+      ArcSort(fst, OLabelCompare<Arc>());
+      fst->SetOutputSymbols(nullptr);
+    }
+  }
+
+  // Returns relabeling pairs (cf. relabel.h::Relabel()). If avoid_collisions is
+  // true, extra pairs are added to ensure no collisions when relabeling
+  // automata that have labels unseen here.
+  void RelabelPairs(std::vector<std::pair<Label, Label>> *pairs,
+                    bool avoid_collisions = false) {
+    pairs->clear();
+    const auto &label2index = *data_->Label2Index();
+    // Maps labels to their new values in [1, label2index().size()].
+    for (auto it = label2index.begin(); it != label2index.end(); ++it) {
+      if (it->second != data_->FinalLabel()) {
+        pairs->push_back(std::make_pair(it->first, it->second));
+      }
+    }
+    if (avoid_collisions) {
+      // Ensures any label in [1, label2index().size()] is mapped either
+      // by the above step or to label2index() + 1 (to avoid collisions).
+      for (size_t i = 1; i <= label2index.size(); ++i) {
+        const auto it = label2index.find(i);
+        if (it == label2index.end() || it->second == data_->FinalLabel()) {
+          pairs->push_back(std::make_pair(i, label2index.size() + 1));
+        }
+      }
+    }
+  }
+
+  // Set current state. Optionally set state associated
+  // with arc iterator to be passed to Reach.
+  void SetState(StateId s, StateId aiter_s = kNoStateId) {
+    s_ = s;
+    if (aiter_s != kNoStateId) {
+      accumulator_->SetState(aiter_s);
+      if (accumulator_->Error()) error_ = true;
+    }
+  }
+
+  // Can reach this label from current state?
+  // Original labels must be transformed by the Relabel methods above.
+  bool Reach(Label label) const {
+    if (label == 0 || error_) return false;
+    return data_->GetIntervalSet(s_).Member(label);
+  }
+
+  // Can reach final state (via epsilon transitions) from this state?
+  bool ReachFinal() const {
+    if (error_) return false;
+    return data_->GetIntervalSet(s_).Member(data_->FinalLabel());
+  }
+
+  // Initialize with secondary FST to be used with Reach(Iterator,...).
+  // If reach_input = true, then arc input labels are considered in
+  // Reach(aiter, ...), o.w. output labels are considered. If copy is true, then
+  // the FST is a copy of the FST used in the previous call to this method
+  // (useful to avoid unnecessary updates).
+  template <class FST>
+  void ReachInit(const FST &fst, bool reach_input, bool copy = false) {
+    reach_fst_input_ = reach_input;
+    if (!fst.Properties(reach_fst_input_ ? kILabelSorted : kOLabelSorted,
+                        true)) {
+      FSTERROR() << "LabelReachable::ReachInit: Fst is not sorted";
+      error_ = true;
+    }
+    accumulator_->Init(fst, copy);
+    if (accumulator_->Error()) error_ = true;
+  }
+
+  // Can reach any arc iterator label between iterator positions
+  // aiter_begin and aiter_end?
+  // Arc iterator labels must be transformed by the Relabel methods
+  // above. If compute_weight is true, user may call ReachWeight().
+  template <class Iterator>
+  bool Reach(Iterator *aiter, ssize_t aiter_begin, ssize_t aiter_end,
+             bool compute_weight) {
+    if (error_) return false;
+    const auto &interval_set = data_->GetIntervalSet(s_);
+    ++ncalls_;
+    nintervals_ += interval_set.Size();
+    reach_begin_ = -1;
+    reach_end_ = -1;
+    reach_weight_ = Weight::Zero();
+    const auto flags = aiter->Flags();  // Save flags to restore them on exit.
+    aiter->SetFlags(kArcNoCache, kArcNoCache);  // Makes caching optional.
+    aiter->Seek(aiter_begin);
+    if (2 * (aiter_end - aiter_begin) < interval_set.Size()) {
+      // Checks each arc against intervals, setting arc iterator flags to only
+      // compute the ilabel or olabel values, since they are the only values
+      // required for most of the arcs processed.
+      aiter->SetFlags(reach_fst_input_ ? kArcILabelValue : kArcOLabelValue,
+                      kArcValueFlags);
+      Label reach_label = kNoLabel;
+      for (auto aiter_pos = aiter_begin; aiter_pos < aiter_end;
+           aiter->Next(), ++aiter_pos) {
+        const auto &arc = aiter->Value();
+        const auto label = reach_fst_input_ ? arc.ilabel : arc.olabel;
+        if (label == reach_label || Reach(label)) {
+          reach_label = label;
+          if (reach_begin_ < 0) reach_begin_ = aiter_pos;
+          reach_end_ = aiter_pos + 1;
+          if (compute_weight) {
+            if (!(aiter->Flags() & kArcWeightValue)) {
+              // If arc.weight wasn't computed by the call to aiter->Value()
+              // above, we need to call aiter->Value() again after having set
+              // the arc iterator flags to compute the arc weight value.
+              aiter->SetFlags(kArcWeightValue, kArcValueFlags);
+              const auto &arcb = aiter->Value();
+              // Call the accumulator.
+              reach_weight_ = accumulator_->Sum(reach_weight_, arcb.weight);
+              // Only ilabel or olabel required to process the following arcs.
+              aiter->SetFlags(
+                  reach_fst_input_ ? kArcILabelValue : kArcOLabelValue,
+                  kArcValueFlags);
+            } else {
+              // Calls the accumulator.
+              reach_weight_ = accumulator_->Sum(reach_weight_, arc.weight);
+            }
+          }
+        }
+      }
+    } else {
+      // Checks each interval against arcs.
+      auto begin_low = aiter_begin;
+      auto end_low = aiter_begin;
+      for (const auto &interval : interval_set) {
+        begin_low = LowerBound(aiter, end_low, aiter_end, interval.begin);
+        end_low = LowerBound(aiter, begin_low, aiter_end, interval.end);
+        if (end_low - begin_low > 0) {
+          if (reach_begin_ < 0) reach_begin_ = begin_low;
+          reach_end_ = end_low;
+          if (compute_weight) {
+            aiter->SetFlags(kArcWeightValue, kArcValueFlags);
+            reach_weight_ =
+                accumulator_->Sum(reach_weight_, aiter, begin_low, end_low);
+          }
+        }
+      }
+    }
+    aiter->SetFlags(flags, kArcFlags);  // Restores original flag values.
+    return reach_begin_ >= 0;
+  }
+
+  // Returns iterator position of first matching arc.
+  ssize_t ReachBegin() const { return reach_begin_; }
+
+  // Returns iterator position one past last matching arc.
+  ssize_t ReachEnd() const { return reach_end_; }
+
+  // Return the sum of the weights for matching arcs. Valid only if
+  // compute_weight was true in Reach() call.
+  Weight ReachWeight() const { return reach_weight_; }
+
+  // Access to the relabeling map. Excludes epsilon (0) label but
+  // includes kNoLabel that is used internally for super-final
+  // transitons.
+  const std::unordered_map<Label, Label> &Label2Index() const {
+    return *data_->Label2Index();
+  }
+
+  const Data *GetData() const { return data_.get(); }
+
+  std::shared_ptr<Data> GetSharedData() const { return data_; }
+
+  bool Error() const { return error_ || accumulator_->Error(); }
+
+ private:
+  // Redirects labeled arcs (input or output labels determined by ReachInput())
+  // to new label-specific final states. Each original final state is
+  // redirected via a transition labeled with kNoLabel to a new
+  // kNoLabel-specific final state. Creates super-initial state for all states
+  // with zero in-degree.
+  void TransformFst() {
+    auto ins = fst_->NumStates();
+    auto ons = ins;
+    std::vector<ssize_t> indeg(ins, 0);
+    // Redirects labeled arcs to new final states.
+    for (StateId s = 0; s < ins; ++s) {
+      for (MutableArcIterator<VectorFst<Arc>> aiter(fst_.get(), s);
+           !aiter.Done(); aiter.Next()) {
+        auto arc = aiter.Value();
+        const auto label = data_->ReachInput() ? arc.ilabel : arc.olabel;
+        if (label) {
+          auto insert_result = label2state_.insert(std::make_pair(label, ons));
+          if (insert_result.second) {
+            indeg.push_back(0);
+            ++ons;
+          }
+          arc.nextstate = label2state_[label];
+          aiter.SetValue(arc);
+        }
+        ++indeg[arc.nextstate];  // Finds in-degrees for next step.
+      }
+      // Redirects final weights to new final state.
+      auto final_weight = fst_->Final(s);
+      if (final_weight != Weight::Zero()) {
+        auto insert_result = label2state_.insert(std::make_pair(kNoLabel, ons));
+        if (insert_result.second) {
+          indeg.push_back(0);
+          ++ons;
+        }
+        const auto nextstate = label2state_[kNoLabel];
+        fst_->EmplaceArc(s, kNoLabel, kNoLabel, std::move(final_weight),
+                         nextstate);
+        ++indeg[nextstate];  // Finds in-degrees for next step.
+        fst_->SetFinal(s, Weight::Zero());
+      }
+    }
+    // Adds new final states to the FST.
+    while (fst_->NumStates() < ons) {
+      StateId s = fst_->AddState();
+      fst_->SetFinal(s, Weight::One());
+    }
+    // Creates a super-initial state for all states with zero in-degree.
+    const auto start = fst_->AddState();
+    fst_->SetStart(start);
+    for (StateId s = 0; s < start; ++s) {
+      if (indeg[s] == 0) {
+        fst_->EmplaceArc(start, 0, 0, Weight::One(), s);
+      }
+    }
+  }
+
+  void FindIntervals(StateId ins) {
+    StateReachable<Arc, Label, LabelIntervalSet> state_reachable(*fst_);
+    if (state_reachable.Error()) {
+      error_ = true;
+      return;
+    }
+    auto &state2index = state_reachable.State2Index();
+    auto &interval_sets = *data_->MutableIntervalSets();
+    interval_sets = state_reachable.IntervalSets();
+    interval_sets.resize(ins);
+    auto &label2index = *data_->Label2Index();
+    for (const auto &kv : label2state_) {
+      Label i = state2index[kv.second];
+      label2index[kv.first] = i;
+      if (kv.first == kNoLabel) data_->SetFinalLabel(i);
+    }
+    label2state_.clear();
+    double nintervals = 0;
+    ssize_t non_intervals = 0;
+    for (StateId s = 0; s < ins; ++s) {
+      nintervals += interval_sets[s].Size();
+      if (interval_sets[s].Size() > 1) {
+        ++non_intervals;
+        VLOG(3) << "state: " << s
+                << " # of intervals: " << interval_sets[s].Size();
+      }
+    }
+    VLOG(2) << "# of states: " << ins;
+    VLOG(2) << "# of intervals: " << nintervals;
+    VLOG(2) << "# of intervals/state: " << nintervals / ins;
+    VLOG(2) << "# of non-interval states: " << non_intervals;
+  }
+
+  template <class Iterator>
+  ssize_t LowerBound(Iterator *aiter, ssize_t aiter_begin, ssize_t aiter_end,
+                     Label match_label) const {
+    // Only needs to compute the ilabel or olabel of arcs when performing the
+    // binary search.
+    aiter->SetFlags(reach_fst_input_ ? kArcILabelValue : kArcOLabelValue,
+                    kArcValueFlags);
+    ssize_t low = aiter_begin;
+    ssize_t high = aiter_end;
+    while (low < high) {
+      const ssize_t mid = low + (high - low) / 2;
+      aiter->Seek(mid);
+      auto label =
+          reach_fst_input_ ? aiter->Value().ilabel : aiter->Value().olabel;
+      if (label < match_label) {
+        low = mid + 1;
+      } else {
+        high = mid;
+      }
+    }
+    aiter->Seek(low);
+    aiter->SetFlags(kArcValueFlags, kArcValueFlags);
+    return low;
+  }
+
+  std::unique_ptr<VectorFst<Arc>> fst_;
+  // Current state
+  StateId s_;
+  // Finds final state for a label
+  std::unordered_map<Label, StateId> label2state_;
+  // Iterator position of first match.
+  ssize_t reach_begin_;
+  // Iterator position after last match.
+  ssize_t reach_end_;
+  // Gives weight sum of arc iterator arcs with reachable labels.
+  Weight reach_weight_;
+  // Shareable data between copies.
+  std::shared_ptr<Data> data_;
+  // Sums arc weights.
+  std::unique_ptr<Accumulator> accumulator_;
+  double ncalls_;
+  double nintervals_;
+  bool reach_fst_input_;
+  bool error_;
+};
+
+}  // namespace fst
+
+#endif  // FST_LABEL_REACHABLE_H_
diff --git a/projects/llm_framework/include/fst/lexicographic-weight.h b/projects/llm_framework/include/fst/lexicographic-weight.h
new file mode 100644
index 00000000..af3c6db6
--- /dev/null
+++ b/projects/llm_framework/include/fst/lexicographic-weight.h
@@ -0,0 +1,173 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Lexicographic weight set and associated semiring operation definitions.
+//
+// A lexicographic weight is a sequence of weights, each of which must have the
+// path property and Times() must be (strongly) cancellative
+// (for all a,b,c != Zero(): Times(c, a) = Times(c, b) => a = b,
+// Times(a, c) = Times(b, c) => a = b).
+// The + operation on two weights a and b is the lexicographically
+// prior of a and b.
+
+#ifndef FST_LEXICOGRAPHIC_WEIGHT_H_
+#define FST_LEXICOGRAPHIC_WEIGHT_H_
+
+#include <cstdlib>
+
+#include <string>
+
+#include <fst/log.h>
+
+#include <fst/pair-weight.h>
+#include <fst/weight.h>
+
+
+namespace fst {
+
+template <class W1, class W2>
+class LexicographicWeight : public PairWeight<W1, W2> {
+ public:
+  using ReverseWeight = LexicographicWeight<typename W1::ReverseWeight,
+                                            typename W2::ReverseWeight>;
+
+  using PairWeight<W1, W2>::Value1;
+  using PairWeight<W1, W2>::Value2;
+  using PairWeight<W1, W2>::SetValue1;
+  using PairWeight<W1, W2>::SetValue2;
+  using PairWeight<W1, W2>::Zero;
+  using PairWeight<W1, W2>::One;
+  using PairWeight<W1, W2>::NoWeight;
+  using PairWeight<W1, W2>::Quantize;
+  using PairWeight<W1, W2>::Reverse;
+
+  LexicographicWeight() {}
+
+  explicit LexicographicWeight(const PairWeight<W1, W2> &w)
+      : PairWeight<W1, W2>(w) {}
+
+  LexicographicWeight(W1 w1, W2 w2) : PairWeight<W1, W2>(w1, w2) {
+    if ((W1::Properties() & kPath) != kPath) {
+      FSTERROR() << "LexicographicWeight must "
+                 << "have the path property: " << W1::Type();
+      SetValue1(W1::NoWeight());
+    }
+    if ((W2::Properties() & kPath) != kPath) {
+      FSTERROR() << "LexicographicWeight must "
+                 << "have the path property: " << W2::Type();
+      SetValue2(W2::NoWeight());
+    }
+  }
+
+  static const LexicographicWeight &Zero() {
+    static const LexicographicWeight zero(PairWeight<W1, W2>::Zero());
+    return zero;
+  }
+
+  static const LexicographicWeight &One() {
+    static const LexicographicWeight one(PairWeight<W1, W2>::One());
+    return one;
+  }
+
+  static const LexicographicWeight &NoWeight() {
+    static const LexicographicWeight no_weight(PairWeight<W1, W2>::NoWeight());
+    return no_weight;
+  }
+
+  static const string &Type() {
+    static const string *const type =
+        new string(W1::Type() + "_LT_" + W2::Type());
+    return *type;
+  }
+
+  bool Member() const {
+    if (!Value1().Member() || !Value2().Member()) return false;
+    // Lexicographic weights cannot mix zeroes and non-zeroes.
+    if (Value1() == W1::Zero() && Value2() == W2::Zero()) return true;
+    if (Value1() != W1::Zero() && Value2() != W2::Zero()) return true;
+    return false;
+  }
+
+  LexicographicWeight Quantize(float delta = kDelta) const {
+    return LexicographicWeight(PairWeight<W1, W2>::Quantize());
+  }
+
+  ReverseWeight Reverse() const {
+    return ReverseWeight(PairWeight<W1, W2>::Reverse());
+  }
+
+  static constexpr uint64 Properties() {
+    return W1::Properties() & W2::Properties() &
+           (kLeftSemiring | kRightSemiring | kPath | kIdempotent |
+            kCommutative);
+  }
+};
+
+template <class W1, class W2>
+inline LexicographicWeight<W1, W2> Plus(const LexicographicWeight<W1, W2> &w,
+                                        const LexicographicWeight<W1, W2> &v) {
+  if (!w.Member() || !v.Member()) {
+    return LexicographicWeight<W1, W2>::NoWeight();
+  }
+  NaturalLess<W1> less1;
+  NaturalLess<W2> less2;
+  if (less1(w.Value1(), v.Value1())) return w;
+  if (less1(v.Value1(), w.Value1())) return v;
+  if (less2(w.Value2(), v.Value2())) return w;
+  if (less2(v.Value2(), w.Value2())) return v;
+  return w;
+}
+
+template <class W1, class W2>
+inline LexicographicWeight<W1, W2> Times(const LexicographicWeight<W1, W2> &w,
+                                         const LexicographicWeight<W1, W2> &v) {
+  return LexicographicWeight<W1, W2>(Times(w.Value1(), v.Value1()),
+                                     Times(w.Value2(), v.Value2()));
+}
+
+template <class W1, class W2>
+inline LexicographicWeight<W1, W2> Divide(const LexicographicWeight<W1, W2> &w,
+                                          const LexicographicWeight<W1, W2> &v,
+                                          DivideType typ = DIVIDE_ANY) {
+  return LexicographicWeight<W1, W2>(Divide(w.Value1(), v.Value1(), typ),
+                                     Divide(w.Value2(), v.Value2(), typ));
+}
+
+// This function object generates weights by calling the underlying generators
+// for the templated weight types, like all other pair weight types. However,
+// for lexicographic weights, we cannot generate zeroes for the two subweights
+// separately: weights are members iff both members are zero or both members
+// are non-zero. This is intended primarily for testing.
+template <class W1, class W2>
+class WeightGenerate<LexicographicWeight<W1, W2>> {
+ public:
+  using Weight = LexicographicWeight<W1, W1>;
+  using Generate1 = WeightGenerate<W1>;
+  using Generate2 = WeightGenerate<W2>;
+
+  explicit WeightGenerate(bool allow_zero = true,
+                          size_t num_random_weights = kNumRandomWeights)
+      : generator1_(false, num_random_weights),
+        generator2_(false, num_random_weights), allow_zero_(allow_zero),
+        num_random_weights_(num_random_weights) {}
+
+  Weight operator()() const {
+    if (allow_zero_) {
+      const int n = rand() % (num_random_weights_ + 1);  // NOLINT
+      if (n == num_random_weights_) return Weight(W1::Zero(), W2::Zero());
+    }
+    return Weight(generator1_(), generator2_());
+  }
+
+ private:
+  const Generate1 generator1_;
+  const Generate2 generator2_;
+  // Permits Zero() and zero divisors.
+  const bool allow_zero_;
+  // The number of alternative random weights.
+  const size_t num_random_weights_;
+};
+
+}  // namespace fst
+
+#endif  // FST_LEXICOGRAPHIC_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/lock.h b/projects/llm_framework/include/fst/lock.h
new file mode 100644
index 00000000..388d42f5
--- /dev/null
+++ b/projects/llm_framework/include/fst/lock.h
@@ -0,0 +1,62 @@
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Google-compatibility locking declarations and inline definitions.
+
+#ifndef FST_LIB_LOCK_H_
+#define FST_LIB_LOCK_H_
+
+#include <mutex>
+
+namespace fst {
+
+using namespace std;
+
+class Mutex {
+ public:
+  Mutex() {}
+
+  inline void Lock() { mu_.lock(); }
+
+  inline void Unlock() { mu_.unlock(); }
+
+ private:
+  std::mutex mu_;
+
+  Mutex(const Mutex &) = delete;
+  Mutex &operator=(const Mutex &) = delete;
+};
+
+class MutexLock {
+ public:
+  explicit MutexLock(Mutex *mu) : mu_(mu) { mu_->Lock(); }
+
+  ~MutexLock() { mu_->Unlock(); }
+
+ private:
+  Mutex *mu_;
+
+  MutexLock(const MutexLock &) = delete;
+  MutexLock &operator=(const MutexLock &) = delete;
+};
+
+// Currently, we don't use a separate reader lock.
+// TODO(kbg): Implement this with std::shared_mutex once C++17 becomes widely
+// available.
+using ReaderMutexLock = MutexLock;
+
+}  // namespace fst
+
+#endif  // FST_LIB_LOCK_H_
diff --git a/projects/llm_framework/include/fst/log.h b/projects/llm_framework/include/fst/log.h
new file mode 100644
index 00000000..6a8f3017
--- /dev/null
+++ b/projects/llm_framework/include/fst/log.h
@@ -0,0 +1,78 @@
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Google-style logging declarations and inline definitions.
+
+#ifndef FST_LIB_LOG_H_
+#define FST_LIB_LOG_H_
+
+#include <cassert>
+#include <iostream>
+#include <string>
+
+#include <fst/flags.h>
+#include <fst/types.h>
+
+using std::string;
+
+DECLARE_int32(v);
+
+class LogMessage {
+ public:
+  LogMessage(const string& type) : fatal_(type == "FATAL") {
+    std::cerr << type << ": ";
+  }
+  ~LogMessage() {
+    std::cerr << std::endl;
+    if (fatal_) exit(1);
+  }
+  std::ostream& stream() { return std::cerr; }
+
+ private:
+  bool fatal_;
+};
+
+// #define LOG(type) LogMessage(#type).stream()
+// #define VLOG(level) if ((level) <= FLAGS_v) LOG(INFO)
+
+// Checks
+inline void FstCheck(bool x, const char* expr, const char* file, int line) {
+  if (!x) {
+    LOG(FATAL) << "Check failed: \"" << expr << "\" file: " << file
+               << " line: " << line;
+  }
+}
+
+// #define CHECK(x) FstCheck(static_cast<bool>(x), #x, __FILE__, __LINE__)
+// #define CHECK_EQ(x, y) CHECK((x) == (y))
+// #define CHECK_LT(x, y) CHECK((x) < (y))
+// #define CHECK_GT(x, y) CHECK((x) > (y))
+// #define CHECK_LE(x, y) CHECK((x) <= (y))
+// #define CHECK_GE(x, y) CHECK((x) >= (y))
+// #define CHECK_NE(x, y) CHECK((x) != (y))
+
+// Debug checks
+// #define DCHECK(x) assert(x)
+// #define DCHECK_EQ(x, y) DCHECK((x) == (y))
+// #define DCHECK_LT(x, y) DCHECK((x) < (y))
+// #define DCHECK_GT(x, y) DCHECK((x) > (y))
+// #define DCHECK_LE(x, y) DCHECK((x) <= (y))
+// #define DCHECK_GE(x, y) DCHECK((x) >= (y))
+// #define DCHECK_NE(x, y) DCHECK((x) != (y))
+
+// Ports
+#define ATTRIBUTE_DEPRECATED __attribute__((deprecated))
+
+#endif  // FST_LIB_LOG_H_
diff --git a/projects/llm_framework/include/fst/lookahead-filter.h b/projects/llm_framework/include/fst/lookahead-filter.h
new file mode 100644
index 00000000..df3fc6c9
--- /dev/null
+++ b/projects/llm_framework/include/fst/lookahead-filter.h
@@ -0,0 +1,623 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Composition filters to support lookahead matchers, useful for improving
+// composition efficiency with certain inputs.
+
+#ifndef FST_LOOKAHEAD_FILTER_H_
+#define FST_LOOKAHEAD_FILTER_H_
+
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/filter-state.h>
+#include <fst/fst.h>
+#include <fst/lookahead-matcher.h>
+
+
+namespace fst {
+
+// Identifies and verifies the capabilities of the matcher to be used for
+// lookahead with the composition filters below. This version is passed two
+// matchers.
+template <class Matcher1, class Matcher2>
+MatchType LookAheadMatchType(const Matcher1 &m1, const Matcher2 &m2) {
+  const auto type1 = m1.Type(false);
+  const auto type2 = m2.Type(false);
+  if (type1 == MATCH_OUTPUT && m1.Flags() & kOutputLookAheadMatcher) {
+    return MATCH_OUTPUT;
+  } else if (type2 == MATCH_INPUT && m2.Flags() & kInputLookAheadMatcher) {
+    return MATCH_INPUT;
+  } else if (m1.Flags() & kOutputLookAheadMatcher &&
+             m1.Type(true) == MATCH_OUTPUT) {
+    return MATCH_OUTPUT;
+  } else if (m2.Flags() & kInputLookAheadMatcher &&
+             m2.Type(true) == MATCH_INPUT) {
+    return MATCH_INPUT;
+  } else {
+    return MATCH_NONE;
+  }
+}
+
+// Identifies and verifies the capabilities of the matcher to be used for
+// lookahead with the composition filters below. This version uses the FST's
+// default matchers.
+template <class Arc>
+MatchType LookAheadMatchType(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {
+  LookAheadMatcher<Fst<Arc>> matcher1(fst1, MATCH_OUTPUT);
+  LookAheadMatcher<Fst<Arc>> matcher2(fst2, MATCH_INPUT);
+  return LookAheadMatchType(matcher1, matcher2);
+}
+
+// LookAheadSelector is a helper class for selecting among possibly distinct
+// FST and matcher types without using a common base class. This lets us avoid
+// virtual function calls. It stores and returns the appropriate FSTs and
+// matcher for lookahead. It is templated on the matcher types. General case
+// has no methods.
+template <class Matcher1, class Matcher2, MatchType MT>
+class LookAheadSelector {};
+
+// Stores and returns the appropriate FST and matcher for lookahead. Specialized
+// for two matchers of same type with the (match) type argument determining
+// which is used for lookahead.
+template <class Matcher, MatchType MT>
+class LookAheadSelector<Matcher, Matcher, MT> {
+ public:
+  using FST = typename Matcher::FST;
+
+  LookAheadSelector(Matcher *lmatcher1, Matcher *lmatcher2, MatchType type)
+      : lmatcher1_(lmatcher1->Copy()),
+        lmatcher2_(lmatcher2->Copy()),
+        type_(type) {}
+
+  LookAheadSelector(const LookAheadSelector<Matcher, Matcher, MT> &selector)
+      : lmatcher1_(selector.lmatcher1_->Copy()),
+        lmatcher2_(selector.lmatcher2_->Copy()),
+        type_(selector.type_) {}
+
+  const FST &GetFst() const {
+    return type_ == MATCH_OUTPUT ? lmatcher2_->GetFst() : lmatcher1_->GetFst();
+  }
+
+  Matcher *GetMatcher() const {
+    return type_ == MATCH_OUTPUT ? lmatcher1_.get() : lmatcher2_.get();
+  }
+
+ private:
+  std::unique_ptr<Matcher> lmatcher1_;
+  std::unique_ptr<Matcher> lmatcher2_;
+  MatchType type_;
+};
+
+// Stores and returns the appropriate FST and matcher for lookahead.
+// Specialized for lookahead on input labels.
+template <class Matcher1, class Matcher2>
+class LookAheadSelector<Matcher1, Matcher2, MATCH_INPUT> {
+ public:
+  using FST1 = typename Matcher1::FST;
+
+  LookAheadSelector(Matcher1 *lmatcher1, Matcher2 *lmatcher2, MatchType)
+      : fst_(lmatcher1->GetFst().Copy()), lmatcher_(lmatcher2->Copy()) {}
+
+  LookAheadSelector(
+      const LookAheadSelector<Matcher1, Matcher2, MATCH_INPUT> &selector)
+      : fst_(selector.fst_->Copy()), lmatcher_(selector.lmatcher_->Copy()) {}
+
+  const FST1 &GetFst() const { return *fst_; }
+
+  Matcher2 *GetMatcher() const { return lmatcher_.get(); }
+
+ private:
+  std::unique_ptr<const FST1> fst_;
+  std::unique_ptr<Matcher2> lmatcher_;
+};
+
+// Stores and returns the appropriate FST and matcher for lookahead.
+// Specialized for lookahead on output labels.
+template <class Matcher1, class Matcher2>
+class LookAheadSelector<Matcher1, Matcher2, MATCH_OUTPUT> {
+ public:
+  using FST2 = typename Matcher2::FST;
+
+  LookAheadSelector(Matcher1 *lmatcher1, Matcher2 *lmatcher2, MatchType)
+      : fst_(lmatcher2->GetFst().Copy()), lmatcher_(lmatcher1->Copy()) {}
+
+  LookAheadSelector(
+      const LookAheadSelector<Matcher1, Matcher2, MATCH_OUTPUT> &selector)
+      : fst_(selector.fst_->Copy()), lmatcher_(selector.lmatcher_->Copy()) {}
+
+  const FST2 &GetFst() const { return *fst_; }
+
+  Matcher1 *GetMatcher() const { return lmatcher_.get(); }
+
+ private:
+  std::unique_ptr<const FST2> fst_;
+  std::unique_ptr<Matcher1> lmatcher_;
+};
+
+// This filter uses a lookahead matcher in FilterArc(arc1, arc2) to examine the
+// future of the composition state (arc1.nextstate, arc2.nextstate), blocking
+// moving forward when its determined to be
+// non-coaccessible. It is templated on an underlying filter, typically the
+// epsilon filter. Which matcher is the lookahead matcher is determined by the
+// template argument MT unless it is MATCH_BOTH. In that case, both matcher
+// arguments must be lookahead matchers of the same type and one will be
+// selected by LookAheadMatchType() based on their capability.
+template <class Filter, class M1 = LookAheadMatcher<typename Filter::FST1>,
+          class M2 = M1, MatchType MT = MATCH_BOTH>
+class LookAheadComposeFilter {
+ public:
+  using Arc = typename Filter::Arc;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FST1 = typename Filter::FST1;
+  using FST2 = typename Filter::FST2;
+  using Matcher1 = typename Filter::Matcher1;
+  using Matcher2 = typename Filter::Matcher2;
+  using FilterState = typename Filter::FilterState;
+
+  LookAheadComposeFilter(const FST1 &fst1, const FST2 &fst2, M1 *matcher1,
+                         M2 *matcher2)
+      : filter_(fst1, fst2, matcher1, matcher2),
+        lookahead_type_(MT == MATCH_BOTH
+                            ? LookAheadMatchType(*filter_.GetMatcher1(),
+                                                 *filter_.GetMatcher2())
+                            : MT),
+        selector_(filter_.GetMatcher1(), filter_.GetMatcher2(),
+                  lookahead_type_),
+        flags_(lookahead_type_ == MATCH_OUTPUT
+                   ? filter_.GetMatcher1()->Flags()
+                   : filter_.GetMatcher2()->Flags()) {
+    if (lookahead_type_ == MATCH_NONE) {
+      FSTERROR() << "LookAheadComposeFilter: 1st argument cannot "
+                 << "match/look-ahead on output labels and 2nd argument "
+                 << "cannot match/look-ahead on input labels";
+    }
+    selector_.GetMatcher()->InitLookAheadFst(selector_.GetFst());
+  }
+
+  LookAheadComposeFilter(
+      const LookAheadComposeFilter<Filter, M1, M2, MT> &filter,
+      bool safe = false)
+      : filter_(filter.filter_, safe),
+        lookahead_type_(filter.lookahead_type_),
+        selector_(filter_.GetMatcher1(), filter_.GetMatcher2(),
+                  lookahead_type_),
+        flags_(filter.flags_) {
+    selector_.GetMatcher()->InitLookAheadFst(selector_.GetFst(), true);
+  }
+
+  FilterState Start() const { return filter_.Start(); }
+
+  void SetState(StateId s1, StateId s2, const FilterState &fs) {
+    filter_.SetState(s1, s2, fs);
+  }
+
+  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
+    lookahead_arc_ = false;
+    const FilterState &fs = filter_.FilterArc(arc1, arc2);
+    if (fs == FilterState::NoState()) return FilterState::NoState();
+    return LookAheadOutput() ? LookAheadFilterArc(arc1, arc2, fs)
+                             : LookAheadFilterArc(arc2, arc1, fs);
+  }
+
+  void FilterFinal(Weight *weight1, Weight *weight2) const {
+    filter_.FilterFinal(weight1, weight2);
+  }
+
+  // Returns matchers; ownership stays with filter.
+
+  Matcher1 *GetMatcher1() { return filter_.GetMatcher1(); }
+
+  Matcher2 *GetMatcher2() { return filter_.GetMatcher2(); }
+
+  const LookAheadSelector<Matcher1, Matcher2, MT> &Selector() const {
+    return selector_;
+  }
+
+  uint64 Properties(uint64 inprops) const {
+    auto outprops = filter_.Properties(inprops);
+    if (lookahead_type_ == MATCH_NONE) outprops |= kError;
+    return outprops;
+  }
+
+  uint32 LookAheadFlags() const { return flags_; }
+
+  bool LookAheadArc() const { return lookahead_arc_; }
+
+  bool LookAheadOutput() const {
+    if (MT == MATCH_OUTPUT) {
+      return true;
+    } else if (MT == MATCH_INPUT) {
+      return false;
+    } else if (lookahead_type_ == MATCH_OUTPUT) {
+      return true;
+    } else {
+      return false;
+    }
+  }
+
+ private:
+  FilterState LookAheadFilterArc(Arc *arca, Arc *arcb,
+                                 const FilterState &fs) const {
+    auto &labela = LookAheadOutput() ? arca->olabel : arca->ilabel;
+    if (labela != 0 && !(flags_ & kLookAheadNonEpsilons)) return fs;
+    if (labela == 0 && !(flags_ & kLookAheadEpsilons)) return fs;
+    lookahead_arc_ = true;
+    selector_.GetMatcher()->SetState(arca->nextstate);
+    return selector_.GetMatcher()->LookAheadFst(selector_.GetFst(),
+                                                arcb->nextstate)
+               ? fs
+               : FilterState::NoState();
+  }
+
+  Filter filter_;             // Underlying filter.
+  MatchType lookahead_type_;  // Lookahead match type.
+  LookAheadSelector<Matcher1, Matcher2, MT> selector_;
+  uint32 flags_;                // Lookahead flags.
+  mutable bool lookahead_arc_;  // Look-ahead performed at last FilterArc()?
+
+  LookAheadComposeFilter &operator=(const LookAheadComposeFilter &) = delete;
+};
+
+// This filter adds weight-pushing to a lookahead composition filter using the
+// LookAheadWeight() method of matcher argument. It is templated on an
+// underlying lookahead filter, typically the basic lookahead filter.
+// Weight-pushing in composition brings weights forward as much as possible
+// based on the lookahead information.
+template <class Filter, class M1 = LookAheadMatcher<typename Filter::FST1>,
+          class M2 = M1, MatchType MT = MATCH_BOTH>
+class PushWeightsComposeFilter {
+ public:
+  using Arc = typename Filter::Arc;
+  using StateId = typename Filter::StateId;
+  using Weight = typename Filter::Weight;
+
+  using FST1 = typename Filter::FST1;
+  using FST2 = typename Filter::FST2;
+  using Matcher1 = typename Filter::Matcher1;
+  using Matcher2 = typename Filter::Matcher2;
+
+  using FilterState1 = typename Filter::FilterState;
+  using FilterState2 = WeightFilterState<Weight>;
+  using FilterState = PairFilterState<FilterState1, FilterState2>;
+
+  PushWeightsComposeFilter(const FST1 &fst1, const FST2 &fst2, M1 *matcher1,
+                           M2 *matcher2)
+      : filter_(fst1, fst2, matcher1, matcher2), fs_(FilterState::NoState()) {}
+
+  PushWeightsComposeFilter(
+      const PushWeightsComposeFilter<Filter, M1, M2, MT> &filter,
+      bool safe = false)
+      : filter_(filter.filter_, safe), fs_(FilterState::NoState()) {}
+
+  FilterState Start() const {
+    return FilterState(filter_.Start(), FilterState2(Weight::One()));
+  }
+
+  void SetState(StateId s1, StateId s2, const FilterState &fs) {
+    fs_ = fs;
+    filter_.SetState(s1, s2, fs.GetState1());
+  }
+
+  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
+    const auto &fs1 = filter_.FilterArc(arc1, arc2);
+    if (fs1 == FilterState1::NoState()) return FilterState::NoState();
+    if (!(LookAheadFlags() & kLookAheadWeight)) {
+      return FilterState(fs1, FilterState2(Weight::One()));
+    }
+    const auto &lweight = filter_.LookAheadArc()
+                              ? Selector().GetMatcher()->LookAheadWeight()
+                              : Weight::One();
+    const auto &fs2 = fs_.GetState2();
+    const auto &fweight = fs2.GetWeight();
+    // Disallows Zero() weight futures.
+    if (lweight == Weight::Zero()) return FilterState::NoState();
+    arc2->weight = Divide(Times(arc2->weight, lweight), fweight);
+    return FilterState(fs1, FilterState2(lweight.Quantize()));
+  }
+
+  void FilterFinal(Weight *weight1, Weight *weight2) const {
+    filter_.FilterFinal(weight1, weight2);
+    if (!(LookAheadFlags() & kLookAheadWeight) || *weight1 == Weight::Zero()) {
+      return;
+    }
+    const auto &fs2 = fs_.GetState2();
+    const auto &fweight = fs2.GetWeight();
+    *weight1 = Divide(*weight1, fweight);
+  }
+
+  // Returns matchers; ownership states with filter.
+
+  Matcher1 *GetMatcher1() { return filter_.GetMatcher1(); }
+
+  Matcher2 *GetMatcher2() { return filter_.GetMatcher2(); }
+
+  const LookAheadSelector<Matcher1, Matcher2, MT> &Selector() const {
+    return filter_.Selector();
+  }
+
+  uint32 LookAheadFlags() const { return filter_.LookAheadFlags(); }
+
+  bool LookAheadArc() const { return filter_.LookAheadArc(); }
+
+  bool LookAheadOutput() const { return filter_.LookAheadOutput(); }
+
+  uint64 Properties(uint64 props) const {
+    return filter_.Properties(props) & kWeightInvariantProperties;
+  }
+
+ private:
+  Filter filter_;   // Underlying filter.
+  FilterState fs_;  // Current filter state.
+
+  PushWeightsComposeFilter &operator=(const PushWeightsComposeFilter &) =
+      delete;
+};
+
+// This filter adds label-pushing to a lookahead composition filter using the
+// LookAheadPrefix() method of the matcher argument. It is templated on an
+// underlying filter, typically the basic lookahead or weight-pushing lookahead
+// filter. Label-pushing in composition matches labels as early as possible
+// based on the lookahead information.
+template <class Filter, class M1 = LookAheadMatcher<typename Filter::FST1>,
+          class M2 = M1, MatchType MT = MATCH_BOTH>
+class PushLabelsComposeFilter {
+ public:
+  using Arc = typename Filter::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FST1 = typename Filter::FST1;
+  using FST2 = typename Filter::FST2;
+  using Matcher1 = MultiEpsMatcher<typename Filter::Matcher1>;
+  using Matcher2 = MultiEpsMatcher<typename Filter::Matcher2>;
+  using FilterState1 = typename Filter::FilterState;
+  using FilterState2 = IntegerFilterState<Label>;
+  using FilterState = PairFilterState<FilterState1, FilterState2>;
+
+  PushLabelsComposeFilter(const FST1 &fst1, const FST2 &fst2, M1 *matcher1,
+                          M2 *matcher2)
+      : filter_(fst1, fst2, matcher1, matcher2),
+        fs_(FilterState::NoState()),
+        fst1_(filter_.GetMatcher1()->GetFst()),
+        fst2_(filter_.GetMatcher2()->GetFst()),
+        matcher1_(fst1_, MATCH_OUTPUT,
+                  filter_.LookAheadOutput() ? kMultiEpsList : kMultiEpsLoop,
+                  filter_.GetMatcher1(), false),
+        matcher2_(fst2_, MATCH_INPUT,
+                  filter_.LookAheadOutput() ? kMultiEpsLoop : kMultiEpsList,
+                  filter_.GetMatcher2(), false) {}
+
+  PushLabelsComposeFilter(
+      const PushLabelsComposeFilter<Filter, M1, M2, MT> &filter,
+      bool safe = false)
+      : filter_(filter.filter_, safe),
+        fs_(FilterState::NoState()),
+        fst1_(filter_.GetMatcher1()->GetFst()),
+        fst2_(filter_.GetMatcher2()->GetFst()),
+        matcher1_(fst1_, MATCH_OUTPUT,
+                  filter_.LookAheadOutput() ? kMultiEpsList : kMultiEpsLoop,
+                  filter_.GetMatcher1(), false),
+        matcher2_(fst2_, MATCH_INPUT,
+                  filter_.LookAheadOutput() ? kMultiEpsLoop : kMultiEpsList,
+                  filter_.GetMatcher2(), false) {}
+
+  FilterState Start() const {
+    return FilterState(filter_.Start(), FilterState2(kNoLabel));
+  }
+
+  void SetState(StateId s1, StateId s2, const FilterState &fs) {
+    fs_ = fs;
+    filter_.SetState(s1, s2, fs.GetState1());
+    if (!(LookAheadFlags() & kLookAheadPrefix)) return;
+    narcsa_ = LookAheadOutput() ? internal::NumArcs(fst1_, s1)
+                                : internal::NumArcs(fst2_, s2);
+    const auto &fs2 = fs_.GetState2();
+    const auto &flabel = fs2.GetState();
+    GetMatcher1()->ClearMultiEpsLabels();
+    GetMatcher2()->ClearMultiEpsLabels();
+    if (flabel != kNoLabel) {                   // Have a lookahead label?
+      GetMatcher1()->AddMultiEpsLabel(flabel);  // Yes, make it a multi-epsilon
+      GetMatcher2()->AddMultiEpsLabel(flabel);  // label so that it matches the
+    }                                           // implicit epsilon arc to be
+  }                                             // modified below when pushing.
+
+  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
+    if (!(LookAheadFlags() & kLookAheadPrefix)) {
+      return FilterState(filter_.FilterArc(arc1, arc2), FilterState2(kNoLabel));
+    }
+    const auto &fs2 = fs_.GetState2();
+    const auto &flabel = fs2.GetState();
+    if (flabel != kNoLabel) {  // Have a lookahead label?
+      return LookAheadOutput() ? PushedLabelFilterArc(arc1, arc2, flabel)
+                               : PushedLabelFilterArc(arc2, arc1, flabel);
+    }
+    const auto &fs1 = filter_.FilterArc(arc1, arc2);
+    if (fs1 == FilterState1::NoState()) return FilterState::NoState();
+    if (!filter_.LookAheadArc())
+      return FilterState(fs1, FilterState2(kNoLabel));
+    return LookAheadOutput() ? PushLabelFilterArc(arc1, arc2, fs1)
+                             : PushLabelFilterArc(arc2, arc1, fs1);
+  }
+
+  void FilterFinal(Weight *weight1, Weight *weight2) const {
+    filter_.FilterFinal(weight1, weight2);
+    if (!(LookAheadFlags() & kLookAheadPrefix) || *weight1 == Weight::Zero()) {
+      return;
+    }
+    const auto &fs2 = fs_.GetState2();
+    const auto &flabel = fs2.GetState();
+    if (flabel != kNoLabel) *weight1 = Weight::Zero();
+  }
+
+  // Returns matchers; ownership states with filter.
+
+  Matcher1 *GetMatcher1() { return &matcher1_; }
+
+  Matcher2 *GetMatcher2() { return &matcher2_; }
+
+  uint64 Properties(uint64 iprops) const {
+    const auto oprops = filter_.Properties(iprops);
+    if (LookAheadOutput()) {
+      return oprops & kOLabelInvariantProperties;
+    } else {
+      return oprops & kILabelInvariantProperties;
+    }
+  }
+
+ private:
+  const LookAheadSelector<typename Filter::Matcher1, typename Filter::Matcher2,
+                          MT>
+      &Selector() const {
+    return filter_.Selector();
+  }
+
+  // Consumes an already pushed label.
+  FilterState PushedLabelFilterArc(Arc *arca, Arc *arcb, Label flabel) const {
+    auto &labela = LookAheadOutput() ? arca->olabel : arca->ilabel;
+    const auto &labelb = LookAheadOutput() ? arcb->ilabel : arcb->olabel;
+    if (labelb != kNoLabel) {
+      return FilterState::NoState();  // Blocks non-(multi-)epsilon label
+    } else if (labela == flabel) {
+      labela = 0;  // Converts match to multi-epsilon to epsilon.
+      return Start();
+    } else if (labela == 0) {
+      if (narcsa_ == 1) return fs_;  // Takes epsilon, keeping state with label.
+      Selector().GetMatcher()->SetState(arca->nextstate);
+      if (Selector().GetMatcher()->LookAheadLabel(flabel)) {
+        return fs_;  // Takes epsilon, keeping state with label.
+      } else {
+        return FilterState::NoState();  // Blocks non-coaccessible path.
+      }
+    } else {
+      return FilterState::NoState();  // Blocks mismatch to multi-epsilon label.
+    }
+  }
+
+  // Pushes a label forward when possible.
+  FilterState PushLabelFilterArc(Arc *arca, Arc *arcb,
+                                 const FilterState1 &fs1) const {
+    auto &labela = LookAheadOutput() ? arca->olabel : arca->ilabel;
+    const auto &labelb = LookAheadOutput() ? arcb->olabel : arcb->ilabel;
+    if (labelb != 0) {  // No place to push.
+      return FilterState(fs1, FilterState2(kNoLabel));
+    }
+    if (labela != 0 &&  // Wrong lookahead prefix type?
+        LookAheadFlags() & kLookAheadNonEpsilonPrefix) {
+      return FilterState(fs1, FilterState2(kNoLabel));
+    }
+    Arc larc(kNoLabel, kNoLabel, Weight::Zero(), kNoStateId);
+    if (Selector().GetMatcher()->LookAheadPrefix(&larc)) {  // Have prefix arc?
+      labela = LookAheadOutput() ? larc.ilabel : larc.olabel;
+      arcb->ilabel = larc.ilabel;  // Goes forward on that arc,
+      arcb->olabel = larc.olabel;  // thus pushing the label.
+      arcb->weight = Times(arcb->weight, larc.weight);
+      arcb->nextstate = larc.nextstate;
+      return FilterState(fs1, FilterState2(labela));
+    } else {
+      return FilterState(fs1, FilterState2(kNoLabel));
+    }
+  }
+
+  uint32 LookAheadFlags() const { return filter_.LookAheadFlags(); }
+
+  bool LookAheadArc() const { return filter_.LookAheadArc(); }
+
+  bool LookAheadOutput() const { return filter_.LookAheadOutput(); }
+
+  Filter filter_;   // Underlying filter.
+  FilterState fs_;  // Current filter state.
+  const FST1 &fst1_;
+  const FST2 &fst2_;
+  Matcher1 matcher1_;  // Multi-epsilon matcher for fst1_.
+  Matcher2 matcher2_;  // Multi-epsilon matcher for fst2_.
+  ssize_t narcsa_;     // Number of arcs leaving look-ahead match FST.
+
+  PushLabelsComposeFilter &operator=(const PushLabelsComposeFilter &) = delete;
+};
+
+// Convenience class for setting up composition with a default lookahead matcher
+// and filter.
+template <class Arc, MatchType type>
+class DefaultLookAhead {
+ public:
+  using M = Matcher<Fst<Arc>>;
+  using ComposeFilter = SequenceComposeFilter<M>;
+  using FstMatcher = M;
+};
+
+// Specializes for MATCH_INPUT to allow lookahead.
+template <class Arc>
+class DefaultLookAhead<Arc, MATCH_INPUT> {
+ public:
+  using M = LookAheadMatcher<Fst<Arc>>;
+  using SF = SequenceComposeFilter<M>;
+  using ComposeFilter = LookAheadComposeFilter<SF, M>;
+  using FstMatcher = M;
+};
+
+// Specializes for MATCH_OUTPUT to allow lookahead.
+template <class Arc>
+class DefaultLookAhead<Arc, MATCH_OUTPUT> {
+ public:
+  using M = LookAheadMatcher<Fst<Arc>>;
+  using SF = AltSequenceComposeFilter<M>;
+  using ComposeFilter = LookAheadComposeFilter<SF, M>;
+  using FstMatcher = M;
+};
+
+// Specializes for StdArc to allow weight and label pushing.
+template <>
+class DefaultLookAhead<StdArc, MATCH_INPUT> {
+ public:
+  using M = LookAheadMatcher<Fst<StdArc>>;
+  using SF = SequenceComposeFilter<M>;
+  using LF = LookAheadComposeFilter<SF, M>;
+  using WF = PushWeightsComposeFilter<LF, M>;
+  using ComposeFilter = PushLabelsComposeFilter<WF, M>;
+  using FstMatcher = M;
+};
+
+// Specializes for StdArc to allow weight and label pushing.
+template <>
+class DefaultLookAhead<StdArc, MATCH_OUTPUT> {
+ public:
+  using M = LookAheadMatcher<Fst<StdArc>>;
+  using SF = AltSequenceComposeFilter<M>;
+  using LF = LookAheadComposeFilter<SF, M>;
+  using WF = PushWeightsComposeFilter<LF, M>;
+  using ComposeFilter = PushLabelsComposeFilter<WF, M>;
+  using FstMatcher = M;
+};
+
+// Specializes for LogArc to allow weight and label pushing.
+template <>
+class DefaultLookAhead<LogArc, MATCH_INPUT> {
+ public:
+  using M = LookAheadMatcher<Fst<LogArc>>;
+  using SF = SequenceComposeFilter<M>;
+  using LF = LookAheadComposeFilter<SF, M>;
+  using WF = PushWeightsComposeFilter<LF, M>;
+  using ComposeFilter = PushLabelsComposeFilter<WF, M>;
+  using FstMatcher = M;
+};
+
+// Specializes for LogArc to allow weight and label pushing.
+template <>
+class DefaultLookAhead<LogArc, MATCH_OUTPUT> {
+ public:
+  using M = LookAheadMatcher<Fst<LogArc>>;
+  using SF = AltSequenceComposeFilter<M>;
+  using LF = LookAheadComposeFilter<SF, M>;
+  using WF = PushWeightsComposeFilter<LF, M>;
+  using ComposeFilter = PushLabelsComposeFilter<WF, M>;
+  using FstMatcher = M;
+};
+
+}  // namespace fst
+
+#endif  // FST_LOOKAHEAD_FILTER_H_
diff --git a/projects/llm_framework/include/fst/lookahead-matcher.h b/projects/llm_framework/include/fst/lookahead-matcher.h
new file mode 100644
index 00000000..d542cfea
--- /dev/null
+++ b/projects/llm_framework/include/fst/lookahead-matcher.h
@@ -0,0 +1,841 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Classes to add lookahead to FST matchers, useful for improving composition
+// efficiency with certain inputs.
+
+#ifndef FST_LOOKAHEAD_MATCHER_H_
+#define FST_LOOKAHEAD_MATCHER_H_
+
+#include <memory>
+#include <utility>
+#include <vector>
+
+#include <fst/flags.h>
+#include <fst/log.h>
+
+#include <fst/add-on.h>
+#include <fst/const-fst.h>
+#include <fst/fst.h>
+#include <fst/label-reachable.h>
+#include <fst/matcher.h>
+
+
+DECLARE_string(save_relabel_ipairs);
+DECLARE_string(save_relabel_opairs);
+
+namespace fst {
+
+// Lookahead matches extend the matcher interface with following additional
+// methods:
+//
+// template <class FST>
+// class LookAheadMatcher {
+//  public:
+//   using Arc = typename FST::Arc;
+//   using Label = typename Arc::Label;
+//   using StateId = typename Arc::StateId;
+//   using Weight = typename Arc::Weight;
+//
+//  // Required constructors.
+//  // This makes a copy of the FST.
+//  LookAheadMatcher(const FST &fst, MatchType match_type);
+//  // This doesn't copy the FST.
+//  LookAheadMatcher(const FST *fst, MatchType match_type);
+//  // This makes a copy of the FST.
+//  // See Copy() below.
+//  LookAheadMatcher(const LookAheadMatcher &matcher, bool safe = false);
+//
+//   // If safe = true, the copy is thread-safe (except the lookahead FST is
+//   // preserved). See Fst<>::Copy() for further doc.
+//   LookaheadMatcher<FST> *Copy(bool safe = false) const override;
+
+//  // Below are methods for looking ahead for a match to a label and more
+//  // generally, to a rational set. Each returns false if there is definitely
+//  // not a match and returns true if there possibly is a match.
+//
+//  // Optionally pre-specifies the lookahead FST that will be passed to
+//  // LookAheadFst() for possible precomputation. If copy is true, then the FST
+//  // argument is a copy of the FST used in the previous call to this method
+//  // (to avoid unnecessary updates).
+//  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) override;
+//
+//  // Are there paths from a state in the lookahead FST that can be read from
+//  // the curent matcher state?
+//  bool LookAheadFst(const Fst<Arc> &fst, StateId s) override;
+//
+//  // Can the label be read from the current matcher state after possibly
+//  // following epsilon transitions?
+//  bool LookAheadLabel(Label label) const override;
+//
+//  // The following methods allow looking ahead for an arbitrary rational set
+//  // of strings, specified by an FST and a state from which to begin the
+//  // matching. If the lookahead FST is a transducer, this looks on the side
+//  // different from the matcher's match_type (cf. composition).
+//  // Is there is a single non-epsilon arc found in the lookahead FST that
+//  // begins the path (after possibly following any epsilons) in the last call
+//  // to LookAheadFst? If so, return true and copy it to the arc argument;
+//  // otherwise, return false. Non-trivial implementations are useful for
+//  // label-pushing in composition.
+//  bool LookAheadPrefix(Arc *arc) override;
+//
+//  // Gives an estimate of the combined weight of the paths in the lookahead
+//  // and matcher FSTs for the last call to LookAheadFst. Non-trivial
+//  // implementations are useful for weight-pushing in composition.
+//  Weight LookAheadWeight() const override;
+// };
+
+// Look-ahead flags.
+// Matcher is a lookahead matcher when match_type is MATCH_INPUT.
+constexpr uint32 kInputLookAheadMatcher = 0x00000010;
+
+// Matcher is a lookahead matcher when match_type is MATCH_OUTPUT.
+constexpr uint32 kOutputLookAheadMatcher = 0x00000020;
+
+// Is a non-trivial implementation of LookAheadWeight() method defined and
+// if so, should it be used?
+constexpr uint32 kLookAheadWeight = 0x00000040;
+
+// Is a non-trivial implementation of LookAheadPrefix() method defined and
+// if so, should it be used?
+constexpr uint32 kLookAheadPrefix = 0x00000080;
+
+// Look-ahead of matcher FST non-epsilon arcs?
+constexpr uint32 kLookAheadNonEpsilons = 0x00000100;
+
+// Look-ahead of matcher FST epsilon arcs?
+constexpr uint32 kLookAheadEpsilons = 0x00000200;
+
+// Ignore epsilon paths for the lookahead prefix? This gives correct results in
+// composition only with an appropriate composition filter since it depends on
+// the filter blocking the ignored paths.
+constexpr uint32 kLookAheadNonEpsilonPrefix = 0x00000400;
+
+// For LabelLookAheadMatcher, save relabeling data to file?
+constexpr uint32 kLookAheadKeepRelabelData = 0x00000800;
+
+// Flags used for lookahead matchers.
+constexpr uint32 kLookAheadFlags = 0x00000ff0;
+
+// LookAhead Matcher interface, templated on the Arc definition; used
+// for lookahead matcher specializations that are returned by the
+// InitMatcher() Fst method.
+template <class Arc>
+class LookAheadMatcherBase : public MatcherBase<Arc> {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  virtual void InitLookAheadFst(const Fst<Arc> &, bool copy = false) = 0;
+  virtual bool LookAheadFst(const Fst<Arc> &, StateId) = 0;
+  virtual bool LookAheadLabel(Label) const = 0;
+
+  // Suggested concrete implementation of lookahead methods.
+
+  bool LookAheadPrefix(Arc *arc) const {
+    if (prefix_arc_.nextstate != kNoStateId) {
+      *arc = prefix_arc_;
+      return true;
+    } else {
+      return false;
+    }
+  }
+
+  Weight LookAheadWeight() const { return weight_; }
+
+ protected:
+  // Concrete implementations for lookahead helper methods.
+
+  void ClearLookAheadWeight() { weight_ = Weight::One(); }
+
+  void SetLookAheadWeight(Weight weight) { weight_ = std::move(weight); }
+
+  void ClearLookAheadPrefix() { prefix_arc_.nextstate = kNoStateId; }
+
+  void SetLookAheadPrefix(Arc arc) { prefix_arc_ = std::move(arc); }
+
+ private:
+  Arc prefix_arc_;
+  Weight weight_;
+};
+
+// Doesn't actually lookahead, just declares that the future looks good.
+template <class M>
+class TrivialLookAheadMatcher
+    : public LookAheadMatcherBase<typename M::FST::Arc> {
+ public:
+  using FST = typename M::FST;
+  using Arc = typename FST::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  // This makes a copy of the FST.
+  TrivialLookAheadMatcher(const FST &fst, MatchType match_type)
+      : matcher_(fst, match_type) {}
+
+  // This doesn't copy the FST.
+  TrivialLookAheadMatcher(const FST *fst, MatchType match_type)
+      : matcher_(fst, match_type) {}
+
+  // This makes a copy of the FST.
+  TrivialLookAheadMatcher(const TrivialLookAheadMatcher<M> &lmatcher,
+                          bool safe = false)
+      : matcher_(lmatcher.matcher_, safe) {}
+
+  TrivialLookAheadMatcher<M> *Copy(bool safe = false) const override {
+    return new TrivialLookAheadMatcher<M>(*this, safe);
+  }
+
+  MatchType Type(bool test) const override { return matcher_.Type(test); }
+
+  void SetState(StateId s) final { return matcher_.SetState(s); }
+
+  bool Find(Label label) final { return matcher_.Find(label); }
+
+  bool Done() const final { return matcher_.Done(); }
+
+  const Arc &Value() const final { return matcher_.Value(); }
+
+  void Next() final { matcher_.Next(); }
+
+  Weight Final(StateId s) const final { return matcher_.Final(s); }
+
+  ssize_t Priority(StateId s) final { return matcher_.Priority(s); }
+
+  const FST &GetFst() const override { return matcher_.GetFst(); }
+
+  uint64 Properties(uint64 props) const override {
+    return matcher_.Properties(props);
+  }
+
+  uint32 Flags() const override {
+    return matcher_.Flags() | kInputLookAheadMatcher | kOutputLookAheadMatcher;
+  }
+
+  // Lookahead methods (all trivial).
+
+  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) override {}
+
+  bool LookAheadFst(const Fst<Arc> &, StateId) final { return true; }
+
+  bool LookAheadLabel(Label) const final { return true; }
+
+  bool LookAheadPrefix(Arc *) const { return false; }
+
+  Weight LookAheadWeight() const { return Weight::One(); }
+
+ private:
+  M matcher_;
+};
+
+// Look-ahead of one transition. Template argument flags accepts flags to
+// control behavior.
+template <class M,
+          uint32 flags = kLookAheadNonEpsilons | kLookAheadEpsilons |
+                         kLookAheadWeight | kLookAheadPrefix>
+class ArcLookAheadMatcher : public LookAheadMatcherBase<typename M::FST::Arc> {
+ public:
+  using FST = typename M::FST;
+  using Arc = typename FST::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using MatcherData = NullAddOn;
+
+  using LookAheadMatcherBase<Arc>::ClearLookAheadWeight;
+  using LookAheadMatcherBase<Arc>::LookAheadWeight;
+  using LookAheadMatcherBase<Arc>::SetLookAheadWeight;
+  using LookAheadMatcherBase<Arc>::ClearLookAheadPrefix;
+  using LookAheadMatcherBase<Arc>::LookAheadPrefix;
+  using LookAheadMatcherBase<Arc>::SetLookAheadPrefix;
+
+  enum : uint32 { kFlags = flags };
+
+  // This makes a copy of the FST.
+  ArcLookAheadMatcher(
+      const FST &fst, MatchType match_type,
+      std::shared_ptr<MatcherData> data = std::shared_ptr<MatcherData>())
+      : matcher_(fst, match_type),
+        fst_(matcher_.GetFst()),
+        lfst_(nullptr),
+        state_(kNoStateId) {}
+
+  // This doesn't copy the FST.
+  ArcLookAheadMatcher(
+      const FST *fst, MatchType match_type,
+      std::shared_ptr<MatcherData> data = std::shared_ptr<MatcherData>())
+      : matcher_(fst, match_type),
+        fst_(matcher_.GetFst()),
+        lfst_(nullptr),
+        state_(kNoStateId) {}
+
+  // This makes a copy of the FST.
+  ArcLookAheadMatcher(const ArcLookAheadMatcher<M, flags> &lmatcher,
+                      bool safe = false)
+      : matcher_(lmatcher.matcher_, safe),
+        fst_(matcher_.GetFst()),
+        lfst_(lmatcher.lfst_),
+        state_(kNoStateId) {}
+
+  // General matcher methods.
+  ArcLookAheadMatcher<M, flags> *Copy(bool safe = false) const override {
+    return new ArcLookAheadMatcher<M, flags>(*this, safe);
+  }
+
+  MatchType Type(bool test) const override { return matcher_.Type(test); }
+
+  void SetState(StateId s) final {
+    state_ = s;
+    matcher_.SetState(s);
+  }
+
+  bool Find(Label label) final { return matcher_.Find(label); }
+
+  bool Done() const final { return matcher_.Done(); }
+
+  const Arc &Value() const final { return matcher_.Value(); }
+
+  void Next() final { matcher_.Next(); }
+
+  Weight Final(StateId s) const final { return matcher_.Final(s); }
+
+  ssize_t Priority(StateId s) final { return matcher_.Priority(s); }
+
+  const FST &GetFst() const override { return fst_; }
+
+  uint64 Properties(uint64 props) const override {
+    return matcher_.Properties(props);
+  }
+
+  uint32 Flags() const override {
+    return matcher_.Flags() | kInputLookAheadMatcher | kOutputLookAheadMatcher |
+           kFlags;
+  }
+
+  const MatcherData *GetData() const { return nullptr; }
+
+  std::shared_ptr<MatcherData> GetSharedData() const {
+    return std::shared_ptr<MatcherData>();
+  }
+
+  // Look-ahead methods.
+
+  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) override {
+    lfst_ = &fst;
+  }
+
+  // Checks if there is a matching (possibly super-final) transition
+  // at (state_, s).
+  bool LookAheadFst(const Fst<Arc> &, StateId) final;
+
+  bool LookAheadLabel(Label label) const final { return matcher_.Find(label); }
+
+ private:
+  mutable M matcher_;
+  const FST &fst_;        // Matcher FST.
+  const Fst<Arc> *lfst_;  // Look-ahead FST.
+  StateId state_;         // Matcher state.
+};
+
+template <class M, uint32 flags>
+bool ArcLookAheadMatcher<M, flags>::LookAheadFst(const Fst<Arc> &fst,
+                                                 StateId s) {
+  if (&fst != lfst_) InitLookAheadFst(fst);
+  bool result = false;
+  ssize_t nprefix = 0;
+  if (kFlags & kLookAheadWeight) ClearLookAheadWeight();
+  if (kFlags & kLookAheadPrefix) ClearLookAheadPrefix();
+  if (fst_.Final(state_) != Weight::Zero() &&
+      lfst_->Final(s) != Weight::Zero()) {
+    if (!(kFlags & (kLookAheadWeight | kLookAheadPrefix))) return true;
+    ++nprefix;
+    if (kFlags & kLookAheadWeight) {
+      SetLookAheadWeight(
+          Plus(LookAheadWeight(), Times(fst_.Final(state_), lfst_->Final(s))));
+    }
+    result = true;
+  }
+  if (matcher_.Find(kNoLabel)) {
+    if (!(kFlags & (kLookAheadWeight | kLookAheadPrefix))) return true;
+    ++nprefix;
+    if (kFlags & kLookAheadWeight) {
+      for (; !matcher_.Done(); matcher_.Next()) {
+        SetLookAheadWeight(Plus(LookAheadWeight(), matcher_.Value().weight));
+      }
+    }
+    result = true;
+  }
+  for (ArcIterator<Fst<Arc>> aiter(*lfst_, s); !aiter.Done(); aiter.Next()) {
+    const auto &arc = aiter.Value();
+    Label label = kNoLabel;
+    switch (matcher_.Type(false)) {
+      case MATCH_INPUT:
+        label = arc.olabel;
+        break;
+      case MATCH_OUTPUT:
+        label = arc.ilabel;
+        break;
+      default:
+        FSTERROR() << "ArcLookAheadMatcher::LookAheadFst: Bad match type";
+        return true;
+    }
+    if (label == 0) {
+      if (!(kFlags & (kLookAheadWeight | kLookAheadPrefix))) return true;
+      if (!(kFlags & kLookAheadNonEpsilonPrefix)) ++nprefix;
+      if (kFlags & kLookAheadWeight) {
+        SetLookAheadWeight(Plus(LookAheadWeight(), arc.weight));
+      }
+      result = true;
+    } else if (matcher_.Find(label)) {
+      if (!(kFlags & (kLookAheadWeight | kLookAheadPrefix))) return true;
+      for (; !matcher_.Done(); matcher_.Next()) {
+        ++nprefix;
+        if (kFlags & kLookAheadWeight) {
+          SetLookAheadWeight(Plus(LookAheadWeight(),
+                                  Times(arc.weight, matcher_.Value().weight)));
+        }
+        if ((kFlags & kLookAheadPrefix) && nprefix == 1)
+          SetLookAheadPrefix(arc);
+      }
+      result = true;
+    }
+  }
+  if (kFlags & kLookAheadPrefix) {
+    if (nprefix == 1) {
+      ClearLookAheadWeight();  // Avoids double counting.
+    } else {
+      ClearLookAheadPrefix();
+    }
+  }
+  return result;
+}
+
+// Template argument flags accepts flags to control behavior. It must include
+// precisely one of kInputLookAheadMatcher or kOutputLookAheadMatcher.
+template <class M,
+          uint32 flags = kLookAheadEpsilons | kLookAheadWeight |
+                         kLookAheadPrefix | kLookAheadNonEpsilonPrefix |
+                         kLookAheadKeepRelabelData,
+          class Accumulator = DefaultAccumulator<typename M::Arc>,
+          class Reachable = LabelReachable<typename M::Arc, Accumulator>>
+class LabelLookAheadMatcher
+    : public LookAheadMatcherBase<typename M::FST::Arc> {
+ public:
+  using FST = typename M::FST;
+  using Arc = typename FST::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using MatcherData = typename Reachable::Data;
+
+  using LookAheadMatcherBase<Arc>::ClearLookAheadWeight;
+  using LookAheadMatcherBase<Arc>::LookAheadWeight;
+  using LookAheadMatcherBase<Arc>::SetLookAheadWeight;
+  using LookAheadMatcherBase<Arc>::ClearLookAheadPrefix;
+  using LookAheadMatcherBase<Arc>::LookAheadPrefix;
+  using LookAheadMatcherBase<Arc>::SetLookAheadPrefix;
+
+  enum : uint32 { kFlags = flags };
+
+  // This makes a copy of the FST.
+  LabelLookAheadMatcher(
+      const FST &fst, MatchType match_type,
+      std::shared_ptr<MatcherData> data = std::shared_ptr<MatcherData>(),
+      Accumulator *accumulator = nullptr)
+      : matcher_(fst, match_type),
+        lfst_(nullptr),
+        state_(kNoStateId),
+        error_(false) {
+    Init(fst, match_type, data, accumulator);
+  }
+
+  // This doesn't copy the FST.
+  LabelLookAheadMatcher(
+      const FST *fst, MatchType match_type,
+      std::shared_ptr<MatcherData> data = std::shared_ptr<MatcherData>(),
+      Accumulator *accumulator = nullptr)
+      : matcher_(fst, match_type),
+        lfst_(nullptr),
+        state_(kNoStateId),
+        error_(false) {
+    Init(*fst, match_type, data, accumulator);
+  }
+
+  // This makes a copy of the FST.
+  LabelLookAheadMatcher(
+      const LabelLookAheadMatcher<M, flags, Accumulator, Reachable> &lmatcher,
+      bool safe = false)
+      : matcher_(lmatcher.matcher_, safe),
+        lfst_(lmatcher.lfst_),
+        label_reachable_(lmatcher.label_reachable_
+                             ? new Reachable(*lmatcher.label_reachable_, safe)
+                             : nullptr),
+        state_(kNoStateId),
+        error_(lmatcher.error_) {}
+
+  LabelLookAheadMatcher<M, flags, Accumulator, Reachable> *Copy(
+      bool safe = false) const override {
+    return new LabelLookAheadMatcher<M, flags, Accumulator, Reachable>(*this,
+                                                                       safe);
+  }
+
+  MatchType Type(bool test) const override { return matcher_.Type(test); }
+
+  void SetState(StateId s) final {
+    if (state_ == s) return;
+    state_ = s;
+    match_set_state_ = false;
+    reach_set_state_ = false;
+  }
+
+  bool Find(Label label) final {
+    if (!match_set_state_) {
+      matcher_.SetState(state_);
+      match_set_state_ = true;
+    }
+    return matcher_.Find(label);
+  }
+
+  bool Done() const final { return matcher_.Done(); }
+
+  const Arc &Value() const final { return matcher_.Value(); }
+
+  void Next() final { matcher_.Next(); }
+
+  Weight Final(StateId s) const final { return matcher_.Final(s); }
+
+  ssize_t Priority(StateId s) final { return matcher_.Priority(s); }
+
+  const FST &GetFst() const override { return matcher_.GetFst(); }
+
+  uint64 Properties(uint64 inprops) const override {
+    auto outprops = matcher_.Properties(inprops);
+    if (error_ || (label_reachable_ && label_reachable_->Error())) {
+      outprops |= kError;
+    }
+    return outprops;
+  }
+
+  uint32 Flags() const override {
+    if (label_reachable_ && label_reachable_->GetData()->ReachInput()) {
+      return matcher_.Flags() | kFlags | kInputLookAheadMatcher;
+    } else if (label_reachable_ && !label_reachable_->GetData()->ReachInput()) {
+      return matcher_.Flags() | kFlags | kOutputLookAheadMatcher;
+    } else {
+      return matcher_.Flags();
+    }
+  }
+
+  const MatcherData *GetData() const {
+    return label_reachable_ ? label_reachable_->GetData() : nullptr;
+  };
+
+  std::shared_ptr<MatcherData> GetSharedData() const {
+    return label_reachable_ ? label_reachable_->GetSharedData()
+                            : std::shared_ptr<MatcherData>();
+  }
+  // Checks if there is a matching (possibly super-final) transition at
+  // (state_, s).
+  template <class LFST>
+  bool LookAheadFst(const LFST &fst, StateId s);
+
+  // Required to make class concrete.
+  bool LookAheadFst(const Fst<Arc> &fst, StateId s) final {
+    return LookAheadFst<Fst<Arc>>(fst, s);
+  }
+
+  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) override {
+    lfst_ = &fst;
+    if (label_reachable_) {
+      const bool reach_input = Type(false) == MATCH_OUTPUT;
+      label_reachable_->ReachInit(fst, reach_input, copy);
+    }
+  }
+
+  template <class LFST>
+  void InitLookAheadFst(const LFST &fst, bool copy = false) {
+    lfst_ = static_cast<const Fst<Arc> *>(&fst);
+    if (label_reachable_) {
+      const bool reach_input = Type(false) == MATCH_OUTPUT;
+      label_reachable_->ReachInit(fst, reach_input, copy);
+    }
+  }
+
+  bool LookAheadLabel(Label label) const final {
+    if (label == 0) return true;
+    if (label_reachable_) {
+      if (!reach_set_state_) {
+        label_reachable_->SetState(state_);
+        reach_set_state_ = true;
+      }
+      return label_reachable_->Reach(label);
+    } else {
+      return true;
+    }
+  }
+
+ private:
+  void Init(const FST &fst, MatchType match_type,
+            std::shared_ptr<MatcherData> data,
+            Accumulator *accumulator) {
+    if (!(kFlags & (kInputLookAheadMatcher | kOutputLookAheadMatcher))) {
+      FSTERROR() << "LabelLookaheadMatcher: Bad matcher flags: " << kFlags;
+      error_ = true;
+    }
+    const bool reach_input = match_type == MATCH_INPUT;
+    if (data) {
+      if (reach_input == data->ReachInput()) {
+        label_reachable_.reset(new Reachable(data, accumulator));
+      }
+    } else if ((reach_input && (kFlags & kInputLookAheadMatcher)) ||
+               (!reach_input && (kFlags & kOutputLookAheadMatcher))) {
+      label_reachable_.reset(new Reachable(fst, reach_input, accumulator,
+                                           kFlags & kLookAheadKeepRelabelData));
+    }
+  }
+
+  mutable M matcher_;
+  const Fst<Arc> *lfst_;                        // Look-ahead FST.
+  std::unique_ptr<Reachable> label_reachable_;  // Label reachability info.
+  StateId state_;                               // Matcher state.
+  bool match_set_state_;                        // matcher_.SetState called?
+  mutable bool reach_set_state_;                // reachable_.SetState called?
+  bool error_;                                  // Error encountered?
+};
+
+template <class M, uint32 flags, class Accumulator, class Reachable>
+template <class LFST>
+inline bool LabelLookAheadMatcher<M, flags, Accumulator,
+                                  Reachable>::LookAheadFst(const LFST &fst,
+                                                           StateId s) {
+  if (static_cast<const Fst<Arc> *>(&fst) != lfst_) InitLookAheadFst(fst);
+  ClearLookAheadWeight();
+  ClearLookAheadPrefix();
+  if (!label_reachable_) return true;
+  label_reachable_->SetState(state_, s);
+  reach_set_state_ = true;
+  bool compute_weight = kFlags & kLookAheadWeight;
+  bool compute_prefix = kFlags & kLookAheadPrefix;
+  ArcIterator<LFST> aiter(fst, s);
+  aiter.SetFlags(kArcNoCache, kArcNoCache);  // Makes caching optional.
+  const bool reach_arc = label_reachable_->Reach(
+      &aiter, 0, internal::NumArcs(*lfst_, s), compute_weight);
+  const auto lfinal = internal::Final(*lfst_, s);
+  const bool reach_final =
+      lfinal != Weight::Zero() && label_reachable_->ReachFinal();
+  if (reach_arc) {
+    const auto begin = label_reachable_->ReachBegin();
+    const auto end = label_reachable_->ReachEnd();
+    if (compute_prefix && end - begin == 1 && !reach_final) {
+      aiter.Seek(begin);
+      SetLookAheadPrefix(aiter.Value());
+      compute_weight = false;
+    } else if (compute_weight) {
+      SetLookAheadWeight(label_reachable_->ReachWeight());
+    }
+  }
+  if (reach_final && compute_weight) {
+    SetLookAheadWeight(reach_arc ? Plus(LookAheadWeight(), lfinal) : lfinal);
+  }
+  return reach_arc || reach_final;
+}
+
+// Label-lookahead relabeling class.
+template <class Arc, class Data = LabelReachableData<typename Arc::Label>>
+class LabelLookAheadRelabeler {
+ public:
+  using Label = typename Arc::Label;
+  using Reachable = LabelReachable<Arc, DefaultAccumulator<Arc>, Data>;
+
+  // Relabels matcher FST (initialization function object).
+  template <typename Impl>
+  explicit LabelLookAheadRelabeler(std::shared_ptr<Impl> *impl);
+
+  // Relabels arbitrary FST. Class LFST should be a label-lookahead FST.
+  template <class LFST>
+  static void Relabel(MutableFst<Arc> *fst, const LFST &mfst,
+                      bool relabel_input) {
+    const auto *data = mfst.GetAddOn();
+    Reachable reachable(data->First() ? data->SharedFirst()
+                                      : data->SharedSecond());
+    reachable.Relabel(fst, relabel_input);
+  }
+
+  // Returns relabeling pairs (cf. relabel.h::Relabel()). Class LFST should be a
+  // label-lookahead FST. If avoid_collisions is true, extra pairs are added to
+  // ensure no collisions when relabeling automata that have labels unseen here.
+  template <class LFST>
+  static void RelabelPairs(const LFST &mfst,
+                           std::vector<std::pair<Label, Label>> *pairs,
+                           bool avoid_collisions = false) {
+    const auto *data = mfst.GetAddOn();
+    Reachable reachable(data->First() ? data->SharedFirst()
+                                      : data->SharedSecond());
+    reachable.RelabelPairs(pairs, avoid_collisions);
+  }
+};
+
+template <class Arc, class Data>
+template <typename Impl>
+inline LabelLookAheadRelabeler<Arc, Data>::LabelLookAheadRelabeler(
+    std::shared_ptr<Impl> *impl) {
+  Fst<Arc> &fst = (*impl)->GetFst();
+  auto data = (*impl)->GetSharedAddOn();
+  const auto name = (*impl)->Type();
+  const bool is_mutable = fst.Properties(kMutable, false);
+  std::unique_ptr<MutableFst<Arc>> mfst;
+  if (is_mutable) {
+    mfst.reset(static_cast<MutableFst<Arc> *>(&fst));
+  } else {
+    mfst.reset(new VectorFst<Arc>(fst));
+  }
+  if (data->First()) {  // reach_input.
+    Reachable reachable(data->SharedFirst());
+    reachable.Relabel(mfst.get(), true);
+    if (!FLAGS_save_relabel_ipairs.empty()) {
+      std::vector<std::pair<Label, Label>> pairs;
+      reachable.RelabelPairs(&pairs, true);
+      WriteLabelPairs(FLAGS_save_relabel_ipairs, pairs);
+    }
+  } else {
+    Reachable reachable(data->SharedSecond());
+    reachable.Relabel(mfst.get(), false);
+    if (!FLAGS_save_relabel_opairs.empty()) {
+      std::vector<std::pair<Label, Label>> pairs;
+      reachable.RelabelPairs(&pairs, true);
+      WriteLabelPairs(FLAGS_save_relabel_opairs, pairs);
+    }
+  }
+  if (!is_mutable) {
+    *impl = std::make_shared<Impl>(*mfst, name);
+    (*impl)->SetAddOn(data);
+  }
+}
+
+// Generic lookahead matcher, templated on the FST definition (a wrapper around
+// a pointer to specific one).
+template <class F>
+class LookAheadMatcher {
+ public:
+  using FST = F;
+  using Arc = typename FST::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using LBase = LookAheadMatcherBase<Arc>;
+
+  // This makes a copy of the FST.
+  LookAheadMatcher(const FST &fst, MatchType match_type)
+      : owned_fst_(fst.Copy()),
+        base_(owned_fst_->InitMatcher(match_type)),
+        lookahead_(false) {
+    if (!base_) base_.reset(new SortedMatcher<FST>(owned_fst_.get(),
+                                                   match_type));
+  }
+
+  // This doesn't copy the FST.
+  LookAheadMatcher(const FST *fst, MatchType match_type)
+      : base_(fst->InitMatcher(match_type)),
+        lookahead_(false) {
+    if (!base_) base_.reset(new SortedMatcher<FST>(fst, match_type));
+  }
+
+  // This makes a copy of the FST.
+  LookAheadMatcher(const LookAheadMatcher<FST> &matcher, bool safe = false)
+      : base_(matcher.base_->Copy(safe)),
+        lookahead_(matcher.lookahead_) { }
+
+  // Takes ownership of base.
+  explicit LookAheadMatcher(MatcherBase<Arc> *base)
+      : base_(base), lookahead_(false) {}
+
+  LookAheadMatcher<FST> *Copy(bool safe = false) const {
+    return new LookAheadMatcher<FST>(*this, safe);
+  }
+
+  MatchType Type(bool test) const { return base_->Type(test); }
+
+  void SetState(StateId s) { base_->SetState(s); }
+
+  bool Find(Label label) { return base_->Find(label); }
+
+  bool Done() const { return base_->Done(); }
+
+  const Arc &Value() const { return base_->Value(); }
+
+  void Next() { base_->Next(); }
+
+  Weight Final(StateId s) const { return base_->Final(s); }
+
+  ssize_t Priority(StateId s) { return base_->Priority(s); }
+
+  const FST &GetFst() const {
+    return static_cast<const FST &>(base_->GetFst());
+  }
+
+  uint64 Properties(uint64 props) const { return base_->Properties(props); }
+
+  uint32 Flags() const { return base_->Flags(); }
+
+  bool LookAheadLabel(Label label) const {
+    if (LookAheadCheck()) {
+      return static_cast<LBase *>(base_.get())->LookAheadLabel(label);
+    } else {
+      return true;
+    }
+  }
+
+  bool LookAheadFst(const Fst<Arc> &fst, StateId s) {
+    if (LookAheadCheck()) {
+      return static_cast<LBase *>(base_.get())->LookAheadFst(fst, s);
+    } else {
+      return true;
+    }
+  }
+
+  Weight LookAheadWeight() const {
+    if (LookAheadCheck()) {
+      return static_cast<LBase *>(base_.get())->LookAheadWeight();
+    } else {
+      return Weight::One();
+    }
+  }
+
+  bool LookAheadPrefix(Arc *arc) const {
+    if (LookAheadCheck()) {
+      return static_cast<LBase *>(base_.get())->LookAheadPrefix(arc);
+    } else {
+      return false;
+    }
+  }
+
+  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) {
+    if (LookAheadCheck()) {
+      static_cast<LBase *>(base_.get())->InitLookAheadFst(fst, copy);
+    }
+  }
+
+ private:
+  bool LookAheadCheck() const {
+    if (!lookahead_) {
+      lookahead_ =
+          base_->Flags() & (kInputLookAheadMatcher | kOutputLookAheadMatcher);
+      if (!lookahead_) {
+        FSTERROR() << "LookAheadMatcher: No look-ahead matcher defined";
+      }
+    }
+    return lookahead_;
+  }
+
+  std::unique_ptr<const FST> owned_fst_;
+  std::unique_ptr<MatcherBase<Arc>> base_;
+  mutable bool lookahead_;
+
+  LookAheadMatcher &operator=(const LookAheadMatcher &) = delete;
+};
+
+}  // namespace fst
+
+#endif  // FST_LOOKAHEAD_MATCHER_H_
diff --git a/projects/llm_framework/include/fst/map.h b/projects/llm_framework/include/fst/map.h
new file mode 100644
index 00000000..5d663ae8
--- /dev/null
+++ b/projects/llm_framework/include/fst/map.h
@@ -0,0 +1,110 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Compatibility file for old-style Map() functions and MapFst class that have
+// been renamed to ArcMap (cf. StateMap).
+
+#ifndef FST_MAP_H_
+#define FST_MAP_H_
+
+
+#include <fst/arc-map.h>
+
+
+namespace fst {
+
+template <class A, class C>
+void Map(MutableFst<A> *fst, C *mapper) {
+  ArcMap(fst, mapper);
+}
+
+template <class A, class C>
+void Map(MutableFst<A> *fst, C mapper) {
+  ArcMap(fst, mapper);
+}
+
+template <class A, class B, class C>
+void Map(const Fst<A> &ifst, MutableFst<B> *ofst, C *mapper) {
+  ArcMap(ifst, ofst, mapper);
+}
+
+template <class A, class B, class C>
+void Map(const Fst<A> &ifst, MutableFst<B> *ofst, C mapper) {
+  ArcMap(ifst, ofst, mapper);
+}
+
+using MapFstOptions = ArcMapFstOptions;
+
+template <class A, class B, class C>
+class MapFst : public ArcMapFst<A, B, C> {
+ public:
+  using FromArc = A;
+  using ToArc = B;
+
+  using StateId = typename ToArc::StateId;
+  using Weight = typename ToArc::Weight;
+
+  using State = CacheState<B>;
+
+  MapFst(const Fst<A> &fst, const C &mapper, const MapFstOptions &opts)
+      : ArcMapFst<A, B, C>(fst, mapper, opts) {}
+
+  MapFst(const Fst<A> &fst, C *mapper, const MapFstOptions &opts)
+      : ArcMapFst<A, B, C>(fst, mapper, opts) {}
+
+  MapFst(const Fst<A> &fst, const C &mapper)
+      : ArcMapFst<A, B, C>(fst, mapper) {}
+
+  MapFst(const Fst<A> &fst, C *mapper) : ArcMapFst<A, B, C>(fst, mapper) {}
+
+  // See Fst<>::Copy() for doc.
+  MapFst(const MapFst<A, B, C> &fst, bool safe = false)
+      : ArcMapFst<A, B, C>(fst, safe) {}
+
+  // Get a copy of this MapFst. See Fst<>::Copy() for further doc.
+  MapFst<A, B, C> *Copy(bool safe = false) const override {
+    return new MapFst(*this, safe);
+  }
+};
+
+// Specialization for MapFst.
+template <class A, class B, class C>
+class StateIterator<MapFst<A, B, C>>
+    : public StateIterator<ArcMapFst<A, B, C>> {
+ public:
+  explicit StateIterator(const ArcMapFst<A, B, C> &fst)
+      : StateIterator<ArcMapFst<A, B, C>>(fst) {}
+};
+
+// Specialization for MapFst.
+template <class A, class B, class C>
+class ArcIterator<MapFst<A, B, C>> : public ArcIterator<ArcMapFst<A, B, C>> {
+ public:
+  ArcIterator(const ArcMapFst<A, B, C> &fst, typename A::StateId s)
+      : ArcIterator<ArcMapFst<A, B, C>>(fst, s) {}
+};
+
+// For backwards compatibility only; use IdentityArcMapper otherwise.
+template <class A>
+struct IdentityMapper {
+  using FromArc = A;
+  using ToArc = A;
+
+  ToArc operator()(const FromArc &arc) const { return arc; }
+
+  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  uint64 Properties(uint64 props) const { return props; }
+};
+
+}  // namespace fst
+
+#endif  // FST_MAP_H_
diff --git a/projects/llm_framework/include/fst/mapped-file.h b/projects/llm_framework/include/fst/mapped-file.h
new file mode 100644
index 00000000..adb33c28
--- /dev/null
+++ b/projects/llm_framework/include/fst/mapped-file.h
@@ -0,0 +1,81 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_MAPPED_FILE_H_
+#define FST_MAPPED_FILE_H_
+
+#include <cstddef>
+#include <istream>
+#include <string>
+
+#include <fst/compat.h>
+#include <fst/flags.h>
+
+namespace fst {
+
+// A memory region is a simple abstraction for allocated memory or data from
+// memory-mapped files. If mmap is null, then data represents an owned region
+// of size bytes. Otherwise, mmap and size refer to the mapping and data is a
+// casted pointer to a region contained within [mmap, mmap + size). If size is
+// 0, then mmap and data refer to a block of memory managed externally by some
+// other allocator. The offset is used when allocating memory to providing
+// padding for alignment.
+struct MemoryRegion {
+  void *data;
+  void *mmap;
+  size_t size;
+  int offset;
+};
+
+class MappedFile {
+ public:
+  ~MappedFile();
+
+  void *mutable_data() const { return region_.data; }
+
+  const void *data() const { return region_.data; }
+
+  // Returns a MappedFile object that contains the contents of the input stream
+  // strm starting from the current file position with size bytes. The memorymap
+  // bool is advisory, and Map will default to allocating and reading. The
+  // source argument needs to contain the filename that was used to open the
+  // input stream.
+  static MappedFile *Map(std::istream *istrm, bool memorymap,
+                         const string &source, size_t size);
+
+  // Returns a MappedFile object that contains the contents of the file referred
+  // to by the file descriptor starting from pos with size bytes. If the
+  // memory mapping fails, nullptr is returned. In contrast to Map(), this
+  // factory function does not backoff to allocating and reading.
+  static MappedFile *MapFromFileDescriptor(int fd, int pos, size_t size);
+
+  // Creates a MappedFile object with a new[]'ed block of memory of size. The
+  // align argument can be used to specify a desired block alignment.
+  // This is RECOMMENDED FOR INTERNAL USE ONLY as it may change in future
+  // releases.
+  static MappedFile *Allocate(size_t size, int align = kArchAlignment);
+
+  // Creates a MappedFile object pointing to a borrowed reference to data. This
+  // block of memory is not owned by the MappedFile object and will not be
+  // freed. This is RECOMMENDED FOR INTERNAL USE ONLY, may change in future
+  // releases.
+  static MappedFile *Borrow(void *data);
+
+  // Alignment required for mapping structures in bytes. Regions of memory that
+  // are not aligned upon a 128-bit boundary are read from the file instead.
+  // This is consistent with the alignment boundary set in ConstFst and
+  // CompactFst.
+  static constexpr int kArchAlignment = 16;
+
+  static constexpr size_t kMaxReadChunk = 256 * 1024 * 1024;  // 256 MB.
+
+ private:
+  explicit MappedFile(const MemoryRegion &region);
+
+  MemoryRegion region_;
+  MappedFile(const MappedFile &) = delete;
+  MappedFile &operator=(const MappedFile &) = delete;
+};
+}  // namespace fst
+
+#endif  // FST_MAPPED_FILE_H_
diff --git a/projects/llm_framework/include/fst/matcher-fst.h b/projects/llm_framework/include/fst/matcher-fst.h
new file mode 100644
index 00000000..61e95820
--- /dev/null
+++ b/projects/llm_framework/include/fst/matcher-fst.h
@@ -0,0 +1,347 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to add a matcher to an FST.
+
+#ifndef FST_MATCHER_FST_H_
+#define FST_MATCHER_FST_H_
+
+#include <memory>
+#include <string>
+
+#include <fst/add-on.h>
+#include <fst/const-fst.h>
+#include <fst/lookahead-matcher.h>
+
+
+namespace fst {
+
+// Writeable matchers have the same interface as Matchers (as defined in
+// matcher.h) along with the following additional methods:
+//
+// template <class F>
+// class Matcher {
+//  public:
+//   using FST = F;
+//   ...
+//   using MatcherData = ...;   // Initialization data.
+//
+//   // Constructor with additional argument for external initialization data;
+//   // matcher increments its reference count on construction and decrements
+//   // the reference count, and deletes once the reference count has reached
+//   // zero.
+//   Matcher(const FST &fst, MatchType type, MatcherData *data);
+//
+//   // Returns pointer to initialization data that can be passed to a Matcher
+//   // constructor.
+//   MatcherData *GetData() const;
+// };
+
+// The matcher initialization data class must also provide the following
+// interface:
+//
+// class MatcherData {
+// public:
+//   // Required copy constructor.
+//   MatcherData(const MatcherData &);
+//
+//   // Required I/O methods.
+//   static MatcherData *Read(std::istream &istrm, const FstReadOptions &opts);
+//   bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const;
+// };
+
+// Trivial (no-op) MatcherFst initializer functor.
+template <class M>
+class NullMatcherFstInit {
+ public:
+  using MatcherData = typename M::MatcherData;
+  using Data = AddOnPair<MatcherData, MatcherData>;
+  using Impl = internal::AddOnImpl<typename M::FST, Data>;
+
+  explicit NullMatcherFstInit(std::shared_ptr<Impl> *) {}
+};
+
+// Class adding a matcher to an FST type. Creates a new FST whose name is given
+// by N. An optional functor Init can be used to initialize the FST. The Data
+// template parameter allows the user to select the type of the add-on.
+template <
+    class F, class M, const char *Name, class Init = NullMatcherFstInit<M>,
+    class Data = AddOnPair<typename M::MatcherData, typename M::MatcherData>>
+class MatcherFst : public ImplToExpandedFst<internal::AddOnImpl<F, Data>> {
+ public:
+  using FST = F;
+  using Arc = typename FST::Arc;
+  using StateId = typename Arc::StateId;
+
+  using FstMatcher = M;
+  using MatcherData = typename FstMatcher::MatcherData;
+
+  using Impl = internal::AddOnImpl<FST, Data>;
+  using D = Data;
+
+  friend class StateIterator<MatcherFst<FST, FstMatcher, Name, Init, Data>>;
+  friend class ArcIterator<MatcherFst<FST, FstMatcher, Name, Init, Data>>;
+
+  MatcherFst() : ImplToExpandedFst<Impl>(std::make_shared<Impl>(FST(), Name)) {}
+
+  explicit MatcherFst(const FST &fst, std::shared_ptr<Data> data = nullptr)
+      : ImplToExpandedFst<Impl>(data ? CreateImpl(fst, Name, data)
+                                     : CreateDataAndImpl(fst, Name)) {}
+
+  explicit MatcherFst(const Fst<Arc> &fst)
+      : ImplToExpandedFst<Impl>(CreateDataAndImpl(fst, Name)) {}
+
+  // See Fst<>::Copy() for doc.
+  MatcherFst(const MatcherFst<FST, FstMatcher, Name, Init, Data> &fst,
+             bool safe = false)
+      : ImplToExpandedFst<Impl>(fst, safe) {}
+
+  // Get a copy of this MatcherFst. See Fst<>::Copy() for further doc.
+  MatcherFst<FST, FstMatcher, Name, Init, Data> *Copy(
+      bool safe = false) const override {
+    return new MatcherFst<FST, FstMatcher, Name, Init, Data>(*this, safe);
+  }
+
+  // Read a MatcherFst from an input stream; return nullptr on error
+  static MatcherFst<FST, M, Name, Init, Data> *Read(
+      std::istream &strm, const FstReadOptions &opts) {
+    auto *impl = Impl::Read(strm, opts);
+    return impl ? new MatcherFst<FST, FstMatcher, Name, Init, Data>(
+                      std::shared_ptr<Impl>(impl))
+                : nullptr;
+  }
+
+  // Read a MatcherFst from a file; return nullptr on error
+  // Empty filename reads from standard input
+  static MatcherFst<FST, FstMatcher, Name, Init, Data> *Read(
+      const string &filename) {
+    auto *impl = ImplToExpandedFst<Impl>::Read(filename);
+    return impl ? new MatcherFst<FST, FstMatcher, Name, Init, Data>(
+                      std::shared_ptr<Impl>(impl))
+                : nullptr;
+  }
+
+  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
+    return GetImpl()->Write(strm, opts);
+  }
+
+  bool Write(const string &filename) const override {
+    return Fst<Arc>::WriteFile(filename);
+  }
+
+  void InitStateIterator(StateIteratorData<Arc> *data) const override {
+    return GetImpl()->InitStateIterator(data);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
+    return GetImpl()->InitArcIterator(s, data);
+  }
+
+  FstMatcher *InitMatcher(MatchType match_type) const override {
+    return new FstMatcher(&GetFst(), match_type, GetSharedData(match_type));
+  }
+
+  const FST &GetFst() const { return GetImpl()->GetFst(); }
+
+  const Data *GetAddOn() const { return GetImpl()->GetAddOn(); }
+
+  std::shared_ptr<Data> GetSharedAddOn() const {
+    return GetImpl()->GetSharedAddOn();
+  }
+
+  const MatcherData *GetData(MatchType match_type) const {
+    const auto *data = GetAddOn();
+    return match_type == MATCH_INPUT ? data->First() : data->Second();
+  }
+
+  std::shared_ptr<MatcherData> GetSharedData(MatchType match_type) const {
+    const auto *data = GetAddOn();
+    return match_type == MATCH_INPUT ? data->SharedFirst()
+                                     : data->SharedSecond();
+  }
+
+ protected:
+  using ImplToFst<Impl, ExpandedFst<Arc>>::GetImpl;
+
+  static std::shared_ptr<Impl> CreateDataAndImpl(const FST &fst,
+                                                 const string &name) {
+    FstMatcher imatcher(fst, MATCH_INPUT);
+    FstMatcher omatcher(fst, MATCH_OUTPUT);
+    return CreateImpl(fst, name,
+                      std::make_shared<Data>(imatcher.GetSharedData(),
+                                             omatcher.GetSharedData()));
+  }
+
+  static std::shared_ptr<Impl> CreateDataAndImpl(const Fst<Arc> &fst,
+                                                 const string &name) {
+    FST result(fst);
+    return CreateDataAndImpl(result, name);
+  }
+
+  static std::shared_ptr<Impl> CreateImpl(const FST &fst, const string &name,
+                                          std::shared_ptr<Data> data) {
+    auto impl = std::make_shared<Impl>(fst, name);
+    impl->SetAddOn(data);
+    Init init(&impl);
+    return impl;
+  }
+
+  explicit MatcherFst(std::shared_ptr<Impl> impl)
+      : ImplToExpandedFst<Impl>(impl) {}
+
+ private:
+  MatcherFst &operator=(const MatcherFst &) = delete;
+};
+
+// Specialization for MatcherFst.
+template <class FST, class M, const char *Name, class Init>
+class StateIterator<MatcherFst<FST, M, Name, Init>>
+    : public StateIterator<FST> {
+ public:
+  explicit StateIterator(const MatcherFst<FST, M, Name, Init> &fst)
+      : StateIterator<FST>(fst.GetImpl()->GetFst()) {}
+};
+
+// Specialization for MatcherFst.
+template <class FST, class M, const char *Name, class Init>
+class ArcIterator<MatcherFst<FST, M, Name, Init>> : public ArcIterator<FST> {
+ public:
+  using StateId = typename FST::Arc::StateId;
+
+  ArcIterator(const MatcherFst<FST, M, Name, Init> &fst,
+              typename FST::Arc::StateId s)
+      : ArcIterator<FST>(fst.GetImpl()->GetFst(), s) {}
+};
+
+// Specialization for MatcherFst.
+template <class F, class M, const char *Name, class Init>
+class Matcher<MatcherFst<F, M, Name, Init>> {
+ public:
+  using FST = MatcherFst<F, M, Name, Init>;
+  using Arc = typename F::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+
+  Matcher(const FST &fst, MatchType match_type)
+      : matcher_(fst.InitMatcher(match_type)) {}
+
+  Matcher(const Matcher<FST> &matcher) : matcher_(matcher.matcher_->Copy()) {}
+
+  Matcher<FST> *Copy() const { return new Matcher<FST>(*this); }
+
+  MatchType Type(bool test) const { return matcher_->Type(test); }
+
+  void SetState(StateId s) { matcher_->SetState(s); }
+
+  bool Find(Label label) { return matcher_->Find(label); }
+
+  bool Done() const { return matcher_->Done(); }
+
+  const Arc &Value() const { return matcher_->Value(); }
+
+  void Next() { matcher_->Next(); }
+
+  uint64 Properties(uint64 props) const { return matcher_->Properties(props); }
+
+  uint32 Flags() const { return matcher_->Flags(); }
+
+ private:
+  std::unique_ptr<M> matcher_;
+};
+
+// Specialization for MatcherFst.
+template <class F, class M, const char *Name, class Init>
+class LookAheadMatcher<MatcherFst<F, M, Name, Init>> {
+ public:
+  using FST = MatcherFst<F, M, Name, Init>;
+  using Arc = typename F::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  LookAheadMatcher(const FST &fst, MatchType match_type)
+      : matcher_(fst.InitMatcher(match_type)) {}
+
+  LookAheadMatcher(const LookAheadMatcher<FST> &matcher, bool safe = false)
+      : matcher_(matcher.matcher_->Copy(safe)) {}
+
+  // General matcher methods.
+  LookAheadMatcher<FST> *Copy(bool safe = false) const {
+    return new LookAheadMatcher<FST>(*this, safe);
+  }
+
+  MatchType Type(bool test) const { return matcher_->Type(test); }
+
+  void SetState(StateId s) { matcher_->SetState(s); }
+
+  bool Find(Label label) { return matcher_->Find(label); }
+
+  bool Done() const { return matcher_->Done(); }
+
+  const Arc &Value() const { return matcher_->Value(); }
+
+  void Next() { matcher_->Next(); }
+
+  const FST &GetFst() const { return matcher_->GetFst(); }
+
+  uint64 Properties(uint64 props) const { return matcher_->Properties(props); }
+
+  uint32 Flags() const { return matcher_->Flags(); }
+
+  bool LookAheadLabel(Label label) const {
+    return matcher_->LookAheadLabel(label);
+  }
+
+  bool LookAheadFst(const Fst<Arc> &fst, StateId s) {
+    return matcher_->LookAheadFst(fst, s);
+  }
+
+  Weight LookAheadWeight() const { return matcher_->LookAheadWeight(); }
+
+  bool LookAheadPrefix(Arc *arc) const {
+    return matcher_->LookAheadPrefix(arc);
+  }
+
+  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) {
+    matcher_->InitLookAheadFst(fst, copy);
+  }
+
+ private:
+  std::unique_ptr<M> matcher_;
+};
+
+// Useful aliases when using StdArc.
+
+extern const char arc_lookahead_fst_type[];
+
+using StdArcLookAheadFst =
+    MatcherFst<ConstFst<StdArc>,
+               ArcLookAheadMatcher<SortedMatcher<ConstFst<StdArc>>>,
+               arc_lookahead_fst_type>;
+
+extern const char ilabel_lookahead_fst_type[];
+extern const char olabel_lookahead_fst_type[];
+
+constexpr auto ilabel_lookahead_flags =
+    kInputLookAheadMatcher | kLookAheadWeight | kLookAheadPrefix |
+    kLookAheadEpsilons | kLookAheadNonEpsilonPrefix;
+
+constexpr auto olabel_lookahead_flags =
+    kOutputLookAheadMatcher | kLookAheadWeight | kLookAheadPrefix |
+    kLookAheadEpsilons | kLookAheadNonEpsilonPrefix;
+
+using StdILabelLookAheadFst = MatcherFst<
+    ConstFst<StdArc>,
+    LabelLookAheadMatcher<SortedMatcher<ConstFst<StdArc>>,
+                          ilabel_lookahead_flags, FastLogAccumulator<StdArc>>,
+    ilabel_lookahead_fst_type, LabelLookAheadRelabeler<StdArc>>;
+
+using StdOLabelLookAheadFst = MatcherFst<
+    ConstFst<StdArc>,
+    LabelLookAheadMatcher<SortedMatcher<ConstFst<StdArc>>,
+                          olabel_lookahead_flags, FastLogAccumulator<StdArc>>,
+    olabel_lookahead_fst_type, LabelLookAheadRelabeler<StdArc>>;
+
+}  // namespace fst
+
+#endif  // FST_MATCHER_FST_H_
diff --git a/projects/llm_framework/include/fst/matcher.h b/projects/llm_framework/include/fst/matcher.h
new file mode 100644
index 00000000..d9528d68
--- /dev/null
+++ b/projects/llm_framework/include/fst/matcher.h
@@ -0,0 +1,1575 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Classes to allow matching labels leaving FST states.
+
+#ifndef FST_MATCHER_H_
+#define FST_MATCHER_H_
+
+#include <algorithm>
+#include <memory>
+#include <unordered_map>
+#include <utility>
+
+#include <fst/log.h>
+
+#include <fst/mutable-fst.h>  // for all internal FST accessors.
+
+
+namespace fst {
+
+// Matchers find and iterate through requested labels at FST states. In the
+// simplest form, these are just some associative map or search keyed on labels.
+// More generally, they may implement matching special labels that represent
+// sets of labels such as sigma (all), rho (rest), or phi (fail). The Matcher
+// interface is:
+//
+// template <class F>
+// class Matcher {
+//  public:
+//   using FST = F;
+//   using Arc = typename FST::Arc;
+//   using Label = typename Arc::Label;
+//   using StateId = typename Arc::StateId;
+//   using Weight = typename Arc::Weight;
+//
+//   // Required constructors. Note:
+//   // -- the constructors that copy the FST arg are useful for
+//   // letting the matcher manage the FST through copies
+//   // (esp with 'safe' copies); e.g. ComposeFst depends on this.
+//   // -- the constructor that does not copy is useful when the
+//   // the FST is mutated during the lifetime of the matcher
+//   // (o.w. the matcher would have its own unmutated deep copy).
+//
+//   // This makes a copy of the FST.
+//   Matcher(const FST &fst, MatchType type);
+//   // This doesn't copy the FST.
+//   Matcher(const FST *fst, MatchType type);
+//   // This makes a copy of the FST.
+//   // See Copy() below.
+//   Matcher(const Matcher &matcher, bool safe = false);
+//
+//   // If safe = true, the copy is thread-safe. See Fst<>::Copy() for
+//   // further doc.
+//   Matcher<FST> *Copy(bool safe = false) const override;
+//
+//   // Returns the match type that can be provided (depending on compatibility
+//   of the input FST). It is either the requested match type, MATCH_NONE, or
+//   MATCH_UNKNOWN. If test is false, a costly testing is avoided, but
+//   MATCH_UNKNOWN may be returned. If test is true, a definite answer is
+//   returned, but may involve more costly computation (e.g., visiting the FST).
+//   MatchType Type(bool test) const override;
+//
+//   // Specifies the current state.
+//   void SetState(StateId s) final;
+//
+//   // Finds matches to a label at the current state, returning true if a match
+//   // found. kNoLabel matches any non-consuming transitions, e.g., epsilon
+//   // transitions, which do not require a matching symbol.
+//   bool Find(Label label) final;
+//
+//   // Iterator methods. Note that initially and after SetState() these have
+//   undefined behavior until Find() is called.
+//
+//   bool Done() const final;
+//
+//   const Arc &Value() const final;
+//
+//   void Next() final;
+//
+//   // Returns final weight of a state.
+//   Weight Final(StateId) const final;
+//
+//   // Indicates preference for being the side used for matching in
+//   // composition. If the value is kRequirePriority, then it is
+//   // mandatory that it be used. Calling this method without passing the
+//   // current state of the matcher invalidates the state of the matcher.
+//   ssize_t Priority(StateId s) final;
+//
+//   // This specifies the known FST properties as viewed from this matcher. It
+//   // takes as argument the input FST's known properties.
+//   uint64 Properties(uint64 props) const override;
+//
+//   // Returns matcher flags.
+//   uint32 Flags() const override;
+//
+//   // Returns matcher FST.
+//   const FST &GetFst() const override;
+// };
+
+// Basic matcher flags.
+
+// Matcher needs to be used as the matching side in composition for
+// at least one state (has kRequirePriority).
+constexpr uint32 kRequireMatch = 0x00000001;
+
+// Flags used for basic matchers (see also lookahead.h).
+constexpr uint32 kMatcherFlags = kRequireMatch;
+
+// Matcher priority that is mandatory.
+constexpr ssize_t kRequirePriority = -1;
+
+// Matcher interface, templated on the Arc definition; used for matcher
+// specializations that are returned by the InitMatcher FST method.
+template <class A>
+class MatcherBase {
+ public:
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  virtual ~MatcherBase() {}
+
+  // Virtual interface.
+
+  virtual MatcherBase<Arc> *Copy(bool safe = false) const = 0;
+  virtual MatchType Type(bool) const = 0;
+  virtual void SetState(StateId) = 0;
+  virtual bool Find(Label) = 0;
+  virtual bool Done() const = 0;
+  virtual const Arc &Value() const = 0;
+  virtual void Next() = 0;
+  virtual const Fst<Arc> &GetFst() const = 0;
+  virtual uint64 Properties(uint64) const = 0;
+
+  // Trivial implementations that can be used by derived classes. Full
+  // devirtualization is expected for any derived class marked final.
+  virtual uint32 Flags() const { return 0; }
+
+  virtual Weight Final(StateId s) const { return internal::Final(GetFst(), s); }
+
+  virtual ssize_t Priority(StateId s) { return internal::NumArcs(GetFst(), s); }
+};
+
+// A matcher that expects sorted labels on the side to be matched.
+// If match_type == MATCH_INPUT, epsilons match the implicit self-loop
+// Arc(kNoLabel, 0, Weight::One(), current_state) as well as any
+// actual epsilon transitions. If match_type == MATCH_OUTPUT, then
+// Arc(0, kNoLabel, Weight::One(), current_state) is instead matched.
+template <class F>
+class SortedMatcher : public MatcherBase<typename F::Arc> {
+ public:
+  using FST = F;
+  using Arc = typename FST::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using MatcherBase<Arc>::Flags;
+  using MatcherBase<Arc>::Properties;
+
+  // Labels >= binary_label will be searched for by binary search;
+  // o.w. linear search is used.
+  // This makes a copy of the FST.
+  SortedMatcher(const FST &fst, MatchType match_type, Label binary_label = 1)
+      : SortedMatcher(fst.Copy(), match_type, binary_label) {
+    owned_fst_.reset(&fst_);
+  }
+
+  // Labels >= binary_label will be searched for by binary search;
+  // o.w. linear search is used.
+  // This doesn't copy the FST.
+  SortedMatcher(const FST *fst, MatchType match_type, Label binary_label = 1)
+      : fst_(*fst),
+        state_(kNoStateId),
+        aiter_(nullptr),
+        match_type_(match_type),
+        binary_label_(binary_label),
+        match_label_(kNoLabel),
+        narcs_(0),
+        loop_(kNoLabel, 0, Weight::One(), kNoStateId),
+        error_(false),
+        aiter_pool_(1) {
+    switch (match_type_) {
+      case MATCH_INPUT:
+      case MATCH_NONE:
+        break;
+      case MATCH_OUTPUT:
+        std::swap(loop_.ilabel, loop_.olabel);
+        break;
+      default:
+        FSTERROR() << "SortedMatcher: Bad match type";
+        match_type_ = MATCH_NONE;
+        error_ = true;
+    }
+  }
+
+  // This makes a copy of the FST.
+  SortedMatcher(const SortedMatcher<FST> &matcher, bool safe = false)
+      : owned_fst_(matcher.fst_.Copy(safe)),
+        fst_(*owned_fst_),
+        state_(kNoStateId),
+        aiter_(nullptr),
+        match_type_(matcher.match_type_),
+        binary_label_(matcher.binary_label_),
+        match_label_(kNoLabel),
+        narcs_(0),
+        loop_(matcher.loop_),
+        error_(matcher.error_),
+        aiter_pool_(1) {}
+
+  ~SortedMatcher() override { Destroy(aiter_, &aiter_pool_); }
+
+  SortedMatcher<FST> *Copy(bool safe = false) const override {
+    return new SortedMatcher<FST>(*this, safe);
+  }
+
+  MatchType Type(bool test) const override {
+    if (match_type_ == MATCH_NONE) return match_type_;
+    const auto true_prop =
+        match_type_ == MATCH_INPUT ? kILabelSorted : kOLabelSorted;
+    const auto false_prop =
+        match_type_ == MATCH_INPUT ? kNotILabelSorted : kNotOLabelSorted;
+    const auto props = fst_.Properties(true_prop | false_prop, test);
+    if (props & true_prop) {
+      return match_type_;
+    } else if (props & false_prop) {
+      return MATCH_NONE;
+    } else {
+      return MATCH_UNKNOWN;
+    }
+  }
+
+  void SetState(StateId s) final {
+    if (state_ == s) return;
+    state_ = s;
+    if (match_type_ == MATCH_NONE) {
+      FSTERROR() << "SortedMatcher: Bad match type";
+      error_ = true;
+    }
+    Destroy(aiter_, &aiter_pool_);
+    aiter_ = new (&aiter_pool_) ArcIterator<FST>(fst_, s);
+    aiter_->SetFlags(kArcNoCache, kArcNoCache);
+    narcs_ = internal::NumArcs(fst_, s);
+    loop_.nextstate = s;
+  }
+
+  bool Find(Label match_label) final {
+    exact_match_ = true;
+    if (error_) {
+      current_loop_ = false;
+      match_label_ = kNoLabel;
+      return false;
+    }
+    current_loop_ = match_label == 0;
+    match_label_ = match_label == kNoLabel ? 0 : match_label;
+    if (Search()) {
+      return true;
+    } else {
+      return current_loop_;
+    }
+  }
+
+  // Positions matcher to the first position where inserting match_label would
+  // maintain the sort order.
+  void LowerBound(Label label) {
+    exact_match_ = false;
+    current_loop_ = false;
+    if (error_) {
+      match_label_ = kNoLabel;
+      return;
+    }
+    match_label_ = label;
+    Search();
+  }
+
+  // After Find(), returns false if no more exact matches.
+  // After LowerBound(), returns false if no more arcs.
+  bool Done() const final {
+    if (current_loop_) return false;
+    if (aiter_->Done()) return true;
+    if (!exact_match_) return false;
+    aiter_->SetFlags(match_type_ == MATCH_INPUT ?
+        kArcILabelValue : kArcOLabelValue,
+        kArcValueFlags);
+    return GetLabel() != match_label_;
+  }
+
+  const Arc &Value() const final {
+    if (current_loop_) return loop_;
+    aiter_->SetFlags(kArcValueFlags, kArcValueFlags);
+    return aiter_->Value();
+  }
+
+  void Next() final {
+    if (current_loop_) {
+      current_loop_ = false;
+    } else {
+      aiter_->Next();
+    }
+  }
+
+  Weight Final(StateId s) const final {
+    return MatcherBase<Arc>::Final(s);
+  }
+
+  ssize_t Priority(StateId s) final {
+    return MatcherBase<Arc>::Priority(s);
+  }
+
+  const FST &GetFst() const override { return fst_; }
+
+  uint64 Properties(uint64 inprops) const override {
+    return inprops | (error_ ? kError : 0);
+  }
+
+  size_t Position() const { return aiter_ ? aiter_->Position() : 0; }
+
+ private:
+  Label GetLabel() const {
+    const auto &arc = aiter_->Value();
+    return match_type_ == MATCH_INPUT ? arc.ilabel : arc.olabel;
+  }
+
+  bool BinarySearch();
+  bool LinearSearch();
+  bool Search();
+
+  std::unique_ptr<const FST> owned_fst_;   // FST ptr if owned.
+  const FST &fst_;           // FST for matching.
+  StateId state_;            // Matcher state.
+  ArcIterator<FST> *aiter_;  // Iterator for current state.
+  MatchType match_type_;     // Type of match to perform.
+  Label binary_label_;       // Least label for binary search.
+  Label match_label_;        // Current label to be matched.
+  size_t narcs_;             // Current state arc count.
+  Arc loop_;                 // For non-consuming symbols.
+  bool current_loop_;        // Current arc is the implicit loop.
+  bool exact_match_;         // Exact match or lower bound?
+  bool error_;               // Error encountered?
+  MemoryPool<ArcIterator<FST>> aiter_pool_;  // Pool of arc iterators.
+};
+
+// Returns true iff match to match_label_. The arc iterator is positioned at the
+// lower bound, that is, the first element greater than or equal to
+// match_label_, or the end if all elements are less than match_label_.
+// If multiple elements are equal to the `match_label_`, returns the rightmost
+// one.
+template <class FST>
+inline bool SortedMatcher<FST>::BinarySearch() {
+  size_t size = narcs_;
+  if (size == 0) {
+    return false;
+  }
+  size_t high = size - 1;
+  while (size > 1) {
+    const size_t half = size / 2;
+    const size_t mid = high - half;
+    aiter_->Seek(mid);
+    if (GetLabel() >= match_label_) {
+      high = mid;
+    }
+    size -= half;
+  }
+  aiter_->Seek(high);
+  const auto label = GetLabel();
+  if (label == match_label_) {
+    return true;
+  }
+  if (label < match_label_) {
+    aiter_->Next();
+  }
+  return false;
+}
+
+// Returns true iff match to match_label_, positioning arc iterator at lower
+// bound.
+template <class FST>
+inline bool SortedMatcher<FST>::LinearSearch() {
+  for (aiter_->Reset(); !aiter_->Done(); aiter_->Next()) {
+    const auto label = GetLabel();
+    if (label == match_label_) return true;
+    if (label > match_label_) break;
+  }
+  return false;
+}
+
+// Returns true iff match to match_label_, positioning arc iterator at lower
+// bound.
+template <class FST>
+inline bool SortedMatcher<FST>::Search() {
+  aiter_->SetFlags(match_type_ == MATCH_INPUT ?
+                   kArcILabelValue : kArcOLabelValue,
+                   kArcValueFlags);
+  if (match_label_ >= binary_label_) {
+    return BinarySearch();
+  } else {
+    return LinearSearch();
+  }
+}
+
+// A matcher that stores labels in a per-state hash table populated upon the
+// first visit to that state. Sorting is not required. Treatment of
+// epsilons are the same as with SortedMatcher.
+template <class F>
+class HashMatcher : public MatcherBase<typename F::Arc> {
+ public:
+  using FST = F;
+  using Arc = typename FST::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using MatcherBase<Arc>::Flags;
+  using MatcherBase<Arc>::Final;
+  using MatcherBase<Arc>::Priority;
+
+  // This makes a copy of the FST.
+  HashMatcher(const FST &fst, MatchType match_type)
+      : HashMatcher(fst.Copy(), match_type) {
+    owned_fst_.reset(&fst_);
+  }
+
+  // This doesn't copy the FST.
+  HashMatcher(const FST *fst, MatchType match_type)
+      : fst_(*fst),
+        state_(kNoStateId),
+        match_type_(match_type),
+        loop_(kNoLabel, 0, Weight::One(), kNoStateId),
+        error_(false),
+        state_table_(std::make_shared<StateTable>()) {
+    switch (match_type_) {
+      case MATCH_INPUT:
+      case MATCH_NONE:
+        break;
+      case MATCH_OUTPUT:
+        std::swap(loop_.ilabel, loop_.olabel);
+        break;
+      default:
+        FSTERROR() << "HashMatcher: Bad match type";
+        match_type_ = MATCH_NONE;
+        error_ = true;
+    }
+  }
+
+  // This makes a copy of the FST.
+  HashMatcher(const HashMatcher<FST> &matcher, bool safe = false)
+      : owned_fst_(matcher.fst_.Copy(safe)),
+        fst_(*owned_fst_),
+        state_(kNoStateId),
+        match_type_(matcher.match_type_),
+        loop_(matcher.loop_),
+        error_(matcher.error_),
+        state_table_(
+            safe ? std::make_shared<StateTable>() : matcher.state_table_) {}
+
+  HashMatcher<FST> *Copy(bool safe = false) const override {
+    return new HashMatcher<FST>(*this, safe);
+  }
+
+  // The argument is ignored as there are no relevant properties to test.
+  MatchType Type(bool test) const override { return match_type_; }
+
+  void SetState(StateId s) final;
+
+  bool Find(Label label) final {
+    current_loop_ = label == 0;
+    if (label == 0) {
+      Search(label);
+      return true;
+    }
+    if (label == kNoLabel) label = 0;
+    return Search(label);
+  }
+
+  bool Done() const final {
+    if (current_loop_) return false;
+    return label_it_ == label_end_;
+  }
+
+  const Arc &Value() const final {
+    if (current_loop_) return loop_;
+    aiter_->Seek(label_it_->second);
+    return aiter_->Value();
+  }
+
+  void Next() final {
+    if (current_loop_) {
+      current_loop_ = false;
+    } else {
+      ++label_it_;
+    }
+  }
+
+  const FST &GetFst() const override { return fst_; }
+
+  uint64 Properties(uint64 inprops) const override {
+    return inprops | (error_ ? kError : 0);
+  }
+
+ private:
+  Label GetLabel() const {
+    const auto &arc = aiter_->Value();
+    return match_type_ == MATCH_INPUT ? arc.ilabel : arc.olabel;
+  }
+
+  bool Search(Label match_label);
+
+  using LabelTable = std::unordered_multimap<Label, size_t>;
+  using StateTable = std::unordered_map<StateId, std::unique_ptr<LabelTable>>;
+
+  std::unique_ptr<const FST> owned_fst_;  // ptr to FST if owned.
+  const FST &fst_;     // FST for matching.
+  StateId state_;      // Matcher state.
+  MatchType match_type_;
+  Arc loop_;            // The implicit loop itself.
+  bool current_loop_;   // Is the current arc the implicit loop?
+  bool error_;          // Error encountered?
+  std::unique_ptr<ArcIterator<FST>> aiter_;
+  std::shared_ptr<StateTable> state_table_;  // Table from state to label table.
+  LabelTable *label_table_;  // Pointer to current state's label table.
+  typename LabelTable::iterator label_it_;   // Position for label.
+  typename LabelTable::iterator label_end_;  // Position for last label + 1.
+};
+
+template <class FST>
+void HashMatcher<FST>::SetState(typename FST::Arc::StateId s) {
+  if (state_ == s) return;
+  // Resets everything for the state.
+  state_ = s;
+  loop_.nextstate = state_;
+  aiter_.reset(new ArcIterator<FST>(fst_, state_));
+  if (match_type_ == MATCH_NONE) {
+    FSTERROR() << "HashMatcher: Bad match type";
+    error_ = true;
+  }
+  // Attempts to insert a new label table.
+  auto it_and_success = state_table_->emplace(
+    state_, std::unique_ptr<LabelTable>(new LabelTable()));
+  // Sets instance's pointer to the label table for this state.
+  label_table_ = it_and_success.first->second.get();
+  // If it already exists, no additional work is done and we simply return.
+  if (!it_and_success.second) return;
+  // Otherwise, populate this new table.
+  // Populates the label table.
+  label_table_->reserve(internal::NumArcs(fst_, state_));
+  const auto aiter_flags =
+      (match_type_ == MATCH_INPUT ? kArcILabelValue : kArcOLabelValue) |
+      kArcNoCache;
+  aiter_->SetFlags(aiter_flags, kArcFlags);
+  for (; !aiter_->Done(); aiter_->Next()) {
+    label_table_->emplace(GetLabel(), aiter_->Position());
+  }
+  aiter_->SetFlags(kArcValueFlags, kArcValueFlags);
+}
+
+template <class FST>
+inline bool HashMatcher<FST>::Search(typename FST::Arc::Label match_label) {
+  auto range = label_table_->equal_range(match_label);
+  label_it_ = range.first;
+  label_end_ = range.second;
+  if (label_it_ == label_end_) return false;
+  aiter_->Seek(label_it_->second);
+  return true;
+}
+
+// Specifies whether we rewrite both the input and output sides during matching.
+enum MatcherRewriteMode {
+  MATCHER_REWRITE_AUTO = 0,  // Rewrites both sides iff acceptor.
+  MATCHER_REWRITE_ALWAYS,
+  MATCHER_REWRITE_NEVER
+};
+
+// For any requested label that doesn't match at a state, this matcher
+// considers the *unique* transition that matches the label 'phi_label'
+// (phi = 'fail'), and recursively looks for a match at its
+// destination.  When 'phi_loop' is true, if no match is found but a
+// phi self-loop is found, then the phi transition found is returned
+// with the phi_label rewritten as the requested label (both sides if
+// an acceptor, or if 'rewrite_both' is true and both input and output
+// labels of the found transition are 'phi_label').  If 'phi_label' is
+// kNoLabel, this special matching is not done.  PhiMatcher is
+// templated itself on a matcher, which is used to perform the
+// underlying matching.  By default, the underlying matcher is
+// constructed by PhiMatcher. The user can instead pass in this
+// object; in that case, PhiMatcher takes its ownership.
+// Phi non-determinism not supported. No non-consuming symbols other
+// than epsilon supported with the underlying template argument matcher.
+template <class M>
+class PhiMatcher : public MatcherBase<typename M::Arc> {
+ public:
+  using FST = typename M::FST;
+  using Arc = typename FST::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  // This makes a copy of the FST (w/o 'matcher' arg).
+  PhiMatcher(const FST &fst, MatchType match_type, Label phi_label = kNoLabel,
+             bool phi_loop = true,
+             MatcherRewriteMode rewrite_mode = MATCHER_REWRITE_AUTO,
+             M *matcher = nullptr)
+      : matcher_(matcher ? matcher : new M(fst, match_type)),
+        match_type_(match_type),
+        phi_label_(phi_label),
+        state_(kNoStateId),
+        phi_loop_(phi_loop),
+        error_(false) {
+    if (match_type == MATCH_BOTH) {
+      FSTERROR() << "PhiMatcher: Bad match type";
+      match_type_ = MATCH_NONE;
+      error_ = true;
+    }
+    if (rewrite_mode == MATCHER_REWRITE_AUTO) {
+      rewrite_both_ = fst.Properties(kAcceptor, true);
+    } else if (rewrite_mode == MATCHER_REWRITE_ALWAYS) {
+      rewrite_both_ = true;
+    } else {
+      rewrite_both_ = false;
+    }
+  }
+
+  // This doesn't copy the FST.
+  PhiMatcher(const FST *fst, MatchType match_type, Label phi_label = kNoLabel,
+             bool phi_loop = true,
+             MatcherRewriteMode rewrite_mode = MATCHER_REWRITE_AUTO,
+             M *matcher = nullptr)
+      : PhiMatcher(*fst, match_type, phi_label, phi_loop, rewrite_mode,
+                   matcher ? matcher : new M(fst, match_type)) { }
+
+
+  // This makes a copy of the FST.
+  PhiMatcher(const PhiMatcher<M> &matcher, bool safe = false)
+      : matcher_(new M(*matcher.matcher_, safe)),
+        match_type_(matcher.match_type_),
+        phi_label_(matcher.phi_label_),
+        rewrite_both_(matcher.rewrite_both_),
+        state_(kNoStateId),
+        phi_loop_(matcher.phi_loop_),
+        error_(matcher.error_) {}
+
+  PhiMatcher<M> *Copy(bool safe = false) const override {
+    return new PhiMatcher<M>(*this, safe);
+  }
+
+  MatchType Type(bool test) const override { return matcher_->Type(test); }
+
+  void SetState(StateId s) final {
+    if (state_ == s) return;
+    matcher_->SetState(s);
+    state_ = s;
+    has_phi_ = phi_label_ != kNoLabel;
+  }
+
+  bool Find(Label match_label) final;
+
+  bool Done() const final { return matcher_->Done(); }
+
+  const Arc &Value() const final {
+    if ((phi_match_ == kNoLabel) && (phi_weight_ == Weight::One())) {
+      return matcher_->Value();
+    } else if (phi_match_ == 0) {  // Virtual epsilon loop.
+      phi_arc_ = Arc(kNoLabel, 0, Weight::One(), state_);
+      if (match_type_ == MATCH_OUTPUT) {
+        std::swap(phi_arc_.ilabel, phi_arc_.olabel);
+      }
+      return phi_arc_;
+    } else {
+      phi_arc_ = matcher_->Value();
+      phi_arc_.weight = Times(phi_weight_, phi_arc_.weight);
+      if (phi_match_ != kNoLabel) {  // Phi loop match.
+        if (rewrite_both_) {
+          if (phi_arc_.ilabel == phi_label_) phi_arc_.ilabel = phi_match_;
+          if (phi_arc_.olabel == phi_label_) phi_arc_.olabel = phi_match_;
+        } else if (match_type_ == MATCH_INPUT) {
+          phi_arc_.ilabel = phi_match_;
+        } else {
+          phi_arc_.olabel = phi_match_;
+        }
+      }
+      return phi_arc_;
+    }
+  }
+
+  void Next() final { matcher_->Next(); }
+
+  Weight Final(StateId s) const final {
+    auto weight = matcher_->Final(s);
+    if (phi_label_ == kNoLabel || weight != Weight::Zero()) {
+      return weight;
+    }
+    weight = Weight::One();
+    matcher_->SetState(s);
+    while (matcher_->Final(s) == Weight::Zero()) {
+      if (!matcher_->Find(phi_label_ == 0 ? -1 : phi_label_)) break;
+      weight = Times(weight, matcher_->Value().weight);
+      if (s == matcher_->Value().nextstate) {
+        return Weight::Zero();  // Does not follow phi self-loops.
+      }
+      s = matcher_->Value().nextstate;
+      matcher_->SetState(s);
+    }
+    weight = Times(weight, matcher_->Final(s));
+    return weight;
+  }
+
+  ssize_t Priority(StateId s) final {
+    if (phi_label_ != kNoLabel) {
+      matcher_->SetState(s);
+      const bool has_phi = matcher_->Find(phi_label_ == 0 ? -1 : phi_label_);
+      return has_phi ? kRequirePriority : matcher_->Priority(s);
+    } else {
+      return matcher_->Priority(s);
+    }
+  }
+
+  const FST &GetFst() const override { return matcher_->GetFst(); }
+
+  uint64 Properties(uint64 props) const override;
+
+  uint32 Flags() const override {
+    if (phi_label_ == kNoLabel || match_type_ == MATCH_NONE) {
+      return matcher_->Flags();
+    }
+    return matcher_->Flags() | kRequireMatch;
+  }
+
+  Label PhiLabel() const { return phi_label_; }
+
+ private:
+  mutable std::unique_ptr<M> matcher_;
+  MatchType match_type_;  // Type of match requested.
+  Label phi_label_;       // Label that represents the phi transition.
+  bool rewrite_both_;     // Rewrite both sides when both are phi_label_?
+  bool has_phi_;          // Are there possibly phis at the current state?
+  Label phi_match_;       // Current label that matches phi loop.
+  mutable Arc phi_arc_;   // Arc to return.
+  StateId state_;         // Matcher state.
+  Weight phi_weight_;     // Product of the weights of phi transitions taken.
+  bool phi_loop_;         // When true, phi self-loop are allowed and treated
+                          // as rho (required for Aho-Corasick).
+  bool error_;            // Error encountered?
+
+  PhiMatcher &operator=(const PhiMatcher &) = delete;
+};
+
+template <class M>
+inline bool PhiMatcher<M>::Find(Label label) {
+  if (label == phi_label_ && phi_label_ != kNoLabel && phi_label_ != 0) {
+    FSTERROR() << "PhiMatcher::Find: bad label (phi): " << phi_label_;
+    error_ = true;
+    return false;
+  }
+  matcher_->SetState(state_);
+  phi_match_ = kNoLabel;
+  phi_weight_ = Weight::One();
+  // If phi_label_ == 0, there are no more true epsilon arcs.
+  if (phi_label_ == 0) {
+    if (label == kNoLabel) {
+      return false;
+    }
+    if (label == 0) {  // but a virtual epsilon loop needs to be returned.
+      if (!matcher_->Find(kNoLabel)) {
+        return matcher_->Find(0);
+      } else {
+        phi_match_ = 0;
+        return true;
+      }
+    }
+  }
+  if (!has_phi_ || label == 0 || label == kNoLabel) {
+    return matcher_->Find(label);
+  }
+  auto s = state_;
+  while (!matcher_->Find(label)) {
+    // Look for phi transition (if phi_label_ == 0, we need to look
+    // for -1 to avoid getting the virtual self-loop)
+    if (!matcher_->Find(phi_label_ == 0 ? -1 : phi_label_)) return false;
+    if (phi_loop_ && matcher_->Value().nextstate == s) {
+      phi_match_ = label;
+      return true;
+    }
+    phi_weight_ = Times(phi_weight_, matcher_->Value().weight);
+    s = matcher_->Value().nextstate;
+    matcher_->Next();
+    if (!matcher_->Done()) {
+      FSTERROR() << "PhiMatcher: Phi non-determinism not supported";
+      error_ = true;
+    }
+    matcher_->SetState(s);
+  }
+  return true;
+}
+
+template <class M>
+inline uint64 PhiMatcher<M>::Properties(uint64 inprops) const {
+  auto outprops = matcher_->Properties(inprops);
+  if (error_) outprops |= kError;
+  if (match_type_ == MATCH_NONE) {
+    return outprops;
+  } else if (match_type_ == MATCH_INPUT) {
+    if (phi_label_ == 0) {
+      outprops &= ~kEpsilons | ~kIEpsilons | ~kOEpsilons;
+      outprops |= kNoEpsilons | kNoIEpsilons;
+    }
+    if (rewrite_both_) {
+      return outprops &
+             ~(kODeterministic | kNonODeterministic | kString | kILabelSorted |
+               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
+    } else {
+      return outprops &
+             ~(kODeterministic | kAcceptor | kString | kILabelSorted |
+               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
+    }
+  } else if (match_type_ == MATCH_OUTPUT) {
+    if (phi_label_ == 0) {
+      outprops &= ~kEpsilons | ~kIEpsilons | ~kOEpsilons;
+      outprops |= kNoEpsilons | kNoOEpsilons;
+    }
+    if (rewrite_both_) {
+      return outprops &
+             ~(kIDeterministic | kNonIDeterministic | kString | kILabelSorted |
+               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
+    } else {
+      return outprops &
+             ~(kIDeterministic | kAcceptor | kString | kILabelSorted |
+               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
+    }
+  } else {
+    // Shouldn't ever get here.
+    FSTERROR() << "PhiMatcher: Bad match type: " << match_type_;
+    return 0;
+  }
+}
+
+// For any requested label that doesn't match at a state, this matcher
+// considers all transitions that match the label 'rho_label' (rho =
+// 'rest').  Each such rho transition found is returned with the
+// rho_label rewritten as the requested label (both sides if an
+// acceptor, or if 'rewrite_both' is true and both input and output
+// labels of the found transition are 'rho_label').  If 'rho_label' is
+// kNoLabel, this special matching is not done.  RhoMatcher is
+// templated itself on a matcher, which is used to perform the
+// underlying matching.  By default, the underlying matcher is
+// constructed by RhoMatcher.  The user can instead pass in this
+// object; in that case, RhoMatcher takes its ownership.
+// No non-consuming symbols other than epsilon supported with
+// the underlying template argument matcher.
+template <class M>
+class RhoMatcher : public MatcherBase<typename M::Arc> {
+ public:
+  using FST = typename M::FST;
+  using Arc = typename FST::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  // This makes a copy of the FST (w/o 'matcher' arg).
+  RhoMatcher(const FST &fst, MatchType match_type, Label rho_label = kNoLabel,
+             MatcherRewriteMode rewrite_mode = MATCHER_REWRITE_AUTO,
+             M *matcher = nullptr)
+      : matcher_(matcher ? matcher : new M(fst, match_type)),
+        match_type_(match_type),
+        rho_label_(rho_label),
+        error_(false),
+        state_(kNoStateId),
+        has_rho_(false) {
+    if (match_type == MATCH_BOTH) {
+      FSTERROR() << "RhoMatcher: Bad match type";
+      match_type_ = MATCH_NONE;
+      error_ = true;
+    }
+    if (rho_label == 0) {
+      FSTERROR() << "RhoMatcher: 0 cannot be used as rho_label";
+      rho_label_ = kNoLabel;
+      error_ = true;
+    }
+    if (rewrite_mode == MATCHER_REWRITE_AUTO) {
+      rewrite_both_ = fst.Properties(kAcceptor, true);
+    } else if (rewrite_mode == MATCHER_REWRITE_ALWAYS) {
+      rewrite_both_ = true;
+    } else {
+      rewrite_both_ = false;
+    }
+  }
+
+  // This doesn't copy the FST.
+  RhoMatcher(const FST *fst, MatchType match_type, Label rho_label = kNoLabel,
+             MatcherRewriteMode rewrite_mode = MATCHER_REWRITE_AUTO,
+             M *matcher = nullptr)
+      : RhoMatcher(*fst, match_type, rho_label, rewrite_mode,
+                   matcher ? matcher : new M(fst, match_type)) { }
+
+  // This makes a copy of the FST.
+  RhoMatcher(const RhoMatcher<M> &matcher, bool safe = false)
+      : matcher_(new M(*matcher.matcher_, safe)),
+        match_type_(matcher.match_type_),
+        rho_label_(matcher.rho_label_),
+        rewrite_both_(matcher.rewrite_both_),
+        error_(matcher.error_),
+        state_(kNoStateId),
+        has_rho_(false) {}
+
+  RhoMatcher<M> *Copy(bool safe = false) const override {
+    return new RhoMatcher<M>(*this, safe);
+  }
+
+  MatchType Type(bool test) const override { return matcher_->Type(test); }
+
+  void SetState(StateId s) final {
+    if (state_ == s) return;
+    state_ = s;
+    matcher_->SetState(s);
+    has_rho_ = rho_label_ != kNoLabel;
+  }
+
+  bool Find(Label label) final {
+    if (label == rho_label_ && rho_label_ != kNoLabel) {
+      FSTERROR() << "RhoMatcher::Find: bad label (rho)";
+      error_ = true;
+      return false;
+    }
+    if (matcher_->Find(label)) {
+      rho_match_ = kNoLabel;
+      return true;
+    } else if (has_rho_ && label != 0 && label != kNoLabel &&
+               (has_rho_ = matcher_->Find(rho_label_))) {
+      rho_match_ = label;
+      return true;
+    } else {
+      return false;
+    }
+  }
+
+  bool Done() const final { return matcher_->Done(); }
+
+  const Arc &Value() const final {
+    if (rho_match_ == kNoLabel) {
+      return matcher_->Value();
+    } else {
+      rho_arc_ = matcher_->Value();
+      if (rewrite_both_) {
+        if (rho_arc_.ilabel == rho_label_) rho_arc_.ilabel = rho_match_;
+        if (rho_arc_.olabel == rho_label_) rho_arc_.olabel = rho_match_;
+      } else if (match_type_ == MATCH_INPUT) {
+        rho_arc_.ilabel = rho_match_;
+      } else {
+        rho_arc_.olabel = rho_match_;
+      }
+      return rho_arc_;
+    }
+  }
+
+  void Next() final { matcher_->Next(); }
+
+  Weight Final(StateId s) const final { return matcher_->Final(s); }
+
+  ssize_t Priority(StateId s) final {
+    state_ = s;
+    matcher_->SetState(s);
+    has_rho_ = matcher_->Find(rho_label_);
+    if (has_rho_) {
+      return kRequirePriority;
+    } else {
+      return matcher_->Priority(s);
+    }
+  }
+
+  const FST &GetFst() const override { return matcher_->GetFst(); }
+
+  uint64 Properties(uint64 props) const override;
+
+  uint32 Flags() const override {
+    if (rho_label_ == kNoLabel || match_type_ == MATCH_NONE) {
+      return matcher_->Flags();
+    }
+    return matcher_->Flags() | kRequireMatch;
+  }
+
+  Label RhoLabel() const { return rho_label_; }
+
+ private:
+  std::unique_ptr<M> matcher_;
+  MatchType match_type_;  // Type of match requested.
+  Label rho_label_;       // Label that represents the rho transition
+  bool rewrite_both_;     // Rewrite both sides when both are rho_label_?
+  Label rho_match_;       // Current label that matches rho transition.
+  mutable Arc rho_arc_;   // Arc to return when rho match.
+  bool error_;            // Error encountered?
+  StateId state_;         // Matcher state.
+  bool has_rho_;          // Are there possibly rhos at the current state?
+};
+
+template <class M>
+inline uint64 RhoMatcher<M>::Properties(uint64 inprops) const {
+  auto outprops = matcher_->Properties(inprops);
+  if (error_) outprops |= kError;
+  if (match_type_ == MATCH_NONE) {
+    return outprops;
+  } else if (match_type_ == MATCH_INPUT) {
+    if (rewrite_both_) {
+      return outprops &
+             ~(kODeterministic | kNonODeterministic | kString | kILabelSorted |
+               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
+    } else {
+      return outprops &
+             ~(kODeterministic | kAcceptor | kString | kILabelSorted |
+               kNotILabelSorted);
+    }
+  } else if (match_type_ == MATCH_OUTPUT) {
+    if (rewrite_both_) {
+      return outprops &
+             ~(kIDeterministic | kNonIDeterministic | kString | kILabelSorted |
+               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
+    } else {
+      return outprops &
+             ~(kIDeterministic | kAcceptor | kString | kOLabelSorted |
+               kNotOLabelSorted);
+    }
+  } else {
+    // Shouldn't ever get here.
+    FSTERROR() << "RhoMatcher: Bad match type: " << match_type_;
+    return 0;
+  }
+}
+
+// For any requested label, this matcher considers all transitions
+// that match the label 'sigma_label' (sigma = "any"), and this in
+// additions to transitions with the requested label.  Each such sigma
+// transition found is returned with the sigma_label rewritten as the
+// requested label (both sides if an acceptor, or if 'rewrite_both' is
+// true and both input and output labels of the found transition are
+// 'sigma_label').  If 'sigma_label' is kNoLabel, this special
+// matching is not done.  SigmaMatcher is templated itself on a
+// matcher, which is used to perform the underlying matching.  By
+// default, the underlying matcher is constructed by SigmaMatcher.
+// The user can instead pass in this object; in that case,
+// SigmaMatcher takes its ownership.  No non-consuming symbols other
+// than epsilon supported with the underlying template argument matcher.
+template <class M>
+class SigmaMatcher : public MatcherBase<typename M::Arc> {
+ public:
+  using FST = typename M::FST;
+  using Arc = typename FST::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  // This makes a copy of the FST (w/o 'matcher' arg).
+  SigmaMatcher(const FST &fst, MatchType match_type,
+               Label sigma_label = kNoLabel,
+               MatcherRewriteMode rewrite_mode = MATCHER_REWRITE_AUTO,
+               M *matcher = nullptr)
+      : matcher_(matcher ? matcher : new M(fst, match_type)),
+        match_type_(match_type),
+        sigma_label_(sigma_label),
+        error_(false),
+        state_(kNoStateId) {
+    if (match_type == MATCH_BOTH) {
+      FSTERROR() << "SigmaMatcher: Bad match type";
+      match_type_ = MATCH_NONE;
+      error_ = true;
+    }
+    if (sigma_label == 0) {
+      FSTERROR() << "SigmaMatcher: 0 cannot be used as sigma_label";
+      sigma_label_ = kNoLabel;
+      error_ = true;
+    }
+    if (rewrite_mode == MATCHER_REWRITE_AUTO) {
+      rewrite_both_ = fst.Properties(kAcceptor, true);
+    } else if (rewrite_mode == MATCHER_REWRITE_ALWAYS) {
+      rewrite_both_ = true;
+    } else {
+      rewrite_both_ = false;
+    }
+  }
+
+  // This doesn't copy the FST.
+  SigmaMatcher(const FST *fst, MatchType match_type,
+               Label sigma_label = kNoLabel,
+               MatcherRewriteMode rewrite_mode = MATCHER_REWRITE_AUTO,
+             M *matcher = nullptr)
+      : SigmaMatcher(*fst, match_type, sigma_label, rewrite_mode,
+                     matcher ? matcher : new M(fst, match_type)) { }
+
+  // This makes a copy of the FST.
+  SigmaMatcher(const SigmaMatcher<M> &matcher, bool safe = false)
+      : matcher_(new M(*matcher.matcher_, safe)),
+        match_type_(matcher.match_type_),
+        sigma_label_(matcher.sigma_label_),
+        rewrite_both_(matcher.rewrite_both_),
+        error_(matcher.error_),
+        state_(kNoStateId) {}
+
+  SigmaMatcher<M> *Copy(bool safe = false) const override {
+    return new SigmaMatcher<M>(*this, safe);
+  }
+
+  MatchType Type(bool test) const override { return matcher_->Type(test); }
+
+  void SetState(StateId s) final {
+    if (state_ == s) return;
+    state_ = s;
+    matcher_->SetState(s);
+    has_sigma_ =
+        (sigma_label_ != kNoLabel) ? matcher_->Find(sigma_label_) : false;
+  }
+
+  bool Find(Label match_label) final {
+    match_label_ = match_label;
+    if (match_label == sigma_label_ && sigma_label_ != kNoLabel) {
+      FSTERROR() << "SigmaMatcher::Find: bad label (sigma)";
+      error_ = true;
+      return false;
+    }
+    if (matcher_->Find(match_label)) {
+      sigma_match_ = kNoLabel;
+      return true;
+    } else if (has_sigma_ && match_label != 0 && match_label != kNoLabel &&
+               matcher_->Find(sigma_label_)) {
+      sigma_match_ = match_label;
+      return true;
+    } else {
+      return false;
+    }
+  }
+
+  bool Done() const final { return matcher_->Done(); }
+
+  const Arc &Value() const final {
+    if (sigma_match_ == kNoLabel) {
+      return matcher_->Value();
+    } else {
+      sigma_arc_ = matcher_->Value();
+      if (rewrite_both_) {
+        if (sigma_arc_.ilabel == sigma_label_) sigma_arc_.ilabel = sigma_match_;
+        if (sigma_arc_.olabel == sigma_label_) sigma_arc_.olabel = sigma_match_;
+      } else if (match_type_ == MATCH_INPUT) {
+        sigma_arc_.ilabel = sigma_match_;
+      } else {
+        sigma_arc_.olabel = sigma_match_;
+      }
+      return sigma_arc_;
+    }
+  }
+
+  void Next() final {
+    matcher_->Next();
+    if (matcher_->Done() && has_sigma_ && (sigma_match_ == kNoLabel) &&
+        (match_label_ > 0)) {
+      matcher_->Find(sigma_label_);
+      sigma_match_ = match_label_;
+    }
+  }
+
+  Weight Final(StateId s) const final { return matcher_->Final(s); }
+
+  ssize_t Priority(StateId s) final {
+    if (sigma_label_ != kNoLabel) {
+      SetState(s);
+      return has_sigma_ ? kRequirePriority : matcher_->Priority(s);
+    } else {
+      return matcher_->Priority(s);
+    }
+  }
+
+  const FST &GetFst() const override { return matcher_->GetFst(); }
+
+  uint64 Properties(uint64 props) const override;
+
+  uint32 Flags() const override {
+    if (sigma_label_ == kNoLabel || match_type_ == MATCH_NONE) {
+      return matcher_->Flags();
+    }
+    return matcher_->Flags() | kRequireMatch;
+  }
+
+  Label SigmaLabel() const { return sigma_label_; }
+
+ private:
+  std::unique_ptr<M> matcher_;
+  MatchType match_type_;   // Type of match requested.
+  Label sigma_label_;      // Label that represents the sigma transition.
+  bool rewrite_both_;      // Rewrite both sides when both are sigma_label_?
+  bool has_sigma_;         // Are there sigmas at the current state?
+  Label sigma_match_;      // Current label that matches sigma transition.
+  mutable Arc sigma_arc_;  // Arc to return when sigma match.
+  Label match_label_;      // Label being matched.
+  bool error_;             // Error encountered?
+  StateId state_;          // Matcher state.
+};
+
+template <class M>
+inline uint64 SigmaMatcher<M>::Properties(uint64 inprops) const {
+  auto outprops = matcher_->Properties(inprops);
+  if (error_) outprops |= kError;
+  if (match_type_ == MATCH_NONE) {
+    return outprops;
+  } else if (rewrite_both_) {
+    return outprops &
+           ~(kIDeterministic | kNonIDeterministic | kODeterministic |
+             kNonODeterministic | kILabelSorted | kNotILabelSorted |
+             kOLabelSorted | kNotOLabelSorted | kString);
+  } else if (match_type_ == MATCH_INPUT) {
+    return outprops &
+           ~(kIDeterministic | kNonIDeterministic | kODeterministic |
+             kNonODeterministic | kILabelSorted | kNotILabelSorted | kString |
+             kAcceptor);
+  } else if (match_type_ == MATCH_OUTPUT) {
+    return outprops &
+           ~(kIDeterministic | kNonIDeterministic | kODeterministic |
+             kNonODeterministic | kOLabelSorted | kNotOLabelSorted | kString |
+             kAcceptor);
+  } else {
+    // Shouldn't ever get here.
+    FSTERROR() << "SigmaMatcher: Bad match type: " << match_type_;
+    return 0;
+  }
+}
+
+// Flags for MultiEpsMatcher.
+
+// Return multi-epsilon arcs for Find(kNoLabel).
+const uint32 kMultiEpsList = 0x00000001;
+
+// Return a kNolabel loop for Find(multi_eps).
+const uint32 kMultiEpsLoop = 0x00000002;
+
+// MultiEpsMatcher: allows treating multiple non-0 labels as
+// non-consuming labels in addition to 0 that is always
+// non-consuming. Precise behavior controlled by 'flags' argument. By
+// default, the underlying matcher is constructed by
+// MultiEpsMatcher. The user can instead pass in this object; in that
+// case, MultiEpsMatcher takes its ownership iff 'own_matcher' is
+// true.
+template <class M>
+class MultiEpsMatcher {
+ public:
+  using FST = typename M::FST;
+  using Arc = typename FST::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  // This makes a copy of the FST (w/o 'matcher' arg).
+  MultiEpsMatcher(const FST &fst, MatchType match_type,
+                  uint32 flags = (kMultiEpsLoop | kMultiEpsList),
+                  M *matcher = nullptr, bool own_matcher = true)
+      : matcher_(matcher ? matcher : new M(fst, match_type)),
+        flags_(flags),
+        own_matcher_(matcher ? own_matcher : true) {
+    Init(match_type);
+  }
+
+  // This doesn't copy the FST.
+  MultiEpsMatcher(const FST *fst, MatchType match_type,
+                  uint32 flags = (kMultiEpsLoop | kMultiEpsList),
+                  M *matcher = nullptr, bool own_matcher = true)
+      : matcher_(matcher ? matcher : new M(fst, match_type)),
+        flags_(flags),
+        own_matcher_(matcher ? own_matcher : true) {
+    Init(match_type);
+  }
+
+  // This makes a copy of the FST.
+  MultiEpsMatcher(const MultiEpsMatcher<M> &matcher, bool safe = false)
+      : matcher_(new M(*matcher.matcher_, safe)),
+        flags_(matcher.flags_),
+        own_matcher_(true),
+        multi_eps_labels_(matcher.multi_eps_labels_),
+        loop_(matcher.loop_) {
+    loop_.nextstate = kNoStateId;
+  }
+
+  ~MultiEpsMatcher() {
+    if (own_matcher_) delete matcher_;
+  }
+
+  MultiEpsMatcher<M> *Copy(bool safe = false) const {
+    return new MultiEpsMatcher<M>(*this, safe);
+  }
+
+  MatchType Type(bool test) const { return matcher_->Type(test); }
+
+  void SetState(StateId state) {
+    matcher_->SetState(state);
+    loop_.nextstate = state;
+  }
+
+  bool Find(Label label);
+
+  bool Done() const { return done_; }
+
+  const Arc &Value() const { return current_loop_ ? loop_ : matcher_->Value(); }
+
+  void Next() {
+    if (!current_loop_) {
+      matcher_->Next();
+      done_ = matcher_->Done();
+      if (done_ && multi_eps_iter_ != multi_eps_labels_.End()) {
+        ++multi_eps_iter_;
+        while ((multi_eps_iter_ != multi_eps_labels_.End()) &&
+               !matcher_->Find(*multi_eps_iter_)) {
+          ++multi_eps_iter_;
+        }
+        if (multi_eps_iter_ != multi_eps_labels_.End()) {
+          done_ = false;
+        } else {
+          done_ = !matcher_->Find(kNoLabel);
+        }
+      }
+    } else {
+      done_ = true;
+    }
+  }
+
+  const FST &GetFst() const { return matcher_->GetFst(); }
+
+  uint64 Properties(uint64 props) const { return matcher_->Properties(props); }
+
+  const M *GetMatcher() const { return matcher_; }
+
+  Weight Final(StateId s) const { return matcher_->Final(s); }
+
+  uint32 Flags() const { return matcher_->Flags(); }
+
+  ssize_t Priority(StateId s) { return matcher_->Priority(s); }
+
+  void AddMultiEpsLabel(Label label) {
+    if (label == 0) {
+      FSTERROR() << "MultiEpsMatcher: Bad multi-eps label: 0";
+    } else {
+      multi_eps_labels_.Insert(label);
+    }
+  }
+
+  void RemoveMultiEpsLabel(Label label) {
+    if (label == 0) {
+      FSTERROR() << "MultiEpsMatcher: Bad multi-eps label: 0";
+    } else {
+      multi_eps_labels_.Erase(label);
+    }
+  }
+
+  void ClearMultiEpsLabels() { multi_eps_labels_.Clear(); }
+
+ private:
+  void Init(MatchType match_type) {
+    if (match_type == MATCH_INPUT) {
+      loop_.ilabel = kNoLabel;
+      loop_.olabel = 0;
+    } else {
+      loop_.ilabel = 0;
+      loop_.olabel = kNoLabel;
+    }
+    loop_.weight = Weight::One();
+    loop_.nextstate = kNoStateId;
+  }
+
+  M *matcher_;
+  uint32 flags_;
+  bool own_matcher_;  // Does this class delete the matcher?
+
+  // Multi-eps label set.
+  CompactSet<Label, kNoLabel> multi_eps_labels_;
+  typename CompactSet<Label, kNoLabel>::const_iterator multi_eps_iter_;
+
+  bool current_loop_;  // Current arc is the implicit loop?
+  mutable Arc loop_;   // For non-consuming symbols.
+  bool done_;          // Matching done?
+
+  MultiEpsMatcher &operator=(const MultiEpsMatcher &) = delete;
+};
+
+template <class M>
+inline bool MultiEpsMatcher<M>::Find(Label label) {
+  multi_eps_iter_ = multi_eps_labels_.End();
+  current_loop_ = false;
+  bool ret;
+  if (label == 0) {
+    ret = matcher_->Find(0);
+  } else if (label == kNoLabel) {
+    if (flags_ & kMultiEpsList) {
+      // Returns all non-consuming arcs (including epsilon).
+      multi_eps_iter_ = multi_eps_labels_.Begin();
+      while ((multi_eps_iter_ != multi_eps_labels_.End()) &&
+             !matcher_->Find(*multi_eps_iter_)) {
+        ++multi_eps_iter_;
+      }
+      if (multi_eps_iter_ != multi_eps_labels_.End()) {
+        ret = true;
+      } else {
+        ret = matcher_->Find(kNoLabel);
+      }
+    } else {
+      // Returns all epsilon arcs.
+      ret = matcher_->Find(kNoLabel);
+    }
+  } else if ((flags_ & kMultiEpsLoop) &&
+             multi_eps_labels_.Find(label) != multi_eps_labels_.End()) {
+    // Returns implicit loop.
+    current_loop_ = true;
+    ret = true;
+  } else {
+    ret = matcher_->Find(label);
+  }
+  done_ = !ret;
+  return ret;
+}
+
+// This class discards any implicit matches (e.g., the implicit epsilon
+// self-loops in the SortedMatcher). Matchers are most often used in
+// composition/intersection where the implicit matches are needed
+// e.g. for epsilon processing. However, if a matcher is simply being
+// used to look-up explicit label matches, this class saves the user
+// from having to check for and discard the unwanted implicit matches
+// themselves.
+template <class M>
+class ExplicitMatcher : public MatcherBase<typename M::Arc> {
+ public:
+  using FST = typename M::FST;
+  using Arc = typename FST::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  // This makes a copy of the FST.
+  ExplicitMatcher(const FST &fst, MatchType match_type, M *matcher = nullptr)
+      : matcher_(matcher ? matcher : new M(fst, match_type)),
+        match_type_(match_type),
+        error_(false) {}
+
+  // This doesn't copy the FST.
+  ExplicitMatcher(const FST *fst, MatchType match_type, M *matcher = nullptr)
+      : matcher_(matcher ? matcher : new M(fst, match_type)),
+        match_type_(match_type),
+        error_(false) {}
+
+  // This makes a copy of the FST.
+  ExplicitMatcher(const ExplicitMatcher<M> &matcher, bool safe = false)
+      : matcher_(new M(*matcher.matcher_, safe)),
+        match_type_(matcher.match_type_),
+        error_(matcher.error_) {}
+
+  ExplicitMatcher<M> *Copy(bool safe = false) const override {
+    return new ExplicitMatcher<M>(*this, safe);
+  }
+
+  MatchType Type(bool test) const override { return matcher_->Type(test); }
+
+  void SetState(StateId s) final { matcher_->SetState(s); }
+
+  bool Find(Label label) final {
+    matcher_->Find(label);
+    CheckArc();
+    return !Done();
+  }
+
+  bool Done() const final { return matcher_->Done(); }
+
+  const Arc &Value() const final { return matcher_->Value(); }
+
+  void Next() final {
+    matcher_->Next();
+    CheckArc();
+  }
+
+  Weight Final(StateId s) const final { return matcher_->Final(s); }
+
+  ssize_t Priority(StateId s) final { return matcher_->Priority(s); }
+
+  const FST &GetFst() const final { return matcher_->GetFst(); }
+
+  uint64 Properties(uint64 inprops) const override {
+    return matcher_->Properties(inprops);
+  }
+
+  const M *GetMatcher() const { return matcher_.get(); }
+
+  uint32 Flags() const override { return matcher_->Flags(); }
+
+ private:
+  // Checks current arc if available and explicit. If not available, stops. If
+  // not explicit, checks next ones.
+  void CheckArc() {
+    for (; !matcher_->Done(); matcher_->Next()) {
+      const auto label = match_type_ == MATCH_INPUT ? matcher_->Value().ilabel
+                                                    : matcher_->Value().olabel;
+      if (label != kNoLabel) return;
+    }
+  }
+
+  std::unique_ptr<M> matcher_;
+  MatchType match_type_;  // Type of match requested.
+  bool error_;            // Error encountered?
+};
+
+// Generic matcher, templated on the FST definition.
+//
+// Here is a typical use:
+//
+//   Matcher<StdFst> matcher(fst, MATCH_INPUT);
+//   matcher.SetState(state);
+//   if (matcher.Find(label))
+//     for (; !matcher.Done(); matcher.Next()) {
+//       auto &arc = matcher.Value();
+//       ...
+//     }
+template <class F>
+class Matcher {
+ public:
+  using FST = F;
+  using Arc = typename F::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  // This makes a copy of the FST.
+  Matcher(const FST &fst, MatchType match_type)
+      : owned_fst_(fst.Copy()),
+        base_(owned_fst_->InitMatcher(match_type)) {
+    if (!base_) base_.reset(new SortedMatcher<FST>(owned_fst_.get(),
+                                                   match_type));
+  }
+
+  // This doesn't copy the FST.
+  Matcher(const FST *fst, MatchType match_type)
+      : base_(fst->InitMatcher(match_type)) {
+    if (!base_) base_.reset(new SortedMatcher<FST>(fst, match_type));
+  }
+
+  // This makes a copy of the FST.
+  Matcher(const Matcher<FST> &matcher, bool safe = false)
+      : base_(matcher.base_->Copy(safe)) { }
+
+  // Takes ownership of the provided matcher.
+  explicit Matcher(MatcherBase<Arc> *base_matcher)
+      : base_(base_matcher) { }
+
+  Matcher<FST> *Copy(bool safe = false) const {
+    return new Matcher<FST>(*this, safe);
+  }
+
+  MatchType Type(bool test) const { return base_->Type(test); }
+
+  void SetState(StateId s) { base_->SetState(s); }
+
+  bool Find(Label label) { return base_->Find(label); }
+
+  bool Done() const { return base_->Done(); }
+
+  const Arc &Value() const { return base_->Value(); }
+
+  void Next() { base_->Next(); }
+
+  const FST &GetFst() const {
+    return static_cast<const FST &>(base_->GetFst());
+  }
+
+  uint64 Properties(uint64 props) const { return base_->Properties(props); }
+
+  Weight Final(StateId s) const { return base_->Final(s); }
+
+  uint32 Flags() const { return base_->Flags() & kMatcherFlags; }
+
+  ssize_t Priority(StateId s) { return base_->Priority(s); }
+
+ private:
+  std::unique_ptr<const FST> owned_fst_;
+  std::unique_ptr<MatcherBase<Arc>> base_;
+};
+
+}  // namespace fst
+
+#endif  // FST_MATCHER_H_
diff --git a/projects/llm_framework/include/fst/memory.h b/projects/llm_framework/include/fst/memory.h
new file mode 100644
index 00000000..c1f0bddc
--- /dev/null
+++ b/projects/llm_framework/include/fst/memory.h
@@ -0,0 +1,443 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// FST memory utilities.
+
+#ifndef FST_MEMORY_H_
+#define FST_MEMORY_H_
+
+#include <list>
+#include <memory>
+#include <utility>
+#include <vector>
+
+#include <fst/types.h>
+#include <fst/log.h>
+#include <fstream>
+
+namespace fst {
+
+// Default block allocation size.
+constexpr int kAllocSize = 64;
+
+// Minimum number of allocations per block.
+constexpr int kAllocFit = 4;
+
+// Base class for MemoryArena that allows (e.g.) MemoryArenaCollection to
+// easily manipulate collections of variously sized arenas.
+class MemoryArenaBase {
+ public:
+  virtual ~MemoryArenaBase() {}
+  virtual size_t Size() const = 0;
+};
+
+namespace internal {
+
+// Allocates 'size' unintialized memory chunks of size object_size from
+// underlying blocks of (at least) size 'block_size * object_size'.
+// All blocks are freed when this class is deleted. Result of allocate() will
+// be aligned to object_size.
+template <size_t object_size>
+class MemoryArenaImpl : public MemoryArenaBase {
+ public:
+  enum { kObjectSize = object_size };
+
+  explicit MemoryArenaImpl(size_t block_size = kAllocSize)
+      : block_size_(block_size * kObjectSize), block_pos_(0) {
+    blocks_.emplace_front(new char[block_size_]);
+  }
+
+  void *Allocate(size_t size) {
+    const auto byte_size = size * kObjectSize;
+    if (byte_size * kAllocFit > block_size_) {
+      // Large block; adds new large block.
+      auto *ptr = new char[byte_size];
+      blocks_.emplace_back(ptr);
+      return ptr;
+    }
+    if (block_pos_ + byte_size > block_size_) {
+      // Doesn't fit; adds new standard block.
+      auto *ptr = new char[block_size_];
+      block_pos_ = 0;
+      blocks_.emplace_front(ptr);
+    }
+    // Fits; uses current block.
+    auto *ptr = blocks_.front().get() + block_pos_;
+    block_pos_ += byte_size;
+    return ptr;
+  }
+
+  size_t Size() const override { return kObjectSize; }
+
+ private:
+  const size_t block_size_;  // Default block size in bytes.
+  size_t block_pos_;   // Current position in block in bytes.
+  std::list<std::unique_ptr<char[]>> blocks_;  // List of allocated blocks.
+};
+
+}  // namespace internal
+
+template <typename T>
+using MemoryArena = internal::MemoryArenaImpl<sizeof(T)>;
+
+// Base class for MemoryPool that allows (e.g.) MemoryPoolCollection to easily
+// manipulate collections of variously sized pools.
+class MemoryPoolBase {
+ public:
+  virtual ~MemoryPoolBase() {}
+  virtual size_t Size() const = 0;
+};
+
+namespace internal {
+
+// Allocates and frees initially uninitialized memory chunks of size
+// object_size. Keeps an internal list of freed chunks that are reused (as is)
+// on the next allocation if available. Chunks are constructed in blocks of size
+// 'pool_size'.
+template <size_t object_size>
+class MemoryPoolImpl : public MemoryPoolBase {
+ public:
+  enum { kObjectSize = object_size };
+
+  struct Link {
+    char buf[kObjectSize];
+    Link *next;
+  };
+
+  explicit MemoryPoolImpl(size_t pool_size)
+      : mem_arena_(pool_size), free_list_(nullptr) {}
+
+  void *Allocate() {
+    if (free_list_ == nullptr) {
+      auto *link = static_cast<Link *>(mem_arena_.Allocate(1));
+      link->next = nullptr;
+      return link;
+    } else {
+      auto *link = free_list_;
+      free_list_ = link->next;
+      return link;
+    }
+  }
+
+  void Free(void *ptr) {
+    if (ptr) {
+      auto *link = static_cast<Link *>(ptr);
+      link->next = free_list_;
+      free_list_ = link;
+    }
+  }
+
+  size_t Size() const override { return kObjectSize; }
+
+ private:
+  MemoryArena<Link> mem_arena_;
+  Link *free_list_;
+
+  MemoryPoolImpl(const MemoryPoolImpl &) = delete;
+  MemoryPoolImpl &operator=(const MemoryPoolImpl &) = delete;
+};
+
+}  // namespace internal
+
+// Allocates and frees initially uninitialized memory chunks of size sizeof(T).
+// All memory is freed when the class is deleted. The result of Allocate() will
+// be suitably memory-aligned. Combined with placement operator new and destroy
+// functions for the T class, this can be used to improve allocation efficiency.
+// See nlp/fst/lib/visit.h (global new) and nlp/fst/lib/dfs-visit.h (class new)
+// for examples.
+template <typename T>
+class MemoryPool : public internal::MemoryPoolImpl<sizeof(T)> {
+ public:
+  // 'pool_size' specifies the size of the initial pool and how it is extended.
+  MemoryPool(size_t pool_size = kAllocSize)
+      : internal::MemoryPoolImpl<sizeof(T)>(pool_size) {}
+};
+
+// Stores a collection of memory arenas.
+class MemoryArenaCollection {
+ public:
+  // 'block_size' specifies the block size of the arenas.
+  explicit MemoryArenaCollection(size_t block_size = kAllocSize)
+      : block_size_(block_size), ref_count_(1) {}
+
+  template <typename T>
+  MemoryArena<T> *Arena() {
+    if (sizeof(T) >= arenas_.size()) arenas_.resize(sizeof(T) + 1);
+    MemoryArenaBase *arena = arenas_[sizeof(T)].get();
+    if (arena == nullptr) {
+      arena = new MemoryArena<T>(block_size_);
+      arenas_[sizeof(T)].reset(arena);
+    }
+    return static_cast<MemoryArena<T> *>(arena);
+  }
+
+  size_t BlockSize() const { return block_size_; }
+
+  size_t RefCount() const { return ref_count_; }
+
+  size_t IncrRefCount() { return ++ref_count_; }
+
+  size_t DecrRefCount() { return --ref_count_; }
+
+ private:
+  size_t block_size_;
+  size_t ref_count_;
+  std::vector<std::unique_ptr<MemoryArenaBase>> arenas_;
+};
+
+// Stores a collection of memory pools
+class MemoryPoolCollection {
+ public:
+  // 'pool_size' specifies the size of initial pool and how it is extended.
+  explicit MemoryPoolCollection(size_t pool_size = kAllocSize)
+      : pool_size_(pool_size), ref_count_(1) {}
+
+  template <typename T>
+  MemoryPool<T> *Pool() {
+    if (sizeof(T) >= pools_.size()) pools_.resize(sizeof(T) + 1);
+    MemoryPoolBase *pool = pools_[sizeof(T)].get();
+    if (pool == nullptr) {
+      pool = new MemoryPool<T>(pool_size_);
+      pools_[sizeof(T)].reset(pool);
+    }
+    return static_cast<MemoryPool<T> *>(pool);
+  }
+
+  size_t PoolSize() const { return pool_size_; }
+
+  size_t RefCount() const { return ref_count_; }
+
+  size_t IncrRefCount() { return ++ref_count_; }
+
+  size_t DecrRefCount() { return --ref_count_; }
+
+ private:
+  size_t pool_size_;
+  size_t ref_count_;
+  std::vector<std::unique_ptr<MemoryPoolBase>> pools_;
+};
+
+// STL allocator using memory arenas. Memory is allocated from underlying
+// blocks of size 'block_size * sizeof(T)'. Memory is freed only when all
+// objects using this allocator are destroyed and there is otherwise no reuse
+// (unlike PoolAllocator).
+//
+// This allocator has object-local state so it should not be used with splicing
+// or swapping operations between objects created with different allocators nor
+// should it be used if copies must be thread-safe. The result of allocate()
+// will be suitably memory-aligned.
+template <typename T>
+class BlockAllocator {
+ public:
+  using Allocator = std::allocator<T>;
+  using size_type = typename Allocator::size_type;
+  using difference_type = typename Allocator::difference_type;
+  using pointer = typename Allocator::pointer;
+  using const_pointer = typename Allocator::const_pointer;
+  using reference = typename Allocator::reference;
+  using const_reference = typename Allocator::const_reference;
+  using value_type = typename Allocator::value_type;
+
+  template <typename U>
+  struct rebind {
+    using other = BlockAllocator<U>;
+  };
+
+  explicit BlockAllocator(size_t block_size = kAllocSize)
+      : arenas_(new MemoryArenaCollection(block_size)) {}
+
+  BlockAllocator(const BlockAllocator<T> &arena_alloc)
+      : arenas_(arena_alloc.Arenas()) {
+    Arenas()->IncrRefCount();
+  }
+
+  template <typename U>
+  explicit BlockAllocator(const BlockAllocator<U> &arena_alloc)
+      : arenas_(arena_alloc.Arenas()) {
+    Arenas()->IncrRefCount();
+  }
+
+  ~BlockAllocator() {
+    if (Arenas()->DecrRefCount() == 0) delete Arenas();
+  }
+
+  pointer address(reference ref) const { return Allocator().address(ref); }
+
+  const_pointer address(const_reference ref) const {
+    return Allocator().address(ref);
+  }
+
+  size_type max_size() const { return Allocator().max_size(); }
+
+  template <class U, class... Args>
+  void construct(U *p, Args &&... args) {
+    Allocator().construct(p, std::forward<Args>(args)...);
+  }
+
+  void destroy(pointer p) { Allocator().destroy(p); }
+
+  pointer allocate(size_type n, const void *hint = nullptr) {
+    if (n * kAllocFit <= kAllocSize) {
+      return static_cast<pointer>(Arena()->Allocate(n));
+    } else {
+      return Allocator().allocate(n, hint);
+    }
+  }
+
+  void deallocate(pointer p, size_type n) {
+    if (n * kAllocFit > kAllocSize) Allocator().deallocate(p, n);
+  }
+
+  MemoryArenaCollection *Arenas() const { return arenas_; }
+
+ private:
+  MemoryArena<T> *Arena() { return arenas_->Arena<T>(); }
+
+  MemoryArenaCollection *arenas_;
+
+  BlockAllocator<T> operator=(const BlockAllocator<T> &);
+};
+
+template <typename T, typename U>
+bool operator==(const BlockAllocator<T> &alloc1,
+                const BlockAllocator<U> &alloc2) {
+  return false;
+}
+
+template <typename T, typename U>
+bool operator!=(const BlockAllocator<T> &alloc1,
+                const BlockAllocator<U> &alloc2) {
+  return true;
+}
+
+// STL allocator using memory pools. Memory is allocated from underlying
+// blocks of size 'block_size * sizeof(T)'. Keeps an internal list of freed
+// chunks thare are reused on the next allocation.
+//
+// This allocator has object-local state so it should not be used with splicing
+// or swapping operations between objects created with different allocators nor
+// should it be used if copies must be thread-safe. The result of allocate()
+// will be suitably memory-aligned.
+template <typename T>
+class PoolAllocator {
+ public:
+  using Allocator = std::allocator<T>;
+  using size_type = typename Allocator::size_type;
+  using difference_type = typename Allocator::difference_type;
+  using pointer = typename Allocator::pointer;
+  using const_pointer = typename Allocator::const_pointer;
+  using reference = typename Allocator::reference;
+  using const_reference = typename Allocator::const_reference;
+  using value_type = typename Allocator::value_type;
+
+  template <typename U>
+  struct rebind {
+    using other = PoolAllocator<U>;
+  };
+
+  explicit PoolAllocator(size_t pool_size = kAllocSize)
+      : pools_(new MemoryPoolCollection(pool_size)) {}
+
+  PoolAllocator(const PoolAllocator<T> &pool_alloc)
+      : pools_(pool_alloc.Pools()) {
+    Pools()->IncrRefCount();
+  }
+
+  template <typename U>
+  explicit PoolAllocator(const PoolAllocator<U> &pool_alloc)
+      : pools_(pool_alloc.Pools()) {
+    Pools()->IncrRefCount();
+  }
+
+  ~PoolAllocator() {
+    if (Pools()->DecrRefCount() == 0) delete Pools();
+  }
+
+  pointer address(reference ref) const { return Allocator().address(ref); }
+
+  const_pointer address(const_reference ref) const {
+    return Allocator().address(ref);
+  }
+
+  size_type max_size() const { return Allocator().max_size(); }
+
+  template <class U, class... Args>
+  void construct(U *p, Args &&... args) {
+    Allocator().construct(p, std::forward<Args>(args)...);
+  }
+
+  void destroy(pointer p) { Allocator().destroy(p); }
+
+  pointer allocate(size_type n, const void *hint = nullptr) {
+    if (n == 1) {
+      return static_cast<pointer>(Pool<1>()->Allocate());
+    } else if (n == 2) {
+      return static_cast<pointer>(Pool<2>()->Allocate());
+    } else if (n <= 4) {
+      return static_cast<pointer>(Pool<4>()->Allocate());
+    } else if (n <= 8) {
+      return static_cast<pointer>(Pool<8>()->Allocate());
+    } else if (n <= 16) {
+      return static_cast<pointer>(Pool<16>()->Allocate());
+    } else if (n <= 32) {
+      return static_cast<pointer>(Pool<32>()->Allocate());
+    } else if (n <= 64) {
+      return static_cast<pointer>(Pool<64>()->Allocate());
+    } else {
+      return Allocator().allocate(n, hint);
+    }
+  }
+
+  void deallocate(pointer p, size_type n) {
+    if (n == 1) {
+      Pool<1>()->Free(p);
+    } else if (n == 2) {
+      Pool<2>()->Free(p);
+    } else if (n <= 4) {
+      Pool<4>()->Free(p);
+    } else if (n <= 8) {
+      Pool<8>()->Free(p);
+    } else if (n <= 16) {
+      Pool<16>()->Free(p);
+    } else if (n <= 32) {
+      Pool<32>()->Free(p);
+    } else if (n <= 64) {
+      Pool<64>()->Free(p);
+    } else {
+      Allocator().deallocate(p, n);
+    }
+  }
+
+  MemoryPoolCollection *Pools() const { return pools_; }
+
+ private:
+  template <int n>
+  struct TN {
+    T buf[n];
+  };
+
+  template <int n>
+  MemoryPool<TN<n>> *Pool() {
+    return pools_->Pool<TN<n>>();
+  }
+
+  MemoryPoolCollection *pools_;
+
+  PoolAllocator<T> operator=(const PoolAllocator<T> &);
+};
+
+template <typename T, typename U>
+bool operator==(const PoolAllocator<T> &alloc1,
+                const PoolAllocator<U> &alloc2) {
+  return false;
+}
+
+template <typename T, typename U>
+bool operator!=(const PoolAllocator<T> &alloc1,
+                const PoolAllocator<U> &alloc2) {
+  return true;
+}
+
+}  // namespace fst
+
+#endif  // FST_MEMORY_H_
diff --git a/projects/llm_framework/include/fst/minimize.h b/projects/llm_framework/include/fst/minimize.h
new file mode 100644
index 00000000..9f17a22a
--- /dev/null
+++ b/projects/llm_framework/include/fst/minimize.h
@@ -0,0 +1,568 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes to minimize an FST.
+
+#ifndef FST_MINIMIZE_H_
+#define FST_MINIMIZE_H_
+
+#include <cmath>
+
+#include <algorithm>
+#include <map>
+#include <queue>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/arcsort.h>
+#include <fst/connect.h>
+#include <fst/dfs-visit.h>
+#include <fst/encode.h>
+#include <fst/factor-weight.h>
+#include <fst/fst.h>
+#include <fst/mutable-fst.h>
+#include <fst/partition.h>
+#include <fst/push.h>
+#include <fst/queue.h>
+#include <fst/reverse.h>
+#include <fst/shortest-distance.h>
+#include <fst/state-map.h>
+
+
+namespace fst {
+namespace internal {
+
+// Comparator for creating partition.
+template <class Arc>
+class StateComparator {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  StateComparator(const Fst<Arc> &fst, const Partition<StateId> &partition)
+      : fst_(fst), partition_(partition) {}
+
+  // Compares state x with state y based on sort criteria.
+  bool operator()(const StateId x, const StateId y) const {
+    // Checks for final state equivalence.
+    const auto xfinal = fst_.Final(x).Hash();
+    const auto yfinal = fst_.Final(y).Hash();
+    if (xfinal < yfinal) {
+      return true;
+    } else if (xfinal > yfinal) {
+      return false;
+    }
+    // Checks for number of arcs.
+    if (fst_.NumArcs(x) < fst_.NumArcs(y)) return true;
+    if (fst_.NumArcs(x) > fst_.NumArcs(y)) return false;
+    // If the number of arcs are equal, checks for arc match.
+    for (ArcIterator<Fst<Arc>> aiter1(fst_, x), aiter2(fst_, y);
+         !aiter1.Done() && !aiter2.Done(); aiter1.Next(), aiter2.Next()) {
+      const auto &arc1 = aiter1.Value();
+      const auto &arc2 = aiter2.Value();
+      if (arc1.ilabel < arc2.ilabel) return true;
+      if (arc1.ilabel > arc2.ilabel) return false;
+      if (partition_.ClassId(arc1.nextstate) <
+          partition_.ClassId(arc2.nextstate))
+        return true;
+      if (partition_.ClassId(arc1.nextstate) >
+          partition_.ClassId(arc2.nextstate))
+        return false;
+    }
+    return false;
+  }
+
+ private:
+  const Fst<Arc> &fst_;
+  const Partition<StateId> &partition_;
+};
+
+// Computes equivalence classes for cyclic unweighted acceptors. For cyclic
+// minimization we use the classic Hopcroft minimization algorithm, which has
+// complexity O(E log V) where E is the number of arcs and V is the number of
+// states.
+//
+// For more information, see:
+//
+//  Hopcroft, J. 1971. An n Log n algorithm for minimizing states in a finite
+//  automaton. Ms, Stanford University.
+//
+// Note: the original presentation of the paper was for a finite automaton (==
+// deterministic, unweighted acceptor), but we also apply it to the
+// nondeterministic case, where it is also applicable as long as the semiring is
+// idempotent (if the semiring is not idempotent, there are some complexities
+// in keeping track of the weight when there are multiple arcs to states that
+// will be merged, and we don't deal with this).
+template <class Arc, class Queue>
+class CyclicMinimizer {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using ClassId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using RevArc = ReverseArc<Arc>;
+
+  explicit CyclicMinimizer(const ExpandedFst<Arc> &fst) {
+    Initialize(fst);
+    Compute(fst);
+  }
+
+  const Partition<StateId> &GetPartition() const { return P_; }
+
+ private:
+  // StateILabelHasher is a hashing object that computes a hash-function
+  // of an FST state that depends only on the set of ilabels on arcs leaving
+  // the state [note: it assumes that the arcs are ilabel-sorted].
+  // In order to work correctly for non-deterministic automata, multiple
+  // instances of the same ilabel count the same as a single instance.
+  class StateILabelHasher {
+   public:
+    explicit StateILabelHasher(const Fst<Arc> &fst) : fst_(fst) {}
+
+    using Label = typename Arc::Label;
+    using StateId = typename Arc::StateId;
+
+    size_t operator()(const StateId s) {
+      const size_t p1 = 7603;
+      const size_t p2 = 433024223;
+      size_t result = p2;
+      size_t current_ilabel = kNoLabel;
+      for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
+        Label this_ilabel = aiter.Value().ilabel;
+        if (this_ilabel != current_ilabel) {  // Ignores repeats.
+          result = p1 * result + this_ilabel;
+          current_ilabel = this_ilabel;
+        }
+      }
+      return result;
+    }
+
+   private:
+    const Fst<Arc> &fst_;
+  };
+
+  class ArcIterCompare {
+   public:
+    explicit ArcIterCompare(const Partition<StateId> &partition)
+        : partition_(partition) {}
+
+    ArcIterCompare(const ArcIterCompare &comp) : partition_(comp.partition_) {}
+
+    // Compares two iterators based on their input labels.
+    bool operator()(const ArcIterator<Fst<RevArc>> *x,
+                    const ArcIterator<Fst<RevArc>> *y) const {
+      const auto &xarc = x->Value();
+      const auto &yarc = y->Value();
+      return xarc.ilabel > yarc.ilabel;
+    }
+
+   private:
+    const Partition<StateId> &partition_;
+  };
+
+  using ArcIterQueue =
+      std::priority_queue<ArcIterator<Fst<RevArc>> *,
+                          std::vector<ArcIterator<Fst<RevArc>> *>,
+                          ArcIterCompare>;
+
+ private:
+  // Prepartitions the space into equivalence classes. We ensure that final and
+  // non-final states always go into different equivalence classes, and we use
+  // class StateILabelHasher to make sure that most of the time, states with
+  // different sets of ilabels on arcs leaving them, go to different partitions.
+  // Note: for the O(n) guarantees we don't rely on the goodness of this
+  // hashing function---it just provides a bonus speedup.
+  void PrePartition(const ExpandedFst<Arc> &fst) {
+    VLOG(5) << "PrePartition";
+    StateId next_class = 0;
+    auto num_states = fst.NumStates();
+    // Allocates a temporary vector to store the initial class mappings, so that
+    // we can allocate the classes all at once.
+    std::vector<StateId> state_to_initial_class(num_states);
+    {
+      // We maintain two maps from hash-value to class---one for final states
+      // (final-prob == One()) and one for non-final states
+      // (final-prob == Zero()). We are processing unweighted acceptors, so the
+      // are the only two possible values.
+      using HashToClassMap = std::unordered_map<size_t, StateId>;
+      HashToClassMap hash_to_class_nonfinal;
+      HashToClassMap hash_to_class_final;
+      StateILabelHasher hasher(fst);
+      for (StateId s = 0; s < num_states; ++s) {
+        size_t hash = hasher(s);
+        HashToClassMap &this_map =
+            (fst.Final(s) != Weight::Zero() ? hash_to_class_final
+                                            : hash_to_class_nonfinal);
+        // Avoids two map lookups by using 'insert' instead of 'find'.
+        auto p = this_map.insert(std::make_pair(hash, next_class));
+        state_to_initial_class[s] = p.second ? next_class++ : p.first->second;
+      }
+      // Lets the unordered_maps go out of scope before we allocate the classes,
+      // to reduce the maximum amount of memory used.
+    }
+    P_.AllocateClasses(next_class);
+    for (StateId s = 0; s < num_states; ++s) {
+      P_.Add(s, state_to_initial_class[s]);
+    }
+    for (StateId c = 0; c < next_class; ++c) L_.Enqueue(c);
+    VLOG(5) << "Initial Partition: " << P_.NumClasses();
+  }
+
+  // Creates inverse transition Tr_ = rev(fst), loops over states in FST and
+  // splits on final, creating two blocks in the partition corresponding to
+  // final, non-final.
+  void Initialize(const ExpandedFst<Arc> &fst) {
+    // Constructs Tr.
+    Reverse(fst, &Tr_);
+    ILabelCompare<RevArc> ilabel_comp;
+    ArcSort(&Tr_, ilabel_comp);
+    // Tells the partition how many elements to allocate. The first state in
+    // Tr_ is super-final state.
+    P_.Initialize(Tr_.NumStates() - 1);
+    // Prepares initial partition.
+    PrePartition(fst);
+    // Allocates arc iterator queue.
+    ArcIterCompare comp(P_);
+    aiter_queue_.reset(new ArcIterQueue(comp));
+  }
+  // Partitions all classes with destination C.
+  void Split(ClassId C) {
+    // Prepares priority queue: opens arc iterator for each state in C, and
+    // inserts into priority queue.
+    for (PartitionIterator<StateId> siter(P_, C); !siter.Done(); siter.Next()) {
+      StateId s = siter.Value();
+      if (Tr_.NumArcs(s + 1)) {
+        aiter_queue_->push(new ArcIterator<Fst<RevArc>>(Tr_, s + 1));
+      }
+    }
+    // Now pops arc iterator from queue, splits entering equivalence class, and
+    // re-inserts updated iterator into queue.
+    Label prev_label = -1;
+    while (!aiter_queue_->empty()) {
+      std::unique_ptr<ArcIterator<Fst<RevArc>>> aiter(aiter_queue_->top());
+      aiter_queue_->pop();
+      if (aiter->Done()) continue;
+      const auto &arc = aiter->Value();
+      auto from_state = aiter->Value().nextstate - 1;
+      auto from_label = arc.ilabel;
+      if (prev_label != from_label) P_.FinalizeSplit(&L_);
+      auto from_class = P_.ClassId(from_state);
+      if (P_.ClassSize(from_class) > 1) P_.SplitOn(from_state);
+      prev_label = from_label;
+      aiter->Next();
+      if (!aiter->Done()) aiter_queue_->push(aiter.release());
+    }
+    P_.FinalizeSplit(&L_);
+  }
+
+  // Main loop for Hopcroft minimization.
+  void Compute(const Fst<Arc> &fst) {
+    // Processes active classes (FIFO, or FILO).
+    while (!L_.Empty()) {
+      const auto C = L_.Head();
+      L_.Dequeue();
+      Split(C);  // Splits on C, all labels in C.
+    }
+  }
+
+ private:
+  // Partioning of states into equivalence classes.
+  Partition<StateId> P_;
+  // Set of active classes to be processed in partition P.
+  Queue L_;
+  // Reverses transition function.
+  VectorFst<RevArc> Tr_;
+  // Priority queue of open arc iterators for all states in the splitter
+  // equivalence class.
+  std::unique_ptr<ArcIterQueue> aiter_queue_;
+};
+
+// Computes equivalence classes for acyclic FST.
+//
+// Complexity:
+//
+//   O(E)
+//
+// where E is the number of arcs.
+//
+// For more information, see:
+//
+// Revuz, D. 1992. Minimization of acyclic deterministic automata in linear
+// time. Theoretical Computer Science 92(1): 181-189.
+template <class Arc>
+class AcyclicMinimizer {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using ClassId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  explicit AcyclicMinimizer(const ExpandedFst<Arc> &fst) {
+    Initialize(fst);
+    Refine(fst);
+  }
+
+  const Partition<StateId> &GetPartition() { return partition_; }
+
+ private:
+  // DFS visitor to compute the height (distance) to final state.
+  class HeightVisitor {
+   public:
+    HeightVisitor() : max_height_(0), num_states_(0) {}
+
+    // Invoked before DFS visit.
+    void InitVisit(const Fst<Arc> &fst) {}
+
+    // Invoked when state is discovered (2nd arg is DFS tree root).
+    bool InitState(StateId s, StateId root) {
+      // Extends height array and initialize height (distance) to 0.
+      for (StateId i = height_.size(); i <= s; ++i) height_.push_back(-1);
+      if (s >= num_states_) num_states_ = s + 1;
+      return true;
+    }
+
+    // Invoked when tree arc examined (to undiscovered state).
+    bool TreeArc(StateId s, const Arc &arc) { return true; }
+
+    // Invoked when back arc examined (to unfinished state).
+    bool BackArc(StateId s, const Arc &arc) { return true; }
+
+    // Invoked when forward or cross arc examined (to finished state).
+    bool ForwardOrCrossArc(StateId s, const Arc &arc) {
+      if (height_[arc.nextstate] + 1 > height_[s]) {
+        height_[s] = height_[arc.nextstate] + 1;
+      }
+      return true;
+    }
+
+    // Invoked when state finished (parent is kNoStateId for tree root).
+    void FinishState(StateId s, StateId parent, const Arc *parent_arc) {
+      if (height_[s] == -1) height_[s] = 0;
+      const auto h = height_[s] + 1;
+      if (parent >= 0) {
+        if (h > height_[parent]) height_[parent] = h;
+        if (h > max_height_) max_height_ = h;
+      }
+    }
+
+    // Invoked after DFS visit.
+    void FinishVisit() {}
+
+    size_t max_height() const { return max_height_; }
+
+    const std::vector<StateId> &height() const { return height_; }
+
+    size_t num_states() const { return num_states_; }
+
+   private:
+    std::vector<StateId> height_;
+    size_t max_height_;
+    size_t num_states_;
+  };
+
+ private:
+  // Cluster states according to height (distance to final state)
+  void Initialize(const Fst<Arc> &fst) {
+    // Computes height (distance to final state).
+    HeightVisitor hvisitor;
+    DfsVisit(fst, &hvisitor);
+    // Creates initial partition based on height.
+    partition_.Initialize(hvisitor.num_states());
+    partition_.AllocateClasses(hvisitor.max_height() + 1);
+    const auto &hstates = hvisitor.height();
+    for (StateId s = 0; s < hstates.size(); ++s) partition_.Add(s, hstates[s]);
+  }
+
+  // Refines states based on arc sort (out degree, arc equivalence).
+  void Refine(const Fst<Arc> &fst) {
+    using EquivalenceMap = std::map<StateId, StateId, StateComparator<Arc>>;
+    StateComparator<Arc> comp(fst, partition_);
+    // Starts with tail (height = 0).
+    auto height = partition_.NumClasses();
+    for (StateId h = 0; h < height; ++h) {
+      EquivalenceMap equiv_classes(comp);
+      // Sorts states within equivalence class.
+      PartitionIterator<StateId> siter(partition_, h);
+      equiv_classes[siter.Value()] = h;
+      for (siter.Next(); !siter.Done(); siter.Next()) {
+        auto insert_result =
+            equiv_classes.insert(std::make_pair(siter.Value(), kNoStateId));
+        if (insert_result.second) {
+          insert_result.first->second = partition_.AddClass();
+        }
+      }
+      // Creates refined partition.
+      for (siter.Reset(); !siter.Done();) {
+        const auto s = siter.Value();
+        const auto old_class = partition_.ClassId(s);
+        const auto new_class = equiv_classes[s];
+        // A move operation can invalidate the iterator, so we first update
+        // the iterator to the next element before we move the current element
+        // out of the list.
+        siter.Next();
+        if (old_class != new_class) partition_.Move(s, new_class);
+      }
+    }
+  }
+
+ private:
+  Partition<StateId> partition_;
+};
+
+// Given a partition and a Mutable FST, merges states of Fst in place (i.e.,
+// destructively). Merging works by taking the first state in a class of the
+// partition to be the representative state for the class. Each arc is then
+// reconnected to this state. All states in the class are merged by adding
+// their arcs to the representative state.
+template <class Arc>
+void MergeStates(const Partition<typename Arc::StateId> &partition,
+                 MutableFst<Arc> *fst) {
+  using StateId = typename Arc::StateId;
+  std::vector<StateId> state_map(partition.NumClasses());
+  for (StateId i = 0; i < partition.NumClasses(); ++i) {
+    PartitionIterator<StateId> siter(partition, i);
+    state_map[i] = siter.Value();  // First state in partition.
+  }
+  // Relabels destination states.
+  for (StateId c = 0; c < partition.NumClasses(); ++c) {
+    for (PartitionIterator<StateId> siter(partition, c); !siter.Done();
+         siter.Next()) {
+      const auto s = siter.Value();
+      for (MutableArcIterator<MutableFst<Arc>> aiter(fst, s); !aiter.Done();
+           aiter.Next()) {
+        auto arc = aiter.Value();
+        arc.nextstate = state_map[partition.ClassId(arc.nextstate)];
+        if (s == state_map[c]) {  // For the first state, just sets destination.
+          aiter.SetValue(arc);
+        } else {
+          fst->AddArc(state_map[c], std::move(arc));
+        }
+      }
+    }
+  }
+  fst->SetStart(state_map[partition.ClassId(fst->Start())]);
+  Connect(fst);
+}
+
+template <class Arc>
+void AcceptorMinimize(MutableFst<Arc> *fst,
+                      bool allow_acyclic_minimization = true) {
+  if (!(fst->Properties(kAcceptor | kUnweighted, true) ==
+        (kAcceptor | kUnweighted))) {
+    FSTERROR() << "FST is not an unweighted acceptor";
+    fst->SetProperties(kError, kError);
+    return;
+  }
+  // Connects FST before minimization, handles disconnected states.
+  Connect(fst);
+  if (fst->NumStates() == 0) return;
+  if (allow_acyclic_minimization && fst->Properties(kAcyclic, true)) {
+    // Acyclic minimization (Revuz).
+    VLOG(2) << "Acyclic minimization";
+    ArcSort(fst, ILabelCompare<Arc>());
+    AcyclicMinimizer<Arc> minimizer(*fst);
+    MergeStates(minimizer.GetPartition(), fst);
+  } else {
+    // Either the FST has cycles, or it's generated from non-deterministic input
+    // (which the Revuz algorithm can't handle), so use the cyclic minimization
+    // algorithm of Hopcroft.
+    VLOG(2) << "Cyclic minimization";
+    CyclicMinimizer<Arc, LifoQueue<typename Arc::StateId>> minimizer(*fst);
+    MergeStates(minimizer.GetPartition(), fst);
+  }
+  // Merges in appropriate semiring
+  ArcUniqueMapper<Arc> mapper(*fst);
+  StateMap(fst, mapper);
+}
+
+}  // namespace internal
+
+// In place minimization of deterministic weighted automata and transducers,
+// and also non-deterministic ones if they use an idempotent semiring.
+// For transducers, if the 'sfst' argument is not null, the algorithm
+// produces a compact factorization of the minimal transducer.
+//
+// In the acyclic deterministic case, we use an algorithm from Revuz that is
+// linear in the number of arcs (edges) in the machine.
+//
+// In the cyclic or non-deterministic case, we use the classical Hopcroft
+// minimization (which was presented for the deterministic case but which
+// also works for non-deterministic FSTs); this has complexity O(e log v).
+//
+template <class Arc>
+void Minimize(MutableFst<Arc> *fst, MutableFst<Arc> *sfst = nullptr,
+              float delta = kShortestDelta, bool allow_nondet = false) {
+  using Weight = typename Arc::Weight;
+  const auto props = fst->Properties(
+      kAcceptor | kIDeterministic | kWeighted | kUnweighted, true);
+  bool allow_acyclic_minimization;
+  if (props & kIDeterministic) {
+    allow_acyclic_minimization = true;
+  } else {
+    // Our approach to minimization of non-deterministic FSTs will only work in
+    // idempotent semirings---for non-deterministic inputs, a state could have
+    // multiple transitions to states that will get merged, and we'd have to
+    // sum their weights. The algorithm doesn't handle that.
+    if (!(Weight::Properties() & kIdempotent)) {
+      fst->SetProperties(kError, kError);
+      FSTERROR() << "Cannot minimize a non-deterministic FST over a "
+                    "non-idempotent semiring";
+      return;
+    } else if (!allow_nondet) {
+      fst->SetProperties(kError, kError);
+      FSTERROR() << "Refusing to minimize a non-deterministic FST with "
+                 << "allow_nondet = false";
+      return;
+    }
+    // The Revuz algorithm won't work for nondeterministic inputs, so if the
+    // input is nondeterministic, we'll have to pass a bool saying not to use
+    // that algorithm. We check at this level rather than in AcceptorMinimize(),
+    // because it's possible that the FST at this level could be deterministic,
+    // but a harmless type of non-determinism could be introduced by Encode()
+    // (thanks to kEncodeWeights, if the FST has epsilons and has a final
+    // weight with weights equal to some epsilon arc.)
+    allow_acyclic_minimization = false;
+  }
+  if (!(props & kAcceptor)) {  // Weighted transducer.
+    VectorFst<GallicArc<Arc, GALLIC_LEFT>> gfst;
+    ArcMap(*fst, &gfst, ToGallicMapper<Arc, GALLIC_LEFT>());
+    fst->DeleteStates();
+    gfst.SetProperties(kAcceptor, kAcceptor);
+    Push(&gfst, REWEIGHT_TO_INITIAL, delta);
+    ArcMap(&gfst, QuantizeMapper<GallicArc<Arc, GALLIC_LEFT>>(delta));
+    EncodeMapper<GallicArc<Arc, GALLIC_LEFT>> encoder(
+        kEncodeLabels | kEncodeWeights, ENCODE);
+    Encode(&gfst, &encoder);
+    internal::AcceptorMinimize(&gfst, allow_acyclic_minimization);
+    Decode(&gfst, encoder);
+    if (!sfst) {
+      FactorWeightFst<GallicArc<Arc, GALLIC_LEFT>,
+                      GallicFactor<typename Arc::Label, Weight, GALLIC_LEFT>>
+          fwfst(gfst);
+      std::unique_ptr<SymbolTable> osyms(
+          fst->OutputSymbols() ? fst->OutputSymbols()->Copy() : nullptr);
+      ArcMap(fwfst, fst, FromGallicMapper<Arc, GALLIC_LEFT>());
+      fst->SetOutputSymbols(osyms.get());
+    } else {
+      sfst->SetOutputSymbols(fst->OutputSymbols());
+      GallicToNewSymbolsMapper<Arc, GALLIC_LEFT> mapper(sfst);
+      ArcMap(gfst, fst, &mapper);
+      fst->SetOutputSymbols(sfst->InputSymbols());
+    }
+  } else if (props & kWeighted) {  // Weighted acceptor.
+    Push(fst, REWEIGHT_TO_INITIAL, delta);
+    ArcMap(fst, QuantizeMapper<Arc>(delta));
+    EncodeMapper<Arc> encoder(kEncodeLabels | kEncodeWeights, ENCODE);
+    Encode(fst, &encoder);
+    internal::AcceptorMinimize(fst, allow_acyclic_minimization);
+    Decode(fst, encoder);
+  } else {  // Unweighted acceptor.
+    internal::AcceptorMinimize(fst, allow_acyclic_minimization);
+  }
+}
+
+}  // namespace fst
+
+#endif  // FST_MINIMIZE_H_
diff --git a/projects/llm_framework/include/fst/mutable-fst.h b/projects/llm_framework/include/fst/mutable-fst.h
new file mode 100644
index 00000000..9031770d
--- /dev/null
+++ b/projects/llm_framework/include/fst/mutable-fst.h
@@ -0,0 +1,398 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Expanded FST augmented with mutators; interface class definition and
+// mutable arc iterator interface.
+
+#ifndef FST_MUTABLE_FST_H_
+#define FST_MUTABLE_FST_H_
+
+#include <stddef.h>
+#include <sys/types.h>
+
+#include <istream>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+#include <fstream>
+
+#include <fst/expanded-fst.h>
+
+
+namespace fst {
+
+template <class Arc>
+struct MutableArcIteratorData;
+
+// Abstract interface for an expanded FST which also supports mutation
+// operations. To modify arcs, use MutableArcIterator.
+template <class A>
+class MutableFst : public ExpandedFst<A> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  virtual MutableFst<Arc> &operator=(const Fst<Arc> &fst) = 0;
+
+  MutableFst<Arc> &operator=(const MutableFst<Arc> &fst) {
+    return operator=(static_cast<const Fst<Arc> &>(fst));
+  }
+
+  // Sets the initial state.
+  virtual void SetStart(StateId) = 0;
+
+  // Sets a state's final weight.
+  virtual void SetFinal(StateId, Weight) = 0;
+
+  // Sets property bits w.r.t. mask.
+  virtual void SetProperties(uint64 props, uint64 mask) = 0;
+
+  // Adds a state and returns its ID.
+  virtual StateId AddState() = 0;
+
+  // Adds an arc to state.
+  virtual void AddArc(StateId, const Arc &arc) = 0;
+
+  // Adds an arc (passed by rvalue reference) to state. Allows subclasses
+  // to optionally implement move semantics. Defaults to lvalue overload.
+  virtual void AddArc(StateId state, Arc &&arc) { AddArc(state, arc); }
+
+  // Deletes some states, preserving original StateId ordering.
+  virtual void DeleteStates(const std::vector<StateId> &) = 0;
+
+  // Delete all states.
+  virtual void DeleteStates() = 0;
+
+  // Delete some arcs at a given state.
+  virtual void DeleteArcs(StateId, size_t n) = 0;
+
+  // Delete all arcs at a given state.
+  virtual void DeleteArcs(StateId) = 0;
+
+  // Optional, best effort only.
+  virtual void ReserveStates(StateId n) {}
+
+  // Optional, best effort only.
+  virtual void ReserveArcs(StateId s, size_t n) {}
+
+  // Returns input label symbol table or nullptr if not specified.
+  const SymbolTable *InputSymbols() const override = 0;
+
+  // Returns output label symbol table or nullptr if not specified.
+  const SymbolTable *OutputSymbols() const override = 0;
+
+  // Returns input label symbol table or nullptr if not specified.
+  virtual SymbolTable *MutableInputSymbols() = 0;
+
+  // Returns output label symbol table or nullptr if not specified.
+  virtual SymbolTable *MutableOutputSymbols() = 0;
+
+  // Sets input label symbol table; pass nullptr to delete table.
+  virtual void SetInputSymbols(const SymbolTable *isyms) = 0;
+
+  // Sets output label symbol table; pass nullptr to delete table.
+  virtual void SetOutputSymbols(const SymbolTable *osyms) = 0;
+
+  // Gets a copy of this MutableFst. See Fst<>::Copy() for further doc.
+  MutableFst<A> *Copy(bool safe = false) const override = 0;
+
+  // Reads a MutableFst from an input stream, returning nullptr on error.
+  static MutableFst<Arc> *Read(std::istream &strm, const FstReadOptions &opts) {
+    FstReadOptions ropts(opts);
+    FstHeader hdr;
+    if (ropts.header) {
+      hdr = *opts.header;
+    } else {
+      if (!hdr.Read(strm, opts.source)) return nullptr;
+      ropts.header = &hdr;
+    }
+    if (!(hdr.Properties() & kMutable)) {
+      LOG(ERROR) << "MutableFst::Read: Not a MutableFst: " << ropts.source;
+      return nullptr;
+    }
+    const auto &fst_type = hdr.FstType();
+    const auto reader = FstRegister<Arc>::GetRegister()->GetReader(fst_type);
+    if (!reader) {
+      LOG(ERROR) << "MutableFst::Read: Unknown FST type \"" << fst_type
+                 << "\" (arc type = \"" << A::Type() << "\"): " << ropts.source;
+      return nullptr;
+    }
+    auto *fst = reader(strm, ropts);
+    if (!fst) return nullptr;
+    return static_cast<MutableFst<Arc> *>(fst);
+  }
+
+  // Reads a MutableFst from a file; returns nullptr on error. An empty
+  // filename results in reading from standard input. If convert is true,
+  // convert to a mutable FST subclass (given by convert_type) in the case
+  // that the input FST is non-mutable.
+  static MutableFst<Arc> *Read(const string &filename, bool convert = false,
+                               const string &convert_type = "vector") {
+    if (convert == false) {
+      if (!filename.empty()) {
+        std::ifstream strm(filename,
+                                std::ios_base::in | std::ios_base::binary);
+        if (!strm) {
+          LOG(ERROR) << "MutableFst::Read: Can't open file: " << filename;
+          return nullptr;
+        }
+        return Read(strm, FstReadOptions(filename));
+      } else {
+        return Read(std::cin, FstReadOptions("standard input"));
+      }
+    } else {  // Converts to 'convert_type' if not mutable.
+      std::unique_ptr<Fst<Arc>> ifst(Fst<Arc>::Read(filename));
+      if (!ifst) return nullptr;
+      if (ifst->Properties(kMutable, false)) {
+        return static_cast<MutableFst<Arc> *>(ifst.release());
+      } else {
+        std::unique_ptr<Fst<Arc>> ofst(Convert(*ifst, convert_type));
+        ifst.reset();
+        if (!ofst) return nullptr;
+        if (!ofst->Properties(kMutable, false)) {
+          LOG(ERROR) << "MutableFst: Bad convert type: " << convert_type;
+        }
+        return static_cast<MutableFst<Arc> *>(ofst.release());
+      }
+    }
+  }
+
+  // For generic mutuble arc iterator construction; not normally called
+  // directly by users.
+  virtual void InitMutableArcIterator(StateId s,
+                                      MutableArcIteratorData<Arc> *data) = 0;
+};
+
+// Mutable arc iterator interface, templated on the Arc definition. This is
+// used by mutable arc iterator specializations that are returned by the
+// InitMutableArcIterator MutableFst method.
+template <class Arc>
+class MutableArcIteratorBase : public ArcIteratorBase<Arc> {
+ public:
+  // Sets current arc.
+  virtual void SetValue(const Arc &) = 0;
+};
+
+template <class Arc>
+struct MutableArcIteratorData {
+  MutableArcIteratorBase<Arc> *base;  // Specific iterator.
+};
+
+// Generic mutable arc iterator, templated on the FST definition; a wrapper
+// around a pointer to a more specific one.
+//
+// Here is a typical use:
+//
+//   for (MutableArcIterator<StdFst> aiter(&fst, s);
+//        !aiter.Done();
+//         aiter.Next()) {
+//     StdArc arc = aiter.Value();
+//     arc.ilabel = 7;
+//     aiter.SetValue(arc);
+//     ...
+//   }
+//
+// This version requires function calls.
+template <class FST>
+class MutableArcIterator {
+ public:
+  using Arc = typename FST::Arc;
+  using StateId = typename Arc::StateId;
+
+  MutableArcIterator(FST *fst, StateId s) {
+    fst->InitMutableArcIterator(s, &data_);
+  }
+
+  ~MutableArcIterator() { delete data_.base; }
+
+  bool Done() const { return data_.base->Done(); }
+
+  const Arc &Value() const { return data_.base->Value(); }
+
+  void Next() { data_.base->Next(); }
+
+  size_t Position() const { return data_.base->Position(); }
+
+  void Reset() { data_.base->Reset(); }
+
+  void Seek(size_t a) { data_.base->Seek(a); }
+
+  void SetValue(const Arc &arc) { data_.base->SetValue(arc); }
+
+  uint32 Flags() const { return data_.base->Flags(); }
+
+  void SetFlags(uint32 flags, uint32 mask) {
+    return data_.base->SetFlags(flags, mask);
+  }
+
+ private:
+  MutableArcIteratorData<Arc> data_;
+
+  MutableArcIterator(const MutableArcIterator &) = delete;
+  MutableArcIterator &operator=(const MutableArcIterator &) = delete;
+};
+
+namespace internal {
+
+// MutableFst<A> case: abstract methods.
+template <class Arc>
+inline typename Arc::Weight Final(const MutableFst<Arc> &fst,
+                                  typename Arc::StateId s) {
+  return fst.Final(s);
+}
+
+template <class Arc>
+inline ssize_t NumArcs(const MutableFst<Arc> &fst, typename Arc::StateId s) {
+  return fst.NumArcs(s);
+}
+
+template <class Arc>
+inline ssize_t NumInputEpsilons(const MutableFst<Arc> &fst,
+                                typename Arc::StateId s) {
+  return fst.NumInputEpsilons(s);
+}
+
+template <class Arc>
+inline ssize_t NumOutputEpsilons(const MutableFst<Arc> &fst,
+                                 typename Arc::StateId s) {
+  return fst.NumOutputEpsilons(s);
+}
+
+}  // namespace internal
+
+// A useful alias when using StdArc.
+using StdMutableFst = MutableFst<StdArc>;
+
+// This is a helper class template useful for attaching a MutableFst interface
+// to its implementation, handling reference counting and COW semantics.
+template <class Impl, class FST = MutableFst<typename Impl::Arc>>
+class ImplToMutableFst : public ImplToExpandedFst<Impl, FST> {
+ public:
+  using Arc = typename Impl::Arc;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using ImplToExpandedFst<Impl, FST>::operator=;
+
+  void SetStart(StateId s) override {
+    MutateCheck();
+    GetMutableImpl()->SetStart(s);
+  }
+
+  void SetFinal(StateId s, Weight weight) override {
+    MutateCheck();
+    GetMutableImpl()->SetFinal(s, std::move(weight));
+  }
+
+  void SetProperties(uint64 props, uint64 mask) override {
+    // Can skip mutate check if extrinsic properties don't change,
+    // since it is then safe to update all (shallow) copies
+    const auto exprops = kExtrinsicProperties & mask;
+    if (GetImpl()->Properties(exprops) != (props & exprops)) MutateCheck();
+    GetMutableImpl()->SetProperties(props, mask);
+  }
+
+  StateId AddState() override {
+    MutateCheck();
+    return GetMutableImpl()->AddState();
+  }
+
+  void AddArc(StateId s, const Arc &arc) override {
+    MutateCheck();
+    GetMutableImpl()->AddArc(s, arc);
+  }
+
+  void AddArc(StateId s, Arc &&arc) override {
+    MutateCheck();
+    GetMutableImpl()->AddArc(s, std::move(arc));
+  }
+
+  void DeleteStates(const std::vector<StateId> &dstates) override {
+    MutateCheck();
+    GetMutableImpl()->DeleteStates(dstates);
+  }
+
+  void DeleteStates() override {
+    if (!Unique()) {
+      const auto *isymbols = GetImpl()->InputSymbols();
+      const auto *osymbols = GetImpl()->OutputSymbols();
+      SetImpl(std::make_shared<Impl>());
+      GetMutableImpl()->SetInputSymbols(isymbols);
+      GetMutableImpl()->SetOutputSymbols(osymbols);
+    } else {
+      GetMutableImpl()->DeleteStates();
+    }
+  }
+
+  void DeleteArcs(StateId s, size_t n) override {
+    MutateCheck();
+    GetMutableImpl()->DeleteArcs(s, n);
+  }
+
+  void DeleteArcs(StateId s) override {
+    MutateCheck();
+    GetMutableImpl()->DeleteArcs(s);
+  }
+
+  void ReserveStates(StateId s) override {
+    MutateCheck();
+    GetMutableImpl()->ReserveStates(s);
+  }
+
+  void ReserveArcs(StateId s, size_t n) override {
+    MutateCheck();
+    GetMutableImpl()->ReserveArcs(s, n);
+  }
+
+  const SymbolTable *InputSymbols() const override {
+    return GetImpl()->InputSymbols();
+  }
+
+  const SymbolTable *OutputSymbols() const override {
+    return GetImpl()->OutputSymbols();
+  }
+
+  SymbolTable *MutableInputSymbols() override {
+    MutateCheck();
+    return GetMutableImpl()->InputSymbols();
+  }
+
+  SymbolTable *MutableOutputSymbols() override {
+    MutateCheck();
+    return GetMutableImpl()->OutputSymbols();
+  }
+
+  void SetInputSymbols(const SymbolTable *isyms) override {
+    MutateCheck();
+    GetMutableImpl()->SetInputSymbols(isyms);
+  }
+
+  void SetOutputSymbols(const SymbolTable *osyms) override {
+    MutateCheck();
+    GetMutableImpl()->SetOutputSymbols(osyms);
+  }
+
+ protected:
+  using ImplToExpandedFst<Impl, FST>::GetImpl;
+  using ImplToExpandedFst<Impl, FST>::GetMutableImpl;
+  using ImplToExpandedFst<Impl, FST>::Unique;
+  using ImplToExpandedFst<Impl, FST>::SetImpl;
+  using ImplToExpandedFst<Impl, FST>::InputSymbols;
+
+  explicit ImplToMutableFst(std::shared_ptr<Impl> impl)
+      : ImplToExpandedFst<Impl, FST>(impl) {}
+
+  ImplToMutableFst(const ImplToMutableFst<Impl, FST> &fst, bool safe)
+      : ImplToExpandedFst<Impl, FST>(fst, safe) {}
+
+  void MutateCheck() {
+    if (!Unique()) SetImpl(std::make_shared<Impl>(*this));
+  }
+};
+
+}  // namespace fst
+
+#endif  // FST_MUTABLE_FST_H_
diff --git a/projects/llm_framework/include/fst/pair-weight.h b/projects/llm_framework/include/fst/pair-weight.h
new file mode 100644
index 00000000..1f2e963c
--- /dev/null
+++ b/projects/llm_framework/include/fst/pair-weight.h
@@ -0,0 +1,155 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Pair weight templated base class for weight classes that contain two weights
+// (e.g. Product, Lexicographic).
+
+#ifndef FST_PAIR_WEIGHT_H_
+#define FST_PAIR_WEIGHT_H_
+
+#include <climits>
+#include <stack>
+#include <string>
+#include <utility>
+
+#include <fst/flags.h>
+#include <fst/log.h>
+
+#include <fst/weight.h>
+
+
+namespace fst {
+
+template <class W1, class W2>
+class PairWeight {
+ public:
+  using ReverseWeight =
+      PairWeight<typename W1::ReverseWeight, typename W2::ReverseWeight>;
+
+  PairWeight() {}
+
+  PairWeight(W1 w1, W2 w2) : value1_(std::move(w1)), value2_(std::move(w2)) {}
+
+  static const PairWeight<W1, W2> &Zero() {
+    static const PairWeight zero(W1::Zero(), W2::Zero());
+    return zero;
+  }
+
+  static const PairWeight<W1, W2> &One() {
+    static const PairWeight one(W1::One(), W2::One());
+    return one;
+  }
+
+  static const PairWeight<W1, W2> &NoWeight() {
+    static const PairWeight no_weight(W1::NoWeight(), W2::NoWeight());
+    return no_weight;
+  }
+
+  std::istream &Read(std::istream &strm) {
+    value1_.Read(strm);
+    return value2_.Read(strm);
+  }
+
+  std::ostream &Write(std::ostream &strm) const {
+    value1_.Write(strm);
+    return value2_.Write(strm);
+  }
+
+  bool Member() const { return value1_.Member() && value2_.Member(); }
+
+  size_t Hash() const {
+    const auto h1 = value1_.Hash();
+    const auto h2 = value2_.Hash();
+    static constexpr int lshift = 5;
+    static constexpr int rshift = CHAR_BIT * sizeof(size_t) - 5;
+    return h1 << lshift ^ h1 >> rshift ^ h2;
+  }
+
+  PairWeight<W1, W2> Quantize(float delta = kDelta) const {
+    return PairWeight<W1, W2>(value1_.Quantize(delta), value2_.Quantize(delta));
+  }
+
+  ReverseWeight Reverse() const {
+    return ReverseWeight(value1_.Reverse(), value2_.Reverse());
+  }
+
+  const W1 &Value1() const { return value1_; }
+
+  const W2 &Value2() const { return value2_; }
+
+  void SetValue1(const W1 &weight) { value1_ = weight; }
+
+  void SetValue2(const W2 &weight) { value2_ = weight; }
+
+ private:
+  W1 value1_;
+  W2 value2_;
+};
+
+template <class W1, class W2>
+inline bool operator==(const PairWeight<W1, W2> &w1,
+                       const PairWeight<W1, W2> &w2) {
+  return w1.Value1() == w2.Value1() && w1.Value2() == w2.Value2();
+}
+
+template <class W1, class W2>
+inline bool operator!=(const PairWeight<W1, W2> &w1,
+                       const PairWeight<W1, W2> &w2) {
+  return w1.Value1() != w2.Value1() || w1.Value2() != w2.Value2();
+}
+
+template <class W1, class W2>
+inline bool ApproxEqual(const PairWeight<W1, W2> &w1,
+                        const PairWeight<W1, W2> &w2, float delta = kDelta) {
+  return ApproxEqual(w1.Value1(), w2.Value1(), delta) &&
+         ApproxEqual(w1.Value2(), w2.Value2(), delta);
+}
+
+template <class W1, class W2>
+inline std::ostream &operator<<(std::ostream &strm,
+                                const PairWeight<W1, W2> &weight) {
+  CompositeWeightWriter writer(strm);
+  writer.WriteBegin();
+  writer.WriteElement(weight.Value1());
+  writer.WriteElement(weight.Value2());
+  writer.WriteEnd();
+  return strm;
+}
+
+template <class W1, class W2>
+inline std::istream &operator>>(std::istream &strm,
+                                PairWeight<W1, W2> &weight) {
+  CompositeWeightReader reader(strm);
+  reader.ReadBegin();
+  W1 w1;
+  reader.ReadElement(&w1);
+  weight.SetValue1(w1);
+  W2 w2;
+  reader.ReadElement(&w2, true);
+  weight.SetValue2(w2);
+  reader.ReadEnd();
+  return strm;
+}
+
+// This function object returns weights by calling the underlying generators
+// and forming a pair. This is intended primarily for testing.
+template <class W1, class W2>
+class WeightGenerate<PairWeight<W1, W2>> {
+ public:
+  using Weight = PairWeight<W1, W2>;
+  using Generate1 = WeightGenerate<W1>;
+  using Generate2 = WeightGenerate<W2>;
+
+  explicit WeightGenerate(bool allow_zero = true)
+      : generate1_(allow_zero), generate2_(allow_zero) {}
+
+  Weight operator()() const { return Weight(generate1_(), generate2_()); }
+
+ private:
+  Generate1 generate1_;
+  Generate2 generate2_;
+};
+
+}  // namespace fst
+
+#endif  // FST_PAIR_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/partition.h b/projects/llm_framework/include/fst/partition.h
new file mode 100644
index 00000000..5dbbe46a
--- /dev/null
+++ b/projects/llm_framework/include/fst/partition.h
@@ -0,0 +1,305 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes to create a partition of states.
+
+#ifndef FST_PARTITION_H_
+#define FST_PARTITION_H_
+
+#include <algorithm>
+#include <vector>
+
+
+#include <fst/queue.h>
+
+
+namespace fst {
+namespace internal {
+
+template <typename T>
+class PartitionIterator;
+
+// Defines a partitioning of elements, used to represent equivalence classes
+// for FST operations like minimization. T must be a signed integer type.
+//
+// The elements are numbered from 0 to num_elements - 1.
+// Initialize(num_elements) sets up the class for a given number of elements.
+// We maintain a partition of these elements into classes. The classes are also
+// numbered from zero; you can add a class with AddClass(), or add them in bulk
+// with AllocateClasses(num_classes). Initially the elements are not assigned
+// to any class; you set up the initial mapping from elements to classes by
+// calling Add(element_id, class_id). You can also move an element to a
+// different class by calling Move(element_id, class_id).
+//
+// We also support a rather specialized interface that allows you to efficiently
+// split classes in the Hopcroft minimization algorithm. This maintains a
+// binary partition of each class.  Let's call these, rather arbitrarily, the
+// 'yes' subset and the 'no' subset of each class, and assume that by default,
+// each element of a class is in its 'no' subset. When one calls
+// SplitOn(element_id), element_id is moved to the 'yes' subset of its class.
+// (If it was already in the 'yes' set, it just stays there). The aim is to
+// enable (later) splitting the class in two in time no greater than the time
+// already spent calling SplitOn() for that class. We keep a list of the classes
+// which have nonempty 'yes' sets, as visited_classes_. When one calls
+// FinalizeSplit(Queue *l), for each class in visited_classes_ whose 'yes'
+// and 'no' sets are both nonempty, it will create a new class consisting of
+// the smaller of the two subsets (and this class will be added to the queue),
+// and the old class will now be the larger of the two subsets. This call also
+// resets all the yes/no partitions so that everything is in the 'no' subsets.
+//
+// One cannot use the Move() function if SplitOn() has been called without
+// a subsequent call to FinalizeSplit()
+template <typename T>
+class Partition {
+ public:
+  Partition() {}
+
+  explicit Partition(T num_elements) { Initialize(num_elements); }
+
+  // Creates an empty partition for num_elements. This means that the elements
+  // are not assigned to a class (i.e class_index = -1); you should set up the
+  // number of classes using AllocateClasses() or AddClass(), and allocate each
+  // element to a class by calling Add(element, class_id).
+  void Initialize(size_t num_elements) {
+    elements_.resize(num_elements);
+    classes_.reserve(num_elements);
+    classes_.clear();
+    yes_counter_ = 1;
+  }
+
+  // Adds a class; returns new number of classes.
+  T AddClass() {
+    auto num_classes = classes_.size();
+    classes_.resize(num_classes + 1);
+    return num_classes;
+  }
+
+  // Adds 'num_classes' new (empty) classes.
+  void AllocateClasses(T num_classes) {
+    classes_.resize(classes_.size() + num_classes);
+  }
+
+  // Adds element_id to class_id. element_id should already have been allocated
+  // by calling Initialize(num_elements)---or the constructor taking
+  // num_elements---with num_elements > element_id. element_id must not
+  // currently be a member of any class; once elements have been added to a
+  // class, use the Move() method to move them from one class to another.
+  void Add(T element_id, T class_id) {
+    auto &this_element = elements_[element_id];
+    auto &this_class = classes_[class_id];
+    ++this_class.size;
+    // Adds the element to the 'no' subset of the class.
+    auto no_head = this_class.no_head;
+    if (no_head >= 0) elements_[no_head].prev_element = element_id;
+    this_class.no_head = element_id;
+    this_element.class_id = class_id;
+    // Adds to the 'no' subset of the class.
+    this_element.yes = 0;
+    this_element.next_element = no_head;
+    this_element.prev_element = -1;
+  }
+
+  // Moves element_id from 'no' subset of its current class to 'no' subset of
+  // class class_id. This may not work correctly if you have called SplitOn()
+  // [for any element] and haven't subsequently called FinalizeSplit().
+  void Move(T element_id, T class_id) {
+    auto elements = &(elements_[0]);
+    auto &element = elements[element_id];
+    auto &old_class = classes_[element.class_id];
+    --old_class.size;
+    // Excises the element from the 'no' list of its old class, where it is
+    // assumed to be.
+    if (element.prev_element >= 0) {
+      elements[element.prev_element].next_element = element.next_element;
+    } else {
+      old_class.no_head = element.next_element;
+    }
+    if (element.next_element >= 0) {
+      elements[element.next_element].prev_element = element.prev_element;
+    }
+    // Adds to new class.
+    Add(element_id, class_id);
+  }
+
+  // Moves element_id to the 'yes' subset of its class if it was in the 'no'
+  // subset, and marks the class as having been visited.
+  void SplitOn(T element_id) {
+    auto elements = &(elements_[0]);
+    auto &element = elements[element_id];
+    if (element.yes == yes_counter_) {
+      return;  // Already in the 'yes' set; nothing to do.
+    }
+    auto class_id = element.class_id;
+    auto &this_class = classes_[class_id];
+    // Excises the element from the 'no' list of its class.
+    if (element.prev_element >= 0) {
+      elements[element.prev_element].next_element = element.next_element;
+    } else {
+      this_class.no_head = element.next_element;
+    }
+    if (element.next_element >= 0) {
+      elements[element.next_element].prev_element = element.prev_element;
+    }
+    // Adds the element to the 'yes' list.
+    if (this_class.yes_head >= 0) {
+      elements[this_class.yes_head].prev_element = element_id;
+    } else {
+      visited_classes_.push_back(class_id);
+    }
+    element.yes = yes_counter_;
+    element.next_element = this_class.yes_head;
+    element.prev_element = -1;
+    this_class.yes_head = element_id;
+    this_class.yes_size++;
+  }
+
+  // This should be called after one has possibly called SplitOn for one or more
+  // elements, thus moving those elements to the 'yes' subset for their class.
+  // For each class that has a nontrivial split (i.e., it's not the case that
+  // all members are in the 'yes' or 'no' subset), this function creates a new
+  // class containing the smaller of the two subsets of elements, leaving the
+  // larger group of elements in the old class. The identifier of the new class
+  // will be added to the queue provided as the pointer L. This method then
+  // moves all elements to the 'no' subset of their class.
+  template <class Queue>
+  void FinalizeSplit(Queue *queue) {
+    for (const auto &visited_class : visited_classes_) {
+      const auto new_class = SplitRefine(visited_class);
+      if (new_class != -1 && queue) queue->Enqueue(new_class);
+    }
+    visited_classes_.clear();
+    // Incrementation sets all the 'yes' members of the elements to false.
+    ++yes_counter_;
+  }
+
+  const T ClassId(T element_id) const { return elements_[element_id].class_id; }
+
+  const size_t ClassSize(T class_id) const { return classes_[class_id].size; }
+
+  const T NumClasses() const { return classes_.size(); }
+
+ private:
+  friend class PartitionIterator<T>;
+
+  // Information about a given element.
+  struct Element {
+    T class_id;      // Class ID of this element.
+    T yes;           // This is to be interpreted as a bool, true if it's in the
+                     // 'yes' set of this class. The interpretation as bool is
+                     // (yes == yes_counter_ ? true : false).
+    T next_element;  // Next element in the 'no' list or 'yes' list of this
+                     // class, whichever of the two we belong to (think of
+                     // this as the 'next' in a doubly-linked list, although
+                     // it is an index into the elements array). Negative
+                     // values corresponds to null.
+    T prev_element;  // Previous element in the 'no' or 'yes' doubly linked
+                     // list. Negative values corresponds to null.
+  };
+
+  // Information about a given class.
+  struct Class {
+    Class() : size(0), yes_size(0), no_head(-1), yes_head(-1) {}
+    T size;      // Total number of elements in this class ('no' plus 'yes'
+                 // subsets).
+    T yes_size;  // Total number of elements of 'yes' subset of this class.
+    T no_head;   // Index of head element of doubly-linked list in 'no' subset.
+                 // Everything is in the 'no' subset until you call SplitOn().
+                 // -1 means no element.
+    T yes_head;  // Index of head element of doubly-linked list in 'yes' subset.
+                 // -1 means no element.
+  };
+
+  // This method, called from FinalizeSplit(), checks whether a class has to
+  // be split (a class will be split only if its 'yes' and 'no' subsets are
+  // both nonempty, but one can assume that since this function was called, the
+  // 'yes' subset is nonempty). It splits by taking the smaller subset and
+  // making it a new class, and leaving the larger subset of elements in the
+  // 'no' subset of the old class. It returns the new class if created, or -1
+  // if none was created.
+  T SplitRefine(T class_id) {
+    auto yes_size = classes_[class_id].yes_size;
+    auto size = classes_[class_id].size;
+    auto no_size = size - yes_size;
+    if (no_size == 0) {
+      // All members are in the 'yes' subset, so we don't have to create a new
+      // class, just move them all to the 'no' subset.
+      classes_[class_id].no_head = classes_[class_id].yes_head;
+      classes_[class_id].yes_head = -1;
+      classes_[class_id].yes_size = 0;
+      return -1;
+    } else {
+      auto new_class_id = classes_.size();
+      classes_.resize(classes_.size() + 1);
+      auto &old_class = classes_[class_id];
+      auto &new_class = classes_[new_class_id];
+      // The new_class will have the values from the constructor.
+      if (no_size < yes_size) {
+        // Moves the 'no' subset to new class ('no' subset).
+        new_class.no_head = old_class.no_head;
+        new_class.size = no_size;
+        // And makes the 'yes' subset of the old class ('no' subset).
+        old_class.no_head = old_class.yes_head;
+        old_class.yes_head = -1;
+        old_class.size = yes_size;
+        old_class.yes_size = 0;
+      } else {
+        // Moves the 'yes' subset to the new class (to the 'no' subset)
+        new_class.size = yes_size;
+        new_class.no_head = old_class.yes_head;
+        // Retains only the 'no' subset in the old class.
+        old_class.size = no_size;
+        old_class.yes_size = 0;
+        old_class.yes_head = -1;
+      }
+      auto elements = &(elements_[0]);
+      // Updates the 'class_id' of all the elements we moved.
+      for (auto e = new_class.no_head; e >= 0; e = elements[e].next_element) {
+        elements[e].class_id = new_class_id;
+      }
+      return new_class_id;
+    }
+  }
+
+  // elements_[i] contains all info about the i'th element.
+  std::vector<Element> elements_;
+  // classes_[i] contains all info about the i'th class.
+  std::vector<Class> classes_;
+  // Set of visited classes to be used in split refine.
+  std::vector<T> visited_classes_;
+  // yes_counter_ is used in interpreting the 'yes' members of class Element.
+  // If element.yes == yes_counter_, we interpret that element as being in the
+  // 'yes' subset of its class. This allows us to, in effect, set all those
+  // bools to false at a stroke by incrementing yes_counter_.
+  T yes_counter_;
+};
+
+// Iterates over members of the 'no' subset of a class in a partition. (When
+// this is used, everything is in the 'no' subset).
+template <typename T>
+class PartitionIterator {
+ public:
+  using Element = typename Partition<T>::Element;
+
+  PartitionIterator(const Partition<T> &partition, T class_id)
+      : partition_(partition),
+        element_id_(partition_.classes_[class_id].no_head),
+        class_id_(class_id) {}
+
+  bool Done() { return element_id_ < 0; }
+
+  const T Value() { return element_id_; }
+
+  void Next() { element_id_ = partition_.elements_[element_id_].next_element; }
+
+  void Reset() { element_id_ = partition_.classes_[class_id_].no_head; }
+
+ private:
+  const Partition<T> &partition_;
+  T element_id_;
+  T class_id_;
+};
+
+}  // namespace internal
+}  // namespace fst
+
+#endif  // FST_PARTITION_H_
diff --git a/projects/llm_framework/include/fst/power-weight.h b/projects/llm_framework/include/fst/power-weight.h
new file mode 100644
index 00000000..f2f3cbdb
--- /dev/null
+++ b/projects/llm_framework/include/fst/power-weight.h
@@ -0,0 +1,168 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Cartesian power weight semiring operation definitions.
+
+#ifndef FST_POWER_WEIGHT_H_
+#define FST_POWER_WEIGHT_H_
+
+#include <string>
+
+#include <fst/tuple-weight.h>
+#include <fst/weight.h>
+
+
+namespace fst {
+
+// Cartesian power semiring: W ^ n
+//
+// Forms:
+//  - a left semimodule when W is a left semiring,
+//  - a right semimodule when W is a right semiring,
+//  - a bisemimodule when W is a semiring,
+//    the free semimodule of rank n over W
+// The Times operation is overloaded to provide the left and right scalar
+// products.
+template <class W, size_t n>
+class PowerWeight : public TupleWeight<W, n> {
+ public:
+  using ReverseWeight = PowerWeight<typename W::ReverseWeight, n>;
+
+  PowerWeight() {}
+
+  explicit PowerWeight(const TupleWeight<W, n> &weight)
+      : TupleWeight<W, n>(weight) {}
+
+  template <class Iterator>
+  PowerWeight(Iterator begin, Iterator end) : TupleWeight<W, n>(begin, end) {}
+
+  // Initialize component `index` to `weight`; initialize all other components
+  // to `default_weight`
+  PowerWeight(size_t index, const W &weight,
+              const W &default_weight = W::Zero())
+      : TupleWeight<W, n>(index, weight, default_weight) {}
+
+  static const PowerWeight &Zero() {
+    static const PowerWeight zero(TupleWeight<W, n>::Zero());
+    return zero;
+  }
+
+  static const PowerWeight &One() {
+    static const PowerWeight one(TupleWeight<W, n>::One());
+    return one;
+  }
+
+  static const PowerWeight &NoWeight() {
+    static const PowerWeight no_weight(TupleWeight<W, n>::NoWeight());
+    return no_weight;
+  }
+
+  static const string &Type() {
+    static const string *const type =
+        new string(W::Type() + "_^" + std::to_string(n));
+    return *type;
+  }
+
+  static constexpr uint64 Properties() {
+    return W::Properties() &
+           (kLeftSemiring | kRightSemiring | kCommutative | kIdempotent);
+  }
+
+  PowerWeight Quantize(float delta = kDelta) const {
+    return PowerWeight(TupleWeight<W, n>::Quantize(delta));
+  }
+
+  ReverseWeight Reverse() const {
+    return ReverseWeight(TupleWeight<W, n>::Reverse());
+  }
+};
+
+// Semiring plus operation.
+template <class W, size_t n>
+inline PowerWeight<W, n> Plus(const PowerWeight<W, n> &w1,
+                              const PowerWeight<W, n> &w2) {
+  PowerWeight<W, n> result;
+  for (size_t i = 0; i < n; ++i) {
+    result.SetValue(i, Plus(w1.Value(i), w2.Value(i)));
+  }
+  return result;
+}
+
+// Semiring times operation.
+template <class W, size_t n>
+inline PowerWeight<W, n> Times(const PowerWeight<W, n> &w1,
+                               const PowerWeight<W, n> &w2) {
+  PowerWeight<W, n> result;
+  for (size_t i = 0; i < n; ++i) {
+    result.SetValue(i, Times(w1.Value(i), w2.Value(i)));
+  }
+  return result;
+}
+
+// Semiring divide operation.
+template <class W, size_t n>
+inline PowerWeight<W, n> Divide(const PowerWeight<W, n> &w1,
+                                const PowerWeight<W, n> &w2,
+                                DivideType type = DIVIDE_ANY) {
+  PowerWeight<W, n> result;
+  for (size_t i = 0; i < n; ++i) {
+    result.SetValue(i, Divide(w1.Value(i), w2.Value(i), type));
+  }
+  return result;
+}
+
+// Semimodule left scalar product.
+template <class W, size_t n>
+inline PowerWeight<W, n> Times(const W &scalar,
+                               const PowerWeight<W, n> &weight) {
+  PowerWeight<W, n> result;
+  for (size_t i = 0; i < n; ++i) {
+    result.SetValue(i, Times(scalar, weight.Value(i)));
+  }
+  return result;
+}
+
+// Semimodule right scalar product.
+template <class W, size_t n>
+inline PowerWeight<W, n> Times(const PowerWeight<W, n> &weight,
+                               const W &scalar) {
+  PowerWeight<W, n> result;
+  for (size_t i = 0; i < n; ++i) {
+    result.SetValue(i, Times(weight.Value(i), scalar));
+  }
+  return result;
+}
+
+// Semimodule dot product.
+template <class W, size_t n>
+inline W DotProduct(const PowerWeight<W, n> &w1, const PowerWeight<W, n> &w2) {
+  W result(W::Zero());
+  for (size_t i = 0; i < n; ++i) {
+    result = Plus(result, Times(w1.Value(i), w2.Value(i)));
+  }
+  return result;
+}
+
+// This function object generates weights over the Cartesian power of rank
+// n over the underlying weight. This is intended primarily for testing.
+template <class W, size_t n>
+class WeightGenerate<PowerWeight<W, n>> {
+ public:
+  using Weight = PowerWeight<W, n>;
+  using Generate = WeightGenerate<W>;
+
+  explicit WeightGenerate(bool allow_zero = true) : generate_(allow_zero) {}
+
+  Weight operator()() const {
+    Weight result;
+    for (size_t i = 0; i < n; ++i) result.SetValue(i, generate_());
+    return result;
+  }
+
+ private:
+  Generate generate_;
+};
+
+}  // namespace fst
+
+#endif  // FST_POWER_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/product-weight.h b/projects/llm_framework/include/fst/product-weight.h
new file mode 100644
index 00000000..56a18be1
--- /dev/null
+++ b/projects/llm_framework/include/fst/product-weight.h
@@ -0,0 +1,107 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Product weight set and associated semiring operation definitions.
+
+#ifndef FST_PRODUCT_WEIGHT_H_
+#define FST_PRODUCT_WEIGHT_H_
+
+#include <string>
+#include <utility>
+
+#include <fst/pair-weight.h>
+#include <fst/weight.h>
+
+
+namespace fst {
+
+// Product semiring: W1 * W2.
+template <class W1, class W2>
+class ProductWeight : public PairWeight<W1, W2> {
+ public:
+  using ReverseWeight =
+      ProductWeight<typename W1::ReverseWeight, typename W2::ReverseWeight>;
+
+  ProductWeight() {}
+
+  explicit ProductWeight(const PairWeight<W1, W2> &weight)
+      : PairWeight<W1, W2>(weight) {}
+
+  ProductWeight(W1 w1, W2 w2)
+      : PairWeight<W1, W2>(std::move(w1), std::move(w2)) {}
+
+  static const ProductWeight &Zero() {
+    static const ProductWeight zero(PairWeight<W1, W2>::Zero());
+    return zero;
+  }
+
+  static const ProductWeight &One() {
+    static const ProductWeight one(PairWeight<W1, W2>::One());
+    return one;
+  }
+
+  static const ProductWeight &NoWeight() {
+    static const ProductWeight no_weight(PairWeight<W1, W2>::NoWeight());
+    return no_weight;
+  }
+
+  static const string &Type() {
+    static const string *const type =
+        new string(W1::Type() + "_X_" + W2::Type());
+    return *type;
+  }
+
+  static constexpr uint64 Properties() {
+    return W1::Properties() & W2::Properties() &
+           (kLeftSemiring | kRightSemiring | kCommutative | kIdempotent);
+  }
+
+  ProductWeight Quantize(float delta = kDelta) const {
+    return ProductWeight(PairWeight<W1, W2>::Quantize(delta));
+  }
+
+  ReverseWeight Reverse() const {
+    return ReverseWeight(PairWeight<W1, W2>::Reverse());
+  }
+};
+
+template <class W1, class W2>
+inline ProductWeight<W1, W2> Plus(const ProductWeight<W1, W2> &w1,
+                                  const ProductWeight<W1, W2> &w2) {
+  return ProductWeight<W1, W2>(Plus(w1.Value1(), w2.Value1()),
+                               Plus(w1.Value2(), w2.Value2()));
+}
+
+template <class W1, class W2>
+inline ProductWeight<W1, W2> Times(const ProductWeight<W1, W2> &w1,
+                                   const ProductWeight<W1, W2> &w2) {
+  return ProductWeight<W1, W2>(Times(w1.Value1(), w2.Value1()),
+                               Times(w1.Value2(), w2.Value2()));
+}
+
+template <class W1, class W2>
+inline ProductWeight<W1, W2> Divide(const ProductWeight<W1, W2> &w1,
+                                    const ProductWeight<W1, W2> &w2,
+                                    DivideType typ = DIVIDE_ANY) {
+  return ProductWeight<W1, W2>(Divide(w1.Value1(), w2.Value1(), typ),
+                               Divide(w1.Value2(), w2.Value2(), typ));
+}
+
+// This function object generates weights by calling the underlying generators
+// for the template weight types, like all other pair weight types. This is
+// intended primarily for testing.
+template <class W1, class W2>
+class WeightGenerate<ProductWeight<W1, W2>> :
+    public WeightGenerate<PairWeight<W1, W2>> {
+ public:
+  using Weight = ProductWeight<W1, W2>;
+  using Generate = WeightGenerate<PairWeight<W1, W2>>;
+
+  explicit WeightGenerate(bool allow_zero = true) : Generate(allow_zero) {}
+
+  Weight operator()() const { return Weight(Generate::operator()()); }
+};
+
+}  // namespace fst
+
+#endif  // FST_PRODUCT_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/project.h b/projects/llm_framework/include/fst/project.h
new file mode 100644
index 00000000..5a82cf14
--- /dev/null
+++ b/projects/llm_framework/include/fst/project.h
@@ -0,0 +1,159 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes to project an FST on to its domain or range.
+
+#ifndef FST_PROJECT_H_
+#define FST_PROJECT_H_
+
+#include <fst/arc-map.h>
+#include <fst/mutable-fst.h>
+
+
+namespace fst {
+
+// This specifies whether to project on input or output.
+enum ProjectType { PROJECT_INPUT = 1, PROJECT_OUTPUT = 2 };
+
+// Mapper to implement projection per arc.
+template <class A>
+class ProjectMapper {
+ public:
+  using FromArc = A;
+  using ToArc = A;
+
+  constexpr explicit ProjectMapper(ProjectType project_type)
+      : project_type_(project_type) {}
+
+  ToArc operator()(const FromArc &arc) const {
+    const auto label = project_type_ == PROJECT_INPUT ? arc.ilabel : arc.olabel;
+    return ToArc(label, label, arc.weight, arc.nextstate);
+  }
+
+  constexpr MapFinalAction FinalAction() const {
+    return MAP_NO_SUPERFINAL;
+  }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return project_type_ == PROJECT_INPUT ? MAP_COPY_SYMBOLS
+                                          : MAP_CLEAR_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return project_type_ == PROJECT_OUTPUT ? MAP_COPY_SYMBOLS
+                                           : MAP_CLEAR_SYMBOLS;
+  }
+
+  constexpr uint64 Properties(uint64 props) const {
+    return ProjectProperties(props, project_type_ == PROJECT_INPUT);
+  }
+
+ private:
+  const ProjectType project_type_;
+};
+
+// Projects an FST onto its domain or range by either copying each arcs' input
+// label to the output label or vice versa.
+//
+// Complexity:
+//
+//   Time: O(V + E)
+//   Space: O(1)
+//
+// where V is the number of states and E is the number of arcs.
+template <class Arc>
+inline void Project(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
+                    ProjectType project_type) {
+  ArcMap(ifst, ofst, ProjectMapper<Arc>(project_type));
+  switch (project_type) {
+    case PROJECT_INPUT:
+      ofst->SetOutputSymbols(ifst.InputSymbols());
+      return;
+    case PROJECT_OUTPUT:
+      ofst->SetInputSymbols(ifst.OutputSymbols());
+      return;
+  }
+}
+
+// Destructive variant of the above.
+template <class Arc>
+inline void Project(MutableFst<Arc> *fst, ProjectType project_type) {
+  ArcMap(fst, ProjectMapper<Arc>(project_type));
+  switch (project_type) {
+    case PROJECT_INPUT:
+      fst->SetOutputSymbols(fst->InputSymbols());
+      return;
+    case PROJECT_OUTPUT:
+      fst->SetInputSymbols(fst->OutputSymbols());
+      return;
+  }
+}
+
+// Projects an FST onto its domain or range by either copying each arc's input
+// label to the output label or vice versa. This version is a delayed FST.
+//
+// Complexity:
+//
+//   Time: O(v + e)
+//   Space: O(1)
+//
+// where v is the number of states visited and e is the number of arcs visited.
+// Constant time and to visit an input state or arc is assumed and exclusive of
+// caching.
+template <class A>
+class ProjectFst : public ArcMapFst<A, A, ProjectMapper<A>> {
+ public:
+  using FromArc = A;
+  using ToArc = A;
+
+  using Impl = internal::ArcMapFstImpl<A, A, ProjectMapper<A>>;
+
+  ProjectFst(const Fst<A> &fst, ProjectType project_type)
+      : ArcMapFst<A, A, ProjectMapper<A>>(fst, ProjectMapper<A>(project_type)) {
+    if (project_type == PROJECT_INPUT) {
+      GetMutableImpl()->SetOutputSymbols(fst.InputSymbols());
+    }
+    if (project_type == PROJECT_OUTPUT) {
+      GetMutableImpl()->SetInputSymbols(fst.OutputSymbols());
+    }
+  }
+
+  // See Fst<>::Copy() for doc.
+  ProjectFst(const ProjectFst<A> &fst, bool safe = false)
+      : ArcMapFst<A, A, ProjectMapper<A>>(fst, safe) {}
+
+  // Gets a copy of this ProjectFst. See Fst<>::Copy() for further doc.
+  ProjectFst<A> *Copy(bool safe = false) const override {
+    return new ProjectFst(*this, safe);
+  }
+
+ private:
+  using ImplToFst<Impl>::GetMutableImpl;
+};
+
+// Specialization for ProjectFst.
+template <class A>
+class StateIterator<ProjectFst<A>>
+    : public StateIterator<ArcMapFst<A, A, ProjectMapper<A>>> {
+ public:
+  explicit StateIterator(const ProjectFst<A> &fst)
+      : StateIterator<ArcMapFst<A, A, ProjectMapper<A>>>(fst) {}
+};
+
+// Specialization for ProjectFst.
+template <class A>
+class ArcIterator<ProjectFst<A>>
+    : public ArcIterator<ArcMapFst<A, A, ProjectMapper<A>>> {
+ public:
+  using StateId = typename A::StateId;
+
+  ArcIterator(const ProjectFst<A> &fst, StateId s)
+      : ArcIterator<ArcMapFst<A, A, ProjectMapper<A>>>(fst, s) {}
+};
+
+// Useful alias when using StdArc.
+using StdProjectFst = ProjectFst<StdArc>;
+
+}  // namespace fst
+
+#endif  // FST_PROJECT_H_
diff --git a/projects/llm_framework/include/fst/properties.h b/projects/llm_framework/include/fst/properties.h
new file mode 100644
index 00000000..157247a6
--- /dev/null
+++ b/projects/llm_framework/include/fst/properties.h
@@ -0,0 +1,468 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// FST property bits.
+
+#ifndef FST_PROPERTIES_H_
+#define FST_PROPERTIES_H_
+
+#include <sys/types.h>
+#include <vector>
+
+#include <fst/compat.h>
+
+namespace fst {
+
+// The property bits here assert facts about an FST. If individual bits are
+// added, then the composite properties below, the property functions and
+// property names in properties.cc, and TestProperties() in test-properties.h
+// should be updated.
+
+// BINARY PROPERTIES
+//
+// For each property below, there is a single bit. If it is set, the property is
+// true. If it is not set, the property is false.
+
+// The Fst is an ExpandedFst.
+constexpr uint64 kExpanded = 0x0000000000000001ULL;
+
+// The Fst is a MutableFst.
+constexpr uint64 kMutable = 0x0000000000000002ULL;
+
+// An error was detected while constructing/using the FST.
+constexpr uint64 kError = 0x0000000000000004ULL;
+
+// TRINARY PROPERTIES
+//
+// For each of these properties below there is a pair of property bits, one
+// positive and one negative. If the positive bit is set, the property is true.
+// If the negative bit is set, the property is false. If neither is set, the
+// property has unknown value. Both should never be simultaneously set. The
+// individual positive and negative bit pairs should be adjacent with the
+// positive bit at an odd and lower position.
+
+// ilabel == olabel for each arc.
+constexpr uint64 kAcceptor = 0x0000000000010000ULL;
+// ilabel != olabel for some arc.
+constexpr uint64 kNotAcceptor = 0x0000000000020000ULL;
+
+// ilabels unique leaving each state.
+constexpr uint64 kIDeterministic = 0x0000000000040000ULL;
+// ilabels not unique leaving some state.
+constexpr uint64 kNonIDeterministic = 0x0000000000080000ULL;
+
+// olabels unique leaving each state.
+constexpr uint64 kODeterministic = 0x0000000000100000ULL;
+// olabels not unique leaving some state.
+constexpr uint64 kNonODeterministic = 0x0000000000200000ULL;
+
+// FST has input/output epsilons.
+constexpr uint64 kEpsilons = 0x0000000000400000ULL;
+// FST has no input/output epsilons.
+constexpr uint64 kNoEpsilons = 0x0000000000800000ULL;
+
+// FST has input epsilons.
+constexpr uint64 kIEpsilons = 0x0000000001000000ULL;
+// FST has no input epsilons.
+constexpr uint64 kNoIEpsilons = 0x0000000002000000ULL;
+
+// FST has output epsilons.
+constexpr uint64 kOEpsilons = 0x0000000004000000ULL;
+// FST has no output epsilons.
+constexpr uint64 kNoOEpsilons = 0x0000000008000000ULL;
+
+// ilabels sorted wrt < for each state.
+constexpr uint64 kILabelSorted = 0x0000000010000000ULL;
+// ilabels not sorted wrt < for some state.
+constexpr uint64 kNotILabelSorted = 0x0000000020000000ULL;
+
+// olabels sorted wrt < for each state.
+constexpr uint64 kOLabelSorted = 0x0000000040000000ULL;
+// olabels not sorted wrt < for some state.
+constexpr uint64 kNotOLabelSorted = 0x0000000080000000ULL;
+
+// Non-trivial arc or final weights.
+constexpr uint64 kWeighted = 0x0000000100000000ULL;
+// Only trivial arc and final weights.
+constexpr uint64 kUnweighted = 0x0000000200000000ULL;
+
+// FST has cycles.
+constexpr uint64 kCyclic = 0x0000000400000000ULL;
+// FST has no cycles.
+constexpr uint64 kAcyclic = 0x0000000800000000ULL;
+
+// FST has cycles containing the initial state.
+constexpr uint64 kInitialCyclic = 0x0000001000000000ULL;
+// FST has no cycles containing the initial state.
+constexpr uint64 kInitialAcyclic = 0x0000002000000000ULL;
+
+// FST is topologically sorted.
+constexpr uint64 kTopSorted = 0x0000004000000000ULL;
+// FST is not topologically sorted.
+constexpr uint64 kNotTopSorted = 0x0000008000000000ULL;
+
+// All states reachable from the initial state.
+constexpr uint64 kAccessible = 0x0000010000000000ULL;
+// Not all states reachable from the initial state.
+constexpr uint64 kNotAccessible = 0x0000020000000000ULL;
+
+// All states can reach a final state.
+constexpr uint64 kCoAccessible = 0x0000040000000000ULL;
+// Not all states can reach a final state.
+constexpr uint64 kNotCoAccessible = 0x0000080000000000ULL;
+
+// If NumStates() > 0, then state 0 is initial, state NumStates() - 1 is final,
+// there is a transition from each non-final state i to state i + 1, and there
+// are no other transitions.
+constexpr uint64 kString = 0x0000100000000000ULL;
+
+// Not a string FST.
+constexpr uint64 kNotString = 0x0000200000000000ULL;
+
+// FST has least one weighted cycle.
+constexpr uint64 kWeightedCycles = 0x0000400000000000ULL;
+
+// Only unweighted cycles.
+constexpr uint64 kUnweightedCycles = 0x0000800000000000ULL;
+
+// COMPOSITE PROPERTIES
+
+// Properties of an empty machine.
+constexpr uint64 kNullProperties =
+    kAcceptor | kIDeterministic | kODeterministic | kNoEpsilons | kNoIEpsilons |
+    kNoOEpsilons | kILabelSorted | kOLabelSorted | kUnweighted | kAcyclic |
+    kInitialAcyclic | kTopSorted | kAccessible | kCoAccessible | kString |
+    kUnweightedCycles;
+
+// Properties that are preserved when an FST is copied.
+constexpr uint64 kCopyProperties =
+    kError | kAcceptor | kNotAcceptor | kIDeterministic | kNonIDeterministic |
+    kODeterministic | kNonODeterministic | kEpsilons | kNoEpsilons |
+    kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons | kILabelSorted |
+    kNotILabelSorted | kOLabelSorted | kNotOLabelSorted | kWeighted |
+    kUnweighted | kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic |
+    kTopSorted | kNotTopSorted | kAccessible | kNotAccessible | kCoAccessible |
+    kNotCoAccessible | kString | kNotString | kWeightedCycles |
+    kUnweightedCycles;
+
+// Properties that are intrinsic to the FST.
+constexpr uint64 kIntrinsicProperties =
+    kExpanded | kMutable | kAcceptor | kNotAcceptor | kIDeterministic |
+    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
+    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
+    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
+    kWeighted | kUnweighted | kCyclic | kAcyclic | kInitialCyclic |
+    kInitialAcyclic | kTopSorted | kNotTopSorted | kAccessible |
+    kNotAccessible | kCoAccessible | kNotCoAccessible | kString | kNotString |
+    kWeightedCycles | kUnweightedCycles;
+
+// Properties that are (potentially) extrinsic to the FST.
+constexpr uint64 kExtrinsicProperties = kError;
+
+// Properties that are preserved when an FST start state is set.
+constexpr uint64 kSetStartProperties =
+    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
+    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
+    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
+    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
+    kWeighted | kUnweighted | kCyclic | kAcyclic | kTopSorted | kNotTopSorted |
+    kCoAccessible | kNotCoAccessible | kWeightedCycles | kUnweightedCycles;
+
+// Properties that are preserved when an FST final weight is set.
+constexpr uint64 kSetFinalProperties =
+    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
+    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
+    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
+    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
+    kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic | kTopSorted |
+    kNotTopSorted | kAccessible | kNotAccessible | kWeightedCycles |
+    kUnweightedCycles;
+
+// Properties that are preserved when an FST state is added.
+constexpr uint64 kAddStateProperties =
+    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
+    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
+    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
+    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
+    kWeighted | kUnweighted | kCyclic | kAcyclic | kInitialCyclic |
+    kInitialAcyclic | kTopSorted | kNotTopSorted | kNotAccessible |
+    kNotCoAccessible | kNotString | kWeightedCycles | kUnweightedCycles;
+
+// Properties that are preserved when an FST arc is added.
+constexpr uint64 kAddArcProperties =
+    kExpanded | kMutable | kError | kNotAcceptor | kNonIDeterministic |
+    kNonODeterministic | kEpsilons | kIEpsilons | kOEpsilons |
+    kNotILabelSorted | kNotOLabelSorted | kWeighted | kCyclic | kInitialCyclic |
+    kNotTopSorted | kAccessible | kCoAccessible | kWeightedCycles;
+
+// Properties that are preserved when an FST arc is set.
+constexpr uint64 kSetArcProperties = kExpanded | kMutable | kError;
+
+// Properties that are preserved when FST states are deleted.
+constexpr uint64 kDeleteStatesProperties =
+    kExpanded | kMutable | kError | kAcceptor | kIDeterministic |
+    kODeterministic | kNoEpsilons | kNoIEpsilons | kNoOEpsilons |
+    kILabelSorted | kOLabelSorted | kUnweighted | kAcyclic | kInitialAcyclic |
+    kTopSorted | kUnweightedCycles;
+
+// Properties that are preserved when FST arcs are deleted.
+constexpr uint64 kDeleteArcsProperties =
+    kExpanded | kMutable | kError | kAcceptor | kIDeterministic |
+    kODeterministic | kNoEpsilons | kNoIEpsilons | kNoOEpsilons |
+    kILabelSorted | kOLabelSorted | kUnweighted | kAcyclic | kInitialAcyclic |
+    kTopSorted | kNotAccessible | kNotCoAccessible | kUnweightedCycles;
+
+// Properties that are preserved when an FST's states are reordered.
+constexpr uint64 kStateSortProperties =
+    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
+    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
+    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
+    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
+    kWeighted | kUnweighted | kCyclic | kAcyclic | kInitialCyclic |
+    kInitialAcyclic | kAccessible | kNotAccessible | kCoAccessible |
+    kNotCoAccessible | kWeightedCycles | kUnweightedCycles;
+
+// Properties that are preserved when an FST's arcs are reordered.
+constexpr uint64 kArcSortProperties =
+    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
+    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
+    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
+    kWeighted | kUnweighted | kCyclic | kAcyclic | kInitialCyclic |
+    kInitialAcyclic | kTopSorted | kNotTopSorted | kAccessible |
+    kNotAccessible | kCoAccessible | kNotCoAccessible | kString | kNotString |
+    kWeightedCycles | kUnweightedCycles;
+
+// Properties that are preserved when an FST's input labels are changed.
+constexpr uint64 kILabelInvariantProperties =
+    kExpanded | kMutable | kError | kODeterministic | kNonODeterministic |
+    kOEpsilons | kNoOEpsilons | kOLabelSorted | kNotOLabelSorted | kWeighted |
+    kUnweighted | kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic |
+    kTopSorted | kNotTopSorted | kAccessible | kNotAccessible | kCoAccessible |
+    kNotCoAccessible | kString | kNotString | kWeightedCycles |
+    kUnweightedCycles;
+
+// Properties that are preserved when an FST's output labels are changed.
+constexpr uint64 kOLabelInvariantProperties =
+    kExpanded | kMutable | kError | kIDeterministic | kNonIDeterministic |
+    kIEpsilons | kNoIEpsilons | kILabelSorted | kNotILabelSorted | kWeighted |
+    kUnweighted | kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic |
+    kTopSorted | kNotTopSorted | kAccessible | kNotAccessible | kCoAccessible |
+    kNotCoAccessible | kString | kNotString | kWeightedCycles |
+    kUnweightedCycles;
+
+// Properties that are preserved when an FST's weights are changed. This
+// assumes that the set of states that are non-final is not changed.
+constexpr uint64 kWeightInvariantProperties =
+    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
+    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
+    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
+    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
+    kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic | kTopSorted |
+    kNotTopSorted | kAccessible | kNotAccessible | kCoAccessible |
+    kNotCoAccessible | kString | kNotString;
+
+// Properties that are preserved when a superfinal state is added and an FST's
+// final weights are directed to it via new transitions.
+constexpr uint64 kAddSuperFinalProperties =
+    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor |
+    kNonIDeterministic | kNonODeterministic | kEpsilons | kIEpsilons |
+    kOEpsilons | kNotILabelSorted | kNotOLabelSorted | kWeighted | kUnweighted |
+    kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic | kNotTopSorted |
+    kNotAccessible | kCoAccessible | kNotCoAccessible | kNotString |
+    kWeightedCycles | kUnweightedCycles;
+
+// Properties that are preserved when a superfinal state is removed and the
+// epsilon transitions directed to it are made final weights.
+constexpr uint64 kRmSuperFinalProperties =
+    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
+    kODeterministic | kNoEpsilons | kNoIEpsilons | kNoOEpsilons |
+    kILabelSorted | kOLabelSorted | kWeighted | kUnweighted | kCyclic |
+    kAcyclic | kInitialCyclic | kInitialAcyclic | kTopSorted | kAccessible |
+    kCoAccessible | kNotCoAccessible | kString | kWeightedCycles |
+    kUnweightedCycles;
+
+// All binary properties.
+constexpr uint64 kBinaryProperties = 0x0000000000000007ULL;
+
+// All trinary properties.
+constexpr uint64 kTrinaryProperties = 0x0000ffffffff0000ULL;
+
+// COMPUTED PROPERTIES
+
+// 1st bit of trinary properties.
+constexpr uint64 kPosTrinaryProperties = kTrinaryProperties &
+    0x5555555555555555ULL;
+
+// 2nd bit of trinary properties.
+constexpr uint64 kNegTrinaryProperties = kTrinaryProperties &
+    0xaaaaaaaaaaaaaaaaULL;
+
+// All properties.
+constexpr uint64 kFstProperties = kBinaryProperties | kTrinaryProperties;
+
+// PROPERTY FUNCTIONS and STRING NAMES (defined in properties.cc).
+
+// Below are functions for getting property bit vectors when executing
+// mutation operations.
+inline uint64 SetStartProperties(uint64 inprops);
+
+template <typename Weight>
+uint64 SetFinalProperties(uint64 inprops, const Weight &old_weight,
+                          const Weight &new_weight);
+
+inline uint64 AddStateProperties(uint64 inprops);
+
+template <typename A>
+uint64 AddArcProperties(uint64 inprops, typename A::StateId s, const A &arc,
+                        const A *prev_arc);
+
+inline uint64 DeleteStatesProperties(uint64 inprops);
+
+inline uint64 DeleteAllStatesProperties(uint64 inprops, uint64 staticProps);
+
+inline uint64 DeleteArcsProperties(uint64 inprops);
+
+uint64 ClosureProperties(uint64 inprops, bool star, bool delayed = false);
+
+uint64 ComplementProperties(uint64 inprops);
+
+uint64 ComposeProperties(uint64 inprops1, uint64 inprops2);
+
+uint64 ConcatProperties(uint64 inprops1, uint64 inprops2, bool delayed = false);
+
+uint64 DeterminizeProperties(uint64 inprops, bool has_subsequential_label,
+                             bool distinct_psubsequential_labels);
+
+uint64 FactorWeightProperties(uint64 inprops);
+
+uint64 InvertProperties(uint64 inprops);
+
+uint64 ProjectProperties(uint64 inprops, bool project_input);
+
+uint64 RandGenProperties(uint64 inprops, bool weighted);
+
+uint64 RelabelProperties(uint64 inprops);
+
+uint64 ReplaceProperties(const std::vector<uint64> &inprops, size_t root,
+                         bool epsilon_on_call, bool epsilon_on_return,
+                         bool out_epsilon_on_call, bool out_epsilon_on_return,
+                         bool replace_transducer, bool no_empty_fst,
+                         bool all_ilabel_sorted, bool all_olabel_sorted,
+                         bool all_negative_or_dense);
+
+uint64 ReverseProperties(uint64 inprops, bool has_superinitial);
+
+uint64 ReweightProperties(uint64 inprops);
+
+uint64 RmEpsilonProperties(uint64 inprops, bool delayed = false);
+
+uint64 ShortestPathProperties(uint64 props, bool tree = false);
+
+uint64 SynchronizeProperties(uint64 inprops);
+
+uint64 UnionProperties(uint64 inprops1, uint64 inprops2, bool delayed = false);
+
+// Definitions of inlined functions.
+
+uint64 SetStartProperties(uint64 inprops) {
+  auto outprops = inprops & kSetStartProperties;
+  if (inprops & kAcyclic) {
+    outprops |= kInitialAcyclic;
+  }
+  return outprops;
+}
+
+uint64 AddStateProperties(uint64 inprops) {
+  return inprops & kAddStateProperties;
+}
+
+uint64 DeleteStatesProperties(uint64 inprops) {
+  return inprops & kDeleteStatesProperties;
+}
+
+uint64 DeleteAllStatesProperties(uint64 inprops, uint64 staticprops) {
+  const auto outprops = inprops & kError;
+  return outprops | kNullProperties | staticprops;
+}
+
+uint64 DeleteArcsProperties(uint64 inprops) {
+  return inprops & kDeleteArcsProperties;
+}
+
+// Definitions of template functions.
+
+template <typename Weight>
+uint64 SetFinalProperties(uint64 inprops, const Weight &old_weight,
+                          const Weight &new_weight) {
+  auto outprops = inprops;
+  if (old_weight != Weight::Zero() && old_weight != Weight::One()) {
+    outprops &= ~kWeighted;
+  }
+  if (new_weight != Weight::Zero() && new_weight != Weight::One()) {
+    outprops |= kWeighted;
+    outprops &= ~kUnweighted;
+  }
+  outprops &= kSetFinalProperties | kWeighted | kUnweighted;
+  return outprops;
+}
+
+/// Gets the properties for the MutableFst::AddArc method.
+///
+/// \param inprops  the current properties of the FST
+/// \param s        the ID of the state to which an arc is being added.
+/// \param arc      the arc being added to the state with the specified ID
+/// \param prev_arc the previously-added (or "last") arc of state s, or nullptr
+//                  if s currently has no arcs.
+template <typename Arc>
+uint64 AddArcProperties(uint64 inprops, typename Arc::StateId s,
+                        const Arc &arc, const Arc *prev_arc) {
+  using Weight = typename Arc::Weight;
+  auto outprops = inprops;
+  if (arc.ilabel != arc.olabel) {
+    outprops |= kNotAcceptor;
+    outprops &= ~kAcceptor;
+  }
+  if (arc.ilabel == 0) {
+    outprops |= kIEpsilons;
+    outprops &= ~kNoIEpsilons;
+    if (arc.olabel == 0) {
+      outprops |= kEpsilons;
+      outprops &= ~kNoEpsilons;
+    }
+  }
+  if (arc.olabel == 0) {
+    outprops |= kOEpsilons;
+    outprops &= ~kNoOEpsilons;
+  }
+  if (prev_arc) {
+    if (prev_arc->ilabel > arc.ilabel) {
+      outprops |= kNotILabelSorted;
+      outprops &= ~kILabelSorted;
+    }
+    if (prev_arc->olabel > arc.olabel) {
+      outprops |= kNotOLabelSorted;
+      outprops &= ~kOLabelSorted;
+    }
+  }
+  if (arc.weight != Weight::Zero() && arc.weight != Weight::One()) {
+    outprops |= kWeighted;
+    outprops &= ~kUnweighted;
+  }
+  if (arc.nextstate <= s) {
+    outprops |= kNotTopSorted;
+    outprops &= ~kTopSorted;
+  }
+  outprops &= kAddArcProperties | kAcceptor | kNoEpsilons | kNoIEpsilons |
+              kNoOEpsilons | kILabelSorted | kOLabelSorted | kUnweighted |
+              kTopSorted;
+  if (outprops & kTopSorted) {
+    outprops |= kAcyclic | kInitialAcyclic;
+  }
+  return outprops;
+}
+
+extern const char *PropertyNames[];
+
+}  // namespace fst
+
+#endif  // FST_PROPERTIES_H_
diff --git a/projects/llm_framework/include/fst/prune.h b/projects/llm_framework/include/fst/prune.h
new file mode 100644
index 00000000..9e7c04bd
--- /dev/null
+++ b/projects/llm_framework/include/fst/prune.h
@@ -0,0 +1,341 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions implementing pruning.
+
+#ifndef FST_PRUNE_H_
+#define FST_PRUNE_H_
+
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/arcfilter.h>
+#include <fst/heap.h>
+#include <fst/shortest-distance.h>
+
+
+namespace fst {
+namespace internal {
+
+template <class StateId, class Weight>
+class PruneCompare {
+ public:
+  PruneCompare(const std::vector<Weight> &idistance,
+               const std::vector<Weight> &fdistance)
+      : idistance_(idistance), fdistance_(fdistance) {}
+
+  bool operator()(const StateId x, const StateId y) const {
+    const auto wx = Times(IDistance(x), FDistance(x));
+    const auto wy = Times(IDistance(y), FDistance(y));
+    return less_(wx, wy);
+  }
+
+ private:
+  Weight IDistance(const StateId s) const {
+    return s < idistance_.size() ? idistance_[s] : Weight::Zero();
+  }
+
+  Weight FDistance(const StateId s) const {
+    return s < fdistance_.size() ? fdistance_[s] : Weight::Zero();
+  }
+
+  const std::vector<Weight> &idistance_;
+  const std::vector<Weight> &fdistance_;
+  NaturalLess<Weight> less_;
+};
+
+}  // namespace internal
+
+template <class Arc, class ArcFilter>
+struct PruneOptions {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  explicit PruneOptions(const Weight &weight_threshold = Weight::Zero(),
+                        StateId state_threshold = kNoStateId,
+                        ArcFilter filter = ArcFilter(),
+                        std::vector<Weight> *distance = nullptr,
+                        float delta = kDelta, bool threshold_initial = false)
+      : weight_threshold(std::move(weight_threshold)),
+        state_threshold(state_threshold),
+        filter(std::move(filter)),
+        distance(distance),
+        delta(delta),
+        threshold_initial(threshold_initial) {}
+
+  // Pruning weight threshold.
+  Weight weight_threshold;
+  // Pruning state threshold.
+  StateId state_threshold;
+  // Arc filter.
+  ArcFilter filter;
+  // If non-zero, passes in pre-computed shortest distance to final states.
+  const std::vector<Weight> *distance;
+  // Determines the degree of convergence required when computing shortest
+  // distances.
+  float delta;
+  // Determines if the shortest path weight is left (true) or right
+  // (false) multiplied by the threshold to get the limit for
+  // keeping a state or arc (matters if the semiring is not
+  // commutative).
+  bool threshold_initial;
+};
+
+// Pruning algorithm: this version modifies its input and it takes an options
+// class as an argument. After pruning the FST contains states and arcs that
+// belong to a successful path in the FST whose weight is no more than the
+// weight of the shortest path Times() the provided weight threshold. When the
+// state threshold is not kNoStateId, the output FST is further restricted to
+// have no more than the number of states in opts.state_threshold. Weights must
+// have the path property. The weight of any cycle needs to be bounded; i.e.,
+//
+//   Plus(weight, Weight::One()) == Weight::One()
+template <class Arc, class ArcFilter,
+          typename std::enable_if<IsPath<typename Arc::Weight>::value>::type * =
+              nullptr>
+void Prune(MutableFst<Arc> *fst, const PruneOptions<Arc, ArcFilter> &opts =
+                                     PruneOptions<Arc, ArcFilter>()) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using StateHeap = Heap<StateId, internal::PruneCompare<StateId, Weight>>;
+  auto ns = fst->NumStates();
+  if (ns < 1) return;
+  std::vector<Weight> idistance(ns, Weight::Zero());
+  std::vector<Weight> tmp;
+  if (!opts.distance) {
+    tmp.reserve(ns);
+    ShortestDistance(*fst, &tmp, true, opts.delta);
+  }
+  const auto *fdistance = opts.distance ? opts.distance : &tmp;
+  if ((opts.state_threshold == 0) || (fdistance->size() <= fst->Start()) ||
+      ((*fdistance)[fst->Start()] == Weight::Zero())) {
+    fst->DeleteStates();
+    return;
+  }
+  internal::PruneCompare<StateId, Weight> compare(idistance, *fdistance);
+  StateHeap heap(compare);
+  std::vector<bool> visited(ns, false);
+  std::vector<size_t> enqueued(ns, StateHeap::kNoKey);
+  std::vector<StateId> dead;
+  dead.push_back(fst->AddState());
+  NaturalLess<Weight> less;
+  auto s = fst->Start();
+  const auto limit = opts.threshold_initial ?
+      Times(opts.weight_threshold, (*fdistance)[s]) :
+      Times((*fdistance)[s], opts.weight_threshold);
+  StateId num_visited = 0;
+
+  if (!less(limit, (*fdistance)[s])) {
+    idistance[s] = Weight::One();
+    enqueued[s] = heap.Insert(s);
+    ++num_visited;
+  }
+  while (!heap.Empty()) {
+    s = heap.Top();
+    heap.Pop();
+    enqueued[s] = StateHeap::kNoKey;
+    visited[s] = true;
+    if (less(limit, Times(idistance[s], fst->Final(s)))) {
+      fst->SetFinal(s, Weight::Zero());
+    }
+    for (MutableArcIterator<MutableFst<Arc>> aiter(fst, s); !aiter.Done();
+         aiter.Next()) {
+      auto arc = aiter.Value();  // Copy intended.
+      if (!opts.filter(arc)) continue;
+      const auto weight = Times(Times(idistance[s], arc.weight),
+                                arc.nextstate < fdistance->size() ?
+                                (*fdistance)[arc.nextstate] : Weight::Zero());
+      if (less(limit, weight)) {
+        arc.nextstate = dead[0];
+        aiter.SetValue(arc);
+        continue;
+      }
+      if (less(Times(idistance[s], arc.weight), idistance[arc.nextstate])) {
+        idistance[arc.nextstate] = Times(idistance[s], arc.weight);
+      }
+      if (visited[arc.nextstate]) continue;
+      if ((opts.state_threshold != kNoStateId) &&
+          (num_visited >= opts.state_threshold)) {
+        continue;
+      }
+      if (enqueued[arc.nextstate] == StateHeap::kNoKey) {
+        enqueued[arc.nextstate] = heap.Insert(arc.nextstate);
+        ++num_visited;
+      } else {
+        heap.Update(enqueued[arc.nextstate], arc.nextstate);
+      }
+    }
+  }
+  for (StateId i = 0; i < visited.size(); ++i) {
+    if (!visited[i]) dead.push_back(i);
+  }
+  fst->DeleteStates(dead);
+}
+
+template <class Arc, class ArcFilter,
+          typename std::enable_if<!IsPath<typename Arc::Weight>::value>::type
+              * = nullptr>
+void Prune(MutableFst<Arc> *fst, const PruneOptions<Arc, ArcFilter> &opts =
+                                     PruneOptions<Arc, ArcFilter>()) {
+  FSTERROR() << "Prune: Weight needs to have the path property: "
+             << Arc::Weight::Type();
+  fst->SetProperties(kError, kError);
+}
+
+// Pruning algorithm: this version modifies its input and takes the
+// pruning threshold as an argument. It deletes states and arcs in the
+// FST that do not belong to a successful path whose weight is more
+// than the weight of the shortest path Times() the provided weight
+// threshold. When the state threshold is not kNoStateId, the output
+// FST is further restricted to have no more than the number of states
+// in opts.state_threshold. Weights must have the path property. The
+// weight of any cycle needs to be bounded; i.e.,
+//
+//   Plus(weight, Weight::One()) == Weight::One()
+template <class Arc>
+void Prune(MutableFst<Arc> *fst, typename Arc::Weight weight_threshold,
+           typename Arc::StateId state_threshold = kNoStateId,
+           float delta = kDelta) {
+  const PruneOptions<Arc, AnyArcFilter<Arc>> opts(
+      weight_threshold, state_threshold, AnyArcFilter<Arc>(), nullptr, delta);
+  Prune(fst, opts);
+}
+
+// Pruning algorithm: this version writes the pruned input FST to an
+// output MutableFst and it takes an options class as an argument. The
+// output FST contains states and arcs that belong to a successful
+// path in the input FST whose weight is more than the weight of the
+// shortest path Times() the provided weight threshold. When the state
+// threshold is not kNoStateId, the output FST is further restricted
+// to have no more than the number of states in
+// opts.state_threshold. Weights have the path property.  The weight
+// of any cycle needs to be bounded; i.e.,
+//
+//   Plus(weight, Weight::One()) == Weight::One()
+template <class Arc, class ArcFilter,
+          typename std::enable_if<IsPath<typename Arc::Weight>::value>::type * =
+              nullptr>
+void Prune(
+    const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
+    const PruneOptions<Arc, ArcFilter> &opts = PruneOptions<Arc, ArcFilter>()) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using StateHeap = Heap<StateId, internal::PruneCompare<StateId, Weight>>;
+  ofst->DeleteStates();
+  ofst->SetInputSymbols(ifst.InputSymbols());
+  ofst->SetOutputSymbols(ifst.OutputSymbols());
+  if (ifst.Start() == kNoStateId) return;
+  NaturalLess<Weight> less;
+  if (less(opts.weight_threshold, Weight::One()) ||
+      (opts.state_threshold == 0)) {
+    return;
+  }
+  std::vector<Weight> idistance;
+  std::vector<Weight> tmp;
+  if (!opts.distance) ShortestDistance(ifst, &tmp, true, opts.delta);
+  const auto *fdistance = opts.distance ? opts.distance : &tmp;
+  if ((fdistance->size() <= ifst.Start()) ||
+      ((*fdistance)[ifst.Start()] == Weight::Zero())) {
+    return;
+  }
+  internal::PruneCompare<StateId, Weight> compare(idistance, *fdistance);
+  StateHeap heap(compare);
+  std::vector<StateId> copy;
+  std::vector<size_t> enqueued;
+  std::vector<bool> visited;
+  auto s = ifst.Start();
+  const auto limit = opts.threshold_initial ?
+      Times(opts.weight_threshold, (*fdistance)[s]) :
+      Times((*fdistance)[s], opts.weight_threshold);
+  while (copy.size() <= s) copy.push_back(kNoStateId);
+  copy[s] = ofst->AddState();
+  ofst->SetStart(copy[s]);
+  while (idistance.size() <= s) idistance.push_back(Weight::Zero());
+  idistance[s] = Weight::One();
+  while (enqueued.size() <= s) {
+    enqueued.push_back(StateHeap::kNoKey);
+    visited.push_back(false);
+  }
+  enqueued[s] = heap.Insert(s);
+  while (!heap.Empty()) {
+    s = heap.Top();
+    heap.Pop();
+    enqueued[s] = StateHeap::kNoKey;
+    visited[s] = true;
+    if (!less(limit, Times(idistance[s], ifst.Final(s)))) {
+      ofst->SetFinal(copy[s], ifst.Final(s));
+    }
+    for (ArcIterator<Fst<Arc>> aiter(ifst, s); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      if (!opts.filter(arc)) continue;
+      const auto weight = Times(Times(idistance[s], arc.weight),
+                                arc.nextstate < fdistance->size() ?
+                                (*fdistance)[arc.nextstate] : Weight::Zero());
+      if (less(limit, weight)) continue;
+      if ((opts.state_threshold != kNoStateId) &&
+          (ofst->NumStates() >= opts.state_threshold)) {
+        continue;
+      }
+      while (idistance.size() <= arc.nextstate) {
+        idistance.push_back(Weight::Zero());
+      }
+      if (less(Times(idistance[s], arc.weight), idistance[arc.nextstate])) {
+        idistance[arc.nextstate] = Times(idistance[s], arc.weight);
+      }
+      while (copy.size() <= arc.nextstate) copy.push_back(kNoStateId);
+      if (copy[arc.nextstate] == kNoStateId) {
+        copy[arc.nextstate] = ofst->AddState();
+      }
+      ofst->AddArc(copy[s], Arc(arc.ilabel, arc.olabel, arc.weight,
+                                copy[arc.nextstate]));
+      while (enqueued.size() <= arc.nextstate) {
+        enqueued.push_back(StateHeap::kNoKey);
+        visited.push_back(false);
+      }
+      if (visited[arc.nextstate]) continue;
+      if (enqueued[arc.nextstate] == StateHeap::kNoKey) {
+        enqueued[arc.nextstate] = heap.Insert(arc.nextstate);
+      } else {
+        heap.Update(enqueued[arc.nextstate], arc.nextstate);
+      }
+    }
+  }
+}
+
+template <class Arc, class ArcFilter,
+          typename std::enable_if<!IsPath<typename Arc::Weight>::value>::type
+              * = nullptr>
+void Prune(const Fst<Arc> &, MutableFst<Arc> *ofst,
+           const PruneOptions<Arc, ArcFilter> &) {
+  FSTERROR() << "Prune: Weight needs to have the path property: "
+             << Arc::Weight::Type();
+  ofst->SetProperties(kError, kError);
+}
+
+// Pruning algorithm: this version writes the pruned input FST to an
+// output MutableFst and simply takes the pruning threshold as an
+// argument. The output FST contains states and arcs that belong to a
+// successful path in the input FST whose weight is no more than the
+// weight of the shortest path Times() the provided weight
+// threshold. When the state threshold is not kNoStateId, the output
+// FST is further restricted to have no more than the number of states
+// in opts.state_threshold. Weights must have the path property. The
+// weight of any cycle needs to be bounded; i.e.,
+//
+// Plus(weight, Weight::One()) = Weight::One();
+template <class Arc>
+void Prune(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
+           typename Arc::Weight weight_threshold,
+           typename Arc::StateId state_threshold = kNoStateId,
+           float delta = kDelta) {
+  const PruneOptions<Arc, AnyArcFilter<Arc>> opts(
+      weight_threshold, state_threshold, AnyArcFilter<Arc>(), nullptr, delta);
+  Prune(ifst, ofst, opts);
+}
+
+}  // namespace fst
+
+#endif  // FST_PRUNE_H_
diff --git a/projects/llm_framework/include/fst/push.h b/projects/llm_framework/include/fst/push.h
new file mode 100644
index 00000000..1f772739
--- /dev/null
+++ b/projects/llm_framework/include/fst/push.h
@@ -0,0 +1,155 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to reweight/push an FST, and utility functions to weigh and reweight
+// an FST.
+
+#ifndef FST_PUSH_H_
+#define FST_PUSH_H_
+
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/arc-map.h>
+#include <fst/factor-weight.h>
+#include <fst/fst.h>
+#include <fst/reweight.h>
+#include <fst/shortest-distance.h>
+
+
+namespace fst {
+
+// Computes the total weight (sum of the weights of all accepting paths) from
+// the output of ShortestDistance, using the shortest distance from the final
+// state when reverse is true and from the initial state otherwise.
+template <class Arc>
+typename Arc::Weight ComputeTotalWeight(
+    const Fst<Arc> &fst, const std::vector<typename Arc::Weight> &distance,
+    bool reverse) {
+  if (reverse) {
+    return fst.Start() < distance.size() ? distance[fst.Start()]
+                                         : Arc::Weight::Zero();
+  }
+  auto sum = Arc::Weight::Zero();
+  for (typename Arc::StateId s = 0; s < distance.size(); ++s) {
+    sum = Plus(sum, Times(distance[s], fst.Final(s)));
+  }
+  return sum;
+}
+
+// Divides the weight of every accepting path by a fixed weight. This weight
+// is also divided at the final state if at_final is true and at the initial
+// state otherwise.
+template <class Arc>
+void RemoveWeight(MutableFst<Arc> *fst, const typename Arc::Weight &weight,
+                  bool at_final) {
+  using Weight = typename Arc::Weight;
+  if ((weight == Weight::One()) || (weight == Weight::Zero())) return;
+  if (at_final) {
+    for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
+         siter.Next()) {
+      fst->SetFinal(siter.Value(),
+                    Divide(fst->Final(siter.Value()), weight, DIVIDE_RIGHT));
+    }
+  } else {
+    const auto start = fst->Start();
+    for (MutableArcIterator<MutableFst<Arc>> aiter(fst, start); !aiter.Done();
+         aiter.Next()) {
+      auto arc = aiter.Value();
+      arc.weight = Divide(arc.weight, weight, DIVIDE_LEFT);
+      aiter.SetValue(arc);
+    }
+    fst->SetFinal(start, Divide(fst->Final(start), weight, DIVIDE_LEFT));
+  }
+}
+
+// Pushes the weights in FST in the direction defined by TYPE. If
+// pushing towards the initial state, the sum of the weight of the
+// outgoing transitions and final weight at a non-initial state is
+// equal to One() in the resulting machine. If pushing towards the
+// final state, the same property holds on the reverse machine.
+//
+// Weight needs to be left distributive when pushing towards the
+// initial state and right distributive when pushing towards the final
+// states.
+template <class Arc>
+void Push(MutableFst<Arc> *fst, ReweightType type, float delta = kDelta,
+          bool remove_total_weight = false) {
+  using Weight = typename Arc::Weight;
+  std::vector<Weight> distance;
+  ShortestDistance(*fst, &distance, type == REWEIGHT_TO_INITIAL, delta);
+  auto total_weight = Weight::One();
+  if (remove_total_weight) {
+    total_weight =
+        ComputeTotalWeight(*fst, distance, type == REWEIGHT_TO_INITIAL);
+  }
+  Reweight(fst, distance, type);
+  if (remove_total_weight) {
+    RemoveWeight(fst, total_weight, type == REWEIGHT_TO_FINAL);
+  }
+}
+
+constexpr uint32 kPushWeights = 0x0001;
+constexpr uint32 kPushLabels = 0x0002;
+constexpr uint32 kPushRemoveTotalWeight = 0x0004;
+constexpr uint32 kPushRemoveCommonAffix = 0x0008;
+
+// Pushes the weights and/or labels of the input FST into the output
+// mutable FST by pushing weights and/or labels (as determined by the
+// ptype argument) towards the initial state or final states (as
+// determined by the rtype template parameter). The weight type must
+// be left distributive when pushing weights towards the initial state, and
+// right distribution when pushing weights towards the final states.
+template <class Arc, ReweightType rtype>
+void Push(const Fst<Arc> &ifst, MutableFst<Arc> *ofst, uint32 ptype,
+          float delta = kDelta) {
+  using Label = typename Arc::Label;
+  using Weight = typename Arc::Weight;
+  if ((ptype & (kPushWeights | kPushLabels)) == kPushWeights) {
+    *ofst = ifst;
+    Push(ofst, rtype, delta, ptype & kPushRemoveTotalWeight);
+  } else if (ptype & kPushLabels) {
+    const auto gtype =
+        rtype == REWEIGHT_TO_INITIAL ? GALLIC_LEFT : GALLIC_RIGHT;
+    using GallicWeight = typename GallicArc<Arc, gtype>::Weight;
+    std::vector<GallicWeight> gdistance;
+    VectorFst<GallicArc<Arc, gtype>> gfst;
+    ArcMap(ifst, &gfst, ToGallicMapper<Arc, gtype>());
+    if (ptype & kPushWeights) {
+      ShortestDistance(gfst, &gdistance, rtype == REWEIGHT_TO_INITIAL, delta);
+    } else {
+      ArcMapFst<Arc, Arc, RmWeightMapper<Arc>> uwfst(ifst,
+                                                      RmWeightMapper<Arc>());
+      ArcMapFst<Arc, GallicArc<Arc, gtype>, ToGallicMapper<Arc, gtype>> guwfst(
+          uwfst, ToGallicMapper<Arc, gtype>());
+      ShortestDistance(guwfst, &gdistance, rtype == REWEIGHT_TO_INITIAL, delta);
+    }
+    auto total_weight = GallicWeight::One();
+    if (ptype & (kPushRemoveTotalWeight | kPushRemoveCommonAffix)) {
+      total_weight =
+          ComputeTotalWeight(gfst, gdistance, rtype == REWEIGHT_TO_INITIAL);
+      total_weight = GallicWeight(
+          ptype & kPushRemoveCommonAffix
+              ? total_weight.Value1()
+              : StringWeight<Label, GallicStringType(gtype)>::One(),
+          ptype & kPushRemoveTotalWeight ? total_weight.Value2()
+                                         : Weight::One());
+    }
+    Reweight(&gfst, gdistance, rtype);
+    if (ptype & (kPushRemoveTotalWeight | kPushRemoveCommonAffix)) {
+      RemoveWeight(&gfst, total_weight, rtype == REWEIGHT_TO_FINAL);
+    }
+    FactorWeightFst<GallicArc<Arc, gtype>, GallicFactor<Label, Weight, gtype>>
+        fwfst(gfst);
+    ArcMap(fwfst, ofst, FromGallicMapper<Arc, gtype>());
+    ofst->SetOutputSymbols(ifst.OutputSymbols());
+  } else {
+    LOG(WARNING) << "Push: pushing type is set to 0, so not pushing";
+    *ofst = ifst;
+  }
+}
+
+}  // namespace fst
+
+#endif  // FST_PUSH_H_
diff --git a/projects/llm_framework/include/fst/queue.h b/projects/llm_framework/include/fst/queue.h
new file mode 100644
index 00000000..f57d176e
--- /dev/null
+++ b/projects/llm_framework/include/fst/queue.h
@@ -0,0 +1,948 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes for various FST state queues with a unified interface.
+
+#ifndef FST_QUEUE_H_
+#define FST_QUEUE_H_
+
+#include <deque>
+#include <memory>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/arcfilter.h>
+#include <fst/connect.h>
+#include <fst/heap.h>
+#include <fst/topsort.h>
+#include <fst/weight.h>
+
+
+namespace fst {
+
+// The Queue interface is:
+//
+// template <class S>
+// class Queue {
+//  public:
+//   using StateId = S;
+//
+//   // Constructor: may need args (e.g., FST, comparator) for some queues.
+//   Queue(...) override;
+//
+//   // Returns the head of the queue.
+//   StateId Head() const override;
+//
+//   // Inserts a state.
+//   void Enqueue(StateId s) override;
+//
+//   // Removes the head of the queue.
+//   void Dequeue() override;
+//
+//   // Updates ordering of state s when weight changes, if necessary.
+//   void Update(StateId s) override;
+//
+//   // Is the queue empty?
+//   bool Empty() const override;
+//
+//   // Removes all states from the queue.
+//   void Clear() override;
+// };
+
+// State queue types.
+enum QueueType {
+  TRIVIAL_QUEUE = 0,         // Single state queue.
+  FIFO_QUEUE = 1,            // First-in, first-out queue.
+  LIFO_QUEUE = 2,            // Last-in, first-out queue.
+  SHORTEST_FIRST_QUEUE = 3,  // Shortest-first queue.
+  TOP_ORDER_QUEUE = 4,       // Topologically-ordered queue.
+  STATE_ORDER_QUEUE = 5,     // State ID-ordered queue.
+  SCC_QUEUE = 6,             // Component graph top-ordered meta-queue.
+  AUTO_QUEUE = 7,            // Auto-selected queue.
+  OTHER_QUEUE = 8
+};
+
+// QueueBase, templated on the StateId, is a virtual base class shared by all
+// queues considered by AutoQueue.
+template <class S>
+class QueueBase {
+ public:
+  using StateId = S;
+
+  virtual ~QueueBase() {}
+
+  // Concrete implementation.
+
+  explicit QueueBase(QueueType type) : queue_type_(type), error_(false) {}
+
+  void SetError(bool error) { error_ = error; }
+
+  bool Error() const { return error_; }
+
+  QueueType Type() const { return queue_type_; }
+
+  // Virtual interface.
+
+  virtual StateId Head() const = 0;
+  virtual void Enqueue(StateId) = 0;
+  virtual void Dequeue() = 0;
+  virtual void Update(StateId) = 0;
+  virtual bool Empty() const = 0;
+  virtual void Clear() = 0;
+
+ private:
+  QueueType queue_type_;
+  bool error_;
+};
+
+// Trivial queue discipline; one may enqueue at most one state at a time. It
+// can be used for strongly connected components with only one state and no
+// self-loops.
+template <class S>
+class TrivialQueue : public QueueBase<S> {
+ public:
+  using StateId = S;
+
+  TrivialQueue() : QueueBase<StateId>(TRIVIAL_QUEUE), front_(kNoStateId) {}
+
+  virtual ~TrivialQueue() = default;
+
+  StateId Head() const final { return front_; }
+
+  void Enqueue(StateId s) final { front_ = s; }
+
+  void Dequeue() final { front_ = kNoStateId; }
+
+  void Update(StateId) final {}
+
+  bool Empty() const final { return front_ == kNoStateId; }
+
+  void Clear() final { front_ = kNoStateId; }
+
+ private:
+  StateId front_;
+};
+
+// First-in, first-out queue discipline.
+//
+// This is not a final class.
+template <class S>
+class FifoQueue : public QueueBase<S> {
+ public:
+  using StateId = S;
+
+  FifoQueue() : QueueBase<StateId>(FIFO_QUEUE) {}
+
+  virtual ~FifoQueue() = default;
+
+  StateId Head() const override { return queue_.back(); }
+
+  void Enqueue(StateId s) override { queue_.push_front(s); }
+
+  void Dequeue() override { queue_.pop_back(); }
+
+  void Update(StateId) override {}
+
+  bool Empty() const override { return queue_.empty(); }
+
+  void Clear() override { queue_.clear(); }
+
+ private:
+  std::deque<StateId> queue_;
+};
+
+// Last-in, first-out queue discipline.
+template <class S>
+class LifoQueue : public QueueBase<S> {
+ public:
+  using StateId = S;
+
+  LifoQueue() : QueueBase<StateId>(LIFO_QUEUE) {}
+
+  virtual ~LifoQueue() = default;
+
+  StateId Head() const final { return queue_.front(); }
+
+  void Enqueue(StateId s) final { queue_.push_front(s); }
+
+  void Dequeue() final { queue_.pop_front(); }
+
+  void Update(StateId) final {}
+
+  bool Empty() const final { return queue_.empty(); }
+
+  void Clear() final { queue_.clear(); }
+
+ private:
+  std::deque<StateId> queue_;
+};
+
+// Shortest-first queue discipline, templated on the StateId and as well as a
+// comparison functor used to compare two StateIds. If a (single) state's order
+// changes, it can be reordered in the queue with a call to Update(). If update
+// is false, call to Update() does not reorder the queue.
+//
+// This is not a final class.
+template <typename S, typename Compare, bool update = true>
+class ShortestFirstQueue : public QueueBase<S> {
+ public:
+  using StateId = S;
+
+  explicit ShortestFirstQueue(Compare comp)
+      : QueueBase<StateId>(SHORTEST_FIRST_QUEUE), heap_(comp) {}
+
+  virtual ~ShortestFirstQueue() = default;
+
+  StateId Head() const override { return heap_.Top(); }
+
+  void Enqueue(StateId s) override {
+    if (update) {
+      for (StateId i = key_.size(); i <= s; ++i) key_.push_back(kNoStateId);
+      key_[s] = heap_.Insert(s);
+    } else {
+      heap_.Insert(s);
+    }
+  }
+
+  void Dequeue() override {
+    if (update) {
+      key_[heap_.Pop()] = kNoStateId;
+    } else {
+      heap_.Pop();
+    }
+  }
+
+  void Update(StateId s) override {
+    if (!update) return;
+    if (s >= key_.size() || key_[s] == kNoStateId) {
+      Enqueue(s);
+    } else {
+      heap_.Update(key_[s], s);
+    }
+  }
+
+  bool Empty() const override { return heap_.Empty(); }
+
+  void Clear() override {
+    heap_.Clear();
+    if (update) key_.clear();
+  }
+
+  const Compare &GetCompare() const { return heap_.GetCompare(); }
+
+ private:
+  Heap<StateId, Compare> heap_;
+  std::vector<ssize_t> key_;
+};
+
+namespace internal {
+
+// Given a vector that maps from states to weights, and a comparison functor
+// for weights, this class defines a comparison function object between states.
+template <typename StateId, typename Less>
+class StateWeightCompare {
+ public:
+  using Weight = typename Less::Weight;
+
+  StateWeightCompare(const std::vector<Weight> &weights, const Less &less)
+      : weights_(weights), less_(less) {}
+
+  bool operator()(const StateId s1, const StateId s2) const {
+    return less_(weights_[s1], weights_[s2]);
+  }
+
+ private:
+  // Borrowed references.
+  const std::vector<Weight> &weights_;
+  const Less &less_;
+};
+
+}  // namespace internal
+
+// Shortest-first queue discipline, templated on the StateId and Weight, is
+// specialized to use the weight's natural order for the comparison function.
+template <typename S, typename Weight>
+class NaturalShortestFirstQueue
+    : public ShortestFirstQueue<
+          S, internal::StateWeightCompare<S, NaturalLess<Weight>>> {
+ public:
+  using StateId = S;
+  using Compare = internal::StateWeightCompare<StateId, NaturalLess<Weight>>;
+
+  explicit NaturalShortestFirstQueue(const std::vector<Weight> &distance)
+      : ShortestFirstQueue<StateId, Compare>(Compare(distance, less_)) {}
+
+  virtual ~NaturalShortestFirstQueue() = default;
+
+ private:
+  // This is non-static because the constructor for non-idempotent weights will
+  // result in an error.
+  const NaturalLess<Weight> less_{};
+};
+
+// In a shortest path computation on a lattice-like FST, we may keep many old
+// nonviable paths as a part of the search. Since the search process always
+// expands the lowest cost path next, that lowest cost path may be a very old
+// nonviable path instead of one we expect to lead to a shortest path.
+//
+// For instance, suppose that the current best path in an alignment has
+// traversed 500 arcs with a cost of 10. We may also have a bad path in
+// the queue that has traversed only 40 arcs but also has a cost of 10.
+// This path is very unlikely to lead to a reasonable alignment, so this queue
+// can prune it from the search space.
+//
+// This queue relies on the caller using a shortest-first exploration order
+// like this:
+//   while (true) {
+//     StateId head = queue.Head();
+//     queue.Dequeue();
+//     for (const auto& arc : GetArcs(fst, head)) {
+//       queue.Enqueue(arc.nextstate);
+//     }
+//   }
+// We use this assumption to guess that there is an arc between Head and the
+// Enqueued state; this is how the number of path steps is measured.
+template <typename S, typename Weight>
+class PruneNaturalShortestFirstQueue
+    : public NaturalShortestFirstQueue<S, Weight> {
+ public:
+  using StateId = S;
+  using Base = NaturalShortestFirstQueue<StateId, Weight>;
+
+  explicit PruneNaturalShortestFirstQueue(const std::vector<Weight> &distance,
+                                          int threshold)
+      : Base(distance),
+        threshold_(threshold),
+        head_steps_(0),
+        max_head_steps_(0) {}
+
+  ~PruneNaturalShortestFirstQueue() override = default;
+
+  StateId Head() const override {
+    const auto head = Base::Head();
+    // Stores the number of steps from the start of the graph to this state
+    // along the shortest-weight path.
+    if (head < steps_.size()) {
+      max_head_steps_ = std::max(steps_[head], max_head_steps_);
+      head_steps_ = steps_[head];
+    }
+    return head;
+  }
+
+  void Enqueue(StateId s) override {
+    // We assume that there is an arc between the Head() state and this
+    // Enqueued state.
+    const ssize_t state_steps = head_steps_ + 1;
+    if (s >= steps_.size()) {
+      steps_.resize(s + 1, state_steps);
+    }
+    // This is the number of arcs in the minimum cost path from Start to s.
+    steps_[s] = state_steps;
+    if (state_steps > (max_head_steps_ - threshold_) || threshold_ < 0) {
+      Base::Enqueue(s);
+    }
+  }
+
+ private:
+  // A dense map from StateId to the number of arcs in the minimum weight
+  // path from Start to this state.
+  std::vector<ssize_t> steps_;
+  // We only keep paths that are within this number of arcs (not weight!)
+  // of the longest path.
+  const ssize_t threshold_;
+
+  // The following are mutable because Head() is const.
+  // The number of arcs traversed in the minimum cost path from the start
+  // state to the current Head() state.
+  mutable ssize_t head_steps_;
+  // The maximum number of arcs traversed by any low-cost path so far.
+  mutable ssize_t max_head_steps_;
+};
+
+// Topological-order queue discipline, templated on the StateId. States are
+// ordered in the queue topologically. The FST must be acyclic.
+template <class S>
+class TopOrderQueue : public QueueBase<S> {
+ public:
+  using StateId = S;
+
+  // This constructor computes the topological order. It accepts an arc filter
+  // to limit the transitions considered in that computation (e.g., only the
+  // epsilon graph).
+  template <class Arc, class ArcFilter>
+  TopOrderQueue(const Fst<Arc> &fst, ArcFilter filter)
+      : QueueBase<StateId>(TOP_ORDER_QUEUE),
+        front_(0),
+        back_(kNoStateId),
+        order_(0),
+        state_(0) {
+    bool acyclic;
+    TopOrderVisitor<Arc> top_order_visitor(&order_, &acyclic);
+    DfsVisit(fst, &top_order_visitor, filter);
+    if (!acyclic) {
+      FSTERROR() << "TopOrderQueue: FST is not acyclic";
+      QueueBase<S>::SetError(true);
+    }
+    state_.resize(order_.size(), kNoStateId);
+  }
+
+  // This constructor is passed the pre-computed topological order.
+  explicit TopOrderQueue(const std::vector<StateId> &order)
+      : QueueBase<StateId>(TOP_ORDER_QUEUE),
+        front_(0),
+        back_(kNoStateId),
+        order_(order),
+        state_(order.size(), kNoStateId) {}
+
+  virtual ~TopOrderQueue() = default;
+
+  StateId Head() const final { return state_[front_]; }
+
+  void Enqueue(StateId s) final {
+    if (front_ > back_) {
+      front_ = back_ = order_[s];
+    } else if (order_[s] > back_) {
+      back_ = order_[s];
+    } else if (order_[s] < front_) {
+      front_ = order_[s];
+    }
+    state_[order_[s]] = s;
+  }
+
+  void Dequeue() final {
+    state_[front_] = kNoStateId;
+    while ((front_ <= back_) && (state_[front_] == kNoStateId)) ++front_;
+  }
+
+  void Update(StateId) final {}
+
+  bool Empty() const final { return front_ > back_; }
+
+  void Clear() final {
+    for (StateId s = front_; s <= back_; ++s) state_[s] = kNoStateId;
+    back_ = kNoStateId;
+    front_ = 0;
+  }
+
+ private:
+  StateId front_;
+  StateId back_;
+  std::vector<StateId> order_;
+  std::vector<StateId> state_;
+};
+
+// State order queue discipline, templated on the StateId. States are ordered in
+// the queue by state ID.
+template <class S>
+class StateOrderQueue : public QueueBase<S> {
+ public:
+  using StateId = S;
+
+  StateOrderQueue()
+      : QueueBase<StateId>(STATE_ORDER_QUEUE), front_(0), back_(kNoStateId) {}
+
+  virtual ~StateOrderQueue() = default;
+
+  StateId Head() const final { return front_; }
+
+  void Enqueue(StateId s) final {
+    if (front_ > back_) {
+      front_ = back_ = s;
+    } else if (s > back_) {
+      back_ = s;
+    } else if (s < front_) {
+      front_ = s;
+    }
+    while (enqueued_.size() <= s) enqueued_.push_back(false);
+    enqueued_[s] = true;
+  }
+
+  void Dequeue() final {
+    enqueued_[front_] = false;
+    while ((front_ <= back_) && (enqueued_[front_] == false)) ++front_;
+  }
+
+  void Update(StateId) final {}
+
+  bool Empty() const final { return front_ > back_; }
+
+  void Clear() final {
+    for (StateId i = front_; i <= back_; ++i) enqueued_[i] = false;
+    front_ = 0;
+    back_ = kNoStateId;
+  }
+
+ private:
+  StateId front_;
+  StateId back_;
+  std::vector<bool> enqueued_;
+};
+
+// SCC topological-order meta-queue discipline, templated on the StateId and a
+// queue used inside each SCC. It visits the SCCs of an FST in topological
+// order. Its constructor is passed the queues to to use within an SCC.
+template <class S, class Queue>
+class SccQueue : public QueueBase<S> {
+ public:
+  using StateId = S;
+
+  // Constructor takes a vector specifying the SCC number per state and a
+  // vector giving the queue to use per SCC number.
+  SccQueue(const std::vector<StateId> &scc,
+           std::vector<std::unique_ptr<Queue>> *queue)
+      : QueueBase<StateId>(SCC_QUEUE),
+        queue_(queue),
+        scc_(scc),
+        front_(0),
+        back_(kNoStateId) {}
+
+  virtual ~SccQueue() = default;
+
+  StateId Head() const final {
+    while ((front_ <= back_) &&
+           (((*queue_)[front_] && (*queue_)[front_]->Empty()) ||
+            (((*queue_)[front_] == nullptr) &&
+             ((front_ >= trivial_queue_.size()) ||
+              (trivial_queue_[front_] == kNoStateId))))) {
+      ++front_;
+    }
+    if ((*queue_)[front_]) {
+      return (*queue_)[front_]->Head();
+    } else {
+      return trivial_queue_[front_];
+    }
+  }
+
+  void Enqueue(StateId s) final {
+    if (front_ > back_) {
+      front_ = back_ = scc_[s];
+    } else if (scc_[s] > back_) {
+      back_ = scc_[s];
+    } else if (scc_[s] < front_) {
+      front_ = scc_[s];
+    }
+    if ((*queue_)[scc_[s]]) {
+      (*queue_)[scc_[s]]->Enqueue(s);
+    } else {
+      while (trivial_queue_.size() <= scc_[s]) {
+        trivial_queue_.push_back(kNoStateId);
+      }
+      trivial_queue_[scc_[s]] = s;
+    }
+  }
+
+  void Dequeue() final {
+    if ((*queue_)[front_]) {
+      (*queue_)[front_]->Dequeue();
+    } else if (front_ < trivial_queue_.size()) {
+      trivial_queue_[front_] = kNoStateId;
+    }
+  }
+
+  void Update(StateId s) final {
+    if ((*queue_)[scc_[s]]) (*queue_)[scc_[s]]->Update(s);
+  }
+
+  bool Empty() const final {
+    // Queues SCC number back_ is not empty unless back_ == front_.
+    if (front_ < back_) {
+      return false;
+    } else if (front_ > back_) {
+      return true;
+    } else if ((*queue_)[front_]) {
+      return (*queue_)[front_]->Empty();
+    } else {
+      return (front_ >= trivial_queue_.size()) ||
+             (trivial_queue_[front_] == kNoStateId);
+    }
+  }
+
+  void Clear() final {
+    for (StateId i = front_; i <= back_; ++i) {
+      if ((*queue_)[i]) {
+        (*queue_)[i]->Clear();
+      } else if (i < trivial_queue_.size()) {
+        trivial_queue_[i] = kNoStateId;
+      }
+    }
+    front_ = 0;
+    back_ = kNoStateId;
+  }
+
+ private:
+  std::vector<std::unique_ptr<Queue>> *queue_;
+  const std::vector<StateId> &scc_;
+  mutable StateId front_;
+  StateId back_;
+  std::vector<StateId> trivial_queue_;
+};
+
+// Automatic queue discipline. It selects a queue discipline for a given FST
+// based on its properties.
+template <class S>
+class AutoQueue : public QueueBase<S> {
+ public:
+  using StateId = S;
+
+  // This constructor takes a state distance vector that, if non-null and if
+  // the Weight type has the path property, will entertain the shortest-first
+  // queue using the natural order w.r.t to the distance.
+  template <class Arc, class ArcFilter>
+  AutoQueue(const Fst<Arc> &fst,
+            const std::vector<typename Arc::Weight> *distance, ArcFilter filter)
+      : QueueBase<StateId>(AUTO_QUEUE) {
+    using Weight = typename Arc::Weight;
+    using Less = NaturalLess<Weight>;
+    using Compare = internal::StateWeightCompare<StateId, Less>;
+    // First checks if the FST is known to have these properties.
+    const auto props =
+        fst.Properties(kAcyclic | kCyclic | kTopSorted | kUnweighted, false);
+    if ((props & kTopSorted) || fst.Start() == kNoStateId) {
+      queue_.reset(new StateOrderQueue<StateId>());
+      VLOG(2) << "AutoQueue: using state-order discipline";
+    } else if (props & kAcyclic) {
+      queue_.reset(new TopOrderQueue<StateId>(fst, filter));
+      VLOG(2) << "AutoQueue: using top-order discipline";
+    } else if ((props & kUnweighted) && (Weight::Properties() & kIdempotent)) {
+      queue_.reset(new LifoQueue<StateId>());
+      VLOG(2) << "AutoQueue: using LIFO discipline";
+    } else {
+      uint64 properties;
+      // Decomposes into strongly-connected components.
+      SccVisitor<Arc> scc_visitor(&scc_, nullptr, nullptr, &properties);
+      DfsVisit(fst, &scc_visitor, filter);
+      auto nscc = *std::max_element(scc_.begin(), scc_.end()) + 1;
+      std::vector<QueueType> queue_types(nscc);
+      std::unique_ptr<Less> less;
+      std::unique_ptr<Compare> comp;
+      if (distance && (Weight::Properties() & kPath) == kPath) {
+        less.reset(new Less);
+        comp.reset(new Compare(*distance, *less));
+      }
+      // Finds the queue type to use per SCC.
+      bool unweighted;
+      bool all_trivial;
+      SccQueueType(fst, scc_, &queue_types, filter, less.get(), &all_trivial,
+                   &unweighted);
+      // If unweighted and semiring is idempotent, uses LIFO queue.
+      if (unweighted) {
+        queue_.reset(new LifoQueue<StateId>());
+        VLOG(2) << "AutoQueue: using LIFO discipline";
+        return;
+      }
+      // If all the SCC are trivial, the FST is acyclic and the scc number gives
+      // the topological order.
+      if (all_trivial) {
+        queue_.reset(new TopOrderQueue<StateId>(scc_));
+        VLOG(2) << "AutoQueue: using top-order discipline";
+        return;
+      }
+      VLOG(2) << "AutoQueue: using SCC meta-discipline";
+      queues_.resize(nscc);
+      for (StateId i = 0; i < nscc; ++i) {
+        switch (queue_types[i]) {
+          case TRIVIAL_QUEUE:
+            queues_[i].reset();
+            VLOG(3) << "AutoQueue: SCC #" << i << ": using trivial discipline";
+            break;
+          case SHORTEST_FIRST_QUEUE:
+            queues_[i].reset(
+                new ShortestFirstQueue<StateId, Compare, false>(*comp));
+            VLOG(3) << "AutoQueue: SCC #" << i
+                    << ": using shortest-first discipline";
+            break;
+          case LIFO_QUEUE:
+            queues_[i].reset(new LifoQueue<StateId>());
+            VLOG(3) << "AutoQueue: SCC #" << i << ": using LIFO discipline";
+            break;
+          case FIFO_QUEUE:
+          default:
+            queues_[i].reset(new FifoQueue<StateId>());
+            VLOG(3) << "AutoQueue: SCC #" << i << ": using FIFO discipine";
+            break;
+        }
+      }
+      queue_.reset(new SccQueue<StateId, QueueBase<StateId>>(scc_, &queues_));
+    }
+  }
+
+  virtual ~AutoQueue() = default;
+
+  StateId Head() const final { return queue_->Head(); }
+
+  void Enqueue(StateId s) final { queue_->Enqueue(s); }
+
+  void Dequeue() final { queue_->Dequeue(); }
+
+  void Update(StateId s) final { queue_->Update(s); }
+
+  bool Empty() const final { return queue_->Empty(); }
+
+  void Clear() final { queue_->Clear(); }
+
+ private:
+  template <class Arc, class ArcFilter, class Less>
+  static void SccQueueType(const Fst<Arc> &fst, const std::vector<StateId> &scc,
+                           std::vector<QueueType> *queue_types,
+                           ArcFilter filter, Less *less, bool *all_trivial,
+                           bool *unweighted);
+
+  std::unique_ptr<QueueBase<StateId>> queue_;
+  std::vector<std::unique_ptr<QueueBase<StateId>>> queues_;
+  std::vector<StateId> scc_;
+};
+
+// Examines the states in an FST's strongly connected components and determines
+// which type of queue to use per SCC. Stores result as a vector of QueueTypes
+// which is assumed to have length equal to the number of SCCs. An arc filter
+// is used to limit the transitions considered (e.g., only the epsilon graph).
+// The argument all_trivial is set to true if every queue is the trivial queue.
+// The argument unweighted is set to true if the semiring is idempotent and all
+// the arc weights are equal to Zero() or One().
+template <class StateId>
+template <class Arc, class ArcFilter, class Less>
+void AutoQueue<StateId>::SccQueueType(const Fst<Arc> &fst,
+                                      const std::vector<StateId> &scc,
+                                      std::vector<QueueType> *queue_type,
+                                      ArcFilter filter, Less *less,
+                                      bool *all_trivial, bool *unweighted) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  *all_trivial = true;
+  *unweighted = true;
+  for (StateId i = 0; i < queue_type->size(); ++i) {
+    (*queue_type)[i] = TRIVIAL_QUEUE;
+  }
+  for (StateIterator<Fst<Arc>> sit(fst); !sit.Done(); sit.Next()) {
+    const auto state = sit.Value();
+    for (ArcIterator<Fst<Arc>> ait(fst, state); !ait.Done(); ait.Next()) {
+      const auto &arc = ait.Value();
+      if (!filter(arc)) continue;
+      if (scc[state] == scc[arc.nextstate]) {
+        auto &type = (*queue_type)[scc[state]];
+        if (!less || ((*less)(arc.weight, Weight::One()))) {
+          type = FIFO_QUEUE;
+        } else if ((type == TRIVIAL_QUEUE) || (type == LIFO_QUEUE)) {
+          if (!(Weight::Properties() & kIdempotent) ||
+              (arc.weight != Weight::Zero() && arc.weight != Weight::One())) {
+            type = SHORTEST_FIRST_QUEUE;
+          } else {
+            type = LIFO_QUEUE;
+          }
+        }
+        if (type != TRIVIAL_QUEUE) *all_trivial = false;
+      }
+      if (!(Weight::Properties() & kIdempotent) ||
+          (arc.weight != Weight::Zero() && arc.weight != Weight::One())) {
+        *unweighted = false;
+      }
+    }
+  }
+}
+
+// An A* estimate is a function object that maps from a state ID to an
+// estimate of the shortest distance to the final states.
+
+// A trivial A* estimate, yielding a queue which behaves the same in Dijkstra's
+// algorithm.
+template <typename StateId, typename Weight>
+struct TrivialAStarEstimate {
+  constexpr Weight operator()(StateId) const { return Weight::One(); }
+};
+
+// A non-trivial A* estimate using a vector of the estimated future costs.
+template <typename StateId, typename Weight>
+class NaturalAStarEstimate {
+ public:
+  NaturalAStarEstimate(const std::vector<Weight> &beta) : beta_(beta) {}
+
+  const Weight &operator()(StateId s) const {
+     return (s < beta_.size()) ? beta_[s] : kZero;
+  }
+
+ private:
+  static constexpr Weight kZero = Weight::Zero();
+
+  const std::vector<Weight> &beta_;
+};
+
+template <typename Arc, typename Weight>
+constexpr Weight NaturalAStarEstimate<Arc, Weight>::kZero;
+
+// Given a vector that maps from states to weights representing the shortest
+// distance from the initial state, a comparison function object between
+// weights, and an estimate of the shortest distance to the final states, this
+// class defines a comparison function object between states.
+template <typename S, typename Less, typename Estimate>
+class AStarWeightCompare {
+ public:
+  using StateId = S;
+  using Weight = typename Less::Weight;
+
+  AStarWeightCompare(const std::vector<Weight> &weights, const Less &less,
+                     const Estimate &estimate)
+      : weights_(weights), less_(less), estimate_(estimate) {}
+
+  bool operator()(StateId s1, StateId s2) const {
+    const auto w1 = Times(weights_[s1], estimate_(s1));
+    const auto w2 = Times(weights_[s2], estimate_(s2));
+    return less_(w1, w2);
+  }
+
+  const Estimate &GetEstimate() const { return estimate_; }
+
+ private:
+  const std::vector<Weight> &weights_;
+  const Less &less_;
+  const Estimate &estimate_;
+};
+
+// A* queue discipline templated on StateId, Weight, and Estimate.
+template <typename S, typename Weight, typename Estimate>
+class NaturalAStarQueue : public ShortestFirstQueue<
+          S, AStarWeightCompare<S, NaturalLess<Weight>, Estimate>> {
+ public:
+  using StateId = S;
+  using Compare = AStarWeightCompare<StateId, NaturalLess<Weight>, Estimate>;
+
+  NaturalAStarQueue(const std::vector<Weight> &distance,
+                    const Estimate &estimate)
+      : ShortestFirstQueue<StateId, Compare>(
+            Compare(distance, less_, estimate)) {}
+
+  ~NaturalAStarQueue() = default;
+
+ private:
+  // This is non-static because the constructor for non-idempotent weights will
+  // result in an error.
+  const NaturalLess<Weight> less_{};
+};
+
+// A state equivalence class is a function object that maps from a state ID to
+// an equivalence class (state) ID. The trivial equivalence class maps a state
+// ID to itself.
+template <typename StateId>
+struct TrivialStateEquivClass {
+  StateId operator()(StateId s) const { return s; }
+};
+
+// Distance-based pruning queue discipline: Enqueues a state only when its
+// shortest distance (so far), as specified by distance, is less than (as
+// specified by comp) the shortest distance Times() the threshold to any state
+// in the same equivalence class, as specified by the functor class_func. The
+// underlying queue discipline is specified by queue. The ownership of queue is
+// given to this class.
+//
+// This is not a final class.
+template <typename Queue, typename Less, typename ClassFnc>
+class PruneQueue : public QueueBase<typename Queue::StateId> {
+ public:
+  using StateId = typename Queue::StateId;
+  using Weight = typename Less::Weight;
+
+  PruneQueue(const std::vector<Weight> &distance, Queue *queue,
+             const Less &less, const ClassFnc &class_fnc, Weight threshold)
+      : QueueBase<StateId>(OTHER_QUEUE),
+        distance_(distance),
+        queue_(queue),
+        less_(less),
+        class_fnc_(class_fnc),
+        threshold_(std::move(threshold)) {}
+
+  virtual ~PruneQueue() = default;
+
+  StateId Head() const override { return queue_->Head(); }
+
+  void Enqueue(StateId s) override {
+    const auto c = class_fnc_(s);
+    if (c >= class_distance_.size()) {
+      class_distance_.resize(c + 1, Weight::Zero());
+    }
+    if (less_(distance_[s], class_distance_[c])) {
+      class_distance_[c] = distance_[s];
+    }
+    // Enqueues only if below threshold limit.
+    const auto limit = Times(class_distance_[c], threshold_);
+    if (less_(distance_[s], limit)) queue_->Enqueue(s);
+  }
+
+  void Dequeue() override { queue_->Dequeue(); }
+
+  void Update(StateId s) override {
+    const auto c = class_fnc_(s);
+    if (less_(distance_[s], class_distance_[c])) {
+      class_distance_[c] = distance_[s];
+    }
+    queue_->Update(s);
+  }
+
+  bool Empty() const override { return queue_->Empty(); }
+
+  void Clear() override { queue_->Clear(); }
+
+ private:
+  const std::vector<Weight> &distance_;  // Shortest distance to state.
+  std::unique_ptr<Queue> queue_;
+  const Less &less_;                    // Borrowed reference.
+  const ClassFnc &class_fnc_;           // Equivalence class functor.
+  Weight threshold_;                    // Pruning weight threshold.
+  std::vector<Weight> class_distance_;  // Shortest distance to class.
+};
+
+// Pruning queue discipline (see above) using the weight's natural order for the
+// comparison function. The ownership of the queue argument is given to this
+// class.
+template <typename Queue, typename Weight, typename ClassFnc>
+class NaturalPruneQueue final
+    : public PruneQueue<Queue, NaturalLess<Weight>, ClassFnc> {
+ public:
+  using StateId = typename Queue::StateId;
+
+  NaturalPruneQueue(const std::vector<Weight> &distance, Queue *queue,
+                    const ClassFnc &class_fnc, Weight threshold)
+      : PruneQueue<Queue, NaturalLess<Weight>, ClassFnc>(
+            distance, queue, NaturalLess<Weight>(), class_fnc, threshold) {}
+
+  virtual ~NaturalPruneQueue() = default;
+};
+
+// Filter-based pruning queue discipline: enqueues a state only if allowed by
+// the filter, specified by the state filter functor argument. The underlying
+// queue discipline is specified by the queue argument. The ownership of the
+// queue is given to this class.
+template <typename Queue, typename Filter>
+class FilterQueue : public QueueBase<typename Queue::StateId> {
+ public:
+  using StateId = typename Queue::StateId;
+
+  FilterQueue(Queue *queue, const Filter &filter)
+      : QueueBase<StateId>(OTHER_QUEUE), queue_(queue), filter_(filter) {}
+
+  virtual ~FilterQueue() = default;
+
+  StateId Head() const final { return queue_->Head(); }
+
+  // Enqueues only if allowed by state filter.
+  void Enqueue(StateId s) final {
+    if (filter_(s)) queue_->Enqueue(s);
+  }
+
+  void Dequeue() final { queue_->Dequeue(); }
+
+  void Update(StateId s) final {}
+
+  bool Empty() const final { return queue_->Empty(); }
+
+  void Clear() final { queue_->Clear(); }
+
+ private:
+  std::unique_ptr<Queue> queue_;
+  const Filter &filter_;
+};
+
+}  // namespace fst
+
+#endif  // FST_QUEUE_H_
diff --git a/projects/llm_framework/include/fst/randequivalent.h b/projects/llm_framework/include/fst/randequivalent.h
new file mode 100644
index 00000000..73108b46
--- /dev/null
+++ b/projects/llm_framework/include/fst/randequivalent.h
@@ -0,0 +1,114 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Tests if two FSTS are equivalent by checking if random strings from one FST
+// are transduced the same by both FSTs.
+
+#ifndef FST_RANDEQUIVALENT_H_
+#define FST_RANDEQUIVALENT_H_
+
+#include <fst/log.h>
+
+#include <fst/arcsort.h>
+#include <fst/compose.h>
+#include <fst/project.h>
+#include <fst/randgen.h>
+#include <fst/shortest-distance.h>
+#include <fst/vector-fst.h>
+
+
+namespace fst {
+
+// Test if two FSTs are stochastically equivalent by randomly generating
+// random paths through the FSTs.
+//
+// For each randomly generated path, the algorithm computes for each
+// of the two FSTs the sum of the weights of all the successful paths
+// sharing the same input and output labels as the considered randomly
+// generated path and checks that these two values are within a user-specified
+// delta. Returns optional error value (when FLAGS_error_fatal = false).
+template <class Arc, class ArcSelector>
+bool RandEquivalent(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+                    int32 num_paths, float delta,
+                    const RandGenOptions<ArcSelector> &opts,
+                    bool *error = nullptr) {
+  using Weight = typename Arc::Weight;
+  if (error) *error = false;
+  // Checks that the symbol table are compatible.
+  if (!CompatSymbols(fst1.InputSymbols(), fst2.InputSymbols()) ||
+      !CompatSymbols(fst1.OutputSymbols(), fst2.OutputSymbols())) {
+    FSTERROR() << "RandEquivalent: Input/output symbol tables of 1st "
+               << "argument do not match input/output symbol tables of 2nd "
+               << "argument";
+    if (error) *error = true;
+    return false;
+  }
+  static const ILabelCompare<Arc> icomp;
+  static const OLabelCompare<Arc> ocomp;
+  VectorFst<Arc> sfst1(fst1);
+  VectorFst<Arc> sfst2(fst2);
+  Connect(&sfst1);
+  Connect(&sfst2);
+  ArcSort(&sfst1, icomp);
+  ArcSort(&sfst2, icomp);
+  bool result = true;
+  for (int32 n = 0; n < num_paths; ++n) {
+    VectorFst<Arc> path;
+    const auto &fst = rand() % 2 ? sfst1 : sfst2;  // NOLINT
+    RandGen(fst, &path, opts);
+    VectorFst<Arc> ipath(path);
+    VectorFst<Arc> opath(path);
+    Project(&ipath, PROJECT_INPUT);
+    Project(&opath, PROJECT_OUTPUT);
+    VectorFst<Arc> cfst1, pfst1;
+    Compose(ipath, sfst1, &cfst1);
+    ArcSort(&cfst1, ocomp);
+    Compose(cfst1, opath, &pfst1);
+    // Gives up if there are epsilon cycles in a non-idempotent semiring.
+    if (!(Weight::Properties() & kIdempotent) &&
+        pfst1.Properties(kCyclic, true)) {
+      continue;
+    }
+    const auto sum1 = ShortestDistance(pfst1);
+    VectorFst<Arc> cfst2;
+    Compose(ipath, sfst2, &cfst2);
+    ArcSort(&cfst2, ocomp);
+    VectorFst<Arc> pfst2;
+    Compose(cfst2, opath, &pfst2);
+    // Gives up if there are epsilon cycles in a non-idempotent semiring.
+    if (!(Weight::Properties() & kIdempotent) &&
+        pfst2.Properties(kCyclic, true)) {
+      continue;
+    }
+    const auto sum2 = ShortestDistance(pfst2);
+    if (!ApproxEqual(sum1, sum2, delta)) {
+      VLOG(1) << "Sum1 = " << sum1;
+      VLOG(1) << "Sum2 = " << sum2;
+      result = false;
+      break;
+    }
+  }
+  if (fst1.Properties(kError, false) || fst2.Properties(kError, false)) {
+    if (error) *error = true;
+    return false;
+  }
+  return result;
+}
+
+// Tests if two FSTs are equivalent by randomly generating a nnum_paths paths
+// (no longer than the path_length) using a user-specified seed, optionally
+// indicating an error setting an optional error argument to true.
+template <class Arc>
+bool RandEquivalent(const Fst<Arc> &fst1, const Fst<Arc> &fst2, int32 num_paths,
+                    float delta = kDelta, time_t seed = time(nullptr),
+                    int32 max_length = std::numeric_limits<int32>::max(),
+                    bool *error = nullptr) {
+  const UniformArcSelector<Arc> uniform_selector(seed);
+  const RandGenOptions<UniformArcSelector<Arc>> opts(uniform_selector,
+                                                     max_length);
+  return RandEquivalent(fst1, fst2, num_paths, delta, opts, error);
+}
+
+}  // namespace fst
+
+#endif  // FST_RANDEQUIVALENT_H_
diff --git a/projects/llm_framework/include/fst/randgen.h b/projects/llm_framework/include/fst/randgen.h
new file mode 100644
index 00000000..5bcd9fd0
--- /dev/null
+++ b/projects/llm_framework/include/fst/randgen.h
@@ -0,0 +1,756 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Classes and functions to generate random paths through an FST.
+
+#ifndef FST_RANDGEN_H_
+#define FST_RANDGEN_H_
+
+#include <math.h>
+#include <stddef.h>
+#include <limits>
+#include <map>
+#include <memory>
+#include <random>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/accumulator.h>
+#include <fst/cache.h>
+#include <fst/dfs-visit.h>
+#include <fst/float-weight.h>
+#include <fst/fst-decl.h>
+#include <fst/fst.h>
+#include <fst/mutable-fst.h>
+#include <fst/properties.h>
+#include <fst/util.h>
+#include <fst/weight.h>
+
+namespace fst {
+
+// The RandGenFst class is roughly similar to ArcMapFst in that it takes two
+// template parameters denoting the input and output arc types. However, it also
+// takes an additional template parameter which specifies a sampler object which
+// samples (with replacement) arcs from an FST state. The sampler in turn takes
+// a template parameter for a selector object which actually chooses the arc.
+//
+// Arc selector functors are used to select a random transition given an FST
+// state s, returning a number N such that 0 <= N <= NumArcs(s). If N is
+// NumArcs(s), then the final weight is selected; otherwise the N-th arc is
+// selected. It is assumed these are not applied to any state which is neither
+// final nor has any arcs leaving it.
+
+// Randomly selects a transition using the uniform distribution. This class is
+// not thread-safe.
+template <class Arc>
+class UniformArcSelector {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  // Constructs a selector with a non-deterministic seed.
+  UniformArcSelector() : rand_(std::random_device()()) {}
+  // Constructs a selector with a given seed.
+  explicit UniformArcSelector(uint64 seed) : rand_(seed) {}
+
+  size_t operator()(const Fst<Arc> &fst, StateId s) const {
+    const auto n = fst.NumArcs(s) + (fst.Final(s) != Weight::Zero());
+    return static_cast<size_t>(
+        std::uniform_int_distribution<>(0, n - 1)(rand_));
+  }
+
+ private:
+  mutable std::mt19937_64 rand_;
+};
+
+// Randomly selects a transition w.r.t. the weights treated as negative log
+// probabilities after normalizing for the total weight leaving the state. Zero
+// transitions are disregarded. It assumed that Arc::Weight::Value() accesses
+// the floating point representation of the weight. This class is not
+// thread-safe.
+template <class Arc>
+class LogProbArcSelector {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  // Constructs a selector with a non-deterministic seed.
+  LogProbArcSelector() : seed_(std::random_device()()), rand_(seed_) {}
+  // Constructs a selector with a given seed.
+  explicit LogProbArcSelector(uint64 seed) : seed_(seed), rand_(seed) {}
+
+  size_t operator()(const Fst<Arc> &fst, StateId s) const {
+    // Finds total weight leaving state.
+    auto sum = Log64Weight::Zero();
+    ArcIterator<Fst<Arc>> aiter(fst, s);
+    for (; !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      sum = Plus(sum, to_log_weight_(arc.weight));
+    }
+    sum = Plus(sum, to_log_weight_(fst.Final(s)));
+    const double threshold =
+        std::uniform_real_distribution<>(0, exp(-sum.Value()))(rand_);
+    auto p = Log64Weight::Zero();
+    size_t n = 0;
+    for (aiter.Reset(); !aiter.Done(); aiter.Next(), ++n) {
+      p = Plus(p, to_log_weight_(aiter.Value().weight));
+      if (exp(-p.Value()) > threshold) return n;
+    }
+    return n;
+  }
+
+  uint64 Seed() const { return seed_; }
+
+ protected:
+  Log64Weight ToLogWeight(const Weight &weight) const {
+    return to_log_weight_(weight);
+  }
+
+  std::mt19937_64 &MutableRand() const { return rand_; }
+
+ private:
+  const uint64 seed_;
+  mutable std::mt19937_64 rand_;
+  const WeightConvert<Weight, Log64Weight> to_log_weight_{};
+};
+
+// Useful alias when using StdArc.
+using StdArcSelector = LogProbArcSelector<StdArc>;
+
+// Same as LogProbArcSelector but use CacheLogAccumulator to cache the weight
+// accumulation computations. This class is not thread-safe.
+template <class Arc>
+class FastLogProbArcSelector : public LogProbArcSelector<Arc> {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using LogProbArcSelector<Arc>::MutableRand;
+  using LogProbArcSelector<Arc>::ToLogWeight;
+  using LogProbArcSelector<Arc>::operator();
+
+  // Constructs a selector with a non-deterministic seed.
+  FastLogProbArcSelector() : LogProbArcSelector<Arc>() {}
+  // Constructs a selector with a given seed.
+  explicit FastLogProbArcSelector(uint64 seed) : LogProbArcSelector<Arc>(
+      seed) {}
+
+  size_t operator()(const Fst<Arc> &fst, StateId s,
+                    CacheLogAccumulator<Arc> *accumulator) const {
+    accumulator->SetState(s);
+    ArcIterator<Fst<Arc>> aiter(fst, s);
+    // Finds total weight leaving state.
+    const double sum =
+        ToLogWeight(accumulator->Sum(fst.Final(s), &aiter, 0, fst.NumArcs(s)))
+            .Value();
+    const double r = -log(std::uniform_real_distribution<>(0, 1)(
+        MutableRand()));
+    Weight w = from_log_weight_(r + sum);
+    aiter.Reset();
+    return accumulator->LowerBound(w, &aiter);
+  }
+
+ private:
+  const WeightConvert<Log64Weight, Weight> from_log_weight_{};
+};
+
+// Random path state info maintained by RandGenFst and passed to samplers.
+template <typename Arc>
+struct RandState {
+  using StateId = typename Arc::StateId;
+
+  StateId state_id;  // Current input FST state.
+  size_t nsamples;   // Number of samples to be sampled at this state.
+  size_t length;     // Length of path to this random state.
+  size_t select;     // Previous sample arc selection.
+  const RandState<Arc> *parent;  // Previous random state on this path.
+
+  explicit RandState(StateId state_id, size_t nsamples = 0, size_t length = 0,
+                     size_t select = 0, const RandState<Arc> *parent = nullptr)
+      : state_id(state_id),
+        nsamples(nsamples),
+        length(length),
+        select(select),
+        parent(parent) {}
+
+  RandState() : RandState(kNoStateId) {}
+};
+
+// This class, given an arc selector, samples, with replacement, multiple random
+// transitions from an FST's state. This is a generic version with a
+// straightforward use of the arc selector. Specializations may be defined for
+// arc selectors for greater efficiency or special behavior.
+template <class Arc, class Selector>
+class ArcSampler {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  // The max_length argument may be interpreted (or ignored) by a selector as
+  // it chooses. This generic version interprets this literally.
+  ArcSampler(const Fst<Arc> &fst, const Selector &selector,
+             int32 max_length = std::numeric_limits<int32>::max())
+      : fst_(fst), selector_(selector), max_length_(max_length) {}
+
+  // Allow updating FST argument; pass only if changed.
+  ArcSampler(const ArcSampler<Arc, Selector> &sampler,
+             const Fst<Arc> *fst = nullptr)
+      : fst_(fst ? *fst : sampler.fst_),
+        selector_(sampler.selector_),
+        max_length_(sampler.max_length_) {
+    Reset();
+  }
+
+  // Samples a fixed number of samples from the given state. The length argument
+  // specifies the length of the path to the state. Returns true if the samples
+  // were collected. No samples may be collected if either there are no
+  // transitions leaving the state and the state is non-final, or if the path
+  // length has been exceeded. Iterator members are provided to read the samples
+  // in the order in which they were collected.
+  bool Sample(const RandState<Arc> &rstate) {
+    sample_map_.clear();
+    if ((fst_.NumArcs(rstate.state_id) == 0 &&
+         fst_.Final(rstate.state_id) == Weight::Zero()) ||
+        rstate.length == max_length_) {
+      Reset();
+      return false;
+    }
+    for (size_t i = 0; i < rstate.nsamples; ++i) {
+      ++sample_map_[selector_(fst_, rstate.state_id)];
+    }
+    Reset();
+    return true;
+  }
+
+  // More samples?
+  bool Done() const { return sample_iter_ == sample_map_.end(); }
+
+  // Gets the next sample.
+  void Next() { ++sample_iter_; }
+
+  std::pair<size_t, size_t> Value() const { return *sample_iter_; }
+
+  void Reset() { sample_iter_ = sample_map_.begin(); }
+
+  bool Error() const { return false; }
+
+ private:
+  const Fst<Arc> &fst_;
+  const Selector &selector_;
+  const int32 max_length_;
+
+  // Stores (N, K) as described for Value().
+  std::map<size_t, size_t> sample_map_;
+  std::map<size_t, size_t>::const_iterator sample_iter_;
+
+  ArcSampler<Arc, Selector> &operator=(const ArcSampler &) = delete;
+};
+
+// Samples one sample of num_to_sample dimensions from a multinomial
+// distribution parameterized by a vector of probabilities. The result
+// container should be pre-initialized (e.g., an empty map or a zeroed vector
+// sized the same as the vector of probabilities.
+// probs.size()).
+template <class Result, class RNG>
+void OneMultinomialSample(const std::vector<double> &probs,
+                          size_t num_to_sample, Result *result, RNG *rng) {
+  // Left-over probability mass.
+  double norm = 0;
+  for (double p : probs) norm += p;
+  // Left-over number of samples needed.
+  for (size_t i = 0; i < probs.size(); ++i) {
+    size_t num_sampled = 0;
+    if (probs[i] > 0) {
+      std::binomial_distribution<> d(num_to_sample, probs[i] / norm);
+      num_sampled = d(*rng);
+    }
+    if (num_sampled != 0) (*result)[i] = num_sampled;
+    norm -= probs[i];
+    num_to_sample -= num_sampled;
+  }
+}
+
+// Specialization for FastLogProbArcSelector.
+template <class Arc>
+class ArcSampler<Arc, FastLogProbArcSelector<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Accumulator = CacheLogAccumulator<Arc>;
+  using Selector = FastLogProbArcSelector<Arc>;
+
+  ArcSampler(const Fst<Arc> &fst, const Selector &selector,
+             int32 max_length = std::numeric_limits<int32>::max())
+      : fst_(fst),
+        selector_(selector),
+        max_length_(max_length),
+        accumulator_(new Accumulator()) {
+    accumulator_->Init(fst);
+    rng_.seed(selector_.Seed());
+  }
+
+  ArcSampler(const ArcSampler<Arc, Selector> &sampler,
+             const Fst<Arc> *fst = nullptr)
+      : fst_(fst ? *fst : sampler.fst_),
+        selector_(sampler.selector_),
+        max_length_(sampler.max_length_) {
+    if (fst) {
+      accumulator_.reset(new Accumulator());
+      accumulator_->Init(*fst);
+    } else {  // Shallow copy.
+      accumulator_.reset(new Accumulator(*sampler.accumulator_));
+    }
+  }
+
+  bool Sample(const RandState<Arc> &rstate) {
+    sample_map_.clear();
+    if ((fst_.NumArcs(rstate.state_id) == 0 &&
+         fst_.Final(rstate.state_id) == Weight::Zero()) ||
+        rstate.length == max_length_) {
+      Reset();
+      return false;
+    }
+    if (fst_.NumArcs(rstate.state_id) + 1 < rstate.nsamples) {
+      MultinomialSample(rstate);
+      Reset();
+      return true;
+    }
+    for (size_t i = 0; i < rstate.nsamples; ++i) {
+      ++sample_map_[selector_(fst_, rstate.state_id, accumulator_.get())];
+    }
+    Reset();
+    return true;
+  }
+
+  bool Done() const { return sample_iter_ == sample_map_.end(); }
+
+  void Next() { ++sample_iter_; }
+
+  std::pair<size_t, size_t> Value() const { return *sample_iter_; }
+
+  void Reset() { sample_iter_ = sample_map_.begin(); }
+
+  bool Error() const { return accumulator_->Error(); }
+
+ private:
+  using RNG = std::mt19937;
+
+  // Sample according to the multinomial distribution of rstate.nsamples draws
+  // from p_.
+  void MultinomialSample(const RandState<Arc> &rstate) {
+    p_.clear();
+    for (ArcIterator<Fst<Arc>> aiter(fst_, rstate.state_id); !aiter.Done();
+         aiter.Next()) {
+      p_.push_back(exp(-to_log_weight_(aiter.Value().weight).Value()));
+    }
+    if (fst_.Final(rstate.state_id) != Weight::Zero()) {
+      p_.push_back(exp(-to_log_weight_(fst_.Final(rstate.state_id)).Value()));
+    }
+    if (rstate.nsamples < std::numeric_limits<RNG::result_type>::max()) {
+      OneMultinomialSample(p_, rstate.nsamples, &sample_map_, &rng_);
+    } else {
+      for (size_t i = 0; i < p_.size(); ++i) {
+        sample_map_[i] = ceil(p_[i] * rstate.nsamples);
+      }
+    }
+  }
+
+  const Fst<Arc> &fst_;
+  const Selector &selector_;
+  const int32 max_length_;
+
+  // Stores (N, K) for Value().
+  std::map<size_t, size_t> sample_map_;
+  std::map<size_t, size_t>::const_iterator sample_iter_;
+
+  std::unique_ptr<Accumulator> accumulator_;
+  RNG rng_;                // Random number generator.
+  std::vector<double> p_;  // Multinomial parameters.
+  const WeightConvert<Weight, Log64Weight> to_log_weight_{};
+};
+
+// Options for random path generation with RandGenFst. The template argument is
+// a sampler, typically the class ArcSampler. Ownership of the sampler is taken
+// by RandGenFst.
+template <class Sampler>
+struct RandGenFstOptions : public CacheOptions {
+  Sampler *sampler;          // How to sample transitions at a state.
+  int32 npath;               // Number of paths to generate.
+  bool weighted;             // Is the output tree weighted by path count, or
+                             // is it just an unweighted DAG?
+  bool remove_total_weight;  // Remove total weight when output is weighted.
+
+  RandGenFstOptions(const CacheOptions &opts, Sampler *sampler, int32 npath = 1,
+                    bool weighted = true, bool remove_total_weight = false)
+      : CacheOptions(opts),
+        sampler(sampler),
+        npath(npath),
+        weighted(weighted),
+        remove_total_weight(remove_total_weight) {}
+};
+
+namespace internal {
+
+// Implementation of RandGenFst.
+template <class FromArc, class ToArc, class Sampler>
+class RandGenFstImpl : public CacheImpl<ToArc> {
+ public:
+  using FstImpl<ToArc>::SetType;
+  using FstImpl<ToArc>::SetProperties;
+  using FstImpl<ToArc>::SetInputSymbols;
+  using FstImpl<ToArc>::SetOutputSymbols;
+
+  using CacheBaseImpl<CacheState<ToArc>>::EmplaceArc;
+  using CacheBaseImpl<CacheState<ToArc>>::HasArcs;
+  using CacheBaseImpl<CacheState<ToArc>>::HasFinal;
+  using CacheBaseImpl<CacheState<ToArc>>::HasStart;
+  using CacheBaseImpl<CacheState<ToArc>>::SetArcs;
+  using CacheBaseImpl<CacheState<ToArc>>::SetFinal;
+  using CacheBaseImpl<CacheState<ToArc>>::SetStart;
+
+  using Label = typename FromArc::Label;
+  using StateId = typename FromArc::StateId;
+  using FromWeight = typename FromArc::Weight;
+
+  using ToWeight = typename ToArc::Weight;
+
+  RandGenFstImpl(const Fst<FromArc> &fst,
+                 const RandGenFstOptions<Sampler> &opts)
+      : CacheImpl<ToArc>(opts),
+        fst_(fst.Copy()),
+        sampler_(opts.sampler),
+        npath_(opts.npath),
+        weighted_(opts.weighted),
+        remove_total_weight_(opts.remove_total_weight),
+        superfinal_(kNoLabel) {
+    SetType("randgen");
+    SetProperties(
+        RandGenProperties(fst.Properties(kFstProperties, false), weighted_),
+        kCopyProperties);
+    SetInputSymbols(fst.InputSymbols());
+    SetOutputSymbols(fst.OutputSymbols());
+  }
+
+  RandGenFstImpl(const RandGenFstImpl &impl)
+      : CacheImpl<ToArc>(impl),
+        fst_(impl.fst_->Copy(true)),
+        sampler_(new Sampler(*impl.sampler_, fst_.get())),
+        npath_(impl.npath_),
+        weighted_(impl.weighted_),
+        superfinal_(kNoLabel) {
+    SetType("randgen");
+    SetProperties(impl.Properties(), kCopyProperties);
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+  }
+
+  StateId Start() {
+    if (!HasStart()) {
+      const auto s = fst_->Start();
+      if (s == kNoStateId) return kNoStateId;
+      SetStart(state_table_.size());
+      state_table_.emplace_back(
+          new RandState<FromArc>(s, npath_, 0, 0, nullptr));
+    }
+    return CacheImpl<ToArc>::Start();
+  }
+
+  ToWeight Final(StateId s) {
+    if (!HasFinal(s)) Expand(s);
+    return CacheImpl<ToArc>::Final(s);
+  }
+
+  size_t NumArcs(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<ToArc>::NumArcs(s);
+  }
+
+  size_t NumInputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<ToArc>::NumInputEpsilons(s);
+  }
+
+  size_t NumOutputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<ToArc>::NumOutputEpsilons(s);
+  }
+
+  uint64 Properties() const override { return Properties(kFstProperties); }
+
+  // Sets error if found, and returns other FST impl properties.
+  uint64 Properties(uint64 mask) const override {
+    if ((mask & kError) &&
+        (fst_->Properties(kError, false) || sampler_->Error())) {
+      SetProperties(kError, kError);
+    }
+    return FstImpl<ToArc>::Properties(mask);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<ToArc> *data) {
+    if (!HasArcs(s)) Expand(s);
+    CacheImpl<ToArc>::InitArcIterator(s, data);
+  }
+
+  // Computes the outgoing transitions from a state, creating new destination
+  // states as needed.
+  void Expand(StateId s) {
+    if (s == superfinal_) {
+      SetFinal(s, ToWeight::One());
+      SetArcs(s);
+      return;
+    }
+    SetFinal(s, ToWeight::Zero());
+    const auto &rstate = *state_table_[s];
+    sampler_->Sample(rstate);
+    ArcIterator<Fst<FromArc>> aiter(*fst_, rstate.state_id);
+    const auto narcs = fst_->NumArcs(rstate.state_id);
+    for (; !sampler_->Done(); sampler_->Next()) {
+      const auto &sample_pair = sampler_->Value();
+      const auto pos = sample_pair.first;
+      const auto count = sample_pair.second;
+      double prob = static_cast<double>(count) / rstate.nsamples;
+      if (pos < narcs) {  // Regular transition.
+        aiter.Seek(sample_pair.first);
+        const auto &aarc = aiter.Value();
+        auto weight =
+            weighted_ ? to_weight_(Log64Weight(-log(prob))) : ToWeight::One();
+        EmplaceArc(s, aarc.ilabel, aarc.olabel, std::move(weight),
+                   state_table_.size());
+        auto *nrstate = new RandState<FromArc>(aarc.nextstate, count,
+                                               rstate.length + 1, pos, &rstate);
+        state_table_.emplace_back(nrstate);
+      } else {  // Super-final transition.
+        if (weighted_) {
+          const auto weight =
+              remove_total_weight_
+                  ? to_weight_(Log64Weight(-log(prob)))
+                  : to_weight_(Log64Weight(-log(prob * npath_)));
+          SetFinal(s, weight);
+        } else {
+          if (superfinal_ == kNoLabel) {
+            superfinal_ = state_table_.size();
+            state_table_.emplace_back(
+                new RandState<FromArc>(kNoStateId, 0, 0, 0, nullptr));
+          }
+          for (size_t n = 0; n < count; ++n) {
+            EmplaceArc(s, 0, 0, ToWeight::One(), superfinal_);
+          }
+        }
+      }
+    }
+    SetArcs(s);
+  }
+
+ private:
+  const std::unique_ptr<Fst<FromArc>> fst_;
+  std::unique_ptr<Sampler> sampler_;
+  const int32 npath_;
+  std::vector<std::unique_ptr<RandState<FromArc>>> state_table_;
+  const bool weighted_;
+  bool remove_total_weight_;
+  StateId superfinal_;
+  const WeightConvert<Log64Weight, ToWeight> to_weight_{};
+};
+
+}  // namespace internal
+
+// FST class to randomly generate paths through an FST, with details controlled
+// by RandGenOptionsFst. Output format is a tree weighted by the path count.
+template <class FromArc, class ToArc, class Sampler>
+class RandGenFst
+    : public ImplToFst<internal::RandGenFstImpl<FromArc, ToArc, Sampler>> {
+ public:
+  using Label = typename FromArc::Label;
+  using StateId = typename FromArc::StateId;
+  using Weight = typename FromArc::Weight;
+
+  using Store = DefaultCacheStore<FromArc>;
+  using State = typename Store::State;
+
+  using Impl = internal::RandGenFstImpl<FromArc, ToArc, Sampler>;
+
+  friend class ArcIterator<RandGenFst<FromArc, ToArc, Sampler>>;
+  friend class StateIterator<RandGenFst<FromArc, ToArc, Sampler>>;
+
+  RandGenFst(const Fst<FromArc> &fst, const RandGenFstOptions<Sampler> &opts)
+      : ImplToFst<Impl>(std::make_shared<Impl>(fst, opts)) {}
+
+  // See Fst<>::Copy() for doc.
+  RandGenFst(const RandGenFst<FromArc, ToArc, Sampler> &fst, bool safe = false)
+      : ImplToFst<Impl>(fst, safe) {}
+
+  // Get a copy of this RandGenFst. See Fst<>::Copy() for further doc.
+  RandGenFst<FromArc, ToArc, Sampler> *Copy(bool safe = false) const override {
+    return new RandGenFst<FromArc, ToArc, Sampler>(*this, safe);
+  }
+
+  inline void InitStateIterator(StateIteratorData<ToArc> *data) const override;
+
+  void InitArcIterator(StateId s, ArcIteratorData<ToArc> *data) const override {
+    GetMutableImpl()->InitArcIterator(s, data);
+  }
+
+ private:
+  using ImplToFst<Impl>::GetImpl;
+  using ImplToFst<Impl>::GetMutableImpl;
+
+  RandGenFst &operator=(const RandGenFst &) = delete;
+};
+
+// Specialization for RandGenFst.
+template <class FromArc, class ToArc, class Sampler>
+class StateIterator<RandGenFst<FromArc, ToArc, Sampler>>
+    : public CacheStateIterator<RandGenFst<FromArc, ToArc, Sampler>> {
+ public:
+  explicit StateIterator(const RandGenFst<FromArc, ToArc, Sampler> &fst)
+      : CacheStateIterator<RandGenFst<FromArc, ToArc, Sampler>>(
+            fst, fst.GetMutableImpl()) {}
+};
+
+// Specialization for RandGenFst.
+template <class FromArc, class ToArc, class Sampler>
+class ArcIterator<RandGenFst<FromArc, ToArc, Sampler>>
+    : public CacheArcIterator<RandGenFst<FromArc, ToArc, Sampler>> {
+ public:
+  using StateId = typename FromArc::StateId;
+
+  ArcIterator(const RandGenFst<FromArc, ToArc, Sampler> &fst, StateId s)
+      : CacheArcIterator<RandGenFst<FromArc, ToArc, Sampler>>(
+            fst.GetMutableImpl(), s) {
+    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
+  }
+};
+
+template <class FromArc, class ToArc, class Sampler>
+inline void RandGenFst<FromArc, ToArc, Sampler>::InitStateIterator(
+    StateIteratorData<ToArc> *data) const {
+  data->base = new StateIterator<RandGenFst<FromArc, ToArc, Sampler>>(*this);
+}
+
+// Options for random path generation.
+template <class Selector>
+struct RandGenOptions {
+  const Selector &selector;  // How an arc is selected at a state.
+  int32 max_length;          // Maximum path length.
+  int32 npath;               // Number of paths to generate.
+  bool weighted;             // Is the output tree weighted by path count, or
+                             // is it just an unweighted DAG?
+  bool remove_total_weight;  // Remove total weight when output is weighted?
+
+  explicit RandGenOptions(const Selector &selector,
+                          int32 max_length = std::numeric_limits<int32>::max(),
+                          int32 npath = 1, bool weighted = false,
+                          bool remove_total_weight = false)
+      : selector(selector),
+        max_length(max_length),
+        npath(npath),
+        weighted(weighted),
+        remove_total_weight(remove_total_weight) {}
+};
+
+namespace internal {
+
+template <class FromArc, class ToArc>
+class RandGenVisitor {
+ public:
+  using StateId = typename FromArc::StateId;
+  using Weight = typename FromArc::Weight;
+
+  explicit RandGenVisitor(MutableFst<ToArc> *ofst) : ofst_(ofst) {}
+
+  void InitVisit(const Fst<FromArc> &ifst) {
+    ifst_ = &ifst;
+    ofst_->DeleteStates();
+    ofst_->SetInputSymbols(ifst.InputSymbols());
+    ofst_->SetOutputSymbols(ifst.OutputSymbols());
+    if (ifst.Properties(kError, false)) ofst_->SetProperties(kError, kError);
+    path_.clear();
+  }
+
+  constexpr bool InitState(StateId, StateId) const { return true; }
+
+  bool TreeArc(StateId, const ToArc &arc) {
+    if (ifst_->Final(arc.nextstate) == Weight::Zero()) {
+      path_.push_back(arc);
+    } else {
+      OutputPath();
+    }
+    return true;
+  }
+
+  bool BackArc(StateId, const FromArc &) {
+    FSTERROR() << "RandGenVisitor: cyclic input";
+    ofst_->SetProperties(kError, kError);
+    return false;
+  }
+
+  bool ForwardOrCrossArc(StateId, const FromArc &) {
+    OutputPath();
+    return true;
+  }
+
+  void FinishState(StateId s, StateId p, const FromArc *) {
+    if (p != kNoStateId && ifst_->Final(s) == Weight::Zero()) path_.pop_back();
+  }
+
+  void FinishVisit() {}
+
+ private:
+  void OutputPath() {
+    if (ofst_->Start() == kNoStateId) {
+      const auto start = ofst_->AddState();
+      ofst_->SetStart(start);
+    }
+    auto src = ofst_->Start();
+    for (size_t i = 0; i < path_.size(); ++i) {
+      const auto dest = ofst_->AddState();
+      const ToArc arc(path_[i].ilabel, path_[i].olabel, Weight::One(), dest);
+      ofst_->AddArc(src, arc);
+      src = dest;
+    }
+    ofst_->SetFinal(src, Weight::One());
+  }
+
+  const Fst<FromArc> *ifst_;
+  MutableFst<ToArc> *ofst_;
+  std::vector<ToArc> path_;
+
+  RandGenVisitor(const RandGenVisitor &) = delete;
+  RandGenVisitor &operator=(const RandGenVisitor &) = delete;
+};
+
+}  // namespace internal
+
+// Randomly generate paths through an FST; details controlled by
+// RandGenOptions.
+template <class FromArc, class ToArc, class Selector>
+void RandGen(const Fst<FromArc> &ifst, MutableFst<ToArc> *ofst,
+             const RandGenOptions<Selector> &opts) {
+  using Sampler = ArcSampler<FromArc, Selector>;
+  auto *sampler = new Sampler(ifst, opts.selector, opts.max_length);
+  RandGenFstOptions<Sampler> fopts(CacheOptions(true, 0), sampler, opts.npath,
+                                   opts.weighted, opts.remove_total_weight);
+  RandGenFst<FromArc, ToArc, Sampler> rfst(ifst, fopts);
+  if (opts.weighted) {
+    *ofst = rfst;
+  } else {
+    internal::RandGenVisitor<FromArc, ToArc> rand_visitor(ofst);
+    DfsVisit(rfst, &rand_visitor);
+  }
+}
+
+// Randomly generate a path through an FST with the uniform distribution
+// over the transitions.
+template <class FromArc, class ToArc>
+void RandGen(const Fst<FromArc> &ifst, MutableFst<ToArc> *ofst) {
+  const UniformArcSelector<FromArc> uniform_selector;
+  RandGenOptions<UniformArcSelector<ToArc>> opts(uniform_selector);
+  RandGen(ifst, ofst, opts);
+}
+
+}  // namespace fst
+
+#endif  // FST_RANDGEN_H_
diff --git a/projects/llm_framework/include/fst/rational.h b/projects/llm_framework/include/fst/rational.h
new file mode 100644
index 00000000..184ebf3f
--- /dev/null
+++ b/projects/llm_framework/include/fst/rational.h
@@ -0,0 +1,307 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// An FST implementation and base interface for delayed unions, concatenations,
+// and closures.
+
+#ifndef FST_RATIONAL_H_
+#define FST_RATIONAL_H_
+
+#include <algorithm>
+#include <string>
+#include <vector>
+
+#include <fst/mutable-fst.h>
+#include <fst/replace.h>
+#include <fst/test-properties.h>
+
+
+namespace fst {
+
+using RationalFstOptions = CacheOptions;
+
+// This specifies whether to add the empty string.
+enum ClosureType {
+  CLOSURE_STAR = 0,  // Add the empty string.
+  CLOSURE_PLUS = 1   // Don't add the empty string.
+};
+
+template <class Arc>
+class RationalFst;
+
+template <class Arc>
+void Union(RationalFst<Arc> *fst1, const Fst<Arc> &fst2);
+
+template <class Arc>
+void Concat(RationalFst<Arc> *fst1, const Fst<Arc> &fst2);
+
+template <class Arc>
+void Concat(const Fst<Arc> &fst1, RationalFst<Arc> *fst2);
+
+template <class Arc>
+void Closure(RationalFst<Arc> *fst, ClosureType closure_type);
+
+namespace internal {
+
+// Implementation class for delayed unions, concatenations and closures.
+template <class A>
+class RationalFstImpl : public FstImpl<A> {
+ public:
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FstImpl<Arc>::SetType;
+  using FstImpl<Arc>::SetProperties;
+  using FstImpl<Arc>::WriteHeader;
+  using FstImpl<Arc>::SetInputSymbols;
+  using FstImpl<Arc>::SetOutputSymbols;
+
+  explicit RationalFstImpl(const RationalFstOptions &opts)
+      : nonterminals_(0), replace_options_(opts, 0) {
+    SetType("rational");
+    fst_tuples_.emplace_back(0, nullptr);
+  }
+
+  RationalFstImpl(const RationalFstImpl<Arc> &impl)
+      : rfst_(impl.rfst_),
+        nonterminals_(impl.nonterminals_),
+        replace_(impl.replace_ ? impl.replace_->Copy(true) : nullptr),
+        replace_options_(impl.replace_options_) {
+    SetType("rational");
+    fst_tuples_.reserve(impl.fst_tuples_.size());
+    for (const auto &pair : impl.fst_tuples_) {
+      fst_tuples_.emplace_back(pair.first,
+                               pair.second ? pair.second->Copy(true) : nullptr);
+    }
+  }
+
+  ~RationalFstImpl() override {
+    for (auto &tuple : fst_tuples_) delete tuple.second;
+  }
+
+  StateId Start() { return Replace()->Start(); }
+
+  Weight Final(StateId s) { return Replace()->Final(s); }
+
+  size_t NumArcs(StateId s) { return Replace()->NumArcs(s); }
+
+  size_t NumInputEpsilons(StateId s) { return Replace()->NumInputEpsilons(s); }
+
+  size_t NumOutputEpsilons(StateId s) {
+    return Replace()->NumOutputEpsilons(s);
+  }
+
+  uint64 Properties() const override { return Properties(kFstProperties); }
+
+  // Sets error if found, and returns other FST impl properties.
+  uint64 Properties(uint64 mask) const override {
+    if ((mask & kError) && Replace()->Properties(kError, false)) {
+      SetProperties(kError, kError);
+    }
+    return FstImpl<Arc>::Properties(mask);
+  }
+
+  // Implementation of UnionFst(fst1, fst2).
+  void InitUnion(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {
+    replace_.reset();
+    const auto props1 = fst1.Properties(kFstProperties, false);
+    const auto props2 = fst2.Properties(kFstProperties, false);
+    SetInputSymbols(fst1.InputSymbols());
+    SetOutputSymbols(fst1.OutputSymbols());
+    rfst_.AddState();
+    rfst_.AddState();
+    rfst_.SetStart(0);
+    rfst_.SetFinal(1, Weight::One());
+    rfst_.SetInputSymbols(fst1.InputSymbols());
+    rfst_.SetOutputSymbols(fst1.OutputSymbols());
+    nonterminals_ = 2;
+    rfst_.EmplaceArc(0, 0, -1, Weight::One(), 1);
+    rfst_.EmplaceArc(0, 0, -2, Weight::One(), 1);
+    fst_tuples_.emplace_back(-1, fst1.Copy());
+    fst_tuples_.emplace_back(-2, fst2.Copy());
+    SetProperties(UnionProperties(props1, props2, true), kCopyProperties);
+  }
+
+  // Implementation of ConcatFst(fst1, fst2).
+  void InitConcat(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {
+    replace_.reset();
+    const auto props1 = fst1.Properties(kFstProperties, false);
+    const auto props2 = fst2.Properties(kFstProperties, false);
+    SetInputSymbols(fst1.InputSymbols());
+    SetOutputSymbols(fst1.OutputSymbols());
+    rfst_.AddState();
+    rfst_.AddState();
+    rfst_.AddState();
+    rfst_.SetStart(0);
+    rfst_.SetFinal(2, Weight::One());
+    rfst_.SetInputSymbols(fst1.InputSymbols());
+    rfst_.SetOutputSymbols(fst1.OutputSymbols());
+    nonterminals_ = 2;
+    rfst_.EmplaceArc(0, 0, -1, Weight::One(), 1);
+    rfst_.EmplaceArc(1, 0, -2, Weight::One(), 2);
+    fst_tuples_.emplace_back(-1, fst1.Copy());
+    fst_tuples_.emplace_back(-2, fst2.Copy());
+    SetProperties(ConcatProperties(props1, props2, true), kCopyProperties);
+  }
+
+  // Implementation of ClosureFst(fst, closure_type).
+  void InitClosure(const Fst<Arc> &fst, ClosureType closure_type) {
+    replace_.reset();
+    const auto props = fst.Properties(kFstProperties, false);
+    SetInputSymbols(fst.InputSymbols());
+    SetOutputSymbols(fst.OutputSymbols());
+    if (closure_type == CLOSURE_STAR) {
+      rfst_.AddState();
+      rfst_.SetStart(0);
+      rfst_.SetFinal(0, Weight::One());
+      rfst_.EmplaceArc(0, 0, -1, Weight::One(), 0);
+    } else {
+      rfst_.AddState();
+      rfst_.AddState();
+      rfst_.SetStart(0);
+      rfst_.SetFinal(1, Weight::One());
+      rfst_.EmplaceArc(0, 0, -1, Weight::One(), 1);
+      rfst_.EmplaceArc(1, 0, 0, Weight::One(), 0);
+    }
+    rfst_.SetInputSymbols(fst.InputSymbols());
+    rfst_.SetOutputSymbols(fst.OutputSymbols());
+    fst_tuples_.emplace_back(-1, fst.Copy());
+    nonterminals_ = 1;
+    SetProperties(ClosureProperties(props, closure_type == CLOSURE_STAR, true),
+                  kCopyProperties);
+  }
+
+  // Implementation of Union(Fst &, RationalFst *).
+  void AddUnion(const Fst<Arc> &fst) {
+    replace_.reset();
+    const auto props1 = FstImpl<A>::Properties();
+    const auto props2 = fst.Properties(kFstProperties, false);
+    VectorFst<Arc> afst;
+    afst.AddState();
+    afst.AddState();
+    afst.SetStart(0);
+    afst.SetFinal(1, Weight::One());
+    ++nonterminals_;
+    afst.EmplaceArc(0, 0, -nonterminals_, Weight::One(), 1);
+    Union(&rfst_, afst);
+    fst_tuples_.emplace_back(-nonterminals_, fst.Copy());
+    SetProperties(UnionProperties(props1, props2, true), kCopyProperties);
+  }
+
+  // Implementation of Concat(Fst &, RationalFst *).
+  void AddConcat(const Fst<Arc> &fst, bool append) {
+    replace_.reset();
+    const auto props1 = FstImpl<A>::Properties();
+    const auto props2 = fst.Properties(kFstProperties, false);
+    VectorFst<Arc> afst;
+    afst.AddState();
+    afst.AddState();
+    afst.SetStart(0);
+    afst.SetFinal(1, Weight::One());
+    ++nonterminals_;
+    afst.EmplaceArc(0, 0, -nonterminals_, Weight::One(), 1);
+    if (append) {
+      Concat(&rfst_, afst);
+    } else {
+      Concat(afst, &rfst_);
+    }
+    fst_tuples_.emplace_back(-nonterminals_, fst.Copy());
+    SetProperties(ConcatProperties(props1, props2, true), kCopyProperties);
+  }
+
+  // Implementation of Closure(RationalFst *, closure_type).
+  void AddClosure(ClosureType closure_type) {
+    replace_.reset();
+    const auto props = FstImpl<A>::Properties();
+    Closure(&rfst_, closure_type);
+    SetProperties(ClosureProperties(props, closure_type == CLOSURE_STAR, true),
+                  kCopyProperties);
+  }
+
+  // Returns the underlying ReplaceFst, preserving ownership of the underlying
+  // object.
+  ReplaceFst<Arc> *Replace() const {
+    if (!replace_) {
+      fst_tuples_[0].second = rfst_.Copy();
+      replace_.reset(new ReplaceFst<Arc>(fst_tuples_, replace_options_));
+    }
+    return replace_.get();
+  }
+
+ private:
+  // Rational topology machine, using negative non-terminals.
+  VectorFst<Arc> rfst_;
+  // Number of nonterminals used.
+  Label nonterminals_;
+  // Contains the nonterminals and their corresponding FSTs.
+  mutable std::vector<std::pair<Label, const Fst<Arc> *>> fst_tuples_;
+  // Underlying ReplaceFst.
+  mutable std::unique_ptr<ReplaceFst<Arc>> replace_;
+  const ReplaceFstOptions<Arc> replace_options_;
+};
+
+}  // namespace internal
+
+// Parent class for the delayed rational operations (union, concatenation, and
+// closure). This class attaches interface to implementation and handles
+// reference counting, delegating most methods to ImplToFst.
+template <class A>
+class RationalFst : public ImplToFst<internal::RationalFstImpl<A>> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+
+  using Impl = internal::RationalFstImpl<Arc>;
+
+  friend class StateIterator<RationalFst<Arc>>;
+  friend class ArcIterator<RationalFst<Arc>>;
+  friend void Union<>(RationalFst<Arc> *fst1, const Fst<Arc> &fst2);
+  friend void Concat<>(RationalFst<Arc> *fst1, const Fst<Arc> &fst2);
+  friend void Concat<>(const Fst<Arc> &fst1, RationalFst<Arc> *fst2);
+  friend void Closure<>(RationalFst<Arc> *fst, ClosureType closure_type);
+
+  void InitStateIterator(StateIteratorData<Arc> *data) const override {
+    GetImpl()->Replace()->InitStateIterator(data);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
+    GetImpl()->Replace()->InitArcIterator(s, data);
+  }
+
+ protected:
+  using ImplToFst<Impl>::GetImpl;
+
+  explicit RationalFst(const RationalFstOptions &opts = RationalFstOptions())
+      : ImplToFst<Impl>(std::make_shared<Impl>(opts)) {}
+
+  // See Fst<>::Copy() for doc.
+  RationalFst(const RationalFst<Arc> &fst, bool safe = false)
+      : ImplToFst<Impl>(fst, safe) {}
+
+ private:
+  RationalFst &operator=(const RationalFst &) = delete;
+};
+
+// Specialization for RationalFst.
+template <class Arc>
+class StateIterator<RationalFst<Arc>> : public StateIterator<ReplaceFst<Arc>> {
+ public:
+  explicit StateIterator(const RationalFst<Arc> &fst)
+      : StateIterator<ReplaceFst<Arc>>(*(fst.GetImpl()->Replace())) {}
+};
+
+// Specialization for RationalFst.
+template <class Arc>
+class ArcIterator<RationalFst<Arc>> : public CacheArcIterator<ReplaceFst<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const RationalFst<Arc> &fst, StateId s)
+      : ArcIterator<ReplaceFst<Arc>>(*(fst.GetImpl()->Replace()), s) {}
+};
+
+}  // namespace fst
+
+#endif  // FST_RATIONAL_H_
diff --git a/projects/llm_framework/include/fst/register.h b/projects/llm_framework/include/fst/register.h
new file mode 100644
index 00000000..2d1a6ea7
--- /dev/null
+++ b/projects/llm_framework/include/fst/register.h
@@ -0,0 +1,115 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Classes for registering derived FST for generic reading.
+
+#ifndef FST_REGISTER_H_
+#define FST_REGISTER_H_
+
+#include <string>
+#include <type_traits>
+
+
+#include <fst/compat.h>
+#include <fst/generic-register.h>
+#include <fst/util.h>
+
+
+#include <fst/types.h>
+#include <fst/log.h>
+
+namespace fst {
+
+template <class Arc>
+class Fst;
+
+struct FstReadOptions;
+
+// This class represents a single entry in a FstRegister
+template <class Arc>
+struct FstRegisterEntry {
+  using Reader = Fst<Arc> *(*)(std::istream &istrm, const FstReadOptions &opts);
+  using Converter = Fst<Arc> *(*)(const Fst<Arc> &fst);
+
+  Reader reader;
+  Converter converter;
+
+  explicit FstRegisterEntry(Reader reader = nullptr,
+                            Converter converter = nullptr)
+      : reader(reader), converter(converter) {}
+};
+
+// This class maintains the correspondence between a string describing
+// an FST type, and its reader and converter.
+template <class Arc>
+class FstRegister
+    : public GenericRegister<string, FstRegisterEntry<Arc>, FstRegister<Arc>> {
+ public:
+  using Reader = typename FstRegisterEntry<Arc>::Reader;
+  using Converter = typename FstRegisterEntry<Arc>::Converter;
+
+  const Reader GetReader(const string &type) const {
+    return this->GetEntry(type).reader;
+  }
+
+  const Converter GetConverter(const string &type) const {
+    return this->GetEntry(type).converter;
+  }
+
+ protected:
+  string ConvertKeyToSoFilename(const string &key) const override {
+    string legal_type(key);
+    ConvertToLegalCSymbol(&legal_type);
+    return legal_type + "-fst.so";
+  }
+};
+
+// This class registers an FST type for generic reading and creating.
+// The type must have a default constructor and a copy constructor from
+// Fst<Arc>.
+template <class FST>
+class FstRegisterer : public GenericRegisterer<FstRegister<typename FST::Arc>> {
+ public:
+  using Arc = typename FST::Arc;
+  using Entry = typename FstRegister<Arc>::Entry;
+  using Reader = typename FstRegister<Arc>::Reader;
+
+  FstRegisterer()
+      : GenericRegisterer<FstRegister<typename FST::Arc>>(FST().Type(),
+                                                          BuildEntry()) {}
+
+ private:
+  static Fst<Arc> *ReadGeneric(
+      std::istream &strm, const FstReadOptions &opts) {
+    static_assert(std::is_base_of<Fst<Arc>, FST>::value,
+                  "FST class does not inherit from Fst<Arc>");
+    return FST::Read(strm, opts);
+  }
+
+  static Entry BuildEntry() {
+    return Entry(&ReadGeneric, &FstRegisterer<FST>::Convert);
+  }
+
+  static Fst<Arc> *Convert(const Fst<Arc> &fst) { return new FST(fst); }
+};
+
+// Convenience macro to generate static FstRegisterer instance.
+#define REGISTER_FST(FST, Arc) \
+  static fst::FstRegisterer<FST<Arc>> FST##_##Arc##_registerer
+
+// Converts an FST to the specified type.
+template <class Arc>
+Fst<Arc> *Convert(const Fst<Arc> &fst, const string &fst_type) {
+  auto *reg = FstRegister<Arc>::GetRegister();
+  const auto converter = reg->GetConverter(fst_type);
+  if (!converter) {
+    FSTERROR() << "Fst::Convert: Unknown FST type " << fst_type << " (arc type "
+               << Arc::Type() << ")";
+    return nullptr;
+  }
+  return converter(fst);
+}
+
+}  // namespace fst
+
+#endif  // FST_REGISTER_H_
diff --git a/projects/llm_framework/include/fst/relabel.h b/projects/llm_framework/include/fst/relabel.h
new file mode 100644
index 00000000..0979b077
--- /dev/null
+++ b/projects/llm_framework/include/fst/relabel.h
@@ -0,0 +1,472 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes to relabel an FST (either on input or output).
+
+#ifndef FST_RELABEL_H_
+#define FST_RELABEL_H_
+
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/cache.h>
+#include <fst/test-properties.h>
+
+
+#include <unordered_map>
+
+namespace fst {
+
+// Relabels either the input labels or output labels. The old to
+// new labels are specified using a vector of std::pair<Label, Label>.
+// Any label associations not specified are assumed to be identity
+// mapping. The destination labels must be valid labels (e.g., not kNoLabel).
+template <class Arc>
+void Relabel(
+    MutableFst<Arc> *fst,
+    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
+        &ipairs,
+    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
+        &opairs) {
+  using Label = typename Arc::Label;
+  const auto props = fst->Properties(kFstProperties, false);
+  // Constructs label-to-label maps.
+  const std::unordered_map<Label, Label> input_map(
+      ipairs.begin(), ipairs.end());
+  const std::unordered_map<Label, Label> output_map(
+      opairs.begin(), opairs.end());
+  for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
+       siter.Next()) {
+    for (MutableArcIterator<MutableFst<Arc>> aiter(fst, siter.Value());
+         !aiter.Done(); aiter.Next()) {
+      auto arc = aiter.Value();
+      // Relabels input.
+      auto it = input_map.find(arc.ilabel);
+      if (it != input_map.end()) {
+        if (it->second == kNoLabel) {
+          FSTERROR() << "Input symbol ID " << arc.ilabel
+                     << " missing from target vocabulary";
+          fst->SetProperties(kError, kError);
+          return;
+        }
+        arc.ilabel = it->second;
+      }
+      // Relabels output.
+      it = output_map.find(arc.olabel);
+      if (it != output_map.end()) {
+        if (it->second == kNoLabel) {
+          FSTERROR() << "Output symbol id " << arc.olabel
+                     << " missing from target vocabulary";
+          fst->SetProperties(kError, kError);
+          return;
+        }
+        arc.olabel = it->second;
+      }
+      aiter.SetValue(arc);
+    }
+  }
+  fst->SetProperties(RelabelProperties(props), kFstProperties);
+}
+
+// Relabels either the input labels or output labels. The old to
+// new labels are specified using pairs of old and new symbol tables.
+// The tables must contain (at least) all labels on the appropriate side of the
+// FST. If the 'unknown_i(o)symbol' is non-empty, it is used to label any
+// missing symbol in new_i(o)symbols table.
+template <class Arc>
+void Relabel(MutableFst<Arc> *fst,
+             const SymbolTable *old_isymbols, const SymbolTable *new_isymbols,
+             const string &unknown_isymbol, bool attach_new_isymbols,
+             const SymbolTable *old_osymbols, const SymbolTable *new_osymbols,
+             const string &unknown_osymbol, bool attach_new_osymbols) {
+  using Label = typename Arc::Label;
+  // Constructs vectors of input-side label pairs.
+  std::vector<std::pair<Label, Label>> ipairs;
+  if (old_isymbols && new_isymbols) {
+    size_t num_missing_syms = 0;
+    Label unknown_ilabel = kNoLabel;
+    if (!unknown_isymbol.empty()) {
+      unknown_ilabel = new_isymbols->Find(unknown_isymbol);
+      if (unknown_ilabel == kNoLabel) {
+        VLOG(1) << "Input symbol '" << unknown_isymbol
+                << "' missing from target symbol table";
+        ++num_missing_syms;
+      }
+    }
+
+    for (SymbolTableIterator siter(*old_isymbols); !siter.Done();
+         siter.Next()) {
+      const auto old_index = siter.Value();
+      const auto symbol = siter.Symbol();
+      auto new_index = new_isymbols->Find(siter.Symbol());
+      if (new_index == kNoLabel) {
+        if (unknown_ilabel != kNoLabel) {
+          new_index = unknown_ilabel;
+        } else {
+          VLOG(1) << "Input symbol ID " << old_index << " symbol '" << symbol
+                  << "' missing from target symbol table";
+          ++num_missing_syms;
+        }
+      }
+      ipairs.push_back(std::make_pair(old_index, new_index));
+    }
+    if (num_missing_syms > 0) {
+      LOG(WARNING) << "Target symbol table missing: " << num_missing_syms
+                   << " input symbols";
+    }
+    if (attach_new_isymbols) fst->SetInputSymbols(new_isymbols);
+  }
+  // Constructs vectors of output-side label pairs.
+  std::vector<std::pair<Label, Label>> opairs;
+  if (old_osymbols && new_osymbols) {
+    size_t num_missing_syms = 0;
+    Label unknown_olabel = kNoLabel;
+    if (!unknown_osymbol.empty()) {
+      unknown_olabel = new_osymbols->Find(unknown_osymbol);
+      if (unknown_olabel == kNoLabel) {
+        VLOG(1) << "Output symbol '" << unknown_osymbol
+                << "' missing from target symbol table";
+        ++num_missing_syms;
+      }
+    }
+
+    for (SymbolTableIterator siter(*old_osymbols); !siter.Done();
+         siter.Next()) {
+      const auto old_index = siter.Value();
+      const auto symbol = siter.Symbol();
+      auto new_index = new_osymbols->Find(siter.Symbol());
+      if (new_index == kNoLabel) {
+        if (unknown_olabel != kNoLabel) {
+          new_index = unknown_olabel;
+        } else {
+          VLOG(1) << "Output symbol ID " << old_index << " symbol '" << symbol
+                  << "' missing from target symbol table";
+          ++num_missing_syms;
+        }
+      }
+      opairs.push_back(std::make_pair(old_index, new_index));
+    }
+    if (num_missing_syms > 0) {
+      LOG(WARNING) << "Target symbol table missing: " << num_missing_syms
+                   << " output symbols";
+    }
+    if (attach_new_osymbols) fst->SetOutputSymbols(new_osymbols);
+  }
+  // Calls relabel using vector of relabel pairs.
+  Relabel(fst, ipairs, opairs);
+}
+
+// Same as previous but no special allowance for unknown symbols. Kept
+// for backward compat.
+template <class Arc>
+void Relabel(MutableFst<Arc> *fst, const SymbolTable *old_isymbols,
+             const SymbolTable *new_isymbols, bool attach_new_isymbols,
+             const SymbolTable *old_osymbols, const SymbolTable *new_osymbols,
+             bool attach_new_osymbols) {
+  Relabel(fst,
+          old_isymbols, new_isymbols, "" /* no unknown isymbol */,
+          attach_new_isymbols,
+          old_osymbols, new_osymbols, "" /* no unknown ioymbol */,
+          attach_new_osymbols);
+}
+
+
+// Relabels either the input labels or output labels. The old to
+// new labels are specified using symbol tables. Any label associations not
+// specified are assumed to be identity mapping.
+template <class Arc>
+void Relabel(MutableFst<Arc> *fst, const SymbolTable *new_isymbols,
+             const SymbolTable *new_osymbols) {
+  Relabel(fst, fst->InputSymbols(), new_isymbols, true, fst->OutputSymbols(),
+          new_osymbols, true);
+}
+
+using RelabelFstOptions = CacheOptions;
+
+template <class Arc>
+class RelabelFst;
+
+namespace internal {
+
+// Relabels an FST from one symbol set to another. Relabeling can either be on
+// input or output space. RelabelFst implements a delayed version of the
+// relabel. Arcs are relabeled on the fly and not cached; i.e., each request is
+// recomputed.
+template <class Arc>
+class RelabelFstImpl : public CacheImpl<Arc> {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Store = DefaultCacheStore<Arc>;
+  using State = typename Store::State;
+
+  using FstImpl<Arc>::SetType;
+  using FstImpl<Arc>::SetProperties;
+  using FstImpl<Arc>::WriteHeader;
+  using FstImpl<Arc>::SetInputSymbols;
+  using FstImpl<Arc>::SetOutputSymbols;
+
+  using CacheImpl<Arc>::PushArc;
+  using CacheImpl<Arc>::HasArcs;
+  using CacheImpl<Arc>::HasFinal;
+  using CacheImpl<Arc>::HasStart;
+  using CacheImpl<Arc>::SetArcs;
+  using CacheImpl<Arc>::SetFinal;
+  using CacheImpl<Arc>::SetStart;
+
+  friend class StateIterator<RelabelFst<Arc>>;
+
+  RelabelFstImpl(const Fst<Arc> &fst,
+                 const std::vector<std::pair<Label, Label>> &ipairs,
+                 const std::vector<std::pair<Label, Label>> &opairs,
+                 const RelabelFstOptions &opts)
+      : CacheImpl<Arc>(opts),
+        fst_(fst.Copy()),
+        input_map_(ipairs.begin(), ipairs.end()),
+        output_map_(opairs.begin(), opairs.end()),
+        relabel_input_(!ipairs.empty()),
+        relabel_output_(!opairs.empty()) {
+    SetProperties(RelabelProperties(fst.Properties(kCopyProperties, false)));
+    SetType("relabel");
+  }
+
+  RelabelFstImpl(const Fst<Arc> &fst,
+                 const SymbolTable *old_isymbols,
+                 const SymbolTable *new_isymbols,
+                 const SymbolTable *old_osymbols,
+                 const SymbolTable *new_osymbols,
+                 const RelabelFstOptions &opts)
+      : CacheImpl<Arc>(opts),
+        fst_(fst.Copy()),
+        relabel_input_(false),
+        relabel_output_(false) {
+    SetType("relabel");
+    SetProperties(RelabelProperties(fst.Properties(kCopyProperties, false)));
+    SetInputSymbols(old_isymbols);
+    SetOutputSymbols(old_osymbols);
+    if (old_isymbols && new_isymbols &&
+        old_isymbols->LabeledCheckSum() != new_isymbols->LabeledCheckSum()) {
+      for (SymbolTableIterator siter(*old_isymbols); !siter.Done();
+           siter.Next()) {
+        input_map_[siter.Value()] = new_isymbols->Find(siter.Symbol());
+      }
+      SetInputSymbols(new_isymbols);
+      relabel_input_ = true;
+    }
+    if (old_osymbols && new_osymbols &&
+        old_osymbols->LabeledCheckSum() != new_osymbols->LabeledCheckSum()) {
+      for (SymbolTableIterator siter(*old_osymbols); !siter.Done();
+           siter.Next()) {
+        output_map_[siter.Value()] = new_osymbols->Find(siter.Symbol());
+      }
+      SetOutputSymbols(new_osymbols);
+      relabel_output_ = true;
+    }
+  }
+
+  RelabelFstImpl(const RelabelFstImpl<Arc> &impl)
+      : CacheImpl<Arc>(impl),
+        fst_(impl.fst_->Copy(true)),
+        input_map_(impl.input_map_),
+        output_map_(impl.output_map_),
+        relabel_input_(impl.relabel_input_),
+        relabel_output_(impl.relabel_output_) {
+    SetType("relabel");
+    SetProperties(impl.Properties(), kCopyProperties);
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+  }
+
+  StateId Start() {
+    if (!HasStart()) SetStart(fst_->Start());
+    return CacheImpl<Arc>::Start();
+  }
+
+  Weight Final(StateId s) {
+    if (!HasFinal(s)) SetFinal(s, fst_->Final(s));
+    return CacheImpl<Arc>::Final(s);
+  }
+
+  size_t NumArcs(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<Arc>::NumArcs(s);
+  }
+
+  size_t NumInputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<Arc>::NumInputEpsilons(s);
+  }
+
+  size_t NumOutputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<Arc>::NumOutputEpsilons(s);
+  }
+
+  uint64 Properties() const override { return Properties(kFstProperties); }
+
+  // Sets error if found, and returns other FST impl properties.
+  uint64 Properties(uint64 mask) const override {
+    if ((mask & kError) && fst_->Properties(kError, false)) {
+      SetProperties(kError, kError);
+    }
+    return FstImpl<Arc>::Properties(mask);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
+    if (!HasArcs(s)) Expand(s);
+    CacheImpl<Arc>::InitArcIterator(s, data);
+  }
+
+  void Expand(StateId s) {
+    for (ArcIterator<Fst<Arc>> aiter(*fst_, s); !aiter.Done(); aiter.Next()) {
+      auto arc = aiter.Value();
+      if (relabel_input_) {
+        auto it = input_map_.find(arc.ilabel);
+        if (it != input_map_.end()) arc.ilabel = it->second;
+      }
+      if (relabel_output_) {
+        auto it = output_map_.find(arc.olabel);
+        if (it != output_map_.end()) {
+          arc.olabel = it->second;
+        }
+      }
+      PushArc(s, std::move(arc));
+    }
+    SetArcs(s);
+  }
+
+ private:
+  std::unique_ptr<const Fst<Arc>> fst_;
+
+  std::unordered_map<Label, Label> input_map_;
+  std::unordered_map<Label, Label> output_map_;
+  bool relabel_input_;
+  bool relabel_output_;
+};
+
+}  // namespace internal
+
+// This class attaches interface to implementation and handles
+// reference counting, delegating most methods to ImplToFst.
+template <class A>
+class RelabelFst : public ImplToFst<internal::RelabelFstImpl<A>> {
+ public:
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Store = DefaultCacheStore<Arc>;
+  using State = typename Store::State;
+  using Impl = internal::RelabelFstImpl<Arc>;
+
+  friend class ArcIterator<RelabelFst<A>>;
+  friend class StateIterator<RelabelFst<A>>;
+
+  RelabelFst(const Fst<Arc> &fst,
+             const std::vector<std::pair<Label, Label>> &ipairs,
+             const std::vector<std::pair<Label, Label>> &opairs,
+             const RelabelFstOptions &opts = RelabelFstOptions())
+      : ImplToFst<Impl>(std::make_shared<Impl>(fst, ipairs, opairs, opts)) {}
+
+  RelabelFst(const Fst<Arc> &fst, const SymbolTable *new_isymbols,
+             const SymbolTable *new_osymbols,
+             const RelabelFstOptions &opts = RelabelFstOptions())
+      : ImplToFst<Impl>(
+            std::make_shared<Impl>(fst, fst.InputSymbols(), new_isymbols,
+                                   fst.OutputSymbols(), new_osymbols, opts)) {}
+
+  RelabelFst(const Fst<Arc> &fst, const SymbolTable *old_isymbols,
+             const SymbolTable *new_isymbols, const SymbolTable *old_osymbols,
+             const SymbolTable *new_osymbols,
+             const RelabelFstOptions &opts = RelabelFstOptions())
+      : ImplToFst<Impl>(std::make_shared<Impl>(fst, old_isymbols, new_isymbols,
+                                               old_osymbols, new_osymbols,
+                                               opts)) {}
+
+  // See Fst<>::Copy() for doc.
+  RelabelFst(const RelabelFst<Arc> &fst, bool safe = false)
+      : ImplToFst<Impl>(fst, safe) {}
+
+  // Gets a copy of this RelabelFst. See Fst<>::Copy() for further doc.
+  RelabelFst<Arc> *Copy(bool safe = false) const override {
+    return new RelabelFst<Arc>(*this, safe);
+  }
+
+  void InitStateIterator(StateIteratorData<Arc> *data) const override;
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
+    return GetMutableImpl()->InitArcIterator(s, data);
+  }
+
+ private:
+  using ImplToFst<Impl>::GetImpl;
+  using ImplToFst<Impl>::GetMutableImpl;
+
+  RelabelFst &operator=(const RelabelFst &) = delete;
+};
+
+// Specialization for RelabelFst.
+template <class Arc>
+class StateIterator<RelabelFst<Arc>> : public StateIteratorBase<Arc> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  explicit StateIterator(const RelabelFst<Arc> &fst)
+      : impl_(fst.GetImpl()), siter_(*impl_->fst_), s_(0) {}
+
+  bool Done() const final { return siter_.Done(); }
+
+  StateId Value() const final { return s_; }
+
+  void Next() final {
+    if (!siter_.Done()) {
+      ++s_;
+      siter_.Next();
+    }
+  }
+
+  void Reset() final {
+    s_ = 0;
+    siter_.Reset();
+  }
+
+ private:
+  const internal::RelabelFstImpl<Arc>* impl_;
+  StateIterator<Fst<Arc>> siter_;
+  StateId s_;
+
+  StateIterator(const StateIterator &) = delete;
+  StateIterator &operator=(const StateIterator &) = delete;
+};
+
+// Specialization for RelabelFst.
+template <class Arc>
+class ArcIterator<RelabelFst<Arc>> : public CacheArcIterator<RelabelFst<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const RelabelFst<Arc> &fst, StateId s)
+      : CacheArcIterator<RelabelFst<Arc>>(fst.GetMutableImpl(), s) {
+    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
+  }
+};
+
+template <class Arc>
+inline void RelabelFst<Arc>::InitStateIterator(
+    StateIteratorData<Arc> *data) const {
+  data->base = new StateIterator<RelabelFst<Arc>>(*this);
+}
+
+// Useful alias when using StdArc.
+using StdRelabelFst = RelabelFst<StdArc>;
+
+}  // namespace fst
+
+#endif  // FST_RELABEL_H_
diff --git a/projects/llm_framework/include/fst/replace-util.h b/projects/llm_framework/include/fst/replace-util.h
new file mode 100644
index 00000000..42c69824
--- /dev/null
+++ b/projects/llm_framework/include/fst/replace-util.h
@@ -0,0 +1,629 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Utility classes for the recursive replacement of FSTs (RTNs).
+
+#ifndef FST_REPLACE_UTIL_H_
+#define FST_REPLACE_UTIL_H_
+
+#include <map>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/connect.h>
+#include <fst/mutable-fst.h>
+#include <fst/topsort.h>
+#include <fst/vector-fst.h>
+
+
+namespace fst {
+
+// This specifies what labels to output on the call or return arc. Note that
+// REPLACE_LABEL_INPUT and REPLACE_LABEL_OUTPUT will produce transducers when
+// applied to acceptors.
+enum ReplaceLabelType {
+  // Epsilon labels on both input and output.
+  REPLACE_LABEL_NEITHER = 1,
+  // Non-epsilon labels on input and epsilon on output.
+  REPLACE_LABEL_INPUT = 2,
+  // Epsilon on input and non-epsilon on output.
+  REPLACE_LABEL_OUTPUT = 3,
+  // Non-epsilon labels on both input and output.
+  REPLACE_LABEL_BOTH = 4
+};
+
+// By default ReplaceUtil will copy the input label of the replace arc.
+// The call_label_type and return_label_type options specify how to manage
+// the labels of the call arc and the return arc of the replace FST
+struct ReplaceUtilOptions {
+  int64 root;                          // Root rule for expansion.
+  ReplaceLabelType call_label_type;    // How to label call arc.
+  ReplaceLabelType return_label_type;  // How to label return arc.
+  int64 return_label;                  // Label to put on return arc.
+
+  explicit ReplaceUtilOptions(
+      int64 root = kNoLabel,
+      ReplaceLabelType call_label_type = REPLACE_LABEL_INPUT,
+      ReplaceLabelType return_label_type = REPLACE_LABEL_NEITHER,
+      int64 return_label = 0)
+      : root(root),
+        call_label_type(call_label_type),
+        return_label_type(return_label_type),
+        return_label(return_label) {}
+
+  // For backwards compatibility.
+  ReplaceUtilOptions(int64 root, bool epsilon_replace_arc)
+      : ReplaceUtilOptions(root,
+                           epsilon_replace_arc ? REPLACE_LABEL_NEITHER
+                                               : REPLACE_LABEL_INPUT) {}
+};
+
+// Every non-terminal on a path appears as the first label on that path in every
+// FST associated with a given SCC of the replace dependency graph. This would
+// be true if the SCC were formed from left-linear grammar rules.
+constexpr uint8 kReplaceSCCLeftLinear = 0x01;
+// Every non-terminal on a path appears as the final label on that path in every
+// FST associated with a given SCC of the replace dependency graph. This would
+// be true if the SCC were formed from right-linear grammar rules.
+constexpr uint8 kReplaceSCCRightLinear = 0x02;
+// The SCC in the replace dependency graph has more than one state or a
+// self-loop.
+constexpr uint8 kReplaceSCCNonTrivial = 0x04;
+
+// Defined in replace.h.
+template <class Arc>
+void Replace(
+    const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>> &,
+    MutableFst<Arc> *, const ReplaceUtilOptions &);
+
+// Utility class for the recursive replacement of FSTs (RTNs). The user provides
+// a set of label/FST pairs at construction. These are used by methods for
+// testing cyclic dependencies and connectedness and doing RTN connection and
+// specific FST replacement by label or for various optimization properties. The
+// modified results can be obtained with the GetFstPairs() or
+// GetMutableFstPairs() methods.
+template <class Arc>
+class ReplaceUtil {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FstPair = std::pair<Label, const Fst<Arc> *>;
+  using MutableFstPair = std::pair<Label, MutableFst<Arc> *>;
+  using NonTerminalHash = std::unordered_map<Label, Label>;
+
+  // Constructs from mutable FSTs; FST ownership is given to ReplaceUtil.
+  ReplaceUtil(const std::vector<MutableFstPair> &fst_pairs,
+              const ReplaceUtilOptions &opts);
+
+  // Constructs from FSTs; FST ownership is retained by caller.
+  ReplaceUtil(const std::vector<FstPair> &fst_pairs,
+              const ReplaceUtilOptions &opts);
+
+  // Constructs from ReplaceFst internals; FST ownership is retained by caller.
+  ReplaceUtil(const std::vector<std::unique_ptr<const Fst<Arc>>> &fst_array,
+              const NonTerminalHash &nonterminal_hash,
+              const ReplaceUtilOptions &opts);
+
+  ~ReplaceUtil() {
+    for (Label i = 0; i < fst_array_.size(); ++i) delete fst_array_[i];
+  }
+
+  // True if the non-terminal dependencies are cyclic. Cyclic dependencies will
+  // result in an unexpandable FST.
+  bool CyclicDependencies() const {
+    GetDependencies(false);
+    return depprops_ & kCyclic;
+  }
+
+  // Returns the strongly-connected component ID in the dependency graph of the
+  // replace FSTS.
+  StateId SCC(Label label) const {
+    GetDependencies(false);
+    const auto it = nonterminal_hash_.find(label);
+    if (it == nonterminal_hash_.end()) return kNoStateId;
+    return depscc_[it->second];
+  }
+
+  // Returns properties for the strongly-connected component in the dependency
+  // graph of the replace FSTs. If the SCC is kReplaceSCCLeftLinear or
+  // kReplaceSCCRightLinear, that SCC can be represented as finite-state despite
+  // any cyclic dependencies, but not by the usual replacement operation (see
+  // fst/extensions/pdt/replace.h).
+  uint8 SCCProperties(StateId scc_id) {
+    GetSCCProperties();
+    return depsccprops_[scc_id];
+  }
+
+  // Returns true if no useless FSTs, states or transitions are present in the
+  // RTN.
+  bool Connected() const {
+    GetDependencies(false);
+    uint64 props = kAccessible | kCoAccessible;
+    for (Label i = 0; i < fst_array_.size(); ++i) {
+      if (!fst_array_[i]) continue;
+      if (fst_array_[i]->Properties(props, true) != props || !depaccess_[i]) {
+        return false;
+      }
+    }
+    return true;
+  }
+
+  // Removes useless FSTs, states and transitions from the RTN.
+  void Connect();
+
+  // Replaces FSTs specified by labels, unless there are cyclic dependencies.
+  void ReplaceLabels(const std::vector<Label> &labels);
+
+  // Replaces FSTs that have at most nstates states, narcs arcs and nnonterm
+  // non-terminals (updating in reverse dependency order), unless there are
+  // cyclic dependencies.
+  void ReplaceBySize(size_t nstates, size_t narcs, size_t nnonterms);
+
+  // Replaces singleton FSTS, unless there are cyclic dependencies.
+  void ReplaceTrivial() { ReplaceBySize(2, 1, 1); }
+
+  // Replaces non-terminals that have at most ninstances instances (updating in
+  // dependency order), unless there are cyclic dependencies.
+  void ReplaceByInstances(size_t ninstances);
+
+  // Replaces non-terminals that have only one instance, unless there are cyclic
+  // dependencies.
+  void ReplaceUnique() { ReplaceByInstances(1); }
+
+  // Returns label/FST pairs, retaining FST ownership.
+  void GetFstPairs(std::vector<FstPair> *fst_pairs);
+
+  // Returns label/mutable FST pairs, giving FST ownership over to the caller.
+  void GetMutableFstPairs(std::vector<MutableFstPair> *mutable_fst_pairs);
+
+ private:
+  // FST statistics.
+  struct ReplaceStats {
+    StateId nstates;  // Number of states.
+    StateId nfinal;   // Number of final states.
+    size_t narcs;     // Number of arcs.
+    Label nnonterms;  // Number of non-terminals in FST.
+    size_t nref;      // Number of non-terminal instances referring to this FST.
+    // Number of times that ith FST references this FST
+    std::map<Label, size_t> inref;
+    // Number of times that this FST references the ith FST
+    std::map<Label, size_t> outref;
+
+    ReplaceStats() : nstates(0), nfinal(0), narcs(0), nnonterms(0), nref(0) {}
+  };
+
+  // Checks that Mutable FSTs exists, creating them if necessary.
+  void CheckMutableFsts();
+
+  // Computes the dependency graph for the RTN, computing dependency statistics
+  // if stats is true.
+  void GetDependencies(bool stats) const;
+
+  void ClearDependencies() const {
+    depfst_.DeleteStates();
+    stats_.clear();
+    depprops_ = 0;
+    depsccprops_.clear();
+    have_stats_ = false;
+  }
+
+  // Gets topological order of dependencies, returning false with cyclic input.
+  bool GetTopOrder(const Fst<Arc> &fst, std::vector<Label> *toporder) const;
+
+  // Updates statistics to reflect the replacement of the jth FST.
+  void UpdateStats(Label j);
+
+  // Computes the properties for the strongly-connected component in the
+  // dependency graph of the replace FSTs.
+  void GetSCCProperties() const;
+
+  Label root_label_;                                  // Root non-terminal.
+  Label root_fst_;                                    // Root FST ID.
+  ReplaceLabelType call_label_type_;                  // See Replace().
+  ReplaceLabelType return_label_type_;                // See Replace().
+  int64 return_label_;                                // See Replace().
+  std::vector<const Fst<Arc> *> fst_array_;           // FST per ID.
+  std::vector<MutableFst<Arc> *> mutable_fst_array_;  // Mutable FST per ID.
+  std::vector<Label> nonterminal_array_;              // FST ID to non-terminal.
+  NonTerminalHash nonterminal_hash_;                  // Non-terminal to FST ID.
+  mutable VectorFst<Arc> depfst_;                     // FST ID dependencies.
+  mutable std::vector<StateId> depscc_;               // FST SCC ID.
+  mutable std::vector<bool> depaccess_;               // FST ID accessibility.
+  mutable uint64 depprops_;                           // Dependency FST props.
+  mutable bool have_stats_;                  // Have dependency statistics?
+  mutable std::vector<ReplaceStats> stats_;  // Per-FST statistics.
+  mutable std::vector<uint8> depsccprops_;   // SCC properties.
+  ReplaceUtil(const ReplaceUtil &) = delete;
+  ReplaceUtil &operator=(const ReplaceUtil &) = delete;
+};
+
+template <class Arc>
+ReplaceUtil<Arc>::ReplaceUtil(const std::vector<MutableFstPair> &fst_pairs,
+                              const ReplaceUtilOptions &opts)
+    : root_label_(opts.root),
+      call_label_type_(opts.call_label_type),
+      return_label_type_(opts.return_label_type),
+      return_label_(opts.return_label),
+      depprops_(0),
+      have_stats_(false) {
+  fst_array_.push_back(nullptr);
+  mutable_fst_array_.push_back(nullptr);
+  nonterminal_array_.push_back(kNoLabel);
+  for (const auto &fst_pair : fst_pairs) {
+    const auto label = fst_pair.first;
+    auto *fst = fst_pair.second;
+    nonterminal_hash_[label] = fst_array_.size();
+    nonterminal_array_.push_back(label);
+    fst_array_.push_back(fst);
+    mutable_fst_array_.push_back(fst);
+  }
+  root_fst_ = nonterminal_hash_[root_label_];
+  if (!root_fst_) {
+    FSTERROR() << "ReplaceUtil: No root FST for label: " << root_label_;
+  }
+}
+
+template <class Arc>
+ReplaceUtil<Arc>::ReplaceUtil(const std::vector<FstPair> &fst_pairs,
+                              const ReplaceUtilOptions &opts)
+    : root_label_(opts.root),
+      call_label_type_(opts.call_label_type),
+      return_label_type_(opts.return_label_type),
+      return_label_(opts.return_label),
+      depprops_(0),
+      have_stats_(false) {
+  fst_array_.push_back(nullptr);
+  nonterminal_array_.push_back(kNoLabel);
+  for (const auto &fst_pair : fst_pairs) {
+    const auto label = fst_pair.first;
+    const auto *fst = fst_pair.second;
+    nonterminal_hash_[label] = fst_array_.size();
+    nonterminal_array_.push_back(label);
+    fst_array_.push_back(fst->Copy());
+  }
+  root_fst_ = nonterminal_hash_[root_label_];
+  if (!root_fst_) {
+    FSTERROR() << "ReplaceUtil: No root FST for label: " << root_label_;
+  }
+}
+
+template <class Arc>
+ReplaceUtil<Arc>::ReplaceUtil(
+    const std::vector<std::unique_ptr<const Fst<Arc>>> &fst_array,
+    const NonTerminalHash &nonterminal_hash, const ReplaceUtilOptions &opts)
+    : root_fst_(opts.root),
+      call_label_type_(opts.call_label_type),
+      return_label_type_(opts.return_label_type),
+      return_label_(opts.return_label),
+      nonterminal_array_(fst_array.size()),
+      nonterminal_hash_(nonterminal_hash),
+      depprops_(0),
+      have_stats_(false) {
+  fst_array_.push_back(nullptr);
+  for (size_t i = 1; i < fst_array.size(); ++i) {
+    fst_array_.push_back(fst_array[i]->Copy());
+  }
+  for (auto it = nonterminal_hash.begin(); it != nonterminal_hash.end(); ++it) {
+    nonterminal_array_[it->second] = it->first;
+  }
+  root_label_ = nonterminal_array_[root_fst_];
+}
+
+template <class Arc>
+void ReplaceUtil<Arc>::GetDependencies(bool stats) const {
+  if (depfst_.NumStates() > 0) {
+    if (stats && !have_stats_) {
+      ClearDependencies();
+    } else {
+      return;
+    }
+  }
+  have_stats_ = stats;
+  if (have_stats_) stats_.reserve(fst_array_.size());
+  for (Label i = 0; i < fst_array_.size(); ++i) {
+    depfst_.AddState();
+    depfst_.SetFinal(i, Weight::One());
+    if (have_stats_) stats_.push_back(ReplaceStats());
+  }
+  depfst_.SetStart(root_fst_);
+  // An arc from each state (representing the FST) to the state representing the
+  // FST being replaced
+  for (Label i = 0; i < fst_array_.size(); ++i) {
+    const auto *ifst = fst_array_[i];
+    if (!ifst) continue;
+    for (StateIterator<Fst<Arc>> siter(*ifst); !siter.Done(); siter.Next()) {
+      const auto s = siter.Value();
+      if (have_stats_) {
+        ++stats_[i].nstates;
+        if (ifst->Final(s) != Weight::Zero()) ++stats_[i].nfinal;
+      }
+      for (ArcIterator<Fst<Arc>> aiter(*ifst, s); !aiter.Done();
+           aiter.Next()) {
+        if (have_stats_) ++stats_[i].narcs;
+        const auto &arc = aiter.Value();
+        auto it = nonterminal_hash_.find(arc.olabel);
+        if (it != nonterminal_hash_.end()) {
+          const auto j = it->second;
+          depfst_.EmplaceArc(i, arc.olabel, arc.olabel, Weight::One(), j);
+          if (have_stats_) {
+            ++stats_[i].nnonterms;
+            ++stats_[j].nref;
+            ++stats_[j].inref[i];
+            ++stats_[i].outref[j];
+          }
+        }
+      }
+    }
+  }
+  // Computes accessibility info.
+  SccVisitor<Arc> scc_visitor(&depscc_, &depaccess_, nullptr, &depprops_);
+  DfsVisit(depfst_, &scc_visitor);
+}
+
+template <class Arc>
+void ReplaceUtil<Arc>::UpdateStats(Label j) {
+  if (!have_stats_) {
+    FSTERROR() << "ReplaceUtil::UpdateStats: Stats not available";
+    return;
+  }
+  if (j == root_fst_) return;  // Can't replace root.
+  for (auto in = stats_[j].inref.begin(); in != stats_[j].inref.end(); ++in) {
+    const auto i = in->first;
+    const auto ni = in->second;
+    stats_[i].nstates += stats_[j].nstates * ni;
+    stats_[i].narcs += (stats_[j].narcs + 1) * ni;
+    stats_[i].nnonterms += (stats_[j].nnonterms - 1) * ni;
+    stats_[i].outref.erase(j);
+    for (auto out = stats_[j].outref.begin(); out != stats_[j].outref.end();
+         ++out) {
+      const auto k = out->first;
+      const auto nk = out->second;
+      stats_[i].outref[k] += ni * nk;
+    }
+  }
+  for (auto out = stats_[j].outref.begin(); out != stats_[j].outref.end();
+       ++out) {
+    const auto k = out->first;
+    const auto nk = out->second;
+    stats_[k].nref -= nk;
+    stats_[k].inref.erase(j);
+    for (auto in = stats_[j].inref.begin(); in != stats_[j].inref.end(); ++in) {
+      const auto i = in->first;
+      const auto ni = in->second;
+      stats_[k].inref[i] += ni * nk;
+      stats_[k].nref += ni * nk;
+    }
+  }
+}
+
+template <class Arc>
+void ReplaceUtil<Arc>::CheckMutableFsts() {
+  if (mutable_fst_array_.empty()) {
+    for (Label i = 0; i < fst_array_.size(); ++i) {
+      if (!fst_array_[i]) {
+        mutable_fst_array_.push_back(nullptr);
+      } else {
+        mutable_fst_array_.push_back(new VectorFst<Arc>(*fst_array_[i]));
+        delete fst_array_[i];
+        fst_array_[i] = mutable_fst_array_[i];
+      }
+    }
+  }
+}
+
+template <class Arc>
+void ReplaceUtil<Arc>::Connect() {
+  CheckMutableFsts();
+  static constexpr auto props = kAccessible | kCoAccessible;
+  for (auto *mutable_fst : mutable_fst_array_) {
+    if (!mutable_fst) continue;
+    if (mutable_fst->Properties(props, false) != props) {
+      fst::Connect(mutable_fst);
+    }
+  }
+  GetDependencies(false);
+  for (Label i = 0; i < mutable_fst_array_.size(); ++i) {
+    auto *fst = mutable_fst_array_[i];
+    if (fst && !depaccess_[i]) {
+      delete fst;
+      fst_array_[i] = nullptr;
+      mutable_fst_array_[i] = nullptr;
+    }
+  }
+  ClearDependencies();
+}
+
+template <class Arc>
+bool ReplaceUtil<Arc>::GetTopOrder(const Fst<Arc> &fst,
+                                   std::vector<Label> *toporder) const {
+  // Finds topological order of dependencies.
+  std::vector<StateId> order;
+  bool acyclic = false;
+  TopOrderVisitor<Arc> top_order_visitor(&order, &acyclic);
+  DfsVisit(fst, &top_order_visitor);
+  if (!acyclic) {
+    LOG(WARNING) << "ReplaceUtil::GetTopOrder: Cyclical label dependencies";
+    return false;
+  }
+  toporder->resize(order.size());
+  for (Label i = 0; i < order.size(); ++i) (*toporder)[order[i]] = i;
+  return true;
+}
+
+template <class Arc>
+void ReplaceUtil<Arc>::ReplaceLabels(const std::vector<Label> &labels) {
+  CheckMutableFsts();
+  std::unordered_set<Label> label_set;
+  for (const auto label : labels) {
+    // Can't replace root.
+    if (label != root_label_) label_set.insert(label);
+  }
+  // Finds FST dependencies restricted to the labels requested.
+  GetDependencies(false);
+  VectorFst<Arc> pfst(depfst_);
+  for (StateId i = 0; i < pfst.NumStates(); ++i) {
+    std::vector<Arc> arcs;
+    for (ArcIterator<VectorFst<Arc>> aiter(pfst, i); !aiter.Done();
+         aiter.Next()) {
+      const auto &arc = aiter.Value();
+      const auto label = nonterminal_array_[arc.nextstate];
+      if (label_set.count(label) > 0) arcs.push_back(arc);
+    }
+    pfst.DeleteArcs(i);
+    for (auto &arc : arcs) pfst.AddArc(i, std::move(arc));
+  }
+  std::vector<Label> toporder;
+  if (!GetTopOrder(pfst, &toporder)) {
+    ClearDependencies();
+    return;
+  }
+  // Visits FSTs in reverse topological order of dependencies and performs
+  // replacements.
+  for (Label o = toporder.size() - 1; o >= 0; --o) {
+    std::vector<FstPair> fst_pairs;
+    auto s = toporder[o];
+    for (ArcIterator<VectorFst<Arc>> aiter(pfst, s); !aiter.Done();
+         aiter.Next()) {
+      const auto &arc = aiter.Value();
+      const auto label = nonterminal_array_[arc.nextstate];
+      const auto *fst = fst_array_[arc.nextstate];
+      fst_pairs.push_back(std::make_pair(label, fst));
+    }
+    if (fst_pairs.empty()) continue;
+    const auto label = nonterminal_array_[s];
+    const auto *fst = fst_array_[s];
+    fst_pairs.push_back(std::make_pair(label, fst));
+    const ReplaceUtilOptions opts(label, call_label_type_, return_label_type_,
+                                  return_label_);
+    Replace(fst_pairs, mutable_fst_array_[s], opts);
+  }
+  ClearDependencies();
+}
+
+template <class Arc>
+void ReplaceUtil<Arc>::ReplaceBySize(size_t nstates, size_t narcs,
+                                     size_t nnonterms) {
+  std::vector<Label> labels;
+  GetDependencies(true);
+  std::vector<Label> toporder;
+  if (!GetTopOrder(depfst_, &toporder)) {
+    ClearDependencies();
+    return;
+  }
+  for (Label o = toporder.size() - 1; o >= 0; --o) {
+    const auto j = toporder[o];
+    if (stats_[j].nstates <= nstates && stats_[j].narcs <= narcs &&
+        stats_[j].nnonterms <= nnonterms) {
+      labels.push_back(nonterminal_array_[j]);
+      UpdateStats(j);
+    }
+  }
+  ReplaceLabels(labels);
+}
+
+template <class Arc>
+void ReplaceUtil<Arc>::ReplaceByInstances(size_t ninstances) {
+  std::vector<Label> labels;
+  GetDependencies(true);
+  std::vector<Label> toporder;
+  if (!GetTopOrder(depfst_, &toporder)) {
+    ClearDependencies();
+    return;
+  }
+  for (Label o = 0; o < toporder.size(); ++o) {
+    const auto j = toporder[o];
+    if (stats_[j].nref <= ninstances) {
+      labels.push_back(nonterminal_array_[j]);
+      UpdateStats(j);
+    }
+  }
+  ReplaceLabels(labels);
+}
+
+template <class Arc>
+void ReplaceUtil<Arc>::GetFstPairs(std::vector<FstPair> *fst_pairs) {
+  CheckMutableFsts();
+  fst_pairs->clear();
+  for (Label i = 0; i < fst_array_.size(); ++i) {
+    const auto label = nonterminal_array_[i];
+    const auto *fst = fst_array_[i];
+    if (!fst) continue;
+    fst_pairs->push_back(std::make_pair(label, fst));
+  }
+}
+
+template <class Arc>
+void ReplaceUtil<Arc>::GetMutableFstPairs(
+    std::vector<MutableFstPair> *mutable_fst_pairs) {
+  CheckMutableFsts();
+  mutable_fst_pairs->clear();
+  for (Label i = 0; i < mutable_fst_array_.size(); ++i) {
+    const auto label = nonterminal_array_[i];
+    const auto *fst = mutable_fst_array_[i];
+    if (!fst) continue;
+    mutable_fst_pairs->push_back(std::make_pair(label, fst->Copy()));
+  }
+}
+
+template <class Arc>
+void ReplaceUtil<Arc>::GetSCCProperties() const {
+  if (!depsccprops_.empty()) return;
+  GetDependencies(false);
+  if (depscc_.empty()) return;
+  for (StateId scc = 0; scc < depscc_.size(); ++scc) {
+    depsccprops_.push_back(kReplaceSCCLeftLinear | kReplaceSCCRightLinear);
+  }
+  if (!(depprops_ & kCyclic)) return;  // No cyclic dependencies.
+  // Checks for self-loops in the dependency graph.
+  for (StateId scc = 0; scc < depscc_.size(); ++scc) {
+    for (ArcIterator<Fst<Arc> > aiter(depfst_, scc);
+         !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      if (arc.nextstate == scc) {  // SCC has a self loop.
+        depsccprops_[scc] |= kReplaceSCCNonTrivial;
+      }
+    }
+  }
+  std::vector<bool> depscc_visited(depscc_.size(), false);
+  for (Label i = 0; i < fst_array_.size(); ++i) {
+    const auto *fst = fst_array_[i];
+    if (!fst) continue;
+    const auto depscc = depscc_[i];
+    if (depscc_visited[depscc]) {  // SCC has more than one state.
+      depsccprops_[depscc] |= kReplaceSCCNonTrivial;
+    }
+    depscc_visited[depscc] = true;
+    std::vector<StateId> fstscc;  // SCCs of the current FST.
+    uint64 fstprops;
+    SccVisitor<Arc> scc_visitor(&fstscc, nullptr, nullptr, &fstprops);
+    DfsVisit(*fst, &scc_visitor);
+    for (StateIterator<Fst<Arc>> siter(*fst); !siter.Done(); siter.Next()) {
+      const auto s = siter.Value();
+      for (ArcIterator<Fst<Arc>> aiter(*fst, s); !aiter.Done(); aiter.Next()) {
+        const auto &arc = aiter.Value();
+        auto it = nonterminal_hash_.find(arc.olabel);
+        if (it == nonterminal_hash_.end() || depscc_[it->second] != depscc) {
+          continue;  // Skips if a terminal or a non-terminal not in SCC.
+        }
+        const bool arc_in_cycle = fstscc[s] == fstscc[arc.nextstate];
+        // Left linear iff all non-terminals are initial.
+        if (s != fst->Start() || arc_in_cycle) {
+          depsccprops_[depscc] &= ~kReplaceSCCLeftLinear;
+        }
+        // Right linear iff all non-terminals are final.
+        if (fst->Final(arc.nextstate) == Weight::Zero() || arc_in_cycle) {
+          depsccprops_[depscc] &= ~kReplaceSCCRightLinear;
+        }
+      }
+    }
+  }
+}
+
+}  // namespace fst
+
+#endif  // FST_REPLACE_UTIL_H_
diff --git a/projects/llm_framework/include/fst/replace.h b/projects/llm_framework/include/fst/replace.h
new file mode 100644
index 00000000..8b6b18ca
--- /dev/null
+++ b/projects/llm_framework/include/fst/replace.h
@@ -0,0 +1,1492 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes for the recursive replacement of FSTs.
+
+#ifndef FST_REPLACE_H_
+#define FST_REPLACE_H_
+
+#include <set>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/cache.h>
+#include <fst/expanded-fst.h>
+#include <fst/fst-decl.h>  // For optional argument declarations.
+#include <fst/fst.h>
+#include <fst/matcher.h>
+#include <fst/replace-util.h>
+#include <fst/state-table.h>
+#include <fst/test-properties.h>
+
+namespace fst {
+
+// Replace state tables have the form:
+//
+// template <class Arc, class P>
+// class ReplaceStateTable {
+//  public:
+//   using Label = typename Arc::Label Label;
+//   using StateId = typename Arc::StateId;
+//
+//   using PrefixId = P;
+//   using StateTuple = ReplaceStateTuple<StateId, PrefixId>;
+//   using StackPrefix = ReplaceStackPrefix<Label, StateId>;
+//
+//   // Required constructor.
+//   ReplaceStateTable(
+//       const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_list,
+//       Label root);
+//
+//   // Required copy constructor that does not copy state.
+//   ReplaceStateTable(const ReplaceStateTable<Arc, PrefixId> &table);
+//
+//   // Looks up state ID by tuple, adding it if it doesn't exist.
+//   StateId FindState(const StateTuple &tuple);
+//
+//   // Looks up state tuple by ID.
+//   const StateTuple &Tuple(StateId id) const;
+//
+//   // Lookus up prefix ID by stack prefix, adding it if it doesn't exist.
+//   PrefixId FindPrefixId(const StackPrefix &stack_prefix);
+//
+//  // Looks up stack prefix by ID.
+//  const StackPrefix &GetStackPrefix(PrefixId id) const;
+// };
+
+// Tuple that uniquely defines a state in replace.
+template <class S, class P>
+struct ReplaceStateTuple {
+  using StateId = S;
+  using PrefixId = P;
+
+  ReplaceStateTuple(PrefixId prefix_id = -1, StateId fst_id = kNoStateId,
+                    StateId fst_state = kNoStateId)
+      : prefix_id(prefix_id), fst_id(fst_id), fst_state(fst_state) {}
+
+  PrefixId prefix_id;  // Index in prefix table.
+  StateId fst_id;      // Current FST being walked.
+  StateId fst_state;   // Current state in FST being walked (not to be
+                       // confused with the thse StateId of the combined FST).
+};
+
+// Equality of replace state tuples.
+template <class StateId, class PrefixId>
+inline bool operator==(const ReplaceStateTuple<StateId, PrefixId> &x,
+                       const ReplaceStateTuple<StateId, PrefixId> &y) {
+  return x.prefix_id == y.prefix_id && x.fst_id == y.fst_id &&
+         x.fst_state == y.fst_state;
+}
+
+// Functor returning true for tuples corresponding to states in the root FST.
+template <class StateId, class PrefixId>
+class ReplaceRootSelector {
+ public:
+  bool operator()(const ReplaceStateTuple<StateId, PrefixId> &tuple) const {
+    return tuple.prefix_id == 0;
+  }
+};
+
+// Functor for fingerprinting replace state tuples.
+template <class StateId, class PrefixId>
+class ReplaceFingerprint {
+ public:
+  explicit ReplaceFingerprint(const std::vector<uint64> *size_array)
+      : size_array_(size_array) {}
+
+  uint64 operator()(const ReplaceStateTuple<StateId, PrefixId> &tuple) const {
+    return tuple.prefix_id * size_array_->back() +
+           size_array_->at(tuple.fst_id - 1) + tuple.fst_state;
+  }
+
+ private:
+  const std::vector<uint64> *size_array_;
+};
+
+// Useful when the fst_state uniquely define the tuple.
+template <class StateId, class PrefixId>
+class ReplaceFstStateFingerprint {
+ public:
+  uint64 operator()(const ReplaceStateTuple<StateId, PrefixId> &tuple) const {
+    return tuple.fst_state;
+  }
+};
+
+// A generic hash function for replace state tuples.
+template <typename S, typename P>
+class ReplaceHash {
+ public:
+  size_t operator()(const ReplaceStateTuple<S, P>& t) const {
+    static constexpr size_t prime0 = 7853;
+    static constexpr size_t prime1 = 7867;
+    return t.prefix_id + t.fst_id * prime0 + t.fst_state * prime1;
+  }
+};
+
+// Container for stack prefix.
+template <class Label, class StateId>
+class ReplaceStackPrefix {
+ public:
+  struct PrefixTuple {
+    PrefixTuple(Label fst_id = kNoLabel, StateId nextstate = kNoStateId)
+        : fst_id(fst_id), nextstate(nextstate) {}
+
+    Label fst_id;
+    StateId nextstate;
+  };
+
+  ReplaceStackPrefix() {}
+
+  ReplaceStackPrefix(const ReplaceStackPrefix &other)
+      : prefix_(other.prefix_) {}
+
+  void Push(StateId fst_id, StateId nextstate) {
+    prefix_.push_back(PrefixTuple(fst_id, nextstate));
+  }
+
+  void Pop() { prefix_.pop_back(); }
+
+  const PrefixTuple &Top() const { return prefix_[prefix_.size() - 1]; }
+
+  size_t Depth() const { return prefix_.size(); }
+
+ public:
+  std::vector<PrefixTuple> prefix_;
+};
+
+// Equality stack prefix classes.
+template <class Label, class StateId>
+inline bool operator==(const ReplaceStackPrefix<Label, StateId> &x,
+                       const ReplaceStackPrefix<Label, StateId> &y) {
+  if (x.prefix_.size() != y.prefix_.size()) return false;
+  for (size_t i = 0; i < x.prefix_.size(); ++i) {
+    if (x.prefix_[i].fst_id != y.prefix_[i].fst_id ||
+        x.prefix_[i].nextstate != y.prefix_[i].nextstate) {
+      return false;
+    }
+  }
+  return true;
+}
+
+// Hash function for stack prefix to prefix id.
+template <class Label, class StateId>
+class ReplaceStackPrefixHash {
+ public:
+  size_t operator()(const ReplaceStackPrefix<Label, StateId> &prefix) const {
+    size_t sum = 0;
+    for (const auto &pair : prefix.prefix_) {
+      static constexpr size_t prime = 7863;
+      sum += pair.fst_id + pair.nextstate * prime;
+    }
+    return sum;
+  }
+};
+
+// Replace state tables.
+
+// A two-level state table for replace. Warning: calls CountStates to compute
+// the number of states of each component FST.
+template <class Arc, class P = ssize_t>
+class VectorHashReplaceStateTable {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+
+  using PrefixId = P;
+
+  using StateTuple = ReplaceStateTuple<StateId, PrefixId>;
+  using StateTable =
+      VectorHashStateTable<ReplaceStateTuple<StateId, PrefixId>,
+                           ReplaceRootSelector<StateId, PrefixId>,
+                           ReplaceFstStateFingerprint<StateId, PrefixId>,
+                           ReplaceFingerprint<StateId, PrefixId>>;
+  using StackPrefix = ReplaceStackPrefix<Label, StateId>;
+  using StackPrefixTable =
+      CompactHashBiTable<PrefixId, StackPrefix,
+                         ReplaceStackPrefixHash<Label, StateId>>;
+
+  VectorHashReplaceStateTable(
+      const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_list,
+      Label root)
+      : root_size_(0) {
+    size_array_.push_back(0);
+    for (const auto &fst_pair : fst_list) {
+      if (fst_pair.first == root) {
+        root_size_ = CountStates(*(fst_pair.second));
+        size_array_.push_back(size_array_.back());
+      } else {
+        size_array_.push_back(size_array_.back() +
+                              CountStates(*(fst_pair.second)));
+      }
+    }
+    state_table_.reset(
+        new StateTable(new ReplaceRootSelector<StateId, PrefixId>,
+                       new ReplaceFstStateFingerprint<StateId, PrefixId>,
+                       new ReplaceFingerprint<StateId, PrefixId>(&size_array_),
+                       root_size_, root_size_ + size_array_.back()));
+  }
+
+  VectorHashReplaceStateTable(
+      const VectorHashReplaceStateTable<Arc, PrefixId> &table)
+      : root_size_(table.root_size_),
+        size_array_(table.size_array_),
+        prefix_table_(table.prefix_table_) {
+    state_table_.reset(
+        new StateTable(new ReplaceRootSelector<StateId, PrefixId>,
+                       new ReplaceFstStateFingerprint<StateId, PrefixId>,
+                       new ReplaceFingerprint<StateId, PrefixId>(&size_array_),
+                       root_size_, root_size_ + size_array_.back()));
+  }
+
+  StateId FindState(const StateTuple &tuple) {
+    return state_table_->FindState(tuple);
+  }
+
+  const StateTuple &Tuple(StateId id) const { return state_table_->Tuple(id); }
+
+  PrefixId FindPrefixId(const StackPrefix &prefix) {
+    return prefix_table_.FindId(prefix);
+  }
+
+  const StackPrefix& GetStackPrefix(PrefixId id) const {
+    return prefix_table_.FindEntry(id);
+  }
+
+ private:
+  StateId root_size_;
+  std::vector<uint64> size_array_;
+  std::unique_ptr<StateTable> state_table_;
+  StackPrefixTable prefix_table_;
+};
+
+// Default replace state table.
+template <class Arc, class P /* = size_t */>
+class DefaultReplaceStateTable
+    : public CompactHashStateTable<ReplaceStateTuple<typename Arc::StateId, P>,
+                                   ReplaceHash<typename Arc::StateId, P>> {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+
+  using PrefixId = P;
+  using StateTuple = ReplaceStateTuple<StateId, PrefixId>;
+  using StateTable =
+      CompactHashStateTable<StateTuple, ReplaceHash<StateId, PrefixId>>;
+  using StackPrefix = ReplaceStackPrefix<Label, StateId>;
+  using StackPrefixTable =
+      CompactHashBiTable<PrefixId, StackPrefix,
+                         ReplaceStackPrefixHash<Label, StateId>>;
+
+  using StateTable::FindState;
+  using StateTable::Tuple;
+
+  DefaultReplaceStateTable(
+      const std::vector<std::pair<Label, const Fst<Arc> *>> &, Label) {}
+
+  DefaultReplaceStateTable(const DefaultReplaceStateTable<Arc, PrefixId> &table)
+      : StateTable(), prefix_table_(table.prefix_table_) {}
+
+  PrefixId FindPrefixId(const StackPrefix &prefix) {
+    return prefix_table_.FindId(prefix);
+  }
+
+  const StackPrefix &GetStackPrefix(PrefixId id) const {
+    return prefix_table_.FindEntry(id);
+  }
+
+ private:
+  StackPrefixTable prefix_table_;
+};
+
+// By default ReplaceFst will copy the input label of the replace arc.
+// The call_label_type and return_label_type options specify how to manage
+// the labels of the call arc and the return arc of the replace FST
+template <class Arc, class StateTable = DefaultReplaceStateTable<Arc>,
+          class CacheStore = DefaultCacheStore<Arc>>
+struct ReplaceFstOptions : CacheImplOptions<CacheStore> {
+  using Label = typename Arc::Label;
+
+  // Index of root rule for expansion.
+  Label root;
+  // How to label call arc.
+  ReplaceLabelType call_label_type = REPLACE_LABEL_INPUT;
+  // How to label return arc.
+  ReplaceLabelType return_label_type = REPLACE_LABEL_NEITHER;
+  // Specifies output label to put on call arc; if kNoLabel, use existing label
+  // on call arc. Otherwise, use this field as the output label.
+  Label call_output_label = kNoLabel;
+  // Specifies label to put on return arc.
+  Label return_label = 0;
+  // Take ownership of input FSTs?
+  bool take_ownership = false;
+  // Pointer to optional pre-constructed state table.
+  StateTable *state_table = nullptr;
+
+  explicit ReplaceFstOptions(const CacheImplOptions<CacheStore> &opts,
+                             Label root = kNoLabel)
+      : CacheImplOptions<CacheStore>(opts), root(root) {}
+
+  explicit ReplaceFstOptions(const CacheOptions &opts, Label root = kNoLabel)
+      : CacheImplOptions<CacheStore>(opts), root(root) {}
+
+  // FIXME(kbg): There are too many constructors here. Come up with a consistent
+  // position for call_output_label (probably the very end) so that it is
+  // possible to express all the remaining constructors with a single
+  // default-argument constructor. Also move clients off of the "backwards
+  // compatibility" constructor, for good.
+
+  explicit ReplaceFstOptions(Label root) : root(root) {}
+
+  explicit ReplaceFstOptions(Label root, ReplaceLabelType call_label_type,
+                             ReplaceLabelType return_label_type,
+                             Label return_label)
+      : root(root),
+        call_label_type(call_label_type),
+        return_label_type(return_label_type),
+        return_label(return_label) {}
+
+  explicit ReplaceFstOptions(Label root, ReplaceLabelType call_label_type,
+                             ReplaceLabelType return_label_type,
+                             Label call_output_label, Label return_label)
+      : root(root),
+        call_label_type(call_label_type),
+        return_label_type(return_label_type),
+        call_output_label(call_output_label),
+        return_label(return_label) {}
+
+  explicit ReplaceFstOptions(const ReplaceUtilOptions &opts)
+      : ReplaceFstOptions(opts.root, opts.call_label_type,
+                          opts.return_label_type, opts.return_label) {}
+
+  ReplaceFstOptions() : root(kNoLabel) {}
+
+  // For backwards compatibility.
+  ReplaceFstOptions(int64 root, bool epsilon_replace_arc)
+      : root(root),
+        call_label_type(epsilon_replace_arc ? REPLACE_LABEL_NEITHER
+                                            : REPLACE_LABEL_INPUT),
+        call_output_label(epsilon_replace_arc ? 0 : kNoLabel) {}
+};
+
+
+// Forward declaration.
+template <class Arc, class StateTable, class CacheStore>
+class ReplaceFstMatcher;
+
+template <class Arc>
+using FstList = std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>;
+
+// Returns true if label type on arc results in epsilon input label.
+inline bool EpsilonOnInput(ReplaceLabelType label_type) {
+  return label_type == REPLACE_LABEL_NEITHER ||
+         label_type == REPLACE_LABEL_OUTPUT;
+}
+
+// Returns true if label type on arc results in epsilon input label.
+inline bool EpsilonOnOutput(ReplaceLabelType label_type) {
+  return label_type == REPLACE_LABEL_NEITHER ||
+         label_type == REPLACE_LABEL_INPUT;
+}
+
+// Returns true if for either the call or return arc ilabel != olabel.
+template <class Label>
+bool ReplaceTransducer(ReplaceLabelType call_label_type,
+                       ReplaceLabelType return_label_type,
+                       Label call_output_label) {
+  return call_label_type == REPLACE_LABEL_INPUT ||
+         call_label_type == REPLACE_LABEL_OUTPUT ||
+         (call_label_type == REPLACE_LABEL_BOTH &&
+          call_output_label != kNoLabel) ||
+         return_label_type == REPLACE_LABEL_INPUT ||
+         return_label_type == REPLACE_LABEL_OUTPUT;
+}
+
+template <class Arc>
+uint64 ReplaceFstProperties(typename Arc::Label root_label,
+                            const FstList<Arc> &fst_list,
+                            ReplaceLabelType call_label_type,
+                            ReplaceLabelType return_label_type,
+                            typename Arc::Label call_output_label,
+                            bool *sorted_and_non_empty) {
+  using Label = typename Arc::Label;
+  std::vector<uint64> inprops;
+  bool all_ilabel_sorted = true;
+  bool all_olabel_sorted = true;
+  bool all_non_empty = true;
+  // All nonterminals are negative?
+  bool all_negative = true;
+  // All nonterminals are positive and form a dense range containing 1?
+  bool dense_range = true;
+  Label root_fst_idx = 0;
+  for (Label i = 0; i < fst_list.size(); ++i) {
+    const auto label = fst_list[i].first;
+    if (label >= 0) all_negative = false;
+    if (label > fst_list.size() || label <= 0) dense_range = false;
+    if (label == root_label) root_fst_idx = i;
+    const auto *fst = fst_list[i].second;
+    if (fst->Start() == kNoStateId) all_non_empty = false;
+    if (!fst->Properties(kILabelSorted, false)) all_ilabel_sorted = false;
+    if (!fst->Properties(kOLabelSorted, false)) all_olabel_sorted = false;
+    inprops.push_back(fst->Properties(kCopyProperties, false));
+  }
+  const auto props = ReplaceProperties(
+      inprops, root_fst_idx, EpsilonOnInput(call_label_type),
+      EpsilonOnInput(return_label_type), EpsilonOnOutput(call_label_type),
+      EpsilonOnOutput(return_label_type),
+      ReplaceTransducer(call_label_type, return_label_type, call_output_label),
+      all_non_empty, all_ilabel_sorted, all_olabel_sorted,
+      all_negative || dense_range);
+  const bool sorted = props & (kILabelSorted | kOLabelSorted);
+  *sorted_and_non_empty = all_non_empty && sorted;
+  return props;
+}
+
+namespace internal {
+
+// The replace implementation class supports a dynamic expansion of a recursive
+// transition network represented as label/FST pairs with dynamic replacable
+// arcs.
+template <class Arc, class StateTable, class CacheStore>
+class ReplaceFstImpl
+    : public CacheBaseImpl<typename CacheStore::State, CacheStore> {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using State = typename CacheStore::State;
+  using CacheImpl = CacheBaseImpl<State, CacheStore>;
+  using PrefixId = typename StateTable::PrefixId;
+  using StateTuple = ReplaceStateTuple<StateId, PrefixId>;
+  using StackPrefix = ReplaceStackPrefix<Label, StateId>;
+  using NonTerminalHash = std::unordered_map<Label, Label>;
+
+  using FstImpl<Arc>::SetType;
+  using FstImpl<Arc>::SetProperties;
+  using FstImpl<Arc>::WriteHeader;
+  using FstImpl<Arc>::SetInputSymbols;
+  using FstImpl<Arc>::SetOutputSymbols;
+  using FstImpl<Arc>::InputSymbols;
+  using FstImpl<Arc>::OutputSymbols;
+
+  using CacheImpl::PushArc;
+  using CacheImpl::HasArcs;
+  using CacheImpl::HasFinal;
+  using CacheImpl::HasStart;
+  using CacheImpl::SetArcs;
+  using CacheImpl::SetFinal;
+  using CacheImpl::SetStart;
+
+  friend class ReplaceFstMatcher<Arc, StateTable, CacheStore>;
+
+  ReplaceFstImpl(const FstList<Arc> &fst_list,
+                 const ReplaceFstOptions<Arc, StateTable, CacheStore> &opts)
+      : CacheImpl(opts),
+        call_label_type_(opts.call_label_type),
+        return_label_type_(opts.return_label_type),
+        call_output_label_(opts.call_output_label),
+        return_label_(opts.return_label),
+        state_table_(opts.state_table ? opts.state_table
+                                      : new StateTable(fst_list, opts.root)) {
+    SetType("replace");
+    // If the label is epsilon, then all replace label options are equivalent,
+    // so we set the label types to NEITHER for simplicity.
+    if (call_output_label_ == 0) call_label_type_ = REPLACE_LABEL_NEITHER;
+    if (return_label_ == 0) return_label_type_ = REPLACE_LABEL_NEITHER;
+    if (!fst_list.empty()) {
+      SetInputSymbols(fst_list[0].second->InputSymbols());
+      SetOutputSymbols(fst_list[0].second->OutputSymbols());
+    }
+    fst_array_.push_back(nullptr);
+    for (Label i = 0; i < fst_list.size(); ++i) {
+      const auto label = fst_list[i].first;
+      const auto *fst = fst_list[i].second;
+      nonterminal_hash_[label] = fst_array_.size();
+      nonterminal_set_.insert(label);
+      fst_array_.emplace_back(opts.take_ownership ? fst : fst->Copy());
+      if (i) {
+        if (!CompatSymbols(InputSymbols(), fst->InputSymbols())) {
+          FSTERROR() << "ReplaceFstImpl: Input symbols of FST " << i
+                     << " do not match input symbols of base FST (0th FST)";
+          SetProperties(kError, kError);
+        }
+        if (!CompatSymbols(OutputSymbols(), fst->OutputSymbols())) {
+          FSTERROR() << "ReplaceFstImpl: Output symbols of FST " << i
+                     << " do not match output symbols of base FST (0th FST)";
+          SetProperties(kError, kError);
+        }
+      }
+    }
+    const auto nonterminal = nonterminal_hash_[opts.root];
+    if ((nonterminal == 0) && (fst_array_.size() > 1)) {
+      FSTERROR() << "ReplaceFstImpl: No FST corresponding to root label "
+                 << opts.root << " in the input tuple vector";
+      SetProperties(kError, kError);
+    }
+    root_ = (nonterminal > 0) ? nonterminal : 1;
+    bool all_non_empty_and_sorted = false;
+    SetProperties(ReplaceFstProperties(opts.root, fst_list, call_label_type_,
+                                       return_label_type_, call_output_label_,
+                                       &all_non_empty_and_sorted));
+    // Enables optional caching as long as sorted and all non-empty.
+    always_cache_ = !all_non_empty_and_sorted;
+    VLOG(2) << "ReplaceFstImpl::ReplaceFstImpl: always_cache = "
+            << (always_cache_ ? "true" : "false");
+  }
+
+  ReplaceFstImpl(const ReplaceFstImpl &impl)
+      : CacheImpl(impl),
+        call_label_type_(impl.call_label_type_),
+        return_label_type_(impl.return_label_type_),
+        call_output_label_(impl.call_output_label_),
+        return_label_(impl.return_label_),
+        always_cache_(impl.always_cache_),
+        state_table_(new StateTable(*(impl.state_table_))),
+        nonterminal_set_(impl.nonterminal_set_),
+        nonterminal_hash_(impl.nonterminal_hash_),
+        root_(impl.root_) {
+    SetType("replace");
+    SetProperties(impl.Properties(), kCopyProperties);
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+    fst_array_.reserve(impl.fst_array_.size());
+    fst_array_.emplace_back(nullptr);
+    for (Label i = 1; i < impl.fst_array_.size(); ++i) {
+      fst_array_.emplace_back(impl.fst_array_[i]->Copy(true));
+    }
+  }
+
+  // Computes the dependency graph of the replace class and returns
+  // true if the dependencies are cyclic. Cyclic dependencies will result
+  // in an un-expandable FST.
+  bool CyclicDependencies() const {
+    const ReplaceUtilOptions opts(root_);
+    ReplaceUtil<Arc> replace_util(fst_array_, nonterminal_hash_, opts);
+    return replace_util.CyclicDependencies();
+  }
+
+  StateId Start() {
+    if (!HasStart()) {
+      if (fst_array_.size() == 1) {
+        SetStart(kNoStateId);
+        return kNoStateId;
+      } else {
+        const auto fst_start = fst_array_[root_]->Start();
+        if (fst_start == kNoStateId) return kNoStateId;
+        const auto prefix = GetPrefixId(StackPrefix());
+        const auto start =
+            state_table_->FindState(StateTuple(prefix, root_, fst_start));
+        SetStart(start);
+        return start;
+      }
+    } else {
+      return CacheImpl::Start();
+    }
+  }
+
+  Weight Final(StateId s) {
+    if (HasFinal(s)) return CacheImpl::Final(s);
+    const auto &tuple = state_table_->Tuple(s);
+    auto weight = Weight::Zero();
+    if (tuple.prefix_id == 0) {
+      const auto fst_state = tuple.fst_state;
+      weight = fst_array_[tuple.fst_id]->Final(fst_state);
+    }
+    if (always_cache_ || HasArcs(s)) SetFinal(s, weight);
+    return weight;
+  }
+
+  size_t NumArcs(StateId s) {
+    if (HasArcs(s)) {
+      return CacheImpl::NumArcs(s);
+    } else if (always_cache_) {  // If always caching, expands and caches state.
+      Expand(s);
+      return CacheImpl::NumArcs(s);
+    } else {  // Otherwise computes the number of arcs without expanding.
+      const auto tuple = state_table_->Tuple(s);
+      if (tuple.fst_state == kNoStateId) return 0;
+      auto num_arcs = fst_array_[tuple.fst_id]->NumArcs(tuple.fst_state);
+      if (ComputeFinalArc(tuple, nullptr)) ++num_arcs;
+      return num_arcs;
+    }
+  }
+
+  // Returns whether a given label is a non-terminal.
+  bool IsNonTerminal(Label label) const {
+    if (label < *nonterminal_set_.begin() ||
+        label > *nonterminal_set_.rbegin()) {
+      return false;
+    } else {
+      return nonterminal_hash_.count(label);
+    }
+    // TODO(allauzen): be smarter and take advantage of all_dense or
+    // all_negative. Also use this in ComputeArc. This would require changes to
+    // Replace so that recursing into an empty FST lead to a non co-accessible
+    // state instead of deleting the arc as done currently. The current use
+    // correct, since labels are sorted if all_non_empty is true.
+  }
+
+  size_t NumInputEpsilons(StateId s) {
+    if (HasArcs(s)) {
+      return CacheImpl::NumInputEpsilons(s);
+    } else if (always_cache_ || !Properties(kILabelSorted)) {
+      // If always caching or if the number of input epsilons is too expensive
+      // to compute without caching (i.e., not ilabel-sorted), then expands and
+      // caches state.
+      Expand(s);
+      return CacheImpl::NumInputEpsilons(s);
+    } else {
+      // Otherwise, computes the number of input epsilons without caching.
+      const auto tuple = state_table_->Tuple(s);
+      if (tuple.fst_state == kNoStateId) return 0;
+      size_t num = 0;
+      if (!EpsilonOnInput(call_label_type_)) {
+        // If EpsilonOnInput(c) is false, all input epsilon arcs
+        // are also input epsilons arcs in the underlying machine.
+        num = fst_array_[tuple.fst_id]->NumInputEpsilons(tuple.fst_state);
+      } else {
+        // Otherwise, one need to consider that all non-terminal arcs
+        // in the underlying machine also become input epsilon arc.
+        ArcIterator<Fst<Arc>> aiter(*fst_array_[tuple.fst_id], tuple.fst_state);
+        for (; !aiter.Done() && ((aiter.Value().ilabel == 0) ||
+                                 IsNonTerminal(aiter.Value().olabel));
+             aiter.Next()) {
+          ++num;
+        }
+      }
+      if (EpsilonOnInput(return_label_type_) &&
+          ComputeFinalArc(tuple, nullptr)) {
+        ++num;
+      }
+      return num;
+    }
+  }
+
+  size_t NumOutputEpsilons(StateId s) {
+    if (HasArcs(s)) {
+      return CacheImpl::NumOutputEpsilons(s);
+    } else if (always_cache_ || !Properties(kOLabelSorted)) {
+      // If always caching or if the number of output epsilons is too expensive
+      // to compute without caching (i.e., not olabel-sorted), then expands and
+      // caches state.
+      Expand(s);
+      return CacheImpl::NumOutputEpsilons(s);
+    } else {
+      // Otherwise, computes the number of output epsilons without caching.
+      const auto tuple = state_table_->Tuple(s);
+      if (tuple.fst_state == kNoStateId) return 0;
+      size_t num = 0;
+      if (!EpsilonOnOutput(call_label_type_)) {
+        // If EpsilonOnOutput(c) is false, all output epsilon arcs are also
+        // output epsilons arcs in the underlying machine.
+        num = fst_array_[tuple.fst_id]->NumOutputEpsilons(tuple.fst_state);
+      } else {
+        // Otherwise, one need to consider that all non-terminal arcs in the
+        // underlying machine also become output epsilon arc.
+        ArcIterator<Fst<Arc>> aiter(*fst_array_[tuple.fst_id], tuple.fst_state);
+        for (; !aiter.Done() && ((aiter.Value().olabel == 0) ||
+                                 IsNonTerminal(aiter.Value().olabel));
+             aiter.Next()) {
+          ++num;
+        }
+      }
+      if (EpsilonOnOutput(return_label_type_) &&
+          ComputeFinalArc(tuple, nullptr)) {
+        ++num;
+      }
+      return num;
+    }
+  }
+
+  uint64 Properties() const override { return Properties(kFstProperties); }
+
+  // Sets error if found, and returns other FST impl properties.
+  uint64 Properties(uint64 mask) const override {
+    if (mask & kError) {
+      for (Label i = 1; i < fst_array_.size(); ++i) {
+        if (fst_array_[i]->Properties(kError, false)) {
+          SetProperties(kError, kError);
+        }
+      }
+    }
+    return FstImpl<Arc>::Properties(mask);
+  }
+
+  // Returns the base arc iterator, and if arcs have not been computed yet,
+  // extends and recurses for new arcs.
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
+    if (!HasArcs(s)) Expand(s);
+    CacheImpl::InitArcIterator(s, data);
+    // TODO(allauzen): Set behaviour of generic iterator.
+    // Warning: ArcIterator<ReplaceFst<A>>::InitCache() relies on current
+    // behaviour.
+  }
+
+  // Extends current state (walk arcs one level deep).
+  void Expand(StateId s) {
+    const auto tuple = state_table_->Tuple(s);
+    if (tuple.fst_state == kNoStateId) {  // Local FST is empty.
+      SetArcs(s);
+      return;
+    }
+    ArcIterator<Fst<Arc>> aiter(*fst_array_[tuple.fst_id], tuple.fst_state);
+    Arc arc;
+    // Creates a final arc when needed.
+    if (ComputeFinalArc(tuple, &arc)) PushArc(s, std::move(arc));
+    // Expands all arcs leaving the state.
+    for (; !aiter.Done(); aiter.Next()) {
+      if (ComputeArc(tuple, aiter.Value(), &arc)) PushArc(s, std::move(arc));
+    }
+    SetArcs(s);
+  }
+
+  void Expand(StateId s, const StateTuple &tuple,
+              const ArcIteratorData<Arc> &data) {
+    if (tuple.fst_state == kNoStateId) {  // Local FST is empty.
+      SetArcs(s);
+      return;
+    }
+    ArcIterator<Fst<Arc>> aiter(data);
+    Arc arc;
+    // Creates a final arc when needed.
+    if (ComputeFinalArc(tuple, &arc)) AddArc(s, arc);
+    // Expands all arcs leaving the state.
+    for (; !aiter.Done(); aiter.Next()) {
+      if (ComputeArc(tuple, aiter.Value(), &arc)) AddArc(s, arc);
+    }
+    SetArcs(s);
+  }
+
+  // If acpp is null, only returns true if a final arcp is required, but does
+  // not actually compute it.
+  bool ComputeFinalArc(const StateTuple &tuple, Arc *arcp,
+                       uint32 flags = kArcValueFlags) {
+    const auto fst_state = tuple.fst_state;
+    if (fst_state == kNoStateId) return false;
+    // If state is final, pops the stack.
+    if (fst_array_[tuple.fst_id]->Final(fst_state) != Weight::Zero() &&
+        tuple.prefix_id) {
+      if (arcp) {
+        arcp->ilabel = (EpsilonOnInput(return_label_type_)) ? 0 : return_label_;
+        arcp->olabel =
+            (EpsilonOnOutput(return_label_type_)) ? 0 : return_label_;
+        if (flags & kArcNextStateValue) {
+          const auto &stack = state_table_->GetStackPrefix(tuple.prefix_id);
+          const auto prefix_id = PopPrefix(stack);
+          const auto &top = stack.Top();
+          arcp->nextstate = state_table_->FindState(
+              StateTuple(prefix_id, top.fst_id, top.nextstate));
+        }
+        if (flags & kArcWeightValue) {
+          arcp->weight = fst_array_[tuple.fst_id]->Final(fst_state);
+        }
+      }
+      return true;
+    } else {
+      return false;
+    }
+  }
+
+  // Computes an arc in the FST corresponding to one in the underlying machine.
+  // Returns false if the underlying arc corresponds to no arc in the resulting
+  // FST.
+  bool ComputeArc(const StateTuple &tuple, const Arc &arc, Arc *arcp,
+                  uint32 flags = kArcValueFlags) {
+    if (!EpsilonOnInput(call_label_type_) &&
+        (flags == (flags & (kArcILabelValue | kArcWeightValue)))) {
+      *arcp = arc;
+      return true;
+    }
+    if (arc.olabel == 0 || arc.olabel < *nonterminal_set_.begin() ||
+        arc.olabel > *nonterminal_set_.rbegin()) {  // Expands local FST.
+      const auto nextstate =
+          flags & kArcNextStateValue
+              ? state_table_->FindState(
+                    StateTuple(tuple.prefix_id, tuple.fst_id, arc.nextstate))
+              : kNoStateId;
+      *arcp = Arc(arc.ilabel, arc.olabel, arc.weight, nextstate);
+    } else {
+      // Checks for non-terminal.
+      const auto it = nonterminal_hash_.find(arc.olabel);
+      if (it != nonterminal_hash_.end()) {  // Recurses into non-terminal.
+        const auto nonterminal = it->second;
+        const auto nt_prefix =
+            PushPrefix(state_table_->GetStackPrefix(tuple.prefix_id),
+                       tuple.fst_id, arc.nextstate);
+        // If the start state is valid, replace; othewise, the arc is implicitly
+        // deleted.
+        const auto nt_start = fst_array_[nonterminal]->Start();
+        if (nt_start != kNoStateId) {
+          const auto nt_nextstate = flags & kArcNextStateValue
+                                        ? state_table_->FindState(StateTuple(
+                                              nt_prefix, nonterminal, nt_start))
+                                        : kNoStateId;
+          const auto ilabel =
+              (EpsilonOnInput(call_label_type_)) ? 0 : arc.ilabel;
+          const auto olabel =
+              (EpsilonOnOutput(call_label_type_))
+                  ? 0
+                  : ((call_output_label_ == kNoLabel) ? arc.olabel
+                                                      : call_output_label_);
+          *arcp = Arc(ilabel, olabel, arc.weight, nt_nextstate);
+        } else {
+          return false;
+        }
+      } else {
+        const auto nextstate =
+            flags & kArcNextStateValue
+                ? state_table_->FindState(
+                      StateTuple(tuple.prefix_id, tuple.fst_id, arc.nextstate))
+                : kNoStateId;
+        *arcp = Arc(arc.ilabel, arc.olabel, arc.weight, nextstate);
+      }
+    }
+    return true;
+  }
+
+  // Returns the arc iterator flags supported by this FST.
+  uint32 ArcIteratorFlags() const {
+    uint32 flags = kArcValueFlags;
+    if (!always_cache_) flags |= kArcNoCache;
+    return flags;
+  }
+
+  StateTable *GetStateTable() const { return state_table_.get(); }
+
+  const Fst<Arc> *GetFst(Label fst_id) const {
+    return fst_array_[fst_id].get();
+  }
+
+  Label GetFstId(Label nonterminal) const {
+    const auto it = nonterminal_hash_.find(nonterminal);
+    if (it == nonterminal_hash_.end()) {
+      FSTERROR() << "ReplaceFstImpl::GetFstId: Nonterminal not found: "
+                 << nonterminal;
+    }
+    return it->second;
+  }
+
+  // Returns true if label type on call arc results in epsilon input label.
+  bool EpsilonOnCallInput() { return EpsilonOnInput(call_label_type_); }
+
+ private:
+  // The unique index into stack prefix table.
+  PrefixId GetPrefixId(const StackPrefix &prefix) {
+    return state_table_->FindPrefixId(prefix);
+  }
+
+  // The prefix ID after a stack pop.
+  PrefixId PopPrefix(StackPrefix prefix) {
+    prefix.Pop();
+    return GetPrefixId(prefix);
+  }
+
+  // The prefix ID after a stack push.
+  PrefixId PushPrefix(StackPrefix prefix, Label fst_id, StateId nextstate) {
+    prefix.Push(fst_id, nextstate);
+    return GetPrefixId(prefix);
+  }
+
+  // Runtime options
+  ReplaceLabelType call_label_type_;    // How to label call arc.
+  ReplaceLabelType return_label_type_;  // How to label return arc.
+  int64 call_output_label_;  // Specifies output label to put on call arc
+  int64 return_label_;       // Specifies label to put on return arc.
+  bool always_cache_;        // Disable optional caching of arc iterator?
+
+  // State table.
+  std::unique_ptr<StateTable> state_table_;
+
+  // Replace components.
+  std::set<Label> nonterminal_set_;
+  NonTerminalHash nonterminal_hash_;
+  std::vector<std::unique_ptr<const Fst<Arc>>> fst_array_;
+  Label root_;
+};
+
+}  // namespace internal
+
+//
+// ReplaceFst supports dynamic replacement of arcs in one FST with another FST.
+// This replacement is recursive. ReplaceFst can be used to support a variety of
+// delayed constructions such as recursive
+// transition networks, union, or closure. It is constructed with an array of
+// FST(s). One FST represents the root (or topology) machine. The root FST
+// refers to other FSTs by recursively replacing arcs labeled as non-terminals
+// with the matching non-terminal FST. Currently the ReplaceFst uses the output
+// symbols of the arcs to determine whether the arc is a non-terminal arc or
+// not. A non-terminal can be any label that is not a non-zero terminal label in
+// the output alphabet.
+//
+// Note that the constructor uses a vector of pairs. These correspond to the
+// tuple of non-terminal Label and corresponding FST. For example to implement
+// the closure operation we need 2 FSTs. The first root FST is a single
+// self-loop arc on the start state.
+//
+// The ReplaceFst class supports an optionally caching arc iterator.
+//
+// The ReplaceFst needs to be built such that it is known to be ilabel- or
+// olabel-sorted (see usage below).
+//
+// Observe that Matcher<Fst<A>> will use the optionally caching arc iterator
+// when available (the FST is ilabel-sorted and matching on the input, or the
+// FST is olabel -orted and matching on the output).  In order to obtain the
+// most efficient behaviour, it is recommended to set call_label_type to
+// REPLACE_LABEL_INPUT or REPLACE_LABEL_BOTH and return_label_type to
+// REPLACE_LABEL_OUTPUT or REPLACE_LABEL_NEITHER. This means that the call arc
+// does not have epsilon on the input side and the return arc has epsilon on the
+// input side) and matching on the input side.
+//
+// This class attaches interface to implementation and handles reference
+// counting, delegating most methods to ImplToFst.
+template <class A, class T /* = DefaultReplaceStateTable<A> */,
+          class CacheStore /* = DefaultCacheStore<A> */>
+class ReplaceFst
+    : public ImplToFst<internal::ReplaceFstImpl<A, T, CacheStore>> {
+ public:
+  using Arc = A;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using StateTable = T;
+  using Store = CacheStore;
+  using State = typename CacheStore::State;
+  using Impl = internal::ReplaceFstImpl<Arc, StateTable, CacheStore>;
+  using CacheImpl = internal::CacheBaseImpl<State, CacheStore>;
+
+  using ImplToFst<Impl>::Properties;
+
+  friend class ArcIterator<ReplaceFst<Arc, StateTable, CacheStore>>;
+  friend class StateIterator<ReplaceFst<Arc, StateTable, CacheStore>>;
+  friend class ReplaceFstMatcher<Arc, StateTable, CacheStore>;
+
+  ReplaceFst(const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_array,
+             Label root)
+      : ImplToFst<Impl>(std::make_shared<Impl>(
+            fst_array, ReplaceFstOptions<Arc, StateTable, CacheStore>(root))) {}
+
+  ReplaceFst(const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_array,
+             const ReplaceFstOptions<Arc, StateTable, CacheStore> &opts)
+      : ImplToFst<Impl>(std::make_shared<Impl>(fst_array, opts)) {}
+
+  // See Fst<>::Copy() for doc.
+  ReplaceFst(const ReplaceFst<Arc, StateTable, CacheStore> &fst,
+             bool safe = false)
+      : ImplToFst<Impl>(fst, safe) {}
+
+  // Get a copy of this ReplaceFst. See Fst<>::Copy() for further doc.
+  ReplaceFst<Arc, StateTable, CacheStore> *Copy(
+      bool safe = false) const override {
+    return new ReplaceFst<Arc, StateTable, CacheStore>(*this, safe);
+  }
+
+  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
+    GetMutableImpl()->InitArcIterator(s, data);
+  }
+
+  MatcherBase<Arc> *InitMatcher(MatchType match_type) const override {
+    if ((GetImpl()->ArcIteratorFlags() & kArcNoCache) &&
+        ((match_type == MATCH_INPUT && Properties(kILabelSorted, false)) ||
+         (match_type == MATCH_OUTPUT && Properties(kOLabelSorted, false)))) {
+      return new ReplaceFstMatcher<Arc, StateTable, CacheStore>
+          (this, match_type);
+    } else {
+      VLOG(2) << "Not using replace matcher";
+      return nullptr;
+    }
+  }
+
+  bool CyclicDependencies() const { return GetImpl()->CyclicDependencies(); }
+
+  const StateTable &GetStateTable() const {
+    return *GetImpl()->GetStateTable();
+  }
+
+  const Fst<Arc> &GetFst(Label nonterminal) const {
+    return *GetImpl()->GetFst(GetImpl()->GetFstId(nonterminal));
+  }
+
+ private:
+  using ImplToFst<Impl>::GetImpl;
+  using ImplToFst<Impl>::GetMutableImpl;
+
+  ReplaceFst &operator=(const ReplaceFst &) = delete;
+};
+
+// Specialization for ReplaceFst.
+template <class Arc, class StateTable, class CacheStore>
+class StateIterator<ReplaceFst<Arc, StateTable, CacheStore>>
+    : public CacheStateIterator<ReplaceFst<Arc, StateTable, CacheStore>> {
+ public:
+  explicit StateIterator(const ReplaceFst<Arc, StateTable, CacheStore> &fst)
+      : CacheStateIterator<ReplaceFst<Arc, StateTable, CacheStore>>(
+            fst, fst.GetMutableImpl()) {}
+};
+
+// Specialization for ReplaceFst, implementing optional caching. It is be used
+// as follows:
+//
+//   ReplaceFst<A> replace;
+//   ArcIterator<ReplaceFst<A>> aiter(replace, s);
+//   // Note: ArcIterator< Fst<A>> is always a caching arc iterator.
+//   aiter.SetFlags(kArcNoCache, kArcNoCache);
+//   // Uses the arc iterator, no arc will be cached, no state will be expanded.
+//   // Arc flags can be used to decide which component of the arc need to be
+//   computed.
+//   aiter.SetFlags(kArcILabelValue, kArcValueFlags);
+//   // Wants the ilabel for this arc.
+//   aiter.Value();  // Does not compute the destination state.
+//   aiter.Next();
+//   aiter.SetFlags(kArcNextStateValue, kArcNextStateValue);
+//   // Wants the ilabel and next state for this arc.
+//   aiter.Value();  // Does compute the destination state and inserts it
+//                   // in the replace state table.
+//   // No additional arcs have been cached at this point.
+template <class Arc, class StateTable, class CacheStore>
+class ArcIterator<ReplaceFst<Arc, StateTable, CacheStore>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  using StateTuple = typename StateTable::StateTuple;
+
+  ArcIterator(const ReplaceFst<Arc, StateTable, CacheStore> &fst, StateId s)
+      : fst_(fst),
+        s_(s),
+        pos_(0),
+        offset_(0),
+        flags_(kArcValueFlags),
+        arcs_(nullptr),
+        data_flags_(0),
+        final_flags_(0) {
+    cache_data_.ref_count = nullptr;
+    local_data_.ref_count = nullptr;
+    // If FST does not support optional caching, forces caching.
+    if (!(fst_.GetImpl()->ArcIteratorFlags() & kArcNoCache) &&
+        !(fst_.GetImpl()->HasArcs(s_))) {
+      fst_.GetMutableImpl()->Expand(s_);
+    }
+    // If state is already cached, use cached arcs array.
+    if (fst_.GetImpl()->HasArcs(s_)) {
+      (fst_.GetImpl())
+          ->internal::template CacheBaseImpl<
+              typename CacheStore::State,
+              CacheStore>::InitArcIterator(s_, &cache_data_);
+      num_arcs_ = cache_data_.narcs;
+      arcs_ = cache_data_.arcs;      // arcs_ is a pointer to the cached arcs.
+      data_flags_ = kArcValueFlags;  // All the arc member values are valid.
+    } else {  // Otherwise delay decision until Value() is called.
+      tuple_ = fst_.GetImpl()->GetStateTable()->Tuple(s_);
+      if (tuple_.fst_state == kNoStateId) {
+        num_arcs_ = 0;
+      } else {
+        // The decision to cache or not to cache has been defered until Value()
+        // or
+        // SetFlags() is called. However, the arc iterator is set up now to be
+        // ready for non-caching in order to keep the Value() method simple and
+        // efficient.
+        const auto *rfst = fst_.GetImpl()->GetFst(tuple_.fst_id);
+        rfst->InitArcIterator(tuple_.fst_state, &local_data_);
+        // arcs_ is a pointer to the arcs in the underlying machine.
+        arcs_ = local_data_.arcs;
+        // Computes the final arc (but not its destination state) if a final arc
+        // is required.
+        bool has_final_arc = fst_.GetMutableImpl()->ComputeFinalArc(
+            tuple_, &final_arc_, kArcValueFlags & ~kArcNextStateValue);
+        // Sets the arc value flags that hold for final_arc_.
+        final_flags_ = kArcValueFlags & ~kArcNextStateValue;
+        // Computes the number of arcs.
+        num_arcs_ = local_data_.narcs;
+        if (has_final_arc) ++num_arcs_;
+        // Sets the offset between the underlying arc positions and the
+        // positions
+        // in the arc iterator.
+        offset_ = num_arcs_ - local_data_.narcs;
+        // Defers the decision to cache or not until Value() or SetFlags() is
+        // called.
+        data_flags_ = 0;
+      }
+    }
+  }
+
+  ~ArcIterator() {
+    if (cache_data_.ref_count) --(*cache_data_.ref_count);
+    if (local_data_.ref_count) --(*local_data_.ref_count);
+  }
+
+  void ExpandAndCache() const  {
+    // TODO(allauzen): revisit this.
+    // fst_.GetImpl()->Expand(s_, tuple_, local_data_);
+    // (fst_.GetImpl())->CacheImpl<A>*>::InitArcIterator(s_,
+    //                                               &cache_data_);
+    //
+    fst_.InitArcIterator(s_, &cache_data_);  // Expand and cache state.
+    arcs_ = cache_data_.arcs;      // arcs_ is a pointer to the cached arcs.
+    data_flags_ = kArcValueFlags;  // All the arc member values are valid.
+    offset_ = 0;                   // No offset.
+  }
+
+  void Init() {
+    if (flags_ & kArcNoCache) {  // If caching is disabled
+      // arcs_ is a pointer to the arcs in the underlying machine.
+      arcs_ = local_data_.arcs;
+      // Sets the arcs value flags that hold for arcs_.
+      data_flags_ = kArcWeightValue;
+      if (!fst_.GetMutableImpl()->EpsilonOnCallInput()) {
+        data_flags_ |= kArcILabelValue;
+      }
+      // Sets the offset between the underlying arc positions and the positions
+      // in the arc iterator.
+      offset_ = num_arcs_ - local_data_.narcs;
+    } else {
+      ExpandAndCache();
+    }
+  }
+
+  bool Done() const { return pos_ >= num_arcs_; }
+
+  const Arc &Value() const {
+    // If data_flags_ is 0, non-caching was not requested.
+    if (!data_flags_) {
+      // TODO(allauzen): Revisit this.
+      if (flags_ & kArcNoCache) {
+        // Should never happen.
+        FSTERROR() << "ReplaceFst: Inconsistent arc iterator flags";
+      }
+      ExpandAndCache();
+    }
+    if (pos_ - offset_ >= 0) {  // The requested arc is not the final arc.
+      const auto &arc = arcs_[pos_ - offset_];
+      if ((data_flags_ & flags_) == (flags_ & kArcValueFlags)) {
+        // If the value flags match the recquired value flags then returns the
+        // arc.
+        return arc;
+      } else {
+        // Otherwise, compute the corresponding arc on-the-fly.
+        fst_.GetMutableImpl()->ComputeArc(tuple_, arc, &arc_,
+                                          flags_ & kArcValueFlags);
+        return arc_;
+      }
+    } else {  // The requested arc is the final arc.
+      if ((final_flags_ & flags_) != (flags_ & kArcValueFlags)) {
+        // If the arc value flags that hold for the final arc do not match the
+        // requested value flags, then
+        // final_arc_ needs to be updated.
+        fst_.GetMutableImpl()->ComputeFinalArc(tuple_, &final_arc_,
+                                               flags_ & kArcValueFlags);
+        final_flags_ = flags_ & kArcValueFlags;
+      }
+      return final_arc_;
+    }
+  }
+
+  void Next() { ++pos_; }
+
+  size_t Position() const { return pos_; }
+
+  void Reset() { pos_ = 0; }
+
+  void Seek(size_t pos) { pos_ = pos; }
+
+  uint32 Flags() const { return flags_; }
+
+  void SetFlags(uint32 flags, uint32 mask) {
+    // Updates the flags taking into account what flags are supported
+    // by the FST.
+    flags_ &= ~mask;
+    flags_ |= (flags & fst_.GetImpl()->ArcIteratorFlags());
+    // If non-caching is not requested (and caching has not already been
+    // performed), then flush data_flags_ to request caching during the next
+    // call to Value().
+    if (!(flags_ & kArcNoCache) && data_flags_ != kArcValueFlags) {
+      if (!fst_.GetImpl()->HasArcs(s_)) data_flags_ = 0;
+    }
+    // If data_flags_ has been flushed but non-caching is requested before
+    // calling Value(), then set up the iterator for non-caching.
+    if ((flags & kArcNoCache) && (!data_flags_)) Init();
+  }
+
+ private:
+  const ReplaceFst<Arc, StateTable, CacheStore> &fst_;  // Reference to the FST.
+  StateId s_;                                           // State in the FST.
+  mutable StateTuple tuple_;  // Tuple corresponding to state_.
+
+  ssize_t pos_;             // Current position.
+  mutable ssize_t offset_;  // Offset between position in iterator and in arcs_.
+  ssize_t num_arcs_;        // Number of arcs at state_.
+  uint32 flags_;            // Behavorial flags for the arc iterator
+  mutable Arc arc_;         // Memory to temporarily store computed arcs.
+
+  mutable ArcIteratorData<Arc> cache_data_;  // Arc iterator data in cache.
+  mutable ArcIteratorData<Arc> local_data_;  // Arc iterator data in local FST.
+
+  mutable const Arc *arcs_;     // Array of arcs.
+  mutable uint32 data_flags_;   // Arc value flags valid for data in arcs_.
+  mutable Arc final_arc_;       // Final arc (when required).
+  mutable uint32 final_flags_;  // Arc value flags valid for final_arc_.
+
+  ArcIterator(const ArcIterator &) = delete;
+  ArcIterator &operator=(const ArcIterator &) = delete;
+};
+
+template <class Arc, class StateTable, class CacheStore>
+class ReplaceFstMatcher : public MatcherBase<Arc> {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FST = ReplaceFst<Arc, StateTable, CacheStore>;
+  using LocalMatcher = MultiEpsMatcher<Matcher<Fst<Arc>>>;
+
+  using StateTuple = typename StateTable::StateTuple;
+
+  // This makes a copy of the FST.
+  ReplaceFstMatcher(const ReplaceFst<Arc, StateTable, CacheStore> &fst,
+                    MatchType match_type)
+      : owned_fst_(fst.Copy()),
+        fst_(*owned_fst_),
+        impl_(fst_.GetMutableImpl()),
+        s_(fst::kNoStateId),
+        match_type_(match_type),
+        current_loop_(false),
+        final_arc_(false),
+        loop_(kNoLabel, 0, Weight::One(), kNoStateId) {
+    if (match_type_ == fst::MATCH_OUTPUT) {
+      std::swap(loop_.ilabel, loop_.olabel);
+    }
+    InitMatchers();
+  }
+
+  // This doesn't copy the FST.
+  ReplaceFstMatcher(const ReplaceFst<Arc, StateTable, CacheStore> *fst,
+                    MatchType match_type)
+      : fst_(*fst),
+        impl_(fst_.GetMutableImpl()),
+        s_(fst::kNoStateId),
+        match_type_(match_type),
+        current_loop_(false),
+        final_arc_(false),
+        loop_(kNoLabel, 0, Weight::One(), kNoStateId) {
+    if (match_type_ == fst::MATCH_OUTPUT) {
+      std::swap(loop_.ilabel, loop_.olabel);
+    }
+    InitMatchers();
+  }
+
+  // This makes a copy of the FST.
+  ReplaceFstMatcher(
+      const ReplaceFstMatcher<Arc, StateTable, CacheStore> &matcher,
+      bool safe = false)
+      : owned_fst_(matcher.fst_.Copy(safe)),
+        fst_(*owned_fst_),
+        impl_(fst_.GetMutableImpl()),
+        s_(fst::kNoStateId),
+        match_type_(matcher.match_type_),
+        current_loop_(false),
+        final_arc_(false),
+        loop_(fst::kNoLabel, 0, Weight::One(), fst::kNoStateId) {
+    if (match_type_ == fst::MATCH_OUTPUT) {
+      std::swap(loop_.ilabel, loop_.olabel);
+    }
+    InitMatchers();
+  }
+
+  // Creates a local matcher for each component FST in the RTN. LocalMatcher is
+  // a multi-epsilon wrapper matcher. MultiEpsilonMatcher is used to match each
+  // non-terminal arc, since these non-terminal
+  // turn into epsilons on recursion.
+  void InitMatchers() {
+    const auto &fst_array = impl_->fst_array_;
+    matcher_.resize(fst_array.size());
+    for (Label i = 0; i < fst_array.size(); ++i) {
+      if (fst_array[i]) {
+        matcher_[i].reset(
+            new LocalMatcher(*fst_array[i], match_type_, kMultiEpsList));
+        auto it = impl_->nonterminal_set_.begin();
+        for (; it != impl_->nonterminal_set_.end(); ++it) {
+          matcher_[i]->AddMultiEpsLabel(*it);
+        }
+      }
+    }
+  }
+
+  ReplaceFstMatcher<Arc, StateTable, CacheStore> *Copy(
+      bool safe = false) const override {
+    return new ReplaceFstMatcher<Arc, StateTable, CacheStore>(*this, safe);
+  }
+
+  MatchType Type(bool test) const override {
+    if (match_type_ == MATCH_NONE) return match_type_;
+    const auto true_prop =
+        match_type_ == MATCH_INPUT ? kILabelSorted : kOLabelSorted;
+    const auto false_prop =
+        match_type_ == MATCH_INPUT ? kNotILabelSorted : kNotOLabelSorted;
+    const auto props = fst_.Properties(true_prop | false_prop, test);
+    if (props & true_prop) {
+      return match_type_;
+    } else if (props & false_prop) {
+      return MATCH_NONE;
+    } else {
+      return MATCH_UNKNOWN;
+    }
+  }
+
+  const Fst<Arc> &GetFst() const override { return fst_; }
+
+  uint64 Properties(uint64 props) const override { return props; }
+
+  // Sets the state from which our matching happens.
+  void SetState(StateId s) final {
+    if (s_ == s) return;
+    s_ = s;
+    tuple_ = impl_->GetStateTable()->Tuple(s_);
+    if (tuple_.fst_state == kNoStateId) {
+      done_ = true;
+      return;
+    }
+    // Gets current matcher, used for non-epsilon matching.
+    current_matcher_ = matcher_[tuple_.fst_id].get();
+    current_matcher_->SetState(tuple_.fst_state);
+    loop_.nextstate = s_;
+    final_arc_ = false;
+  }
+
+  // Searches for label from previous set state. If label == 0, first
+  // hallucinate an epsilon loop; otherwise use the underlying matcher to
+  // search for the label or epsilons. Note since the ReplaceFst recursion
+  // on non-terminal arcs causes epsilon transitions to be created we use
+  // MultiEpsilonMatcher to search for possible matches of non-terminals. If the
+  // component FST
+  // reaches a final state we also need to add the exiting final arc.
+  bool Find(Label label) final {
+    bool found = false;
+    label_ = label;
+    if (label_ == 0 || label_ == kNoLabel) {
+      // Computes loop directly, avoiding Replace::ComputeArc.
+      if (label_ == 0) {
+        current_loop_ = true;
+        found = true;
+      }
+      // Searches for matching multi-epsilons.
+      final_arc_ = impl_->ComputeFinalArc(tuple_, nullptr);
+      found = current_matcher_->Find(kNoLabel) || final_arc_ || found;
+    } else {
+      // Searches on a sub machine directly using sub machine matcher.
+      found = current_matcher_->Find(label_);
+    }
+    return found;
+  }
+
+  bool Done() const final {
+    return !current_loop_ && !final_arc_ && current_matcher_->Done();
+  }
+
+  const Arc &Value() const final {
+    if (current_loop_) return loop_;
+    if (final_arc_) {
+      impl_->ComputeFinalArc(tuple_, &arc_);
+      return arc_;
+    }
+    const auto &component_arc = current_matcher_->Value();
+    impl_->ComputeArc(tuple_, component_arc, &arc_);
+    return arc_;
+  }
+
+  void Next() final {
+    if (current_loop_) {
+      current_loop_ = false;
+      return;
+    }
+    if (final_arc_) {
+      final_arc_ = false;
+      return;
+    }
+    current_matcher_->Next();
+  }
+
+  ssize_t Priority(StateId s) final { return fst_.NumArcs(s); }
+
+ private:
+  std::unique_ptr<const ReplaceFst<Arc, StateTable, CacheStore>> owned_fst_;
+  const ReplaceFst<Arc, StateTable, CacheStore> &fst_;
+  internal::ReplaceFstImpl<Arc, StateTable, CacheStore> *impl_;
+  LocalMatcher *current_matcher_;
+  std::vector<std::unique_ptr<LocalMatcher>> matcher_;
+  StateId s_;             // Current state.
+  Label label_;           // Current label.
+  MatchType match_type_;  // Supplied by caller.
+  mutable bool done_;
+  mutable bool current_loop_;  // Current arc is the implicit loop.
+  mutable bool final_arc_;     // Current arc for exiting recursion.
+  mutable StateTuple tuple_;   // Tuple corresponding to state_.
+  mutable Arc arc_;
+  Arc loop_;
+
+  ReplaceFstMatcher &operator=(const ReplaceFstMatcher &) = delete;
+};
+
+template <class Arc, class StateTable, class CacheStore>
+inline void ReplaceFst<Arc, StateTable, CacheStore>::InitStateIterator(
+    StateIteratorData<Arc> *data) const {
+  data->base =
+      new StateIterator<ReplaceFst<Arc, StateTable, CacheStore>>(*this);
+}
+
+using StdReplaceFst = ReplaceFst<StdArc>;
+
+// Recursively replaces arcs in the root FSTs with other FSTs.
+// This version writes the result of replacement to an output MutableFst.
+//
+// Replace supports replacement of arcs in one Fst with another FST. This
+// replacement is recursive. Replace takes an array of FST(s). One FST
+// represents the root (or topology) machine. The root FST refers to other FSTs
+// by recursively replacing arcs labeled as non-terminals with the matching
+// non-terminal FST. Currently Replace uses the output symbols of the arcs to
+// determine whether the arc is a non-terminal arc or not. A non-terminal can be
+// any label that is not a non-zero terminal label in the output alphabet.
+//
+// Note that input argument is a vector of pairs. These correspond to the tuple
+// of non-terminal Label and corresponding FST.
+template <class Arc>
+void Replace(const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
+                 &ifst_array,
+             MutableFst<Arc> *ofst,
+             ReplaceFstOptions<Arc> opts = ReplaceFstOptions<Arc>()) {
+  opts.gc = true;
+  opts.gc_limit = 0;  // Caches only the last state for fastest copy.
+  *ofst = ReplaceFst<Arc>(ifst_array, opts);
+}
+
+template <class Arc>
+void Replace(const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
+                 &ifst_array,
+             MutableFst<Arc> *ofst, const ReplaceUtilOptions &opts) {
+  Replace(ifst_array, ofst, ReplaceFstOptions<Arc>(opts));
+}
+
+// For backwards compatibility.
+template <class Arc>
+void Replace(const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
+                 &ifst_array,
+             MutableFst<Arc> *ofst, typename Arc::Label root,
+             bool epsilon_on_replace) {
+  Replace(ifst_array, ofst, ReplaceFstOptions<Arc>(root, epsilon_on_replace));
+}
+
+template <class Arc>
+void Replace(const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
+                 &ifst_array,
+             MutableFst<Arc> *ofst, typename Arc::Label root) {
+  Replace(ifst_array, ofst, ReplaceFstOptions<Arc>(root));
+}
+
+}  // namespace fst
+
+#endif  // FST_REPLACE_H_
diff --git a/projects/llm_framework/include/fst/reverse.h b/projects/llm_framework/include/fst/reverse.h
new file mode 100644
index 00000000..7c7c89db
--- /dev/null
+++ b/projects/llm_framework/include/fst/reverse.h
@@ -0,0 +1,116 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes to sort arcs in an FST.
+
+#ifndef FST_REVERSE_H_
+#define FST_REVERSE_H_
+
+#include <algorithm>
+#include <vector>
+
+#include <fst/cache.h>
+
+
+namespace fst {
+
+// Reverses an FST. The reversed result is written to an output mutable FST.
+// If A transduces string x to y with weight a, then the reverse of A
+// transduces the reverse of x to the reverse of y with weight a.Reverse().
+//
+// Typically, a = a.Reverse() and an arc is its own reverse (e.g., for
+// TropicalWeight or LogWeight). In general, e.g., when the weights only form a
+// left or right semiring, the output arc type must match the input arc type
+// except having the reversed Weight type.
+//
+// When require_superinitial is false, a superinitial state is not created in
+// the reversed FST iff the input FST has exactly one final state (which becomes
+// the initial state of the reversed FST) with a final weight of semiring One,
+// or if it does not belong to any cycle. When require_superinitial is true, a
+// superinitial state is always created.
+template <class FromArc, class ToArc>
+void Reverse(const Fst<FromArc> &ifst, MutableFst<ToArc> *ofst,
+             bool require_superinitial = true) {
+  using StateId = typename FromArc::StateId;
+  using FromWeight = typename FromArc::Weight;
+  using ToWeight = typename ToArc::Weight;
+  ofst->DeleteStates();
+  ofst->SetInputSymbols(ifst.InputSymbols());
+  ofst->SetOutputSymbols(ifst.OutputSymbols());
+  if (ifst.Properties(kExpanded, false)) {
+    ofst->ReserveStates(CountStates(ifst) + 1);
+  }
+  StateId istart = ifst.Start();
+  StateId ostart = kNoStateId;
+  StateId offset = 0;
+  uint64 dfs_iprops = 0;
+  uint64 dfs_oprops = 0;
+  if (!require_superinitial) {
+    for (StateIterator<Fst<FromArc>> siter(ifst); !siter.Done(); siter.Next()) {
+      const auto s = siter.Value();
+      if (ifst.Final(s) == FromWeight::Zero()) continue;
+      if (ostart != kNoStateId) {
+        ostart = kNoStateId;
+        break;
+      } else {
+        ostart = s;
+      }
+    }
+    if (ostart != kNoStateId && ifst.Final(ostart) != FromWeight::One()) {
+      std::vector<StateId> scc;
+      SccVisitor<FromArc> scc_visitor(&scc, nullptr, nullptr, &dfs_iprops);
+      DfsVisit(ifst, &scc_visitor);
+      if (count(scc.begin(), scc.end(), scc[ostart]) > 1) {
+        ostart = kNoStateId;
+      } else {
+        for (ArcIterator<Fst<FromArc>> aiter(ifst, ostart); !aiter.Done();
+             aiter.Next()) {
+          if (aiter.Value().nextstate == ostart) {
+            ostart = kNoStateId;
+            break;
+          }
+        }
+      }
+      if (ostart != kNoStateId) dfs_oprops = kInitialAcyclic;
+    }
+  }
+  if (ostart == kNoStateId) {  // Super-initial requested or needed.
+    ostart = ofst->AddState();
+    offset = 1;
+  }
+  for (StateIterator<Fst<FromArc>> siter(ifst); !siter.Done(); siter.Next()) {
+    const auto is = siter.Value();
+    const auto os = is + offset;
+    while (ofst->NumStates() <= os) ofst->AddState();
+    if (is == istart) ofst->SetFinal(os, ToWeight::One());
+    const auto weight = ifst.Final(is);
+    if ((weight != FromWeight::Zero()) && (offset == 1)) {
+      const ToArc oarc(0, 0, weight.Reverse(), os);
+      ofst->AddArc(0, oarc);
+    }
+    for (ArcIterator<Fst<FromArc>> aiter(ifst, is); !aiter.Done();
+         aiter.Next()) {
+      const auto &iarc = aiter.Value();
+      const auto nos = iarc.nextstate + offset;
+      auto weight = iarc.weight.Reverse();
+      if (!offset && (nos == ostart)) {
+        weight = Times(ifst.Final(ostart).Reverse(), weight);
+      }
+      const ToArc oarc(iarc.ilabel, iarc.olabel, weight, os);
+      while (ofst->NumStates() <= nos) ofst->AddState();
+      ofst->AddArc(nos, oarc);
+    }
+  }
+  ofst->SetStart(ostart);
+  if (offset == 0 && ostart == istart) {
+    ofst->SetFinal(ostart, ifst.Final(ostart).Reverse());
+  }
+  const auto iprops = ifst.Properties(kCopyProperties, false) | dfs_iprops;
+  const auto oprops = ofst->Properties(kFstProperties, false) | dfs_oprops;
+  ofst->SetProperties(ReverseProperties(iprops, offset == 1) | oprops,
+                      kFstProperties);
+}
+
+}  // namespace fst
+
+#endif  // FST_REVERSE_H_
diff --git a/projects/llm_framework/include/fst/reweight.h b/projects/llm_framework/include/fst/reweight.h
new file mode 100644
index 00000000..64e68cb7
--- /dev/null
+++ b/projects/llm_framework/include/fst/reweight.h
@@ -0,0 +1,127 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Function to reweight an FST.
+
+#ifndef FST_REWEIGHT_H_
+#define FST_REWEIGHT_H_
+
+#include <vector>
+#include <fst/log.h>
+
+#include <fst/mutable-fst.h>
+
+
+namespace fst {
+
+enum ReweightType { REWEIGHT_TO_INITIAL, REWEIGHT_TO_FINAL };
+
+// Reweights an FST according to a vector of potentials in a given direction.
+// The weight must be left distributive when reweighting towards the initial
+// state and right distributive when reweighting towards the final states.
+//
+// An arc of weight w, with an origin state of potential p and destination state
+// of potential q, is reweighted by p^-1 \otimes (w \otimes q) when reweighting
+// torwards the initial state, and by (p \otimes w) \otimes q^-1 when
+// reweighting towards the final states.
+template <class Arc>
+void Reweight(MutableFst<Arc> *fst,
+              const std::vector<typename Arc::Weight> &potential,
+              ReweightType type) {
+  using Weight = typename Arc::Weight;
+  if (fst->NumStates() == 0) return;
+  // TODO(kbg): Make this a compile-time static_assert once we have a pleasant
+  // way to "deregister" this operation for non-distributive semirings so an
+  // informative error message is produced.
+  if (type == REWEIGHT_TO_FINAL && !(Weight::Properties() & kRightSemiring)) {
+    FSTERROR() << "Reweight: Reweighting to the final states requires "
+               << "Weight to be right distributive: " << Weight::Type();
+    fst->SetProperties(kError, kError);
+    return;
+  }
+  // TODO(kbg): Make this a compile-time static_assert once we have a pleasant
+  // way to "deregister" this operation for non-distributive semirings so an
+  // informative error message is produced.
+  if (type == REWEIGHT_TO_INITIAL && !(Weight::Properties() & kLeftSemiring)) {
+    FSTERROR() << "Reweight: Reweighting to the initial state requires "
+               << "Weight to be left distributive: " << Weight::Type();
+    fst->SetProperties(kError, kError);
+    return;
+  }
+  StateIterator<MutableFst<Arc>> siter(*fst);
+  for (; !siter.Done(); siter.Next()) {
+    const auto s = siter.Value();
+    if (s == potential.size()) break;
+    const auto &weight = potential[s];
+    if (weight != Weight::Zero()) {
+      for (MutableArcIterator<MutableFst<Arc>> aiter(fst, s); !aiter.Done();
+           aiter.Next()) {
+        auto arc = aiter.Value();
+        if (arc.nextstate >= potential.size()) continue;
+        const auto &nextweight = potential[arc.nextstate];
+        if (nextweight == Weight::Zero()) continue;
+        if (type == REWEIGHT_TO_INITIAL) {
+          arc.weight =
+              Divide(Times(arc.weight, nextweight), weight, DIVIDE_LEFT);
+        }
+        if (type == REWEIGHT_TO_FINAL) {
+          arc.weight =
+              Divide(Times(weight, arc.weight), nextweight, DIVIDE_RIGHT);
+        }
+        aiter.SetValue(arc);
+      }
+      if (type == REWEIGHT_TO_INITIAL) {
+        fst->SetFinal(s, Divide(fst->Final(s), weight, DIVIDE_LEFT));
+      }
+    }
+    if (type == REWEIGHT_TO_FINAL) {
+      fst->SetFinal(s, Times(weight, fst->Final(s)));
+    }
+  }
+  // This handles elements past the end of the potentials array.
+  for (; !siter.Done(); siter.Next()) {
+    const auto s = siter.Value();
+    if (type == REWEIGHT_TO_FINAL) {
+      fst->SetFinal(s, Times(Weight::Zero(), fst->Final(s)));
+    }
+  }
+  const auto startweight = fst->Start() < potential.size()
+                               ? potential[fst->Start()]
+                               : Weight::Zero();
+  if ((startweight != Weight::One()) && (startweight != Weight::Zero())) {
+    if (fst->Properties(kInitialAcyclic, true) & kInitialAcyclic) {
+      const auto s = fst->Start();
+      for (MutableArcIterator<MutableFst<Arc>> aiter(fst, s); !aiter.Done();
+           aiter.Next()) {
+        auto arc = aiter.Value();
+        if (type == REWEIGHT_TO_INITIAL) {
+          arc.weight = Times(startweight, arc.weight);
+        } else {
+          arc.weight = Times(Divide(Weight::One(), startweight, DIVIDE_RIGHT),
+                             arc.weight);
+        }
+        aiter.SetValue(arc);
+      }
+      if (type == REWEIGHT_TO_INITIAL) {
+        fst->SetFinal(s, Times(startweight, fst->Final(s)));
+      } else {
+        fst->SetFinal(s, Times(Divide(Weight::One(), startweight, DIVIDE_RIGHT),
+                               fst->Final(s)));
+      }
+    } else {
+      const auto s = fst->AddState();
+      const auto weight =
+          (type == REWEIGHT_TO_INITIAL)
+              ? startweight
+              : Divide(Weight::One(), startweight, DIVIDE_RIGHT);
+      fst->AddArc(s, Arc(0, 0, weight, fst->Start()));
+      fst->SetStart(s);
+    }
+  }
+  fst->SetProperties(ReweightProperties(fst->Properties(kFstProperties, false)),
+                     kFstProperties);
+}
+
+}  // namespace fst
+
+#endif  // FST_REWEIGHT_H_
diff --git a/projects/llm_framework/include/fst/rmepsilon.h b/projects/llm_framework/include/fst/rmepsilon.h
new file mode 100644
index 00000000..5135bf2d
--- /dev/null
+++ b/projects/llm_framework/include/fst/rmepsilon.h
@@ -0,0 +1,548 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes that implemement epsilon-removal.
+
+#ifndef FST_RMEPSILON_H_
+#define FST_RMEPSILON_H_
+
+#include <forward_list>
+#include <stack>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/arcfilter.h>
+#include <fst/cache.h>
+#include <fst/connect.h>
+#include <fst/factor-weight.h>
+#include <fst/invert.h>
+#include <fst/prune.h>
+#include <fst/queue.h>
+#include <fst/shortest-distance.h>
+#include <fst/topsort.h>
+
+
+namespace fst {
+
+template <class Arc, class Queue>
+class RmEpsilonOptions
+    : public ShortestDistanceOptions<Arc, Queue, EpsilonArcFilter<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  bool connect;             // Connect output
+  Weight weight_threshold;  // Pruning weight threshold.
+  StateId state_threshold;  // Pruning state threshold.
+
+  explicit RmEpsilonOptions(Queue *queue, float delta = kShortestDelta,
+                            bool connect = true,
+                            Weight weight_threshold = Weight::Zero(),
+                            StateId state_threshold = kNoStateId)
+      : ShortestDistanceOptions<Arc, Queue, EpsilonArcFilter<Arc>>(
+            queue, EpsilonArcFilter<Arc>(), kNoStateId, delta),
+        connect(connect),
+        weight_threshold(std::move(weight_threshold)),
+        state_threshold(state_threshold) {}
+};
+
+namespace internal {
+
+// Computation state of the epsilon-removal algorithm.
+template <class Arc, class Queue>
+class RmEpsilonState {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  RmEpsilonState(const Fst<Arc> &fst, std::vector<Weight> *distance,
+                 const RmEpsilonOptions<Arc, Queue> &opts)
+      : fst_(fst),
+        distance_(distance),
+        sd_state_(fst_, distance, opts, true),
+        expand_id_(0) {}
+
+  void Expand(StateId s);
+
+  std::vector<Arc> &Arcs() { return arcs_; }
+
+  const Weight &Final() const { return final_; }
+
+  bool Error() const { return sd_state_.Error(); }
+
+ private:
+  struct Element {
+    Label ilabel;
+    Label olabel;
+    StateId nextstate;
+
+    Element() {}
+
+    Element(Label ilabel, Label olabel, StateId nexstate)
+        : ilabel(ilabel), olabel(olabel), nextstate(nexstate) {}
+  };
+
+  struct ElementHash {
+   public:
+    size_t operator()(const Element &element) const {
+      static constexpr size_t prime0 = 7853;
+      static constexpr size_t prime1 = 7867;
+      return static_cast<size_t>(element.nextstate) +
+             static_cast<size_t>(element.ilabel) * prime0 +
+             static_cast<size_t>(element.olabel) * prime1;
+    }
+  };
+
+  class ElementEqual {
+   public:
+    bool operator()(const Element &e1, const Element &e2) const {
+      return (e1.ilabel == e2.ilabel) && (e1.olabel == e2.olabel) &&
+             (e1.nextstate == e2.nextstate);
+    }
+  };
+
+  using ElementMap = std::unordered_map<Element, std::pair<StateId, size_t>,
+                                        ElementHash, ElementEqual>;
+
+  const Fst<Arc> &fst_;
+  // Distance from state being expanded in epsilon-closure.
+  std::vector<Weight> *distance_;
+  // Shortest distance algorithm computation state.
+  internal::ShortestDistanceState<Arc, Queue, EpsilonArcFilter<Arc>> sd_state_;
+  // Maps an element to a pair corresponding to a position in the arcs vector
+  // of the state being expanded. The element corresopnds to the position in
+  // the arcs_ vector if p.first is equal to the state being expanded.
+  ElementMap element_map_;
+  EpsilonArcFilter<Arc> eps_filter_;
+  std::stack<StateId> eps_queue_;  // Queue used to visit the epsilon-closure.
+  std::vector<bool> visited_;      // True if the state has been visited.
+  std::forward_list<StateId> visited_states_;  // List of visited states.
+  std::vector<Arc> arcs_;                      // Arcs of state being expanded.
+  Weight final_;       // Final weight of state being expanded.
+  StateId expand_id_;  // Unique ID for each call to Expand
+
+  RmEpsilonState(const RmEpsilonState &) = delete;
+  RmEpsilonState &operator=(const RmEpsilonState &) = delete;
+};
+
+template <class Arc, class Queue>
+void RmEpsilonState<Arc, Queue>::Expand(typename Arc::StateId source) {
+  final_ = Weight::Zero();
+  arcs_.clear();
+  sd_state_.ShortestDistance(source);
+  if (sd_state_.Error()) return;
+  eps_queue_.push(source);
+  while (!eps_queue_.empty()) {
+    const auto state = eps_queue_.top();
+    eps_queue_.pop();
+    while (visited_.size() <= state) visited_.push_back(false);
+    if (visited_[state]) continue;
+    visited_[state] = true;
+    visited_states_.push_front(state);
+    for (ArcIterator<Fst<Arc>> aiter(fst_, state); !aiter.Done();
+         aiter.Next()) {
+      auto arc = aiter.Value();
+      arc.weight = Times((*distance_)[state], arc.weight);
+      if (eps_filter_(arc)) {
+        while (visited_.size() <= arc.nextstate) visited_.push_back(false);
+        if (!visited_[arc.nextstate]) eps_queue_.push(arc.nextstate);
+      } else {
+        const Element element(arc.ilabel, arc.olabel, arc.nextstate);
+        auto insert_result = element_map_.insert(
+            std::make_pair(element, std::make_pair(expand_id_, arcs_.size())));
+        if (insert_result.second) {
+          arcs_.push_back(std::move(arc));
+        } else {
+          if (insert_result.first->second.first == expand_id_) {
+            auto &weight = arcs_[insert_result.first->second.second].weight;
+            weight = Plus(weight, arc.weight);
+          } else {
+            insert_result.first->second.first = expand_id_;
+            insert_result.first->second.second = arcs_.size();
+            arcs_.push_back(std::move(arc));
+          }
+        }
+      }
+    }
+    final_ = Plus(final_, Times((*distance_)[state], fst_.Final(state)));
+  }
+  while (!visited_states_.empty()) {
+    visited_[visited_states_.front()] = false;
+    visited_states_.pop_front();
+  }
+  ++expand_id_;
+}
+
+}  // namespace internal
+
+// Removes epsilon-transitions (when both the input and output label are an
+// epsilon) from a transducer. The result will be an equivalent FST that has no
+// such epsilon transitions. This version modifies its input. It allows fine
+// control via the options argument; see below for a simpler interface.
+//
+// The distance vector will be used to hold the shortest distances during the
+// epsilon-closure computation. The state queue discipline and convergence delta
+// are taken in the options argument.
+template <class Arc, class Queue>
+void RmEpsilon(MutableFst<Arc> *fst,
+               std::vector<typename Arc::Weight> *distance,
+               const RmEpsilonOptions<Arc, Queue> &opts) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  if (fst->Start() == kNoStateId) return;
+  // noneps_in[s] will be set to true iff s admits a non-epsilon incoming
+  // transition or is the start state.
+  std::vector<bool> noneps_in(fst->NumStates(), false);
+  noneps_in[fst->Start()] = true;
+  for (size_t i = 0; i < fst->NumStates(); ++i) {
+    for (ArcIterator<Fst<Arc>> aiter(*fst, i); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      if (arc.ilabel != 0 || arc.olabel != 0) {
+        noneps_in[arc.nextstate] = true;
+      }
+    }
+  }
+  // States sorted in topological order when (acyclic) or generic topological
+  // order (cyclic).
+  std::vector<StateId> states;
+  states.reserve(fst->NumStates());
+  if (fst->Properties(kTopSorted, false) & kTopSorted) {
+    for (size_t i = 0; i < fst->NumStates(); i++) states.push_back(i);
+  } else if (fst->Properties(kAcyclic, false) & kAcyclic) {
+    std::vector<StateId> order;
+    bool acyclic;
+    TopOrderVisitor<Arc> top_order_visitor(&order, &acyclic);
+    DfsVisit(*fst, &top_order_visitor, EpsilonArcFilter<Arc>());
+    // Sanity check: should be acyclic if property bit is set.
+    if (!acyclic) {
+      FSTERROR() << "RmEpsilon: Inconsistent acyclic property bit";
+      fst->SetProperties(kError, kError);
+      return;
+    }
+    states.resize(order.size());
+    for (StateId i = 0; i < order.size(); i++) states[order[i]] = i;
+  } else {
+    uint64 props;
+    std::vector<StateId> scc;
+    SccVisitor<Arc> scc_visitor(&scc, nullptr, nullptr, &props);
+    DfsVisit(*fst, &scc_visitor, EpsilonArcFilter<Arc>());
+    std::vector<StateId> first(scc.size(), kNoStateId);
+    std::vector<StateId> next(scc.size(), kNoStateId);
+    for (StateId i = 0; i < scc.size(); i++) {
+      if (first[scc[i]] != kNoStateId) next[i] = first[scc[i]];
+      first[scc[i]] = i;
+    }
+    for (StateId i = 0; i < first.size(); i++) {
+      for (auto j = first[i]; j != kNoStateId; j = next[j]) {
+        states.push_back(j);
+      }
+    }
+  }
+  internal::RmEpsilonState<Arc, Queue> rmeps_state(*fst, distance, opts);
+  while (!states.empty()) {
+    const auto state = states.back();
+    states.pop_back();
+    if (!noneps_in[state] &&
+        (opts.connect || opts.weight_threshold != Weight::Zero() ||
+         opts.state_threshold != kNoStateId)) {
+      continue;
+    }
+    rmeps_state.Expand(state);
+    fst->SetFinal(state, rmeps_state.Final());
+    fst->DeleteArcs(state);
+    auto &arcs = rmeps_state.Arcs();
+    fst->ReserveArcs(state, arcs.size());
+    while (!arcs.empty()) {
+      fst->AddArc(state, arcs.back());
+      arcs.pop_back();
+    }
+  }
+  if (opts.connect || opts.weight_threshold != Weight::Zero() ||
+      opts.state_threshold != kNoStateId) {
+    for (size_t s = 0; s < fst->NumStates(); ++s) {
+      if (!noneps_in[s]) fst->DeleteArcs(s);
+    }
+  }
+  if (rmeps_state.Error()) fst->SetProperties(kError, kError);
+  fst->SetProperties(
+      RmEpsilonProperties(fst->Properties(kFstProperties, false)),
+      kFstProperties);
+  if (opts.weight_threshold != Weight::Zero() ||
+      opts.state_threshold != kNoStateId) {
+    Prune(fst, opts.weight_threshold, opts.state_threshold);
+  }
+  if (opts.connect && opts.weight_threshold == Weight::Zero() &&
+      opts.state_threshold == kNoStateId) {
+    Connect(fst);
+  }
+}
+
+// Removes epsilon-transitions (when both the input and output label
+// are an epsilon) from a transducer. The result will be an equivalent
+// FST that has no such epsilon transitions. This version modifies its
+// input. It has a simplified interface; see above for a version that
+// allows finer control.
+//
+// Complexity:
+//
+// - Time:
+//
+//   Unweighted: O(v^2 + ve).
+//   Acyclic: O(v^2 + V e).
+//   Tropical semiring: O(v^2 log V + ve).
+//   General: exponential.
+//
+// - Space: O(vE)
+//
+// where v is the number of states visited and e is the number of arcs visited.
+//
+// For more information, see:
+//
+// Mohri, M. 2002. Generic epsilon-removal and input epsilon-normalization
+// algorithms for weighted transducers. International Journal of Computer
+// Science 13(1): 129-143.
+template <class Arc>
+void RmEpsilon(MutableFst<Arc> *fst, bool connect = true,
+               typename Arc::Weight weight_threshold = Arc::Weight::Zero(),
+               typename Arc::StateId state_threshold = kNoStateId,
+               float delta = kShortestDelta) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  std::vector<Weight> distance;
+  AutoQueue<StateId> state_queue(*fst, &distance, EpsilonArcFilter<Arc>());
+  RmEpsilonOptions<Arc, AutoQueue<StateId>> opts(
+      &state_queue, delta, connect, weight_threshold, state_threshold);
+  RmEpsilon(fst, &distance, opts);
+}
+
+struct RmEpsilonFstOptions : CacheOptions {
+  float delta;
+
+  explicit RmEpsilonFstOptions(const CacheOptions &opts,
+                               float delta = kShortestDelta)
+      : CacheOptions(opts), delta(delta) {}
+
+  explicit RmEpsilonFstOptions(float delta = kShortestDelta) : delta(delta) {}
+};
+
+namespace internal {
+
+// Implementation of delayed RmEpsilonFst.
+template <class Arc>
+class RmEpsilonFstImpl : public CacheImpl<Arc> {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Store = DefaultCacheStore<Arc>;
+  using State = typename Store::State;
+
+  using FstImpl<Arc>::Properties;
+  using FstImpl<Arc>::SetType;
+  using FstImpl<Arc>::SetProperties;
+  using FstImpl<Arc>::SetInputSymbols;
+  using FstImpl<Arc>::SetOutputSymbols;
+
+  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
+  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
+  using CacheBaseImpl<CacheState<Arc>>::HasStart;
+  using CacheBaseImpl<CacheState<Arc>>::PushArc;
+  using CacheBaseImpl<CacheState<Arc>>::SetArcs;
+  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
+  using CacheBaseImpl<CacheState<Arc>>::SetStart;
+
+  RmEpsilonFstImpl(const Fst<Arc> &fst, const RmEpsilonFstOptions &opts)
+      : CacheImpl<Arc>(opts),
+        fst_(fst.Copy()),
+        delta_(opts.delta),
+        rmeps_state_(
+            *fst_, &distance_,
+            RmEpsilonOptions<Arc, FifoQueue<StateId>>(&queue_, delta_, false)) {
+    SetType("rmepsilon");
+    SetProperties(
+        RmEpsilonProperties(fst.Properties(kFstProperties, false), true),
+        kCopyProperties);
+    SetInputSymbols(fst.InputSymbols());
+    SetOutputSymbols(fst.OutputSymbols());
+  }
+
+  RmEpsilonFstImpl(const RmEpsilonFstImpl &impl)
+      : CacheImpl<Arc>(impl),
+        fst_(impl.fst_->Copy(true)),
+        delta_(impl.delta_),
+        rmeps_state_(
+            *fst_, &distance_,
+            RmEpsilonOptions<Arc, FifoQueue<StateId>>(&queue_, delta_, false)) {
+    SetType("rmepsilon");
+    SetProperties(impl.Properties(), kCopyProperties);
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+  }
+
+  StateId Start() {
+    if (!HasStart()) SetStart(fst_->Start());
+    return CacheImpl<Arc>::Start();
+  }
+
+  Weight Final(StateId s) {
+    if (!HasFinal(s)) Expand(s);
+    return CacheImpl<Arc>::Final(s);
+  }
+
+  size_t NumArcs(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<Arc>::NumArcs(s);
+  }
+
+  size_t NumInputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<Arc>::NumInputEpsilons(s);
+  }
+
+  size_t NumOutputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<Arc>::NumOutputEpsilons(s);
+  }
+
+  uint64 Properties() const override { return Properties(kFstProperties); }
+
+  // Sets error if found and returns other FST impl properties.
+  uint64 Properties(uint64 mask) const override {
+    if ((mask & kError) &&
+        (fst_->Properties(kError, false) || rmeps_state_.Error())) {
+      SetProperties(kError, kError);
+    }
+    return FstImpl<Arc>::Properties(mask);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
+    if (!HasArcs(s)) Expand(s);
+    CacheImpl<Arc>::InitArcIterator(s, data);
+  }
+
+  void Expand(StateId s) {
+    rmeps_state_.Expand(s);
+    SetFinal(s, rmeps_state_.Final());
+    auto &arcs = rmeps_state_.Arcs();
+    while (!arcs.empty()) {
+      PushArc(s, std::move(arcs.back()));
+      arcs.pop_back();
+    }
+    SetArcs(s);
+  }
+
+ private:
+  std::unique_ptr<const Fst<Arc>> fst_;
+  float delta_;
+  std::vector<Weight> distance_;
+  FifoQueue<StateId> queue_;
+  internal::RmEpsilonState<Arc, FifoQueue<StateId>> rmeps_state_;
+};
+
+}  // namespace internal
+
+// Removes epsilon-transitions (when both the input and output label are an
+// epsilon) from a transducer. The result will be an equivalent FST that has no
+// such epsilon transitions. This version is a
+// delayed FST.
+//
+// Complexity:
+//
+// - Time:
+//   Unweighted: O(v^2 + ve).
+//   General: exponential.
+//
+// - Space: O(vE)
+//
+// where v is the number of states visited and e is the number of arcs visited.
+// Constant time to visit an input state or arc is assumed and exclusive of
+// caching.
+//
+// For more information, see:
+//
+// Mohri, M. 2002. Generic epsilon-removal and input epsilon-normalization
+// algorithms for weighted transducers. International Journal of Computer
+// Science 13(1): 129-143.
+//
+// This class attaches interface to implementation and handles
+// reference counting, delegating most methods to ImplToFst.
+template <class A>
+class RmEpsilonFst : public ImplToFst<internal::RmEpsilonFstImpl<A>> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+
+  using Store = DefaultCacheStore<Arc>;
+  using State = typename Store::State;
+  using Impl = internal::RmEpsilonFstImpl<Arc>;
+
+  friend class ArcIterator<RmEpsilonFst<Arc>>;
+  friend class StateIterator<RmEpsilonFst<Arc>>;
+
+  explicit RmEpsilonFst(const Fst<Arc> &fst)
+      : ImplToFst<Impl>(std::make_shared<Impl>(fst, RmEpsilonFstOptions())) {}
+
+  RmEpsilonFst(const Fst<A> &fst, const RmEpsilonFstOptions &opts)
+      : ImplToFst<Impl>(std::make_shared<Impl>(fst, opts)) {}
+
+  // See Fst<>::Copy() for doc.
+  RmEpsilonFst(const RmEpsilonFst<Arc> &fst, bool safe = false)
+      : ImplToFst<Impl>(fst, safe) {}
+
+  // Get a copy of this RmEpsilonFst. See Fst<>::Copy() for further doc.
+  RmEpsilonFst<Arc> *Copy(bool safe = false) const override {
+    return new RmEpsilonFst<Arc>(*this, safe);
+  }
+
+  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
+    GetMutableImpl()->InitArcIterator(s, data);
+  }
+
+ private:
+  using ImplToFst<Impl>::GetImpl;
+  using ImplToFst<Impl>::GetMutableImpl;
+
+  RmEpsilonFst &operator=(const RmEpsilonFst &) = delete;
+};
+
+// Specialization for RmEpsilonFst.
+template <class Arc>
+class StateIterator<RmEpsilonFst<Arc>>
+    : public CacheStateIterator<RmEpsilonFst<Arc>> {
+ public:
+  explicit StateIterator(const RmEpsilonFst<Arc> &fst)
+      : CacheStateIterator<RmEpsilonFst<Arc>>(fst, fst.GetMutableImpl()) {}
+};
+
+// Specialization for RmEpsilonFst.
+template <class Arc>
+class ArcIterator<RmEpsilonFst<Arc>>
+    : public CacheArcIterator<RmEpsilonFst<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const RmEpsilonFst<Arc> &fst, StateId s)
+      : CacheArcIterator<RmEpsilonFst<Arc>>(fst.GetMutableImpl(), s) {
+    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
+  }
+};
+
+template <class Arc>
+inline void RmEpsilonFst<Arc>::InitStateIterator(
+    StateIteratorData<Arc> *data) const {
+  data->base = new StateIterator<RmEpsilonFst<Arc>>(*this);
+}
+
+// Useful alias when using StdArc.
+using StdRmEpsilonFst = RmEpsilonFst<StdArc>;
+
+}  // namespace fst
+
+#endif  // FST_RMEPSILON_H_
diff --git a/projects/llm_framework/include/fst/rmfinalepsilon.h b/projects/llm_framework/include/fst/rmfinalepsilon.h
new file mode 100644
index 00000000..87e3a714
--- /dev/null
+++ b/projects/llm_framework/include/fst/rmfinalepsilon.h
@@ -0,0 +1,80 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Function to remove of final states that have epsilon-only input arcs.
+
+#ifndef FST_RMFINALEPSILON_H_
+#define FST_RMFINALEPSILON_H_
+
+#include <unordered_set>
+#include <vector>
+
+#include <fst/connect.h>
+#include <fst/mutable-fst.h>
+
+
+namespace fst {
+
+// Removes final states that have epsilon-only input arcs.
+template <class Arc>
+void RmFinalEpsilon(MutableFst<Arc> *fst) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  // Determines the coaccesibility of states.
+  std::vector<bool> access;
+  std::vector<bool> coaccess;
+  uint64 props = 0;
+  SccVisitor<Arc> scc_visitor(nullptr, &access, &coaccess, &props);
+  DfsVisit(*fst, &scc_visitor);
+  // Finds potential list of removable final states. These are final states that
+  // have no outgoing transitions or final states that have a non-coaccessible
+  // future.
+  std::unordered_set<StateId> finals;
+  for (StateIterator<Fst<Arc>> siter(*fst); !siter.Done(); siter.Next()) {
+    const auto s = siter.Value();
+    if (fst->Final(s) != Weight::Zero()) {
+      bool future_coaccess = false;
+      for (ArcIterator<Fst<Arc>> aiter(*fst, s); !aiter.Done(); aiter.Next()) {
+        const auto &arc = aiter.Value();
+        if (coaccess[arc.nextstate]) {
+          future_coaccess = true;
+          break;
+        }
+      }
+      if (!future_coaccess) finals.insert(s);
+    }
+  }
+  // Moves the final weight.
+  std::vector<Arc> arcs;
+  for (StateIterator<Fst<Arc>> siter(*fst); !siter.Done(); siter.Next()) {
+    const auto s = siter.Value();
+    auto weight = fst->Final(s);
+    arcs.clear();
+    for (ArcIterator<Fst<Arc>> aiter(*fst, s); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      // Next state is in the list of finals.
+      if (finals.find(arc.nextstate) != finals.end()) {
+        // Sums up all epsilon arcs.
+        if (arc.ilabel == 0 && arc.olabel == 0) {
+          weight = Plus(Times(fst->Final(arc.nextstate), arc.weight), weight);
+        } else {
+          arcs.push_back(arc);
+        }
+      } else {
+        arcs.push_back(arc);
+      }
+    }
+    // If some arcs (epsilon arcs) were deleted, delete all arcs and add back
+    // only the non-epsilon arcs.
+    if (arcs.size() < fst->NumArcs(s)) {
+      fst->DeleteArcs(s);
+      fst->SetFinal(s, weight);
+      for (const auto &arc : arcs) fst->AddArc(s, arc);
+    }
+  }
+  Connect(fst);
+}
+
+}  // namespace fst
+
+#endif  // FST_RMFINALEPSILON_H_
diff --git a/projects/llm_framework/include/fst/script/arc-class.h b/projects/llm_framework/include/fst/script/arc-class.h
new file mode 100644
index 00000000..551266d7
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/arc-class.h
@@ -0,0 +1,40 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_ARC_CLASS_H_
+#define FST_SCRIPT_ARC_CLASS_H_
+
+#include <fst/script/weight-class.h>
+
+namespace fst {
+namespace script {
+
+// A struct representing an arc while ignoring arc type. It is passed as an
+// argument to AddArc.
+
+struct ArcClass {
+  template <class Arc>
+  explicit ArcClass(const Arc &arc)
+      : ilabel(arc.ilabel), olabel(arc.olabel), weight(arc.weight),
+        nextstate(arc.nextstate) {}
+
+  ArcClass(int64 ilabel, int64 olabel, const WeightClass &weight,
+           int64 nextstate)
+      : ilabel(ilabel), olabel(olabel), weight(weight), nextstate(nextstate) {}
+
+  template <class Arc>
+  Arc GetArc() const {
+    return Arc(ilabel, olabel, *(weight.GetWeight<typename Arc::Weight>()),
+               nextstate);
+  }
+
+  int64 ilabel;
+  int64 olabel;
+  WeightClass weight;
+  int64 nextstate;
+};
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_ARC_CLASS_H_
diff --git a/projects/llm_framework/include/fst/script/arciterator-class.h b/projects/llm_framework/include/fst/script/arciterator-class.h
new file mode 100644
index 00000000..8e4ca4f8
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/arciterator-class.h
@@ -0,0 +1,212 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_ARCITERATOR_CLASS_H_
+#define FST_SCRIPT_ARCITERATOR_CLASS_H_
+
+#include <memory>
+#include <utility>
+
+#include <fst/fstlib.h>
+#include <fst/script/fst-class.h>
+
+// Scripting API support for ArcIterator.
+//
+// A call to Value() causes the underlying arc to be used to construct the
+// associated ArcClass.
+
+namespace fst {
+namespace script {
+
+// Non-mutable arc iterators.
+
+// Virtual interface implemented by each concrete ArcIteratorImpl<F>.
+class ArcIteratorImplBase {
+ public:
+  virtual bool Done() const = 0;
+  virtual uint32 Flags() const = 0;
+  virtual void Next() = 0;
+  virtual size_t Position() const = 0;
+  virtual void Reset() = 0;
+  virtual void Seek(size_t a) = 0;
+  virtual void SetFlags(uint32 flags, uint32 mask) = 0;
+  virtual ArcClass Value() const = 0;
+  virtual ~ArcIteratorImplBase() {}
+};
+
+// Templated implementation.
+template <class Arc>
+class ArcIteratorClassImpl : public ArcIteratorImplBase {
+ public:
+  explicit ArcIteratorClassImpl(const Fst<Arc> &fst, int64 s)
+      : aiter_(fst, s) {}
+
+  bool Done() const final { return aiter_.Done(); }
+
+  uint32 Flags() const final { return aiter_.Flags(); }
+
+  void Next() final { aiter_.Next(); }
+
+  size_t Position() const final { return aiter_.Position(); }
+
+  void Reset() final { aiter_.Reset(); }
+
+  void Seek(size_t a) final { aiter_.Seek(a); }
+
+  void SetFlags(uint32 flags, uint32 mask) final {
+    aiter_.SetFlags(flags, mask);
+  }
+
+  // This is returned by value because it has not yet been constructed, and
+  // is likely to participate in return-value optimization.
+  ArcClass Value() const final { return ArcClass(aiter_.Value()); }
+
+  ~ArcIteratorClassImpl() final {}
+
+ private:
+  ArcIterator<Fst<Arc>> aiter_;
+};
+
+class ArcIteratorClass;
+
+using InitArcIteratorClassArgs =
+    std::tuple<const FstClass &, int64, ArcIteratorClass *>;
+
+// Untemplated user-facing class holding a templated pimpl.
+class ArcIteratorClass {
+ public:
+  ArcIteratorClass(const FstClass &fst, int64 s);
+
+  template <class Arc>
+  ArcIteratorClass(const Fst<Arc> &fst, int64 s)
+      : impl_(new ArcIteratorClassImpl<Arc>(fst, s)) {}
+
+  bool Done() const { return impl_->Done(); }
+
+  uint32 Flags() const { return impl_->Flags(); }
+
+  void Next() { impl_->Next(); }
+
+  size_t Position() const { return impl_->Position(); }
+
+  void Reset() { impl_->Reset(); }
+
+  void Seek(size_t a) { impl_->Seek(a); }
+
+  void SetFlags(uint32 flags, uint32 mask) { impl_->SetFlags(flags, mask); }
+
+  ArcClass Value() const { return impl_->Value(); }
+
+  template <class Arc>
+  friend void InitArcIteratorClass(InitArcIteratorClassArgs *args);
+
+ private:
+  std::unique_ptr<ArcIteratorImplBase> impl_;
+};
+
+template <class Arc>
+void InitArcIteratorClass(InitArcIteratorClassArgs *args) {
+  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
+  std::get<2>(*args)->impl_.reset(
+      new ArcIteratorClassImpl<Arc>(fst, std::get<1>(*args)));
+}
+
+// Mutable arc iterators.
+
+// Virtual interface implemented by each concrete MutableArcIteratorImpl<F>.
+class MutableArcIteratorImplBase : public ArcIteratorImplBase {
+ public:
+  virtual void SetValue(const ArcClass &) = 0;
+
+  ~MutableArcIteratorImplBase() override {}
+};
+
+// Templated implementation.
+template <class Arc>
+class MutableArcIteratorClassImpl
+    : public MutableArcIteratorImplBase {
+ public:
+  explicit MutableArcIteratorClassImpl(MutableFst<Arc> *fst, int64 s)
+      : aiter_(fst, s) {}
+
+  bool Done() const final { return aiter_.Done(); }
+
+  uint32 Flags() const final { return aiter_.Flags(); }
+
+  void Next() final { aiter_.Next(); }
+
+  size_t Position() const final { return aiter_.Position(); }
+
+  void Reset() final { aiter_.Reset(); }
+
+  void Seek(size_t a) final { aiter_.Seek(a); }
+
+  void SetFlags(uint32 flags, uint32 mask) final {
+    aiter_.SetFlags(flags, mask);
+  }
+
+  void SetValue(const Arc &arc) { aiter_.SetValue(arc); }
+
+  void SetValue(const ArcClass &ac) final { aiter_.SetValue(ac.GetArc<Arc>()); }
+
+  // This is returned by value because it has not yet been constructed, and
+  // is likely to participate in return-value optimization.
+  ArcClass Value() const final { return ArcClass(aiter_.Value()); }
+
+  ~MutableArcIteratorClassImpl() override {}
+
+ private:
+  MutableArcIterator<MutableFst<Arc>> aiter_;
+};
+
+class MutableArcIteratorClass;
+
+using InitMutableArcIteratorClassArgs =
+    std::tuple<MutableFstClass *, int64, MutableArcIteratorClass *>;
+
+// Untemplated user-facing class holding a templated pimpl.
+class MutableArcIteratorClass {
+ public:
+  MutableArcIteratorClass(MutableFstClass *fst, int64 s);
+
+  template <class Arc>
+  MutableArcIteratorClass(MutableFst<Arc> *fst, int64 s)
+      : impl_(new MutableArcIteratorClassImpl<Arc>(fst, s)) {}
+
+  bool Done() const { return impl_->Done(); }
+
+  uint32 Flags() const { return impl_->Flags(); }
+
+  void Next() { impl_->Next(); }
+
+  size_t Position() const { return impl_->Position(); }
+
+  void Reset() { impl_->Reset(); }
+
+  void Seek(size_t a) { impl_->Seek(a); }
+
+  void SetFlags(uint32 flags, uint32 mask) { impl_->SetFlags(flags, mask); }
+
+  void SetValue(const ArcClass &ac) { impl_->SetValue(ac); }
+
+  ArcClass Value() const { return impl_->Value(); }
+
+  template <class Arc>
+  friend void InitMutableArcIteratorClass(
+      InitMutableArcIteratorClassArgs *args);
+
+ private:
+  std::unique_ptr<MutableArcIteratorImplBase> impl_;
+};
+
+template <class Arc>
+void InitMutableArcIteratorClass(InitMutableArcIteratorClassArgs *args) {
+  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
+  std::get<2>(*args)->impl_.reset(
+      new MutableArcIteratorClassImpl<Arc>(fst, std::get<1>(*args)));
+}
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_ARCITERATOR_CLASS_H_
diff --git a/projects/llm_framework/include/fst/script/arcsort.h b/projects/llm_framework/include/fst/script/arcsort.h
new file mode 100644
index 00000000..3e56fe5c
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/arcsort.h
@@ -0,0 +1,44 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_ARCSORT_H_
+#define FST_SCRIPT_ARCSORT_H_
+
+#include <utility>
+
+#include <fst/arcsort.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+enum ArcSortType {
+  ILABEL_SORT,
+  OLABEL_SORT
+};
+
+using ArcSortArgs = std::pair<MutableFstClass *, ArcSortType>;
+
+template <class Arc>
+void ArcSort(ArcSortArgs *args) {
+  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
+  switch (std::get<1>(*args)) {
+    case ILABEL_SORT: {
+      const ILabelCompare<Arc> icomp;
+      ArcSort(fst, icomp);
+      return;
+    }
+    case OLABEL_SORT: {
+      const OLabelCompare<Arc> ocomp;
+      ArcSort(fst, ocomp);
+      return;
+    }
+  }
+}
+
+void ArcSort(MutableFstClass *ofst, ArcSortType);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_ARCSORT_H_
diff --git a/projects/llm_framework/include/fst/script/arg-packs.h b/projects/llm_framework/include/fst/script/arg-packs.h
new file mode 100644
index 00000000..93cd4a05
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/arg-packs.h
@@ -0,0 +1,37 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// std::pair and std::tuple are used for the arguments of FstClass operations.
+//
+// If a function with a return value is required, use the WithReturnValue
+// template as follows:
+//
+// WithReturnValue<bool, std::tuple<...>>
+
+#ifndef FST_SCRIPT_ARG_PACKS_H_
+#define FST_SCRIPT_ARG_PACKS_H_
+
+#include <type_traits>
+
+namespace fst {
+namespace script {
+
+// Tack this on to an existing type to add a return value. The syntax for
+// accessing the args is then slightly more stilted, as you must do an extra
+// member access (since the args are stored as a member of this class).
+
+template <class Retval, class ArgTuple>
+struct WithReturnValue {
+  // Avoid reference-to-reference if ArgTuple is a reference.
+  using Args = typename std::remove_reference<ArgTuple>::type;
+
+  Retval retval;
+  const Args &args;
+
+  explicit WithReturnValue(const Args &args) : args(args) {}
+};
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_ARG_PACKS_H_
diff --git a/projects/llm_framework/include/fst/script/closure.h b/projects/llm_framework/include/fst/script/closure.h
new file mode 100644
index 00000000..7c68604a
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/closure.h
@@ -0,0 +1,28 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_CLOSURE_H_
+#define FST_SCRIPT_CLOSURE_H_
+
+#include <utility>
+
+#include <fst/closure.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using ClosureArgs = std::pair<MutableFstClass *, const ClosureType>;
+
+template <class Arc>
+void Closure(ClosureArgs *args) {
+  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
+  Closure(fst, std::get<1>(*args));
+}
+
+void Closure(MutableFstClass *ofst, ClosureType closure_type);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_CLOSURE_H_
diff --git a/projects/llm_framework/include/fst/script/compile-impl.h b/projects/llm_framework/include/fst/script/compile-impl.h
new file mode 100644
index 00000000..943a0b72
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/compile-impl.h
@@ -0,0 +1,217 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to to compile a binary FST from textual input.
+
+#ifndef FST_SCRIPT_COMPILE_IMPL_H_
+#define FST_SCRIPT_COMPILE_IMPL_H_
+
+#include <iostream>
+#include <memory>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#include <fst/fst.h>
+#include <fst/util.h>
+#include <fst/vector-fst.h>
+#include <unordered_map>
+
+DECLARE_string(fst_field_separator);
+
+namespace fst {
+
+// Compile a binary Fst from textual input, helper class for fstcompile.cc
+// WARNING: Stand-alone use of this class not recommended, most code should
+// read/write using the binary format which is much more efficient.
+template <class Arc>
+class FstCompiler {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  // WARNING: use of negative labels not recommended as it may cause conflicts.
+  // If add_symbols_ is true, then the symbols will be dynamically added to the
+  // symbol tables. This is only useful if you set the (i/o)keep flag to attach
+  // the final symbol table, or use the accessors. (The input symbol tables are
+  // const and therefore not changed.)
+  FstCompiler(std::istream &istrm, const string &source,  // NOLINT
+              const SymbolTable *isyms, const SymbolTable *osyms,
+              const SymbolTable *ssyms, bool accep, bool ikeep,
+              bool okeep, bool nkeep, bool allow_negative_labels = false) {
+    std::unique_ptr<SymbolTable> misyms(isyms ? isyms->Copy() : nullptr);
+    std::unique_ptr<SymbolTable> mosyms(osyms ? osyms->Copy() : nullptr);
+    std::unique_ptr<SymbolTable> mssyms(ssyms ? ssyms->Copy() : nullptr);
+    Init(istrm, source, misyms.get(), mosyms.get(), mssyms.get(), accep,
+         ikeep, okeep, nkeep, allow_negative_labels, false);
+  }
+
+  FstCompiler(std::istream &istrm, const string &source,  // NOLINT
+              SymbolTable *isyms, SymbolTable *osyms, SymbolTable *ssyms,
+              bool accep, bool ikeep, bool okeep, bool nkeep,
+              bool allow_negative_labels, bool add_symbols) {
+    Init(istrm, source, isyms, osyms, ssyms, accep, ikeep, okeep, nkeep,
+         allow_negative_labels, add_symbols);
+  }
+
+  void Init(std::istream &istrm, const string &source,  // NOLINT
+            SymbolTable *isyms, SymbolTable *osyms, SymbolTable *ssyms,
+            bool accep, bool ikeep, bool okeep, bool nkeep,
+            bool allow_negative_labels, bool add_symbols) {
+    nline_ = 0;
+    source_ = source;
+    isyms_ = isyms;
+    osyms_ = osyms;
+    ssyms_ = ssyms;
+    nstates_ = 0;
+    keep_state_numbering_ = nkeep;
+    allow_negative_labels_ = allow_negative_labels;
+    add_symbols_ = add_symbols;
+    bool start_state_populated = false;
+    char line[kLineLen];
+    const string separator = FLAGS_fst_field_separator + "\n";
+    while (istrm.getline(line, kLineLen)) {
+      ++nline_;
+      std::vector<char *> col;
+      SplitString(line, separator.c_str(), &col, true);
+      if (col.empty() || col[0][0] == '\0')
+        continue;
+      if (col.size() > 5 || (col.size() > 4 && accep) ||
+          (col.size() == 3 && !accep)) {
+        FSTERROR() << "FstCompiler: Bad number of columns, source = " << source_
+                   << ", line = " << nline_;
+        fst_.SetProperties(kError, kError);
+        return;
+      }
+      StateId s = StrToStateId(col[0]);
+      while (s >= fst_.NumStates()) fst_.AddState();
+      if (!start_state_populated) {
+        fst_.SetStart(s);
+        start_state_populated = true;
+      }
+
+      Arc arc;
+      StateId d = s;
+      switch (col.size()) {
+        case 1:
+          fst_.SetFinal(s, Weight::One());
+          break;
+        case 2:
+          fst_.SetFinal(s, StrToWeight(col[1], true));
+          break;
+        case 3:
+          arc.nextstate = d = StrToStateId(col[1]);
+          arc.ilabel = StrToILabel(col[2]);
+          arc.olabel = arc.ilabel;
+          arc.weight = Weight::One();
+          fst_.AddArc(s, arc);
+          break;
+        case 4:
+          arc.nextstate = d = StrToStateId(col[1]);
+          arc.ilabel = StrToILabel(col[2]);
+          if (accep) {
+            arc.olabel = arc.ilabel;
+            arc.weight = StrToWeight(col[3], true);
+          } else {
+            arc.olabel = StrToOLabel(col[3]);
+            arc.weight = Weight::One();
+          }
+          fst_.AddArc(s, arc);
+          break;
+        case 5:
+          arc.nextstate = d = StrToStateId(col[1]);
+          arc.ilabel = StrToILabel(col[2]);
+          arc.olabel = StrToOLabel(col[3]);
+          arc.weight = StrToWeight(col[4], true);
+          fst_.AddArc(s, arc);
+      }
+      while (d >= fst_.NumStates()) fst_.AddState();
+    }
+    if (ikeep) fst_.SetInputSymbols(isyms);
+    if (okeep) fst_.SetOutputSymbols(osyms);
+  }
+
+  const VectorFst<Arc> &Fst() const { return fst_; }
+
+ private:
+  // Maximum line length in text file.
+  static constexpr int kLineLen = 8096;
+
+  StateId StrToId(const char *s, SymbolTable *syms, const char *name,
+                  bool allow_negative = false) const {
+    StateId n = 0;
+    if (syms) {
+      n = (add_symbols_) ? syms->AddSymbol(s) : syms->Find(s);
+      if (n == -1 || (!allow_negative && n < 0)) {
+        FSTERROR() << "FstCompiler: Symbol \"" << s
+                   << "\" is not mapped to any integer " << name
+                   << ", symbol table = " << syms->Name()
+                   << ", source = " << source_ << ", line = " << nline_;
+        fst_.SetProperties(kError, kError);
+      }
+    } else {
+      char *p;
+      n = strtoll(s, &p, 10);
+      if (p < s + strlen(s) || (!allow_negative && n < 0)) {
+        FSTERROR() << "FstCompiler: Bad " << name << " integer = \"" << s
+                   << "\", source = " << source_ << ", line = " << nline_;
+        fst_.SetProperties(kError, kError);
+      }
+    }
+    return n;
+  }
+
+  StateId StrToStateId(const char *s) {
+    StateId n = StrToId(s, ssyms_, "state ID");
+    if (keep_state_numbering_) return n;
+    // Remaps state IDs to make dense set.
+    const auto it = states_.find(n);
+    if (it == states_.end()) {
+      states_[n] = nstates_;
+      return nstates_++;
+    } else {
+      return it->second;
+    }
+  }
+
+  StateId StrToILabel(const char *s) const {
+    return StrToId(s, isyms_, "arc ilabel", allow_negative_labels_);
+  }
+
+  StateId StrToOLabel(const char *s) const {
+    return StrToId(s, osyms_, "arc olabel", allow_negative_labels_);
+  }
+
+  Weight StrToWeight(const char *s, bool allow_zero) const {
+    Weight w;
+    std::istringstream strm(s);
+    strm >> w;
+    if (!strm || (!allow_zero && w == Weight::Zero())) {
+      FSTERROR() << "FstCompiler: Bad weight = \"" << s
+                 << "\", source = " << source_ << ", line = " << nline_;
+      fst_.SetProperties(kError, kError);
+      w = Weight::NoWeight();
+    }
+    return w;
+  }
+
+  mutable VectorFst<Arc> fst_;
+  size_t nline_;
+  string source_;       // Text FST source name.
+  SymbolTable *isyms_;  // ilabel symbol table (not owned).
+  SymbolTable *osyms_;  // olabel symbol table (not owned).
+  SymbolTable *ssyms_;  // slabel symbol table (not owned).
+  std::unordered_map<StateId, StateId> states_;  // State ID map.
+  StateId nstates_;                              // Number of seen states.
+  bool keep_state_numbering_;
+  bool allow_negative_labels_;  // Not recommended; may cause conflicts.
+  bool add_symbols_;            // Add to symbol tables on-the fly.
+
+  FstCompiler(const FstCompiler &) = delete;
+  FstCompiler &operator=(const FstCompiler &) = delete;
+};
+
+}  // namespace fst
+
+#endif  // FST_SCRIPT_COMPILE_IMPL_H_
diff --git a/projects/llm_framework/include/fst/script/compile.h b/projects/llm_framework/include/fst/script/compile.h
new file mode 100644
index 00000000..c82ed477
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/compile.h
@@ -0,0 +1,98 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_COMPILE_H_
+#define FST_SCRIPT_COMPILE_H_
+
+#include <istream>
+#include <memory>
+
+#include <fst/script/arg-packs.h>
+#include <fst/script/compile-impl.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+// This operation exists in two forms. 1 is a void operation which writes the
+// compiled machine to disk; 2 returns an FstClass. I/O should normally be done
+// using the binary format for efficiency, so users are STRONGLY ENCOURAGED to
+// use 1 or to construct FSTs using the C++ FST mutation operations.
+
+// Note: it is safe to pass these strings as references because
+// this struct is only used to pass them deeper in the call graph.
+// Be sure you understand why this is so before using this struct
+// for anything else!
+struct CompileFstInnerArgs {
+  std::istream &istrm;
+  const string &source;
+  const string &fst_type;
+  const fst::SymbolTable *isyms;
+  const fst::SymbolTable *osyms;
+  const fst::SymbolTable *ssyms;
+  const bool accep;
+  const bool ikeep;
+  const bool okeep;
+  const bool nkeep;
+  const bool allow_negative_labels;
+
+  CompileFstInnerArgs(std::istream &istrm, const string &source,
+                      const string &fst_type, const fst::SymbolTable *isyms,
+                      const fst::SymbolTable *osyms,
+                      const fst::SymbolTable *ssyms, bool accep, bool ikeep,
+                      bool okeep, bool nkeep,
+                      bool allow_negative_labels = false)
+      : istrm(istrm),
+        source(source),
+        fst_type(fst_type),
+        isyms(isyms),
+        osyms(osyms),
+        ssyms(ssyms),
+        accep(accep),
+        ikeep(ikeep),
+        okeep(okeep),
+        nkeep(nkeep),
+        allow_negative_labels(allow_negative_labels) {}
+};
+
+using CompileFstArgs = WithReturnValue<FstClass *, CompileFstInnerArgs>;
+
+template <class Arc>
+void CompileFstInternal(CompileFstArgs *args) {
+  using fst::Convert;
+  using fst::Fst;
+  using fst::FstCompiler;
+  FstCompiler<Arc> fstcompiler(
+      args->args.istrm, args->args.source, args->args.isyms, args->args.osyms,
+      args->args.ssyms, args->args.accep, args->args.ikeep, args->args.okeep,
+      args->args.nkeep, args->args.allow_negative_labels);
+  const Fst<Arc> *fst = &fstcompiler.Fst();
+  std::unique_ptr<const Fst<Arc>> owned_fst;
+  if (args->args.fst_type != "vector") {
+    owned_fst.reset(Convert<Arc>(*fst, args->args.fst_type));
+    if (!owned_fst) {
+      FSTERROR() << "Failed to convert FST to desired type: "
+                 << args->args.fst_type;
+    }
+    fst = owned_fst.get();
+  }
+  args->retval = fst ? new FstClass(*fst) : nullptr;
+}
+
+void CompileFst(std::istream &istrm, const string &source, const string &dest,
+                const string &fst_type, const string &arc_type,
+                const SymbolTable *isyms, const SymbolTable *osyms,
+                const SymbolTable *ssyms, bool accep, bool ikeep, bool okeep,
+                bool nkeep, bool allow_negative_labels);
+
+FstClass *CompileFstInternal(std::istream &istrm, const string &source,
+                             const string &fst_type, const string &arc_type,
+                             const SymbolTable *isyms, const SymbolTable *osyms,
+                             const SymbolTable *ssyms, bool accep, bool ikeep,
+                             bool okeep, bool nkeep,
+                             bool allow_negative_labels);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_COMPILE_H_
diff --git a/projects/llm_framework/include/fst/script/compose.h b/projects/llm_framework/include/fst/script/compose.h
new file mode 100644
index 00000000..a1735803
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/compose.h
@@ -0,0 +1,34 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_COMPOSE_H_
+#define FST_SCRIPT_COMPOSE_H_
+
+#include <tuple>
+
+#include <fst/compose.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using ComposeArgs = std::tuple<const FstClass &, const FstClass &,
+                               MutableFstClass *, const ComposeOptions &>;
+
+template <class Arc>
+void Compose(ComposeArgs *args) {
+  const Fst<Arc> &ifst1 = *(std::get<0>(*args).GetFst<Arc>());
+  const Fst<Arc> &ifst2 = *(std::get<1>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
+  const auto &opts = std::get<3>(*args);
+  Compose(ifst1, ifst2, ofst, opts);
+}
+
+void Compose(const FstClass &ifst1, const FstClass &ifst2,
+             MutableFstClass *ofst,
+             const ComposeOptions &opts = ComposeOptions());
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_COMPOSE_H_
diff --git a/projects/llm_framework/include/fst/script/concat.h b/projects/llm_framework/include/fst/script/concat.h
new file mode 100644
index 00000000..4bf8dc61
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/concat.h
@@ -0,0 +1,40 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_CONCAT_H_
+#define FST_SCRIPT_CONCAT_H_
+
+#include <utility>
+
+#include <fst/concat.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using ConcatArgs1 = std::pair<MutableFstClass *, const FstClass &>;
+
+template <class Arc>
+void Concat(ConcatArgs1 *args) {
+  MutableFst<Arc> *ofst = std::get<0>(*args)->GetMutableFst<Arc>();
+  const Fst<Arc> &ifst = *(std::get<1>(*args).GetFst<Arc>());
+  Concat(ofst, ifst);
+}
+
+using ConcatArgs2 = std::pair<const FstClass &, MutableFstClass *>;
+
+template <class Arc>
+void Concat(ConcatArgs2 *args) {
+  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
+  Concat(ifst, ofst);
+}
+
+void Concat(MutableFstClass *ofst, const FstClass &ifst);
+
+void Concat(const FstClass &ifst, MutableFstClass *ofst);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_CONCAT_H_
diff --git a/projects/llm_framework/include/fst/script/connect.h b/projects/llm_framework/include/fst/script/connect.h
new file mode 100644
index 00000000..030102ac
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/connect.h
@@ -0,0 +1,23 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_CONNECT_H_
+#define FST_SCRIPT_CONNECT_H_
+
+#include <fst/connect.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+template <class Arc>
+void Connect(MutableFstClass *fst) {
+  Connect(fst->GetMutableFst<Arc>());
+}
+
+void Connect(MutableFstClass *fst);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_CONNECT_H_
diff --git a/projects/llm_framework/include/fst/script/convert.h b/projects/llm_framework/include/fst/script/convert.h
new file mode 100644
index 00000000..1a6eeaa3
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/convert.h
@@ -0,0 +1,35 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_CONVERT_H_
+#define FST_SCRIPT_CONVERT_H_
+
+#include <memory>
+#include <string>
+#include <utility>
+
+#include <fst/register.h>
+#include <fst/script/arg-packs.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using ConvertInnerArgs = std::pair<const FstClass &, const string &>;
+
+using ConvertArgs = WithReturnValue<FstClass *, ConvertInnerArgs>;
+
+template <class Arc>
+void Convert(ConvertArgs *args) {
+  const Fst<Arc> &fst = *(std::get<0>(args->args).GetFst<Arc>());
+  const string &new_type = std::get<1>(args->args);
+  std::unique_ptr<Fst<Arc>> result(Convert(fst, new_type));
+  args->retval = result ? new FstClass(*result) : nullptr;
+}
+
+FstClass *Convert(const FstClass &fst, const string &new_type);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_CONVERT_H_
diff --git a/projects/llm_framework/include/fst/script/decode.h b/projects/llm_framework/include/fst/script/decode.h
new file mode 100644
index 00000000..09f25391
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/decode.h
@@ -0,0 +1,49 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_DECODE_H_
+#define FST_SCRIPT_DECODE_H_
+
+#include <memory>
+#include <string>
+#include <utility>
+
+#include <fst/encode.h>
+#include <fst/script/encodemapper-class.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using DecodeArgs1 = std::pair<MutableFstClass *, const string &>;
+
+template <class Arc>
+void Decode(DecodeArgs1 *args) {
+  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
+  std::unique_ptr<EncodeMapper<Arc>> decoder(
+      EncodeMapper<Arc>::Read(std::get<1>(*args), DECODE));
+  if (!decoder) {
+    fst->SetProperties(kError, kError);
+    return;
+  }
+  Decode(fst, *decoder);
+}
+
+using DecodeArgs2 = std::pair<MutableFstClass *, const EncodeMapperClass &>;
+
+template <class Arc>
+void Decode(DecodeArgs2 *args) {
+  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
+  const EncodeMapper<Arc> &encoder =
+      *(std::get<1>(*args).GetEncodeMapper<Arc>());
+  Decode(fst, encoder);
+}
+
+void Decode(MutableFstClass *fst, const string &coder_fname);
+
+void Decode(MutableFstClass *fst, const EncodeMapperClass &encoder);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_DECODE_H_
diff --git a/projects/llm_framework/include/fst/script/determinize.h b/projects/llm_framework/include/fst/script/determinize.h
new file mode 100644
index 00000000..383a8fe4
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/determinize.h
@@ -0,0 +1,59 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_DETERMINIZE_H_
+#define FST_SCRIPT_DETERMINIZE_H_
+
+#include <tuple>
+
+#include <fst/determinize.h>
+#include <fst/script/fst-class.h>
+#include <fst/script/weight-class.h>
+
+namespace fst {
+namespace script {
+
+struct DeterminizeOptions {
+  const float delta;
+  const WeightClass &weight_threshold;
+  const int64 state_threshold;
+  const int64 subsequential_label;
+  const DeterminizeType det_type;
+  const bool increment_subsequential_label;
+
+  DeterminizeOptions(float delta, const WeightClass &weight_threshold,
+                     int64 state_threshold = kNoStateId,
+                     int64 subsequential_label = 0,
+                     DeterminizeType det_type = DETERMINIZE_FUNCTIONAL,
+                     bool increment_subsequential_label = false)
+      : delta(delta),
+        weight_threshold(weight_threshold),
+        state_threshold(state_threshold),
+        subsequential_label(subsequential_label),
+        det_type(det_type),
+        increment_subsequential_label(increment_subsequential_label) {}
+};
+
+using DeterminizeArgs = std::tuple<const FstClass &, MutableFstClass *,
+                                   const DeterminizeOptions &>;
+
+template <class Arc>
+void Determinize(DeterminizeArgs *args) {
+  using Weight = typename Arc::Weight;
+  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
+  const auto &opts = std::get<2>(*args);
+  const auto weight_threshold = *(opts.weight_threshold.GetWeight<Weight>());
+  const fst::DeterminizeOptions<Arc> detargs(opts.delta, weight_threshold,
+      opts.state_threshold, opts.subsequential_label, opts.det_type,
+      opts.increment_subsequential_label);
+  Determinize(ifst, ofst, detargs);
+}
+
+void Determinize(const FstClass &ifst, MutableFstClass *ofst,
+                 const DeterminizeOptions &opts);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_DETERMINIZE_H_
diff --git a/projects/llm_framework/include/fst/script/difference.h b/projects/llm_framework/include/fst/script/difference.h
new file mode 100644
index 00000000..7af6200a
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/difference.h
@@ -0,0 +1,35 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_DIFFERENCE_H_
+#define FST_SCRIPT_DIFFERENCE_H_
+
+#include <tuple>
+
+#include <fst/difference.h>
+#include <fst/script/compose.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using DifferenceArgs = std::tuple<const FstClass &, const FstClass &,
+                                  MutableFstClass *, const ComposeOptions &>;
+
+template <class Arc>
+void Difference(DifferenceArgs *args) {
+  const Fst<Arc> &ifst1 = *(std::get<0>(*args).GetFst<Arc>());
+  const Fst<Arc> &ifst2 = *(std::get<1>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
+  const auto &opts = std::get<3>(*args);
+  Difference(ifst1, ifst2, ofst, opts);
+}
+
+void Difference(const FstClass &ifst1, const FstClass &ifst2,
+                MutableFstClass *ofst,
+                const ComposeOptions &opts = ComposeOptions());
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_DIFFERENCE_H_
diff --git a/projects/llm_framework/include/fst/script/disambiguate.h b/projects/llm_framework/include/fst/script/disambiguate.h
new file mode 100644
index 00000000..acc1fba2
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/disambiguate.h
@@ -0,0 +1,54 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_DISAMBIGUATE_H_
+#define FST_SCRIPT_DISAMBIGUATE_H_
+
+#include <tuple>
+#include <utility>
+
+#include <fst/disambiguate.h>
+#include <fst/script/fst-class.h>
+#include <fst/script/weight-class.h>
+
+namespace fst {
+namespace script {
+
+struct DisambiguateOptions {
+  const float delta;
+  const WeightClass &weight_threshold;
+  const int64 state_threshold;
+  const int64 subsequential_label;
+
+  DisambiguateOptions(float delta, const WeightClass &weight_threshold,
+                      int64 state_threshold = kNoStateId,
+                      int64 subsequential_label = 0)
+      : delta(delta),
+        weight_threshold(weight_threshold),
+        state_threshold(state_threshold),
+        subsequential_label(subsequential_label) {}
+};
+
+using DisambiguateArgs = std::tuple<const FstClass &, MutableFstClass *,
+                                     const DisambiguateOptions &>;
+
+template <class Arc>
+void Disambiguate(DisambiguateArgs *args) {
+  using Weight = typename Arc::Weight;
+  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
+  const auto &opts = std::get<2>(*args);
+  const auto weight_threshold = *(opts.weight_threshold.GetWeight<Weight>());
+  const fst::DisambiguateOptions<Arc> disargs(opts.delta, weight_threshold,
+                                                  opts.state_threshold,
+                                                  opts.subsequential_label);
+  Disambiguate(ifst, ofst, disargs);
+}
+
+void Disambiguate(const FstClass &ifst, MutableFstClass *ofst,
+                  const DisambiguateOptions &opts);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_DISAMBIGUATE_H_
diff --git a/projects/llm_framework/include/fst/script/draw-impl.h b/projects/llm_framework/include/fst/script/draw-impl.h
new file mode 100644
index 00000000..f204b2e6
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/draw-impl.h
@@ -0,0 +1,227 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to draw a binary FST by producing a text file in dot format, a helper
+// class to fstdraw.cc.
+
+#ifndef FST_SCRIPT_DRAW_IMPL_H_
+#define FST_SCRIPT_DRAW_IMPL_H_
+
+#include <ostream>
+#include <sstream>
+#include <string>
+
+#include <fst/fst.h>
+#include <fst/util.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+
+// Print a binary FST in GraphViz textual format (helper class for fstdraw.cc).
+// WARNING: Stand-alone use not recommend.
+template <class Arc>
+class FstDrawer {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  FstDrawer(const Fst<Arc> &fst, const SymbolTable *isyms,
+            const SymbolTable *osyms, const SymbolTable *ssyms, bool accep,
+            const string &title, float width, float height, bool portrait,
+            bool vertical, float ranksep, float nodesep, int fontsize,
+            int precision, const string &float_format, bool show_weight_one)
+      : fst_(fst),
+        isyms_(isyms),
+        osyms_(osyms),
+        ssyms_(ssyms),
+        accep_(accep && fst.Properties(kAcceptor, true)),
+        ostrm_(nullptr),
+        title_(title),
+        width_(width),
+        height_(height),
+        portrait_(portrait),
+        vertical_(vertical),
+        ranksep_(ranksep),
+        nodesep_(nodesep),
+        fontsize_(fontsize),
+        precision_(precision),
+        float_format_(float_format),
+        show_weight_one_(show_weight_one) {}
+
+  // Draws FST to an output buffer.
+  void Draw(std::ostream *strm, const string &dest) {
+    ostrm_ = strm;
+    SetStreamState(ostrm_);
+    dest_ = dest;
+    StateId start = fst_.Start();
+    if (start == kNoStateId) return;
+    PrintString("digraph FST {\n");
+    if (vertical_) {
+      PrintString("rankdir = BT;\n");
+    } else {
+      PrintString("rankdir = LR;\n");
+    }
+    PrintString("size = \"");
+    Print(width_);
+    PrintString(",");
+    Print(height_);
+    PrintString("\";\n");
+    if (!dest_.empty()) PrintString("label = \"" + title_ + "\";\n");
+    PrintString("center = 1;\n");
+    if (portrait_) {
+      PrintString("orientation = Portrait;\n");
+    } else {
+      PrintString("orientation = Landscape;\n");
+    }
+    PrintString("ranksep = \"");
+    Print(ranksep_);
+    PrintString("\";\n");
+    PrintString("nodesep = \"");
+    Print(nodesep_);
+    PrintString("\";\n");
+    // Initial state first.
+    DrawState(start);
+    for (StateIterator<Fst<Arc>> siter(fst_); !siter.Done(); siter.Next()) {
+      const auto s = siter.Value();
+      if (s != start) DrawState(s);
+    }
+    PrintString("}\n");
+  }
+
+ private:
+  void SetStreamState(std::ostream* strm) const {
+    strm->precision(precision_);
+    if (float_format_ == "e")
+        strm->setf(std::ios_base::scientific, std::ios_base::floatfield);
+    if (float_format_ == "f")
+        strm->setf(std::ios_base::fixed, std::ios_base::floatfield);
+    // O.w. defaults to "g" per standard lib.
+  }
+
+  void PrintString(const string &str) const { *ostrm_ << str; }
+
+  // Escapes backslash and double quote if these occur in the string. Dot will
+  // not deal gracefully with these if they are not escaped.
+  static string Escape(const string &str) {
+    string ns;
+    for (char c : str) {
+      if (c == '\\' || c == '"') ns.push_back('\\');
+      ns.push_back(c);
+    }
+    return ns;
+  }
+
+  void PrintId(StateId id, const SymbolTable *syms, const char *name) const {
+    if (syms) {
+      auto symbol = syms->Find(id);
+      if (symbol.empty()) {
+        FSTERROR() << "FstDrawer: Integer " << id
+                   << " is not mapped to any textual symbol"
+                   << ", symbol table = " << syms->Name()
+                   << ", destination = " << dest_;
+        symbol = "?";
+      }
+      PrintString(Escape(symbol));
+    } else {
+      PrintString(std::to_string(id));
+    }
+  }
+
+  void PrintStateId(StateId s) const { PrintId(s, ssyms_, "state ID"); }
+
+  void PrintILabel(Label label) const {
+    PrintId(label, isyms_, "arc input label");
+  }
+
+  void PrintOLabel(Label label) const {
+    PrintId(label, osyms_, "arc output label");
+  }
+
+  void PrintWeight(Weight w) const {
+    // Weight may have double quote characters in it, so escape it.
+    PrintString(Escape(ToString(w)));
+  }
+
+  template <class T>
+  void Print(T t) const { *ostrm_ << t; }
+
+  template <class T>
+  string ToString(T t) const {
+    std::stringstream ss;
+    SetStreamState(&ss);
+    ss << t;
+    return ss.str();
+  }
+
+  void DrawState(StateId s) const {
+    Print(s);
+    PrintString(" [label = \"");
+    PrintStateId(s);
+    const auto weight = fst_.Final(s);
+    if (weight != Weight::Zero()) {
+      if (show_weight_one_ || (weight != Weight::One())) {
+        PrintString("/");
+        PrintWeight(weight);
+      }
+      PrintString("\", shape = doublecircle,");
+    } else {
+      PrintString("\", shape = circle,");
+    }
+    if (s == fst_.Start()) {
+      PrintString(" style = bold,");
+    } else {
+      PrintString(" style = solid,");
+    }
+    PrintString(" fontsize = ");
+    Print(fontsize_);
+    PrintString("]\n");
+    for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      PrintString("\t");
+      Print(s);
+      PrintString(" -> ");
+      Print(arc.nextstate);
+      PrintString(" [label = \"");
+      PrintILabel(arc.ilabel);
+      if (!accep_) {
+        PrintString(":");
+        PrintOLabel(arc.olabel);
+      }
+      if (show_weight_one_ || (arc.weight != Weight::One())) {
+        PrintString("/");
+        PrintWeight(arc.weight);
+      }
+      PrintString("\", fontsize = ");
+      Print(fontsize_);
+      PrintString("];\n");
+    }
+  }
+
+  const Fst<Arc> &fst_;
+  const SymbolTable *isyms_;  // ilabel symbol table.
+  const SymbolTable *osyms_;  // olabel symbol table.
+  const SymbolTable *ssyms_;  // slabel symbol table.
+  bool accep_;                // Print as acceptor when possible.
+  std::ostream *ostrm_;       // Drawn FST destination.
+  string dest_;               // Drawn FST destination name.
+
+  string title_;
+  float width_;
+  float height_;
+  bool portrait_;
+  bool vertical_;
+  float ranksep_;
+  float nodesep_;
+  int fontsize_;
+  int precision_;
+  string float_format_;
+  bool show_weight_one_;
+
+  FstDrawer(const FstDrawer &) = delete;
+  FstDrawer &operator=(const FstDrawer &) = delete;
+};
+
+}  // namespace fst
+
+#endif  // FST_SCRIPT_DRAW_IMPL_H_
diff --git a/projects/llm_framework/include/fst/script/draw.h b/projects/llm_framework/include/fst/script/draw.h
new file mode 100644
index 00000000..cb37df1e
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/draw.h
@@ -0,0 +1,85 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_DRAW_H_
+#define FST_SCRIPT_DRAW_H_
+
+#include <ostream>
+
+#include <fst/script/draw-impl.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+// Note: it is safe to pass these strings as references because
+// this struct is only used to pass them deeper in the call graph.
+// Be sure you understand why this is so before using this struct
+// for anything else!
+struct FstDrawerArgs {
+  const FstClass &fst;
+  const SymbolTable *isyms;
+  const SymbolTable *osyms;
+  const SymbolTable *ssyms;
+  const bool accep;
+  const string &title;
+  const float width;
+  const float height;
+  const bool portrait;
+  const bool vertical;
+  const float ranksep;
+  const float nodesep;
+  const int fontsize;
+  const int precision;
+  const string &float_format;  // NOLINT
+  const bool show_weight_one;
+  std::ostream *ostrm;
+  const string &dest;
+
+  FstDrawerArgs(const FstClass &fst, const SymbolTable *isyms,
+                const SymbolTable *osyms, const SymbolTable *ssyms, bool accep,
+                const string &title, float width, float height, bool portrait,
+                bool vertical, float ranksep, float nodesep, int fontsize,
+                int precision, const string &float_format,
+                bool show_weight_one, std::ostream *ostrm,  const string &dest)
+      : fst(fst),
+        isyms(isyms),
+        osyms(osyms),
+        ssyms(ssyms),
+        accep(accep),
+        title(title),
+        width(width),
+        height(height),
+        portrait(portrait),
+        vertical(vertical),
+        ranksep(ranksep),
+        nodesep(nodesep),
+        fontsize(fontsize),
+        precision(precision),
+        float_format(float_format),
+        show_weight_one(show_weight_one),
+        ostrm(ostrm),
+        dest(dest) {}
+};
+
+template <class Arc>
+void DrawFst(FstDrawerArgs *args) {
+  const Fst<Arc> &fst = *(args->fst.GetFst<Arc>());
+  FstDrawer<Arc> fstdrawer(fst, args->isyms, args->osyms, args->ssyms,
+      args->accep, args->title, args->width, args->height, args->portrait,
+      args->vertical, args->ranksep, args->nodesep, args->fontsize,
+      args->precision, args->float_format, args->show_weight_one);
+  fstdrawer.Draw(args->ostrm, args->dest);
+}
+
+void DrawFst(const FstClass &fst, const SymbolTable *isyms,
+             const SymbolTable *osyms, const SymbolTable *ssyms, bool accep,
+             const string &title, float width, float height, bool portrait,
+             bool vertical, float ranksep, float nodesep, int fontsize,
+             int precision, const string &float_format, bool show_weight_one,
+             std::ostream *ostrm, const string &dest);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_DRAW_H_
diff --git a/projects/llm_framework/include/fst/script/encode.h b/projects/llm_framework/include/fst/script/encode.h
new file mode 100644
index 00000000..6a869680
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/encode.h
@@ -0,0 +1,51 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_ENCODE_H_
+#define FST_SCRIPT_ENCODE_H_
+
+#include <memory>
+#include <string>
+#include <tuple>
+#include <utility>
+
+#include <fst/encode.h>
+#include <fst/script/encodemapper-class.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using EncodeArgs1 = std::tuple<MutableFstClass *, uint32, bool, const string &>;
+
+template <class Arc>
+void Encode(EncodeArgs1 *args) {
+  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
+  const string &coder_fname = std::get<3>(*args);
+  // If true, reuse encode from disk. If false, make a new encoder and just use
+  // the filename argument as the destination state.
+  std::unique_ptr<EncodeMapper<Arc>> encoder(
+      std::get<2>(*args) ? EncodeMapper<Arc>::Read(coder_fname, ENCODE)
+                         : new EncodeMapper<Arc>(std::get<1>(*args), ENCODE));
+  Encode(fst, encoder.get());
+  if (!std::get<2>(*args)) encoder->Write(coder_fname);
+}
+
+using EncodeArgs2 = std::pair<MutableFstClass *, EncodeMapperClass *>;
+
+template <class Arc>
+void Encode(EncodeArgs2 *args) {
+  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
+  EncodeMapper<Arc> *encoder = std::get<1>(*args)->GetEncodeMapper<Arc>();
+  Encode(fst, encoder);
+}
+
+void Encode(MutableFstClass *fst, uint32 flags, bool reuse_encoder,
+            const string &coder_fname);
+
+void Encode(MutableFstClass *fst, EncodeMapperClass *encoder);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_ENCODE_H_
diff --git a/projects/llm_framework/include/fst/script/encodemapper-class.h b/projects/llm_framework/include/fst/script/encodemapper-class.h
new file mode 100644
index 00000000..b62824f1
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/encodemapper-class.h
@@ -0,0 +1,169 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_ENCODEMAPPER_CLASS_H_
+#define FST_SCRIPT_ENCODEMAPPER_CLASS_H_
+
+#include <memory>
+#include <string>
+#include <iostream>
+
+#include <fst/fstlib.h>
+#include <fst/script/arc-class.h>
+#include <fst/script/fst-class.h>
+
+// Scripting API support for EncodeMapper.
+
+namespace fst {
+namespace script {
+
+// Virtual interface implemented by each concrete EncodeMapperClassImpl<A>.
+class EncodeMapperImplBase {
+ public:
+  // Returns an encoded ArcClass.
+  virtual ArcClass operator()(const ArcClass &a) = 0;
+  virtual const string &ArcType() const = 0;
+  virtual uint32 Flags() const = 0;
+  virtual uint64 Properties(uint64 inprops) = 0;
+  virtual EncodeType Type() const = 0;
+  virtual const SymbolTable *InputSymbols() const = 0;
+  virtual const SymbolTable *OutputSymbols() const = 0;
+  virtual void SetInputSymbols(const SymbolTable *syms) = 0;
+  virtual void SetOutputSymbols(const SymbolTable *syms) = 0;
+  virtual const string &WeightType() const = 0;
+  virtual ~EncodeMapperImplBase() {}
+};
+
+// Templated implementation.
+template <class Arc>
+class EncodeMapperClassImpl : public EncodeMapperImplBase {
+ public:
+  EncodeMapperClassImpl(uint32 flags, EncodeType type)
+      : encoder_(flags, type) {}
+
+  ArcClass operator()(const ArcClass &a) final;
+
+  const string &ArcType() const final { return Arc::Type(); }
+
+  uint32 Flags() const final { return encoder_.Flags(); }
+
+  uint64 Properties(uint64 inprops) final {
+    return encoder_.Properties(inprops);
+  }
+
+  EncodeType Type() const final { return encoder_.Type(); }
+
+  const SymbolTable *InputSymbols() const final {
+    return encoder_.InputSymbols();
+  }
+
+  const SymbolTable *OutputSymbols() const final {
+    return encoder_.OutputSymbols();
+  }
+
+  void SetInputSymbols(const SymbolTable *syms) final {
+    encoder_.SetInputSymbols(syms);
+  }
+
+  void SetOutputSymbols(const SymbolTable *syms) final {
+    encoder_.SetOutputSymbols(syms);
+  }
+
+  const string &WeightType() const final { return Arc::Weight::Type(); }
+
+  ~EncodeMapperClassImpl() override {}
+
+  EncodeMapper<Arc> *GetImpl() const { return &encoder_; }
+
+  EncodeMapper<Arc> *GetImpl() { return &encoder_; }
+
+ private:
+  EncodeMapper<Arc> encoder_;
+};
+
+// This is returned by value because it is very likely to undergo return-value
+// optimization.
+template <class Arc>
+inline ArcClass EncodeMapperClassImpl<Arc>::operator()(const ArcClass &a) {
+  Arc arc(a.ilabel, a.olabel, *(a.weight.GetWeight<typename Arc::Weight>()),
+          a.nextstate);
+  return ArcClass(encoder_(arc));
+}
+
+class EncodeMapperClass;
+
+using InitEncodeMapperClassArgs =
+    std::tuple<uint32, EncodeType, EncodeMapperClass *>;
+
+class EncodeMapperClass {
+ public:
+  EncodeMapperClass(const string &arc_type, uint32 flags, EncodeType type);
+
+  template <class Arc>
+  EncodeMapperClass(uint32 flags, EncodeType type)
+      : impl_(new EncodeMapperClassImpl<Arc>(flags, type)) {}
+
+  ArcClass operator()(const ArcClass &arc) { return (*impl_)(arc); }
+
+  const string &ArcType() const { return impl_->ArcType(); }
+
+  uint32 Flags() const { return impl_->Flags(); }
+
+  uint64 Properties(uint64 inprops) { return impl_->Properties(inprops); }
+
+  EncodeType Type() const { return impl_->Type(); }
+
+  const SymbolTable *InputSymbols() const { return impl_->InputSymbols(); }
+
+  const SymbolTable *OutputSymbols() const { return impl_->OutputSymbols(); }
+
+  void SetInputSymbols(const SymbolTable *syms) {
+    impl_->SetInputSymbols(syms);
+  }
+
+  void SetOutputSymbols(const SymbolTable *syms) {
+    impl_->SetOutputSymbols(syms);
+  }
+
+  const string &WeightType() const { return impl_->WeightType(); }
+
+  template <class Arc>
+  friend void InitEncodeMapperClass(InitEncodeMapperClassArgs *args);
+
+  // Naturally, this exists in non-const and const forms. Encoding arcs or FSTs
+  // mutates the underlying encoder; decoding them does not.
+
+  template <class Arc>
+  EncodeMapper<Arc> *GetEncodeMapper() {
+    if (Arc::Type() != ArcType()) {
+      return nullptr;
+    } else {
+      auto *typed_impl = static_cast<EncodeMapperClassImpl<Arc> *>(impl_.get());
+      return typed_impl->GetImpl();
+    }
+  }
+
+  template <class Arc>
+  const EncodeMapper<Arc> *GetEncodeMapper() const {
+    if (Arc::Type() != ArcType()) {
+      return nullptr;
+    } else {
+      auto *typed_impl = static_cast<EncodeMapperClassImpl<Arc> *>(impl_.get());
+      return typed_impl->GetImpl();
+    }
+  }
+
+ private:
+  std::unique_ptr<EncodeMapperImplBase> impl_;
+};
+
+template <class Arc>
+void InitEncodeMapperClass(InitEncodeMapperClassArgs *args) {
+  std::get<2>(*args)->impl_.reset(
+      new EncodeMapperClassImpl<Arc>(std::get<0>(*args), std::get<1>(*args)));
+}
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_ENCODEMAPPER_CLASS_H_
diff --git a/projects/llm_framework/include/fst/script/epsnormalize.h b/projects/llm_framework/include/fst/script/epsnormalize.h
new file mode 100644
index 00000000..b55fefae
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/epsnormalize.h
@@ -0,0 +1,31 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_EPSNORMALIZE_H_
+#define FST_SCRIPT_EPSNORMALIZE_H_
+
+#include <tuple>
+
+#include <fst/epsnormalize.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using EpsNormalizeArgs = std::tuple<const FstClass &, MutableFstClass *,
+                                    EpsNormalizeType>;
+
+template <class Arc>
+void EpsNormalize(EpsNormalizeArgs *args) {
+  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
+  EpsNormalize(ifst, ofst, std::get<2>(*args));
+}
+
+void EpsNormalize(const FstClass &ifst, MutableFstClass *ofst,
+                  EpsNormalizeType norm_type = EPS_NORM_INPUT);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_EPSNORMALIZE_H_
diff --git a/projects/llm_framework/include/fst/script/equal.h b/projects/llm_framework/include/fst/script/equal.h
new file mode 100644
index 00000000..79ea9aa4
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/equal.h
@@ -0,0 +1,32 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_EQUAL_H_
+#define FST_SCRIPT_EQUAL_H_
+
+#include <tuple>
+
+#include <fst/equal.h>
+#include <fst/script/arg-packs.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using EqualInnerArgs = std::tuple<const FstClass &, const FstClass &, float>;
+
+using EqualArgs = WithReturnValue<bool, EqualInnerArgs>;
+
+template <class Arc>
+void Equal(EqualArgs *args) {
+  const Fst<Arc> &fst1 = *(std::get<0>(args->args).GetFst<Arc>());
+  const Fst<Arc> &fst2 = *(std::get<1>(args->args).GetFst<Arc>());
+  args->retval = Equal(fst1, fst2, std::get<2>(args->args));
+}
+
+bool Equal(const FstClass &fst1, const FstClass &fst2, float delta = kDelta);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_EQUAL_H_
diff --git a/projects/llm_framework/include/fst/script/equivalent.h b/projects/llm_framework/include/fst/script/equivalent.h
new file mode 100644
index 00000000..7cdff45e
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/equivalent.h
@@ -0,0 +1,34 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_EQUIVALENT_H_
+#define FST_SCRIPT_EQUIVALENT_H_
+
+#include <tuple>
+
+#include <fst/equivalent.h>
+#include <fst/script/arg-packs.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using EquivalentInnerArgs = std::tuple<const FstClass &, const FstClass &,
+                                       float>;
+
+using EquivalentArgs = WithReturnValue<bool, EquivalentInnerArgs>;
+
+template <class Arc>
+void Equivalent(EquivalentArgs *args) {
+  const Fst<Arc> &fst1 = *(std::get<0>(args->args).GetFst<Arc>());
+  const Fst<Arc> &fst2 = *(std::get<1>(args->args).GetFst<Arc>());
+  args->retval = Equivalent(fst1, fst2, std::get<2>(args->args));
+}
+
+bool Equivalent(const FstClass &fst1, const FstClass &fst2,
+                float delta = kDelta);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_EQUIVALENT_H_
diff --git a/projects/llm_framework/include/fst/script/fst-class.h b/projects/llm_framework/include/fst/script/fst-class.h
new file mode 100644
index 00000000..07319fc7
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/fst-class.h
@@ -0,0 +1,530 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_FST_CLASS_H_
+#define FST_SCRIPT_FST_CLASS_H_
+
+#include <algorithm>
+#include <limits>
+#include <string>
+#include <type_traits>
+
+#include <fst/expanded-fst.h>
+#include <fst/fst.h>
+#include <fst/mutable-fst.h>
+#include <fst/vector-fst.h>
+#include <fst/script/arc-class.h>
+#include <fst/script/weight-class.h>
+
+// Classes to support "boxing" all existing types of FST arcs in a single
+// FstClass which hides the arc types. This allows clients to load
+// and work with FSTs without knowing the arc type. These classes are only
+// recommended for use in high-level scripting applications. Most users should
+// use the lower-level templated versions corresponding to these classes.
+
+namespace fst {
+namespace script {
+
+// Abstract base class defining the set of functionalities implemented in all
+// impls and passed through by all bases. Below FstClassBase the class
+// hierarchy bifurcates; FstClassImplBase serves as the base class for all
+// implementations (of which FstClassImpl is currently the only one) and
+// FstClass serves as the base class for all interfaces.
+
+class FstClassBase {
+ public:
+  virtual const string &ArcType() const = 0;
+  virtual WeightClass Final(int64) const = 0;
+  virtual const string &FstType() const = 0;
+  virtual const SymbolTable *InputSymbols() const = 0;
+  virtual size_t NumArcs(int64) const = 0;
+  virtual size_t NumInputEpsilons(int64) const = 0;
+  virtual size_t NumOutputEpsilons(int64) const = 0;
+  virtual const SymbolTable *OutputSymbols() const = 0;
+  virtual uint64 Properties(uint64, bool) const = 0;
+  virtual int64 Start() const = 0;
+  virtual const string &WeightType() const = 0;
+  virtual bool ValidStateId(int64) const = 0;
+  virtual bool Write(const string &) const = 0;
+  virtual bool Write(std::ostream &, const string &) const = 0;
+  virtual ~FstClassBase() {}
+};
+
+// Adds all the MutableFst methods.
+class FstClassImplBase : public FstClassBase {
+ public:
+  virtual bool AddArc(int64, const ArcClass &) = 0;
+  virtual int64 AddState() = 0;
+  virtual FstClassImplBase *Copy() = 0;
+  virtual bool DeleteArcs(int64, size_t) = 0;
+  virtual bool DeleteArcs(int64) = 0;
+  virtual bool DeleteStates(const std::vector<int64> &) = 0;
+  virtual void DeleteStates() = 0;
+  virtual SymbolTable *MutableInputSymbols() = 0;
+  virtual SymbolTable *MutableOutputSymbols() = 0;
+  virtual int64 NumStates() const = 0;
+  virtual bool ReserveArcs(int64, size_t) = 0;
+  virtual void ReserveStates(int64) = 0;
+  virtual void SetInputSymbols(SymbolTable *) = 0;
+  virtual bool SetFinal(int64, const WeightClass &) = 0;
+  virtual void SetOutputSymbols(SymbolTable *) = 0;
+  virtual void SetProperties(uint64, uint64) = 0;
+  virtual bool SetStart(int64) = 0;
+  ~FstClassImplBase() override {}
+};
+
+// Containiner class wrapping an Fst<Arc>, hiding its arc type. Whether this
+// Fst<Arc> pointer refers to a special kind of FST (e.g. a MutableFst) is
+// known by the type of interface class that owns the pointer to this
+// container.
+
+template <class Arc>
+class FstClassImpl : public FstClassImplBase {
+ public:
+  explicit FstClassImpl(Fst<Arc> *impl, bool should_own = false)
+      : impl_(should_own ? impl : impl->Copy()) {}
+
+  explicit FstClassImpl(const Fst<Arc> &impl) : impl_(impl.Copy()) {}
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  bool AddArc(int64 s, const ArcClass &ac) final {
+    if (!ValidStateId(s)) return false;
+    // Note that we do not check that the destination state is valid, so users
+    // can add arcs before they add the corresponding states. Verify can be
+    // used to determine whether any arc has a nonexisting destination.
+    Arc arc(ac.ilabel, ac.olabel, *ac.weight.GetWeight<typename Arc::Weight>(),
+            ac.nextstate);
+    static_cast<MutableFst<Arc> *>(impl_.get())->AddArc(s, arc);
+    return true;
+  }
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  int64 AddState() final {
+    return static_cast<MutableFst<Arc> *>(impl_.get())->AddState();
+  }
+
+  const string &ArcType() const final { return Arc::Type(); }
+
+  FstClassImpl *Copy() final { return new FstClassImpl<Arc>(impl_.get()); }
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  bool DeleteArcs(int64 s, size_t n) final {
+    if (!ValidStateId(s)) return false;
+    static_cast<MutableFst<Arc> *>(impl_.get())->DeleteArcs(s, n);
+    return true;
+  }
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  bool DeleteArcs(int64 s) final {
+    if (!ValidStateId(s)) return false;
+    static_cast<MutableFst<Arc> *>(impl_.get())->DeleteArcs(s);
+    return true;
+  }
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  bool DeleteStates(const std::vector<int64> &dstates) final {
+    for (const auto &state : dstates)
+      if (!ValidStateId(state)) return false;
+    // Warning: calling this method with any integers beyond the precision of
+    // the underlying FST will result in truncation.
+    std::vector<typename Arc::StateId> typed_dstates(dstates.size());
+    std::copy(dstates.begin(), dstates.end(), typed_dstates.begin());
+    static_cast<MutableFst<Arc> *>(impl_.get())->DeleteStates(typed_dstates);
+    return true;
+  }
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  void DeleteStates() final {
+    static_cast<MutableFst<Arc> *>(impl_.get())->DeleteStates();
+  }
+
+  WeightClass Final(int64 s) const final {
+    if (!ValidStateId(s)) return WeightClass::NoWeight(WeightType());
+    WeightClass w(impl_->Final(s));
+    return w;
+  }
+
+  const string &FstType() const final { return impl_->Type(); }
+
+  const SymbolTable *InputSymbols() const final {
+    return impl_->InputSymbols();
+  }
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  SymbolTable *MutableInputSymbols() final {
+    return static_cast<MutableFst<Arc> *>(impl_.get())->MutableInputSymbols();
+  }
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  SymbolTable *MutableOutputSymbols() final {
+    return static_cast<MutableFst<Arc> *>(impl_.get())->MutableOutputSymbols();
+  }
+
+  // Signals failure by returning size_t max.
+  size_t NumArcs(int64 s) const final {
+    return ValidStateId(s) ? impl_->NumArcs(s)
+                           : std::numeric_limits<size_t>::max();
+  }
+
+  // Signals failure by returning size_t max.
+  size_t NumInputEpsilons(int64 s) const final {
+    return ValidStateId(s) ? impl_->NumInputEpsilons(s)
+                           : std::numeric_limits<size_t>::max();
+  }
+
+  // Signals failure by returning size_t max.
+  size_t NumOutputEpsilons(int64 s) const final {
+    return ValidStateId(s) ? impl_->NumOutputEpsilons(s)
+                           : std::numeric_limits<size_t>::max();
+  }
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  int64 NumStates() const final {
+    return static_cast<MutableFst<Arc> *>(impl_.get())->NumStates();
+  }
+
+  uint64 Properties(uint64 mask, bool test) const final {
+    return impl_->Properties(mask, test);
+  }
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  bool ReserveArcs(int64 s, size_t n) final {
+    if (!ValidStateId(s)) return false;
+    static_cast<MutableFst<Arc> *>(impl_.get())->ReserveArcs(s, n);
+    return true;
+  }
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  void ReserveStates(int64 s) final {
+    static_cast<MutableFst<Arc> *>(impl_.get())->ReserveStates(s);
+  }
+
+  const SymbolTable *OutputSymbols() const final {
+    return impl_->OutputSymbols();
+  }
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  void SetInputSymbols(SymbolTable *isyms) final {
+    static_cast<MutableFst<Arc> *>(impl_.get())->SetInputSymbols(isyms);
+  }
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  bool SetFinal(int64 s, const WeightClass &weight) final {
+    if (!ValidStateId(s)) return false;
+    static_cast<MutableFst<Arc> *>(impl_.get())
+        ->SetFinal(s, *weight.GetWeight<typename Arc::Weight>());
+    return true;
+  }
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  void SetOutputSymbols(SymbolTable *osyms) final {
+    static_cast<MutableFst<Arc> *>(impl_.get())->SetOutputSymbols(osyms);
+  }
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  void SetProperties(uint64 props, uint64 mask) final {
+    static_cast<MutableFst<Arc> *>(impl_.get())->SetProperties(props, mask);
+  }
+
+  // Warning: calling this method casts the FST to a mutable FST.
+  bool SetStart(int64 s) final {
+    if (!ValidStateId(s)) return false;
+    static_cast<MutableFst<Arc> *>(impl_.get())->SetStart(s);
+    return true;
+  }
+
+  int64 Start() const final { return impl_->Start(); }
+
+  bool ValidStateId(int64 s) const final {
+    // This cowardly refuses to count states if the FST is not yet expanded.
+    if (!Properties(kExpanded, true)) {
+      FSTERROR() << "Cannot get number of states for unexpanded FST";
+      return false;
+    }
+    // If the FST is already expanded, CountStates calls NumStates.
+    if (s < 0 || s >= CountStates(*impl_)) {
+      FSTERROR() << "State ID " << s << " not valid";
+      return false;
+    }
+    return true;
+  }
+
+  const string &WeightType() const final { return Arc::Weight::Type(); }
+
+  bool Write(const string &fname) const final { return impl_->Write(fname); }
+
+  bool Write(std::ostream &ostr, const string &fname) const final {
+    const FstWriteOptions opts(fname);
+    return impl_->Write(ostr, opts);
+  }
+
+  ~FstClassImpl() override {}
+
+  Fst<Arc> *GetImpl() const { return impl_.get(); }
+
+ private:
+  std::unique_ptr<Fst<Arc>> impl_;
+};
+
+// BASE CLASS DEFINITIONS
+
+class MutableFstClass;
+
+class FstClass : public FstClassBase {
+ public:
+  FstClass() : impl_(nullptr) {}
+
+  template <class Arc>
+  explicit FstClass(const Fst<Arc> &fst) : impl_(new FstClassImpl<Arc>(fst)) {}
+
+  FstClass(const FstClass &other)
+      : impl_(other.impl_ == nullptr ? nullptr : other.impl_->Copy()) {}
+
+  FstClass &operator=(const FstClass &other) {
+    impl_.reset(other.impl_ == nullptr ? nullptr : other.impl_->Copy());
+    return *this;
+  }
+
+  WeightClass Final(int64 s) const final { return impl_->Final(s); }
+
+  const string &ArcType() const final { return impl_->ArcType(); }
+
+  const string &FstType() const final { return impl_->FstType(); }
+
+  const SymbolTable *InputSymbols() const final {
+    return impl_->InputSymbols();
+  }
+
+  size_t NumArcs(int64 s) const final { return impl_->NumArcs(s); }
+
+  size_t NumInputEpsilons(int64 s) const final {
+    return impl_->NumInputEpsilons(s);
+  }
+
+  size_t NumOutputEpsilons(int64 s) const final {
+    return impl_->NumOutputEpsilons(s);
+  }
+
+  const SymbolTable *OutputSymbols() const final {
+    return impl_->OutputSymbols();
+  }
+
+  uint64 Properties(uint64 mask, bool test) const final {
+    // Special handling for FSTs with a null impl.
+    if (!impl_) return kError & mask;
+    return impl_->Properties(mask, test);
+  }
+
+  static FstClass *Read(const string &fname);
+
+  static FstClass *Read(std::istream &istrm, const string &source);
+
+  int64 Start() const final { return impl_->Start(); }
+
+  bool ValidStateId(int64 s) const final { return impl_->ValidStateId(s); }
+
+  const string &WeightType() const final { return impl_->WeightType(); }
+
+  // Helper that logs an ERROR if the weight type of an FST and a WeightClass
+  // don't match.
+
+  bool WeightTypesMatch(const WeightClass &weight, const string &op_name) const;
+
+  bool Write(const string &fname) const final { return impl_->Write(fname); }
+
+  bool Write(std::ostream &ostr, const string &fname) const final {
+    return impl_->Write(ostr, fname);
+  }
+
+  ~FstClass() override {}
+
+  // These methods are required by IO registration.
+
+  template <class Arc>
+  static FstClassImplBase *Convert(const FstClass &other) {
+    FSTERROR() << "Doesn't make sense to convert any class to type FstClass";
+    return nullptr;
+  }
+
+  template <class Arc>
+  static FstClassImplBase *Create() {
+    FSTERROR() << "Doesn't make sense to create an FstClass with a "
+               << "particular arc type";
+    return nullptr;
+  }
+
+  template <class Arc>
+  const Fst<Arc> *GetFst() const {
+    if (Arc::Type() != ArcType()) {
+      return nullptr;
+    } else {
+      FstClassImpl<Arc> *typed_impl =
+          static_cast<FstClassImpl<Arc> *>(impl_.get());
+      return typed_impl->GetImpl();
+    }
+  }
+
+  template <class Arc>
+  static FstClass *Read(std::istream &stream, const FstReadOptions &opts) {
+    if (!opts.header) {
+      LOG(ERROR) << "FstClass::Read: Options header not specified";
+      return nullptr;
+    }
+    const FstHeader &hdr = *opts.header;
+    if (hdr.Properties() & kMutable) {
+      return ReadTypedFst<MutableFstClass, MutableFst<Arc>>(stream, opts);
+    } else {
+      return ReadTypedFst<FstClass, Fst<Arc>>(stream, opts);
+    }
+  }
+
+ protected:
+  explicit FstClass(FstClassImplBase *impl) : impl_(impl) {}
+
+  const FstClassImplBase *GetImpl() const { return impl_.get(); }
+
+  FstClassImplBase *GetImpl() { return impl_.get(); }
+
+  // Generic template method for reading an arc-templated FST of type
+  // UnderlyingT, and returning it wrapped as FstClassT, with appropriat
+  // error checking. Called from arc-templated Read() static methods.
+  template <class FstClassT, class UnderlyingT>
+  static FstClassT *ReadTypedFst(std::istream &stream,
+                                 const FstReadOptions &opts) {
+    std::unique_ptr<UnderlyingT> u(UnderlyingT::Read(stream, opts));
+    return u ? new FstClassT(*u) : nullptr;
+  }
+
+ private:
+  std::unique_ptr<FstClassImplBase> impl_;
+};
+
+// Specific types of FstClass with special properties
+
+class MutableFstClass : public FstClass {
+ public:
+  bool AddArc(int64 s, const ArcClass &ac) {
+    if (!WeightTypesMatch(ac.weight, "AddArc")) return false;
+    return GetImpl()->AddArc(s, ac);
+  }
+
+  int64 AddState() { return GetImpl()->AddState(); }
+
+  bool DeleteArcs(int64 s, size_t n) { return GetImpl()->DeleteArcs(s, n); }
+
+  bool DeleteArcs(int64 s) { return GetImpl()->DeleteArcs(s); }
+
+  bool DeleteStates(const std::vector<int64> &dstates) {
+    return GetImpl()->DeleteStates(dstates);
+  }
+
+  void DeleteStates() { GetImpl()->DeleteStates(); }
+
+  SymbolTable *MutableInputSymbols() {
+    return GetImpl()->MutableInputSymbols();
+  }
+
+  SymbolTable *MutableOutputSymbols() {
+    return GetImpl()->MutableOutputSymbols();
+  }
+
+  int64 NumStates() const { return GetImpl()->NumStates(); }
+
+  bool ReserveArcs(int64 s, size_t n) { return GetImpl()->ReserveArcs(s, n); }
+
+  void ReserveStates(int64 s) { GetImpl()->ReserveStates(s); }
+
+  static MutableFstClass *Read(const string &fname, bool convert = false);
+
+  void SetInputSymbols(SymbolTable *isyms) {
+    GetImpl()->SetInputSymbols(isyms);
+  }
+
+  bool SetFinal(int64 s, const WeightClass &weight) {
+    if (!WeightTypesMatch(weight, "SetFinal")) return false;
+    return GetImpl()->SetFinal(s, weight);
+  }
+
+  void SetOutputSymbols(SymbolTable *osyms) {
+    GetImpl()->SetOutputSymbols(osyms);
+  }
+
+  void SetProperties(uint64 props, uint64 mask) {
+    GetImpl()->SetProperties(props, mask);
+  }
+
+  bool SetStart(int64 s) { return GetImpl()->SetStart(s); }
+
+  template <class Arc>
+  explicit MutableFstClass(const MutableFst<Arc> &fst) : FstClass(fst) {}
+
+  // These methods are required by IO registration.
+
+  template <class Arc>
+  static FstClassImplBase *Convert(const FstClass &other) {
+    FSTERROR() << "Doesn't make sense to convert any class to type "
+               << "MutableFstClass";
+    return nullptr;
+  }
+
+  template <class Arc>
+  static FstClassImplBase *Create() {
+    FSTERROR() << "Doesn't make sense to create a MutableFstClass with a "
+               << "particular arc type";
+    return nullptr;
+  }
+
+  template <class Arc>
+  MutableFst<Arc> *GetMutableFst() {
+    Fst<Arc> *fst = const_cast<Fst<Arc> *>(this->GetFst<Arc>());
+    MutableFst<Arc> *mfst = static_cast<MutableFst<Arc> *>(fst);
+    return mfst;
+  }
+
+  template <class Arc>
+  static MutableFstClass *Read(std::istream &stream,
+                               const FstReadOptions &opts) {
+    std::unique_ptr<MutableFst<Arc>> mfst(MutableFst<Arc>::Read(stream, opts));
+    return mfst ? new MutableFstClass(*mfst) : nullptr;
+  }
+
+ protected:
+  explicit MutableFstClass(FstClassImplBase *impl) : FstClass(impl) {}
+};
+
+class VectorFstClass : public MutableFstClass {
+ public:
+  explicit VectorFstClass(FstClassImplBase *impl) : MutableFstClass(impl) {}
+
+  explicit VectorFstClass(const FstClass &other);
+
+  explicit VectorFstClass(const string &arc_type);
+
+  static VectorFstClass *Read(const string &fname);
+
+  template <class Arc>
+  static VectorFstClass *Read(std::istream &stream,
+                              const FstReadOptions &opts) {
+    std::unique_ptr<VectorFst<Arc>> mfst(VectorFst<Arc>::Read(stream, opts));
+    return mfst ? new VectorFstClass(*mfst) : nullptr;
+  }
+
+  template <class Arc>
+  explicit VectorFstClass(const VectorFst<Arc> &fst) : MutableFstClass(fst) {}
+
+  template <class Arc>
+  static FstClassImplBase *Convert(const FstClass &other) {
+    return new FstClassImpl<Arc>(new VectorFst<Arc>(*other.GetFst<Arc>()),
+                                 true);
+  }
+
+  template <class Arc>
+  static FstClassImplBase *Create() {
+    return new FstClassImpl<Arc>(new VectorFst<Arc>(), true);
+  }
+};
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_FST_CLASS_H_
diff --git a/projects/llm_framework/include/fst/script/fstscript-decl.h b/projects/llm_framework/include/fst/script/fstscript-decl.h
new file mode 100644
index 00000000..294d0159
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/fstscript-decl.h
@@ -0,0 +1,32 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Forward declarations for the FST and FST script classes.
+
+#ifndef FST_SCRIPT_FSTSCRIPT_DECL_H_
+#define FST_SCRIPT_FSTSCRIPT_DECL_H_
+
+#include <fst/fst-decl.h>
+
+namespace fst {
+namespace script {
+
+class ArcClass;
+
+class ArcIteratorClass;
+class MutableArcIteratorClass;
+
+class EncodeMapperClass;
+
+class FstClass;
+class MutableFstClass;
+class VectorFstClass;
+
+class StateIteratorClass;
+
+class WeightClass;
+
+}  // namespace script
+}  // namespace fst;
+
+#endif  // FST_SCRIPT_FSTSCRIPT_DECL_H_
diff --git a/projects/llm_framework/include/fst/script/fstscript.h b/projects/llm_framework/include/fst/script/fstscript.h
new file mode 100644
index 00000000..45f16175
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/fstscript.h
@@ -0,0 +1,155 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// The FST script interface permits users to interact with FSTs without knowing
+// their arc type. It does this by mapping compile-time polymorphism (in the
+// form of a arc-templated FST types) onto a shared virtual interface. It also
+// supports arc extension via a DSO interface. Due to the overhead of virtual
+// dispatch and registered function lookups, the script API is somewhat slower
+// then library API provided by types like StdVectorFst, but has the advantage
+// that it is designed not to crash (and to provide useful debugging
+// information) upon common user errors like passing invalid indices or
+// attempting comparison of incompatible FSTs. It is used both by the FST
+// binaries and the Python extension.
+//
+// This header includes all of the FST script functionality.
+
+#ifndef FST_SCRIPT_FSTSCRIPT_H_
+#define FST_SCRIPT_FSTSCRIPT_H_
+
+// Major classes
+#include <fst/script/arciterator-class.h>
+#include <fst/script/encodemapper-class.h>
+#include <fst/script/fst-class.h>
+#include <fst/script/stateiterator-class.h>
+#include <fst/script/text-io.h>
+#include <fst/script/weight-class.h>
+
+// Flag-to-enum parsers.
+#include <fst/script/getters.h>
+// Templates like Operation<> and Apply<>.
+#include <fst/script/script-impl.h>
+
+// Operations.
+#include <fst/script/arcsort.h>
+#include <fst/script/closure.h>
+#include <fst/script/compile.h>
+#include <fst/script/compose.h>
+#include <fst/script/concat.h>
+#include <fst/script/connect.h>
+#include <fst/script/convert.h>
+#include <fst/script/decode.h>
+#include <fst/script/determinize.h>
+#include <fst/script/difference.h>
+#include <fst/script/disambiguate.h>
+#include <fst/script/draw.h>
+#include <fst/script/encode.h>
+#include <fst/script/epsnormalize.h>
+#include <fst/script/equal.h>
+#include <fst/script/equivalent.h>
+#include <fst/script/info.h>
+#include <fst/script/intersect.h>
+#include <fst/script/invert.h>
+#include <fst/script/isomorphic.h>
+#include <fst/script/map.h>
+#include <fst/script/minimize.h>
+#include <fst/script/print.h>
+#include <fst/script/project.h>
+#include <fst/script/prune.h>
+#include <fst/script/push.h>
+#include <fst/script/randequivalent.h>
+#include <fst/script/randgen.h>
+#include <fst/script/relabel.h>
+#include <fst/script/replace.h>
+#include <fst/script/reverse.h>
+#include <fst/script/reweight.h>
+#include <fst/script/rmepsilon.h>
+#include <fst/script/shortest-distance.h>
+#include <fst/script/shortest-path.h>
+#include <fst/script/synchronize.h>
+#include <fst/script/topsort.h>
+#include <fst/script/union.h>
+#include <fst/script/verify.h>
+
+// This class is necessary because registering each of the operations
+// separately overfills the stack, as there's so many of them.
+namespace fst {
+namespace script {
+
+template <class Arc>
+class AllFstOperationsRegisterer {
+ public:
+  AllFstOperationsRegisterer() {
+    RegisterBatch1();
+    RegisterBatch2();
+  }
+
+ private:
+  void RegisterBatch1() {
+    REGISTER_FST_OPERATION(ArcSort, Arc, ArcSortArgs);
+    REGISTER_FST_OPERATION(Closure, Arc, ClosureArgs);
+    REGISTER_FST_OPERATION(CompileFstInternal, Arc, CompileFstArgs);
+    REGISTER_FST_OPERATION(Compose, Arc, ComposeArgs);
+    REGISTER_FST_OPERATION(Concat, Arc, ConcatArgs1);
+    REGISTER_FST_OPERATION(Concat, Arc, ConcatArgs2);
+    REGISTER_FST_OPERATION(Connect, Arc, MutableFstClass);
+    REGISTER_FST_OPERATION(Convert, Arc, ConvertArgs);
+    REGISTER_FST_OPERATION(Decode, Arc, DecodeArgs1);
+    REGISTER_FST_OPERATION(Decode, Arc, DecodeArgs2);
+    REGISTER_FST_OPERATION(Determinize, Arc, DeterminizeArgs);
+    REGISTER_FST_OPERATION(Difference, Arc, DifferenceArgs);
+    REGISTER_FST_OPERATION(Disambiguate, Arc, DisambiguateArgs);
+    REGISTER_FST_OPERATION(DrawFst, Arc, FstDrawerArgs);
+    REGISTER_FST_OPERATION(Encode, Arc, EncodeArgs1);
+    REGISTER_FST_OPERATION(Encode, Arc, EncodeArgs2);
+    REGISTER_FST_OPERATION(EpsNormalize, Arc, EpsNormalizeArgs);
+    REGISTER_FST_OPERATION(Equal, Arc, EqualArgs);
+    REGISTER_FST_OPERATION(Equivalent, Arc, EquivalentArgs);
+    REGISTER_FST_OPERATION(PrintFstInfo, Arc, InfoArgs);
+    REGISTER_FST_OPERATION(GetFstInfo, Arc, GetInfoArgs);
+    REGISTER_FST_OPERATION(InitArcIteratorClass, Arc,
+                           InitArcIteratorClassArgs);
+    REGISTER_FST_OPERATION(InitEncodeMapperClass, Arc,
+                           InitEncodeMapperClassArgs);
+    REGISTER_FST_OPERATION(InitMutableArcIteratorClass, Arc,
+                           InitMutableArcIteratorClassArgs);
+    REGISTER_FST_OPERATION(InitStateIteratorClass, Arc,
+                           InitStateIteratorClassArgs);
+  }
+
+  void RegisterBatch2() {
+    REGISTER_FST_OPERATION(Intersect, Arc, IntersectArgs);
+    REGISTER_FST_OPERATION(Invert, Arc, MutableFstClass);
+    REGISTER_FST_OPERATION(Map, Arc, MapArgs);
+    REGISTER_FST_OPERATION(Minimize, Arc, MinimizeArgs);
+    REGISTER_FST_OPERATION(PrintFst, Arc, FstPrinterArgs);
+    REGISTER_FST_OPERATION(Project, Arc, ProjectArgs);
+    REGISTER_FST_OPERATION(Prune, Arc, PruneArgs1);
+    REGISTER_FST_OPERATION(Prune, Arc, PruneArgs2);
+    REGISTER_FST_OPERATION(Push, Arc, PushArgs1);
+    REGISTER_FST_OPERATION(Push, Arc, PushArgs2);
+    REGISTER_FST_OPERATION(RandEquivalent, Arc, RandEquivalentArgs);
+    REGISTER_FST_OPERATION(RandGen, Arc, RandGenArgs);
+    REGISTER_FST_OPERATION(Relabel, Arc, RelabelArgs1);
+    REGISTER_FST_OPERATION(Relabel, Arc, RelabelArgs2);
+    REGISTER_FST_OPERATION(Replace, Arc, ReplaceArgs);
+    REGISTER_FST_OPERATION(Reverse, Arc, ReverseArgs);
+    REGISTER_FST_OPERATION(Reweight, Arc, ReweightArgs);
+    REGISTER_FST_OPERATION(RmEpsilon, Arc, RmEpsilonArgs);
+    REGISTER_FST_OPERATION(ShortestDistance, Arc, ShortestDistanceArgs1);
+    REGISTER_FST_OPERATION(ShortestDistance, Arc, ShortestDistanceArgs2);
+    REGISTER_FST_OPERATION(ShortestPath, Arc, ShortestPathArgs);
+    REGISTER_FST_OPERATION(Synchronize, Arc, SynchronizeArgs);
+    REGISTER_FST_OPERATION(TopSort, Arc, TopSortArgs);
+    REGISTER_FST_OPERATION(Union, Arc, UnionArgs);
+    REGISTER_FST_OPERATION(Verify, Arc, VerifyArgs);
+  }
+};
+
+}  // namespace script
+}  // namespace fst
+
+#define REGISTER_FST_OPERATIONS(Arc) \
+  AllFstOperationsRegisterer<Arc> register_all_fst_operations##Arc;
+
+#endif  // FST_SCRIPT_FSTSCRIPT_H_
diff --git a/projects/llm_framework/include/fst/script/getters.h b/projects/llm_framework/include/fst/script/getters.h
new file mode 100644
index 00000000..5cd727e8
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/getters.h
@@ -0,0 +1,76 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Getters for converting command-line arguments into the appropriate enums
+// or bitmasks, with the simplest ones defined as inline.
+
+#ifndef FST_SCRIPT_GETTERS_H_
+#define FST_SCRIPT_GETTERS_H_
+
+#include <string>
+
+#include <fst/compose.h>          // For ComposeFilter.
+#include <fst/determinize.h>      // For DeterminizeType.
+#include <fst/encode.h>           // For kEncodeLabels (etc.).
+#include <fst/epsnormalize.h>     // For EpsNormalizeType.
+#include <fst/project.h>          // For ProjectType.
+#include <fst/push.h>             // For kPushWeights (etc.).
+#include <fst/queue.h>            // For QueueType.
+#include <fst/rational.h>         // For ClosureType.
+#include <fst/script/arcsort.h>       // For ArcSortType.
+#include <fst/script/map.h>           // For MapType.
+#include <fst/script/script-impl.h>   // For RandArcSelection.
+
+#include <fst/log.h>
+
+namespace fst {
+namespace script {
+
+bool GetArcSortType(const string &str, ArcSortType *sort_type);
+
+inline ClosureType GetClosureType(bool closure_plus) {
+  return closure_plus ? CLOSURE_PLUS : CLOSURE_STAR;
+}
+
+bool GetComposeFilter(const string &str, ComposeFilter *compose_filter);
+
+bool GetDeterminizeType(const string &str, DeterminizeType *det_type);
+
+inline uint32 GetEncodeFlags(bool encode_labels, bool encode_weights) {
+  return (encode_labels ? kEncodeLabels : 0) |
+         (encode_weights ? kEncodeWeights : 0);
+}
+
+inline EpsNormalizeType GetEpsNormalizeType(bool eps_norm_output) {
+  return eps_norm_output ? EPS_NORM_OUTPUT : EPS_NORM_INPUT;
+}
+
+bool GetMapType(const string &str, MapType *map_type);
+
+inline ProjectType GetProjectType(bool project_output) {
+  return project_output ? PROJECT_OUTPUT : PROJECT_INPUT;
+}
+
+inline uint32 GetPushFlags(bool push_weights, bool push_labels,
+                           bool remove_total_weight, bool remove_common_affix) {
+  return ((push_weights ? kPushWeights : 0) |
+          (push_labels ? kPushLabels : 0) |
+          (remove_total_weight ? kPushRemoveTotalWeight : 0) |
+          (remove_common_affix ? kPushRemoveCommonAffix : 0));
+}
+
+bool GetQueueType(const string &str, QueueType *queue_type);
+
+bool GetRandArcSelection(const string &str, RandArcSelection *ras);
+
+bool GetReplaceLabelType(const string &str, bool epsilon_on_replace,
+                         ReplaceLabelType *rlt);
+
+inline ReweightType GetReweightType(bool to_final) {
+  return to_final ? REWEIGHT_TO_FINAL : REWEIGHT_TO_INITIAL;
+}
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_GETTERS_H_
diff --git a/projects/llm_framework/include/fst/script/info-impl.h b/projects/llm_framework/include/fst/script/info-impl.h
new file mode 100644
index 00000000..e8956498
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/info-impl.h
@@ -0,0 +1,314 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to compute various information about FSTs, a helper class for
+// fstinfo.cc.
+
+#ifndef FST_SCRIPT_INFO_IMPL_H_
+#define FST_SCRIPT_INFO_IMPL_H_
+
+#include <map>
+#include <string>
+#include <vector>
+
+#include <fst/connect.h>
+#include <fst/dfs-visit.h>
+#include <fst/fst.h>
+#include <fst/lookahead-matcher.h>
+#include <fst/matcher.h>
+#include <fst/queue.h>
+#include <fst/test-properties.h>
+#include <fst/verify.h>
+#include <fst/visit.h>
+
+namespace fst {
+
+// Compute various information about FSTs, helper class for fstinfo.cc.
+// WARNING: Stand-alone use of this class is not recommended, most code
+// should call directly the relevant library functions: Fst<Arc>::NumStates,
+// Fst<Arc>::NumArcs, TestProperties, etc.
+class FstInfo {
+ public:
+  FstInfo() {}
+
+  // When info_type is "short" (or "auto" and not an ExpandedFst) then only
+  // minimal info is computed and can be requested.
+  template <typename Arc>
+  FstInfo(const Fst<Arc> &fst, bool test_properties,
+          const string &arc_filter_type = "any",
+          const string &info_type = "auto", bool verify = true)
+      : fst_type_(fst.Type()),
+        input_symbols_(fst.InputSymbols() ? fst.InputSymbols()->Name()
+                                          : "none"),
+        output_symbols_(fst.OutputSymbols() ? fst.OutputSymbols()->Name()
+                                            : "none"),
+        nstates_(0),
+        narcs_(0),
+        start_(kNoStateId),
+        nfinal_(0),
+        nepsilons_(0),
+        niepsilons_(0),
+        noepsilons_(0),
+        ilabel_mult_(0.0),
+        olabel_mult_(0.0),
+        naccess_(0),
+        ncoaccess_(0),
+        nconnect_(0),
+        ncc_(0),
+        nscc_(0),
+        input_match_type_(MATCH_NONE),
+        output_match_type_(MATCH_NONE),
+        input_lookahead_(false),
+        output_lookahead_(false),
+        properties_(0),
+        arc_filter_type_(arc_filter_type),
+        long_info_(true),
+        arc_type_(Arc::Type()) {
+    using Label = typename Arc::Label;
+    using StateId = typename Arc::StateId;
+    using Weight = typename Arc::Weight;
+    if (info_type == "long") {
+      long_info_ = true;
+    } else if (info_type == "short") {
+      long_info_ = false;
+    } else if (info_type == "auto") {
+      long_info_ = fst.Properties(kExpanded, false);
+    } else {
+      FSTERROR() << "Bad info type: " << info_type;
+      return;
+    }
+    if (!long_info_) return;
+    // If the FST is not sane, we return.
+    if (verify && !Verify(fst)) {
+      FSTERROR() << "FstInfo: Verify: FST not well-formed";
+      return;
+    }
+    start_ = fst.Start();
+    properties_ = fst.Properties(kFstProperties, test_properties);
+    for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
+      ++nstates_;
+      const auto s = siter.Value();
+      if (fst.Final(s) != Weight::Zero()) ++nfinal_;
+      std::map<Label, size_t> ilabel_count;
+      std::map<Label, size_t> olabel_count;
+      for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
+        const auto &arc = aiter.Value();
+        ++narcs_;
+        if (arc.ilabel == 0 && arc.olabel == 0) ++nepsilons_;
+        if (arc.ilabel == 0) ++niepsilons_;
+        if (arc.olabel == 0) ++noepsilons_;
+        ++ilabel_count[arc.ilabel];
+        ++olabel_count[arc.olabel];
+      }
+      for (auto it = ilabel_count.begin(); it != ilabel_count.end(); ++it) {
+        ilabel_mult_ += it->second * it->second;
+      }
+      for (auto it = olabel_count.begin(); it != olabel_count.end(); ++it) {
+        olabel_mult_ += it->second * it->second;
+      }
+    }
+    if (narcs_ > 0) {
+      ilabel_mult_ /= narcs_;
+      olabel_mult_ /= narcs_;
+    }
+    {
+      std::vector<StateId> cc;
+      CcVisitor<Arc> cc_visitor(&cc);
+      FifoQueue<StateId> fifo_queue;
+      if (arc_filter_type == "any") {
+        Visit(fst, &cc_visitor, &fifo_queue);
+      } else if (arc_filter_type == "epsilon") {
+        Visit(fst, &cc_visitor, &fifo_queue, EpsilonArcFilter<Arc>());
+      } else if (arc_filter_type == "iepsilon") {
+        Visit(fst, &cc_visitor, &fifo_queue, InputEpsilonArcFilter<Arc>());
+      } else if (arc_filter_type == "oepsilon") {
+        Visit(fst, &cc_visitor, &fifo_queue, OutputEpsilonArcFilter<Arc>());
+      } else {
+        FSTERROR() << "Bad arc filter type: " << arc_filter_type;
+        return;
+      }
+      for (StateId s = 0; s < cc.size(); ++s) {
+        if (cc[s] >= ncc_) ncc_ = cc[s] + 1;
+      }
+    }
+    {
+      std::vector<StateId> scc;
+      std::vector<bool> access, coaccess;
+      uint64 props = 0;
+      SccVisitor<Arc> scc_visitor(&scc, &access, &coaccess, &props);
+      if (arc_filter_type == "any") {
+        DfsVisit(fst, &scc_visitor);
+      } else if (arc_filter_type == "epsilon") {
+        DfsVisit(fst, &scc_visitor, EpsilonArcFilter<Arc>());
+      } else if (arc_filter_type == "iepsilon") {
+        DfsVisit(fst, &scc_visitor, InputEpsilonArcFilter<Arc>());
+      } else if (arc_filter_type == "oepsilon") {
+        DfsVisit(fst, &scc_visitor, OutputEpsilonArcFilter<Arc>());
+      } else {
+        FSTERROR() << "Bad arc filter type: " << arc_filter_type;
+        return;
+      }
+      for (StateId s = 0; s < scc.size(); ++s) {
+        if (access[s]) ++naccess_;
+        if (coaccess[s]) ++ncoaccess_;
+        if (access[s] && coaccess[s]) ++nconnect_;
+        if (scc[s] >= nscc_) nscc_ = scc[s] + 1;
+      }
+    }
+    LookAheadMatcher<Fst<Arc>> imatcher(fst, MATCH_INPUT);
+    input_match_type_ = imatcher.Type(test_properties);
+    input_lookahead_ = imatcher.Flags() & kInputLookAheadMatcher;
+    LookAheadMatcher<Fst<Arc>> omatcher(fst, MATCH_OUTPUT);
+    output_match_type_ = omatcher.Type(test_properties);
+    output_lookahead_ = omatcher.Flags() & kOutputLookAheadMatcher;
+  }
+
+  // Short info.
+
+  const string &FstType() const { return fst_type_; }
+
+  const string &ArcType() const { return arc_type_; }
+
+  const string &InputSymbols() const { return input_symbols_; }
+
+  const string &OutputSymbols() const { return output_symbols_; }
+
+  bool LongInfo() const { return long_info_; }
+
+  const string &ArcFilterType() const { return arc_filter_type_; }
+
+  // Long info.
+
+  MatchType InputMatchType() const {
+    CheckLong();
+    return input_match_type_;
+  }
+
+  MatchType OutputMatchType() const {
+    CheckLong();
+    return output_match_type_;
+  }
+
+  bool InputLookAhead() const {
+    CheckLong();
+    return input_lookahead_;
+  }
+
+  bool OutputLookAhead() const {
+    CheckLong();
+    return output_lookahead_;
+  }
+
+  int64 NumStates() const {
+    CheckLong();
+    return nstates_;
+  }
+
+  size_t NumArcs() const {
+    CheckLong();
+    return narcs_;
+  }
+
+  int64 Start() const {
+    CheckLong();
+    return start_;
+  }
+
+  size_t NumFinal() const {
+    CheckLong();
+    return nfinal_;
+  }
+
+  size_t NumEpsilons() const {
+    CheckLong();
+    return nepsilons_;
+  }
+
+  size_t NumInputEpsilons() const {
+    CheckLong();
+    return niepsilons_;
+  }
+
+  size_t NumOutputEpsilons() const {
+    CheckLong();
+    return noepsilons_;
+  }
+
+  double InputLabelMultiplicity() const {
+    CheckLong();
+    return ilabel_mult_;
+  }
+
+  double OutputLabelMultiplicity() const {
+    CheckLong();
+    return olabel_mult_;
+  }
+
+  size_t NumAccessible() const {
+    CheckLong();
+    return naccess_;
+  }
+
+  size_t NumCoAccessible() const {
+    CheckLong();
+    return ncoaccess_;
+  }
+
+  size_t NumConnected() const {
+    CheckLong();
+    return nconnect_;
+  }
+
+  size_t NumCc() const {
+    CheckLong();
+    return ncc_;
+  }
+
+  size_t NumScc() const {
+    CheckLong();
+    return nscc_;
+  }
+
+  uint64 Properties() const {
+    CheckLong();
+    return properties_;
+  }
+
+ private:
+  void CheckLong() const {
+    if (!long_info_)
+      FSTERROR() << "FstInfo: Method only available with long info signature";
+  }
+
+  string fst_type_;
+  string input_symbols_;
+  string output_symbols_;
+  int64 nstates_;
+  size_t narcs_;
+  int64 start_;
+  size_t nfinal_;
+  size_t nepsilons_;
+  size_t niepsilons_;
+  size_t noepsilons_;
+  double ilabel_mult_;
+  double olabel_mult_;
+  size_t naccess_;
+  size_t ncoaccess_;
+  size_t nconnect_;
+  size_t ncc_;
+  size_t nscc_;
+  MatchType input_match_type_;
+  MatchType output_match_type_;
+  bool input_lookahead_;
+  bool output_lookahead_;
+  uint64 properties_;
+  string arc_filter_type_;
+  bool long_info_;
+  string arc_type_;
+};
+
+void PrintFstInfoImpl(const FstInfo &fstinfo, bool pipe = false);
+
+}  // namespace fst
+
+#endif  // FST_SCRIPT_INFO_IMPL_H_
diff --git a/projects/llm_framework/include/fst/script/info.h b/projects/llm_framework/include/fst/script/info.h
new file mode 100644
index 00000000..039d06d8
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/info.h
@@ -0,0 +1,50 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_INFO_H_
+#define FST_SCRIPT_INFO_H_
+
+#include <string>
+#include <tuple>
+
+#include <fst/script/arg-packs.h>
+#include <fst/script/fst-class.h>
+#include <fst/script/info-impl.h>
+
+namespace fst {
+namespace script {
+
+using InfoArgs = std::tuple<const FstClass &, bool, const string &,
+                            const string &, bool, bool>;
+
+template <class Arc>
+void PrintFstInfo(InfoArgs *args) {
+  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
+  const FstInfo fstinfo(fst, std::get<1>(*args), std::get<2>(*args),
+                        std::get<3>(*args), std::get<4>(*args));
+  PrintFstInfoImpl(fstinfo, std::get<5>(*args));
+  if (std::get<5>(*args)) fst.Write("");
+}
+
+void PrintFstInfo(const FstClass &f, bool test_properties,
+                  const string &arc_filter, const string &info_type, bool pipe,
+                  bool verify);
+
+using GetInfoArgs = std::tuple<const FstClass &, bool, const string &,
+                               const string &, bool, FstInfo *>;
+
+template <class Arc>
+void GetFstInfo(GetInfoArgs *args) {
+  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
+  *(std::get<5>(*args)) = FstInfo(fst, std::get<1>(*args), std::get<2>(*args),
+                                  std::get<3>(*args), std::get<4>(*args));
+}
+
+void GetFstInfo(const FstClass &fst, bool test_properties,
+                const string &arc_filter, const string &info_type, bool verify,
+                FstInfo *info);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_INFO_H_
diff --git a/projects/llm_framework/include/fst/script/intersect.h b/projects/llm_framework/include/fst/script/intersect.h
new file mode 100644
index 00000000..229bd56f
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/intersect.h
@@ -0,0 +1,35 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_INTERSECT_H_
+#define FST_SCRIPT_INTERSECT_H_
+
+#include <tuple>
+
+#include <fst/intersect.h>
+#include <fst/script/compose.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using IntersectArgs = std::tuple<const FstClass &, const FstClass &,
+                                 MutableFstClass *, const ComposeOptions &>;
+
+template <class Arc>
+void Intersect(IntersectArgs *args) {
+  const Fst<Arc> &ifst1 = *(std::get<0>(*args).GetFst<Arc>());
+  const Fst<Arc> &ifst2 = *(std::get<1>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
+  const auto &opts = std::get<3>(*args);
+  Intersect(ifst1, ifst2, ofst, opts);
+}
+
+void Intersect(const FstClass &ifst, const FstClass &ifst2,
+               MutableFstClass *ofst,
+               const ComposeOptions &opts = ComposeOptions());
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_INTERSECT_H_
diff --git a/projects/llm_framework/include/fst/script/invert.h b/projects/llm_framework/include/fst/script/invert.h
new file mode 100644
index 00000000..5bc31317
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/invert.h
@@ -0,0 +1,23 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_INVERT_H_
+#define FST_SCRIPT_INVERT_H_
+
+#include <fst/invert.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+template <class Arc>
+void Invert(MutableFstClass *fst) {
+  Invert(fst->GetMutableFst<Arc>());
+}
+
+void Invert(MutableFstClass *fst);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_INVERT_H_
diff --git a/projects/llm_framework/include/fst/script/isomorphic.h b/projects/llm_framework/include/fst/script/isomorphic.h
new file mode 100644
index 00000000..94ea77f9
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/isomorphic.h
@@ -0,0 +1,34 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_ISOMORPHIC_H_
+#define FST_SCRIPT_ISOMORPHIC_H_
+
+#include <tuple>
+
+#include <fst/isomorphic.h>
+#include <fst/script/arg-packs.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using IsomorphicInnerArgs = std::tuple<const FstClass &, const FstClass &,
+                                       float>;
+
+using IsomorphicArgs = WithReturnValue<bool, IsomorphicInnerArgs>;
+
+template <class Arc>
+void Isomorphic(IsomorphicArgs *args) {
+  const Fst<Arc> &fst1 = *(std::get<0>(args->args).GetFst<Arc>());
+  const Fst<Arc> &fst2 = *(std::get<1>(args->args).GetFst<Arc>());
+  args->retval = Isomorphic(fst1, fst2, std::get<2>(args->args));
+}
+
+bool Isomorphic(const FstClass &fst1, const FstClass &fst2,
+                float delta = kDelta);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_ISOMORPHIC_H_
diff --git a/projects/llm_framework/include/fst/script/map.h b/projects/llm_framework/include/fst/script/map.h
new file mode 100644
index 00000000..158d98aa
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/map.h
@@ -0,0 +1,158 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_MAP_H_
+#define FST_SCRIPT_MAP_H_
+
+#include <memory>
+#include <tuple>
+
+#include <fst/arc-map.h>
+#include <fst/state-map.h>
+#include <fst/script/arg-packs.h>
+#include <fst/script/fst-class.h>
+#include <fst/script/weight-class.h>
+
+namespace fst {
+namespace script {
+
+template <class M>
+Fst<typename M::ToArc> *ArcMap(const Fst<typename M::FromArc> &fst,
+                               const M &mapper) {
+  using ToArc = typename M::ToArc;
+  auto *ofst = new VectorFst<ToArc>;
+  ArcMap(fst, ofst, mapper);
+  return ofst;
+}
+
+template <class M>
+Fst<typename M::ToArc> *StateMap(const Fst<typename M::FromArc> &fst,
+                                 const M &mapper) {
+  using ToArc = typename M::ToArc;
+  auto *ofst = new VectorFst<ToArc>;
+  StateMap(fst, ofst, mapper);
+  return ofst;
+}
+
+enum MapType {
+  ARC_SUM_MAPPER,
+  ARC_UNIQUE_MAPPER,
+  IDENTITY_MAPPER,
+  INPUT_EPSILON_MAPPER,
+  INVERT_MAPPER,
+  OUTPUT_EPSILON_MAPPER,
+  PLUS_MAPPER,
+  POWER_MAPPER,
+  QUANTIZE_MAPPER,
+  RMWEIGHT_MAPPER,
+  SUPERFINAL_MAPPER,
+  TIMES_MAPPER,
+  TO_LOG_MAPPER,
+  TO_LOG64_MAPPER,
+  TO_STD_MAPPER
+};
+
+using MapInnerArgs =
+    std::tuple<const FstClass &, MapType, float, double, const WeightClass &>;
+
+using MapArgs = WithReturnValue<FstClass *, MapInnerArgs>;
+
+template <class Arc>
+void Map(MapArgs *args) {
+  using Weight = typename Arc::Weight;
+  const Fst<Arc> &ifst = *(std::get<0>(args->args).GetFst<Arc>());
+  const auto map_type = std::get<1>(args->args);
+  switch (map_type) {
+    case ARC_SUM_MAPPER: {
+      std::unique_ptr<Fst<Arc>> ofst(StateMap(ifst, ArcSumMapper<Arc>(ifst)));
+      args->retval = new FstClass(*ofst);
+      return;
+    }
+    case ARC_UNIQUE_MAPPER: {
+      std::unique_ptr<Fst<Arc>> ofst(
+          StateMap(ifst, ArcUniqueMapper<Arc>(ifst)));
+      args->retval = new FstClass(*ofst);
+      return;
+    }
+    case IDENTITY_MAPPER: {
+      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, IdentityArcMapper<Arc>()));
+      args->retval = new FstClass(*ofst);
+      return;
+    }
+    case INPUT_EPSILON_MAPPER: {
+      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, InputEpsilonMapper<Arc>()));
+      args->retval = new FstClass(*ofst);
+      return;
+    }
+    case INVERT_MAPPER: {
+      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, InvertWeightMapper<Arc>()));
+      args->retval = new FstClass(*ofst);
+      return;
+    }
+    case OUTPUT_EPSILON_MAPPER: {
+      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, OutputEpsilonMapper<Arc>()));
+      args->retval = new FstClass(*ofst);
+      return;
+    }
+    case PLUS_MAPPER: {
+      const auto weight = *(std::get<4>(args->args).GetWeight<Weight>());
+      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, PlusMapper<Arc>(weight)));
+      args->retval = new FstClass(*ofst);
+      return;
+    }
+    case POWER_MAPPER: {
+      const auto power = std::get<3>(args->args);
+      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, PowerMapper<Arc>(power)));
+      args->retval = new FstClass(*ofst);
+      return;
+    }
+    case QUANTIZE_MAPPER: {
+      const auto delta = std::get<2>(args->args);
+      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, QuantizeMapper<Arc>(delta)));
+      args->retval = new FstClass(*ofst);
+      return;
+    }
+    case RMWEIGHT_MAPPER: {
+      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, RmWeightMapper<Arc>()));
+      args->retval = new FstClass(*ofst);
+      return;
+    }
+    case SUPERFINAL_MAPPER: {
+      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, SuperFinalMapper<Arc>()));
+      args->retval = new FstClass(*ofst);
+      return;
+    }
+    case TIMES_MAPPER: {
+      const auto weight = *(std::get<4>(args->args).GetWeight<Weight>());
+      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, TimesMapper<Arc>(weight)));
+      args->retval = new FstClass(*ofst);
+      return;
+    }
+    case TO_LOG_MAPPER: {
+      std::unique_ptr<Fst<LogArc>> ofst(
+          ArcMap(ifst, WeightConvertMapper<Arc, LogArc>()));
+      args->retval = new FstClass(*ofst);
+      return;
+    }
+    case TO_LOG64_MAPPER: {
+      std::unique_ptr<Fst<Log64Arc>> ofst(
+          ArcMap(ifst, WeightConvertMapper<Arc, Log64Arc>()));
+      args->retval = new FstClass(*ofst);
+      return;
+    }
+    case TO_STD_MAPPER: {
+      std::unique_ptr<Fst<StdArc>> ofst(
+          ArcMap(ifst, WeightConvertMapper<Arc, StdArc>()));
+      args->retval = new FstClass(*ofst);
+      return;
+    }
+  }
+}
+
+FstClass *Map(const FstClass &ifst, MapType map_type, float delta, double power,
+              const WeightClass &weight);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_MAP_H_
diff --git a/projects/llm_framework/include/fst/script/minimize.h b/projects/llm_framework/include/fst/script/minimize.h
new file mode 100644
index 00000000..773e8ef1
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/minimize.h
@@ -0,0 +1,33 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_MINIMIZE_H_
+#define FST_SCRIPT_MINIMIZE_H_
+
+#include <tuple>
+
+#include <fst/minimize.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using MinimizeArgs = std::tuple<MutableFstClass *, MutableFstClass *, float,
+                                bool>;
+
+template <class Arc>
+void Minimize(MinimizeArgs *args) {
+  MutableFst<Arc> *ofst1 = std::get<0>(*args)->GetMutableFst<Arc>();
+  MutableFst<Arc> *ofst2 = (std::get<1>(*args) ?
+                            std::get<1>(*args)->GetMutableFst<Arc>() :
+                            nullptr);
+  Minimize(ofst1, ofst2, std::get<2>(*args), std::get<3>(*args));
+}
+
+void Minimize(MutableFstClass *ofst1, MutableFstClass *ofst2 = nullptr,
+              float delta = kShortestDelta, bool allow_nondet = false);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_MINIMIZE_H_
diff --git a/projects/llm_framework/include/fst/script/print-impl.h b/projects/llm_framework/include/fst/script/print-impl.h
new file mode 100644
index 00000000..539c6d8f
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/print-impl.h
@@ -0,0 +1,132 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Stand-alone class to print out binary FSTs in the AT&T format, a helper
+// class for fstprint.cc.
+
+#ifndef FST_SCRIPT_PRINT_IMPL_H_
+#define FST_SCRIPT_PRINT_IMPL_H_
+
+#include <ostream>
+#include <sstream>
+#include <string>
+
+#include <fst/fstlib.h>
+#include <fst/util.h>
+
+DECLARE_string(fst_field_separator);
+
+namespace fst {
+
+// Print a binary FST in textual format (helper class for fstprint.cc).
+// WARNING: Stand-alone use of this class not recommended, most code should
+// read/write using the binary format which is much more efficient.
+template <class Arc>
+class FstPrinter {
+ public:
+  using StateId = typename Arc::StateId;
+  using Label = typename Arc::Label;
+  using Weight = typename Arc::Weight;
+
+  FstPrinter(const Fst<Arc> &fst, const SymbolTable *isyms,
+             const SymbolTable *osyms, const SymbolTable *ssyms, bool accep,
+             bool show_weight_one, const string &field_separator,
+             const string &missing_symbol = "")
+      : fst_(fst),
+        isyms_(isyms),
+        osyms_(osyms),
+        ssyms_(ssyms),
+        accep_(accep && fst.Properties(kAcceptor, true)),
+        ostrm_(nullptr),
+        show_weight_one_(show_weight_one),
+        sep_(field_separator),
+        missing_symbol_(missing_symbol) {}
+
+  // Prints FST to an output stream.
+  void Print(std::ostream *ostrm, const string &dest) {
+    ostrm_ = ostrm;
+    dest_ = dest;
+    const auto start = fst_.Start();
+    if (start == kNoStateId) return;
+    // Initial state first.
+    PrintState(start);
+    for (StateIterator<Fst<Arc>> siter(fst_); !siter.Done(); siter.Next()) {
+      const auto s = siter.Value();
+      if (s != start) PrintState(s);
+    }
+  }
+
+ private:
+  void PrintId(StateId id, const SymbolTable *syms, const char *name) const {
+    if (syms) {
+      string symbol = syms->Find(id);
+      if (symbol.empty()) {
+        if (missing_symbol_.empty()) {
+          FSTERROR() << "FstPrinter: Integer " << id
+                     << " is not mapped to any textual symbol"
+                     << ", symbol table = " << syms->Name()
+                     << ", destination = " << dest_;
+          symbol = "?";
+        } else {
+          symbol = missing_symbol_;
+        }
+      }
+      *ostrm_ << symbol;
+    } else {
+      *ostrm_ << id;
+    }
+  }
+
+  void PrintStateId(StateId s) const { PrintId(s, ssyms_, "state ID"); }
+
+  void PrintILabel(Label l) const { PrintId(l, isyms_, "arc input label"); }
+
+  void PrintOLabel(Label l) const { PrintId(l, osyms_, "arc output label"); }
+
+  void PrintState(StateId s) const {
+    bool output = false;
+    for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      PrintStateId(s);
+      *ostrm_ << sep_;
+      PrintStateId(arc.nextstate);
+      *ostrm_ << sep_;
+      PrintILabel(arc.ilabel);
+      if (!accep_) {
+        *ostrm_ << sep_;
+        PrintOLabel(arc.olabel);
+      }
+      if (show_weight_one_ || arc.weight != Weight::One())
+        *ostrm_ << sep_ << arc.weight;
+      *ostrm_ << "\n";
+      output = true;
+    }
+    const auto weight = fst_.Final(s);
+    if (weight != Weight::Zero() || !output) {
+      PrintStateId(s);
+      if (show_weight_one_ || weight != Weight::One()) {
+        *ostrm_ << sep_ << weight;
+      }
+      *ostrm_ << "\n";
+    }
+  }
+
+  const Fst<Arc> &fst_;
+  const SymbolTable *isyms_;  // ilabel symbol table.
+  const SymbolTable *osyms_;  // olabel symbol table.
+  const SymbolTable *ssyms_;  // slabel symbol table.
+  bool accep_;                // Print as acceptor when possible?
+  std::ostream *ostrm_;       // Text FST destination.
+  string dest_;               // Text FST destination name.
+  bool show_weight_one_;      // Print weights equal to Weight::One()?
+  string sep_;                // Separator character between fields.
+  string missing_symbol_;     // Symbol to print when lookup fails (default
+                              // "" means raise error).
+                              //
+  FstPrinter(const FstPrinter &) = delete;
+  FstPrinter &operator=(const FstPrinter &) = delete;
+};
+
+}  // namespace fst
+
+#endif  // FST_SCRIPT_PRINT_IMPL_H_
diff --git a/projects/llm_framework/include/fst/script/print.h b/projects/llm_framework/include/fst/script/print.h
new file mode 100644
index 00000000..687606b3
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/print.h
@@ -0,0 +1,79 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_PRINT_H_
+#define FST_SCRIPT_PRINT_H_
+
+#include <ostream>
+
+#include <fst/flags.h>
+#include <fst/script/fst-class.h>
+#include <fst/script/print-impl.h>
+
+DECLARE_string(fst_field_separator);
+
+namespace fst {
+namespace script {
+
+// Note: it is safe to pass these strings as references because
+// this struct is only used to pass them deeper in the call graph.
+// Be sure you understand why this is so before using this struct
+// for anything else!
+struct FstPrinterArgs {
+  const FstClass &fst;
+  const SymbolTable *isyms;
+  const SymbolTable *osyms;
+  const SymbolTable *ssyms;
+  const bool accept;
+  const bool show_weight_one;
+  std::ostream *ostrm;
+  const string &dest;
+  const string &sep;  // NOLINT
+  const string &missing_symbol;
+
+  FstPrinterArgs(const FstClass &fst, const SymbolTable *isyms,
+                 const SymbolTable *osyms, const SymbolTable *ssyms,
+                 bool accept, bool show_weight_one, std::ostream *ostrm,
+                 const string &dest, const string &sep,
+                 const string &missing_sym = "")
+      : fst(fst),
+        isyms(isyms),
+        osyms(osyms),
+        ssyms(ssyms),
+        accept(accept),
+        show_weight_one(show_weight_one),
+        ostrm(ostrm),
+        dest(dest),
+        sep(sep),
+        missing_symbol(missing_sym) {}
+};
+
+template <class Arc>
+void PrintFst(FstPrinterArgs *args) {
+  const Fst<Arc> &fst = *(args->fst.GetFst<Arc>());
+  FstPrinter<Arc> fstprinter(fst, args->isyms, args->osyms, args->ssyms,
+                             args->accept, args->show_weight_one, args->sep,
+                             args->missing_symbol);
+  fstprinter.Print(args->ostrm, args->dest);
+}
+
+void PrintFst(const FstClass &fst, std::ostream &ostrm, const string &dest,
+              const SymbolTable *isyms, const SymbolTable *osyms,
+              const SymbolTable *ssyms, bool accept, bool show_weight_one,
+              const string &missing_sym = "");
+
+// The same, but with more sensible defaults.
+template <class Arc>
+void PrintFst(const Fst<Arc> &fst, std::ostream &ostrm, const string &dest = "",
+              const SymbolTable *isyms = nullptr,
+              const SymbolTable *osyms = nullptr,
+              const SymbolTable *ssyms = nullptr) {
+  const string sep = FLAGS_fst_field_separator.substr(0, 1);
+  FstPrinter<Arc> fstprinter(fst, isyms, osyms, ssyms, true, true, sep);
+  fstprinter.Print(&ostrm, dest);
+}
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_PRINT_H_
diff --git a/projects/llm_framework/include/fst/script/project.h b/projects/llm_framework/include/fst/script/project.h
new file mode 100644
index 00000000..13edeb1d
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/project.h
@@ -0,0 +1,28 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_PROJECT_H_
+#define FST_SCRIPT_PROJECT_H_
+
+#include <utility>
+
+#include <fst/project.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using ProjectArgs = std::pair<MutableFstClass *, ProjectType>;
+
+template <class Arc>
+void Project(ProjectArgs *args) {
+  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
+  Project(fst, std::get<1>(*args));
+}
+
+void Project(MutableFstClass *fst, ProjectType project_type);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_PROJECT_H_
diff --git a/projects/llm_framework/include/fst/script/prune.h b/projects/llm_framework/include/fst/script/prune.h
new file mode 100644
index 00000000..ed10b540
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/prune.h
@@ -0,0 +1,51 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_PRUNE_H_
+#define FST_SCRIPT_PRUNE_H_
+
+#include <tuple>
+#include <utility>
+
+#include <fst/prune.h>
+#include <fst/script/fst-class.h>
+#include <fst/script/weight-class.h>
+
+namespace fst {
+namespace script {
+
+using PruneArgs1 = std::tuple<const FstClass &, MutableFstClass *,
+                              const WeightClass &, int64, float>;
+
+template <class Arc>
+void Prune(PruneArgs1 *args) {
+  using Weight = typename Arc::Weight;
+  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
+  const auto weight_threshold = *(std::get<2>(*args).GetWeight<Weight>());
+  Prune(ifst, ofst, weight_threshold, std::get<3>(*args), std::get<4>(*args));
+}
+
+using PruneArgs2 = std::tuple<MutableFstClass *, const WeightClass &, int64,
+                               float>;
+
+template <class Arc>
+void Prune(PruneArgs2 *args) {
+  using Weight = typename Arc::Weight;
+  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
+  const auto weight_threshold = *(std::get<1>(*args).GetWeight<Weight>());
+  Prune(fst, weight_threshold, std::get<2>(*args), std::get<3>(*args));
+}
+
+void Prune(const FstClass &ifst, MutableFstClass *ofst,
+           const WeightClass &weight_threshold,
+           int64 state_threshold = kNoStateId,
+           float delta = kDelta);
+
+void Prune(MutableFstClass *fst, const WeightClass &weight_threshold,
+           int64 state_threshold = kNoStateId, float delta = kDelta);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_PRUNE_H_
diff --git a/projects/llm_framework/include/fst/script/push.h b/projects/llm_framework/include/fst/script/push.h
new file mode 100644
index 00000000..018cd8f8
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/push.h
@@ -0,0 +1,53 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_PUSH_H_
+#define FST_SCRIPT_PUSH_H_
+
+#include <tuple>
+
+#include <fst/push.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using PushArgs1 = std::tuple<MutableFstClass *, ReweightType, float, bool>;
+
+template <class Arc>
+void Push(PushArgs1 *args) {
+  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
+  Push(fst, std::get<1>(*args), std::get<2>(*args), std::get<3>(*args));
+}
+
+using PushArgs2 = std::tuple<const FstClass &, MutableFstClass *, uint32,
+                             ReweightType, float>;
+
+template <class Arc>
+void Push(PushArgs2 *args) {
+  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
+  switch (std::get<3>(*args)) {
+    case REWEIGHT_TO_FINAL: {
+      Push<Arc, REWEIGHT_TO_FINAL>(ifst, ofst, std::get<2>(*args),
+                                   std::get<4>(*args));
+      return;
+    }
+    case REWEIGHT_TO_INITIAL: {
+      Push<Arc, REWEIGHT_TO_INITIAL>(ifst, ofst, std::get<2>(*args),
+                                     std::get<4>(*args));
+      return;
+    }
+  }
+}
+
+void Push(MutableFstClass *fst, ReweightType rew_type, float delta = kDelta,
+          bool remove_total_weight = false);
+
+void Push(const FstClass &ifst, MutableFstClass *ofst, uint32 flags,
+          ReweightType rew_type, float delta = kDelta);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_PUSH_H_
diff --git a/projects/llm_framework/include/fst/script/randequivalent.h b/projects/llm_framework/include/fst/script/randequivalent.h
new file mode 100644
index 00000000..945f8a06
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/randequivalent.h
@@ -0,0 +1,67 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_RANDEQUIVALENT_H_
+#define FST_SCRIPT_RANDEQUIVALENT_H_
+
+#include <climits>
+#include <ctime>
+
+#include <tuple>
+
+#include <fst/randequivalent.h>
+#include <fst/script/arg-packs.h>
+#include <fst/script/fst-class.h>
+#include <fst/script/script-impl.h>
+
+namespace fst {
+namespace script {
+
+using RandEquivalentInnerArgs = std::tuple<const FstClass &, const FstClass &,
+    int32, float, time_t, const RandGenOptions<RandArcSelection> &>;
+
+using RandEquivalentArgs = WithReturnValue<bool, RandEquivalentInnerArgs>;
+
+template <class Arc>
+void RandEquivalent(RandEquivalentArgs *args) {
+  const Fst<Arc> &fst1 = *(std::get<0>(args->args).GetFst<Arc>());
+  const Fst<Arc> &fst2 = *(std::get<1>(args->args).GetFst<Arc>());
+  const auto seed = std::get<4>(args->args);
+  const auto &opts = std::get<5>(args->args);
+  switch (opts.selector) {
+    case UNIFORM_ARC_SELECTOR: {
+      const UniformArcSelector<Arc> selector(seed);
+      const RandGenOptions<UniformArcSelector<Arc>> ropts(selector,
+                                                          opts.max_length);
+      args->retval = RandEquivalent(fst1, fst2, std::get<2>(args->args),
+                                    std::get<3>(args->args), ropts);
+      return;
+    }
+    case FAST_LOG_PROB_ARC_SELECTOR: {
+      const FastLogProbArcSelector<Arc> selector(seed);
+      const RandGenOptions<FastLogProbArcSelector<Arc>> ropts(selector,
+                                                              opts.max_length);
+      args->retval = RandEquivalent(fst1, fst2, std::get<2>(args->args),
+                                    std::get<3>(args->args), ropts);
+      return;
+    }
+    case LOG_PROB_ARC_SELECTOR: {
+      const LogProbArcSelector<Arc> selector(seed);
+      const RandGenOptions<LogProbArcSelector<Arc>> ropts(selector,
+                                                          opts.max_length);
+      args->retval = RandEquivalent(fst1, fst2, std::get<2>(args->args),
+                                    std::get<3>(args->args), ropts);
+      return;
+    }
+  }
+}
+
+bool RandEquivalent(const FstClass &fst1, const FstClass &fst2, int32 npath = 1,
+    float delta = kDelta, time_t seed = time(nullptr),
+    const RandGenOptions<RandArcSelection> &opts =
+        RandGenOptions<RandArcSelection>(UNIFORM_ARC_SELECTOR));
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_RANDEQUIVALENT_H_
diff --git a/projects/llm_framework/include/fst/script/randgen.h b/projects/llm_framework/include/fst/script/randgen.h
new file mode 100644
index 00000000..5ce79d01
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/randgen.h
@@ -0,0 +1,63 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_RANDGEN_H_
+#define FST_SCRIPT_RANDGEN_H_
+
+#include <ctime>
+
+#include <tuple>
+
+#include <fst/randgen.h>
+#include <fst/script/fst-class.h>
+#include <fst/script/script-impl.h>
+
+namespace fst {
+namespace script {
+
+using RandGenArgs = std::tuple<const FstClass &, MutableFstClass *, time_t,
+                               const RandGenOptions<RandArcSelection> &>;
+
+template <class Arc>
+void RandGen(RandGenArgs *args) {
+  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
+  const time_t seed = std::get<2>(*args);
+  const auto &opts = std::get<3>(*args);
+  switch (opts.selector) {
+    case UNIFORM_ARC_SELECTOR: {
+      const UniformArcSelector<Arc> selector(seed);
+      const RandGenOptions<UniformArcSelector<Arc>> ropts(
+          selector, opts.max_length, opts.npath, opts.weighted,
+          opts.remove_total_weight);
+      RandGen(ifst, ofst, ropts);
+      return;
+    }
+    case FAST_LOG_PROB_ARC_SELECTOR: {
+      const FastLogProbArcSelector<Arc> selector(seed);
+      const RandGenOptions<FastLogProbArcSelector<Arc>> ropts(
+          selector, opts.max_length, opts.npath, opts.weighted,
+          opts.remove_total_weight);
+      RandGen(ifst, ofst, ropts);
+      return;
+    }
+    case LOG_PROB_ARC_SELECTOR: {
+      const LogProbArcSelector<Arc> selector(seed);
+      const RandGenOptions<LogProbArcSelector<Arc>> ropts(
+          selector, opts.max_length, opts.npath, opts.weighted,
+          opts.remove_total_weight);
+      RandGen(ifst, ofst, ropts);
+      return;
+    }
+  }
+}
+
+void RandGen(const FstClass &ifst, MutableFstClass *ofst,
+             time_t seed = time(nullptr),
+             const RandGenOptions<RandArcSelection> &opts =
+                 RandGenOptions<RandArcSelection>(UNIFORM_ARC_SELECTOR));
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_RANDGEN_H_
diff --git a/projects/llm_framework/include/fst/script/register.h b/projects/llm_framework/include/fst/script/register.h
new file mode 100644
index 00000000..d66e7ade
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/register.h
@@ -0,0 +1,99 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_REGISTER_H_
+#define FST_SCRIPT_REGISTER_H_
+
+#include <istream>
+#include <string>
+
+#include <fst/generic-register.h>
+#include <fst/script/fst-class.h>
+#include <fst/script/weight-class.h>
+
+// Holds methods and classes responsible for maintaining
+// the register for FstClass arc types.
+
+namespace fst {
+namespace script {
+
+// Registers for reading and converting various kinds of FST classes.
+
+// This class definition is to avoid a nested class definition inside the
+// IORegistration struct.
+
+template <class Reader, class Creator, class Converter>
+struct FstClassRegEntry {
+  Reader reader;
+  Creator creator;
+  Converter converter;
+
+  FstClassRegEntry(Reader r, Creator cr, Converter co)
+      : reader(r), creator(cr), converter(co) {}
+
+  FstClassRegEntry()
+      : reader(nullptr), creator(nullptr), converter(nullptr) {}
+};
+
+template <class Reader, class Creator, class Converter>
+class FstClassIORegister
+    : public GenericRegister<string,
+                             FstClassRegEntry<Reader, Creator, Converter>,
+                             FstClassIORegister<Reader, Creator, Converter>> {
+ public:
+  Reader GetReader(const string &arc_type) const {
+    return this->GetEntry(arc_type).reader;
+  }
+
+  Creator GetCreator(const string &arc_type) const {
+    return this->GetEntry(arc_type).creator;
+  }
+
+  Converter GetConverter(const string &arc_type) const {
+    return this->GetEntry(arc_type).converter;
+  }
+
+ protected:
+  string ConvertKeyToSoFilename(const string &key) const final {
+    string legal_type(key);
+    ConvertToLegalCSymbol(&legal_type);
+    return legal_type + "-arc.so";
+  }
+};
+
+// Struct containing everything needed to register a particular type
+// of FST class (e.g., a plain FstClass, or a MutableFstClass, etc.).
+template <class FstClassType>
+struct IORegistration {
+  using Reader = FstClassType *(*)(std::istream &stream,
+                                   const FstReadOptions &opts);
+
+  using Creator = FstClassImplBase *(*)();
+
+  using Converter = FstClassImplBase *(*)(const FstClass &other);
+
+  using Entry = FstClassRegEntry<Reader, Creator, Converter>;
+
+  // FST class Register.
+  using Register = FstClassIORegister<Reader, Creator, Converter>;
+
+  // FST class Register-er.
+  using Registerer =
+      GenericRegisterer<FstClassIORegister<Reader, Creator, Converter>>;
+};
+
+#define REGISTER_FST_CLASS(Class, Arc)                                   \
+  static IORegistration<Class>::Registerer Class##_##Arc##_registerer(   \
+      Arc::Type(),                                                       \
+      IORegistration<Class>::Entry(Class::Read<Arc>, Class::Create<Arc>, \
+                                   Class::Convert<Arc>))
+
+#define REGISTER_FST_CLASSES(Arc)           \
+  REGISTER_FST_CLASS(FstClass, Arc);        \
+  REGISTER_FST_CLASS(MutableFstClass, Arc); \
+  REGISTER_FST_CLASS(VectorFstClass, Arc);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_REGISTER_H_
diff --git a/projects/llm_framework/include/fst/script/relabel.h b/projects/llm_framework/include/fst/script/relabel.h
new file mode 100644
index 00000000..74443490
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/relabel.h
@@ -0,0 +1,64 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_RELABEL_H_
+#define FST_SCRIPT_RELABEL_H_
+
+#include <algorithm>
+#include <tuple>
+#include <utility>
+#include <vector>
+
+#include <fst/relabel.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using RelabelArgs1 = std::tuple<MutableFstClass *, const SymbolTable *,
+                                const SymbolTable *, const string &, bool,
+                                const SymbolTable *, const SymbolTable *,
+                                const string &, bool>;
+
+template <class Arc>
+void Relabel(RelabelArgs1 *args) {
+  MutableFst<Arc> *ofst = std::get<0>(*args)->GetMutableFst<Arc>();
+  Relabel(ofst, std::get<1>(*args), std::get<2>(*args), std::get<3>(*args),
+          std::get<4>(*args), std::get<5>(*args), std::get<6>(*args),
+          std::get<7>(*args), std::get<8>(*args));
+}
+
+using LabelPair = std::pair<int64, int64>;
+
+using RelabelArgs2 = std::tuple<MutableFstClass *,
+                                const std::vector<LabelPair> &,
+                                const std::vector<LabelPair> &>;
+
+template <class Arc>
+void Relabel(RelabelArgs2 *args) {
+  MutableFst<Arc> *ofst = std::get<0>(*args)->GetMutableFst<Arc>();
+  using LabelPair = std::pair<typename Arc::Label, typename Arc::Label>;
+  // In case the MutableFstClass::Label is not the same as Arc::Label,
+  // make a copy.
+  std::vector<LabelPair> typed_ipairs(std::get<1>(*args).size());
+  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
+            typed_ipairs.begin());
+  std::vector<LabelPair> typed_opairs(std::get<2>(*args).size());
+  std::copy(std::get<2>(*args).begin(), std::get<2>(*args).end(),
+            typed_opairs.begin());
+  Relabel(ofst, typed_ipairs, typed_opairs);
+}
+
+void Relabel(MutableFstClass *ofst,
+             const SymbolTable *old_isymbols, const SymbolTable *new_isymbols,
+             const string &unknown_isymbol,  bool attach_new_isymbols,
+             const SymbolTable *old_osymbols, const SymbolTable *new_osymbols,
+             const string &unknown_osymbol, bool attach_new_osymbols);
+
+void Relabel(MutableFstClass *ofst, const std::vector<LabelPair> &ipairs,
+             const std::vector<LabelPair> &opairs);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_RELABEL_H_
diff --git a/projects/llm_framework/include/fst/script/replace.h b/projects/llm_framework/include/fst/script/replace.h
new file mode 100644
index 00000000..926ece24
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/replace.h
@@ -0,0 +1,72 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_REPLACE_H_
+#define FST_SCRIPT_REPLACE_H_
+
+#include <tuple>
+#include <utility>
+#include <vector>
+
+#include <fst/replace.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+struct ReplaceOptions {
+  const int64 root;                          // Root rule for expansion.
+  const ReplaceLabelType call_label_type;    // How to label call arc.
+  const ReplaceLabelType return_label_type;  // How to label return arc.
+  const int64 return_label;                  // Specifies return arc label.
+
+  explicit ReplaceOptions(int64 root,
+      ReplaceLabelType call_label_type = REPLACE_LABEL_INPUT,
+      ReplaceLabelType return_label_type = REPLACE_LABEL_NEITHER,
+      int64 return_label = 0)
+      : root(root),
+        call_label_type(call_label_type),
+        return_label_type(return_label_type),
+        return_label(return_label) {}
+};
+
+using LabelFstClassPair = std::pair<int64, const FstClass *>;
+
+using ReplaceArgs = std::tuple<const std::vector<LabelFstClassPair> &,
+                               MutableFstClass *, const ReplaceOptions &>;
+
+template <class Arc>
+void Replace(ReplaceArgs *args) {
+  using LabelFstPair = std::pair<typename Arc::Label, const Fst<Arc> *>;
+  // Now that we know the arc type, we construct a vector of
+  // std::pair<real label, real fst> that the real Replace will use.
+  const auto &untyped_pairs = std::get<0>(*args);
+  std::vector<LabelFstPair> typed_pairs;
+  typed_pairs.reserve(untyped_pairs.size());
+  for (const auto &untyped_pair : untyped_pairs) {
+    typed_pairs.emplace_back(untyped_pair.first,  // Converts label.
+                             untyped_pair.second->GetFst<Arc>());
+  }
+  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
+  const auto &opts = std::get<2>(*args);
+  ReplaceFstOptions<Arc> typed_opts(opts.root, opts.call_label_type,
+                                    opts.return_label_type, opts.return_label);
+  ReplaceFst<Arc> rfst(typed_pairs, typed_opts);
+  // Checks for cyclic dependencies before attempting expansion.
+  if (rfst.CyclicDependencies()) {
+    FSTERROR() << "Replace: Cyclic dependencies detected; cannot expand";
+    ofst->SetProperties(kError, kError);
+    return;
+  }
+  typed_opts.gc = true;     // Caching options to speed up batch copy.
+  typed_opts.gc_limit = 0;
+  *ofst = rfst;
+}
+
+void Replace(const std::vector<LabelFstClassPair> &pairs,
+             MutableFstClass *ofst, const ReplaceOptions &opts);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_REPLACE_H_
diff --git a/projects/llm_framework/include/fst/script/reverse.h b/projects/llm_framework/include/fst/script/reverse.h
new file mode 100644
index 00000000..badd96b5
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/reverse.h
@@ -0,0 +1,30 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_REVERSE_H_
+#define FST_SCRIPT_REVERSE_H_
+
+#include <tuple>
+
+#include <fst/reverse.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using ReverseArgs = std::tuple<const FstClass &, MutableFstClass *, bool>;
+
+template <class Arc>
+void Reverse(ReverseArgs *args) {
+  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
+  Reverse(ifst, ofst, std::get<2>(*args));
+}
+
+void Reverse(const FstClass &ifst, MutableFstClass *ofst,
+             bool require_superinitial = true);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_REVERSE_H_
diff --git a/projects/llm_framework/include/fst/script/reweight.h b/projects/llm_framework/include/fst/script/reweight.h
new file mode 100644
index 00000000..3893ad85
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/reweight.h
@@ -0,0 +1,37 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_REWEIGHT_H_
+#define FST_SCRIPT_REWEIGHT_H_
+
+#include <tuple>
+#include <vector>
+
+#include <fst/reweight.h>
+#include <fst/script/fst-class.h>
+#include <fst/script/script-impl.h>
+#include <fst/script/weight-class.h>
+
+namespace fst {
+namespace script {
+
+using ReweightArgs = std::tuple<MutableFstClass *,
+                                const std::vector<WeightClass> &, ReweightType>;
+
+template <class Arc>
+void Reweight(ReweightArgs *args) {
+  using Weight = typename Arc::Weight;
+  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
+  const std::vector<WeightClass> &potentials = std::get<1>(*args);
+  std::vector<Weight> typed_potentials;
+  internal::CopyWeights(potentials, &typed_potentials);
+  Reweight(fst, typed_potentials, std::get<2>(*args));
+}
+
+void Reweight(MutableFstClass *fst, const std::vector<WeightClass> &potentials,
+              ReweightType reweight_type);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_REWEIGHT_H_
diff --git a/projects/llm_framework/include/fst/script/rmepsilon.h b/projects/llm_framework/include/fst/script/rmepsilon.h
new file mode 100644
index 00000000..42986c85
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/rmepsilon.h
@@ -0,0 +1,109 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_RMEPSILON_H_
+#define FST_SCRIPT_RMEPSILON_H_
+
+#include <utility>
+#include <vector>
+
+#include <fst/queue.h>
+#include <fst/rmepsilon.h>
+#include <fst/script/fst-class.h>
+#include <fst/script/shortest-distance.h>
+#include <fst/script/weight-class.h>
+
+namespace fst {
+namespace script {
+
+struct RmEpsilonOptions : public ShortestDistanceOptions {
+  const bool connect;
+  const WeightClass &weight_threshold;
+  const int64 state_threshold;
+
+  RmEpsilonOptions(QueueType queue_type, bool connect,
+                   const WeightClass &weight_threshold,
+                   int64 state_threshold = kNoStateId, float delta = kDelta)
+      : ShortestDistanceOptions(queue_type, EPSILON_ARC_FILTER, kNoStateId,
+                                delta),
+        connect(connect),
+        weight_threshold(weight_threshold),
+        state_threshold(state_threshold) {}
+};
+
+namespace internal {
+
+// Code to implement switching on queue types.
+
+template <class Arc, class Queue>
+void RmEpsilon(MutableFst<Arc> *fst,
+               std::vector<typename Arc::Weight> *distance,
+               const RmEpsilonOptions &opts, Queue *queue) {
+  using Weight = typename Arc::Weight;
+  const fst::RmEpsilonOptions<Arc, Queue> ropts(
+      queue, opts.delta, opts.connect,
+      *opts.weight_threshold.GetWeight<Weight>(), opts.state_threshold);
+  RmEpsilon(fst, distance, ropts);
+}
+
+template <class Arc>
+void RmEpsilon(MutableFst<Arc> *fst, const RmEpsilonOptions &opts) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  std::vector<Weight> distance;
+  switch (opts.queue_type) {
+    case AUTO_QUEUE: {
+      AutoQueue<StateId> queue(*fst, &distance, EpsilonArcFilter<Arc>());
+      RmEpsilon(fst, &distance, opts, &queue);
+      return;
+    }
+    case FIFO_QUEUE: {
+      FifoQueue<StateId> queue;
+      RmEpsilon(fst, &distance, opts, &queue);
+      return;
+    }
+    case LIFO_QUEUE: {
+      LifoQueue<StateId> queue;
+      RmEpsilon(fst, &distance, opts, &queue);
+      return;
+    }
+    case SHORTEST_FIRST_QUEUE: {
+      NaturalShortestFirstQueue<StateId, Weight> queue(distance);
+      RmEpsilon(fst, &distance, opts, &queue);
+      return;
+    }
+    case STATE_ORDER_QUEUE: {
+      StateOrderQueue<StateId> queue;
+      RmEpsilon(fst, &distance, opts, &queue);
+      return;
+    }
+    case TOP_ORDER_QUEUE: {
+      TopOrderQueue<StateId> queue(*fst, EpsilonArcFilter<Arc>());
+      internal::RmEpsilon(fst, &distance, opts, &queue);
+      return;
+    }
+    default: {
+      FSTERROR() << "RmEpsilon: Unknown queue type: " << opts.queue_type;
+      fst->SetProperties(kError, kError);
+      return;
+    }
+  }
+}
+
+}  // namespace internal
+
+using RmEpsilonArgs = std::pair<MutableFstClass *, const RmEpsilonOptions &>;
+
+template <class Arc>
+void RmEpsilon(RmEpsilonArgs *args) {
+  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
+  const auto &opts = std::get<1>(*args);
+  internal::RmEpsilon(fst, opts);
+}
+
+void RmEpsilon(MutableFstClass *fst, const RmEpsilonOptions &opts);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_RMEPSILON_H_
diff --git a/projects/llm_framework/include/fst/script/script-impl.h b/projects/llm_framework/include/fst/script/script-impl.h
new file mode 100644
index 00000000..33c2853a
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/script-impl.h
@@ -0,0 +1,211 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// This file defines the registration mechanism for new operations.
+// These operations are designed to enable scripts to work with FST classes
+// at a high level.
+//
+// If you have a new arc type and want these operations to work with FSTs
+// with that arc type, see below for the registration steps
+// you must take.
+//
+// These methods are only recommended for use in high-level scripting
+// applications. Most users should use the lower-level templated versions
+// corresponding to these.
+//
+// If you have a new arc type you'd like these operations to work with,
+// use the REGISTER_FST_OPERATIONS macro defined in fstscript.h.
+//
+// If you have a custom operation you'd like to define, you need four
+// components. In the following, assume you want to create a new operation
+// with the signature
+//
+//    void Foo(const FstClass &ifst, MutableFstClass *ofst);
+//
+//  You need:
+//
+//  1) A way to bundle the args that your new Foo operation will take, as
+//     a single struct. The template structs in arg-packs.h provide a handy
+//     way to do this. In Foo's case, that might look like this:
+//
+//       using FooArgs = std::pair<const FstClass &, MutableFstClass *>;
+//
+//     Note: this package of args is going to be passed by non-const pointer.
+//
+//  2) A function template that is able to perform Foo, given the args and
+//     arc type. Yours might look like this:
+//
+//       template<class Arc>
+//       void Foo(FooArgs *args) {
+//          // Pulls out the actual, arc-templated FSTs.
+//          const Fst<Arc> &ifst = std::get<0>(*args).GetFst<Arc>();
+//          MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
+//          // Actually perform Foo on ifst and ofst.
+//       }
+//
+//  3) a client-facing function for your operation. This would look like
+//     the following:
+//
+//     void Foo(const FstClass &ifst, MutableFstClass *ofst) {
+//       // Check that the arc types of the FSTs match
+//       if (!ArcTypesMatch(ifst, *ofst, "Foo")) return;
+//       // package the args
+//       FooArgs args(ifst, ofst);
+//       // Finally, call the operation
+//       Apply<Operation<FooArgs>>("Foo", ifst->ArcType(), &args);
+//     }
+//
+//  The Apply<> function template takes care of the link between 2 and 3,
+//  provided you also have:
+//
+//  4) A registration for your new operation, on the arc types you care about.
+//     This can be provided easily by the REGISTER_FST_OPERATION macro in
+//     operations.h:
+//
+//       REGISTER_FST_OPERATION(Foo, StdArc, FooArgs);
+//       REGISTER_FST_OPERATION(Foo, MyArc, FooArgs);
+//       // .. etc
+//
+//
+//  That's it! Now when you call Foo(const FstClass &, MutableFstClass *),
+//  it dispatches (in #3) via the Apply<> function to the correct
+//  instantiation of the template function in #2.
+//
+
+#ifndef FST_SCRIPT_SCRIPT_IMPL_H_
+#define FST_SCRIPT_SCRIPT_IMPL_H_
+
+// This file contains general-purpose templates which are used in the
+// implementation of the operations.
+
+#include <string>
+#include <utility>
+
+#include <fst/generic-register.h>
+#include <fst/script/fst-class.h>
+
+#include <fst/log.h>
+
+namespace fst {
+namespace script {
+
+enum RandArcSelection {
+  UNIFORM_ARC_SELECTOR,
+  LOG_PROB_ARC_SELECTOR,
+  FAST_LOG_PROB_ARC_SELECTOR
+};
+
+// A generic register for operations with various kinds of signatures.
+// Needed since every function signature requires a new registration class.
+// The std::pair<string, string> is understood to be the operation name and arc
+// type; subclasses (or typedefs) need only provide the operation signature.
+
+template <class OperationSignature>
+class GenericOperationRegister
+    : public GenericRegister<std::pair<string, string>, OperationSignature,
+                             GenericOperationRegister<OperationSignature>> {
+ public:
+  void RegisterOperation(const string &operation_name, const string &arc_type,
+                         OperationSignature op) {
+    this->SetEntry(std::make_pair(operation_name, arc_type), op);
+  }
+
+  OperationSignature GetOperation(const string &operation_name,
+                                  const string &arc_type) {
+    return this->GetEntry(std::make_pair(operation_name, arc_type));
+  }
+
+ protected:
+  string ConvertKeyToSoFilename(
+      const std::pair<string, string> &key) const final {
+    // Uses the old-style FST for now.
+    string legal_type(key.second);  // The arc type.
+    ConvertToLegalCSymbol(&legal_type);
+    return legal_type + "-arc.so";
+  }
+};
+
+// Operation package: everything you need to register a new type of operation.
+// The ArgPack should be the type that's passed into each wrapped function;
+// for instance, it might be a struct containing all the args. It's always
+// passed by pointer, so const members should be used to enforce constness where
+// it's needed. Return values should be implemented as a member of ArgPack as
+// well.
+
+template <class Args>
+struct Operation {
+  using ArgPack = Args;
+
+  using OpType = void (*)(ArgPack *args);
+
+  // The register (hash) type.
+  using Register = GenericOperationRegister<OpType>;
+
+  // The register-er type
+  using Registerer = GenericRegisterer<Register>;
+};
+
+// Macro for registering new types of operations.
+
+#define REGISTER_FST_OPERATION(Op, Arc, ArgPack)               \
+  static fst::script::Operation<ArgPack>::Registerer       \
+      arc_dispatched_operation_##ArgPack##Op##Arc##_registerer \
+      (std::make_pair(#Op, Arc::Type()), Op<Arc>)
+
+// Template function to apply an operation by name.
+
+template <class OpReg>
+void Apply(const string &op_name, const string &arc_type,
+           typename OpReg::ArgPack *args) {
+  const auto op = OpReg::Register::GetRegister()->GetOperation(op_name,
+                                                               arc_type);
+  if (!op) {
+    FSTERROR() << "No operation found for " << op_name << " on "
+               << "arc type " << arc_type;
+    return;
+  }
+  op(args);
+}
+
+namespace internal {
+
+// Helper that logs to ERROR if the arc types of m and n don't match,
+// assuming that both m and n implement .ArcType(). The op_name argument is
+// used to construct the error message.
+template <class M, class N>
+bool ArcTypesMatch(const M &m, const N &n, const string &op_name) {
+  if (m.ArcType() != n.ArcType()) {
+    FSTERROR() << "Arguments with non-matching arc types passed to "
+               << op_name << ":\t" << m.ArcType() << " and " << n.ArcType();
+    return false;
+  }
+  return true;
+}
+
+// From untyped to typed weights.
+template <class Weight>
+void CopyWeights(const std::vector<WeightClass> &weights,
+                 std::vector<Weight> *typed_weights) {
+  typed_weights->clear();
+  typed_weights->reserve(weights.size());
+  for (const auto &weight : weights) {
+    typed_weights->push_back(*weight.GetWeight<Weight>());
+  }
+}
+
+// From typed to untyped weights.
+template <class Weight>
+void CopyWeights(const std::vector<Weight> &typed_weights,
+                 std::vector<WeightClass> *weights) {
+  weights->clear();
+  weights->reserve(typed_weights.size());
+  for (const auto &typed_weight : typed_weights) {
+    weights->emplace_back(typed_weight);
+  }
+}
+
+}  // namespace internal
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_SCRIPT_IMPL_H_
diff --git a/projects/llm_framework/include/fst/script/shortest-distance.h b/projects/llm_framework/include/fst/script/shortest-distance.h
new file mode 100644
index 00000000..a44a6c9b
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/shortest-distance.h
@@ -0,0 +1,214 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_SHORTEST_DISTANCE_H_
+#define FST_SCRIPT_SHORTEST_DISTANCE_H_
+
+#include <tuple>
+#include <vector>
+
+#include <fst/queue.h>
+#include <fst/shortest-distance.h>
+#include <fst/script/fst-class.h>
+#include <fst/script/prune.h>
+#include <fst/script/script-impl.h>
+#include <fst/script/weight-class.h>
+
+namespace fst {
+namespace script {
+
+enum ArcFilterType {
+  ANY_ARC_FILTER,
+  EPSILON_ARC_FILTER,
+  INPUT_EPSILON_ARC_FILTER,
+  OUTPUT_EPSILON_ARC_FILTER
+};
+
+struct ShortestDistanceOptions {
+  const QueueType queue_type;
+  const ArcFilterType arc_filter_type;
+  const int64 source;
+  const float delta;
+
+  ShortestDistanceOptions(QueueType queue_type, ArcFilterType arc_filter_type,
+                          int64 source, float delta)
+      : queue_type(queue_type),
+        arc_filter_type(arc_filter_type),
+        source(source),
+        delta(delta) {}
+};
+
+namespace internal {
+
+// Code to implement switching on queue and arc filter types.
+
+template <class Arc, class Queue, class ArcFilter>
+struct QueueConstructor {
+  using Weight = typename Arc::Weight;
+
+  static Queue *Construct(const Fst<Arc> &, const std::vector<Weight> *) {
+    return new Queue();
+  }
+};
+
+// Specializations to support queues with different constructors.
+
+template <class Arc, class ArcFilter>
+struct QueueConstructor<Arc, AutoQueue<typename Arc::StateId>, ArcFilter> {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  //  template<class Arc, class ArcFilter>
+  static AutoQueue<StateId> *Construct(const Fst<Arc> &fst,
+                                       const std::vector<Weight> *distance) {
+    return new AutoQueue<StateId>(fst, distance, ArcFilter());
+  }
+};
+
+template <class Arc, class ArcFilter>
+struct QueueConstructor<
+    Arc, NaturalShortestFirstQueue<typename Arc::StateId, typename Arc::Weight>,
+    ArcFilter> {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  static NaturalShortestFirstQueue<StateId, Weight> *Construct(
+      const Fst<Arc> &, const std::vector<Weight> *distance) {
+    return new NaturalShortestFirstQueue<StateId, Weight>(*distance);
+  }
+};
+
+template <class Arc, class ArcFilter>
+struct QueueConstructor<Arc, TopOrderQueue<typename Arc::StateId>, ArcFilter> {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  static TopOrderQueue<StateId> *Construct(const Fst<Arc> &fst,
+                                           const std::vector<Weight> *) {
+    return new TopOrderQueue<StateId>(fst, ArcFilter());
+  }
+};
+
+template <class Arc, class Queue, class ArcFilter>
+void ShortestDistance(const Fst<Arc> &fst,
+                      std::vector<typename Arc::Weight> *distance,
+                      const ShortestDistanceOptions &opts) {
+  std::unique_ptr<Queue> queue(
+      QueueConstructor<Arc, Queue, ArcFilter>::Construct(fst, distance));
+  const fst::ShortestDistanceOptions<Arc, Queue, ArcFilter> sopts(
+      queue.get(), ArcFilter(), opts.source, opts.delta);
+  ShortestDistance(fst, distance, sopts);
+}
+
+template <class Arc, class Queue>
+void ShortestDistance(const Fst<Arc> &fst,
+                      std::vector<typename Arc::Weight> *distance,
+                      const ShortestDistanceOptions &opts) {
+  switch (opts.arc_filter_type) {
+    case ANY_ARC_FILTER: {
+      ShortestDistance<Arc, Queue, AnyArcFilter<Arc>>(fst, distance, opts);
+      return;
+    }
+    case EPSILON_ARC_FILTER: {
+      ShortestDistance<Arc, Queue, EpsilonArcFilter<Arc>>(fst, distance, opts);
+      return;
+    }
+    case INPUT_EPSILON_ARC_FILTER: {
+      ShortestDistance<Arc, Queue, InputEpsilonArcFilter<Arc>>(fst, distance,
+                                                               opts);
+      return;
+    }
+    case OUTPUT_EPSILON_ARC_FILTER: {
+      ShortestDistance<Arc, Queue, OutputEpsilonArcFilter<Arc>>(fst, distance,
+                                                                opts);
+      return;
+    }
+    default: {
+      FSTERROR() << "ShortestDistance: Unknown arc filter type: "
+                 << opts.arc_filter_type;
+      distance->clear();
+      distance->resize(1, Arc::Weight::NoWeight());
+      return;
+    }
+  }
+}
+
+}  // namespace internal
+
+using ShortestDistanceArgs1 =
+    std::tuple<const FstClass &, std::vector<WeightClass> *,
+               const ShortestDistanceOptions &>;
+
+template <class Arc>
+void ShortestDistance(ShortestDistanceArgs1 *args) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
+  const auto &opts = std::get<2>(*args);
+  std::vector<Weight> typed_distance;
+  switch (opts.queue_type) {
+    case AUTO_QUEUE: {
+      internal::ShortestDistance<Arc, AutoQueue<StateId>>(fst, &typed_distance,
+                                                          opts);
+      break;
+    }
+    case FIFO_QUEUE: {
+      internal::ShortestDistance<Arc, FifoQueue<StateId>>(fst, &typed_distance,
+                                                          opts);
+      break;
+    }
+    case LIFO_QUEUE: {
+      internal::ShortestDistance<Arc, LifoQueue<StateId>>(fst, &typed_distance,
+                                                          opts);
+      break;
+    }
+    case SHORTEST_FIRST_QUEUE: {
+      internal::ShortestDistance<Arc,
+                                 NaturalShortestFirstQueue<StateId, Weight>>(
+          fst, &typed_distance, opts);
+      break;
+    }
+    case STATE_ORDER_QUEUE: {
+      internal::ShortestDistance<Arc, StateOrderQueue<StateId>>(
+          fst, &typed_distance, opts);
+      break;
+    }
+    case TOP_ORDER_QUEUE: {
+      internal::ShortestDistance<Arc, TopOrderQueue<StateId>>(
+          fst, &typed_distance, opts);
+      break;
+    }
+    default: {
+      FSTERROR() << "ShortestDistance: Unknown queue type: " << opts.queue_type;
+      typed_distance.clear();
+      typed_distance.resize(1, Arc::Weight::NoWeight());
+      break;
+    }
+  }
+  internal::CopyWeights(typed_distance, std::get<1>(*args));
+}
+
+using ShortestDistanceArgs2 =
+    std::tuple<const FstClass &, std::vector<WeightClass> *, bool, double>;
+
+template <class Arc>
+void ShortestDistance(ShortestDistanceArgs2 *args) {
+  using Weight = typename Arc::Weight;
+  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
+  std::vector<Weight> typed_distance;
+  ShortestDistance(fst, &typed_distance, std::get<2>(*args),
+                   std::get<3>(*args));
+  internal::CopyWeights(typed_distance, std::get<1>(*args));
+}
+
+void ShortestDistance(const FstClass &fst, std::vector<WeightClass> *distance,
+                      const ShortestDistanceOptions &opts);
+
+void ShortestDistance(const FstClass &ifst, std::vector<WeightClass> *distance,
+                      bool reverse = false,
+                      double delta = fst::kShortestDelta);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_SHORTEST_DISTANCE_H_
diff --git a/projects/llm_framework/include/fst/script/shortest-path.h b/projects/llm_framework/include/fst/script/shortest-path.h
new file mode 100644
index 00000000..86bc88da
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/shortest-path.h
@@ -0,0 +1,116 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_SHORTEST_PATH_H_
+#define FST_SCRIPT_SHORTEST_PATH_H_
+
+#include <memory>
+#include <vector>
+
+#include <fst/shortest-path.h>
+#include <fst/script/fst-class.h>
+#include <fst/script/shortest-distance.h>
+#include <fst/script/weight-class.h>
+
+namespace fst {
+namespace script {
+
+// Slightly simplified interface: `has_distance` and `first_path` are disabled.
+
+struct ShortestPathOptions : public ShortestDistanceOptions {
+  const int32 nshortest;
+  const bool unique;
+  const WeightClass &weight_threshold;
+  const int64 state_threshold;
+
+  ShortestPathOptions(QueueType queue_type, int32 nshortest, bool unique,
+                      float delta, const WeightClass &weight_threshold,
+                      int64 state_threshold = kNoStateId)
+      : ShortestDistanceOptions(queue_type, ANY_ARC_FILTER, kNoStateId, delta),
+        nshortest(nshortest),
+        unique(unique),
+        weight_threshold(weight_threshold),
+        state_threshold(state_threshold) {}
+};
+
+namespace internal {
+
+// Code to implement switching on queue types.
+
+template <class Arc, class Queue>
+void ShortestPath(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
+                  std::vector<typename Arc::Weight> *distance,
+                  const ShortestPathOptions &opts) {
+  using ArcFilter = AnyArcFilter<Arc>;
+  using Weight = typename Arc::Weight;
+  const std::unique_ptr<Queue> queue(
+      QueueConstructor<Arc, Queue, ArcFilter>::Construct(ifst, distance));
+  const fst::ShortestPathOptions<Arc, Queue, ArcFilter> sopts(
+      queue.get(), ArcFilter(), opts.nshortest, opts.unique,
+      /* has_distance=*/false, opts.delta, /* first_path=*/false,
+      *opts.weight_threshold.GetWeight<Weight>(), opts.state_threshold);
+  ShortestPath(ifst, ofst, distance, sopts);
+}
+
+template <class Arc>
+void ShortestPath(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
+                  const ShortestPathOptions &opts) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  std::vector<Weight> distance;
+  switch (opts.queue_type) {
+    case AUTO_QUEUE: {
+      ShortestPath<Arc, AutoQueue<StateId>>(ifst, ofst, &distance, opts);
+      return;
+    }
+    case FIFO_QUEUE: {
+      ShortestPath<Arc, FifoQueue<StateId>>(ifst, ofst, &distance, opts);
+      return;
+    }
+    case LIFO_QUEUE: {
+      ShortestPath<Arc, LifoQueue<StateId>>(ifst, ofst, &distance, opts);
+      return;
+    }
+    case SHORTEST_FIRST_QUEUE: {
+      ShortestPath<Arc, NaturalShortestFirstQueue<StateId, Weight>>(ifst, ofst,
+                                                                    &distance,
+                                                                    opts);
+      return;
+    }
+    case STATE_ORDER_QUEUE: {
+      ShortestPath<Arc, StateOrderQueue<StateId>>(ifst, ofst, &distance, opts);
+      return;
+    }
+    case TOP_ORDER_QUEUE: {
+      ShortestPath<Arc, TopOrderQueue<StateId>>(ifst, ofst, &distance, opts);
+      return;
+    }
+    default: {
+      FSTERROR() << "ShortestPath: Unknown queue type: "
+                 << opts.queue_type;
+      ofst->SetProperties(kError, kError);
+      return;
+    }
+  }
+}
+
+}  // namespace internal
+
+using ShortestPathArgs = std::tuple<const FstClass &, MutableFstClass *,
+                                    const ShortestPathOptions &>;
+
+template <class Arc>
+void ShortestPath(ShortestPathArgs *args) {
+  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
+  const ShortestPathOptions &opts = std::get<2>(*args);
+  internal::ShortestPath(ifst, ofst, opts);
+}
+
+void ShortestPath(const FstClass &ifst, MutableFstClass *ofst,
+                  const ShortestPathOptions &opts);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_SHORTEST_PATH_H_
diff --git a/projects/llm_framework/include/fst/script/stateiterator-class.h b/projects/llm_framework/include/fst/script/stateiterator-class.h
new file mode 100644
index 00000000..f6fddfe6
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/stateiterator-class.h
@@ -0,0 +1,85 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_STATEITERATOR_CLASS_H_
+#define FST_SCRIPT_STATEITERATOR_CLASS_H_
+
+#include <memory>
+
+#include <fst/fstlib.h>
+#include <fst/script/fst-class.h>
+
+// Scripting API support for StateIterator.
+
+namespace fst {
+namespace script {
+
+// Virtual interface implemented by each concrete StateIteratorImpl<F>.
+class StateIteratorImplBase {
+ public:
+  virtual bool Done() const = 0;
+  virtual int64 Value() const = 0;
+  virtual void Next() = 0;
+  virtual void Reset() = 0;
+  virtual ~StateIteratorImplBase() {}
+};
+
+// Templated implementation.
+template <class Arc>
+class StateIteratorClassImpl : public StateIteratorImplBase {
+ public:
+  explicit StateIteratorClassImpl(const Fst<Arc> &fst) : siter_(fst) {}
+
+  bool Done() const final { return siter_.Done(); }
+
+  int64 Value() const final { return siter_.Value(); }
+
+  void Next() final { siter_.Next(); }
+
+  void Reset() final { siter_.Reset(); }
+
+  ~StateIteratorClassImpl() override {}
+
+ private:
+  StateIterator<Fst<Arc>> siter_;
+};
+
+class StateIteratorClass;
+
+using InitStateIteratorClassArgs =
+    std::pair<const FstClass &, StateIteratorClass *>;
+
+// Untemplated user-facing class holding a templated pimpl.
+class StateIteratorClass {
+ public:
+  explicit StateIteratorClass(const FstClass &fst);
+
+  template <class Arc>
+  explicit StateIteratorClass(const Fst<Arc> &fst)
+      : impl_(new StateIteratorClassImpl<Arc>(fst)) {}
+
+  bool Done() const { return impl_->Done(); }
+
+  int64 Value() const { return impl_->Value(); }
+
+  void Next() { impl_->Next(); }
+
+  void Reset() { impl_->Reset(); }
+
+  template <class Arc>
+  friend void InitStateIteratorClass(InitStateIteratorClassArgs *args);
+
+ private:
+  std::unique_ptr<StateIteratorImplBase> impl_;
+};
+
+template <class Arc>
+void InitStateIteratorClass(InitStateIteratorClassArgs *args) {
+  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
+  std::get<1>(*args)->impl_.reset(new StateIteratorClassImpl<Arc>(fst));
+}
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_STATEITERATOR_CLASS_H_
diff --git a/projects/llm_framework/include/fst/script/synchronize.h b/projects/llm_framework/include/fst/script/synchronize.h
new file mode 100644
index 00000000..01df151a
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/synchronize.h
@@ -0,0 +1,29 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_SYNCHRONIZE_H_
+#define FST_SCRIPT_SYNCHRONIZE_H_
+
+#include <utility>
+
+#include <fst/synchronize.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using SynchronizeArgs = std::pair<const FstClass &, MutableFstClass *>;
+
+template <class Arc>
+void Synchronize(SynchronizeArgs *args) {
+  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
+  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
+  Synchronize(ifst, ofst);
+}
+
+void Synchronize(const FstClass &ifst, MutableFstClass *ofst);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_SYNCHRONIZE_H_
diff --git a/projects/llm_framework/include/fst/script/text-io.h b/projects/llm_framework/include/fst/script/text-io.h
new file mode 100644
index 00000000..464bf885
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/text-io.h
@@ -0,0 +1,28 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Utilities for reading and writing textual strings representing states,
+// labels, and weights and files specifying label-label pairs and potentials
+// (state-weight pairs).
+
+#ifndef FST_SCRIPT_TEXT_IO_H__
+#define FST_SCRIPT_TEXT_IO_H__
+
+#include <string>
+#include <vector>
+
+#include <fst/script/weight-class.h>
+
+namespace fst {
+namespace script {
+
+bool ReadPotentials(const string &weight_type, const string &filename,
+                    std::vector<WeightClass> *potentials);
+
+bool WritePotentials(const string &filename,
+                     const std::vector<WeightClass> &potentials);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_TEXT_IO_H__
diff --git a/projects/llm_framework/include/fst/script/topsort.h b/projects/llm_framework/include/fst/script/topsort.h
new file mode 100644
index 00000000..fb6738d7
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/topsort.h
@@ -0,0 +1,26 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_TOPSORT_H_
+#define FST_SCRIPT_TOPSORT_H_
+
+#include <fst/topsort.h>
+#include <fst/script/arg-packs.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using TopSortArgs = WithReturnValue<bool, MutableFstClass *>;
+
+template <class Arc>
+void TopSort(TopSortArgs *args) {
+  args->retval = TopSort(args->args->GetMutableFst<Arc>());
+}
+
+bool TopSort(MutableFstClass *fst);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_TOPSORT_H_
diff --git a/projects/llm_framework/include/fst/script/union.h b/projects/llm_framework/include/fst/script/union.h
new file mode 100644
index 00000000..9493e2b1
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/union.h
@@ -0,0 +1,29 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_UNION_H_
+#define FST_SCRIPT_UNION_H_
+
+#include <utility>
+
+#include <fst/union.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using UnionArgs = std::pair<MutableFstClass *, const FstClass &>;
+
+template <class Arc>
+void Union(UnionArgs *args) {
+  MutableFst<Arc> *fst1 = std::get<0>(*args)->GetMutableFst<Arc>();
+  const Fst<Arc> &fst2 = *(std::get<1>(*args).GetFst<Arc>());
+  Union(fst1, fst2);
+}
+
+void Union(MutableFstClass *fst1, const FstClass &fst2);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_UNION_H_
diff --git a/projects/llm_framework/include/fst/script/verify.h b/projects/llm_framework/include/fst/script/verify.h
new file mode 100644
index 00000000..52f58641
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/verify.h
@@ -0,0 +1,27 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SCRIPT_VERIFY_H_
+#define FST_SCRIPT_VERIFY_H_
+
+#include <fst/verify.h>
+#include <fst/script/arg-packs.h>
+#include <fst/script/fst-class.h>
+
+namespace fst {
+namespace script {
+
+using VerifyArgs = WithReturnValue<bool, const FstClass &>;
+
+template <class Arc>
+void Verify(VerifyArgs *args) {
+  const Fst<Arc> &fst = *(args->args.GetFst<Arc>());
+  args->retval = Verify(fst);
+}
+
+bool Verify(const FstClass &fst);
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_VERIFY_H_
diff --git a/projects/llm_framework/include/fst/script/weight-class.h b/projects/llm_framework/include/fst/script/weight-class.h
new file mode 100644
index 00000000..6dadf92c
--- /dev/null
+++ b/projects/llm_framework/include/fst/script/weight-class.h
@@ -0,0 +1,235 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Represents a generic weight in an FST; that is, represents a specific type
+// of weight underneath while hiding that type from a client.
+
+#ifndef FST_SCRIPT_WEIGHT_CLASS_H_
+#define FST_SCRIPT_WEIGHT_CLASS_H_
+
+#include <memory>
+#include <ostream>
+#include <string>
+
+#include <fst/arc.h>
+#include <fst/generic-register.h>
+#include <fst/util.h>
+#include <fst/weight.h>
+
+namespace fst {
+namespace script {
+
+class WeightImplBase {
+ public:
+  virtual WeightImplBase *Copy() const = 0;
+  virtual void Print(std::ostream *o) const = 0;
+  virtual const string &Type() const = 0;
+  virtual string ToString() const = 0;
+  virtual bool Member() const = 0;
+  virtual bool operator==(const WeightImplBase &other) const = 0;
+  virtual bool operator!=(const WeightImplBase &other) const = 0;
+  virtual WeightImplBase &PlusEq(const WeightImplBase &other) = 0;
+  virtual WeightImplBase &TimesEq(const WeightImplBase &other) = 0;
+  virtual WeightImplBase &DivideEq(const WeightImplBase &other) = 0;
+  virtual WeightImplBase &PowerEq(size_t n) = 0;
+  virtual ~WeightImplBase() {}
+};
+
+template <class W>
+class WeightClassImpl : public WeightImplBase {
+ public:
+  explicit WeightClassImpl(const W &weight) : weight_(weight) {}
+
+  WeightClassImpl<W> *Copy() const final {
+    return new WeightClassImpl<W>(weight_);
+  }
+
+  const string &Type() const final { return W::Type(); }
+
+  void Print(std::ostream *ostrm) const final { *ostrm << weight_; }
+
+  string ToString() const final {
+    string str;
+    WeightToStr(weight_, &str);
+    return str;
+  }
+
+  bool Member() const final { return weight_.Member(); }
+
+  bool operator==(const WeightImplBase &other) const final {
+    const auto *typed_other = static_cast<const WeightClassImpl<W> *>(&other);
+    return weight_ == typed_other->weight_;
+  }
+
+  bool operator!=(const WeightImplBase &other) const final {
+    return !(*this == other);
+  }
+
+  WeightClassImpl<W> &PlusEq(const WeightImplBase &other) final {
+    const auto *typed_other = static_cast<const WeightClassImpl<W> *>(&other);
+    weight_ = Plus(weight_, typed_other->weight_);
+    return *this;
+  }
+
+  WeightClassImpl<W> &TimesEq(const WeightImplBase &other) final {
+    const auto *typed_other = static_cast<const WeightClassImpl<W> *>(&other);
+    weight_ = Times(weight_, typed_other->weight_);
+    return *this;
+  }
+
+  WeightClassImpl<W> &DivideEq(const WeightImplBase &other) final {
+    const auto *typed_other = static_cast<const WeightClassImpl<W> *>(&other);
+    weight_ = Divide(weight_, typed_other->weight_);
+    return *this;
+  }
+
+  WeightClassImpl<W> &PowerEq(size_t n) final {
+    weight_ = Power(weight_, n);
+    return *this;
+  }
+
+  W *GetImpl() { return &weight_; }
+
+ private:
+  W weight_;
+};
+
+
+class WeightClass {
+ public:
+  WeightClass() = default;
+
+  template <class W>
+  explicit WeightClass(const W &weight)
+      : impl_(new WeightClassImpl<W>(weight)) {}
+
+  template <class W>
+  explicit WeightClass(const WeightClassImpl<W> &impl)
+      : impl_(new WeightClassImpl<W>(impl)) {}
+
+  WeightClass(const string &weight_type, const string &weight_str);
+
+  WeightClass(const WeightClass &other)
+      : impl_(other.impl_ ? other.impl_->Copy() : nullptr) {}
+
+  WeightClass &operator=(const WeightClass &other) {
+    impl_.reset(other.impl_ ? other.impl_->Copy() : nullptr);
+    return *this;
+  }
+
+  static constexpr const char *__ZERO__ = "__ZERO__";  // NOLINT
+
+  static WeightClass Zero(const string &weight_type);
+
+  static constexpr const char *__ONE__ = "__ONE__";  // NOLINT
+
+  static WeightClass One(const string &weight_type);
+
+  static constexpr const char *__NOWEIGHT__ = "__NOWEIGHT__";  // NOLINT
+
+  static WeightClass NoWeight(const string &weight_type);
+
+  template <class W>
+  const W *GetWeight() const {
+    if (W::Type() != impl_->Type()) {
+       return nullptr;
+    } else {
+      auto *typed_impl = static_cast<WeightClassImpl<W> *>(impl_.get());
+      return typed_impl->GetImpl();
+    }
+  }
+
+  string ToString() const { return (impl_) ? impl_->ToString() : "none"; }
+
+  const string &Type() const {
+    if (impl_) return impl_->Type();
+    static const string *const no_type = new string("none");
+    return *no_type;
+  }
+
+  bool Member() const { return impl_ && impl_->Member(); }
+
+  bool WeightTypesMatch(const WeightClass &other, const string &op_name) const;
+
+  friend bool operator==(const WeightClass &lhs, const WeightClass &rhs);
+
+  friend WeightClass Plus(const WeightClass &lhs, const WeightClass &rhs);
+
+  friend WeightClass Times(const WeightClass &lhs, const WeightClass &rhs);
+
+  friend WeightClass Divide(const WeightClass &lhs, const WeightClass &rhs);
+
+  friend WeightClass Power(const WeightClass &w, size_t n);
+
+ private:
+  const WeightImplBase *GetImpl() const { return impl_.get(); }
+
+  WeightImplBase *GetImpl() { return impl_.get(); }
+
+  std::unique_ptr<WeightImplBase> impl_;
+
+  friend std::ostream &operator<<(std::ostream &o, const WeightClass &c);
+};
+
+bool operator==(const WeightClass &lhs, const WeightClass &rhs);
+
+bool operator!=(const WeightClass &lhs, const WeightClass &rhs);
+
+WeightClass Plus(const WeightClass &lhs, const WeightClass &rhs);
+
+WeightClass Times(const WeightClass &lhs, const WeightClass &rhs);
+
+WeightClass Divide(const WeightClass &lhs, const WeightClass &rhs);
+
+WeightClass Power(const WeightClass &w, size_t n);
+
+std::ostream &operator<<(std::ostream &o, const WeightClass &c);
+
+// Registration for generic weight types.
+
+using StrToWeightImplBaseT = WeightImplBase *(*)(const string &str,
+                                                 const string &src,
+                                                 size_t nline);
+
+template <class W>
+WeightImplBase *StrToWeightImplBase(const string &str, const string &src,
+                                    size_t nline) {
+  if (str == WeightClass::__ZERO__)
+    return new WeightClassImpl<W>(W::Zero());
+  else if (str == WeightClass::__ONE__)
+    return new WeightClassImpl<W>(W::One());
+  else if (str == WeightClass::__NOWEIGHT__)
+    return new WeightClassImpl<W>(W::NoWeight());
+  return new WeightClassImpl<W>(StrToWeight<W>(str, src, nline));
+}
+
+class WeightClassRegister : public GenericRegister<string, StrToWeightImplBaseT,
+                                                   WeightClassRegister> {
+ protected:
+  string ConvertKeyToSoFilename(const string &key) const final {
+    string legal_type(key);
+    ConvertToLegalCSymbol(&legal_type);
+    return legal_type + ".so";
+  }
+};
+
+using WeightClassRegisterer = GenericRegisterer<WeightClassRegister>;
+
+// Internal version; needs to be called by wrapper in order for macro args to
+// expand.
+#define REGISTER_FST_WEIGHT__(Weight, line)                \
+  static WeightClassRegisterer weight_registerer##_##line( \
+      Weight::Type(), StrToWeightImplBase<Weight>)
+
+// This layer is where __FILE__ and __LINE__ are expanded.
+#define REGISTER_FST_WEIGHT_EXPANDER(Weight, line) \
+  REGISTER_FST_WEIGHT__(Weight, line)
+
+// Macro for registering new weight types; clients call this.
+#define REGISTER_FST_WEIGHT(Weight) \
+  REGISTER_FST_WEIGHT_EXPANDER(Weight, __LINE__)
+
+}  // namespace script
+}  // namespace fst
+
+#endif  // FST_SCRIPT_WEIGHT_CLASS_H_
diff --git a/projects/llm_framework/include/fst/set-weight.h b/projects/llm_framework/include/fst/set-weight.h
new file mode 100644
index 00000000..dd665f0d
--- /dev/null
+++ b/projects/llm_framework/include/fst/set-weight.h
@@ -0,0 +1,618 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Weights consisting of sets (of integral Labels) and
+// associated semiring operation definitions using intersect
+// and union.
+
+#ifndef FST_SET_WEIGHT_H_
+#define FST_SET_WEIGHT_H_
+
+#include <cstdlib>
+
+#include <algorithm>
+#include <list>
+#include <string>
+#include <vector>
+
+#include <fst/union-weight.h>
+#include <fst/weight.h>
+
+
+namespace fst {
+
+constexpr int kSetEmpty = 0;         // Label for the empty set.
+constexpr int kSetUniv = -1;         // Label for the universal set.
+constexpr int kSetBad = -2;          // Label for a non-set.
+constexpr char kSetSeparator = '_';  // Label separator in sets.
+
+// Determines whether to use (intersect, union) or (union, intersect)
+// as (+, *) for the semiring. SET_INTERSECT_UNION_RESTRICTED is a
+// restricted version of (intersect, union) that requires summed
+// arguments to be equal (or an error is signalled), useful for
+// algorithms that require a unique labelled path weight.  SET_BOOLEAN
+// treats all non-Zero() elements as equivalent (with Zero() ==
+// UnivSet()), useful for algorithms that don't really depend on the
+// detailed sets.
+enum SetType { SET_INTERSECT_UNION = 0,
+               SET_UNION_INTERSECT = 1,
+               SET_INTERSECT_UNION_RESTRICT = 2,
+               SET_BOOLEAN = 3 };
+
+template <class>
+class SetWeightIterator;
+
+// Set semiring of integral labels.
+template <typename Label_, SetType S = SET_INTERSECT_UNION>
+class SetWeight {
+ public:
+  using Label = Label_;
+  using ReverseWeight = SetWeight<Label, S>;
+  using Iterator = SetWeightIterator<SetWeight>;
+  friend class SetWeightIterator<SetWeight>;
+  // Allow type-converting copy and move constructors private access.
+  template <typename L2, SetType S2>
+  friend class SetWeight;
+
+  SetWeight() {}
+
+  // Input should be positive, sorted and unique.
+  template <typename Iterator>
+  SetWeight(const Iterator &begin, const Iterator &end) {
+    for (auto iter = begin; iter != end; ++iter) PushBack(*iter);
+  }
+
+  // Input should be positive. (Non-positive value has
+  // special internal meaning w.r.t. integral constants above.)
+  explicit SetWeight(Label label) { PushBack(label); }
+
+  template <SetType S2>
+  explicit SetWeight(const SetWeight<Label, S2> &w)
+    : first_(w.first_), rest_(w.rest_) {}
+
+  template <SetType S2>
+  explicit SetWeight(SetWeight<Label, S2> &&w)
+    : first_(w.first_), rest_(std::move(w.rest_)) { w.Clear(); }
+
+  template <SetType S2>
+  SetWeight &operator=(const SetWeight<Label, S2> &w) {
+    first_ = w.first_;
+    rest_ = w.rest_;
+    return *this;
+  }
+
+  template <SetType S2>
+  SetWeight &operator=(SetWeight<Label, S2> &&w) {
+    first_ = w.first_;
+    rest_ = std::move(w.rest_);
+    w.Clear();
+    return *this;
+  }
+
+  static const SetWeight &Zero() {
+    return S == SET_UNION_INTERSECT ? EmptySet() : UnivSet();
+  }
+
+  static const SetWeight &One() {
+    return S == SET_UNION_INTERSECT ? UnivSet() : EmptySet();
+  }
+
+  static const SetWeight &NoWeight() {
+    static const auto *const no_weight = new SetWeight(Label(kSetBad));
+    return *no_weight;
+  }
+
+  static const string &Type() {
+    static const string *const type = new string(
+        S == SET_UNION_INTERSECT
+        ? "union_intersect_set"
+        : (S == SET_INTERSECT_UNION
+           ? "intersect_union_set"
+           : (S == SET_INTERSECT_UNION_RESTRICT
+              ? "restricted_set_intersect_union"
+              : "boolean_set")));
+    return *type;
+  }
+
+  bool Member() const;
+
+  std::istream &Read(std::istream &strm);
+
+  std::ostream &Write(std::ostream &strm) const;
+
+  size_t Hash() const;
+
+  SetWeight Quantize(float delta = kDelta) const { return *this; }
+
+  ReverseWeight Reverse() const;
+
+  static constexpr uint64 Properties() {
+    return kIdempotent | kLeftSemiring | kRightSemiring | kCommutative;
+  }
+
+  // These operations combined with the SetWeightIterator
+  // provide the access and mutation of the set internal elements.
+
+  // The empty set.
+  static const SetWeight &EmptySet() {
+    static const auto *const empty = new SetWeight(Label(kSetEmpty));
+    return *empty;
+  }
+
+  // The univeral set.
+  static const SetWeight &UnivSet() {
+    static const auto *const univ = new SetWeight(Label(kSetUniv));
+    return *univ;
+  }
+
+  // Clear existing SetWeight.
+  void Clear() {
+    first_ = kSetEmpty;
+    rest_.clear();
+  }
+
+  size_t Size() const { return first_ == kSetEmpty ? 0 : rest_.size() + 1; }
+
+  Label Back() {
+    if (rest_.empty()) {
+      return first_;
+    } else {
+      return rest_.back();
+    }
+  }
+
+  // Caller must add in sort order and be unique (or error signalled).
+  // Input should also be positive. Non-positive value (for the first
+  // push) has special internal meaning w.r.t. integral constants above.
+  void PushBack(Label label) {
+    if (first_ == kSetEmpty) {
+      first_ = label;
+    } else {
+      if (label <= Back() || label <= 0) {
+        FSTERROR() << "SetWeight: labels must be positive, added"
+                   << " in sort order and be unique.";
+        rest_.push_back(Label(kSetBad));
+      }
+      rest_.push_back(label);
+    }
+  }
+
+ private:
+  Label first_ = kSetEmpty;  // First label in set (kSetEmpty if empty).
+  std::list<Label> rest_;    // Remaining labels in set.
+};
+
+// Traverses set in forward direction.
+template <class SetWeight_>
+class SetWeightIterator {
+ public:
+  using Weight = SetWeight_;
+  using Label = typename Weight::Label;
+
+  explicit SetWeightIterator(const Weight &w)
+      : first_(w.first_), rest_(w.rest_), init_(true), iter_(rest_.begin()) {}
+
+  bool Done() const {
+    if (init_) {
+      return first_ == kSetEmpty;
+    } else {
+      return iter_ == rest_.end();
+    }
+  }
+
+  const Label &Value() const { return init_ ? first_ : *iter_; }
+
+  void Next() {
+    if (init_) {
+      init_ = false;
+    } else {
+      ++iter_;
+    }
+  }
+
+  void Reset() {
+    init_ = true;
+    iter_ = rest_.begin();
+  }
+
+ private:
+  const Label &first_;
+  const decltype(Weight::rest_) &rest_;
+  bool init_;  // In the initialized state?
+  typename std::remove_reference<decltype(Weight::rest_)>::type::const_iterator iter_;
+};
+
+
+// SetWeight member functions follow that require SetWeightIterator
+
+template <typename Label, SetType S>
+inline std::istream &SetWeight<Label, S>::Read(std::istream &strm) {
+  Clear();
+  int32 size;
+  ReadType(strm, &size);
+  for (int32 i = 0; i < size; ++i) {
+    Label label;
+    ReadType(strm, &label);
+    PushBack(label);
+  }
+  return strm;
+}
+
+template <typename Label, SetType S>
+inline std::ostream &SetWeight<Label, S>::Write(std::ostream &strm) const {
+  const int32 size = Size();
+  WriteType(strm, size);
+  for (Iterator iter(*this); !iter.Done(); iter.Next()) {
+    WriteType(strm, iter.Value());
+  }
+  return strm;
+}
+
+template <typename Label, SetType S>
+inline bool SetWeight<Label, S>::Member() const {
+  Iterator iter(*this);
+  return iter.Value() != Label(kSetBad);
+}
+
+template <typename Label, SetType S>
+inline typename SetWeight<Label, S>::ReverseWeight
+SetWeight<Label, S>::Reverse() const {
+  return *this;
+}
+
+template <typename Label, SetType S>
+inline size_t SetWeight<Label, S>::Hash() const {
+  using Weight = SetWeight<Label, S>;
+  if (S == SET_BOOLEAN) {
+    return *this == Weight::Zero() ? 0 : 1;
+  } else {
+    size_t h = 0;
+    for (Iterator iter(*this); !iter.Done(); iter.Next()) {
+      h ^= h << 1 ^ iter.Value();
+    }
+    return h;
+  }
+}
+
+// Default ==
+template <typename Label, SetType S>
+inline bool operator==(const SetWeight<Label, S> &w1,
+                       const SetWeight<Label, S> &w2) {
+  if (w1.Size() != w2.Size()) return false;
+  using Iterator = typename SetWeight<Label, S>::Iterator;
+  Iterator iter1(w1);
+  Iterator iter2(w2);
+  for (; !iter1.Done(); iter1.Next(), iter2.Next()) {
+    if (iter1.Value() != iter2.Value()) return false;
+  }
+  return true;
+}
+
+// Boolean ==
+template <typename Label>
+inline bool operator==(const SetWeight<Label, SET_BOOLEAN> &w1,
+                       const SetWeight<Label, SET_BOOLEAN> &w2) {
+  // x == kSetEmpty if x \nin {kUnivSet, kSetBad}
+  if (!w1.Member() || !w2.Member()) return false;
+  using Iterator = typename SetWeight<Label, SET_BOOLEAN>::Iterator;
+  Iterator iter1(w1);
+  Iterator iter2(w2);
+  Label label1 = iter1.Done() ? kSetEmpty : iter1.Value();
+  Label label2 = iter2.Done() ? kSetEmpty : iter2.Value();
+  if (label1 == kSetUniv) return label2 == kSetUniv;
+  if (label2 == kSetUniv) return label1 == kSetUniv;
+  return true;
+}
+
+template <typename Label, SetType S>
+inline bool operator!=(const SetWeight<Label, S> &w1,
+                       const SetWeight<Label, S> &w2) {
+  return !(w1 == w2);
+}
+
+template <typename Label, SetType S>
+inline bool ApproxEqual(const SetWeight<Label, S> &w1,
+                        const SetWeight<Label, S> &w2,
+                        float delta = kDelta) {
+  return w1 == w2;
+}
+
+template <typename Label, SetType S>
+inline std::ostream &operator<<(std::ostream &strm,
+                                const SetWeight<Label, S> &weight) {
+  typename SetWeight<Label, S>::Iterator iter(weight);
+  if (iter.Done()) {
+    return strm << "EmptySet";
+  } else if (iter.Value() == Label(kSetUniv)) {
+    return strm << "UnivSet";
+  } else if (iter.Value() == Label(kSetBad)) {
+    return strm << "BadSet";
+  } else {
+    for (size_t i = 0; !iter.Done(); ++i, iter.Next()) {
+      if (i > 0) strm << kSetSeparator;
+      strm << iter.Value();
+    }
+  }
+  return strm;
+}
+
+template <typename Label, SetType S>
+inline std::istream &operator>>(std::istream &strm,
+                                SetWeight<Label, S> &weight) {
+  string str;
+  strm >> str;
+  using Weight = SetWeight<Label, S>;
+  if (str == "EmptySet") {
+    weight = Weight(Label(kSetEmpty));
+  } else if (str == "UnivSet") {
+    weight = Weight(Label(kSetUniv));
+  } else {
+    weight.Clear();
+    char *p = nullptr;
+    for (const char *cs = str.c_str(); !p || *p != '\0'; cs = p + 1) {
+      const Label label = strtoll(cs, &p, 10);
+      if (p == cs || (*p != 0 && *p != kSetSeparator)) {
+        strm.clear(std::ios::badbit);
+        break;
+      }
+      weight.PushBack(label);
+    }
+  }
+  return strm;
+}
+
+template <typename Label, SetType S>
+inline SetWeight<Label, S> Union(
+    const SetWeight<Label, S> &w1,
+    const SetWeight<Label, S> &w2) {
+  using Weight = SetWeight<Label, S>;
+  using Iterator = typename SetWeight<Label, S>::Iterator;
+  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
+  if (w1 == Weight::EmptySet()) return w2;
+  if (w2 == Weight::EmptySet()) return w1;
+  if (w1 == Weight::UnivSet()) return w1;
+  if (w2 == Weight::UnivSet()) return w2;
+  Iterator it1(w1);
+  Iterator it2(w2);
+  Weight result;
+  while (!it1.Done() && !it2.Done()) {
+    const auto v1 = it1.Value();
+    const auto v2 = it2.Value();
+    if (v1 < v2) {
+      result.PushBack(v1);
+      it1.Next();
+    } else if (v1 > v2) {
+      result.PushBack(v2);
+      it2.Next();
+    } else {
+      result.PushBack(v1);
+      it1.Next();
+      it2.Next();
+    }
+  }
+  for (; !it1.Done(); it1.Next()) result.PushBack(it1.Value());
+  for (; !it2.Done(); it2.Next()) result.PushBack(it2.Value());
+  return result;
+}
+
+template <typename Label, SetType S>
+inline SetWeight<Label, S> Intersect(
+    const SetWeight<Label, S> &w1,
+    const SetWeight<Label, S> &w2) {
+  using Weight = SetWeight<Label, S>;
+  using Iterator = typename SetWeight<Label, S>::Iterator;
+  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
+  if (w1 == Weight::EmptySet()) return w1;
+  if (w2 == Weight::EmptySet()) return w2;
+  if (w1 == Weight::UnivSet()) return w2;
+  if (w2 == Weight::UnivSet()) return w1;
+  Iterator it1(w1);
+  Iterator it2(w2);
+  Weight result;
+  while (!it1.Done() && !it2.Done()) {
+    const auto v1 = it1.Value();
+    const auto v2 = it2.Value();
+    if (v1 < v2) {
+      it1.Next();
+    } else if (v1 > v2) {
+      it2.Next();
+    } else {
+      result.PushBack(v1);
+      it1.Next();
+      it2.Next();
+    }
+  }
+  return result;
+}
+
+template <typename Label, SetType S>
+inline SetWeight<Label, S> Difference(
+    const SetWeight<Label, S> &w1,
+    const SetWeight<Label, S> &w2) {
+  using Weight = SetWeight<Label, S>;
+  using Iterator = typename SetWeight<Label, S>::Iterator;
+  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
+  if (w1 == Weight::EmptySet()) return w1;
+  if (w2 == Weight::EmptySet()) return w1;
+  if (w2 == Weight::UnivSet()) return Weight::EmptySet();
+  Iterator it1(w1);
+  Iterator it2(w2);
+  Weight result;
+  while (!it1.Done() && !it2.Done()) {
+    const auto v1 = it1.Value();
+    const auto v2 = it2.Value();
+    if (v1 < v2) {
+      result.PushBack(v1);
+      it1.Next();
+    } else if (v1 > v2) {
+      it2.Next();
+    } else {
+      it1.Next();
+      it2.Next();
+    }
+  }
+  for (; !it1.Done(); it1.Next()) result.PushBack(it1.Value());
+  return result;
+}
+
+// Default: Plus = Intersect.
+template <typename Label, SetType S>
+inline SetWeight<Label, S> Plus(
+    const SetWeight<Label, S> &w1,
+    const SetWeight<Label, S> &w2) {
+  return Intersect(w1, w2);
+}
+
+// Plus = Union.
+template <typename Label>
+inline SetWeight<Label, SET_UNION_INTERSECT> Plus(
+    const SetWeight<Label, SET_UNION_INTERSECT> &w1,
+    const SetWeight<Label, SET_UNION_INTERSECT> &w2) {
+  return Union(w1, w2);
+}
+
+// Plus = Set equality is required (for non-Zero() input). The
+// restriction is useful (e.g., in determinization) to ensure the input
+// has a unique labelled path weight.
+template <typename Label>
+inline SetWeight<Label, SET_INTERSECT_UNION_RESTRICT> Plus(
+    const SetWeight<Label, SET_INTERSECT_UNION_RESTRICT> &w1,
+    const SetWeight<Label, SET_INTERSECT_UNION_RESTRICT> &w2) {
+  using Weight = SetWeight<Label, SET_INTERSECT_UNION_RESTRICT>;
+  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
+  if (w1 == Weight::Zero()) return w2;
+  if (w2 == Weight::Zero()) return w1;
+  if (w1 != w2) {
+    FSTERROR() << "SetWeight::Plus: Unequal arguments "
+               << "(non-unique labelled path weights?)"
+               << " w1 = " << w1 << " w2 = " << w2;
+    return Weight::NoWeight();
+  }
+  return w1;
+}
+
+// Plus = Or.
+template <typename Label>
+inline SetWeight<Label, SET_BOOLEAN> Plus(
+    const SetWeight<Label, SET_BOOLEAN> &w1,
+    const SetWeight<Label, SET_BOOLEAN> &w2) {
+  using Weight = SetWeight<Label, SET_BOOLEAN>;
+  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
+  if (w1 == Weight::One()) return w1;
+  if (w2 == Weight::One()) return w2;
+  return Weight::Zero();
+}
+
+// Default: Times = Union.
+template <typename Label, SetType S>
+inline SetWeight<Label, S> Times(
+    const SetWeight<Label, S> &w1,
+    const SetWeight<Label, S> &w2) {
+  return Union(w1, w2);
+}
+
+// Times = Intersect.
+template <typename Label>
+inline SetWeight<Label, SET_UNION_INTERSECT> Times(
+    const SetWeight<Label, SET_UNION_INTERSECT> &w1,
+    const SetWeight<Label, SET_UNION_INTERSECT> &w2) {
+  return Intersect(w1, w2);
+}
+
+// Times = And.
+template <typename Label>
+inline SetWeight<Label, SET_BOOLEAN> Times(
+    const SetWeight<Label, SET_BOOLEAN> &w1,
+    const SetWeight<Label, SET_BOOLEAN> &w2) {
+  using Weight = SetWeight<Label, SET_BOOLEAN>;
+  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
+  if (w1 == Weight::One()) return w2;
+  return w1;
+}
+
+// Divide = Difference.
+template <typename Label, SetType S>
+inline SetWeight<Label, S> Divide(const SetWeight<Label, S> &w1,
+                                  const SetWeight<Label, S> &w2,
+                                  DivideType divide_type = DIVIDE_ANY) {
+  return Difference(w1, w2);
+}
+
+// Divide = dividend (or the universal set if the
+// dividend == divisor).
+template <typename Label>
+inline SetWeight<Label, SET_UNION_INTERSECT> Divide(
+    const SetWeight<Label, SET_UNION_INTERSECT> &w1,
+    const SetWeight<Label, SET_UNION_INTERSECT> &w2,
+    DivideType divide_type = DIVIDE_ANY) {
+  using Weight = SetWeight<Label, SET_UNION_INTERSECT>;
+  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
+  if (w1 == w2) return Weight::UnivSet();
+  return w1;
+}
+
+// Divide = Or Not.
+template <typename Label>
+inline SetWeight<Label, SET_BOOLEAN> Divide(
+    const SetWeight<Label, SET_BOOLEAN> &w1,
+    const SetWeight<Label, SET_BOOLEAN> &w2,
+    DivideType divide_type = DIVIDE_ANY) {
+  using Weight = SetWeight<Label, SET_BOOLEAN>;
+  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
+  if (w1 == Weight::One()) return w1;
+  if (w2 == Weight::Zero()) return Weight::One();
+  return Weight::Zero();
+}
+
+// Converts between different set types.
+template <typename Label, SetType S1, SetType S2>
+struct WeightConvert<SetWeight<Label, S1>, SetWeight<Label, S2>> {
+  SetWeight<Label, S2> operator()(const SetWeight<Label, S1> &w1) const {
+    using Iterator = SetWeightIterator<SetWeight<Label, S1>>;
+    SetWeight<Label, S2> w2;
+    for (Iterator iter(w1); !iter.Done(); iter.Next())
+      w2.PushBack(iter.Value());
+    return w2;
+  }
+};
+
+// This function object generates SetWeights that are random integer sets
+// from {1, ... , alphabet_size}^{0, max_set_length} U { Zero }. This is
+// intended primarily for testing.
+template <class Label, SetType S>
+class WeightGenerate<SetWeight<Label, S>> {
+ public:
+  using Weight = SetWeight<Label, S>;
+
+  explicit WeightGenerate(bool allow_zero = true,
+                          size_t alphabet_size = kNumRandomWeights,
+                          size_t max_set_length = kNumRandomWeights)
+      : allow_zero_(allow_zero),
+        alphabet_size_(alphabet_size),
+        max_set_length_(max_set_length) {}
+
+  Weight operator()() const {
+    const size_t n = rand() % (max_set_length_ + allow_zero_);  // NOLINT
+    if (allow_zero_ && n == max_set_length_) return Weight::Zero();
+    std::vector<Label> labels;
+    for (size_t i = 0; i < n; ++i) {
+      labels.push_back(rand() % alphabet_size_ + 1);  // NOLINT
+    }
+    std::sort(labels.begin(), labels.end());
+    const auto labels_end = std::unique(labels.begin(), labels.end());
+    labels.resize(labels_end - labels.begin());
+    return Weight(labels.begin(), labels.end());
+  }
+
+ private:
+  // Permits Zero() and zero divisors.
+  const bool allow_zero_;
+  // Alphabet size for random weights.
+  const size_t alphabet_size_;
+  // Number of alternative random weights.
+  const size_t max_set_length_;
+};
+
+}  // namespace fst
+
+#endif  // FST_SET_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/shortest-distance.h b/projects/llm_framework/include/fst/shortest-distance.h
new file mode 100644
index 00000000..b1c5b26c
--- /dev/null
+++ b/projects/llm_framework/include/fst/shortest-distance.h
@@ -0,0 +1,351 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes to find shortest distance in an FST.
+
+#ifndef FST_SHORTEST_DISTANCE_H_
+#define FST_SHORTEST_DISTANCE_H_
+
+#include <cstddef>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/arcfilter.h>
+#include <fst/cache.h>
+#include <fst/queue.h>
+#include <fst/reverse.h>
+#include <fst/test-properties.h>
+
+
+namespace fst {
+
+// A representable float for shortest distance and shortest path algorithms.
+constexpr float kShortestDelta = 1e-6;
+
+template <class Arc, class Queue, class ArcFilter>
+struct ShortestDistanceOptions {
+  using StateId = typename Arc::StateId;
+
+  Queue *state_queue;    // Queue discipline used; owned by caller.
+  ArcFilter arc_filter;  // Arc filter (e.g., limit to only epsilon graph).
+  StateId source;        // If kNoStateId, use the FST's initial state.
+  float delta;           // Determines the degree of convergence required
+  bool first_path;       // For a semiring with the path property (o.w.
+                         // undefined), compute the shortest-distances along
+                         // along the first path to a final state found
+                         // by the algorithm. That path is the shortest-path
+                         // only if the FST has a unique final state (or all
+                         // the final states have the same final weight), the
+                         // queue discipline is shortest-first and all the
+                         // weights in the FST are between One() and Zero()
+                         // according to NaturalLess.
+
+  ShortestDistanceOptions(Queue *state_queue, ArcFilter arc_filter,
+                          StateId source = kNoStateId,
+                          float delta = kShortestDelta,
+                          bool first_path = false)
+      : state_queue(state_queue),
+        arc_filter(arc_filter),
+        source(source),
+        delta(delta),
+        first_path(first_path) {}
+};
+
+namespace internal {
+
+// Computation state of the shortest-distance algorithm. Reusable information
+// is maintained across calls to member function ShortestDistance(source) when
+// retain is true for improved efficiency when calling multiple times from
+// different source states (e.g., in epsilon removal). Contrary to the usual
+// conventions, fst may not be freed before this class. Vector distance
+// should not be modified by the user between these calls. The Error() method
+// returns true iff an error was encountered.
+template <class Arc, class Queue, class ArcFilter>
+class ShortestDistanceState {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  ShortestDistanceState(
+      const Fst<Arc> &fst, std::vector<Weight> *distance,
+      const ShortestDistanceOptions<Arc, Queue, ArcFilter> &opts, bool retain)
+      : fst_(fst),
+        distance_(distance),
+        state_queue_(opts.state_queue),
+        arc_filter_(opts.arc_filter),
+        delta_(opts.delta),
+        first_path_(opts.first_path),
+        retain_(retain),
+        source_id_(0),
+        error_(false) {
+    distance_->clear();
+    if (fst.Properties(kExpanded, false) == kExpanded) {
+      const auto num_states = CountStates(fst);
+      distance_->reserve(num_states);
+      adder_.reserve(num_states);
+      radder_.reserve(num_states);
+      enqueued_.reserve(num_states);
+    }
+  }
+
+  void ShortestDistance(StateId source);
+
+  bool Error() const { return error_; }
+
+ private:
+  void EnsureDistanceIndexIsValid(std::size_t index) {
+    while (distance_->size() <= index) {
+      distance_->push_back(Weight::Zero());
+      adder_.push_back(Adder<Weight>());
+      radder_.push_back(Adder<Weight>());
+      enqueued_.push_back(false);
+    }
+    DCHECK_LT(index, distance_->size());
+  }
+
+  void EnsureSourcesIndexIsValid(std::size_t index) {
+    while (sources_.size() <= index) {
+      sources_.push_back(kNoStateId);
+    }
+    DCHECK_LT(index, sources_.size());
+  }
+
+  const Fst<Arc> &fst_;
+  std::vector<Weight> *distance_;
+  Queue *state_queue_;
+  ArcFilter arc_filter_;
+  const float delta_;
+  const bool first_path_;
+  const bool retain_;  // Retain and reuse information across calls.
+
+  std::vector<Adder<Weight>> adder_;   // Sums distance_ accurately.
+  std::vector<Adder<Weight>> radder_;  // Relaxation distance.
+  std::vector<bool> enqueued_;         // Is state enqueued?
+  std::vector<StateId> sources_;       // Source ID for ith state in distance_,
+                                       // (r)adder_, and enqueued_ if retained.
+  StateId source_id_;                  // Unique ID characterizing each call.
+  bool error_;
+};
+
+// Compute the shortest distance; if source is kNoStateId, uses the initial
+// state of the FST.
+template <class Arc, class Queue, class ArcFilter>
+void ShortestDistanceState<Arc, Queue, ArcFilter>::ShortestDistance(
+    StateId source) {
+  if (fst_.Start() == kNoStateId) {
+    if (fst_.Properties(kError, false)) error_ = true;
+    return;
+  }
+  if (!(Weight::Properties() & kRightSemiring)) {
+    FSTERROR() << "ShortestDistance: Weight needs to be right distributive: "
+               << Weight::Type();
+    error_ = true;
+    return;
+  }
+  if (first_path_ && !(Weight::Properties() & kPath)) {
+    FSTERROR() << "ShortestDistance: The first_path option is disallowed when "
+               << "Weight does not have the path property: " << Weight::Type();
+    error_ = true;
+    return;
+  }
+  state_queue_->Clear();
+  if (!retain_) {
+    distance_->clear();
+    adder_.clear();
+    radder_.clear();
+    enqueued_.clear();
+  }
+  if (source == kNoStateId) source = fst_.Start();
+  EnsureDistanceIndexIsValid(source);
+  if (retain_) {
+    EnsureSourcesIndexIsValid(source);
+    sources_[source] = source_id_;
+  }
+  (*distance_)[source] = Weight::One();
+  adder_[source].Reset(Weight::One());
+  radder_[source].Reset(Weight::One());
+  enqueued_[source] = true;
+  state_queue_->Enqueue(source);
+  while (!state_queue_->Empty()) {
+    const auto state = state_queue_->Head();
+    state_queue_->Dequeue();
+    EnsureDistanceIndexIsValid(state);
+    if (first_path_ && (fst_.Final(state) != Weight::Zero())) break;
+    enqueued_[state] = false;
+    const auto r = radder_[state].Sum();
+    radder_[state].Reset();
+    for (ArcIterator<Fst<Arc>> aiter(fst_, state); !aiter.Done();
+         aiter.Next()) {
+      const auto &arc = aiter.Value();
+      const auto nextstate = arc.nextstate;
+      if (!arc_filter_(arc)) continue;
+      EnsureDistanceIndexIsValid(nextstate);
+      if (retain_) {
+        EnsureSourcesIndexIsValid(nextstate);
+        if (sources_[nextstate] != source_id_) {
+          (*distance_)[nextstate] = Weight::Zero();
+          adder_[nextstate].Reset();
+          radder_[nextstate].Reset();
+          enqueued_[nextstate] = false;
+          sources_[nextstate] = source_id_;
+        }
+      }
+      auto &nd = (*distance_)[nextstate];
+      auto &na = adder_[nextstate];
+      auto &nr = radder_[nextstate];
+      auto weight = Times(r, arc.weight);
+      if (!ApproxEqual(nd, Plus(nd, weight), delta_)) {
+        nd = na.Add(weight);
+        nr.Add(weight);
+        if (!nd.Member() || !nr.Sum().Member()) {
+          error_ = true;
+          return;
+        }
+        if (!enqueued_[nextstate]) {
+          state_queue_->Enqueue(nextstate);
+          enqueued_[nextstate] = true;
+        } else {
+          state_queue_->Update(nextstate);
+        }
+      }
+    }
+  }
+  ++source_id_;
+  if (fst_.Properties(kError, false)) error_ = true;
+}
+
+}  // namespace internal
+
+// Shortest-distance algorithm: this version allows fine control
+// via the options argument. See below for a simpler interface.
+//
+// This computes the shortest distance from the opts.source state to each
+// visited state S and stores the value in the distance vector. An
+// unvisited state S has distance Zero(), which will be stored in the
+// distance vector if S is less than the maximum visited state. The state
+// queue discipline, arc filter, and convergence delta are taken in the
+// options argument. The distance vector will contain a unique element for
+// which Member() is false if an error was encountered.
+//
+// The weights must must be right distributive and k-closed (i.e., 1 +
+// x + x^2 + ... + x^(k +1) = 1 + x + x^2 + ... + x^k).
+//
+// Complexity:
+//
+// Depends on properties of the semiring and the queue discipline.
+//
+// For more information, see:
+//
+// Mohri, M. 2002. Semiring framework and algorithms for shortest-distance
+// problems, Journal of Automata, Languages and
+// Combinatorics 7(3): 321-350, 2002.
+template <class Arc, class Queue, class ArcFilter>
+void ShortestDistance(
+    const Fst<Arc> &fst, std::vector<typename Arc::Weight> *distance,
+    const ShortestDistanceOptions<Arc, Queue, ArcFilter> &opts) {
+  internal::ShortestDistanceState<Arc, Queue, ArcFilter> sd_state(fst, distance,
+                                                                  opts, false);
+  sd_state.ShortestDistance(opts.source);
+  if (sd_state.Error()) {
+    distance->assign(1, Arc::Weight::NoWeight());
+  }
+}
+
+// Shortest-distance algorithm: simplified interface. See above for a version
+// that permits finer control.
+//
+// If reverse is false, this computes the shortest distance from the initial
+// state to each state S and stores the value in the distance vector. If
+// reverse is true, this computes the shortest distance from each state to the
+// final states. An unvisited state S has distance Zero(), which will be stored
+// in the distance vector if S is less than the maximum visited state. The
+// state queue discipline is automatically-selected. The distance vector will
+// contain a unique element for which Member() is false if an error was
+// encountered.
+//
+// The weights must must be right (left) distributive if reverse is false (true)
+// and k-closed (i.e., 1 + x + x^2 + ... + x^(k +1) = 1 + x + x^2 + ... + x^k).
+//
+// Arc weights must satisfy the property that the sum of the weights of one or
+// more paths from some state S to T is never Zero(). In particular, arc weights
+// are never Zero().
+//
+// Complexity:
+//
+// Depends on properties of the semiring and the queue discipline.
+//
+// For more information, see:
+//
+// Mohri, M. 2002. Semiring framework and algorithms for
+// shortest-distance problems, Journal of Automata, Languages and
+// Combinatorics 7(3): 321-350, 2002.
+template <class Arc>
+void ShortestDistance(const Fst<Arc> &fst,
+                      std::vector<typename Arc::Weight> *distance,
+                      bool reverse = false, float delta = kShortestDelta) {
+  using StateId = typename Arc::StateId;
+  if (!reverse) {
+    AnyArcFilter<Arc> arc_filter;
+    AutoQueue<StateId> state_queue(fst, distance, arc_filter);
+    const ShortestDistanceOptions<Arc, AutoQueue<StateId>, AnyArcFilter<Arc>>
+        opts(&state_queue, arc_filter, kNoStateId, delta);
+    ShortestDistance(fst, distance, opts);
+  } else {
+    using ReverseArc = ReverseArc<Arc>;
+    using ReverseWeight = typename ReverseArc::Weight;
+    AnyArcFilter<ReverseArc> rarc_filter;
+    VectorFst<ReverseArc> rfst;
+    Reverse(fst, &rfst);
+    std::vector<ReverseWeight> rdistance;
+    AutoQueue<StateId> state_queue(rfst, &rdistance, rarc_filter);
+    const ShortestDistanceOptions<ReverseArc, AutoQueue<StateId>,
+                                  AnyArcFilter<ReverseArc>>
+        ropts(&state_queue, rarc_filter, kNoStateId, delta);
+    ShortestDistance(rfst, &rdistance, ropts);
+    distance->clear();
+    if (rdistance.size() == 1 && !rdistance[0].Member()) {
+      distance->assign(1, Arc::Weight::NoWeight());
+      return;
+    }
+    DCHECK_GE(rdistance.size(), 1);  // reversing added one state
+    distance->reserve(rdistance.size() - 1);
+    while (distance->size() < rdistance.size() - 1) {
+      distance->push_back(rdistance[distance->size() + 1].Reverse());
+    }
+  }
+}
+
+// Return the sum of the weight of all successful paths in an FST, i.e., the
+// shortest-distance from the initial state to the final states. Returns a
+// weight such that Member() is false if an error was encountered.
+template <class Arc>
+typename Arc::Weight ShortestDistance(const Fst<Arc> &fst,
+                                      float delta = kShortestDelta) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  std::vector<Weight> distance;
+  if (Weight::Properties() & kRightSemiring) {
+    ShortestDistance(fst, &distance, false, delta);
+    if (distance.size() == 1 && !distance[0].Member()) {
+      return Arc::Weight::NoWeight();
+    }
+    Adder<Weight> adder;  // maintains cumulative sum accurately
+    for (StateId state = 0; state < distance.size(); ++state) {
+      adder.Add(Times(distance[state], fst.Final(state)));
+    }
+    return adder.Sum();
+  } else {
+    ShortestDistance(fst, &distance, true, delta);
+    const auto state = fst.Start();
+    if (distance.size() == 1 && !distance[0].Member()) {
+      return Arc::Weight::NoWeight();
+    }
+    return state != kNoStateId && state < distance.size() ? distance[state]
+                                                          : Weight::Zero();
+  }
+}
+
+}  // namespace fst
+
+#endif  // FST_SHORTEST_DISTANCE_H_
diff --git a/projects/llm_framework/include/fst/shortest-path.h b/projects/llm_framework/include/fst/shortest-path.h
new file mode 100644
index 00000000..7bd04711
--- /dev/null
+++ b/projects/llm_framework/include/fst/shortest-path.h
@@ -0,0 +1,549 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions to find shortest paths in an FST.
+
+#ifndef FST_SHORTEST_PATH_H_
+#define FST_SHORTEST_PATH_H_
+
+#include <functional>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/cache.h>
+#include <fst/determinize.h>
+#include <fst/queue.h>
+#include <fst/shortest-distance.h>
+#include <fst/test-properties.h>
+
+
+namespace fst {
+
+template <class Arc, class Queue, class ArcFilter>
+struct ShortestPathOptions
+    : public ShortestDistanceOptions<Arc, Queue, ArcFilter> {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  int32 nshortest;    // Returns n-shortest paths.
+  bool unique;        // Only returns paths with distinct input strings.
+  bool has_distance;  // Distance vector already contains the
+                      // shortest distance from the initial state.
+  bool first_path;    // Single shortest path stops after finding the first
+                      // path to a final state; that path is the shortest path
+                      // only when:
+                      // (1) using the ShortestFirstQueue with all the weights
+                      // in the FST being between One() and Zero() according to
+                      // NaturalLess or when
+                      // (2) using the NaturalAStarQueue with an admissible
+                      // and consistent estimate.
+  Weight weight_threshold;  // Pruning weight threshold.
+  StateId state_threshold;  // Pruning state threshold.
+
+  ShortestPathOptions(Queue *queue, ArcFilter filter, int32 nshortest = 1,
+                      bool unique = false, bool has_distance = false,
+                      float delta = kShortestDelta, bool first_path = false,
+                      Weight weight_threshold = Weight::Zero(),
+                      StateId state_threshold = kNoStateId)
+      : ShortestDistanceOptions<Arc, Queue, ArcFilter>(queue, filter,
+                                                       kNoStateId, delta),
+        nshortest(nshortest),
+        unique(unique),
+        has_distance(has_distance),
+        first_path(first_path),
+        weight_threshold(std::move(weight_threshold)),
+        state_threshold(state_threshold) {}
+};
+
+namespace internal {
+
+constexpr size_t kNoArc = -1;
+
+// Helper function for SingleShortestPath building the shortest path as a left-
+// to-right machine backwards from the best final state. It takes the input
+// FST passed to SingleShortestPath and the parent vector and f_parent returned
+// by that function, and builds the result into the provided output mutable FS
+// This is not normally called by users; see ShortestPath instead.
+template <class Arc>
+void SingleShortestPathBacktrace(
+    const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
+    const std::vector<std::pair<typename Arc::StateId, size_t>> &parent,
+    typename Arc::StateId f_parent) {
+  using StateId = typename Arc::StateId;
+  ofst->DeleteStates();
+  ofst->SetInputSymbols(ifst.InputSymbols());
+  ofst->SetOutputSymbols(ifst.OutputSymbols());
+  StateId s_p = kNoStateId;
+  StateId d_p = kNoStateId;
+  for (StateId state = f_parent, d = kNoStateId; state != kNoStateId;
+       d = state, state = parent[state].first) {
+    d_p = s_p;
+    s_p = ofst->AddState();
+    if (d == kNoStateId) {
+      ofst->SetFinal(s_p, ifst.Final(f_parent));
+    } else {
+      ArcIterator<Fst<Arc>> aiter(ifst, state);
+      aiter.Seek(parent[d].second);
+      auto arc = aiter.Value();
+      arc.nextstate = d_p;
+      ofst->AddArc(s_p, std::move(arc));
+    }
+  }
+  ofst->SetStart(s_p);
+  if (ifst.Properties(kError, false)) ofst->SetProperties(kError, kError);
+  ofst->SetProperties(
+      ShortestPathProperties(ofst->Properties(kFstProperties, false), true),
+      kFstProperties);
+}
+
+// Helper function for SingleShortestPath building a tree of shortest paths to
+// every final state in the input FST. It takes the input FST and parent values
+// computed by SingleShortestPath and builds into the output mutable FST the
+// subtree of ifst that consists only of the best paths to all final states.
+// This is not normally called by users; see ShortestPath instead.
+template <class Arc>
+void SingleShortestTree(
+    const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
+    const std::vector<std::pair<typename Arc::StateId, size_t>> &parent) {
+  ofst->DeleteStates();
+  ofst->SetInputSymbols(ifst.InputSymbols());
+  ofst->SetOutputSymbols(ifst.OutputSymbols());
+  ofst->SetStart(ifst.Start());
+  for (StateIterator<Fst<Arc>> siter(ifst); !siter.Done(); siter.Next()) {
+    ofst->AddState();
+    ofst->SetFinal(siter.Value(), ifst.Final(siter.Value()));
+  }
+  for (const auto &pair : parent) {
+    if (pair.first != kNoStateId && pair.second != kNoArc) {
+      ArcIterator<Fst<Arc>> aiter(ifst, pair.first);
+      aiter.Seek(pair.second);
+      ofst->AddArc(pair.first, aiter.Value());
+    }
+  }
+  if (ifst.Properties(kError, false)) ofst->SetProperties(kError, kError);
+  ofst->SetProperties(
+      ShortestPathProperties(ofst->Properties(kFstProperties, false), true),
+      kFstProperties);
+}
+
+// Implements the stopping criterion when ShortestPathOptions::first_path
+// is set to true:
+//   operator()(s, d, f) == true
+//   iff every successful path through state 's' has a cost greater or equal
+//   to 'f' under the assumption that 'd' is the shortest distance to state 's'.
+// Correct when using the ShortestFirstQueue with all the weights in the FST
+// being between One() and Zero() according to NaturalLess
+template <typename S, typename W, typename Queue>
+struct FirstPathSelect {
+  FirstPathSelect(const Queue &) {}
+  bool operator()(S s, W d, W f) const { return f == Plus(d, f); }
+};
+
+// Specialisation for A*.
+// Correct when the estimate is admissible and consistent.
+template <typename S, typename W, typename Estimate>
+class FirstPathSelect<S, W, NaturalAStarQueue<S, W, Estimate>> {
+ public:
+  using Queue = NaturalAStarQueue<S, W, Estimate>;
+
+  FirstPathSelect(const Queue &state_queue)
+    : estimate_(state_queue.GetCompare().GetEstimate()) {}
+
+  bool operator()(S s, W d, W f) const {
+    return f == Plus(Times(d, estimate_(s)), f);
+  }
+
+ private:
+  const Estimate &estimate_;
+};
+
+// Shortest-path algorithm. It builds the output mutable FST so that it contains
+// the shortest path in the input FST; distance returns the shortest distances
+// from the source state to each state in the input FST, and the options struct
+// is
+// used to specify options such as the queue discipline, the arc filter and
+// delta. The super_final option is an output parameter indicating the final
+// state, and the parent argument is used for the storage of the backtrace path
+// for each state 1 to n, (i.e., the best previous state and the arc that
+// transition to state n.) The shortest path is the lowest weight path w.r.t.
+// the natural semiring order. The weights need to be right distributive and
+// have the path (kPath) property. False is returned if an error is encountered.
+//
+// This is not normally called by users; see ShortestPath instead (with n = 1).
+template <class Arc, class Queue, class ArcFilter>
+bool SingleShortestPath(
+    const Fst<Arc> &ifst, std::vector<typename Arc::Weight> *distance,
+    const ShortestPathOptions<Arc, Queue, ArcFilter> &opts,
+    typename Arc::StateId *f_parent,
+    std::vector<std::pair<typename Arc::StateId, size_t>> *parent) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  static_assert(IsPath<Weight>::value, "Weight must have path property.");
+  static_assert((Weight::Properties() & kRightSemiring) == kRightSemiring,
+                "Weight must be right distributive.");
+  parent->clear();
+  *f_parent = kNoStateId;
+  if (ifst.Start() == kNoStateId) return true;
+  std::vector<bool> enqueued;
+  auto state_queue = opts.state_queue;
+  const auto source = (opts.source == kNoStateId) ? ifst.Start() : opts.source;
+  bool final_seen = false;
+  auto f_distance = Weight::Zero();
+  distance->clear();
+  state_queue->Clear();
+  while (distance->size() < source) {
+    distance->push_back(Weight::Zero());
+    enqueued.push_back(false);
+    parent->emplace_back(kNoStateId, kNoArc);
+  }
+  distance->push_back(Weight::One());
+  parent->emplace_back(kNoStateId, kNoArc);
+  state_queue->Enqueue(source);
+  enqueued.push_back(true);
+  while (!state_queue->Empty()) {
+    const auto s = state_queue->Head();
+    state_queue->Dequeue();
+    enqueued[s] = false;
+    const auto sd = (*distance)[s];
+    // If we are using a shortest queue, no other path is going to be shorter
+    // than f_distance at this point.
+    using FirstPath = FirstPathSelect<StateId, Weight, Queue>;
+    if (opts.first_path && final_seen &&
+        FirstPath(*state_queue)(s, sd, f_distance)) {
+      break;
+    }
+    if (ifst.Final(s) != Weight::Zero()) {
+      const auto plus = Plus(f_distance, Times(sd, ifst.Final(s)));
+      if (f_distance != plus) {
+        f_distance = plus;
+        *f_parent = s;
+      }
+      if (!f_distance.Member()) return false;
+      final_seen = true;
+    }
+    for (ArcIterator<Fst<Arc>> aiter(ifst, s); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      while (distance->size() <= arc.nextstate) {
+        distance->push_back(Weight::Zero());
+        enqueued.push_back(false);
+        parent->emplace_back(kNoStateId, kNoArc);
+      }
+      auto &nd = (*distance)[arc.nextstate];
+      const auto weight = Times(sd, arc.weight);
+      if (nd != Plus(nd, weight)) {
+        nd = Plus(nd, weight);
+        if (!nd.Member()) return false;
+        (*parent)[arc.nextstate] = std::make_pair(s, aiter.Position());
+        if (!enqueued[arc.nextstate]) {
+          state_queue->Enqueue(arc.nextstate);
+          enqueued[arc.nextstate] = true;
+        } else {
+          state_queue->Update(arc.nextstate);
+        }
+      }
+    }
+  }
+  return true;
+}
+
+template <class StateId, class Weight>
+class ShortestPathCompare {
+ public:
+  ShortestPathCompare(const std::vector<std::pair<StateId, Weight>> &pairs,
+                      const std::vector<Weight> &distance, StateId superfinal,
+                      float delta)
+      : pairs_(pairs),
+        distance_(distance),
+        superfinal_(superfinal),
+        delta_(delta) {}
+
+  bool operator()(const StateId x, const StateId y) const {
+    const auto &px = pairs_[x];
+    const auto &py = pairs_[y];
+    const auto wx = Times(PWeight(px.first), px.second);
+    const auto wy = Times(PWeight(py.first), py.second);
+    // Penalize complete paths to ensure correct results with inexact weights.
+    // This forms a strict weak order so long as ApproxEqual(a, b) =>
+    // ApproxEqual(a, c) for all c s.t. less_(a, c) && less_(c, b).
+    if (px.first == superfinal_ && py.first != superfinal_) {
+      return less_(wy, wx) || ApproxEqual(wx, wy, delta_);
+    } else if (py.first == superfinal_ && px.first != superfinal_) {
+      return less_(wy, wx) && !ApproxEqual(wx, wy, delta_);
+    } else {
+      return less_(wy, wx);
+    }
+  }
+
+ private:
+  Weight PWeight(StateId state) const {
+    return (state == superfinal_)
+               ? Weight::One()
+               : (state < distance_.size()) ? distance_[state] : Weight::Zero();
+  }
+
+  const std::vector<std::pair<StateId, Weight>> &pairs_;
+  const std::vector<Weight> &distance_;
+  const StateId superfinal_;
+  const float delta_;
+  NaturalLess<Weight> less_;
+};
+
+// N-Shortest-path algorithm: implements the core n-shortest path algorithm.
+// The output is built reversed. See below for versions with more options and
+// *not reversed*.
+//
+// The output mutable FST contains the REVERSE of n'shortest paths in the input
+// FST; distance must contain the shortest distance from each state to a final
+// state in the input FST; delta is the convergence delta.
+//
+// The n-shortest paths are the n-lowest weight paths w.r.t. the natural
+// semiring order. The single path that can be read from the ith of at most n
+// transitions leaving the initial state of the input FST is the ith shortest
+// path. Disregarding the initial state and initial transitions, the
+// n-shortest paths, in fact, form a tree rooted at the single final state.
+//
+// The weights need to be left and right distributive (kSemiring) and have the
+// path (kPath) property.
+//
+// Arc weights must satisfy the property that the sum of the weights of one or
+// more paths from some state S to T is never Zero(). In particular, arc weights
+// are never Zero().
+//
+// For more information, see:
+//
+// Mohri, M, and Riley, M. 2002. An efficient algorithm for the n-best-strings
+// problem. In Proc. ICSLP.
+//
+// The algorithm relies on the shortest-distance algorithm. There are some
+// issues with the pseudo-code as written in the paper (viz., line 11).
+//
+// IMPLEMENTATION NOTE: The input FST can be a delayed FST and at any state in
+// its expansion the values of distance vector need only be defined at that time
+// for the states that are known to exist.
+template <class Arc, class RevArc>
+void NShortestPath(const Fst<RevArc> &ifst, MutableFst<Arc> *ofst,
+                   const std::vector<typename Arc::Weight> &distance,
+                   int32 nshortest, float delta = kShortestDelta,
+                   typename Arc::Weight weight_threshold = Arc::Weight::Zero(),
+                   typename Arc::StateId state_threshold = kNoStateId) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using Pair = std::pair<StateId, Weight>;
+  static_assert((Weight::Properties() & kPath) == kPath,
+                "Weight must have path property.");
+  static_assert((Weight::Properties() & kSemiring) == kSemiring,
+                "Weight must be distributive.");
+  if (nshortest <= 0) return;
+  ofst->DeleteStates();
+  ofst->SetInputSymbols(ifst.InputSymbols());
+  ofst->SetOutputSymbols(ifst.OutputSymbols());
+  // Each state in ofst corresponds to a path with weight w from the initial
+  // state of ifst to a state s in ifst, that can be characterized by a pair
+  // (s, w). The vector pairs maps each state in ofst to the corresponding
+  // pair maps states in ofst to the corresponding pair (s, w).
+  std::vector<Pair> pairs;
+  // The supefinal state is denoted by kNoStateId. The distance from the
+  // superfinal state to the final state is semiring One, so
+  // `distance[kNoStateId]` is not needed.
+  const ShortestPathCompare<StateId, Weight> compare(pairs, distance,
+                                                     kNoStateId, delta);
+  const NaturalLess<Weight> less;
+  if (ifst.Start() == kNoStateId || distance.size() <= ifst.Start() ||
+      distance[ifst.Start()] == Weight::Zero() ||
+      less(weight_threshold, Weight::One()) || state_threshold == 0) {
+    if (ifst.Properties(kError, false)) ofst->SetProperties(kError, kError);
+    return;
+  }
+  ofst->SetStart(ofst->AddState());
+  const auto final_state = ofst->AddState();
+  ofst->SetFinal(final_state, Weight::One());
+  while (pairs.size() <= final_state) {
+    pairs.push_back(std::make_pair(kNoStateId, Weight::Zero()));
+  }
+  pairs[final_state] = std::make_pair(ifst.Start(), Weight::One());
+  std::vector<StateId> heap;
+  heap.push_back(final_state);
+  const auto limit = Times(distance[ifst.Start()], weight_threshold);
+  // r[s + 1], s state in fst, is the number of states in ofst which
+  // corresponding pair contains s, i.e., it is number of paths computed so far
+  // to s. Valid for s == kNoStateId (the superfinal state).
+  std::vector<int> r;
+  while (!heap.empty()) {
+    std::pop_heap(heap.begin(), heap.end(), compare);
+    const auto state = heap.back();
+    const auto p = pairs[state];
+    heap.pop_back();
+    const auto d =
+        (p.first == kNoStateId)
+            ? Weight::One()
+            : (p.first < distance.size()) ? distance[p.first] : Weight::Zero();
+    if (less(limit, Times(d, p.second)) ||
+        (state_threshold != kNoStateId &&
+         ofst->NumStates() >= state_threshold)) {
+      continue;
+    }
+    while (r.size() <= p.first + 1) r.push_back(0);
+    ++r[p.first + 1];
+    if (p.first == kNoStateId) {
+      ofst->AddArc(ofst->Start(), Arc(0, 0, Weight::One(), state));
+    }
+    if ((p.first == kNoStateId) && (r[p.first + 1] == nshortest)) break;
+    if (r[p.first + 1] > nshortest) continue;
+    if (p.first == kNoStateId) continue;
+    for (ArcIterator<Fst<RevArc>> aiter(ifst, p.first); !aiter.Done();
+         aiter.Next()) {
+      const auto &rarc = aiter.Value();
+      Arc arc(rarc.ilabel, rarc.olabel, rarc.weight.Reverse(), rarc.nextstate);
+      const auto weight = Times(p.second, arc.weight);
+      const auto next = ofst->AddState();
+      pairs.push_back(std::make_pair(arc.nextstate, weight));
+      arc.nextstate = state;
+      ofst->AddArc(next, std::move(arc));
+      heap.push_back(next);
+      std::push_heap(heap.begin(), heap.end(), compare);
+    }
+    const auto final_weight = ifst.Final(p.first).Reverse();
+    if (final_weight != Weight::Zero()) {
+      const auto weight = Times(p.second, final_weight);
+      const auto next = ofst->AddState();
+      pairs.push_back(std::make_pair(kNoStateId, weight));
+      ofst->AddArc(next, Arc(0, 0, final_weight, state));
+      heap.push_back(next);
+      std::push_heap(heap.begin(), heap.end(), compare);
+    }
+  }
+  Connect(ofst);
+  if (ifst.Properties(kError, false)) ofst->SetProperties(kError, kError);
+  ofst->SetProperties(
+      ShortestPathProperties(ofst->Properties(kFstProperties, false)),
+      kFstProperties);
+}
+
+}  // namespace internal
+
+// N-Shortest-path algorithm: this version allows finer control via the options
+// argument. See below for a simpler interface. The output mutable FST contains
+// the n-shortest paths in the input FST; the distance argument is used to
+// return the shortest distances from the source state to each state in the
+// input FST, and the options struct is used to specify the number of paths to
+// return, whether they need to have distinct input strings, the queue
+// discipline, the arc filter and the convergence delta.
+//
+// The n-shortest paths are the n-lowest weight paths w.r.t. the natural
+// semiring order. The single path that can be read from the ith of at most n
+// transitions leaving the initial state of the output FST is the ith shortest
+// path.
+// Disregarding the initial state and initial transitions, The n-shortest paths,
+// in fact, form a tree rooted at the single final state.
+//
+// The weights need to be right distributive and have the path (kPath) property.
+// They need to be left distributive as well for nshortest > 1.
+//
+// For more information, see:
+//
+// Mohri, M, and Riley, M. 2002. An efficient algorithm for the n-best-strings
+// problem. In Proc. ICSLP.
+//
+// The algorithm relies on the shortest-distance algorithm. There are some
+// issues with the pseudo-code as written in the paper (viz., line 11).
+template <class Arc, class Queue, class ArcFilter,
+          typename std::enable_if<IsPath<typename Arc::Weight>::value>::type * =
+              nullptr>
+void ShortestPath(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
+                  std::vector<typename Arc::Weight> *distance,
+                  const ShortestPathOptions<Arc, Queue, ArcFilter> &opts) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using RevArc = ReverseArc<Arc>;
+  if (opts.nshortest == 1) {
+    std::vector<std::pair<StateId, size_t>> parent;
+    StateId f_parent;
+    if (internal::SingleShortestPath(ifst, distance, opts, &f_parent,
+                                     &parent)) {
+      internal::SingleShortestPathBacktrace(ifst, ofst, parent, f_parent);
+    } else {
+      ofst->SetProperties(kError, kError);
+    }
+    return;
+  }
+  if (opts.nshortest <= 0) return;
+  if (!opts.has_distance) {
+    ShortestDistance(ifst, distance, opts);
+    if (distance->size() == 1 && !(*distance)[0].Member()) {
+      ofst->SetProperties(kError, kError);
+      return;
+    }
+  }
+  // Algorithm works on the reverse of 'fst'; 'distance' is the distance to the
+  // final state in 'rfst', 'ofst' is built as the reverse of the tree of
+  // n-shortest path in 'rfst'.
+  VectorFst<RevArc> rfst;
+  Reverse(ifst, &rfst);
+  auto d = Weight::Zero();
+  for (ArcIterator<VectorFst<RevArc>> aiter(rfst, 0); !aiter.Done();
+       aiter.Next()) {
+    const auto &arc = aiter.Value();
+    const auto state = arc.nextstate - 1;
+    if (state < distance->size()) {
+      d = Plus(d, Times(arc.weight.Reverse(), (*distance)[state]));
+    }
+  }
+  // TODO(kbg): Avoid this expensive vector operation.
+  distance->insert(distance->begin(), d);
+  if (!opts.unique) {
+    internal::NShortestPath(rfst, ofst, *distance, opts.nshortest, opts.delta,
+                            opts.weight_threshold, opts.state_threshold);
+  } else {
+    std::vector<Weight> ddistance;
+    DeterminizeFstOptions<RevArc> dopts(opts.delta);
+    DeterminizeFst<RevArc> dfst(rfst, distance, &ddistance, dopts);
+    internal::NShortestPath(dfst, ofst, ddistance, opts.nshortest, opts.delta,
+                            opts.weight_threshold, opts.state_threshold);
+  }
+  // TODO(kbg): Avoid this expensive vector operation.
+  distance->erase(distance->begin());
+}
+
+template <class Arc, class Queue, class ArcFilter,
+          typename std::enable_if<!IsPath<typename Arc::Weight>::value>::type
+              * = nullptr>
+void ShortestPath(const Fst<Arc> &, MutableFst<Arc> *ofst,
+                  std::vector<typename Arc::Weight> *,
+                  const ShortestPathOptions<Arc, Queue, ArcFilter> &) {
+  FSTERROR() << "ShortestPath: Weight needs to have the "
+             << "path property and be distributive: " << Arc::Weight::Type();
+  ofst->SetProperties(kError, kError);
+}
+
+// Shortest-path algorithm: simplified interface. See above for a version that
+// allows finer control. The output mutable FST contains the n-shortest paths
+// in the input FST. The queue discipline is automatically selected. When unique
+// is true, only paths with distinct input label sequences are returned.
+//
+// The n-shortest paths are the n-lowest weight paths w.r.t. the natural
+// semiring order. The single path that can be read from the ith of at most n
+// transitions leaving the initial state of the output FST is the ith best path.
+// The weights need to be right distributive and have the path (kPath) property.
+template <class Arc>
+void ShortestPath(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
+                  int32 nshortest = 1, bool unique = false,
+                  bool first_path = false,
+                  typename Arc::Weight weight_threshold = Arc::Weight::Zero(),
+                  typename Arc::StateId state_threshold = kNoStateId,
+                  float delta = kShortestDelta) {
+  using StateId = typename Arc::StateId;
+  std::vector<typename Arc::Weight> distance;
+  AnyArcFilter<Arc> arc_filter;
+  AutoQueue<StateId> state_queue(ifst, &distance, arc_filter);
+  const ShortestPathOptions<Arc, AutoQueue<StateId>, AnyArcFilter<Arc>> opts(
+      &state_queue, arc_filter, nshortest, unique, false, delta, first_path,
+      weight_threshold, state_threshold);
+  ShortestPath(ifst, ofst, &distance, opts);
+}
+
+}  // namespace fst
+
+#endif  // FST_SHORTEST_PATH_H_
diff --git a/projects/llm_framework/include/fst/signed-log-weight.h b/projects/llm_framework/include/fst/signed-log-weight.h
new file mode 100644
index 00000000..41339bc8
--- /dev/null
+++ b/projects/llm_framework/include/fst/signed-log-weight.h
@@ -0,0 +1,440 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// LogWeight along with sign information that represents the value X in the
+// linear domain as <sign(X), -ln(|X|)>
+//
+// The sign is a TropicalWeight:
+//  positive, TropicalWeight.Value() > 0.0, recommended value 1.0
+//  negative, TropicalWeight.Value() <= 0.0, recommended value -1.0
+
+#ifndef FST_SIGNED_LOG_WEIGHT_H_
+#define FST_SIGNED_LOG_WEIGHT_H_
+
+#include <cstdlib>
+
+#include <fst/float-weight.h>
+#include <fst/pair-weight.h>
+#include <fst/product-weight.h>
+
+
+namespace fst {
+template <class T>
+class SignedLogWeightTpl : public PairWeight<TropicalWeight, LogWeightTpl<T>> {
+ public:
+  using X1 = TropicalWeight;
+  using X2 = LogWeightTpl<T>;
+  using ReverseWeight = SignedLogWeightTpl;
+
+  using PairWeight<X1, X2>::Value1;
+  using PairWeight<X1, X2>::Value2;
+
+  SignedLogWeightTpl() noexcept : PairWeight<X1, X2>() {}
+
+  explicit SignedLogWeightTpl(const PairWeight<X1, X2> &w)
+      : PairWeight<X1, X2>(w) {}
+
+  SignedLogWeightTpl(const X1 &x1, const X2 &x2) : PairWeight<X1, X2>(x1, x2) {}
+
+  static const SignedLogWeightTpl &Zero() {
+    static const SignedLogWeightTpl zero(X1(1.0), X2::Zero());
+    return zero;
+  }
+
+  static const SignedLogWeightTpl &One() {
+    static const SignedLogWeightTpl one(X1(1.0), X2::One());
+    return one;
+  }
+
+  static const SignedLogWeightTpl &NoWeight() {
+    static const SignedLogWeightTpl no_weight(X1(1.0), X2::NoWeight());
+    return no_weight;
+  }
+
+  static const string &Type() {
+    static const string *const type =
+        new string("signed_log_" + X1::Type() + "_" + X2::Type());
+    return *type;
+  }
+
+  bool IsPositive() const { return Value1().Value() > 0; }
+
+  SignedLogWeightTpl Quantize(float delta = kDelta) const {
+    return SignedLogWeightTpl(PairWeight<X1, X2>::Quantize(delta));
+  }
+
+  ReverseWeight Reverse() const {
+    return SignedLogWeightTpl(PairWeight<X1, X2>::Reverse());
+  }
+
+  bool Member() const { return PairWeight<X1, X2>::Member(); }
+
+  // Neither idempotent nor path.
+  static constexpr uint64 Properties() {
+    return kLeftSemiring | kRightSemiring | kCommutative;
+  }
+
+  size_t Hash() const {
+    size_t h1;
+    if (Value2() == X2::Zero() || IsPositive()) {
+      h1 = TropicalWeight(1.0).Hash();
+    } else {
+      h1 = TropicalWeight(-1.0).Hash();
+    }
+    size_t h2 = Value2().Hash();
+    static constexpr int lshift = 5;
+    static constexpr int rshift = CHAR_BIT * sizeof(size_t) - 5;
+    return h1 << lshift ^ h1 >> rshift ^ h2;
+  }
+};
+
+template <class T>
+inline SignedLogWeightTpl<T> Plus(const SignedLogWeightTpl<T> &w1,
+                                  const SignedLogWeightTpl<T> &w2) {
+  using X1 = TropicalWeight;
+  using X2 = LogWeightTpl<T>;
+  if (!w1.Member() || !w2.Member()) return SignedLogWeightTpl<T>::NoWeight();
+  const auto s1 = w1.IsPositive();
+  const auto s2 = w2.IsPositive();
+  const bool equal = (s1 == s2);
+  const auto f1 = w1.Value2().Value();
+  const auto f2 = w2.Value2().Value();
+  if (f1 == FloatLimits<T>::PosInfinity()) {
+    return w2;
+  } else if (f2 == FloatLimits<T>::PosInfinity()) {
+    return w1;
+  } else if (f1 == f2) {
+    if (equal) {
+      return SignedLogWeightTpl<T>(X1(w1.Value1()), X2(f2 - log(2.0F)));
+    } else {
+      return SignedLogWeightTpl<T>::Zero();
+    }
+  } else if (f1 > f2) {
+    if (equal) {
+      return SignedLogWeightTpl<T>(X1(w1.Value1()),
+                                   X2(f2 - internal::LogPosExp(f1 - f2)));
+    } else {
+      return SignedLogWeightTpl<T>(X1(w2.Value1()),
+                                   X2((f2 - internal::LogNegExp(f1 - f2))));
+    }
+  } else {
+    if (equal) {
+      return SignedLogWeightTpl<T>(X1(w2.Value1()),
+                                   X2((f1 - internal::LogPosExp(f2 - f1))));
+    } else {
+      return SignedLogWeightTpl<T>(X1(w1.Value1()),
+                                   X2((f1 - internal::LogNegExp(f2 - f1))));
+    }
+  }
+}
+
+template <class T>
+inline SignedLogWeightTpl<T> Minus(const SignedLogWeightTpl<T> &w1,
+                                   const SignedLogWeightTpl<T> &w2) {
+  SignedLogWeightTpl<T> minus_w2(-w2.Value1().Value(), w2.Value2());
+  return Plus(w1, minus_w2);
+}
+
+template <class T>
+inline SignedLogWeightTpl<T> Times(const SignedLogWeightTpl<T> &w1,
+                                   const SignedLogWeightTpl<T> &w2) {
+  using X2 = LogWeightTpl<T>;
+  if (!w1.Member() || !w2.Member()) return SignedLogWeightTpl<T>::NoWeight();
+  const auto s1 = w1.IsPositive();
+  const auto s2 = w2.IsPositive();
+  const auto f1 = w1.Value2().Value();
+  const auto f2 = w2.Value2().Value();
+  if (s1 == s2) {
+    return SignedLogWeightTpl<T>(TropicalWeight(1.0), X2(f1 + f2));
+  } else {
+    return SignedLogWeightTpl<T>(TropicalWeight(-1.0), X2(f1 + f2));
+  }
+}
+
+template <class T>
+inline SignedLogWeightTpl<T> Divide(const SignedLogWeightTpl<T> &w1,
+                                    const SignedLogWeightTpl<T> &w2,
+                                    DivideType typ = DIVIDE_ANY) {
+  using X2 = LogWeightTpl<T>;
+  if (!w1.Member() || !w2.Member()) return SignedLogWeightTpl<T>::NoWeight();
+  const auto s1 = w1.IsPositive();
+  const auto s2 = w2.IsPositive();
+  const auto f1 = w1.Value2().Value();
+  const auto f2 = w2.Value2().Value();
+  if (f2 == FloatLimits<T>::PosInfinity()) {
+    return SignedLogWeightTpl<T>(TropicalWeight(1.0),
+                                 X2(FloatLimits<T>::NumberBad()));
+  } else if (f1 == FloatLimits<T>::PosInfinity()) {
+    return SignedLogWeightTpl<T>(TropicalWeight(1.0),
+                                 X2(FloatLimits<T>::PosInfinity()));
+  } else if (s1 == s2) {
+    return SignedLogWeightTpl<T>(TropicalWeight(1.0), X2(f1 - f2));
+  } else {
+    return SignedLogWeightTpl<T>(TropicalWeight(-1.0), X2(f1 - f2));
+  }
+}
+
+template <class T>
+inline bool ApproxEqual(const SignedLogWeightTpl<T> &w1,
+                        const SignedLogWeightTpl<T> &w2, float delta = kDelta) {
+  using X2 = LogWeightTpl<T>;
+  if (w1.IsPositive() == w2.IsPositive()) {
+    return ApproxEqual(w1.Value2(), w2.Value2(), delta);
+  } else {
+    return ApproxEqual(w1.Value2(), X2::Zero(), delta)
+        && ApproxEqual(w2.Value2(), X2::Zero(), delta);
+  }
+}
+
+template <class T>
+inline bool operator==(const SignedLogWeightTpl<T> &w1,
+                       const SignedLogWeightTpl<T> &w2) {
+  using X2 = LogWeightTpl<T>;
+  if (w1.IsPositive() == w2.IsPositive()) {
+    return w1.Value2() == w2.Value2();
+  } else {
+    return w1.Value2() == X2::Zero()
+        && w2.Value2() == X2::Zero();
+  }
+}
+
+template <class T>
+inline bool operator!=(const SignedLogWeightTpl<T> &w1,
+                       const SignedLogWeightTpl<T> &w2) {
+  return !(w1 == w2);
+}
+
+// Single-precision signed-log weight.
+using SignedLogWeight = SignedLogWeightTpl<float>;
+
+// Double-precision signed-log weight.
+using SignedLog64Weight = SignedLogWeightTpl<double>;
+
+template <class W1, class W2>
+bool SignedLogConvertCheck(W1 weight) {
+  if (weight.Value1().Value() < 0.0) {
+    FSTERROR() << "WeightConvert: Can't convert weight " << weight
+               << " from " << W1::Type() << " to " << W2::Type();
+    return false;
+  }
+  return true;
+}
+
+// Specialization using the Kahan compensated summation
+template <class T>
+class Adder<SignedLogWeightTpl<T>> {
+ public:
+  using Weight = SignedLogWeightTpl<T>;
+  using X1 = TropicalWeight;
+  using X2 = LogWeightTpl<T>;
+
+  explicit Adder(Weight w = Weight::Zero())
+      : ssum_(w.IsPositive()),
+        sum_(w.Value2().Value()),
+        c_(0.0) { }
+
+  Weight Add(const Weight &w) {
+    const auto sw = w.IsPositive();
+    const auto f = w.Value2().Value();
+    const bool equal = (ssum_ == sw);
+
+    if (!Sum().Member() || f == FloatLimits<T>::PosInfinity()) {
+      return Sum();
+    } else if (!w.Member() || sum_ == FloatLimits<T>::PosInfinity()) {
+      sum_ = f;
+      ssum_ = sw;
+      c_ = 0.0;
+    } else if (f == sum_) {
+      if (equal) {
+        sum_ = internal::KahanLogSum(sum_, f, &c_);
+      } else {
+        sum_ = FloatLimits<T>::PosInfinity();
+        ssum_ = true;
+        c_ = 0.0;
+      }
+    } else if (f > sum_) {
+      if (equal) {
+        sum_ = internal::KahanLogSum(sum_, f, &c_);
+      } else {
+        sum_ = internal::KahanLogDiff(sum_, f, &c_);
+      }
+    } else {
+      if (equal) {
+        sum_ = internal::KahanLogSum(f, sum_, &c_);
+      } else {
+        sum_ = internal::KahanLogDiff(f, sum_, &c_);
+        ssum_ = sw;
+      }
+    }
+    return Sum();
+  }
+
+  Weight Sum() { return Weight(X1(ssum_ ? 1.0 : -1.0), X2(sum_)); }
+
+  void Reset(Weight w = Weight::Zero()) {
+    ssum_ = w.IsPositive();
+    sum_ = w.Value2().Value();
+    c_ = 0.0;
+  }
+
+ private:
+  bool ssum_;   // true iff sign of sum is positive
+  double sum_;  // unsigned sum
+  double c_;    // Kahan compensation
+};
+
+// Converts to tropical.
+template <>
+struct WeightConvert<SignedLogWeight, TropicalWeight> {
+  TropicalWeight operator()(const SignedLogWeight &weight) const {
+    if (!SignedLogConvertCheck<SignedLogWeight, TropicalWeight>(weight)) {
+      return TropicalWeight::NoWeight();
+    }
+    return TropicalWeight(weight.Value2().Value());
+  }
+};
+
+template <>
+struct WeightConvert<SignedLog64Weight, TropicalWeight> {
+  TropicalWeight operator()(const SignedLog64Weight &weight) const {
+    if (!SignedLogConvertCheck<SignedLog64Weight, TropicalWeight>(weight)) {
+      return TropicalWeight::NoWeight();
+    }
+    return TropicalWeight(weight.Value2().Value());
+  }
+};
+
+// Converts to log.
+template <>
+struct WeightConvert<SignedLogWeight, LogWeight> {
+  LogWeight operator()(const SignedLogWeight &weight) const {
+    if (!SignedLogConvertCheck<SignedLogWeight, LogWeight>(weight)) {
+      return LogWeight::NoWeight();
+    }
+    return LogWeight(weight.Value2().Value());
+  }
+};
+
+template <>
+struct WeightConvert<SignedLog64Weight, LogWeight> {
+  LogWeight operator()(const SignedLog64Weight &weight) const {
+    if (!SignedLogConvertCheck<SignedLog64Weight, LogWeight>(weight)) {
+      return LogWeight::NoWeight();
+    }
+    return LogWeight(weight.Value2().Value());
+  }
+};
+
+// Converts to log64.
+template <>
+struct WeightConvert<SignedLogWeight, Log64Weight> {
+  Log64Weight operator()(const SignedLogWeight &weight) const {
+    if (!SignedLogConvertCheck<SignedLogWeight, Log64Weight>(weight)) {
+      return Log64Weight::NoWeight();
+    }
+    return Log64Weight(weight.Value2().Value());
+  }
+};
+
+template <>
+struct WeightConvert<SignedLog64Weight, Log64Weight> {
+  Log64Weight operator()(const SignedLog64Weight &weight) const {
+    if (!SignedLogConvertCheck<SignedLog64Weight, Log64Weight>(weight)) {
+      return Log64Weight::NoWeight();
+    }
+    return Log64Weight(weight.Value2().Value());
+  }
+};
+
+// Converts to signed log.
+template <>
+struct WeightConvert<TropicalWeight, SignedLogWeight> {
+  SignedLogWeight operator()(const TropicalWeight &weight) const {
+    return SignedLogWeight(1.0, weight.Value());
+  }
+};
+
+template <>
+struct WeightConvert<LogWeight, SignedLogWeight> {
+  SignedLogWeight operator()(const LogWeight &weight) const {
+    return SignedLogWeight(1.0, weight.Value());
+  }
+};
+
+template <>
+struct WeightConvert<Log64Weight, SignedLogWeight> {
+  SignedLogWeight operator()(const Log64Weight &weight) const {
+    return SignedLogWeight(1.0, weight.Value());
+  }
+};
+
+template <>
+struct WeightConvert<SignedLog64Weight, SignedLogWeight> {
+  SignedLogWeight operator()(const SignedLog64Weight &weight) const {
+    return SignedLogWeight(weight.Value1(), weight.Value2().Value());
+  }
+};
+
+// Converts to signed log64.
+template <>
+struct WeightConvert<TropicalWeight, SignedLog64Weight> {
+  SignedLog64Weight operator()(const TropicalWeight &weight) const {
+    return SignedLog64Weight(1.0, weight.Value());
+  }
+};
+
+template <>
+struct WeightConvert<LogWeight, SignedLog64Weight> {
+  SignedLog64Weight operator()(const LogWeight &weight) const {
+    return SignedLog64Weight(1.0, weight.Value());
+  }
+};
+
+template <>
+struct WeightConvert<Log64Weight, SignedLog64Weight> {
+  SignedLog64Weight operator()(const Log64Weight &weight) const {
+    return SignedLog64Weight(1.0, weight.Value());
+  }
+};
+
+template <>
+struct WeightConvert<SignedLogWeight, SignedLog64Weight> {
+  SignedLog64Weight operator()(const SignedLogWeight &weight) const {
+    return SignedLog64Weight(weight.Value1(), weight.Value2().Value());
+  }
+};
+
+// This function object returns SignedLogWeightTpl<T>'s that are random integers
+// chosen from [0, num_random_weights) times a random sign. This is intended
+// primarily for testing.
+template <class T>
+class WeightGenerate<SignedLogWeightTpl<T>> {
+ public:
+  using Weight = SignedLogWeightTpl<T>;
+  using X1 = typename Weight::X1;
+  using X2 = typename Weight::X2;
+
+  explicit WeightGenerate(bool allow_zero = true,
+                          size_t num_random_weights = kNumRandomWeights)
+    : allow_zero_(allow_zero), num_random_weights_(num_random_weights) {}
+
+  Weight operator()() const {
+    static const X1 negative_one(-1.0);
+    static const X1 positive_one(+1.0);
+    const int m = rand() % 2;                                    // NOLINT
+    const int n = rand() % (num_random_weights_ + allow_zero_);  // NOLINT
+    return Weight((m == 0) ? negative_one : positive_one,
+                  (allow_zero_ && n == num_random_weights_) ?
+                   X2::Zero() : X2(n));
+  }
+
+ private:
+  // Permits Zero() and zero divisors.
+  const bool allow_zero_;
+  // Number of alternative random weights.
+  const size_t num_random_weights_;
+};
+
+}  // namespace fst
+
+#endif  // FST_SIGNED_LOG_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/sparse-power-weight.h b/projects/llm_framework/include/fst/sparse-power-weight.h
new file mode 100644
index 00000000..b7f20981
--- /dev/null
+++ b/projects/llm_framework/include/fst/sparse-power-weight.h
@@ -0,0 +1,209 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Cartesian power weight semiring operation definitions, using
+// SparseTupleWeight as underlying representation.
+
+#ifndef FST_SPARSE_POWER_WEIGHT_H_
+#define FST_SPARSE_POWER_WEIGHT_H_
+
+#include <climits>
+#include <string>
+
+#include <fst/sparse-tuple-weight.h>
+#include <fst/weight.h>
+
+
+namespace fst {
+
+// Sparse cartesian power semiring: W ^ n
+//
+// Forms:
+//
+//  - a left semimodule when W is a left semiring,
+//  - a right semimodule when W is a right semiring,
+//  - a bisemimodule when W is a semiring,
+//    the free semimodule of rank n over W
+//
+// The Times operation is overloaded to provide the left and right scalar
+// products.
+//
+// K is the key value type. kNoKey (-1) is reserved for internal use
+template <class W, class K = int>
+class SparsePowerWeight : public SparseTupleWeight<W, K> {
+ public:
+  using ReverseWeight = SparsePowerWeight<typename W::ReverseWeight, K>;
+
+  SparsePowerWeight() {}
+
+  explicit SparsePowerWeight(const SparseTupleWeight<W, K> &weight)
+      : SparseTupleWeight<W, K>(weight) {}
+
+  template <class Iterator>
+  SparsePowerWeight(Iterator begin, Iterator end)
+      : SparseTupleWeight<W, K>(begin, end) {}
+
+  // Initialize component `key` to `weight`, with `default_weight` for all
+  // other components.
+  SparsePowerWeight(const K &key, const W &weight,
+                    const W &default_weight = W::Zero())
+      : SparseTupleWeight<W, K>(key, weight, default_weight) {}
+
+  static const SparsePowerWeight &Zero() {
+    static const SparsePowerWeight zero(SparseTupleWeight<W, K>::Zero());
+    return zero;
+  }
+
+  static const SparsePowerWeight &One() {
+    static const SparsePowerWeight one(SparseTupleWeight<W, K>::One());
+    return one;
+  }
+
+  static const SparsePowerWeight &NoWeight() {
+    static const SparsePowerWeight no_weight(
+        SparseTupleWeight<W, K>::NoWeight());
+    return no_weight;
+  }
+
+  // Overide this: Overwrite the Type method to reflect the key type if using
+  // a non-default key type.
+  static const string &Type() {
+    static const string *const type = [] {
+      string type = W::Type() + "_^n";
+      if (sizeof(K) != sizeof(uint32)) {
+        type += "_" + std::to_string(CHAR_BIT * sizeof(K));
+      }
+      return new string(type);
+    }();
+    return *type;
+  }
+
+  static constexpr uint64 Properties() {
+    return W::Properties() &
+           (kLeftSemiring | kRightSemiring | kCommutative | kIdempotent);
+  }
+
+  SparsePowerWeight Quantize(float delta = kDelta) const {
+    return SparsePowerWeight(SparseTupleWeight<W, K>::Quantize(delta));
+  }
+
+  ReverseWeight Reverse() const {
+    return ReverseWeight(SparseTupleWeight<W, K>::Reverse());
+  }
+};
+
+template <class W, class K, class M>
+inline SparsePowerWeight<W, K> SparsePowerWeightMap(
+    const SparsePowerWeight<W, K> &w1,
+    const SparsePowerWeight<W, K> &w2,
+    const M &operator_mapper) {
+  SparsePowerWeight<W, K> result;
+  SparseTupleWeightMap(&result, w1, w2, operator_mapper);
+  return result;
+}
+
+// Semimodule plus operation.
+template <class W, class K>
+inline SparsePowerWeight<W, K> Plus(const SparsePowerWeight<W, K> &w1,
+                                    const SparsePowerWeight<W, K> &w2) {
+  return SparsePowerWeightMap(w1, w2, [](const K &k, const W &v1, const W &v2) {
+    return Plus(v1, v2);
+  });
+}
+
+// Semimodule times operation.
+template <class W, class K>
+inline SparsePowerWeight<W, K> Times(const SparsePowerWeight<W, K> &w1,
+                                     const SparsePowerWeight<W, K> &w2) {
+  return SparsePowerWeightMap(w1, w2, [](const K &k, const W &v1, const W &v2) {
+    return Times(v1, v2);
+  });
+}
+
+// Semimodule divide operation.
+template <class W, class K>
+inline SparsePowerWeight<W, K> Divide(const SparsePowerWeight<W, K> &w1,
+                                      const SparsePowerWeight<W, K> &w2,
+                                      DivideType type = DIVIDE_ANY) {
+  return SparsePowerWeightMap(w1, w2,
+                              [type](const K &k, const W &v1, const W &v2) {
+                                return Divide(v1, v2, type);
+                              });
+}
+
+// Semimodule dot product operation.
+template <class W, class K>
+inline const W &DotProduct(const SparsePowerWeight<W, K> &w1,
+                           const SparsePowerWeight<W, K> &w2) {
+  const SparsePowerWeight<W, K> product = Times(w1, w2);
+  W result(W::Zero());
+  for (SparseTupleWeightIterator<W, K> it(product); !it.Done(); it.Next()) {
+    result = Plus(result, it.Value().second);
+  }
+  return result;
+}
+
+template <class W, class K>
+inline bool ApproxEqual(const SparsePowerWeight<W, K> &w1,
+                        const SparsePowerWeight<W, K> &w2,
+                        float delta = kDelta) {
+  auto result = SparsePowerWeightMap(
+      w1, w2, [delta](const K &k, const W &v1, const W &v2) {
+        return ApproxEqual(v1, v2, delta) ? W::One() : W::Zero();
+      });
+  return result == SparsePowerWeight<W, K>::One();
+}
+
+template <class W, class K>
+inline SparsePowerWeight<W, K> Times(const W &k,
+                                     const SparsePowerWeight<W, K> &w2) {
+  const SparseTupleWeight<W, K> t2(k);
+  const SparsePowerWeight<W, K> w1(t2);
+  return Times(w1, w2);
+}
+
+template <class W, class K>
+inline SparsePowerWeight<W, K> Times(const SparsePowerWeight<W, K> &w1,
+                                     const W &k) {
+  const SparseTupleWeight<W, K> t2(k);
+  const SparsePowerWeight<W, K> w2(t2);
+  return Times(w1, w2);
+}
+
+template <class W, class K>
+inline SparsePowerWeight<W, K> Divide(const SparsePowerWeight<W, K> &w1,
+                                      const W &k,
+                                      DivideType divide_type = DIVIDE_ANY) {
+  const SparseTupleWeight<W, K> t2(k);
+  const SparsePowerWeight<W, K> w2(t2);
+  return Divide(w1, w2, divide_type);
+}
+
+// This function object generates weights over the Cartesian power of rank
+// n over the underlying weight. This is intended primarily for testing.
+template <class W, class K>
+class WeightGenerate<SparsePowerWeight<W, K>> {
+ public:
+  using Weight = SparsePowerWeight<W, K>;
+  using Generate = WeightGenerate<W>;
+
+  explicit WeightGenerate(bool allow_zero = true,
+                          size_t sparse_power_rank = 3)
+      : generate_(allow_zero), sparse_power_rank_(sparse_power_rank) {}
+
+  Weight operator()() const {
+    Weight weight;
+    for (size_t i = 1; i <= sparse_power_rank_; ++i) {
+      weight.PushBack(i, generate_(), true);
+    }
+    return weight;
+  }
+
+ private:
+  const Generate generate_;
+  const size_t sparse_power_rank_;
+};
+
+}  // namespace fst
+
+#endif  // FST_SPARSE_POWER_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/sparse-tuple-weight.h b/projects/llm_framework/include/fst/sparse-tuple-weight.h
new file mode 100644
index 00000000..68c9c663
--- /dev/null
+++ b/projects/llm_framework/include/fst/sparse-tuple-weight.h
@@ -0,0 +1,422 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Sparse version of tuple-weight, based on tuple-weight.h.
+// Internally stores sparse key, value pairs in linked list. The default value
+// element is the assumed value of unset keys. Internal singleton
+// implementation that stores first key, value pair as a initialized member
+// variable to avoid unnecessary allocation on heap. Use
+// SparseTupleWeightIterator to iterate through the key,value pairs. Note:
+// this does NOT iterate through the default value.
+//
+// Sparse tuple weight set operation definitions.
+
+#ifndef FST_SPARSE_TUPLE_WEIGHT_H_
+#define FST_SPARSE_TUPLE_WEIGHT_H_
+
+#include <algorithm>
+#include <list>
+#include <stack>
+#include <string>
+#include <unordered_map>
+#include <utility>
+
+
+#include <fst/weight.h>
+
+
+namespace fst {
+
+template <class W, class K>
+class SparseTupleWeightIterator;
+
+// Arbitrary dimension tuple weight, stored as a sorted linked-list.
+// W is any weight class, and K is the key value type. kNoKey (-1) is reserved
+// for internal use.
+template <class W, class K = int>
+class SparseTupleWeight {
+ public:
+  using ReverseWeight = SparseTupleWeight<typename W::ReverseWeight, K>;
+
+  using Iterator = SparseTupleWeightIterator<W, K>;
+  using Pair = std::pair<K, W>;
+  using Weight = W;
+  using Index = K;
+
+  constexpr static K kNoKey = -1;
+
+  SparseTupleWeight() { Init(); }
+
+  ~SparseTupleWeight() noexcept = default;
+
+  template <class Iterator>
+  SparseTupleWeight(Iterator begin, Iterator end) {
+    Init();
+    // Assumes input iterator is sorted.
+    for (auto it = begin; it != end; ++it) PushBack(*it);
+  }
+
+  // Initialize component `key` to `weight`, with `default_weight` for all
+  // other components.
+  SparseTupleWeight(const K &key, const W &weight, const W &default_weight)
+      : default_(default_weight),
+        first_(weight == default_weight ? kNoKey : key, weight) {}
+
+  explicit SparseTupleWeight(const W &weight) { Init(weight); }
+
+  SparseTupleWeight(const SparseTupleWeight &weight) {
+    Init(weight.DefaultValue());
+    SetDefaultValue(weight.DefaultValue());
+    for (Iterator it(weight); !it.Done(); it.Next()) {
+      PushBack(it.Value());
+    }
+  }
+
+  SparseTupleWeight(SparseTupleWeight &&weight) noexcept
+      // Don't move the default, so weight.default_ is still valid.
+      : default_(weight.default_),  // NOLINT
+        first_(std::move(weight.first_)),
+        rest_(std::move(weight.rest_)) {
+    // move leaves the source in a valid but unspecified state.
+    // Make sure the source weight is empty.
+    weight.first_ = Pair(kNoKey, W::NoWeight());
+    weight.rest_.clear();
+  }
+
+  static const SparseTupleWeight &Zero() {
+    static const SparseTupleWeight zero(W::Zero());
+    return zero;
+  }
+
+  static const SparseTupleWeight &One() {
+    static const SparseTupleWeight one(W::One());
+    return one;
+  }
+
+  static const SparseTupleWeight &NoWeight() {
+    static const SparseTupleWeight no_weight(W::NoWeight());
+    return no_weight;
+  }
+
+  std::istream &Read(std::istream &strm) {
+    ReadType(strm, &default_);
+    ReadType(strm, &first_);
+    return ReadType(strm, &rest_);
+  }
+
+  std::ostream &Write(std::ostream &strm) const {
+    WriteType(strm, default_);
+    WriteType(strm, first_);
+    return WriteType(strm, rest_);
+  }
+
+  SparseTupleWeight &operator=(const SparseTupleWeight &weight) {
+    if (this == &weight) return *this;  // Checks for identity.
+    Init(weight.DefaultValue());
+    for (Iterator it(weight); !it.Done(); it.Next()) {
+      PushBack(it.Value());
+    }
+    return *this;
+  }
+
+  SparseTupleWeight &operator=(SparseTupleWeight &&weight) noexcept {
+    if (this == &weight) return *this;  // Checks for identity.
+    // Don't move the default, so weight.default_ is still valid.
+    default_ = weight.default_;
+    first_ = std::move(weight.first_);
+    rest_ = std::move(weight.rest_);
+    // move leaves the source in a valid but unspecified state.
+    // Make sure the source weight is empty.
+    weight.first_ = Pair(kNoKey, W::NoWeight());
+    weight.rest_.clear();
+    return *this;
+  }
+
+  bool Member() const {
+    if (!DefaultValue().Member()) return false;
+    for (Iterator it(*this); !it.Done(); it.Next()) {
+      if (!it.Value().second.Member()) return false;
+    }
+    return true;
+  }
+
+  // Assumes H() function exists for the hash of the key value.
+  size_t Hash() const {
+    size_t h = 0;
+    static const std::hash<K> H;
+    for (Iterator it(*this); !it.Done(); it.Next()) {
+      h = 5 * h + H(it.Value().first);
+      h = 13 * h + it.Value().second.Hash();
+    }
+    return h;
+  }
+
+  SparseTupleWeight Quantize(float delta = kDelta) const {
+    SparseTupleWeight weight;
+    for (Iterator it(*this); !it.Done(); it.Next()) {
+      weight.PushBack(it.Value().first, it.Value().second.Quantize(delta));
+    }
+    return weight;
+  }
+
+  ReverseWeight Reverse() const {
+    ReverseWeight weight(DefaultValue().Reverse());
+    for (Iterator it(*this); !it.Done(); it.Next()) {
+      weight.PushBack(it.Value().first, it.Value().second.Reverse());
+    }
+    return weight;
+  }
+
+  void Init(const W &default_value = W::Zero()) {
+    first_ = Pair(kNoKey, W::NoWeight());
+    // Initialized to the reserved key value.
+    default_ = default_value;
+    rest_.clear();
+  }
+
+  size_t Size() const {
+    if (first_.first == kNoKey) {
+      return 0;
+    } else {
+      return rest_.size() + 1;
+    }
+  }
+
+  inline void PushBack(const K &key, const W &weight,
+                       bool default_value_check = true) {
+    PushBack(std::make_pair(key, weight), default_value_check);
+  }
+
+  inline void PushBack(const Pair &pair, bool default_value_check = true) {
+    if (default_value_check && pair.second == default_) return;
+    if (first_.first == kNoKey) {
+      first_ = pair;
+    } else {
+      rest_.push_back(pair);
+    }
+  }
+
+  // Returns the `key`-th component, or the default value if not set.
+  const W &Value(const K &key) const {
+    // TODO(rybach): Consider binary search.
+    Iterator iter(*this);
+    for (; !iter.Done() && iter.Value().first < key; iter.Next()) continue;
+    return !iter.Done() && iter.Value().first == key ? iter.Value().second
+                                                     : DefaultValue();
+  }
+
+  void SetValue(const K &key, const W &w) {
+    if (w == DefaultValue()) {
+      ClearValue(key);
+    } else {
+      SetValueToNonDefault(key, w);
+    }
+  }
+
+  void SetDefaultValue(const W &value) { default_ = value; }
+
+  const W &DefaultValue() const { return default_; }
+
+ private:
+  void SetValueToNonDefault(const K &key, const W &w) {
+    // Don't use SparseTupleWeightIterator, since that's const.
+    if (first_.first == kNoKey) {
+      first_ = Pair(key, w);
+    } else if (key < first_.first) {
+      rest_.push_front(first_);
+      first_ = Pair(key, w);
+    } else if (key == first_.first) {
+      first_.second = w;
+    } else {
+      const auto i =
+          std::find_if(rest_.begin(), rest_.end(),
+                       [key](const Pair &p) { return p.first >= key; });
+      if (i != rest_.end() && i->first == key) {
+        i->second = w;
+      } else {
+        rest_.insert(i, Pair(key, w));
+      }
+    }
+  }
+
+  // Removes the weight value for `key`, having the effect of setting
+  // it to `DefaultValue()`.
+  void ClearValue(const K &key) {
+    if (key == first_.first) {
+      if (!rest_.empty()) {
+        first_ = rest_.front();
+        rest_.pop_front();
+      } else {
+        first_.first = kNoKey;
+      }
+    } else if (key > first_.first) {
+      const auto i =
+          std::find_if(rest_.begin(), rest_.end(),
+                       [key](const Pair &p) { return p.first >= key; });
+      if (i != rest_.end() && i->first == key) {
+        rest_.erase(i);
+      }
+    }
+  }
+
+  // Assumed default value of uninitialized keys, by default W::Zero().
+  W default_;
+
+  // Key values pairs are first stored in first_, then fill rest_ this way we
+  // can avoid dynamic allocation in the common case where the weight is a
+  // single key/value pair.
+  Pair first_;
+  std::list<Pair> rest_;
+
+  friend class SparseTupleWeightIterator<W, K>;
+};
+
+// Declare storage for kNoKey since it is passed by reference.
+template <class W, class K>
+constexpr K SparseTupleWeight<W, K>::kNoKey;
+
+template <class W, class K>
+class SparseTupleWeightIterator {
+ public:
+  using Pair = typename SparseTupleWeight<W, K>::Pair;
+  using const_iterator = typename std::list<Pair>::const_iterator;
+  using iterator = typename std::list<Pair>::iterator;
+
+  explicit SparseTupleWeightIterator(const SparseTupleWeight<W, K> &weight)
+      : first_(weight.first_),
+        rest_(weight.rest_),
+        init_(true),
+        iter_(rest_.begin()) {}
+
+  bool Done() const {
+    if (init_) {
+      return first_.first == SparseTupleWeight<W, K>::kNoKey;
+    } else {
+      return iter_ == rest_.end();
+    }
+  }
+
+  const Pair &Value() const { return init_ ? first_ : *iter_; }
+
+  void Next() {
+    if (init_) {
+      init_ = false;
+    } else {
+      ++iter_;
+    }
+  }
+
+  void Reset() {
+    init_ = true;
+    iter_ = rest_.begin();
+  }
+
+ private:
+  const Pair &first_;
+  const std::list<Pair> &rest_;
+  bool init_;  // In the initialized state?
+  const_iterator iter_;
+};
+
+// M must be callable as a function W(K, W, W).
+// K will be kNoKey when mapping the default value.
+template <class W, class K, class M>
+inline void SparseTupleWeightMap(SparseTupleWeight<W, K> *result,
+                                 const SparseTupleWeight<W, K> &w1,
+                                 const SparseTupleWeight<W, K> &w2,
+                                 const M &operator_mapper) {
+  SparseTupleWeightIterator<W, K> w1_it(w1);
+  SparseTupleWeightIterator<W, K> w2_it(w2);
+  const auto &v1_def = w1.DefaultValue();
+  const auto &v2_def = w2.DefaultValue();
+  result->SetDefaultValue(
+      operator_mapper(SparseTupleWeight<W, K>::kNoKey, v1_def, v2_def));
+  while (!w1_it.Done() || !w2_it.Done()) {
+    const auto &k1 = (w1_it.Done()) ? w2_it.Value().first : w1_it.Value().first;
+    const auto &k2 = (w2_it.Done()) ? w1_it.Value().first : w2_it.Value().first;
+    const auto &v1 = (w1_it.Done()) ? v1_def : w1_it.Value().second;
+    const auto &v2 = (w2_it.Done()) ? v2_def : w2_it.Value().second;
+    if (k1 == k2) {
+      result->PushBack(k1, operator_mapper(k1, v1, v2));
+      if (!w1_it.Done()) w1_it.Next();
+      if (!w2_it.Done()) w2_it.Next();
+    } else if (k1 < k2) {
+      result->PushBack(k1, operator_mapper(k1, v1, v2_def));
+      w1_it.Next();
+    } else {
+      result->PushBack(k2, operator_mapper(k2, v1_def, v2));
+      w2_it.Next();
+    }
+  }
+}
+
+template <class W, class K>
+inline bool operator==(const SparseTupleWeight<W, K> &w1,
+                       const SparseTupleWeight<W, K> &w2) {
+  const auto &v1_def = w1.DefaultValue();
+  const auto &v2_def = w2.DefaultValue();
+  if (v1_def != v2_def) return false;
+  SparseTupleWeightIterator<W, K> w1_it(w1);
+  SparseTupleWeightIterator<W, K> w2_it(w2);
+  while (!w1_it.Done() || !w2_it.Done()) {
+    const auto &k1 = (w1_it.Done()) ? w2_it.Value().first : w1_it.Value().first;
+    const auto &k2 = (w2_it.Done()) ? w1_it.Value().first : w2_it.Value().first;
+    const auto &v1 = (w1_it.Done()) ? v1_def : w1_it.Value().second;
+    const auto &v2 = (w2_it.Done()) ? v2_def : w2_it.Value().second;
+    if (k1 == k2) {
+      if (v1 != v2) return false;
+      if (!w1_it.Done()) w1_it.Next();
+      if (!w2_it.Done()) w2_it.Next();
+    } else if (k1 < k2) {
+      if (v1 != v2_def) return false;
+      w1_it.Next();
+    } else {
+      if (v1_def != v2) return false;
+      w2_it.Next();
+    }
+  }
+  return true;
+}
+
+template <class W, class K>
+inline bool operator!=(const SparseTupleWeight<W, K> &w1,
+                       const SparseTupleWeight<W, K> &w2) {
+  return !(w1 == w2);
+}
+
+template <class W, class K>
+inline std::ostream &operator<<(std::ostream &strm,
+                                const SparseTupleWeight<W, K> &weight) {
+  CompositeWeightWriter writer(strm);
+  writer.WriteBegin();
+  writer.WriteElement(weight.DefaultValue());
+  for (SparseTupleWeightIterator<W, K> it(weight); !it.Done(); it.Next()) {
+    writer.WriteElement(it.Value().first);
+    writer.WriteElement(it.Value().second);
+  }
+  writer.WriteEnd();
+  return strm;
+}
+
+template <class W, class K>
+inline std::istream &operator>>(std::istream &strm,
+                                SparseTupleWeight<W, K> &weight) {
+  CompositeWeightReader reader(strm);
+  reader.ReadBegin();
+  W def;
+  bool more = reader.ReadElement(&def);
+  weight.Init(def);
+  while (more) {
+    K key;
+    reader.ReadElement(&key);
+    W v;
+    more = reader.ReadElement(&v);
+    weight.PushBack(key, v);
+  }
+  reader.ReadEnd();
+  return strm;
+}
+
+}  // namespace fst
+
+#endif  // FST_SPARSE_TUPLE_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/state-map.h b/projects/llm_framework/include/fst/state-map.h
new file mode 100644
index 00000000..065f1c0b
--- /dev/null
+++ b/projects/llm_framework/include/fst/state-map.h
@@ -0,0 +1,613 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to map over/transform states e.g., sort transitions.
+//
+// Consider using when operation does not change the number of states.
+
+#ifndef FST_STATE_MAP_H_
+#define FST_STATE_MAP_H_
+
+#include <algorithm>
+#include <string>
+#include <utility>
+
+#include <fst/log.h>
+
+#include <fst/arc-map.h>
+#include <fst/cache.h>
+#include <fst/mutable-fst.h>
+
+
+namespace fst {
+
+// StateMapper Interface. The class determines how states are mapped; useful for
+// implementing operations that do not change the number of states.
+//
+// class StateMapper {
+//  public:
+//   using FromArc = A;
+//   using ToArc = B;
+//
+//   // Typical constructor.
+//   StateMapper(const Fst<A> &fst);
+//
+//   // Required copy constructor that allows updating FST argument;
+//   // pass only if relevant and changed.
+//   StateMapper(const StateMapper &mapper, const Fst<A> *fst = 0);
+//
+//   // Specifies initial state of result.
+//   B::StateId Start() const;
+//   // Specifies state's final weight in result.
+//   B::Weight Final(B::StateId state) const;
+//
+//   // These methods iterate through a state's arcs in result.
+//
+//   // Specifies state to iterate over.
+//   void SetState(B::StateId state);
+//
+//   // End of arcs?
+//   bool Done() const;
+//
+//   // Current arc.
+//   const B &Value() const;
+//
+//   // Advances to next arc (when !Done)
+//   void Next();
+//
+//   // Specifies input symbol table action the mapper requires (see above).
+//   MapSymbolsAction InputSymbolsAction() const;
+//
+//   // Specifies output symbol table action the mapper requires (see above).
+//   MapSymbolsAction OutputSymbolsAction() const;
+//
+//   // This specifies the known properties of an FST mapped by this
+//   // mapper. It takes as argument the input FST's known properties.
+//   uint64 Properties(uint64 props) const;
+// };
+//
+// We include a various state map versions below. One dimension of variation is
+// whether the mapping mutates its input, writes to a new result FST, or is an
+// on-the-fly Fst. Another dimension is how we pass the mapper. We allow passing
+// the mapper by pointer for cases that we need to change the state of the
+// user's mapper. We also include map versions that pass the mapper by value or
+// const reference when this suffices.
+
+// Maps an arc type A using a mapper function object C, passed by pointer. This
+// version modifies the input FST.
+template <class A, class C>
+void StateMap(MutableFst<A> *fst, C *mapper) {
+  if (mapper->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
+    fst->SetInputSymbols(nullptr);
+  }
+  if (mapper->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
+    fst->SetOutputSymbols(nullptr);
+  }
+  if (fst->Start() == kNoStateId) return;
+  const auto props = fst->Properties(kFstProperties, false);
+  fst->SetStart(mapper->Start());
+  for (StateIterator<Fst<A>> siter(*fst); !siter.Done(); siter.Next()) {
+    const auto state = siter.Value();
+    mapper->SetState(state);
+    fst->DeleteArcs(state);
+    for (; !mapper->Done(); mapper->Next()) {
+      fst->AddArc(state, mapper->Value());
+    }
+    fst->SetFinal(state, mapper->Final(state));
+  }
+  fst->SetProperties(mapper->Properties(props), kFstProperties);
+}
+
+// Maps an arc type A using a mapper function object C, passed by value.
+// This version modifies the input FST.
+template <class A, class C>
+void StateMap(MutableFst<A> *fst, C mapper) {
+  StateMap(fst, &mapper);
+}
+
+// Maps an arc type A to an arc type B using mapper functor C, passed by
+// pointer. This version writes to an output FST.
+template <class A, class B, class C>
+void StateMap(const Fst<A> &ifst, MutableFst<B> *ofst, C *mapper) {
+  ofst->DeleteStates();
+  if (mapper->InputSymbolsAction() == MAP_COPY_SYMBOLS) {
+    ofst->SetInputSymbols(ifst.InputSymbols());
+  } else if (mapper->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
+    ofst->SetInputSymbols(nullptr);
+  }
+  if (mapper->OutputSymbolsAction() == MAP_COPY_SYMBOLS) {
+    ofst->SetOutputSymbols(ifst.OutputSymbols());
+  } else if (mapper->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
+    ofst->SetOutputSymbols(nullptr);
+  }
+  const auto iprops = ifst.Properties(kCopyProperties, false);
+  if (ifst.Start() == kNoStateId) {
+    if (iprops & kError) ofst->SetProperties(kError, kError);
+    return;
+  }
+  // Adds all states.
+  if (ifst.Properties(kExpanded, false)) ofst->ReserveStates(CountStates(ifst));
+  for (StateIterator<Fst<A>> siter(ifst); !siter.Done(); siter.Next()) {
+    ofst->AddState();
+  }
+  ofst->SetStart(mapper->Start());
+  for (StateIterator<Fst<A>> siter(ifst); !siter.Done(); siter.Next()) {
+    const auto state = siter.Value();
+    mapper->SetState(state);
+    for (; !mapper->Done(); mapper->Next()) {
+      ofst->AddArc(state, mapper->Value());
+    }
+    ofst->SetFinal(state, mapper->Final(state));
+  }
+  const auto oprops = ofst->Properties(kFstProperties, false);
+  ofst->SetProperties(mapper->Properties(iprops) | oprops, kFstProperties);
+}
+
+// Maps an arc type A to an arc type B using mapper functor object C, passed by
+// value. This version writes to an output FST.
+template <class A, class B, class C>
+void StateMap(const Fst<A> &ifst, MutableFst<B> *ofst, C mapper) {
+  StateMap(ifst, ofst, &mapper);
+}
+
+using StateMapFstOptions = CacheOptions;
+
+template <class A, class B, class C>
+class StateMapFst;
+
+// Facade around StateIteratorBase<A> inheriting from StateIteratorBase<B>.
+template <class A, class B>
+class StateMapStateIteratorBase : public StateIteratorBase<B> {
+ public:
+  using Arc = B;
+  using StateId = typename Arc::StateId;
+
+  explicit StateMapStateIteratorBase(StateIteratorBase<A> *base)
+      : base_(base) {}
+
+  bool Done() const final { return base_->Done(); }
+
+  StateId Value() const final { return base_->Value(); }
+
+  void Next() final { base_->Next(); }
+
+  void Reset() final { base_->Reset(); }
+
+ private:
+  std::unique_ptr<StateIteratorBase<A>> base_;
+
+  StateMapStateIteratorBase() = delete;
+};
+
+namespace internal {
+
+// Implementation of delayed StateMapFst.
+template <class A, class B, class C>
+class StateMapFstImpl : public CacheImpl<B> {
+ public:
+  using Arc = B;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FstImpl<B>::SetType;
+  using FstImpl<B>::SetProperties;
+  using FstImpl<B>::SetInputSymbols;
+  using FstImpl<B>::SetOutputSymbols;
+
+  using CacheImpl<B>::PushArc;
+  using CacheImpl<B>::HasArcs;
+  using CacheImpl<B>::HasFinal;
+  using CacheImpl<B>::HasStart;
+  using CacheImpl<B>::SetArcs;
+  using CacheImpl<B>::SetFinal;
+  using CacheImpl<B>::SetStart;
+
+  friend class StateIterator<StateMapFst<A, B, C>>;
+
+  StateMapFstImpl(const Fst<A> &fst, const C &mapper,
+                  const StateMapFstOptions &opts)
+      : CacheImpl<B>(opts),
+        fst_(fst.Copy()),
+        mapper_(new C(mapper, fst_.get())),
+        own_mapper_(true) {
+    Init();
+  }
+
+  StateMapFstImpl(const Fst<A> &fst, C *mapper, const StateMapFstOptions &opts)
+      : CacheImpl<B>(opts),
+        fst_(fst.Copy()),
+        mapper_(mapper),
+        own_mapper_(false) {
+    Init();
+  }
+
+  StateMapFstImpl(const StateMapFstImpl<A, B, C> &impl)
+      : CacheImpl<B>(impl),
+        fst_(impl.fst_->Copy(true)),
+        mapper_(new C(*impl.mapper_, fst_.get())),
+        own_mapper_(true) {
+    Init();
+  }
+
+  ~StateMapFstImpl() override {
+    if (own_mapper_) delete mapper_;
+  }
+
+  StateId Start() {
+    if (!HasStart()) SetStart(mapper_->Start());
+    return CacheImpl<B>::Start();
+  }
+
+  Weight Final(StateId state) {
+    if (!HasFinal(state)) SetFinal(state, mapper_->Final(state));
+    return CacheImpl<B>::Final(state);
+  }
+
+  size_t NumArcs(StateId state) {
+    if (!HasArcs(state)) Expand(state);
+    return CacheImpl<B>::NumArcs(state);
+  }
+
+  size_t NumInputEpsilons(StateId state) {
+    if (!HasArcs(state)) Expand(state);
+    return CacheImpl<B>::NumInputEpsilons(state);
+  }
+
+  size_t NumOutputEpsilons(StateId state) {
+    if (!HasArcs(state)) Expand(state);
+    return CacheImpl<B>::NumOutputEpsilons(state);
+  }
+
+  void InitStateIterator(StateIteratorData<B> *datb) const {
+    StateIteratorData<A> data;
+    fst_->InitStateIterator(&data);
+    datb->base = data.base ? new StateMapStateIteratorBase<A, B>(data.base)
+        : nullptr;
+    datb->nstates = data.nstates;
+  }
+
+  void InitArcIterator(StateId state, ArcIteratorData<B> *data) {
+    if (!HasArcs(state)) Expand(state);
+    CacheImpl<B>::InitArcIterator(state, data);
+  }
+
+  uint64 Properties() const override { return Properties(kFstProperties); }
+
+  uint64 Properties(uint64 mask) const override {
+    if ((mask & kError) && (fst_->Properties(kError, false) ||
+                            (mapper_->Properties(0) & kError))) {
+      SetProperties(kError, kError);
+    }
+    return FstImpl<Arc>::Properties(mask);
+  }
+
+  void Expand(StateId state) {
+    // Adds exiting arcs.
+    for (mapper_->SetState(state); !mapper_->Done(); mapper_->Next()) {
+      PushArc(state, mapper_->Value());
+    }
+    SetArcs(state);
+  }
+
+  const Fst<A> *GetFst() const { return fst_.get(); }
+
+ private:
+  void Init() {
+    SetType("statemap");
+    if (mapper_->InputSymbolsAction() == MAP_COPY_SYMBOLS) {
+      SetInputSymbols(fst_->InputSymbols());
+    } else if (mapper_->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
+      SetInputSymbols(nullptr);
+    }
+    if (mapper_->OutputSymbolsAction() == MAP_COPY_SYMBOLS) {
+      SetOutputSymbols(fst_->OutputSymbols());
+    } else if (mapper_->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
+      SetOutputSymbols(nullptr);
+    }
+    const auto props = fst_->Properties(kCopyProperties, false);
+    SetProperties(mapper_->Properties(props));
+  }
+
+  std::unique_ptr<const Fst<A>> fst_;
+  C *mapper_;
+  bool own_mapper_;
+};
+
+}  // namespace internal
+
+// Maps an arc type A to an arc type B using Mapper function object
+// C. This version is a delayed FST.
+template <class A, class B, class C>
+class StateMapFst : public ImplToFst<internal::StateMapFstImpl<A, B, C>> {
+ public:
+  friend class ArcIterator<StateMapFst<A, B, C>>;
+
+  using Arc = B;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using Store = DefaultCacheStore<Arc>;
+  using State = typename Store::State;
+  using Impl = internal::StateMapFstImpl<A, B, C>;
+
+  StateMapFst(const Fst<A> &fst, const C &mapper,
+              const StateMapFstOptions &opts)
+      : ImplToFst<Impl>(std::make_shared<Impl>(fst, mapper, opts)) {}
+
+  StateMapFst(const Fst<A> &fst, C *mapper, const StateMapFstOptions &opts)
+      : ImplToFst<Impl>(std::make_shared<Impl>(fst, mapper, opts)) {}
+
+  StateMapFst(const Fst<A> &fst, const C &mapper)
+      : ImplToFst<Impl>(
+            std::make_shared<Impl>(fst, mapper, StateMapFstOptions())) {}
+
+  StateMapFst(const Fst<A> &fst, C *mapper)
+      : ImplToFst<Impl>(
+            std::make_shared<Impl>(fst, mapper, StateMapFstOptions())) {}
+
+  // See Fst<>::Copy() for doc.
+  StateMapFst(const StateMapFst<A, B, C> &fst, bool safe = false)
+      : ImplToFst<Impl>(fst, safe) {}
+
+  // Get a copy of this StateMapFst. See Fst<>::Copy() for further doc.
+  StateMapFst<A, B, C> *Copy(bool safe = false) const override {
+    return new StateMapFst<A, B, C>(*this, safe);
+  }
+
+  void InitStateIterator(StateIteratorData<B> *data) const override {
+    GetImpl()->InitStateIterator(data);
+  }
+
+  void InitArcIterator(StateId state, ArcIteratorData<B> *data) const override {
+    GetMutableImpl()->InitArcIterator(state, data);
+  }
+
+ protected:
+  using ImplToFst<Impl>::GetImpl;
+  using ImplToFst<Impl>::GetMutableImpl;
+
+ private:
+  StateMapFst &operator=(const StateMapFst &) = delete;
+};
+
+// Specialization for StateMapFst.
+template <class A, class B, class C>
+class ArcIterator<StateMapFst<A, B, C>>
+    : public CacheArcIterator<StateMapFst<A, B, C>> {
+ public:
+  using StateId = typename A::StateId;
+
+  ArcIterator(const StateMapFst<A, B, C> &fst, StateId state)
+      : CacheArcIterator<StateMapFst<A, B, C>>(fst.GetMutableImpl(), state) {
+    if (!fst.GetImpl()->HasArcs(state)) fst.GetMutableImpl()->Expand(state);
+  }
+};
+
+// Utility mappers.
+
+// Mapper that returns its input.
+template <class Arc>
+class IdentityStateMapper {
+ public:
+  using FromArc = Arc;
+  using ToArc = Arc;
+
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  explicit IdentityStateMapper(const Fst<Arc> &fst) : fst_(fst) {}
+
+  // Allows updating FST argument; pass only if changed.
+  IdentityStateMapper(const IdentityStateMapper<Arc> &mapper,
+                      const Fst<Arc> *fst = nullptr)
+      : fst_(fst ? *fst : mapper.fst_) {}
+
+  StateId Start() const { return fst_.Start(); }
+
+  Weight Final(StateId state) const { return fst_.Final(state); }
+
+  void SetState(StateId state) {
+    aiter_.reset(new ArcIterator<Fst<Arc>>(fst_, state));
+  }
+
+  bool Done() const { return aiter_->Done(); }
+
+  const Arc &Value() const { return aiter_->Value(); }
+
+  void Next() { aiter_->Next(); }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  uint64 Properties(uint64 props) const { return props; }
+
+ private:
+  const Fst<Arc> &fst_;
+  std::unique_ptr<ArcIterator<Fst<Arc>>> aiter_;
+};
+
+template <class Arc>
+class ArcSumMapper {
+ public:
+  using FromArc = Arc;
+  using ToArc = Arc;
+
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  explicit ArcSumMapper(const Fst<Arc> &fst) : fst_(fst), i_(0) {}
+
+  // Allows updating FST argument; pass only if changed.
+  ArcSumMapper(const ArcSumMapper<Arc> &mapper, const Fst<Arc> *fst = nullptr)
+      : fst_(fst ? *fst : mapper.fst_), i_(0) {}
+
+  StateId Start() const { return fst_.Start(); }
+
+  Weight Final(StateId state) const { return fst_.Final(state); }
+
+  void SetState(StateId state) {
+    i_ = 0;
+    arcs_.clear();
+    arcs_.reserve(fst_.NumArcs(state));
+    for (ArcIterator<Fst<Arc>> aiter(fst_, state); !aiter.Done();
+         aiter.Next()) {
+      arcs_.push_back(aiter.Value());
+    }
+    // First sorts the exiting arcs by input label, output label and destination
+    // state and then sums weights of arcs with the same input label, output
+    // label, and destination state.
+    std::sort(arcs_.begin(), arcs_.end(), comp_);
+    size_t narcs = 0;
+    for (const auto &arc : arcs_) {
+      if (narcs > 0 && equal_(arc, arcs_[narcs - 1])) {
+        arcs_[narcs - 1].weight = Plus(arcs_[narcs - 1].weight, arc.weight);
+      } else {
+        arcs_[narcs] = arc;
+        ++narcs;
+      }
+    }
+    arcs_.resize(narcs);
+  }
+
+  bool Done() const { return i_ >= arcs_.size(); }
+
+  const Arc &Value() const { return arcs_[i_]; }
+
+  void Next() { ++i_; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  uint64 Properties(uint64 props) const {
+    return props & kArcSortProperties & kDeleteArcsProperties &
+           kWeightInvariantProperties;
+  }
+
+ private:
+  struct Compare {
+    bool operator()(const Arc &x, const Arc &y) const {
+      if (x.ilabel < y.ilabel) return true;
+      if (x.ilabel > y.ilabel) return false;
+      if (x.olabel < y.olabel) return true;
+      if (x.olabel > y.olabel) return false;
+      if (x.nextstate < y.nextstate) return true;
+      if (x.nextstate > y.nextstate) return false;
+      return false;
+    }
+  };
+
+  struct Equal {
+    bool operator()(const Arc &x, const Arc &y) const {
+      return (x.ilabel == y.ilabel && x.olabel == y.olabel &&
+              x.nextstate == y.nextstate);
+    }
+  };
+
+  const Fst<Arc> &fst_;
+  Compare comp_;
+  Equal equal_;
+  std::vector<Arc> arcs_;
+  ssize_t i_;  // Current arc position.
+
+  ArcSumMapper &operator=(const ArcSumMapper &) = delete;
+};
+
+template <class Arc>
+class ArcUniqueMapper {
+ public:
+  using FromArc = Arc;
+  using ToArc = Arc;
+
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  explicit ArcUniqueMapper(const Fst<Arc> &fst) : fst_(fst), i_(0) {}
+
+  // Allows updating FST argument; pass only if changed.
+  ArcUniqueMapper(const ArcUniqueMapper<Arc> &mapper,
+                  const Fst<Arc> *fst = nullptr)
+      : fst_(fst ? *fst : mapper.fst_), i_(0) {}
+
+  StateId Start() const { return fst_.Start(); }
+
+  Weight Final(StateId state) const { return fst_.Final(state); }
+
+  void SetState(StateId state) {
+    i_ = 0;
+    arcs_.clear();
+    arcs_.reserve(fst_.NumArcs(state));
+    for (ArcIterator<Fst<Arc>> aiter(fst_, state); !aiter.Done();
+         aiter.Next()) {
+      arcs_.push_back(aiter.Value());
+    }
+    // First sorts the exiting arcs by input label, output label and destination
+    // state and then uniques identical arcs.
+    std::sort(arcs_.begin(), arcs_.end(), comp_);
+    arcs_.erase(std::unique(arcs_.begin(), arcs_.end(), equal_), arcs_.end());
+  }
+
+  bool Done() const { return i_ >= arcs_.size(); }
+
+  const Arc &Value() const { return arcs_[i_]; }
+
+  void Next() { ++i_; }
+
+  constexpr MapSymbolsAction InputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  constexpr MapSymbolsAction OutputSymbolsAction() const {
+    return MAP_COPY_SYMBOLS;
+  }
+
+  uint64 Properties(uint64 props) const {
+    return props & kArcSortProperties & kDeleteArcsProperties;
+  }
+
+ private:
+  struct Compare {
+    bool operator()(const Arc &x, const Arc &y) const {
+      if (x.ilabel < y.ilabel) return true;
+      if (x.ilabel > y.ilabel) return false;
+      if (x.olabel < y.olabel) return true;
+      if (x.olabel > y.olabel) return false;
+      if (x.nextstate < y.nextstate) return true;
+      if (x.nextstate > y.nextstate) return false;
+      return false;
+    }
+  };
+
+  struct Equal {
+    bool operator()(const Arc &x, const Arc &y) const {
+      return (x.ilabel == y.ilabel && x.olabel == y.olabel &&
+              x.nextstate == y.nextstate && x.weight == y.weight);
+    }
+  };
+
+  const Fst<Arc> &fst_;
+  Compare comp_;
+  Equal equal_;
+  std::vector<Arc> arcs_;
+  size_t i_;  // Current arc position.
+
+  ArcUniqueMapper &operator=(const ArcUniqueMapper &) = delete;
+};
+
+// Useful aliases when using StdArc.
+
+using StdArcSumMapper = ArcSumMapper<StdArc>;
+
+using StdArcUniqueMapper = ArcUniqueMapper<StdArc>;
+
+}  // namespace fst
+
+#endif  // FST_STATE_MAP_H_
diff --git a/projects/llm_framework/include/fst/state-reachable.h b/projects/llm_framework/include/fst/state-reachable.h
new file mode 100644
index 00000000..36b5559a
--- /dev/null
+++ b/projects/llm_framework/include/fst/state-reachable.h
@@ -0,0 +1,224 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Class to determine whether a given (final) state can be reached from some
+// other given state.
+
+#ifndef FST_STATE_REACHABLE_H_
+#define FST_STATE_REACHABLE_H_
+
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/connect.h>
+#include <fst/dfs-visit.h>
+#include <fst/fst.h>
+#include <fst/interval-set.h>
+#include <fst/vector-fst.h>
+
+
+namespace fst {
+
+// Computes the (final) states reachable from a given state in an FST. After
+// this visitor has been called, a final state f can be reached from a state
+// s iff (*isets)[s].Member(state2index[f]) is true, where (*isets[s]) is a
+// set of half-open inteval of final state indices and state2index[f] maps from
+// a final state to its index. If state2index is empty, it is filled-in with
+// suitable indices. If it is non-empty, those indices are used; in this case,
+// the final states must have out-degree 0.
+template <class Arc, class I = typename Arc::StateId, class S = IntervalSet<I>>
+class IntervalReachVisitor {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Index = I;
+  using ISet = S;
+  using Interval = typename ISet::Interval;
+
+  IntervalReachVisitor(const Fst<Arc> &fst, std::vector<S> *isets,
+                       std::vector<Index> *state2index)
+      : fst_(fst),
+        isets_(isets),
+        state2index_(state2index),
+        index_(state2index->empty() ? 1 : -1),
+        error_(false) {
+    isets_->clear();
+  }
+
+  void InitVisit(const Fst<Arc> &) { error_ = false; }
+
+  bool InitState(StateId s, StateId r) {
+    while (isets_->size() <= s) isets_->push_back(S());
+    while (state2index_->size() <= s) state2index_->push_back(-1);
+    if (fst_.Final(s) != Weight::Zero()) {
+      // Create tree interval.
+      auto *intervals = (*isets_)[s].MutableIntervals();
+      if (index_ < 0) {  // Uses state2index_ map to set index.
+        if (fst_.NumArcs(s) > 0) {
+          FSTERROR() << "IntervalReachVisitor: state2index map must be empty "
+                     << "for this FST";
+          error_ = true;
+          return false;
+        }
+        const auto index = (*state2index_)[s];
+        if (index < 0) {
+          FSTERROR() << "IntervalReachVisitor: state2index map incomplete";
+          error_ = true;
+          return false;
+        }
+        intervals->push_back(Interval(index, index + 1));
+      } else {  // Use pre-order index.
+        intervals->push_back(Interval(index_, index_ + 1));
+        (*state2index_)[s] = index_++;
+      }
+    }
+    return true;
+  }
+
+  constexpr bool TreeArc(StateId, const Arc &) const { return true; }
+
+  bool BackArc(StateId s, const Arc &arc) {
+    FSTERROR() << "IntervalReachVisitor: Cyclic input";
+    error_ = true;
+    return false;
+  }
+
+  bool ForwardOrCrossArc(StateId s, const Arc &arc) {
+    // Non-tree interval.
+    (*isets_)[s].Union((*isets_)[arc.nextstate]);
+    return true;
+  }
+
+  void FinishState(StateId s, StateId p, const Arc *) {
+    if (index_ >= 0 && fst_.Final(s) != Weight::Zero()) {
+      auto *intervals = (*isets_)[s].MutableIntervals();
+      (*intervals)[0].end = index_;  // Updates tree interval end.
+    }
+    (*isets_)[s].Normalize();
+    if (p != kNoStateId) {
+      (*isets_)[p].Union((*isets_)[s]);  // Propagates intervals to parent.
+    }
+  }
+
+  void FinishVisit() {}
+
+  bool Error() const { return error_; }
+
+ private:
+  const Fst<Arc> &fst_;
+  std::vector<ISet> *isets_;
+  std::vector<Index> *state2index_;
+  Index index_;
+  bool error_;
+};
+
+// Tests reachability of final states from a given state. To test for
+// reachability from a state s, first do SetState(s). Then a final state f can
+// be reached from state s of FST iff Reach(f) is true. The input can be cyclic,
+// but no cycle may contain a final state.
+template <class Arc, class I = typename Arc::StateId, class S = IntervalSet<I>>
+class StateReachable {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Index = I;
+  using ISet = S;
+  using Interval = typename ISet::Interval;
+
+  explicit StateReachable(const Fst<Arc> &fst) : error_(false) {
+    if (fst.Properties(kAcyclic, true)) {
+      AcyclicStateReachable(fst);
+    } else {
+      CyclicStateReachable(fst);
+    }
+  }
+
+  explicit StateReachable(const StateReachable<Arc> &reachable) {
+    FSTERROR() << "Copy constructor for state reachable class "
+               << "not implemented.";
+    error_ = true;
+  }
+
+  // Sets current state.
+  void SetState(StateId s) { s_ = s; }
+
+  // Can reach this final state from current state?
+  bool Reach(StateId s) {
+    if (s >= state2index_.size()) return false;
+    const auto i = state2index_[s];
+    if (i < 0) {
+      FSTERROR() << "StateReachable: State non-final: " << s;
+      error_ = true;
+      return false;
+    }
+    return isets_[s_].Member(i);
+  }
+
+  // Access to the state-to-index mapping. Unassigned states have index -1.
+  std::vector<Index> &State2Index() { return state2index_; }
+
+  // Access to the interval sets. These specify the reachability to the final
+  // states as intervals of the final state indices.
+  const std::vector<ISet> &IntervalSets() { return isets_; }
+
+  bool Error() const { return error_; }
+
+ private:
+  void AcyclicStateReachable(const Fst<Arc> &fst) {
+    IntervalReachVisitor<Arc, StateId, ISet> reach_visitor(fst, &isets_,
+                                                           &state2index_);
+    DfsVisit(fst, &reach_visitor);
+    if (reach_visitor.Error()) error_ = true;
+  }
+
+  void CyclicStateReachable(const Fst<Arc> &fst) {
+    // Finds state reachability on the acyclic condensation FST.
+    VectorFst<Arc> cfst;
+    std::vector<StateId> scc;
+    Condense(fst, &cfst, &scc);
+    StateReachable reachable(cfst);
+    if (reachable.Error()) {
+      error_ = true;
+      return;
+    }
+    // Gets the number of states per SCC.
+    std::vector<size_t> nscc;
+    for (StateId s = 0; s < scc.size(); ++s) {
+      const auto c = scc[s];
+      while (c >= nscc.size()) nscc.push_back(0);
+      ++nscc[c];
+    }
+    // Constructs the interval sets and state index mapping for the original
+    // FST from the condensation FST.
+    state2index_.resize(scc.size(), -1);
+    isets_.resize(scc.size());
+    for (StateId s = 0; s < scc.size(); ++s) {
+      const auto c = scc[s];
+      isets_[s] = reachable.IntervalSets()[c];
+      state2index_[s] = reachable.State2Index()[c];
+      // Checks that each final state in an input FST is not contained in a
+      // cycle (i.e., not in a non-trivial SCC).
+      if (cfst.Final(c) != Weight::Zero() && nscc[c] > 1) {
+        FSTERROR() << "StateReachable: Final state contained in a cycle";
+        error_ = true;
+        return;
+      }
+    }
+  }
+
+  StateId s_;                       // Current state.
+  std::vector<ISet> isets_;         // Interval sets per state.
+  std::vector<Index> state2index_;  // Finds index for a final state.
+  bool error_;
+
+  StateReachable &operator=(const StateReachable &) = delete;
+};
+
+}  // namespace fst
+
+#endif  // FST_STATE_REACHABLE_H_
diff --git a/projects/llm_framework/include/fst/state-table.h b/projects/llm_framework/include/fst/state-table.h
new file mode 100644
index 00000000..a5067592
--- /dev/null
+++ b/projects/llm_framework/include/fst/state-table.h
@@ -0,0 +1,494 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Classes for representing the mapping between state tuples and state IDs.
+
+#ifndef FST_STATE_TABLE_H_
+#define FST_STATE_TABLE_H_
+
+#include <deque>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/bi-table.h>
+#include <fst/expanded-fst.h>
+#include <fst/filter-state.h>
+
+
+namespace fst {
+
+// State tables determine the bijective mapping between state tuples (e.g., in
+// composition, triples of two FST states and a composition filter state) and
+// their corresponding state IDs. They are classes, templated on state tuples,
+// with the following interface:
+//
+// template <class T>
+// class StateTable {
+//  public:
+//   using StateTuple = T;
+//
+//   // Required constructors.
+//   StateTable();
+//
+//   StateTable(const StateTable &);
+//
+//   // Looks up state ID by tuple. If it doesn't exist, then add it.
+//   StateId FindState(const StateTuple &tuple);
+//
+//   // Looks up state tuple by state ID.
+//   const StateTuple<StateId> &Tuple(StateId s) const;
+//
+//   // # of stored tuples.
+//   StateId Size() const;
+// };
+//
+// A state tuple has the form:
+//
+// template <class S>
+// struct StateTuple {
+//   using StateId = S;
+//
+//   // Required constructors.
+//
+//   StateTuple();
+//
+//   StateTuple(const StateTuple &tuple);
+// };
+
+// An implementation using a hash map for the tuple to state ID mapping. The
+// state tuple T must support operator==.
+template <class T, class H>
+class HashStateTable : public HashBiTable<typename T::StateId, T, H> {
+ public:
+  using StateTuple = T;
+  using StateId = typename StateTuple::StateId;
+
+  using HashBiTable<StateId, StateTuple, H>::FindId;
+  using HashBiTable<StateId, StateTuple, H>::FindEntry;
+  using HashBiTable<StateId, StateTuple, H>::Size;
+
+  HashStateTable() : HashBiTable<StateId, StateTuple, H>() {}
+
+  explicit HashStateTable(size_t table_size)
+      : HashBiTable<StateId, StateTuple, H>(table_size) {}
+
+  StateId FindState(const StateTuple &tuple) { return FindId(tuple); }
+
+  const StateTuple &Tuple(StateId s) const { return FindEntry(s); }
+};
+
+// An implementation using a hash map for the tuple to state ID mapping. The
+// state tuple T must support operator==.
+template <class T, class H>
+class CompactHashStateTable
+    : public CompactHashBiTable<typename T::StateId, T, H> {
+ public:
+  using StateTuple = T;
+  using StateId = typename StateTuple::StateId;
+
+  using CompactHashBiTable<StateId, StateTuple, H>::FindId;
+  using CompactHashBiTable<StateId, StateTuple, H>::FindEntry;
+  using CompactHashBiTable<StateId, StateTuple, H>::Size;
+
+  CompactHashStateTable() : CompactHashBiTable<StateId, StateTuple, H>() {}
+
+  explicit CompactHashStateTable(size_t table_size)
+      : CompactHashBiTable<StateId, StateTuple, H>(table_size) {}
+
+  StateId FindState(const StateTuple &tuple) { return FindId(tuple); }
+
+  const StateTuple &Tuple(StateId s) const { return FindEntry(s); }
+};
+
+// An implementation using a vector for the tuple to state mapping. It is
+// passed a fingerprint functor that should fingerprint tuples uniquely to an
+// integer that can used as a vector index. Normally, VectorStateTable
+// constructs the fingerprint functor. Alternately, the user can pass this
+// object, in which case the table takes ownership.
+template <class T, class FP>
+class VectorStateTable : public VectorBiTable<typename T::StateId, T, FP> {
+ public:
+  using StateTuple = T;
+  using StateId = typename StateTuple::StateId;
+
+  using VectorBiTable<StateId, StateTuple, FP>::FindId;
+  using VectorBiTable<StateId, StateTuple, FP>::FindEntry;
+  using VectorBiTable<StateId, StateTuple, FP>::Size;
+  using VectorBiTable<StateId, StateTuple, FP>::Fingerprint;
+
+  explicit VectorStateTable(FP *fingerprint = nullptr, size_t table_size = 0)
+      : VectorBiTable<StateId, StateTuple, FP>(fingerprint, table_size) {}
+
+  StateId FindState(const StateTuple &tuple) { return FindId(tuple); }
+
+  const StateTuple &Tuple(StateId s) const { return FindEntry(s); }
+};
+
+// An implementation using a vector and a compact hash table. The selection
+// functor returns true for tuples to be hashed in the vector. The fingerprint
+// functor should fingerprint tuples uniquely to an integer that can be used as
+// a vector index. A hash functor is used when hashing tuples into the compact
+// hash table.
+template <class T, class Select, class FP, class H>
+class VectorHashStateTable
+    : public VectorHashBiTable<typename T::StateId, T, Select, FP, H> {
+ public:
+  using StateTuple = T;
+  using StateId = typename StateTuple::StateId;
+
+  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::FindId;
+  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::FindEntry;
+  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::Size;
+  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::Selector;
+  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::Fingerprint;
+  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::Hash;
+
+  VectorHashStateTable(Select *select, FP *fingerprint, H *hash,
+                       size_t vector_size = 0, size_t tuple_size = 0)
+      : VectorHashBiTable<StateId, StateTuple, Select, FP, H>(
+            select, fingerprint, hash, vector_size, tuple_size) {}
+
+  StateId FindState(const StateTuple &tuple) { return FindId(tuple); }
+
+  const StateTuple &Tuple(StateId s) const { return FindEntry(s); }
+};
+
+// An implementation using a hash map to map from tuples to state IDs. This
+// version permits erasing of states. The state tuple's default constructor
+// must produce a tuple that will never be seen and the table must suppor
+// operator==.
+template <class T, class H>
+class ErasableStateTable : public ErasableBiTable<typename T::StateId, T, H> {
+ public:
+  using StateTuple = T;
+  using StateId = typename StateTuple::StateId;
+
+  using ErasableBiTable<StateId, StateTuple, H>::FindId;
+  using ErasableBiTable<StateId, StateTuple, H>::FindEntry;
+  using ErasableBiTable<StateId, StateTuple, H>::Size;
+  using ErasableBiTable<StateId, StateTuple, H>::Erase;
+
+  ErasableStateTable() : ErasableBiTable<StateId, StateTuple, H>() {}
+
+  StateId FindState(const StateTuple &tuple) { return FindId(tuple); }
+
+  const StateTuple &Tuple(StateId s) const { return FindEntry(s); }
+};
+
+// The composition state table has the form:
+//
+// template <class Arc, class FilterState>
+// class ComposeStateTable {
+//  public:
+//   using StateId = typename Arc::StateId;
+//
+//   // Required constructors.
+//
+//   ComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2);
+//   ComposeStateTable(const ComposeStateTable<Arc, FilterState> &table);
+//
+//   // Looks up a state ID by tuple, adding it if doesn't exist.
+//   StateId FindState(const StateTuple &tuple);
+//
+//   // Looks up a tuple by state ID.
+//   const ComposeStateTuple<StateId> &Tuple(StateId s) const;
+//
+//   // The number of of stored tuples.
+//   StateId Size() const;
+//
+//   // Return true if error was encountered.
+//   bool Error() const;
+// };
+//
+// The following interface is used to represent the composition state.
+//
+// template <class S, class FS>
+// class CompositionStateTuple {
+//  public:
+//   using StateId = typename StateId;
+//   using FS = FilterState;
+//
+//   // Required constructors.
+//   StateTuple();
+//   StateTuple(StateId s1, StateId s2, const FilterState &fs);
+//
+//   StateId StateId1() const;
+//   StateId StateId2() const;
+//
+//   FilterState GetFilterState() const;
+//
+//   std::pair<StateId, StateId> StatePair() const;
+//
+//   size_t Hash() const;
+//
+//   friend bool operator==(const StateTuple& x, const StateTuple &y);
+// }
+//
+template <typename S, typename FS>
+class DefaultComposeStateTuple {
+ public:
+  using StateId = S;
+  using FilterState = FS;
+
+  DefaultComposeStateTuple()
+      : state_pair_(kNoStateId, kNoStateId), fs_(FilterState::NoState()) {}
+
+  DefaultComposeStateTuple(StateId s1, StateId s2, const FilterState &fs)
+      : state_pair_(s1, s2), fs_(fs) {}
+
+  StateId StateId1() const { return state_pair_.first; }
+
+  StateId StateId2() const { return state_pair_.second; }
+
+  FilterState GetFilterState() const { return fs_; }
+
+  const std::pair<StateId, StateId> &StatePair() const { return state_pair_; }
+
+  friend bool operator==(const DefaultComposeStateTuple &x,
+                         const DefaultComposeStateTuple &y) {
+    return (&x == &y) || (x.state_pair_ == y.state_pair_ && x.fs_ == y.fs_);
+  }
+
+  size_t Hash() const {
+    return static_cast<size_t>(StateId1()) +
+           static_cast<size_t>(StateId2()) * 7853u +
+           GetFilterState().Hash() * 7867u;
+  }
+
+ private:
+  std::pair<StateId, StateId> state_pair_;
+  FilterState fs_;  // State of composition filter.
+};
+
+// Specialization for TrivialFilterState that does not explicitely store the
+// filter state since it is always the unique non-blocking state.
+template <typename S>
+class DefaultComposeStateTuple<S, TrivialFilterState> {
+ public:
+  using StateId = S;
+  using FilterState = TrivialFilterState;
+
+  DefaultComposeStateTuple()
+      : state_pair_(kNoStateId, kNoStateId) {}
+
+  DefaultComposeStateTuple(StateId s1, StateId s2, const FilterState &)
+      : state_pair_(s1, s2) {}
+
+  StateId StateId1() const { return state_pair_.first; }
+
+  StateId StateId2() const { return state_pair_.second; }
+
+  FilterState GetFilterState() const { return FilterState(true); }
+
+  const std::pair<StateId, StateId> &StatePair() const { return state_pair_; }
+
+  friend bool operator==(const DefaultComposeStateTuple &x,
+                         const DefaultComposeStateTuple &y) {
+    return (&x == &y) || (x.state_pair_ == y.state_pair_);
+  }
+
+  size_t Hash() const { return StateId1() + StateId2() * 7853; }
+
+ private:
+  std::pair<StateId, StateId> state_pair_;
+};
+
+// Hashing of composition state tuples.
+template <typename T>
+class ComposeHash {
+ public:
+  size_t operator()(const T &t) const { return t.Hash(); }
+};
+
+// A HashStateTable over composition tuples.
+template <typename Arc, typename FilterState,
+          typename StateTuple =
+              DefaultComposeStateTuple<typename Arc::StateId, FilterState>,
+          typename StateTable =
+              CompactHashStateTable<StateTuple, ComposeHash<StateTuple>>>
+class GenericComposeStateTable : public StateTable {
+ public:
+  using StateId = typename Arc::StateId;
+
+  GenericComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {}
+
+  GenericComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+                           size_t table_size)
+      : StateTable(table_size) {}
+
+  constexpr bool Error() const { return false; }
+
+ private:
+  GenericComposeStateTable &operator=(const GenericComposeStateTable &table) =
+      delete;
+};
+
+//  Fingerprint for general composition tuples.
+template <typename StateTuple>
+class ComposeFingerprint {
+ public:
+  using StateId = typename StateTuple::StateId;
+
+  // Required but suboptimal constructor.
+  ComposeFingerprint() : mult1_(8192), mult2_(8192) {
+    LOG(WARNING) << "TupleFingerprint: # of FST states should be provided.";
+  }
+
+  // Constructor is provided the sizes of the input FSTs.
+  ComposeFingerprint(StateId nstates1, StateId nstates2)
+      : mult1_(nstates1), mult2_(nstates1 * nstates2) {}
+
+  size_t operator()(const StateTuple &tuple) {
+    return tuple.StateId1() + tuple.StateId2() * mult1_ +
+           tuple.GetFilterState().Hash() * mult2_;
+  }
+
+ private:
+  const ssize_t mult1_;
+  const ssize_t mult2_;
+};
+
+// Useful when the first composition state determines the tuple.
+template <typename StateTuple>
+class ComposeState1Fingerprint {
+ public:
+  size_t operator()(const StateTuple &tuple) { return tuple.StateId1(); }
+};
+
+// Useful when the second composition state determines the tuple.
+template <typename StateTuple>
+class ComposeState2Fingerprint {
+ public:
+  size_t operator()(const StateTuple &tuple) { return tuple.StateId2(); }
+};
+
+// A VectorStateTable over composition tuples. This can be used when the
+// product of number of states in FST1 and FST2 (and the composition filter
+// state hash) is manageable. If the FSTs are not expanded FSTs, they will
+// first have their states counted.
+template <typename Arc, typename StateTuple>
+class ProductComposeStateTable
+    : public VectorStateTable<StateTuple, ComposeFingerprint<StateTuple>> {
+ public:
+  using StateId = typename Arc::StateId;
+  using StateTable =
+      VectorStateTable<StateTuple, ComposeFingerprint<StateTuple>>;
+
+  ProductComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+                           size_t table_size = 0)
+      : StateTable(new ComposeFingerprint<StateTuple>(CountStates(fst1),
+                                                      CountStates(fst2)),
+                   table_size) {}
+
+  ProductComposeStateTable(
+      const ProductComposeStateTable<Arc, StateTuple> &table)
+      : StateTable(new ComposeFingerprint<StateTuple>(table.Fingerprint())) {}
+
+  constexpr bool Error() const { return false; }
+
+ private:
+  ProductComposeStateTable &operator=(const ProductComposeStateTable &table) =
+      delete;
+};
+
+// A vector-backed table over composition tuples which can be used when the
+// first FST is a string (i.e., satisfies kString property) and the second is
+// deterministic and epsilon-free. It should be used with a composition filter
+// that creates at most one filter state per tuple under these conditions (e.g.,
+// SequenceComposeFilter or MatchComposeFilter).
+template <typename Arc, typename StateTuple>
+class StringDetComposeStateTable
+    : public VectorStateTable<StateTuple,
+                              ComposeState1Fingerprint<StateTuple>> {
+ public:
+  using StateId = typename Arc::StateId;
+  using StateTable =
+      VectorStateTable<StateTuple, ComposeState1Fingerprint<StateTuple>>;
+
+  StringDetComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2)
+      : error_(false) {
+    static constexpr auto props2 = kIDeterministic | kNoIEpsilons;
+    if (fst1.Properties(kString, true) != kString) {
+      FSTERROR() << "StringDetComposeStateTable: 1st FST is not a string";
+      error_ = true;
+    } else if (fst2.Properties(props2, true) != props2) {
+      FSTERROR() << "StringDetComposeStateTable: 2nd FST is not deterministic "
+                    "and epsilon-free";
+      error_ = true;
+    }
+  }
+
+  StringDetComposeStateTable(
+      const StringDetComposeStateTable<Arc, StateTuple> &table)
+      : StateTable(table), error_(table.error_) {}
+
+  bool Error() const { return error_; }
+
+ private:
+  bool error_;
+
+  StringDetComposeStateTable &operator=(const StringDetComposeStateTable &) =
+      delete;
+};
+
+// A vector-backed table over composition tuples which can be used when the
+// first FST is deterministic and epsilon-free and the second is a string (i.e.,
+// satisfies kString). It should be used with a composition filter that creates
+// at most one filter state per tuple under these conditions (e.g.,
+// SequenceComposeFilter or MatchComposeFilter).
+template <typename Arc, typename StateTuple>
+class DetStringComposeStateTable
+    : public VectorStateTable<StateTuple,
+                              ComposeState2Fingerprint<StateTuple>> {
+ public:
+  using StateId = typename Arc::StateId;
+  using StateTable =
+      VectorStateTable<StateTuple, ComposeState2Fingerprint<StateTuple>>;
+
+  DetStringComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2)
+      : error_(false) {
+    static constexpr auto props = kODeterministic | kNoOEpsilons;
+    if (fst1.Properties(props, true) != props) {
+      FSTERROR() << "StringDetComposeStateTable: 1st FST is not "
+                 << "input-deterministic and epsilon-free";
+      error_ = true;
+    } else if (fst2.Properties(kString, true) != kString) {
+      FSTERROR() << "DetStringComposeStateTable: 2nd FST is not a string";
+      error_ = true;
+    }
+  }
+
+  DetStringComposeStateTable(
+      const DetStringComposeStateTable<Arc, StateTuple> &table)
+      : StateTable(table), error_(table.error_) {}
+
+  bool Error() const { return error_; }
+
+ private:
+  bool error_;
+
+  DetStringComposeStateTable &operator=(const DetStringComposeStateTable &) =
+      delete;
+};
+
+// An erasable table over composition tuples. The Erase(StateId) method can be
+// called if the user either is sure that composition will never return to that
+// tuple or doesn't care that if it does, it is assigned a new state ID.
+template <typename Arc, typename StateTuple>
+class ErasableComposeStateTable
+    : public ErasableStateTable<StateTuple, ComposeHash<StateTuple>> {
+ public:
+  ErasableComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {}
+
+  constexpr bool Error() const { return false; }
+
+ private:
+  ErasableComposeStateTable &operator=(const ErasableComposeStateTable &table) =
+      delete;
+};
+
+}  // namespace fst
+
+#endif  // FST_STATE_TABLE_H_
diff --git a/projects/llm_framework/include/fst/statesort.h b/projects/llm_framework/include/fst/statesort.h
new file mode 100644
index 00000000..346c7d32
--- /dev/null
+++ b/projects/llm_framework/include/fst/statesort.h
@@ -0,0 +1,74 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Function to sort states of an FST.
+
+#ifndef FST_STATESORT_H_
+#define FST_STATESORT_H_
+
+#include <algorithm>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/mutable-fst.h>
+
+
+namespace fst {
+
+// Sorts the input states of an FST. order[i] gives the the state ID after
+// sorting that corresponds to the state ID i before sorting; it must
+// therefore be a permutation of the input FST's states ID sequence.
+template <class Arc>
+void StateSort(MutableFst<Arc> *fst,
+               const std::vector<typename Arc::StateId> &order) {
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  if (order.size() != fst->NumStates()) {
+    FSTERROR() << "StateSort: Bad order vector size: " << order.size();
+    fst->SetProperties(kError, kError);
+    return;
+  }
+  if (fst->Start() == kNoStateId) return;
+  const auto props = fst->Properties(kStateSortProperties, false);
+  std::vector<bool> done(order.size(), false);
+  std::vector<Arc> arcsa;
+  std::vector<Arc> arcsb;
+  fst->SetStart(order[fst->Start()]);
+  for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
+       siter.Next()) {
+    auto s1 = siter.Value();
+    StateId s2;
+    if (done[s1]) continue;
+    auto final1 = fst->Final(s1);
+    auto final2 = Weight::Zero();
+    arcsa.clear();
+    for (ArcIterator<MutableFst<Arc>> aiter(*fst, s1); !aiter.Done();
+         aiter.Next()) {
+      arcsa.push_back(aiter.Value());
+    }
+    for (; !done[s1]; s1 = s2, final1 = final2, std::swap(arcsa, arcsb)) {
+      s2 = order[s1];
+      if (!done[s2]) {
+        final2 = fst->Final(s2);
+        arcsb.clear();
+        for (ArcIterator<MutableFst<Arc>> aiter(*fst, s2); !aiter.Done();
+             aiter.Next()) {
+          arcsb.push_back(aiter.Value());
+        }
+      }
+      fst->SetFinal(s2, final1);
+      fst->DeleteArcs(s2);
+      for (auto arc : arcsa) {  // Copy intended.
+        arc.nextstate = order[arc.nextstate];
+        fst->AddArc(s2, arc);
+      }
+      done[s1] = true;
+    }
+  }
+  fst->SetProperties(props, kFstProperties);
+}
+
+}  // namespace fst
+
+#endif  // FST_STATESORT_H_
diff --git a/projects/llm_framework/include/fst/string-weight.h b/projects/llm_framework/include/fst/string-weight.h
new file mode 100644
index 00000000..fb3e70f1
--- /dev/null
+++ b/projects/llm_framework/include/fst/string-weight.h
@@ -0,0 +1,807 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// String weight set and associated semiring operation definitions.
+
+#ifndef FST_STRING_WEIGHT_H_
+#define FST_STRING_WEIGHT_H_
+
+#include <cstdlib>
+
+#include <list>
+#include <string>
+#include <vector>
+
+#include <fst/product-weight.h>
+#include <fst/union-weight.h>
+#include <fst/weight.h>
+
+
+namespace fst {
+
+constexpr int kStringInfinity = -1;     // Label for the infinite string.
+constexpr int kStringBad = -2;          // Label for a non-string.
+constexpr char kStringSeparator = '_';  // Label separator in strings.
+
+// Determines whether to use left or right string semiring.  Includes a
+// 'restricted' version that signals an error if proper prefixes/suffixes
+// would otherwise be returned by Plus, useful with various
+// algorithms that require functional transducer input with the
+// string semirings.
+enum StringType { STRING_LEFT = 0, STRING_RIGHT = 1, STRING_RESTRICT = 2 };
+
+constexpr StringType ReverseStringType(StringType s) {
+  return s == STRING_LEFT ? STRING_RIGHT
+                          : (s == STRING_RIGHT ? STRING_LEFT : STRING_RESTRICT);
+}
+
+template <class>
+class StringWeightIterator;
+template <class>
+class StringWeightReverseIterator;
+
+// String semiring: (longest_common_prefix/suffix, ., Infinity, Epsilon)
+template <typename Label_, StringType S = STRING_LEFT>
+class StringWeight {
+ public:
+  using Label = Label_;
+  using ReverseWeight = StringWeight<Label, ReverseStringType(S)>;
+  using Iterator = StringWeightIterator<StringWeight>;
+  using ReverseIterator = StringWeightReverseIterator<StringWeight>;
+
+  friend class StringWeightIterator<StringWeight>;
+  friend class StringWeightReverseIterator<StringWeight>;
+
+  StringWeight() {}
+
+  template <typename Iterator>
+  StringWeight(const Iterator &begin, const Iterator &end) {
+    for (auto iter = begin; iter != end; ++iter) PushBack(*iter);
+  }
+
+  explicit StringWeight(Label label) { PushBack(label); }
+
+  static const StringWeight &Zero() {
+    static const auto *const zero = new StringWeight(Label(kStringInfinity));
+    return *zero;
+  }
+
+  static const StringWeight &One() {
+    static const auto *const one = new StringWeight();
+    return *one;
+  }
+
+  static const StringWeight &NoWeight() {
+    static const auto *const no_weight = new StringWeight(Label(kStringBad));
+    return *no_weight;
+  }
+
+  static const string &Type() {
+    static const string *const type = new string(
+        S == STRING_LEFT
+            ? "left_string"
+            : (S == STRING_RIGHT ? "right_string" : "restricted_string"));
+    return *type;
+  }
+
+  bool Member() const;
+
+  std::istream &Read(std::istream &strm);
+
+  std::ostream &Write(std::ostream &strm) const;
+
+  size_t Hash() const;
+
+  StringWeight Quantize(float delta = kDelta) const { return *this; }
+
+  ReverseWeight Reverse() const;
+
+  static constexpr uint64 Properties() {
+    return kIdempotent |
+           (S == STRING_LEFT ? kLeftSemiring
+                             : (S == STRING_RIGHT
+                                    ? kRightSemiring
+                                    : /* S == STRING_RESTRICT */ kLeftSemiring |
+                                          kRightSemiring));
+  }
+
+  // These operations combined with the StringWeightIterator and
+  // StringWeightReverseIterator provide the access and mutation of the string
+  // internal elements.
+
+  // Clear existing StringWeight.
+  void Clear() {
+    first_ = 0;
+    rest_.clear();
+  }
+
+  size_t Size() const { return first_ ? rest_.size() + 1 : 0; }
+
+  void PushFront(Label label) {
+    if (first_) rest_.push_front(first_);
+    first_ = label;
+  }
+
+  void PushBack(Label label) {
+    if (!first_) {
+      first_ = label;
+    } else {
+      rest_.push_back(label);
+    }
+  }
+
+ private:
+  Label first_ = 0;        // First label in string (0 if empty).
+  std::list<Label> rest_;  // Remaining labels in string.
+};
+
+// Traverses string in forward direction.
+template <class StringWeight_>
+class StringWeightIterator {
+ public:
+  using Weight = StringWeight_;
+  using Label = typename Weight::Label;
+
+  explicit StringWeightIterator(const Weight &w)
+      : first_(w.first_), rest_(w.rest_), init_(true), iter_(rest_.begin()) {}
+
+  bool Done() const {
+    if (init_) {
+      return first_ == 0;
+    } else {
+      return iter_ == rest_.end();
+    }
+  }
+
+  const Label &Value() const { return init_ ? first_ : *iter_; }
+
+  void Next() {
+    if (init_) {
+      init_ = false;
+    } else {
+      ++iter_;
+    }
+  }
+
+  void Reset() {
+    init_ = true;
+    iter_ = rest_.begin();
+  }
+
+ private:
+  const Label &first_;
+  const decltype(Weight::rest_) &rest_;
+  bool init_;  // In the initialized state?
+  typename std::remove_reference<decltype (Weight::rest_)>::type::const_iterator iter_;
+};
+
+// Traverses string in backward direction.
+template <class StringWeight_>
+class StringWeightReverseIterator {
+ public:
+  using Weight = StringWeight_;
+  using Label = typename Weight::Label;
+
+  explicit StringWeightReverseIterator(const Weight &w)
+      : first_(w.first_),
+        rest_(w.rest_),
+        fin_(first_ == Label()),
+        iter_(rest_.rbegin()) {}
+
+  bool Done() const { return fin_; }
+
+  const Label &Value() const { return iter_ == rest_.rend() ? first_ : *iter_; }
+
+  void Next() {
+    if (iter_ == rest_.rend()) {
+      fin_ = true;
+    } else {
+      ++iter_;
+    }
+  }
+
+  void Reset() {
+    fin_ = false;
+    iter_ = rest_.rbegin();
+  }
+
+ private:
+  const Label &first_;
+  const decltype(Weight::rest_) &rest_;
+  bool fin_;  // In the final state?
+  typename std::remove_reference<decltype (Weight::rest_)>::type::const_reverse_iterator iter_;
+};
+
+// StringWeight member functions follow that require
+// StringWeightIterator or StringWeightReverseIterator.
+
+template <typename Label, StringType S>
+inline std::istream &StringWeight<Label, S>::Read(std::istream &strm) {
+  Clear();
+  int32 size;
+  ReadType(strm, &size);
+  for (int32 i = 0; i < size; ++i) {
+    Label label;
+    ReadType(strm, &label);
+    PushBack(label);
+  }
+  return strm;
+}
+
+template <typename Label, StringType S>
+inline std::ostream &StringWeight<Label, S>::Write(std::ostream &strm) const {
+  const int32 size = Size();
+  WriteType(strm, size);
+  for (Iterator iter(*this); !iter.Done(); iter.Next()) {
+    WriteType(strm, iter.Value());
+  }
+  return strm;
+}
+
+template <typename Label, StringType S>
+inline bool StringWeight<Label, S>::Member() const {
+  Iterator iter(*this);
+  return iter.Value() != Label(kStringBad);
+}
+
+template <typename Label, StringType S>
+inline typename StringWeight<Label, S>::ReverseWeight
+StringWeight<Label, S>::Reverse() const {
+  ReverseWeight rweight;
+  for (Iterator iter(*this); !iter.Done(); iter.Next()) {
+    rweight.PushFront(iter.Value());
+  }
+  return rweight;
+}
+
+template <typename Label, StringType S>
+inline size_t StringWeight<Label, S>::Hash() const {
+  size_t h = 0;
+  for (Iterator iter(*this); !iter.Done(); iter.Next()) {
+    h ^= h << 1 ^ iter.Value();
+  }
+  return h;
+}
+
+template <typename Label, StringType S>
+inline bool operator==(const StringWeight<Label, S> &w1,
+                       const StringWeight<Label, S> &w2) {
+  if (w1.Size() != w2.Size()) return false;
+  using Iterator = typename StringWeight<Label, S>::Iterator;
+  Iterator iter1(w1);
+  Iterator iter2(w2);
+  for (; !iter1.Done(); iter1.Next(), iter2.Next()) {
+    if (iter1.Value() != iter2.Value()) return false;
+  }
+  return true;
+}
+
+template <typename Label, StringType S>
+inline bool operator!=(const StringWeight<Label, S> &w1,
+                       const StringWeight<Label, S> &w2) {
+  return !(w1 == w2);
+}
+
+template <typename Label, StringType S>
+inline bool ApproxEqual(const StringWeight<Label, S> &w1,
+                        const StringWeight<Label, S> &w2,
+                        float delta = kDelta) {
+  return w1 == w2;
+}
+
+template <typename Label, StringType S>
+inline std::ostream &operator<<(std::ostream &strm,
+                                const StringWeight<Label, S> &weight) {
+  typename StringWeight<Label, S>::Iterator iter(weight);
+  if (iter.Done()) {
+    return strm << "Epsilon";
+  } else if (iter.Value() == Label(kStringInfinity)) {
+    return strm << "Infinity";
+  } else if (iter.Value() == Label(kStringBad)) {
+    return strm << "BadString";
+  } else {
+    for (size_t i = 0; !iter.Done(); ++i, iter.Next()) {
+      if (i > 0) strm << kStringSeparator;
+      strm << iter.Value();
+    }
+  }
+  return strm;
+}
+
+template <typename Label, StringType S>
+inline std::istream &operator>>(std::istream &strm,
+                                StringWeight<Label, S> &weight) {
+  string str;
+  strm >> str;
+  using Weight = StringWeight<Label, S>;
+  if (str == "Infinity") {
+    weight = Weight::Zero();
+  } else if (str == "Epsilon") {
+    weight = Weight::One();
+  } else {
+    weight.Clear();
+    char *p = nullptr;
+    for (const char *cs = str.c_str(); !p || *p != '\0'; cs = p + 1) {
+      const Label label = strtoll(cs, &p, 10);
+      if (p == cs || (*p != 0 && *p != kStringSeparator)) {
+        strm.clear(std::ios::badbit);
+        break;
+      }
+      weight.PushBack(label);
+    }
+  }
+  return strm;
+}
+
+// Default is for the restricted string semiring. String equality is required
+// (for non-Zero() input). The restriction is used (e.g., in determinization)
+// to ensure the input is functional.
+template <typename Label, StringType S>
+inline StringWeight<Label, S> Plus(const StringWeight<Label, S> &w1,
+                                   const StringWeight<Label, S> &w2) {
+  using Weight = StringWeight<Label, S>;
+  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
+  if (w1 == Weight::Zero()) return w2;
+  if (w2 == Weight::Zero()) return w1;
+  if (w1 != w2) {
+    FSTERROR() << "StringWeight::Plus: Unequal arguments "
+               << "(non-functional FST?)"
+               << " w1 = " << w1 << " w2 = " << w2;
+    return Weight::NoWeight();
+  }
+  return w1;
+}
+
+// Longest common prefix for left string semiring.
+template <typename Label>
+inline StringWeight<Label, STRING_LEFT> Plus(
+    const StringWeight<Label, STRING_LEFT> &w1,
+    const StringWeight<Label, STRING_LEFT> &w2) {
+  using Weight = StringWeight<Label, STRING_LEFT>;
+  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
+  if (w1 == Weight::Zero()) return w2;
+  if (w2 == Weight::Zero()) return w1;
+  Weight sum;
+  typename Weight::Iterator iter1(w1);
+  typename Weight::Iterator iter2(w2);
+  for (; !iter1.Done() && !iter2.Done() && iter1.Value() == iter2.Value();
+       iter1.Next(), iter2.Next()) {
+    sum.PushBack(iter1.Value());
+  }
+  return sum;
+}
+
+// Longest common suffix for right string semiring.
+template <typename Label>
+inline StringWeight<Label, STRING_RIGHT> Plus(
+    const StringWeight<Label, STRING_RIGHT> &w1,
+    const StringWeight<Label, STRING_RIGHT> &w2) {
+  using Weight = StringWeight<Label, STRING_RIGHT>;
+  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
+  if (w1 == Weight::Zero()) return w2;
+  if (w2 == Weight::Zero()) return w1;
+  Weight sum;
+  typename Weight::ReverseIterator iter1(w1);
+  typename Weight::ReverseIterator iter2(w2);
+  for (; !iter1.Done() && !iter2.Done() && iter1.Value() == iter2.Value();
+       iter1.Next(), iter2.Next()) {
+    sum.PushFront(iter1.Value());
+  }
+  return sum;
+}
+
+template <typename Label, StringType S>
+inline StringWeight<Label, S> Times(const StringWeight<Label, S> &w1,
+                                    const StringWeight<Label, S> &w2) {
+  using Weight = StringWeight<Label, S>;
+  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
+  if (w1 == Weight::Zero() || w2 == Weight::Zero()) return Weight::Zero();
+  Weight product(w1);
+  for (typename Weight::Iterator iter(w2); !iter.Done(); iter.Next()) {
+    product.PushBack(iter.Value());
+  }
+  return product;
+}
+
+// Left division in a left string semiring.
+template <typename Label, StringType S>
+inline StringWeight<Label, S> DivideLeft(const StringWeight<Label, S> &w1,
+                                         const StringWeight<Label, S> &w2) {
+  using Weight = StringWeight<Label, S>;
+  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
+  if (w2 == Weight::Zero()) {
+    return Weight(Label(kStringBad));
+  } else if (w1 == Weight::Zero()) {
+    return Weight::Zero();
+  }
+  Weight result;
+  typename Weight::Iterator iter(w1);
+  size_t i = 0;
+  for (; !iter.Done() && i < w2.Size(); iter.Next(), ++i) {
+  }
+  for (; !iter.Done(); iter.Next()) result.PushBack(iter.Value());
+  return result;
+}
+
+// Right division in a right string semiring.
+template <typename Label, StringType S>
+inline StringWeight<Label, S> DivideRight(const StringWeight<Label, S> &w1,
+                                          const StringWeight<Label, S> &w2) {
+  using Weight = StringWeight<Label, S>;
+  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
+  if (w2 == Weight::Zero()) {
+    return Weight(Label(kStringBad));
+  } else if (w1 == Weight::Zero()) {
+    return Weight::Zero();
+  }
+  Weight result;
+  typename Weight::ReverseIterator iter(w1);
+  size_t i = 0;
+  for (; !iter.Done() && i < w2.Size(); iter.Next(), ++i) {
+  }
+  for (; !iter.Done(); iter.Next()) result.PushFront(iter.Value());
+  return result;
+}
+
+// Default is the restricted string semiring.
+template <typename Label, StringType S>
+inline StringWeight<Label, S> Divide(const StringWeight<Label, S> &w1,
+                                     const StringWeight<Label, S> &w2,
+                                     DivideType divide_type) {
+  using Weight = StringWeight<Label, S>;
+  if (divide_type == DIVIDE_LEFT) {
+    return DivideLeft(w1, w2);
+  } else if (divide_type == DIVIDE_RIGHT) {
+    return DivideRight(w1, w2);
+  } else {
+    FSTERROR() << "StringWeight::Divide: "
+               << "Only explicit left or right division is defined "
+               << "for the " << Weight::Type() << " semiring";
+    return Weight::NoWeight();
+  }
+}
+
+// Left division in the left string semiring.
+template <typename Label>
+inline StringWeight<Label, STRING_LEFT> Divide(
+    const StringWeight<Label, STRING_LEFT> &w1,
+    const StringWeight<Label, STRING_LEFT> &w2, DivideType divide_type) {
+  if (divide_type != DIVIDE_LEFT) {
+    FSTERROR() << "StringWeight::Divide: Only left division is defined "
+               << "for the left string semiring";
+    return StringWeight<Label, STRING_LEFT>::NoWeight();
+  }
+  return DivideLeft(w1, w2);
+}
+
+// Right division in the right string semiring.
+template <typename Label>
+inline StringWeight<Label, STRING_RIGHT> Divide(
+    const StringWeight<Label, STRING_RIGHT> &w1,
+    const StringWeight<Label, STRING_RIGHT> &w2, DivideType divide_type) {
+  if (divide_type != DIVIDE_RIGHT) {
+    FSTERROR() << "StringWeight::Divide: Only right division is defined "
+               << "for the right string semiring";
+    return StringWeight<Label, STRING_RIGHT>::NoWeight();
+  }
+  return DivideRight(w1, w2);
+}
+
+// This function object generates StringWeights that are random integer strings
+// from {1, ... , alphabet_size)^{0, max_string_length} U { Zero }. This is
+// intended primarily for testing.
+template <class Label, StringType S>
+class WeightGenerate<StringWeight<Label, S>> {
+ public:
+  using Weight = StringWeight<Label, S>;
+
+  explicit WeightGenerate(bool allow_zero = true,
+                          size_t alphabet_size = kNumRandomWeights,
+                          size_t max_string_length = kNumRandomWeights)
+      : allow_zero_(allow_zero),
+        alphabet_size_(alphabet_size),
+        max_string_length_(max_string_length) {}
+
+  Weight operator()() const {
+    size_t n = rand() % (max_string_length_ + allow_zero_);  // NOLINT
+    if (allow_zero_ && n == max_string_length_) return Weight::Zero();
+    std::vector<Label> labels;
+    labels.reserve(n);
+    for (size_t i = 0; i < n; ++i) {
+      labels.push_back(rand() % alphabet_size_ + 1);  // NOLINT
+    }
+    return Weight(labels.begin(), labels.end());
+  }
+
+ private:
+  // Permits Zero() and zero divisors.
+  const bool allow_zero_;
+  // Alphabet size for random weights.
+  const size_t alphabet_size_;
+  // Number of alternative random weights.
+  const size_t max_string_length_;
+};
+
+// Determines whether to use left, right, or (general) gallic semiring. Includes
+// a restricted version that signals an error if proper string prefixes or
+// suffixes would otherwise be returned by string Plus. This is useful with
+// algorithms that require functional transducer input. Also includes min
+// version that changes the Plus to keep only the lowest W weight string.
+enum GallicType {
+  GALLIC_LEFT = 0,
+  GALLIC_RIGHT = 1,
+  GALLIC_RESTRICT = 2,
+  GALLIC_MIN = 3,
+  GALLIC = 4
+};
+
+constexpr StringType GallicStringType(GallicType g) {
+  return g == GALLIC_LEFT
+             ? STRING_LEFT
+             : (g == GALLIC_RIGHT ? STRING_RIGHT : STRING_RESTRICT);
+}
+
+constexpr GallicType ReverseGallicType(GallicType g) {
+  return g == GALLIC_LEFT
+             ? GALLIC_RIGHT
+             : (g == GALLIC_RIGHT
+                    ? GALLIC_LEFT
+                    : (g == GALLIC_RESTRICT
+                           ? GALLIC_RESTRICT
+                           : (g == GALLIC_MIN ? GALLIC_MIN : GALLIC)));
+}
+
+// Product of string weight and an arbitraryy weight.
+template <class Label, class W, GallicType G = GALLIC_LEFT>
+struct GallicWeight
+    : public ProductWeight<StringWeight<Label, GallicStringType(G)>, W> {
+  using ReverseWeight =
+      GallicWeight<Label, typename W::ReverseWeight, ReverseGallicType(G)>;
+  using SW = StringWeight<Label, GallicStringType(G)>;
+
+  using ProductWeight<SW, W>::Properties;
+
+  GallicWeight() {}
+
+  GallicWeight(SW w1, W w2) : ProductWeight<SW, W>(w1, w2) {}
+
+  explicit GallicWeight(const string &s, int *nread = nullptr)
+      : ProductWeight<SW, W>(s, nread) {}
+
+  explicit GallicWeight(const ProductWeight<SW, W> &w)
+      : ProductWeight<SW, W>(w) {}
+
+  static const GallicWeight &Zero() {
+    static const GallicWeight zero(ProductWeight<SW, W>::Zero());
+    return zero;
+  }
+
+  static const GallicWeight &One() {
+    static const GallicWeight one(ProductWeight<SW, W>::One());
+    return one;
+  }
+
+  static const GallicWeight &NoWeight() {
+    static const GallicWeight no_weight(ProductWeight<SW, W>::NoWeight());
+    return no_weight;
+  }
+
+  static const string &Type() {
+    static const string *const type = new string(
+        G == GALLIC_LEFT
+            ? "left_gallic"
+            : (G == GALLIC_RIGHT
+                   ? "right_gallic"
+                   : (G == GALLIC_RESTRICT
+                          ? "restricted_gallic"
+                          : (G == GALLIC_MIN ? "min_gallic" : "gallic"))));
+    return *type;
+  }
+
+  GallicWeight Quantize(float delta = kDelta) const {
+    return GallicWeight(ProductWeight<SW, W>::Quantize(delta));
+  }
+
+  ReverseWeight Reverse() const {
+    return ReverseWeight(ProductWeight<SW, W>::Reverse());
+  }
+};
+
+// Default plus.
+template <class Label, class W, GallicType G>
+inline GallicWeight<Label, W, G> Plus(const GallicWeight<Label, W, G> &w,
+                                      const GallicWeight<Label, W, G> &v) {
+  return GallicWeight<Label, W, G>(Plus(w.Value1(), v.Value1()),
+                                   Plus(w.Value2(), v.Value2()));
+}
+
+// Min gallic plus.
+template <class Label, class W>
+inline GallicWeight<Label, W, GALLIC_MIN> Plus(
+    const GallicWeight<Label, W, GALLIC_MIN> &w1,
+    const GallicWeight<Label, W, GALLIC_MIN> &w2) {
+  static const NaturalLess<W> less;
+  return less(w1.Value2(), w2.Value2()) ? w1 : w2;
+}
+
+template <class Label, class W, GallicType G>
+inline GallicWeight<Label, W, G> Times(const GallicWeight<Label, W, G> &w,
+                                       const GallicWeight<Label, W, G> &v) {
+  return GallicWeight<Label, W, G>(Times(w.Value1(), v.Value1()),
+                                   Times(w.Value2(), v.Value2()));
+}
+
+template <class Label, class W, GallicType G>
+inline GallicWeight<Label, W, G> Divide(const GallicWeight<Label, W, G> &w,
+                                        const GallicWeight<Label, W, G> &v,
+                                        DivideType divide_type = DIVIDE_ANY) {
+  return GallicWeight<Label, W, G>(Divide(w.Value1(), v.Value1(), divide_type),
+                                   Divide(w.Value2(), v.Value2(), divide_type));
+}
+
+// This function object generates gallic weights by calling an underlying
+// product weight generator. This is intended primarily for testing.
+template <class Label, class W, GallicType G>
+class WeightGenerate<GallicWeight<Label, W, G>>
+    : public WeightGenerate<
+          ProductWeight<StringWeight<Label, GallicStringType(G)>, W>> {
+ public:
+  using Weight = GallicWeight<Label, W, G>;
+  using Generate = WeightGenerate<
+      ProductWeight<StringWeight<Label, GallicStringType(G)>, W>>;
+
+  explicit WeightGenerate(bool allow_zero = true) : generate_(allow_zero) {}
+
+  Weight operator()() const { return Weight(generate_()); }
+
+ private:
+  const Generate generate_;
+};
+
+// Union weight options for (general) GALLIC type.
+template <class Label, class W>
+struct GallicUnionWeightOptions {
+  using ReverseOptions = GallicUnionWeightOptions<Label, W>;
+  using GW = GallicWeight<Label, W, GALLIC_RESTRICT>;
+  using SW = StringWeight<Label, GallicStringType(GALLIC_RESTRICT)>;
+  using SI = StringWeightIterator<SW>;
+
+  // Military order.
+  struct Compare {
+    bool operator()(const GW &w1, const GW &w2) const {
+      const SW &s1 = w1.Value1();
+      const SW &s2 = w2.Value1();
+      if (s1.Size() < s2.Size()) return true;
+      if (s1.Size() > s2.Size()) return false;
+      SI iter1(s1);
+      SI iter2(s2);
+      while (!iter1.Done()) {
+        const auto l1 = iter1.Value();
+        const auto l2 = iter2.Value();
+        if (l1 < l2) return true;
+        if (l1 > l2) return false;
+        iter1.Next();
+        iter2.Next();
+      }
+      return false;
+    }
+  };
+
+  // Adds W weights when string part equal.
+  struct Merge {
+    GW operator()(const GW &w1, const GW &w2) const {
+      return GW(w1.Value1(), Plus(w1.Value2(), w2.Value2()));
+    }
+  };
+};
+
+// Specialization for the (general) GALLIC type.
+template <class Label, class W>
+struct GallicWeight<Label, W, GALLIC>
+    : public UnionWeight<GallicWeight<Label, W, GALLIC_RESTRICT>,
+                         GallicUnionWeightOptions<Label, W>> {
+  using GW = GallicWeight<Label, W, GALLIC_RESTRICT>;
+  using SW = StringWeight<Label, GallicStringType(GALLIC_RESTRICT)>;
+  using SI = StringWeightIterator<SW>;
+  using UW = UnionWeight<GW, GallicUnionWeightOptions<Label, W>>;
+  using UI = UnionWeightIterator<GW, GallicUnionWeightOptions<Label, W>>;
+  using ReverseWeight = GallicWeight<Label, W, GALLIC>;
+
+  using UW::Properties;
+
+  GallicWeight() {}
+
+  // Copy constructor.
+  GallicWeight(const UW &weight) : UW(weight) {}  // NOLINT
+
+  // Singleton constructors: create a GALLIC weight containing a single
+  // GALLIC_RESTRICT weight. Takes as argument (1) a GALLIC_RESTRICT weight or
+  // (2) the two components of a GALLIC_RESTRICT weight.
+  explicit GallicWeight(const GW &weight) : UW(weight) {}
+
+  GallicWeight(SW w1, W w2) : UW(GW(w1, w2)) {}
+
+  explicit GallicWeight(const string &str, int *nread = nullptr)
+      : UW(str, nread) {}
+
+  static const GallicWeight<Label, W, GALLIC> &Zero() {
+    static const GallicWeight<Label, W, GALLIC> zero(UW::Zero());
+    return zero;
+  }
+
+  static const GallicWeight<Label, W, GALLIC> &One() {
+    static const GallicWeight<Label, W, GALLIC> one(UW::One());
+    return one;
+  }
+
+  static const GallicWeight<Label, W, GALLIC> &NoWeight() {
+    static const GallicWeight<Label, W, GALLIC> no_weight(UW::NoWeight());
+    return no_weight;
+  }
+
+  static const string &Type() {
+    static const string *const type = new string("gallic");
+    return *type;
+  }
+
+  GallicWeight<Label, W, GALLIC> Quantize(float delta = kDelta) const {
+    return UW::Quantize(delta);
+  }
+
+  ReverseWeight Reverse() const { return UW::Reverse(); }
+};
+
+// (General) gallic plus.
+template <class Label, class W>
+inline GallicWeight<Label, W, GALLIC> Plus(
+    const GallicWeight<Label, W, GALLIC> &w1,
+    const GallicWeight<Label, W, GALLIC> &w2) {
+  using GW = GallicWeight<Label, W, GALLIC_RESTRICT>;
+  using UW = UnionWeight<GW, GallicUnionWeightOptions<Label, W>>;
+  return Plus(static_cast<UW>(w1), static_cast<UW>(w2));
+}
+
+// (General) gallic times.
+template <class Label, class W>
+inline GallicWeight<Label, W, GALLIC> Times(
+    const GallicWeight<Label, W, GALLIC> &w1,
+    const GallicWeight<Label, W, GALLIC> &w2) {
+  using GW = GallicWeight<Label, W, GALLIC_RESTRICT>;
+  using UW = UnionWeight<GW, GallicUnionWeightOptions<Label, W>>;
+  return Times(static_cast<UW>(w1), static_cast<UW>(w2));
+}
+
+// (General) gallic divide.
+template <class Label, class W>
+inline GallicWeight<Label, W, GALLIC> Divide(
+    const GallicWeight<Label, W, GALLIC> &w1,
+    const GallicWeight<Label, W, GALLIC> &w2,
+    DivideType divide_type = DIVIDE_ANY) {
+  using GW = GallicWeight<Label, W, GALLIC_RESTRICT>;
+  using UW = UnionWeight<GW, GallicUnionWeightOptions<Label, W>>;
+  return Divide(static_cast<UW>(w1), static_cast<UW>(w2), divide_type);
+}
+
+// This function object generates gallic weights by calling an underlying
+// union weight generator. This is intended primarily for testing.
+template <class Label, class W>
+class WeightGenerate<GallicWeight<Label, W, GALLIC>>
+    : public WeightGenerate<UnionWeight<GallicWeight<Label, W, GALLIC_RESTRICT>,
+                                        GallicUnionWeightOptions<Label, W>>> {
+ public:
+  using Weight = GallicWeight<Label, W, GALLIC>;
+  using Generate =
+      WeightGenerate<UnionWeight<GallicWeight<Label, W, GALLIC_RESTRICT>,
+                                 GallicUnionWeightOptions<Label, W>>>;
+
+  explicit WeightGenerate(bool allow_zero = true) : generate_(allow_zero) {}
+
+  Weight operator()() const { return Weight(generate_()); }
+
+ private:
+  const Generate generate_;
+};
+
+}  // namespace fst
+
+#endif  // FST_STRING_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/string.h b/projects/llm_framework/include/fst/string.h
new file mode 100644
index 00000000..07ebfddd
--- /dev/null
+++ b/projects/llm_framework/include/fst/string.h
@@ -0,0 +1,286 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Utilities to convert strings into FSTs.
+
+#ifndef FST_STRING_H_
+#define FST_STRING_H_
+
+#include <memory>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#include <fst/flags.h>
+#include <fst/log.h>
+
+#include <fst/compact-fst.h>
+#include <fst/icu.h>
+#include <fst/mutable-fst.h>
+#include <fst/util.h>
+
+
+DECLARE_string(fst_field_separator);
+
+namespace fst {
+
+enum StringTokenType { SYMBOL = 1, BYTE = 2, UTF8 = 3 };
+
+namespace internal {
+
+template <class Label>
+bool ConvertSymbolToLabel(const char *str, const SymbolTable *syms,
+                          Label unknown_label, bool allow_negative,
+                          Label *output) {
+  int64 n;
+  if (syms) {
+    n = syms->Find(str);
+    if ((n == -1) && (unknown_label != kNoLabel)) n = unknown_label;
+    if (n == -1 || (!allow_negative && n < 0)) {
+      LOG(ERROR) << "ConvertSymbolToLabel: Symbol \"" << str
+                 << "\" is not mapped to any integer label, symbol table = "
+                 << syms->Name();
+      return false;
+    }
+  } else {
+    char *p;
+    n = strtoll(str, &p, 10);
+    if (p < str + strlen(str) || (!allow_negative && n < 0)) {
+      LOG(ERROR) << "ConvertSymbolToLabel: Bad label integer "
+                 << "= \"" << str << "\"";
+      return false;
+    }
+  }
+  *output = n;
+  return true;
+}
+
+template <class Label>
+bool ConvertStringToLabels(const string &str, StringTokenType token_type,
+                           const SymbolTable *syms, Label unknown_label,
+                           bool allow_negative, std::vector<Label> *labels) {
+  labels->clear();
+  if (token_type == StringTokenType::BYTE) {
+    labels->reserve(str.size());
+    return ByteStringToLabels(str, labels);
+  } else if (token_type == StringTokenType::UTF8) {
+    return UTF8StringToLabels(str, labels);
+  } else {
+    std::unique_ptr<char[]> c_str(new char[str.size() + 1]);
+    str.copy(c_str.get(), str.size());
+    c_str[str.size()] = 0;
+    std::vector<char *> vec;
+    const string separator = "\n" + FLAGS_fst_field_separator;
+    SplitString(c_str.get(), separator.c_str(), &vec, true);
+    for (const char *c : vec) {
+      Label label;
+      if (!ConvertSymbolToLabel(c, syms, unknown_label, allow_negative,
+                                &label)) {
+        return false;
+      }
+      labels->push_back(label);
+    }
+  }
+  return true;
+}
+
+// The sep string is used as a separator between symbols, unless it is nullptr,
+// in which case the last character of FLAGS_fst_field_separator is used.
+template <class Label>
+bool LabelsToSymbolString(const std::vector<Label> &labels, string *str,
+                          const SymbolTable &syms,
+                          const string *sep = nullptr) {
+  std::stringstream ostrm;
+  string delim = "";
+  for (auto label : labels) {
+    ostrm << delim;
+    const string &symbol = syms.Find(label);
+    if (symbol.empty()) {
+      LOG(ERROR) << "LabelToSymbolString: Label " << label
+                 << " is not mapped onto any textual symbol in symbol table "
+                 << syms.Name();
+      return false;
+    }
+    ostrm << symbol;
+    delim = sep != nullptr ? *sep : string(1, FLAGS_fst_field_separator.back());
+  }
+  *str = ostrm.str();
+  return !!ostrm;
+}
+
+}  // namespace internal
+
+// TODO(kbg): Move this just before StringPrinter.
+// Converts an FST to a vector of output labels. To get input labels, use
+// Project or Invert. Returns true on success. Use only with string FSTs; may
+// loop for non-string FSTs.
+template <class Arc>
+bool StringFstToOutputLabels(const Fst<Arc> &fst,
+                             std::vector<typename Arc::Label> *labels) {
+  labels->clear();
+  auto s = fst.Start();
+  if (s == kNoStateId) {
+    LOG(ERROR) << "StringFstToOutputLabels: Invalid start state";
+    return false;
+  }
+  while (fst.Final(s) == Arc::Weight::Zero()) {
+    ArcIterator<Fst<Arc>> aiter(fst, s);
+    if (aiter.Done()) {
+      LOG(ERROR) << "StringFstToOutputLabels: Does not reach final state";
+      return false;
+    }
+    const auto &arc = aiter.Value();
+    labels->push_back(arc.olabel);
+    s = arc.nextstate;
+    aiter.Next();
+    if (!aiter.Done()) {
+      LOG(ERROR) << "StringFstToOutputLabels: State has multiple outgoing arcs";
+      return false;
+    }
+  }
+  if (fst.NumArcs(s) != 0) {
+    LOG(ERROR) << "FstToOutputLabels: Final state has outgoing arc(s)";
+    return false;
+  }
+  return true;
+}
+
+// TODO(kbg): Move this just before StringPrinter.
+// Converts a list of symbols to a string. Returns true on success. If the token
+// type is SYMBOL and sep is provided, it is used to separate textual symbols.
+// If the token type is SYMBOL and it is not provided, the last character of
+// FLAGS_fst_field_separator is used.
+template <class Label>
+bool LabelsToString(const std::vector<Label> &labels, string *str,
+                    StringTokenType ttype = BYTE,
+                    const SymbolTable *syms = nullptr,
+                    const string *sep = nullptr) {
+  switch (ttype) {
+    case StringTokenType::BYTE: {
+      return LabelsToByteString(labels, str);
+    }
+    case StringTokenType::UTF8: {
+      return LabelsToUTF8String(labels, str);
+    }
+    case StringTokenType::SYMBOL: {
+      return internal::LabelsToSymbolString(labels, str, *syms, sep);
+    }
+  }
+  return false;
+}
+
+// Functor for compiling a string in an FST.
+template <class Arc>
+class StringCompiler {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  explicit StringCompiler(StringTokenType token_type,
+                          const SymbolTable *syms = nullptr,
+                          Label unknown_label = kNoLabel,
+                          bool allow_negative = false)
+      : token_type_(token_type),
+        syms_(syms),
+        unknown_label_(unknown_label),
+        allow_negative_(allow_negative) {}
+
+  // Compiles string into an FST.
+  template <class FST>
+  bool operator()(const string &str, FST *fst) const {
+    std::vector<Label> labels;
+    if (!internal::ConvertStringToLabels(str, token_type_, syms_,
+                                         unknown_label_, allow_negative_,
+                                         &labels)) {
+      return false;
+    }
+    Compile(labels, fst);
+    return true;
+  }
+
+  template <class FST>
+  bool operator()(const string &str, FST *fst, Weight weight) const {
+    std::vector<Label> labels;
+    if (!internal::ConvertStringToLabels(str, token_type_, syms_,
+                                         unknown_label_, allow_negative_,
+                                         &labels)) {
+      return false;
+    }
+    Compile(labels, fst, std::move(weight));
+    return true;
+  }
+
+ private:
+  void Compile(const std::vector<Label> &labels, MutableFst<Arc> *fst,
+               Weight weight = Weight::One()) const {
+    fst->DeleteStates();
+    while (fst->NumStates() <= labels.size()) fst->AddState();
+    for (StateId i = 0; i < labels.size(); ++i) {
+      fst->AddArc(i, Arc(labels[i], labels[i], Weight::One(), i + 1));
+    }
+    fst->SetStart(0);
+    fst->SetFinal(labels.size(), std::move(weight));
+  }
+
+  template <class Unsigned>
+  void Compile(const std::vector<Label> &labels,
+               CompactStringFst<Arc, Unsigned> *fst) const {
+    fst->SetCompactElements(labels.begin(), labels.end());
+  }
+
+  template <class Unsigned>
+  void Compile(const std::vector<Label> &labels,
+               CompactWeightedStringFst<Arc, Unsigned> *fst,
+               const Weight &weight = Weight::One()) const {
+    std::vector<std::pair<Label, Weight>> compacts;
+    compacts.reserve(labels.size() + 1);
+    for (StateId i = 0; i < static_cast<StateId>(labels.size()) - 1; ++i) {
+      compacts.emplace_back(labels[i], Weight::One());
+    }
+    compacts.emplace_back(!labels.empty() ? labels.back() : kNoLabel, weight);
+    fst->SetCompactElements(compacts.begin(), compacts.end());
+  }
+
+  const StringTokenType token_type_;
+  const SymbolTable *syms_;    // Symbol table (used when token type is symbol).
+  const Label unknown_label_;  // Label for token missing from symbol table.
+  const bool allow_negative_;  // Negative labels allowed?
+
+  StringCompiler(const StringCompiler &) = delete;
+  StringCompiler &operator=(const StringCompiler &) = delete;
+};
+
+// Helpers for StringPrinter.
+
+// Functor for printing a string FST as a string.
+template <class Arc>
+class StringPrinter {
+ public:
+  using Label = typename Arc::Label;
+
+  explicit StringPrinter(StringTokenType token_type,
+                         const SymbolTable *syms = nullptr)
+      : token_type_(token_type), syms_(syms) {}
+
+  // Converts the FST into a string. With SYMBOL token type, sep is used as a
+  // separator between symbols, unless it is nullptr, in which case the last
+  // character of FLAGS_fst_field_separator is used. Returns true on success.
+  bool operator()(const Fst<Arc> &fst, string *str,
+                  const string *sep = nullptr) const {
+    std::vector<Label> labels;
+    return (StringFstToOutputLabels(fst, &labels) &&
+            LabelsToString(labels, str, token_type_, syms_, sep));
+  }
+
+ private:
+  const StringTokenType token_type_;
+  const SymbolTable *syms_;
+
+  StringPrinter(const StringPrinter &) = delete;
+  StringPrinter &operator=(const StringPrinter &) = delete;
+};
+
+}  // namespace fst
+
+#endif  // FST_STRING_H_
diff --git a/projects/llm_framework/include/fst/symbol-table-ops.h b/projects/llm_framework/include/fst/symbol-table-ops.h
new file mode 100644
index 00000000..38f9d819
--- /dev/null
+++ b/projects/llm_framework/include/fst/symbol-table-ops.h
@@ -0,0 +1,76 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+
+#ifndef FST_SYMBOL_TABLE_OPS_H_
+#define FST_SYMBOL_TABLE_OPS_H_
+
+#include <string>
+#include <unordered_set>
+#include <vector>
+
+
+#include <fst/fst.h>
+#include <fst/symbol-table.h>
+
+
+namespace fst {
+
+// Returns a minimal symbol table containing only symbols referenced by the
+// passed fst.  Symbols preserve their original numbering, so fst does not
+// require relabeling.
+template <class Arc>
+SymbolTable *PruneSymbolTable(const Fst<Arc> &fst, const SymbolTable &syms,
+                              bool input) {
+  std::unordered_set<typename Arc::Label> seen;
+  seen.insert(0);  // Always keep epsilon.
+  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
+    for (ArcIterator<Fst<Arc>> aiter(fst, siter.Value()); !aiter.Done();
+         aiter.Next()) {
+      const auto sym = (input) ? aiter.Value().ilabel : aiter.Value().olabel;
+      seen.insert(sym);
+    }
+  }
+  auto *pruned = new SymbolTable(syms.Name() + "_pruned");
+  for (SymbolTableIterator stiter(syms); !stiter.Done(); stiter.Next()) {
+    const auto label = stiter.Value();
+    if (seen.count(label)) pruned->AddSymbol(stiter.Symbol(), label);
+  }
+  return pruned;
+}
+
+// Relabels a symbol table to make it a contiguous mapping.
+SymbolTable *CompactSymbolTable(const SymbolTable &syms);
+
+// Merges two SymbolTables, all symbols from left will be merged into right
+// with the same ids.  Symbols in right that have conflicting ids with those
+// in left will be assigned to value assigned from the left SymbolTable.
+// The returned symbol table will never modify symbol assignments from the left
+// side, but may do so on the right.  If right_relabel_output is non-NULL, it
+// will be assigned true if the symbols from the right table needed to be
+// reassigned.
+// A potential use case is to Compose two Fst's that have different symbol
+// tables.  You can reconcile them in the following way:
+//   Fst<Arc> a, b;
+//   bool relabel;
+//   std::unique_ptr<SymbolTable> bnew(MergeSymbolTable(a.OutputSymbols(),
+//                                     b.InputSymbols(), &relabel);
+//   if (relabel) {
+//     Relabel(b, bnew.get(), nullptr);
+//   }
+//   b.SetInputSymbols(bnew);
+SymbolTable *MergeSymbolTable(const SymbolTable &left, const SymbolTable &right,
+                              bool *right_relabel_output = nullptr);
+
+// Read the symbol table from any Fst::Read()able file, without loading the
+// corresponding Fst.  Returns nullptr if the Fst does not contain a symbol
+// table or the symbol table cannot be read.
+SymbolTable *FstReadSymbols(const string &filename, bool input);
+
+// Adds a contiguous range of symbols to a symbol table using a simple prefix
+// for the string, returning false if the inserted symbol string clashes with
+// any currently present.
+bool AddAuxiliarySymbols(const string &prefix, int64 start_label,
+                         int64 nlabels, SymbolTable *syms);
+
+}  // namespace fst
+#endif  // FST_SYMBOL_TABLE_OPS_H_
diff --git a/projects/llm_framework/include/fst/symbol-table.h b/projects/llm_framework/include/fst/symbol-table.h
new file mode 100644
index 00000000..2884bc3b
--- /dev/null
+++ b/projects/llm_framework/include/fst/symbol-table.h
@@ -0,0 +1,445 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Classes to provide symbol-to-integer and integer-to-symbol mappings.
+
+#ifndef FST_SYMBOL_TABLE_H_
+#define FST_SYMBOL_TABLE_H_
+
+#include <functional>
+#include <ios>
+#include <iostream>
+#include <memory>
+#include <sstream>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include <fst/compat.h>
+#include <fst/flags.h>
+#include <fst/log.h>
+#include <fstream>
+#include <map>
+
+DECLARE_bool(fst_compat_symbols);
+
+namespace fst {
+
+constexpr int64 kNoSymbol = -1;
+
+// WARNING: Reading via symbol table read options should
+//          not be used. This is a temporary work around for
+//          reading symbol ranges of previously stored symbol sets.
+struct SymbolTableReadOptions {
+  SymbolTableReadOptions() {}
+
+  SymbolTableReadOptions(
+      std::vector<std::pair<int64, int64>> string_hash_ranges,
+      const string &source)
+      : string_hash_ranges(std::move(string_hash_ranges)), source(source) {}
+
+  std::vector<std::pair<int64, int64>> string_hash_ranges;
+  string source;
+};
+
+struct SymbolTableTextOptions {
+  explicit SymbolTableTextOptions(bool allow_negative_labels = false);
+
+  bool allow_negative_labels;
+  string fst_field_separator;
+};
+
+namespace internal {
+
+extern const int kLineLen;
+
+// List of symbols with a dense hash for looking up symbol index, rehashing at
+// 75% occupancy.
+class DenseSymbolMap {
+ public:
+  DenseSymbolMap();
+
+  DenseSymbolMap(const DenseSymbolMap &x);
+
+  std::pair<int64, bool> InsertOrFind(const string &key);
+
+  int64 Find(const string &key) const;
+
+  size_t Size() const { return symbols_.size(); }
+
+  const string &GetSymbol(size_t idx) const { return symbols_[idx]; }
+
+  void RemoveSymbol(size_t idx);
+
+ private:
+  // num_buckets must be power of 2.
+  void Rehash(size_t num_buckets);
+
+  int64 empty_;
+  std::vector<string> symbols_;
+  std::hash<string> str_hash_;
+  std::vector<int64> buckets_;
+  uint64 hash_mask_;
+};
+
+class SymbolTableImpl {
+ public:
+  explicit SymbolTableImpl(const string &name)
+      : name_(name),
+        available_key_(0),
+        dense_key_limit_(0),
+        check_sum_finalized_(false) {}
+
+  SymbolTableImpl(const SymbolTableImpl &impl)
+      : name_(impl.name_),
+        available_key_(impl.available_key_),
+        dense_key_limit_(impl.dense_key_limit_),
+        symbols_(impl.symbols_),
+        idx_key_(impl.idx_key_),
+        key_map_(impl.key_map_),
+        check_sum_finalized_(false) {}
+
+  int64 AddSymbol(const string &symbol, int64 key);
+
+  int64 AddSymbol(const string &symbol) {
+    return AddSymbol(symbol, available_key_);
+  }
+
+  // Removes the symbol with the given key. The removal is costly
+  // (O(NumSymbols)) and may reduce the efficiency of Find() because of a
+  // potentially reduced size of the dense key interval.
+  void RemoveSymbol(int64 key);
+
+  static SymbolTableImpl *ReadText(
+      std::istream &strm, const string &name,
+      const SymbolTableTextOptions &opts = SymbolTableTextOptions());
+
+  static SymbolTableImpl* Read(std::istream &strm,
+                               const SymbolTableReadOptions &opts);
+
+  bool Write(std::ostream &strm) const;
+
+  // Return the string associated with the key. If the key is out of
+  // range (<0, >max), return an empty string.
+  string Find(int64 key) const {
+    int64 idx = key;
+    if (key < 0 || key >= dense_key_limit_) {
+      const auto it = key_map_.find(key);
+      if (it == key_map_.end()) return "";
+      idx = it->second;
+    }
+    if (idx < 0 || idx >= symbols_.Size()) return "";
+    return symbols_.GetSymbol(idx);
+  }
+
+  // Returns the key associated with the symbol; if the symbol
+  // does not exists, returns kNoSymbol.
+  int64 Find(const string &symbol) const {
+    int64 idx = symbols_.Find(symbol);
+    if (idx == kNoSymbol || idx < dense_key_limit_) return idx;
+    return idx_key_[idx - dense_key_limit_];
+  }
+
+  bool Member(int64 key) const { return !Find(key).empty(); }
+
+  bool Member(const string &symbol) const { return Find(symbol) != kNoSymbol; }
+
+  int64 GetNthKey(ssize_t pos) const {
+    if (pos < 0 || pos >= symbols_.Size()) return kNoSymbol;
+    if (pos < dense_key_limit_) return pos;
+    return Find(symbols_.GetSymbol(pos));
+  }
+
+  const string &Name() const { return name_; }
+
+  void SetName(const string &new_name) { name_ = new_name; }
+
+  const string &CheckSum() const {
+    MaybeRecomputeCheckSum();
+    return check_sum_string_;
+  }
+
+  const string &LabeledCheckSum() const {
+    MaybeRecomputeCheckSum();
+    return labeled_check_sum_string_;
+  }
+
+  int64 AvailableKey() const { return available_key_; }
+
+  size_t NumSymbols() const { return symbols_.Size(); }
+
+ private:
+  // Recomputes the checksums (both of them) if we've had changes since the last
+  // computation (i.e., if check_sum_finalized_ is false).
+  // Takes ~2.5 microseconds (dbg) or ~230 nanoseconds (opt) on a 2.67GHz Xeon
+  // if the checksum is up-to-date (requiring no recomputation).
+  void MaybeRecomputeCheckSum() const;
+
+  string name_;
+  int64 available_key_;
+  int64 dense_key_limit_;
+
+  DenseSymbolMap symbols_;
+  // Maps index to key for index >= dense_key_limit:
+  //   key = idx_key_[index - dense_key_limit]
+  std::vector<int64> idx_key_;
+  // Maps key to index for key >= dense_key_limit_.
+  //  index = key_map_[key]
+  std::map<int64, int64> key_map_;
+
+  mutable bool check_sum_finalized_;
+  mutable string check_sum_string_;
+  mutable string labeled_check_sum_string_;
+  mutable Mutex check_sum_mutex_;
+};
+
+}  // namespace internal
+
+// Symbol (string) to integer (and reverse) mapping.
+//
+// The SymbolTable implements the mappings of labels to strings and reverse.
+// SymbolTables are used to describe the alphabet of the input and output
+// labels for arcs in a Finite State Transducer.
+//
+// SymbolTables are reference-counted and can therefore be shared across
+// multiple machines. For example a language model grammar G, with a
+// SymbolTable for the words in the language model can share this symbol
+// table with the lexical representation L o G.
+class SymbolTable {
+ public:
+  // Constructs symbol table with an optional name.
+  explicit SymbolTable(const string &name = "<unspecified>")
+      : impl_(std::make_shared<internal::SymbolTableImpl>(name)) {}
+
+  virtual ~SymbolTable() {}
+
+  // Reads a text representation of the symbol table from an istream. Pass a
+  // name to give the resulting SymbolTable.
+  static SymbolTable *ReadText(
+      std::istream &strm, const string &name,
+      const SymbolTableTextOptions &opts = SymbolTableTextOptions()) {
+    auto *impl = internal::SymbolTableImpl::ReadText(strm, name, opts);
+    return impl ? new SymbolTable(impl) : nullptr;
+  }
+
+  // Reads a text representation of the symbol table.
+  static SymbolTable *ReadText(const string &filename,
+      const SymbolTableTextOptions &opts = SymbolTableTextOptions()) {
+    std::ifstream strm(filename, std::ios_base::in);
+    if (!strm.good()) {
+      LOG(ERROR) << "SymbolTable::ReadText: Can't open file " << filename;
+      return nullptr;
+    }
+    return ReadText(strm, filename, opts);
+  }
+
+  // WARNING: Reading via symbol table read options should not be used. This is
+  // a temporary work-around.
+  static SymbolTable* Read(std::istream &strm,
+                           const SymbolTableReadOptions &opts) {
+    auto *impl = internal::SymbolTableImpl::Read(strm, opts);
+    return (impl) ? new SymbolTable(impl) : nullptr;
+  }
+
+  // Reads a binary dump of the symbol table from a stream.
+  static SymbolTable *Read(std::istream &strm,
+                           const string &source) {
+    SymbolTableReadOptions opts;
+    opts.source = source;
+    return Read(strm, opts);
+  }
+
+  // Reads a binary dump of the symbol table.
+  static SymbolTable *Read(const string& filename) {
+    std::ifstream strm(filename,
+                            std::ios_base::in | std::ios_base::binary);
+    if (!strm.good()) {
+      LOG(ERROR) << "SymbolTable::Read: Can't open file " << filename;
+      return nullptr;
+    }
+    return Read(strm, filename);
+  }
+
+  // DERIVABLE INTERFACE
+
+  // Creates a reference counted copy.
+  virtual SymbolTable *Copy() const { return new SymbolTable(*this); }
+
+  // Adds a symbol with given key to table. A symbol table also keeps track of
+  // the last available key (highest key value in the symbol table).
+  virtual int64 AddSymbol(const string &symbol, int64 key) {
+    MutateCheck();
+    return impl_->AddSymbol(symbol, key);
+  }
+
+  // Adds a symbol to the table. The associated value key is automatically
+  // assigned by the symbol table.
+  virtual int64 AddSymbol(const string &symbol) {
+    MutateCheck();
+    return impl_->AddSymbol(symbol);
+  }
+
+  // Adds another symbol table to this table. All key values will be offset
+  // by the current available key (highest key value in the symbol table).
+  // Note string symbols with the same key value will still have the same
+  // key value after the symbol table has been merged, but a different
+  // value. Adding symbol tables do not result in changes in the base table.
+  virtual void AddTable(const SymbolTable &table);
+
+  // Returns the current available key (i.e., highest key + 1) in the symbol
+  // table.
+  virtual int64 AvailableKey() const { return impl_->AvailableKey(); }
+
+  // Return the label-agnostic MD5 check-sum for this table. All new symbols
+  // added to the table will result in an updated checksum. Deprecated.
+  virtual const string &CheckSum() const { return impl_->CheckSum(); }
+
+  virtual int64 GetNthKey(ssize_t pos) const { return impl_->GetNthKey(pos); }
+
+  // Returns the string associated with the key; if the key is out of
+  // range (<0, >max), returns an empty string.
+  virtual string Find(int64 key) const { return impl_->Find(key); }
+
+  // Returns the key associated with the symbol; if the symbol does not exist,
+  // kNoSymbol is returned.
+  virtual int64 Find(const string &symbol) const { return impl_->Find(symbol); }
+
+  // Same as CheckSum(), but returns an label-dependent version.
+  virtual const string &LabeledCheckSum() const {
+    return impl_->LabeledCheckSum();
+  }
+
+  virtual bool Member(int64 key) const { return impl_->Member(key); }
+
+  virtual bool Member(const string &symbol) const {
+    return impl_->Member(symbol);
+  }
+
+  // Returns the name of the symbol table.
+  virtual const string &Name() const { return impl_->Name(); }
+
+  // Returns the current number of symbols in table (not necessarily equal to
+  // AvailableKey()).
+  virtual size_t NumSymbols() const { return impl_->NumSymbols(); }
+
+  virtual void RemoveSymbol(int64 key) {
+    MutateCheck();
+    return impl_->RemoveSymbol(key);
+  }
+
+  // Sets the name of the symbol table.
+  virtual void SetName(const string &new_name) {
+    MutateCheck();
+    impl_->SetName(new_name);
+  }
+
+  virtual bool Write(std::ostream &strm) const { return impl_->Write(strm); }
+
+  virtual bool Write(const string &filename) const {
+    std::ofstream strm(filename,
+                             std::ios_base::out | std::ios_base::binary);
+    if (!strm.good()) {
+      LOG(ERROR) << "SymbolTable::Write: Can't open file " << filename;
+      return false;
+    }
+    return Write(strm);
+  }
+
+  // Dump a text representation of the symbol table via a stream.
+  virtual bool WriteText(std::ostream &strm,
+      const SymbolTableTextOptions &opts = SymbolTableTextOptions()) const;
+
+  // Dump an text representation of the symbol table.
+  virtual bool WriteText(const string &filename) const {
+    std::ofstream strm(filename);
+    if (!strm.good()) {
+      LOG(ERROR) << "SymbolTable::WriteText: Can't open file " << filename;
+      return false;
+    }
+    return WriteText(strm);
+  }
+
+ private:
+  explicit SymbolTable(internal::SymbolTableImpl *impl) : impl_(impl) {}
+
+  void MutateCheck() {
+    if (!impl_.unique()) impl_.reset(new internal::SymbolTableImpl(*impl_));
+  }
+
+  const internal::SymbolTableImpl *Impl() const { return impl_.get(); }
+
+ private:
+  std::shared_ptr<internal::SymbolTableImpl> impl_;
+};
+
+// Iterator class for symbols in a symbol table.
+class SymbolTableIterator {
+ public:
+  explicit SymbolTableIterator(const SymbolTable &table)
+      : table_(table),
+        pos_(0),
+        nsymbols_(table.NumSymbols()),
+        key_(table.GetNthKey(0)) {}
+
+  ~SymbolTableIterator() {}
+
+  // Returns whether iterator is done.
+  bool Done() const { return (pos_ == nsymbols_); }
+
+  // Return the key of the current symbol.
+  int64 Value() const { return key_; }
+
+  // Return the string of the current symbol.
+  string Symbol() const { return table_.Find(key_); }
+
+  // Advances iterator.
+  void Next() {
+    ++pos_;
+    if (pos_ < nsymbols_) key_ = table_.GetNthKey(pos_);
+  }
+
+  // Resets iterator.
+  void Reset() {
+    pos_ = 0;
+    key_ = table_.GetNthKey(0);
+  }
+
+ private:
+  const SymbolTable &table_;
+  ssize_t pos_;
+  size_t nsymbols_;
+  int64 key_;
+};
+
+// Relabels a symbol table as specified by the input vector of pairs
+// (old label, new label). The new symbol table only retains symbols
+// for which a relabeling is explicitly specified.
+//
+// TODO(allauzen): consider adding options to allow for some form of implicit
+// identity relabeling.
+template <class Label>
+SymbolTable *RelabelSymbolTable(const SymbolTable *table,
+    const std::vector<std::pair<Label, Label>> &pairs) {
+  auto *new_table = new SymbolTable(
+      table->Name().empty() ? string()
+                            : (string("relabeled_") + table->Name()));
+  for (const auto &pair : pairs) {
+    new_table->AddSymbol(table->Find(pair.first), pair.second);
+  }
+  return new_table;
+}
+
+// Returns true if the two symbol tables have equal checksums. Passing in
+// nullptr for either table always returns true.
+bool CompatSymbols(const SymbolTable *syms1, const SymbolTable *syms2,
+                   bool warning = true);
+
+// Symbol Table serialization.
+
+void SymbolTableToString(const SymbolTable *table, string *result);
+
+SymbolTable *StringToSymbolTable(const string &str);
+
+}  // namespace fst
+
+#endif  // FST_SYMBOL_TABLE_H_
diff --git a/projects/llm_framework/include/fst/synchronize.h b/projects/llm_framework/include/fst/synchronize.h
new file mode 100644
index 00000000..0c1c6862
--- /dev/null
+++ b/projects/llm_framework/include/fst/synchronize.h
@@ -0,0 +1,414 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Synchronize an FST with bounded delay.
+
+#ifndef FST_SYNCHRONIZE_H_
+#define FST_SYNCHRONIZE_H_
+
+#include <algorithm>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+
+#include <fst/cache.h>
+#include <fst/test-properties.h>
+
+
+namespace fst {
+
+using SynchronizeFstOptions = CacheOptions;
+
+namespace internal {
+
+// Implementation class for SynchronizeFst.
+// TODO(kbg,sorenj): Refactor to guarantee thread-safety.
+
+template <class Arc>
+class SynchronizeFstImpl : public CacheImpl<Arc> {
+ public:
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FstImpl<Arc>::SetType;
+  using FstImpl<Arc>::SetProperties;
+  using FstImpl<Arc>::SetInputSymbols;
+  using FstImpl<Arc>::SetOutputSymbols;
+
+  using CacheBaseImpl<CacheState<Arc>>::EmplaceArc;
+  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
+  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
+  using CacheBaseImpl<CacheState<Arc>>::HasStart;
+  using CacheBaseImpl<CacheState<Arc>>::SetArcs;
+  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
+  using CacheBaseImpl<CacheState<Arc>>::SetStart;
+
+  using String = basic_string<Label>;
+
+  struct Element {
+    Element() {}
+
+    Element(StateId state_, const String *i, const String *o)
+        : state(state_), istring(i), ostring(o) {}
+
+    StateId state;          // Input state ID.
+    const String *istring;  // Residual input labels.
+    const String *ostring;  // Residual output labels.
+    // Residual strings are represented by const pointers to
+    // basic_string<Label> and are stored in a hash_set. The pointed
+    // memory is owned by the hash_set string_set_.
+  };
+
+  SynchronizeFstImpl(const Fst<Arc> &fst, const SynchronizeFstOptions &opts)
+      : CacheImpl<Arc>(opts), fst_(fst.Copy()) {
+    SetType("synchronize");
+    const auto props = fst.Properties(kFstProperties, false);
+    SetProperties(SynchronizeProperties(props), kCopyProperties);
+    SetInputSymbols(fst.InputSymbols());
+    SetOutputSymbols(fst.OutputSymbols());
+  }
+
+  SynchronizeFstImpl(const SynchronizeFstImpl &impl)
+      : CacheImpl<Arc>(impl), fst_(impl.fst_->Copy(true)) {
+    SetType("synchronize");
+    SetProperties(impl.Properties(), kCopyProperties);
+    SetInputSymbols(impl.InputSymbols());
+    SetOutputSymbols(impl.OutputSymbols());
+  }
+
+  ~SynchronizeFstImpl() override {
+    for (const auto *ptr : string_set_) delete ptr;
+  }
+
+  StateId Start() {
+    if (!HasStart()) {
+      auto start = fst_->Start();
+      if (start == kNoStateId) return kNoStateId;
+      const auto *empty = FindString(new String());
+      start = FindState(Element(fst_->Start(), empty, empty));
+      SetStart(start);
+    }
+    return CacheImpl<Arc>::Start();
+  }
+
+  Weight Final(StateId s) {
+    if (!HasFinal(s)) {
+      const auto &element = elements_[s];
+      const auto weight = element.state == kNoStateId
+                              ? Weight::One()
+                              : fst_->Final(element.state);
+      if ((weight != Weight::Zero()) && (element.istring)->empty() &&
+          (element.ostring)->empty()) {
+        SetFinal(s, weight);
+      } else {
+        SetFinal(s, Weight::Zero());
+      }
+    }
+    return CacheImpl<Arc>::Final(s);
+  }
+
+  size_t NumArcs(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<Arc>::NumArcs(s);
+  }
+
+  size_t NumInputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<Arc>::NumInputEpsilons(s);
+  }
+
+  size_t NumOutputEpsilons(StateId s) {
+    if (!HasArcs(s)) Expand(s);
+    return CacheImpl<Arc>::NumOutputEpsilons(s);
+  }
+
+  uint64 Properties() const override { return Properties(kFstProperties); }
+
+  // Sets error if found, returning other FST impl properties.
+  uint64 Properties(uint64 mask) const override {
+    if ((mask & kError) && fst_->Properties(kError, false)) {
+      SetProperties(kError, kError);
+    }
+    return FstImpl<Arc>::Properties(mask);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
+    if (!HasArcs(s)) Expand(s);
+    CacheImpl<Arc>::InitArcIterator(s, data);
+  }
+
+  // Returns the first character of the string obtained by concatenating the
+  // string and the label.
+  Label Car(const String *str, Label label = 0) const {
+    if (!str->empty()) {
+      return (*str)[0];
+    } else {
+      return label;
+    }
+  }
+
+  // Computes the residual string obtained by removing the first
+  // character in the concatenation of the string and the label.
+  const String *Cdr(const String *str, Label label = 0) {
+    auto *r = new String();
+    for (size_t i = 1; i < str->size(); ++i) r->push_back((*str)[i]);
+    if (label && !(str->empty())) r->push_back(label);
+    return FindString(r);
+  }
+
+  // Computes the concatenation of the string and the label.
+  const String *Concat(const String *str, Label label = 0) {
+    auto *r = new String();
+    for (size_t i = 0; i < str->size(); ++i) r->push_back((*str)[i]);
+    if (label) r->push_back(label);
+    return FindString(r);
+  }
+
+  // Tests if the concatenation of the string and label is empty.
+  bool Empty(const String *str, Label label = 0) const {
+    if (str->empty()) {
+      return label == 0;
+    } else {
+      return false;
+    }
+  }
+
+  // Finds the string pointed by s in the hash set. Transfers the pointer
+  // ownership to the hash set.
+  const String *FindString(const String *str) {
+    const auto insert_result = string_set_.insert(str);
+    if (!insert_result.second) {
+      delete str;
+    }
+    return *insert_result.first;
+  }
+
+  // Finds state corresponding to an element. Creates new state if element
+  // is not found.
+  StateId FindState(const Element &element) {
+    const auto insert_result =
+        element_map_.insert(std::make_pair(element, elements_.size()));
+    if (insert_result.second) {
+      elements_.push_back(element);
+    }
+    return insert_result.first->second;
+  }
+
+  // Computes the outgoing transitions from a state, creating new destination
+  // states as needed.
+  void Expand(StateId s) {
+    const auto element = elements_[s];
+    if (element.state != kNoStateId) {
+      for (ArcIterator<Fst<Arc>> aiter(*fst_, element.state); !aiter.Done();
+           aiter.Next()) {
+        const auto &arc = aiter.Value();
+        if (!Empty(element.istring, arc.ilabel) &&
+            !Empty(element.ostring, arc.olabel)) {
+          const auto *istring = Cdr(element.istring, arc.ilabel);
+          const auto *ostring = Cdr(element.ostring, arc.olabel);
+          EmplaceArc(s, Car(element.istring, arc.ilabel),
+                     Car(element.ostring, arc.olabel), arc.weight,
+                     FindState(Element(arc.nextstate, istring, ostring)));
+        } else {
+          const auto *istring = Concat(element.istring, arc.ilabel);
+          const auto *ostring = Concat(element.ostring, arc.olabel);
+          EmplaceArc(s, 0, 0, arc.weight,
+                     FindState(Element(arc.nextstate, istring, ostring)));
+        }
+      }
+    }
+    const auto weight = element.state == kNoStateId
+                            ? Weight::One()
+                            : fst_->Final(element.state);
+    if ((weight != Weight::Zero()) &&
+        ((element.istring)->size() + (element.ostring)->size() > 0)) {
+      const auto *istring = Cdr(element.istring);
+      const auto *ostring = Cdr(element.ostring);
+      EmplaceArc(s, Car(element.istring), Car(element.ostring), weight,
+                 FindState(Element(kNoStateId, istring, ostring)));
+    }
+    SetArcs(s);
+  }
+
+ private:
+  // Equality function for Elements; assumes strings have been hashed.
+  class ElementEqual {
+   public:
+    bool operator()(const Element &x, const Element &y) const {
+      return x.state == y.state && x.istring == y.istring &&
+             x.ostring == y.ostring;
+    }
+  };
+
+  // Hash function for Elements to FST states.
+  class ElementKey {
+   public:
+    size_t operator()(const Element &x) const {
+      size_t key = x.state;
+      key = (key << 1) ^ (x.istring)->size();
+      for (size_t i = 0; i < (x.istring)->size(); ++i) {
+        key = (key << 1) ^ (*x.istring)[i];
+      }
+      key = (key << 1) ^ (x.ostring)->size();
+      for (size_t i = 0; i < (x.ostring)->size(); ++i) {
+        key = (key << 1) ^ (*x.ostring)[i];
+      }
+      return key;
+    }
+  };
+
+  // Equality function for strings.
+  class StringEqual {
+   public:
+    bool operator()(const String *const &x, const String *const &y) const {
+      if (x->size() != y->size()) return false;
+      for (size_t i = 0; i < x->size(); ++i) {
+        if ((*x)[i] != (*y)[i]) return false;
+      }
+      return true;
+    }
+  };
+
+  // Hash function for set of strings
+  class StringKey {
+   public:
+    size_t operator()(const String *const &x) const {
+      size_t key = x->size();
+      for (size_t i = 0; i < x->size(); ++i) key = (key << 1) ^ (*x)[i];
+      return key;
+    }
+  };
+
+  using ElementMap =
+      std::unordered_map<Element, StateId, ElementKey, ElementEqual>;
+  using StringSet = std::unordered_set<const String *, StringKey, StringEqual>;
+
+  std::unique_ptr<const Fst<Arc>> fst_;
+  std::vector<Element> elements_;  // Maps FST state to Elements.
+  ElementMap element_map_;         // Maps Elements to FST state.
+  StringSet string_set_;
+};
+
+}  // namespace internal
+
+// Synchronizes a transducer. This version is a delayed FST. The result is an
+// equivalent FST that has the property that during the traversal of a path,
+// the delay is either zero or strictly increasing, where the delay is the
+// difference between the number of non-epsilon output labels and input labels
+// along the path.
+//
+// For the algorithm to terminate, the input transducer must have bounded
+// delay, i.e., the delay of every cycle must be zero.
+//
+// Complexity:
+//
+// - A has bounded delay: exponential.
+// - A does not have bounded delay: does not terminate.
+//
+// For more information, see:
+//
+// Mohri, M. 2003. Edit-distance of weighted automata: General definitions and
+// algorithms. International Journal of Computer Science 14(6): 957-982.
+//
+// This class attaches interface to implementation and handles reference
+// counting, delegating most methods to ImplToFst.
+template <class A>
+class SynchronizeFst : public ImplToFst<internal::SynchronizeFstImpl<A>> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using Store = DefaultCacheStore<Arc>;
+  using State = typename Store::State;
+  using Impl = internal::SynchronizeFstImpl<A>;
+
+  friend class ArcIterator<SynchronizeFst<A>>;
+  friend class StateIterator<SynchronizeFst<A>>;
+
+  explicit SynchronizeFst(
+      const Fst<A> &fst,
+      const SynchronizeFstOptions &opts = SynchronizeFstOptions())
+      : ImplToFst<Impl>(std::make_shared<Impl>(fst, opts)) {}
+
+  // See Fst<>::Copy() for doc.
+  SynchronizeFst(const SynchronizeFst<Arc> &fst, bool safe = false)
+      : ImplToFst<Impl>(fst, safe) {}
+
+  // Gets a copy of this SynchronizeFst. See Fst<>::Copy() for further doc.
+  SynchronizeFst<Arc> *Copy(bool safe = false) const override {
+    return new SynchronizeFst<Arc>(*this, safe);
+  }
+
+  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
+    GetMutableImpl()->InitArcIterator(s, data);
+  }
+
+ private:
+  using ImplToFst<Impl>::GetImpl;
+  using ImplToFst<Impl>::GetMutableImpl;
+
+  SynchronizeFst &operator=(const SynchronizeFst &) = delete;
+};
+
+// Specialization for SynchronizeFst.
+template <class Arc>
+class StateIterator<SynchronizeFst<Arc>>
+    : public CacheStateIterator<SynchronizeFst<Arc>> {
+ public:
+  explicit StateIterator(const SynchronizeFst<Arc> &fst)
+      : CacheStateIterator<SynchronizeFst<Arc>>(fst, fst.GetMutableImpl()) {}
+};
+
+// Specialization for SynchronizeFst.
+template <class Arc>
+class ArcIterator<SynchronizeFst<Arc>>
+    : public CacheArcIterator<SynchronizeFst<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const SynchronizeFst<Arc> &fst, StateId s)
+      : CacheArcIterator<SynchronizeFst<Arc>>(fst.GetMutableImpl(), s) {
+    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
+  }
+};
+
+template <class Arc>
+inline void SynchronizeFst<Arc>::InitStateIterator(
+    StateIteratorData<Arc> *data) const {
+  data->base = new StateIterator<SynchronizeFst<Arc>>(*this);
+}
+
+// Synchronizes a transducer. This version writes the synchronized result to a
+// MutableFst. The result will be an equivalent FST that has the property that
+// during the traversal of a path, the delay is either zero or strictly
+// increasing, where the delay is the difference between the number of
+// non-epsilon output labels and input labels along the path.
+//
+// For the algorithm to terminate, the input transducer must have bounded
+// delay, i.e., the delay of every cycle must be zero.
+//
+// Complexity:
+//
+// - A has bounded delay: exponential.
+// - A does not have bounded delay: does not terminate.
+//
+// For more information, see:
+//
+// Mohri, M. 2003. Edit-distance of weighted automata: General definitions and
+// algorithms. International Journal of Computer Science 14(6): 957-982.
+template <class Arc>
+void Synchronize(const Fst<Arc> &ifst, MutableFst<Arc> *ofst) {
+  // Caches only the last state for fastest copy.
+  const SynchronizeFstOptions opts(FLAGS_fst_default_cache_gc, 0);
+  *ofst = SynchronizeFst<Arc>(ifst, opts);
+}
+
+}  // namespace fst
+
+#endif  // FST_SYNCHRONIZE_H_
diff --git a/projects/llm_framework/include/fst/test-properties.h b/projects/llm_framework/include/fst/test-properties.h
new file mode 100644
index 00000000..677ed01b
--- /dev/null
+++ b/projects/llm_framework/include/fst/test-properties.h
@@ -0,0 +1,246 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions to manipulate and test property bits.
+
+#ifndef FST_TEST_PROPERTIES_H_
+#define FST_TEST_PROPERTIES_H_
+
+#include <unordered_set>
+
+#include <fst/flags.h>
+#include <fst/log.h>
+
+#include <fst/connect.h>
+#include <fst/dfs-visit.h>
+
+
+DECLARE_bool(fst_verify_properties);
+
+namespace fst {
+// namespace internal {
+
+// For a binary property, the bit is always returned set. For a trinary (i.e.,
+// two-bit) property, both bits are returned set iff either corresponding input
+// bit is set.
+inline uint64 KnownProperties(uint64 props) {
+  return kBinaryProperties | (props & kTrinaryProperties) |
+         ((props & kPosTrinaryProperties) << 1) |
+         ((props & kNegTrinaryProperties) >> 1);
+}
+
+// Tests compatibility between two sets of properties.
+inline bool CompatProperties(uint64 props1, uint64 props2) {
+  const auto known_props1 = KnownProperties(props1);
+  const auto known_props2 = KnownProperties(props2);
+  const auto known_props = known_props1 & known_props2;
+  const auto incompat_props = (props1 & known_props) ^ (props2 & known_props);
+  if (incompat_props) {
+    uint64 prop = 1;
+    for (int i = 0; i < 64; ++i, prop <<= 1) {
+      if (prop & incompat_props) {
+        LOG(ERROR) << "CompatProperties: Mismatch: " << PropertyNames[i]
+                   << ": props1 = " << (props1 & prop ? "true" : "false")
+                   << ", props2 = " << (props2 & prop ? "true" : "false");
+      }
+    }
+    return false;
+  } else {
+    return true;
+  }
+}
+
+// Computes FST property values defined in properties.h. The value of each
+// property indicated in the mask will be determined and returned (these will
+// never be unknown here). In the course of determining the properties
+// specifically requested in the mask, certain other properties may be
+// determined (those with little additional expense) and their values will be
+// returned as well. The complete set of known properties (whether true or
+// false) determined by this operation will be assigned to the the value pointed
+// to by KNOWN. If 'use_stored' is true, pre-computed FST properties may be used
+// when possible. 'mask & required_mask' is used to determine whether the stored
+// propertoes can be used. This routine is seldom called directly; instead it is
+// used to implement fst.Properties(mask, true).
+template <class Arc>
+uint64 ComputeProperties(const Fst<Arc> &fst, uint64 mask, uint64 *known,
+                         bool use_stored) {
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  const auto fst_props = fst.Properties(kFstProperties, false);  // FST-stored.
+  // Check stored FST properties first if allowed.
+  if (use_stored) {
+    const auto known_props = KnownProperties(fst_props);
+    // If FST contains required info, return it.
+    if ((known_props & mask) == mask) {
+      if (known) *known = known_props;
+      return fst_props;
+    }
+  }
+  // Computes (trinary) properties explicitly.
+  // Initialize with binary properties (already known).
+  uint64 comp_props = fst_props & kBinaryProperties;
+  // Computes these trinary properties with a DFS. We compute only those that
+  // need a DFS here, since we otherwise would like to avoid a DFS since its
+  // stack could grow large.
+  uint64 dfs_props = kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic |
+                     kAccessible | kNotAccessible | kCoAccessible |
+                     kNotCoAccessible;
+  std::vector<StateId> scc;
+  if (mask & (dfs_props | kWeightedCycles | kUnweightedCycles)) {
+    SccVisitor<Arc> scc_visitor(&scc, nullptr, nullptr, &comp_props);
+    DfsVisit(fst, &scc_visitor);
+  }
+  // Computes any remaining trinary properties via a state and arcs iterations
+  if (mask & ~(kBinaryProperties | dfs_props)) {
+    comp_props |= kAcceptor | kNoEpsilons | kNoIEpsilons | kNoOEpsilons |
+                  kILabelSorted | kOLabelSorted | kUnweighted | kTopSorted |
+                  kString;
+    if (mask & (kIDeterministic | kNonIDeterministic)) {
+      comp_props |= kIDeterministic;
+    }
+    if (mask & (kODeterministic | kNonODeterministic)) {
+      comp_props |= kODeterministic;
+    }
+    if (mask & (dfs_props | kWeightedCycles | kUnweightedCycles)) {
+      comp_props |= kUnweightedCycles;
+    }
+    std::unique_ptr<std::unordered_set<Label>> ilabels;
+    std::unique_ptr<std::unordered_set<Label>> olabels;
+    StateId nfinal = 0;
+    for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
+      StateId s = siter.Value();
+      Arc prev_arc;
+      // Creates these only if we need to.
+      if (mask & (kIDeterministic | kNonIDeterministic)) {
+        ilabels.reset(new std::unordered_set<Label>());
+      }
+      if (mask & (kODeterministic | kNonODeterministic)) {
+        olabels.reset(new std::unordered_set<Label>());
+      }
+      bool first_arc = true;
+      for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
+        const auto &arc = aiter.Value();
+        if (ilabels && ilabels->find(arc.ilabel) != ilabels->end()) {
+          comp_props |= kNonIDeterministic;
+          comp_props &= ~kIDeterministic;
+        }
+        if (olabels && olabels->find(arc.olabel) != olabels->end()) {
+          comp_props |= kNonODeterministic;
+          comp_props &= ~kODeterministic;
+        }
+        if (arc.ilabel != arc.olabel) {
+          comp_props |= kNotAcceptor;
+          comp_props &= ~kAcceptor;
+        }
+        if (arc.ilabel == 0 && arc.olabel == 0) {
+          comp_props |= kEpsilons;
+          comp_props &= ~kNoEpsilons;
+        }
+        if (arc.ilabel == 0) {
+          comp_props |= kIEpsilons;
+          comp_props &= ~kNoIEpsilons;
+        }
+        if (arc.olabel == 0) {
+          comp_props |= kOEpsilons;
+          comp_props &= ~kNoOEpsilons;
+        }
+        if (!first_arc) {
+          if (arc.ilabel < prev_arc.ilabel) {
+            comp_props |= kNotILabelSorted;
+            comp_props &= ~kILabelSorted;
+          }
+          if (arc.olabel < prev_arc.olabel) {
+            comp_props |= kNotOLabelSorted;
+            comp_props &= ~kOLabelSorted;
+          }
+        }
+        if (arc.weight != Weight::One() && arc.weight != Weight::Zero()) {
+          comp_props |= kWeighted;
+          comp_props &= ~kUnweighted;
+          if ((comp_props & kUnweightedCycles) &&
+              scc[s] == scc[arc.nextstate]) {
+            comp_props |= kWeightedCycles;
+            comp_props &= ~kUnweightedCycles;
+          }
+        }
+        if (arc.nextstate <= s) {
+          comp_props |= kNotTopSorted;
+          comp_props &= ~kTopSorted;
+        }
+        if (arc.nextstate != s + 1) {
+          comp_props |= kNotString;
+          comp_props &= ~kString;
+        }
+        prev_arc = arc;
+        first_arc = false;
+        if (ilabels) ilabels->insert(arc.ilabel);
+        if (olabels) olabels->insert(arc.olabel);
+      }
+
+      if (nfinal > 0) {  // Final state not last.
+        comp_props |= kNotString;
+        comp_props &= ~kString;
+      }
+      const auto final_weight = fst.Final(s);
+      if (final_weight != Weight::Zero()) {  // Final state.
+        if (final_weight != Weight::One()) {
+          comp_props |= kWeighted;
+          comp_props &= ~kUnweighted;
+        }
+        ++nfinal;
+      } else {  // Non-final state.
+        if (fst.NumArcs(s) != 1) {
+          comp_props |= kNotString;
+          comp_props &= ~kString;
+        }
+      }
+    }
+    if (fst.Start() != kNoStateId && fst.Start() != 0) {
+      comp_props |= kNotString;
+      comp_props &= ~kString;
+    }
+  }
+  if (known) *known = KnownProperties(comp_props);
+  return comp_props;
+}
+
+// This is a wrapper around ComputeProperties that will cause a fatal error if
+// the stored properties and the computed properties are incompatible when
+// FLAGS_fst_verify_properties is true. This routine is seldom called directly;
+// instead it is used to implement fst.Properties(mask, true).
+template <class Arc>
+uint64 TestProperties(const Fst<Arc> &fst, uint64 mask, uint64 *known) {
+  if (FLAGS_fst_verify_properties) {
+    const auto stored_props = fst.Properties(kFstProperties, false);
+    const auto computed_props = ComputeProperties(fst, mask, known, false);
+    if (!CompatProperties(stored_props, computed_props)) {
+      FSTERROR() << "TestProperties: stored FST properties incorrect"
+                 << " (stored: props1, computed: props2)";
+    }
+    return computed_props;
+  } else {
+    return ComputeProperties(fst, mask, known, true);
+  }
+}
+
+// If all the properties of 'fst' corresponding to 'check_mask' are known,
+// returns the stored properties. Otherwise, the properties corresponding to
+// both 'check_mask' and 'test_mask' are computed. This is used to check for
+// newly-added properties that might not be set in old binary files.
+template <class Arc>
+uint64 CheckProperties(const Fst<Arc> &fst, uint64 check_mask,
+                       uint64 test_mask) {
+  auto props = fst.Properties(kFstProperties, false);
+  if (FLAGS_fst_verify_properties) {
+    props = TestProperties(fst, check_mask | test_mask, nullptr);
+  } else if ((KnownProperties(props) & check_mask) != check_mask) {
+    props = ComputeProperties(fst, check_mask | test_mask, nullptr, false);
+  }
+  return props & (check_mask | test_mask);
+}
+
+//}  // namespace internal
+}  // namespace fst
+
+#endif  // FST_TEST_PROPERTIES_H_
diff --git a/projects/llm_framework/include/fst/test/algo_test.h b/projects/llm_framework/include/fst/test/algo_test.h
new file mode 100644
index 00000000..4c5854a0
--- /dev/null
+++ b/projects/llm_framework/include/fst/test/algo_test.h
@@ -0,0 +1,1414 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Regression test for various FST algorithms.
+
+#ifndef FST_TEST_ALGO_TEST_H_
+#define FST_TEST_ALGO_TEST_H_
+
+#include <fst/log.h>
+
+#include <fst/fstlib.h>
+#include <fst/test/rand-fst.h>
+
+DECLARE_int32(repeat);  // defined in ./algo_test.cc
+
+namespace fst {
+
+// Mapper to change input and output label of every transition into
+// epsilons.
+template <class A>
+class EpsMapper {
+ public:
+  EpsMapper() {}
+
+  A operator()(const A &arc) const {
+    return A(0, 0, arc.weight, arc.nextstate);
+  }
+
+  uint64 Properties(uint64 props) const {
+    props &= ~kNotAcceptor;
+    props |= kAcceptor;
+    props &= ~kNoIEpsilons & ~kNoOEpsilons & ~kNoEpsilons;
+    props |= kIEpsilons | kOEpsilons | kEpsilons;
+    props &= ~kNotILabelSorted & ~kNotOLabelSorted;
+    props |= kILabelSorted | kOLabelSorted;
+    return props;
+  }
+
+  MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
+
+  MapSymbolsAction InputSymbolsAction() const { return MAP_COPY_SYMBOLS; }
+
+  MapSymbolsAction OutputSymbolsAction() const { return MAP_COPY_SYMBOLS; }
+};
+
+// Generic - no lookahead.
+template <class Arc>
+void LookAheadCompose(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
+                      MutableFst<Arc> *ofst) {
+  Compose(ifst1, ifst2, ofst);
+}
+
+// Specialized and epsilon olabel acyclic - lookahead.
+void LookAheadCompose(const Fst<StdArc> &ifst1, const Fst<StdArc> &ifst2,
+                      MutableFst<StdArc> *ofst) {
+  std::vector<StdArc::StateId> order;
+  bool acyclic;
+  TopOrderVisitor<StdArc> visitor(&order, &acyclic);
+  DfsVisit(ifst1, &visitor, OutputEpsilonArcFilter<StdArc>());
+  if (acyclic) {  // no ifst1 output epsilon cycles?
+    StdOLabelLookAheadFst lfst1(ifst1);
+    StdVectorFst lfst2(ifst2);
+    LabelLookAheadRelabeler<StdArc>::Relabel(&lfst2, lfst1, true);
+    Compose(lfst1, lfst2, ofst);
+  } else {
+    Compose(ifst1, ifst2, ofst);
+  }
+}
+
+// This class tests a variety of identities and properties that must
+// hold for various algorithms on weighted FSTs.
+template <class Arc, class WeightGenerator>
+class WeightedTester {
+ public:
+  typedef typename Arc::Label Label;
+  typedef typename Arc::StateId StateId;
+  typedef typename Arc::Weight Weight;
+
+  WeightedTester(time_t seed, const Fst<Arc> &zero_fst, const Fst<Arc> &one_fst,
+                 const Fst<Arc> &univ_fst, WeightGenerator *weight_generator)
+      : seed_(seed),
+        zero_fst_(zero_fst),
+        one_fst_(one_fst),
+        univ_fst_(univ_fst),
+        weight_generator_(weight_generator) {}
+
+  void Test(const Fst<Arc> &T1, const Fst<Arc> &T2, const Fst<Arc> &T3) {
+    TestRational(T1, T2, T3);
+    TestMap(T1);
+    TestCompose(T1, T2, T3);
+    TestSort(T1);
+    TestOptimize(T1);
+    TestSearch(T1);
+  }
+
+ private:
+  // Tests rational operations with identities
+  void TestRational(const Fst<Arc> &T1, const Fst<Arc> &T2,
+                    const Fst<Arc> &T3) {
+    {
+      VLOG(1) << "Check destructive and delayed union are equivalent.";
+      VectorFst<Arc> U1(T1);
+      Union(&U1, T2);
+      UnionFst<Arc> U2(T1, T2);
+      CHECK(Equiv(U1, U2));
+    }
+
+    {
+      VLOG(1) << "Check destructive and delayed concatenation are equivalent.";
+      VectorFst<Arc> C1(T1);
+      Concat(&C1, T2);
+      ConcatFst<Arc> C2(T1, T2);
+      CHECK(Equiv(C1, C2));
+      VectorFst<Arc> C3(T2);
+      Concat(T1, &C3);
+      CHECK(Equiv(C3, C2));
+    }
+
+    {
+      VLOG(1) << "Check destructive and delayed closure* are equivalent.";
+      VectorFst<Arc> C1(T1);
+      Closure(&C1, CLOSURE_STAR);
+      ClosureFst<Arc> C2(T1, CLOSURE_STAR);
+      CHECK(Equiv(C1, C2));
+    }
+
+    {
+      VLOG(1) << "Check destructive and delayed closure+ are equivalent.";
+      VectorFst<Arc> C1(T1);
+      Closure(&C1, CLOSURE_PLUS);
+      ClosureFst<Arc> C2(T1, CLOSURE_PLUS);
+      CHECK(Equiv(C1, C2));
+    }
+
+    {
+      VLOG(1) << "Check union is associative (destructive).";
+      VectorFst<Arc> U1(T1);
+      Union(&U1, T2);
+      Union(&U1, T3);
+
+      VectorFst<Arc> U3(T2);
+      Union(&U3, T3);
+      VectorFst<Arc> U4(T1);
+      Union(&U4, U3);
+
+      CHECK(Equiv(U1, U4));
+    }
+
+    {
+      VLOG(1) << "Check union is associative (delayed).";
+      UnionFst<Arc> U1(T1, T2);
+      UnionFst<Arc> U2(U1, T3);
+
+      UnionFst<Arc> U3(T2, T3);
+      UnionFst<Arc> U4(T1, U3);
+
+      CHECK(Equiv(U2, U4));
+    }
+
+    {
+      VLOG(1) << "Check union is associative (destructive delayed).";
+      UnionFst<Arc> U1(T1, T2);
+      Union(&U1, T3);
+
+      UnionFst<Arc> U3(T2, T3);
+      UnionFst<Arc> U4(T1, U3);
+
+      CHECK(Equiv(U1, U4));
+    }
+
+    {
+      VLOG(1) << "Check concatenation is associative (destructive).";
+      VectorFst<Arc> C1(T1);
+      Concat(&C1, T2);
+      Concat(&C1, T3);
+
+      VectorFst<Arc> C3(T2);
+      Concat(&C3, T3);
+      VectorFst<Arc> C4(T1);
+      Concat(&C4, C3);
+
+      CHECK(Equiv(C1, C4));
+    }
+
+    {
+      VLOG(1) << "Check concatenation is associative (delayed).";
+      ConcatFst<Arc> C1(T1, T2);
+      ConcatFst<Arc> C2(C1, T3);
+
+      ConcatFst<Arc> C3(T2, T3);
+      ConcatFst<Arc> C4(T1, C3);
+
+      CHECK(Equiv(C2, C4));
+    }
+
+    {
+      VLOG(1) << "Check concatenation is associative (destructive delayed).";
+      ConcatFst<Arc> C1(T1, T2);
+      Concat(&C1, T3);
+
+      ConcatFst<Arc> C3(T2, T3);
+      ConcatFst<Arc> C4(T1, C3);
+
+      CHECK(Equiv(C1, C4));
+    }
+
+    if (Weight::Properties() & kLeftSemiring) {
+      VLOG(1) << "Check concatenation left distributes"
+              << " over union (destructive).";
+
+      VectorFst<Arc> U1(T1);
+      Union(&U1, T2);
+      VectorFst<Arc> C1(T3);
+      Concat(&C1, U1);
+
+      VectorFst<Arc> C2(T3);
+      Concat(&C2, T1);
+      VectorFst<Arc> C3(T3);
+      Concat(&C3, T2);
+      VectorFst<Arc> U2(C2);
+      Union(&U2, C3);
+
+      CHECK(Equiv(C1, U2));
+    }
+
+    if (Weight::Properties() & kRightSemiring) {
+      VLOG(1) << "Check concatenation right distributes"
+              << " over union (destructive).";
+      VectorFst<Arc> U1(T1);
+      Union(&U1, T2);
+      VectorFst<Arc> C1(U1);
+      Concat(&C1, T3);
+
+      VectorFst<Arc> C2(T1);
+      Concat(&C2, T3);
+      VectorFst<Arc> C3(T2);
+      Concat(&C3, T3);
+      VectorFst<Arc> U2(C2);
+      Union(&U2, C3);
+
+      CHECK(Equiv(C1, U2));
+    }
+
+    if (Weight::Properties() & kLeftSemiring) {
+      VLOG(1) << "Check concatenation left distributes over union (delayed).";
+      UnionFst<Arc> U1(T1, T2);
+      ConcatFst<Arc> C1(T3, U1);
+
+      ConcatFst<Arc> C2(T3, T1);
+      ConcatFst<Arc> C3(T3, T2);
+      UnionFst<Arc> U2(C2, C3);
+
+      CHECK(Equiv(C1, U2));
+    }
+
+    if (Weight::Properties() & kRightSemiring) {
+      VLOG(1) << "Check concatenation right distributes over union (delayed).";
+      UnionFst<Arc> U1(T1, T2);
+      ConcatFst<Arc> C1(U1, T3);
+
+      ConcatFst<Arc> C2(T1, T3);
+      ConcatFst<Arc> C3(T2, T3);
+      UnionFst<Arc> U2(C2, C3);
+
+      CHECK(Equiv(C1, U2));
+    }
+
+    if (Weight::Properties() & kLeftSemiring) {
+      VLOG(1) << "Check T T* == T+ (destructive).";
+      VectorFst<Arc> S(T1);
+      Closure(&S, CLOSURE_STAR);
+      VectorFst<Arc> C(T1);
+      Concat(&C, S);
+
+      VectorFst<Arc> P(T1);
+      Closure(&P, CLOSURE_PLUS);
+
+      CHECK(Equiv(C, P));
+    }
+
+    if (Weight::Properties() & kRightSemiring) {
+      VLOG(1) << "Check T* T == T+ (destructive).";
+      VectorFst<Arc> S(T1);
+      Closure(&S, CLOSURE_STAR);
+      VectorFst<Arc> C(S);
+      Concat(&C, T1);
+
+      VectorFst<Arc> P(T1);
+      Closure(&P, CLOSURE_PLUS);
+
+      CHECK(Equiv(C, P));
+    }
+
+    if (Weight::Properties() & kLeftSemiring) {
+      VLOG(1) << "Check T T* == T+ (delayed).";
+      ClosureFst<Arc> S(T1, CLOSURE_STAR);
+      ConcatFst<Arc> C(T1, S);
+
+      ClosureFst<Arc> P(T1, CLOSURE_PLUS);
+
+      CHECK(Equiv(C, P));
+    }
+
+    if (Weight::Properties() & kRightSemiring) {
+      VLOG(1) << "Check T* T == T+ (delayed).";
+      ClosureFst<Arc> S(T1, CLOSURE_STAR);
+      ConcatFst<Arc> C(S, T1);
+
+      ClosureFst<Arc> P(T1, CLOSURE_PLUS);
+
+      CHECK(Equiv(C, P));
+    }
+  }
+
+  // Tests map-based operations.
+  void TestMap(const Fst<Arc> &T) {
+    {
+      VLOG(1) << "Check destructive and delayed projection are equivalent.";
+      VectorFst<Arc> P1(T);
+      Project(&P1, PROJECT_INPUT);
+      ProjectFst<Arc> P2(T, PROJECT_INPUT);
+      CHECK(Equiv(P1, P2));
+    }
+
+    {
+      VLOG(1) << "Check destructive and delayed inversion are equivalent.";
+      VectorFst<Arc> I1(T);
+      Invert(&I1);
+      InvertFst<Arc> I2(T);
+      CHECK(Equiv(I1, I2));
+    }
+
+    {
+      VLOG(1) << "Check Pi_1(T) = Pi_2(T^-1) (destructive).";
+      VectorFst<Arc> P1(T);
+      VectorFst<Arc> I1(T);
+      Project(&P1, PROJECT_INPUT);
+      Invert(&I1);
+      Project(&I1, PROJECT_OUTPUT);
+      CHECK(Equiv(P1, I1));
+    }
+
+    {
+      VLOG(1) << "Check Pi_2(T) = Pi_1(T^-1) (destructive).";
+      VectorFst<Arc> P1(T);
+      VectorFst<Arc> I1(T);
+      Project(&P1, PROJECT_OUTPUT);
+      Invert(&I1);
+      Project(&I1, PROJECT_INPUT);
+      CHECK(Equiv(P1, I1));
+    }
+
+    {
+      VLOG(1) << "Check Pi_1(T) = Pi_2(T^-1) (delayed).";
+      ProjectFst<Arc> P1(T, PROJECT_INPUT);
+      InvertFst<Arc> I1(T);
+      ProjectFst<Arc> P2(I1, PROJECT_OUTPUT);
+      CHECK(Equiv(P1, P2));
+    }
+
+    {
+      VLOG(1) << "Check Pi_2(T) = Pi_1(T^-1) (delayed).";
+      ProjectFst<Arc> P1(T, PROJECT_OUTPUT);
+      InvertFst<Arc> I1(T);
+      ProjectFst<Arc> P2(I1, PROJECT_INPUT);
+      CHECK(Equiv(P1, P2));
+    }
+
+    {
+      VLOG(1) << "Check destructive relabeling";
+      static const int kNumLabels = 10;
+      // set up relabeling pairs
+      std::vector<Label> labelset(kNumLabels);
+      for (size_t i = 0; i < kNumLabels; ++i) labelset[i] = i;
+      for (size_t i = 0; i < kNumLabels; ++i) {
+        using std::swap;
+        swap(labelset[i], labelset[rand() % kNumLabels]);
+      }
+
+      std::vector<std::pair<Label, Label>> ipairs1(kNumLabels);
+      std::vector<std::pair<Label, Label>> opairs1(kNumLabels);
+      for (size_t i = 0; i < kNumLabels; ++i) {
+        ipairs1[i] = std::make_pair(i, labelset[i]);
+        opairs1[i] = std::make_pair(labelset[i], i);
+      }
+      VectorFst<Arc> R(T);
+      Relabel(&R, ipairs1, opairs1);
+
+      std::vector<std::pair<Label, Label>> ipairs2(kNumLabels);
+      std::vector<std::pair<Label, Label>> opairs2(kNumLabels);
+      for (size_t i = 0; i < kNumLabels; ++i) {
+        ipairs2[i] = std::make_pair(labelset[i], i);
+        opairs2[i] = std::make_pair(i, labelset[i]);
+      }
+      Relabel(&R, ipairs2, opairs2);
+      CHECK(Equiv(R, T));
+
+      VLOG(1) << "Check on-the-fly relabeling";
+      RelabelFst<Arc> Rdelay(T, ipairs1, opairs1);
+
+      RelabelFst<Arc> RRdelay(Rdelay, ipairs2, opairs2);
+      CHECK(Equiv(RRdelay, T));
+    }
+
+    {
+      VLOG(1) << "Check encoding/decoding (destructive).";
+      VectorFst<Arc> D(T);
+      uint32 encode_props = 0;
+      if (rand() % 2) encode_props |= kEncodeLabels;
+      if (rand() % 2) encode_props |= kEncodeWeights;
+      EncodeMapper<Arc> encoder(encode_props, ENCODE);
+      Encode(&D, &encoder);
+      Decode(&D, encoder);
+      CHECK(Equiv(D, T));
+    }
+
+    {
+      VLOG(1) << "Check encoding/decoding (delayed).";
+      uint32 encode_props = 0;
+      if (rand() % 2) encode_props |= kEncodeLabels;
+      if (rand() % 2) encode_props |= kEncodeWeights;
+      EncodeMapper<Arc> encoder(encode_props, ENCODE);
+      EncodeFst<Arc> E(T, &encoder);
+      VectorFst<Arc> Encoded(E);
+      DecodeFst<Arc> D(Encoded, encoder);
+      CHECK(Equiv(D, T));
+    }
+
+    {
+      VLOG(1) << "Check gallic mappers (constructive).";
+      ToGallicMapper<Arc> to_mapper;
+      FromGallicMapper<Arc> from_mapper;
+      VectorFst<GallicArc<Arc>> G;
+      VectorFst<Arc> F;
+      ArcMap(T, &G, to_mapper);
+      ArcMap(G, &F, from_mapper);
+      CHECK(Equiv(T, F));
+    }
+
+    {
+      VLOG(1) << "Check gallic mappers (delayed).";
+      ToGallicMapper<Arc> to_mapper;
+      FromGallicMapper<Arc> from_mapper;
+      ArcMapFst<Arc, GallicArc<Arc>, ToGallicMapper<Arc>> G(T, to_mapper);
+      ArcMapFst<GallicArc<Arc>, Arc, FromGallicMapper<Arc>> F(G, from_mapper);
+      CHECK(Equiv(T, F));
+    }
+  }
+
+  // Tests compose-based operations.
+  void TestCompose(const Fst<Arc> &T1, const Fst<Arc> &T2, const Fst<Arc> &T3) {
+    if (!(Weight::Properties() & kCommutative)) return;
+
+    VectorFst<Arc> S1(T1);
+    VectorFst<Arc> S2(T2);
+    VectorFst<Arc> S3(T3);
+
+    ILabelCompare<Arc> icomp;
+    OLabelCompare<Arc> ocomp;
+
+    ArcSort(&S1, ocomp);
+    ArcSort(&S2, ocomp);
+    ArcSort(&S3, icomp);
+
+    {
+      VLOG(1) << "Check composition is associative.";
+      ComposeFst<Arc> C1(S1, S2);
+      ComposeFst<Arc> C2(C1, S3);
+      ComposeFst<Arc> C3(S2, S3);
+      ComposeFst<Arc> C4(S1, C3);
+
+      CHECK(Equiv(C2, C4));
+    }
+
+    {
+      VLOG(1) << "Check composition left distributes over union.";
+      UnionFst<Arc> U1(S2, S3);
+      ComposeFst<Arc> C1(S1, U1);
+
+      ComposeFst<Arc> C2(S1, S2);
+      ComposeFst<Arc> C3(S1, S3);
+      UnionFst<Arc> U2(C2, C3);
+
+      CHECK(Equiv(C1, U2));
+    }
+
+    {
+      VLOG(1) << "Check composition right distributes over union.";
+      UnionFst<Arc> U1(S1, S2);
+      ComposeFst<Arc> C1(U1, S3);
+
+      ComposeFst<Arc> C2(S1, S3);
+      ComposeFst<Arc> C3(S2, S3);
+      UnionFst<Arc> U2(C2, C3);
+
+      CHECK(Equiv(C1, U2));
+    }
+
+    VectorFst<Arc> A1(S1);
+    VectorFst<Arc> A2(S2);
+    VectorFst<Arc> A3(S3);
+    Project(&A1, PROJECT_OUTPUT);
+    Project(&A2, PROJECT_INPUT);
+    Project(&A3, PROJECT_INPUT);
+
+    {
+      VLOG(1) << "Check intersection is commutative.";
+      IntersectFst<Arc> I1(A1, A2);
+      IntersectFst<Arc> I2(A2, A1);
+      CHECK(Equiv(I1, I2));
+    }
+
+    {
+      VLOG(1) << "Check all epsilon filters leads to equivalent results.";
+      typedef Matcher<Fst<Arc>> M;
+      ComposeFst<Arc> C1(S1, S2);
+      ComposeFst<Arc> C2(
+          S1, S2, ComposeFstOptions<Arc, M, AltSequenceComposeFilter<M>>());
+      ComposeFst<Arc> C3(S1, S2,
+                         ComposeFstOptions<Arc, M, MatchComposeFilter<M>>());
+
+      CHECK(Equiv(C1, C2));
+      CHECK(Equiv(C1, C3));
+
+      if ((Weight::Properties() & kIdempotent) ||
+          S1.Properties(kNoOEpsilons, false) ||
+          S2.Properties(kNoIEpsilons, false)) {
+        ComposeFst<Arc> C4(
+            S1, S2, ComposeFstOptions<Arc, M, TrivialComposeFilter<M>>());
+        CHECK(Equiv(C1, C4));
+        ComposeFst<Arc> C5(
+            S1, S2, ComposeFstOptions<Arc, M, NoMatchComposeFilter<M>>());
+        CHECK(Equiv(C1, C5));
+      }
+
+      if (S1.Properties(kNoOEpsilons, false) &&
+          S2.Properties(kNoIEpsilons, false)) {
+        ComposeFst<Arc> C6(S1, S2,
+                           ComposeFstOptions<Arc, M, NullComposeFilter<M>>());
+        CHECK(Equiv(C1, C6));
+      }
+    }
+
+    {
+      VLOG(1) << "Check look-ahead filters lead to equivalent results.";
+      VectorFst<Arc> C1, C2;
+      Compose(S1, S2, &C1);
+      LookAheadCompose(S1, S2, &C2);
+      CHECK(Equiv(C1, C2));
+    }
+  }
+
+  // Tests sorting operations
+  void TestSort(const Fst<Arc> &T) {
+    ILabelCompare<Arc> icomp;
+    OLabelCompare<Arc> ocomp;
+
+    {
+      VLOG(1) << "Check arc sorted Fst is equivalent to its input.";
+      VectorFst<Arc> S1(T);
+      ArcSort(&S1, icomp);
+      CHECK(Equiv(T, S1));
+    }
+
+    {
+      VLOG(1) << "Check destructive and delayed arcsort are equivalent.";
+      VectorFst<Arc> S1(T);
+      ArcSort(&S1, icomp);
+      ArcSortFst<Arc, ILabelCompare<Arc>> S2(T, icomp);
+      CHECK(Equiv(S1, S2));
+    }
+
+    {
+      VLOG(1) << "Check ilabel sorting vs. olabel sorting with inversions.";
+      VectorFst<Arc> S1(T);
+      VectorFst<Arc> S2(T);
+      ArcSort(&S1, icomp);
+      Invert(&S2);
+      ArcSort(&S2, ocomp);
+      Invert(&S2);
+      CHECK(Equiv(S1, S2));
+    }
+
+    {
+      VLOG(1) << "Check topologically sorted Fst is equivalent to its input.";
+      VectorFst<Arc> S1(T);
+      TopSort(&S1);
+      CHECK(Equiv(T, S1));
+    }
+
+    {
+      VLOG(1) << "Check reverse(reverse(T)) = T";
+      for (int i = 0; i < 2; ++i) {
+        VectorFst<ReverseArc<Arc>> R1;
+        VectorFst<Arc> R2;
+        bool require_superinitial = i == 1;
+        Reverse(T, &R1, require_superinitial);
+        Reverse(R1, &R2, require_superinitial);
+        CHECK(Equiv(T, R2));
+      }
+    }
+  }
+
+  // Tests optimization operations
+  void TestOptimize(const Fst<Arc> &T) {
+    uint64 tprops = T.Properties(kFstProperties, true);
+    uint64 wprops = Weight::Properties();
+
+    VectorFst<Arc> A(T);
+    Project(&A, PROJECT_INPUT);
+    {
+      VLOG(1) << "Check connected FST is equivalent to its input.";
+      VectorFst<Arc> C1(T);
+      Connect(&C1);
+      CHECK(Equiv(T, C1));
+    }
+
+    if ((wprops & kSemiring) == kSemiring &&
+        (tprops & kAcyclic || wprops & kIdempotent)) {
+      VLOG(1) << "Check epsilon-removed FST is equivalent to its input.";
+      VectorFst<Arc> R1(T);
+      RmEpsilon(&R1);
+      CHECK(Equiv(T, R1));
+
+      VLOG(1) << "Check destructive and delayed epsilon removal"
+              << "are equivalent.";
+      RmEpsilonFst<Arc> R2(T);
+      CHECK(Equiv(R1, R2));
+
+      VLOG(1) << "Check an FST with a large proportion"
+              << " of epsilon transitions:";
+      // Maps all transitions of T to epsilon-transitions and append
+      // a non-epsilon transition.
+      VectorFst<Arc> U;
+      ArcMap(T, &U, EpsMapper<Arc>());
+      VectorFst<Arc> V;
+      V.SetStart(V.AddState());
+      Arc arc(1, 1, Weight::One(), V.AddState());
+      V.AddArc(V.Start(), arc);
+      V.SetFinal(arc.nextstate, Weight::One());
+      Concat(&U, V);
+      // Check that epsilon-removal preserves the shortest-distance
+      // from the initial state to the final states.
+      std::vector<Weight> d;
+      ShortestDistance(U, &d, true);
+      Weight w = U.Start() < d.size() ? d[U.Start()] : Weight::Zero();
+      VectorFst<Arc> U1(U);
+      RmEpsilon(&U1);
+      ShortestDistance(U1, &d, true);
+      Weight w1 = U1.Start() < d.size() ? d[U1.Start()] : Weight::Zero();
+      CHECK(ApproxEqual(w, w1, kTestDelta));
+      RmEpsilonFst<Arc> U2(U);
+      ShortestDistance(U2, &d, true);
+      Weight w2 = U2.Start() < d.size() ? d[U2.Start()] : Weight::Zero();
+      CHECK(ApproxEqual(w, w2, kTestDelta));
+    }
+
+    if ((wprops & kSemiring) == kSemiring && tprops & kAcyclic) {
+      VLOG(1) << "Check determinized FSA is equivalent to its input.";
+      DeterminizeFst<Arc> D(A);
+      CHECK(Equiv(A, D));
+
+      {
+        VLOG(1) << "Check determinized FST is equivalent to its input.";
+        DeterminizeFstOptions<Arc> opts;
+        opts.type = DETERMINIZE_NONFUNCTIONAL;
+        DeterminizeFst<Arc> DT(T, opts);
+        CHECK(Equiv(T, DT));
+      }
+
+      if ((wprops & (kPath | kCommutative)) == (kPath | kCommutative)) {
+        VLOG(1) << "Check pruning in determinization";
+        VectorFst<Arc> P;
+        Weight threshold = (*weight_generator_)();
+        DeterminizeOptions<Arc> opts;
+        opts.weight_threshold = threshold;
+        Determinize(A, &P, opts);
+        CHECK(P.Properties(kIDeterministic, true));
+        CHECK(PruneEquiv(A, P, threshold));
+      }
+
+      if ((wprops & kPath) == kPath) {
+        VLOG(1) << "Check min-determinization";
+
+        // Ensures no input epsilons
+        VectorFst<Arc> R(T);
+        std::vector<std::pair<Label, Label>> ipairs, opairs;
+        ipairs.push_back(std::pair<Label, Label>(0, 1));
+        Relabel(&R, ipairs, opairs);
+
+        VectorFst<Arc> M;
+        DeterminizeOptions<Arc> opts;
+        opts.type = DETERMINIZE_DISAMBIGUATE;
+        Determinize(R, &M, opts);
+        CHECK(M.Properties(kIDeterministic, true));
+        CHECK(MinRelated(M, R));
+      }
+
+      int n;
+      {
+        VLOG(1) << "Check size(min(det(A))) <= size(det(A))"
+                << " and  min(det(A)) equiv det(A)";
+        VectorFst<Arc> M(D);
+        n = M.NumStates();
+        Minimize(&M, static_cast<MutableFst<Arc> *>(nullptr), kDelta);
+        CHECK(Equiv(D, M));
+        CHECK(M.NumStates() <= n);
+        n = M.NumStates();
+      }
+
+      if (n && (wprops & kIdempotent) == kIdempotent &&
+          A.Properties(kNoEpsilons, true)) {
+        VLOG(1) << "Check that Revuz's algorithm leads to the"
+                << " same number of states as Brozozowski's algorithm";
+
+        // Skip test if A is the empty machine or contains epsilons or
+        // if the semiring is not idempotent (to avoid floating point
+        // errors)
+        VectorFst<Arc> R;
+        Reverse(A, &R);
+        RmEpsilon(&R);
+        DeterminizeFst<Arc> DR(R);
+        VectorFst<Arc> RD;
+        Reverse(DR, &RD);
+        DeterminizeFst<Arc> DRD(RD);
+        VectorFst<Arc> M(DRD);
+        CHECK_EQ(n + 1, M.NumStates());  // Accounts for the epsilon transition
+                                         // to the initial state
+      }
+    }
+
+    if ((wprops & kSemiring) == kSemiring && tprops & kAcyclic) {
+      VLOG(1) << "Check disambiguated FSA is equivalent to its input.";
+      VectorFst<Arc> R(A), D;
+      RmEpsilon(&R);
+      Disambiguate(R, &D);
+      CHECK(Equiv(R, D));
+      VLOG(1) << "Check disambiguated FSA is unambiguous";
+      CHECK(Unambiguous(D));
+
+      /* TODO(riley): find out why this fails
+      if ((wprops & (kPath | kCommutative)) == (kPath | kCommutative)) {
+        VLOG(1)  << "Check pruning in disambiguation";
+        VectorFst<Arc> P;
+        Weight threshold = (*weight_generator_)();
+        DisambiguateOptions<Arc> opts;
+        opts.weight_threshold = threshold;
+        Disambiguate(R, &P, opts);
+        CHECK(Unambiguous(P));
+        CHECK(PruneEquiv(A, P, threshold));
+      }
+      */
+    }
+
+    if (Arc::Type() == LogArc::Type() || Arc::Type() == StdArc::Type()) {
+      VLOG(1) << "Check reweight(T) equiv T";
+      std::vector<Weight> potential;
+      VectorFst<Arc> RI(T);
+      VectorFst<Arc> RF(T);
+      while (potential.size() < RI.NumStates())
+        potential.push_back((*weight_generator_)());
+
+      Reweight(&RI, potential, REWEIGHT_TO_INITIAL);
+      CHECK(Equiv(T, RI));
+
+      Reweight(&RF, potential, REWEIGHT_TO_FINAL);
+      CHECK(Equiv(T, RF));
+    }
+
+    if ((wprops & kIdempotent) || (tprops & kAcyclic)) {
+      VLOG(1) << "Check pushed FST is equivalent to input FST.";
+      // Pushing towards the final state.
+      if (wprops & kRightSemiring) {
+        VectorFst<Arc> P1;
+        Push<Arc, REWEIGHT_TO_FINAL>(T, &P1, kPushLabels);
+        CHECK(Equiv(T, P1));
+
+        VectorFst<Arc> P2;
+        Push<Arc, REWEIGHT_TO_FINAL>(T, &P2, kPushWeights);
+        CHECK(Equiv(T, P2));
+
+        VectorFst<Arc> P3;
+        Push<Arc, REWEIGHT_TO_FINAL>(T, &P3, kPushLabels | kPushWeights);
+        CHECK(Equiv(T, P3));
+      }
+
+      // Pushing towards the initial state.
+      if (wprops & kLeftSemiring) {
+        VectorFst<Arc> P1;
+        Push<Arc, REWEIGHT_TO_INITIAL>(T, &P1, kPushLabels);
+        CHECK(Equiv(T, P1));
+
+        VectorFst<Arc> P2;
+        Push<Arc, REWEIGHT_TO_INITIAL>(T, &P2, kPushWeights);
+        CHECK(Equiv(T, P2));
+        VectorFst<Arc> P3;
+        Push<Arc, REWEIGHT_TO_INITIAL>(T, &P3, kPushLabels | kPushWeights);
+        CHECK(Equiv(T, P3));
+      }
+    }
+
+    if ((wprops & (kPath | kCommutative)) == (kPath | kCommutative)) {
+      VLOG(1) << "Check pruning algorithm";
+      {
+        VLOG(1) << "Check equiv. of constructive and destructive algorithms";
+        Weight thresold = (*weight_generator_)();
+        VectorFst<Arc> P1(T);
+        Prune(&P1, thresold);
+        VectorFst<Arc> P2;
+        Prune(T, &P2, thresold);
+        CHECK(Equiv(P1, P2));
+      }
+
+      {
+        VLOG(1) << "Check prune(reverse) equiv reverse(prune)";
+        Weight thresold = (*weight_generator_)();
+        VectorFst<ReverseArc<Arc>> R;
+        VectorFst<Arc> P1(T);
+        VectorFst<Arc> P2;
+        Prune(&P1, thresold);
+        Reverse(T, &R);
+        Prune(&R, thresold.Reverse());
+        Reverse(R, &P2);
+        CHECK(Equiv(P1, P2));
+      }
+      {
+        VLOG(1) << "Check: ShortestDistance(A - prune(A))"
+                << " > ShortestDistance(A) times Threshold";
+        Weight threshold = (*weight_generator_)();
+        VectorFst<Arc> P;
+        Prune(A, &P, threshold);
+        CHECK(PruneEquiv(A, P, threshold));
+      }
+    }
+    if (tprops & kAcyclic) {
+      VLOG(1) << "Check synchronize(T) equiv T";
+      SynchronizeFst<Arc> S(T);
+      CHECK(Equiv(T, S));
+    }
+  }
+
+  // Tests search operations
+  void TestSearch(const Fst<Arc> &T) {
+    uint64 wprops = Weight::Properties();
+
+    VectorFst<Arc> A(T);
+    Project(&A, PROJECT_INPUT);
+
+    if ((wprops & (kPath | kRightSemiring)) == (kPath | kRightSemiring)) {
+      VLOG(1) << "Check 1-best weight.";
+      VectorFst<Arc> path;
+      ShortestPath(T, &path);
+      Weight tsum = ShortestDistance(T);
+      Weight psum = ShortestDistance(path);
+      CHECK(ApproxEqual(tsum, psum, kTestDelta));
+    }
+
+    if ((wprops & (kPath | kSemiring)) == (kPath | kSemiring)) {
+      VLOG(1) << "Check n-best weights";
+      VectorFst<Arc> R(A);
+      RmEpsilon(&R, /*connect=*/ true, Arc::Weight::Zero(), kNoStateId,
+                kDelta);
+      int nshortest = rand() % kNumRandomShortestPaths + 2;
+      VectorFst<Arc> paths;
+      ShortestPath(R, &paths, nshortest, /*unique=*/ true,
+                   /*first_path=*/ false, Weight::Zero(), kNumShortestStates,
+                   kDelta);
+      std::vector<Weight> distance;
+      ShortestDistance(paths, &distance, true, kDelta);
+      StateId pstart = paths.Start();
+      if (pstart != kNoStateId) {
+        ArcIterator<Fst<Arc>> piter(paths, pstart);
+        for (; !piter.Done(); piter.Next()) {
+          StateId s = piter.Value().nextstate;
+          Weight nsum = s < distance.size()
+                            ? Times(piter.Value().weight, distance[s])
+                            : Weight::Zero();
+          VectorFst<Arc> path;
+          ShortestPath(R, &path, 1, false, false, Weight::Zero(), kNoStateId,
+                       kDelta);
+          Weight dsum = ShortestDistance(path, kDelta);
+          CHECK(ApproxEqual(nsum, dsum, kTestDelta));
+          ArcMap(&path, RmWeightMapper<Arc>());
+          VectorFst<Arc> S;
+          Difference(R, path, &S);
+          R = S;
+        }
+      }
+    }
+  }
+
+  // Tests if two FSTS are equivalent by checking if random
+  // strings from one FST are transduced the same by both FSTs.
+  template <class A>
+  bool Equiv(const Fst<A> &fst1, const Fst<A> &fst2) {
+    VLOG(1) << "Check FSTs for sanity (including property bits).";
+    CHECK(Verify(fst1));
+    CHECK(Verify(fst2));
+
+    // Ensures seed used once per instantiation.
+    static UniformArcSelector<A> uniform_selector(seed_);
+    RandGenOptions<UniformArcSelector<A>> opts(uniform_selector,
+                                               kRandomPathLength);
+    return RandEquivalent(fst1, fst2, kNumRandomPaths, kTestDelta, opts);
+  }
+
+  // Tests FSA is unambiguous
+  bool Unambiguous(const Fst<Arc> &fst) {
+    VectorFst<StdArc> sfst, dfst;
+    VectorFst<LogArc> lfst1, lfst2;
+    Map(fst, &sfst, RmWeightMapper<Arc, StdArc>());
+    Determinize(sfst, &dfst);
+    Map(fst, &lfst1, RmWeightMapper<Arc, LogArc>());
+    Map(dfst, &lfst2, RmWeightMapper<StdArc, LogArc>());
+    return Equiv(lfst1, lfst2);
+  }
+
+  // Ensures input-epsilon free transducers fst1 and fst2 have the
+  // same domain and that for each string pair '(is, os)' in fst1,
+  // '(is, os)' is the minimum weight match to 'is' in fst2.
+  template <class A>
+  bool MinRelated(const Fst<A> &fst1, const Fst<A> &fst2) {
+    // Same domain
+    VectorFst<Arc> P1(fst1), P2(fst2);
+    Project(&P1, PROJECT_INPUT);
+    Project(&P2, PROJECT_INPUT);
+    if (!Equiv(P1, P2)) {
+      LOG(ERROR) << "Inputs not equivalent";
+      return false;
+    }
+
+    // Ensures seed used once per instantiation.
+    static UniformArcSelector<A> uniform_selector(seed_);
+    RandGenOptions<UniformArcSelector<A>> opts(uniform_selector,
+                                               kRandomPathLength);
+
+    VectorFst<Arc> path, paths1, paths2;
+    for (ssize_t n = 0; n < kNumRandomPaths; ++n) {
+      RandGen(fst1, &path, opts);
+      Invert(&path);
+      Map(&path, RmWeightMapper<Arc>());
+      Compose(path, fst2, &paths1);
+      Weight sum1 = ShortestDistance(paths1);
+      Compose(paths1, path, &paths2);
+      Weight sum2 = ShortestDistance(paths2);
+      if (!ApproxEqual(Plus(sum1, sum2), sum2, kTestDelta)) {
+        LOG(ERROR) << "Sums not equivalent: " << sum1 << " " << sum2;
+        return false;
+      }
+    }
+    return true;
+  }
+
+  // Tests ShortestDistance(A - P) >=
+  // ShortestDistance(A) times Threshold.
+  template <class A>
+  bool PruneEquiv(const Fst<A> &fst, const Fst<A> &pfst, Weight threshold) {
+    VLOG(1) << "Check FSTs for sanity (including property bits).";
+    CHECK(Verify(fst));
+    CHECK(Verify(pfst));
+
+    DifferenceFst<Arc> D(fst, DeterminizeFst<Arc>(RmEpsilonFst<Arc>(
+                                  ArcMapFst<Arc, Arc, RmWeightMapper<Arc>>(
+                                      pfst, RmWeightMapper<Arc>()))));
+    Weight sum1 = Times(ShortestDistance(fst), threshold);
+    Weight sum2 = ShortestDistance(D);
+    return ApproxEqual(Plus(sum1, sum2), sum1, kTestDelta);
+  }
+
+  // Random seed.
+  int seed_;
+  // FST with no states
+  VectorFst<Arc> zero_fst_;
+  // FST with one state that accepts epsilon.
+  VectorFst<Arc> one_fst_;
+  // FST with one state that accepts all strings.
+  VectorFst<Arc> univ_fst_;
+  // Generates weights used in testing.
+  WeightGenerator *weight_generator_;
+  // Maximum random path length.
+  static const int kRandomPathLength;
+  // Number of random paths to explore.
+  static const int kNumRandomPaths;
+  // Maximum number of nshortest paths.
+  static const int kNumRandomShortestPaths;
+  // Maximum number of nshortest states.
+  static const int kNumShortestStates;
+  // Delta for equivalence tests.
+  static const float kTestDelta;
+
+  WeightedTester(const WeightedTester &) = delete;
+  WeightedTester &operator=(const WeightedTester &) = delete;
+};
+
+template <class A, class WG>
+const int WeightedTester<A, WG>::kRandomPathLength = 25;
+
+template <class A, class WG>
+const int WeightedTester<A, WG>::kNumRandomPaths = 100;
+
+template <class A, class WG>
+const int WeightedTester<A, WG>::kNumRandomShortestPaths = 100;
+
+template <class A, class WG>
+const int WeightedTester<A, WG>::kNumShortestStates = 10000;
+
+template <class A, class WG>
+const float WeightedTester<A, WG>::kTestDelta = .05;
+
+// This class tests a variety of identities and properties that must
+// hold for various algorithms on unweighted FSAs and that are not tested
+// by WeightedTester. Only the specialization does anything interesting.
+template <class Arc>
+class UnweightedTester {
+ public:
+  UnweightedTester(const Fst<Arc> &zero_fsa, const Fst<Arc> &one_fsa,
+                   const Fst<Arc> &univ_fsa) {}
+
+  void Test(const Fst<Arc> &A1, const Fst<Arc> &A2, const Fst<Arc> &A3) {}
+};
+
+// Specialization for StdArc. This should work for any commutative,
+// idempotent semiring when restricted to the unweighted case
+// (being isomorphic to the boolean semiring).
+template <>
+class UnweightedTester<StdArc> {
+ public:
+  typedef StdArc Arc;
+  typedef Arc::Label Label;
+  typedef Arc::StateId StateId;
+  typedef Arc::Weight Weight;
+
+  UnweightedTester(const Fst<Arc> &zero_fsa, const Fst<Arc> &one_fsa,
+                   const Fst<Arc> &univ_fsa)
+      : zero_fsa_(zero_fsa), one_fsa_(one_fsa), univ_fsa_(univ_fsa) {}
+
+  void Test(const Fst<Arc> &A1, const Fst<Arc> &A2, const Fst<Arc> &A3) {
+    TestRational(A1, A2, A3);
+    TestIntersect(A1, A2, A3);
+    TestOptimize(A1);
+  }
+
+ private:
+  // Tests rational operations with identities
+  void TestRational(const Fst<Arc> &A1, const Fst<Arc> &A2,
+                    const Fst<Arc> &A3) {
+    {
+      VLOG(1) << "Check the union contains its arguments (destructive).";
+      VectorFst<Arc> U(A1);
+      Union(&U, A2);
+
+      CHECK(Subset(A1, U));
+      CHECK(Subset(A2, U));
+    }
+
+    {
+      VLOG(1) << "Check the union contains its arguments (delayed).";
+      UnionFst<Arc> U(A1, A2);
+
+      CHECK(Subset(A1, U));
+      CHECK(Subset(A2, U));
+    }
+
+    {
+      VLOG(1) << "Check if A^n c A* (destructive).";
+      VectorFst<Arc> C(one_fsa_);
+      int n = rand() % 5;
+      for (int i = 0; i < n; ++i) Concat(&C, A1);
+
+      VectorFst<Arc> S(A1);
+      Closure(&S, CLOSURE_STAR);
+      CHECK(Subset(C, S));
+    }
+
+    {
+      VLOG(1) << "Check if A^n c A* (delayed).";
+      int n = rand() % 5;
+      Fst<Arc> *C = new VectorFst<Arc>(one_fsa_);
+      for (int i = 0; i < n; ++i) {
+        ConcatFst<Arc> *F = new ConcatFst<Arc>(*C, A1);
+        delete C;
+        C = F;
+      }
+      ClosureFst<Arc> S(A1, CLOSURE_STAR);
+      CHECK(Subset(*C, S));
+      delete C;
+    }
+  }
+
+  // Tests intersect-based operations.
+  void TestIntersect(const Fst<Arc> &A1, const Fst<Arc> &A2,
+                     const Fst<Arc> &A3) {
+    VectorFst<Arc> S1(A1);
+    VectorFst<Arc> S2(A2);
+    VectorFst<Arc> S3(A3);
+
+    ILabelCompare<Arc> comp;
+
+    ArcSort(&S1, comp);
+    ArcSort(&S2, comp);
+    ArcSort(&S3, comp);
+
+    {
+      VLOG(1) << "Check the intersection is contained in its arguments.";
+      IntersectFst<Arc> I1(S1, S2);
+      CHECK(Subset(I1, S1));
+      CHECK(Subset(I1, S2));
+    }
+
+    {
+      VLOG(1) << "Check union distributes over intersection.";
+      IntersectFst<Arc> I1(S1, S2);
+      UnionFst<Arc> U1(I1, S3);
+
+      UnionFst<Arc> U2(S1, S3);
+      UnionFst<Arc> U3(S2, S3);
+      ArcSortFst<Arc, ILabelCompare<Arc>> S4(U3, comp);
+      IntersectFst<Arc> I2(U2, S4);
+
+      CHECK(Equiv(U1, I2));
+    }
+
+    VectorFst<Arc> C1;
+    VectorFst<Arc> C2;
+    Complement(S1, &C1);
+    Complement(S2, &C2);
+    ArcSort(&C1, comp);
+    ArcSort(&C2, comp);
+
+    {
+      VLOG(1) << "Check S U S' = Sigma*";
+      UnionFst<Arc> U(S1, C1);
+      CHECK(Equiv(U, univ_fsa_));
+    }
+
+    {
+      VLOG(1) << "Check S n S' = {}";
+      IntersectFst<Arc> I(S1, C1);
+      CHECK(Equiv(I, zero_fsa_));
+    }
+
+    {
+      VLOG(1) << "Check (S1' U S2') == (S1 n S2)'";
+      UnionFst<Arc> U(C1, C2);
+
+      IntersectFst<Arc> I(S1, S2);
+      VectorFst<Arc> C3;
+      Complement(I, &C3);
+      CHECK(Equiv(U, C3));
+    }
+
+    {
+      VLOG(1) << "Check (S1' n S2') == (S1 U S2)'";
+      IntersectFst<Arc> I(C1, C2);
+
+      UnionFst<Arc> U(S1, S2);
+      VectorFst<Arc> C3;
+      Complement(U, &C3);
+      CHECK(Equiv(I, C3));
+    }
+  }
+
+  // Tests optimization operations
+  void TestOptimize(const Fst<Arc> &A) {
+    {
+      VLOG(1) << "Check determinized FSA is equivalent to its input.";
+      DeterminizeFst<Arc> D(A);
+      CHECK(Equiv(A, D));
+    }
+
+    {
+      VLOG(1) << "Check disambiguated FSA is equivalent to its input.";
+      VectorFst<Arc> R(A), D;
+      RmEpsilon(&R);
+
+      Disambiguate(R, &D);
+      CHECK(Equiv(R, D));
+    }
+
+    {
+      VLOG(1) << "Check minimized FSA is equivalent to its input.";
+      int n;
+      {
+        RmEpsilonFst<Arc> R(A);
+        DeterminizeFst<Arc> D(R);
+        VectorFst<Arc> M(D);
+        Minimize(&M, static_cast<MutableFst<Arc> *>(nullptr), kDelta);
+        CHECK(Equiv(A, M));
+        n = M.NumStates();
+      }
+
+      if (n) {  // Skip test if A is the empty machine
+        VLOG(1) << "Check that Hopcroft's and Revuz's algorithms lead to the"
+                << " same number of states as Brozozowski's algorithm";
+        VectorFst<Arc> R;
+        Reverse(A, &R);
+        RmEpsilon(&R);
+        DeterminizeFst<Arc> DR(R);
+        VectorFst<Arc> RD;
+        Reverse(DR, &RD);
+        DeterminizeFst<Arc> DRD(RD);
+        VectorFst<Arc> M(DRD);
+        CHECK_EQ(n + 1, M.NumStates());  // Accounts for the epsilon transition
+                                         // to the initial state
+      }
+    }
+  }
+
+  // Tests if two FSAS are equivalent.
+  bool Equiv(const Fst<Arc> &fsa1, const Fst<Arc> &fsa2) {
+    VLOG(1) << "Check FSAs for sanity (including property bits).";
+    CHECK(Verify(fsa1));
+    CHECK(Verify(fsa2));
+
+    VectorFst<Arc> vfsa1(fsa1);
+    VectorFst<Arc> vfsa2(fsa2);
+    RmEpsilon(&vfsa1);
+    RmEpsilon(&vfsa2);
+    DeterminizeFst<Arc> dfa1(vfsa1);
+    DeterminizeFst<Arc> dfa2(vfsa2);
+
+    // Test equivalence using union-find algorithm
+    bool equiv1 = Equivalent(dfa1, dfa2);
+
+    // Test equivalence by checking if (S1 - S2) U (S2 - S1) is empty
+    ILabelCompare<Arc> comp;
+    VectorFst<Arc> sdfa1(dfa1);
+    ArcSort(&sdfa1, comp);
+    VectorFst<Arc> sdfa2(dfa2);
+    ArcSort(&sdfa2, comp);
+
+    DifferenceFst<Arc> dfsa1(sdfa1, sdfa2);
+    DifferenceFst<Arc> dfsa2(sdfa2, sdfa1);
+
+    VectorFst<Arc> ufsa(dfsa1);
+    Union(&ufsa, dfsa2);
+    Connect(&ufsa);
+    bool equiv2 = ufsa.NumStates() == 0;
+
+    // Check two equivalence tests match
+    CHECK((equiv1 && equiv2) || (!equiv1 && !equiv2));
+
+    return equiv1;
+  }
+
+  // Tests if FSA1 is a subset of FSA2 (disregarding weights).
+  bool Subset(const Fst<Arc> &fsa1, const Fst<Arc> &fsa2) {
+    VLOG(1) << "Check FSAs (incl. property bits) for sanity";
+    CHECK(Verify(fsa1));
+    CHECK(Verify(fsa2));
+
+    VectorFst<StdArc> vfsa1;
+    VectorFst<StdArc> vfsa2;
+    RmEpsilon(&vfsa1);
+    RmEpsilon(&vfsa2);
+    ILabelCompare<StdArc> comp;
+    ArcSort(&vfsa1, comp);
+    ArcSort(&vfsa2, comp);
+    IntersectFst<StdArc> ifsa(vfsa1, vfsa2);
+    DeterminizeFst<StdArc> dfa1(vfsa1);
+    DeterminizeFst<StdArc> dfa2(ifsa);
+    return Equivalent(dfa1, dfa2);
+  }
+
+  // Returns complement Fsa
+  void Complement(const Fst<Arc> &ifsa, MutableFst<Arc> *ofsa) {
+    RmEpsilonFst<Arc> rfsa(ifsa);
+    DeterminizeFst<Arc> dfa(rfsa);
+    DifferenceFst<Arc> cfsa(univ_fsa_, dfa);
+    *ofsa = cfsa;
+  }
+
+  // FSA with no states
+  VectorFst<Arc> zero_fsa_;
+
+  // FSA with one state that accepts epsilon.
+  VectorFst<Arc> one_fsa_;
+
+  // FSA with one state that accepts all strings.
+  VectorFst<Arc> univ_fsa_;
+};
+
+// This class tests a variety of identities and properties that must
+// hold for various FST algorithms. It randomly generates FSTs, using
+// function object 'weight_generator' to select weights. 'WeightTester'
+// and 'UnweightedTester' are then called.
+template <class Arc, class WeightGenerator>
+class AlgoTester {
+ public:
+  typedef typename Arc::Label Label;
+  typedef typename Arc::StateId StateId;
+  typedef typename Arc::Weight Weight;
+
+  AlgoTester(WeightGenerator generator, int seed)
+      : weight_generator_(generator) {
+    one_fst_.AddState();
+    one_fst_.SetStart(0);
+    one_fst_.SetFinal(0, Weight::One());
+
+    univ_fst_.AddState();
+    univ_fst_.SetStart(0);
+    univ_fst_.SetFinal(0, Weight::One());
+    for (int i = 0; i < kNumRandomLabels; ++i)
+      univ_fst_.AddArc(0, Arc(i, i, Weight::One(), 0));
+
+    weighted_tester_ = new WeightedTester<Arc, WeightGenerator>(
+        seed, zero_fst_, one_fst_, univ_fst_, &weight_generator_);
+
+    unweighted_tester_ =
+        new UnweightedTester<Arc>(zero_fst_, one_fst_, univ_fst_);
+  }
+
+  ~AlgoTester() {
+    delete weighted_tester_;
+    delete unweighted_tester_;
+  }
+
+  void MakeRandFst(MutableFst<Arc> *fst) {
+    RandFst<Arc, WeightGenerator>(kNumRandomStates, kNumRandomArcs,
+                                  kNumRandomLabels, kAcyclicProb,
+                                  &weight_generator_, fst);
+  }
+
+  void Test() {
+    VLOG(1) << "weight type = " << Weight::Type();
+
+    for (int i = 0; i < FLAGS_repeat; ++i) {
+      // Random transducers
+      VectorFst<Arc> T1;
+      VectorFst<Arc> T2;
+      VectorFst<Arc> T3;
+      MakeRandFst(&T1);
+      MakeRandFst(&T2);
+      MakeRandFst(&T3);
+      weighted_tester_->Test(T1, T2, T3);
+
+      VectorFst<Arc> A1(T1);
+      VectorFst<Arc> A2(T2);
+      VectorFst<Arc> A3(T3);
+      Project(&A1, PROJECT_OUTPUT);
+      Project(&A2, PROJECT_INPUT);
+      Project(&A3, PROJECT_INPUT);
+      ArcMap(&A1, rm_weight_mapper_);
+      ArcMap(&A2, rm_weight_mapper_);
+      ArcMap(&A3, rm_weight_mapper_);
+      unweighted_tester_->Test(A1, A2, A3);
+    }
+  }
+
+ private:
+  // Generates weights used in testing.
+  WeightGenerator weight_generator_;
+
+  // FST with no states
+  VectorFst<Arc> zero_fst_;
+
+  // FST with one state that accepts epsilon.
+  VectorFst<Arc> one_fst_;
+
+  // FST with one state that accepts all strings.
+  VectorFst<Arc> univ_fst_;
+
+  // Tests weighted FSTs
+  WeightedTester<Arc, WeightGenerator> *weighted_tester_;
+
+  // Tests unweighted FSTs
+  UnweightedTester<Arc> *unweighted_tester_;
+
+  // Mapper to remove weights from an Fst
+  RmWeightMapper<Arc> rm_weight_mapper_;
+
+  // Maximum number of states in random test Fst.
+  static const int kNumRandomStates;
+
+  // Maximum number of arcs in random test Fst.
+  static const int kNumRandomArcs;
+
+  // Number of alternative random labels.
+  static const int kNumRandomLabels;
+
+  // Probability to force an acyclic Fst
+  static const float kAcyclicProb;
+
+  // Maximum random path length.
+  static const int kRandomPathLength;
+
+  // Number of random paths to explore.
+  static const int kNumRandomPaths;
+
+  AlgoTester(const AlgoTester &) = delete;
+  AlgoTester &operator=(const AlgoTester &) = delete;
+};
+
+template <class A, class G>
+const int AlgoTester<A, G>::kNumRandomStates = 10;
+
+template <class A, class G>
+const int AlgoTester<A, G>::kNumRandomArcs = 25;
+
+template <class A, class G>
+const int AlgoTester<A, G>::kNumRandomLabels = 5;
+
+template <class A, class G>
+const float AlgoTester<A, G>::kAcyclicProb = .25;
+
+template <class A, class G>
+const int AlgoTester<A, G>::kRandomPathLength = 25;
+
+template <class A, class G>
+const int AlgoTester<A, G>::kNumRandomPaths = 100;
+
+}  // namespace fst
+
+#endif  // FST_TEST_ALGO_TEST_H_
diff --git a/projects/llm_framework/include/fst/test/fst_test.h b/projects/llm_framework/include/fst/test/fst_test.h
new file mode 100644
index 00000000..7d536d90
--- /dev/null
+++ b/projects/llm_framework/include/fst/test/fst_test.h
@@ -0,0 +1,318 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Regression test for FST classes.
+
+#ifndef FST_TEST_FST_TEST_H_
+#define FST_TEST_FST_TEST_H_
+
+#include <fst/equal.h>
+#include <fstream>
+#include <fst/matcher.h>
+#include <fst/vector-fst.h>
+#include <fst/verify.h>
+
+DECLARE_string(tmpdir);
+
+namespace fst {
+
+// This tests an Fst F that is assumed to have a copy method from an
+// arbitrary Fst. Some test functions make further assumptions mostly
+// obvious from their name. These tests are written as member temple
+// functions that take a test fst as its argument so that different
+// Fsts in the interface hierarchy can be tested separately and so
+// that we can instantiate only those tests that make sense for a
+// particular Fst.
+template <class F>
+class FstTester {
+ public:
+  typedef typename F::Arc Arc;
+  typedef typename Arc::StateId StateId;
+  typedef typename Arc::Weight Weight;
+  typedef typename Arc::Label Label;
+
+  FstTester() {
+    VectorFst<Arc> vfst;
+    InitFst(&vfst, 128);
+    testfst_ = new F(vfst);
+  }
+
+  explicit FstTester(F *testfst) : testfst_(testfst) {}
+
+  ~FstTester() { delete testfst_; }
+
+  // This verifies the contents described in InitFst() using
+  // methods defined in a generic Fst.
+  template <class G>
+  void TestBase(const G &fst) const {
+    CHECK(Verify(fst));
+    CHECK_EQ(fst.Start(), 0);
+    StateId ns = 0;
+    StateIterator<G> siter(fst);
+    Matcher<G> matcher(fst, MATCH_INPUT);
+    MatchType match_type = matcher.Type(true);
+    for (; !siter.Done(); siter.Next()) {
+    }
+    for (siter.Reset(); !siter.Done(); siter.Next()) {
+      StateId s = siter.Value();
+      matcher.SetState(s);
+      CHECK_EQ(fst.Final(s), NthWeight(s));
+      size_t na = 0;
+      ArcIterator<G> aiter(fst, s);
+      for (; !aiter.Done(); aiter.Next()) {
+      }
+      for (aiter.Reset(); !aiter.Done(); aiter.Next()) {
+        ++na;
+        const Arc &arc = aiter.Value();
+        CHECK_EQ(arc.ilabel, na);
+        CHECK_EQ(arc.olabel, 0);
+        CHECK_EQ(arc.weight, NthWeight(na));
+        CHECK_EQ(arc.nextstate, s);
+        if (match_type == MATCH_INPUT) {
+          CHECK(matcher.Find(arc.ilabel));
+          CHECK_EQ(matcher.Value().ilabel, arc.ilabel);
+        }
+      }
+      CHECK_EQ(na, s);
+      CHECK_EQ(na, aiter.Position());
+      CHECK_EQ(fst.NumArcs(s), s);
+      CHECK_EQ(fst.NumInputEpsilons(s), 0);
+      CHECK_EQ(fst.NumOutputEpsilons(s), s);
+      CHECK(!matcher.Find(s + 1));     // out-of-range
+      CHECK(!matcher.Find(kNoLabel));  // no explicit epsilons
+      CHECK(matcher.Find(0));
+      CHECK_EQ(matcher.Value().ilabel, kNoLabel);  // implicit epsilon loop
+      ++ns;
+    }
+    CHECK(fst.Properties(kNotAcceptor, true));
+    CHECK(fst.Properties(kOEpsilons, true));
+  }
+
+  void TestBase() const { TestBase(*testfst_); }
+
+  // This verifies methods specfic to an ExpandedFst.
+  template <class G>
+  void TestExpanded(const G &fst) const {
+    StateId ns = 0;
+    for (StateIterator<G> siter(fst); !siter.Done(); siter.Next()) {
+      ++ns;
+    }
+    CHECK_EQ(fst.NumStates(), ns);
+    CHECK(fst.Properties(kExpanded, false));
+  }
+
+  void TestExpanded() const { TestExpanded(*testfst_); }
+
+  // This verifies methods specific to a MutableFst.
+  template <class G>
+  void TestMutable(G *fst) const {
+    for (StateIterator<G> siter(*fst); !siter.Done(); siter.Next()) {
+      StateId s = siter.Value();
+      size_t na = 0;
+      size_t ni = fst->NumInputEpsilons(s);
+      MutableArcIterator<G> aiter(fst, s);
+      for (; !aiter.Done(); aiter.Next()) {
+      }
+      for (aiter.Reset(); !aiter.Done(); aiter.Next()) {
+        ++na;
+        Arc arc = aiter.Value();
+        arc.ilabel = 0;
+        aiter.SetValue(arc);
+        arc = aiter.Value();
+        CHECK_EQ(arc.ilabel, 0);
+        CHECK_EQ(fst->NumInputEpsilons(s), ni + 1);
+        arc.ilabel = na;
+        aiter.SetValue(arc);
+        CHECK_EQ(fst->NumInputEpsilons(s), ni);
+      }
+    }
+
+    G *cfst1 = fst->Copy();
+    cfst1->DeleteStates();
+    CHECK_EQ(cfst1->NumStates(), 0);
+    delete cfst1;
+
+    G *cfst2 = fst->Copy();
+    for (StateIterator<G> siter(*cfst2); !siter.Done(); siter.Next()) {
+      StateId s = siter.Value();
+      cfst2->DeleteArcs(s);
+      CHECK_EQ(cfst2->NumArcs(s), 0);
+      CHECK_EQ(cfst2->NumInputEpsilons(s), 0);
+      CHECK_EQ(cfst2->NumOutputEpsilons(s), 0);
+    }
+    delete cfst2;
+  }
+
+  void TestMutable() { TestMutable(testfst_); }
+
+  // This verifies the copy methods.
+  template <class G>
+  void TestAssign(G *fst) const {
+    // Assignment from G
+    G afst1;
+    afst1 = *fst;
+    CHECK(Equal(*fst, afst1));
+
+    // Assignment from Fst
+    G afst2;
+    afst2 = *static_cast<const Fst<Arc> *>(fst);
+    CHECK(Equal(*fst, afst2));
+
+    // Assignment from self
+    afst2.operator=(afst2);
+    CHECK(Equal(*fst, afst2));
+  }
+
+  void TestAssign() { TestAssign(testfst_); }
+
+  // This verifies the copy methods.
+  template <class G>
+  void TestCopy(const G &fst) const {
+    // Copy from G
+    G c1fst(fst);
+    TestBase(c1fst);
+
+    // Copy from Fst
+    const G c2fst(static_cast<const Fst<Arc> &>(fst));
+    TestBase(c2fst);
+
+    // Copy from self
+    const G *c3fst = fst.Copy();
+    TestBase(*c3fst);
+    delete c3fst;
+  }
+
+  void TestCopy() const { TestCopy(*testfst_); }
+
+  // This verifies the read/write methods.
+  template <class G>
+  void TestIO(const G &fst) const {
+    const string filename = FLAGS_tmpdir + "/test.fst";
+    const string aligned = FLAGS_tmpdir + "/aligned.fst";
+    {
+      // write/read
+      CHECK(fst.Write(filename));
+      G *ffst = G::Read(filename);
+      CHECK(ffst);
+      TestBase(*ffst);
+      delete ffst;
+    }
+
+    {
+      // generic read/cast/test
+      Fst<Arc> *gfst = Fst<Arc>::Read(filename);
+      CHECK(gfst);
+      G *dfst = static_cast<G *>(gfst);
+      TestBase(*dfst);
+
+      // generic write/read/test
+      CHECK(gfst->Write(filename));
+      Fst<Arc> *hfst = Fst<Arc>::Read(filename);
+      CHECK(hfst);
+      TestBase(*hfst);
+      delete gfst;
+      delete hfst;
+    }
+
+    {
+      // check mmaping by first writing the file with the aligned attribute set
+      {
+        std::ofstream ostr(aligned);
+        FstWriteOptions opts;
+        opts.source = aligned;
+        opts.align = true;
+        CHECK(fst.Write(ostr, opts));
+      }
+      std::ifstream istr(aligned);
+      FstReadOptions opts;
+      opts.mode = FstReadOptions::ReadMode("map");
+      opts.source = aligned;
+      G *gfst = G::Read(istr, opts);
+      CHECK(gfst);
+      TestBase(*gfst);
+      delete gfst;
+    }
+
+    // check mmaping of unaligned files to make sure it does not fail.
+    {
+      {
+        std::ofstream ostr(aligned);
+        FstWriteOptions opts;
+        opts.source = aligned;
+        opts.align = false;
+        CHECK(fst.Write(ostr, opts));
+      }
+      std::ifstream istr(aligned);
+      FstReadOptions opts;
+      opts.mode = FstReadOptions::ReadMode("map");
+      opts.source = aligned;
+      G *gfst = G::Read(istr, opts);
+      CHECK(gfst);
+      TestBase(*gfst);
+      delete gfst;
+    }
+
+    // expanded write/read/test
+    if (fst.Properties(kExpanded, false)) {
+      ExpandedFst<Arc> *efst = ExpandedFst<Arc>::Read(filename);
+      CHECK(efst);
+      TestBase(*efst);
+      TestExpanded(*efst);
+      delete efst;
+    }
+
+    // mutable write/read/test
+    if (fst.Properties(kMutable, false)) {
+      MutableFst<Arc> *mfst = MutableFst<Arc>::Read(filename);
+      CHECK(mfst);
+      TestBase(*mfst);
+      TestExpanded(*mfst);
+      TestMutable(mfst);
+      delete mfst;
+    }
+  }
+
+  void TestIO() const { TestIO(*testfst_); }
+
+ private:
+  // This constructs test FSTs. Given a mutable FST, will leave
+  // the FST as follows:
+  // (I) NumStates() = nstates
+  // (II) Start() = 0
+  // (III) Final(s) =  NthWeight(s)
+  // (IV) For state s:
+  //     (a) NumArcs(s) == s
+  //     (b) For ith arc of s:
+  //         (1) ilabel = i
+  //         (2) olabel = 0
+  //         (3) weight = NthWeight(i)
+  //         (4) nextstate = s
+  void InitFst(MutableFst<Arc> *fst, size_t nstates) const {
+    fst->DeleteStates();
+    CHECK_GT(nstates, 0);
+
+    for (StateId s = 0; s < nstates; ++s) {
+      fst->AddState();
+      fst->SetFinal(s, NthWeight(s));
+      for (size_t i = 1; i <= s; ++i) {
+        Arc arc(i, 0, NthWeight(i), s);
+        fst->AddArc(s, arc);
+      }
+    }
+
+    fst->SetStart(0);
+  }
+
+  // Generates One() + ... + One() (n times)
+  Weight NthWeight(int n) const {
+    Weight w = Weight::Zero();
+    for (int i = 0; i < n; ++i) w = Plus(w, Weight::One());
+    return w;
+  }
+
+  F *testfst_;  // what we're testing
+};
+
+}  // namespace fst
+
+#endif  // FST_TEST_FST_TEST_H_
diff --git a/projects/llm_framework/include/fst/test/rand-fst.h b/projects/llm_framework/include/fst/test/rand-fst.h
new file mode 100644
index 00000000..f2f34c67
--- /dev/null
+++ b/projects/llm_framework/include/fst/test/rand-fst.h
@@ -0,0 +1,90 @@
+#ifndef FST_TEST_RAND_FST_H_
+#define FST_TEST_RAND_FST_H_
+
+#include <fst/log.h>
+#include <fst/mutable-fst.h>
+#include <fst/verify.h>
+
+namespace fst {
+
+// Generates a random FST.
+template <class Arc, class WeightGenerator>
+void RandFst(const int num_random_states, const int num_random_arcs,
+             const int num_random_labels, const float acyclic_prob,
+             WeightGenerator *weight_generator, MutableFst<Arc> *fst) {
+  typedef typename Arc::Label Label;
+  typedef typename Arc::StateId StateId;
+  typedef typename Arc::Weight Weight;
+
+  // Determines direction of the arcs wrt state numbering. This way we
+  // can force acyclicity when desired.
+  enum ArcDirection {
+    ANY_DIRECTION = 0,
+    FORWARD_DIRECTION = 1,
+    REVERSE_DIRECTION = 2,
+    NUM_DIRECTIONS = 3
+  };
+
+  ArcDirection arc_direction = ANY_DIRECTION;
+  if (rand() / (RAND_MAX + 1.0) < acyclic_prob)
+    arc_direction = rand() % 2 ? FORWARD_DIRECTION : REVERSE_DIRECTION;
+
+  fst->DeleteStates();
+  StateId ns = rand() % num_random_states;
+
+  if (ns == 0) return;
+  for (StateId s = 0; s < ns; ++s) fst->AddState();
+
+  StateId start = rand() % ns;
+  fst->SetStart(start);
+
+  size_t na = rand() % num_random_arcs;
+  for (size_t n = 0; n < na; ++n) {
+    StateId s = rand() % ns;
+    Arc arc;
+    arc.ilabel = rand() % num_random_labels;
+    arc.olabel = rand() % num_random_labels;
+    arc.weight = (*weight_generator)();
+    arc.nextstate = rand() % ns;
+
+    if ((arc_direction == FORWARD_DIRECTION ||
+         arc_direction == REVERSE_DIRECTION) &&
+        s == arc.nextstate) {
+      continue;  // skips self-loops
+    }
+
+    if ((arc_direction == FORWARD_DIRECTION && s > arc.nextstate) ||
+        (arc_direction == REVERSE_DIRECTION && s < arc.nextstate)) {
+      StateId t = s;  // reverses arcs
+      s = arc.nextstate;
+      arc.nextstate = t;
+    }
+
+    fst->AddArc(s, arc);
+  }
+
+  StateId nf = rand() % (ns + 1);
+  for (StateId n = 0; n < nf; ++n) {
+    StateId s = rand() % ns;
+    Weight final = (*weight_generator)();
+    fst->SetFinal(s, final);
+  }
+  VLOG(1) << "Check FST for sanity (including property bits).";
+  CHECK(Verify(*fst));
+
+  // Get/compute all properties.
+  uint64 props = fst->Properties(kFstProperties, true);
+
+  // Select random set of properties to be unknown.
+  uint64 mask = 0;
+  for (int n = 0; n < 8; ++n) {
+    mask |= rand() & 0xff;
+    mask <<= 8;
+  }
+  mask &= ~kTrinaryProperties;
+  fst->SetProperties(props & ~mask, mask);
+}
+
+}  // namespace fst
+
+#endif  // FST_TEST_RAND_FST_H_
diff --git a/projects/llm_framework/include/fst/test/weight-tester.h b/projects/llm_framework/include/fst/test/weight-tester.h
new file mode 100644
index 00000000..b7de665e
--- /dev/null
+++ b/projects/llm_framework/include/fst/test/weight-tester.h
@@ -0,0 +1,207 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Utility class for regression testing of FST weights.
+
+#ifndef FST_TEST_WEIGHT_TESTER_H_
+#define FST_TEST_WEIGHT_TESTER_H_
+
+#include <iostream>
+#include <sstream>
+
+#include <utility>
+
+#include <fst/log.h>
+#include <fst/weight.h>
+
+namespace fst {
+
+// This class tests a variety of identities and properties that must
+// hold for the Weight class to be well-defined. It calls function object
+// WEIGHT_GENERATOR to select weights that are used in the tests.
+template <class Weight, class WeightGenerator>
+class WeightTester {
+ public:
+  WeightTester(WeightGenerator generator)
+      : weight_generator_(std::move(generator)) {}
+
+  void Test(int iterations, bool test_division = true) {
+    for (int i = 0; i < iterations; ++i) {
+      // Selects the test weights.
+      const Weight w1(weight_generator_());
+      const Weight w2(weight_generator_());
+      const Weight w3(weight_generator_());
+
+      VLOG(1) << "weight type = " << Weight::Type();
+      VLOG(1) << "w1 = " << w1;
+      VLOG(1) << "w2 = " << w2;
+      VLOG(1) << "w3 = " << w3;
+
+      TestSemiring(w1, w2, w3);
+      if (test_division) TestDivision(w1, w2);
+      TestReverse(w1, w2);
+      TestEquality(w1, w2, w3);
+      TestIO(w1);
+      TestCopy(w1);
+    }
+  }
+
+ private:
+  // Note in the tests below we use ApproxEqual rather than == and add
+  // kDelta to inequalities where the weights might be inexact.
+
+  // Tests (Plus, Times, Zero, One) defines a commutative semiring.
+  void TestSemiring(Weight w1, Weight w2, Weight w3) {
+    // Checks that the operations are closed.
+    CHECK(Plus(w1, w2).Member());
+    CHECK(Times(w1, w2).Member());
+
+    // Checks that the operations are associative.
+    CHECK(ApproxEqual(Plus(w1, Plus(w2, w3)), Plus(Plus(w1, w2), w3)));
+    CHECK(ApproxEqual(Times(w1, Times(w2, w3)), Times(Times(w1, w2), w3)));
+
+    // Checks the identity elements.
+    CHECK(Plus(w1, Weight::Zero()) == w1);
+    CHECK(Plus(Weight::Zero(), w1) == w1);
+    CHECK(Times(w1, Weight::One()) == w1);
+    CHECK(Times(Weight::One(), w1) == w1);
+
+    // Check the no weight element.
+    CHECK(!Weight::NoWeight().Member());
+    CHECK(!Plus(w1, Weight::NoWeight()).Member());
+    CHECK(!Plus(Weight::NoWeight(), w1).Member());
+    CHECK(!Times(w1, Weight::NoWeight()).Member());
+    CHECK(!Times(Weight::NoWeight(), w1).Member());
+
+    // Checks that the operations commute.
+    CHECK(ApproxEqual(Plus(w1, w2), Plus(w2, w1)));
+
+    if (Weight::Properties() & kCommutative)
+      CHECK(ApproxEqual(Times(w1, w2), Times(w2, w1)));
+
+    // Checks Zero() is the annihilator.
+    CHECK(Times(w1, Weight::Zero()) == Weight::Zero());
+    CHECK(Times(Weight::Zero(), w1) == Weight::Zero());
+
+    // Check Power(w, 0) is Weight::One()
+    CHECK(Power(w1, 0) == Weight::One());
+
+    // Check Power(w, 1) is w
+    CHECK(Power(w1, 1) == w1);
+
+    // Check Power(w, 3) is Times(w, Times(w, w))
+    CHECK(Power(w1, 3) == Times(w1, Times(w1, w1)));
+
+    // Checks distributivity.
+    if (Weight::Properties() & kLeftSemiring) {
+      CHECK(ApproxEqual(Times(w1, Plus(w2, w3)),
+                        Plus(Times(w1, w2), Times(w1, w3))));
+    }
+    if (Weight::Properties() & kRightSemiring)
+      CHECK(ApproxEqual(Times(Plus(w1, w2), w3),
+                        Plus(Times(w1, w3), Times(w2, w3))));
+
+    if (Weight::Properties() & kIdempotent) CHECK(Plus(w1, w1) == w1);
+
+    if (Weight::Properties() & kPath)
+      CHECK(Plus(w1, w2) == w1 || Plus(w1, w2) == w2);
+
+    // Ensure weights form a left or right semiring.
+    CHECK(Weight::Properties() & (kLeftSemiring | kRightSemiring));
+
+    // Check when Times() is commutative that it is marked as a semiring.
+    if (Weight::Properties() & kCommutative)
+      CHECK(Weight::Properties() & kSemiring);
+  }
+
+  // Tests division operation.
+  void TestDivision(Weight w1, Weight w2) {
+    Weight p = Times(w1, w2);
+
+    if (Weight::Properties() & kLeftSemiring) {
+      Weight d = Divide(p, w1, DIVIDE_LEFT);
+      if (d.Member()) CHECK(ApproxEqual(p, Times(w1, d)));
+      CHECK(!Divide(w1, Weight::NoWeight(), DIVIDE_LEFT).Member());
+      CHECK(!Divide(Weight::NoWeight(), w1, DIVIDE_LEFT).Member());
+    }
+
+    if (Weight::Properties() & kRightSemiring) {
+      Weight d = Divide(p, w2, DIVIDE_RIGHT);
+      if (d.Member()) CHECK(ApproxEqual(p, Times(d, w2)));
+      CHECK(!Divide(w1, Weight::NoWeight(), DIVIDE_RIGHT).Member());
+      CHECK(!Divide(Weight::NoWeight(), w1, DIVIDE_RIGHT).Member());
+    }
+
+    if (Weight::Properties() & kCommutative) {
+      Weight d = Divide(p, w1, DIVIDE_RIGHT);
+      if (d.Member()) CHECK(ApproxEqual(p, Times(d, w1)));
+    }
+  }
+
+  // Tests reverse operation.
+  void TestReverse(Weight w1, Weight w2) {
+    typedef typename Weight::ReverseWeight ReverseWeight;
+
+    ReverseWeight rw1 = w1.Reverse();
+    ReverseWeight rw2 = w2.Reverse();
+
+    CHECK(rw1.Reverse() == w1);
+    CHECK(Plus(w1, w2).Reverse() == Plus(rw1, rw2));
+    CHECK(Times(w1, w2).Reverse() == Times(rw2, rw1));
+  }
+
+  // Tests == is an equivalence relation.
+  void TestEquality(Weight w1, Weight w2, Weight w3) {
+    // Checks reflexivity.
+    CHECK(w1 == w1);
+
+    // Checks symmetry.
+    CHECK((w1 == w2) == (w2 == w1));
+
+    // Checks transitivity.
+    if (w1 == w2 && w2 == w3) CHECK(w1 == w3);
+  }
+
+  // Tests binary serialization and textual I/O.
+  void TestIO(Weight w) {
+    // Tests binary I/O
+    {
+      std::ostringstream os;
+      w.Write(os);
+      os.flush();
+      std::istringstream is(os.str());
+      Weight v;
+      v.Read(is);
+      CHECK_EQ(w, v);
+    }
+
+    // Tests textual I/O.
+    {
+      std::ostringstream os;
+      os << w;
+      std::istringstream is(os.str());
+      Weight v(Weight::One());
+      is >> v;
+      CHECK(ApproxEqual(w, v));
+    }
+  }
+
+  // Tests copy constructor and assignment operator
+  void TestCopy(Weight w) {
+    Weight x = w;
+    CHECK(w == x);
+
+    x = Weight(w);
+    CHECK(w == x);
+
+    x.operator=(x);
+    CHECK(w == x);
+  }
+
+  // Generates weights used in testing.
+  WeightGenerator weight_generator_;
+};
+
+}  // namespace fst
+
+#endif  // FST_TEST_WEIGHT_TESTER_H_
diff --git a/projects/llm_framework/include/fst/topsort.h b/projects/llm_framework/include/fst/topsort.h
new file mode 100644
index 00000000..cae3e154
--- /dev/null
+++ b/projects/llm_framework/include/fst/topsort.h
@@ -0,0 +1,95 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Topological sort of FSTs.
+
+#ifndef FST_TOPSORT_H_
+#define FST_TOPSORT_H_
+
+#include <memory>
+#include <vector>
+
+
+#include <fst/dfs-visit.h>
+#include <fst/fst.h>
+#include <fst/statesort.h>
+
+
+namespace fst {
+
+// DFS visitor class to return topological ordering.
+template <class Arc>
+class TopOrderVisitor {
+ public:
+  using StateId = typename Arc::StateId;
+
+  // If acyclic, order[i] gives the topological position of StateId i;
+  // otherwise it is unchanged. acyclic_ will be true iff the FST has no
+  // cycles. The caller retains ownership of the state order vector.
+  TopOrderVisitor(std::vector<StateId> *order, bool *acyclic)
+      : order_(order), acyclic_(acyclic) {}
+
+  void InitVisit(const Fst<Arc> &fst) {
+    finish_.reset(new std::vector<StateId>());
+    *acyclic_ = true;
+  }
+
+  constexpr bool InitState(StateId, StateId) const { return true; }
+
+  constexpr bool TreeArc(StateId, const Arc &) const { return true; }
+
+  bool BackArc(StateId, const Arc &) { return (*acyclic_ = false); }
+
+  constexpr bool ForwardOrCrossArc(StateId, const Arc &) const { return true; }
+
+  void FinishState(StateId s, StateId, const Arc *) { finish_->push_back(s); }
+
+  void FinishVisit() {
+    if (*acyclic_) {
+      order_->clear();
+      for (StateId s = 0; s < finish_->size(); ++s) {
+        order_->push_back(kNoStateId);
+      }
+      for (StateId s = 0; s < finish_->size(); ++s) {
+        (*order_)[(*finish_)[finish_->size() - s - 1]] = s;
+      }
+    }
+    finish_.reset();
+  }
+
+ private:
+  std::vector<StateId> *order_;
+  bool *acyclic_;
+  // States in finish-time order.
+  std::unique_ptr<std::vector<StateId>> finish_;
+};
+
+// Topologically sorts its input if acyclic, modifying it. Otherwise, the input
+// is unchanged. When sorted, all transitions are from lower to higher state
+// IDs.
+//
+// Complexity:
+//
+//   Time:  O(V + E)
+//   Space: O(V + E)
+//
+// where V is the number of states and E is the number of arcs.
+template <class Arc>
+bool TopSort(MutableFst<Arc> *fst) {
+  std::vector<typename Arc::StateId> order;
+  bool acyclic;
+  TopOrderVisitor<Arc> top_order_visitor(&order, &acyclic);
+  DfsVisit(*fst, &top_order_visitor);
+  if (acyclic) {
+    StateSort(fst, order);
+    fst->SetProperties(kAcyclic | kInitialAcyclic | kTopSorted,
+                       kAcyclic | kInitialAcyclic | kTopSorted);
+  } else {
+    fst->SetProperties(kCyclic | kNotTopSorted, kCyclic | kNotTopSorted);
+  }
+  return acyclic;
+}
+
+}  // namespace fst
+
+#endif  // FST_TOPSORT_H_
diff --git a/projects/llm_framework/include/fst/tuple-weight.h b/projects/llm_framework/include/fst/tuple-weight.h
new file mode 100644
index 00000000..c8ecab76
--- /dev/null
+++ b/projects/llm_framework/include/fst/tuple-weight.h
@@ -0,0 +1,163 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Tuple weight set operation definitions.
+
+#ifndef FST_TUPLE_WEIGHT_H_
+#define FST_TUPLE_WEIGHT_H_
+
+#include <algorithm>
+#include <array>
+#include <functional>
+#include <string>
+#include <vector>
+
+#include <fst/flags.h>
+#include <fst/log.h>
+
+#include <fst/weight.h>
+
+
+namespace fst {
+
+// n-tuple weight, element of the n-th Cartesian power of W.
+template <class W, size_t n>
+class TupleWeight {
+ public:
+  using ReverseWeight = TupleWeight<typename W::ReverseWeight, n>;
+
+  using Weight = W;
+  using Index = size_t;
+
+  template <class Iterator>
+  TupleWeight(Iterator begin, Iterator end) {
+    std::copy(begin, end, values_.begin());
+  }
+
+  explicit TupleWeight(const W &weight = W::Zero()) { values_.fill(weight); }
+
+  // Initialize component `index` to `weight`; initialize all other components
+  // to `default_weight`
+  TupleWeight(Index index, const W &weight, const W &default_weight)
+      : TupleWeight(default_weight) {
+    values_[index] = weight;
+  }
+
+  static const TupleWeight<W, n> &Zero() {
+    static const TupleWeight<W, n> zero(W::Zero());
+    return zero;
+  }
+
+  static const TupleWeight<W, n> &One() {
+    static const TupleWeight<W, n> one(W::One());
+    return one;
+  }
+
+  static const TupleWeight<W, n> &NoWeight() {
+    static const TupleWeight<W, n> no_weight(W::NoWeight());
+    return no_weight;
+  }
+
+  constexpr static size_t Length() { return n; }
+
+  std::istream &Read(std::istream &istrm) {
+    for (size_t i = 0; i < n; ++i) values_[i].Read(istrm);
+    return istrm;
+  }
+
+  std::ostream &Write(std::ostream &ostrm) const {
+    for (size_t i = 0; i < n; ++i) values_[i].Write(ostrm);
+    return ostrm;
+  }
+
+  bool Member() const {
+    return std::all_of(values_.begin(), values_.end(),
+                       std::mem_fn(&W::Member));
+  }
+
+  size_t Hash() const {
+    uint64 hash = 0;
+    for (size_t i = 0; i < n; ++i) hash = 5 * hash + values_[i].Hash();
+    return size_t(hash);
+  }
+
+  TupleWeight<W, n> Quantize(float delta = kDelta) const {
+    TupleWeight<W, n> weight;
+    for (size_t i = 0; i < n; ++i) {
+      weight.values_[i] = values_[i].Quantize(delta);
+    }
+    return weight;
+  }
+
+  ReverseWeight Reverse() const {
+    TupleWeight<W, n> w;
+    for (size_t i = 0; i < n; ++i) w.values_[i] = values_[i].Reverse();
+    return w;
+  }
+
+  const W &Value(size_t i) const { return values_[i]; }
+
+  void SetValue(size_t i, const W &w) { values_[i] = w; }
+
+ private:
+  std::array<W, n> values_;
+};
+
+template <class W, size_t n>
+inline bool operator==(const TupleWeight<W, n> &w1,
+                       const TupleWeight<W, n> &w2) {
+  for (size_t i = 0; i < n; ++i) {
+    if (w1.Value(i) != w2.Value(i)) return false;
+  }
+  return true;
+}
+
+template <class W, size_t n>
+inline bool operator!=(const TupleWeight<W, n> &w1,
+                       const TupleWeight<W, n> &w2) {
+  for (size_t i = 0; i < n; ++i) {
+    if (w1.Value(i) != w2.Value(i)) return true;
+  }
+  return false;
+}
+
+template <class W, size_t n>
+inline bool ApproxEqual(const TupleWeight<W, n> &w1,
+                        const TupleWeight<W, n> &w2, float delta = kDelta) {
+  for (size_t i = 0; i < n; ++i) {
+    if (!ApproxEqual(w1.Value(i), w2.Value(i), delta)) return false;
+  }
+  return true;
+}
+
+template <class W, size_t n>
+inline std::ostream &operator<<(std::ostream &strm,
+                                const TupleWeight<W, n> &w) {
+  CompositeWeightWriter writer(strm);
+  writer.WriteBegin();
+  for (size_t i = 0; i < n; ++i) writer.WriteElement(w.Value(i));
+  writer.WriteEnd();
+  return strm;
+}
+
+template <class W, size_t n>
+inline std::istream &operator>>(std::istream &strm, TupleWeight<W, n> &w) {
+  CompositeWeightReader reader(strm);
+  reader.ReadBegin();
+  W v;
+  // Reads first n-1 elements.
+  static_assert(n > 0, "Size must be positive.");
+  for (size_t i = 0; i < n - 1; ++i) {
+    reader.ReadElement(&v);
+    w.SetValue(i, v);
+  }
+  // Reads n-th element.
+  reader.ReadElement(&v, true);
+  w.SetValue(n - 1, v);
+  reader.ReadEnd();
+  return strm;
+}
+
+}  // namespace fst
+
+#endif  // FST_TUPLE_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/types.h b/projects/llm_framework/include/fst/types.h
new file mode 100644
index 00000000..9c0b7998
--- /dev/null
+++ b/projects/llm_framework/include/fst/types.h
@@ -0,0 +1,41 @@
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Various type definitions (mostly for Google compatibility).
+
+#include <cstddef>       // For std::ptrdiff_t.
+#include <cstdlib>       // for ssize_t.
+#include <cstdint>       // for ?int*_t.
+
+#ifndef FST_LIB_TYPES_H_
+#define FST_LIB_TYPES_H_
+
+using int8 = int8_t;
+using int16 = int16_t;
+using int32 = int32_t;
+using int64 = int64_t;
+
+using uint8 = uint8_t;
+using uint16 = uint16_t;
+using uint32 = uint32_t;
+using uint64 = uint64_t;
+
+#ifdef _MSC_VER
+// Not really Windows-specific: they should have used ptrdiff_t in the first
+// place. But on Windows there has never been ssize_t.
+using ssize_t = std::ptrdiff_t;
+#endif  // _MSC_VER
+
+#endif  // FST_LIB_TYPES_H_
diff --git a/projects/llm_framework/include/fst/union-find.h b/projects/llm_framework/include/fst/union-find.h
new file mode 100644
index 00000000..b5c7f10b
--- /dev/null
+++ b/projects/llm_framework/include/fst/union-find.h
@@ -0,0 +1,84 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Union-find algorithm for dense sets of non-negative integers, implemented
+// using disjoint tree forests with rank heuristics and path compression.
+
+#ifndef FST_UNION_FIND_H_
+#define FST_UNION_FIND_H_
+
+#include <stack>
+#include <vector>
+
+namespace fst {
+
+// Union-Find algorithm for dense sets of non-negative integers.
+template <class T>
+class UnionFind {
+ public:
+  // Creates a disjoint set forest for the range [0; max); 'fail' is a value
+  // indicating that an element hasn't been initialized using MakeSet(...).
+  // The upper bound of the range can be reset (increased) using MakeSet(...).
+  UnionFind(T max, T fail) : parent_(max, fail), rank_(max), fail_(fail) {}
+
+  // Finds the representative of the set 'item' belongs to, performing path
+  // compression if necessary.
+  T FindSet(T item) {
+    if (item >= parent_.size() || item == fail_ || parent_[item] == fail_) {
+      return fail_;
+    }
+    auto *p = &parent_[item];
+    for (; *p != item; item = *p, p = &parent_[item]) exec_stack_.push(p);
+    for (; !exec_stack_.empty(); exec_stack_.pop()) *exec_stack_.top() = *p;
+    return *p;
+  }
+
+  // Creates the (destructive) union of the sets x and y belong to.
+  void Union(T x, T y) { Link(FindSet(x), FindSet(y)); }
+
+  // Initialization of an element: creates a singleton set containing 'item'.
+  // The range [0; max) is reset if item >= max.
+  T MakeSet(T item) {
+    if (item >= parent_.size()) {
+      // New value in parent_ should be initialized to fail_.
+      const auto nitem = item > 0 ? 2 * item : 2;
+      parent_.resize(nitem, fail_);
+      rank_.resize(nitem);
+    }
+    parent_[item] = item;
+    return item;
+  }
+
+  // Initialization of all elements starting from 0 to max - 1 to distinct sets.
+  void MakeAllSet(T max) {
+    parent_.resize(max);
+    for (T item = 0; item < max; ++item) parent_[item] = item;
+  }
+
+ private:
+  // Links trees rooted in 'x' and 'y'.
+  void Link(T x, T y) {
+    if (x == y) return;
+    if (rank_[x] > rank_[y]) {
+      parent_[y] = x;
+    } else {
+      parent_[x] = y;
+      if (rank_[x] == rank_[y]) {
+        ++rank_[y];
+      }
+    }
+  }
+
+  UnionFind(const UnionFind &) = delete;
+
+  UnionFind &operator=(const UnionFind &) = delete;
+
+  std::vector<T> parent_;       // Parent nodes.
+  std::vector<int> rank_;       // Rank of an element = min. depth in tree.
+  T fail_;                      // Value indicating lookup failure.
+  std::stack<T *> exec_stack_;  // Used for path compression.
+};
+
+}  // namespace fst
+
+#endif  // FST_UNION_FIND_H_
diff --git a/projects/llm_framework/include/fst/union-weight.h b/projects/llm_framework/include/fst/union-weight.h
new file mode 100644
index 00000000..bb2dea96
--- /dev/null
+++ b/projects/llm_framework/include/fst/union-weight.h
@@ -0,0 +1,505 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Union weight set and associated semiring operation definitions.
+//
+// TODO(riley): add in normalizer functor
+
+#ifndef FST_UNION_WEIGHT_H_
+#define FST_UNION_WEIGHT_H_
+
+#include <cstdlib>
+
+#include <iostream>
+#include <list>
+#include <sstream>
+#include <string>
+#include <utility>
+
+#include <fst/weight.h>
+
+
+namespace fst {
+
+// Example UnionWeightOptions for UnionWeight template below. The Merge
+// operation is used to collapse elements of the set and the Compare function
+// to efficiently implement the merge. In the simplest case, merge would just
+// apply with equality of set elements so the result is a set (and not a
+// multiset). More generally, this can be used to maintain the multiplicity or
+// other such weight associated with the set elements (cf. Gallic weights).
+
+// template <class W>
+// struct UnionWeightOptions {
+//   // Comparison function C is a total order on W that is monotonic w.r.t. to
+//   // Times: for all a, b,c != Zero(): C(a, b) => C(ca, cb) and is
+//   // anti-monotonic w.r.rt to Divide: C(a, b) => C(c/b, c/a).
+//   //
+//   // For all a, b: only one of C(a, b), C(b, a) or a ~ b must true where
+//   // ~ is an equivalence relation on W. Also we require a ~ b iff
+//   // a.Reverse() ~ b.Reverse().
+//   using Compare = NaturalLess<W>;
+//
+//   // How to combine two weights if a ~ b as above. For all a, b: a ~ b =>
+//   // merge(a, b) ~ a, Merge must define a semiring endomorphism from the
+//   // unmerged weight sets to the merged weight sets.
+//   struct Merge {
+//     W operator()(const W &w1, const W &w2) const { return w1; }
+//   };
+//
+//   // For ReverseWeight.
+//   using ReverseOptions = UnionWeightOptions<ReverseWeight>;
+// };
+
+template <class W, class O>
+class UnionWeight;
+
+template <class W, class O>
+class UnionWeightIterator;
+
+template <class W, class O>
+class UnionWeightReverseIterator;
+
+template <class W, class O>
+bool operator==(const UnionWeight<W, O> &, const UnionWeight<W, O> &);
+
+// Semiring that uses Times() and One() from W and union and the empty set
+// for Plus() and Zero(), respectively. Template argument O specifies the union
+// weight options as above.
+template <class W, class O>
+class UnionWeight {
+ public:
+  using Weight = W;
+  using Compare = typename O::Compare;
+  using Merge = typename O::Merge;
+
+  using ReverseWeight =
+      UnionWeight<typename W::ReverseWeight, typename O::ReverseOptions>;
+
+  friend class UnionWeightIterator<W, O>;
+  friend class UnionWeightReverseIterator<W, O>;
+  friend bool operator==
+      <>(const UnionWeight<W, O> &, const UnionWeight<W, O> &);
+
+  // Sets represented as first_ weight + rest_ weights. Uses first_ as
+  // NoWeight() to indicate the union weight Zero() ask the empty set. Uses
+  // rest_ containing NoWeight() to indicate the union weight NoWeight().
+  UnionWeight() : first_(W::NoWeight()) {}
+
+  explicit UnionWeight(W weight) : first_(weight) {
+    if (weight == W::NoWeight()) rest_.push_back(weight);
+  }
+
+  static const UnionWeight<W, O> &Zero() {
+    static const UnionWeight<W, O> zero(W::NoWeight());
+    return zero;
+  }
+
+  static const UnionWeight<W, O> &One() {
+    static const UnionWeight<W, O> one(W::One());
+    return one;
+  }
+
+  static const UnionWeight<W, O> &NoWeight() {
+    static const UnionWeight<W, O> no_weight(W::Zero(), W::NoWeight());
+    return no_weight;
+  }
+
+  static const string &Type() {
+    static const string *const type = new string(W::Type() + "_union");
+    return *type;
+  }
+
+  static constexpr uint64 Properties() {
+    return W::Properties() &
+           (kLeftSemiring | kRightSemiring | kCommutative | kIdempotent);
+  }
+
+  bool Member() const;
+
+  std::istream &Read(std::istream &strm);
+
+  std::ostream &Write(std::ostream &strm) const;
+
+  size_t Hash() const;
+
+  UnionWeight<W, O> Quantize(float delta = kDelta) const;
+
+  ReverseWeight Reverse() const;
+
+  // These operations combined with the UnionWeightIterator and
+  // UnionWeightReverseIterator provide the access and mutation of the union
+  // weight internal elements.
+
+  // Common initializer among constructors; clears existing UnionWeight.
+  void Clear() {
+    first_ = W::NoWeight();
+    rest_.clear();
+  }
+
+  size_t Size() const { return first_.Member() ? rest_.size() + 1 : 0; }
+
+  const W &Back() const { return rest_.empty() ? first_ : rest_.back(); }
+
+  // When srt is true, assumes elements added sorted w.r.t Compare and merging
+  // of weights performed as needed. Otherwise, just ensures first_ is the
+  // least element wrt Compare.
+  void PushBack(W weight, bool srt);
+
+  // Sorts the elements of the set. Assumes that first_, if present, is the
+  // least element.
+  void Sort() { rest_.sort(comp_); }
+
+ private:
+  W &Back() {
+    if (rest_.empty()) {
+      return first_;
+    } else {
+      return rest_.back();
+    }
+  }
+
+  UnionWeight(W w1, W w2) : first_(std::move(w1)), rest_(1, std::move(w2)) {}
+
+  W first_;            // First weight in set.
+  std::list<W> rest_;  // Remaining weights in set.
+  Compare comp_;
+  Merge merge_;
+};
+
+template <class W, class O>
+void UnionWeight<W, O>::PushBack(W weight, bool srt) {
+  if (!weight.Member()) {
+    rest_.push_back(std::move(weight));
+  } else if (!first_.Member()) {
+    first_ = std::move(weight);
+  } else if (srt) {
+    auto &back = Back();
+    if (comp_(back, weight)) {
+      rest_.push_back(std::move(weight));
+    } else {
+      back = merge_(back, std::move(weight));
+    }
+  } else {
+    if (comp_(first_, weight)) {
+      rest_.push_back(std::move(weight));
+    } else {
+      rest_.push_back(first_);
+      first_ = std::move(weight);
+    }
+  }
+}
+
+// Traverses union weight in the forward direction.
+template <class W, class O>
+class UnionWeightIterator {
+ public:
+  explicit UnionWeightIterator(const UnionWeight<W, O> &weight)
+      : first_(weight.first_),
+        rest_(weight.rest_),
+        init_(true),
+        it_(rest_.begin()) {}
+
+  bool Done() const { return init_ ? !first_.Member() : it_ == rest_.end(); }
+
+  const W &Value() const { return init_ ? first_ : *it_; }
+
+  void Next() {
+    if (init_) {
+      init_ = false;
+    } else {
+      ++it_;
+    }
+  }
+
+  void Reset() {
+    init_ = true;
+    it_ = rest_.begin();
+  }
+
+ private:
+  const W &first_;
+  const std::list<W> &rest_;
+  bool init_;  // in the initialized state?
+  typename std::list<W>::const_iterator it_;
+};
+
+// Traverses union weight in backward direction.
+template <typename L, class O>
+class UnionWeightReverseIterator {
+ public:
+  explicit UnionWeightReverseIterator(const UnionWeight<L, O> &weight)
+      : first_(weight.first_),
+        rest_(weight.rest_),
+        fin_(!first_.Member()),
+        it_(rest_.rbegin()) {}
+
+  bool Done() const { return fin_; }
+
+  const L &Value() const { return it_ == rest_.rend() ? first_ : *it_; }
+
+  void Next() {
+    if (it_ == rest_.rend()) {
+      fin_ = true;
+    } else {
+      ++it_;
+    }
+  }
+
+  void Reset() {
+    fin_ = !first_.Member();
+    it_ = rest_.rbegin();
+  }
+
+ private:
+  const L &first_;
+  const std::list<L> &rest_;
+  bool fin_;  // in the final state?
+  typename std::list<L>::const_reverse_iterator it_;
+};
+
+// UnionWeight member functions follow that require UnionWeightIterator.
+template <class W, class O>
+inline std::istream &UnionWeight<W, O>::Read(std::istream &istrm) {
+  Clear();
+  int32 size;
+  ReadType(istrm, &size);
+  for (int i = 0; i < size; ++i) {
+    W weight;
+    ReadType(istrm, &weight);
+    PushBack(weight, true);
+  }
+  return istrm;
+}
+
+template <class W, class O>
+inline std::ostream &UnionWeight<W, O>::Write(std::ostream &ostrm) const {
+  const int32 size = Size();
+  WriteType(ostrm, size);
+  for (UnionWeightIterator<W, O> it(*this); !it.Done(); it.Next()) {
+    WriteType(ostrm, it.Value());
+  }
+  return ostrm;
+}
+
+template <class W, class O>
+inline bool UnionWeight<W, O>::Member() const {
+  if (Size() <= 1) return true;
+  for (UnionWeightIterator<W, O> it(*this); !it.Done(); it.Next()) {
+    if (!it.Value().Member()) return false;
+  }
+  return true;
+}
+
+template <class W, class O>
+inline UnionWeight<W, O> UnionWeight<W, O>::Quantize(float delta) const {
+  UnionWeight<W, O> weight;
+  for (UnionWeightIterator<W, O> it(*this); !it.Done(); it.Next()) {
+    weight.PushBack(it.Value().Quantize(delta), true);
+  }
+  return weight;
+}
+
+template <class W, class O>
+inline typename UnionWeight<W, O>::ReverseWeight UnionWeight<W, O>::Reverse()
+    const {
+  ReverseWeight weight;
+  for (UnionWeightIterator<W, O> it(*this); !it.Done(); it.Next()) {
+    weight.PushBack(it.Value().Reverse(), false);
+  }
+  weight.Sort();
+  return weight;
+}
+
+template <class W, class O>
+inline size_t UnionWeight<W, O>::Hash() const {
+  size_t h = 0;
+  static constexpr int lshift = 5;
+  static constexpr int rshift = CHAR_BIT * sizeof(size_t) - lshift;
+  for (UnionWeightIterator<W, O> it(*this); !it.Done(); it.Next()) {
+    h = h << lshift ^ h >> rshift ^ it.Value().Hash();
+  }
+  return h;
+}
+
+// Requires union weight has been canonicalized.
+template <class W, class O>
+inline bool operator==(const UnionWeight<W, O> &w1,
+                       const UnionWeight<W, O> &w2) {
+  if (w1.Size() != w2.Size()) return false;
+  UnionWeightIterator<W, O> it1(w1);
+  UnionWeightIterator<W, O> it2(w2);
+  for (; !it1.Done(); it1.Next(), it2.Next()) {
+    if (it1.Value() != it2.Value()) return false;
+  }
+  return true;
+}
+
+// Requires union weight has been canonicalized.
+template <class W, class O>
+inline bool operator!=(const UnionWeight<W, O> &w1,
+                       const UnionWeight<W, O> &w2) {
+  return !(w1 == w2);
+}
+
+// Requires union weight has been canonicalized.
+template <class W, class O>
+inline bool ApproxEqual(const UnionWeight<W, O> &w1,
+                        const UnionWeight<W, O> &w2, float delta = kDelta) {
+  if (w1.Size() != w2.Size()) return false;
+  UnionWeightIterator<W, O> it1(w1);
+  UnionWeightIterator<W, O> it2(w2);
+  for (; !it1.Done(); it1.Next(), it2.Next()) {
+    if (!ApproxEqual(it1.Value(), it2.Value(), delta)) return false;
+  }
+  return true;
+}
+
+template <class W, class O>
+inline std::ostream &operator<<(std::ostream &ostrm,
+                                const UnionWeight<W, O> &weight) {
+  UnionWeightIterator<W, O> it(weight);
+  if (it.Done()) {
+    return ostrm << "EmptySet";
+  } else if (!weight.Member()) {
+    return ostrm << "BadSet";
+  } else {
+    CompositeWeightWriter writer(ostrm);
+    writer.WriteBegin();
+    for (; !it.Done(); it.Next()) writer.WriteElement(it.Value());
+    writer.WriteEnd();
+  }
+  return ostrm;
+}
+
+template <class W, class O>
+inline std::istream &operator>>(std::istream &istrm,
+                                UnionWeight<W, O> &weight) {
+  string s;
+  istrm >> s;
+  if (s == "EmptySet") {
+    weight = UnionWeight<W, O>::Zero();
+  } else if (s == "BadSet") {
+    weight = UnionWeight<W, O>::NoWeight();
+  } else {
+    weight = UnionWeight<W, O>::Zero();
+    std::istringstream sstrm(s);
+    CompositeWeightReader reader(sstrm);
+    reader.ReadBegin();
+    bool more = true;
+    while (more) {
+      W v;
+      more = reader.ReadElement(&v);
+      weight.PushBack(v, true);
+    }
+    reader.ReadEnd();
+  }
+  return istrm;
+}
+
+template <class W, class O>
+inline UnionWeight<W, O> Plus(const UnionWeight<W, O> &w1,
+                              const UnionWeight<W, O> &w2) {
+  if (!w1.Member() || !w2.Member()) return UnionWeight<W, O>::NoWeight();
+  if (w1 == UnionWeight<W, O>::Zero()) return w2;
+  if (w2 == UnionWeight<W, O>::Zero()) return w1;
+  UnionWeightIterator<W, O> it1(w1);
+  UnionWeightIterator<W, O> it2(w2);
+  UnionWeight<W, O> sum;
+  typename O::Compare comp;
+  while (!it1.Done() && !it2.Done()) {
+    const auto v1 = it1.Value();
+    const auto v2 = it2.Value();
+    if (comp(v1, v2)) {
+      sum.PushBack(v1, true);
+      it1.Next();
+    } else {
+      sum.PushBack(v2, true);
+      it2.Next();
+    }
+  }
+  for (; !it1.Done(); it1.Next()) sum.PushBack(it1.Value(), true);
+  for (; !it2.Done(); it2.Next()) sum.PushBack(it2.Value(), true);
+  return sum;
+}
+
+template <class W, class O>
+inline UnionWeight<W, O> Times(const UnionWeight<W, O> &w1,
+                               const UnionWeight<W, O> &w2) {
+  if (!w1.Member() || !w2.Member()) return UnionWeight<W, O>::NoWeight();
+  if (w1 == UnionWeight<W, O>::Zero() || w2 == UnionWeight<W, O>::Zero()) {
+    return UnionWeight<W, O>::Zero();
+  }
+  UnionWeightIterator<W, O> it1(w1);
+  UnionWeightIterator<W, O> it2(w2);
+  UnionWeight<W, O> prod1;
+  for (; !it1.Done(); it1.Next()) {
+    UnionWeight<W, O> prod2;
+    for (; !it2.Done(); it2.Next()) {
+      prod2.PushBack(Times(it1.Value(), it2.Value()), true);
+    }
+    prod1 = Plus(prod1, prod2);
+    it2.Reset();
+  }
+  return prod1;
+}
+
+template <class W, class O>
+inline UnionWeight<W, O> Divide(const UnionWeight<W, O> &w1,
+                                const UnionWeight<W, O> &w2, DivideType typ) {
+  if (!w1.Member() || !w2.Member()) return UnionWeight<W, O>::NoWeight();
+  if (w1 == UnionWeight<W, O>::Zero() || w2 == UnionWeight<W, O>::Zero()) {
+    return UnionWeight<W, O>::Zero();
+  }
+  UnionWeightIterator<W, O> it1(w1);
+  UnionWeightReverseIterator<W, O> it2(w2);
+  UnionWeight<W, O> quot;
+  if (w1.Size() == 1) {
+    for (; !it2.Done(); it2.Next()) {
+      quot.PushBack(Divide(it1.Value(), it2.Value(), typ), true);
+    }
+  } else if (w2.Size() == 1) {
+    for (; !it1.Done(); it1.Next()) {
+      quot.PushBack(Divide(it1.Value(), it2.Value(), typ), true);
+    }
+  } else {
+    quot = UnionWeight<W, O>::NoWeight();
+  }
+  return quot;
+}
+
+// This function object generates weights over the union of weights for the
+// underlying generators for the template weight types. This is intended
+// primarily for testing.
+template <class W, class O>
+class WeightGenerate<UnionWeight<W, O>> {
+ public:
+  using Weight = UnionWeight<W, O>;
+  using Generate = WeightGenerate<W>;
+
+  explicit WeightGenerate(bool allow_zero = true,
+                          size_t num_random_weights = kNumRandomWeights)
+      : generate_(false), allow_zero_(allow_zero),
+        num_random_weights_(num_random_weights) {}
+
+  Weight operator()() const {
+    const int n = rand() % (num_random_weights_ + 1);  // NOLINT
+    if (allow_zero_ && n == num_random_weights_) {
+      return Weight::Zero();
+    } else if (n % 2 == 0) {
+      return Weight(generate_());
+    } else {
+      return Plus(Weight(generate_()), Weight(generate_()));
+    }
+  }
+
+ private:
+  Generate generate_;
+  // Permits Zero() and zero divisors.
+  bool allow_zero_;
+  // The number of alternative random weights.
+  const size_t num_random_weights_;
+};
+
+}  // namespace fst
+
+#endif  // FST_UNION_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/union.h b/projects/llm_framework/include/fst/union.h
new file mode 100644
index 00000000..257099b8
--- /dev/null
+++ b/projects/llm_framework/include/fst/union.h
@@ -0,0 +1,157 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Functions and classes to compute the union of two FSTs.
+
+#ifndef FST_UNION_H_
+#define FST_UNION_H_
+
+#include <algorithm>
+#include <utility>
+#include <vector>
+
+#include <fst/mutable-fst.h>
+#include <fst/rational.h>
+
+
+namespace fst {
+
+// Computes the union (sum) of two FSTs. This version writes the union to an
+// output MutableFst. If A transduces string x to y with weight a and B
+// transduces string w to v with weight b, then their union transduces x to y
+// with weight a and w to v with weight b.
+//
+// Complexity:
+//
+//   Time: (V_2 + E_2)
+//   Space: O(V_2 + E_2)
+//
+// where Vi is the number of states, and Ei is the number of arcs, in the ith
+// FST.
+template <class Arc>
+void Union(MutableFst<Arc> *fst1, const Fst<Arc> &fst2) {
+  using Weight = typename Arc::Weight;
+  // Checks for symbol table compatibility.
+  if (!CompatSymbols(fst1->InputSymbols(), fst2.InputSymbols()) ||
+      !CompatSymbols(fst1->OutputSymbols(), fst2.OutputSymbols())) {
+    FSTERROR() << "Union: Input/output symbol tables of 1st argument "
+               << "do not match input/output symbol tables of 2nd argument";
+    fst1->SetProperties(kError, kError);
+    return;
+  }
+  const auto numstates1 = fst1->NumStates();
+  const bool initial_acyclic1 = fst1->Properties(kInitialAcyclic, true);
+  const auto props1 = fst1->Properties(kFstProperties, false);
+  const auto props2 = fst2.Properties(kFstProperties, false);
+  const auto start2 = fst2.Start();
+  if (start2 == kNoStateId) {
+    if (props2 & kError) fst1->SetProperties(kError, kError);
+    return;
+  }
+  if (fst2.Properties(kExpanded, false)) {
+    fst1->ReserveStates(numstates1 + CountStates(fst2) +
+                        (initial_acyclic1 ? 0 : 1));
+  }
+  for (StateIterator<Fst<Arc>> siter(fst2); !siter.Done(); siter.Next()) {
+    const auto s1 = fst1->AddState();
+    const auto s2 = siter.Value();
+    fst1->SetFinal(s1, fst2.Final(s2));
+    fst1->ReserveArcs(s1, fst2.NumArcs(s2));
+    for (ArcIterator<Fst<Arc>> aiter(fst2, s2); !aiter.Done(); aiter.Next()) {
+      auto arc = aiter.Value();  // Copy intended.
+      arc.nextstate += numstates1;
+      fst1->AddArc(s1, std::move(arc));
+    }
+  }
+  const auto start1 = fst1->Start();
+  if (start1 == kNoStateId) {
+    fst1->SetStart(start2);
+    fst1->SetProperties(props2, kCopyProperties);
+    return;
+  }
+  if (initial_acyclic1) {
+    fst1->AddArc(start1, Arc(0, 0, Weight::One(), start2 + numstates1));
+  } else {
+    const auto nstart1 = fst1->AddState();
+    fst1->SetStart(nstart1);
+    fst1->AddArc(nstart1, Arc(0, 0, Weight::One(), start1));
+    fst1->AddArc(nstart1, Arc(0, 0, Weight::One(), start2 + numstates1));
+  }
+  fst1->SetProperties(UnionProperties(props1, props2), kFstProperties);
+}
+
+// Computes the union of two FSTs, modifying the RationalFst argument.
+template <class Arc>
+void Union(RationalFst<Arc> *fst1, const Fst<Arc> &fst2) {
+  fst1->GetMutableImpl()->AddUnion(fst2);
+}
+
+using UnionFstOptions = RationalFstOptions;
+
+// Computes the union (sum) of two FSTs. This version is a delayed FST. If A
+// transduces string x to y with weight a and B transduces string w to v with
+// weight b, then their union transduces x to y with weight a and w to v with
+// weight b.
+//
+// Complexity:
+//
+//   Time: O(v_1 + e_1 + v_2 + e_2)
+//   Space: O(v_1 + v_2)
+//
+// where vi is the number of states visited, and ei is the number of arcs
+// visited, in the ith FST. Constant time and space to visit an input state or
+// arc is assumed and exclusive of caching.
+template <class A>
+class UnionFst : public RationalFst<A> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  UnionFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {
+    GetMutableImpl()->InitUnion(fst1, fst2);
+  }
+
+  UnionFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
+           const UnionFstOptions &opts)
+      : RationalFst<Arc>(opts) {
+    GetMutableImpl()->InitUnion(fst1, fst2);
+  }
+
+  // See Fst<>::Copy() for doc.
+  UnionFst(const UnionFst<Arc> &fst, bool safe = false)
+      : RationalFst<Arc>(fst, safe) {}
+
+  // Gets a copy of this UnionFst. See Fst<>::Copy() for further doc.
+  UnionFst<Arc> *Copy(bool safe = false) const override {
+    return new UnionFst<Arc>(*this, safe);
+  }
+
+ private:
+  using ImplToFst<internal::RationalFstImpl<Arc>>::GetImpl;
+  using ImplToFst<internal::RationalFstImpl<Arc>>::GetMutableImpl;
+};
+
+// Specialization for UnionFst.
+template <class Arc>
+class StateIterator<UnionFst<Arc>> : public StateIterator<RationalFst<Arc>> {
+ public:
+  explicit StateIterator(const UnionFst<Arc> &fst)
+      : StateIterator<RationalFst<Arc>>(fst) {}
+};
+
+// Specialization for UnionFst.
+template <class Arc>
+class ArcIterator<UnionFst<Arc>> : public ArcIterator<RationalFst<Arc>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const UnionFst<Arc> &fst, StateId s)
+      : ArcIterator<RationalFst<Arc>>(fst, s) {}
+};
+
+using StdUnionFst = UnionFst<StdArc>;
+
+}  // namespace fst
+
+#endif  // FST_UNION_H_
diff --git a/projects/llm_framework/include/fst/util.h b/projects/llm_framework/include/fst/util.h
new file mode 100644
index 00000000..c7520213
--- /dev/null
+++ b/projects/llm_framework/include/fst/util.h
@@ -0,0 +1,400 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// FST utility inline definitions.
+
+#ifndef FST_UTIL_H_
+#define FST_UTIL_H_
+
+#include <iostream>
+#include <iterator>
+#include <list>
+#include <map>
+#include <set>
+#include <sstream>
+#include <string>
+#include <type_traits>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+
+#include <fst/compat.h>
+#include <fst/types.h>
+#include <fst/log.h>
+#include <fstream>
+
+#include <fst/flags.h>
+#include <unordered_map>
+
+
+// Utility for error handling.
+
+DECLARE_bool(fst_error_fatal);
+
+#define FSTERROR() \
+  (FLAGS_fst_error_fatal ? LOG(FATAL) : LOG(ERROR))
+
+namespace fst {
+
+// Utility for type I/O.
+
+// Reads types from an input stream.
+
+// Generic case.
+template <class T,
+    typename std::enable_if<std::is_class<T>::value, T>::type* = nullptr>
+inline std::istream &ReadType(std::istream &strm, T *t) {
+  return t->Read(strm);
+}
+
+// Numeric (boolean, integral, floating-point) case.
+template <class T,
+    typename std::enable_if<std::is_arithmetic<T>::value, T>::type* = nullptr>
+inline std::istream &ReadType(std::istream &strm, T *t) {
+  return strm.read(reinterpret_cast<char *>(t), sizeof(T)); \
+}
+
+// String case.
+inline std::istream &ReadType(std::istream &strm, string *s) {  // NOLINT
+  s->clear();
+  int32 ns = 0;
+  strm.read(reinterpret_cast<char *>(&ns), sizeof(ns));
+  for (int32 i = 0; i < ns; ++i) {
+    char c;
+    strm.read(&c, 1);
+    *s += c;
+  }
+  return strm;
+}
+
+// Pair case.
+template <typename S, typename T>
+inline std::istream &ReadType(std::istream &strm, std::pair<S, T> *p) {
+  ReadType(strm, &p->first);
+  ReadType(strm, &p->second);
+  return strm;
+}
+
+template <typename S, typename T>
+inline std::istream &ReadType(std::istream &strm, std::pair<const S, T> *p) {
+  ReadType(strm, const_cast<S *>(&p->first));
+  ReadType(strm, &p->second);
+  return strm;
+}
+
+namespace internal {
+template <class C, class ReserveFn>
+std::istream &ReadContainerType(std::istream &strm, C *c, ReserveFn reserve) {
+  c->clear();
+  int64 n = 0;
+  ReadType(strm, &n);
+  reserve(c, n);
+  auto insert = std::inserter(*c, c->begin());
+  for (int64 i = 0; i < n; ++i) {
+    typename C::value_type value;
+    ReadType(strm, &value);
+    *insert = value;
+  }
+  return strm;
+}
+}  // namespace internal
+
+template <typename T, typename A>
+std::istream &ReadType(std::istream &strm, std::vector<T,A> *c) {
+  return internal::ReadContainerType(
+      strm, c, [](decltype(c) v, int n) { v->reserve(n); });
+}
+
+template <typename T, typename A>
+std::istream &ReadType(std::istream &strm, std::list<T,A> *c) {
+  return internal::ReadContainerType(strm, c, [](decltype(c) v, int n) {});
+}
+
+template <typename T, typename L, typename A>
+std::istream &ReadType(std::istream &strm, std::set<T,L,A> *c) {
+  return internal::ReadContainerType(strm, c, [](decltype(c) v, int n) {});
+}
+
+template <typename K, typename V, typename L, typename A>
+std::istream &ReadType(std::istream &strm, std::map<K,V,L,A> *c) {
+  return internal::ReadContainerType(strm, c, [](decltype(c) v, int n) {});
+}
+
+template <typename T, typename H, typename E, typename A>
+std::istream &ReadType(std::istream &strm, std::unordered_set<T,H,E,A> *c) {
+  return internal::ReadContainerType(
+      strm, c, [](decltype(c) v, int n) { v->reserve(n); });
+}
+
+template <typename K, typename V, typename H, typename E, typename A>
+std::istream &ReadType(std::istream &strm, std::unordered_map<K,V,H,E,A> *c) {
+  return internal::ReadContainerType(
+      strm, c, [](decltype(c) v, int n) { v->reserve(n); });
+}
+
+// Writes types to an output stream.
+
+// Generic case.
+template <class T,
+    typename std::enable_if<std::is_class<T>::value, T>::type* = nullptr>
+inline std::ostream &WriteType(std::ostream &strm, const T t) {
+  t.Write(strm);
+  return strm;
+}
+
+// Numeric (boolean, integral, floating-point) case.
+template <class T,
+    typename std::enable_if<std::is_arithmetic<T>::value, T>::type* = nullptr>
+inline std::ostream &WriteType(std::ostream &strm, const T t) {
+  return strm.write(reinterpret_cast<const char *>(&t), sizeof(T));
+}
+
+// String case.
+inline std::ostream &WriteType(std::ostream &strm, const string &s) {  // NOLINT
+  int32 ns = s.size();
+  strm.write(reinterpret_cast<const char *>(&ns), sizeof(ns));
+  return strm.write(s.data(), ns);
+}
+
+// Pair case.
+template <typename S, typename T>
+inline std::ostream &WriteType(std::ostream &strm,
+                               const std::pair<S, T> &p) {  // NOLINT
+  WriteType(strm, p.first);
+  WriteType(strm, p.second);
+  return strm;
+}
+
+namespace internal {
+template <class C>
+std::ostream &WriteContainer(std::ostream &strm, const C &c) {
+  const int64 n = c.size();
+  WriteType(strm, n);
+  for (const auto &e : c) {
+    WriteType(strm, e);
+  }
+  return strm;
+}
+}  // namespace internal
+
+template <typename T, typename A>
+std::ostream &WriteType(std::ostream &strm, const std::vector<T,A> &c) {
+  return internal::WriteContainer(strm, c);
+}
+
+template <typename T, typename A>
+std::ostream &WriteType(std::ostream &strm, const std::list<T,A> &c) {
+  return internal::WriteContainer(strm, c);
+}
+
+template <typename T, typename L, typename A>
+std::ostream &WriteType(std::ostream &strm, const std::set<T,L,A> &c) {
+  return internal::WriteContainer(strm, c);
+}
+
+template <typename K, typename V, typename L, typename A>
+std::ostream &WriteType(std::ostream &strm, const std::map<K,V,L,A> &c) {
+  return internal::WriteContainer(strm, c);
+}
+
+template <typename T, typename H, typename E, typename A>
+std::ostream &WriteType(std::ostream &strm, const std::unordered_set<T,H,E,A> &c) {
+  return internal::WriteContainer(strm, c);
+}
+
+template <typename K, typename V, typename H, typename E, typename A>
+std::ostream &WriteType(std::ostream &strm, const std::unordered_map<K,V,H,E,A> &c) {
+  return internal::WriteContainer(strm, c);
+}
+
+// Utilities for converting between int64 or Weight and string.
+
+int64 StrToInt64(const string &s, const string &src, size_t nline,
+                 bool allow_negative, bool *error = nullptr);
+
+template <typename Weight>
+Weight StrToWeight(const string &s, const string &src, size_t nline) {
+  Weight w;
+  std::istringstream strm(s);
+  strm >> w;
+  if (!strm) {
+    FSTERROR() << "StrToWeight: Bad weight = \"" << s << "\", source = " << src
+               << ", line = " << nline;
+    return Weight::NoWeight();
+  }
+  return w;
+}
+
+template <typename Weight>
+void WeightToStr(Weight w, string *s) {
+  std::ostringstream strm;
+  strm.precision(9);
+  strm << w;
+  s->append(strm.str().data(), strm.str().size());
+}
+
+// Utilities for reading/writing integer pairs (typically labels)
+
+// Modifies line using a vector of pointers to a buffer beginning with line.
+void SplitString(char *line, const char *delim, std::vector<char *> *vec,
+                 bool omit_empty_strings);
+
+template <typename I>
+bool ReadIntPairs(const string &filename, std::vector<std::pair<I, I>> *pairs,
+                  bool allow_negative = false) {
+  std::ifstream strm(filename, std::ios_base::in);
+  if (!strm) {
+    LOG(ERROR) << "ReadIntPairs: Can't open file: " << filename;
+    return false;
+  }
+  const int kLineLen = 8096;
+  char line[kLineLen];
+  size_t nline = 0;
+  pairs->clear();
+  while (strm.getline(line, kLineLen)) {
+    ++nline;
+    std::vector<char *> col;
+    SplitString(line, "\n\t ", &col, true);
+    // empty line or comment?
+    if (col.empty() || col[0][0] == '\0' || col[0][0] == '#') continue;
+    if (col.size() != 2) {
+      LOG(ERROR) << "ReadIntPairs: Bad number of columns, "
+                 << "file = " << filename << ", line = " << nline;
+      return false;
+    }
+    bool err;
+    I i1 = StrToInt64(col[0], filename, nline, allow_negative, &err);
+    if (err) return false;
+    I i2 = StrToInt64(col[1], filename, nline, allow_negative, &err);
+    if (err) return false;
+    pairs->push_back(std::make_pair(i1, i2));
+  }
+  return true;
+}
+
+template <typename I>
+bool WriteIntPairs(const string &filename,
+                   const std::vector<std::pair<I, I>> &pairs) {
+  std::ostream *strm = &std::cout;
+  if (!filename.empty()) {
+    strm = new std::ofstream(filename);
+    if (!*strm) {
+      LOG(ERROR) << "WriteIntPairs: Can't open file: " << filename;
+      return false;
+    }
+  }
+  for (ssize_t n = 0; n < pairs.size(); ++n) {
+    *strm << pairs[n].first << "\t" << pairs[n].second << "\n";
+  }
+  if (!*strm) {
+    LOG(ERROR) << "WriteIntPairs: Write failed: "
+               << (filename.empty() ? "standard output" : filename);
+    return false;
+  }
+  if (strm != &std::cout) delete strm;
+  return true;
+}
+
+// Utilities for reading/writing label pairs.
+
+template <typename Label>
+bool ReadLabelPairs(const string &filename,
+                    std::vector<std::pair<Label, Label>> *pairs,
+                    bool allow_negative = false) {
+  return ReadIntPairs(filename, pairs, allow_negative);
+}
+
+template <typename Label>
+bool WriteLabelPairs(const string &filename,
+                     const std::vector<std::pair<Label, Label>> &pairs) {
+  return WriteIntPairs(filename, pairs);
+}
+
+// Utilities for converting a type name to a legal C symbol.
+
+void ConvertToLegalCSymbol(string *s);
+
+// Utilities for stream I/O.
+
+bool AlignInput(std::istream &strm);
+bool AlignOutput(std::ostream &strm);
+
+// An associative container for which testing membership is faster than an STL
+// set if members are restricted to an interval that excludes most non-members.
+// A Key must have ==, !=, and < operators defined. Element NoKey should be a
+// key that marks an uninitialized key and is otherwise unused. Find() returns
+// an STL const_iterator to the match found, otherwise it equals End().
+template <class Key, Key NoKey>
+class CompactSet {
+ public:
+  using const_iterator = typename std::set<Key>::const_iterator;
+
+  CompactSet() : min_key_(NoKey), max_key_(NoKey) {}
+
+  CompactSet(const CompactSet<Key, NoKey> &compact_set)
+      : set_(compact_set.set_),
+        min_key_(compact_set.min_key_),
+        max_key_(compact_set.max_key_) {}
+
+  void Insert(Key key) {
+    set_.insert(key);
+    if (min_key_ == NoKey || key < min_key_) min_key_ = key;
+    if (max_key_ == NoKey || max_key_ < key) max_key_ = key;
+  }
+
+  void Erase(Key key) {
+    set_.erase(key);
+    if (set_.empty()) {
+      min_key_ = max_key_ = NoKey;
+    } else if (key == min_key_) {
+      ++min_key_;
+    } else if (key == max_key_) {
+      --max_key_;
+    }
+  }
+
+  void Clear() {
+    set_.clear();
+    min_key_ = max_key_ = NoKey;
+  }
+
+  const_iterator Find(Key key) const {
+    if (min_key_ == NoKey || key < min_key_ || max_key_ < key) {
+      return set_.end();
+    } else {
+      return set_.find(key);
+    }
+  }
+
+  bool Member(Key key) const {
+    if (min_key_ == NoKey || key < min_key_ || max_key_ < key) {
+      return false;  // out of range
+    } else if (min_key_ != NoKey && max_key_ + 1 == min_key_ + set_.size()) {
+      return true;  // dense range
+    } else {
+      return set_.count(key);
+    }
+  }
+
+  const_iterator Begin() const { return set_.begin(); }
+
+  const_iterator End() const { return set_.end(); }
+
+  // All stored keys are greater than or equal to this value.
+  Key LowerBound() const { return min_key_; }
+
+  // All stored keys are less than or equal to this value.
+  Key UpperBound() const { return max_key_; }
+
+ private:
+  std::set<Key> set_;
+  Key min_key_;
+  Key max_key_;
+
+  void operator=(const CompactSet &) = delete;
+};
+
+}  // namespace fst
+
+#endif  // FST_UTIL_H_
diff --git a/projects/llm_framework/include/fst/vector-fst.h b/projects/llm_framework/include/fst/vector-fst.h
new file mode 100644
index 00000000..7514bc55
--- /dev/null
+++ b/projects/llm_framework/include/fst/vector-fst.h
@@ -0,0 +1,796 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Simple concrete, mutable FST whose states and arcs are stored in STL vectors.
+
+#ifndef FST_VECTOR_FST_H_
+#define FST_VECTOR_FST_H_
+
+#include <string>
+#include <utility>
+#include <vector>
+
+#include <fst/log.h>
+
+#include <fst/fst-decl.h>  // For optional argument declarations
+#include <fst/mutable-fst.h>
+#include <fst/test-properties.h>
+
+
+namespace fst {
+
+template <class A, class S>
+class VectorFst;
+
+template <class F, class G>
+void Cast(const F &, G *);
+
+// Arcs (of type A) implemented by an STL vector per state. M specifies Arc
+// allocator (default declared in fst-decl.h).
+template <class A, class M /* = std::allocator<A> */>
+class VectorState {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+  using ArcAllocator = M;
+  using StateAllocator =
+      typename ArcAllocator::template rebind<VectorState<Arc, M>>::other;
+
+  // Provide STL allocator for arcs.
+  explicit VectorState(const ArcAllocator &alloc)
+      : final_(Weight::Zero()), niepsilons_(0), noepsilons_(0), arcs_(alloc) {}
+
+  VectorState(const VectorState<A, M> &state, const ArcAllocator &alloc)
+      : final_(state.Final()),
+        niepsilons_(state.NumInputEpsilons()),
+        noepsilons_(state.NumOutputEpsilons()),
+        arcs_(state.arcs_.begin(), state.arcs_.end(), alloc) {}
+
+  void Reset() {
+    final_ = Weight::Zero();
+    niepsilons_ = 0;
+    noepsilons_ = 0;
+    arcs_.clear();
+  }
+
+  Weight Final() const { return final_; }
+
+  size_t NumInputEpsilons() const { return niepsilons_; }
+
+  size_t NumOutputEpsilons() const { return noepsilons_; }
+
+  size_t NumArcs() const { return arcs_.size(); }
+
+  const Arc &GetArc(size_t n) const { return arcs_[n]; }
+
+  const Arc *Arcs() const { return !arcs_.empty() ? &arcs_[0] : nullptr; }
+
+  Arc *MutableArcs() { return !arcs_.empty() ? &arcs_[0] : nullptr; }
+
+  void ReserveArcs(size_t n) { arcs_.reserve(n); }
+
+  void SetFinal(Weight weight) { final_ = std::move(weight); }
+
+  void SetNumInputEpsilons(size_t n) { niepsilons_ = n; }
+
+  void SetNumOutputEpsilons(size_t n) { noepsilons_ = n; }
+
+  void AddArc(const Arc &arc) {
+    IncrementNumEpsilons(arc);
+    arcs_.push_back(arc);
+  }
+
+  void AddArc(Arc &&arc) {
+    IncrementNumEpsilons(arc);
+    arcs_.push_back(std::move(arc));
+  }
+
+  template <class... T>
+  void EmplaceArc(T&&... ctor_args) {
+    arcs_.emplace_back(std::forward<T>(ctor_args)...);
+    IncrementNumEpsilons(arcs_.back());
+  }
+
+  void SetArc(const Arc &arc, size_t n) {
+    if (arcs_[n].ilabel == 0) --niepsilons_;
+    if (arcs_[n].olabel == 0) --noepsilons_;
+    IncrementNumEpsilons(arc);
+    arcs_[n] = arc;
+  }
+
+  void DeleteArcs() {
+    niepsilons_ = 0;
+    noepsilons_ = 0;
+    arcs_.clear();
+  }
+
+  void DeleteArcs(size_t n) {
+    for (size_t i = 0; i < n; ++i) {
+      if (arcs_.back().ilabel == 0) --niepsilons_;
+      if (arcs_.back().olabel == 0) --noepsilons_;
+      arcs_.pop_back();
+    }
+  }
+
+  // For state class allocation.
+  void *operator new(size_t size, StateAllocator *alloc) {
+    return alloc->allocate(1);
+  }
+
+  // For state destruction and memory freeing.
+  static void Destroy(VectorState<A, M> *state, StateAllocator *alloc) {
+    if (state) {
+      state->~VectorState<A, M>();
+      alloc->deallocate(state, 1);
+    }
+  }
+
+ private:
+  // Update the number of epsilons as a result of having added an arc.
+  void IncrementNumEpsilons(const Arc &arc) {
+    if (arc.ilabel == 0) ++niepsilons_;
+    if (arc.olabel == 0) ++noepsilons_;
+  }
+
+  Weight final_;                       // Final weight.
+  size_t niepsilons_;                  // # of input epsilons
+  size_t noepsilons_;                  // # of output epsilons
+  std::vector<A, ArcAllocator> arcs_;  // Arc container.
+};
+
+namespace internal {
+
+// States are implemented by STL vectors, templated on the
+// State definition. This does not manage the Fst properties.
+template <class S>
+class VectorFstBaseImpl : public FstImpl<typename S::Arc> {
+ public:
+  using State = S;
+  using Arc = typename State::Arc;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  VectorFstBaseImpl() : start_(kNoStateId) {}
+
+  ~VectorFstBaseImpl() override {
+    for (size_t s = 0; s < states_.size(); ++s) {
+      State::Destroy(states_[s], &state_alloc_);
+    }
+  }
+
+  // Copying is not permitted.
+  VectorFstBaseImpl(const VectorFstBaseImpl<S> &) = delete;
+  VectorFstBaseImpl<S> &operator=(const VectorFstBaseImpl<S> &) = delete;
+
+  // Moving is permitted.
+  VectorFstBaseImpl(VectorFstBaseImpl<S> &&impl) noexcept
+      : FstImpl<typename S::Arc>(),
+        states_(std::move(impl.states_)),
+        start_(impl.start_) {
+    impl.states_.clear();
+    impl.start_ = kNoStateId;
+  }
+
+  VectorFstBaseImpl<S> &operator=(VectorFstBaseImpl<S> &&impl) noexcept {
+    states_ = std::move(impl.states_);
+    start_ = impl.start_;
+    impl.states_.clear();
+    impl.start_ = kNoStateId;
+    return *this;
+  }
+
+  StateId Start() const { return start_; }
+
+  Weight Final(StateId state) const { return states_[state]->Final(); }
+
+  StateId NumStates() const { return states_.size(); }
+
+  size_t NumArcs(StateId state) const { return states_[state]->NumArcs(); }
+
+  size_t NumInputEpsilons(StateId state) const {
+    return GetState(state)->NumInputEpsilons();
+  }
+
+  size_t NumOutputEpsilons(StateId state) const {
+    return GetState(state)->NumOutputEpsilons();
+  }
+
+  void SetStart(StateId state) { start_ = state; }
+
+  void SetFinal(StateId state, Weight weight) {
+    states_[state]->SetFinal(std::move(weight));
+  }
+
+  StateId AddState() {
+    states_.push_back(new (&state_alloc_) State(arc_alloc_));
+    return states_.size() - 1;
+  }
+
+  StateId AddState(State *state) {
+    states_.push_back(state);
+    return states_.size() - 1;
+  }
+
+  void AddArc(StateId state, const Arc &arc) { states_[state]->AddArc(arc); }
+
+  void AddArc(StateId state, Arc &&arc) {
+    states_[state]->AddArc(std::move(arc));
+  }
+
+  template <class... T>
+  void EmplaceArc(StateId state, T&&... ctor_args) {
+    states_[state]->EmplaceArc(std::forward<T>(ctor_args)...);
+  }
+
+  void DeleteStates(const std::vector<StateId> &dstates) {
+    std::vector<StateId> newid(states_.size(), 0);
+    for (size_t i = 0; i < dstates.size(); ++i) newid[dstates[i]] = kNoStateId;
+    StateId nstates = 0;
+    for (StateId state = 0; state < states_.size(); ++state) {
+      if (newid[state] != kNoStateId) {
+        newid[state] = nstates;
+        if (state != nstates) states_[nstates] = states_[state];
+        ++nstates;
+      } else {
+        State::Destroy(states_[state], &state_alloc_);
+      }
+    }
+    states_.resize(nstates);
+    for (StateId state = 0; state < states_.size(); ++state) {
+      auto *arcs = states_[state]->MutableArcs();
+      size_t narcs = 0;
+      auto nieps = states_[state]->NumInputEpsilons();
+      auto noeps = states_[state]->NumOutputEpsilons();
+      for (size_t i = 0; i < states_[state]->NumArcs(); ++i) {
+        const auto t = newid[arcs[i].nextstate];
+        if (t != kNoStateId) {
+          arcs[i].nextstate = t;
+          if (i != narcs) arcs[narcs] = arcs[i];
+          ++narcs;
+        } else {
+          if (arcs[i].ilabel == 0) --nieps;
+          if (arcs[i].olabel == 0) --noeps;
+        }
+      }
+      states_[state]->DeleteArcs(states_[state]->NumArcs() - narcs);
+      states_[state]->SetNumInputEpsilons(nieps);
+      states_[state]->SetNumOutputEpsilons(noeps);
+    }
+    if (Start() != kNoStateId) SetStart(newid[Start()]);
+  }
+
+  void DeleteStates() {
+    for (size_t state = 0; state < states_.size(); ++state) {
+      State::Destroy(states_[state], &state_alloc_);
+    }
+    states_.clear();
+    SetStart(kNoStateId);
+  }
+
+  void DeleteArcs(StateId state, size_t n) { states_[state]->DeleteArcs(n); }
+
+  void DeleteArcs(StateId state) { states_[state]->DeleteArcs(); }
+
+  State *GetState(StateId state) { return states_[state]; }
+
+  const State *GetState(StateId state) const { return states_[state]; }
+
+  void SetState(StateId state, State *vstate) { states_[state] = vstate; }
+
+  void ReserveStates(StateId n) { states_.reserve(n); }
+
+  void ReserveArcs(StateId state, size_t n) { states_[state]->ReserveArcs(n); }
+
+  // Provide information needed for generic state iterator.
+  void InitStateIterator(StateIteratorData<Arc> *data) const {
+    data->base = nullptr;
+    data->nstates = states_.size();
+  }
+
+  // Provide information needed for generic arc iterator.
+  void InitArcIterator(StateId state, ArcIteratorData<Arc> *data) const {
+    data->base = nullptr;
+    data->narcs = states_[state]->NumArcs();
+    data->arcs = states_[state]->Arcs();
+    data->ref_count = nullptr;
+  }
+
+ private:
+  std::vector<State *> states_;                 // States represenation.
+  StateId start_;                               // Initial state.
+  typename State::StateAllocator state_alloc_;  // For state allocation.
+  typename State::ArcAllocator arc_alloc_;      // For arc allocation.
+};
+
+// This is a VectorFstBaseImpl container that holds VectorStates and manages FST
+// properties.
+template <class S>
+class VectorFstImpl : public VectorFstBaseImpl<S> {
+ public:
+  using State = S;
+  using Arc = typename State::Arc;
+  using Label = typename Arc::Label;
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  using FstImpl<Arc>::SetInputSymbols;
+  using FstImpl<Arc>::SetOutputSymbols;
+  using FstImpl<Arc>::SetType;
+  using FstImpl<Arc>::SetProperties;
+  using FstImpl<Arc>::Properties;
+
+  using VectorFstBaseImpl<S>::Start;
+  using VectorFstBaseImpl<S>::NumStates;
+  using VectorFstBaseImpl<S>::GetState;
+  using VectorFstBaseImpl<S>::ReserveArcs;
+
+  friend class MutableArcIterator<VectorFst<Arc, S>>;
+
+  using BaseImpl = VectorFstBaseImpl<S>;
+
+  VectorFstImpl() {
+    SetType("vector");
+    SetProperties(kNullProperties | kStaticProperties);
+  }
+
+  explicit VectorFstImpl(const Fst<Arc> &fst);
+
+  static VectorFstImpl<S> *Read(std::istream &strm, const FstReadOptions &opts);
+
+  void SetStart(StateId state) {
+    BaseImpl::SetStart(state);
+    SetProperties(SetStartProperties(Properties()));
+  }
+
+  void SetFinal(StateId state, Weight weight) {
+    const auto old_weight = BaseImpl::Final(state);
+    const auto properties =
+        SetFinalProperties(Properties(), old_weight, weight);
+    BaseImpl::SetFinal(state, std::move(weight));
+    SetProperties(properties);
+  }
+
+  StateId AddState() {
+    const auto state = BaseImpl::AddState();
+    SetProperties(AddStateProperties(Properties()));
+    return state;
+  }
+
+  void AddArc(StateId state, const Arc &arc) {
+    BaseImpl::AddArc(state, arc);
+    UpdatePropertiesAfterAddArc(state);
+  }
+
+  void AddArc(StateId state, Arc &&arc) {
+    BaseImpl::AddArc(state, std::move(arc));
+    UpdatePropertiesAfterAddArc(state);
+  }
+
+  template <class... T>
+  void EmplaceArc(StateId state, T&&... ctor_args) {
+    BaseImpl::EmplaceArc(state, std::forward<T>(ctor_args)...);
+    UpdatePropertiesAfterAddArc(state);
+  }
+
+  void DeleteStates(const std::vector<StateId> &dstates) {
+    BaseImpl::DeleteStates(dstates);
+    SetProperties(DeleteStatesProperties(Properties()));
+  }
+
+  void DeleteStates() {
+    BaseImpl::DeleteStates();
+    SetProperties(DeleteAllStatesProperties(Properties(), kStaticProperties));
+  }
+
+  void DeleteArcs(StateId state, size_t n) {
+    BaseImpl::DeleteArcs(state, n);
+    SetProperties(DeleteArcsProperties(Properties()));
+  }
+
+  void DeleteArcs(StateId state) {
+    BaseImpl::DeleteArcs(state);
+    SetProperties(DeleteArcsProperties(Properties()));
+  }
+
+  // Properties always true of this FST class
+  static constexpr uint64 kStaticProperties = kExpanded | kMutable;
+
+ private:
+  void UpdatePropertiesAfterAddArc(StateId state) {
+    auto *vstate = GetState(state);
+    const size_t num_arcs{vstate->NumArcs()};
+    if (num_arcs) {
+      const auto &arc = vstate->GetArc(num_arcs - 1);
+      const auto *parc = (num_arcs < 2)
+                         ? nullptr
+                         : &(vstate->GetArc(num_arcs - 2));
+      SetProperties(AddArcProperties(Properties(), state, arc, parc));
+    }
+  }
+
+  // Minimum file format version supported.
+  static constexpr int kMinFileVersion = 2;
+};
+
+template <class S>
+constexpr uint64 VectorFstImpl<S>::kStaticProperties;
+
+template <class S>
+constexpr int VectorFstImpl<S>::kMinFileVersion;
+
+template <class S>
+VectorFstImpl<S>::VectorFstImpl(const Fst<Arc> &fst) {
+  SetType("vector");
+  SetInputSymbols(fst.InputSymbols());
+  SetOutputSymbols(fst.OutputSymbols());
+  BaseImpl::SetStart(fst.Start());
+  if (fst.Properties(kExpanded, false)) {
+    BaseImpl::ReserveStates(CountStates(fst));
+  }
+  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
+    const auto state = siter.Value();
+    BaseImpl::AddState();
+    BaseImpl::SetFinal(state, fst.Final(state));
+    ReserveArcs(state, fst.NumArcs(state));
+    for (ArcIterator<Fst<Arc>> aiter(fst, state); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      BaseImpl::AddArc(state, arc);
+    }
+  }
+  SetProperties(fst.Properties(kCopyProperties, false) | kStaticProperties);
+}
+
+template <class S>
+VectorFstImpl<S> *VectorFstImpl<S>::Read(std::istream &strm,
+                                         const FstReadOptions &opts) {
+  std::unique_ptr<VectorFstImpl<S>> impl(new VectorFstImpl());
+  FstHeader hdr;
+  if (!impl->ReadHeader(strm, opts, kMinFileVersion, &hdr)) return nullptr;
+  impl->BaseImpl::SetStart(hdr.Start());
+  if (hdr.NumStates() != kNoStateId) impl->ReserveStates(hdr.NumStates());
+  StateId state = 0;
+  for (; hdr.NumStates() == kNoStateId || state < hdr.NumStates(); ++state) {
+    Weight weight;
+    if (!weight.Read(strm)) break;
+    impl->BaseImpl::AddState();
+    auto *vstate = impl->GetState(state);
+    vstate->SetFinal(weight);
+    int64 narcs;
+    ReadType(strm, &narcs);
+    if (!strm) {
+      LOG(ERROR) << "VectorFst::Read: Read failed: " << opts.source;
+      return nullptr;
+    }
+    impl->ReserveArcs(state, narcs);
+    for (int64 i = 0; i < narcs; ++i) {
+      Arc arc;
+      ReadType(strm, &arc.ilabel);
+      ReadType(strm, &arc.olabel);
+      arc.weight.Read(strm);
+      ReadType(strm, &arc.nextstate);
+      if (!strm) {
+        LOG(ERROR) << "VectorFst::Read: Read failed: " << opts.source;
+        return nullptr;
+      }
+      impl->BaseImpl::AddArc(state, std::move(arc));
+    }
+  }
+  if (hdr.NumStates() != kNoStateId && state != hdr.NumStates()) {
+    LOG(ERROR) << "VectorFst::Read: Unexpected end of file: " << opts.source;
+    return nullptr;
+  }
+  return impl.release();
+}
+
+}  // namespace internal
+
+// Simple concrete, mutable FST. This class attaches interface to implementation
+// and handles reference counting, delegating most methods to ImplToMutableFst.
+// Also supports ReserveStates and ReserveArcs methods (cf. STL vector methods).
+// The second optional template argument gives the State definition.
+template <class A, class S /* = VectorState<A> */>
+class VectorFst : public ImplToMutableFst<internal::VectorFstImpl<S>> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+
+  using State = S;
+  using Impl = internal::VectorFstImpl<State>;
+
+  friend class StateIterator<VectorFst<Arc, State>>;
+  friend class ArcIterator<VectorFst<Arc, State>>;
+  friend class MutableArcIterator<VectorFst<A, S>>;
+
+  template <class F, class G>
+  friend void Cast(const F &, G *);
+
+  VectorFst() : ImplToMutableFst<Impl>(std::make_shared<Impl>()) {}
+
+  explicit VectorFst(const Fst<Arc> &fst)
+      : ImplToMutableFst<Impl>(std::make_shared<Impl>(fst)) {}
+
+  VectorFst(const VectorFst<Arc, State> &fst, bool safe = false)
+      : ImplToMutableFst<Impl>(fst) {}
+
+  VectorFst(VectorFst<Arc, State> &&) noexcept;
+
+  // Get a copy of this VectorFst. See Fst<>::Copy() for further doc.
+  VectorFst<Arc, State> *Copy(bool safe = false) const override {
+    return new VectorFst<Arc, State>(*this, safe);
+  }
+
+  VectorFst<Arc, State> &operator=(const VectorFst<Arc, State> &) = default;
+
+  VectorFst<Arc, State> &operator=(VectorFst<Arc, State> &&) noexcept;
+
+  VectorFst<Arc, State> &operator=(const Fst<Arc> &fst) override {
+    if (this != &fst) SetImpl(std::make_shared<Impl>(fst));
+    return *this;
+  }
+
+  template <class... T>
+  void EmplaceArc(StateId state, T&&... ctor_args) {
+    MutateCheck();
+    GetMutableImpl()->EmplaceArc(state, std::forward<T>(ctor_args)...);
+  }
+
+  // Reads a VectorFst from an input stream, returning nullptr on error.
+  static VectorFst<Arc, State> *Read(std::istream &strm,
+                                     const FstReadOptions &opts) {
+    auto *impl = Impl::Read(strm, opts);
+    return impl ? new VectorFst<Arc, State>(std::shared_ptr<Impl>(impl))
+                : nullptr;
+  }
+
+  // Read a VectorFst from a file, returning nullptr on error; empty filename
+  // reads from standard input.
+  static VectorFst<Arc, State> *Read(const string &filename) {
+    auto *impl = ImplToExpandedFst<Impl, MutableFst<Arc>>::Read(filename);
+    return impl ? new VectorFst<Arc, State>(std::shared_ptr<Impl>(impl))
+                : nullptr;
+  }
+
+  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
+    return WriteFst(*this, strm, opts);
+  }
+
+  bool Write(const string &filename) const override {
+    return Fst<Arc>::WriteFile(filename);
+  }
+
+  template <class FST>
+  static bool WriteFst(const FST &fst, std::ostream &strm,
+                       const FstWriteOptions &opts);
+
+  void InitStateIterator(StateIteratorData<Arc> *data) const override {
+    GetImpl()->InitStateIterator(data);
+  }
+
+  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
+    GetImpl()->InitArcIterator(s, data);
+  }
+
+  inline void InitMutableArcIterator(StateId s,
+                                     MutableArcIteratorData<Arc> *) override;
+
+  using ImplToMutableFst<Impl, MutableFst<Arc>>::ReserveArcs;
+  using ImplToMutableFst<Impl, MutableFst<Arc>>::ReserveStates;
+
+ private:
+  using ImplToMutableFst<Impl, MutableFst<Arc>>::GetImpl;
+  using ImplToMutableFst<Impl, MutableFst<Arc>>::GetMutableImpl;
+  using ImplToMutableFst<Impl, MutableFst<Arc>>::MutateCheck;
+  using ImplToMutableFst<Impl, MutableFst<Arc>>::SetImpl;
+
+  explicit VectorFst(std::shared_ptr<Impl> impl)
+      : ImplToMutableFst<Impl>(impl) {}
+};
+
+template <class Arc, class State>
+inline VectorFst<Arc, State>::VectorFst(
+    VectorFst<Arc, State> &&fst) noexcept = default;
+
+template <class Arc, class State>
+inline VectorFst<Arc, State> &VectorFst<Arc, State>::operator=(
+    VectorFst<Arc, State> &&fst) noexcept = default;
+
+// Writes FST to file in Vector format, potentially with a pass over the machine
+// before writing to compute number of states.
+template <class Arc, class State>
+template <class FST>
+bool VectorFst<Arc, State>::WriteFst(const FST &fst, std::ostream &strm,
+                                     const FstWriteOptions &opts) {
+  static constexpr int file_version = 2;
+  bool update_header = true;
+  FstHeader hdr;
+  hdr.SetStart(fst.Start());
+  hdr.SetNumStates(kNoStateId);
+  std::streampos start_offset = 0;
+  if (fst.Properties(kExpanded, false) || opts.stream_write ||
+      (start_offset = strm.tellp()) != -1) {
+    hdr.SetNumStates(CountStates(fst));
+    update_header = false;
+  }
+  const auto properties =
+      fst.Properties(kCopyProperties, false) | Impl::kStaticProperties;
+  internal::FstImpl<Arc>::WriteFstHeader(fst, strm, opts, file_version,
+                                         "vector", properties, &hdr);
+  StateId num_states = 0;
+  for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
+    const auto s = siter.Value();
+    fst.Final(s).Write(strm);
+    const int64 narcs = fst.NumArcs(s);
+    WriteType(strm, narcs);
+    for (ArcIterator<FST> aiter(fst, s); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      WriteType(strm, arc.ilabel);
+      WriteType(strm, arc.olabel);
+      arc.weight.Write(strm);
+      WriteType(strm, arc.nextstate);
+    }
+    ++num_states;
+  }
+  strm.flush();
+  if (!strm) {
+    LOG(ERROR) << "VectorFst::Write: Write failed: " << opts.source;
+    return false;
+  }
+  if (update_header) {
+    hdr.SetNumStates(num_states);
+    return internal::FstImpl<Arc>::UpdateFstHeader(
+        fst, strm, opts, file_version, "vector", properties, &hdr,
+        start_offset);
+  } else {
+    if (num_states != hdr.NumStates()) {
+      LOG(ERROR) << "Inconsistent number of states observed during write";
+      return false;
+    }
+  }
+  return true;
+}
+
+// Specialization for VectorFst; see generic version in fst.h for sample usage
+// (but use the VectorFst type instead). This version should inline.
+template <class Arc, class State>
+class StateIterator<VectorFst<Arc, State>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  explicit StateIterator(const VectorFst<Arc, State> &fst)
+      : nstates_(fst.GetImpl()->NumStates()), s_(0) {}
+
+  bool Done() const { return s_ >= nstates_; }
+
+  StateId Value() const { return s_; }
+
+  void Next() { ++s_; }
+
+  void Reset() { s_ = 0; }
+
+ private:
+  const StateId nstates_;
+  StateId s_;
+};
+
+// Specialization for VectorFst; see generic version in fst.h for sample usage
+// (but use the VectorFst type instead). This version should inline.
+template <class Arc, class State>
+class ArcIterator<VectorFst<Arc, State>> {
+ public:
+  using StateId = typename Arc::StateId;
+
+  ArcIterator(const VectorFst<Arc, State> &fst, StateId s)
+      : arcs_(fst.GetImpl()->GetState(s)->Arcs()),
+        narcs_(fst.GetImpl()->GetState(s)->NumArcs()),
+        i_(0) {}
+
+  bool Done() const { return i_ >= narcs_; }
+
+  const Arc &Value() const { return arcs_[i_]; }
+
+  void Next() { ++i_; }
+
+  void Reset() { i_ = 0; }
+
+  void Seek(size_t a) { i_ = a; }
+
+  size_t Position() const { return i_; }
+
+  constexpr uint32 Flags() const { return kArcValueFlags; }
+
+  void SetFlags(uint32, uint32) {}
+
+ private:
+  const Arc *arcs_;
+  size_t narcs_;
+  size_t i_;
+};
+
+// Specialization for VectorFst; see generic version in mutable-fst.h for sample
+// usage (but use the VectorFst type instead). This version should inline.
+template <class Arc, class State>
+class MutableArcIterator<VectorFst<Arc, State>>
+    : public MutableArcIteratorBase<Arc> {
+ public:
+  using StateId = typename Arc::StateId;
+  using Weight = typename Arc::Weight;
+
+  MutableArcIterator(VectorFst<Arc, State> *fst, StateId s) : i_(0) {
+    fst->MutateCheck();
+    state_ = fst->GetMutableImpl()->GetState(s);
+    properties_ = &fst->GetImpl()->properties_;
+  }
+
+  bool Done() const final { return i_ >= state_->NumArcs(); }
+
+  const Arc &Value() const final { return state_->GetArc(i_); }
+
+  void Next() final { ++i_; }
+
+  size_t Position() const final { return i_; }
+
+  void Reset() final { i_ = 0; }
+
+  void Seek(size_t a) final { i_ = a; }
+
+  void SetValue(const Arc &arc) final {
+    const auto &oarc = state_->GetArc(i_);
+    if (oarc.ilabel != oarc.olabel) *properties_ &= ~kNotAcceptor;
+    if (oarc.ilabel == 0) {
+      *properties_ &= ~kIEpsilons;
+      if (oarc.olabel == 0) *properties_ &= ~kEpsilons;
+    }
+    if (oarc.olabel == 0) *properties_ &= ~kOEpsilons;
+    if (oarc.weight != Weight::Zero() && oarc.weight != Weight::One()) {
+      *properties_ &= ~kWeighted;
+    }
+    state_->SetArc(arc, i_);
+    if (arc.ilabel != arc.olabel) {
+      *properties_ |= kNotAcceptor;
+      *properties_ &= ~kAcceptor;
+    }
+    if (arc.ilabel == 0) {
+      *properties_ |= kIEpsilons;
+      *properties_ &= ~kNoIEpsilons;
+      if (arc.olabel == 0) {
+        *properties_ |= kEpsilons;
+        *properties_ &= ~kNoEpsilons;
+      }
+    }
+    if (arc.olabel == 0) {
+      *properties_ |= kOEpsilons;
+      *properties_ &= ~kNoOEpsilons;
+    }
+    if (arc.weight != Weight::Zero() && arc.weight != Weight::One()) {
+      *properties_ |= kWeighted;
+      *properties_ &= ~kUnweighted;
+    }
+    *properties_ &= kSetArcProperties | kAcceptor | kNotAcceptor | kEpsilons |
+                    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons |
+                    kNoOEpsilons | kWeighted | kUnweighted;
+  }
+
+  uint32 Flags() const final { return kArcValueFlags; }
+
+  void SetFlags(uint32, uint32) final {}
+
+ private:
+  State *state_;
+  uint64 *properties_;
+  size_t i_;
+};
+
+// Provides information needed for the generic mutable arc iterator.
+template <class Arc, class State>
+inline void VectorFst<Arc, State>::InitMutableArcIterator(
+    StateId s, MutableArcIteratorData<Arc> *data) {
+  data->base = new MutableArcIterator<VectorFst<Arc, State>>(this, s);
+}
+
+// A useful alias when using StdArc.
+using StdVectorFst = VectorFst<StdArc>;
+
+}  // namespace fst
+
+#endif  // FST_VECTOR_FST_H_
diff --git a/projects/llm_framework/include/fst/verify.h b/projects/llm_framework/include/fst/verify.h
new file mode 100644
index 00000000..2ea8a64a
--- /dev/null
+++ b/projects/llm_framework/include/fst/verify.h
@@ -0,0 +1,100 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Function to verify an FST's contents.
+
+#ifndef FST_VERIFY_H_
+#define FST_VERIFY_H_
+
+#include <fst/log.h>
+
+#include <fst/fst.h>
+#include <fst/test-properties.h>
+
+
+namespace fst {
+
+// Verifies that an Fst's contents are sane.
+template <class Arc>
+bool Verify(const Fst<Arc> &fst, bool allow_negative_labels = false) {
+  using StateId = typename Arc::StateId;
+  const auto start = fst.Start();
+  const auto *isyms = fst.InputSymbols();
+  const auto *osyms = fst.OutputSymbols();
+  // Count states
+  StateId ns = 0;
+  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) ++ns;
+  if (start == kNoStateId && ns > 0) {
+    LOG(ERROR) << "Verify: FST start state ID not set";
+    return false;
+  } else if (start >= ns) {
+    LOG(ERROR) << "Verify: FST start state ID exceeds number of states";
+    return false;
+  }
+  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
+    auto state = siter.Value();
+    size_t na = 0;
+    for (ArcIterator<Fst<Arc>> aiter(fst, state); !aiter.Done(); aiter.Next()) {
+      const auto &arc = aiter.Value();
+      if (!allow_negative_labels && arc.ilabel < 0) {
+        LOG(ERROR) << "Verify: FST input label ID of arc at position " << na
+                   << " of state " << state << " is negative";
+        return false;
+      } else if (isyms && isyms->Find(arc.ilabel).empty()) {
+        LOG(ERROR) << "Verify: FST input label ID " << arc.ilabel
+                   << " of arc at position " << na << " of state " << state
+                   << " is missing from input symbol table \"" << isyms->Name()
+                   << "\"";
+        return false;
+      } else if (!allow_negative_labels && arc.olabel < 0) {
+        LOG(ERROR) << "Verify: FST output label ID of arc at position " << na
+                   << " of state " << state << " is negative";
+        return false;
+      } else if (osyms && osyms->Find(arc.olabel).empty()) {
+        LOG(ERROR) << "Verify: FST output label ID " << arc.olabel
+                   << " of arc at position " << na << " of state " << state
+                   << " is missing from output symbol table \"" << osyms->Name()
+                   << "\"";
+        return false;
+      } else if (!arc.weight.Member()) {
+        LOG(ERROR) << "Verify: FST weight of arc at position " << na
+                   << " of state " << state << " is invalid";
+        return false;
+      } else if (arc.nextstate < 0) {
+        LOG(ERROR) << "Verify: FST destination state ID of arc at position "
+                   << na << " of state " << state << " is negative";
+        return false;
+      } else if (arc.nextstate >= ns) {
+        LOG(ERROR) << "Verify: FST destination state ID of arc at position "
+                   << na << " of state " << state
+                   << " exceeds number of states";
+        return false;
+      }
+      ++na;
+    }
+    if (!fst.Final(state).Member()) {
+      LOG(ERROR) << "Verify: FST final weight of state " << state
+                 << " is invalid";
+      return false;
+    }
+  }
+  const auto fst_props = fst.Properties(kFstProperties, false);
+  if (fst_props & kError) {
+    LOG(ERROR) << "Verify: FST error property is set";
+    return false;
+  }
+  uint64 known_props;
+  uint64 test_props =
+      ComputeProperties(fst, kFstProperties, &known_props, false);
+  if (!CompatProperties(fst_props, test_props)) {
+    LOG(ERROR) << "Verify: Stored FST properties incorrect "
+               << "(props1 = stored props, props2 = tested)";
+    return false;
+  } else {
+    return true;
+  }
+}
+
+}  // namespace fst
+
+#endif  // FST_VERIFY_H_
diff --git a/projects/llm_framework/include/fst/visit.h b/projects/llm_framework/include/fst/visit.h
new file mode 100644
index 00000000..c6658047
--- /dev/null
+++ b/projects/llm_framework/include/fst/visit.h
@@ -0,0 +1,321 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// Queue-dependent visitation of finite-state transducers. See also dfs-visit.h.
+
+#ifndef FST_VISIT_H_
+#define FST_VISIT_H_
+
+
+#include <fst/arcfilter.h>
+#include <fst/mutable-fst.h>
+
+
+namespace fst {
+
+// Visitor Interface: class determining actions taken during a visit. If any of
+// the boolean member functions return false, the visit is aborted by first
+// calling FinishState() on all unfinished (grey) states and then calling
+// FinishVisit().
+//
+// Note this is more general than the visitor interface in dfs-visit.h but lacks
+// some DFS-specific behavior.
+//
+// template <class Arc>
+// class Visitor {
+//  public:
+//   using StateId = typename Arc::StateId;
+//
+//   Visitor(T *return_data);
+//
+//   // Invoked before visit.
+//   void InitVisit(const Fst<Arc> &fst);
+//
+//   // Invoked when state discovered (2nd arg is visitation root).
+//   bool InitState(StateId s, StateId root);
+//
+//   // Invoked when arc to white/undiscovered state examined.
+//   bool WhiteArc(StateId s, const Arc &arc);
+//
+//   // Invoked when arc to grey/unfinished state examined.
+//   bool GreyArc(StateId s, const Arc &arc);
+//
+//   // Invoked when arc to black/finished state examined.
+//   bool BlackArc(StateId s, const Arc &arc);
+//
+//   // Invoked when state finished.
+//   void FinishState(StateId s);
+//
+//   // Invoked after visit.
+//   void FinishVisit();
+// };
+
+// Performs queue-dependent visitation. Visitor class argument determines
+// actions and contains any return data. ArcFilter determines arcs that are
+// considered. If 'access_only' is true, performs visitation only to states
+// accessible from the initial state.
+template <class FST, class Visitor, class Queue, class ArcFilter>
+void Visit(const FST &fst, Visitor *visitor, Queue *queue, ArcFilter filter,
+           bool access_only = false) {
+  using Arc = typename FST::Arc;
+  using StateId = typename Arc::StateId;
+  visitor->InitVisit(fst);
+  const auto start = fst.Start();
+  if (start == kNoStateId) {
+    visitor->FinishVisit();
+    return;
+  }
+  // An FST's state's visit color.
+  static constexpr uint8 kWhiteState = 0x01;  // Undiscovered.
+  static constexpr uint8 kGreyState = 0x02;   // Discovered & unfinished.
+  static constexpr uint8 kBlackState = 0x04;  // Finished.
+  // We destroy an iterator as soon as possible and mark it so.
+  static constexpr uint8 kArcIterDone = 0x08;
+  std::vector<uint8> state_status;
+  std::vector<ArcIterator<FST> *> arc_iterator;
+  MemoryPool<ArcIterator<FST>> aiter_pool;
+  StateId nstates = start + 1;  // Number of known states in general case.
+  bool expanded = false;
+  if (fst.Properties(kExpanded, false)) {  // Tests if expanded, then uses
+    nstates = CountStates(fst);            // ExpandedFst::NumStates().
+    expanded = true;
+  }
+  state_status.resize(nstates, kWhiteState);
+  arc_iterator.resize(nstates);
+  StateIterator<Fst<Arc>> siter(fst);
+  // Continues visit while true.
+  bool visit = true;
+  // Iterates over trees in visit forest.
+  for (auto root = start; visit && root < nstates;) {
+    visit = visitor->InitState(root, root);
+    state_status[root] = kGreyState;
+    queue->Enqueue(root);
+    while (!queue->Empty()) {
+      auto state = queue->Head();
+      if (state >= state_status.size()) {
+        nstates = state + 1;
+        state_status.resize(nstates, kWhiteState);
+        arc_iterator.resize(nstates);
+      }
+      // Creates arc iterator if needed.
+      if (!arc_iterator[state] && !(state_status[state] & kArcIterDone) &&
+          visit) {
+        arc_iterator[state] = new (&aiter_pool) ArcIterator<FST>(fst, state);
+      }
+      // Deletes arc iterator if done.
+      auto *aiter = arc_iterator[state];
+      if ((aiter && aiter->Done()) || !visit) {
+        Destroy(aiter, &aiter_pool);
+        arc_iterator[state] = nullptr;
+        state_status[state] |= kArcIterDone;
+      }
+      // Dequeues state and marks black if done.
+      if (state_status[state] & kArcIterDone) {
+        queue->Dequeue();
+        visitor->FinishState(state);
+        state_status[state] = kBlackState;
+        continue;
+      }
+      const auto &arc = aiter->Value();
+      if (arc.nextstate >= state_status.size()) {
+        nstates = arc.nextstate + 1;
+        state_status.resize(nstates, kWhiteState);
+        arc_iterator.resize(nstates);
+      }
+      // Visits respective arc types.
+      if (filter(arc)) {
+        // Enqueues destination state and marks grey if white.
+        if (state_status[arc.nextstate] == kWhiteState) {
+          visit = visitor->WhiteArc(state, arc);
+          if (!visit) continue;
+          visit = visitor->InitState(arc.nextstate, root);
+          state_status[arc.nextstate] = kGreyState;
+          queue->Enqueue(arc.nextstate);
+        } else if (state_status[arc.nextstate] == kBlackState) {
+          visit = visitor->BlackArc(state, arc);
+        } else {
+          visit = visitor->GreyArc(state, arc);
+        }
+      }
+      aiter->Next();
+      // Destroys an iterator ASAP for efficiency.
+      if (aiter->Done()) {
+        Destroy(aiter, &aiter_pool);
+        arc_iterator[state] = nullptr;
+        state_status[state] |= kArcIterDone;
+      }
+    }
+    if (access_only) break;
+    // Finds next tree root.
+    for (root = (root == start) ? 0 : root + 1;
+         root < nstates && state_status[root] != kWhiteState; ++root) {
+    }
+    // Check for a state beyond the largest known state.
+    if (!expanded && root == nstates) {
+      for (; !siter.Done(); siter.Next()) {
+        if (siter.Value() == nstates) {
+          ++nstates;
+          state_status.push_back(kWhiteState);
+          arc_iterator.push_back(nullptr);
+          break;
+        }
+      }
+    }
+  }
+  visitor->FinishVisit();
+}
+
+template <class Arc, class Visitor, class Queue>
+inline void Visit(const Fst<Arc> &fst, Visitor *visitor, Queue *queue) {
+  Visit(fst, visitor, queue, AnyArcFilter<Arc>());
+}
+
+// Copies input FST to mutable FST following queue order.
+template <class A>
+class CopyVisitor {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+
+  explicit CopyVisitor(MutableFst<Arc> *ofst) : ifst_(nullptr), ofst_(ofst) {}
+
+  void InitVisit(const Fst<A> &ifst) {
+    ifst_ = &ifst;
+    ofst_->DeleteStates();
+    ofst_->SetStart(ifst_->Start());
+  }
+
+  bool InitState(StateId state, StateId) {
+    while (ofst_->NumStates() <= state) ofst_->AddState();
+    return true;
+  }
+
+  bool WhiteArc(StateId state, const Arc &arc) {
+    ofst_->AddArc(state, arc);
+    return true;
+  }
+
+  bool GreyArc(StateId state, const Arc &arc) {
+    ofst_->AddArc(state, arc);
+    return true;
+  }
+
+  bool BlackArc(StateId state, const Arc &arc) {
+    ofst_->AddArc(state, arc);
+    return true;
+  }
+
+  void FinishState(StateId state) {
+    ofst_->SetFinal(state, ifst_->Final(state));
+  }
+
+  void FinishVisit() {}
+
+ private:
+  const Fst<Arc> *ifst_;
+  MutableFst<Arc> *ofst_;
+};
+
+// Visits input FST up to a state limit following queue order.
+template <class A>
+class PartialVisitor {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+
+  explicit PartialVisitor(StateId maxvisit)
+      : fst_(nullptr), maxvisit_(maxvisit) {}
+
+  void InitVisit(const Fst<A> &ifst) {
+    fst_ = &ifst;
+    ninit_ = 0;
+    nfinish_ = 0;
+  }
+
+  bool InitState(StateId state, StateId root) {
+    ++ninit_;
+    return ninit_ <= maxvisit_;
+  }
+
+  bool WhiteArc(StateId state, const Arc &arc) { return true; }
+
+  bool GreyArc(StateId state, const Arc &arc) { return true; }
+
+  bool BlackArc(StateId state, const Arc &arc) { return true; }
+
+  void FinishState(StateId state) {
+    fst_->Final(state);  // Visits super-final arc.
+    ++nfinish_;
+  }
+
+  void FinishVisit() {}
+
+  StateId NumInitialized() { return ninit_; }
+
+  StateId NumFinished() { return nfinish_; }
+
+ private:
+  const Fst<Arc> *fst_;
+  StateId maxvisit_;
+  StateId ninit_;
+  StateId nfinish_;
+};
+
+// Copies input FST to mutable FST up to a state limit following queue order.
+template <class A>
+class PartialCopyVisitor : public CopyVisitor<A> {
+ public:
+  using Arc = A;
+  using StateId = typename Arc::StateId;
+
+  using CopyVisitor<A>::WhiteArc;
+
+  PartialCopyVisitor(MutableFst<Arc> *ofst, StateId maxvisit,
+                     bool copy_grey = true, bool copy_black = true)
+      : CopyVisitor<A>(ofst), maxvisit_(maxvisit),
+        copy_grey_(copy_grey), copy_black_(copy_black) {}
+
+  void InitVisit(const Fst<A> &ifst) {
+    CopyVisitor<A>::InitVisit(ifst);
+    ninit_ = 0;
+    nfinish_ = 0;
+  }
+
+  bool InitState(StateId state, StateId root) {
+    CopyVisitor<A>::InitState(state, root);
+    ++ninit_;
+    return ninit_ <= maxvisit_;
+  }
+
+  bool GreyArc(StateId state, const Arc &arc) {
+    if (copy_grey_) return CopyVisitor<A>::GreyArc(state, arc);
+    return true;
+  }
+
+  bool BlackArc(StateId state, const Arc &arc) {
+    if (copy_black_) return CopyVisitor<A>::BlackArc(state, arc);
+    return true;
+  }
+
+  void FinishState(StateId state) {
+    CopyVisitor<A>::FinishState(state);
+    ++nfinish_;
+  }
+
+  void FinishVisit() {}
+
+  StateId NumInitialized() { return ninit_; }
+
+  StateId NumFinished() { return nfinish_; }
+
+ private:
+  StateId maxvisit_;
+  StateId ninit_;
+  StateId nfinish_;
+  const bool copy_grey_;
+  const bool copy_black_;
+};
+
+}  // namespace fst
+
+#endif  // FST_VISIT_H_
diff --git a/projects/llm_framework/include/fst/weight.h b/projects/llm_framework/include/fst/weight.h
new file mode 100644
index 00000000..ea012ec6
--- /dev/null
+++ b/projects/llm_framework/include/fst/weight.h
@@ -0,0 +1,389 @@
+// See www.openfst.org for extensive documentation on this weighted
+// finite-state transducer library.
+//
+// General weight set and associated semiring operation definitions.
+
+#ifndef FST_WEIGHT_H_
+#define FST_WEIGHT_H_
+
+#include <cctype>
+#include <cmath>
+#include <iostream>
+#include <sstream>
+#include <type_traits>
+#include <utility>
+
+#include <fst/compat.h>
+#include <fst/log.h>
+
+#include <fst/util.h>
+
+
+DECLARE_string(fst_weight_parentheses);
+DECLARE_string(fst_weight_separator);
+
+namespace fst {
+
+// A semiring is specified by two binary operations Plus and Times and two
+// designated elements Zero and One with the following properties:
+//
+//   Plus: associative, commutative, and has Zero as its identity.
+//
+//   Times: associative and has identity One, distributes w.r.t. Plus, and
+//     has Zero as an annihilator:
+//          Times(Zero(), a) == Times(a, Zero()) = Zero().
+//
+// A left semiring distributes on the left; a right semiring is similarly
+// defined.
+//
+// A Weight class must have binary functions Plus and Times and static member
+// functions Zero() and One() and these must form (at least) a left or right
+// semiring.
+//
+// In addition, the following should be defined for a Weight:
+//
+//   Member: predicate on set membership.
+//
+//   NoWeight: static member function that returns an element that is
+//      not a set member; used to signal an error.
+//
+//   >>: reads textual representation of a weight.
+//
+//   <<: prints textual representation of a weight.
+//
+//   Read(istream &istrm): reads binary representation of a weight.
+//
+//   Write(ostream &ostrm): writes binary representation of a weight.
+//
+//   Hash: maps weight to size_t.
+//
+//   ApproxEqual: approximate equality (for inexact weights)
+//
+//   Quantize: quantizes w.r.t delta (for inexact weights)
+//
+//   Divide:
+//     - In a left semiring, for all a, b, b', c:
+//       if Times(a, b) = c, Divide(c, a, DIVIDE_LEFT) = b' and b'.Member(),
+//       then Times(a, b') = c.
+//     - In a right semiring, for all a, a', b, c:
+//       if Times(a, b) = c, Divide(c, b, DIVIDE_RIGHT) = a' and a'.Member(),
+//       then Times(a', b) = c.
+//     - In a commutative semiring,
+//        * for all a, c:
+//          Divide(c, a, DIVIDE_ANY) = Divide(c, a, DIVIDE_LEFT)
+//           = Divide(c, a, DIVIDE_RIGHT)
+//        * for all a, b, b', c:
+//          if Times(a, b), Divide(c, a, DIVIDE_ANY) = b' and b'.Member(),
+//          then Times(a, b') = c
+//     - In the case where there exist no b such that c = Times(a, b), the
+//       return value of Divide(c, a, DIVIDE_LEFT) is unspecified. Returning
+//       Weight::NoWeight() is recommemded but not required in order to
+//       allow the most efficient implementation.
+//     - All algorithms in this library only call Divide(c, a) when it is
+//       guaranteed that there exists a b such that c = Times(a, b).
+//
+//   ReverseWeight: the type of the corresponding reverse weight.
+//
+//     Typically the same type as Weight for a (both left and right) semiring.
+//     For the left string semiring, it is the right string semiring.
+//
+//   Reverse: a mapping from Weight to ReverseWeight s.t.
+//
+//     --> Reverse(Reverse(a)) = a
+//     --> Reverse(Plus(a, b)) = Plus(Reverse(a), Reverse(b))
+//     --> Reverse(Times(a, b)) = Times(Reverse(b), Reverse(a))
+//     Typically the identity mapping in a (both left and right) semiring.
+//     In the left string semiring, it maps to the reverse string in the right
+//     string semiring.
+//
+//   Properties: specifies additional properties that hold:
+//      LeftSemiring: indicates weights form a left semiring.
+//      RightSemiring: indicates weights form a right semiring.
+//      Commutative: for all a,b: Times(a,b) == Times(b,a)
+//      Idempotent: for all a: Plus(a, a) == a.
+//      Path: for all a, b: Plus(a, b) == a or Plus(a, b) == b.
+
+// CONSTANT DEFINITIONS
+
+// A representable float near .001.
+constexpr float kDelta = 1.0F / 1024.0F;
+
+// For all a, b, c: Times(c, Plus(a, b)) = Plus(Times(c, a), Times(c, b)).
+constexpr uint64 kLeftSemiring = 0x0000000000000001ULL;
+
+// For all a, b, c: Times(Plus(a, b), c) = Plus(Times(a, c), Times(b, c)).
+constexpr uint64 kRightSemiring = 0x0000000000000002ULL;
+
+constexpr uint64 kSemiring = kLeftSemiring | kRightSemiring;
+
+// For all a, b: Times(a, b) = Times(b, a).
+constexpr uint64 kCommutative = 0x0000000000000004ULL;
+
+// For all a: Plus(a, a) = a.
+constexpr uint64 kIdempotent = 0x0000000000000008ULL;
+
+// For all a, b: Plus(a, b) = a or Plus(a, b) = b.
+constexpr uint64 kPath = 0x0000000000000010ULL;
+
+// For random weight generation: default number of distinct weights.
+// This is also used for a few other weight generation defaults.
+constexpr size_t kNumRandomWeights = 5;
+
+// Weight property boolean constants needed for SFINAE.
+
+// MSVC compiler bug workaround: an expression containing W::Properties() cannot
+// be directly used as a value argument to std::enable_if or integral_constant.
+// WeightPropertiesThunk<W>::Properties works instead, however.
+namespace bug {
+template <class W>
+struct WeightPropertiesThunk {
+  WeightPropertiesThunk() = delete;
+  constexpr static const uint64 Properties = W::Properties();
+};
+
+template <class W, uint64 props>
+using TestWeightProperties = std::integral_constant<bool,
+        (WeightPropertiesThunk<W>::Properties & props) == props>;
+}  // namespace bug
+
+template <class W>
+using IsIdempotent = bug::TestWeightProperties<W, kIdempotent>;
+
+template <class W>
+using IsPath = bug::TestWeightProperties<W, kPath>;
+
+
+// Determines direction of division.
+enum DivideType {
+  DIVIDE_LEFT,   // left division
+  DIVIDE_RIGHT,  // right division
+  DIVIDE_ANY
+};  // division in a commutative semiring
+
+// NATURAL ORDER
+//
+// By definition:
+//
+//                 a <= b iff a + b = a
+//
+// The natural order is a negative partial order iff the semiring is
+// idempotent. It is trivially monotonic for plus. It is left
+// (resp. right) monotonic for times iff the semiring is left
+// (resp. right) distributive. It is a total order iff the semiring
+// has the path property.
+//
+// For more information, see:
+//
+// Mohri, M. 2002. Semiring framework and algorithms for shortest-distance
+// problems, Journal of Automata, Languages and
+// Combinatorics 7(3): 321-350, 2002.
+//
+// We define the strict version of this order below.
+
+template <class W>
+class NaturalLess {
+public:
+  using Weight = W;
+
+  NaturalLess() {
+    if (!(W::Properties() & kIdempotent)) {
+      FSTERROR() << "NaturalLess: Weight type is not idempotent: " << W::Type();
+    }
+  }
+
+  bool operator()(const W &w1, const W &w2) const {
+    return (Plus(w1, w2) == w1) && w1 != w2;
+  }
+};
+
+// Power is the iterated product for arbitrary semirings such that Power(w, 0)
+// is One() for the semiring, and Power(w, n) = Times(Power(w, n - 1), w).
+template <class Weight>
+Weight Power(const Weight &weight, size_t n) {
+  auto result = Weight::One();
+  for (size_t i = 0; i < n; ++i) result = Times(result, weight);
+  return result;
+}
+
+// Simple default adder class. Specializations might be more complex.
+template <class Weight>
+class Adder {
+ public:
+  explicit Adder(Weight w = Weight::Zero()) : sum_(w) { }
+
+  Weight Add(const Weight &w) {
+    sum_ = Plus(sum_, w);
+    return sum_;
+  }
+
+  Weight Sum() { return sum_; }
+
+  void Reset(Weight w = Weight::Zero()) { sum_ = w; }
+
+ private:
+  Weight sum_;
+};
+
+// General weight converter: raises error.
+template <class W1, class W2>
+struct WeightConvert {
+  W2 operator()(W1 w1) const {
+    FSTERROR() << "WeightConvert: Can't convert weight from \"" << W1::Type()
+               << "\" to \"" << W2::Type();
+    return W2::NoWeight();
+  }
+};
+
+// Specialized weight converter to self.
+template <class W>
+struct WeightConvert<W, W> {
+  W operator()(W weight) const { return weight; }
+};
+
+// General random weight generator: raises error.
+template <class W>
+struct WeightGenerate {
+  W operator()() const {
+    FSTERROR() << "WeightGenerate: No random generator for " << W::Type();
+    return W::NoWeight();
+  }
+};
+
+namespace internal {
+
+class CompositeWeightIO {
+ public:
+  CompositeWeightIO();
+  CompositeWeightIO(char separator, std::pair<char, char> parentheses);
+
+  std::pair<char, char> parentheses() const {
+    return {open_paren_, close_paren_};
+  }
+  char separator() const { return separator_; }
+
+  bool error() const { return error_; }
+
+ protected:
+  const char separator_;
+  const char open_paren_;
+  const char close_paren_;
+
+ private:
+  bool error_;
+};
+
+}  // namespace internal
+
+// Helper class for writing textual composite weights.
+class CompositeWeightWriter : public internal::CompositeWeightIO {
+ public:
+  // Uses configuration from flags (FLAGS_fst_weight_separator,
+  // FLAGS_fst_weight_parentheses).
+  explicit CompositeWeightWriter(std::ostream &ostrm);
+
+  // parentheses defines the opening and closing parenthesis characters.
+  // Set parentheses = {0, 0} to disable writing parenthesis.
+  CompositeWeightWriter(std::ostream &ostrm, char separator,
+                        std::pair<char, char> parentheses);
+
+  CompositeWeightWriter(const CompositeWeightWriter &) = delete;
+  CompositeWeightWriter &operator=(const CompositeWeightWriter &) = delete;
+
+  // Writes open parenthesis to a stream if option selected.
+  void WriteBegin();
+
+  // Writes element to a stream.
+  template <class T>
+  void WriteElement(const T &comp) {
+    if (i_++ > 0) ostrm_ << separator_;
+    ostrm_ << comp;
+  }
+
+  // Writes close parenthesis to a stream if option selected.
+  void WriteEnd();
+
+ private:
+  std::ostream &ostrm_;
+  int i_ = 0;  // Element position.
+};
+
+// Helper class for reading textual composite weights. Elements are separated by
+// a separator character. There must be at least one element per textual
+// representation.  Parentheses characters should be set if the composite
+// weights themselves contain composite weights to ensure proper parsing.
+class CompositeWeightReader : public internal::CompositeWeightIO {
+ public:
+  // Uses configuration from flags (FLAGS_fst_weight_separator,
+  // FLAGS_fst_weight_parentheses).
+  explicit CompositeWeightReader(std::istream &istrm);
+
+  // parentheses defines the opening and closing parenthesis characters.
+  // Set parentheses = {0, 0} to disable reading parenthesis.
+  CompositeWeightReader(std::istream &istrm, char separator,
+                        std::pair<char, char> parentheses);
+
+  CompositeWeightReader(const CompositeWeightReader &) = delete;
+  CompositeWeightReader &operator=(const CompositeWeightReader &) = delete;
+
+  // Reads open parenthesis from a stream if option selected.
+  void ReadBegin();
+
+  // Reads element from a stream. The second argument, when true, indicates that
+  // this will be the last element (allowing more forgiving formatting of the
+  // last element). Returns false when last element is read.
+  template <class T>
+  bool ReadElement(T *comp, bool last = false);
+
+  // Finalizes reading.
+  void ReadEnd();
+
+ private:
+  std::istream &istrm_;  // Input stream.
+  int c_ = 0;            // Last character read, or EOF.
+  int depth_ = 0;        // Weight parentheses depth.
+};
+
+template <class T>
+inline bool CompositeWeightReader::ReadElement(T *comp, bool last) {
+  string s;
+  const bool has_parens = open_paren_ != 0;
+  while ((c_ != std::istream::traits_type::eof()) && !std::isspace(c_) &&
+         (c_ != separator_ || depth_ > 1 || last) &&
+         (c_ != close_paren_ || depth_ != 1)) {
+    s += c_;
+    // If parentheses encountered before separator, they must be matched.
+    if (has_parens && c_ == open_paren_) {
+      ++depth_;
+    } else if (has_parens && c_ == close_paren_) {
+      // Failure on unmatched parentheses.
+      if (depth_ == 0) {
+        FSTERROR() << "CompositeWeightReader: Unmatched close paren: "
+                   << "Is the fst_weight_parentheses flag set correctly?";
+        istrm_.clear(std::ios::badbit);
+        return false;
+      }
+      --depth_;
+    }
+    c_ = istrm_.get();
+  }
+  if (s.empty()) {
+    FSTERROR() << "CompositeWeightReader: Empty element: "
+               << "Is the fst_weight_parentheses flag set correctly?";
+    istrm_.clear(std::ios::badbit);
+    return false;
+  }
+  std::istringstream istrm(s);
+  istrm >> *comp;
+  // Skips separator/close parenthesis.
+  if (c_ != std::istream::traits_type::eof() && !std::isspace(c_)) {
+    c_ = istrm_.get();
+  }
+  const bool is_eof = c_ == std::istream::traits_type::eof();
+  // Clears fail bit if just EOF.
+  if (is_eof && !istrm_.bad()) istrm_.clear(std::ios::eofbit);
+  return !is_eof && !std::isspace(c_);
+}
+
+}  // namespace fst
+
+#endif  // FST_WEIGHT_H_
diff --git a/projects/llm_framework/include/gflags/defines.h b/projects/llm_framework/include/gflags/defines.h
new file mode 100644
index 00000000..944ed7db
--- /dev/null
+++ b/projects/llm_framework/include/gflags/defines.h
@@ -0,0 +1,48 @@
+/* Generated from defines.h.in during build configuration using CMake. */
+
+// Note: This header file is only used internally. It is not part of public interface!
+//       Any cmakedefine is defined using the -D flag instead when Bazel is used.
+//       For Bazel, this file is thus not used to avoid a private file in $(GENDIR).
+
+#ifndef GFLAGS_DEFINES_H_
+#define GFLAGS_DEFINES_H_
+
+
+// Define if you build this library for a MS Windows OS.
+/* #undef OS_WINDOWS */
+
+// Define if you have the <stdint.h> header file.
+#define HAVE_STDINT_H
+
+// Define if you have the <sys/types.h> header file.
+#define HAVE_SYS_TYPES_H
+
+// Define if you have the <inttypes.h> header file.
+#define HAVE_INTTYPES_H
+
+// Define if you have the <sys/stat.h> header file.
+#define HAVE_SYS_STAT_H
+
+// Define if you have the <unistd.h> header file.
+#define HAVE_UNISTD_H
+
+// Define if you have the <fnmatch.h> header file.
+#define HAVE_FNMATCH_H
+
+// Define if you have the <shlwapi.h> header file (Windows 2000/XP).
+/* #undef HAVE_SHLWAPI_H */
+
+// Define if you have the strtoll function.
+#define HAVE_STRTOLL
+
+// Define if you have the strtoq function.
+/* #undef HAVE_STRTOQ */
+
+// Define if you have the <pthread.h> header file.
+/* #undef HAVE_PTHREAD */
+
+// Define if your pthread library defines the type pthread_rwlock_t
+/* #undef HAVE_RWLOCK */
+
+
+#endif // GFLAGS_DEFINES_H_
diff --git a/projects/llm_framework/include/gflags/gflags.h b/projects/llm_framework/include/gflags/gflags.h
new file mode 100644
index 00000000..9273da8d
--- /dev/null
+++ b/projects/llm_framework/include/gflags/gflags.h
@@ -0,0 +1,626 @@
+// Copyright (c) 2006, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// ---
+// Revamped and reorganized by Craig Silverstein
+//
+// This is the file that should be included by any file which declares
+// or defines a command line flag or wants to parse command line flags
+// or print a program usage message (which will include information about
+// flags).  Executive summary, in the form of an example foo.cc file:
+//
+//    #include "foo.h"         // foo.h has a line "DECLARE_int32(start);"
+//    #include "validators.h"  // hypothetical file defining ValidateIsFile()
+//
+//    DEFINE_int32(end, 1000, "The last record to read");
+//
+//    DEFINE_string(filename, "my_file.txt", "The file to read");
+//    // Crash if the specified file does not exist.
+//    static bool dummy = RegisterFlagValidator(&FLAGS_filename,
+//                                              &ValidateIsFile);
+//
+//    DECLARE_bool(verbose); // some other file has a DEFINE_bool(verbose, ...)
+//
+//    void MyFunc() {
+//      if (FLAGS_verbose) printf("Records %d-%d\n", FLAGS_start, FLAGS_end);
+//    }
+//
+//    Then, at the command-line:
+//       ./foo --noverbose --start=5 --end=100
+//
+// For more details, see
+//    doc/gflags.html
+//
+// --- A note about thread-safety:
+//
+// We describe many functions in this routine as being thread-hostile,
+// thread-compatible, or thread-safe.  Here are the meanings we use:
+//
+// thread-safe: it is safe for multiple threads to call this routine
+//   (or, when referring to a class, methods of this class)
+//   concurrently.
+// thread-hostile: it is not safe for multiple threads to call this
+//   routine (or methods of this class) concurrently.  In gflags,
+//   most thread-hostile routines are intended to be called early in,
+//   or even before, main() -- that is, before threads are spawned.
+// thread-compatible: it is safe for multiple threads to read from
+//   this variable (when applied to variables), or to call const
+//   methods of this class (when applied to classes), as long as no
+//   other thread is writing to the variable or calling non-const
+//   methods of this class.
+
+#ifndef GFLAGS_GFLAGS_H_
+#define GFLAGS_GFLAGS_H_
+
+#include <string>
+#include <vector>
+
+#include "gflags/gflags_declare.h" // IWYU pragma: export
+
+
+// We always want to export variables defined in user code
+#ifndef GFLAGS_DLL_DEFINE_FLAG
+#  if GFLAGS_IS_A_DLL && defined(_MSC_VER)
+#    define GFLAGS_DLL_DEFINE_FLAG __declspec(dllexport)
+#  else
+#    define GFLAGS_DLL_DEFINE_FLAG
+#  endif
+#endif
+
+
+namespace GFLAGS_NAMESPACE {
+
+
+// --------------------------------------------------------------------
+// To actually define a flag in a file, use DEFINE_bool,
+// DEFINE_string, etc. at the bottom of this file.  You may also find
+// it useful to register a validator with the flag.  This ensures that
+// when the flag is parsed from the commandline, or is later set via
+// SetCommandLineOption, we call the validation function. It is _not_
+// called when you assign the value to the flag directly using the = operator.
+//
+// The validation function should return true if the flag value is valid, and
+// false otherwise. If the function returns false for the new setting of the
+// flag, the flag will retain its current value. If it returns false for the
+// default value, ParseCommandLineFlags() will die.
+//
+// This function is safe to call at global construct time (as in the
+// example below).
+//
+// Example use:
+//    static bool ValidatePort(const char* flagname, int32 value) {
+//       if (value > 0 && value < 32768)   // value is ok
+//         return true;
+//       printf("Invalid value for --%s: %d\n", flagname, (int)value);
+//       return false;
+//    }
+//    DEFINE_int32(port, 0, "What port to listen on");
+//    static bool dummy = RegisterFlagValidator(&FLAGS_port, &ValidatePort);
+
+// Returns true if successfully registered, false if not (because the
+// first argument doesn't point to a command-line flag, or because a
+// validator is already registered for this flag).
+extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const bool*        flag, bool (*validate_fn)(const char*, bool));
+extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const int32*       flag, bool (*validate_fn)(const char*, int32));
+extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const uint32*      flag, bool (*validate_fn)(const char*, uint32));
+extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const int64*       flag, bool (*validate_fn)(const char*, int64));
+extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const uint64*      flag, bool (*validate_fn)(const char*, uint64));
+extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const double*      flag, bool (*validate_fn)(const char*, double));
+extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const std::string* flag, bool (*validate_fn)(const char*, const std::string&));
+
+// Convenience macro for the registration of a flag validator
+#define DEFINE_validator(name, validator) \
+    static const bool name##_validator_registered = \
+            GFLAGS_NAMESPACE::RegisterFlagValidator(&FLAGS_##name, validator)
+
+
+// --------------------------------------------------------------------
+// These methods are the best way to get access to info about the
+// list of commandline flags.  Note that these routines are pretty slow.
+//   GetAllFlags: mostly-complete info about the list, sorted by file.
+//   ShowUsageWithFlags: pretty-prints the list to stdout (what --help does)
+//   ShowUsageWithFlagsRestrict: limit to filenames with restrict as a substr
+//
+// In addition to accessing flags, you can also access argv[0] (the program
+// name) and argv (the entire commandline), which we sock away a copy of.
+// These variables are static, so you should only set them once.
+//
+// No need to export this data only structure from DLL, avoiding VS warning 4251.
+struct CommandLineFlagInfo {
+  std::string name;            // the name of the flag
+  std::string type;            // the type of the flag: int32, etc
+  std::string description;     // the "help text" associated with the flag
+  std::string current_value;   // the current value, as a string
+  std::string default_value;   // the default value, as a string
+  std::string filename;        // 'cleaned' version of filename holding the flag
+  bool has_validator_fn;       // true if RegisterFlagValidator called on this flag
+  bool is_default;             // true if the flag has the default value and
+                               // has not been set explicitly from the cmdline
+                               // or via SetCommandLineOption
+  const void* flag_ptr;        // pointer to the flag's current value (i.e. FLAGS_foo)
+};
+
+// Using this inside of a validator is a recipe for a deadlock.
+// TODO(user) Fix locking when validators are running, to make it safe to
+// call validators during ParseAllFlags.
+// Also make sure then to uncomment the corresponding unit test in
+// gflags_unittest.sh
+extern GFLAGS_DLL_DECL void GetAllFlags(std::vector<CommandLineFlagInfo>* OUTPUT);
+// These two are actually defined in gflags_reporting.cc.
+extern GFLAGS_DLL_DECL void ShowUsageWithFlags(const char *argv0);  // what --help does
+extern GFLAGS_DLL_DECL void ShowUsageWithFlagsRestrict(const char *argv0, const char *restrict);
+
+// Create a descriptive string for a flag.
+// Goes to some trouble to make pretty line breaks.
+extern GFLAGS_DLL_DECL std::string DescribeOneFlag(const CommandLineFlagInfo& flag);
+
+// Thread-hostile; meant to be called before any threads are spawned.
+extern GFLAGS_DLL_DECL void SetArgv(int argc, const char** argv);
+
+// The following functions are thread-safe as long as SetArgv() is
+// only called before any threads start.
+extern GFLAGS_DLL_DECL const std::vector<std::string>& GetArgvs();
+extern GFLAGS_DLL_DECL const char* GetArgv();                      // all of argv as a string
+extern GFLAGS_DLL_DECL const char* GetArgv0();                     // only argv0
+extern GFLAGS_DLL_DECL uint32 GetArgvSum();                        // simple checksum of argv
+extern GFLAGS_DLL_DECL const char* ProgramInvocationName();        // argv0, or "UNKNOWN" if not set
+extern GFLAGS_DLL_DECL const char* ProgramInvocationShortName();   // basename(argv0)
+
+// ProgramUsage() is thread-safe as long as SetUsageMessage() is only
+// called before any threads start.
+extern GFLAGS_DLL_DECL const char* ProgramUsage();                 // string set by SetUsageMessage()
+
+// VersionString() is thread-safe as long as SetVersionString() is only
+// called before any threads start.
+extern GFLAGS_DLL_DECL const char* VersionString();                // string set by SetVersionString()
+
+
+
+// --------------------------------------------------------------------
+// Normally you access commandline flags by just saying "if (FLAGS_foo)"
+// or whatever, and set them by calling "FLAGS_foo = bar" (or, more
+// commonly, via the DEFINE_foo macro).  But if you need a bit more
+// control, we have programmatic ways to get/set the flags as well.
+// These programmatic ways to access flags are thread-safe, but direct
+// access is only thread-compatible.
+
+// Return true iff the flagname was found.
+// OUTPUT is set to the flag's value, or unchanged if we return false.
+extern GFLAGS_DLL_DECL bool GetCommandLineOption(const char* name, std::string* OUTPUT);
+
+// Return true iff the flagname was found. OUTPUT is set to the flag's
+// CommandLineFlagInfo or unchanged if we return false.
+extern GFLAGS_DLL_DECL bool GetCommandLineFlagInfo(const char* name, CommandLineFlagInfo* OUTPUT);
+
+// Return the CommandLineFlagInfo of the flagname.  exit() if name not found.
+// Example usage, to check if a flag's value is currently the default value:
+//   if (GetCommandLineFlagInfoOrDie("foo").is_default) ...
+extern GFLAGS_DLL_DECL CommandLineFlagInfo GetCommandLineFlagInfoOrDie(const char* name);
+
+enum GFLAGS_DLL_DECL FlagSettingMode {
+  // update the flag's value (can call this multiple times).
+  SET_FLAGS_VALUE,
+  // update the flag's value, but *only if* it has not yet been updated
+  // with SET_FLAGS_VALUE, SET_FLAG_IF_DEFAULT, or "FLAGS_xxx = nondef".
+  SET_FLAG_IF_DEFAULT,
+  // set the flag's default value to this.  If the flag has not yet updated
+  // yet (via SET_FLAGS_VALUE, SET_FLAG_IF_DEFAULT, or "FLAGS_xxx = nondef")
+  // change the flag's current value to the new default value as well.
+  SET_FLAGS_DEFAULT
+};
+
+// Set a particular flag ("command line option").  Returns a string
+// describing the new value that the option has been set to.  The
+// return value API is not well-specified, so basically just depend on
+// it to be empty if the setting failed for some reason -- the name is
+// not a valid flag name, or the value is not a valid value -- and
+// non-empty else.
+
+// SetCommandLineOption uses set_mode == SET_FLAGS_VALUE (the common case)
+extern GFLAGS_DLL_DECL std::string SetCommandLineOption        (const char* name, const char* value);
+extern GFLAGS_DLL_DECL std::string SetCommandLineOptionWithMode(const char* name, const char* value, FlagSettingMode set_mode);
+
+
+// --------------------------------------------------------------------
+// Saves the states (value, default value, whether the user has set
+// the flag, registered validators, etc) of all flags, and restores
+// them when the FlagSaver is destroyed.  This is very useful in
+// tests, say, when you want to let your tests change the flags, but
+// make sure that they get reverted to the original states when your
+// test is complete.
+//
+// Example usage:
+//   void TestFoo() {
+//     FlagSaver s1;
+//     FLAG_foo = false;
+//     FLAG_bar = "some value";
+//
+//     // test happens here.  You can return at any time
+//     // without worrying about restoring the FLAG values.
+//   }
+//
+// Note: This class is marked with GFLAGS_ATTRIBUTE_UNUSED because all
+// the work is done in the constructor and destructor, so in the standard
+// usage example above, the compiler would complain that it's an
+// unused variable.
+//
+// This class is thread-safe.  However, its destructor writes to
+// exactly the set of flags that have changed value during its
+// lifetime, so concurrent _direct_ access to those flags
+// (i.e. FLAGS_foo instead of {Get,Set}CommandLineOption()) is unsafe.
+
+class GFLAGS_DLL_DECL FlagSaver {
+ public:
+  FlagSaver();
+  ~FlagSaver();
+
+ private:
+  class FlagSaverImpl* impl_;   // we use pimpl here to keep API steady
+
+  FlagSaver(const FlagSaver&);  // no copying!
+  void operator=(const FlagSaver&);
+}__attribute((unused));
+
+// --------------------------------------------------------------------
+// Some deprecated or hopefully-soon-to-be-deprecated functions.
+
+// This is often used for logging.  TODO(csilvers): figure out a better way
+extern GFLAGS_DLL_DECL std::string CommandlineFlagsIntoString();
+// Usually where this is used, a FlagSaver should be used instead.
+extern GFLAGS_DLL_DECL
+bool ReadFlagsFromString(const std::string& flagfilecontents,
+                         const char* prog_name,
+                         bool errors_are_fatal);  // uses SET_FLAGS_VALUE
+
+// These let you manually implement --flagfile functionality.
+// DEPRECATED.
+extern GFLAGS_DLL_DECL bool AppendFlagsIntoFile(const std::string& filename, const char* prog_name);
+extern GFLAGS_DLL_DECL bool ReadFromFlagsFile(const std::string& filename, const char* prog_name, bool errors_are_fatal);   // uses SET_FLAGS_VALUE
+
+
+// --------------------------------------------------------------------
+// Useful routines for initializing flags from the environment.
+// In each case, if 'varname' does not exist in the environment
+// return defval.  If 'varname' does exist but is not valid
+// (e.g., not a number for an int32 flag), abort with an error.
+// Otherwise, return the value.  NOTE: for booleans, for true use
+// 't' or 'T' or 'true' or '1', for false 'f' or 'F' or 'false' or '0'.
+
+extern GFLAGS_DLL_DECL bool BoolFromEnv(const char *varname, bool defval);
+extern GFLAGS_DLL_DECL int32 Int32FromEnv(const char *varname, int32 defval);
+extern GFLAGS_DLL_DECL uint32 Uint32FromEnv(const char *varname, uint32 defval);
+extern GFLAGS_DLL_DECL int64 Int64FromEnv(const char *varname, int64 defval);
+extern GFLAGS_DLL_DECL uint64 Uint64FromEnv(const char *varname, uint64 defval);
+extern GFLAGS_DLL_DECL double DoubleFromEnv(const char *varname, double defval);
+extern GFLAGS_DLL_DECL const char *StringFromEnv(const char *varname, const char *defval);
+
+
+// --------------------------------------------------------------------
+// The next two functions parse gflags from main():
+
+// Set the "usage" message for this program.  For example:
+//   string usage("This program does nothing.  Sample usage:\n");
+//   usage += argv[0] + " <uselessarg1> <uselessarg2>";
+//   SetUsageMessage(usage);
+// Do not include commandline flags in the usage: we do that for you!
+// Thread-hostile; meant to be called before any threads are spawned.
+extern GFLAGS_DLL_DECL void SetUsageMessage(const std::string& usage);
+
+// Sets the version string, which is emitted with --version.
+// For instance: SetVersionString("1.3");
+// Thread-hostile; meant to be called before any threads are spawned.
+extern GFLAGS_DLL_DECL void SetVersionString(const std::string& version);
+
+
+// Looks for flags in argv and parses them.  Rearranges argv to put
+// flags first, or removes them entirely if remove_flags is true.
+// If a flag is defined more than once in the command line or flag
+// file, the last definition is used.  Returns the index (into argv)
+// of the first non-flag argument.
+// See top-of-file for more details on this function.
+#ifndef SWIG   // In swig, use ParseCommandLineFlagsScript() instead.
+extern GFLAGS_DLL_DECL uint32 ParseCommandLineFlags(int *argc, char*** argv, bool remove_flags);
+#endif
+
+
+// Calls to ParseCommandLineNonHelpFlags and then to
+// HandleCommandLineHelpFlags can be used instead of a call to
+// ParseCommandLineFlags during initialization, in order to allow for
+// changing default values for some FLAGS (via
+// e.g. SetCommandLineOptionWithMode calls) between the time of
+// command line parsing and the time of dumping help information for
+// the flags as a result of command line parsing.  If a flag is
+// defined more than once in the command line or flag file, the last
+// definition is used.  Returns the index (into argv) of the first
+// non-flag argument.  (If remove_flags is true, will always return 1.)
+extern GFLAGS_DLL_DECL uint32 ParseCommandLineNonHelpFlags(int *argc, char*** argv, bool remove_flags);
+
+// This is actually defined in gflags_reporting.cc.
+// This function is misnamed (it also handles --version, etc.), but
+// it's too late to change that now. :-(
+extern GFLAGS_DLL_DECL void HandleCommandLineHelpFlags();   // in gflags_reporting.cc
+
+// Allow command line reparsing.  Disables the error normally
+// generated when an unknown flag is found, since it may be found in a
+// later parse.  Thread-hostile; meant to be called before any threads
+// are spawned.
+extern GFLAGS_DLL_DECL void AllowCommandLineReparsing();
+
+// Reparse the flags that have not yet been recognized.  Only flags
+// registered since the last parse will be recognized.  Any flag value
+// must be provided as part of the argument using "=", not as a
+// separate command line argument that follows the flag argument.
+// Intended for handling flags from dynamically loaded libraries,
+// since their flags are not registered until they are loaded.
+extern GFLAGS_DLL_DECL void ReparseCommandLineNonHelpFlags();
+
+// Clean up memory allocated by flags.  This is only needed to reduce
+// the quantity of "potentially leaked" reports emitted by memory
+// debugging tools such as valgrind.  It is not required for normal
+// operation, or for the google perftools heap-checker.  It must only
+// be called when the process is about to exit, and all threads that
+// might access flags are quiescent.  Referencing flags after this is
+// called will have unexpected consequences.  This is not safe to run
+// when multiple threads might be running: the function is
+// thread-hostile.
+extern GFLAGS_DLL_DECL void ShutDownCommandLineFlags();
+
+
+// --------------------------------------------------------------------
+// Now come the command line flag declaration/definition macros that
+// will actually be used.  They're kind of hairy.  A major reason
+// for this is initialization: we want people to be able to access
+// variables in global constructors and have that not crash, even if
+// their global constructor runs before the global constructor here.
+// (Obviously, we can't guarantee the flags will have the correct
+// default value in that case, but at least accessing them is safe.)
+// The only way to do that is have flags point to a static buffer.
+// So we make one, using a union to ensure proper alignment, and
+// then use placement-new to actually set up the flag with the
+// correct default value.  In the same vein, we have to worry about
+// flag access in global destructors, so FlagRegisterer has to be
+// careful never to destroy the flag-values it constructs.
+//
+// Note that when we define a flag variable FLAGS_<name>, we also
+// preemptively define a junk variable, FLAGS_no<name>.  This is to
+// cause a link-time error if someone tries to define 2 flags with
+// names like "logging" and "nologging".  We do this because a bool
+// flag FLAG can be set from the command line to true with a "-FLAG"
+// argument, and to false with a "-noFLAG" argument, and so this can
+// potentially avert confusion.
+//
+// We also put flags into their own namespace.  It is purposefully
+// named in an opaque way that people should have trouble typing
+// directly.  The idea is that DEFINE puts the flag in the weird
+// namespace, and DECLARE imports the flag from there into the current
+// namespace.  The net result is to force people to use DECLARE to get
+// access to a flag, rather than saying "extern GFLAGS_DLL_DECL bool FLAGS_whatever;"
+// or some such instead.  We want this so we can put extra
+// functionality (like sanity-checking) in DECLARE if we want, and
+// make sure it is picked up everywhere.
+//
+// We also put the type of the variable in the namespace, so that
+// people can't DECLARE_int32 something that they DEFINE_bool'd
+// elsewhere.
+
+class GFLAGS_DLL_DECL FlagRegisterer {
+ public:
+  // We instantiate this template ctor for all supported types,
+  // so it is possible to place implementation of the FlagRegisterer ctor in
+  // .cc file.
+  // Calling this constructor with unsupported type will produce linker error.
+  template <typename FlagType>
+  FlagRegisterer(const char* name,
+                 const char* help, const char* filename,
+                 FlagType* current_storage, FlagType* defvalue_storage);
+};
+
+// Force compiler to not generate code for the given template specialization.
+#if defined(_MSC_VER) && _MSC_VER < 1800 // Visual Studio 2013 version 12.0
+  #define GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(type)
+#else
+  #define GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(type)                  \
+    extern template GFLAGS_DLL_DECL FlagRegisterer::FlagRegisterer(  \
+        const char* name, const char* help, const char* filename,    \
+        type* current_storage, type* defvalue_storage)
+#endif
+
+// Do this for all supported flag types.
+GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(bool);
+GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(int32);
+GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(uint32);
+GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(int64);
+GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(uint64);
+GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(double);
+GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(std::string);
+
+#undef GFLAGS_DECLARE_FLAG_REGISTERER_CTOR
+
+// If your application #defines STRIP_FLAG_HELP to a non-zero value
+// before #including this file, we remove the help message from the
+// binary file. This can reduce the size of the resulting binary
+// somewhat, and may also be useful for security reasons.
+
+extern GFLAGS_DLL_DECL const char kStrippedFlagHelp[];
+
+
+} // namespace GFLAGS_NAMESPACE
+
+
+#ifndef SWIG  // In swig, ignore the main flag declarations
+
+#if defined(STRIP_FLAG_HELP) && STRIP_FLAG_HELP > 0
+// Need this construct to avoid the 'defined but not used' warning.
+#define MAYBE_STRIPPED_HELP(txt) \
+   (false ? (txt) : GFLAGS_NAMESPACE::kStrippedFlagHelp)
+#else
+#define MAYBE_STRIPPED_HELP(txt) txt
+#endif
+
+// Each command-line flag has two variables associated with it: one
+// with the current value, and one with the default value.  However,
+// we have a third variable, which is where value is assigned; it's a
+// constant.  This guarantees that FLAG_##value is initialized at
+// static initialization time (e.g. before program-start) rather than
+// than global construction time (which is after program-start but
+// before main), at least when 'value' is a compile-time constant.  We
+// use a small trick for the "default value" variable, and call it
+// FLAGS_no<name>.  This serves the second purpose of assuring a
+// compile error if someone tries to define a flag named no<name>
+// which is illegal (--foo and --nofoo both affect the "foo" flag).
+#define DEFINE_VARIABLE(type, shorttype, name, value, help)             \
+  namespace fL##shorttype {                                             \
+    static const type FLAGS_nono##name = value;                         \
+    /* We always want to export defined variables, dll or no */         \
+    GFLAGS_DLL_DEFINE_FLAG type FLAGS_##name = FLAGS_nono##name;        \
+    static type FLAGS_no##name = FLAGS_nono##name;                      \
+    static GFLAGS_NAMESPACE::FlagRegisterer o_##name(                   \
+      #name, MAYBE_STRIPPED_HELP(help), __FILE__,                       \
+      &FLAGS_##name, &FLAGS_no##name);                                  \
+  }                                                                     \
+  using fL##shorttype::FLAGS_##name
+
+// For DEFINE_bool, we want to do the extra check that the passed-in
+// value is actually a bool, and not a string or something that can be
+// coerced to a bool.  These declarations (no definition needed!) will
+// help us do that, and never evaluate From, which is important.
+// We'll use 'sizeof(IsBool(val))' to distinguish. This code requires
+// that the compiler have different sizes for bool & double. Since
+// this is not guaranteed by the standard, we check it with a
+// COMPILE_ASSERT.
+namespace fLB {
+struct CompileAssert {};
+typedef CompileAssert expected_sizeof_double_neq_sizeof_bool[
+                      (sizeof(double) != sizeof(bool)) ? 1 : -1];
+template<typename From> double GFLAGS_DLL_DECL IsBoolFlag(const From& from);
+GFLAGS_DLL_DECL bool IsBoolFlag(bool from);
+}  // namespace fLB
+
+// Here are the actual DEFINE_*-macros. The respective DECLARE_*-macros
+// are in a separate include, gflags_declare.h, for reducing
+// the physical transitive size for DECLARE use.
+#define DEFINE_bool(name, val, txt)                                     \
+  namespace fLB {                                                       \
+    typedef ::fLB::CompileAssert FLAG_##name##_value_is_not_a_bool[     \
+            (sizeof(::fLB::IsBoolFlag(val)) != sizeof(double))? 1: -1]; \
+  }                                                                     \
+  DEFINE_VARIABLE(bool, B, name, val, txt)
+
+#define DEFINE_int32(name, val, txt) \
+   DEFINE_VARIABLE(GFLAGS_NAMESPACE::int32, I, \
+                   name, val, txt)
+
+#define DEFINE_uint32(name,val, txt) \
+   DEFINE_VARIABLE(GFLAGS_NAMESPACE::uint32, U, \
+                   name, val, txt)
+
+#define DEFINE_int64(name, val, txt) \
+   DEFINE_VARIABLE(GFLAGS_NAMESPACE::int64, I64, \
+                   name, val, txt)
+
+#define DEFINE_uint64(name,val, txt) \
+   DEFINE_VARIABLE(GFLAGS_NAMESPACE::uint64, U64, \
+                   name, val, txt)
+
+#define DEFINE_double(name, val, txt) \
+   DEFINE_VARIABLE(double, D, name, val, txt)
+
+// Strings are trickier, because they're not a POD, so we can't
+// construct them at static-initialization time (instead they get
+// constructed at global-constructor time, which is much later).  To
+// try to avoid crashes in that case, we use a char buffer to store
+// the string, which we can static-initialize, and then placement-new
+// into it later.  It's not perfect, but the best we can do.
+
+namespace fLS {
+
+inline clstring* dont_pass0toDEFINE_string(char *stringspot,
+                                           const char *value) {
+  return new(stringspot) clstring(value);
+}
+inline clstring* dont_pass0toDEFINE_string(char *stringspot,
+                                           const clstring &value) {
+  return new(stringspot) clstring(value);
+}
+inline clstring* dont_pass0toDEFINE_string(char *stringspot,
+                                           int value);
+
+// Auxiliary class used to explicitly call destructor of string objects
+// allocated using placement new during static program deinitialization.
+// The destructor MUST be an inline function such that the explicit
+// destruction occurs in the same compilation unit as the placement new.
+class StringFlagDestructor {
+  void *current_storage_;
+  void *defvalue_storage_;
+
+public: 
+
+  StringFlagDestructor(void *current, void *defvalue)
+  : current_storage_(current), defvalue_storage_(defvalue) {}
+
+  ~StringFlagDestructor() {
+    reinterpret_cast<clstring*>(current_storage_ )->~clstring();
+    reinterpret_cast<clstring*>(defvalue_storage_)->~clstring();
+  }
+};
+
+}  // namespace fLS
+
+// We need to define a var named FLAGS_no##name so people don't define
+// --string and --nostring.  And we need a temporary place to put val
+// so we don't have to evaluate it twice.  Two great needs that go
+// great together!
+// The weird 'using' + 'extern' inside the fLS namespace is to work around
+// an unknown compiler bug/issue with the gcc 4.2.1 on SUSE 10.  See
+//    http://code.google.com/p/google-gflags/issues/detail?id=20
+#define DEFINE_string(name, val, txt)                                       \
+  namespace fLS {                                                           \
+    using ::fLS::clstring;                                                  \
+    using ::fLS::StringFlagDestructor;                                      \
+    static union { void* align; char s[sizeof(clstring)]; } s_##name[2];    \
+    clstring* const FLAGS_no##name = ::fLS::                                \
+                                   dont_pass0toDEFINE_string(s_##name[0].s, \
+                                                             val);          \
+    static GFLAGS_NAMESPACE::FlagRegisterer o_##name(                       \
+        #name, MAYBE_STRIPPED_HELP(txt), __FILE__,                          \
+        FLAGS_no##name, new (s_##name[1].s) clstring(*FLAGS_no##name));     \
+    static StringFlagDestructor d_##name(s_##name[0].s, s_##name[1].s);     \
+    extern GFLAGS_DLL_DEFINE_FLAG clstring& FLAGS_##name;                   \
+    using fLS::FLAGS_##name;                                                \
+    clstring& FLAGS_##name = *FLAGS_no##name;                               \
+  }                                                                         \
+  using fLS::FLAGS_##name
+
+#endif  // SWIG
+
+
+
+
+
+#endif  // GFLAGS_GFLAGS_H_
diff --git a/projects/llm_framework/include/gflags/gflags_completions.h b/projects/llm_framework/include/gflags/gflags_completions.h
new file mode 100644
index 00000000..2fa0db6d
--- /dev/null
+++ b/projects/llm_framework/include/gflags/gflags_completions.h
@@ -0,0 +1,121 @@
+// Copyright (c) 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// ---
+
+//
+// Implement helpful bash-style command line flag completions
+//
+// ** Functional API:
+// HandleCommandLineCompletions() should be called early during
+// program startup, but after command line flag code has been
+// initialized, such as the beginning of HandleCommandLineHelpFlags().
+// It checks the value of the flag --tab_completion_word.  If this
+// flag is empty, nothing happens here.  If it contains a string,
+// however, then HandleCommandLineCompletions() will hijack the
+// process, attempting to identify the intention behind this
+// completion.  Regardless of the outcome of this deduction, the
+// process will be terminated, similar to --helpshort flag
+// handling.
+//
+// ** Overview of Bash completions:
+// Bash can be told to programatically determine completions for the
+// current 'cursor word'.  It does this by (in this case) invoking a
+// command with some additional arguments identifying the command
+// being executed, the word being completed, and the previous word
+// (if any).  Bash then expects a sequence of output lines to be
+// printed to stdout.  If these lines all contain a common prefix
+// longer than the cursor word, bash will replace the cursor word
+// with that common prefix, and display nothing.  If there isn't such
+// a common prefix, bash will display the lines in pages using 'more'.
+//
+// ** Strategy taken for command line completions:
+// If we can deduce either the exact flag intended, or a common flag
+// prefix, we'll output exactly that.  Otherwise, if information
+// must be displayed to the user, we'll take the opportunity to add
+// some helpful information beyond just the flag name (specifically,
+// we'll include the default flag value and as much of the flag's
+// description as can fit on a single terminal line width, as specified
+// by the flag --tab_completion_columns).  Furthermore, we'll try to
+// make bash order the output such that the most useful or relevent
+// flags are the most likely to be shown at the top.
+//
+// ** Additional features:
+// To assist in finding that one really useful flag, substring matching
+// was implemented.  Before pressing a <TAB> to get completion for the
+// current word, you can append one or more '?' to the flag to do
+// substring matching.  Here's the semantics:
+//   --foo<TAB>     Show me all flags with names prefixed by 'foo'
+//   --foo?<TAB>    Show me all flags with 'foo' somewhere in the name
+//   --foo??<TAB>   Same as prior case, but also search in module
+//                  definition path for 'foo'
+//   --foo???<TAB>  Same as prior case, but also search in flag
+//                  descriptions for 'foo'
+// Finally, we'll trim the output to a relatively small number of
+// flags to keep bash quiet about the verbosity of output.  If one
+// really wanted to see all possible matches, appending a '+' to the
+// search word will force the exhaustive list of matches to be printed.
+//
+// ** How to have bash accept completions from a binary:
+// Bash requires that it be informed about each command that programmatic
+// completion should be enabled for.  Example addition to a .bashrc
+// file would be (your path to gflags_completions.sh file may differ):
+
+/*
+$ complete -o bashdefault -o default -o nospace -C                            \
+ '/home/build/eng/bash/bash_completions.sh --tab_completion_columns $COLUMNS' \
+  time  env  binary_name  another_binary  [...]
+*/
+
+// This would allow the following to work:
+//   $ /path/to/binary_name --vmodule<TAB>
+// Or:
+//   $ ./bin/path/another_binary --gfs_u<TAB>
+// (etc)
+//
+// Sadly, it appears that bash gives no easy way to force this behavior for
+// all commands.  That's where the "time" in the above example comes in.
+// If you haven't specifically added a command to the list of completion
+// supported commands, you can still get completions by prefixing the
+// entire command with "env".
+//   $ env /some/brand/new/binary --vmod<TAB>
+// Assuming that "binary" is a newly compiled binary, this should still
+// produce the expected completion output.
+
+
+#ifndef GFLAGS_COMPLETIONS_H_
+#define GFLAGS_COMPLETIONS_H_
+
+namespace gflags {
+
+extern void HandleCommandLineCompletions(void);
+
+}
+
+#endif  // GFLAGS_COMPLETIONS_H_
diff --git a/projects/llm_framework/include/gflags/gflags_declare.h b/projects/llm_framework/include/gflags/gflags_declare.h
new file mode 100644
index 00000000..69cf1129
--- /dev/null
+++ b/projects/llm_framework/include/gflags/gflags_declare.h
@@ -0,0 +1,156 @@
+// Copyright (c) 1999, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// ---
+//
+// Revamped and reorganized by Craig Silverstein
+//
+// This is the file that should be included by any file which declares
+// command line flag.
+
+#ifndef GFLAGS_DECLARE_H_
+#define GFLAGS_DECLARE_H_
+
+
+// ---------------------------------------------------------------------------
+// Namespace of gflags library symbols.
+#define GFLAGS_NAMESPACE gflags
+
+// ---------------------------------------------------------------------------
+// Windows DLL import/export.
+
+// Whether gflags library is a DLL.
+//
+// Set to 1 by default when the shared gflags library was built on Windows.
+// Must be overwritten when this header file is used with the optionally also
+// built static library instead; set by CMake's INTERFACE_COMPILE_DEFINITIONS.
+#ifndef GFLAGS_IS_A_DLL
+#  define GFLAGS_IS_A_DLL 1
+#endif
+
+// We always want to import the symbols of the gflags library.
+#ifndef GFLAGS_DLL_DECL
+#  if GFLAGS_IS_A_DLL && defined(_MSC_VER)
+#    define GFLAGS_DLL_DECL __declspec(dllimport)
+#  elif defined(__GNUC__) && __GNUC__ >= 4
+#    define GFLAGS_DLL_DECL __attribute__((visibility("default")))
+#  else
+#    define GFLAGS_DLL_DECL
+#  endif
+#endif
+
+// We always want to import variables declared in user code.
+#ifndef GFLAGS_DLL_DECLARE_FLAG
+#  if GFLAGS_IS_A_DLL && defined(_MSC_VER)
+#    define GFLAGS_DLL_DECLARE_FLAG __declspec(dllimport)
+#  elif defined(__GNUC__) && __GNUC__ >= 4
+#    define GFLAGS_DLL_DECLARE_FLAG __attribute__((visibility("default")))
+#  else
+#    define GFLAGS_DLL_DECLARE_FLAG
+#  endif
+#endif
+
+// ---------------------------------------------------------------------------
+// Flag types
+#include <string>
+#if 1
+#  include <stdint.h>                   // the normal place uint32_t is defined
+#elif 1
+#  include <sys/types.h>                // the normal place u_int32_t is defined
+#elif 1
+#  include <inttypes.h>                 // a third place for uint32_t or u_int32_t
+#endif
+
+namespace GFLAGS_NAMESPACE {
+
+#if 1 // C99
+typedef int32_t          int32;
+typedef uint32_t         uint32;
+typedef int64_t          int64;
+typedef uint64_t         uint64;
+#elif 0 // BSD
+typedef int32_t          int32;
+typedef u_int32_t        uint32;
+typedef int64_t          int64;
+typedef u_int64_t        uint64;
+#elif 0 // Windows
+typedef __int32          int32;
+typedef unsigned __int32 uint32;
+typedef __int64          int64;
+typedef unsigned __int64 uint64;
+#else
+#  error Do not know how to define a 32-bit integer quantity on your system
+#endif
+
+} // namespace GFLAGS_NAMESPACE
+
+
+namespace fLS {
+
+// The meaning of "string" might be different between now and when the
+// macros below get invoked (e.g., if someone is experimenting with
+// other string implementations that get defined after this file is
+// included).  Save the current meaning now and use it in the macros.
+typedef std::string clstring;
+
+} // namespace fLS
+
+
+#define DECLARE_VARIABLE(type, shorttype, name) \
+  /* We always want to import declared variables, dll or no */ \
+  namespace fL##shorttype { extern GFLAGS_DLL_DECLARE_FLAG type FLAGS_##name; } \
+  using fL##shorttype::FLAGS_##name
+
+#define DECLARE_bool(name) \
+  DECLARE_VARIABLE(bool, B, name)
+
+#define DECLARE_int32(name) \
+  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::int32, I, name)
+
+#define DECLARE_uint32(name) \
+  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::uint32, U, name)
+
+#define DECLARE_int64(name) \
+  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::int64, I64, name)
+
+#define DECLARE_uint64(name) \
+  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::uint64, U64, name)
+
+#define DECLARE_double(name) \
+  DECLARE_VARIABLE(double, D, name)
+
+#define DECLARE_string(name) \
+  /* We always want to import declared variables, dll or no */ \
+  namespace fLS { \
+  extern GFLAGS_DLL_DECLARE_FLAG ::fLS::clstring& FLAGS_##name; \
+  } \
+  using fLS::FLAGS_##name
+
+
+#endif  // GFLAGS_DECLARE_H_
diff --git a/projects/llm_framework/include/glog/log_severity.h b/projects/llm_framework/include/glog/log_severity.h
new file mode 100644
index 00000000..99945a42
--- /dev/null
+++ b/projects/llm_framework/include/glog/log_severity.h
@@ -0,0 +1,92 @@
+// Copyright (c) 2007, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+#ifndef BASE_LOG_SEVERITY_H__
+#define BASE_LOG_SEVERITY_H__
+
+// Annoying stuff for windows -- makes sure clients can import these functions
+#ifndef GOOGLE_GLOG_DLL_DECL
+# if defined(_WIN32) && !defined(__CYGWIN__)
+#   define GOOGLE_GLOG_DLL_DECL  __declspec(dllimport)
+# else
+#   define GOOGLE_GLOG_DLL_DECL
+# endif
+#endif
+
+// Variables of type LogSeverity are widely taken to lie in the range
+// [0, NUM_SEVERITIES-1].  Be careful to preserve this assumption if
+// you ever need to change their values or add a new severity.
+typedef int LogSeverity;
+
+const int GLOG_INFO = 0, GLOG_WARNING = 1, GLOG_ERROR = 2, GLOG_FATAL = 3,
+  NUM_SEVERITIES = 4;
+#ifndef GLOG_NO_ABBREVIATED_SEVERITIES
+# ifdef ERROR
+#  error ERROR macro is defined. Define GLOG_NO_ABBREVIATED_SEVERITIES before including logging.h. See the document for detail.
+# endif
+const int INFO = GLOG_INFO, WARNING = GLOG_WARNING,
+  ERROR = GLOG_ERROR, FATAL = GLOG_FATAL;
+#endif
+
+// DFATAL is FATAL in debug mode, ERROR in normal mode
+#ifdef NDEBUG
+#define DFATAL_LEVEL ERROR
+#else
+#define DFATAL_LEVEL FATAL
+#endif
+
+extern GOOGLE_GLOG_DLL_DECL const char* const LogSeverityNames[NUM_SEVERITIES];
+
+// NDEBUG usage helpers related to (RAW_)DCHECK:
+//
+// DEBUG_MODE is for small !NDEBUG uses like
+//   if (DEBUG_MODE) foo.CheckThatFoo();
+// instead of substantially more verbose
+//   #ifndef NDEBUG
+//     foo.CheckThatFoo();
+//   #endif
+//
+// IF_DEBUG_MODE is for small !NDEBUG uses like
+//   IF_DEBUG_MODE( string error; )
+//   DCHECK(Foo(&error)) << error;
+// instead of substantially more verbose
+//   #ifndef NDEBUG
+//     string error;
+//     DCHECK(Foo(&error)) << error;
+//   #endif
+//
+#ifdef NDEBUG
+enum { DEBUG_MODE = 0 };
+#define IF_DEBUG_MODE(x)
+#else
+enum { DEBUG_MODE = 1 };
+#define IF_DEBUG_MODE(x) x
+#endif
+
+#endif  // BASE_LOG_SEVERITY_H__
diff --git a/projects/llm_framework/include/glog/logging.h b/projects/llm_framework/include/glog/logging.h
new file mode 100644
index 00000000..4cd247e1
--- /dev/null
+++ b/projects/llm_framework/include/glog/logging.h
@@ -0,0 +1,1662 @@
+// Copyright (c) 1999, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: Ray Sidney
+//
+// This file contains #include information about logging-related stuff.
+// Pretty much everybody needs to #include this file so that they can
+// log various happenings.
+//
+#ifndef _LOGGING_H_
+#define _LOGGING_H_
+
+#include <errno.h>
+#include <string.h>
+#include <time.h>
+#include <iosfwd>
+#include <ostream>
+#include <sstream>
+#include <string>
+#if 1
+# include <unistd.h>
+#endif
+#include <vector>
+
+#if defined(_MSC_VER)
+#define GLOG_MSVC_PUSH_DISABLE_WARNING(n) __pragma(warning(push)) \
+                                     __pragma(warning(disable:n))
+#define GLOG_MSVC_POP_WARNING() __pragma(warning(pop))
+#else
+#define GLOG_MSVC_PUSH_DISABLE_WARNING(n)
+#define GLOG_MSVC_POP_WARNING()
+#endif
+
+// Annoying stuff for windows -- makes sure clients can import these functions
+#ifndef GOOGLE_GLOG_DLL_DECL
+# if defined(_WIN32) && !defined(__CYGWIN__)
+#   define GOOGLE_GLOG_DLL_DECL  __declspec(dllimport)
+# else
+#   define GOOGLE_GLOG_DLL_DECL
+# endif
+#endif
+
+// We care a lot about number of bits things take up.  Unfortunately,
+// systems define their bit-specific ints in a lot of different ways.
+// We use our own way, and have a typedef to get there.
+// Note: these commands below may look like "#if 1" or "#if 0", but
+// that's because they were constructed that way at ./configure time.
+// Look at logging.h.in to see how they're calculated (based on your config).
+#if 1
+#include <stdint.h>             // the normal place uint16_t is defined
+#endif
+#if 1
+#include <sys/types.h>          // the normal place u_int16_t is defined
+#endif
+#if 1
+#include <inttypes.h>           // a third place for uint16_t or u_int16_t
+#endif
+
+#if 0
+#include <gflags/gflags.h>
+#endif
+
+namespace google {
+
+#if 1      // the C99 format
+typedef int32_t int32;
+typedef uint32_t uint32;
+typedef int64_t int64;
+typedef uint64_t uint64;
+#elif 1   // the BSD format
+typedef int32_t int32;
+typedef u_int32_t uint32;
+typedef int64_t int64;
+typedef u_int64_t uint64;
+#elif 0    // the windows (vc7) format
+typedef __int32 int32;
+typedef unsigned __int32 uint32;
+typedef __int64 int64;
+typedef unsigned __int64 uint64;
+#else
+#error Do not know how to define a 32-bit integer quantity on your system
+#endif
+
+}
+
+// The global value of GOOGLE_STRIP_LOG. All the messages logged to
+// LOG(XXX) with severity less than GOOGLE_STRIP_LOG will not be displayed.
+// If it can be determined at compile time that the message will not be
+// printed, the statement will be compiled out.
+//
+// Example: to strip out all INFO and WARNING messages, use the value
+// of 2 below. To make an exception for WARNING messages from a single
+// file, add "#define GOOGLE_STRIP_LOG 1" to that file _before_ including
+// base/logging.h
+#ifndef GOOGLE_STRIP_LOG
+#define GOOGLE_STRIP_LOG 0
+#endif
+
+// GCC can be told that a certain branch is not likely to be taken (for
+// instance, a CHECK failure), and use that information in static analysis.
+// Giving it this information can help it optimize for the common case in
+// the absence of better information (ie. -fprofile-arcs).
+//
+#ifndef GOOGLE_PREDICT_BRANCH_NOT_TAKEN
+#if 1
+#define GOOGLE_PREDICT_BRANCH_NOT_TAKEN(x) (__builtin_expect(x, 0))
+#else
+#define GOOGLE_PREDICT_BRANCH_NOT_TAKEN(x) x
+#endif
+#endif
+
+#ifndef GOOGLE_PREDICT_FALSE
+#if 1
+#define GOOGLE_PREDICT_FALSE(x) (__builtin_expect(x, 0))
+#else
+#define GOOGLE_PREDICT_FALSE(x) x
+#endif
+#endif
+
+#ifndef GOOGLE_PREDICT_TRUE
+#if 1
+#define GOOGLE_PREDICT_TRUE(x) (__builtin_expect(!!(x), 1))
+#else
+#define GOOGLE_PREDICT_TRUE(x) x
+#endif
+#endif
+
+
+// Make a bunch of macros for logging.  The way to log things is to stream
+// things to LOG(<a particular severity level>).  E.g.,
+//
+//   LOG(INFO) << "Found " << num_cookies << " cookies";
+//
+// You can capture log messages in a string, rather than reporting them
+// immediately:
+//
+//   vector<string> errors;
+//   LOG_STRING(ERROR, &errors) << "Couldn't parse cookie #" << cookie_num;
+//
+// This pushes back the new error onto 'errors'; if given a NULL pointer,
+// it reports the error via LOG(ERROR).
+//
+// You can also do conditional logging:
+//
+//   LOG_IF(INFO, num_cookies > 10) << "Got lots of cookies";
+//
+// You can also do occasional logging (log every n'th occurrence of an
+// event):
+//
+//   LOG_EVERY_N(INFO, 10) << "Got the " << google::COUNTER << "th cookie";
+//
+// The above will cause log messages to be output on the 1st, 11th, 21st, ...
+// times it is executed.  Note that the special google::COUNTER value is used
+// to identify which repetition is happening.
+//
+// You can also do occasional conditional logging (log every n'th
+// occurrence of an event, when condition is satisfied):
+//
+//   LOG_IF_EVERY_N(INFO, (size > 1024), 10) << "Got the " << google::COUNTER
+//                                           << "th big cookie";
+//
+// You can log messages the first N times your code executes a line. E.g.
+//
+//   LOG_FIRST_N(INFO, 20) << "Got the " << google::COUNTER << "th cookie";
+//
+// Outputs log messages for the first 20 times it is executed.
+//
+// Analogous SYSLOG, SYSLOG_IF, and SYSLOG_EVERY_N macros are available.
+// These log to syslog as well as to the normal logs.  If you use these at
+// all, you need to be aware that syslog can drastically reduce performance,
+// especially if it is configured for remote logging!  Don't use these
+// unless you fully understand this and have a concrete need to use them.
+// Even then, try to minimize your use of them.
+//
+// There are also "debug mode" logging macros like the ones above:
+//
+//   DLOG(INFO) << "Found cookies";
+//
+//   DLOG_IF(INFO, num_cookies > 10) << "Got lots of cookies";
+//
+//   DLOG_EVERY_N(INFO, 10) << "Got the " << google::COUNTER << "th cookie";
+//
+// All "debug mode" logging is compiled away to nothing for non-debug mode
+// compiles.
+//
+// We also have
+//
+//   LOG_ASSERT(assertion);
+//   DLOG_ASSERT(assertion);
+//
+// which is syntactic sugar for {,D}LOG_IF(FATAL, assert fails) << assertion;
+//
+// There are "verbose level" logging macros.  They look like
+//
+//   VLOG(1) << "I'm printed when you run the program with --v=1 or more";
+//   VLOG(2) << "I'm printed when you run the program with --v=2 or more";
+//
+// These always log at the INFO log level (when they log at all).
+// The verbose logging can also be turned on module-by-module.  For instance,
+//    --vmodule=mapreduce=2,file=1,gfs*=3 --v=0
+// will cause:
+//   a. VLOG(2) and lower messages to be printed from mapreduce.{h,cc}
+//   b. VLOG(1) and lower messages to be printed from file.{h,cc}
+//   c. VLOG(3) and lower messages to be printed from files prefixed with "gfs"
+//   d. VLOG(0) and lower messages to be printed from elsewhere
+//
+// The wildcarding functionality shown by (c) supports both '*' (match
+// 0 or more characters) and '?' (match any single character) wildcards.
+//
+// There's also VLOG_IS_ON(n) "verbose level" condition macro. To be used as
+//
+//   if (VLOG_IS_ON(2)) {
+//     // do some logging preparation and logging
+//     // that can't be accomplished with just VLOG(2) << ...;
+//   }
+//
+// There are also VLOG_IF, VLOG_EVERY_N and VLOG_IF_EVERY_N "verbose level"
+// condition macros for sample cases, when some extra computation and
+// preparation for logs is not needed.
+//   VLOG_IF(1, (size > 1024))
+//      << "I'm printed when size is more than 1024 and when you run the "
+//         "program with --v=1 or more";
+//   VLOG_EVERY_N(1, 10)
+//      << "I'm printed every 10th occurrence, and when you run the program "
+//         "with --v=1 or more. Present occurence is " << google::COUNTER;
+//   VLOG_IF_EVERY_N(1, (size > 1024), 10)
+//      << "I'm printed on every 10th occurence of case when size is more "
+//         " than 1024, when you run the program with --v=1 or more. ";
+//         "Present occurence is " << google::COUNTER;
+//
+// The supported severity levels for macros that allow you to specify one
+// are (in increasing order of severity) INFO, WARNING, ERROR, and FATAL.
+// Note that messages of a given severity are logged not only in the
+// logfile for that severity, but also in all logfiles of lower severity.
+// E.g., a message of severity FATAL will be logged to the logfiles of
+// severity FATAL, ERROR, WARNING, and INFO.
+//
+// There is also the special severity of DFATAL, which logs FATAL in
+// debug mode, ERROR in normal mode.
+//
+// Very important: logging a message at the FATAL severity level causes
+// the program to terminate (after the message is logged).
+//
+// Unless otherwise specified, logs will be written to the filename
+// "<program name>.<hostname>.<user name>.log.<severity level>.", followed
+// by the date, time, and pid (you can't prevent the date, time, and pid
+// from being in the filename).
+//
+// The logging code takes two flags:
+//     --v=#           set the verbose level
+//     --logtostderr   log all the messages to stderr instead of to logfiles
+
+// LOG LINE PREFIX FORMAT
+//
+// Log lines have this form:
+//
+//     Lmmdd hh:mm:ss.uuuuuu threadid file:line] msg...
+//
+// where the fields are defined as follows:
+//
+//   L                A single character, representing the log level
+//                    (eg 'I' for INFO)
+//   mm               The month (zero padded; ie May is '05')
+//   dd               The day (zero padded)
+//   hh:mm:ss.uuuuuu  Time in hours, minutes and fractional seconds
+//   threadid         The space-padded thread ID as returned by GetTID()
+//                    (this matches the PID on Linux)
+//   file             The file name
+//   line             The line number
+//   msg              The user-supplied message
+//
+// Example:
+//
+//   I1103 11:57:31.739339 24395 google.cc:2341] Command line: ./some_prog
+//   I1103 11:57:31.739403 24395 google.cc:2342] Process id 24395
+//
+// NOTE: although the microseconds are useful for comparing events on
+// a single machine, clocks on different machines may not be well
+// synchronized.  Hence, use caution when comparing the low bits of
+// timestamps from different machines.
+
+#ifndef DECLARE_VARIABLE
+#define MUST_UNDEF_GFLAGS_DECLARE_MACROS
+#define DECLARE_VARIABLE(type, shorttype, name, tn)                     \
+  namespace fL##shorttype {                                             \
+    extern GOOGLE_GLOG_DLL_DECL type FLAGS_##name;                      \
+  }                                                                     \
+  using fL##shorttype::FLAGS_##name
+
+// bool specialization
+#define DECLARE_bool(name) \
+  DECLARE_VARIABLE(bool, B, name, bool)
+
+// int32 specialization
+#define DECLARE_int32(name) \
+  DECLARE_VARIABLE(google::int32, I, name, int32)
+
+// Special case for string, because we have to specify the namespace
+// std::string, which doesn't play nicely with our FLAG__namespace hackery.
+#define DECLARE_string(name)                                            \
+  namespace fLS {                                                       \
+    extern GOOGLE_GLOG_DLL_DECL std::string& FLAGS_##name;              \
+  }                                                                     \
+  using fLS::FLAGS_##name
+#endif
+
+// Set whether log messages go to stderr instead of logfiles
+DECLARE_bool(logtostderr);
+
+// Set whether log messages go to stderr in addition to logfiles.
+DECLARE_bool(alsologtostderr);
+
+// Set color messages logged to stderr (if supported by terminal).
+DECLARE_bool(colorlogtostderr);
+
+// Log messages at a level >= this flag are automatically sent to
+// stderr in addition to log files.
+DECLARE_int32(stderrthreshold);
+
+// Set whether the log prefix should be prepended to each line of output.
+DECLARE_bool(log_prefix);
+
+// Log messages at a level <= this flag are buffered.
+// Log messages at a higher level are flushed immediately.
+DECLARE_int32(logbuflevel);
+
+// Sets the maximum number of seconds which logs may be buffered for.
+DECLARE_int32(logbufsecs);
+
+// Log suppression level: messages logged at a lower level than this
+// are suppressed.
+DECLARE_int32(minloglevel);
+
+// If specified, logfiles are written into this directory instead of the
+// default logging directory.
+DECLARE_string(log_dir);
+
+// Set the log file mode.
+DECLARE_int32(logfile_mode);
+
+// Sets the path of the directory into which to put additional links
+// to the log files.
+DECLARE_string(log_link);
+
+DECLARE_int32(v);  // in vlog_is_on.cc
+
+// Sets the maximum log file size (in MB).
+DECLARE_int32(max_log_size);
+
+// Sets whether to avoid logging to the disk if the disk is full.
+DECLARE_bool(stop_logging_if_full_disk);
+
+#ifdef MUST_UNDEF_GFLAGS_DECLARE_MACROS
+#undef MUST_UNDEF_GFLAGS_DECLARE_MACROS
+#undef DECLARE_VARIABLE
+#undef DECLARE_bool
+#undef DECLARE_int32
+#undef DECLARE_string
+#endif
+
+// Log messages below the GOOGLE_STRIP_LOG level will be compiled away for
+// security reasons. See LOG(severtiy) below.
+
+// A few definitions of macros that don't generate much code.  Since
+// LOG(INFO) and its ilk are used all over our code, it's
+// better to have compact code for these operations.
+
+#if GOOGLE_STRIP_LOG == 0
+#define COMPACT_GOOGLE_LOG_INFO google::LogMessage( \
+      __FILE__, __LINE__)
+#define LOG_TO_STRING_INFO(message) google::LogMessage( \
+      __FILE__, __LINE__, google::GLOG_INFO, message)
+#else
+#define COMPACT_GOOGLE_LOG_INFO google::NullStream()
+#define LOG_TO_STRING_INFO(message) google::NullStream()
+#endif
+
+#if GOOGLE_STRIP_LOG <= 1
+#define COMPACT_GOOGLE_LOG_WARNING google::LogMessage( \
+      __FILE__, __LINE__, google::GLOG_WARNING)
+#define LOG_TO_STRING_WARNING(message) google::LogMessage( \
+      __FILE__, __LINE__, google::GLOG_WARNING, message)
+#else
+#define COMPACT_GOOGLE_LOG_WARNING google::NullStream()
+#define LOG_TO_STRING_WARNING(message) google::NullStream()
+#endif
+
+#if GOOGLE_STRIP_LOG <= 2
+#define COMPACT_GOOGLE_LOG_ERROR google::LogMessage( \
+      __FILE__, __LINE__, google::GLOG_ERROR)
+#define LOG_TO_STRING_ERROR(message) google::LogMessage( \
+      __FILE__, __LINE__, google::GLOG_ERROR, message)
+#else
+#define COMPACT_GOOGLE_LOG_ERROR google::NullStream()
+#define LOG_TO_STRING_ERROR(message) google::NullStream()
+#endif
+
+#if GOOGLE_STRIP_LOG <= 3
+#define COMPACT_GOOGLE_LOG_FATAL google::LogMessageFatal( \
+      __FILE__, __LINE__)
+#define LOG_TO_STRING_FATAL(message) google::LogMessage( \
+      __FILE__, __LINE__, google::GLOG_FATAL, message)
+#else
+#define COMPACT_GOOGLE_LOG_FATAL google::NullStreamFatal()
+#define LOG_TO_STRING_FATAL(message) google::NullStreamFatal()
+#endif
+
+#if defined(NDEBUG) && !defined(DCHECK_ALWAYS_ON)
+#define DCHECK_IS_ON() 0
+#else
+#define DCHECK_IS_ON() 1
+#endif
+
+// For DFATAL, we want to use LogMessage (as opposed to
+// LogMessageFatal), to be consistent with the original behavior.
+#if !DCHECK_IS_ON()
+#define COMPACT_GOOGLE_LOG_DFATAL COMPACT_GOOGLE_LOG_ERROR
+#elif GOOGLE_STRIP_LOG <= 3
+#define COMPACT_GOOGLE_LOG_DFATAL google::LogMessage( \
+      __FILE__, __LINE__, google::GLOG_FATAL)
+#else
+#define COMPACT_GOOGLE_LOG_DFATAL google::NullStreamFatal()
+#endif
+
+#define GOOGLE_LOG_INFO(counter) google::LogMessage(__FILE__, __LINE__, google::GLOG_INFO, counter, &google::LogMessage::SendToLog)
+#define SYSLOG_INFO(counter) \
+  google::LogMessage(__FILE__, __LINE__, google::GLOG_INFO, counter, \
+  &google::LogMessage::SendToSyslogAndLog)
+#define GOOGLE_LOG_WARNING(counter)  \
+  google::LogMessage(__FILE__, __LINE__, google::GLOG_WARNING, counter, \
+  &google::LogMessage::SendToLog)
+#define SYSLOG_WARNING(counter)  \
+  google::LogMessage(__FILE__, __LINE__, google::GLOG_WARNING, counter, \
+  &google::LogMessage::SendToSyslogAndLog)
+#define GOOGLE_LOG_ERROR(counter)  \
+  google::LogMessage(__FILE__, __LINE__, google::GLOG_ERROR, counter, \
+  &google::LogMessage::SendToLog)
+#define SYSLOG_ERROR(counter)  \
+  google::LogMessage(__FILE__, __LINE__, google::GLOG_ERROR, counter, \
+  &google::LogMessage::SendToSyslogAndLog)
+#define GOOGLE_LOG_FATAL(counter) \
+  google::LogMessage(__FILE__, __LINE__, google::GLOG_FATAL, counter, \
+  &google::LogMessage::SendToLog)
+#define SYSLOG_FATAL(counter) \
+  google::LogMessage(__FILE__, __LINE__, google::GLOG_FATAL, counter, \
+  &google::LogMessage::SendToSyslogAndLog)
+#define GOOGLE_LOG_DFATAL(counter) \
+  google::LogMessage(__FILE__, __LINE__, google::DFATAL_LEVEL, counter, \
+  &google::LogMessage::SendToLog)
+#define SYSLOG_DFATAL(counter) \
+  google::LogMessage(__FILE__, __LINE__, google::DFATAL_LEVEL, counter, \
+  &google::LogMessage::SendToSyslogAndLog)
+
+#if defined(WIN32) || defined(_WIN32) || defined(__WIN32__) || defined(__CYGWIN__) || defined(__CYGWIN32__)
+// A very useful logging macro to log windows errors:
+#define LOG_SYSRESULT(result) \
+  if (FAILED(HRESULT_FROM_WIN32(result))) { \
+    LPSTR message = NULL; \
+    LPSTR msg = reinterpret_cast<LPSTR>(&message); \
+    DWORD message_length = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | \
+                         FORMAT_MESSAGE_FROM_SYSTEM, \
+                         0, result, 0, msg, 100, NULL); \
+    if (message_length > 0) { \
+      google::LogMessage(__FILE__, __LINE__, google::GLOG_ERROR, 0, \
+          &google::LogMessage::SendToLog).stream() \
+          << reinterpret_cast<const char*>(message); \
+      LocalFree(message); \
+    } \
+  }
+#endif
+
+// We use the preprocessor's merging operator, "##", so that, e.g.,
+// LOG(INFO) becomes the token GOOGLE_LOG_INFO.  There's some funny
+// subtle difference between ostream member streaming functions (e.g.,
+// ostream::operator<<(int) and ostream non-member streaming functions
+// (e.g., ::operator<<(ostream&, string&): it turns out that it's
+// impossible to stream something like a string directly to an unnamed
+// ostream. We employ a neat hack by calling the stream() member
+// function of LogMessage which seems to avoid the problem.
+#define LOG(severity) COMPACT_GOOGLE_LOG_ ## severity.stream()
+#define SYSLOG(severity) SYSLOG_ ## severity(0).stream()
+
+namespace google {
+
+// They need the definitions of integer types.
+#include "glog/log_severity.h"
+#include "glog/vlog_is_on.h"
+
+// Initialize google's logging library. You will see the program name
+// specified by argv0 in log outputs.
+GOOGLE_GLOG_DLL_DECL void InitGoogleLogging(const char* argv0);
+
+// Shutdown google's logging library.
+GOOGLE_GLOG_DLL_DECL void ShutdownGoogleLogging();
+
+// Install a function which will be called after LOG(FATAL).
+GOOGLE_GLOG_DLL_DECL void InstallFailureFunction(void (*fail_func)());
+
+class LogSink;  // defined below
+
+// If a non-NULL sink pointer is given, we push this message to that sink.
+// For LOG_TO_SINK we then do normal LOG(severity) logging as well.
+// This is useful for capturing messages and passing/storing them
+// somewhere more specific than the global log of the process.
+// Argument types:
+//   LogSink* sink;
+//   LogSeverity severity;
+// The cast is to disambiguate NULL arguments.
+#define LOG_TO_SINK(sink, severity) \
+  google::LogMessage(                                    \
+      __FILE__, __LINE__,                                               \
+      google::GLOG_ ## severity,                         \
+      static_cast<google::LogSink*>(sink), true).stream()
+#define LOG_TO_SINK_BUT_NOT_TO_LOGFILE(sink, severity)                  \
+  google::LogMessage(                                    \
+      __FILE__, __LINE__,                                               \
+      google::GLOG_ ## severity,                         \
+      static_cast<google::LogSink*>(sink), false).stream()
+
+// If a non-NULL string pointer is given, we write this message to that string.
+// We then do normal LOG(severity) logging as well.
+// This is useful for capturing messages and storing them somewhere more
+// specific than the global log of the process.
+// Argument types:
+//   string* message;
+//   LogSeverity severity;
+// The cast is to disambiguate NULL arguments.
+// NOTE: LOG(severity) expands to LogMessage().stream() for the specified
+// severity.
+#define LOG_TO_STRING(severity, message) \
+  LOG_TO_STRING_##severity(static_cast<string*>(message)).stream()
+
+// If a non-NULL pointer is given, we push the message onto the end
+// of a vector of strings; otherwise, we report it with LOG(severity).
+// This is handy for capturing messages and perhaps passing them back
+// to the caller, rather than reporting them immediately.
+// Argument types:
+//   LogSeverity severity;
+//   vector<string> *outvec;
+// The cast is to disambiguate NULL arguments.
+#define LOG_STRING(severity, outvec) \
+  LOG_TO_STRING_##severity(static_cast<std::vector<std::string>*>(outvec)).stream()
+
+#define LOG_IF(severity, condition) \
+  static_cast<void>(0),             \
+  !(condition) ? (void) 0 : google::LogMessageVoidify() & LOG(severity)
+#define SYSLOG_IF(severity, condition) \
+  static_cast<void>(0),                \
+  !(condition) ? (void) 0 : google::LogMessageVoidify() & SYSLOG(severity)
+
+#define LOG_ASSERT(condition)  \
+  LOG_IF(FATAL, !(condition)) << "Assert failed: " #condition
+#define SYSLOG_ASSERT(condition) \
+  SYSLOG_IF(FATAL, !(condition)) << "Assert failed: " #condition
+
+// CHECK dies with a fatal error if condition is not true.  It is *not*
+// controlled by DCHECK_IS_ON(), so the check will be executed regardless of
+// compilation mode.  Therefore, it is safe to do things like:
+//    CHECK(fp->Write(x) == 4)
+#define CHECK(condition)  \
+      LOG_IF(FATAL, GOOGLE_PREDICT_BRANCH_NOT_TAKEN(!(condition))) \
+             << "Check failed: " #condition " "
+
+// A container for a string pointer which can be evaluated to a bool -
+// true iff the pointer is NULL.
+struct CheckOpString {
+  CheckOpString(std::string* str) : str_(str) { }
+  // No destructor: if str_ is non-NULL, we're about to LOG(FATAL),
+  // so there's no point in cleaning up str_.
+  operator bool() const {
+    return GOOGLE_PREDICT_BRANCH_NOT_TAKEN(str_ != NULL);
+  }
+  std::string* str_;
+};
+
+// Function is overloaded for integral types to allow static const
+// integrals declared in classes and not defined to be used as arguments to
+// CHECK* macros. It's not encouraged though.
+template <class T>
+inline const T&       GetReferenceableValue(const T&           t) { return t; }
+inline char           GetReferenceableValue(char               t) { return t; }
+inline unsigned char  GetReferenceableValue(unsigned char      t) { return t; }
+inline signed char    GetReferenceableValue(signed char        t) { return t; }
+inline short          GetReferenceableValue(short              t) { return t; }
+inline unsigned short GetReferenceableValue(unsigned short     t) { return t; }
+inline int            GetReferenceableValue(int                t) { return t; }
+inline unsigned int   GetReferenceableValue(unsigned int       t) { return t; }
+inline long           GetReferenceableValue(long               t) { return t; }
+inline unsigned long  GetReferenceableValue(unsigned long      t) { return t; }
+inline long long      GetReferenceableValue(long long          t) { return t; }
+inline unsigned long long GetReferenceableValue(unsigned long long t) {
+  return t;
+}
+
+// This is a dummy class to define the following operator.
+struct DummyClassToDefineOperator {};
+
+}
+
+// Define global operator<< to declare using ::operator<<.
+// This declaration will allow use to use CHECK macros for user
+// defined classes which have operator<< (e.g., stl_logging.h).
+inline std::ostream& operator<<(
+    std::ostream& out, const google::DummyClassToDefineOperator&) {
+  return out;
+}
+
+namespace google {
+
+// This formats a value for a failing CHECK_XX statement.  Ordinarily,
+// it uses the definition for operator<<, with a few special cases below.
+template <typename T>
+inline void MakeCheckOpValueString(std::ostream* os, const T& v) {
+  (*os) << v;
+}
+
+// Overrides for char types provide readable values for unprintable
+// characters.
+template <> GOOGLE_GLOG_DLL_DECL
+void MakeCheckOpValueString(std::ostream* os, const char& v);
+template <> GOOGLE_GLOG_DLL_DECL
+void MakeCheckOpValueString(std::ostream* os, const signed char& v);
+template <> GOOGLE_GLOG_DLL_DECL
+void MakeCheckOpValueString(std::ostream* os, const unsigned char& v);
+
+// Build the error message string. Specify no inlining for code size.
+template <typename T1, typename T2>
+std::string* MakeCheckOpString(const T1& v1, const T2& v2, const char* exprtext)
+    __attribute__((noinline));
+
+namespace base {
+namespace internal {
+
+// If "s" is less than base_logging::INFO, returns base_logging::INFO.
+// If "s" is greater than base_logging::FATAL, returns
+// base_logging::ERROR.  Otherwise, returns "s".
+LogSeverity NormalizeSeverity(LogSeverity s);
+
+}  // namespace internal
+
+// A helper class for formatting "expr (V1 vs. V2)" in a CHECK_XX
+// statement.  See MakeCheckOpString for sample usage.  Other
+// approaches were considered: use of a template method (e.g.,
+// base::BuildCheckOpString(exprtext, base::Print<T1>, &v1,
+// base::Print<T2>, &v2), however this approach has complications
+// related to volatile arguments and function-pointer arguments).
+class GOOGLE_GLOG_DLL_DECL CheckOpMessageBuilder {
+ public:
+  // Inserts "exprtext" and " (" to the stream.
+  explicit CheckOpMessageBuilder(const char *exprtext);
+  // Deletes "stream_".
+  ~CheckOpMessageBuilder();
+  // For inserting the first variable.
+  std::ostream* ForVar1() { return stream_; }
+  // For inserting the second variable (adds an intermediate " vs. ").
+  std::ostream* ForVar2();
+  // Get the result (inserts the closing ")").
+  std::string* NewString();
+
+ private:
+  std::ostringstream *stream_;
+};
+
+}  // namespace base
+
+template <typename T1, typename T2>
+std::string* MakeCheckOpString(const T1& v1, const T2& v2, const char* exprtext) {
+  base::CheckOpMessageBuilder comb(exprtext);
+  MakeCheckOpValueString(comb.ForVar1(), v1);
+  MakeCheckOpValueString(comb.ForVar2(), v2);
+  return comb.NewString();
+}
+
+// Helper functions for CHECK_OP macro.
+// The (int, int) specialization works around the issue that the compiler
+// will not instantiate the template version of the function on values of
+// unnamed enum type - see comment below.
+#define DEFINE_CHECK_OP_IMPL(name, op) \
+  template <typename T1, typename T2> \
+  inline std::string* name##Impl(const T1& v1, const T2& v2,    \
+                            const char* exprtext) { \
+    if (GOOGLE_PREDICT_TRUE(v1 op v2)) return NULL; \
+    else return MakeCheckOpString(v1, v2, exprtext); \
+  } \
+  inline std::string* name##Impl(int v1, int v2, const char* exprtext) { \
+    return name##Impl<int, int>(v1, v2, exprtext); \
+  }
+
+// We use the full name Check_EQ, Check_NE, etc. in case the file including
+// base/logging.h provides its own #defines for the simpler names EQ, NE, etc.
+// This happens if, for example, those are used as token names in a
+// yacc grammar.
+DEFINE_CHECK_OP_IMPL(Check_EQ, ==)  // Compilation error with CHECK_EQ(NULL, x)?
+DEFINE_CHECK_OP_IMPL(Check_NE, !=)  // Use CHECK(x == NULL) instead.
+DEFINE_CHECK_OP_IMPL(Check_LE, <=)
+DEFINE_CHECK_OP_IMPL(Check_LT, < )
+DEFINE_CHECK_OP_IMPL(Check_GE, >=)
+DEFINE_CHECK_OP_IMPL(Check_GT, > )
+#undef DEFINE_CHECK_OP_IMPL
+
+// Helper macro for binary operators.
+// Don't use this macro directly in your code, use CHECK_EQ et al below.
+
+#if defined(STATIC_ANALYSIS)
+// Only for static analysis tool to know that it is equivalent to assert
+#define CHECK_OP_LOG(name, op, val1, val2, log) CHECK((val1) op (val2))
+#elif DCHECK_IS_ON()
+// In debug mode, avoid constructing CheckOpStrings if possible,
+// to reduce the overhead of CHECK statments by 2x.
+// Real DCHECK-heavy tests have seen 1.5x speedups.
+
+// The meaning of "string" might be different between now and
+// when this macro gets invoked (e.g., if someone is experimenting
+// with other string implementations that get defined after this
+// file is included).  Save the current meaning now and use it
+// in the macro.
+typedef std::string _Check_string;
+#define CHECK_OP_LOG(name, op, val1, val2, log)                         \
+  while (google::_Check_string* _result =                \
+         google::Check##name##Impl(                      \
+             google::GetReferenceableValue(val1),        \
+             google::GetReferenceableValue(val2),        \
+             #val1 " " #op " " #val2))                                  \
+    log(__FILE__, __LINE__,                                             \
+        google::CheckOpString(_result)).stream()
+#else
+// In optimized mode, use CheckOpString to hint to compiler that
+// the while condition is unlikely.
+#define CHECK_OP_LOG(name, op, val1, val2, log)                         \
+  while (google::CheckOpString _result =                 \
+         google::Check##name##Impl(                      \
+             google::GetReferenceableValue(val1),        \
+             google::GetReferenceableValue(val2),        \
+             #val1 " " #op " " #val2))                                  \
+    log(__FILE__, __LINE__, _result).stream()
+#endif  // STATIC_ANALYSIS, DCHECK_IS_ON()
+
+#if GOOGLE_STRIP_LOG <= 3
+#define CHECK_OP(name, op, val1, val2) \
+  CHECK_OP_LOG(name, op, val1, val2, google::LogMessageFatal)
+#else
+#define CHECK_OP(name, op, val1, val2) \
+  CHECK_OP_LOG(name, op, val1, val2, google::NullStreamFatal)
+#endif // STRIP_LOG <= 3
+
+// Equality/Inequality checks - compare two values, and log a FATAL message
+// including the two values when the result is not as expected.  The values
+// must have operator<<(ostream, ...) defined.
+//
+// You may append to the error message like so:
+//   CHECK_NE(1, 2) << ": The world must be ending!";
+//
+// We are very careful to ensure that each argument is evaluated exactly
+// once, and that anything which is legal to pass as a function argument is
+// legal here.  In particular, the arguments may be temporary expressions
+// which will end up being destroyed at the end of the apparent statement,
+// for example:
+//   CHECK_EQ(string("abc")[1], 'b');
+//
+// WARNING: These don't compile correctly if one of the arguments is a pointer
+// and the other is NULL. To work around this, simply static_cast NULL to the
+// type of the desired pointer.
+
+#define CHECK_EQ(val1, val2) CHECK_OP(_EQ, ==, val1, val2)
+#define CHECK_NE(val1, val2) CHECK_OP(_NE, !=, val1, val2)
+#define CHECK_LE(val1, val2) CHECK_OP(_LE, <=, val1, val2)
+#define CHECK_LT(val1, val2) CHECK_OP(_LT, < , val1, val2)
+#define CHECK_GE(val1, val2) CHECK_OP(_GE, >=, val1, val2)
+#define CHECK_GT(val1, val2) CHECK_OP(_GT, > , val1, val2)
+
+// Check that the input is non NULL.  This very useful in constructor
+// initializer lists.
+
+#define CHECK_NOTNULL(val) \
+  google::CheckNotNull(__FILE__, __LINE__, "'" #val "' Must be non NULL", (val))
+
+// Helper functions for string comparisons.
+// To avoid bloat, the definitions are in logging.cc.
+#define DECLARE_CHECK_STROP_IMPL(func, expected) \
+  GOOGLE_GLOG_DLL_DECL std::string* Check##func##expected##Impl( \
+      const char* s1, const char* s2, const char* names);
+DECLARE_CHECK_STROP_IMPL(strcmp, true)
+DECLARE_CHECK_STROP_IMPL(strcmp, false)
+DECLARE_CHECK_STROP_IMPL(strcasecmp, true)
+DECLARE_CHECK_STROP_IMPL(strcasecmp, false)
+#undef DECLARE_CHECK_STROP_IMPL
+
+// Helper macro for string comparisons.
+// Don't use this macro directly in your code, use CHECK_STREQ et al below.
+#define CHECK_STROP(func, op, expected, s1, s2) \
+  while (google::CheckOpString _result = \
+         google::Check##func##expected##Impl((s1), (s2), \
+                                     #s1 " " #op " " #s2)) \
+    LOG(FATAL) << *_result.str_
+
+
+// String (char*) equality/inequality checks.
+// CASE versions are case-insensitive.
+//
+// Note that "s1" and "s2" may be temporary strings which are destroyed
+// by the compiler at the end of the current "full expression"
+// (e.g. CHECK_STREQ(Foo().c_str(), Bar().c_str())).
+
+#define CHECK_STREQ(s1, s2) CHECK_STROP(strcmp, ==, true, s1, s2)
+#define CHECK_STRNE(s1, s2) CHECK_STROP(strcmp, !=, false, s1, s2)
+#define CHECK_STRCASEEQ(s1, s2) CHECK_STROP(strcasecmp, ==, true, s1, s2)
+#define CHECK_STRCASENE(s1, s2) CHECK_STROP(strcasecmp, !=, false, s1, s2)
+
+#define CHECK_INDEX(I,A) CHECK(I < (sizeof(A)/sizeof(A[0])))
+#define CHECK_BOUND(B,A) CHECK(B <= (sizeof(A)/sizeof(A[0])))
+
+#define CHECK_DOUBLE_EQ(val1, val2)              \
+  do {                                           \
+    CHECK_LE((val1), (val2)+0.000000000000001L); \
+    CHECK_GE((val1), (val2)-0.000000000000001L); \
+  } while (0)
+
+#define CHECK_NEAR(val1, val2, margin)           \
+  do {                                           \
+    CHECK_LE((val1), (val2)+(margin));           \
+    CHECK_GE((val1), (val2)-(margin));           \
+  } while (0)
+
+// perror()..googly style!
+//
+// PLOG() and PLOG_IF() and PCHECK() behave exactly like their LOG* and
+// CHECK equivalents with the addition that they postpend a description
+// of the current state of errno to their output lines.
+
+#define PLOG(severity) GOOGLE_PLOG(severity, 0).stream()
+
+#define GOOGLE_PLOG(severity, counter)  \
+  google::ErrnoLogMessage( \
+      __FILE__, __LINE__, google::GLOG_ ## severity, counter, \
+      &google::LogMessage::SendToLog)
+
+#define PLOG_IF(severity, condition) \
+  static_cast<void>(0),              \
+  !(condition) ? (void) 0 : google::LogMessageVoidify() & PLOG(severity)
+
+// A CHECK() macro that postpends errno if the condition is false. E.g.
+//
+// if (poll(fds, nfds, timeout) == -1) { PCHECK(errno == EINTR); ... }
+#define PCHECK(condition)  \
+      PLOG_IF(FATAL, GOOGLE_PREDICT_BRANCH_NOT_TAKEN(!(condition))) \
+              << "Check failed: " #condition " "
+
+// A CHECK() macro that lets you assert the success of a function that
+// returns -1 and sets errno in case of an error. E.g.
+//
+// CHECK_ERR(mkdir(path, 0700));
+//
+// or
+//
+// int fd = open(filename, flags); CHECK_ERR(fd) << ": open " << filename;
+#define CHECK_ERR(invocation)                                          \
+PLOG_IF(FATAL, GOOGLE_PREDICT_BRANCH_NOT_TAKEN((invocation) == -1))    \
+        << #invocation
+
+// Use macro expansion to create, for each use of LOG_EVERY_N(), static
+// variables with the __LINE__ expansion as part of the variable name.
+#define LOG_EVERY_N_VARNAME(base, line) LOG_EVERY_N_VARNAME_CONCAT(base, line)
+#define LOG_EVERY_N_VARNAME_CONCAT(base, line) base ## line
+
+#define LOG_OCCURRENCES LOG_EVERY_N_VARNAME(occurrences_, __LINE__)
+#define LOG_OCCURRENCES_MOD_N LOG_EVERY_N_VARNAME(occurrences_mod_n_, __LINE__)
+
+#define SOME_KIND_OF_LOG_EVERY_N(severity, n, what_to_do) \
+  static int LOG_OCCURRENCES = 0, LOG_OCCURRENCES_MOD_N = 0; \
+  ++LOG_OCCURRENCES; \
+  if (++LOG_OCCURRENCES_MOD_N > n) LOG_OCCURRENCES_MOD_N -= n; \
+  if (LOG_OCCURRENCES_MOD_N == 1) \
+    google::LogMessage( \
+        __FILE__, __LINE__, google::GLOG_ ## severity, LOG_OCCURRENCES, \
+        &what_to_do).stream()
+
+#define SOME_KIND_OF_LOG_IF_EVERY_N(severity, condition, n, what_to_do) \
+  static int LOG_OCCURRENCES = 0, LOG_OCCURRENCES_MOD_N = 0; \
+  ++LOG_OCCURRENCES; \
+  if (condition && \
+      ((LOG_OCCURRENCES_MOD_N=(LOG_OCCURRENCES_MOD_N + 1) % n) == (1 % n))) \
+    google::LogMessage( \
+        __FILE__, __LINE__, google::GLOG_ ## severity, LOG_OCCURRENCES, \
+                 &what_to_do).stream()
+
+#define SOME_KIND_OF_PLOG_EVERY_N(severity, n, what_to_do) \
+  static int LOG_OCCURRENCES = 0, LOG_OCCURRENCES_MOD_N = 0; \
+  ++LOG_OCCURRENCES; \
+  if (++LOG_OCCURRENCES_MOD_N > n) LOG_OCCURRENCES_MOD_N -= n; \
+  if (LOG_OCCURRENCES_MOD_N == 1) \
+    google::ErrnoLogMessage( \
+        __FILE__, __LINE__, google::GLOG_ ## severity, LOG_OCCURRENCES, \
+        &what_to_do).stream()
+
+#define SOME_KIND_OF_LOG_FIRST_N(severity, n, what_to_do) \
+  static int LOG_OCCURRENCES = 0; \
+  if (LOG_OCCURRENCES <= n) \
+    ++LOG_OCCURRENCES; \
+  if (LOG_OCCURRENCES <= n) \
+    google::LogMessage( \
+        __FILE__, __LINE__, google::GLOG_ ## severity, LOG_OCCURRENCES, \
+        &what_to_do).stream()
+
+namespace glog_internal_namespace_ {
+template <bool>
+struct CompileAssert {
+};
+struct CrashReason;
+
+// Returns true if FailureSignalHandler is installed.
+// Needs to be exported since it's used by the signalhandler_unittest.
+GOOGLE_GLOG_DLL_DECL bool IsFailureSignalHandlerInstalled();
+}  // namespace glog_internal_namespace_
+
+#define LOG_EVERY_N(severity, n)                                        \
+  SOME_KIND_OF_LOG_EVERY_N(severity, (n), google::LogMessage::SendToLog)
+
+#define SYSLOG_EVERY_N(severity, n) \
+  SOME_KIND_OF_LOG_EVERY_N(severity, (n), google::LogMessage::SendToSyslogAndLog)
+
+#define PLOG_EVERY_N(severity, n) \
+  SOME_KIND_OF_PLOG_EVERY_N(severity, (n), google::LogMessage::SendToLog)
+
+#define LOG_FIRST_N(severity, n) \
+  SOME_KIND_OF_LOG_FIRST_N(severity, (n), google::LogMessage::SendToLog)
+
+#define LOG_IF_EVERY_N(severity, condition, n) \
+  SOME_KIND_OF_LOG_IF_EVERY_N(severity, (condition), (n), google::LogMessage::SendToLog)
+
+// We want the special COUNTER value available for LOG_EVERY_X()'ed messages
+enum PRIVATE_Counter {COUNTER};
+
+#ifdef GLOG_NO_ABBREVIATED_SEVERITIES
+// wingdi.h defines ERROR to be 0. When we call LOG(ERROR), it gets
+// substituted with 0, and it expands to COMPACT_GOOGLE_LOG_0. To allow us
+// to keep using this syntax, we define this macro to do the same thing
+// as COMPACT_GOOGLE_LOG_ERROR.
+#define COMPACT_GOOGLE_LOG_0 COMPACT_GOOGLE_LOG_ERROR
+#define SYSLOG_0 SYSLOG_ERROR
+#define LOG_TO_STRING_0 LOG_TO_STRING_ERROR
+// Needed for LOG_IS_ON(ERROR).
+const LogSeverity GLOG_0 = GLOG_ERROR;
+#else
+// Users may include windows.h after logging.h without
+// GLOG_NO_ABBREVIATED_SEVERITIES nor WIN32_LEAN_AND_MEAN.
+// For this case, we cannot detect if ERROR is defined before users
+// actually use ERROR. Let's make an undefined symbol to warn users.
+# define GLOG_ERROR_MSG ERROR_macro_is_defined_Define_GLOG_NO_ABBREVIATED_SEVERITIES_before_including_logging_h_See_the_document_for_detail
+# define COMPACT_GOOGLE_LOG_0 GLOG_ERROR_MSG
+# define SYSLOG_0 GLOG_ERROR_MSG
+# define LOG_TO_STRING_0 GLOG_ERROR_MSG
+# define GLOG_0 GLOG_ERROR_MSG
+#endif
+
+// Plus some debug-logging macros that get compiled to nothing for production
+
+#if DCHECK_IS_ON()
+
+#define DLOG(severity) LOG(severity)
+#define DVLOG(verboselevel) VLOG(verboselevel)
+#define DLOG_IF(severity, condition) LOG_IF(severity, condition)
+#define DLOG_EVERY_N(severity, n) LOG_EVERY_N(severity, n)
+#define DLOG_IF_EVERY_N(severity, condition, n) \
+  LOG_IF_EVERY_N(severity, condition, n)
+#define DLOG_ASSERT(condition) LOG_ASSERT(condition)
+
+// debug-only checking.  executed if DCHECK_IS_ON().
+#define DCHECK(condition) CHECK(condition)
+#define DCHECK_EQ(val1, val2) CHECK_EQ(val1, val2)
+#define DCHECK_NE(val1, val2) CHECK_NE(val1, val2)
+#define DCHECK_LE(val1, val2) CHECK_LE(val1, val2)
+#define DCHECK_LT(val1, val2) CHECK_LT(val1, val2)
+#define DCHECK_GE(val1, val2) CHECK_GE(val1, val2)
+#define DCHECK_GT(val1, val2) CHECK_GT(val1, val2)
+#define DCHECK_NOTNULL(val) CHECK_NOTNULL(val)
+#define DCHECK_STREQ(str1, str2) CHECK_STREQ(str1, str2)
+#define DCHECK_STRCASEEQ(str1, str2) CHECK_STRCASEEQ(str1, str2)
+#define DCHECK_STRNE(str1, str2) CHECK_STRNE(str1, str2)
+#define DCHECK_STRCASENE(str1, str2) CHECK_STRCASENE(str1, str2)
+
+#else  // !DCHECK_IS_ON()
+
+#define DLOG(severity)  \
+  static_cast<void>(0), \
+  true ? (void) 0 : google::LogMessageVoidify() & LOG(severity)
+
+#define DVLOG(verboselevel)             \
+  static_cast<void>(0),                 \
+  (true || !VLOG_IS_ON(verboselevel)) ? \
+      (void) 0 : google::LogMessageVoidify() & LOG(INFO)
+
+#define DLOG_IF(severity, condition) \
+  static_cast<void>(0),              \
+  (true || !(condition)) ? (void) 0 : google::LogMessageVoidify() & LOG(severity)
+
+#define DLOG_EVERY_N(severity, n) \
+  static_cast<void>(0),           \
+  true ? (void) 0 : google::LogMessageVoidify() & LOG(severity)
+
+#define DLOG_IF_EVERY_N(severity, condition, n) \
+  static_cast<void>(0),                         \
+  (true || !(condition))? (void) 0 : google::LogMessageVoidify() & LOG(severity)
+
+#define DLOG_ASSERT(condition) \
+  static_cast<void>(0),        \
+  true ? (void) 0 : LOG_ASSERT(condition)
+
+// MSVC warning C4127: conditional expression is constant
+#define DCHECK(condition) \
+  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
+  while (false) \
+    GLOG_MSVC_POP_WARNING() CHECK(condition)
+
+#define DCHECK_EQ(val1, val2) \
+  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
+  while (false) \
+    GLOG_MSVC_POP_WARNING() CHECK_EQ(val1, val2)
+
+#define DCHECK_NE(val1, val2) \
+  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
+  while (false) \
+    GLOG_MSVC_POP_WARNING() CHECK_NE(val1, val2)
+
+#define DCHECK_LE(val1, val2) \
+  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
+  while (false) \
+    GLOG_MSVC_POP_WARNING() CHECK_LE(val1, val2)
+
+#define DCHECK_LT(val1, val2) \
+  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
+  while (false) \
+    GLOG_MSVC_POP_WARNING() CHECK_LT(val1, val2)
+
+#define DCHECK_GE(val1, val2) \
+  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
+  while (false) \
+    GLOG_MSVC_POP_WARNING() CHECK_GE(val1, val2)
+
+#define DCHECK_GT(val1, val2) \
+  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
+  while (false) \
+    GLOG_MSVC_POP_WARNING() CHECK_GT(val1, val2)
+
+// You may see warnings in release mode if you don't use the return
+// value of DCHECK_NOTNULL. Please just use DCHECK for such cases.
+#define DCHECK_NOTNULL(val) (val)
+
+#define DCHECK_STREQ(str1, str2) \
+  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
+  while (false) \
+    GLOG_MSVC_POP_WARNING() CHECK_STREQ(str1, str2)
+
+#define DCHECK_STRCASEEQ(str1, str2) \
+  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
+  while (false) \
+    GLOG_MSVC_POP_WARNING() CHECK_STRCASEEQ(str1, str2)
+
+#define DCHECK_STRNE(str1, str2) \
+  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
+  while (false) \
+    GLOG_MSVC_POP_WARNING() CHECK_STRNE(str1, str2)
+
+#define DCHECK_STRCASENE(str1, str2) \
+  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
+  while (false) \
+    GLOG_MSVC_POP_WARNING() CHECK_STRCASENE(str1, str2)
+
+#endif  // DCHECK_IS_ON()
+
+// Log only in verbose mode.
+
+#define VLOG(verboselevel) LOG_IF(INFO, VLOG_IS_ON(verboselevel))
+
+#define VLOG_IF(verboselevel, condition) \
+  LOG_IF(INFO, (condition) && VLOG_IS_ON(verboselevel))
+
+#define VLOG_EVERY_N(verboselevel, n) \
+  LOG_IF_EVERY_N(INFO, VLOG_IS_ON(verboselevel), n)
+
+#define VLOG_IF_EVERY_N(verboselevel, condition, n) \
+  LOG_IF_EVERY_N(INFO, (condition) && VLOG_IS_ON(verboselevel), n)
+
+namespace base_logging {
+
+// LogMessage::LogStream is a std::ostream backed by this streambuf.
+// This class ignores overflow and leaves two bytes at the end of the
+// buffer to allow for a '\n' and '\0'.
+class GOOGLE_GLOG_DLL_DECL LogStreamBuf : public std::streambuf {
+ public:
+  // REQUIREMENTS: "len" must be >= 2 to account for the '\n' and '\0'.
+  LogStreamBuf(char *buf, int len) {
+    setp(buf, buf + len - 2);
+  }
+
+  // This effectively ignores overflow.
+  virtual int_type overflow(int_type ch) {
+    return ch;
+  }
+
+  // Legacy public ostrstream method.
+  size_t pcount() const { return pptr() - pbase(); }
+  char* pbase() const { return std::streambuf::pbase(); }
+};
+
+}  // namespace base_logging
+
+//
+// This class more or less represents a particular log message.  You
+// create an instance of LogMessage and then stream stuff to it.
+// When you finish streaming to it, ~LogMessage is called and the
+// full message gets streamed to the appropriate destination.
+//
+// You shouldn't actually use LogMessage's constructor to log things,
+// though.  You should use the LOG() macro (and variants thereof)
+// above.
+class GOOGLE_GLOG_DLL_DECL LogMessage {
+public:
+  enum {
+    // Passing kNoLogPrefix for the line number disables the
+    // log-message prefix. Useful for using the LogMessage
+    // infrastructure as a printing utility. See also the --log_prefix
+    // flag for controlling the log-message prefix on an
+    // application-wide basis.
+    kNoLogPrefix = -1
+  };
+
+  // LogStream inherit from non-DLL-exported class (std::ostrstream)
+  // and VC++ produces a warning for this situation.
+  // However, MSDN says "C4275 can be ignored in Microsoft Visual C++
+  // 2005 if you are deriving from a type in the Standard C++ Library"
+  // http://msdn.microsoft.com/en-us/library/3tdb471s(VS.80).aspx
+  // Let's just ignore the warning.
+GLOG_MSVC_PUSH_DISABLE_WARNING(4275)
+  class GOOGLE_GLOG_DLL_DECL LogStream : public std::ostream {
+GLOG_MSVC_POP_WARNING()
+  public:
+    LogStream(char *buf, int len, int ctr)
+        : std::ostream(NULL),
+          streambuf_(buf, len),
+          ctr_(ctr),
+          self_(this) {
+      rdbuf(&streambuf_);
+    }
+
+    int ctr() const { return ctr_; }
+    void set_ctr(int ctr) { ctr_ = ctr; }
+    LogStream* self() const { return self_; }
+
+    // Legacy std::streambuf methods.
+    size_t pcount() const { return streambuf_.pcount(); }
+    char* pbase() const { return streambuf_.pbase(); }
+    char* str() const { return pbase(); }
+
+  private:
+    LogStream(const LogStream&);
+    LogStream& operator=(const LogStream&);
+    base_logging::LogStreamBuf streambuf_;
+    int ctr_;  // Counter hack (for the LOG_EVERY_X() macro)
+    LogStream *self_;  // Consistency check hack
+  };
+
+public:
+  // icc 8 requires this typedef to avoid an internal compiler error.
+  typedef void (LogMessage::*SendMethod)();
+
+  LogMessage(const char* file, int line, LogSeverity severity, int ctr,
+             SendMethod send_method);
+
+  // Two special constructors that generate reduced amounts of code at
+  // LOG call sites for common cases.
+
+  // Used for LOG(INFO): Implied are:
+  // severity = INFO, ctr = 0, send_method = &LogMessage::SendToLog.
+  //
+  // Using this constructor instead of the more complex constructor above
+  // saves 19 bytes per call site.
+  LogMessage(const char* file, int line);
+
+  // Used for LOG(severity) where severity != INFO.  Implied
+  // are: ctr = 0, send_method = &LogMessage::SendToLog
+  //
+  // Using this constructor instead of the more complex constructor above
+  // saves 17 bytes per call site.
+  LogMessage(const char* file, int line, LogSeverity severity);
+
+  // Constructor to log this message to a specified sink (if not NULL).
+  // Implied are: ctr = 0, send_method = &LogMessage::SendToSinkAndLog if
+  // also_send_to_log is true, send_method = &LogMessage::SendToSink otherwise.
+  LogMessage(const char* file, int line, LogSeverity severity, LogSink* sink,
+             bool also_send_to_log);
+
+  // Constructor where we also give a vector<string> pointer
+  // for storing the messages (if the pointer is not NULL).
+  // Implied are: ctr = 0, send_method = &LogMessage::SaveOrSendToLog.
+  LogMessage(const char* file, int line, LogSeverity severity,
+             std::vector<std::string>* outvec);
+
+  // Constructor where we also give a string pointer for storing the
+  // message (if the pointer is not NULL).  Implied are: ctr = 0,
+  // send_method = &LogMessage::WriteToStringAndLog.
+  LogMessage(const char* file, int line, LogSeverity severity,
+             std::string* message);
+
+  // A special constructor used for check failures
+  LogMessage(const char* file, int line, const CheckOpString& result);
+
+  ~LogMessage();
+
+  // Flush a buffered message to the sink set in the constructor.  Always
+  // called by the destructor, it may also be called from elsewhere if
+  // needed.  Only the first call is actioned; any later ones are ignored.
+  void Flush();
+
+  // An arbitrary limit on the length of a single log message.  This
+  // is so that streaming can be done more efficiently.
+  static const size_t kMaxLogMessageLen;
+
+  // Theses should not be called directly outside of logging.*,
+  // only passed as SendMethod arguments to other LogMessage methods:
+  void SendToLog();  // Actually dispatch to the logs
+  void SendToSyslogAndLog();  // Actually dispatch to syslog and the logs
+
+  // Call abort() or similar to perform LOG(FATAL) crash.
+  static void __attribute__((noreturn)) Fail();
+
+  std::ostream& stream();
+
+  int preserved_errno() const;
+
+  // Must be called without the log_mutex held.  (L < log_mutex)
+  static int64 num_messages(int severity);
+
+  struct LogMessageData;
+
+private:
+  // Fully internal SendMethod cases:
+  void SendToSinkAndLog();  // Send to sink if provided and dispatch to the logs
+  void SendToSink();  // Send to sink if provided, do nothing otherwise.
+
+  // Write to string if provided and dispatch to the logs.
+  void WriteToStringAndLog();
+
+  void SaveOrSendToLog();  // Save to stringvec if provided, else to logs
+
+  void Init(const char* file, int line, LogSeverity severity,
+            void (LogMessage::*send_method)());
+
+  // Used to fill in crash information during LOG(FATAL) failures.
+  void RecordCrashReason(glog_internal_namespace_::CrashReason* reason);
+
+  // Counts of messages sent at each priority:
+  static int64 num_messages_[NUM_SEVERITIES];  // under log_mutex
+
+  // We keep the data in a separate struct so that each instance of
+  // LogMessage uses less stack space.
+  LogMessageData* allocated_;
+  LogMessageData* data_;
+
+  friend class LogDestination;
+
+  LogMessage(const LogMessage&);
+  void operator=(const LogMessage&);
+};
+
+// This class happens to be thread-hostile because all instances share
+// a single data buffer, but since it can only be created just before
+// the process dies, we don't worry so much.
+class GOOGLE_GLOG_DLL_DECL LogMessageFatal : public LogMessage {
+ public:
+  LogMessageFatal(const char* file, int line);
+  LogMessageFatal(const char* file, int line, const CheckOpString& result);
+  __attribute__((noreturn)) ~LogMessageFatal();
+};
+
+// A non-macro interface to the log facility; (useful
+// when the logging level is not a compile-time constant).
+inline void LogAtLevel(int const severity, std::string const &msg) {
+  LogMessage(__FILE__, __LINE__, severity).stream() << msg;
+}
+
+// A macro alternative of LogAtLevel. New code may want to use this
+// version since there are two advantages: 1. this version outputs the
+// file name and the line number where this macro is put like other
+// LOG macros, 2. this macro can be used as C++ stream.
+#define LOG_AT_LEVEL(severity) google::LogMessage(__FILE__, __LINE__, severity).stream()
+
+// Check if it's compiled in C++11 mode.
+//
+// GXX_EXPERIMENTAL_CXX0X is defined by gcc and clang up to at least
+// gcc-4.7 and clang-3.1 (2011-12-13).  __cplusplus was defined to 1
+// in gcc before 4.7 (Crosstool 16) and clang before 3.1, but is
+// defined according to the language version in effect thereafter.
+// Microsoft Visual Studio 14 (2015) sets __cplusplus==199711 despite
+// reasonably good C++11 support, so we set LANG_CXX for it and
+// newer versions (_MSC_VER >= 1900).
+#if (defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103L || \
+     (defined(_MSC_VER) && _MSC_VER >= 1900))
+// Helper for CHECK_NOTNULL().
+//
+// In C++11, all cases can be handled by a single function. Since the value
+// category of the argument is preserved (also for rvalue references),
+// member initializer lists like the one below will compile correctly:
+//
+//   Foo()
+//     : x_(CHECK_NOTNULL(MethodReturningUniquePtr())) {}
+template <typename T>
+T CheckNotNull(const char* file, int line, const char* names, T&& t) {
+ if (t == nullptr) {
+   LogMessageFatal(file, line, new std::string(names));
+ }
+ return std::forward<T>(t);
+}
+
+#else
+
+// A small helper for CHECK_NOTNULL().
+template <typename T>
+T* CheckNotNull(const char *file, int line, const char *names, T* t) {
+  if (t == NULL) {
+    LogMessageFatal(file, line, new std::string(names));
+  }
+  return t;
+}
+#endif
+
+// Allow folks to put a counter in the LOG_EVERY_X()'ed messages. This
+// only works if ostream is a LogStream. If the ostream is not a
+// LogStream you'll get an assert saying as much at runtime.
+GOOGLE_GLOG_DLL_DECL std::ostream& operator<<(std::ostream &os,
+                                              const PRIVATE_Counter&);
+
+
+// Derived class for PLOG*() above.
+class GOOGLE_GLOG_DLL_DECL ErrnoLogMessage : public LogMessage {
+ public:
+
+  ErrnoLogMessage(const char* file, int line, LogSeverity severity, int ctr,
+                  void (LogMessage::*send_method)());
+
+  // Postpends ": strerror(errno) [errno]".
+  ~ErrnoLogMessage();
+
+ private:
+  ErrnoLogMessage(const ErrnoLogMessage&);
+  void operator=(const ErrnoLogMessage&);
+};
+
+
+// This class is used to explicitly ignore values in the conditional
+// logging macros.  This avoids compiler warnings like "value computed
+// is not used" and "statement has no effect".
+
+class GOOGLE_GLOG_DLL_DECL LogMessageVoidify {
+ public:
+  LogMessageVoidify() { }
+  // This has to be an operator with a precedence lower than << but
+  // higher than ?:
+  void operator&(std::ostream&) { }
+};
+
+
+// Flushes all log files that contains messages that are at least of
+// the specified severity level.  Thread-safe.
+GOOGLE_GLOG_DLL_DECL void FlushLogFiles(LogSeverity min_severity);
+
+// Flushes all log files that contains messages that are at least of
+// the specified severity level. Thread-hostile because it ignores
+// locking -- used for catastrophic failures.
+GOOGLE_GLOG_DLL_DECL void FlushLogFilesUnsafe(LogSeverity min_severity);
+
+//
+// Set the destination to which a particular severity level of log
+// messages is sent.  If base_filename is "", it means "don't log this
+// severity".  Thread-safe.
+//
+GOOGLE_GLOG_DLL_DECL void SetLogDestination(LogSeverity severity,
+                                            const char* base_filename);
+
+//
+// Set the basename of the symlink to the latest log file at a given
+// severity.  If symlink_basename is empty, do not make a symlink.  If
+// you don't call this function, the symlink basename is the
+// invocation name of the program.  Thread-safe.
+//
+GOOGLE_GLOG_DLL_DECL void SetLogSymlink(LogSeverity severity,
+                                        const char* symlink_basename);
+
+//
+// Used to send logs to some other kind of destination
+// Users should subclass LogSink and override send to do whatever they want.
+// Implementations must be thread-safe because a shared instance will
+// be called from whichever thread ran the LOG(XXX) line.
+class GOOGLE_GLOG_DLL_DECL LogSink {
+ public:
+  virtual ~LogSink();
+
+  // Sink's logging logic (message_len is such as to exclude '\n' at the end).
+  // This method can't use LOG() or CHECK() as logging system mutex(s) are held
+  // during this call.
+  virtual void send(LogSeverity severity, const char* full_filename,
+                    const char* base_filename, int line,
+                    const struct ::tm* tm_time,
+                    const char* message, size_t message_len) = 0;
+
+  // Redefine this to implement waiting for
+  // the sink's logging logic to complete.
+  // It will be called after each send() returns,
+  // but before that LogMessage exits or crashes.
+  // By default this function does nothing.
+  // Using this function one can implement complex logic for send()
+  // that itself involves logging; and do all this w/o causing deadlocks and
+  // inconsistent rearrangement of log messages.
+  // E.g. if a LogSink has thread-specific actions, the send() method
+  // can simply add the message to a queue and wake up another thread that
+  // handles real logging while itself making some LOG() calls;
+  // WaitTillSent() can be implemented to wait for that logic to complete.
+  // See our unittest for an example.
+  virtual void WaitTillSent();
+
+  // Returns the normal text output of the log message.
+  // Can be useful to implement send().
+  static std::string ToString(LogSeverity severity, const char* file, int line,
+                              const struct ::tm* tm_time,
+                              const char* message, size_t message_len);
+};
+
+// Add or remove a LogSink as a consumer of logging data.  Thread-safe.
+GOOGLE_GLOG_DLL_DECL void AddLogSink(LogSink *destination);
+GOOGLE_GLOG_DLL_DECL void RemoveLogSink(LogSink *destination);
+
+//
+// Specify an "extension" added to the filename specified via
+// SetLogDestination.  This applies to all severity levels.  It's
+// often used to append the port we're listening on to the logfile
+// name.  Thread-safe.
+//
+GOOGLE_GLOG_DLL_DECL void SetLogFilenameExtension(
+    const char* filename_extension);
+
+//
+// Make it so that all log messages of at least a particular severity
+// are logged to stderr (in addition to logging to the usual log
+// file(s)).  Thread-safe.
+//
+GOOGLE_GLOG_DLL_DECL void SetStderrLogging(LogSeverity min_severity);
+
+//
+// Make it so that all log messages go only to stderr.  Thread-safe.
+//
+GOOGLE_GLOG_DLL_DECL void LogToStderr();
+
+//
+// Make it so that all log messages of at least a particular severity are
+// logged via email to a list of addresses (in addition to logging to the
+// usual log file(s)).  The list of addresses is just a string containing
+// the email addresses to send to (separated by spaces, say).  Thread-safe.
+//
+GOOGLE_GLOG_DLL_DECL void SetEmailLogging(LogSeverity min_severity,
+                                          const char* addresses);
+
+// A simple function that sends email. dest is a commma-separated
+// list of addressess.  Thread-safe.
+GOOGLE_GLOG_DLL_DECL bool SendEmail(const char *dest,
+                                    const char *subject, const char *body);
+
+GOOGLE_GLOG_DLL_DECL const std::vector<std::string>& GetLoggingDirectories();
+
+// For tests only:  Clear the internal [cached] list of logging directories to
+// force a refresh the next time GetLoggingDirectories is called.
+// Thread-hostile.
+void TestOnly_ClearLoggingDirectoriesList();
+
+// Returns a set of existing temporary directories, which will be a
+// subset of the directories returned by GetLogginDirectories().
+// Thread-safe.
+GOOGLE_GLOG_DLL_DECL void GetExistingTempDirectories(
+    std::vector<std::string>* list);
+
+// Print any fatal message again -- useful to call from signal handler
+// so that the last thing in the output is the fatal message.
+// Thread-hostile, but a race is unlikely.
+GOOGLE_GLOG_DLL_DECL void ReprintFatalMessage();
+
+// Truncate a log file that may be the append-only output of multiple
+// processes and hence can't simply be renamed/reopened (typically a
+// stdout/stderr).  If the file "path" is > "limit" bytes, copy the
+// last "keep" bytes to offset 0 and truncate the rest. Since we could
+// be racing with other writers, this approach has the potential to
+// lose very small amounts of data. For security, only follow symlinks
+// if the path is /proc/self/fd/*
+GOOGLE_GLOG_DLL_DECL void TruncateLogFile(const char *path,
+                                          int64 limit, int64 keep);
+
+// Truncate stdout and stderr if they are over the value specified by
+// --max_log_size; keep the final 1MB.  This function has the same
+// race condition as TruncateLogFile.
+GOOGLE_GLOG_DLL_DECL void TruncateStdoutStderr();
+
+// Return the string representation of the provided LogSeverity level.
+// Thread-safe.
+GOOGLE_GLOG_DLL_DECL const char* GetLogSeverityName(LogSeverity severity);
+
+// ---------------------------------------------------------------------
+// Implementation details that are not useful to most clients
+// ---------------------------------------------------------------------
+
+// A Logger is the interface used by logging modules to emit entries
+// to a log.  A typical implementation will dump formatted data to a
+// sequence of files.  We also provide interfaces that will forward
+// the data to another thread so that the invoker never blocks.
+// Implementations should be thread-safe since the logging system
+// will write to them from multiple threads.
+
+namespace base {
+
+class GOOGLE_GLOG_DLL_DECL Logger {
+ public:
+  virtual ~Logger();
+
+  // Writes "message[0,message_len-1]" corresponding to an event that
+  // occurred at "timestamp".  If "force_flush" is true, the log file
+  // is flushed immediately.
+  //
+  // The input message has already been formatted as deemed
+  // appropriate by the higher level logging facility.  For example,
+  // textual log messages already contain timestamps, and the
+  // file:linenumber header.
+  virtual void Write(bool force_flush,
+                     time_t timestamp,
+                     const char* message,
+                     int message_len) = 0;
+
+  // Flush any buffered messages
+  virtual void Flush() = 0;
+
+  // Get the current LOG file size.
+  // The returned value is approximate since some
+  // logged data may not have been flushed to disk yet.
+  virtual uint32 LogSize() = 0;
+};
+
+// Get the logger for the specified severity level.  The logger
+// remains the property of the logging module and should not be
+// deleted by the caller.  Thread-safe.
+extern GOOGLE_GLOG_DLL_DECL Logger* GetLogger(LogSeverity level);
+
+// Set the logger for the specified severity level.  The logger
+// becomes the property of the logging module and should not
+// be deleted by the caller.  Thread-safe.
+extern GOOGLE_GLOG_DLL_DECL void SetLogger(LogSeverity level, Logger* logger);
+
+}
+
+// glibc has traditionally implemented two incompatible versions of
+// strerror_r(). There is a poorly defined convention for picking the
+// version that we want, but it is not clear whether it even works with
+// all versions of glibc.
+// So, instead, we provide this wrapper that automatically detects the
+// version that is in use, and then implements POSIX semantics.
+// N.B. In addition to what POSIX says, we also guarantee that "buf" will
+// be set to an empty string, if this function failed. This means, in most
+// cases, you do not need to check the error code and you can directly
+// use the value of "buf". It will never have an undefined value.
+// DEPRECATED: Use StrError(int) instead.
+GOOGLE_GLOG_DLL_DECL int posix_strerror_r(int err, char *buf, size_t len);
+
+// A thread-safe replacement for strerror(). Returns a string describing the
+// given POSIX error code.
+GOOGLE_GLOG_DLL_DECL std::string StrError(int err);
+
+// A class for which we define operator<<, which does nothing.
+class GOOGLE_GLOG_DLL_DECL NullStream : public LogMessage::LogStream {
+ public:
+  // Initialize the LogStream so the messages can be written somewhere
+  // (they'll never be actually displayed). This will be needed if a
+  // NullStream& is implicitly converted to LogStream&, in which case
+  // the overloaded NullStream::operator<< will not be invoked.
+  NullStream() : LogMessage::LogStream(message_buffer_, 1, 0) { }
+  NullStream(const char* /*file*/, int /*line*/,
+             const CheckOpString& /*result*/) :
+      LogMessage::LogStream(message_buffer_, 1, 0) { }
+  NullStream &stream() { return *this; }
+ private:
+  // A very short buffer for messages (which we discard anyway). This
+  // will be needed if NullStream& converted to LogStream& (e.g. as a
+  // result of a conditional expression).
+  char message_buffer_[2];
+};
+
+// Do nothing. This operator is inline, allowing the message to be
+// compiled away. The message will not be compiled away if we do
+// something like (flag ? LOG(INFO) : LOG(ERROR)) << message; when
+// SKIP_LOG=WARNING. In those cases, NullStream will be implicitly
+// converted to LogStream and the message will be computed and then
+// quietly discarded.
+template<class T>
+inline NullStream& operator<<(NullStream &str, const T &) { return str; }
+
+// Similar to NullStream, but aborts the program (without stack
+// trace), like LogMessageFatal.
+class GOOGLE_GLOG_DLL_DECL NullStreamFatal : public NullStream {
+ public:
+  NullStreamFatal() { }
+  NullStreamFatal(const char* file, int line, const CheckOpString& result) :
+      NullStream(file, line, result) { }
+  __attribute__((noreturn)) ~NullStreamFatal() throw () { _exit(1); }
+};
+
+// Install a signal handler that will dump signal information and a stack
+// trace when the program crashes on certain signals.  We'll install the
+// signal handler for the following signals.
+//
+// SIGSEGV, SIGILL, SIGFPE, SIGABRT, SIGBUS, and SIGTERM.
+//
+// By default, the signal handler will write the failure dump to the
+// standard error.  You can customize the destination by installing your
+// own writer function by InstallFailureWriter() below.
+//
+// Note on threading:
+//
+// The function should be called before threads are created, if you want
+// to use the failure signal handler for all threads.  The stack trace
+// will be shown only for the thread that receives the signal.  In other
+// words, stack traces of other threads won't be shown.
+GOOGLE_GLOG_DLL_DECL void InstallFailureSignalHandler();
+
+// Installs a function that is used for writing the failure dump.  "data"
+// is the pointer to the beginning of a message to be written, and "size"
+// is the size of the message.  You should not expect the data is
+// terminated with '\0'.
+GOOGLE_GLOG_DLL_DECL void InstallFailureWriter(
+    void (*writer)(const char* data, int size));
+
+}
+
+#endif // _LOGGING_H_
diff --git a/projects/llm_framework/include/glog/raw_logging.h b/projects/llm_framework/include/glog/raw_logging.h
new file mode 100644
index 00000000..cf3f27d9
--- /dev/null
+++ b/projects/llm_framework/include/glog/raw_logging.h
@@ -0,0 +1,180 @@
+// Copyright (c) 2006, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: Maxim Lifantsev
+//
+// Thread-safe logging routines that do not allocate any memory or
+// acquire any locks, and can therefore be used by low-level memory
+// allocation and synchronization code.
+
+#ifndef BASE_RAW_LOGGING_H_
+#define BASE_RAW_LOGGING_H_
+
+#include <time.h>
+
+namespace google {
+
+#include "glog/log_severity.h"
+#include "glog/vlog_is_on.h"
+
+// Annoying stuff for windows -- makes sure clients can import these functions
+#ifndef GOOGLE_GLOG_DLL_DECL
+# if defined(_WIN32) && !defined(__CYGWIN__)
+#   define GOOGLE_GLOG_DLL_DECL  __declspec(dllimport)
+# else
+#   define GOOGLE_GLOG_DLL_DECL
+# endif
+#endif
+
+// This is similar to LOG(severity) << format... and VLOG(level) << format..,
+// but
+// * it is to be used ONLY by low-level modules that can't use normal LOG()
+// * it is desiged to be a low-level logger that does not allocate any
+//   memory and does not need any locks, hence:
+// * it logs straight and ONLY to STDERR w/o buffering
+// * it uses an explicit format and arguments list
+// * it will silently chop off really long message strings
+// Usage example:
+//   RAW_LOG(ERROR, "Failed foo with %i: %s", status, error);
+//   RAW_VLOG(3, "status is %i", status);
+// These will print an almost standard log lines like this to stderr only:
+//   E0821 211317 file.cc:123] RAW: Failed foo with 22: bad_file
+//   I0821 211317 file.cc:142] RAW: status is 20
+#define RAW_LOG(severity, ...) \
+  do { \
+    switch (google::GLOG_ ## severity) {  \
+      case 0: \
+        RAW_LOG_INFO(__VA_ARGS__); \
+        break; \
+      case 1: \
+        RAW_LOG_WARNING(__VA_ARGS__); \
+        break; \
+      case 2: \
+        RAW_LOG_ERROR(__VA_ARGS__); \
+        break; \
+      case 3: \
+        RAW_LOG_FATAL(__VA_ARGS__); \
+        break; \
+      default: \
+        break; \
+    } \
+  } while (0)
+
+// The following STRIP_LOG testing is performed in the header file so that it's
+// possible to completely compile out the logging code and the log messages.
+#if STRIP_LOG == 0
+#define RAW_VLOG(verboselevel, ...) \
+  do { \
+    if (VLOG_IS_ON(verboselevel)) { \
+      RAW_LOG_INFO(__VA_ARGS__); \
+    } \
+  } while (0)
+#else
+#define RAW_VLOG(verboselevel, ...) RawLogStub__(0, __VA_ARGS__)
+#endif // STRIP_LOG == 0
+
+#if STRIP_LOG == 0
+#define RAW_LOG_INFO(...) google::RawLog__(google::GLOG_INFO, \
+                                   __FILE__, __LINE__, __VA_ARGS__)
+#else
+#define RAW_LOG_INFO(...) google::RawLogStub__(0, __VA_ARGS__)
+#endif // STRIP_LOG == 0
+
+#if STRIP_LOG <= 1
+#define RAW_LOG_WARNING(...) google::RawLog__(google::GLOG_WARNING,   \
+                                      __FILE__, __LINE__, __VA_ARGS__)
+#else
+#define RAW_LOG_WARNING(...) google::RawLogStub__(0, __VA_ARGS__)
+#endif // STRIP_LOG <= 1
+
+#if STRIP_LOG <= 2
+#define RAW_LOG_ERROR(...) google::RawLog__(google::GLOG_ERROR,       \
+                                    __FILE__, __LINE__, __VA_ARGS__)
+#else
+#define RAW_LOG_ERROR(...) google::RawLogStub__(0, __VA_ARGS__)
+#endif // STRIP_LOG <= 2
+
+#if STRIP_LOG <= 3
+#define RAW_LOG_FATAL(...) google::RawLog__(google::GLOG_FATAL,       \
+                                    __FILE__, __LINE__, __VA_ARGS__)
+#else
+#define RAW_LOG_FATAL(...) \
+  do { \
+    google::RawLogStub__(0, __VA_ARGS__);        \
+    exit(1); \
+  } while (0)
+#endif // STRIP_LOG <= 3
+
+// Similar to CHECK(condition) << message,
+// but for low-level modules: we use only RAW_LOG that does not allocate memory.
+// We do not want to provide args list here to encourage this usage:
+//   if (!cond)  RAW_LOG(FATAL, "foo ...", hard_to_compute_args);
+// so that the args are not computed when not needed.
+#define RAW_CHECK(condition, message)                                   \
+  do {                                                                  \
+    if (!(condition)) {                                                 \
+      RAW_LOG(FATAL, "Check %s failed: %s", #condition, message);       \
+    }                                                                   \
+  } while (0)
+
+// Debug versions of RAW_LOG and RAW_CHECK
+#ifndef NDEBUG
+
+#define RAW_DLOG(severity, ...) RAW_LOG(severity, __VA_ARGS__)
+#define RAW_DCHECK(condition, message) RAW_CHECK(condition, message)
+
+#else  // NDEBUG
+
+#define RAW_DLOG(severity, ...)                                 \
+  while (false)                                                 \
+    RAW_LOG(severity, __VA_ARGS__)
+#define RAW_DCHECK(condition, message) \
+  while (false) \
+    RAW_CHECK(condition, message)
+
+#endif  // NDEBUG
+
+// Stub log function used to work around for unused variable warnings when
+// building with STRIP_LOG > 0.
+static inline void RawLogStub__(int /* ignored */, ...) {
+}
+
+// Helper function to implement RAW_LOG and RAW_VLOG
+// Logs format... at "severity" level, reporting it
+// as called from file:line.
+// This does not allocate memory or acquire locks.
+GOOGLE_GLOG_DLL_DECL void RawLog__(LogSeverity severity,
+                                   const char* file,
+                                   int line,
+                                   const char* format, ...)
+   ;
+
+}
+
+#endif  // BASE_RAW_LOGGING_H_
diff --git a/projects/llm_framework/include/glog/stl_logging.h b/projects/llm_framework/include/glog/stl_logging.h
new file mode 100644
index 00000000..40a15aa4
--- /dev/null
+++ b/projects/llm_framework/include/glog/stl_logging.h
@@ -0,0 +1,220 @@
+// Copyright (c) 2003, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Stream output operators for STL containers; to be used for logging *only*.
+// Inclusion of this file lets you do:
+//
+// list<string> x;
+// LOG(INFO) << "data: " << x;
+// vector<int> v1, v2;
+// CHECK_EQ(v1, v2);
+//
+// If you want to use this header file with hash maps or slist, you
+// need to define macros before including this file:
+//
+// - GLOG_STL_LOGGING_FOR_UNORDERED     - <unordered_map> and <unordered_set>
+// - GLOG_STL_LOGGING_FOR_TR1_UNORDERED - <tr1/unordered_(map|set)>
+// - GLOG_STL_LOGGING_FOR_EXT_HASH      - <ext/hash_(map|set)>
+// - GLOG_STL_LOGGING_FOR_EXT_SLIST     - <ext/slist>
+//
+
+#ifndef UTIL_GTL_STL_LOGGING_INL_H_
+#define UTIL_GTL_STL_LOGGING_INL_H_
+
+#if !1
+# error We do not support stl_logging for this compiler
+#endif
+
+#include <deque>
+#include <list>
+#include <map>
+#include <ostream>
+#include <set>
+#include <utility>
+#include <vector>
+
+#ifdef GLOG_STL_LOGGING_FOR_UNORDERED
+# include <unordered_map>
+# include <unordered_set>
+#endif
+
+#ifdef GLOG_STL_LOGGING_FOR_TR1_UNORDERED
+# include <tr1/unordered_map>
+# include <tr1/unordered_set>
+#endif
+
+#ifdef GLOG_STL_LOGGING_FOR_EXT_HASH
+# include <ext/hash_set>
+# include <ext/hash_map>
+#endif
+#ifdef GLOG_STL_LOGGING_FOR_EXT_SLIST
+# include <ext/slist>
+#endif
+
+// Forward declare these two, and define them after all the container streams
+// operators so that we can recurse from pair -> container -> container -> pair
+// properly.
+template<class First, class Second>
+std::ostream& operator<<(std::ostream& out, const std::pair<First, Second>& p);
+
+namespace google {
+
+template<class Iter>
+void PrintSequence(std::ostream& out, Iter begin, Iter end);
+
+}
+
+#define OUTPUT_TWO_ARG_CONTAINER(Sequence) \
+template<class T1, class T2> \
+inline std::ostream& operator<<(std::ostream& out, \
+                                const Sequence<T1, T2>& seq) { \
+  google::PrintSequence(out, seq.begin(), seq.end()); \
+  return out; \
+}
+
+OUTPUT_TWO_ARG_CONTAINER(std::vector)
+OUTPUT_TWO_ARG_CONTAINER(std::deque)
+OUTPUT_TWO_ARG_CONTAINER(std::list)
+#ifdef GLOG_STL_LOGGING_FOR_EXT_SLIST
+OUTPUT_TWO_ARG_CONTAINER(__gnu_cxx::slist)
+#endif
+
+#undef OUTPUT_TWO_ARG_CONTAINER
+
+#define OUTPUT_THREE_ARG_CONTAINER(Sequence) \
+template<class T1, class T2, class T3> \
+inline std::ostream& operator<<(std::ostream& out, \
+                                const Sequence<T1, T2, T3>& seq) { \
+  google::PrintSequence(out, seq.begin(), seq.end()); \
+  return out; \
+}
+
+OUTPUT_THREE_ARG_CONTAINER(std::set)
+OUTPUT_THREE_ARG_CONTAINER(std::multiset)
+
+#undef OUTPUT_THREE_ARG_CONTAINER
+
+#define OUTPUT_FOUR_ARG_CONTAINER(Sequence) \
+template<class T1, class T2, class T3, class T4> \
+inline std::ostream& operator<<(std::ostream& out, \
+                                const Sequence<T1, T2, T3, T4>& seq) { \
+  google::PrintSequence(out, seq.begin(), seq.end()); \
+  return out; \
+}
+
+OUTPUT_FOUR_ARG_CONTAINER(std::map)
+OUTPUT_FOUR_ARG_CONTAINER(std::multimap)
+#ifdef GLOG_STL_LOGGING_FOR_UNORDERED
+OUTPUT_FOUR_ARG_CONTAINER(std::unordered_set)
+OUTPUT_FOUR_ARG_CONTAINER(std::unordered_multiset)
+#endif
+#ifdef GLOG_STL_LOGGING_FOR_TR1_UNORDERED
+OUTPUT_FOUR_ARG_CONTAINER(std::tr1::unordered_set)
+OUTPUT_FOUR_ARG_CONTAINER(std::tr1::unordered_multiset)
+#endif
+#ifdef GLOG_STL_LOGGING_FOR_EXT_HASH
+OUTPUT_FOUR_ARG_CONTAINER(__gnu_cxx::hash_set)
+OUTPUT_FOUR_ARG_CONTAINER(__gnu_cxx::hash_multiset)
+#endif
+
+#undef OUTPUT_FOUR_ARG_CONTAINER
+
+#define OUTPUT_FIVE_ARG_CONTAINER(Sequence) \
+template<class T1, class T2, class T3, class T4, class T5> \
+inline std::ostream& operator<<(std::ostream& out, \
+                                const Sequence<T1, T2, T3, T4, T5>& seq) { \
+  google::PrintSequence(out, seq.begin(), seq.end()); \
+  return out; \
+}
+
+#ifdef GLOG_STL_LOGGING_FOR_UNORDERED
+OUTPUT_FIVE_ARG_CONTAINER(std::unordered_map)
+OUTPUT_FIVE_ARG_CONTAINER(std::unordered_multimap)
+#endif
+#ifdef GLOG_STL_LOGGING_FOR_TR1_UNORDERED
+OUTPUT_FIVE_ARG_CONTAINER(std::tr1::unordered_map)
+OUTPUT_FIVE_ARG_CONTAINER(std::tr1::unordered_multimap)
+#endif
+#ifdef GLOG_STL_LOGGING_FOR_EXT_HASH
+OUTPUT_FIVE_ARG_CONTAINER(__gnu_cxx::hash_map)
+OUTPUT_FIVE_ARG_CONTAINER(__gnu_cxx::hash_multimap)
+#endif
+
+#undef OUTPUT_FIVE_ARG_CONTAINER
+
+template<class First, class Second>
+inline std::ostream& operator<<(std::ostream& out,
+                                const std::pair<First, Second>& p) {
+  out << '(' << p.first << ", " << p.second << ')';
+  return out;
+}
+
+namespace google {
+
+template<class Iter>
+inline void PrintSequence(std::ostream& out, Iter begin, Iter end) {
+  // Output at most 100 elements -- appropriate if used for logging.
+  for (int i = 0; begin != end && i < 100; ++i, ++begin) {
+    if (i > 0) out << ' ';
+    out << *begin;
+  }
+  if (begin != end) {
+    out << " ...";
+  }
+}
+
+}
+
+// Note that this is technically undefined behavior! We are adding things into
+// the std namespace for a reason though -- we are providing new operations on
+// types which are themselves defined with this namespace. Without this, these
+// operator overloads cannot be found via ADL. If these definitions are not
+// found via ADL, they must be #included before they're used, which requires
+// this header to be included before apparently independent other headers.
+//
+// For example, base/logging.h defines various template functions to implement
+// CHECK_EQ(x, y) and stream x and y into the log in the event the check fails.
+// It does so via the function template MakeCheckOpValueString:
+//   template<class T>
+//   void MakeCheckOpValueString(strstream* ss, const T& v) {
+//     (*ss) << v;
+//   }
+// Because 'glog/logging.h' is included before 'glog/stl_logging.h',
+// subsequent CHECK_EQ(v1, v2) for vector<...> typed variable v1 and v2 can only
+// find these operator definitions via ADL.
+//
+// Even this solution has problems -- it may pull unintended operators into the
+// namespace as well, allowing them to also be found via ADL, and creating code
+// that only works with a particular order of includes. Long term, we need to
+// move all of the *definitions* into namespace std, bet we need to ensure no
+// one references them first. This lets us take that step. We cannot define them
+// in both because that would create ambiguous overloads when both are found.
+namespace std { using ::operator<<; }
+
+#endif  // UTIL_GTL_STL_LOGGING_INL_H_
diff --git a/projects/llm_framework/include/glog/vlog_is_on.h b/projects/llm_framework/include/glog/vlog_is_on.h
new file mode 100644
index 00000000..02b0b867
--- /dev/null
+++ b/projects/llm_framework/include/glog/vlog_is_on.h
@@ -0,0 +1,129 @@
+// Copyright (c) 1999, 2007, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: Ray Sidney and many others
+//
+// Defines the VLOG_IS_ON macro that controls the variable-verbosity
+// conditional logging.
+//
+// It's used by VLOG and VLOG_IF in logging.h
+// and by RAW_VLOG in raw_logging.h to trigger the logging.
+//
+// It can also be used directly e.g. like this:
+//   if (VLOG_IS_ON(2)) {
+//     // do some logging preparation and logging
+//     // that can't be accomplished e.g. via just VLOG(2) << ...;
+//   }
+//
+// The truth value that VLOG_IS_ON(level) returns is determined by 
+// the three verbosity level flags:
+//   --v=<n>  Gives the default maximal active V-logging level;
+//            0 is the default.
+//            Normally positive values are used for V-logging levels.
+//   --vmodule=<str>  Gives the per-module maximal V-logging levels to override
+//                    the value given by --v.
+//                    E.g. "my_module=2,foo*=3" would change the logging level
+//                    for all code in source files "my_module.*" and "foo*.*"
+//                    ("-inl" suffixes are also disregarded for this matching).
+//
+// SetVLOGLevel helper function is provided to do limited dynamic control over
+// V-logging by overriding the per-module settings given via --vmodule flag.
+//
+// CAVEAT: --vmodule functionality is not available in non gcc compilers.
+//
+
+#ifndef BASE_VLOG_IS_ON_H_
+#define BASE_VLOG_IS_ON_H_
+
+#include "glog/log_severity.h"
+
+// Annoying stuff for windows -- makes sure clients can import these functions
+#ifndef GOOGLE_GLOG_DLL_DECL
+# if defined(_WIN32) && !defined(__CYGWIN__)
+#   define GOOGLE_GLOG_DLL_DECL  __declspec(dllimport)
+# else
+#   define GOOGLE_GLOG_DLL_DECL
+# endif
+#endif
+
+#if defined(__GNUC__)
+// We emit an anonymous static int* variable at every VLOG_IS_ON(n) site.
+// (Normally) the first time every VLOG_IS_ON(n) site is hit,
+// we determine what variable will dynamically control logging at this site:
+// it's either FLAGS_v or an appropriate internal variable
+// matching the current source file that represents results of
+// parsing of --vmodule flag and/or SetVLOGLevel calls.
+#define VLOG_IS_ON(verboselevel)                                \
+  __extension__  \
+  ({ static google::int32* vlocal__ = &google::kLogSiteUninitialized;           \
+     google::int32 verbose_level__ = (verboselevel);                    \
+     (*vlocal__ >= verbose_level__) &&                          \
+     ((vlocal__ != &google::kLogSiteUninitialized) ||                   \
+      (google::InitVLOG3__(&vlocal__, &FLAGS_v,                         \
+                   __FILE__, verbose_level__))); })
+#else
+// GNU extensions not available, so we do not support --vmodule.
+// Dynamic value of FLAGS_v always controls the logging level.
+#define VLOG_IS_ON(verboselevel) (FLAGS_v >= (verboselevel))
+#endif
+
+// Set VLOG(_IS_ON) level for module_pattern to log_level.
+// This lets us dynamically control what is normally set by the --vmodule flag.
+// Returns the level that previously applied to module_pattern.
+// NOTE: To change the log level for VLOG(_IS_ON) sites
+//	 that have already executed after/during InitGoogleLogging,
+//	 one needs to supply the exact --vmodule pattern that applied to them.
+//       (If no --vmodule pattern applied to them
+//       the value of FLAGS_v will continue to control them.)
+extern GOOGLE_GLOG_DLL_DECL int SetVLOGLevel(const char* module_pattern,
+                                             int log_level);
+
+// Various declarations needed for VLOG_IS_ON above: =========================
+
+// Special value used to indicate that a VLOG_IS_ON site has not been
+// initialized.  We make this a large value, so the common-case check
+// of "*vlocal__ >= verbose_level__" in VLOG_IS_ON definition
+// passes in such cases and InitVLOG3__ is then triggered.
+extern google::int32 kLogSiteUninitialized;
+
+// Helper routine which determines the logging info for a particalur VLOG site.
+//   site_flag     is the address of the site-local pointer to the controlling
+//                 verbosity level
+//   site_default  is the default to use for *site_flag
+//   fname         is the current source file name
+//   verbose_level is the argument to VLOG_IS_ON
+// We will return the return value for VLOG_IS_ON
+// and if possible set *site_flag appropriately.
+extern GOOGLE_GLOG_DLL_DECL bool InitVLOG3__(
+    google::int32** site_flag,
+    google::int32* site_default,
+    const char* fname,
+    google::int32 verbose_level);
+
+#endif  // BASE_VLOG_IS_ON_H_
diff --git a/projects/llm_framework/main_melotts/SConstruct b/projects/llm_framework/main_melotts/SConstruct
index 6663ca30..87886e09 100644
--- a/projects/llm_framework/main_melotts/SConstruct
+++ b/projects/llm_framework/main_melotts/SConstruct
@@ -25,9 +25,12 @@ REQUIREMENTS += ['samplerate']
 
 INCLUDE += [ADir('../include')]
 INCLUDE += [ADir('src/runner'), ADir('../include/onnxruntime/core/session')]
-
+LINK_SEARCH_PATH += [ADir('../static_lib/wetext')]
 LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/onnx')]
-LDFLAGS += ['-l:libcargs.a', '-l:libonnxruntime.a']
+LDFLAGS += ['-l:libcargs.a', '-l:libonnxruntime.a','-l:libglog.so','-l:libfst.so']
+
+
+LDFLAGS += []
 
 STATIC_FILES += Glob('mode_*.json')
 
diff --git a/projects/llm_framework/main_melotts/mode_melotts-en-default.json b/projects/llm_framework/main_melotts/mode_melotts-en-default.json
index 8b161169..54c64b1c 100644
--- a/projects/llm_framework/main_melotts/mode_melotts-en-default.json
+++ b/projects/llm_framework/main_melotts/mode_melotts-en-default.json
@@ -21,6 +21,8 @@
         "gbin": "g-en-def.bin",
         "tokens": "tokens-en.txt",
         "lexicon": "lexicon-en.txt",
+        "tagger": "en_tn_tagger.fst",
+        "verbalizer": "en_tn_verbalizer.fst",
         "spacker_speed": 1.2,
         "mode_rate": 44100,
         "audio_rate": 16000,
diff --git a/projects/llm_framework/main_melotts/mode_melotts-en-us.json b/projects/llm_framework/main_melotts/mode_melotts-en-us.json
index 6a375c93..d6320873 100644
--- a/projects/llm_framework/main_melotts/mode_melotts-en-us.json
+++ b/projects/llm_framework/main_melotts/mode_melotts-en-us.json
@@ -1,9 +1,9 @@
 {
     "mode": "melotts-en-us",
     "type": "tts",
-    "homepage":"https://huggingface.co/myshell-ai/MeloTTS-English",
-    "compile_flage":"pulsar2 build --input decoder-en.onnx --config config_decoder_u16.json --output_dir decoder-en --output_name decoder-en.axmodel --target_hardware AX620E --npu_mode NPU2 --compiler.check 0",
-    "pulsar_version":"3.4-3dfd5692",
+    "homepage": "https://huggingface.co/myshell-ai/MeloTTS-English",
+    "compile_flage": "pulsar2 build --input decoder-en.onnx --config config_decoder_u16.json --output_dir decoder-en --output_name decoder-en.axmodel --target_hardware AX620E --npu_mode NPU2 --compiler.check 0",
+    "pulsar_version": "3.4-3dfd5692",
     "capabilities": [
         "tts",
         "English"
@@ -21,6 +21,8 @@
         "gbin": "g-en.bin",
         "tokens": "tokens.txt",
         "lexicon": "lexicon.txt",
+        "tagger": "en_tn_tagger.fst",
+        "verbalizer": "en_tn_verbalizer.fst",
         "spacker_speed": 1.0,
         "mode_rate": 44100,
         "audio_rate": 16000,
diff --git a/projects/llm_framework/main_melotts/mode_melotts-ja-jp.json b/projects/llm_framework/main_melotts/mode_melotts-ja-jp.json
index d2df3e12..0b93f91e 100644
--- a/projects/llm_framework/main_melotts/mode_melotts-ja-jp.json
+++ b/projects/llm_framework/main_melotts/mode_melotts-ja-jp.json
@@ -21,6 +21,8 @@
         "gbin": "g-jp.bin",
         "tokens": "tokens-jp.txt",
         "lexicon": "lexicon-jp.txt",
+        "tagger": "ja_tn_tagger.fst",
+        "verbalizer": "ja_tn_verbalizer.fst",
         "spacker_speed": 1.1,
         "mode_rate": 44100,
         "audio_rate": 16000,
diff --git a/projects/llm_framework/main_melotts/mode_melotts-zh-cn.json b/projects/llm_framework/main_melotts/mode_melotts-zh-cn.json
index b5edfe02..17867b92 100644
--- a/projects/llm_framework/main_melotts/mode_melotts-zh-cn.json
+++ b/projects/llm_framework/main_melotts/mode_melotts-zh-cn.json
@@ -21,6 +21,8 @@
         "gbin": "g-zh_mix_en.bin",
         "tokens": "tokens.txt",
         "lexicon": "lexicon.txt",
+        "tagger": "zh_tn_tagger.fst",
+        "verbalizer": "zh_tn_verbalizer.fst",
         "spacker_speed": 1.1,
         "mode_rate": 44100,
         "audio_rate": 16000,
diff --git a/projects/llm_framework/main_melotts/src/main.cpp b/projects/llm_framework/main_melotts/src/main.cpp
index 5c371d97..6400ead6 100644
--- a/projects/llm_framework/main_melotts/src/main.cpp
+++ b/projects/llm_framework/main_melotts/src/main.cpp
@@ -9,7 +9,6 @@
 #include "Lexicon.hpp"
 #include <ax_sys_api.h>
 #include "AudioFile.h"
-#include "Lexicon.hpp"
 
 #include <signal.h>
 #include <sys/stat.h>
@@ -44,6 +43,8 @@ typedef struct {
     std::string tokens;
     std::string gbin;
     std::string sentence;
+    std::string tagger;
+    std::string verbalizer;
     float spacker_speed = 1.0;
     int mode_rate       = 44100;
     int audio_rate      = 16000;
@@ -169,17 +170,22 @@ class llm_task {
             CONFIG_AUTO_SET(file_body["mode_param"], length_scale);
             CONFIG_AUTO_SET(file_body["mode_param"], noise_scale_w);
             CONFIG_AUTO_SET(file_body["mode_param"], sdp_ratio);
-            mode_config_.tokens  = base_model + mode_config_.tokens;
-            mode_config_.gbin    = base_model + mode_config_.gbin;
-            mode_config_.encoder = base_model + mode_config_.encoder;
-            mode_config_.decoder = base_model + mode_config_.decoder;
-            mode_config_.lexicon = base_model + mode_config_.lexicon;
+            CONFIG_AUTO_SET(file_body["mode_param"], tagger);
+            CONFIG_AUTO_SET(file_body["mode_param"], verbalizer);
+            mode_config_.tokens     = base_model + mode_config_.tokens;
+            mode_config_.gbin       = base_model + mode_config_.gbin;
+            mode_config_.encoder    = base_model + mode_config_.encoder;
+            mode_config_.decoder    = base_model + mode_config_.decoder;
+            mode_config_.lexicon    = base_model + mode_config_.lexicon;
+            mode_config_.tagger     = base_model + mode_config_.tagger;
+            mode_config_.verbalizer = base_model + mode_config_.verbalizer;
             if (config_body.contains("awake_delay"))
                 awake_delay_ = config_body["awake_delay"].get<int>();
             else if (file_body["mode_param"].contains("awake_delay"))
                 awake_delay_ = file_body["mode_param"]["awake_delay"];
             // Load lexicon
-            lexicon_ = std::make_unique<Lexicon>(mode_config_.lexicon, mode_config_.tokens);
+            lexicon_ = std::make_unique<Lexicon>(mode_config_.lexicon, mode_config_.tokens, mode_config_.tagger,
+                                                 mode_config_.verbalizer);
             // Read g.bin
             g_matrix.resize(256, 0);
             FILE *fp = fopen(mode_config_.gbin.c_str(), "rb");
diff --git a/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp b/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp
index 29c31817..134f64c4 100644
--- a/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp
+++ b/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp
@@ -8,8 +8,10 @@
 #include <cassert>
 #include <iostream>
 #include "../../../../../SDK/components/utilities/include/sample_log.h"
+#include "processor/wetext_processor.h"
+
 // Debug logging switch - set to true to enable debug logs
-static bool DEBUG_LOGGING = true;
+static bool DEBUG_LOGGING = false;
 // Macro for debug logging
 #define DEBUG_LOG(fmt, ...)            \
     do {                               \
@@ -36,16 +38,23 @@ class Lexicon {
     std::pair<std::vector<int>, std::vector<int>> unknown_token;
     std::unordered_map<int, std::string> reverse_tokens;
 
+    wetext::Processor* m_processor;
+
 public:
     // Setter for debug logging
     static void setDebugLogging(bool enable)
     {
         DEBUG_LOGGING = enable;
     }
-    Lexicon(const std::string& lexicon_filename, const std::string& tokens_filename) : max_phrase_length(0)
+    Lexicon(const std::string& lexicon_filename, const std::string& tokens_filename, const std::string& tagger_filename,
+            const std::string& verbalizer_filename)
+        : max_phrase_length(0)
     {
-        DEBUG_LOG("Dictionary loading: %s Pronunciation table loading: %s", tokens_filename.c_str(),
-                  lexicon_filename.c_str());
+        DEBUG_LOG("Dictionary loading: %s Pronunciation table loading: %s tagger_filename: %s verbalizer_filename: %s",
+                  tokens_filename.c_str(), lexicon_filename.c_str(), tagger_filename.c_str(),
+                  verbalizer_filename.c_str());
+
+        m_processor = new wetext::Processor(tagger_filename, verbalizer_filename);
 
         std::unordered_map<std::string, int> tokens;
         std::ifstream ifs(tokens_filename);
@@ -198,6 +207,12 @@ class Lexicon {
     void convert(const std::string& text, std::vector<int>& phones, std::vector<int>& tones)
     {
         DEBUG_LOG("\nStarting text processing: \"%s\"", text.c_str());
+
+        std::string taggedText = m_processor->Tag(text);
+        DEBUG_LOG("\taggedText processing: \"%s\"", taggedText.c_str());
+        std::string normalizedText = m_processor->Verbalize(taggedText);
+        DEBUG_LOG("\normalizedText processing: \"%s\"", normalizedText.c_str());
+
         DEBUG_LOG("=======Matching Results=======");
         DEBUG_LOG("Unit\t|\tPhonemes\t|\tTones");
         DEBUG_LOG("-----------------------------");
@@ -205,7 +220,7 @@ class Lexicon {
         tones.insert(tones.end(), unknown_token.second.begin(), unknown_token.second.end());
         DEBUG_LOG("<BOS>\t|\t%s\t|\t%s", phonesToString(unknown_token.first).c_str(),
                   tonesToString(unknown_token.second).c_str());
-        auto chars = splitEachChar(text);
+        auto chars = splitEachChar(normalizedText);
         int i      = 0;
         while (i < chars.size()) {
             if (is_english(chars[i])) {
diff --git a/projects/llm_framework/main_melotts/src/runner/base64.cpp b/projects/llm_framework/main_melotts/src/runner/base64.cpp
index 5e0fd6ad..e8e1add3 100644
--- a/projects/llm_framework/main_melotts/src/runner/base64.cpp
+++ b/projects/llm_framework/main_melotts/src/runner/base64.cpp
@@ -1,17 +1,13 @@
 #include "base64.h"
 
 static uint8 alphabet_map[] = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
-static uint8 reverse_map[] =
-{
-255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-	255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-	255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 62, 255, 255, 255, 63,
-	52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 255, 255, 255, 255, 255, 255,
-	255, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14,
-	15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 255, 255, 255, 255, 255,
-	255, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40,
-	41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 255, 255, 255, 255, 255
-};
+static uint8 reverse_map[]  = {
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 62,
+    255, 255, 255, 63,  52,  53,  54,  55,  56,  57,  58,  59,  60,  61,  255, 255, 255, 255, 255, 255, 255, 0,
+    1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,  15,  16,  17,  18,  19,  20,  21,  22,
+    23,  24,  25,  255, 255, 255, 255, 255, 255, 26,  27,  28,  29,  30,  31,  32,  33,  34,  35,  36,  37,  38,
+    39,  40,  41,  42,  43,  44,  45,  46,  47,  48,  49,  50,  51,  255, 255, 255, 255, 255};
 
 // //GB2312到UTF-8的转换
 // char* G2U(const char* gb2312)
@@ -42,26 +38,27 @@ static uint8 reverse_map[] =
 // 	return str;
 // }
 
-// uint32 base64_encode(char* input, uint8* encode)
+// base64_uint32 base64_encode(char* input, uint8* encode)
 // {
 // 	//1、包含中文的字符串 字符编码（windows默认是gbk）转换成unicode
-	
+
 // 	//2、字符编码方式是utf-8的二进制
 // 	// uint8* text = (uint8*)G2U(input);
-// 	uint32 text_len = (uint32)strlen((char*)input);
+// 	base64_uint32 text_len = (base64_uint32)strlen((char*)input);
 
-// 	uint32 i, j;
+// 	base64_uint32 i, j;
 // 	for (i = 0, j = 0; i + 3 <= text_len; i += 3)
 // 	{
-// 		encode[j++] = alphabet_map[text[i] >> 2];                             //取出第一个字符的前6位并找出对应的结果字符
-// 		encode[j++] = alphabet_map[((text[i] << 4) & 0x30) | (text[i + 1] >> 4)];     //将第一个字符的后2位与第二个字符的前4位进行组合并找到对应的结果字符
-// 		encode[j++] = alphabet_map[((text[i + 1] << 2) & 0x3c) | (text[i + 2] >> 6)];   //将第二个字符的后4位与第三个字符的前2位组合并找出对应的结果字符
-// 		encode[j++] = alphabet_map[text[i + 2] & 0x3f];                         //取出第三个字符的后6位并找出结果字符
+// 		encode[j++] = alphabet_map[text[i] >> 2]; //取出第一个字符的前6位并找出对应的结果字符 		encode[j++] =
+// alphabet_map[((text[i] << 4) & 0x30) | (text[i + 1] >> 4)];
+// //将第一个字符的后2位与第二个字符的前4位进行组合并找到对应的结果字符 		encode[j++] = alphabet_map[((text[i + 1] << 2) &
+// 0x3c) | (text[i + 2] >> 6)];   //将第二个字符的后4位与第三个字符的前2位组合并找出对应的结果字符 		encode[j++] =
+// alphabet_map[text[i + 2] & 0x3f];                         //取出第三个字符的后6位并找出结果字符
 // 	}
 
 // 	if (i < text_len)
 // 	{
-// 		uint32 tail = text_len - i;
+// 		base64_uint32 tail = text_len - i;
 // 		if (tail == 1)
 // 		{
 // 			encode[j++] = alphabet_map[text[i] >> 2];
@@ -81,40 +78,41 @@ static uint8 reverse_map[] =
 // 	return j;
 // }
 
-int base64_decode(const uint8* code, uint32 code_len, char* str)
+int base64_decode(const uint8* code, base64_uint32 code_len, char* str)
 {
-	uint8 plain[1024];
-	assert((code_len & 0x03) == 0);  //如果它的条件返回错误，则终止程序执行。4的倍数。
+    uint8 plain[1024];
+    assert((code_len & 0x03) == 0);  // 如果它的条件返回错误，则终止程序执行。4的倍数。
 
-	uint32 i, j = 0;
-	uint8 quad[4];
-	for (i = 0; i < code_len; i += 4)
-	{
-		for (uint32 k = 0; k < 4; k++)
-		{
-			quad[k] = reverse_map[code[i + k]];//分组，每组四个分别依次转换为base64表内的十进制数
-		}
+    base64_uint32 i, j = 0;
+    uint8 quad[4];
+    for (i = 0; i < code_len; i += 4) {
+        for (base64_uint32 k = 0; k < 4; k++) {
+            quad[k] = reverse_map[code[i + k]];  // 分组，每组四个分别依次转换为base64表内的十进制数
+        }
 
-		assert(quad[0] < 64 && quad[1] < 64);
+        assert(quad[0] < 64 && quad[1] < 64);
 
-		plain[j++] = (quad[0] << 2) | (quad[1] >> 4); //取出第一个字符对应base64表的十进制数的前6位与第二个字符对应base64表的十进制数的前2位进行组合
+        plain[j++] =
+            (quad[0] << 2) |
+            (quad[1] >>
+             4);  // 取出第一个字符对应base64表的十进制数的前6位与第二个字符对应base64表的十进制数的前2位进行组合
 
-		if (quad[2] >= 64)
-			break;
-		else if (quad[3] >= 64)
-		{
-			plain[j++] = (quad[1] << 4) | (quad[2] >> 2); //取出第二个字符对应base64表的十进制数的后4位与第三个字符对应base64表的十进制数的前4位进行组合
-			break;
-		}
-		else
-		{
-			plain[j++] = (quad[1] << 4) | (quad[2] >> 2);
-			plain[j++] = (quad[2] << 6) | quad[3];//取出第三个字符对应base64表的十进制数的后2位与第4个字符进行组合
-		}
-	}
-	plain[j] = 0;
-	// char str[1024] = "";
-	strcpy(str, (char*)plain);
-	// strcpy_s(str, sizeof(plain), U2G(str));
-	return j;
+        if (quad[2] >= 64)
+            break;
+        else if (quad[3] >= 64) {
+            plain[j++] =
+                (quad[1] << 4) |
+                (quad[2] >>
+                 2);  // 取出第二个字符对应base64表的十进制数的后4位与第三个字符对应base64表的十进制数的前4位进行组合
+            break;
+        } else {
+            plain[j++] = (quad[1] << 4) | (quad[2] >> 2);
+            plain[j++] = (quad[2] << 6) | quad[3];  // 取出第三个字符对应base64表的十进制数的后2位与第4个字符进行组合
+        }
+    }
+    plain[j] = 0;
+    // char str[1024] = "";
+    strcpy(str, (char*)plain);
+    // strcpy_s(str, sizeof(plain), U2G(str));
+    return j;
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/src/runner/base64.h b/projects/llm_framework/main_melotts/src/runner/base64.h
index 8e3dcb6c..f7a01c7e 100644
--- a/projects/llm_framework/main_melotts/src/runner/base64.h
+++ b/projects/llm_framework/main_melotts/src/runner/base64.h
@@ -2,10 +2,10 @@
 
 #include <string.h>
 #include <stdlib.h>
-#include <wchar.h> 
+#include <wchar.h>
 #include <assert.h>
 #include <iostream>
-typedef unsigned char     uint8;
-typedef unsigned long    uint32;
-// uint32 base64_encode(char* input, uint8* encode);
-int base64_decode(const uint8* code, uint32 code_len, char* str);
\ No newline at end of file
+typedef unsigned char uint8;
+typedef unsigned long base64_uint32;
+// base64_uint32 base64_encode(char* input, uint8* encode);
+int base64_decode(const uint8* code, base64_uint32 code_len, char* str);
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/src/runner/processor/CMakeLists.txt b/projects/llm_framework/main_melotts/src/runner/processor/CMakeLists.txt
new file mode 100644
index 00000000..a2e8d97c
--- /dev/null
+++ b/projects/llm_framework/main_melotts/src/runner/processor/CMakeLists.txt
@@ -0,0 +1,13 @@
+add_library(wetext_processor STATIC
+  wetext_processor.cc
+  wetext_token_parser.cc
+)
+if(ANDROID)
+  target_link_libraries(wetext_processor PUBLIC fst wetext_utils)
+else()
+  if(MSVC)
+    target_link_libraries(wetext_processor PUBLIC fst wetext_utils)
+  else()
+    target_link_libraries(wetext_processor PUBLIC dl fst wetext_utils)
+  endif()
+endif()
diff --git a/projects/llm_framework/main_melotts/src/runner/processor/wetext_processor.cc b/projects/llm_framework/main_melotts/src/runner/processor/wetext_processor.cc
new file mode 100644
index 00000000..eec45a24
--- /dev/null
+++ b/projects/llm_framework/main_melotts/src/runner/processor/wetext_processor.cc
@@ -0,0 +1,86 @@
+// Copyright (c) 2022 Zhendong Peng (pzd17@tsinghua.org.cn)
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "processor/wetext_processor.h"
+
+using fst::StringTokenType;
+
+namespace wetext {
+Processor::Processor(const std::string& tagger_path, const std::string& verbalizer_path)
+{
+    tagger_.reset(StdVectorFst::Read(tagger_path));
+    verbalizer_.reset(StdVectorFst::Read(verbalizer_path));
+    compiler_ = std::make_shared<StringCompiler<StdArc>>(StringTokenType::BYTE);
+    printer_  = std::make_shared<StringPrinter<StdArc>>(StringTokenType::BYTE);
+
+    if (tagger_path.find("zh_tn_") != tagger_path.npos) {
+        parse_type_ = ParseType::kZH_TN;
+    } else if (tagger_path.find("zh_itn_") != tagger_path.npos) {
+        parse_type_ = ParseType::kZH_ITN;
+    } else if (tagger_path.find("en_tn_") != tagger_path.npos) {
+        parse_type_ = ParseType::kEN_TN;
+    } else if (tagger_path.find("ja_tn_") != tagger_path.npos) {
+        parse_type_ = ParseType::kZH_TN;  // 如果是日语的文件开始，也使用中文的规则进行转换
+    } else {
+        LOG(FATAL) << "Invalid fst prefix, prefix should contain" << " either \"_tn_\" or \"_itn_\".";
+    }
+}
+
+std::string Processor::ShortestPath(const StdVectorFst& lattice)
+{
+    StdVectorFst shortest_path;
+    fst::ShortestPath(lattice, &shortest_path, 1, true);
+
+    std::string output;
+    printer_->operator()(shortest_path, &output);
+    return output;
+}
+
+std::string Processor::Compose(const std::string& input, const StdVectorFst* fst)
+{
+    StdVectorFst input_fst;
+    compiler_->operator()(input, &input_fst);
+
+    StdVectorFst lattice;
+    fst::Compose(input_fst, *fst, &lattice);
+    return ShortestPath(lattice);
+}
+
+std::string Processor::Tag(const std::string& input)
+{
+    if (input.empty()) {
+        return "";
+    }
+    return Compose(input, tagger_.get());
+}
+
+std::string Processor::Verbalize(const std::string& input)
+{
+    if (input.empty()) {
+        return "";
+    }
+    TokenParser parser(parse_type_);
+    std::string output = parser.Reorder(input);
+
+    output = Compose(output, verbalizer_.get());
+    output.erase(std::remove(output.begin(), output.end(), '\0'), output.end());
+    return output;
+}
+
+std::string Processor::Normalize(const std::string& input)
+{
+    return Verbalize(Tag(input));
+}
+
+}  // namespace wetext
diff --git a/projects/llm_framework/main_melotts/src/runner/processor/wetext_processor.h b/projects/llm_framework/main_melotts/src/runner/processor/wetext_processor.h
new file mode 100644
index 00000000..e11d307e
--- /dev/null
+++ b/projects/llm_framework/main_melotts/src/runner/processor/wetext_processor.h
@@ -0,0 +1,51 @@
+// Copyright (c) 2022 Zhendong Peng (pzd17@tsinghua.org.cn)
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef PROCESSOR_WETEXT_PROCESSOR_H_
+#define PROCESSOR_WETEXT_PROCESSOR_H_
+
+#include <memory>
+#include <string>
+
+#include "fst/fstlib.h"
+
+#include "processor/wetext_token_parser.h"
+
+using fst::StdArc;
+using fst::StdVectorFst;
+using fst::StringCompiler;
+using fst::StringPrinter;
+
+namespace wetext {
+class Processor {
+ public:
+  Processor(const std::string& tagger_path, const std::string& verbalizer_path);
+  std::string Tag(const std::string& input);
+  std::string Verbalize(const std::string& input);
+  std::string Normalize(const std::string& input);
+
+ private:
+  std::string ShortestPath(const StdVectorFst& lattice);
+  std::string Compose(const std::string& input, const StdVectorFst* fst);
+
+  ParseType parse_type_;
+  std::shared_ptr<StdVectorFst> tagger_ = nullptr;
+  std::shared_ptr<StdVectorFst> verbalizer_ = nullptr;
+  std::shared_ptr<StringCompiler<StdArc>> compiler_ = nullptr;
+  std::shared_ptr<StringPrinter<StdArc>> printer_ = nullptr;
+};
+
+}  // namespace wetext
+
+#endif  // PROCESSOR_WETEXT_PROCESSOR_H_
diff --git a/projects/llm_framework/main_melotts/src/runner/processor/wetext_token_parser.cc b/projects/llm_framework/main_melotts/src/runner/processor/wetext_token_parser.cc
new file mode 100644
index 00000000..a600eead
--- /dev/null
+++ b/projects/llm_framework/main_melotts/src/runner/processor/wetext_token_parser.cc
@@ -0,0 +1,161 @@
+// Copyright (c) 2022 Zhendong Peng (pzd17@tsinghua.org.cn)
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "processor/wetext_token_parser.h"
+
+#include "utils/wetext_log.h"
+#include "utils/wetext_string.h"
+
+namespace wetext {
+const char EOS[] = "<EOS>";
+const std::set<std::string> UTF8_WHITESPACE = {" ", "\t", "\n", "\r",
+                                               "\x0b\x0c"};
+const std::set<std::string> ASCII_LETTERS = {
+    "a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n",
+    "o", "p", "q", "r", "s", "t", "u", "v", "w", "x", "y", "z", "A", "B",
+    "C", "D", "E", "F", "G", "H", "I", "J", "K", "L", "M", "N", "O", "P",
+    "Q", "R", "S", "T", "U", "V", "W", "X", "Y", "Z", "_"};
+const std::unordered_map<std::string, std::vector<std::string>> ZH_TN_ORDERS = {
+    {"date", {"year", "month", "day"}},
+    {"fraction", {"denominator", "numerator"}},
+    {"measure", {"denominator", "numerator", "value"}},
+    {"money", {"value", "currency"}},
+    {"time", {"noon", "hour", "minute", "second"}}};
+const std::unordered_map<std::string, std::vector<std::string>> EN_TN_ORDERS = {
+    {"date", {"preserve_order", "text", "day", "month", "year"}},
+    {"money", {"integer_part", "fractional_part", "quantity", "currency_maj"}}};
+const std::unordered_map<std::string, std::vector<std::string>> ZH_ITN_ORDERS =
+    {{"date", {"year", "month", "day"}},
+     {"fraction", {"sign", "numerator", "denominator"}},
+     {"measure", {"numerator", "denominator", "value"}},
+     {"money", {"currency", "value", "decimal"}},
+     {"time", {"hour", "minute", "second", "noon"}}};
+
+TokenParser::TokenParser(ParseType type) {
+  if (type == ParseType::kZH_TN) {
+    orders_ = ZH_TN_ORDERS;
+  } else if (type == ParseType::kZH_ITN) {
+    orders_ = ZH_ITN_ORDERS;
+  } else if (type == ParseType::kEN_TN) {
+    orders_ = EN_TN_ORDERS;
+  } else {
+    LOG(FATAL) << "Invalid order";
+  }
+}
+
+void TokenParser::Load(const std::string& input) {
+  wetext::SplitUTF8StringToChars(input, &text_);
+  CHECK_GT(text_.size(), 0);
+  index_ = 0;
+  ch_ = text_[0];
+}
+
+bool TokenParser::Read() {
+  if (index_ < text_.size() - 1) {
+    index_ += 1;
+    ch_ = text_[index_];
+    return true;
+  }
+  ch_ = EOS;
+  return false;
+}
+
+bool TokenParser::ParseWs() {
+  bool not_eos = ch_ != EOS;
+  while (not_eos && ch_ == " ") {
+    not_eos = Read();
+  }
+  return not_eos;
+}
+
+bool TokenParser::ParseChar(const std::string& exp) {
+  if (ch_ == exp) {
+    Read();
+    return true;
+  }
+  return false;
+}
+
+bool TokenParser::ParseChars(const std::string& exp) {
+  bool ok = false;
+  std::vector<std::string> chars;
+  wetext::SplitUTF8StringToChars(exp, &chars);
+  for (const auto& x : chars) {
+    ok |= ParseChar(x);
+  }
+  return ok;
+}
+
+std::string TokenParser::ParseKey() {
+  CHECK_NE(ch_, EOS);
+  CHECK_EQ(UTF8_WHITESPACE.count(ch_), 0);
+
+  std::string key = "";
+  while (ASCII_LETTERS.count(ch_) > 0) {
+    key += ch_;
+    Read();
+  }
+  return key;
+}
+
+std::string TokenParser::ParseValue() {
+  CHECK_NE(ch_, EOS);
+  bool escape = false;
+
+  std::string value = "";
+  while (ch_ != "\"") {
+    value += ch_;
+    escape = ch_ == "\\";
+    Read();
+    if (escape) {
+      escape = false;
+      value += ch_;
+      Read();
+    }
+  }
+  return value;
+}
+
+void TokenParser::Parse(const std::string& input) {
+  Load(input);
+  while (ParseWs()) {
+    std::string name = ParseKey();
+    ParseChars(" { ");
+
+    Token token(name);
+    while (ParseWs()) {
+      if (ch_ == "}") {
+        ParseChar("}");
+        break;
+      }
+      std::string key = ParseKey();
+      ParseChars(": \"");
+      std::string value = ParseValue();
+      ParseChar("\"");
+      token.Append(key, value);
+    }
+    tokens_.emplace_back(token);
+  }
+}
+
+std::string TokenParser::Reorder(const std::string& input) {
+  Parse(input);
+  std::string output = "";
+  for (auto& token : tokens_) {
+    output += token.String(orders_) + " ";
+  }
+  return Trim(output);
+}
+
+}  // namespace wetext
diff --git a/projects/llm_framework/main_melotts/src/runner/processor/wetext_token_parser.h b/projects/llm_framework/main_melotts/src/runner/processor/wetext_token_parser.h
new file mode 100644
index 00000000..34aba979
--- /dev/null
+++ b/projects/llm_framework/main_melotts/src/runner/processor/wetext_token_parser.h
@@ -0,0 +1,94 @@
+// Copyright (c) 2022 Zhendong Peng (pzd17@tsinghua.org.cn)
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef PROCESSOR_WETEXT_TOKEN_PARSER_H_
+#define PROCESSOR_WETEXT_TOKEN_PARSER_H_
+
+#include <set>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+namespace wetext {
+
+extern const char EOS[];
+extern const std::set<std::string> UTF8_WHITESPACE;
+extern const std::set<std::string> ASCII_LETTERS;
+extern const std::unordered_map<std::string, std::vector<std::string>>
+    ZH_TN_ORDERS;
+extern const std::unordered_map<std::string, std::vector<std::string>>
+    ZH_ITN_ORDERS;
+extern const std::unordered_map<std::string, std::vector<std::string>>
+    EN_TN_ORDERS;
+
+struct Token {
+  std::string name;
+  std::vector<std::string> order;
+  std::unordered_map<std::string, std::string> members;
+
+  explicit Token(const std::string& name) : name(name) {}
+
+  void Append(const std::string& key, const std::string& value) {
+    order.emplace_back(key);
+    members[key] = value;
+  }
+
+  std::string String(
+      const std::unordered_map<std::string, std::vector<std::string>>& orders) {
+    std::string output = name + " {";
+    if (orders.count(name) > 0) {
+      order = orders.at(name);
+    }
+
+    for (const auto& key : order) {
+      if (members.count(key) == 0) {
+        continue;
+      }
+      output += " " + key + ": \"" + members[key] + "\"";
+    }
+    return output + " }";
+  }
+};
+
+enum ParseType {
+  kZH_TN = 0x00,   // Chinese Text Normalization
+  kZH_ITN = 0x01,  // Chinese Inverse Text Normalization
+  kEN_TN = 0x02    // English Text Normalization
+};
+
+class TokenParser {
+ public:
+  explicit TokenParser(ParseType type);
+  std::string Reorder(const std::string& input);
+
+ private:
+  void Load(const std::string& input);
+  bool Read();
+  bool ParseWs();
+  bool ParseChar(const std::string& exp);
+  bool ParseChars(const std::string& exp);
+  std::string ParseKey();
+  std::string ParseValue();
+  void Parse(const std::string& input);
+
+  int index_;
+  std::string ch_;
+  std::vector<std::string> text_;
+  std::vector<Token> tokens_;
+  std::unordered_map<std::string, std::vector<std::string>> orders_;
+};
+
+}  // namespace wetext
+
+#endif  // PROCESSOR_WETEXT_TOKEN_PARSER_H_
diff --git a/projects/llm_framework/main_melotts/src/runner/utils/CMakeLists.txt b/projects/llm_framework/main_melotts/src/runner/utils/CMakeLists.txt
new file mode 100644
index 00000000..30071f4c
--- /dev/null
+++ b/projects/llm_framework/main_melotts/src/runner/utils/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_library(wetext_utils STATIC wetext_string.cc)
+
+target_link_libraries(wetext_utils PUBLIC glog)
diff --git a/projects/llm_framework/main_melotts/src/runner/utils/wetext_flags.h b/projects/llm_framework/main_melotts/src/runner/utils/wetext_flags.h
new file mode 100644
index 00000000..c1d30df3
--- /dev/null
+++ b/projects/llm_framework/main_melotts/src/runner/utils/wetext_flags.h
@@ -0,0 +1,23 @@
+// Copyright (c) 2022 Zhendong Peng (pzd17@tsinghua.org.cn)
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef UTILS_WETEXT_FLAGS_H_
+#define UTILS_WETEXT_FLAGS_H_
+
+// Because openfst is a dynamic library compiled with gflags/glog, we must use
+// the gflags/glog from openfst to avoid them linked both statically and
+// dynamically into the executable.
+#include "fst/flags.h"
+
+#endif  // UTILS_WETEXT_FLAGS_H_
diff --git a/projects/llm_framework/main_melotts/src/runner/utils/wetext_log.h b/projects/llm_framework/main_melotts/src/runner/utils/wetext_log.h
new file mode 100644
index 00000000..b47a6a48
--- /dev/null
+++ b/projects/llm_framework/main_melotts/src/runner/utils/wetext_log.h
@@ -0,0 +1,23 @@
+// Copyright (c) 2022 Zhendong Peng (pzd17@tsinghua.org.cn)
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef UTILS_WETEXT_LOG_H_
+#define UTILS_WETEXT_LOG_H_
+
+// Because openfst is a dynamic library compiled with gflags/glog, we must use
+// the gflags/glog from openfst to avoid them linked both statically and
+// dynamically into the executable.
+#include "fst/log.h"
+
+#endif  // UTILS_WETEXT_LOG_H_
diff --git a/projects/llm_framework/main_melotts/src/runner/utils/wetext_string.cc b/projects/llm_framework/main_melotts/src/runner/utils/wetext_string.cc
new file mode 100644
index 00000000..4df9ec91
--- /dev/null
+++ b/projects/llm_framework/main_melotts/src/runner/utils/wetext_string.cc
@@ -0,0 +1,89 @@
+// Copyright (c) 2022 Zhendong Peng (pzd17@tsinghua.org.cn)
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "utils/wetext_string.h"
+
+#include "utils/wetext_log.h"
+
+namespace wetext {
+const char* WHITESPACE = " \n\r\t\f\v";
+
+int UTF8CharLength(char ch) {
+  int num_bytes = 1;
+  CHECK_LE((ch & 0xF8), 0xF0);
+  if ((ch & 0x80) == 0x00) {
+    // The first 128 characters (US-ASCII) in UTF-8 format only need one byte.
+    num_bytes = 1;
+  } else if ((ch & 0xE0) == 0xC0) {
+    // The next 1,920 characters need two bytes to encode,
+    // which covers the remainder of almost all Latin-script alphabets.
+    num_bytes = 2;
+  } else if ((ch & 0xF0) == 0xE0) {
+    // Three bytes are needed for characters in the rest of
+    // the Basic Multilingual Plane, which contains virtually all characters
+    // in common use, including most Chinese, Japanese and Korean characters.
+    num_bytes = 3;
+  } else if ((ch & 0xF8) == 0xF0) {
+    // Four bytes are needed for characters in the other planes of Unicode,
+    // which include less common CJK characters, various historic scripts,
+    // mathematical symbols, and emoji (pictographic symbols).
+    num_bytes = 4;
+  }
+  return num_bytes;
+}
+
+int UTF8StringLength(const std::string& str) {
+  int len = 0;
+  int num_bytes = 1;
+  for (size_t i = 0; i < str.length(); i += num_bytes) {
+    num_bytes = UTF8CharLength(str[i]);
+    ++len;
+  }
+  return len;
+}
+
+void SplitUTF8StringToChars(const std::string& str,
+                            std::vector<std::string>* chars) {
+  chars->clear();
+  int num_bytes = 1;
+  for (size_t i = 0; i < str.length(); i += num_bytes) {
+    num_bytes = UTF8CharLength(str[i]);
+    chars->push_back(str.substr(i, num_bytes));
+  }
+}
+
+std::string Ltrim(const std::string& str) {
+  size_t start = str.find_first_not_of(WHITESPACE);
+  return (start == std::string::npos) ? "" : str.substr(start);
+}
+
+std::string Rtrim(const std::string& str) {
+  size_t end = str.find_last_not_of(WHITESPACE);
+  return end == std::string::npos ? "" : str.substr(0, end + 1);
+}
+
+std::string Trim(const std::string& str) { return Rtrim(Ltrim(str)); }
+
+void Split(const std::string& str, const std::string& delim,
+           std::vector<std::string>* output) {
+  std::string s = str;
+  size_t pos = 0;
+  while ((pos = s.find(delim)) != std::string::npos) {
+    output->emplace_back(s.substr(0, pos));
+    s.erase(0, pos + delim.length());
+  }
+  output->emplace_back(s);
+}
+
+}  // namespace wetext
diff --git a/projects/llm_framework/main_melotts/src/runner/utils/wetext_string.h b/projects/llm_framework/main_melotts/src/runner/utils/wetext_string.h
new file mode 100644
index 00000000..ae890d60
--- /dev/null
+++ b/projects/llm_framework/main_melotts/src/runner/utils/wetext_string.h
@@ -0,0 +1,42 @@
+// Copyright (c) 2022 Zhendong Peng (pzd17@tsinghua.org.cn)
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef UTILS_WETEXT_STRING_H_
+#define UTILS_WETEXT_STRING_H_
+
+#include <string>
+#include <vector>
+
+namespace wetext {
+extern const char* WHITESPACE;
+
+int UTF8CharLength(char ch);
+
+int UTF8StringLength(const std::string& str);
+
+void SplitUTF8StringToChars(const std::string& str,
+                            std::vector<std::string>* chars);
+
+std::string Ltrim(const std::string& str);
+
+std::string Rtrim(const std::string& str);
+
+std::string Trim(const std::string& str);
+
+void Split(const std::string& str, const std::string& delim,
+           std::vector<std::string>* output);
+
+}  // namespace wetext
+
+#endif  // UTILS_WETEXT_STRING_H_

From 9a20dd028de0dd4d98da4ab14af576e9f6d68b3f Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 16 May 2025 14:13:49 +0800
Subject: [PATCH 04/79] [update] update melotts, update static_lib verison

---
 projects/llm_framework/SConstruct              | 2 +-
 projects/llm_framework/main_melotts/SConstruct | 2 +-
 projects/llm_framework/tools/llm_pack.py       | 2 +-
 3 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/projects/llm_framework/SConstruct b/projects/llm_framework/SConstruct
index 5d014ef6..7282f87e 100644
--- a/projects/llm_framework/SConstruct
+++ b/projects/llm_framework/SConstruct
@@ -5,7 +5,7 @@ import shutil
 os.environ['SDK_PATH'] = os.path.normpath(str(Path(os.getcwd())/'..'/'..'/'SDK'))
 os.environ['EXT_COMPONENTS_PATH'] = os.path.normpath(str(Path(os.getcwd())/'..'/'..'/'ext_components'))
 
-version = 'v0.0.7'
+version = 'v0.0.8'
 static_lib = 'static_lib'
 update = False
 
diff --git a/projects/llm_framework/main_melotts/SConstruct b/projects/llm_framework/main_melotts/SConstruct
index 87886e09..b9993031 100644
--- a/projects/llm_framework/main_melotts/SConstruct
+++ b/projects/llm_framework/main_melotts/SConstruct
@@ -34,7 +34,7 @@ LDFLAGS += []
 
 STATIC_FILES += Glob('mode_*.json')
 
-env['COMPONENTS'].append({'target':'llm_melotts-1.7',
+env['COMPONENTS'].append({'target':'llm_melotts-1.8',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 9fd6c580..3db22980 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -359,7 +359,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-asr':[create_bin_deb,'llm-asr', '1.6', src_folder, revision],
         'llm-llm':[create_bin_deb,'llm-llm', '1.8', src_folder, revision],
         'llm-tts':[create_bin_deb,'llm-tts', '1.6', src_folder, revision],
-        'llm-melotts':[create_bin_deb,'llm-melotts', '1.7', src_folder, revision],
+        'llm-melotts':[create_bin_deb,'llm-melotts', '1.8', src_folder, revision],
         'llm-camera':[create_bin_deb,'llm-camera', '1.8', src_folder, revision, 'lib-llm'],
         'llm-vlm':[create_bin_deb,'llm-vlm', '1.7', src_folder, revision],
         'llm-yolo':[create_bin_deb,'llm-yolo', '1.8', src_folder, revision],

From daeaf4b830d0d60e5ed9456f53e38de38d4ae3d0 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 16 May 2025 16:20:19 +0800
Subject: [PATCH 05/79] [update] update lib-llm version, update melotts model
 version.

---
 projects/llm_framework/main/SConstruct         |  2 ++
 projects/llm_framework/main_melotts/SConstruct |  6 ++----
 projects/llm_framework/tools/llm_pack.py       | 10 +++++-----
 3 files changed, 9 insertions(+), 9 deletions(-)

diff --git a/projects/llm_framework/main/SConstruct b/projects/llm_framework/main/SConstruct
index 77237295..a95f5367 100644
--- a/projects/llm_framework/main/SConstruct
+++ b/projects/llm_framework/main/SConstruct
@@ -24,6 +24,8 @@ STATIC_FILES += [AFile('../static_lib/sherpa/ncnn/libsherpa-ncnn-core.so'),
                  AFile('../static_lib/sherpa/ncnn/libncnn.so'),
                  AFile('../static_lib/libtts.so'),
                  AFile('../static_lib/sherpa/ncnn/libkaldi-native-fbank-core.so'),
+                 AFile('../static_lib/wetext/libglog.so.0'),
+                 AFile('../static_lib/wetext/libfst.so.16'),
                  ]
 
 env['COMPONENTS'].append({'target':'static_file-1.0',
diff --git a/projects/llm_framework/main_melotts/SConstruct b/projects/llm_framework/main_melotts/SConstruct
index b9993031..8faa8d86 100644
--- a/projects/llm_framework/main_melotts/SConstruct
+++ b/projects/llm_framework/main_melotts/SConstruct
@@ -27,10 +27,8 @@ INCLUDE += [ADir('../include')]
 INCLUDE += [ADir('src/runner'), ADir('../include/onnxruntime/core/session')]
 LINK_SEARCH_PATH += [ADir('../static_lib/wetext')]
 LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/onnx')]
-LDFLAGS += ['-l:libcargs.a', '-l:libonnxruntime.a','-l:libglog.so','-l:libfst.so']
-
-
-LDFLAGS += []
+LDFLAGS += ['-l:libcargs.a', '-l:libonnxruntime.a']
+REQUIREMENTS += ['glog', 'fst']
 
 STATIC_FILES += Glob('mode_*.json')
 
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 3db22980..20a04248 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -225,7 +225,7 @@ def create_data_deb(package_name, version, src_folder, revision = 'm5stack1', de
     shutil.rmtree(deb_folder)
     return package_name + " creat success!"
 
-def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', depends = 'lib-llm (>= 1.7)'):
+def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', depends = 'lib-llm (>= 1.8)'):
     bin_files = glob.glob(os.path.join(src_folder, package_name.replace("-", "_") + "-*"))
     version_info = 0.0
     print(os.path.join(src_folder, package_name + "-*"))
@@ -352,7 +352,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         cpu_count = cpu_count - 2
     # cpu_count = 50
     Tasks = {
-        'lib-llm':[create_lib_deb,'lib-llm', '1.7', src_folder, revision],
+        'lib-llm':[create_lib_deb,'lib-llm', '1.8', src_folder, revision],
         'llm-sys':[create_bin_deb,'llm-sys', '1.6', src_folder, revision],
         'llm-audio':[create_bin_deb,'llm-audio', '1.6', src_folder, revision],
         'llm-kws':[create_bin_deb,'llm-kws', '1.7', src_folder, revision],
@@ -376,10 +376,10 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-model-sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01':[create_data_deb,'llm-model-sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01', '0.3', src_folder, revision],
         'llm-model-single-speaker-english-fast':[create_data_deb,'llm-model-single-speaker-english-fast', '0.3', src_folder, revision],
         'llm-model-single-speaker-fast':[create_data_deb,'llm-model-single-speaker-fast', '0.3', src_folder, revision],
-        'llm-model-melotts-zh-cn':[create_data_deb,'llm-model-melotts-zh-cn', '0.5', src_folder, revision],
+        'llm-model-melotts-zh-cn':[create_data_deb,'llm-model-melotts-zh-cn', '0.6', src_folder, revision],
         'llm-model-melotts-en-us':[create_data_deb,'llm-model-melotts-en-us', '0.5', src_folder, revision],
-        'llm-model-melotts-en-default':[create_data_deb,'llm-model-melotts-en-default', '0.5', src_folder, revision],
-        'llm-model-melotts-ja-jp':[create_data_deb,'llm-model-melotts-ja-jp', '0.5', src_folder, revision],
+        'llm-model-melotts-en-default':[create_data_deb,'llm-model-melotts-en-default', '0.6', src_folder, revision],
+        'llm-model-melotts-ja-jp':[create_data_deb,'llm-model-melotts-ja-jp', '0.6', src_folder, revision],
         'llm-model-yolo11n':[create_data_deb,'llm-model-yolo11n', data_version, src_folder, revision],
         'llm-model-yolo11n-pose':[create_data_deb,'llm-model-yolo11n-pose', '0.3', src_folder, revision],
         'llm-model-yolo11n-hand-pose':[create_data_deb,'llm-model-yolo11n-hand-pose', '0.3', src_folder, revision],

From 00c0533b86a15dacf7a5446cfd1c153b4fd32856 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 16 May 2025 18:48:49 +0800
Subject: [PATCH 06/79] [update] update libonnxruntime.so

---
 projects/llm_framework/main/SConstruct         | 1 +
 projects/llm_framework/main_kws/SConstruct     | 6 ++++--
 projects/llm_framework/main_melotts/SConstruct | 4 ++--
 projects/llm_framework/main_vad/SConstruct     | 7 ++++---
 projects/llm_framework/tools/llm_pack.py       | 4 ++--
 5 files changed, 13 insertions(+), 9 deletions(-)

diff --git a/projects/llm_framework/main/SConstruct b/projects/llm_framework/main/SConstruct
index a95f5367..72d44864 100644
--- a/projects/llm_framework/main/SConstruct
+++ b/projects/llm_framework/main/SConstruct
@@ -26,6 +26,7 @@ STATIC_FILES += [AFile('../static_lib/sherpa/ncnn/libsherpa-ncnn-core.so'),
                  AFile('../static_lib/sherpa/ncnn/libkaldi-native-fbank-core.so'),
                  AFile('../static_lib/wetext/libglog.so.0'),
                  AFile('../static_lib/wetext/libfst.so.16'),
+                 AFile('../static_lib/libonnxruntime.so.1'),
                  ]
 
 env['COMPONENTS'].append({'target':'static_file-1.0',
diff --git a/projects/llm_framework/main_kws/SConstruct b/projects/llm_framework/main_kws/SConstruct
index c09ca41a..b417fc2b 100644
--- a/projects/llm_framework/main_kws/SConstruct
+++ b/projects/llm_framework/main_kws/SConstruct
@@ -29,10 +29,12 @@ INCLUDE += [ADir('../include/sherpa'),
             ]
 
 LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/onnx')]
-LDFLAGS += ['-l:libcargs.a', '-l:libonnxruntime.a',
+LDFLAGS += ['-l:libcargs.a',
             '-l:libsherpa-onnx-core.a', '-l:libkaldi-native-fbank-core.a',
             '-l:libkaldi-decoder-core.a', '-l:libssentencepiece_core.a']
 
+REQUIREMENTS += ['onnxruntime']
+
 STATIC_FILES += [os.path.join(python_venv, 'sherpa-onnx')]
 STATIC_FILES += Glob('llm-kws_text2token.py')
 STATIC_FILES += Glob('mode_*.json')
@@ -55,7 +57,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_kws-1.7',
+env['COMPONENTS'].append({'target':'llm_kws-1.8',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_melotts/SConstruct b/projects/llm_framework/main_melotts/SConstruct
index 8faa8d86..69d9d464 100644
--- a/projects/llm_framework/main_melotts/SConstruct
+++ b/projects/llm_framework/main_melotts/SConstruct
@@ -26,10 +26,10 @@ REQUIREMENTS += ['samplerate']
 INCLUDE += [ADir('../include')]
 INCLUDE += [ADir('src/runner'), ADir('../include/onnxruntime/core/session')]
 LINK_SEARCH_PATH += [ADir('../static_lib/wetext')]
-LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/onnx')]
-LDFLAGS += ['-l:libcargs.a', '-l:libonnxruntime.a']
 REQUIREMENTS += ['glog', 'fst']
 
+REQUIREMENTS += ['onnxruntime']
+
 STATIC_FILES += Glob('mode_*.json')
 
 env['COMPONENTS'].append({'target':'llm_melotts-1.8',
diff --git a/projects/llm_framework/main_vad/SConstruct b/projects/llm_framework/main_vad/SConstruct
index 2493e159..bbc289d5 100644
--- a/projects/llm_framework/main_vad/SConstruct
+++ b/projects/llm_framework/main_vad/SConstruct
@@ -23,12 +23,13 @@ LINK_SEARCH_PATH += [ADir('../static_lib')]
 INCLUDE += [ADir('../include/sherpa')]
 
 LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/onnx')]
-LDFLAGS += ['-l:libsherpa-onnx-core.a',
-            '-l:libonnxruntime.a']
+LDFLAGS += ['-l:libsherpa-onnx-core.a']
+
+REQUIREMENTS += ['onnxruntime']
 
 STATIC_FILES += Glob('mode_*.json')
 
-env['COMPONENTS'].append({'target':'llm_vad-1.6',
+env['COMPONENTS'].append({'target':'llm_vad-1.7',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 20a04248..e593f873 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -355,7 +355,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'lib-llm':[create_lib_deb,'lib-llm', '1.8', src_folder, revision],
         'llm-sys':[create_bin_deb,'llm-sys', '1.6', src_folder, revision],
         'llm-audio':[create_bin_deb,'llm-audio', '1.6', src_folder, revision],
-        'llm-kws':[create_bin_deb,'llm-kws', '1.7', src_folder, revision],
+        'llm-kws':[create_bin_deb,'llm-kws', '1.8', src_folder, revision],
         'llm-asr':[create_bin_deb,'llm-asr', '1.6', src_folder, revision],
         'llm-llm':[create_bin_deb,'llm-llm', '1.8', src_folder, revision],
         'llm-tts':[create_bin_deb,'llm-tts', '1.6', src_folder, revision],
@@ -365,7 +365,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-yolo':[create_bin_deb,'llm-yolo', '1.8', src_folder, revision],
         'llm-skel':[create_bin_deb,'llm-skel', version, src_folder, revision],
         'llm-depth-anything':[create_bin_deb,'llm-depth-anything', '1.6', src_folder, revision],
-        'llm-vad':[create_bin_deb,'llm-vad', '1.6', src_folder, revision],
+        'llm-vad':[create_bin_deb,'llm-vad', '1.7', src_folder, revision],
         'llm-whisper':[create_bin_deb,'llm-whisper', '1.7', src_folder, revision],
         'llm-openai-api':[create_bin_deb,'llm-openai-api', '1.7', src_folder, revision],
         'llm-model-audio-en-us':[create_data_deb,'llm-model-audio-en-us', data_version, src_folder, revision],

From f775786033855e254d0710651b2d073b21a339f7 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Tue, 20 May 2025 14:19:44 +0800
Subject: [PATCH 07/79] [update] add en-au, en-br, en-india, en-us model.
 Format code.

---
 .../main_melotts/mode_melotts-en-au.json      | 31 +++++++
 .../main_melotts/mode_melotts-en-br.json      | 31 +++++++
 .../main_melotts/mode_melotts-en-india.json   | 31 +++++++
 .../main_melotts/mode_melotts-en-us.json      | 16 ++--
 .../llm_framework/main_melotts/src/main.cpp   |  9 +-
 .../main_melotts/src/runner/Lexicon.hpp       | 85 ++++++++-----------
 projects/llm_framework/tools/llm_pack.py      |  5 +-
 7 files changed, 145 insertions(+), 63 deletions(-)
 create mode 100644 projects/llm_framework/main_melotts/mode_melotts-en-au.json
 create mode 100644 projects/llm_framework/main_melotts/mode_melotts-en-br.json
 create mode 100644 projects/llm_framework/main_melotts/mode_melotts-en-india.json

diff --git a/projects/llm_framework/main_melotts/mode_melotts-en-au.json b/projects/llm_framework/main_melotts/mode_melotts-en-au.json
new file mode 100644
index 00000000..17b71a94
--- /dev/null
+++ b/projects/llm_framework/main_melotts/mode_melotts-en-au.json
@@ -0,0 +1,31 @@
+{
+    "mode": "melotts-en-au",
+    "type": "tts",
+    "homepage": "https://github.com/ml-inory/melotts.axera/tree/main/model_convert",
+    "compile_flage": "pulsar2 build --input decoder-en.onnx --config config_decoder_u16.json --output_dir decoder_en-au --output_name decoder-en-au.axmodel --target_hardware AX620E --npu_mode NPU2 --compiler.check 0",
+    "pulsar_version": "4.0-64a0e58f",
+    "capabilities": [
+        "tts",
+        "English"
+    ],
+    "input_type": [
+        "tts.utf-8"
+    ],
+    "output_type": [
+        "tts.wav",
+        "sys.play.0_1"
+    ],
+    "mode_param": {
+        "encoder": "encoder-en-au.ort",
+        "decoder": "decoder-en-au.axmodel",
+        "gbin": "g-en-au.bin",
+        "tokens": "tokens-en.txt",
+        "lexicon": "lexicon-en.txt",
+        "tagger": "en_tn_tagger.fst",
+        "verbalizer": "en_tn_verbalizer.fst",
+        "spacker_speed": 1.2,
+        "mode_rate": 44100,
+        "audio_rate": 16000,
+        "awake_delay": 1000
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/mode_melotts-en-br.json b/projects/llm_framework/main_melotts/mode_melotts-en-br.json
new file mode 100644
index 00000000..d5e68979
--- /dev/null
+++ b/projects/llm_framework/main_melotts/mode_melotts-en-br.json
@@ -0,0 +1,31 @@
+{
+    "mode": "melotts-en-br",
+    "type": "tts",
+    "homepage": "https://github.com/ml-inory/melotts.axera/tree/main/model_convert",
+    "compile_flage": "pulsar2 build --input decoder-en.onnx --config config_decoder_u16.json --output_dir decoder_en-br --output_name decoder-en-br.axmodel --target_hardware AX620E --npu_mode NPU2 --compiler.check 0",
+    "pulsar_version": "4.0-64a0e58f",
+    "capabilities": [
+        "tts",
+        "English"
+    ],
+    "input_type": [
+        "tts.utf-8"
+    ],
+    "output_type": [
+        "tts.wav",
+        "sys.play.0_1"
+    ],
+    "mode_param": {
+        "encoder": "encoder-en-br.ort",
+        "decoder": "decoder-en-br.axmodel",
+        "gbin": "g-en-br.bin",
+        "tokens": "tokens-en.txt",
+        "lexicon": "lexicon-en.txt",
+        "tagger": "en_tn_tagger.fst",
+        "verbalizer": "en_tn_verbalizer.fst",
+        "spacker_speed": 1.2,
+        "mode_rate": 44100,
+        "audio_rate": 16000,
+        "awake_delay": 1000
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/mode_melotts-en-india.json b/projects/llm_framework/main_melotts/mode_melotts-en-india.json
new file mode 100644
index 00000000..e39c1318
--- /dev/null
+++ b/projects/llm_framework/main_melotts/mode_melotts-en-india.json
@@ -0,0 +1,31 @@
+{
+    "mode": "melotts-en-india",
+    "type": "tts",
+    "homepage": "https://github.com/ml-inory/melotts.axera/tree/main/model_convert",
+    "compile_flage": "pulsar2 build --input decoder-en.onnx --config config_decoder_u16.json --output_dir decoder_en-india --output_name decoder-en-india.axmodel --target_hardware AX620E",
+    "pulsar_version": "4.0-64a0e58f",
+    "capabilities": [
+        "tts",
+        "English"
+    ],
+    "input_type": [
+        "tts.utf-8"
+    ],
+    "output_type": [
+        "tts.wav",
+        "sys.play.0_1"
+    ],
+    "mode_param": {
+        "encoder": "encoder-en-india.ort",
+        "decoder": "decoder-en-india.axmodel",
+        "gbin": "g-en-india.bin",
+        "tokens": "tokens-en.txt",
+        "lexicon": "lexicon-en.txt",
+        "tagger": "en_tn_tagger.fst",
+        "verbalizer": "en_tn_verbalizer.fst",
+        "spacker_speed": 1.2,
+        "mode_rate": 44100,
+        "audio_rate": 16000,
+        "awake_delay": 1000
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/mode_melotts-en-us.json b/projects/llm_framework/main_melotts/mode_melotts-en-us.json
index d6320873..52eed57e 100644
--- a/projects/llm_framework/main_melotts/mode_melotts-en-us.json
+++ b/projects/llm_framework/main_melotts/mode_melotts-en-us.json
@@ -2,8 +2,8 @@
     "mode": "melotts-en-us",
     "type": "tts",
     "homepage": "https://huggingface.co/myshell-ai/MeloTTS-English",
-    "compile_flage": "pulsar2 build --input decoder-en.onnx --config config_decoder_u16.json --output_dir decoder-en --output_name decoder-en.axmodel --target_hardware AX620E --npu_mode NPU2 --compiler.check 0",
-    "pulsar_version": "3.4-3dfd5692",
+    "compile_flage": "pulsar2 build --input decoder-en.onnx --config config_decoder_u16.json --output_dir decoder_en-us --output_name decoder-en-us.axmodel --target_hardware AX620E --npu_mode NPU2 --compiler.check 0",
+    "pulsar_version": "4.0-64a0e58f",
     "capabilities": [
         "tts",
         "English"
@@ -16,14 +16,14 @@
         "sys.play.0_1"
     ],
     "mode_param": {
-        "encoder": "encoder-en.ort",
-        "decoder": "decoder-en.axmodel",
-        "gbin": "g-en.bin",
-        "tokens": "tokens.txt",
-        "lexicon": "lexicon.txt",
+        "encoder": "encoder-en-us.ort",
+        "decoder": "decoder-en-us.axmodel",
+        "gbin": "g-en-us.bin",
+        "tokens": "tokens-en.txt",
+        "lexicon": "lexicon-en.txt",
         "tagger": "en_tn_tagger.fst",
         "verbalizer": "en_tn_verbalizer.fst",
-        "spacker_speed": 1.0,
+        "spacker_speed": 1.2,
         "mode_rate": 44100,
         "audio_rate": 16000,
         "awake_delay": 1000
diff --git a/projects/llm_framework/main_melotts/src/main.cpp b/projects/llm_framework/main_melotts/src/main.cpp
index 0875be97..f80d9074 100644
--- a/projects/llm_framework/main_melotts/src/main.cpp
+++ b/projects/llm_framework/main_melotts/src/main.cpp
@@ -190,7 +190,7 @@ class llm_task {
             g_matrix.resize(256, 0);
             FILE *fp = fopen(mode_config_.gbin.c_str(), "rb");
             if (!fp) {
-                printf("Open %s failed!\n", mode_config_.gbin.c_str());
+                SLOGE("Open %s failed!", mode_config_.gbin.c_str());
                 return -3;
             }
             fread(g_matrix.data(), sizeof(float), g_matrix.size(), fp);
@@ -198,11 +198,11 @@ class llm_task {
             encoder_ = std::make_unique<OnnxWrapper>();
             decoder_ = std::make_unique<EngineWrapper>();
             if (0 != encoder_->Init(mode_config_.encoder)) {
-                printf("encoder init failed!\n");
+                SLOGE("encoder init failed!");
                 return -4;
             }
             if (0 != decoder_->Init(mode_config_.decoder.c_str())) {
-                printf("Init decoder model failed!\n");
+                SLOGE("Init decoder model failed!");
                 return -5;
             }
         } catch (...) {
@@ -398,7 +398,6 @@ class llm_task {
                         }
                     }
 
-
                     int aligned_start = audio_start + best_offset;
 
                     std::vector<float> crossfade_region(sola_buffer_frame);
@@ -457,7 +456,6 @@ class llm_task {
                 pcmlist.resize(audio_len);
             }
 
-
             double src_ratio =
                 static_cast<double>(mode_config_.audio_rate) / static_cast<double>(mode_config_.mode_rate);
             std::vector<float> tmp_pcm((pcmlist.size() * src_ratio + 1));
@@ -465,7 +463,6 @@ class llm_task {
 
             resample_audio(pcmlist.data(), pcmlist.size(), tmp_pcm.data(), &len, src_ratio);
 
-
             wav_pcm_data.reserve(len);
             std::transform(tmp_pcm.begin(), tmp_pcm.begin() + len, std::back_inserter(wav_pcm_data),
                            [](const auto val) { return static_cast<int16_t>(val * INT16_MAX); });
diff --git a/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp b/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp
index 134f64c4..de28da8d 100644
--- a/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp
+++ b/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp
@@ -10,15 +10,6 @@
 #include "../../../../../SDK/components/utilities/include/sample_log.h"
 #include "processor/wetext_processor.h"
 
-// Debug logging switch - set to true to enable debug logs
-static bool DEBUG_LOGGING = false;
-// Macro for debug logging
-#define DEBUG_LOG(fmt, ...)            \
-    do {                               \
-        if (DEBUG_LOGGING) {           \
-            SLOGI(fmt, ##__VA_ARGS__); \
-        }                              \
-    } while (0)
 std::vector<std::string> split(const std::string& s, char delim)
 {
     std::vector<std::string> result;
@@ -31,6 +22,7 @@ std::vector<std::string> split(const std::string& s, char delim)
     }
     return result;
 }
+
 class Lexicon {
 private:
     std::unordered_map<std::string, std::pair<std::vector<int>, std::vector<int>>> lexicon;
@@ -41,18 +33,12 @@ class Lexicon {
     wetext::Processor* m_processor;
 
 public:
-    // Setter for debug logging
-    static void setDebugLogging(bool enable)
-    {
-        DEBUG_LOGGING = enable;
-    }
     Lexicon(const std::string& lexicon_filename, const std::string& tokens_filename, const std::string& tagger_filename,
             const std::string& verbalizer_filename)
         : max_phrase_length(0)
     {
-        DEBUG_LOG("Dictionary loading: %s Pronunciation table loading: %s tagger_filename: %s verbalizer_filename: %s",
-                  tokens_filename.c_str(), lexicon_filename.c_str(), tagger_filename.c_str(),
-                  verbalizer_filename.c_str());
+        SLOGD("Dictionary loading: %s Pronunciation table loading: %s tagger_filename: %s verbalizer_filename: %s",
+              tokens_filename.c_str(), lexicon_filename.c_str(), tagger_filename.c_str(), verbalizer_filename.c_str());
 
         m_processor = new wetext::Processor(tagger_filename, verbalizer_filename);
 
@@ -106,8 +92,8 @@ class Lexicon {
         lexicon["。"] = lexicon["."];
         lexicon["！"] = lexicon["!"];
         lexicon["？"] = lexicon["?"];
-        DEBUG_LOG("Dictionary loading complete, containing %zu entries, longest phrase length: %zu", lexicon.size(),
-                  max_phrase_length);
+        SLOGD("Dictionary loading complete, containing %zu entries, longest phrase length: %zu", lexicon.size(),
+              max_phrase_length);
     }
 
     std::vector<std::string> splitEachChar(const std::string& text)
@@ -136,15 +122,17 @@ class Lexicon {
     {
         return s.size() == 1 && ((s[0] >= 'A' && s[0] <= 'Z') || (s[0] >= 'a' && s[0] <= 'z'));
     }
+
     bool is_english_token_char(const std::string& s)
     {
         if (s.size() != 1) return false;
         char c = s[0];
         return (c >= 'A' && c <= 'Z') || (c >= 'a' && c <= 'z') || (c >= '0' && c <= '9') || c == '-' || c == '_';
     }
+
     void process_unknown_english(const std::string& word, std::vector<int>& phones, std::vector<int>& tones)
     {
-        DEBUG_LOG("Processing unknown term: %s", word.c_str());
+        SLOGD("Processing unknown term: %s", word.c_str());
         std::string orig_word = word;
         std::vector<std::string> parts;
         std::vector<std::string> phonetic_parts;
@@ -163,7 +151,7 @@ class Lexicon {
                     tones.insert(tones.end(), sub_tones.begin(), sub_tones.end());
                     parts.push_back(sub_word);
                     phonetic_parts.push_back(phonesToString(sub_phones));
-                    DEBUG_LOG("  Matched: '%s' -> %s", sub_word.c_str(), phonesToString(sub_phones).c_str());
+                    SLOGD("  Matched: '%s' -> %s", sub_word.c_str(), phonesToString(sub_phones).c_str());
                     start += len;
                     matched = true;
                     break;
@@ -180,13 +168,13 @@ class Lexicon {
                     tones.insert(tones.end(), char_tones.begin(), char_tones.end());
                     parts.push_back(single_char);
                     phonetic_parts.push_back(phonesToString(char_phones));
-                    DEBUG_LOG("  Single char: '%s' -> %s", single_char.c_str(), phonesToString(char_phones).c_str());
+                    SLOGD("  Single char: '%s' -> %s", single_char.c_str(), phonesToString(char_phones).c_str());
                 } else {
                     phones.insert(phones.end(), unknown_token.first.begin(), unknown_token.first.end());
                     tones.insert(tones.end(), unknown_token.second.begin(), unknown_token.second.end());
                     parts.push_back(single_char);
                     phonetic_parts.push_back("_unknown_");
-                    DEBUG_LOG("  Unknown: '%s'", single_char.c_str());
+                    SLOGD("  Unknown: '%s'", single_char.c_str());
                 }
                 start++;
             }
@@ -200,26 +188,25 @@ class Lexicon {
             parts_str += parts[i];
             phonetic_str += phonetic_parts[i];
         }
-        DEBUG_LOG("%s\t|\tDecomposed: %s\t|\tPhonetics: %s", orig_word.c_str(), parts_str.c_str(),
-                  phonetic_str.c_str());
+        SLOGD("%s\t|\tDecomposed: %s\t|\tPhonetics: %s", orig_word.c_str(), parts_str.c_str(), phonetic_str.c_str());
     }
 
     void convert(const std::string& text, std::vector<int>& phones, std::vector<int>& tones)
     {
-        DEBUG_LOG("\nStarting text processing: \"%s\"", text.c_str());
+        SLOGD("\nStarting text processing: \"%s\"", text.c_str());
 
         std::string taggedText = m_processor->Tag(text);
-        DEBUG_LOG("\taggedText processing: \"%s\"", taggedText.c_str());
+        SLOGD("\taggedText processing: \"%s\"", taggedText.c_str());
         std::string normalizedText = m_processor->Verbalize(taggedText);
-        DEBUG_LOG("\normalizedText processing: \"%s\"", normalizedText.c_str());
+        SLOGD("\normalizedText processing: \"%s\"", normalizedText.c_str());
 
-        DEBUG_LOG("=======Matching Results=======");
-        DEBUG_LOG("Unit\t|\tPhonemes\t|\tTones");
-        DEBUG_LOG("-----------------------------");
+        SLOGD("=======Matching Results=======");
+        SLOGD("Unit\t|\tPhonemes\t|\tTones");
+        SLOGD("-----------------------------");
         phones.insert(phones.end(), unknown_token.first.begin(), unknown_token.first.end());
         tones.insert(tones.end(), unknown_token.second.begin(), unknown_token.second.end());
-        DEBUG_LOG("<BOS>\t|\t%s\t|\t%s", phonesToString(unknown_token.first).c_str(),
-                  tonesToString(unknown_token.second).c_str());
+        SLOGD("<BOS>\t|\t%s\t|\t%s", phonesToString(unknown_token.first).c_str(),
+              tonesToString(unknown_token.second).c_str());
         auto chars = splitEachChar(normalizedText);
         int i      = 0;
         while (i < chars.size()) {
@@ -236,8 +223,8 @@ class Lexicon {
                     auto& [eng_phones, eng_tones] = lexicon[eng_word];
                     phones.insert(phones.end(), eng_phones.begin(), eng_phones.end());
                     tones.insert(tones.end(), eng_tones.begin(), eng_tones.end());
-                    DEBUG_LOG("%s\t|\t%s\t|\t%s", orig_word.c_str(), phonesToString(eng_phones).c_str(),
-                              tonesToString(eng_tones).c_str());
+                    SLOGD("%s\t|\t%s\t|\t%s", orig_word.c_str(), phonesToString(eng_phones).c_str(),
+                          tonesToString(eng_tones).c_str());
                 } else {
                     process_unknown_english(orig_word, phones, tones);
                 }
@@ -256,8 +243,8 @@ class Lexicon {
                     auto& [phrase_phones, phrase_tones] = lexicon[phrase];
                     phones.insert(phones.end(), phrase_phones.begin(), phrase_phones.end());
                     tones.insert(tones.end(), phrase_tones.begin(), phrase_tones.end());
-                    DEBUG_LOG("%s\t|\t%s\t|\t%s", phrase.c_str(), phonesToString(phrase_phones).c_str(),
-                              tonesToString(phrase_tones).c_str());
+                    SLOGD("%s\t|\t%s\t|\t%s", phrase.c_str(), phonesToString(phrase_phones).c_str(),
+                          tonesToString(phrase_tones).c_str());
                     i += len;
                     matched = true;
                     break;
@@ -279,25 +266,25 @@ class Lexicon {
                     auto& [char_phones, char_tones] = lexicon[s];
                     phones.insert(phones.end(), char_phones.begin(), char_phones.end());
                     tones.insert(tones.end(), char_tones.begin(), char_tones.end());
-                    DEBUG_LOG("%s\t|\t%s\t|\t%s", orig_char.c_str(), phonesToString(char_phones).c_str(),
-                              tonesToString(char_tones).c_str());
+                    SLOGD("%s\t|\t%s\t|\t%s", orig_char.c_str(), phonesToString(char_phones).c_str(),
+                          tonesToString(char_tones).c_str());
                 } else {
                     phones.insert(phones.end(), unknown_token.first.begin(), unknown_token.first.end());
                     tones.insert(tones.end(), unknown_token.second.begin(), unknown_token.second.end());
-                    DEBUG_LOG("%s\t|\t%s (Not matched)\t|\t%s", orig_char.c_str(),
-                              phonesToString(unknown_token.first).c_str(), tonesToString(unknown_token.second).c_str());
+                    SLOGD("%s\t|\t%s (Not matched)\t|\t%s", orig_char.c_str(),
+                          phonesToString(unknown_token.first).c_str(), tonesToString(unknown_token.second).c_str());
                 }
             }
         }
         phones.insert(phones.end(), unknown_token.first.begin(), unknown_token.first.end());
         tones.insert(tones.end(), unknown_token.second.begin(), unknown_token.second.end());
-        DEBUG_LOG("<EOS>\t|\t%s\t|\t%s", phonesToString(unknown_token.first).c_str(),
-                  tonesToString(unknown_token.second).c_str());
-        DEBUG_LOG("\nProcessing Summary:");
-        DEBUG_LOG("Original text: %s", text.c_str());
-        DEBUG_LOG("Phonemes: %s", phonesToString(phones).c_str());
-        DEBUG_LOG("Tones: %s", tonesToString(tones).c_str());
-        DEBUG_LOG("====================");
+        SLOGD("<EOS>\t|\t%s\t|\t%s", phonesToString(unknown_token.first).c_str(),
+              tonesToString(unknown_token.second).c_str());
+        SLOGD("\nProcessing Summary:");
+        SLOGD("Original text: %s", text.c_str());
+        SLOGD("Phonemes: %s", phonesToString(phones).c_str());
+        SLOGD("Tones: %s", tonesToString(tones).c_str());
+        SLOGD("====================");
     }
 
 private:
@@ -316,6 +303,7 @@ class Lexicon {
         phones.insert(phones.end(), phones_and_tones.first.begin(), phones_and_tones.first.end());
         tones.insert(tones.end(), phones_and_tones.second.begin(), phones_and_tones.second.end());
     }
+
     std::string phonesToString(const std::vector<int>& phones)
     {
         std::string result;
@@ -329,6 +317,7 @@ class Lexicon {
         }
         return result;
     }
+
     std::string tonesToString(const std::vector<int>& tones)
     {
         std::string result;
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index e593f873..c188bfe8 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -377,8 +377,11 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-model-single-speaker-english-fast':[create_data_deb,'llm-model-single-speaker-english-fast', '0.3', src_folder, revision],
         'llm-model-single-speaker-fast':[create_data_deb,'llm-model-single-speaker-fast', '0.3', src_folder, revision],
         'llm-model-melotts-zh-cn':[create_data_deb,'llm-model-melotts-zh-cn', '0.6', src_folder, revision],
-        'llm-model-melotts-en-us':[create_data_deb,'llm-model-melotts-en-us', '0.5', src_folder, revision],
+        'llm-model-melotts-en-au':[create_data_deb,'llm-model-melotts-en-au', '0.6', src_folder, revision],
+        'llm-model-melotts-en-br':[create_data_deb,'llm-model-melotts-en-br', '0.6', src_folder, revision],
         'llm-model-melotts-en-default':[create_data_deb,'llm-model-melotts-en-default', '0.6', src_folder, revision],
+        'llm-model-melotts-en-india':[create_data_deb,'llm-model-melotts-en-india', '0.6', src_folder, revision],
+        'llm-model-melotts-en-us':[create_data_deb,'llm-model-melotts-en-us', '0.6', src_folder, revision],
         'llm-model-melotts-ja-jp':[create_data_deb,'llm-model-melotts-ja-jp', '0.6', src_folder, revision],
         'llm-model-yolo11n':[create_data_deb,'llm-model-yolo11n', data_version, src_folder, revision],
         'llm-model-yolo11n-pose':[create_data_deb,'llm-model-yolo11n-pose', '0.3', src_folder, revision],

From 8acb179b634afc26c1a723c8dd948524918e65bb Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Tue, 20 May 2025 15:06:11 +0800
Subject: [PATCH 08/79] [fix] Handles the situation where Either tagger or
 verbalizer file does not exist.

---
 .../llm_framework/main_melotts/src/main.cpp   | 12 ++-
 .../main_melotts/src/runner/Lexicon.hpp       | 78 +++++++++++++++++--
 2 files changed, 78 insertions(+), 12 deletions(-)

diff --git a/projects/llm_framework/main_melotts/src/main.cpp b/projects/llm_framework/main_melotts/src/main.cpp
index f80d9074..6b8e1c15 100644
--- a/projects/llm_framework/main_melotts/src/main.cpp
+++ b/projects/llm_framework/main_melotts/src/main.cpp
@@ -183,10 +183,14 @@ class llm_task {
                 awake_delay_ = config_body["awake_delay"].get<int>();
             else if (file_body["mode_param"].contains("awake_delay"))
                 awake_delay_ = file_body["mode_param"]["awake_delay"];
-            // Load lexicon
-            lexicon_ = std::make_unique<Lexicon>(mode_config_.lexicon, mode_config_.tokens, mode_config_.tagger,
-                                                 mode_config_.verbalizer);
-            // Read g.bin
+
+            if (!std::filesystem::exists(mode_config_.tagger) || !std::filesystem::exists(mode_config_.verbalizer)) {
+                SLOGW("Either tagger or verbalizer file does not exist, using alternative lexicon.");
+                lexicon_ = std::make_unique<Lexicon>(mode_config_.lexicon, mode_config_.tokens);
+            } else {
+                lexicon_ = std::make_unique<Lexicon>(mode_config_.lexicon, mode_config_.tokens, mode_config_.tagger,
+                                                     mode_config_.verbalizer);
+            }
             g_matrix.resize(256, 0);
             FILE *fp = fopen(mode_config_.gbin.c_str(), "rb");
             if (!fp) {
diff --git a/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp b/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp
index de28da8d..1884c929 100644
--- a/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp
+++ b/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp
@@ -30,7 +30,7 @@ class Lexicon {
     std::pair<std::vector<int>, std::vector<int>> unknown_token;
     std::unordered_map<int, std::string> reverse_tokens;
 
-    wetext::Processor* m_processor;
+    wetext::Processor* m_processor = nullptr;
 
 public:
     Lexicon(const std::string& lexicon_filename, const std::string& tokens_filename, const std::string& tagger_filename,
@@ -96,6 +96,65 @@ class Lexicon {
               max_phrase_length);
     }
 
+    Lexicon(const std::string& lexicon_filename, const std::string& tokens_filename) : max_phrase_length(0)
+    {
+        SLOGD("Dictionary loading: %s Pronunciation table loading: %s", tokens_filename.c_str(),
+              lexicon_filename.c_str());
+
+        std::unordered_map<std::string, int> tokens;
+        std::ifstream ifs(tokens_filename);
+        assert(ifs.is_open());
+        std::string line;
+        while (std::getline(ifs, line)) {
+            auto splitted_line = split(line, ' ');
+            if (splitted_line.size() >= 2) {
+                int token_id = std::stoi(splitted_line[1]);
+                tokens.insert({splitted_line[0], token_id});
+                reverse_tokens[token_id] = splitted_line[0];
+            }
+        }
+        ifs.close();
+        ifs.open(lexicon_filename);
+        assert(ifs.is_open());
+        while (std::getline(ifs, line)) {
+            auto splitted_line = split(line, ' ');
+            if (splitted_line.empty()) continue;
+            std::string word_or_phrase = splitted_line[0];
+            auto chars                 = splitEachChar(word_or_phrase);
+            max_phrase_length          = std::max(max_phrase_length, chars.size());
+            size_t phone_tone_len      = splitted_line.size() - 1;
+            size_t half_len            = phone_tone_len / 2;
+            std::vector<int> phones, tones;
+            for (size_t i = 0; i < phone_tone_len; i++) {
+                auto phone_or_tone = splitted_line[i + 1];
+                if (i < half_len) {
+                    if (tokens.find(phone_or_tone) != tokens.end()) {
+                        phones.push_back(tokens[phone_or_tone]);
+                    }
+                } else {
+                    tones.push_back(std::stoi(phone_or_tone));
+                }
+            }
+            lexicon[word_or_phrase] = std::make_pair(phones, tones);
+        }
+        const std::vector<std::string> punctuation{"!", "?", "…", ",", ".", "'", "-"};
+        for (const auto& p : punctuation) {
+            if (tokens.find(p) != tokens.end()) {
+                int i      = tokens[p];
+                lexicon[p] = std::make_pair(std::vector<int>{i}, std::vector<int>{0});
+            }
+        }
+        assert(tokens.find("_") != tokens.end());
+        unknown_token = std::make_pair(std::vector<int>{tokens["_"]}, std::vector<int>{0});
+        lexicon[" "]  = unknown_token;
+        lexicon["，"] = lexicon[","];
+        lexicon["。"] = lexicon["."];
+        lexicon["！"] = lexicon["!"];
+        lexicon["？"] = lexicon["?"];
+        SLOGD("Dictionary loading complete, containing %zu entries, longest phrase length: %zu", lexicon.size(),
+              max_phrase_length);
+    }
+
     std::vector<std::string> splitEachChar(const std::string& text)
     {
         std::vector<std::string> words;
@@ -195,14 +254,17 @@ class Lexicon {
     {
         SLOGD("\nStarting text processing: \"%s\"", text.c_str());
 
-        std::string taggedText = m_processor->Tag(text);
-        SLOGD("\taggedText processing: \"%s\"", taggedText.c_str());
-        std::string normalizedText = m_processor->Verbalize(taggedText);
-        SLOGD("\normalizedText processing: \"%s\"", normalizedText.c_str());
+        std::string normalizedText;
+        if (m_processor) {
+            std::string taggedText = m_processor->Tag(text);
+            SLOGD("\taggedText processing: \"%s\"", taggedText.c_str());
+            normalizedText = m_processor->Verbalize(taggedText);
+            SLOGD("\tnormalizedText processing: \"%s\"", normalizedText.c_str());
+        } else {
+            SLOGD("m_processor is not initialized, skipping tag and verbalize steps.");
+            normalizedText = text;
+        }
 
-        SLOGD("=======Matching Results=======");
-        SLOGD("Unit\t|\tPhonemes\t|\tTones");
-        SLOGD("-----------------------------");
         phones.insert(phones.end(), unknown_token.first.begin(), unknown_token.first.end());
         tones.insert(tones.end(), unknown_token.second.begin(), unknown_token.second.end());
         SLOGD("<BOS>\t|\t%s\t|\t%s", phonesToString(unknown_token.first).c_str(),

From f67506c66238d74278c38cc1bc61961207f29211 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Tue, 20 May 2025 16:16:32 +0800
Subject: [PATCH 09/79] [update] update melotts-es-es model

---
 .../main_melotts/mode_melotts-es-es.json      | 29 +++++++++++++++++++
 projects/llm_framework/tools/llm_pack.py      |  1 +
 2 files changed, 30 insertions(+)
 create mode 100644 projects/llm_framework/main_melotts/mode_melotts-es-es.json

diff --git a/projects/llm_framework/main_melotts/mode_melotts-es-es.json b/projects/llm_framework/main_melotts/mode_melotts-es-es.json
new file mode 100644
index 00000000..4a65a57e
--- /dev/null
+++ b/projects/llm_framework/main_melotts/mode_melotts-es-es.json
@@ -0,0 +1,29 @@
+{
+    "mode": "melotts-es-es",
+    "type": "tts",
+    "homepage": "https://github.com/ml-inory/melotts.axera/tree/main/model_convert",
+    "compile_flage": "pulsar2 build --input decoder-es.onnx --config config_decoder_u16.json --output_dir decoder_es --output_name decoder-es.axmodel --target_hardware AX620E --npu_mode NPU2 --compiler.check 0",
+    "pulsar_version": "4.0-64a0e58f",
+    "capabilities": [
+        "tts",
+        "Spanish"
+    ],
+    "input_type": [
+        "tts.utf-8"
+    ],
+    "output_type": [
+        "tts.wav",
+        "sys.play.0_1"
+    ],
+    "mode_param": {
+        "encoder": "encoder-es.ort",
+        "decoder": "decoder-es.axmodel",
+        "gbin": "g-es.bin",
+        "tokens": "tokens-es.txt",
+        "lexicon": "lexicon-es.txt",
+        "spacker_speed": 1.2,
+        "mode_rate": 44100,
+        "audio_rate": 16000,
+        "awake_delay": 1000
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index c188bfe8..b27cd597 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -383,6 +383,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-model-melotts-en-india':[create_data_deb,'llm-model-melotts-en-india', '0.6', src_folder, revision],
         'llm-model-melotts-en-us':[create_data_deb,'llm-model-melotts-en-us', '0.6', src_folder, revision],
         'llm-model-melotts-ja-jp':[create_data_deb,'llm-model-melotts-ja-jp', '0.6', src_folder, revision],
+        'llm-model-melotts-es-es':[create_data_deb,'llm-model-melotts-es-es', '0.5', src_folder, revision],
         'llm-model-yolo11n':[create_data_deb,'llm-model-yolo11n', data_version, src_folder, revision],
         'llm-model-yolo11n-pose':[create_data_deb,'llm-model-yolo11n-pose', '0.3', src_folder, revision],
         'llm-model-yolo11n-hand-pose':[create_data_deb,'llm-model-yolo11n-hand-pose', '0.3', src_folder, revision],

From 764bca125836471f12cd5f274a668066e6bafbac Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Tue, 20 May 2025 17:57:37 +0800
Subject: [PATCH 10/79] [update] update model list

---
 README_zh.md | 37 +++++++++++++++++++++++++++++++++++++
 1 file changed, 37 insertions(+)

diff --git a/README_zh.md b/README_zh.md
index 4216b0aa..e80293e6 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -58,7 +58,44 @@ StackFlow 语音助手的主要工作模式：
 ## 模型列表
 | 模型名 | 模型类型 | 模型大小 | 模型能力 | 模型配置文件 | 计算单元 |
 | :----: | :----: | :----: | :----: | :----: | :----: |
+| [silero-vad](https://github.com/snakers4/silero-vad) | VAD | 3.3M | 语音活动检测 | [mode_silero-vad.json](projects/llm_framework/main_vad/mode_silero-vad.json) | CPU |
 | [sherpa-onnx-kws-zipformer-gigaspeech-3.3M-2024-01-01](https://github.com/k2-fsa/sherpa-onnx/releases/download/kws-models/sherpa-onnx-kws-zipformer-gigaspeech-3.3M-2024-01-01.tar.bz2) | KWS | 6.4M | 关键词识别 | [mode_sherpa-onnx-kws-zipformer-gigaspeech-3.3M-2024-01-01.json](projects/llm_framework/main_kws/mode_sherpa-onnx-kws-zipformer-gigaspeech-3.3M-2024-01-01.json) | CPU |
+| [sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01](https://github.com/k2-fsa/sherpa-onnx/releases/download/kws-models/sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01.tar.bz2) | KWS | 5.7M | 关键词识别 | [mode_sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01.json](projects/llm_framework/main_kws/mode_sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01.json) | CPU |
+| [sherpa-ncnn-streaming-zipformer-20M-2023-02-17](https://huggingface.co/desh2608/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-small) | ASR | 40M | 语音识别 | [mode_sherpa-ncnn-streaming-zipformer-20M-2023-02-17.json](projects/llm_framework/main_asr/mode_sherpa-ncnn-streaming-zipformer-20M-2023-02-17.json) | CPU |
+| [sherpa-ncnn-streaming-zipformer-zh-14M-2023-02-23](https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/pruned_transducer_stateless7_streaming) | ASR | 24M | 语音识别 | [mode_sherpa-ncnn-streaming-zipformer-zh-14M-2023-02-23.json](projects/llm_framework/main_asr/mode_sherpa-ncnn-streaming-zipformer-zh-14M-2023-02-23.json) | CPU |
+| [whisper-tiny](https://huggingface.co/openai/whisper-tiny) | ASR | 201M | 语音识别 | [mode_whisper-tiny.json](projects/llm_framework/main_whisper/mode_whisper-tiny.json) | NPU |
+| [whisper-base](https://huggingface.co/openai/whisper-base) | ASR | 309M | 语音识别 | [mode_whisper-base.json](projects/llm_framework/main_whisper/mode_whisper-base.json) | NPU |
+| [whisper-small](https://huggingface.co/openai/whisper-small) | ASR | 725M | 语音识别 | [mode_whisper-small.json](projects/llm_framework/main_whisper/mode_whisper-small.json) | NPU |
+| [single-speaker-fast](https://github.com/huakunyang/SummerTTS) | TTS | 77M | 语音生成 | [mode_whisper-tiny.json](projects/llm_framework/main_tts/mode_single-speaker-fast.json) | CPU |
+| [single-speaker-english-fast](https://github.com/huakunyang/SummerTTS) | TTS | 60M | 语音生成 | [mode_whisper-tiny.json](projects/llm_framework/main_tts/mode_single-speaker-english-fast.json) | CPU |
+| [melotts-en-au](https://huggingface.co/myshell-ai/MeloTTS-English) | TTS | 102M | 语音生成 | [mode_melotts-en-au.json](projects/llm_framework/main_melotts/mode_melotts-en-au.json) | NPU |
+| [melotts-en-br](https://huggingface.co/myshell-ai/MeloTTS-English) | TTS | 102M | 语音生成 | [mode_melotts-en-au.json](projects/llm_framework/main_melotts/mode_melotts-en-br.json) | NPU |
+| [melotts-en-default](https://huggingface.co/myshell-ai/MeloTTS-English) | TTS | 102M | 语音生成 | [mode_melotts-en-india.json](projects/llm_framework/main_melotts/mode_melotts-en-default.json) | NPU |
+| [melotts-en-us](https://huggingface.co/myshell-ai/MeloTTS-English) | TTS | 102M | 语音生成 | [mode_melotts-en-au.json](projects/llm_framework/main_melotts/mode_melotts-en-us.json) | NPU |
+| [melotts-es-es](https://huggingface.co/myshell-ai/MeloTTS-Spanish) | TTS | 83M | 语音生成 | [mode_melotts-es-es.json](projects/llm_framework/main_melotts/mode_melotts-es-es.json) | NPU |
+| [melotts-ja-jp](https://huggingface.co/myshell-ai/MeloTTS-Japanese) | TTS | 83M | 语音生成 | [mode_melotts-ja-jp.json](projects/llm_framework/main_melotts/mode_melotts-ja-jp.json) | NPU |
+| [melotts-zh-cn](https://huggingface.co/myshell-ai/MeloTTS-Chinese) | TTS | 86M | 语音生成 | [mode_melotts-zh-cn.json](projects/llm_framework/main_melotts/mode_melotts-zh-cn.json) | NPU |
+| [deepseek-r1-1.5B-ax630c](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B) | LLM | 2.0G | 文本生成 | [mode_deepseek-r1-1.5B-ax630c.json](projects/llm_framework/main_llm/models/mode_deepseek-r1-1.5B-ax630c.json) | NPU |
+| [deepseek-r1-1.5B-p256-ax630c](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B) | LLM | 2.0G | 文本生成 | [mode_deepseek-r1-1.5B-p256-ax630c.json](projects/llm_framework/main_llm/models/mode_deepseek-r1-1.5B-p256-ax630c.json) | NPU |
+| [llama3.2-1B-p256-ax630c](https://huggingface.co/meta-llama/Llama-3.2-1B) | LLM | 1.7G | 文本生成 | [mode_llama3.2-1B-p256-ax630c.json](projects/llm_framework/main_llm/models/mode_llama3.2-1B-p256-ax630c.json) | NPU |
+| [llama3.2-1B-prefill-ax630c](https://huggingface.co/meta-llama/Llama-3.2-1B) | LLM | 1.7G | 文本生成 | [mode_llama3.2-1B-prefill-ax630c.json](projects/llm_framework/main_llm/models/mode_llama3.2-1B-prefill-ax630c.json) | NPU |
+| [openbuddy-llama3.2-1B-ax630c](https://huggingface.co/OpenBuddy/openbuddy-llama3.2-1b-v23.1-131k) | LLM | 1.7G | 文本生成 | [mode_openbuddy-llama3.2-1B-ax630c.json](projects/llm_framework/main_llm/models/mode_openbuddy-llama3.2-1B-ax630c.json) | NPU |
+| [qwen2.5-0.5B-Int4-ax630c](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct-GPTQ-Int4) | LLM | 626M | 文本生成 | [mode_qwen2.5-0.5B-Int4-ax630c.json](projects/llm_framework/main_llm/models/mode_qwen2.5-0.5B-Int4-ax630c.json) | NPU |
+| [qwen2.5-0.5B-p256-ax630c](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) | LLM | 760M | 文本生成 | [mode_qwen2.5-0.5B-p256-ax630c.json](projects/llm_framework/main_llm/models/mode_qwen2.5-0.5B-p256-ax630c.json) | NPU |
+| [qwen2.5-0.5B-prefill-20e](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) | LLM | 758M | 文本生成 | [mode_qwen2.5-0.5B-prefill-20e.json](projects/llm_framework/main_llm/models/mode_qwen2.5-0.5B-prefill-20e.json) | NPU |
+| [qwen2.5-1.5B-Int4-ax630c](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GPTQ-Int4) | LLM | 1.5G | 文本生成 | [mode_qwen2.5-1.5B-Int4-ax630c.json](projects/llm_framework/main_llm/models/mode_qwen2.5-1.5B-Int4-ax630c.json) | NPU |
+| [qwen2.5-1.5B-p256-ax630c](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) | LLM | 2.0G | 文本生成 | [mode_qwen2.5-1.5B-p256-ax630c.json](projects/llm_framework/main_llm/models/mode_qwen2.5-1.5B-p256-ax630c.json) | NPU |
+| [qwen2.5-1.5B-ax630c](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) | LLM | 2.0G | 文本生成 | [mode_qwen2.5-1.5B-ax630c.json](projects/llm_framework/main_llm/models/mode_qwen2.5-1.5B-ax630c.json) | NPU |
+| [qwen2.5-coder-0.5B-ax630c](https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B-Instruct) | LLM | 756M | 文本生成 | [mode_qwen2.5-coder-0.5B-ax630c.json](projects/llm_framework/main_llm/models/mode_qwen2.5-coder-0.5B-ax630c.json) | NPU |
+| [qwen3-0.6B-ax630c](https://huggingface.co/AXERA-TECH/InternVL2_5-1B) | LLM | 917M | 文本生成 | [mode_qwen3-0.6B-ax630c.json](projects/llm_framework/main_llm/models/mode_qwen3-0.6B-ax630c.json) | NPU |
+| [mode_internvl2.5-1B-364-ax630c](https://huggingface.co/Qwen/Qwen3-0.6B) | VLM | 1.2G | 多模态文本生成 | [mode_internvl2.5-1B-364-ax630c.json](projects/llm_framework/main_vlm/models/mode_internvl2.5-1B-364-ax630c.json) | NPU |
+| [smolvlm-256M-ax630c](https://huggingface.co/HuggingFaceTB/SmolVLM-256M-Instruct) | VLM | 330M | 多模态文本生成 | [mode_smolvlm-256M-ax630c.json](projects/llm_framework/main_vlm/models/mode_smolvlm-256M-ax630c.json) | NPU |
+| [smolvlm-500M-ax630c](https://huggingface.co/HuggingFaceTB/SmolVLM-500M-Instruct) | VLM | 605M | 多模态文本生成 | [mode_smolvlm-256M-ax630c.json](projects/llm_framework/main_vlm/models/mode_smolvlm-500M-ax630c.json) | NPU |
+| [yolo11n](https://github.com/ultralytics/ultralytics) | CV | 2.8M | 目标检测 | [mode_yolo11n.json](projects/llm_framework/main_yolo/mode_yolo11n.json) | NPU |
+| [yolo11n-seg](https://github.com/ultralytics/ultralytics) | CV | 3.0M | 实例分割 | [mode_yolo11n-seg.json](projects/llm_framework/main_yolo/mode_yolo11n-seg.json) | NPU |
+| [yolo11n-pose](https://github.com/ultralytics/ultralytics) | CV | 3.1M | 姿态检测 | [mode_yolo11n-pose.json](projects/llm_framework/main_yolo/mode_yolo11n-pose.json) | NPU |
+| [yolo11n-hand-pose](https://github.com/ultralytics/ultralytics) | CV | 3.2M | 姿态检测 | [mode_yolo11n-hand-pose.json](projects/llm_framework/main_yolo/mode_yolo11n-hand-pose.json) | NPU |
+| [depth-anything-ax630c](https://github.com/DepthAnything/Depth-Anything-V2) | CV | 29M | 单目深度估计 | [mode_depth-anything-ax630c.json](projects/llm_framework/main_depth_anything/mode_depth-anything-ax630c.json) | NPU |
 
 ## 环境要求 ##
 当前 StackFlow 的 AI 单元是建立在 AXERA 加速平台之上的，主要的芯片平台为 ax630c、ax650n。系统要求为 ubuntu。

From aa10381ad85c4b6c8b98152cecc4465529a05074 Mon Sep 17 00:00:00 2001
From: nyasu3w <125797829+nyasu3w@users.noreply.github.com>
Date: Tue, 3 Jun 2025 23:26:19 +0900
Subject: [PATCH 11/79] add trigger method to llm_kws

---
 projects/llm_framework/main_kws/src/main.cpp | 42 ++++++++++++++++++++
 1 file changed, 42 insertions(+)

diff --git a/projects/llm_framework/main_kws/src/main.cpp b/projects/llm_framework/main_kws/src/main.cpp
index 9c05ce82..38e030c7 100644
--- a/projects/llm_framework/main_kws/src/main.cpp
+++ b/projects/llm_framework/main_kws/src/main.cpp
@@ -247,6 +247,13 @@ class llm_task {
         }
     }
 
+    void trigger()
+    {
+        if (out_callback_) {
+            out_callback_("True");
+        }
+    }
+
     bool delete_model()
     {
         spotter_.reset();
@@ -284,6 +291,9 @@ class llm_kws : public StackFlow {
     llm_kws() : StackFlow("kws")
     {
         task_count_ = 1;
+        rpc_ctx_->register_rpc_action("trigger",
+                       std::bind(&llm_kws::trigger, this, std::placeholders::_1, std::placeholders::_2));
+
     }
 
     void play_awake_wav(const std::string &wav_file)
@@ -529,6 +539,38 @@ class llm_kws : public StackFlow {
         return 0;
     }
 
+    std::string  trigger(pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data>& rawdata0)
+    {
+        const std::string rawdata = rawdata0->string();
+        int pos = rawdata.find("{");
+//        SLOGI("llm_kws::trigger:json:%s", rawdata.substr(pos).c_str());
+
+        nlohmann::json error_body;
+        nlohmann::json data;
+        try {
+            data = nlohmann::json::parse(rawdata.substr(pos));
+        } catch (...) {
+            SLOGE("setup json format error.");
+            error_body["code"]    = -2;
+            error_body["message"] = "json format error.";
+            send("None", "None", error_body, "kws");
+            return LLM_NONE;
+        }
+        auto work_id = data["work_id"].get<std::string>();
+
+        int work_id_num = sample_get_work_id_num(work_id);
+        if (llm_task_.find(work_id_num) == llm_task_.end()) {
+            error_body["code"]    = -6;
+            error_body["message"] = "Unit Does Not Exist";
+            send("None", "None", error_body, work_id);
+            return LLM_NONE;
+        }
+
+        llm_task_[work_id_num]->trigger();
+        return LLM_NONE;
+    }
+
+
     ~llm_kws()
     {
         while (1) {

From b9401b2817a1874c79cc012302ad4b362e90aea2 Mon Sep 17 00:00:00 2001
From: dianjixz <18637716021@163.com>
Date: Thu, 5 Jun 2025 19:28:50 +0800
Subject: [PATCH 12/79] [update] llm trigger Standardization.

---
 projects/llm_framework/main_kws/src/main.cpp | 88 ++++++++++++--------
 1 file changed, 55 insertions(+), 33 deletions(-)

diff --git a/projects/llm_framework/main_kws/src/main.cpp b/projects/llm_framework/main_kws/src/main.cpp
index 38e030c7..4ea6abcd 100644
--- a/projects/llm_framework/main_kws/src/main.cpp
+++ b/projects/llm_framework/main_kws/src/main.cpp
@@ -283,6 +283,7 @@ class llm_task {
 
 class llm_kws : public StackFlow {
 private:
+    enum { EVENT_TRIGGER = EVENT_EXPORT + 1 };
     int task_count_;
     std::string audio_url_;
     std::unordered_map<int, std::shared_ptr<llm_task>> llm_task_;
@@ -291,9 +292,13 @@ class llm_kws : public StackFlow {
     llm_kws() : StackFlow("kws")
     {
         task_count_ = 1;
-        rpc_ctx_->register_rpc_action("trigger",
-                       std::bind(&llm_kws::trigger, this, std::placeholders::_1, std::placeholders::_2));
-
+        event_queue_.appendListener(EVENT_TRIGGER, std::bind(&llm_kws::trigger, this, std::placeholders::_1));
+        rpc_ctx_->register_rpc_action(
+            "trigger", [this](pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) -> std::string {
+                this->event_queue_.enqueue(EVENT_TRIGGER,
+                                           std::make_shared<stackflow_data>(data->get_param(0), data->get_param(1)));
+                return LLM_NONE;
+            });
     }
 
     void play_awake_wav(const std::string &wav_file)
@@ -440,9 +445,11 @@ class llm_kws : public StackFlow {
             return -1;
         }
 
-        int work_id_num   = sample_get_work_id_num(work_id);
-        auto llm_channel  = get_channel(work_id);
-        auto llm_task_obj = std::make_shared<llm_task>(work_id);
+        int work_id_num                             = sample_get_work_id_num(work_id);
+        auto llm_channel                            = get_channel(work_id);
+        auto llm_task_obj                           = std::make_shared<llm_task>(work_id);
+        std::weak_ptr<llm_task> _llm_task_obj       = llm_task_obj;
+        std::weak_ptr<llm_channel_obj> _llm_channel = llm_channel;
         nlohmann::json config_body;
         try {
             config_body = nlohmann::json::parse(data);
@@ -458,17 +465,22 @@ class llm_kws : public StackFlow {
             llm_channel->set_output(llm_task_obj->enoutput_);
             llm_channel->set_stream(llm_task_obj->enstream_);
             llm_task_obj->play_awake_wav = std::bind(&llm_kws::play_awake_wav, this, std::placeholders::_1);
-            llm_task_obj->set_output([llm_task_obj, llm_channel](const std::string &data) {
-                llm_channel->send(llm_task_obj->response_format_, true, LLM_NO_ERROR);
+            llm_task_obj->set_output([_llm_task_obj, _llm_channel](const std::string &data) {
+                auto llm_task_obj = _llm_task_obj.lock();
+                auto llm_channel  = _llm_channel.lock();
+                if (llm_task_obj && llm_channel) {
+                    llm_channel->send(llm_task_obj->response_format_, true, LLM_NO_ERROR);
+                }
             });
 
             for (const auto input : llm_task_obj->inputs_) {
                 if (input.find("sys") != std::string::npos) {
-                    audio_url_                            = unit_call("audio", "cap", "None");
-                    std::weak_ptr<llm_task> _llm_task_obj = llm_task_obj;
-                    llm_channel->subscriber(audio_url_, [_llm_task_obj](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
-                        _llm_task_obj.lock()->sys_pcm_on_data(raw->string());
-                    });
+                    audio_url_ = unit_call("audio", "cap", "None");
+                    llm_channel->subscriber(audio_url_,
+                                            [_llm_task_obj](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
+                                                auto llm_task_obj = _llm_task_obj.lock();
+                                                if (llm_task_obj) llm_task_obj->sys_pcm_on_data(raw->string());
+                                            });
                     llm_task_obj->audio_flage_ = true;
                 } else if (input.find("kws") != std::string::npos) {
                     llm_channel->subscriber_work_id(
@@ -539,38 +551,48 @@ class llm_kws : public StackFlow {
         return 0;
     }
 
-    std::string  trigger(pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data>& rawdata0)
+    std::string trigger(const std::shared_ptr<void> &arg)
     {
-        const std::string rawdata = rawdata0->string();
-        int pos = rawdata.find("{");
-//        SLOGI("llm_kws::trigger:json:%s", rawdata.substr(pos).c_str());
-
-        nlohmann::json error_body;
-        nlohmann::json data;
-        try {
-            data = nlohmann::json::parse(rawdata.substr(pos));
-        } catch (...) {
-            SLOGE("setup json format error.");
-            error_body["code"]    = -2;
-            error_body["message"] = "json format error.";
-            send("None", "None", error_body, "kws");
+        std::shared_ptr<stackflow_data> originalPtr = std::static_pointer_cast<stackflow_data>(arg);
+        std::string zmq_url                         = originalPtr->string(0);
+        std::string data                            = originalPtr->string(1);
+        std::string work_id                         = sample_json_str_get(data, "work_id");
+        if (work_id.length() == 0) {
+            nlohmann::json out_body;
+            out_body["request_id"]       = sample_json_str_get(data, "request_id");
+            out_body["work_id"]          = "kws";
+            out_body["created"]          = time(NULL);
+            out_body["object"]           = "";
+            out_body["data"]             = "";
+            out_body["error"]["code"]    = -2;
+            out_body["error"]["message"] = "json format error.";
+            pzmq _zmq(zmq_url, ZMQ_PUSH);
+            std::string out = out_body.dump();
+            out += "\n";
+            _zmq.send_data(out);
             return LLM_NONE;
         }
-        auto work_id = data["work_id"].get<std::string>();
 
         int work_id_num = sample_get_work_id_num(work_id);
         if (llm_task_.find(work_id_num) == llm_task_.end()) {
-            error_body["code"]    = -6;
-            error_body["message"] = "Unit Does Not Exist";
-            send("None", "None", error_body, work_id);
+            nlohmann::json out_body;
+            out_body["request_id"]       = sample_json_str_get(data, "request_id");
+            out_body["work_id"]          = "kws";
+            out_body["created"]          = time(NULL);
+            out_body["object"]           = "";
+            out_body["data"]             = "";
+            out_body["error"]["code"]    = -6;
+            out_body["error"]["message"] = "Unit Does Not Exist";
+            pzmq _zmq(zmq_url, ZMQ_PUSH);
+            std::string out = out_body.dump();
+            out += "\n";
+            _zmq.send_data(out);
             return LLM_NONE;
         }
-
         llm_task_[work_id_num]->trigger();
         return LLM_NONE;
     }
 
-
     ~llm_kws()
     {
         while (1) {

From 61c69a33623081885e6fb990120f03a4c86d3c46 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Tue, 10 Jun 2025 09:21:08 +0800
Subject: [PATCH 13/79] [update] update docs

---
 projects/llm_framework/README.md                     | 2 +-
 projects/llm_framework/tools/test_tools/test-melo.py | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/projects/llm_framework/README.md b/projects/llm_framework/README.md
index e99981c3..10699700 100644
--- a/projects/llm_framework/README.md
+++ b/projects/llm_framework/README.md
@@ -98,7 +98,7 @@ send :
     "action": "setup",
     "object": "melotts.setup",
     "data": {
-        "model": "melotts_zh-cn",
+        "model": "melotts-zh-cn",
         "response_format": "sys.pcm",
         "input": "tts.utf-8",
         "enoutput": false
diff --git a/projects/llm_framework/tools/test_tools/test-melo.py b/projects/llm_framework/tools/test_tools/test-melo.py
index b257f2fc..02135bfa 100644
--- a/projects/llm_framework/tools/test_tools/test-melo.py
+++ b/projects/llm_framework/tools/test_tools/test-melo.py
@@ -45,7 +45,7 @@ def create_melotts_setup_data(request_id="melotts_setup"):
         "action": "setup",
         "object": "melotts.setup",
         "data": {
-            "model": "melotts_zh-cn",
+            "model": "melotts-zh-cn",
             "response_format": "sys.pcm",
             "input": "tts.utf-8",
             "enoutput": False

From 2d0cd697a9005e7075109b0a6c54475c6cbbd152 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Wed, 18 Jun 2025 11:41:23 +0800
Subject: [PATCH 14/79] [update] vlm add task_camera_data

---
 projects/llm_framework/main_vlm/src/main.cpp | 118 ++++++++++++++++---
 1 file changed, 102 insertions(+), 16 deletions(-)

diff --git a/projects/llm_framework/main_vlm/src/main.cpp b/projects/llm_framework/main_vlm/src/main.cpp
index b625b5f9..e7aeb36f 100644
--- a/projects/llm_framework/main_vlm/src/main.cpp
+++ b/projects/llm_framework/main_vlm/src/main.cpp
@@ -15,19 +15,25 @@
 #include <fstream>
 #include <stdexcept>
 #include "../../../../SDK/components/utilities/include/sample_log.h"
+#include "thread_safe_list.h"
 
 using namespace StackFlows;
 
 int main_exit_flage = 0;
 static void __sigint(int iSigNo)
 {
-    SLOGW("llm_sys will be exit!");
+    SLOGW("llm_vlm will be exit!");
     main_exit_flage = 1;
 }
 
 static std::string base_model_path_;
 static std::string base_model_config_path_;
 
+typedef struct {
+    cv::Mat inference_src;
+    bool inference_bgr2rgb;
+} inference_async_par;
+
 typedef std::function<void(const std::string &data, bool finish)> task_callback_t;
 
 #define CONFIG_AUTO_SET(obj, key)             \
@@ -56,6 +62,8 @@ class llm_task {
     task_callback_t out_callback_;
     bool enoutput_;
     bool enstream_;
+    bool encamera_;
+    thread_safe::list<inference_async_par> async_list_;
 
     void set_output(task_callback_t out_callback)
     {
@@ -222,10 +230,46 @@ class llm_task {
         return oss_prompt.str();
     }
 
+    int inference_async(cv::Mat &src, bool bgr2rgb = true)
+    {
+        if (async_list_.size() < 1) {
+            inference_async_par par;
+            par.inference_src     = src.clone();
+            par.inference_bgr2rgb = bgr2rgb;
+            async_list_.put(par);
+        }
+        return async_list_.size();
+    }
+
+    bool inference_raw_yuv(const std::string &msg)
+    {
+        if (msg.size() != 320 * 320 * 2) {
+            throw std::string("img size error");
+        }
+        cv::Mat camera_data(320, 320, CV_8UC2, (void *)msg.data());
+        cv::Mat rgb;
+        cv::cvtColor(camera_data, rgb, cv::COLOR_YUV2RGB_YUYV);
+        return inference_async(rgb, true) ? false : true;
+    }
+
     void inference(const std::string &msg)
     {
         try {
-            if (image_data_.empty()) {
+            if (encamera_) {
+                inference_async_par par;
+                async_list_.get();  // discard buffered frames
+                par = async_list_.get();
+                if (par.inference_src.empty()) return;
+                if (par.inference_bgr2rgb) {
+                    cv::Mat rgb;
+                    cv::cvtColor(par.inference_src, rgb, cv::COLOR_BGR2RGB);
+                    par.inference_src = rgb;
+                }
+                lLaMa_->Encode(par.inference_src, img_embed);
+                lLaMa_->Encode(img_embed, prompt_data_, prompt_complete(msg));
+                std::string out = lLaMa_->Run(prompt_data_);
+                if (out_callback_) out_callback_(out, true);
+            } else if (image_data_.empty()) {
                 lLaMa_->Encode(prompt_data_, prompt_complete(msg));
                 std::string out = lLaMa_->Run(prompt_data_);
                 if (out_callback_) out_callback_(out, true);
@@ -302,13 +346,13 @@ std::atomic<unsigned int> llm_task::next_port_{8090};
 
 #undef CONFIG_AUTO_SET
 
-class llm_llm : public StackFlow {
+class llm_vlm : public StackFlow {
 private:
     int task_count_;
     std::unordered_map<int, std::shared_ptr<llm_task>> llm_task_;
 
 public:
-    llm_llm() : StackFlow("vlm")
+    llm_vlm() : StackFlow("vlm")
     {
         task_count_ = 2;
     }
@@ -447,6 +491,23 @@ class llm_llm : public StackFlow {
         llm_task_obj->lLaMa_->Stop();
     }
 
+    void task_camera_data(const std::weak_ptr<llm_task> llm_task_obj_weak,
+                          const std::weak_ptr<llm_channel_obj> llm_channel_weak, const std::string &data)
+    {
+        nlohmann::json error_body;
+        auto llm_task_obj = llm_task_obj_weak.lock();
+        auto llm_channel  = llm_channel_weak.lock();
+        if (!(llm_task_obj && llm_channel)) {
+            SLOGE("Model run failed.");
+            return;
+        }
+        try {
+            llm_task_obj->inference_raw_yuv(data);
+        } catch (...) {
+            SLOGE("data format error");
+        }
+    }
+
     int setup(const std::string &work_id, const std::string &object, const std::string &data) override
     {
         nlohmann::json error_body;
@@ -476,26 +537,38 @@ class llm_llm : public StackFlow {
             llm_channel->set_output(llm_task_obj->enoutput_);
             llm_channel->set_stream(llm_task_obj->enstream_);
 
-            llm_task_obj->set_output(std::bind(&llm_llm::task_output, this, std::weak_ptr<llm_task>(llm_task_obj),
+            llm_task_obj->set_output(std::bind(&llm_vlm::task_output, this, std::weak_ptr<llm_task>(llm_task_obj),
                                                std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
                                                std::placeholders::_2));
 
             for (const auto input : llm_task_obj->inputs_) {
                 if (input.find("vlm") != std::string::npos) {
                     llm_channel->subscriber_work_id(
-                        "", std::bind(&llm_llm::task_user_data, this, std::weak_ptr<llm_task>(llm_task_obj),
+                        "", std::bind(&llm_vlm::task_user_data, this, std::weak_ptr<llm_task>(llm_task_obj),
                                       std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
                                       std::placeholders::_2));
                 } else if (input.find("asr") != std::string::npos) {
                     llm_channel->subscriber_work_id(
-                        input, std::bind(&llm_llm::task_asr_data, this, std::weak_ptr<llm_task>(llm_task_obj),
+                        input, std::bind(&llm_vlm::task_asr_data, this, std::weak_ptr<llm_task>(llm_task_obj),
                                          std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
                                          std::placeholders::_2));
                 } else if (input.find("kws") != std::string::npos) {
                     llm_channel->subscriber_work_id(
-                        input, std::bind(&llm_llm::kws_awake, this, std::weak_ptr<llm_task>(llm_task_obj),
+                        input, std::bind(&llm_vlm::kws_awake, this, std::weak_ptr<llm_task>(llm_task_obj),
                                          std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
                                          std::placeholders::_2));
+                } else if (input.find("camera") != std::string::npos) {
+                    llm_task_obj->encamera_    = true;
+                    std::string input_url_name = input + ".out_port";
+                    std::string input_url      = unit_call("sys", "sql_select", input_url_name);
+                    if (!input_url.empty()) {
+                        std::weak_ptr<llm_task> _llm_task_obj       = llm_task_obj;
+                        std::weak_ptr<llm_channel_obj> _llm_channel = llm_channel;
+                        llm_channel->subscriber(input_url, [this, _llm_task_obj, _llm_channel](
+                                                               pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
+                            this->task_camera_data(_llm_task_obj, _llm_channel, raw->string());
+                        });
+                    }
                 }
             }
             llm_task_[work_id_num] = llm_task_obj;
@@ -513,7 +586,7 @@ class llm_llm : public StackFlow {
 
     void link(const std::string &work_id, const std::string &object, const std::string &data) override
     {
-        SLOGI("llm_llm::link:%s", data.c_str());
+        SLOGI("llm_vlm::link:%s", data.c_str());
         int ret = 1;
         nlohmann::json error_body;
         int work_id_num = sample_get_work_id_num(work_id);
@@ -528,15 +601,28 @@ class llm_llm : public StackFlow {
         if (data.find("asr") != std::string::npos) {
             ret = llm_channel->subscriber_work_id(
                 data,
-                std::bind(&llm_llm::task_asr_data, this, std::weak_ptr<llm_task>(llm_task_obj),
+                std::bind(&llm_vlm::task_asr_data, this, std::weak_ptr<llm_task>(llm_task_obj),
                           std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1, std::placeholders::_2));
             llm_task_obj->inputs_.push_back(data);
         } else if (data.find("kws") != std::string::npos) {
             ret = llm_channel->subscriber_work_id(
                 data,
-                std::bind(&llm_llm::kws_awake, this, std::weak_ptr<llm_task>(llm_task_obj),
+                std::bind(&llm_vlm::kws_awake, this, std::weak_ptr<llm_task>(llm_task_obj),
                           std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1, std::placeholders::_2));
             llm_task_obj->inputs_.push_back(data);
+        } else if (data.find("camera") != std::string::npos) {
+            llm_task_obj->encamera_    = true;
+            std::string input_url_name = data + ".out_port";
+            std::string input_url      = unit_call("sys", "sql_select", input_url_name);
+            if (!input_url.empty()) {
+                std::weak_ptr<llm_task> _llm_task_obj       = llm_task_obj;
+                std::weak_ptr<llm_channel_obj> _llm_channel = llm_channel;
+                llm_channel->subscriber(
+                    input_url, [this, _llm_task_obj, _llm_channel](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
+                        this->task_camera_data(_llm_task_obj, _llm_channel, raw->string());
+                    });
+            }
+            llm_task_obj->inputs_.push_back(data);
         }
         if (ret) {
             error_body["code"]    = -20;
@@ -550,7 +636,7 @@ class llm_llm : public StackFlow {
 
     void unlink(const std::string &work_id, const std::string &object, const std::string &data) override
     {
-        SLOGI("llm_llm::unlink:%s", data.c_str());
+        SLOGI("llm_vlm::unlink:%s", data.c_str());
         int ret = 0;
         nlohmann::json error_body;
         int work_id_num = sample_get_work_id_num(work_id);
@@ -575,7 +661,7 @@ class llm_llm : public StackFlow {
 
     void taskinfo(const std::string &work_id, const std::string &object, const std::string &data) override
     {
-        SLOGI("llm_llm::taskinfo:%s", data.c_str());
+        SLOGI("llm_vlm::taskinfo:%s", data.c_str());
         nlohmann::json req_body;
         int work_id_num = sample_get_work_id_num(work_id);
         if (WORK_ID_NONE == work_id_num) {
@@ -602,7 +688,7 @@ class llm_llm : public StackFlow {
 
     int exit(const std::string &work_id, const std::string &object, const std::string &data) override
     {
-        SLOGI("llm_llm::exit:%s", data.c_str());
+        SLOGI("llm_vlm::exit:%s", data.c_str());
 
         nlohmann::json error_body;
         int work_id_num = sample_get_work_id_num(work_id);
@@ -621,7 +707,7 @@ class llm_llm : public StackFlow {
         return 0;
     }
 
-    ~llm_llm()
+    ~llm_vlm()
     {
         while (1) {
             auto iteam = llm_task_.begin();
@@ -641,7 +727,7 @@ int main(int argc, char *argv[])
     signal(SIGTERM, __sigint);
     signal(SIGINT, __sigint);
     mkdir("/tmp/llm", 0777);
-    llm_llm llm;
+    llm_vlm llm;
     while (!main_exit_flage) {
         sleep(1);
     }

From 357a6f11574c27b15ead26b35a8529dab55c833e Mon Sep 17 00:00:00 2001
From: dianjixz <18637716021@163.com>
Date: Mon, 23 Jun 2025 11:54:54 +0800
Subject: [PATCH 15/79] [update] llm-camera axera camera add custom_config

---
 .../llm_framework/main_camera/camera.json     |   4 +-
 .../main_camera/src/axera_camera.c            |  55 +-
 .../main_camera/src/axera_camera.h            |  41 +-
 .../llm_framework/main_camera/src/camera.h    |   3 +-
 .../llm_framework/main_camera/src/main.cpp    | 690 +++++++++---------
 .../main_camera/src/v4l2_camera.c             |   2 +-
 6 files changed, 419 insertions(+), 376 deletions(-)

diff --git a/projects/llm_framework/main_camera/camera.json b/projects/llm_framework/main_camera/camera.json
index 22fbd880..081fe9ba 100644
--- a/projects/llm_framework/main_camera/camera.json
+++ b/projects/llm_framework/main_camera/camera.json
@@ -13,7 +13,9 @@
         "image.jpeg.base64"
     ],
     "cap_param": {
-        "None": "None"
+        "VinParam.eSysMode": 1,
+        "VinParam.eHdrMode": 1,
+        "VinParam.bAiispEnable": 1
     },
     "jpeg_config_param": {
         "stVencAttr.enType": 26,
diff --git a/projects/llm_framework/main_camera/src/axera_camera.c b/projects/llm_framework/main_camera/src/axera_camera.c
index 841a6b56..00172883 100644
--- a/projects/llm_framework/main_camera/src/axera_camera.c
+++ b/projects/llm_framework/main_camera/src/axera_camera.c
@@ -115,27 +115,7 @@ AX_VIN_CHN_ATTR_T gSc850slChn0Attr = {
     .tFrameRateCtrl = {AX_INVALID_FRMRATE, AX_INVALID_FRMRATE},
 };
 
-typedef enum {
-    SAMPLE_VIN_NONE                      = -1,
-    SAMPLE_VIN_SINGLE_DUMMY              = 0,
-    SAMPLE_VIN_SINGLE_OS04A10            = 1,
-    SAMPLE_VIN_DOUBLE_OS04A10            = 2,
-    SAMPLE_VIN_SINGLE_SC450AI            = 3,
-    SAMPLE_VIN_DOUBLE_SC450AI            = 4,
-    SAMPLE_VIN_DOUBLE_OS04A10_AND_BT656  = 5,
-    SAMPLE_VIN_SINGLE_S5KJN1SQ03         = 6,
-    SAMPLE_VIN_SINGLE_OS04A10_DCG_HDR    = 7,
-    SAMPLE_VIN_SINGLE_OS04A10_DCG_VS_HDR = 8,
-    SYS_CASE_SINGLE_DVP                  = 20,
-    SYS_CASE_SINGLE_BT601                = 21,
-    SYS_CASE_SINGLE_BT656                = 22,
-    SYS_CASE_SINGLE_BT1120               = 23,
-    SYS_CASE_SINGLE_LVDS                 = 24,
-    SYS_CASE_SINGLE_OS04A10_ONLINE       = 25,
-    SMARTSENS_SC850SL                    = 13,
-    SAMPLE_VIN_SINGLE_SC850SL            = 26,
-    SAMPLE_VIN_BUTT
-} SAMPLE_VIN_CASE_E;
+
 
 struct axera_camera_index_t {
     char name[48];
@@ -157,14 +137,7 @@ struct axera_camera_index_t {
                           {"axera_single_os04a10_online", SYS_CASE_SINGLE_OS04A10_ONLINE},
                           {"axera_single_sc850sl", SAMPLE_VIN_SINGLE_SC850SL}};
 
-typedef struct {
-    SAMPLE_VIN_CASE_E eSysCase;
-    COMMON_VIN_MODE_E eSysMode;
-    AX_SNS_HDR_MODE_E eHdrMode;
-    SAMPLE_LOAD_RAW_NODE_E eLoadRawNode;
-    AX_BOOL bAiispEnable;
-    AX_S32 nDumpFrameNum;
-} SAMPLE_VIN_PARAM_T;
+
 
 /* comm pool */
 COMMON_SYS_POOL_CFG_T gtSysCommPoolSingleDummySdr[] = {
@@ -1194,11 +1167,13 @@ struct axera_camera_t {
     AX_VIDEO_FRAME_T out_img;
     int Chn;
     AX_VENC_CHN_ATTR_T stVencChnAttr;
+    AX_VENC_CHN_ATTR_T stJPEGVencChnAttr;
     AX_IVPS_PIPELINE_ATTR_T stPipelineAttr;
     AX_RTSP_HANDLE pRtspHandle;
     AX_RTSP_ATTR_T stRtspAttr[MAX_RTSP_MAX_CHANNEL_NUM];
     pthread_t venc_thread_id_;
     int venc_run_;
+    int venc_jpeg_run_;
 } axera_obj = {0};
 
 static int camera_capture_callback_set(struct camera_t *camera, vcamera_frame_get pcallback)
@@ -1448,6 +1423,12 @@ void init_rtsp(AX_VENC_CHN_ATTR_T *stVencChnAttr)
     axera_obj.venc_run_     = 1;
 }
 
+void init_jpeg(AX_VENC_CHN_ATTR_T *stVencChnAttr)
+{
+    axera_obj.stJPEGVencChnAttr = *stVencChnAttr;
+    axera_obj.venc_jpeg_run_    = 1;
+}
+
 static int SAMPLE_IVPS_Init(AX_S32 nGrpId, camera_t *camera)
 {
     AX_S32 s32Ret                          = 0, nChn;
@@ -1611,6 +1592,7 @@ int axera_camera_open_from(camera_t *camera)
     int Ret = -1;
     AX_S32 axRet;
     if (camera == NULL) return -1;
+    axera_config_t *axera_config = (axera_config_t *)camera->custom_config_;
     /* Check whether the camera is already open or in an error state */
     SLOGI("Open camera %s...", camera->dev_name_);
     if (camera->state_ & AX_SENSOR_CAM_OPEN) {
@@ -1629,9 +1611,9 @@ int axera_camera_open_from(camera_t *camera)
         return -10;
     }
 
-    axera_obj.VinParam.eSysMode     = COMMON_VIN_SENSOR;
-    axera_obj.VinParam.eHdrMode     = AX_SNS_LINEAR_MODE;
-    axera_obj.VinParam.bAiispEnable = AX_TRUE;
+    axera_obj.VinParam.eSysMode     = axera_config->VinParam.eSysMode; // COMMON_VIN_SENSOR;
+    axera_obj.VinParam.eHdrMode     = axera_config->VinParam.eHdrMode; // AX_SNS_LINEAR_MODE;
+    axera_obj.VinParam.bAiispEnable = axera_config->VinParam.bAiispEnable; // AX_TRUE;
     // axera_obj.gCams.tChnAttr
     __sample_case_config(&axera_obj.gCams, &axera_obj.VinParam, &axera_obj.tCommonArgs, &axera_obj.tPrivArgs);
     COMMON_SYS_Init(&axera_obj.tCommonArgs);
@@ -1682,7 +1664,7 @@ int axera_camera_open_from(camera_t *camera)
     return -1;
 }
 
-camera_t *axera_camera_open(const char *pdev_name, int width, int height, int fps)
+camera_t *axera_camera_open(const char *pdev_name, int width, int height, int fps, void *config)
 {
     int Ret          = -1;
     camera_t *camera = (camera_t *)malloc(sizeof(camera_t));
@@ -1698,9 +1680,10 @@ camera_t *axera_camera_open(const char *pdev_name, int width, int height, int fp
     memset(camera->dev_name_, 0, CONFIG_DEVNAME_LEN);
     memcpy(camera->dev_name_, pdev_name, CopyLen);
 
-    camera->width_       = width;
-    camera->height_      = height;
-    camera->capture_fps_ = fps;
+    camera->width_         = width;
+    camera->height_        = height;
+    camera->capture_fps_   = fps;
+    camera->custom_config_ = config;
 
     Ret = axera_camera_open_from(camera);
     if (Ret) {
diff --git a/projects/llm_framework/main_camera/src/axera_camera.h b/projects/llm_framework/main_camera/src/axera_camera.h
index 6b8644d3..124693e2 100644
--- a/projects/llm_framework/main_camera/src/axera_camera.h
+++ b/projects/llm_framework/main_camera/src/axera_camera.h
@@ -6,15 +6,52 @@
 #ifndef AXERA_CAMERA_H
 #define AXERA_CAMERA_H
 #include "common_venc.h"
+#include "common_vin.h"
 #if __cplusplus
 extern "C" {
 #endif
+
+typedef enum {
+    SAMPLE_VIN_NONE                      = -1,
+    SAMPLE_VIN_SINGLE_DUMMY              = 0,
+    SAMPLE_VIN_SINGLE_OS04A10            = 1,
+    SAMPLE_VIN_DOUBLE_OS04A10            = 2,
+    SAMPLE_VIN_SINGLE_SC450AI            = 3,
+    SAMPLE_VIN_DOUBLE_SC450AI            = 4,
+    SAMPLE_VIN_DOUBLE_OS04A10_AND_BT656  = 5,
+    SAMPLE_VIN_SINGLE_S5KJN1SQ03         = 6,
+    SAMPLE_VIN_SINGLE_OS04A10_DCG_HDR    = 7,
+    SAMPLE_VIN_SINGLE_OS04A10_DCG_VS_HDR = 8,
+    SYS_CASE_SINGLE_DVP                  = 20,
+    SYS_CASE_SINGLE_BT601                = 21,
+    SYS_CASE_SINGLE_BT656                = 22,
+    SYS_CASE_SINGLE_BT1120               = 23,
+    SYS_CASE_SINGLE_LVDS                 = 24,
+    SYS_CASE_SINGLE_OS04A10_ONLINE       = 25,
+    SMARTSENS_SC850SL                    = 13,
+    SAMPLE_VIN_SINGLE_SC850SL            = 26,
+    SAMPLE_VIN_BUTT
+} SAMPLE_VIN_CASE_E;
+
+typedef struct {
+    SAMPLE_VIN_CASE_E eSysCase;
+    COMMON_VIN_MODE_E eSysMode;
+    AX_SNS_HDR_MODE_E eHdrMode;
+    SAMPLE_LOAD_RAW_NODE_E eLoadRawNode;
+    AX_BOOL bAiispEnable;
+    AX_S32 nDumpFrameNum;
+} SAMPLE_VIN_PARAM_T;
+
+typedef struct axera_config_t {
+    SAMPLE_VIN_PARAM_T VinParam;
+} axera_config_t;
+
 /**
  * Open the axera_camera
  * @pdev_name Device node
  * Return value: NULL for failure
  */
-camera_t* axera_camera_open(const char* pdev_name, int width, int height, int fps);
+camera_t* axera_camera_open(const char* pdev_name, int width, int height, int fps, void* config);
 
 /**
  * Open the axera_camera from config
@@ -30,7 +67,7 @@ int axera_camera_open_from(camera_t* camera);
 int axera_camera_close(camera_t* camera);
 
 void init_rtsp(AX_VENC_CHN_ATTR_T *stVencChnAttr);
-void init_jpeg();
+void init_jpeg(AX_VENC_CHN_ATTR_T *stVencChnAttr);
 
 #if __cplusplus
 }
diff --git a/projects/llm_framework/main_camera/src/camera.h b/projects/llm_framework/main_camera/src/camera.h
index 8aab5310..607c9ffa 100644
--- a/projects/llm_framework/main_camera/src/camera.h
+++ b/projects/llm_framework/main_camera/src/camera.h
@@ -36,6 +36,7 @@ typedef struct camera_t {
                             3 Error */
     pthread_t capture_thread_id_;
     void* ctx_;
+    void* custom_config_;
     /**
      * Set capture frame callback
      * Return value: 0 for success, -1 for failure
@@ -59,7 +60,7 @@ typedef struct camera_t {
  * @pdev_name Device node
  * Return value: NULL for failure
  */
-camera_t* camera_open(const char* pdev_name, int width, int height, int fps);
+camera_t* camera_open(const char* pdev_name, int width, int height, int fps, void* config);
 
 /**
  * Open the camera from config
diff --git a/projects/llm_framework/main_camera/src/main.cpp b/projects/llm_framework/main_camera/src/main.cpp
index dad26993..3565b01d 100644
--- a/projects/llm_framework/main_camera/src/main.cpp
+++ b/projects/llm_framework/main_camera/src/main.cpp
@@ -61,20 +61,34 @@ static void __sigint(int iSigNo)
 
 typedef std::function<void(const void *, int)> task_callback_t;
 
-typedef camera_t *(*hal_camera_open_fun)(const char *pdev_name, int width, int height, int fps);
+typedef camera_t *(*hal_camera_open_fun)(const char *pdev_name, int width, int height, int fps, void *config);
 typedef int (*hal_camera_close_fun)(camera_t *camera);
+typedef bool (*hal_parse_config_fun)(const nlohmann::json &config_body, const nlohmann::json &file_body,
+                                     void **custom_config);
+
+#define CONFIG_OBJECT stVencChnAttr
 
 #define CONFIG_AUTO_SET(obj, key)              \
     if (config_body.contains(#key))            \
-        stVencChnAttr.key = config_body[#key]; \
+        CONFIG_OBJECT.key = config_body[#key]; \
     else if (obj.contains(#key))               \
-        stVencChnAttr.key = obj[#key];
+        CONFIG_OBJECT.key = obj[#key];
+
+#define CONFIG_AUTO_SET_DEFAULT(obj, key, default_value) \
+    if (config_body.contains(#key))                      \
+        CONFIG_OBJECT.key = config_body[#key];           \
+    else if (obj.contains(#key))                         \
+        CONFIG_OBJECT.key = obj[#key];                   \
+    else                                                 \
+        CONFIG_OBJECT.key = default_value;
 
 class llm_task {
 private:
     camera_t *cam;
+    void *cam_config;
     hal_camera_open_fun hal_camera_open;
     hal_camera_close_fun hal_camera_close;
+    hal_parse_config_fun hal_parse_config;
 
 public:
     std::string response_format_;
@@ -159,6 +173,311 @@ class llm_task {
         out_callback_ = out_callback;
     }
 
+    static bool parse_axera_config(const nlohmann::json &config_body, const nlohmann::json &file_body,
+                                   void **custom_config)
+    {
+        std::string rtsp_config;
+        static axera_config_t axera_config;
+        memset(&axera_config, 0, sizeof(axera_config_t));
+        if (config_body.contains("rtsp")) {
+            rtsp_config = config_body.at("rtsp");
+        }
+        int frame_width  = config_body.at("frame_width");
+        int frame_height = config_body.at("frame_height");
+        *custom_config   = (void *)&axera_config;
+
+        {
+#undef CONFIG_OBJECT
+#define CONFIG_OBJECT axera_config
+            CONFIG_AUTO_SET_DEFAULT(file_body["cap_param"], VinParam.eSysMode, COMMON_VIN_SENSOR);
+            CONFIG_AUTO_SET_DEFAULT(file_body["cap_param"], VinParam.eHdrMode, AX_SNS_LINEAR_MODE);
+            CONFIG_AUTO_SET_DEFAULT(file_body["cap_param"], VinParam.bAiispEnable, AX_TRUE);
+        }
+
+        if (rtsp_config.empty() == false) {
+            AX_VENC_CHN_ATTR_T stVencChnAttr;
+            memset(&stVencChnAttr, 0, sizeof(AX_VENC_CHN_ATTR_T));
+#undef CONFIG_OBJECT
+#define CONFIG_OBJECT stVencChnAttr
+            if (rtsp_config.find("h264") != std::string::npos) {
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.enType);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u32MaxPicWidth);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u32MaxPicHeight);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.enMemSource);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u32BufSize);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.enProfile);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.enLevel);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.enTier);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u32PicWidthSrc);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u32PicHeightSrc);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.stCropCfg.bEnable);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.stCropCfg.stRect.s32X);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.stCropCfg.stRect.s32Y);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.stCropCfg.stRect.u32Width);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.stCropCfg.stRect.u32Height);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.enRotation);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.enLinkMode);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.bDeBreathEffect);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.bRefRingbuf);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.s32StopWaitTime);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u8InFifoDepth);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u8OutFifoDepth);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u32SliceNum);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.stAttrH265e.bRcnRefShareBuf);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.enRcMode);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.s32FirstFrameStartQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stFrameRate.fSrcFrameRate);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stFrameRate.fDstFrameRate);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32Gop);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32StatTime);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32BitRate);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32MinQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32MaxQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32MinIQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32MaxIQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32MaxIprop);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32MinIprop);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.s32IntraQpDelta);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.s32DeBreathQpDelta);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32IdrQpDeltaRange);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.stQpmapInfo.enCtbRcMode);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.stQpmapInfo.enQpmapQpType);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.stQpmapInfo.enQpmapBlockType);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.stQpmapInfo.enQpmapBlockUnit);
+
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32Gop);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32StatTime);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32MaxBitRate);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.enVQ);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32MaxQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32MinQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32MaxIQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32MinIQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.s32IntraQpDelta);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.s32DeBreathQpDelta);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32IdrQpDeltaRange);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.stQpmapInfo.enCtbRcMode);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.stQpmapInfo.enQpmapQpType);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.stQpmapInfo.enQpmapBlockType);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.stQpmapInfo.enQpmapBlockUnit);
+
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32Gop);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32StatTime);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32MaxBitRate);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32MaxQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32MinQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32MaxIQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32MinIQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.s32IntraQpDelta);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.s32DeBreathQpDelta);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32IdrQpDeltaRange);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.stQpmapInfo.enCtbRcMode);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.stQpmapInfo.enQpmapQpType);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.stQpmapInfo.enQpmapBlockType);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.stQpmapInfo.enQpmapBlockUnit);
+
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QVbr.u32Gop);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QVbr.u32StatTime);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QVbr.u32TargetBitRate);
+
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32Gop);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32StatTime);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MaxQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MinQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MaxIQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MinIQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MinQpDelta);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MaxQpDelta);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.s32DeBreathQpDelta);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32IdrQpDeltaRange);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MaxIprop);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MinIprop);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MaxBitRate);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32ShortTermStatTime);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32LongTermStatTime);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32LongTermMaxBitrate);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32LongTermMinBitrate);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32ExtraBitPercent);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32LongTermStatTimeUnit);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.s32IntraQpDelta);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.stQpmapInfo.enCtbRcMode);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.stQpmapInfo.enQpmapQpType);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.stQpmapInfo.enQpmapBlockType);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.stQpmapInfo.enQpmapBlockUnit);
+
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264FixQp.u32Gop);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264FixQp.u32IQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264FixQp.u32PQp);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264FixQp.u32BQp);
+
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QpMap.u32Gop);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QpMap.u32StatTime);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QpMap.u32TargetBitRate);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QpMap.stQpmapInfo.enCtbRcMode);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QpMap.stQpmapInfo.enQpmapQpType);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QpMap.stQpmapInfo.enQpmapBlockType);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QpMap.stQpmapInfo.enQpmapBlockUnit);
+
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stGopAttr.enGopMode);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stGopAttr.stNormalP.stPicConfig.s32QpOffset);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stGopAttr.stNormalP.stPicConfig.f32QpFactor);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stGopAttr.stOneLTR.stPicConfig.s32QpOffset);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stGopAttr.stOneLTR.stPicConfig.f32QpFactor);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stGopAttr.stOneLTR.stPicSpecialConfig.s32QpOffset);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stGopAttr.stOneLTR.stPicSpecialConfig.f32QpFactor);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stGopAttr.stOneLTR.stPicSpecialConfig.s32Interval);
+                CONFIG_AUTO_SET(file_body["h264_config_param"], stGopAttr.stSvcT.u32GopSize);
+            } else if (rtsp_config.find("h265") != std::string::npos) {
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.enType);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u32MaxPicWidth);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u32MaxPicHeight);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.enMemSource);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u32BufSize);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.enProfile);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.enLevel);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.enTier);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u32PicWidthSrc);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u32PicHeightSrc);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.stCropCfg.bEnable);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.stCropCfg.stRect.s32X);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.stCropCfg.stRect.s32Y);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.stCropCfg.stRect.u32Width);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.stCropCfg.stRect.u32Height);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.enRotation);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.enLinkMode);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.bDeBreathEffect);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.bRefRingbuf);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.s32StopWaitTime);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u8InFifoDepth);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u8OutFifoDepth);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u32SliceNum);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.stAttrH265e.bRcnRefShareBuf);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.enRcMode);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.s32FirstFrameStartQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stFrameRate.fSrcFrameRate);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stFrameRate.fDstFrameRate);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32Gop);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32StatTime);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32BitRate);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32MinQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32MaxQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32MinIQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32MaxIQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32MaxIprop);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32MinIprop);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.s32IntraQpDelta);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.s32DeBreathQpDelta);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32IdrQpDeltaRange);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.stQpmapInfo.enCtbRcMode);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.stQpmapInfo.enQpmapQpType);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.stQpmapInfo.enQpmapBlockType);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.stQpmapInfo.enQpmapBlockUnit);
+
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32Gop);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32StatTime);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32MaxBitRate);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.enVQ);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32MaxQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32MinQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32MaxIQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32MinIQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.s32IntraQpDelta);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.s32DeBreathQpDelta);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32IdrQpDeltaRange);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.stQpmapInfo.enCtbRcMode);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.stQpmapInfo.enQpmapQpType);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.stQpmapInfo.enQpmapBlockType);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.stQpmapInfo.enQpmapBlockUnit);
+
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32Gop);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32StatTime);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32MaxBitRate);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32MaxQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32MinQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32MaxIQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32MinIQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.s32IntraQpDelta);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.s32DeBreathQpDelta);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32IdrQpDeltaRange);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.stQpmapInfo.enCtbRcMode);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.stQpmapInfo.enQpmapQpType);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.stQpmapInfo.enQpmapBlockType);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.stQpmapInfo.enQpmapBlockUnit);
+
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QVbr.u32Gop);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QVbr.u32StatTime);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QVbr.u32TargetBitRate);
+
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32Gop);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32StatTime);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MaxQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MinQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MaxIQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MinIQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MinQpDelta);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MaxQpDelta);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.s32DeBreathQpDelta);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32IdrQpDeltaRange);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MaxIprop);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MinIprop);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MaxBitRate);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32ShortTermStatTime);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32LongTermStatTime);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32LongTermMaxBitrate);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32LongTermMinBitrate);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32ExtraBitPercent);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32LongTermStatTimeUnit);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.s32IntraQpDelta);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.stQpmapInfo.enCtbRcMode);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.stQpmapInfo.enQpmapQpType);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.stQpmapInfo.enQpmapBlockType);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.stQpmapInfo.enQpmapBlockUnit);
+
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265FixQp.u32Gop);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265FixQp.u32IQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265FixQp.u32PQp);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265FixQp.u32BQp);
+
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QpMap.u32Gop);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QpMap.u32StatTime);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QpMap.u32TargetBitRate);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QpMap.stQpmapInfo.enCtbRcMode);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QpMap.stQpmapInfo.enQpmapQpType);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QpMap.stQpmapInfo.enQpmapBlockType);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QpMap.stQpmapInfo.enQpmapBlockUnit);
+
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stGopAttr.enGopMode);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stGopAttr.stNormalP.stPicConfig.s32QpOffset);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stGopAttr.stNormalP.stPicConfig.f32QpFactor);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stGopAttr.stOneLTR.stPicConfig.s32QpOffset);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stGopAttr.stOneLTR.stPicConfig.f32QpFactor);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stGopAttr.stOneLTR.stPicSpecialConfig.s32QpOffset);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stGopAttr.stOneLTR.stPicSpecialConfig.f32QpFactor);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stGopAttr.stOneLTR.stPicSpecialConfig.s32Interval);
+                CONFIG_AUTO_SET(file_body["h265_config_param"], stGopAttr.stSvcT.u32GopSize);
+            }
+
+            try {
+                std::regex pattern(R"(rtsp\.(\d+)[xX-](\d+)\.h(264|265))");
+                std::smatch matches;
+                if (std::regex_search(rtsp_config, matches, pattern)) {
+                    if (matches.size() >= 3) {
+                        stVencChnAttr.stVencAttr.u32PicWidthSrc  = std::stoi(matches[1].str());
+                        stVencChnAttr.stVencAttr.u32PicHeightSrc = std::stoi(matches[2].str());
+                    }
+                }
+            } catch (...) {
+                return true;
+            }
+            if ((stVencChnAttr.stVencAttr.u32PicWidthSrc < frame_width) ||
+                (stVencChnAttr.stVencAttr.u32PicHeightSrc < frame_height)) {
+                return true;
+            }
+            init_rtsp(&stVencChnAttr);
+        }
+        return false;
+    }
+
     bool parse_config(const nlohmann::json &config_body)
     {
         try {
@@ -167,9 +486,6 @@ class llm_task {
             devname_         = config_body.at("input");
             frame_width_     = config_body.at("frame_width");
             frame_height_    = config_body.at("frame_height");
-            if (config_body.contains("rtsp")) {
-                rtsp_config_ = config_body.at("rtsp");
-            }
             if (config_body.contains("enable_webstream")) {
                 enable_webstream_ = config_body.at("enable_webstream");
             } else {
@@ -185,340 +501,43 @@ class llm_task {
         if (devname_.find("/dev/video") != std::string::npos) {
             hal_camera_open  = camera_open;
             hal_camera_close = camera_close;
+            hal_parse_config = NULL;
         } else if (devname_.find("axera_") != std::string::npos) {
             hal_camera_open  = axera_camera_open;
             hal_camera_close = axera_camera_close;
-            if (!rtsp_config_.empty()) {
-                nlohmann::json error_body;
-                nlohmann::json file_body;
-                std::string base_model_path;
-                std::string base_model_config_path;
-                std::list<std::string> config_file_paths =
-                    get_config_file_paths(base_model_path, base_model_config_path, "camera");
-                try {
-                    for (auto file_name : config_file_paths) {
-                        std::ifstream config_file(file_name);
-                        if (!config_file.is_open()) {
-                            SLOGW("config file :%s miss", file_name.c_str());
-                            continue;
-                        }
-                        SLOGI("config file :%s read", file_name.c_str());
-                        config_file >> file_body;
-                        config_file.close();
-                        break;
-                    }
-                    if (file_body.empty()) {
-                        SLOGE("all config file miss");
-                        return true;
-                    }
-                    AX_VENC_CHN_ATTR_T stVencChnAttr;
-                    memset(&stVencChnAttr, 0, sizeof(AX_VENC_CHN_ATTR_T));
-                    if (rtsp_config_.find("h264") != std::string::npos) {
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.enType);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u32MaxPicWidth);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u32MaxPicHeight);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.enMemSource);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u32BufSize);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.enProfile);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.enLevel);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.enTier);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u32PicWidthSrc);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u32PicHeightSrc);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.stCropCfg.bEnable);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.stCropCfg.stRect.s32X);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.stCropCfg.stRect.s32Y);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.stCropCfg.stRect.u32Width);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.stCropCfg.stRect.u32Height);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.enRotation);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.enLinkMode);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.bDeBreathEffect);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.bRefRingbuf);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.s32StopWaitTime);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u8InFifoDepth);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u8OutFifoDepth);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.u32SliceNum);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stVencAttr.stAttrH265e.bRcnRefShareBuf);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.enRcMode);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.s32FirstFrameStartQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stFrameRate.fSrcFrameRate);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stFrameRate.fDstFrameRate);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32Gop);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32StatTime);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32BitRate);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32MinQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32MaxQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32MinIQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32MaxIQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32MaxIprop);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32MinIprop);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.s32IntraQpDelta);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.s32DeBreathQpDelta);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.u32IdrQpDeltaRange);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.stQpmapInfo.enCtbRcMode);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Cbr.stQpmapInfo.enQpmapQpType);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"],
-                                        stRcAttr.stH264Cbr.stQpmapInfo.enQpmapBlockType);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"],
-                                        stRcAttr.stH264Cbr.stQpmapInfo.enQpmapBlockUnit);
-
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32Gop);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32StatTime);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32MaxBitRate);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.enVQ);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32MaxQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32MinQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32MaxIQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32MinIQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.s32IntraQpDelta);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.s32DeBreathQpDelta);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.u32IdrQpDeltaRange);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.stQpmapInfo.enCtbRcMode);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264Vbr.stQpmapInfo.enQpmapQpType);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"],
-                                        stRcAttr.stH264Vbr.stQpmapInfo.enQpmapBlockType);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"],
-                                        stRcAttr.stH264Vbr.stQpmapInfo.enQpmapBlockUnit);
-
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32Gop);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32StatTime);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32MaxBitRate);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32MaxQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32MinQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32MaxIQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32MinIQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.s32IntraQpDelta);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.s32DeBreathQpDelta);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.u32IdrQpDeltaRange);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.stQpmapInfo.enCtbRcMode);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264AVbr.stQpmapInfo.enQpmapQpType);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"],
-                                        stRcAttr.stH264AVbr.stQpmapInfo.enQpmapBlockType);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"],
-                                        stRcAttr.stH264AVbr.stQpmapInfo.enQpmapBlockUnit);
-
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QVbr.u32Gop);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QVbr.u32StatTime);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QVbr.u32TargetBitRate);
-
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32Gop);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32StatTime);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MaxQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MinQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MaxIQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MinIQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MinQpDelta);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MaxQpDelta);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.s32DeBreathQpDelta);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32IdrQpDeltaRange);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MaxIprop);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MinIprop);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32MaxBitRate);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32ShortTermStatTime);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32LongTermStatTime);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32LongTermMaxBitrate);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32LongTermMinBitrate);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32ExtraBitPercent);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.u32LongTermStatTimeUnit);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.s32IntraQpDelta);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.stQpmapInfo.enCtbRcMode);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264CVbr.stQpmapInfo.enQpmapQpType);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"],
-                                        stRcAttr.stH264CVbr.stQpmapInfo.enQpmapBlockType);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"],
-                                        stRcAttr.stH264CVbr.stQpmapInfo.enQpmapBlockUnit);
-
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264FixQp.u32Gop);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264FixQp.u32IQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264FixQp.u32PQp);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264FixQp.u32BQp);
-
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QpMap.u32Gop);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QpMap.u32StatTime);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QpMap.u32TargetBitRate);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QpMap.stQpmapInfo.enCtbRcMode);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stRcAttr.stH264QpMap.stQpmapInfo.enQpmapQpType);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"],
-                                        stRcAttr.stH264QpMap.stQpmapInfo.enQpmapBlockType);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"],
-                                        stRcAttr.stH264QpMap.stQpmapInfo.enQpmapBlockUnit);
-
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stGopAttr.enGopMode);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stGopAttr.stNormalP.stPicConfig.s32QpOffset);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stGopAttr.stNormalP.stPicConfig.f32QpFactor);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stGopAttr.stOneLTR.stPicConfig.s32QpOffset);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stGopAttr.stOneLTR.stPicConfig.f32QpFactor);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"],
-                                        stGopAttr.stOneLTR.stPicSpecialConfig.s32QpOffset);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"],
-                                        stGopAttr.stOneLTR.stPicSpecialConfig.f32QpFactor);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"],
-                                        stGopAttr.stOneLTR.stPicSpecialConfig.s32Interval);
-                        CONFIG_AUTO_SET(file_body["h264_config_param"], stGopAttr.stSvcT.u32GopSize);
-                    } else if (rtsp_config_.find("h265") != std::string::npos) {
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.enType);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u32MaxPicWidth);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u32MaxPicHeight);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.enMemSource);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u32BufSize);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.enProfile);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.enLevel);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.enTier);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u32PicWidthSrc);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u32PicHeightSrc);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.stCropCfg.bEnable);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.stCropCfg.stRect.s32X);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.stCropCfg.stRect.s32Y);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.stCropCfg.stRect.u32Width);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.stCropCfg.stRect.u32Height);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.enRotation);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.enLinkMode);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.bDeBreathEffect);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.bRefRingbuf);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.s32StopWaitTime);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u8InFifoDepth);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u8OutFifoDepth);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.u32SliceNum);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stVencAttr.stAttrH265e.bRcnRefShareBuf);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.enRcMode);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.s32FirstFrameStartQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stFrameRate.fSrcFrameRate);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stFrameRate.fDstFrameRate);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32Gop);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32StatTime);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32BitRate);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32MinQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32MaxQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32MinIQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32MaxIQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32MaxIprop);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32MinIprop);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.s32IntraQpDelta);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.s32DeBreathQpDelta);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.u32IdrQpDeltaRange);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.stQpmapInfo.enCtbRcMode);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Cbr.stQpmapInfo.enQpmapQpType);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"],
-                                        stRcAttr.stH265Cbr.stQpmapInfo.enQpmapBlockType);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"],
-                                        stRcAttr.stH265Cbr.stQpmapInfo.enQpmapBlockUnit);
-
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32Gop);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32StatTime);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32MaxBitRate);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.enVQ);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32MaxQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32MinQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32MaxIQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32MinIQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.s32IntraQpDelta);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.s32DeBreathQpDelta);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.u32IdrQpDeltaRange);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.stQpmapInfo.enCtbRcMode);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265Vbr.stQpmapInfo.enQpmapQpType);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"],
-                                        stRcAttr.stH265Vbr.stQpmapInfo.enQpmapBlockType);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"],
-                                        stRcAttr.stH265Vbr.stQpmapInfo.enQpmapBlockUnit);
-
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32Gop);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32StatTime);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32MaxBitRate);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32MaxQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32MinQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32MaxIQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32MinIQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.s32IntraQpDelta);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.s32DeBreathQpDelta);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.u32IdrQpDeltaRange);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.stQpmapInfo.enCtbRcMode);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265AVbr.stQpmapInfo.enQpmapQpType);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"],
-                                        stRcAttr.stH265AVbr.stQpmapInfo.enQpmapBlockType);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"],
-                                        stRcAttr.stH265AVbr.stQpmapInfo.enQpmapBlockUnit);
-
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QVbr.u32Gop);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QVbr.u32StatTime);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QVbr.u32TargetBitRate);
-
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32Gop);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32StatTime);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MaxQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MinQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MaxIQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MinIQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MinQpDelta);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MaxQpDelta);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.s32DeBreathQpDelta);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32IdrQpDeltaRange);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MaxIprop);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MinIprop);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32MaxBitRate);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32ShortTermStatTime);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32LongTermStatTime);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32LongTermMaxBitrate);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32LongTermMinBitrate);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32ExtraBitPercent);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.u32LongTermStatTimeUnit);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.s32IntraQpDelta);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.stQpmapInfo.enCtbRcMode);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265CVbr.stQpmapInfo.enQpmapQpType);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"],
-                                        stRcAttr.stH265CVbr.stQpmapInfo.enQpmapBlockType);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"],
-                                        stRcAttr.stH265CVbr.stQpmapInfo.enQpmapBlockUnit);
-
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265FixQp.u32Gop);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265FixQp.u32IQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265FixQp.u32PQp);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265FixQp.u32BQp);
-
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QpMap.u32Gop);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QpMap.u32StatTime);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QpMap.u32TargetBitRate);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QpMap.stQpmapInfo.enCtbRcMode);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stRcAttr.stH265QpMap.stQpmapInfo.enQpmapQpType);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"],
-                                        stRcAttr.stH265QpMap.stQpmapInfo.enQpmapBlockType);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"],
-                                        stRcAttr.stH265QpMap.stQpmapInfo.enQpmapBlockUnit);
-
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stGopAttr.enGopMode);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stGopAttr.stNormalP.stPicConfig.s32QpOffset);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stGopAttr.stNormalP.stPicConfig.f32QpFactor);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stGopAttr.stOneLTR.stPicConfig.s32QpOffset);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stGopAttr.stOneLTR.stPicConfig.f32QpFactor);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"],
-                                        stGopAttr.stOneLTR.stPicSpecialConfig.s32QpOffset);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"],
-                                        stGopAttr.stOneLTR.stPicSpecialConfig.f32QpFactor);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"],
-                                        stGopAttr.stOneLTR.stPicSpecialConfig.s32Interval);
-                        CONFIG_AUTO_SET(file_body["h265_config_param"], stGopAttr.stSvcT.u32GopSize);
-                    }
-                    try {
-                        std::regex pattern(R"(rtsp\.(\d+)[xX-](\d+)\.h(264|265))");
-                        std::smatch matches;
-                        if (std::regex_search(rtsp_config_, matches, pattern)) {
-                            if (matches.size() >= 3) {
-                                stVencChnAttr.stVencAttr.u32PicWidthSrc  = std::stoi(matches[1].str());
-                                stVencChnAttr.stVencAttr.u32PicHeightSrc = std::stoi(matches[2].str());
-                            }
-                        }
-                    } catch (...) {
-                        return true;
-                    }
-                    if ((stVencChnAttr.stVencAttr.u32PicWidthSrc < frame_width_) ||
-                        (stVencChnAttr.stVencAttr.u32PicHeightSrc < frame_height_)) {
-                        return true;
+            hal_parse_config = llm_task::parse_axera_config;
+        } else {
+            return true;
+        }
+        {
+            nlohmann::json error_body;
+            nlohmann::json file_body;
+            std::string base_model_path;
+            std::string base_model_config_path;
+            std::list<std::string> config_file_paths =
+                get_config_file_paths(base_model_path, base_model_config_path, "camera");
+            try {
+                for (auto file_name : config_file_paths) {
+                    std::ifstream config_file(file_name);
+                    if (!config_file.is_open()) {
+                        SLOGW("config file :%s miss", file_name.c_str());
+                        continue;
                     }
-                    init_rtsp(&stVencChnAttr);
-                } catch (...) {
+                    SLOGI("config file :%s read", file_name.c_str());
+                    config_file >> file_body;
+                    config_file.close();
+                    break;
+                }
+                if (file_body.empty()) {
+                    SLOGE("all config file miss");
                     return true;
                 }
+                if (hal_parse_config) {
+                    if (hal_parse_config(config_body, file_body, &cam_config)) return true;
+                }
+            } catch (...) {
+                return true;
             }
-        } else {
-            return true;
         }
 
         return false;
@@ -530,7 +549,7 @@ class llm_task {
             return -1;
         }
         try {
-            cam = hal_camera_open(devname_.c_str(), frame_width_, frame_height_, 30);
+            cam = hal_camera_open(devname_.c_str(), frame_width_, frame_height_, 30, cam_config);
             if (cam == NULL) {
                 printf("Camera open failed \n");
                 return -1;
@@ -553,7 +572,8 @@ class llm_task {
 
     llm_task(const std::string &workid)
     {
-        cam = NULL;
+        cam        = NULL;
+        cam_config = NULL;
     }
 
     void start()
diff --git a/projects/llm_framework/main_camera/src/v4l2_camera.c b/projects/llm_framework/main_camera/src/v4l2_camera.c
index c9efbf71..33f3fe46 100644
--- a/projects/llm_framework/main_camera/src/v4l2_camera.c
+++ b/projects/llm_framework/main_camera/src/v4l2_camera.c
@@ -317,7 +317,7 @@ int camera_open_from(camera_t* camera)
     return -1;
 }
 
-camera_t* camera_open(const char* pdev_name, int width, int height, int fps)
+camera_t* camera_open(const char* pdev_name, int width, int height, int fps, void* config)
 {
     int Ret          = -1;
     camera_t* camera = (camera_t*)malloc(sizeof(camera_t));

From 57b143756f24fa2818997a826dc72108160ce46e Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Tue, 24 Jun 2025 19:23:49 +0800
Subject: [PATCH 16/79] [update] depth_anything use async inference, move
 ax_engine init.fix link return value.

---
 .../main_depth_anything/src/EngineWrapper.cpp |  11 +-
 .../main_depth_anything/src/EngineWrapper.hpp |   2 +-
 .../main_depth_anything/src/main.cpp          | 115 ++++++++++--------
 projects/llm_framework/main_vlm/src/main.cpp  |   1 +
 .../main_yolo/src/EngineWrapper.cpp           |  11 +-
 .../main_yolo/src/EngineWrapper.hpp           |   2 +-
 projects/llm_framework/main_yolo/src/main.cpp |  30 ++---
 7 files changed, 102 insertions(+), 70 deletions(-)

diff --git a/projects/llm_framework/main_depth_anything/src/EngineWrapper.cpp b/projects/llm_framework/main_depth_anything/src/EngineWrapper.cpp
index 7a7ec61f..21c51e49 100644
--- a/projects/llm_framework/main_depth_anything/src/EngineWrapper.cpp
+++ b/projects/llm_framework/main_depth_anything/src/EngineWrapper.cpp
@@ -178,10 +178,19 @@ static AX_S32 CheckModelVNpu(const std::string &strModel, const AX_ENGINE_MODEL_
 }
 #endif
 
-int EngineWrapper::Init(const char *strModelPath, uint32_t nNpuType)
+int EngineWrapper::Init(const char *strModelPath, uint32_t nNpuType, uint32_t npuMode)
 {
     AX_S32 ret = 0;
 
+    // 0. Init AX_ENGINE
+    AX_ENGINE_NPU_ATTR_T npu_attr;
+    memset(&npu_attr, 0, sizeof(npu_attr));
+    npu_attr.eHardMode = static_cast<AX_ENGINE_NPU_MODE_T>(npuMode);
+    ret                = AX_ENGINE_Init(&npu_attr);
+    if (0 != ret) {
+        fprintf(stderr, "Init ax-engine failed{0x%8x}.\n", ret);
+    }
+
     // 1. load model
     AX_BOOL bLoadModelUseCmm     = AX_TRUE;
     AX_CHAR *pModelBufferVirAddr = nullptr;
diff --git a/projects/llm_framework/main_depth_anything/src/EngineWrapper.hpp b/projects/llm_framework/main_depth_anything/src/EngineWrapper.hpp
index c3f848ea..07eb43d6 100644
--- a/projects/llm_framework/main_depth_anything/src/EngineWrapper.hpp
+++ b/projects/llm_framework/main_depth_anything/src/EngineWrapper.hpp
@@ -30,7 +30,7 @@ class EngineWrapper {
         Release();
     }
 
-    int Init(const char* strModelPath, uint32_t nNpuType = 0);
+    int Init(const char* strModelPath, uint32_t nNpuType = 0, uint32_t npuMode = 0);
 
     int SetInput(void* pInput, int index);
 
diff --git a/projects/llm_framework/main_depth_anything/src/main.cpp b/projects/llm_framework/main_depth_anything/src/main.cpp
index 7685bc83..ed82a61a 100644
--- a/projects/llm_framework/main_depth_anything/src/main.cpp
+++ b/projects/llm_framework/main_depth_anything/src/main.cpp
@@ -9,8 +9,8 @@
 #include <ax_sys_api.h>
 #include <sys/stat.h>
 #include <fstream>
-
 #include "../../../../SDK/components/utilities/include/sample_log.h"
+#include "thread_safe_list.h"
 
 using namespace StackFlows;
 
@@ -27,13 +27,15 @@ static std::string base_model_config_path_;
 typedef struct {
     std::string depth_anything_model;
     std::string model_type = "detect";
-    std::vector<std::string> cls_name;
-    int img_h            = 640;
-    int img_w            = 640;
-    int cls_num          = 80;
-    float pron_threshold = 0.45f;
-    float nms_threshold  = 0.45;
-} yolo_config;
+    int img_h              = 640;
+    int img_w              = 640;
+    uint32_t npu_type      = 0;
+} depth_anything_config;
+
+typedef struct {
+    cv::Mat inference_src;
+    bool inference_bgr2rgb;
+} inference_async_par;
 
 typedef std::function<void(const std::string &data, bool finish)> task_callback_t;
 
@@ -46,7 +48,7 @@ typedef std::function<void(const std::string &data, bool finish)> task_callback_
 class llm_task {
 private:
 public:
-    yolo_config mode_config_;
+    depth_anything_config mode_config_;
     std::string model_;
     std::unique_ptr<EngineWrapper> depth_anything_;
     std::string response_format_;
@@ -57,7 +59,8 @@ class llm_task {
     static int ax_init_flage_;
     task_callback_t out_callback_;
     std::atomic_bool camera_flage_;
-    std::mutex inference_mtx_;
+    std::unique_ptr<std::thread> inference_run_;
+    thread_safe::list<inference_async_par> async_list_;
 
     bool parse_config(const nlohmann::json &config_body)
     {
@@ -113,10 +116,11 @@ class llm_task {
             CONFIG_AUTO_SET(file_body["mode_param"], img_h);
             CONFIG_AUTO_SET(file_body["mode_param"], img_w);
             CONFIG_AUTO_SET(file_body["mode_param"], model_type);
+            CONFIG_AUTO_SET(file_body["mode_param"], npu_type);
             mode_config_.depth_anything_model = base_model + mode_config_.depth_anything_model;
             depth_anything_                   = std::make_unique<EngineWrapper>();
-            if (0 != depth_anything_->Init(mode_config_.depth_anything_model.c_str())) {
-                SLOGE("Init yolo_model model failed!\n");
+            if (0 != depth_anything_->Init(mode_config_.depth_anything_model.c_str(), 0, mode_config_.npu_type)) {
+                SLOGE("Init depth_anything_model model failed!\n");
                 return -5;
             }
         } catch (...) {
@@ -140,54 +144,63 @@ class llm_task {
 
     bool inference_decode(const std::string &msg)
     {
-        if (inference_mtx_.try_lock())
-            std::lock_guard<std::mutex> guard(inference_mtx_, std::adopt_lock);
-        else
-            return true;
         cv::Mat src = cv::imdecode(std::vector<uint8_t>(msg.begin(), msg.end()), cv::IMREAD_COLOR);
         if (src.empty()) return true;
-        return inference(src);
+        return inference_async(src) ? false : true;
     }
 
     bool inference_raw_yuv(const std::string &msg)
     {
-        if (inference_mtx_.try_lock())
-            std::lock_guard<std::mutex> guard(inference_mtx_, std::adopt_lock);
-        else
-            return true;
         if (msg.size() != mode_config_.img_w * mode_config_.img_h * 2) {
             throw std::string("img size error");
         }
         cv::Mat camera_data(mode_config_.img_h, mode_config_.img_w, CV_8UC2, (void *)msg.data());
         cv::Mat rgb;
         cv::cvtColor(camera_data, rgb, cv::COLOR_YUV2RGB_YUYV);
-        return inference(rgb, true);
+        return inference_async(rgb, true) ? false : true;
     }
 
     bool inference_raw_rgb(const std::string &msg)
     {
-        if (inference_mtx_.try_lock())
-            std::lock_guard<std::mutex> guard(inference_mtx_, std::adopt_lock);
-        else
-            return true;
         if (msg.size() != mode_config_.img_w * mode_config_.img_h * 3) {
             throw std::string("img size error");
         }
         cv::Mat camera_data(mode_config_.img_h, mode_config_.img_w, CV_8UC3, (void *)msg.data());
-        return inference(camera_data, false);
+        return inference_async(camera_data, false) ? false : true;
     }
 
     bool inference_raw_bgr(const std::string &msg)
     {
-        if (inference_mtx_.try_lock())
-            std::lock_guard<std::mutex> guard(inference_mtx_, std::adopt_lock);
-        else
-            return true;
         if (msg.size() != mode_config_.img_w * mode_config_.img_h * 3) {
             throw std::string("img size error");
         }
         cv::Mat camera_data(mode_config_.img_h, mode_config_.img_w, CV_8UC3, (void *)msg.data());
-        return inference(camera_data);
+        return inference_async(camera_data) ? false : true;
+    }
+
+    void run()
+    {
+        inference_async_par par;
+        for (;;) {
+            {
+                par = async_list_.get();
+                if (par.inference_src.empty()) break;
+                inference(par.inference_src, par.inference_bgr2rgb);
+            }
+        }
+    }
+
+    int inference_async(cv::Mat &src, bool bgr2rgb = true)
+    {
+        if (async_list_.size() < 3) {
+            inference_async_par par;
+            par.inference_src     = src.clone();
+            par.inference_bgr2rgb = bgr2rgb;
+            async_list_.put(par);
+        } else {
+            SLOGE("inference list is full\n");
+        }
+        return async_list_.size();
     }
 
     bool inference(cv::Mat &src, bool bgr2rgb = true)
@@ -195,7 +208,7 @@ class llm_task {
         try {
             int ret = -1;
             std::vector<uint8_t> image(mode_config_.img_w * mode_config_.img_h * 3, 0);
-            common::get_input_data_no_letterbox(src, image, mode_config_.img_h, mode_config_.img_w, bgr2rgb);
+            common::get_input_data_letterbox(src, image, mode_config_.img_h, mode_config_.img_w, bgr2rgb);
             cv::Mat img_mat(mode_config_.img_h, mode_config_.img_w, CV_8UC3, image.data());
             depth_anything_->SetInput((void *)image.data(), 0);
             if (0 != depth_anything_->Run()) {
@@ -206,7 +219,7 @@ class llm_task {
             depth_anything_->Post_Process(img_mat, mode_config_.model_type, depth_anything_output);
             if (out_callback_) out_callback_(depth_anything_output, true);
         } catch (...) {
-            SLOGW("yolo_->Run have error!");
+            SLOGW("depth_anything_->Run have error!");
             return true;
         }
         return false;
@@ -219,12 +232,6 @@ class llm_task {
             if (0 != ret) {
                 fprintf(stderr, "AX_SYS_Init failed! ret = 0x%x\n", ret);
             }
-            AX_ENGINE_NPU_ATTR_T npu_attr;
-            memset(&npu_attr, 0, sizeof(npu_attr));
-            ret = AX_ENGINE_Init(&npu_attr);
-            if (0 != ret) {
-                fprintf(stderr, "Init ax-engine failed{0x%8x}.\n", ret);
-            }
         }
         ax_init_flage_++;
     }
@@ -234,7 +241,6 @@ class llm_task {
         if (ax_init_flage_ > 0) {
             --ax_init_flage_;
             if (!ax_init_flage_) {
-                AX_ENGINE_Deinit();
                 AX_SYS_Deinit();
             }
         }
@@ -243,15 +249,25 @@ class llm_task {
     llm_task(const std::string &workid)
     {
         _ax_init();
+        inference_run_ = std::make_unique<std::thread>(std::bind(&llm_task::run, this));
     }
 
     void start()
     {
+        if (!inference_run_) {
+            inference_run_ = std::make_unique<std::thread>(std::bind(&llm_task::run, this));
+        }
     }
 
     void stop()
     {
+        if (inference_run_) {
+            inference_async_par par;
+            async_list_.put(par);
+            inference_run_->join();
+            inference_run_.reset();
         }
+    }
 
     ~llm_task()
     {
@@ -418,10 +434,10 @@ class llm_depth_anything : public StackFlow {
                     if (!input_url.empty()) {
                         std::weak_ptr<llm_task> _llm_task_obj       = llm_task_obj;
                         std::weak_ptr<llm_channel_obj> _llm_channel = llm_channel;
-                        llm_channel->subscriber(
-                            input_url, [this, _llm_task_obj, _llm_channel](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
-                                this->task_camera_data(_llm_task_obj, _llm_channel, raw->string());
-                            });
+                        llm_channel->subscriber(input_url, [this, _llm_task_obj, _llm_channel](
+                                                               pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
+                            this->task_camera_data(_llm_task_obj, _llm_channel, raw->string());
+                        });
                     }
                 }
                 llm_task_[work_id_num] = llm_task_obj;
@@ -465,10 +481,11 @@ class llm_depth_anything : public StackFlow {
             if (!input_url.empty()) {
                 std::weak_ptr<llm_task> _llm_task_obj       = llm_task_obj;
                 std::weak_ptr<llm_channel_obj> _llm_channel = llm_channel;
-                llm_channel->subscriber(input_url,
-                                        [this, _llm_task_obj, _llm_channel](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
-                                            this->task_camera_data(_llm_task_obj, _llm_channel, raw->string());
-                                        });
+                llm_channel->subscriber(
+                    input_url, [this, _llm_task_obj, _llm_channel](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
+                        this->task_camera_data(_llm_task_obj, _llm_channel, raw->string());
+                    });
+                ret = 0;
             }
             llm_task_obj->inputs_.push_back(data);
         }
diff --git a/projects/llm_framework/main_vlm/src/main.cpp b/projects/llm_framework/main_vlm/src/main.cpp
index e7aeb36f..6ee6dd67 100644
--- a/projects/llm_framework/main_vlm/src/main.cpp
+++ b/projects/llm_framework/main_vlm/src/main.cpp
@@ -621,6 +621,7 @@ class llm_vlm : public StackFlow {
                     input_url, [this, _llm_task_obj, _llm_channel](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
                         this->task_camera_data(_llm_task_obj, _llm_channel, raw->string());
                     });
+                ret = 0;
             }
             llm_task_obj->inputs_.push_back(data);
         }
diff --git a/projects/llm_framework/main_yolo/src/EngineWrapper.cpp b/projects/llm_framework/main_yolo/src/EngineWrapper.cpp
index 84085db0..99e448d1 100644
--- a/projects/llm_framework/main_yolo/src/EngineWrapper.cpp
+++ b/projects/llm_framework/main_yolo/src/EngineWrapper.cpp
@@ -179,10 +179,19 @@ static AX_S32 CheckModelVNpu(const std::string& strModel, const AX_ENGINE_MODEL_
 }
 #endif
 
-int EngineWrapper::Init(const char* strModelPath, uint32_t nNpuType)
+int EngineWrapper::Init(const char* strModelPath, uint32_t nNpuType, uint32_t npuMode)
 {
     AX_S32 ret = 0;
 
+    // 0. Init AX_ENGINE
+    AX_ENGINE_NPU_ATTR_T npu_attr;
+    memset(&npu_attr, 0, sizeof(npu_attr));
+    npu_attr.eHardMode = static_cast<AX_ENGINE_NPU_MODE_T>(npuMode);
+    ret                = AX_ENGINE_Init(&npu_attr);
+    if (0 != ret) {
+        fprintf(stderr, "Init ax-engine failed{0x%8x}.\n", ret);
+    }
+
     // 1. load model
     AX_BOOL bLoadModelUseCmm     = AX_TRUE;
     AX_CHAR* pModelBufferVirAddr = nullptr;
diff --git a/projects/llm_framework/main_yolo/src/EngineWrapper.hpp b/projects/llm_framework/main_yolo/src/EngineWrapper.hpp
index 5d42a07e..b049a6fe 100644
--- a/projects/llm_framework/main_yolo/src/EngineWrapper.hpp
+++ b/projects/llm_framework/main_yolo/src/EngineWrapper.hpp
@@ -43,7 +43,7 @@ class EngineWrapper {
         Release();
     }
 
-    int Init(const char* strModelPath, uint32_t nNpuType = 0);
+    int Init(const char* strModelPath, uint32_t nNpuType = 0, uint32_t npuMode = 0);
 
     int SetInput(void* pInput, int index);
 
diff --git a/projects/llm_framework/main_yolo/src/main.cpp b/projects/llm_framework/main_yolo/src/main.cpp
index 208a3757..bf259999 100644
--- a/projects/llm_framework/main_yolo/src/main.cpp
+++ b/projects/llm_framework/main_yolo/src/main.cpp
@@ -40,6 +40,7 @@ typedef struct {
     int point_num        = 17;
     float pron_threshold = 0.45f;
     float nms_threshold  = 0.45;
+    uint32_t npu_type    = 0;
 } yolo_config;
 
 typedef struct {
@@ -131,9 +132,10 @@ class llm_task {
             CONFIG_AUTO_SET(file_body["mode_param"], cls_num);
             CONFIG_AUTO_SET(file_body["mode_param"], point_num);
             CONFIG_AUTO_SET(file_body["mode_param"], model_type);
+            CONFIG_AUTO_SET(file_body["mode_param"], npu_type);
             mode_config_.yolo_model = base_model + mode_config_.yolo_model;
             yolo_                   = std::make_unique<EngineWrapper>();
-            if (0 != yolo_->Init(mode_config_.yolo_model.c_str())) {
+            if (0 != yolo_->Init(mode_config_.yolo_model.c_str(), 0, mode_config_.npu_type)) {
                 SLOGE("Init yolo_model model failed!\n");
                 return -5;
             }
@@ -293,12 +295,6 @@ class llm_task {
             if (0 != ret) {
                 fprintf(stderr, "AX_SYS_Init failed! ret = 0x%x\n", ret);
             }
-            AX_ENGINE_NPU_ATTR_T npu_attr;
-            memset(&npu_attr, 0, sizeof(npu_attr));
-            ret = AX_ENGINE_Init(&npu_attr);
-            if (0 != ret) {
-                fprintf(stderr, "Init ax-engine failed{0x%8x}.\n", ret);
-            }
         }
         ax_init_flage_++;
     }
@@ -308,7 +304,6 @@ class llm_task {
         if (ax_init_flage_ > 0) {
             --ax_init_flage_;
             if (!ax_init_flage_) {
-                AX_ENGINE_Deinit();
                 AX_SYS_Deinit();
             }
         }
@@ -355,7 +350,7 @@ class llm_yolo : public StackFlow {
 public:
     llm_yolo() : StackFlow("yolo")
     {
-        task_count_ = 1;
+        task_count_ = 2;
     }
 
     void task_output(const std::weak_ptr<llm_task> llm_task_obj_weak,
@@ -503,10 +498,10 @@ class llm_yolo : public StackFlow {
                     if (!input_url.empty()) {
                         std::weak_ptr<llm_task> _llm_task_obj       = llm_task_obj;
                         std::weak_ptr<llm_channel_obj> _llm_channel = llm_channel;
-                        llm_channel->subscriber(
-                            input_url, [this, _llm_task_obj, _llm_channel](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
-                                this->task_camera_data(_llm_task_obj, _llm_channel, raw->string());
-                            });
+                        llm_channel->subscriber(input_url, [this, _llm_task_obj, _llm_channel](
+                                                               pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
+                            this->task_camera_data(_llm_task_obj, _llm_channel, raw->string());
+                        });
                     }
                 }
                 llm_task_[work_id_num] = llm_task_obj;
@@ -550,10 +545,11 @@ class llm_yolo : public StackFlow {
             if (!input_url.empty()) {
                 std::weak_ptr<llm_task> _llm_task_obj       = llm_task_obj;
                 std::weak_ptr<llm_channel_obj> _llm_channel = llm_channel;
-                llm_channel->subscriber(input_url,
-                                        [this, _llm_task_obj, _llm_channel](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
-                                            this->task_camera_data(_llm_task_obj, _llm_channel, raw->string());
-                                        });
+                llm_channel->subscriber(
+                    input_url, [this, _llm_task_obj, _llm_channel](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
+                        this->task_camera_data(_llm_task_obj, _llm_channel, raw->string());
+                    });
+                ret = 0;
             }
             llm_task_obj->inputs_.push_back(data);
         }

From b7e62dd6d61848e9d77f6d374c97f38ff072ed3f Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Wed, 25 Jun 2025 17:09:35 +0800
Subject: [PATCH 17/79] [update] update llm-depth-anything version, llm-yolo
 version. fix libonnxruntime.so Missing

---
 projects/llm_framework/SConstruct                     | 2 +-
 projects/llm_framework/main/SConstruct                | 1 +
 projects/llm_framework/main_depth_anything/SConstruct | 2 +-
 projects/llm_framework/main_yolo/SConstruct           | 2 +-
 projects/llm_framework/tools/llm_pack.py              | 9 +++++++--
 5 files changed, 11 insertions(+), 5 deletions(-)

diff --git a/projects/llm_framework/SConstruct b/projects/llm_framework/SConstruct
index 7282f87e..6c4575b5 100644
--- a/projects/llm_framework/SConstruct
+++ b/projects/llm_framework/SConstruct
@@ -5,7 +5,7 @@ import shutil
 os.environ['SDK_PATH'] = os.path.normpath(str(Path(os.getcwd())/'..'/'..'/'SDK'))
 os.environ['EXT_COMPONENTS_PATH'] = os.path.normpath(str(Path(os.getcwd())/'..'/'..'/'ext_components'))
 
-version = 'v0.0.8'
+version = 'v0.0.9'
 static_lib = 'static_lib'
 update = False
 
diff --git a/projects/llm_framework/main/SConstruct b/projects/llm_framework/main/SConstruct
index 72d44864..0c82732f 100644
--- a/projects/llm_framework/main/SConstruct
+++ b/projects/llm_framework/main/SConstruct
@@ -27,6 +27,7 @@ STATIC_FILES += [AFile('../static_lib/sherpa/ncnn/libsherpa-ncnn-core.so'),
                  AFile('../static_lib/wetext/libglog.so.0'),
                  AFile('../static_lib/wetext/libfst.so.16'),
                  AFile('../static_lib/libonnxruntime.so.1'),
+                AFile('../static_lib/libonnxruntime.so.1.14.0')
                  ]
 
 env['COMPONENTS'].append({'target':'static_file-1.0',
diff --git a/projects/llm_framework/main_depth_anything/SConstruct b/projects/llm_framework/main_depth_anything/SConstruct
index a90887f6..66caf78a 100644
--- a/projects/llm_framework/main_depth_anything/SConstruct
+++ b/projects/llm_framework/main_depth_anything/SConstruct
@@ -30,7 +30,7 @@ STATIC_LIB += static_file * 2
 STATIC_FILES += Glob('mode_*.json')
 
 
-env['COMPONENTS'].append({'target':'llm_depth_anything-1.6',
+env['COMPONENTS'].append({'target':'llm_depth_anything-1.7',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_yolo/SConstruct b/projects/llm_framework/main_yolo/SConstruct
index 99f1e1ef..7c588e53 100644
--- a/projects/llm_framework/main_yolo/SConstruct
+++ b/projects/llm_framework/main_yolo/SConstruct
@@ -39,7 +39,7 @@ STATIC_FILES += Glob('mode_*.json')
 #                 AFile('../static_lib/libbz2.so.1.0')]
 # DEFINITIONS += ["-DENABLE_BACKWARD"]
 
-env['COMPONENTS'].append({'target':'llm_yolo-1.8',
+env['COMPONENTS'].append({'target':'llm_yolo-1.9',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index b27cd597..1e954272 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -362,9 +362,9 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-melotts':[create_bin_deb,'llm-melotts', '1.8', src_folder, revision],
         'llm-camera':[create_bin_deb,'llm-camera', '1.8', src_folder, revision, 'lib-llm'],
         'llm-vlm':[create_bin_deb,'llm-vlm', '1.7', src_folder, revision],
-        'llm-yolo':[create_bin_deb,'llm-yolo', '1.8', src_folder, revision],
+        'llm-yolo':[create_bin_deb,'llm-yolo', '1.9', src_folder, revision],
         'llm-skel':[create_bin_deb,'llm-skel', version, src_folder, revision],
-        'llm-depth-anything':[create_bin_deb,'llm-depth-anything', '1.6', src_folder, revision],
+        'llm-depth-anything':[create_bin_deb,'llm-depth-anything', '1.7', src_folder, revision],
         'llm-vad':[create_bin_deb,'llm-vad', '1.7', src_folder, revision],
         'llm-whisper':[create_bin_deb,'llm-whisper', '1.7', src_folder, revision],
         'llm-openai-api':[create_bin_deb,'llm-openai-api', '1.7', src_folder, revision],
@@ -385,10 +385,15 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-model-melotts-ja-jp':[create_data_deb,'llm-model-melotts-ja-jp', '0.6', src_folder, revision],
         'llm-model-melotts-es-es':[create_data_deb,'llm-model-melotts-es-es', '0.5', src_folder, revision],
         'llm-model-yolo11n':[create_data_deb,'llm-model-yolo11n', data_version, src_folder, revision],
+        'llm-model-yolo11n-npu1':[create_data_deb,'llm-model-yolo11n-npu1', '0.4', src_folder, revision],
         'llm-model-yolo11n-pose':[create_data_deb,'llm-model-yolo11n-pose', '0.3', src_folder, revision],
+        'llm-model-yolo11n-pose-npu1':[create_data_deb,'llm-model-yolo11n-pose-npu1', '0.4', src_folder, revision],
         'llm-model-yolo11n-hand-pose':[create_data_deb,'llm-model-yolo11n-hand-pose', '0.3', src_folder, revision],
+        'llm-model-yolo11n-hand-pose-npu1':[create_data_deb,'llm-model-yolo11n-hand-pose-npu1', '0.4', src_folder, revision],
         'llm-model-yolo11n-seg':[create_data_deb,'llm-model-yolo11n-seg', '0.3', src_folder, revision],
+        'llm-model-yolo11n-seg-npu1':[create_data_deb,'llm-model-yolo11n-seg-npu1', '0.4', src_folder, revision],
         'llm-model-depth-anything-ax630c':[create_data_deb,'llm-model-depth-anything-ax630c', '0.4', src_folder, revision],
+        'llm-model-depth-anything-npu1-ax630c':[create_data_deb,'llm-model-depth-anything-npu1-ax630c', '0.4', src_folder, revision],
         'llm-model-whisper-tiny':[create_data_deb,'llm-model-whisper-tiny', '0.4', src_folder, revision],
         'llm-model-whisper-base':[create_data_deb,'llm-model-whisper-base', '0.4', src_folder, revision],
         'llm-model-whisper-small':[create_data_deb,'llm-model-whisper-small', '0.4', src_folder, revision],

From 592fd9e16c564cd14b349b3236fbd143fe31c63d Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Wed, 25 Jun 2025 18:04:17 +0800
Subject: [PATCH 18/79] [update] update llm-camera version

---
 projects/llm_framework/main_camera/SConstruct | 2 +-
 projects/llm_framework/tools/llm_pack.py      | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/projects/llm_framework/main_camera/SConstruct b/projects/llm_framework/main_camera/SConstruct
index eb5190c7..92f525e9 100644
--- a/projects/llm_framework/main_camera/SConstruct
+++ b/projects/llm_framework/main_camera/SConstruct
@@ -68,7 +68,7 @@ STATIC_LIB += static_file * 4
 STATIC_FILES += [AFile('camera.json')]
 STATIC_FILES += Glob('mode_*.json')
 
-env['COMPONENTS'].append({'target':'llm_camera-1.8',
+env['COMPONENTS'].append({'target':'llm_camera-1.9',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 1e954272..599676c9 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -360,7 +360,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-llm':[create_bin_deb,'llm-llm', '1.8', src_folder, revision],
         'llm-tts':[create_bin_deb,'llm-tts', '1.6', src_folder, revision],
         'llm-melotts':[create_bin_deb,'llm-melotts', '1.8', src_folder, revision],
-        'llm-camera':[create_bin_deb,'llm-camera', '1.8', src_folder, revision, 'lib-llm'],
+        'llm-camera':[create_bin_deb,'llm-camera', '1.9', src_folder, revision, 'lib-llm'],
         'llm-vlm':[create_bin_deb,'llm-vlm', '1.7', src_folder, revision],
         'llm-yolo':[create_bin_deb,'llm-yolo', '1.9', src_folder, revision],
         'llm-skel':[create_bin_deb,'llm-skel', version, src_folder, revision],

From cccddd2273017a8d6e94046dc01abf0bcc5231c2 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 26 Jun 2025 10:44:43 +0800
Subject: [PATCH 19/79] [update] update llm-vlm version

---
 projects/llm_framework/main_vlm/SConstruct | 2 +-
 projects/llm_framework/tools/llm_pack.py   | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/projects/llm_framework/main_vlm/SConstruct b/projects/llm_framework/main_vlm/SConstruct
index d1046f23..bba83f8f 100644
--- a/projects/llm_framework/main_vlm/SConstruct
+++ b/projects/llm_framework/main_vlm/SConstruct
@@ -73,7 +73,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_vlm-1.7',
+env['COMPONENTS'].append({'target':'llm_vlm-1.8',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 599676c9..1d2a89f0 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -361,7 +361,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-tts':[create_bin_deb,'llm-tts', '1.6', src_folder, revision],
         'llm-melotts':[create_bin_deb,'llm-melotts', '1.8', src_folder, revision],
         'llm-camera':[create_bin_deb,'llm-camera', '1.9', src_folder, revision, 'lib-llm'],
-        'llm-vlm':[create_bin_deb,'llm-vlm', '1.7', src_folder, revision],
+        'llm-vlm':[create_bin_deb,'llm-vlm', '1.8', src_folder, revision],
         'llm-yolo':[create_bin_deb,'llm-yolo', '1.9', src_folder, revision],
         'llm-skel':[create_bin_deb,'llm-skel', version, src_folder, revision],
         'llm-depth-anything':[create_bin_deb,'llm-depth-anything', '1.7', src_folder, revision],

From b0743f068a6b86dff84e3ac685b679c92870a99f Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 27 Jun 2025 15:19:53 +0800
Subject: [PATCH 20/79] [update] update model list & add npu1 model.

---
 README_zh.md                                  |   5 +
 .../mode_depth-anything-npu1-ax630c.json      |  25 ++++
 projects/llm_framework/main_yolo/SConstruct   |   2 +-
 .../models/mode_yolo11n-hand-pose-npu1.json   |  33 ++++++
 .../{ => models}/mode_yolo11n-hand-pose.json  |   0
 .../main_yolo/models/mode_yolo11n-npu1.json   | 111 ++++++++++++++++++
 .../models/mode_yolo11n-pose-npu1.json        |  33 ++++++
 .../{ => models}/mode_yolo11n-pose.json       |   0
 .../models/mode_yolo11n-seg-npu1.json         | 111 ++++++++++++++++++
 .../{ => models}/mode_yolo11n-seg.json        |   0
 .../main_yolo/{ => models}/mode_yolo11n.json  |   0
 11 files changed, 319 insertions(+), 1 deletion(-)
 create mode 100644 projects/llm_framework/main_depth_anything/mode_depth-anything-npu1-ax630c.json
 create mode 100644 projects/llm_framework/main_yolo/models/mode_yolo11n-hand-pose-npu1.json
 rename projects/llm_framework/main_yolo/{ => models}/mode_yolo11n-hand-pose.json (100%)
 create mode 100644 projects/llm_framework/main_yolo/models/mode_yolo11n-npu1.json
 create mode 100644 projects/llm_framework/main_yolo/models/mode_yolo11n-pose-npu1.json
 rename projects/llm_framework/main_yolo/{ => models}/mode_yolo11n-pose.json (100%)
 create mode 100644 projects/llm_framework/main_yolo/models/mode_yolo11n-seg-npu1.json
 rename projects/llm_framework/main_yolo/{ => models}/mode_yolo11n-seg.json (100%)
 rename projects/llm_framework/main_yolo/{ => models}/mode_yolo11n.json (100%)

diff --git a/README_zh.md b/README_zh.md
index e80293e6..a2f7e4a6 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -92,10 +92,15 @@ StackFlow 语音助手的主要工作模式：
 | [smolvlm-256M-ax630c](https://huggingface.co/HuggingFaceTB/SmolVLM-256M-Instruct) | VLM | 330M | 多模态文本生成 | [mode_smolvlm-256M-ax630c.json](projects/llm_framework/main_vlm/models/mode_smolvlm-256M-ax630c.json) | NPU |
 | [smolvlm-500M-ax630c](https://huggingface.co/HuggingFaceTB/SmolVLM-500M-Instruct) | VLM | 605M | 多模态文本生成 | [mode_smolvlm-256M-ax630c.json](projects/llm_framework/main_vlm/models/mode_smolvlm-500M-ax630c.json) | NPU |
 | [yolo11n](https://github.com/ultralytics/ultralytics) | CV | 2.8M | 目标检测 | [mode_yolo11n.json](projects/llm_framework/main_yolo/mode_yolo11n.json) | NPU |
+| [yolo11n-npu1](https://github.com/ultralytics/ultralytics) | CV | 2.8M | 目标检测 | [mode_yolo11n-npu1.json](projects/llm_framework/main_yolo/mode_yolo11n-npu1.json) | NPU |
 | [yolo11n-seg](https://github.com/ultralytics/ultralytics) | CV | 3.0M | 实例分割 | [mode_yolo11n-seg.json](projects/llm_framework/main_yolo/mode_yolo11n-seg.json) | NPU |
+| [yolo11n-seg-npu1](https://github.com/ultralytics/ultralytics) | CV | 3.0M | 实例分割 | [mode_yolo11n-seg-npu1.json](projects/llm_framework/main_yolo/mode_yolo11n-seg-npu1.json) | NPU |
 | [yolo11n-pose](https://github.com/ultralytics/ultralytics) | CV | 3.1M | 姿态检测 | [mode_yolo11n-pose.json](projects/llm_framework/main_yolo/mode_yolo11n-pose.json) | NPU |
+| [yolo11n-pose-npu1](https://github.com/ultralytics/ultralytics) | CV | 3.1M | 姿态检测 | [mode_yolo11n-pose-npu1.json](projects/llm_framework/main_yolo/mode_yolo11n-pose-npu1.json) | NPU |
 | [yolo11n-hand-pose](https://github.com/ultralytics/ultralytics) | CV | 3.2M | 姿态检测 | [mode_yolo11n-hand-pose.json](projects/llm_framework/main_yolo/mode_yolo11n-hand-pose.json) | NPU |
+| [yolo11n-hand-pose-npu1](https://github.com/ultralytics/ultralytics) | CV | 3.2M | 姿态检测 | [mode_yolo11n-hand-pose-npu1.json](projects/llm_framework/main_yolo/mode_yolo11n-hand-pose-npu1.json) | NPU |
 | [depth-anything-ax630c](https://github.com/DepthAnything/Depth-Anything-V2) | CV | 29M | 单目深度估计 | [mode_depth-anything-ax630c.json](projects/llm_framework/main_depth_anything/mode_depth-anything-ax630c.json) | NPU |
+| [depth-anything-npu1-ax630c](https://github.com/DepthAnything/Depth-Anything-V2) | CV | 29M | 单目深度估计 | [mode_depth-anything-npu1-ax630c.json](projects/llm_framework/main_depth_anything/mode_depth-anything-npu1-ax630c.json) | NPU |
 
 ## 环境要求 ##
 当前 StackFlow 的 AI 单元是建立在 AXERA 加速平台之上的，主要的芯片平台为 ax630c、ax650n。系统要求为 ubuntu。
diff --git a/projects/llm_framework/main_depth_anything/mode_depth-anything-npu1-ax630c.json b/projects/llm_framework/main_depth_anything/mode_depth-anything-npu1-ax630c.json
new file mode 100644
index 00000000..079ff56d
--- /dev/null
+++ b/projects/llm_framework/main_depth_anything/mode_depth-anything-npu1-ax630c.json
@@ -0,0 +1,25 @@
+{
+    "mode":"depth-anything-npu1-ax630c",
+    "type":"cv",
+    "homepage":"https://github.com/DepthAnything/Depth-Anything-V2",
+    "compile_flage":"pulsar2 build --input cv/model/depth_anything_vits14-sim-384x256.onnx --config cv/config/depth_anything_u8.json --output_dir cv/axmodel/depth-anything-u8-npu1 --output_name depth_anything.axmodel --target_hardware AX620E --compiler.check 0 --npu_mode NPU1",
+    "pulsar_version":"3.3-f0b32d03",
+    "capabilities":[
+        "Segmentation"
+    ],
+    "input_type":[
+        "cv.jpeg.base64"
+    ],
+    "output_type":[
+        "cv.jpeg.base64"
+    ],
+    "mode_param":{
+        "depth_anything_model":"depth-anything-npu1-ax630c.axmodel",
+        "model_type":"segment",
+        "img_h":256,
+        "img_w":384,
+        "npu_type":1
+    },
+    "mode_param_bak":{
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_yolo/SConstruct b/projects/llm_framework/main_yolo/SConstruct
index 7c588e53..54502a54 100644
--- a/projects/llm_framework/main_yolo/SConstruct
+++ b/projects/llm_framework/main_yolo/SConstruct
@@ -27,7 +27,7 @@ static_file = Glob('../static_lib/module-llm/libabsl_*')
 static_file = Glob('../static_lib/libopencv-4.6-aarch64-none/lib/lib*')
 STATIC_LIB += static_file * 2
 
-STATIC_FILES += Glob('mode_*.json')
+STATIC_FILES += Glob('models/mode_*.json')
 
 
 
diff --git a/projects/llm_framework/main_yolo/models/mode_yolo11n-hand-pose-npu1.json b/projects/llm_framework/main_yolo/models/mode_yolo11n-hand-pose-npu1.json
new file mode 100644
index 00000000..877e8f33
--- /dev/null
+++ b/projects/llm_framework/main_yolo/models/mode_yolo11n-hand-pose-npu1.json
@@ -0,0 +1,33 @@
+{
+    "mode":"yolo11n-hand-pose-npu1",
+    "type":"cv",
+    "homepage":"https://github.com/ultralytics/ultralytics",
+    "compile_flage":"pulsar2 build --input cv/model/yolo11n-hand-pose-cut.onnx --config cv/config/yolo11_hand_pose_config.json --output_dir cv/axmodel/yolo11n-hand-pose-npu1 --output_name yolo11n-hand-pose-npu1.axmodel --target_hardware AX620E --compiler.check 0 --npu_mode NPU1",
+    "pulsar_version":"3.3-f0b32d03",
+    "capabilities":[
+        "Pose"
+    ],
+    "input_type":[
+        "yolo.jpeg.base64"
+    ],
+    "output_type":[
+        "yolo.box",
+        "yolo.boxV2"
+    ],
+    "mode_param":{
+        "yolo_model":"yolo11n-hand-pose-npu1.axmodel",
+        "model_type":"pose",
+        "img_h":320,
+        "img_w":320,
+        "cls_num":1,
+        "point_num":21,
+        "pron_threshold":0.45,
+        "nms_threshold":0.45,
+        "npu_type":1,
+        "cls_name":[
+            "hand"
+        ]
+    },
+    "mode_param_bak":{
+    }
+}
diff --git a/projects/llm_framework/main_yolo/mode_yolo11n-hand-pose.json b/projects/llm_framework/main_yolo/models/mode_yolo11n-hand-pose.json
similarity index 100%
rename from projects/llm_framework/main_yolo/mode_yolo11n-hand-pose.json
rename to projects/llm_framework/main_yolo/models/mode_yolo11n-hand-pose.json
diff --git a/projects/llm_framework/main_yolo/models/mode_yolo11n-npu1.json b/projects/llm_framework/main_yolo/models/mode_yolo11n-npu1.json
new file mode 100644
index 00000000..76fee735
--- /dev/null
+++ b/projects/llm_framework/main_yolo/models/mode_yolo11n-npu1.json
@@ -0,0 +1,111 @@
+{
+    "mode":"yolo11n-npu1",
+    "type":"cv",
+    "homepage":"https://github.com/ultralytics/ultralytics",
+    "compile_flage":"pulsar2 build --input cv/model/yolo11n.onnx --config cv/config/yolo11_config.json --output_dir cv/axmodel/yolo11n-npu1 --output_name yolo11n-npu1.axmodel --target_hardware AX620E --compiler.check 0 --npu_mode NPU1",
+    "pulsar_version":"3.3-f0b32d03",
+    "capabilities":[
+        "Detection"
+    ],
+    "input_type":[
+        "yolo.jpeg.base64"
+    ],
+    "output_type":[
+        "yolo.box",
+        "yolo.boxV2"
+    ],
+    "mode_param":{
+        "yolo_model":"yolo11n-npu1.axmodel",
+        "model_type":"detect",
+        "img_h":320,
+        "img_w":320,
+        "cls_num":80,
+        "pron_threshold":0.45,
+        "nms_threshold":0.45,
+        "npu_type":1,
+        "cls_name":[
+            "person",
+            "bicycle",
+            "car",
+            "motorcycle",
+            "airplane",
+            "bus",
+            "train",
+            "truck",
+            "boat",
+            "traffic light",
+            "fire hydrant",
+            "stop sign",
+            "parking meter",
+            "bench",
+            "bird",
+            "cat",
+            "dog",
+            "horse",
+            "sheep",
+            "cow",
+            "elephant",
+            "bear",
+            "zebra",
+            "giraffe",
+            "backpack",
+            "umbrella",
+            "handbag",
+            "tie",
+            "suitcase",
+            "frisbee",
+            "skis",
+            "snowboard",
+            "sports ball",
+            "kite",
+            "baseball bat",
+            "baseball glove",
+            "skateboard",
+            "surfboard",
+            "tennis racket",
+            "bottle",
+            "wine glass",
+            "cup",
+            "fork",
+            "knife",
+            "spoon",
+            "bowl",
+            "banana",
+            "apple",
+            "sandwich",
+            "orange",
+            "broccoli",
+            "carrot",
+            "hot dog",
+            "pizza",
+            "donut",
+            "cake",
+            "chair",
+            "couch",
+            "potted plant",
+            "bed",
+            "dining table",
+            "toilet",
+            "tv",
+            "laptop",
+            "mouse",
+            "remote",
+            "keyboard",
+            "cell phone",
+            "microwave",
+            "oven",
+            "toaster",
+            "sink",
+            "refrigerator",
+            "book",
+            "clock",
+            "vase",
+            "scissors",
+            "teddy bear",
+            "hair drier",
+            "toothbrush"
+        ]
+    },
+    "mode_param_bak":{
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_yolo/models/mode_yolo11n-pose-npu1.json b/projects/llm_framework/main_yolo/models/mode_yolo11n-pose-npu1.json
new file mode 100644
index 00000000..8a7cbeaf
--- /dev/null
+++ b/projects/llm_framework/main_yolo/models/mode_yolo11n-pose-npu1.json
@@ -0,0 +1,33 @@
+{
+    "mode":"yolo11n-pose-npu1",
+    "type":"cv",
+    "homepage":"https://github.com/ultralytics/ultralytics",
+    "compile_flage":"pulsar2 build --input cv/model/yolo11n-pose-cut.onnx --config cv/config/yolo11_pose_config.json --output_dir cv/axmodel/yolo11n-pose-npu1 --output_name yolo11n-pose-npu1.axmodel --target_hardware AX620E --compiler.check 0 --npu_mode NPU1",
+    "pulsar_version":"3.3-f0b32d03",
+    "capabilities":[
+        "Pose"
+    ],
+    "input_type":[
+        "yolo.jpeg.base64"
+    ],
+    "output_type":[
+        "yolo.box",
+        "yolo.boxV2"
+    ],
+    "mode_param":{
+        "yolo_model":"yolo11n-pose-npu1.axmodel",
+        "model_type":"pose",
+        "img_h":320,
+        "img_w":320,
+        "cls_num":1,
+        "point_num":17,
+        "pron_threshold":0.45,
+        "nms_threshold":0.45,
+        "npu_type":1,
+        "cls_name":[
+            "person"
+        ]
+    },
+    "mode_param_bak":{
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_yolo/mode_yolo11n-pose.json b/projects/llm_framework/main_yolo/models/mode_yolo11n-pose.json
similarity index 100%
rename from projects/llm_framework/main_yolo/mode_yolo11n-pose.json
rename to projects/llm_framework/main_yolo/models/mode_yolo11n-pose.json
diff --git a/projects/llm_framework/main_yolo/models/mode_yolo11n-seg-npu1.json b/projects/llm_framework/main_yolo/models/mode_yolo11n-seg-npu1.json
new file mode 100644
index 00000000..a0b8df2d
--- /dev/null
+++ b/projects/llm_framework/main_yolo/models/mode_yolo11n-seg-npu1.json
@@ -0,0 +1,111 @@
+{
+    "mode":"yolo11n-seg-npu1",
+    "type":"cv",
+    "homepage":"https://github.com/ultralytics/ultralytics",
+    "compile_flage":"pulsar2 build --input cv/model/yolo11n-seg-cut.onnx --config cv/config/yolo11_seg_config.json --output_dir cv/axmodel/yolo11n-seg-npu1 --output_name yolo11n-seg-npu1.axmodel --target_hardware AX620E --compiler.check 0 --npu_mode NPU1",
+    "pulsar_version":"3.3-f0b32d03",
+    "capabilities":[
+        "Segmentation"
+    ],
+    "input_type":[
+        "yolo.jpeg.base64"
+    ],
+    "output_type":[
+        "yolo.box",
+        "yolo.boxV2"
+    ],
+    "mode_param":{
+        "yolo_model":"yolo11n-seg-npu1.axmodel",
+        "model_type":"segment",
+        "img_h":320,
+        "img_w":320,
+        "cls_num":80,
+        "pron_threshold":0.45,
+        "nms_threshold":0.45,
+        "npu_type":1,
+        "cls_name":[
+            "person",
+            "bicycle",
+            "car",
+            "motorcycle",
+            "airplane",
+            "bus",
+            "train",
+            "truck",
+            "boat",
+            "traffic light",
+            "fire hydrant",
+            "stop sign",
+            "parking meter",
+            "bench",
+            "bird",
+            "cat",
+            "dog",
+            "horse",
+            "sheep",
+            "cow",
+            "elephant",
+            "bear",
+            "zebra",
+            "giraffe",
+            "backpack",
+            "umbrella",
+            "handbag",
+            "tie",
+            "suitcase",
+            "frisbee",
+            "skis",
+            "snowboard",
+            "sports ball",
+            "kite",
+            "baseball bat",
+            "baseball glove",
+            "skateboard",
+            "surfboard",
+            "tennis racket",
+            "bottle",
+            "wine glass",
+            "cup",
+            "fork",
+            "knife",
+            "spoon",
+            "bowl",
+            "banana",
+            "apple",
+            "sandwich",
+            "orange",
+            "broccoli",
+            "carrot",
+            "hot dog",
+            "pizza",
+            "donut",
+            "cake",
+            "chair",
+            "couch",
+            "potted plant",
+            "bed",
+            "dining table",
+            "toilet",
+            "tv",
+            "laptop",
+            "mouse",
+            "remote",
+            "keyboard",
+            "cell phone",
+            "microwave",
+            "oven",
+            "toaster",
+            "sink",
+            "refrigerator",
+            "book",
+            "clock",
+            "vase",
+            "scissors",
+            "teddy bear",
+            "hair drier",
+            "toothbrush"
+        ]
+    },
+    "mode_param_bak":{
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_yolo/mode_yolo11n-seg.json b/projects/llm_framework/main_yolo/models/mode_yolo11n-seg.json
similarity index 100%
rename from projects/llm_framework/main_yolo/mode_yolo11n-seg.json
rename to projects/llm_framework/main_yolo/models/mode_yolo11n-seg.json
diff --git a/projects/llm_framework/main_yolo/mode_yolo11n.json b/projects/llm_framework/main_yolo/models/mode_yolo11n.json
similarity index 100%
rename from projects/llm_framework/main_yolo/mode_yolo11n.json
rename to projects/llm_framework/main_yolo/models/mode_yolo11n.json

From 629e822e0aad37af3c130b042959f8cb6dd4512d Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 27 Jun 2025 18:25:09 +0800
Subject: [PATCH 21/79] [update] update docs

---
 .../llm_camera_en.md                          |  1 +
 .../llm_camera_zh.md                          |  1 +
 doc/projects_llm_framework_doc/llm_vlm_en.md  | 40 +++++++++++++++++--
 doc/projects_llm_framework_doc/llm_vlm_zh.md  | 40 +++++++++++++++++--
 4 files changed, 74 insertions(+), 8 deletions(-)

diff --git a/doc/projects_llm_framework_doc/llm_camera_en.md b/doc/projects_llm_framework_doc/llm_camera_en.md
index 25a0f05c..ddac27e0 100644
--- a/doc/projects_llm_framework_doc/llm_camera_en.md
+++ b/doc/projects_llm_framework_doc/llm_camera_en.md
@@ -37,6 +37,7 @@ Send JSON:
 - enoutput: Whether to enable user result output. If you do not need to obtain camera images, do not enable this parameter, as the video stream will increase the communication pressure on the channel.
 - enable_webstream: Whether to enable webstream output, webstream will listen on tcp:8989 port, and once a client connection is received, it will push jpeg images in HTTP protocol multipart/x-mixed-replace type.
 - rtsp: Whether to enable rtsp stream output, rtsp will establish an RTSP TCP server at rtsp://{DevIp}:8554/axstream0, and you can pull the video stream from this port using the RTSP protocol. The video stream format is 1280x720 H265. Note that this video stream is only valid on the AX630C MIPI camera, and the UVC camera cannot use RTSP.
+- VinParam.bAiispEnable: Whether to enable AI-ISP, enabled by default. Set to 0 to disable, only valid when using AX630C MIPI camera.
 
 Response JSON:
 
diff --git a/doc/projects_llm_framework_doc/llm_camera_zh.md b/doc/projects_llm_framework_doc/llm_camera_zh.md
index 5675df39..4bb3cb6c 100644
--- a/doc/projects_llm_framework_doc/llm_camera_zh.md
+++ b/doc/projects_llm_framework_doc/llm_camera_zh.md
@@ -37,6 +37,7 @@
 - enoutput：是否起用用户结果输出。如果不需要获取摄像头图片，请不要开启该参数，视频流会增加信道的通信压力。
 - enable_webstream：是否启用 webstream 流输出，webstream 会监听 tcp:8989 端口，一但收到客户端连接，将会以 HTTP 协议 multipart/x-mixed-replace 类型推送 jpeg 图片。
 - rtsp：是否启用 rtsp 流输出，rtsp 会建立一个 rtsp://{DevIp}:8554/axstream0 RTSP TCP 服务端，可使用RTSP 协议向该端口拉取视频流。视频流的格式为 1280x720 H265。注意，该视频流只在 AX630C MIPI 摄像头上有效，UVC 摄像头无法使用 RTSP。
+- VinParam.bAiispEnable：是否开启 AI-ISP，默认开启。关闭为 0，仅在使用 AX630C MIPI 摄像头时有效。
 
 响应 json：
 
diff --git a/doc/projects_llm_framework_doc/llm_vlm_en.md b/doc/projects_llm_framework_doc/llm_vlm_en.md
index 76196696..bbaa0d9c 100644
--- a/doc/projects_llm_framework_doc/llm_vlm_en.md
+++ b/doc/projects_llm_framework_doc/llm_vlm_en.md
@@ -15,7 +15,7 @@ Send the following JSON:
   "action": "setup",
   "object": "vlm.setup",
   "data": {
-    "model": "internvl2.5-1B-ax630c",
+    "model": "internvl2.5-1B-364-ax630c",
     "response_format": "vlm.utf-8.stream",
     "input": "vlm.utf-8",
     "enoutput": true,
@@ -29,7 +29,7 @@ Send the following JSON:
 - work_id: Set to `vlm` when configuring the unit.
 - action: The method being called is `setup`.
 - object: Data type being transferred is `vlm.setup`.
-- model: The model used is `internvl2.5-1B-ax630c`, a multimodal model.
+- model: The model used is `internvl2.5-1B-364-ax630c`, a multimodal model.
 - response_format: The output is in `vlm.utf-8.stream`, a UTF-8 stream format.
 - input: The input is `vlm.utf-8`, representing user input.
 - enoutput: Specifies whether to enable user output.
@@ -250,7 +250,7 @@ Example:
   "action": "setup",
   "object": "vlm.setup",
   "data": {
-    "model": "internvl2.5-1B-ax630c",
+    "model": "internvl2.5-1B-364-ax630c",
     "response_format": "vlm.utf-8.stream",
     "input": [
       "vlm.utf-8",
@@ -264,6 +264,38 @@ Example:
 }
 ```
 
+Linking the Output of the llm-camera Unit.
+
+Sending JSON:
+
+```json
+{
+  "request_id": "3",
+  "work_id": "vlm.1003",
+  "action": "link",
+  "object": "work_id",
+  "data": "camera.1000"
+}
+```
+
+Response JSON:
+
+```json
+{
+    "created": 1750992545,
+    "data": "None",
+    "error": {
+        "code": 0,
+        "message": ""
+    },
+    "object": "None",
+    "request_id": "3",
+    "work_id": "vlm.1003"
+}
+```
+
+> **Ensure that the camera is properly configured and ready for operation when performing the link action. If using the AX630C MIPI camera, configure it in AI-ISP disabled mode during the initialization of llm-camera.**
+
 ## unlink
 
 Unlink units.
@@ -447,7 +479,7 @@ Response JSON:
       "vlm.utf-8",
       "kws.1000"
     ],
-    "model": "internvl2.5-1B-ax630c",
+    "model": "internvl2.5-1B-364-ax630c",
     "response_format": "vlm.utf-8.stream"
   },
   "error": {
diff --git a/doc/projects_llm_framework_doc/llm_vlm_zh.md b/doc/projects_llm_framework_doc/llm_vlm_zh.md
index 161f48c1..ce797e5e 100644
--- a/doc/projects_llm_framework_doc/llm_vlm_zh.md
+++ b/doc/projects_llm_framework_doc/llm_vlm_zh.md
@@ -15,7 +15,7 @@
   "action": "setup",
   "object": "vlm.setup",
   "data": {
-    "model": "internvl2.5-1B-ax630c",
+    "model": "internvl2.5-1B-364-ax630c",
     "response_format": "vlm.utf-8.stream",
     "input": "vlm.utf-8",
     "enoutput": true,
@@ -29,7 +29,7 @@
 - work_id：配置单元时，为 `vlm`。
 - action：调用的方法为 `setup`。
 - object：传输的数据类型为 `vlm.setup`。
-- model：使用的模型为 `internvl2.5-1B-ax630c` 多模态模型。
+- model：使用的模型为 `internvl2.5-1B-364-ax630c` 多模态模型。
 - response_format：返回结果为 `vlm.utf-8.stream`, utf-8 的流式输出。
 - input：输入的为 `vlm.utf-8`,代表的是从用户输入。
 - enoutput：是否起用用户结果输出。
@@ -248,7 +248,7 @@ error::code 为 0 表示执行成功。
   "action": "setup",
   "object": "vlm.setup",
   "data": {
-    "model": "internvl2.5-1B-ax630c",
+    "model": "internvl2.5-1B-364-ax630c",
     "response_format": "vlm.utf-8.stream",
     "input": [
       "vlm.utf-8",
@@ -262,6 +262,38 @@ error::code 为 0 表示执行成功。
 }
 ```
 
+链接 llm-camera 单元的输出。
+
+发送 json：
+
+```json
+{
+  "request_id": "3",
+  "work_id": "vlm.1003",
+  "action": "link",
+  "object": "work_id",
+  "data": "camera.1000"
+}
+```
+
+响应 json：
+
+```json
+{
+    "created": 1750992545,
+    "data": "None",
+    "error": {
+        "code": 0,
+        "message": ""
+    },
+    "object": "None",
+    "request_id": "3",
+    "work_id": "vlm.1003"
+}
+```
+
+> **link 时必须保证 camera 此时已经配置好进入工作状态。当同时使用 AX630C MIPI 摄像头，需要在 llm-camera 初始化的时候配置为 AI-ISP 关闭模式。**
+
 ## unlink
 
 取消链接。
@@ -445,7 +477,7 @@ error::code 为 0 表示执行成功。
       "vlm.utf-8",
       "kws.1000"
     ],
-    "model": "internvl2.5-1B-ax630c",
+    "model": "internvl2.5-1B-364-ax630c",
     "response_format": "vlm.utf-8.stream"
   },
   "error": {

From d29e074afcac048fbdeb7c34b300c27fe6ca2a12 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 27 Jun 2025 18:30:33 +0800
Subject: [PATCH 22/79] [update] update ax650 model config, melotts model.

---
 .../mode_deepseek-r1-1.5B-Int4-ax650.json     |  35 +++
 .../mode_deepseek-r1-7B-Int4-ax650.json       |  35 +++
 .../models/mode_qwen2.5-0.5B-Int4-ax650.json  |  35 +++
 .../models/mode_qwen2.5-1.5B-Int4-ax650.json  |  35 +++
 .../models/mode_qwen2.5-3B-Int4-ax650.json    |  35 +++
 .../models/mode_qwen2.5-7B-Int4-ax650.json    |  35 +++
 .../tokenizer_deepseek-r1-1.5B-Int4-ax650.py  | 131 +++++++++
 .../tokenizer_deepseek-r1-7B-Int4-ax650.py    | 131 +++++++++
 .../tokenizer_qwen2.5-0.5B-Int4-ax650.py      | 131 +++++++++
 .../tokenizer_qwen2.5-1.5B-Int4-ax650.py      | 131 +++++++++
 .../tokenizer_qwen2.5-3B-Int4-ax650.py        | 131 +++++++++
 .../tokenizer_qwen2.5-7B-Int4-ax650.py        | 131 +++++++++
 .../llm_framework/main_melotts/SConstruct     |   2 +-
 .../{ => models}/mode_melotts-en-au.json      |   0
 .../{ => models}/mode_melotts-en-br.json      |   0
 .../{ => models}/mode_melotts-en-default.json |   0
 .../{ => models}/mode_melotts-en-india.json   |   0
 .../{ => models}/mode_melotts-en-us.json      |   0
 .../{ => models}/mode_melotts-es-es.json      |   0
 .../{ => models}/mode_melotts-ja-jp.json      |   0
 .../{ => models}/mode_melotts-zh-cn.json      |   0
 .../models/mode_internvl2.5-1B-448-ax650.json |  35 +++
 .../models/mode_smolvlm-256M-ax650.json       |  35 +++
 .../models/mode_smolvlm-500M-ax650.json       |  37 +++
 .../tokenizer_internvl2.5-1B-448-ax650.py     | 138 ++++++++++
 .../scripts/tokenizer_smolvlm-256M-ax650.py   | 248 ++++++++++++++++++
 .../llm_framework/main_whisper/SConstruct     |   2 +-
 .../models/mode_whisper-base-ax650.json       |  42 +++
 .../{ => models}/mode_whisper-base.json       |   0
 .../models/mode_whisper-small-ax650.json      |  42 +++
 .../{ => models}/mode_whisper-small.json      |   0
 .../models/mode_whisper-tiny-ax650.json       |  42 +++
 .../{ => models}/mode_whisper-tiny.json       |   0
 33 files changed, 1617 insertions(+), 2 deletions(-)
 create mode 100644 projects/llm_framework/main_llm/models/mode_deepseek-r1-1.5B-Int4-ax650.json
 create mode 100644 projects/llm_framework/main_llm/models/mode_deepseek-r1-7B-Int4-ax650.json
 create mode 100644 projects/llm_framework/main_llm/models/mode_qwen2.5-0.5B-Int4-ax650.json
 create mode 100644 projects/llm_framework/main_llm/models/mode_qwen2.5-1.5B-Int4-ax650.json
 create mode 100644 projects/llm_framework/main_llm/models/mode_qwen2.5-3B-Int4-ax650.json
 create mode 100644 projects/llm_framework/main_llm/models/mode_qwen2.5-7B-Int4-ax650.json
 create mode 100644 projects/llm_framework/main_llm/scripts/tokenizer_deepseek-r1-1.5B-Int4-ax650.py
 create mode 100644 projects/llm_framework/main_llm/scripts/tokenizer_deepseek-r1-7B-Int4-ax650.py
 create mode 100644 projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-0.5B-Int4-ax650.py
 create mode 100644 projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-1.5B-Int4-ax650.py
 create mode 100644 projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-3B-Int4-ax650.py
 create mode 100644 projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-7B-Int4-ax650.py
 rename projects/llm_framework/main_melotts/{ => models}/mode_melotts-en-au.json (100%)
 rename projects/llm_framework/main_melotts/{ => models}/mode_melotts-en-br.json (100%)
 rename projects/llm_framework/main_melotts/{ => models}/mode_melotts-en-default.json (100%)
 rename projects/llm_framework/main_melotts/{ => models}/mode_melotts-en-india.json (100%)
 rename projects/llm_framework/main_melotts/{ => models}/mode_melotts-en-us.json (100%)
 rename projects/llm_framework/main_melotts/{ => models}/mode_melotts-es-es.json (100%)
 rename projects/llm_framework/main_melotts/{ => models}/mode_melotts-ja-jp.json (100%)
 rename projects/llm_framework/main_melotts/{ => models}/mode_melotts-zh-cn.json (100%)
 create mode 100644 projects/llm_framework/main_vlm/models/mode_internvl2.5-1B-448-ax650.json
 create mode 100644 projects/llm_framework/main_vlm/models/mode_smolvlm-256M-ax650.json
 create mode 100644 projects/llm_framework/main_vlm/models/mode_smolvlm-500M-ax650.json
 create mode 100644 projects/llm_framework/main_vlm/scripts/tokenizer_internvl2.5-1B-448-ax650.py
 create mode 100644 projects/llm_framework/main_vlm/scripts/tokenizer_smolvlm-256M-ax650.py
 create mode 100644 projects/llm_framework/main_whisper/models/mode_whisper-base-ax650.json
 rename projects/llm_framework/main_whisper/{ => models}/mode_whisper-base.json (100%)
 create mode 100644 projects/llm_framework/main_whisper/models/mode_whisper-small-ax650.json
 rename projects/llm_framework/main_whisper/{ => models}/mode_whisper-small.json (100%)
 create mode 100644 projects/llm_framework/main_whisper/models/mode_whisper-tiny-ax650.json
 rename projects/llm_framework/main_whisper/{ => models}/mode_whisper-tiny.json (100%)

diff --git a/projects/llm_framework/main_llm/models/mode_deepseek-r1-1.5B-Int4-ax650.json b/projects/llm_framework/main_llm/models/mode_deepseek-r1-1.5B-Int4-ax650.json
new file mode 100644
index 00000000..8eeeb580
--- /dev/null
+++ b/projects/llm_framework/main_llm/models/mode_deepseek-r1-1.5B-Int4-ax650.json
@@ -0,0 +1,35 @@
+{
+    "mode":"deepseek-r1-1.5B-Int4-ax650",
+    "type":"llm",
+    "homepage":"https://huggingface.co/AXERA-TECH/DeepSeek-R1-Distill-Qwen-1.5B-GPTQ-Int4",
+    "capabilities":[
+        "text_generation",
+        "chat"
+    ],
+    "input_type":[
+        "llm.utf-8",
+        "llm.utf-8.stream",
+        "llm.chat_completion",
+        "llm.chat_completion.stream"
+    ],
+    "output_type":[
+        "llm.utf-8",
+        "llm.utf-8.stream"
+    ],
+    "mode_param":{
+        "tokenizer_type":2,
+        "filename_tokenizer_model":"http://localhost:8080",
+        "filename_tokens_embed":"model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel":"qwen2_post.axmodel",
+        "template_filename_axmodel":"qwen2_p128_l%d_together.axmodel",
+        "b_use_topk":false,
+        "b_bos":false,
+        "b_eos":false,
+        "axmodel_num":28,
+        "tokens_embed_num":151936,
+        "tokens_embed_size":1536,
+        "b_use_mmap_load_embed":true,
+        "b_dynamic_load_axmodel_layer":false,
+        "ext_scripts":["tokenizer_deepseek-r1-1.5B-Int4-ax650.py"]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_llm/models/mode_deepseek-r1-7B-Int4-ax650.json b/projects/llm_framework/main_llm/models/mode_deepseek-r1-7B-Int4-ax650.json
new file mode 100644
index 00000000..757683c0
--- /dev/null
+++ b/projects/llm_framework/main_llm/models/mode_deepseek-r1-7B-Int4-ax650.json
@@ -0,0 +1,35 @@
+{
+    "mode":"deepseek-r1-7B-Int4-ax650",
+    "type":"llm",
+    "homepage":"https://huggingface.co/AXERA-TECH/DeepSeek-R1-Distill-Qwen-7B-GPTQ-Int4",
+    "capabilities":[
+        "text_generation",
+        "chat"
+    ],
+    "input_type":[
+        "llm.utf-8",
+        "llm.utf-8.stream",
+        "llm.chat_completion",
+        "llm.chat_completion.stream"
+    ],
+    "output_type":[
+        "llm.utf-8",
+        "llm.utf-8.stream"
+    ],
+    "mode_param":{
+        "tokenizer_type":2,
+        "filename_tokenizer_model":"http://localhost:8080",
+        "filename_tokens_embed":"model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel":"qwen2_post.axmodel",
+        "template_filename_axmodel":"qwen2_p128_l%d_together.axmodel",
+        "b_use_topk":false,
+        "b_bos":false,
+        "b_eos":false,
+        "axmodel_num":28,
+        "tokens_embed_num":152064,
+        "tokens_embed_size":3584,
+        "b_use_mmap_load_embed":true,
+        "b_dynamic_load_axmodel_layer":false,
+        "ext_scripts":["tokenizer_deepseek-r1-7B-Int4-ax650.py"]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_llm/models/mode_qwen2.5-0.5B-Int4-ax650.json b/projects/llm_framework/main_llm/models/mode_qwen2.5-0.5B-Int4-ax650.json
new file mode 100644
index 00000000..528ae793
--- /dev/null
+++ b/projects/llm_framework/main_llm/models/mode_qwen2.5-0.5B-Int4-ax650.json
@@ -0,0 +1,35 @@
+{
+    "mode":"qwen2.5-0.5B-Int4-ax650",
+    "type":"llm",
+    "homepage":"https://huggingface.co/AXERA-TECH/Qwen2.5-0.5B-Instruct-GPTQ-Int4",
+    "capabilities":[
+        "text_generation",
+        "chat"
+    ],
+    "input_type":[
+        "llm.utf-8",
+        "llm.utf-8.stream",
+        "llm.chat_completion",
+        "llm.chat_completion.stream"
+    ],
+    "output_type":[
+        "llm.utf-8",
+        "llm.utf-8.stream"
+    ],
+    "mode_param":{
+        "tokenizer_type":2,
+        "filename_tokenizer_model":"http://localhost:8080",
+        "filename_tokens_embed":"model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel":"qwen2_post.axmodel",
+        "template_filename_axmodel":"qwen2_p128_l%d_together.axmodel",
+        "b_use_topk":false,
+        "b_bos":false,
+        "b_eos":false,
+        "axmodel_num":24,
+        "tokens_embed_num":151936,
+        "tokens_embed_size":896,
+        "b_use_mmap_load_embed":true,
+        "b_dynamic_load_axmodel_layer":false,
+        "ext_scripts":["tokenizer_qwen2.5-0.5B-Int4-ax650.py"]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_llm/models/mode_qwen2.5-1.5B-Int4-ax650.json b/projects/llm_framework/main_llm/models/mode_qwen2.5-1.5B-Int4-ax650.json
new file mode 100644
index 00000000..6e50ca2c
--- /dev/null
+++ b/projects/llm_framework/main_llm/models/mode_qwen2.5-1.5B-Int4-ax650.json
@@ -0,0 +1,35 @@
+{
+    "mode":"qwen2.5-1.5B-Int4-ax650",
+    "type":"llm",
+    "homepage":"https://huggingface.co/AXERA-TECH/Qwen2.5-1.5B-Instruct-GPTQ-Int4",
+    "capabilities":[
+        "text_generation",
+        "chat"
+    ],
+    "input_type":[
+        "llm.utf-8",
+        "llm.utf-8.stream",
+        "llm.chat_completion",
+        "llm.chat_completion.stream"
+    ],
+    "output_type":[
+        "llm.utf-8",
+        "llm.utf-8.stream"
+    ],
+    "mode_param":{
+        "tokenizer_type":2,
+        "filename_tokenizer_model":"http://localhost:8080",
+        "filename_tokens_embed":"model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel":"qwen2_post.axmodel",
+        "template_filename_axmodel":"qwen2_p128_l%d_together.axmodel",
+        "b_use_topk":false,
+        "b_bos":false,
+        "b_eos":false,
+        "axmodel_num":28,
+        "tokens_embed_num":151936,
+        "tokens_embed_size":1536,
+        "b_use_mmap_load_embed":true,
+        "b_dynamic_load_axmodel_layer":false,
+        "ext_scripts":["tokenizer_qwen2.5-1.5B-Int4-ax650.py"]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_llm/models/mode_qwen2.5-3B-Int4-ax650.json b/projects/llm_framework/main_llm/models/mode_qwen2.5-3B-Int4-ax650.json
new file mode 100644
index 00000000..e11b7d82
--- /dev/null
+++ b/projects/llm_framework/main_llm/models/mode_qwen2.5-3B-Int4-ax650.json
@@ -0,0 +1,35 @@
+{
+    "mode":"qwen2.5-3B-Int4-ax650",
+    "type":"llm",
+    "homepage":"https://huggingface.co/AXERA-TECH/Qwen2.5-3B-Instruct-GPTQ-Int4",
+    "capabilities":[
+        "text_generation",
+        "chat"
+    ],
+    "input_type":[
+        "llm.utf-8",
+        "llm.utf-8.stream",
+        "llm.chat_completion",
+        "llm.chat_completion.stream"
+    ],
+    "output_type":[
+        "llm.utf-8",
+        "llm.utf-8.stream"
+    ],
+    "mode_param":{
+        "tokenizer_type":2,
+        "filename_tokenizer_model":"http://localhost:8080",
+        "filename_tokens_embed":"model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel":"qwen2_post.axmodel",
+        "template_filename_axmodel":"qwen2_p128_l%d_together.axmodel",
+        "b_use_topk":false,
+        "b_bos":false,
+        "b_eos":false,
+        "axmodel_num":36,
+        "tokens_embed_num":151936,
+        "tokens_embed_size":2048,
+        "b_use_mmap_load_embed":true,
+        "b_dynamic_load_axmodel_layer":false,
+        "ext_scripts":["tokenizer_qwen2.5-3B-Int4-ax650.py"]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_llm/models/mode_qwen2.5-7B-Int4-ax650.json b/projects/llm_framework/main_llm/models/mode_qwen2.5-7B-Int4-ax650.json
new file mode 100644
index 00000000..8da02c1b
--- /dev/null
+++ b/projects/llm_framework/main_llm/models/mode_qwen2.5-7B-Int4-ax650.json
@@ -0,0 +1,35 @@
+{
+    "mode":"qwen2.5-7B-Int4-ax650",
+    "type":"llm",
+    "homepage":"https://huggingface.co/AXERA-TECH/Qwen2.5-7B-Instruct-GPTQ-Int4",
+    "capabilities":[
+        "text_generation",
+        "chat"
+    ],
+    "input_type":[
+        "llm.utf-8",
+        "llm.utf-8.stream",
+        "llm.chat_completion",
+        "llm.chat_completion.stream"
+    ],
+    "output_type":[
+        "llm.utf-8",
+        "llm.utf-8.stream"
+    ],
+    "mode_param":{
+        "tokenizer_type":2,
+        "filename_tokenizer_model":"http://localhost:8080",
+        "filename_tokens_embed":"model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel":"qwen2_post.axmodel",
+        "template_filename_axmodel":"qwen2_p128_l%d_together.axmodel",
+        "b_use_topk":false,
+        "b_bos":false,
+        "b_eos":false,
+        "axmodel_num":28,
+        "tokens_embed_num":152064,
+        "tokens_embed_size":3584,
+        "b_use_mmap_load_embed":true,
+        "b_dynamic_load_axmodel_layer":false,
+        "ext_scripts":["tokenizer_qwen2.5-7B-Int4-ax650.py"]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_llm/scripts/tokenizer_deepseek-r1-1.5B-Int4-ax650.py b/projects/llm_framework/main_llm/scripts/tokenizer_deepseek-r1-1.5B-Int4-ax650.py
new file mode 100644
index 00000000..3af58240
--- /dev/null
+++ b/projects/llm_framework/main_llm/scripts/tokenizer_deepseek-r1-1.5B-Int4-ax650.py
@@ -0,0 +1,131 @@
+from transformers import AutoTokenizer, PreTrainedTokenizerFast
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+
+class Tokenizer_Http():
+
+    def __init__(self, model_id):
+        self.tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+    def encode(self, prompt, content):
+        messages = [
+            {"role": "system", "content": content},
+            {"role": "user", "content": prompt}
+        ]
+        text = self.tokenizer.apply_chat_template(
+            messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        print(text)
+        token_ids = self.tokenizer.encode(text)
+        return token_ids
+
+    def decode(self, token_ids):
+        return self.tokenizer.decode(token_ids)
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+    
+    @property
+    def bos_token(self):
+        return self.tokenizer.bos_token
+
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+
+class Request(BaseHTTPRequestHandler):
+    #通过类继承，新定义类
+    timeout = 5
+    server_version = 'Apache'
+
+    def do_GET(self):
+        print(self.path)
+        #在新类中定义get的内容（当客户端向该服务端使用get请求时，本服务端将如下运行）
+        self.send_response(200)
+        self.send_header("type", "get")  #设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == '/bos_id':
+            bos_id = tokenizer.bos_id
+            # print(bos_id)
+            # to json
+            if bos_id is None:
+                msg = json.dumps({'bos_id': -1})
+            else:
+                msg = json.dumps({'bos_id': bos_id})
+        elif self.path == '/eos_id':
+            eos_id = tokenizer.eos_id
+            if eos_id is None:
+                msg = json.dumps({'eos_id': -1})
+            else:
+                msg = json.dumps({'eos_id': eos_id})
+        else:
+            msg = 'error'
+
+        print(msg)
+        msg = str(msg).encode()  #转为str再转为byte格式
+
+        self.wfile.write(msg)  #将byte格式的信息返回给客户端
+
+    def do_POST(self):
+        #在新类中定义post的内容（当客户端向该服务端使用post请求时，本服务端将如下运行）
+        data = self.rfile.read(int(
+            self.headers['content-length']))  #获取从客户端传入的参数（byte格式）
+        data = data.decode()  #将byte格式转为str格式
+
+        self.send_response(200)
+        self.send_header("type", "post")  #设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == '/encode':
+            req = json.loads(data)
+            prompt = req['text']
+
+            token_ids = tokenizer.encode(prompt, args.content)
+            if token_ids is None:
+                msg = json.dumps({'token_ids': -1})
+            else:
+                msg = json.dumps({'token_ids': token_ids})
+
+        elif self.path == '/decode':
+            req = json.loads(data)
+            token_ids = req['token_ids']
+            text = tokenizer.decode(token_ids)
+            if text is None:
+                msg = json.dumps({'text': ""})
+            else:
+                msg = json.dumps({'text': text})
+        else:
+            msg = 'error'
+        print(msg)
+        msg = str(msg).encode()  #转为str再转为byte格式
+
+        self.wfile.write(msg)  #将byte格式的信息返回给客户端
+
+
+if __name__ == "__main__":
+
+    args = argparse.ArgumentParser()
+    args.add_argument('--host', type=str, default='localhost')
+    args.add_argument('--port', type=int, default=8080)
+    args.add_argument('--model_id', type=str, default='deepseek_tokenizer')
+    args.add_argument('--content', type=str, default='You are a helpful assistant.')
+    args = args.parse_args()
+
+    tokenizer = Tokenizer_Http(args.model_id)
+
+    # print(tokenizer.bos_id, tokenizer.bos_token, tokenizer.eos_id, tokenizer.eos_token)
+    # print(tokenizer.encode("hello world", args.content))
+
+    host = (args.host, args.port)  #设定地址与端口号，'localhost'等价于'127.0.0.1'
+    print('http://%s:%s' % host)
+    server = HTTPServer(host, Request)  #根据地址端口号和新定义的类，创建服务器实例
+    server.serve_forever()  #开启服务
diff --git a/projects/llm_framework/main_llm/scripts/tokenizer_deepseek-r1-7B-Int4-ax650.py b/projects/llm_framework/main_llm/scripts/tokenizer_deepseek-r1-7B-Int4-ax650.py
new file mode 100644
index 00000000..3af58240
--- /dev/null
+++ b/projects/llm_framework/main_llm/scripts/tokenizer_deepseek-r1-7B-Int4-ax650.py
@@ -0,0 +1,131 @@
+from transformers import AutoTokenizer, PreTrainedTokenizerFast
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+
+class Tokenizer_Http():
+
+    def __init__(self, model_id):
+        self.tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+    def encode(self, prompt, content):
+        messages = [
+            {"role": "system", "content": content},
+            {"role": "user", "content": prompt}
+        ]
+        text = self.tokenizer.apply_chat_template(
+            messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        print(text)
+        token_ids = self.tokenizer.encode(text)
+        return token_ids
+
+    def decode(self, token_ids):
+        return self.tokenizer.decode(token_ids)
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+    
+    @property
+    def bos_token(self):
+        return self.tokenizer.bos_token
+
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+
+class Request(BaseHTTPRequestHandler):
+    #通过类继承，新定义类
+    timeout = 5
+    server_version = 'Apache'
+
+    def do_GET(self):
+        print(self.path)
+        #在新类中定义get的内容（当客户端向该服务端使用get请求时，本服务端将如下运行）
+        self.send_response(200)
+        self.send_header("type", "get")  #设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == '/bos_id':
+            bos_id = tokenizer.bos_id
+            # print(bos_id)
+            # to json
+            if bos_id is None:
+                msg = json.dumps({'bos_id': -1})
+            else:
+                msg = json.dumps({'bos_id': bos_id})
+        elif self.path == '/eos_id':
+            eos_id = tokenizer.eos_id
+            if eos_id is None:
+                msg = json.dumps({'eos_id': -1})
+            else:
+                msg = json.dumps({'eos_id': eos_id})
+        else:
+            msg = 'error'
+
+        print(msg)
+        msg = str(msg).encode()  #转为str再转为byte格式
+
+        self.wfile.write(msg)  #将byte格式的信息返回给客户端
+
+    def do_POST(self):
+        #在新类中定义post的内容（当客户端向该服务端使用post请求时，本服务端将如下运行）
+        data = self.rfile.read(int(
+            self.headers['content-length']))  #获取从客户端传入的参数（byte格式）
+        data = data.decode()  #将byte格式转为str格式
+
+        self.send_response(200)
+        self.send_header("type", "post")  #设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == '/encode':
+            req = json.loads(data)
+            prompt = req['text']
+
+            token_ids = tokenizer.encode(prompt, args.content)
+            if token_ids is None:
+                msg = json.dumps({'token_ids': -1})
+            else:
+                msg = json.dumps({'token_ids': token_ids})
+
+        elif self.path == '/decode':
+            req = json.loads(data)
+            token_ids = req['token_ids']
+            text = tokenizer.decode(token_ids)
+            if text is None:
+                msg = json.dumps({'text': ""})
+            else:
+                msg = json.dumps({'text': text})
+        else:
+            msg = 'error'
+        print(msg)
+        msg = str(msg).encode()  #转为str再转为byte格式
+
+        self.wfile.write(msg)  #将byte格式的信息返回给客户端
+
+
+if __name__ == "__main__":
+
+    args = argparse.ArgumentParser()
+    args.add_argument('--host', type=str, default='localhost')
+    args.add_argument('--port', type=int, default=8080)
+    args.add_argument('--model_id', type=str, default='deepseek_tokenizer')
+    args.add_argument('--content', type=str, default='You are a helpful assistant.')
+    args = args.parse_args()
+
+    tokenizer = Tokenizer_Http(args.model_id)
+
+    # print(tokenizer.bos_id, tokenizer.bos_token, tokenizer.eos_id, tokenizer.eos_token)
+    # print(tokenizer.encode("hello world", args.content))
+
+    host = (args.host, args.port)  #设定地址与端口号，'localhost'等价于'127.0.0.1'
+    print('http://%s:%s' % host)
+    server = HTTPServer(host, Request)  #根据地址端口号和新定义的类，创建服务器实例
+    server.serve_forever()  #开启服务
diff --git a/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-0.5B-Int4-ax650.py b/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-0.5B-Int4-ax650.py
new file mode 100644
index 00000000..4fded69c
--- /dev/null
+++ b/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-0.5B-Int4-ax650.py
@@ -0,0 +1,131 @@
+from transformers import AutoTokenizer, PreTrainedTokenizerFast
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+
+class Tokenizer_Http():
+
+    def __init__(self, model_id):
+        self.tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+    def encode(self, prompt, content):
+        messages = [
+            {"role": "system", "content": content},
+            {"role": "user", "content": prompt}
+        ]
+        text = self.tokenizer.apply_chat_template(
+            messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        print(text)
+        token_ids = self.tokenizer.encode(text)
+        return token_ids
+
+    def decode(self, token_ids):
+        return self.tokenizer.decode(token_ids)
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+    
+    @property
+    def bos_token(self):
+        return self.tokenizer.bos_token
+
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+
+class Request(BaseHTTPRequestHandler):
+    #通过类继承，新定义类
+    timeout = 5
+    server_version = 'Apache'
+
+    def do_GET(self):
+        print(self.path)
+        #在新类中定义get的内容（当客户端向该服务端使用get请求时，本服务端将如下运行）
+        self.send_response(200)
+        self.send_header("type", "get")  #设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == '/bos_id':
+            bos_id = tokenizer.bos_id
+            # print(bos_id)
+            # to json
+            if bos_id is None:
+                msg = json.dumps({'bos_id': -1})
+            else:
+                msg = json.dumps({'bos_id': bos_id})
+        elif self.path == '/eos_id':
+            eos_id = tokenizer.eos_id
+            if eos_id is None:
+                msg = json.dumps({'eos_id': -1})
+            else:
+                msg = json.dumps({'eos_id': eos_id})
+        else:
+            msg = 'error'
+
+        print(msg)
+        msg = str(msg).encode()  #转为str再转为byte格式
+
+        self.wfile.write(msg)  #将byte格式的信息返回给客户端
+
+    def do_POST(self):
+        #在新类中定义post的内容（当客户端向该服务端使用post请求时，本服务端将如下运行）
+        data = self.rfile.read(int(
+            self.headers['content-length']))  #获取从客户端传入的参数（byte格式）
+        data = data.decode()  #将byte格式转为str格式
+
+        self.send_response(200)
+        self.send_header("type", "post")  #设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == '/encode':
+            req = json.loads(data)
+            prompt = req['text']
+
+            token_ids = tokenizer.encode(prompt, args.content)
+            if token_ids is None:
+                msg = json.dumps({'token_ids': -1})
+            else:
+                msg = json.dumps({'token_ids': token_ids})
+
+        elif self.path == '/decode':
+            req = json.loads(data)
+            token_ids = req['token_ids']
+            text = tokenizer.decode(token_ids)
+            if text is None:
+                msg = json.dumps({'text': ""})
+            else:
+                msg = json.dumps({'text': text})
+        else:
+            msg = 'error'
+        print(msg)
+        msg = str(msg).encode()  #转为str再转为byte格式
+
+        self.wfile.write(msg)  #将byte格式的信息返回给客户端
+
+
+if __name__ == "__main__":
+
+    args = argparse.ArgumentParser()
+    args.add_argument('--host', type=str, default='localhost')
+    args.add_argument('--port', type=int, default=8080)
+    args.add_argument('--model_id', type=str, default='qwen2.5_coder_tokenizer')
+    args.add_argument('--content', type=str, default='You are Qwen, created by Alibaba Cloud. You are a helpful assistant.')
+    args = args.parse_args()
+
+    tokenizer = Tokenizer_Http(args.model_id)
+
+    # print(tokenizer.bos_id, tokenizer.bos_token, tokenizer.eos_id, tokenizer.eos_token)
+    # print(tokenizer.encode("hello world", args.content))
+
+    host = (args.host, args.port)  #设定地址与端口号，'localhost'等价于'127.0.0.1'
+    print('http://%s:%s' % host)
+    server = HTTPServer(host, Request)  #根据地址端口号和新定义的类，创建服务器实例
+    server.serve_forever()  #开启服务
diff --git a/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-1.5B-Int4-ax650.py b/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-1.5B-Int4-ax650.py
new file mode 100644
index 00000000..4fded69c
--- /dev/null
+++ b/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-1.5B-Int4-ax650.py
@@ -0,0 +1,131 @@
+from transformers import AutoTokenizer, PreTrainedTokenizerFast
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+
+class Tokenizer_Http():
+
+    def __init__(self, model_id):
+        self.tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+    def encode(self, prompt, content):
+        messages = [
+            {"role": "system", "content": content},
+            {"role": "user", "content": prompt}
+        ]
+        text = self.tokenizer.apply_chat_template(
+            messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        print(text)
+        token_ids = self.tokenizer.encode(text)
+        return token_ids
+
+    def decode(self, token_ids):
+        return self.tokenizer.decode(token_ids)
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+    
+    @property
+    def bos_token(self):
+        return self.tokenizer.bos_token
+
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+
+class Request(BaseHTTPRequestHandler):
+    #通过类继承，新定义类
+    timeout = 5
+    server_version = 'Apache'
+
+    def do_GET(self):
+        print(self.path)
+        #在新类中定义get的内容（当客户端向该服务端使用get请求时，本服务端将如下运行）
+        self.send_response(200)
+        self.send_header("type", "get")  #设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == '/bos_id':
+            bos_id = tokenizer.bos_id
+            # print(bos_id)
+            # to json
+            if bos_id is None:
+                msg = json.dumps({'bos_id': -1})
+            else:
+                msg = json.dumps({'bos_id': bos_id})
+        elif self.path == '/eos_id':
+            eos_id = tokenizer.eos_id
+            if eos_id is None:
+                msg = json.dumps({'eos_id': -1})
+            else:
+                msg = json.dumps({'eos_id': eos_id})
+        else:
+            msg = 'error'
+
+        print(msg)
+        msg = str(msg).encode()  #转为str再转为byte格式
+
+        self.wfile.write(msg)  #将byte格式的信息返回给客户端
+
+    def do_POST(self):
+        #在新类中定义post的内容（当客户端向该服务端使用post请求时，本服务端将如下运行）
+        data = self.rfile.read(int(
+            self.headers['content-length']))  #获取从客户端传入的参数（byte格式）
+        data = data.decode()  #将byte格式转为str格式
+
+        self.send_response(200)
+        self.send_header("type", "post")  #设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == '/encode':
+            req = json.loads(data)
+            prompt = req['text']
+
+            token_ids = tokenizer.encode(prompt, args.content)
+            if token_ids is None:
+                msg = json.dumps({'token_ids': -1})
+            else:
+                msg = json.dumps({'token_ids': token_ids})
+
+        elif self.path == '/decode':
+            req = json.loads(data)
+            token_ids = req['token_ids']
+            text = tokenizer.decode(token_ids)
+            if text is None:
+                msg = json.dumps({'text': ""})
+            else:
+                msg = json.dumps({'text': text})
+        else:
+            msg = 'error'
+        print(msg)
+        msg = str(msg).encode()  #转为str再转为byte格式
+
+        self.wfile.write(msg)  #将byte格式的信息返回给客户端
+
+
+if __name__ == "__main__":
+
+    args = argparse.ArgumentParser()
+    args.add_argument('--host', type=str, default='localhost')
+    args.add_argument('--port', type=int, default=8080)
+    args.add_argument('--model_id', type=str, default='qwen2.5_coder_tokenizer')
+    args.add_argument('--content', type=str, default='You are Qwen, created by Alibaba Cloud. You are a helpful assistant.')
+    args = args.parse_args()
+
+    tokenizer = Tokenizer_Http(args.model_id)
+
+    # print(tokenizer.bos_id, tokenizer.bos_token, tokenizer.eos_id, tokenizer.eos_token)
+    # print(tokenizer.encode("hello world", args.content))
+
+    host = (args.host, args.port)  #设定地址与端口号，'localhost'等价于'127.0.0.1'
+    print('http://%s:%s' % host)
+    server = HTTPServer(host, Request)  #根据地址端口号和新定义的类，创建服务器实例
+    server.serve_forever()  #开启服务
diff --git a/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-3B-Int4-ax650.py b/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-3B-Int4-ax650.py
new file mode 100644
index 00000000..4fded69c
--- /dev/null
+++ b/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-3B-Int4-ax650.py
@@ -0,0 +1,131 @@
+from transformers import AutoTokenizer, PreTrainedTokenizerFast
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+
+class Tokenizer_Http():
+
+    def __init__(self, model_id):
+        self.tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+    def encode(self, prompt, content):
+        messages = [
+            {"role": "system", "content": content},
+            {"role": "user", "content": prompt}
+        ]
+        text = self.tokenizer.apply_chat_template(
+            messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        print(text)
+        token_ids = self.tokenizer.encode(text)
+        return token_ids
+
+    def decode(self, token_ids):
+        return self.tokenizer.decode(token_ids)
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+    
+    @property
+    def bos_token(self):
+        return self.tokenizer.bos_token
+
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+
+class Request(BaseHTTPRequestHandler):
+    #通过类继承，新定义类
+    timeout = 5
+    server_version = 'Apache'
+
+    def do_GET(self):
+        print(self.path)
+        #在新类中定义get的内容（当客户端向该服务端使用get请求时，本服务端将如下运行）
+        self.send_response(200)
+        self.send_header("type", "get")  #设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == '/bos_id':
+            bos_id = tokenizer.bos_id
+            # print(bos_id)
+            # to json
+            if bos_id is None:
+                msg = json.dumps({'bos_id': -1})
+            else:
+                msg = json.dumps({'bos_id': bos_id})
+        elif self.path == '/eos_id':
+            eos_id = tokenizer.eos_id
+            if eos_id is None:
+                msg = json.dumps({'eos_id': -1})
+            else:
+                msg = json.dumps({'eos_id': eos_id})
+        else:
+            msg = 'error'
+
+        print(msg)
+        msg = str(msg).encode()  #转为str再转为byte格式
+
+        self.wfile.write(msg)  #将byte格式的信息返回给客户端
+
+    def do_POST(self):
+        #在新类中定义post的内容（当客户端向该服务端使用post请求时，本服务端将如下运行）
+        data = self.rfile.read(int(
+            self.headers['content-length']))  #获取从客户端传入的参数（byte格式）
+        data = data.decode()  #将byte格式转为str格式
+
+        self.send_response(200)
+        self.send_header("type", "post")  #设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == '/encode':
+            req = json.loads(data)
+            prompt = req['text']
+
+            token_ids = tokenizer.encode(prompt, args.content)
+            if token_ids is None:
+                msg = json.dumps({'token_ids': -1})
+            else:
+                msg = json.dumps({'token_ids': token_ids})
+
+        elif self.path == '/decode':
+            req = json.loads(data)
+            token_ids = req['token_ids']
+            text = tokenizer.decode(token_ids)
+            if text is None:
+                msg = json.dumps({'text': ""})
+            else:
+                msg = json.dumps({'text': text})
+        else:
+            msg = 'error'
+        print(msg)
+        msg = str(msg).encode()  #转为str再转为byte格式
+
+        self.wfile.write(msg)  #将byte格式的信息返回给客户端
+
+
+if __name__ == "__main__":
+
+    args = argparse.ArgumentParser()
+    args.add_argument('--host', type=str, default='localhost')
+    args.add_argument('--port', type=int, default=8080)
+    args.add_argument('--model_id', type=str, default='qwen2.5_coder_tokenizer')
+    args.add_argument('--content', type=str, default='You are Qwen, created by Alibaba Cloud. You are a helpful assistant.')
+    args = args.parse_args()
+
+    tokenizer = Tokenizer_Http(args.model_id)
+
+    # print(tokenizer.bos_id, tokenizer.bos_token, tokenizer.eos_id, tokenizer.eos_token)
+    # print(tokenizer.encode("hello world", args.content))
+
+    host = (args.host, args.port)  #设定地址与端口号，'localhost'等价于'127.0.0.1'
+    print('http://%s:%s' % host)
+    server = HTTPServer(host, Request)  #根据地址端口号和新定义的类，创建服务器实例
+    server.serve_forever()  #开启服务
diff --git a/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-7B-Int4-ax650.py b/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-7B-Int4-ax650.py
new file mode 100644
index 00000000..4fded69c
--- /dev/null
+++ b/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-7B-Int4-ax650.py
@@ -0,0 +1,131 @@
+from transformers import AutoTokenizer, PreTrainedTokenizerFast
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+
+class Tokenizer_Http():
+
+    def __init__(self, model_id):
+        self.tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+    def encode(self, prompt, content):
+        messages = [
+            {"role": "system", "content": content},
+            {"role": "user", "content": prompt}
+        ]
+        text = self.tokenizer.apply_chat_template(
+            messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        print(text)
+        token_ids = self.tokenizer.encode(text)
+        return token_ids
+
+    def decode(self, token_ids):
+        return self.tokenizer.decode(token_ids)
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+    
+    @property
+    def bos_token(self):
+        return self.tokenizer.bos_token
+
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+
+class Request(BaseHTTPRequestHandler):
+    #通过类继承，新定义类
+    timeout = 5
+    server_version = 'Apache'
+
+    def do_GET(self):
+        print(self.path)
+        #在新类中定义get的内容（当客户端向该服务端使用get请求时，本服务端将如下运行）
+        self.send_response(200)
+        self.send_header("type", "get")  #设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == '/bos_id':
+            bos_id = tokenizer.bos_id
+            # print(bos_id)
+            # to json
+            if bos_id is None:
+                msg = json.dumps({'bos_id': -1})
+            else:
+                msg = json.dumps({'bos_id': bos_id})
+        elif self.path == '/eos_id':
+            eos_id = tokenizer.eos_id
+            if eos_id is None:
+                msg = json.dumps({'eos_id': -1})
+            else:
+                msg = json.dumps({'eos_id': eos_id})
+        else:
+            msg = 'error'
+
+        print(msg)
+        msg = str(msg).encode()  #转为str再转为byte格式
+
+        self.wfile.write(msg)  #将byte格式的信息返回给客户端
+
+    def do_POST(self):
+        #在新类中定义post的内容（当客户端向该服务端使用post请求时，本服务端将如下运行）
+        data = self.rfile.read(int(
+            self.headers['content-length']))  #获取从客户端传入的参数（byte格式）
+        data = data.decode()  #将byte格式转为str格式
+
+        self.send_response(200)
+        self.send_header("type", "post")  #设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == '/encode':
+            req = json.loads(data)
+            prompt = req['text']
+
+            token_ids = tokenizer.encode(prompt, args.content)
+            if token_ids is None:
+                msg = json.dumps({'token_ids': -1})
+            else:
+                msg = json.dumps({'token_ids': token_ids})
+
+        elif self.path == '/decode':
+            req = json.loads(data)
+            token_ids = req['token_ids']
+            text = tokenizer.decode(token_ids)
+            if text is None:
+                msg = json.dumps({'text': ""})
+            else:
+                msg = json.dumps({'text': text})
+        else:
+            msg = 'error'
+        print(msg)
+        msg = str(msg).encode()  #转为str再转为byte格式
+
+        self.wfile.write(msg)  #将byte格式的信息返回给客户端
+
+
+if __name__ == "__main__":
+
+    args = argparse.ArgumentParser()
+    args.add_argument('--host', type=str, default='localhost')
+    args.add_argument('--port', type=int, default=8080)
+    args.add_argument('--model_id', type=str, default='qwen2.5_coder_tokenizer')
+    args.add_argument('--content', type=str, default='You are Qwen, created by Alibaba Cloud. You are a helpful assistant.')
+    args = args.parse_args()
+
+    tokenizer = Tokenizer_Http(args.model_id)
+
+    # print(tokenizer.bos_id, tokenizer.bos_token, tokenizer.eos_id, tokenizer.eos_token)
+    # print(tokenizer.encode("hello world", args.content))
+
+    host = (args.host, args.port)  #设定地址与端口号，'localhost'等价于'127.0.0.1'
+    print('http://%s:%s' % host)
+    server = HTTPServer(host, Request)  #根据地址端口号和新定义的类，创建服务器实例
+    server.serve_forever()  #开启服务
diff --git a/projects/llm_framework/main_melotts/SConstruct b/projects/llm_framework/main_melotts/SConstruct
index 69d9d464..1701cb76 100644
--- a/projects/llm_framework/main_melotts/SConstruct
+++ b/projects/llm_framework/main_melotts/SConstruct
@@ -30,7 +30,7 @@ REQUIREMENTS += ['glog', 'fst']
 
 REQUIREMENTS += ['onnxruntime']
 
-STATIC_FILES += Glob('mode_*.json')
+STATIC_FILES += Glob('models/mode_*.json')
 
 env['COMPONENTS'].append({'target':'llm_melotts-1.8',
                           'SRCS':SRCS,
diff --git a/projects/llm_framework/main_melotts/mode_melotts-en-au.json b/projects/llm_framework/main_melotts/models/mode_melotts-en-au.json
similarity index 100%
rename from projects/llm_framework/main_melotts/mode_melotts-en-au.json
rename to projects/llm_framework/main_melotts/models/mode_melotts-en-au.json
diff --git a/projects/llm_framework/main_melotts/mode_melotts-en-br.json b/projects/llm_framework/main_melotts/models/mode_melotts-en-br.json
similarity index 100%
rename from projects/llm_framework/main_melotts/mode_melotts-en-br.json
rename to projects/llm_framework/main_melotts/models/mode_melotts-en-br.json
diff --git a/projects/llm_framework/main_melotts/mode_melotts-en-default.json b/projects/llm_framework/main_melotts/models/mode_melotts-en-default.json
similarity index 100%
rename from projects/llm_framework/main_melotts/mode_melotts-en-default.json
rename to projects/llm_framework/main_melotts/models/mode_melotts-en-default.json
diff --git a/projects/llm_framework/main_melotts/mode_melotts-en-india.json b/projects/llm_framework/main_melotts/models/mode_melotts-en-india.json
similarity index 100%
rename from projects/llm_framework/main_melotts/mode_melotts-en-india.json
rename to projects/llm_framework/main_melotts/models/mode_melotts-en-india.json
diff --git a/projects/llm_framework/main_melotts/mode_melotts-en-us.json b/projects/llm_framework/main_melotts/models/mode_melotts-en-us.json
similarity index 100%
rename from projects/llm_framework/main_melotts/mode_melotts-en-us.json
rename to projects/llm_framework/main_melotts/models/mode_melotts-en-us.json
diff --git a/projects/llm_framework/main_melotts/mode_melotts-es-es.json b/projects/llm_framework/main_melotts/models/mode_melotts-es-es.json
similarity index 100%
rename from projects/llm_framework/main_melotts/mode_melotts-es-es.json
rename to projects/llm_framework/main_melotts/models/mode_melotts-es-es.json
diff --git a/projects/llm_framework/main_melotts/mode_melotts-ja-jp.json b/projects/llm_framework/main_melotts/models/mode_melotts-ja-jp.json
similarity index 100%
rename from projects/llm_framework/main_melotts/mode_melotts-ja-jp.json
rename to projects/llm_framework/main_melotts/models/mode_melotts-ja-jp.json
diff --git a/projects/llm_framework/main_melotts/mode_melotts-zh-cn.json b/projects/llm_framework/main_melotts/models/mode_melotts-zh-cn.json
similarity index 100%
rename from projects/llm_framework/main_melotts/mode_melotts-zh-cn.json
rename to projects/llm_framework/main_melotts/models/mode_melotts-zh-cn.json
diff --git a/projects/llm_framework/main_vlm/models/mode_internvl2.5-1B-448-ax650.json b/projects/llm_framework/main_vlm/models/mode_internvl2.5-1B-448-ax650.json
new file mode 100644
index 00000000..ecc6c071
--- /dev/null
+++ b/projects/llm_framework/main_vlm/models/mode_internvl2.5-1B-448-ax650.json
@@ -0,0 +1,35 @@
+{
+    "mode":"internvl2.5-1B-448-ax650",
+    "type":"vlm",
+    "homepage":"https://huggingface.co/AXERA-TECH/InternVL2_5-1B",
+    "capabilities":[
+        "text_generation",
+        "chat"
+    ],
+    "input_type":[
+        "vlm.chat_completion",
+        "vlm.chat_completion.stream"
+    ],
+    "output_type":[
+        "vlm.utf-8",
+        "vlm.utf-8.stream"
+    ],
+    "mode_param":{
+        "tokenizer_type":2,
+        "filename_tokenizer_model":"http://localhost:8080",
+        "filename_tokens_embed":"model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel":"qwen2_post.axmodel",
+        "template_filename_axmodel":"qwen2_p320_l%d_together.axmodel",
+        "filename_vpm_resampler_axmodedl":"vit_intern_2_5_sim_space2depth_nhwc.axmodel",
+        "b_use_topk":false,
+        "b_bos":false,
+        "b_eos":false,
+        "axmodel_num":24,
+        "tokens_embed_num":151674,
+        "img_token_id":151667,
+        "tokens_embed_size":896,
+        "b_use_mmap_load_embed":true,
+        "b_dynamic_load_axmodel_layer":false,
+        "ext_scripts":["tokenizer_internvl2.5-1B-448-ax650.py"]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/models/mode_smolvlm-256M-ax650.json b/projects/llm_framework/main_vlm/models/mode_smolvlm-256M-ax650.json
new file mode 100644
index 00000000..179fa6b6
--- /dev/null
+++ b/projects/llm_framework/main_vlm/models/mode_smolvlm-256M-ax650.json
@@ -0,0 +1,35 @@
+{
+    "mode":"smolvlm-256M-ax650",
+    "type":"vlm",
+    "homepage":"https://huggingface.co/AXERA-TECH/SmolVLM-256M-Instruct",
+    "capabilities":[
+        "text_generation",
+        "chat"
+    ],
+    "input_type":[
+        "vlm.chat_completion",
+        "vlm.chat_completion.stream"
+    ],
+    "output_type":[
+        "vlm.utf-8",
+        "vlm.utf-8.stream"
+    ],
+    "mode_param":{
+        "tokenizer_type":2,
+        "filename_tokenizer_model":"http://localhost:8080",
+        "filename_tokens_embed":"model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel":"llama_post.axmodel",
+        "template_filename_axmodel":"llama_p128_l%d_together.axmodel",
+        "filename_vpm_resampler_axmodedl":"SmolVLM-256M-Instruct_vision_nhwc.axmodel",
+        "b_use_topk":false,
+        "b_bos":false,
+        "b_eos":false,
+        "axmodel_num":30,
+        "tokens_embed_num":49280,
+        "img_token_id":49190,
+        "tokens_embed_size":576,
+        "b_use_mmap_load_embed":true,
+        "b_dynamic_load_axmodel_layer":false,
+        "ext_scripts":["tokenizer_smolvlm-256M-ax650.py"]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/models/mode_smolvlm-500M-ax650.json b/projects/llm_framework/main_vlm/models/mode_smolvlm-500M-ax650.json
new file mode 100644
index 00000000..3ce09a08
--- /dev/null
+++ b/projects/llm_framework/main_vlm/models/mode_smolvlm-500M-ax650.json
@@ -0,0 +1,37 @@
+{
+    "mode":"smolvlm-500M-ax630c",
+    "type":"vlm",
+    "homepage":"https://huggingface.co/HuggingFaceTB/SmolVLM-500M-Instruct",
+    "compile_flage":"pulsar2 build --input HuggingFaceTB/SmolVLM-500M-w8a16/SmolVLM-500M-Instruct_vision.onnx --config AXERA/SmolVLM-256M-Instruct.axera/model_convert/config.json --output_dir HuggingFaceTB/SmolVLM-500M-w8a16/build-output --output_name SmolVLM-500M-Instruct_vision.axmodel --target_hardware AX620E --compiler.check 0 --npu_mode NPU2",
+    "pulsar_version":"3.4-983bb35e",    
+    "capabilities":[
+        "text_generation",
+        "chat"
+    ],
+    "input_type":[
+        "vlm.chat_completion",
+        "vlm.chat_completion.stream"
+    ],
+    "output_type":[
+        "vlm.utf-8",
+        "vlm.utf-8.stream"
+    ],
+    "mode_param":{
+        "tokenizer_type":2,
+        "filename_tokenizer_model":"http://localhost:8080",
+        "filename_tokens_embed":"model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel":"llama_post.axmodel",
+        "template_filename_axmodel":"llama_p128_l%d_together.axmodel",
+        "filename_vpm_resampler_axmodedl":"SmolVLM-500M-Instruct_vision.axmodel",
+        "b_use_topk":false,
+        "b_bos":false,
+        "b_eos":false,
+        "axmodel_num":32,
+        "tokens_embed_num":49280,
+        "img_token_id":49190,
+        "tokens_embed_size":960,
+        "b_use_mmap_load_embed":true,
+        "b_dynamic_load_axmodel_layer":false,
+        "ext_scripts":["tokenizer_smolvlm-500M-ax630c.py"]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/scripts/tokenizer_internvl2.5-1B-448-ax650.py b/projects/llm_framework/main_vlm/scripts/tokenizer_internvl2.5-1B-448-ax650.py
new file mode 100644
index 00000000..bb27674b
--- /dev/null
+++ b/projects/llm_framework/main_vlm/scripts/tokenizer_internvl2.5-1B-448-ax650.py
@@ -0,0 +1,138 @@
+from transformers import AutoTokenizer, PreTrainedTokenizerFast
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+
+
+class Tokenizer_Http:
+
+    def __init__(self, model_id):
+        self.tokenizer = AutoTokenizer.from_pretrained(
+            model_id, trust_remote_code=True, use_fast=False
+        )
+
+    def encode(self, prompt, content):
+        prompt = f"<|im_start|>system\n{content}<|im_end|><|im_start|>user\n{prompt}<|im_end|><|im_start|>assistant\n"
+        input_ids = self.tokenizer.encode(prompt)
+        return input_ids
+
+    def encode_vpm(self, prompt, content="Please describe the image shortly."):
+        prompt = f"<|im_start|>system\n{content}<|im_end|><|im_start|>user\n<img>" + "<IMG_CONTEXT>" * 256 + f"</img>\n{prompt}<|im_end|><|im_start|>assistant\n"
+        input_ids = self.tokenizer.encode(prompt)
+        return input_ids
+
+    def decode(self, token_ids):
+        return self.tokenizer.decode(token_ids, clean_up_tokenization_spaces=False)
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+
+    @property
+    def bos_token(self):
+        return self.tokenizer.bos_token
+
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+
+class Request(BaseHTTPRequestHandler):
+    # 通过类继承，新定义类
+    timeout = 5
+    server_version = "Apache"
+
+    def do_GET(self):
+        print(self.path)
+        # 在新类中定义get的内容（当客户端向该服务端使用get请求时，本服务端将如下运行）
+        self.send_response(200)
+        self.send_header("type", "get")  # 设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == "/bos_id":
+            bos_id = tokenizer.bos_id
+            # print(bos_id)
+            # to json
+            if bos_id is None:
+                msg = json.dumps({"bos_id": -1})
+            else:
+                msg = json.dumps({"bos_id": bos_id})
+        elif self.path == "/eos_id":
+            eos_id = tokenizer.eos_id
+            if eos_id is None:
+                msg = json.dumps({"eos_id": -1})
+            else:
+                msg = json.dumps({"eos_id": eos_id})
+        else:
+            msg = "error"
+
+        print(msg)
+        msg = str(msg).encode()  # 转为str再转为byte格式
+
+        self.wfile.write(msg)  # 将byte格式的信息返回给客户端
+
+    def do_POST(self):
+        # 在新类中定义post的内容（当客户端向该服务端使用post请求时，本服务端将如下运行）
+        data = self.rfile.read(
+            int(self.headers["content-length"])
+        )  # 获取从客户端传入的参数（byte格式）
+        data = data.decode()  # 将byte格式转为str格式
+
+        self.send_response(200)
+        self.send_header("type", "post")  # 设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == "/encode":
+            req = json.loads(data)
+            print(req)
+            prompt = req["text"]
+            b_img_prompt = False
+            if "img_prompt" in req:
+                b_img_prompt = req["img_prompt"]
+            if b_img_prompt:
+                token_ids = tokenizer.encode_vpm(prompt)
+            else:
+                token_ids = tokenizer.encode(prompt, args.content)
+            if token_ids is None:
+                msg = json.dumps({"token_ids": -1})
+            else:
+                msg = json.dumps({"token_ids": token_ids})
+
+        elif self.path == "/decode":
+            req = json.loads(data)
+            token_ids = req["token_ids"]
+            text = tokenizer.decode(token_ids)
+            if text is None:
+                msg = json.dumps({"text": ""})
+            else:
+                msg = json.dumps({"text": text})
+        else:
+            msg = "error"
+        print(msg)
+        msg = str(msg).encode()  # 转为str再转为byte格式
+
+        self.wfile.write(msg)  # 将byte格式的信息返回给客户端
+
+
+if __name__ == "__main__":
+
+    args = argparse.ArgumentParser()
+    args.add_argument("--host", type=str, default="localhost")
+    args.add_argument("--port", type=int, default=8080)
+    args.add_argument('--model_id', type=str, default='internvl2_tokenizer')
+    args.add_argument('--content', type=str, default='你是由上海人工智能实验室联合商汤科技开发的书生多模态大模型，英文名叫InternVL, 是一个有用无害的人工智能助手。')
+    args = args.parse_args()
+
+    tokenizer = Tokenizer_Http(args.model_id)
+
+
+    # print(tokenizer.bos_id, tokenizer.bos_token, tokenizer.eos_id, tokenizer.eos_token)
+    # print(tokenizer.encode("hello world", args.content))
+
+    host = (args.host, args.port)  # 设定地址与端口号，'localhost'等价于'127.0.0.1'
+    print("http://%s:%s" % host)
+    server = HTTPServer(host, Request)  # 根据地址端口号和新定义的类，创建服务器实例
+    server.serve_forever()  # 开启服务
diff --git a/projects/llm_framework/main_vlm/scripts/tokenizer_smolvlm-256M-ax650.py b/projects/llm_framework/main_vlm/scripts/tokenizer_smolvlm-256M-ax650.py
new file mode 100644
index 00000000..560a71f3
--- /dev/null
+++ b/projects/llm_framework/main_vlm/scripts/tokenizer_smolvlm-256M-ax650.py
@@ -0,0 +1,248 @@
+from transformers import AutoTokenizer, PreTrainedTokenizerFast
+from transformers.tokenization_utils_base import AddedToken
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+
+def _prompt_split_image(
+    image_seq_len,
+    image_rows,
+    image_cols,
+    fake_token_around_image,
+    image_token,
+    global_img_token,
+):
+    """Prompt with expanded image tokens for when the image is split into patches."""
+    text_split_images = ""
+    for n_h in range(image_rows):
+        for n_w in range(image_cols):
+            text_split_images += (
+                f"{fake_token_around_image}"
+                + f"<row_{n_h + 1}_col_{n_w + 1}>"
+                + f"{image_token}" * image_seq_len
+            )
+        text_split_images += "\n"
+
+    text_split_images += (
+        f"\n{fake_token_around_image}"
+        + f"{global_img_token}"
+        + f"{image_token}" * image_seq_len
+        + f"{fake_token_around_image}"
+    )
+    return text_split_images
+
+
+def _prompt_single_image(
+    image_seq_len, fake_token_around_image, image_token, global_img_token
+):
+    """Prompt with expanded image tokens for a single image."""
+    return (
+        f"{fake_token_around_image}"
+        + f"{global_img_token}"
+        + f"{image_token}" * image_seq_len
+        + f"{fake_token_around_image}"
+    )
+
+
+def get_image_prompt_string(
+    image_rows,
+    image_cols,
+    image_seq_len,
+    fake_token_around_image,
+    image_token,
+    global_img_token,
+):
+    if image_rows == 0 and image_cols == 0:
+        return _prompt_single_image(
+            image_seq_len,
+            fake_token_around_image=fake_token_around_image,
+            image_token=image_token,
+            global_img_token=global_img_token,
+        )
+    return _prompt_split_image(
+        image_seq_len,
+        image_rows,
+        image_cols,
+        fake_token_around_image,
+        image_token,
+        global_img_token,
+    )
+
+class Tokenizer_Http:
+
+    def __init__(self, model_id):
+        self.tokenizer = AutoTokenizer.from_pretrained(
+            model_id, trust_remote_code=True, use_fast=False
+        )
+
+    def encode(self, prompt, content):
+        prompt = f"<|im_start|>User:{content}<end_of_utterance>\nAssistant:"
+        input_ids = self.tokenizer(prompt)
+        return input_ids["input_ids"]
+
+    def encode_vpm(self, prompt, content="Please describe the image shortly."):
+        prompt = f"<|im_start|>User:<image>{prompt}<end_of_utterance>\nAssistant:"
+        text = [prompt]
+        image_rows = [[0]]
+        image_cols = [[0]]
+        image_seq_len = 64
+        image_token = "<image>"
+        fake_image_token = "<fake_token_around_image>"
+        global_img_token = "<global-img>"
+        prompt_strings = []
+        for sample, sample_rows, sample_cols in zip(text, image_rows, image_cols):
+            # Replace the image token with fake tokens around the expanded image token sequence of length `image_seq_len`
+            image_prompt_strings = []
+            for n_rows, n_cols in zip(sample_rows, sample_cols):
+                image_prompt_string = get_image_prompt_string(
+                    n_rows,
+                    n_cols,
+                    image_seq_len,
+                    image_token=image_token,
+                    fake_token_around_image=fake_image_token,
+                    global_img_token=global_img_token,
+                )
+                image_prompt_strings.append(image_prompt_string)
+
+            split_sample = sample.split(image_token)
+            if len(split_sample) == 0:
+                raise ValueError("The image token should be present in the text.")
+
+            # Place in the image prompt strings where the image tokens are
+            sample = split_sample[0]
+            for i, image_prompt_string in enumerate(image_prompt_strings):
+                sample += image_prompt_string + split_sample[i + 1]
+            prompt_strings.append(sample)
+
+        fake_image_token = AddedToken(fake_image_token, normalized=False, special=True)
+        image_token = AddedToken(image_token, normalized=False, special=True)
+        end_of_utterance_token = AddedToken(
+            "<end_of_utterance>", normalized=False, special=True
+        )
+        tokens_to_add = {
+            "additional_special_tokens": [
+                fake_image_token,
+                image_token,
+                end_of_utterance_token,
+            ]
+        }
+        self.tokenizer.add_special_tokens(tokens_to_add)
+
+        input_ids = self.tokenizer(prompt_strings)["input_ids"][0]
+        return input_ids
+
+    def decode(self, token_ids):
+        return self.tokenizer.decode(token_ids, clean_up_tokenization_spaces=False)
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+
+    @property
+    def bos_token(self):
+        return self.tokenizer.bos_token
+
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+
+class Request(BaseHTTPRequestHandler):
+    # 通过类继承，新定义类
+    timeout = 5
+    server_version = "Apache"
+
+    def do_GET(self):
+        print(self.path)
+        # 在新类中定义get的内容（当客户端向该服务端使用get请求时，本服务端将如下运行）
+        self.send_response(200)
+        self.send_header("type", "get")  # 设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == "/bos_id":
+            bos_id = tokenizer.bos_id
+            # print(bos_id)
+            # to json
+            if bos_id is None:
+                msg = json.dumps({"bos_id": -1})
+            else:
+                msg = json.dumps({"bos_id": bos_id})
+        elif self.path == "/eos_id":
+            eos_id = tokenizer.eos_id
+            if eos_id is None:
+                msg = json.dumps({"eos_id": -1})
+            else:
+                msg = json.dumps({"eos_id": eos_id})
+        else:
+            msg = "error"
+
+        print(msg)
+        msg = str(msg).encode()  # 转为str再转为byte格式
+
+        self.wfile.write(msg)  # 将byte格式的信息返回给客户端
+
+    def do_POST(self):
+        # 在新类中定义post的内容（当客户端向该服务端使用post请求时，本服务端将如下运行）
+        data = self.rfile.read(
+            int(self.headers["content-length"])
+        )  # 获取从客户端传入的参数（byte格式）
+        data = data.decode()  # 将byte格式转为str格式
+
+        self.send_response(200)
+        self.send_header("type", "post")  # 设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == "/encode":
+            req = json.loads(data)
+            print(req)
+            prompt = req["text"]
+            b_img_prompt = False
+            if "img_prompt" in req:
+                b_img_prompt = req["img_prompt"]
+            if b_img_prompt:
+                token_ids = tokenizer.encode_vpm(prompt)
+            else:
+                token_ids = tokenizer.encode(prompt, args.content)
+            if token_ids is None:
+                msg = json.dumps({"token_ids": -1})
+            else:
+                msg = json.dumps({"token_ids": token_ids})
+
+        elif self.path == "/decode":
+            req = json.loads(data)
+            token_ids = req["token_ids"]
+            text = tokenizer.decode(token_ids)
+            if text is None:
+                msg = json.dumps({"text": ""})
+            else:
+                msg = json.dumps({"text": text})
+        else:
+            msg = "error"
+        print(msg)
+        msg = str(msg).encode()  # 转为str再转为byte格式
+
+        self.wfile.write(msg)  # 将byte格式的信息返回给客户端
+
+
+if __name__ == "__main__":
+
+    args = argparse.ArgumentParser()
+    args.add_argument("--host", type=str, default="localhost")
+    args.add_argument("--port", type=int, default=8080)
+    args.add_argument('--model_id', type=str, default='internvl2_tokenizer')
+    args.add_argument('--content', type=str, default='')
+    args = args.parse_args()
+
+    tokenizer = Tokenizer_Http(args.model_id)
+
+
+    # print(tokenizer.bos_id, tokenizer.bos_token, tokenizer.eos_id, tokenizer.eos_token)
+    # print(tokenizer.encode("hello world", args.content))
+
+    host = (args.host, args.port)  # 设定地址与端口号，'localhost'等价于'127.0.0.1'
+    print("http://%s:%s" % host)
+    server = HTTPServer(host, Request)  # 根据地址端口号和新定义的类，创建服务器实例
+    server.serve_forever()  # 开启服务
diff --git a/projects/llm_framework/main_whisper/SConstruct b/projects/llm_framework/main_whisper/SConstruct
index 4c61edce..6395e923 100644
--- a/projects/llm_framework/main_whisper/SConstruct
+++ b/projects/llm_framework/main_whisper/SConstruct
@@ -31,7 +31,7 @@ LINK_SEARCH_PATH += [ADir('../static_lib/opencc/lib')]
 # LDFLAGS += ['-l:libcargs.a', '-l:libonnxruntime.a']
 LDFLAGS += ['-l:libopencc.a', '-l:libmarisa.a']
 
-STATIC_FILES += Glob('mode_*.json')
+STATIC_FILES += Glob('models/mode_*.json')
 
 env['COMPONENTS'].append({'target':'llm_whisper-1.7',
                           'SRCS':SRCS,
diff --git a/projects/llm_framework/main_whisper/models/mode_whisper-base-ax650.json b/projects/llm_framework/main_whisper/models/mode_whisper-base-ax650.json
new file mode 100644
index 00000000..a9c22be6
--- /dev/null
+++ b/projects/llm_framework/main_whisper/models/mode_whisper-base-ax650.json
@@ -0,0 +1,42 @@
+{
+    "mode": "whisper-base-ax650",
+    "type": "asr",
+    "homepage":"https://huggingface.co/openai/whisper-base",
+    "capabilities": [
+        "Automatic_Speech_Recognition",
+        "English",
+        "Chinese",
+        "Japanese"
+    ],
+    "input_type": [
+        "sys.pcm"
+    ],
+    "output_type": [
+        "asr.utf-8"
+    ],
+    "mode_param": {
+        "model_type": "base",
+        "language": "en",
+        "encoder": "base-encoder.axmodel",
+        "decoder_main": "base-decoder-main.axmodel",
+        "decoder_loop": "base-decoder-loop.axmodel",
+        "positional_embedding": "base-positional_embedding.bin",
+        "tokens": "base-tokens.txt",
+        "t2s": "t2s.json",
+        "whisper_sample_rate": 16000,
+        "whisper_n_fft": 400,
+        "awake_delay": 1000,
+        "whisper_hop_length": 160,
+        "whisper_chunk_size": 30,
+        "whisper_n_mels": 80,
+        "whisper_sot": 50258,
+        "whisper_eot": 50257,
+        "whisper_blank": 220,
+        "whisper_no_timestamps": 50363,
+        "whisper_no_speech": 50362,
+        "whisper_translate": 50358,
+        "whisper_transcribe": 50359,
+        "whisper_vocab_size": 51865,
+        "whisper_n_text_ctx": 448
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_whisper/mode_whisper-base.json b/projects/llm_framework/main_whisper/models/mode_whisper-base.json
similarity index 100%
rename from projects/llm_framework/main_whisper/mode_whisper-base.json
rename to projects/llm_framework/main_whisper/models/mode_whisper-base.json
diff --git a/projects/llm_framework/main_whisper/models/mode_whisper-small-ax650.json b/projects/llm_framework/main_whisper/models/mode_whisper-small-ax650.json
new file mode 100644
index 00000000..8ef8f1c7
--- /dev/null
+++ b/projects/llm_framework/main_whisper/models/mode_whisper-small-ax650.json
@@ -0,0 +1,42 @@
+{
+    "mode": "whisper-small-ax650",
+    "type": "asr",
+    "homepage":"https://huggingface.co/openai/whisper-small",
+    "capabilities": [
+        "Automatic_Speech_Recognition",
+        "English",
+        "Chinese",
+        "Japanese"
+    ],
+    "input_type": [
+        "sys.pcm"
+    ],
+    "output_type": [
+        "asr.utf-8"
+    ],
+    "mode_param": {
+        "model_type": "small",
+        "language": "en",
+        "encoder": "small-encoder.axmodel",
+        "decoder_main": "small-decoder-main.axmodel",
+        "decoder_loop": "small-decoder-loop.axmodel",
+        "positional_embedding": "small-positional_embedding.bin",
+        "tokens": "small-tokens.txt",
+        "t2s": "t2s.json",
+        "whisper_sample_rate": 16000,
+        "whisper_n_fft": 400,
+        "awake_delay": 1000,
+        "whisper_hop_length": 160,
+        "whisper_chunk_size": 30,
+        "whisper_n_mels": 80,
+        "whisper_sot": 50258,
+        "whisper_eot": 50257,
+        "whisper_blank": 220,
+        "whisper_no_timestamps": 50363,
+        "whisper_no_speech": 50362,
+        "whisper_translate": 50358,
+        "whisper_transcribe": 50359,
+        "whisper_vocab_size": 51865,
+        "whisper_n_text_ctx": 448
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_whisper/mode_whisper-small.json b/projects/llm_framework/main_whisper/models/mode_whisper-small.json
similarity index 100%
rename from projects/llm_framework/main_whisper/mode_whisper-small.json
rename to projects/llm_framework/main_whisper/models/mode_whisper-small.json
diff --git a/projects/llm_framework/main_whisper/models/mode_whisper-tiny-ax650.json b/projects/llm_framework/main_whisper/models/mode_whisper-tiny-ax650.json
new file mode 100644
index 00000000..b44e7b47
--- /dev/null
+++ b/projects/llm_framework/main_whisper/models/mode_whisper-tiny-ax650.json
@@ -0,0 +1,42 @@
+{
+    "mode": "whisper-tiny-ax650",
+    "type": "asr",
+    "homepage":"https://huggingface.co/openai/whisper-tiny",
+    "capabilities": [
+        "Automatic_Speech_Recognition",
+        "English",
+        "Chinese",
+        "Japanese"
+    ],
+    "input_type": [
+        "sys.pcm"
+    ],
+    "output_type": [
+        "asr.utf-8"
+    ],
+    "mode_param": {
+        "model_type": "tiny",
+        "language": "en",
+        "encoder": "tiny-encoder.axmodel",
+        "decoder_main": "tiny-decoder-main.axmodel",
+        "decoder_loop": "tiny-decoder-loop.axmodel",
+        "positional_embedding": "tiny-positional_embedding.bin",
+        "tokens": "tiny-tokens.txt",
+        "t2s": "t2s.json",
+        "whisper_sample_rate": 16000,
+        "whisper_n_fft": 400,
+        "awake_delay": 1000,
+        "whisper_hop_length": 160,
+        "whisper_chunk_size": 30,
+        "whisper_n_mels": 80,
+        "whisper_sot": 50258,
+        "whisper_eot": 50257,
+        "whisper_blank": 220,
+        "whisper_no_timestamps": 50363,
+        "whisper_no_speech": 50362,
+        "whisper_translate": 50358,
+        "whisper_transcribe": 50359,
+        "whisper_vocab_size": 51865,
+        "whisper_n_text_ctx": 448
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_whisper/mode_whisper-tiny.json b/projects/llm_framework/main_whisper/models/mode_whisper-tiny.json
similarity index 100%
rename from projects/llm_framework/main_whisper/mode_whisper-tiny.json
rename to projects/llm_framework/main_whisper/models/mode_whisper-tiny.json

From 90fae78817446612fb174087191d095333ad0dea Mon Sep 17 00:00:00 2001
From: dianjixz <18637716021@163.com>
Date: Tue, 1 Jul 2025 12:17:26 +0800
Subject: [PATCH 23/79] [update] main_audio add 630c kit default param &&
 StackFlow add send_raw_to_pub

---
 .../StackFlow/stackflow/StackFlow.cpp         |   5 +
 .../StackFlow/stackflow/StackFlow.h           |   1 +
 projects/llm_framework/main_audio/SConstruct  |   2 +-
 .../llm_framework/main_audio/audio_kit.json   | 115 ++++++++++++++++++
 .../llm_framework/main_audio/src/main.cpp     |  22 ++--
 5 files changed, 130 insertions(+), 15 deletions(-)
 create mode 100644 projects/llm_framework/main_audio/audio_kit.json

diff --git a/ext_components/StackFlow/stackflow/StackFlow.cpp b/ext_components/StackFlow/stackflow/StackFlow.cpp
index 2991162e..561e20f8 100644
--- a/ext_components/StackFlow/stackflow/StackFlow.cpp
+++ b/ext_components/StackFlow/stackflow/StackFlow.cpp
@@ -117,6 +117,11 @@ int llm_channel_obj::send_raw_to_pub(const std::string &raw)
     return zmq_[-1]->send_data(raw);
 }
 
+int llm_channel_obj::send_raw_to_pub(const char *data, int size)
+{
+    return zmq_[-1]->send_data(data, size);
+}
+
 int llm_channel_obj::send_raw_to_usr(const std::string &raw)
 {
     if (zmq_[-2]) {
diff --git a/ext_components/StackFlow/stackflow/StackFlow.h b/ext_components/StackFlow/stackflow/StackFlow.h
index 25ac6743..c3424fa2 100644
--- a/ext_components/StackFlow/stackflow/StackFlow.h
+++ b/ext_components/StackFlow/stackflow/StackFlow.h
@@ -106,6 +106,7 @@ class llm_channel_obj {
     void stop_subscriber(const std::string &zmq_url);
     int check_zmq_errno(void *ctx, void *com, int code);
     int send_raw_to_pub(const std::string &raw);
+    int send_raw_to_pub(const char *data, int size);
     int send_raw_to_usr(const std::string &raw);
     template <typename T, typename U>
     int output_data(const std::string &object, const T &data, const U &error_msg)
diff --git a/projects/llm_framework/main_audio/SConstruct b/projects/llm_framework/main_audio/SConstruct
index 27a04b68..18d7ca05 100644
--- a/projects/llm_framework/main_audio/SConstruct
+++ b/projects/llm_framework/main_audio/SConstruct
@@ -26,7 +26,7 @@ LINK_SEARCH_PATH += [ADir('../static_lib')]
 REQUIREMENTS += ['ax_sys', 'ax_interpreter', 'ax_audio', 'ax_audio_3a', 'ax_fdk', 'ax_opus']
 REQUIREMENTS += ['tinyalsa', 'opus', 'samplerate', 'fdk-aac']
 
-STATIC_FILES += [AFile('audio.json')]
+STATIC_FILES += [AFile('audio.json'), AFile('audio_kit.json')]
 STATIC_FILES += Glob('mode_*.json')
 
 env['COMPONENTS'].append({'target':'llm_audio-1.6',
diff --git a/projects/llm_framework/main_audio/audio_kit.json b/projects/llm_framework/main_audio/audio_kit.json
new file mode 100644
index 00000000..1c7a85fd
--- /dev/null
+++ b/projects/llm_framework/main_audio/audio_kit.json
@@ -0,0 +1,115 @@
+{
+    "mode": "None",
+    "type": "audio",
+    "capabilities": [
+        "play",
+        "cap"
+    ],
+    "input_type": [
+        "rpc.audio.wav.base64",
+        "rpc.audio.pcm.base64"
+    ],
+    "output_type": [
+        "audio.pcm.stream"
+    ],
+    "play_param": {
+        "card": 0,
+        "device": 1,
+        "volume": 0.5,
+        "channel": 1,
+        "rate": 16000,
+        "bit": 16,
+        "stPoolConfig.MetaSize": 8192,
+        "stPoolConfig.BlkSize": 32768,
+        "stPoolConfig.BlkCnt": 37,
+        "stPoolConfig.IsMergeMode": 0,
+        "stPoolConfig.CacheMode": 0,
+        "stPoolConfig.PartitionName": "anonymous",
+        "stAttr.enBitwidth": 1,
+        "stAttr.enSoundmode": 0,
+        "stAttr.u32ChnCnt": 2,
+        "stAttr.enLinkMode": 0,
+        "stAttr.enSamplerate": 16000,
+        "stAttr.U32Depth": 30,
+        "stAttr.u32PeriodSize": 160,
+        "stAttr.u32PeriodCount": 8,
+        "stAttr.bInsertSilence": 0,
+        "stVqeAttr.s32SampleRate": 16000,
+        "stVqeAttr.u32FrameSamples": 160,
+        "stVqeAttr.stNsCfg.bNsEnable": 0,
+        "stVqeAttr.stNsCfg.enAggressivenessLevel": 2,
+        "stVqeAttr.stAgcCfg.bAgcEnable": 0,
+        "stVqeAttr.stAgcCfg.enAgcMode": 2,
+        "stVqeAttr.stAgcCfg.s16TargetLevel": -3,
+        "stVqeAttr.stAgcCfg.s16Gain": 9,
+        "stHpfAttr.bEnable": 0,
+        "stHpfAttr.s32GainDb": -3,
+        "stHpfAttr.s32Samplerate": 16000,
+        "stHpfAttr.s32Freq": 200,
+        "stLpfAttr.bEnable": 0,
+        "stLpfAttr.s32GainDb": 0,
+        "stLpfAttr.s32Samplerate": 16000,
+        "stLpfAttr.s32Freq": 3000,
+        "stEqAttr.bEnable": 0,
+        "stEqAttr.s32GainDb[0]": -10,
+        "stEqAttr.s32GainDb[1]": -3,
+        "stEqAttr.s32GainDb[2]": 3,
+        "stEqAttr.s32GainDb[3]": 5,
+        "stEqAttr.s32GainDb[4]": 10,
+        "stEqAttr.s32Samplerate": 16000,
+        "gResample": 0,
+        "enInSampleRate": 16000,
+        "gInstant": 0,
+        "gInsertSilence": 0
+    },
+    "cap_param": {
+        "sys_pcm_cap_channel": "ipc:///tmp/llm/pcm.cap.socket",
+        "card": 0,
+        "device": 0,
+        "volume": 1.0,
+        "channel": 1,
+        "rate": 16000,
+        "bit": 16,
+        "stPoolConfig.MetaSize": 8192,
+        "stPoolConfig.BlkSize": 7680,
+        "stPoolConfig.BlkCnt": 33,
+        "stPoolConfig.IsMergeMode": 0,
+        "stPoolConfig.CacheMode": 0,
+        "stPoolConfig.PartitionName": "anonymous",
+        "aistAttr.enBitwidth": 1,
+        "aistAttr.enLinkMode": 0,
+        "aistAttr.enSamplerate": 16000,
+        "aistAttr.enLayoutMode": 1,
+        "aistAttr.U32Depth": 30,
+        "aistAttr.u32PeriodSize": 160,
+        "aistAttr.u32PeriodCount": 8,
+        "aistAttr.u32ChnCnt": 2,
+        "aistVqeAttr.s32SampleRate": 16000,
+        "aistVqeAttr.u32FrameSamples": 160,
+        "aistVqeAttr.stNsCfg.bNsEnable": 1,
+        "aistVqeAttr.stNsCfg.enAggressivenessLevel": 2,
+        "aistVqeAttr.stAgcCfg.bAgcEnable": 0,
+        "aistVqeAttr.stAgcCfg.enAgcMode": 2,
+        "aistVqeAttr.stAgcCfg.s16TargetLevel": -3,
+        "aistVqeAttr.stAgcCfg.s16Gain": 9,
+        "aistVqeAttr.stAecCfg.enAecMode": 2,
+        "stHpfAttr.bEnable": 0,
+        "stHpfAttr.s32GainDb": -3,
+        "stHpfAttr.s32Samplerate": 16000,
+        "stHpfAttr.s32Freq": 200,
+        "stLpfAttr.bEnable": 0,
+        "stLpfAttr.s32GainDb": 0,
+        "stLpfAttr.s32Samplerate": 16000,
+        "stLpfAttr.s32Freq": 3000,
+        "stEqAttr.bEnable": 0,
+        "stEqAttr.s32GainDb[0]": -10,
+        "stEqAttr.s32GainDb[1]": -3,
+        "stEqAttr.s32GainDb[2]": 3,
+        "stEqAttr.s32GainDb[3]": 5,
+        "stEqAttr.s32GainDb[4]": 10,
+        "stEqAttr.s32Samplerate": 16000,
+        "gResample": 0,
+        "enOutSampleRate": 16000,
+        "gDbDetection": 0
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_audio/src/main.cpp b/projects/llm_framework/main_audio/src/main.cpp
index 80da841c..e398e8ad 100644
--- a/projects/llm_framework/main_audio/src/main.cpp
+++ b/projects/llm_framework/main_audio/src/main.cpp
@@ -142,8 +142,13 @@ class llm_audio : public StackFlow {
         nlohmann::json error_body;
         std::string base_model_path;
         std::string base_model_config_path;
-        std::list<std::string> config_file_paths =
-            get_config_file_paths(base_model_path, base_model_config_path, "audio");
+        std::list<std::string> config_file_paths;
+        if (access("/sys/devices/platform/soc/4851000.i2c/i2c-1/1-0043", F_OK) == 0) {
+            config_file_paths = get_config_file_paths(base_model_path, base_model_config_path, "audio_kit");
+        } else {
+            config_file_paths = get_config_file_paths(base_model_path, base_model_config_path, "audio");
+        }
+
         try {
             config_body = nlohmann::json::parse(data);
             for (auto file_name : config_file_paths) {
@@ -241,18 +246,7 @@ class llm_audio : public StackFlow {
                 CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.enBitwidth);
                 CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.enLinkMode);
                 CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.enSamplerate);
-
-                if (config_body.contains("aistAttr.enLayoutMode"))
-                    mode_config_.aistAttr.enLayoutMode = config_body["aistAttr.enLayoutMode"];
-                else if (file_body["cap_param"].contains("aistAttr.enLayoutMode")) {
-                    mode_config_.aistAttr.enLayoutMode = file_body["cap_param"]["aistAttr.enLayoutMode"];
-                    if (access("/sys/devices/platform/soc/4851000.i2c/i2c-1/1-0043", F_OK) == 0) {
-                        if (mode_config_.aistAttr.enLayoutMode == AX_AI_REF_MIC) {
-                            mode_config_.aistAttr.enLayoutMode = AX_AI_MIC_REF;
-                        }
-                    }
-                }
-
+                CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.enLayoutMode);
                 CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.U32Depth);
                 CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.u32PeriodSize);
                 CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.u32PeriodCount);

From 59958869a0adec852067acce4a6a61b941f680b4 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Tue, 1 Jul 2025 15:40:30 +0800
Subject: [PATCH 24/79] [update] main_audio add tinyalsa API cap function.

---
 ext_components/ax_msp/SConstruct              |   1 +
 .../llm_framework/main_audio/src/alsa_audio.c | 136 ++++++++++++++++++
 .../llm_framework/main_audio/src/alsa_audio.h |  15 ++
 .../llm_framework/main_audio/src/main.cpp     |   1 +
 4 files changed, 153 insertions(+)
 create mode 100644 projects/llm_framework/main_audio/src/alsa_audio.c
 create mode 100644 projects/llm_framework/main_audio/src/alsa_audio.h

diff --git a/ext_components/ax_msp/SConstruct b/ext_components/ax_msp/SConstruct
index a77c7d46..3c5ecbe3 100644
--- a/ext_components/ax_msp/SConstruct
+++ b/ext_components/ax_msp/SConstruct
@@ -29,6 +29,7 @@ if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
     for dirn in third_party:
         INCLUDE.append(os.path.join(MSP_PATH,'third-party',dirn,'include'))
         LINK_SEARCH_PATH.append(os.path.join(MSP_PATH,'third-party',dirn,'lib/arm64/glibc'))
+    INCLUDE.append(os.path.join(MSP_PATH,'third-party/tinyalsa/include/tinyalsa'))
     INCLUDE.append(os.path.join(MSP_PATH,'third-party/live/out/arm64/glibc/include'))
     LINK_SEARCH_PATH.append(os.path.join(MSP_PATH,'third-party/live/out/arm64/glibc/lib'))
     INCLUDE.append(os.path.join(MSP_PATH,'third-party/openssl/arm64/include'))
diff --git a/projects/llm_framework/main_audio/src/alsa_audio.c b/projects/llm_framework/main_audio/src/alsa_audio.c
new file mode 100644
index 00000000..73aca0f4
--- /dev/null
+++ b/projects/llm_framework/main_audio/src/alsa_audio.c
@@ -0,0 +1,136 @@
+#include "alsa_audio.h"
+#include "samplerate.h"
+#include <pcm.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <errno.h>
+#include <string.h>
+
+static int gcapLoopExit = 0;
+
+void alsa_cap_start(unsigned int card, unsigned int device, float Volume, int channel, int rate, int bit,
+                    AUDIOCallback callback)
+{
+    struct pcm_config config;
+    unsigned int pcm_open_flags;
+    struct pcm *pcm;
+    char *buffer;
+    unsigned int size;
+    unsigned int frames_read;
+    unsigned int total_frames_read;
+    unsigned int bytes_per_frame;
+
+    memset(&config, 0, sizeof(config));
+    config.channels          = channel;
+    config.rate              = 48000;  // TODO: 部分USB MIC仅支持48k，暂时固定采集为48k
+    config.period_size       = 512;
+    config.period_count      = 4;
+    config.format            = PCM_FORMAT_S16_LE;
+    config.start_threshold   = 0;
+    config.stop_threshold    = 0;
+    config.silence_threshold = 0;
+
+    pcm_open_flags = PCM_IN;
+    if (1) pcm_open_flags |= PCM_MMAP;
+
+    pcm = pcm_open(card, device, pcm_open_flags, &config);
+    if (!pcm || !pcm_is_ready(pcm)) {
+        fprintf(stderr, "Unable to open PCM device (%s)\n", pcm_get_error(pcm));
+        return;
+    }
+
+    size   = pcm_frames_to_bytes(pcm, pcm_get_buffer_size(pcm));
+    buffer = malloc(size);
+    if (!buffer) {
+        fprintf(stderr, "Unable to allocate %u bytes\n", size);
+        pcm_close(pcm);
+        return;
+    }
+
+    if (1) {
+        printf("Capturing sample: %u ch, %u hz, %u bit\n", channel, rate, pcm_format_to_bits(PCM_FORMAT_S16_LE));
+    }
+
+    bytes_per_frame   = pcm_frames_to_bytes(pcm, 1);
+    total_frames_read = 0;
+
+    SRC_STATE *src_state = NULL;
+    float *in_float = NULL, *out_float = NULL;
+    int in_frames  = pcm_get_buffer_size(pcm);
+    int out_frames = (int)((float)in_frames * ((float)rate / 48000.0f) + 1);
+    int out_bytes  = out_frames * channel * sizeof(short);
+
+    if (rate != 48000) {
+        src_state = src_new(SRC_SINC_FASTEST, channel, NULL);
+        in_float  = malloc(in_frames * channel * sizeof(float));
+        out_float = malloc(out_frames * channel * sizeof(float));
+        if (!src_state || !in_float || !out_float) {
+            fprintf(stderr, "Unable to allocate resample buffers\n");
+            free(buffer);
+            if (in_float) free(in_float);
+            if (out_float) free(out_float);
+            if (src_state) src_delete(src_state);
+            pcm_close(pcm);
+            return;
+        }
+    }
+
+    while (!gcapLoopExit) {
+        int ret = pcm_readi(pcm, buffer, in_frames);
+        if (ret < 0) {
+            fprintf(stderr, "Error capturing samples - %d (%s)\n", errno, strerror(errno));
+            break;
+        }
+        frames_read = ret;
+        total_frames_read += frames_read;
+
+        if (rate == 48000) {
+            callback(buffer, frames_read * bytes_per_frame);
+        } else {
+            short *in_short = (short *)buffer;
+            for (int i = 0; i < frames_read * channel; ++i) {
+                in_float[i] = in_short[i] / 32768.0f;
+            }
+            SRC_DATA src_data;
+            src_data.data_in       = in_float;
+            src_data.input_frames  = frames_read;
+            src_data.data_out      = out_float;
+            src_data.output_frames = out_frames;
+            src_data.src_ratio     = (double)rate / 48000.0;
+            src_data.end_of_input  = 0;
+            int error              = src_process(src_state, &src_data);
+            if (error) {
+                fprintf(stderr, "SRC error: %s\n", src_strerror(error));
+                break;
+            }
+            // float转short
+            short *out_short = malloc(src_data.output_frames_gen * channel * sizeof(short));
+            for (int i = 0; i < src_data.output_frames_gen * channel; ++i) {
+                float sample = out_float[i];
+                if (sample > 1.0f) sample = 1.0f;
+                if (sample < -1.0f) sample = -1.0f;
+                out_short[i] = (short)(sample * 32767.0f);
+            }
+            callback((const char *)out_short, src_data.output_frames_gen * channel * sizeof(short));
+            free(out_short);
+        }
+    }
+
+    if (rate != 48000) {
+        free(in_float);
+        free(out_float);
+        src_delete(src_state);
+    }
+    free(buffer);
+    pcm_close(pcm);
+}
+
+void alsa_close_cap()
+{
+    gcapLoopExit = 1;
+}
+
+int alsa_cap_status()
+{
+    return gcapLoopExit;
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_audio/src/alsa_audio.h b/projects/llm_framework/main_audio/src/alsa_audio.h
new file mode 100644
index 00000000..95c469ab
--- /dev/null
+++ b/projects/llm_framework/main_audio/src/alsa_audio.h
@@ -0,0 +1,15 @@
+#pragma once
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef void (*AUDIOCallback)(const char *data, int size);
+
+void alsa_cap_start(unsigned int card, unsigned int device, float Volume, int channel, int rate, int bit,
+                    AUDIOCallback callback);
+void alsa_close_cap();
+
+#ifdef __cplusplus
+}
+#endif
\ No newline at end of file
diff --git a/projects/llm_framework/main_audio/src/main.cpp b/projects/llm_framework/main_audio/src/main.cpp
index e398e8ad..e7007538 100644
--- a/projects/llm_framework/main_audio/src/main.cpp
+++ b/projects/llm_framework/main_audio/src/main.cpp
@@ -22,6 +22,7 @@ static void __sigint(int iSigNo)
 }
 
 #include "sample_audio.h"
+#include "alsa_audio.h"
 
 #define CONFIG_AUTO_SET(obj, key)             \
     if (config_body.contains(#key))           \

From 9abe069db6c7d1ff29e46ccba6f5b2e5e19bd78a Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 18 Jul 2025 14:12:12 +0800
Subject: [PATCH 25/79] [update] KWS sets multiple keywords, fix melotts File
 missing, Optimize generated audio

---
 projects/llm_framework/main_kws/src/main.cpp  |  76 +++++--
 .../llm_framework/main_melotts/src/main.cpp   | 212 +++++++-----------
 .../main_melotts/src/runner/Lexicon.hpp       |   8 +-
 3 files changed, 141 insertions(+), 155 deletions(-)

diff --git a/projects/llm_framework/main_kws/src/main.cpp b/projects/llm_framework/main_kws/src/main.cpp
index 4ea6abcd..5b8a21bb 100644
--- a/projects/llm_framework/main_kws/src/main.cpp
+++ b/projects/llm_framework/main_kws/src/main.cpp
@@ -34,6 +34,8 @@ static void __sigint(int iSigNo)
 static std::string base_model_path_;
 static std::string base_model_config_path_;
 
+typedef std::function<void(const std::string &data, bool finish)> task_callback_t;
+
 #define CONFIG_AUTO_SET(obj, key)             \
     if (config_body.contains(#key))           \
         mode_config_.key = config_body[#key]; \
@@ -50,16 +52,17 @@ class llm_task {
     std::string model_;
     std::string response_format_;
     std::vector<std::string> inputs_;
-    std::string kws_;
+    std::vector<std::string> kws_;
     bool enoutput_;
+    bool enoutput_json_;
     bool enstream_;
     bool enwake_audio_;
     std::atomic_bool audio_flage_;
+    task_callback_t out_callback_;
     int delay_audio_frame_ = 100;
     buffer_t *pcmdata;
     std::string wake_wav_file_;
 
-    std::function<void(const std::string &)> out_callback_;
     std::function<void(const std::string &)> play_awake_wav;
 
     bool parse_config(const nlohmann::json &config_body)
@@ -68,7 +71,6 @@ class llm_task {
             model_           = config_body.at("model");
             response_format_ = config_body.at("response_format");
             enoutput_        = config_body.at("enoutput");
-            kws_             = config_body.at("kws");
             if (config_body.contains("enwake_audio")) {
                 enwake_audio_ = config_body["enwake_audio"];
             } else {
@@ -83,6 +85,16 @@ class llm_task {
                     }
                 }
             }
+            if (config_body.contains("kws")) {
+                if (config_body["kws"].is_string()) {
+                    kws_.push_back(config_body["kws"].get<std::string>());
+                } else if (config_body["kws"].is_array()) {
+                    for (auto _in : config_body["kws"]) {
+                        kws_.push_back(_in.get<std::string>());
+                    }
+                }
+            }
+            enoutput_json_ = response_format_.find("json") == std::string::npos ? false : true;
         } catch (...) {
             SLOGE("setup config_body error");
             return true;
@@ -173,7 +185,9 @@ class llm_task {
             mode_config_.keywords_file                   = base_model + mode_config_.keywords_file;
 
             std::ofstream temp_awake_key("/tmp/kws_awake.txt.tmp");
-            temp_awake_key << kws_;
+            for (const auto &keyword : kws_) {
+                temp_awake_key << keyword << std::endl;
+            }
             temp_awake_key.close();
             std::ostringstream awake_key_compile_cmd;
             if (file_exists("/opt/m5stack/scripts/text2token.py"))
@@ -206,7 +220,7 @@ class llm_task {
         return 0;
     }
 
-    void set_output(std::function<void(const std::string &)> out_callback)
+    void set_output(task_callback_t out_callback)
     {
         out_callback_ = out_callback;
     }
@@ -242,16 +256,17 @@ class llm_task {
                 play_awake_wav(wake_wav_file_);
             }
             if (out_callback_) {
-                out_callback_("True");
+                if (enoutput_json_)
+                    out_callback_(r.AsJsonString(), true);
+                else
+                    out_callback_("", true);
             }
         }
     }
 
     void trigger()
     {
-        if (out_callback_) {
-            out_callback_("True");
-        }
+        if (out_callback_) out_callback_("", true);
     }
 
     bool delete_model()
@@ -301,6 +316,39 @@ class llm_kws : public StackFlow {
             });
     }
 
+    void task_output(const std::weak_ptr<llm_task> llm_task_obj_weak,
+                     const std::weak_ptr<llm_channel_obj> llm_channel_weak, const std::string &data, bool finish)
+    {
+        auto llm_task_obj = llm_task_obj_weak.lock();
+        auto llm_channel  = llm_channel_weak.lock();
+        if (!(llm_task_obj && llm_channel)) {
+            return;
+        }
+        std::string tmp_msg1;
+        const std::string *next_data = &data;
+        if (data.empty()) {
+            llm_channel->send(llm_task_obj->response_format_, true, LLM_NO_ERROR);
+            return;
+        }
+        if (finish) {
+            tmp_msg1  = data + ".";
+            next_data = &tmp_msg1;
+        }
+        if (llm_channel->enstream_) {
+            static int count = 0;
+            nlohmann::json data_body;
+            data_body["index"]  = count++;
+            data_body["delta"]  = (*next_data);
+            data_body["finish"] = finish;
+            if (finish) count = 0;
+            SLOGI("send stream:%s", next_data->c_str());
+            llm_channel->send(llm_task_obj->response_format_, data_body, LLM_NO_ERROR);
+        } else if (finish) {
+            SLOGI("send utf-8:%s", next_data->c_str());
+            llm_channel->send(llm_task_obj->response_format_, (*next_data), LLM_NO_ERROR);
+        }
+    }
+
     void play_awake_wav(const std::string &wav_file)
     {
         FILE *fp = fopen(wav_file.c_str(), "rb");
@@ -465,13 +513,9 @@ class llm_kws : public StackFlow {
             llm_channel->set_output(llm_task_obj->enoutput_);
             llm_channel->set_stream(llm_task_obj->enstream_);
             llm_task_obj->play_awake_wav = std::bind(&llm_kws::play_awake_wav, this, std::placeholders::_1);
-            llm_task_obj->set_output([_llm_task_obj, _llm_channel](const std::string &data) {
-                auto llm_task_obj = _llm_task_obj.lock();
-                auto llm_channel  = _llm_channel.lock();
-                if (llm_task_obj && llm_channel) {
-                    llm_channel->send(llm_task_obj->response_format_, true, LLM_NO_ERROR);
-                }
-            });
+            llm_task_obj->set_output(std::bind(&llm_kws::task_output, this, std::weak_ptr<llm_task>(llm_task_obj),
+                                               std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
+                                               std::placeholders::_2));
 
             for (const auto input : llm_task_obj->inputs_) {
                 if (input.find("sys") != std::string::npos) {
diff --git a/projects/llm_framework/main_melotts/src/main.cpp b/projects/llm_framework/main_melotts/src/main.cpp
index 6b8e1c15..c4e2122c 100644
--- a/projects/llm_framework/main_melotts/src/main.cpp
+++ b/projects/llm_framework/main_melotts/src/main.cpp
@@ -184,7 +184,10 @@ class llm_task {
             else if (file_body["mode_param"].contains("awake_delay"))
                 awake_delay_ = file_body["mode_param"]["awake_delay"];
 
-            if (!std::filesystem::exists(mode_config_.tagger) || !std::filesystem::exists(mode_config_.verbalizer)) {
+            if (!std::filesystem::exists(mode_config_.tagger) ||
+                !std::filesystem::is_regular_file(mode_config_.tagger) ||
+                !std::filesystem::exists(mode_config_.verbalizer) ||
+                !std::filesystem::is_regular_file(mode_config_.verbalizer)) {
                 SLOGW("Either tagger or verbalizer file does not exist, using alternative lexicon.");
                 lexicon_ = std::make_unique<Lexicon>(mode_config_.lexicon, mode_config_.tokens);
             } else {
@@ -281,195 +284,132 @@ class llm_task {
             int dec_len         = zp_size / zp_shape[1];
             int audio_slice_len = decoder_->GetOutputSize(0) / sizeof(float);
 
-            const int pad_frames        = 24;
-            const int samples_per_frame = 512;
+            const int overlap_size = 1024;
+            const int fade_size    = 512;
 
-            const int effective_frames = dec_len - 2 * pad_frames;
+            int dec_slice_num = static_cast<int>(std::ceil(static_cast<double>(zp_shape[2]) / dec_len));
 
-            int dec_slice_num =
-                static_cast<int>(std::ceil(static_cast<double>(zp_shape[2]) / static_cast<double>(effective_frames)));
-
-            const int sola_buffer_frame = pad_frames * samples_per_frame;
-            const int sola_search_frame = pad_frames * samples_per_frame;
-            const int block_frame       = (dec_len - 2 * pad_frames) * samples_per_frame;
-
-            std::vector<float> fade_in_window(sola_buffer_frame);
-            std::vector<float> fade_out_window(sola_buffer_frame);
-
-            for (int i = 0; i < sola_buffer_frame; i++) {
-                fade_in_window[i]  = static_cast<float>(i) / sola_buffer_frame;
-                fade_out_window[i] = 1.0f - fade_in_window[i];
+            std::vector<float> fade_in(fade_size);
+            std::vector<float> fade_out(fade_size);
+            for (int i = 0; i < fade_size; i++) {
+                float t     = static_cast<float>(i) / fade_size;
+                fade_in[i]  = t;
+                fade_out[i] = 1.0f - t;
             }
 
-            std::vector<float> sola_buffer(sola_buffer_frame, 0.0f);
-            bool first_frame = true;
-
             std::vector<float> pcmlist;
+            std::vector<float> previous_tail;
 
             for (int i = 0; i < dec_slice_num; i++) {
-                int input_start = i * effective_frames;
-                if (i > 0) {
-                    input_start -= pad_frames;
-                }
-                input_start = std::max(0, input_start);
-
-                int actual_len = std::min(dec_len, static_cast<int>(zp_shape[2] - input_start));
-
-                int output_start_frame, output_end_frame;
-
-                if (i == 0) {
-                    output_start_frame = 0;
-                    output_end_frame   = effective_frames - 1;
-                } else if (i == dec_slice_num - 1) {
-                    output_start_frame = i * effective_frames;
-                    output_end_frame   = static_cast<int>(zp_shape[2]) - 1;
-                } else {
-                    output_start_frame = i * effective_frames;
-                    output_end_frame   = (i + 1) * effective_frames - 1;
-                }
+                int input_start = i * dec_len;
+                int actual_size = std::min(dec_len, static_cast<int>(zp_shape[2] - input_start));
 
                 std::vector<float> zp(zp_size, 0);
-
                 for (int n = 0; n < zp_shape[1]; n++) {
-                    int copy_size = std::min(actual_len, static_cast<int>(zp_shape[2] - input_start));
-                    if (copy_size > 0) {
+                    if (actual_size > 0) {
                         memcpy(zp.data() + n * dec_len, zp_data + n * zp_shape[2] + input_start,
-                               sizeof(float) * copy_size);
+                               sizeof(float) * actual_size);
                     }
                 }
 
-                std::vector<float> decoder_output(audio_slice_len);
                 decoder_->SetInput(zp.data(), 0);
                 decoder_->SetInput(g_matrix.data(), 1);
 
                 if (0 != decoder_->Run()) {
+                    SLOGE("Decoder run failed at slice %d", i);
                     throw std::string("decoder_ RunSync error");
                 }
 
+                std::vector<float> decoder_output(audio_slice_len);
                 decoder_->GetOutput(decoder_output.data(), 0);
 
-                if (first_frame) {
-                    int audio_start = 0;
-                    int audio_len   = decoder_output.size() - sola_buffer_frame;
-                    audio_len       = std::max(0, audio_len);
-
-                    if (audio_len > 0) {
-                        pcmlist.insert(pcmlist.end(), decoder_output.begin() + audio_start,
-                                       decoder_output.begin() + audio_start + audio_len);
-                    }
+                if (i == 0) {
+                    int main_part_size = static_cast<int>(decoder_output.size()) - overlap_size;
+                    main_part_size     = std::max(0, main_part_size);
 
-                    int buffer_start = audio_len;
+                    pcmlist.insert(pcmlist.end(), decoder_output.begin(), decoder_output.begin() + main_part_size);
 
-                    if (buffer_start + sola_buffer_frame <= decoder_output.size()) {
-                        std::copy(decoder_output.begin() + buffer_start,
-                                  decoder_output.begin() + buffer_start + sola_buffer_frame, sola_buffer.begin());
-                    } else {
-                        int available = static_cast<int>(decoder_output.size() - buffer_start);
-                        if (available > 0) {
-                            std::copy(decoder_output.begin() + buffer_start, decoder_output.end(), sola_buffer.begin());
-                            std::fill(sola_buffer.begin() + available, sola_buffer.end(), 0.0f);
-                        } else {
-                            std::fill(sola_buffer.begin(), sola_buffer.end(), 0.0f);
-                        }
+                    if (decoder_output.size() > main_part_size) {
+                        previous_tail.assign(decoder_output.begin() + main_part_size, decoder_output.end());
                     }
 
-                    first_frame = false;
-
                 } else {
-                    int audio_start = pad_frames * samples_per_frame;
-
-                    std::vector<float> search_window(sola_buffer_frame + sola_search_frame);
-                    std::copy(decoder_output.begin() + audio_start,
-                              decoder_output.begin() + audio_start + search_window.size(), search_window.begin());
-
-                    int best_offset        = 0;
-                    float best_correlation = -1.0;
-
-                    for (int offset = 0; offset <= sola_search_frame; offset++) {
-                        float correlation = 0.0;
-                        float energy      = 0.0;
-
-                        for (int j = 0; j < sola_buffer_frame; j++) {
-                            correlation += sola_buffer[j] * search_window[j + offset];
-                            energy += search_window[j + offset] * search_window[j + offset];
-                        }
+                    if (previous_tail.empty()) {
+                        pcmlist.insert(pcmlist.end(), decoder_output.begin(), decoder_output.end());
+                        continue;
+                    }
 
-                        float normalized_correlation = (energy > 1e-8) ? correlation / std::sqrt(energy) : 0.0f;
+                    int blend_size = std::min(
+                        {fade_size, static_cast<int>(previous_tail.size()), static_cast<int>(decoder_output.size())});
 
-                        if (normalized_correlation > best_correlation) {
-                            best_correlation = normalized_correlation;
-                            best_offset      = offset;
-                        }
+                    std::vector<float> blended_region(blend_size);
+                    for (int j = 0; j < blend_size; j++) {
+                        blended_region[j] = previous_tail[j] * fade_out[j * fade_size / blend_size] +
+                                            decoder_output[j] * fade_in[j * fade_size / blend_size];
                     }
 
-                    int aligned_start = audio_start + best_offset;
-
-                    std::vector<float> crossfade_region(sola_buffer_frame);
+                    pcmlist.insert(pcmlist.end(), blended_region.begin(), blended_region.end());
 
-                    for (int j = 0; j < sola_buffer_frame; j++) {
-                        crossfade_region[j] =
-                            decoder_output[aligned_start + j] * fade_in_window[j] + sola_buffer[j] * fade_out_window[j];
+                    if (static_cast<int>(previous_tail.size()) > blend_size) {
+                        pcmlist.insert(pcmlist.end(), previous_tail.begin() + blend_size, previous_tail.end());
                     }
 
-                    pcmlist.insert(pcmlist.end(), crossfade_region.begin(), crossfade_region.end());
-
-                    int remaining_start = aligned_start + sola_buffer_frame;
+                    int current_remaining_start = blend_size;
+                    int current_remaining_size  = static_cast<int>(decoder_output.size()) - current_remaining_start;
 
                     if (i == dec_slice_num - 1) {
-                        int total_expected_samples = audio_len * samples_per_frame / 512;
-                        int processed_samples      = static_cast<int>(pcmlist.size());
-                        int remaining_needed       = total_expected_samples - processed_samples;
-                        remaining_needed           = std::max(0, remaining_needed);
+                        int total_expected     = audio_len;
+                        int current_total      = static_cast<int>(pcmlist.size());
+                        current_remaining_size = std::min(current_remaining_size, total_expected - current_total);
+                    }
 
-                        int remaining_len =
-                            std::min(remaining_needed, static_cast<int>(decoder_output.size() - remaining_start));
+                    if (current_remaining_size > overlap_size && i < dec_slice_num - 1) {
+                        int main_part_size = current_remaining_size - overlap_size;
 
-                        if (remaining_len > 0) {
-                            pcmlist.insert(pcmlist.end(), decoder_output.begin() + remaining_start,
-                                           decoder_output.begin() + remaining_start + remaining_len);
-                        }
+                        pcmlist.insert(pcmlist.end(), decoder_output.begin() + current_remaining_start,
+                                       decoder_output.begin() + current_remaining_start + main_part_size);
 
+                        previous_tail.assign(decoder_output.begin() + current_remaining_start + main_part_size,
+                                             decoder_output.begin() + current_remaining_start + current_remaining_size);
                     } else {
-                        int remaining_len = (dec_len - 2 * pad_frames) * samples_per_frame - sola_buffer_frame;
-                        remaining_len =
-                            std::min(remaining_len, static_cast<int>(decoder_output.size() - remaining_start));
-
-                        if (remaining_len > 0) {
-                            pcmlist.insert(pcmlist.end(), decoder_output.begin() + remaining_start,
-                                           decoder_output.begin() + remaining_start + remaining_len);
-                        }
-
-                        int buffer_start = remaining_start + remaining_len;
-
-                        if (buffer_start + sola_buffer_frame <= decoder_output.size()) {
-                            std::copy(decoder_output.begin() + buffer_start,
-                                      decoder_output.begin() + buffer_start + sola_buffer_frame, sola_buffer.begin());
-                        } else {
-                            int avail = static_cast<int>(decoder_output.size() - buffer_start);
-                            if (avail > 0) {
-                                std::copy(decoder_output.begin() + buffer_start, decoder_output.end(),
-                                          sola_buffer.begin());
-                            }
-                            std::fill(sola_buffer.begin() + avail, sola_buffer.end(), 0.0f);
+                        if (current_remaining_size > 0) {
+                            pcmlist.insert(pcmlist.end(), decoder_output.begin() + current_remaining_start,
+                                           decoder_output.begin() + current_remaining_start + current_remaining_size);
                         }
+                        previous_tail.clear();
                     }
                 }
+
+                if (static_cast<int>(pcmlist.size()) >= audio_len) {
+                    break;
+                }
             }
 
-            if (pcmlist.size() > audio_len) {
+            if (static_cast<int>(pcmlist.size()) > audio_len) {
                 pcmlist.resize(audio_len);
             }
 
+            float max_val  = 0.0f;
+            int clip_count = 0;
+            for (float sample : pcmlist) {
+                max_val = std::max(max_val, std::abs(sample));
+                if (std::abs(sample) > 0.95f) clip_count++;
+            }
+
             double src_ratio =
                 static_cast<double>(mode_config_.audio_rate) / static_cast<double>(mode_config_.mode_rate);
-            std::vector<float> tmp_pcm((pcmlist.size() * src_ratio + 1));
+            std::vector<float> tmp_pcm(static_cast<size_t>(pcmlist.size() * src_ratio + 1));
             int len;
-
             resample_audio(pcmlist.data(), pcmlist.size(), tmp_pcm.data(), &len, src_ratio);
 
             wav_pcm_data.reserve(len);
-            std::transform(tmp_pcm.begin(), tmp_pcm.begin() + len, std::back_inserter(wav_pcm_data),
-                           [](const auto val) { return static_cast<int16_t>(val * INT16_MAX); });
+            for (int i = 0; i < len; i++) {
+                float val = tmp_pcm[i];
+                if (std::abs(val) > 0.95f) {
+                    val = val > 0 ? 0.95f : -0.95f;
+                }
+                wav_pcm_data.push_back(static_cast<int16_t>(val * INT16_MAX));
+            }
 
             if (out_callback_) {
                 out_callback_(
@@ -478,8 +418,10 @@ class llm_task {
             }
 
         } catch (const std::exception &e) {
+            SLOGE("Exception: %s", e.what());
             return true;
         } catch (...) {
+            SLOGE("Unknown exception occurred");
             return true;
         }
         return false;
diff --git a/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp b/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp
index 1884c929..cf2c1fcb 100644
--- a/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp
+++ b/projects/llm_framework/main_melotts/src/runner/Lexicon.hpp
@@ -37,7 +37,7 @@ class Lexicon {
             const std::string& verbalizer_filename)
         : max_phrase_length(0)
     {
-        SLOGD("Dictionary loading: %s Pronunciation table loading: %s tagger_filename: %s verbalizer_filename: %s",
+        SLOGI("Dictionary loading: %s Pronunciation table loading: %s tagger_filename: %s verbalizer_filename: %s",
               tokens_filename.c_str(), lexicon_filename.c_str(), tagger_filename.c_str(), verbalizer_filename.c_str());
 
         m_processor = new wetext::Processor(tagger_filename, verbalizer_filename);
@@ -92,13 +92,13 @@ class Lexicon {
         lexicon["。"] = lexicon["."];
         lexicon["！"] = lexicon["!"];
         lexicon["？"] = lexicon["?"];
-        SLOGD("Dictionary loading complete, containing %zu entries, longest phrase length: %zu", lexicon.size(),
+        SLOGI("Dictionary loading complete, containing %zu entries, longest phrase length: %zu", lexicon.size(),
               max_phrase_length);
     }
 
     Lexicon(const std::string& lexicon_filename, const std::string& tokens_filename) : max_phrase_length(0)
     {
-        SLOGD("Dictionary loading: %s Pronunciation table loading: %s", tokens_filename.c_str(),
+        SLOGI("Dictionary loading: %s Pronunciation table loading: %s", tokens_filename.c_str(),
               lexicon_filename.c_str());
 
         std::unordered_map<std::string, int> tokens;
@@ -151,7 +151,7 @@ class Lexicon {
         lexicon["。"] = lexicon["."];
         lexicon["！"] = lexicon["!"];
         lexicon["？"] = lexicon["?"];
-        SLOGD("Dictionary loading complete, containing %zu entries, longest phrase length: %zu", lexicon.size(),
+        SLOGI("Dictionary loading complete, containing %zu entries, longest phrase length: %zu", lexicon.size(),
               max_phrase_length);
     }
 

From c25b4f0a093fcca647b2851b3472c36e9510384f Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Wed, 23 Jul 2025 14:26:22 +0800
Subject: [PATCH 26/79] [fix] Fix caching causing audio issues

---
 .../llm_framework/main_melotts/src/runner/EngineWrapper.cpp     | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/projects/llm_framework/main_melotts/src/runner/EngineWrapper.cpp b/projects/llm_framework/main_melotts/src/runner/EngineWrapper.cpp
index 0dda3e7b..0394ee48 100644
--- a/projects/llm_framework/main_melotts/src/runner/EngineWrapper.cpp
+++ b/projects/llm_framework/main_melotts/src/runner/EngineWrapper.cpp
@@ -292,7 +292,7 @@ int EngineWrapper::Init(const char* strModelPath, uint32_t nNpuType)
 
     // 6. prepare io
     // AX_U32 nIoDepth = (stCtx.vecOutputBufferFlag.size() == 0) ? 1 : stCtx.vecOutputBufferFlag.size();
-    ret = utils::prepare_io(strModelPath, m_io_info, m_io, utils::IO_BUFFER_STRATEGY_CACHED);
+    ret = utils::prepare_io(strModelPath, m_io_info, m_io, utils::IO_BUFFER_STRATEGY_DEFAULT);
     if (0 != ret) {
         printf("prepare io failed!\n");
         utils::free_io(m_io);

From cfbfd62c4acf4f915b4498a173fb6c7ac0394a41 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 14 Aug 2025 09:15:49 +0800
Subject: [PATCH 27/79] [update] update docs

---
 doc/projects_llm_framework_doc/llm_kws_en.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/doc/projects_llm_framework_doc/llm_kws_en.md b/doc/projects_llm_framework_doc/llm_kws_en.md
index 4ab85b47..84b2a666 100644
--- a/doc/projects_llm_framework_doc/llm_kws_en.md
+++ b/doc/projects_llm_framework_doc/llm_kws_en.md
@@ -34,7 +34,7 @@ Send JSON:
 - response_format: The result returned is in `kws.bool` format.
 - input: The input is `sys.pcm`, representing system audio.
 - enoutput: Whether to enable user result output.
-- kws: The Chinese wake-up word is `"你好你好"`.
+- kws: The English wake-up word is `"HELLO"`. It must be capital letters.
 - enwake_audio: Whether to enable wake-up audio output. Default is true.
 
 Response JSON:

From a916ca04d64746153da363dc966ebc6730468a1d Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 21 Aug 2025 19:08:20 +0800
Subject: [PATCH 28/79] [update] Reduce buffer frames

---
 projects/llm_framework/main_asr/src/main.cpp | 6 +++---
 projects/llm_framework/main_kws/src/main.cpp | 4 ++--
 projects/llm_framework/main_vad/src/main.cpp | 4 ++--
 3 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/projects/llm_framework/main_asr/src/main.cpp b/projects/llm_framework/main_asr/src/main.cpp
index c3bd64f2..4a5f9703 100644
--- a/projects/llm_framework/main_asr/src/main.cpp
+++ b/projects/llm_framework/main_asr/src/main.cpp
@@ -58,7 +58,7 @@ class llm_task {
     std::atomic_bool audio_flage_;
     std::atomic_bool awake_flage_;
     int awake_delay_       = 50;
-    int delay_audio_frame_ = 100;
+    int delay_audio_frame_ = 10;
     buffer_t *pcmdata;
 
     std::function<void(void)> pause;
@@ -183,11 +183,11 @@ class llm_task {
     {
         static int count = 0;
         if (count < delay_audio_frame_) {
-            buffer_write_char(pcmdata, raw.c_str(), raw.length());
+            buffer_write_char(pcmdata, raw.data(), raw.length());
             count++;
             return;
         }
-        buffer_write_char(pcmdata, raw.c_str(), raw.length());
+        buffer_write_char(pcmdata, raw.data(), raw.length());
         buffer_position_set(pcmdata, 0);
         count = 0;
         std::vector<float> floatSamples;
diff --git a/projects/llm_framework/main_kws/src/main.cpp b/projects/llm_framework/main_kws/src/main.cpp
index 5b8a21bb..18e4dfd2 100644
--- a/projects/llm_framework/main_kws/src/main.cpp
+++ b/projects/llm_framework/main_kws/src/main.cpp
@@ -59,7 +59,7 @@ class llm_task {
     bool enwake_audio_;
     std::atomic_bool audio_flage_;
     task_callback_t out_callback_;
-    int delay_audio_frame_ = 100;
+    int delay_audio_frame_ = 10;
     buffer_t *pcmdata;
     std::string wake_wav_file_;
 
@@ -233,7 +233,7 @@ class llm_task {
             count++;
             return;
         }
-        buffer_write_char(pcmdata, raw.c_str(), raw.length());
+        buffer_write_char(pcmdata, raw.data(), raw.length());
         buffer_position_set(pcmdata, 0);
         count = 0;
         std::vector<float> floatSamples;
diff --git a/projects/llm_framework/main_vad/src/main.cpp b/projects/llm_framework/main_vad/src/main.cpp
index 3bd53f3f..20a67eb3 100644
--- a/projects/llm_framework/main_vad/src/main.cpp
+++ b/projects/llm_framework/main_vad/src/main.cpp
@@ -60,7 +60,7 @@ class llm_task {
     std::string superior_id_;
     task_callback_t out_callback_;
     int awake_delay_       = 50;
-    int delay_audio_frame_ = 100;
+    int delay_audio_frame_ = 3;
     buffer_t *pcmdata;
     std::string wake_wav_file_;
 
@@ -158,7 +158,7 @@ class llm_task {
             count++;
             return;
         }
-        buffer_write_char(pcmdata, raw.c_str(), raw.length());
+        buffer_write_char(pcmdata, raw.data(), raw.length());
         buffer_position_set(pcmdata, 0);
         count = 0;
         std::vector<float> floatSamples;

From 73c4a49b05125c2484940bc15ce2e1948366f7c6 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 22 Aug 2025 17:38:39 +0800
Subject: [PATCH 29/79] [update] ModuleLLM support ctx model, add HomeAssistant
 model, add model post process config.

---
 .../mode_qwen2.5-HA-0.5B-ctx-ax630c.json      |   38 +
 .../tokenizer_qwen2.5-HA-0.5B-ctx-ax630c.py   |  203 ++++
 projects/llm_framework/main_llm/src/main.cpp  |  148 ++-
 .../llm_framework/main_llm/src/runner/LLM.hpp | 1066 +++++++++++++++--
 .../main_llm/src/runner/LLMPostprocess.hpp    |   46 +-
 .../src/runner/Tokenizer/Tokenizer.cpp        |  618 +++++-----
 .../src/runner/Tokenizer/Tokenizer.hpp        |   90 +-
 .../ax_model_runner/ax_model_runner.hpp       |    3 +
 .../main_llm/src/runner/utils/http_utils.hpp  |  102 ++
 projects/llm_framework/tools/llm_pack.py      |    1 +
 10 files changed, 1826 insertions(+), 489 deletions(-)
 create mode 100644 projects/llm_framework/main_llm/models/mode_qwen2.5-HA-0.5B-ctx-ax630c.json
 create mode 100644 projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-HA-0.5B-ctx-ax630c.py
 create mode 100644 projects/llm_framework/main_llm/src/runner/utils/http_utils.hpp

diff --git a/projects/llm_framework/main_llm/models/mode_qwen2.5-HA-0.5B-ctx-ax630c.json b/projects/llm_framework/main_llm/models/mode_qwen2.5-HA-0.5B-ctx-ax630c.json
new file mode 100644
index 00000000..1481d454
--- /dev/null
+++ b/projects/llm_framework/main_llm/models/mode_qwen2.5-HA-0.5B-ctx-ax630c.json
@@ -0,0 +1,38 @@
+{
+    "mode":"qwen2.5-HA-0.5B-ctx-ax630c",
+    "type":"llm",
+    "homepage":"https://huggingface.co/yunyu1258/qwen2.5-0.5b-ha",
+    "compile_flage":"pulsar2 llm_build --input_path Qwen/qwen2.5-0.5b-ha --output_path Qwen/qwen2.5-0.5B-p1024-ha-ax630c --hidden_state_type bf16 --prefill_len 128 --kv_cache_len 1280 --last_kv_cache_len 128 --last_kv_cache_len 512 --last_kv_cache_len 1024 --chip AX620E --parallel 24",
+    "pulsar_version":"4.1-patch1-c37957c7",
+    "capabilities":[
+        "text_generation",
+        "chat"
+    ],
+    "input_type":[
+        "llm.utf-8",
+        "llm.utf-8.stream",
+        "llm.chat_completion",
+        "llm.chat_completion.stream"
+    ],
+    "output_type":[
+        "llm.utf-8",
+        "llm.utf-8.stream"
+    ],
+    "mode_param":{
+        "tokenizer_type":2,
+        "url_tokenizer_model":"http://localhost:8080",
+        "filename_tokens_embed":"model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel":"qwen2_post.axmodel",
+        "template_filename_axmodel":"qwen2_p128_l%d_together.axmodel",
+        "b_use_topk":false,
+        "b_bos":false,
+        "b_eos":false,
+        "axmodel_num":24,
+        "tokens_embed_num":151936,
+        "tokens_embed_size":896,
+        "b_use_mmap_load_embed":true,
+        "b_dynamic_load_axmodel_layer":false,
+        "precompute_len":1202,
+        "ext_scripts":["tokenizer_qwen2.5-HA-0.5B-ctx-ax630c.py"]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-HA-0.5B-ctx-ax630c.py b/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-HA-0.5B-ctx-ax630c.py
new file mode 100644
index 00000000..4a63e88a
--- /dev/null
+++ b/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-HA-0.5B-ctx-ax630c.py
@@ -0,0 +1,203 @@
+from transformers import AutoTokenizer, PreTrainedTokenizerFast
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+import uuid
+
+# 全局字典：存储 uid 到 Tokenizer_Http 实例的映射
+tokenizers = {}
+
+class Tokenizer_Http():
+    def __init__(self, model_id):
+        self.tokenizer = AutoTokenizer.from_pretrained(model_id)
+        self.messages = [
+            {"role": "system", "content": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant."},
+        ]
+        self.token_ids = []
+        
+        self.token_ids_cache = []
+
+    def encode(self, prompt, last_reply=None):
+        if last_reply is not None:
+            self.messages.append({"role": "assistant", "content": last_reply})
+            text = self.tokenizer.apply_chat_template(
+                self.messages,
+                tokenize=False,
+                add_generation_prompt=True
+            )
+            # print("生成的文本:\n============\n", text, "============\n")
+            self.token_ids = self.tokenizer.encode(text)[:-3]
+        self.messages.append({"role": "user", "content": prompt})
+        
+        text = self.tokenizer.apply_chat_template(
+            self.messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        print("生成的文本:\n============\n", text, "============\n")
+        token_ids = self.tokenizer.encode(text)
+        # 找出新增部分
+        diff = token_ids[len(self.token_ids):]
+        self.token_ids = token_ids
+        print(self.decode(diff))
+        return token_ids, diff
+
+    def decode(self, token_ids):
+        self.token_ids_cache += token_ids
+        text = self.tokenizer.decode(self.token_ids_cache)
+        if "\ufffd" in text:
+            print("text 中包含非法字符")
+            return ""
+        else:
+            self.token_ids_cache.clear()
+            return text
+        
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+    
+    @property
+    def bos_token(self):
+        return self.tokenizer.bos_token
+
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+    
+    def reset(self, system_prompt=None):
+        if system_prompt is None:
+            system_prompt = args.content
+        self.messages = [
+            {"role": "system", "content": system_prompt},
+        ]
+        text = self.tokenizer.apply_chat_template(
+            self.messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        token_ids = self.tokenizer.encode(text)[:-3]
+        self.token_ids = token_ids
+        print(self.decode(token_ids))
+        return token_ids
+        
+
+class Request(BaseHTTPRequestHandler):
+    timeout = 5
+    server_version = 'Apache'
+
+    def do_GET(self):
+        print("GET 请求路径:", self.path)
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.end_headers()
+
+        # 新增接口：获取 uid
+        if '/get_uid' in self.path:
+            new_uid = str(uuid.uuid4())
+            print("新 uid:", new_uid)
+            # 为该 uid 创建一个新的 Tokenizer_Http 实例
+            tokenizers[new_uid] = Tokenizer_Http(args.model_id)
+            msg = json.dumps({'uid': new_uid})
+        elif '/bos_id' in self.path:
+            # 获取 uid 参数（例如 ?uid=xxx）
+            uid = self.get_query_param("uid")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                bos_id = instance.bos_id
+                msg = json.dumps({'bos_id': bos_id if bos_id is not None else -1})
+        elif '/eos_id' in self.path:
+            uid = self.get_query_param("uid")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                eos_id = instance.eos_id
+                msg = json.dumps({'eos_id': eos_id if eos_id is not None else -1})
+        else:
+            msg = json.dumps({'error': 'Invalid GET endpoint'})
+
+        print("响应消息:", msg)
+        self.wfile.write(msg.encode())
+
+    def do_POST(self):
+        content_length = int(self.headers.get('content-length', 0))
+        data = self.rfile.read(content_length).decode()
+        print("POST 请求路径:", self.path)
+        print("接收到的数据:", data)
+        req = json.loads(data)
+
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.end_headers()
+
+        if '/encode' in self.path:
+            # 请求数据中必须包含 uid, text, 和可选的 last_reply
+            uid = req.get('uid')
+            prompt = req.get('text')
+            last_reply = req.get('last_reply')
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                token_ids, diff = instance.encode(prompt, last_reply)
+                msg = json.dumps({'token_ids': token_ids, 'diff': diff})
+        elif '/decode' in self.path:
+            uid = req.get('uid')
+            token_ids = req.get('token_ids')
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                text = instance.decode(token_ids)
+                msg = json.dumps({'text': text})
+        elif '/reset' in self.path:
+            uid = req.get("uid")
+            system_prompt = req.get("system_prompt")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                if system_prompt is not None:
+                    print("system_prompt:", system_prompt)
+                    token_ids = instance.reset(system_prompt)
+                    msg = json.dumps({'token_ids': token_ids})
+                else:
+                    token_ids = instance.reset()
+                    msg = json.dumps({'token_ids': token_ids})
+        else:
+            msg = json.dumps({'error': 'Invalid POST endpoint'})
+
+        print("响应消息:", msg)
+        self.wfile.write(msg.encode())
+
+    def get_query_param(self, key):
+        """
+        辅助函数：从 GET 请求的 URL 中获取查询参数的值
+        例如：/bos_id?uid=xxx
+        """
+        from urllib.parse import urlparse, parse_qs
+        query = urlparse(self.path).query
+        params = parse_qs(query)
+        values = params.get(key)
+        return values[0] if values else None
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--host', type=str, default='0.0.0.0')
+    parser.add_argument('--port', type=int, default=12345)
+    parser.add_argument('--model_id', type=str, default='qwen3_1.7B_tokenizer')
+    parser.add_argument('--content', type=str, default='You are Qwen, created by Alibaba Cloud. You are a helpful assistant.')
+
+    args = parser.parse_args()
+
+    host = (args.host, args.port)
+    print('Server running at http://%s:%s' % host)
+    server = HTTPServer(host, Request)
+    server.serve_forever()
diff --git a/projects/llm_framework/main_llm/src/main.cpp b/projects/llm_framework/main_llm/src/main.cpp
index 7cbb9fbe..e198cc9c 100644
--- a/projects/llm_framework/main_llm/src/main.cpp
+++ b/projects/llm_framework/main_llm/src/main.cpp
@@ -55,10 +55,17 @@ class llm_task {
     enum inference_status { INFERENCE_NONE = 0, INFERENCE_RUNNING };
     LLMAttrType mode_config_;
     std::unique_ptr<LLM> lLaMa_;
+    std::unique_ptr<LLM_CTX> lLaMa_ctx_;
     std::string model_;
     std::string response_format_;
     std::vector<std::string> inputs_;
     std::string prompt_;
+    std::string last_reply;
+    std::vector<unsigned short> prompt_data;
+    std::vector<int> tokens_ids, tokens_diff;
+    std::vector<std::vector<unsigned short>> k_caches, v_caches;
+    int precompute_len = 0;
+    std::vector<int> _token_ids;
     task_callback_t out_callback_;
     bool enoutput_;
     bool enstream_;
@@ -125,10 +132,10 @@ class llm_task {
 
             CONFIG_AUTO_SET(file_body["mode_param"], tokenizer_type);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_tokenizer_model);
+            CONFIG_AUTO_SET(file_body["mode_param"], url_tokenizer_model);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_tokens_embed);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_post_axmodel);
             CONFIG_AUTO_SET(file_body["mode_param"], template_filename_axmodel);
-            CONFIG_AUTO_SET(file_body["mode_param"], b_use_topk);
             CONFIG_AUTO_SET(file_body["mode_param"], b_bos);
             CONFIG_AUTO_SET(file_body["mode_param"], b_eos);
             CONFIG_AUTO_SET(file_body["mode_param"], axmodel_num);
@@ -137,61 +144,119 @@ class llm_task {
             CONFIG_AUTO_SET(file_body["mode_param"], b_use_mmap_load_embed);
             CONFIG_AUTO_SET(file_body["mode_param"], b_dynamic_load_axmodel_layer);
             CONFIG_AUTO_SET(file_body["mode_param"], max_token_len);
+            CONFIG_AUTO_SET(file_body["mode_param"], enable_temperature);
             CONFIG_AUTO_SET(file_body["mode_param"], temperature);
+            CONFIG_AUTO_SET(file_body["mode_param"], enable_top_p_sampling);
             CONFIG_AUTO_SET(file_body["mode_param"], top_p);
+            CONFIG_AUTO_SET(file_body["mode_param"], enable_top_k_sampling);
+            CONFIG_AUTO_SET(file_body["mode_param"], top_k);
+            CONFIG_AUTO_SET(file_body["mode_param"], enable_repetition_penalty);
+            CONFIG_AUTO_SET(file_body["mode_param"], repetition_penalty);
+            CONFIG_AUTO_SET(file_body["mode_param"], penalty_window);
+            CONFIG_AUTO_SET(file_body["mode_param"], precompute_len);
+            {
+                auto has_http = [](const std::string &s) { return s.find("http") != std::string::npos; };
+
+                auto find_tokenizer_file = [this]() -> std::string {
+                    const std::string base = "/opt/m5stack/scripts/";
+                    const std::string a    = base + model_ + "_tokenizer.py";
+                    if (file_exists(a)) return a;
+                    const std::string b = base + "tokenizer_" + model_ + ".py";
+                    if (file_exists(b)) return b;
+                    SLOGE("%s or %s not found!", a.c_str(), b.c_str());
+                    return {};
+                };
+
+                auto start_tokenizer_server = [&](const std::string &tokenizer_file) {
+                    if (tokenizer_file.empty()) return;
+                    if (tokenizer_server_flage_.load()) return;
 
-            if (mode_config_.filename_tokenizer_model.find("http:") != std::string::npos) {
-                mode_config_.filename_tokenizer_model = "http://localhost:" + std::to_string(port_);
-                std::string tokenizer_file;
-                if (file_exists(std::string("/opt/m5stack/scripts/") + model_ + std::string("_tokenizer.py"))) {
-                    tokenizer_file = std::string("/opt/m5stack/scripts/") + model_ + std::string("_tokenizer.py");
-                } else if (file_exists(std::string("/opt/m5stack/scripts/") + std::string("tokenizer_") + model_ +
-                                       std::string(".py"))) {
-                    tokenizer_file =
-                        std::string("/opt/m5stack/scripts/") + std::string("tokenizer_") + model_ + std::string(".py");
-                } else {
-                    std::string __log = model_ + std::string("_tokenizer.py");
-                    __log += " or ";
-                    __log += std::string("tokenizer_") + model_ + std::string(".py");
-                    __log += " not found!";
-                    SLOGE("%s", __log.c_str());
-                }
-                if (!tokenizer_server_flage_.load()) {
                     tokenizer_pid_ = fork();
                     if (tokenizer_pid_ == 0) {
                         setenv("PYTHONPATH", "/opt/m5stack/lib/llm/site-packages", 1);
+                        const std::string port_str = std::to_string(port_);
+                        const std::string model_id = base_model + "tokenizer";
+
                         execl("/usr/bin/python3", "python3", tokenizer_file.c_str(), "--host", "localhost", "--port",
-                              std::to_string(port_).c_str(), "--model_id", (base_model + "tokenizer").c_str(),
-                              "--content", ("'" + prompt_ + "'").c_str(), nullptr);
+                              port_str.c_str(), "--model_id", model_id.c_str(), "--content", prompt_.c_str(),
+                              (char *)nullptr);
+
                         perror("execl failed");
-                        exit(1);
+                        _exit(1);
                     }
+
                     tokenizer_server_flage_.store(true);
                     SLOGI("port_=%s model_id=%s content=%s", std::to_string(port_).c_str(),
-                          (base_model + "tokenizer").c_str(), ("'" + prompt_ + "'").c_str());
+                          (base_model + std::string("tokenizer")).c_str(), prompt_.c_str());
+
                     std::this_thread::sleep_for(std::chrono::seconds(15));
+                };
+
+                auto process_field = [&](std::string &field, const char *name_for_log) -> bool {
+                    if (!has_http(field)) return false;
+
+                    field                            = "http://localhost:" + std::to_string(port_);
+                    const std::string tokenizer_file = find_tokenizer_file();
+                    start_tokenizer_server(tokenizer_file);
+                    SLOGI("%s: %s", name_for_log, field.c_str());
+                    return true;
+                };
+
+                if (!process_field(mode_config_.filename_tokenizer_model, "filename_tokenizer_model") &&
+                    !process_field(mode_config_.url_tokenizer_model, "url_tokenizer_model")) {
+                    mode_config_.filename_tokenizer_model = base_model + mode_config_.filename_tokenizer_model;
+                    SLOGE("filename_tokenizer_model: %s", mode_config_.filename_tokenizer_model.c_str());
                 }
-            } else {
-                mode_config_.filename_tokenizer_model = base_model + mode_config_.filename_tokenizer_model;
             }
-            SLOGI("filename_tokenizer_model: %s", mode_config_.filename_tokenizer_model.c_str());
             mode_config_.filename_tokens_embed     = base_model + mode_config_.filename_tokens_embed;
             mode_config_.filename_post_axmodel     = base_model + mode_config_.filename_post_axmodel;
             mode_config_.template_filename_axmodel = base_model + mode_config_.template_filename_axmodel;
-
             mode_config_.runing_callback = [this](int *p_token, int n_token, const char *p_str, float token_per_sec,
                                                   void *reserve) {
                 if (this->out_callback_) {
                     this->out_callback_(std::string(p_str), false);
                 }
             };
-            lLaMa_ = std::make_unique<LLM>();
-            if (!lLaMa_->Init(mode_config_)) {
-                lLaMa_->Deinit();
-                lLaMa_.reset();
-                return -2;
+
+            if (mode_config_.precompute_len > 0) {
+                lLaMa_ctx_ = std::make_unique<LLM_CTX>();
+                if (!lLaMa_ctx_->Init(mode_config_)) {
+                    lLaMa_ctx_->Deinit();
+                    lLaMa_ctx_.reset();
+                    return -2;
+                }
+            } else {
+                lLaMa_ = std::make_unique<LLM>();
+                if (!lLaMa_->Init(mode_config_)) {
+                    lLaMa_->Deinit();
+                    lLaMa_.reset();
+                    return -2;
+                }
             }
 
+            if (lLaMa_ctx_) {
+                lLaMa_ctx_->SetSystemPrompt(mode_config_.system_prompt, _token_ids);
+                std::string kvcache_path = "/tmp/.llm/";
+                if (!kvcache_path.empty() && kvcache_path != "") {
+                    if (lLaMa_ctx_->load_kvcache(kvcache_path, mode_config_.axmodel_num, k_caches, v_caches,
+                                                 mode_config_.system_prompt, precompute_len)) {
+                        ALOGI("load kvcache from path: %s success,precompute_len: %d", kvcache_path.c_str(),
+                              precompute_len);
+                    } else {
+                        ALOGW("load kvcache from path: %s failed,generate kvcache", kvcache_path.c_str());
+                        lLaMa_ctx_->GenerateKVCachePrefill(_token_ids, k_caches, v_caches, precompute_len);
+                        if (!lLaMa_ctx_->save_kvcache(kvcache_path, mode_config_.system_prompt, precompute_len,
+                                                      k_caches, v_caches)) {
+                            ALOGE("save kvcache failed");
+                        }
+                        ALOGI("generate kvcache to path: %s", kvcache_path.c_str());
+                    }
+                } else {
+                    lLaMa_ctx_->GenerateKVCachePrefill(_token_ids, k_caches, v_caches, precompute_len);
+                }
+                ALOGI("precompute_len: %d", precompute_len);
+                ALOGI("system_prompt: %s", mode_config_.system_prompt.c_str());
+            }
         } catch (...) {
             SLOGE("config false");
             return -3;
@@ -253,8 +318,25 @@ class llm_task {
     {
 #if 1
         try {
-            std::string out = lLaMa_->Run(prompt_complete(msg));
-            if (out_callback_) out_callback_(out, true);
+            if (lLaMa_) {
+                std::string out = lLaMa_->Run(prompt_complete(msg));
+                if (out_callback_) out_callback_(out, true);
+            }
+
+            if (lLaMa_ctx_) {
+                lLaMa_ctx_->Encode(prompt_data, prompt_complete(msg), last_reply, tokens_ids, tokens_diff);
+                if (auto ret = lLaMa_ctx_->SetKVCache(k_caches, v_caches, precompute_len, tokens_diff.size());
+                    ret != 0) {
+                    ALOGE("SetKVCache failed: %d,the context may be full,input \"reset\" to reset context", ret);
+                    // raise;
+                    lLaMa_ctx_->SetSystemPrompt(mode_config_.system_prompt, _token_ids);
+                    lLaMa_ctx_->GenerateKVCachePrefill(_token_ids, k_caches, v_caches, precompute_len);
+                    lLaMa_ctx_->SetKVCache(k_caches, v_caches, precompute_len, tokens_diff.size());
+                }
+                last_reply = lLaMa_ctx_->Run(prompt_data);
+                lLaMa_ctx_->GetKVCache(k_caches, v_caches, precompute_len);
+                if (out_callback_) out_callback_(last_reply, true);
+            }
         } catch (...) {
             SLOGW("lLaMa_->Run have error!");
         }
diff --git a/projects/llm_framework/main_llm/src/runner/LLM.hpp b/projects/llm_framework/main_llm/src/runner/LLM.hpp
index 7d700a53..5639fb89 100644
--- a/projects/llm_framework/main_llm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_llm/src/runner/LLM.hpp
@@ -15,48 +15,57 @@
 #include <ax_sys_api.h>
 
 #include <arm_neon.h>
+#define ALIGN_DOWN(x, a) ((x) & ~((a) - 1))
 
 // typedef void (*LLMRuningCallback)(int *p_token, int n_token, const char *p_str, float token_per_sec, void *reserve);
 typedef std::function<void(int *, int, const char *, float, void *)> LLMRuningCallback;
 
 struct LLMAttrType {
+    std::string system_prompt;
+
     std::string template_filename_axmodel = "tinyllama-int8/tinyllama_l%d.axmodel";
+    std::string post_config_path          = "post_config.json";
     int axmodel_num                       = 22;
 
-    // std::string template_prefill_filename_axmodel = "minicpmv/prefill_axmodel/minicpm_p96_l%d.axmodel";
-    // int prefill_axmodel_num = 40;
-    int prefill_token_num = 96;  // auto calc
+    int prefill_token_num     = 96;
+    int prefill_max_token_num = 512;
 
     std::string filename_post_axmodel = "tinyllama-int8/tinyllama_post.axmodel";
 
-    bool b_use_topk = false;
-
-    // std::string filename_vpm_resampler_axmodedl = "minicpmv/vpm_resampler_version0_fp16.axmodel";
-    // int vpm_width = 280;
-    // int vpm_height = 280;
-
     TokenizerType tokenizer_type         = TKT_LLaMa;
     std::string filename_tokenizer_model = "tokenizer.model";
-    bool b_bos = true, b_eos = false;
+    std::string url_tokenizer_model;
+    bool b_bos                        = true;
+    bool b_eos                        = false;
     std::string filename_tokens_embed = "tinyllama.model.embed_tokens.weight.bfloat16.bin";
     int tokens_embed_num              = 32000;
     int tokens_embed_size             = 2048;
 
-    int max_token_len = 127;  // auto calc
+    int max_token_len = 127;
+    int kv_cache_num  = 1024;
+    int kv_cache_size = 256;
 
-    int kv_cache_num  = 1024;  // auto calc
-    int kv_cache_size = 256;   // auto calc
+    int precompute_len = 0;
+    std::vector<int> prefill_max_kv_cache_num_grp;
+    int prefill_grpid = -1;
 
-    float temperature                 = 0.7f;
-    float top_p                       = 0.9f;
-    bool b_use_mmap_load_embed        = false;
-    bool b_dynamic_load_axmodel_layer = false;
+    bool enable_temperature = false;
+    float temperature       = 0.7f;
+
+    bool enable_top_p_sampling = false;
+    float top_p                = 0.7f;
 
-    bool b_use_mmap_load_layer = true;
+    bool enable_top_k_sampling = false;
+    int top_k                  = 50;
 
-    std::string post_config_path = "post_config.json";
+    bool enable_repetition_penalty = false;
+    float repetition_penalty       = 1.2f;
+    int penalty_window             = 50;
+
+    bool b_use_mmap_load_embed        = false;
+    bool b_dynamic_load_axmodel_layer = false;
+    bool b_use_mmap_load_layer        = true;
 
-    // bool b_live_print = true;
     LLMRuningCallback runing_callback = nullptr;
     void *reserve                     = nullptr;
 };
@@ -97,28 +106,7 @@ class LLM {
             logits[i]         = *reinterpret_cast<float *>(&proc);
         }
 
-        // postprocess.set_temperature(true, 0.9f);
-        // // postprocess.set_repetition_penalty(true, 1.1f);
-        // postprocess.set_top_k_sampling(true, 10);
-        // // postprocess.set_top_p_sampling(true, 0.9f);
-
         return postprocess.apply(logits, history);
-
-        // float max_val = -MAXFLOAT;
-        // int max_index = 0;
-        // for (int i = 0; i < n; i++)
-        // {
-        //     unsigned int proc = p[i] << 16;
-        //     float tmp = *reinterpret_cast<float *>(&proc);
-        //     if (tmp > max_val)
-        //     {
-        //         max_val = tmp;
-        //         max_index = i;
-        //     }
-        // }
-        // if (val)
-        //     *val = max_val;
-        // return max_index;
     }
 
 public:
@@ -133,17 +121,6 @@ class LLM {
             return false;
         }
         update_cqdm(&cqdm, 0, "count", "tokenizer init ok");
-        // test code
-        // {
-        //     std::vector<int> output;
-        //     tokenizer.Encode("Today is National", output);
-        //     // print output
-        //     for (size_t i = 0; i < output.size(); i++)
-        //     {
-        //         printf("%d ", output[i]);
-        //     }
-        //     printf("\n");
-        // }
 
         if (!embed_selector.Init(attr.filename_tokens_embed, attr.tokens_embed_num, attr.tokens_embed_size,
                                  attr.b_use_mmap_load_embed)) {
@@ -152,20 +129,7 @@ class LLM {
             return false;
         }
         update_cqdm(&cqdm, 1, "count", "embed_selector init ok");
-        // test code
-        // {
-        //     std::vector<unsigned short> embed = embed_selector.getByIndex(123);
-        //     printf("embed size: %d\n", embed.size());
-        //     for (int i = 0; i < embed.size(); i++)
-        //     {
-        //         bfloat16 bf16 = bfloat16(embed[i]);
-        //         float val = bf16;
-        //         printf("%d %0.22f\n", embed[i], val);
-        //     }
-        // }
-
         llama_layers.resize(attr.axmodel_num);
-        // prefill_layers.resize(attr.prefill_axmodel_num);
 
         char axmodel_path[1024];
         for (int i = 0; i < attr.axmodel_num; i++) {
@@ -227,8 +191,6 @@ class LLM {
             int max_token_len   = llama_layers[0].layer.get_input("mask").nSize / sizeof(unsigned short) - 1;
             _attr.max_token_len = max_token_len > _attr.max_token_len ? _attr.max_token_len : max_token_len;
             ALOGI("max_token_len : %d", _attr.max_token_len);
-            // auto &input_k_cache = llama_layers[0].layer.get_input("K_cache");
-            // auto &output_k_cache_out = llama_layers[0].layer.get_output("K_cache_out");
             _attr.kv_cache_size = llama_layers[0].layer.get_output("K_cache_out").nSize / sizeof(unsigned short);
             _attr.kv_cache_num =
                 llama_layers[0].layer.get_input("K_cache").nSize / _attr.kv_cache_size / sizeof(unsigned short);
@@ -245,11 +207,29 @@ class LLM {
             auto &layer = llama_layers[0];
             layer.layer.deinit();
         }
+        nlohmann::json dynamic_config;
+
+        dynamic_config["enable_temperature"] = _attr.enable_temperature;
+        dynamic_config["temperature"]        = _attr.temperature;
+
+        dynamic_config["enable_repetition_penalty"] = _attr.enable_repetition_penalty;
+        dynamic_config["repetition_penalty"]        = _attr.repetition_penalty;
+        dynamic_config["penalty_window"]            = _attr.penalty_window;
+
+        dynamic_config["enable_top_p_sampling"] = _attr.enable_top_p_sampling;
+        dynamic_config["top_p"]                 = _attr.top_p;
+
+        dynamic_config["enable_top_k_sampling"] = _attr.enable_top_k_sampling;
+        dynamic_config["top_k"]                 = _attr.top_k;
 
         if (!postprocess.load_config(attr.post_config_path)) {
             ALOGW("load postprocess config(%s) failed", attr.post_config_path.c_str());
         }
 
+        if (!postprocess.load_config(dynamic_config)) {
+            ALOGW("load postprocess config(%s) failed", dynamic_config.dump(4).c_str());
+        }
+
         // Reset();
         ALOGI("LLM init ok");
         return true;
@@ -269,12 +249,6 @@ class LLM {
         embed_selector.Deinit();
     }
 
-    // void Reset()
-    // {
-    //     k_caches.resize(_attr.axmodel_num, std::vector<unsigned short>(_attr.kv_cache_num * _attr.kv_cache_size, 0));
-    //     v_caches.resize(_attr.axmodel_num, std::vector<unsigned short>(_attr.kv_cache_num * _attr.kv_cache_size, 0));
-    // }
-
     void Stop()
     {
         b_stop = true;
@@ -282,9 +256,11 @@ class LLM {
 
     int Encode(std::vector<unsigned short> &out_embed, std::string prompt = "What is in the image?")
     {
-        std::vector<int> input_ids = tokenizer->Encode(prompt, true);
+        ImageInfo img_info;
+        img_info.img_prompt        = false;
+        std::vector<int> input_ids = tokenizer->Encode(prompt, img_info);
         if (input_ids.size() > _attr.prefill_token_num) {
-            ALOGE("input_ids(%d) > prefill_token_num(%d)", input_ids.size(), _attr.prefill_token_num);
+            ALOGE("input_ids(%ld) > prefill_token_num(%d)", input_ids.size(), _attr.prefill_token_num);
             return -1;
         }
         out_embed.resize(input_ids.size() * _attr.tokens_embed_size);
@@ -306,7 +282,7 @@ class LLM {
         return Run(test_embed);
     }
 
-    std::string Run(std::vector<unsigned short> test_embed)
+    std::string Run(std::vector<unsigned short> &test_embed)
     {
         b_stop = false;
         std::string final_out;
@@ -391,20 +367,8 @@ class LLM {
             if (_attr.b_dynamic_load_axmodel_layer) {
                 layer.layer.deinit();
             }
-            // ALOGI("%f %f %f %f %f", bfloat16(embed[0]).fp32(), bfloat16(embed[1]).fp32(), bfloat16(embed[2]).fp32(),
-            // bfloat16(embed[3]).fp32(), bfloat16(embed[4]).fp32());
         }
 
-        // ALOGI("prefill time cost: %.2f s", t_cost.cost() / 1000);
-
-        // print token_ids
-        // printf("%s\n", input_str.c_str());
-        // for (size_t i = 0; i < token_ids.size(); i++)
-        // {
-        //     printf("%d ", token_ids[i]);
-        // }
-        // printf("\n");
-
         int next_token = -1;
         t_cqdm cqdm    = create_cqdm(_attr.max_token_len, 32);
         std::vector<unsigned short> embed(_attr.tokens_embed_size, 0);
@@ -417,19 +381,15 @@ class LLM {
             auto &input = llama_post.get_input("input");
             memcpy(input.pVirAddr, embed.data(), embed.size() * sizeof(unsigned short));
             llama_post.inference();
+
             int max_index;
-            if (_attr.b_use_topk) {
-                AX_SYS_MinvalidateCache(llama_post.get_output("indices").phyAddr,
-                                        llama_post.get_output("indices").pVirAddr,
-                                        llama_post.get_output("indices").nSize);
-                max_index = *(int *)llama_post.get_output("indices").pVirAddr;
-            } else {
-                auto &output_post = llama_post.get_output("output");
-                AX_SYS_MinvalidateCache(output_post.phyAddr, output_post.pVirAddr, output_post.nSize);
-                unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
-                float max_val            = -MAXFLOAT;
-                max_index = post_process(postprocess, post_out, _attr.tokens_embed_num, token_ids, &max_val);
-            }
+
+            auto &output_post = llama_post.get_output("output");
+            AX_SYS_MinvalidateCache(output_post.phyAddr, output_post.pVirAddr, output_post.nSize);
+            unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
+            float max_val            = -MAXFLOAT;
+            max_index                = post_process(postprocess, post_out, _attr.tokens_embed_num, token_ids, &max_val);
+
             next_token = max_index;
 
             token_ids.push_back(max_index);
@@ -513,18 +473,13 @@ class LLM {
                 memcpy(input.pVirAddr, embed.data(), embed.size() * sizeof(unsigned short));
                 llama_post.inference();
                 int max_index;
-                if (_attr.b_use_topk) {
-                    AX_SYS_MinvalidateCache(llama_post.get_output("indices").phyAddr,
-                                            llama_post.get_output("indices").pVirAddr,
-                                            llama_post.get_output("indices").nSize);
-                    max_index = *(int *)llama_post.get_output("indices").pVirAddr;
-                } else {
-                    auto &output_post = llama_post.get_output("output");
-                    AX_SYS_MinvalidateCache(output_post.phyAddr, output_post.pVirAddr, output_post.nSize);
-                    unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
-                    float max_val            = -MAXFLOAT;
-                    max_index = post_process(postprocess, post_out, _attr.tokens_embed_num, token_ids, &max_val);
-                }
+
+                auto &output_post = llama_post.get_output("output");
+                AX_SYS_MinvalidateCache(output_post.phyAddr, output_post.pVirAddr, output_post.nSize);
+                unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
+                float max_val            = -MAXFLOAT;
+                max_index = post_process(postprocess, post_out, _attr.tokens_embed_num, token_ids, &max_val);
+
                 next_token = max_index;
 
                 if (tokenizer->isEnd(max_index)) {
@@ -574,3 +529,890 @@ class LLM {
         return final_out;
     }
 };
+
+class LLM_CTX {
+private:
+    std::shared_ptr<BaseTokenizer> tokenizer;
+    LLaMaEmbedSelector embed_selector;
+
+    LLMAttrType _attr;
+
+    struct LLMLayer {
+        ax_runner_ax650 layer;
+        std::string filename;
+        MMap layer_buffer;
+        std::vector<char> layer_buffer_vec;
+    };
+
+    std::vector<LLMLayer> llama_layers;
+    ax_runner_ax650 llama_post;
+
+    //
+    int decode_grpid = 0;
+
+    // ax_runner_ax650 vpm_resampler;
+
+    // std::vector<std::vector<unsigned short>> k_caches, v_caches;
+
+    bool b_stop = false;
+
+    LLMPostprocess postprocess;
+    static int post_process(LLMPostprocess &postprocess, unsigned short *p, int n, std::vector<int> &history,
+                            float *val = 0)
+    {
+        std::vector<float> logits(n);
+        for (int i = 0; i < n; i++) {
+            unsigned int proc = p[i] << 16;
+            logits[i]         = *reinterpret_cast<float *>(&proc);
+        }
+
+        return postprocess.apply(logits, history);
+    }
+
+public:
+    bool Init(LLMAttrType attr)
+    {
+        ALOGI("LLM init start");
+        t_cqdm cqdm = create_cqdm(attr.axmodel_num + 3, 32);
+        this->_attr = attr;
+        tokenizer   = CreateTokenizer(attr.tokenizer_type);
+        if (!tokenizer->Init(attr.url_tokenizer_model)) {
+            ALOGE("tokenizer.Init(%s) failed", attr.url_tokenizer_model.c_str());
+            return false;
+        }
+        std::vector<int> _token_ids;
+        tokenizer->Reset(attr.system_prompt, _token_ids);
+        update_cqdm(&cqdm, 0, "count", "tokenizer init ok");
+
+        if (!embed_selector.Init(attr.filename_tokens_embed, attr.tokens_embed_num, attr.tokens_embed_size,
+                                 attr.b_use_mmap_load_embed)) {
+            ALOGE("embed_selector.Init(%s, %d, %d) failed", attr.filename_tokens_embed.c_str(), attr.tokens_embed_num,
+                  attr.tokens_embed_size);
+            return false;
+        }
+        update_cqdm(&cqdm, 1, "count", "embed_selector init ok");
+
+        llama_layers.resize(attr.axmodel_num);
+        // prefill_layers.resize(attr.prefill_axmodel_num);
+
+        char axmodel_path[1024];
+        for (int i = 0; i < attr.axmodel_num; i++) {
+            sprintf(axmodel_path, attr.template_filename_axmodel.c_str(), i);
+            llama_layers[i].filename = axmodel_path;
+
+            int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), false);
+            if (ret != 0) {
+                ALOGE("init axmodel(%s) failed", llama_layers[i].filename.c_str());
+                return false;
+            }
+            int remain_cmm = get_remaining_cmm_size();
+            sprintf(axmodel_path, "init %d axmodel ok,remain_cmm(%d MB)", i, remain_cmm);
+            update_cqdm(&cqdm, i + 2, "count", axmodel_path);
+        }
+
+        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), false);
+        if (ret != 0) {
+            ALOGE("init post axmodel(%s) failed", attr.filename_post_axmodel.c_str());
+            return false;
+        }
+        int remain_cmm = get_remaining_cmm_size();
+        sprintf(axmodel_path, "init post axmodel ok,remain_cmm(%d MB)", remain_cmm);
+        update_cqdm(&cqdm, attr.axmodel_num + 2, "count", axmodel_path);
+
+        {
+            _attr.max_token_len = llama_layers[0].layer.get_input("mask").nSize / sizeof(unsigned short) - 1;
+            ALOGI("max_token_len : %d", _attr.max_token_len);
+            // auto &input_k_cache = llama_layers[0].layer.get_input("K_cache");
+            // auto &output_k_cache_out = llama_layers[0].layer.get_output("K_cache_out");
+            _attr.kv_cache_size = llama_layers[0].layer.get_output("K_cache_out").nSize / sizeof(unsigned short);
+            _attr.kv_cache_num =
+                llama_layers[0].layer.get_input("K_cache").nSize / _attr.kv_cache_size / sizeof(unsigned short);
+            ALOGI("kv_cache_size : %d, kv_cache_num: %d", _attr.kv_cache_size, _attr.kv_cache_num);
+            if (_attr.max_token_len > _attr.kv_cache_num) {
+                ALOGE("max_token_len(%d) > kv_cache_num(%d)", _attr.max_token_len, _attr.kv_cache_num);
+                return false;
+            }
+
+            _attr.prefill_token_num = llama_layers[0].layer.get_input(1, "indices").vShape[1];
+            ALOGI("prefill_token_num : %d", _attr.prefill_token_num);
+            for (size_t i = 0; i < llama_layers[0].layer.get_num_input_groups() - 1; i++) {
+                int prefill_max_kv_cache_num = llama_layers[0].layer.get_input(i + 1, "K_cache").vShape[1];
+                ALOGI("grp: %d, prefill_max_token_num : %d", i + 1, prefill_max_kv_cache_num);
+                _attr.prefill_max_kv_cache_num_grp.push_back(prefill_max_kv_cache_num);
+            }
+            _attr.prefill_max_token_num =
+                _attr.prefill_max_kv_cache_num_grp[_attr.prefill_max_kv_cache_num_grp.size() - 1];
+            ALOGI("prefill_max_token_num : %d", _attr.prefill_max_token_num);
+        }
+
+        nlohmann::json dynamic_config;
+
+        dynamic_config["enable_temperature"] = _attr.enable_temperature;
+        dynamic_config["temperature"]        = _attr.temperature;
+
+        dynamic_config["enable_repetition_penalty"] = _attr.enable_repetition_penalty;
+        dynamic_config["repetition_penalty"]        = _attr.repetition_penalty;
+        dynamic_config["penalty_window"]            = _attr.penalty_window;
+
+        dynamic_config["enable_top_p_sampling"] = _attr.enable_top_p_sampling;
+        dynamic_config["top_p"]                 = _attr.top_p;
+
+        dynamic_config["enable_top_k_sampling"] = _attr.enable_top_k_sampling;
+        dynamic_config["top_k"]                 = _attr.top_k;
+
+        if (!postprocess.load_config(attr.post_config_path)) {
+            ALOGW("load postprocess config(%s) failed", attr.post_config_path.c_str());
+        }
+
+        if (!postprocess.load_config(dynamic_config)) {
+            ALOGW("load postprocess config(%s) failed", dynamic_config.dump(4).c_str());
+        }
+
+        // Reset();
+        ALOGI("LLM init ok");
+        return true;
+    }
+
+    LLMAttrType *getAttr()
+    {
+        return &_attr;
+    }
+
+    LLMPostprocess *getPostprocess()
+    {
+        return &postprocess;
+    }
+
+    void Deinit()
+    {
+        for (int i = 0; i < _attr.axmodel_num; i++) {
+            llama_layers[i].layer.release();
+        }
+        llama_post.release();
+        embed_selector.Deinit();
+    }
+
+    void Stop()
+    {
+        b_stop = true;
+    }
+
+    int SetSystemPrompt(std::string system_prompt, std::vector<int> &_token_ids)
+    {
+        tokenizer->Reset(system_prompt, _token_ids);
+        _attr.system_prompt         = system_prompt;
+        _attr.prefill_max_token_num = _attr.prefill_max_kv_cache_num_grp[_attr.prefill_max_kv_cache_num_grp.size() - 1];
+        return 0;
+    }
+
+    int GenerateKVCachePrefill(std::vector<int> &_token_ids, std::vector<std::vector<unsigned short>> &k_caches,
+                               std::vector<std::vector<unsigned short>> &v_caches, int &precompute_len)
+    {
+        bfloat16 bf16       = -65536.f;
+        int input_embed_num = _token_ids.size();
+        precompute_len      = _token_ids.size();
+
+        k_caches.resize(_attr.axmodel_num);
+        v_caches.resize(_attr.axmodel_num);
+        int prefill_split_num = ceil((double)input_embed_num / _attr.prefill_token_num);
+
+        int prefill_grpid = _attr.prefill_max_kv_cache_num_grp.size();
+
+        for (size_t i = 0; i < _attr.prefill_max_kv_cache_num_grp.size(); i++) {
+            if (input_embed_num <= _attr.prefill_max_kv_cache_num_grp[i]) {
+                prefill_grpid = i + 1;
+                break;
+            }
+        }
+        ALOGI("input token num : %d, prefill_split_num : %d prefill_grpid : %d", input_embed_num, prefill_split_num,
+              prefill_grpid);
+
+        // clear kv cache
+        for (size_t i = 0; i < _attr.axmodel_num; i++) {
+            memset((void *)llama_layers[i].layer.get_input(prefill_grpid, "K_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(prefill_grpid, "K_cache").nSize);
+            memset((void *)llama_layers[i].layer.get_input(prefill_grpid, "V_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(prefill_grpid, "V_cache").nSize);
+        }
+
+        if (input_embed_num == 0) {
+            for (size_t i = 0; i < _attr.axmodel_num; i++) {
+                k_caches[i].resize(precompute_len * _attr.kv_cache_size);
+                v_caches[i].resize(precompute_len * _attr.kv_cache_size);
+            }
+            ALOGI("input token num is 0, skip");
+            return 0;
+        }
+
+        int kv_cache_num = _attr.prefill_max_kv_cache_num_grp[prefill_grpid - 1];
+
+        std::vector<unsigned short> test_embed;
+        test_embed.resize(_token_ids.size() * _attr.tokens_embed_size);
+
+        for (size_t i = 0; i < _token_ids.size(); i++) {
+            embed_selector.getByIndex(_token_ids[i], test_embed.data() + i * _attr.tokens_embed_size);
+        }
+
+        for (size_t p = 0; p < prefill_split_num; p++) {
+            std::vector<unsigned short> mask_tmp;
+            mask_tmp.resize(1 * _attr.prefill_token_num * (kv_cache_num + _attr.prefill_token_num), bf16.data);
+            int input_num_token = _attr.prefill_token_num;
+            if (p == prefill_split_num - 1) {
+                input_num_token = input_embed_num - p * _attr.prefill_token_num;
+            }
+
+            ALOGI("input_num_token:%d", input_num_token);
+            for (size_t i = 0; i < _attr.prefill_token_num; i++) {
+                if (i < input_num_token) {
+                    int mask_current_start = kv_cache_num;
+                    auto mask_ptr          = mask_tmp.data() + i * (kv_cache_num + _attr.prefill_token_num);
+
+                    for (int j = 0; j < p * _attr.prefill_token_num; j++) {
+                        mask_ptr[j] = 0;
+                    }
+
+                    for (int j = mask_current_start; j < mask_current_start + i + 1; j++) {
+                        mask_ptr[j] = 0;
+                    }
+                }
+            }
+
+            std::vector<unsigned short> embed_tmp(_attr.prefill_token_num * _attr.tokens_embed_size, 0);
+            if (p == (prefill_split_num - 1)) {
+                memcpy(
+                    embed_tmp.data(), test_embed.data() + p * _attr.prefill_token_num * _attr.tokens_embed_size,
+                    (input_embed_num - p * _attr.prefill_token_num) * _attr.tokens_embed_size * sizeof(unsigned short));
+            } else {
+                memcpy(embed_tmp.data(), test_embed.data() + p * _attr.prefill_token_num * _attr.tokens_embed_size,
+                       _attr.prefill_token_num * _attr.tokens_embed_size * sizeof(unsigned short));
+            }
+
+            for (unsigned int m = 0; m < _attr.axmodel_num; m++) {
+                auto &layer = llama_layers[m];
+                // set indices
+                auto &input_indices             = layer.layer.get_input(prefill_grpid, "indices");
+                unsigned int *input_indices_ptr = (unsigned int *)input_indices.pVirAddr;
+                memset(input_indices_ptr, 0, input_indices.nSize);
+                int idx = 0;
+                for (unsigned int i = p * _attr.prefill_token_num; i < (p + 1) * _attr.prefill_token_num; i++) {
+                    input_indices_ptr[idx] = i;
+                    idx++;
+                }
+
+                // set mask
+                auto &input_mask = layer.layer.get_input(prefill_grpid, "mask");
+                memcpy((void *)input_mask.pVirAddr, (void *)mask_tmp.data(), mask_tmp.size() * sizeof(unsigned short));
+
+                auto &input_input = layer.layer.get_input(prefill_grpid, "input");
+                memcpy((void *)input_input.pVirAddr, embed_tmp.data(), embed_tmp.size() * sizeof(unsigned short));
+
+                layer.layer.inference(prefill_grpid);
+
+                auto &input_decoder_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
+                auto &input_decoder_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
+
+                auto &input_prefill_k_cache = layer.layer.get_input(prefill_grpid, "K_cache");
+                auto &input_prefill_v_cache = layer.layer.get_input(prefill_grpid, "V_cache");
+
+                auto &output_k_cache = layer.layer.get_output(prefill_grpid, "K_cache_out");
+                auto &output_v_cache = layer.layer.get_output(prefill_grpid, "V_cache_out");
+
+                int kv_offset = (p * _attr.prefill_token_num) * _attr.kv_cache_size;
+
+                memcpy((unsigned short *)input_decoder_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.prefill_token_num * _attr.kv_cache_size);
+
+                memcpy((unsigned short *)input_decoder_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.prefill_token_num * _attr.kv_cache_size);
+
+                memcpy((unsigned short *)input_prefill_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.prefill_token_num * _attr.kv_cache_size);
+
+                memcpy((unsigned short *)input_prefill_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.prefill_token_num * _attr.kv_cache_size);
+
+                auto &output = layer.layer.get_output(prefill_grpid, "output");
+                memcpy(embed_tmp.data(), (void *)output.pVirAddr, embed_tmp.size() * sizeof(unsigned short));
+            }
+        }
+
+        for (size_t i = 0; i < _attr.axmodel_num; i++) {
+            auto &layer = llama_layers[i];
+            k_caches[i].resize(precompute_len * _attr.kv_cache_size);
+            v_caches[i].resize(precompute_len * _attr.kv_cache_size);
+            auto &input_k_cache = layer.layer.get_input(prefill_grpid, "K_cache");
+            auto &input_v_cache = layer.layer.get_input(prefill_grpid, "V_cache");
+            memcpy((void *)k_caches[i].data(), (void *)input_k_cache.pVirAddr,
+                   precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+            memcpy((void *)v_caches[i].data(), (void *)input_v_cache.pVirAddr,
+                   precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+        }
+
+        return 0;
+    }
+
+    int GenerateKVCache(std::vector<int> &_token_ids)
+    {
+        // clear kv cache
+        for (size_t i = 0; i < _attr.axmodel_num; i++) {
+            memset((void *)llama_layers[i].layer.get_input(decode_grpid, "K_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(decode_grpid, "K_cache").nSize);
+            memset((void *)llama_layers[i].layer.get_input(decode_grpid, "V_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(decode_grpid, "V_cache").nSize);
+        }
+
+        bfloat16 bf16 = -65536.f;
+        std::vector<unsigned short> mask(_attr.kv_cache_num + 1, bf16.data);
+        mask[_attr.kv_cache_num] = 0;
+        std::vector<unsigned short> embed;
+
+        int next_token = _token_ids[0];
+
+        t_cqdm cqdm = create_cqdm(_token_ids.size(), 32);
+
+        for (unsigned int indices = 0; indices < _token_ids.size(); indices++) {
+            // ALOGI("out %d %d", indices, next_token);
+            embed_selector.getByIndex(next_token, embed);
+
+            for (int m = 0; m < _attr.axmodel_num; m++) {
+                if (b_stop) {
+                    break;
+                }
+
+                auto &layer = llama_layers[m];
+
+                auto &input_k_cache               = layer.layer.get_input(decode_grpid, "K_cache");
+                unsigned short *input_k_cache_ptr = (unsigned short *)input_k_cache.pVirAddr;
+                auto &input_v_cache               = layer.layer.get_input(decode_grpid, "V_cache");
+                unsigned short *input_v_cache_ptr = (unsigned short *)input_v_cache.pVirAddr;
+
+                auto &input_indices = layer.layer.get_input(decode_grpid, "indices");
+                memcpy(input_indices.pVirAddr, &indices, sizeof(indices));
+
+                auto &input_mask = layer.layer.get_input(decode_grpid, "mask");
+                memcpy(input_mask.pVirAddr, mask.data(), mask.size() * sizeof(unsigned short));
+
+                auto &input_input = layer.layer.get_input(decode_grpid, "input");
+                memcpy(input_input.pVirAddr, embed.data(), embed.size() * sizeof(unsigned short));
+
+                layer.layer.inference(decode_grpid);
+
+                auto &output_k_cache = layer.layer.get_output(decode_grpid, "K_cache_out");
+                memcpy(input_k_cache_ptr + indices * _attr.kv_cache_size, output_k_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.kv_cache_size);
+
+                auto &output_v_cache = layer.layer.get_output(decode_grpid, "V_cache_out");
+                memcpy(input_v_cache_ptr + indices * _attr.kv_cache_size, output_v_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.kv_cache_size);
+
+                auto &output = layer.layer.get_output(decode_grpid, "output");
+                memcpy(embed.data(), output.pVirAddr, embed.size() * sizeof(unsigned short));
+            }
+            mask[indices] = 0;
+            next_token    = _token_ids[indices + 1];
+            update_cqdm(&cqdm, indices, "token", "");
+            // ALOGI("");
+        }
+        return 0;
+    }
+
+    int GetKVCache(std::vector<std::vector<unsigned short>> &k_caches,
+                   std::vector<std::vector<unsigned short>> &v_caches, int &precompute_len)
+    {
+        bfloat16 bf16 = -65536.f;
+        std::vector<unsigned short> mask(_attr.kv_cache_num + 1, bf16.data);
+        auto &input_mask = llama_layers[0].layer.get_input(decode_grpid, "mask");
+        memcpy(mask.data(), (void *)input_mask.pVirAddr, input_mask.nSize);
+        for (size_t i = 0; i < mask.size(); i++) {
+            if (mask[i] == bf16.data) {
+                precompute_len = i + 1;
+                break;
+            }
+        }
+        ALOGI("precompute_len:%d, remaining:%d", precompute_len,
+              _attr.prefill_max_kv_cache_num_grp[_attr.prefill_max_kv_cache_num_grp.size() - 1] - precompute_len);
+        k_caches.resize(_attr.axmodel_num);
+        v_caches.resize(_attr.axmodel_num);
+        for (size_t i = 0; i < _attr.axmodel_num; i++) {
+            auto &layer = llama_layers[i];
+            k_caches[i].resize(precompute_len * _attr.kv_cache_size);
+            v_caches[i].resize(precompute_len * _attr.kv_cache_size);
+            auto &input_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
+            auto &input_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
+            memcpy((void *)k_caches[i].data(), (void *)input_k_cache.pVirAddr,
+                   precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+            memcpy((void *)v_caches[i].data(), (void *)input_v_cache.pVirAddr,
+                   precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+        }
+
+        _attr.prefill_max_token_num = _attr.prefill_max_kv_cache_num_grp[_attr.prefill_max_kv_cache_num_grp.size() - 1];
+
+        return 0;
+    }
+
+    int SetKVCache(std::vector<std::vector<unsigned short>> &k_caches,
+                   std::vector<std::vector<unsigned short>> &v_caches, int precompute_len, int input_num_token)
+    {
+        _attr.precompute_len = precompute_len;
+        for (size_t i = 0; i < _attr.prefill_max_kv_cache_num_grp.size(); i++) {
+            if (_attr.precompute_len + input_num_token <= _attr.prefill_max_kv_cache_num_grp[i]) {
+                _attr.prefill_grpid = i + 1;
+                break;
+            }
+        }
+        int kv_cache_num = _attr.prefill_max_kv_cache_num_grp[_attr.prefill_grpid - 1];
+        ALOGI("prefill_grpid:%d kv_cache_num:%d precompute_len:%d input_num_token:%d", _attr.prefill_grpid,
+              kv_cache_num, precompute_len, input_num_token);
+
+        _attr.prefill_max_token_num =
+            ALIGN_DOWN(_attr.prefill_max_token_num - _attr.precompute_len, _attr.prefill_token_num);
+        ALOGI("current prefill_max_token_num:%d", _attr.prefill_max_token_num);
+
+        if (precompute_len == 0) {
+            ALOGI("first run");
+            return 0;
+        }
+
+        if (precompute_len + input_num_token > kv_cache_num) {
+            ALOGE("precompute_len(%d) + input_num_token(%d) > _attr.prefill_max_kv_cache_num_grp[%d]", precompute_len,
+                  input_num_token, _attr.prefill_grpid - 1);
+            return -1;
+        }
+
+        if (input_num_token > _attr.prefill_max_token_num) {
+            ALOGE("input_num_token(%d) > _attr.prefill_max_token_num(%d)", input_num_token,
+                  _attr.prefill_max_token_num);
+            return -1;
+        }
+
+        if (k_caches.size() != v_caches.size()) {
+            ALOGE("k_caches.size(%d) != v_caches.size(%d)", k_caches.size(), v_caches.size());
+            return -1;
+        }
+
+        if (k_caches.size() != _attr.axmodel_num) {
+            ALOGE("k_caches.size(%d) != _attr.axmodel_num(%d)", k_caches.size(), _attr.axmodel_num);
+            return -1;
+        }
+
+        // clear kv cache
+        for (size_t i = 0; i < _attr.axmodel_num; i++) {
+            memset((void *)llama_layers[i].layer.get_input(_attr.prefill_grpid, "K_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(_attr.prefill_grpid, "K_cache").nSize);
+            memset((void *)llama_layers[i].layer.get_input(_attr.prefill_grpid, "V_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(_attr.prefill_grpid, "V_cache").nSize);
+
+            memset((void *)llama_layers[i].layer.get_input(decode_grpid, "K_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(decode_grpid, "K_cache").nSize);
+            memset((void *)llama_layers[i].layer.get_input(decode_grpid, "V_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(decode_grpid, "V_cache").nSize);
+        }
+
+        // int prefill_grpid = llama_layers[0].layer.get_num_input_groups() - 1;
+
+        for (unsigned int m = 0; m < _attr.axmodel_num; m++) {
+            auto &layer = llama_layers[m];
+
+            auto &k_cache = k_caches[m];
+            auto &v_cache = v_caches[m];
+
+            if (k_cache.size() != _attr.precompute_len * _attr.kv_cache_size) {
+                ALOGE("k_cache.size(%d) != precompute_len(%d) * _attr.kv_cache_size(%d)", k_cache.size(),
+                      _attr.precompute_len, _attr.kv_cache_size);
+                return -1;
+            }
+            if (v_cache.size() < _attr.precompute_len * _attr.kv_cache_size) {
+                ALOGE("v_cache.size(%d) < precompute_len(%d) * _attr.kv_cache_size(%d)", v_cache.size(),
+                      _attr.precompute_len, _attr.kv_cache_size);
+                return -1;
+            }
+
+            // set kv cache inputs
+            {
+                auto &input_k_cache               = layer.layer.get_input(_attr.prefill_grpid, "K_cache");
+                unsigned short *input_k_cache_ptr = (unsigned short *)input_k_cache.pVirAddr;
+                auto &input_v_cache               = layer.layer.get_input(_attr.prefill_grpid, "V_cache");
+                unsigned short *input_v_cache_ptr = (unsigned short *)input_v_cache.pVirAddr;
+
+                memcpy(input_k_cache_ptr, k_cache.data(),
+                       _attr.precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+                memcpy(input_v_cache_ptr, v_cache.data(),
+                       _attr.precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+            }
+
+            {
+                auto &input_k_cache               = layer.layer.get_input(decode_grpid, "K_cache");
+                unsigned short *input_k_cache_ptr = (unsigned short *)input_k_cache.pVirAddr;
+                auto &input_v_cache               = layer.layer.get_input(decode_grpid, "V_cache");
+                unsigned short *input_v_cache_ptr = (unsigned short *)input_v_cache.pVirAddr;
+
+                memcpy(input_k_cache_ptr, k_cache.data(),
+                       _attr.precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+                memcpy(input_v_cache_ptr, v_cache.data(),
+                       _attr.precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+            }
+        }
+
+        return 0;
+    }
+
+    bool save_kvcache(std::string target_path, std::string system_prompt, int precompute_len,
+                      std::vector<std::vector<unsigned short>> &k_caches,
+                      std::vector<std::vector<unsigned short>> &v_caches)
+    {
+        for (size_t i = 0; i < k_caches.size(); i++) {
+            std::string k_cache_path = target_path + "/k_cache_" + std::to_string(i) + ".bin";
+            std::string v_cache_path = target_path + "/v_cache_" + std::to_string(i) + ".bin";
+            std::ofstream k_cache_file(k_cache_path);
+            std::ofstream v_cache_file(v_cache_path);
+            if (!k_cache_file.is_open() || !v_cache_file.is_open()) {
+                ALOGE("save kvcache failed");
+                return false;
+            }
+            k_cache_file.write((char *)k_caches[i].data(), k_caches[i].size() * sizeof(unsigned short));
+            v_cache_file.write((char *)v_caches[i].data(), v_caches[i].size() * sizeof(unsigned short));
+            k_cache_file.close();
+            v_cache_file.close();
+        }
+        nlohmann::json j;
+        j["system_prompt"]      = system_prompt;
+        j["precompute_len"]     = precompute_len;
+        std::string config_path = target_path + "/config.json";
+        std::ofstream config_file(config_path);
+        config_file << j.dump();
+        config_file.close();
+        return true;
+    }
+
+    bool load_kvcache(std::string target_path, int axmodel_num, std::vector<std::vector<unsigned short>> &k_caches,
+                      std::vector<std::vector<unsigned short>> &v_caches, std::string &system_prompt,
+                      int &precompute_len)
+    {
+        k_caches.resize(axmodel_num);
+        v_caches.resize(axmodel_num);
+        for (size_t i = 0; i < k_caches.size(); i++) {
+            std::string k_cache_path = target_path + "/k_cache_" + std::to_string(i) + ".bin";
+            std::string v_cache_path = target_path + "/v_cache_" + std::to_string(i) + ".bin";
+            if (file_exist(k_cache_path) && file_exist(v_cache_path)) {
+                std::vector<unsigned short> k_cache;
+                std::vector<unsigned short> v_cache;
+                std::ifstream k_cache_file(k_cache_path);
+                std::ifstream v_cache_file(v_cache_path);
+
+                k_cache_file.seekg(0, std::ios::end);
+                k_cache.resize(k_cache_file.tellg() / sizeof(unsigned short));
+                k_cache_file.seekg(0, std::ios::beg);
+
+                v_cache_file.seekg(0, std::ios::end);
+                v_cache.resize(v_cache_file.tellg() / sizeof(unsigned short));
+                v_cache_file.seekg(0, std::ios::beg);
+
+                k_cache_file.read((char *)k_cache.data(), k_cache.size() * sizeof(unsigned short));
+                v_cache_file.read((char *)v_cache.data(), v_cache.size() * sizeof(unsigned short));
+
+                k_cache_file.close();
+                v_cache_file.close();
+                k_caches[i] = k_cache;
+                v_caches[i] = v_cache;
+            } else {
+                ALOGE("k_cache %s or v_cache %s not exist", k_cache_path.c_str(), v_cache_path.c_str());
+                return false;
+            }
+        }
+
+        std::string config_path = target_path + "/config.json";
+        if (file_exist(config_path)) {
+            std::ifstream config_file(config_path);
+            nlohmann::json j;
+            config_file >> j;
+            system_prompt  = j["system_prompt"].get<std::string>();
+            precompute_len = j["precompute_len"].get<int>();
+            config_file.close();
+        } else {
+            ALOGE("config %s not exist", config_path.c_str());
+            return false;
+        }
+        return true;
+    }
+
+    int Encode(std::vector<unsigned short> &out_embed, std::string prompt, std::string last_reply,
+               std::vector<int> &tokens_ids, std::vector<int> &tokens_diff)
+    {
+        ImageInfo img_info;
+        img_info.img_prompt = false;
+        if (!tokenizer->Encode(prompt, last_reply, tokens_ids, tokens_diff, img_info)) {
+            ALOGE("encode failed");
+            return -1;
+        }
+
+        out_embed.resize(tokens_diff.size() * _attr.tokens_embed_size);
+
+        for (size_t i = 0; i < tokens_diff.size(); i++) {
+            embed_selector.getByIndex(tokens_diff[i], out_embed.data() + i * _attr.tokens_embed_size);
+        }
+
+        return 0;
+    }
+
+    std::string Run(std::vector<unsigned short> test_embed)
+    {
+        b_stop = false;
+        std::string final_out;
+
+        bfloat16 bf16 = -65536.f;
+        std::vector<unsigned short> mask(_attr.kv_cache_num + 1, bf16.data);
+        std::vector<unsigned short> embed(_attr.tokens_embed_size, 0);
+        int kv_cache_num = _attr.prefill_max_kv_cache_num_grp[_attr.prefill_grpid - 1];
+
+        std::vector<int> cached_token;
+        std::vector<int> token_ids;
+
+        int input_embed_num   = test_embed.size() / _attr.tokens_embed_size;
+        int prefill_split_num = ceil((double)input_embed_num / _attr.prefill_token_num);
+        ALOGI("input token num : %d, prefill_split_num : %d", input_embed_num, prefill_split_num);
+
+        mask[_attr.kv_cache_num] = 0;
+        for (size_t i = 0; i < _attr.precompute_len + input_embed_num; i++) {
+            mask[i] = 0;
+        }
+        timer t_cost;
+        timer ttft_timer;
+        ttft_timer.start();
+
+        for (size_t p = 0; p < prefill_split_num; p++) {
+            if (b_stop) {
+                break;
+            }
+
+            std::vector<unsigned short> mask_tmp;
+            mask_tmp.resize(1 * _attr.prefill_token_num * (kv_cache_num + _attr.prefill_token_num), bf16.data);
+            int input_num_token = _attr.prefill_token_num;
+            if (p == prefill_split_num - 1) {
+                input_num_token = input_embed_num - p * _attr.prefill_token_num;
+            }
+
+            ALOGI("input_num_token:%d", input_num_token);
+            for (size_t i = 0; i < _attr.prefill_token_num; i++) {
+                if (i < input_num_token) {
+                    int mask_current_start = kv_cache_num;
+                    auto mask_ptr          = mask_tmp.data() + i * (kv_cache_num + _attr.prefill_token_num);
+
+                    for (int j = 0; j < _attr.precompute_len + p * _attr.prefill_token_num; j++) {
+                        mask_ptr[j] = 0;
+                    }
+
+                    for (int j = mask_current_start; j < mask_current_start + i + 1; j++) {
+                        mask_ptr[j] = 0;
+                    }
+                }
+            }
+
+            std::vector<unsigned short> embed_tmp(_attr.prefill_token_num * _attr.tokens_embed_size, 0);
+            if (p == (prefill_split_num - 1)) {
+                memcpy(
+                    embed_tmp.data(), test_embed.data() + p * _attr.prefill_token_num * _attr.tokens_embed_size,
+                    (input_embed_num - p * _attr.prefill_token_num) * _attr.tokens_embed_size * sizeof(unsigned short));
+            } else {
+                memcpy(embed_tmp.data(), test_embed.data() + p * _attr.prefill_token_num * _attr.tokens_embed_size,
+                       _attr.prefill_token_num * _attr.tokens_embed_size * sizeof(unsigned short));
+            }
+
+            for (unsigned int m = 0; m < _attr.axmodel_num; m++) {
+                if (b_stop) {
+                    break;
+                }
+
+                auto &layer = llama_layers[m];
+
+                // set indices
+                auto &input_indices             = layer.layer.get_input(_attr.prefill_grpid, "indices");
+                unsigned int *input_indices_ptr = (unsigned int *)input_indices.pVirAddr;
+                memset(input_indices_ptr, 0, input_indices.nSize);
+                int idx = 0;
+                for (unsigned int i = _attr.precompute_len + p * _attr.prefill_token_num;
+                     i < _attr.precompute_len + (p + 1) * _attr.prefill_token_num; i++) {
+                    input_indices_ptr[idx] = i;
+                    idx++;
+                }
+
+                // set mask
+                auto &input_mask = layer.layer.get_input(_attr.prefill_grpid, "mask");
+                memcpy((void *)input_mask.pVirAddr, (void *)mask_tmp.data(), mask_tmp.size() * sizeof(unsigned short));
+
+                // set input
+                auto &input_input = layer.layer.get_input(_attr.prefill_grpid, "input");
+                memcpy((void *)input_input.pVirAddr, embed_tmp.data(), embed_tmp.size() * sizeof(unsigned short));
+
+                layer.layer.inference(_attr.prefill_grpid);
+
+                auto &input_decoder_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
+                auto &input_decoder_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
+
+                auto &input_prefill_k_cache = layer.layer.get_input(_attr.prefill_grpid, "K_cache");
+                auto &input_prefill_v_cache = layer.layer.get_input(_attr.prefill_grpid, "V_cache");
+
+                auto &output_k_cache = layer.layer.get_output(_attr.prefill_grpid, "K_cache_out");
+                auto &output_v_cache = layer.layer.get_output(_attr.prefill_grpid, "V_cache_out");
+
+                int kv_offset = (_attr.precompute_len + p * _attr.prefill_token_num) * _attr.kv_cache_size;
+
+                memcpy((unsigned short *)input_decoder_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
+                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+
+                memcpy((unsigned short *)input_decoder_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
+                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+
+                memcpy((unsigned short *)input_prefill_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
+                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+
+                memcpy((unsigned short *)input_prefill_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
+                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+
+                auto &output = layer.layer.get_output(_attr.prefill_grpid, "output");
+                memcpy(embed_tmp.data(), (void *)output.pVirAddr, embed_tmp.size() * sizeof(unsigned short));
+            }
+            if (p == (prefill_split_num - 1)) {
+                memcpy(embed.data(),
+                       embed_tmp.data() + (input_embed_num - p * _attr.prefill_token_num - 1) * _attr.tokens_embed_size,
+                       _attr.tokens_embed_size * sizeof(unsigned short));
+            }
+        }
+
+        int next_token = -1;
+        t_cqdm cqdm    = create_cqdm(_attr.max_token_len, 32);
+
+        {
+            // post process
+            auto &input = llama_post.get_input("input");
+            memcpy(input.pVirAddr, embed.data(), embed.size() * sizeof(unsigned short));
+            llama_post.inference();
+            int max_index;
+
+            auto &output_post = llama_post.get_output("output");
+            // AX_SYS_MinvalidateCache(output_post.phyAddr, output_post.pVirAddr, output_post.nSize);
+            unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
+            float max_val            = -MAXFLOAT;
+            // max_index = FindMax(post_out, _attr.tokens_embed_num, &max_val);
+            max_index = post_process(postprocess, post_out, _attr.tokens_embed_num, token_ids, nullptr);
+
+            next_token = max_index;
+
+            token_ids.push_back(max_index);
+            cached_token.push_back(max_index);
+            ALOGI("ttft: %.2f ms", ttft_timer.cost());
+        }
+        t_cost.start();
+
+        bool b_hit_eos = false;
+        for (unsigned int indices = _attr.precompute_len + input_embed_num; indices < _attr.max_token_len; indices++) {
+            if (b_stop) {
+                break;
+            }
+
+            // ALOGI("out %d %d", indices, next_token);
+            embed_selector.getByIndex(next_token, embed);
+            // ALOGI("%f %f %f %f %f", bfloat16(embed[0]).fp32(), bfloat16(embed[1]).fp32(), bfloat16(embed[2]).fp32(),
+            // bfloat16(embed[3]).fp32(), bfloat16(embed[4]).fp32());
+
+            for (int m = 0; m < _attr.axmodel_num; m++) {
+                if (b_stop) {
+                    break;
+                }
+
+                auto &layer = llama_layers[m];
+
+                auto &input_k_cache               = layer.layer.get_input(decode_grpid, "K_cache");
+                unsigned short *input_k_cache_ptr = (unsigned short *)input_k_cache.pVirAddr;
+                // memcpy(input_k_cache.pVirAddr, k_caches[m].data(), sizeof(unsigned short) * k_caches[m].size());
+                auto &input_v_cache               = layer.layer.get_input(decode_grpid, "V_cache");
+                unsigned short *input_v_cache_ptr = (unsigned short *)input_v_cache.pVirAddr;
+                // memcpy(input_v_cache.pVirAddr, v_caches[m].data(), sizeof(unsigned short) * v_caches[m].size());
+
+                auto &input_indices = layer.layer.get_input(decode_grpid, "indices");
+                memcpy(input_indices.pVirAddr, &indices, sizeof(indices));
+
+                auto &input_mask = layer.layer.get_input(decode_grpid, "mask");
+                memcpy(input_mask.pVirAddr, mask.data(), mask.size() * sizeof(unsigned short));
+
+                auto &input_input = layer.layer.get_input(decode_grpid, "input");
+                memcpy(input_input.pVirAddr, embed.data(), embed.size() * sizeof(unsigned short));
+
+                layer.layer.inference(decode_grpid);
+
+                auto &output_k_cache = layer.layer.get_output(decode_grpid, "K_cache_out");
+                // AX_SYS_MinvalidateCache(output_k_cache.phyAddr, output_k_cache.pVirAddr, output_k_cache.nSize);
+                memcpy(input_k_cache_ptr + indices * _attr.kv_cache_size, output_k_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.kv_cache_size);
+
+                auto &output_v_cache = layer.layer.get_output(decode_grpid, "V_cache_out");
+                // AX_SYS_MinvalidateCache(output_v_cache.phyAddr, output_v_cache.pVirAddr, output_v_cache.nSize);
+                memcpy(input_v_cache_ptr + indices * _attr.kv_cache_size, output_v_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.kv_cache_size);
+
+                auto &output = layer.layer.get_output(decode_grpid, "output");
+                // AX_SYS_MinvalidateCache(output.phyAddr, output.pVirAddr, output.nSize);
+                memcpy(embed.data(), output.pVirAddr, embed.size() * sizeof(unsigned short));
+
+                // ALOGI("%f %f %f %f %f", bfloat16(embed[0]).fp32(), bfloat16(embed[1]).fp32(),
+                // bfloat16(embed[2]).fp32(), bfloat16(embed[3]).fp32(), bfloat16(embed[4]).fp32());
+            }
+            // ALOGI("");
+            mask[indices] = 0;
+            {
+                // post process
+                auto &input = llama_post.get_input("input");
+                memcpy(input.pVirAddr, embed.data(), embed.size() * sizeof(unsigned short));
+                llama_post.inference();
+                int max_index;
+
+                auto &output_post = llama_post.get_output("output");
+                // AX_SYS_MinvalidateCache(output_post.phyAddr, output_post.pVirAddr, output_post.nSize);
+                unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
+                float max_val            = -MAXFLOAT;
+                // max_index = FindMax(post_out, _attr.tokens_embed_num, &max_val);
+                max_index = post_process(postprocess, post_out, _attr.tokens_embed_num, token_ids, nullptr);
+
+                next_token = max_index;
+
+                if (tokenizer->isEnd(max_index)) {
+                    if (cached_token.size() && _attr.runing_callback) {
+                        float t_cost_ms     = t_cost.cost();
+                        float token_per_sec = token_ids.size() / (t_cost_ms / 1000);
+                        auto tmp_out        = tokenizer->Decode(cached_token);
+                        _attr.runing_callback(cached_token.data(), cached_token.size(), tmp_out.c_str(), token_per_sec,
+                                              _attr.reserve);
+                        cached_token.clear();
+                    }
+                    b_hit_eos = true;
+                    break;
+                }
+                token_ids.push_back(max_index);
+
+                if (_attr.runing_callback) {
+                    cached_token.push_back(max_index);
+                    if (cached_token.size() >= 3) {
+                        float t_cost_ms     = t_cost.cost();
+                        float token_per_sec = token_ids.size() / (t_cost_ms / 1000);
+                        auto tmp_out        = tokenizer->Decode(cached_token);
+                        _attr.runing_callback(cached_token.data(), cached_token.size(), tmp_out.c_str(), token_per_sec,
+                                              _attr.reserve);
+                        cached_token.clear();
+                    }
+                }
+            }
+
+            if (_attr.runing_callback == nullptr) update_cqdm(&cqdm, indices, "token", "");
+            if (b_hit_eos) {
+                break;
+            }
+        }
+        printf("\n\n");
+        fflush(stdout);
+        float t_cost_ms = t_cost.cost();
+        ALOGN("hit eos,avg %.2f token/s\n", token_ids.size() / (t_cost_ms / 1000));
+
+        final_out = tokenizer->Decode(token_ids);
+
+        return final_out;
+    }
+};
\ No newline at end of file
diff --git a/projects/llm_framework/main_llm/src/runner/LLMPostprocess.hpp b/projects/llm_framework/main_llm/src/runner/LLMPostprocess.hpp
index c98205e4..b7d50156 100644
--- a/projects/llm_framework/main_llm/src/runner/LLMPostprocess.hpp
+++ b/projects/llm_framework/main_llm/src/runner/LLMPostprocess.hpp
@@ -242,10 +242,11 @@ class LLMPostprocess
         this->temperature = temperature;
     }
 
-    void set_repetition_penalty(bool enable, float penalty)
+    void set_repetition_penalty(bool enable, float penalty, int penalty_window)
     {
         enable_repetition_penalty = enable;
         this->repetition_penalty = penalty;
+        this->penalty_window = penalty_window;
     }
 
     void set_diversity_penalty(bool enable, const std::vector<int> &common_phrases, float penalty)
@@ -295,6 +296,49 @@ class LLMPostprocess
         return true;
     }
 
+    bool load_config(const nlohmann::json& config)
+    {
+        if (config.is_null()) {
+            ALOGE("config is null or invalid");
+            return false;
+        }
+
+        ALOGI("load config: \n%s\n", config.dump(4).c_str());
+
+        if (config.contains("enable_temperature")) {
+            enable_temperature = config["enable_temperature"].get<bool>();
+        }
+        if (config.contains("temperature")) {
+            temperature = config["temperature"].get<float>();
+        }
+
+        if (config.contains("enable_repetition_penalty")) {
+            enable_repetition_penalty = config["enable_repetition_penalty"].get<bool>();
+        }
+        if (config.contains("repetition_penalty")) {
+            repetition_penalty = config["repetition_penalty"].get<float>();
+        }
+        if (config.contains("penalty_window")) {
+            penalty_window = config["penalty_window"].get<int>();
+        }
+
+        if (config.contains("enable_top_p_sampling")) {
+            enable_top_p_sampling = config["enable_top_p_sampling"].get<bool>();
+        }
+        if (config.contains("top_p")) {
+            top_p = config["top_p"].get<float>();
+        }
+
+        if (config.contains("enable_top_k_sampling")) {
+            enable_top_k_sampling = config["enable_top_k_sampling"].get<bool>();
+        }
+        if (config.contains("top_k")) {
+            top_k = config["top_k"].get<int>();
+        }
+
+        return true;
+    }
+
     int apply(std::vector<float> &logits, const std::vector<int> &history)
     {
         if (enable_temperature)
diff --git a/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.cpp b/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.cpp
index 31aa272d..f676efdd 100644
--- a/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.cpp
+++ b/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.cpp
@@ -8,6 +8,7 @@
 // #include "chatglm.h"
 
 #include "httplib.h"
+#include "http_utils.hpp"
 #include "json.hpp"
 
 #include "sample_log.h"
@@ -21,8 +22,7 @@
 #include <cstring>
 #include <csignal>
 
-class TokenizerLLaMa : public BaseTokenizer
-{
+class TokenizerLLaMa : public BaseTokenizer {
 protected:
     sentencepiece::SentencePieceProcessor sp;
     bool _b_bos, _b_eos;
@@ -33,8 +33,7 @@ class TokenizerLLaMa : public BaseTokenizer
     bool Init(std::string model_path, bool b_bos = true, bool b_eos = false) override
     {
         auto ret = sp.Load(model_path);
-        if (!ret.ok())
-        {
+        if (!ret.ok()) {
             ALOGE("%s", ret.error_message());
             return false;
         }
@@ -44,43 +43,37 @@ class TokenizerLLaMa : public BaseTokenizer
         return ret.ok();
     }
 
-    bool Encode(std::string input, std::vector<int> &output, bool b_img_prompt = false) override
+    bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) override
     {
         auto ret = sp.Encode(input, &output);
-        if (!ret.ok())
-        {
+        if (!ret.ok()) {
             ALOGE("%s", ret.error_message());
             return false;
         }
-        if (_b_bos)
-        {
+        if (_b_bos) {
             output.insert(output.begin(), sp.bos_id());
         }
-        if (_b_eos)
-        {
+        if (_b_eos) {
             output.push_back(sp.eos_id());
         }
         return true;
     }
 
-    std::vector<int> Encode(std::string input, bool b_img_prompt = false) override
+    std::vector<int> Encode(std::string input, ImageInfo img_info) override
     {
         std::vector<int> output;
-        Encode(input, output, b_img_prompt);
+        Encode(input, output, img_info);
         return output;
     }
 
-    std::string Decode(const std::vector<int> input) override
+    std::string Decode(const std::vector<int> &input) override
     {
         sentencepiece::SentencePieceText spt;
         sp.Decode(input, &spt);
         std::string out = spt.pieces()[0].piece();
-        if (*(unsigned short *)out.data() == 38626)
-        {
+        if (*(unsigned short *)out.data() == 38626) {
             return " " + spt.text();
-        }
-        else
-        {
+        } else {
             return spt.text();
         }
     }
@@ -98,40 +91,32 @@ class TokenizerLLaMa : public BaseTokenizer
     {
         std::ostringstream oss_prompt;
         int messages_len = messages_.size();
-        for(auto &message : messages_)
-        {
-            messages_len --;
-            switch (message.first)
-            {
-            case ROLE_USER:
-            {
-                oss_prompt << "<|user|>\n" << message.second << "</s>";
-            }
-            break;
-            case ROLE_SYSTEM:
-            break;
-            case ROLE_ASSISTANT:
-            break;
-            case ROLE_ASSISTANT_HELP:
-            {
-                if(messages_len == 0)
-                {
-                    oss_prompt << "<|assistant|>\n";
-                }
-            }
-            break;
-            default:
-                break;
+        for (auto &message : messages_) {
+            messages_len--;
+            switch (message.first) {
+                case ROLE_USER: {
+                    oss_prompt << "<|user|>\n" << message.second << "</s>";
+                } break;
+                case ROLE_SYSTEM:
+                    break;
+                case ROLE_ASSISTANT:
+                    break;
+                case ROLE_ASSISTANT_HELP: {
+                    if (messages_len == 0) {
+                        oss_prompt << "<|assistant|>\n";
+                    }
+                } break;
+                default:
+                    break;
             }
         }
         return oss_prompt.str();
     }
 };
 
-class TokenizerMINICPM : public TokenizerLLaMa
-{
+class TokenizerMINICPM : public TokenizerLLaMa {
 public:
-    std::string Decode(const std::vector<int> input) override
+    std::string Decode(const std::vector<int> &input) override
     {
         sentencepiece::SentencePieceText spt;
         sp.Decode(input, &spt);
@@ -141,38 +126,30 @@ class TokenizerMINICPM : public TokenizerLLaMa
     {
         std::ostringstream oss_prompt;
         int messages_len = messages_.size();
-        for(auto &message : messages_)
-        {
-            messages_len --;
-            switch (message.first)
-            {
-            case ROLE_USER:
-            {
-                oss_prompt << "<用户>" << message.second;
-            }
-            break;
-            case ROLE_SYSTEM:
-            break;
-            case ROLE_ASSISTANT:
-            break;
-            case ROLE_ASSISTANT_HELP:
-            {
-                if(messages_len == 0)
-                {
-                    oss_prompt << "<AI>";
-                }
-            }
-            break;
-            default:
-                break;
+        for (auto &message : messages_) {
+            messages_len--;
+            switch (message.first) {
+                case ROLE_USER: {
+                    oss_prompt << "<用户>" << message.second;
+                } break;
+                case ROLE_SYSTEM:
+                    break;
+                case ROLE_ASSISTANT:
+                    break;
+                case ROLE_ASSISTANT_HELP: {
+                    if (messages_len == 0) {
+                        oss_prompt << "<AI>";
+                    }
+                } break;
+                default:
+                    break;
             }
         }
         return oss_prompt.str();
     }
 };
 
-class TokenizerPhi3 : public BaseTokenizer
-{
+class TokenizerPhi3 : public BaseTokenizer {
     sentencepiece::SentencePieceProcessor sp;
     bool _b_bos, _b_eos;
 
@@ -182,8 +159,7 @@ class TokenizerPhi3 : public BaseTokenizer
     bool Init(std::string model_path, bool b_bos = true, bool b_eos = false) override
     {
         auto ret = sp.Load(model_path);
-        if (!ret.ok())
-        {
+        if (!ret.ok()) {
             ALOGE("%s", ret.error_message());
             return false;
         }
@@ -193,46 +169,40 @@ class TokenizerPhi3 : public BaseTokenizer
         return ret.ok();
     }
 
-    bool Encode(std::string input, std::vector<int> &output, bool b_img_prompt = false) override
+    bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) override
     {
         auto ret = sp.Encode(input, &output);
-        if (!ret.ok())
-        {
+        if (!ret.ok()) {
             ALOGE("%s", ret.error_message());
             return false;
         }
-        output.insert(output.begin(), 32010); //"<|user|>"
-        output.push_back(32007);              //"<|end|>"
-        output.push_back(32001);              //"<|assistant|>"
-        if (_b_bos)
-        {
+        output.insert(output.begin(), 32010);  //"<|user|>"
+        output.push_back(32007);               //"<|end|>"
+        output.push_back(32001);               //"<|assistant|>"
+        if (_b_bos) {
             output.insert(output.begin(), sp.bos_id());
         }
-        if (_b_eos)
-        {
+        if (_b_eos) {
             output.push_back(sp.eos_id());
         }
         return true;
     }
 
-    std::vector<int> Encode(std::string input, bool b_img_prompt = false) override
+    std::vector<int> Encode(std::string input, ImageInfo img_info) override
     {
         std::vector<int> output;
-        Encode(input, output, b_img_prompt);
+        Encode(input, output, img_info);
         return output;
     }
 
-    std::string Decode(const std::vector<int> input) override
+    std::string Decode(const std::vector<int> &input) override
     {
         sentencepiece::SentencePieceText spt;
         sp.Decode(input, &spt);
         std::string out = spt.pieces()[0].piece();
-        if (*(unsigned short *)out.data() == 38626)
-        {
+        if (*(unsigned short *)out.data() == 38626) {
             return " " + spt.text();
-        }
-        else
-        {
+        } else {
             return spt.text();
         }
     }
@@ -251,42 +221,35 @@ class TokenizerPhi3 : public BaseTokenizer
     {
         return id == GetEosID() || id > 31999;
     }
+
     std::string apply_chat_template() override
     {
         std::ostringstream oss_prompt;
         int messages_len = messages_.size();
-        for(auto &message : messages_)
-        {
-            messages_len --;
-            switch (message.first)
-            {
-            case ROLE_USER:
-            {
-                oss_prompt << message.second;
-            }
-            break;
-            case ROLE_SYSTEM:
-            break;
-            case ROLE_ASSISTANT:
-            break;
-            case ROLE_ASSISTANT_HELP:
-            {
-                if(messages_len == 0)
-                {
-                    oss_prompt << " ";
-                }
-            }
-            break;
-            default:
-                break;
+        for (auto &message : messages_) {
+            messages_len--;
+            switch (message.first) {
+                case ROLE_USER: {
+                    oss_prompt << message.second;
+                } break;
+                case ROLE_SYSTEM:
+                    break;
+                case ROLE_ASSISTANT:
+                    break;
+                case ROLE_ASSISTANT_HELP: {
+                    if (messages_len == 0) {
+                        oss_prompt << " ";
+                    }
+                } break;
+                default:
+                    break;
             }
         }
         return oss_prompt.str();
     }
 };
 
-class TokenizerQwen : public BaseTokenizer
-{
+class TokenizerQwen : public BaseTokenizer {
     std::shared_ptr<QwenTokenizer> sp;
     bool _b_bos, _b_eos;
 
@@ -295,8 +258,7 @@ class TokenizerQwen : public BaseTokenizer
 public:
     bool Init(std::string model_path, bool b_bos = true, bool b_eos = false) override
     {
-        if (!file_exist(model_path))
-        {
+        if (!file_exist(model_path)) {
             ALOGE("tokenizer model file(%s) not exist", model_path.c_str());
             return false;
         }
@@ -308,14 +270,12 @@ class TokenizerQwen : public BaseTokenizer
         return true;
     }
 
-    bool Encode(std::string input, std::vector<int> &output, bool b_img_prompt = false) override
+    bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) override
     {
-        if (_b_bos)
-        {
+        if (_b_bos) {
             // input += "<|im_start|>";
         }
-        if (_b_eos)
-        {
+        if (_b_eos) {
             input += "<|endoftext|>";
         }
         output = sp->encode(input, 1024);
@@ -323,14 +283,14 @@ class TokenizerQwen : public BaseTokenizer
         return true;
     }
 
-    std::vector<int> Encode(std::string input, bool b_img_prompt = false) override
+    std::vector<int> Encode(std::string input, ImageInfo img_info) override
     {
         std::vector<int> output;
-        Encode(input, output, b_img_prompt);
+        Encode(input, output, img_info);
         return output;
     }
 
-    std::string Decode(const std::vector<int> input) override
+    std::string Decode(const std::vector<int> &input) override
     {
         return sp->decode(input);
     }
@@ -348,108 +308,32 @@ class TokenizerQwen : public BaseTokenizer
     {
         std::ostringstream oss_prompt;
         int messages_len = messages_.size();
-        for(auto &message : messages_)
-        {
-            messages_len --;
-            switch (message.first)
-            {
-            case ROLE_USER:
-            {
-                oss_prompt << "<|im_start|>user\n" << message.second << "<|im_end|>\n";
-            }
-            break;
-            case ROLE_SYSTEM:
-            {
-                oss_prompt << "<|im_start|>system\n" << message.second << ".<|im_end|>\n";
-            }
-            break;
-            case ROLE_ASSISTANT:
-            {
-                oss_prompt << "<|im_start|>assistant\n" << message.second << ".<|im_end|>\n";
-            }
-            break;
-            case ROLE_ASSISTANT_HELP:
-            {
-                if(messages_len == 0)
-                {
-                    oss_prompt << "<|im_start|>assistant\n";
-                }
-            }
-            break;
-            default:
-                break;
+        for (auto &message : messages_) {
+            messages_len--;
+            switch (message.first) {
+                case ROLE_USER: {
+                    oss_prompt << "<|im_start|>user\n" << message.second << "<|im_end|>\n";
+                } break;
+                case ROLE_SYSTEM: {
+                    oss_prompt << "<|im_start|>system\n" << message.second << ".<|im_end|>\n";
+                } break;
+                case ROLE_ASSISTANT: {
+                    oss_prompt << "<|im_start|>assistant\n" << message.second << ".<|im_end|>\n";
+                } break;
+                case ROLE_ASSISTANT_HELP: {
+                    if (messages_len == 0) {
+                        oss_prompt << "<|im_start|>assistant\n";
+                    }
+                } break;
+                default:
+                    break;
             }
         }
         return oss_prompt.str();
     }
 };
 
-// class TokenizerGLM3 : public BaseTokenizer
-// {
-//     std::shared_ptr<chatglm::ChatGLM3Tokenizer> sp;
-//     bool _b_bos, _b_eos;
-
-// private:
-//     /* data */
-// public:
-//     bool Init(std::string model_path, bool b_bos = true, bool b_eos = false) override
-//     {
-//         if (!file_exist(model_path))
-//         {
-//             ALOGE("tokenizer model file(%s) not exist", model_path.c_str());
-//             return false;
-//         }
-//         // std::vector<char> sp_model_data;
-//         // read_file(model_path, sp_model_data);
-//         // std::string_view serialized_model_proto(sp_model_data.data(), sp_model_data.size());
-
-//         sp.reset(new chatglm::ChatGLM3Tokenizer(model_path));
-
-//         this->_b_bos = b_bos;
-//         this->_b_eos = b_eos;
-//         return true;
-//     }
-
-//     bool Encode(std::string input, std::vector<int> &output) override
-//     {
-//         if (_b_bos)
-//         {
-//             // input += "<|im_start|>";
-//         }
-//         if (_b_eos)
-//         {
-//             // input += "<|endoftext|>";
-//         }
-//         output = sp->encode(input, 1024);
-
-//         return true;
-//     }
-
-//     std::vector<int> Encode(std::string input) override
-//     {
-//         std::vector<int> output;
-//         Encode(input, output);
-//         return output;
-//     }
-
-//     std::string Decode(const std::vector<int> input) override
-//     {
-//         return sp->decode(input);
-//     }
-
-//     int GetBosID() override
-//     {
-//         return sp->sp.bos_id();
-//     }
-
-//     int GetEosID() override
-//     {
-//         return sp->sp.eos_id();
-//     }
-// };
-
-class Tokenizer_Http : public BaseTokenizer
-{
+class Tokenizer_Http : public BaseTokenizer {
     std::shared_ptr<httplib::Client> cli;
     bool _b_bos, _b_eos;
 
@@ -457,14 +341,15 @@ class Tokenizer_Http : public BaseTokenizer
 
     int bos_id, eos_id;
 
+    std::string uid;
+
 private:
     /* data */
 public:
     bool Init(std::string model_path = "http://localhost:8080", bool b_bos = true, bool b_eos = false) override
     {
         base_url = model_path;
-        try
-        {
+        try {
             cli = std::make_shared<httplib::Client>(base_url);
             cli->set_connection_timeout(1);
             cli->set_read_timeout(1);
@@ -472,30 +357,26 @@ class Tokenizer_Http : public BaseTokenizer
             {
                 auto ret = cli->Get("/bos_id");
                 auto rep = ret.value();
-                if (rep.status != 200)
-                {
+                if (rep.status != 200) {
                     ALOGE("get bos_id failed, status: %d", rep.status);
                     return false;
                 }
                 nlohmann::json j = nlohmann::json::parse(rep.body);
-                bos_id = j["bos_id"];
+                bos_id           = j["bos_id"];
             }
 
             {
                 auto ret = cli->Get("/eos_id");
                 auto rep = ret.value();
-                if (rep.status != 200)
-                {
+                if (rep.status != 200) {
                     ALOGE("get eos_id failed, status: %d", rep.status);
                     return false;
                 }
                 nlohmann::json j = nlohmann::json::parse(rep.body);
-                eos_id = j["eos_id"];
+                eos_id           = j["eos_id"];
             }
             printf("bos_id: %d, eos_id: %d\n", bos_id, eos_id);
-        }
-        catch (const std::exception &e)
-        {
+        } catch (const std::exception &e) {
             std::cerr << e.what() << '\n';
             return false;
         }
@@ -505,77 +386,201 @@ class Tokenizer_Http : public BaseTokenizer
         return true;
     }
 
-    bool Encode(std::string input, std::vector<int> &output, bool b_img_prompt = false) override
+    bool Init(std::string model_path = "http://localhost:8080") override
+    {
+        base_url = model_path;
+        if (!test_connect_http(base_url, 10)) {
+            ALOGE("connect %s failed", base_url.c_str());
+            return false;
+        } else {
+            ALOGI("connect %s ok", base_url.c_str());
+        }
+
+        cli = std::make_shared<httplib::Client>(base_url);
+        cli->set_connection_timeout(10);
+        cli->set_read_timeout(10);
+        cli->set_write_timeout(10);
+
+        int try_count = 10;
+        int count     = try_count;
+        while (count-- > 0) {
+            try {
+                auto ret = cli->Get("/get_uid");
+                auto rep = ret.value();
+                if (rep.status != 200) {
+                    ALOGE("get uid failed, status: %d", rep.status);
+                    return false;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                uid              = j["uid"];
+                ALOGI("uid: %s", uid.c_str());
+                break;
+            } catch (const std::exception &e) {
+                std::cerr << e.what() << '\n';
+            }
+            std::this_thread::sleep_for(std::chrono::seconds(1));
+            ALOGE("get uid failed, try again %d/%d", count, try_count);
+        }
+
+        count = 10;
+        while (count-- > 0) {
+            try {
+                auto ret = cli->Get("/bos_id?uid=" + uid);
+                auto rep = ret.value();
+                if (rep.status != 200) {
+                    ALOGE("get bos_id failed, status: %d", rep.status);
+                    return false;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                bos_id           = j["bos_id"];
+                break;
+            } catch (const std::exception &e) {
+                std::cerr << e.what() << '\n';
+            }
+            std::this_thread::sleep_for(std::chrono::seconds(1));
+            ALOGE("get bos_id failed, try again %d/%d", count, try_count);
+        }
+
+        count = 10;
+        while (count-- > 0) {
+            try {
+                auto ret = cli->Get("/eos_id?uid=" + uid);
+                auto rep = ret.value();
+                if (rep.status != 200) {
+                    ALOGE("get eos_id failed, status: %d", rep.status);
+                    return false;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                eos_id           = j["eos_id"];
+                break;
+            } catch (const std::exception &e) {
+                std::cerr << e.what() << '\n';
+            }
+            std::this_thread::sleep_for(std::chrono::seconds(1));
+            ALOGE("get eos_id failed, try again %d/%d", count, try_count);
+        }
+
+        printf("bos_id: %d, eos_id: %d\n", bos_id, eos_id);
+
+        return true;
+    }
+
+    bool Encode(std::string input, std::string last_reply, std::vector<int> &tokens, std::vector<int> &tokens_diff,
+                ImageInfo img_info) override
     {
         nlohmann::json j;
+        j["uid"]  = uid;
         j["text"] = input;
-        j["img_prompt"] = b_img_prompt;
-        auto ret = cli->Post("/encode", j.dump(), "application/json");
-        auto rep = ret.value();
-        if (rep.status != 200)
-        {
+        if (!last_reply.empty() and last_reply != "") {
+            j["last_reply"] = last_reply;
+        }
+
+        j["img_prompt"] = img_info.img_prompt;
+        auto ret        = cli->Post("/encode", j.dump(), "application/json");
+        auto rep        = ret.value();
+        if (rep.status != 200) {
             ALOGE("encode failed, status: %d", rep.status);
             return false;
         }
         nlohmann::json j2;
-        try
-        {
+        try {
             j2 = nlohmann::json::parse(rep.body);
+        } catch (const std::exception &e) {
+            ALOGE("json parse failed: %s", e.what());
+            ALOGE("%s", rep.body.c_str());
+            return false;
+        }
+
+        std::vector<int> _token_ids   = j2["token_ids"];
+        std::vector<int> _tokens_diff = j2["diff"];
+
+        tokens      = _token_ids;
+        tokens_diff = _tokens_diff;
+
+        return true;
+    }
+
+    bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) override
+    {
+        nlohmann::json j;
+        j["text"]       = input;
+        j["img_prompt"] = img_info.img_prompt;
+        auto ret        = cli->Post("/encode", j.dump(), "application/json");
+        auto rep        = ret.value();
+        if (rep.status != 200) {
+            ALOGE("encode failed, status: %d", rep.status);
+            return false;
         }
-        catch (const std::exception &e)
-        {
+        nlohmann::json j2;
+        try {
+            j2 = nlohmann::json::parse(rep.body);
+        } catch (const std::exception &e) {
             ALOGE("json parse failed: %s", e.what());
             ALOGE("%s", rep.body.c_str());
             return false;
         }
 
         std::vector<int> out = j2["token_ids"];
-        output = out;
-        // output = sp->encode(input, 1024);
-        if (_b_bos)
-        {
+        output               = out;
+
+        if (_b_bos) {
             output.insert(output.begin(), bos_id);
         }
-        if (_b_eos)
-        {
+        if (_b_eos) {
             output.push_back(eos_id);
         }
 
         return true;
     }
 
-    std::vector<int> Encode(std::string input, bool b_img_prompt = false) override
+    bool Reset(std::string system_prompt, std::vector<int> &tokens) override
+    {
+        nlohmann::json j;
+        j["uid"] = uid;
+        if (!system_prompt.empty() and system_prompt != "") {
+            j["system_prompt"] = system_prompt;
+        }
+
+        auto ret = cli->Post("/reset", j.dump(), "application/json");
+        auto rep = ret.value();
+        if (rep.status != 200) {
+            ALOGE("reset failed, status: %d", rep.status);
+            return false;
+        }
+        nlohmann::json j_rep        = nlohmann::json::parse(rep.body);
+        std::vector<int> _token_ids = j_rep["token_ids"];
+        tokens                      = _token_ids;
+        return true;
+    }
+
+    std::vector<int> Encode(std::string input, ImageInfo img_info) override
     {
         std::vector<int> output;
-        Encode(input, output, b_img_prompt);
+        Encode(input, output, img_info);
         return output;
     }
 
-    std::string Decode(const std::vector<int> input) override
+    std::string Decode(const std::vector<int> &input) override
     {
-        int cnt = 2;
+        int cnt             = 2;
         std::string out_str = "";
-        while (cnt--)
-        {
+        while (cnt--) {
             nlohmann::json j;
             j["token_ids"] = input;
-            auto ret = cli->Post("/decode", j.dump(), "application/json");
-            auto rep = ret.value();
-            if (rep.status != 200)
-            {
+            j["uid"]       = uid;
+            auto ret       = cli->Post("/decode", j.dump(), "application/json");
+            auto rep       = ret.value();
+            if (rep.status != 200) {
                 ALOGE("decode failed, status: %d, try again", rep.status);
                 ALOGE("%s", rep.body.c_str());
                 usleep(1000 * 1000);
                 continue;
             }
-            try
-            {
+            try {
                 nlohmann::json j2 = nlohmann::json::parse(rep.body);
-                out_str = j2["text"];
+                out_str           = j2["text"];
                 break;
-            }
-            catch (const std::exception &e)
-            {
+            } catch (const std::exception &e) {
                 ALOGE("json parse failed: %s, try again", e.what());
                 ALOGE("%s", rep.body.c_str());
                 usleep(1000 * 1000);
@@ -594,34 +599,25 @@ class Tokenizer_Http : public BaseTokenizer
     {
         return eos_id;
     }
+
     std::string apply_chat_template() override
     {
         std::ostringstream oss_prompt;
         int messages_len = messages_.size();
-        for(auto &message : messages_)
-        {
-            messages_len --;
-            switch (message.first)
-            {
-            case ROLE_USER:
-            {
-                oss_prompt << message.second ;
-            }
-            break;
-            case ROLE_SYSTEM:
-            {
-            }
-            break;
-            case ROLE_ASSISTANT:
-            {
-            }
-            break;
-            case ROLE_ASSISTANT_HELP:
-            {
-            }
-            break;
-            default:
-                break;
+        for (auto &message : messages_) {
+            messages_len--;
+            switch (message.first) {
+                case ROLE_USER: {
+                    oss_prompt << message.second;
+                } break;
+                case ROLE_SYSTEM: {
+                } break;
+                case ROLE_ASSISTANT: {
+                } break;
+                case ROLE_ASSISTANT_HELP: {
+                } break;
+                default:
+                    break;
             }
         }
         return oss_prompt.str();
@@ -801,7 +797,7 @@ class Tokenizer_Auto : public BaseTokenizer {
         return ret.success;
     }
 
-    bool Encode(std::string input, std::vector<int> &output, bool b_img_prompt = false) override
+    bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) override
     {
         nlohmann::json rpcobj;
         rpcobj["method"] = "encode";
@@ -815,14 +811,14 @@ class Tokenizer_Auto : public BaseTokenizer {
         return ret.success;
     }
 
-    std::vector<int> Encode(std::string input, bool b_img_prompt = false) override
+    std::vector<int> Encode(std::string input, ImageInfo img_info) override
     {
         std::vector<int> output;
-        Encode(input, output, b_img_prompt);
+        Encode(input, output, img_info);
         return output;
     }
 
-    std::string Decode(const std::vector<int> input) override
+    std::string Decode(const std::vector<int> &input) override
     {
         nlohmann::json rpcobj;
         rpcobj["method"] = "decode";
@@ -897,22 +893,20 @@ class Tokenizer_Auto : public BaseTokenizer {
 
 std::shared_ptr<BaseTokenizer> CreateTokenizer(TokenizerType type)
 {
-    switch (type)
-    {
-    case TKT_LLaMa:
-        return std::make_shared<TokenizerLLaMa>();
-    case TKT_MINICPM:
-        return std::make_shared<TokenizerMINICPM>();
-    case TKT_HTTP:
-        return std::make_shared<Tokenizer_Http>();
-    case TKT_Qwen:
-        return std::make_shared<TokenizerQwen>();
-    case TKT_Phi3:
-        return std::make_shared<TokenizerPhi3>();
-    case TKT_AUTO:
-        return std::make_shared<Tokenizer_Auto>();
-    default:
-        return nullptr;
+    switch (type) {
+        case TKT_LLaMa:
+            return std::make_shared<TokenizerLLaMa>();
+        case TKT_MINICPM:
+            return std::make_shared<TokenizerMINICPM>();
+        case TKT_HTTP:
+            return std::make_shared<Tokenizer_Http>();
+        case TKT_Qwen:
+            return std::make_shared<TokenizerQwen>();
+        case TKT_Phi3:
+            return std::make_shared<TokenizerPhi3>();
+        case TKT_AUTO:
+            return std::make_shared<Tokenizer_Auto>();
+        default:
+            return nullptr;
     }
 }
-
diff --git a/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.hpp b/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.hpp
index ccb8c4a6..ab545999 100644
--- a/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.hpp
+++ b/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.hpp
@@ -5,48 +5,76 @@
 #include <list>
 #include <utility>
 #include <iostream>
-enum TokenizerType
-{
-    TKT_LLaMa,
-    TKT_Qwen,
-    TKT_HTTP,
-    TKT_Phi3,
-    TKT_MINICPM,
-    TKT_AUTO,
-    TKT_END
+enum TokenizerType { TKT_LLaMa, TKT_Qwen, TKT_HTTP, TKT_Phi3, TKT_MINICPM, TKT_AUTO, TKT_END };
+
+enum TokenizeRole {
+    ROLE_USER,           // 用户输入
+    ROLE_SYSTEM,         // 提示词
+    ROLE_TOOL,           // 工具
+    ROLE_IPYTHON,        // 工具
+    ROLE_ASSISTANT,      // 助手回复
+    ROLE_ASSISTANT_HELP  // 询问句
 };
 
-enum TokenizeRole{
-    ROLE_USER,//用户输入
-    ROLE_SYSTEM,//提示词
-    ROLE_TOOL,  //工具
-    ROLE_IPYTHON,  //工具
-    ROLE_ASSISTANT,//助手回复
-    ROLE_ASSISTANT_HELP// 询问句
+struct ImageInfo {
+    int imgsz       = 448;
+    int num_img     = 1;
+    bool img_prompt = false;
 };
 
-class BaseTokenizer
-{
-public:
-    std::list<std::pair<enum TokenizeRole, std::string>> messages_;
-    void messages_clean() {messages_.clear();};
+class BaseTokenizer {
 public:
-    virtual bool Init(std::string model_path, bool b_bos = true, bool b_eos = false) = 0;
-    virtual bool Encode(std::string input, std::vector<int> &output, bool b_img_prompt = false) = 0;
-    virtual std::vector<int> Encode(std::string input, bool b_img_prompt = false) = 0;
-    virtual std::string Decode(const std::vector<int> input) = 0;
-    virtual int GetBosID() = 0;
-    virtual int GetEosID() = 0;
+    std::list<std::pair<TokenizeRole, std::string>> messages_;
+
+    void messages_clean()
+    {
+        messages_.clear();
+    }
+
+    virtual ~BaseTokenizer() = default;
+
+    virtual bool Init(std::string model_path)
+    {
+        return false;
+    };
+
+    virtual bool Init(std::string model_path, bool b_bos, bool b_eos)
+    {
+        return false;
+    };
+
+    virtual bool Reset(std::string system_prompt, std::vector<int> &tokens)
+    {
+        return false;
+    };
+
+    virtual bool Encode(std::string input, std::string last_reply, std::vector<int> &tokens,
+                        std::vector<int> &tokens_diff, ImageInfo img_info)
+    {
+        return false;
+    };
+
+    virtual bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) = 0;
+    virtual std::vector<int> Encode(std::string input, ImageInfo img_info)               = 0;
+    virtual std::string Decode(const std::vector<int> &input)                            = 0;
+    virtual int GetBosID()                                                               = 0;
+    virtual int GetEosID()                                                               = 0;
+
     virtual std::string apply_chat_template() = 0;
-    virtual std::string messages_complete(enum TokenizeRole role, const std::string &content = ""){
+
+    virtual std::string messages_complete(TokenizeRole role, const std::string &content = "")
+    {
         messages_.push_back(std::make_pair(role, content));
-        // std::cout << "messages_complete role:" << role << "content:" << content << std::endl;
-        if(ROLE_ASSISTANT_HELP == role)
+        if (ROLE_ASSISTANT_HELP == role)
             return apply_chat_template();
         else
             return "";
     }
-    virtual bool isEnd(int id) { return id == GetEosID(); }
+
+    virtual bool isEnd(int id)
+    {
+        return id == GetEosID();
+    }
 };
 
 std::shared_ptr<BaseTokenizer> CreateTokenizer(TokenizerType type);
\ No newline at end of file
diff --git a/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner.hpp b/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner.hpp
index 551f65bd..c9ed4fa3 100644
--- a/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner.hpp
+++ b/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner.hpp
@@ -61,6 +61,9 @@ class ax_runner_base
     int get_num_inputs() { return minput_tensors.size(); };
     int get_num_outputs() { return moutput_tensors.size(); };
 
+    int get_num_input_groups() { return mgroup_input_tensors.size(); };
+    int get_num_output_groups() { return mgroup_output_tensors.size(); };
+
     const ax_runner_tensor_t &get_input(int idx) { return minput_tensors[idx]; }
     const ax_runner_tensor_t *get_inputs_ptr() { return minput_tensors.data(); }
     const ax_runner_tensor_t &get_input(std::string name)
diff --git a/projects/llm_framework/main_llm/src/runner/utils/http_utils.hpp b/projects/llm_framework/main_llm/src/runner/utils/http_utils.hpp
new file mode 100644
index 00000000..5041b6a1
--- /dev/null
+++ b/projects/llm_framework/main_llm/src/runner/utils/http_utils.hpp
@@ -0,0 +1,102 @@
+#pragma once
+#include <iostream>
+#include <string>
+#include <cstring>
+#include <sys/socket.h>
+#include <arpa/inet.h>
+#include <unistd.h>
+#include <chrono>
+#include <thread>
+
+/**
+ * @brief Attempts to establish a TCP connection to a specified host and port.
+ *
+ * This function creates a socket and tries to connect to a server specified by
+ * the host and port parameters. It returns true if the connection is successful,
+ * otherwise it returns false and outputs an error message to standard error.
+ *
+ * @param host The IP address of the server to connect to.
+ * @param port The port number of the server to connect to.
+ *
+ * @return true if the connection is successfully established, false otherwise.
+ */
+static bool test_connect(const std::string &host, int port)
+{
+    int sock = socket(AF_INET, SOCK_STREAM, 0);
+    if (sock < 0)
+    {
+        // std::cerr << "Socket creation failed\n";
+        return false;
+    }
+
+    sockaddr_in server_addr{};
+    server_addr.sin_family = AF_INET;
+    server_addr.sin_port = htons(port);
+
+    if (inet_pton(AF_INET, host.c_str(), &server_addr.sin_addr) <= 0)
+    {
+        // std::cerr << "IP address conversion failed\n";
+        close(sock);
+        return false;
+    }
+
+    if (connect(sock, (struct sockaddr *)&server_addr, sizeof(server_addr)) < 0)
+    {
+        // std::cerr << "Connection failed\n";
+        close(sock);
+        return false;
+    }
+
+    close(sock);
+    return true;
+}
+
+/**
+ * @brief Attempts to establish an HTTP connection to a specified URL with a timeout.
+ *
+ * This function parses the provided HTTP URL to extract the host and port information,
+ * and attempts to establish a TCP connection using the `test_connect` function.
+ * It retries the connection until the specified timeout is reached.
+ *
+ * @param http_url The HTTP URL of the server to connect to.
+ * @param timeout The maximum number of seconds to keep attempting the connection.
+ *
+ * @return true if the connection is successfully established within the timeout period,
+ * false otherwise.
+ */
+static bool test_connect_http(const std::string &http_url, int timeout)
+{
+    size_t pos = http_url.find("://");
+    if (pos == std::string::npos)
+        return false;
+    std::string host = http_url.substr(pos + 3);
+    pos = host.find('/');
+    if (pos != std::string::npos)
+        host = host.substr(0, pos);
+    pos = host.find(':');
+    int port = 80;
+    if (pos != std::string::npos)
+    {
+        port = std::stoi(host.substr(pos + 1));
+        host = host.substr(0, pos);
+    }
+    else
+    {
+        return false;
+    }
+
+    if (host == "localhost")
+        host = "127.0.0.1";
+    int tmp = timeout;
+    while (timeout--)
+    {
+        if (test_connect(host, port))
+            return true;
+        std::this_thread::sleep_for(std::chrono::seconds(1));
+        printf("\033[1;30;31m"
+               "connect failed %s, try again in %2d/%2d \n"
+               "\033[0m",
+               http_url.c_str(), timeout, tmp);
+    }
+    return false;
+}
\ No newline at end of file
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 1d2a89f0..1783e536 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -401,6 +401,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-model-qwen2.5-0.5B-prefill-20e':[create_data_deb,'llm-model-qwen2.5-0.5B-prefill-20e', data_version, src_folder, revision],
         'llm-model-qwen2.5-0.5B-p256-ax630c':[create_data_deb,'llm-model-qwen2.5-0.5B-p256-ax630c', '0.4', src_folder, revision],
         'llm-model-qwen2.5-0.5B-Int4-ax630c':[create_data_deb,'llm-model-qwen2.5-0.5B-Int4-ax630c', '0.4', src_folder, revision],
+        'llm-model-qwen2.5-HA-0.5B-ctx-ax630c':[create_data_deb,'llm-model-qwen2.5-HA-0.5B-ctx-ax630c', '0.5', src_folder, revision],
         'llm-model-qwen2.5-1.5B-ax630c':[create_data_deb,'llm-model-qwen2.5-1.5B-ax630c', '0.3', src_folder, revision],
         'llm-model-qwen2.5-1.5B-p256-ax630c':[create_data_deb,'llm-model-qwen2.5-1.5B-p256-ax630c', '0.4', src_folder, revision],
         'llm-model-qwen2.5-1.5B-Int4-ax630c':[create_data_deb,'llm-model-qwen2.5-1.5B-Int4-ax630c', '0.4', src_folder, revision],

From 9167b6e546d12353224e26b03c9db042a1a0d05d Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Tue, 26 Aug 2025 17:20:31 +0800
Subject: [PATCH 30/79] [update] update llm_vlm encoder.  update audio cache.

---
 projects/llm_framework/main_asr/src/main.cpp  |   14 +-
 projects/llm_framework/main_kws/src/main.cpp  |   16 +-
 projects/llm_framework/main_llm/src/main.cpp  |    1 +
 projects/llm_framework/main_vad/src/main.cpp  |   13 +-
 projects/llm_framework/main_vlm/src/main.cpp  |  226 +-
 .../llm_framework/main_vlm/src/runner/LLM.hpp | 1167 ++++++++-
 .../src/runner/Tokenizer/QwenTokenizer.cpp    |  127 -
 .../src/runner/Tokenizer/QwenTokenizer.hpp    |   34 -
 .../src/runner/Tokenizer/Tokenizer.cpp        |  576 ++---
 .../src/runner/Tokenizer/Tokenizer.hpp        |   21 +-
 .../main_vlm/src/runner/Tokenizer/base64.h    |   54 -
 .../main_vlm/src/runner/Tokenizer/tiktoken.h  |  326 ---
 .../src/runner/Tokenizer/unordered_dense.h    | 2240 -----------------
 .../ax_model_runner/ax_model_runner.hpp       |    3 +
 .../main_vlm/src/runner/utils/http_utils.hpp  |  102 +
 .../llm_framework/main_whisper/src/main.cpp   |   28 +-
 16 files changed, 1725 insertions(+), 3223 deletions(-)
 delete mode 100644 projects/llm_framework/main_vlm/src/runner/Tokenizer/QwenTokenizer.cpp
 delete mode 100644 projects/llm_framework/main_vlm/src/runner/Tokenizer/QwenTokenizer.hpp
 delete mode 100644 projects/llm_framework/main_vlm/src/runner/Tokenizer/base64.h
 delete mode 100644 projects/llm_framework/main_vlm/src/runner/Tokenizer/tiktoken.h
 delete mode 100644 projects/llm_framework/main_vlm/src/runner/Tokenizer/unordered_dense.h
 create mode 100644 projects/llm_framework/main_vlm/src/runner/utils/http_utils.hpp

diff --git a/projects/llm_framework/main_asr/src/main.cpp b/projects/llm_framework/main_asr/src/main.cpp
index 4a5f9703..ae51982b 100644
--- a/projects/llm_framework/main_asr/src/main.cpp
+++ b/projects/llm_framework/main_asr/src/main.cpp
@@ -58,7 +58,7 @@ class llm_task {
     std::atomic_bool audio_flage_;
     std::atomic_bool awake_flage_;
     int awake_delay_       = 50;
-    int delay_audio_frame_ = 10;
+    int delay_audio_frame_ = 11;
     buffer_t *pcmdata;
 
     std::function<void(void)> pause;
@@ -187,18 +187,20 @@ class llm_task {
             count++;
             return;
         }
-        buffer_write_char(pcmdata, raw.data(), raw.length());
         buffer_position_set(pcmdata, 0);
-        count = 0;
+
         std::vector<float> floatSamples;
         {
             int16_t audio_val;
-            while (buffer_read_u16(pcmdata, (unsigned short *)&audio_val, 1)) {
-                float normalizedSample = (float)audio_val / INT16_MAX;
+            while (buffer_read_i16(pcmdata, &audio_val, 1)) {
+                float normalizedSample = static_cast<float>(audio_val) / INT16_MAX;
                 floatSamples.push_back(normalizedSample);
             }
         }
-        buffer_position_set(pcmdata, 0);
+
+        buffer_resize(pcmdata, 0);
+        count = 0;
+
         if (awake_flage_ && recognizer_stream_) {
             recognizer_stream_.reset();
             awake_flage_ = false;
diff --git a/projects/llm_framework/main_kws/src/main.cpp b/projects/llm_framework/main_kws/src/main.cpp
index 18e4dfd2..8c724599 100644
--- a/projects/llm_framework/main_kws/src/main.cpp
+++ b/projects/llm_framework/main_kws/src/main.cpp
@@ -59,7 +59,7 @@ class llm_task {
     bool enwake_audio_;
     std::atomic_bool audio_flage_;
     task_callback_t out_callback_;
-    int delay_audio_frame_ = 10;
+    int delay_audio_frame_ = 11;
     buffer_t *pcmdata;
     std::string wake_wav_file_;
 
@@ -229,22 +229,24 @@ class llm_task {
     {
         static int count = 0;
         if (count < delay_audio_frame_) {
-            buffer_write_char(pcmdata, raw.c_str(), raw.length());
+            buffer_write_char(pcmdata, raw.data(), raw.length());
             count++;
             return;
         }
-        buffer_write_char(pcmdata, raw.data(), raw.length());
         buffer_position_set(pcmdata, 0);
-        count = 0;
+
         std::vector<float> floatSamples;
         {
             int16_t audio_val;
-            while (buffer_read_u16(pcmdata, (unsigned short *)&audio_val, 1)) {
-                float normalizedSample = (float)audio_val / INT16_MAX;
+            while (buffer_read_i16(pcmdata, &audio_val, 1)) {
+                float normalizedSample = static_cast<float>(audio_val) / INT16_MAX;
                 floatSamples.push_back(normalizedSample);
             }
         }
-        buffer_position_set(pcmdata, 0);
+
+        buffer_resize(pcmdata, 0);
+        count = 0;
+
         spotter_stream_->AcceptWaveform(mode_config_.feat_config.sampling_rate, floatSamples.data(),
                                         floatSamples.size());
         while (spotter_->IsReady(spotter_stream_.get())) {
diff --git a/projects/llm_framework/main_llm/src/main.cpp b/projects/llm_framework/main_llm/src/main.cpp
index e198cc9c..dcb64fb8 100644
--- a/projects/llm_framework/main_llm/src/main.cpp
+++ b/projects/llm_framework/main_llm/src/main.cpp
@@ -130,6 +130,7 @@ class llm_task {
             std::string base_model = base_model_path_ + model_ + "/";
             SLOGI("base_model %s", base_model.c_str());
 
+            CONFIG_AUTO_SET(file_body["mode_param"], system_prompt);   
             CONFIG_AUTO_SET(file_body["mode_param"], tokenizer_type);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_tokenizer_model);
             CONFIG_AUTO_SET(file_body["mode_param"], url_tokenizer_model);
diff --git a/projects/llm_framework/main_vad/src/main.cpp b/projects/llm_framework/main_vad/src/main.cpp
index 20a67eb3..1c49eb33 100644
--- a/projects/llm_framework/main_vad/src/main.cpp
+++ b/projects/llm_framework/main_vad/src/main.cpp
@@ -60,7 +60,7 @@ class llm_task {
     std::string superior_id_;
     task_callback_t out_callback_;
     int awake_delay_       = 50;
-    int delay_audio_frame_ = 3;
+    int delay_audio_frame_ = 4;
     buffer_t *pcmdata;
     std::string wake_wav_file_;
 
@@ -158,18 +158,19 @@ class llm_task {
             count++;
             return;
         }
-        buffer_write_char(pcmdata, raw.data(), raw.length());
         buffer_position_set(pcmdata, 0);
-        count = 0;
+
         std::vector<float> floatSamples;
         {
             int16_t audio_val;
-            while (buffer_read_u16(pcmdata, (unsigned short *)&audio_val, 1)) {
-                float normalizedSample = (float)audio_val / INT16_MAX;
+            while (buffer_read_i16(pcmdata, &audio_val, 1)) {
+                float normalizedSample = static_cast<float>(audio_val) / INT16_MAX;
                 floatSamples.push_back(normalizedSample);
             }
         }
-        buffer_position_set(pcmdata, 0);
+        buffer_resize(pcmdata, 0);
+        count = 0;
+
         vad_->AcceptWaveform(floatSamples.data(), floatSamples.size());
 
         if (vad_->IsSpeechDetected() && !printed) {
diff --git a/projects/llm_framework/main_vlm/src/main.cpp b/projects/llm_framework/main_vlm/src/main.cpp
index 6ee6dd67..83a56e51 100644
--- a/projects/llm_framework/main_vlm/src/main.cpp
+++ b/projects/llm_framework/main_vlm/src/main.cpp
@@ -52,6 +52,7 @@ class llm_task {
 public:
     LLMAttrType mode_config_;
     std::unique_ptr<LLM> lLaMa_;
+    std::unique_ptr<LLM_CTX> lLaMa_ctx_;
     std::string model_;
     std::string response_format_;
     std::vector<std::string> inputs_;
@@ -59,6 +60,11 @@ class llm_task {
     std::vector<unsigned char> image_data_;
     std::vector<unsigned short> img_embed;
     std::string prompt_;
+    std::string last_reply;
+    std::vector<int> tokens_ids, tokens_diff;
+    std::vector<std::vector<unsigned short>> k_caches, v_caches;
+    int precompute_len = 0;
+    std::vector<int> _token_ids;
     task_callback_t out_callback_;
     bool enoutput_;
     bool enstream_;
@@ -122,8 +128,10 @@ class llm_task {
             std::string base_model = base_model_path_ + model_ + "/";
             SLOGI("base_model %s", base_model.c_str());
 
+            CONFIG_AUTO_SET(file_body["mode_param"], system_prompt);
             CONFIG_AUTO_SET(file_body["mode_param"], tokenizer_type);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_tokenizer_model);
+            CONFIG_AUTO_SET(file_body["mode_param"], url_tokenizer_model);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_tokens_embed);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_post_axmodel);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_vpm_resampler_axmodedl);
@@ -139,64 +147,123 @@ class llm_task {
             CONFIG_AUTO_SET(file_body["mode_param"], b_use_mmap_load_embed);
             CONFIG_AUTO_SET(file_body["mode_param"], b_dynamic_load_axmodel_layer);
             CONFIG_AUTO_SET(file_body["mode_param"], max_token_len);
+            CONFIG_AUTO_SET(file_body["mode_param"], enable_temperature);
             CONFIG_AUTO_SET(file_body["mode_param"], temperature);
+            CONFIG_AUTO_SET(file_body["mode_param"], enable_top_p_sampling);
             CONFIG_AUTO_SET(file_body["mode_param"], top_p);
+            CONFIG_AUTO_SET(file_body["mode_param"], enable_top_k_sampling);
+            CONFIG_AUTO_SET(file_body["mode_param"], top_k);
+            CONFIG_AUTO_SET(file_body["mode_param"], enable_repetition_penalty);
+            CONFIG_AUTO_SET(file_body["mode_param"], repetition_penalty);
+            CONFIG_AUTO_SET(file_body["mode_param"], penalty_window);
             CONFIG_AUTO_SET(file_body["mode_param"], vpm_width);
             CONFIG_AUTO_SET(file_body["mode_param"], vpm_height);
+            CONFIG_AUTO_SET(file_body["mode_param"], precompute_len);
+
+            {
+                auto has_http = [](const std::string &s) { return s.find("http") != std::string::npos; };
+
+                auto find_tokenizer_file = [this]() -> std::string {
+                    const std::string base = "/opt/m5stack/scripts/";
+                    const std::string a    = base + model_ + "_tokenizer.py";
+                    if (file_exists(a)) return a;
+                    const std::string b = base + "tokenizer_" + model_ + ".py";
+                    if (file_exists(b)) return b;
+                    SLOGE("%s or %s not found!", a.c_str(), b.c_str());
+                    return {};
+                };
+
+                auto start_tokenizer_server = [&](const std::string &tokenizer_file) {
+                    if (tokenizer_file.empty()) return;
+                    if (tokenizer_server_flage_.load()) return;
 
-            if (mode_config_.filename_tokenizer_model.find("http:") != std::string::npos) {
-                mode_config_.filename_tokenizer_model = "http://localhost:" + std::to_string(port_);
-                std::string tokenizer_file;
-                if (file_exists(std::string("/opt/m5stack/scripts/") + model_ + std::string("_tokenizer.py"))) {
-                    tokenizer_file = std::string("/opt/m5stack/scripts/") + model_ + std::string("_tokenizer.py");
-                } else if (file_exists(std::string("/opt/m5stack/scripts/") + std::string("tokenizer_") + model_ +
-                                       std::string(".py"))) {
-                    tokenizer_file =
-                        std::string("/opt/m5stack/scripts/") + std::string("tokenizer_") + model_ + std::string(".py");
-                } else {
-                    std::string __log = model_ + std::string("_tokenizer.py");
-                    __log += " or ";
-                    __log += std::string("tokenizer_") + model_ + std::string(".py");
-                    __log += " not found!";
-                    SLOGE("%s", __log.c_str());
-                }
-                if (!tokenizer_server_flage_.load()) {
                     tokenizer_pid_ = fork();
                     if (tokenizer_pid_ == 0) {
-                        setenv("PYTHONPATH", "/opt/m5stack/lib/vlm/site-packages", 1);
+                        setenv("PYTHONPATH", "/opt/m5stack/lib/llm/site-packages", 1);
+                        const std::string port_str = std::to_string(port_);
+                        const std::string model_id = base_model + "tokenizer";
+
                         execl("/usr/bin/python3", "python3", tokenizer_file.c_str(), "--host", "localhost", "--port",
-                              std::to_string(port_).c_str(), "--model_id", (base_model + "tokenizer").c_str(),
-                              "--content", ("'" + prompt_ + "'").c_str(), nullptr);
+                              port_str.c_str(), "--model_id", model_id.c_str(), "--content", prompt_.c_str(),
+                              (char *)nullptr);
+
                         perror("execl failed");
-                        exit(1);
+                        _exit(1);
                     }
+
                     tokenizer_server_flage_.store(true);
                     SLOGI("port_=%s model_id=%s content=%s", std::to_string(port_).c_str(),
-                          (base_model + "tokenizer").c_str(), ("'" + prompt_ + "'").c_str());
+                          (base_model + std::string("tokenizer")).c_str(), prompt_.c_str());
+
                     std::this_thread::sleep_for(std::chrono::seconds(15));
+                };
+
+                auto process_field = [&](std::string &field, const char *name_for_log) -> bool {
+                    if (!has_http(field)) return false;
+
+                    field                            = "http://localhost:" + std::to_string(port_);
+                    const std::string tokenizer_file = find_tokenizer_file();
+                    start_tokenizer_server(tokenizer_file);
+                    SLOGI("%s: %s", name_for_log, field.c_str());
+                    return true;
+                };
+
+                if (!process_field(mode_config_.filename_tokenizer_model, "filename_tokenizer_model") &&
+                    !process_field(mode_config_.url_tokenizer_model, "url_tokenizer_model")) {
+                    mode_config_.filename_tokenizer_model = base_model + mode_config_.filename_tokenizer_model;
+                    SLOGE("filename_tokenizer_model: %s", mode_config_.filename_tokenizer_model.c_str());
                 }
-            } else {
-                mode_config_.filename_tokenizer_model = base_model + mode_config_.filename_tokenizer_model;
             }
-            SLOGI("filename_tokenizer_model: %s", mode_config_.filename_tokenizer_model.c_str());
             mode_config_.filename_tokens_embed           = base_model + mode_config_.filename_tokens_embed;
             mode_config_.filename_post_axmodel           = base_model + mode_config_.filename_post_axmodel;
-            mode_config_.filename_vpm_resampler_axmodedl = base_model + mode_config_.filename_vpm_resampler_axmodedl;
             mode_config_.template_filename_axmodel       = base_model + mode_config_.template_filename_axmodel;
-
+            mode_config_.filename_vpm_resampler_axmodedl = base_model + mode_config_.filename_vpm_resampler_axmodedl;
             mode_config_.runing_callback = [this](int *p_token, int n_token, const char *p_str, float token_per_sec,
                                                   void *reserve) {
                 if (this->out_callback_) {
                     this->out_callback_(std::string(p_str), false);
                 }
             };
-            lLaMa_ = std::make_unique<LLM>();
-            if (!lLaMa_->Init(mode_config_)) {
-                lLaMa_->Deinit();
-                lLaMa_.reset();
-                return -2;
+
+            if (mode_config_.precompute_len > 0) {
+                lLaMa_ctx_ = std::make_unique<LLM_CTX>();
+                if (!lLaMa_ctx_->Init(mode_config_)) {
+                    lLaMa_ctx_->Deinit();
+                    lLaMa_ctx_.reset();
+                    return -2;
+                }
+            } else {
+                lLaMa_ = std::make_unique<LLM>();
+                if (!lLaMa_->Init(mode_config_)) {
+                    lLaMa_->Deinit();
+                    lLaMa_.reset();
+                    return -2;
+                }
             }
 
+            if (lLaMa_ctx_) {
+                lLaMa_ctx_->SetSystemPrompt(mode_config_.system_prompt, _token_ids);
+                std::string kvcache_path = "/tmp/.vlm/";
+                if (!kvcache_path.empty() && kvcache_path != "") {
+                    if (lLaMa_ctx_->load_kvcache(kvcache_path, mode_config_.axmodel_num, k_caches, v_caches,
+                                                 mode_config_.system_prompt, precompute_len)) {
+                        ALOGI("load kvcache from path: %s success,precompute_len: %d", kvcache_path.c_str(),
+                              precompute_len);
+                    } else {
+                        ALOGW("load kvcache from path: %s failed,generate kvcache", kvcache_path.c_str());
+                        lLaMa_ctx_->GenerateKVCachePrefill(_token_ids, k_caches, v_caches, precompute_len);
+                        if (!lLaMa_ctx_->save_kvcache(kvcache_path, mode_config_.system_prompt, precompute_len,
+                                                      k_caches, v_caches)) {
+                            ALOGE("save kvcache failed");
+                        }
+                        ALOGI("generate kvcache to path: %s", kvcache_path.c_str());
+                    }
+                } else {
+                    lLaMa_ctx_->GenerateKVCachePrefill(_token_ids, k_caches, v_caches, precompute_len);
+                }
+                ALOGI("precompute_len: %d", precompute_len);
+                ALOGI("system_prompt: %s", mode_config_.system_prompt.c_str());
+            }
         } catch (...) {
             SLOGE("config false");
             return -3;
@@ -208,19 +275,6 @@ class llm_task {
     {
         std::ostringstream oss_prompt;
         switch (mode_config_.tokenizer_type) {
-            case TKT_LLaMa:
-                oss_prompt << "<|user|>\n" << input << "</s><|assistant|>\n";
-                break;
-            case TKT_MINICPM:
-                oss_prompt << "<用户>" << input << "<AI>";
-                break;
-            case TKT_Phi3:
-                oss_prompt << input << " ";
-                break;
-            case TKT_Qwen:
-                oss_prompt << "<|im_start|>system\n" << prompt_ << ".<|im_end|>";
-                oss_prompt << "\n<|im_start|>user\n" << input << "<|im_end|>\n<|im_start|>assistant\n";
-                break;
             case TKT_HTTP:
             default:
                 oss_prompt << input;
@@ -255,32 +309,68 @@ class llm_task {
     void inference(const std::string &msg)
     {
         try {
-            if (encamera_) {
-                inference_async_par par;
-                async_list_.get();  // discard buffered frames
-                par = async_list_.get();
-                if (par.inference_src.empty()) return;
-                if (par.inference_bgr2rgb) {
-                    cv::Mat rgb;
-                    cv::cvtColor(par.inference_src, rgb, cv::COLOR_BGR2RGB);
-                    par.inference_src = rgb;
+            if (lLaMa_) {
+                if (encamera_) {
+                    inference_async_par par;
+                    async_list_.get();  // discard buffered frames
+                    par = async_list_.get();
+                    if (par.inference_src.empty()) return;
+                    if (par.inference_bgr2rgb) {
+                        cv::Mat rgb;
+                        cv::cvtColor(par.inference_src, rgb, cv::COLOR_BGR2RGB);
+                        par.inference_src = rgb;
+                    }
+                    lLaMa_->Encode(par.inference_src, img_embed);
+                    lLaMa_->Encode(img_embed, prompt_data_, prompt_complete(msg));
+                    std::string out = lLaMa_->Run(prompt_data_);
+                    if (out_callback_) out_callback_(out, true);
+                } else if (image_data_.empty()) {
+                    lLaMa_->Encode(prompt_data_, prompt_complete(msg));
+                    std::string out = lLaMa_->Run(prompt_data_);
+                    if (out_callback_) out_callback_(out, true);
+                } else {
+                    cv::Mat src = cv::imdecode(image_data_, cv::IMREAD_COLOR);
+                    if (src.empty()) return;
+                    image_data_.clear();
+                    lLaMa_->Encode(src, img_embed);
+                    lLaMa_->Encode(img_embed, prompt_data_, prompt_complete(msg));
+                    std::string out = lLaMa_->Run(prompt_data_);
+                    if (out_callback_) out_callback_(out, true);
+                }
+            }
+
+            if (lLaMa_ctx_) {
+                if (image_data_.empty()) {
+                    lLaMa_ctx_->Encode(prompt_data_, prompt_complete(prompt_), last_reply, tokens_ids, tokens_diff);
+                    if (auto ret = lLaMa_ctx_->SetKVCache(k_caches, v_caches, precompute_len, tokens_diff.size());
+                        ret != 0) {
+                        ALOGE("SetKVCache failed: %d,the context may be full,input \"reset\" to reset context", ret);
+                        return;
+                    }
+                    last_reply = lLaMa_ctx_->Run(prompt_data_);
+                    lLaMa_ctx_->GetKVCache(k_caches, v_caches, precompute_len);
                 }
-                lLaMa_->Encode(par.inference_src, img_embed);
-                lLaMa_->Encode(img_embed, prompt_data_, prompt_complete(msg));
-                std::string out = lLaMa_->Run(prompt_data_);
-                if (out_callback_) out_callback_(out, true);
-            } else if (image_data_.empty()) {
-                lLaMa_->Encode(prompt_data_, prompt_complete(msg));
-                std::string out = lLaMa_->Run(prompt_data_);
-                if (out_callback_) out_callback_(out, true);
-            } else {
                 cv::Mat src = cv::imdecode(image_data_, cv::IMREAD_COLOR);
                 if (src.empty()) return;
                 image_data_.clear();
-                lLaMa_->Encode(src, img_embed);
-                lLaMa_->Encode(img_embed, prompt_data_, prompt_complete(msg));
-                std::string out = lLaMa_->Run(prompt_data_);
-                if (out_callback_) out_callback_(out, true);
+                std::vector<unsigned short> img_embed;
+                if (auto ret = lLaMa_ctx_->Encode(src, img_embed); ret != 0) {
+                    ALOGE("lLaMaCtx.Encode failed");
+                    return;
+                }
+                if (auto ret =
+                        lLaMa_ctx_->Encode(img_embed, prompt_data_, prompt_complete(prompt_), tokens_ids, tokens_diff);
+                    ret != 0) {
+                    ALOGE("lLaMaCtx.Encode failed");
+                    return;
+                }
+                if (auto ret = lLaMa_ctx_->SetKVCache(k_caches, v_caches, precompute_len, tokens_diff.size());
+                    ret != 0) {
+                    ALOGE("SetKVCache failed: %d,the context may be full,input \"reset\" to reset context", ret);
+                    return;
+                }
+                last_reply = lLaMa_ctx_->Run(prompt_data_);
+                lLaMa_ctx_->GetKVCache(k_caches, v_caches, precompute_len);
             }
         } catch (...) {
             SLOGW("lLaMa_->Run have error!");
@@ -333,11 +423,9 @@ class llm_task {
         if (tokenizer_pid_ != -1) {
             kill(tokenizer_pid_, SIGTERM);
             waitpid(tokenizer_pid_, nullptr, WNOHANG);
-            // tokenizer_pid_ = -1;
         }
         if (lLaMa_) {
             lLaMa_->Deinit();
-            // lLaMa_.reset();
         }
     }
 };
diff --git a/projects/llm_framework/main_vlm/src/runner/LLM.hpp b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
index 2cbbf388..f7a8b560 100644
--- a/projects/llm_framework/main_vlm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
@@ -14,39 +14,59 @@
 #include "ax_sys_api.h"
 #include "LLMPostprocess.hpp"
 
+#define ALIGN_DOWN(x, a) ((x) & ~((a) - 1))
+
 typedef std::function<void(int *, int, const char *, float, void *)> LLMRuningCallback;
 
 struct LLMAttrType {
+    std::string system_prompt;
     std::string template_filename_axmodel = "tinyllama-int8/tinyllama_l%d.axmodel";
     int axmodel_num                       = 22;
 
-    // std::string template_prefill_filename_axmodel = "minicpmv/prefill_axmodel/minicpm_p96_l%d.axmodel";
-    // int prefill_axmodel_num = 40;
-    int prefill_token_num = 96;
-
-    std::string filename_post_axmodel = "tinyllama-int8/tinyllama_post.axmodel";
-
+    std::string filename_post_axmodel           = "tinyllama-int8/tinyllama_post.axmodel";
+    std::string filename_image_encoder_axmodedl = "minicpmv/vpm_resampler_version0_fp16.axmodel";
     std::string filename_vpm_encoder_axmodedl   = "minicpmv/vpm_resampler_version0_fp16.axmodel";
     std::string filename_vpm_resampler_axmodedl = "minicpmv/vpm_resampler_version0_fp16.axmodel";
-    int vpm_width                               = 280;
-    int vpm_height                              = 280;
-    bool b_vpm_two_stage                        = false;
+
+    int image_encoder_width  = 448;
+    int image_encoder_height = 448;
+    int vpm_width            = 280;
+    int vpm_height           = 280;
+    bool b_vpm_two_stage     = false;
+
+    int prefill_token_num     = 96;
+    int prefill_max_token_num = 512;
+    std::vector<int> prefill_max_kv_cache_num_grp;
+    int precompute_len = 0;
+    int prefill_grpid  = -1;
 
     TokenizerType tokenizer_type         = TKT_LLaMa;
     std::string filename_tokenizer_model = "tokenizer.model";
+    std::string url_tokenizer_model;
     bool b_bos = true, b_eos = false;
     std::string filename_tokens_embed = "tinyllama.model.embed_tokens.weight.bfloat16.bin";
     int tokens_embed_num              = 32000;
-    int img_token_id                  = 151667;  // InternVL2.5
+    int img_token_id                  = 151667;
     int tokens_embed_size             = 2048;
 
-    int max_token_len = 127;  // auto calc
+    int max_token_len = 127;
+
+    int kv_cache_num  = 1024;
+    int kv_cache_size = 256;
+
+    bool enable_temperature = false;
+    float temperature       = 0.7f;
+
+    bool enable_top_p_sampling = false;
+    float top_p                = 0.7f;
 
-    int kv_cache_num  = 1024;  // auto calc
-    int kv_cache_size = 256;   // auto calc
+    bool enable_top_k_sampling = false;
+    int top_k                  = 50;
+
+    bool enable_repetition_penalty = false;
+    float repetition_penalty       = 1.2f;
+    int penalty_window             = 50;
 
-    float temperature                 = 0.7f;
-    float top_p                       = 0.9f;
     bool b_use_mmap_load_embed        = false;
     bool b_dynamic_load_axmodel_layer = false;
 
@@ -58,6 +78,11 @@ struct LLMAttrType {
     // bool b_live_print = true;
     LLMRuningCallback runing_callback = nullptr;
     void *reserve                     = nullptr;
+
+    int IMAGE_CONTEXT_TOKEN       = 151667;
+    int IMAGE_START_TOKEN         = 151665;
+    int IMAGE_ENCODER_INPUT_NCHW  = -1;
+    int IMAGE_ENCODER_OUTPUT_BF16 = -1;
 };
 
 class LLM {
@@ -82,8 +107,6 @@ class LLM {
     int prefill_grpid = 1;
     int decode_grpid  = 0;
 
-    // std::vector<std::vector<unsigned short>> k_caches, v_caches;
-
     bool b_stop = false;
 
     LLMPostprocess postprocess;
@@ -111,17 +134,6 @@ class LLM {
             return false;
         }
         update_cqdm(&cqdm, 0, "count", "tokenizer init ok");
-        // test code
-        // {
-        //     std::vector<int> output;
-        //     tokenizer.Encode("Today is National", output);
-        //     // print output
-        //     for (size_t i = 0; i < output.size(); i++)
-        //     {
-        //         printf("%d ", output[i]);
-        //     }
-        //     printf("\n");
-        // }
 
         if (!embed_selector.Init(attr.filename_tokens_embed, attr.tokens_embed_num, attr.tokens_embed_size,
                                  attr.b_use_mmap_load_embed)) {
@@ -130,20 +142,8 @@ class LLM {
             return false;
         }
         update_cqdm(&cqdm, 1, "count", "embed_selector init ok");
-        // test code
-        // {
-        //     std::vector<unsigned short> embed = embed_selector.getByIndex(123);
-        //     printf("embed size: %d\n", embed.size());
-        //     for (int i = 0; i < embed.size(); i++)
-        //     {
-        //         bfloat16 bf16 = bfloat16(embed[i]);
-        //         float val = bf16;
-        //         printf("%d %0.22f\n", embed[i], val);
-        //     }
-        // }
 
         llama_layers.resize(attr.axmodel_num);
-        // prefill_layers.resize(attr.prefill_axmodel_num);
 
         char axmodel_path[1024];
         for (int i = 0; i < attr.axmodel_num; i++) {
@@ -230,8 +230,6 @@ class LLM {
             int max_token_len   = llama_layers[0].layer.get_input("mask").nSize / sizeof(unsigned short) - 1;
             _attr.max_token_len = max_token_len > _attr.max_token_len ? _attr.max_token_len : max_token_len;
             ALOGI("max_token_len : %d", _attr.max_token_len);
-            // auto &input_k_cache = llama_layers[0].layer.get_input("K_cache");
-            // auto &output_k_cache_out = llama_layers[0].layer.get_output("K_cache_out");
             _attr.kv_cache_size = llama_layers[0].layer.get_output("K_cache_out").nSize / sizeof(unsigned short);
             _attr.kv_cache_num =
                 llama_layers[0].layer.get_input("K_cache").nSize / _attr.kv_cache_size / sizeof(unsigned short);
@@ -308,14 +306,15 @@ class LLM {
             out_embed[i] = bfloat16(output_data[i]).data;
         }
 
-        // memcpy(out_embed.data(), vpm_resampler.get_output(0).pVirAddr, vpm_resampler.get_output(0).nSize);
         ALOGI("image encode time : %f ms, size : %d", t.cost(), out_embed.size());
         return 0;
     }
 
     int Encode(std::vector<unsigned short> &out_embed, std::string prompt = "What is in the image?")
     {
-        std::vector<int> input_ids = tokenizer->Encode(prompt, false);
+        ImageInfo img_info;
+        img_info.img_prompt        = false;
+        std::vector<int> input_ids = tokenizer->Encode(prompt, img_info);
         if (input_ids.size() > _attr.prefill_token_num) {
             ALOGE("input_ids(%d) > prefill_token_num(%d)", input_ids.size(), _attr.prefill_token_num);
             return -1;
@@ -326,16 +325,17 @@ class LLM {
             embed_selector.getByIndex(input_ids[i], out_embed.data() + i * _attr.tokens_embed_size);
         }
 
-        // memcpy(out_embed.data() + 5 * _attr.tokens_embed_size, vpm_resampler.get_output(0).pVirAddr,
-        // vpm_resampler.get_output(0).nSize);
-
         return 0;
     }
 
     int Encode(std::vector<unsigned short> &img_embed, std::vector<unsigned short> &out_embed,
                std::string prompt = "What is in the image?")
     {
-        std::vector<int> input_ids = tokenizer->Encode(prompt, true);
+        ImageInfo img_info;
+        img_info.img_prompt        = true;
+        img_info.num_img           = 1;
+        img_info.imgsz             = _attr.image_encoder_width;
+        std::vector<int> input_ids = tokenizer->Encode(prompt, img_info);
 
         // constexpr int img_token_id = 49190;	// smolvlm
         // constexpr int img_token_id = 151667; // InternVL2.5
@@ -362,12 +362,6 @@ class LLM {
             return -1;
         }
 
-        // for (size_t i = 0; i < input_ids.size(); i++)
-        // {
-        //     printf("%d ", input_ids[i]);
-        // }
-        // printf("\n");
-
         if (input_ids.size() > _attr.prefill_token_num) {
             ALOGE("input_ids(%d) > prefill_token_num(%d)", input_ids.size(), _attr.prefill_token_num);
             return -1;
@@ -658,3 +652,1068 @@ class LLM {
         return final_out;
     }
 };
+
+class LLM_CTX {
+private:
+    std::shared_ptr<BaseTokenizer> tokenizer;
+    LLaMaEmbedSelector embed_selector;
+
+    LLMAttrType _attr;
+
+    struct LLMLayer {
+        ax_runner_ax650 layer;
+        std::string filename;
+        MMap layer_buffer;
+        std::vector<char> layer_buffer_vec;
+    };
+
+    std::vector<LLMLayer> llama_layers;
+    ax_runner_ax650 llama_post;
+    ax_runner_ax650 image_encoder;
+
+    //
+    int decode_grpid = 0;
+
+    bool b_stop = false;
+
+    LLMPostprocess postprocess;
+    static int post_process(LLMPostprocess &postprocess, unsigned short *p, int n, std::vector<int> &history,
+                            float *val = 0)
+    {
+        std::vector<float> logits(n);
+        for (int i = 0; i < n; i++) {
+            unsigned int proc = p[i] << 16;
+            logits[i]         = *reinterpret_cast<float *>(&proc);
+        }
+
+        return postprocess.apply(logits, history);
+    }
+
+public:
+    bool Init(LLMAttrType attr)
+    {
+        ALOGI("LLM init start");
+        t_cqdm cqdm = create_cqdm(attr.axmodel_num + 3, 32);
+        this->_attr = attr;
+        tokenizer   = CreateTokenizer(attr.tokenizer_type);
+        if (!tokenizer->Init(attr.url_tokenizer_model)) {
+            ALOGE("tokenizer.Init(%s) failed", attr.url_tokenizer_model.c_str());
+            return false;
+        }
+        std::vector<int> _token_ids;
+        tokenizer->Reset(attr.system_prompt, _token_ids);
+        update_cqdm(&cqdm, 0, "count", "tokenizer init ok");
+
+        if (!embed_selector.Init(attr.filename_tokens_embed, attr.tokens_embed_num, attr.tokens_embed_size,
+                                 attr.b_use_mmap_load_embed)) {
+            ALOGE("embed_selector.Init(%s, %d, %d) failed", attr.filename_tokens_embed.c_str(), attr.tokens_embed_num,
+                  attr.tokens_embed_size);
+            return false;
+        }
+        update_cqdm(&cqdm, 1, "count", "embed_selector init ok");
+
+        llama_layers.resize(attr.axmodel_num);
+
+        char axmodel_path[1024];
+        for (int i = 0; i < attr.axmodel_num; i++) {
+            sprintf(axmodel_path, attr.template_filename_axmodel.c_str(), i);
+            llama_layers[i].filename = axmodel_path;
+
+            int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), false);
+            if (ret != 0) {
+                ALOGE("init axmodel(%s) failed", llama_layers[i].filename.c_str());
+                return false;
+            }
+            int remain_cmm = get_remaining_cmm_size();
+            sprintf(axmodel_path, "init %d axmodel ok,remain_cmm(%d MB)", i, remain_cmm);
+            update_cqdm(&cqdm, i + 2, "count", axmodel_path);
+        }
+
+        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), false);
+        if (ret != 0) {
+            ALOGE("init post axmodel(%s) failed", attr.filename_post_axmodel.c_str());
+            return false;
+        }
+        int remain_cmm = get_remaining_cmm_size();
+        sprintf(axmodel_path, "init post axmodel ok,remain_cmm(%d MB)", remain_cmm);
+        update_cqdm(&cqdm, attr.axmodel_num + 2, "count", axmodel_path);
+
+        ret = image_encoder.init(attr.filename_image_encoder_axmodedl.c_str());
+        if (ret != 0) {
+            ALOGE("init vpm axmodel(%s) failed", attr.filename_image_encoder_axmodedl.c_str());
+            return false;
+        }
+
+        _attr.IMAGE_CONTEXT_TOKEN = tokenizer->GetImgContextID();
+        _attr.IMAGE_START_TOKEN   = tokenizer->GetImgStartID();
+
+        ALOGI("IMAGE_CONTEXT_TOKEN: %d, IMAGE_START_TOKEN: %d", _attr.IMAGE_CONTEXT_TOKEN, _attr.IMAGE_START_TOKEN);
+
+        _attr.IMAGE_ENCODER_INPUT_NCHW = -1;
+        for (size_t i = 1; i < image_encoder.get_input(0).vShape.size(); i++) {
+            if (image_encoder.get_input(0).vShape[i] == 3) {
+                if (i == 1) {
+                    _attr.IMAGE_ENCODER_INPUT_NCHW = 1;
+                } else if (i == 3) {
+                    _attr.IMAGE_ENCODER_INPUT_NCHW = 0;
+                }
+            }
+        }
+        if (_attr.IMAGE_ENCODER_INPUT_NCHW == -1) {
+            ALOGE("image encoder input nchw or nhwc not found");
+            return false;
+        }
+
+        if (_attr.IMAGE_ENCODER_INPUT_NCHW) {
+            ALOGI("image encoder input nchw@float32");
+            _attr.image_encoder_height = image_encoder.get_input(0).vShape[2];
+            _attr.image_encoder_width  = image_encoder.get_input(0).vShape[3];
+        } else {
+            ALOGI("image encoder input nhwc@uint8");
+            _attr.image_encoder_height = image_encoder.get_input(0).vShape[1];
+            _attr.image_encoder_width  = image_encoder.get_input(0).vShape[2];
+        }
+
+        if (_attr.image_encoder_height != _attr.image_encoder_width) {
+            ALOGE("image encoder height != width");
+            return false;
+        }
+        int output_elem_size = 1;
+        for (int i = 0; i < image_encoder.get_output(0).vShape.size(); i++) {
+            output_elem_size *= image_encoder.get_output(0).vShape[i];
+        }
+
+        if (output_elem_size * 2 == image_encoder.get_output(0).nSize) {
+            _attr.IMAGE_ENCODER_OUTPUT_BF16 = 1;
+            ALOGI("image encoder output bf16");
+        } else if (output_elem_size * 4 == image_encoder.get_output(0).nSize) {
+            _attr.IMAGE_ENCODER_OUTPUT_BF16 = 0;
+            ALOGI("image encoder output float32");
+        } else {
+            ALOGE("image encoder output not support");
+            return false;
+        }
+
+        printf("\n");
+
+        {
+            ALOGI("image_encoder_height : %d, image_encoder_width: %d", _attr.image_encoder_height,
+                  _attr.image_encoder_width);
+            _attr.max_token_len = llama_layers[0].layer.get_input("mask").nSize / sizeof(unsigned short) - 1;
+            ALOGI("max_token_len : %d", _attr.max_token_len);
+            _attr.kv_cache_size = llama_layers[0].layer.get_output("K_cache_out").nSize / sizeof(unsigned short);
+            _attr.kv_cache_num =
+                llama_layers[0].layer.get_input("K_cache").nSize / _attr.kv_cache_size / sizeof(unsigned short);
+            ALOGI("kv_cache_size : %d, kv_cache_num: %d", _attr.kv_cache_size, _attr.kv_cache_num);
+            if (_attr.max_token_len > _attr.kv_cache_num) {
+                ALOGE("max_token_len(%d) > kv_cache_num(%d)", _attr.max_token_len, _attr.kv_cache_num);
+                return false;
+            }
+
+            _attr.prefill_token_num = llama_layers[0].layer.get_input(1, "indices").vShape[1];
+            ALOGI("prefill_token_num : %d", _attr.prefill_token_num);
+            for (size_t i = 0; i < llama_layers[0].layer.get_num_input_groups() - 1; i++) {
+                int prefill_max_kv_cache_num = llama_layers[0].layer.get_input(i + 1, "K_cache").vShape[1];
+                ALOGI("grp: %d, prefill_max_token_num : %d", i + 1, prefill_max_kv_cache_num);
+                _attr.prefill_max_kv_cache_num_grp.push_back(prefill_max_kv_cache_num);
+            }
+            _attr.prefill_max_token_num =
+                _attr.prefill_max_kv_cache_num_grp[_attr.prefill_max_kv_cache_num_grp.size() - 1];
+            ALOGI("prefill_max_token_num : %d", _attr.prefill_max_token_num);
+        }
+
+        if (!postprocess.load_config(attr.post_config_path)) {
+            ALOGW("load postprocess config(%s) failed", attr.post_config_path.c_str());
+        }
+
+        ALOGI("LLM init ok");
+        return true;
+    }
+
+    LLMAttrType *getAttr()
+    {
+        return &_attr;
+    }
+
+    LLMPostprocess *getPostprocess()
+    {
+        return &postprocess;
+    }
+
+    void Deinit()
+    {
+        for (int i = 0; i < _attr.axmodel_num; i++) {
+            llama_layers[i].layer.release();
+        }
+        llama_post.release();
+        embed_selector.Deinit();
+    }
+
+    void Stop()
+    {
+        b_stop = true;
+    }
+
+    int SetSystemPrompt(std::string system_prompt, std::vector<int> &_token_ids)
+    {
+        tokenizer->Reset(system_prompt, _token_ids);
+        _attr.system_prompt         = system_prompt;
+        _attr.prefill_max_token_num = _attr.prefill_max_kv_cache_num_grp[_attr.prefill_max_kv_cache_num_grp.size() - 1];
+        return 0;
+    }
+
+    int GenerateKVCachePrefill(std::vector<int> &_token_ids, std::vector<std::vector<unsigned short>> &k_caches,
+                               std::vector<std::vector<unsigned short>> &v_caches, int &precompute_len)
+    {
+        bfloat16 bf16       = -65536.f;
+        int input_embed_num = _token_ids.size();
+        precompute_len      = _token_ids.size();
+
+        k_caches.resize(_attr.axmodel_num);
+        v_caches.resize(_attr.axmodel_num);
+        int prefill_split_num = ceil((double)input_embed_num / _attr.prefill_token_num);
+
+        int prefill_grpid = _attr.prefill_max_kv_cache_num_grp.size();
+
+        for (size_t i = 0; i < _attr.prefill_max_kv_cache_num_grp.size(); i++) {
+            if (input_embed_num <= _attr.prefill_max_kv_cache_num_grp[i]) {
+                prefill_grpid = i + 1;
+                break;
+            }
+        }
+        ALOGI("input token num : %d, prefill_split_num : %d prefill_grpid : %d", input_embed_num, prefill_split_num,
+              prefill_grpid);
+
+        for (size_t i = 0; i < _attr.axmodel_num; i++) {
+            memset((void *)llama_layers[i].layer.get_input(prefill_grpid, "K_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(prefill_grpid, "K_cache").nSize);
+            memset((void *)llama_layers[i].layer.get_input(prefill_grpid, "V_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(prefill_grpid, "V_cache").nSize);
+        }
+
+        if (input_embed_num == 0) {
+            for (size_t i = 0; i < _attr.axmodel_num; i++) {
+                k_caches[i].resize(precompute_len * _attr.kv_cache_size);
+                v_caches[i].resize(precompute_len * _attr.kv_cache_size);
+            }
+            ALOGI("input token num is 0, skip");
+            return 0;
+        }
+
+        int kv_cache_num = _attr.prefill_max_kv_cache_num_grp[prefill_grpid - 1];
+
+        std::vector<unsigned short> test_embed;
+        test_embed.resize(_token_ids.size() * _attr.tokens_embed_size);
+
+        for (size_t i = 0; i < _token_ids.size(); i++) {
+            embed_selector.getByIndex(_token_ids[i], test_embed.data() + i * _attr.tokens_embed_size);
+        }
+
+        for (size_t p = 0; p < prefill_split_num; p++) {
+            std::vector<unsigned short> mask_tmp;
+            mask_tmp.resize(1 * _attr.prefill_token_num * (kv_cache_num + _attr.prefill_token_num), bf16.data);
+            int input_num_token = _attr.prefill_token_num;
+            if (p == prefill_split_num - 1) {
+                input_num_token = input_embed_num - p * _attr.prefill_token_num;
+            }
+
+            ALOGI("input_num_token:%d", input_num_token);
+            for (size_t i = 0; i < _attr.prefill_token_num; i++) {
+                if (i < input_num_token) {
+                    int mask_current_start = kv_cache_num;
+                    auto mask_ptr          = mask_tmp.data() + i * (kv_cache_num + _attr.prefill_token_num);
+
+                    for (int j = 0; j < p * _attr.prefill_token_num; j++) {
+                        mask_ptr[j] = 0;
+                    }
+
+                    for (int j = mask_current_start; j < mask_current_start + i + 1; j++) {
+                        mask_ptr[j] = 0;
+                    }
+                }
+            }
+
+            std::vector<unsigned short> embed_tmp(_attr.prefill_token_num * _attr.tokens_embed_size, 0);
+            if (p == (prefill_split_num - 1)) {
+                memcpy(
+                    embed_tmp.data(), test_embed.data() + p * _attr.prefill_token_num * _attr.tokens_embed_size,
+                    (input_embed_num - p * _attr.prefill_token_num) * _attr.tokens_embed_size * sizeof(unsigned short));
+            } else {
+                memcpy(embed_tmp.data(), test_embed.data() + p * _attr.prefill_token_num * _attr.tokens_embed_size,
+                       _attr.prefill_token_num * _attr.tokens_embed_size * sizeof(unsigned short));
+            }
+
+            for (unsigned int m = 0; m < _attr.axmodel_num; m++) {
+                auto &layer = llama_layers[m];
+                // set indices
+                auto &input_indices             = layer.layer.get_input(prefill_grpid, "indices");
+                unsigned int *input_indices_ptr = (unsigned int *)input_indices.pVirAddr;
+                memset(input_indices_ptr, 0, input_indices.nSize);
+                int idx = 0;
+                for (unsigned int i = p * _attr.prefill_token_num; i < (p + 1) * _attr.prefill_token_num; i++) {
+                    input_indices_ptr[idx] = i;
+                    idx++;
+                }
+
+                // set mask
+                auto &input_mask = layer.layer.get_input(prefill_grpid, "mask");
+                memcpy((void *)input_mask.pVirAddr, (void *)mask_tmp.data(), mask_tmp.size() * sizeof(unsigned short));
+
+                auto &input_input = layer.layer.get_input(prefill_grpid, "input");
+                memcpy((void *)input_input.pVirAddr, embed_tmp.data(), embed_tmp.size() * sizeof(unsigned short));
+
+                layer.layer.inference(prefill_grpid);
+
+                auto &input_decoder_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
+                auto &input_decoder_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
+
+                auto &input_prefill_k_cache = layer.layer.get_input(prefill_grpid, "K_cache");
+                auto &input_prefill_v_cache = layer.layer.get_input(prefill_grpid, "V_cache");
+
+                auto &output_k_cache = layer.layer.get_output(prefill_grpid, "K_cache_out");
+                auto &output_v_cache = layer.layer.get_output(prefill_grpid, "V_cache_out");
+
+                int kv_offset = (p * _attr.prefill_token_num) * _attr.kv_cache_size;
+
+                memcpy((unsigned short *)input_decoder_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.prefill_token_num * _attr.kv_cache_size);
+
+                memcpy((unsigned short *)input_decoder_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.prefill_token_num * _attr.kv_cache_size);
+
+                memcpy((unsigned short *)input_prefill_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.prefill_token_num * _attr.kv_cache_size);
+
+                memcpy((unsigned short *)input_prefill_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.prefill_token_num * _attr.kv_cache_size);
+
+                auto &output = layer.layer.get_output(prefill_grpid, "output");
+                memcpy(embed_tmp.data(), (void *)output.pVirAddr, embed_tmp.size() * sizeof(unsigned short));
+            }
+        }
+
+        for (size_t i = 0; i < _attr.axmodel_num; i++) {
+            auto &layer = llama_layers[i];
+            k_caches[i].resize(precompute_len * _attr.kv_cache_size);
+            v_caches[i].resize(precompute_len * _attr.kv_cache_size);
+            auto &input_k_cache = layer.layer.get_input(prefill_grpid, "K_cache");
+            auto &input_v_cache = layer.layer.get_input(prefill_grpid, "V_cache");
+            memcpy((void *)k_caches[i].data(), (void *)input_k_cache.pVirAddr,
+                   precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+            memcpy((void *)v_caches[i].data(), (void *)input_v_cache.pVirAddr,
+                   precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+        }
+
+        return 0;
+    }
+
+    int GenerateKVCache(std::vector<int> &_token_ids)
+    {
+        for (size_t i = 0; i < _attr.axmodel_num; i++) {
+            memset((void *)llama_layers[i].layer.get_input(decode_grpid, "K_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(decode_grpid, "K_cache").nSize);
+            memset((void *)llama_layers[i].layer.get_input(decode_grpid, "V_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(decode_grpid, "V_cache").nSize);
+        }
+
+        bfloat16 bf16 = -65536.f;
+        std::vector<unsigned short> mask(_attr.kv_cache_num + 1, bf16.data);
+        mask[_attr.kv_cache_num] = 0;
+        std::vector<unsigned short> embed;
+
+        int next_token = _token_ids[0];
+
+        t_cqdm cqdm = create_cqdm(_token_ids.size(), 32);
+
+        for (unsigned int indices = 0; indices < _token_ids.size(); indices++) {
+            embed_selector.getByIndex(next_token, embed);
+
+            for (int m = 0; m < _attr.axmodel_num; m++) {
+                if (b_stop) {
+                    break;
+                }
+
+                auto &layer = llama_layers[m];
+
+                auto &input_k_cache               = layer.layer.get_input(decode_grpid, "K_cache");
+                unsigned short *input_k_cache_ptr = (unsigned short *)input_k_cache.pVirAddr;
+                auto &input_v_cache               = layer.layer.get_input(decode_grpid, "V_cache");
+                unsigned short *input_v_cache_ptr = (unsigned short *)input_v_cache.pVirAddr;
+
+                auto &input_indices = layer.layer.get_input(decode_grpid, "indices");
+                memcpy(input_indices.pVirAddr, &indices, sizeof(indices));
+
+                auto &input_mask = layer.layer.get_input(decode_grpid, "mask");
+                memcpy(input_mask.pVirAddr, mask.data(), mask.size() * sizeof(unsigned short));
+
+                auto &input_input = layer.layer.get_input(decode_grpid, "input");
+                memcpy(input_input.pVirAddr, embed.data(), embed.size() * sizeof(unsigned short));
+
+                layer.layer.inference(decode_grpid);
+
+                auto &output_k_cache = layer.layer.get_output(decode_grpid, "K_cache_out");
+                memcpy(input_k_cache_ptr + indices * _attr.kv_cache_size, output_k_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.kv_cache_size);
+
+                auto &output_v_cache = layer.layer.get_output(decode_grpid, "V_cache_out");
+                memcpy(input_v_cache_ptr + indices * _attr.kv_cache_size, output_v_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.kv_cache_size);
+
+                auto &output = layer.layer.get_output(decode_grpid, "output");
+                memcpy(embed.data(), output.pVirAddr, embed.size() * sizeof(unsigned short));
+            }
+            mask[indices] = 0;
+            next_token    = _token_ids[indices + 1];
+            update_cqdm(&cqdm, indices, "token", "");
+            // ALOGI("");
+        }
+        return 0;
+    }
+
+    int GetKVCache(std::vector<std::vector<unsigned short>> &k_caches,
+                   std::vector<std::vector<unsigned short>> &v_caches, int &precompute_len)
+    {
+        bfloat16 bf16 = -65536.f;
+        std::vector<unsigned short> mask(_attr.kv_cache_num + 1, bf16.data);
+        auto &input_mask = llama_layers[0].layer.get_input(decode_grpid, "mask");
+        memcpy(mask.data(), (void *)input_mask.pVirAddr, input_mask.nSize);
+        for (size_t i = 0; i < mask.size(); i++) {
+            if (mask[i] == bf16.data) {
+                precompute_len = i + 1;
+                break;
+            }
+        }
+        ALOGI("precompute_len:%d, remaining:%d", precompute_len,
+              _attr.prefill_max_kv_cache_num_grp[_attr.prefill_max_kv_cache_num_grp.size() - 1] - precompute_len);
+        k_caches.resize(_attr.axmodel_num);
+        v_caches.resize(_attr.axmodel_num);
+        for (size_t i = 0; i < _attr.axmodel_num; i++) {
+            auto &layer = llama_layers[i];
+            k_caches[i].resize(precompute_len * _attr.kv_cache_size);
+            v_caches[i].resize(precompute_len * _attr.kv_cache_size);
+            auto &input_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
+            auto &input_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
+            memcpy((void *)k_caches[i].data(), (void *)input_k_cache.pVirAddr,
+                   precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+            memcpy((void *)v_caches[i].data(), (void *)input_v_cache.pVirAddr,
+                   precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+        }
+
+        _attr.prefill_max_token_num = _attr.prefill_max_kv_cache_num_grp[_attr.prefill_max_kv_cache_num_grp.size() - 1];
+
+        return 0;
+    }
+
+    int SetKVCache(std::vector<std::vector<unsigned short>> &k_caches,
+                   std::vector<std::vector<unsigned short>> &v_caches, int precompute_len, int input_num_token)
+    {
+        _attr.precompute_len = precompute_len;
+        for (size_t i = 0; i < _attr.prefill_max_kv_cache_num_grp.size(); i++) {
+            if (_attr.precompute_len + input_num_token <= _attr.prefill_max_kv_cache_num_grp[i]) {
+                _attr.prefill_grpid = i + 1;
+                break;
+            }
+        }
+        int kv_cache_num = _attr.prefill_max_kv_cache_num_grp[_attr.prefill_grpid - 1];
+        ALOGI("prefill_grpid:%d kv_cache_num:%d precompute_len:%d input_num_token:%d", _attr.prefill_grpid,
+              kv_cache_num, precompute_len, input_num_token);
+
+        _attr.prefill_max_token_num =
+            ALIGN_DOWN(_attr.prefill_max_token_num - _attr.precompute_len, _attr.prefill_token_num);
+        ALOGI("current prefill_max_token_num:%d", _attr.prefill_max_token_num);
+
+        if (precompute_len == 0) {
+            ALOGI("first run");
+            return 0;
+        }
+
+        if (precompute_len + input_num_token > kv_cache_num) {
+            ALOGE("precompute_len(%d) + input_num_token(%d) > _attr.prefill_max_kv_cache_num_grp[%d]", precompute_len,
+                  input_num_token, _attr.prefill_grpid - 1);
+            return -1;
+        }
+
+        if (input_num_token > _attr.prefill_max_token_num) {
+            ALOGE("input_num_token(%d) > _attr.prefill_max_token_num(%d)", input_num_token,
+                  _attr.prefill_max_token_num);
+            return -1;
+        }
+
+        if (k_caches.size() != v_caches.size()) {
+            ALOGE("k_caches.size(%d) != v_caches.size(%d)", k_caches.size(), v_caches.size());
+            return -1;
+        }
+
+        if (k_caches.size() != _attr.axmodel_num) {
+            ALOGE("k_caches.size(%d) != _attr.axmodel_num(%d)", k_caches.size(), _attr.axmodel_num);
+            return -1;
+        }
+
+        // clear kv cache
+        for (size_t i = 0; i < _attr.axmodel_num; i++) {
+            memset((void *)llama_layers[i].layer.get_input(_attr.prefill_grpid, "K_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(_attr.prefill_grpid, "K_cache").nSize);
+            memset((void *)llama_layers[i].layer.get_input(_attr.prefill_grpid, "V_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(_attr.prefill_grpid, "V_cache").nSize);
+
+            memset((void *)llama_layers[i].layer.get_input(decode_grpid, "K_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(decode_grpid, "K_cache").nSize);
+            memset((void *)llama_layers[i].layer.get_input(decode_grpid, "V_cache").pVirAddr, 0,
+                   llama_layers[i].layer.get_input(decode_grpid, "V_cache").nSize);
+        }
+
+        for (unsigned int m = 0; m < _attr.axmodel_num; m++) {
+            auto &layer = llama_layers[m];
+
+            auto &k_cache = k_caches[m];
+            auto &v_cache = v_caches[m];
+
+            if (k_cache.size() != _attr.precompute_len * _attr.kv_cache_size) {
+                ALOGE("k_cache.size(%d) != precompute_len(%d) * _attr.kv_cache_size(%d)", k_cache.size(),
+                      _attr.precompute_len, _attr.kv_cache_size);
+                return -1;
+            }
+            if (v_cache.size() < _attr.precompute_len * _attr.kv_cache_size) {
+                ALOGE("v_cache.size(%d) < precompute_len(%d) * _attr.kv_cache_size(%d)", v_cache.size(),
+                      _attr.precompute_len, _attr.kv_cache_size);
+                return -1;
+            }
+
+            // set kv cache inputs
+            {
+                auto &input_k_cache               = layer.layer.get_input(_attr.prefill_grpid, "K_cache");
+                unsigned short *input_k_cache_ptr = (unsigned short *)input_k_cache.pVirAddr;
+                auto &input_v_cache               = layer.layer.get_input(_attr.prefill_grpid, "V_cache");
+                unsigned short *input_v_cache_ptr = (unsigned short *)input_v_cache.pVirAddr;
+
+                memcpy(input_k_cache_ptr, k_cache.data(),
+                       _attr.precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+                memcpy(input_v_cache_ptr, v_cache.data(),
+                       _attr.precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+            }
+
+            {
+                auto &input_k_cache               = layer.layer.get_input(decode_grpid, "K_cache");
+                unsigned short *input_k_cache_ptr = (unsigned short *)input_k_cache.pVirAddr;
+                auto &input_v_cache               = layer.layer.get_input(decode_grpid, "V_cache");
+                unsigned short *input_v_cache_ptr = (unsigned short *)input_v_cache.pVirAddr;
+
+                memcpy(input_k_cache_ptr, k_cache.data(),
+                       _attr.precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+                memcpy(input_v_cache_ptr, v_cache.data(),
+                       _attr.precompute_len * _attr.kv_cache_size * sizeof(unsigned short));
+            }
+        }
+
+        return 0;
+    }
+
+    bool save_kvcache(std::string target_path, std::string system_prompt, int precompute_len,
+                      std::vector<std::vector<unsigned short>> &k_caches,
+                      std::vector<std::vector<unsigned short>> &v_caches)
+    {
+        for (size_t i = 0; i < k_caches.size(); i++) {
+            std::string k_cache_path = target_path + "/k_cache_" + std::to_string(i) + ".bin";
+            std::string v_cache_path = target_path + "/v_cache_" + std::to_string(i) + ".bin";
+            std::ofstream k_cache_file(k_cache_path);
+            std::ofstream v_cache_file(v_cache_path);
+            if (!k_cache_file.is_open() || !v_cache_file.is_open()) {
+                ALOGE("save kvcache failed");
+                return false;
+            }
+            k_cache_file.write((char *)k_caches[i].data(), k_caches[i].size() * sizeof(unsigned short));
+            v_cache_file.write((char *)v_caches[i].data(), v_caches[i].size() * sizeof(unsigned short));
+            k_cache_file.close();
+            v_cache_file.close();
+        }
+        nlohmann::json j;
+        j["system_prompt"]      = system_prompt;
+        j["precompute_len"]     = precompute_len;
+        std::string config_path = target_path + "/config.json";
+        std::ofstream config_file(config_path);
+        config_file << j.dump();
+        config_file.close();
+        return true;
+    }
+
+    bool load_kvcache(std::string target_path, int axmodel_num, std::vector<std::vector<unsigned short>> &k_caches,
+                      std::vector<std::vector<unsigned short>> &v_caches, std::string &system_prompt,
+                      int &precompute_len)
+    {
+        k_caches.resize(axmodel_num);
+        v_caches.resize(axmodel_num);
+        for (size_t i = 0; i < k_caches.size(); i++) {
+            std::string k_cache_path = target_path + "/k_cache_" + std::to_string(i) + ".bin";
+            std::string v_cache_path = target_path + "/v_cache_" + std::to_string(i) + ".bin";
+            if (file_exist(k_cache_path) && file_exist(v_cache_path)) {
+                std::vector<unsigned short> k_cache;
+                std::vector<unsigned short> v_cache;
+                std::ifstream k_cache_file(k_cache_path);
+                std::ifstream v_cache_file(v_cache_path);
+
+                k_cache_file.seekg(0, std::ios::end);
+                k_cache.resize(k_cache_file.tellg() / sizeof(unsigned short));
+                k_cache_file.seekg(0, std::ios::beg);
+
+                v_cache_file.seekg(0, std::ios::end);
+                v_cache.resize(v_cache_file.tellg() / sizeof(unsigned short));
+                v_cache_file.seekg(0, std::ios::beg);
+
+                k_cache_file.read((char *)k_cache.data(), k_cache.size() * sizeof(unsigned short));
+                v_cache_file.read((char *)v_cache.data(), v_cache.size() * sizeof(unsigned short));
+
+                k_cache_file.close();
+                v_cache_file.close();
+                k_caches[i] = k_cache;
+                v_caches[i] = v_cache;
+            } else {
+                ALOGE("k_cache %s or v_cache %s not exist", k_cache_path.c_str(), v_cache_path.c_str());
+                return false;
+            }
+        }
+
+        std::string config_path = target_path + "/config.json";
+        if (file_exist(config_path)) {
+            std::ifstream config_file(config_path);
+            nlohmann::json j;
+            config_file >> j;
+            system_prompt  = j["system_prompt"].get<std::string>();
+            precompute_len = j["precompute_len"].get<int>();
+            config_file.close();
+        } else {
+            ALOGE("config %s not exist", config_path.c_str());
+            return false;
+        }
+        return true;
+    }
+
+    int Encode(cv::Mat src, std::vector<unsigned short> &out_embed)
+    {
+        timer t;
+        t.start();
+        if (_attr.IMAGE_ENCODER_INPUT_NCHW) {
+            std::vector<float> mean  = {0.485, 0.456, 0.406};
+            std::vector<float> scale = {0.229, 0.224, 0.225};
+
+            cv::Mat dst;
+            cv::resize(src, dst, cv::Size(_attr.image_encoder_width, _attr.image_encoder_height));
+            cv::cvtColor(dst, dst, cv::COLOR_BGR2RGB);
+
+            float *input_data = (float *)image_encoder.get_input(0).pVirAddr;
+
+            unsigned char *img_data = dst.data;
+            int letterbox_rows      = dst.rows;
+            int letterbox_cols      = dst.cols;
+
+            for (int h = 0; h < letterbox_rows; h++) {
+                for (int w = 0; w < letterbox_cols; w++) {
+                    for (int c = 0; c < 3; c++) {
+                        int in_index          = h * letterbox_cols * 3 + w * 3 + c;
+                        int out_index         = c * letterbox_rows * letterbox_cols + h * letterbox_cols + w;
+                        input_data[out_index] = (float(img_data[in_index]) / 255.0 - mean[c]) / scale[c];
+                    }
+                }
+            }
+            image_encoder.inference();
+        } else {
+            cv::Mat dst;
+            cv::resize(src, dst, cv::Size(_attr.image_encoder_width, _attr.image_encoder_height));
+            cv::cvtColor(dst, dst, cv::COLOR_BGR2RGB);
+            void *data = image_encoder.get_input(0).pVirAddr;
+            memcpy(data, dst.data, dst.rows * dst.cols * 3);
+            image_encoder.inference();
+        }
+
+        int size = 1;
+        for (size_t i = 0; i < image_encoder.get_output(0).vShape.size(); i++) {
+            size *= image_encoder.get_output(0).vShape[i];
+        }
+
+        out_embed.resize(size);
+
+        if (_attr.IMAGE_ENCODER_OUTPUT_BF16)
+            memcpy(out_embed.data(), image_encoder.get_output(0).pVirAddr, image_encoder.get_output(0).nSize);
+        else {
+            float *out_data = (float *)image_encoder.get_output(0).pVirAddr;
+            for (size_t i = 0; i < size; i++) {
+                out_embed[i] = bfloat16(out_data[i]).data;
+            }
+        }
+
+        ALOGI("image encode time : %0.2f ms, size : %ld", t.cost(), out_embed.size());
+        return 0;
+    }
+
+    int Encode(std::vector<cv::Mat> srcs, std::vector<std::vector<unsigned short>> &out_embeds)
+    {
+        out_embeds.resize(srcs.size());
+        for (size_t i = 0; i < srcs.size(); i++) {
+            auto ret = Encode(srcs[i], out_embeds[i]);
+            if (ret != 0) {
+                ALOGE("Encode image failed");
+                return -1;
+            }
+        }
+
+        return 0;
+    }
+
+    int Encode(std::vector<unsigned short> &out_embed, std::string prompt = "What is in the image?")
+    {
+        ImageInfo img_info;
+        img_info.img_prompt        = false;
+        std::vector<int> input_ids = tokenizer->Encode(prompt, img_info);
+        if (input_ids.size() > _attr.prefill_token_num) {
+            ALOGE("input_ids(%ld) > prefill_token_num(%d)", input_ids.size(), _attr.prefill_token_num);
+            return -1;
+        }
+        out_embed.resize(input_ids.size() * _attr.tokens_embed_size);
+
+        for (size_t i = 0; i < input_ids.size(); i++) {
+            embed_selector.getByIndex(input_ids[i], out_embed.data() + i * _attr.tokens_embed_size);
+        }
+
+        return 0;
+    }
+
+    int Encode(std::vector<std::vector<unsigned short>> &imgs_embed, std::vector<unsigned short> &out_embed,
+               std::string prompt, std::vector<int> &tokens_ids, std::vector<int> &tokens_diff)
+    {
+        ImageInfo img_info;
+        img_info.img_prompt        = true;
+        img_info.num_img           = imgs_embed.size();
+        img_info.imgsz             = _attr.image_encoder_width;
+        std::vector<int> input_ids = tokenizer->Encode_ctx(prompt, img_info, tokens_ids, tokens_diff);
+
+        std::vector<int> img_start_index;
+        for (size_t i = 0; i < input_ids.size(); i++) {
+            if (input_ids[i] == _attr.IMAGE_START_TOKEN) {
+                img_start_index.push_back(i);
+            }
+        }
+
+        if (img_start_index.size() != imgs_embed.size()) {
+            ALOGE("img_start_index.size() != imgs_embed.size(), img_start_index.size() : %ld, imgs_embed.size() : %ld",
+                  img_start_index.size(), imgs_embed.size());
+
+            printf("input_ids : ");
+            for (size_t i = 0; i < input_ids.size(); i++) {
+                printf("%d ", input_ids[i]);
+            }
+            printf("\n");
+
+            return -1;
+        }
+
+        if (input_ids.size() > _attr.prefill_max_token_num) {
+            ALOGE("input_ids(%ld) > prefill_max_token_num(%d)", input_ids.size(), _attr.prefill_max_token_num);
+            return -1;
+        }
+        out_embed.resize(input_ids.size() * _attr.tokens_embed_size);
+
+        for (size_t i = 0; i < input_ids.size(); i++) {
+            embed_selector.getByIndex(input_ids[i], out_embed.data() + i * _attr.tokens_embed_size);
+        }
+        for (size_t i = 0; i < imgs_embed.size(); i++) {
+            int offset      = img_start_index[i] + 1;
+            auto &img_embed = imgs_embed[i];
+
+            int img_context_count = 0;
+            for (size_t j = offset; j < input_ids.size(); j++) {
+                if (input_ids[j] == _attr.IMAGE_CONTEXT_TOKEN) {
+                    img_context_count++;
+                } else {
+                    break;
+                }
+            }
+
+            if (img_context_count != img_embed.size() / _attr.tokens_embed_size) {
+                ALOGE("img_context_count(%d) != img_embed.size() / tokens_embed_size(%ld)", img_context_count,
+                      img_embed.size() / _attr.tokens_embed_size);
+                return -1;
+            }
+
+            memcpy(out_embed.data() + offset * _attr.tokens_embed_size, img_embed.data(),
+                   img_embed.size() * sizeof(unsigned short));
+            ALOGI("idx:%ld offset : %d out_embed.size() : %ld", i, offset, out_embed.size());
+        }
+
+        return 0;
+    }
+
+    int Encode(std::vector<unsigned short> &img_embed, std::vector<unsigned short> &out_embed, std::string prompt,
+               std::vector<int> &tokens_ids, std::vector<int> &tokens_diff)
+    {
+        std::vector<std::vector<unsigned short>> imgs_embed = {img_embed};
+        return Encode(imgs_embed, out_embed, prompt, tokens_ids, tokens_diff);
+    }
+
+    int Encode(std::vector<unsigned short> &out_embed, std::string prompt, std::string last_reply,
+               std::vector<int> &tokens_ids, std::vector<int> &tokens_diff)
+    {
+        ImageInfo img_info;
+        img_info.img_prompt = false;
+        if (!tokenizer->Encode(prompt, last_reply, tokens_ids, tokens_diff, img_info)) {
+            ALOGE("encode failed");
+            return -1;
+        }
+
+        out_embed.resize(tokens_diff.size() * _attr.tokens_embed_size);
+
+        for (size_t i = 0; i < tokens_diff.size(); i++) {
+            embed_selector.getByIndex(tokens_diff[i], out_embed.data() + i * _attr.tokens_embed_size);
+        }
+
+        return 0;
+    }
+
+    std::string Run(std::vector<unsigned short> test_embed)
+    {
+        b_stop = false;
+        std::string final_out;
+
+        bfloat16 bf16 = -65536.f;
+        std::vector<unsigned short> mask(_attr.kv_cache_num + 1, bf16.data);
+        std::vector<unsigned short> embed(_attr.tokens_embed_size, 0);
+        int kv_cache_num = _attr.prefill_max_kv_cache_num_grp[_attr.prefill_grpid - 1];
+
+        std::vector<int> cached_token;
+        std::vector<int> token_ids;
+
+        int input_embed_num   = test_embed.size() / _attr.tokens_embed_size;
+        int prefill_split_num = ceil((double)input_embed_num / _attr.prefill_token_num);
+        ALOGI("input token num : %d, prefill_split_num : %d", input_embed_num, prefill_split_num);
+
+        mask[_attr.kv_cache_num] = 0;
+        for (size_t i = 0; i < _attr.precompute_len + input_embed_num; i++) {
+            mask[i] = 0;
+        }
+        timer t_cost;
+        timer ttft_timer;
+        ttft_timer.start();
+
+        for (size_t p = 0; p < prefill_split_num; p++) {
+            if (b_stop) {
+                break;
+            }
+
+            std::vector<unsigned short> mask_tmp;
+            mask_tmp.resize(1 * _attr.prefill_token_num * (kv_cache_num + _attr.prefill_token_num), bf16.data);
+            int input_num_token = _attr.prefill_token_num;
+            if (p == prefill_split_num - 1) {
+                input_num_token = input_embed_num - p * _attr.prefill_token_num;
+            }
+
+            ALOGI("input_num_token:%d", input_num_token);
+            for (size_t i = 0; i < _attr.prefill_token_num; i++) {
+                if (i < input_num_token) {
+                    int mask_current_start = kv_cache_num;
+                    auto mask_ptr          = mask_tmp.data() + i * (kv_cache_num + _attr.prefill_token_num);
+
+                    for (int j = 0; j < _attr.precompute_len + p * _attr.prefill_token_num; j++) {
+                        mask_ptr[j] = 0;
+                    }
+
+                    for (int j = mask_current_start; j < mask_current_start + i + 1; j++) {
+                        mask_ptr[j] = 0;
+                    }
+                }
+            }
+
+            std::vector<unsigned short> embed_tmp(_attr.prefill_token_num * _attr.tokens_embed_size, 0);
+            if (p == (prefill_split_num - 1)) {
+                memcpy(
+                    embed_tmp.data(), test_embed.data() + p * _attr.prefill_token_num * _attr.tokens_embed_size,
+                    (input_embed_num - p * _attr.prefill_token_num) * _attr.tokens_embed_size * sizeof(unsigned short));
+            } else {
+                memcpy(embed_tmp.data(), test_embed.data() + p * _attr.prefill_token_num * _attr.tokens_embed_size,
+                       _attr.prefill_token_num * _attr.tokens_embed_size * sizeof(unsigned short));
+            }
+
+            for (unsigned int m = 0; m < _attr.axmodel_num; m++) {
+                if (b_stop) {
+                    break;
+                }
+
+                auto &layer = llama_layers[m];
+
+                // set indices
+                auto &input_indices             = layer.layer.get_input(_attr.prefill_grpid, "indices");
+                unsigned int *input_indices_ptr = (unsigned int *)input_indices.pVirAddr;
+                memset(input_indices_ptr, 0, input_indices.nSize);
+                int idx = 0;
+                for (unsigned int i = _attr.precompute_len + p * _attr.prefill_token_num;
+                     i < _attr.precompute_len + (p + 1) * _attr.prefill_token_num; i++) {
+                    input_indices_ptr[idx] = i;
+                    idx++;
+                }
+
+                // set mask
+                auto &input_mask = layer.layer.get_input(_attr.prefill_grpid, "mask");
+                memcpy((void *)input_mask.pVirAddr, (void *)mask_tmp.data(), mask_tmp.size() * sizeof(unsigned short));
+
+                // set input
+                auto &input_input = layer.layer.get_input(_attr.prefill_grpid, "input");
+                memcpy((void *)input_input.pVirAddr, embed_tmp.data(), embed_tmp.size() * sizeof(unsigned short));
+
+                layer.layer.inference(_attr.prefill_grpid);
+
+                auto &input_decoder_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
+                auto &input_decoder_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
+
+                auto &input_prefill_k_cache = layer.layer.get_input(_attr.prefill_grpid, "K_cache");
+                auto &input_prefill_v_cache = layer.layer.get_input(_attr.prefill_grpid, "V_cache");
+
+                auto &output_k_cache = layer.layer.get_output(_attr.prefill_grpid, "K_cache_out");
+                auto &output_v_cache = layer.layer.get_output(_attr.prefill_grpid, "V_cache_out");
+
+                int kv_offset = (_attr.precompute_len + p * _attr.prefill_token_num) * _attr.kv_cache_size;
+
+                memcpy((unsigned short *)input_decoder_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
+                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+
+                memcpy((unsigned short *)input_decoder_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
+                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+
+                memcpy((unsigned short *)input_prefill_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
+                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+
+                memcpy((unsigned short *)input_prefill_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
+                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+
+                auto &output = layer.layer.get_output(_attr.prefill_grpid, "output");
+                memcpy(embed_tmp.data(), (void *)output.pVirAddr, embed_tmp.size() * sizeof(unsigned short));
+            }
+            if (p == (prefill_split_num - 1)) {
+                memcpy(embed.data(),
+                       embed_tmp.data() + (input_embed_num - p * _attr.prefill_token_num - 1) * _attr.tokens_embed_size,
+                       _attr.tokens_embed_size * sizeof(unsigned short));
+            }
+        }
+
+        int next_token = -1;
+        t_cqdm cqdm    = create_cqdm(_attr.max_token_len, 32);
+
+        {
+            // post process
+            auto &input = llama_post.get_input("input");
+            memcpy(input.pVirAddr, embed.data(), embed.size() * sizeof(unsigned short));
+            llama_post.inference();
+            int max_index;
+
+            auto &output_post = llama_post.get_output("output");
+            // AX_SYS_MinvalidateCache(output_post.phyAddr, output_post.pVirAddr, output_post.nSize);
+            unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
+            float max_val            = -MAXFLOAT;
+            // max_index = FindMax(post_out, _attr.tokens_embed_num, &max_val);
+            max_index = post_process(postprocess, post_out, _attr.tokens_embed_num, token_ids, nullptr);
+
+            next_token = max_index;
+
+            token_ids.push_back(max_index);
+            cached_token.push_back(max_index);
+            ALOGI("ttft: %.2f ms", ttft_timer.cost());
+        }
+        t_cost.start();
+
+        bool b_hit_eos = false;
+        for (unsigned int indices = _attr.precompute_len + input_embed_num; indices < _attr.max_token_len; indices++) {
+            if (b_stop) {
+                break;
+            }
+
+            embed_selector.getByIndex(next_token, embed);
+
+            for (int m = 0; m < _attr.axmodel_num; m++) {
+                if (b_stop) {
+                    break;
+                }
+
+                auto &layer = llama_layers[m];
+
+                auto &input_k_cache               = layer.layer.get_input(decode_grpid, "K_cache");
+                unsigned short *input_k_cache_ptr = (unsigned short *)input_k_cache.pVirAddr;
+                auto &input_v_cache               = layer.layer.get_input(decode_grpid, "V_cache");
+                unsigned short *input_v_cache_ptr = (unsigned short *)input_v_cache.pVirAddr;
+
+                auto &input_indices = layer.layer.get_input(decode_grpid, "indices");
+                memcpy(input_indices.pVirAddr, &indices, sizeof(indices));
+
+                auto &input_mask = layer.layer.get_input(decode_grpid, "mask");
+                memcpy(input_mask.pVirAddr, mask.data(), mask.size() * sizeof(unsigned short));
+
+                auto &input_input = layer.layer.get_input(decode_grpid, "input");
+                memcpy(input_input.pVirAddr, embed.data(), embed.size() * sizeof(unsigned short));
+
+                layer.layer.inference(decode_grpid);
+
+                auto &output_k_cache = layer.layer.get_output(decode_grpid, "K_cache_out");
+                memcpy(input_k_cache_ptr + indices * _attr.kv_cache_size, output_k_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.kv_cache_size);
+
+                auto &output_v_cache = layer.layer.get_output(decode_grpid, "V_cache_out");
+                memcpy(input_v_cache_ptr + indices * _attr.kv_cache_size, output_v_cache.pVirAddr,
+                       sizeof(unsigned short) * _attr.kv_cache_size);
+
+                auto &output = layer.layer.get_output(decode_grpid, "output");
+                memcpy(embed.data(), output.pVirAddr, embed.size() * sizeof(unsigned short));
+            }
+            mask[indices] = 0;
+            {
+                // post process
+                auto &input = llama_post.get_input("input");
+                memcpy(input.pVirAddr, embed.data(), embed.size() * sizeof(unsigned short));
+                llama_post.inference();
+                int max_index;
+
+                auto &output_post = llama_post.get_output("output");
+                // AX_SYS_MinvalidateCache(output_post.phyAddr, output_post.pVirAddr, output_post.nSize);
+                unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
+                float max_val            = -MAXFLOAT;
+                // max_index = FindMax(post_out, _attr.tokens_embed_num, &max_val);
+                max_index = post_process(postprocess, post_out, _attr.tokens_embed_num, token_ids, nullptr);
+
+                next_token = max_index;
+
+                if (tokenizer->isEnd(max_index)) {
+                    if (cached_token.size() && _attr.runing_callback) {
+                        float t_cost_ms     = t_cost.cost();
+                        float token_per_sec = token_ids.size() / (t_cost_ms / 1000);
+                        auto tmp_out        = tokenizer->Decode(cached_token);
+                        _attr.runing_callback(cached_token.data(), cached_token.size(), tmp_out.c_str(), token_per_sec,
+                                              _attr.reserve);
+                        cached_token.clear();
+                    }
+                    b_hit_eos = true;
+                    break;
+                }
+                token_ids.push_back(max_index);
+
+                if (_attr.runing_callback) {
+                    cached_token.push_back(max_index);
+                    if (cached_token.size() >= 3) {
+                        float t_cost_ms     = t_cost.cost();
+                        float token_per_sec = token_ids.size() / (t_cost_ms / 1000);
+                        auto tmp_out        = tokenizer->Decode(cached_token);
+                        _attr.runing_callback(cached_token.data(), cached_token.size(), tmp_out.c_str(), token_per_sec,
+                                              _attr.reserve);
+                        cached_token.clear();
+                    }
+                }
+            }
+
+            if (_attr.runing_callback == nullptr) update_cqdm(&cqdm, indices, "token", "");
+            if (b_hit_eos) {
+                break;
+            }
+        }
+        printf("\n\n");
+        fflush(stdout);
+        float t_cost_ms = t_cost.cost();
+        ALOGN("hit eos,avg %.2f token/s\n", token_ids.size() / (t_cost_ms / 1000));
+
+        final_out = tokenizer->Decode(token_ids);
+
+        return final_out;
+    }
+};
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/Tokenizer/QwenTokenizer.cpp b/projects/llm_framework/main_vlm/src/runner/Tokenizer/QwenTokenizer.cpp
deleted file mode 100644
index 3b723ea1..00000000
--- a/projects/llm_framework/main_vlm/src/runner/Tokenizer/QwenTokenizer.cpp
+++ /dev/null
@@ -1,127 +0,0 @@
-#include "QwenTokenizer.hpp"
-
-#include <sstream>
-#include <fstream>
-#include <unordered_map>
-
-#include "sample_log.h"
-#include "base64.h"
-
-static const std::string PAT_STR = R"((?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?:$|[^\S])|\s+)";
-
-static std::pair<std::string, int> _parse(const std::string &line)
-{
-    auto pos = line.find(" ");
-    if (pos == std::string::npos)
-    {
-        throw std::runtime_error("invalid encoder line: " + line);
-    }
-
-    auto token = base64::decode({line.data(), pos});
-    int rank = 0;
-    try
-    {
-        rank = std::stoul(line.substr(pos + 1));
-    }
-    catch (const std::exception &)
-    {
-        throw std::runtime_error("invalid encoder rank: " + line);
-    }
-
-    return {std::move(token), rank};
-}
-
-QwenTokenizer::QwenTokenizer(const std::string &tiktoken_path, const QwenConfig &config)
-{
-    std::ifstream file(tiktoken_path);
-    if (!file)
-    {
-        throw std::runtime_error("failed to open encoder file: " + tiktoken_path);
-    }
-
-    ankerl::unordered_dense::map<std::string, int> encoder;
-    std::string line;
-    while (std::getline(file, line))
-    {
-        auto [token, rank] = _parse(line);
-
-        if (!encoder.emplace(std::move(token), rank).second)
-        {
-            throw std::runtime_error("duplicate item: " + line);
-        }
-    }
-
-    std::vector<std::string> special_tokens_s{"<|endoftext|>", "<|im_start|>", "<|im_end|>"};
-    char buffer[14];
-    for (size_t i = 0; i < 205; i++)
-    {
-        snprintf(buffer, 14, "<|extra_%zu|>", i);
-        special_tokens_s.push_back(buffer);
-    }
-    size_t encoder_size = encoder.size();
-    ankerl::unordered_dense::map<std::string, int> special_tokens;
-    special_tokens.reserve(special_tokens_s.size());
-    for (size_t i = 0; i < special_tokens_s.size(); i++)
-    {
-        special_tokens[special_tokens_s[i]] = encoder_size + i;
-    }
-
-    tokenizer = tiktoken::tiktoken(std::move(encoder), special_tokens, PAT_STR);
-    eos_token_id = config.eos_token_id;
-    im_start_id = config.im_start_id;
-    im_end_id = config.im_end_id;
-}
-
-auto QwenTokenizer::build_prompt(const std::vector<std::string> &history) const -> std::string
-{
-    if (history.size() % 2 == 1)
-    {
-        ALOGE("invalid history size %d", history.size());
-        return "";
-    }
-
-    std::ostringstream oss_prompt;
-    oss_prompt << "<|im_start|>system\nYou are a helpful assistant.<|im_end|>";
-    for (size_t i = 0; i < history.size() - 1; i += 2)
-    {
-        oss_prompt << "\n<|im_start|>user\n"
-                   << history[i] << "<|im_end|>\n<|im_start|>" << history[i + 1] << "<|im_end|>";
-    }
-    oss_prompt << "\n<|im_start|>user\n"
-               << history.back() << "<|im_end|>\n<|im_start|>assistant\n";
-
-    return oss_prompt.str();
-}
-
-auto QwenTokenizer::encode(const std::string &text, int max_length) const -> std::vector<int>
-{
-    auto ids = tokenizer.encode(text);
-    if ((int)ids.size() > max_length)
-    {
-        ids.erase(ids.begin(), ids.end() - max_length);
-    }
-    return ids;
-}
-
-auto QwenTokenizer::decode(const std::vector<int> &ids) const -> std::string
-{
-    std::vector<int> normal_ids(ids);
-    normal_ids.erase(std::remove_if(normal_ids.begin(), normal_ids.end(), [this](int id)
-                                    { return is_special_id(id); }),
-                     normal_ids.end());
-    auto text = tokenizer.decode(normal_ids);
-    return text;
-}
-
-auto QwenTokenizer::encode_history(
-    const std::vector<std::string> &history, int max_length) const -> std::vector<int>
-{
-    std::string prompt = build_prompt(history);
-    std::vector<int> input_ids = encode(prompt, max_length);
-    return input_ids;
-}
-
-auto QwenTokenizer::is_special_id(int id) const -> bool
-{
-    return id == eos_token_id || id == im_start_id || id == im_end_id;
-}
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/Tokenizer/QwenTokenizer.hpp b/projects/llm_framework/main_vlm/src/runner/Tokenizer/QwenTokenizer.hpp
deleted file mode 100644
index a217f9c2..00000000
--- a/projects/llm_framework/main_vlm/src/runner/Tokenizer/QwenTokenizer.hpp
+++ /dev/null
@@ -1,34 +0,0 @@
-#pragma once
-#include <string>
-#include <vector>
-#include "tiktoken.h"
-
-struct QwenConfig
-{
-    // int pad_token_id;
-    // for tokenizer
-    int eos_token_id = 151643;
-    int im_start_id = 151644;
-    int im_end_id = 151645;
-};
-
-class QwenTokenizer
-{
-public:
-    QwenTokenizer(const std::string &tiktoken_path, const QwenConfig &config);
-
-    auto encode(const std::string &text, int max_length) const -> std::vector<int>;
-
-    auto decode(const std::vector<int> &ids) const -> std::string;
-
-    auto encode_history(const std::vector<std::string> &history, int max_length) const -> std::vector<int>;
-
-    auto build_prompt(const std::vector<std::string> &history) const -> std::string;
-
-    auto is_special_id(int id) const -> bool;
-
-    tiktoken::tiktoken tokenizer;
-    int eos_token_id;
-    int im_start_id;
-    int im_end_id;
-};
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp b/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp
index 56600968..38974cb6 100644
--- a/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp
+++ b/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp
@@ -1,327 +1,242 @@
 #include "Tokenizer.hpp"
 
-#include "sentencepiece_processor.h"
-#include "builtin_pb/sentencepiece.pb.h"
-
-#include "QwenTokenizer.hpp"
-
-// #include "chatglm.h"
-
 #include "httplib.h"
+#include "http_utils.hpp"
 #include "json.hpp"
 
 #include "sample_log.h"
 #include "string_utility.hpp"
 #include "memory_utils.hpp"
 
-class TokenizerLLaMa : public BaseTokenizer
+class Tokenizer_Http : public BaseTokenizer
 {
-protected:
-    sentencepiece::SentencePieceProcessor sp;
+    std::shared_ptr<httplib::Client> cli;
     bool _b_bos, _b_eos;
 
+    std::string base_url;
+
+    int bos_id, eos_id;
+
+    std::string uid;
+    int img_start_token, img_context_token;
+
 private:
     /* data */
 public:
-    bool Init(std::string model_path, bool b_bos = true, bool b_eos = false) override
+    bool Init(std::string model_path) override
     {
-        auto ret = sp.Load(model_path);
-        if (!ret.ok())
-        {
-            ALOGE("%s", ret.error_message());
-            return false;
-        }
-
-        this->_b_bos = b_bos;
-        this->_b_eos = b_eos;
-        return ret.ok();
-    }
-
-    bool Encode(std::string input, std::vector<int> &output, bool b_img_prompt = false) override
-    {
-        auto ret = sp.Encode(input, &output);
-        if (!ret.ok())
+        base_url = model_path;
+        if (!test_connect_http(base_url, 10))
         {
-            ALOGE("%s", ret.error_message());
+            ALOGE("connect %s failed", base_url.c_str());
             return false;
         }
-        if (_b_bos)
-        {
-            output.insert(output.begin(), sp.bos_id());
-        }
-        if (_b_eos)
-        {
-            output.push_back(sp.eos_id());
-        }
-        return true;
-    }
-
-    std::vector<int> Encode(std::string input, bool b_img_prompt = false) override
-    {
-        std::vector<int> output;
-        Encode(input, output, b_img_prompt);
-        return output;
-    }
-
-    std::string Decode(const std::vector<int> input) override
-    {
-        sentencepiece::SentencePieceText spt;
-        sp.Decode(input, &spt);
-        std::string out = spt.pieces()[0].piece();
-        if (*(unsigned short *)out.data() == 38626)
-        {
-            return " " + spt.text();
-        }
         else
         {
-            return spt.text();
+            ALOGI("connect %s ok", base_url.c_str());
         }
-    }
-
-    int GetBosID() override
-    {
-        return sp.bos_id();
-    }
-
-    int GetEosID() override
-    {
-        return sp.eos_id();
-    }
-};
-
-class TokenizerMINICPM : public TokenizerLLaMa
-{
-public:
-    std::string Decode(const std::vector<int> input) override
-    {
-        sentencepiece::SentencePieceText spt;
-        sp.Decode(input, &spt);
-        return spt.text();
-    }
-};
 
-class TokenizerPhi3 : public BaseTokenizer
-{
-    sentencepiece::SentencePieceProcessor sp;
-    bool _b_bos, _b_eos;
+        cli = std::make_shared<httplib::Client>(base_url);
+        cli->set_connection_timeout(10);
+        cli->set_read_timeout(10);
+        cli->set_write_timeout(10);
 
-private:
-    /* data */
-public:
-    bool Init(std::string model_path, bool b_bos = true, bool b_eos = false) override
-    {
-        auto ret = sp.Load(model_path);
-        if (!ret.ok())
+        int try_count = 10;
+        int count = try_count;
+        while (count-- > 0)
         {
-            ALOGE("%s", ret.error_message());
-            return false;
+            try
+            {
+                auto ret = cli->Get("/get_uid");
+                auto rep = ret.value();
+                if (rep.status != 200)
+                {
+                    ALOGE("get uid failed, status: %d", rep.status);
+                    return false;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                uid = j["uid"];
+                ALOGI("uid: %s", uid.c_str());
+                break;
+            }
+            catch (const std::exception &e)
+            {
+                std::cerr << e.what() << '\n';
+            }
+            std::this_thread::sleep_for(std::chrono::seconds(1));
+            ALOGE("get uid failed, try again %d/%d", count, try_count);
         }
 
-        this->_b_bos = b_bos;
-        this->_b_eos = b_eos;
-        return ret.ok();
-    }
-
-    bool Encode(std::string input, std::vector<int> &output, bool b_img_prompt = false) override
-    {
-        auto ret = sp.Encode(input, &output);
-        if (!ret.ok())
+        count = 10;
+        while (count-- > 0)
         {
-            ALOGE("%s", ret.error_message());
-            return false;
-        }
-        output.insert(output.begin(), 32010); //"<|user|>"
-        output.push_back(32007);              //"<|end|>"
-        output.push_back(32001);              //"<|assistant|>"
-        if (_b_bos)
-        {
-            output.insert(output.begin(), sp.bos_id());
+            try
+            {
+                auto ret = cli->Get("/bos_id?uid=" + uid);
+                auto rep = ret.value();
+                if (rep.status != 200)
+                {
+                    ALOGE("get bos_id failed, status: %d", rep.status);
+                    return false;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                bos_id = j["bos_id"];
+                break;
+            }
+            catch (const std::exception &e)
+            {
+                std::cerr << e.what() << '\n';
+            }
+            std::this_thread::sleep_for(std::chrono::seconds(1));
+            ALOGE("get bos_id failed, try again %d/%d", count, try_count);
         }
-        if (_b_eos)
+
+        count = 10;
+        while (count-- > 0)
         {
-            output.push_back(sp.eos_id());
+            try
+            {
+                auto ret = cli->Get("/eos_id?uid=" + uid);
+                auto rep = ret.value();
+                if (rep.status != 200)
+                {
+                    ALOGE("get eos_id failed, status: %d", rep.status);
+                    return false;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                eos_id = j["eos_id"];
+                break;
+            }
+            catch (const std::exception &e)
+            {
+                std::cerr << e.what() << '\n';
+            }
+            std::this_thread::sleep_for(std::chrono::seconds(1));
+            ALOGE("get eos_id failed, try again %d/%d", count, try_count);
         }
-        return true;
-    }
 
-    std::vector<int> Encode(std::string input, bool b_img_prompt = false) override
-    {
-        std::vector<int> output;
-        Encode(input, output, b_img_prompt);
-        return output;
-    }
-
-    std::string Decode(const std::vector<int> input) override
-    {
-        sentencepiece::SentencePieceText spt;
-        sp.Decode(input, &spt);
-        std::string out = spt.pieces()[0].piece();
-        if (*(unsigned short *)out.data() == 38626)
+        count = 10;
+        while (count-- > 0)
         {
-            return " " + spt.text();
+            try
+            {
+                auto ret = cli->Get("/img_start_token?uid=" + uid);
+                if (!ret) {
+                    ALOGE("get img_start_token failed, no response");
+                    continue;
+                }
+                auto rep = ret.value();
+                if (rep.status != 200)
+                {
+                    ALOGE("get img_start_token failed, status: %d", rep.status);
+                    continue;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                img_start_token = j["img_start_token"];
+                ALOGI("img_start_token: %d", img_start_token);
+                break;
+            }
+            catch (const std::exception &e)
+            {
+                std::cerr << "Exception: " << e.what() << '\n';
+            }
+            std::this_thread::sleep_for(std::chrono::seconds(1));
+            ALOGE("get img_start_token failed, try again %d/%d", count, try_count);
         }
-        else
+
+        count = 10;
+        while (count-- > 0)
         {
-            return spt.text();
+            try
+            {
+                auto ret = cli->Get("/img_context_token?uid=" + uid);
+                if (!ret) {
+                    ALOGE("get img_context_token failed, no response");
+                    continue;
+                }
+                auto rep = ret.value();
+                if (rep.status != 200)
+                {
+                    ALOGE("get img_context_token failed, status: %d", rep.status);
+                    continue;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                img_context_token = j["img_context_token"];
+                ALOGI("img_context_token: %d", img_context_token);
+                break;
+            }
+            catch (const std::exception &e)
+            {
+                std::cerr << "Exception: " << e.what() << '\n';
+            }
+            std::this_thread::sleep_for(std::chrono::seconds(1));
+            ALOGE("get img_context_token failed, try again %d/%d", count, try_count);
         }
-    }
 
-    int GetBosID() override
-    {
-        return sp.bos_id();
-    }
+        printf("bos_id: %d, eos_id: %d\n", bos_id, eos_id);
+        printf("img_start_token: %d, img_context_token: %d\n", img_start_token, img_context_token);
 
-    int GetEosID() override
-    {
-        return 32007;
+        return true;
     }
 
-    bool isEnd(int id) override
+    bool Init(std::string model_path = "http://localhost:8080", bool b_bos = true, bool b_eos = false) override
     {
-        return id == GetEosID() || id > 31999;
-    }
-};
-
-class TokenizerQwen : public BaseTokenizer
-{
-    std::shared_ptr<QwenTokenizer> sp;
-    bool _b_bos, _b_eos;
+        base_url = model_path;
+        try
+        {
+            cli = std::make_shared<httplib::Client>(base_url);
+            cli->set_connection_timeout(1);
+            cli->set_read_timeout(1);
+            cli->set_write_timeout(1);
+            {
+                auto ret = cli->Get("/bos_id");
+                auto rep = ret.value();
+                if (rep.status != 200)
+                {
+                    ALOGE("get bos_id failed, status: %d", rep.status);
+                    return false;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                bos_id = j["bos_id"];
+            }
 
-private:
-    /* data */
-public:
-    bool Init(std::string model_path, bool b_bos = true, bool b_eos = false) override
-    {
-        if (!file_exist(model_path))
+            {
+                auto ret = cli->Get("/eos_id");
+                auto rep = ret.value();
+                if (rep.status != 200)
+                {
+                    ALOGE("get eos_id failed, status: %d", rep.status);
+                    return false;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                eos_id = j["eos_id"];
+            }
+            printf("bos_id: %d, eos_id: %d\n", bos_id, eos_id);
+        }
+        catch (const std::exception &e)
         {
-            ALOGE("tokenizer model file(%s) not exist", model_path.c_str());
+            std::cerr << e.what() << '\n';
             return false;
         }
 
-        sp.reset(new QwenTokenizer(model_path, QwenConfig()));
-
         this->_b_bos = b_bos;
         this->_b_eos = b_eos;
         return true;
     }
 
-    bool Encode(std::string input, std::vector<int> &output, bool b_img_prompt = false) override
+    bool Init_new(std::string model_path, bool b_bos, bool b_eos) override
     {
-        if (_b_bos)
+        base_url = model_path;
+        if (!test_connect_http(base_url, 10))
         {
-            // input += "<|im_start|>";
+            ALOGE("connect %s failed", base_url.c_str());
+            return false;
         }
-        if (_b_eos)
+        else
         {
-            input += "<|endoftext|>";
+            ALOGI("connect %s ok", base_url.c_str());
         }
-        output = sp->encode(input, 1024);
-
-        return true;
-    }
-
-    std::vector<int> Encode(std::string input, bool b_img_prompt = false) override
-    {
-        std::vector<int> output;
-        Encode(input, output, b_img_prompt);
-        return output;
-    }
-
-    std::string Decode(const std::vector<int> input) override
-    {
-        return sp->decode(input);
-    }
-
-    int GetBosID() override
-    {
-        return -1;
-    }
-
-    int GetEosID() override
-    {
-        return sp->eos_token_id;
-    }
-};
-
-// class TokenizerGLM3 : public BaseTokenizer
-// {
-//     std::shared_ptr<chatglm::ChatGLM3Tokenizer> sp;
-//     bool _b_bos, _b_eos;
-
-// private:
-//     /* data */
-// public:
-//     bool Init(std::string model_path, bool b_bos = true, bool b_eos = false) override
-//     {
-//         if (!file_exist(model_path))
-//         {
-//             ALOGE("tokenizer model file(%s) not exist", model_path.c_str());
-//             return false;
-//         }
-//         // std::vector<char> sp_model_data;
-//         // read_file(model_path, sp_model_data);
-//         // std::string_view serialized_model_proto(sp_model_data.data(), sp_model_data.size());
-
-//         sp.reset(new chatglm::ChatGLM3Tokenizer(model_path));
-
-//         this->_b_bos = b_bos;
-//         this->_b_eos = b_eos;
-//         return true;
-//     }
-
-//     bool Encode(std::string input, std::vector<int> &output) override
-//     {
-//         if (_b_bos)
-//         {
-//             // input += "<|im_start|>";
-//         }
-//         if (_b_eos)
-//         {
-//             // input += "<|endoftext|>";
-//         }
-//         output = sp->encode(input, 1024);
-
-//         return true;
-//     }
-
-//     std::vector<int> Encode(std::string input) override
-//     {
-//         std::vector<int> output;
-//         Encode(input, output);
-//         return output;
-//     }
-
-//     std::string Decode(const std::vector<int> input) override
-//     {
-//         return sp->decode(input);
-//     }
-
-//     int GetBosID() override
-//     {
-//         return sp->sp.bos_id();
-//     }
-
-//     int GetEosID() override
-//     {
-//         return sp->sp.eos_id();
-//     }
-// };
-
-class Tokenizer_Http : public BaseTokenizer
-{
-    std::shared_ptr<httplib::Client> cli;
-    bool _b_bos, _b_eos;
 
-    std::string base_url;
-
-    int bos_id, eos_id;
-
-private:
-    /* data */
-public:
-    bool Init(std::string model_path = "http://localhost:8080", bool b_bos = true, bool b_eos = false) override
-    {
-        base_url = model_path;
         try
         {
             cli = std::make_shared<httplib::Client>(base_url);
@@ -351,7 +266,33 @@ class Tokenizer_Http : public BaseTokenizer
                 nlohmann::json j = nlohmann::json::parse(rep.body);
                 eos_id = j["eos_id"];
             }
-            printf("bos_id: %d, eos_id: %d\n", bos_id, eos_id);
+            ALOGI("bos_id: %d, eos_id: %d", bos_id, eos_id);
+
+            {
+                auto ret = cli->Get("/img_start_token");
+                auto rep = ret.value();
+                if (rep.status != 200)
+                {
+                    ALOGE("get img_start_token failed, status: %d", rep.status);
+                    return false;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                img_start_token = j["img_start_token"];
+            }
+            ALOGI("img_start_token: %d", img_start_token);
+
+            {
+                auto ret = cli->Get("/img_context_token");
+                auto rep = ret.value();
+                if (rep.status != 200)
+                {
+                    ALOGE("get img_context_token failed, status: %d", rep.status);
+                    return false;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                img_context_token = j["img_context_token"];
+            }
+            ALOGI("img_context_token: %d", img_context_token);
         }
         catch (const std::exception &e)
         {
@@ -364,11 +305,75 @@ class Tokenizer_Http : public BaseTokenizer
         return true;
     }
 
-    bool Encode(std::string input, std::vector<int> &output, bool b_img_prompt = false) override
+    bool Reset(std::string system_prompt, std::vector<int> &tokens) override
+    {
+        nlohmann::json j;
+        j["uid"] = uid;
+        if (!system_prompt.empty() and system_prompt != "")
+        {
+            j["system_prompt"] = system_prompt;
+        }
+
+        auto ret = cli->Post("/reset", j.dump(), "application/json");
+        auto rep = ret.value();
+        if (rep.status != 200)
+        {
+            ALOGE("reset failed, status: %d", rep.status);
+            return false;
+        }
+        nlohmann::json j_rep = nlohmann::json::parse(rep.body);
+        std::vector<int> _token_ids = j_rep["token_ids"];
+        tokens = _token_ids;
+        return true;
+    }
+
+    bool Encode(std::string input, std::string last_reply, std::vector<int> &tokens, std::vector<int> &tokens_diff, ImageInfo img_info) override
     {
         nlohmann::json j;
+        j["uid"] = uid;
         j["text"] = input;
-        j["img_prompt"] = b_img_prompt;
+        j["img_prompt"] = img_info.img_prompt;
+        j["imgsz"] = img_info.imgsz;
+        j["num_img"] = img_info.num_img;
+        if (!last_reply.empty() and last_reply != "")
+        {
+            j["last_reply"] = last_reply;
+        }
+        auto ret = cli->Post("/encode", j.dump(), "application/json");
+        auto rep = ret.value();
+        if (rep.status != 200)
+        {
+            ALOGE("encode failed, status: %d", rep.status);
+            return false;
+        }
+        nlohmann::json j2;
+        try
+        {
+            j2 = nlohmann::json::parse(rep.body);
+        }
+        catch (const std::exception &e)
+        {
+            ALOGE("json parse failed: %s", e.what());
+            ALOGE("%s", rep.body.c_str());
+            return false;
+        }
+
+        std::vector<int> _token_ids = j2["token_ids"];
+        std::vector<int> _tokens_diff = j2["diff"];
+
+        tokens = _token_ids;
+        tokens_diff = _tokens_diff;
+
+        return true;
+    }
+
+    bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) override
+    {
+        nlohmann::json j;
+        j["text"] = input;
+        j["img_prompt"] = img_info.img_prompt;
+        j["imgsz"] = img_info.imgsz;
+        j["num_img"] = img_info.num_img;
         auto ret = cli->Post("/encode", j.dump(), "application/json");
         auto rep = ret.value();
         if (rep.status != 200)
@@ -403,10 +408,17 @@ class Tokenizer_Http : public BaseTokenizer
         return true;
     }
 
-    std::vector<int> Encode(std::string input, bool b_img_prompt = false) override
+    std::vector<int> Encode(std::string input, ImageInfo img_info) override
     {
         std::vector<int> output;
-        Encode(input, output, b_img_prompt);
+        Encode(input, output, img_info);
+        return output;
+    }
+
+    std::vector<int> Encode_ctx(std::string input, ImageInfo img_info, std::vector<int> &tokens_ids, std::vector<int> &tokens_diff) override
+    {
+        std::vector<int> output;
+        Encode(input, "", output, tokens_diff, img_info);
         return output;
     }
 
@@ -418,6 +430,7 @@ class Tokenizer_Http : public BaseTokenizer
         {
             nlohmann::json j;
             j["token_ids"] = input;
+            j["uid"] = uid;
             auto ret = cli->Post("/decode", j.dump(), "application/json");
             auto rep = ret.value();
             if (rep.status != 200)
@@ -453,23 +466,26 @@ class Tokenizer_Http : public BaseTokenizer
     {
         return eos_id;
     }
+
+    int GetImgStartID() override
+    {
+        return img_start_token;
+    }
+
+    int GetImgContextID() override
+    {
+        return img_context_token;
+    }
 };
 
 std::shared_ptr<BaseTokenizer> CreateTokenizer(TokenizerType type)
 {
     switch (type)
     {
-    case TKT_LLaMa:
-        return std::make_shared<TokenizerLLaMa>();
-    case TKT_MINICPM:
-        return std::make_shared<TokenizerMINICPM>();
     case TKT_HTTP:
         return std::make_shared<Tokenizer_Http>();
-    case TKT_Qwen:
-        return std::make_shared<TokenizerQwen>();
-    case TKT_Phi3:
-        return std::make_shared<TokenizerPhi3>();
     default:
+        ALOGE("unknown tokenizer type: %d", type);
         return nullptr;
     }
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.hpp b/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.hpp
index c6ad3d8c..e82bdfd8 100644
--- a/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.hpp
@@ -9,19 +9,32 @@ enum TokenizerType
     TKT_Qwen,
     TKT_HTTP,
     TKT_Phi3,
-    TKT_MINICPM,
     TKT_END
 };
 
+struct ImageInfo
+{
+    int imgsz = 448;
+    int num_img = 1;
+    bool img_prompt = false;
+};
+
 class BaseTokenizer
 {
 public:
-    virtual bool Init(std::string model_path, bool b_bos = true, bool b_eos = false) = 0;
-    virtual bool Encode(std::string input, std::vector<int> &output, bool b_img_prompt = false) = 0;
-    virtual std::vector<int> Encode(std::string input, bool b_img_prompt = false) = 0;
+    virtual bool Init(std::string model_path) = 0;
+    virtual bool Init(std::string model_path, bool b_bos, bool b_eos) = 0;
+    virtual bool Init_new(std::string model_path, bool b_bos, bool b_eos) = 0;
+    virtual bool Reset(std::string system_prompt, std::vector<int> &tokens) = 0;
+    virtual bool Encode(std::string input, std::string last_reply, std::vector<int> &tokens, std::vector<int> &tokens_diff, ImageInfo img_info) = 0;
+    virtual bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) = 0;
+    virtual std::vector<int> Encode(std::string input, ImageInfo img_info) = 0;
+    virtual std::vector<int> Encode_ctx(std::string input, ImageInfo img_info, std::vector<int> &tokens_ids, std::vector<int> &tokens_diff) = 0;
     virtual std::string Decode(const std::vector<int> input) = 0;
     virtual int GetBosID() = 0;
     virtual int GetEosID() = 0;
+    virtual int GetImgStartID() = 0;
+    virtual int GetImgContextID() = 0;
 
     virtual bool isEnd(int id) { return id == GetEosID(); }
 };
diff --git a/projects/llm_framework/main_vlm/src/runner/Tokenizer/base64.h b/projects/llm_framework/main_vlm/src/runner/Tokenizer/base64.h
deleted file mode 100644
index c42b4828..00000000
--- a/projects/llm_framework/main_vlm/src/runner/Tokenizer/base64.h
+++ /dev/null
@@ -1,54 +0,0 @@
-#pragma once
-
-#include <string>
-#include <string_view>
-#include <stdexcept>
-
-namespace base64
-{
-
-    static auto pos_of_char(const unsigned char chr) -> size_t
-    {
-        if (chr >= 'A' && chr <= 'Z')
-            return chr - 'A';
-        else if (chr >= 'a' && chr <= 'z')
-            return chr - 'a' + ('Z' - 'A') + 1;
-        else if (chr >= '0' && chr <= '9')
-            return chr - '0' + ('Z' - 'A') + ('z' - 'a') + 2;
-        else if (chr == '+' || chr == '-')
-            return 62;
-        else if (chr == '/' || chr == '_')
-            return 63;
-        else
-            throw std::runtime_error("Input is not valid base64-encoded data.");
-    }
-
-    inline auto decode(std::string_view s) -> std::string
-    {
-        if (s.empty())
-            throw std::runtime_error("empty input");
-        size_t length = s.length();
-        size_t idx = 0;
-
-        std::string out;
-        out.reserve(length / 4 * 3);
-
-        while (idx < length)
-        {
-            size_t pos_of_char_1 = pos_of_char(s.at(idx + 1));
-            out.push_back(static_cast<std::string::value_type>(((pos_of_char(s.at(idx + 0))) << 2) + ((pos_of_char_1 & 0x30) >> 4)));
-            if ((idx + 2 < length) && s.at(idx + 2) != '=' && s.at(idx + 2) != '.')
-            {
-                size_t pos_of_char_2 = pos_of_char(s.at(idx + 2));
-                out.push_back(static_cast<std::string::value_type>(((pos_of_char_1 & 0x0f) << 4) + ((pos_of_char_2 & 0x3c) >> 2)));
-                if ((idx + 3 < length) && s.at(idx + 3) != '=' && s.at(idx + 3) != '.')
-                {
-                    out.push_back(static_cast<std::string::value_type>(((pos_of_char_2 & 0x03) << 6) + pos_of_char(s.at(idx + 3))));
-                }
-            }
-            idx += 4;
-        }
-        return out;
-    }
-
-} // namespace base64
diff --git a/projects/llm_framework/main_vlm/src/runner/Tokenizer/tiktoken.h b/projects/llm_framework/main_vlm/src/runner/Tokenizer/tiktoken.h
deleted file mode 100644
index 62738502..00000000
--- a/projects/llm_framework/main_vlm/src/runner/Tokenizer/tiktoken.h
+++ /dev/null
@@ -1,326 +0,0 @@
-#pragma once
-
-#include <re2/re2.h>
-#include "unordered_dense.h"
-
-#include <cassert>
-#include <limits>
-#include <optional>
-#include <regex>
-#include <string>
-#include <unordered_map>
-#include <unordered_set>
-#include <vector>
-
-namespace tiktoken
-{
-
-	static auto _byte_pair_merge(
-		const std::string &piece,
-		const ankerl::unordered_dense::map<std::string, int> &ranks,
-		std::function<int(int, int)> func) -> std::vector<int>
-	{
-		std::vector<std::pair<int, int>> parts;
-		parts.reserve(piece.size() + 1);
-		for (auto idx = 0U; idx < piece.size() + 1; ++idx)
-		{
-			parts.emplace_back(idx, std::numeric_limits<int>::max());
-		}
-
-		auto get_rank = [&piece, &ranks](
-							const std::vector<std::pair<int, int>> &parts,
-							int start_idx,
-							int skip) -> std::optional<int>
-		{
-			if (start_idx + skip + 2 < parts.size())
-			{
-				auto s = parts[start_idx].first;
-				auto e = parts[start_idx + skip + 2].first;
-				auto key = piece.substr(s, e - s);
-				auto iter = ranks.find(key);
-				if (iter != ranks.end())
-				{
-					return iter->second;
-				}
-			}
-			return std::nullopt;
-		};
-
-		for (auto i = 0U; i < parts.size() - 2; ++i)
-		{
-			auto rank = get_rank(parts, i, 0);
-			if (rank)
-			{
-				assert(*rank != std::numeric_limits<int>::max());
-				parts[i].second = *rank;
-			}
-		}
-
-		while (true)
-		{
-			if (parts.size() == 1)
-				break;
-
-			auto min_rank = std::make_pair<int, int>(std::numeric_limits<int>::max(), 0);
-			for (auto i = 0U; i < parts.size() - 1; ++i)
-			{
-				auto rank = parts[i].second;
-				if (rank < min_rank.first)
-				{
-					min_rank = {rank, i};
-				}
-			}
-
-			if (min_rank.first != std::numeric_limits<int>::max())
-			{
-				auto i = min_rank.second;
-				auto rank = get_rank(parts, i, 1);
-				if (rank)
-				{
-					parts[i].second = *rank;
-				}
-				else
-				{
-					parts[i].second = std::numeric_limits<int>::max();
-				}
-				if (i > 0)
-				{
-					auto rank = get_rank(parts, i - 1, 1);
-					if (rank)
-					{
-						parts[i - 1].second = *rank;
-					}
-					else
-					{
-						parts[i - 1].second = std::numeric_limits<int>::max();
-					}
-				}
-
-				parts.erase(parts.begin() + (i + 1));
-			}
-			else
-			{
-				break;
-			}
-		}
-		std::vector<int> out;
-		out.reserve(parts.size() - 1);
-		for (auto i = 0U; i < parts.size() - 1; ++i)
-		{
-			out.push_back(func(parts[i].first, parts[i + 1].first));
-		}
-		return out;
-	}
-
-	static auto byte_pair_encode(
-		const std::string &piece,
-		const ankerl::unordered_dense::map<std::string, int> &ranks) -> std::vector<int>
-	{
-		if (piece.size() == 1)
-		{
-			return {ranks.at(piece)};
-		}
-
-		auto func = [&piece, &ranks](int start, int stop) -> int
-		{
-			std::string key = piece.substr(start, stop - start);
-			return ranks.at(key);
-		};
-
-		return _byte_pair_merge(piece, ranks, func);
-	}
-
-	class tiktoken
-	{
-	public:
-		tiktoken() = default;
-		tiktoken(
-			ankerl::unordered_dense::map<std::string, int> encoder,
-			ankerl::unordered_dense::map<std::string, int> special_encoder,
-			const std::string &pattern)
-		{
-			regex_ = std::make_unique<re2::RE2>("(" + pattern + ")");
-
-			std::string special_pattern;
-			for (const auto &item : special_encoder)
-			{
-				if (!special_pattern.empty())
-				{
-					special_pattern += "|";
-				}
-				special_pattern += re2::RE2::QuoteMeta(item.first);
-			}
-			if (special_pattern.empty())
-			{
-				special_regex_ = nullptr;
-			}
-			else
-			{
-				special_regex_ = std::make_unique<re2::RE2>("(" + special_pattern + ")");
-			}
-
-			encoder_ = std::move(encoder);
-			special_tokens_encoder = std::move(special_encoder);
-
-			for (const auto &[k, v] : encoder_)
-			{
-				decoder_.emplace(v, k);
-			}
-			assert(encoder_.size() != decoder_.size() && "Encoder and decoder must be of equal length; maybe you had duplicate token indices in your encoder?");
-
-			for (const auto &[k, v] : special_tokens_encoder)
-			{
-				special_tokens_decoder.emplace(v, k);
-			}
-		}
-
-		auto encode_ordinary(const std::string &text) const -> std::vector<int>
-		{
-			return _encode_ordinary_native(text);
-		}
-
-		auto encode(const std::string &text) const -> std::vector<int>
-		{
-			return _encode_native(text, special_tokens_encoder).first;
-		}
-
-		auto encode_single_piece(const std::string &text) const -> std::vector<int>
-		{
-			auto iter = encoder_.find(text);
-			if (iter != encoder_.end())
-			{
-				return {iter->second};
-			}
-			return byte_pair_encode(text, encoder_);
-		}
-
-		auto decode(const std::vector<int> &tokens) const -> std::string
-		{
-			return _decode_native(tokens);
-		}
-
-	private:
-		auto split_with_allowed_special_token(
-			re2::StringPiece &input,
-			const ankerl::unordered_dense::map<std::string, int> &allowed_special) const -> std::pair<std::optional<std::string>, re2::StringPiece>
-		{
-			if (special_regex_ == nullptr)
-				return {std::nullopt, input};
-
-			auto start = input.begin();
-			std::string special;
-			while (true)
-			{
-				if (!re2::RE2::FindAndConsume(&input, *special_regex_, &special))
-				{
-					break;
-				}
-
-				if (allowed_special.count(special) == 1)
-				{
-					return {std::move(special), re2::StringPiece(start, input.begin() - start - special.size())};
-				}
-			}
-
-			return {std::nullopt, input};
-		}
-
-		auto _encode_ordinary_native(const std::string &text) const -> std::vector<int>
-		{
-			std::vector<int> ret;
-			re2::StringPiece input(text);
-
-			std::string piece;
-			while (re2::RE2::FindAndConsume(&input, *regex_, &piece))
-			{
-				auto iter = encoder_.find(piece);
-				if (iter != encoder_.end())
-				{
-					ret.push_back(iter->second);
-					continue;
-				}
-				auto tokens = byte_pair_encode(piece, encoder_);
-				ret.insert(ret.end(), tokens.begin(), tokens.end());
-			}
-			return ret;
-		}
-
-		auto _encode_native(
-			const std::string &text,
-			const ankerl::unordered_dense::map<std::string, int> &allowed_special) const -> std::pair<std::vector<int>, int>
-		{
-			std::vector<int> ret;
-			int last_piece_token_len = 0;
-			re2::StringPiece input(text);
-
-			while (true)
-			{
-				auto [special, sub_input] = split_with_allowed_special_token(input, allowed_special);
-				std::string piece;
-				while (re2::RE2::FindAndConsume(&sub_input, *regex_, &piece))
-				{
-					auto iter = encoder_.find(piece);
-					if (iter != encoder_.end())
-					{
-						last_piece_token_len = 1;
-						ret.push_back(iter->second);
-						continue;
-					}
-					auto tokens = byte_pair_encode(piece, encoder_);
-					last_piece_token_len = tokens.size();
-					ret.insert(ret.end(), tokens.begin(), tokens.end());
-				}
-
-				if (special)
-				{
-					int token = special_tokens_encoder.at(*special);
-					ret.push_back(token);
-					last_piece_token_len = 0;
-				}
-				else
-				{
-					break;
-				}
-			}
-
-			return {ret, last_piece_token_len};
-		}
-
-		auto _decode_native(const std::vector<int> &tokens) const -> std::string
-		{
-			std::string ret;
-			ret.reserve(tokens.size() * 2);
-			for (auto token : tokens)
-			{
-				std::string token_bytes;
-				auto iter = decoder_.find(token);
-				if (iter != decoder_.end())
-				{
-					token_bytes = iter->second;
-				}
-				else
-				{
-					iter = special_tokens_decoder.find(token);
-					if (iter != special_tokens_decoder.end())
-					{
-						token_bytes = iter->second;
-					}
-					else
-					{
-						throw std::runtime_error("unknown token: " + std::to_string(token));
-					}
-				}
-				ret += token_bytes;
-			}
-			return ret;
-		}
-
-		ankerl::unordered_dense::map<std::string, int> encoder_;
-		ankerl::unordered_dense::map<std::string, int> special_tokens_encoder;
-		ankerl::unordered_dense::map<int, std::string> decoder_;
-		ankerl::unordered_dense::map<int, std::string> special_tokens_decoder;
-		std::unique_ptr<re2::RE2> regex_;
-		std::unique_ptr<re2::RE2> special_regex_;
-	};
-
-} // namespace tiktoken
diff --git a/projects/llm_framework/main_vlm/src/runner/Tokenizer/unordered_dense.h b/projects/llm_framework/main_vlm/src/runner/Tokenizer/unordered_dense.h
deleted file mode 100644
index af23690b..00000000
--- a/projects/llm_framework/main_vlm/src/runner/Tokenizer/unordered_dense.h
+++ /dev/null
@@ -1,2240 +0,0 @@
-///////////////////////// ankerl::unordered_dense::{map, set} /////////////////////////
-
-// A fast & densely stored hashmap and hashset based on robin-hood backward shift deletion.
-// Version 4.1.2
-// https://github.com/martinus/unordered_dense
-//
-// Licensed under the MIT License <http://opensource.org/licenses/MIT>.
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2022-2023 Martin Leitner-Ankerl <martin.ankerl@gmail.com>
-//
-// Permission is hereby granted, free of charge, to any person obtaining a copy
-// of this software and associated documentation files (the "Software"), to deal
-// in the Software without restriction, including without limitation the rights
-// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-// copies of the Software, and to permit persons to whom the Software is
-// furnished to do so, subject to the following conditions:
-//
-// The above copyright notice and this permission notice shall be included in all
-// copies or substantial portions of the Software.
-//
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-// SOFTWARE.
-
-#ifndef ANKERL_UNORDERED_DENSE_H
-#define ANKERL_UNORDERED_DENSE_H
-
-// see https://semver.org/spec/v2.0.0.html
-#define ANKERL_UNORDERED_DENSE_VERSION_MAJOR 4 // NOLINT(cppcoreguidelines-macro-usage) incompatible API changes
-#define ANKERL_UNORDERED_DENSE_VERSION_MINOR 1 // NOLINT(cppcoreguidelines-macro-usage) backwards compatible functionality
-#define ANKERL_UNORDERED_DENSE_VERSION_PATCH 2 // NOLINT(cppcoreguidelines-macro-usage) backwards compatible bug fixes
-
-// API versioning with inline namespace, see https://www.foonathan.net/2018/11/inline-namespaces/
-
-// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
-#define ANKERL_UNORDERED_DENSE_VERSION_CONCAT1(major, minor, patch) v##major##_##minor##_##patch
-// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
-#define ANKERL_UNORDERED_DENSE_VERSION_CONCAT(major, minor, patch) ANKERL_UNORDERED_DENSE_VERSION_CONCAT1(major, minor, patch)
-#define ANKERL_UNORDERED_DENSE_NAMESPACE   \
-    ANKERL_UNORDERED_DENSE_VERSION_CONCAT( \
-        ANKERL_UNORDERED_DENSE_VERSION_MAJOR, ANKERL_UNORDERED_DENSE_VERSION_MINOR, ANKERL_UNORDERED_DENSE_VERSION_PATCH)
-
-#if defined(_MSVC_LANG)
-#define ANKERL_UNORDERED_DENSE_CPP_VERSION _MSVC_LANG
-#else
-#define ANKERL_UNORDERED_DENSE_CPP_VERSION __cplusplus
-#endif
-
-#if defined(__GNUC__)
-// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
-#define ANKERL_UNORDERED_DENSE_PACK(decl) decl __attribute__((__packed__))
-#elif defined(_MSC_VER)
-// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
-#define ANKERL_UNORDERED_DENSE_PACK(decl) __pragma(pack(push, 1)) decl __pragma(pack(pop))
-#endif
-
-// exceptions
-#if defined(__cpp_exceptions) || defined(__EXCEPTIONS) || defined(_CPPUNWIND)
-#define ANKERL_UNORDERED_DENSE_HAS_EXCEPTIONS() 1 // NOLINT(cppcoreguidelines-macro-usage)
-#else
-#define ANKERL_UNORDERED_DENSE_HAS_EXCEPTIONS() 0 // NOLINT(cppcoreguidelines-macro-usage)
-#endif
-#ifdef _MSC_VER
-#define ANKERL_UNORDERED_DENSE_NOINLINE __declspec(noinline)
-#else
-#define ANKERL_UNORDERED_DENSE_NOINLINE __attribute__((noinline))
-#endif
-
-// defined in unordered_dense.cpp
-#if !defined(ANKERL_UNORDERED_DENSE_EXPORT)
-#define ANKERL_UNORDERED_DENSE_EXPORT
-#endif
-
-#if ANKERL_UNORDERED_DENSE_CPP_VERSION < 201703L
-#error ankerl::unordered_dense requires C++17 or higher
-#else
-#include <array>            // for array
-#include <cstdint>          // for uint64_t, uint32_t, uint8_t, UINT64_C
-#include <cstring>          // for size_t, memcpy, memset
-#include <functional>       // for equal_to, hash
-#include <initializer_list> // for initializer_list
-#include <iterator>         // for pair, distance
-#include <limits>           // for numeric_limits
-#include <memory>           // for allocator, allocator_traits, shared_ptr
-#include <stdexcept>        // for out_of_range
-#include <string>           // for basic_string
-#include <string_view>      // for basic_string_view, hash
-#include <tuple>            // for forward_as_tuple
-#include <type_traits>      // for enable_if_t, declval, conditional_t, ena...
-#include <utility>          // for forward, exchange, pair, as_const, piece...
-#include <vector>           // for vector
-#if ANKERL_UNORDERED_DENSE_HAS_EXCEPTIONS() == 0
-#include <cstdlib> // for abort
-#endif
-
-#if defined(__has_include)
-#if __has_include(<memory_resource>)
-#define ANKERL_UNORDERED_DENSE_PMR std::pmr // NOLINT(cppcoreguidelines-macro-usage)
-#include <memory_resource>                  // for polymorphic_allocator
-#elif __has_include(<experimental/memory_resource>)
-#define ANKERL_UNORDERED_DENSE_PMR std::experimental::pmr // NOLINT(cppcoreguidelines-macro-usage)
-#include <experimental/memory_resource>                   // for polymorphic_allocator
-#endif
-#endif
-
-#if defined(_MSC_VER) && defined(_M_X64)
-#include <intrin.h>
-#pragma intrinsic(_umul128)
-#endif
-
-#if defined(__GNUC__) || defined(__INTEL_COMPILER) || defined(__clang__)
-#define ANKERL_UNORDERED_DENSE_LIKELY(x) __builtin_expect(x, 1)   // NOLINT(cppcoreguidelines-macro-usage)
-#define ANKERL_UNORDERED_DENSE_UNLIKELY(x) __builtin_expect(x, 0) // NOLINT(cppcoreguidelines-macro-usage)
-#else
-#define ANKERL_UNORDERED_DENSE_LIKELY(x) (x)   // NOLINT(cppcoreguidelines-macro-usage)
-#define ANKERL_UNORDERED_DENSE_UNLIKELY(x) (x) // NOLINT(cppcoreguidelines-macro-usage)
-#endif
-
-namespace ankerl::unordered_dense
-{
-    inline namespace ANKERL_UNORDERED_DENSE_NAMESPACE
-    {
-
-        namespace detail
-        {
-
-#if ANKERL_UNORDERED_DENSE_HAS_EXCEPTIONS()
-
-            // make sure this is not inlined as it is slow and dramatically enlarges code, thus making other
-            // inlinings more difficult. Throws are also generally the slow path.
-            [[noreturn]] inline ANKERL_UNORDERED_DENSE_NOINLINE void on_error_key_not_found()
-            {
-                throw std::out_of_range("ankerl::unordered_dense::map::at(): key not found");
-            }
-            [[noreturn]] inline ANKERL_UNORDERED_DENSE_NOINLINE void on_error_bucket_overflow()
-            {
-                throw std::overflow_error("ankerl::unordered_dense: reached max bucket size, cannot increase size");
-            }
-            [[noreturn]] inline ANKERL_UNORDERED_DENSE_NOINLINE void on_error_too_many_elements()
-            {
-                throw std::out_of_range("ankerl::unordered_dense::map::replace(): too many elements");
-            }
-
-#else
-
-            [[noreturn]] inline void on_error_key_not_found()
-            {
-                abort();
-            }
-            [[noreturn]] inline void on_error_bucket_overflow()
-            {
-                abort();
-            }
-            [[noreturn]] inline void on_error_too_many_elements()
-            {
-                abort();
-            }
-
-#endif
-
-        } // namespace detail
-
-        // hash ///////////////////////////////////////////////////////////////////////
-
-        // This is a stripped-down implementation of wyhash: https://github.com/wangyi-fudan/wyhash
-        // No big-endian support (because different values on different machines don't matter),
-        // hardcodes seed and the secret, reformats the code, and clang-tidy fixes.
-        namespace detail::wyhash
-        {
-
-            inline void mum(uint64_t *a, uint64_t *b)
-            {
-#if defined(__SIZEOF_INT128__)
-                __uint128_t r = *a;
-                r *= *b;
-                *a = static_cast<uint64_t>(r);
-                *b = static_cast<uint64_t>(r >> 64U);
-#elif defined(_MSC_VER) && defined(_M_X64)
-                *a = _umul128(*a, *b, b);
-#else
-                uint64_t ha = *a >> 32U;
-                uint64_t hb = *b >> 32U;
-                uint64_t la = static_cast<uint32_t>(*a);
-                uint64_t lb = static_cast<uint32_t>(*b);
-                uint64_t hi{};
-                uint64_t lo{};
-                uint64_t rh = ha * hb;
-                uint64_t rm0 = ha * lb;
-                uint64_t rm1 = hb * la;
-                uint64_t rl = la * lb;
-                uint64_t t = rl + (rm0 << 32U);
-                auto c = static_cast<uint64_t>(t < rl);
-                lo = t + (rm1 << 32U);
-                c += static_cast<uint64_t>(lo < t);
-                hi = rh + (rm0 >> 32U) + (rm1 >> 32U) + c;
-                *a = lo;
-                *b = hi;
-#endif
-            }
-
-            // multiply and xor mix function, aka MUM
-            [[nodiscard]] inline auto mix(uint64_t a, uint64_t b) -> uint64_t
-            {
-                mum(&a, &b);
-                return a ^ b;
-            }
-
-            // read functions. WARNING: we don't care about endianness, so results are different on big endian!
-            [[nodiscard]] inline auto r8(const uint8_t *p) -> uint64_t
-            {
-                uint64_t v{};
-                std::memcpy(&v, p, 8U);
-                return v;
-            }
-
-            [[nodiscard]] inline auto r4(const uint8_t *p) -> uint64_t
-            {
-                uint32_t v{};
-                std::memcpy(&v, p, 4);
-                return v;
-            }
-
-            // reads 1, 2, or 3 bytes
-            [[nodiscard]] inline auto r3(const uint8_t *p, size_t k) -> uint64_t
-            {
-                return (static_cast<uint64_t>(p[0]) << 16U) | (static_cast<uint64_t>(p[k >> 1U]) << 8U) | p[k - 1];
-            }
-
-            [[maybe_unused]] [[nodiscard]] inline auto hash(void const *key, size_t len) -> uint64_t
-            {
-                static constexpr auto secret = std::array{UINT64_C(0xa0761d6478bd642f),
-                                                          UINT64_C(0xe7037ed1a0b428db),
-                                                          UINT64_C(0x8ebc6af09c88c6e3),
-                                                          UINT64_C(0x589965cc75374cc3)};
-
-                auto const *p = static_cast<uint8_t const *>(key);
-                uint64_t seed = secret[0];
-                uint64_t a{};
-                uint64_t b{};
-                if (ANKERL_UNORDERED_DENSE_LIKELY(len <= 16))
-                {
-                    if (ANKERL_UNORDERED_DENSE_LIKELY(len >= 4))
-                    {
-                        a = (r4(p) << 32U) | r4(p + ((len >> 3U) << 2U));
-                        b = (r4(p + len - 4) << 32U) | r4(p + len - 4 - ((len >> 3U) << 2U));
-                    }
-                    else if (ANKERL_UNORDERED_DENSE_LIKELY(len > 0))
-                    {
-                        a = r3(p, len);
-                        b = 0;
-                    }
-                    else
-                    {
-                        a = 0;
-                        b = 0;
-                    }
-                }
-                else
-                {
-                    size_t i = len;
-                    if (ANKERL_UNORDERED_DENSE_UNLIKELY(i > 48))
-                    {
-                        uint64_t see1 = seed;
-                        uint64_t see2 = seed;
-                        do
-                        {
-                            seed = mix(r8(p) ^ secret[1], r8(p + 8) ^ seed);
-                            see1 = mix(r8(p + 16) ^ secret[2], r8(p + 24) ^ see1);
-                            see2 = mix(r8(p + 32) ^ secret[3], r8(p + 40) ^ see2);
-                            p += 48;
-                            i -= 48;
-                        } while (ANKERL_UNORDERED_DENSE_LIKELY(i > 48));
-                        seed ^= see1 ^ see2;
-                    }
-                    while (ANKERL_UNORDERED_DENSE_UNLIKELY(i > 16))
-                    {
-                        seed = mix(r8(p) ^ secret[1], r8(p + 8) ^ seed);
-                        i -= 16;
-                        p += 16;
-                    }
-                    a = r8(p + i - 16);
-                    b = r8(p + i - 8);
-                }
-
-                return mix(secret[1] ^ len, mix(a ^ secret[1], b ^ seed));
-            }
-
-            [[nodiscard]] inline auto hash(uint64_t x) -> uint64_t
-            {
-                return detail::wyhash::mix(x, UINT64_C(0x9E3779B97F4A7C15));
-            }
-
-        } // namespace detail::wyhash
-
-        ANKERL_UNORDERED_DENSE_EXPORT template <typename T, typename Enable = void>
-        struct hash
-        {
-            auto operator()(T const &obj) const noexcept(noexcept(std::declval<std::hash<T>>().operator()(std::declval<T const &>())))
-                -> uint64_t
-            {
-                return std::hash<T>{}(obj);
-            }
-        };
-
-        template <typename CharT>
-        struct hash<std::basic_string<CharT>>
-        {
-            using is_avalanching = void;
-            auto operator()(std::basic_string<CharT> const &str) const noexcept -> uint64_t
-            {
-                return detail::wyhash::hash(str.data(), sizeof(CharT) * str.size());
-            }
-        };
-
-        template <typename CharT>
-        struct hash<std::basic_string_view<CharT>>
-        {
-            using is_avalanching = void;
-            auto operator()(std::basic_string_view<CharT> const &sv) const noexcept -> uint64_t
-            {
-                return detail::wyhash::hash(sv.data(), sizeof(CharT) * sv.size());
-            }
-        };
-
-        template <class T>
-        struct hash<T *>
-        {
-            using is_avalanching = void;
-            auto operator()(T *ptr) const noexcept -> uint64_t
-            {
-                // NOLINTNEXTLINE(cppcoreguidelines-pro-type-reinterpret-cast)
-                return detail::wyhash::hash(reinterpret_cast<uintptr_t>(ptr));
-            }
-        };
-
-        template <class T>
-        struct hash<std::unique_ptr<T>>
-        {
-            using is_avalanching = void;
-            auto operator()(std::unique_ptr<T> const &ptr) const noexcept -> uint64_t
-            {
-                // NOLINTNEXTLINE(cppcoreguidelines-pro-type-reinterpret-cast)
-                return detail::wyhash::hash(reinterpret_cast<uintptr_t>(ptr.get()));
-            }
-        };
-
-        template <class T>
-        struct hash<std::shared_ptr<T>>
-        {
-            using is_avalanching = void;
-            auto operator()(std::shared_ptr<T> const &ptr) const noexcept -> uint64_t
-            {
-                // NOLINTNEXTLINE(cppcoreguidelines-pro-type-reinterpret-cast)
-                return detail::wyhash::hash(reinterpret_cast<uintptr_t>(ptr.get()));
-            }
-        };
-
-        template <typename Enum>
-        struct hash<Enum, typename std::enable_if<std::is_enum<Enum>::value>::type>
-        {
-            using is_avalanching = void;
-            auto operator()(Enum e) const noexcept -> uint64_t
-            {
-                using underlying = typename std::underlying_type_t<Enum>;
-                return detail::wyhash::hash(static_cast<underlying>(e));
-            }
-        };
-
-// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
-#define ANKERL_UNORDERED_DENSE_HASH_STATICCAST(T)                    \
-    template <>                                                      \
-    struct hash<T>                                                   \
-    {                                                                \
-        using is_avalanching = void;                                 \
-        auto operator()(T const &obj) const noexcept -> uint64_t     \
-        {                                                            \
-            return detail::wyhash::hash(static_cast<uint64_t>(obj)); \
-        }                                                            \
-    }
-
-#if defined(__GNUC__) && !defined(__clang__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wuseless-cast"
-#endif
-        // see https://en.cppreference.com/w/cpp/utility/hash
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(bool);
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(char);
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(signed char);
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(unsigned char);
-#if ANKERL_UNORDERED_DENSE_CPP_VERSION >= 202002L && defined(__cpp_char8_t)
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(char8_t);
-#endif
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(char16_t);
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(char32_t);
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(wchar_t);
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(short);
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(unsigned short);
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(int);
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(unsigned int);
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(long);
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(long long);
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(unsigned long);
-        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(unsigned long long);
-
-#if defined(__GNUC__) && !defined(__clang__)
-#pragma GCC diagnostic pop
-#endif
-
-        // bucket_type //////////////////////////////////////////////////////////
-
-        namespace bucket_type
-        {
-
-            struct standard
-            {
-                static constexpr uint32_t dist_inc = 1U << 8U;             // skip 1 byte fingerprint
-                static constexpr uint32_t fingerprint_mask = dist_inc - 1; // mask for 1 byte of fingerprint
-
-                uint32_t m_dist_and_fingerprint; // upper 3 byte: distance to original bucket. lower byte: fingerprint from hash
-                uint32_t m_value_idx;            // index into the m_values vector.
-            };
-
-            ANKERL_UNORDERED_DENSE_PACK(struct big {
-                static constexpr uint32_t dist_inc = 1U << 8U;             // skip 1 byte fingerprint
-                static constexpr uint32_t fingerprint_mask = dist_inc - 1; // mask for 1 byte of fingerprint
-
-                uint32_t m_dist_and_fingerprint; // upper 3 byte: distance to original bucket. lower byte: fingerprint from hash
-                size_t m_value_idx;              // index into the m_values vector.
-            });
-
-        } // namespace bucket_type
-
-        namespace detail
-        {
-
-            struct nonesuch
-            {
-            };
-
-            template <class Default, class AlwaysVoid, template <class...> class Op, class... Args>
-            struct detector
-            {
-                using value_t = std::false_type;
-                using type = Default;
-            };
-
-            template <class Default, template <class...> class Op, class... Args>
-            struct detector<Default, std::void_t<Op<Args...>>, Op, Args...>
-            {
-                using value_t = std::true_type;
-                using type = Op<Args...>;
-            };
-
-            template <template <class...> class Op, class... Args>
-            using is_detected = typename detail::detector<detail::nonesuch, void, Op, Args...>::value_t;
-
-            template <template <class...> class Op, class... Args>
-            constexpr bool is_detected_v = is_detected<Op, Args...>::value;
-
-            template <typename T>
-            using detect_avalanching = typename T::is_avalanching;
-
-            template <typename T>
-            using detect_is_transparent = typename T::is_transparent;
-
-            template <typename T>
-            using detect_iterator = typename T::iterator;
-
-            template <typename T>
-            using detect_reserve = decltype(std::declval<T &>().reserve(size_t{}));
-
-            // enable_if helpers
-
-            template <typename Mapped>
-            constexpr bool is_map_v = !std::is_void_v<Mapped>;
-
-            // clang-format off
-template <typename Hash, typename KeyEqual>
-constexpr bool is_transparent_v = is_detected_v<detect_is_transparent, Hash> && is_detected_v<detect_is_transparent, KeyEqual>;
-            // clang-format on
-
-            template <typename From, typename To1, typename To2>
-            constexpr bool is_neither_convertible_v = !std::is_convertible_v<From, To1> && !std::is_convertible_v<From, To2>;
-
-            template <typename T>
-            constexpr bool has_reserve = is_detected_v<detect_reserve, T>;
-
-            // base type for map has mapped_type
-            template <class T>
-            struct base_table_type_map
-            {
-                using mapped_type = T;
-            };
-
-            // base type for set doesn't have mapped_type
-            struct base_table_type_set
-            {
-            };
-
-        } // namespace detail
-
-        // Very much like std::deque, but faster for indexing (in most cases). As of now this doesn't implement the full std::vector
-        // API, but merely what's necessary to work as an underlying container for ankerl::unordered_dense::{map, set}.
-        // It allocates blocks of equal size and puts them into the m_blocks vector. That means it can grow simply by adding a new
-        // block to the back of m_blocks, and doesn't double its size like an std::vector. The disadvantage is that memory is not
-        // linear and thus there is one more indirection necessary for indexing.
-        template <typename T, typename Allocator = std::allocator<T>, size_t MaxSegmentSizeBytes = 4096>
-        class segmented_vector
-        {
-            template <bool IsConst>
-            class iter_t;
-
-        public:
-            using allocator_type = Allocator;
-            using pointer = typename std::allocator_traits<allocator_type>::pointer;
-            using const_pointer = typename std::allocator_traits<allocator_type>::const_pointer;
-            using difference_type = typename std::allocator_traits<allocator_type>::difference_type;
-            using value_type = T;
-            using size_type = std::size_t;
-            using reference = T &;
-            using const_reference = T const &;
-            using iterator = iter_t<false>;
-            using const_iterator = iter_t<true>;
-
-        private:
-            using vec_alloc = typename std::allocator_traits<Allocator>::template rebind_alloc<pointer>;
-            std::vector<pointer, vec_alloc> m_blocks{};
-            size_t m_size{};
-
-            // Calculates the maximum number for x in  (s << x) <= max_val
-            static constexpr auto num_bits_closest(size_t max_val, size_t s) -> size_t
-            {
-                auto f = size_t{0};
-                while (s << (f + 1) <= max_val)
-                {
-                    ++f;
-                }
-                return f;
-            }
-
-            using self_t = segmented_vector<T, Allocator, MaxSegmentSizeBytes>;
-            static constexpr auto num_bits = num_bits_closest(MaxSegmentSizeBytes, sizeof(T));
-            static constexpr auto num_elements_in_block = 1U << num_bits;
-            static constexpr auto mask = num_elements_in_block - 1U;
-
-            /**
-             * Iterator class doubles as const_iterator and iterator
-             */
-            template <bool IsConst>
-            class iter_t
-            {
-                using ptr_t = typename std::conditional_t<IsConst, segmented_vector::const_pointer const *, segmented_vector::pointer *>;
-                ptr_t m_data{};
-                size_t m_idx{};
-
-                template <bool B>
-                friend class iter_t;
-
-            public:
-                using difference_type = segmented_vector::difference_type;
-                using value_type = T;
-                using reference = typename std::conditional_t<IsConst, value_type const &, value_type &>;
-                using pointer = typename std::conditional_t<IsConst, segmented_vector::const_pointer, segmented_vector::pointer>;
-                using iterator_category = std::forward_iterator_tag;
-
-                iter_t() noexcept = default;
-
-                template <bool OtherIsConst, typename = typename std::enable_if<IsConst && !OtherIsConst>::type>
-                // NOLINTNEXTLINE(google-explicit-constructor,hicpp-explicit-conversions)
-                constexpr iter_t(iter_t<OtherIsConst> const &other) noexcept
-                    : m_data(other.m_data), m_idx(other.m_idx)
-                {
-                }
-
-                constexpr iter_t(ptr_t data, size_t idx) noexcept
-                    : m_data(data), m_idx(idx) {}
-
-                template <bool OtherIsConst, typename = typename std::enable_if<IsConst && !OtherIsConst>::type>
-                constexpr auto operator=(iter_t<OtherIsConst> const &other) noexcept -> iter_t &
-                {
-                    m_data = other.m_data;
-                    m_idx = other.m_idx;
-                    return *this;
-                }
-
-                constexpr auto operator++() noexcept -> iter_t &
-                {
-                    ++m_idx;
-                    return *this;
-                }
-
-                constexpr auto operator+(difference_type diff) noexcept -> iter_t
-                {
-                    return {m_data, static_cast<size_t>(static_cast<difference_type>(m_idx) + diff)};
-                }
-
-                template <bool OtherIsConst>
-                constexpr auto operator-(iter_t<OtherIsConst> const &other) noexcept -> difference_type
-                {
-                    return static_cast<difference_type>(m_idx) - static_cast<difference_type>(other.m_idx);
-                }
-
-                constexpr auto operator*() const noexcept -> reference
-                {
-                    return m_data[m_idx >> num_bits][m_idx & mask];
-                }
-
-                constexpr auto operator->() const noexcept -> pointer
-                {
-                    return &m_data[m_idx >> num_bits][m_idx & mask];
-                }
-
-                template <bool O>
-                constexpr auto operator==(iter_t<O> const &o) const noexcept -> bool
-                {
-                    return m_idx == o.m_idx;
-                }
-
-                template <bool O>
-                constexpr auto operator!=(iter_t<O> const &o) const noexcept -> bool
-                {
-                    return !(*this == o);
-                }
-            };
-
-            // slow path: need to allocate a new segment every once in a while
-            void increase_capacity()
-            {
-                auto ba = Allocator(m_blocks.get_allocator());
-                pointer block = std::allocator_traits<Allocator>::allocate(ba, num_elements_in_block);
-                m_blocks.push_back(block);
-            }
-
-            // Moves everything from other
-            void append_everything_from(segmented_vector &&other)
-            {
-                reserve(size() + other.size());
-                for (auto &&o : other)
-                {
-                    emplace_back(std::move(o));
-                }
-            }
-
-            // Copies everything from other
-            void append_everything_from(segmented_vector const &other)
-            {
-                reserve(size() + other.size());
-                for (auto const &o : other)
-                {
-                    emplace_back(o);
-                }
-            }
-
-            void dealloc()
-            {
-                auto ba = Allocator(m_blocks.get_allocator());
-                for (auto ptr : m_blocks)
-                {
-                    std::allocator_traits<Allocator>::deallocate(ba, ptr, num_elements_in_block);
-                }
-            }
-
-            [[nodiscard]] static constexpr auto calc_num_blocks_for_capacity(size_t capacity)
-            {
-                return (capacity + num_elements_in_block - 1U) / num_elements_in_block;
-            }
-
-        public:
-            segmented_vector() = default;
-
-            // NOLINTNEXTLINE(google-explicit-constructor,hicpp-explicit-conversions)
-            segmented_vector(Allocator alloc)
-                : m_blocks(vec_alloc(alloc)) {}
-
-            segmented_vector(segmented_vector &&other, Allocator alloc)
-                : segmented_vector(alloc)
-            {
-                *this = std::move(other);
-            }
-
-            segmented_vector(segmented_vector const &other, Allocator alloc)
-                : m_blocks(vec_alloc(alloc))
-            {
-                append_everything_from(other);
-            }
-
-            segmented_vector(segmented_vector &&other) noexcept
-                : segmented_vector(std::move(other), get_allocator()) {}
-
-            segmented_vector(segmented_vector const &other)
-            {
-                append_everything_from(other);
-            }
-
-            auto operator=(segmented_vector const &other) -> segmented_vector &
-            {
-                if (this == &other)
-                {
-                    return *this;
-                }
-                clear();
-                append_everything_from(other);
-                return *this;
-            }
-
-            auto operator=(segmented_vector &&other) noexcept -> segmented_vector &
-            {
-                clear();
-                dealloc();
-                if (other.get_allocator() == get_allocator())
-                {
-                    m_blocks = std::move(other.m_blocks);
-                    m_size = std::exchange(other.m_size, {});
-                }
-                else
-                {
-                    // make sure to construct with other's allocator!
-                    m_blocks = std::vector<pointer, vec_alloc>(vec_alloc(other.get_allocator()));
-                    append_everything_from(std::move(other));
-                }
-                return *this;
-            }
-
-            ~segmented_vector()
-            {
-                clear();
-                dealloc();
-            }
-
-            [[nodiscard]] constexpr auto size() const -> size_t
-            {
-                return m_size;
-            }
-
-            [[nodiscard]] constexpr auto capacity() const -> size_t
-            {
-                return m_blocks.size() * num_elements_in_block;
-            }
-
-            // Indexing is highly performance critical
-            [[nodiscard]] constexpr auto operator[](size_t i) const noexcept -> T const &
-            {
-                return m_blocks[i >> num_bits][i & mask];
-            }
-
-            [[nodiscard]] constexpr auto operator[](size_t i) noexcept -> T &
-            {
-                return m_blocks[i >> num_bits][i & mask];
-            }
-
-            [[nodiscard]] constexpr auto begin() -> iterator
-            {
-                return {m_blocks.data(), 0U};
-            }
-            [[nodiscard]] constexpr auto begin() const -> const_iterator
-            {
-                return {m_blocks.data(), 0U};
-            }
-            [[nodiscard]] constexpr auto cbegin() const -> const_iterator
-            {
-                return {m_blocks.data(), 0U};
-            }
-
-            [[nodiscard]] constexpr auto end() -> iterator
-            {
-                return {m_blocks.data(), m_size};
-            }
-            [[nodiscard]] constexpr auto end() const -> const_iterator
-            {
-                return {m_blocks.data(), m_size};
-            }
-            [[nodiscard]] constexpr auto cend() const -> const_iterator
-            {
-                return {m_blocks.data(), m_size};
-            }
-
-            [[nodiscard]] constexpr auto back() -> reference
-            {
-                return operator[](m_size - 1);
-            }
-            [[nodiscard]] constexpr auto back() const -> const_reference
-            {
-                return operator[](m_size - 1);
-            }
-
-            void pop_back()
-            {
-                back().~T();
-                --m_size;
-            }
-
-            [[nodiscard]] auto empty() const
-            {
-                return 0 == m_size;
-            }
-
-            void reserve(size_t new_capacity)
-            {
-                m_blocks.reserve(calc_num_blocks_for_capacity(new_capacity));
-                while (new_capacity > capacity())
-                {
-                    increase_capacity();
-                }
-            }
-
-            [[nodiscard]] auto get_allocator() const -> allocator_type
-            {
-                return allocator_type{m_blocks.get_allocator()};
-            }
-
-            template <class... Args>
-            auto emplace_back(Args &&...args) -> reference
-            {
-                if (m_size == capacity())
-                {
-                    increase_capacity();
-                }
-                auto *ptr = static_cast<void *>(&operator[](m_size));
-                auto &ref = *new (ptr) T(std::forward<Args>(args)...);
-                ++m_size;
-                return ref;
-            }
-
-            void clear()
-            {
-                if constexpr (!std::is_trivially_destructible_v<T>)
-                {
-                    for (size_t i = 0, s = size(); i < s; ++i)
-                    {
-                        operator[](i).~T();
-                    }
-                }
-                m_size = 0;
-            }
-
-            void shrink_to_fit()
-            {
-                auto ba = Allocator(m_blocks.get_allocator());
-                auto num_blocks_required = calc_num_blocks_for_capacity(m_size);
-                while (m_blocks.size() > num_blocks_required)
-                {
-                    std::allocator_traits<Allocator>::deallocate(ba, m_blocks.back(), num_elements_in_block);
-                    m_blocks.pop_back();
-                }
-                m_blocks.shrink_to_fit();
-            }
-        };
-
-        namespace detail
-        {
-
-            // This is it, the table. Doubles as map and set, and uses `void` for T when its used as a set.
-            template <class Key,
-                      class T, // when void, treat it as a set.
-                      class Hash,
-                      class KeyEqual,
-                      class AllocatorOrContainer,
-                      class Bucket,
-                      bool IsSegmented>
-            class table : public std::conditional_t<is_map_v<T>, base_table_type_map<T>, base_table_type_set>
-            {
-                using underlying_value_type = typename std::conditional_t<is_map_v<T>, std::pair<Key, T>, Key>;
-                using underlying_container_type = std::conditional_t<IsSegmented,
-                                                                     segmented_vector<underlying_value_type, AllocatorOrContainer>,
-                                                                     std::vector<underlying_value_type, AllocatorOrContainer>>;
-
-            public:
-                using value_container_type = std::
-                    conditional_t<is_detected_v<detect_iterator, AllocatorOrContainer>, AllocatorOrContainer, underlying_container_type>;
-
-            private:
-                using bucket_alloc =
-                    typename std::allocator_traits<typename value_container_type::allocator_type>::template rebind_alloc<Bucket>;
-                using bucket_alloc_traits = std::allocator_traits<bucket_alloc>;
-
-                static constexpr uint8_t initial_shifts = 64 - 3; // 2^(64-m_shift) number of buckets
-                static constexpr float default_max_load_factor = 0.8F;
-
-            public:
-                using key_type = Key;
-                using value_type = typename value_container_type::value_type;
-                using size_type = typename value_container_type::size_type;
-                using difference_type = typename value_container_type::difference_type;
-                using hasher = Hash;
-                using key_equal = KeyEqual;
-                using allocator_type = typename value_container_type::allocator_type;
-                using reference = typename value_container_type::reference;
-                using const_reference = typename value_container_type::const_reference;
-                using pointer = typename value_container_type::pointer;
-                using const_pointer = typename value_container_type::const_pointer;
-                using const_iterator = typename value_container_type::const_iterator;
-                using iterator = std::conditional_t<is_map_v<T>, typename value_container_type::iterator, const_iterator>;
-                using bucket_type = Bucket;
-
-            private:
-                using value_idx_type = decltype(Bucket::m_value_idx);
-                using dist_and_fingerprint_type = decltype(Bucket::m_dist_and_fingerprint);
-
-                static_assert(std::is_trivially_destructible_v<Bucket>, "assert there's no need to call destructor / std::destroy");
-                static_assert(std::is_trivially_copyable_v<Bucket>, "assert we can just memset / memcpy");
-
-                value_container_type m_values{}; // Contains all the key-value pairs in one densely stored container. No holes.
-                using bucket_pointer = typename std::allocator_traits<bucket_alloc>::pointer;
-                bucket_pointer m_buckets{};
-                size_t m_num_buckets = 0;
-                size_t m_max_bucket_capacity = 0;
-                float m_max_load_factor = default_max_load_factor;
-                Hash m_hash{};
-                KeyEqual m_equal{};
-                uint8_t m_shifts = initial_shifts;
-
-                [[nodiscard]] auto next(value_idx_type bucket_idx) const -> value_idx_type
-                {
-                    return ANKERL_UNORDERED_DENSE_UNLIKELY(bucket_idx + 1U == m_num_buckets)
-                               ? 0
-                               : static_cast<value_idx_type>(bucket_idx + 1U);
-                }
-
-                // Helper to access bucket through pointer types
-                [[nodiscard]] static constexpr auto at(bucket_pointer bucket_ptr, size_t offset) -> Bucket &
-                {
-                    return *(bucket_ptr + static_cast<typename std::allocator_traits<bucket_alloc>::difference_type>(offset));
-                }
-
-                // use the dist_inc and dist_dec functions so that uint16_t types work without warning
-                [[nodiscard]] static constexpr auto dist_inc(dist_and_fingerprint_type x) -> dist_and_fingerprint_type
-                {
-                    return static_cast<dist_and_fingerprint_type>(x + Bucket::dist_inc);
-                }
-
-                [[nodiscard]] static constexpr auto dist_dec(dist_and_fingerprint_type x) -> dist_and_fingerprint_type
-                {
-                    return static_cast<dist_and_fingerprint_type>(x - Bucket::dist_inc);
-                }
-
-                // The goal of mixed_hash is to always produce a high quality 64bit hash.
-                template <typename K>
-                [[nodiscard]] constexpr auto mixed_hash(K const &key) const -> uint64_t
-                {
-                    if constexpr (is_detected_v<detect_avalanching, Hash>)
-                    {
-                        // we know that the hash is good because is_avalanching.
-                        if constexpr (sizeof(decltype(m_hash(key))) < sizeof(uint64_t))
-                        {
-                            // 32bit hash and is_avalanching => multiply with a constant to avalanche bits upwards
-                            return m_hash(key) * UINT64_C(0x9ddfea08eb382d69);
-                        }
-                        else
-                        {
-                            // 64bit and is_avalanching => only use the hash itself.
-                            return m_hash(key);
-                        }
-                    }
-                    else
-                    {
-                        // not is_avalanching => apply wyhash
-                        return wyhash::hash(m_hash(key));
-                    }
-                }
-
-                [[nodiscard]] constexpr auto dist_and_fingerprint_from_hash(uint64_t hash) const -> dist_and_fingerprint_type
-                {
-                    return Bucket::dist_inc | (static_cast<dist_and_fingerprint_type>(hash) & Bucket::fingerprint_mask);
-                }
-
-                [[nodiscard]] constexpr auto bucket_idx_from_hash(uint64_t hash) const -> value_idx_type
-                {
-                    return static_cast<value_idx_type>(hash >> m_shifts);
-                }
-
-                [[nodiscard]] static constexpr auto get_key(value_type const &vt) -> key_type const &
-                {
-                    if constexpr (is_map_v<T>)
-                    {
-                        return vt.first;
-                    }
-                    else
-                    {
-                        return vt;
-                    }
-                }
-
-                template <typename K>
-                [[nodiscard]] auto next_while_less(K const &key) const -> Bucket
-                {
-                    auto hash = mixed_hash(key);
-                    auto dist_and_fingerprint = dist_and_fingerprint_from_hash(hash);
-                    auto bucket_idx = bucket_idx_from_hash(hash);
-
-                    while (dist_and_fingerprint < at(m_buckets, bucket_idx).m_dist_and_fingerprint)
-                    {
-                        dist_and_fingerprint = dist_inc(dist_and_fingerprint);
-                        bucket_idx = next(bucket_idx);
-                    }
-                    return {dist_and_fingerprint, bucket_idx};
-                }
-
-                void place_and_shift_up(Bucket bucket, value_idx_type place)
-                {
-                    while (0 != at(m_buckets, place).m_dist_and_fingerprint)
-                    {
-                        bucket = std::exchange(at(m_buckets, place), bucket);
-                        bucket.m_dist_and_fingerprint = dist_inc(bucket.m_dist_and_fingerprint);
-                        place = next(place);
-                    }
-                    at(m_buckets, place) = bucket;
-                }
-
-                [[nodiscard]] static constexpr auto calc_num_buckets(uint8_t shifts) -> size_t
-                {
-                    return (std::min)(max_bucket_count(), size_t{1} << (64U - shifts));
-                }
-
-                [[nodiscard]] constexpr auto calc_shifts_for_size(size_t s) const -> uint8_t
-                {
-                    auto shifts = initial_shifts;
-                    while (shifts > 0 && static_cast<size_t>(static_cast<float>(calc_num_buckets(shifts)) * max_load_factor()) < s)
-                    {
-                        --shifts;
-                    }
-                    return shifts;
-                }
-
-                // assumes m_values has data, m_buckets=m_buckets_end=nullptr, m_shifts is INITIAL_SHIFTS
-                void copy_buckets(table const &other)
-                {
-                    if (!empty())
-                    {
-                        m_shifts = other.m_shifts;
-                        allocate_buckets_from_shift();
-                        std::memcpy(m_buckets, other.m_buckets, sizeof(Bucket) * bucket_count());
-                    }
-                }
-
-                /**
-                 * True when no element can be added any more without increasing the size
-                 */
-                [[nodiscard]] auto is_full() const -> bool
-                {
-                    return size() >= m_max_bucket_capacity;
-                }
-
-                void deallocate_buckets()
-                {
-                    auto ba = bucket_alloc(m_values.get_allocator());
-                    if (nullptr != m_buckets)
-                    {
-                        bucket_alloc_traits::deallocate(ba, m_buckets, bucket_count());
-                        m_buckets = nullptr;
-                    }
-                    m_num_buckets = 0;
-                    m_max_bucket_capacity = 0;
-                }
-
-                void allocate_buckets_from_shift()
-                {
-                    auto ba = bucket_alloc(m_values.get_allocator());
-                    m_num_buckets = calc_num_buckets(m_shifts);
-                    m_buckets = bucket_alloc_traits::allocate(ba, m_num_buckets);
-                    if (m_num_buckets == max_bucket_count())
-                    {
-                        // reached the maximum, make sure we can use each bucket
-                        m_max_bucket_capacity = max_bucket_count();
-                    }
-                    else
-                    {
-                        m_max_bucket_capacity = static_cast<value_idx_type>(static_cast<float>(m_num_buckets) * max_load_factor());
-                    }
-                }
-
-                void clear_buckets()
-                {
-                    if (m_buckets != nullptr)
-                    {
-                        std::memset(&*m_buckets, 0, sizeof(Bucket) * bucket_count());
-                    }
-                }
-
-                void clear_and_fill_buckets_from_values()
-                {
-                    clear_buckets();
-                    for (value_idx_type value_idx = 0, end_idx = static_cast<value_idx_type>(m_values.size()); value_idx < end_idx;
-                         ++value_idx)
-                    {
-                        auto const &key = get_key(m_values[value_idx]);
-                        auto [dist_and_fingerprint, bucket] = next_while_less(key);
-
-                        // we know for certain that key has not yet been inserted, so no need to check it.
-                        place_and_shift_up({dist_and_fingerprint, value_idx}, bucket);
-                    }
-                }
-
-                void increase_size()
-                {
-                    if (ANKERL_UNORDERED_DENSE_UNLIKELY(m_max_bucket_capacity == max_bucket_count()))
-                    {
-                        on_error_bucket_overflow();
-                    }
-                    --m_shifts;
-                    deallocate_buckets();
-                    allocate_buckets_from_shift();
-                    clear_and_fill_buckets_from_values();
-                }
-
-                void do_erase(value_idx_type bucket_idx)
-                {
-                    auto const value_idx_to_remove = at(m_buckets, bucket_idx).m_value_idx;
-
-                    // shift down until either empty or an element with correct spot is found
-                    auto next_bucket_idx = next(bucket_idx);
-                    while (at(m_buckets, next_bucket_idx).m_dist_and_fingerprint >= Bucket::dist_inc * 2)
-                    {
-                        at(m_buckets, bucket_idx) = {dist_dec(at(m_buckets, next_bucket_idx).m_dist_and_fingerprint),
-                                                     at(m_buckets, next_bucket_idx).m_value_idx};
-                        bucket_idx = std::exchange(next_bucket_idx, next(next_bucket_idx));
-                    }
-                    at(m_buckets, bucket_idx) = {};
-
-                    // update m_values
-                    if (value_idx_to_remove != m_values.size() - 1)
-                    {
-                        // no luck, we'll have to replace the value with the last one and update the index accordingly
-                        auto &val = m_values[value_idx_to_remove];
-                        val = std::move(m_values.back());
-
-                        // update the values_idx of the moved entry. No need to play the info game, just look until we find the values_idx
-                        auto mh = mixed_hash(get_key(val));
-                        bucket_idx = bucket_idx_from_hash(mh);
-
-                        auto const values_idx_back = static_cast<value_idx_type>(m_values.size() - 1);
-                        while (values_idx_back != at(m_buckets, bucket_idx).m_value_idx)
-                        {
-                            bucket_idx = next(bucket_idx);
-                        }
-                        at(m_buckets, bucket_idx).m_value_idx = value_idx_to_remove;
-                    }
-                    m_values.pop_back();
-                }
-
-                template <typename K>
-                auto do_erase_key(K &&key) -> size_t
-                {
-                    if (empty())
-                    {
-                        return 0;
-                    }
-
-                    auto [dist_and_fingerprint, bucket_idx] = next_while_less(key);
-
-                    while (dist_and_fingerprint == at(m_buckets, bucket_idx).m_dist_and_fingerprint &&
-                           !m_equal(key, get_key(m_values[at(m_buckets, bucket_idx).m_value_idx])))
-                    {
-                        dist_and_fingerprint = dist_inc(dist_and_fingerprint);
-                        bucket_idx = next(bucket_idx);
-                    }
-
-                    if (dist_and_fingerprint != at(m_buckets, bucket_idx).m_dist_and_fingerprint)
-                    {
-                        return 0;
-                    }
-                    do_erase(bucket_idx);
-                    return 1;
-                }
-
-                template <class K, class M>
-                auto do_insert_or_assign(K &&key, M &&mapped) -> std::pair<iterator, bool>
-                {
-                    auto it_isinserted = try_emplace(std::forward<K>(key), std::forward<M>(mapped));
-                    if (!it_isinserted.second)
-                    {
-                        it_isinserted.first->second = std::forward<M>(mapped);
-                    }
-                    return it_isinserted;
-                }
-
-                template <typename K, typename... Args>
-                auto do_place_element(dist_and_fingerprint_type dist_and_fingerprint, value_idx_type bucket_idx, K &&key, Args &&...args)
-                    -> std::pair<iterator, bool>
-                {
-
-                    // emplace the new value. If that throws an exception, no harm done; index is still in a valid state
-                    m_values.emplace_back(std::piecewise_construct,
-                                          std::forward_as_tuple(std::forward<K>(key)),
-                                          std::forward_as_tuple(std::forward<Args>(args)...));
-
-                    // place element and shift up until we find an empty spot
-                    auto value_idx = static_cast<value_idx_type>(m_values.size() - 1);
-                    place_and_shift_up({dist_and_fingerprint, value_idx}, bucket_idx);
-                    return {begin() + static_cast<difference_type>(value_idx), true};
-                }
-
-                template <typename K, typename... Args>
-                auto do_try_emplace(K &&key, Args &&...args) -> std::pair<iterator, bool>
-                {
-                    if (ANKERL_UNORDERED_DENSE_UNLIKELY(is_full()))
-                    {
-                        increase_size();
-                    }
-
-                    auto hash = mixed_hash(key);
-                    auto dist_and_fingerprint = dist_and_fingerprint_from_hash(hash);
-                    auto bucket_idx = bucket_idx_from_hash(hash);
-
-                    while (true)
-                    {
-                        auto *bucket = &at(m_buckets, bucket_idx);
-                        if (dist_and_fingerprint == bucket->m_dist_and_fingerprint)
-                        {
-                            if (m_equal(key, get_key(m_values[bucket->m_value_idx])))
-                            {
-                                return {begin() + static_cast<difference_type>(bucket->m_value_idx), false};
-                            }
-                        }
-                        else if (dist_and_fingerprint > bucket->m_dist_and_fingerprint)
-                        {
-                            return do_place_element(dist_and_fingerprint, bucket_idx, std::forward<K>(key), std::forward<Args>(args)...);
-                        }
-                        dist_and_fingerprint = dist_inc(dist_and_fingerprint);
-                        bucket_idx = next(bucket_idx);
-                    }
-                }
-
-                template <typename K>
-                auto do_find(K const &key) -> iterator
-                {
-                    if (ANKERL_UNORDERED_DENSE_UNLIKELY(empty()))
-                    {
-                        return end();
-                    }
-
-                    auto mh = mixed_hash(key);
-                    auto dist_and_fingerprint = dist_and_fingerprint_from_hash(mh);
-                    auto bucket_idx = bucket_idx_from_hash(mh);
-                    auto *bucket = &at(m_buckets, bucket_idx);
-
-                    // unrolled loop. *Always* check a few directly, then enter the loop. This is faster.
-                    if (dist_and_fingerprint == bucket->m_dist_and_fingerprint && m_equal(key, get_key(m_values[bucket->m_value_idx])))
-                    {
-                        return begin() + static_cast<difference_type>(bucket->m_value_idx);
-                    }
-                    dist_and_fingerprint = dist_inc(dist_and_fingerprint);
-                    bucket_idx = next(bucket_idx);
-                    bucket = &at(m_buckets, bucket_idx);
-
-                    if (dist_and_fingerprint == bucket->m_dist_and_fingerprint && m_equal(key, get_key(m_values[bucket->m_value_idx])))
-                    {
-                        return begin() + static_cast<difference_type>(bucket->m_value_idx);
-                    }
-                    dist_and_fingerprint = dist_inc(dist_and_fingerprint);
-                    bucket_idx = next(bucket_idx);
-                    bucket = &at(m_buckets, bucket_idx);
-
-                    while (true)
-                    {
-                        if (dist_and_fingerprint == bucket->m_dist_and_fingerprint)
-                        {
-                            if (m_equal(key, get_key(m_values[bucket->m_value_idx])))
-                            {
-                                return begin() + static_cast<difference_type>(bucket->m_value_idx);
-                            }
-                        }
-                        else if (dist_and_fingerprint > bucket->m_dist_and_fingerprint)
-                        {
-                            return end();
-                        }
-                        dist_and_fingerprint = dist_inc(dist_and_fingerprint);
-                        bucket_idx = next(bucket_idx);
-                        bucket = &at(m_buckets, bucket_idx);
-                    }
-                }
-
-                template <typename K>
-                auto do_find(K const &key) const -> const_iterator
-                {
-                    return const_cast<table *>(this)->do_find(key); // NOLINT(cppcoreguidelines-pro-type-const-cast)
-                }
-
-                template <typename K, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
-                auto do_at(K const &key) -> Q &
-                {
-                    if (auto it = find(key); ANKERL_UNORDERED_DENSE_LIKELY(end() != it))
-                    {
-                        return it->second;
-                    }
-                    on_error_key_not_found();
-                }
-
-                template <typename K, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
-                auto do_at(K const &key) const -> Q const &
-                {
-                    return const_cast<table *>(this)->at(key); // NOLINT(cppcoreguidelines-pro-type-const-cast)
-                }
-
-            public:
-                table()
-                    : table(0) {}
-
-                explicit table(size_t bucket_count,
-                               Hash const &hash = Hash(),
-                               KeyEqual const &equal = KeyEqual(),
-                               allocator_type const &alloc_or_container = allocator_type())
-                    : m_values(alloc_or_container), m_hash(hash), m_equal(equal)
-                {
-                    if (0 != bucket_count)
-                    {
-                        reserve(bucket_count);
-                    }
-                }
-
-                table(size_t bucket_count, allocator_type const &alloc)
-                    : table(bucket_count, Hash(), KeyEqual(), alloc) {}
-
-                table(size_t bucket_count, Hash const &hash, allocator_type const &alloc)
-                    : table(bucket_count, hash, KeyEqual(), alloc) {}
-
-                explicit table(allocator_type const &alloc)
-                    : table(0, Hash(), KeyEqual(), alloc) {}
-
-                template <class InputIt>
-                table(InputIt first,
-                      InputIt last,
-                      size_type bucket_count = 0,
-                      Hash const &hash = Hash(),
-                      KeyEqual const &equal = KeyEqual(),
-                      allocator_type const &alloc = allocator_type())
-                    : table(bucket_count, hash, equal, alloc)
-                {
-                    insert(first, last);
-                }
-
-                template <class InputIt>
-                table(InputIt first, InputIt last, size_type bucket_count, allocator_type const &alloc)
-                    : table(first, last, bucket_count, Hash(), KeyEqual(), alloc) {}
-
-                template <class InputIt>
-                table(InputIt first, InputIt last, size_type bucket_count, Hash const &hash, allocator_type const &alloc)
-                    : table(first, last, bucket_count, hash, KeyEqual(), alloc) {}
-
-                table(table const &other)
-                    : table(other, other.m_values.get_allocator()) {}
-
-                table(table const &other, allocator_type const &alloc)
-                    : m_values(other.m_values, alloc), m_max_load_factor(other.m_max_load_factor), m_hash(other.m_hash), m_equal(other.m_equal)
-                {
-                    copy_buckets(other);
-                }
-
-                table(table &&other) noexcept
-                    : table(std::move(other), other.m_values.get_allocator()) {}
-
-                table(table &&other, allocator_type const &alloc) noexcept
-                    : m_values(alloc)
-                {
-                    *this = std::move(other);
-                }
-
-                table(std::initializer_list<value_type> ilist,
-                      size_t bucket_count = 0,
-                      Hash const &hash = Hash(),
-                      KeyEqual const &equal = KeyEqual(),
-                      allocator_type const &alloc = allocator_type())
-                    : table(bucket_count, hash, equal, alloc)
-                {
-                    insert(ilist);
-                }
-
-                table(std::initializer_list<value_type> ilist, size_type bucket_count, allocator_type const &alloc)
-                    : table(ilist, bucket_count, Hash(), KeyEqual(), alloc) {}
-
-                table(std::initializer_list<value_type> init, size_type bucket_count, Hash const &hash, allocator_type const &alloc)
-                    : table(init, bucket_count, hash, KeyEqual(), alloc) {}
-
-                ~table()
-                {
-                    if (nullptr != m_buckets)
-                    {
-                        auto ba = bucket_alloc(m_values.get_allocator());
-                        bucket_alloc_traits::deallocate(ba, m_buckets, bucket_count());
-                    }
-                }
-
-                auto operator=(table const &other) -> table &
-                {
-                    if (&other != this)
-                    {
-                        deallocate_buckets(); // deallocate before m_values is set (might have another allocator)
-                        m_values = other.m_values;
-                        m_max_load_factor = other.m_max_load_factor;
-                        m_hash = other.m_hash;
-                        m_equal = other.m_equal;
-                        m_shifts = initial_shifts;
-                        copy_buckets(other);
-                    }
-                    return *this;
-                }
-
-                auto operator=(table &&other) noexcept(
-                    noexcept(std::is_nothrow_move_assignable_v<value_container_type> && std::is_nothrow_move_assignable_v<Hash> &&
-                             std::is_nothrow_move_assignable_v<KeyEqual>)) -> table &
-                {
-                    if (&other != this)
-                    {
-                        deallocate_buckets(); // deallocate before m_values is set (might have another allocator)
-                        m_values = std::move(other.m_values);
-                        other.m_values.clear();
-
-                        // we can only reuse m_buckets when both maps have the same allocator!
-                        if (get_allocator() == other.get_allocator())
-                        {
-                            m_buckets = std::exchange(other.m_buckets, nullptr);
-                            m_num_buckets = std::exchange(other.m_num_buckets, 0);
-                            m_max_bucket_capacity = std::exchange(other.m_max_bucket_capacity, 0);
-                            m_shifts = std::exchange(other.m_shifts, initial_shifts);
-                            m_max_load_factor = std::exchange(other.m_max_load_factor, default_max_load_factor);
-                            m_hash = std::exchange(other.m_hash, {});
-                            m_equal = std::exchange(other.m_equal, {});
-                        }
-                        else
-                        {
-                            // set max_load_factor *before* copying the other's buckets, so we have the same
-                            // behavior
-                            m_max_load_factor = other.m_max_load_factor;
-
-                            // copy_buckets sets m_buckets, m_num_buckets, m_max_bucket_capacity, m_shifts
-                            copy_buckets(other);
-                            // clear's the other's buckets so other is now already usable.
-                            other.clear_buckets();
-                            m_hash = other.m_hash;
-                            m_equal = other.m_equal;
-                        }
-                        // map "other" is now already usable, it's empty.
-                    }
-                    return *this;
-                }
-
-                auto operator=(std::initializer_list<value_type> ilist) -> table &
-                {
-                    clear();
-                    insert(ilist);
-                    return *this;
-                }
-
-                auto get_allocator() const noexcept -> allocator_type
-                {
-                    return m_values.get_allocator();
-                }
-
-                // iterators //////////////////////////////////////////////////////////////
-
-                auto begin() noexcept -> iterator
-                {
-                    return m_values.begin();
-                }
-
-                auto begin() const noexcept -> const_iterator
-                {
-                    return m_values.begin();
-                }
-
-                auto cbegin() const noexcept -> const_iterator
-                {
-                    return m_values.cbegin();
-                }
-
-                auto end() noexcept -> iterator
-                {
-                    return m_values.end();
-                }
-
-                auto cend() const noexcept -> const_iterator
-                {
-                    return m_values.cend();
-                }
-
-                auto end() const noexcept -> const_iterator
-                {
-                    return m_values.end();
-                }
-
-                // capacity ///////////////////////////////////////////////////////////////
-
-                [[nodiscard]] auto empty() const noexcept -> bool
-                {
-                    return m_values.empty();
-                }
-
-                [[nodiscard]] auto size() const noexcept -> size_t
-                {
-                    return m_values.size();
-                }
-
-                [[nodiscard]] static constexpr auto max_size() noexcept -> size_t
-                {
-                    if constexpr ((std::numeric_limits<value_idx_type>::max)() == (std::numeric_limits<size_t>::max)())
-                    {
-                        return size_t{1} << (sizeof(value_idx_type) * 8 - 1);
-                    }
-                    else
-                    {
-                        return size_t{1} << (sizeof(value_idx_type) * 8);
-                    }
-                }
-
-                // modifiers //////////////////////////////////////////////////////////////
-
-                void clear()
-                {
-                    m_values.clear();
-                    clear_buckets();
-                }
-
-                auto insert(value_type const &value) -> std::pair<iterator, bool>
-                {
-                    return emplace(value);
-                }
-
-                auto insert(value_type &&value) -> std::pair<iterator, bool>
-                {
-                    return emplace(std::move(value));
-                }
-
-                template <class P, std::enable_if_t<std::is_constructible_v<value_type, P &&>, bool> = true>
-                auto insert(P &&value) -> std::pair<iterator, bool>
-                {
-                    return emplace(std::forward<P>(value));
-                }
-
-                auto insert(const_iterator /*hint*/, value_type const &value) -> iterator
-                {
-                    return insert(value).first;
-                }
-
-                auto insert(const_iterator /*hint*/, value_type &&value) -> iterator
-                {
-                    return insert(std::move(value)).first;
-                }
-
-                template <class P, std::enable_if_t<std::is_constructible_v<value_type, P &&>, bool> = true>
-                auto insert(const_iterator /*hint*/, P &&value) -> iterator
-                {
-                    return insert(std::forward<P>(value)).first;
-                }
-
-                template <class InputIt>
-                void insert(InputIt first, InputIt last)
-                {
-                    while (first != last)
-                    {
-                        insert(*first);
-                        ++first;
-                    }
-                }
-
-                void insert(std::initializer_list<value_type> ilist)
-                {
-                    insert(ilist.begin(), ilist.end());
-                }
-
-                // nonstandard API: *this is emptied.
-                // Also see "A Standard flat_map" https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2022/p0429r9.pdf
-                auto extract() && -> value_container_type
-                {
-                    return std::move(m_values);
-                }
-
-                // nonstandard API:
-                // Discards the internally held container and replaces it with the one passed. Erases non-unique elements.
-                auto replace(value_container_type &&container)
-                {
-                    if (ANKERL_UNORDERED_DENSE_UNLIKELY(container.size() > max_size()))
-                    {
-                        on_error_too_many_elements();
-                    }
-                    auto shifts = calc_shifts_for_size(container.size());
-                    if (0 == m_num_buckets || shifts < m_shifts || container.get_allocator() != m_values.get_allocator())
-                    {
-                        m_shifts = shifts;
-                        deallocate_buckets();
-                        allocate_buckets_from_shift();
-                    }
-                    clear_buckets();
-
-                    m_values = std::move(container);
-
-                    // can't use clear_and_fill_buckets_from_values() because container elements might not be unique
-                    auto value_idx = value_idx_type{};
-
-                    // loop until we reach the end of the container. duplicated entries will be replaced with back().
-                    while (value_idx != static_cast<value_idx_type>(m_values.size()))
-                    {
-                        auto const &key = get_key(m_values[value_idx]);
-
-                        auto hash = mixed_hash(key);
-                        auto dist_and_fingerprint = dist_and_fingerprint_from_hash(hash);
-                        auto bucket_idx = bucket_idx_from_hash(hash);
-
-                        bool key_found = false;
-                        while (true)
-                        {
-                            auto const &bucket = at(m_buckets, bucket_idx);
-                            if (dist_and_fingerprint > bucket.m_dist_and_fingerprint)
-                            {
-                                break;
-                            }
-                            if (dist_and_fingerprint == bucket.m_dist_and_fingerprint &&
-                                m_equal(key, get_key(m_values[bucket.m_value_idx])))
-                            {
-                                key_found = true;
-                                break;
-                            }
-                            dist_and_fingerprint = dist_inc(dist_and_fingerprint);
-                            bucket_idx = next(bucket_idx);
-                        }
-
-                        if (key_found)
-                        {
-                            if (value_idx != static_cast<value_idx_type>(m_values.size() - 1))
-                            {
-                                m_values[value_idx] = std::move(m_values.back());
-                            }
-                            m_values.pop_back();
-                        }
-                        else
-                        {
-                            place_and_shift_up({dist_and_fingerprint, value_idx}, bucket_idx);
-                            ++value_idx;
-                        }
-                    }
-                }
-
-                template <class M, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
-                auto insert_or_assign(Key const &key, M &&mapped) -> std::pair<iterator, bool>
-                {
-                    return do_insert_or_assign(key, std::forward<M>(mapped));
-                }
-
-                template <class M, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
-                auto insert_or_assign(Key &&key, M &&mapped) -> std::pair<iterator, bool>
-                {
-                    return do_insert_or_assign(std::move(key), std::forward<M>(mapped));
-                }
-
-                template <typename K,
-                          typename M,
-                          typename Q = T,
-                          typename H = Hash,
-                          typename KE = KeyEqual,
-                          std::enable_if_t<is_map_v<Q> && is_transparent_v<H, KE>, bool> = true>
-                auto insert_or_assign(K &&key, M &&mapped) -> std::pair<iterator, bool>
-                {
-                    return do_insert_or_assign(std::forward<K>(key), std::forward<M>(mapped));
-                }
-
-                template <class M, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
-                auto insert_or_assign(const_iterator /*hint*/, Key const &key, M &&mapped) -> iterator
-                {
-                    return do_insert_or_assign(key, std::forward<M>(mapped)).first;
-                }
-
-                template <class M, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
-                auto insert_or_assign(const_iterator /*hint*/, Key &&key, M &&mapped) -> iterator
-                {
-                    return do_insert_or_assign(std::move(key), std::forward<M>(mapped)).first;
-                }
-
-                template <typename K,
-                          typename M,
-                          typename Q = T,
-                          typename H = Hash,
-                          typename KE = KeyEqual,
-                          std::enable_if_t<is_map_v<Q> && is_transparent_v<H, KE>, bool> = true>
-                auto insert_or_assign(const_iterator /*hint*/, K &&key, M &&mapped) -> iterator
-                {
-                    return do_insert_or_assign(std::forward<K>(key), std::forward<M>(mapped)).first;
-                }
-
-                // Single arguments for unordered_set can be used without having to construct the value_type
-                template <class K,
-                          typename Q = T,
-                          typename H = Hash,
-                          typename KE = KeyEqual,
-                          std::enable_if_t<!is_map_v<Q> && is_transparent_v<H, KE>, bool> = true>
-                auto emplace(K &&key) -> std::pair<iterator, bool>
-                {
-                    if (is_full())
-                    {
-                        increase_size();
-                    }
-
-                    auto hash = mixed_hash(key);
-                    auto dist_and_fingerprint = dist_and_fingerprint_from_hash(hash);
-                    auto bucket_idx = bucket_idx_from_hash(hash);
-
-                    while (dist_and_fingerprint <= at(m_buckets, bucket_idx).m_dist_and_fingerprint)
-                    {
-                        if (dist_and_fingerprint == at(m_buckets, bucket_idx).m_dist_and_fingerprint &&
-                            m_equal(key, m_values[at(m_buckets, bucket_idx).m_value_idx]))
-                        {
-                            // found it, return without ever actually creating anything
-                            return {begin() + static_cast<difference_type>(at(m_buckets, bucket_idx).m_value_idx), false};
-                        }
-                        dist_and_fingerprint = dist_inc(dist_and_fingerprint);
-                        bucket_idx = next(bucket_idx);
-                    }
-
-                    // value is new, insert element first, so when exception happens we are in a valid state
-                    m_values.emplace_back(std::forward<K>(key));
-                    // now place the bucket and shift up until we find an empty spot
-                    auto value_idx = static_cast<value_idx_type>(m_values.size() - 1);
-                    place_and_shift_up({dist_and_fingerprint, value_idx}, bucket_idx);
-                    return {begin() + static_cast<difference_type>(value_idx), true};
-                }
-
-                template <class... Args>
-                auto emplace(Args &&...args) -> std::pair<iterator, bool>
-                {
-                    if (is_full())
-                    {
-                        increase_size();
-                    }
-
-                    // we have to instantiate the value_type to be able to access the key.
-                    // 1. emplace_back the object so it is constructed. 2. If the key is already there, pop it later in the loop.
-                    auto &key = get_key(m_values.emplace_back(std::forward<Args>(args)...));
-                    auto hash = mixed_hash(key);
-                    auto dist_and_fingerprint = dist_and_fingerprint_from_hash(hash);
-                    auto bucket_idx = bucket_idx_from_hash(hash);
-
-                    while (dist_and_fingerprint <= at(m_buckets, bucket_idx).m_dist_and_fingerprint)
-                    {
-                        if (dist_and_fingerprint == at(m_buckets, bucket_idx).m_dist_and_fingerprint &&
-                            m_equal(key, get_key(m_values[at(m_buckets, bucket_idx).m_value_idx])))
-                        {
-                            m_values.pop_back(); // value was already there, so get rid of it
-                            return {begin() + static_cast<difference_type>(at(m_buckets, bucket_idx).m_value_idx), false};
-                        }
-                        dist_and_fingerprint = dist_inc(dist_and_fingerprint);
-                        bucket_idx = next(bucket_idx);
-                    }
-
-                    // value is new, place the bucket and shift up until we find an empty spot
-                    auto value_idx = static_cast<value_idx_type>(m_values.size() - 1);
-                    place_and_shift_up({dist_and_fingerprint, value_idx}, bucket_idx);
-
-                    return {begin() + static_cast<difference_type>(value_idx), true};
-                }
-
-                template <class... Args>
-                auto emplace_hint(const_iterator /*hint*/, Args &&...args) -> iterator
-                {
-                    return emplace(std::forward<Args>(args)...).first;
-                }
-
-                template <class... Args, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
-                auto try_emplace(Key const &key, Args &&...args) -> std::pair<iterator, bool>
-                {
-                    return do_try_emplace(key, std::forward<Args>(args)...);
-                }
-
-                template <class... Args, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
-                auto try_emplace(Key &&key, Args &&...args) -> std::pair<iterator, bool>
-                {
-                    return do_try_emplace(std::move(key), std::forward<Args>(args)...);
-                }
-
-                template <class... Args, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
-                auto try_emplace(const_iterator /*hint*/, Key const &key, Args &&...args) -> iterator
-                {
-                    return do_try_emplace(key, std::forward<Args>(args)...).first;
-                }
-
-                template <class... Args, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
-                auto try_emplace(const_iterator /*hint*/, Key &&key, Args &&...args) -> iterator
-                {
-                    return do_try_emplace(std::move(key), std::forward<Args>(args)...).first;
-                }
-
-                template <
-                    typename K,
-                    typename... Args,
-                    typename Q = T,
-                    typename H = Hash,
-                    typename KE = KeyEqual,
-                    std::enable_if_t<is_map_v<Q> && is_transparent_v<H, KE> && is_neither_convertible_v<K &&, iterator, const_iterator>,
-                                     bool> = true>
-                auto try_emplace(K &&key, Args &&...args) -> std::pair<iterator, bool>
-                {
-                    return do_try_emplace(std::forward<K>(key), std::forward<Args>(args)...);
-                }
-
-                template <
-                    typename K,
-                    typename... Args,
-                    typename Q = T,
-                    typename H = Hash,
-                    typename KE = KeyEqual,
-                    std::enable_if_t<is_map_v<Q> && is_transparent_v<H, KE> && is_neither_convertible_v<K &&, iterator, const_iterator>,
-                                     bool> = true>
-                auto try_emplace(const_iterator /*hint*/, K &&key, Args &&...args) -> iterator
-                {
-                    return do_try_emplace(std::forward<K>(key), std::forward<Args>(args)...).first;
-                }
-
-                auto erase(iterator it) -> iterator
-                {
-                    auto hash = mixed_hash(get_key(*it));
-                    auto bucket_idx = bucket_idx_from_hash(hash);
-
-                    auto const value_idx_to_remove = static_cast<value_idx_type>(it - cbegin());
-                    while (at(m_buckets, bucket_idx).m_value_idx != value_idx_to_remove)
-                    {
-                        bucket_idx = next(bucket_idx);
-                    }
-
-                    do_erase(bucket_idx);
-                    return begin() + static_cast<difference_type>(value_idx_to_remove);
-                }
-
-                template <typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
-                auto erase(const_iterator it) -> iterator
-                {
-                    return erase(begin() + (it - cbegin()));
-                }
-
-                auto erase(const_iterator first, const_iterator last) -> iterator
-                {
-                    auto const idx_first = first - cbegin();
-                    auto const idx_last = last - cbegin();
-                    auto const first_to_last = std::distance(first, last);
-                    auto const last_to_end = std::distance(last, cend());
-
-                    // remove elements from left to right which moves elements from the end back
-                    auto const mid = idx_first + (std::min)(first_to_last, last_to_end);
-                    auto idx = idx_first;
-                    while (idx != mid)
-                    {
-                        erase(begin() + idx);
-                        ++idx;
-                    }
-
-                    // all elements from the right are moved, now remove the last element until all done
-                    idx = idx_last;
-                    while (idx != mid)
-                    {
-                        --idx;
-                        erase(begin() + idx);
-                    }
-
-                    return begin() + idx_first;
-                }
-
-                auto erase(Key const &key) -> size_t
-                {
-                    return do_erase_key(key);
-                }
-
-                template <class K, class H = Hash, class KE = KeyEqual, std::enable_if_t<is_transparent_v<H, KE>, bool> = true>
-                auto erase(K &&key) -> size_t
-                {
-                    return do_erase_key(std::forward<K>(key));
-                }
-
-                void swap(table &other) noexcept(noexcept(std::is_nothrow_swappable_v<value_container_type> &&
-                                                          std::is_nothrow_swappable_v<Hash> && std::is_nothrow_swappable_v<KeyEqual>))
-                {
-                    using std::swap;
-                    swap(other, *this);
-                }
-
-                // lookup /////////////////////////////////////////////////////////////////
-
-                template <typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
-                auto at(key_type const &key) -> Q &
-                {
-                    return do_at(key);
-                }
-
-                template <typename K,
-                          typename Q = T,
-                          typename H = Hash,
-                          typename KE = KeyEqual,
-                          std::enable_if_t<is_map_v<Q> && is_transparent_v<H, KE>, bool> = true>
-                auto at(K const &key) -> Q &
-                {
-                    return do_at(key);
-                }
-
-                template <typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
-                auto at(key_type const &key) const -> Q const &
-                {
-                    return do_at(key);
-                }
-
-                template <typename K,
-                          typename Q = T,
-                          typename H = Hash,
-                          typename KE = KeyEqual,
-                          std::enable_if_t<is_map_v<Q> && is_transparent_v<H, KE>, bool> = true>
-                auto at(K const &key) const -> Q const &
-                {
-                    return do_at(key);
-                }
-
-                template <typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
-                auto operator[](Key const &key) -> Q &
-                {
-                    return try_emplace(key).first->second;
-                }
-
-                template <typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
-                auto operator[](Key &&key) -> Q &
-                {
-                    return try_emplace(std::move(key)).first->second;
-                }
-
-                template <typename K,
-                          typename Q = T,
-                          typename H = Hash,
-                          typename KE = KeyEqual,
-                          std::enable_if_t<is_map_v<Q> && is_transparent_v<H, KE>, bool> = true>
-                auto operator[](K &&key) -> Q &
-                {
-                    return try_emplace(std::forward<K>(key)).first->second;
-                }
-
-                auto count(Key const &key) const -> size_t
-                {
-                    return find(key) == end() ? 0 : 1;
-                }
-
-                template <class K, class H = Hash, class KE = KeyEqual, std::enable_if_t<is_transparent_v<H, KE>, bool> = true>
-                auto count(K const &key) const -> size_t
-                {
-                    return find(key) == end() ? 0 : 1;
-                }
-
-                auto find(Key const &key) -> iterator
-                {
-                    return do_find(key);
-                }
-
-                auto find(Key const &key) const -> const_iterator
-                {
-                    return do_find(key);
-                }
-
-                template <class K, class H = Hash, class KE = KeyEqual, std::enable_if_t<is_transparent_v<H, KE>, bool> = true>
-                auto find(K const &key) -> iterator
-                {
-                    return do_find(key);
-                }
-
-                template <class K, class H = Hash, class KE = KeyEqual, std::enable_if_t<is_transparent_v<H, KE>, bool> = true>
-                auto find(K const &key) const -> const_iterator
-                {
-                    return do_find(key);
-                }
-
-                auto contains(Key const &key) const -> bool
-                {
-                    return find(key) != end();
-                }
-
-                template <class K, class H = Hash, class KE = KeyEqual, std::enable_if_t<is_transparent_v<H, KE>, bool> = true>
-                auto contains(K const &key) const -> bool
-                {
-                    return find(key) != end();
-                }
-
-                auto equal_range(Key const &key) -> std::pair<iterator, iterator>
-                {
-                    auto it = do_find(key);
-                    return {it, it == end() ? end() : it + 1};
-                }
-
-                auto equal_range(const Key &key) const -> std::pair<const_iterator, const_iterator>
-                {
-                    auto it = do_find(key);
-                    return {it, it == end() ? end() : it + 1};
-                }
-
-                template <class K, class H = Hash, class KE = KeyEqual, std::enable_if_t<is_transparent_v<H, KE>, bool> = true>
-                auto equal_range(K const &key) -> std::pair<iterator, iterator>
-                {
-                    auto it = do_find(key);
-                    return {it, it == end() ? end() : it + 1};
-                }
-
-                template <class K, class H = Hash, class KE = KeyEqual, std::enable_if_t<is_transparent_v<H, KE>, bool> = true>
-                auto equal_range(K const &key) const -> std::pair<const_iterator, const_iterator>
-                {
-                    auto it = do_find(key);
-                    return {it, it == end() ? end() : it + 1};
-                }
-
-                // bucket interface ///////////////////////////////////////////////////////
-
-                auto bucket_count() const noexcept -> size_t
-                { // NOLINT(modernize-use-nodiscard)
-                    return m_num_buckets;
-                }
-
-                static constexpr auto max_bucket_count() noexcept -> size_t
-                { // NOLINT(modernize-use-nodiscard)
-                    return max_size();
-                }
-
-                // hash policy ////////////////////////////////////////////////////////////
-
-                [[nodiscard]] auto load_factor() const -> float
-                {
-                    return bucket_count() ? static_cast<float>(size()) / static_cast<float>(bucket_count()) : 0.0F;
-                }
-
-                [[nodiscard]] auto max_load_factor() const -> float
-                {
-                    return m_max_load_factor;
-                }
-
-                void max_load_factor(float ml)
-                {
-                    m_max_load_factor = ml;
-                    if (m_num_buckets != max_bucket_count())
-                    {
-                        m_max_bucket_capacity = static_cast<value_idx_type>(static_cast<float>(bucket_count()) * max_load_factor());
-                    }
-                }
-
-                void rehash(size_t count)
-                {
-                    count = (std::min)(count, max_size());
-                    auto shifts = calc_shifts_for_size((std::max)(count, size()));
-                    if (shifts != m_shifts)
-                    {
-                        m_shifts = shifts;
-                        deallocate_buckets();
-                        m_values.shrink_to_fit();
-                        allocate_buckets_from_shift();
-                        clear_and_fill_buckets_from_values();
-                    }
-                }
-
-                void reserve(size_t capa)
-                {
-                    capa = (std::min)(capa, max_size());
-                    if constexpr (has_reserve<value_container_type>)
-                    {
-                        // std::deque doesn't have reserve(). Make sure we only call when available
-                        m_values.reserve(capa);
-                    }
-                    auto shifts = calc_shifts_for_size((std::max)(capa, size()));
-                    if (0 == m_num_buckets || shifts < m_shifts)
-                    {
-                        m_shifts = shifts;
-                        deallocate_buckets();
-                        allocate_buckets_from_shift();
-                        clear_and_fill_buckets_from_values();
-                    }
-                }
-
-                // observers //////////////////////////////////////////////////////////////
-
-                auto hash_function() const -> hasher
-                {
-                    return m_hash;
-                }
-
-                auto key_eq() const -> key_equal
-                {
-                    return m_equal;
-                }
-
-                // nonstandard API: expose the underlying values container
-                [[nodiscard]] auto values() const noexcept -> value_container_type const &
-                {
-                    return m_values;
-                }
-
-                // non-member functions ///////////////////////////////////////////////////
-
-                friend auto operator==(table const &a, table const &b) -> bool
-                {
-                    if (&a == &b)
-                    {
-                        return true;
-                    }
-                    if (a.size() != b.size())
-                    {
-                        return false;
-                    }
-                    for (auto const &b_entry : b)
-                    {
-                        auto it = a.find(get_key(b_entry));
-                        if constexpr (is_map_v<T>)
-                        {
-                            // map: check that key is here, then also check that value is the same
-                            if (a.end() == it || !(b_entry.second == it->second))
-                            {
-                                return false;
-                            }
-                        }
-                        else
-                        {
-                            // set: only check that the key is here
-                            if (a.end() == it)
-                            {
-                                return false;
-                            }
-                        }
-                    }
-                    return true;
-                }
-
-                friend auto operator!=(table const &a, table const &b) -> bool
-                {
-                    return !(a == b);
-                }
-            };
-
-        } // namespace detail
-
-        ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
-                                                class T,
-                                                class Hash = hash<Key>,
-                                                class KeyEqual = std::equal_to<Key>,
-                                                class AllocatorOrContainer = std::allocator<std::pair<Key, T>>,
-                                                class Bucket = bucket_type::standard>
-        using map = detail::table<Key, T, Hash, KeyEqual, AllocatorOrContainer, Bucket, false>;
-
-        ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
-                                                class T,
-                                                class Hash = hash<Key>,
-                                                class KeyEqual = std::equal_to<Key>,
-                                                class AllocatorOrContainer = std::allocator<std::pair<Key, T>>,
-                                                class Bucket = bucket_type::standard>
-        using segmented_map = detail::table<Key, T, Hash, KeyEqual, AllocatorOrContainer, Bucket, true>;
-
-        ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
-                                                class Hash = hash<Key>,
-                                                class KeyEqual = std::equal_to<Key>,
-                                                class AllocatorOrContainer = std::allocator<Key>,
-                                                class Bucket = bucket_type::standard>
-        using set = detail::table<Key, void, Hash, KeyEqual, AllocatorOrContainer, Bucket, false>;
-
-        ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
-                                                class Hash = hash<Key>,
-                                                class KeyEqual = std::equal_to<Key>,
-                                                class AllocatorOrContainer = std::allocator<Key>,
-                                                class Bucket = bucket_type::standard>
-        using segmented_set = detail::table<Key, void, Hash, KeyEqual, AllocatorOrContainer, Bucket, true>;
-
-#if defined(ANKERL_UNORDERED_DENSE_PMR)
-
-        namespace pmr
-        {
-
-            ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
-                                                    class T,
-                                                    class Hash = hash<Key>,
-                                                    class KeyEqual = std::equal_to<Key>,
-                                                    class Bucket = bucket_type::standard>
-            using map =
-                detail::table<Key, T, Hash, KeyEqual, ANKERL_UNORDERED_DENSE_PMR::polymorphic_allocator<std::pair<Key, T>>, Bucket, false>;
-
-            ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
-                                                    class T,
-                                                    class Hash = hash<Key>,
-                                                    class KeyEqual = std::equal_to<Key>,
-                                                    class Bucket = bucket_type::standard>
-            using segmented_map =
-                detail::table<Key, T, Hash, KeyEqual, ANKERL_UNORDERED_DENSE_PMR::polymorphic_allocator<std::pair<Key, T>>, Bucket, true>;
-
-            ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
-                                                    class Hash = hash<Key>,
-                                                    class KeyEqual = std::equal_to<Key>,
-                                                    class Bucket = bucket_type::standard>
-            using set = detail::table<Key, void, Hash, KeyEqual, ANKERL_UNORDERED_DENSE_PMR::polymorphic_allocator<Key>, Bucket, false>;
-
-            ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
-                                                    class Hash = hash<Key>,
-                                                    class KeyEqual = std::equal_to<Key>,
-                                                    class Bucket = bucket_type::standard>
-            using segmented_set =
-                detail::table<Key, void, Hash, KeyEqual, ANKERL_UNORDERED_DENSE_PMR::polymorphic_allocator<Key>, Bucket, true>;
-
-        } // namespace pmr
-
-#endif
-
-        // deduction guides ///////////////////////////////////////////////////////////
-
-        // deduction guides for alias templates are only possible since C++20
-        // see https://en.cppreference.com/w/cpp/language/class_template_argument_deduction
-
-    } // namespace ANKERL_UNORDERED_DENSE_NAMESPACE
-} // namespace ankerl::unordered_dense
-
-// std extensions /////////////////////////////////////////////////////////////
-
-namespace std
-{ // NOLINT(cert-dcl58-cpp)
-
-    ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
-                                            class T,
-                                            class Hash,
-                                            class KeyEqual,
-                                            class AllocatorOrContainer,
-                                            class Bucket,
-                                            class Pred,
-                                            bool IsSegmented>
-    // NOLINTNEXTLINE(cert-dcl58-cpp)
-    auto erase_if(ankerl::unordered_dense::detail::table<Key, T, Hash, KeyEqual, AllocatorOrContainer, Bucket, IsSegmented> &map,
-                  Pred pred) -> size_t
-    {
-        using map_t = ankerl::unordered_dense::detail::table<Key, T, Hash, KeyEqual, AllocatorOrContainer, Bucket, IsSegmented>;
-
-        // going back to front because erase() invalidates the end iterator
-        auto const old_size = map.size();
-        auto idx = old_size;
-        while (idx)
-        {
-            --idx;
-            auto it = map.begin() + static_cast<typename map_t::difference_type>(idx);
-            if (pred(*it))
-            {
-                map.erase(it);
-            }
-        }
-
-        return old_size - map.size();
-    }
-
-} // namespace std
-
-#endif
-#endif
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner.hpp b/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner.hpp
index 94fa3c00..d8c78d84 100644
--- a/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner.hpp
@@ -61,6 +61,9 @@ class ax_runner_base
     int get_num_inputs() { return minput_tensors.size(); };
     int get_num_outputs() { return moutput_tensors.size(); };
 
+    int get_num_input_groups() { return mgroup_input_tensors.size(); };
+    int get_num_output_groups() { return mgroup_output_tensors.size(); };
+
     const ax_runner_tensor_t &get_input(int idx) { return minput_tensors[idx]; }
     const ax_runner_tensor_t *get_inputs_ptr() { return minput_tensors.data(); }
     const ax_runner_tensor_t &get_input(std::string name)
diff --git a/projects/llm_framework/main_vlm/src/runner/utils/http_utils.hpp b/projects/llm_framework/main_vlm/src/runner/utils/http_utils.hpp
new file mode 100644
index 00000000..5041b6a1
--- /dev/null
+++ b/projects/llm_framework/main_vlm/src/runner/utils/http_utils.hpp
@@ -0,0 +1,102 @@
+#pragma once
+#include <iostream>
+#include <string>
+#include <cstring>
+#include <sys/socket.h>
+#include <arpa/inet.h>
+#include <unistd.h>
+#include <chrono>
+#include <thread>
+
+/**
+ * @brief Attempts to establish a TCP connection to a specified host and port.
+ *
+ * This function creates a socket and tries to connect to a server specified by
+ * the host and port parameters. It returns true if the connection is successful,
+ * otherwise it returns false and outputs an error message to standard error.
+ *
+ * @param host The IP address of the server to connect to.
+ * @param port The port number of the server to connect to.
+ *
+ * @return true if the connection is successfully established, false otherwise.
+ */
+static bool test_connect(const std::string &host, int port)
+{
+    int sock = socket(AF_INET, SOCK_STREAM, 0);
+    if (sock < 0)
+    {
+        // std::cerr << "Socket creation failed\n";
+        return false;
+    }
+
+    sockaddr_in server_addr{};
+    server_addr.sin_family = AF_INET;
+    server_addr.sin_port = htons(port);
+
+    if (inet_pton(AF_INET, host.c_str(), &server_addr.sin_addr) <= 0)
+    {
+        // std::cerr << "IP address conversion failed\n";
+        close(sock);
+        return false;
+    }
+
+    if (connect(sock, (struct sockaddr *)&server_addr, sizeof(server_addr)) < 0)
+    {
+        // std::cerr << "Connection failed\n";
+        close(sock);
+        return false;
+    }
+
+    close(sock);
+    return true;
+}
+
+/**
+ * @brief Attempts to establish an HTTP connection to a specified URL with a timeout.
+ *
+ * This function parses the provided HTTP URL to extract the host and port information,
+ * and attempts to establish a TCP connection using the `test_connect` function.
+ * It retries the connection until the specified timeout is reached.
+ *
+ * @param http_url The HTTP URL of the server to connect to.
+ * @param timeout The maximum number of seconds to keep attempting the connection.
+ *
+ * @return true if the connection is successfully established within the timeout period,
+ * false otherwise.
+ */
+static bool test_connect_http(const std::string &http_url, int timeout)
+{
+    size_t pos = http_url.find("://");
+    if (pos == std::string::npos)
+        return false;
+    std::string host = http_url.substr(pos + 3);
+    pos = host.find('/');
+    if (pos != std::string::npos)
+        host = host.substr(0, pos);
+    pos = host.find(':');
+    int port = 80;
+    if (pos != std::string::npos)
+    {
+        port = std::stoi(host.substr(pos + 1));
+        host = host.substr(0, pos);
+    }
+    else
+    {
+        return false;
+    }
+
+    if (host == "localhost")
+        host = "127.0.0.1";
+    int tmp = timeout;
+    while (timeout--)
+    {
+        if (test_connect(host, port))
+            return true;
+        std::this_thread::sleep_for(std::chrono::seconds(1));
+        printf("\033[1;30;31m"
+               "connect failed %s, try again in %2d/%2d \n"
+               "\033[0m",
+               http_url.c_str(), timeout, tmp);
+    }
+    return false;
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_whisper/src/main.cpp b/projects/llm_framework/main_whisper/src/main.cpp
index cc3aaa96..51c6c6e1 100644
--- a/projects/llm_framework/main_whisper/src/main.cpp
+++ b/projects/llm_framework/main_whisper/src/main.cpp
@@ -94,12 +94,12 @@ class llm_task {
     std::atomic_bool superior_flage_;
     std::atomic_bool audio_flage_;
     std::atomic_bool awake_flage_;
-    std::atomic_bool endpoint_flage_;
+    std::atomic_bool endpoint_flage_ = true;
     std::string superior_id_;
     static int ax_init_flage_;
     task_callback_t out_callback_;
     int awake_delay_       = 50;
-    int delay_audio_frame_ = 1000;
+    int delay_audio_frame_ = 3001;
     buffer_t *pcmdata;
 
     std::function<void(void)> pause;
@@ -339,24 +339,24 @@ class llm_task {
         double start, end;
         double start_all, end_all;
         if (count < delay_audio_frame_) {
-            buffer_write_char(pcmdata, raw.c_str(), raw.length());
+            buffer_write_char(pcmdata, raw.data(), raw.length());
             count++;
             if (endpoint_flage_) return;
         }
         endpoint_flage_ = true;
-        if (delay_audio_frame_ == 0) buffer_resize(pcmdata, 0);
-        buffer_write_char(pcmdata, raw.c_str(), raw.length());
+
         buffer_position_set(pcmdata, 0);
-        count = 0;
+
         std::vector<float> floatSamples;
         {
             int16_t audio_val;
-            while (buffer_read_u16(pcmdata, (unsigned short *)&audio_val, 1)) {
-                float normalizedSample = (float)audio_val / INT16_MAX;
+            while (buffer_read_i16(pcmdata, &audio_val, 1)) {
+                float normalizedSample = static_cast<float>(audio_val) / INT16_MAX;
                 floatSamples.push_back(normalizedSample);
             }
         }
-        buffer_position_set(pcmdata, 0);
+        buffer_resize(pcmdata, 0);
+        count = 0;
 
         if (WHISPER_N_TEXT_STATE_MAP.find(mode_config_.model_type) == WHISPER_N_TEXT_STATE_MAP.end()) {
             fprintf(stderr, "Can NOT find n_text_state for model_type: %s\n", mode_config_.model_type.c_str());
@@ -773,9 +773,7 @@ class llm_whisper : public StackFlow {
         if (!(llm_task_obj && llm_channel)) {
             return;
         }
-        if (data == "true" || data == "false") {
-            llm_task_obj->endpoint_flage_ = (data == "true");
-        }
+        if (data == "false") llm_task_obj->endpoint_flage_ = false;
     }
 
     void work(const std::string &work_id, const std::string &object, const std::string &data) override
@@ -866,7 +864,6 @@ class llm_whisper : public StackFlow {
                                          std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
                                          std::placeholders::_2));
                 } else if (input.find("vad") != std::string::npos) {
-                    llm_task_obj->endpoint_flage_ = true;
                     // task_pause(work_id, "");
                     llm_channel->subscriber_work_id(
                         input, std::bind(&llm_whisper::vad_endpoint, this, std::weak_ptr<llm_task>(llm_task_obj),
@@ -917,11 +914,10 @@ class llm_whisper : public StackFlow {
                                              std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1, std::placeholders::_2));
             llm_task_obj->inputs_.push_back(data);
         } else if (data.find("vad") != std::string::npos) {
-            llm_task_obj->endpoint_flage_ = true;
-            ret                           = llm_channel->subscriber_work_id(
+            ret = llm_channel->subscriber_work_id(
                 data,
                 std::bind(&llm_whisper::vad_endpoint, this, std::weak_ptr<llm_task>(llm_task_obj),
-                                                    std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1, std::placeholders::_2));
+                          std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1, std::placeholders::_2));
         }
         if (ret) {
             error_body["code"]    = -20;

From 9a14d45799280d9558129a9357fdc340441bd2c7 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Wed, 27 Aug 2025 16:10:46 +0800
Subject: [PATCH 31/79] [update] support ax650. add ax650 model.

---
 ext_components/ax_msp/SConstruct              |  57 +++++++++
 projects/llm_framework/main_audio/SConstruct  |   6 +-
 .../llm_framework/main_audio/src/main.cpp     |  27 ++++-
 .../main_audio/src/sample_audio.c             |  27 ++++-
 .../main_audio/src/sample_audio.h             |   5 +
 projects/llm_framework/main_camera/SConstruct |  10 +-
 .../main_camera/src/axera_camera.c            |  20 ++--
 .../llm_framework/main_camera/src/main.cpp    |  17 ++-
 .../mode_depth-anything-ax6350.json           |  24 ++++
 .../main_depth_anything/src/EngineWrapper.cpp |   2 +-
 .../models/mode_melotts-en-au-ax650.json      |  31 +++++
 .../models/mode_melotts-en-br-ax650.json      |  31 +++++
 .../models/mode_melotts-en-default-ax650.json |  31 +++++
 .../models/mode_melotts-en-india-ax650.json   |  31 +++++
 .../models/mode_melotts-en-us-ax650.json      |  31 +++++
 .../models/mode_melotts-es-es-ax650.json      |  29 +++++
 .../models/mode_melotts-ja-jp-ax650.json      |  31 +++++
 .../models/mode_melotts-zh-cn-ax650.json      |  31 +++++
 .../main_melotts/src/runner/EngineWrapper.cpp |   2 +-
 projects/llm_framework/main_skel/SConstruct   |   9 +-
 projects/llm_framework/main_skel/src/main.c   |   8 +-
 .../main_whisper/src/runner/EngineWrapper.cpp |   2 +-
 .../main_yolo/models/mode_yolo11n-ax650.json  | 110 ++++++++++++++++++
 .../models/mode_yolo11n-hand-pose-ax650.json  |  32 +++++
 .../models/mode_yolo11n-pose-ax650.json       |  32 +++++
 .../models/mode_yolo11n-seg-ax650.json        | 110 ++++++++++++++++++
 .../main_yolo/src/EngineWrapper.cpp           |   2 +-
 projects/llm_framework/tools/llm_pack.py      |  61 +++++++++-
 28 files changed, 769 insertions(+), 40 deletions(-)
 create mode 100644 projects/llm_framework/main_depth_anything/mode_depth-anything-ax6350.json
 create mode 100644 projects/llm_framework/main_melotts/models/mode_melotts-en-au-ax650.json
 create mode 100644 projects/llm_framework/main_melotts/models/mode_melotts-en-br-ax650.json
 create mode 100644 projects/llm_framework/main_melotts/models/mode_melotts-en-default-ax650.json
 create mode 100644 projects/llm_framework/main_melotts/models/mode_melotts-en-india-ax650.json
 create mode 100644 projects/llm_framework/main_melotts/models/mode_melotts-en-us-ax650.json
 create mode 100644 projects/llm_framework/main_melotts/models/mode_melotts-es-es-ax650.json
 create mode 100644 projects/llm_framework/main_melotts/models/mode_melotts-ja-jp-ax650.json
 create mode 100644 projects/llm_framework/main_melotts/models/mode_melotts-zh-cn-ax650.json
 create mode 100644 projects/llm_framework/main_yolo/models/mode_yolo11n-ax650.json
 create mode 100644 projects/llm_framework/main_yolo/models/mode_yolo11n-hand-pose-ax650.json
 create mode 100644 projects/llm_framework/main_yolo/models/mode_yolo11n-pose-ax650.json
 create mode 100644 projects/llm_framework/main_yolo/models/mode_yolo11n-seg-ax650.json

diff --git a/ext_components/ax_msp/SConstruct b/ext_components/ax_msp/SConstruct
index 3c5ecbe3..c4e1ad0d 100644
--- a/ext_components/ax_msp/SConstruct
+++ b/ext_components/ax_msp/SConstruct
@@ -55,3 +55,60 @@ if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
                             'REGISTER':'static'
                             })
 
+if 'CONFIG_AX_650N_MSP_ENABLED' in os.environ:
+    MSP_PATH = ''
+    if 'CONFIG_MSP_V3_0_0_ENABLED' in os.environ:
+        MSP_PATH = os.environ.get('EXT_MSP_PATH', check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_ax650n_msp.tar.gz", 'm5stack_ax650n_msp.tar.gz'))
+    
+    SRCS=[]
+    INCLUDE=[]
+    PRIVATE_INCLUDE=[]
+    REQUIREMENTS=[]
+    STATIC_LIB=[]
+    DYNAMIC_LIB=[]
+    DEFINITIONS=[]
+    DEFINITIONS_PRIVATE=['']
+    LDFLAGS=[]
+    LINK_SEARCH_PATH=[]
+    
+    env["MSP_PATH"] = MSP_PATH
+    # , os.path.join(MSP_PATH, 'sample/common')
+    INCLUDE += [os.path.join(MSP_PATH, 'out/include'), os.path.join(MSP_PATH, 'sample'), os.path.join(MSP_PATH, 'sample/rtsp')]
+    LINK_SEARCH_PATH += [os.path.join(MSP_PATH, 'out/lib')]
+    REQUIREMENTS += ['dl', 'm', 'pthread']
+    third_party = ['drm','faac','fdk-aac','ffmpeg','libsamplerate','mp4','opencv','opus','tinyalsa']
+    for dirn in third_party:
+        INCLUDE.append(os.path.join(MSP_PATH,'third-party',dirn,'include'))
+        LINK_SEARCH_PATH.append(os.path.join(MSP_PATH,'third-party',dirn,'lib/arm64/glibc'))
+    INCLUDE.append(os.path.join(MSP_PATH,'third-party/tinyalsa/include/tinyalsa'))
+    INCLUDE.append(os.path.join(MSP_PATH,'third-party/live/out/include/liveMedia'))
+    INCLUDE.append(os.path.join(MSP_PATH,'third-party/live/out/include/groupsock'))
+    INCLUDE.append(os.path.join(MSP_PATH,'third-party/live/out/include/UsageEnvironment'))
+    INCLUDE.append(os.path.join(MSP_PATH,'third-party/live/out/include/BasicUsageEnvironment'))
+    INCLUDE.append(os.path.join(MSP_PATH,'component/isp_proton/sensor/platform/tcm'))
+    LINK_SEARCH_PATH.append(os.path.join(MSP_PATH,'third-party/libsamplerate/lib'))
+    INCLUDE.append(os.path.join(MSP_PATH,'third-party/openssl/include'))
+    LINK_SEARCH_PATH.append(os.path.join(MSP_PATH,'third-party/openssl/arm64/glibc/lib'))
+    LINK_SEARCH_PATH.append(os.path.join(MSP_PATH,'third-party/tinyalsa/lib'))
+    LINK_SEARCH_PATH.append(os.path.join(MSP_PATH,'third-party/fdk-aac/lib'))
+    LINK_SEARCH_PATH.append(os.path.join(MSP_PATH,'third-party/opus/lib'))
+    LINK_SEARCH_PATH.append(os.path.join(MSP_PATH,'app/lib'))
+
+    if 'CONFIG_SAMPLE_COMMON_ENABLED' in os.environ:
+        if 'CONFIG_MSP_V3_0_0_ENABLED' in os.environ:
+            INCLUDE.append(os.path.join(MSP_PATH,'sample/common'))
+            SRCS += AGlob(os.path.join(MSP_PATH,'sample/common/*.c*'))
+
+    env['COMPONENTS'].append({'target':os.path.basename(env['component_dir']),
+                            'SRCS':SRCS,
+                            'INCLUDE':INCLUDE,
+                            'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
+                            'REQUIREMENTS':REQUIREMENTS,
+                            'STATIC_LIB':STATIC_LIB,
+                            'DYNAMIC_LIB':DYNAMIC_LIB,
+                            'DEFINITIONS':DEFINITIONS,
+                            'DEFINITIONS_PRIVATE':DEFINITIONS_PRIVATE,
+                            'LDFLAGS':LDFLAGS,
+                            'LINK_SEARCH_PATH':LINK_SEARCH_PATH,
+                            'REGISTER':'static'
+                            })
\ No newline at end of file
diff --git a/projects/llm_framework/main_audio/SConstruct b/projects/llm_framework/main_audio/SConstruct
index 18d7ca05..5491339b 100644
--- a/projects/llm_framework/main_audio/SConstruct
+++ b/projects/llm_framework/main_audio/SConstruct
@@ -22,9 +22,9 @@ DEFINITIONS += ['-std=c++17']
 LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
 # INCLUDE += [ADir('../include'), ADir('../include/cppzmq'), ADir('../include/libzmq')]
 LINK_SEARCH_PATH += [ADir('../static_lib')]
-
-REQUIREMENTS += ['ax_sys', 'ax_interpreter', 'ax_audio', 'ax_audio_3a', 'ax_fdk', 'ax_opus']
-REQUIREMENTS += ['tinyalsa', 'opus', 'samplerate', 'fdk-aac']
+REQUIREMENTS += ['ax_sys', 'ax_audio', 'ax_audio_3a', 'tinyalsa', 'fdk-aac', 'opus', 'samplerate']
+if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
+    REQUIREMENTS += ['ax_interpreter', 'ax_fdk', 'ax_opus']
 
 STATIC_FILES += [AFile('audio.json'), AFile('audio_kit.json')]
 STATIC_FILES += Glob('mode_*.json')
diff --git a/projects/llm_framework/main_audio/src/main.cpp b/projects/llm_framework/main_audio/src/main.cpp
index e7007538..fdd25063 100644
--- a/projects/llm_framework/main_audio/src/main.cpp
+++ b/projects/llm_framework/main_audio/src/main.cpp
@@ -69,8 +69,13 @@ class llm_audio : public StackFlow {
 
     void hw_cap()
     {
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
         ax_cap_start(cap_config.card, cap_config.device, cap_config.volume, cap_config.channel, cap_config.rate,
                      cap_config.bit, llm_audio::on_cap_sample);
+#else
+        alsa_cap_start(cap_config.card, cap_config.device, cap_config.volume, cap_config.channel, cap_config.rate,
+                       cap_config.bit, llm_audio::on_cap_sample);
+#endif
     }
 
     void _play(const std::string &audio_data)
@@ -143,13 +148,17 @@ class llm_audio : public StackFlow {
         nlohmann::json error_body;
         std::string base_model_path;
         std::string base_model_config_path;
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
         std::list<std::string> config_file_paths;
         if (access("/sys/devices/platform/soc/4851000.i2c/i2c-1/1-0043", F_OK) == 0) {
             config_file_paths = get_config_file_paths(base_model_path, base_model_config_path, "audio_kit");
         } else {
             config_file_paths = get_config_file_paths(base_model_path, base_model_config_path, "audio");
         }
-
+#else
+        std::list<std::string> config_file_paths =
+            get_config_file_paths(base_model_path, base_model_config_path, "audio");
+#endif
         try {
             config_body = nlohmann::json::parse(data);
             for (auto file_name : config_file_paths) {
@@ -201,6 +210,7 @@ class llm_audio : public StackFlow {
                 CONFIG_AUTO_SET(file_body["play_param"], stVqeAttr.stAgcCfg.enAgcMode);
                 CONFIG_AUTO_SET(file_body["play_param"], stVqeAttr.stAgcCfg.s16TargetLevel);
                 CONFIG_AUTO_SET(file_body["play_param"], stVqeAttr.stAgcCfg.s16Gain);
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
                 CONFIG_AUTO_SET(file_body["play_param"], stHpfAttr.bEnable);
                 CONFIG_AUTO_SET(file_body["play_param"], stHpfAttr.s32GainDb);
                 CONFIG_AUTO_SET(file_body["play_param"], stHpfAttr.s32Freq);
@@ -215,6 +225,7 @@ class llm_audio : public StackFlow {
                 CONFIG_AUTO_SET(file_body["play_param"], stEqAttr.s32GainDb[3]);
                 CONFIG_AUTO_SET(file_body["play_param"], stEqAttr.s32GainDb[4]);
                 CONFIG_AUTO_SET(file_body["play_param"], stEqAttr.s32Samplerate);
+#endif
                 CONFIG_AUTO_SET(file_body["play_param"], gResample);
                 CONFIG_AUTO_SET(file_body["play_param"], enInSampleRate);
                 CONFIG_AUTO_SET(file_body["play_param"], gInstant);
@@ -247,7 +258,17 @@ class llm_audio : public StackFlow {
                 CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.enBitwidth);
                 CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.enLinkMode);
                 CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.enSamplerate);
-                CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.enLayoutMode);
+                // CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.enLayoutMode);
+                if (config_body.contains("aistAttr.enLayoutMode"))
+                    mode_config_.aistAttr.enLayoutMode = config_body["aistAttr.enLayoutMode"];
+                else if (file_body["cap_param"].contains("aistAttr.enLayoutMode")) {
+                    mode_config_.aistAttr.enLayoutMode = file_body["cap_param"]["aistAttr.enLayoutMode"];
+                    if (access("/sys/devices/platform/soc/4851000.i2c/i2c-1/1-0043", F_OK) == 0) {
+                        if (mode_config_.aistAttr.enLayoutMode == AX_AI_REF_MIC) {
+                            mode_config_.aistAttr.enLayoutMode = AX_AI_MIC_REF;
+                        }
+                    }
+                }
                 CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.U32Depth);
                 CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.u32PeriodSize);
                 CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.u32PeriodCount);
@@ -261,6 +282,7 @@ class llm_audio : public StackFlow {
                 CONFIG_AUTO_SET(file_body["cap_param"], aistVqeAttr.stAgcCfg.s16TargetLevel);
                 CONFIG_AUTO_SET(file_body["cap_param"], aistVqeAttr.stAgcCfg.s16Gain);
                 CONFIG_AUTO_SET(file_body["cap_param"], aistVqeAttr.stAecCfg.enAecMode);
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
                 CONFIG_AUTO_SET(file_body["cap_param"], stHpfAttr.bEnable);
                 CONFIG_AUTO_SET(file_body["cap_param"], stHpfAttr.s32GainDb);
                 CONFIG_AUTO_SET(file_body["cap_param"], stHpfAttr.s32Samplerate);
@@ -276,6 +298,7 @@ class llm_audio : public StackFlow {
                 CONFIG_AUTO_SET(file_body["cap_param"], stEqAttr.s32GainDb[3]);
                 CONFIG_AUTO_SET(file_body["cap_param"], stEqAttr.s32GainDb[4]);
                 CONFIG_AUTO_SET(file_body["cap_param"], stEqAttr.s32Samplerate);
+#endif
                 CONFIG_AUTO_SET(file_body["cap_param"], gResample);
                 CONFIG_AUTO_SET(file_body["cap_param"], enOutSampleRate);
                 CONFIG_AUTO_SET(file_body["cap_param"], gDbDetection);
diff --git a/projects/llm_framework/main_audio/src/sample_audio.c b/projects/llm_framework/main_audio/src/sample_audio.c
index 400d46be..ab78b6c6 100644
--- a/projects/llm_framework/main_audio/src/sample_audio.c
+++ b/projects/llm_framework/main_audio/src/sample_audio.c
@@ -116,6 +116,7 @@ void ax_play(unsigned int card, unsigned int device, float Volume, int channel,
             goto AO_DEINIT;
         }
     }
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
     if (play_config.stHpfAttr.bEnable) {
         ret = AX_ACODEC_TxHpfSetAttr(card, &play_config.stHpfAttr);
         if (ret) {
@@ -152,10 +153,14 @@ void ax_play(unsigned int card, unsigned int device, float Volume, int channel,
             goto DIS_LPF;
         }
     }
+#endif
     ret = AX_AO_EnableDev(card, device);
     if (ret) {
         printf("AX_AO_EnableDev failed! ret = %x \n", ret);
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
         goto DIS_EQ;
+#endif
+        goto DIS_AO_DEVICE;
     }
     if (play_config.gResample) {
         ret = AX_AO_EnableResample(card, device, play_config.enInSampleRate);
@@ -184,7 +189,11 @@ void ax_play(unsigned int card, unsigned int device, float Volume, int channel,
                 AX_POOL_ReleaseBlock(stFrmInfo.u32BlkId);
                 break;
             }
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
             ret = AX_AO_SendFrame(card, device, &stFrmInfo, NULL, 0.f, -1);
+#else
+            ret = AX_AO_SendFrame(card, device, &stFrmInfo, -1);
+#endif
             if (ret != AX_SUCCESS) {
                 AX_POOL_ReleaseBlock(stFrmInfo.u32BlkId);
                 printf("AX_AO_SendFrame error, ret: %x\n", ret);
@@ -205,6 +214,7 @@ void ax_play(unsigned int card, unsigned int device, float Volume, int channel,
         while (1) {
             ret = AX_AO_QueryDevStat(card, device, &stStatus);
             if (stStatus.u32DevBusyNum == 0) {
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
                 if (play_config.gInsertSilence) {
                     break;
                 } else {
@@ -212,11 +222,15 @@ void ax_play(unsigned int card, unsigned int device, float Volume, int channel,
                         break;
                     }
                 }
+#else
+                break;
+#endif
             }
             usleep(10 * 1000);
         }
     }
     printf("ao success.\n");
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
 DIS_EQ:
     if (play_config.stEqAttr.bEnable) {
         ret = AX_ACODEC_TxEqDisable(card);
@@ -238,7 +252,7 @@ void ax_play(unsigned int card, unsigned int device, float Volume, int channel,
             printf("AX_ACODEC_TxHpfDisable failed! ret= %x\n", ret);
         }
     }
-
+#endif
 DIS_AO_DEVICE:
     ret = AX_AO_DisableDev(card, device);
     if (ret) {
@@ -329,7 +343,7 @@ void ax_cap_start(unsigned int card, unsigned int device, float Volume, int chan
         }
         outRate = rate;
     }
-
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
     if (cap_config.stHpfAttr.bEnable) {
         ret = AX_ACODEC_RxHpfSetAttr(card, &cap_config.stHpfAttr);
         if (ret) {
@@ -366,11 +380,14 @@ void ax_cap_start(unsigned int card, unsigned int device, float Volume, int chan
             goto DIS_LPF;
         }
     }
-
+#endif
     ret = AX_AI_EnableDev(card, device);
     if (ret) {
         printf("AX_AI_EnableDev failed! ret = %x \n", ret);
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
         goto DIS_EQ;
+#endif
+        goto AI_DETACHPOOL;
     }
 
     if (cap_config.gResample) {
@@ -415,7 +432,7 @@ void ax_cap_start(unsigned int card, unsigned int device, float Volume, int chan
         printf("AX_AI_DisableDev failed! ret= %x\n", ret);
         goto FREE_SYS;
     }
-
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
 DIS_EQ:
     if (cap_config.stEqAttr.bEnable) {
         ret = AX_ACODEC_RxEqDisable(card);
@@ -437,7 +454,7 @@ void ax_cap_start(unsigned int card, unsigned int device, float Volume, int chan
             printf("AX_ACODEC_RxHpfDisable failed! ret= %x\n", ret);
         }
     }
-
+#endif
 AI_DETACHPOOL:
     ret = AX_AI_DetachPool(card, device);
     if (AX_SUCCESS != ret) {
diff --git a/projects/llm_framework/main_audio/src/sample_audio.h b/projects/llm_framework/main_audio/src/sample_audio.h
index 037b8b9f..29833fbc 100644
--- a/projects/llm_framework/main_audio/src/sample_audio.h
+++ b/projects/llm_framework/main_audio/src/sample_audio.h
@@ -7,8 +7,11 @@
 #include "ax_ai_api.h"
 #include "ax_ao_api.h"
 #include "ax_global_type.h"
+#include <global_config.h>
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
 #include "ax_acodec_api.h"
 #include "ax_aac.h"
+#endif
 
 typedef struct {
     unsigned int card;
@@ -26,9 +29,11 @@ typedef struct {
         AX_AP_DNVQE_ATTR_T stVqeAttr;
         AX_AP_UPTALKVQE_ATTR_T aistVqeAttr;
     };
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
     AX_ACODEC_FREQ_ATTR_T stHpfAttr;
     AX_ACODEC_FREQ_ATTR_T stLpfAttr;
     AX_ACODEC_EQ_ATTR_T stEqAttr;
+#endif
     int gResample;
     union {
         AX_AUDIO_SAMPLE_RATE_E enInSampleRate;
diff --git a/projects/llm_framework/main_camera/SConstruct b/projects/llm_framework/main_camera/SConstruct
index 92f525e9..a68311d9 100644
--- a/projects/llm_framework/main_camera/SConstruct
+++ b/projects/llm_framework/main_camera/SConstruct
@@ -41,11 +41,11 @@ SRCS += Glob(os.path.join(env["MSP_PATH"], 'sample/rtsp/*.c*') )
 # INCLUDE.append(ADir('ax_sample'))
 # SRCS += Glob('ax_sample/*.c*')
 
-REQUIREMENTS += ['ax_sys', 'ax_interpreter']
-REQUIREMENTS += ['ax_engine', 'ax_interpreter', 'ax_sys', 'ax_vo']
-REQUIREMENTS += ['ax_ae', 'ax_skel', 'ax_venc', 'ax_mipi', 'ax_ives', 'ax_ivps', 'ax_proton', 'ax_audio', 'tinyalsa']
-REQUIREMENTS += ['ax_audio_3a','samplerate', 'ax_fdk', 'fdk-aac', 'ax_awb', 'ax_af', 'ax_nt_stream', 'ax_nt_ctrl']
-REQUIREMENTS += ['liveMedia', 'UsageEnvironment', 'groupsock', 'BasicUsageEnvironment']
+REQUIREMENTS += ['ax_sys', 'ax_interpreter', 'ax_engine']
+if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
+    REQUIREMENTS += ['ax_ae', 'ax_vo', 'ax_skel', 'ax_venc', 'ax_mipi', 'ax_ives', 'ax_ivps', 'ax_proton', 'ax_audio', 'tinyalsa']
+    REQUIREMENTS += ['ax_audio_3a','samplerate', 'ax_fdk', 'fdk-aac', 'ax_awb', 'ax_af', 'ax_nt_stream', 'ax_nt_ctrl']
+    REQUIREMENTS += ['liveMedia', 'UsageEnvironment', 'groupsock', 'BasicUsageEnvironment']
 
 INCLUDE += [ADir('../include/opencv4')]
 static_file = []
diff --git a/projects/llm_framework/main_camera/src/axera_camera.c b/projects/llm_framework/main_camera/src/axera_camera.c
index 00172883..f6a02cb2 100644
--- a/projects/llm_framework/main_camera/src/axera_camera.c
+++ b/projects/llm_framework/main_camera/src/axera_camera.c
@@ -4,7 +4,6 @@
  * SPDX-License-Identifier: MIT
  */
 #include "camera.h"
-#include "axera_camera.h"
 #include <assert.h>
 #include <errno.h>
 #include <fcntl.h>
@@ -18,6 +17,9 @@
 #include <unistd.h>
 #include <dlfcn.h>
 #include <sys/prctl.h>
+#include <global_config.h>
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
+#include "axera_camera.h"
 #include "AXRtspWrapper.h"
 #include "ax_venc_api.h"
 #include "ax_global_type.h"
@@ -29,6 +31,9 @@
 #include "common_vin.h"
 #include "ax_sys_api.h"
 #include "ax_ivps_api.h"
+#endif
+
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
 
 #ifndef ALIGN_UP
 #define ALIGN_UP(x, a) ((((x) + ((a) - 1)) / a) * a)
@@ -115,8 +120,6 @@ AX_VIN_CHN_ATTR_T gSc850slChn0Attr = {
     .tFrameRateCtrl = {AX_INVALID_FRMRATE, AX_INVALID_FRMRATE},
 };
 
-
-
 struct axera_camera_index_t {
     char name[48];
     SAMPLE_VIN_CASE_E index;
@@ -137,8 +140,6 @@ struct axera_camera_index_t {
                           {"axera_single_os04a10_online", SYS_CASE_SINGLE_OS04A10_ONLINE},
                           {"axera_single_sc850sl", SAMPLE_VIN_SINGLE_SC850SL}};
 
-
-
 /* comm pool */
 COMMON_SYS_POOL_CFG_T gtSysCommPoolSingleDummySdr[] = {
     {2688, 1520, 2688, AX_FORMAT_BAYER_RAW_16BPP, 6},                              /*vin raw16 use */
@@ -1611,9 +1612,9 @@ int axera_camera_open_from(camera_t *camera)
         return -10;
     }
 
-    axera_obj.VinParam.eSysMode     = axera_config->VinParam.eSysMode; // COMMON_VIN_SENSOR;
-    axera_obj.VinParam.eHdrMode     = axera_config->VinParam.eHdrMode; // AX_SNS_LINEAR_MODE;
-    axera_obj.VinParam.bAiispEnable = axera_config->VinParam.bAiispEnable; // AX_TRUE;
+    axera_obj.VinParam.eSysMode     = axera_config->VinParam.eSysMode;      // COMMON_VIN_SENSOR;
+    axera_obj.VinParam.eHdrMode     = axera_config->VinParam.eHdrMode;      // AX_SNS_LINEAR_MODE;
+    axera_obj.VinParam.bAiispEnable = axera_config->VinParam.bAiispEnable;  // AX_TRUE;
     // axera_obj.gCams.tChnAttr
     __sample_case_config(&axera_obj.gCams, &axera_obj.VinParam, &axera_obj.tCommonArgs, &axera_obj.tPrivArgs);
     COMMON_SYS_Init(&axera_obj.tCommonArgs);
@@ -1720,4 +1721,5 @@ int axera_camera_close(camera_t *camera)
     SLOGI("camera closed");
 
     return 0;
-}
\ No newline at end of file
+}
+#endif
\ No newline at end of file
diff --git a/projects/llm_framework/main_camera/src/main.cpp b/projects/llm_framework/main_camera/src/main.cpp
index 3565b01d..45f730bc 100644
--- a/projects/llm_framework/main_camera/src/main.cpp
+++ b/projects/llm_framework/main_camera/src/main.cpp
@@ -13,7 +13,10 @@
 #include <iostream>
 #include "../../../../SDK/components/utilities/include/sample_log.h"
 #include "camera.h"
+#include <global_config.h>
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
 #include "axera_camera.h"
+#endif
 #include <glob.h>
 #include <opencv2/opencv.hpp>
 #include "hv/TcpServer.h"
@@ -172,7 +175,7 @@ class llm_task {
     {
         out_callback_ = out_callback;
     }
-
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
     static bool parse_axera_config(const nlohmann::json &config_body, const nlohmann::json &file_body,
                                    void **custom_config)
     {
@@ -473,11 +476,13 @@ class llm_task {
                 (stVencChnAttr.stVencAttr.u32PicHeightSrc < frame_height)) {
                 return true;
             }
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
             init_rtsp(&stVencChnAttr);
+#endif
         }
         return false;
     }
-
+#endif
     bool parse_config(const nlohmann::json &config_body)
     {
         try {
@@ -502,11 +507,15 @@ class llm_task {
             hal_camera_open  = camera_open;
             hal_camera_close = camera_close;
             hal_parse_config = NULL;
-        } else if (devname_.find("axera_") != std::string::npos) {
+        }
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
+        else if (devname_.find("axera_") != std::string::npos) {
             hal_camera_open  = axera_camera_open;
             hal_camera_close = axera_camera_close;
             hal_parse_config = llm_task::parse_axera_config;
-        } else {
+        }
+#endif
+        else {
             return true;
         }
         {
diff --git a/projects/llm_framework/main_depth_anything/mode_depth-anything-ax6350.json b/projects/llm_framework/main_depth_anything/mode_depth-anything-ax6350.json
new file mode 100644
index 00000000..447f6f41
--- /dev/null
+++ b/projects/llm_framework/main_depth_anything/mode_depth-anything-ax6350.json
@@ -0,0 +1,24 @@
+{
+    "mode":"depth-anything-ax650",
+    "type":"cv",
+    "homepage":"https://huggingface.co/AXERA-TECH/Depth-Anything-V2",
+    "compile_flage":"pulsar2 build --input depth_anything_v2_vits.onnx --config config.json --output_dir build-output --output_name depth_anything_v2_vits.axmodel --target_hardware AX650 --compiler.check 0",
+    "pulsar_version":"3.3-ae03a08f",
+    "capabilities":[
+        "Segmentation"
+    ],
+    "input_type":[
+        "cv.jpeg.base64"
+    ],
+    "output_type":[
+        "cv.jpeg.base64"
+    ],
+    "mode_param":{
+        "depth_anything_model":"depth_anything.axmodel",
+        "model_type":"segment",
+        "img_h":518,
+        "img_w":518
+    },
+    "mode_param_bak":{
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_depth_anything/src/EngineWrapper.cpp b/projects/llm_framework/main_depth_anything/src/EngineWrapper.cpp
index 21c51e49..1aad9678 100644
--- a/projects/llm_framework/main_depth_anything/src/EngineWrapper.cpp
+++ b/projects/llm_framework/main_depth_anything/src/EngineWrapper.cpp
@@ -32,7 +32,7 @@ typedef enum axNPU_TYPE_E {
     AX_BL_VNPU_2   = (1 << 4)  /* running under BIG-LITTLE VNPU2 */
 } AX_NPU_TYPE_E;
 
-#if defined(CHIP_AX650)
+#if defined(CONFIG_AX_650N_MSP_ENABLED)
 static AX_S32 CheckModelVNpu(const std::string &strModel, const AX_ENGINE_MODEL_TYPE_T &eModelType,
                              const AX_S32 &nNpuType, AX_U32 &nNpuSet)
 {
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-en-au-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-en-au-ax650.json
new file mode 100644
index 00000000..8b94e7ef
--- /dev/null
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-en-au-ax650.json
@@ -0,0 +1,31 @@
+{
+    "mode": "melotts-en-au-ax650",
+    "type": "tts",
+    "homepage": "https://github.com/ml-inory/melotts.axera/tree/main/model_convert",
+    "compile_flage": "pulsar2 build --input decoder-en.onnx --config config_decoder_u16_en-au.json --output_dir decoder-au-ax650 --output_name decoder-en-au.axmodel --target_hardware AX650 --npu_mode NPU3 --compiler.check 0",
+    "pulsar_version": "4.0-64a0e58f",
+    "capabilities": [
+        "tts",
+        "English"
+    ],
+    "input_type": [
+        "tts.utf-8"
+    ],
+    "output_type": [
+        "tts.wav",
+        "sys.play.0_1"
+    ],
+    "mode_param": {
+        "encoder": "encoder-en-au.ort",
+        "decoder": "decoder-en-au.axmodel",
+        "gbin": "g-en-au.bin",
+        "tokens": "tokens-en.txt",
+        "lexicon": "lexicon-en.txt",
+        "tagger": "en_tn_tagger.fst",
+        "verbalizer": "en_tn_verbalizer.fst",
+        "spacker_speed": 1.2,
+        "mode_rate": 44100,
+        "audio_rate": 16000,
+        "awake_delay": 1000
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-en-br-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-en-br-ax650.json
new file mode 100644
index 00000000..ef43e582
--- /dev/null
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-en-br-ax650.json
@@ -0,0 +1,31 @@
+{
+    "mode": "melotts-en-br-ax650",
+    "type": "tts",
+    "homepage": "https://github.com/ml-inory/melotts.axera/tree/main/model_convert",
+    "compile_flage": "pulsar2 build --input decoder-en.onnx --config config_decoder_u16_en-br.json --output_dir decoder-br-ax650 --output_name decoder-en-br.axmodel --target_hardware AX650 --npu_mode NPU3 --compiler.check 0",
+    "pulsar_version": "4.0-64a0e58f",
+    "capabilities": [
+        "tts",
+        "English"
+    ],
+    "input_type": [
+        "tts.utf-8"
+    ],
+    "output_type": [
+        "tts.wav",
+        "sys.play.0_1"
+    ],
+    "mode_param": {
+        "encoder": "encoder-en-br.ort",
+        "decoder": "decoder-en-br.axmodel",
+        "gbin": "g-en-br.bin",
+        "tokens": "tokens-en.txt",
+        "lexicon": "lexicon-en.txt",
+        "tagger": "en_tn_tagger.fst",
+        "verbalizer": "en_tn_verbalizer.fst",
+        "spacker_speed": 1.2,
+        "mode_rate": 44100,
+        "audio_rate": 16000,
+        "awake_delay": 1000
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-en-default-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-en-default-ax650.json
new file mode 100644
index 00000000..3e19b401
--- /dev/null
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-en-default-ax650.json
@@ -0,0 +1,31 @@
+{
+    "mode": "melotts-en-default-ax650",
+    "type": "tts",
+    "homepage": "https://github.com/ml-inory/melotts.axera/tree/main/model_convert",
+    "compile_flage": "pulsar2 build --input decoder-en.onnx --config config_decoder_u16_en-default.json --output_dir decoder-default-ax650 --output_name decoder-en-default.axmodel --target_hardware AX650 --npu_mode NPU3 --compiler.check 0",
+    "pulsar_version": "4.0-64a0e58f",
+    "capabilities": [
+        "tts",
+        "English"
+    ],
+    "input_type": [
+        "tts.utf-8"
+    ],
+    "output_type": [
+        "tts.wav",
+        "sys.play.0_1"
+    ],
+    "mode_param": {
+        "encoder": "encoder-en-default.ort",
+        "decoder": "decoder-en-default.axmodel",
+        "gbin": "g-en-default.bin",
+        "tokens": "tokens-en.txt",
+        "lexicon": "lexicon-en.txt",
+        "tagger": "en_tn_tagger.fst",
+        "verbalizer": "en_tn_verbalizer.fst",
+        "spacker_speed": 1.2,
+        "mode_rate": 44100,
+        "audio_rate": 16000,
+        "awake_delay": 1000
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-en-india-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-en-india-ax650.json
new file mode 100644
index 00000000..ee0cab64
--- /dev/null
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-en-india-ax650.json
@@ -0,0 +1,31 @@
+{
+    "mode": "melotts-en-india-ax650",
+    "type": "tts",
+    "homepage": "https://github.com/ml-inory/melotts.axera/tree/main/model_convert",
+    "compile_flage": "pulsar2 build --input decoder-en.onnx --config config_decoder_u16_en-india.json --output_dir decoder-india-ax650 --output_name decoder-en-india.axmodel --target_hardware AX650 --npu_mode NPU3 --compiler.check 0",
+    "pulsar_version": "4.0-64a0e58f",
+    "capabilities": [
+        "tts",
+        "English"
+    ],
+    "input_type": [
+        "tts.utf-8"
+    ],
+    "output_type": [
+        "tts.wav",
+        "sys.play.0_1"
+    ],
+    "mode_param": {
+        "encoder": "encoder-en-india.ort",
+        "decoder": "decoder-en-india.axmodel",
+        "gbin": "g-en-india.bin",
+        "tokens": "tokens-en.txt",
+        "lexicon": "lexicon-en.txt",
+        "tagger": "en_tn_tagger.fst",
+        "verbalizer": "en_tn_verbalizer.fst",
+        "spacker_speed": 1.2,
+        "mode_rate": 44100,
+        "audio_rate": 16000,
+        "awake_delay": 1000
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-en-us-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-en-us-ax650.json
new file mode 100644
index 00000000..6ca7181e
--- /dev/null
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-en-us-ax650.json
@@ -0,0 +1,31 @@
+{
+    "mode": "melotts-en-us-ax650",
+    "type": "tts",
+    "homepage": "https://github.com/ml-inory/melotts.axera/tree/main/model_convert",
+    "compile_flage": "pulsar2 build --input decoder-en.onnx --config config_decoder_u16_en-us.json --output_dir decoder-us-ax650 --output_name decoder-en-us.axmodel --target_hardware AX650 --npu_mode NPU3 --compiler.check 0",
+    "pulsar_version": "4.0-64a0e58f",
+    "capabilities": [
+        "tts",
+        "English"
+    ],
+    "input_type": [
+        "tts.utf-8"
+    ],
+    "output_type": [
+        "tts.wav",
+        "sys.play.0_1"
+    ],
+    "mode_param": {
+        "encoder": "encoder-en-us.ort",
+        "decoder": "decoder-en-us.axmodel",
+        "gbin": "g-en-us.bin",
+        "tokens": "tokens-en.txt",
+        "lexicon": "lexicon-en.txt",
+        "tagger": "en_tn_tagger.fst",
+        "verbalizer": "en_tn_verbalizer.fst",
+        "spacker_speed": 1.2,
+        "mode_rate": 44100,
+        "audio_rate": 16000,
+        "awake_delay": 1000
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-es-es-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-es-es-ax650.json
new file mode 100644
index 00000000..7012574b
--- /dev/null
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-es-es-ax650.json
@@ -0,0 +1,29 @@
+{
+    "mode": "melotts-es-es-ax650",
+    "type": "tts",
+    "homepage": "https://github.com/ml-inory/melotts.axera/tree/main/model_convert",
+    "compile_flage": "pulsar2 build --input decoder-es.onnx --config config_decoder_u16_es-es.json --output_dir decoder-es-ax650 --output_name decoder-es-es.axmodel --target_hardware AX650 --npu_mode NPU3 --compiler.check 0",
+    "pulsar_version": "4.0-64a0e58f",
+    "capabilities": [
+        "tts",
+        "Spanish"
+    ],
+    "input_type": [
+        "tts.utf-8"
+    ],
+    "output_type": [
+        "tts.wav",
+        "sys.play.0_1"
+    ],
+    "mode_param": {
+        "encoder": "encoder-es-es.ort",
+        "decoder": "decoder-es-es.axmodel",
+        "gbin": "g-es.bin",
+        "tokens": "tokens-es.txt",
+        "lexicon": "lexicon-es.txt",
+        "spacker_speed": 1.2,
+        "mode_rate": 44100,
+        "audio_rate": 16000,
+        "awake_delay": 1000
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-ja-jp-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-ja-jp-ax650.json
new file mode 100644
index 00000000..49f26aa7
--- /dev/null
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-ja-jp-ax650.json
@@ -0,0 +1,31 @@
+{
+    "mode": "melotts-ja-jp-ax650",
+    "type": "tts",
+    "homepage": "https://github.com/ml-inory/melotts.axera/tree/main/model_convert",
+    "compile_flage": "pulsar2 build --input decoder-jp.onnx --config config_decoder_u16_ja-jp.json --output_dir decoder-ja-ax650 --output_name decoder-ja-jp.axmodel --target_hardware AX650 --npu_mode NPU3 --compiler.check 0",
+    "pulsar_version": "4.0-64a0e58f",
+    "capabilities": [
+        "tts",
+        "Japanese"
+    ],
+    "input_type": [
+        "tts.utf-8"
+    ],
+    "output_type": [
+        "tts.wav",
+        "sys.play.0_1"
+    ],
+    "mode_param": {
+        "encoder": "encoder-ja-jp.ort",
+        "decoder": "decoder-ja-jp.axmodel",
+        "gbin": "g-jp.bin",
+        "tokens": "tokens-jp.txt",
+        "lexicon": "lexicon-jp.txt",
+        "tagger": "ja_tn_tagger.fst",
+        "verbalizer": "ja_tn_verbalizer.fst",
+        "spacker_speed": 1.1,
+        "mode_rate": 44100,
+        "audio_rate": 16000,
+        "awake_delay": 1000
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-zh-cn-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-zh-cn-ax650.json
new file mode 100644
index 00000000..41060977
--- /dev/null
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-zh-cn-ax650.json
@@ -0,0 +1,31 @@
+{
+    "mode": "melotts-zh-cn-ax650",
+    "type": "tts",
+    "homepage": "https://github.com/ml-inory/melotts.axera/tree/main/model_convert",
+    "compile_flage": "pulsar2 build --input decoder-zh.onnx --config config_decoder_u16_zh-cn.json --output_dir decoder-zh-ax650 --output_name decoder-zh-cn.axmodel --target_hardware AX650 --npu_mode NPU3 --compiler.check 0",
+    "pulsar_version": "4.0-64a0e58f",
+    "capabilities": [
+        "tts",
+        "Chinese"
+    ],
+    "input_type": [
+        "tts.utf-8"
+    ],
+    "output_type": [
+        "tts.wav",
+        "sys.play.0_1"
+    ],
+    "mode_param": {
+        "encoder": "encoder-zh-cn.ort",
+        "decoder": "decoder-zh-cn.axmodel",
+        "gbin": "g-zh_mix_en.bin",
+        "tokens": "tokens.txt",
+        "lexicon": "lexicon.txt",
+        "tagger": "zh_tn_tagger.fst",
+        "verbalizer": "zh_tn_verbalizer.fst",
+        "spacker_speed": 1.1,
+        "mode_rate": 44100,
+        "audio_rate": 16000,
+        "awake_delay": 1000
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/src/runner/EngineWrapper.cpp b/projects/llm_framework/main_melotts/src/runner/EngineWrapper.cpp
index 0394ee48..bc7590af 100644
--- a/projects/llm_framework/main_melotts/src/runner/EngineWrapper.cpp
+++ b/projects/llm_framework/main_melotts/src/runner/EngineWrapper.cpp
@@ -32,7 +32,7 @@ typedef enum axNPU_TYPE_E {
     AX_BL_VNPU_2 = (1 << 4)    /* running under BIG-LITTLE VNPU2 */
 } AX_NPU_TYPE_E;
 
-#if defined(CHIP_AX650)
+#if defined(CONFIG_AX_650N_MSP_ENABLED)
 static AX_S32 CheckModelVNpu(const std::string &strModel, const AX_ENGINE_MODEL_TYPE_T &eModelType, const AX_S32 &nNpuType, AX_U32 &nNpuSet) {
     AX_ENGINE_NPU_ATTR_T stNpuAttr;
     memset(&stNpuAttr, 0x00, sizeof(stNpuAttr));
diff --git a/projects/llm_framework/main_skel/SConstruct b/projects/llm_framework/main_skel/SConstruct
index bae9b432..55e508b6 100644
--- a/projects/llm_framework/main_skel/SConstruct
+++ b/projects/llm_framework/main_skel/SConstruct
@@ -3,8 +3,10 @@ import os
 Import('env')
 with open(env['PROJECT_TOOL_S']) as f:
     exec(f.read())
-
-SRCS = append_srcs_dir(ADir('src'))
+if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
+    SRCS = append_srcs_dir(ADir('src'))
+else:
+    SRCS = Glob('src/main.c')
 INCLUDE = [ADir('include'), ADir('.')]
 PRIVATE_INCLUDE = []
 REQUIREMENTS = ['pthread', 'utilities', 'ax_msp', 'eventpp', 'StackFlow', 'ax-samples']
@@ -21,7 +23,8 @@ DEFINITIONS += ['-std=c++17']
 LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 REQUIREMENTS += ['ax_engine', 'ax_interpreter', 'ax_sys']
-REQUIREMENTS += ['ax_skel', 'ax_engine', 'ax_venc', 'ax_vdec', 'ax_ivps', 'ax_proton']
+if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
+    REQUIREMENTS += ['ax_skel', 'ax_engine', 'ax_venc', 'ax_vdec', 'ax_ivps', 'ax_proton']
 
 INCLUDE += [ADir('../include'), ADir('src/runner'), ADir('../include/opencv4')]
 
diff --git a/projects/llm_framework/main_skel/src/main.c b/projects/llm_framework/main_skel/src/main.c
index 144d3a46..cd835ca8 100644
--- a/projects/llm_framework/main_skel/src/main.c
+++ b/projects/llm_framework/main_skel/src/main.c
@@ -7,7 +7,8 @@
  * written consent of Axera Semiconductor Co., Ltd.
  *
  **************************************************************************************************/
-
+#include <global_config.h>
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
@@ -2251,3 +2252,8 @@ AX_S32 main(AX_S32 argc, AX_CHAR *argv[])
 
     return (0 != nRet) ? -1 : 0;
 }
+#else
+int main(){
+    return 0;
+}
+#endif
\ No newline at end of file
diff --git a/projects/llm_framework/main_whisper/src/runner/EngineWrapper.cpp b/projects/llm_framework/main_whisper/src/runner/EngineWrapper.cpp
index 6d219645..bb0e7f8c 100644
--- a/projects/llm_framework/main_whisper/src/runner/EngineWrapper.cpp
+++ b/projects/llm_framework/main_whisper/src/runner/EngineWrapper.cpp
@@ -33,7 +33,7 @@ typedef enum axNPU_TYPE_E {
     AX_BL_VNPU_2 = (1 << 4)    /* running under BIG-LITTLE VNPU2 */
 } AX_NPU_TYPE_E;
 
-#if defined(CHIP_AX650)
+#if defined(CONFIG_AX_650N_MSP_ENABLED)
 static AX_S32 CheckModelVNpu(const std::string &strModel, const AX_ENGINE_MODEL_TYPE_T &eModelType, const AX_S32 &nNpuType, AX_U32 &nNpuSet) {
     AX_ENGINE_NPU_ATTR_T stNpuAttr;
     memset(&stNpuAttr, 0x00, sizeof(stNpuAttr));
diff --git a/projects/llm_framework/main_yolo/models/mode_yolo11n-ax650.json b/projects/llm_framework/main_yolo/models/mode_yolo11n-ax650.json
new file mode 100644
index 00000000..cad0e05d
--- /dev/null
+++ b/projects/llm_framework/main_yolo/models/mode_yolo11n-ax650.json
@@ -0,0 +1,110 @@
+{
+    "mode":"yolo11n-ax650",
+    "type":"cv",
+    "homepage":"https://github.com/ultralytics/ultralytics",
+    "compile_flage":"pulsar2 build  --target_hardware AX650 --input Workspace/cv/model/yolo11n.onnx --output_dir yolo11n --config Workspace/cv/config/yolo11_config.json --output_name yolo11n.axmodel --npu_mode NPU3",
+    "pulsar_version":"4.1-patch1-c37957c7",
+    "capabilities":[
+        "Detection"
+    ],
+    "input_type":[
+        "yolo.jpeg.base64"
+    ],
+    "output_type":[
+        "yolo.box",
+        "yolo.boxV2"
+    ],
+    "mode_param":{
+        "yolo_model":"yolo11n.axmodel",
+        "model_type":"detect",
+        "img_h":320,
+        "img_w":320,
+        "cls_num":80,
+        "pron_threshold":0.45,
+        "nms_threshold":0.45,
+        "cls_name":[
+            "person",
+            "bicycle",
+            "car",
+            "motorcycle",
+            "airplane",
+            "bus",
+            "train",
+            "truck",
+            "boat",
+            "traffic light",
+            "fire hydrant",
+            "stop sign",
+            "parking meter",
+            "bench",
+            "bird",
+            "cat",
+            "dog",
+            "horse",
+            "sheep",
+            "cow",
+            "elephant",
+            "bear",
+            "zebra",
+            "giraffe",
+            "backpack",
+            "umbrella",
+            "handbag",
+            "tie",
+            "suitcase",
+            "frisbee",
+            "skis",
+            "snowboard",
+            "sports ball",
+            "kite",
+            "baseball bat",
+            "baseball glove",
+            "skateboard",
+            "surfboard",
+            "tennis racket",
+            "bottle",
+            "wine glass",
+            "cup",
+            "fork",
+            "knife",
+            "spoon",
+            "bowl",
+            "banana",
+            "apple",
+            "sandwich",
+            "orange",
+            "broccoli",
+            "carrot",
+            "hot dog",
+            "pizza",
+            "donut",
+            "cake",
+            "chair",
+            "couch",
+            "potted plant",
+            "bed",
+            "dining table",
+            "toilet",
+            "tv",
+            "laptop",
+            "mouse",
+            "remote",
+            "keyboard",
+            "cell phone",
+            "microwave",
+            "oven",
+            "toaster",
+            "sink",
+            "refrigerator",
+            "book",
+            "clock",
+            "vase",
+            "scissors",
+            "teddy bear",
+            "hair drier",
+            "toothbrush"
+        ]
+    },
+    "mode_param_bak":{
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_yolo/models/mode_yolo11n-hand-pose-ax650.json b/projects/llm_framework/main_yolo/models/mode_yolo11n-hand-pose-ax650.json
new file mode 100644
index 00000000..0ab718e2
--- /dev/null
+++ b/projects/llm_framework/main_yolo/models/mode_yolo11n-hand-pose-ax650.json
@@ -0,0 +1,32 @@
+{
+    "mode":"yolo11n-hand-pose-ax650",
+    "type":"cv",
+    "homepage":"https://github.com/ultralytics/ultralytics",
+    "compile_flage":"pulsar2 build  --target_hardware AX650 --input Workspace/cv/model/yolo11n-hand-pose-cut.onnx --output_dir yolo11n-hand-pose-ax650 --config Workspace/cv/config/yolo11_hand_pose_config.json --output_name yolo11n-hand-pose.axmodel --npu_mode NPU3",
+    "pulsar_version":"4.1-patch1-c37957c7",
+    "capabilities":[
+        "Pose"
+    ],
+    "input_type":[
+        "yolo.jpeg.base64"
+    ],
+    "output_type":[
+        "yolo.box",
+        "yolo.boxV2"
+    ],
+    "mode_param":{
+        "yolo_model":"yolo11n-hand-pose.axmodel",
+        "model_type":"pose",
+        "img_h":320,
+        "img_w":320,
+        "cls_num":1,
+        "point_num":21,
+        "pron_threshold":0.45,
+        "nms_threshold":0.45,
+        "cls_name":[
+            "hand"
+        ]
+    },
+    "mode_param_bak":{
+    }
+}
diff --git a/projects/llm_framework/main_yolo/models/mode_yolo11n-pose-ax650.json b/projects/llm_framework/main_yolo/models/mode_yolo11n-pose-ax650.json
new file mode 100644
index 00000000..bc7d009c
--- /dev/null
+++ b/projects/llm_framework/main_yolo/models/mode_yolo11n-pose-ax650.json
@@ -0,0 +1,32 @@
+{
+    "mode":"yolo11n-pose-ax650",
+    "type":"cv",
+    "homepage":"https://github.com/ultralytics/ultralytics",
+    "compile_flage":"pulsar2 build  --target_hardware AX650 --input Workspace/cv/model/yolo11n-seg-cut.onnx --output_dir yolo11n-seg-ax650 --config Workspace/cv/config/yolo11_seg_config.json --output_name yolo11n-seg.axmodel --npu_mode NPU3",
+    "pulsar_version":"4.1-patch1-c37957c7",
+    "capabilities":[
+        "Pose"
+    ],
+    "input_type":[
+        "yolo.jpeg.base64"
+    ],
+    "output_type":[
+        "yolo.box",
+        "yolo.boxV2"
+    ],
+    "mode_param":{
+        "yolo_model":"yolo11n-pose.axmodel",
+        "model_type":"pose",
+        "img_h":320,
+        "img_w":320,
+        "cls_num":1,
+        "point_num":17,
+        "pron_threshold":0.45,
+        "nms_threshold":0.45,
+        "cls_name":[
+            "person"
+        ]
+    },
+    "mode_param_bak":{
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_yolo/models/mode_yolo11n-seg-ax650.json b/projects/llm_framework/main_yolo/models/mode_yolo11n-seg-ax650.json
new file mode 100644
index 00000000..50e21d08
--- /dev/null
+++ b/projects/llm_framework/main_yolo/models/mode_yolo11n-seg-ax650.json
@@ -0,0 +1,110 @@
+{
+    "mode":"yolo11n-seg-ax650",
+    "type":"cv",
+    "homepage":"https://github.com/ultralytics/ultralytics",
+    "compile_flage":"pulsar2 build  --target_hardware AX650 --input Workspace/cv/model/yolo11n-pose-cut.onnx --output_dir yolo11n-pose-ax650 --config Workspace/cv/config/yolo11_pose_config.json --output_name yolo11n-pose.axmodel --npu_mode NPU3",
+    "pulsar_version":"4.1-patch1-c37957c7",
+    "capabilities":[
+        "Segmentation"
+    ],
+    "input_type":[
+        "yolo.jpeg.base64"
+    ],
+    "output_type":[
+        "yolo.box",
+        "yolo.boxV2"
+    ],
+    "mode_param":{
+        "yolo_model":"yolo11n-seg.axmodel",
+        "model_type":"segment",
+        "img_h":320,
+        "img_w":320,
+        "cls_num":80,
+        "pron_threshold":0.45,
+        "nms_threshold":0.45,
+        "cls_name":[
+            "person",
+            "bicycle",
+            "car",
+            "motorcycle",
+            "airplane",
+            "bus",
+            "train",
+            "truck",
+            "boat",
+            "traffic light",
+            "fire hydrant",
+            "stop sign",
+            "parking meter",
+            "bench",
+            "bird",
+            "cat",
+            "dog",
+            "horse",
+            "sheep",
+            "cow",
+            "elephant",
+            "bear",
+            "zebra",
+            "giraffe",
+            "backpack",
+            "umbrella",
+            "handbag",
+            "tie",
+            "suitcase",
+            "frisbee",
+            "skis",
+            "snowboard",
+            "sports ball",
+            "kite",
+            "baseball bat",
+            "baseball glove",
+            "skateboard",
+            "surfboard",
+            "tennis racket",
+            "bottle",
+            "wine glass",
+            "cup",
+            "fork",
+            "knife",
+            "spoon",
+            "bowl",
+            "banana",
+            "apple",
+            "sandwich",
+            "orange",
+            "broccoli",
+            "carrot",
+            "hot dog",
+            "pizza",
+            "donut",
+            "cake",
+            "chair",
+            "couch",
+            "potted plant",
+            "bed",
+            "dining table",
+            "toilet",
+            "tv",
+            "laptop",
+            "mouse",
+            "remote",
+            "keyboard",
+            "cell phone",
+            "microwave",
+            "oven",
+            "toaster",
+            "sink",
+            "refrigerator",
+            "book",
+            "clock",
+            "vase",
+            "scissors",
+            "teddy bear",
+            "hair drier",
+            "toothbrush"
+        ]
+    },
+    "mode_param_bak":{
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_yolo/src/EngineWrapper.cpp b/projects/llm_framework/main_yolo/src/EngineWrapper.cpp
index 99e448d1..72f16a6f 100644
--- a/projects/llm_framework/main_yolo/src/EngineWrapper.cpp
+++ b/projects/llm_framework/main_yolo/src/EngineWrapper.cpp
@@ -33,7 +33,7 @@ typedef enum axNPU_TYPE_E {
     AX_BL_VNPU_2   = (1 << 4)  /* running under BIG-LITTLE VNPU2 */
 } AX_NPU_TYPE_E;
 
-#if defined(CHIP_AX650)
+#if defined(CONFIG_AX_650N_MSP_ENABLED)
 static AX_S32 CheckModelVNpu(const std::string& strModel, const AX_ENGINE_MODEL_TYPE_T& eModelType,
                              const AX_S32& nNpuType, AX_U32& nNpuSet)
 {
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 1783e536..fc0097d2 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -352,6 +352,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         cpu_count = cpu_count - 2
     # cpu_count = 50
     Tasks = {
+        # Unit
         'lib-llm':[create_lib_deb,'lib-llm', '1.8', src_folder, revision],
         'llm-sys':[create_bin_deb,'llm-sys', '1.6', src_folder, revision],
         'llm-audio':[create_bin_deb,'llm-audio', '1.6', src_folder, revision],
@@ -368,14 +369,21 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-vad':[create_bin_deb,'llm-vad', '1.7', src_folder, revision],
         'llm-whisper':[create_bin_deb,'llm-whisper', '1.7', src_folder, revision],
         'llm-openai-api':[create_bin_deb,'llm-openai-api', '1.7', src_folder, revision],
+        # keyword spotting Audio file
         'llm-model-audio-en-us':[create_data_deb,'llm-model-audio-en-us', data_version, src_folder, revision],
         'llm-model-audio-zh-cn':[create_data_deb,'llm-model-audio-zh-cn', data_version, src_folder, revision],
+        # Kaldi model        
         'llm-model-sherpa-ncnn-streaming-zipformer-20M-2023-02-17':[create_data_deb,'llm-model-sherpa-ncnn-streaming-zipformer-20M-2023-02-17', data_version, src_folder, revision],
         'llm-model-sherpa-ncnn-streaming-zipformer-zh-14M-2023-02-23':[create_data_deb,'llm-model-sherpa-ncnn-streaming-zipformer-zh-14M-2023-02-23', data_version, src_folder, revision],
         'llm-model-sherpa-onnx-kws-zipformer-gigaspeech-3.3M-2024-01-01':[create_data_deb,'llm-model-sherpa-onnx-kws-zipformer-gigaspeech-3.3M-2024-01-01', '0.3', src_folder, revision],
         'llm-model-sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01':[create_data_deb,'llm-model-sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01', '0.3', src_folder, revision],
+        # TTS model
         'llm-model-single-speaker-english-fast':[create_data_deb,'llm-model-single-speaker-english-fast', '0.3', src_folder, revision],
         'llm-model-single-speaker-fast':[create_data_deb,'llm-model-single-speaker-fast', '0.3', src_folder, revision],
+        # VAD model
+        'llm-model-silero-vad':[create_data_deb,'llm-model-silero-vad', '0.4', src_folder, revision],
+        # MeloTTS model
+        ## AX630C
         'llm-model-melotts-zh-cn':[create_data_deb,'llm-model-melotts-zh-cn', '0.6', src_folder, revision],
         'llm-model-melotts-en-au':[create_data_deb,'llm-model-melotts-en-au', '0.6', src_folder, revision],
         'llm-model-melotts-en-br':[create_data_deb,'llm-model-melotts-en-br', '0.6', src_folder, revision],
@@ -384,20 +392,47 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-model-melotts-en-us':[create_data_deb,'llm-model-melotts-en-us', '0.6', src_folder, revision],
         'llm-model-melotts-ja-jp':[create_data_deb,'llm-model-melotts-ja-jp', '0.6', src_folder, revision],
         'llm-model-melotts-es-es':[create_data_deb,'llm-model-melotts-es-es', '0.5', src_folder, revision],
+        ## AX650
+        'llm-model-melotts-zh-cn-ax650':[create_data_deb,'llm-model-melotts-zh-cn-ax650', '0.6', src_folder, revision],
+        'llm-model-melotts-en-au-ax650':[create_data_deb,'llm-model-melotts-en-au-ax650', '0.6', src_folder, revision],
+        'llm-model-melotts-en-br-ax650':[create_data_deb,'llm-model-melotts-en-br-ax650', '0.6', src_folder, revision],
+        'llm-model-melotts-en-default-ax650':[create_data_deb,'llm-model-melotts-en-default-ax650', '0.6', src_folder, revision],
+        'llm-model-melotts-en-india-ax650':[create_data_deb,'llm-model-melotts-en-india-ax650', '0.6', src_folder, revision],
+        'llm-model-melotts-en-us-ax650':[create_data_deb,'llm-model-melotts-en-us-ax650', '0.6', src_folder, revision],
+        'llm-model-melotts-ja-jp-ax650':[create_data_deb,'llm-model-melotts-ja-jp-ax650', '0.6', src_folder, revision],
+        'llm-model-melotts-es-es-ax650':[create_data_deb,'llm-model-melotts-es-es-ax650', '0.5', src_folder, revision],
+        # Yolo model
+        ## AX630C
         'llm-model-yolo11n':[create_data_deb,'llm-model-yolo11n', data_version, src_folder, revision],
-        'llm-model-yolo11n-npu1':[create_data_deb,'llm-model-yolo11n-npu1', '0.4', src_folder, revision],
         'llm-model-yolo11n-pose':[create_data_deb,'llm-model-yolo11n-pose', '0.3', src_folder, revision],
-        'llm-model-yolo11n-pose-npu1':[create_data_deb,'llm-model-yolo11n-pose-npu1', '0.4', src_folder, revision],
         'llm-model-yolo11n-hand-pose':[create_data_deb,'llm-model-yolo11n-hand-pose', '0.3', src_folder, revision],
-        'llm-model-yolo11n-hand-pose-npu1':[create_data_deb,'llm-model-yolo11n-hand-pose-npu1', '0.4', src_folder, revision],
         'llm-model-yolo11n-seg':[create_data_deb,'llm-model-yolo11n-seg', '0.3', src_folder, revision],
+        'llm-model-yolo11n-npu1':[create_data_deb,'llm-model-yolo11n-npu1', '0.4', src_folder, revision],
+        'llm-model-yolo11n-pose-npu1':[create_data_deb,'llm-model-yolo11n-pose-npu1', '0.4', src_folder, revision],
+        'llm-model-yolo11n-hand-pose-npu1':[create_data_deb,'llm-model-yolo11n-hand-pose-npu1', '0.4', src_folder, revision],
         'llm-model-yolo11n-seg-npu1':[create_data_deb,'llm-model-yolo11n-seg-npu1', '0.4', src_folder, revision],
+        ## AX650
+        'llm-model-yolo11n-ax650':[create_data_deb,'llm-model-yolo11n-ax650', '0.4', src_folder, revision],
+        'llm-model-yolo11n-pose-ax650':[create_data_deb,'llm-model-yolo11n-pose-ax650', '0.4', src_folder, revision],
+        'llm-model-yolo11n-hand-pose-ax650':[create_data_deb,'llm-model-yolo11n-hand-pose-ax650', '0.4', src_folder, revision],
+        'llm-model-yolo11n-seg-ax650':[create_data_deb,'llm-model-yolo11n-seg-ax650', '0.4', src_folder, revision],
+        # DepthAnything model
+        ## AX630C
         'llm-model-depth-anything-ax630c':[create_data_deb,'llm-model-depth-anything-ax630c', '0.4', src_folder, revision],
         'llm-model-depth-anything-npu1-ax630c':[create_data_deb,'llm-model-depth-anything-npu1-ax630c', '0.4', src_folder, revision],
+        ## AX650
+        'llm-model-depth-anything-ax650':[create_data_deb,'llm-model-depth-anything-ax650', '0.4', src_folder, revision],
+        # Whisper model
+        ## AX630C
         'llm-model-whisper-tiny':[create_data_deb,'llm-model-whisper-tiny', '0.4', src_folder, revision],
         'llm-model-whisper-base':[create_data_deb,'llm-model-whisper-base', '0.4', src_folder, revision],
         'llm-model-whisper-small':[create_data_deb,'llm-model-whisper-small', '0.4', src_folder, revision],
-        'llm-model-silero-vad':[create_data_deb,'llm-model-silero-vad', '0.4', src_folder, revision],
+        ## AX650
+        'llm-model-whisper-tiny-ax650':[create_data_deb,'llm-model-whisper-tiny-ax650', '0.4', src_folder, revision],
+        'llm-model-whisper-base-ax650':[create_data_deb,'llm-model-whisper-base-ax650', '0.4', src_folder, revision],
+        'llm-model-whisper-small-ax650':[create_data_deb,'llm-model-whisper-small-ax650', '0.4', src_folder, revision],
+        # Qwen model
+        ## AX630C
         'llm-model-qwen2.5-0.5B-prefill-20e':[create_data_deb,'llm-model-qwen2.5-0.5B-prefill-20e', data_version, src_folder, revision],
         'llm-model-qwen2.5-0.5B-p256-ax630c':[create_data_deb,'llm-model-qwen2.5-0.5B-p256-ax630c', '0.4', src_folder, revision],
         'llm-model-qwen2.5-0.5B-Int4-ax630c':[create_data_deb,'llm-model-qwen2.5-0.5B-Int4-ax630c', '0.4', src_folder, revision],
@@ -407,14 +442,32 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-model-qwen2.5-1.5B-Int4-ax630c':[create_data_deb,'llm-model-qwen2.5-1.5B-Int4-ax630c', '0.4', src_folder, revision],
         'llm-model-qwen2.5-coder-0.5B-ax630c':[create_data_deb,'llm-model-qwen2.5-coder-0.5B-ax630c', data_version, src_folder, revision],
         'llm-model-qwen3-0.6B-ax630c':[create_data_deb,'llm-model-qwen3-0.6B-ax630c', '0.4', src_folder, revision],
+        ## AX650
+        'llm-model-qwen2.5-0.5B-Int4-ax650':[create_data_deb,'llm-model-qwen2.5-0.5B-Int4-ax650', '0.4', src_folder, revision],
+        'llm-model-qwen2.5-HA-0.5B-ctx-ax650':[create_data_deb,'llm-model-qwen2.5-HA-0.5B-ctx-ax650', '0.5', src_folder, revision],
+        'llm-model-qwen2.5-1.5B-Int4-ax650':[create_data_deb,'llm-model-qwen2.5-1.5B-Int4-ax650', '0.4', src_folder, revision],
+        'llm-model-qwen2.5-3B-Int4-ax650':[create_data_deb,'llm-model-qwen2.5-3B-Int4-ax650', '0.4', src_folder, revision],
+        'llm-model-qwen2.5-7B-Int4-ax650':[create_data_deb,'llm-model-qwen2.5-7B-Int4-ax650', '0.4', src_folder, revision],
+        # Llama model
         'llm-model-llama3.2-1B-prefill-ax630c':[create_data_deb,'llm-model-llama3.2-1B-prefill-ax630c', data_version, src_folder, revision],
         'llm-model-llama3.2-1B-p256-ax630c':[create_data_deb,'llm-model-llama3.2-1B-p256-ax630c', '0.4', src_folder, revision],
         'llm-model-openbuddy-llama3.2-1B-ax630c':[create_data_deb,'llm-model-openbuddy-llama3.2-1B-ax630c', data_version, src_folder, revision],
+        # InternVL model
         'llm-model-internvl2.5-1B-ax630c':[create_data_deb,'llm-model-internvl2.5-1B-ax630c', '0.4', src_folder, revision],
         'llm-model-internvl2.5-1B-364-ax630c':[create_data_deb,'llm-model-internvl2.5-1B-364-ax630c', '0.4', src_folder, revision],
+        'llm-model-internvl3-1B-448-ax650':[create_data_deb,'llm-model-internvl3-1B-448-ax650', '0.6', src_folder, revision],
+        # DeepSeek model
+        ## AX630C
         'llm-model-deepseek-r1-1.5B-ax630c':[create_data_deb,'llm-model-deepseek-r1-1.5B-ax630c', '0.3', src_folder, revision],
         'llm-model-deepseek-r1-1.5B-p256-ax630c':[create_data_deb,'llm-model-deepseek-r1-1.5B-p256-ax630c', '0.4', src_folder, revision],
+        ## AX650
+        'llm-model-deepseek-r1-1.5B-Int4-ax650':[create_data_deb,'llm-model-deepseek-r1-1.5B-Int4-ax650', '0.4', src_folder, revision],#
+        'llm-model-deepseek-r1-7B-Int4-ax650':[create_data_deb,'llm-model-deepseek-r1-7B-Int4-ax650', '0.4', src_folder, revision],#
+        # Smolvlm model
+        ## AX630C
         'llm-model-smolvlm-256M-ax630c':[create_data_deb,'llm-model-smolvlm-256M-ax630c', '0.4', src_folder, revision],
+        'llm-model-smolvlm-256M-ax650':[create_data_deb,'llm-model-smolvlm-256M-ax650', '0.4', src_folder, revision],
+        ## AX650
         'llm-model-smolvlm-500M-ax630c':[create_data_deb,'llm-model-smolvlm-500M-ax630c', '0.4', src_folder, revision],
         # 'llm-model-qwen2-0.5B-prefill-20e':[create_data_deb,'llm-model-qwen2-0.5B-prefill-20e', data_version, src_folder, revision],
         # 'llm-model-qwen2-1.5B-prefill-20e':[create_data_deb,'llm-model-qwen2-1.5B-prefill-20e', data_version, src_folder, revision]

From 9d816fe8275577f07d10d2cd89484f7a9debd4ac Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 28 Aug 2025 16:38:46 +0800
Subject: [PATCH 32/79] [update]  ensure that a frame is written

---
 projects/llm_framework/main_asr/src/main.cpp     | 4 +++-
 projects/llm_framework/main_kws/src/main.cpp     | 4 +++-
 projects/llm_framework/main_vad/src/main.cpp     | 4 +++-
 projects/llm_framework/main_whisper/src/main.cpp | 5 +++--
 4 files changed, 12 insertions(+), 5 deletions(-)

diff --git a/projects/llm_framework/main_asr/src/main.cpp b/projects/llm_framework/main_asr/src/main.cpp
index ae51982b..d2dff9ba 100644
--- a/projects/llm_framework/main_asr/src/main.cpp
+++ b/projects/llm_framework/main_asr/src/main.cpp
@@ -58,7 +58,7 @@ class llm_task {
     std::atomic_bool audio_flage_;
     std::atomic_bool awake_flage_;
     int awake_delay_       = 50;
-    int delay_audio_frame_ = 11;
+    int delay_audio_frame_ = 10;
     buffer_t *pcmdata;
 
     std::function<void(void)> pause;
@@ -187,6 +187,7 @@ class llm_task {
             count++;
             return;
         }
+        buffer_write_char(pcmdata, raw.data(), raw.length());
         buffer_position_set(pcmdata, 0);
 
         std::vector<float> floatSamples;
@@ -523,6 +524,7 @@ class llm_asr : public StackFlow {
                                             });
                     llm_task_obj->audio_flage_ = true;
                 } else if (input.find("asr") != std::string::npos) {
+                    llm_task_obj->delay_audio_frame_ = 0;
                     llm_channel->subscriber_work_id(
                         "", std::bind(&llm_asr::task_user_data, this, std::weak_ptr<llm_task>(llm_task_obj),
                                       std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
diff --git a/projects/llm_framework/main_kws/src/main.cpp b/projects/llm_framework/main_kws/src/main.cpp
index 8c724599..9a4f8fc7 100644
--- a/projects/llm_framework/main_kws/src/main.cpp
+++ b/projects/llm_framework/main_kws/src/main.cpp
@@ -59,7 +59,7 @@ class llm_task {
     bool enwake_audio_;
     std::atomic_bool audio_flage_;
     task_callback_t out_callback_;
-    int delay_audio_frame_ = 11;
+    int delay_audio_frame_ = 10;
     buffer_t *pcmdata;
     std::string wake_wav_file_;
 
@@ -233,6 +233,7 @@ class llm_task {
             count++;
             return;
         }
+        buffer_write_char(pcmdata, raw.data(), raw.length());
         buffer_position_set(pcmdata, 0);
 
         std::vector<float> floatSamples;
@@ -529,6 +530,7 @@ class llm_kws : public StackFlow {
                                             });
                     llm_task_obj->audio_flage_ = true;
                 } else if (input.find("kws") != std::string::npos) {
+                    llm_task_obj->delay_audio_frame_ = 0;
                     llm_channel->subscriber_work_id(
                         "", std::bind(&llm_kws::task_user_data, this, std::weak_ptr<llm_task>(llm_task_obj),
                                       std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
diff --git a/projects/llm_framework/main_vad/src/main.cpp b/projects/llm_framework/main_vad/src/main.cpp
index 1c49eb33..d1414a71 100644
--- a/projects/llm_framework/main_vad/src/main.cpp
+++ b/projects/llm_framework/main_vad/src/main.cpp
@@ -60,7 +60,7 @@ class llm_task {
     std::string superior_id_;
     task_callback_t out_callback_;
     int awake_delay_       = 50;
-    int delay_audio_frame_ = 4;
+    int delay_audio_frame_ = 3;
     buffer_t *pcmdata;
     std::string wake_wav_file_;
 
@@ -158,6 +158,7 @@ class llm_task {
             count++;
             return;
         }
+        buffer_write_char(pcmdata, raw.c_str(), raw.length());
         buffer_position_set(pcmdata, 0);
 
         std::vector<float> floatSamples;
@@ -430,6 +431,7 @@ class llm_vad : public StackFlow {
                                             });
                     llm_task_obj->audio_flage_ = true;
                 } else if (input.find("vad") != std::string::npos) {
+                    llm_task_obj->delay_audio_frame_ = 0;
                     llm_channel->subscriber_work_id(
                         "", std::bind(&llm_vad::task_user_data, this, std::weak_ptr<llm_task>(llm_task_obj),
                                       std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
diff --git a/projects/llm_framework/main_whisper/src/main.cpp b/projects/llm_framework/main_whisper/src/main.cpp
index 51c6c6e1..9c957d84 100644
--- a/projects/llm_framework/main_whisper/src/main.cpp
+++ b/projects/llm_framework/main_whisper/src/main.cpp
@@ -99,7 +99,7 @@ class llm_task {
     static int ax_init_flage_;
     task_callback_t out_callback_;
     int awake_delay_       = 50;
-    int delay_audio_frame_ = 3001;
+    int delay_audio_frame_ = 3000;
     buffer_t *pcmdata;
 
     std::function<void(void)> pause;
@@ -343,6 +343,7 @@ class llm_task {
             count++;
             if (endpoint_flage_) return;
         }
+        buffer_write_char(pcmdata, raw.data(), raw.length());
         endpoint_flage_ = true;
 
         buffer_position_set(pcmdata, 0);
@@ -851,7 +852,7 @@ class llm_whisper : public StackFlow {
                                             });
                     llm_task_obj->audio_flage_ = true;
                 } else if (input.find("whisper") != std::string::npos) {
-                    if (input.find("stream.base64") != std::string::npos) llm_task_obj->delay_audio_frame_ = 0;
+                    llm_task_obj->delay_audio_frame_ = 0;
                     llm_channel->subscriber_work_id(
                         "", std::bind(&llm_whisper::task_user_data, this, std::weak_ptr<llm_task>(llm_task_obj),
                                       std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,

From 92b10acf3ebd0f5792f5a8fd96f7ae0c6ccf1e80 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 29 Aug 2025 16:22:35 +0800
Subject: [PATCH 33/79] [update] add internvl3-1B-ax630c model update main_vlm

---
 .../scripts/tokenizer_internvl3-1B-ax630c.py  | 241 ++++++++++++++++++
 projects/llm_framework/main_vlm/src/main.cpp  |  49 ++--
 .../llm_framework/main_vlm/src/runner/LLM.hpp |  39 +--
 3 files changed, 269 insertions(+), 60 deletions(-)
 create mode 100644 projects/llm_framework/main_vlm/scripts/tokenizer_internvl3-1B-ax630c.py

diff --git a/projects/llm_framework/main_vlm/scripts/tokenizer_internvl3-1B-ax630c.py b/projects/llm_framework/main_vlm/scripts/tokenizer_internvl3-1B-ax630c.py
new file mode 100644
index 00000000..cf6a3dc4
--- /dev/null
+++ b/projects/llm_framework/main_vlm/scripts/tokenizer_internvl3-1B-ax630c.py
@@ -0,0 +1,241 @@
+from transformers import AutoTokenizer
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+import uuid
+
+tokenizers = {}
+
+class Tokenizer_Http():
+    def __init__(self, model_id):
+        self.tokenizer = AutoTokenizer.from_pretrained(model_id)
+        self.messages = [
+            {"role": "system", "content": "你是由上海人工智能实验室联合商汤科技开发的书生多模态大模型，英文名叫InternVL, 是一个有用无害的人工智能助手。"},
+        ]
+        self.token_ids = []
+        self.token_ids_cache = []
+
+    def encode(self, prompt, last_reply=None):
+        if last_reply is not None:
+            self.messages.append({"role": "assistant", "content": last_reply})
+            text = self.tokenizer.apply_chat_template(
+                self.messages,
+                tokenize=False,
+                add_generation_prompt=True
+            )
+            self.token_ids = self.tokenizer.encode(text)[:-3]
+        self.messages.append({"role": "user", "content": prompt})
+
+        text = self.tokenizer.apply_chat_template(
+            self.messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        token_ids = self.tokenizer.encode(text)
+        diff = token_ids[len(self.token_ids):]
+        self.token_ids = token_ids
+        return token_ids, diff
+
+    def encode_with_image(self, question: str, num_of_images: int, imgsz: int, last_reply=None):
+        if last_reply is not None:
+            self.messages.append({"role": "assistant", "content": last_reply})
+
+        # 根据图片尺寸设定 context_len
+        if imgsz == 448:
+            context_len = 256
+        elif imgsz == 224:
+            context_len = 64
+        else:
+            print(f"Unsupported imgsz: {imgsz}")
+            return None, None
+
+        # 拼接带图片的用户输入
+        question_with_images = question
+        if num_of_images > 0:
+            for _ in range(num_of_images):
+                question_with_images += "\n<img>" + "<IMG_CONTEXT>" * context_len + "</img>\n"
+
+        self.messages.append({"role": "user", "content": question_with_images})
+
+        text = self.tokenizer.apply_chat_template(
+            self.messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        token_ids = self.tokenizer.encode(text)
+        diff = token_ids[len(self.token_ids):]
+        self.token_ids = token_ids
+        return token_ids, diff
+
+    def decode(self, token_ids):
+        self.token_ids_cache += token_ids
+        text = self.tokenizer.decode(self.token_ids_cache)
+        if "\ufffd" in text:
+            print("Text 中包含非法字符")
+            return ""
+        else:
+            self.token_ids_cache.clear()
+            return text
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+    @property
+    def bos_token(self):
+        return self.tokenizer.bos_token
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+    @property
+    def img_start_token(self):
+        return self.tokenizer.encode("<img>")[0]
+    @property
+    def img_context_token(self):
+        return self.tokenizer.encode("<IMG_CONTEXT>")[0]
+
+    def reset(self, system_prompt=None):
+        if system_prompt is None:
+            system_prompt = args.content
+        self.messages = [
+            {"role": "system", "content": system_prompt},
+        ]
+        text = self.tokenizer.apply_chat_template(
+            self.messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        token_ids = self.tokenizer.encode(text)[:-3]
+        self.token_ids = token_ids
+        print(self.decode(token_ids))
+        return token_ids
+
+
+class Request(BaseHTTPRequestHandler):
+    timeout = 5
+    server_version = 'Apache'
+
+    def do_GET(self):
+        print(self.path)
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.end_headers()
+        if '/get_uid' in self.path:
+            new_uid = str(uuid.uuid4())
+            print("新 uid:", new_uid)
+            tokenizers[new_uid] = Tokenizer_Http(args.model_id)
+            msg = json.dumps({'uid': new_uid})
+        elif '/bos_id' in self.path:
+            uid = self.get_query_param("uid")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                msg = json.dumps({'bos_id': instance.bos_id if instance.bos_id is not None else -1})
+        elif '/eos_id' in self.path:
+            uid = self.get_query_param("uid")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                msg = json.dumps({'eos_id': instance.eos_id if instance.eos_id is not None else -1})
+        elif '/img_start_token' in self.path:
+            uid = self.get_query_param("uid")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                msg = json.dumps({'img_start_token': instance.img_start_token})
+
+        elif '/img_context_token' in self.path:
+            uid = self.get_query_param("uid")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                msg = json.dumps({'img_context_token': instance.img_context_token})
+        else:
+            msg = json.dumps({'error': 'Invalid GET endpoint'})
+        print(msg)
+        self.wfile.write(msg.encode())
+
+    def do_POST(self):
+        content_length = int(self.headers.get('content-length', 0))
+        data = self.rfile.read(content_length).decode()
+        req = json.loads(data)
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.end_headers()
+
+        if '/encode' in self.path:
+            uid = req.get('uid')
+            prompt = req.get('text')
+            last_reply = req.get('last_reply')
+            b_img_prompt = False
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if 'img_prompt' in req:
+                b_img_prompt = req['img_prompt']
+            if b_img_prompt:
+                num_img = req['num_img']
+                imgsz = req['imgsz']
+
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                if b_img_prompt:
+                    token_ids, diff = instance.encode_with_image(prompt, num_img, imgsz, last_reply)
+                else:
+                    token_ids, diff = instance.encode(prompt, last_reply)
+                msg = json.dumps({'token_ids': token_ids, 'diff': diff})
+
+        elif '/decode' in self.path:
+            uid = req.get('uid')
+            token_ids = req.get('token_ids')
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                text = instance.decode(token_ids)
+                msg = json.dumps({'text': text})
+
+        elif '/reset' in self.path:
+            uid = req.get("uid")
+            system_prompt = req.get("system_prompt")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                if system_prompt is not None:
+                    print("system_prompt:", system_prompt)
+                    token_ids = instance.reset(system_prompt)
+                else:
+                    token_ids = instance.reset()
+                msg = json.dumps({'token_ids': token_ids})
+
+        else:
+            msg = json.dumps({'error': 'Invalid POST endpoint'})
+
+        self.wfile.write(msg.encode())
+
+    def get_query_param(self, key):
+        from urllib.parse import urlparse, parse_qs
+        query = urlparse(self.path).query
+        params = parse_qs(query)
+        values = params.get(key)
+        return values[0] if values else None
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", type=str, default="localhost")
+    parser.add_argument("--port", type=int, default=8080)
+    parser.add_argument('--model_id', type=str, default='internvl3_tokenizer')
+    parser.add_argument('--content', type=str, default='你是由上海人工智能实验室联合商汤科技开发的书生多模态大模型，英文名叫InternVL, 是一个有用无害的人工智能助手。')
+    args = parser.parse_args()
+
+    host = (args.host, args.port)
+    print("http://%s:%s" % host)
+    server = HTTPServer(host, Request)
+    server.serve_forever()
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/main.cpp b/projects/llm_framework/main_vlm/src/main.cpp
index 83a56e51..0aaaae66 100644
--- a/projects/llm_framework/main_vlm/src/main.cpp
+++ b/projects/llm_framework/main_vlm/src/main.cpp
@@ -135,6 +135,7 @@ class llm_task {
             CONFIG_AUTO_SET(file_body["mode_param"], filename_tokens_embed);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_post_axmodel);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_vpm_resampler_axmodedl);
+            CONFIG_AUTO_SET(file_body["mode_param"], filename_image_encoder_axmodedl);
             CONFIG_AUTO_SET(file_body["mode_param"], template_filename_axmodel);
             CONFIG_AUTO_SET(file_body["mode_param"], b_use_topk);
             CONFIG_AUTO_SET(file_body["mode_param"], b_vpm_two_stage);
@@ -218,6 +219,7 @@ class llm_task {
             mode_config_.filename_post_axmodel           = base_model + mode_config_.filename_post_axmodel;
             mode_config_.template_filename_axmodel       = base_model + mode_config_.template_filename_axmodel;
             mode_config_.filename_vpm_resampler_axmodedl = base_model + mode_config_.filename_vpm_resampler_axmodedl;
+            mode_config_.filename_image_encoder_axmodedl = base_model + mode_config_.filename_image_encoder_axmodedl;
             mode_config_.runing_callback = [this](int *p_token, int n_token, const char *p_str, float token_per_sec,
                                                   void *reserve) {
                 if (this->out_callback_) {
@@ -341,7 +343,7 @@ class llm_task {
 
             if (lLaMa_ctx_) {
                 if (image_data_.empty()) {
-                    lLaMa_ctx_->Encode(prompt_data_, prompt_complete(prompt_), last_reply, tokens_ids, tokens_diff);
+                    lLaMa_ctx_->Encode(prompt_data_, prompt_complete(msg), last_reply, tokens_ids, tokens_diff);
                     if (auto ret = lLaMa_ctx_->SetKVCache(k_caches, v_caches, precompute_len, tokens_diff.size());
                         ret != 0) {
                         ALOGE("SetKVCache failed: %d,the context may be full,input \"reset\" to reset context", ret);
@@ -349,28 +351,31 @@ class llm_task {
                     }
                     last_reply = lLaMa_ctx_->Run(prompt_data_);
                     lLaMa_ctx_->GetKVCache(k_caches, v_caches, precompute_len);
+                    if (out_callback_) out_callback_(last_reply, true);
+                } else {
+                    cv::Mat src = cv::imdecode(image_data_, cv::IMREAD_COLOR);
+                    if (src.empty()) return;
+                    image_data_.clear();
+                    std::vector<unsigned short> img_embed;
+                    if (auto ret = lLaMa_ctx_->Encode(src, img_embed); ret != 0) {
+                        ALOGE("lLaMaCtx.Encode failed");
+                        return;
+                    }
+                    if (auto ret =
+                            lLaMa_ctx_->Encode(img_embed, prompt_data_, prompt_complete(msg), tokens_ids, tokens_diff);
+                        ret != 0) {
+                        ALOGE("lLaMaCtx.Encode failed");
+                        return;
+                    }
+                    if (auto ret = lLaMa_ctx_->SetKVCache(k_caches, v_caches, precompute_len, tokens_diff.size());
+                        ret != 0) {
+                        ALOGE("SetKVCache failed: %d,the context may be full,input \"reset\" to reset context", ret);
+                        return;
+                    }
+                    last_reply = lLaMa_ctx_->Run(prompt_data_);
+                    lLaMa_ctx_->GetKVCache(k_caches, v_caches, precompute_len);
+                    if (out_callback_) out_callback_(last_reply, true);
                 }
-                cv::Mat src = cv::imdecode(image_data_, cv::IMREAD_COLOR);
-                if (src.empty()) return;
-                image_data_.clear();
-                std::vector<unsigned short> img_embed;
-                if (auto ret = lLaMa_ctx_->Encode(src, img_embed); ret != 0) {
-                    ALOGE("lLaMaCtx.Encode failed");
-                    return;
-                }
-                if (auto ret =
-                        lLaMa_ctx_->Encode(img_embed, prompt_data_, prompt_complete(prompt_), tokens_ids, tokens_diff);
-                    ret != 0) {
-                    ALOGE("lLaMaCtx.Encode failed");
-                    return;
-                }
-                if (auto ret = lLaMa_ctx_->SetKVCache(k_caches, v_caches, precompute_len, tokens_diff.size());
-                    ret != 0) {
-                    ALOGE("SetKVCache failed: %d,the context may be full,input \"reset\" to reset context", ret);
-                    return;
-                }
-                last_reply = lLaMa_ctx_->Run(prompt_data_);
-                lLaMa_ctx_->GetKVCache(k_caches, v_caches, precompute_len);
             }
         } catch (...) {
             SLOGW("lLaMa_->Run have error!");
diff --git a/projects/llm_framework/main_vlm/src/runner/LLM.hpp b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
index f7a8b560..c7cb156c 100644
--- a/projects/llm_framework/main_vlm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
@@ -469,20 +469,8 @@ class LLM {
             if (_attr.b_dynamic_load_axmodel_layer) {
                 layer.layer.deinit();
             }
-            // ALOGI("%f %f %f %f %f", bfloat16(embed[0]).fp32(), bfloat16(embed[1]).fp32(), bfloat16(embed[2]).fp32(),
-            // bfloat16(embed[3]).fp32(), bfloat16(embed[4]).fp32());
         }
 
-        // ALOGI("prefill time cost: %.2f s", t_cost.cost() / 1000);
-
-        // print token_ids
-        // printf("%s\n", input_str.c_str());
-        // for (size_t i = 0; i < token_ids.size(); i++)
-        // {
-        //     printf("%d ", token_ids[i]);
-        // }
-        // printf("\n");
-
         int next_token = -1;
         t_cqdm cqdm    = create_cqdm(_attr.max_token_len, 32);
         std::vector<unsigned short> embed(_attr.tokens_embed_size, 0);
@@ -522,10 +510,7 @@ class LLM {
                 break;
             }
 
-            // ALOGI("out %d %d", indices, next_token);
             embed_selector.getByIndex(next_token, embed);
-            // ALOGI("%f %f %f %f %f", bfloat16(embed[0]).fp32(), bfloat16(embed[1]).fp32(), bfloat16(embed[2]).fp32(),
-            // bfloat16(embed[3]).fp32(), bfloat16(embed[4]).fp32());
 
             for (int m = 0; m < _attr.axmodel_num; m++) {
                 if (b_stop) {
@@ -580,10 +565,8 @@ class LLM {
                 if (_attr.b_dynamic_load_axmodel_layer) {
                     layer.layer.deinit();
                 }
-                // ALOGI("%f %f %f %f %f", bfloat16(embed[0]).fp32(), bfloat16(embed[1]).fp32(),
-                // bfloat16(embed[2]).fp32(), bfloat16(embed[3]).fp32(), bfloat16(embed[4]).fp32());
             }
-            // ALOGI("");
+
             mask[indices] = 0;
             {
                 // post process
@@ -644,9 +627,6 @@ class LLM {
         float t_cost_ms = t_cost.cost();
         ALOGN("hit eos,avg %.2f token/s\n", token_ids.size() / (t_cost_ms / 1000));
 
-        // 去掉 len_of_input 那部分
-        // token_ids.erase(token_ids.begin(), token_ids.begin() + len_of_input);
-
         final_out = tokenizer->Decode(token_ids);
 
         return final_out;
@@ -1358,23 +1338,6 @@ class LLM_CTX {
         return 0;
     }
 
-    int Encode(std::vector<unsigned short> &out_embed, std::string prompt = "What is in the image?")
-    {
-        ImageInfo img_info;
-        img_info.img_prompt        = false;
-        std::vector<int> input_ids = tokenizer->Encode(prompt, img_info);
-        if (input_ids.size() > _attr.prefill_token_num) {
-            ALOGE("input_ids(%ld) > prefill_token_num(%d)", input_ids.size(), _attr.prefill_token_num);
-            return -1;
-        }
-        out_embed.resize(input_ids.size() * _attr.tokens_embed_size);
-
-        for (size_t i = 0; i < input_ids.size(); i++) {
-            embed_selector.getByIndex(input_ids[i], out_embed.data() + i * _attr.tokens_embed_size);
-        }
-
-        return 0;
-    }
 
     int Encode(std::vector<std::vector<unsigned short>> &imgs_embed, std::vector<unsigned short> &out_embed,
                std::string prompt, std::vector<int> &tokens_ids, std::vector<int> &tokens_diff)

From 57404bc4efda7aac92c406dd8b2b60c01805541e Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 29 Aug 2025 16:42:36 +0800
Subject: [PATCH 34/79] [update] add internvl3-1B config file, update
 postprocess.

---
 .../models/mode_internvl3-1B-ax630c.json      | 37 +++++++++++++++
 .../llm_framework/main_vlm/src/runner/LLM.hpp | 19 +++++++-
 .../main_vlm/src/runner/LLMPostprocess.hpp    | 46 ++++++++++++++++++-
 3 files changed, 100 insertions(+), 2 deletions(-)
 create mode 100644 projects/llm_framework/main_vlm/models/mode_internvl3-1B-ax630c.json

diff --git a/projects/llm_framework/main_vlm/models/mode_internvl3-1B-ax630c.json b/projects/llm_framework/main_vlm/models/mode_internvl3-1B-ax630c.json
new file mode 100644
index 00000000..50848e20
--- /dev/null
+++ b/projects/llm_framework/main_vlm/models/mode_internvl3-1B-ax630c.json
@@ -0,0 +1,37 @@
+{
+    "mode": "internvl3-1B-ax630c",
+    "type": "vlm",
+    "homepage": "https://huggingface.co/AXERA-TECH/InternVL3-1B",
+    "capabilities": [
+        "text_generation",
+        "chat"
+    ],
+    "input_type": [
+        "vlm.chat_completion",
+        "vlm.chat_completion.stream"
+    ],
+    "output_type": [
+        "vlm.utf-8",
+        "vlm.utf-8.stream"
+    ],
+    "mode_param": {
+        "tokenizer_type": 2,
+        "url_tokenizer_model": "http://localhost:8080",
+        "filename_tokens_embed": "model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel": "qwen2_post.axmodel",
+        "template_filename_axmodel": "qwen2_p128_l%d_together.axmodel",
+        "filename_image_encoder_axmodedl": "internvl3_1b_vit.axmodel",
+        "b_bos": false,
+        "b_eos": false,
+        "axmodel_num": 24,
+        "tokens_embed_num": 151674,
+        "img_token_id": 151667,
+        "tokens_embed_size": 896,
+        "b_use_mmap_load_embed": true,
+        "b_dynamic_load_axmodel_layer": false,
+        "precompute_len": 1024,
+        "ext_scripts": [
+            "tokenizer_internvl3-1B-ax630c.py"
+        ]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/LLM.hpp b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
index c7cb156c..2fd7bff5 100644
--- a/projects/llm_framework/main_vlm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
@@ -801,11 +801,29 @@ class LLM_CTX {
                 _attr.prefill_max_kv_cache_num_grp[_attr.prefill_max_kv_cache_num_grp.size() - 1];
             ALOGI("prefill_max_token_num : %d", _attr.prefill_max_token_num);
         }
+        nlohmann::json dynamic_config;
+
+        dynamic_config["enable_temperature"] = _attr.enable_temperature;
+        dynamic_config["temperature"]        = _attr.temperature;
+
+        dynamic_config["enable_repetition_penalty"] = _attr.enable_repetition_penalty;
+        dynamic_config["repetition_penalty"]        = _attr.repetition_penalty;
+        dynamic_config["penalty_window"]            = _attr.penalty_window;
+
+        dynamic_config["enable_top_p_sampling"] = _attr.enable_top_p_sampling;
+        dynamic_config["top_p"]                 = _attr.top_p;
+
+        dynamic_config["enable_top_k_sampling"] = _attr.enable_top_k_sampling;
+        dynamic_config["top_k"]                 = _attr.top_k;
 
         if (!postprocess.load_config(attr.post_config_path)) {
             ALOGW("load postprocess config(%s) failed", attr.post_config_path.c_str());
         }
 
+        if (!postprocess.load_config(dynamic_config)) {
+            ALOGW("load postprocess config(%s) failed", dynamic_config.dump(4).c_str());
+        }
+
         ALOGI("LLM init ok");
         return true;
     }
@@ -1338,7 +1356,6 @@ class LLM_CTX {
         return 0;
     }
 
-
     int Encode(std::vector<std::vector<unsigned short>> &imgs_embed, std::vector<unsigned short> &out_embed,
                std::string prompt, std::vector<int> &tokens_ids, std::vector<int> &tokens_diff)
     {
diff --git a/projects/llm_framework/main_vlm/src/runner/LLMPostprocess.hpp b/projects/llm_framework/main_vlm/src/runner/LLMPostprocess.hpp
index c98205e4..b7d50156 100644
--- a/projects/llm_framework/main_vlm/src/runner/LLMPostprocess.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/LLMPostprocess.hpp
@@ -242,10 +242,11 @@ class LLMPostprocess
         this->temperature = temperature;
     }
 
-    void set_repetition_penalty(bool enable, float penalty)
+    void set_repetition_penalty(bool enable, float penalty, int penalty_window)
     {
         enable_repetition_penalty = enable;
         this->repetition_penalty = penalty;
+        this->penalty_window = penalty_window;
     }
 
     void set_diversity_penalty(bool enable, const std::vector<int> &common_phrases, float penalty)
@@ -295,6 +296,49 @@ class LLMPostprocess
         return true;
     }
 
+    bool load_config(const nlohmann::json& config)
+    {
+        if (config.is_null()) {
+            ALOGE("config is null or invalid");
+            return false;
+        }
+
+        ALOGI("load config: \n%s\n", config.dump(4).c_str());
+
+        if (config.contains("enable_temperature")) {
+            enable_temperature = config["enable_temperature"].get<bool>();
+        }
+        if (config.contains("temperature")) {
+            temperature = config["temperature"].get<float>();
+        }
+
+        if (config.contains("enable_repetition_penalty")) {
+            enable_repetition_penalty = config["enable_repetition_penalty"].get<bool>();
+        }
+        if (config.contains("repetition_penalty")) {
+            repetition_penalty = config["repetition_penalty"].get<float>();
+        }
+        if (config.contains("penalty_window")) {
+            penalty_window = config["penalty_window"].get<int>();
+        }
+
+        if (config.contains("enable_top_p_sampling")) {
+            enable_top_p_sampling = config["enable_top_p_sampling"].get<bool>();
+        }
+        if (config.contains("top_p")) {
+            top_p = config["top_p"].get<float>();
+        }
+
+        if (config.contains("enable_top_k_sampling")) {
+            enable_top_k_sampling = config["enable_top_k_sampling"].get<bool>();
+        }
+        if (config.contains("top_k")) {
+            top_k = config["top_k"].get<int>();
+        }
+
+        return true;
+    }
+
     int apply(std::vector<float> &logits, const std::vector<int> &history)
     {
         if (enable_temperature)

From 2de874ce577a11b3b87363e4f3b3f90d032c05a5 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Wed, 3 Sep 2025 19:05:07 +0800
Subject: [PATCH 35/79] [update] update llm & vlm

---
 projects/llm_framework/main_llm/src/main.cpp  | 12 ++++-
 projects/llm_framework/main_vlm/src/main.cpp  | 50 +++++++++++++++----
 .../llm_framework/main_vlm/src/runner/LLM.hpp | 37 ++++++++------
 3 files changed, 71 insertions(+), 28 deletions(-)

diff --git a/projects/llm_framework/main_llm/src/main.cpp b/projects/llm_framework/main_llm/src/main.cpp
index dcb64fb8..a1cb066c 100644
--- a/projects/llm_framework/main_llm/src/main.cpp
+++ b/projects/llm_framework/main_llm/src/main.cpp
@@ -29,7 +29,7 @@ using namespace StackFlows;
 int main_exit_flage = 0;
 static void __sigint(int iSigNo)
 {
-    SLOGW("llm_sys will be exit!");
+    SLOGW("llm_llm will be exit!");
     main_exit_flage = 1;
 }
 
@@ -130,7 +130,7 @@ class llm_task {
             std::string base_model = base_model_path_ + model_ + "/";
             SLOGI("base_model %s", base_model.c_str());
 
-            CONFIG_AUTO_SET(file_body["mode_param"], system_prompt);   
+            CONFIG_AUTO_SET(file_body["mode_param"], system_prompt);
             CONFIG_AUTO_SET(file_body["mode_param"], tokenizer_type);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_tokenizer_model);
             CONFIG_AUTO_SET(file_body["mode_param"], url_tokenizer_model);
@@ -325,6 +325,14 @@ class llm_task {
             }
 
             if (lLaMa_ctx_) {
+                if (msg == "reset") {
+                    lLaMa_ctx_->SetSystemPrompt(mode_config_.system_prompt, _token_ids);
+                    lLaMa_ctx_->GenerateKVCachePrefill(_token_ids, k_caches, v_caches, precompute_len);
+                    last_reply.clear();
+                    if (out_callback_) out_callback_("Context has been reset.", true);
+                    return;
+                }
+
                 lLaMa_ctx_->Encode(prompt_data, prompt_complete(msg), last_reply, tokens_ids, tokens_diff);
                 if (auto ret = lLaMa_ctx_->SetKVCache(k_caches, v_caches, precompute_len, tokens_diff.size());
                     ret != 0) {
diff --git a/projects/llm_framework/main_vlm/src/main.cpp b/projects/llm_framework/main_vlm/src/main.cpp
index 0aaaae66..6460cff1 100644
--- a/projects/llm_framework/main_vlm/src/main.cpp
+++ b/projects/llm_framework/main_vlm/src/main.cpp
@@ -58,6 +58,8 @@ class llm_task {
     std::vector<std::string> inputs_;
     std::vector<unsigned short> prompt_data_;
     std::vector<unsigned char> image_data_;
+    std::vector<std::vector<unsigned char>> images_data;
+    std::vector<cv::Mat> mats;
     std::vector<unsigned short> img_embed;
     std::string prompt_;
     std::string last_reply;
@@ -134,8 +136,8 @@ class llm_task {
             CONFIG_AUTO_SET(file_body["mode_param"], url_tokenizer_model);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_tokens_embed);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_post_axmodel);
-            CONFIG_AUTO_SET(file_body["mode_param"], filename_vpm_resampler_axmodedl);
-            CONFIG_AUTO_SET(file_body["mode_param"], filename_image_encoder_axmodedl);
+            CONFIG_AUTO_SET(file_body["mode_param"], filename_vpm_resampler_axmodel);
+            CONFIG_AUTO_SET(file_body["mode_param"], filename_image_encoder_axmodel);
             CONFIG_AUTO_SET(file_body["mode_param"], template_filename_axmodel);
             CONFIG_AUTO_SET(file_body["mode_param"], b_use_topk);
             CONFIG_AUTO_SET(file_body["mode_param"], b_vpm_two_stage);
@@ -215,11 +217,11 @@ class llm_task {
                     SLOGE("filename_tokenizer_model: %s", mode_config_.filename_tokenizer_model.c_str());
                 }
             }
-            mode_config_.filename_tokens_embed           = base_model + mode_config_.filename_tokens_embed;
-            mode_config_.filename_post_axmodel           = base_model + mode_config_.filename_post_axmodel;
-            mode_config_.template_filename_axmodel       = base_model + mode_config_.template_filename_axmodel;
-            mode_config_.filename_vpm_resampler_axmodedl = base_model + mode_config_.filename_vpm_resampler_axmodedl;
-            mode_config_.filename_image_encoder_axmodedl = base_model + mode_config_.filename_image_encoder_axmodedl;
+            mode_config_.filename_tokens_embed          = base_model + mode_config_.filename_tokens_embed;
+            mode_config_.filename_post_axmodel          = base_model + mode_config_.filename_post_axmodel;
+            mode_config_.template_filename_axmodel      = base_model + mode_config_.template_filename_axmodel;
+            mode_config_.filename_vpm_resampler_axmodel = base_model + mode_config_.filename_vpm_resampler_axmodel;
+            mode_config_.filename_image_encoder_axmodel = base_model + mode_config_.filename_image_encoder_axmodel;
             mode_config_.runing_callback = [this](int *p_token, int n_token, const char *p_str, float token_per_sec,
                                                   void *reserve) {
                 if (this->out_callback_) {
@@ -342,17 +344,38 @@ class llm_task {
             }
 
             if (lLaMa_ctx_) {
+                if (msg == "reset") {
+                    lLaMa_ctx_->SetSystemPrompt(mode_config_.system_prompt, _token_ids);
+                    lLaMa_ctx_->GenerateKVCachePrefill(_token_ids, k_caches, v_caches, precompute_len);
+                    last_reply.clear();
+                    mats.clear();
+                    if (out_callback_) out_callback_("Context has been reset.", true);
+                    return;
+                }
+
                 if (image_data_.empty()) {
                     lLaMa_ctx_->Encode(prompt_data_, prompt_complete(msg), last_reply, tokens_ids, tokens_diff);
                     if (auto ret = lLaMa_ctx_->SetKVCache(k_caches, v_caches, precompute_len, tokens_diff.size());
                         ret != 0) {
-                        ALOGE("SetKVCache failed: %d,the context may be full,input \"reset\" to reset context", ret);
-                        return;
+                        ALOGW("The context full,Reset context");
+                        lLaMa_ctx_->SetSystemPrompt(mode_config_.system_prompt, _token_ids);
+                        lLaMa_ctx_->GenerateKVCachePrefill(_token_ids, k_caches, v_caches, precompute_len);
+                        lLaMa_ctx_->SetKVCache(k_caches, v_caches, precompute_len, tokens_diff.size());
                     }
                     last_reply = lLaMa_ctx_->Run(prompt_data_);
                     lLaMa_ctx_->GetKVCache(k_caches, v_caches, precompute_len);
                     if (out_callback_) out_callback_(last_reply, true);
                 } else {
+                    for (const auto &img_buf : images_data) {
+                        cv::Mat src = cv::imdecode(img_buf, cv::IMREAD_COLOR);
+                        if (src.empty()) {
+                            std::cerr << "Decode failed!" << std::endl;
+                            continue;
+                        }
+                        mats.push_back(src);
+                    }
+                    if (mats.empty()) return;
+                    images_data.clear();
                     cv::Mat src = cv::imdecode(image_data_, cv::IMREAD_COLOR);
                     if (src.empty()) return;
                     image_data_.clear();
@@ -361,6 +384,7 @@ class llm_task {
                         ALOGE("lLaMaCtx.Encode failed");
                         return;
                     }
+                    mats.clear();
                     if (auto ret =
                             lLaMa_ctx_->Encode(img_embed, prompt_data_, prompt_complete(msg), tokens_ids, tokens_diff);
                         ret != 0) {
@@ -369,8 +393,11 @@ class llm_task {
                     }
                     if (auto ret = lLaMa_ctx_->SetKVCache(k_caches, v_caches, precompute_len, tokens_diff.size());
                         ret != 0) {
-                        ALOGE("SetKVCache failed: %d,the context may be full,input \"reset\" to reset context", ret);
-                        return;
+                        ALOGW("The context full,Reset context");
+                        lLaMa_ctx_->SetSystemPrompt(mode_config_.system_prompt, _token_ids);
+                        lLaMa_ctx_->GenerateKVCachePrefill(_token_ids, k_caches, v_caches, precompute_len);
+                        lLaMa_ctx_->SetKVCache(k_caches, v_caches, precompute_len, tokens_diff.size());
+                        lLaMa_ctx_->ClearImgsEmbed();
                     }
                     last_reply = lLaMa_ctx_->Run(prompt_data_);
                     lLaMa_ctx_->GetKVCache(k_caches, v_caches, precompute_len);
@@ -549,6 +576,7 @@ class llm_vlm : public StackFlow {
         }
         if (object.find("jpeg") != std::string::npos) {
             llm_task_obj->image_data_.assign(next_data->begin(), next_data->end());
+            llm_task_obj->images_data.emplace_back(next_data->begin(), next_data->end());
             return;
         }
         llm_task_obj->inference((*next_data));
diff --git a/projects/llm_framework/main_vlm/src/runner/LLM.hpp b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
index 2fd7bff5..ff232473 100644
--- a/projects/llm_framework/main_vlm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
@@ -23,10 +23,10 @@ struct LLMAttrType {
     std::string template_filename_axmodel = "tinyllama-int8/tinyllama_l%d.axmodel";
     int axmodel_num                       = 22;
 
-    std::string filename_post_axmodel           = "tinyllama-int8/tinyllama_post.axmodel";
-    std::string filename_image_encoder_axmodedl = "minicpmv/vpm_resampler_version0_fp16.axmodel";
-    std::string filename_vpm_encoder_axmodedl   = "minicpmv/vpm_resampler_version0_fp16.axmodel";
-    std::string filename_vpm_resampler_axmodedl = "minicpmv/vpm_resampler_version0_fp16.axmodel";
+    std::string filename_post_axmodel          = "tinyllama-int8/tinyllama_post.axmodel";
+    std::string filename_image_encoder_axmodel = "minicpmv/vpm_resampler_version0_fp16.axmodel";
+    std::string filename_vpm_encoder_axmodel   = "minicpmv/vpm_resampler_version0_fp16.axmodel";
+    std::string filename_vpm_resampler_axmodel = "minicpmv/vpm_resampler_version0_fp16.axmodel";
 
     int image_encoder_width  = 448;
     int image_encoder_height = 448;
@@ -184,24 +184,24 @@ class LLM {
         update_cqdm(&cqdm, attr.axmodel_num + 2, "count", axmodel_path);
 
         if (_attr.b_vpm_two_stage) {
-            ret = vpm_encoder.init(attr.filename_vpm_encoder_axmodedl.c_str(), false);
+            ret = vpm_encoder.init(attr.filename_vpm_encoder_axmodel.c_str(), false);
             if (ret != 0) {
-                ALOGE("init vpm axmodel(%s) failed", attr.filename_vpm_encoder_axmodedl.c_str());
+                ALOGE("init vpm axmodel(%s) failed", attr.filename_vpm_encoder_axmodel.c_str());
                 return false;
             }
 
-            ret = vpm_resampler.init(attr.filename_vpm_resampler_axmodedl.c_str(), false);
+            ret = vpm_resampler.init(attr.filename_vpm_resampler_axmodel.c_str(), false);
             if (ret != 0) {
-                ALOGE("init vpm axmodel(%s) failed", attr.filename_vpm_resampler_axmodedl.c_str());
+                ALOGE("init vpm axmodel(%s) failed", attr.filename_vpm_resampler_axmodel.c_str());
                 return false;
             }
 
             _attr.vpm_height = vpm_encoder.get_input(0).vShape[1];
             _attr.vpm_width  = vpm_encoder.get_input(0).vShape[2];
         } else {
-            ret = vpm_resampler.init(attr.filename_vpm_resampler_axmodedl.c_str(), false);
+            ret = vpm_resampler.init(attr.filename_vpm_resampler_axmodel.c_str(), false);
             if (ret != 0) {
-                ALOGE("init vpm axmodel(%s) failed", attr.filename_vpm_resampler_axmodedl.c_str());
+                ALOGE("init vpm axmodel(%s) failed", attr.filename_vpm_resampler_axmodel.c_str());
                 return false;
             }
             _attr.vpm_height = vpm_resampler.get_input(0).vShape[1];
@@ -637,6 +637,7 @@ class LLM_CTX {
 private:
     std::shared_ptr<BaseTokenizer> tokenizer;
     LLaMaEmbedSelector embed_selector;
+    std::vector<std::vector<unsigned short>> imgs_embed_;
 
     LLMAttrType _attr;
 
@@ -718,9 +719,9 @@ class LLM_CTX {
         sprintf(axmodel_path, "init post axmodel ok,remain_cmm(%d MB)", remain_cmm);
         update_cqdm(&cqdm, attr.axmodel_num + 2, "count", axmodel_path);
 
-        ret = image_encoder.init(attr.filename_image_encoder_axmodedl.c_str());
+        ret = image_encoder.init(attr.filename_image_encoder_axmodel.c_str());
         if (ret != 0) {
-            ALOGE("init vpm axmodel(%s) failed", attr.filename_image_encoder_axmodedl.c_str());
+            ALOGE("init vpm axmodel(%s) failed", attr.filename_image_encoder_axmodel.c_str());
             return false;
         }
 
@@ -1424,8 +1425,9 @@ class LLM_CTX {
     int Encode(std::vector<unsigned short> &img_embed, std::vector<unsigned short> &out_embed, std::string prompt,
                std::vector<int> &tokens_ids, std::vector<int> &tokens_diff)
     {
-        std::vector<std::vector<unsigned short>> imgs_embed = {img_embed};
-        return Encode(imgs_embed, out_embed, prompt, tokens_ids, tokens_diff);
+        // std::vector<std::vector<unsigned short>> imgs_embed = {img_embed};
+        imgs_embed_.push_back(img_embed);
+        return Encode(imgs_embed_, out_embed, prompt, tokens_ids, tokens_diff);
     }
 
     int Encode(std::vector<unsigned short> &out_embed, std::string prompt, std::string last_reply,
@@ -1447,7 +1449,12 @@ class LLM_CTX {
         return 0;
     }
 
-    std::string Run(std::vector<unsigned short> test_embed)
+    void ClearImgsEmbed()
+    {
+        imgs_embed_.clear();
+    }
+
+    std::string Run(std::vector<unsigned short> &test_embed)
     {
         b_stop = false;
         std::string final_out;

From b6d6e95eefc9e713e0439dc8eea1357e4b2d1791 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 4 Sep 2025 11:02:02 +0800
Subject: [PATCH 36/79] [update] move public include into static_lib, update
 llm & vlm

---
 projects/llm_framework/SConstruct             |      2 +-
 .../abseil-cpp/absl/algorithm/algorithm.h     |     64 -
 .../abseil-cpp/absl/algorithm/container.h     |   1760 -
 .../include/abseil-cpp/absl/base/attributes.h |    921 -
 .../include/abseil-cpp/absl/base/call_once.h  |    225 -
 .../include/abseil-cpp/absl/base/casts.h      |    180 -
 .../include/abseil-cpp/absl/base/config.h     |   1000 -
 .../include/abseil-cpp/absl/base/const_init.h |     76 -
 .../absl/base/dynamic_annotations.h           |    496 -
 .../absl/base/internal/atomic_hook.h          |    200 -
 .../base/internal/atomic_hook_test_helper.h   |     34 -
 .../absl/base/internal/cycleclock.h           |    144 -
 .../absl/base/internal/cycleclock_config.h    |     55 -
 .../absl/base/internal/direct_mmap.h          |    170 -
 .../absl/base/internal/dynamic_annotations.h  |    398 -
 .../abseil-cpp/absl/base/internal/endian.h    |    283 -
 .../absl/base/internal/errno_saver.h          |     43 -
 .../base/internal/exception_safety_testing.h  |   1109 -
 .../absl/base/internal/exception_testing.h    |     42 -
 .../absl/base/internal/fast_type_id.h         |     50 -
 .../abseil-cpp/absl/base/internal/hide_ptr.h  |     51 -
 .../abseil-cpp/absl/base/internal/identity.h  |     39 -
 .../absl/base/internal/inline_variable.h      |    108 -
 .../base/internal/inline_variable_testing.h   |     46 -
 .../abseil-cpp/absl/base/internal/invoke.h    |    241 -
 .../absl/base/internal/low_level_alloc.h      |    127 -
 .../absl/base/internal/low_level_scheduling.h |    134 -
 .../absl/base/internal/nullability_impl.h     |    106 -
 .../absl/base/internal/per_thread_tls.h       |     52 -
 .../absl/base/internal/pretty_function.h      |     33 -
 .../absl/base/internal/raw_logging.h          |    217 -
 .../absl/base/internal/scheduling_mode.h      |     58 -
 .../absl/base/internal/scoped_set_env.h       |     45 -
 .../abseil-cpp/absl/base/internal/spinlock.h  |    274 -
 .../absl/base/internal/spinlock_wait.h        |     95 -
 .../abseil-cpp/absl/base/internal/strerror.h  |     39 -
 .../abseil-cpp/absl/base/internal/sysinfo.h   |     74 -
 .../absl/base/internal/thread_identity.h      |    269 -
 .../absl/base/internal/throw_delegate.h       |     75 -
 .../absl/base/internal/tsan_mutex_interface.h |     68 -
 .../absl/base/internal/unaligned_access.h     |     89 -
 .../absl/base/internal/unscaledcycleclock.h   |     96 -
 .../base/internal/unscaledcycleclock_config.h |     62 -
 .../abseil-cpp/absl/base/log_severity.h       |    185 -
 .../include/abseil-cpp/absl/base/macros.h     |    177 -
 .../abseil-cpp/absl/base/no_destructor.h      |    218 -
 .../abseil-cpp/absl/base/nullability.h        |    224 -
 .../abseil-cpp/absl/base/optimization.h       |    305 -
 .../include/abseil-cpp/absl/base/options.h    |    258 -
 .../abseil-cpp/absl/base/policy_checks.h      |    113 -
 .../include/abseil-cpp/absl/base/port.h       |     25 -
 .../include/abseil-cpp/absl/base/prefetch.h   |    209 -
 .../abseil-cpp/absl/base/thread_annotations.h |    333 -
 .../include/abseil-cpp/absl/cleanup/cleanup.h |    140 -
 .../absl/cleanup/internal/cleanup.h           |    100 -
 .../abseil-cpp/absl/container/btree_map.h     |    887 -
 .../abseil-cpp/absl/container/btree_set.h     |    821 -
 .../abseil-cpp/absl/container/btree_test.h    |    166 -
 .../abseil-cpp/absl/container/fixed_array.h   |    556 -
 .../abseil-cpp/absl/container/flat_hash_map.h |    638 -
 .../abseil-cpp/absl/container/flat_hash_set.h |    529 -
 .../absl/container/inlined_vector.h           |   1015 -
 .../absl/container/internal/btree.h           |   3053 -
 .../absl/container/internal/btree_container.h |    763 -
 .../absl/container/internal/common.h          |    207 -
 .../container/internal/common_policy_traits.h |    143 -
 .../container/internal/compressed_tuple.h     |    271 -
 .../container/internal/container_memory.h     |    491 -
 .../internal/hash_function_defaults.h         |    276 -
 .../internal/hash_generator_testing.h         |    182 -
 .../container/internal/hash_policy_testing.h  |    183 -
 .../container/internal/hash_policy_traits.h   |    207 -
 .../absl/container/internal/hashtable_debug.h |    102 -
 .../internal/hashtable_debug_hooks.h          |     85 -
 .../container/internal/hashtablez_sampler.h   |    257 -
 .../absl/container/internal/inlined_vector.h  |   1114 -
 .../absl/container/internal/layout.h          |    728 -
 .../container/internal/node_slot_policy.h     |     95 -
 .../absl/container/internal/raw_hash_map.h    |    226 -
 .../absl/container/internal/raw_hash_set.h    |   3904 -
 .../absl/container/internal/test_allocator.h  |    387 -
 .../internal/test_instance_tracker.h          |    274 -
 .../absl/container/internal/tracked.h         |     83 -
 .../internal/unordered_map_constructor_test.h |    494 -
 .../internal/unordered_map_lookup_test.h      |    117 -
 .../internal/unordered_map_members_test.h     |     87 -
 .../internal/unordered_map_modifiers_test.h   |    352 -
 .../internal/unordered_set_constructor_test.h |    496 -
 .../internal/unordered_set_lookup_test.h      |     91 -
 .../internal/unordered_set_members_test.h     |     86 -
 .../internal/unordered_set_modifiers_test.h   |    221 -
 .../abseil-cpp/absl/container/node_hash_map.h |    629 -
 .../abseil-cpp/absl/container/node_hash_set.h |    524 -
 .../include/abseil-cpp/absl/crc/crc32c.h      |    190 -
 .../abseil-cpp/absl/crc/internal/cpu_detect.h |     63 -
 .../abseil-cpp/absl/crc/internal/crc.h        |     83 -
 .../internal/crc32_x86_arm_combined_simd.h    |    299 -
 .../abseil-cpp/absl/crc/internal/crc32c.h     |     39 -
 .../absl/crc/internal/crc32c_inline.h         |     72 -
 .../absl/crc/internal/crc_cord_state.h        |    159 -
 .../absl/crc/internal/crc_internal.h          |    177 -
 .../abseil-cpp/absl/crc/internal/crc_memcpy.h |    122 -
 .../internal/non_temporal_arm_intrinsics.h    |     79 -
 .../absl/crc/internal/non_temporal_memcpy.h   |    180 -
 .../absl/debugging/failure_signal_handler.h   |    121 -
 .../debugging/internal/address_is_readable.h  |     32 -
 .../absl/debugging/internal/demangle.h        |     73 -
 .../absl/debugging/internal/elf_mem_image.h   |    140 -
 .../absl/debugging/internal/examine_stack.h   |     64 -
 .../debugging/internal/stack_consumption.h    |     50 -
 .../debugging/internal/stacktrace_config.h    |     88 -
 .../absl/debugging/internal/symbolize.h       |    153 -
 .../absl/debugging/internal/vdso_support.h    |    158 -
 .../abseil-cpp/absl/debugging/leak_check.h    |    150 -
 .../abseil-cpp/absl/debugging/stacktrace.h    |    231 -
 .../abseil-cpp/absl/debugging/symbolize.h     |     99 -
 .../abseil-cpp/absl/flags/commandlineflag.h   |    211 -
 .../include/abseil-cpp/absl/flags/config.h    |     68 -
 .../include/abseil-cpp/absl/flags/declare.h   |     68 -
 .../include/abseil-cpp/absl/flags/flag.h      |    303 -
 .../absl/flags/internal/commandlineflag.h     |     68 -
 .../abseil-cpp/absl/flags/internal/flag.h     |    805 -
 .../abseil-cpp/absl/flags/internal/parse.h    |     70 -
 .../absl/flags/internal/path_util.h           |     62 -
 .../flags/internal/private_handle_accessor.h  |     61 -
 .../absl/flags/internal/program_name.h        |     50 -
 .../abseil-cpp/absl/flags/internal/registry.h |     97 -
 .../absl/flags/internal/sequence_lock.h       |    187 -
 .../abseil-cpp/absl/flags/internal/usage.h    |    106 -
 .../abseil-cpp/absl/flags/marshalling.h       |    361 -
 .../include/abseil-cpp/absl/flags/parse.h     |    130 -
 .../abseil-cpp/absl/flags/reflection.h        |     90 -
 .../include/abseil-cpp/absl/flags/usage.h     |     43 -
 .../abseil-cpp/absl/flags/usage_config.h      |    135 -
 .../absl/functional/any_invocable.h           |    324 -
 .../abseil-cpp/absl/functional/bind_front.h   |    194 -
 .../abseil-cpp/absl/functional/function_ref.h |    151 -
 .../absl/functional/internal/any_invocable.h  |    891 -
 .../absl/functional/internal/front_binder.h   |     95 -
 .../absl/functional/internal/function_ref.h   |    116 -
 .../abseil-cpp/absl/functional/overload.h     |     82 -
 .../include/abseil-cpp/absl/hash/hash.h       |    424 -
 .../abseil-cpp/absl/hash/hash_testing.h       |    378 -
 .../abseil-cpp/absl/hash/internal/city.h      |     78 -
 .../abseil-cpp/absl/hash/internal/hash.h      |   1383 -
 .../abseil-cpp/absl/hash/internal/hash_test.h |     87 -
 .../absl/hash/internal/low_level_hash.h       |     54 -
 .../absl/hash/internal/spy_hash_state.h       |    266 -
 .../include/abseil-cpp/absl/log/absl_check.h  |    117 -
 .../include/abseil-cpp/absl/log/absl_log.h    |    115 -
 .../abseil-cpp/absl/log/absl_vlog_is_on.h     |     93 -
 .../include/abseil-cpp/absl/log/check.h       |    209 -
 .../include/abseil-cpp/absl/log/die_if_null.h |     76 -
 .../include/abseil-cpp/absl/log/flags.h       |     43 -
 .../include/abseil-cpp/absl/log/globals.h     |    218 -
 .../include/abseil-cpp/absl/log/initialize.h  |     45 -
 .../absl/log/internal/append_truncated.h      |     47 -
 .../abseil-cpp/absl/log/internal/check_impl.h |    150 -
 .../abseil-cpp/absl/log/internal/check_op.h   |    420 -
 .../abseil-cpp/absl/log/internal/conditions.h |    239 -
 .../abseil-cpp/absl/log/internal/config.h     |     45 -
 .../abseil-cpp/absl/log/internal/flags.h      |     59 -
 .../abseil-cpp/absl/log/internal/fnmatch.h    |     35 -
 .../abseil-cpp/absl/log/internal/globals.h    |    101 -
 .../abseil-cpp/absl/log/internal/log_format.h |     78 -
 .../abseil-cpp/absl/log/internal/log_impl.h   |    282 -
 .../absl/log/internal/log_message.h           |    375 -
 .../absl/log/internal/log_sink_set.h          |     54 -
 .../abseil-cpp/absl/log/internal/nullguard.h  |     88 -
 .../abseil-cpp/absl/log/internal/nullstream.h |    136 -
 .../abseil-cpp/absl/log/internal/proto.h      |    288 -
 .../abseil-cpp/absl/log/internal/strip.h      |     72 -
 .../abseil-cpp/absl/log/internal/structured.h |     58 -
 .../absl/log/internal/test_actions.h          |     90 -
 .../absl/log/internal/test_helpers.h          |     71 -
 .../absl/log/internal/test_matchers.h         |     94 -
 .../absl/log/internal/vlog_config.h           |    163 -
 .../abseil-cpp/absl/log/internal/voidify.h    |     44 -
 .../include/abseil-cpp/absl/log/log.h         |    361 -
 .../include/abseil-cpp/absl/log/log_entry.h   |    221 -
 .../include/abseil-cpp/absl/log/log_sink.h    |     71 -
 .../abseil-cpp/absl/log/log_sink_registry.h   |     61 -
 .../abseil-cpp/absl/log/log_streamer.h        |    181 -
 .../abseil-cpp/absl/log/scoped_mock_log.h     |    197 -
 .../include/abseil-cpp/absl/log/structured.h  |     70 -
 .../include/abseil-cpp/absl/log/vlog_is_on.h  |     72 -
 .../include/abseil-cpp/absl/memory/memory.h   |    278 -
 .../abseil-cpp/absl/meta/type_traits.h        |    580 -
 .../include/abseil-cpp/absl/numeric/bits.h    |    196 -
 .../include/abseil-cpp/absl/numeric/int128.h  |   1157 -
 .../abseil-cpp/absl/numeric/internal/bits.h   |    358 -
 .../absl/numeric/internal/representation.h    |     55 -
 .../profiling/internal/exponential_biased.h   |    130 -
 .../profiling/internal/periodic_sampler.h     |    211 -
 .../absl/profiling/internal/sample_recorder.h |    253 -
 .../absl/random/bernoulli_distribution.h      |    200 -
 .../absl/random/beta_distribution.h           |    427 -
 .../abseil-cpp/absl/random/bit_gen_ref.h      |    186 -
 .../absl/random/discrete_distribution.h       |    247 -
 .../abseil-cpp/absl/random/distributions.h    |    452 -
 .../absl/random/exponential_distribution.h    |    165 -
 .../absl/random/gaussian_distribution.h       |    275 -
 .../absl/random/internal/chi_square.h         |     89 -
 .../random/internal/distribution_caller.h     |     95 -
 .../random/internal/distribution_test_util.h  |    113 -
 .../absl/random/internal/explicit_seed_seq.h  |     92 -
 .../absl/random/internal/fast_uniform_bits.h  |    271 -
 .../absl/random/internal/fastmath.h           |     57 -
 .../absl/random/internal/generate_real.h      |    144 -
 .../random/internal/iostream_state_saver.h    |    245 -
 .../absl/random/internal/mock_helpers.h       |    135 -
 .../absl/random/internal/mock_overload_set.h  |    100 -
 .../absl/random/internal/nanobenchmark.h      |    172 -
 .../absl/random/internal/nonsecure_base.h     |    161 -
 .../absl/random/internal/pcg_engine.h         |    287 -
 .../absl/random/internal/platform.h           |    171 -
 .../absl/random/internal/pool_urbg.h          |    131 -
 .../abseil-cpp/absl/random/internal/randen.h  |     96 -
 .../absl/random/internal/randen_detect.h      |     33 -
 .../absl/random/internal/randen_engine.h      |    264 -
 .../absl/random/internal/randen_hwaes.h       |     50 -
 .../absl/random/internal/randen_slow.h        |     40 -
 .../absl/random/internal/randen_traits.h      |     88 -
 .../absl/random/internal/salted_seed_seq.h    |    165 -
 .../absl/random/internal/seed_material.h      |    104 -
 .../absl/random/internal/sequence_urbg.h      |     60 -
 .../abseil-cpp/absl/random/internal/traits.h  |    149 -
 .../absl/random/internal/uniform_helper.h     |    244 -
 .../absl/random/internal/wide_multiply.h      |     96 -
 .../random/log_uniform_int_distribution.h     |    256 -
 .../absl/random/mock_distributions.h          |    266 -
 .../abseil-cpp/absl/random/mocking_bit_gen.h  |    240 -
 .../absl/random/poisson_distribution.h        |    261 -
 .../include/abseil-cpp/absl/random/random.h   |    189 -
 .../absl/random/seed_gen_exception.h          |     55 -
 .../abseil-cpp/absl/random/seed_sequences.h   |    113 -
 .../absl/random/uniform_int_distribution.h    |    275 -
 .../absl/random/uniform_real_distribution.h   |    202 -
 .../absl/random/zipf_distribution.h           |    272 -
 .../absl/status/internal/status_internal.h    |    132 -
 .../absl/status/internal/statusor_internal.h  |    435 -
 .../include/abseil-cpp/absl/status/status.h   |    943 -
 .../absl/status/status_payload_printer.h      |     52 -
 .../include/abseil-cpp/absl/status/statusor.h |    830 -
 .../include/abseil-cpp/absl/strings/ascii.h   |    244 -
 .../abseil-cpp/absl/strings/charconv.h        |    123 -
 .../include/abseil-cpp/absl/strings/charset.h |    164 -
 .../include/abseil-cpp/absl/strings/cord.h    |   1707 -
 .../abseil-cpp/absl/strings/cord_analysis.h   |     63 -
 .../abseil-cpp/absl/strings/cord_buffer.h     |    572 -
 .../absl/strings/cord_test_helpers.h          |    122 -
 .../absl/strings/cordz_test_helpers.h         |    153 -
 .../abseil-cpp/absl/strings/escaping.h        |    186 -
 .../absl/strings/has_absl_stringify.h         |     63 -
 .../absl/strings/has_ostream_operator.h       |     42 -
 .../absl/strings/internal/charconv_bigint.h   |    423 -
 .../absl/strings/internal/charconv_parse.h    |     99 -
 .../absl/strings/internal/cord_data_edge.h    |     63 -
 .../absl/strings/internal/cord_internal.h     |    912 -
 .../absl/strings/internal/cord_rep_btree.h    |    944 -
 .../internal/cord_rep_btree_navigator.h       |    267 -
 .../strings/internal/cord_rep_btree_reader.h  |    212 -
 .../absl/strings/internal/cord_rep_consume.h  |     47 -
 .../absl/strings/internal/cord_rep_crc.h      |    103 -
 .../absl/strings/internal/cord_rep_flat.h     |    195 -
 .../strings/internal/cord_rep_test_util.h     |    205 -
 .../absl/strings/internal/cordz_functions.h   |     77 -
 .../absl/strings/internal/cordz_handle.h      |     98 -
 .../absl/strings/internal/cordz_info.h        |    298 -
 .../strings/internal/cordz_sample_token.h     |     97 -
 .../absl/strings/internal/cordz_statistics.h  |     88 -
 .../strings/internal/cordz_update_scope.h     |     71 -
 .../strings/internal/cordz_update_tracker.h   |    123 -
 .../internal/damerau_levenshtein_distance.h   |     34 -
 .../absl/strings/internal/escaping.h          |     57 -
 .../strings/internal/escaping_test_common.h   |    133 -
 .../strings/internal/has_absl_stringify.h     |     44 -
 .../absl/strings/internal/memutil.h           |     40 -
 .../strings/internal/numbers_test_common.h    |    184 -
 .../absl/strings/internal/ostringstream.h     |    114 -
 .../absl/strings/internal/pow10_helper.h      |     40 -
 .../strings/internal/resize_uninitialized.h   |    119 -
 .../absl/strings/internal/stl_type_traits.h   |    248 -
 .../absl/strings/internal/str_format/arg.h    |    671 -
 .../absl/strings/internal/str_format/bind.h   |    237 -
 .../strings/internal/str_format/checker.h     |    100 -
 .../internal/str_format/constexpr_parser.h    |    357 -
 .../strings/internal/str_format/extension.h   |    456 -
 .../internal/str_format/float_conversion.h    |     37 -
 .../absl/strings/internal/str_format/output.h |     97 -
 .../absl/strings/internal/str_format/parser.h |    269 -
 .../absl/strings/internal/str_join_internal.h |    325 -
 .../strings/internal/str_split_internal.h     |    485 -
 .../absl/strings/internal/string_constant.h   |     72 -
 .../absl/strings/internal/stringify_sink.h    |     57 -
 .../abseil-cpp/absl/strings/internal/utf8.h   |     50 -
 .../include/abseil-cpp/absl/strings/match.h   |    119 -
 .../include/abseil-cpp/absl/strings/numbers.h |    456 -
 .../include/abseil-cpp/absl/strings/str_cat.h |    636 -
 .../abseil-cpp/absl/strings/str_format.h      |    887 -
 .../abseil-cpp/absl/strings/str_join.h        |    287 -
 .../abseil-cpp/absl/strings/str_replace.h     |    222 -
 .../abseil-cpp/absl/strings/str_split.h       |    565 -
 .../abseil-cpp/absl/strings/string_view.h     |    769 -
 .../include/abseil-cpp/absl/strings/strip.h   |     96 -
 .../abseil-cpp/absl/strings/substitute.h      |    767 -
 .../abseil-cpp/absl/synchronization/barrier.h |     79 -
 .../absl/synchronization/blocking_counter.h   |    101 -
 .../internal/create_thread_identity.h         |     56 -
 .../absl/synchronization/internal/futex.h     |    177 -
 .../synchronization/internal/futex_waiter.h   |     63 -
 .../synchronization/internal/graphcycles.h    |    141 -
 .../synchronization/internal/kernel_timeout.h |    178 -
 .../synchronization/internal/per_thread_sem.h |    119 -
 .../synchronization/internal/pthread_waiter.h |     60 -
 .../synchronization/internal/sem_waiter.h     |     65 -
 .../synchronization/internal/stdcpp_waiter.h  |     56 -
 .../synchronization/internal/thread_pool.h    |     96 -
 .../absl/synchronization/internal/waiter.h    |     69 -
 .../synchronization/internal/waiter_base.h    |     90 -
 .../synchronization/internal/win32_waiter.h   |     72 -
 .../abseil-cpp/absl/synchronization/mutex.h   |   1218 -
 .../absl/synchronization/notification.h       |    123 -
 .../include/abseil-cpp/absl/time/civil_time.h |    589 -
 .../include/abseil-cpp/absl/time/clock.h      |     78 -
 .../internal/cctz/include/cctz/civil_time.h   |    332 -
 .../cctz/include/cctz/civil_time_detail.h     |    632 -
 .../internal/cctz/include/cctz/time_zone.h    |    460 -
 .../cctz/include/cctz/zone_info_source.h      |    102 -
 .../time/internal/cctz/src/time_zone_fixed.h  |     52 -
 .../time/internal/cctz/src/time_zone_if.h     |     80 -
 .../time/internal/cctz/src/time_zone_impl.h   |     97 -
 .../time/internal/cctz/src/time_zone_info.h   |    128 -
 .../time/internal/cctz/src/time_zone_libc.h   |     60 -
 .../time/internal/cctz/src/time_zone_posix.h  |    132 -
 .../absl/time/internal/cctz/src/tzfile.h      |    118 -
 .../abseil-cpp/absl/time/internal/test_util.h |     33 -
 .../include/abseil-cpp/absl/time/time.h       |   1811 -
 .../include/abseil-cpp/absl/types/any.h       |    519 -
 .../abseil-cpp/absl/types/bad_any_cast.h      |     75 -
 .../absl/types/bad_optional_access.h          |     78 -
 .../absl/types/bad_variant_access.h           |     82 -
 .../include/abseil-cpp/absl/types/compare.h   |    505 -
 .../abseil-cpp/absl/types/internal/optional.h |    352 -
 .../abseil-cpp/absl/types/internal/span.h     |    139 -
 .../abseil-cpp/absl/types/internal/variant.h  |   1635 -
 .../include/abseil-cpp/absl/types/optional.h  |    779 -
 .../include/abseil-cpp/absl/types/span.h      |    753 -
 .../include/abseil-cpp/absl/types/variant.h   |    861 -
 .../absl/utility/internal/if_constexpr.h      |     70 -
 .../include/abseil-cpp/absl/utility/utility.h |    268 -
 .../abseil-cpp/ci/absl_alternate_options.h    |     30 -
 projects/llm_framework/include/archive.h      |   1214 -
 .../llm_framework/include/archive_entry.h     |    731 -
 projects/llm_framework/include/dwarf.h        |   1029 -
 .../include/elfutils/elf-knowledge.h          |    101 -
 .../include/elfutils/known-dwarf.h            |    782 -
 .../llm_framework/include/elfutils/libdw.h    |   1111 -
 .../llm_framework/include/elfutils/libdwelf.h |    147 -
 .../llm_framework/include/elfutils/libdwfl.h  |    802 -
 .../llm_framework/include/elfutils/version.h  |     38 -
 .../llm_framework/include/fst/accumulator.h   |    903 -
 projects/llm_framework/include/fst/add-on.h   |    248 -
 .../llm_framework/include/fst/arc-arena.h     |    232 -
 projects/llm_framework/include/fst/arc-map.h  |   1285 -
 projects/llm_framework/include/fst/arc.h      |    317 -
 .../llm_framework/include/fst/arcfilter.h     |     93 -
 projects/llm_framework/include/fst/arcsort.h  |    211 -
 projects/llm_framework/include/fst/bi-table.h |    480 -
 projects/llm_framework/include/fst/cache.h    |   1327 -
 projects/llm_framework/include/fst/closure.h  |    134 -
 .../llm_framework/include/fst/compact-fst.h   |   1564 -
 projects/llm_framework/include/fst/compat.h   |    130 -
 .../llm_framework/include/fst/complement.h    |    277 -
 .../include/fst/compose-filter.h              |    571 -
 projects/llm_framework/include/fst/compose.h  |   1035 -
 projects/llm_framework/include/fst/concat.h   |    220 -
 projects/llm_framework/include/fst/config.h   |      3 -
 .../llm_framework/include/fst/config.h.in     |     11 -
 projects/llm_framework/include/fst/connect.h  |    323 -
 .../llm_framework/include/fst/const-fst.h     |    485 -
 .../llm_framework/include/fst/determinize.h   |   1093 -
 .../llm_framework/include/fst/dfs-visit.h     |    202 -
 .../llm_framework/include/fst/difference.h    |    205 -
 .../llm_framework/include/fst/disambiguate.h  |    564 -
 projects/llm_framework/include/fst/edit-fst.h |    702 -
 projects/llm_framework/include/fst/encode.h   |    556 -
 .../llm_framework/include/fst/epsnormalize.h  |     61 -
 projects/llm_framework/include/fst/equal.h    |    169 -
 .../llm_framework/include/fst/equivalent.h    |    230 -
 .../llm_framework/include/fst/expanded-fst.h  |    179 -
 .../include/fst/expectation-weight.h          |    134 -
 .../fst/extensions/compress/compress-script.h |     53 -
 .../fst/extensions/compress/compress.h        |    906 -
 .../include/fst/extensions/compress/elias.h   |     97 -
 .../include/fst/extensions/compress/gzfile.h  |    127 -
 .../include/fst/extensions/compress/randmod.h |    102 -
 .../fst/extensions/far/compile-strings.h      |    260 -
 .../include/fst/extensions/far/create.h       |     46 -
 .../include/fst/extensions/far/equal.h        |     69 -
 .../include/fst/extensions/far/extract.h      |    118 -
 .../include/fst/extensions/far/far-class.h    |    258 -
 .../include/fst/extensions/far/far.h          |    481 -
 .../include/fst/extensions/far/farlib.h       |     19 -
 .../include/fst/extensions/far/farscript.h    |    269 -
 .../include/fst/extensions/far/getters.h      |     30 -
 .../include/fst/extensions/far/info.h         |    147 -
 .../include/fst/extensions/far/isomorphic.h   |     69 -
 .../fst/extensions/far/print-strings.h        |    105 -
 .../include/fst/extensions/far/script-impl.h  |     23 -
 .../include/fst/extensions/far/stlist.h       |    273 -
 .../include/fst/extensions/far/sttable.h      |    353 -
 .../linear/linear-fst-data-builder.h          |   1074 -
 .../fst/extensions/linear/linear-fst-data.h   |    526 -
 .../fst/extensions/linear/linear-fst.h        |   1173 -
 .../fst/extensions/linear/linearscript.h      |    391 -
 .../fst/extensions/linear/loglinear-apply.h   |     77 -
 .../include/fst/extensions/linear/trie.h      |    444 -
 .../include/fst/extensions/mpdt/compose.h     |    267 -
 .../include/fst/extensions/mpdt/expand.h      |    335 -
 .../include/fst/extensions/mpdt/info.h        |    190 -
 .../include/fst/extensions/mpdt/mpdt.h        |    357 -
 .../include/fst/extensions/mpdt/mpdtlib.h     |     18 -
 .../include/fst/extensions/mpdt/mpdtscript.h  |    156 -
 .../fst/extensions/mpdt/read_write_utils.h    |     86 -
 .../include/fst/extensions/mpdt/reverse.h     |     54 -
 .../fst/extensions/ngram/bitmap-index.h       |    168 -
 .../include/fst/extensions/ngram/ngram-fst.h  |   1027 -
 .../include/fst/extensions/ngram/nthbit.h     |     49 -
 .../include/fst/extensions/pdt/collection.h   |    107 -
 .../include/fst/extensions/pdt/compose.h      |    493 -
 .../include/fst/extensions/pdt/expand.h       |    933 -
 .../include/fst/extensions/pdt/getters.h      |     22 -
 .../include/fst/extensions/pdt/info.h         |    152 -
 .../include/fst/extensions/pdt/paren.h        |    440 -
 .../include/fst/extensions/pdt/pdt.h          |    165 -
 .../include/fst/extensions/pdt/pdtlib.h       |     19 -
 .../include/fst/extensions/pdt/pdtscript.h    |    244 -
 .../include/fst/extensions/pdt/replace.h      |    827 -
 .../include/fst/extensions/pdt/reverse.h      |     38 -
 .../fst/extensions/pdt/shortest-path.h        |    715 -
 .../include/fst/extensions/special/phi-fst.h  |    183 -
 .../include/fst/extensions/special/rho-fst.h  |    172 -
 .../fst/extensions/special/sigma-fst.h        |    176 -
 .../llm_framework/include/fst/factor-weight.h |    496 -
 .../llm_framework/include/fst/filter-state.h  |    199 -
 projects/llm_framework/include/fst/flags.h    |    219 -
 .../llm_framework/include/fst/float-weight.h  |    858 -
 projects/llm_framework/include/fst/fst-decl.h |    254 -
 projects/llm_framework/include/fst/fst.h      |   1007 -
 projects/llm_framework/include/fst/fstlib.h   |    130 -
 .../include/fst/generic-register.h            |    126 -
 projects/llm_framework/include/fst/heap.h     |    168 -
 projects/llm_framework/include/fst/icu.h      |    129 -
 .../llm_framework/include/fst/intersect.h     |    181 -
 .../llm_framework/include/fst/interval-set.h  |    398 -
 projects/llm_framework/include/fst/invert.h   |    139 -
 .../llm_framework/include/fst/isomorphic.h    |    183 -
 .../include/fst/label-reachable.h             |    511 -
 .../include/fst/lexicographic-weight.h        |    173 -
 projects/llm_framework/include/fst/lock.h     |     62 -
 projects/llm_framework/include/fst/log.h      |     78 -
 .../include/fst/lookahead-filter.h            |    623 -
 .../include/fst/lookahead-matcher.h           |    841 -
 projects/llm_framework/include/fst/map.h      |    110 -
 .../llm_framework/include/fst/mapped-file.h   |     81 -
 .../llm_framework/include/fst/matcher-fst.h   |    347 -
 projects/llm_framework/include/fst/matcher.h  |   1575 -
 projects/llm_framework/include/fst/memory.h   |    443 -
 projects/llm_framework/include/fst/minimize.h |    568 -
 .../llm_framework/include/fst/mutable-fst.h   |    398 -
 .../llm_framework/include/fst/pair-weight.h   |    155 -
 .../llm_framework/include/fst/partition.h     |    305 -
 .../llm_framework/include/fst/power-weight.h  |    168 -
 .../include/fst/product-weight.h              |    107 -
 projects/llm_framework/include/fst/project.h  |    159 -
 .../llm_framework/include/fst/properties.h    |    468 -
 projects/llm_framework/include/fst/prune.h    |    341 -
 projects/llm_framework/include/fst/push.h     |    155 -
 projects/llm_framework/include/fst/queue.h    |    948 -
 .../include/fst/randequivalent.h              |    114 -
 projects/llm_framework/include/fst/randgen.h  |    756 -
 projects/llm_framework/include/fst/rational.h |    307 -
 projects/llm_framework/include/fst/register.h |    115 -
 projects/llm_framework/include/fst/relabel.h  |    472 -
 .../llm_framework/include/fst/replace-util.h  |    629 -
 projects/llm_framework/include/fst/replace.h  |   1492 -
 projects/llm_framework/include/fst/reverse.h  |    116 -
 projects/llm_framework/include/fst/reweight.h |    127 -
 .../llm_framework/include/fst/rmepsilon.h     |    548 -
 .../include/fst/rmfinalepsilon.h              |     80 -
 .../include/fst/script/arc-class.h            |     40 -
 .../include/fst/script/arciterator-class.h    |    212 -
 .../include/fst/script/arcsort.h              |     44 -
 .../include/fst/script/arg-packs.h            |     37 -
 .../include/fst/script/closure.h              |     28 -
 .../include/fst/script/compile-impl.h         |    217 -
 .../include/fst/script/compile.h              |     98 -
 .../include/fst/script/compose.h              |     34 -
 .../llm_framework/include/fst/script/concat.h |     40 -
 .../include/fst/script/connect.h              |     23 -
 .../include/fst/script/convert.h              |     35 -
 .../llm_framework/include/fst/script/decode.h |     49 -
 .../include/fst/script/determinize.h          |     59 -
 .../include/fst/script/difference.h           |     35 -
 .../include/fst/script/disambiguate.h         |     54 -
 .../include/fst/script/draw-impl.h            |    227 -
 .../llm_framework/include/fst/script/draw.h   |     85 -
 .../llm_framework/include/fst/script/encode.h |     51 -
 .../include/fst/script/encodemapper-class.h   |    169 -
 .../include/fst/script/epsnormalize.h         |     31 -
 .../llm_framework/include/fst/script/equal.h  |     32 -
 .../include/fst/script/equivalent.h           |     34 -
 .../include/fst/script/fst-class.h            |    530 -
 .../include/fst/script/fstscript-decl.h       |     32 -
 .../include/fst/script/fstscript.h            |    155 -
 .../include/fst/script/getters.h              |     76 -
 .../include/fst/script/info-impl.h            |    314 -
 .../llm_framework/include/fst/script/info.h   |     50 -
 .../include/fst/script/intersect.h            |     35 -
 .../llm_framework/include/fst/script/invert.h |     23 -
 .../include/fst/script/isomorphic.h           |     34 -
 .../llm_framework/include/fst/script/map.h    |    158 -
 .../include/fst/script/minimize.h             |     33 -
 .../include/fst/script/print-impl.h           |    132 -
 .../llm_framework/include/fst/script/print.h  |     79 -
 .../include/fst/script/project.h              |     28 -
 .../llm_framework/include/fst/script/prune.h  |     51 -
 .../llm_framework/include/fst/script/push.h   |     53 -
 .../include/fst/script/randequivalent.h       |     67 -
 .../include/fst/script/randgen.h              |     63 -
 .../include/fst/script/register.h             |     99 -
 .../include/fst/script/relabel.h              |     64 -
 .../include/fst/script/replace.h              |     72 -
 .../include/fst/script/reverse.h              |     30 -
 .../include/fst/script/reweight.h             |     37 -
 .../include/fst/script/rmepsilon.h            |    109 -
 .../include/fst/script/script-impl.h          |    211 -
 .../include/fst/script/shortest-distance.h    |    214 -
 .../include/fst/script/shortest-path.h        |    116 -
 .../include/fst/script/stateiterator-class.h  |     85 -
 .../include/fst/script/synchronize.h          |     29 -
 .../include/fst/script/text-io.h              |     28 -
 .../include/fst/script/topsort.h              |     26 -
 .../llm_framework/include/fst/script/union.h  |     29 -
 .../llm_framework/include/fst/script/verify.h |     27 -
 .../include/fst/script/weight-class.h         |    235 -
 .../llm_framework/include/fst/set-weight.h    |    618 -
 .../include/fst/shortest-distance.h           |    351 -
 .../llm_framework/include/fst/shortest-path.h |    549 -
 .../include/fst/signed-log-weight.h           |    440 -
 .../include/fst/sparse-power-weight.h         |    209 -
 .../include/fst/sparse-tuple-weight.h         |    422 -
 .../llm_framework/include/fst/state-map.h     |    613 -
 .../include/fst/state-reachable.h             |    224 -
 .../llm_framework/include/fst/state-table.h   |    494 -
 .../llm_framework/include/fst/statesort.h     |     74 -
 .../llm_framework/include/fst/string-weight.h |    807 -
 projects/llm_framework/include/fst/string.h   |    286 -
 .../include/fst/symbol-table-ops.h            |     76 -
 .../llm_framework/include/fst/symbol-table.h  |    445 -
 .../llm_framework/include/fst/synchronize.h   |    414 -
 .../include/fst/test-properties.h             |    246 -
 .../include/fst/test/algo_test.h              |   1414 -
 .../llm_framework/include/fst/test/fst_test.h |    318 -
 .../llm_framework/include/fst/test/rand-fst.h |     90 -
 .../include/fst/test/weight-tester.h          |    207 -
 projects/llm_framework/include/fst/topsort.h  |     95 -
 .../llm_framework/include/fst/tuple-weight.h  |    163 -
 projects/llm_framework/include/fst/types.h    |     41 -
 .../llm_framework/include/fst/union-find.h    |     84 -
 .../llm_framework/include/fst/union-weight.h  |    505 -
 projects/llm_framework/include/fst/union.h    |    157 -
 projects/llm_framework/include/fst/util.h     |    400 -
 .../llm_framework/include/fst/vector-fst.h    |    796 -
 projects/llm_framework/include/fst/verify.h   |    100 -
 projects/llm_framework/include/fst/visit.h    |    321 -
 projects/llm_framework/include/fst/weight.h   |    389 -
 projects/llm_framework/include/gelf.h         |    342 -
 .../llm_framework/include/gflags/defines.h    |     48 -
 .../llm_framework/include/gflags/gflags.h     |    626 -
 .../include/gflags/gflags_completions.h       |    121 -
 .../include/gflags/gflags_declare.h           |    156 -
 .../llm_framework/include/glog/log_severity.h |     92 -
 projects/llm_framework/include/glog/logging.h |   1662 -
 .../llm_framework/include/glog/raw_logging.h  |    180 -
 .../llm_framework/include/glog/stl_logging.h  |    220 -
 .../llm_framework/include/glog/vlog_is_on.h   |    129 -
 projects/llm_framework/include/libdw.h        |   1111 -
 projects/llm_framework/include/libelf.h       |    522 -
 .../onnxruntime/core/common/basic_types.h     |     19 -
 .../onnxruntime/core/common/code_location.h   |     58 -
 .../include/onnxruntime/core/common/common.h  |    287 -
 .../core/common/const_pointer_container.h     |     85 -
 .../onnxruntime/core/common/denormal.h        |     12 -
 .../core/common/eigen_common_wrapper.h        |     62 -
 .../onnxruntime/core/common/exceptions.h      |     71 -
 .../core/common/gpu_profiler_common.h         |    472 -
 .../include/onnxruntime/core/common/gsl.h     |      6 -
 .../onnxruntime/core/common/hash_combine.h    |     21 -
 .../core/common/inlined_containers.h          |    175 -
 .../core/common/inlined_containers_fwd.h      |    147 -
 .../onnxruntime/core/common/logging/capture.h |    115 -
 .../onnxruntime/core/common/logging/isink.h   |     41 -
 .../onnxruntime/core/common/logging/logging.h |    337 -
 .../onnxruntime/core/common/logging/macros.h  |    278 -
 .../core/common/logging/severity.h            |     22 -
 .../onnxruntime/core/common/make_string.h     |    126 -
 .../include/onnxruntime/core/common/narrow.h  |     77 -
 .../onnxruntime/core/common/optional.h        |     23 -
 .../onnxruntime/core/common/parse_string.h    |     85 -
 .../onnxruntime/core/common/profiler_common.h |     93 -
 .../onnxruntime/core/common/span_utils.h      |     88 -
 .../onnxruntime/core/common/spin_pause.h      |     28 -
 .../include/onnxruntime/core/common/status.h  |    195 -
 .../onnxruntime/core/common/string_helper.h   |     11 -
 .../onnxruntime/core/framework/alloc_kind.h   |     36 -
 .../onnxruntime/core/framework/allocator.h    |    194 -
 .../core/framework/buffer_deleter.h           |     36 -
 .../core/framework/customregistry.h           |     60 -
 .../onnxruntime/core/framework/data_types.h   |   1062 -
 .../core/framework/data_types_internal.h      |    569 -
 .../onnxruntime/core/framework/endian.h       |     27 -
 .../core/framework/execution_provider.h       |    332 -
 .../onnxruntime/core/framework/float16.h      |    159 -
 .../core/framework/framework_common.h         |     22 -
 .../onnxruntime/core/framework/func_api.h     |     27 -
 .../core/framework/kernel_def_builder.h       |    353 -
 .../core/framework/kernel_registry.h          |     91 -
 .../onnxruntime/core/framework/op_kernel.h    |    387 -
 .../core/framework/op_kernel_context.h        |    237 -
 .../core/framework/op_kernel_info.h           |     63 -
 .../core/framework/op_node_proto_helper.h     |    167 -
 .../onnxruntime/core/framework/ort_value.h    |    123 -
 .../onnxruntime/core/framework/ortdevice.h    |     74 -
 .../core/framework/ortmemoryinfo.h            |     87 -
 .../core/framework/provider_options.h         |     18 -
 .../core/framework/provider_options_utils.h   |    164 -
 .../core/framework/provider_shutdown.h        |      8 -
 .../onnxruntime/core/framework/run_options.h  |     49 -
 .../core/framework/sparse_tensor.h            |    534 -
 .../core/framework/stream_handles.h           |    170 -
 .../onnxruntime/core/framework/tensor.h       |    323 -
 .../onnxruntime/core/framework/tensor_shape.h |    203 -
 .../framework/to_tensor_proto_element_type.h  |     79 -
 .../onnxruntime/core/graph/basic_types.h      |     50 -
 .../onnxruntime/core/graph/constants.h        |     49 -
 .../include/onnxruntime/core/graph/function.h |     31 -
 .../include/onnxruntime/core/graph/graph.h    |   1710 -
 .../onnxruntime/core/graph/graph_nodes.h      |    182 -
 .../onnxruntime/core/graph/graph_viewer.h     |    222 -
 .../core/graph/indexed_sub_graph.h            |     84 -
 .../include/onnxruntime/core/graph/node_arg.h |    134 -
 .../onnxruntime/core/graph/schema_registry.h  |    163 -
 .../core/optimizer/graph_transformer.h        |     73 -
 .../core/optimizer/graph_transformer_config.h |     73 -
 .../core/optimizer/graph_transformer_level.h  |     19 -
 .../core/optimizer/graph_transformer_utils.h  |     86 -
 .../onnxruntime/core/optimizer/rewrite_rule.h |     89 -
 .../optimizer/rule_based_graph_transformer.h  |     82 -
 .../core/providers/cpu/cpu_provider_factory.h |     19 -
 .../onnxruntime/core/session/environment.h    |     96 -
 .../experimental_onnxruntime_cxx_api.h        |     69 -
 .../experimental_onnxruntime_cxx_inline.h     |    108 -
 .../core/session/onnxruntime_c_api.h          |   3987 -
 .../core/session/onnxruntime_cxx_api.h        |   1876 -
 .../core/session/onnxruntime_cxx_inline.h     |   1874 -
 .../onnxruntime_run_options_config_keys.h     |     32 -
 .../onnxruntime_session_options_config_keys.h |    186 -
 .../onnxruntime/core/session/snippets.dox     |      8 -
 .../include/opencv4/opencv2/core.hpp          |   3344 -
 .../include/opencv4/opencv2/core/affine.hpp   |    678 -
 .../include/opencv4/opencv2/core/async.hpp    |    105 -
 .../include/opencv4/opencv2/core/base.hpp     |    664 -
 .../opencv4/opencv2/core/bindings_utils.hpp   |    248 -
 .../opencv4/opencv2/core/bufferpool.hpp       |     40 -
 .../include/opencv4/opencv2/core/check.hpp    |    160 -
 .../include/opencv4/opencv2/core/core.hpp     |     48 -
 .../include/opencv4/opencv2/core/core_c.h     |   3125 -
 .../include/opencv4/opencv2/core/cuda.hpp     |   1270 -
 .../include/opencv4/opencv2/core/cuda.inl.hpp |    723 -
 .../opencv4/opencv2/core/cuda/block.hpp       |    211 -
 .../opencv2/core/cuda/border_interpolate.hpp  |    722 -
 .../opencv4/opencv2/core/cuda/color.hpp       |    309 -
 .../opencv4/opencv2/core/cuda/common.hpp      |    123 -
 .../opencv2/core/cuda/datamov_utils.hpp       |    113 -
 .../opencv2/core/cuda/detail/color_detail.hpp |   2018 -
 .../opencv2/core/cuda/detail/reduce.hpp       |    365 -
 .../core/cuda/detail/reduce_key_val.hpp       |    502 -
 .../core/cuda/detail/transform_detail.hpp     |    392 -
 .../core/cuda/detail/type_traits_detail.hpp   |    191 -
 .../core/cuda/detail/vec_distance_detail.hpp  |    121 -
 .../opencv2/core/cuda/dynamic_smem.hpp        |     88 -
 .../opencv4/opencv2/core/cuda/emulation.hpp   |    269 -
 .../opencv4/opencv2/core/cuda/filters.hpp     |    293 -
 .../opencv4/opencv2/core/cuda/funcattrib.hpp  |     79 -
 .../opencv4/opencv2/core/cuda/functional.hpp  |    805 -
 .../opencv4/opencv2/core/cuda/limits.hpp      |    128 -
 .../opencv4/opencv2/core/cuda/reduce.hpp      |    209 -
 .../opencv2/core/cuda/saturate_cast.hpp       |    292 -
 .../opencv4/opencv2/core/cuda/scan.hpp        |    258 -
 .../opencv2/core/cuda/simd_functions.hpp      |    869 -
 .../opencv4/opencv2/core/cuda/transform.hpp   |     75 -
 .../opencv4/opencv2/core/cuda/type_traits.hpp |     90 -
 .../opencv4/opencv2/core/cuda/utility.hpp     |    230 -
 .../opencv2/core/cuda/vec_distance.hpp        |    232 -
 .../opencv4/opencv2/core/cuda/vec_math.hpp    |    923 -
 .../opencv4/opencv2/core/cuda/vec_traits.hpp  |    288 -
 .../opencv4/opencv2/core/cuda/warp.hpp        |    139 -
 .../opencv4/opencv2/core/cuda/warp_reduce.hpp |     76 -
 .../opencv2/core/cuda/warp_shuffle.hpp        |    162 -
 .../opencv2/core/cuda_stream_accessor.hpp     |     86 -
 .../opencv4/opencv2/core/cuda_types.hpp       |    144 -
 .../opencv4/opencv2/core/cv_cpu_dispatch.h    |    368 -
 .../opencv4/opencv2/core/cv_cpu_helper.h      |    508 -
 .../include/opencv4/opencv2/core/cvdef.h      |    967 -
 .../include/opencv4/opencv2/core/cvstd.hpp    |    190 -
 .../opencv4/opencv2/core/cvstd.inl.hpp        |    197 -
 .../opencv4/opencv2/core/cvstd_wrapper.hpp    |    154 -
 .../opencv2/core/detail/async_promise.hpp     |     71 -
 .../core/detail/dispatch_helper.impl.hpp      |     49 -
 .../opencv2/core/detail/exception_ptr.hpp     |     27 -
 .../include/opencv4/opencv2/core/directx.hpp  |    184 -
 .../opencv4/opencv2/core/dualquaternion.hpp   |    979 -
 .../opencv2/core/dualquaternion.inl.hpp       |    487 -
 .../include/opencv4/opencv2/core/eigen.hpp    |    402 -
 .../opencv4/opencv2/core/fast_math.hpp        |    411 -
 .../include/opencv4/opencv2/core/hal/hal.hpp  |    256 -
 .../opencv4/opencv2/core/hal/interface.h      |    190 -
 .../opencv4/opencv2/core/hal/intrin.hpp       |    706 -
 .../opencv4/opencv2/core/hal/intrin_avx.hpp   |   3177 -
 .../opencv2/core/hal/intrin_avx512.hpp        |   3090 -
 .../opencv4/opencv2/core/hal/intrin_cpp.hpp   |   3320 -
 .../opencv2/core/hal/intrin_forward.hpp       |    191 -
 .../opencv4/opencv2/core/hal/intrin_msa.hpp   |   1887 -
 .../opencv4/opencv2/core/hal/intrin_neon.hpp  |   2613 -
 .../opencv4/opencv2/core/hal/intrin_rvv.hpp   |   2887 -
 .../opencv2/core/hal/intrin_rvv071.hpp        |   2545 -
 .../opencv4/opencv2/core/hal/intrin_sse.hpp   |   3467 -
 .../opencv2/core/hal/intrin_sse_em.hpp        |    180 -
 .../opencv4/opencv2/core/hal/intrin_vsx.hpp   |   1608 -
 .../opencv4/opencv2/core/hal/intrin_wasm.hpp  |   2782 -
 .../opencv4/opencv2/core/hal/msa_macros.h     |   1558 -
 .../opencv2/core/hal/simd_utils.impl.hpp      |    146 -
 .../include/opencv4/opencv2/core/mat.hpp      |   3766 -
 .../include/opencv4/opencv2/core/mat.inl.hpp  |   3422 -
 .../include/opencv4/opencv2/core/matx.hpp     |   1508 -
 .../opencv4/opencv2/core/neon_utils.hpp       |    128 -
 .../include/opencv4/opencv2/core/ocl.hpp      |    917 -
 .../opencv4/opencv2/core/ocl_genbase.hpp      |     69 -
 .../opencv4/opencv2/core/opencl/ocl_defs.hpp  |     82 -
 .../opencv2/core/opencl/opencl_info.hpp       |    212 -
 .../opencv2/core/opencl/opencl_svm.hpp        |     81 -
 .../runtime/autogenerated/opencl_clblas.hpp   |    602 -
 .../runtime/autogenerated/opencl_clfft.hpp    |    146 -
 .../runtime/autogenerated/opencl_core.hpp     |    371 -
 .../autogenerated/opencl_core_wrappers.hpp    |    272 -
 .../runtime/autogenerated/opencl_gl.hpp       |     62 -
 .../autogenerated/opencl_gl_wrappers.hpp      |     42 -
 .../core/opencl/runtime/opencl_clblas.hpp     |     53 -
 .../core/opencl/runtime/opencl_clfft.hpp      |     53 -
 .../core/opencl/runtime/opencl_core.hpp       |     84 -
 .../opencl/runtime/opencl_core_wrappers.hpp   |     47 -
 .../opencv2/core/opencl/runtime/opencl_gl.hpp |     53 -
 .../opencl/runtime/opencl_gl_wrappers.hpp     |     47 -
 .../core/opencl/runtime/opencl_svm_20.hpp     |     48 -
 .../opencl/runtime/opencl_svm_definitions.hpp |     42 -
 .../runtime/opencl_svm_hsa_extension.hpp      |    166 -
 .../include/opencv4/opencv2/core/opengl.hpp   |    725 -
 .../opencv4/opencv2/core/operations.hpp       |    594 -
 .../include/opencv4/opencv2/core/optim.hpp    |    302 -
 .../include/opencv4/opencv2/core/ovx.hpp      |     28 -
 .../parallel/backend/parallel_for.openmp.hpp  |     72 -
 .../parallel/backend/parallel_for.tbb.hpp     |    153 -
 .../core/parallel/parallel_backend.hpp        |     90 -
 .../opencv4/opencv2/core/persistence.hpp      |   1350 -
 .../opencv4/opencv2/core/quaternion.hpp       |   1696 -
 .../opencv4/opencv2/core/quaternion.inl.hpp   |   1063 -
 .../include/opencv4/opencv2/core/saturate.hpp |    179 -
 .../opencv4/opencv2/core/simd_intrinsics.hpp  |     87 -
 .../opencv4/opencv2/core/softfloat.hpp        |    514 -
 .../opencv4/opencv2/core/sse_utils.hpp        |    652 -
 .../include/opencv4/opencv2/core/traits.hpp   |    417 -
 .../include/opencv4/opencv2/core/types.hpp    |   2439 -
 .../include/opencv4/opencv2/core/types_c.h    |   2126 -
 .../include/opencv4/opencv2/core/utility.hpp  |   1229 -
 .../opencv2/core/utils/allocator_stats.hpp    |     29 -
 .../core/utils/allocator_stats.impl.hpp       |    158 -
 .../opencv4/opencv2/core/utils/filesystem.hpp |     82 -
 .../opencv2/core/utils/instrumentation.hpp    |    125 -
 .../opencv2/core/utils/logger.defines.hpp     |     42 -
 .../opencv4/opencv2/core/utils/logger.hpp     |    218 -
 .../opencv4/opencv2/core/utils/logtag.hpp     |     28 -
 .../opencv4/opencv2/core/utils/tls.hpp        |    235 -
 .../opencv4/opencv2/core/utils/trace.hpp      |    252 -
 .../include/opencv4/opencv2/core/va_intel.hpp |     75 -
 .../include/opencv4/opencv2/core/version.hpp  |     26 -
 .../opencv4/opencv2/core/vsx_utils.hpp        |   1042 -
 .../include/opencv4/opencv2/cvconfig.h        |    149 -
 .../include/opencv4/opencv2/highgui.hpp       |    868 -
 .../opencv4/opencv2/highgui/highgui.hpp       |     48 -
 .../opencv4/opencv2/highgui/highgui_c.h       |    251 -
 .../include/opencv4/opencv2/imgcodecs.hpp     |    330 -
 .../opencv4/opencv2/imgcodecs/imgcodecs.hpp   |     48 -
 .../opencv4/opencv2/imgcodecs/imgcodecs_c.h   |      1 -
 .../include/opencv4/opencv2/imgcodecs/ios.h   |     59 -
 .../opencv2/imgcodecs/legacy/constants_c.h    |     54 -
 .../opencv4/opencv2/imgcodecs/macosx.h        |     20 -
 .../include/opencv4/opencv2/imgproc.hpp       |   4994 -
 .../opencv4/opencv2/imgproc/bindings.hpp      |     34 -
 .../opencv2/imgproc/detail/gcgraph.hpp        |    395 -
 .../opencv4/opencv2/imgproc/hal/hal.hpp       |    246 -
 .../opencv4/opencv2/imgproc/hal/interface.h   |     46 -
 .../opencv4/opencv2/imgproc/imgproc.hpp       |     48 -
 .../opencv4/opencv2/imgproc/imgproc_c.h       |   1177 -
 .../opencv4/opencv2/imgproc/segmentation.hpp  |    141 -
 .../include/opencv4/opencv2/imgproc/types_c.h |    659 -
 .../include/opencv4/opencv2/opencv.hpp        |     95 -
 .../opencv4/opencv2/opencv_modules.hpp        |     19 -
 .../include/opencv4/opencv2/videoio.hpp       |   1129 -
 .../include/opencv4/opencv2/videoio/cap_ios.h |    150 -
 .../opencv2/videoio/legacy/constants_c.h      |    434 -
 .../opencv4/opencv2/videoio/registry.hpp      |     72 -
 .../opencv4/opencv2/videoio/videoio.hpp       |     48 -
 .../opencv4/opencv2/videoio/videoio_c.h       |    153 -
 .../protobuf-lite/google/protobuf/any.h       |    150 -
 .../protobuf-lite/google/protobuf/arena.h     |    694 -
 .../google/protobuf/arena_impl.h              |    491 -
 .../google/protobuf/arenastring.h             |    375 -
 .../google/protobuf/descriptor.h              |   2324 -
 .../google/protobuf/extension_set.h           |   1591 -
 .../google/protobuf/extension_set_inl.h       |    276 -
 .../protobuf/generated_enum_reflection.h      |     98 -
 .../google/protobuf/generated_enum_util.h     |     83 -
 .../protobuf/generated_message_table_driven.h |    336 -
 .../generated_message_table_driven_lite.h     |    802 -
 .../google/protobuf/generated_message_util.h  |    270 -
 .../protobuf-lite/google/protobuf/has_bits.h  |    116 -
 .../google/protobuf/implicit_weak_message.h   |    190 -
 .../google/protobuf/io/coded_stream.h         |   1714 -
 .../google/protobuf/io/io_win32.h             |    139 -
 .../google/protobuf/io/zero_copy_stream.h     |    253 -
 .../protobuf/io/zero_copy_stream_impl.h       |    327 -
 .../protobuf/io/zero_copy_stream_impl_lite.h  |    408 -
 .../protobuf-lite/google/protobuf/map.h       |   1362 -
 .../google/protobuf/map_entry_lite.h          |    653 -
 .../google/protobuf/map_field_lite.h          |    183 -
 .../google/protobuf/map_type_handler.h        |    688 -
 .../google/protobuf/message_lite.h            |    621 -
 .../google/protobuf/metadata_lite.h           |    248 -
 .../google/protobuf/parse_context.h           |    869 -
 .../protobuf-lite/google/protobuf/port.h      |     43 -
 .../google/protobuf/port_def.inc              |    606 -
 .../google/protobuf/port_undef.inc            |    116 -
 .../google/protobuf/repeated_field.h          |   2847 -
 .../google/protobuf/stubs/bytestream.h        |    351 -
 .../google/protobuf/stubs/callback.h          |    583 -
 .../google/protobuf/stubs/casts.h             |    139 -
 .../google/protobuf/stubs/common.h            |    202 -
 .../google/protobuf/stubs/hash.h              |    114 -
 .../google/protobuf/stubs/int128.h            |    387 -
 .../google/protobuf/stubs/logging.h           |    241 -
 .../google/protobuf/stubs/macros.h            |    120 -
 .../google/protobuf/stubs/map_util.h          |    769 -
 .../google/protobuf/stubs/mutex.h             |    186 -
 .../google/protobuf/stubs/once.h              |     55 -
 .../google/protobuf/stubs/platform_macros.h   |    134 -
 .../google/protobuf/stubs/port.h              |    405 -
 .../google/protobuf/stubs/status.h            |    125 -
 .../google/protobuf/stubs/statusor.h          |    272 -
 .../google/protobuf/stubs/stl_util.h          |     71 -
 .../google/protobuf/stubs/stringpiece.h       |    489 -
 .../google/protobuf/stubs/stringprintf.h      |     85 -
 .../google/protobuf/stubs/strutil.h           |    953 -
 .../google/protobuf/stubs/time.h              |     80 -
 .../google/protobuf/unknown_field_set.h       |    411 -
 .../google/protobuf/wire_format_lite.h        |   1866 -
 .../llm_framework/include/re2/re2/bitmap256.h |     86 -
 .../include/re2/re2/filtered_re2.h            |    115 -
 .../llm_framework/include/re2/re2/pod_array.h |     55 -
 .../llm_framework/include/re2/re2/prefilter.h |    167 -
 .../include/re2/re2/prefilter_tree.h          |    152 -
 projects/llm_framework/include/re2/re2/prog.h |    466 -
 projects/llm_framework/include/re2/re2/re2.h  |   1078 -
 .../llm_framework/include/re2/re2/regexp.h    |    664 -
 projects/llm_framework/include/re2/re2/set.h  |     86 -
 .../include/re2/re2/sparse_array.h            |    392 -
 .../include/re2/re2/sparse_set.h              |    264 -
 .../include/re2/re2/stringpiece.h             |     18 -
 .../re2/re2/testing/exhaustive_tester.h       |    104 -
 .../re2/re2/testing/regexp_generator.h        |     76 -
 .../re2/re2/testing/string_generator.h        |     75 -
 .../include/re2/re2/testing/tester.h          |    121 -
 .../include/re2/re2/unicode_casefold.h        |     77 -
 .../include/re2/re2/unicode_groups.h          |     66 -
 .../include/re2/re2/walker-inl.h              |    248 -
 .../llm_framework/include/re2/util/logging.h  |    109 -
 .../include/re2/util/malloc_counter.h         |     19 -
 .../llm_framework/include/re2/util/pcre.h     |    671 -
 .../llm_framework/include/re2/util/strutil.h  |     16 -
 projects/llm_framework/include/re2/util/utf.h |     44 -
 .../include/sentencepiece/bpe_model.h         |     52 -
 .../include/sentencepiece/bpe_model_trainer.h |    130 -
 .../include/sentencepiece/builder.h           |    148 -
 .../builtin_pb/sentencepiece.pb.h             |   1041 -
 .../builtin_pb/sentencepiece_model.pb.h       |   5030 -
 .../include/sentencepiece/char_model.h        |     34 -
 .../sentencepiece/char_model_trainer.h        |     37 -
 .../include/sentencepiece/common.h            |    177 -
 .../include/sentencepiece/filesystem.h        |     59 -
 .../include/sentencepiece/freelist.h          |     90 -
 .../include/sentencepiece/init.h              |     62 -
 .../include/sentencepiece/model_factory.h     |     31 -
 .../include/sentencepiece/model_interface.h   |    249 -
 .../sentencepiece/normalization_rule.h        | 137363 ---------------
 .../include/sentencepiece/normalizer.h        |    159 -
 .../sentencepiece/pretokenizer_for_training.h |     63 -
 .../sentencepiece/sentencepiece_processor.h   |    752 -
 .../sentencepiece/sentencepiece_trainer.h     |    206 -
 .../include/sentencepiece/spec_parser.h       |    284 -
 .../include/sentencepiece/testharness.h       |    227 -
 .../include/sentencepiece/trainer_factory.h   |     33 -
 .../include/sentencepiece/trainer_interface.h |    183 -
 .../include/sentencepiece/unicode_script.h    |    165 -
 .../sentencepiece/unicode_script_map.h        |   1955 -
 .../include/sentencepiece/unigram_model.h     |    217 -
 .../sentencepiece/unigram_model_trainer.h     |    115 -
 .../include/sentencepiece/util.h              |    442 -
 .../include/sentencepiece/word_model.h        |     34 -
 .../sentencepiece/word_model_trainer.h        |     41 -
 .../include/sherpa/sherpa-ncnn/allocator.h    |    444 -
 .../include/sherpa/sherpa-ncnn/benchmark.h    |     39 -
 .../include/sherpa/sherpa-ncnn/blob.h         |     44 -
 .../include/sherpa/sherpa-ncnn/c_api.h        |    370 -
 .../include/sherpa/sherpa-ncnn/command.h      |    134 -
 .../include/sherpa/sherpa-ncnn/cpu.h          |    178 -
 .../include/sherpa/sherpa-ncnn/csrc/alsa.h    |     76 -
 .../sherpa/sherpa-ncnn/csrc/context-graph.h   |     91 -
 .../sherpa-ncnn/csrc/conv-emformer-model.h    |    107 -
 .../include/sherpa/sherpa-ncnn/csrc/decoder.h |    102 -
 .../include/sherpa/sherpa-ncnn/csrc/display.h |    108 -
 .../sherpa/sherpa-ncnn/csrc/endpoint.h        |     84 -
 .../sherpa/sherpa-ncnn/csrc/features.h        |     82 -
 .../sherpa-ncnn/csrc/greedy-search-decoder.h  |     47 -
 .../sherpa/sherpa-ncnn/csrc/hypothesis.h      |    133 -
 .../sherpa/sherpa-ncnn/csrc/lstm-model.h      |    132 -
 .../include/sherpa/sherpa-ncnn/csrc/math.h    |    120 -
 .../sherpa/sherpa-ncnn/csrc/meta-data.h       |     72 -
 .../sherpa/sherpa-ncnn/csrc/microphone.h      |     34 -
 .../include/sherpa/sherpa-ncnn/csrc/model.h   |    147 -
 .../csrc/modified-beam-search-decoder.h       |     55 -
 .../sherpa-ncnn/csrc/poolingmodulenoproj.h    |     43 -
 .../sherpa/sherpa-ncnn/csrc/recognizer.h      |    126 -
 .../sherpa/sherpa-ncnn/csrc/resample.h        |    144 -
 .../sherpa/sherpa-ncnn/csrc/simpleupsample.h  |     52 -
 .../include/sherpa/sherpa-ncnn/csrc/stack.h   |     48 -
 .../include/sherpa/sherpa-ncnn/csrc/stream.h  |    108 -
 .../sherpa/sherpa-ncnn/csrc/symbol-table.h    |     74 -
 .../sherpa/sherpa-ncnn/csrc/tensorasstrided.h |     48 -
 .../sherpa/sherpa-ncnn/csrc/wave-reader.h     |     45 -
 .../sherpa/sherpa-ncnn/csrc/zipformer-model.h |    111 -
 .../include/sherpa/sherpa-ncnn/datareader.h   |    122 -
 .../include/sherpa/sherpa-ncnn/gpu.h          |    480 -
 .../include/sherpa/sherpa-ncnn/layer.h        |    230 -
 .../include/sherpa/sherpa-ncnn/layer/absval.h |     32 -
 .../include/sherpa/sherpa-ncnn/layer/argmax.h |     38 -
 .../sherpa/sherpa-ncnn/layer/arm/absval_arm.h |     32 -
 .../sherpa-ncnn/layer/arm/arm_activation.h    |    221 -
 .../sherpa-ncnn/layer/arm/arm_usability.h     |    770 -
 .../sherpa-ncnn/layer/arm/batchnorm_arm.h     |     41 -
 .../sherpa/sherpa-ncnn/layer/arm/bias_arm.h   |     30 -
 .../sherpa-ncnn/layer/arm/binaryop_arm.h      |     44 -
 .../sherpa/sherpa-ncnn/layer/arm/cast_arm.h   |     32 -
 .../sherpa/sherpa-ncnn/layer/arm/cast_bf16.h  |    309 -
 .../sherpa/sherpa-ncnn/layer/arm/cast_fp16.h  |    325 -
 .../sherpa/sherpa-ncnn/layer/arm/clip_arm.h   |     40 -
 .../sherpa/sherpa-ncnn/layer/arm/concat_arm.h |     35 -
 .../sherpa-ncnn/layer/arm/convolution1d_arm.h |     54 -
 .../layer/arm/convolution1d_packed.h          |   1293 -
 .../layer/arm/convolution1d_packed_bf16s.h    |   1330 -
 .../layer/arm/convolution1d_packed_fp16s.h    |   1883 -
 .../sherpa-ncnn/layer/arm/convolution_1x1.h   |   2598 -
 .../sherpa-ncnn/layer/arm/convolution_2x2.h   |    443 -
 .../sherpa-ncnn/layer/arm/convolution_3x3.h   |   3013 -
 .../layer/arm/convolution_3x3_int8.h          |   1190 -
 .../layer/arm/convolution_3x3_pack1to4.h      |   1449 -
 .../arm/convolution_3x3_pack1to4_bf16s.h      |   3732 -
 .../arm/convolution_3x3_pack1to4_fp16s.h      |    646 -
 .../arm/convolution_3x3_pack1to8_fp16s.h      |    646 -
 .../layer/arm/convolution_3x3_pack4.h         |   1072 -
 .../layer/arm/convolution_3x3_pack4_bf16s.h   |   3003 -
 .../layer/arm/convolution_3x3_pack4_fp16s.h   |    574 -
 .../layer/arm/convolution_3x3_pack4to1.h      |   1153 -
 .../layer/arm/convolution_3x3_pack8_fp16s.h   |   1952 -
 .../layer/arm/convolution_3x3_winograd.h      |   9405 -
 .../arm/convolution_3x3_winograd_bf16s.h      |   4541 -
 .../arm/convolution_3x3_winograd_fp16s.h      |   5986 -
 .../layer/arm/convolution_3x3_winograd_int8.h |   5719 -
 .../sherpa-ncnn/layer/arm/convolution_4x4.h   |    400 -
 .../sherpa-ncnn/layer/arm/convolution_5x5.h   |   1406 -
 .../layer/arm/convolution_5x5_pack4.h         |   5007 -
 .../layer/arm/convolution_5x5_pack4_bf16s.h   |  12974 --
 .../layer/arm/convolution_5x5_pack8_fp16s.h   |   3449 -
 .../sherpa-ncnn/layer/arm/convolution_7x7.h   |   1401 -
 .../layer/arm/convolution_7x7_pack1to4.h      |   2109 -
 .../arm/convolution_7x7_pack1to4_bf16s.h      |   5575 -
 .../arm/convolution_7x7_pack1to8_fp16s.h      |   1155 -
 .../sherpa-ncnn/layer/arm/convolution_arm.h   |     76 -
 .../layer/arm/convolution_im2col_gemm.h       |   6904 -
 .../layer/arm/convolution_im2col_gemm_bf16s.h |   6085 -
 .../arm/convolution_im2col_gemm_bf16s_fp16s.h |   1147 -
 .../layer/arm/convolution_im2col_gemm_fp16s.h |   3221 -
 .../layer/arm/convolution_im2col_gemm_int8.h  |  13541 --
 .../layer/arm/convolution_packed.h            |   1321 -
 .../layer/arm/convolution_packed_bf16s.h      |   1358 -
 .../layer/arm/convolution_packed_fp16s.h      |   1933 -
 .../layer/arm/convolution_packed_int8.h       |   1573 -
 .../layer/arm/convolutiondepthwise_3x3.h      |   1076 -
 .../arm/convolutiondepthwise_3x3_fp16s.h      |    457 -
 .../layer/arm/convolutiondepthwise_3x3_int8.h |   1798 -
 .../arm/convolutiondepthwise_3x3_pack4.h      |   1290 -
 .../convolutiondepthwise_3x3_pack4_bf16s.h    |   1501 -
 .../convolutiondepthwise_3x3_pack8_fp16s.h    |    909 -
 .../arm/convolutiondepthwise_3x3_pack8_int8.h |    482 -
 .../layer/arm/convolutiondepthwise_5x5.h      |   2250 -
 .../arm/convolutiondepthwise_5x5_pack4.h      |   1737 -
 .../convolutiondepthwise_5x5_pack4_bf16s.h    |   3818 -
 .../convolutiondepthwise_5x5_pack8_fp16s.h    |   1426 -
 .../layer/arm/convolutiondepthwise_arm.h      |     61 -
 .../sherpa/sherpa-ncnn/layer/arm/crop_arm.h   |     34 -
 .../sherpa-ncnn/layer/arm/deconvolution_3x3.h |    267 -
 .../sherpa-ncnn/layer/arm/deconvolution_4x4.h |    334 -
 .../layer/arm/deconvolution_4x4_fp16s.h       |    159 -
 .../sherpa-ncnn/layer/arm/deconvolution_arm.h |     57 -
 .../layer/arm/deconvolutiondepthwise_arm.h    |     55 -
 .../sherpa-ncnn/layer/arm/dequantize_arm.h    |     41 -
 .../sherpa-ncnn/layer/arm/dropout_arm.h       |     32 -
 .../sherpa-ncnn/layer/arm/eltwise_arm.h       |     41 -
 .../sherpa-ncnn/layer/arm/flatten_arm.h       |     36 -
 .../sherpa/sherpa-ncnn/layer/arm/gelu_arm.h   |     42 -
 .../sherpa/sherpa-ncnn/layer/arm/gemm_arm.h   |     54 -
 .../sherpa/sherpa-ncnn/layer/arm/gemm_bf16s.h |   3530 -
 .../sherpa-ncnn/layer/arm/gemm_bf16s_fp16s.h  |   1584 -
 .../sherpa/sherpa-ncnn/layer/arm/gemm_fp16s.h |   3945 -
 .../sherpa/sherpa-ncnn/layer/arm/gru_arm.h    |     53 -
 .../sherpa-ncnn/layer/arm/hardsigmoid_arm.h   |     41 -
 .../sherpa-ncnn/layer/arm/hardswish_arm.h     |     41 -
 .../sherpa-ncnn/layer/arm/innerproduct_arm.h  |     64 -
 .../layer/arm/innerproduct_fp16s.h            |    697 -
 .../layer/arm/innerproduct_gemm_fp16s.h       |    458 -
 .../sherpa-ncnn/layer/arm/instancenorm_arm.h  |     40 -
 .../sherpa/sherpa-ncnn/layer/arm/interp_arm.h |     41 -
 .../sherpa-ncnn/layer/arm/interp_bicubic.h    |    261 -
 .../layer/arm/interp_bicubic_bf16s.h          |    188 -
 .../layer/arm/interp_bicubic_fp16s.h          |    436 -
 .../layer/arm/interp_bicubic_pack4.h          |    272 -
 .../layer/arm/interp_bicubic_pack4_bf16s.h    |    272 -
 .../layer/arm/interp_bicubic_pack4_fp16s.h    |    529 -
 .../layer/arm/interp_bicubic_pack8_fp16s.h    |    271 -
 .../sherpa-ncnn/layer/arm/interp_bilinear.h   |    222 -
 .../layer/arm/interp_bilinear_bf16s.h         |    135 -
 .../layer/arm/interp_bilinear_fp16s.h         |    274 -
 .../layer/arm/interp_bilinear_pack4.h         |    120 -
 .../layer/arm/interp_bilinear_pack4_bf16s.h   |    120 -
 .../layer/arm/interp_bilinear_pack4_fp16s.h   |    227 -
 .../layer/arm/interp_bilinear_pack8_fp16s.h   |    120 -
 .../sherpa/sherpa-ncnn/layer/arm/lrn_arm.h    |     30 -
 .../sherpa/sherpa-ncnn/layer/arm/lstm_arm.h   |     53 -
 .../sherpa/sherpa-ncnn/layer/arm/matmul_arm.h |     38 -
 .../sherpa/sherpa-ncnn/layer/arm/mish_arm.h   |     41 -
 .../layer/arm/multiheadattention_arm.h        |     46 -
 .../sherpa-ncnn/layer/arm/neon_mathfun.h      |    399 -
 .../layer/arm/neon_mathfun_fp16s.h            |    633 -
 .../sherpa-ncnn/layer/arm/neon_mathfun_tanh.h |     82 -
 .../sherpa-ncnn/layer/arm/packing_arm.h       |     36 -
 .../sherpa-ncnn/layer/arm/padding_arm.h       |     50 -
 .../sherpa-ncnn/layer/arm/padding_pack4.h     |    562 -
 .../layer/arm/padding_pack4_bf16s_fp16s.h     |    548 -
 .../layer/arm/padding_pack8_fp16s.h           |    376 -
 .../layer/arm/padding_pack8_int8.h            |    554 -
 .../sherpa-ncnn/layer/arm/pixelshuffle_arm.h  |     35 -
 .../sherpa-ncnn/layer/arm/pooling_2x2.h       |    113 -
 .../sherpa-ncnn/layer/arm/pooling_2x2_pack4.h |    144 -
 .../layer/arm/pooling_2x2_pack4_bf16s.h       |    194 -
 .../sherpa-ncnn/layer/arm/pooling_3x3.h       |    191 -
 .../sherpa-ncnn/layer/arm/pooling_3x3_pack4.h |    333 -
 .../layer/arm/pooling_3x3_pack4_bf16s.h       |    448 -
 .../sherpa-ncnn/layer/arm/pooling_arm.h       |     42 -
 .../sherpa/sherpa-ncnn/layer/arm/prelu_arm.h  |     41 -
 .../sherpa-ncnn/layer/arm/quantize_arm.h      |     41 -
 .../sherpa/sherpa-ncnn/layer/arm/relu_arm.h   |     41 -
 .../sherpa-ncnn/layer/arm/requantize_arm.h    |     33 -
 .../layer/arm/requantize_leakyrelu_pack4.h    |    275 -
 .../layer/arm/requantize_leakyrelu_pack8.h    |    202 -
 .../layer/arm/requantize_relu_pack4.h         |    271 -
 .../layer/arm/requantize_relu_pack8.h         |    200 -
 .../sherpa-ncnn/layer/arm/reshape_arm.h       |     35 -
 .../sherpa/sherpa-ncnn/layer/arm/rnn_arm.h    |     53 -
 .../sherpa/sherpa-ncnn/layer/arm/scale_arm.h  |     32 -
 .../sherpa/sherpa-ncnn/layer/arm/selu_arm.h   |     30 -
 .../layer/arm/shufflechannel_arm.h            |     35 -
 .../sherpa-ncnn/layer/arm/sigmoid_arm.h       |     41 -
 .../sherpa/sherpa-ncnn/layer/arm/slice_arm.h  |     35 -
 .../sherpa-ncnn/layer/arm/softmax_arm.h       |     40 -
 .../sherpa/sherpa-ncnn/layer/arm/swish_arm.h  |     41 -
 .../sherpa/sherpa-ncnn/layer/arm/tanh_arm.h   |     41 -
 .../sherpa-ncnn/layer/arm/unaryop_arm.h       |     40 -
 .../sherpa/sherpa-ncnn/layer/batchnorm.h      |     50 -
 .../include/sherpa/sherpa-ncnn/layer/bias.h   |     43 -
 .../sherpa/sherpa-ncnn/layer/binaryop.h       |     60 -
 .../include/sherpa/sherpa-ncnn/layer/bnll.h   |     34 -
 .../include/sherpa/sherpa-ncnn/layer/cast.h   |     44 -
 .../include/sherpa/sherpa-ncnn/layer/celu.h   |     37 -
 .../include/sherpa/sherpa-ncnn/layer/clip.h   |     38 -
 .../include/sherpa/sherpa-ncnn/layer/concat.h |     37 -
 .../sherpa/sherpa-ncnn/layer/convolution.h    |     82 -
 .../sherpa/sherpa-ncnn/layer/convolution1d.h  |     65 -
 .../sherpa/sherpa-ncnn/layer/convolution3d.h  |     68 -
 .../sherpa-ncnn/layer/convolutiondepthwise.h  |     83 -
 .../layer/convolutiondepthwise1d.h            |     66 -
 .../layer/convolutiondepthwise3d.h            |     69 -
 .../include/sherpa/sherpa-ncnn/layer/copyto.h |     48 -
 .../include/sherpa/sherpa-ncnn/layer/crop.h   |     67 -
 .../sherpa/sherpa-ncnn/layer/cumulativesum.h  |     37 -
 .../sherpa/sherpa-ncnn/layer/deconvolution.h  |     72 -
 .../sherpa-ncnn/layer/deconvolution1d.h       |     64 -
 .../sherpa-ncnn/layer/deconvolution3d.h       |     74 -
 .../layer/deconvolutiondepthwise.h            |     73 -
 .../layer/deconvolutiondepthwise1d.h          |     65 -
 .../layer/deconvolutiondepthwise3d.h          |     75 -
 .../sherpa/sherpa-ncnn/layer/deepcopy.h       |     32 -
 .../sherpa-ncnn/layer/deformableconv2d.h      |     61 -
 .../sherpa/sherpa-ncnn/layer/dequantize.h     |     43 -
 .../sherpa-ncnn/layer/detectionoutput.h       |     42 -
 .../include/sherpa/sherpa-ncnn/layer/diag.h   |     37 -
 .../sherpa/sherpa-ncnn/layer/dropout.h        |     37 -
 .../include/sherpa/sherpa-ncnn/layer/einsum.h |     39 -
 .../sherpa/sherpa-ncnn/layer/eltwise.h        |     46 -
 .../include/sherpa/sherpa-ncnn/layer/elu.h    |     37 -
 .../include/sherpa/sherpa-ncnn/layer/embed.h  |     48 -
 .../include/sherpa/sherpa-ncnn/layer/erf.h    |     32 -
 .../include/sherpa/sherpa-ncnn/layer/exp.h    |     39 -
 .../sherpa/sherpa-ncnn/layer/expanddims.h     |     41 -
 .../sherpa/sherpa-ncnn/layer/flatten.h        |     32 -
 .../include/sherpa/sherpa-ncnn/layer/fold.h   |     48 -
 .../sherpa-ncnn/layer/fused_activation.h      |    139 -
 .../include/sherpa/sherpa-ncnn/layer/gelu.h   |     37 -
 .../include/sherpa/sherpa-ncnn/layer/gemm.h   |     65 -
 .../include/sherpa/sherpa-ncnn/layer/glu.h    |     38 -
 .../sherpa/sherpa-ncnn/layer/gridsample.h     |     56 -
 .../sherpa/sherpa-ncnn/layer/groupnorm.h      |     47 -
 .../include/sherpa/sherpa-ncnn/layer/gru.h    |     47 -
 .../sherpa/sherpa-ncnn/layer/hardsigmoid.h    |     37 -
 .../sherpa/sherpa-ncnn/layer/hardswish.h      |     37 -
 .../sherpa/sherpa-ncnn/layer/innerproduct.h   |     63 -
 .../include/sherpa/sherpa-ncnn/layer/input.h  |     45 -
 .../sherpa/sherpa-ncnn/layer/instancenorm.h   |     46 -
 .../include/sherpa/sherpa-ncnn/layer/interp.h |     46 -
 .../sherpa/sherpa-ncnn/layer/layernorm.h      |     46 -
 .../include/sherpa/sherpa-ncnn/layer/log.h    |     39 -
 .../layer/loongarch/absval_loongarch.h        |     32 -
 .../layer/loongarch/batchnorm_loongarch.h     |     32 -
 .../layer/loongarch/bias_loongarch.h          |     30 -
 .../layer/loongarch/binaryop_loongarch.h      |     34 -
 .../layer/loongarch/cast_loongarch.h          |     32 -
 .../layer/loongarch/clip_loongarch.h          |     32 -
 .../layer/loongarch/concat_loongarch.h        |     32 -
 .../layer/loongarch/convolution1d_loongarch.h |     41 -
 .../layer/loongarch/convolution_1x1.h         |     26 -
 .../layer/loongarch/convolution_1x1_int8.h    |     83 -
 .../loongarch/convolution_1x1_pack1to4_int8.h |     83 -
 .../layer/loongarch/convolution_1x1_pack4.h   |     65 -
 .../loongarch/convolution_1x1_pack4to1.h      |     65 -
 .../loongarch/convolution_1x1_pack8to1_int8.h |     65 -
 .../loongarch/convolution_1x1_pack8to4_int8.h |     65 -
 .../layer/loongarch/convolution_3x3.h         |    412 -
 .../layer/loongarch/convolution_3x3_int8.h    |    252 -
 .../loongarch/convolution_3x3_pack1to4.h      |    812 -
 .../layer/loongarch/convolution_3x3_pack4.h   |    425 -
 .../loongarch/convolution_3x3_pack8to1_int8.h |    177 -
 .../loongarch/convolution_3x3_pack8to4_int8.h |    161 -
 .../loongarch/convolution_7x7_pack1to4.h      |    652 -
 .../layer/loongarch/convolution_int8.h        |     82 -
 .../layer/loongarch/convolution_loongarch.h   |     56 -
 .../layer/loongarch/convolution_pack1to4.h    |     90 -
 .../loongarch/convolution_pack1to4_int8.h     |     87 -
 .../layer/loongarch/convolution_pack4.h       |    102 -
 .../layer/loongarch/convolution_pack4to1.h    |     94 -
 .../loongarch/convolution_pack8to1_int8.h     |     87 -
 .../loongarch/convolution_pack8to4_int8.h     |    120 -
 .../layer/loongarch/convolution_sgemm.h       |    650 -
 .../layer/loongarch/convolution_sgemm_int8.h  |    800 -
 .../convolution_sgemm_pack1to4_int8.h         |    481 -
 .../layer/loongarch/convolution_sgemm_pack4.h |    519 -
 .../loongarch/convolution_sgemm_pack4to1.h    |    667 -
 .../convolution_sgemm_pack8to1_int8.h         |    458 -
 .../convolution_sgemm_pack8to4_int8.h         |    324 -
 .../loongarch/convolution_winograd_dot.h      |    495 -
 .../loongarch/convolution_winograd_dot_int8.h |    594 -
 .../convolution_winograd_dot_pack4.h          |    448 -
 .../convolution_winograd_dot_pack8to1_int8.h  |    363 -
 .../convolution_winograd_dot_pack8to4_int8.h  |    233 -
 .../convolution_winograd_transform.h          |    405 -
 .../convolution_winograd_transform_int8.h     |    229 -
 .../convolution_winograd_transform_pack4.h    |    730 -
 ...onvolution_winograd_transform_pack4_int8.h |    166 -
 ...onvolution_winograd_transform_pack8_int8.h |    132 -
 .../loongarch/convolutiondepthwise_3x3.h      |    193 -
 .../convolutiondepthwise_3x3_pack4.h          |    464 -
 .../convolutiondepthwise_5x5_pack4.h          |    511 -
 .../convolutiondepthwise_loongarch.h          |     50 -
 .../layer/loongarch/crop_loongarch.h          |     34 -
 .../layer/loongarch/deconvolution_loongarch.h |     40 -
 .../layer/loongarch/deconvolution_pack1to4.h  |     99 -
 .../layer/loongarch/deconvolution_pack4.h     |    106 -
 .../layer/loongarch/deconvolution_pack4to1.h  |    101 -
 .../deconvolutiondepthwise_loongarch.h        |     45 -
 .../layer/loongarch/dequantize_loongarch.h    |     32 -
 .../layer/loongarch/dropout_loongarch.h       |     32 -
 .../layer/loongarch/eltwise_loongarch.h       |     32 -
 .../layer/loongarch/flatten_loongarch.h       |     35 -
 .../layer/loongarch/hardsigmoid_loongarch.h   |     32 -
 .../layer/loongarch/hardswish_loongarch.h     |     32 -
 .../layer/loongarch/innerproduct_loongarch.h  |     54 -
 .../layer/loongarch/interp_bicubic.h          |    261 -
 .../layer/loongarch/interp_bicubic_pack4.h    |    286 -
 .../layer/loongarch/interp_bilinear.h         |    172 -
 .../layer/loongarch/interp_bilinear_pack4.h   |    123 -
 .../layer/loongarch/interp_loongarch.h        |     32 -
 .../layer/loongarch/loongarch_activation.h    |     70 -
 .../layer/loongarch/loongarch_usability.h     |    235 -
 .../sherpa-ncnn/layer/loongarch/lsx_mathfun.h |    272 -
 .../layer/loongarch/mish_loongarch.h          |     32 -
 .../layer/loongarch/packing_loongarch.h       |     35 -
 .../layer/loongarch/padding_loongarch.h       |     35 -
 .../layer/loongarch/padding_pack4.h           |    213 -
 .../layer/loongarch/padding_pack8_int8.h      |    171 -
 .../layer/loongarch/pooling_loongarch.h       |     33 -
 .../layer/loongarch/prelu_loongarch.h         |     32 -
 .../layer/loongarch/quantize_loongarch.h      |     32 -
 .../layer/loongarch/relu_loongarch.h          |     32 -
 .../loongarch/requantize_leakyrelu_pack4.h    |    271 -
 .../loongarch/requantize_leakyrelu_pack8.h    |    188 -
 .../layer/loongarch/requantize_loongarch.h    |     32 -
 .../layer/loongarch/requantize_relu_pack4.h   |    267 -
 .../layer/loongarch/requantize_relu_pack8.h   |    186 -
 .../layer/loongarch/sigmoid_loongarch.h       |     32 -
 .../layer/loongarch/slice_loongarch.h         |     32 -
 .../layer/loongarch/softmax_loongarch.h       |     30 -
 .../layer/loongarch/swish_loongarch.h         |     32 -
 .../layer/loongarch/tanh_loongarch.h          |     32 -
 .../layer/loongarch/unaryop_loongarch.h       |     32 -
 .../include/sherpa/sherpa-ncnn/layer/lrn.h    |     48 -
 .../include/sherpa/sherpa-ncnn/layer/lstm.h   |     49 -
 .../include/sherpa/sherpa-ncnn/layer/matmul.h |     37 -
 .../sherpa/sherpa-ncnn/layer/memorydata.h     |     45 -
 .../sherpa-ncnn/layer/mips/absval_mips.h      |     32 -
 .../sherpa-ncnn/layer/mips/batchnorm_mips.h   |     32 -
 .../sherpa/sherpa-ncnn/layer/mips/bias_mips.h |     30 -
 .../sherpa-ncnn/layer/mips/binaryop_mips.h    |     34 -
 .../sherpa/sherpa-ncnn/layer/mips/cast_mips.h |     32 -
 .../sherpa/sherpa-ncnn/layer/mips/clip_mips.h |     32 -
 .../sherpa-ncnn/layer/mips/concat_mips.h      |     32 -
 .../layer/mips/convolution1d_mips.h           |     41 -
 .../sherpa-ncnn/layer/mips/convolution_1x1.h  |     26 -
 .../layer/mips/convolution_1x1_int8.h         |     83 -
 .../mips/convolution_1x1_pack1to4_int8.h      |     83 -
 .../layer/mips/convolution_1x1_pack4.h        |     65 -
 .../layer/mips/convolution_1x1_pack4to1.h     |     65 -
 .../mips/convolution_1x1_pack8to1_int8.h      |     65 -
 .../mips/convolution_1x1_pack8to4_int8.h      |     65 -
 .../sherpa-ncnn/layer/mips/convolution_3x3.h  |    412 -
 .../layer/mips/convolution_3x3_int8.h         |    285 -
 .../layer/mips/convolution_3x3_pack1to4.h     |    812 -
 .../layer/mips/convolution_3x3_pack4.h        |    425 -
 .../mips/convolution_3x3_pack8to1_int8.h      |    177 -
 .../mips/convolution_3x3_pack8to4_int8.h      |    161 -
 .../layer/mips/convolution_7x7_pack1to4.h     |    652 -
 .../sherpa-ncnn/layer/mips/convolution_int8.h |     82 -
 .../sherpa-ncnn/layer/mips/convolution_mips.h |     56 -
 .../layer/mips/convolution_pack1to4.h         |     90 -
 .../layer/mips/convolution_pack1to4_int8.h    |     87 -
 .../layer/mips/convolution_pack4.h            |    102 -
 .../layer/mips/convolution_pack4to1.h         |     94 -
 .../layer/mips/convolution_pack8to1_int8.h    |     87 -
 .../layer/mips/convolution_pack8to4_int8.h    |    120 -
 .../layer/mips/convolution_sgemm.h            |    650 -
 .../layer/mips/convolution_sgemm_int8.h       |   1179 -
 .../mips/convolution_sgemm_pack1to4_int8.h    |    481 -
 .../layer/mips/convolution_sgemm_pack4.h      |    519 -
 .../layer/mips/convolution_sgemm_pack4to1.h   |    666 -
 .../mips/convolution_sgemm_pack8to1_int8.h    |    458 -
 .../mips/convolution_sgemm_pack8to4_int8.h    |    324 -
 .../layer/mips/convolution_winograd_dot.h     |    495 -
 .../mips/convolution_winograd_dot_int8.h      |    883 -
 .../mips/convolution_winograd_dot_pack4.h     |    448 -
 .../convolution_winograd_dot_pack8to1_int8.h  |    363 -
 .../convolution_winograd_dot_pack8to4_int8.h  |    233 -
 .../mips/convolution_winograd_transform.h     |    405 -
 .../convolution_winograd_transform_int8.h     |    229 -
 .../convolution_winograd_transform_pack4.h    |    730 -
 ...onvolution_winograd_transform_pack4_int8.h |    166 -
 ...onvolution_winograd_transform_pack8_int8.h |    132 -
 .../layer/mips/convolutiondepthwise_3x3.h     |    193 -
 .../mips/convolutiondepthwise_3x3_pack4.h     |    464 -
 .../mips/convolutiondepthwise_5x5_pack4.h     |    511 -
 .../layer/mips/convolutiondepthwise_mips.h    |     50 -
 .../sherpa/sherpa-ncnn/layer/mips/crop_mips.h |     34 -
 .../layer/mips/deconvolution_mips.h           |     40 -
 .../layer/mips/deconvolution_pack1to4.h       |     99 -
 .../layer/mips/deconvolution_pack4.h          |    106 -
 .../layer/mips/deconvolution_pack4to1.h       |    101 -
 .../layer/mips/deconvolutiondepthwise_mips.h  |     45 -
 .../sherpa-ncnn/layer/mips/dequantize_mips.h  |     32 -
 .../sherpa-ncnn/layer/mips/dropout_mips.h     |     32 -
 .../sherpa-ncnn/layer/mips/eltwise_mips.h     |     32 -
 .../sherpa-ncnn/layer/mips/flatten_mips.h     |     35 -
 .../sherpa-ncnn/layer/mips/hardsigmoid_mips.h |     32 -
 .../sherpa-ncnn/layer/mips/hardswish_mips.h   |     32 -
 .../layer/mips/innerproduct_mips.h            |     54 -
 .../sherpa-ncnn/layer/mips/interp_bicubic.h   |    261 -
 .../layer/mips/interp_bicubic_pack4.h         |    286 -
 .../sherpa-ncnn/layer/mips/interp_bilinear.h  |    172 -
 .../layer/mips/interp_bilinear_pack4.h        |    123 -
 .../sherpa-ncnn/layer/mips/interp_mips.h      |     32 -
 .../sherpa-ncnn/layer/mips/loongson_mmi.h     |    268 -
 .../sherpa-ncnn/layer/mips/mips_activation.h  |     70 -
 .../sherpa-ncnn/layer/mips/mips_usability.h   |    251 -
 .../sherpa/sherpa-ncnn/layer/mips/mish_mips.h |     32 -
 .../sherpa-ncnn/layer/mips/msa_mathfun.h      |    270 -
 .../sherpa-ncnn/layer/mips/packing_mips.h     |     35 -
 .../sherpa-ncnn/layer/mips/padding_mips.h     |     35 -
 .../sherpa-ncnn/layer/mips/padding_pack4.h    |    213 -
 .../layer/mips/padding_pack8_int8.h           |    171 -
 .../sherpa-ncnn/layer/mips/pooling_mips.h     |     33 -
 .../sherpa-ncnn/layer/mips/prelu_mips.h       |     32 -
 .../sherpa-ncnn/layer/mips/quantize_mips.h    |     32 -
 .../sherpa/sherpa-ncnn/layer/mips/relu_mips.h |     32 -
 .../layer/mips/requantize_leakyrelu_pack4.h   |    267 -
 .../layer/mips/requantize_leakyrelu_pack8.h   |    188 -
 .../sherpa-ncnn/layer/mips/requantize_mips.h  |     32 -
 .../layer/mips/requantize_relu_pack4.h        |    263 -
 .../layer/mips/requantize_relu_pack8.h        |    186 -
 .../sherpa-ncnn/layer/mips/sigmoid_mips.h     |     32 -
 .../sherpa-ncnn/layer/mips/slice_mips.h       |     32 -
 .../sherpa-ncnn/layer/mips/softmax_mips.h     |     30 -
 .../sherpa-ncnn/layer/mips/swish_mips.h       |     32 -
 .../sherpa/sherpa-ncnn/layer/mips/tanh_mips.h |     32 -
 .../sherpa-ncnn/layer/mips/unaryop_mips.h     |     32 -
 .../include/sherpa/sherpa-ncnn/layer/mish.h   |     32 -
 .../sherpa-ncnn/layer/multiheadattention.h    |     53 -
 .../include/sherpa/sherpa-ncnn/layer/mvn.h    |     39 -
 .../include/sherpa/sherpa-ncnn/layer/noop.h   |     32 -
 .../sherpa/sherpa-ncnn/layer/normalize.h      |     51 -
 .../sherpa/sherpa-ncnn/layer/packing.h        |     52 -
 .../sherpa/sherpa-ncnn/layer/padding.h        |     50 -
 .../sherpa/sherpa-ncnn/layer/permute.h        |     37 -
 .../sherpa/sherpa-ncnn/layer/pixelshuffle.h   |     38 -
 .../sherpa/sherpa-ncnn/layer/pooling.h        |     61 -
 .../sherpa/sherpa-ncnn/layer/pooling1d.h      |     56 -
 .../sherpa/sherpa-ncnn/layer/pooling3d.h      |     66 -
 .../include/sherpa/sherpa-ncnn/layer/power.h  |     39 -
 .../include/sherpa/sherpa-ncnn/layer/prelu.h  |     40 -
 .../sherpa/sherpa-ncnn/layer/priorbox.h       |     49 -
 .../sherpa/sherpa-ncnn/layer/proposal.h       |     48 -
 .../sherpa/sherpa-ncnn/layer/psroipooling.h   |     40 -
 .../sherpa/sherpa-ncnn/layer/quantize.h       |     40 -
 .../sherpa/sherpa-ncnn/layer/reduction.h      |     57 -
 .../include/sherpa/sherpa-ncnn/layer/relu.h   |     37 -
 .../include/sherpa/sherpa-ncnn/layer/reorg.h  |     38 -
 .../sherpa/sherpa-ncnn/layer/requantize.h     |     57 -
 .../sherpa/sherpa-ncnn/layer/reshape.h        |     49 -
 .../sherpa-ncnn/layer/riscv/absval_riscv.h    |     37 -
 .../sherpa-ncnn/layer/riscv/batchnorm_riscv.h |     37 -
 .../sherpa-ncnn/layer/riscv/binaryop_riscv.h  |     42 -
 .../sherpa-ncnn/layer/riscv/cast_riscv.h      |     32 -
 .../sherpa-ncnn/layer/riscv/clip_riscv.h      |     38 -
 .../sherpa-ncnn/layer/riscv/concat_riscv.h    |     35 -
 .../layer/riscv/convolution1d_riscv.h         |     52 -
 .../sherpa-ncnn/layer/riscv/convolution_1x1.h |     26 -
 .../layer/riscv/convolution_1x1_fp16s.h       |     26 -
 .../layer/riscv/convolution_1x1_pack1ton.h    |     26 -
 .../riscv/convolution_1x1_pack1ton_fp16s.h    |     26 -
 .../layer/riscv/convolution_1x1_packn.h       |     68 -
 .../layer/riscv/convolution_1x1_packn_fp16s.h |     68 -
 .../layer/riscv/convolution_1x1_packnto1.h    |     68 -
 .../riscv/convolution_1x1_packnto1_fp16s.h    |     68 -
 .../sherpa-ncnn/layer/riscv/convolution_3x3.h |    413 -
 .../layer/riscv/convolution_3x3_pack1ton.h    |    566 -
 .../riscv/convolution_3x3_pack1ton_fp16s.h    |    566 -
 .../layer/riscv/convolution_3x3_packn.h       |    432 -
 .../layer/riscv/convolution_3x3_packn_fp16s.h |    432 -
 .../layer/riscv/convolution_7x7_pack1ton.h    |   1007 -
 .../riscv/convolution_7x7_pack1ton_fp16s.h    |   1007 -
 .../layer/riscv/convolution_fp16s.h           |     90 -
 .../layer/riscv/convolution_pack1ton.h        |     93 -
 .../layer/riscv/convolution_pack1ton_fp16s.h  |    173 -
 .../layer/riscv/convolution_packn.h           |     97 -
 .../layer/riscv/convolution_packn_fp16s.h     |    186 -
 .../layer/riscv/convolution_packnto1.h        |     97 -
 .../layer/riscv/convolution_packnto1_fp16s.h  |    191 -
 .../layer/riscv/convolution_riscv.h           |     55 -
 .../layer/riscv/convolution_sgemm.h           |    675 -
 .../layer/riscv/convolution_sgemm_fp16s.h     |    540 -
 .../layer/riscv/convolution_sgemm_pack1ton.h  |    136 -
 .../riscv/convolution_sgemm_pack1ton_fp16s.h  |    136 -
 .../layer/riscv/convolution_sgemm_packn.h     |    414 -
 .../riscv/convolution_sgemm_packn_fp16s.h     |    486 -
 .../layer/riscv/convolution_sgemm_packnto1.h  |    728 -
 .../riscv/convolution_sgemm_packnto1_fp16s.h  |    729 -
 .../layer/riscv/convolution_winograd_dot.h    |    518 -
 .../riscv/convolution_winograd_dot_packn.h    |    317 -
 .../convolution_winograd_dot_packn_fp16s.h    |    317 -
 .../riscv/convolution_winograd_transform.h    |    402 -
 .../convolution_winograd_transform_packn.h    |    745 -
 ...nvolution_winograd_transform_packn_fp16s.h |    745 -
 .../layer/riscv/convolutiondepthwise_3x3.h    |    193 -
 .../riscv/convolutiondepthwise_3x3_packn.h    |    444 -
 .../convolutiondepthwise_3x3_packn_fp16s.h    |    444 -
 .../riscv/convolutiondepthwise_5x5_packn.h    |    492 -
 .../convolutiondepthwise_5x5_packn_fp16s.h    |    492 -
 .../layer/riscv/convolutiondepthwise_riscv.h  |     54 -
 .../sherpa-ncnn/layer/riscv/crop_riscv.h      |     34 -
 .../layer/riscv/deconvolution_fp16s.h         |     99 -
 .../layer/riscv/deconvolution_pack1ton.h      |    101 -
 .../riscv/deconvolution_pack1ton_fp16s.h      |    189 -
 .../layer/riscv/deconvolution_packn.h         |    103 -
 .../layer/riscv/deconvolution_packn_fp16s.h   |    193 -
 .../layer/riscv/deconvolution_packnto1.h      |    104 -
 .../riscv/deconvolution_packnto1_fp16s.h      |    205 -
 .../layer/riscv/deconvolution_riscv.h         |     50 -
 .../riscv/deconvolutiondepthwise_riscv.h      |     53 -
 .../sherpa-ncnn/layer/riscv/dropout_riscv.h   |     35 -
 .../sherpa-ncnn/layer/riscv/flatten_riscv.h   |     36 -
 .../sherpa-ncnn/layer/riscv/gelu_riscv.h      |     32 -
 .../sherpa-ncnn/layer/riscv/gemm_riscv.h      |     43 -
 .../sherpa-ncnn/layer/riscv/gru_riscv.h       |     48 -
 .../layer/riscv/hardsigmoid_riscv.h           |     37 -
 .../sherpa-ncnn/layer/riscv/hardswish_riscv.h |     40 -
 .../layer/riscv/innerproduct_riscv.h          |     50 -
 .../layer/riscv/instancenorm_riscv.h          |     36 -
 .../sherpa-ncnn/layer/riscv/interp_bicubic.h  |    261 -
 .../layer/riscv/interp_bicubic_fp16s.h        |    436 -
 .../layer/riscv/interp_bicubic_packn.h        |    242 -
 .../layer/riscv/interp_bicubic_packn_fp16s.h  |    471 -
 .../sherpa-ncnn/layer/riscv/interp_bilinear.h |    222 -
 .../layer/riscv/interp_bilinear_fp16s.h       |    252 -
 .../layer/riscv/interp_bilinear_packn.h       |    120 -
 .../layer/riscv/interp_bilinear_packn_fp16s.h |    227 -
 .../sherpa-ncnn/layer/riscv/interp_riscv.h    |     38 -
 .../sherpa-ncnn/layer/riscv/mish_riscv.h      |     38 -
 .../sherpa-ncnn/layer/riscv/packing_riscv.h   |     36 -
 .../sherpa-ncnn/layer/riscv/padding_packn.h   |    231 -
 .../sherpa-ncnn/layer/riscv/padding_riscv.h   |     49 -
 .../sherpa-ncnn/layer/riscv/pooling_riscv.h   |     39 -
 .../sherpa-ncnn/layer/riscv/prelu_riscv.h     |     38 -
 .../sherpa-ncnn/layer/riscv/relu_riscv.h      |     37 -
 .../layer/riscv/riscv_activation.h            |     81 -
 .../sherpa-ncnn/layer/riscv/riscv_usability.h |    619 -
 .../sherpa-ncnn/layer/riscv/rvv_mathfun.h     |    583 -
 .../layer/riscv/rvv_mathfun_fp16s.h           |    419 -
 .../sherpa-ncnn/layer/riscv/selu_riscv.h      |     30 -
 .../sherpa-ncnn/layer/riscv/sigmoid_riscv.h   |     38 -
 .../sherpa-ncnn/layer/riscv/softmax_riscv.h   |     32 -
 .../sherpa-ncnn/layer/riscv/swish_riscv.h     |     38 -
 .../sherpa-ncnn/layer/riscv/tanh_riscv.h      |     38 -
 .../sherpa-ncnn/layer/riscv/unaryop_riscv.h   |     37 -
 .../include/sherpa/sherpa-ncnn/layer/rnn.h    |     47 -
 .../sherpa/sherpa-ncnn/layer/roialign.h       |     42 -
 .../sherpa/sherpa-ncnn/layer/roipooling.h     |     39 -
 .../include/sherpa/sherpa-ncnn/layer/scale.h  |     46 -
 .../include/sherpa/sherpa-ncnn/layer/selu.h   |     38 -
 .../include/sherpa/sherpa-ncnn/layer/shrink.h |     37 -
 .../sherpa/sherpa-ncnn/layer/shufflechannel.h |     38 -
 .../sherpa/sherpa-ncnn/layer/sigmoid.h        |     32 -
 .../include/sherpa/sherpa-ncnn/layer/slice.h  |     39 -
 .../sherpa/sherpa-ncnn/layer/softmax.h        |     37 -
 .../sherpa/sherpa-ncnn/layer/softplus.h       |     32 -
 .../include/sherpa/sherpa-ncnn/layer/split.h  |     32 -
 .../include/sherpa/sherpa-ncnn/layer/spp.h    |     45 -
 .../sherpa/sherpa-ncnn/layer/squeeze.h        |     41 -
 .../sherpa-ncnn/layer/statisticspooling.h     |     36 -
 .../include/sherpa/sherpa-ncnn/layer/swish.h  |     32 -
 .../include/sherpa/sherpa-ncnn/layer/tanh.h   |     32 -
 .../sherpa/sherpa-ncnn/layer/threshold.h      |     37 -
 .../include/sherpa/sherpa-ncnn/layer/tile.h   |     39 -
 .../sherpa/sherpa-ncnn/layer/unaryop.h        |     62 -
 .../include/sherpa/sherpa-ncnn/layer/unfold.h |     50 -
 .../sherpa-ncnn/layer/vulkan/absval_vulkan.h  |     42 -
 .../layer/vulkan/batchnorm_vulkan.h           |     49 -
 .../layer/vulkan/binaryop_vulkan.h            |     55 -
 .../sherpa-ncnn/layer/vulkan/cast_vulkan.h    |     45 -
 .../sherpa-ncnn/layer/vulkan/celu_vulkan.h    |     42 -
 .../sherpa-ncnn/layer/vulkan/clip_vulkan.h    |     42 -
 .../sherpa-ncnn/layer/vulkan/concat_vulkan.h  |     45 -
 .../layer/vulkan/convolution1d_vulkan.h       |     55 -
 .../layer/vulkan/convolution_vulkan.h         |     77 -
 .../vulkan/convolutiondepthwise_vulkan.h      |     68 -
 .../sherpa-ncnn/layer/vulkan/crop_vulkan.h    |     53 -
 .../layer/vulkan/deconvolution_vulkan.h       |     59 -
 .../vulkan/deconvolutiondepthwise_vulkan.h    |     68 -
 .../layer/vulkan/deepcopy_vulkan.h            |     42 -
 .../sherpa-ncnn/layer/vulkan/dropout_vulkan.h |     42 -
 .../sherpa-ncnn/layer/vulkan/eltwise_vulkan.h |     42 -
 .../sherpa-ncnn/layer/vulkan/elu_vulkan.h     |     42 -
 .../sherpa-ncnn/layer/vulkan/erf_vulkan.h     |     42 -
 .../sherpa-ncnn/layer/vulkan/flatten_vulkan.h |     45 -
 .../sherpa-ncnn/layer/vulkan/gelu_vulkan.h    |     42 -
 .../sherpa-ncnn/layer/vulkan/gemm_vulkan.h    |     56 -
 .../layer/vulkan/hardsigmoid_vulkan.h         |     42 -
 .../layer/vulkan/hardswish_vulkan.h           |     42 -
 .../layer/vulkan/innerproduct_vulkan.h        |     58 -
 .../layer/vulkan/instancenorm_vulkan.h        |     66 -
 .../sherpa-ncnn/layer/vulkan/interp_vulkan.h  |     50 -
 .../sherpa-ncnn/layer/vulkan/lrn_vulkan.h     |     49 -
 .../layer/vulkan/memorydata_vulkan.h          |     42 -
 .../sherpa-ncnn/layer/vulkan/mish_vulkan.h    |     42 -
 .../layer/vulkan/multiheadattention_vulkan.h  |     57 -
 .../sherpa-ncnn/layer/vulkan/noop_vulkan.h    |     34 -
 .../layer/vulkan/normalize_vulkan.h           |     58 -
 .../sherpa-ncnn/layer/vulkan/packing_vulkan.h |     50 -
 .../sherpa-ncnn/layer/vulkan/padding_vulkan.h |     61 -
 .../sherpa-ncnn/layer/vulkan/permute_vulkan.h |     48 -
 .../layer/vulkan/pixelshuffle_vulkan.h        |     45 -
 .../sherpa-ncnn/layer/vulkan/pooling_vulkan.h |     60 -
 .../sherpa-ncnn/layer/vulkan/prelu_vulkan.h   |     47 -
 .../layer/vulkan/priorbox_vulkan.h            |     45 -
 .../sherpa-ncnn/layer/vulkan/relu_vulkan.h    |     42 -
 .../sherpa-ncnn/layer/vulkan/reorg_vulkan.h   |     45 -
 .../sherpa-ncnn/layer/vulkan/reshape_vulkan.h |     55 -
 .../sherpa-ncnn/layer/vulkan/scale_vulkan.h   |     51 -
 .../layer/vulkan/shufflechannel_vulkan.h      |     42 -
 .../sherpa-ncnn/layer/vulkan/sigmoid_vulkan.h |     42 -
 .../sherpa-ncnn/layer/vulkan/slice_vulkan.h   |     45 -
 .../sherpa-ncnn/layer/vulkan/softmax_vulkan.h |     53 -
 .../sherpa-ncnn/layer/vulkan/split_vulkan.h   |     34 -
 .../sherpa-ncnn/layer/vulkan/swish_vulkan.h   |     42 -
 .../sherpa-ncnn/layer/vulkan/tanh_vulkan.h    |     42 -
 .../sherpa-ncnn/layer/vulkan/unaryop_vulkan.h |     42 -
 .../sherpa-ncnn/layer/x86/avx512_mathfun.h    |    859 -
 .../sherpa-ncnn/layer/x86/avx_mathfun.h       |   1090 -
 .../sherpa-ncnn/layer/x86/batchnorm_x86.h     |     32 -
 .../sherpa/sherpa-ncnn/layer/x86/bias_x86.h   |     30 -
 .../sherpa-ncnn/layer/x86/binaryop_x86.h      |     34 -
 .../sherpa/sherpa-ncnn/layer/x86/bnll_x86.h   |     33 -
 .../sherpa/sherpa-ncnn/layer/x86/cast_bf16.h  |    159 -
 .../sherpa/sherpa-ncnn/layer/x86/cast_fp16.h  |    138 -
 .../sherpa/sherpa-ncnn/layer/x86/cast_x86.h   |     32 -
 .../sherpa/sherpa-ncnn/layer/x86/clip_x86.h   |     32 -
 .../sherpa/sherpa-ncnn/layer/x86/concat_x86.h |     32 -
 .../layer/x86/convolution1d_packed.h          |   2819 -
 .../sherpa-ncnn/layer/x86/convolution1d_x86.h |     40 -
 .../sherpa-ncnn/layer/x86/convolution_1x1.h   |    207 -
 .../layer/x86/convolution_2x2_pack8.h         |    384 -
 .../sherpa-ncnn/layer/x86/convolution_3x3.h   |    209 -
 .../layer/x86/convolution_3x3_int8.h          |    147 -
 .../layer/x86/convolution_3x3_pack16to1.h     |    100 -
 .../layer/x86/convolution_3x3_pack1to4.h      |   1283 -
 .../layer/x86/convolution_3x3_pack1to8.h      |   1555 -
 .../layer/x86/convolution_3x3_pack8.h         |    718 -
 .../layer/x86/convolution_3x3_pack8to1.h      |     98 -
 .../layer/x86/convolution_3x3_winograd.h      |   7380 -
 .../layer/x86/convolution_3x3_winograd_int8.h |   6407 -
 .../sherpa-ncnn/layer/x86/convolution_5x5.h   |    209 -
 .../layer/x86/convolution_im2col_gemm_int8.h  |   8895 -
 .../layer/x86/convolution_packed.h            |   2881 -
 .../layer/x86/convolution_packed_int8.h       |   5386 -
 .../sherpa-ncnn/layer/x86/convolution_x86.h   |     63 -
 .../layer/x86/convolutiondepthwise_3x3.h      |    199 -
 .../layer/x86/convolutiondepthwise_3x3_int8.h |    405 -
 .../x86/convolutiondepthwise_3x3_pack16.h     |    733 -
 .../x86/convolutiondepthwise_3x3_pack4.h      |    581 -
 .../x86/convolutiondepthwise_3x3_pack8.h      |    795 -
 .../x86/convolutiondepthwise_5x5_pack16.h     |    317 -
 .../x86/convolutiondepthwise_5x5_pack4.h      |    486 -
 .../x86/convolutiondepthwise_5x5_pack8.h      |    317 -
 .../layer/x86/convolutiondepthwise_x86.h      |     50 -
 .../sherpa/sherpa-ncnn/layer/x86/crop_x86.h   |     34 -
 .../layer/x86/deconvolution_pack16.h          |    142 -
 .../layer/x86/deconvolution_pack16to1.h       |    101 -
 .../layer/x86/deconvolution_pack16to4.h       |    142 -
 .../layer/x86/deconvolution_pack16to8.h       |    142 -
 .../layer/x86/deconvolution_pack1to16.h       |     99 -
 .../layer/x86/deconvolution_pack1to4.h        |     99 -
 .../layer/x86/deconvolution_pack1to8.h        |     99 -
 .../layer/x86/deconvolution_pack4.h           |    106 -
 .../layer/x86/deconvolution_pack4to1.h        |    101 -
 .../layer/x86/deconvolution_pack4to16.h       |    106 -
 .../layer/x86/deconvolution_pack4to8.h        |    106 -
 .../layer/x86/deconvolution_pack8.h           |    118 -
 .../layer/x86/deconvolution_pack8to1.h        |    101 -
 .../layer/x86/deconvolution_pack8to16.h       |    118 -
 .../layer/x86/deconvolution_pack8to4.h        |    118 -
 .../sherpa-ncnn/layer/x86/deconvolution_x86.h |     43 -
 .../layer/x86/deconvolutiondepthwise_x86.h    |     45 -
 .../layer/x86/deformableconv2d_pack16.h       |    435 -
 .../layer/x86/deformableconv2d_pack16to1.h    |    370 -
 .../layer/x86/deformableconv2d_pack16to4.h    |    435 -
 .../layer/x86/deformableconv2d_pack16to8.h    |    435 -
 .../layer/x86/deformableconv2d_pack1to16.h    |    195 -
 .../layer/x86/deformableconv2d_pack1to4.h     |    195 -
 .../layer/x86/deformableconv2d_pack1to8.h     |    195 -
 .../layer/x86/deformableconv2d_pack4.h        |    243 -
 .../layer/x86/deformableconv2d_pack4to1.h     |    211 -
 .../layer/x86/deformableconv2d_pack4to16.h    |    243 -
 .../layer/x86/deformableconv2d_pack4to8.h     |    243 -
 .../layer/x86/deformableconv2d_pack8.h        |    307 -
 .../layer/x86/deformableconv2d_pack8to1.h     |    259 -
 .../layer/x86/deformableconv2d_pack8to16.h    |    307 -
 .../layer/x86/deformableconv2d_pack8to4.h     |    307 -
 .../layer/x86/deformableconv2d_x86.h          |     42 -
 .../sherpa-ncnn/layer/x86/dequantize_x86.h    |     32 -
 .../sherpa-ncnn/layer/x86/dropout_x86.h       |     32 -
 .../sherpa-ncnn/layer/x86/eltwise_x86.h       |     32 -
 .../sherpa/sherpa-ncnn/layer/x86/elu_x86.h    |     32 -
 .../sherpa-ncnn/layer/x86/flatten_x86.h       |     35 -
 .../sherpa/sherpa-ncnn/layer/x86/gelu_x86.h   |     33 -
 .../sherpa/sherpa-ncnn/layer/x86/gemm_x86.h   |     40 -
 .../gridsample_bicubic_apply_interpolation.h  |    288 -
 .../x86/gridsample_bicubic_compute_blob.h     |    299 -
 .../gridsample_bilinear_apply_interpolation.h |    370 -
 .../x86/gridsample_bilinear_compute_blob.h    |    623 -
 .../layer/x86/gridsample_compute_blob.h       |    145 -
 .../gridsample_nearest_apply_interpolation.h  |    126 -
 .../x86/gridsample_nearest_compute_blob.h     |    315 -
 .../sherpa-ncnn/layer/x86/gridsample_x86.h    |     32 -
 .../sherpa-ncnn/layer/x86/groupnorm_x86.h     |     32 -
 .../sherpa-ncnn/layer/x86/hardsigmoid_x86.h   |     32 -
 .../sherpa-ncnn/layer/x86/hardswish_x86.h     |     32 -
 .../sherpa-ncnn/layer/x86/innerproduct_fp.h   |   1401 -
 .../layer/x86/innerproduct_gemm_fp.h          |   1285 -
 .../sherpa-ncnn/layer/x86/innerproduct_x86.h  |     54 -
 .../sherpa-ncnn/layer/x86/interp_bicubic.h    |    310 -
 .../layer/x86/interp_bicubic_pack16.h         |    286 -
 .../layer/x86/interp_bicubic_pack4.h          |    286 -
 .../layer/x86/interp_bicubic_pack8.h          |    286 -
 .../sherpa-ncnn/layer/x86/interp_bilinear.h   |    171 -
 .../layer/x86/interp_bilinear_pack16.h        |    123 -
 .../layer/x86/interp_bilinear_pack4.h         |    123 -
 .../layer/x86/interp_bilinear_pack8.h         |    123 -
 .../sherpa/sherpa-ncnn/layer/x86/interp_x86.h |     32 -
 .../sherpa-ncnn/layer/x86/layernorm_x86.h     |     32 -
 .../sherpa/sherpa-ncnn/layer/x86/lrn_x86.h    |     30 -
 .../sherpa/sherpa-ncnn/layer/x86/lstm_x86.h   |     41 -
 .../sherpa/sherpa-ncnn/layer/x86/matmul_x86.h |     38 -
 .../sherpa/sherpa-ncnn/layer/x86/mish_x86.h   |     32 -
 .../layer/x86/multiheadattention_x86.h        |     46 -
 .../sherpa-ncnn/layer/x86/packing_x86.h       |     35 -
 .../sherpa-ncnn/layer/x86/padding_pack16.h    |    212 -
 .../sherpa-ncnn/layer/x86/padding_pack4.h     |    212 -
 .../sherpa-ncnn/layer/x86/padding_pack8.h     |    212 -
 .../layer/x86/padding_pack8_int8.h            |    171 -
 .../sherpa-ncnn/layer/x86/padding_x86.h       |     35 -
 .../sherpa-ncnn/layer/x86/pooling_2x2.h       |     83 -
 .../layer/x86/pooling_2x2_pack16.h            |     60 -
 .../sherpa-ncnn/layer/x86/pooling_2x2_pack4.h |     61 -
 .../sherpa-ncnn/layer/x86/pooling_2x2_pack8.h |     61 -
 .../layer/x86/pooling_3x3_pack16.h            |    193 -
 .../sherpa-ncnn/layer/x86/pooling_3x3_pack4.h |    116 -
 .../sherpa-ncnn/layer/x86/pooling_3x3_pack8.h |    194 -
 .../sherpa-ncnn/layer/x86/pooling_x86.h       |     37 -
 .../sherpa/sherpa-ncnn/layer/x86/prelu_x86.h  |     32 -
 .../sherpa-ncnn/layer/x86/quantize_x86.h      |     32 -
 .../sherpa/sherpa-ncnn/layer/x86/relu_x86.h   |     35 -
 .../sherpa-ncnn/layer/x86/requantize_x86.h    |     32 -
 .../sherpa-ncnn/layer/x86/reshape_x86.h       |     32 -
 .../sherpa-ncnn/layer/x86/roialign_x86.h      |     32 -
 .../sherpa/sherpa-ncnn/layer/x86/scale_x86.h  |     32 -
 .../sherpa/sherpa-ncnn/layer/x86/selu_x86.h   |     32 -
 .../layer/x86/shufflechannel_x86.h            |     32 -
 .../sherpa-ncnn/layer/x86/sigmoid_x86.h       |     32 -
 .../sherpa/sherpa-ncnn/layer/x86/slice_x86.h  |     32 -
 .../sherpa-ncnn/layer/x86/softmax_x86.h       |     32 -
 .../sherpa-ncnn/layer/x86/sse_mathfun.h       |   1160 -
 .../sherpa/sherpa-ncnn/layer/x86/swish_x86.h  |     32 -
 .../sherpa/sherpa-ncnn/layer/x86/tanh_x86.h   |     32 -
 .../sherpa-ncnn/layer/x86/unaryop_x86.h       |     32 -
 .../sherpa-ncnn/layer/x86/x86_activation.h    |    327 -
 .../sherpa-ncnn/layer/x86/x86_usability.h     |   1344 -
 .../layer/x86/yolov3detectionoutput_x86.h     |     32 -
 .../sherpa-ncnn/layer/yolodetectionoutput.h   |     46 -
 .../sherpa-ncnn/layer/yolov3detectionoutput.h |     62 -
 .../sherpa/sherpa-ncnn/layer_declaration.h    |    125 -
 .../sherpa/sherpa-ncnn/layer_registry.h       |   1619 -
 .../sherpa-ncnn/layer_shader_registry.h       |      5 -
 .../sherpa-ncnn/layer_shader_spv_data.h       |      5 -
 .../sherpa/sherpa-ncnn/layer_shader_type.h    |     29 -
 .../sherpa-ncnn/layer_shader_type_enum.h      |      5 -
 .../include/sherpa/sherpa-ncnn/layer_type.h   |     30 -
 .../sherpa/sherpa-ncnn/layer_type_enum.h      |    109 -
 .../include/sherpa/sherpa-ncnn/mat.h          |   1847 -
 .../sherpa-ncnn/mat_pixel_drawing_font.h      |   3870 -
 .../include/sherpa/sherpa-ncnn/modelbin.h     |     80 -
 .../include/sherpa/sherpa-ncnn/ncnn_export.h  |     42 -
 .../include/sherpa/sherpa-ncnn/net.h          |    275 -
 .../include/sherpa/sherpa-ncnn/option.h       |    158 -
 .../include/sherpa/sherpa-ncnn/paramdict.h    |     73 -
 .../include/sherpa/sherpa-ncnn/pipeline.h     |    111 -
 .../sherpa/sherpa-ncnn/pipelinecache.h        |     81 -
 .../include/sherpa/sherpa-ncnn/platform.h     |    325 -
 .../include/sherpa/sherpa-ncnn/ruapu.h        |    302 -
 .../include/sherpa/sherpa-ncnn/simplemath.h   |    103 -
 .../include/sherpa/sherpa-ncnn/simpleocv.h    |    503 -
 .../include/sherpa/sherpa-ncnn/simpleomp.h    |     53 -
 .../include/sherpa/sherpa-ncnn/simplestl.h    |    565 -
 .../include/sherpa/sherpa-ncnn/simplevk.h     |   2207 -
 .../include/sherpa/sherpa-ncnn/stb_image.h    |   8362 -
 .../sherpa/sherpa-ncnn/stb_image_write.h      |   1724 -
 .../sherpa/sherpa-ncnn/vulkan_header_fix.h    |   1222 -
 .../sherpa/sherpa-onnx/csrc/alsa-play.h       |     37 -
 .../include/sherpa/sherpa-onnx/csrc/alsa.h    |     46 -
 .../sherpa-onnx/csrc/audio-tagging-ced-impl.h |    111 -
 .../sherpa-onnx/csrc/audio-tagging-impl.h     |     39 -
 .../csrc/audio-tagging-label-file.h           |     39 -
 .../csrc/audio-tagging-model-config.h         |     42 -
 .../csrc/audio-tagging-zipformer-impl.h       |    118 -
 .../sherpa/sherpa-onnx/csrc/audio-tagging.h   |     74 -
 .../sherpa/sherpa-onnx/csrc/base64-decode.h   |     19 -
 .../include/sherpa/sherpa-onnx/csrc/cat.h     |     28 -
 .../sherpa/sherpa-onnx/csrc/circular-buffer.h |     61 -
 .../sherpa/sherpa-onnx/csrc/context-graph.h   |     92 -
 .../include/sherpa/sherpa-onnx/csrc/display.h |     88 -
 .../sherpa/sherpa-onnx/csrc/endpoint.h        |     76 -
 .../sherpa-onnx/csrc/fast-clustering-config.h |     39 -
 .../sherpa/sherpa-onnx/csrc/fast-clustering.h |     43 -
 .../sherpa/sherpa-onnx/csrc/features.h        |    137 -
 .../sherpa/sherpa-onnx/csrc/file-utils.h      |     28 -
 .../sherpa/sherpa-onnx/csrc/fst-utils.h       |     18 -
 .../sherpa/sherpa-onnx/csrc/hypothesis.h      |    166 -
 .../sherpa/sherpa-onnx/csrc/jieba-lexicon.h   |     36 -
 .../sherpa-onnx/csrc/keyword-spotter-impl.h   |     48 -
 .../csrc/keyword-spotter-transducer-impl.h    |    350 -
 .../sherpa/sherpa-onnx/csrc/keyword-spotter.h |    153 -
 .../include/sherpa/sherpa-onnx/csrc/lexicon.h |     66 -
 .../include/sherpa/sherpa-onnx/csrc/log.h     |    378 -
 .../include/sherpa/sherpa-onnx/csrc/macros.h  |    188 -
 .../include/sherpa/sherpa-onnx/csrc/math.h    |    123 -
 .../sherpa-onnx/csrc/melo-tts-lexicon.h       |     49 -
 .../sherpa/sherpa-onnx/csrc/microphone.h      |     18 -
 .../sherpa-onnx/csrc/offline-ced-model.h      |     56 -
 .../offline-ct-transformer-model-meta-data.h  |     29 -
 .../csrc/offline-ct-transformer-model.h       |     59 -
 .../sherpa-onnx/csrc/offline-ctc-decoder.h    |     50 -
 .../csrc/offline-ctc-fst-decoder-config.h     |     32 -
 .../csrc/offline-ctc-fst-decoder.h            |     33 -
 .../csrc/offline-ctc-greedy-search-decoder.h  |     28 -
 .../sherpa-onnx/csrc/offline-ctc-model.h      |     71 -
 .../sherpa-onnx/csrc/offline-lm-config.h      |     39 -
 .../sherpa/sherpa-onnx/csrc/offline-lm.h      |     50 -
 .../sherpa-onnx/csrc/offline-model-config.h   |     93 -
 .../csrc/offline-moonshine-decoder.h          |     34 -
 .../offline-moonshine-greedy-search-decoder.h |     29 -
 .../csrc/offline-moonshine-model-config.h     |     37 -
 .../csrc/offline-moonshine-model.h            |     87 -
 .../offline-nemo-enc-dec-ctc-model-config.h   |     28 -
 .../csrc/offline-nemo-enc-dec-ctc-model.h     |     83 -
 .../csrc/offline-paraformer-decoder.h         |     42 -
 ...offline-paraformer-greedy-search-decoder.h |     29 -
 .../csrc/offline-paraformer-model-config.h    |     28 -
 .../csrc/offline-paraformer-model.h           |     74 -
 .../offline-punctuation-ct-transformer-impl.h |    197 -
 .../csrc/offline-punctuation-impl.h           |     36 -
 .../csrc/offline-punctuation-model-config.h   |     38 -
 .../sherpa-onnx/csrc/offline-punctuation.h    |     57 -
 .../csrc/offline-recognizer-ctc-impl.h        |    269 -
 .../csrc/offline-recognizer-impl.h            |     61 -
 .../csrc/offline-recognizer-moonshine-impl.h  |    143 -
 .../csrc/offline-recognizer-paraformer-impl.h |    267 -
 .../offline-recognizer-sense-voice-impl.h     |    364 -
 .../csrc/offline-recognizer-transducer-impl.h |    304 -
 .../offline-recognizer-transducer-nemo-impl.h |    189 -
 .../csrc/offline-recognizer-whisper-impl.h    |    181 -
 .../sherpa-onnx/csrc/offline-recognizer.h     |    131 -
 .../sherpa/sherpa-onnx/csrc/offline-rnn-lm.h  |     44 -
 .../csrc/offline-sense-voice-model-config.h   |     39 -
 .../offline-sense-voice-model-meta-data.h     |     50 -
 .../csrc/offline-sense-voice-model.h          |     55 -
 .../csrc/offline-speaker-diarization-impl.h   |     39 -
 ...ffline-speaker-diarization-pyannote-impl.h |    743 -
 .../csrc/offline-speaker-diarization-result.h |     67 -
 .../csrc/offline-speaker-diarization.h        |     84 -
 ...ffline-speaker-segmentation-model-config.h |     40 -
 ...eaker-segmentation-pyannote-model-config.h |     30 -
 ...er-segmentation-pyannote-model-meta-data.h |     29 -
 ...line-speaker-segmentation-pyannote-model.h |     44 -
 .../sherpa/sherpa-onnx/csrc/offline-stream.h  |    106 -
 .../sherpa-onnx/csrc/offline-tdnn-ctc-model.h |     59 -
 .../csrc/offline-tdnn-model-config.h          |     28 -
 .../csrc/offline-telespeech-ctc-model.h       |     74 -
 .../csrc/offline-transducer-decoder.h         |     43 -
 ...offline-transducer-greedy-search-decoder.h |     34 -
 ...ne-transducer-greedy-search-nemo-decoder.h |     33 -
 .../csrc/offline-transducer-model-config.h    |     34 -
 .../csrc/offline-transducer-model.h           |    104 -
 ...-transducer-modified-beam-search-decoder.h |     47 -
 .../csrc/offline-transducer-nemo-model.h      |     98 -
 .../csrc/offline-tts-character-frontend.h     |     48 -
 .../sherpa-onnx/csrc/offline-tts-frontend.h   |     56 -
 .../sherpa-onnx/csrc/offline-tts-impl.h       |     39 -
 .../csrc/offline-tts-model-config.h           |     40 -
 .../sherpa-onnx/csrc/offline-tts-vits-impl.h  |    493 -
 .../csrc/offline-tts-vits-model-config.h      |     59 -
 .../csrc/offline-tts-vits-model-metadata.h    |     49 -
 .../sherpa-onnx/csrc/offline-tts-vits-model.h |     51 -
 .../sherpa/sherpa-onnx/csrc/offline-tts.h     |    100 -
 .../csrc/offline-websocket-server-impl.h      |    205 -
 .../csrc/offline-wenet-ctc-model-config.h     |     28 -
 .../csrc/offline-wenet-ctc-model.h            |     73 -
 .../csrc/offline-whisper-decoder.h            |     43 -
 .../offline-whisper-greedy-search-decoder.h   |     33 -
 .../csrc/offline-whisper-model-config.h       |     60 -
 .../sherpa-onnx/csrc/offline-whisper-model.h  |    115 -
 ...ine-zipformer-audio-tagging-model-config.h |     29 -
 .../offline-zipformer-audio-tagging-model.h   |     64 -
 .../csrc/offline-zipformer-ctc-model-config.h |     32 -
 .../csrc/offline-zipformer-ctc-model.h        |     62 -
 .../csrc/online-cnn-bilstm-model-meta-data.h  |     25 -
 .../csrc/online-cnn-bilstm-model.h            |     62 -
 .../csrc/online-conformer-transducer-model.h  |    100 -
 .../sherpa-onnx/csrc/online-ctc-decoder.h     |     63 -
 .../csrc/online-ctc-fst-decoder-config.h      |     32 -
 .../sherpa-onnx/csrc/online-ctc-fst-decoder.h |     39 -
 .../csrc/online-ctc-greedy-search-decoder.h   |     29 -
 .../sherpa-onnx/csrc/online-ctc-model.h       |     84 -
 .../sherpa-onnx/csrc/online-lm-config.h       |     42 -
 .../sherpa/sherpa-onnx/csrc/online-lm.h       |     63 -
 .../csrc/online-lstm-transducer-model.h       |     95 -
 .../sherpa-onnx/csrc/online-model-config.h    |     86 -
 .../csrc/online-nemo-ctc-model-config.h       |     28 -
 .../sherpa-onnx/csrc/online-nemo-ctc-model.h  |     75 -
 .../csrc/online-paraformer-decoder.h          |     23 -
 .../csrc/online-paraformer-model-config.h     |     31 -
 .../csrc/online-paraformer-model.h            |     70 -
 .../csrc/online-punctuation-cnn-bilstm-impl.h |    278 -
 .../csrc/online-punctuation-impl.h            |     37 -
 .../csrc/online-punctuation-model-config.h    |     42 -
 .../sherpa-onnx/csrc/online-punctuation.h     |     57 -
 .../csrc/online-recognizer-ctc-impl.h         |    316 -
 .../sherpa-onnx/csrc/online-recognizer-impl.h |     71 -
 .../csrc/online-recognizer-paraformer-impl.h  |    475 -
 .../csrc/online-recognizer-transducer-impl.h  |    495 -
 .../online-recognizer-transducer-nemo-impl.h  |    251 -
 .../sherpa-onnx/csrc/online-recognizer.h      |    211 -
 .../sherpa/sherpa-onnx/csrc/online-rnn-lm.h   |     68 -
 .../sherpa/sherpa-onnx/csrc/online-stream.h   |    120 -
 .../csrc/online-transducer-decoder.h          |    111 -
 .../online-transducer-greedy-search-decoder.h |     42 -
 ...ne-transducer-greedy-search-nemo-decoder.h |     34 -
 .../csrc/online-transducer-model-config.h     |     32 -
 .../csrc/online-transducer-model.h            |    145 -
 ...-transducer-modified-beam-search-decoder.h |     64 -
 .../csrc/online-transducer-nemo-model.h       |    124 -
 .../csrc/online-websocket-server-impl.h       |    181 -
 .../csrc/online-wenet-ctc-model-config.h      |     38 -
 .../sherpa-onnx/csrc/online-wenet-ctc-model.h |     75 -
 .../csrc/online-zipformer-transducer-model.h  |     99 -
 .../csrc/online-zipformer2-ctc-model-config.h |     29 -
 .../csrc/online-zipformer2-ctc-model.h        |     74 -
 .../csrc/online-zipformer2-transducer-model.h |    110 -
 .../sherpa/sherpa-onnx/csrc/onnx-utils.h      |    142 -
 .../sherpa/sherpa-onnx/csrc/packed-sequence.h |     52 -
 .../sherpa/sherpa-onnx/csrc/pad-sequence.h    |     31 -
 .../sherpa/sherpa-onnx/csrc/parse-options.h   |    259 -
 .../csrc/piper-phonemize-lexicon.h            |     38 -
 .../sherpa/sherpa-onnx/csrc/provider-config.h |     95 -
 .../sherpa/sherpa-onnx/csrc/provider.h        |     36 -
 .../sherpa/sherpa-onnx/csrc/resample.h        |    144 -
 .../include/sherpa/sherpa-onnx/csrc/session.h |     36 -
 .../csrc/silero-vad-model-config.h            |     46 -
 .../sherpa-onnx/csrc/silero-vad-model.h       |     55 -
 .../include/sherpa/sherpa-onnx/csrc/slice.h   |     48 -
 ...speaker-embedding-extractor-general-impl.h |    117 -
 .../csrc/speaker-embedding-extractor-impl.h   |     38 -
 ...aker-embedding-extractor-model-meta-data.h |     28 -
 .../csrc/speaker-embedding-extractor-model.h  |     41 -
 .../speaker-embedding-extractor-nemo-impl.h   |    145 -
 ...embedding-extractor-nemo-model-meta-data.h |     28 -
 .../speaker-embedding-extractor-nemo-model.h  |     44 -
 .../csrc/speaker-embedding-extractor.h        |     71 -
 .../csrc/speaker-embedding-manager.h          |    120 -
 .../spoken-language-identification-impl.h     |     38 -
 ...ken-language-identification-whisper-impl.h |    133 -
 .../csrc/spoken-language-identification.h     |     99 -
 .../include/sherpa/sherpa-onnx/csrc/stack.h   |     29 -
 .../sherpa/sherpa-onnx/csrc/symbol-table.h    |     71 -
 .../sherpa/sherpa-onnx/csrc/tee-stream.h      |     61 -
 .../sherpa/sherpa-onnx/csrc/text-utils.h      |    129 -
 .../csrc/transducer-keyword-decoder.h         |     62 -
 .../sherpa/sherpa-onnx/csrc/transpose.h       |     32 -
 .../include/sherpa/sherpa-onnx/csrc/unbind.h  |     28 -
 .../include/sherpa/sherpa-onnx/csrc/utils.h   |     62 -
 .../sherpa-onnx/csrc/vad-model-config.h       |     42 -
 .../sherpa/sherpa-onnx/csrc/vad-model.h       |     47 -
 .../csrc/voice-activity-detector.h            |     53 -
 .../sherpa/sherpa-onnx/csrc/wave-reader.h     |     31 -
 .../sherpa/sherpa-onnx/csrc/wave-writer.h     |     27 -
 .../kaldi-decoder/csrc/decodable-ctc.h        |     47 -
 .../kaldi-decoder/csrc/decodable-itf.h        |    106 -
 .../sherpa-onnx/kaldi-decoder/csrc/eigen.h    |     39 -
 .../kaldi-decoder/csrc/faster-decoder.h       |    204 -
 .../kaldi-decoder/csrc/hash-list-inl.h        |    207 -
 .../kaldi-decoder/csrc/hash-list.h            |    133 -
 .../kaldi-decoder/csrc/kaldi-math.h           |     48 -
 .../csrc/lattice-faster-decoder.h             |    274 -
 .../csrc/lattice-simple-decoder.h             |    324 -
 .../sherpa-onnx/kaldi-decoder/csrc/log.h      |    100 -
 .../kaldi-decoder/csrc/simple-decoder.h       |    138 -
 .../kaldi-decoder/csrc/stl-utils.h            |    209 -
 .../sherpa-onnx/kaldifst/csrc/basic-filebuf.h |    999 -
 .../kaldifst/csrc/const-integer-set-inl.h     |     76 -
 .../kaldifst/csrc/const-integer-set.h         |     85 -
 .../sherpa-onnx/kaldifst/csrc/context-fst.h   |    274 -
 .../kaldifst/csrc/deterministic-fst-inl.h     |    276 -
 .../kaldifst/csrc/deterministic-fst.h         |    121 -
 .../kaldifst/csrc/determinize-star-inl.h      |   1116 -
 .../kaldifst/csrc/determinize-star.h          |     47 -
 .../kaldifst/csrc/fstext-utils-inl.h          |    612 -
 .../sherpa-onnx/kaldifst/csrc/fstext-utils.h  |    154 -
 .../kaldifst/csrc/grammar-context-fst.h       |     45 -
 .../sherpa-onnx/kaldifst/csrc/io-funcs-inl.h  |     54 -
 .../sherpa-onnx/kaldifst/csrc/io-funcs.h      |     27 -
 .../kaldifst/csrc/kaldi-fst-io-inl.h          |    202 -
 .../sherpa-onnx/kaldifst/csrc/kaldi-fst-io.h  |    150 -
 .../kaldifst/csrc/kaldi-holder-inl.h          |     76 -
 .../sherpa-onnx/kaldifst/csrc/kaldi-holder.h  |     39 -
 .../sherpa-onnx/kaldifst/csrc/kaldi-io-inl.h  |     40 -
 .../sherpa-onnx/kaldifst/csrc/kaldi-io.h      |    227 -
 .../sherpa-onnx/kaldifst/csrc/kaldi-math.h    |     29 -
 .../sherpa-onnx/kaldifst/csrc/kaldi-pipebuf.h |     88 -
 .../kaldifst/csrc/kaldi-semaphore.h           |     51 -
 .../kaldifst/csrc/kaldi-table-inl.h           |   2703 -
 .../sherpa-onnx/kaldifst/csrc/kaldi-table.h   |    451 -
 .../kaldifst/csrc/lattice-utils-inl.h         |     42 -
 .../sherpa-onnx/kaldifst/csrc/lattice-utils.h |     50 -
 .../kaldifst/csrc/lattice-weight.h            |    898 -
 .../sherpa/sherpa-onnx/kaldifst/csrc/log.h    |    100 -
 .../sherpa-onnx/kaldifst/csrc/parse-options.h |     26 -
 .../kaldifst/csrc/pre-determinize-inl.h       |     94 -
 .../kaldifst/csrc/pre-determinize.h           |     44 -
 .../kaldifst/csrc/remove-eps-local-inl.h      |    306 -
 .../kaldifst/csrc/remove-eps-local.h          |     45 -
 .../sherpa-onnx/kaldifst/csrc/stl-utils.h     |    105 -
 .../sherpa-onnx/kaldifst/csrc/table-matcher.h |    381 -
 .../kaldifst/csrc/text-normalizer.h           |     36 -
 .../sherpa-onnx/kaldifst/csrc/text-utils.h    |    127 -
 .../onnxruntime-src/cpu_provider_factory.h    |     19 -
 .../onnxruntime-src/onnxruntime_c_api.h       |   4717 -
 .../onnxruntime-src/onnxruntime_cxx_api.h     |   2372 -
 .../onnxruntime-src/onnxruntime_cxx_inline.h  |   2075 -
 .../onnxruntime-src/onnxruntime_float16.h     |    540 -
 .../onnxruntime_run_options_config_keys.h     |     32 -
 .../onnxruntime_session_options_config_keys.h |    258 -
 .../onnxruntime-src/provider_options.h        |     18 -
 .../sherpa-onnx/openfst-src/fst/accumulator.h |    894 -
 .../sherpa-onnx/openfst-src/fst/add-on.h      |    248 -
 .../sherpa-onnx/openfst-src/fst/arc-arena.h   |    232 -
 .../sherpa-onnx/openfst-src/fst/arc-map.h     |   1275 -
 .../sherpa/sherpa-onnx/openfst-src/fst/arc.h  |    310 -
 .../sherpa-onnx/openfst-src/fst/arcfilter.h   |     93 -
 .../sherpa-onnx/openfst-src/fst/arcsort.h     |    211 -
 .../sherpa-onnx/openfst-src/fst/bi-table.h    |    473 -
 .../sherpa-onnx/openfst-src/fst/cache.h       |   1277 -
 .../sherpa-onnx/openfst-src/fst/closure.h     |    136 -
 .../sherpa-onnx/openfst-src/fst/compact-fst.h |   1566 -
 .../sherpa-onnx/openfst-src/fst/compat.h      |    130 -
 .../sherpa-onnx/openfst-src/fst/complement.h  |    277 -
 .../openfst-src/fst/compose-filter.h          |    508 -
 .../sherpa-onnx/openfst-src/fst/compose.h     |   1013 -
 .../sherpa-onnx/openfst-src/fst/concat.h      |    223 -
 .../sherpa-onnx/openfst-src/fst/config.h      |      3 -
 .../sherpa-onnx/openfst-src/fst/connect.h     |    314 -
 .../sherpa-onnx/openfst-src/fst/const-fst.h   |    488 -
 .../sherpa-onnx/openfst-src/fst/determinize.h |   1086 -
 .../sherpa-onnx/openfst-src/fst/dfs-visit.h   |    204 -
 .../sherpa-onnx/openfst-src/fst/difference.h  |    177 -
 .../openfst-src/fst/disambiguate.h            |    564 -
 .../sherpa-onnx/openfst-src/fst/edit-fst.h    |    702 -
 .../sherpa-onnx/openfst-src/fst/encode.h      |    556 -
 .../openfst-src/fst/epsnormalize.h            |     58 -
 .../sherpa-onnx/openfst-src/fst/equal.h       |    139 -
 .../sherpa-onnx/openfst-src/fst/equivalent.h  |    230 -
 .../openfst-src/fst/expanded-fst.h            |    183 -
 .../openfst-src/fst/expectation-weight.h      |    137 -
 .../fst/extensions/compress/compress-script.h |     54 -
 .../fst/extensions/compress/compress.h        |    904 -
 .../fst/extensions/compress/elias.h           |     97 -
 .../fst/extensions/compress/gzfile.h          |    126 -
 .../fst/extensions/compress/randmod.h         |    101 -
 .../fst/extensions/far/compile-strings.h      |    260 -
 .../openfst-src/fst/extensions/far/create.h   |     46 -
 .../openfst-src/fst/extensions/far/equal.h    |     69 -
 .../openfst-src/fst/extensions/far/extract.h  |    117 -
 .../fst/extensions/far/far-class.h            |    270 -
 .../openfst-src/fst/extensions/far/far.h      |    520 -
 .../openfst-src/fst/extensions/far/farlib.h   |     19 -
 .../fst/extensions/far/farscript.h            |    269 -
 .../openfst-src/fst/extensions/far/getters.h  |     30 -
 .../openfst-src/fst/extensions/far/info.h     |    147 -
 .../fst/extensions/far/isomorphic.h           |     69 -
 .../fst/extensions/far/print-strings.h        |    105 -
 .../fst/extensions/far/script-impl.h          |     23 -
 .../openfst-src/fst/extensions/far/stlist.h   |    267 -
 .../openfst-src/fst/extensions/far/sttable.h  |    414 -
 .../linear/linear-fst-data-builder.h          |   1074 -
 .../fst/extensions/linear/linear-fst-data.h   |    526 -
 .../fst/extensions/linear/linear-fst.h        |   1147 -
 .../fst/extensions/linear/linearscript.h      |    391 -
 .../fst/extensions/linear/loglinear-apply.h   |     77 -
 .../openfst-src/fst/extensions/linear/trie.h  |    444 -
 .../openfst-src/fst/extensions/mpdt/compose.h |    267 -
 .../openfst-src/fst/extensions/mpdt/expand.h  |    335 -
 .../openfst-src/fst/extensions/mpdt/info.h    |    190 -
 .../openfst-src/fst/extensions/mpdt/mpdt.h    |    357 -
 .../openfst-src/fst/extensions/mpdt/mpdtlib.h |     18 -
 .../fst/extensions/mpdt/mpdtscript.h          |    156 -
 .../fst/extensions/mpdt/read_write_utils.h    |     86 -
 .../openfst-src/fst/extensions/mpdt/reverse.h |     54 -
 .../fst/extensions/ngram/bitmap-index.h       |    168 -
 .../fst/extensions/ngram/ngram-fst.h          |   1010 -
 .../openfst-src/fst/extensions/ngram/nthbit.h |     33 -
 .../fst/extensions/pdt/collection.h           |    107 -
 .../openfst-src/fst/extensions/pdt/compose.h  |    476 -
 .../openfst-src/fst/extensions/pdt/expand.h   |    933 -
 .../openfst-src/fst/extensions/pdt/getters.h  |     22 -
 .../openfst-src/fst/extensions/pdt/info.h     |    152 -
 .../openfst-src/fst/extensions/pdt/paren.h    |    440 -
 .../openfst-src/fst/extensions/pdt/pdt.h      |    165 -
 .../openfst-src/fst/extensions/pdt/pdtlib.h   |     19 -
 .../fst/extensions/pdt/pdtscript.h            |    244 -
 .../openfst-src/fst/extensions/pdt/replace.h  |    827 -
 .../openfst-src/fst/extensions/pdt/reverse.h  |     38 -
 .../fst/extensions/pdt/shortest-path.h        |    715 -
 .../fst/extensions/special/phi-fst.h          |    171 -
 .../fst/extensions/special/rho-fst.h          |    160 -
 .../fst/extensions/special/sigma-fst.h        |    163 -
 .../openfst-src/fst/factor-weight.h           |    496 -
 .../openfst-src/fst/filter-state.h            |    199 -
 .../sherpa-onnx/openfst-src/fst/flags.h       |    243 -
 .../openfst-src/fst/float-weight.h            |    780 -
 .../sherpa-onnx/openfst-src/fst/fst-decl.h    |    251 -
 .../sherpa/sherpa-onnx/openfst-src/fst/fst.h  |    960 -
 .../sherpa-onnx/openfst-src/fst/fstlib.h      |    130 -
 .../openfst-src/fst/generic-register.h        |    128 -
 .../sherpa/sherpa-onnx/openfst-src/fst/heap.h |    166 -
 .../sherpa/sherpa-onnx/openfst-src/fst/icu.h  |    108 -
 .../sherpa-onnx/openfst-src/fst/intersect.h   |    153 -
 .../openfst-src/fst/interval-set.h            |    398 -
 .../sherpa-onnx/openfst-src/fst/invert.h      |    119 -
 .../sherpa-onnx/openfst-src/fst/isomorphic.h  |    178 -
 .../openfst-src/fst/label-reachable.h         |    511 -
 .../openfst-src/fst/lexicographic-weight.h    |    173 -
 .../sherpa/sherpa-onnx/openfst-src/fst/lock.h |     62 -
 .../sherpa/sherpa-onnx/openfst-src/fst/log.h  |     82 -
 .../openfst-src/fst/lookahead-filter.h        |    623 -
 .../openfst-src/fst/lookahead-matcher.h       |    786 -
 .../sherpa/sherpa-onnx/openfst-src/fst/map.h  |    110 -
 .../sherpa-onnx/openfst-src/fst/mapped-file.h |     75 -
 .../sherpa-onnx/openfst-src/fst/matcher-fst.h |    347 -
 .../sherpa-onnx/openfst-src/fst/matcher.h     |   1458 -
 .../sherpa-onnx/openfst-src/fst/memory.h      |    419 -
 .../sherpa-onnx/openfst-src/fst/minimize.h    |    567 -
 .../sherpa-onnx/openfst-src/fst/mutable-fst.h |    389 -
 .../sherpa-onnx/openfst-src/fst/pair-weight.h |    164 -
 .../sherpa-onnx/openfst-src/fst/partition.h   |    305 -
 .../openfst-src/fst/power-weight.h            |    168 -
 .../openfst-src/fst/product-weight.h          |    107 -
 .../sherpa-onnx/openfst-src/fst/project.h     |    138 -
 .../sherpa-onnx/openfst-src/fst/properties.h  |    467 -
 .../sherpa-onnx/openfst-src/fst/prune.h       |    325 -
 .../sherpa/sherpa-onnx/openfst-src/fst/push.h |    155 -
 .../sherpa-onnx/openfst-src/fst/queue.h       |    858 -
 .../openfst-src/fst/randequivalent.h          |    114 -
 .../sherpa-onnx/openfst-src/fst/randgen.h     |    750 -
 .../sherpa-onnx/openfst-src/fst/rational.h    |    307 -
 .../sherpa-onnx/openfst-src/fst/register.h    |    115 -
 .../sherpa-onnx/openfst-src/fst/relabel.h     |    478 -
 .../openfst-src/fst/replace-util.h            |    629 -
 .../sherpa-onnx/openfst-src/fst/replace.h     |   1470 -
 .../sherpa-onnx/openfst-src/fst/reverse.h     |    116 -
 .../sherpa-onnx/openfst-src/fst/reweight.h    |    127 -
 .../sherpa-onnx/openfst-src/fst/rmepsilon.h   |    548 -
 .../openfst-src/fst/rmfinalepsilon.h          |     80 -
 .../openfst-src/fst/script/arc-class.h        |     40 -
 .../fst/script/arciterator-class.h            |    212 -
 .../openfst-src/fst/script/arcsort.h          |     44 -
 .../openfst-src/fst/script/arg-packs.h        |     37 -
 .../openfst-src/fst/script/closure.h          |     28 -
 .../openfst-src/fst/script/compile-impl.h     |    217 -
 .../openfst-src/fst/script/compile.h          |     98 -
 .../openfst-src/fst/script/compose.h          |     34 -
 .../openfst-src/fst/script/concat.h           |     40 -
 .../openfst-src/fst/script/connect.h          |     23 -
 .../openfst-src/fst/script/convert.h          |     35 -
 .../openfst-src/fst/script/decode.h           |     49 -
 .../openfst-src/fst/script/determinize.h      |     59 -
 .../openfst-src/fst/script/difference.h       |     35 -
 .../openfst-src/fst/script/disambiguate.h     |     54 -
 .../openfst-src/fst/script/draw-impl.h        |    227 -
 .../sherpa-onnx/openfst-src/fst/script/draw.h |     85 -
 .../openfst-src/fst/script/encode.h           |     51 -
 .../fst/script/encodemapper-class.h           |    169 -
 .../openfst-src/fst/script/epsnormalize.h     |     31 -
 .../openfst-src/fst/script/equal.h            |     32 -
 .../openfst-src/fst/script/equivalent.h       |     34 -
 .../openfst-src/fst/script/fst-class.h        |    533 -
 .../openfst-src/fst/script/fstscript-decl.h   |     32 -
 .../openfst-src/fst/script/fstscript.h        |    155 -
 .../openfst-src/fst/script/getters.h          |     76 -
 .../openfst-src/fst/script/info-impl.h        |    314 -
 .../sherpa-onnx/openfst-src/fst/script/info.h |     50 -
 .../openfst-src/fst/script/intersect.h        |     35 -
 .../openfst-src/fst/script/invert.h           |     23 -
 .../openfst-src/fst/script/isomorphic.h       |     34 -
 .../sherpa-onnx/openfst-src/fst/script/map.h  |    150 -
 .../openfst-src/fst/script/minimize.h         |     33 -
 .../openfst-src/fst/script/print-impl.h       |    132 -
 .../openfst-src/fst/script/print.h            |     79 -
 .../openfst-src/fst/script/project.h          |     28 -
 .../openfst-src/fst/script/prune.h            |     51 -
 .../sherpa-onnx/openfst-src/fst/script/push.h |     53 -
 .../openfst-src/fst/script/randequivalent.h   |     67 -
 .../openfst-src/fst/script/randgen.h          |     63 -
 .../openfst-src/fst/script/register.h         |    114 -
 .../openfst-src/fst/script/relabel.h          |     64 -
 .../openfst-src/fst/script/replace.h          |     72 -
 .../openfst-src/fst/script/reverse.h          |     30 -
 .../openfst-src/fst/script/reweight.h         |     37 -
 .../openfst-src/fst/script/rmepsilon.h        |    114 -
 .../openfst-src/fst/script/script-impl.h      |    211 -
 .../fst/script/shortest-distance.h            |    217 -
 .../openfst-src/fst/script/shortest-path.h    |    116 -
 .../fst/script/stateiterator-class.h          |     85 -
 .../openfst-src/fst/script/synchronize.h      |     29 -
 .../openfst-src/fst/script/text-io.h          |     28 -
 .../openfst-src/fst/script/topsort.h          |     26 -
 .../openfst-src/fst/script/union.h            |     29 -
 .../openfst-src/fst/script/verify.h           |     27 -
 .../openfst-src/fst/script/weight-class.h     |    230 -
 .../sherpa-onnx/openfst-src/fst/set-weight.h  |    592 -
 .../openfst-src/fst/shortest-distance.h       |    336 -
 .../openfst-src/fst/shortest-path.h           |    515 -
 .../openfst-src/fst/signed-log-weight.h       |    436 -
 .../openfst-src/fst/sparse-power-weight.h     |    235 -
 .../openfst-src/fst/sparse-tuple-weight.h     |    411 -
 .../sherpa-onnx/openfst-src/fst/state-map.h   |    613 -
 .../openfst-src/fst/state-reachable.h         |    224 -
 .../sherpa-onnx/openfst-src/fst/state-table.h |    494 -
 .../sherpa-onnx/openfst-src/fst/statesort.h   |     74 -
 .../openfst-src/fst/string-weight.h           |    807 -
 .../sherpa-onnx/openfst-src/fst/string.h      |    270 -
 .../openfst-src/fst/symbol-table-ops.h        |     76 -
 .../openfst-src/fst/symbol-table.h            |    466 -
 .../sherpa-onnx/openfst-src/fst/synchronize.h |    414 -
 .../openfst-src/fst/test-properties.h         |    246 -
 .../sherpa-onnx/openfst-src/fst/topsort.h     |     95 -
 .../openfst-src/fst/tuple-weight.h            |    170 -
 .../sherpa-onnx/openfst-src/fst/types.h       |     41 -
 .../sherpa-onnx/openfst-src/fst/union-find.h  |     84 -
 .../openfst-src/fst/union-weight.h            |    505 -
 .../sherpa-onnx/openfst-src/fst/union.h       |    158 -
 .../sherpa/sherpa-onnx/openfst-src/fst/util.h |    399 -
 .../sherpa-onnx/openfst-src/fst/vector-fst.h  |    717 -
 .../sherpa-onnx/openfst-src/fst/verify.h      |    102 -
 .../sherpa-onnx/openfst-src/fst/visit.h       |    321 -
 .../sherpa-onnx/openfst-src/fst/weight.h      |    358 -
 .../llm_framework/include/utils/checker.h     |     32 -
 projects/llm_framework/include/utils/io.hpp   |    601 -
 projects/llm_framework/include/utils/logger.h |     82 -
 projects/llm_framework/main/SConstruct        |      3 +-
 projects/llm_framework/main_asr/SConstruct    |      2 +-
 projects/llm_framework/main_audio/SConstruct  |      1 -
 projects/llm_framework/main_camera/SConstruct |     11 +-
 .../main_depth_anything/SConstruct            |      2 +-
 projects/llm_framework/main_kws/SConstruct    |      8 +-
 projects/llm_framework/main_llm/SConstruct    |      8 +-
 projects/llm_framework/main_llm/src/main.cpp  |     22 +-
 .../llm_framework/main_melotts/SConstruct     |      4 +-
 projects/llm_framework/main_skel/SConstruct   |      2 +-
 projects/llm_framework/main_vad/SConstruct    |      2 +-
 projects/llm_framework/main_vlm/SConstruct    |     10 +-
 .../models/mode_internvl3-1B-448-ax650.json   |     39 +
 .../tokenizer_internvl3-1B-448-ax650.py       |    241 +
 projects/llm_framework/main_vlm/src/main.cpp  |     34 +-
 .../llm_framework/main_whisper/SConstruct     |      4 +-
 projects/llm_framework/main_yolo/SConstruct   |      2 +-
 2185 files changed, 353 insertions(+), 881750 deletions(-)
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/algorithm/algorithm.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/algorithm/container.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/attributes.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/call_once.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/casts.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/config.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/const_init.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/dynamic_annotations.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/atomic_hook.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/atomic_hook_test_helper.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/cycleclock.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/cycleclock_config.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/direct_mmap.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/dynamic_annotations.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/endian.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/errno_saver.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/exception_safety_testing.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/exception_testing.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/fast_type_id.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/hide_ptr.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/identity.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/inline_variable.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/inline_variable_testing.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/invoke.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/low_level_alloc.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/low_level_scheduling.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/nullability_impl.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/per_thread_tls.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/pretty_function.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/raw_logging.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/scheduling_mode.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/scoped_set_env.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/spinlock.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/spinlock_wait.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/strerror.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/sysinfo.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/thread_identity.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/throw_delegate.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/tsan_mutex_interface.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/unaligned_access.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/unscaledcycleclock.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/internal/unscaledcycleclock_config.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/log_severity.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/macros.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/no_destructor.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/nullability.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/optimization.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/options.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/policy_checks.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/port.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/prefetch.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/base/thread_annotations.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/cleanup/cleanup.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/cleanup/internal/cleanup.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/btree_map.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/btree_set.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/btree_test.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/fixed_array.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/flat_hash_map.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/flat_hash_set.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/inlined_vector.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/btree.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/btree_container.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/common.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/common_policy_traits.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/compressed_tuple.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/container_memory.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_function_defaults.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_generator_testing.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_policy_testing.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_policy_traits.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/hashtable_debug.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/hashtable_debug_hooks.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/hashtablez_sampler.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/inlined_vector.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/layout.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/node_slot_policy.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/raw_hash_map.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/raw_hash_set.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/test_allocator.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/test_instance_tracker.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/tracked.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_constructor_test.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_lookup_test.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_members_test.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_modifiers_test.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_constructor_test.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_lookup_test.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_members_test.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_modifiers_test.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/node_hash_map.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/container/node_hash_set.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/crc/crc32c.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/crc/internal/cpu_detect.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc32_x86_arm_combined_simd.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc32c.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc32c_inline.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc_cord_state.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc_internal.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc_memcpy.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/crc/internal/non_temporal_arm_intrinsics.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/crc/internal/non_temporal_memcpy.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/debugging/failure_signal_handler.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/debugging/internal/address_is_readable.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/debugging/internal/demangle.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/debugging/internal/elf_mem_image.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/debugging/internal/examine_stack.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/debugging/internal/stack_consumption.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/debugging/internal/stacktrace_config.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/debugging/internal/symbolize.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/debugging/internal/vdso_support.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/debugging/leak_check.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/debugging/stacktrace.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/debugging/symbolize.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/commandlineflag.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/config.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/declare.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/flag.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/internal/commandlineflag.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/internal/flag.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/internal/parse.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/internal/path_util.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/internal/private_handle_accessor.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/internal/program_name.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/internal/registry.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/internal/sequence_lock.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/internal/usage.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/marshalling.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/parse.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/reflection.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/usage.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/flags/usage_config.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/functional/any_invocable.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/functional/bind_front.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/functional/function_ref.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/functional/internal/any_invocable.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/functional/internal/front_binder.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/functional/internal/function_ref.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/functional/overload.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/hash/hash.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/hash/hash_testing.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/hash/internal/city.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/hash/internal/hash.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/hash/internal/hash_test.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/hash/internal/low_level_hash.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/hash/internal/spy_hash_state.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/absl_check.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/absl_log.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/absl_vlog_is_on.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/check.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/die_if_null.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/flags.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/globals.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/initialize.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/append_truncated.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/check_impl.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/check_op.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/conditions.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/config.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/flags.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/fnmatch.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/globals.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/log_format.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/log_impl.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/log_message.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/log_sink_set.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/nullguard.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/nullstream.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/proto.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/strip.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/structured.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/test_actions.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/test_helpers.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/test_matchers.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/vlog_config.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/internal/voidify.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/log.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/log_entry.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/log_sink.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/log_sink_registry.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/log_streamer.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/scoped_mock_log.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/structured.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/log/vlog_is_on.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/memory/memory.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/meta/type_traits.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/numeric/bits.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/numeric/int128.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/numeric/internal/bits.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/numeric/internal/representation.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/profiling/internal/exponential_biased.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/profiling/internal/periodic_sampler.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/profiling/internal/sample_recorder.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/bernoulli_distribution.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/beta_distribution.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/bit_gen_ref.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/discrete_distribution.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/distributions.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/exponential_distribution.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/gaussian_distribution.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/chi_square.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/distribution_caller.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/distribution_test_util.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/explicit_seed_seq.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/fast_uniform_bits.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/fastmath.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/generate_real.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/iostream_state_saver.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/mock_helpers.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/mock_overload_set.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/nanobenchmark.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/nonsecure_base.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/pcg_engine.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/platform.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/pool_urbg.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/randen.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_detect.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_engine.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_hwaes.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_slow.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_traits.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/salted_seed_seq.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/seed_material.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/sequence_urbg.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/traits.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/uniform_helper.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/internal/wide_multiply.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/log_uniform_int_distribution.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/mock_distributions.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/mocking_bit_gen.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/poisson_distribution.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/random.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/seed_gen_exception.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/seed_sequences.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/uniform_int_distribution.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/uniform_real_distribution.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/random/zipf_distribution.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/status/internal/status_internal.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/status/internal/statusor_internal.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/status/status.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/status/status_payload_printer.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/status/statusor.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/ascii.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/charconv.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/charset.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/cord.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/cord_analysis.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/cord_buffer.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/cord_test_helpers.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/cordz_test_helpers.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/escaping.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/has_absl_stringify.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/has_ostream_operator.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/charconv_bigint.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/charconv_parse.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_data_edge.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_internal.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_btree.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_btree_navigator.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_btree_reader.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_consume.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_crc.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_flat.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_test_util.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_functions.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_handle.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_info.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_sample_token.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_statistics.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_update_scope.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_update_tracker.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/damerau_levenshtein_distance.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/escaping.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/escaping_test_common.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/has_absl_stringify.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/memutil.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/numbers_test_common.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/ostringstream.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/pow10_helper.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/resize_uninitialized.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/stl_type_traits.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/arg.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/bind.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/checker.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/constexpr_parser.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/extension.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/float_conversion.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/output.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/parser.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_join_internal.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_split_internal.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/string_constant.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/stringify_sink.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/internal/utf8.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/match.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/numbers.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/str_cat.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/str_format.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/str_join.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/str_replace.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/str_split.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/string_view.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/strip.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/strings/substitute.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/barrier.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/blocking_counter.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/create_thread_identity.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/futex.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/futex_waiter.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/graphcycles.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/kernel_timeout.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/per_thread_sem.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/pthread_waiter.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/sem_waiter.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/stdcpp_waiter.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/thread_pool.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/waiter.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/waiter_base.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/win32_waiter.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/mutex.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/synchronization/notification.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/time/civil_time.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/time/clock.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/civil_time.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/civil_time_detail.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/time_zone.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/zone_info_source.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_fixed.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_if.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_impl.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_info.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_libc.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_posix.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/tzfile.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/time/internal/test_util.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/time/time.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/types/any.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/types/bad_any_cast.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/types/bad_optional_access.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/types/bad_variant_access.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/types/compare.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/types/internal/optional.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/types/internal/span.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/types/internal/variant.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/types/optional.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/types/span.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/types/variant.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/utility/internal/if_constexpr.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/absl/utility/utility.h
 delete mode 100644 projects/llm_framework/include/abseil-cpp/ci/absl_alternate_options.h
 delete mode 100644 projects/llm_framework/include/archive.h
 delete mode 100644 projects/llm_framework/include/archive_entry.h
 delete mode 100644 projects/llm_framework/include/dwarf.h
 delete mode 100644 projects/llm_framework/include/elfutils/elf-knowledge.h
 delete mode 100644 projects/llm_framework/include/elfutils/known-dwarf.h
 delete mode 100644 projects/llm_framework/include/elfutils/libdw.h
 delete mode 100644 projects/llm_framework/include/elfutils/libdwelf.h
 delete mode 100644 projects/llm_framework/include/elfutils/libdwfl.h
 delete mode 100644 projects/llm_framework/include/elfutils/version.h
 delete mode 100644 projects/llm_framework/include/fst/accumulator.h
 delete mode 100644 projects/llm_framework/include/fst/add-on.h
 delete mode 100644 projects/llm_framework/include/fst/arc-arena.h
 delete mode 100644 projects/llm_framework/include/fst/arc-map.h
 delete mode 100644 projects/llm_framework/include/fst/arc.h
 delete mode 100644 projects/llm_framework/include/fst/arcfilter.h
 delete mode 100644 projects/llm_framework/include/fst/arcsort.h
 delete mode 100644 projects/llm_framework/include/fst/bi-table.h
 delete mode 100644 projects/llm_framework/include/fst/cache.h
 delete mode 100644 projects/llm_framework/include/fst/closure.h
 delete mode 100644 projects/llm_framework/include/fst/compact-fst.h
 delete mode 100644 projects/llm_framework/include/fst/compat.h
 delete mode 100644 projects/llm_framework/include/fst/complement.h
 delete mode 100644 projects/llm_framework/include/fst/compose-filter.h
 delete mode 100644 projects/llm_framework/include/fst/compose.h
 delete mode 100644 projects/llm_framework/include/fst/concat.h
 delete mode 100644 projects/llm_framework/include/fst/config.h
 delete mode 100644 projects/llm_framework/include/fst/config.h.in
 delete mode 100644 projects/llm_framework/include/fst/connect.h
 delete mode 100644 projects/llm_framework/include/fst/const-fst.h
 delete mode 100644 projects/llm_framework/include/fst/determinize.h
 delete mode 100644 projects/llm_framework/include/fst/dfs-visit.h
 delete mode 100644 projects/llm_framework/include/fst/difference.h
 delete mode 100644 projects/llm_framework/include/fst/disambiguate.h
 delete mode 100644 projects/llm_framework/include/fst/edit-fst.h
 delete mode 100644 projects/llm_framework/include/fst/encode.h
 delete mode 100644 projects/llm_framework/include/fst/epsnormalize.h
 delete mode 100644 projects/llm_framework/include/fst/equal.h
 delete mode 100644 projects/llm_framework/include/fst/equivalent.h
 delete mode 100644 projects/llm_framework/include/fst/expanded-fst.h
 delete mode 100644 projects/llm_framework/include/fst/expectation-weight.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/compress/compress-script.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/compress/compress.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/compress/elias.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/compress/gzfile.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/compress/randmod.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/far/compile-strings.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/far/create.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/far/equal.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/far/extract.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/far/far-class.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/far/far.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/far/farlib.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/far/farscript.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/far/getters.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/far/info.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/far/isomorphic.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/far/print-strings.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/far/script-impl.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/far/stlist.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/far/sttable.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/linear/linear-fst-data-builder.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/linear/linear-fst-data.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/linear/linear-fst.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/linear/linearscript.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/linear/loglinear-apply.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/linear/trie.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/mpdt/compose.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/mpdt/expand.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/mpdt/info.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/mpdt/mpdt.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/mpdt/mpdtlib.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/mpdt/mpdtscript.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/mpdt/read_write_utils.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/mpdt/reverse.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/ngram/bitmap-index.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/ngram/ngram-fst.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/ngram/nthbit.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/pdt/collection.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/pdt/compose.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/pdt/expand.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/pdt/getters.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/pdt/info.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/pdt/paren.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/pdt/pdt.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/pdt/pdtlib.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/pdt/pdtscript.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/pdt/replace.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/pdt/reverse.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/pdt/shortest-path.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/special/phi-fst.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/special/rho-fst.h
 delete mode 100644 projects/llm_framework/include/fst/extensions/special/sigma-fst.h
 delete mode 100644 projects/llm_framework/include/fst/factor-weight.h
 delete mode 100644 projects/llm_framework/include/fst/filter-state.h
 delete mode 100644 projects/llm_framework/include/fst/flags.h
 delete mode 100644 projects/llm_framework/include/fst/float-weight.h
 delete mode 100644 projects/llm_framework/include/fst/fst-decl.h
 delete mode 100644 projects/llm_framework/include/fst/fst.h
 delete mode 100644 projects/llm_framework/include/fst/fstlib.h
 delete mode 100644 projects/llm_framework/include/fst/generic-register.h
 delete mode 100644 projects/llm_framework/include/fst/heap.h
 delete mode 100644 projects/llm_framework/include/fst/icu.h
 delete mode 100644 projects/llm_framework/include/fst/intersect.h
 delete mode 100644 projects/llm_framework/include/fst/interval-set.h
 delete mode 100644 projects/llm_framework/include/fst/invert.h
 delete mode 100644 projects/llm_framework/include/fst/isomorphic.h
 delete mode 100644 projects/llm_framework/include/fst/label-reachable.h
 delete mode 100644 projects/llm_framework/include/fst/lexicographic-weight.h
 delete mode 100644 projects/llm_framework/include/fst/lock.h
 delete mode 100644 projects/llm_framework/include/fst/log.h
 delete mode 100644 projects/llm_framework/include/fst/lookahead-filter.h
 delete mode 100644 projects/llm_framework/include/fst/lookahead-matcher.h
 delete mode 100644 projects/llm_framework/include/fst/map.h
 delete mode 100644 projects/llm_framework/include/fst/mapped-file.h
 delete mode 100644 projects/llm_framework/include/fst/matcher-fst.h
 delete mode 100644 projects/llm_framework/include/fst/matcher.h
 delete mode 100644 projects/llm_framework/include/fst/memory.h
 delete mode 100644 projects/llm_framework/include/fst/minimize.h
 delete mode 100644 projects/llm_framework/include/fst/mutable-fst.h
 delete mode 100644 projects/llm_framework/include/fst/pair-weight.h
 delete mode 100644 projects/llm_framework/include/fst/partition.h
 delete mode 100644 projects/llm_framework/include/fst/power-weight.h
 delete mode 100644 projects/llm_framework/include/fst/product-weight.h
 delete mode 100644 projects/llm_framework/include/fst/project.h
 delete mode 100644 projects/llm_framework/include/fst/properties.h
 delete mode 100644 projects/llm_framework/include/fst/prune.h
 delete mode 100644 projects/llm_framework/include/fst/push.h
 delete mode 100644 projects/llm_framework/include/fst/queue.h
 delete mode 100644 projects/llm_framework/include/fst/randequivalent.h
 delete mode 100644 projects/llm_framework/include/fst/randgen.h
 delete mode 100644 projects/llm_framework/include/fst/rational.h
 delete mode 100644 projects/llm_framework/include/fst/register.h
 delete mode 100644 projects/llm_framework/include/fst/relabel.h
 delete mode 100644 projects/llm_framework/include/fst/replace-util.h
 delete mode 100644 projects/llm_framework/include/fst/replace.h
 delete mode 100644 projects/llm_framework/include/fst/reverse.h
 delete mode 100644 projects/llm_framework/include/fst/reweight.h
 delete mode 100644 projects/llm_framework/include/fst/rmepsilon.h
 delete mode 100644 projects/llm_framework/include/fst/rmfinalepsilon.h
 delete mode 100644 projects/llm_framework/include/fst/script/arc-class.h
 delete mode 100644 projects/llm_framework/include/fst/script/arciterator-class.h
 delete mode 100644 projects/llm_framework/include/fst/script/arcsort.h
 delete mode 100644 projects/llm_framework/include/fst/script/arg-packs.h
 delete mode 100644 projects/llm_framework/include/fst/script/closure.h
 delete mode 100644 projects/llm_framework/include/fst/script/compile-impl.h
 delete mode 100644 projects/llm_framework/include/fst/script/compile.h
 delete mode 100644 projects/llm_framework/include/fst/script/compose.h
 delete mode 100644 projects/llm_framework/include/fst/script/concat.h
 delete mode 100644 projects/llm_framework/include/fst/script/connect.h
 delete mode 100644 projects/llm_framework/include/fst/script/convert.h
 delete mode 100644 projects/llm_framework/include/fst/script/decode.h
 delete mode 100644 projects/llm_framework/include/fst/script/determinize.h
 delete mode 100644 projects/llm_framework/include/fst/script/difference.h
 delete mode 100644 projects/llm_framework/include/fst/script/disambiguate.h
 delete mode 100644 projects/llm_framework/include/fst/script/draw-impl.h
 delete mode 100644 projects/llm_framework/include/fst/script/draw.h
 delete mode 100644 projects/llm_framework/include/fst/script/encode.h
 delete mode 100644 projects/llm_framework/include/fst/script/encodemapper-class.h
 delete mode 100644 projects/llm_framework/include/fst/script/epsnormalize.h
 delete mode 100644 projects/llm_framework/include/fst/script/equal.h
 delete mode 100644 projects/llm_framework/include/fst/script/equivalent.h
 delete mode 100644 projects/llm_framework/include/fst/script/fst-class.h
 delete mode 100644 projects/llm_framework/include/fst/script/fstscript-decl.h
 delete mode 100644 projects/llm_framework/include/fst/script/fstscript.h
 delete mode 100644 projects/llm_framework/include/fst/script/getters.h
 delete mode 100644 projects/llm_framework/include/fst/script/info-impl.h
 delete mode 100644 projects/llm_framework/include/fst/script/info.h
 delete mode 100644 projects/llm_framework/include/fst/script/intersect.h
 delete mode 100644 projects/llm_framework/include/fst/script/invert.h
 delete mode 100644 projects/llm_framework/include/fst/script/isomorphic.h
 delete mode 100644 projects/llm_framework/include/fst/script/map.h
 delete mode 100644 projects/llm_framework/include/fst/script/minimize.h
 delete mode 100644 projects/llm_framework/include/fst/script/print-impl.h
 delete mode 100644 projects/llm_framework/include/fst/script/print.h
 delete mode 100644 projects/llm_framework/include/fst/script/project.h
 delete mode 100644 projects/llm_framework/include/fst/script/prune.h
 delete mode 100644 projects/llm_framework/include/fst/script/push.h
 delete mode 100644 projects/llm_framework/include/fst/script/randequivalent.h
 delete mode 100644 projects/llm_framework/include/fst/script/randgen.h
 delete mode 100644 projects/llm_framework/include/fst/script/register.h
 delete mode 100644 projects/llm_framework/include/fst/script/relabel.h
 delete mode 100644 projects/llm_framework/include/fst/script/replace.h
 delete mode 100644 projects/llm_framework/include/fst/script/reverse.h
 delete mode 100644 projects/llm_framework/include/fst/script/reweight.h
 delete mode 100644 projects/llm_framework/include/fst/script/rmepsilon.h
 delete mode 100644 projects/llm_framework/include/fst/script/script-impl.h
 delete mode 100644 projects/llm_framework/include/fst/script/shortest-distance.h
 delete mode 100644 projects/llm_framework/include/fst/script/shortest-path.h
 delete mode 100644 projects/llm_framework/include/fst/script/stateiterator-class.h
 delete mode 100644 projects/llm_framework/include/fst/script/synchronize.h
 delete mode 100644 projects/llm_framework/include/fst/script/text-io.h
 delete mode 100644 projects/llm_framework/include/fst/script/topsort.h
 delete mode 100644 projects/llm_framework/include/fst/script/union.h
 delete mode 100644 projects/llm_framework/include/fst/script/verify.h
 delete mode 100644 projects/llm_framework/include/fst/script/weight-class.h
 delete mode 100644 projects/llm_framework/include/fst/set-weight.h
 delete mode 100644 projects/llm_framework/include/fst/shortest-distance.h
 delete mode 100644 projects/llm_framework/include/fst/shortest-path.h
 delete mode 100644 projects/llm_framework/include/fst/signed-log-weight.h
 delete mode 100644 projects/llm_framework/include/fst/sparse-power-weight.h
 delete mode 100644 projects/llm_framework/include/fst/sparse-tuple-weight.h
 delete mode 100644 projects/llm_framework/include/fst/state-map.h
 delete mode 100644 projects/llm_framework/include/fst/state-reachable.h
 delete mode 100644 projects/llm_framework/include/fst/state-table.h
 delete mode 100644 projects/llm_framework/include/fst/statesort.h
 delete mode 100644 projects/llm_framework/include/fst/string-weight.h
 delete mode 100644 projects/llm_framework/include/fst/string.h
 delete mode 100644 projects/llm_framework/include/fst/symbol-table-ops.h
 delete mode 100644 projects/llm_framework/include/fst/symbol-table.h
 delete mode 100644 projects/llm_framework/include/fst/synchronize.h
 delete mode 100644 projects/llm_framework/include/fst/test-properties.h
 delete mode 100644 projects/llm_framework/include/fst/test/algo_test.h
 delete mode 100644 projects/llm_framework/include/fst/test/fst_test.h
 delete mode 100644 projects/llm_framework/include/fst/test/rand-fst.h
 delete mode 100644 projects/llm_framework/include/fst/test/weight-tester.h
 delete mode 100644 projects/llm_framework/include/fst/topsort.h
 delete mode 100644 projects/llm_framework/include/fst/tuple-weight.h
 delete mode 100644 projects/llm_framework/include/fst/types.h
 delete mode 100644 projects/llm_framework/include/fst/union-find.h
 delete mode 100644 projects/llm_framework/include/fst/union-weight.h
 delete mode 100644 projects/llm_framework/include/fst/union.h
 delete mode 100644 projects/llm_framework/include/fst/util.h
 delete mode 100644 projects/llm_framework/include/fst/vector-fst.h
 delete mode 100644 projects/llm_framework/include/fst/verify.h
 delete mode 100644 projects/llm_framework/include/fst/visit.h
 delete mode 100644 projects/llm_framework/include/fst/weight.h
 delete mode 100644 projects/llm_framework/include/gelf.h
 delete mode 100644 projects/llm_framework/include/gflags/defines.h
 delete mode 100644 projects/llm_framework/include/gflags/gflags.h
 delete mode 100644 projects/llm_framework/include/gflags/gflags_completions.h
 delete mode 100644 projects/llm_framework/include/gflags/gflags_declare.h
 delete mode 100644 projects/llm_framework/include/glog/log_severity.h
 delete mode 100644 projects/llm_framework/include/glog/logging.h
 delete mode 100644 projects/llm_framework/include/glog/raw_logging.h
 delete mode 100644 projects/llm_framework/include/glog/stl_logging.h
 delete mode 100644 projects/llm_framework/include/glog/vlog_is_on.h
 delete mode 100644 projects/llm_framework/include/libdw.h
 delete mode 100644 projects/llm_framework/include/libelf.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/basic_types.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/code_location.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/common.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/const_pointer_container.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/denormal.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/eigen_common_wrapper.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/exceptions.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/gpu_profiler_common.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/gsl.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/hash_combine.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/inlined_containers.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/inlined_containers_fwd.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/logging/capture.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/logging/isink.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/logging/logging.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/logging/macros.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/logging/severity.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/make_string.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/narrow.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/optional.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/parse_string.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/profiler_common.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/span_utils.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/spin_pause.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/status.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/common/string_helper.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/alloc_kind.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/allocator.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/buffer_deleter.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/customregistry.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/data_types.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/data_types_internal.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/endian.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/execution_provider.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/float16.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/framework_common.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/func_api.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/kernel_def_builder.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/kernel_registry.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/op_kernel.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/op_kernel_context.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/op_kernel_info.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/op_node_proto_helper.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/ort_value.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/ortdevice.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/ortmemoryinfo.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/provider_options.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/provider_options_utils.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/provider_shutdown.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/run_options.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/sparse_tensor.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/stream_handles.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/tensor.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/tensor_shape.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/framework/to_tensor_proto_element_type.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/graph/basic_types.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/graph/constants.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/graph/function.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/graph/graph.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/graph/graph_nodes.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/graph/graph_viewer.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/graph/indexed_sub_graph.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/graph/node_arg.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/graph/schema_registry.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer_config.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer_level.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer_utils.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/optimizer/rewrite_rule.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/optimizer/rule_based_graph_transformer.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/providers/cpu/cpu_provider_factory.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/session/environment.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/session/experimental_onnxruntime_cxx_api.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/session/experimental_onnxruntime_cxx_inline.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/session/onnxruntime_c_api.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/session/onnxruntime_cxx_api.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/session/onnxruntime_cxx_inline.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/session/onnxruntime_run_options_config_keys.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/session/onnxruntime_session_options_config_keys.h
 delete mode 100644 projects/llm_framework/include/onnxruntime/core/session/snippets.dox
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/affine.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/async.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/base.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/bindings_utils.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/bufferpool.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/check.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/core.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/core_c.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda.inl.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/block.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/border_interpolate.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/color.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/common.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/datamov_utils.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/color_detail.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/reduce.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/reduce_key_val.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/transform_detail.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/type_traits_detail.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/vec_distance_detail.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/dynamic_smem.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/emulation.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/filters.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/funcattrib.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/functional.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/limits.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/reduce.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/saturate_cast.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/scan.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/simd_functions.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/transform.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/type_traits.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/utility.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/vec_distance.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/vec_math.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/vec_traits.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/warp.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/warp_reduce.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda/warp_shuffle.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda_stream_accessor.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cuda_types.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cv_cpu_dispatch.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cv_cpu_helper.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cvdef.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cvstd.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cvstd.inl.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/cvstd_wrapper.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/detail/async_promise.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/detail/dispatch_helper.impl.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/detail/exception_ptr.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/directx.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/dualquaternion.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/dualquaternion.inl.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/eigen.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/fast_math.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/hal.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/interface.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/intrin.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_avx.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_avx512.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_cpp.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_forward.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_msa.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_neon.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_rvv.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_rvv071.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_sse.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_sse_em.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_vsx.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_wasm.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/msa_macros.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/hal/simd_utils.impl.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/mat.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/mat.inl.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/matx.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/neon_utils.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/ocl.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/ocl_genbase.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/ocl_defs.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/opencl_info.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/opencl_svm.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_clblas.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_clfft.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_core.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_core_wrappers.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_gl.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_gl_wrappers.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_clblas.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_clfft.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_core.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_core_wrappers.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_gl.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_gl_wrappers.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_svm_20.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_svm_definitions.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_svm_hsa_extension.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/opengl.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/operations.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/optim.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/ovx.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/parallel/backend/parallel_for.openmp.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/parallel/backend/parallel_for.tbb.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/parallel/parallel_backend.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/persistence.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/quaternion.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/quaternion.inl.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/saturate.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/simd_intrinsics.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/softfloat.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/sse_utils.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/traits.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/types.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/types_c.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/utility.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/utils/allocator_stats.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/utils/allocator_stats.impl.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/utils/filesystem.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/utils/instrumentation.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/utils/logger.defines.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/utils/logger.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/utils/logtag.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/utils/tls.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/utils/trace.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/va_intel.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/version.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/core/vsx_utils.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/cvconfig.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/highgui.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/highgui/highgui.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/highgui/highgui_c.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/imgcodecs.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/imgcodecs/imgcodecs.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/imgcodecs/imgcodecs_c.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/imgcodecs/ios.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/imgcodecs/legacy/constants_c.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/imgcodecs/macosx.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/imgproc.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/imgproc/bindings.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/imgproc/detail/gcgraph.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/imgproc/hal/hal.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/imgproc/hal/interface.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/imgproc/imgproc.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/imgproc/imgproc_c.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/imgproc/segmentation.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/imgproc/types_c.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/opencv.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/opencv_modules.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/videoio.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/videoio/cap_ios.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/videoio/legacy/constants_c.h
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/videoio/registry.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/videoio/videoio.hpp
 delete mode 100644 projects/llm_framework/include/opencv4/opencv2/videoio/videoio_c.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/any.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/arena.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/arena_impl.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/arenastring.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/descriptor.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/extension_set.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/extension_set_inl.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/generated_enum_reflection.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/generated_enum_util.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/generated_message_table_driven.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/generated_message_table_driven_lite.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/generated_message_util.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/has_bits.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/implicit_weak_message.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/io/coded_stream.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/io/io_win32.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/io/zero_copy_stream.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/io/zero_copy_stream_impl.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/io/zero_copy_stream_impl_lite.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/map.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/map_entry_lite.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/map_field_lite.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/map_type_handler.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/message_lite.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/metadata_lite.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/parse_context.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/port.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/port_def.inc
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/port_undef.inc
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/repeated_field.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/bytestream.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/callback.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/casts.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/common.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/hash.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/int128.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/logging.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/macros.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/map_util.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/mutex.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/once.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/platform_macros.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/port.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/status.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/statusor.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/stl_util.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/stringpiece.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/stringprintf.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/strutil.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/time.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/unknown_field_set.h
 delete mode 100644 projects/llm_framework/include/protobuf-lite/google/protobuf/wire_format_lite.h
 delete mode 100644 projects/llm_framework/include/re2/re2/bitmap256.h
 delete mode 100644 projects/llm_framework/include/re2/re2/filtered_re2.h
 delete mode 100644 projects/llm_framework/include/re2/re2/pod_array.h
 delete mode 100644 projects/llm_framework/include/re2/re2/prefilter.h
 delete mode 100644 projects/llm_framework/include/re2/re2/prefilter_tree.h
 delete mode 100644 projects/llm_framework/include/re2/re2/prog.h
 delete mode 100644 projects/llm_framework/include/re2/re2/re2.h
 delete mode 100644 projects/llm_framework/include/re2/re2/regexp.h
 delete mode 100644 projects/llm_framework/include/re2/re2/set.h
 delete mode 100644 projects/llm_framework/include/re2/re2/sparse_array.h
 delete mode 100644 projects/llm_framework/include/re2/re2/sparse_set.h
 delete mode 100644 projects/llm_framework/include/re2/re2/stringpiece.h
 delete mode 100644 projects/llm_framework/include/re2/re2/testing/exhaustive_tester.h
 delete mode 100644 projects/llm_framework/include/re2/re2/testing/regexp_generator.h
 delete mode 100644 projects/llm_framework/include/re2/re2/testing/string_generator.h
 delete mode 100644 projects/llm_framework/include/re2/re2/testing/tester.h
 delete mode 100644 projects/llm_framework/include/re2/re2/unicode_casefold.h
 delete mode 100644 projects/llm_framework/include/re2/re2/unicode_groups.h
 delete mode 100644 projects/llm_framework/include/re2/re2/walker-inl.h
 delete mode 100644 projects/llm_framework/include/re2/util/logging.h
 delete mode 100644 projects/llm_framework/include/re2/util/malloc_counter.h
 delete mode 100644 projects/llm_framework/include/re2/util/pcre.h
 delete mode 100644 projects/llm_framework/include/re2/util/strutil.h
 delete mode 100644 projects/llm_framework/include/re2/util/utf.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/bpe_model.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/bpe_model_trainer.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/builder.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/builtin_pb/sentencepiece.pb.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/builtin_pb/sentencepiece_model.pb.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/char_model.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/char_model_trainer.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/common.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/filesystem.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/freelist.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/init.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/model_factory.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/model_interface.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/normalization_rule.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/normalizer.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/pretokenizer_for_training.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/sentencepiece_processor.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/sentencepiece_trainer.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/spec_parser.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/testharness.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/trainer_factory.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/trainer_interface.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/unicode_script.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/unicode_script_map.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/unigram_model.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/unigram_model_trainer.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/util.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/word_model.h
 delete mode 100644 projects/llm_framework/include/sentencepiece/word_model_trainer.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/allocator.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/benchmark.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/blob.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/c_api.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/command.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/cpu.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/alsa.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/context-graph.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/conv-emformer-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/display.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/endpoint.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/features.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/greedy-search-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/hypothesis.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/lstm-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/math.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/meta-data.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/microphone.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/modified-beam-search-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/poolingmodulenoproj.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/recognizer.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/resample.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/simpleupsample.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/stack.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/stream.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/symbol-table.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/tensorasstrided.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/wave-reader.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/zipformer-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/datareader.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/gpu.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/absval.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/argmax.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/absval_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/arm_activation.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/arm_usability.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/batchnorm_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/bias_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/binaryop_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/cast_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/cast_bf16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/cast_fp16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/clip_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/concat_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_packed.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_packed_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_packed_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_1x1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_2x2.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to4_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to4_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to8_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack8_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_4x4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5_pack4_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5_pack8_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7_pack1to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7_pack1to4_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7_pack1to8_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_bf16s_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack4_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack8_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack8_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5_pack4_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5_pack8_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/crop_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_3x3.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_4x4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_4x4_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolutiondepthwise_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/dequantize_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/dropout_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/eltwise_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/flatten_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gelu_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_bf16s_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gru_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/hardsigmoid_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/hardswish_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/innerproduct_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/innerproduct_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/innerproduct_gemm_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/instancenorm_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack4_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack4_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack8_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack4_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack4_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack8_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/lrn_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/lstm_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/matmul_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/mish_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/multiheadattention_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/neon_mathfun.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/neon_mathfun_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/neon_mathfun_tanh.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/packing_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack4_bf16s_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack8_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack8_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pixelshuffle_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_2x2.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_2x2_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_2x2_pack4_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_3x3.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_3x3_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_3x3_pack4_bf16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/prelu_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/quantize_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/relu_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_leakyrelu_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_leakyrelu_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_relu_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_relu_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/reshape_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/rnn_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/scale_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/selu_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/shufflechannel_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/sigmoid_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/slice_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/softmax_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/swish_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/tanh_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/unaryop_arm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/batchnorm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/bias.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/binaryop.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/bnll.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/cast.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/celu.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/clip.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/concat.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolution.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolution1d.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolution3d.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolutiondepthwise.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolutiondepthwise1d.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolutiondepthwise3d.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/copyto.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/crop.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/cumulativesum.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolution.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolution1d.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolution3d.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolutiondepthwise.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolutiondepthwise1d.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolutiondepthwise3d.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deepcopy.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deformableconv2d.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/dequantize.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/detectionoutput.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/diag.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/dropout.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/einsum.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/eltwise.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/elu.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/embed.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/erf.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/exp.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/expanddims.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/flatten.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/fold.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/fused_activation.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gelu.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gemm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/glu.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gridsample.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/groupnorm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gru.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/hardsigmoid.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/hardswish.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/innerproduct.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/input.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/instancenorm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/interp.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/layernorm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/log.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/absval_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/batchnorm_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/bias_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/binaryop_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/cast_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/clip_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/concat_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution1d_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack1to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack4to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack8to1_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack8to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack1to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack8to1_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack8to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_7x7_pack1to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack1to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack1to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack4to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack8to1_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack8to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack1to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack4to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack8to1_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack8to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_pack8to1_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_pack8to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_pack4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_pack8_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_3x3.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_3x3_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_5x5_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/crop_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_pack1to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_pack4to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolutiondepthwise_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/dequantize_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/dropout_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/eltwise_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/flatten_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/hardsigmoid_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/hardswish_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/innerproduct_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bicubic.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bicubic_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bilinear.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bilinear_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/loongarch_activation.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/loongarch_usability.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/lsx_mathfun.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/mish_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/packing_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/padding_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/padding_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/padding_pack8_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/pooling_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/prelu_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/quantize_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/relu_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_leakyrelu_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_leakyrelu_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_relu_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_relu_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/sigmoid_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/slice_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/softmax_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/swish_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/tanh_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/unaryop_loongarch.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/lrn.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/lstm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/matmul.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/memorydata.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/absval_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/batchnorm_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/bias_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/binaryop_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/cast_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/clip_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/concat_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution1d_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack1to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack4to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack8to1_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack8to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack1to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack8to1_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack8to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_7x7_pack1to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack1to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack1to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack4to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack8to1_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack8to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack1to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack4to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack8to1_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack8to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_pack8to1_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_pack8to4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_pack4_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_pack8_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_3x3.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_3x3_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_5x5_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/crop_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_pack1to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_pack4to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolutiondepthwise_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/dequantize_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/dropout_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/eltwise_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/flatten_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/hardsigmoid_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/hardswish_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/innerproduct_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bicubic.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bicubic_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bilinear.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bilinear_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/loongson_mmi.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/mips_activation.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/mips_usability.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/mish_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/msa_mathfun.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/packing_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/padding_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/padding_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/padding_pack8_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/pooling_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/prelu_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/quantize_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/relu_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_leakyrelu_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_leakyrelu_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_relu_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_relu_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/sigmoid_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/slice_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/softmax_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/swish_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/tanh_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/unaryop_mips.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mish.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/multiheadattention.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mvn.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/noop.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/normalize.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/packing.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/padding.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/permute.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pixelshuffle.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pooling.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pooling1d.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pooling3d.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/power.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/prelu.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/priorbox.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/proposal.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/psroipooling.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/quantize.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/reduction.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/relu.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/reorg.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/requantize.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/reshape.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/absval_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/batchnorm_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/binaryop_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/cast_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/clip_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/concat_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution1d_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_pack1ton.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_pack1ton_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packn.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packn_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packnto1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packnto1_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_pack1ton.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_pack1ton_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_packn.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_packn_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_7x7_pack1ton.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_7x7_pack1ton_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_pack1ton.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_pack1ton_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packn.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packn_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packnto1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packnto1_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_pack1ton.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_pack1ton_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packn.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packn_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packnto1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packnto1_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_dot.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_dot_packn.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_dot_packn_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_transform.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_transform_packn.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_transform_packn_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_3x3.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_3x3_packn.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_3x3_packn_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_5x5_packn.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_5x5_packn_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/crop_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_pack1ton.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_pack1ton_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packn.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packn_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packnto1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packnto1_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolutiondepthwise_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/dropout_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/flatten_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/gelu_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/gemm_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/gru_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/hardsigmoid_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/hardswish_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/innerproduct_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/instancenorm_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic_packn.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic_packn_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear_packn.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear_packn_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/mish_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/packing_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/padding_packn.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/padding_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/pooling_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/prelu_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/relu_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/riscv_activation.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/riscv_usability.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/rvv_mathfun.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/rvv_mathfun_fp16s.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/selu_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/sigmoid_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/softmax_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/swish_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/tanh_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/unaryop_riscv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/rnn.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/roialign.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/roipooling.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/scale.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/selu.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/shrink.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/shufflechannel.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/sigmoid.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/slice.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/softmax.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/softplus.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/split.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/spp.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/squeeze.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/statisticspooling.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/swish.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/tanh.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/threshold.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/tile.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/unaryop.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/unfold.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/absval_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/batchnorm_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/binaryop_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/cast_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/celu_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/clip_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/concat_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/convolution1d_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/convolution_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/convolutiondepthwise_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/crop_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/deconvolution_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/deconvolutiondepthwise_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/deepcopy_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/dropout_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/eltwise_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/elu_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/erf_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/flatten_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/gelu_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/gemm_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/hardsigmoid_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/hardswish_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/innerproduct_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/instancenorm_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/interp_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/lrn_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/memorydata_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/mish_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/multiheadattention_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/noop_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/normalize_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/packing_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/padding_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/permute_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/pixelshuffle_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/pooling_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/prelu_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/priorbox_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/relu_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/reorg_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/reshape_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/scale_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/shufflechannel_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/sigmoid_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/slice_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/softmax_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/split_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/swish_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/tanh_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/unaryop_vulkan.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/avx512_mathfun.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/avx_mathfun.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/batchnorm_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/bias_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/binaryop_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/bnll_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/cast_bf16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/cast_fp16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/cast_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/clip_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/concat_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution1d_packed.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution1d_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_1x1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_2x2_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack16to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack1to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack1to8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack8to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_winograd.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_winograd_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_5x5.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_im2col_gemm_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_packed.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_packed_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_pack16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_5x5_pack16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_5x5_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_5x5_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/crop_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16to8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack1to16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack1to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack1to8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4to16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4to8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8to16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolutiondepthwise_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16to8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack1to16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack1to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack1to8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4to16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4to8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8to1.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8to16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8to4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/dequantize_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/dropout_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/eltwise_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/elu_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/flatten_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gelu_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gemm_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bicubic_apply_interpolation.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bicubic_compute_blob.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bilinear_apply_interpolation.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bilinear_compute_blob.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_compute_blob.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_nearest_apply_interpolation.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_nearest_compute_blob.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/groupnorm_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/hardsigmoid_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/hardswish_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/innerproduct_fp.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/innerproduct_gemm_fp.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/innerproduct_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic_pack16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear_pack16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/layernorm_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/lrn_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/lstm_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/matmul_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/mish_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/multiheadattention_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/packing_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack8_int8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2_pack16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_3x3_pack16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_3x3_pack4.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_3x3_pack8.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/prelu_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/quantize_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/relu_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/requantize_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/reshape_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/roialign_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/scale_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/selu_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/shufflechannel_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/sigmoid_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/slice_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/softmax_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/sse_mathfun.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/swish_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/tanh_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/unaryop_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/x86_activation.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/x86_usability.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/yolov3detectionoutput_x86.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/yolodetectionoutput.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer/yolov3detectionoutput.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer_declaration.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer_registry.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_registry.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_spv_data.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_type.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_type_enum.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer_type.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/layer_type_enum.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/mat.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/mat_pixel_drawing_font.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/modelbin.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/ncnn_export.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/net.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/option.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/paramdict.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/pipeline.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/pipelinecache.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/platform.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/ruapu.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/simplemath.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/simpleocv.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/simpleomp.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/simplestl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/simplevk.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/stb_image.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/stb_image_write.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-ncnn/vulkan_header_fix.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/alsa-play.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/alsa.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-ced-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-label-file.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-zipformer-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/base64-decode.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/cat.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/circular-buffer.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/context-graph.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/display.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/endpoint.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/fast-clustering-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/fast-clustering.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/features.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/file-utils.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/fst-utils.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/hypothesis.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/jieba-lexicon.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/keyword-spotter-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/keyword-spotter-transducer-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/keyword-spotter.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/lexicon.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/log.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/macros.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/math.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/melo-tts-lexicon.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/microphone.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ced-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ct-transformer-model-meta-data.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ct-transformer-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-fst-decoder-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-fst-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-greedy-search-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-lm-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-lm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-greedy-search-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-greedy-search-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation-ct-transformer-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-ctc-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-moonshine-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-paraformer-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-sense-voice-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-transducer-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-transducer-nemo-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-whisper-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-rnn-lm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-sense-voice-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-sense-voice-model-meta-data.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-sense-voice-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization-pyannote-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization-result.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model-meta-data.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-stream.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tdnn-ctc-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tdnn-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-telespeech-ctc-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-greedy-search-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-greedy-search-nemo-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-modified-beam-search-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-nemo-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-character-frontend.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-frontend.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-model-metadata.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-websocket-server-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-wenet-ctc-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-wenet-ctc-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-audio-tagging-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-audio-tagging-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-ctc-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-ctc-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-cnn-bilstm-model-meta-data.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-cnn-bilstm-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-conformer-transducer-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-fst-decoder-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-fst-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-greedy-search-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-lm-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-lm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-lstm-transducer-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-nemo-ctc-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-nemo-ctc-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-paraformer-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-paraformer-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-paraformer-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation-cnn-bilstm-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-ctc-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-paraformer-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-transducer-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-transducer-nemo-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-rnn-lm.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-stream.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-greedy-search-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-greedy-search-nemo-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-modified-beam-search-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-nemo-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-websocket-server-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-wenet-ctc-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-wenet-ctc-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer-transducer-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer2-ctc-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer2-ctc-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer2-transducer-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/onnx-utils.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/packed-sequence.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/pad-sequence.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/parse-options.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/piper-phonemize-lexicon.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/provider-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/provider.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/resample.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/session.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/silero-vad-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/silero-vad-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/slice.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-general-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-model-meta-data.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-nemo-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-nemo-model-meta-data.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-nemo-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-manager.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/spoken-language-identification-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/spoken-language-identification-whisper-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/spoken-language-identification.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/stack.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/symbol-table.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/tee-stream.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/text-utils.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/transducer-keyword-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/transpose.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/unbind.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/utils.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/vad-model-config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/vad-model.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/voice-activity-detector.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/wave-reader.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/csrc/wave-writer.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/decodable-ctc.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/decodable-itf.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/eigen.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/faster-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/hash-list-inl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/hash-list.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/kaldi-math.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/lattice-faster-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/lattice-simple-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/log.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/simple-decoder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/stl-utils.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/basic-filebuf.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/const-integer-set-inl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/const-integer-set.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/context-fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/deterministic-fst-inl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/deterministic-fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/determinize-star-inl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/determinize-star.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/fstext-utils-inl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/fstext-utils.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/grammar-context-fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/io-funcs-inl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/io-funcs.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-fst-io-inl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-fst-io.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-holder-inl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-holder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-io-inl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-io.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-math.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-pipebuf.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-semaphore.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-table-inl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-table.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/lattice-utils-inl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/lattice-utils.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/lattice-weight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/log.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/parse-options.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/pre-determinize-inl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/pre-determinize.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/remove-eps-local-inl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/remove-eps-local.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/stl-utils.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/table-matcher.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/text-normalizer.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/text-utils.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/cpu_provider_factory.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_c_api.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_cxx_api.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_cxx_inline.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_float16.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_run_options_config_keys.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_session_options_config_keys.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/provider_options.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/accumulator.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/add-on.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arc-arena.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arc-map.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arc.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arcfilter.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arcsort.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/bi-table.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/cache.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/closure.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compact-fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compat.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/complement.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compose-filter.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compose.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/concat.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/config.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/connect.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/const-fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/determinize.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/dfs-visit.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/difference.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/disambiguate.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/edit-fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/encode.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/epsnormalize.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/equal.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/equivalent.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/expanded-fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/expectation-weight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/compress-script.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/compress.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/elias.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/gzfile.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/randmod.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/compile-strings.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/create.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/equal.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/extract.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/far-class.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/far.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/farlib.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/farscript.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/getters.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/info.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/isomorphic.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/print-strings.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/script-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/stlist.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/sttable.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linear-fst-data-builder.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linear-fst-data.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linear-fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linearscript.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/loglinear-apply.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/trie.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/compose.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/expand.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/info.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/mpdt.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/mpdtlib.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/mpdtscript.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/read_write_utils.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/reverse.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/ngram/bitmap-index.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/ngram/ngram-fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/ngram/nthbit.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/collection.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/compose.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/expand.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/getters.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/info.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/paren.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/pdt.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/pdtlib.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/pdtscript.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/replace.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/reverse.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/shortest-path.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/special/phi-fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/special/rho-fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/special/sigma-fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/factor-weight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/filter-state.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/flags.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/float-weight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/fst-decl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/fstlib.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/generic-register.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/heap.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/icu.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/intersect.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/interval-set.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/invert.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/isomorphic.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/label-reachable.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lexicographic-weight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lock.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/log.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lookahead-filter.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lookahead-matcher.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/map.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/mapped-file.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/matcher-fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/matcher.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/memory.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/minimize.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/mutable-fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/pair-weight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/partition.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/power-weight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/product-weight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/project.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/properties.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/prune.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/push.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/queue.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/randequivalent.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/randgen.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/rational.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/register.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/relabel.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/replace-util.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/replace.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/reverse.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/reweight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/rmepsilon.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/rmfinalepsilon.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arc-class.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arciterator-class.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arcsort.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arg-packs.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/closure.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/compile-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/compile.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/compose.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/concat.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/connect.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/convert.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/decode.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/determinize.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/difference.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/disambiguate.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/draw-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/draw.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/encode.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/encodemapper-class.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/epsnormalize.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/equal.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/equivalent.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/fst-class.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/fstscript-decl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/fstscript.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/getters.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/info-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/info.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/intersect.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/invert.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/isomorphic.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/map.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/minimize.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/print-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/print.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/project.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/prune.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/push.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/randequivalent.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/randgen.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/register.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/relabel.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/replace.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/reverse.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/reweight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/rmepsilon.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/script-impl.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/shortest-distance.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/shortest-path.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/stateiterator-class.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/synchronize.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/text-io.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/topsort.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/union.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/verify.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/weight-class.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/set-weight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/shortest-distance.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/shortest-path.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/signed-log-weight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/sparse-power-weight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/sparse-tuple-weight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/state-map.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/state-reachable.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/state-table.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/statesort.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/string-weight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/string.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/symbol-table-ops.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/symbol-table.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/synchronize.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/test-properties.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/topsort.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/tuple-weight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/types.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/union-find.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/union-weight.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/union.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/util.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/vector-fst.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/verify.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/visit.h
 delete mode 100644 projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/weight.h
 delete mode 100644 projects/llm_framework/include/utils/checker.h
 delete mode 100644 projects/llm_framework/include/utils/io.hpp
 delete mode 100644 projects/llm_framework/include/utils/logger.h
 create mode 100644 projects/llm_framework/main_vlm/models/mode_internvl3-1B-448-ax650.json
 create mode 100644 projects/llm_framework/main_vlm/scripts/tokenizer_internvl3-1B-448-ax650.py

diff --git a/projects/llm_framework/SConstruct b/projects/llm_framework/SConstruct
index 6c4575b5..21134df7 100644
--- a/projects/llm_framework/SConstruct
+++ b/projects/llm_framework/SConstruct
@@ -5,7 +5,7 @@ import shutil
 os.environ['SDK_PATH'] = os.path.normpath(str(Path(os.getcwd())/'..'/'..'/'SDK'))
 os.environ['EXT_COMPONENTS_PATH'] = os.path.normpath(str(Path(os.getcwd())/'..'/'..'/'ext_components'))
 
-version = 'v0.0.9'
+version = 'v0.1.0'
 static_lib = 'static_lib'
 update = False
 
diff --git a/projects/llm_framework/include/abseil-cpp/absl/algorithm/algorithm.h b/projects/llm_framework/include/abseil-cpp/absl/algorithm/algorithm.h
deleted file mode 100644
index 59aeed7d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/algorithm/algorithm.h
+++ /dev/null
@@ -1,64 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: algorithm.h
-// -----------------------------------------------------------------------------
-//
-// This header file contains Google extensions to the standard <algorithm> C++
-// header.
-
-#ifndef ABSL_ALGORITHM_ALGORITHM_H_
-#define ABSL_ALGORITHM_ALGORITHM_H_
-
-#include <algorithm>
-#include <iterator>
-#include <type_traits>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// equal()
-// rotate()
-//
-// Historical note: Abseil once provided implementations of these algorithms
-// prior to their adoption in C++14. New code should prefer to use the std
-// variants.
-//
-// See the documentation for the STL <algorithm> header for more information:
-// https://en.cppreference.com/w/cpp/header/algorithm
-using std::equal;
-using std::rotate;
-
-// linear_search()
-//
-// Performs a linear search for `value` using the iterator `first` up to
-// but not including `last`, returning true if [`first`, `last`) contains an
-// element equal to `value`.
-//
-// A linear search is of O(n) complexity which is guaranteed to make at most
-// n = (`last` - `first`) comparisons. A linear search over short containers
-// may be faster than a binary search, even when the container is sorted.
-template <typename InputIterator, typename EqualityComparable>
-bool linear_search(InputIterator first, InputIterator last,
-                   const EqualityComparable& value) {
-  return std::find(first, last, value) != last;
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_ALGORITHM_ALGORITHM_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/algorithm/container.h b/projects/llm_framework/include/abseil-cpp/absl/algorithm/container.h
deleted file mode 100644
index c7bafae1..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/algorithm/container.h
+++ /dev/null
@@ -1,1760 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: container.h
-// -----------------------------------------------------------------------------
-//
-// This header file provides Container-based versions of algorithmic functions
-// within the C++ standard library. The following standard library sets of
-// functions are covered within this file:
-//
-//   * Algorithmic <iterator> functions
-//   * Algorithmic <numeric> functions
-//   * <algorithm> functions
-//
-// The standard library functions operate on iterator ranges; the functions
-// within this API operate on containers, though many return iterator ranges.
-//
-// All functions within this API are named with a `c_` prefix. Calls such as
-// `absl::c_xx(container, ...) are equivalent to std:: functions such as
-// `std::xx(std::begin(cont), std::end(cont), ...)`. Functions that act on
-// iterators but not conceptually on iterator ranges (e.g. `std::iter_swap`)
-// have no equivalent here.
-//
-// For template parameter and variable naming, `C` indicates the container type
-// to which the function is applied, `Pred` indicates the predicate object type
-// to be used by the function and `T` indicates the applicable element type.
-
-#ifndef ABSL_ALGORITHM_CONTAINER_H_
-#define ABSL_ALGORITHM_CONTAINER_H_
-
-#include <algorithm>
-#include <cassert>
-#include <iterator>
-#include <numeric>
-#include <type_traits>
-#include <unordered_map>
-#include <unordered_set>
-#include <utility>
-#include <vector>
-
-#include "absl/algorithm/algorithm.h"
-#include "absl/base/macros.h"
-#include "absl/base/nullability.h"
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_algorithm_internal {
-
-// NOTE: it is important to defer to ADL lookup for building with C++ modules,
-// especially for headers like <valarray> which are not visible from this file
-// but specialize std::begin and std::end.
-using std::begin;
-using std::end;
-
-// The type of the iterator given by begin(c) (possibly std::begin(c)).
-// ContainerIter<const vector<T>> gives vector<T>::const_iterator,
-// while ContainerIter<vector<T>> gives vector<T>::iterator.
-template <typename C>
-using ContainerIter = decltype(begin(std::declval<C&>()));
-
-// An MSVC bug involving template parameter substitution requires us to use
-// decltype() here instead of just std::pair.
-template <typename C1, typename C2>
-using ContainerIterPairType =
-    decltype(std::make_pair(ContainerIter<C1>(), ContainerIter<C2>()));
-
-template <typename C>
-using ContainerDifferenceType = decltype(std::distance(
-    std::declval<ContainerIter<C>>(), std::declval<ContainerIter<C>>()));
-
-template <typename C>
-using ContainerPointerType =
-    typename std::iterator_traits<ContainerIter<C>>::pointer;
-
-// container_algorithm_internal::c_begin and
-// container_algorithm_internal::c_end are abbreviations for proper ADL
-// lookup of std::begin and std::end, i.e.
-//   using std::begin;
-//   using std::end;
-//   std::foo(begin(c), end(c));
-// becomes
-//   std::foo(container_algorithm_internal::begin(c),
-//            container_algorithm_internal::end(c));
-// These are meant for internal use only.
-
-template <typename C>
-ContainerIter<C> c_begin(C& c) {
-  return begin(c);
-}
-
-template <typename C>
-ContainerIter<C> c_end(C& c) {
-  return end(c);
-}
-
-template <typename T>
-struct IsUnorderedContainer : std::false_type {};
-
-template <class Key, class T, class Hash, class KeyEqual, class Allocator>
-struct IsUnorderedContainer<
-    std::unordered_map<Key, T, Hash, KeyEqual, Allocator>> : std::true_type {};
-
-template <class Key, class Hash, class KeyEqual, class Allocator>
-struct IsUnorderedContainer<std::unordered_set<Key, Hash, KeyEqual, Allocator>>
-    : std::true_type {};
-
-}  // namespace container_algorithm_internal
-
-// PUBLIC API
-
-//------------------------------------------------------------------------------
-// Abseil algorithm.h functions
-//------------------------------------------------------------------------------
-
-// c_linear_search()
-//
-// Container-based version of absl::linear_search() for performing a linear
-// search within a container.
-template <typename C, typename EqualityComparable>
-bool c_linear_search(const C& c, EqualityComparable&& value) {
-  return linear_search(container_algorithm_internal::c_begin(c),
-                       container_algorithm_internal::c_end(c),
-                       std::forward<EqualityComparable>(value));
-}
-
-//------------------------------------------------------------------------------
-// <iterator> algorithms
-//------------------------------------------------------------------------------
-
-// c_distance()
-//
-// Container-based version of the <iterator> `std::distance()` function to
-// return the number of elements within a container.
-template <typename C>
-container_algorithm_internal::ContainerDifferenceType<const C> c_distance(
-    const C& c) {
-  return std::distance(container_algorithm_internal::c_begin(c),
-                       container_algorithm_internal::c_end(c));
-}
-
-//------------------------------------------------------------------------------
-// <algorithm> Non-modifying sequence operations
-//------------------------------------------------------------------------------
-
-// c_all_of()
-//
-// Container-based version of the <algorithm> `std::all_of()` function to
-// test if all elements within a container satisfy a condition.
-template <typename C, typename Pred>
-bool c_all_of(const C& c, Pred&& pred) {
-  return std::all_of(container_algorithm_internal::c_begin(c),
-                     container_algorithm_internal::c_end(c),
-                     std::forward<Pred>(pred));
-}
-
-// c_any_of()
-//
-// Container-based version of the <algorithm> `std::any_of()` function to
-// test if any element in a container fulfills a condition.
-template <typename C, typename Pred>
-bool c_any_of(const C& c, Pred&& pred) {
-  return std::any_of(container_algorithm_internal::c_begin(c),
-                     container_algorithm_internal::c_end(c),
-                     std::forward<Pred>(pred));
-}
-
-// c_none_of()
-//
-// Container-based version of the <algorithm> `std::none_of()` function to
-// test if no elements in a container fulfill a condition.
-template <typename C, typename Pred>
-bool c_none_of(const C& c, Pred&& pred) {
-  return std::none_of(container_algorithm_internal::c_begin(c),
-                      container_algorithm_internal::c_end(c),
-                      std::forward<Pred>(pred));
-}
-
-// c_for_each()
-//
-// Container-based version of the <algorithm> `std::for_each()` function to
-// apply a function to a container's elements.
-template <typename C, typename Function>
-decay_t<Function> c_for_each(C&& c, Function&& f) {
-  return std::for_each(container_algorithm_internal::c_begin(c),
-                       container_algorithm_internal::c_end(c),
-                       std::forward<Function>(f));
-}
-
-// c_find()
-//
-// Container-based version of the <algorithm> `std::find()` function to find
-// the first element containing the passed value within a container value.
-template <typename C, typename T>
-container_algorithm_internal::ContainerIter<C> c_find(C& c, T&& value) {
-  return std::find(container_algorithm_internal::c_begin(c),
-                   container_algorithm_internal::c_end(c),
-                   std::forward<T>(value));
-}
-
-// c_find_if()
-//
-// Container-based version of the <algorithm> `std::find_if()` function to find
-// the first element in a container matching the given condition.
-template <typename C, typename Pred>
-container_algorithm_internal::ContainerIter<C> c_find_if(C& c, Pred&& pred) {
-  return std::find_if(container_algorithm_internal::c_begin(c),
-                      container_algorithm_internal::c_end(c),
-                      std::forward<Pred>(pred));
-}
-
-// c_find_if_not()
-//
-// Container-based version of the <algorithm> `std::find_if_not()` function to
-// find the first element in a container not matching the given condition.
-template <typename C, typename Pred>
-container_algorithm_internal::ContainerIter<C> c_find_if_not(C& c,
-                                                             Pred&& pred) {
-  return std::find_if_not(container_algorithm_internal::c_begin(c),
-                          container_algorithm_internal::c_end(c),
-                          std::forward<Pred>(pred));
-}
-
-// c_find_end()
-//
-// Container-based version of the <algorithm> `std::find_end()` function to
-// find the last subsequence within a container.
-template <typename Sequence1, typename Sequence2>
-container_algorithm_internal::ContainerIter<Sequence1> c_find_end(
-    Sequence1& sequence, Sequence2& subsequence) {
-  return std::find_end(container_algorithm_internal::c_begin(sequence),
-                       container_algorithm_internal::c_end(sequence),
-                       container_algorithm_internal::c_begin(subsequence),
-                       container_algorithm_internal::c_end(subsequence));
-}
-
-// Overload of c_find_end() for using a predicate evaluation other than `==` as
-// the function's test condition.
-template <typename Sequence1, typename Sequence2, typename BinaryPredicate>
-container_algorithm_internal::ContainerIter<Sequence1> c_find_end(
-    Sequence1& sequence, Sequence2& subsequence, BinaryPredicate&& pred) {
-  return std::find_end(container_algorithm_internal::c_begin(sequence),
-                       container_algorithm_internal::c_end(sequence),
-                       container_algorithm_internal::c_begin(subsequence),
-                       container_algorithm_internal::c_end(subsequence),
-                       std::forward<BinaryPredicate>(pred));
-}
-
-// c_find_first_of()
-//
-// Container-based version of the <algorithm> `std::find_first_of()` function to
-// find the first element within the container that is also within the options
-// container.
-template <typename C1, typename C2>
-container_algorithm_internal::ContainerIter<C1> c_find_first_of(C1& container,
-                                                                C2& options) {
-  return std::find_first_of(container_algorithm_internal::c_begin(container),
-                            container_algorithm_internal::c_end(container),
-                            container_algorithm_internal::c_begin(options),
-                            container_algorithm_internal::c_end(options));
-}
-
-// Overload of c_find_first_of() for using a predicate evaluation other than
-// `==` as the function's test condition.
-template <typename C1, typename C2, typename BinaryPredicate>
-container_algorithm_internal::ContainerIter<C1> c_find_first_of(
-    C1& container, C2& options, BinaryPredicate&& pred) {
-  return std::find_first_of(container_algorithm_internal::c_begin(container),
-                            container_algorithm_internal::c_end(container),
-                            container_algorithm_internal::c_begin(options),
-                            container_algorithm_internal::c_end(options),
-                            std::forward<BinaryPredicate>(pred));
-}
-
-// c_adjacent_find()
-//
-// Container-based version of the <algorithm> `std::adjacent_find()` function to
-// find equal adjacent elements within a container.
-template <typename Sequence>
-container_algorithm_internal::ContainerIter<Sequence> c_adjacent_find(
-    Sequence& sequence) {
-  return std::adjacent_find(container_algorithm_internal::c_begin(sequence),
-                            container_algorithm_internal::c_end(sequence));
-}
-
-// Overload of c_adjacent_find() for using a predicate evaluation other than
-// `==` as the function's test condition.
-template <typename Sequence, typename BinaryPredicate>
-container_algorithm_internal::ContainerIter<Sequence> c_adjacent_find(
-    Sequence& sequence, BinaryPredicate&& pred) {
-  return std::adjacent_find(container_algorithm_internal::c_begin(sequence),
-                            container_algorithm_internal::c_end(sequence),
-                            std::forward<BinaryPredicate>(pred));
-}
-
-// c_count()
-//
-// Container-based version of the <algorithm> `std::count()` function to count
-// values that match within a container.
-template <typename C, typename T>
-container_algorithm_internal::ContainerDifferenceType<const C> c_count(
-    const C& c, T&& value) {
-  return std::count(container_algorithm_internal::c_begin(c),
-                    container_algorithm_internal::c_end(c),
-                    std::forward<T>(value));
-}
-
-// c_count_if()
-//
-// Container-based version of the <algorithm> `std::count_if()` function to
-// count values matching a condition within a container.
-template <typename C, typename Pred>
-container_algorithm_internal::ContainerDifferenceType<const C> c_count_if(
-    const C& c, Pred&& pred) {
-  return std::count_if(container_algorithm_internal::c_begin(c),
-                       container_algorithm_internal::c_end(c),
-                       std::forward<Pred>(pred));
-}
-
-// c_mismatch()
-//
-// Container-based version of the <algorithm> `std::mismatch()` function to
-// return the first element where two ordered containers differ. Applies `==` to
-// the first N elements of `c1` and `c2`, where N = min(size(c1), size(c2)).
-template <typename C1, typename C2>
-container_algorithm_internal::ContainerIterPairType<C1, C2> c_mismatch(C1& c1,
-                                                                       C2& c2) {
-  return std::mismatch(container_algorithm_internal::c_begin(c1),
-                       container_algorithm_internal::c_end(c1),
-                       container_algorithm_internal::c_begin(c2),
-                       container_algorithm_internal::c_end(c2));
-}
-
-// Overload of c_mismatch() for using a predicate evaluation other than `==` as
-// the function's test condition. Applies `pred`to the first N elements of `c1`
-// and `c2`, where N = min(size(c1), size(c2)).
-template <typename C1, typename C2, typename BinaryPredicate>
-container_algorithm_internal::ContainerIterPairType<C1, C2> c_mismatch(
-    C1& c1, C2& c2, BinaryPredicate pred) {
-  return std::mismatch(container_algorithm_internal::c_begin(c1),
-                       container_algorithm_internal::c_end(c1),
-                       container_algorithm_internal::c_begin(c2),
-                       container_algorithm_internal::c_end(c2), pred);
-}
-
-// c_equal()
-//
-// Container-based version of the <algorithm> `std::equal()` function to
-// test whether two containers are equal.
-template <typename C1, typename C2>
-bool c_equal(const C1& c1, const C2& c2) {
-  return std::equal(container_algorithm_internal::c_begin(c1),
-                    container_algorithm_internal::c_end(c1),
-                    container_algorithm_internal::c_begin(c2),
-                    container_algorithm_internal::c_end(c2));
-}
-
-// Overload of c_equal() for using a predicate evaluation other than `==` as
-// the function's test condition.
-template <typename C1, typename C2, typename BinaryPredicate>
-bool c_equal(const C1& c1, const C2& c2, BinaryPredicate&& pred) {
-  return std::equal(container_algorithm_internal::c_begin(c1),
-                    container_algorithm_internal::c_end(c1),
-                    container_algorithm_internal::c_begin(c2),
-                    container_algorithm_internal::c_end(c2),
-                    std::forward<BinaryPredicate>(pred));
-}
-
-// c_is_permutation()
-//
-// Container-based version of the <algorithm> `std::is_permutation()` function
-// to test whether a container is a permutation of another.
-template <typename C1, typename C2>
-bool c_is_permutation(const C1& c1, const C2& c2) {
-  return std::is_permutation(container_algorithm_internal::c_begin(c1),
-                             container_algorithm_internal::c_end(c1),
-                             container_algorithm_internal::c_begin(c2),
-                             container_algorithm_internal::c_end(c2));
-}
-
-// Overload of c_is_permutation() for using a predicate evaluation other than
-// `==` as the function's test condition.
-template <typename C1, typename C2, typename BinaryPredicate>
-bool c_is_permutation(const C1& c1, const C2& c2, BinaryPredicate&& pred) {
-  return std::is_permutation(container_algorithm_internal::c_begin(c1),
-                             container_algorithm_internal::c_end(c1),
-                             container_algorithm_internal::c_begin(c2),
-                             container_algorithm_internal::c_end(c2),
-                             std::forward<BinaryPredicate>(pred));
-}
-
-// c_search()
-//
-// Container-based version of the <algorithm> `std::search()` function to search
-// a container for a subsequence.
-template <typename Sequence1, typename Sequence2>
-container_algorithm_internal::ContainerIter<Sequence1> c_search(
-    Sequence1& sequence, Sequence2& subsequence) {
-  return std::search(container_algorithm_internal::c_begin(sequence),
-                     container_algorithm_internal::c_end(sequence),
-                     container_algorithm_internal::c_begin(subsequence),
-                     container_algorithm_internal::c_end(subsequence));
-}
-
-// Overload of c_search() for using a predicate evaluation other than
-// `==` as the function's test condition.
-template <typename Sequence1, typename Sequence2, typename BinaryPredicate>
-container_algorithm_internal::ContainerIter<Sequence1> c_search(
-    Sequence1& sequence, Sequence2& subsequence, BinaryPredicate&& pred) {
-  return std::search(container_algorithm_internal::c_begin(sequence),
-                     container_algorithm_internal::c_end(sequence),
-                     container_algorithm_internal::c_begin(subsequence),
-                     container_algorithm_internal::c_end(subsequence),
-                     std::forward<BinaryPredicate>(pred));
-}
-
-// c_search_n()
-//
-// Container-based version of the <algorithm> `std::search_n()` function to
-// search a container for the first sequence of N elements.
-template <typename Sequence, typename Size, typename T>
-container_algorithm_internal::ContainerIter<Sequence> c_search_n(
-    Sequence& sequence, Size count, T&& value) {
-  return std::search_n(container_algorithm_internal::c_begin(sequence),
-                       container_algorithm_internal::c_end(sequence), count,
-                       std::forward<T>(value));
-}
-
-// Overload of c_search_n() for using a predicate evaluation other than
-// `==` as the function's test condition.
-template <typename Sequence, typename Size, typename T,
-          typename BinaryPredicate>
-container_algorithm_internal::ContainerIter<Sequence> c_search_n(
-    Sequence& sequence, Size count, T&& value, BinaryPredicate&& pred) {
-  return std::search_n(container_algorithm_internal::c_begin(sequence),
-                       container_algorithm_internal::c_end(sequence), count,
-                       std::forward<T>(value),
-                       std::forward<BinaryPredicate>(pred));
-}
-
-//------------------------------------------------------------------------------
-// <algorithm> Modifying sequence operations
-//------------------------------------------------------------------------------
-
-// c_copy()
-//
-// Container-based version of the <algorithm> `std::copy()` function to copy a
-// container's elements into an iterator.
-template <typename InputSequence, typename OutputIterator>
-OutputIterator c_copy(const InputSequence& input, OutputIterator output) {
-  return std::copy(container_algorithm_internal::c_begin(input),
-                   container_algorithm_internal::c_end(input), output);
-}
-
-// c_copy_n()
-//
-// Container-based version of the <algorithm> `std::copy_n()` function to copy a
-// container's first N elements into an iterator.
-template <typename C, typename Size, typename OutputIterator>
-OutputIterator c_copy_n(const C& input, Size n, OutputIterator output) {
-  return std::copy_n(container_algorithm_internal::c_begin(input), n, output);
-}
-
-// c_copy_if()
-//
-// Container-based version of the <algorithm> `std::copy_if()` function to copy
-// a container's elements satisfying some condition into an iterator.
-template <typename InputSequence, typename OutputIterator, typename Pred>
-OutputIterator c_copy_if(const InputSequence& input, OutputIterator output,
-                         Pred&& pred) {
-  return std::copy_if(container_algorithm_internal::c_begin(input),
-                      container_algorithm_internal::c_end(input), output,
-                      std::forward<Pred>(pred));
-}
-
-// c_copy_backward()
-//
-// Container-based version of the <algorithm> `std::copy_backward()` function to
-// copy a container's elements in reverse order into an iterator.
-template <typename C, typename BidirectionalIterator>
-BidirectionalIterator c_copy_backward(const C& src,
-                                      BidirectionalIterator dest) {
-  return std::copy_backward(container_algorithm_internal::c_begin(src),
-                            container_algorithm_internal::c_end(src), dest);
-}
-
-// c_move()
-//
-// Container-based version of the <algorithm> `std::move()` function to move
-// a container's elements into an iterator.
-template <typename C, typename OutputIterator>
-OutputIterator c_move(C&& src, OutputIterator dest) {
-  return std::move(container_algorithm_internal::c_begin(src),
-                   container_algorithm_internal::c_end(src), dest);
-}
-
-// c_move_backward()
-//
-// Container-based version of the <algorithm> `std::move_backward()` function to
-// move a container's elements into an iterator in reverse order.
-template <typename C, typename BidirectionalIterator>
-BidirectionalIterator c_move_backward(C&& src, BidirectionalIterator dest) {
-  return std::move_backward(container_algorithm_internal::c_begin(src),
-                            container_algorithm_internal::c_end(src), dest);
-}
-
-// c_swap_ranges()
-//
-// Container-based version of the <algorithm> `std::swap_ranges()` function to
-// swap a container's elements with another container's elements. Swaps the
-// first N elements of `c1` and `c2`, where N = min(size(c1), size(c2)).
-template <typename C1, typename C2>
-container_algorithm_internal::ContainerIter<C2> c_swap_ranges(C1& c1, C2& c2) {
-  auto first1 = container_algorithm_internal::c_begin(c1);
-  auto last1 = container_algorithm_internal::c_end(c1);
-  auto first2 = container_algorithm_internal::c_begin(c2);
-  auto last2 = container_algorithm_internal::c_end(c2);
-
-  using std::swap;
-  for (; first1 != last1 && first2 != last2; ++first1, (void)++first2) {
-    swap(*first1, *first2);
-  }
-  return first2;
-}
-
-// c_transform()
-//
-// Container-based version of the <algorithm> `std::transform()` function to
-// transform a container's elements using the unary operation, storing the
-// result in an iterator pointing to the last transformed element in the output
-// range.
-template <typename InputSequence, typename OutputIterator, typename UnaryOp>
-OutputIterator c_transform(const InputSequence& input, OutputIterator output,
-                           UnaryOp&& unary_op) {
-  return std::transform(container_algorithm_internal::c_begin(input),
-                        container_algorithm_internal::c_end(input), output,
-                        std::forward<UnaryOp>(unary_op));
-}
-
-// Overload of c_transform() for performing a transformation using a binary
-// predicate. Applies `binary_op` to the first N elements of `c1` and `c2`,
-// where N = min(size(c1), size(c2)).
-template <typename InputSequence1, typename InputSequence2,
-          typename OutputIterator, typename BinaryOp>
-OutputIterator c_transform(const InputSequence1& input1,
-                           const InputSequence2& input2, OutputIterator output,
-                           BinaryOp&& binary_op) {
-  auto first1 = container_algorithm_internal::c_begin(input1);
-  auto last1 = container_algorithm_internal::c_end(input1);
-  auto first2 = container_algorithm_internal::c_begin(input2);
-  auto last2 = container_algorithm_internal::c_end(input2);
-  for (; first1 != last1 && first2 != last2;
-       ++first1, (void)++first2, ++output) {
-    *output = binary_op(*first1, *first2);
-  }
-
-  return output;
-}
-
-// c_replace()
-//
-// Container-based version of the <algorithm> `std::replace()` function to
-// replace a container's elements of some value with a new value. The container
-// is modified in place.
-template <typename Sequence, typename T>
-void c_replace(Sequence& sequence, const T& old_value, const T& new_value) {
-  std::replace(container_algorithm_internal::c_begin(sequence),
-               container_algorithm_internal::c_end(sequence), old_value,
-               new_value);
-}
-
-// c_replace_if()
-//
-// Container-based version of the <algorithm> `std::replace_if()` function to
-// replace a container's elements of some value with a new value based on some
-// condition. The container is modified in place.
-template <typename C, typename Pred, typename T>
-void c_replace_if(C& c, Pred&& pred, T&& new_value) {
-  std::replace_if(container_algorithm_internal::c_begin(c),
-                  container_algorithm_internal::c_end(c),
-                  std::forward<Pred>(pred), std::forward<T>(new_value));
-}
-
-// c_replace_copy()
-//
-// Container-based version of the <algorithm> `std::replace_copy()` function to
-// replace a container's elements of some value with a new value  and return the
-// results within an iterator.
-template <typename C, typename OutputIterator, typename T>
-OutputIterator c_replace_copy(const C& c, OutputIterator result, T&& old_value,
-                              T&& new_value) {
-  return std::replace_copy(container_algorithm_internal::c_begin(c),
-                           container_algorithm_internal::c_end(c), result,
-                           std::forward<T>(old_value),
-                           std::forward<T>(new_value));
-}
-
-// c_replace_copy_if()
-//
-// Container-based version of the <algorithm> `std::replace_copy_if()` function
-// to replace a container's elements of some value with a new value based on
-// some condition, and return the results within an iterator.
-template <typename C, typename OutputIterator, typename Pred, typename T>
-OutputIterator c_replace_copy_if(const C& c, OutputIterator result, Pred&& pred,
-                                 const T& new_value) {
-  return std::replace_copy_if(container_algorithm_internal::c_begin(c),
-                              container_algorithm_internal::c_end(c), result,
-                              std::forward<Pred>(pred), new_value);
-}
-
-// c_fill()
-//
-// Container-based version of the <algorithm> `std::fill()` function to fill a
-// container with some value.
-template <typename C, typename T>
-void c_fill(C& c, const T& value) {
-  std::fill(container_algorithm_internal::c_begin(c),
-            container_algorithm_internal::c_end(c), value);
-}
-
-// c_fill_n()
-//
-// Container-based version of the <algorithm> `std::fill_n()` function to fill
-// the first N elements in a container with some value.
-template <typename C, typename Size, typename T>
-void c_fill_n(C& c, Size n, const T& value) {
-  std::fill_n(container_algorithm_internal::c_begin(c), n, value);
-}
-
-// c_generate()
-//
-// Container-based version of the <algorithm> `std::generate()` function to
-// assign a container's elements to the values provided by the given generator.
-template <typename C, typename Generator>
-void c_generate(C& c, Generator&& gen) {
-  std::generate(container_algorithm_internal::c_begin(c),
-                container_algorithm_internal::c_end(c),
-                std::forward<Generator>(gen));
-}
-
-// c_generate_n()
-//
-// Container-based version of the <algorithm> `std::generate_n()` function to
-// assign a container's first N elements to the values provided by the given
-// generator.
-template <typename C, typename Size, typename Generator>
-container_algorithm_internal::ContainerIter<C> c_generate_n(C& c, Size n,
-                                                            Generator&& gen) {
-  return std::generate_n(container_algorithm_internal::c_begin(c), n,
-                         std::forward<Generator>(gen));
-}
-
-// Note: `c_xx()` <algorithm> container versions for `remove()`, `remove_if()`,
-// and `unique()` are omitted, because it's not clear whether or not such
-// functions should call erase on their supplied sequences afterwards. Either
-// behavior would be surprising for a different set of users.
-
-// c_remove_copy()
-//
-// Container-based version of the <algorithm> `std::remove_copy()` function to
-// copy a container's elements while removing any elements matching the given
-// `value`.
-template <typename C, typename OutputIterator, typename T>
-OutputIterator c_remove_copy(const C& c, OutputIterator result,
-                             const T& value) {
-  return std::remove_copy(container_algorithm_internal::c_begin(c),
-                          container_algorithm_internal::c_end(c), result,
-                          value);
-}
-
-// c_remove_copy_if()
-//
-// Container-based version of the <algorithm> `std::remove_copy_if()` function
-// to copy a container's elements while removing any elements matching the given
-// condition.
-template <typename C, typename OutputIterator, typename Pred>
-OutputIterator c_remove_copy_if(const C& c, OutputIterator result,
-                                Pred&& pred) {
-  return std::remove_copy_if(container_algorithm_internal::c_begin(c),
-                             container_algorithm_internal::c_end(c), result,
-                             std::forward<Pred>(pred));
-}
-
-// c_unique_copy()
-//
-// Container-based version of the <algorithm> `std::unique_copy()` function to
-// copy a container's elements while removing any elements containing duplicate
-// values.
-template <typename C, typename OutputIterator>
-OutputIterator c_unique_copy(const C& c, OutputIterator result) {
-  return std::unique_copy(container_algorithm_internal::c_begin(c),
-                          container_algorithm_internal::c_end(c), result);
-}
-
-// Overload of c_unique_copy() for using a predicate evaluation other than
-// `==` for comparing uniqueness of the element values.
-template <typename C, typename OutputIterator, typename BinaryPredicate>
-OutputIterator c_unique_copy(const C& c, OutputIterator result,
-                             BinaryPredicate&& pred) {
-  return std::unique_copy(container_algorithm_internal::c_begin(c),
-                          container_algorithm_internal::c_end(c), result,
-                          std::forward<BinaryPredicate>(pred));
-}
-
-// c_reverse()
-//
-// Container-based version of the <algorithm> `std::reverse()` function to
-// reverse a container's elements.
-template <typename Sequence>
-void c_reverse(Sequence& sequence) {
-  std::reverse(container_algorithm_internal::c_begin(sequence),
-               container_algorithm_internal::c_end(sequence));
-}
-
-// c_reverse_copy()
-//
-// Container-based version of the <algorithm> `std::reverse()` function to
-// reverse a container's elements and write them to an iterator range.
-template <typename C, typename OutputIterator>
-OutputIterator c_reverse_copy(const C& sequence, OutputIterator result) {
-  return std::reverse_copy(container_algorithm_internal::c_begin(sequence),
-                           container_algorithm_internal::c_end(sequence),
-                           result);
-}
-
-// c_rotate()
-//
-// Container-based version of the <algorithm> `std::rotate()` function to
-// shift a container's elements leftward such that the `middle` element becomes
-// the first element in the container.
-template <typename C,
-          typename Iterator = container_algorithm_internal::ContainerIter<C>>
-Iterator c_rotate(C& sequence, Iterator middle) {
-  return absl::rotate(container_algorithm_internal::c_begin(sequence), middle,
-                      container_algorithm_internal::c_end(sequence));
-}
-
-// c_rotate_copy()
-//
-// Container-based version of the <algorithm> `std::rotate_copy()` function to
-// shift a container's elements leftward such that the `middle` element becomes
-// the first element in a new iterator range.
-template <typename C, typename OutputIterator>
-OutputIterator c_rotate_copy(
-    const C& sequence,
-    container_algorithm_internal::ContainerIter<const C> middle,
-    OutputIterator result) {
-  return std::rotate_copy(container_algorithm_internal::c_begin(sequence),
-                          middle, container_algorithm_internal::c_end(sequence),
-                          result);
-}
-
-// c_shuffle()
-//
-// Container-based version of the <algorithm> `std::shuffle()` function to
-// randomly shuffle elements within the container using a `gen()` uniform random
-// number generator.
-template <typename RandomAccessContainer, typename UniformRandomBitGenerator>
-void c_shuffle(RandomAccessContainer& c, UniformRandomBitGenerator&& gen) {
-  std::shuffle(container_algorithm_internal::c_begin(c),
-               container_algorithm_internal::c_end(c),
-               std::forward<UniformRandomBitGenerator>(gen));
-}
-
-// c_sample()
-//
-// Container-based version of the <algorithm> `std::sample()` function to
-// randomly sample elements from the container without replacement using a
-// `gen()` uniform random number generator and write them to an iterator range.
-template <typename C, typename OutputIterator, typename Distance,
-          typename UniformRandomBitGenerator>
-OutputIterator c_sample(const C& c, OutputIterator result, Distance n,
-                        UniformRandomBitGenerator&& gen) {
-#if defined(__cpp_lib_sample) && __cpp_lib_sample >= 201603L
-  return std::sample(container_algorithm_internal::c_begin(c),
-                     container_algorithm_internal::c_end(c), result, n,
-                     std::forward<UniformRandomBitGenerator>(gen));
-#else
-  // Fall back to a stable selection-sampling implementation.
-  auto first = container_algorithm_internal::c_begin(c);
-  Distance unsampled_elements = c_distance(c);
-  n = (std::min)(n, unsampled_elements);
-  for (; n != 0; ++first) {
-    Distance r =
-        std::uniform_int_distribution<Distance>(0, --unsampled_elements)(gen);
-    if (r < n) {
-      *result++ = *first;
-      --n;
-    }
-  }
-  return result;
-#endif
-}
-
-//------------------------------------------------------------------------------
-// <algorithm> Partition functions
-//------------------------------------------------------------------------------
-
-// c_is_partitioned()
-//
-// Container-based version of the <algorithm> `std::is_partitioned()` function
-// to test whether all elements in the container for which `pred` returns `true`
-// precede those for which `pred` is `false`.
-template <typename C, typename Pred>
-bool c_is_partitioned(const C& c, Pred&& pred) {
-  return std::is_partitioned(container_algorithm_internal::c_begin(c),
-                             container_algorithm_internal::c_end(c),
-                             std::forward<Pred>(pred));
-}
-
-// c_partition()
-//
-// Container-based version of the <algorithm> `std::partition()` function
-// to rearrange all elements in a container in such a way that all elements for
-// which `pred` returns `true` precede all those for which it returns `false`,
-// returning an iterator to the first element of the second group.
-template <typename C, typename Pred>
-container_algorithm_internal::ContainerIter<C> c_partition(C& c, Pred&& pred) {
-  return std::partition(container_algorithm_internal::c_begin(c),
-                        container_algorithm_internal::c_end(c),
-                        std::forward<Pred>(pred));
-}
-
-// c_stable_partition()
-//
-// Container-based version of the <algorithm> `std::stable_partition()` function
-// to rearrange all elements in a container in such a way that all elements for
-// which `pred` returns `true` precede all those for which it returns `false`,
-// preserving the relative ordering between the two groups. The function returns
-// an iterator to the first element of the second group.
-template <typename C, typename Pred>
-container_algorithm_internal::ContainerIter<C> c_stable_partition(C& c,
-                                                                  Pred&& pred) {
-  return std::stable_partition(container_algorithm_internal::c_begin(c),
-                               container_algorithm_internal::c_end(c),
-                               std::forward<Pred>(pred));
-}
-
-// c_partition_copy()
-//
-// Container-based version of the <algorithm> `std::partition_copy()` function
-// to partition a container's elements and return them into two iterators: one
-// for which `pred` returns `true`, and one for which `pred` returns `false.`
-
-template <typename C, typename OutputIterator1, typename OutputIterator2,
-          typename Pred>
-std::pair<OutputIterator1, OutputIterator2> c_partition_copy(
-    const C& c, OutputIterator1 out_true, OutputIterator2 out_false,
-    Pred&& pred) {
-  return std::partition_copy(container_algorithm_internal::c_begin(c),
-                             container_algorithm_internal::c_end(c), out_true,
-                             out_false, std::forward<Pred>(pred));
-}
-
-// c_partition_point()
-//
-// Container-based version of the <algorithm> `std::partition_point()` function
-// to return the first element of an already partitioned container for which
-// the given `pred` is not `true`.
-template <typename C, typename Pred>
-container_algorithm_internal::ContainerIter<C> c_partition_point(C& c,
-                                                                 Pred&& pred) {
-  return std::partition_point(container_algorithm_internal::c_begin(c),
-                              container_algorithm_internal::c_end(c),
-                              std::forward<Pred>(pred));
-}
-
-//------------------------------------------------------------------------------
-// <algorithm> Sorting functions
-//------------------------------------------------------------------------------
-
-// c_sort()
-//
-// Container-based version of the <algorithm> `std::sort()` function
-// to sort elements in ascending order of their values.
-template <typename C>
-void c_sort(C& c) {
-  std::sort(container_algorithm_internal::c_begin(c),
-            container_algorithm_internal::c_end(c));
-}
-
-// Overload of c_sort() for performing a `comp` comparison other than the
-// default `operator<`.
-template <typename C, typename LessThan>
-void c_sort(C& c, LessThan&& comp) {
-  std::sort(container_algorithm_internal::c_begin(c),
-            container_algorithm_internal::c_end(c),
-            std::forward<LessThan>(comp));
-}
-
-// c_stable_sort()
-//
-// Container-based version of the <algorithm> `std::stable_sort()` function
-// to sort elements in ascending order of their values, preserving the order
-// of equivalents.
-template <typename C>
-void c_stable_sort(C& c) {
-  std::stable_sort(container_algorithm_internal::c_begin(c),
-                   container_algorithm_internal::c_end(c));
-}
-
-// Overload of c_stable_sort() for performing a `comp` comparison other than the
-// default `operator<`.
-template <typename C, typename LessThan>
-void c_stable_sort(C& c, LessThan&& comp) {
-  std::stable_sort(container_algorithm_internal::c_begin(c),
-                   container_algorithm_internal::c_end(c),
-                   std::forward<LessThan>(comp));
-}
-
-// c_is_sorted()
-//
-// Container-based version of the <algorithm> `std::is_sorted()` function
-// to evaluate whether the given container is sorted in ascending order.
-template <typename C>
-bool c_is_sorted(const C& c) {
-  return std::is_sorted(container_algorithm_internal::c_begin(c),
-                        container_algorithm_internal::c_end(c));
-}
-
-// c_is_sorted() overload for performing a `comp` comparison other than the
-// default `operator<`.
-template <typename C, typename LessThan>
-bool c_is_sorted(const C& c, LessThan&& comp) {
-  return std::is_sorted(container_algorithm_internal::c_begin(c),
-                        container_algorithm_internal::c_end(c),
-                        std::forward<LessThan>(comp));
-}
-
-// c_partial_sort()
-//
-// Container-based version of the <algorithm> `std::partial_sort()` function
-// to rearrange elements within a container such that elements before `middle`
-// are sorted in ascending order.
-template <typename RandomAccessContainer>
-void c_partial_sort(
-    RandomAccessContainer& sequence,
-    container_algorithm_internal::ContainerIter<RandomAccessContainer> middle) {
-  std::partial_sort(container_algorithm_internal::c_begin(sequence), middle,
-                    container_algorithm_internal::c_end(sequence));
-}
-
-// Overload of c_partial_sort() for performing a `comp` comparison other than
-// the default `operator<`.
-template <typename RandomAccessContainer, typename LessThan>
-void c_partial_sort(
-    RandomAccessContainer& sequence,
-    container_algorithm_internal::ContainerIter<RandomAccessContainer> middle,
-    LessThan&& comp) {
-  std::partial_sort(container_algorithm_internal::c_begin(sequence), middle,
-                    container_algorithm_internal::c_end(sequence),
-                    std::forward<LessThan>(comp));
-}
-
-// c_partial_sort_copy()
-//
-// Container-based version of the <algorithm> `std::partial_sort_copy()`
-// function to sort the elements in the given range `result` within the larger
-// `sequence` in ascending order (and using `result` as the output parameter).
-// At most min(result.last - result.first, sequence.last - sequence.first)
-// elements from the sequence will be stored in the result.
-template <typename C, typename RandomAccessContainer>
-container_algorithm_internal::ContainerIter<RandomAccessContainer>
-c_partial_sort_copy(const C& sequence, RandomAccessContainer& result) {
-  return std::partial_sort_copy(container_algorithm_internal::c_begin(sequence),
-                                container_algorithm_internal::c_end(sequence),
-                                container_algorithm_internal::c_begin(result),
-                                container_algorithm_internal::c_end(result));
-}
-
-// Overload of c_partial_sort_copy() for performing a `comp` comparison other
-// than the default `operator<`.
-template <typename C, typename RandomAccessContainer, typename LessThan>
-container_algorithm_internal::ContainerIter<RandomAccessContainer>
-c_partial_sort_copy(const C& sequence, RandomAccessContainer& result,
-                    LessThan&& comp) {
-  return std::partial_sort_copy(container_algorithm_internal::c_begin(sequence),
-                                container_algorithm_internal::c_end(sequence),
-                                container_algorithm_internal::c_begin(result),
-                                container_algorithm_internal::c_end(result),
-                                std::forward<LessThan>(comp));
-}
-
-// c_is_sorted_until()
-//
-// Container-based version of the <algorithm> `std::is_sorted_until()` function
-// to return the first element within a container that is not sorted in
-// ascending order as an iterator.
-template <typename C>
-container_algorithm_internal::ContainerIter<C> c_is_sorted_until(C& c) {
-  return std::is_sorted_until(container_algorithm_internal::c_begin(c),
-                              container_algorithm_internal::c_end(c));
-}
-
-// Overload of c_is_sorted_until() for performing a `comp` comparison other than
-// the default `operator<`.
-template <typename C, typename LessThan>
-container_algorithm_internal::ContainerIter<C> c_is_sorted_until(
-    C& c, LessThan&& comp) {
-  return std::is_sorted_until(container_algorithm_internal::c_begin(c),
-                              container_algorithm_internal::c_end(c),
-                              std::forward<LessThan>(comp));
-}
-
-// c_nth_element()
-//
-// Container-based version of the <algorithm> `std::nth_element()` function
-// to rearrange the elements within a container such that the `nth` element
-// would be in that position in an ordered sequence; other elements may be in
-// any order, except that all preceding `nth` will be less than that element,
-// and all following `nth` will be greater than that element.
-template <typename RandomAccessContainer>
-void c_nth_element(
-    RandomAccessContainer& sequence,
-    container_algorithm_internal::ContainerIter<RandomAccessContainer> nth) {
-  std::nth_element(container_algorithm_internal::c_begin(sequence), nth,
-                   container_algorithm_internal::c_end(sequence));
-}
-
-// Overload of c_nth_element() for performing a `comp` comparison other than
-// the default `operator<`.
-template <typename RandomAccessContainer, typename LessThan>
-void c_nth_element(
-    RandomAccessContainer& sequence,
-    container_algorithm_internal::ContainerIter<RandomAccessContainer> nth,
-    LessThan&& comp) {
-  std::nth_element(container_algorithm_internal::c_begin(sequence), nth,
-                   container_algorithm_internal::c_end(sequence),
-                   std::forward<LessThan>(comp));
-}
-
-//------------------------------------------------------------------------------
-// <algorithm> Binary Search
-//------------------------------------------------------------------------------
-
-// c_lower_bound()
-//
-// Container-based version of the <algorithm> `std::lower_bound()` function
-// to return an iterator pointing to the first element in a sorted container
-// which does not compare less than `value`.
-template <typename Sequence, typename T>
-container_algorithm_internal::ContainerIter<Sequence> c_lower_bound(
-    Sequence& sequence, const T& value) {
-  return std::lower_bound(container_algorithm_internal::c_begin(sequence),
-                          container_algorithm_internal::c_end(sequence), value);
-}
-
-// Overload of c_lower_bound() for performing a `comp` comparison other than
-// the default `operator<`.
-template <typename Sequence, typename T, typename LessThan>
-container_algorithm_internal::ContainerIter<Sequence> c_lower_bound(
-    Sequence& sequence, const T& value, LessThan&& comp) {
-  return std::lower_bound(container_algorithm_internal::c_begin(sequence),
-                          container_algorithm_internal::c_end(sequence), value,
-                          std::forward<LessThan>(comp));
-}
-
-// c_upper_bound()
-//
-// Container-based version of the <algorithm> `std::upper_bound()` function
-// to return an iterator pointing to the first element in a sorted container
-// which is greater than `value`.
-template <typename Sequence, typename T>
-container_algorithm_internal::ContainerIter<Sequence> c_upper_bound(
-    Sequence& sequence, const T& value) {
-  return std::upper_bound(container_algorithm_internal::c_begin(sequence),
-                          container_algorithm_internal::c_end(sequence), value);
-}
-
-// Overload of c_upper_bound() for performing a `comp` comparison other than
-// the default `operator<`.
-template <typename Sequence, typename T, typename LessThan>
-container_algorithm_internal::ContainerIter<Sequence> c_upper_bound(
-    Sequence& sequence, const T& value, LessThan&& comp) {
-  return std::upper_bound(container_algorithm_internal::c_begin(sequence),
-                          container_algorithm_internal::c_end(sequence), value,
-                          std::forward<LessThan>(comp));
-}
-
-// c_equal_range()
-//
-// Container-based version of the <algorithm> `std::equal_range()` function
-// to return an iterator pair pointing to the first and last elements in a
-// sorted container which compare equal to `value`.
-template <typename Sequence, typename T>
-container_algorithm_internal::ContainerIterPairType<Sequence, Sequence>
-c_equal_range(Sequence& sequence, const T& value) {
-  return std::equal_range(container_algorithm_internal::c_begin(sequence),
-                          container_algorithm_internal::c_end(sequence), value);
-}
-
-// Overload of c_equal_range() for performing a `comp` comparison other than
-// the default `operator<`.
-template <typename Sequence, typename T, typename LessThan>
-container_algorithm_internal::ContainerIterPairType<Sequence, Sequence>
-c_equal_range(Sequence& sequence, const T& value, LessThan&& comp) {
-  return std::equal_range(container_algorithm_internal::c_begin(sequence),
-                          container_algorithm_internal::c_end(sequence), value,
-                          std::forward<LessThan>(comp));
-}
-
-// c_binary_search()
-//
-// Container-based version of the <algorithm> `std::binary_search()` function
-// to test if any element in the sorted container contains a value equivalent to
-// 'value'.
-template <typename Sequence, typename T>
-bool c_binary_search(const Sequence& sequence, const T& value) {
-  return std::binary_search(container_algorithm_internal::c_begin(sequence),
-                            container_algorithm_internal::c_end(sequence),
-                            value);
-}
-
-// Overload of c_binary_search() for performing a `comp` comparison other than
-// the default `operator<`.
-template <typename Sequence, typename T, typename LessThan>
-bool c_binary_search(const Sequence& sequence, const T& value,
-                     LessThan&& comp) {
-  return std::binary_search(container_algorithm_internal::c_begin(sequence),
-                            container_algorithm_internal::c_end(sequence),
-                            value, std::forward<LessThan>(comp));
-}
-
-//------------------------------------------------------------------------------
-// <algorithm> Merge functions
-//------------------------------------------------------------------------------
-
-// c_merge()
-//
-// Container-based version of the <algorithm> `std::merge()` function
-// to merge two sorted containers into a single sorted iterator.
-template <typename C1, typename C2, typename OutputIterator>
-OutputIterator c_merge(const C1& c1, const C2& c2, OutputIterator result) {
-  return std::merge(container_algorithm_internal::c_begin(c1),
-                    container_algorithm_internal::c_end(c1),
-                    container_algorithm_internal::c_begin(c2),
-                    container_algorithm_internal::c_end(c2), result);
-}
-
-// Overload of c_merge() for performing a `comp` comparison other than
-// the default `operator<`.
-template <typename C1, typename C2, typename OutputIterator, typename LessThan>
-OutputIterator c_merge(const C1& c1, const C2& c2, OutputIterator result,
-                       LessThan&& comp) {
-  return std::merge(container_algorithm_internal::c_begin(c1),
-                    container_algorithm_internal::c_end(c1),
-                    container_algorithm_internal::c_begin(c2),
-                    container_algorithm_internal::c_end(c2), result,
-                    std::forward<LessThan>(comp));
-}
-
-// c_inplace_merge()
-//
-// Container-based version of the <algorithm> `std::inplace_merge()` function
-// to merge a supplied iterator `middle` into a container.
-template <typename C>
-void c_inplace_merge(C& c,
-                     container_algorithm_internal::ContainerIter<C> middle) {
-  std::inplace_merge(container_algorithm_internal::c_begin(c), middle,
-                     container_algorithm_internal::c_end(c));
-}
-
-// Overload of c_inplace_merge() for performing a merge using a `comp` other
-// than `operator<`.
-template <typename C, typename LessThan>
-void c_inplace_merge(C& c,
-                     container_algorithm_internal::ContainerIter<C> middle,
-                     LessThan&& comp) {
-  std::inplace_merge(container_algorithm_internal::c_begin(c), middle,
-                     container_algorithm_internal::c_end(c),
-                     std::forward<LessThan>(comp));
-}
-
-// c_includes()
-//
-// Container-based version of the <algorithm> `std::includes()` function
-// to test whether a sorted container `c1` entirely contains another sorted
-// container `c2`.
-template <typename C1, typename C2>
-bool c_includes(const C1& c1, const C2& c2) {
-  return std::includes(container_algorithm_internal::c_begin(c1),
-                       container_algorithm_internal::c_end(c1),
-                       container_algorithm_internal::c_begin(c2),
-                       container_algorithm_internal::c_end(c2));
-}
-
-// Overload of c_includes() for performing a merge using a `comp` other than
-// `operator<`.
-template <typename C1, typename C2, typename LessThan>
-bool c_includes(const C1& c1, const C2& c2, LessThan&& comp) {
-  return std::includes(container_algorithm_internal::c_begin(c1),
-                       container_algorithm_internal::c_end(c1),
-                       container_algorithm_internal::c_begin(c2),
-                       container_algorithm_internal::c_end(c2),
-                       std::forward<LessThan>(comp));
-}
-
-// c_set_union()
-//
-// Container-based version of the <algorithm> `std::set_union()` function
-// to return an iterator containing the union of two containers; duplicate
-// values are not copied into the output.
-template <typename C1, typename C2, typename OutputIterator,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C1>::value,
-              void>::type,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C2>::value,
-              void>::type>
-OutputIterator c_set_union(const C1& c1, const C2& c2, OutputIterator output) {
-  return std::set_union(container_algorithm_internal::c_begin(c1),
-                        container_algorithm_internal::c_end(c1),
-                        container_algorithm_internal::c_begin(c2),
-                        container_algorithm_internal::c_end(c2), output);
-}
-
-// Overload of c_set_union() for performing a merge using a `comp` other than
-// `operator<`.
-template <typename C1, typename C2, typename OutputIterator, typename LessThan,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C1>::value,
-              void>::type,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C2>::value,
-              void>::type>
-OutputIterator c_set_union(const C1& c1, const C2& c2, OutputIterator output,
-                           LessThan&& comp) {
-  return std::set_union(container_algorithm_internal::c_begin(c1),
-                        container_algorithm_internal::c_end(c1),
-                        container_algorithm_internal::c_begin(c2),
-                        container_algorithm_internal::c_end(c2), output,
-                        std::forward<LessThan>(comp));
-}
-
-// c_set_intersection()
-//
-// Container-based version of the <algorithm> `std::set_intersection()` function
-// to return an iterator containing the intersection of two sorted containers.
-template <typename C1, typename C2, typename OutputIterator,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C1>::value,
-              void>::type,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C2>::value,
-              void>::type>
-OutputIterator c_set_intersection(const C1& c1, const C2& c2,
-                                  OutputIterator output) {
-  // In debug builds, ensure that both containers are sorted with respect to the
-  // default comparator. std::set_intersection requires the containers be sorted
-  // using operator<.
-  assert(absl::c_is_sorted(c1));
-  assert(absl::c_is_sorted(c2));
-  return std::set_intersection(container_algorithm_internal::c_begin(c1),
-                               container_algorithm_internal::c_end(c1),
-                               container_algorithm_internal::c_begin(c2),
-                               container_algorithm_internal::c_end(c2), output);
-}
-
-// Overload of c_set_intersection() for performing a merge using a `comp` other
-// than `operator<`.
-template <typename C1, typename C2, typename OutputIterator, typename LessThan,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C1>::value,
-              void>::type,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C2>::value,
-              void>::type>
-OutputIterator c_set_intersection(const C1& c1, const C2& c2,
-                                  OutputIterator output, LessThan&& comp) {
-  // In debug builds, ensure that both containers are sorted with respect to the
-  // default comparator. std::set_intersection requires the containers be sorted
-  // using the same comparator.
-  assert(absl::c_is_sorted(c1, comp));
-  assert(absl::c_is_sorted(c2, comp));
-  return std::set_intersection(container_algorithm_internal::c_begin(c1),
-                               container_algorithm_internal::c_end(c1),
-                               container_algorithm_internal::c_begin(c2),
-                               container_algorithm_internal::c_end(c2), output,
-                               std::forward<LessThan>(comp));
-}
-
-// c_set_difference()
-//
-// Container-based version of the <algorithm> `std::set_difference()` function
-// to return an iterator containing elements present in the first container but
-// not in the second.
-template <typename C1, typename C2, typename OutputIterator,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C1>::value,
-              void>::type,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C2>::value,
-              void>::type>
-OutputIterator c_set_difference(const C1& c1, const C2& c2,
-                                OutputIterator output) {
-  return std::set_difference(container_algorithm_internal::c_begin(c1),
-                             container_algorithm_internal::c_end(c1),
-                             container_algorithm_internal::c_begin(c2),
-                             container_algorithm_internal::c_end(c2), output);
-}
-
-// Overload of c_set_difference() for performing a merge using a `comp` other
-// than `operator<`.
-template <typename C1, typename C2, typename OutputIterator, typename LessThan,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C1>::value,
-              void>::type,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C2>::value,
-              void>::type>
-OutputIterator c_set_difference(const C1& c1, const C2& c2,
-                                OutputIterator output, LessThan&& comp) {
-  return std::set_difference(container_algorithm_internal::c_begin(c1),
-                             container_algorithm_internal::c_end(c1),
-                             container_algorithm_internal::c_begin(c2),
-                             container_algorithm_internal::c_end(c2), output,
-                             std::forward<LessThan>(comp));
-}
-
-// c_set_symmetric_difference()
-//
-// Container-based version of the <algorithm> `std::set_symmetric_difference()`
-// function to return an iterator containing elements present in either one
-// container or the other, but not both.
-template <typename C1, typename C2, typename OutputIterator,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C1>::value,
-              void>::type,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C2>::value,
-              void>::type>
-OutputIterator c_set_symmetric_difference(const C1& c1, const C2& c2,
-                                          OutputIterator output) {
-  return std::set_symmetric_difference(
-      container_algorithm_internal::c_begin(c1),
-      container_algorithm_internal::c_end(c1),
-      container_algorithm_internal::c_begin(c2),
-      container_algorithm_internal::c_end(c2), output);
-}
-
-// Overload of c_set_symmetric_difference() for performing a merge using a
-// `comp` other than `operator<`.
-template <typename C1, typename C2, typename OutputIterator, typename LessThan,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C1>::value,
-              void>::type,
-          typename = typename std::enable_if<
-              !container_algorithm_internal::IsUnorderedContainer<C2>::value,
-              void>::type>
-OutputIterator c_set_symmetric_difference(const C1& c1, const C2& c2,
-                                          OutputIterator output,
-                                          LessThan&& comp) {
-  return std::set_symmetric_difference(
-      container_algorithm_internal::c_begin(c1),
-      container_algorithm_internal::c_end(c1),
-      container_algorithm_internal::c_begin(c2),
-      container_algorithm_internal::c_end(c2), output,
-      std::forward<LessThan>(comp));
-}
-
-//------------------------------------------------------------------------------
-// <algorithm> Heap functions
-//------------------------------------------------------------------------------
-
-// c_push_heap()
-//
-// Container-based version of the <algorithm> `std::push_heap()` function
-// to push a value onto a container heap.
-template <typename RandomAccessContainer>
-void c_push_heap(RandomAccessContainer& sequence) {
-  std::push_heap(container_algorithm_internal::c_begin(sequence),
-                 container_algorithm_internal::c_end(sequence));
-}
-
-// Overload of c_push_heap() for performing a push operation on a heap using a
-// `comp` other than `operator<`.
-template <typename RandomAccessContainer, typename LessThan>
-void c_push_heap(RandomAccessContainer& sequence, LessThan&& comp) {
-  std::push_heap(container_algorithm_internal::c_begin(sequence),
-                 container_algorithm_internal::c_end(sequence),
-                 std::forward<LessThan>(comp));
-}
-
-// c_pop_heap()
-//
-// Container-based version of the <algorithm> `std::pop_heap()` function
-// to pop a value from a heap container.
-template <typename RandomAccessContainer>
-void c_pop_heap(RandomAccessContainer& sequence) {
-  std::pop_heap(container_algorithm_internal::c_begin(sequence),
-                container_algorithm_internal::c_end(sequence));
-}
-
-// Overload of c_pop_heap() for performing a pop operation on a heap using a
-// `comp` other than `operator<`.
-template <typename RandomAccessContainer, typename LessThan>
-void c_pop_heap(RandomAccessContainer& sequence, LessThan&& comp) {
-  std::pop_heap(container_algorithm_internal::c_begin(sequence),
-                container_algorithm_internal::c_end(sequence),
-                std::forward<LessThan>(comp));
-}
-
-// c_make_heap()
-//
-// Container-based version of the <algorithm> `std::make_heap()` function
-// to make a container a heap.
-template <typename RandomAccessContainer>
-void c_make_heap(RandomAccessContainer& sequence) {
-  std::make_heap(container_algorithm_internal::c_begin(sequence),
-                 container_algorithm_internal::c_end(sequence));
-}
-
-// Overload of c_make_heap() for performing heap comparisons using a
-// `comp` other than `operator<`
-template <typename RandomAccessContainer, typename LessThan>
-void c_make_heap(RandomAccessContainer& sequence, LessThan&& comp) {
-  std::make_heap(container_algorithm_internal::c_begin(sequence),
-                 container_algorithm_internal::c_end(sequence),
-                 std::forward<LessThan>(comp));
-}
-
-// c_sort_heap()
-//
-// Container-based version of the <algorithm> `std::sort_heap()` function
-// to sort a heap into ascending order (after which it is no longer a heap).
-template <typename RandomAccessContainer>
-void c_sort_heap(RandomAccessContainer& sequence) {
-  std::sort_heap(container_algorithm_internal::c_begin(sequence),
-                 container_algorithm_internal::c_end(sequence));
-}
-
-// Overload of c_sort_heap() for performing heap comparisons using a
-// `comp` other than `operator<`
-template <typename RandomAccessContainer, typename LessThan>
-void c_sort_heap(RandomAccessContainer& sequence, LessThan&& comp) {
-  std::sort_heap(container_algorithm_internal::c_begin(sequence),
-                 container_algorithm_internal::c_end(sequence),
-                 std::forward<LessThan>(comp));
-}
-
-// c_is_heap()
-//
-// Container-based version of the <algorithm> `std::is_heap()` function
-// to check whether the given container is a heap.
-template <typename RandomAccessContainer>
-bool c_is_heap(const RandomAccessContainer& sequence) {
-  return std::is_heap(container_algorithm_internal::c_begin(sequence),
-                      container_algorithm_internal::c_end(sequence));
-}
-
-// Overload of c_is_heap() for performing heap comparisons using a
-// `comp` other than `operator<`
-template <typename RandomAccessContainer, typename LessThan>
-bool c_is_heap(const RandomAccessContainer& sequence, LessThan&& comp) {
-  return std::is_heap(container_algorithm_internal::c_begin(sequence),
-                      container_algorithm_internal::c_end(sequence),
-                      std::forward<LessThan>(comp));
-}
-
-// c_is_heap_until()
-//
-// Container-based version of the <algorithm> `std::is_heap_until()` function
-// to find the first element in a given container which is not in heap order.
-template <typename RandomAccessContainer>
-container_algorithm_internal::ContainerIter<RandomAccessContainer>
-c_is_heap_until(RandomAccessContainer& sequence) {
-  return std::is_heap_until(container_algorithm_internal::c_begin(sequence),
-                            container_algorithm_internal::c_end(sequence));
-}
-
-// Overload of c_is_heap_until() for performing heap comparisons using a
-// `comp` other than `operator<`
-template <typename RandomAccessContainer, typename LessThan>
-container_algorithm_internal::ContainerIter<RandomAccessContainer>
-c_is_heap_until(RandomAccessContainer& sequence, LessThan&& comp) {
-  return std::is_heap_until(container_algorithm_internal::c_begin(sequence),
-                            container_algorithm_internal::c_end(sequence),
-                            std::forward<LessThan>(comp));
-}
-
-//------------------------------------------------------------------------------
-//  <algorithm> Min/max
-//------------------------------------------------------------------------------
-
-// c_min_element()
-//
-// Container-based version of the <algorithm> `std::min_element()` function
-// to return an iterator pointing to the element with the smallest value, using
-// `operator<` to make the comparisons.
-template <typename Sequence>
-container_algorithm_internal::ContainerIter<Sequence> c_min_element(
-    Sequence& sequence) {
-  return std::min_element(container_algorithm_internal::c_begin(sequence),
-                          container_algorithm_internal::c_end(sequence));
-}
-
-// Overload of c_min_element() for performing a `comp` comparison other than
-// `operator<`.
-template <typename Sequence, typename LessThan>
-container_algorithm_internal::ContainerIter<Sequence> c_min_element(
-    Sequence& sequence, LessThan&& comp) {
-  return std::min_element(container_algorithm_internal::c_begin(sequence),
-                          container_algorithm_internal::c_end(sequence),
-                          std::forward<LessThan>(comp));
-}
-
-// c_max_element()
-//
-// Container-based version of the <algorithm> `std::max_element()` function
-// to return an iterator pointing to the element with the largest value, using
-// `operator<` to make the comparisons.
-template <typename Sequence>
-container_algorithm_internal::ContainerIter<Sequence> c_max_element(
-    Sequence& sequence) {
-  return std::max_element(container_algorithm_internal::c_begin(sequence),
-                          container_algorithm_internal::c_end(sequence));
-}
-
-// Overload of c_max_element() for performing a `comp` comparison other than
-// `operator<`.
-template <typename Sequence, typename LessThan>
-container_algorithm_internal::ContainerIter<Sequence> c_max_element(
-    Sequence& sequence, LessThan&& comp) {
-  return std::max_element(container_algorithm_internal::c_begin(sequence),
-                          container_algorithm_internal::c_end(sequence),
-                          std::forward<LessThan>(comp));
-}
-
-// c_minmax_element()
-//
-// Container-based version of the <algorithm> `std::minmax_element()` function
-// to return a pair of iterators pointing to the elements containing the
-// smallest and largest values, respectively, using `operator<` to make the
-// comparisons.
-template <typename C>
-container_algorithm_internal::ContainerIterPairType<C, C> c_minmax_element(
-    C& c) {
-  return std::minmax_element(container_algorithm_internal::c_begin(c),
-                             container_algorithm_internal::c_end(c));
-}
-
-// Overload of c_minmax_element() for performing `comp` comparisons other than
-// `operator<`.
-template <typename C, typename LessThan>
-container_algorithm_internal::ContainerIterPairType<C, C> c_minmax_element(
-    C& c, LessThan&& comp) {
-  return std::minmax_element(container_algorithm_internal::c_begin(c),
-                             container_algorithm_internal::c_end(c),
-                             std::forward<LessThan>(comp));
-}
-
-//------------------------------------------------------------------------------
-//  <algorithm> Lexicographical Comparisons
-//------------------------------------------------------------------------------
-
-// c_lexicographical_compare()
-//
-// Container-based version of the <algorithm> `std::lexicographical_compare()`
-// function to lexicographically compare (e.g. sort words alphabetically) two
-// container sequences. The comparison is performed using `operator<`. Note
-// that capital letters ("A-Z") have ASCII values less than lowercase letters
-// ("a-z").
-template <typename Sequence1, typename Sequence2>
-bool c_lexicographical_compare(const Sequence1& sequence1,
-                               const Sequence2& sequence2) {
-  return std::lexicographical_compare(
-      container_algorithm_internal::c_begin(sequence1),
-      container_algorithm_internal::c_end(sequence1),
-      container_algorithm_internal::c_begin(sequence2),
-      container_algorithm_internal::c_end(sequence2));
-}
-
-// Overload of c_lexicographical_compare() for performing a lexicographical
-// comparison using a `comp` operator instead of `operator<`.
-template <typename Sequence1, typename Sequence2, typename LessThan>
-bool c_lexicographical_compare(const Sequence1& sequence1,
-                               const Sequence2& sequence2, LessThan&& comp) {
-  return std::lexicographical_compare(
-      container_algorithm_internal::c_begin(sequence1),
-      container_algorithm_internal::c_end(sequence1),
-      container_algorithm_internal::c_begin(sequence2),
-      container_algorithm_internal::c_end(sequence2),
-      std::forward<LessThan>(comp));
-}
-
-// c_next_permutation()
-//
-// Container-based version of the <algorithm> `std::next_permutation()` function
-// to rearrange a container's elements into the next lexicographically greater
-// permutation.
-template <typename C>
-bool c_next_permutation(C& c) {
-  return std::next_permutation(container_algorithm_internal::c_begin(c),
-                               container_algorithm_internal::c_end(c));
-}
-
-// Overload of c_next_permutation() for performing a lexicographical
-// comparison using a `comp` operator instead of `operator<`.
-template <typename C, typename LessThan>
-bool c_next_permutation(C& c, LessThan&& comp) {
-  return std::next_permutation(container_algorithm_internal::c_begin(c),
-                               container_algorithm_internal::c_end(c),
-                               std::forward<LessThan>(comp));
-}
-
-// c_prev_permutation()
-//
-// Container-based version of the <algorithm> `std::prev_permutation()` function
-// to rearrange a container's elements into the next lexicographically lesser
-// permutation.
-template <typename C>
-bool c_prev_permutation(C& c) {
-  return std::prev_permutation(container_algorithm_internal::c_begin(c),
-                               container_algorithm_internal::c_end(c));
-}
-
-// Overload of c_prev_permutation() for performing a lexicographical
-// comparison using a `comp` operator instead of `operator<`.
-template <typename C, typename LessThan>
-bool c_prev_permutation(C& c, LessThan&& comp) {
-  return std::prev_permutation(container_algorithm_internal::c_begin(c),
-                               container_algorithm_internal::c_end(c),
-                               std::forward<LessThan>(comp));
-}
-
-//------------------------------------------------------------------------------
-// <numeric> algorithms
-//------------------------------------------------------------------------------
-
-// c_iota()
-//
-// Container-based version of the <numeric> `std::iota()` function
-// to compute successive values of `value`, as if incremented with `++value`
-// after each element is written, and write them to the container.
-template <typename Sequence, typename T>
-void c_iota(Sequence& sequence, const T& value) {
-  std::iota(container_algorithm_internal::c_begin(sequence),
-            container_algorithm_internal::c_end(sequence), value);
-}
-
-// c_accumulate()
-//
-// Container-based version of the <numeric> `std::accumulate()` function
-// to accumulate the element values of a container to `init` and return that
-// accumulation by value.
-//
-// Note: Due to a language technicality this function has return type
-// absl::decay_t<T>. As a user of this function you can casually read
-// this as "returns T by value" and assume it does the right thing.
-template <typename Sequence, typename T>
-decay_t<T> c_accumulate(const Sequence& sequence, T&& init) {
-  return std::accumulate(container_algorithm_internal::c_begin(sequence),
-                         container_algorithm_internal::c_end(sequence),
-                         std::forward<T>(init));
-}
-
-// Overload of c_accumulate() for using a binary operations other than
-// addition for computing the accumulation.
-template <typename Sequence, typename T, typename BinaryOp>
-decay_t<T> c_accumulate(const Sequence& sequence, T&& init,
-                        BinaryOp&& binary_op) {
-  return std::accumulate(container_algorithm_internal::c_begin(sequence),
-                         container_algorithm_internal::c_end(sequence),
-                         std::forward<T>(init),
-                         std::forward<BinaryOp>(binary_op));
-}
-
-// c_inner_product()
-//
-// Container-based version of the <numeric> `std::inner_product()` function
-// to compute the cumulative inner product of container element pairs.
-//
-// Note: Due to a language technicality this function has return type
-// absl::decay_t<T>. As a user of this function you can casually read
-// this as "returns T by value" and assume it does the right thing.
-template <typename Sequence1, typename Sequence2, typename T>
-decay_t<T> c_inner_product(const Sequence1& factors1, const Sequence2& factors2,
-                           T&& sum) {
-  return std::inner_product(container_algorithm_internal::c_begin(factors1),
-                            container_algorithm_internal::c_end(factors1),
-                            container_algorithm_internal::c_begin(factors2),
-                            std::forward<T>(sum));
-}
-
-// Overload of c_inner_product() for using binary operations other than
-// `operator+` (for computing the accumulation) and `operator*` (for computing
-// the product between the two container's element pair).
-template <typename Sequence1, typename Sequence2, typename T,
-          typename BinaryOp1, typename BinaryOp2>
-decay_t<T> c_inner_product(const Sequence1& factors1, const Sequence2& factors2,
-                           T&& sum, BinaryOp1&& op1, BinaryOp2&& op2) {
-  return std::inner_product(container_algorithm_internal::c_begin(factors1),
-                            container_algorithm_internal::c_end(factors1),
-                            container_algorithm_internal::c_begin(factors2),
-                            std::forward<T>(sum), std::forward<BinaryOp1>(op1),
-                            std::forward<BinaryOp2>(op2));
-}
-
-// c_adjacent_difference()
-//
-// Container-based version of the <numeric> `std::adjacent_difference()`
-// function to compute the difference between each element and the one preceding
-// it and write it to an iterator.
-template <typename InputSequence, typename OutputIt>
-OutputIt c_adjacent_difference(const InputSequence& input,
-                               OutputIt output_first) {
-  return std::adjacent_difference(container_algorithm_internal::c_begin(input),
-                                  container_algorithm_internal::c_end(input),
-                                  output_first);
-}
-
-// Overload of c_adjacent_difference() for using a binary operation other than
-// subtraction to compute the adjacent difference.
-template <typename InputSequence, typename OutputIt, typename BinaryOp>
-OutputIt c_adjacent_difference(const InputSequence& input,
-                               OutputIt output_first, BinaryOp&& op) {
-  return std::adjacent_difference(container_algorithm_internal::c_begin(input),
-                                  container_algorithm_internal::c_end(input),
-                                  output_first, std::forward<BinaryOp>(op));
-}
-
-// c_partial_sum()
-//
-// Container-based version of the <numeric> `std::partial_sum()` function
-// to compute the partial sum of the elements in a sequence and write them
-// to an iterator. The partial sum is the sum of all element values so far in
-// the sequence.
-template <typename InputSequence, typename OutputIt>
-OutputIt c_partial_sum(const InputSequence& input, OutputIt output_first) {
-  return std::partial_sum(container_algorithm_internal::c_begin(input),
-                          container_algorithm_internal::c_end(input),
-                          output_first);
-}
-
-// Overload of c_partial_sum() for using a binary operation other than addition
-// to compute the "partial sum".
-template <typename InputSequence, typename OutputIt, typename BinaryOp>
-OutputIt c_partial_sum(const InputSequence& input, OutputIt output_first,
-                       BinaryOp&& op) {
-  return std::partial_sum(container_algorithm_internal::c_begin(input),
-                          container_algorithm_internal::c_end(input),
-                          output_first, std::forward<BinaryOp>(op));
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_ALGORITHM_CONTAINER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/attributes.h b/projects/llm_framework/include/abseil-cpp/absl/base/attributes.h
deleted file mode 100644
index 38086a82..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/attributes.h
+++ /dev/null
@@ -1,921 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// This header file defines macros for declaring attributes for functions,
-// types, and variables.
-//
-// These macros are used within Abseil and allow the compiler to optimize, where
-// applicable, certain function calls.
-//
-// Most macros here are exposing GCC or Clang features, and are stubbed out for
-// other compilers.
-//
-// GCC attributes documentation:
-//   https://gcc.gnu.org/onlinedocs/gcc-4.7.0/gcc/Function-Attributes.html
-//   https://gcc.gnu.org/onlinedocs/gcc-4.7.0/gcc/Variable-Attributes.html
-//   https://gcc.gnu.org/onlinedocs/gcc-4.7.0/gcc/Type-Attributes.html
-//
-// Most attributes in this file are already supported by GCC 4.7. However, some
-// of them are not supported in older version of Clang. Thus, we check
-// `__has_attribute()` first. If the check fails, we check if we are on GCC and
-// assume the attribute exists on GCC (which is verified on GCC 4.7).
-
-#ifndef ABSL_BASE_ATTRIBUTES_H_
-#define ABSL_BASE_ATTRIBUTES_H_
-
-#include "absl/base/config.h"
-
-// ABSL_HAVE_ATTRIBUTE
-//
-// A function-like feature checking macro that is a wrapper around
-// `__has_attribute`, which is defined by GCC 5+ and Clang and evaluates to a
-// nonzero constant integer if the attribute is supported or 0 if not.
-//
-// It evaluates to zero if `__has_attribute` is not defined by the compiler.
-//
-// GCC: https://gcc.gnu.org/gcc-5/changes.html
-// Clang: https://clang.llvm.org/docs/LanguageExtensions.html
-#ifdef __has_attribute
-#define ABSL_HAVE_ATTRIBUTE(x) __has_attribute(x)
-#else
-#define ABSL_HAVE_ATTRIBUTE(x) 0
-#endif
-
-// ABSL_HAVE_CPP_ATTRIBUTE
-//
-// A function-like feature checking macro that accepts C++11 style attributes.
-// It's a wrapper around `__has_cpp_attribute`, defined by ISO C++ SD-6
-// (https://en.cppreference.com/w/cpp/experimental/feature_test). If we don't
-// find `__has_cpp_attribute`, will evaluate to 0.
-#if defined(__cplusplus) && defined(__has_cpp_attribute)
-// NOTE: requiring __cplusplus above should not be necessary, but
-// works around https://bugs.llvm.org/show_bug.cgi?id=23435.
-#define ABSL_HAVE_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
-#else
-#define ABSL_HAVE_CPP_ATTRIBUTE(x) 0
-#endif
-
-// -----------------------------------------------------------------------------
-// Function Attributes
-// -----------------------------------------------------------------------------
-//
-// GCC: https://gcc.gnu.org/onlinedocs/gcc/Function-Attributes.html
-// Clang: https://clang.llvm.org/docs/AttributeReference.html
-
-// ABSL_PRINTF_ATTRIBUTE
-// ABSL_SCANF_ATTRIBUTE
-//
-// Tells the compiler to perform `printf` format string checking if the
-// compiler supports it; see the 'format' attribute in
-// <https://gcc.gnu.org/onlinedocs/gcc-4.7.0/gcc/Function-Attributes.html>.
-//
-// Note: As the GCC manual states, "[s]ince non-static C++ methods
-// have an implicit 'this' argument, the arguments of such methods
-// should be counted from two, not one."
-#if ABSL_HAVE_ATTRIBUTE(format) || (defined(__GNUC__) && !defined(__clang__))
-#define ABSL_PRINTF_ATTRIBUTE(string_index, first_to_check) \
-  __attribute__((__format__(__printf__, string_index, first_to_check)))
-#define ABSL_SCANF_ATTRIBUTE(string_index, first_to_check) \
-  __attribute__((__format__(__scanf__, string_index, first_to_check)))
-#else
-#define ABSL_PRINTF_ATTRIBUTE(string_index, first_to_check)
-#define ABSL_SCANF_ATTRIBUTE(string_index, first_to_check)
-#endif
-
-// ABSL_ATTRIBUTE_ALWAYS_INLINE
-// ABSL_ATTRIBUTE_NOINLINE
-//
-// Forces functions to either inline or not inline. Introduced in gcc 3.1.
-#if ABSL_HAVE_ATTRIBUTE(always_inline) || \
-    (defined(__GNUC__) && !defined(__clang__))
-#define ABSL_ATTRIBUTE_ALWAYS_INLINE __attribute__((always_inline))
-#define ABSL_HAVE_ATTRIBUTE_ALWAYS_INLINE 1
-#else
-#define ABSL_ATTRIBUTE_ALWAYS_INLINE
-#endif
-
-#if ABSL_HAVE_ATTRIBUTE(noinline) || (defined(__GNUC__) && !defined(__clang__))
-#define ABSL_ATTRIBUTE_NOINLINE __attribute__((noinline))
-#define ABSL_HAVE_ATTRIBUTE_NOINLINE 1
-#else
-#define ABSL_ATTRIBUTE_NOINLINE
-#endif
-
-// ABSL_ATTRIBUTE_NO_TAIL_CALL
-//
-// Prevents the compiler from optimizing away stack frames for functions which
-// end in a call to another function.
-#if ABSL_HAVE_ATTRIBUTE(disable_tail_calls)
-#define ABSL_HAVE_ATTRIBUTE_NO_TAIL_CALL 1
-#define ABSL_ATTRIBUTE_NO_TAIL_CALL __attribute__((disable_tail_calls))
-#elif defined(__GNUC__) && !defined(__clang__) && !defined(__e2k__)
-#define ABSL_HAVE_ATTRIBUTE_NO_TAIL_CALL 1
-#define ABSL_ATTRIBUTE_NO_TAIL_CALL \
-  __attribute__((optimize("no-optimize-sibling-calls")))
-#else
-#define ABSL_ATTRIBUTE_NO_TAIL_CALL
-#define ABSL_HAVE_ATTRIBUTE_NO_TAIL_CALL 0
-#endif
-
-// ABSL_ATTRIBUTE_WEAK
-//
-// Tags a function as weak for the purposes of compilation and linking.
-// Weak attributes did not work properly in LLVM's Windows backend before
-// 9.0.0, so disable them there. See https://bugs.llvm.org/show_bug.cgi?id=37598
-// for further information.
-// The MinGW compiler doesn't complain about the weak attribute until the link
-// step, presumably because Windows doesn't use ELF binaries.
-#if (ABSL_HAVE_ATTRIBUTE(weak) ||                                         \
-     (defined(__GNUC__) && !defined(__clang__))) &&                       \
-    (!defined(_WIN32) || (defined(__clang__) && __clang_major__ >= 9)) && \
-    !defined(__MINGW32__)
-#undef ABSL_ATTRIBUTE_WEAK
-#define ABSL_ATTRIBUTE_WEAK __attribute__((weak))
-#define ABSL_HAVE_ATTRIBUTE_WEAK 1
-#else
-#define ABSL_ATTRIBUTE_WEAK
-#define ABSL_HAVE_ATTRIBUTE_WEAK 0
-#endif
-
-// ABSL_ATTRIBUTE_NONNULL
-//
-// Tells the compiler either (a) that a particular function parameter
-// should be a non-null pointer, or (b) that all pointer arguments should
-// be non-null.
-//
-// Note: As the GCC manual states, "[s]ince non-static C++ methods
-// have an implicit 'this' argument, the arguments of such methods
-// should be counted from two, not one."
-//
-// Args are indexed starting at 1.
-//
-// For non-static class member functions, the implicit `this` argument
-// is arg 1, and the first explicit argument is arg 2. For static class member
-// functions, there is no implicit `this`, and the first explicit argument is
-// arg 1.
-//
-// Example:
-//
-//   /* arg_a cannot be null, but arg_b can */
-//   void Function(void* arg_a, void* arg_b) ABSL_ATTRIBUTE_NONNULL(1);
-//
-//   class C {
-//     /* arg_a cannot be null, but arg_b can */
-//     void Method(void* arg_a, void* arg_b) ABSL_ATTRIBUTE_NONNULL(2);
-//
-//     /* arg_a cannot be null, but arg_b can */
-//     static void StaticMethod(void* arg_a, void* arg_b)
-//     ABSL_ATTRIBUTE_NONNULL(1);
-//   };
-//
-// If no arguments are provided, then all pointer arguments should be non-null.
-//
-//  /* No pointer arguments may be null. */
-//  void Function(void* arg_a, void* arg_b, int arg_c) ABSL_ATTRIBUTE_NONNULL();
-//
-// NOTE: The GCC nonnull attribute actually accepts a list of arguments, but
-// ABSL_ATTRIBUTE_NONNULL does not.
-#if ABSL_HAVE_ATTRIBUTE(nonnull) || (defined(__GNUC__) && !defined(__clang__))
-#define ABSL_ATTRIBUTE_NONNULL(arg_index) __attribute__((nonnull(arg_index)))
-#else
-#define ABSL_ATTRIBUTE_NONNULL(...)
-#endif
-
-// ABSL_ATTRIBUTE_NORETURN
-//
-// Tells the compiler that a given function never returns.
-#if ABSL_HAVE_ATTRIBUTE(noreturn) || (defined(__GNUC__) && !defined(__clang__))
-#define ABSL_ATTRIBUTE_NORETURN __attribute__((noreturn))
-#elif defined(_MSC_VER)
-#define ABSL_ATTRIBUTE_NORETURN __declspec(noreturn)
-#else
-#define ABSL_ATTRIBUTE_NORETURN
-#endif
-
-// ABSL_ATTRIBUTE_NO_SANITIZE_ADDRESS
-//
-// Tells the AddressSanitizer (or other memory testing tools) to ignore a given
-// function. Useful for cases when a function reads random locations on stack,
-// calls _exit from a cloned subprocess, deliberately accesses buffer
-// out of bounds or does other scary things with memory.
-// NOTE: GCC supports AddressSanitizer(asan) since 4.8.
-// https://gcc.gnu.org/gcc-4.8/changes.html
-#if defined(ABSL_HAVE_ADDRESS_SANITIZER) && \
-    ABSL_HAVE_ATTRIBUTE(no_sanitize_address)
-#define ABSL_ATTRIBUTE_NO_SANITIZE_ADDRESS __attribute__((no_sanitize_address))
-#elif defined(ABSL_HAVE_ADDRESS_SANITIZER) && defined(_MSC_VER) && \
-    _MSC_VER >= 1928
-// https://docs.microsoft.com/en-us/cpp/cpp/no-sanitize-address
-#define ABSL_ATTRIBUTE_NO_SANITIZE_ADDRESS __declspec(no_sanitize_address)
-#elif defined(ABSL_HAVE_HWADDRESS_SANITIZER) && ABSL_HAVE_ATTRIBUTE(no_sanitize)
-// HWAddressSanitizer is a sanitizer similar to AddressSanitizer, which uses CPU
-// features to detect similar bugs with less CPU and memory overhead.
-// NOTE: GCC supports HWAddressSanitizer(hwasan) since 11.
-// https://gcc.gnu.org/gcc-11/changes.html
-#define ABSL_ATTRIBUTE_NO_SANITIZE_ADDRESS \
-  __attribute__((no_sanitize("hwaddress")))
-#else
-#define ABSL_ATTRIBUTE_NO_SANITIZE_ADDRESS
-#endif
-
-// ABSL_ATTRIBUTE_NO_SANITIZE_MEMORY
-//
-// Tells the MemorySanitizer to relax the handling of a given function. All "Use
-// of uninitialized value" warnings from such functions will be suppressed, and
-// all values loaded from memory will be considered fully initialized.  This
-// attribute is similar to the ABSL_ATTRIBUTE_NO_SANITIZE_ADDRESS attribute
-// above, but deals with initialized-ness rather than addressability issues.
-// NOTE: MemorySanitizer(msan) is supported by Clang but not GCC.
-#if ABSL_HAVE_ATTRIBUTE(no_sanitize_memory)
-#define ABSL_ATTRIBUTE_NO_SANITIZE_MEMORY __attribute__((no_sanitize_memory))
-#else
-#define ABSL_ATTRIBUTE_NO_SANITIZE_MEMORY
-#endif
-
-// ABSL_ATTRIBUTE_NO_SANITIZE_THREAD
-//
-// Tells the ThreadSanitizer to not instrument a given function.
-// NOTE: GCC supports ThreadSanitizer(tsan) since 4.8.
-// https://gcc.gnu.org/gcc-4.8/changes.html
-#if ABSL_HAVE_ATTRIBUTE(no_sanitize_thread)
-#define ABSL_ATTRIBUTE_NO_SANITIZE_THREAD __attribute__((no_sanitize_thread))
-#else
-#define ABSL_ATTRIBUTE_NO_SANITIZE_THREAD
-#endif
-
-// ABSL_ATTRIBUTE_NO_SANITIZE_UNDEFINED
-//
-// Tells the UndefinedSanitizer to ignore a given function. Useful for cases
-// where certain behavior (eg. division by zero) is being used intentionally.
-// NOTE: GCC supports UndefinedBehaviorSanitizer(ubsan) since 4.9.
-// https://gcc.gnu.org/gcc-4.9/changes.html
-#if ABSL_HAVE_ATTRIBUTE(no_sanitize_undefined)
-#define ABSL_ATTRIBUTE_NO_SANITIZE_UNDEFINED \
-  __attribute__((no_sanitize_undefined))
-#elif ABSL_HAVE_ATTRIBUTE(no_sanitize)
-#define ABSL_ATTRIBUTE_NO_SANITIZE_UNDEFINED \
-  __attribute__((no_sanitize("undefined")))
-#else
-#define ABSL_ATTRIBUTE_NO_SANITIZE_UNDEFINED
-#endif
-
-// ABSL_ATTRIBUTE_NO_SANITIZE_CFI
-//
-// Tells the ControlFlowIntegrity sanitizer to not instrument a given function.
-// See https://clang.llvm.org/docs/ControlFlowIntegrity.html for details.
-#if ABSL_HAVE_ATTRIBUTE(no_sanitize) && defined(__llvm__)
-#define ABSL_ATTRIBUTE_NO_SANITIZE_CFI __attribute__((no_sanitize("cfi")))
-#else
-#define ABSL_ATTRIBUTE_NO_SANITIZE_CFI
-#endif
-
-// ABSL_ATTRIBUTE_NO_SANITIZE_SAFESTACK
-//
-// Tells the SafeStack to not instrument a given function.
-// See https://clang.llvm.org/docs/SafeStack.html for details.
-#if ABSL_HAVE_ATTRIBUTE(no_sanitize)
-#define ABSL_ATTRIBUTE_NO_SANITIZE_SAFESTACK \
-  __attribute__((no_sanitize("safe-stack")))
-#else
-#define ABSL_ATTRIBUTE_NO_SANITIZE_SAFESTACK
-#endif
-
-// ABSL_ATTRIBUTE_RETURNS_NONNULL
-//
-// Tells the compiler that a particular function never returns a null pointer.
-#if ABSL_HAVE_ATTRIBUTE(returns_nonnull)
-#define ABSL_ATTRIBUTE_RETURNS_NONNULL __attribute__((returns_nonnull))
-#else
-#define ABSL_ATTRIBUTE_RETURNS_NONNULL
-#endif
-
-// ABSL_HAVE_ATTRIBUTE_SECTION
-//
-// Indicates whether labeled sections are supported. Weak symbol support is
-// a prerequisite. Labeled sections are not supported on Darwin/iOS.
-#ifdef ABSL_HAVE_ATTRIBUTE_SECTION
-#error ABSL_HAVE_ATTRIBUTE_SECTION cannot be directly set
-#elif (ABSL_HAVE_ATTRIBUTE(section) ||                \
-       (defined(__GNUC__) && !defined(__clang__))) && \
-    !defined(__APPLE__) && ABSL_HAVE_ATTRIBUTE_WEAK
-#define ABSL_HAVE_ATTRIBUTE_SECTION 1
-
-// ABSL_ATTRIBUTE_SECTION
-//
-// Tells the compiler/linker to put a given function into a section and define
-// `__start_ ## name` and `__stop_ ## name` symbols to bracket the section.
-// This functionality is supported by GNU linker.  Any function annotated with
-// `ABSL_ATTRIBUTE_SECTION` must not be inlined, or it will be placed into
-// whatever section its caller is placed into.
-//
-#ifndef ABSL_ATTRIBUTE_SECTION
-#define ABSL_ATTRIBUTE_SECTION(name) \
-  __attribute__((section(#name))) __attribute__((noinline))
-#endif
-
-// ABSL_ATTRIBUTE_SECTION_VARIABLE
-//
-// Tells the compiler/linker to put a given variable into a section and define
-// `__start_ ## name` and `__stop_ ## name` symbols to bracket the section.
-// This functionality is supported by GNU linker.
-#ifndef ABSL_ATTRIBUTE_SECTION_VARIABLE
-#ifdef _AIX
-// __attribute__((section(#name))) on AIX is achieved by using the `.csect`
-// psudo op which includes an additional integer as part of its syntax indcating
-// alignment. If data fall under different alignments then you might get a
-// compilation error indicating a `Section type conflict`.
-#define ABSL_ATTRIBUTE_SECTION_VARIABLE(name)
-#else
-#define ABSL_ATTRIBUTE_SECTION_VARIABLE(name) __attribute__((section(#name)))
-#endif
-#endif
-
-// ABSL_DECLARE_ATTRIBUTE_SECTION_VARS
-//
-// A weak section declaration to be used as a global declaration
-// for ABSL_ATTRIBUTE_SECTION_START|STOP(name) to compile and link
-// even without functions with ABSL_ATTRIBUTE_SECTION(name).
-// ABSL_DEFINE_ATTRIBUTE_SECTION should be in the exactly one file; it's
-// a no-op on ELF but not on Mach-O.
-//
-#ifndef ABSL_DECLARE_ATTRIBUTE_SECTION_VARS
-#define ABSL_DECLARE_ATTRIBUTE_SECTION_VARS(name)   \
-  extern char __start_##name[] ABSL_ATTRIBUTE_WEAK; \
-  extern char __stop_##name[] ABSL_ATTRIBUTE_WEAK
-#endif
-#ifndef ABSL_DEFINE_ATTRIBUTE_SECTION_VARS
-#define ABSL_INIT_ATTRIBUTE_SECTION_VARS(name)
-#define ABSL_DEFINE_ATTRIBUTE_SECTION_VARS(name)
-#endif
-
-// ABSL_ATTRIBUTE_SECTION_START
-//
-// Returns `void*` pointers to start/end of a section of code with
-// functions having ABSL_ATTRIBUTE_SECTION(name).
-// Returns 0 if no such functions exist.
-// One must ABSL_DECLARE_ATTRIBUTE_SECTION_VARS(name) for this to compile and
-// link.
-//
-#define ABSL_ATTRIBUTE_SECTION_START(name) \
-  (reinterpret_cast<void *>(__start_##name))
-#define ABSL_ATTRIBUTE_SECTION_STOP(name) \
-  (reinterpret_cast<void *>(__stop_##name))
-
-#else  // !ABSL_HAVE_ATTRIBUTE_SECTION
-
-#define ABSL_HAVE_ATTRIBUTE_SECTION 0
-
-// provide dummy definitions
-#define ABSL_ATTRIBUTE_SECTION(name)
-#define ABSL_ATTRIBUTE_SECTION_VARIABLE(name)
-#define ABSL_INIT_ATTRIBUTE_SECTION_VARS(name)
-#define ABSL_DEFINE_ATTRIBUTE_SECTION_VARS(name)
-#define ABSL_DECLARE_ATTRIBUTE_SECTION_VARS(name)
-#define ABSL_ATTRIBUTE_SECTION_START(name) (reinterpret_cast<void *>(0))
-#define ABSL_ATTRIBUTE_SECTION_STOP(name) (reinterpret_cast<void *>(0))
-
-#endif  // ABSL_ATTRIBUTE_SECTION
-
-// ABSL_ATTRIBUTE_STACK_ALIGN_FOR_OLD_LIBC
-//
-// Support for aligning the stack on 32-bit x86.
-#if ABSL_HAVE_ATTRIBUTE(force_align_arg_pointer) || \
-    (defined(__GNUC__) && !defined(__clang__))
-#if defined(__i386__)
-#define ABSL_ATTRIBUTE_STACK_ALIGN_FOR_OLD_LIBC \
-  __attribute__((force_align_arg_pointer))
-#define ABSL_REQUIRE_STACK_ALIGN_TRAMPOLINE (0)
-#elif defined(__x86_64__)
-#define ABSL_REQUIRE_STACK_ALIGN_TRAMPOLINE (1)
-#define ABSL_ATTRIBUTE_STACK_ALIGN_FOR_OLD_LIBC
-#else  // !__i386__ && !__x86_64
-#define ABSL_REQUIRE_STACK_ALIGN_TRAMPOLINE (0)
-#define ABSL_ATTRIBUTE_STACK_ALIGN_FOR_OLD_LIBC
-#endif  // __i386__
-#else
-#define ABSL_ATTRIBUTE_STACK_ALIGN_FOR_OLD_LIBC
-#define ABSL_REQUIRE_STACK_ALIGN_TRAMPOLINE (0)
-#endif
-
-// ABSL_MUST_USE_RESULT
-//
-// Tells the compiler to warn about unused results.
-//
-// For code or headers that are assured to only build with C++17 and up, prefer
-// just using the standard `[[nodiscard]]` directly over this macro.
-//
-// When annotating a function, it must appear as the first part of the
-// declaration or definition. The compiler will warn if the return value from
-// such a function is unused:
-//
-//   ABSL_MUST_USE_RESULT Sprocket* AllocateSprocket();
-//   AllocateSprocket();  // Triggers a warning.
-//
-// When annotating a class, it is equivalent to annotating every function which
-// returns an instance.
-//
-//   class ABSL_MUST_USE_RESULT Sprocket {};
-//   Sprocket();  // Triggers a warning.
-//
-//   Sprocket MakeSprocket();
-//   MakeSprocket();  // Triggers a warning.
-//
-// Note that references and pointers are not instances:
-//
-//   Sprocket* SprocketPointer();
-//   SprocketPointer();  // Does *not* trigger a warning.
-//
-// ABSL_MUST_USE_RESULT allows using cast-to-void to suppress the unused result
-// warning. For that, warn_unused_result is used only for clang but not for gcc.
-// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66425
-//
-// Note: past advice was to place the macro after the argument list.
-//
-// TODO(b/176172494): Use ABSL_HAVE_CPP_ATTRIBUTE(nodiscard) when all code is
-// compliant with the stricter [[nodiscard]].
-#if defined(__clang__) && ABSL_HAVE_ATTRIBUTE(warn_unused_result)
-#define ABSL_MUST_USE_RESULT __attribute__((warn_unused_result))
-#else
-#define ABSL_MUST_USE_RESULT
-#endif
-
-// ABSL_ATTRIBUTE_HOT, ABSL_ATTRIBUTE_COLD
-//
-// Tells GCC that a function is hot or cold. GCC can use this information to
-// improve static analysis, i.e. a conditional branch to a cold function
-// is likely to be not-taken.
-// This annotation is used for function declarations.
-//
-// Example:
-//
-//   int foo() ABSL_ATTRIBUTE_HOT;
-#if ABSL_HAVE_ATTRIBUTE(hot) || (defined(__GNUC__) && !defined(__clang__))
-#define ABSL_ATTRIBUTE_HOT __attribute__((hot))
-#else
-#define ABSL_ATTRIBUTE_HOT
-#endif
-
-#if ABSL_HAVE_ATTRIBUTE(cold) || (defined(__GNUC__) && !defined(__clang__))
-#define ABSL_ATTRIBUTE_COLD __attribute__((cold))
-#else
-#define ABSL_ATTRIBUTE_COLD
-#endif
-
-// ABSL_XRAY_ALWAYS_INSTRUMENT, ABSL_XRAY_NEVER_INSTRUMENT, ABSL_XRAY_LOG_ARGS
-//
-// We define the ABSL_XRAY_ALWAYS_INSTRUMENT and ABSL_XRAY_NEVER_INSTRUMENT
-// macro used as an attribute to mark functions that must always or never be
-// instrumented by XRay. Currently, this is only supported in Clang/LLVM.
-//
-// For reference on the LLVM XRay instrumentation, see
-// http://llvm.org/docs/XRay.html.
-//
-// A function with the XRAY_ALWAYS_INSTRUMENT macro attribute in its declaration
-// will always get the XRay instrumentation sleds. These sleds may introduce
-// some binary size and runtime overhead and must be used sparingly.
-//
-// These attributes only take effect when the following conditions are met:
-//
-//   * The file/target is built in at least C++11 mode, with a Clang compiler
-//     that supports XRay attributes.
-//   * The file/target is built with the -fxray-instrument flag set for the
-//     Clang/LLVM compiler.
-//   * The function is defined in the translation unit (the compiler honors the
-//     attribute in either the definition or the declaration, and must match).
-//
-// There are cases when, even when building with XRay instrumentation, users
-// might want to control specifically which functions are instrumented for a
-// particular build using special-case lists provided to the compiler. These
-// special case lists are provided to Clang via the
-// -fxray-always-instrument=... and -fxray-never-instrument=... flags. The
-// attributes in source take precedence over these special-case lists.
-//
-// To disable the XRay attributes at build-time, users may define
-// ABSL_NO_XRAY_ATTRIBUTES. Do NOT define ABSL_NO_XRAY_ATTRIBUTES on specific
-// packages/targets, as this may lead to conflicting definitions of functions at
-// link-time.
-//
-// XRay isn't currently supported on Android:
-// https://github.com/android/ndk/issues/368
-#if ABSL_HAVE_CPP_ATTRIBUTE(clang::xray_always_instrument) && \
-    !defined(ABSL_NO_XRAY_ATTRIBUTES) && !defined(__ANDROID__)
-#define ABSL_XRAY_ALWAYS_INSTRUMENT [[clang::xray_always_instrument]]
-#define ABSL_XRAY_NEVER_INSTRUMENT [[clang::xray_never_instrument]]
-#if ABSL_HAVE_CPP_ATTRIBUTE(clang::xray_log_args)
-#define ABSL_XRAY_LOG_ARGS(N) \
-  [[clang::xray_always_instrument, clang::xray_log_args(N)]]
-#else
-#define ABSL_XRAY_LOG_ARGS(N) [[clang::xray_always_instrument]]
-#endif
-#else
-#define ABSL_XRAY_ALWAYS_INSTRUMENT
-#define ABSL_XRAY_NEVER_INSTRUMENT
-#define ABSL_XRAY_LOG_ARGS(N)
-#endif
-
-// ABSL_ATTRIBUTE_REINITIALIZES
-//
-// Indicates that a member function reinitializes the entire object to a known
-// state, independent of the previous state of the object.
-//
-// The clang-tidy check bugprone-use-after-move allows member functions marked
-// with this attribute to be called on objects that have been moved from;
-// without the attribute, this would result in a use-after-move warning.
-#if ABSL_HAVE_CPP_ATTRIBUTE(clang::reinitializes)
-#define ABSL_ATTRIBUTE_REINITIALIZES [[clang::reinitializes]]
-#else
-#define ABSL_ATTRIBUTE_REINITIALIZES
-#endif
-
-// -----------------------------------------------------------------------------
-// Variable Attributes
-// -----------------------------------------------------------------------------
-
-// ABSL_ATTRIBUTE_UNUSED
-//
-// Prevents the compiler from complaining about variables that appear unused.
-//
-// For code or headers that are assured to only build with C++17 and up, prefer
-// just using the standard '[[maybe_unused]]' directly over this macro.
-//
-// Due to differences in positioning requirements between the old, compiler
-// specific __attribute__ syntax and the now standard [[maybe_unused]], this
-// macro does not attempt to take advantage of '[[maybe_unused]]'.
-#if ABSL_HAVE_ATTRIBUTE(unused) || (defined(__GNUC__) && !defined(__clang__))
-#undef ABSL_ATTRIBUTE_UNUSED
-#define ABSL_ATTRIBUTE_UNUSED __attribute__((__unused__))
-#else
-#define ABSL_ATTRIBUTE_UNUSED
-#endif
-
-// ABSL_ATTRIBUTE_INITIAL_EXEC
-//
-// Tells the compiler to use "initial-exec" mode for a thread-local variable.
-// See http://people.redhat.com/drepper/tls.pdf for the gory details.
-#if ABSL_HAVE_ATTRIBUTE(tls_model) || (defined(__GNUC__) && !defined(__clang__))
-#define ABSL_ATTRIBUTE_INITIAL_EXEC __attribute__((tls_model("initial-exec")))
-#else
-#define ABSL_ATTRIBUTE_INITIAL_EXEC
-#endif
-
-// ABSL_ATTRIBUTE_PACKED
-//
-// Instructs the compiler not to use natural alignment for a tagged data
-// structure, but instead to reduce its alignment to 1.
-//
-// Therefore, DO NOT APPLY THIS ATTRIBUTE TO STRUCTS CONTAINING ATOMICS. Doing
-// so can cause atomic variables to be mis-aligned and silently violate
-// atomicity on x86.
-//
-// This attribute can either be applied to members of a structure or to a
-// structure in its entirety. Applying this attribute (judiciously) to a
-// structure in its entirety to optimize the memory footprint of very
-// commonly-used structs is fine. Do not apply this attribute to a structure in
-// its entirety if the purpose is to control the offsets of the members in the
-// structure. Instead, apply this attribute only to structure members that need
-// it.
-//
-// When applying ABSL_ATTRIBUTE_PACKED only to specific structure members the
-// natural alignment of structure members not annotated is preserved. Aligned
-// member accesses are faster than non-aligned member accesses even if the
-// targeted microprocessor supports non-aligned accesses.
-#if ABSL_HAVE_ATTRIBUTE(packed) || (defined(__GNUC__) && !defined(__clang__))
-#define ABSL_ATTRIBUTE_PACKED __attribute__((__packed__))
-#else
-#define ABSL_ATTRIBUTE_PACKED
-#endif
-
-// ABSL_ATTRIBUTE_FUNC_ALIGN
-//
-// Tells the compiler to align the function start at least to certain
-// alignment boundary
-#if ABSL_HAVE_ATTRIBUTE(aligned) || (defined(__GNUC__) && !defined(__clang__))
-#define ABSL_ATTRIBUTE_FUNC_ALIGN(bytes) __attribute__((aligned(bytes)))
-#else
-#define ABSL_ATTRIBUTE_FUNC_ALIGN(bytes)
-#endif
-
-// ABSL_FALLTHROUGH_INTENDED
-//
-// Annotates implicit fall-through between switch labels, allowing a case to
-// indicate intentional fallthrough and turn off warnings about any lack of a
-// `break` statement. The ABSL_FALLTHROUGH_INTENDED macro should be followed by
-// a semicolon and can be used in most places where `break` can, provided that
-// no statements exist between it and the next switch label.
-//
-// Example:
-//
-//  switch (x) {
-//    case 40:
-//    case 41:
-//      if (truth_is_out_there) {
-//        ++x;
-//        ABSL_FALLTHROUGH_INTENDED;  // Use instead of/along with annotations
-//                                    // in comments
-//      } else {
-//        return x;
-//      }
-//    case 42:
-//      ...
-//
-// Notes: When supported, GCC and Clang can issue a warning on switch labels
-// with unannotated fallthrough using the warning `-Wimplicit-fallthrough`. See
-// clang documentation on language extensions for details:
-// https://clang.llvm.org/docs/AttributeReference.html#fallthrough-clang-fallthrough
-//
-// When used with unsupported compilers, the ABSL_FALLTHROUGH_INTENDED macro has
-// no effect on diagnostics. In any case this macro has no effect on runtime
-// behavior and performance of code.
-
-#ifdef ABSL_FALLTHROUGH_INTENDED
-#error "ABSL_FALLTHROUGH_INTENDED should not be defined."
-#elif ABSL_HAVE_CPP_ATTRIBUTE(fallthrough)
-#define ABSL_FALLTHROUGH_INTENDED [[fallthrough]]
-#elif ABSL_HAVE_CPP_ATTRIBUTE(clang::fallthrough)
-#define ABSL_FALLTHROUGH_INTENDED [[clang::fallthrough]]
-#elif ABSL_HAVE_CPP_ATTRIBUTE(gnu::fallthrough)
-#define ABSL_FALLTHROUGH_INTENDED [[gnu::fallthrough]]
-#else
-#define ABSL_FALLTHROUGH_INTENDED \
-  do {                            \
-  } while (0)
-#endif
-
-// ABSL_DEPRECATED()
-//
-// Marks a deprecated class, struct, enum, function, method and variable
-// declarations. The macro argument is used as a custom diagnostic message (e.g.
-// suggestion of a better alternative).
-//
-// For code or headers that are assured to only build with C++14 and up, prefer
-// just using the standard `[[deprecated("message")]]` directly over this macro.
-//
-// Examples:
-//
-//   class ABSL_DEPRECATED("Use Bar instead") Foo {...};
-//
-//   ABSL_DEPRECATED("Use Baz() instead") void Bar() {...}
-//
-//   template <typename T>
-//   ABSL_DEPRECATED("Use DoThat() instead")
-//   void DoThis();
-//
-//   enum FooEnum {
-//     kBar ABSL_DEPRECATED("Use kBaz instead"),
-//   };
-//
-// Every usage of a deprecated entity will trigger a warning when compiled with
-// GCC/Clang's `-Wdeprecated-declarations` option. Google's production toolchain
-// turns this warning off by default, instead relying on clang-tidy to report
-// new uses of deprecated code.
-#if ABSL_HAVE_ATTRIBUTE(deprecated)
-#define ABSL_DEPRECATED(message) __attribute__((deprecated(message)))
-#else
-#define ABSL_DEPRECATED(message)
-#endif
-
-// When deprecating Abseil code, it is sometimes necessary to turn off the
-// warning within Abseil, until the deprecated code is actually removed. The
-// deprecated code can be surrounded with these directives to achieve that
-// result.
-//
-// class ABSL_DEPRECATED("Use Bar instead") Foo;
-//
-// ABSL_INTERNAL_DISABLE_DEPRECATED_DECLARATION_WARNING
-// Baz ComputeBazFromFoo(Foo f);
-// ABSL_INTERNAL_RESTORE_DEPRECATED_DECLARATION_WARNING
-#if defined(__GNUC__) || defined(__clang__)
-// Clang also supports these GCC pragmas.
-#define ABSL_INTERNAL_DISABLE_DEPRECATED_DECLARATION_WARNING \
-  _Pragma("GCC diagnostic push")             \
-  _Pragma("GCC diagnostic ignored \"-Wdeprecated-declarations\"")
-#define ABSL_INTERNAL_RESTORE_DEPRECATED_DECLARATION_WARNING \
-  _Pragma("GCC diagnostic pop")
-#else
-#define ABSL_INTERNAL_DISABLE_DEPRECATED_DECLARATION_WARNING
-#define ABSL_INTERNAL_RESTORE_DEPRECATED_DECLARATION_WARNING
-#endif  // defined(__GNUC__) || defined(__clang__)
-
-// ABSL_CONST_INIT
-//
-// A variable declaration annotated with the `ABSL_CONST_INIT` attribute will
-// not compile (on supported platforms) unless the variable has a constant
-// initializer. This is useful for variables with static and thread storage
-// duration, because it guarantees that they will not suffer from the so-called
-// "static init order fiasco".
-//
-// This attribute must be placed on the initializing declaration of the
-// variable. Some compilers will give a -Wmissing-constinit warning when this
-// attribute is placed on some other declaration but missing from the
-// initializing declaration.
-//
-// In some cases (notably with thread_local variables), `ABSL_CONST_INIT` can
-// also be used in a non-initializing declaration to tell the compiler that a
-// variable is already initialized, reducing overhead that would otherwise be
-// incurred by a hidden guard variable. Thus annotating all declarations with
-// this attribute is recommended to potentially enhance optimization.
-//
-// Example:
-//
-//   class MyClass {
-//    public:
-//     ABSL_CONST_INIT static MyType my_var;
-//   };
-//
-//   ABSL_CONST_INIT MyType MyClass::my_var = MakeMyType(...);
-//
-// For code or headers that are assured to only build with C++20 and up, prefer
-// just using the standard `constinit` keyword directly over this macro.
-//
-// Note that this attribute is redundant if the variable is declared constexpr.
-#if defined(__cpp_constinit) && __cpp_constinit >= 201907L
-#define ABSL_CONST_INIT constinit
-#elif ABSL_HAVE_CPP_ATTRIBUTE(clang::require_constant_initialization)
-#define ABSL_CONST_INIT [[clang::require_constant_initialization]]
-#else
-#define ABSL_CONST_INIT
-#endif
-
-// ABSL_ATTRIBUTE_PURE_FUNCTION
-//
-// ABSL_ATTRIBUTE_PURE_FUNCTION is used to annotate declarations of "pure"
-// functions. A function is pure if its return value is only a function of its
-// arguments. The pure attribute prohibits a function from modifying the state
-// of the program that is observable by means other than inspecting the
-// function's return value. Declaring such functions with the pure attribute
-// allows the compiler to avoid emitting some calls in repeated invocations of
-// the function with the same argument values.
-//
-// Example:
-//
-//  ABSL_ATTRIBUTE_PURE_FUNCTION std::string FormatTime(Time t);
-#if ABSL_HAVE_CPP_ATTRIBUTE(gnu::pure)
-#define ABSL_ATTRIBUTE_PURE_FUNCTION [[gnu::pure]]
-#elif ABSL_HAVE_ATTRIBUTE(pure)
-#define ABSL_ATTRIBUTE_PURE_FUNCTION __attribute__((pure))
-#else
-// If the attribute isn't defined, we'll fallback to ABSL_MUST_USE_RESULT since
-// pure functions are useless if its return is ignored.
-#define ABSL_ATTRIBUTE_PURE_FUNCTION ABSL_MUST_USE_RESULT
-#endif
-
-// ABSL_ATTRIBUTE_CONST_FUNCTION
-//
-// ABSL_ATTRIBUTE_CONST_FUNCTION is used to annotate declarations of "const"
-// functions. A const function is similar to a pure function, with one
-// exception: Pure functions may return value that depend on a non-volatile
-// object that isn't provided as a function argument, while the const function
-// is guaranteed to return the same result given the same arguments.
-//
-// Example:
-//
-//  ABSL_ATTRIBUTE_CONST_FUNCTION int64_t ToInt64Milliseconds(Duration d);
-#if defined(_MSC_VER) && !defined(__clang__)
-// Put the MSVC case first since MSVC seems to parse const as a C++ keyword.
-#define ABSL_ATTRIBUTE_CONST_FUNCTION ABSL_ATTRIBUTE_PURE_FUNCTION
-#elif ABSL_HAVE_CPP_ATTRIBUTE(gnu::const)
-#define ABSL_ATTRIBUTE_CONST_FUNCTION [[gnu::const]]
-#elif ABSL_HAVE_ATTRIBUTE(const)
-#define ABSL_ATTRIBUTE_CONST_FUNCTION __attribute__((const))
-#else
-// Since const functions are more restrictive pure function, we'll fallback to a
-// pure function if the const attribute is not handled.
-#define ABSL_ATTRIBUTE_CONST_FUNCTION ABSL_ATTRIBUTE_PURE_FUNCTION
-#endif
-
-// ABSL_ATTRIBUTE_LIFETIME_BOUND indicates that a resource owned by a function
-// parameter or implicit object parameter is retained by the return value of the
-// annotated function (or, for a parameter of a constructor, in the value of the
-// constructed object). This attribute causes warnings to be produced if a
-// temporary object does not live long enough.
-//
-// When applied to a reference parameter, the referenced object is assumed to be
-// retained by the return value of the function. When applied to a non-reference
-// parameter (for example, a pointer or a class type), all temporaries
-// referenced by the parameter are assumed to be retained by the return value of
-// the function.
-//
-// See also the upstream documentation:
-// https://clang.llvm.org/docs/AttributeReference.html#lifetimebound
-#if ABSL_HAVE_CPP_ATTRIBUTE(clang::lifetimebound)
-#define ABSL_ATTRIBUTE_LIFETIME_BOUND [[clang::lifetimebound]]
-#elif ABSL_HAVE_ATTRIBUTE(lifetimebound)
-#define ABSL_ATTRIBUTE_LIFETIME_BOUND __attribute__((lifetimebound))
-#else
-#define ABSL_ATTRIBUTE_LIFETIME_BOUND
-#endif
-
-// ABSL_ATTRIBUTE_TRIVIAL_ABI
-// Indicates that a type is "trivially relocatable" -- meaning it can be
-// relocated without invoking the constructor/destructor, using a form of move
-// elision.
-//
-// From a memory safety point of view, putting aside destructor ordering, it's
-// safe to apply ABSL_ATTRIBUTE_TRIVIAL_ABI if an object's location
-// can change over the course of its lifetime: if a constructor can be run one
-// place, and then the object magically teleports to another place where some
-// methods are run, and then the object teleports to yet another place where it
-// is destroyed. This is notably not true for self-referential types, where the
-// move-constructor must keep the self-reference up to date. If the type changed
-// location without invoking the move constructor, it would have a dangling
-// self-reference.
-//
-// The use of this teleporting machinery means that the number of paired
-// move/destroy operations can change, and so it is a bad idea to apply this to
-// a type meant to count the number of moves.
-//
-// Warning: applying this can, rarely, break callers. Objects passed by value
-// will be destroyed at the end of the call, instead of the end of the
-// full-expression containing the call. In addition, it changes the ABI
-// of functions accepting this type by value (e.g. to pass in registers).
-//
-// See also the upstream documentation:
-// https://clang.llvm.org/docs/AttributeReference.html#trivial-abi
-//
-// b/321691395 - This is currently disabled in open-source builds since
-// compiler support differs. If system libraries compiled with GCC are mixed
-// with libraries compiled with Clang, types will have different ideas about
-// their ABI, leading to hard to debug crashes.
-#define ABSL_ATTRIBUTE_TRIVIAL_ABI
-
-// ABSL_ATTRIBUTE_NO_UNIQUE_ADDRESS
-//
-// Indicates a data member can be optimized to occupy no space (if it is empty)
-// and/or its tail padding can be used for other members.
-//
-// For code that is assured to only build with C++20 or later, prefer using
-// the standard attribute `[[no_unique_address]]` directly instead of this
-// macro.
-//
-// https://devblogs.microsoft.com/cppblog/msvc-cpp20-and-the-std-cpp20-switch/#c20-no_unique_address
-// Current versions of MSVC have disabled `[[no_unique_address]]` since it
-// breaks ABI compatibility, but offers `[[msvc::no_unique_address]]` for
-// situations when it can be assured that it is desired. Since Abseil does not
-// claim ABI compatibility in mixed builds, we can offer it unconditionally.
-#if defined(_MSC_VER) && _MSC_VER >= 1929
-#define ABSL_ATTRIBUTE_NO_UNIQUE_ADDRESS [[msvc::no_unique_address]]
-#elif ABSL_HAVE_CPP_ATTRIBUTE(no_unique_address)
-#define ABSL_ATTRIBUTE_NO_UNIQUE_ADDRESS [[no_unique_address]]
-#else
-#define ABSL_ATTRIBUTE_NO_UNIQUE_ADDRESS
-#endif
-
-// ABSL_ATTRIBUTE_UNINITIALIZED
-//
-// GCC and Clang support a flag `-ftrivial-auto-var-init=<option>` (<option>
-// can be "zero" or "pattern") that can be used to initialize automatic stack
-// variables. Variables with this attribute will be left uninitialized,
-// overriding the compiler flag.
-//
-// See https://clang.llvm.org/docs/AttributeReference.html#uninitialized
-// and https://gcc.gnu.org/onlinedocs/gcc/Common-Variable-Attributes.html#index-uninitialized-variable-attribute
-#if ABSL_HAVE_CPP_ATTRIBUTE(clang::uninitialized)
-#define ABSL_ATTRIBUTE_UNINITIALIZED [[clang::uninitialized]]
-#elif ABSL_HAVE_CPP_ATTRIBUTE(gnu::uninitialized)
-#define ABSL_ATTRIBUTE_UNINITIALIZED [[gnu::uninitialized]]
-#elif ABSL_HAVE_ATTRIBUTE(uninitialized)
-#define ABSL_ATTRIBUTE_UNINITIALIZED __attribute__((uninitialized))
-#else
-#define ABSL_ATTRIBUTE_UNINITIALIZED
-#endif
-
-// ABSL_ATTRIBUTE_WARN_UNUSED
-//
-// Compilers routinely warn about trivial variables that are unused.  For
-// non-trivial types, this warning is suppressed since the
-// constructor/destructor may be intentional and load-bearing, for example, with
-// a RAII scoped lock.
-//
-// For example:
-//
-// class ABSL_ATTRIBUTE_WARN_UNUSED MyType {
-//  public:
-//   MyType();
-//   ~MyType();
-// };
-//
-// void foo() {
-//   // Warns with ABSL_ATTRIBUTE_WARN_UNUSED attribute present.
-//   MyType unused;
-// }
-//
-// See https://clang.llvm.org/docs/AttributeReference.html#warn-unused and
-// https://gcc.gnu.org/onlinedocs/gcc/C_002b_002b-Attributes.html#index-warn_005funused-type-attribute
-#if ABSL_HAVE_CPP_ATTRIBUTE(gnu::warn_unused)
-#define ABSL_ATTRIBUTE_WARN_UNUSED [[gnu::warn_unused]]
-#else
-#define ABSL_ATTRIBUTE_WARN_UNUSED
-#endif
-
-#endif  // ABSL_BASE_ATTRIBUTES_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/call_once.h b/projects/llm_framework/include/abseil-cpp/absl/base/call_once.h
deleted file mode 100644
index 7b0e69cc..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/call_once.h
+++ /dev/null
@@ -1,225 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: call_once.h
-// -----------------------------------------------------------------------------
-//
-// This header file provides an Abseil version of `std::call_once` for invoking
-// a given function at most once, across all threads. This Abseil version is
-// faster than the C++11 version and incorporates the C++17 argument-passing
-// fix, so that (for example) non-const references may be passed to the invoked
-// function.
-
-#ifndef ABSL_BASE_CALL_ONCE_H_
-#define ABSL_BASE_CALL_ONCE_H_
-
-#include <algorithm>
-#include <atomic>
-#include <cstdint>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/internal/invoke.h"
-#include "absl/base/internal/low_level_scheduling.h"
-#include "absl/base/internal/raw_logging.h"
-#include "absl/base/internal/scheduling_mode.h"
-#include "absl/base/internal/spinlock_wait.h"
-#include "absl/base/macros.h"
-#include "absl/base/nullability.h"
-#include "absl/base/optimization.h"
-#include "absl/base/port.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-class once_flag;
-
-namespace base_internal {
-absl::Nonnull<std::atomic<uint32_t>*> ControlWord(
-    absl::Nonnull<absl::once_flag*> flag);
-}  // namespace base_internal
-
-// call_once()
-//
-// For all invocations using a given `once_flag`, invokes a given `fn` exactly
-// once across all threads. The first call to `call_once()` with a particular
-// `once_flag` argument (that does not throw an exception) will run the
-// specified function with the provided `args`; other calls with the same
-// `once_flag` argument will not run the function, but will wait
-// for the provided function to finish running (if it is still running).
-//
-// This mechanism provides a safe, simple, and fast mechanism for one-time
-// initialization in a multi-threaded process.
-//
-// Example:
-//
-// class MyInitClass {
-//  public:
-//  ...
-//  mutable absl::once_flag once_;
-//
-//  MyInitClass* init() const {
-//    absl::call_once(once_, &MyInitClass::Init, this);
-//    return ptr_;
-//  }
-//
-template <typename Callable, typename... Args>
-void call_once(absl::once_flag& flag, Callable&& fn, Args&&... args);
-
-// once_flag
-//
-// Objects of this type are used to distinguish calls to `call_once()` and
-// ensure the provided function is only invoked once across all threads. This
-// type is not copyable or movable. However, it has a `constexpr`
-// constructor, and is safe to use as a namespace-scoped global variable.
-class once_flag {
- public:
-  constexpr once_flag() : control_(0) {}
-  once_flag(const once_flag&) = delete;
-  once_flag& operator=(const once_flag&) = delete;
-
- private:
-  friend absl::Nonnull<std::atomic<uint32_t>*> base_internal::ControlWord(
-      absl::Nonnull<once_flag*> flag);
-  std::atomic<uint32_t> control_;
-};
-
-//------------------------------------------------------------------------------
-// End of public interfaces.
-// Implementation details follow.
-//------------------------------------------------------------------------------
-
-namespace base_internal {
-
-// Like call_once, but uses KERNEL_ONLY scheduling. Intended to be used to
-// initialize entities used by the scheduler implementation.
-template <typename Callable, typename... Args>
-void LowLevelCallOnce(absl::Nonnull<absl::once_flag*> flag, Callable&& fn,
-                      Args&&... args);
-
-// Disables scheduling while on stack when scheduling mode is non-cooperative.
-// No effect for cooperative scheduling modes.
-class SchedulingHelper {
- public:
-  explicit SchedulingHelper(base_internal::SchedulingMode mode) : mode_(mode) {
-    if (mode_ == base_internal::SCHEDULE_KERNEL_ONLY) {
-      guard_result_ = base_internal::SchedulingGuard::DisableRescheduling();
-    }
-  }
-
-  ~SchedulingHelper() {
-    if (mode_ == base_internal::SCHEDULE_KERNEL_ONLY) {
-      base_internal::SchedulingGuard::EnableRescheduling(guard_result_);
-    }
-  }
-
- private:
-  base_internal::SchedulingMode mode_;
-  bool guard_result_ = false;
-};
-
-// Bit patterns for call_once state machine values.  Internal implementation
-// detail, not for use by clients.
-//
-// The bit patterns are arbitrarily chosen from unlikely values, to aid in
-// debugging.  However, kOnceInit must be 0, so that a zero-initialized
-// once_flag will be valid for immediate use.
-enum {
-  kOnceInit = 0,
-  kOnceRunning = 0x65C2937B,
-  kOnceWaiter = 0x05A308D2,
-  // A very small constant is chosen for kOnceDone so that it fit in a single
-  // compare with immediate instruction for most common ISAs.  This is verified
-  // for x86, POWER and ARM.
-  kOnceDone = 221,    // Random Number
-};
-
-template <typename Callable, typename... Args>
-ABSL_ATTRIBUTE_NOINLINE void CallOnceImpl(
-    absl::Nonnull<std::atomic<uint32_t>*> control,
-    base_internal::SchedulingMode scheduling_mode, Callable&& fn,
-    Args&&... args) {
-#ifndef NDEBUG
-  {
-    uint32_t old_control = control->load(std::memory_order_relaxed);
-    if (old_control != kOnceInit &&
-        old_control != kOnceRunning &&
-        old_control != kOnceWaiter &&
-        old_control != kOnceDone) {
-      ABSL_RAW_LOG(FATAL, "Unexpected value for control word: 0x%lx",
-                   static_cast<unsigned long>(old_control));  // NOLINT
-    }
-  }
-#endif  // NDEBUG
-  static const base_internal::SpinLockWaitTransition trans[] = {
-      {kOnceInit, kOnceRunning, true},
-      {kOnceRunning, kOnceWaiter, false},
-      {kOnceDone, kOnceDone, true}};
-
-  // Must do this before potentially modifying control word's state.
-  base_internal::SchedulingHelper maybe_disable_scheduling(scheduling_mode);
-  // Short circuit the simplest case to avoid procedure call overhead.
-  // The base_internal::SpinLockWait() call returns either kOnceInit or
-  // kOnceDone. If it returns kOnceDone, it must have loaded the control word
-  // with std::memory_order_acquire and seen a value of kOnceDone.
-  uint32_t old_control = kOnceInit;
-  if (control->compare_exchange_strong(old_control, kOnceRunning,
-                                       std::memory_order_relaxed) ||
-      base_internal::SpinLockWait(control, ABSL_ARRAYSIZE(trans), trans,
-                                  scheduling_mode) == kOnceInit) {
-    base_internal::invoke(std::forward<Callable>(fn),
-                          std::forward<Args>(args)...);
-    old_control =
-        control->exchange(base_internal::kOnceDone, std::memory_order_release);
-    if (old_control == base_internal::kOnceWaiter) {
-      base_internal::SpinLockWake(control, true);
-    }
-  }  // else *control is already kOnceDone
-}
-
-inline absl::Nonnull<std::atomic<uint32_t>*> ControlWord(
-    absl::Nonnull<once_flag*> flag) {
-  return &flag->control_;
-}
-
-template <typename Callable, typename... Args>
-void LowLevelCallOnce(absl::Nonnull<absl::once_flag*> flag, Callable&& fn,
-                      Args&&... args) {
-  std::atomic<uint32_t>* once = base_internal::ControlWord(flag);
-  uint32_t s = once->load(std::memory_order_acquire);
-  if (ABSL_PREDICT_FALSE(s != base_internal::kOnceDone)) {
-    base_internal::CallOnceImpl(once, base_internal::SCHEDULE_KERNEL_ONLY,
-                                std::forward<Callable>(fn),
-                                std::forward<Args>(args)...);
-  }
-}
-
-}  // namespace base_internal
-
-template <typename Callable, typename... Args>
-void call_once(absl::once_flag& flag, Callable&& fn, Args&&... args) {
-  std::atomic<uint32_t>* once = base_internal::ControlWord(&flag);
-  uint32_t s = once->load(std::memory_order_acquire);
-  if (ABSL_PREDICT_FALSE(s != base_internal::kOnceDone)) {
-    base_internal::CallOnceImpl(
-        once, base_internal::SCHEDULE_COOPERATIVE_AND_KERNEL,
-        std::forward<Callable>(fn), std::forward<Args>(args)...);
-  }
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_CALL_ONCE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/casts.h b/projects/llm_framework/include/abseil-cpp/absl/base/casts.h
deleted file mode 100644
index e0b11bbe..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/casts.h
+++ /dev/null
@@ -1,180 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: casts.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines casting templates to fit use cases not covered by
-// the standard casts provided in the C++ standard. As with all cast operations,
-// use these with caution and only if alternatives do not exist.
-
-#ifndef ABSL_BASE_CASTS_H_
-#define ABSL_BASE_CASTS_H_
-
-#include <cstring>
-#include <memory>
-#include <type_traits>
-#include <utility>
-
-#if defined(__cpp_lib_bit_cast) && __cpp_lib_bit_cast >= 201806L
-#include <bit>  // For std::bit_cast.
-#endif  // defined(__cpp_lib_bit_cast) && __cpp_lib_bit_cast >= 201806L
-
-#include "absl/base/internal/identity.h"
-#include "absl/base/macros.h"
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// implicit_cast()
-//
-// Performs an implicit conversion between types following the language
-// rules for implicit conversion; if an implicit conversion is otherwise
-// allowed by the language in the given context, this function performs such an
-// implicit conversion.
-//
-// Example:
-//
-//   // If the context allows implicit conversion:
-//   From from;
-//   To to = from;
-//
-//   // Such code can be replaced by:
-//   implicit_cast<To>(from);
-//
-// An `implicit_cast()` may also be used to annotate numeric type conversions
-// that, although safe, may produce compiler warnings (such as `long` to `int`).
-// Additionally, an `implicit_cast()` is also useful within return statements to
-// indicate a specific implicit conversion is being undertaken.
-//
-// Example:
-//
-//   return implicit_cast<double>(size_in_bytes) / capacity_;
-//
-// Annotating code with `implicit_cast()` allows you to explicitly select
-// particular overloads and template instantiations, while providing a safer
-// cast than `reinterpret_cast()` or `static_cast()`.
-//
-// Additionally, an `implicit_cast()` can be used to allow upcasting within a
-// type hierarchy where incorrect use of `static_cast()` could accidentally
-// allow downcasting.
-//
-// Finally, an `implicit_cast()` can be used to perform implicit conversions
-// from unrelated types that otherwise couldn't be implicitly cast directly;
-// C++ will normally only implicitly cast "one step" in such conversions.
-//
-// That is, if C is a type which can be implicitly converted to B, with B being
-// a type that can be implicitly converted to A, an `implicit_cast()` can be
-// used to convert C to B (which the compiler can then implicitly convert to A
-// using language rules).
-//
-// Example:
-//
-//   // Assume an object C is convertible to B, which is implicitly convertible
-//   // to A
-//   A a = implicit_cast<B>(C);
-//
-// Such implicit cast chaining may be useful within template logic.
-template <typename To>
-constexpr To implicit_cast(typename absl::internal::type_identity_t<To> to) {
-  return to;
-}
-
-// bit_cast()
-//
-// Creates a value of the new type `Dest` whose representation is the same as
-// that of the argument, which is of (deduced) type `Source` (a "bitwise cast";
-// every bit in the value representation of the result is equal to the
-// corresponding bit in the object representation of the source). Source and
-// destination types must be of the same size, and both types must be trivially
-// copyable.
-//
-// As with most casts, use with caution. A `bit_cast()` might be needed when you
-// need to treat a value as the value of some other type, for example, to access
-// the individual bits of an object which are not normally accessible through
-// the object's type, such as for working with the binary representation of a
-// floating point value:
-//
-//   float f = 3.14159265358979;
-//   int i = bit_cast<int>(f);
-//   // i = 0x40490fdb
-//
-// Reinterpreting and accessing a value directly as a different type (as shown
-// below) usually results in undefined behavior.
-//
-// Example:
-//
-//   // WRONG
-//   float f = 3.14159265358979;
-//   int i = reinterpret_cast<int&>(f);    // Wrong
-//   int j = *reinterpret_cast<int*>(&f);  // Equally wrong
-//   int k = *bit_cast<int*>(&f);          // Equally wrong
-//
-// Reinterpret-casting results in undefined behavior according to the ISO C++
-// specification, section [basic.lval]. Roughly, this section says: if an object
-// in memory has one type, and a program accesses it with a different type, the
-// result is undefined behavior for most "different type".
-//
-// Using bit_cast on a pointer and then dereferencing it is no better than using
-// reinterpret_cast. You should only use bit_cast on the value itself.
-//
-// Such casting results in type punning: holding an object in memory of one type
-// and reading its bits back using a different type. A `bit_cast()` avoids this
-// issue by copying the object representation to a new value, which avoids
-// introducing this undefined behavior (since the original value is never
-// accessed in the wrong way).
-//
-// The requirements of `absl::bit_cast` are more strict than that of
-// `std::bit_cast` unless compiler support is available. Specifically, without
-// compiler support, this implementation also requires `Dest` to be
-// default-constructible. In C++20, `absl::bit_cast` is replaced by
-// `std::bit_cast`.
-#if defined(__cpp_lib_bit_cast) && __cpp_lib_bit_cast >= 201806L
-
-using std::bit_cast;
-
-#else  // defined(__cpp_lib_bit_cast) && __cpp_lib_bit_cast >= 201806L
-
-template <
-    typename Dest, typename Source,
-    typename std::enable_if<sizeof(Dest) == sizeof(Source) &&
-                                std::is_trivially_copyable<Source>::value &&
-                                std::is_trivially_copyable<Dest>::value
-#if !ABSL_HAVE_BUILTIN(__builtin_bit_cast)
-                                && std::is_default_constructible<Dest>::value
-#endif  // !ABSL_HAVE_BUILTIN(__builtin_bit_cast)
-                            ,
-                            int>::type = 0>
-#if ABSL_HAVE_BUILTIN(__builtin_bit_cast)
-inline constexpr Dest bit_cast(const Source& source) {
-  return __builtin_bit_cast(Dest, source);
-}
-#else  // ABSL_HAVE_BUILTIN(__builtin_bit_cast)
-inline Dest bit_cast(const Source& source) {
-  Dest dest;
-  memcpy(static_cast<void*>(std::addressof(dest)),
-         static_cast<const void*>(std::addressof(source)), sizeof(dest));
-  return dest;
-}
-#endif  // ABSL_HAVE_BUILTIN(__builtin_bit_cast)
-
-#endif  // defined(__cpp_lib_bit_cast) && __cpp_lib_bit_cast >= 201806L
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_CASTS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/config.h b/projects/llm_framework/include/abseil-cpp/absl/base/config.h
deleted file mode 100644
index 5fa9f0ef..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/config.h
+++ /dev/null
@@ -1,1000 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: config.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines a set of macros for checking the presence of
-// important compiler and platform features. Such macros can be used to
-// produce portable code by parameterizing compilation based on the presence or
-// lack of a given feature.
-//
-// We define a "feature" as some interface we wish to program to: for example,
-// a library function or system call. A value of `1` indicates support for
-// that feature; any other value indicates the feature support is undefined.
-//
-// Example:
-//
-// Suppose a programmer wants to write a program that uses the 'mmap()' system
-// call. The Abseil macro for that feature (`ABSL_HAVE_MMAP`) allows you to
-// selectively include the `mmap.h` header and bracket code using that feature
-// in the macro:
-//
-//   #include "absl/base/config.h"
-//
-//   #ifdef ABSL_HAVE_MMAP
-//   #include "sys/mman.h"
-//   #endif  //ABSL_HAVE_MMAP
-//
-//   ...
-//   #ifdef ABSL_HAVE_MMAP
-//   void *ptr = mmap(...);
-//   ...
-//   #endif  // ABSL_HAVE_MMAP
-
-#ifndef ABSL_BASE_CONFIG_H_
-#define ABSL_BASE_CONFIG_H_
-
-// Included for the __GLIBC__ macro (or similar macros on other systems).
-#include <limits.h>
-
-#ifdef __cplusplus
-// Included for __GLIBCXX__, _LIBCPP_VERSION
-#include <cstddef>
-#endif  // __cplusplus
-
-// ABSL_INTERNAL_CPLUSPLUS_LANG
-//
-// MSVC does not set the value of __cplusplus correctly, but instead uses
-// _MSVC_LANG as a stand-in.
-// https://docs.microsoft.com/en-us/cpp/preprocessor/predefined-macros
-//
-// However, there are reports that MSVC even sets _MSVC_LANG incorrectly at
-// times, for example:
-// https://github.com/microsoft/vscode-cpptools/issues/1770
-// https://reviews.llvm.org/D70996
-//
-// For this reason, this symbol is considered INTERNAL and code outside of
-// Abseil must not use it.
-#if defined(_MSVC_LANG)
-#define ABSL_INTERNAL_CPLUSPLUS_LANG _MSVC_LANG
-#elif defined(__cplusplus)
-#define ABSL_INTERNAL_CPLUSPLUS_LANG __cplusplus
-#endif
-
-#if defined(ABSL_INTERNAL_CPLUSPLUS_LANG) && \
-    ABSL_INTERNAL_CPLUSPLUS_LANG >= 202002L
-// Include library feature test macros.
-#include <version>
-#endif
-
-#if defined(__APPLE__)
-// Included for TARGET_OS_IPHONE, __IPHONE_OS_VERSION_MIN_REQUIRED,
-// __IPHONE_8_0.
-#include <Availability.h>
-#include <TargetConditionals.h>
-#endif
-
-#include "absl/base/options.h"
-#include "absl/base/policy_checks.h"
-
-// Abseil long-term support (LTS) releases will define
-// `ABSL_LTS_RELEASE_VERSION` to the integer representing the date string of the
-// LTS release version, and will define `ABSL_LTS_RELEASE_PATCH_LEVEL` to the
-// integer representing the patch-level for that release.
-//
-// For example, for LTS release version "20300401.2", this would give us
-// ABSL_LTS_RELEASE_VERSION == 20300401 && ABSL_LTS_RELEASE_PATCH_LEVEL == 2
-//
-// These symbols will not be defined in non-LTS code.
-//
-// Abseil recommends that clients live-at-head. Therefore, if you are using
-// these symbols to assert a minimum version requirement, we recommend you do it
-// as
-//
-// #if defined(ABSL_LTS_RELEASE_VERSION) && ABSL_LTS_RELEASE_VERSION < 20300401
-// #error Project foo requires Abseil LTS version >= 20300401
-// #endif
-//
-// The `defined(ABSL_LTS_RELEASE_VERSION)` part of the check excludes
-// live-at-head clients from the minimum version assertion.
-//
-// See https://abseil.io/about/releases for more information on Abseil release
-// management.
-//
-// LTS releases can be obtained from
-// https://github.com/abseil/abseil-cpp/releases.
-#undef ABSL_LTS_RELEASE_VERSION
-#undef ABSL_LTS_RELEASE_PATCH_LEVEL
-
-// Helper macro to convert a CPP variable to a string literal.
-#define ABSL_INTERNAL_DO_TOKEN_STR(x) #x
-#define ABSL_INTERNAL_TOKEN_STR(x) ABSL_INTERNAL_DO_TOKEN_STR(x)
-
-// -----------------------------------------------------------------------------
-// Abseil namespace annotations
-// -----------------------------------------------------------------------------
-
-// ABSL_NAMESPACE_BEGIN/ABSL_NAMESPACE_END
-//
-// An annotation placed at the beginning/end of each `namespace absl` scope.
-// This is used to inject an inline namespace.
-//
-// The proper way to write Abseil code in the `absl` namespace is:
-//
-// namespace absl {
-// ABSL_NAMESPACE_BEGIN
-//
-// void Foo();  // absl::Foo().
-//
-// ABSL_NAMESPACE_END
-// }  // namespace absl
-//
-// Users of Abseil should not use these macros, because users of Abseil should
-// not write `namespace absl {` in their own code for any reason.  (Abseil does
-// not support forward declarations of its own types, nor does it support
-// user-provided specialization of Abseil templates.  Code that violates these
-// rules may be broken without warning.)
-#if !defined(ABSL_OPTION_USE_INLINE_NAMESPACE) || \
-    !defined(ABSL_OPTION_INLINE_NAMESPACE_NAME)
-#error options.h is misconfigured.
-#endif
-
-// Check that ABSL_OPTION_INLINE_NAMESPACE_NAME is neither "head" nor ""
-#if defined(__cplusplus) && ABSL_OPTION_USE_INLINE_NAMESPACE == 1
-
-#define ABSL_INTERNAL_INLINE_NAMESPACE_STR \
-  ABSL_INTERNAL_TOKEN_STR(ABSL_OPTION_INLINE_NAMESPACE_NAME)
-
-static_assert(ABSL_INTERNAL_INLINE_NAMESPACE_STR[0] != '\0',
-              "options.h misconfigured: ABSL_OPTION_INLINE_NAMESPACE_NAME must "
-              "not be empty.");
-static_assert(ABSL_INTERNAL_INLINE_NAMESPACE_STR[0] != 'h' ||
-                  ABSL_INTERNAL_INLINE_NAMESPACE_STR[1] != 'e' ||
-                  ABSL_INTERNAL_INLINE_NAMESPACE_STR[2] != 'a' ||
-                  ABSL_INTERNAL_INLINE_NAMESPACE_STR[3] != 'd' ||
-                  ABSL_INTERNAL_INLINE_NAMESPACE_STR[4] != '\0',
-              "options.h misconfigured: ABSL_OPTION_INLINE_NAMESPACE_NAME must "
-              "be changed to a new, unique identifier name.");
-
-#endif
-
-#if ABSL_OPTION_USE_INLINE_NAMESPACE == 0
-#define ABSL_NAMESPACE_BEGIN
-#define ABSL_NAMESPACE_END
-#define ABSL_INTERNAL_C_SYMBOL(x) x
-#elif ABSL_OPTION_USE_INLINE_NAMESPACE == 1
-#define ABSL_NAMESPACE_BEGIN \
-  inline namespace ABSL_OPTION_INLINE_NAMESPACE_NAME {
-#define ABSL_NAMESPACE_END }
-#define ABSL_INTERNAL_C_SYMBOL_HELPER_2(x, v) x##_##v
-#define ABSL_INTERNAL_C_SYMBOL_HELPER_1(x, v) \
-  ABSL_INTERNAL_C_SYMBOL_HELPER_2(x, v)
-#define ABSL_INTERNAL_C_SYMBOL(x) \
-  ABSL_INTERNAL_C_SYMBOL_HELPER_1(x, ABSL_OPTION_INLINE_NAMESPACE_NAME)
-#else
-#error options.h is misconfigured.
-#endif
-
-// -----------------------------------------------------------------------------
-// Compiler Feature Checks
-// -----------------------------------------------------------------------------
-
-// ABSL_HAVE_BUILTIN()
-//
-// Checks whether the compiler supports a Clang Feature Checking Macro, and if
-// so, checks whether it supports the provided builtin function "x" where x
-// is one of the functions noted in
-// https://clang.llvm.org/docs/LanguageExtensions.html
-//
-// Note: Use this macro to avoid an extra level of #ifdef __has_builtin check.
-// http://releases.llvm.org/3.3/tools/clang/docs/LanguageExtensions.html
-#ifdef __has_builtin
-#define ABSL_HAVE_BUILTIN(x) __has_builtin(x)
-#else
-#define ABSL_HAVE_BUILTIN(x) 0
-#endif
-
-#ifdef __has_feature
-#define ABSL_HAVE_FEATURE(f) __has_feature(f)
-#else
-#define ABSL_HAVE_FEATURE(f) 0
-#endif
-
-// Portable check for GCC minimum version:
-// https://gcc.gnu.org/onlinedocs/cpp/Common-Predefined-Macros.html
-#if defined(__GNUC__) && defined(__GNUC_MINOR__)
-#define ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(x, y) \
-  (__GNUC__ > (x) || __GNUC__ == (x) && __GNUC_MINOR__ >= (y))
-#else
-#define ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(x, y) 0
-#endif
-
-#if defined(__clang__) && defined(__clang_major__) && defined(__clang_minor__)
-#define ABSL_INTERNAL_HAVE_MIN_CLANG_VERSION(x, y) \
-  (__clang_major__ > (x) || __clang_major__ == (x) && __clang_minor__ >= (y))
-#else
-#define ABSL_INTERNAL_HAVE_MIN_CLANG_VERSION(x, y) 0
-#endif
-
-// ABSL_HAVE_TLS is defined to 1 when __thread should be supported.
-// We assume __thread is supported on Linux or Asylo when compiled with Clang or
-// compiled against libstdc++ with _GLIBCXX_HAVE_TLS defined.
-#ifdef ABSL_HAVE_TLS
-#error ABSL_HAVE_TLS cannot be directly set
-#elif (defined(__linux__) || defined(__ASYLO__)) && \
-    (defined(__clang__) || defined(_GLIBCXX_HAVE_TLS))
-#define ABSL_HAVE_TLS 1
-#endif
-
-// ABSL_HAVE_STD_IS_TRIVIALLY_DESTRUCTIBLE
-//
-// Checks whether `std::is_trivially_destructible<T>` is supported.
-#ifdef ABSL_HAVE_STD_IS_TRIVIALLY_DESTRUCTIBLE
-#error ABSL_HAVE_STD_IS_TRIVIALLY_DESTRUCTIBLE cannot be directly set
-#define ABSL_HAVE_STD_IS_TRIVIALLY_DESTRUCTIBLE 1
-#endif
-
-// ABSL_HAVE_STD_IS_TRIVIALLY_CONSTRUCTIBLE
-//
-// Checks whether `std::is_trivially_default_constructible<T>` and
-// `std::is_trivially_copy_constructible<T>` are supported.
-#ifdef ABSL_HAVE_STD_IS_TRIVIALLY_CONSTRUCTIBLE
-#error ABSL_HAVE_STD_IS_TRIVIALLY_CONSTRUCTIBLE cannot be directly set
-#else
-#define ABSL_HAVE_STD_IS_TRIVIALLY_CONSTRUCTIBLE 1
-#endif
-
-// ABSL_HAVE_STD_IS_TRIVIALLY_ASSIGNABLE
-//
-// Checks whether `std::is_trivially_copy_assignable<T>` is supported.
-#ifdef ABSL_HAVE_STD_IS_TRIVIALLY_ASSIGNABLE
-#error ABSL_HAVE_STD_IS_TRIVIALLY_ASSIGNABLE cannot be directly set
-#else
-#define ABSL_HAVE_STD_IS_TRIVIALLY_ASSIGNABLE 1
-#endif
-
-// ABSL_HAVE_STD_IS_TRIVIALLY_COPYABLE
-//
-// Checks whether `std::is_trivially_copyable<T>` is supported.
-#ifdef ABSL_HAVE_STD_IS_TRIVIALLY_COPYABLE
-#error ABSL_HAVE_STD_IS_TRIVIALLY_COPYABLE cannot be directly set
-#define ABSL_HAVE_STD_IS_TRIVIALLY_COPYABLE 1
-#endif
-
-// ABSL_HAVE_THREAD_LOCAL
-//
-// Checks whether C++11's `thread_local` storage duration specifier is
-// supported.
-#ifdef ABSL_HAVE_THREAD_LOCAL
-#error ABSL_HAVE_THREAD_LOCAL cannot be directly set
-#elif defined(__APPLE__)
-// Notes:
-// * Xcode's clang did not support `thread_local` until version 8, and
-//   even then not for all iOS < 9.0.
-// * Xcode 9.3 started disallowing `thread_local` for 32-bit iOS simulator
-//   targeting iOS 9.x.
-// * Xcode 10 moves the deployment target check for iOS < 9.0 to link time
-//   making ABSL_HAVE_FEATURE unreliable there.
-//
-#if ABSL_HAVE_FEATURE(cxx_thread_local) && \
-    !(TARGET_OS_IPHONE && __IPHONE_OS_VERSION_MIN_REQUIRED < __IPHONE_9_0)
-#define ABSL_HAVE_THREAD_LOCAL 1
-#endif
-#else  // !defined(__APPLE__)
-#define ABSL_HAVE_THREAD_LOCAL 1
-#endif
-
-// There are platforms for which TLS should not be used even though the compiler
-// makes it seem like it's supported (Android NDK < r12b for example).
-// This is primarily because of linker problems and toolchain misconfiguration:
-// Abseil does not intend to support this indefinitely. Currently, the newest
-// toolchain that we intend to support that requires this behavior is the
-// r11 NDK - allowing for a 5 year support window on that means this option
-// is likely to be removed around June of 2021.
-// TLS isn't supported until NDK r12b per
-// https://developer.android.com/ndk/downloads/revision_history.html
-// Since NDK r16, `__NDK_MAJOR__` and `__NDK_MINOR__` are defined in
-// <android/ndk-version.h>. For NDK < r16, users should define these macros,
-// e.g. `-D__NDK_MAJOR__=11 -D__NKD_MINOR__=0` for NDK r11.
-#if defined(__ANDROID__) && defined(__clang__)
-#if __has_include(<android/ndk-version.h>)
-#include <android/ndk-version.h>
-#endif  // __has_include(<android/ndk-version.h>)
-#if defined(__ANDROID__) && defined(__clang__) && defined(__NDK_MAJOR__) && \
-    defined(__NDK_MINOR__) &&                                               \
-    ((__NDK_MAJOR__ < 12) || ((__NDK_MAJOR__ == 12) && (__NDK_MINOR__ < 1)))
-#undef ABSL_HAVE_TLS
-#undef ABSL_HAVE_THREAD_LOCAL
-#endif
-#endif  // defined(__ANDROID__) && defined(__clang__)
-
-// ABSL_HAVE_INTRINSIC_INT128
-//
-// Checks whether the __int128 compiler extension for a 128-bit integral type is
-// supported.
-//
-// Note: __SIZEOF_INT128__ is defined by Clang and GCC when __int128 is
-// supported, but we avoid using it in certain cases:
-// * On Clang:
-//   * Building using Clang for Windows, where the Clang runtime library has
-//     128-bit support only on LP64 architectures, but Windows is LLP64.
-// * On Nvidia's nvcc:
-//   * nvcc also defines __GNUC__ and __SIZEOF_INT128__, but not all versions
-//     actually support __int128.
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-#error ABSL_HAVE_INTRINSIC_INT128 cannot be directly set
-#elif defined(__SIZEOF_INT128__)
-#if (defined(__clang__) && !defined(_WIN32)) ||           \
-    (defined(__CUDACC__) && __CUDACC_VER_MAJOR__ >= 9) || \
-    (defined(__GNUC__) && !defined(__clang__) && !defined(__CUDACC__))
-#define ABSL_HAVE_INTRINSIC_INT128 1
-#elif defined(__CUDACC__)
-// __CUDACC_VER__ is a full version number before CUDA 9, and is defined to a
-// string explaining that it has been removed starting with CUDA 9. We use
-// nested #ifs because there is no short-circuiting in the preprocessor.
-// NOTE: `__CUDACC__` could be undefined while `__CUDACC_VER__` is defined.
-#if __CUDACC_VER__ >= 70000
-#define ABSL_HAVE_INTRINSIC_INT128 1
-#endif  // __CUDACC_VER__ >= 70000
-#endif  // defined(__CUDACC__)
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-
-// ABSL_HAVE_EXCEPTIONS
-//
-// Checks whether the compiler both supports and enables exceptions. Many
-// compilers support a "no exceptions" mode that disables exceptions.
-//
-// Generally, when ABSL_HAVE_EXCEPTIONS is not defined:
-//
-// * Code using `throw` and `try` may not compile.
-// * The `noexcept` specifier will still compile and behave as normal.
-// * The `noexcept` operator may still return `false`.
-//
-// For further details, consult the compiler's documentation.
-#ifdef ABSL_HAVE_EXCEPTIONS
-#error ABSL_HAVE_EXCEPTIONS cannot be directly set.
-#elif ABSL_INTERNAL_HAVE_MIN_CLANG_VERSION(3, 6)
-// Clang >= 3.6
-#if ABSL_HAVE_FEATURE(cxx_exceptions)
-#define ABSL_HAVE_EXCEPTIONS 1
-#endif  // ABSL_HAVE_FEATURE(cxx_exceptions)
-#elif defined(__clang__)
-// Clang < 3.6
-// http://releases.llvm.org/3.6.0/tools/clang/docs/ReleaseNotes.html#the-exceptions-macro
-#if defined(__EXCEPTIONS) && ABSL_HAVE_FEATURE(cxx_exceptions)
-#define ABSL_HAVE_EXCEPTIONS 1
-#endif  // defined(__EXCEPTIONS) && ABSL_HAVE_FEATURE(cxx_exceptions)
-// Handle remaining special cases and default to exceptions being supported.
-#elif !(defined(__GNUC__) && !defined(__cpp_exceptions)) && \
-    !(defined(_MSC_VER) && !defined(_CPPUNWIND))
-#define ABSL_HAVE_EXCEPTIONS 1
-#endif
-
-// -----------------------------------------------------------------------------
-// Platform Feature Checks
-// -----------------------------------------------------------------------------
-
-// Currently supported operating systems and associated preprocessor
-// symbols:
-//
-//   Linux and Linux-derived           __linux__
-//   Android                           __ANDROID__ (implies __linux__)
-//   Linux (non-Android)               __linux__ && !__ANDROID__
-//   Darwin (macOS and iOS)            __APPLE__
-//   Akaros (http://akaros.org)        __ros__
-//   Windows                           _WIN32
-//   NaCL                              __native_client__
-//   AsmJS                             __asmjs__
-//   WebAssembly (Emscripten)          __EMSCRIPTEN__
-//   Fuchsia                           __Fuchsia__
-//
-// Note that since Android defines both __ANDROID__ and __linux__, one
-// may probe for either Linux or Android by simply testing for __linux__.
-
-// ABSL_HAVE_MMAP
-//
-// Checks whether the platform has an mmap(2) implementation as defined in
-// POSIX.1-2001.
-#ifdef ABSL_HAVE_MMAP
-#error ABSL_HAVE_MMAP cannot be directly set
-#elif defined(__linux__) || defined(__APPLE__) || defined(__FreeBSD__) ||    \
-    defined(_AIX) || defined(__ros__) || defined(__native_client__) ||       \
-    defined(__asmjs__) || defined(__EMSCRIPTEN__) || defined(__Fuchsia__) || \
-    defined(__sun) || defined(__ASYLO__) || defined(__myriad2__) ||          \
-    defined(__HAIKU__) || defined(__OpenBSD__) || defined(__NetBSD__) ||     \
-    defined(__QNX__) || defined(__VXWORKS__) || defined(__hexagon__)
-#define ABSL_HAVE_MMAP 1
-#endif
-
-// ABSL_HAVE_PTHREAD_GETSCHEDPARAM
-//
-// Checks whether the platform implements the pthread_(get|set)schedparam(3)
-// functions as defined in POSIX.1-2001.
-#ifdef ABSL_HAVE_PTHREAD_GETSCHEDPARAM
-#error ABSL_HAVE_PTHREAD_GETSCHEDPARAM cannot be directly set
-#elif defined(__linux__) || defined(__APPLE__) || defined(__FreeBSD__) || \
-    defined(_AIX) || defined(__ros__) || defined(__OpenBSD__) ||          \
-    defined(__NetBSD__) || defined(__VXWORKS__)
-#define ABSL_HAVE_PTHREAD_GETSCHEDPARAM 1
-#endif
-
-// ABSL_HAVE_SCHED_GETCPU
-//
-// Checks whether sched_getcpu is available.
-#ifdef ABSL_HAVE_SCHED_GETCPU
-#error ABSL_HAVE_SCHED_GETCPU cannot be directly set
-#elif defined(__linux__)
-#define ABSL_HAVE_SCHED_GETCPU 1
-#endif
-
-// ABSL_HAVE_SCHED_YIELD
-//
-// Checks whether the platform implements sched_yield(2) as defined in
-// POSIX.1-2001.
-#ifdef ABSL_HAVE_SCHED_YIELD
-#error ABSL_HAVE_SCHED_YIELD cannot be directly set
-#elif defined(__linux__) || defined(__ros__) || defined(__native_client__) || \
-    defined(__VXWORKS__)
-#define ABSL_HAVE_SCHED_YIELD 1
-#endif
-
-// ABSL_HAVE_SEMAPHORE_H
-//
-// Checks whether the platform supports the <semaphore.h> header and sem_init(3)
-// family of functions as standardized in POSIX.1-2001.
-//
-// Note: While Apple provides <semaphore.h> for both iOS and macOS, it is
-// explicitly deprecated and will cause build failures if enabled for those
-// platforms.  We side-step the issue by not defining it here for Apple
-// platforms.
-#ifdef ABSL_HAVE_SEMAPHORE_H
-#error ABSL_HAVE_SEMAPHORE_H cannot be directly set
-#elif defined(__linux__) || defined(__ros__) || defined(__VXWORKS__)
-#define ABSL_HAVE_SEMAPHORE_H 1
-#endif
-
-// ABSL_HAVE_ALARM
-//
-// Checks whether the platform supports the <signal.h> header and alarm(2)
-// function as standardized in POSIX.1-2001.
-#ifdef ABSL_HAVE_ALARM
-#error ABSL_HAVE_ALARM cannot be directly set
-#elif defined(__GOOGLE_GRTE_VERSION__)
-// feature tests for Google's GRTE
-#define ABSL_HAVE_ALARM 1
-#elif defined(__GLIBC__)
-// feature test for glibc
-#define ABSL_HAVE_ALARM 1
-#elif defined(_MSC_VER)
-// feature tests for Microsoft's library
-#elif defined(__MINGW32__)
-// mingw32 doesn't provide alarm(2):
-// https://osdn.net/projects/mingw/scm/git/mingw-org-wsl/blobs/5.2-trunk/mingwrt/include/unistd.h
-// mingw-w64 provides a no-op implementation:
-// https://sourceforge.net/p/mingw-w64/mingw-w64/ci/master/tree/mingw-w64-crt/misc/alarm.c
-#elif defined(__EMSCRIPTEN__)
-// emscripten doesn't support signals
-#elif defined(__wasi__)
-// WASI doesn't support signals
-#elif defined(__Fuchsia__)
-// Signals don't exist on fuchsia.
-#elif defined(__native_client__)
-// Signals don't exist on hexagon/QuRT
-#elif defined(__hexagon__)
-#else
-// other standard libraries
-#define ABSL_HAVE_ALARM 1
-#endif
-
-// ABSL_IS_LITTLE_ENDIAN
-// ABSL_IS_BIG_ENDIAN
-//
-// Checks the endianness of the platform.
-//
-// Notes: uses the built in endian macros provided by GCC (since 4.6) and
-// Clang (since 3.2); see
-// https://gcc.gnu.org/onlinedocs/cpp/Common-Predefined-Macros.html.
-// Otherwise, if _WIN32, assume little endian. Otherwise, bail with an error.
-#if defined(ABSL_IS_BIG_ENDIAN)
-#error "ABSL_IS_BIG_ENDIAN cannot be directly set."
-#endif
-#if defined(ABSL_IS_LITTLE_ENDIAN)
-#error "ABSL_IS_LITTLE_ENDIAN cannot be directly set."
-#endif
-
-#if (defined(__BYTE_ORDER__) && defined(__ORDER_LITTLE_ENDIAN__) && \
-     __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
-#define ABSL_IS_LITTLE_ENDIAN 1
-#elif defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && \
-    __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
-#define ABSL_IS_BIG_ENDIAN 1
-#elif defined(_WIN32)
-#define ABSL_IS_LITTLE_ENDIAN 1
-#else
-#error "absl endian detection needs to be set up for your compiler"
-#endif
-
-// macOS < 10.13 and iOS < 12 don't support <any>, <optional>, or <variant>
-// because the libc++ shared library shipped on the system doesn't have the
-// requisite exported symbols.  See
-// https://github.com/abseil/abseil-cpp/issues/207 and
-// https://developer.apple.com/documentation/xcode_release_notes/xcode_10_release_notes
-//
-// libc++ spells out the availability requirements in the file
-// llvm-project/libcxx/include/__config via the #define
-// _LIBCPP_AVAILABILITY_BAD_OPTIONAL_ACCESS. The set of versions has been
-// modified a few times, via
-// https://github.com/llvm/llvm-project/commit/7fb40e1569dd66292b647f4501b85517e9247953
-// and
-// https://github.com/llvm/llvm-project/commit/0bc451e7e137c4ccadcd3377250874f641ca514a
-// The second has the actually correct versions, thus, is what we copy here.
-#if defined(__APPLE__) &&                                         \
-    ((defined(__ENVIRONMENT_MAC_OS_X_VERSION_MIN_REQUIRED__) &&   \
-      __ENVIRONMENT_MAC_OS_X_VERSION_MIN_REQUIRED__ < 101300) ||  \
-     (defined(__ENVIRONMENT_IPHONE_OS_VERSION_MIN_REQUIRED__) &&  \
-      __ENVIRONMENT_IPHONE_OS_VERSION_MIN_REQUIRED__ < 120000) || \
-     (defined(__ENVIRONMENT_WATCH_OS_VERSION_MIN_REQUIRED__) &&   \
-      __ENVIRONMENT_WATCH_OS_VERSION_MIN_REQUIRED__ < 50000) ||   \
-     (defined(__ENVIRONMENT_TV_OS_VERSION_MIN_REQUIRED__) &&      \
-      __ENVIRONMENT_TV_OS_VERSION_MIN_REQUIRED__ < 120000))
-#define ABSL_INTERNAL_APPLE_CXX17_TYPES_UNAVAILABLE 1
-#else
-#define ABSL_INTERNAL_APPLE_CXX17_TYPES_UNAVAILABLE 0
-#endif
-
-// ABSL_HAVE_STD_ANY
-//
-// Checks whether C++17 std::any is available.
-#ifdef ABSL_HAVE_STD_ANY
-#error "ABSL_HAVE_STD_ANY cannot be directly set."
-#elif defined(__cpp_lib_any) && __cpp_lib_any >= 201606L
-#define ABSL_HAVE_STD_ANY 1
-#elif defined(ABSL_INTERNAL_CPLUSPLUS_LANG) && \
-    ABSL_INTERNAL_CPLUSPLUS_LANG >= 201703L && \
-    !ABSL_INTERNAL_APPLE_CXX17_TYPES_UNAVAILABLE
-#define ABSL_HAVE_STD_ANY 1
-#endif
-
-// ABSL_HAVE_STD_OPTIONAL
-//
-// Checks whether C++17 std::optional is available.
-#ifdef ABSL_HAVE_STD_OPTIONAL
-#error "ABSL_HAVE_STD_OPTIONAL cannot be directly set."
-#elif defined(__cpp_lib_optional) && __cpp_lib_optional >= 202106L
-#define ABSL_HAVE_STD_OPTIONAL 1
-#elif defined(ABSL_INTERNAL_CPLUSPLUS_LANG) && \
-    ABSL_INTERNAL_CPLUSPLUS_LANG >= 201703L && \
-    !ABSL_INTERNAL_APPLE_CXX17_TYPES_UNAVAILABLE
-#define ABSL_HAVE_STD_OPTIONAL 1
-#endif
-
-// ABSL_HAVE_STD_VARIANT
-//
-// Checks whether C++17 std::variant is available.
-#ifdef ABSL_HAVE_STD_VARIANT
-#error "ABSL_HAVE_STD_VARIANT cannot be directly set."
-#elif defined(__cpp_lib_variant) && __cpp_lib_variant >= 201606L
-#define ABSL_HAVE_STD_VARIANT 1
-#elif defined(ABSL_INTERNAL_CPLUSPLUS_LANG) && \
-    ABSL_INTERNAL_CPLUSPLUS_LANG >= 201703L && \
-    !ABSL_INTERNAL_APPLE_CXX17_TYPES_UNAVAILABLE
-#define ABSL_HAVE_STD_VARIANT 1
-#endif
-
-// ABSL_HAVE_STD_STRING_VIEW
-//
-// Checks whether C++17 std::string_view is available.
-#ifdef ABSL_HAVE_STD_STRING_VIEW
-#error "ABSL_HAVE_STD_STRING_VIEW cannot be directly set."
-#elif defined(__cpp_lib_string_view) && __cpp_lib_string_view >= 201606L
-#define ABSL_HAVE_STD_STRING_VIEW 1
-#elif defined(ABSL_INTERNAL_CPLUSPLUS_LANG) && \
-    ABSL_INTERNAL_CPLUSPLUS_LANG >= 201703L
-#define ABSL_HAVE_STD_STRING_VIEW 1
-#endif
-
-// ABSL_HAVE_STD_ORDERING
-//
-// Checks whether C++20 std::{partial,weak,strong}_ordering are available.
-//
-// __cpp_lib_three_way_comparison is missing on libc++
-// (https://github.com/llvm/llvm-project/issues/73953) so treat it as defined
-// when building in C++20 mode.
-#ifdef ABSL_HAVE_STD_ORDERING
-#error "ABSL_HAVE_STD_ORDERING cannot be directly set."
-#elif (defined(__cpp_lib_three_way_comparison) &&    \
-       __cpp_lib_three_way_comparison >= 201907L) || \
-    (defined(ABSL_INTERNAL_CPLUSPLUS_LANG) &&        \
-     ABSL_INTERNAL_CPLUSPLUS_LANG >= 202002L)
-#define ABSL_HAVE_STD_ORDERING 1
-#endif
-
-// ABSL_USES_STD_ANY
-//
-// Indicates whether absl::any is an alias for std::any.
-#if !defined(ABSL_OPTION_USE_STD_ANY)
-#error options.h is misconfigured.
-#elif ABSL_OPTION_USE_STD_ANY == 0 || \
-    (ABSL_OPTION_USE_STD_ANY == 2 && !defined(ABSL_HAVE_STD_ANY))
-#undef ABSL_USES_STD_ANY
-#elif ABSL_OPTION_USE_STD_ANY == 1 || \
-    (ABSL_OPTION_USE_STD_ANY == 2 && defined(ABSL_HAVE_STD_ANY))
-#define ABSL_USES_STD_ANY 1
-#else
-#error options.h is misconfigured.
-#endif
-
-// ABSL_USES_STD_OPTIONAL
-//
-// Indicates whether absl::optional is an alias for std::optional.
-#if !defined(ABSL_OPTION_USE_STD_OPTIONAL)
-#error options.h is misconfigured.
-#elif ABSL_OPTION_USE_STD_OPTIONAL == 0 || \
-    (ABSL_OPTION_USE_STD_OPTIONAL == 2 && !defined(ABSL_HAVE_STD_OPTIONAL))
-#undef ABSL_USES_STD_OPTIONAL
-#elif ABSL_OPTION_USE_STD_OPTIONAL == 1 || \
-    (ABSL_OPTION_USE_STD_OPTIONAL == 2 && defined(ABSL_HAVE_STD_OPTIONAL))
-#define ABSL_USES_STD_OPTIONAL 1
-#else
-#error options.h is misconfigured.
-#endif
-
-// ABSL_USES_STD_VARIANT
-//
-// Indicates whether absl::variant is an alias for std::variant.
-#if !defined(ABSL_OPTION_USE_STD_VARIANT)
-#error options.h is misconfigured.
-#elif ABSL_OPTION_USE_STD_VARIANT == 0 || \
-    (ABSL_OPTION_USE_STD_VARIANT == 2 && !defined(ABSL_HAVE_STD_VARIANT))
-#undef ABSL_USES_STD_VARIANT
-#elif ABSL_OPTION_USE_STD_VARIANT == 1 || \
-    (ABSL_OPTION_USE_STD_VARIANT == 2 && defined(ABSL_HAVE_STD_VARIANT))
-#define ABSL_USES_STD_VARIANT 1
-#else
-#error options.h is misconfigured.
-#endif
-
-// ABSL_USES_STD_STRING_VIEW
-//
-// Indicates whether absl::string_view is an alias for std::string_view.
-#if !defined(ABSL_OPTION_USE_STD_STRING_VIEW)
-#error options.h is misconfigured.
-#elif ABSL_OPTION_USE_STD_STRING_VIEW == 0 || \
-    (ABSL_OPTION_USE_STD_STRING_VIEW == 2 &&  \
-     !defined(ABSL_HAVE_STD_STRING_VIEW))
-#undef ABSL_USES_STD_STRING_VIEW
-#elif ABSL_OPTION_USE_STD_STRING_VIEW == 1 || \
-    (ABSL_OPTION_USE_STD_STRING_VIEW == 2 &&  \
-     defined(ABSL_HAVE_STD_STRING_VIEW))
-#define ABSL_USES_STD_STRING_VIEW 1
-#else
-#error options.h is misconfigured.
-#endif
-
-// ABSL_USES_STD_ORDERING
-//
-// Indicates whether absl::{partial,weak,strong}_ordering are aliases for the
-// std:: ordering types.
-#if !defined(ABSL_OPTION_USE_STD_ORDERING)
-#error options.h is misconfigured.
-#elif ABSL_OPTION_USE_STD_ORDERING == 0 || \
-    (ABSL_OPTION_USE_STD_ORDERING == 2 && !defined(ABSL_HAVE_STD_ORDERING))
-#undef ABSL_USES_STD_ORDERING
-#elif ABSL_OPTION_USE_STD_ORDERING == 1 || \
-    (ABSL_OPTION_USE_STD_ORDERING == 2 && defined(ABSL_HAVE_STD_ORDERING))
-#define ABSL_USES_STD_ORDERING 1
-#else
-#error options.h is misconfigured.
-#endif
-
-// In debug mode, MSVC 2017's std::variant throws a EXCEPTION_ACCESS_VIOLATION
-// SEH exception from emplace for variant<SomeStruct> when constructing the
-// struct can throw. This defeats some of variant_test and
-// variant_exception_safety_test.
-#if defined(_MSC_VER) && _MSC_VER >= 1700 && defined(_DEBUG)
-#define ABSL_INTERNAL_MSVC_2017_DBG_MODE
-#endif
-
-// ABSL_INTERNAL_MANGLED_NS
-// ABSL_INTERNAL_MANGLED_BACKREFERENCE
-//
-// Internal macros for building up mangled names in our internal fork of CCTZ.
-// This implementation detail is only needed and provided for the MSVC build.
-//
-// These macros both expand to string literals.  ABSL_INTERNAL_MANGLED_NS is
-// the mangled spelling of the `absl` namespace, and
-// ABSL_INTERNAL_MANGLED_BACKREFERENCE is a back-reference integer representing
-// the proper count to skip past the CCTZ fork namespace names.  (This number
-// is one larger when there is an inline namespace name to skip.)
-#if defined(_MSC_VER)
-#if ABSL_OPTION_USE_INLINE_NAMESPACE == 0
-#define ABSL_INTERNAL_MANGLED_NS "absl"
-#define ABSL_INTERNAL_MANGLED_BACKREFERENCE "5"
-#else
-#define ABSL_INTERNAL_MANGLED_NS \
-  ABSL_INTERNAL_TOKEN_STR(ABSL_OPTION_INLINE_NAMESPACE_NAME) "@absl"
-#define ABSL_INTERNAL_MANGLED_BACKREFERENCE "6"
-#endif
-#endif
-
-// ABSL_DLL
-//
-// When building Abseil as a DLL, this macro expands to `__declspec(dllexport)`
-// so we can annotate symbols appropriately as being exported. When used in
-// headers consuming a DLL, this macro expands to `__declspec(dllimport)` so
-// that consumers know the symbol is defined inside the DLL. In all other cases,
-// the macro expands to nothing.
-#if defined(_MSC_VER)
-#if defined(ABSL_BUILD_DLL)
-#define ABSL_DLL __declspec(dllexport)
-#elif defined(ABSL_CONSUME_DLL)
-#define ABSL_DLL __declspec(dllimport)
-#else
-#define ABSL_DLL
-#endif
-#else
-#define ABSL_DLL
-#endif  // defined(_MSC_VER)
-
-#if defined(_MSC_VER)
-#if defined(ABSL_BUILD_TEST_DLL)
-#define ABSL_TEST_DLL __declspec(dllexport)
-#elif defined(ABSL_CONSUME_TEST_DLL)
-#define ABSL_TEST_DLL __declspec(dllimport)
-#else
-#define ABSL_TEST_DLL
-#endif
-#else
-#define ABSL_TEST_DLL
-#endif  // defined(_MSC_VER)
-
-// ABSL_HAVE_MEMORY_SANITIZER
-//
-// MemorySanitizer (MSan) is a detector of uninitialized reads. It consists of
-// a compiler instrumentation module and a run-time library.
-#ifdef ABSL_HAVE_MEMORY_SANITIZER
-#error "ABSL_HAVE_MEMORY_SANITIZER cannot be directly set."
-#elif !defined(__native_client__) && ABSL_HAVE_FEATURE(memory_sanitizer)
-#define ABSL_HAVE_MEMORY_SANITIZER 1
-#endif
-
-// ABSL_HAVE_THREAD_SANITIZER
-//
-// ThreadSanitizer (TSan) is a fast data race detector.
-#ifdef ABSL_HAVE_THREAD_SANITIZER
-#error "ABSL_HAVE_THREAD_SANITIZER cannot be directly set."
-#elif defined(__SANITIZE_THREAD__)
-#define ABSL_HAVE_THREAD_SANITIZER 1
-#elif ABSL_HAVE_FEATURE(thread_sanitizer)
-#define ABSL_HAVE_THREAD_SANITIZER 1
-#endif
-
-// ABSL_HAVE_ADDRESS_SANITIZER
-//
-// AddressSanitizer (ASan) is a fast memory error detector.
-#ifdef ABSL_HAVE_ADDRESS_SANITIZER
-#error "ABSL_HAVE_ADDRESS_SANITIZER cannot be directly set."
-#elif defined(__SANITIZE_ADDRESS__)
-#define ABSL_HAVE_ADDRESS_SANITIZER 1
-#elif ABSL_HAVE_FEATURE(address_sanitizer)
-#define ABSL_HAVE_ADDRESS_SANITIZER 1
-#endif
-
-// ABSL_HAVE_HWADDRESS_SANITIZER
-//
-// Hardware-Assisted AddressSanitizer (or HWASAN) is even faster than asan
-// memory error detector which can use CPU features like ARM TBI, Intel LAM or
-// AMD UAI.
-#ifdef ABSL_HAVE_HWADDRESS_SANITIZER
-#error "ABSL_HAVE_HWADDRESS_SANITIZER cannot be directly set."
-#elif defined(__SANITIZE_HWADDRESS__)
-#define ABSL_HAVE_HWADDRESS_SANITIZER 1
-#elif ABSL_HAVE_FEATURE(hwaddress_sanitizer)
-#define ABSL_HAVE_HWADDRESS_SANITIZER 1
-#endif
-
-// ABSL_HAVE_DATAFLOW_SANITIZER
-//
-// Dataflow Sanitizer (or DFSAN) is a generalised dynamic data flow analysis.
-#ifdef ABSL_HAVE_DATAFLOW_SANITIZER
-#error "ABSL_HAVE_DATAFLOW_SANITIZER cannot be directly set."
-#elif defined(DATAFLOW_SANITIZER)
-// GCC provides no method for detecting the presence of the standalone
-// DataFlowSanitizer (-fsanitize=dataflow), so GCC users of -fsanitize=dataflow
-// should also use -DDATAFLOW_SANITIZER.
-#define ABSL_HAVE_DATAFLOW_SANITIZER 1
-#elif ABSL_HAVE_FEATURE(dataflow_sanitizer)
-#define ABSL_HAVE_DATAFLOW_SANITIZER 1
-#endif
-
-// ABSL_HAVE_LEAK_SANITIZER
-//
-// LeakSanitizer (or lsan) is a detector of memory leaks.
-// https://clang.llvm.org/docs/LeakSanitizer.html
-// https://github.com/google/sanitizers/wiki/AddressSanitizerLeakSanitizer
-//
-// The macro ABSL_HAVE_LEAK_SANITIZER can be used to detect at compile-time
-// whether the LeakSanitizer is potentially available. However, just because the
-// LeakSanitizer is available does not mean it is active. Use the
-// always-available run-time interface in //absl/debugging/leak_check.h for
-// interacting with LeakSanitizer.
-#ifdef ABSL_HAVE_LEAK_SANITIZER
-#error "ABSL_HAVE_LEAK_SANITIZER cannot be directly set."
-#elif defined(LEAK_SANITIZER)
-// GCC provides no method for detecting the presence of the standalone
-// LeakSanitizer (-fsanitize=leak), so GCC users of -fsanitize=leak should also
-// use -DLEAK_SANITIZER.
-#define ABSL_HAVE_LEAK_SANITIZER 1
-// Clang standalone LeakSanitizer (-fsanitize=leak)
-#elif ABSL_HAVE_FEATURE(leak_sanitizer)
-#define ABSL_HAVE_LEAK_SANITIZER 1
-#elif defined(ABSL_HAVE_ADDRESS_SANITIZER)
-// GCC or Clang using the LeakSanitizer integrated into AddressSanitizer.
-#define ABSL_HAVE_LEAK_SANITIZER 1
-#endif
-
-// ABSL_HAVE_CLASS_TEMPLATE_ARGUMENT_DEDUCTION
-//
-// Class template argument deduction is a language feature added in C++17.
-#ifdef ABSL_HAVE_CLASS_TEMPLATE_ARGUMENT_DEDUCTION
-#error "ABSL_HAVE_CLASS_TEMPLATE_ARGUMENT_DEDUCTION cannot be directly set."
-#elif defined(__cpp_deduction_guides)
-#define ABSL_HAVE_CLASS_TEMPLATE_ARGUMENT_DEDUCTION 1
-#endif
-
-// ABSL_INTERNAL_NEED_REDUNDANT_CONSTEXPR_DECL
-//
-// Prior to C++17, static constexpr variables defined in classes required a
-// separate definition outside of the class body, for example:
-//
-// class Foo {
-//   static constexpr int kBar = 0;
-// };
-// constexpr int Foo::kBar;
-//
-// In C++17, these variables defined in classes are considered inline variables,
-// and the extra declaration is redundant. Since some compilers warn on the
-// extra declarations, ABSL_INTERNAL_NEED_REDUNDANT_CONSTEXPR_DECL can be used
-// conditionally ignore them:
-//
-// #ifdef ABSL_INTERNAL_NEED_REDUNDANT_CONSTEXPR_DECL
-// constexpr int Foo::kBar;
-// #endif
-#if defined(ABSL_INTERNAL_CPLUSPLUS_LANG) && \
-    ABSL_INTERNAL_CPLUSPLUS_LANG < 201703L
-#define ABSL_INTERNAL_NEED_REDUNDANT_CONSTEXPR_DECL 1
-#endif
-
-// `ABSL_INTERNAL_HAS_RTTI` determines whether abseil is being compiled with
-// RTTI support.
-#ifdef ABSL_INTERNAL_HAS_RTTI
-#error ABSL_INTERNAL_HAS_RTTI cannot be directly set
-#elif ABSL_HAVE_FEATURE(cxx_rtti)
-#define ABSL_INTERNAL_HAS_RTTI 1
-#elif defined(__GNUC__) && defined(__GXX_RTTI)
-#define ABSL_INTERNAL_HAS_RTTI 1
-#elif defined(_MSC_VER) && defined(_CPPRTTI)
-#define ABSL_INTERNAL_HAS_RTTI 1
-#elif !defined(__GNUC__) && !defined(_MSC_VER)
-// Unknown compiler, default to RTTI
-#define ABSL_INTERNAL_HAS_RTTI 1
-#endif
-
-// `ABSL_INTERNAL_HAS_CXA_DEMANGLE` determines whether `abi::__cxa_demangle` is
-// available.
-#ifdef ABSL_INTERNAL_HAS_CXA_DEMANGLE
-#error ABSL_INTERNAL_HAS_CXA_DEMANGLE cannot be directly set
-#elif defined(OS_ANDROID) && (defined(__i386__) || defined(__x86_64__))
-#define ABSL_INTERNAL_HAS_CXA_DEMANGLE 0
-#elif defined(__GNUC__) && !defined(__mips__)
-#define ABSL_INTERNAL_HAS_CXA_DEMANGLE 1
-#elif defined(__clang__) && !defined(_MSC_VER)
-#define ABSL_INTERNAL_HAS_CXA_DEMANGLE 1
-#endif
-
-// ABSL_INTERNAL_HAVE_SSE is used for compile-time detection of SSE support.
-// See https://gcc.gnu.org/onlinedocs/gcc/x86-Options.html for an overview of
-// which architectures support the various x86 instruction sets.
-#ifdef ABSL_INTERNAL_HAVE_SSE
-#error ABSL_INTERNAL_HAVE_SSE cannot be directly set
-#elif defined(__SSE__)
-#define ABSL_INTERNAL_HAVE_SSE 1
-#elif (defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)) && \
-    !defined(_M_ARM64EC)
-// MSVC only defines _M_IX86_FP for x86 32-bit code, and _M_IX86_FP >= 1
-// indicates that at least SSE was targeted with the /arch:SSE option.
-// All x86-64 processors support SSE, so support can be assumed.
-// https://docs.microsoft.com/en-us/cpp/preprocessor/predefined-macros
-#define ABSL_INTERNAL_HAVE_SSE 1
-#endif
-
-// ABSL_INTERNAL_HAVE_SSE2 is used for compile-time detection of SSE2 support.
-// See https://gcc.gnu.org/onlinedocs/gcc/x86-Options.html for an overview of
-// which architectures support the various x86 instruction sets.
-#ifdef ABSL_INTERNAL_HAVE_SSE2
-#error ABSL_INTERNAL_HAVE_SSE2 cannot be directly set
-#elif defined(__SSE2__)
-#define ABSL_INTERNAL_HAVE_SSE2 1
-#elif (defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 2)) && \
-    !defined(_M_ARM64EC)
-// MSVC only defines _M_IX86_FP for x86 32-bit code, and _M_IX86_FP >= 2
-// indicates that at least SSE2 was targeted with the /arch:SSE2 option.
-// All x86-64 processors support SSE2, so support can be assumed.
-// https://docs.microsoft.com/en-us/cpp/preprocessor/predefined-macros
-#define ABSL_INTERNAL_HAVE_SSE2 1
-#endif
-
-// ABSL_INTERNAL_HAVE_SSSE3 is used for compile-time detection of SSSE3 support.
-// See https://gcc.gnu.org/onlinedocs/gcc/x86-Options.html for an overview of
-// which architectures support the various x86 instruction sets.
-//
-// MSVC does not have a mode that targets SSSE3 at compile-time. To use SSSE3
-// with MSVC requires either assuming that the code will only every run on CPUs
-// that support SSSE3, otherwise __cpuid() can be used to detect support at
-// runtime and fallback to a non-SSSE3 implementation when SSSE3 is unsupported
-// by the CPU.
-#ifdef ABSL_INTERNAL_HAVE_SSSE3
-#error ABSL_INTERNAL_HAVE_SSSE3 cannot be directly set
-#elif defined(__SSSE3__)
-#define ABSL_INTERNAL_HAVE_SSSE3 1
-#endif
-
-// ABSL_INTERNAL_HAVE_ARM_NEON is used for compile-time detection of NEON (ARM
-// SIMD).
-//
-// If __CUDA_ARCH__ is defined, then we are compiling CUDA code in device mode.
-// In device mode, NEON intrinsics are not available, regardless of host
-// platform.
-// https://llvm.org/docs/CompileCudaWithLLVM.html#detecting-clang-vs-nvcc-from-code
-#ifdef ABSL_INTERNAL_HAVE_ARM_NEON
-#error ABSL_INTERNAL_HAVE_ARM_NEON cannot be directly set
-#elif defined(__ARM_NEON) && !defined(__CUDA_ARCH__)
-#define ABSL_INTERNAL_HAVE_ARM_NEON 1
-#endif
-
-// ABSL_HAVE_CONSTANT_EVALUATED is used for compile-time detection of
-// constant evaluation support through `absl::is_constant_evaluated`.
-#ifdef ABSL_HAVE_CONSTANT_EVALUATED
-#error ABSL_HAVE_CONSTANT_EVALUATED cannot be directly set
-#endif
-#ifdef __cpp_lib_is_constant_evaluated
-#define ABSL_HAVE_CONSTANT_EVALUATED 1
-#elif ABSL_HAVE_BUILTIN(__builtin_is_constant_evaluated)
-#define ABSL_HAVE_CONSTANT_EVALUATED 1
-#endif
-
-// ABSL_INTERNAL_EMSCRIPTEN_VERSION combines Emscripten's three version macros
-// into an integer that can be compared against.
-#ifdef ABSL_INTERNAL_EMSCRIPTEN_VERSION
-#error ABSL_INTERNAL_EMSCRIPTEN_VERSION cannot be directly set
-#endif
-#ifdef __EMSCRIPTEN__
-#include <emscripten/version.h>
-#ifdef __EMSCRIPTEN_major__
-#if __EMSCRIPTEN_minor__ >= 1000
-#error __EMSCRIPTEN_minor__ is too big to fit in ABSL_INTERNAL_EMSCRIPTEN_VERSION
-#endif
-#if __EMSCRIPTEN_tiny__ >= 1000
-#error __EMSCRIPTEN_tiny__ is too big to fit in ABSL_INTERNAL_EMSCRIPTEN_VERSION
-#endif
-#define ABSL_INTERNAL_EMSCRIPTEN_VERSION                              \
-  ((__EMSCRIPTEN_major__) * 1000000 + (__EMSCRIPTEN_minor__) * 1000 + \
-   (__EMSCRIPTEN_tiny__))
-#endif
-#endif
-
-#endif  // ABSL_BASE_CONFIG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/const_init.h b/projects/llm_framework/include/abseil-cpp/absl/base/const_init.h
deleted file mode 100644
index 16520b61..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/const_init.h
+++ /dev/null
@@ -1,76 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// kConstInit
-// -----------------------------------------------------------------------------
-//
-// A constructor tag used to mark an object as safe for use as a global
-// variable, avoiding the usual lifetime issues that can affect globals.
-
-#ifndef ABSL_BASE_CONST_INIT_H_
-#define ABSL_BASE_CONST_INIT_H_
-
-#include "absl/base/config.h"
-
-// In general, objects with static storage duration (such as global variables)
-// can trigger tricky object lifetime situations.  Attempting to access them
-// from the constructors or destructors of other global objects can result in
-// undefined behavior, unless their constructors and destructors are designed
-// with this issue in mind.
-//
-// The normal way to deal with this issue in C++11 is to use constant
-// initialization and trivial destructors.
-//
-// Constant initialization is guaranteed to occur before any other code
-// executes.  Constructors that are declared 'constexpr' are eligible for
-// constant initialization.  You can annotate a variable declaration with the
-// ABSL_CONST_INIT macro to express this intent.  For compilers that support
-// it, this annotation will cause a compilation error for declarations that
-// aren't subject to constant initialization (perhaps because a runtime value
-// was passed as a constructor argument).
-//
-// On program shutdown, lifetime issues can be avoided on global objects by
-// ensuring that they contain  trivial destructors.  A class has a trivial
-// destructor unless it has a user-defined destructor, a virtual method or base
-// class, or a data member or base class with a non-trivial destructor of its
-// own.  Objects with static storage duration and a trivial destructor are not
-// cleaned up on program shutdown, and are thus safe to access from other code
-// running during shutdown.
-//
-// For a few core Abseil classes, we make a best effort to allow for safe global
-// instances, even though these classes have non-trivial destructors.  These
-// objects can be created with the absl::kConstInit tag.  For example:
-//   ABSL_CONST_INIT absl::Mutex global_mutex(absl::kConstInit);
-//
-// The line above declares a global variable of type absl::Mutex which can be
-// accessed at any point during startup or shutdown.  global_mutex's destructor
-// will still run, but will not invalidate the object.  Note that C++ specifies
-// that accessing an object after its destructor has run results in undefined
-// behavior, but this pattern works on the toolchains we support.
-//
-// The absl::kConstInit tag should only be used to define objects with static
-// or thread_local storage duration.
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-enum ConstInitType {
-  kConstInit,
-};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_CONST_INIT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/dynamic_annotations.h b/projects/llm_framework/include/abseil-cpp/absl/base/dynamic_annotations.h
deleted file mode 100644
index 7ba8912e..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/dynamic_annotations.h
+++ /dev/null
@@ -1,496 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-// This file defines dynamic annotations for use with dynamic analysis tool
-// such as valgrind, PIN, etc.
-//
-// Dynamic annotation is a source code annotation that affects the generated
-// code (that is, the annotation is not a comment). Each such annotation is
-// attached to a particular instruction and/or to a particular object (address)
-// in the program.
-//
-// The annotations that should be used by users are macros in all upper-case
-// (e.g., ABSL_ANNOTATE_THREAD_NAME).
-//
-// Actual implementation of these macros may differ depending on the dynamic
-// analysis tool being used.
-//
-// This file supports the following configurations:
-// - Dynamic Annotations enabled (with static thread-safety warnings disabled).
-//   In this case, macros expand to functions implemented by Thread Sanitizer,
-//   when building with TSan. When not provided an external implementation,
-//   dynamic_annotations.cc provides no-op implementations.
-//
-// - Static Clang thread-safety warnings enabled.
-//   When building with a Clang compiler that supports thread-safety warnings,
-//   a subset of annotations can be statically-checked at compile-time. We
-//   expand these macros to static-inline functions that can be analyzed for
-//   thread-safety, but afterwards elided when building the final binary.
-//
-// - All annotations are disabled.
-//   If neither Dynamic Annotations nor Clang thread-safety warnings are
-//   enabled, then all annotation-macros expand to empty.
-
-#ifndef ABSL_BASE_DYNAMIC_ANNOTATIONS_H_
-#define ABSL_BASE_DYNAMIC_ANNOTATIONS_H_
-
-#include <stddef.h>
-#include <stdint.h>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#ifdef __cplusplus
-#include "absl/base/macros.h"
-#endif
-
-#ifdef ABSL_HAVE_HWADDRESS_SANITIZER
-#include <sanitizer/hwasan_interface.h>
-#endif
-
-// TODO(rogeeff): Remove after the backward compatibility period.
-#include "absl/base/internal/dynamic_annotations.h"  // IWYU pragma: export
-
-// -------------------------------------------------------------------------
-// Decide which features are enabled.
-
-#ifdef ABSL_HAVE_THREAD_SANITIZER
-
-#define ABSL_INTERNAL_RACE_ANNOTATIONS_ENABLED 1
-#define ABSL_INTERNAL_READS_ANNOTATIONS_ENABLED 1
-#define ABSL_INTERNAL_WRITES_ANNOTATIONS_ENABLED 1
-#define ABSL_INTERNAL_ANNOTALYSIS_ENABLED 0
-#define ABSL_INTERNAL_READS_WRITES_ANNOTATIONS_ENABLED 1
-
-#else
-
-#define ABSL_INTERNAL_RACE_ANNOTATIONS_ENABLED 0
-#define ABSL_INTERNAL_READS_ANNOTATIONS_ENABLED 0
-#define ABSL_INTERNAL_WRITES_ANNOTATIONS_ENABLED 0
-
-// Clang provides limited support for static thread-safety analysis through a
-// feature called Annotalysis. We configure macro-definitions according to
-// whether Annotalysis support is available. When running in opt-mode, GCC
-// will issue a warning, if these attributes are compiled. Only include them
-// when compiling using Clang.
-
-#if defined(__clang__)
-#define ABSL_INTERNAL_ANNOTALYSIS_ENABLED 1
-#if !defined(SWIG)
-#define ABSL_INTERNAL_IGNORE_READS_ATTRIBUTE_ENABLED 1
-#endif
-#else
-#define ABSL_INTERNAL_ANNOTALYSIS_ENABLED 0
-#endif
-
-// Read/write annotations are enabled in Annotalysis mode; disabled otherwise.
-#define ABSL_INTERNAL_READS_WRITES_ANNOTATIONS_ENABLED \
-  ABSL_INTERNAL_ANNOTALYSIS_ENABLED
-
-#endif  // ABSL_HAVE_THREAD_SANITIZER
-
-#ifdef __cplusplus
-#define ABSL_INTERNAL_BEGIN_EXTERN_C extern "C" {
-#define ABSL_INTERNAL_END_EXTERN_C }  // extern "C"
-#define ABSL_INTERNAL_GLOBAL_SCOPED(F) ::F
-#define ABSL_INTERNAL_STATIC_INLINE inline
-#else
-#define ABSL_INTERNAL_BEGIN_EXTERN_C  // empty
-#define ABSL_INTERNAL_END_EXTERN_C    // empty
-#define ABSL_INTERNAL_GLOBAL_SCOPED(F) F
-#define ABSL_INTERNAL_STATIC_INLINE static inline
-#endif
-
-// -------------------------------------------------------------------------
-// Define race annotations.
-
-#if ABSL_INTERNAL_RACE_ANNOTATIONS_ENABLED == 1
-// Some of the symbols used in this section (e.g. AnnotateBenignRaceSized) are
-// defined by the compiler-based sanitizer implementation, not by the Abseil
-// library. Therefore they do not use ABSL_INTERNAL_C_SYMBOL.
-
-// -------------------------------------------------------------
-// Annotations that suppress errors. It is usually better to express the
-// program's synchronization using the other annotations, but these can be used
-// when all else fails.
-
-// Report that we may have a benign race at `pointer`, with size
-// "sizeof(*(pointer))". `pointer` must be a non-void* pointer. Insert at the
-// point where `pointer` has been allocated, preferably close to the point
-// where the race happens. See also ABSL_ANNOTATE_BENIGN_RACE_STATIC.
-#define ABSL_ANNOTATE_BENIGN_RACE(pointer, description) \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateBenignRaceSized)  \
-  (__FILE__, __LINE__, pointer, sizeof(*(pointer)), description)
-
-// Same as ABSL_ANNOTATE_BENIGN_RACE(`address`, `description`), but applies to
-// the memory range [`address`, `address`+`size`).
-#define ABSL_ANNOTATE_BENIGN_RACE_SIZED(address, size, description) \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateBenignRaceSized)              \
-  (__FILE__, __LINE__, address, size, description)
-
-// Enable (`enable`!=0) or disable (`enable`==0) race detection for all threads.
-// This annotation could be useful if you want to skip expensive race analysis
-// during some period of program execution, e.g. during initialization.
-#define ABSL_ANNOTATE_ENABLE_RACE_DETECTION(enable)        \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateEnableRaceDetection) \
-  (__FILE__, __LINE__, enable)
-
-// -------------------------------------------------------------
-// Annotations useful for debugging.
-
-// Report the current thread `name` to a race detector.
-#define ABSL_ANNOTATE_THREAD_NAME(name) \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateThreadName)(__FILE__, __LINE__, name)
-
-// -------------------------------------------------------------
-// Annotations useful when implementing locks. They are not normally needed by
-// modules that merely use locks. The `lock` argument is a pointer to the lock
-// object.
-
-// Report that a lock has been created at address `lock`.
-#define ABSL_ANNOTATE_RWLOCK_CREATE(lock) \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateRWLockCreate)(__FILE__, __LINE__, lock)
-
-// Report that a linker initialized lock has been created at address `lock`.
-#ifdef ABSL_HAVE_THREAD_SANITIZER
-#define ABSL_ANNOTATE_RWLOCK_CREATE_STATIC(lock)          \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateRWLockCreateStatic) \
-  (__FILE__, __LINE__, lock)
-#else
-#define ABSL_ANNOTATE_RWLOCK_CREATE_STATIC(lock) \
-  ABSL_ANNOTATE_RWLOCK_CREATE(lock)
-#endif
-
-// Report that the lock at address `lock` is about to be destroyed.
-#define ABSL_ANNOTATE_RWLOCK_DESTROY(lock) \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateRWLockDestroy)(__FILE__, __LINE__, lock)
-
-// Report that the lock at address `lock` has been acquired.
-// `is_w`=1 for writer lock, `is_w`=0 for reader lock.
-#define ABSL_ANNOTATE_RWLOCK_ACQUIRED(lock, is_w)     \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateRWLockAcquired) \
-  (__FILE__, __LINE__, lock, is_w)
-
-// Report that the lock at address `lock` is about to be released.
-// `is_w`=1 for writer lock, `is_w`=0 for reader lock.
-#define ABSL_ANNOTATE_RWLOCK_RELEASED(lock, is_w)     \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateRWLockReleased) \
-  (__FILE__, __LINE__, lock, is_w)
-
-// Apply ABSL_ANNOTATE_BENIGN_RACE_SIZED to a static variable `static_var`.
-#define ABSL_ANNOTATE_BENIGN_RACE_STATIC(static_var, description)      \
-  namespace {                                                          \
-  class static_var##_annotator {                                       \
-   public:                                                             \
-    static_var##_annotator() {                                         \
-      ABSL_ANNOTATE_BENIGN_RACE_SIZED(&static_var, sizeof(static_var), \
-                                      #static_var ": " description);   \
-    }                                                                  \
-  };                                                                   \
-  static static_var##_annotator the##static_var##_annotator;           \
-  }  // namespace
-
-// Function prototypes of annotations provided by the compiler-based sanitizer
-// implementation.
-ABSL_INTERNAL_BEGIN_EXTERN_C
-void AnnotateRWLockCreate(const char* file, int line,
-                          const volatile void* lock);
-void AnnotateRWLockCreateStatic(const char* file, int line,
-                                const volatile void* lock);
-void AnnotateRWLockDestroy(const char* file, int line,
-                           const volatile void* lock);
-void AnnotateRWLockAcquired(const char* file, int line,
-                            const volatile void* lock, long is_w);  // NOLINT
-void AnnotateRWLockReleased(const char* file, int line,
-                            const volatile void* lock, long is_w);  // NOLINT
-void AnnotateBenignRace(const char* file, int line,
-                        const volatile void* address, const char* description);
-void AnnotateBenignRaceSized(const char* file, int line,
-                             const volatile void* address, size_t size,
-                             const char* description);
-void AnnotateThreadName(const char* file, int line, const char* name);
-void AnnotateEnableRaceDetection(const char* file, int line, int enable);
-ABSL_INTERNAL_END_EXTERN_C
-
-#else  // ABSL_INTERNAL_RACE_ANNOTATIONS_ENABLED == 0
-
-#define ABSL_ANNOTATE_RWLOCK_CREATE(lock)                            // empty
-#define ABSL_ANNOTATE_RWLOCK_CREATE_STATIC(lock)                     // empty
-#define ABSL_ANNOTATE_RWLOCK_DESTROY(lock)                           // empty
-#define ABSL_ANNOTATE_RWLOCK_ACQUIRED(lock, is_w)                    // empty
-#define ABSL_ANNOTATE_RWLOCK_RELEASED(lock, is_w)                    // empty
-#define ABSL_ANNOTATE_BENIGN_RACE(address, description)              // empty
-#define ABSL_ANNOTATE_BENIGN_RACE_SIZED(address, size, description)  // empty
-#define ABSL_ANNOTATE_THREAD_NAME(name)                              // empty
-#define ABSL_ANNOTATE_ENABLE_RACE_DETECTION(enable)                  // empty
-#define ABSL_ANNOTATE_BENIGN_RACE_STATIC(static_var, description)    // empty
-
-#endif  // ABSL_INTERNAL_RACE_ANNOTATIONS_ENABLED
-
-// -------------------------------------------------------------------------
-// Define memory annotations.
-
-#ifdef ABSL_HAVE_MEMORY_SANITIZER
-
-#include <sanitizer/msan_interface.h>
-
-#define ABSL_ANNOTATE_MEMORY_IS_INITIALIZED(address, size) \
-  __msan_unpoison(address, size)
-
-#define ABSL_ANNOTATE_MEMORY_IS_UNINITIALIZED(address, size) \
-  __msan_allocated_memory(address, size)
-
-#else  // !defined(ABSL_HAVE_MEMORY_SANITIZER)
-
-// TODO(rogeeff): remove this branch
-#ifdef ABSL_HAVE_THREAD_SANITIZER
-#define ABSL_ANNOTATE_MEMORY_IS_INITIALIZED(address, size) \
-  do {                                                     \
-    (void)(address);                                       \
-    (void)(size);                                          \
-  } while (0)
-#define ABSL_ANNOTATE_MEMORY_IS_UNINITIALIZED(address, size) \
-  do {                                                       \
-    (void)(address);                                         \
-    (void)(size);                                            \
-  } while (0)
-#else
-
-#define ABSL_ANNOTATE_MEMORY_IS_INITIALIZED(address, size)    // empty
-#define ABSL_ANNOTATE_MEMORY_IS_UNINITIALIZED(address, size)  // empty
-
-#endif
-
-#endif  // ABSL_HAVE_MEMORY_SANITIZER
-
-// -------------------------------------------------------------------------
-// Define IGNORE_READS_BEGIN/_END attributes.
-
-#if defined(ABSL_INTERNAL_IGNORE_READS_ATTRIBUTE_ENABLED)
-
-#define ABSL_INTERNAL_IGNORE_READS_BEGIN_ATTRIBUTE \
-  __attribute((exclusive_lock_function("*")))
-#define ABSL_INTERNAL_IGNORE_READS_END_ATTRIBUTE \
-  __attribute((unlock_function("*")))
-
-#else  // !defined(ABSL_INTERNAL_IGNORE_READS_ATTRIBUTE_ENABLED)
-
-#define ABSL_INTERNAL_IGNORE_READS_BEGIN_ATTRIBUTE  // empty
-#define ABSL_INTERNAL_IGNORE_READS_END_ATTRIBUTE    // empty
-
-#endif  // defined(ABSL_INTERNAL_IGNORE_READS_ATTRIBUTE_ENABLED)
-
-// -------------------------------------------------------------------------
-// Define IGNORE_READS_BEGIN/_END annotations.
-
-#if ABSL_INTERNAL_READS_ANNOTATIONS_ENABLED == 1
-// Some of the symbols used in this section (e.g. AnnotateIgnoreReadsBegin) are
-// defined by the compiler-based implementation, not by the Abseil
-// library. Therefore they do not use ABSL_INTERNAL_C_SYMBOL.
-
-// Request the analysis tool to ignore all reads in the current thread until
-// ABSL_ANNOTATE_IGNORE_READS_END is called. Useful to ignore intentional racey
-// reads, while still checking other reads and all writes.
-// See also ABSL_ANNOTATE_UNPROTECTED_READ.
-#define ABSL_ANNOTATE_IGNORE_READS_BEGIN()              \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateIgnoreReadsBegin) \
-  (__FILE__, __LINE__)
-
-// Stop ignoring reads.
-#define ABSL_ANNOTATE_IGNORE_READS_END()              \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateIgnoreReadsEnd) \
-  (__FILE__, __LINE__)
-
-// Function prototypes of annotations provided by the compiler-based sanitizer
-// implementation.
-ABSL_INTERNAL_BEGIN_EXTERN_C
-void AnnotateIgnoreReadsBegin(const char* file, int line)
-    ABSL_INTERNAL_IGNORE_READS_BEGIN_ATTRIBUTE;
-void AnnotateIgnoreReadsEnd(const char* file,
-                            int line) ABSL_INTERNAL_IGNORE_READS_END_ATTRIBUTE;
-ABSL_INTERNAL_END_EXTERN_C
-
-#elif defined(ABSL_INTERNAL_ANNOTALYSIS_ENABLED)
-
-// When Annotalysis is enabled without Dynamic Annotations, the use of
-// static-inline functions allows the annotations to be read at compile-time,
-// while still letting the compiler elide the functions from the final build.
-//
-// TODO(delesley) -- The exclusive lock here ignores writes as well, but
-// allows IGNORE_READS_AND_WRITES to work properly.
-
-#define ABSL_ANNOTATE_IGNORE_READS_BEGIN()                          \
-  ABSL_INTERNAL_GLOBAL_SCOPED(                                      \
-      ABSL_INTERNAL_C_SYMBOL(AbslInternalAnnotateIgnoreReadsBegin)) \
-  ()
-
-#define ABSL_ANNOTATE_IGNORE_READS_END()                          \
-  ABSL_INTERNAL_GLOBAL_SCOPED(                                    \
-      ABSL_INTERNAL_C_SYMBOL(AbslInternalAnnotateIgnoreReadsEnd)) \
-  ()
-
-ABSL_INTERNAL_STATIC_INLINE void ABSL_INTERNAL_C_SYMBOL(
-    AbslInternalAnnotateIgnoreReadsBegin)()
-    ABSL_INTERNAL_IGNORE_READS_BEGIN_ATTRIBUTE {}
-
-ABSL_INTERNAL_STATIC_INLINE void ABSL_INTERNAL_C_SYMBOL(
-    AbslInternalAnnotateIgnoreReadsEnd)()
-    ABSL_INTERNAL_IGNORE_READS_END_ATTRIBUTE {}
-
-#else
-
-#define ABSL_ANNOTATE_IGNORE_READS_BEGIN()  // empty
-#define ABSL_ANNOTATE_IGNORE_READS_END()    // empty
-
-#endif
-
-// -------------------------------------------------------------------------
-// Define IGNORE_WRITES_BEGIN/_END annotations.
-
-#if ABSL_INTERNAL_WRITES_ANNOTATIONS_ENABLED == 1
-
-// Similar to ABSL_ANNOTATE_IGNORE_READS_BEGIN, but ignore writes instead.
-#define ABSL_ANNOTATE_IGNORE_WRITES_BEGIN() \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateIgnoreWritesBegin)(__FILE__, __LINE__)
-
-// Stop ignoring writes.
-#define ABSL_ANNOTATE_IGNORE_WRITES_END() \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateIgnoreWritesEnd)(__FILE__, __LINE__)
-
-// Function prototypes of annotations provided by the compiler-based sanitizer
-// implementation.
-ABSL_INTERNAL_BEGIN_EXTERN_C
-void AnnotateIgnoreWritesBegin(const char* file, int line);
-void AnnotateIgnoreWritesEnd(const char* file, int line);
-ABSL_INTERNAL_END_EXTERN_C
-
-#else
-
-#define ABSL_ANNOTATE_IGNORE_WRITES_BEGIN()  // empty
-#define ABSL_ANNOTATE_IGNORE_WRITES_END()    // empty
-
-#endif
-
-// -------------------------------------------------------------------------
-// Define the ABSL_ANNOTATE_IGNORE_READS_AND_WRITES_* annotations using the more
-// primitive annotations defined above.
-//
-//     Instead of doing
-//        ABSL_ANNOTATE_IGNORE_READS_BEGIN();
-//        ... = x;
-//        ABSL_ANNOTATE_IGNORE_READS_END();
-//     one can use
-//        ... = ABSL_ANNOTATE_UNPROTECTED_READ(x);
-
-#if defined(ABSL_INTERNAL_READS_WRITES_ANNOTATIONS_ENABLED)
-
-// Start ignoring all memory accesses (both reads and writes).
-#define ABSL_ANNOTATE_IGNORE_READS_AND_WRITES_BEGIN() \
-  do {                                                \
-    ABSL_ANNOTATE_IGNORE_READS_BEGIN();               \
-    ABSL_ANNOTATE_IGNORE_WRITES_BEGIN();              \
-  } while (0)
-
-// Stop ignoring both reads and writes.
-#define ABSL_ANNOTATE_IGNORE_READS_AND_WRITES_END() \
-  do {                                              \
-    ABSL_ANNOTATE_IGNORE_WRITES_END();              \
-    ABSL_ANNOTATE_IGNORE_READS_END();               \
-  } while (0)
-
-#ifdef __cplusplus
-// ABSL_ANNOTATE_UNPROTECTED_READ is the preferred way to annotate racey reads.
-#define ABSL_ANNOTATE_UNPROTECTED_READ(x) \
-  absl::base_internal::AnnotateUnprotectedRead(x)
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-template <typename T>
-inline T AnnotateUnprotectedRead(const volatile T& x) {  // NOLINT
-  ABSL_ANNOTATE_IGNORE_READS_BEGIN();
-  T res = x;
-  ABSL_ANNOTATE_IGNORE_READS_END();
-  return res;
-}
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-#endif
-
-#else
-
-#define ABSL_ANNOTATE_IGNORE_READS_AND_WRITES_BEGIN()  // empty
-#define ABSL_ANNOTATE_IGNORE_READS_AND_WRITES_END()    // empty
-#define ABSL_ANNOTATE_UNPROTECTED_READ(x) (x)
-
-#endif
-
-// -------------------------------------------------------------------------
-// Address sanitizer annotations
-
-#ifdef ABSL_HAVE_ADDRESS_SANITIZER
-// Describe the current state of a contiguous container such as e.g.
-// std::vector or std::string. For more details see
-// sanitizer/common_interface_defs.h, which is provided by the compiler.
-#include <sanitizer/common_interface_defs.h>
-
-#define ABSL_ANNOTATE_CONTIGUOUS_CONTAINER(beg, end, old_mid, new_mid) \
-  __sanitizer_annotate_contiguous_container(beg, end, old_mid, new_mid)
-#define ABSL_ADDRESS_SANITIZER_REDZONE(name) \
-  struct {                                   \
-    alignas(8) char x[8];                    \
-  } name
-
-#else
-
-#define ABSL_ANNOTATE_CONTIGUOUS_CONTAINER(beg, end, old_mid, new_mid)  // empty
-#define ABSL_ADDRESS_SANITIZER_REDZONE(name) static_assert(true, "")
-
-#endif  // ABSL_HAVE_ADDRESS_SANITIZER
-
-// -------------------------------------------------------------------------
-// HWAddress sanitizer annotations
-
-#ifdef __cplusplus
-namespace absl {
-#ifdef ABSL_HAVE_HWADDRESS_SANITIZER
-// Under HWASAN changes the tag of the pointer.
-template <typename T>
-T* HwasanTagPointer(T* ptr, uintptr_t tag) {
-  return reinterpret_cast<T*>(__hwasan_tag_pointer(ptr, tag));
-}
-#else
-template <typename T>
-T* HwasanTagPointer(T* ptr, uintptr_t) {
-  return ptr;
-}
-#endif
-}  // namespace absl
-#endif
-
-// -------------------------------------------------------------------------
-// Undefine the macros intended only for this file.
-
-#undef ABSL_INTERNAL_RACE_ANNOTATIONS_ENABLED
-#undef ABSL_INTERNAL_READS_ANNOTATIONS_ENABLED
-#undef ABSL_INTERNAL_WRITES_ANNOTATIONS_ENABLED
-#undef ABSL_INTERNAL_ANNOTALYSIS_ENABLED
-#undef ABSL_INTERNAL_READS_WRITES_ANNOTATIONS_ENABLED
-#undef ABSL_INTERNAL_BEGIN_EXTERN_C
-#undef ABSL_INTERNAL_END_EXTERN_C
-#undef ABSL_INTERNAL_STATIC_INLINE
-
-#endif  // ABSL_BASE_DYNAMIC_ANNOTATIONS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/atomic_hook.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/atomic_hook.h
deleted file mode 100644
index ae21cd7f..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/atomic_hook.h
+++ /dev/null
@@ -1,200 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_BASE_INTERNAL_ATOMIC_HOOK_H_
-#define ABSL_BASE_INTERNAL_ATOMIC_HOOK_H_
-
-#include <atomic>
-#include <cassert>
-#include <cstdint>
-#include <utility>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-
-#if defined(_MSC_VER) && !defined(__clang__)
-#define ABSL_HAVE_WORKING_CONSTEXPR_STATIC_INIT 0
-#else
-#define ABSL_HAVE_WORKING_CONSTEXPR_STATIC_INIT 1
-#endif
-
-#if defined(_MSC_VER)
-#define ABSL_HAVE_WORKING_ATOMIC_POINTER 0
-#else
-#define ABSL_HAVE_WORKING_ATOMIC_POINTER 1
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-template <typename T>
-class AtomicHook;
-
-// To workaround AtomicHook not being constant-initializable on some platforms,
-// prefer to annotate instances with `ABSL_INTERNAL_ATOMIC_HOOK_ATTRIBUTES`
-// instead of `ABSL_CONST_INIT`.
-#if ABSL_HAVE_WORKING_CONSTEXPR_STATIC_INIT
-#define ABSL_INTERNAL_ATOMIC_HOOK_ATTRIBUTES ABSL_CONST_INIT
-#else
-#define ABSL_INTERNAL_ATOMIC_HOOK_ATTRIBUTES
-#endif
-
-// `AtomicHook` is a helper class, templatized on a raw function pointer type,
-// for implementing Abseil customization hooks.  It is a callable object that
-// dispatches to the registered hook.  Objects of type `AtomicHook` must have
-// static or thread storage duration.
-//
-// A default constructed object performs a no-op (and returns a default
-// constructed object) if no hook has been registered.
-//
-// Hooks can be pre-registered via constant initialization, for example:
-//
-// ABSL_INTERNAL_ATOMIC_HOOK_ATTRIBUTES static AtomicHook<void(*)()>
-//     my_hook(DefaultAction);
-//
-// and then changed at runtime via a call to `Store()`.
-//
-// Reads and writes guarantee memory_order_acquire/memory_order_release
-// semantics.
-template <typename ReturnType, typename... Args>
-class AtomicHook<ReturnType (*)(Args...)> {
- public:
-  using FnPtr = ReturnType (*)(Args...);
-
-  // Constructs an object that by default performs a no-op (and
-  // returns a default constructed object) when no hook as been registered.
-  constexpr AtomicHook() : AtomicHook(DummyFunction) {}
-
-  // Constructs an object that by default dispatches to/returns the
-  // pre-registered default_fn when no hook has been registered at runtime.
-#if ABSL_HAVE_WORKING_ATOMIC_POINTER && ABSL_HAVE_WORKING_CONSTEXPR_STATIC_INIT
-  explicit constexpr AtomicHook(FnPtr default_fn)
-      : hook_(default_fn), default_fn_(default_fn) {}
-#elif ABSL_HAVE_WORKING_CONSTEXPR_STATIC_INIT
-  explicit constexpr AtomicHook(FnPtr default_fn)
-      : hook_(kUninitialized), default_fn_(default_fn) {}
-#else
-  // As of January 2020, on all known versions of MSVC this constructor runs in
-  // the global constructor sequence.  If `Store()` is called by a dynamic
-  // initializer, we want to preserve the value, even if this constructor runs
-  // after the call to `Store()`.  If not, `hook_` will be
-  // zero-initialized by the linker and we have no need to set it.
-  // https://developercommunity.visualstudio.com/content/problem/336946/class-with-constexpr-constructor-not-using-static.html
-  explicit constexpr AtomicHook(FnPtr default_fn)
-      : /* hook_(deliberately omitted), */ default_fn_(default_fn) {
-    static_assert(kUninitialized == 0, "here we rely on zero-initialization");
-  }
-#endif
-
-  // Stores the provided function pointer as the value for this hook.
-  //
-  // This is intended to be called once.  Multiple calls are legal only if the
-  // same function pointer is provided for each call.  The store is implemented
-  // as a memory_order_release operation, and read accesses are implemented as
-  // memory_order_acquire.
-  void Store(FnPtr fn) {
-    bool success = DoStore(fn);
-    static_cast<void>(success);
-    assert(success);
-  }
-
-  // Invokes the registered callback.  If no callback has yet been registered, a
-  // default-constructed object of the appropriate type is returned instead.
-  template <typename... CallArgs>
-  ReturnType operator()(CallArgs&&... args) const {
-    return DoLoad()(std::forward<CallArgs>(args)...);
-  }
-
-  // Returns the registered callback, or nullptr if none has been registered.
-  // Useful if client code needs to conditionalize behavior based on whether a
-  // callback was registered.
-  //
-  // Note that atomic_hook.Load()() and atomic_hook() have different semantics:
-  // operator()() will perform a no-op if no callback was registered, while
-  // Load()() will dereference a null function pointer.  Prefer operator()() to
-  // Load()() unless you must conditionalize behavior on whether a hook was
-  // registered.
-  FnPtr Load() const {
-    FnPtr ptr = DoLoad();
-    return (ptr == DummyFunction) ? nullptr : ptr;
-  }
-
- private:
-  static ReturnType DummyFunction(Args...) {
-    return ReturnType();
-  }
-
-  // Current versions of MSVC (as of September 2017) have a broken
-  // implementation of std::atomic<T*>:  Its constructor attempts to do the
-  // equivalent of a reinterpret_cast in a constexpr context, which is not
-  // allowed.
-  //
-  // This causes an issue when building with LLVM under Windows.  To avoid this,
-  // we use a less-efficient, intptr_t-based implementation on Windows.
-#if ABSL_HAVE_WORKING_ATOMIC_POINTER
-  // Return the stored value, or DummyFunction if no value has been stored.
-  FnPtr DoLoad() const { return hook_.load(std::memory_order_acquire); }
-
-  // Store the given value.  Returns false if a different value was already
-  // stored to this object.
-  bool DoStore(FnPtr fn) {
-    assert(fn);
-    FnPtr expected = default_fn_;
-    const bool store_succeeded = hook_.compare_exchange_strong(
-        expected, fn, std::memory_order_acq_rel, std::memory_order_acquire);
-    const bool same_value_already_stored = (expected == fn);
-    return store_succeeded || same_value_already_stored;
-  }
-
-  std::atomic<FnPtr> hook_;
-#else  // !ABSL_HAVE_WORKING_ATOMIC_POINTER
-  // Use a sentinel value unlikely to be the address of an actual function.
-  static constexpr intptr_t kUninitialized = 0;
-
-  static_assert(sizeof(intptr_t) >= sizeof(FnPtr),
-                "intptr_t can't contain a function pointer");
-
-  FnPtr DoLoad() const {
-    const intptr_t value = hook_.load(std::memory_order_acquire);
-    if (value == kUninitialized) {
-      return default_fn_;
-    }
-    return reinterpret_cast<FnPtr>(value);
-  }
-
-  bool DoStore(FnPtr fn) {
-    assert(fn);
-    const auto value = reinterpret_cast<intptr_t>(fn);
-    intptr_t expected = kUninitialized;
-    const bool store_succeeded = hook_.compare_exchange_strong(
-        expected, value, std::memory_order_acq_rel, std::memory_order_acquire);
-    const bool same_value_already_stored = (expected == value);
-    return store_succeeded || same_value_already_stored;
-  }
-
-  std::atomic<intptr_t> hook_;
-#endif
-
-  const FnPtr default_fn_;
-};
-
-#undef ABSL_HAVE_WORKING_ATOMIC_POINTER
-#undef ABSL_HAVE_WORKING_CONSTEXPR_STATIC_INIT
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_ATOMIC_HOOK_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/atomic_hook_test_helper.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/atomic_hook_test_helper.h
deleted file mode 100644
index c72015ef..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/atomic_hook_test_helper.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_BASE_INTERNAL_ATOMIC_HOOK_TEST_HELPER_H_
-#define ABSL_BASE_INTERNAL_ATOMIC_HOOK_TEST_HELPER_H_
-
-#include "absl/base/internal/atomic_hook.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace atomic_hook_internal {
-
-using VoidF = void (*)();
-extern absl::base_internal::AtomicHook<VoidF> func;
-extern int default_func_calls;
-void DefaultFunc();
-void RegisterFunc(VoidF func);
-
-}  // namespace atomic_hook_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_ATOMIC_HOOK_TEST_HELPER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/cycleclock.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/cycleclock.h
deleted file mode 100644
index cbfdf579..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/cycleclock.h
+++ /dev/null
@@ -1,144 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-// -----------------------------------------------------------------------------
-// File: cycleclock.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines a `CycleClock`, which yields the value and frequency
-// of a cycle counter that increments at a rate that is approximately constant.
-//
-// NOTE:
-//
-// The cycle counter frequency is not necessarily related to the core clock
-// frequency and should not be treated as such. That is, `CycleClock` cycles are
-// not necessarily "CPU cycles" and code should not rely on that behavior, even
-// if experimentally observed.
-//
-// An arbitrary offset may have been added to the counter at power on.
-//
-// On some platforms, the rate and offset of the counter may differ
-// slightly when read from different CPUs of a multiprocessor. Usually,
-// we try to ensure that the operating system adjusts values periodically
-// so that values agree approximately.   If you need stronger guarantees,
-// consider using alternate interfaces.
-//
-// The CPU is not required to maintain the ordering of a cycle counter read
-// with respect to surrounding instructions.
-
-#ifndef ABSL_BASE_INTERNAL_CYCLECLOCK_H_
-#define ABSL_BASE_INTERNAL_CYCLECLOCK_H_
-
-#include <atomic>
-#include <cstdint>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/internal/cycleclock_config.h"
-#include "absl/base/internal/unscaledcycleclock.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-using CycleClockSourceFunc = int64_t (*)();
-
-// -----------------------------------------------------------------------------
-// CycleClock
-// -----------------------------------------------------------------------------
-class CycleClock {
- public:
-  // CycleClock::Now()
-  //
-  // Returns the value of a cycle counter that counts at a rate that is
-  // approximately constant.
-  static int64_t Now();
-
-  // CycleClock::Frequency()
-  //
-  // Returns the amount by which `CycleClock::Now()` increases per second. Note
-  // that this value may not necessarily match the core CPU clock frequency.
-  static double Frequency();
-
- private:
-#if ABSL_USE_UNSCALED_CYCLECLOCK
-  static CycleClockSourceFunc LoadCycleClockSource();
-
-  static constexpr int32_t kShift = kCycleClockShift;
-  static constexpr double kFrequencyScale = kCycleClockFrequencyScale;
-
-  ABSL_CONST_INIT static std::atomic<CycleClockSourceFunc> cycle_clock_source_;
-#endif  //  ABSL_USE_UNSCALED_CYCLECLOC
-
-  CycleClock() = delete;  // no instances
-  CycleClock(const CycleClock&) = delete;
-  CycleClock& operator=(const CycleClock&) = delete;
-
-  friend class CycleClockSource;
-};
-
-class CycleClockSource {
- private:
-  // CycleClockSource::Register()
-  //
-  // Register a function that provides an alternate source for the unscaled CPU
-  // cycle count value. The source function must be async signal safe, must not
-  // call CycleClock::Now(), and must have a frequency that matches that of the
-  // unscaled clock used by CycleClock. A nullptr value resets CycleClock to use
-  // the default source.
-  static void Register(CycleClockSourceFunc source);
-};
-
-#if ABSL_USE_UNSCALED_CYCLECLOCK
-
-inline CycleClockSourceFunc CycleClock::LoadCycleClockSource() {
-#if !defined(__x86_64__)
-  // Optimize for the common case (no callback) by first doing a relaxed load;
-  // this is significantly faster on non-x86 platforms.
-  if (cycle_clock_source_.load(std::memory_order_relaxed) == nullptr) {
-    return nullptr;
-  }
-#endif  // !defined(__x86_64__)
-
-  // This corresponds to the store(std::memory_order_release) in
-  // CycleClockSource::Register, and makes sure that any updates made prior to
-  // registering the callback are visible to this thread before the callback
-  // is invoked.
-  return cycle_clock_source_.load(std::memory_order_acquire);
-}
-
-// Accessing globals in inlined code in Window DLLs is problematic.
-#ifndef _WIN32
-inline int64_t CycleClock::Now() {
-  auto fn = LoadCycleClockSource();
-  if (fn == nullptr) {
-    return base_internal::UnscaledCycleClock::Now() >> kShift;
-  }
-  return fn() >> kShift;
-}
-#endif
-
-inline double CycleClock::Frequency() {
-  return kFrequencyScale * base_internal::UnscaledCycleClock::Frequency();
-}
-
-#endif  // ABSL_USE_UNSCALED_CYCLECLOCK
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_CYCLECLOCK_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/cycleclock_config.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/cycleclock_config.h
deleted file mode 100644
index 191112b5..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/cycleclock_config.h
+++ /dev/null
@@ -1,55 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_BASE_INTERNAL_CYCLECLOCK_CONFIG_H_
-#define ABSL_BASE_INTERNAL_CYCLECLOCK_CONFIG_H_
-
-#include <cstdint>
-
-#include "absl/base/config.h"
-#include "absl/base/internal/inline_variable.h"
-#include "absl/base/internal/unscaledcycleclock_config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-#if ABSL_USE_UNSCALED_CYCLECLOCK
-#ifdef NDEBUG
-#ifdef ABSL_INTERNAL_UNSCALED_CYCLECLOCK_FREQUENCY_IS_CPU_FREQUENCY
-// Not debug mode and the UnscaledCycleClock frequency is the CPU
-// frequency.  Scale the CycleClock to prevent overflow if someone
-// tries to represent the time as cycles since the Unix epoch.
-ABSL_INTERNAL_INLINE_CONSTEXPR(int32_t, kCycleClockShift, 1);
-#else
-// Not debug mode and the UnscaledCycleClock isn't operating at the
-// raw CPU frequency. There is no need to do any scaling, so don't
-// needlessly sacrifice precision.
-ABSL_INTERNAL_INLINE_CONSTEXPR(int32_t, kCycleClockShift, 0);
-#endif
-#else   // NDEBUG
-// In debug mode use a different shift to discourage depending on a
-// particular shift value.
-ABSL_INTERNAL_INLINE_CONSTEXPR(int32_t, kCycleClockShift, 2);
-#endif  // NDEBUG
-
-ABSL_INTERNAL_INLINE_CONSTEXPR(double, kCycleClockFrequencyScale,
-                               1.0 / (1 << kCycleClockShift));
-#endif  //  ABSL_USE_UNSCALED_CYCLECLOC
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_CYCLECLOCK_CONFIG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/direct_mmap.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/direct_mmap.h
deleted file mode 100644
index 1beb2ee4..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/direct_mmap.h
+++ /dev/null
@@ -1,170 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Functions for directly invoking mmap() via syscall, avoiding the case where
-// mmap() has been locally overridden.
-
-#ifndef ABSL_BASE_INTERNAL_DIRECT_MMAP_H_
-#define ABSL_BASE_INTERNAL_DIRECT_MMAP_H_
-
-#include "absl/base/config.h"
-
-#ifdef ABSL_HAVE_MMAP
-
-#include <sys/mman.h>
-
-#ifdef __linux__
-
-#include <sys/types.h>
-#ifdef __BIONIC__
-#include <sys/syscall.h>
-#else
-#include <syscall.h>
-#endif
-
-#include <linux/unistd.h>
-#include <unistd.h>
-#include <cerrno>
-#include <cstdarg>
-#include <cstdint>
-
-#ifdef __mips__
-// Include definitions of the ABI currently in use.
-#if defined(__BIONIC__) || !defined(__GLIBC__)
-// Android doesn't have sgidefs.h, but does have asm/sgidefs.h, which has the
-// definitions we need.
-#include <asm/sgidefs.h>
-#else
-#include <sgidefs.h>
-#endif  // __BIONIC__ || !__GLIBC__
-#endif  // __mips__
-
-// SYS_mmap and SYS_munmap are not defined in Android.
-#ifdef __BIONIC__
-extern "C" void* __mmap2(void*, size_t, int, int, int, size_t);
-#if defined(__NR_mmap) && !defined(SYS_mmap)
-#define SYS_mmap __NR_mmap
-#endif
-#ifndef SYS_munmap
-#define SYS_munmap __NR_munmap
-#endif
-#endif  // __BIONIC__
-
-#if defined(__NR_mmap2) && !defined(SYS_mmap2)
-#define SYS_mmap2 __NR_mmap2
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-// Platform specific logic extracted from
-// https://chromium.googlesource.com/linux-syscall-support/+/master/linux_syscall_support.h
-inline void* DirectMmap(void* start, size_t length, int prot, int flags, int fd,
-                        off_t offset) noexcept {
-#if defined(__i386__) || defined(__ARM_ARCH_3__) || defined(__ARM_EABI__) || \
-    defined(__m68k__) || defined(__sh__) ||                                  \
-    (defined(__hppa__) && !defined(__LP64__)) ||                             \
-    (defined(__mips__) && _MIPS_SIM == _MIPS_SIM_ABI32) ||                   \
-    (defined(__PPC__) && !defined(__PPC64__)) ||                             \
-    (defined(__riscv) && __riscv_xlen == 32) ||                              \
-    (defined(__s390__) && !defined(__s390x__)) ||                            \
-    (defined(__sparc__) && !defined(__arch64__))
-  // On these architectures, implement mmap with mmap2.
-  static int pagesize = 0;
-  if (pagesize == 0) {
-#if defined(__wasm__) || defined(__asmjs__)
-    pagesize = getpagesize();
-#else
-    pagesize = sysconf(_SC_PAGESIZE);
-#endif
-  }
-  if (offset < 0 || offset % pagesize != 0) {
-    errno = EINVAL;
-    return MAP_FAILED;
-  }
-#ifdef __BIONIC__
-  // SYS_mmap2 has problems on Android API level <= 16.
-  // Workaround by invoking __mmap2() instead.
-  return __mmap2(start, length, prot, flags, fd,
-                 static_cast<size_t>(offset / pagesize));
-#else
-  return reinterpret_cast<void*>(
-      syscall(SYS_mmap2, start, length, prot, flags, fd,
-              static_cast<unsigned long>(offset / pagesize)));  // NOLINT
-#endif
-#elif defined(__s390x__)
-  // On s390x, mmap() arguments are passed in memory.
-  unsigned long buf[6] = {reinterpret_cast<unsigned long>(start),  // NOLINT
-                          static_cast<unsigned long>(length),      // NOLINT
-                          static_cast<unsigned long>(prot),        // NOLINT
-                          static_cast<unsigned long>(flags),       // NOLINT
-                          static_cast<unsigned long>(fd),          // NOLINT
-                          static_cast<unsigned long>(offset)};     // NOLINT
-  return reinterpret_cast<void*>(syscall(SYS_mmap, buf));
-#elif defined(__x86_64__)
-// The x32 ABI has 32 bit longs, but the syscall interface is 64 bit.
-// We need to explicitly cast to an unsigned 64 bit type to avoid implicit
-// sign extension.  We can't cast pointers directly because those are
-// 32 bits, and gcc will dump ugly warnings about casting from a pointer
-// to an integer of a different size. We also need to make sure __off64_t
-// isn't truncated to 32-bits under x32.
-#define MMAP_SYSCALL_ARG(x) ((uint64_t)(uintptr_t)(x))
-  return reinterpret_cast<void*>(
-      syscall(SYS_mmap, MMAP_SYSCALL_ARG(start), MMAP_SYSCALL_ARG(length),
-              MMAP_SYSCALL_ARG(prot), MMAP_SYSCALL_ARG(flags),
-              MMAP_SYSCALL_ARG(fd), static_cast<uint64_t>(offset)));
-#undef MMAP_SYSCALL_ARG
-#else  // Remaining 64-bit aritectures.
-  static_assert(sizeof(unsigned long) == 8, "Platform is not 64-bit");
-  return reinterpret_cast<void*>(
-      syscall(SYS_mmap, start, length, prot, flags, fd, offset));
-#endif
-}
-
-inline int DirectMunmap(void* start, size_t length) {
-  return static_cast<int>(syscall(SYS_munmap, start, length));
-}
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#else  // !__linux__
-
-// For non-linux platforms where we have mmap, just dispatch directly to the
-// actual mmap()/munmap() methods.
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-inline void* DirectMmap(void* start, size_t length, int prot, int flags, int fd,
-                        off_t offset) {
-  return mmap(start, length, prot, flags, fd, offset);
-}
-
-inline int DirectMunmap(void* start, size_t length) {
-  return munmap(start, length);
-}
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // __linux__
-
-#endif  // ABSL_HAVE_MMAP
-
-#endif  // ABSL_BASE_INTERNAL_DIRECT_MMAP_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/dynamic_annotations.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/dynamic_annotations.h
deleted file mode 100644
index b23c5ec1..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/dynamic_annotations.h
+++ /dev/null
@@ -1,398 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-// This file defines dynamic annotations for use with dynamic analysis tool
-// such as valgrind, PIN, etc.
-//
-// Dynamic annotation is a source code annotation that affects the generated
-// code (that is, the annotation is not a comment). Each such annotation is
-// attached to a particular instruction and/or to a particular object (address)
-// in the program.
-//
-// The annotations that should be used by users are macros in all upper-case
-// (e.g., ANNOTATE_THREAD_NAME).
-//
-// Actual implementation of these macros may differ depending on the dynamic
-// analysis tool being used.
-//
-// This file supports the following configurations:
-// - Dynamic Annotations enabled (with static thread-safety warnings disabled).
-//   In this case, macros expand to functions implemented by Thread Sanitizer,
-//   when building with TSan. When not provided an external implementation,
-//   dynamic_annotations.cc provides no-op implementations.
-//
-// - Static Clang thread-safety warnings enabled.
-//   When building with a Clang compiler that supports thread-safety warnings,
-//   a subset of annotations can be statically-checked at compile-time. We
-//   expand these macros to static-inline functions that can be analyzed for
-//   thread-safety, but afterwards elided when building the final binary.
-//
-// - All annotations are disabled.
-//   If neither Dynamic Annotations nor Clang thread-safety warnings are
-//   enabled, then all annotation-macros expand to empty.
-
-#ifndef ABSL_BASE_INTERNAL_DYNAMIC_ANNOTATIONS_H_
-#define ABSL_BASE_INTERNAL_DYNAMIC_ANNOTATIONS_H_
-
-#include <stddef.h>
-
-#include "absl/base/config.h"
-
-// -------------------------------------------------------------------------
-// Decide which features are enabled
-
-#ifndef DYNAMIC_ANNOTATIONS_ENABLED
-#define DYNAMIC_ANNOTATIONS_ENABLED 0
-#endif
-
-#if defined(__clang__) && !defined(SWIG)
-#define ABSL_INTERNAL_IGNORE_READS_ATTRIBUTE_ENABLED 1
-#endif
-
-#if DYNAMIC_ANNOTATIONS_ENABLED != 0
-
-#define ABSL_INTERNAL_RACE_ANNOTATIONS_ENABLED 1
-#define ABSL_INTERNAL_READS_ANNOTATIONS_ENABLED 1
-#define ABSL_INTERNAL_WRITES_ANNOTATIONS_ENABLED 1
-#define ABSL_INTERNAL_ANNOTALYSIS_ENABLED 0
-#define ABSL_INTERNAL_READS_WRITES_ANNOTATIONS_ENABLED 1
-
-#else
-
-#define ABSL_INTERNAL_RACE_ANNOTATIONS_ENABLED 0
-#define ABSL_INTERNAL_READS_ANNOTATIONS_ENABLED 0
-#define ABSL_INTERNAL_WRITES_ANNOTATIONS_ENABLED 0
-
-// Clang provides limited support for static thread-safety analysis through a
-// feature called Annotalysis. We configure macro-definitions according to
-// whether Annotalysis support is available. When running in opt-mode, GCC
-// will issue a warning, if these attributes are compiled. Only include them
-// when compiling using Clang.
-
-// ANNOTALYSIS_ENABLED == 1 when IGNORE_READ_ATTRIBUTE_ENABLED == 1
-#define ABSL_INTERNAL_ANNOTALYSIS_ENABLED \
-  defined(ABSL_INTERNAL_IGNORE_READS_ATTRIBUTE_ENABLED)
-// Read/write annotations are enabled in Annotalysis mode; disabled otherwise.
-#define ABSL_INTERNAL_READS_WRITES_ANNOTATIONS_ENABLED \
-  ABSL_INTERNAL_ANNOTALYSIS_ENABLED
-#endif
-
-// Memory annotations are also made available to LLVM's Memory Sanitizer
-#if defined(ABSL_HAVE_MEMORY_SANITIZER) && !defined(__native_client__)
-#define ABSL_INTERNAL_MEMORY_ANNOTATIONS_ENABLED 1
-#endif
-
-#ifndef ABSL_INTERNAL_MEMORY_ANNOTATIONS_ENABLED
-#define ABSL_INTERNAL_MEMORY_ANNOTATIONS_ENABLED 0
-#endif
-
-#ifdef __cplusplus
-#define ABSL_INTERNAL_BEGIN_EXTERN_C extern "C" {
-#define ABSL_INTERNAL_END_EXTERN_C }  // extern "C"
-#define ABSL_INTERNAL_GLOBAL_SCOPED(F) ::F
-#define ABSL_INTERNAL_STATIC_INLINE inline
-#else
-#define ABSL_INTERNAL_BEGIN_EXTERN_C  // empty
-#define ABSL_INTERNAL_END_EXTERN_C    // empty
-#define ABSL_INTERNAL_GLOBAL_SCOPED(F) F
-#define ABSL_INTERNAL_STATIC_INLINE static inline
-#endif
-
-// -------------------------------------------------------------------------
-// Define race annotations.
-
-#if ABSL_INTERNAL_RACE_ANNOTATIONS_ENABLED == 1
-
-// -------------------------------------------------------------
-// Annotations that suppress errors. It is usually better to express the
-// program's synchronization using the other annotations, but these can be used
-// when all else fails.
-
-// Report that we may have a benign race at `pointer`, with size
-// "sizeof(*(pointer))". `pointer` must be a non-void* pointer. Insert at the
-// point where `pointer` has been allocated, preferably close to the point
-// where the race happens. See also ANNOTATE_BENIGN_RACE_STATIC.
-#define ANNOTATE_BENIGN_RACE(pointer, description)     \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateBenignRaceSized) \
-  (__FILE__, __LINE__, pointer, sizeof(*(pointer)), description)
-
-// Same as ANNOTATE_BENIGN_RACE(`address`, `description`), but applies to
-// the memory range [`address`, `address`+`size`).
-#define ANNOTATE_BENIGN_RACE_SIZED(address, size, description) \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateBenignRaceSized)         \
-  (__FILE__, __LINE__, address, size, description)
-
-// Enable (`enable`!=0) or disable (`enable`==0) race detection for all threads.
-// This annotation could be useful if you want to skip expensive race analysis
-// during some period of program execution, e.g. during initialization.
-#define ANNOTATE_ENABLE_RACE_DETECTION(enable)             \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateEnableRaceDetection) \
-  (__FILE__, __LINE__, enable)
-
-// -------------------------------------------------------------
-// Annotations useful for debugging.
-
-// Report the current thread `name` to a race detector.
-#define ANNOTATE_THREAD_NAME(name) \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateThreadName)(__FILE__, __LINE__, name)
-
-// -------------------------------------------------------------
-// Annotations useful when implementing locks. They are not normally needed by
-// modules that merely use locks. The `lock` argument is a pointer to the lock
-// object.
-
-// Report that a lock has been created at address `lock`.
-#define ANNOTATE_RWLOCK_CREATE(lock) \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateRWLockCreate)(__FILE__, __LINE__, lock)
-
-// Report that a linker initialized lock has been created at address `lock`.
-#ifdef ABSL_HAVE_THREAD_SANITIZER
-#define ANNOTATE_RWLOCK_CREATE_STATIC(lock)               \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateRWLockCreateStatic) \
-  (__FILE__, __LINE__, lock)
-#else
-#define ANNOTATE_RWLOCK_CREATE_STATIC(lock) ANNOTATE_RWLOCK_CREATE(lock)
-#endif
-
-// Report that the lock at address `lock` is about to be destroyed.
-#define ANNOTATE_RWLOCK_DESTROY(lock) \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateRWLockDestroy)(__FILE__, __LINE__, lock)
-
-// Report that the lock at address `lock` has been acquired.
-// `is_w`=1 for writer lock, `is_w`=0 for reader lock.
-#define ANNOTATE_RWLOCK_ACQUIRED(lock, is_w)          \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateRWLockAcquired) \
-  (__FILE__, __LINE__, lock, is_w)
-
-// Report that the lock at address `lock` is about to be released.
-// `is_w`=1 for writer lock, `is_w`=0 for reader lock.
-#define ANNOTATE_RWLOCK_RELEASED(lock, is_w)          \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateRWLockReleased) \
-  (__FILE__, __LINE__, lock, is_w)
-
-// Apply ANNOTATE_BENIGN_RACE_SIZED to a static variable `static_var`.
-#define ANNOTATE_BENIGN_RACE_STATIC(static_var, description)      \
-  namespace {                                                     \
-  class static_var##_annotator {                                  \
-   public:                                                        \
-    static_var##_annotator() {                                    \
-      ANNOTATE_BENIGN_RACE_SIZED(&static_var, sizeof(static_var), \
-                                 #static_var ": " description);   \
-    }                                                             \
-  };                                                              \
-  static static_var##_annotator the##static_var##_annotator;      \
-  }  // namespace
-
-#else  // ABSL_INTERNAL_RACE_ANNOTATIONS_ENABLED == 0
-
-#define ANNOTATE_RWLOCK_CREATE(lock)                            // empty
-#define ANNOTATE_RWLOCK_CREATE_STATIC(lock)                     // empty
-#define ANNOTATE_RWLOCK_DESTROY(lock)                           // empty
-#define ANNOTATE_RWLOCK_ACQUIRED(lock, is_w)                    // empty
-#define ANNOTATE_RWLOCK_RELEASED(lock, is_w)                    // empty
-#define ANNOTATE_BENIGN_RACE(address, description)              // empty
-#define ANNOTATE_BENIGN_RACE_SIZED(address, size, description)  // empty
-#define ANNOTATE_THREAD_NAME(name)                              // empty
-#define ANNOTATE_ENABLE_RACE_DETECTION(enable)                  // empty
-#define ANNOTATE_BENIGN_RACE_STATIC(static_var, description)    // empty
-
-#endif  // ABSL_INTERNAL_RACE_ANNOTATIONS_ENABLED
-
-// -------------------------------------------------------------------------
-// Define memory annotations.
-
-#if ABSL_INTERNAL_MEMORY_ANNOTATIONS_ENABLED == 1
-
-#include <sanitizer/msan_interface.h>
-
-#define ANNOTATE_MEMORY_IS_INITIALIZED(address, size) \
-  __msan_unpoison(address, size)
-
-#define ANNOTATE_MEMORY_IS_UNINITIALIZED(address, size) \
-  __msan_allocated_memory(address, size)
-
-#else  // ABSL_INTERNAL_MEMORY_ANNOTATIONS_ENABLED == 0
-
-#if DYNAMIC_ANNOTATIONS_ENABLED == 1
-#define ANNOTATE_MEMORY_IS_INITIALIZED(address, size) \
-  do {                                                \
-    (void)(address);                                  \
-    (void)(size);                                     \
-  } while (0)
-#define ANNOTATE_MEMORY_IS_UNINITIALIZED(address, size) \
-  do {                                                  \
-    (void)(address);                                    \
-    (void)(size);                                       \
-  } while (0)
-#else
-#define ANNOTATE_MEMORY_IS_INITIALIZED(address, size)    // empty
-#define ANNOTATE_MEMORY_IS_UNINITIALIZED(address, size)  // empty
-#endif
-
-#endif  // ABSL_INTERNAL_MEMORY_ANNOTATIONS_ENABLED
-
-// -------------------------------------------------------------------------
-// Define IGNORE_READS_BEGIN/_END attributes.
-
-#if defined(ABSL_INTERNAL_IGNORE_READS_ATTRIBUTE_ENABLED)
-
-#define ABSL_INTERNAL_IGNORE_READS_BEGIN_ATTRIBUTE \
-  __attribute((exclusive_lock_function("*")))
-#define ABSL_INTERNAL_IGNORE_READS_END_ATTRIBUTE \
-  __attribute((unlock_function("*")))
-
-#else  // !defined(ABSL_INTERNAL_IGNORE_READS_ATTRIBUTE_ENABLED)
-
-#define ABSL_INTERNAL_IGNORE_READS_BEGIN_ATTRIBUTE  // empty
-#define ABSL_INTERNAL_IGNORE_READS_END_ATTRIBUTE    // empty
-
-#endif  // defined(ABSL_INTERNAL_IGNORE_READS_ATTRIBUTE_ENABLED)
-
-// -------------------------------------------------------------------------
-// Define IGNORE_READS_BEGIN/_END annotations.
-
-#if ABSL_INTERNAL_READS_ANNOTATIONS_ENABLED == 1
-
-// Request the analysis tool to ignore all reads in the current thread until
-// ANNOTATE_IGNORE_READS_END is called. Useful to ignore intentional racey
-// reads, while still checking other reads and all writes.
-// See also ANNOTATE_UNPROTECTED_READ.
-#define ANNOTATE_IGNORE_READS_BEGIN() \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateIgnoreReadsBegin)(__FILE__, __LINE__)
-
-// Stop ignoring reads.
-#define ANNOTATE_IGNORE_READS_END() \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateIgnoreReadsEnd)(__FILE__, __LINE__)
-
-#elif defined(ABSL_INTERNAL_ANNOTALYSIS_ENABLED)
-
-// When Annotalysis is enabled without Dynamic Annotations, the use of
-// static-inline functions allows the annotations to be read at compile-time,
-// while still letting the compiler elide the functions from the final build.
-//
-// TODO(delesley) -- The exclusive lock here ignores writes as well, but
-// allows IGNORE_READS_AND_WRITES to work properly.
-
-#define ANNOTATE_IGNORE_READS_BEGIN() \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AbslInternalAnnotateIgnoreReadsBegin)()
-
-#define ANNOTATE_IGNORE_READS_END() \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AbslInternalAnnotateIgnoreReadsEnd)()
-
-#else
-
-#define ANNOTATE_IGNORE_READS_BEGIN()  // empty
-#define ANNOTATE_IGNORE_READS_END()    // empty
-
-#endif
-
-// -------------------------------------------------------------------------
-// Define IGNORE_WRITES_BEGIN/_END annotations.
-
-#if ABSL_INTERNAL_WRITES_ANNOTATIONS_ENABLED == 1
-
-// Similar to ANNOTATE_IGNORE_READS_BEGIN, but ignore writes instead.
-#define ANNOTATE_IGNORE_WRITES_BEGIN() \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateIgnoreWritesBegin)(__FILE__, __LINE__)
-
-// Stop ignoring writes.
-#define ANNOTATE_IGNORE_WRITES_END() \
-  ABSL_INTERNAL_GLOBAL_SCOPED(AnnotateIgnoreWritesEnd)(__FILE__, __LINE__)
-
-#else
-
-#define ANNOTATE_IGNORE_WRITES_BEGIN()  // empty
-#define ANNOTATE_IGNORE_WRITES_END()    // empty
-
-#endif
-
-// -------------------------------------------------------------------------
-// Define the ANNOTATE_IGNORE_READS_AND_WRITES_* annotations using the more
-// primitive annotations defined above.
-//
-//     Instead of doing
-//        ANNOTATE_IGNORE_READS_BEGIN();
-//        ... = x;
-//        ANNOTATE_IGNORE_READS_END();
-//     one can use
-//        ... = ANNOTATE_UNPROTECTED_READ(x);
-
-#if defined(ABSL_INTERNAL_READS_WRITES_ANNOTATIONS_ENABLED)
-
-// Start ignoring all memory accesses (both reads and writes).
-#define ANNOTATE_IGNORE_READS_AND_WRITES_BEGIN() \
-  do {                                           \
-    ANNOTATE_IGNORE_READS_BEGIN();               \
-    ANNOTATE_IGNORE_WRITES_BEGIN();              \
-  } while (0)
-
-// Stop ignoring both reads and writes.
-#define ANNOTATE_IGNORE_READS_AND_WRITES_END() \
-  do {                                         \
-    ANNOTATE_IGNORE_WRITES_END();              \
-    ANNOTATE_IGNORE_READS_END();               \
-  } while (0)
-
-#ifdef __cplusplus
-// ANNOTATE_UNPROTECTED_READ is the preferred way to annotate racey reads.
-#define ANNOTATE_UNPROTECTED_READ(x) \
-  absl::base_internal::AnnotateUnprotectedRead(x)
-
-#endif
-
-#else
-
-#define ANNOTATE_IGNORE_READS_AND_WRITES_BEGIN()  // empty
-#define ANNOTATE_IGNORE_READS_AND_WRITES_END()    // empty
-#define ANNOTATE_UNPROTECTED_READ(x) (x)
-
-#endif
-
-// -------------------------------------------------------------------------
-// Address sanitizer annotations
-
-#ifdef ABSL_HAVE_ADDRESS_SANITIZER
-// Describe the current state of a contiguous container such as e.g.
-// std::vector or std::string. For more details see
-// sanitizer/common_interface_defs.h, which is provided by the compiler.
-#include <sanitizer/common_interface_defs.h>
-
-#define ANNOTATE_CONTIGUOUS_CONTAINER(beg, end, old_mid, new_mid) \
-  __sanitizer_annotate_contiguous_container(beg, end, old_mid, new_mid)
-#define ADDRESS_SANITIZER_REDZONE(name)    \
-  struct {                                 \
-    char x[8] __attribute__((aligned(8))); \
-  } name
-
-#else
-
-#define ANNOTATE_CONTIGUOUS_CONTAINER(beg, end, old_mid, new_mid)
-#define ADDRESS_SANITIZER_REDZONE(name) static_assert(true, "")
-
-#endif  // ABSL_HAVE_ADDRESS_SANITIZER
-
-// -------------------------------------------------------------------------
-// Undefine the macros intended only for this file.
-
-#undef ABSL_INTERNAL_RACE_ANNOTATIONS_ENABLED
-#undef ABSL_INTERNAL_MEMORY_ANNOTATIONS_ENABLED
-#undef ABSL_INTERNAL_READS_ANNOTATIONS_ENABLED
-#undef ABSL_INTERNAL_WRITES_ANNOTATIONS_ENABLED
-#undef ABSL_INTERNAL_ANNOTALYSIS_ENABLED
-#undef ABSL_INTERNAL_READS_WRITES_ANNOTATIONS_ENABLED
-#undef ABSL_INTERNAL_BEGIN_EXTERN_C
-#undef ABSL_INTERNAL_END_EXTERN_C
-#undef ABSL_INTERNAL_STATIC_INLINE
-
-#endif  // ABSL_BASE_INTERNAL_DYNAMIC_ANNOTATIONS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/endian.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/endian.h
deleted file mode 100644
index 943f3d97..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/endian.h
+++ /dev/null
@@ -1,283 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_BASE_INTERNAL_ENDIAN_H_
-#define ABSL_BASE_INTERNAL_ENDIAN_H_
-
-#include <cstdint>
-#include <cstdlib>
-
-#include "absl/base/casts.h"
-#include "absl/base/config.h"
-#include "absl/base/internal/unaligned_access.h"
-#include "absl/base/nullability.h"
-#include "absl/base/port.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-inline uint64_t gbswap_64(uint64_t host_int) {
-#if ABSL_HAVE_BUILTIN(__builtin_bswap64) || defined(__GNUC__)
-  return __builtin_bswap64(host_int);
-#elif defined(_MSC_VER)
-  return _byteswap_uint64(host_int);
-#else
-  return (((host_int & uint64_t{0xFF}) << 56) |
-          ((host_int & uint64_t{0xFF00}) << 40) |
-          ((host_int & uint64_t{0xFF0000}) << 24) |
-          ((host_int & uint64_t{0xFF000000}) << 8) |
-          ((host_int & uint64_t{0xFF00000000}) >> 8) |
-          ((host_int & uint64_t{0xFF0000000000}) >> 24) |
-          ((host_int & uint64_t{0xFF000000000000}) >> 40) |
-          ((host_int & uint64_t{0xFF00000000000000}) >> 56));
-#endif
-}
-
-inline uint32_t gbswap_32(uint32_t host_int) {
-#if ABSL_HAVE_BUILTIN(__builtin_bswap32) || defined(__GNUC__)
-  return __builtin_bswap32(host_int);
-#elif defined(_MSC_VER)
-  return _byteswap_ulong(host_int);
-#else
-  return (((host_int & uint32_t{0xFF}) << 24) |
-          ((host_int & uint32_t{0xFF00}) << 8) |
-          ((host_int & uint32_t{0xFF0000}) >> 8) |
-          ((host_int & uint32_t{0xFF000000}) >> 24));
-#endif
-}
-
-inline uint16_t gbswap_16(uint16_t host_int) {
-#if ABSL_HAVE_BUILTIN(__builtin_bswap16) || defined(__GNUC__)
-  return __builtin_bswap16(host_int);
-#elif defined(_MSC_VER)
-  return _byteswap_ushort(host_int);
-#else
-  return (((host_int & uint16_t{0xFF}) << 8) |
-          ((host_int & uint16_t{0xFF00}) >> 8));
-#endif
-}
-
-#ifdef ABSL_IS_LITTLE_ENDIAN
-
-// Portable definitions for htonl (host-to-network) and friends on little-endian
-// architectures.
-inline uint16_t ghtons(uint16_t x) { return gbswap_16(x); }
-inline uint32_t ghtonl(uint32_t x) { return gbswap_32(x); }
-inline uint64_t ghtonll(uint64_t x) { return gbswap_64(x); }
-
-#elif defined ABSL_IS_BIG_ENDIAN
-
-// Portable definitions for htonl (host-to-network) etc on big-endian
-// architectures. These definitions are simpler since the host byte order is the
-// same as network byte order.
-inline uint16_t ghtons(uint16_t x) { return x; }
-inline uint32_t ghtonl(uint32_t x) { return x; }
-inline uint64_t ghtonll(uint64_t x) { return x; }
-
-#else
-#error \
-    "Unsupported byte order: Either ABSL_IS_BIG_ENDIAN or " \
-       "ABSL_IS_LITTLE_ENDIAN must be defined"
-#endif  // byte order
-
-inline uint16_t gntohs(uint16_t x) { return ghtons(x); }
-inline uint32_t gntohl(uint32_t x) { return ghtonl(x); }
-inline uint64_t gntohll(uint64_t x) { return ghtonll(x); }
-
-// Utilities to convert numbers between the current hosts's native byte
-// order and little-endian byte order
-//
-// Load/Store methods are alignment safe
-namespace little_endian {
-// Conversion functions.
-#ifdef ABSL_IS_LITTLE_ENDIAN
-
-inline uint16_t FromHost16(uint16_t x) { return x; }
-inline uint16_t ToHost16(uint16_t x) { return x; }
-
-inline uint32_t FromHost32(uint32_t x) { return x; }
-inline uint32_t ToHost32(uint32_t x) { return x; }
-
-inline uint64_t FromHost64(uint64_t x) { return x; }
-inline uint64_t ToHost64(uint64_t x) { return x; }
-
-inline constexpr bool IsLittleEndian() { return true; }
-
-#elif defined ABSL_IS_BIG_ENDIAN
-
-inline uint16_t FromHost16(uint16_t x) { return gbswap_16(x); }
-inline uint16_t ToHost16(uint16_t x) { return gbswap_16(x); }
-
-inline uint32_t FromHost32(uint32_t x) { return gbswap_32(x); }
-inline uint32_t ToHost32(uint32_t x) { return gbswap_32(x); }
-
-inline uint64_t FromHost64(uint64_t x) { return gbswap_64(x); }
-inline uint64_t ToHost64(uint64_t x) { return gbswap_64(x); }
-
-inline constexpr bool IsLittleEndian() { return false; }
-
-#endif /* ENDIAN */
-
-inline uint8_t FromHost(uint8_t x) { return x; }
-inline uint16_t FromHost(uint16_t x) { return FromHost16(x); }
-inline uint32_t FromHost(uint32_t x) { return FromHost32(x); }
-inline uint64_t FromHost(uint64_t x) { return FromHost64(x); }
-inline uint8_t ToHost(uint8_t x) { return x; }
-inline uint16_t ToHost(uint16_t x) { return ToHost16(x); }
-inline uint32_t ToHost(uint32_t x) { return ToHost32(x); }
-inline uint64_t ToHost(uint64_t x) { return ToHost64(x); }
-
-inline int8_t FromHost(int8_t x) { return x; }
-inline int16_t FromHost(int16_t x) {
-  return bit_cast<int16_t>(FromHost16(bit_cast<uint16_t>(x)));
-}
-inline int32_t FromHost(int32_t x) {
-  return bit_cast<int32_t>(FromHost32(bit_cast<uint32_t>(x)));
-}
-inline int64_t FromHost(int64_t x) {
-  return bit_cast<int64_t>(FromHost64(bit_cast<uint64_t>(x)));
-}
-inline int8_t ToHost(int8_t x) { return x; }
-inline int16_t ToHost(int16_t x) {
-  return bit_cast<int16_t>(ToHost16(bit_cast<uint16_t>(x)));
-}
-inline int32_t ToHost(int32_t x) {
-  return bit_cast<int32_t>(ToHost32(bit_cast<uint32_t>(x)));
-}
-inline int64_t ToHost(int64_t x) {
-  return bit_cast<int64_t>(ToHost64(bit_cast<uint64_t>(x)));
-}
-
-// Functions to do unaligned loads and stores in little-endian order.
-inline uint16_t Load16(absl::Nonnull<const void *> p) {
-  return ToHost16(ABSL_INTERNAL_UNALIGNED_LOAD16(p));
-}
-
-inline void Store16(absl::Nonnull<void *> p, uint16_t v) {
-  ABSL_INTERNAL_UNALIGNED_STORE16(p, FromHost16(v));
-}
-
-inline uint32_t Load32(absl::Nonnull<const void *> p) {
-  return ToHost32(ABSL_INTERNAL_UNALIGNED_LOAD32(p));
-}
-
-inline void Store32(absl::Nonnull<void *> p, uint32_t v) {
-  ABSL_INTERNAL_UNALIGNED_STORE32(p, FromHost32(v));
-}
-
-inline uint64_t Load64(absl::Nonnull<const void *> p) {
-  return ToHost64(ABSL_INTERNAL_UNALIGNED_LOAD64(p));
-}
-
-inline void Store64(absl::Nonnull<void *> p, uint64_t v) {
-  ABSL_INTERNAL_UNALIGNED_STORE64(p, FromHost64(v));
-}
-
-}  // namespace little_endian
-
-// Utilities to convert numbers between the current hosts's native byte
-// order and big-endian byte order (same as network byte order)
-//
-// Load/Store methods are alignment safe
-namespace big_endian {
-#ifdef ABSL_IS_LITTLE_ENDIAN
-
-inline uint16_t FromHost16(uint16_t x) { return gbswap_16(x); }
-inline uint16_t ToHost16(uint16_t x) { return gbswap_16(x); }
-
-inline uint32_t FromHost32(uint32_t x) { return gbswap_32(x); }
-inline uint32_t ToHost32(uint32_t x) { return gbswap_32(x); }
-
-inline uint64_t FromHost64(uint64_t x) { return gbswap_64(x); }
-inline uint64_t ToHost64(uint64_t x) { return gbswap_64(x); }
-
-inline constexpr bool IsLittleEndian() { return true; }
-
-#elif defined ABSL_IS_BIG_ENDIAN
-
-inline uint16_t FromHost16(uint16_t x) { return x; }
-inline uint16_t ToHost16(uint16_t x) { return x; }
-
-inline uint32_t FromHost32(uint32_t x) { return x; }
-inline uint32_t ToHost32(uint32_t x) { return x; }
-
-inline uint64_t FromHost64(uint64_t x) { return x; }
-inline uint64_t ToHost64(uint64_t x) { return x; }
-
-inline constexpr bool IsLittleEndian() { return false; }
-
-#endif /* ENDIAN */
-
-inline uint8_t FromHost(uint8_t x) { return x; }
-inline uint16_t FromHost(uint16_t x) { return FromHost16(x); }
-inline uint32_t FromHost(uint32_t x) { return FromHost32(x); }
-inline uint64_t FromHost(uint64_t x) { return FromHost64(x); }
-inline uint8_t ToHost(uint8_t x) { return x; }
-inline uint16_t ToHost(uint16_t x) { return ToHost16(x); }
-inline uint32_t ToHost(uint32_t x) { return ToHost32(x); }
-inline uint64_t ToHost(uint64_t x) { return ToHost64(x); }
-
-inline int8_t FromHost(int8_t x) { return x; }
-inline int16_t FromHost(int16_t x) {
-  return bit_cast<int16_t>(FromHost16(bit_cast<uint16_t>(x)));
-}
-inline int32_t FromHost(int32_t x) {
-  return bit_cast<int32_t>(FromHost32(bit_cast<uint32_t>(x)));
-}
-inline int64_t FromHost(int64_t x) {
-  return bit_cast<int64_t>(FromHost64(bit_cast<uint64_t>(x)));
-}
-inline int8_t ToHost(int8_t x) { return x; }
-inline int16_t ToHost(int16_t x) {
-  return bit_cast<int16_t>(ToHost16(bit_cast<uint16_t>(x)));
-}
-inline int32_t ToHost(int32_t x) {
-  return bit_cast<int32_t>(ToHost32(bit_cast<uint32_t>(x)));
-}
-inline int64_t ToHost(int64_t x) {
-  return bit_cast<int64_t>(ToHost64(bit_cast<uint64_t>(x)));
-}
-
-// Functions to do unaligned loads and stores in big-endian order.
-inline uint16_t Load16(absl::Nonnull<const void *> p) {
-  return ToHost16(ABSL_INTERNAL_UNALIGNED_LOAD16(p));
-}
-
-inline void Store16(absl::Nonnull<void *> p, uint16_t v) {
-  ABSL_INTERNAL_UNALIGNED_STORE16(p, FromHost16(v));
-}
-
-inline uint32_t Load32(absl::Nonnull<const void *> p) {
-  return ToHost32(ABSL_INTERNAL_UNALIGNED_LOAD32(p));
-}
-
-inline void Store32(absl::Nonnull<void *>p, uint32_t v) {
-  ABSL_INTERNAL_UNALIGNED_STORE32(p, FromHost32(v));
-}
-
-inline uint64_t Load64(absl::Nonnull<const void *> p) {
-  return ToHost64(ABSL_INTERNAL_UNALIGNED_LOAD64(p));
-}
-
-inline void Store64(absl::Nonnull<void *> p, uint64_t v) {
-  ABSL_INTERNAL_UNALIGNED_STORE64(p, FromHost64(v));
-}
-
-}  // namespace big_endian
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_ENDIAN_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/errno_saver.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/errno_saver.h
deleted file mode 100644
index 251de510..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/errno_saver.h
+++ /dev/null
@@ -1,43 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_BASE_INTERNAL_ERRNO_SAVER_H_
-#define ABSL_BASE_INTERNAL_ERRNO_SAVER_H_
-
-#include <cerrno>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-// `ErrnoSaver` captures the value of `errno` upon construction and restores it
-// upon deletion.  It is used in low-level code and must be super fast.  Do not
-// add instrumentation, even in debug modes.
-class ErrnoSaver {
- public:
-  ErrnoSaver() : saved_errno_(errno) {}
-  ~ErrnoSaver() { errno = saved_errno_; }
-  int operator()() const { return saved_errno_; }
-
- private:
-  const int saved_errno_;
-};
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_ERRNO_SAVER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/exception_safety_testing.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/exception_safety_testing.h
deleted file mode 100644
index c1061544..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/exception_safety_testing.h
+++ /dev/null
@@ -1,1109 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-// Utilities for testing exception-safety
-
-#ifndef ABSL_BASE_INTERNAL_EXCEPTION_SAFETY_TESTING_H_
-#define ABSL_BASE_INTERNAL_EXCEPTION_SAFETY_TESTING_H_
-
-#include "absl/base/config.h"
-
-#ifdef ABSL_HAVE_EXCEPTIONS
-
-#include <cstddef>
-#include <cstdint>
-#include <functional>
-#include <initializer_list>
-#include <iosfwd>
-#include <string>
-#include <tuple>
-#include <unordered_map>
-
-#include "gtest/gtest.h"
-#include "absl/base/internal/pretty_function.h"
-#include "absl/memory/memory.h"
-#include "absl/meta/type_traits.h"
-#include "absl/strings/string_view.h"
-#include "absl/strings/substitute.h"
-#include "absl/utility/utility.h"
-
-namespace testing {
-
-enum class TypeSpec;
-enum class AllocSpec;
-
-constexpr TypeSpec operator|(TypeSpec a, TypeSpec b) {
-  using T = absl::underlying_type_t<TypeSpec>;
-  return static_cast<TypeSpec>(static_cast<T>(a) | static_cast<T>(b));
-}
-
-constexpr TypeSpec operator&(TypeSpec a, TypeSpec b) {
-  using T = absl::underlying_type_t<TypeSpec>;
-  return static_cast<TypeSpec>(static_cast<T>(a) & static_cast<T>(b));
-}
-
-constexpr AllocSpec operator|(AllocSpec a, AllocSpec b) {
-  using T = absl::underlying_type_t<AllocSpec>;
-  return static_cast<AllocSpec>(static_cast<T>(a) | static_cast<T>(b));
-}
-
-constexpr AllocSpec operator&(AllocSpec a, AllocSpec b) {
-  using T = absl::underlying_type_t<AllocSpec>;
-  return static_cast<AllocSpec>(static_cast<T>(a) & static_cast<T>(b));
-}
-
-namespace exceptions_internal {
-
-std::string GetSpecString(TypeSpec);
-std::string GetSpecString(AllocSpec);
-
-struct NoThrowTag {};
-struct StrongGuaranteeTagType {};
-
-// A simple exception class.  We throw this so that test code can catch
-// exceptions specifically thrown by ThrowingValue.
-class TestException {
- public:
-  explicit TestException(absl::string_view msg) : msg_(msg) {}
-  virtual ~TestException() {}
-  virtual const char* what() const noexcept { return msg_.c_str(); }
-
- private:
-  std::string msg_;
-};
-
-// TestBadAllocException exists because allocation functions must throw an
-// exception which can be caught by a handler of std::bad_alloc.  We use a child
-// class of std::bad_alloc so we can customise the error message, and also
-// derive from TestException so we don't accidentally end up catching an actual
-// bad_alloc exception in TestExceptionSafety.
-class TestBadAllocException : public std::bad_alloc, public TestException {
- public:
-  explicit TestBadAllocException(absl::string_view msg) : TestException(msg) {}
-  using TestException::what;
-};
-
-extern int countdown;
-
-// Allows the countdown variable to be set manually (defaulting to the initial
-// value of 0)
-inline void SetCountdown(int i = 0) { countdown = i; }
-// Sets the countdown to the terminal value -1
-inline void UnsetCountdown() { SetCountdown(-1); }
-
-void MaybeThrow(absl::string_view msg, bool throw_bad_alloc = false);
-
-testing::AssertionResult FailureMessage(const TestException& e,
-                                        int countdown) noexcept;
-
-struct TrackedAddress {
-  bool is_alive;
-  std::string description;
-};
-
-// Inspects the constructions and destructions of anything inheriting from
-// TrackedObject. This allows us to safely "leak" TrackedObjects, as
-// ConstructorTracker will destroy everything left over in its destructor.
-class ConstructorTracker {
- public:
-  explicit ConstructorTracker(int count) : countdown_(count) {
-    assert(current_tracker_instance_ == nullptr);
-    current_tracker_instance_ = this;
-  }
-
-  ~ConstructorTracker() {
-    assert(current_tracker_instance_ == this);
-    current_tracker_instance_ = nullptr;
-
-    for (auto& it : address_map_) {
-      void* address = it.first;
-      TrackedAddress& tracked_address = it.second;
-      if (tracked_address.is_alive) {
-        ADD_FAILURE() << ErrorMessage(address, tracked_address.description,
-                                      countdown_, "Object was not destroyed.");
-      }
-    }
-  }
-
-  static void ObjectConstructed(void* address, std::string description) {
-    if (!CurrentlyTracking()) return;
-
-    TrackedAddress& tracked_address =
-        current_tracker_instance_->address_map_[address];
-    if (tracked_address.is_alive) {
-      ADD_FAILURE() << ErrorMessage(
-          address, tracked_address.description,
-          current_tracker_instance_->countdown_,
-          "Object was re-constructed. Current object was constructed by " +
-              description);
-    }
-    tracked_address = {true, std::move(description)};
-  }
-
-  static void ObjectDestructed(void* address) {
-    if (!CurrentlyTracking()) return;
-
-    auto it = current_tracker_instance_->address_map_.find(address);
-    // Not tracked. Ignore.
-    if (it == current_tracker_instance_->address_map_.end()) return;
-
-    TrackedAddress& tracked_address = it->second;
-    if (!tracked_address.is_alive) {
-      ADD_FAILURE() << ErrorMessage(address, tracked_address.description,
-                                    current_tracker_instance_->countdown_,
-                                    "Object was re-destroyed.");
-    }
-    tracked_address.is_alive = false;
-  }
-
- private:
-  static bool CurrentlyTracking() {
-    return current_tracker_instance_ != nullptr;
-  }
-
-  static std::string ErrorMessage(void* address,
-                                  const std::string& address_description,
-                                  int countdown,
-                                  const std::string& error_description) {
-    return absl::Substitute(
-        "With coundtown at $0:\n"
-        "  $1\n"
-        "  Object originally constructed by $2\n"
-        "  Object address: $3\n",
-        countdown, error_description, address_description, address);
-  }
-
-  std::unordered_map<void*, TrackedAddress> address_map_;
-  int countdown_;
-
-  static ConstructorTracker* current_tracker_instance_;
-};
-
-class TrackedObject {
- public:
-  TrackedObject(const TrackedObject&) = delete;
-  TrackedObject(TrackedObject&&) = delete;
-
- protected:
-  explicit TrackedObject(std::string description) {
-    ConstructorTracker::ObjectConstructed(this, std::move(description));
-  }
-
-  ~TrackedObject() noexcept { ConstructorTracker::ObjectDestructed(this); }
-};
-}  // namespace exceptions_internal
-
-extern exceptions_internal::NoThrowTag nothrow_ctor;
-
-extern exceptions_internal::StrongGuaranteeTagType strong_guarantee;
-
-// A test class which is convertible to bool.  The conversion can be
-// instrumented to throw at a controlled time.
-class ThrowingBool {
- public:
-  ThrowingBool(bool b) noexcept : b_(b) {}  // NOLINT(runtime/explicit)
-  operator bool() const {                   // NOLINT
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return b_;
-  }
-
- private:
-  bool b_;
-};
-
-/*
- * Configuration enum for the ThrowingValue type that defines behavior for the
- * lifetime of the instance. Use testing::nothrow_ctor to prevent the integer
- * constructor from throwing.
- *
- * kEverythingThrows: Every operation can throw an exception
- * kNoThrowCopy: Copy construction and copy assignment will not throw
- * kNoThrowMove: Move construction and move assignment will not throw
- * kNoThrowNew: Overloaded operators new and new[] will not throw
- */
-enum class TypeSpec {
-  kEverythingThrows = 0,
-  kNoThrowCopy = 1,
-  kNoThrowMove = 1 << 1,
-  kNoThrowNew = 1 << 2,
-};
-
-/*
- * A testing class instrumented to throw an exception at a controlled time.
- *
- * ThrowingValue implements a slightly relaxed version of the Regular concept --
- * that is it's a value type with the expected semantics.  It also implements
- * arithmetic operations.  It doesn't implement member and pointer operators
- * like operator-> or operator[].
- *
- * ThrowingValue can be instrumented to have certain operations be noexcept by
- * using compile-time bitfield template arguments.  That is, to make an
- * ThrowingValue which has noexcept move construction/assignment and noexcept
- * copy construction/assignment, use the following:
- *   ThrowingValue<testing::kNoThrowMove | testing::kNoThrowCopy> my_thrwr{val};
- */
-template <TypeSpec Spec = TypeSpec::kEverythingThrows>
-class ThrowingValue : private exceptions_internal::TrackedObject {
-  static constexpr bool IsSpecified(TypeSpec spec) {
-    return static_cast<bool>(Spec & spec);
-  }
-
-  static constexpr int kDefaultValue = 0;
-  static constexpr int kBadValue = 938550620;
-
- public:
-  ThrowingValue() : TrackedObject(GetInstanceString(kDefaultValue)) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    dummy_ = kDefaultValue;
-  }
-
-  ThrowingValue(const ThrowingValue& other) noexcept(
-      IsSpecified(TypeSpec::kNoThrowCopy))
-      : TrackedObject(GetInstanceString(other.dummy_)) {
-    if (!IsSpecified(TypeSpec::kNoThrowCopy)) {
-      exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    }
-    dummy_ = other.dummy_;
-  }
-
-  ThrowingValue(ThrowingValue&& other) noexcept(
-      IsSpecified(TypeSpec::kNoThrowMove))
-      : TrackedObject(GetInstanceString(other.dummy_)) {
-    if (!IsSpecified(TypeSpec::kNoThrowMove)) {
-      exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    }
-    dummy_ = other.dummy_;
-  }
-
-  explicit ThrowingValue(int i) : TrackedObject(GetInstanceString(i)) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    dummy_ = i;
-  }
-
-  ThrowingValue(int i, exceptions_internal::NoThrowTag) noexcept
-      : TrackedObject(GetInstanceString(i)), dummy_(i) {}
-
-  // absl expects nothrow destructors
-  ~ThrowingValue() noexcept = default;
-
-  ThrowingValue& operator=(const ThrowingValue& other) noexcept(
-      IsSpecified(TypeSpec::kNoThrowCopy)) {
-    dummy_ = kBadValue;
-    if (!IsSpecified(TypeSpec::kNoThrowCopy)) {
-      exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    }
-    dummy_ = other.dummy_;
-    return *this;
-  }
-
-  ThrowingValue& operator=(ThrowingValue&& other) noexcept(
-      IsSpecified(TypeSpec::kNoThrowMove)) {
-    dummy_ = kBadValue;
-    if (!IsSpecified(TypeSpec::kNoThrowMove)) {
-      exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    }
-    dummy_ = other.dummy_;
-    return *this;
-  }
-
-  // Arithmetic Operators
-  ThrowingValue operator+(const ThrowingValue& other) const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return ThrowingValue(dummy_ + other.dummy_, nothrow_ctor);
-  }
-
-  ThrowingValue operator+() const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return ThrowingValue(dummy_, nothrow_ctor);
-  }
-
-  ThrowingValue operator-(const ThrowingValue& other) const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return ThrowingValue(dummy_ - other.dummy_, nothrow_ctor);
-  }
-
-  ThrowingValue operator-() const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return ThrowingValue(-dummy_, nothrow_ctor);
-  }
-
-  ThrowingValue& operator++() {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    ++dummy_;
-    return *this;
-  }
-
-  ThrowingValue operator++(int) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    auto out = ThrowingValue(dummy_, nothrow_ctor);
-    ++dummy_;
-    return out;
-  }
-
-  ThrowingValue& operator--() {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    --dummy_;
-    return *this;
-  }
-
-  ThrowingValue operator--(int) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    auto out = ThrowingValue(dummy_, nothrow_ctor);
-    --dummy_;
-    return out;
-  }
-
-  ThrowingValue operator*(const ThrowingValue& other) const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return ThrowingValue(dummy_ * other.dummy_, nothrow_ctor);
-  }
-
-  ThrowingValue operator/(const ThrowingValue& other) const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return ThrowingValue(dummy_ / other.dummy_, nothrow_ctor);
-  }
-
-  ThrowingValue operator%(const ThrowingValue& other) const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return ThrowingValue(dummy_ % other.dummy_, nothrow_ctor);
-  }
-
-  ThrowingValue operator<<(int shift) const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return ThrowingValue(dummy_ << shift, nothrow_ctor);
-  }
-
-  ThrowingValue operator>>(int shift) const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return ThrowingValue(dummy_ >> shift, nothrow_ctor);
-  }
-
-  // Comparison Operators
-  // NOTE: We use `ThrowingBool` instead of `bool` because most STL
-  // types/containers requires T to be convertible to bool.
-  friend ThrowingBool operator==(const ThrowingValue& a,
-                                 const ThrowingValue& b) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return a.dummy_ == b.dummy_;
-  }
-  friend ThrowingBool operator!=(const ThrowingValue& a,
-                                 const ThrowingValue& b) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return a.dummy_ != b.dummy_;
-  }
-  friend ThrowingBool operator<(const ThrowingValue& a,
-                                const ThrowingValue& b) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return a.dummy_ < b.dummy_;
-  }
-  friend ThrowingBool operator<=(const ThrowingValue& a,
-                                 const ThrowingValue& b) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return a.dummy_ <= b.dummy_;
-  }
-  friend ThrowingBool operator>(const ThrowingValue& a,
-                                const ThrowingValue& b) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return a.dummy_ > b.dummy_;
-  }
-  friend ThrowingBool operator>=(const ThrowingValue& a,
-                                 const ThrowingValue& b) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return a.dummy_ >= b.dummy_;
-  }
-
-  // Logical Operators
-  ThrowingBool operator!() const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return !dummy_;
-  }
-
-  ThrowingBool operator&&(const ThrowingValue& other) const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return dummy_ && other.dummy_;
-  }
-
-  ThrowingBool operator||(const ThrowingValue& other) const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return dummy_ || other.dummy_;
-  }
-
-  // Bitwise Logical Operators
-  ThrowingValue operator~() const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return ThrowingValue(~dummy_, nothrow_ctor);
-  }
-
-  ThrowingValue operator&(const ThrowingValue& other) const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return ThrowingValue(dummy_ & other.dummy_, nothrow_ctor);
-  }
-
-  ThrowingValue operator|(const ThrowingValue& other) const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return ThrowingValue(dummy_ | other.dummy_, nothrow_ctor);
-  }
-
-  ThrowingValue operator^(const ThrowingValue& other) const {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return ThrowingValue(dummy_ ^ other.dummy_, nothrow_ctor);
-  }
-
-  // Compound Assignment operators
-  ThrowingValue& operator+=(const ThrowingValue& other) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    dummy_ += other.dummy_;
-    return *this;
-  }
-
-  ThrowingValue& operator-=(const ThrowingValue& other) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    dummy_ -= other.dummy_;
-    return *this;
-  }
-
-  ThrowingValue& operator*=(const ThrowingValue& other) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    dummy_ *= other.dummy_;
-    return *this;
-  }
-
-  ThrowingValue& operator/=(const ThrowingValue& other) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    dummy_ /= other.dummy_;
-    return *this;
-  }
-
-  ThrowingValue& operator%=(const ThrowingValue& other) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    dummy_ %= other.dummy_;
-    return *this;
-  }
-
-  ThrowingValue& operator&=(const ThrowingValue& other) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    dummy_ &= other.dummy_;
-    return *this;
-  }
-
-  ThrowingValue& operator|=(const ThrowingValue& other) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    dummy_ |= other.dummy_;
-    return *this;
-  }
-
-  ThrowingValue& operator^=(const ThrowingValue& other) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    dummy_ ^= other.dummy_;
-    return *this;
-  }
-
-  ThrowingValue& operator<<=(int shift) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    dummy_ <<= shift;
-    return *this;
-  }
-
-  ThrowingValue& operator>>=(int shift) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    dummy_ >>= shift;
-    return *this;
-  }
-
-  // Pointer operators
-  void operator&() const = delete;  // NOLINT(runtime/operator)
-
-  // Stream operators
-  friend std::ostream& operator<<(std::ostream& os, const ThrowingValue& tv) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return os << GetInstanceString(tv.dummy_);
-  }
-
-  friend std::istream& operator>>(std::istream& is, const ThrowingValue&) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    return is;
-  }
-
-  // Memory management operators
-  static void* operator new(size_t s) noexcept(
-      IsSpecified(TypeSpec::kNoThrowNew)) {
-    if (!IsSpecified(TypeSpec::kNoThrowNew)) {
-      exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION, true);
-    }
-    return ::operator new(s);
-  }
-
-  static void* operator new[](size_t s) noexcept(
-      IsSpecified(TypeSpec::kNoThrowNew)) {
-    if (!IsSpecified(TypeSpec::kNoThrowNew)) {
-      exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION, true);
-    }
-    return ::operator new[](s);
-  }
-
-  template <typename... Args>
-  static void* operator new(size_t s, Args&&... args) noexcept(
-      IsSpecified(TypeSpec::kNoThrowNew)) {
-    if (!IsSpecified(TypeSpec::kNoThrowNew)) {
-      exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION, true);
-    }
-    return ::operator new(s, std::forward<Args>(args)...);
-  }
-
-  template <typename... Args>
-  static void* operator new[](size_t s, Args&&... args) noexcept(
-      IsSpecified(TypeSpec::kNoThrowNew)) {
-    if (!IsSpecified(TypeSpec::kNoThrowNew)) {
-      exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION, true);
-    }
-    return ::operator new[](s, std::forward<Args>(args)...);
-  }
-
-  // Abseil doesn't support throwing overloaded operator delete.  These are
-  // provided so a throwing operator-new can clean up after itself.
-  void operator delete(void* p) noexcept { ::operator delete(p); }
-
-  template <typename... Args>
-  void operator delete(void* p, Args&&... args) noexcept {
-    ::operator delete(p, std::forward<Args>(args)...);
-  }
-
-  void operator delete[](void* p) noexcept { return ::operator delete[](p); }
-
-  template <typename... Args>
-  void operator delete[](void* p, Args&&... args) noexcept {
-    return ::operator delete[](p, std::forward<Args>(args)...);
-  }
-
-  // Non-standard access to the actual contained value.  No need for this to
-  // throw.
-  int& Get() noexcept { return dummy_; }
-  const int& Get() const noexcept { return dummy_; }
-
- private:
-  static std::string GetInstanceString(int dummy) {
-    return absl::StrCat("ThrowingValue<",
-                        exceptions_internal::GetSpecString(Spec), ">(", dummy,
-                        ")");
-  }
-
-  int dummy_;
-};
-// While not having to do with exceptions, explicitly delete comma operator, to
-// make sure we don't use it on user-supplied types.
-template <TypeSpec Spec, typename T>
-void operator,(const ThrowingValue<Spec>&, T&&) = delete;
-template <TypeSpec Spec, typename T>
-void operator,(T&&, const ThrowingValue<Spec>&) = delete;
-
-/*
- * Configuration enum for the ThrowingAllocator type that defines behavior for
- * the lifetime of the instance.
- *
- * kEverythingThrows: Calls to the member functions may throw
- * kNoThrowAllocate: Calls to the member functions will not throw
- */
-enum class AllocSpec {
-  kEverythingThrows = 0,
-  kNoThrowAllocate = 1,
-};
-
-/*
- * An allocator type which is instrumented to throw at a controlled time, or not
- * to throw, using AllocSpec. The supported settings are the default of every
- * function which is allowed to throw in a conforming allocator possibly
- * throwing, or nothing throws, in line with the ABSL_ALLOCATOR_THROWS
- * configuration macro.
- */
-template <typename T, AllocSpec Spec = AllocSpec::kEverythingThrows>
-class ThrowingAllocator : private exceptions_internal::TrackedObject {
-  static constexpr bool IsSpecified(AllocSpec spec) {
-    return static_cast<bool>(Spec & spec);
-  }
-
- public:
-  using pointer = T*;
-  using const_pointer = const T*;
-  using reference = T&;
-  using const_reference = const T&;
-  using void_pointer = void*;
-  using const_void_pointer = const void*;
-  using value_type = T;
-  using size_type = size_t;
-  using difference_type = ptrdiff_t;
-
-  using is_nothrow =
-      std::integral_constant<bool, Spec == AllocSpec::kNoThrowAllocate>;
-  using propagate_on_container_copy_assignment = std::true_type;
-  using propagate_on_container_move_assignment = std::true_type;
-  using propagate_on_container_swap = std::true_type;
-  using is_always_equal = std::false_type;
-
-  ThrowingAllocator() : TrackedObject(GetInstanceString(next_id_)) {
-    exceptions_internal::MaybeThrow(ABSL_PRETTY_FUNCTION);
-    dummy_ = std::make_shared<const int>(next_id_++);
-  }
-
-  template <typename U>
-  ThrowingAllocator(const ThrowingAllocator<U, Spec>& other) noexcept  // NOLINT
-      : TrackedObject(GetInstanceString(*other.State())),
-        dummy_(other.State()) {}
-
-  // According to C++11 standard [17.6.3.5], Table 28, the move/copy ctors of
-  // allocator shall not exit via an exception, thus they are marked noexcept.
-  ThrowingAllocator(const ThrowingAllocator& other) noexcept
-      : TrackedObject(GetInstanceString(*other.State())),
-        dummy_(other.State()) {}
-
-  template <typename U>
-  ThrowingAllocator(ThrowingAllocator<U, Spec>&& other) noexcept  // NOLINT
-      : TrackedObject(GetInstanceString(*other.State())),
-        dummy_(std::move(other.State())) {}
-
-  ThrowingAllocator(ThrowingAllocator&& other) noexcept
-      : TrackedObject(GetInstanceString(*other.State())),
-        dummy_(std::move(other.State())) {}
-
-  ~ThrowingAllocator() noexcept = default;
-
-  ThrowingAllocator& operator=(const ThrowingAllocator& other) noexcept {
-    dummy_ = other.State();
-    return *this;
-  }
-
-  template <typename U>
-  ThrowingAllocator& operator=(
-      const ThrowingAllocator<U, Spec>& other) noexcept {
-    dummy_ = other.State();
-    return *this;
-  }
-
-  template <typename U>
-  ThrowingAllocator& operator=(ThrowingAllocator<U, Spec>&& other) noexcept {
-    dummy_ = std::move(other.State());
-    return *this;
-  }
-
-  template <typename U>
-  struct rebind {
-    using other = ThrowingAllocator<U, Spec>;
-  };
-
-  pointer allocate(size_type n) noexcept(
-      IsSpecified(AllocSpec::kNoThrowAllocate)) {
-    ReadStateAndMaybeThrow(ABSL_PRETTY_FUNCTION);
-    return static_cast<pointer>(::operator new(n * sizeof(T)));
-  }
-
-  pointer allocate(size_type n, const_void_pointer) noexcept(
-      IsSpecified(AllocSpec::kNoThrowAllocate)) {
-    return allocate(n);
-  }
-
-  void deallocate(pointer ptr, size_type) noexcept {
-    ReadState();
-    ::operator delete(static_cast<void*>(ptr));
-  }
-
-  template <typename U, typename... Args>
-  void construct(U* ptr, Args&&... args) noexcept(
-      IsSpecified(AllocSpec::kNoThrowAllocate)) {
-    ReadStateAndMaybeThrow(ABSL_PRETTY_FUNCTION);
-    ::new (static_cast<void*>(ptr)) U(std::forward<Args>(args)...);
-  }
-
-  template <typename U>
-  void destroy(U* p) noexcept {
-    ReadState();
-    p->~U();
-  }
-
-  size_type max_size() const noexcept {
-    return (std::numeric_limits<difference_type>::max)() / sizeof(value_type);
-  }
-
-  ThrowingAllocator select_on_container_copy_construction() noexcept(
-      IsSpecified(AllocSpec::kNoThrowAllocate)) {
-    ReadStateAndMaybeThrow(ABSL_PRETTY_FUNCTION);
-    return *this;
-  }
-
-  template <typename U>
-  bool operator==(const ThrowingAllocator<U, Spec>& other) const noexcept {
-    return dummy_ == other.dummy_;
-  }
-
-  template <typename U>
-  bool operator!=(const ThrowingAllocator<U, Spec>& other) const noexcept {
-    return dummy_ != other.dummy_;
-  }
-
-  template <typename, AllocSpec>
-  friend class ThrowingAllocator;
-
- private:
-  static std::string GetInstanceString(int dummy) {
-    return absl::StrCat("ThrowingAllocator<",
-                        exceptions_internal::GetSpecString(Spec), ">(", dummy,
-                        ")");
-  }
-
-  const std::shared_ptr<const int>& State() const { return dummy_; }
-  std::shared_ptr<const int>& State() { return dummy_; }
-
-  void ReadState() {
-    // we know that this will never be true, but the compiler doesn't, so this
-    // should safely force a read of the value.
-    if (*dummy_ < 0) std::abort();
-  }
-
-  void ReadStateAndMaybeThrow(absl::string_view msg) const {
-    if (!IsSpecified(AllocSpec::kNoThrowAllocate)) {
-      exceptions_internal::MaybeThrow(
-          absl::Substitute("Allocator id $0 threw from $1", *dummy_, msg));
-    }
-  }
-
-  static int next_id_;
-  std::shared_ptr<const int> dummy_;
-};
-
-template <typename T, AllocSpec Spec>
-int ThrowingAllocator<T, Spec>::next_id_ = 0;
-
-// Tests for resource leaks by attempting to construct a T using args repeatedly
-// until successful, using the countdown method.  Side effects can then be
-// tested for resource leaks.
-template <typename T, typename... Args>
-void TestThrowingCtor(Args&&... args) {
-  struct Cleanup {
-    ~Cleanup() { exceptions_internal::UnsetCountdown(); }
-  } c;
-  for (int count = 0;; ++count) {
-    exceptions_internal::ConstructorTracker ct(count);
-    exceptions_internal::SetCountdown(count);
-    try {
-      T temp(std::forward<Args>(args)...);
-      static_cast<void>(temp);
-      break;
-    } catch (const exceptions_internal::TestException&) {
-    }
-  }
-}
-
-// Tests the nothrow guarantee of the provided nullary operation. If the an
-// exception is thrown, the result will be AssertionFailure(). Otherwise, it
-// will be AssertionSuccess().
-template <typename Operation>
-testing::AssertionResult TestNothrowOp(const Operation& operation) {
-  struct Cleanup {
-    Cleanup() { exceptions_internal::SetCountdown(); }
-    ~Cleanup() { exceptions_internal::UnsetCountdown(); }
-  } c;
-  try {
-    operation();
-    return testing::AssertionSuccess();
-  } catch (const exceptions_internal::TestException&) {
-    return testing::AssertionFailure()
-           << "TestException thrown during call to operation() when nothrow "
-              "guarantee was expected.";
-  } catch (...) {
-    return testing::AssertionFailure()
-           << "Unknown exception thrown during call to operation() when "
-              "nothrow guarantee was expected.";
-  }
-}
-
-namespace exceptions_internal {
-
-// Dummy struct for ExceptionSafetyTestBuilder<> partial state.
-struct UninitializedT {};
-
-template <typename T>
-class DefaultFactory {
- public:
-  explicit DefaultFactory(const T& t) : t_(t) {}
-  std::unique_ptr<T> operator()() const { return absl::make_unique<T>(t_); }
-
- private:
-  T t_;
-};
-
-template <size_t LazyContractsCount, typename LazyFactory,
-          typename LazyOperation>
-using EnableIfTestable = typename absl::enable_if_t<
-    LazyContractsCount != 0 &&
-    !std::is_same<LazyFactory, UninitializedT>::value &&
-    !std::is_same<LazyOperation, UninitializedT>::value>;
-
-template <typename Factory = UninitializedT,
-          typename Operation = UninitializedT, typename... Contracts>
-class ExceptionSafetyTestBuilder;
-
-}  // namespace exceptions_internal
-
-/*
- * Constructs an empty ExceptionSafetyTestBuilder. All
- * ExceptionSafetyTestBuilder objects are immutable and all With[thing] mutation
- * methods return new instances of ExceptionSafetyTestBuilder.
- *
- * In order to test a T for exception safety, a factory for that T, a testable
- * operation, and at least one contract callback returning an assertion
- * result must be applied using the respective methods.
- */
-exceptions_internal::ExceptionSafetyTestBuilder<> MakeExceptionSafetyTester();
-
-namespace exceptions_internal {
-template <typename T>
-struct IsUniquePtr : std::false_type {};
-
-template <typename T, typename D>
-struct IsUniquePtr<std::unique_ptr<T, D>> : std::true_type {};
-
-template <typename Factory>
-struct FactoryPtrTypeHelper {
-  using type = decltype(std::declval<const Factory&>()());
-
-  static_assert(IsUniquePtr<type>::value, "Factories must return a unique_ptr");
-};
-
-template <typename Factory>
-using FactoryPtrType = typename FactoryPtrTypeHelper<Factory>::type;
-
-template <typename Factory>
-using FactoryElementType = typename FactoryPtrType<Factory>::element_type;
-
-template <typename T>
-class ExceptionSafetyTest {
-  using Factory = std::function<std::unique_ptr<T>()>;
-  using Operation = std::function<void(T*)>;
-  using Contract = std::function<AssertionResult(T*)>;
-
- public:
-  template <typename... Contracts>
-  explicit ExceptionSafetyTest(const Factory& f, const Operation& op,
-                               const Contracts&... contracts)
-      : factory_(f), operation_(op), contracts_{WrapContract(contracts)...} {}
-
-  AssertionResult Test() const {
-    for (int count = 0;; ++count) {
-      exceptions_internal::ConstructorTracker ct(count);
-
-      for (const auto& contract : contracts_) {
-        auto t_ptr = factory_();
-        try {
-          SetCountdown(count);
-          operation_(t_ptr.get());
-          // Unset for the case that the operation throws no exceptions, which
-          // would leave the countdown set and break the *next* exception safety
-          // test after this one.
-          UnsetCountdown();
-          return AssertionSuccess();
-        } catch (const exceptions_internal::TestException& e) {
-          if (!contract(t_ptr.get())) {
-            return AssertionFailure() << e.what() << " failed contract check";
-          }
-        }
-      }
-    }
-  }
-
- private:
-  template <typename ContractFn>
-  Contract WrapContract(const ContractFn& contract) {
-    return [contract](T* t_ptr) { return AssertionResult(contract(t_ptr)); };
-  }
-
-  Contract WrapContract(StrongGuaranteeTagType) {
-    return [this](T* t_ptr) { return AssertionResult(*factory_() == *t_ptr); };
-  }
-
-  Factory factory_;
-  Operation operation_;
-  std::vector<Contract> contracts_;
-};
-
-/*
- * Builds a tester object that tests if performing a operation on a T follows
- * exception safety guarantees. Verification is done via contract assertion
- * callbacks applied to T instances post-throw.
- *
- * Template parameters for ExceptionSafetyTestBuilder:
- *
- * - Factory: The factory object (passed in via tester.WithFactory(...) or
- *   tester.WithInitialValue(...)) must be invocable with the signature
- *   `std::unique_ptr<T> operator()() const` where T is the type being tested.
- *   It is used for reliably creating identical T instances to test on.
- *
- * - Operation: The operation object (passed in via tester.WithOperation(...)
- *   or tester.Test(...)) must be invocable with the signature
- *   `void operator()(T*) const` where T is the type being tested. It is used
- *   for performing steps on a T instance that may throw and that need to be
- *   checked for exception safety. Each call to the operation will receive a
- *   fresh T instance so it's free to modify and destroy the T instances as it
- *   pleases.
- *
- * - Contracts...: The contract assertion callback objects (passed in via
- *   tester.WithContracts(...)) must be invocable with the signature
- *   `testing::AssertionResult operator()(T*) const` where T is the type being
- *   tested. Contract assertion callbacks are provided T instances post-throw.
- *   They must return testing::AssertionSuccess when the type contracts of the
- *   provided T instance hold. If the type contracts of the T instance do not
- *   hold, they must return testing::AssertionFailure. Execution order of
- *   Contracts... is unspecified. They will each individually get a fresh T
- *   instance so they are free to modify and destroy the T instances as they
- *   please.
- */
-template <typename Factory, typename Operation, typename... Contracts>
-class ExceptionSafetyTestBuilder {
- public:
-  /*
-   * Returns a new ExceptionSafetyTestBuilder with an included T factory based
-   * on the provided T instance. The existing factory will not be included in
-   * the newly created tester instance. The created factory returns a new T
-   * instance by copy-constructing the provided const T& t.
-   *
-   * Preconditions for tester.WithInitialValue(const T& t):
-   *
-   * - The const T& t object must be copy-constructible where T is the type
-   *   being tested. For non-copy-constructible objects, use the method
-   *   tester.WithFactory(...).
-   */
-  template <typename T>
-  ExceptionSafetyTestBuilder<DefaultFactory<T>, Operation, Contracts...>
-  WithInitialValue(const T& t) const {
-    return WithFactory(DefaultFactory<T>(t));
-  }
-
-  /*
-   * Returns a new ExceptionSafetyTestBuilder with the provided T factory
-   * included. The existing factory will not be included in the newly-created
-   * tester instance. This method is intended for use with types lacking a copy
-   * constructor. Types that can be copy-constructed should instead use the
-   * method tester.WithInitialValue(...).
-   */
-  template <typename NewFactory>
-  ExceptionSafetyTestBuilder<absl::decay_t<NewFactory>, Operation, Contracts...>
-  WithFactory(const NewFactory& new_factory) const {
-    return {new_factory, operation_, contracts_};
-  }
-
-  /*
-   * Returns a new ExceptionSafetyTestBuilder with the provided testable
-   * operation included. The existing operation will not be included in the
-   * newly created tester.
-   */
-  template <typename NewOperation>
-  ExceptionSafetyTestBuilder<Factory, absl::decay_t<NewOperation>, Contracts...>
-  WithOperation(const NewOperation& new_operation) const {
-    return {factory_, new_operation, contracts_};
-  }
-
-  /*
-   * Returns a new ExceptionSafetyTestBuilder with the provided MoreContracts...
-   * combined with the Contracts... that were already included in the instance
-   * on which the method was called. Contracts... cannot be removed or replaced
-   * once added to an ExceptionSafetyTestBuilder instance. A fresh object must
-   * be created in order to get an empty Contracts... list.
-   *
-   * In addition to passing in custom contract assertion callbacks, this method
-   * accepts `testing::strong_guarantee` as an argument which checks T instances
-   * post-throw against freshly created T instances via operator== to verify
-   * that any state changes made during the execution of the operation were
-   * properly rolled back.
-   */
-  template <typename... MoreContracts>
-  ExceptionSafetyTestBuilder<Factory, Operation, Contracts...,
-                             absl::decay_t<MoreContracts>...>
-  WithContracts(const MoreContracts&... more_contracts) const {
-    return {
-        factory_, operation_,
-        std::tuple_cat(contracts_, std::tuple<absl::decay_t<MoreContracts>...>(
-                                       more_contracts...))};
-  }
-
-  /*
-   * Returns a testing::AssertionResult that is the reduced result of the
-   * exception safety algorithm. The algorithm short circuits and returns
-   * AssertionFailure after the first contract callback returns an
-   * AssertionFailure. Otherwise, if all contract callbacks return an
-   * AssertionSuccess, the reduced result is AssertionSuccess.
-   *
-   * The passed-in testable operation will not be saved in a new tester instance
-   * nor will it modify/replace the existing tester instance. This is useful
-   * when each operation being tested is unique and does not need to be reused.
-   *
-   * Preconditions for tester.Test(const NewOperation& new_operation):
-   *
-   * - May only be called after at least one contract assertion callback and a
-   *   factory or initial value have been provided.
-   */
-  template <
-      typename NewOperation,
-      typename = EnableIfTestable<sizeof...(Contracts), Factory, NewOperation>>
-  testing::AssertionResult Test(const NewOperation& new_operation) const {
-    return TestImpl(new_operation, absl::index_sequence_for<Contracts...>());
-  }
-
-  /*
-   * Returns a testing::AssertionResult that is the reduced result of the
-   * exception safety algorithm. The algorithm short circuits and returns
-   * AssertionFailure after the first contract callback returns an
-   * AssertionFailure. Otherwise, if all contract callbacks return an
-   * AssertionSuccess, the reduced result is AssertionSuccess.
-   *
-   * Preconditions for tester.Test():
-   *
-   * - May only be called after at least one contract assertion callback, a
-   *   factory or initial value and a testable operation have been provided.
-   */
-  template <
-      typename LazyOperation = Operation,
-      typename = EnableIfTestable<sizeof...(Contracts), Factory, LazyOperation>>
-  testing::AssertionResult Test() const {
-    return Test(operation_);
-  }
-
- private:
-  template <typename, typename, typename...>
-  friend class ExceptionSafetyTestBuilder;
-
-  friend ExceptionSafetyTestBuilder<> testing::MakeExceptionSafetyTester();
-
-  ExceptionSafetyTestBuilder() {}
-
-  ExceptionSafetyTestBuilder(const Factory& f, const Operation& o,
-                             const std::tuple<Contracts...>& i)
-      : factory_(f), operation_(o), contracts_(i) {}
-
-  template <typename SelectedOperation, size_t... Indices>
-  testing::AssertionResult TestImpl(SelectedOperation selected_operation,
-                                    absl::index_sequence<Indices...>) const {
-    return ExceptionSafetyTest<FactoryElementType<Factory>>(
-               factory_, selected_operation, std::get<Indices>(contracts_)...)
-        .Test();
-  }
-
-  Factory factory_;
-  Operation operation_;
-  std::tuple<Contracts...> contracts_;
-};
-
-}  // namespace exceptions_internal
-
-}  // namespace testing
-
-#endif  // ABSL_HAVE_EXCEPTIONS
-
-#endif  // ABSL_BASE_INTERNAL_EXCEPTION_SAFETY_TESTING_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/exception_testing.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/exception_testing.h
deleted file mode 100644
index 01b54655..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/exception_testing.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-// Testing utilities for ABSL types which throw exceptions.
-
-#ifndef ABSL_BASE_INTERNAL_EXCEPTION_TESTING_H_
-#define ABSL_BASE_INTERNAL_EXCEPTION_TESTING_H_
-
-#include "gtest/gtest.h"
-#include "absl/base/config.h"
-
-// ABSL_BASE_INTERNAL_EXPECT_FAIL tests either for a specified thrown exception
-// if exceptions are enabled, or for death with a specified text in the error
-// message
-#ifdef ABSL_HAVE_EXCEPTIONS
-
-#define ABSL_BASE_INTERNAL_EXPECT_FAIL(expr, exception_t, text) \
-  EXPECT_THROW(expr, exception_t)
-
-#elif defined(__ANDROID__)
-// Android asserts do not log anywhere that gtest can currently inspect.
-// So we expect exit, but cannot match the message.
-#define ABSL_BASE_INTERNAL_EXPECT_FAIL(expr, exception_t, text) \
-  EXPECT_DEATH(expr, ".*")
-#else
-#define ABSL_BASE_INTERNAL_EXPECT_FAIL(expr, exception_t, text) \
-  EXPECT_DEATH_IF_SUPPORTED(expr, text)
-
-#endif
-
-#endif  // ABSL_BASE_INTERNAL_EXCEPTION_TESTING_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/fast_type_id.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/fast_type_id.h
deleted file mode 100644
index a547b3a8..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/fast_type_id.h
+++ /dev/null
@@ -1,50 +0,0 @@
-//
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_BASE_INTERNAL_FAST_TYPE_ID_H_
-#define ABSL_BASE_INTERNAL_FAST_TYPE_ID_H_
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-template <typename Type>
-struct FastTypeTag {
-  constexpr static char dummy_var = 0;
-};
-
-#ifdef ABSL_INTERNAL_NEED_REDUNDANT_CONSTEXPR_DECL
-template <typename Type>
-constexpr char FastTypeTag<Type>::dummy_var;
-#endif
-
-// FastTypeId<Type>() evaluates at compile/link-time to a unique pointer for the
-// passed-in type. These are meant to be good match for keys into maps or
-// straight up comparisons.
-using FastTypeIdType = const void*;
-
-template <typename Type>
-constexpr inline FastTypeIdType FastTypeId() {
-  return &FastTypeTag<Type>::dummy_var;
-}
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_FAST_TYPE_ID_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/hide_ptr.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/hide_ptr.h
deleted file mode 100644
index 1dba8090..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/hide_ptr.h
+++ /dev/null
@@ -1,51 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_BASE_INTERNAL_HIDE_PTR_H_
-#define ABSL_BASE_INTERNAL_HIDE_PTR_H_
-
-#include <cstdint>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-// Arbitrary value with high bits set. Xor'ing with it is unlikely
-// to map one valid pointer to another valid pointer.
-constexpr uintptr_t HideMask() {
-  return (uintptr_t{0xF03A5F7BU} << (sizeof(uintptr_t) - 4) * 8) | 0xF03A5F7BU;
-}
-
-// Hide a pointer from the leak checker. For internal use only.
-// Differs from absl::IgnoreLeak(ptr) in that absl::IgnoreLeak(ptr) causes ptr
-// and all objects reachable from ptr to be ignored by the leak checker.
-template <class T>
-inline uintptr_t HidePtr(T* ptr) {
-  return reinterpret_cast<uintptr_t>(ptr) ^ HideMask();
-}
-
-// Return a pointer that has been hidden from the leak checker.
-// For internal use only.
-template <class T>
-inline T* UnhidePtr(uintptr_t hidden) {
-  return reinterpret_cast<T*>(hidden ^ HideMask());
-}
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_HIDE_PTR_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/identity.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/identity.h
deleted file mode 100644
index 365207b7..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/identity.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_BASE_INTERNAL_IDENTITY_H_
-#define ABSL_BASE_INTERNAL_IDENTITY_H_
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace internal {
-
-// This is a back-fill of C++20's `std::type_identity`.
-template <typename T>
-struct type_identity {
-  typedef T type;
-};
-
-// This is a back-fill of C++20's `std::type_identity_t`.
-template <typename T>
-using type_identity_t = typename type_identity<T>::type;
-
-}  // namespace internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_IDENTITY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/inline_variable.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/inline_variable.h
deleted file mode 100644
index 09daf0f5..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/inline_variable.h
+++ /dev/null
@@ -1,108 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_BASE_INTERNAL_INLINE_VARIABLE_H_
-#define ABSL_BASE_INTERNAL_INLINE_VARIABLE_H_
-
-#include <type_traits>
-
-#include "absl/base/internal/identity.h"
-
-// File:
-//   This file define a macro that allows the creation of or emulation of C++17
-//   inline variables based on whether or not the feature is supported.
-
-////////////////////////////////////////////////////////////////////////////////
-// Macro: ABSL_INTERNAL_INLINE_CONSTEXPR(type, name, init)
-//
-// Description:
-//   Expands to the equivalent of an inline constexpr instance of the specified
-//   `type` and `name`, initialized to the value `init`. If the compiler being
-//   used is detected as supporting actual inline variables as a language
-//   feature, then the macro expands to an actual inline variable definition.
-//
-// Requires:
-//   `type` is a type that is usable in an extern variable declaration.
-//
-// Requires: `name` is a valid identifier
-//
-// Requires:
-//   `init` is an expression that can be used in the following definition:
-//     constexpr type name = init;
-//
-// Usage:
-//
-//   // Equivalent to: `inline constexpr size_t variant_npos = -1;`
-//   ABSL_INTERNAL_INLINE_CONSTEXPR(size_t, variant_npos, -1);
-//
-// Differences in implementation:
-//   For a direct, language-level inline variable, decltype(name) will be the
-//   type that was specified along with const qualification, whereas for
-//   emulated inline variables, decltype(name) may be different (in practice
-//   it will likely be a reference type).
-////////////////////////////////////////////////////////////////////////////////
-
-#ifdef __cpp_inline_variables
-
-// Clang's -Wmissing-variable-declarations option erroneously warned that
-// inline constexpr objects need to be pre-declared. This has now been fixed,
-// but we will need to support this workaround for people building with older
-// versions of clang.
-//
-// Bug: https://bugs.llvm.org/show_bug.cgi?id=35862
-//
-// Note:
-//   type_identity_t is used here so that the const and name are in the
-//   appropriate place for pointer types, reference types, function pointer
-//   types, etc..
-#if defined(__clang__)
-#define ABSL_INTERNAL_EXTERN_DECL(type, name) \
-  extern const ::absl::internal::type_identity_t<type> name;
-#else  // Otherwise, just define the macro to do nothing.
-#define ABSL_INTERNAL_EXTERN_DECL(type, name)
-#endif  // defined(__clang__)
-
-// See above comment at top of file for details.
-#define ABSL_INTERNAL_INLINE_CONSTEXPR(type, name, init) \
-  ABSL_INTERNAL_EXTERN_DECL(type, name)                  \
-  inline constexpr ::absl::internal::type_identity_t<type> name = init
-
-#else
-
-// See above comment at top of file for details.
-//
-// Note:
-//   type_identity_t is used here so that the const and name are in the
-//   appropriate place for pointer types, reference types, function pointer
-//   types, etc..
-#define ABSL_INTERNAL_INLINE_CONSTEXPR(var_type, name, init)                 \
-  template <class /*AbslInternalDummy*/ = void>                              \
-  struct AbslInternalInlineVariableHolder##name {                            \
-    static constexpr ::absl::internal::type_identity_t<var_type> kInstance = \
-        init;                                                                \
-  };                                                                         \
-                                                                             \
-  template <class AbslInternalDummy>                                         \
-  constexpr ::absl::internal::type_identity_t<var_type>                      \
-      AbslInternalInlineVariableHolder##name<AbslInternalDummy>::kInstance;  \
-                                                                             \
-  static constexpr const ::absl::internal::type_identity_t<var_type>&        \
-      name = /* NOLINT */                                                    \
-      AbslInternalInlineVariableHolder##name<>::kInstance;                   \
-  static_assert(sizeof(void (*)(decltype(name))) != 0,                       \
-                "Silence unused variable warnings.")
-
-#endif  // __cpp_inline_variables
-
-#endif  // ABSL_BASE_INTERNAL_INLINE_VARIABLE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/inline_variable_testing.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/inline_variable_testing.h
deleted file mode 100644
index f3c81459..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/inline_variable_testing.h
+++ /dev/null
@@ -1,46 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_BASE_INTERNAL_INLINE_VARIABLE_TESTING_H_
-#define ABSL_BASE_INTERNAL_INLINE_VARIABLE_TESTING_H_
-
-#include "absl/base/internal/inline_variable.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace inline_variable_testing_internal {
-
-struct Foo {
-  int value = 5;
-};
-
-ABSL_INTERNAL_INLINE_CONSTEXPR(Foo, inline_variable_foo, {});
-ABSL_INTERNAL_INLINE_CONSTEXPR(Foo, other_inline_variable_foo, {});
-
-ABSL_INTERNAL_INLINE_CONSTEXPR(int, inline_variable_int, 5);
-ABSL_INTERNAL_INLINE_CONSTEXPR(int, other_inline_variable_int, 5);
-
-ABSL_INTERNAL_INLINE_CONSTEXPR(void(*)(), inline_variable_fun_ptr, nullptr);
-
-const Foo& get_foo_a();
-const Foo& get_foo_b();
-
-const int& get_int_a();
-const int& get_int_b();
-
-}  // namespace inline_variable_testing_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_INLINE_VARIABLE_TESTING_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/invoke.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/invoke.h
deleted file mode 100644
index 643c2a42..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/invoke.h
+++ /dev/null
@@ -1,241 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// absl::base_internal::invoke(f, args...) is an implementation of
-// INVOKE(f, args...) from section [func.require] of the C++ standard.
-// When compiled as C++17 and later versions, it is implemented as an alias of
-// std::invoke.
-//
-// [func.require]
-// Define INVOKE (f, t1, t2, ..., tN) as follows:
-// 1. (t1.*f)(t2, ..., tN) when f is a pointer to a member function of a class T
-//    and t1 is an object of type T or a reference to an object of type T or a
-//    reference to an object of a type derived from T;
-// 2. ((*t1).*f)(t2, ..., tN) when f is a pointer to a member function of a
-//    class T and t1 is not one of the types described in the previous item;
-// 3. t1.*f when N == 1 and f is a pointer to member data of a class T and t1 is
-//    an object of type T or a reference to an object of type T or a reference
-//    to an object of a type derived from T;
-// 4. (*t1).*f when N == 1 and f is a pointer to member data of a class T and t1
-//    is not one of the types described in the previous item;
-// 5. f(t1, t2, ..., tN) in all other cases.
-//
-// The implementation is SFINAE-friendly: substitution failure within invoke()
-// isn't an error.
-
-#ifndef ABSL_BASE_INTERNAL_INVOKE_H_
-#define ABSL_BASE_INTERNAL_INVOKE_H_
-
-#include "absl/base/config.h"
-
-#if ABSL_INTERNAL_CPLUSPLUS_LANG >= 201703L
-
-#include <functional>
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-using std::invoke;
-using std::invoke_result_t;
-using std::is_invocable_r;
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#else  // ABSL_INTERNAL_CPLUSPLUS_LANG >= 201703L
-
-#include <algorithm>
-#include <type_traits>
-#include <utility>
-
-#include "absl/meta/type_traits.h"
-
-// The following code is internal implementation detail.  See the comment at the
-// top of this file for the API documentation.
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-// The five classes below each implement one of the clauses from the definition
-// of INVOKE. The inner class template Accept<F, Args...> checks whether the
-// clause is applicable; static function template Invoke(f, args...) does the
-// invocation.
-//
-// By separating the clause selection logic from invocation we make sure that
-// Invoke() does exactly what the standard says.
-
-template <typename Derived>
-struct StrippedAccept {
-  template <typename... Args>
-  struct Accept : Derived::template AcceptImpl<typename std::remove_cv<
-                      typename std::remove_reference<Args>::type>::type...> {};
-};
-
-// (t1.*f)(t2, ..., tN) when f is a pointer to a member function of a class T
-// and t1 is an object of type T or a reference to an object of type T or a
-// reference to an object of a type derived from T.
-struct MemFunAndRef : StrippedAccept<MemFunAndRef> {
-  template <typename... Args>
-  struct AcceptImpl : std::false_type {};
-
-  template <typename MemFunType, typename C, typename Obj, typename... Args>
-  struct AcceptImpl<MemFunType C::*, Obj, Args...>
-      : std::integral_constant<bool, std::is_base_of<C, Obj>::value &&
-                                         absl::is_function<MemFunType>::value> {
-  };
-
-  template <typename MemFun, typename Obj, typename... Args>
-  static decltype((std::declval<Obj>().*
-                   std::declval<MemFun>())(std::declval<Args>()...))
-  Invoke(MemFun&& mem_fun, Obj&& obj, Args&&... args) {
-// Ignore bogus GCC warnings on this line.
-// See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101436 for similar example.
-#if ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(11, 0)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Warray-bounds"
-#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
-#endif
-    return (std::forward<Obj>(obj).*
-            std::forward<MemFun>(mem_fun))(std::forward<Args>(args)...);
-#if ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(11, 0)
-#pragma GCC diagnostic pop
-#endif
-  }
-};
-
-// ((*t1).*f)(t2, ..., tN) when f is a pointer to a member function of a
-// class T and t1 is not one of the types described in the previous item.
-struct MemFunAndPtr : StrippedAccept<MemFunAndPtr> {
-  template <typename... Args>
-  struct AcceptImpl : std::false_type {};
-
-  template <typename MemFunType, typename C, typename Ptr, typename... Args>
-  struct AcceptImpl<MemFunType C::*, Ptr, Args...>
-      : std::integral_constant<bool, !std::is_base_of<C, Ptr>::value &&
-                                         absl::is_function<MemFunType>::value> {
-  };
-
-  template <typename MemFun, typename Ptr, typename... Args>
-  static decltype(((*std::declval<Ptr>()).*
-                   std::declval<MemFun>())(std::declval<Args>()...))
-  Invoke(MemFun&& mem_fun, Ptr&& ptr, Args&&... args) {
-    return ((*std::forward<Ptr>(ptr)).*
-            std::forward<MemFun>(mem_fun))(std::forward<Args>(args)...);
-  }
-};
-
-// t1.*f when N == 1 and f is a pointer to member data of a class T and t1 is
-// an object of type T or a reference to an object of type T or a reference
-// to an object of a type derived from T.
-struct DataMemAndRef : StrippedAccept<DataMemAndRef> {
-  template <typename... Args>
-  struct AcceptImpl : std::false_type {};
-
-  template <typename R, typename C, typename Obj>
-  struct AcceptImpl<R C::*, Obj>
-      : std::integral_constant<bool, std::is_base_of<C, Obj>::value &&
-                                         !absl::is_function<R>::value> {};
-
-  template <typename DataMem, typename Ref>
-  static decltype(std::declval<Ref>().*std::declval<DataMem>()) Invoke(
-      DataMem&& data_mem, Ref&& ref) {
-    return std::forward<Ref>(ref).*std::forward<DataMem>(data_mem);
-  }
-};
-
-// (*t1).*f when N == 1 and f is a pointer to member data of a class T and t1
-// is not one of the types described in the previous item.
-struct DataMemAndPtr : StrippedAccept<DataMemAndPtr> {
-  template <typename... Args>
-  struct AcceptImpl : std::false_type {};
-
-  template <typename R, typename C, typename Ptr>
-  struct AcceptImpl<R C::*, Ptr>
-      : std::integral_constant<bool, !std::is_base_of<C, Ptr>::value &&
-                                         !absl::is_function<R>::value> {};
-
-  template <typename DataMem, typename Ptr>
-  static decltype((*std::declval<Ptr>()).*std::declval<DataMem>()) Invoke(
-      DataMem&& data_mem, Ptr&& ptr) {
-    return (*std::forward<Ptr>(ptr)).*std::forward<DataMem>(data_mem);
-  }
-};
-
-// f(t1, t2, ..., tN) in all other cases.
-struct Callable {
-  // Callable doesn't have Accept because it's the last clause that gets picked
-  // when none of the previous clauses are applicable.
-  template <typename F, typename... Args>
-  static decltype(std::declval<F>()(std::declval<Args>()...)) Invoke(
-      F&& f, Args&&... args) {
-    return std::forward<F>(f)(std::forward<Args>(args)...);
-  }
-};
-
-// Resolves to the first matching clause.
-template <typename... Args>
-struct Invoker {
-  typedef typename std::conditional<
-      MemFunAndRef::Accept<Args...>::value, MemFunAndRef,
-      typename std::conditional<
-          MemFunAndPtr::Accept<Args...>::value, MemFunAndPtr,
-          typename std::conditional<
-              DataMemAndRef::Accept<Args...>::value, DataMemAndRef,
-              typename std::conditional<DataMemAndPtr::Accept<Args...>::value,
-                                        DataMemAndPtr, Callable>::type>::type>::
-          type>::type type;
-};
-
-// The result type of Invoke<F, Args...>.
-template <typename F, typename... Args>
-using invoke_result_t = decltype(Invoker<F, Args...>::type::Invoke(
-    std::declval<F>(), std::declval<Args>()...));
-
-// Invoke(f, args...) is an implementation of INVOKE(f, args...) from section
-// [func.require] of the C++ standard.
-template <typename F, typename... Args>
-invoke_result_t<F, Args...> invoke(F&& f, Args&&... args) {
-  return Invoker<F, Args...>::type::Invoke(std::forward<F>(f),
-                                           std::forward<Args>(args)...);
-}
-
-template <typename AlwaysVoid, typename, typename, typename...>
-struct IsInvocableRImpl : std::false_type {};
-
-template <typename R, typename F, typename... Args>
-struct IsInvocableRImpl<
-    absl::void_t<absl::base_internal::invoke_result_t<F, Args...> >, R, F,
-    Args...>
-    : std::integral_constant<
-          bool,
-          std::is_convertible<absl::base_internal::invoke_result_t<F, Args...>,
-                              R>::value ||
-              std::is_void<R>::value> {};
-
-// Type trait whose member `value` is true if invoking `F` with `Args` is valid,
-// and either the return type is convertible to `R`, or `R` is void.
-// C++11-compatible version of `std::is_invocable_r`.
-template <typename R, typename F, typename... Args>
-using is_invocable_r = IsInvocableRImpl<void, R, F, Args...>;
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_INTERNAL_CPLUSPLUS_LANG >= 201703L
-
-#endif  // ABSL_BASE_INTERNAL_INVOKE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/low_level_alloc.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/low_level_alloc.h
deleted file mode 100644
index c2f1f25d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/low_level_alloc.h
+++ /dev/null
@@ -1,127 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_BASE_INTERNAL_LOW_LEVEL_ALLOC_H_
-#define ABSL_BASE_INTERNAL_LOW_LEVEL_ALLOC_H_
-
-// A simple thread-safe memory allocator that does not depend on
-// mutexes or thread-specific data.  It is intended to be used
-// sparingly, and only when malloc() would introduce an unwanted
-// dependency, such as inside the heap-checker, or the Mutex
-// implementation.
-
-// IWYU pragma: private, include "base/low_level_alloc.h"
-
-#include <sys/types.h>
-
-#include <cstdint>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-
-// LowLevelAlloc requires that the platform support low-level
-// allocation of virtual memory. Platforms lacking this cannot use
-// LowLevelAlloc.
-#ifdef ABSL_LOW_LEVEL_ALLOC_MISSING
-#error ABSL_LOW_LEVEL_ALLOC_MISSING cannot be directly set
-#elif !defined(ABSL_HAVE_MMAP) && !defined(_WIN32)
-#define ABSL_LOW_LEVEL_ALLOC_MISSING 1
-#endif
-
-// Using LowLevelAlloc with kAsyncSignalSafe isn't supported on Windows or
-// asm.js / WebAssembly.
-// See https://kripken.github.io/emscripten-site/docs/porting/pthreads.html
-// for more information.
-#ifdef ABSL_LOW_LEVEL_ALLOC_ASYNC_SIGNAL_SAFE_MISSING
-#error ABSL_LOW_LEVEL_ALLOC_ASYNC_SIGNAL_SAFE_MISSING cannot be directly set
-#elif defined(_WIN32) || defined(__asmjs__) || defined(__wasm__) || \
-    defined(__hexagon__)
-#define ABSL_LOW_LEVEL_ALLOC_ASYNC_SIGNAL_SAFE_MISSING 1
-#endif
-
-#include <cstddef>
-
-#include "absl/base/port.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-class LowLevelAlloc {
- public:
-  struct Arena;       // an arena from which memory may be allocated
-
-  // Returns a pointer to a block of at least "request" bytes
-  // that have been newly allocated from the specific arena.
-  // for Alloc() call the DefaultArena() is used.
-  // Returns 0 if passed request==0.
-  // Does not return 0 under other circumstances; it crashes if memory
-  // is not available.
-  static void *Alloc(size_t request) ABSL_ATTRIBUTE_SECTION(malloc_hook);
-  static void *AllocWithArena(size_t request, Arena *arena)
-      ABSL_ATTRIBUTE_SECTION(malloc_hook);
-
-  // Deallocates a region of memory that was previously allocated with
-  // Alloc().   Does nothing if passed 0.   "s" must be either 0,
-  // or must have been returned from a call to Alloc() and not yet passed to
-  // Free() since that call to Alloc().  The space is returned to the arena
-  // from which it was allocated.
-  static void Free(void *s) ABSL_ATTRIBUTE_SECTION(malloc_hook);
-
-  // ABSL_ATTRIBUTE_SECTION(malloc_hook) for Alloc* and Free
-  // are to put all callers of MallocHook::Invoke* in this module
-  // into special section,
-  // so that MallocHook::GetCallerStackTrace can function accurately.
-
-  // Create a new arena.
-  // The root metadata for the new arena is allocated in the
-  // meta_data_arena; the DefaultArena() can be passed for meta_data_arena.
-  // These values may be ored into flags:
-  enum {
-    // Report calls to Alloc() and Free() via the MallocHook interface.
-    // Set in the DefaultArena.
-    kCallMallocHook = 0x0001,
-
-#ifndef ABSL_LOW_LEVEL_ALLOC_ASYNC_SIGNAL_SAFE_MISSING
-    // Make calls to Alloc(), Free() be async-signal-safe. Not set in
-    // DefaultArena(). Not supported on all platforms.
-    kAsyncSignalSafe = 0x0002,
-#endif
-  };
-  // Construct a new arena.  The allocation of the underlying metadata honors
-  // the provided flags.  For example, the call NewArena(kAsyncSignalSafe)
-  // is itself async-signal-safe, as well as generatating an arena that provides
-  // async-signal-safe Alloc/Free.
-  static Arena *NewArena(uint32_t flags);
-
-  // Destroys an arena allocated by NewArena and returns true,
-  // provided no allocated blocks remain in the arena.
-  // If allocated blocks remain in the arena, does nothing and
-  // returns false.
-  // It is illegal to attempt to destroy the DefaultArena().
-  static bool DeleteArena(Arena *arena);
-
-  // The default arena that always exists.
-  static Arena *DefaultArena();
-
- private:
-  LowLevelAlloc();      // no instances
-};
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_LOW_LEVEL_ALLOC_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/low_level_scheduling.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/low_level_scheduling.h
deleted file mode 100644
index 9baccc06..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/low_level_scheduling.h
+++ /dev/null
@@ -1,134 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Core interfaces and definitions used by by low-level interfaces such as
-// SpinLock.
-
-#ifndef ABSL_BASE_INTERNAL_LOW_LEVEL_SCHEDULING_H_
-#define ABSL_BASE_INTERNAL_LOW_LEVEL_SCHEDULING_H_
-
-#include "absl/base/internal/raw_logging.h"
-#include "absl/base/internal/scheduling_mode.h"
-#include "absl/base/macros.h"
-
-// The following two declarations exist so SchedulingGuard may friend them with
-// the appropriate language linkage.  These callbacks allow libc internals, such
-// as function level statics, to schedule cooperatively when locking.
-extern "C" bool __google_disable_rescheduling(void);
-extern "C" void __google_enable_rescheduling(bool disable_result);
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-class CondVar;
-class Mutex;
-
-namespace synchronization_internal {
-int MutexDelay(int32_t c, int mode);
-}  // namespace synchronization_internal
-
-namespace base_internal {
-
-class SchedulingHelper;  // To allow use of SchedulingGuard.
-class SpinLock;          // To allow use of SchedulingGuard.
-
-// SchedulingGuard
-// Provides guard semantics that may be used to disable cooperative rescheduling
-// of the calling thread within specific program blocks.  This is used to
-// protect resources (e.g. low-level SpinLocks or Domain code) that cooperative
-// scheduling depends on.
-//
-// Domain implementations capable of rescheduling in reaction to involuntary
-// kernel thread actions (e.g blocking due to a pagefault or syscall) must
-// guarantee that an annotated thread is not allowed to (cooperatively)
-// reschedule until the annotated region is complete.
-//
-// It is an error to attempt to use a cooperatively scheduled resource (e.g.
-// Mutex) within a rescheduling-disabled region.
-//
-// All methods are async-signal safe.
-class SchedulingGuard {
- public:
-  // Returns true iff the calling thread may be cooperatively rescheduled.
-  static bool ReschedulingIsAllowed();
-  SchedulingGuard(const SchedulingGuard&) = delete;
-  SchedulingGuard& operator=(const SchedulingGuard&) = delete;
-
- private:
-  // Disable cooperative rescheduling of the calling thread.  It may still
-  // initiate scheduling operations (e.g. wake-ups), however, it may not itself
-  // reschedule.  Nestable.  The returned result is opaque, clients should not
-  // attempt to interpret it.
-  // REQUIRES: Result must be passed to a pairing EnableScheduling().
-  static bool DisableRescheduling();
-
-  // Marks the end of a rescheduling disabled region, previously started by
-  // DisableRescheduling().
-  // REQUIRES: Pairs with innermost call (and result) of DisableRescheduling().
-  static void EnableRescheduling(bool disable_result);
-
-  // A scoped helper for {Disable, Enable}Rescheduling().
-  // REQUIRES: destructor must run in same thread as constructor.
-  struct ScopedDisable {
-    ScopedDisable() { disabled = SchedulingGuard::DisableRescheduling(); }
-    ~ScopedDisable() { SchedulingGuard::EnableRescheduling(disabled); }
-
-    bool disabled;
-  };
-
-  // A scoped helper to enable rescheduling temporarily.
-  // REQUIRES: destructor must run in same thread as constructor.
-  class ScopedEnable {
-   public:
-    ScopedEnable();
-    ~ScopedEnable();
-
-   private:
-    int scheduling_disabled_depth_;
-  };
-
-  // Access to SchedulingGuard is explicitly permitted.
-  friend class absl::CondVar;
-  friend class absl::Mutex;
-  friend class SchedulingHelper;
-  friend class SpinLock;
-  friend int absl::synchronization_internal::MutexDelay(int32_t c, int mode);
-};
-
-//------------------------------------------------------------------------------
-// End of public interfaces.
-//------------------------------------------------------------------------------
-
-inline bool SchedulingGuard::ReschedulingIsAllowed() {
-  return false;
-}
-
-inline bool SchedulingGuard::DisableRescheduling() {
-  return false;
-}
-
-inline void SchedulingGuard::EnableRescheduling(bool /* disable_result */) {
-  return;
-}
-
-inline SchedulingGuard::ScopedEnable::ScopedEnable()
-    : scheduling_disabled_depth_(0) {}
-inline SchedulingGuard::ScopedEnable::~ScopedEnable() {
-  ABSL_RAW_CHECK(scheduling_disabled_depth_ == 0, "disable unused warning");
-}
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_LOW_LEVEL_SCHEDULING_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/nullability_impl.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/nullability_impl.h
deleted file mode 100644
index 36e1b33d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/nullability_impl.h
+++ /dev/null
@@ -1,106 +0,0 @@
-// Copyright 2023 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_BASE_INTERNAL_NULLABILITY_IMPL_H_
-#define ABSL_BASE_INTERNAL_NULLABILITY_IMPL_H_
-
-#include <memory>
-#include <type_traits>
-
-#include "absl/base/attributes.h"
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-
-namespace nullability_internal {
-
-// `IsNullabilityCompatible` checks whether its first argument is a class
-// explicitly tagged as supporting nullability annotations. The tag is the type
-// declaration `absl_nullability_compatible`.
-template <typename, typename = void>
-struct IsNullabilityCompatible : std::false_type {};
-
-template <typename T>
-struct IsNullabilityCompatible<
-    T, absl::void_t<typename T::absl_nullability_compatible>> : std::true_type {
-};
-
-template <typename T>
-constexpr bool IsSupportedType = IsNullabilityCompatible<T>::value;
-
-template <typename T>
-constexpr bool IsSupportedType<T*> = true;
-
-template <typename T, typename U>
-constexpr bool IsSupportedType<T U::*> = true;
-
-template <typename T, typename... Deleter>
-constexpr bool IsSupportedType<std::unique_ptr<T, Deleter...>> = true;
-
-template <typename T>
-constexpr bool IsSupportedType<std::shared_ptr<T>> = true;
-
-template <typename T>
-struct EnableNullable {
-  static_assert(nullability_internal::IsSupportedType<std::remove_cv_t<T>>,
-                "Template argument must be a raw or supported smart pointer "
-                "type. See absl/base/nullability.h.");
-  using type = T;
-};
-
-template <typename T>
-struct EnableNonnull {
-  static_assert(nullability_internal::IsSupportedType<std::remove_cv_t<T>>,
-                "Template argument must be a raw or supported smart pointer "
-                "type. See absl/base/nullability.h.");
-  using type = T;
-};
-
-template <typename T>
-struct EnableNullabilityUnknown {
-  static_assert(nullability_internal::IsSupportedType<std::remove_cv_t<T>>,
-                "Template argument must be a raw or supported smart pointer "
-                "type. See absl/base/nullability.h.");
-  using type = T;
-};
-
-// Note: we do not apply Clang nullability attributes (e.g. _Nullable).  These
-// only support raw pointers, and conditionally enabling them only for raw
-// pointers inhibits template arg deduction.  Ideally, they would support all
-// pointer-like types.
-template <typename T, typename = typename EnableNullable<T>::type>
-using NullableImpl
-#if ABSL_HAVE_CPP_ATTRIBUTE(clang::annotate)
-    [[clang::annotate("Nullable")]]
-#endif
-    = T;
-
-template <typename T, typename = typename EnableNonnull<T>::type>
-using NonnullImpl
-#if ABSL_HAVE_CPP_ATTRIBUTE(clang::annotate)
-    [[clang::annotate("Nonnull")]]
-#endif
-    = T;
-
-template <typename T, typename = typename EnableNullabilityUnknown<T>::type>
-using NullabilityUnknownImpl
-#if ABSL_HAVE_CPP_ATTRIBUTE(clang::annotate)
-    [[clang::annotate("Nullability_Unspecified")]]
-#endif
-    = T;
-
-}  // namespace nullability_internal
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_NULLABILITY_IMPL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/per_thread_tls.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/per_thread_tls.h
deleted file mode 100644
index cf5e97a0..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/per_thread_tls.h
+++ /dev/null
@@ -1,52 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_BASE_INTERNAL_PER_THREAD_TLS_H_
-#define ABSL_BASE_INTERNAL_PER_THREAD_TLS_H_
-
-// This header defines two macros:
-//
-// If the platform supports thread-local storage:
-//
-// * ABSL_PER_THREAD_TLS_KEYWORD is the C keyword needed to declare a
-//   thread-local variable
-// * ABSL_PER_THREAD_TLS is 1
-//
-// Otherwise:
-//
-// * ABSL_PER_THREAD_TLS_KEYWORD is empty
-// * ABSL_PER_THREAD_TLS is 0
-//
-// Microsoft C supports thread-local storage.
-// GCC supports it if the appropriate version of glibc is available,
-// which the programmer can indicate by defining ABSL_HAVE_TLS
-
-#include "absl/base/port.h"  // For ABSL_HAVE_TLS
-
-#if defined(ABSL_PER_THREAD_TLS)
-#error ABSL_PER_THREAD_TLS cannot be directly set
-#elif defined(ABSL_PER_THREAD_TLS_KEYWORD)
-#error ABSL_PER_THREAD_TLS_KEYWORD cannot be directly set
-#elif defined(ABSL_HAVE_TLS)
-#define ABSL_PER_THREAD_TLS_KEYWORD __thread
-#define ABSL_PER_THREAD_TLS 1
-#elif defined(_MSC_VER)
-#define ABSL_PER_THREAD_TLS_KEYWORD __declspec(thread)
-#define ABSL_PER_THREAD_TLS 1
-#else
-#define ABSL_PER_THREAD_TLS_KEYWORD
-#define ABSL_PER_THREAD_TLS 0
-#endif
-
-#endif  // ABSL_BASE_INTERNAL_PER_THREAD_TLS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/pretty_function.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/pretty_function.h
deleted file mode 100644
index 35d51676..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/pretty_function.h
+++ /dev/null
@@ -1,33 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_BASE_INTERNAL_PRETTY_FUNCTION_H_
-#define ABSL_BASE_INTERNAL_PRETTY_FUNCTION_H_
-
-// ABSL_PRETTY_FUNCTION
-//
-// In C++11, __func__ gives the undecorated name of the current function.  That
-// is, "main", not "int main()".  Various compilers give extra macros to get the
-// decorated function name, including return type and arguments, to
-// differentiate between overload sets.  ABSL_PRETTY_FUNCTION is a portable
-// version of these macros which forwards to the correct macro on each compiler.
-#if defined(_MSC_VER)
-#define ABSL_PRETTY_FUNCTION __FUNCSIG__
-#elif defined(__GNUC__)
-#define ABSL_PRETTY_FUNCTION __PRETTY_FUNCTION__
-#else
-#error "Unsupported compiler"
-#endif
-
-#endif  // ABSL_BASE_INTERNAL_PRETTY_FUNCTION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/raw_logging.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/raw_logging.h
deleted file mode 100644
index d7cfbc57..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/raw_logging.h
+++ /dev/null
@@ -1,217 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Thread-safe logging routines that do not allocate any memory or
-// acquire any locks, and can therefore be used by low-level memory
-// allocation, synchronization, and signal-handling code.
-
-#ifndef ABSL_BASE_INTERNAL_RAW_LOGGING_H_
-#define ABSL_BASE_INTERNAL_RAW_LOGGING_H_
-
-#include <string>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/internal/atomic_hook.h"
-#include "absl/base/log_severity.h"
-#include "absl/base/macros.h"
-#include "absl/base/optimization.h"
-#include "absl/base/port.h"
-
-// This is similar to LOG(severity) << format..., but
-// * it is to be used ONLY by low-level modules that can't use normal LOG()
-// * it is designed to be a low-level logger that does not allocate any
-//   memory and does not need any locks, hence:
-// * it logs straight and ONLY to STDERR w/o buffering
-// * it uses an explicit printf-format and arguments list
-// * it will silently chop off really long message strings
-// Usage example:
-//   ABSL_RAW_LOG(ERROR, "Failed foo with %i: %s", status, error);
-// This will print an almost standard log line like this to stderr only:
-//   E0821 211317 file.cc:123] RAW: Failed foo with 22: bad_file
-
-#define ABSL_RAW_LOG(severity, ...)                                            \
-  do {                                                                         \
-    constexpr const char* absl_raw_log_internal_basename =                     \
-        ::absl::raw_log_internal::Basename(__FILE__, sizeof(__FILE__) - 1);    \
-    ::absl::raw_log_internal::RawLog(ABSL_RAW_LOG_INTERNAL_##severity,         \
-                                     absl_raw_log_internal_basename, __LINE__, \
-                                     __VA_ARGS__);                             \
-    ABSL_RAW_LOG_INTERNAL_MAYBE_UNREACHABLE_##severity;                        \
-  } while (0)
-
-// Similar to CHECK(condition) << message, but for low-level modules:
-// we use only ABSL_RAW_LOG that does not allocate memory.
-// We do not want to provide args list here to encourage this usage:
-//   if (!cond)  ABSL_RAW_LOG(FATAL, "foo ...", hard_to_compute_args);
-// so that the args are not computed when not needed.
-#define ABSL_RAW_CHECK(condition, message)                             \
-  do {                                                                 \
-    if (ABSL_PREDICT_FALSE(!(condition))) {                            \
-      ABSL_RAW_LOG(FATAL, "Check %s failed: %s", #condition, message); \
-    }                                                                  \
-  } while (0)
-
-// ABSL_INTERNAL_LOG and ABSL_INTERNAL_CHECK work like the RAW variants above,
-// except that if the richer log library is linked into the binary, we dispatch
-// to that instead.  This is potentially useful for internal logging and
-// assertions, where we are using RAW_LOG neither for its async-signal-safety
-// nor for its non-allocating nature, but rather because raw logging has very
-// few other dependencies.
-//
-// The API is a subset of the above: each macro only takes two arguments.  Use
-// StrCat if you need to build a richer message.
-#define ABSL_INTERNAL_LOG(severity, message)                              \
-  do {                                                                    \
-    constexpr const char* absl_raw_log_internal_filename = __FILE__;      \
-    ::absl::raw_log_internal::internal_log_function(                      \
-        ABSL_RAW_LOG_INTERNAL_##severity, absl_raw_log_internal_filename, \
-        __LINE__, message);                                               \
-    ABSL_RAW_LOG_INTERNAL_MAYBE_UNREACHABLE_##severity;                   \
-  } while (0)
-
-#define ABSL_INTERNAL_CHECK(condition, message)                    \
-  do {                                                             \
-    if (ABSL_PREDICT_FALSE(!(condition))) {                        \
-      std::string death_message = "Check " #condition " failed: "; \
-      death_message += std::string(message);                       \
-      ABSL_INTERNAL_LOG(FATAL, death_message);                     \
-    }                                                              \
-  } while (0)
-
-#ifndef NDEBUG
-
-#define ABSL_RAW_DLOG(severity, ...) ABSL_RAW_LOG(severity, __VA_ARGS__)
-#define ABSL_RAW_DCHECK(condition, message) ABSL_RAW_CHECK(condition, message)
-
-#else  // NDEBUG
-
-#define ABSL_RAW_DLOG(severity, ...)                   \
-  while (false) ABSL_RAW_LOG(severity, __VA_ARGS__)
-#define ABSL_RAW_DCHECK(condition, message) \
-  while (false) ABSL_RAW_CHECK(condition, message)
-
-#endif  // NDEBUG
-
-#define ABSL_RAW_LOG_INTERNAL_INFO ::absl::LogSeverity::kInfo
-#define ABSL_RAW_LOG_INTERNAL_WARNING ::absl::LogSeverity::kWarning
-#define ABSL_RAW_LOG_INTERNAL_ERROR ::absl::LogSeverity::kError
-#define ABSL_RAW_LOG_INTERNAL_FATAL ::absl::LogSeverity::kFatal
-#define ABSL_RAW_LOG_INTERNAL_DFATAL ::absl::kLogDebugFatal
-#define ABSL_RAW_LOG_INTERNAL_LEVEL(severity) \
-  ::absl::NormalizeLogSeverity(severity)
-
-#define ABSL_RAW_LOG_INTERNAL_MAYBE_UNREACHABLE_INFO
-#define ABSL_RAW_LOG_INTERNAL_MAYBE_UNREACHABLE_WARNING
-#define ABSL_RAW_LOG_INTERNAL_MAYBE_UNREACHABLE_ERROR
-#define ABSL_RAW_LOG_INTERNAL_MAYBE_UNREACHABLE_FATAL ABSL_UNREACHABLE()
-#define ABSL_RAW_LOG_INTERNAL_MAYBE_UNREACHABLE_DFATAL
-#define ABSL_RAW_LOG_INTERNAL_MAYBE_UNREACHABLE_LEVEL(severity)
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace raw_log_internal {
-
-// Helper function to implement ABSL_RAW_LOG
-// Logs format... at "severity" level, reporting it
-// as called from file:line.
-// This does not allocate memory or acquire locks.
-void RawLog(absl::LogSeverity severity, const char* file, int line,
-            const char* format, ...) ABSL_PRINTF_ATTRIBUTE(4, 5);
-
-// Writes the provided buffer directly to stderr, in a signal-safe, low-level
-// manner.  Preserves errno.
-void AsyncSignalSafeWriteError(const char* s, size_t len);
-
-// compile-time function to get the "base" filename, that is, the part of
-// a filename after the last "/" or "\" path separator.  The search starts at
-// the end of the string; the second parameter is the length of the string.
-constexpr const char* Basename(const char* fname, int offset) {
-  return offset == 0 || fname[offset - 1] == '/' || fname[offset - 1] == '\\'
-             ? fname + offset
-             : Basename(fname, offset - 1);
-}
-
-// For testing only.
-// Returns true if raw logging is fully supported. When it is not
-// fully supported, no messages will be emitted, but a log at FATAL
-// severity will cause an abort.
-//
-// TODO(gfalcon): Come up with a better name for this method.
-bool RawLoggingFullySupported();
-
-// Function type for a raw_log customization hook for suppressing messages
-// by severity, and for writing custom prefixes on non-suppressed messages.
-//
-// The installed hook is called for every raw log invocation.  The message will
-// be logged to stderr only if the hook returns true.  FATAL errors will cause
-// the process to abort, even if writing to stderr is suppressed.  The hook is
-// also provided with an output buffer, where it can write a custom log message
-// prefix.
-//
-// The raw_log system does not allocate memory or grab locks.  User-provided
-// hooks must avoid these operations, and must not throw exceptions.
-//
-// 'severity' is the severity level of the message being written.
-// 'file' and 'line' are the file and line number where the ABSL_RAW_LOG macro
-// was located.
-// 'buf' and 'buf_size' are pointers to the buffer and buffer size.  If the
-// hook writes a prefix, it must increment *buf and decrement *buf_size
-// accordingly.
-using LogFilterAndPrefixHook = bool (*)(absl::LogSeverity severity,
-                                        const char* file, int line, char** buf,
-                                        int* buf_size);
-
-// Function type for a raw_log customization hook called to abort a process
-// when a FATAL message is logged.  If the provided AbortHook() returns, the
-// logging system will call abort().
-//
-// 'file' and 'line' are the file and line number where the ABSL_RAW_LOG macro
-// was located.
-// The NUL-terminated logged message lives in the buffer between 'buf_start'
-// and 'buf_end'.  'prefix_end' points to the first non-prefix character of the
-// buffer (as written by the LogFilterAndPrefixHook.)
-//
-// The lifetime of the filename and message buffers will not end while the
-// process remains alive.
-using AbortHook = void (*)(const char* file, int line, const char* buf_start,
-                           const char* prefix_end, const char* buf_end);
-
-// Internal logging function for ABSL_INTERNAL_LOG to dispatch to.
-//
-// TODO(gfalcon): When string_view no longer depends on base, change this
-// interface to take its message as a string_view instead.
-using InternalLogFunction = void (*)(absl::LogSeverity severity,
-                                     const char* file, int line,
-                                     const std::string& message);
-
-ABSL_INTERNAL_ATOMIC_HOOK_ATTRIBUTES ABSL_DLL extern base_internal::AtomicHook<
-    InternalLogFunction>
-    internal_log_function;
-
-// Registers hooks of the above types.  Only a single hook of each type may be
-// registered.  It is an error to call these functions multiple times with
-// different input arguments.
-//
-// These functions are safe to call at any point during initialization; they do
-// not block or malloc, and are async-signal safe.
-void RegisterLogFilterAndPrefixHook(LogFilterAndPrefixHook func);
-void RegisterAbortHook(AbortHook func);
-void RegisterInternalLogFunction(InternalLogFunction func);
-
-}  // namespace raw_log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_RAW_LOGGING_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/scheduling_mode.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/scheduling_mode.h
deleted file mode 100644
index 8be5ab6d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/scheduling_mode.h
+++ /dev/null
@@ -1,58 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Core interfaces and definitions used by by low-level interfaces such as
-// SpinLock.
-
-#ifndef ABSL_BASE_INTERNAL_SCHEDULING_MODE_H_
-#define ABSL_BASE_INTERNAL_SCHEDULING_MODE_H_
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-// Used to describe how a thread may be scheduled.  Typically associated with
-// the declaration of a resource supporting synchronized access.
-//
-// SCHEDULE_COOPERATIVE_AND_KERNEL:
-// Specifies that when waiting, a cooperative thread (e.g. a Fiber) may
-// reschedule (using base::scheduling semantics); allowing other cooperative
-// threads to proceed.
-//
-// SCHEDULE_KERNEL_ONLY: (Also described as "non-cooperative")
-// Specifies that no cooperative scheduling semantics may be used, even if the
-// current thread is itself cooperatively scheduled.  This means that
-// cooperative threads will NOT allow other cooperative threads to execute in
-// their place while waiting for a resource of this type.  Host operating system
-// semantics (e.g. a futex) may still be used.
-//
-// When optional, clients should strongly prefer SCHEDULE_COOPERATIVE_AND_KERNEL
-// by default.  SCHEDULE_KERNEL_ONLY should only be used for resources on which
-// base::scheduling (e.g. the implementation of a Scheduler) may depend.
-//
-// NOTE: Cooperative resources may not be nested below non-cooperative ones.
-// This means that it is invalid to to acquire a SCHEDULE_COOPERATIVE_AND_KERNEL
-// resource if a SCHEDULE_KERNEL_ONLY resource is already held.
-enum SchedulingMode {
-  SCHEDULE_KERNEL_ONLY = 0,         // Allow scheduling only the host OS.
-  SCHEDULE_COOPERATIVE_AND_KERNEL,  // Also allow cooperative scheduling.
-};
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_SCHEDULING_MODE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/scoped_set_env.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/scoped_set_env.h
deleted file mode 100644
index 19ec7b5d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/scoped_set_env.h
+++ /dev/null
@@ -1,45 +0,0 @@
-//
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_BASE_INTERNAL_SCOPED_SET_ENV_H_
-#define ABSL_BASE_INTERNAL_SCOPED_SET_ENV_H_
-
-#include <string>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-class ScopedSetEnv {
- public:
-  ScopedSetEnv(const char* var_name, const char* new_value);
-  ~ScopedSetEnv();
-
- private:
-  std::string var_name_;
-  std::string old_value_;
-
-  // True if the environment variable was initially not set.
-  bool was_unset_;
-};
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_SCOPED_SET_ENV_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/spinlock.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/spinlock.h
deleted file mode 100644
index 301c3e3b..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/spinlock.h
+++ /dev/null
@@ -1,274 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-//  Most users requiring mutual exclusion should use Mutex.
-//  SpinLock is provided for use in two situations:
-//   - for use by Abseil internal code that Mutex itself depends on
-//   - for async signal safety (see below)
-
-// SpinLock with a base_internal::SchedulingMode::SCHEDULE_KERNEL_ONLY is async
-// signal safe. If a spinlock is used within a signal handler, all code that
-// acquires the lock must ensure that the signal cannot arrive while they are
-// holding the lock. Typically, this is done by blocking the signal.
-//
-// Threads waiting on a SpinLock may be woken in an arbitrary order.
-
-#ifndef ABSL_BASE_INTERNAL_SPINLOCK_H_
-#define ABSL_BASE_INTERNAL_SPINLOCK_H_
-
-#include <atomic>
-#include <cstdint>
-
-#include "absl/base/attributes.h"
-#include "absl/base/const_init.h"
-#include "absl/base/dynamic_annotations.h"
-#include "absl/base/internal/low_level_scheduling.h"
-#include "absl/base/internal/raw_logging.h"
-#include "absl/base/internal/scheduling_mode.h"
-#include "absl/base/internal/tsan_mutex_interface.h"
-#include "absl/base/thread_annotations.h"
-
-namespace tcmalloc {
-namespace tcmalloc_internal {
-
-class AllocationGuardSpinLockHolder;
-
-}  // namespace tcmalloc_internal
-}  // namespace tcmalloc
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-class ABSL_LOCKABLE ABSL_ATTRIBUTE_WARN_UNUSED SpinLock {
- public:
-  SpinLock() : lockword_(kSpinLockCooperative) {
-    ABSL_TSAN_MUTEX_CREATE(this, __tsan_mutex_not_static);
-  }
-
-  // Constructors that allow non-cooperative spinlocks to be created for use
-  // inside thread schedulers.  Normal clients should not use these.
-  explicit SpinLock(base_internal::SchedulingMode mode);
-
-  // Constructor for global SpinLock instances.  See absl/base/const_init.h.
-  constexpr SpinLock(absl::ConstInitType, base_internal::SchedulingMode mode)
-      : lockword_(IsCooperative(mode) ? kSpinLockCooperative : 0) {}
-
-  // For global SpinLock instances prefer trivial destructor when possible.
-  // Default but non-trivial destructor in some build configurations causes an
-  // extra static initializer.
-#ifdef ABSL_INTERNAL_HAVE_TSAN_INTERFACE
-  ~SpinLock() { ABSL_TSAN_MUTEX_DESTROY(this, __tsan_mutex_not_static); }
-#else
-  ~SpinLock() = default;
-#endif
-
-  // Acquire this SpinLock.
-  inline void Lock() ABSL_EXCLUSIVE_LOCK_FUNCTION() {
-    ABSL_TSAN_MUTEX_PRE_LOCK(this, 0);
-    if (!TryLockImpl()) {
-      SlowLock();
-    }
-    ABSL_TSAN_MUTEX_POST_LOCK(this, 0, 0);
-  }
-
-  // Try to acquire this SpinLock without blocking and return true if the
-  // acquisition was successful.  If the lock was not acquired, false is
-  // returned.  If this SpinLock is free at the time of the call, TryLock
-  // will return true with high probability.
-  inline bool TryLock() ABSL_EXCLUSIVE_TRYLOCK_FUNCTION(true) {
-    ABSL_TSAN_MUTEX_PRE_LOCK(this, __tsan_mutex_try_lock);
-    bool res = TryLockImpl();
-    ABSL_TSAN_MUTEX_POST_LOCK(
-        this, __tsan_mutex_try_lock | (res ? 0 : __tsan_mutex_try_lock_failed),
-        0);
-    return res;
-  }
-
-  // Release this SpinLock, which must be held by the calling thread.
-  inline void Unlock() ABSL_UNLOCK_FUNCTION() {
-    ABSL_TSAN_MUTEX_PRE_UNLOCK(this, 0);
-    uint32_t lock_value = lockword_.load(std::memory_order_relaxed);
-    lock_value = lockword_.exchange(lock_value & kSpinLockCooperative,
-                                    std::memory_order_release);
-
-    if ((lock_value & kSpinLockDisabledScheduling) != 0) {
-      base_internal::SchedulingGuard::EnableRescheduling(true);
-    }
-    if ((lock_value & kWaitTimeMask) != 0) {
-      // Collect contentionz profile info, and speed the wakeup of any waiter.
-      // The wait_cycles value indicates how long this thread spent waiting
-      // for the lock.
-      SlowUnlock(lock_value);
-    }
-    ABSL_TSAN_MUTEX_POST_UNLOCK(this, 0);
-  }
-
-  // Determine if the lock is held.  When the lock is held by the invoking
-  // thread, true will always be returned. Intended to be used as
-  // CHECK(lock.IsHeld()).
-  inline bool IsHeld() const {
-    return (lockword_.load(std::memory_order_relaxed) & kSpinLockHeld) != 0;
-  }
-
-  // Return immediately if this thread holds the SpinLock exclusively.
-  // Otherwise, report an error by crashing with a diagnostic.
-  inline void AssertHeld() const ABSL_ASSERT_EXCLUSIVE_LOCK() {
-    if (!IsHeld()) {
-      ABSL_RAW_LOG(FATAL, "thread should hold the lock on SpinLock");
-    }
-  }
-
- protected:
-  // These should not be exported except for testing.
-
-  // Store number of cycles between wait_start_time and wait_end_time in a
-  // lock value.
-  static uint32_t EncodeWaitCycles(int64_t wait_start_time,
-                                   int64_t wait_end_time);
-
-  // Extract number of wait cycles in a lock value.
-  static int64_t DecodeWaitCycles(uint32_t lock_value);
-
-  // Provide access to protected method above.  Use for testing only.
-  friend struct SpinLockTest;
-  friend class tcmalloc::tcmalloc_internal::AllocationGuardSpinLockHolder;
-
- private:
-  // lockword_ is used to store the following:
-  //
-  // bit[0] encodes whether a lock is being held.
-  // bit[1] encodes whether a lock uses cooperative scheduling.
-  // bit[2] encodes whether the current lock holder disabled scheduling when
-  //        acquiring the lock. Only set when kSpinLockHeld is also set.
-  // bit[3:31] encodes time a lock spent on waiting as a 29-bit unsigned int.
-  //        This is set by the lock holder to indicate how long it waited on
-  //        the lock before eventually acquiring it. The number of cycles is
-  //        encoded as a 29-bit unsigned int, or in the case that the current
-  //        holder did not wait but another waiter is queued, the LSB
-  //        (kSpinLockSleeper) is set. The implementation does not explicitly
-  //        track the number of queued waiters beyond this. It must always be
-  //        assumed that waiters may exist if the current holder was required to
-  //        queue.
-  //
-  // Invariant: if the lock is not held, the value is either 0 or
-  // kSpinLockCooperative.
-  static constexpr uint32_t kSpinLockHeld = 1;
-  static constexpr uint32_t kSpinLockCooperative = 2;
-  static constexpr uint32_t kSpinLockDisabledScheduling = 4;
-  static constexpr uint32_t kSpinLockSleeper = 8;
-  // Includes kSpinLockSleeper.
-  static constexpr uint32_t kWaitTimeMask =
-      ~(kSpinLockHeld | kSpinLockCooperative | kSpinLockDisabledScheduling);
-
-  // Returns true if the provided scheduling mode is cooperative.
-  static constexpr bool IsCooperative(
-      base_internal::SchedulingMode scheduling_mode) {
-    return scheduling_mode == base_internal::SCHEDULE_COOPERATIVE_AND_KERNEL;
-  }
-
-  bool IsCooperative() const {
-    return lockword_.load(std::memory_order_relaxed) & kSpinLockCooperative;
-  }
-
-  uint32_t TryLockInternal(uint32_t lock_value, uint32_t wait_cycles);
-  void SlowLock() ABSL_ATTRIBUTE_COLD;
-  void SlowUnlock(uint32_t lock_value) ABSL_ATTRIBUTE_COLD;
-  uint32_t SpinLoop();
-
-  inline bool TryLockImpl() {
-    uint32_t lock_value = lockword_.load(std::memory_order_relaxed);
-    return (TryLockInternal(lock_value, 0) & kSpinLockHeld) == 0;
-  }
-
-  std::atomic<uint32_t> lockword_;
-
-  SpinLock(const SpinLock&) = delete;
-  SpinLock& operator=(const SpinLock&) = delete;
-};
-
-// Corresponding locker object that arranges to acquire a spinlock for
-// the duration of a C++ scope.
-//
-// TODO(b/176172494): Use only [[nodiscard]] when baseline is raised.
-// TODO(b/6695610): Remove forward declaration when #ifdef is no longer needed.
-#if ABSL_HAVE_CPP_ATTRIBUTE(nodiscard)
-class [[nodiscard]] SpinLockHolder;
-#else
-class ABSL_MUST_USE_RESULT ABSL_ATTRIBUTE_TRIVIAL_ABI SpinLockHolder;
-#endif
-
-class ABSL_SCOPED_LOCKABLE SpinLockHolder {
- public:
-  inline explicit SpinLockHolder(SpinLock* l) ABSL_EXCLUSIVE_LOCK_FUNCTION(l)
-      : lock_(l) {
-    l->Lock();
-  }
-  inline ~SpinLockHolder() ABSL_UNLOCK_FUNCTION() { lock_->Unlock(); }
-
-  SpinLockHolder(const SpinLockHolder&) = delete;
-  SpinLockHolder& operator=(const SpinLockHolder&) = delete;
-
- private:
-  SpinLock* lock_;
-};
-
-// Register a hook for profiling support.
-//
-// The function pointer registered here will be called whenever a spinlock is
-// contended.  The callback is given an opaque handle to the contended spinlock
-// and the number of wait cycles.  This is thread-safe, but only a single
-// profiler can be registered.  It is an error to call this function multiple
-// times with different arguments.
-void RegisterSpinLockProfiler(void (*fn)(const void* lock,
-                                         int64_t wait_cycles));
-
-//------------------------------------------------------------------------------
-// Public interface ends here.
-//------------------------------------------------------------------------------
-
-// If (result & kSpinLockHeld) == 0, then *this was successfully locked.
-// Otherwise, returns last observed value for lockword_.
-inline uint32_t SpinLock::TryLockInternal(uint32_t lock_value,
-                                          uint32_t wait_cycles) {
-  if ((lock_value & kSpinLockHeld) != 0) {
-    return lock_value;
-  }
-
-  uint32_t sched_disabled_bit = 0;
-  if ((lock_value & kSpinLockCooperative) == 0) {
-    // For non-cooperative locks we must make sure we mark ourselves as
-    // non-reschedulable before we attempt to CompareAndSwap.
-    if (base_internal::SchedulingGuard::DisableRescheduling()) {
-      sched_disabled_bit = kSpinLockDisabledScheduling;
-    }
-  }
-
-  if (!lockword_.compare_exchange_strong(
-          lock_value,
-          kSpinLockHeld | lock_value | wait_cycles | sched_disabled_bit,
-          std::memory_order_acquire, std::memory_order_relaxed)) {
-    base_internal::SchedulingGuard::EnableRescheduling(sched_disabled_bit != 0);
-  }
-
-  return lock_value;
-}
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_SPINLOCK_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/spinlock_wait.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/spinlock_wait.h
deleted file mode 100644
index 9a1adcda..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/spinlock_wait.h
+++ /dev/null
@@ -1,95 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_BASE_INTERNAL_SPINLOCK_WAIT_H_
-#define ABSL_BASE_INTERNAL_SPINLOCK_WAIT_H_
-
-// Operations to make atomic transitions on a word, and to allow
-// waiting for those transitions to become possible.
-
-#include <stdint.h>
-#include <atomic>
-
-#include "absl/base/internal/scheduling_mode.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-// SpinLockWait() waits until it can perform one of several transitions from
-// "from" to "to".  It returns when it performs a transition where done==true.
-struct SpinLockWaitTransition {
-  uint32_t from;
-  uint32_t to;
-  bool done;
-};
-
-// Wait until *w can transition from trans[i].from to trans[i].to for some i
-// satisfying 0<=i<n && trans[i].done, atomically make the transition,
-// then return the old value of *w.   Make any other atomic transitions
-// where !trans[i].done, but continue waiting.
-//
-// Wakeups for threads blocked on SpinLockWait do not respect priorities.
-uint32_t SpinLockWait(std::atomic<uint32_t> *w, int n,
-                      const SpinLockWaitTransition trans[],
-                      SchedulingMode scheduling_mode);
-
-// If possible, wake some thread that has called SpinLockDelay(w, ...). If `all`
-// is true, wake all such threads. On some systems, this may be a no-op; on
-// those systems, threads calling SpinLockDelay() will always wake eventually
-// even if SpinLockWake() is never called.
-void SpinLockWake(std::atomic<uint32_t> *w, bool all);
-
-// Wait for an appropriate spin delay on iteration "loop" of a
-// spin loop on location *w, whose previously observed value was "value".
-// SpinLockDelay() may do nothing, may yield the CPU, may sleep a clock tick,
-// or may wait for a call to SpinLockWake(w).
-void SpinLockDelay(std::atomic<uint32_t> *w, uint32_t value, int loop,
-                   base_internal::SchedulingMode scheduling_mode);
-
-// Helper used by AbslInternalSpinLockDelay.
-// Returns a suggested delay in nanoseconds for iteration number "loop".
-int SpinLockSuggestedDelayNS(int loop);
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-// In some build configurations we pass --detect-odr-violations to the
-// gold linker.  This causes it to flag weak symbol overrides as ODR
-// violations.  Because ODR only applies to C++ and not C,
-// --detect-odr-violations ignores symbols not mangled with C++ names.
-// By changing our extension points to be extern "C", we dodge this
-// check.
-extern "C" {
-void ABSL_INTERNAL_C_SYMBOL(AbslInternalSpinLockWake)(std::atomic<uint32_t> *w,
-                                                      bool all);
-void ABSL_INTERNAL_C_SYMBOL(AbslInternalSpinLockDelay)(
-    std::atomic<uint32_t> *w, uint32_t value, int loop,
-    absl::base_internal::SchedulingMode scheduling_mode);
-}
-
-inline void absl::base_internal::SpinLockWake(std::atomic<uint32_t> *w,
-                                              bool all) {
-  ABSL_INTERNAL_C_SYMBOL(AbslInternalSpinLockWake)(w, all);
-}
-
-inline void absl::base_internal::SpinLockDelay(
-    std::atomic<uint32_t> *w, uint32_t value, int loop,
-    absl::base_internal::SchedulingMode scheduling_mode) {
-  ABSL_INTERNAL_C_SYMBOL(AbslInternalSpinLockDelay)
-  (w, value, loop, scheduling_mode);
-}
-
-#endif  // ABSL_BASE_INTERNAL_SPINLOCK_WAIT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/strerror.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/strerror.h
deleted file mode 100644
index 35009736..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/strerror.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_BASE_INTERNAL_STRERROR_H_
-#define ABSL_BASE_INTERNAL_STRERROR_H_
-
-#include <string>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-// A portable and thread-safe alternative to C89's `strerror`.
-//
-// The C89 specification of `strerror` is not suitable for use in a
-// multi-threaded application as the returned string may be changed by calls to
-// `strerror` from another thread.  The many non-stdlib alternatives differ
-// enough in their names, availability, and semantics to justify this wrapper
-// around them.  `errno` will not be modified by a call to `absl::StrError`.
-std::string StrError(int errnum);
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_STRERROR_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/sysinfo.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/sysinfo.h
deleted file mode 100644
index 119cf1f0..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/sysinfo.h
+++ /dev/null
@@ -1,74 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// This file includes routines to find out characteristics
-// of the machine a program is running on.  It is undoubtedly
-// system-dependent.
-
-// Functions listed here that accept a pid_t as an argument act on the
-// current process if the pid_t argument is 0
-// All functions here are thread-hostile due to file caching unless
-// commented otherwise.
-
-#ifndef ABSL_BASE_INTERNAL_SYSINFO_H_
-#define ABSL_BASE_INTERNAL_SYSINFO_H_
-
-#ifndef _WIN32
-#include <sys/types.h>
-#endif
-
-#include <cstdint>
-
-#include "absl/base/config.h"
-#include "absl/base/port.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-// Nominal core processor cycles per second of each processor.   This is _not_
-// necessarily the frequency of the CycleClock counter (see cycleclock.h)
-// Thread-safe.
-double NominalCPUFrequency();
-
-// Number of logical processors (hyperthreads) in system. Thread-safe.
-int NumCPUs();
-
-// Return the thread id of the current thread, as told by the system.
-// No two currently-live threads implemented by the OS shall have the same ID.
-// Thread ids of exited threads may be reused.   Multiple user-level threads
-// may have the same thread ID if multiplexed on the same OS thread.
-//
-// On Linux, you may send a signal to the resulting ID with kill().  However,
-// it is recommended for portability that you use pthread_kill() instead.
-#ifdef _WIN32
-// On Windows, process id and thread id are of the same type according to the
-// return types of GetProcessId() and GetThreadId() are both DWORD, an unsigned
-// 32-bit type.
-using pid_t = uint32_t;
-#endif
-pid_t GetTID();
-
-// Like GetTID(), but caches the result in thread-local storage in order
-// to avoid unnecessary system calls. Note that there are some cases where
-// one must call through to GetTID directly, which is why this exists as a
-// separate function. For example, GetCachedTID() is not safe to call in
-// an asynchronous signal-handling context nor right after a call to fork().
-pid_t GetCachedTID();
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_SYSINFO_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/thread_identity.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/thread_identity.h
deleted file mode 100644
index b6e917ce..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/thread_identity.h
+++ /dev/null
@@ -1,269 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Each active thread has an ThreadIdentity that may represent the thread in
-// various level interfaces.  ThreadIdentity objects are never deallocated.
-// When a thread terminates, its ThreadIdentity object may be reused for a
-// thread created later.
-
-#ifndef ABSL_BASE_INTERNAL_THREAD_IDENTITY_H_
-#define ABSL_BASE_INTERNAL_THREAD_IDENTITY_H_
-
-#ifndef _WIN32
-#include <pthread.h>
-// Defines __GOOGLE_GRTE_VERSION__ (via glibc-specific features.h) when
-// supported.
-#include <unistd.h>
-#endif
-
-#include <atomic>
-#include <cstdint>
-
-#include "absl/base/config.h"
-#include "absl/base/internal/per_thread_tls.h"
-#include "absl/base/optimization.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-struct SynchLocksHeld;
-struct SynchWaitParams;
-
-namespace base_internal {
-
-class SpinLock;
-struct ThreadIdentity;
-
-// Used by the implementation of absl::Mutex and absl::CondVar.
-struct PerThreadSynch {
-  // The internal representation of absl::Mutex and absl::CondVar rely
-  // on the alignment of PerThreadSynch. Both store the address of the
-  // PerThreadSynch in the high-order bits of their internal state,
-  // which means the low kLowZeroBits of the address of PerThreadSynch
-  // must be zero.
-  static constexpr int kLowZeroBits = 8;
-  static constexpr int kAlignment = 1 << kLowZeroBits;
-
-  // Returns the associated ThreadIdentity.
-  // This can be implemented as a cast because we guarantee
-  // PerThreadSynch is the first element of ThreadIdentity.
-  ThreadIdentity* thread_identity() {
-    return reinterpret_cast<ThreadIdentity*>(this);
-  }
-
-  PerThreadSynch* next;  // Circular waiter queue; initialized to 0.
-  PerThreadSynch* skip;  // If non-zero, all entries in Mutex queue
-                         // up to and including "skip" have same
-                         // condition as this, and will be woken later
-  bool may_skip;         // if false while on mutex queue, a mutex unlocker
-                         // is using this PerThreadSynch as a terminator.  Its
-                         // skip field must not be filled in because the loop
-                         // might then skip over the terminator.
-  bool wake;             // This thread is to be woken from a Mutex.
-  // If "x" is on a waiter list for a mutex, "x->cond_waiter" is true iff the
-  // waiter is waiting on the mutex as part of a CV Wait or Mutex Await.
-  //
-  // The value of "x->cond_waiter" is meaningless if "x" is not on a
-  // Mutex waiter list.
-  bool cond_waiter;
-  bool maybe_unlocking;  // Valid at head of Mutex waiter queue;
-                         // true if UnlockSlow could be searching
-                         // for a waiter to wake.  Used for an optimization
-                         // in Enqueue().  true is always a valid value.
-                         // Can be reset to false when the unlocker or any
-                         // writer releases the lock, or a reader fully
-                         // releases the lock.  It may not be set to false
-                         // by a reader that decrements the count to
-                         // non-zero. protected by mutex spinlock
-  bool suppress_fatal_errors;  // If true, try to proceed even in the face
-                               // of broken invariants.  This is used within
-                               // fatal signal handlers to improve the
-                               // chances of debug logging information being
-                               // output successfully.
-  int priority;                // Priority of thread (updated every so often).
-
-  // State values:
-  //   kAvailable: This PerThreadSynch is available.
-  //   kQueued: This PerThreadSynch is unavailable, it's currently queued on a
-  //            Mutex or CondVar waistlist.
-  //
-  // Transitions from kQueued to kAvailable require a release
-  // barrier. This is needed as a waiter may use "state" to
-  // independently observe that it's no longer queued.
-  //
-  // Transitions from kAvailable to kQueued require no barrier, they
-  // are externally ordered by the Mutex.
-  enum State { kAvailable, kQueued };
-  std::atomic<State> state;
-
-  // The wait parameters of the current wait.  waitp is null if the
-  // thread is not waiting. Transitions from null to non-null must
-  // occur before the enqueue commit point (state = kQueued in
-  // Enqueue() and CondVarEnqueue()). Transitions from non-null to
-  // null must occur after the wait is finished (state = kAvailable in
-  // Mutex::Block() and CondVar::WaitCommon()). This field may be
-  // changed only by the thread that describes this PerThreadSynch.  A
-  // special case is Fer(), which calls Enqueue() on another thread,
-  // but with an identical SynchWaitParams pointer, thus leaving the
-  // pointer unchanged.
-  SynchWaitParams* waitp;
-
-  intptr_t readers;  // Number of readers in mutex.
-
-  // When priority will next be read (cycles).
-  int64_t next_priority_read_cycles;
-
-  // Locks held; used during deadlock detection.
-  // Allocated in Synch_GetAllLocks() and freed in ReclaimThreadIdentity().
-  SynchLocksHeld* all_locks;
-};
-
-// The instances of this class are allocated in NewThreadIdentity() with an
-// alignment of PerThreadSynch::kAlignment.
-//
-// NOTE: The layout of fields in this structure is critical, please do not
-//       add, remove, or modify the field placements without fully auditing the
-//       layout.
-struct ThreadIdentity {
-  // Must be the first member.  The Mutex implementation requires that
-  // the PerThreadSynch object associated with each thread is
-  // PerThreadSynch::kAlignment aligned.  We provide this alignment on
-  // ThreadIdentity itself.
-  PerThreadSynch per_thread_synch;
-
-  // Private: Reserved for absl::synchronization_internal::Waiter.
-  struct WaiterState {
-    alignas(void*) char data[256];
-  } waiter_state;
-
-  // Used by PerThreadSem::{Get,Set}ThreadBlockedCounter().
-  std::atomic<int>* blocked_count_ptr;
-
-  // The following variables are mostly read/written just by the
-  // thread itself.  The only exception is that these are read by
-  // a ticker thread as a hint.
-  std::atomic<int> ticker;      // Tick counter, incremented once per second.
-  std::atomic<int> wait_start;  // Ticker value when thread started waiting.
-  std::atomic<bool> is_idle;    // Has thread become idle yet?
-
-  ThreadIdentity* next;
-};
-
-// Returns the ThreadIdentity object representing the calling thread; guaranteed
-// to be unique for its lifetime.  The returned object will remain valid for the
-// program's lifetime; although it may be re-assigned to a subsequent thread.
-// If one does not exist, return nullptr instead.
-//
-// Does not malloc(*), and is async-signal safe.
-// [*] Technically pthread_setspecific() does malloc on first use; however this
-// is handled internally within tcmalloc's initialization already. Note that
-// darwin does *not* use tcmalloc, so this can catch you if using MallocHooks
-// on Apple platforms. Whatever function is calling your MallocHooks will need
-// to watch for recursion on Apple platforms.
-//
-// New ThreadIdentity objects can be constructed and associated with a thread
-// by calling GetOrCreateCurrentThreadIdentity() in per-thread-sem.h.
-ThreadIdentity* CurrentThreadIdentityIfPresent();
-
-using ThreadIdentityReclaimerFunction = void (*)(void*);
-
-// Sets the current thread identity to the given value.  'reclaimer' is a
-// pointer to the global function for cleaning up instances on thread
-// destruction.
-void SetCurrentThreadIdentity(ThreadIdentity* identity,
-                              ThreadIdentityReclaimerFunction reclaimer);
-
-// Removes the currently associated ThreadIdentity from the running thread.
-// This must be called from inside the ThreadIdentityReclaimerFunction, and only
-// from that function.
-void ClearCurrentThreadIdentity();
-
-// May be chosen at compile time via: -DABSL_FORCE_THREAD_IDENTITY_MODE=<mode
-// index>
-#ifdef ABSL_THREAD_IDENTITY_MODE_USE_POSIX_SETSPECIFIC
-#error ABSL_THREAD_IDENTITY_MODE_USE_POSIX_SETSPECIFIC cannot be directly set
-#else
-#define ABSL_THREAD_IDENTITY_MODE_USE_POSIX_SETSPECIFIC 0
-#endif
-
-#ifdef ABSL_THREAD_IDENTITY_MODE_USE_TLS
-#error ABSL_THREAD_IDENTITY_MODE_USE_TLS cannot be directly set
-#else
-#define ABSL_THREAD_IDENTITY_MODE_USE_TLS 1
-#endif
-
-#ifdef ABSL_THREAD_IDENTITY_MODE_USE_CPP11
-#error ABSL_THREAD_IDENTITY_MODE_USE_CPP11 cannot be directly set
-#else
-#define ABSL_THREAD_IDENTITY_MODE_USE_CPP11 2
-#endif
-
-#ifdef ABSL_THREAD_IDENTITY_MODE
-#error ABSL_THREAD_IDENTITY_MODE cannot be directly set
-#elif defined(ABSL_FORCE_THREAD_IDENTITY_MODE)
-#define ABSL_THREAD_IDENTITY_MODE ABSL_FORCE_THREAD_IDENTITY_MODE
-#elif defined(_WIN32) && !defined(__MINGW32__)
-#define ABSL_THREAD_IDENTITY_MODE ABSL_THREAD_IDENTITY_MODE_USE_CPP11
-#elif defined(__APPLE__) && defined(ABSL_HAVE_THREAD_LOCAL)
-#define ABSL_THREAD_IDENTITY_MODE ABSL_THREAD_IDENTITY_MODE_USE_CPP11
-#elif ABSL_PER_THREAD_TLS && defined(__GOOGLE_GRTE_VERSION__) && \
-    (__GOOGLE_GRTE_VERSION__ >= 20140228L)
-// Support for async-safe TLS was specifically added in GRTEv4.  It's not
-// present in the upstream eglibc.
-// Note:  Current default for production systems.
-#define ABSL_THREAD_IDENTITY_MODE ABSL_THREAD_IDENTITY_MODE_USE_TLS
-#else
-#define ABSL_THREAD_IDENTITY_MODE \
-  ABSL_THREAD_IDENTITY_MODE_USE_POSIX_SETSPECIFIC
-#endif
-
-#if ABSL_THREAD_IDENTITY_MODE == ABSL_THREAD_IDENTITY_MODE_USE_TLS || \
-    ABSL_THREAD_IDENTITY_MODE == ABSL_THREAD_IDENTITY_MODE_USE_CPP11
-
-#if ABSL_PER_THREAD_TLS
-ABSL_CONST_INIT extern ABSL_PER_THREAD_TLS_KEYWORD ThreadIdentity*
-    thread_identity_ptr;
-#elif defined(ABSL_HAVE_THREAD_LOCAL)
-ABSL_CONST_INIT extern thread_local ThreadIdentity* thread_identity_ptr;
-#else
-#error Thread-local storage not detected on this platform
-#endif
-
-// thread_local variables cannot be in headers exposed by DLLs or in certain
-// build configurations on Apple platforms. However, it is important for
-// performance reasons in general that `CurrentThreadIdentityIfPresent` be
-// inlined. In the other cases we opt to have the function not be inlined. Note
-// that `CurrentThreadIdentityIfPresent` is declared above so we can exclude
-// this entire inline definition.
-#if !defined(__APPLE__) && !defined(ABSL_BUILD_DLL) && \
-    !defined(ABSL_CONSUME_DLL)
-#define ABSL_INTERNAL_INLINE_CURRENT_THREAD_IDENTITY_IF_PRESENT 1
-#endif
-
-#ifdef ABSL_INTERNAL_INLINE_CURRENT_THREAD_IDENTITY_IF_PRESENT
-inline ThreadIdentity* CurrentThreadIdentityIfPresent() {
-  return thread_identity_ptr;
-}
-#endif
-
-#elif ABSL_THREAD_IDENTITY_MODE != \
-    ABSL_THREAD_IDENTITY_MODE_USE_POSIX_SETSPECIFIC
-#error Unknown ABSL_THREAD_IDENTITY_MODE
-#endif
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_THREAD_IDENTITY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/throw_delegate.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/throw_delegate.h
deleted file mode 100644
index 075f5272..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/throw_delegate.h
+++ /dev/null
@@ -1,75 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_BASE_INTERNAL_THROW_DELEGATE_H_
-#define ABSL_BASE_INTERNAL_THROW_DELEGATE_H_
-
-#include <string>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-// Helper functions that allow throwing exceptions consistently from anywhere.
-// The main use case is for header-based libraries (eg templates), as they will
-// be built by many different targets with their own compiler options.
-// In particular, this will allow a safe way to throw exceptions even if the
-// caller is compiled with -fno-exceptions.  This is intended for implementing
-// things like map<>::at(), which the standard documents as throwing an
-// exception on error.
-//
-// Using other techniques like #if tricks could lead to ODR violations.
-//
-// You shouldn't use it unless you're writing code that you know will be built
-// both with and without exceptions and you need to conform to an interface
-// that uses exceptions.
-
-[[noreturn]] void ThrowStdLogicError(const std::string& what_arg);
-[[noreturn]] void ThrowStdLogicError(const char* what_arg);
-[[noreturn]] void ThrowStdInvalidArgument(const std::string& what_arg);
-[[noreturn]] void ThrowStdInvalidArgument(const char* what_arg);
-[[noreturn]] void ThrowStdDomainError(const std::string& what_arg);
-[[noreturn]] void ThrowStdDomainError(const char* what_arg);
-[[noreturn]] void ThrowStdLengthError(const std::string& what_arg);
-[[noreturn]] void ThrowStdLengthError(const char* what_arg);
-[[noreturn]] void ThrowStdOutOfRange(const std::string& what_arg);
-[[noreturn]] void ThrowStdOutOfRange(const char* what_arg);
-[[noreturn]] void ThrowStdRuntimeError(const std::string& what_arg);
-[[noreturn]] void ThrowStdRuntimeError(const char* what_arg);
-[[noreturn]] void ThrowStdRangeError(const std::string& what_arg);
-[[noreturn]] void ThrowStdRangeError(const char* what_arg);
-[[noreturn]] void ThrowStdOverflowError(const std::string& what_arg);
-[[noreturn]] void ThrowStdOverflowError(const char* what_arg);
-[[noreturn]] void ThrowStdUnderflowError(const std::string& what_arg);
-[[noreturn]] void ThrowStdUnderflowError(const char* what_arg);
-
-[[noreturn]] void ThrowStdBadFunctionCall();
-[[noreturn]] void ThrowStdBadAlloc();
-
-// ThrowStdBadArrayNewLength() cannot be consistently supported because
-// std::bad_array_new_length is missing in libstdc++ until 4.9.0.
-// https://gcc.gnu.org/onlinedocs/gcc-4.8.3/libstdc++/api/a01379_source.html
-// https://gcc.gnu.org/onlinedocs/gcc-4.9.0/libstdc++/api/a01327_source.html
-// libcxx (as of 3.2) and msvc (as of 2015) both have it.
-// [[noreturn]] void ThrowStdBadArrayNewLength();
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_INTERNAL_THROW_DELEGATE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/tsan_mutex_interface.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/tsan_mutex_interface.h
deleted file mode 100644
index 39207d8a..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/tsan_mutex_interface.h
+++ /dev/null
@@ -1,68 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// This file is intended solely for spinlock.h.
-// It provides ThreadSanitizer annotations for custom mutexes.
-// See <sanitizer/tsan_interface.h> for meaning of these annotations.
-
-#ifndef ABSL_BASE_INTERNAL_TSAN_MUTEX_INTERFACE_H_
-#define ABSL_BASE_INTERNAL_TSAN_MUTEX_INTERFACE_H_
-
-#include "absl/base/config.h"
-
-// ABSL_INTERNAL_HAVE_TSAN_INTERFACE
-// Macro intended only for internal use.
-//
-// Checks whether LLVM Thread Sanitizer interfaces are available.
-// First made available in LLVM 5.0 (Sep 2017).
-#ifdef ABSL_INTERNAL_HAVE_TSAN_INTERFACE
-#error "ABSL_INTERNAL_HAVE_TSAN_INTERFACE cannot be directly set."
-#endif
-
-#if defined(ABSL_HAVE_THREAD_SANITIZER) && defined(__has_include)
-#if __has_include(<sanitizer/tsan_interface.h>)
-#define ABSL_INTERNAL_HAVE_TSAN_INTERFACE 1
-#endif
-#endif
-
-#ifdef ABSL_INTERNAL_HAVE_TSAN_INTERFACE
-#include <sanitizer/tsan_interface.h>
-
-#define ABSL_TSAN_MUTEX_CREATE __tsan_mutex_create
-#define ABSL_TSAN_MUTEX_DESTROY __tsan_mutex_destroy
-#define ABSL_TSAN_MUTEX_PRE_LOCK __tsan_mutex_pre_lock
-#define ABSL_TSAN_MUTEX_POST_LOCK __tsan_mutex_post_lock
-#define ABSL_TSAN_MUTEX_PRE_UNLOCK __tsan_mutex_pre_unlock
-#define ABSL_TSAN_MUTEX_POST_UNLOCK __tsan_mutex_post_unlock
-#define ABSL_TSAN_MUTEX_PRE_SIGNAL __tsan_mutex_pre_signal
-#define ABSL_TSAN_MUTEX_POST_SIGNAL __tsan_mutex_post_signal
-#define ABSL_TSAN_MUTEX_PRE_DIVERT __tsan_mutex_pre_divert
-#define ABSL_TSAN_MUTEX_POST_DIVERT __tsan_mutex_post_divert
-
-#else
-
-#define ABSL_TSAN_MUTEX_CREATE(...)
-#define ABSL_TSAN_MUTEX_DESTROY(...)
-#define ABSL_TSAN_MUTEX_PRE_LOCK(...)
-#define ABSL_TSAN_MUTEX_POST_LOCK(...)
-#define ABSL_TSAN_MUTEX_PRE_UNLOCK(...)
-#define ABSL_TSAN_MUTEX_POST_UNLOCK(...)
-#define ABSL_TSAN_MUTEX_PRE_SIGNAL(...)
-#define ABSL_TSAN_MUTEX_POST_SIGNAL(...)
-#define ABSL_TSAN_MUTEX_PRE_DIVERT(...)
-#define ABSL_TSAN_MUTEX_POST_DIVERT(...)
-
-#endif
-
-#endif  // ABSL_BASE_INTERNAL_TSAN_MUTEX_INTERFACE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/unaligned_access.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/unaligned_access.h
deleted file mode 100644
index 4fea4574..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/unaligned_access.h
+++ /dev/null
@@ -1,89 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_BASE_INTERNAL_UNALIGNED_ACCESS_H_
-#define ABSL_BASE_INTERNAL_UNALIGNED_ACCESS_H_
-
-#include <string.h>
-
-#include <cstdint>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/nullability.h"
-
-// unaligned APIs
-
-// Portable handling of unaligned loads, stores, and copies.
-
-// The unaligned API is C++ only.  The declarations use C++ features
-// (namespaces, inline) which are absent or incompatible in C.
-#if defined(__cplusplus)
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-inline uint16_t UnalignedLoad16(absl::Nonnull<const void *> p) {
-  uint16_t t;
-  memcpy(&t, p, sizeof t);
-  return t;
-}
-
-inline uint32_t UnalignedLoad32(absl::Nonnull<const void *> p) {
-  uint32_t t;
-  memcpy(&t, p, sizeof t);
-  return t;
-}
-
-inline uint64_t UnalignedLoad64(absl::Nonnull<const void *> p) {
-  uint64_t t;
-  memcpy(&t, p, sizeof t);
-  return t;
-}
-
-inline void UnalignedStore16(absl::Nonnull<void *> p, uint16_t v) {
-  memcpy(p, &v, sizeof v);
-}
-
-inline void UnalignedStore32(absl::Nonnull<void *> p, uint32_t v) {
-  memcpy(p, &v, sizeof v);
-}
-
-inline void UnalignedStore64(absl::Nonnull<void *> p, uint64_t v) {
-  memcpy(p, &v, sizeof v);
-}
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#define ABSL_INTERNAL_UNALIGNED_LOAD16(_p) \
-  (absl::base_internal::UnalignedLoad16(_p))
-#define ABSL_INTERNAL_UNALIGNED_LOAD32(_p) \
-  (absl::base_internal::UnalignedLoad32(_p))
-#define ABSL_INTERNAL_UNALIGNED_LOAD64(_p) \
-  (absl::base_internal::UnalignedLoad64(_p))
-
-#define ABSL_INTERNAL_UNALIGNED_STORE16(_p, _val) \
-  (absl::base_internal::UnalignedStore16(_p, _val))
-#define ABSL_INTERNAL_UNALIGNED_STORE32(_p, _val) \
-  (absl::base_internal::UnalignedStore32(_p, _val))
-#define ABSL_INTERNAL_UNALIGNED_STORE64(_p, _val) \
-  (absl::base_internal::UnalignedStore64(_p, _val))
-
-#endif  // defined(__cplusplus), end of unaligned API
-
-#endif  // ABSL_BASE_INTERNAL_UNALIGNED_ACCESS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/unscaledcycleclock.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/unscaledcycleclock.h
deleted file mode 100644
index cc1276ba..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/unscaledcycleclock.h
+++ /dev/null
@@ -1,96 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// UnscaledCycleClock
-//    An UnscaledCycleClock yields the value and frequency of a cycle counter
-//    that increments at a rate that is approximately constant.
-//    This class is for internal use only, you should consider using CycleClock
-//    instead.
-//
-// Notes:
-// The cycle counter frequency is not necessarily the core clock frequency.
-// That is, CycleCounter cycles are not necessarily "CPU cycles".
-//
-// An arbitrary offset may have been added to the counter at power on.
-//
-// On some platforms, the rate and offset of the counter may differ
-// slightly when read from different CPUs of a multiprocessor.  Usually,
-// we try to ensure that the operating system adjusts values periodically
-// so that values agree approximately.   If you need stronger guarantees,
-// consider using alternate interfaces.
-//
-// The CPU is not required to maintain the ordering of a cycle counter read
-// with respect to surrounding instructions.
-
-#ifndef ABSL_BASE_INTERNAL_UNSCALEDCYCLECLOCK_H_
-#define ABSL_BASE_INTERNAL_UNSCALEDCYCLECLOCK_H_
-
-#include <cstdint>
-
-#if defined(__APPLE__)
-#include <TargetConditionals.h>
-#endif
-
-#include "absl/base/config.h"
-#include "absl/base/internal/unscaledcycleclock_config.h"
-
-#if ABSL_USE_UNSCALED_CYCLECLOCK
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace time_internal {
-class UnscaledCycleClockWrapperForGetCurrentTime;
-}  // namespace time_internal
-
-namespace base_internal {
-class CycleClock;
-class UnscaledCycleClockWrapperForInitializeFrequency;
-
-class UnscaledCycleClock {
- private:
-  UnscaledCycleClock() = delete;
-
-  // Return the value of a cycle counter that counts at a rate that is
-  // approximately constant.
-  static int64_t Now();
-
-  // Return the how much UnscaledCycleClock::Now() increases per second.
-  // This is not necessarily the core CPU clock frequency.
-  // It may be the nominal value report by the kernel, rather than a measured
-  // value.
-  static double Frequency();
-
-  // Allowed users
-  friend class base_internal::CycleClock;
-  friend class time_internal::UnscaledCycleClockWrapperForGetCurrentTime;
-  friend class base_internal::UnscaledCycleClockWrapperForInitializeFrequency;
-};
-
-#if defined(__x86_64__)
-
-inline int64_t UnscaledCycleClock::Now() {
-  uint64_t low, high;
-  __asm__ volatile("rdtsc" : "=a"(low), "=d"(high));
-  return static_cast<int64_t>((high << 32) | low);
-}
-
-#endif
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_USE_UNSCALED_CYCLECLOCK
-
-#endif  // ABSL_BASE_INTERNAL_UNSCALEDCYCLECLOCK_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/internal/unscaledcycleclock_config.h b/projects/llm_framework/include/abseil-cpp/absl/base/internal/unscaledcycleclock_config.h
deleted file mode 100644
index 24b324ac..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/internal/unscaledcycleclock_config.h
+++ /dev/null
@@ -1,62 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_BASE_INTERNAL_UNSCALEDCYCLECLOCK_CONFIG_H_
-#define ABSL_BASE_INTERNAL_UNSCALEDCYCLECLOCK_CONFIG_H_
-
-#if defined(__APPLE__)
-#include <TargetConditionals.h>
-#endif
-
-// The following platforms have an implementation of a hardware counter.
-#if defined(__i386__) || defined(__x86_64__) || defined(__aarch64__) || \
-    defined(__powerpc__) || defined(__ppc__) || defined(__riscv) ||     \
-    defined(_M_IX86) || (defined(_M_X64) && !defined(_M_ARM64EC))
-#define ABSL_HAVE_UNSCALED_CYCLECLOCK_IMPLEMENTATION 1
-#else
-#define ABSL_HAVE_UNSCALED_CYCLECLOCK_IMPLEMENTATION 0
-#endif
-
-// The following platforms often disable access to the hardware
-// counter (through a sandbox) even if the underlying hardware has a
-// usable counter. The CycleTimer interface also requires a *scaled*
-// CycleClock that runs at atleast 1 MHz. We've found some Android
-// ARM64 devices where this is not the case, so we disable it by
-// default on Android ARM64.
-#if defined(__native_client__) || (defined(__APPLE__)) || \
-    (defined(__ANDROID__) && defined(__aarch64__))
-#define ABSL_USE_UNSCALED_CYCLECLOCK_DEFAULT 0
-#else
-#define ABSL_USE_UNSCALED_CYCLECLOCK_DEFAULT 1
-#endif
-
-// UnscaledCycleClock is an optional internal feature.
-// Use "#if ABSL_USE_UNSCALED_CYCLECLOCK" to test for its presence.
-// Can be overridden at compile-time via -DABSL_USE_UNSCALED_CYCLECLOCK=0|1
-#if !defined(ABSL_USE_UNSCALED_CYCLECLOCK)
-#define ABSL_USE_UNSCALED_CYCLECLOCK               \
-  (ABSL_HAVE_UNSCALED_CYCLECLOCK_IMPLEMENTATION && \
-   ABSL_USE_UNSCALED_CYCLECLOCK_DEFAULT)
-#endif
-
-#if ABSL_USE_UNSCALED_CYCLECLOCK
-// This macro can be used to test if UnscaledCycleClock::Frequency()
-// is NominalCPUFrequency() on a particular platform.
-#if (defined(__i386__) || defined(__x86_64__) || defined(__riscv) || \
-     defined(_M_IX86) || defined(_M_X64))
-#define ABSL_INTERNAL_UNSCALED_CYCLECLOCK_FREQUENCY_IS_CPU_FREQUENCY
-#endif
-#endif
-
-#endif  // ABSL_BASE_INTERNAL_UNSCALEDCYCLECLOCK_CONFIG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/log_severity.h b/projects/llm_framework/include/abseil-cpp/absl/base/log_severity.h
deleted file mode 100644
index de9702ab..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/log_severity.h
+++ /dev/null
@@ -1,185 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_BASE_LOG_SEVERITY_H_
-#define ABSL_BASE_LOG_SEVERITY_H_
-
-#include <array>
-#include <ostream>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// absl::LogSeverity
-//
-// Four severity levels are defined. Logging APIs should terminate the program
-// when a message is logged at severity `kFatal`; the other levels have no
-// special semantics.
-//
-// Values other than the four defined levels (e.g. produced by `static_cast`)
-// are valid, but their semantics when passed to a function, macro, or flag
-// depend on the function, macro, or flag. The usual behavior is to normalize
-// such values to a defined severity level, however in some cases values other
-// than the defined levels are useful for comparison.
-//
-// Example:
-//
-//   // Effectively disables all logging:
-//   SetMinLogLevel(static_cast<absl::LogSeverity>(100));
-//
-// Abseil flags may be defined with type `LogSeverity`. Dependency layering
-// constraints require that the `AbslParseFlag()` overload be declared and
-// defined in the flags library itself rather than here. The `AbslUnparseFlag()`
-// overload is defined there as well for consistency.
-//
-// absl::LogSeverity Flag String Representation
-//
-// An `absl::LogSeverity` has a string representation used for parsing
-// command-line flags based on the enumerator name (e.g. `kFatal`) or
-// its unprefixed name (without the `k`) in any case-insensitive form. (E.g.
-// "FATAL", "fatal" or "Fatal" are all valid.) Unparsing such flags produces an
-// unprefixed string representation in all caps (e.g. "FATAL") or an integer.
-//
-// Additionally, the parser accepts arbitrary integers (as if the type were
-// `int`).
-//
-// Examples:
-//
-//   --my_log_level=kInfo
-//   --my_log_level=INFO
-//   --my_log_level=info
-//   --my_log_level=0
-//
-// `DFATAL` and `kLogDebugFatal` are similarly accepted.
-//
-// Unparsing a flag produces the same result as `absl::LogSeverityName()` for
-// the standard levels and a base-ten integer otherwise.
-enum class LogSeverity : int {
-  kInfo = 0,
-  kWarning = 1,
-  kError = 2,
-  kFatal = 3,
-};
-
-// LogSeverities()
-//
-// Returns an iterable of all standard `absl::LogSeverity` values, ordered from
-// least to most severe.
-constexpr std::array<absl::LogSeverity, 4> LogSeverities() {
-  return {{absl::LogSeverity::kInfo, absl::LogSeverity::kWarning,
-           absl::LogSeverity::kError, absl::LogSeverity::kFatal}};
-}
-
-// `absl::kLogDebugFatal` equals `absl::LogSeverity::kFatal` in debug builds
-// (i.e. when `NDEBUG` is not defined) and `absl::LogSeverity::kError`
-// otherwise.  Avoid ODR-using this variable as it has internal linkage and thus
-// distinct storage in different TUs.
-#ifdef NDEBUG
-static constexpr absl::LogSeverity kLogDebugFatal = absl::LogSeverity::kError;
-#else
-static constexpr absl::LogSeverity kLogDebugFatal = absl::LogSeverity::kFatal;
-#endif
-
-// LogSeverityName()
-//
-// Returns the all-caps string representation (e.g. "INFO") of the specified
-// severity level if it is one of the standard levels and "UNKNOWN" otherwise.
-constexpr const char* LogSeverityName(absl::LogSeverity s) {
-  switch (s) {
-    case absl::LogSeverity::kInfo: return "INFO";
-    case absl::LogSeverity::kWarning: return "WARNING";
-    case absl::LogSeverity::kError: return "ERROR";
-    case absl::LogSeverity::kFatal: return "FATAL";
-  }
-  return "UNKNOWN";
-}
-
-// NormalizeLogSeverity()
-//
-// Values less than `kInfo` normalize to `kInfo`; values greater than `kFatal`
-// normalize to `kError` (**NOT** `kFatal`).
-constexpr absl::LogSeverity NormalizeLogSeverity(absl::LogSeverity s) {
-  absl::LogSeverity n = s;
-  if (n < absl::LogSeverity::kInfo) n = absl::LogSeverity::kInfo;
-  if (n > absl::LogSeverity::kFatal) n = absl::LogSeverity::kError;
-  return n;
-}
-constexpr absl::LogSeverity NormalizeLogSeverity(int s) {
-  return absl::NormalizeLogSeverity(static_cast<absl::LogSeverity>(s));
-}
-
-// operator<<
-//
-// The exact representation of a streamed `absl::LogSeverity` is deliberately
-// unspecified; do not rely on it.
-std::ostream& operator<<(std::ostream& os, absl::LogSeverity s);
-
-// Enums representing a lower bound for LogSeverity. APIs that only operate on
-// messages of at least a certain level (for example, `SetMinLogLevel()`) use
-// this type to specify that level. absl::LogSeverityAtLeast::kInfinity is
-// a level above all threshold levels and therefore no log message will
-// ever meet this threshold.
-enum class LogSeverityAtLeast : int {
-  kInfo = static_cast<int>(absl::LogSeverity::kInfo),
-  kWarning = static_cast<int>(absl::LogSeverity::kWarning),
-  kError = static_cast<int>(absl::LogSeverity::kError),
-  kFatal = static_cast<int>(absl::LogSeverity::kFatal),
-  kInfinity = 1000,
-};
-
-std::ostream& operator<<(std::ostream& os, absl::LogSeverityAtLeast s);
-
-// Enums representing an upper bound for LogSeverity. APIs that only operate on
-// messages of at most a certain level (for example, buffer all messages at or
-// below a certain level) use this type to specify that level.
-// absl::LogSeverityAtMost::kNegativeInfinity is a level below all threshold
-// levels and therefore will exclude all log messages.
-enum class LogSeverityAtMost : int {
-  kNegativeInfinity = -1000,
-  kInfo = static_cast<int>(absl::LogSeverity::kInfo),
-  kWarning = static_cast<int>(absl::LogSeverity::kWarning),
-  kError = static_cast<int>(absl::LogSeverity::kError),
-  kFatal = static_cast<int>(absl::LogSeverity::kFatal),
-};
-
-std::ostream& operator<<(std::ostream& os, absl::LogSeverityAtMost s);
-
-#define COMPOP(op1, op2, T)                                         \
-  constexpr bool operator op1(absl::T lhs, absl::LogSeverity rhs) { \
-    return static_cast<absl::LogSeverity>(lhs) op1 rhs;             \
-  }                                                                 \
-  constexpr bool operator op2(absl::LogSeverity lhs, absl::T rhs) { \
-    return lhs op2 static_cast<absl::LogSeverity>(rhs);             \
-  }
-
-// Comparisons between `LogSeverity` and `LogSeverityAtLeast`/
-// `LogSeverityAtMost` are only supported in one direction.
-// Valid checks are:
-//   LogSeverity >= LogSeverityAtLeast
-//   LogSeverity < LogSeverityAtLeast
-//   LogSeverity <= LogSeverityAtMost
-//   LogSeverity > LogSeverityAtMost
-COMPOP(>, <, LogSeverityAtLeast)
-COMPOP(<=, >=, LogSeverityAtLeast)
-COMPOP(<, >, LogSeverityAtMost)
-COMPOP(>=, <=, LogSeverityAtMost)
-#undef COMPOP
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_LOG_SEVERITY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/macros.h b/projects/llm_framework/include/abseil-cpp/absl/base/macros.h
deleted file mode 100644
index 33c1cd1e..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/macros.h
+++ /dev/null
@@ -1,177 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: macros.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines the set of language macros used within Abseil code.
-// For the set of macros used to determine supported compilers and platforms,
-// see absl/base/config.h instead.
-//
-// This code is compiled directly on many platforms, including client
-// platforms like Windows, Mac, and embedded systems.  Before making
-// any changes here, make sure that you're not breaking any platforms.
-
-#ifndef ABSL_BASE_MACROS_H_
-#define ABSL_BASE_MACROS_H_
-
-#include <cassert>
-#include <cstddef>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/optimization.h"
-#include "absl/base/port.h"
-
-// ABSL_ARRAYSIZE()
-//
-// Returns the number of elements in an array as a compile-time constant, which
-// can be used in defining new arrays. If you use this macro on a pointer by
-// mistake, you will get a compile-time error.
-#define ABSL_ARRAYSIZE(array) \
-  (sizeof(::absl::macros_internal::ArraySizeHelper(array)))
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace macros_internal {
-// Note: this internal template function declaration is used by ABSL_ARRAYSIZE.
-// The function doesn't need a definition, as we only use its type.
-template <typename T, size_t N>
-auto ArraySizeHelper(const T (&array)[N]) -> char (&)[N];
-}  // namespace macros_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-// ABSL_BAD_CALL_IF()
-//
-// Used on a function overload to trap bad calls: any call that matches the
-// overload will cause a compile-time error. This macro uses a clang-specific
-// "enable_if" attribute, as described at
-// https://clang.llvm.org/docs/AttributeReference.html#enable-if
-//
-// Overloads which use this macro should be bracketed by
-// `#ifdef ABSL_BAD_CALL_IF`.
-//
-// Example:
-//
-//   int isdigit(int c);
-//   #ifdef ABSL_BAD_CALL_IF
-//   int isdigit(int c)
-//     ABSL_BAD_CALL_IF(c <= -1 || c > 255,
-//                       "'c' must have the value of an unsigned char or EOF");
-//   #endif // ABSL_BAD_CALL_IF
-#if ABSL_HAVE_ATTRIBUTE(enable_if)
-#define ABSL_BAD_CALL_IF(expr, msg) \
-  __attribute__((enable_if(expr, "Bad call trap"), unavailable(msg)))
-#endif
-
-// ABSL_ASSERT()
-//
-// In C++11, `assert` can't be used portably within constexpr functions.
-// ABSL_ASSERT functions as a runtime assert but works in C++11 constexpr
-// functions.  Example:
-//
-// constexpr double Divide(double a, double b) {
-//   return ABSL_ASSERT(b != 0), a / b;
-// }
-//
-// This macro is inspired by
-// https://akrzemi1.wordpress.com/2017/05/18/asserts-in-constexpr-functions/
-#if defined(NDEBUG)
-#define ABSL_ASSERT(expr) \
-  (false ? static_cast<void>(expr) : static_cast<void>(0))
-#else
-#define ABSL_ASSERT(expr)                           \
-  (ABSL_PREDICT_TRUE((expr)) ? static_cast<void>(0) \
-                             : [] { assert(false && #expr); }())  // NOLINT
-#endif
-
-// `ABSL_INTERNAL_HARDENING_ABORT()` controls how `ABSL_HARDENING_ASSERT()`
-// aborts the program in release mode (when NDEBUG is defined). The
-// implementation should abort the program as quickly as possible and ideally it
-// should not be possible to ignore the abort request.
-#define ABSL_INTERNAL_HARDENING_ABORT()   \
-  do {                                    \
-    ABSL_INTERNAL_IMMEDIATE_ABORT_IMPL(); \
-    ABSL_INTERNAL_UNREACHABLE_IMPL();     \
-  } while (false)
-
-// ABSL_HARDENING_ASSERT()
-//
-// `ABSL_HARDENING_ASSERT()` is like `ABSL_ASSERT()`, but used to implement
-// runtime assertions that should be enabled in hardened builds even when
-// `NDEBUG` is defined.
-//
-// When `NDEBUG` is not defined, `ABSL_HARDENING_ASSERT()` is identical to
-// `ABSL_ASSERT()`.
-//
-// See `ABSL_OPTION_HARDENED` in `absl/base/options.h` for more information on
-// hardened mode.
-#if ABSL_OPTION_HARDENED == 1 && defined(NDEBUG)
-#define ABSL_HARDENING_ASSERT(expr)                 \
-  (ABSL_PREDICT_TRUE((expr)) ? static_cast<void>(0) \
-                             : [] { ABSL_INTERNAL_HARDENING_ABORT(); }())
-#else
-#define ABSL_HARDENING_ASSERT(expr) ABSL_ASSERT(expr)
-#endif
-
-#ifdef ABSL_HAVE_EXCEPTIONS
-#define ABSL_INTERNAL_TRY try
-#define ABSL_INTERNAL_CATCH_ANY catch (...)
-#define ABSL_INTERNAL_RETHROW do { throw; } while (false)
-#else  // ABSL_HAVE_EXCEPTIONS
-#define ABSL_INTERNAL_TRY if (true)
-#define ABSL_INTERNAL_CATCH_ANY else if (false)
-#define ABSL_INTERNAL_RETHROW do {} while (false)
-#endif  // ABSL_HAVE_EXCEPTIONS
-
-// ABSL_DEPRECATE_AND_INLINE()
-//
-// Marks a function or type alias as deprecated and tags it to be picked up for
-// automated refactoring by go/cpp-inliner. It can added to inline function
-// definitions or type aliases. It should only be used within a header file. It
-// differs from `ABSL_DEPRECATED` in the following ways:
-//
-// 1. New uses of the function or type will be discouraged via Tricorder
-//    warnings.
-// 2. If enabled via `METADATA`, automated changes will be sent out inlining the
-//    functions's body or replacing the type where it is used.
-//
-// For example:
-//
-// ABSL_DEPRECATE_AND_INLINE() inline int OldFunc(int x) {
-//   return NewFunc(x, 0);
-// }
-//
-// will mark `OldFunc` as deprecated, and the go/cpp-inliner service will
-// replace calls to `OldFunc(x)` with calls to `NewFunc(x, 0)`. Once all calls
-// to `OldFunc` have been replaced, `OldFunc` can be deleted.
-//
-// See go/cpp-inliner for more information.
-//
-// Note: go/cpp-inliner is Google-internal service for automated refactoring.
-// While open-source users do not have access to this service, the macro is
-// provided for compatibility, and so that users receive deprecation warnings.
-#if ABSL_HAVE_CPP_ATTRIBUTE(deprecated) && \
-    ABSL_HAVE_CPP_ATTRIBUTE(clang::annotate)
-#define ABSL_DEPRECATE_AND_INLINE() [[deprecated, clang::annotate("inline-me")]]
-#elif ABSL_HAVE_CPP_ATTRIBUTE(deprecated)
-#define ABSL_DEPRECATE_AND_INLINE() [[deprecated]]
-#else
-#define ABSL_DEPRECATE_AND_INLINE()
-#endif
-
-#endif  // ABSL_BASE_MACROS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/no_destructor.h b/projects/llm_framework/include/abseil-cpp/absl/base/no_destructor.h
deleted file mode 100644
index 7b46456c..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/no_destructor.h
+++ /dev/null
@@ -1,218 +0,0 @@
-// Copyright 2023 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: no_destructor.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines the absl::NoDestructor<T> wrapper for defining a
-// static type that does not need to be destructed upon program exit. Instead,
-// such an object survives during program exit (and can be safely accessed at
-// any time).
-//
-// Objects of such type, if constructed safely and under the right conditions,
-// provide two main benefits over other alternatives:
-//
-//   * Global objects not normally allowed due to concerns of destruction order
-//     (i.e. no "complex globals") can be safely allowed, provided that such
-//     objects can be constant initialized.
-//   * Function scope static objects can be optimized to avoid heap allocation,
-//     pointer chasing, and allow lazy construction.
-//
-// See below for complete details.
-
-
-#ifndef ABSL_BASE_NO_DESTRUCTOR_H_
-#define ABSL_BASE_NO_DESTRUCTOR_H_
-
-#include <new>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/config.h"
-#include "absl/base/nullability.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// absl::NoDestructor<T>
-//
-// NoDestructor<T> is a wrapper around an object of type T that behaves as an
-// object of type T but never calls T's destructor. NoDestructor<T> makes it
-// safer and/or more efficient to use such objects in static storage contexts:
-// as global or function scope static variables.
-//
-// An instance of absl::NoDestructor<T> has similar type semantics to an
-// instance of T:
-//
-// * Constructs in the same manner as an object of type T through perfect
-//   forwarding.
-// * Provides pointer/reference semantic access to the object of type T via
-//   `->`, `*`, and `get()`.
-//   (Note that `const NoDestructor<T>` works like a pointer to const `T`.)
-//
-// An object of type NoDestructor<T> should be defined in static storage:
-// as either a global static object, or as a function scope static variable.
-//
-// Additionally, NoDestructor<T> provides the following benefits:
-//
-// * Never calls T's destructor for the object
-// * If the object is a function-local static variable, the type can be
-//   lazily constructed.
-//
-// An object of type NoDestructor<T> is "trivially destructible" in the notion
-// that its destructor is never run. Provided that an object of this type can be
-// safely initialized and does not need to be cleaned up on program shutdown,
-// NoDestructor<T> allows you to define global static variables, since Google's
-// C++ style guide ban on such objects doesn't apply to objects that are
-// trivially destructible.
-//
-// Usage as Global Static Variables
-//
-// NoDestructor<T> allows declaration of a global object with a non-trivial
-// constructor in static storage without needing to add a destructor.
-// However, such objects still need to worry about initialization order, so
-// such objects should be const initialized:
-//
-//    // Global or namespace scope.
-//    constinit absl::NoDestructor<MyRegistry> reg{"foo", "bar", 8008};
-//
-// Note that if your object already has a trivial destructor, you don't need to
-// use NoDestructor<T>.
-//
-// Usage as Function Scope Static Variables
-//
-// Function static objects will be lazily initialized within static storage:
-//
-//    // Function scope.
-//    const std::string& MyString() {
-//      static const absl::NoDestructor<std::string> x("foo");
-//      return *x;
-//    }
-//
-// For function static variables, NoDestructor avoids heap allocation and can be
-// inlined in static storage, resulting in exactly-once, thread-safe
-// construction of an object, and very fast access thereafter (the cost is a few
-// extra cycles).
-//
-// Using NoDestructor<T> in this manner is generally better than other patterns
-// which require pointer chasing:
-//
-//   // Prefer using absl::NoDestructor<T> instead for the static variable.
-//   const std::string& MyString() {
-//     static const std::string* x = new std::string("foo");
-//     return *x;
-//   }
-//
-template <typename T>
-class NoDestructor {
- public:
-  // Forwards arguments to the T's constructor: calls T(args...).
-  template <typename... Ts,
-            // Disable this overload when it might collide with copy/move.
-            typename std::enable_if<!std::is_same<void(std::decay_t<Ts>&...),
-                                                  void(NoDestructor&)>::value,
-                                    int>::type = 0>
-  explicit constexpr NoDestructor(Ts&&... args)
-      : impl_(std::forward<Ts>(args)...) {}
-
-  // Forwards copy and move construction for T. Enables usage like this:
-  //   static NoDestructor<std::array<string, 3>> x{{{"1", "2", "3"}}};
-  //   static NoDestructor<std::vector<int>> x{{1, 2, 3}};
-  explicit constexpr NoDestructor(const T& x) : impl_(x) {}
-  explicit constexpr NoDestructor(T&& x)
-      : impl_(std::move(x)) {}
-
-  // No copying.
-  NoDestructor(const NoDestructor&) = delete;
-  NoDestructor& operator=(const NoDestructor&) = delete;
-
-  // Pretend to be a smart pointer to T with deep constness.
-  // Never returns a null pointer.
-  T& operator*() { return *get(); }
-  absl::Nonnull<T*> operator->() { return get(); }
-  absl::Nonnull<T*> get() { return impl_.get(); }
-  const T& operator*() const { return *get(); }
-  absl::Nonnull<const T*> operator->() const { return get(); }
-  absl::Nonnull<const T*> get() const { return impl_.get(); }
-
- private:
-  class DirectImpl {
-   public:
-    template <typename... Args>
-    explicit constexpr DirectImpl(Args&&... args)
-        : value_(std::forward<Args>(args)...) {}
-    absl::Nonnull<const T*> get() const { return &value_; }
-    absl::Nonnull<T*> get() { return &value_; }
-
-   private:
-    T value_;
-  };
-
-  class PlacementImpl {
-   public:
-    template <typename... Args>
-    explicit PlacementImpl(Args&&... args) {
-      new (&space_) T(std::forward<Args>(args)...);
-    }
-    absl::Nonnull<const T*> get() const {
-      return Launder(reinterpret_cast<const T*>(&space_));
-    }
-    absl::Nonnull<T*> get() { return Launder(reinterpret_cast<T*>(&space_)); }
-
-   private:
-    template <typename P>
-    static absl::Nonnull<P*> Launder(absl::Nonnull<P*> p) {
-#if defined(__cpp_lib_launder) && __cpp_lib_launder >= 201606L
-      return std::launder(p);
-#elif ABSL_HAVE_BUILTIN(__builtin_launder)
-      return __builtin_launder(p);
-#else
-      // When `std::launder` or equivalent are not available, we rely on
-      // undefined behavior, which works as intended on Abseil's officially
-      // supported platforms as of Q3 2023.
-#if defined(__GNUC__) && !defined(__clang__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#endif
-      return p;
-#if defined(__GNUC__) && !defined(__clang__)
-#pragma GCC diagnostic pop
-#endif
-#endif
-    }
-
-    alignas(T) unsigned char space_[sizeof(T)];
-  };
-
-  // If the object is trivially destructible we use a member directly to avoid
-  // potential once-init runtime initialization. It somewhat defeats the
-  // purpose of NoDestructor in this case, but this makes the class more
-  // friendly to generic code.
-  std::conditional_t<std::is_trivially_destructible<T>::value, DirectImpl,
-                     PlacementImpl>
-      impl_;
-};
-
-#ifdef ABSL_HAVE_CLASS_TEMPLATE_ARGUMENT_DEDUCTION
-// Provide 'Class Template Argument Deduction': the type of NoDestructor's T
-// will be the same type as the argument passed to NoDestructor's constructor.
-template <typename T>
-NoDestructor(T) -> NoDestructor<T>;
-#endif  // ABSL_HAVE_CLASS_TEMPLATE_ARGUMENT_DEDUCTION
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_NO_DESTRUCTOR_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/nullability.h b/projects/llm_framework/include/abseil-cpp/absl/base/nullability.h
deleted file mode 100644
index 6f49b6f5..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/nullability.h
+++ /dev/null
@@ -1,224 +0,0 @@
-// Copyright 2023 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: nullability.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines a set of "templated annotations" for designating the
-// expected nullability of pointers. These annotations allow you to designate
-// pointers in one of three classification states:
-//
-//  * "Non-null" (for pointers annotated `Nonnull<T>`), indicating that it is
-//    invalid for the given pointer to ever be null.
-//  * "Nullable" (for pointers annotated `Nullable<T>`), indicating that it is
-//    valid for the given pointer to be null.
-//  * "Unknown" (for pointers annotated `NullabilityUnknown<T>`), indicating
-//    that the given pointer has not been yet classified as either nullable or
-//    non-null. This is the default state of unannotated pointers.
-//
-// NOTE: unannotated pointers implicitly bear the annotation
-// `NullabilityUnknown<T>`; you should rarely, if ever, see this annotation used
-// in the codebase explicitly.
-//
-// -----------------------------------------------------------------------------
-// Nullability and Contracts
-// -----------------------------------------------------------------------------
-//
-// These nullability annotations allow you to more clearly specify contracts on
-// software components by narrowing the *preconditions*, *postconditions*, and
-// *invariants* of pointer state(s) in any given interface. It then depends on
-// context who is responsible for fulfilling the annotation's requirements.
-//
-// For example, a function may receive a pointer argument. Designating that
-// pointer argument as "non-null" tightens the precondition of the contract of
-// that function. It is then the responsibility of anyone calling such a
-// function to ensure that the passed pointer is not null.
-//
-// Similarly, a function may have a pointer as a return value. Designating that
-// return value as "non-null" tightens the postcondition of the contract of that
-// function. In this case, however, it is the responsibility of the function
-// itself to ensure that the returned pointer is not null.
-//
-// Clearly defining these contracts allows providers (and consumers) of such
-// pointers to have more confidence in their null state. If a function declares
-// a return value as "non-null", for example, the caller should not need to
-// check whether the returned value is `nullptr`; it can simply assume the
-// pointer is valid.
-//
-// Of course most interfaces already have expectations on the nullability state
-// of pointers, and these expectations are, in effect, a contract; often,
-// however, those contracts are either poorly or partially specified, assumed,
-// or misunderstood. These nullability annotations are designed to allow you to
-// formalize those contracts within the codebase.
-//
-// -----------------------------------------------------------------------------
-// Using Nullability Annotations
-// -----------------------------------------------------------------------------
-//
-// It is important to note that these annotations are not distinct strong
-// *types*. They are alias templates defined to be equal to the underlying
-// pointer type. A pointer annotated `Nonnull<T*>`, for example, is simply a
-// pointer of type `T*`. Each annotation acts as a form of documentation about
-// the contract for the given pointer. Each annotation requires providers or
-// consumers of these pointers across API boundaries to take appropriate steps
-// when setting or using these pointers:
-//
-// * "Non-null" pointers should never be null. It is the responsibility of the
-//   provider of this pointer to ensure that the pointer may never be set to
-//   null. Consumers of such pointers can treat such pointers as non-null.
-// * "Nullable" pointers may or may not be null. Consumers of such pointers
-//   should precede any usage of that pointer (e.g. a dereference operation)
-//   with a a `nullptr` check.
-// * "Unknown" pointers may be either "non-null" or "nullable" but have not been
-//   definitively determined to be in either classification state. Providers of
-//   such pointers across API boundaries should determine --  over time -- to
-//   annotate the pointer in either of the above two states. Consumers of such
-//   pointers across an API boundary should continue to treat such pointers as
-//   they currently do.
-//
-// Example:
-//
-// // PaySalary() requires the passed pointer to an `Employee` to be non-null.
-// void PaySalary(absl::Nonnull<Employee *> e) {
-//   pay(e->salary);  // OK to dereference
-// }
-//
-// // CompleteTransaction() guarantees the returned pointer to an `Account` to
-// // be non-null.
-// absl::Nonnull<Account *> balance CompleteTransaction(double fee) {
-// ...
-// }
-//
-// // Note that specifying a nullability annotation does not prevent someone
-// // from violating the contract:
-//
-// Nullable<Employee *> find(Map& employees, std::string_view name);
-//
-// void g(Map& employees) {
-//   Employee *e = find(employees, "Pat");
-//   // `e` can now be null.
-//   PaySalary(e); // Violates contract, but compiles!
-// }
-//
-// Nullability annotations, in other words, are useful for defining and
-// narrowing contracts; *enforcement* of those contracts depends on use and any
-// additional (static or dynamic analysis) tooling.
-//
-// NOTE: The "unknown" annotation state indicates that a pointer's contract has
-// not yet been positively identified. The unknown state therefore acts as a
-// form of documentation of your technical debt, and a codebase that adopts
-// nullability annotations should aspire to annotate every pointer as either
-// "non-null" or "nullable".
-//
-// -----------------------------------------------------------------------------
-// Applicability of Nullability Annotations
-// -----------------------------------------------------------------------------
-//
-// By default, nullability annotations are applicable to raw and smart
-// pointers. User-defined types can indicate compatibility with nullability
-// annotations by providing an `absl_nullability_compatible` nested type. The
-// actual definition of this inner type is not relevant as it is used merely as
-// a marker. It is common to use a using declaration of
-// `absl_nullability_compatible` set to void.
-//
-// // Example:
-// struct MyPtr {
-//   using absl_nullability_compatible = void;
-//   ...
-// };
-//
-// DISCLAIMER:
-// ===========================================================================
-// These nullability annotations are primarily a human readable signal about the
-// intended contract of the pointer. They are not *types* and do not currently
-// provide any correctness guarantees. For example, a pointer annotated as
-// `Nonnull<T*>` is *not guaranteed* to be non-null, and the compiler won't
-// alert or prevent assignment of a `Nullable<T*>` to a `Nonnull<T*>`.
-// ===========================================================================
-#ifndef ABSL_BASE_NULLABILITY_H_
-#define ABSL_BASE_NULLABILITY_H_
-
-#include "absl/base/internal/nullability_impl.h"
-
-namespace absl {
-
-// absl::Nonnull
-//
-// The indicated pointer is never null. It is the responsibility of the provider
-// of this pointer across an API boundary to ensure that the pointer is never be
-// set to null. Consumers of this pointer across an API boundary may safely
-// dereference the pointer.
-//
-// Example:
-//
-// // `employee` is designated as not null.
-// void PaySalary(absl::Nonnull<Employee *> employee) {
-//   pay(*employee);  // OK to dereference
-// }
-template <typename T>
-using Nonnull = nullability_internal::NonnullImpl<T>;
-
-// absl::Nullable
-//
-// The indicated pointer may, by design, be either null or non-null. Consumers
-// of this pointer across an API boundary should perform a `nullptr` check
-// before performing any operation using the pointer.
-//
-// Example:
-//
-// // `employee` may  be null.
-// void PaySalary(absl::Nullable<Employee *> employee) {
-//   if (employee != nullptr) {
-//     Pay(*employee);  // OK to dereference
-//   }
-// }
-template <typename T>
-using Nullable = nullability_internal::NullableImpl<T>;
-
-// absl::NullabilityUnknown (default)
-//
-// The indicated pointer has not yet been determined to be definitively
-// "non-null" or "nullable." Providers of such pointers across API boundaries
-// should, over time, annotate such pointers as either "non-null" or "nullable."
-// Consumers of these pointers across an API boundary should treat such pointers
-// with the same caution they treat currently unannotated pointers. Most
-// existing code will have "unknown"  pointers, which should eventually be
-// migrated into one of the above two nullability states: `Nonnull<T>` or
-//  `Nullable<T>`.
-//
-// NOTE: Because this annotation is the global default state, pointers without
-// any annotation are assumed to have "unknown" semantics. This assumption is
-// designed to minimize churn and reduce clutter within the codebase.
-//
-// Example:
-//
-// // `employee`s nullability state is unknown.
-// void PaySalary(absl::NullabilityUnknown<Employee *> employee) {
-//   Pay(*employee); // Potentially dangerous. API provider should investigate.
-// }
-//
-// Note that a pointer without an annotation, by default, is assumed to have the
-// annotation `NullabilityUnknown`.
-//
-// // `employee`s nullability state is unknown.
-// void PaySalary(Employee* employee) {
-//   Pay(*employee); // Potentially dangerous. API provider should investigate.
-// }
-template <typename T>
-using NullabilityUnknown = nullability_internal::NullabilityUnknownImpl<T>;
-
-}  // namespace absl
-
-#endif  // ABSL_BASE_NULLABILITY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/optimization.h b/projects/llm_framework/include/abseil-cpp/absl/base/optimization.h
deleted file mode 100644
index f9859958..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/optimization.h
+++ /dev/null
@@ -1,305 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: optimization.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines portable macros for performance optimization.
-
-#ifndef ABSL_BASE_OPTIMIZATION_H_
-#define ABSL_BASE_OPTIMIZATION_H_
-
-#include <assert.h>
-
-#include "absl/base/config.h"
-#include "absl/base/options.h"
-
-// ABSL_BLOCK_TAIL_CALL_OPTIMIZATION
-//
-// Instructs the compiler to avoid optimizing tail-call recursion. This macro is
-// useful when you wish to preserve the existing function order within a stack
-// trace for logging, debugging, or profiling purposes.
-//
-// Example:
-//
-//   int f() {
-//     int result = g();
-//     ABSL_BLOCK_TAIL_CALL_OPTIMIZATION();
-//     return result;
-//   }
-#if defined(__pnacl__)
-#define ABSL_BLOCK_TAIL_CALL_OPTIMIZATION() if (volatile int x = 0) { (void)x; }
-#elif defined(__clang__)
-// Clang will not tail call given inline volatile assembly.
-#define ABSL_BLOCK_TAIL_CALL_OPTIMIZATION() __asm__ __volatile__("")
-#elif defined(__GNUC__)
-// GCC will not tail call given inline volatile assembly.
-#define ABSL_BLOCK_TAIL_CALL_OPTIMIZATION() __asm__ __volatile__("")
-#elif defined(_MSC_VER)
-#include <intrin.h>
-// The __nop() intrinsic blocks the optimisation.
-#define ABSL_BLOCK_TAIL_CALL_OPTIMIZATION() __nop()
-#else
-#define ABSL_BLOCK_TAIL_CALL_OPTIMIZATION() if (volatile int x = 0) { (void)x; }
-#endif
-
-// ABSL_CACHELINE_SIZE
-//
-// Explicitly defines the size of the L1 cache for purposes of alignment.
-// Setting the cacheline size allows you to specify that certain objects be
-// aligned on a cacheline boundary with `ABSL_CACHELINE_ALIGNED` declarations.
-// (See below.)
-//
-// NOTE: this macro should be replaced with the following C++17 features, when
-// those are generally available:
-//
-//   * `std::hardware_constructive_interference_size`
-//   * `std::hardware_destructive_interference_size`
-//
-// See http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0154r1.html
-// for more information.
-#if defined(__GNUC__)
-// Cache line alignment
-#if defined(__i386__) || defined(__x86_64__)
-#define ABSL_CACHELINE_SIZE 64
-#elif defined(__powerpc64__)
-#define ABSL_CACHELINE_SIZE 128
-#elif defined(__aarch64__)
-// We would need to read special register ctr_el0 to find out L1 dcache size.
-// This value is a good estimate based on a real aarch64 machine.
-#define ABSL_CACHELINE_SIZE 64
-#elif defined(__arm__)
-// Cache line sizes for ARM: These values are not strictly correct since
-// cache line sizes depend on implementations, not architectures.  There
-// are even implementations with cache line sizes configurable at boot
-// time.
-#if defined(__ARM_ARCH_5T__)
-#define ABSL_CACHELINE_SIZE 32
-#elif defined(__ARM_ARCH_7A__)
-#define ABSL_CACHELINE_SIZE 64
-#endif
-#endif
-#endif
-
-#ifndef ABSL_CACHELINE_SIZE
-// A reasonable default guess.  Note that overestimates tend to waste more
-// space, while underestimates tend to waste more time.
-#define ABSL_CACHELINE_SIZE 64
-#endif
-
-// ABSL_CACHELINE_ALIGNED
-//
-// Indicates that the declared object be cache aligned using
-// `ABSL_CACHELINE_SIZE` (see above). Cacheline aligning objects allows you to
-// load a set of related objects in the L1 cache for performance improvements.
-// Cacheline aligning objects properly allows constructive memory sharing and
-// prevents destructive (or "false") memory sharing.
-//
-// NOTE: callers should replace uses of this macro with `alignas()` using
-// `std::hardware_constructive_interference_size` and/or
-// `std::hardware_destructive_interference_size` when C++17 becomes available to
-// them.
-//
-// See http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0154r1.html
-// for more information.
-//
-// On some compilers, `ABSL_CACHELINE_ALIGNED` expands to an `__attribute__`
-// or `__declspec` attribute. For compilers where this is not known to work,
-// the macro expands to nothing.
-//
-// No further guarantees are made here. The result of applying the macro
-// to variables and types is always implementation-defined.
-//
-// WARNING: It is easy to use this attribute incorrectly, even to the point
-// of causing bugs that are difficult to diagnose, crash, etc. It does not
-// of itself guarantee that objects are aligned to a cache line.
-//
-// NOTE: Some compilers are picky about the locations of annotations such as
-// this attribute, so prefer to put it at the beginning of your declaration.
-// For example,
-//
-//   ABSL_CACHELINE_ALIGNED static Foo* foo = ...
-//
-//   class ABSL_CACHELINE_ALIGNED Bar { ...
-//
-// Recommendations:
-//
-// 1) Consult compiler documentation; this comment is not kept in sync as
-//    toolchains evolve.
-// 2) Verify your use has the intended effect. This often requires inspecting
-//    the generated machine code.
-// 3) Prefer applying this attribute to individual variables. Avoid
-//    applying it to types. This tends to localize the effect.
-#if defined(__clang__) || defined(__GNUC__)
-#define ABSL_CACHELINE_ALIGNED __attribute__((aligned(ABSL_CACHELINE_SIZE)))
-#elif defined(_MSC_VER)
-#define ABSL_CACHELINE_ALIGNED __declspec(align(ABSL_CACHELINE_SIZE))
-#else
-#define ABSL_CACHELINE_ALIGNED
-#endif
-
-// ABSL_PREDICT_TRUE, ABSL_PREDICT_FALSE
-//
-// Enables the compiler to prioritize compilation using static analysis for
-// likely paths within a boolean branch.
-//
-// Example:
-//
-//   if (ABSL_PREDICT_TRUE(expression)) {
-//     return result;                        // Faster if more likely
-//   } else {
-//     return 0;
-//   }
-//
-// Compilers can use the information that a certain branch is not likely to be
-// taken (for instance, a CHECK failure) to optimize for the common case in
-// the absence of better information (ie. compiling gcc with `-fprofile-arcs`).
-//
-// Recommendation: Modern CPUs dynamically predict branch execution paths,
-// typically with accuracy greater than 97%. As a result, annotating every
-// branch in a codebase is likely counterproductive; however, annotating
-// specific branches that are both hot and consistently mispredicted is likely
-// to yield performance improvements.
-#if ABSL_HAVE_BUILTIN(__builtin_expect) || \
-    (defined(__GNUC__) && !defined(__clang__))
-#define ABSL_PREDICT_FALSE(x) (__builtin_expect(false || (x), false))
-#define ABSL_PREDICT_TRUE(x) (__builtin_expect(false || (x), true))
-#else
-#define ABSL_PREDICT_FALSE(x) (x)
-#define ABSL_PREDICT_TRUE(x) (x)
-#endif
-
-// `ABSL_INTERNAL_IMMEDIATE_ABORT_IMPL()` aborts the program in the fastest
-// possible way, with no attempt at logging. One use is to implement hardening
-// aborts with ABSL_OPTION_HARDENED.  Since this is an internal symbol, it
-// should not be used directly outside of Abseil.
-#if ABSL_HAVE_BUILTIN(__builtin_trap) || \
-    (defined(__GNUC__) && !defined(__clang__))
-#define ABSL_INTERNAL_IMMEDIATE_ABORT_IMPL() __builtin_trap()
-#else
-#define ABSL_INTERNAL_IMMEDIATE_ABORT_IMPL() abort()
-#endif
-
-// `ABSL_INTERNAL_UNREACHABLE_IMPL()` is the platform specific directive to
-// indicate that a statement is unreachable, and to allow the compiler to
-// optimize accordingly. Clients should use `ABSL_UNREACHABLE()`, which is
-// defined below.
-#if defined(__cpp_lib_unreachable) && __cpp_lib_unreachable >= 202202L
-#define ABSL_INTERNAL_UNREACHABLE_IMPL() std::unreachable()
-#elif defined(__GNUC__) || ABSL_HAVE_BUILTIN(__builtin_unreachable)
-#define ABSL_INTERNAL_UNREACHABLE_IMPL() __builtin_unreachable()
-#elif ABSL_HAVE_BUILTIN(__builtin_assume)
-#define ABSL_INTERNAL_UNREACHABLE_IMPL() __builtin_assume(false)
-#elif defined(_MSC_VER)
-#define ABSL_INTERNAL_UNREACHABLE_IMPL() __assume(false)
-#else
-#define ABSL_INTERNAL_UNREACHABLE_IMPL()
-#endif
-
-// `ABSL_UNREACHABLE()` is an unreachable statement.  A program which reaches
-// one has undefined behavior, and the compiler may optimize accordingly.
-#if ABSL_OPTION_HARDENED == 1 && defined(NDEBUG)
-// Abort in hardened mode to avoid dangerous undefined behavior.
-#define ABSL_UNREACHABLE()                \
-  do {                                    \
-    ABSL_INTERNAL_IMMEDIATE_ABORT_IMPL(); \
-    ABSL_INTERNAL_UNREACHABLE_IMPL();     \
-  } while (false)
-#else
-// The assert only fires in debug mode to aid in debugging.
-// When NDEBUG is defined, reaching ABSL_UNREACHABLE() is undefined behavior.
-#define ABSL_UNREACHABLE()                       \
-  do {                                           \
-    /* NOLINTNEXTLINE: misc-static-assert */     \
-    assert(false && "ABSL_UNREACHABLE reached"); \
-    ABSL_INTERNAL_UNREACHABLE_IMPL();            \
-  } while (false)
-#endif
-
-// ABSL_ASSUME(cond)
-//
-// Informs the compiler that a condition is always true and that it can assume
-// it to be true for optimization purposes.
-//
-// WARNING: If the condition is false, the program can produce undefined and
-// potentially dangerous behavior.
-//
-// In !NDEBUG mode, the condition is checked with an assert().
-//
-// NOTE: The expression must not have side effects, as it may only be evaluated
-// in some compilation modes and not others. Some compilers may issue a warning
-// if the compiler cannot prove the expression has no side effects. For example,
-// the expression should not use a function call since the compiler cannot prove
-// that a function call does not have side effects.
-//
-// Example:
-//
-//   int x = ...;
-//   ABSL_ASSUME(x >= 0);
-//   // The compiler can optimize the division to a simple right shift using the
-//   // assumption specified above.
-//   int y = x / 16;
-//
-#if !defined(NDEBUG)
-#define ABSL_ASSUME(cond) assert(cond)
-#elif ABSL_HAVE_BUILTIN(__builtin_assume)
-#define ABSL_ASSUME(cond) __builtin_assume(cond)
-#elif defined(_MSC_VER)
-#define ABSL_ASSUME(cond) __assume(cond)
-#elif defined(__cpp_lib_unreachable) && __cpp_lib_unreachable >= 202202L
-#define ABSL_ASSUME(cond)            \
-  do {                               \
-    if (!(cond)) std::unreachable(); \
-  } while (false)
-#elif defined(__GNUC__) || ABSL_HAVE_BUILTIN(__builtin_unreachable)
-#define ABSL_ASSUME(cond)                 \
-  do {                                    \
-    if (!(cond)) __builtin_unreachable(); \
-  } while (false)
-#else
-#define ABSL_ASSUME(cond)               \
-  do {                                  \
-    static_cast<void>(false && (cond)); \
-  } while (false)
-#endif
-
-// ABSL_INTERNAL_UNIQUE_SMALL_NAME(cond)
-// This macro forces small unique name on a static file level symbols like
-// static local variables or static functions. This is intended to be used in
-// macro definitions to optimize the cost of generated code. Do NOT use it on
-// symbols exported from translation unit since it may cause a link time
-// conflict.
-//
-// Example:
-//
-// #define MY_MACRO(txt)
-// namespace {
-//  char VeryVeryLongVarName[] ABSL_INTERNAL_UNIQUE_SMALL_NAME() = txt;
-//  const char* VeryVeryLongFuncName() ABSL_INTERNAL_UNIQUE_SMALL_NAME();
-//  const char* VeryVeryLongFuncName() { return txt; }
-// }
-//
-
-#if defined(__GNUC__)
-#define ABSL_INTERNAL_UNIQUE_SMALL_NAME2(x) #x
-#define ABSL_INTERNAL_UNIQUE_SMALL_NAME1(x) ABSL_INTERNAL_UNIQUE_SMALL_NAME2(x)
-#define ABSL_INTERNAL_UNIQUE_SMALL_NAME() \
-  asm(ABSL_INTERNAL_UNIQUE_SMALL_NAME1(.absl.__COUNTER__))
-#else
-#define ABSL_INTERNAL_UNIQUE_SMALL_NAME()
-#endif
-
-#endif  // ABSL_BASE_OPTIMIZATION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/options.h b/projects/llm_framework/include/abseil-cpp/absl/base/options.h
deleted file mode 100644
index a169658f..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/options.h
+++ /dev/null
@@ -1,258 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: options.h
-// -----------------------------------------------------------------------------
-//
-// This file contains Abseil configuration options for setting specific
-// implementations instead of letting Abseil determine which implementation to
-// use at compile-time. Setting these options may be useful for package or build
-// managers who wish to guarantee ABI stability within binary builds (which are
-// otherwise difficult to enforce).
-//
-// *** IMPORTANT NOTICE FOR PACKAGE MANAGERS:  It is important that
-// maintainers of package managers who wish to package Abseil read and
-// understand this file! ***
-//
-// Abseil contains a number of possible configuration endpoints, based on
-// parameters such as the detected platform, language version, or command-line
-// flags used to invoke the underlying binary. As is the case with all
-// libraries, binaries which contain Abseil code must ensure that separate
-// packages use the same compiled copy of Abseil to avoid a diamond dependency
-// problem, which can occur if two packages built with different Abseil
-// configuration settings are linked together. Diamond dependency problems in
-// C++ may manifest as violations to the One Definition Rule (ODR) (resulting in
-// linker errors), or undefined behavior (resulting in crashes).
-//
-// Diamond dependency problems can be avoided if all packages utilize the same
-// exact version of Abseil. Building from source code with the same compilation
-// parameters is the easiest way to avoid such dependency problems. However, for
-// package managers who cannot control such compilation parameters, we are
-// providing the file to allow you to inject ABI (Application Binary Interface)
-// stability across builds. Settings options in this file will neither change
-// API nor ABI, providing a stable copy of Abseil between packages.
-//
-// Care must be taken to keep options within these configurations isolated
-// from any other dynamic settings, such as command-line flags which could alter
-// these options. This file is provided specifically to help build and package
-// managers provide a stable copy of Abseil within their libraries and binaries;
-// other developers should not have need to alter the contents of this file.
-//
-// -----------------------------------------------------------------------------
-// Usage
-// -----------------------------------------------------------------------------
-//
-// For any particular package release, set the appropriate definitions within
-// this file to whatever value makes the most sense for your package(s). Note
-// that, by default, most of these options, at the moment, affect the
-// implementation of types; future options may affect other implementation
-// details.
-//
-// NOTE: the defaults within this file all assume that Abseil can select the
-// proper Abseil implementation at compile-time, which will not be sufficient
-// to guarantee ABI stability to package managers.
-
-#ifndef ABSL_BASE_OPTIONS_H_
-#define ABSL_BASE_OPTIONS_H_
-
-// -----------------------------------------------------------------------------
-// Type Compatibility Options
-// -----------------------------------------------------------------------------
-//
-// ABSL_OPTION_USE_STD_ANY
-//
-// This option controls whether absl::any is implemented as an alias to
-// std::any, or as an independent implementation.
-//
-// A value of 0 means to use Abseil's implementation.  This requires only C++11
-// support, and is expected to work on every toolchain we support.
-//
-// A value of 1 means to use an alias to std::any.  This requires that all code
-// using Abseil is built in C++17 mode or later.
-//
-// A value of 2 means to detect the C++ version being used to compile Abseil,
-// and use an alias only if a working std::any is available.  This option is
-// useful when you are building your entire program, including all of its
-// dependencies, from source.  It should not be used otherwise -- for example,
-// if you are distributing Abseil in a binary package manager -- since in
-// mode 2, absl::any will name a different type, with a different mangled name
-// and binary layout, depending on the compiler flags passed by the end user.
-// For more info, see https://abseil.io/about/design/dropin-types.
-//
-// User code should not inspect this macro.  To check in the preprocessor if
-// absl::any is a typedef of std::any, use the feature macro ABSL_USES_STD_ANY.
-
-#define ABSL_OPTION_USE_STD_ANY 2
-
-
-// ABSL_OPTION_USE_STD_OPTIONAL
-//
-// This option controls whether absl::optional is implemented as an alias to
-// std::optional, or as an independent implementation.
-//
-// A value of 0 means to use Abseil's implementation.  This requires only C++11
-// support, and is expected to work on every toolchain we support.
-//
-// A value of 1 means to use an alias to std::optional.  This requires that all
-// code using Abseil is built in C++17 mode or later.
-//
-// A value of 2 means to detect the C++ version being used to compile Abseil,
-// and use an alias only if a working std::optional is available.  This option
-// is useful when you are building your program from source.  It should not be
-// used otherwise -- for example, if you are distributing Abseil in a binary
-// package manager -- since in mode 2, absl::optional will name a different
-// type, with a different mangled name and binary layout, depending on the
-// compiler flags passed by the end user.  For more info, see
-// https://abseil.io/about/design/dropin-types.
-
-// User code should not inspect this macro.  To check in the preprocessor if
-// absl::optional is a typedef of std::optional, use the feature macro
-// ABSL_USES_STD_OPTIONAL.
-
-#define ABSL_OPTION_USE_STD_OPTIONAL 2
-
-
-// ABSL_OPTION_USE_STD_STRING_VIEW
-//
-// This option controls whether absl::string_view is implemented as an alias to
-// std::string_view, or as an independent implementation.
-//
-// A value of 0 means to use Abseil's implementation.  This requires only C++11
-// support, and is expected to work on every toolchain we support.
-//
-// A value of 1 means to use an alias to std::string_view.  This requires that
-// all code using Abseil is built in C++17 mode or later.
-//
-// A value of 2 means to detect the C++ version being used to compile Abseil,
-// and use an alias only if a working std::string_view is available.  This
-// option is useful when you are building your program from source.  It should
-// not be used otherwise -- for example, if you are distributing Abseil in a
-// binary package manager -- since in mode 2, absl::string_view will name a
-// different type, with a different mangled name and binary layout, depending on
-// the compiler flags passed by the end user.  For more info, see
-// https://abseil.io/about/design/dropin-types.
-//
-// User code should not inspect this macro.  To check in the preprocessor if
-// absl::string_view is a typedef of std::string_view, use the feature macro
-// ABSL_USES_STD_STRING_VIEW.
-
-#define ABSL_OPTION_USE_STD_STRING_VIEW 2
-
-// ABSL_OPTION_USE_STD_VARIANT
-//
-// This option controls whether absl::variant is implemented as an alias to
-// std::variant, or as an independent implementation.
-//
-// A value of 0 means to use Abseil's implementation.  This requires only C++11
-// support, and is expected to work on every toolchain we support.
-//
-// A value of 1 means to use an alias to std::variant.  This requires that all
-// code using Abseil is built in C++17 mode or later.
-//
-// A value of 2 means to detect the C++ version being used to compile Abseil,
-// and use an alias only if a working std::variant is available.  This option
-// is useful when you are building your program from source.  It should not be
-// used otherwise -- for example, if you are distributing Abseil in a binary
-// package manager -- since in mode 2, absl::variant will name a different
-// type, with a different mangled name and binary layout, depending on the
-// compiler flags passed by the end user.  For more info, see
-// https://abseil.io/about/design/dropin-types.
-//
-// User code should not inspect this macro.  To check in the preprocessor if
-// absl::variant is a typedef of std::variant, use the feature macro
-// ABSL_USES_STD_VARIANT.
-
-#define ABSL_OPTION_USE_STD_VARIANT 2
-
-// ABSL_OPTION_USE_STD_ORDERING
-//
-// This option controls whether absl::{partial,weak,strong}_ordering are
-// implemented as aliases to the std:: ordering types, or as an independent
-// implementation.
-//
-// A value of 0 means to use Abseil's implementation.  This requires only C++11
-// support, and is expected to work on every toolchain we support.
-//
-// A value of 1 means to use aliases.  This requires that all code using Abseil
-// is built in C++20 mode or later.
-//
-// A value of 2 means to detect the C++ version being used to compile Abseil,
-// and use an alias only if working std:: ordering types are available.  This
-// option is useful when you are building your program from source.  It should
-// not be used otherwise -- for example, if you are distributing Abseil in a
-// binary package manager -- since in mode 2, they will name different types,
-// with different mangled names and binary layout, depending on the compiler
-// flags passed by the end user.  For more info, see
-// https://abseil.io/about/design/dropin-types.
-//
-// User code should not inspect this macro.  To check in the preprocessor if
-// the ordering types are aliases of std:: ordering types, use the feature macro
-// ABSL_USES_STD_ORDERING.
-
-#define ABSL_OPTION_USE_STD_ORDERING 2
-
-// ABSL_OPTION_USE_INLINE_NAMESPACE
-// ABSL_OPTION_INLINE_NAMESPACE_NAME
-//
-// These options controls whether all entities in the absl namespace are
-// contained within an inner inline namespace.  This does not affect the
-// user-visible API of Abseil, but it changes the mangled names of all symbols.
-//
-// This can be useful as a version tag if you are distributing Abseil in
-// precompiled form.  This will prevent a binary library build of Abseil with
-// one inline namespace being used with headers configured with a different
-// inline namespace name.  Binary packagers are reminded that Abseil does not
-// guarantee any ABI stability in Abseil, so any update of Abseil or
-// configuration change in such a binary package should be combined with a
-// new, unique value for the inline namespace name.
-//
-// A value of 0 means not to use inline namespaces.
-//
-// A value of 1 means to use an inline namespace with the given name inside
-// namespace absl.  If this is set, ABSL_OPTION_INLINE_NAMESPACE_NAME must also
-// be changed to a new, unique identifier name.  In particular "head" is not
-// allowed.
-
-#define ABSL_OPTION_USE_INLINE_NAMESPACE 0
-#define ABSL_OPTION_INLINE_NAMESPACE_NAME head
-
-// ABSL_OPTION_HARDENED
-//
-// This option enables a "hardened" build in release mode (in this context,
-// release mode is defined as a build where the `NDEBUG` macro is defined).
-//
-// A value of 0 means that "hardened" mode is not enabled.
-//
-// A value of 1 means that "hardened" mode is enabled.
-//
-// Hardened builds have additional security checks enabled when `NDEBUG` is
-// defined. Defining `NDEBUG` is normally used to turn `assert()` macro into a
-// no-op, as well as disabling other bespoke program consistency checks. By
-// defining ABSL_OPTION_HARDENED to 1, a select set of checks remain enabled in
-// release mode. These checks guard against programming errors that may lead to
-// security vulnerabilities. In release mode, when one of these programming
-// errors is encountered, the program will immediately abort, possibly without
-// any attempt at logging.
-//
-// The checks enabled by this option are not free; they do incur runtime cost.
-//
-// The checks enabled by this option are always active when `NDEBUG` is not
-// defined, even in the case when ABSL_OPTION_HARDENED is defined to 0. The
-// checks enabled by this option may abort the program in a different way and
-// log additional information when `NDEBUG` is not defined.
-
-#define ABSL_OPTION_HARDENED 0
-
-#endif  // ABSL_BASE_OPTIONS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/policy_checks.h b/projects/llm_framework/include/abseil-cpp/absl/base/policy_checks.h
deleted file mode 100644
index 372e848d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/policy_checks.h
+++ /dev/null
@@ -1,113 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: policy_checks.h
-// -----------------------------------------------------------------------------
-//
-// This header enforces a minimum set of policies at build time, such as the
-// supported compiler and library versions. Unsupported configurations are
-// reported with `#error`. This enforcement is best effort, so successfully
-// compiling this header does not guarantee a supported configuration.
-
-#ifndef ABSL_BASE_POLICY_CHECKS_H_
-#define ABSL_BASE_POLICY_CHECKS_H_
-
-// Included for the __GLIBC_PREREQ macro used below.
-#include <limits.h>
-
-// Included for the _STLPORT_VERSION macro used below.
-#if defined(__cplusplus)
-#include <cstddef>
-#endif
-
-// -----------------------------------------------------------------------------
-// Operating System Check
-// -----------------------------------------------------------------------------
-
-#if defined(__CYGWIN__)
-#error "Cygwin is not supported."
-#endif
-
-// -----------------------------------------------------------------------------
-// Toolchain Check
-// -----------------------------------------------------------------------------
-
-// We support Visual Studio 2019 (MSVC++ 16.0) and later.
-// This minimum will go up.
-#if defined(_MSC_VER) && _MSC_VER < 1920 && !defined(__clang__)
-#error "This package requires Visual Studio 2019 (MSVC++ 16.0) or higher."
-#endif
-
-// We support GCC 7 and later.
-// This minimum will go up.
-#if defined(__GNUC__) && !defined(__clang__)
-#if __GNUC__ < 7
-#error "This package requires GCC 7 or higher."
-#endif
-#endif
-
-// We support Apple Xcode clang 4.2.1 (version 421.11.65) and later.
-// This corresponds to Apple Xcode version 4.5.
-// This minimum will go up.
-#if defined(__apple_build_version__) && __apple_build_version__ < 4211165
-#error "This package requires __apple_build_version__ of 4211165 or higher."
-#endif
-
-// -----------------------------------------------------------------------------
-// C++ Version Check
-// -----------------------------------------------------------------------------
-
-// Enforce C++14 as the minimum.
-#if defined(_MSVC_LANG)
-#if _MSVC_LANG < 201402L
-#error "C++ versions less than C++14 are not supported."
-#endif  // _MSVC_LANG < 201402L
-#elif defined(__cplusplus)
-#if __cplusplus < 201402L
-#error "C++ versions less than C++14 are not supported."
-#endif  // __cplusplus < 201402L
-#endif
-
-// -----------------------------------------------------------------------------
-// Standard Library Check
-// -----------------------------------------------------------------------------
-
-#if defined(_STLPORT_VERSION)
-#error "STLPort is not supported."
-#endif
-
-// -----------------------------------------------------------------------------
-// `char` Size Check
-// -----------------------------------------------------------------------------
-
-// Abseil currently assumes CHAR_BIT == 8. If you would like to use Abseil on a
-// platform where this is not the case, please provide us with the details about
-// your platform so we can consider relaxing this requirement.
-#if CHAR_BIT != 8
-#error "Abseil assumes CHAR_BIT == 8."
-#endif
-
-// -----------------------------------------------------------------------------
-// `int` Size Check
-// -----------------------------------------------------------------------------
-
-// Abseil currently assumes that an int is 4 bytes. If you would like to use
-// Abseil on a platform where this is not the case, please provide us with the
-// details about your platform so we can consider relaxing this requirement.
-#if INT_MAX < 2147483647
-#error "Abseil assumes that int is at least 4 bytes. "
-#endif
-
-#endif  // ABSL_BASE_POLICY_CHECKS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/port.h b/projects/llm_framework/include/abseil-cpp/absl/base/port.h
deleted file mode 100644
index 5bc4d6cd..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/port.h
+++ /dev/null
@@ -1,25 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// This files is a forwarding header for other headers containing various
-// portability macros and functions.
-
-#ifndef ABSL_BASE_PORT_H_
-#define ABSL_BASE_PORT_H_
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/optimization.h"
-
-#endif  // ABSL_BASE_PORT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/prefetch.h b/projects/llm_framework/include/abseil-cpp/absl/base/prefetch.h
deleted file mode 100644
index 482cde3a..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/prefetch.h
+++ /dev/null
@@ -1,209 +0,0 @@
-// Copyright 2023 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: prefetch.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines prefetch functions to prefetch memory contents
-// into the first level cache (L1) for the current CPU. The prefetch logic
-// offered in this header is limited to prefetching first level cachelines
-// only, and is aimed at relatively 'simple' prefetching logic.
-//
-#ifndef ABSL_BASE_PREFETCH_H_
-#define ABSL_BASE_PREFETCH_H_
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-
-#if defined(ABSL_INTERNAL_HAVE_SSE)
-#include <xmmintrin.h>
-#endif
-
-#if defined(_MSC_VER)
-#include <intrin.h>
-#if defined(ABSL_INTERNAL_HAVE_SSE)
-#pragma intrinsic(_mm_prefetch)
-#endif
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// Moves data into the L1 cache before it is read, or "prefetches" it.
-//
-// The value of `addr` is the address of the memory to prefetch. If
-// the target and compiler support it, data prefetch instructions are
-// generated. If the prefetch is done some time before the memory is
-// read, it may be in the cache by the time the read occurs.
-//
-// This method prefetches data with the highest degree of temporal locality;
-// data is prefetched where possible into all levels of the cache.
-//
-// Incorrect or gratuitous use of this function can degrade performance.
-// Use this function only when representative benchmarks show an improvement.
-//
-// Example:
-//
-//  // Computes incremental checksum for `data`.
-//  int ComputeChecksum(int sum, absl::string_view data);
-//
-//  // Computes cumulative checksum for all values in `data`
-//  int ComputeChecksum(absl::Span<const std::string> data) {
-//    int sum = 0;
-//    auto it = data.begin();
-//    auto pit = data.begin();
-//    auto end = data.end();
-//    for (int dist = 8; dist > 0 && pit != data.end(); --dist, ++pit) {
-//      absl::PrefetchToLocalCache(pit->data());
-//    }
-//    for (; pit != end; ++pit, ++it) {
-//      sum = ComputeChecksum(sum, *it);
-//      absl::PrefetchToLocalCache(pit->data());
-//    }
-//    for (; it != end; ++it) {
-//      sum = ComputeChecksum(sum, *it);
-//    }
-//    return sum;
-//  }
-//
-void PrefetchToLocalCache(const void* addr);
-
-// Moves data into the L1 cache before it is read, or "prefetches" it.
-//
-// This function is identical to `PrefetchToLocalCache()` except that it has
-// non-temporal locality: the fetched data should not be left in any of the
-// cache tiers. This is useful for cases where the data is used only once /
-// short term, for example, invoking a destructor on an object.
-//
-// Incorrect or gratuitous use of this function can degrade performance.
-// Use this function only when representative benchmarks show an improvement.
-//
-// Example:
-//
-//  template <typename Iterator>
-//  void DestroyPointers(Iterator begin, Iterator end) {
-//    size_t distance = std::min(8U, bars.size());
-//
-//    int dist = 8;
-//    auto prefetch_it = begin;
-//    while (prefetch_it != end && --dist;) {
-//      absl::PrefetchToLocalCacheNta(*prefetch_it++);
-//    }
-//    while (prefetch_it != end) {
-//      delete *begin++;
-//      absl::PrefetchToLocalCacheNta(*prefetch_it++);
-//    }
-//    while (begin != end) {
-//      delete *begin++;
-//    }
-//  }
-//
-void PrefetchToLocalCacheNta(const void* addr);
-
-// Moves data into the L1 cache with the intent to modify it.
-//
-// This function is similar to `PrefetchToLocalCache()` except that it
-// prefetches cachelines with an 'intent to modify' This typically includes
-// invalidating cache entries for this address in all other cache tiers, and an
-// exclusive access intent.
-//
-// Incorrect or gratuitous use of this function can degrade performance. As this
-// function can invalidate cached cachelines on other caches and computer cores,
-// incorrect usage of this function can have an even greater negative impact
-// than incorrect regular prefetches.
-// Use this function only when representative benchmarks show an improvement.
-//
-// Example:
-//
-//  void* Arena::Allocate(size_t size) {
-//    void* ptr = AllocateBlock(size);
-//    absl::PrefetchToLocalCacheForWrite(ptr);
-//    return ptr;
-//  }
-//
-void PrefetchToLocalCacheForWrite(const void* addr);
-
-#if ABSL_HAVE_BUILTIN(__builtin_prefetch) || defined(__GNUC__)
-
-#define ABSL_HAVE_PREFETCH 1
-
-// See __builtin_prefetch:
-// https://gcc.gnu.org/onlinedocs/gcc/Other-Builtins.html.
-//
-ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCache(
-    const void* addr) {
-  __builtin_prefetch(addr, 0, 3);
-}
-
-ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCacheNta(
-    const void* addr) {
-  __builtin_prefetch(addr, 0, 0);
-}
-
-ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCacheForWrite(
-    const void* addr) {
-  // [x86] gcc/clang don't generate PREFETCHW for __builtin_prefetch(.., 1)
-  // unless -march=broadwell or newer; this is not generally the default, so we
-  // manually emit prefetchw. PREFETCHW is recognized as a no-op on older Intel
-  // processors and has been present on AMD processors since the K6-2.
-#if defined(__x86_64__) && !defined(__PRFCHW__)
-  asm("prefetchw %0" : : "m"(*reinterpret_cast<const char*>(addr)));
-#else
-  __builtin_prefetch(addr, 1, 3);
-#endif
-}
-
-#elif defined(ABSL_INTERNAL_HAVE_SSE)
-
-#define ABSL_HAVE_PREFETCH 1
-
-ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCache(
-    const void* addr) {
-  _mm_prefetch(reinterpret_cast<const char*>(addr), _MM_HINT_T0);
-}
-
-ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCacheNta(
-    const void* addr) {
-  _mm_prefetch(reinterpret_cast<const char*>(addr), _MM_HINT_NTA);
-}
-
-ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCacheForWrite(
-    const void* addr) {
-#if defined(_MM_HINT_ET0)
-  _mm_prefetch(reinterpret_cast<const char*>(addr), _MM_HINT_ET0);
-#elif !defined(_MSC_VER) && defined(__x86_64__)
-  // _MM_HINT_ET0 is not universally supported. As we commented further
-  // up, PREFETCHW is recognized as a no-op on older Intel processors
-  // and has been present on AMD processors since the K6-2. We have this
-  // disabled for MSVC compilers as this miscompiles on older MSVC compilers.
-  asm("prefetchw %0" : : "m"(*reinterpret_cast<const char*>(addr)));
-#endif
-}
-
-#else
-
-ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCache(
-    const void* addr) {}
-ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCacheNta(
-    const void* addr) {}
-ABSL_ATTRIBUTE_ALWAYS_INLINE inline void PrefetchToLocalCacheForWrite(
-    const void* addr) {}
-
-#endif
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_PREFETCH_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/base/thread_annotations.h b/projects/llm_framework/include/abseil-cpp/absl/base/thread_annotations.h
deleted file mode 100644
index 4a3f3e33..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/base/thread_annotations.h
+++ /dev/null
@@ -1,333 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: thread_annotations.h
-// -----------------------------------------------------------------------------
-//
-// This header file contains macro definitions for thread safety annotations
-// that allow developers to document the locking policies of multi-threaded
-// code. The annotations can also help program analysis tools to identify
-// potential thread safety issues.
-//
-// These annotations are implemented using compiler attributes. Using the macros
-// defined here instead of raw attributes allow for portability and future
-// compatibility.
-//
-// When referring to mutexes in the arguments of the attributes, you should
-// use variable names or more complex expressions (e.g. my_object->mutex_)
-// that evaluate to a concrete mutex object whenever possible. If the mutex
-// you want to refer to is not in scope, you may use a member pointer
-// (e.g. &MyClass::mutex_) to refer to a mutex in some (unknown) object.
-
-#ifndef ABSL_BASE_THREAD_ANNOTATIONS_H_
-#define ABSL_BASE_THREAD_ANNOTATIONS_H_
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-
-// ABSL_GUARDED_BY()
-//
-// Documents if a shared field or global variable needs to be protected by a
-// mutex. ABSL_GUARDED_BY() allows the user to specify a particular mutex that
-// should be held when accessing the annotated variable.
-//
-// Although this annotation (and ABSL_PT_GUARDED_BY, below) cannot be applied to
-// local variables, a local variable and its associated mutex can often be
-// combined into a small class or struct, thereby allowing the annotation.
-//
-// Example:
-//
-//   class Foo {
-//     Mutex mu_;
-//     int p1_ ABSL_GUARDED_BY(mu_);
-//     ...
-//   };
-#if ABSL_HAVE_ATTRIBUTE(guarded_by)
-#define ABSL_GUARDED_BY(x) __attribute__((guarded_by(x)))
-#else
-#define ABSL_GUARDED_BY(x)
-#endif
-
-// ABSL_PT_GUARDED_BY()
-//
-// Documents if the memory location pointed to by a pointer should be guarded
-// by a mutex when dereferencing the pointer.
-//
-// Example:
-//   class Foo {
-//     Mutex mu_;
-//     int *p1_ ABSL_PT_GUARDED_BY(mu_);
-//     ...
-//   };
-//
-// Note that a pointer variable to a shared memory location could itself be a
-// shared variable.
-//
-// Example:
-//
-//   // `q_`, guarded by `mu1_`, points to a shared memory location that is
-//   // guarded by `mu2_`:
-//   int *q_ ABSL_GUARDED_BY(mu1_) ABSL_PT_GUARDED_BY(mu2_);
-#if ABSL_HAVE_ATTRIBUTE(pt_guarded_by)
-#define ABSL_PT_GUARDED_BY(x) __attribute__((pt_guarded_by(x)))
-#else
-#define ABSL_PT_GUARDED_BY(x)
-#endif
-
-// ABSL_ACQUIRED_AFTER() / ABSL_ACQUIRED_BEFORE()
-//
-// Documents the acquisition order between locks that can be held
-// simultaneously by a thread. For any two locks that need to be annotated
-// to establish an acquisition order, only one of them needs the annotation.
-// (i.e. You don't have to annotate both locks with both ABSL_ACQUIRED_AFTER
-// and ABSL_ACQUIRED_BEFORE.)
-//
-// As with ABSL_GUARDED_BY, this is only applicable to mutexes that are shared
-// fields or global variables.
-//
-// Example:
-//
-//   Mutex m1_;
-//   Mutex m2_ ABSL_ACQUIRED_AFTER(m1_);
-#if ABSL_HAVE_ATTRIBUTE(acquired_after)
-#define ABSL_ACQUIRED_AFTER(...) __attribute__((acquired_after(__VA_ARGS__)))
-#else
-#define ABSL_ACQUIRED_AFTER(...)
-#endif
-
-#if ABSL_HAVE_ATTRIBUTE(acquired_before)
-#define ABSL_ACQUIRED_BEFORE(...) __attribute__((acquired_before(__VA_ARGS__)))
-#else
-#define ABSL_ACQUIRED_BEFORE(...)
-#endif
-
-// ABSL_EXCLUSIVE_LOCKS_REQUIRED() / ABSL_SHARED_LOCKS_REQUIRED()
-//
-// Documents a function that expects a mutex to be held prior to entry.
-// The mutex is expected to be held both on entry to, and exit from, the
-// function.
-//
-// An exclusive lock allows read-write access to the guarded data member(s), and
-// only one thread can acquire a lock exclusively at any one time. A shared lock
-// allows read-only access, and any number of threads can acquire a shared lock
-// concurrently.
-//
-// Generally, non-const methods should be annotated with
-// ABSL_EXCLUSIVE_LOCKS_REQUIRED, while const methods should be annotated with
-// ABSL_SHARED_LOCKS_REQUIRED.
-//
-// Example:
-//
-//   Mutex mu1, mu2;
-//   int a ABSL_GUARDED_BY(mu1);
-//   int b ABSL_GUARDED_BY(mu2);
-//
-//   void foo() ABSL_EXCLUSIVE_LOCKS_REQUIRED(mu1, mu2) { ... }
-//   void bar() const ABSL_SHARED_LOCKS_REQUIRED(mu1, mu2) { ... }
-#if ABSL_HAVE_ATTRIBUTE(exclusive_locks_required)
-#define ABSL_EXCLUSIVE_LOCKS_REQUIRED(...) \
-  __attribute__((exclusive_locks_required(__VA_ARGS__)))
-#else
-#define ABSL_EXCLUSIVE_LOCKS_REQUIRED(...)
-#endif
-
-#if ABSL_HAVE_ATTRIBUTE(shared_locks_required)
-#define ABSL_SHARED_LOCKS_REQUIRED(...) \
-  __attribute__((shared_locks_required(__VA_ARGS__)))
-#else
-#define ABSL_SHARED_LOCKS_REQUIRED(...)
-#endif
-
-// ABSL_LOCKS_EXCLUDED()
-//
-// Documents the locks that cannot be held by callers of this function, as they
-// might be acquired by this function (Abseil's `Mutex` locks are
-// non-reentrant).
-#if ABSL_HAVE_ATTRIBUTE(locks_excluded)
-#define ABSL_LOCKS_EXCLUDED(...) __attribute__((locks_excluded(__VA_ARGS__)))
-#else
-#define ABSL_LOCKS_EXCLUDED(...)
-#endif
-
-// ABSL_LOCK_RETURNED()
-//
-// Documents a function that returns a mutex without acquiring it.  For example,
-// a public getter method that returns a pointer to a private mutex should
-// be annotated with ABSL_LOCK_RETURNED.
-#if ABSL_HAVE_ATTRIBUTE(lock_returned)
-#define ABSL_LOCK_RETURNED(x) __attribute__((lock_returned(x)))
-#else
-#define ABSL_LOCK_RETURNED(x)
-#endif
-
-// ABSL_LOCKABLE
-//
-// Documents if a class/type is a lockable type (such as the `Mutex` class).
-#if ABSL_HAVE_ATTRIBUTE(lockable)
-#define ABSL_LOCKABLE __attribute__((lockable))
-#else
-#define ABSL_LOCKABLE
-#endif
-
-// ABSL_SCOPED_LOCKABLE
-//
-// Documents if a class does RAII locking (such as the `MutexLock` class).
-// The constructor should use `LOCK_FUNCTION()` to specify the mutex that is
-// acquired, and the destructor should use `UNLOCK_FUNCTION()` with no
-// arguments; the analysis will assume that the destructor unlocks whatever the
-// constructor locked.
-#if ABSL_HAVE_ATTRIBUTE(scoped_lockable)
-#define ABSL_SCOPED_LOCKABLE __attribute__((scoped_lockable))
-#else
-#define ABSL_SCOPED_LOCKABLE
-#endif
-
-// ABSL_EXCLUSIVE_LOCK_FUNCTION()
-//
-// Documents functions that acquire a lock in the body of a function, and do
-// not release it.
-#if ABSL_HAVE_ATTRIBUTE(exclusive_lock_function)
-#define ABSL_EXCLUSIVE_LOCK_FUNCTION(...) \
-  __attribute__((exclusive_lock_function(__VA_ARGS__)))
-#else
-#define ABSL_EXCLUSIVE_LOCK_FUNCTION(...)
-#endif
-
-// ABSL_SHARED_LOCK_FUNCTION()
-//
-// Documents functions that acquire a shared (reader) lock in the body of a
-// function, and do not release it.
-#if ABSL_HAVE_ATTRIBUTE(shared_lock_function)
-#define ABSL_SHARED_LOCK_FUNCTION(...) \
-  __attribute__((shared_lock_function(__VA_ARGS__)))
-#else
-#define ABSL_SHARED_LOCK_FUNCTION(...)
-#endif
-
-// ABSL_UNLOCK_FUNCTION()
-//
-// Documents functions that expect a lock to be held on entry to the function,
-// and release it in the body of the function.
-#if ABSL_HAVE_ATTRIBUTE(unlock_function)
-#define ABSL_UNLOCK_FUNCTION(...) __attribute__((unlock_function(__VA_ARGS__)))
-#else
-#define ABSL_UNLOCK_FUNCTION(...)
-#endif
-
-// ABSL_EXCLUSIVE_TRYLOCK_FUNCTION() / ABSL_SHARED_TRYLOCK_FUNCTION()
-//
-// Documents functions that try to acquire a lock, and return success or failure
-// (or a non-boolean value that can be interpreted as a boolean).
-// The first argument should be `true` for functions that return `true` on
-// success, or `false` for functions that return `false` on success. The second
-// argument specifies the mutex that is locked on success. If unspecified, this
-// mutex is assumed to be `this`.
-#if ABSL_HAVE_ATTRIBUTE(exclusive_trylock_function)
-#define ABSL_EXCLUSIVE_TRYLOCK_FUNCTION(...) \
-  __attribute__((exclusive_trylock_function(__VA_ARGS__)))
-#else
-#define ABSL_EXCLUSIVE_TRYLOCK_FUNCTION(...)
-#endif
-
-#if ABSL_HAVE_ATTRIBUTE(shared_trylock_function)
-#define ABSL_SHARED_TRYLOCK_FUNCTION(...) \
-  __attribute__((shared_trylock_function(__VA_ARGS__)))
-#else
-#define ABSL_SHARED_TRYLOCK_FUNCTION(...)
-#endif
-
-// ABSL_ASSERT_EXCLUSIVE_LOCK() / ABSL_ASSERT_SHARED_LOCK()
-//
-// Documents functions that dynamically check to see if a lock is held, and fail
-// if it is not held.
-#if ABSL_HAVE_ATTRIBUTE(assert_exclusive_lock)
-#define ABSL_ASSERT_EXCLUSIVE_LOCK(...) \
-  __attribute__((assert_exclusive_lock(__VA_ARGS__)))
-#else
-#define ABSL_ASSERT_EXCLUSIVE_LOCK(...)
-#endif
-
-#if ABSL_HAVE_ATTRIBUTE(assert_shared_lock)
-#define ABSL_ASSERT_SHARED_LOCK(...) \
-  __attribute__((assert_shared_lock(__VA_ARGS__)))
-#else
-#define ABSL_ASSERT_SHARED_LOCK(...)
-#endif
-
-// ABSL_NO_THREAD_SAFETY_ANALYSIS
-//
-// Turns off thread safety checking within the body of a particular function.
-// This annotation is used to mark functions that are known to be correct, but
-// the locking behavior is more complicated than the analyzer can handle.
-#if ABSL_HAVE_ATTRIBUTE(no_thread_safety_analysis)
-#define ABSL_NO_THREAD_SAFETY_ANALYSIS \
-  __attribute__((no_thread_safety_analysis))
-#else
-#define ABSL_NO_THREAD_SAFETY_ANALYSIS
-#endif
-
-//------------------------------------------------------------------------------
-// Tool-Supplied Annotations
-//------------------------------------------------------------------------------
-
-// ABSL_TS_UNCHECKED should be placed around lock expressions that are not valid
-// C++ syntax, but which are present for documentation purposes.  These
-// annotations will be ignored by the analysis.
-#define ABSL_TS_UNCHECKED(x) ""
-
-// ABSL_TS_FIXME is used to mark lock expressions that are not valid C++ syntax.
-// It is used by automated tools to mark and disable invalid expressions.
-// The annotation should either be fixed, or changed to ABSL_TS_UNCHECKED.
-#define ABSL_TS_FIXME(x) ""
-
-// Like ABSL_NO_THREAD_SAFETY_ANALYSIS, this turns off checking within the body
-// of a particular function.  However, this attribute is used to mark functions
-// that are incorrect and need to be fixed.  It is used by automated tools to
-// avoid breaking the build when the analysis is updated.
-// Code owners are expected to eventually fix the routine.
-#define ABSL_NO_THREAD_SAFETY_ANALYSIS_FIXME ABSL_NO_THREAD_SAFETY_ANALYSIS
-
-// Similar to ABSL_NO_THREAD_SAFETY_ANALYSIS_FIXME, this macro marks a
-// ABSL_GUARDED_BY annotation that needs to be fixed, because it is producing
-// thread safety warning. It disables the ABSL_GUARDED_BY.
-#define ABSL_GUARDED_BY_FIXME(x)
-
-// Disables warnings for a single read operation.  This can be used to avoid
-// warnings when it is known that the read is not actually involved in a race,
-// but the compiler cannot confirm that.
-#define ABSL_TS_UNCHECKED_READ(x) absl::base_internal::ts_unchecked_read(x)
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace base_internal {
-
-// Takes a reference to a guarded data member, and returns an unguarded
-// reference.
-// Do not use this function directly, use ABSL_TS_UNCHECKED_READ instead.
-template <typename T>
-inline const T& ts_unchecked_read(const T& v) ABSL_NO_THREAD_SAFETY_ANALYSIS {
-  return v;
-}
-
-template <typename T>
-inline T& ts_unchecked_read(T& v) ABSL_NO_THREAD_SAFETY_ANALYSIS {
-  return v;
-}
-
-}  // namespace base_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_BASE_THREAD_ANNOTATIONS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/cleanup/cleanup.h b/projects/llm_framework/include/abseil-cpp/absl/cleanup/cleanup.h
deleted file mode 100644
index 960ccd08..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/cleanup/cleanup.h
+++ /dev/null
@@ -1,140 +0,0 @@
-// Copyright 2021 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: cleanup.h
-// -----------------------------------------------------------------------------
-//
-// `absl::Cleanup` implements the scope guard idiom, invoking the contained
-// callback's `operator()() &&` on scope exit.
-//
-// Example:
-//
-// ```
-//   absl::Status CopyGoodData(const char* source_path, const char* sink_path) {
-//     FILE* source_file = fopen(source_path, "r");
-//     if (source_file == nullptr) {
-//       return absl::NotFoundError("No source file");  // No cleanups execute
-//     }
-//
-//     // C++17 style cleanup using class template argument deduction
-//     absl::Cleanup source_closer = [source_file] { fclose(source_file); };
-//
-//     FILE* sink_file = fopen(sink_path, "w");
-//     if (sink_file == nullptr) {
-//       return absl::NotFoundError("No sink file");  // First cleanup executes
-//     }
-//
-//     // C++11 style cleanup using the factory function
-//     auto sink_closer = absl::MakeCleanup([sink_file] { fclose(sink_file); });
-//
-//     Data data;
-//     while (ReadData(source_file, &data)) {
-//       if (!data.IsGood()) {
-//         absl::Status result = absl::FailedPreconditionError("Read bad data");
-//         return result;  // Both cleanups execute
-//       }
-//       SaveData(sink_file, &data);
-//     }
-//
-//     return absl::OkStatus();  // Both cleanups execute
-//   }
-// ```
-//
-// Methods:
-//
-// `std::move(cleanup).Cancel()` will prevent the callback from executing.
-//
-// `std::move(cleanup).Invoke()` will execute the callback early, before
-// destruction, and prevent the callback from executing in the destructor.
-//
-// Usage:
-//
-// `absl::Cleanup` is not an interface type. It is only intended to be used
-// within the body of a function. It is not a value type and instead models a
-// control flow construct. Check out `defer` in Golang for something similar.
-
-#ifndef ABSL_CLEANUP_CLEANUP_H_
-#define ABSL_CLEANUP_CLEANUP_H_
-
-#include <utility>
-
-#include "absl/base/config.h"
-#include "absl/base/macros.h"
-#include "absl/cleanup/internal/cleanup.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-template <typename Arg, typename Callback = void()>
-class ABSL_MUST_USE_RESULT Cleanup final {
-  static_assert(cleanup_internal::WasDeduced<Arg>(),
-                "Explicit template parameters are not supported.");
-
-  static_assert(cleanup_internal::ReturnsVoid<Callback>(),
-                "Callbacks that return values are not supported.");
-
- public:
-  Cleanup(Callback callback) : storage_(std::move(callback)) {}  // NOLINT
-
-  Cleanup(Cleanup&& other) = default;
-
-  void Cancel() && {
-    ABSL_HARDENING_ASSERT(storage_.IsCallbackEngaged());
-    storage_.DestroyCallback();
-  }
-
-  void Invoke() && {
-    ABSL_HARDENING_ASSERT(storage_.IsCallbackEngaged());
-    storage_.InvokeCallback();
-    storage_.DestroyCallback();
-  }
-
-  ~Cleanup() {
-    if (storage_.IsCallbackEngaged()) {
-      storage_.InvokeCallback();
-      storage_.DestroyCallback();
-    }
-  }
-
- private:
-  cleanup_internal::Storage<Callback> storage_;
-};
-
-// `absl::Cleanup c = /* callback */;`
-//
-// C++17 type deduction API for creating an instance of `absl::Cleanup`
-#if defined(ABSL_HAVE_CLASS_TEMPLATE_ARGUMENT_DEDUCTION)
-template <typename Callback>
-Cleanup(Callback callback) -> Cleanup<cleanup_internal::Tag, Callback>;
-#endif  // defined(ABSL_HAVE_CLASS_TEMPLATE_ARGUMENT_DEDUCTION)
-
-// `auto c = absl::MakeCleanup(/* callback */);`
-//
-// C++11 type deduction API for creating an instance of `absl::Cleanup`
-template <typename... Args, typename Callback>
-absl::Cleanup<cleanup_internal::Tag, Callback> MakeCleanup(Callback callback) {
-  static_assert(cleanup_internal::WasDeduced<cleanup_internal::Tag, Args...>(),
-                "Explicit template parameters are not supported.");
-
-  static_assert(cleanup_internal::ReturnsVoid<Callback>(),
-                "Callbacks that return values are not supported.");
-
-  return {std::move(callback)};
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CLEANUP_CLEANUP_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/cleanup/internal/cleanup.h b/projects/llm_framework/include/abseil-cpp/absl/cleanup/internal/cleanup.h
deleted file mode 100644
index 2783fcb7..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/cleanup/internal/cleanup.h
+++ /dev/null
@@ -1,100 +0,0 @@
-// Copyright 2021 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CLEANUP_INTERNAL_CLEANUP_H_
-#define ABSL_CLEANUP_INTERNAL_CLEANUP_H_
-
-#include <new>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/internal/invoke.h"
-#include "absl/base/macros.h"
-#include "absl/base/thread_annotations.h"
-#include "absl/utility/utility.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-namespace cleanup_internal {
-
-struct Tag {};
-
-template <typename Arg, typename... Args>
-constexpr bool WasDeduced() {
-  return (std::is_same<cleanup_internal::Tag, Arg>::value) &&
-         (sizeof...(Args) == 0);
-}
-
-template <typename Callback>
-constexpr bool ReturnsVoid() {
-  return (std::is_same<base_internal::invoke_result_t<Callback>, void>::value);
-}
-
-template <typename Callback>
-class Storage {
- public:
-  Storage() = delete;
-
-  explicit Storage(Callback callback) {
-    // Placement-new into a character buffer is used for eager destruction when
-    // the cleanup is invoked or cancelled. To ensure this optimizes well, the
-    // behavior is implemented locally instead of using an absl::optional.
-    ::new (GetCallbackBuffer()) Callback(std::move(callback));
-    is_callback_engaged_ = true;
-  }
-
-  Storage(Storage&& other) {
-    ABSL_HARDENING_ASSERT(other.IsCallbackEngaged());
-
-    ::new (GetCallbackBuffer()) Callback(std::move(other.GetCallback()));
-    is_callback_engaged_ = true;
-
-    other.DestroyCallback();
-  }
-
-  Storage(const Storage& other) = delete;
-
-  Storage& operator=(Storage&& other) = delete;
-
-  Storage& operator=(const Storage& other) = delete;
-
-  void* GetCallbackBuffer() { return static_cast<void*>(+callback_buffer_); }
-
-  Callback& GetCallback() {
-    return *reinterpret_cast<Callback*>(GetCallbackBuffer());
-  }
-
-  bool IsCallbackEngaged() const { return is_callback_engaged_; }
-
-  void DestroyCallback() {
-    is_callback_engaged_ = false;
-    GetCallback().~Callback();
-  }
-
-  void InvokeCallback() ABSL_NO_THREAD_SAFETY_ANALYSIS {
-    std::move(GetCallback())();
-  }
-
- private:
-  bool is_callback_engaged_;
-  alignas(Callback) char callback_buffer_[sizeof(Callback)];
-};
-
-}  // namespace cleanup_internal
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CLEANUP_INTERNAL_CLEANUP_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/btree_map.h b/projects/llm_framework/include/abseil-cpp/absl/container/btree_map.h
deleted file mode 100644
index 0f62f0bd..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/btree_map.h
+++ /dev/null
@@ -1,887 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: btree_map.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines B-tree maps: sorted associative containers mapping
-// keys to values.
-//
-//     * `absl::btree_map<>`
-//     * `absl::btree_multimap<>`
-//
-// These B-tree types are similar to the corresponding types in the STL
-// (`std::map` and `std::multimap`) and generally conform to the STL interfaces
-// of those types. However, because they are implemented using B-trees, they
-// are more efficient in most situations.
-//
-// Unlike `std::map` and `std::multimap`, which are commonly implemented using
-// red-black tree nodes, B-tree maps use more generic B-tree nodes able to hold
-// multiple values per node. Holding multiple values per node often makes
-// B-tree maps perform better than their `std::map` counterparts, because
-// multiple entries can be checked within the same cache hit.
-//
-// However, these types should not be considered drop-in replacements for
-// `std::map` and `std::multimap` as there are some API differences, which are
-// noted in this header file. The most consequential differences with respect to
-// migrating to b-tree from the STL types are listed in the next paragraph.
-// Other API differences are minor.
-//
-// Importantly, insertions and deletions may invalidate outstanding iterators,
-// pointers, and references to elements. Such invalidations are typically only
-// an issue if insertion and deletion operations are interleaved with the use of
-// more than one iterator, pointer, or reference simultaneously.  For this
-// reason, `insert()`, `erase()`, and `extract_and_get_next()` return a valid
-// iterator at the current position. Another important difference is that
-// key-types must be copy-constructible.
-//
-// Another API difference is that btree iterators can be subtracted, and this
-// is faster than using std::distance.
-
-#ifndef ABSL_CONTAINER_BTREE_MAP_H_
-#define ABSL_CONTAINER_BTREE_MAP_H_
-
-#include "absl/base/attributes.h"
-#include "absl/container/internal/btree.h"  // IWYU pragma: export
-#include "absl/container/internal/btree_container.h"  // IWYU pragma: export
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-namespace container_internal {
-
-template <typename Key, typename Data, typename Compare, typename Alloc,
-          int TargetNodeSize, bool IsMulti>
-struct map_params;
-
-}  // namespace container_internal
-
-// absl::btree_map<>
-//
-// An `absl::btree_map<K, V>` is an ordered associative container of
-// unique keys and associated values designed to be a more efficient replacement
-// for `std::map` (in most cases).
-//
-// Keys are sorted using an (optional) comparison function, which defaults to
-// `std::less<K>`.
-//
-// An `absl::btree_map<K, V>` uses a default allocator of
-// `std::allocator<std::pair<const K, V>>` to allocate (and deallocate)
-// nodes, and construct and destruct values within those nodes. You may
-// instead specify a custom allocator `A` (which in turn requires specifying a
-// custom comparator `C`) as in `absl::btree_map<K, V, C, A>`.
-//
-template <typename Key, typename Value, typename Compare = std::less<Key>,
-          typename Alloc = std::allocator<std::pair<const Key, Value>>>
-class btree_map
-    : public container_internal::btree_map_container<
-          container_internal::btree<container_internal::map_params<
-              Key, Value, Compare, Alloc, /*TargetNodeSize=*/256,
-              /*IsMulti=*/false>>> {
-  using Base = typename btree_map::btree_map_container;
-
- public:
-  // Constructors and Assignment Operators
-  //
-  // A `btree_map` supports the same overload set as `std::map`
-  // for construction and assignment:
-  //
-  // * Default constructor
-  //
-  //   absl::btree_map<int, std::string> map1;
-  //
-  // * Initializer List constructor
-  //
-  //   absl::btree_map<int, std::string> map2 =
-  //       {{1, "huey"}, {2, "dewey"}, {3, "louie"},};
-  //
-  // * Copy constructor
-  //
-  //   absl::btree_map<int, std::string> map3(map2);
-  //
-  // * Copy assignment operator
-  //
-  //  absl::btree_map<int, std::string> map4;
-  //  map4 = map3;
-  //
-  // * Move constructor
-  //
-  //   // Move is guaranteed efficient
-  //   absl::btree_map<int, std::string> map5(std::move(map4));
-  //
-  // * Move assignment operator
-  //
-  //   // May be efficient if allocators are compatible
-  //   absl::btree_map<int, std::string> map6;
-  //   map6 = std::move(map5);
-  //
-  // * Range constructor
-  //
-  //   std::vector<std::pair<int, std::string>> v = {{1, "a"}, {2, "b"}};
-  //   absl::btree_map<int, std::string> map7(v.begin(), v.end());
-  btree_map() {}
-  using Base::Base;
-
-  // btree_map::begin()
-  //
-  // Returns an iterator to the beginning of the `btree_map`.
-  using Base::begin;
-
-  // btree_map::cbegin()
-  //
-  // Returns a const iterator to the beginning of the `btree_map`.
-  using Base::cbegin;
-
-  // btree_map::end()
-  //
-  // Returns an iterator to the end of the `btree_map`.
-  using Base::end;
-
-  // btree_map::cend()
-  //
-  // Returns a const iterator to the end of the `btree_map`.
-  using Base::cend;
-
-  // btree_map::empty()
-  //
-  // Returns whether or not the `btree_map` is empty.
-  using Base::empty;
-
-  // btree_map::max_size()
-  //
-  // Returns the largest theoretical possible number of elements within a
-  // `btree_map` under current memory constraints. This value can be thought
-  // of as the largest value of `std::distance(begin(), end())` for a
-  // `btree_map<Key, T>`.
-  using Base::max_size;
-
-  // btree_map::size()
-  //
-  // Returns the number of elements currently within the `btree_map`.
-  using Base::size;
-
-  // btree_map::clear()
-  //
-  // Removes all elements from the `btree_map`. Invalidates any references,
-  // pointers, or iterators referring to contained elements.
-  using Base::clear;
-
-  // btree_map::erase()
-  //
-  // Erases elements within the `btree_map`. If an erase occurs, any references,
-  // pointers, or iterators are invalidated.
-  // Overloads are listed below.
-  //
-  // iterator erase(iterator position):
-  // iterator erase(const_iterator position):
-  //
-  //   Erases the element at `position` of the `btree_map`, returning
-  //   the iterator pointing to the element after the one that was erased
-  //   (or end() if none exists).
-  //
-  // iterator erase(const_iterator first, const_iterator last):
-  //
-  //   Erases the elements in the open interval [`first`, `last`), returning
-  //   the iterator pointing to the element after the interval that was erased
-  //   (or end() if none exists).
-  //
-  // template <typename K> size_type erase(const K& key):
-  //
-  //   Erases the element with the matching key, if it exists, returning the
-  //   number of elements erased (0 or 1).
-  using Base::erase;
-
-  // btree_map::insert()
-  //
-  // Inserts an element of the specified value into the `btree_map`,
-  // returning an iterator pointing to the newly inserted element, provided that
-  // an element with the given key does not already exist. If an insertion
-  // occurs, any references, pointers, or iterators are invalidated.
-  // Overloads are listed below.
-  //
-  // std::pair<iterator,bool> insert(const value_type& value):
-  //
-  //   Inserts a value into the `btree_map`. Returns a pair consisting of an
-  //   iterator to the inserted element (or to the element that prevented the
-  //   insertion) and a bool denoting whether the insertion took place.
-  //
-  // std::pair<iterator,bool> insert(value_type&& value):
-  //
-  //   Inserts a moveable value into the `btree_map`. Returns a pair
-  //   consisting of an iterator to the inserted element (or to the element that
-  //   prevented the insertion) and a bool denoting whether the insertion took
-  //   place.
-  //
-  // iterator insert(const_iterator hint, const value_type& value):
-  // iterator insert(const_iterator hint, value_type&& value):
-  //
-  //   Inserts a value, using the position of `hint` as a non-binding suggestion
-  //   for where to begin the insertion search. Returns an iterator to the
-  //   inserted element, or to the existing element that prevented the
-  //   insertion.
-  //
-  // void insert(InputIterator first, InputIterator last):
-  //
-  //   Inserts a range of values [`first`, `last`).
-  //
-  // void insert(std::initializer_list<init_type> ilist):
-  //
-  //   Inserts the elements within the initializer list `ilist`.
-  using Base::insert;
-
-  // btree_map::insert_or_assign()
-  //
-  // Inserts an element of the specified value into the `btree_map` provided
-  // that a value with the given key does not already exist, or replaces the
-  // corresponding mapped type with the forwarded `obj` argument if a key for
-  // that value already exists, returning an iterator pointing to the newly
-  // inserted element. Overloads are listed below.
-  //
-  // pair<iterator, bool> insert_or_assign(const key_type& k, M&& obj):
-  // pair<iterator, bool> insert_or_assign(key_type&& k, M&& obj):
-  //
-  //   Inserts/Assigns (or moves) the element of the specified key into the
-  //   `btree_map`. If the returned bool is true, insertion took place, and if
-  //   it's false, assignment took place.
-  //
-  // iterator insert_or_assign(const_iterator hint,
-  //                           const key_type& k, M&& obj):
-  // iterator insert_or_assign(const_iterator hint, key_type&& k, M&& obj):
-  //
-  //   Inserts/Assigns (or moves) the element of the specified key into the
-  //   `btree_map` using the position of `hint` as a non-binding suggestion
-  //   for where to begin the insertion search.
-  using Base::insert_or_assign;
-
-  // btree_map::emplace()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `btree_map`, provided that no element with the given key
-  // already exists.
-  //
-  // The element may be constructed even if there already is an element with the
-  // key in the container, in which case the newly constructed element will be
-  // destroyed immediately. Prefer `try_emplace()` unless your key is not
-  // copyable or moveable.
-  //
-  // If an insertion occurs, any references, pointers, or iterators are
-  // invalidated.
-  using Base::emplace;
-
-  // btree_map::emplace_hint()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `btree_map`, using the position of `hint` as a non-binding
-  // suggestion for where to begin the insertion search, and only inserts
-  // provided that no element with the given key already exists.
-  //
-  // The element may be constructed even if there already is an element with the
-  // key in the container, in which case the newly constructed element will be
-  // destroyed immediately. Prefer `try_emplace()` unless your key is not
-  // copyable or moveable.
-  //
-  // If an insertion occurs, any references, pointers, or iterators are
-  // invalidated.
-  using Base::emplace_hint;
-
-  // btree_map::try_emplace()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `btree_map`, provided that no element with the given key
-  // already exists. Unlike `emplace()`, if an element with the given key
-  // already exists, we guarantee that no element is constructed.
-  //
-  // If an insertion occurs, any references, pointers, or iterators are
-  // invalidated.
-  //
-  // Overloads are listed below.
-  //
-  //   std::pair<iterator, bool> try_emplace(const key_type& k, Args&&... args):
-  //   std::pair<iterator, bool> try_emplace(key_type&& k, Args&&... args):
-  //
-  // Inserts (via copy or move) the element of the specified key into the
-  // `btree_map`.
-  //
-  //   iterator try_emplace(const_iterator hint,
-  //                        const key_type& k, Args&&... args):
-  //   iterator try_emplace(const_iterator hint, key_type&& k, Args&&... args):
-  //
-  // Inserts (via copy or move) the element of the specified key into the
-  // `btree_map` using the position of `hint` as a non-binding suggestion
-  // for where to begin the insertion search.
-  using Base::try_emplace;
-
-  // btree_map::extract()
-  //
-  // Extracts the indicated element, erasing it in the process, and returns it
-  // as a C++17-compatible node handle. Any references, pointers, or iterators
-  // are invalidated. Overloads are listed below.
-  //
-  // node_type extract(const_iterator position):
-  //
-  //   Extracts the element at the indicated position and returns a node handle
-  //   owning that extracted data.
-  //
-  // template <typename K> node_type extract(const K& k):
-  //
-  //   Extracts the element with the key matching the passed key value and
-  //   returns a node handle owning that extracted data. If the `btree_map`
-  //   does not contain an element with a matching key, this function returns an
-  //   empty node handle.
-  //
-  // NOTE: when compiled in an earlier version of C++ than C++17,
-  // `node_type::key()` returns a const reference to the key instead of a
-  // mutable reference. We cannot safely return a mutable reference without
-  // std::launder (which is not available before C++17).
-  //
-  // NOTE: In this context, `node_type` refers to the C++17 concept of a
-  // move-only type that owns and provides access to the elements in associative
-  // containers (https://en.cppreference.com/w/cpp/container/node_handle).
-  // It does NOT refer to the data layout of the underlying btree.
-  using Base::extract;
-
-  // btree_map::extract_and_get_next()
-  //
-  // Extracts the indicated element, erasing it in the process, and returns it
-  // as a C++17-compatible node handle along with an iterator to the next
-  // element.
-  //
-  // extract_and_get_next_return_type extract_and_get_next(
-  //     const_iterator position):
-  //
-  //   Extracts the element at the indicated position, returns a struct
-  //   containing a member named `node`: a node handle owning that extracted
-  //   data and a member named `next`: an iterator pointing to the next element
-  //   in the btree.
-  using Base::extract_and_get_next;
-
-  // btree_map::merge()
-  //
-  // Extracts elements from a given `source` btree_map into this
-  // `btree_map`. If the destination `btree_map` already contains an
-  // element with an equivalent key, that element is not extracted.
-  using Base::merge;
-
-  // btree_map::swap(btree_map& other)
-  //
-  // Exchanges the contents of this `btree_map` with those of the `other`
-  // btree_map, avoiding invocation of any move, copy, or swap operations on
-  // individual elements.
-  //
-  // All iterators and references on the `btree_map` remain valid, excepting
-  // for the past-the-end iterator, which is invalidated.
-  using Base::swap;
-
-  // btree_map::at()
-  //
-  // Returns a reference to the mapped value of the element with key equivalent
-  // to the passed key.
-  using Base::at;
-
-  // btree_map::contains()
-  //
-  // template <typename K> bool contains(const K& key) const:
-  //
-  // Determines whether an element comparing equal to the given `key` exists
-  // within the `btree_map`, returning `true` if so or `false` otherwise.
-  //
-  // Supports heterogeneous lookup, provided that the map has a compatible
-  // heterogeneous comparator.
-  using Base::contains;
-
-  // btree_map::count()
-  //
-  // template <typename K> size_type count(const K& key) const:
-  //
-  // Returns the number of elements comparing equal to the given `key` within
-  // the `btree_map`. Note that this function will return either `1` or `0`
-  // since duplicate elements are not allowed within a `btree_map`.
-  //
-  // Supports heterogeneous lookup, provided that the map has a compatible
-  // heterogeneous comparator.
-  using Base::count;
-
-  // btree_map::equal_range()
-  //
-  // Returns a half-open range [first, last), defined by a `std::pair` of two
-  // iterators, containing all elements with the passed key in the `btree_map`.
-  using Base::equal_range;
-
-  // btree_map::find()
-  //
-  // template <typename K> iterator find(const K& key):
-  // template <typename K> const_iterator find(const K& key) const:
-  //
-  // Finds an element with the passed `key` within the `btree_map`.
-  //
-  // Supports heterogeneous lookup, provided that the map has a compatible
-  // heterogeneous comparator.
-  using Base::find;
-
-  // btree_map::lower_bound()
-  //
-  // template <typename K> iterator lower_bound(const K& key):
-  // template <typename K> const_iterator lower_bound(const K& key) const:
-  //
-  // Finds the first element with a key that is not less than `key` within the
-  // `btree_map`.
-  //
-  // Supports heterogeneous lookup, provided that the map has a compatible
-  // heterogeneous comparator.
-  using Base::lower_bound;
-
-  // btree_map::upper_bound()
-  //
-  // template <typename K> iterator upper_bound(const K& key):
-  // template <typename K> const_iterator upper_bound(const K& key) const:
-  //
-  // Finds the first element with a key that is greater than `key` within the
-  // `btree_map`.
-  //
-  // Supports heterogeneous lookup, provided that the map has a compatible
-  // heterogeneous comparator.
-  using Base::upper_bound;
-
-  // btree_map::operator[]()
-  //
-  // Returns a reference to the value mapped to the passed key within the
-  // `btree_map`, performing an `insert()` if the key does not already
-  // exist.
-  //
-  // If an insertion occurs, any references, pointers, or iterators are
-  // invalidated. Otherwise iterators are not affected and references are not
-  // invalidated. Overloads are listed below.
-  //
-  // T& operator[](key_type&& key):
-  // T& operator[](const key_type& key):
-  //
-  //   Inserts a value_type object constructed in-place if the element with the
-  //   given key does not exist.
-  using Base::operator[];
-
-  // btree_map::get_allocator()
-  //
-  // Returns the allocator function associated with this `btree_map`.
-  using Base::get_allocator;
-
-  // btree_map::key_comp();
-  //
-  // Returns the key comparator associated with this `btree_map`.
-  using Base::key_comp;
-
-  // btree_map::value_comp();
-  //
-  // Returns the value comparator associated with this `btree_map`.
-  using Base::value_comp;
-};
-
-// absl::swap(absl::btree_map<>, absl::btree_map<>)
-//
-// Swaps the contents of two `absl::btree_map` containers.
-template <typename K, typename V, typename C, typename A>
-void swap(btree_map<K, V, C, A> &x, btree_map<K, V, C, A> &y) {
-  return x.swap(y);
-}
-
-// absl::erase_if(absl::btree_map<>, Pred)
-//
-// Erases all elements that satisfy the predicate pred from the container.
-// Returns the number of erased elements.
-template <typename K, typename V, typename C, typename A, typename Pred>
-typename btree_map<K, V, C, A>::size_type erase_if(
-    btree_map<K, V, C, A> &map, Pred pred) {
-  return container_internal::btree_access::erase_if(map, std::move(pred));
-}
-
-// absl::btree_multimap
-//
-// An `absl::btree_multimap<K, V>` is an ordered associative container of
-// keys and associated values designed to be a more efficient replacement for
-// `std::multimap` (in most cases). Unlike `absl::btree_map`, a B-tree multimap
-// allows multiple elements with equivalent keys.
-//
-// Keys are sorted using an (optional) comparison function, which defaults to
-// `std::less<K>`.
-//
-// An `absl::btree_multimap<K, V>` uses a default allocator of
-// `std::allocator<std::pair<const K, V>>` to allocate (and deallocate)
-// nodes, and construct and destruct values within those nodes. You may
-// instead specify a custom allocator `A` (which in turn requires specifying a
-// custom comparator `C`) as in `absl::btree_multimap<K, V, C, A>`.
-//
-template <typename Key, typename Value, typename Compare = std::less<Key>,
-          typename Alloc = std::allocator<std::pair<const Key, Value>>>
-class btree_multimap
-    : public container_internal::btree_multimap_container<
-          container_internal::btree<container_internal::map_params<
-              Key, Value, Compare, Alloc, /*TargetNodeSize=*/256,
-              /*IsMulti=*/true>>> {
-  using Base = typename btree_multimap::btree_multimap_container;
-
- public:
-  // Constructors and Assignment Operators
-  //
-  // A `btree_multimap` supports the same overload set as `std::multimap`
-  // for construction and assignment:
-  //
-  // * Default constructor
-  //
-  //   absl::btree_multimap<int, std::string> map1;
-  //
-  // * Initializer List constructor
-  //
-  //   absl::btree_multimap<int, std::string> map2 =
-  //       {{1, "huey"}, {2, "dewey"}, {3, "louie"},};
-  //
-  // * Copy constructor
-  //
-  //   absl::btree_multimap<int, std::string> map3(map2);
-  //
-  // * Copy assignment operator
-  //
-  //  absl::btree_multimap<int, std::string> map4;
-  //  map4 = map3;
-  //
-  // * Move constructor
-  //
-  //   // Move is guaranteed efficient
-  //   absl::btree_multimap<int, std::string> map5(std::move(map4));
-  //
-  // * Move assignment operator
-  //
-  //   // May be efficient if allocators are compatible
-  //   absl::btree_multimap<int, std::string> map6;
-  //   map6 = std::move(map5);
-  //
-  // * Range constructor
-  //
-  //   std::vector<std::pair<int, std::string>> v = {{1, "a"}, {2, "b"}};
-  //   absl::btree_multimap<int, std::string> map7(v.begin(), v.end());
-  btree_multimap() {}
-  using Base::Base;
-
-  // btree_multimap::begin()
-  //
-  // Returns an iterator to the beginning of the `btree_multimap`.
-  using Base::begin;
-
-  // btree_multimap::cbegin()
-  //
-  // Returns a const iterator to the beginning of the `btree_multimap`.
-  using Base::cbegin;
-
-  // btree_multimap::end()
-  //
-  // Returns an iterator to the end of the `btree_multimap`.
-  using Base::end;
-
-  // btree_multimap::cend()
-  //
-  // Returns a const iterator to the end of the `btree_multimap`.
-  using Base::cend;
-
-  // btree_multimap::empty()
-  //
-  // Returns whether or not the `btree_multimap` is empty.
-  using Base::empty;
-
-  // btree_multimap::max_size()
-  //
-  // Returns the largest theoretical possible number of elements within a
-  // `btree_multimap` under current memory constraints. This value can be
-  // thought of as the largest value of `std::distance(begin(), end())` for a
-  // `btree_multimap<Key, T>`.
-  using Base::max_size;
-
-  // btree_multimap::size()
-  //
-  // Returns the number of elements currently within the `btree_multimap`.
-  using Base::size;
-
-  // btree_multimap::clear()
-  //
-  // Removes all elements from the `btree_multimap`. Invalidates any references,
-  // pointers, or iterators referring to contained elements.
-  using Base::clear;
-
-  // btree_multimap::erase()
-  //
-  // Erases elements within the `btree_multimap`. If an erase occurs, any
-  // references, pointers, or iterators are invalidated.
-  // Overloads are listed below.
-  //
-  // iterator erase(iterator position):
-  // iterator erase(const_iterator position):
-  //
-  //   Erases the element at `position` of the `btree_multimap`, returning
-  //   the iterator pointing to the element after the one that was erased
-  //   (or end() if none exists).
-  //
-  // iterator erase(const_iterator first, const_iterator last):
-  //
-  //   Erases the elements in the open interval [`first`, `last`), returning
-  //   the iterator pointing to the element after the interval that was erased
-  //   (or end() if none exists).
-  //
-  // template <typename K> size_type erase(const K& key):
-  //
-  //   Erases the elements matching the key, if any exist, returning the
-  //   number of elements erased.
-  using Base::erase;
-
-  // btree_multimap::insert()
-  //
-  // Inserts an element of the specified value into the `btree_multimap`,
-  // returning an iterator pointing to the newly inserted element.
-  // Any references, pointers, or iterators are invalidated.  Overloads are
-  // listed below.
-  //
-  // iterator insert(const value_type& value):
-  //
-  //   Inserts a value into the `btree_multimap`, returning an iterator to the
-  //   inserted element.
-  //
-  // iterator insert(value_type&& value):
-  //
-  //   Inserts a moveable value into the `btree_multimap`, returning an iterator
-  //   to the inserted element.
-  //
-  // iterator insert(const_iterator hint, const value_type& value):
-  // iterator insert(const_iterator hint, value_type&& value):
-  //
-  //   Inserts a value, using the position of `hint` as a non-binding suggestion
-  //   for where to begin the insertion search. Returns an iterator to the
-  //   inserted element.
-  //
-  // void insert(InputIterator first, InputIterator last):
-  //
-  //   Inserts a range of values [`first`, `last`).
-  //
-  // void insert(std::initializer_list<init_type> ilist):
-  //
-  //   Inserts the elements within the initializer list `ilist`.
-  using Base::insert;
-
-  // btree_multimap::emplace()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `btree_multimap`. Any references, pointers, or iterators are
-  // invalidated.
-  using Base::emplace;
-
-  // btree_multimap::emplace_hint()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `btree_multimap`, using the position of `hint` as a non-binding
-  // suggestion for where to begin the insertion search.
-  //
-  // Any references, pointers, or iterators are invalidated.
-  using Base::emplace_hint;
-
-  // btree_multimap::extract()
-  //
-  // Extracts the indicated element, erasing it in the process, and returns it
-  // as a C++17-compatible node handle. Overloads are listed below.
-  //
-  // node_type extract(const_iterator position):
-  //
-  //   Extracts the element at the indicated position and returns a node handle
-  //   owning that extracted data.
-  //
-  // template <typename K> node_type extract(const K& k):
-  //
-  //   Extracts the element with the key matching the passed key value and
-  //   returns a node handle owning that extracted data. If the `btree_multimap`
-  //   does not contain an element with a matching key, this function returns an
-  //   empty node handle.
-  //
-  // NOTE: when compiled in an earlier version of C++ than C++17,
-  // `node_type::key()` returns a const reference to the key instead of a
-  // mutable reference. We cannot safely return a mutable reference without
-  // std::launder (which is not available before C++17).
-  //
-  // NOTE: In this context, `node_type` refers to the C++17 concept of a
-  // move-only type that owns and provides access to the elements in associative
-  // containers (https://en.cppreference.com/w/cpp/container/node_handle).
-  // It does NOT refer to the data layout of the underlying btree.
-  using Base::extract;
-
-  // btree_multimap::extract_and_get_next()
-  //
-  // Extracts the indicated element, erasing it in the process, and returns it
-  // as a C++17-compatible node handle along with an iterator to the next
-  // element.
-  //
-  // extract_and_get_next_return_type extract_and_get_next(
-  //     const_iterator position):
-  //
-  //   Extracts the element at the indicated position, returns a struct
-  //   containing a member named `node`: a node handle owning that extracted
-  //   data and a member named `next`: an iterator pointing to the next element
-  //   in the btree.
-  using Base::extract_and_get_next;
-
-  // btree_multimap::merge()
-  //
-  // Extracts all elements from a given `source` btree_multimap into this
-  // `btree_multimap`.
-  using Base::merge;
-
-  // btree_multimap::swap(btree_multimap& other)
-  //
-  // Exchanges the contents of this `btree_multimap` with those of the `other`
-  // btree_multimap, avoiding invocation of any move, copy, or swap operations
-  // on individual elements.
-  //
-  // All iterators and references on the `btree_multimap` remain valid,
-  // excepting for the past-the-end iterator, which is invalidated.
-  using Base::swap;
-
-  // btree_multimap::contains()
-  //
-  // template <typename K> bool contains(const K& key) const:
-  //
-  // Determines whether an element comparing equal to the given `key` exists
-  // within the `btree_multimap`, returning `true` if so or `false` otherwise.
-  //
-  // Supports heterogeneous lookup, provided that the map has a compatible
-  // heterogeneous comparator.
-  using Base::contains;
-
-  // btree_multimap::count()
-  //
-  // template <typename K> size_type count(const K& key) const:
-  //
-  // Returns the number of elements comparing equal to the given `key` within
-  // the `btree_multimap`.
-  //
-  // Supports heterogeneous lookup, provided that the map has a compatible
-  // heterogeneous comparator.
-  using Base::count;
-
-  // btree_multimap::equal_range()
-  //
-  // Returns a half-open range [first, last), defined by a `std::pair` of two
-  // iterators, containing all elements with the passed key in the
-  // `btree_multimap`.
-  using Base::equal_range;
-
-  // btree_multimap::find()
-  //
-  // template <typename K> iterator find(const K& key):
-  // template <typename K> const_iterator find(const K& key) const:
-  //
-  // Finds an element with the passed `key` within the `btree_multimap`.
-  //
-  // Supports heterogeneous lookup, provided that the map has a compatible
-  // heterogeneous comparator.
-  using Base::find;
-
-  // btree_multimap::lower_bound()
-  //
-  // template <typename K> iterator lower_bound(const K& key):
-  // template <typename K> const_iterator lower_bound(const K& key) const:
-  //
-  // Finds the first element with a key that is not less than `key` within the
-  // `btree_multimap`.
-  //
-  // Supports heterogeneous lookup, provided that the map has a compatible
-  // heterogeneous comparator.
-  using Base::lower_bound;
-
-  // btree_multimap::upper_bound()
-  //
-  // template <typename K> iterator upper_bound(const K& key):
-  // template <typename K> const_iterator upper_bound(const K& key) const:
-  //
-  // Finds the first element with a key that is greater than `key` within the
-  // `btree_multimap`.
-  //
-  // Supports heterogeneous lookup, provided that the map has a compatible
-  // heterogeneous comparator.
-  using Base::upper_bound;
-
-  // btree_multimap::get_allocator()
-  //
-  // Returns the allocator function associated with this `btree_multimap`.
-  using Base::get_allocator;
-
-  // btree_multimap::key_comp();
-  //
-  // Returns the key comparator associated with this `btree_multimap`.
-  using Base::key_comp;
-
-  // btree_multimap::value_comp();
-  //
-  // Returns the value comparator associated with this `btree_multimap`.
-  using Base::value_comp;
-};
-
-// absl::swap(absl::btree_multimap<>, absl::btree_multimap<>)
-//
-// Swaps the contents of two `absl::btree_multimap` containers.
-template <typename K, typename V, typename C, typename A>
-void swap(btree_multimap<K, V, C, A> &x, btree_multimap<K, V, C, A> &y) {
-  return x.swap(y);
-}
-
-// absl::erase_if(absl::btree_multimap<>, Pred)
-//
-// Erases all elements that satisfy the predicate pred from the container.
-// Returns the number of erased elements.
-template <typename K, typename V, typename C, typename A, typename Pred>
-typename btree_multimap<K, V, C, A>::size_type erase_if(
-    btree_multimap<K, V, C, A> &map, Pred pred) {
-  return container_internal::btree_access::erase_if(map, std::move(pred));
-}
-
-namespace container_internal {
-
-// A parameters structure for holding the type parameters for a btree_map.
-// Compare and Alloc should be nothrow copy-constructible.
-template <typename Key, typename Data, typename Compare, typename Alloc,
-          int TargetNodeSize, bool IsMulti>
-struct map_params : common_params<Key, Compare, Alloc, TargetNodeSize, IsMulti,
-                                  /*IsMap=*/true, map_slot_policy<Key, Data>> {
-  using super_type = typename map_params::common_params;
-  using mapped_type = Data;
-  // This type allows us to move keys when it is safe to do so. It is safe
-  // for maps in which value_type and mutable_value_type are layout compatible.
-  using slot_policy = typename super_type::slot_policy;
-  using slot_type = typename super_type::slot_type;
-  using value_type = typename super_type::value_type;
-  using init_type = typename super_type::init_type;
-
-  template <typename V>
-  static auto key(const V &value ABSL_ATTRIBUTE_LIFETIME_BOUND)
-      -> decltype((value.first)) {
-    return value.first;
-  }
-  static const Key &key(const slot_type *s) { return slot_policy::key(s); }
-  static const Key &key(slot_type *s) { return slot_policy::key(s); }
-  // For use in node handle.
-  static auto mutable_key(slot_type *s)
-      -> decltype(slot_policy::mutable_key(s)) {
-    return slot_policy::mutable_key(s);
-  }
-  static mapped_type &value(value_type *value) { return value->second; }
-};
-
-}  // namespace container_internal
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_BTREE_MAP_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/btree_set.h b/projects/llm_framework/include/abseil-cpp/absl/container/btree_set.h
deleted file mode 100644
index 51dc42b7..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/btree_set.h
+++ /dev/null
@@ -1,821 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: btree_set.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines B-tree sets: sorted associative containers of
-// values.
-//
-//     * `absl::btree_set<>`
-//     * `absl::btree_multiset<>`
-//
-// These B-tree types are similar to the corresponding types in the STL
-// (`std::set` and `std::multiset`) and generally conform to the STL interfaces
-// of those types. However, because they are implemented using B-trees, they
-// are more efficient in most situations.
-//
-// Unlike `std::set` and `std::multiset`, which are commonly implemented using
-// red-black tree nodes, B-tree sets use more generic B-tree nodes able to hold
-// multiple values per node. Holding multiple values per node often makes
-// B-tree sets perform better than their `std::set` counterparts, because
-// multiple entries can be checked within the same cache hit.
-//
-// However, these types should not be considered drop-in replacements for
-// `std::set` and `std::multiset` as there are some API differences, which are
-// noted in this header file. The most consequential differences with respect to
-// migrating to b-tree from the STL types are listed in the next paragraph.
-// Other API differences are minor.
-//
-// Importantly, insertions and deletions may invalidate outstanding iterators,
-// pointers, and references to elements. Such invalidations are typically only
-// an issue if insertion and deletion operations are interleaved with the use of
-// more than one iterator, pointer, or reference simultaneously. For this
-// reason, `insert()`, `erase()`, and `extract_and_get_next()` return a valid
-// iterator at the current position.
-//
-// Another API difference is that btree iterators can be subtracted, and this
-// is faster than using std::distance.
-
-#ifndef ABSL_CONTAINER_BTREE_SET_H_
-#define ABSL_CONTAINER_BTREE_SET_H_
-
-#include "absl/container/internal/btree.h"  // IWYU pragma: export
-#include "absl/container/internal/btree_container.h"  // IWYU pragma: export
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-namespace container_internal {
-
-template <typename Key>
-struct set_slot_policy;
-
-template <typename Key, typename Compare, typename Alloc, int TargetNodeSize,
-          bool IsMulti>
-struct set_params;
-
-}  // namespace container_internal
-
-// absl::btree_set<>
-//
-// An `absl::btree_set<K>` is an ordered associative container of unique key
-// values designed to be a more efficient replacement for `std::set` (in most
-// cases).
-//
-// Keys are sorted using an (optional) comparison function, which defaults to
-// `std::less<K>`.
-//
-// An `absl::btree_set<K>` uses a default allocator of `std::allocator<K>` to
-// allocate (and deallocate) nodes, and construct and destruct values within
-// those nodes. You may instead specify a custom allocator `A` (which in turn
-// requires specifying a custom comparator `C`) as in
-// `absl::btree_set<K, C, A>`.
-//
-template <typename Key, typename Compare = std::less<Key>,
-          typename Alloc = std::allocator<Key>>
-class btree_set
-    : public container_internal::btree_set_container<
-          container_internal::btree<container_internal::set_params<
-              Key, Compare, Alloc, /*TargetNodeSize=*/256,
-              /*IsMulti=*/false>>> {
-  using Base = typename btree_set::btree_set_container;
-
- public:
-  // Constructors and Assignment Operators
-  //
-  // A `btree_set` supports the same overload set as `std::set`
-  // for construction and assignment:
-  //
-  // * Default constructor
-  //
-  //   absl::btree_set<std::string> set1;
-  //
-  // * Initializer List constructor
-  //
-  //   absl::btree_set<std::string> set2 =
-  //       {{"huey"}, {"dewey"}, {"louie"},};
-  //
-  // * Copy constructor
-  //
-  //   absl::btree_set<std::string> set3(set2);
-  //
-  // * Copy assignment operator
-  //
-  //  absl::btree_set<std::string> set4;
-  //  set4 = set3;
-  //
-  // * Move constructor
-  //
-  //   // Move is guaranteed efficient
-  //   absl::btree_set<std::string> set5(std::move(set4));
-  //
-  // * Move assignment operator
-  //
-  //   // May be efficient if allocators are compatible
-  //   absl::btree_set<std::string> set6;
-  //   set6 = std::move(set5);
-  //
-  // * Range constructor
-  //
-  //   std::vector<std::string> v = {"a", "b"};
-  //   absl::btree_set<std::string> set7(v.begin(), v.end());
-  btree_set() {}
-  using Base::Base;
-
-  // btree_set::begin()
-  //
-  // Returns an iterator to the beginning of the `btree_set`.
-  using Base::begin;
-
-  // btree_set::cbegin()
-  //
-  // Returns a const iterator to the beginning of the `btree_set`.
-  using Base::cbegin;
-
-  // btree_set::end()
-  //
-  // Returns an iterator to the end of the `btree_set`.
-  using Base::end;
-
-  // btree_set::cend()
-  //
-  // Returns a const iterator to the end of the `btree_set`.
-  using Base::cend;
-
-  // btree_set::empty()
-  //
-  // Returns whether or not the `btree_set` is empty.
-  using Base::empty;
-
-  // btree_set::max_size()
-  //
-  // Returns the largest theoretical possible number of elements within a
-  // `btree_set` under current memory constraints. This value can be thought
-  // of as the largest value of `std::distance(begin(), end())` for a
-  // `btree_set<Key>`.
-  using Base::max_size;
-
-  // btree_set::size()
-  //
-  // Returns the number of elements currently within the `btree_set`.
-  using Base::size;
-
-  // btree_set::clear()
-  //
-  // Removes all elements from the `btree_set`. Invalidates any references,
-  // pointers, or iterators referring to contained elements.
-  using Base::clear;
-
-  // btree_set::erase()
-  //
-  // Erases elements within the `btree_set`. Overloads are listed below.
-  //
-  // iterator erase(iterator position):
-  // iterator erase(const_iterator position):
-  //
-  //   Erases the element at `position` of the `btree_set`, returning
-  //   the iterator pointing to the element after the one that was erased
-  //   (or end() if none exists).
-  //
-  // iterator erase(const_iterator first, const_iterator last):
-  //
-  //   Erases the elements in the open interval [`first`, `last`), returning
-  //   the iterator pointing to the element after the interval that was erased
-  //   (or end() if none exists).
-  //
-  // template <typename K> size_type erase(const K& key):
-  //
-  //   Erases the element with the matching key, if it exists, returning the
-  //   number of elements erased (0 or 1).
-  using Base::erase;
-
-  // btree_set::insert()
-  //
-  // Inserts an element of the specified value into the `btree_set`,
-  // returning an iterator pointing to the newly inserted element, provided that
-  // an element with the given key does not already exist. If an insertion
-  // occurs, any references, pointers, or iterators are invalidated.
-  // Overloads are listed below.
-  //
-  // std::pair<iterator,bool> insert(const value_type& value):
-  //
-  //   Inserts a value into the `btree_set`. Returns a pair consisting of an
-  //   iterator to the inserted element (or to the element that prevented the
-  //   insertion) and a bool denoting whether the insertion took place.
-  //
-  // std::pair<iterator,bool> insert(value_type&& value):
-  //
-  //   Inserts a moveable value into the `btree_set`. Returns a pair
-  //   consisting of an iterator to the inserted element (or to the element that
-  //   prevented the insertion) and a bool denoting whether the insertion took
-  //   place.
-  //
-  // iterator insert(const_iterator hint, const value_type& value):
-  // iterator insert(const_iterator hint, value_type&& value):
-  //
-  //   Inserts a value, using the position of `hint` as a non-binding suggestion
-  //   for where to begin the insertion search. Returns an iterator to the
-  //   inserted element, or to the existing element that prevented the
-  //   insertion.
-  //
-  // void insert(InputIterator first, InputIterator last):
-  //
-  //   Inserts a range of values [`first`, `last`).
-  //
-  // void insert(std::initializer_list<init_type> ilist):
-  //
-  //   Inserts the elements within the initializer list `ilist`.
-  using Base::insert;
-
-  // btree_set::emplace()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `btree_set`, provided that no element with the given key
-  // already exists.
-  //
-  // The element may be constructed even if there already is an element with the
-  // key in the container, in which case the newly constructed element will be
-  // destroyed immediately.
-  //
-  // If an insertion occurs, any references, pointers, or iterators are
-  // invalidated.
-  using Base::emplace;
-
-  // btree_set::emplace_hint()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `btree_set`, using the position of `hint` as a non-binding
-  // suggestion for where to begin the insertion search, and only inserts
-  // provided that no element with the given key already exists.
-  //
-  // The element may be constructed even if there already is an element with the
-  // key in the container, in which case the newly constructed element will be
-  // destroyed immediately.
-  //
-  // If an insertion occurs, any references, pointers, or iterators are
-  // invalidated.
-  using Base::emplace_hint;
-
-  // btree_set::extract()
-  //
-  // Extracts the indicated element, erasing it in the process, and returns it
-  // as a C++17-compatible node handle. Any references, pointers, or iterators
-  // are invalidated. Overloads are listed below.
-  //
-  // node_type extract(const_iterator position):
-  //
-  //   Extracts the element at the indicated position and returns a node handle
-  //   owning that extracted data.
-  //
-  // template <typename K> node_type extract(const K& k):
-  //
-  //   Extracts the element with the key matching the passed key value and
-  //   returns a node handle owning that extracted data. If the `btree_set`
-  //   does not contain an element with a matching key, this function returns an
-  //   empty node handle.
-  //
-  // NOTE: In this context, `node_type` refers to the C++17 concept of a
-  // move-only type that owns and provides access to the elements in associative
-  // containers (https://en.cppreference.com/w/cpp/container/node_handle).
-  // It does NOT refer to the data layout of the underlying btree.
-  using Base::extract;
-
-  // btree_set::extract_and_get_next()
-  //
-  // Extracts the indicated element, erasing it in the process, and returns it
-  // as a C++17-compatible node handle along with an iterator to the next
-  // element.
-  //
-  // extract_and_get_next_return_type extract_and_get_next(
-  //     const_iterator position):
-  //
-  //   Extracts the element at the indicated position, returns a struct
-  //   containing a member named `node`: a node handle owning that extracted
-  //   data and a member named `next`: an iterator pointing to the next element
-  //   in the btree.
-  using Base::extract_and_get_next;
-
-  // btree_set::merge()
-  //
-  // Extracts elements from a given `source` btree_set into this
-  // `btree_set`. If the destination `btree_set` already contains an
-  // element with an equivalent key, that element is not extracted.
-  using Base::merge;
-
-  // btree_set::swap(btree_set& other)
-  //
-  // Exchanges the contents of this `btree_set` with those of the `other`
-  // btree_set, avoiding invocation of any move, copy, or swap operations on
-  // individual elements.
-  //
-  // All iterators and references on the `btree_set` remain valid, excepting
-  // for the past-the-end iterator, which is invalidated.
-  using Base::swap;
-
-  // btree_set::contains()
-  //
-  // template <typename K> bool contains(const K& key) const:
-  //
-  // Determines whether an element comparing equal to the given `key` exists
-  // within the `btree_set`, returning `true` if so or `false` otherwise.
-  //
-  // Supports heterogeneous lookup, provided that the set has a compatible
-  // heterogeneous comparator.
-  using Base::contains;
-
-  // btree_set::count()
-  //
-  // template <typename K> size_type count(const K& key) const:
-  //
-  // Returns the number of elements comparing equal to the given `key` within
-  // the `btree_set`. Note that this function will return either `1` or `0`
-  // since duplicate elements are not allowed within a `btree_set`.
-  //
-  // Supports heterogeneous lookup, provided that the set has a compatible
-  // heterogeneous comparator.
-  using Base::count;
-
-  // btree_set::equal_range()
-  //
-  // Returns a closed range [first, last], defined by a `std::pair` of two
-  // iterators, containing all elements with the passed key in the
-  // `btree_set`.
-  using Base::equal_range;
-
-  // btree_set::find()
-  //
-  // template <typename K> iterator find(const K& key):
-  // template <typename K> const_iterator find(const K& key) const:
-  //
-  // Finds an element with the passed `key` within the `btree_set`.
-  //
-  // Supports heterogeneous lookup, provided that the set has a compatible
-  // heterogeneous comparator.
-  using Base::find;
-
-  // btree_set::lower_bound()
-  //
-  // template <typename K> iterator lower_bound(const K& key):
-  // template <typename K> const_iterator lower_bound(const K& key) const:
-  //
-  // Finds the first element that is not less than `key` within the `btree_set`.
-  //
-  // Supports heterogeneous lookup, provided that the set has a compatible
-  // heterogeneous comparator.
-  using Base::lower_bound;
-
-  // btree_set::upper_bound()
-  //
-  // template <typename K> iterator upper_bound(const K& key):
-  // template <typename K> const_iterator upper_bound(const K& key) const:
-  //
-  // Finds the first element that is greater than `key` within the `btree_set`.
-  //
-  // Supports heterogeneous lookup, provided that the set has a compatible
-  // heterogeneous comparator.
-  using Base::upper_bound;
-
-  // btree_set::get_allocator()
-  //
-  // Returns the allocator function associated with this `btree_set`.
-  using Base::get_allocator;
-
-  // btree_set::key_comp();
-  //
-  // Returns the key comparator associated with this `btree_set`.
-  using Base::key_comp;
-
-  // btree_set::value_comp();
-  //
-  // Returns the value comparator associated with this `btree_set`. The keys to
-  // sort the elements are the values themselves, therefore `value_comp` and its
-  // sibling member function `key_comp` are equivalent.
-  using Base::value_comp;
-};
-
-// absl::swap(absl::btree_set<>, absl::btree_set<>)
-//
-// Swaps the contents of two `absl::btree_set` containers.
-template <typename K, typename C, typename A>
-void swap(btree_set<K, C, A> &x, btree_set<K, C, A> &y) {
-  return x.swap(y);
-}
-
-// absl::erase_if(absl::btree_set<>, Pred)
-//
-// Erases all elements that satisfy the predicate pred from the container.
-// Returns the number of erased elements.
-template <typename K, typename C, typename A, typename Pred>
-typename btree_set<K, C, A>::size_type erase_if(btree_set<K, C, A> &set,
-                                                Pred pred) {
-  return container_internal::btree_access::erase_if(set, std::move(pred));
-}
-
-// absl::btree_multiset<>
-//
-// An `absl::btree_multiset<K>` is an ordered associative container of
-// keys and associated values designed to be a more efficient replacement
-// for `std::multiset` (in most cases). Unlike `absl::btree_set`, a B-tree
-// multiset allows equivalent elements.
-//
-// Keys are sorted using an (optional) comparison function, which defaults to
-// `std::less<K>`.
-//
-// An `absl::btree_multiset<K>` uses a default allocator of `std::allocator<K>`
-// to allocate (and deallocate) nodes, and construct and destruct values within
-// those nodes. You may instead specify a custom allocator `A` (which in turn
-// requires specifying a custom comparator `C`) as in
-// `absl::btree_multiset<K, C, A>`.
-//
-template <typename Key, typename Compare = std::less<Key>,
-          typename Alloc = std::allocator<Key>>
-class btree_multiset
-    : public container_internal::btree_multiset_container<
-          container_internal::btree<container_internal::set_params<
-              Key, Compare, Alloc, /*TargetNodeSize=*/256,
-              /*IsMulti=*/true>>> {
-  using Base = typename btree_multiset::btree_multiset_container;
-
- public:
-  // Constructors and Assignment Operators
-  //
-  // A `btree_multiset` supports the same overload set as `std::set`
-  // for construction and assignment:
-  //
-  // * Default constructor
-  //
-  //   absl::btree_multiset<std::string> set1;
-  //
-  // * Initializer List constructor
-  //
-  //   absl::btree_multiset<std::string> set2 =
-  //       {{"huey"}, {"dewey"}, {"louie"},};
-  //
-  // * Copy constructor
-  //
-  //   absl::btree_multiset<std::string> set3(set2);
-  //
-  // * Copy assignment operator
-  //
-  //  absl::btree_multiset<std::string> set4;
-  //  set4 = set3;
-  //
-  // * Move constructor
-  //
-  //   // Move is guaranteed efficient
-  //   absl::btree_multiset<std::string> set5(std::move(set4));
-  //
-  // * Move assignment operator
-  //
-  //   // May be efficient if allocators are compatible
-  //   absl::btree_multiset<std::string> set6;
-  //   set6 = std::move(set5);
-  //
-  // * Range constructor
-  //
-  //   std::vector<std::string> v = {"a", "b"};
-  //   absl::btree_multiset<std::string> set7(v.begin(), v.end());
-  btree_multiset() {}
-  using Base::Base;
-
-  // btree_multiset::begin()
-  //
-  // Returns an iterator to the beginning of the `btree_multiset`.
-  using Base::begin;
-
-  // btree_multiset::cbegin()
-  //
-  // Returns a const iterator to the beginning of the `btree_multiset`.
-  using Base::cbegin;
-
-  // btree_multiset::end()
-  //
-  // Returns an iterator to the end of the `btree_multiset`.
-  using Base::end;
-
-  // btree_multiset::cend()
-  //
-  // Returns a const iterator to the end of the `btree_multiset`.
-  using Base::cend;
-
-  // btree_multiset::empty()
-  //
-  // Returns whether or not the `btree_multiset` is empty.
-  using Base::empty;
-
-  // btree_multiset::max_size()
-  //
-  // Returns the largest theoretical possible number of elements within a
-  // `btree_multiset` under current memory constraints. This value can be
-  // thought of as the largest value of `std::distance(begin(), end())` for a
-  // `btree_multiset<Key>`.
-  using Base::max_size;
-
-  // btree_multiset::size()
-  //
-  // Returns the number of elements currently within the `btree_multiset`.
-  using Base::size;
-
-  // btree_multiset::clear()
-  //
-  // Removes all elements from the `btree_multiset`. Invalidates any references,
-  // pointers, or iterators referring to contained elements.
-  using Base::clear;
-
-  // btree_multiset::erase()
-  //
-  // Erases elements within the `btree_multiset`. Overloads are listed below.
-  //
-  // iterator erase(iterator position):
-  // iterator erase(const_iterator position):
-  //
-  //   Erases the element at `position` of the `btree_multiset`, returning
-  //   the iterator pointing to the element after the one that was erased
-  //   (or end() if none exists).
-  //
-  // iterator erase(const_iterator first, const_iterator last):
-  //
-  //   Erases the elements in the open interval [`first`, `last`), returning
-  //   the iterator pointing to the element after the interval that was erased
-  //   (or end() if none exists).
-  //
-  // template <typename K> size_type erase(const K& key):
-  //
-  //   Erases the elements matching the key, if any exist, returning the
-  //   number of elements erased.
-  using Base::erase;
-
-  // btree_multiset::insert()
-  //
-  // Inserts an element of the specified value into the `btree_multiset`,
-  // returning an iterator pointing to the newly inserted element.
-  // Any references, pointers, or iterators are invalidated.  Overloads are
-  // listed below.
-  //
-  // iterator insert(const value_type& value):
-  //
-  //   Inserts a value into the `btree_multiset`, returning an iterator to the
-  //   inserted element.
-  //
-  // iterator insert(value_type&& value):
-  //
-  //   Inserts a moveable value into the `btree_multiset`, returning an iterator
-  //   to the inserted element.
-  //
-  // iterator insert(const_iterator hint, const value_type& value):
-  // iterator insert(const_iterator hint, value_type&& value):
-  //
-  //   Inserts a value, using the position of `hint` as a non-binding suggestion
-  //   for where to begin the insertion search. Returns an iterator to the
-  //   inserted element.
-  //
-  // void insert(InputIterator first, InputIterator last):
-  //
-  //   Inserts a range of values [`first`, `last`).
-  //
-  // void insert(std::initializer_list<init_type> ilist):
-  //
-  //   Inserts the elements within the initializer list `ilist`.
-  using Base::insert;
-
-  // btree_multiset::emplace()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `btree_multiset`. Any references, pointers, or iterators are
-  // invalidated.
-  using Base::emplace;
-
-  // btree_multiset::emplace_hint()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `btree_multiset`, using the position of `hint` as a non-binding
-  // suggestion for where to begin the insertion search.
-  //
-  // Any references, pointers, or iterators are invalidated.
-  using Base::emplace_hint;
-
-  // btree_multiset::extract()
-  //
-  // Extracts the indicated element, erasing it in the process, and returns it
-  // as a C++17-compatible node handle. Overloads are listed below.
-  //
-  // node_type extract(const_iterator position):
-  //
-  //   Extracts the element at the indicated position and returns a node handle
-  //   owning that extracted data.
-  //
-  // template <typename K> node_type extract(const K& k):
-  //
-  //   Extracts the element with the key matching the passed key value and
-  //   returns a node handle owning that extracted data. If the `btree_multiset`
-  //   does not contain an element with a matching key, this function returns an
-  //   empty node handle.
-  //
-  // NOTE: In this context, `node_type` refers to the C++17 concept of a
-  // move-only type that owns and provides access to the elements in associative
-  // containers (https://en.cppreference.com/w/cpp/container/node_handle).
-  // It does NOT refer to the data layout of the underlying btree.
-  using Base::extract;
-
-  // btree_multiset::extract_and_get_next()
-  //
-  // Extracts the indicated element, erasing it in the process, and returns it
-  // as a C++17-compatible node handle along with an iterator to the next
-  // element.
-  //
-  // extract_and_get_next_return_type extract_and_get_next(
-  //     const_iterator position):
-  //
-  //   Extracts the element at the indicated position, returns a struct
-  //   containing a member named `node`: a node handle owning that extracted
-  //   data and a member named `next`: an iterator pointing to the next element
-  //   in the btree.
-  using Base::extract_and_get_next;
-
-  // btree_multiset::merge()
-  //
-  // Extracts all elements from a given `source` btree_multiset into this
-  // `btree_multiset`.
-  using Base::merge;
-
-  // btree_multiset::swap(btree_multiset& other)
-  //
-  // Exchanges the contents of this `btree_multiset` with those of the `other`
-  // btree_multiset, avoiding invocation of any move, copy, or swap operations
-  // on individual elements.
-  //
-  // All iterators and references on the `btree_multiset` remain valid,
-  // excepting for the past-the-end iterator, which is invalidated.
-  using Base::swap;
-
-  // btree_multiset::contains()
-  //
-  // template <typename K> bool contains(const K& key) const:
-  //
-  // Determines whether an element comparing equal to the given `key` exists
-  // within the `btree_multiset`, returning `true` if so or `false` otherwise.
-  //
-  // Supports heterogeneous lookup, provided that the set has a compatible
-  // heterogeneous comparator.
-  using Base::contains;
-
-  // btree_multiset::count()
-  //
-  // template <typename K> size_type count(const K& key) const:
-  //
-  // Returns the number of elements comparing equal to the given `key` within
-  // the `btree_multiset`.
-  //
-  // Supports heterogeneous lookup, provided that the set has a compatible
-  // heterogeneous comparator.
-  using Base::count;
-
-  // btree_multiset::equal_range()
-  //
-  // Returns a closed range [first, last], defined by a `std::pair` of two
-  // iterators, containing all elements with the passed key in the
-  // `btree_multiset`.
-  using Base::equal_range;
-
-  // btree_multiset::find()
-  //
-  // template <typename K> iterator find(const K& key):
-  // template <typename K> const_iterator find(const K& key) const:
-  //
-  // Finds an element with the passed `key` within the `btree_multiset`.
-  //
-  // Supports heterogeneous lookup, provided that the set has a compatible
-  // heterogeneous comparator.
-  using Base::find;
-
-  // btree_multiset::lower_bound()
-  //
-  // template <typename K> iterator lower_bound(const K& key):
-  // template <typename K> const_iterator lower_bound(const K& key) const:
-  //
-  // Finds the first element that is not less than `key` within the
-  // `btree_multiset`.
-  //
-  // Supports heterogeneous lookup, provided that the set has a compatible
-  // heterogeneous comparator.
-  using Base::lower_bound;
-
-  // btree_multiset::upper_bound()
-  //
-  // template <typename K> iterator upper_bound(const K& key):
-  // template <typename K> const_iterator upper_bound(const K& key) const:
-  //
-  // Finds the first element that is greater than `key` within the
-  // `btree_multiset`.
-  //
-  // Supports heterogeneous lookup, provided that the set has a compatible
-  // heterogeneous comparator.
-  using Base::upper_bound;
-
-  // btree_multiset::get_allocator()
-  //
-  // Returns the allocator function associated with this `btree_multiset`.
-  using Base::get_allocator;
-
-  // btree_multiset::key_comp();
-  //
-  // Returns the key comparator associated with this `btree_multiset`.
-  using Base::key_comp;
-
-  // btree_multiset::value_comp();
-  //
-  // Returns the value comparator associated with this `btree_multiset`. The
-  // keys to sort the elements are the values themselves, therefore `value_comp`
-  // and its sibling member function `key_comp` are equivalent.
-  using Base::value_comp;
-};
-
-// absl::swap(absl::btree_multiset<>, absl::btree_multiset<>)
-//
-// Swaps the contents of two `absl::btree_multiset` containers.
-template <typename K, typename C, typename A>
-void swap(btree_multiset<K, C, A> &x, btree_multiset<K, C, A> &y) {
-  return x.swap(y);
-}
-
-// absl::erase_if(absl::btree_multiset<>, Pred)
-//
-// Erases all elements that satisfy the predicate pred from the container.
-// Returns the number of erased elements.
-template <typename K, typename C, typename A, typename Pred>
-typename btree_multiset<K, C, A>::size_type erase_if(
-   btree_multiset<K, C, A> & set, Pred pred) {
-  return container_internal::btree_access::erase_if(set, std::move(pred));
-}
-
-namespace container_internal {
-
-// This type implements the necessary functions from the
-// absl::container_internal::slot_type interface for btree_(multi)set.
-template <typename Key>
-struct set_slot_policy {
-  using slot_type = Key;
-  using value_type = Key;
-  using mutable_value_type = Key;
-
-  static value_type &element(slot_type *slot) { return *slot; }
-  static const value_type &element(const slot_type *slot) { return *slot; }
-
-  template <typename Alloc, class... Args>
-  static void construct(Alloc *alloc, slot_type *slot, Args &&...args) {
-    absl::allocator_traits<Alloc>::construct(*alloc, slot,
-                                             std::forward<Args>(args)...);
-  }
-
-  template <typename Alloc>
-  static void construct(Alloc *alloc, slot_type *slot, slot_type *other) {
-    absl::allocator_traits<Alloc>::construct(*alloc, slot, std::move(*other));
-  }
-
-  template <typename Alloc>
-  static void construct(Alloc *alloc, slot_type *slot, const slot_type *other) {
-    absl::allocator_traits<Alloc>::construct(*alloc, slot, *other);
-  }
-
-  template <typename Alloc>
-  static void destroy(Alloc *alloc, slot_type *slot) {
-    absl::allocator_traits<Alloc>::destroy(*alloc, slot);
-  }
-};
-
-// A parameters structure for holding the type parameters for a btree_set.
-// Compare and Alloc should be nothrow copy-constructible.
-template <typename Key, typename Compare, typename Alloc, int TargetNodeSize,
-          bool IsMulti>
-struct set_params : common_params<Key, Compare, Alloc, TargetNodeSize, IsMulti,
-                                  /*IsMap=*/false, set_slot_policy<Key>> {
-  using value_type = Key;
-  using slot_type = typename set_params::common_params::slot_type;
-
-  template <typename V>
-  static const V &key(const V &value) {
-    return value;
-  }
-  static const Key &key(const slot_type *slot) { return *slot; }
-  static const Key &key(slot_type *slot) { return *slot; }
-};
-
-}  // namespace container_internal
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_BTREE_SET_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/btree_test.h b/projects/llm_framework/include/abseil-cpp/absl/container/btree_test.h
deleted file mode 100644
index 62490807..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/btree_test.h
+++ /dev/null
@@ -1,166 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_BTREE_TEST_H_
-#define ABSL_CONTAINER_BTREE_TEST_H_
-
-#include <algorithm>
-#include <cassert>
-#include <random>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "absl/container/btree_map.h"
-#include "absl/container/btree_set.h"
-#include "absl/container/flat_hash_set.h"
-#include "absl/strings/cord.h"
-#include "absl/time/time.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-// Like remove_const but propagates the removal through std::pair.
-template <typename T>
-struct remove_pair_const {
-  using type = typename std::remove_const<T>::type;
-};
-template <typename T, typename U>
-struct remove_pair_const<std::pair<T, U> > {
-  using type = std::pair<typename remove_pair_const<T>::type,
-                         typename remove_pair_const<U>::type>;
-};
-
-// Utility class to provide an accessor for a key given a value. The default
-// behavior is to treat the value as a pair and return the first element.
-template <typename K, typename V>
-struct KeyOfValue {
-  struct type {
-    const K& operator()(const V& p) const { return p.first; }
-  };
-};
-
-// Partial specialization of KeyOfValue class for when the key and value are
-// the same type such as in set<> and btree_set<>.
-template <typename K>
-struct KeyOfValue<K, K> {
-  struct type {
-    const K& operator()(const K& k) const { return k; }
-  };
-};
-
-inline char* GenerateDigits(char buf[16], unsigned val, unsigned maxval) {
-  assert(val <= maxval);
-  constexpr unsigned kBase = 64;  // avoid integer division.
-  unsigned p = 15;
-  buf[p--] = 0;
-  while (maxval > 0) {
-    buf[p--] = ' ' + (val % kBase);
-    val /= kBase;
-    maxval /= kBase;
-  }
-  return buf + p + 1;
-}
-
-template <typename K>
-struct Generator {
-  int maxval;
-  explicit Generator(int m) : maxval(m) {}
-  K operator()(int i) const {
-    assert(i <= maxval);
-    return K(i);
-  }
-};
-
-template <>
-struct Generator<absl::Time> {
-  int maxval;
-  explicit Generator(int m) : maxval(m) {}
-  absl::Time operator()(int i) const { return absl::FromUnixMillis(i); }
-};
-
-template <>
-struct Generator<std::string> {
-  int maxval;
-  explicit Generator(int m) : maxval(m) {}
-  std::string operator()(int i) const {
-    char buf[16];
-    return GenerateDigits(buf, i, maxval);
-  }
-};
-
-template <>
-struct Generator<Cord> {
-  int maxval;
-  explicit Generator(int m) : maxval(m) {}
-  Cord operator()(int i) const {
-    char buf[16];
-    return Cord(GenerateDigits(buf, i, maxval));
-  }
-};
-
-template <typename T, typename U>
-struct Generator<std::pair<T, U> > {
-  Generator<typename remove_pair_const<T>::type> tgen;
-  Generator<typename remove_pair_const<U>::type> ugen;
-
-  explicit Generator(int m) : tgen(m), ugen(m) {}
-  std::pair<T, U> operator()(int i) const {
-    return std::make_pair(tgen(i), ugen(i));
-  }
-};
-
-// Generate n values for our tests and benchmarks. Value range is [0, maxval].
-inline std::vector<int> GenerateNumbersWithSeed(int n, int maxval, int seed) {
-  // NOTE: Some tests rely on generated numbers not changing between test runs.
-  // We use std::minstd_rand0 because it is well-defined, but don't use
-  // std::uniform_int_distribution because platforms use different algorithms.
-  std::minstd_rand0 rng(seed);
-
-  std::vector<int> values;
-  absl::flat_hash_set<int> unique_values;
-  if (values.size() < n) {
-    for (int i = values.size(); i < n; i++) {
-      int value;
-      do {
-        value = static_cast<int>(rng()) % (maxval + 1);
-      } while (!unique_values.insert(value).second);
-
-      values.push_back(value);
-    }
-  }
-  return values;
-}
-
-// Generates n values in the range [0, maxval].
-template <typename V>
-std::vector<V> GenerateValuesWithSeed(int n, int maxval, int seed) {
-  const std::vector<int> nums = GenerateNumbersWithSeed(n, maxval, seed);
-  Generator<V> gen(maxval);
-  std::vector<V> vec;
-
-  vec.reserve(n);
-  for (int i = 0; i < n; i++) {
-    vec.push_back(gen(nums[i]));
-  }
-
-  return vec;
-}
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_BTREE_TEST_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/fixed_array.h b/projects/llm_framework/include/abseil-cpp/absl/container/fixed_array.h
deleted file mode 100644
index 9f1c813d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/fixed_array.h
+++ /dev/null
@@ -1,556 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: fixed_array.h
-// -----------------------------------------------------------------------------
-//
-// A `FixedArray<T>` represents a non-resizable array of `T` where the length of
-// the array can be determined at run-time. It is a good replacement for
-// non-standard and deprecated uses of `alloca()` and variable length arrays
-// within the GCC extension. (See
-// https://gcc.gnu.org/onlinedocs/gcc/Variable-Length.html).
-//
-// `FixedArray` allocates small arrays inline, keeping performance fast by
-// avoiding heap operations. It also helps reduce the chances of
-// accidentally overflowing your stack if large input is passed to
-// your function.
-
-#ifndef ABSL_CONTAINER_FIXED_ARRAY_H_
-#define ABSL_CONTAINER_FIXED_ARRAY_H_
-
-#include <algorithm>
-#include <cassert>
-#include <cstddef>
-#include <initializer_list>
-#include <iterator>
-#include <limits>
-#include <memory>
-#include <new>
-#include <type_traits>
-
-#include "absl/algorithm/algorithm.h"
-#include "absl/base/config.h"
-#include "absl/base/dynamic_annotations.h"
-#include "absl/base/internal/throw_delegate.h"
-#include "absl/base/macros.h"
-#include "absl/base/optimization.h"
-#include "absl/base/port.h"
-#include "absl/container/internal/compressed_tuple.h"
-#include "absl/memory/memory.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-constexpr static auto kFixedArrayUseDefault = static_cast<size_t>(-1);
-
-// -----------------------------------------------------------------------------
-// FixedArray
-// -----------------------------------------------------------------------------
-//
-// A `FixedArray` provides a run-time fixed-size array, allocating a small array
-// inline for efficiency.
-//
-// Most users should not specify the `N` template parameter and let `FixedArray`
-// automatically determine the number of elements to store inline based on
-// `sizeof(T)`. If `N` is specified, the `FixedArray` implementation will use
-// inline storage for arrays with a length <= `N`.
-//
-// Note that a `FixedArray` constructed with a `size_type` argument will
-// default-initialize its values by leaving trivially constructible types
-// uninitialized (e.g. int, int[4], double), and others default-constructed.
-// This matches the behavior of c-style arrays and `std::array`, but not
-// `std::vector`.
-template <typename T, size_t N = kFixedArrayUseDefault,
-          typename A = std::allocator<T>>
-class FixedArray {
-  static_assert(!std::is_array<T>::value || std::extent<T>::value > 0,
-                "Arrays with unknown bounds cannot be used with FixedArray.");
-
-  static constexpr size_t kInlineBytesDefault = 256;
-
-  using AllocatorTraits = std::allocator_traits<A>;
-  // std::iterator_traits isn't guaranteed to be SFINAE-friendly until C++17,
-  // but this seems to be mostly pedantic.
-  template <typename Iterator>
-  using EnableIfForwardIterator = absl::enable_if_t<std::is_convertible<
-      typename std::iterator_traits<Iterator>::iterator_category,
-      std::forward_iterator_tag>::value>;
-  static constexpr bool NoexceptCopyable() {
-    return std::is_nothrow_copy_constructible<StorageElement>::value &&
-           absl::allocator_is_nothrow<allocator_type>::value;
-  }
-  static constexpr bool NoexceptMovable() {
-    return std::is_nothrow_move_constructible<StorageElement>::value &&
-           absl::allocator_is_nothrow<allocator_type>::value;
-  }
-  static constexpr bool DefaultConstructorIsNonTrivial() {
-    return !absl::is_trivially_default_constructible<StorageElement>::value;
-  }
-
- public:
-  using allocator_type = typename AllocatorTraits::allocator_type;
-  using value_type = typename AllocatorTraits::value_type;
-  using pointer = typename AllocatorTraits::pointer;
-  using const_pointer = typename AllocatorTraits::const_pointer;
-  using reference = value_type&;
-  using const_reference = const value_type&;
-  using size_type = typename AllocatorTraits::size_type;
-  using difference_type = typename AllocatorTraits::difference_type;
-  using iterator = pointer;
-  using const_iterator = const_pointer;
-  using reverse_iterator = std::reverse_iterator<iterator>;
-  using const_reverse_iterator = std::reverse_iterator<const_iterator>;
-
-  static constexpr size_type inline_elements =
-      (N == kFixedArrayUseDefault ? kInlineBytesDefault / sizeof(value_type)
-                                  : static_cast<size_type>(N));
-
-  FixedArray(const FixedArray& other) noexcept(NoexceptCopyable())
-      : FixedArray(other,
-                   AllocatorTraits::select_on_container_copy_construction(
-                       other.storage_.alloc())) {}
-
-  FixedArray(const FixedArray& other,
-             const allocator_type& a) noexcept(NoexceptCopyable())
-      : FixedArray(other.begin(), other.end(), a) {}
-
-  FixedArray(FixedArray&& other) noexcept(NoexceptMovable())
-      : FixedArray(std::move(other), other.storage_.alloc()) {}
-
-  FixedArray(FixedArray&& other,
-             const allocator_type& a) noexcept(NoexceptMovable())
-      : FixedArray(std::make_move_iterator(other.begin()),
-                   std::make_move_iterator(other.end()), a) {}
-
-  // Creates an array object that can store `n` elements.
-  // Note that trivially constructible elements will be uninitialized.
-  explicit FixedArray(size_type n, const allocator_type& a = allocator_type())
-      : storage_(n, a) {
-    if (DefaultConstructorIsNonTrivial()) {
-      memory_internal::ConstructRange(storage_.alloc(), storage_.begin(),
-                                      storage_.end());
-    }
-  }
-
-  // Creates an array initialized with `n` copies of `val`.
-  FixedArray(size_type n, const value_type& val,
-             const allocator_type& a = allocator_type())
-      : storage_(n, a) {
-    memory_internal::ConstructRange(storage_.alloc(), storage_.begin(),
-                                    storage_.end(), val);
-  }
-
-  // Creates an array initialized with the size and contents of `init_list`.
-  FixedArray(std::initializer_list<value_type> init_list,
-             const allocator_type& a = allocator_type())
-      : FixedArray(init_list.begin(), init_list.end(), a) {}
-
-  // Creates an array initialized with the elements from the input
-  // range. The array's size will always be `std::distance(first, last)`.
-  // REQUIRES: Iterator must be a forward_iterator or better.
-  template <typename Iterator, EnableIfForwardIterator<Iterator>* = nullptr>
-  FixedArray(Iterator first, Iterator last,
-             const allocator_type& a = allocator_type())
-      : storage_(std::distance(first, last), a) {
-    memory_internal::CopyRange(storage_.alloc(), storage_.begin(), first, last);
-  }
-
-  ~FixedArray() noexcept {
-    for (auto* cur = storage_.begin(); cur != storage_.end(); ++cur) {
-      AllocatorTraits::destroy(storage_.alloc(), cur);
-    }
-  }
-
-  // Assignments are deleted because they break the invariant that the size of a
-  // `FixedArray` never changes.
-  void operator=(FixedArray&&) = delete;
-  void operator=(const FixedArray&) = delete;
-
-  // FixedArray::size()
-  //
-  // Returns the length of the fixed array.
-  size_type size() const { return storage_.size(); }
-
-  // FixedArray::max_size()
-  //
-  // Returns the largest possible value of `std::distance(begin(), end())` for a
-  // `FixedArray<T>`. This is equivalent to the most possible addressable bytes
-  // over the number of bytes taken by T.
-  constexpr size_type max_size() const {
-    return (std::numeric_limits<difference_type>::max)() / sizeof(value_type);
-  }
-
-  // FixedArray::empty()
-  //
-  // Returns whether or not the fixed array is empty.
-  bool empty() const { return size() == 0; }
-
-  // FixedArray::memsize()
-  //
-  // Returns the memory size of the fixed array in bytes.
-  size_t memsize() const { return size() * sizeof(value_type); }
-
-  // FixedArray::data()
-  //
-  // Returns a const T* pointer to elements of the `FixedArray`. This pointer
-  // can be used to access (but not modify) the contained elements.
-  const_pointer data() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return AsValueType(storage_.begin());
-  }
-
-  // Overload of FixedArray::data() to return a T* pointer to elements of the
-  // fixed array. This pointer can be used to access and modify the contained
-  // elements.
-  pointer data() ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return AsValueType(storage_.begin());
-  }
-
-  // FixedArray::operator[]
-  //
-  // Returns a reference the ith element of the fixed array.
-  // REQUIRES: 0 <= i < size()
-  reference operator[](size_type i) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(i < size());
-    return data()[i];
-  }
-
-  // Overload of FixedArray::operator()[] to return a const reference to the
-  // ith element of the fixed array.
-  // REQUIRES: 0 <= i < size()
-  const_reference operator[](size_type i) const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(i < size());
-    return data()[i];
-  }
-
-  // FixedArray::at
-  //
-  // Bounds-checked access.  Returns a reference to the ith element of the fixed
-  // array, or throws std::out_of_range
-  reference at(size_type i) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    if (ABSL_PREDICT_FALSE(i >= size())) {
-      base_internal::ThrowStdOutOfRange("FixedArray::at failed bounds check");
-    }
-    return data()[i];
-  }
-
-  // Overload of FixedArray::at() to return a const reference to the ith element
-  // of the fixed array.
-  const_reference at(size_type i) const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    if (ABSL_PREDICT_FALSE(i >= size())) {
-      base_internal::ThrowStdOutOfRange("FixedArray::at failed bounds check");
-    }
-    return data()[i];
-  }
-
-  // FixedArray::front()
-  //
-  // Returns a reference to the first element of the fixed array.
-  reference front() ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(!empty());
-    return data()[0];
-  }
-
-  // Overload of FixedArray::front() to return a reference to the first element
-  // of a fixed array of const values.
-  const_reference front() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(!empty());
-    return data()[0];
-  }
-
-  // FixedArray::back()
-  //
-  // Returns a reference to the last element of the fixed array.
-  reference back() ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(!empty());
-    return data()[size() - 1];
-  }
-
-  // Overload of FixedArray::back() to return a reference to the last element
-  // of a fixed array of const values.
-  const_reference back() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(!empty());
-    return data()[size() - 1];
-  }
-
-  // FixedArray::begin()
-  //
-  // Returns an iterator to the beginning of the fixed array.
-  iterator begin() ABSL_ATTRIBUTE_LIFETIME_BOUND { return data(); }
-
-  // Overload of FixedArray::begin() to return a const iterator to the
-  // beginning of the fixed array.
-  const_iterator begin() const ABSL_ATTRIBUTE_LIFETIME_BOUND { return data(); }
-
-  // FixedArray::cbegin()
-  //
-  // Returns a const iterator to the beginning of the fixed array.
-  const_iterator cbegin() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return begin();
-  }
-
-  // FixedArray::end()
-  //
-  // Returns an iterator to the end of the fixed array.
-  iterator end() ABSL_ATTRIBUTE_LIFETIME_BOUND { return data() + size(); }
-
-  // Overload of FixedArray::end() to return a const iterator to the end of the
-  // fixed array.
-  const_iterator end() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return data() + size();
-  }
-
-  // FixedArray::cend()
-  //
-  // Returns a const iterator to the end of the fixed array.
-  const_iterator cend() const ABSL_ATTRIBUTE_LIFETIME_BOUND { return end(); }
-
-  // FixedArray::rbegin()
-  //
-  // Returns a reverse iterator from the end of the fixed array.
-  reverse_iterator rbegin() ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return reverse_iterator(end());
-  }
-
-  // Overload of FixedArray::rbegin() to return a const reverse iterator from
-  // the end of the fixed array.
-  const_reverse_iterator rbegin() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return const_reverse_iterator(end());
-  }
-
-  // FixedArray::crbegin()
-  //
-  // Returns a const reverse iterator from the end of the fixed array.
-  const_reverse_iterator crbegin() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return rbegin();
-  }
-
-  // FixedArray::rend()
-  //
-  // Returns a reverse iterator from the beginning of the fixed array.
-  reverse_iterator rend() ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return reverse_iterator(begin());
-  }
-
-  // Overload of FixedArray::rend() for returning a const reverse iterator
-  // from the beginning of the fixed array.
-  const_reverse_iterator rend() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return const_reverse_iterator(begin());
-  }
-
-  // FixedArray::crend()
-  //
-  // Returns a reverse iterator from the beginning of the fixed array.
-  const_reverse_iterator crend() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return rend();
-  }
-
-  // FixedArray::fill()
-  //
-  // Assigns the given `value` to all elements in the fixed array.
-  void fill(const value_type& val) { std::fill(begin(), end(), val); }
-
-  // Relational operators. Equality operators are elementwise using
-  // `operator==`, while order operators order FixedArrays lexicographically.
-  friend bool operator==(const FixedArray& lhs, const FixedArray& rhs) {
-    return std::equal(lhs.begin(), lhs.end(), rhs.begin(), rhs.end());
-  }
-
-  friend bool operator!=(const FixedArray& lhs, const FixedArray& rhs) {
-    return !(lhs == rhs);
-  }
-
-  friend bool operator<(const FixedArray& lhs, const FixedArray& rhs) {
-    return std::lexicographical_compare(lhs.begin(), lhs.end(), rhs.begin(),
-                                        rhs.end());
-  }
-
-  friend bool operator>(const FixedArray& lhs, const FixedArray& rhs) {
-    return rhs < lhs;
-  }
-
-  friend bool operator<=(const FixedArray& lhs, const FixedArray& rhs) {
-    return !(rhs < lhs);
-  }
-
-  friend bool operator>=(const FixedArray& lhs, const FixedArray& rhs) {
-    return !(lhs < rhs);
-  }
-
-  template <typename H>
-  friend H AbslHashValue(H h, const FixedArray& v) {
-    return H::combine(H::combine_contiguous(std::move(h), v.data(), v.size()),
-                      v.size());
-  }
-
- private:
-  // StorageElement
-  //
-  // For FixedArrays with a C-style-array value_type, StorageElement is a POD
-  // wrapper struct called StorageElementWrapper that holds the value_type
-  // instance inside. This is needed for construction and destruction of the
-  // entire array regardless of how many dimensions it has. For all other cases,
-  // StorageElement is just an alias of value_type.
-  //
-  // Maintainer's Note: The simpler solution would be to simply wrap value_type
-  // in a struct whether it's an array or not. That causes some paranoid
-  // diagnostics to misfire, believing that 'data()' returns a pointer to a
-  // single element, rather than the packed array that it really is.
-  // e.g.:
-  //
-  //     FixedArray<char> buf(1);
-  //     sprintf(buf.data(), "foo");
-  //
-  //     error: call to int __builtin___sprintf_chk(etc...)
-  //     will always overflow destination buffer [-Werror]
-  //
-  template <typename OuterT, typename InnerT = absl::remove_extent_t<OuterT>,
-            size_t InnerN = std::extent<OuterT>::value>
-  struct StorageElementWrapper {
-    InnerT array[InnerN];
-  };
-
-  using StorageElement =
-      absl::conditional_t<std::is_array<value_type>::value,
-                          StorageElementWrapper<value_type>, value_type>;
-
-  static pointer AsValueType(pointer ptr) { return ptr; }
-  static pointer AsValueType(StorageElementWrapper<value_type>* ptr) {
-    return std::addressof(ptr->array);
-  }
-
-  static_assert(sizeof(StorageElement) == sizeof(value_type), "");
-  static_assert(alignof(StorageElement) == alignof(value_type), "");
-
-  class NonEmptyInlinedStorage {
-   public:
-    StorageElement* data() { return reinterpret_cast<StorageElement*>(buff_); }
-    void AnnotateConstruct(size_type n);
-    void AnnotateDestruct(size_type n);
-
-#ifdef ABSL_HAVE_ADDRESS_SANITIZER
-    void* RedzoneBegin() { return &redzone_begin_; }
-    void* RedzoneEnd() { return &redzone_end_ + 1; }
-#endif  // ABSL_HAVE_ADDRESS_SANITIZER
-
-   private:
-    ABSL_ADDRESS_SANITIZER_REDZONE(redzone_begin_);
-    alignas(StorageElement) char buff_[sizeof(StorageElement[inline_elements])];
-    ABSL_ADDRESS_SANITIZER_REDZONE(redzone_end_);
-  };
-
-  class EmptyInlinedStorage {
-   public:
-    StorageElement* data() { return nullptr; }
-    void AnnotateConstruct(size_type) {}
-    void AnnotateDestruct(size_type) {}
-  };
-
-  using InlinedStorage =
-      absl::conditional_t<inline_elements == 0, EmptyInlinedStorage,
-                          NonEmptyInlinedStorage>;
-
-  // Storage
-  //
-  // An instance of Storage manages the inline and out-of-line memory for
-  // instances of FixedArray. This guarantees that even when construction of
-  // individual elements fails in the FixedArray constructor body, the
-  // destructor for Storage will still be called and out-of-line memory will be
-  // properly deallocated.
-  //
-  class Storage : public InlinedStorage {
-   public:
-    Storage(size_type n, const allocator_type& a)
-        : size_alloc_(n, a), data_(InitializeData()) {}
-
-    ~Storage() noexcept {
-      if (UsingInlinedStorage(size())) {
-        InlinedStorage::AnnotateDestruct(size());
-      } else {
-        AllocatorTraits::deallocate(alloc(), AsValueType(begin()), size());
-      }
-    }
-
-    size_type size() const { return size_alloc_.template get<0>(); }
-    StorageElement* begin() const { return data_; }
-    StorageElement* end() const { return begin() + size(); }
-    allocator_type& alloc() { return size_alloc_.template get<1>(); }
-    const allocator_type& alloc() const {
-      return size_alloc_.template get<1>();
-    }
-
-   private:
-    static bool UsingInlinedStorage(size_type n) {
-      return n <= inline_elements;
-    }
-
-#ifdef ABSL_HAVE_ADDRESS_SANITIZER
-    ABSL_ATTRIBUTE_NOINLINE
-#endif  // ABSL_HAVE_ADDRESS_SANITIZER
-    StorageElement* InitializeData() {
-      if (UsingInlinedStorage(size())) {
-        InlinedStorage::AnnotateConstruct(size());
-        return InlinedStorage::data();
-      } else {
-        return reinterpret_cast<StorageElement*>(
-            AllocatorTraits::allocate(alloc(), size()));
-      }
-    }
-
-    // `CompressedTuple` takes advantage of EBCO for stateless `allocator_type`s
-    container_internal::CompressedTuple<size_type, allocator_type> size_alloc_;
-    StorageElement* data_;
-  };
-
-  Storage storage_;
-};
-
-#ifdef ABSL_INTERNAL_NEED_REDUNDANT_CONSTEXPR_DECL
-template <typename T, size_t N, typename A>
-constexpr size_t FixedArray<T, N, A>::kInlineBytesDefault;
-
-template <typename T, size_t N, typename A>
-constexpr typename FixedArray<T, N, A>::size_type
-    FixedArray<T, N, A>::inline_elements;
-#endif
-
-template <typename T, size_t N, typename A>
-void FixedArray<T, N, A>::NonEmptyInlinedStorage::AnnotateConstruct(
-    typename FixedArray<T, N, A>::size_type n) {
-#ifdef ABSL_HAVE_ADDRESS_SANITIZER
-  if (!n) return;
-  ABSL_ANNOTATE_CONTIGUOUS_CONTAINER(data(), RedzoneEnd(), RedzoneEnd(),
-                                     data() + n);
-  ABSL_ANNOTATE_CONTIGUOUS_CONTAINER(RedzoneBegin(), data(), data(),
-                                     RedzoneBegin());
-#endif  // ABSL_HAVE_ADDRESS_SANITIZER
-  static_cast<void>(n);  // Mark used when not in asan mode
-}
-
-template <typename T, size_t N, typename A>
-void FixedArray<T, N, A>::NonEmptyInlinedStorage::AnnotateDestruct(
-    typename FixedArray<T, N, A>::size_type n) {
-#ifdef ABSL_HAVE_ADDRESS_SANITIZER
-  if (!n) return;
-  ABSL_ANNOTATE_CONTIGUOUS_CONTAINER(data(), RedzoneEnd(), data() + n,
-                                     RedzoneEnd());
-  ABSL_ANNOTATE_CONTIGUOUS_CONTAINER(RedzoneBegin(), data(), RedzoneBegin(),
-                                     data());
-#endif  // ABSL_HAVE_ADDRESS_SANITIZER
-  static_cast<void>(n);  // Mark used when not in asan mode
-}
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_FIXED_ARRAY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/flat_hash_map.h b/projects/llm_framework/include/abseil-cpp/absl/container/flat_hash_map.h
deleted file mode 100644
index a33c794f..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/flat_hash_map.h
+++ /dev/null
@@ -1,638 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: flat_hash_map.h
-// -----------------------------------------------------------------------------
-//
-// An `absl::flat_hash_map<K, V>` is an unordered associative container of
-// unique keys and associated values designed to be a more efficient replacement
-// for `std::unordered_map`. Like `unordered_map`, search, insertion, and
-// deletion of map elements can be done as an `O(1)` operation. However,
-// `flat_hash_map` (and other unordered associative containers known as the
-// collection of Abseil "Swiss tables") contain other optimizations that result
-// in both memory and computation advantages.
-//
-// In most cases, your default choice for a hash map should be a map of type
-// `flat_hash_map`.
-
-#ifndef ABSL_CONTAINER_FLAT_HASH_MAP_H_
-#define ABSL_CONTAINER_FLAT_HASH_MAP_H_
-
-#include <cstddef>
-#include <new>
-#include <type_traits>
-#include <utility>
-
-#include "absl/algorithm/container.h"
-#include "absl/base/macros.h"
-#include "absl/container/internal/container_memory.h"
-#include "absl/container/internal/hash_function_defaults.h"  // IWYU pragma: export
-#include "absl/container/internal/raw_hash_map.h"  // IWYU pragma: export
-#include "absl/memory/memory.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-template <class K, class V>
-struct FlatHashMapPolicy;
-}  // namespace container_internal
-
-// -----------------------------------------------------------------------------
-// absl::flat_hash_map
-// -----------------------------------------------------------------------------
-//
-// An `absl::flat_hash_map<K, V>` is an unordered associative container which
-// has been optimized for both speed and memory footprint in most common use
-// cases. Its interface is similar to that of `std::unordered_map<K, V>` with
-// the following notable differences:
-//
-// * Requires keys that are CopyConstructible
-// * Requires values that are MoveConstructible
-// * Supports heterogeneous lookup, through `find()`, `operator[]()` and
-//   `insert()`, provided that the map is provided a compatible heterogeneous
-//   hashing function and equality operator. See below for details.
-// * Invalidates any references and pointers to elements within the table after
-//   `rehash()` and when the table is moved.
-// * Contains a `capacity()` member function indicating the number of element
-//   slots (open, deleted, and empty) within the hash map.
-// * Returns `void` from the `erase(iterator)` overload.
-//
-// By default, `flat_hash_map` uses the `absl::Hash` hashing framework.
-// All fundamental and Abseil types that support the `absl::Hash` framework have
-// a compatible equality operator for comparing insertions into `flat_hash_map`.
-// If your type is not yet supported by the `absl::Hash` framework, see
-// absl/hash/hash.h for information on extending Abseil hashing to user-defined
-// types.
-//
-// Using `absl::flat_hash_map` at interface boundaries in dynamically loaded
-// libraries (e.g. .dll, .so) is unsupported due to way `absl::Hash` values may
-// be randomized across dynamically loaded libraries.
-//
-// To achieve heterogeneous lookup for custom types either `Hash` and `Eq` type
-// parameters can be used or `T` should have public inner types
-// `absl_container_hash` and (optionally) `absl_container_eq`. In either case,
-// `typename Hash::is_transparent` and `typename Eq::is_transparent` should be
-// well-formed. Both types are basically functors:
-// * `Hash` should support `size_t operator()(U val) const` that returns a hash
-// for the given `val`.
-// * `Eq` should support `bool operator()(U lhs, V rhs) const` that returns true
-// if `lhs` is equal to `rhs`.
-//
-// In most cases `T` needs only to provide the `absl_container_hash`. In this
-// case `std::equal_to<void>` will be used instead of `eq` part.
-//
-// NOTE: A `flat_hash_map` stores its value types directly inside its
-// implementation array to avoid memory indirection. Because a `flat_hash_map`
-// is designed to move data when rehashed, map values will not retain pointer
-// stability. If you require pointer stability, or if your values are large,
-// consider using `absl::flat_hash_map<Key, std::unique_ptr<Value>>` instead.
-// If your types are not moveable or you require pointer stability for keys,
-// consider `absl::node_hash_map`.
-//
-// Example:
-//
-//   // Create a flat hash map of three strings (that map to strings)
-//   absl::flat_hash_map<std::string, std::string> ducks =
-//     {{"a", "huey"}, {"b", "dewey"}, {"c", "louie"}};
-//
-//  // Insert a new element into the flat hash map
-//  ducks.insert({"d", "donald"});
-//
-//  // Force a rehash of the flat hash map
-//  ducks.rehash(0);
-//
-//  // Find the element with the key "b"
-//  std::string search_key = "b";
-//  auto result = ducks.find(search_key);
-//  if (result != ducks.end()) {
-//    std::cout << "Result: " << result->second << std::endl;
-//  }
-template <class K, class V,
-          class Hash = absl::container_internal::hash_default_hash<K>,
-          class Eq = absl::container_internal::hash_default_eq<K>,
-          class Allocator = std::allocator<std::pair<const K, V>>>
-class flat_hash_map : public absl::container_internal::raw_hash_map<
-                          absl::container_internal::FlatHashMapPolicy<K, V>,
-                          Hash, Eq, Allocator> {
-  using Base = typename flat_hash_map::raw_hash_map;
-
- public:
-  // Constructors and Assignment Operators
-  //
-  // A flat_hash_map supports the same overload set as `std::unordered_map`
-  // for construction and assignment:
-  //
-  // *  Default constructor
-  //
-  //    // No allocation for the table's elements is made.
-  //    absl::flat_hash_map<int, std::string> map1;
-  //
-  // * Initializer List constructor
-  //
-  //   absl::flat_hash_map<int, std::string> map2 =
-  //       {{1, "huey"}, {2, "dewey"}, {3, "louie"},};
-  //
-  // * Copy constructor
-  //
-  //   absl::flat_hash_map<int, std::string> map3(map2);
-  //
-  // * Copy assignment operator
-  //
-  //  // Hash functor and Comparator are copied as well
-  //  absl::flat_hash_map<int, std::string> map4;
-  //  map4 = map3;
-  //
-  // * Move constructor
-  //
-  //   // Move is guaranteed efficient
-  //   absl::flat_hash_map<int, std::string> map5(std::move(map4));
-  //
-  // * Move assignment operator
-  //
-  //   // May be efficient if allocators are compatible
-  //   absl::flat_hash_map<int, std::string> map6;
-  //   map6 = std::move(map5);
-  //
-  // * Range constructor
-  //
-  //   std::vector<std::pair<int, std::string>> v = {{1, "a"}, {2, "b"}};
-  //   absl::flat_hash_map<int, std::string> map7(v.begin(), v.end());
-  flat_hash_map() {}
-  using Base::Base;
-
-  // flat_hash_map::begin()
-  //
-  // Returns an iterator to the beginning of the `flat_hash_map`.
-  using Base::begin;
-
-  // flat_hash_map::cbegin()
-  //
-  // Returns a const iterator to the beginning of the `flat_hash_map`.
-  using Base::cbegin;
-
-  // flat_hash_map::cend()
-  //
-  // Returns a const iterator to the end of the `flat_hash_map`.
-  using Base::cend;
-
-  // flat_hash_map::end()
-  //
-  // Returns an iterator to the end of the `flat_hash_map`.
-  using Base::end;
-
-  // flat_hash_map::capacity()
-  //
-  // Returns the number of element slots (assigned, deleted, and empty)
-  // available within the `flat_hash_map`.
-  //
-  // NOTE: this member function is particular to `absl::flat_hash_map` and is
-  // not provided in the `std::unordered_map` API.
-  using Base::capacity;
-
-  // flat_hash_map::empty()
-  //
-  // Returns whether or not the `flat_hash_map` is empty.
-  using Base::empty;
-
-  // flat_hash_map::max_size()
-  //
-  // Returns the largest theoretical possible number of elements within a
-  // `flat_hash_map` under current memory constraints. This value can be thought
-  // of the largest value of `std::distance(begin(), end())` for a
-  // `flat_hash_map<K, V>`.
-  using Base::max_size;
-
-  // flat_hash_map::size()
-  //
-  // Returns the number of elements currently within the `flat_hash_map`.
-  using Base::size;
-
-  // flat_hash_map::clear()
-  //
-  // Removes all elements from the `flat_hash_map`. Invalidates any references,
-  // pointers, or iterators referring to contained elements.
-  //
-  // NOTE: this operation may shrink the underlying buffer. To avoid shrinking
-  // the underlying buffer call `erase(begin(), end())`.
-  using Base::clear;
-
-  // flat_hash_map::erase()
-  //
-  // Erases elements within the `flat_hash_map`. Erasing does not trigger a
-  // rehash. Overloads are listed below.
-  //
-  // void erase(const_iterator pos):
-  //
-  //   Erases the element at `position` of the `flat_hash_map`, returning
-  //   `void`.
-  //
-  //   NOTE: returning `void` in this case is different than that of STL
-  //   containers in general and `std::unordered_map` in particular (which
-  //   return an iterator to the element following the erased element). If that
-  //   iterator is needed, simply post increment the iterator:
-  //
-  //     map.erase(it++);
-  //
-  // iterator erase(const_iterator first, const_iterator last):
-  //
-  //   Erases the elements in the open interval [`first`, `last`), returning an
-  //   iterator pointing to `last`. The special case of calling
-  //   `erase(begin(), end())` resets the reserved growth such that if
-  //   `reserve(N)` has previously been called and there has been no intervening
-  //   call to `clear()`, then after calling `erase(begin(), end())`, it is safe
-  //   to assume that inserting N elements will not cause a rehash.
-  //
-  // size_type erase(const key_type& key):
-  //
-  //   Erases the element with the matching key, if it exists, returning the
-  //   number of elements erased (0 or 1).
-  using Base::erase;
-
-  // flat_hash_map::insert()
-  //
-  // Inserts an element of the specified value into the `flat_hash_map`,
-  // returning an iterator pointing to the newly inserted element, provided that
-  // an element with the given key does not already exist. If rehashing occurs
-  // due to the insertion, all iterators are invalidated. Overloads are listed
-  // below.
-  //
-  // std::pair<iterator,bool> insert(const init_type& value):
-  //
-  //   Inserts a value into the `flat_hash_map`. Returns a pair consisting of an
-  //   iterator to the inserted element (or to the element that prevented the
-  //   insertion) and a bool denoting whether the insertion took place.
-  //
-  // std::pair<iterator,bool> insert(T&& value):
-  // std::pair<iterator,bool> insert(init_type&& value):
-  //
-  //   Inserts a moveable value into the `flat_hash_map`. Returns a pair
-  //   consisting of an iterator to the inserted element (or to the element that
-  //   prevented the insertion) and a bool denoting whether the insertion took
-  //   place.
-  //
-  // iterator insert(const_iterator hint, const init_type& value):
-  // iterator insert(const_iterator hint, T&& value):
-  // iterator insert(const_iterator hint, init_type&& value);
-  //
-  //   Inserts a value, using the position of `hint` as a non-binding suggestion
-  //   for where to begin the insertion search. Returns an iterator to the
-  //   inserted element, or to the existing element that prevented the
-  //   insertion.
-  //
-  // void insert(InputIterator first, InputIterator last):
-  //
-  //   Inserts a range of values [`first`, `last`).
-  //
-  //   NOTE: Although the STL does not specify which element may be inserted if
-  //   multiple keys compare equivalently, for `flat_hash_map` we guarantee the
-  //   first match is inserted.
-  //
-  // void insert(std::initializer_list<init_type> ilist):
-  //
-  //   Inserts the elements within the initializer list `ilist`.
-  //
-  //   NOTE: Although the STL does not specify which element may be inserted if
-  //   multiple keys compare equivalently within the initializer list, for
-  //   `flat_hash_map` we guarantee the first match is inserted.
-  using Base::insert;
-
-  // flat_hash_map::insert_or_assign()
-  //
-  // Inserts an element of the specified value into the `flat_hash_map` provided
-  // that a value with the given key does not already exist, or replaces it with
-  // the element value if a key for that value already exists, returning an
-  // iterator pointing to the newly inserted element.  If rehashing occurs due
-  // to the insertion, all existing iterators are invalidated. Overloads are
-  // listed below.
-  //
-  // pair<iterator, bool> insert_or_assign(const init_type& k, T&& obj):
-  // pair<iterator, bool> insert_or_assign(init_type&& k, T&& obj):
-  //
-  //   Inserts/Assigns (or moves) the element of the specified key into the
-  //   `flat_hash_map`.
-  //
-  // iterator insert_or_assign(const_iterator hint,
-  //                           const init_type& k, T&& obj):
-  // iterator insert_or_assign(const_iterator hint, init_type&& k, T&& obj):
-  //
-  //   Inserts/Assigns (or moves) the element of the specified key into the
-  //   `flat_hash_map` using the position of `hint` as a non-binding suggestion
-  //   for where to begin the insertion search.
-  using Base::insert_or_assign;
-
-  // flat_hash_map::emplace()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `flat_hash_map`, provided that no element with the given key
-  // already exists.
-  //
-  // The element may be constructed even if there already is an element with the
-  // key in the container, in which case the newly constructed element will be
-  // destroyed immediately. Prefer `try_emplace()` unless your key is not
-  // copyable or moveable.
-  //
-  // If rehashing occurs due to the insertion, all iterators are invalidated.
-  using Base::emplace;
-
-  // flat_hash_map::emplace_hint()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `flat_hash_map`, using the position of `hint` as a non-binding
-  // suggestion for where to begin the insertion search, and only inserts
-  // provided that no element with the given key already exists.
-  //
-  // The element may be constructed even if there already is an element with the
-  // key in the container, in which case the newly constructed element will be
-  // destroyed immediately. Prefer `try_emplace()` unless your key is not
-  // copyable or moveable.
-  //
-  // If rehashing occurs due to the insertion, all iterators are invalidated.
-  using Base::emplace_hint;
-
-  // flat_hash_map::try_emplace()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `flat_hash_map`, provided that no element with the given key
-  // already exists. Unlike `emplace()`, if an element with the given key
-  // already exists, we guarantee that no element is constructed.
-  //
-  // If rehashing occurs due to the insertion, all iterators are invalidated.
-  // Overloads are listed below.
-  //
-  //   pair<iterator, bool> try_emplace(const key_type& k, Args&&... args):
-  //   pair<iterator, bool> try_emplace(key_type&& k, Args&&... args):
-  //
-  // Inserts (via copy or move) the element of the specified key into the
-  // `flat_hash_map`.
-  //
-  //   iterator try_emplace(const_iterator hint,
-  //                        const key_type& k, Args&&... args):
-  //   iterator try_emplace(const_iterator hint, key_type&& k, Args&&... args):
-  //
-  // Inserts (via copy or move) the element of the specified key into the
-  // `flat_hash_map` using the position of `hint` as a non-binding suggestion
-  // for where to begin the insertion search.
-  //
-  // All `try_emplace()` overloads make the same guarantees regarding rvalue
-  // arguments as `std::unordered_map::try_emplace()`, namely that these
-  // functions will not move from rvalue arguments if insertions do not happen.
-  using Base::try_emplace;
-
-  // flat_hash_map::extract()
-  //
-  // Extracts the indicated element, erasing it in the process, and returns it
-  // as a C++17-compatible node handle. Overloads are listed below.
-  //
-  // node_type extract(const_iterator position):
-  //
-  //   Extracts the key,value pair of the element at the indicated position and
-  //   returns a node handle owning that extracted data.
-  //
-  // node_type extract(const key_type& x):
-  //
-  //   Extracts the key,value pair of the element with a key matching the passed
-  //   key value and returns a node handle owning that extracted data. If the
-  //   `flat_hash_map` does not contain an element with a matching key, this
-  //   function returns an empty node handle.
-  //
-  // NOTE: when compiled in an earlier version of C++ than C++17,
-  // `node_type::key()` returns a const reference to the key instead of a
-  // mutable reference. We cannot safely return a mutable reference without
-  // std::launder (which is not available before C++17).
-  using Base::extract;
-
-  // flat_hash_map::merge()
-  //
-  // Extracts elements from a given `source` flat hash map into this
-  // `flat_hash_map`. If the destination `flat_hash_map` already contains an
-  // element with an equivalent key, that element is not extracted.
-  using Base::merge;
-
-  // flat_hash_map::swap(flat_hash_map& other)
-  //
-  // Exchanges the contents of this `flat_hash_map` with those of the `other`
-  // flat hash map, avoiding invocation of any move, copy, or swap operations on
-  // individual elements.
-  //
-  // All iterators and references on the `flat_hash_map` remain valid, excepting
-  // for the past-the-end iterator, which is invalidated.
-  //
-  // `swap()` requires that the flat hash map's hashing and key equivalence
-  // functions be Swappable, and are exchanged using unqualified calls to
-  // non-member `swap()`. If the map's allocator has
-  // `std::allocator_traits<allocator_type>::propagate_on_container_swap::value`
-  // set to `true`, the allocators are also exchanged using an unqualified call
-  // to non-member `swap()`; otherwise, the allocators are not swapped.
-  using Base::swap;
-
-  // flat_hash_map::rehash(count)
-  //
-  // Rehashes the `flat_hash_map`, setting the number of slots to be at least
-  // the passed value. If the new number of slots increases the load factor more
-  // than the current maximum load factor
-  // (`count` < `size()` / `max_load_factor()`), then the new number of slots
-  // will be at least `size()` / `max_load_factor()`.
-  //
-  // To force a rehash, pass rehash(0).
-  //
-  // NOTE: unlike behavior in `std::unordered_map`, references are also
-  // invalidated upon a `rehash()`.
-  using Base::rehash;
-
-  // flat_hash_map::reserve(count)
-  //
-  // Sets the number of slots in the `flat_hash_map` to the number needed to
-  // accommodate at least `count` total elements without exceeding the current
-  // maximum load factor, and may rehash the container if needed.
-  using Base::reserve;
-
-  // flat_hash_map::at()
-  //
-  // Returns a reference to the mapped value of the element with key equivalent
-  // to the passed key.
-  using Base::at;
-
-  // flat_hash_map::contains()
-  //
-  // Determines whether an element with a key comparing equal to the given `key`
-  // exists within the `flat_hash_map`, returning `true` if so or `false`
-  // otherwise.
-  using Base::contains;
-
-  // flat_hash_map::count(const Key& key) const
-  //
-  // Returns the number of elements with a key comparing equal to the given
-  // `key` within the `flat_hash_map`. note that this function will return
-  // either `1` or `0` since duplicate keys are not allowed within a
-  // `flat_hash_map`.
-  using Base::count;
-
-  // flat_hash_map::equal_range()
-  //
-  // Returns a closed range [first, last], defined by a `std::pair` of two
-  // iterators, containing all elements with the passed key in the
-  // `flat_hash_map`.
-  using Base::equal_range;
-
-  // flat_hash_map::find()
-  //
-  // Finds an element with the passed `key` within the `flat_hash_map`.
-  using Base::find;
-
-  // flat_hash_map::operator[]()
-  //
-  // Returns a reference to the value mapped to the passed key within the
-  // `flat_hash_map`, performing an `insert()` if the key does not already
-  // exist.
-  //
-  // If an insertion occurs and results in a rehashing of the container, all
-  // iterators are invalidated. Otherwise iterators are not affected and
-  // references are not invalidated. Overloads are listed below.
-  //
-  // T& operator[](const Key& key):
-  //
-  //   Inserts an init_type object constructed in-place if the element with the
-  //   given key does not exist.
-  //
-  // T& operator[](Key&& key):
-  //
-  //   Inserts an init_type object constructed in-place provided that an element
-  //   with the given key does not exist.
-  using Base::operator[];
-
-  // flat_hash_map::bucket_count()
-  //
-  // Returns the number of "buckets" within the `flat_hash_map`. Note that
-  // because a flat hash map contains all elements within its internal storage,
-  // this value simply equals the current capacity of the `flat_hash_map`.
-  using Base::bucket_count;
-
-  // flat_hash_map::load_factor()
-  //
-  // Returns the current load factor of the `flat_hash_map` (the average number
-  // of slots occupied with a value within the hash map).
-  using Base::load_factor;
-
-  // flat_hash_map::max_load_factor()
-  //
-  // Manages the maximum load factor of the `flat_hash_map`. Overloads are
-  // listed below.
-  //
-  // float flat_hash_map::max_load_factor()
-  //
-  //   Returns the current maximum load factor of the `flat_hash_map`.
-  //
-  // void flat_hash_map::max_load_factor(float ml)
-  //
-  //   Sets the maximum load factor of the `flat_hash_map` to the passed value.
-  //
-  //   NOTE: This overload is provided only for API compatibility with the STL;
-  //   `flat_hash_map` will ignore any set load factor and manage its rehashing
-  //   internally as an implementation detail.
-  using Base::max_load_factor;
-
-  // flat_hash_map::get_allocator()
-  //
-  // Returns the allocator function associated with this `flat_hash_map`.
-  using Base::get_allocator;
-
-  // flat_hash_map::hash_function()
-  //
-  // Returns the hashing function used to hash the keys within this
-  // `flat_hash_map`.
-  using Base::hash_function;
-
-  // flat_hash_map::key_eq()
-  //
-  // Returns the function used for comparing keys equality.
-  using Base::key_eq;
-};
-
-// erase_if(flat_hash_map<>, Pred)
-//
-// Erases all elements that satisfy the predicate `pred` from the container `c`.
-// Returns the number of erased elements.
-template <typename K, typename V, typename H, typename E, typename A,
-          typename Predicate>
-typename flat_hash_map<K, V, H, E, A>::size_type erase_if(
-    flat_hash_map<K, V, H, E, A>& c, Predicate pred) {
-  return container_internal::EraseIf(pred, &c);
-}
-
-namespace container_internal {
-
-template <class K, class V>
-struct FlatHashMapPolicy {
-  using slot_policy = container_internal::map_slot_policy<K, V>;
-  using slot_type = typename slot_policy::slot_type;
-  using key_type = K;
-  using mapped_type = V;
-  using init_type = std::pair</*non const*/ key_type, mapped_type>;
-
-  template <class Allocator, class... Args>
-  static void construct(Allocator* alloc, slot_type* slot, Args&&... args) {
-    slot_policy::construct(alloc, slot, std::forward<Args>(args)...);
-  }
-
-  // Returns std::true_type in case destroy is trivial.
-  template <class Allocator>
-  static auto destroy(Allocator* alloc, slot_type* slot) {
-    return slot_policy::destroy(alloc, slot);
-  }
-
-  template <class Allocator>
-  static auto transfer(Allocator* alloc, slot_type* new_slot,
-                       slot_type* old_slot) {
-    return slot_policy::transfer(alloc, new_slot, old_slot);
-  }
-
-  template <class F, class... Args>
-  static decltype(absl::container_internal::DecomposePair(
-      std::declval<F>(), std::declval<Args>()...))
-  apply(F&& f, Args&&... args) {
-    return absl::container_internal::DecomposePair(std::forward<F>(f),
-                                                   std::forward<Args>(args)...);
-  }
-
-  template <class Hash>
-  static constexpr HashSlotFn get_hash_slot_fn() {
-    return memory_internal::IsLayoutCompatible<K, V>::value
-               ? &TypeErasedApplyToSlotFn<Hash, K>
-               : nullptr;
-  }
-
-  static size_t space_used(const slot_type*) { return 0; }
-
-  static std::pair<const K, V>& element(slot_type* slot) { return slot->value; }
-
-  static V& value(std::pair<const K, V>* kv) { return kv->second; }
-  static const V& value(const std::pair<const K, V>* kv) { return kv->second; }
-};
-
-}  // namespace container_internal
-
-namespace container_algorithm_internal {
-
-// Specialization of trait in absl/algorithm/container.h
-template <class Key, class T, class Hash, class KeyEqual, class Allocator>
-struct IsUnorderedContainer<
-    absl::flat_hash_map<Key, T, Hash, KeyEqual, Allocator>> : std::true_type {};
-
-}  // namespace container_algorithm_internal
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_FLAT_HASH_MAP_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/flat_hash_set.h b/projects/llm_framework/include/abseil-cpp/absl/container/flat_hash_set.h
deleted file mode 100644
index 5f72f954..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/flat_hash_set.h
+++ /dev/null
@@ -1,529 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: flat_hash_set.h
-// -----------------------------------------------------------------------------
-//
-// An `absl::flat_hash_set<T>` is an unordered associative container designed to
-// be a more efficient replacement for `std::unordered_set`. Like
-// `unordered_set`, search, insertion, and deletion of set elements can be done
-// as an `O(1)` operation. However, `flat_hash_set` (and other unordered
-// associative containers known as the collection of Abseil "Swiss tables")
-// contain other optimizations that result in both memory and computation
-// advantages.
-//
-// In most cases, your default choice for a hash set should be a set of type
-// `flat_hash_set`.
-#ifndef ABSL_CONTAINER_FLAT_HASH_SET_H_
-#define ABSL_CONTAINER_FLAT_HASH_SET_H_
-
-#include <cstddef>
-#include <memory>
-#include <type_traits>
-#include <utility>
-
-#include "absl/algorithm/container.h"
-#include "absl/base/macros.h"
-#include "absl/container/internal/container_memory.h"
-#include "absl/container/internal/hash_function_defaults.h"  // IWYU pragma: export
-#include "absl/container/internal/raw_hash_set.h"  // IWYU pragma: export
-#include "absl/memory/memory.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-template <typename T>
-struct FlatHashSetPolicy;
-}  // namespace container_internal
-
-// -----------------------------------------------------------------------------
-// absl::flat_hash_set
-// -----------------------------------------------------------------------------
-//
-// An `absl::flat_hash_set<T>` is an unordered associative container which has
-// been optimized for both speed and memory footprint in most common use cases.
-// Its interface is similar to that of `std::unordered_set<T>` with the
-// following notable differences:
-//
-// * Requires keys that are CopyConstructible
-// * Supports heterogeneous lookup, through `find()` and `insert()`, provided
-//   that the set is provided a compatible heterogeneous hashing function and
-//   equality operator. See below for details.
-// * Invalidates any references and pointers to elements within the table after
-//   `rehash()` and when the table is moved.
-// * Contains a `capacity()` member function indicating the number of element
-//   slots (open, deleted, and empty) within the hash set.
-// * Returns `void` from the `erase(iterator)` overload.
-//
-// By default, `flat_hash_set` uses the `absl::Hash` hashing framework. All
-// fundamental and Abseil types that support the `absl::Hash` framework have a
-// compatible equality operator for comparing insertions into `flat_hash_set`.
-// If your type is not yet supported by the `absl::Hash` framework, see
-// absl/hash/hash.h for information on extending Abseil hashing to user-defined
-// types.
-//
-// Using `absl::flat_hash_set` at interface boundaries in dynamically loaded
-// libraries (e.g. .dll, .so) is unsupported due to way `absl::Hash` values may
-// be randomized across dynamically loaded libraries.
-//
-// To achieve heterogeneous lookup for custom types either `Hash` and `Eq` type
-// parameters can be used or `T` should have public inner types
-// `absl_container_hash` and (optionally) `absl_container_eq`. In either case,
-// `typename Hash::is_transparent` and `typename Eq::is_transparent` should be
-// well-formed. Both types are basically functors:
-// * `Hash` should support `size_t operator()(U val) const` that returns a hash
-// for the given `val`.
-// * `Eq` should support `bool operator()(U lhs, V rhs) const` that returns true
-// if `lhs` is equal to `rhs`.
-//
-// In most cases `T` needs only to provide the `absl_container_hash`. In this
-// case `std::equal_to<void>` will be used instead of `eq` part.
-//
-// NOTE: A `flat_hash_set` stores its keys directly inside its implementation
-// array to avoid memory indirection. Because a `flat_hash_set` is designed to
-// move data when rehashed, set keys will not retain pointer stability. If you
-// require pointer stability, consider using
-// `absl::flat_hash_set<std::unique_ptr<T>>`. If your type is not moveable and
-// you require pointer stability, consider `absl::node_hash_set` instead.
-//
-// Example:
-//
-//   // Create a flat hash set of three strings
-//   absl::flat_hash_set<std::string> ducks =
-//     {"huey", "dewey", "louie"};
-//
-//  // Insert a new element into the flat hash set
-//  ducks.insert("donald");
-//
-//  // Force a rehash of the flat hash set
-//  ducks.rehash(0);
-//
-//  // See if "dewey" is present
-//  if (ducks.contains("dewey")) {
-//    std::cout << "We found dewey!" << std::endl;
-//  }
-template <class T, class Hash = absl::container_internal::hash_default_hash<T>,
-          class Eq = absl::container_internal::hash_default_eq<T>,
-          class Allocator = std::allocator<T>>
-class flat_hash_set
-    : public absl::container_internal::raw_hash_set<
-          absl::container_internal::FlatHashSetPolicy<T>, Hash, Eq, Allocator> {
-  using Base = typename flat_hash_set::raw_hash_set;
-
- public:
-  // Constructors and Assignment Operators
-  //
-  // A flat_hash_set supports the same overload set as `std::unordered_set`
-  // for construction and assignment:
-  //
-  // *  Default constructor
-  //
-  //    // No allocation for the table's elements is made.
-  //    absl::flat_hash_set<std::string> set1;
-  //
-  // * Initializer List constructor
-  //
-  //   absl::flat_hash_set<std::string> set2 =
-  //       {{"huey"}, {"dewey"}, {"louie"},};
-  //
-  // * Copy constructor
-  //
-  //   absl::flat_hash_set<std::string> set3(set2);
-  //
-  // * Copy assignment operator
-  //
-  //  // Hash functor and Comparator are copied as well
-  //  absl::flat_hash_set<std::string> set4;
-  //  set4 = set3;
-  //
-  // * Move constructor
-  //
-  //   // Move is guaranteed efficient
-  //   absl::flat_hash_set<std::string> set5(std::move(set4));
-  //
-  // * Move assignment operator
-  //
-  //   // May be efficient if allocators are compatible
-  //   absl::flat_hash_set<std::string> set6;
-  //   set6 = std::move(set5);
-  //
-  // * Range constructor
-  //
-  //   std::vector<std::string> v = {"a", "b"};
-  //   absl::flat_hash_set<std::string> set7(v.begin(), v.end());
-  flat_hash_set() {}
-  using Base::Base;
-
-  // flat_hash_set::begin()
-  //
-  // Returns an iterator to the beginning of the `flat_hash_set`.
-  using Base::begin;
-
-  // flat_hash_set::cbegin()
-  //
-  // Returns a const iterator to the beginning of the `flat_hash_set`.
-  using Base::cbegin;
-
-  // flat_hash_set::cend()
-  //
-  // Returns a const iterator to the end of the `flat_hash_set`.
-  using Base::cend;
-
-  // flat_hash_set::end()
-  //
-  // Returns an iterator to the end of the `flat_hash_set`.
-  using Base::end;
-
-  // flat_hash_set::capacity()
-  //
-  // Returns the number of element slots (assigned, deleted, and empty)
-  // available within the `flat_hash_set`.
-  //
-  // NOTE: this member function is particular to `absl::flat_hash_set` and is
-  // not provided in the `std::unordered_set` API.
-  using Base::capacity;
-
-  // flat_hash_set::empty()
-  //
-  // Returns whether or not the `flat_hash_set` is empty.
-  using Base::empty;
-
-  // flat_hash_set::max_size()
-  //
-  // Returns the largest theoretical possible number of elements within a
-  // `flat_hash_set` under current memory constraints. This value can be thought
-  // of the largest value of `std::distance(begin(), end())` for a
-  // `flat_hash_set<T>`.
-  using Base::max_size;
-
-  // flat_hash_set::size()
-  //
-  // Returns the number of elements currently within the `flat_hash_set`.
-  using Base::size;
-
-  // flat_hash_set::clear()
-  //
-  // Removes all elements from the `flat_hash_set`. Invalidates any references,
-  // pointers, or iterators referring to contained elements.
-  //
-  // NOTE: this operation may shrink the underlying buffer. To avoid shrinking
-  // the underlying buffer call `erase(begin(), end())`.
-  using Base::clear;
-
-  // flat_hash_set::erase()
-  //
-  // Erases elements within the `flat_hash_set`. Erasing does not trigger a
-  // rehash. Overloads are listed below.
-  //
-  // void erase(const_iterator pos):
-  //
-  //   Erases the element at `position` of the `flat_hash_set`, returning
-  //   `void`.
-  //
-  //   NOTE: returning `void` in this case is different than that of STL
-  //   containers in general and `std::unordered_set` in particular (which
-  //   return an iterator to the element following the erased element). If that
-  //   iterator is needed, simply post increment the iterator:
-  //
-  //     set.erase(it++);
-  //
-  // iterator erase(const_iterator first, const_iterator last):
-  //
-  //   Erases the elements in the open interval [`first`, `last`), returning an
-  //   iterator pointing to `last`. The special case of calling
-  //   `erase(begin(), end())` resets the reserved growth such that if
-  //   `reserve(N)` has previously been called and there has been no intervening
-  //   call to `clear()`, then after calling `erase(begin(), end())`, it is safe
-  //   to assume that inserting N elements will not cause a rehash.
-  //
-  // size_type erase(const key_type& key):
-  //
-  //   Erases the element with the matching key, if it exists, returning the
-  //   number of elements erased (0 or 1).
-  using Base::erase;
-
-  // flat_hash_set::insert()
-  //
-  // Inserts an element of the specified value into the `flat_hash_set`,
-  // returning an iterator pointing to the newly inserted element, provided that
-  // an element with the given key does not already exist. If rehashing occurs
-  // due to the insertion, all iterators are invalidated. Overloads are listed
-  // below.
-  //
-  // std::pair<iterator,bool> insert(const T& value):
-  //
-  //   Inserts a value into the `flat_hash_set`. Returns a pair consisting of an
-  //   iterator to the inserted element (or to the element that prevented the
-  //   insertion) and a bool denoting whether the insertion took place.
-  //
-  // std::pair<iterator,bool> insert(T&& value):
-  //
-  //   Inserts a moveable value into the `flat_hash_set`. Returns a pair
-  //   consisting of an iterator to the inserted element (or to the element that
-  //   prevented the insertion) and a bool denoting whether the insertion took
-  //   place.
-  //
-  // iterator insert(const_iterator hint, const T& value):
-  // iterator insert(const_iterator hint, T&& value):
-  //
-  //   Inserts a value, using the position of `hint` as a non-binding suggestion
-  //   for where to begin the insertion search. Returns an iterator to the
-  //   inserted element, or to the existing element that prevented the
-  //   insertion.
-  //
-  // void insert(InputIterator first, InputIterator last):
-  //
-  //   Inserts a range of values [`first`, `last`).
-  //
-  //   NOTE: Although the STL does not specify which element may be inserted if
-  //   multiple keys compare equivalently, for `flat_hash_set` we guarantee the
-  //   first match is inserted.
-  //
-  // void insert(std::initializer_list<T> ilist):
-  //
-  //   Inserts the elements within the initializer list `ilist`.
-  //
-  //   NOTE: Although the STL does not specify which element may be inserted if
-  //   multiple keys compare equivalently within the initializer list, for
-  //   `flat_hash_set` we guarantee the first match is inserted.
-  using Base::insert;
-
-  // flat_hash_set::emplace()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `flat_hash_set`, provided that no element with the given key
-  // already exists.
-  //
-  // The element may be constructed even if there already is an element with the
-  // key in the container, in which case the newly constructed element will be
-  // destroyed immediately.
-  //
-  // If rehashing occurs due to the insertion, all iterators are invalidated.
-  using Base::emplace;
-
-  // flat_hash_set::emplace_hint()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `flat_hash_set`, using the position of `hint` as a non-binding
-  // suggestion for where to begin the insertion search, and only inserts
-  // provided that no element with the given key already exists.
-  //
-  // The element may be constructed even if there already is an element with the
-  // key in the container, in which case the newly constructed element will be
-  // destroyed immediately.
-  //
-  // If rehashing occurs due to the insertion, all iterators are invalidated.
-  using Base::emplace_hint;
-
-  // flat_hash_set::extract()
-  //
-  // Extracts the indicated element, erasing it in the process, and returns it
-  // as a C++17-compatible node handle. Overloads are listed below.
-  //
-  // node_type extract(const_iterator position):
-  //
-  //   Extracts the element at the indicated position and returns a node handle
-  //   owning that extracted data.
-  //
-  // node_type extract(const key_type& x):
-  //
-  //   Extracts the element with the key matching the passed key value and
-  //   returns a node handle owning that extracted data. If the `flat_hash_set`
-  //   does not contain an element with a matching key, this function returns an
-  //   empty node handle.
-  using Base::extract;
-
-  // flat_hash_set::merge()
-  //
-  // Extracts elements from a given `source` flat hash set into this
-  // `flat_hash_set`. If the destination `flat_hash_set` already contains an
-  // element with an equivalent key, that element is not extracted.
-  using Base::merge;
-
-  // flat_hash_set::swap(flat_hash_set& other)
-  //
-  // Exchanges the contents of this `flat_hash_set` with those of the `other`
-  // flat hash set, avoiding invocation of any move, copy, or swap operations on
-  // individual elements.
-  //
-  // All iterators and references on the `flat_hash_set` remain valid, excepting
-  // for the past-the-end iterator, which is invalidated.
-  //
-  // `swap()` requires that the flat hash set's hashing and key equivalence
-  // functions be Swappable, and are exchanged using unqualified calls to
-  // non-member `swap()`. If the set's allocator has
-  // `std::allocator_traits<allocator_type>::propagate_on_container_swap::value`
-  // set to `true`, the allocators are also exchanged using an unqualified call
-  // to non-member `swap()`; otherwise, the allocators are not swapped.
-  using Base::swap;
-
-  // flat_hash_set::rehash(count)
-  //
-  // Rehashes the `flat_hash_set`, setting the number of slots to be at least
-  // the passed value. If the new number of slots increases the load factor more
-  // than the current maximum load factor
-  // (`count` < `size()` / `max_load_factor()`), then the new number of slots
-  // will be at least `size()` / `max_load_factor()`.
-  //
-  // To force a rehash, pass rehash(0).
-  //
-  // NOTE: unlike behavior in `std::unordered_set`, references are also
-  // invalidated upon a `rehash()`.
-  using Base::rehash;
-
-  // flat_hash_set::reserve(count)
-  //
-  // Sets the number of slots in the `flat_hash_set` to the number needed to
-  // accommodate at least `count` total elements without exceeding the current
-  // maximum load factor, and may rehash the container if needed.
-  using Base::reserve;
-
-  // flat_hash_set::contains()
-  //
-  // Determines whether an element comparing equal to the given `key` exists
-  // within the `flat_hash_set`, returning `true` if so or `false` otherwise.
-  using Base::contains;
-
-  // flat_hash_set::count(const Key& key) const
-  //
-  // Returns the number of elements comparing equal to the given `key` within
-  // the `flat_hash_set`. note that this function will return either `1` or `0`
-  // since duplicate elements are not allowed within a `flat_hash_set`.
-  using Base::count;
-
-  // flat_hash_set::equal_range()
-  //
-  // Returns a closed range [first, last], defined by a `std::pair` of two
-  // iterators, containing all elements with the passed key in the
-  // `flat_hash_set`.
-  using Base::equal_range;
-
-  // flat_hash_set::find()
-  //
-  // Finds an element with the passed `key` within the `flat_hash_set`.
-  using Base::find;
-
-  // flat_hash_set::bucket_count()
-  //
-  // Returns the number of "buckets" within the `flat_hash_set`. Note that
-  // because a flat hash set contains all elements within its internal storage,
-  // this value simply equals the current capacity of the `flat_hash_set`.
-  using Base::bucket_count;
-
-  // flat_hash_set::load_factor()
-  //
-  // Returns the current load factor of the `flat_hash_set` (the average number
-  // of slots occupied with a value within the hash set).
-  using Base::load_factor;
-
-  // flat_hash_set::max_load_factor()
-  //
-  // Manages the maximum load factor of the `flat_hash_set`. Overloads are
-  // listed below.
-  //
-  // float flat_hash_set::max_load_factor()
-  //
-  //   Returns the current maximum load factor of the `flat_hash_set`.
-  //
-  // void flat_hash_set::max_load_factor(float ml)
-  //
-  //   Sets the maximum load factor of the `flat_hash_set` to the passed value.
-  //
-  //   NOTE: This overload is provided only for API compatibility with the STL;
-  //   `flat_hash_set` will ignore any set load factor and manage its rehashing
-  //   internally as an implementation detail.
-  using Base::max_load_factor;
-
-  // flat_hash_set::get_allocator()
-  //
-  // Returns the allocator function associated with this `flat_hash_set`.
-  using Base::get_allocator;
-
-  // flat_hash_set::hash_function()
-  //
-  // Returns the hashing function used to hash the keys within this
-  // `flat_hash_set`.
-  using Base::hash_function;
-
-  // flat_hash_set::key_eq()
-  //
-  // Returns the function used for comparing keys equality.
-  using Base::key_eq;
-};
-
-// erase_if(flat_hash_set<>, Pred)
-//
-// Erases all elements that satisfy the predicate `pred` from the container `c`.
-// Returns the number of erased elements.
-template <typename T, typename H, typename E, typename A, typename Predicate>
-typename flat_hash_set<T, H, E, A>::size_type erase_if(
-    flat_hash_set<T, H, E, A>& c, Predicate pred) {
-  return container_internal::EraseIf(pred, &c);
-}
-
-namespace container_internal {
-
-template <class T>
-struct FlatHashSetPolicy {
-  using slot_type = T;
-  using key_type = T;
-  using init_type = T;
-  using constant_iterators = std::true_type;
-
-  template <class Allocator, class... Args>
-  static void construct(Allocator* alloc, slot_type* slot, Args&&... args) {
-    absl::allocator_traits<Allocator>::construct(*alloc, slot,
-                                                 std::forward<Args>(args)...);
-  }
-
-  // Return std::true_type in case destroy is trivial.
-  template <class Allocator>
-  static auto destroy(Allocator* alloc, slot_type* slot) {
-    absl::allocator_traits<Allocator>::destroy(*alloc, slot);
-    return IsDestructionTrivial<Allocator, slot_type>();
-  }
-
-  static T& element(slot_type* slot) { return *slot; }
-
-  template <class F, class... Args>
-  static decltype(absl::container_internal::DecomposeValue(
-      std::declval<F>(), std::declval<Args>()...))
-  apply(F&& f, Args&&... args) {
-    return absl::container_internal::DecomposeValue(
-        std::forward<F>(f), std::forward<Args>(args)...);
-  }
-
-  static size_t space_used(const T*) { return 0; }
-
-  template <class Hash>
-  static constexpr HashSlotFn get_hash_slot_fn() {
-    return &TypeErasedApplyToSlotFn<Hash, T>;
-  }
-};
-}  // namespace container_internal
-
-namespace container_algorithm_internal {
-
-// Specialization of trait in absl/algorithm/container.h
-template <class Key, class Hash, class KeyEqual, class Allocator>
-struct IsUnorderedContainer<absl::flat_hash_set<Key, Hash, KeyEqual, Allocator>>
-    : std::true_type {};
-
-}  // namespace container_algorithm_internal
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_FLAT_HASH_SET_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/inlined_vector.h b/projects/llm_framework/include/abseil-cpp/absl/container/inlined_vector.h
deleted file mode 100644
index 974b6521..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/inlined_vector.h
+++ /dev/null
@@ -1,1015 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: inlined_vector.h
-// -----------------------------------------------------------------------------
-//
-// This header file contains the declaration and definition of an "inlined
-// vector" which behaves in an equivalent fashion to a `std::vector`, except
-// that storage for small sequences of the vector are provided inline without
-// requiring any heap allocation.
-//
-// An `absl::InlinedVector<T, N>` specifies the default capacity `N` as one of
-// its template parameters. Instances where `size() <= N` hold contained
-// elements in inline space. Typically `N` is very small so that sequences that
-// are expected to be short do not require allocations.
-//
-// An `absl::InlinedVector` does not usually require a specific allocator. If
-// the inlined vector grows beyond its initial constraints, it will need to
-// allocate (as any normal `std::vector` would). This is usually performed with
-// the default allocator (defined as `std::allocator<T>`). Optionally, a custom
-// allocator type may be specified as `A` in `absl::InlinedVector<T, N, A>`.
-
-#ifndef ABSL_CONTAINER_INLINED_VECTOR_H_
-#define ABSL_CONTAINER_INLINED_VECTOR_H_
-
-#include <algorithm>
-#include <cstddef>
-#include <cstdlib>
-#include <cstring>
-#include <initializer_list>
-#include <iterator>
-#include <memory>
-#include <type_traits>
-#include <utility>
-
-#include "absl/algorithm/algorithm.h"
-#include "absl/base/internal/throw_delegate.h"
-#include "absl/base/macros.h"
-#include "absl/base/optimization.h"
-#include "absl/base/port.h"
-#include "absl/container/internal/inlined_vector.h"
-#include "absl/memory/memory.h"
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-// -----------------------------------------------------------------------------
-// InlinedVector
-// -----------------------------------------------------------------------------
-//
-// An `absl::InlinedVector` is designed to be a drop-in replacement for
-// `std::vector` for use cases where the vector's size is sufficiently small
-// that it can be inlined. If the inlined vector does grow beyond its estimated
-// capacity, it will trigger an initial allocation on the heap, and will behave
-// as a `std::vector`. The API of the `absl::InlinedVector` within this file is
-// designed to cover the same API footprint as covered by `std::vector`.
-template <typename T, size_t N, typename A = std::allocator<T>>
-class InlinedVector {
-  static_assert(N > 0, "`absl::InlinedVector` requires an inlined capacity.");
-
-  using Storage = inlined_vector_internal::Storage<T, N, A>;
-
-  template <typename TheA>
-  using AllocatorTraits = inlined_vector_internal::AllocatorTraits<TheA>;
-  template <typename TheA>
-  using MoveIterator = inlined_vector_internal::MoveIterator<TheA>;
-  template <typename TheA>
-  using IsMoveAssignOk = inlined_vector_internal::IsMoveAssignOk<TheA>;
-
-  template <typename TheA, typename Iterator>
-  using IteratorValueAdapter =
-      inlined_vector_internal::IteratorValueAdapter<TheA, Iterator>;
-  template <typename TheA>
-  using CopyValueAdapter = inlined_vector_internal::CopyValueAdapter<TheA>;
-  template <typename TheA>
-  using DefaultValueAdapter =
-      inlined_vector_internal::DefaultValueAdapter<TheA>;
-
-  template <typename Iterator>
-  using EnableIfAtLeastForwardIterator = absl::enable_if_t<
-      inlined_vector_internal::IsAtLeastForwardIterator<Iterator>::value, int>;
-  template <typename Iterator>
-  using DisableIfAtLeastForwardIterator = absl::enable_if_t<
-      !inlined_vector_internal::IsAtLeastForwardIterator<Iterator>::value, int>;
-
-  using MemcpyPolicy = typename Storage::MemcpyPolicy;
-  using ElementwiseAssignPolicy = typename Storage::ElementwiseAssignPolicy;
-  using ElementwiseConstructPolicy =
-      typename Storage::ElementwiseConstructPolicy;
-  using MoveAssignmentPolicy = typename Storage::MoveAssignmentPolicy;
-
- public:
-  using allocator_type = A;
-  using value_type = inlined_vector_internal::ValueType<A>;
-  using pointer = inlined_vector_internal::Pointer<A>;
-  using const_pointer = inlined_vector_internal::ConstPointer<A>;
-  using size_type = inlined_vector_internal::SizeType<A>;
-  using difference_type = inlined_vector_internal::DifferenceType<A>;
-  using reference = inlined_vector_internal::Reference<A>;
-  using const_reference = inlined_vector_internal::ConstReference<A>;
-  using iterator = inlined_vector_internal::Iterator<A>;
-  using const_iterator = inlined_vector_internal::ConstIterator<A>;
-  using reverse_iterator = inlined_vector_internal::ReverseIterator<A>;
-  using const_reverse_iterator =
-      inlined_vector_internal::ConstReverseIterator<A>;
-
-  // ---------------------------------------------------------------------------
-  // InlinedVector Constructors and Destructor
-  // ---------------------------------------------------------------------------
-
-  // Creates an empty inlined vector with a value-initialized allocator.
-  InlinedVector() noexcept(noexcept(allocator_type())) : storage_() {}
-
-  // Creates an empty inlined vector with a copy of `allocator`.
-  explicit InlinedVector(const allocator_type& allocator) noexcept
-      : storage_(allocator) {}
-
-  // Creates an inlined vector with `n` copies of `value_type()`.
-  explicit InlinedVector(size_type n,
-                         const allocator_type& allocator = allocator_type())
-      : storage_(allocator) {
-    storage_.Initialize(DefaultValueAdapter<A>(), n);
-  }
-
-  // Creates an inlined vector with `n` copies of `v`.
-  InlinedVector(size_type n, const_reference v,
-                const allocator_type& allocator = allocator_type())
-      : storage_(allocator) {
-    storage_.Initialize(CopyValueAdapter<A>(std::addressof(v)), n);
-  }
-
-  // Creates an inlined vector with copies of the elements of `list`.
-  InlinedVector(std::initializer_list<value_type> list,
-                const allocator_type& allocator = allocator_type())
-      : InlinedVector(list.begin(), list.end(), allocator) {}
-
-  // Creates an inlined vector with elements constructed from the provided
-  // forward iterator range [`first`, `last`).
-  //
-  // NOTE: the `enable_if` prevents ambiguous interpretation between a call to
-  // this constructor with two integral arguments and a call to the above
-  // `InlinedVector(size_type, const_reference)` constructor.
-  template <typename ForwardIterator,
-            EnableIfAtLeastForwardIterator<ForwardIterator> = 0>
-  InlinedVector(ForwardIterator first, ForwardIterator last,
-                const allocator_type& allocator = allocator_type())
-      : storage_(allocator) {
-    storage_.Initialize(IteratorValueAdapter<A, ForwardIterator>(first),
-                        static_cast<size_t>(std::distance(first, last)));
-  }
-
-  // Creates an inlined vector with elements constructed from the provided input
-  // iterator range [`first`, `last`).
-  template <typename InputIterator,
-            DisableIfAtLeastForwardIterator<InputIterator> = 0>
-  InlinedVector(InputIterator first, InputIterator last,
-                const allocator_type& allocator = allocator_type())
-      : storage_(allocator) {
-    std::copy(first, last, std::back_inserter(*this));
-  }
-
-  // Creates an inlined vector by copying the contents of `other` using
-  // `other`'s allocator.
-  InlinedVector(const InlinedVector& other)
-      : InlinedVector(other, other.storage_.GetAllocator()) {}
-
-  // Creates an inlined vector by copying the contents of `other` using the
-  // provided `allocator`.
-  InlinedVector(const InlinedVector& other, const allocator_type& allocator)
-      : storage_(allocator) {
-    // Fast path: if the other vector is empty, there's nothing for us to do.
-    if (other.empty()) {
-      return;
-    }
-
-    // Fast path: if the value type is trivially copy constructible, we know the
-    // allocator doesn't do anything fancy, and there is nothing on the heap
-    // then we know it is legal for us to simply memcpy the other vector's
-    // inlined bytes to form our copy of its elements.
-    if (absl::is_trivially_copy_constructible<value_type>::value &&
-        std::is_same<A, std::allocator<value_type>>::value &&
-        !other.storage_.GetIsAllocated()) {
-      storage_.MemcpyFrom(other.storage_);
-      return;
-    }
-
-    storage_.InitFrom(other.storage_);
-  }
-
-  // Creates an inlined vector by moving in the contents of `other` without
-  // allocating. If `other` contains allocated memory, the newly-created inlined
-  // vector will take ownership of that memory. However, if `other` does not
-  // contain allocated memory, the newly-created inlined vector will perform
-  // element-wise move construction of the contents of `other`.
-  //
-  // NOTE: since no allocation is performed for the inlined vector in either
-  // case, the `noexcept(...)` specification depends on whether moving the
-  // underlying objects can throw. It is assumed assumed that...
-  //  a) move constructors should only throw due to allocation failure.
-  //  b) if `value_type`'s move constructor allocates, it uses the same
-  //     allocation function as the inlined vector's allocator.
-  // Thus, the move constructor is non-throwing if the allocator is non-throwing
-  // or `value_type`'s move constructor is specified as `noexcept`.
-  InlinedVector(InlinedVector&& other) noexcept(
-      absl::allocator_is_nothrow<allocator_type>::value ||
-      std::is_nothrow_move_constructible<value_type>::value)
-      : storage_(other.storage_.GetAllocator()) {
-    // Fast path: if the value type can be trivially relocated (i.e. moved from
-    // and destroyed), and we know the allocator doesn't do anything fancy, then
-    // it's safe for us to simply adopt the contents of the storage for `other`
-    // and remove its own reference to them. It's as if we had individually
-    // move-constructed each value and then destroyed the original.
-    if (absl::is_trivially_relocatable<value_type>::value &&
-        std::is_same<A, std::allocator<value_type>>::value) {
-      storage_.MemcpyFrom(other.storage_);
-      other.storage_.SetInlinedSize(0);
-      return;
-    }
-
-    // Fast path: if the other vector is on the heap, we can simply take over
-    // its allocation.
-    if (other.storage_.GetIsAllocated()) {
-      storage_.SetAllocation({other.storage_.GetAllocatedData(),
-                              other.storage_.GetAllocatedCapacity()});
-      storage_.SetAllocatedSize(other.storage_.GetSize());
-
-      other.storage_.SetInlinedSize(0);
-      return;
-    }
-
-    // Otherwise we must move each element individually.
-    IteratorValueAdapter<A, MoveIterator<A>> other_values(
-        MoveIterator<A>(other.storage_.GetInlinedData()));
-
-    inlined_vector_internal::ConstructElements<A>(
-        storage_.GetAllocator(), storage_.GetInlinedData(), other_values,
-        other.storage_.GetSize());
-
-    storage_.SetInlinedSize(other.storage_.GetSize());
-  }
-
-  // Creates an inlined vector by moving in the contents of `other` with a copy
-  // of `allocator`.
-  //
-  // NOTE: if `other`'s allocator is not equal to `allocator`, even if `other`
-  // contains allocated memory, this move constructor will still allocate. Since
-  // allocation is performed, this constructor can only be `noexcept` if the
-  // specified allocator is also `noexcept`.
-  InlinedVector(
-      InlinedVector&& other,
-      const allocator_type&
-          allocator) noexcept(absl::allocator_is_nothrow<allocator_type>::value)
-      : storage_(allocator) {
-    // Fast path: if the value type can be trivially relocated (i.e. moved from
-    // and destroyed), and we know the allocator doesn't do anything fancy, then
-    // it's safe for us to simply adopt the contents of the storage for `other`
-    // and remove its own reference to them. It's as if we had individually
-    // move-constructed each value and then destroyed the original.
-    if (absl::is_trivially_relocatable<value_type>::value &&
-        std::is_same<A, std::allocator<value_type>>::value) {
-      storage_.MemcpyFrom(other.storage_);
-      other.storage_.SetInlinedSize(0);
-      return;
-    }
-
-    // Fast path: if the other vector is on the heap and shared the same
-    // allocator, we can simply take over its allocation.
-    if ((storage_.GetAllocator() == other.storage_.GetAllocator()) &&
-        other.storage_.GetIsAllocated()) {
-      storage_.SetAllocation({other.storage_.GetAllocatedData(),
-                              other.storage_.GetAllocatedCapacity()});
-      storage_.SetAllocatedSize(other.storage_.GetSize());
-
-      other.storage_.SetInlinedSize(0);
-      return;
-    }
-
-    // Otherwise we must move each element individually.
-    storage_.Initialize(
-        IteratorValueAdapter<A, MoveIterator<A>>(MoveIterator<A>(other.data())),
-        other.size());
-  }
-
-  ~InlinedVector() {}
-
-  // ---------------------------------------------------------------------------
-  // InlinedVector Member Accessors
-  // ---------------------------------------------------------------------------
-
-  // `InlinedVector::empty()`
-  //
-  // Returns whether the inlined vector contains no elements.
-  bool empty() const noexcept { return !size(); }
-
-  // `InlinedVector::size()`
-  //
-  // Returns the number of elements in the inlined vector.
-  size_type size() const noexcept { return storage_.GetSize(); }
-
-  // `InlinedVector::max_size()`
-  //
-  // Returns the maximum number of elements the inlined vector can hold.
-  size_type max_size() const noexcept {
-    // One bit of the size storage is used to indicate whether the inlined
-    // vector contains allocated memory. As a result, the maximum size that the
-    // inlined vector can express is the minimum of the limit of how many
-    // objects we can allocate and std::numeric_limits<size_type>::max() / 2.
-    return (std::min)(AllocatorTraits<A>::max_size(storage_.GetAllocator()),
-                      (std::numeric_limits<size_type>::max)() / 2);
-  }
-
-  // `InlinedVector::capacity()`
-  //
-  // Returns the number of elements that could be stored in the inlined vector
-  // without requiring a reallocation.
-  //
-  // NOTE: for most inlined vectors, `capacity()` should be equal to the
-  // template parameter `N`. For inlined vectors which exceed this capacity,
-  // they will no longer be inlined and `capacity()` will equal the capactity of
-  // the allocated memory.
-  size_type capacity() const noexcept {
-    return storage_.GetIsAllocated() ? storage_.GetAllocatedCapacity()
-                                     : storage_.GetInlinedCapacity();
-  }
-
-  // `InlinedVector::data()`
-  //
-  // Returns a `pointer` to the elements of the inlined vector. This pointer
-  // can be used to access and modify the contained elements.
-  //
-  // NOTE: only elements within [`data()`, `data() + size()`) are valid.
-  pointer data() noexcept ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return storage_.GetIsAllocated() ? storage_.GetAllocatedData()
-                                     : storage_.GetInlinedData();
-  }
-
-  // Overload of `InlinedVector::data()` that returns a `const_pointer` to the
-  // elements of the inlined vector. This pointer can be used to access but not
-  // modify the contained elements.
-  //
-  // NOTE: only elements within [`data()`, `data() + size()`) are valid.
-  const_pointer data() const noexcept ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return storage_.GetIsAllocated() ? storage_.GetAllocatedData()
-                                     : storage_.GetInlinedData();
-  }
-
-  // `InlinedVector::operator[](...)`
-  //
-  // Returns a `reference` to the `i`th element of the inlined vector.
-  reference operator[](size_type i) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(i < size());
-    return data()[i];
-  }
-
-  // Overload of `InlinedVector::operator[](...)` that returns a
-  // `const_reference` to the `i`th element of the inlined vector.
-  const_reference operator[](size_type i) const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(i < size());
-    return data()[i];
-  }
-
-  // `InlinedVector::at(...)`
-  //
-  // Returns a `reference` to the `i`th element of the inlined vector.
-  //
-  // NOTE: if `i` is not within the required range of `InlinedVector::at(...)`,
-  // in both debug and non-debug builds, `std::out_of_range` will be thrown.
-  reference at(size_type i) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    if (ABSL_PREDICT_FALSE(i >= size())) {
-      base_internal::ThrowStdOutOfRange(
-          "`InlinedVector::at(size_type)` failed bounds check");
-    }
-    return data()[i];
-  }
-
-  // Overload of `InlinedVector::at(...)` that returns a `const_reference` to
-  // the `i`th element of the inlined vector.
-  //
-  // NOTE: if `i` is not within the required range of `InlinedVector::at(...)`,
-  // in both debug and non-debug builds, `std::out_of_range` will be thrown.
-  const_reference at(size_type i) const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    if (ABSL_PREDICT_FALSE(i >= size())) {
-      base_internal::ThrowStdOutOfRange(
-          "`InlinedVector::at(size_type) const` failed bounds check");
-    }
-    return data()[i];
-  }
-
-  // `InlinedVector::front()`
-  //
-  // Returns a `reference` to the first element of the inlined vector.
-  reference front() ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(!empty());
-    return data()[0];
-  }
-
-  // Overload of `InlinedVector::front()` that returns a `const_reference` to
-  // the first element of the inlined vector.
-  const_reference front() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(!empty());
-    return data()[0];
-  }
-
-  // `InlinedVector::back()`
-  //
-  // Returns a `reference` to the last element of the inlined vector.
-  reference back() ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(!empty());
-    return data()[size() - 1];
-  }
-
-  // Overload of `InlinedVector::back()` that returns a `const_reference` to the
-  // last element of the inlined vector.
-  const_reference back() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(!empty());
-    return data()[size() - 1];
-  }
-
-  // `InlinedVector::begin()`
-  //
-  // Returns an `iterator` to the beginning of the inlined vector.
-  iterator begin() noexcept ABSL_ATTRIBUTE_LIFETIME_BOUND { return data(); }
-
-  // Overload of `InlinedVector::begin()` that returns a `const_iterator` to
-  // the beginning of the inlined vector.
-  const_iterator begin() const noexcept ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return data();
-  }
-
-  // `InlinedVector::end()`
-  //
-  // Returns an `iterator` to the end of the inlined vector.
-  iterator end() noexcept ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return data() + size();
-  }
-
-  // Overload of `InlinedVector::end()` that returns a `const_iterator` to the
-  // end of the inlined vector.
-  const_iterator end() const noexcept ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return data() + size();
-  }
-
-  // `InlinedVector::cbegin()`
-  //
-  // Returns a `const_iterator` to the beginning of the inlined vector.
-  const_iterator cbegin() const noexcept ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return begin();
-  }
-
-  // `InlinedVector::cend()`
-  //
-  // Returns a `const_iterator` to the end of the inlined vector.
-  const_iterator cend() const noexcept ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return end();
-  }
-
-  // `InlinedVector::rbegin()`
-  //
-  // Returns a `reverse_iterator` from the end of the inlined vector.
-  reverse_iterator rbegin() noexcept ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return reverse_iterator(end());
-  }
-
-  // Overload of `InlinedVector::rbegin()` that returns a
-  // `const_reverse_iterator` from the end of the inlined vector.
-  const_reverse_iterator rbegin() const noexcept ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return const_reverse_iterator(end());
-  }
-
-  // `InlinedVector::rend()`
-  //
-  // Returns a `reverse_iterator` from the beginning of the inlined vector.
-  reverse_iterator rend() noexcept ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return reverse_iterator(begin());
-  }
-
-  // Overload of `InlinedVector::rend()` that returns a `const_reverse_iterator`
-  // from the beginning of the inlined vector.
-  const_reverse_iterator rend() const noexcept ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return const_reverse_iterator(begin());
-  }
-
-  // `InlinedVector::crbegin()`
-  //
-  // Returns a `const_reverse_iterator` from the end of the inlined vector.
-  const_reverse_iterator crbegin() const noexcept
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return rbegin();
-  }
-
-  // `InlinedVector::crend()`
-  //
-  // Returns a `const_reverse_iterator` from the beginning of the inlined
-  // vector.
-  const_reverse_iterator crend() const noexcept ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return rend();
-  }
-
-  // `InlinedVector::get_allocator()`
-  //
-  // Returns a copy of the inlined vector's allocator.
-  allocator_type get_allocator() const { return storage_.GetAllocator(); }
-
-  // ---------------------------------------------------------------------------
-  // InlinedVector Member Mutators
-  // ---------------------------------------------------------------------------
-
-  // `InlinedVector::operator=(...)`
-  //
-  // Replaces the elements of the inlined vector with copies of the elements of
-  // `list`.
-  InlinedVector& operator=(std::initializer_list<value_type> list) {
-    assign(list.begin(), list.end());
-
-    return *this;
-  }
-
-  // Overload of `InlinedVector::operator=(...)` that replaces the elements of
-  // the inlined vector with copies of the elements of `other`.
-  InlinedVector& operator=(const InlinedVector& other) {
-    if (ABSL_PREDICT_TRUE(this != std::addressof(other))) {
-      const_pointer other_data = other.data();
-      assign(other_data, other_data + other.size());
-    }
-
-    return *this;
-  }
-
-  // Overload of `InlinedVector::operator=(...)` that moves the elements of
-  // `other` into the inlined vector.
-  //
-  // NOTE: as a result of calling this overload, `other` is left in a valid but
-  // unspecified state.
-  InlinedVector& operator=(InlinedVector&& other) {
-    if (ABSL_PREDICT_TRUE(this != std::addressof(other))) {
-      MoveAssignment(MoveAssignmentPolicy{}, std::move(other));
-    }
-
-    return *this;
-  }
-
-  // `InlinedVector::assign(...)`
-  //
-  // Replaces the contents of the inlined vector with `n` copies of `v`.
-  void assign(size_type n, const_reference v) {
-    storage_.Assign(CopyValueAdapter<A>(std::addressof(v)), n);
-  }
-
-  // Overload of `InlinedVector::assign(...)` that replaces the contents of the
-  // inlined vector with copies of the elements of `list`.
-  void assign(std::initializer_list<value_type> list) {
-    assign(list.begin(), list.end());
-  }
-
-  // Overload of `InlinedVector::assign(...)` to replace the contents of the
-  // inlined vector with the range [`first`, `last`).
-  //
-  // NOTE: this overload is for iterators that are "forward" category or better.
-  template <typename ForwardIterator,
-            EnableIfAtLeastForwardIterator<ForwardIterator> = 0>
-  void assign(ForwardIterator first, ForwardIterator last) {
-    storage_.Assign(IteratorValueAdapter<A, ForwardIterator>(first),
-                    static_cast<size_t>(std::distance(first, last)));
-  }
-
-  // Overload of `InlinedVector::assign(...)` to replace the contents of the
-  // inlined vector with the range [`first`, `last`).
-  //
-  // NOTE: this overload is for iterators that are "input" category.
-  template <typename InputIterator,
-            DisableIfAtLeastForwardIterator<InputIterator> = 0>
-  void assign(InputIterator first, InputIterator last) {
-    size_type i = 0;
-    for (; i < size() && first != last; ++i, static_cast<void>(++first)) {
-      data()[i] = *first;
-    }
-
-    erase(data() + i, data() + size());
-    std::copy(first, last, std::back_inserter(*this));
-  }
-
-  // `InlinedVector::resize(...)`
-  //
-  // Resizes the inlined vector to contain `n` elements.
-  //
-  // NOTE: If `n` is smaller than `size()`, extra elements are destroyed. If `n`
-  // is larger than `size()`, new elements are value-initialized.
-  void resize(size_type n) {
-    ABSL_HARDENING_ASSERT(n <= max_size());
-    storage_.Resize(DefaultValueAdapter<A>(), n);
-  }
-
-  // Overload of `InlinedVector::resize(...)` that resizes the inlined vector to
-  // contain `n` elements.
-  //
-  // NOTE: if `n` is smaller than `size()`, extra elements are destroyed. If `n`
-  // is larger than `size()`, new elements are copied-constructed from `v`.
-  void resize(size_type n, const_reference v) {
-    ABSL_HARDENING_ASSERT(n <= max_size());
-    storage_.Resize(CopyValueAdapter<A>(std::addressof(v)), n);
-  }
-
-  // `InlinedVector::insert(...)`
-  //
-  // Inserts a copy of `v` at `pos`, returning an `iterator` to the newly
-  // inserted element.
-  iterator insert(const_iterator pos,
-                  const_reference v) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return emplace(pos, v);
-  }
-
-  // Overload of `InlinedVector::insert(...)` that inserts `v` at `pos` using
-  // move semantics, returning an `iterator` to the newly inserted element.
-  iterator insert(const_iterator pos,
-                  value_type&& v) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return emplace(pos, std::move(v));
-  }
-
-  // Overload of `InlinedVector::insert(...)` that inserts `n` contiguous copies
-  // of `v` starting at `pos`, returning an `iterator` pointing to the first of
-  // the newly inserted elements.
-  iterator insert(const_iterator pos, size_type n,
-                  const_reference v) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(pos >= begin());
-    ABSL_HARDENING_ASSERT(pos <= end());
-
-    if (ABSL_PREDICT_TRUE(n != 0)) {
-      value_type dealias = v;
-      // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102329#c2
-      // It appears that GCC thinks that since `pos` is a const pointer and may
-      // point to uninitialized memory at this point, a warning should be
-      // issued. But `pos` is actually only used to compute an array index to
-      // write to.
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
-#endif
-      return storage_.Insert(pos, CopyValueAdapter<A>(std::addressof(dealias)),
-                             n);
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic pop
-#endif
-    } else {
-      return const_cast<iterator>(pos);
-    }
-  }
-
-  // Overload of `InlinedVector::insert(...)` that inserts copies of the
-  // elements of `list` starting at `pos`, returning an `iterator` pointing to
-  // the first of the newly inserted elements.
-  iterator insert(const_iterator pos, std::initializer_list<value_type> list)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert(pos, list.begin(), list.end());
-  }
-
-  // Overload of `InlinedVector::insert(...)` that inserts the range [`first`,
-  // `last`) starting at `pos`, returning an `iterator` pointing to the first
-  // of the newly inserted elements.
-  //
-  // NOTE: this overload is for iterators that are "forward" category or better.
-  template <typename ForwardIterator,
-            EnableIfAtLeastForwardIterator<ForwardIterator> = 0>
-  iterator insert(const_iterator pos, ForwardIterator first,
-                  ForwardIterator last) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(pos >= begin());
-    ABSL_HARDENING_ASSERT(pos <= end());
-
-    if (ABSL_PREDICT_TRUE(first != last)) {
-      return storage_.Insert(
-          pos, IteratorValueAdapter<A, ForwardIterator>(first),
-          static_cast<size_type>(std::distance(first, last)));
-    } else {
-      return const_cast<iterator>(pos);
-    }
-  }
-
-  // Overload of `InlinedVector::insert(...)` that inserts the range [`first`,
-  // `last`) starting at `pos`, returning an `iterator` pointing to the first
-  // of the newly inserted elements.
-  //
-  // NOTE: this overload is for iterators that are "input" category.
-  template <typename InputIterator,
-            DisableIfAtLeastForwardIterator<InputIterator> = 0>
-  iterator insert(const_iterator pos, InputIterator first,
-                  InputIterator last) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(pos >= begin());
-    ABSL_HARDENING_ASSERT(pos <= end());
-
-    size_type index = static_cast<size_type>(std::distance(cbegin(), pos));
-    for (size_type i = index; first != last; ++i, static_cast<void>(++first)) {
-      insert(data() + i, *first);
-    }
-
-    return iterator(data() + index);
-  }
-
-  // `InlinedVector::emplace(...)`
-  //
-  // Constructs and inserts an element using `args...` in the inlined vector at
-  // `pos`, returning an `iterator` pointing to the newly emplaced element.
-  template <typename... Args>
-  iterator emplace(const_iterator pos,
-                   Args&&... args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(pos >= begin());
-    ABSL_HARDENING_ASSERT(pos <= end());
-
-    value_type dealias(std::forward<Args>(args)...);
-    // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102329#c2
-    // It appears that GCC thinks that since `pos` is a const pointer and may
-    // point to uninitialized memory at this point, a warning should be
-    // issued. But `pos` is actually only used to compute an array index to
-    // write to.
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
-#endif
-    return storage_.Insert(pos,
-                           IteratorValueAdapter<A, MoveIterator<A>>(
-                               MoveIterator<A>(std::addressof(dealias))),
-                           1);
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic pop
-#endif
-  }
-
-  // `InlinedVector::emplace_back(...)`
-  //
-  // Constructs and inserts an element using `args...` in the inlined vector at
-  // `end()`, returning a `reference` to the newly emplaced element.
-  template <typename... Args>
-  reference emplace_back(Args&&... args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return storage_.EmplaceBack(std::forward<Args>(args)...);
-  }
-
-  // `InlinedVector::push_back(...)`
-  //
-  // Inserts a copy of `v` in the inlined vector at `end()`.
-  void push_back(const_reference v) { static_cast<void>(emplace_back(v)); }
-
-  // Overload of `InlinedVector::push_back(...)` for inserting `v` at `end()`
-  // using move semantics.
-  void push_back(value_type&& v) {
-    static_cast<void>(emplace_back(std::move(v)));
-  }
-
-  // `InlinedVector::pop_back()`
-  //
-  // Destroys the element at `back()`, reducing the size by `1`.
-  void pop_back() noexcept {
-    ABSL_HARDENING_ASSERT(!empty());
-
-    AllocatorTraits<A>::destroy(storage_.GetAllocator(), data() + (size() - 1));
-    storage_.SubtractSize(1);
-  }
-
-  // `InlinedVector::erase(...)`
-  //
-  // Erases the element at `pos`, returning an `iterator` pointing to where the
-  // erased element was located.
-  //
-  // NOTE: may return `end()`, which is not dereferenceable.
-  iterator erase(const_iterator pos) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(pos >= begin());
-    ABSL_HARDENING_ASSERT(pos < end());
-
-    // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102329#c2
-    // It appears that GCC thinks that since `pos` is a const pointer and may
-    // point to uninitialized memory at this point, a warning should be
-    // issued. But `pos` is actually only used to compute an array index to
-    // write to.
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
-#pragma GCC diagnostic ignored "-Wuninitialized"
-#endif
-    return storage_.Erase(pos, pos + 1);
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic pop
-#endif
-  }
-
-  // Overload of `InlinedVector::erase(...)` that erases every element in the
-  // range [`from`, `to`), returning an `iterator` pointing to where the first
-  // erased element was located.
-  //
-  // NOTE: may return `end()`, which is not dereferenceable.
-  iterator erase(const_iterator from,
-                 const_iterator to) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(from >= begin());
-    ABSL_HARDENING_ASSERT(from <= to);
-    ABSL_HARDENING_ASSERT(to <= end());
-
-    if (ABSL_PREDICT_TRUE(from != to)) {
-      return storage_.Erase(from, to);
-    } else {
-      return const_cast<iterator>(from);
-    }
-  }
-
-  // `InlinedVector::clear()`
-  //
-  // Destroys all elements in the inlined vector, setting the size to `0` and
-  // deallocating any held memory.
-  void clear() noexcept {
-    inlined_vector_internal::DestroyAdapter<A>::DestroyElements(
-        storage_.GetAllocator(), data(), size());
-    storage_.DeallocateIfAllocated();
-
-    storage_.SetInlinedSize(0);
-  }
-
-  // `InlinedVector::reserve(...)`
-  //
-  // Ensures that there is enough room for at least `n` elements.
-  void reserve(size_type n) { storage_.Reserve(n); }
-
-  // `InlinedVector::shrink_to_fit()`
-  //
-  // Attempts to reduce memory usage by moving elements to (or keeping elements
-  // in) the smallest available buffer sufficient for containing `size()`
-  // elements.
-  //
-  // If `size()` is sufficiently small, the elements will be moved into (or kept
-  // in) the inlined space.
-  void shrink_to_fit() {
-    if (storage_.GetIsAllocated()) {
-      storage_.ShrinkToFit();
-    }
-  }
-
-  // `InlinedVector::swap(...)`
-  //
-  // Swaps the contents of the inlined vector with `other`.
-  void swap(InlinedVector& other) {
-    if (ABSL_PREDICT_TRUE(this != std::addressof(other))) {
-      storage_.Swap(std::addressof(other.storage_));
-    }
-  }
-
- private:
-  template <typename H, typename TheT, size_t TheN, typename TheA>
-  friend H AbslHashValue(H h, const absl::InlinedVector<TheT, TheN, TheA>& a);
-
-  void MoveAssignment(MemcpyPolicy, InlinedVector&& other) {
-    // Assumption check: we shouldn't be told to use memcpy to implement move
-    // assignment unless we have trivially destructible elements and an
-    // allocator that does nothing fancy.
-    static_assert(absl::is_trivially_destructible<value_type>::value, "");
-    static_assert(std::is_same<A, std::allocator<value_type>>::value, "");
-
-    // Throw away our existing heap allocation, if any. There is no need to
-    // destroy the existing elements one by one because we know they are
-    // trivially destructible.
-    storage_.DeallocateIfAllocated();
-
-    // Adopt the other vector's inline elements or heap allocation.
-    storage_.MemcpyFrom(other.storage_);
-    other.storage_.SetInlinedSize(0);
-  }
-
-  // Destroy our existing elements, if any, and adopt the heap-allocated
-  // elements of the other vector.
-  //
-  // REQUIRES: other.storage_.GetIsAllocated()
-  void DestroyExistingAndAdopt(InlinedVector&& other) {
-    ABSL_HARDENING_ASSERT(other.storage_.GetIsAllocated());
-
-    inlined_vector_internal::DestroyAdapter<A>::DestroyElements(
-        storage_.GetAllocator(), data(), size());
-    storage_.DeallocateIfAllocated();
-
-    storage_.MemcpyFrom(other.storage_);
-    other.storage_.SetInlinedSize(0);
-  }
-
-  void MoveAssignment(ElementwiseAssignPolicy, InlinedVector&& other) {
-    // Fast path: if the other vector is on the heap then we don't worry about
-    // actually move-assigning each element. Instead we only throw away our own
-    // existing elements and adopt the heap allocation of the other vector.
-    if (other.storage_.GetIsAllocated()) {
-      DestroyExistingAndAdopt(std::move(other));
-      return;
-    }
-
-    storage_.Assign(IteratorValueAdapter<A, MoveIterator<A>>(
-                        MoveIterator<A>(other.storage_.GetInlinedData())),
-                    other.size());
-  }
-
-  void MoveAssignment(ElementwiseConstructPolicy, InlinedVector&& other) {
-    // Fast path: if the other vector is on the heap then we don't worry about
-    // actually move-assigning each element. Instead we only throw away our own
-    // existing elements and adopt the heap allocation of the other vector.
-    if (other.storage_.GetIsAllocated()) {
-      DestroyExistingAndAdopt(std::move(other));
-      return;
-    }
-
-    inlined_vector_internal::DestroyAdapter<A>::DestroyElements(
-        storage_.GetAllocator(), data(), size());
-    storage_.DeallocateIfAllocated();
-
-    IteratorValueAdapter<A, MoveIterator<A>> other_values(
-        MoveIterator<A>(other.storage_.GetInlinedData()));
-    inlined_vector_internal::ConstructElements<A>(
-        storage_.GetAllocator(), storage_.GetInlinedData(), other_values,
-        other.storage_.GetSize());
-    storage_.SetInlinedSize(other.storage_.GetSize());
-  }
-
-  Storage storage_;
-};
-
-// -----------------------------------------------------------------------------
-// InlinedVector Non-Member Functions
-// -----------------------------------------------------------------------------
-
-// `swap(...)`
-//
-// Swaps the contents of two inlined vectors.
-template <typename T, size_t N, typename A>
-void swap(absl::InlinedVector<T, N, A>& a,
-          absl::InlinedVector<T, N, A>& b) noexcept(noexcept(a.swap(b))) {
-  a.swap(b);
-}
-
-// `operator==(...)`
-//
-// Tests for value-equality of two inlined vectors.
-template <typename T, size_t N, typename A>
-bool operator==(const absl::InlinedVector<T, N, A>& a,
-                const absl::InlinedVector<T, N, A>& b) {
-  auto a_data = a.data();
-  auto b_data = b.data();
-  return std::equal(a_data, a_data + a.size(), b_data, b_data + b.size());
-}
-
-// `operator!=(...)`
-//
-// Tests for value-inequality of two inlined vectors.
-template <typename T, size_t N, typename A>
-bool operator!=(const absl::InlinedVector<T, N, A>& a,
-                const absl::InlinedVector<T, N, A>& b) {
-  return !(a == b);
-}
-
-// `operator<(...)`
-//
-// Tests whether the value of an inlined vector is less than the value of
-// another inlined vector using a lexicographical comparison algorithm.
-template <typename T, size_t N, typename A>
-bool operator<(const absl::InlinedVector<T, N, A>& a,
-               const absl::InlinedVector<T, N, A>& b) {
-  auto a_data = a.data();
-  auto b_data = b.data();
-  return std::lexicographical_compare(a_data, a_data + a.size(), b_data,
-                                      b_data + b.size());
-}
-
-// `operator>(...)`
-//
-// Tests whether the value of an inlined vector is greater than the value of
-// another inlined vector using a lexicographical comparison algorithm.
-template <typename T, size_t N, typename A>
-bool operator>(const absl::InlinedVector<T, N, A>& a,
-               const absl::InlinedVector<T, N, A>& b) {
-  return b < a;
-}
-
-// `operator<=(...)`
-//
-// Tests whether the value of an inlined vector is less than or equal to the
-// value of another inlined vector using a lexicographical comparison algorithm.
-template <typename T, size_t N, typename A>
-bool operator<=(const absl::InlinedVector<T, N, A>& a,
-                const absl::InlinedVector<T, N, A>& b) {
-  return !(b < a);
-}
-
-// `operator>=(...)`
-//
-// Tests whether the value of an inlined vector is greater than or equal to the
-// value of another inlined vector using a lexicographical comparison algorithm.
-template <typename T, size_t N, typename A>
-bool operator>=(const absl::InlinedVector<T, N, A>& a,
-                const absl::InlinedVector<T, N, A>& b) {
-  return !(a < b);
-}
-
-// `AbslHashValue(...)`
-//
-// Provides `absl::Hash` support for `absl::InlinedVector`. It is uncommon to
-// call this directly.
-template <typename H, typename T, size_t N, typename A>
-H AbslHashValue(H h, const absl::InlinedVector<T, N, A>& a) {
-  auto size = a.size();
-  return H::combine(H::combine_contiguous(std::move(h), a.data(), size), size);
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INLINED_VECTOR_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/btree.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/btree.h
deleted file mode 100644
index fd7860da..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/btree.h
+++ /dev/null
@@ -1,3053 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-// A btree implementation of the STL set and map interfaces. A btree is smaller
-// and generally also faster than STL set/map (refer to the benchmarks below).
-// The red-black tree implementation of STL set/map has an overhead of 3
-// pointers (left, right and parent) plus the node color information for each
-// stored value. So a set<int32_t> consumes 40 bytes for each value stored in
-// 64-bit mode. This btree implementation stores multiple values on fixed
-// size nodes (usually 256 bytes) and doesn't store child pointers for leaf
-// nodes. The result is that a btree_set<int32_t> may use much less memory per
-// stored value. For the random insertion benchmark in btree_bench.cc, a
-// btree_set<int32_t> with node-size of 256 uses 5.1 bytes per stored value.
-//
-// The packing of multiple values on to each node of a btree has another effect
-// besides better space utilization: better cache locality due to fewer cache
-// lines being accessed. Better cache locality translates into faster
-// operations.
-//
-// CAVEATS
-//
-// Insertions and deletions on a btree can cause splitting, merging or
-// rebalancing of btree nodes. And even without these operations, insertions
-// and deletions on a btree will move values around within a node. In both
-// cases, the result is that insertions and deletions can invalidate iterators
-// pointing to values other than the one being inserted/deleted. Therefore, this
-// container does not provide pointer stability. This is notably different from
-// STL set/map which takes care to not invalidate iterators on insert/erase
-// except, of course, for iterators pointing to the value being erased.  A
-// partial workaround when erasing is available: erase() returns an iterator
-// pointing to the item just after the one that was erased (or end() if none
-// exists).
-
-#ifndef ABSL_CONTAINER_INTERNAL_BTREE_H_
-#define ABSL_CONTAINER_INTERNAL_BTREE_H_
-
-#include <algorithm>
-#include <cassert>
-#include <cstddef>
-#include <cstdint>
-#include <cstring>
-#include <functional>
-#include <iterator>
-#include <limits>
-#include <new>
-#include <string>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/internal/raw_logging.h"
-#include "absl/base/macros.h"
-#include "absl/container/internal/common.h"
-#include "absl/container/internal/common_policy_traits.h"
-#include "absl/container/internal/compressed_tuple.h"
-#include "absl/container/internal/container_memory.h"
-#include "absl/container/internal/layout.h"
-#include "absl/memory/memory.h"
-#include "absl/meta/type_traits.h"
-#include "absl/strings/cord.h"
-#include "absl/strings/string_view.h"
-#include "absl/types/compare.h"
-#include "absl/utility/utility.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-#ifdef ABSL_BTREE_ENABLE_GENERATIONS
-#error ABSL_BTREE_ENABLE_GENERATIONS cannot be directly set
-#elif defined(ABSL_HAVE_ADDRESS_SANITIZER) || \
-    defined(ABSL_HAVE_HWADDRESS_SANITIZER) || \
-    defined(ABSL_HAVE_MEMORY_SANITIZER)
-// When compiled in sanitizer mode, we add generation integers to the nodes and
-// iterators. When iterators are used, we validate that the container has not
-// been mutated since the iterator was constructed.
-#define ABSL_BTREE_ENABLE_GENERATIONS
-#endif
-
-#ifdef ABSL_BTREE_ENABLE_GENERATIONS
-constexpr bool BtreeGenerationsEnabled() { return true; }
-#else
-constexpr bool BtreeGenerationsEnabled() { return false; }
-#endif
-
-template <typename Compare, typename T, typename U>
-using compare_result_t = absl::result_of_t<const Compare(const T &, const U &)>;
-
-// A helper class that indicates if the Compare parameter is a key-compare-to
-// comparator.
-template <typename Compare, typename T>
-using btree_is_key_compare_to =
-    std::is_convertible<compare_result_t<Compare, T, T>, absl::weak_ordering>;
-
-struct StringBtreeDefaultLess {
-  using is_transparent = void;
-
-  StringBtreeDefaultLess() = default;
-
-  // Compatibility constructor.
-  StringBtreeDefaultLess(std::less<std::string>) {}        // NOLINT
-  StringBtreeDefaultLess(std::less<absl::string_view>) {}  // NOLINT
-
-  // Allow converting to std::less for use in key_comp()/value_comp().
-  explicit operator std::less<std::string>() const { return {}; }
-  explicit operator std::less<absl::string_view>() const { return {}; }
-  explicit operator std::less<absl::Cord>() const { return {}; }
-
-  absl::weak_ordering operator()(absl::string_view lhs,
-                                 absl::string_view rhs) const {
-    return compare_internal::compare_result_as_ordering(lhs.compare(rhs));
-  }
-  StringBtreeDefaultLess(std::less<absl::Cord>) {}  // NOLINT
-  absl::weak_ordering operator()(const absl::Cord &lhs,
-                                 const absl::Cord &rhs) const {
-    return compare_internal::compare_result_as_ordering(lhs.Compare(rhs));
-  }
-  absl::weak_ordering operator()(const absl::Cord &lhs,
-                                 absl::string_view rhs) const {
-    return compare_internal::compare_result_as_ordering(lhs.Compare(rhs));
-  }
-  absl::weak_ordering operator()(absl::string_view lhs,
-                                 const absl::Cord &rhs) const {
-    return compare_internal::compare_result_as_ordering(-rhs.Compare(lhs));
-  }
-};
-
-struct StringBtreeDefaultGreater {
-  using is_transparent = void;
-
-  StringBtreeDefaultGreater() = default;
-
-  StringBtreeDefaultGreater(std::greater<std::string>) {}        // NOLINT
-  StringBtreeDefaultGreater(std::greater<absl::string_view>) {}  // NOLINT
-
-  // Allow converting to std::greater for use in key_comp()/value_comp().
-  explicit operator std::greater<std::string>() const { return {}; }
-  explicit operator std::greater<absl::string_view>() const { return {}; }
-  explicit operator std::greater<absl::Cord>() const { return {}; }
-
-  absl::weak_ordering operator()(absl::string_view lhs,
-                                 absl::string_view rhs) const {
-    return compare_internal::compare_result_as_ordering(rhs.compare(lhs));
-  }
-  StringBtreeDefaultGreater(std::greater<absl::Cord>) {}  // NOLINT
-  absl::weak_ordering operator()(const absl::Cord &lhs,
-                                 const absl::Cord &rhs) const {
-    return compare_internal::compare_result_as_ordering(rhs.Compare(lhs));
-  }
-  absl::weak_ordering operator()(const absl::Cord &lhs,
-                                 absl::string_view rhs) const {
-    return compare_internal::compare_result_as_ordering(-lhs.Compare(rhs));
-  }
-  absl::weak_ordering operator()(absl::string_view lhs,
-                                 const absl::Cord &rhs) const {
-    return compare_internal::compare_result_as_ordering(rhs.Compare(lhs));
-  }
-};
-
-// See below comments for checked_compare.
-template <typename Compare, bool is_class = std::is_class<Compare>::value>
-struct checked_compare_base : Compare {
-  using Compare::Compare;
-  explicit checked_compare_base(Compare c) : Compare(std::move(c)) {}
-  const Compare &comp() const { return *this; }
-};
-template <typename Compare>
-struct checked_compare_base<Compare, false> {
-  explicit checked_compare_base(Compare c) : compare(std::move(c)) {}
-  const Compare &comp() const { return compare; }
-  Compare compare;
-};
-
-// A mechanism for opting out of checked_compare for use only in btree_test.cc.
-struct BtreeTestOnlyCheckedCompareOptOutBase {};
-
-// A helper class to adapt the specified comparator for two use cases:
-// (1) When using common Abseil string types with common comparison functors,
-// convert a boolean comparison into a three-way comparison that returns an
-// `absl::weak_ordering`. This helper class is specialized for
-// less<std::string>, greater<std::string>, less<string_view>,
-// greater<string_view>, less<absl::Cord>, and greater<absl::Cord>.
-// (2) Adapt the comparator to diagnose cases of non-strict-weak-ordering (see
-// https://en.cppreference.com/w/cpp/named_req/Compare) in debug mode. Whenever
-// a comparison is made, we will make assertions to verify that the comparator
-// is valid.
-template <typename Compare, typename Key>
-struct key_compare_adapter {
-  // Inherit from checked_compare_base to support function pointers and also
-  // keep empty-base-optimization (EBO) support for classes.
-  // Note: we can't use CompressedTuple here because that would interfere
-  // with the EBO for `btree::rightmost_`. `btree::rightmost_` is itself a
-  // CompressedTuple and nested `CompressedTuple`s don't support EBO.
-  // TODO(b/214288561): use CompressedTuple instead once it supports EBO for
-  // nested `CompressedTuple`s.
-  struct checked_compare : checked_compare_base<Compare> {
-   private:
-    using Base = typename checked_compare::checked_compare_base;
-    using Base::comp;
-
-    // If possible, returns whether `t` is equivalent to itself. We can only do
-    // this for `Key`s because we can't be sure that it's safe to call
-    // `comp()(k, k)` otherwise. Even if SFINAE allows it, there could be a
-    // compilation failure inside the implementation of the comparison operator.
-    bool is_self_equivalent(const Key &k) const {
-      // Note: this works for both boolean and three-way comparators.
-      return comp()(k, k) == 0;
-    }
-    // If we can't compare `t` with itself, returns true unconditionally.
-    template <typename T>
-    bool is_self_equivalent(const T &) const {
-      return true;
-    }
-
-   public:
-    using Base::Base;
-    checked_compare(Compare comp) : Base(std::move(comp)) {}  // NOLINT
-
-    // Allow converting to Compare for use in key_comp()/value_comp().
-    explicit operator Compare() const { return comp(); }
-
-    template <typename T, typename U,
-              absl::enable_if_t<
-                  std::is_same<bool, compare_result_t<Compare, T, U>>::value,
-                  int> = 0>
-    bool operator()(const T &lhs, const U &rhs) const {
-      // NOTE: if any of these assertions fail, then the comparator does not
-      // establish a strict-weak-ordering (see
-      // https://en.cppreference.com/w/cpp/named_req/Compare).
-      assert(is_self_equivalent(lhs));
-      assert(is_self_equivalent(rhs));
-      const bool lhs_comp_rhs = comp()(lhs, rhs);
-      assert(!lhs_comp_rhs || !comp()(rhs, lhs));
-      return lhs_comp_rhs;
-    }
-
-    template <
-        typename T, typename U,
-        absl::enable_if_t<std::is_convertible<compare_result_t<Compare, T, U>,
-                                              absl::weak_ordering>::value,
-                          int> = 0>
-    absl::weak_ordering operator()(const T &lhs, const U &rhs) const {
-      // NOTE: if any of these assertions fail, then the comparator does not
-      // establish a strict-weak-ordering (see
-      // https://en.cppreference.com/w/cpp/named_req/Compare).
-      assert(is_self_equivalent(lhs));
-      assert(is_self_equivalent(rhs));
-      const absl::weak_ordering lhs_comp_rhs = comp()(lhs, rhs);
-#ifndef NDEBUG
-      const absl::weak_ordering rhs_comp_lhs = comp()(rhs, lhs);
-      if (lhs_comp_rhs > 0) {
-        assert(rhs_comp_lhs < 0 && "lhs_comp_rhs > 0 -> rhs_comp_lhs < 0");
-      } else if (lhs_comp_rhs == 0) {
-        assert(rhs_comp_lhs == 0 && "lhs_comp_rhs == 0 -> rhs_comp_lhs == 0");
-      } else {
-        assert(rhs_comp_lhs > 0 && "lhs_comp_rhs < 0 -> rhs_comp_lhs > 0");
-      }
-#endif
-      return lhs_comp_rhs;
-    }
-  };
-  using type = absl::conditional_t<
-      std::is_base_of<BtreeTestOnlyCheckedCompareOptOutBase, Compare>::value,
-      Compare, checked_compare>;
-};
-
-template <>
-struct key_compare_adapter<std::less<std::string>, std::string> {
-  using type = StringBtreeDefaultLess;
-};
-
-template <>
-struct key_compare_adapter<std::greater<std::string>, std::string> {
-  using type = StringBtreeDefaultGreater;
-};
-
-template <>
-struct key_compare_adapter<std::less<absl::string_view>, absl::string_view> {
-  using type = StringBtreeDefaultLess;
-};
-
-template <>
-struct key_compare_adapter<std::greater<absl::string_view>, absl::string_view> {
-  using type = StringBtreeDefaultGreater;
-};
-
-template <>
-struct key_compare_adapter<std::less<absl::Cord>, absl::Cord> {
-  using type = StringBtreeDefaultLess;
-};
-
-template <>
-struct key_compare_adapter<std::greater<absl::Cord>, absl::Cord> {
-  using type = StringBtreeDefaultGreater;
-};
-
-// Detects an 'absl_btree_prefer_linear_node_search' member. This is
-// a protocol used as an opt-in or opt-out of linear search.
-//
-//  For example, this would be useful for key types that wrap an integer
-//  and define their own cheap operator<(). For example:
-//
-//   class K {
-//    public:
-//     using absl_btree_prefer_linear_node_search = std::true_type;
-//     ...
-//    private:
-//     friend bool operator<(K a, K b) { return a.k_ < b.k_; }
-//     int k_;
-//   };
-//
-//   btree_map<K, V> m;  // Uses linear search
-//
-// If T has the preference tag, then it has a preference.
-// Btree will use the tag's truth value.
-template <typename T, typename = void>
-struct has_linear_node_search_preference : std::false_type {};
-template <typename T, typename = void>
-struct prefers_linear_node_search : std::false_type {};
-template <typename T>
-struct has_linear_node_search_preference<
-    T, absl::void_t<typename T::absl_btree_prefer_linear_node_search>>
-    : std::true_type {};
-template <typename T>
-struct prefers_linear_node_search<
-    T, absl::void_t<typename T::absl_btree_prefer_linear_node_search>>
-    : T::absl_btree_prefer_linear_node_search {};
-
-template <typename Compare, typename Key>
-constexpr bool compare_has_valid_result_type() {
-  using compare_result_type = compare_result_t<Compare, Key, Key>;
-  return std::is_same<compare_result_type, bool>::value ||
-         std::is_convertible<compare_result_type, absl::weak_ordering>::value;
-}
-
-template <typename original_key_compare, typename value_type>
-class map_value_compare {
-  template <typename Params>
-  friend class btree;
-
-  // Note: this `protected` is part of the API of std::map::value_compare. See
-  // https://en.cppreference.com/w/cpp/container/map/value_compare.
- protected:
-  explicit map_value_compare(original_key_compare c) : comp(std::move(c)) {}
-
-  original_key_compare comp;  // NOLINT
-
- public:
-  auto operator()(const value_type &lhs, const value_type &rhs) const
-      -> decltype(comp(lhs.first, rhs.first)) {
-    return comp(lhs.first, rhs.first);
-  }
-};
-
-template <typename Key, typename Compare, typename Alloc, int TargetNodeSize,
-          bool IsMulti, bool IsMap, typename SlotPolicy>
-struct common_params : common_policy_traits<SlotPolicy> {
-  using original_key_compare = Compare;
-
-  // If Compare is a common comparator for a string-like type, then we adapt it
-  // to use heterogeneous lookup and to be a key-compare-to comparator.
-  // We also adapt the comparator to diagnose invalid comparators in debug mode.
-  // We disable this when `Compare` is invalid in a way that will cause
-  // adaptation to fail (having invalid return type) so that we can give a
-  // better compilation failure in static_assert_validation. If we don't do
-  // this, then there will be cascading compilation failures that are confusing
-  // for users.
-  using key_compare =
-      absl::conditional_t<!compare_has_valid_result_type<Compare, Key>(),
-                          Compare,
-                          typename key_compare_adapter<Compare, Key>::type>;
-
-  static constexpr bool kIsKeyCompareStringAdapted =
-      std::is_same<key_compare, StringBtreeDefaultLess>::value ||
-      std::is_same<key_compare, StringBtreeDefaultGreater>::value;
-  static constexpr bool kIsKeyCompareTransparent =
-      IsTransparent<original_key_compare>::value || kIsKeyCompareStringAdapted;
-
-  // A type which indicates if we have a key-compare-to functor or a plain old
-  // key-compare functor.
-  using is_key_compare_to = btree_is_key_compare_to<key_compare, Key>;
-
-  using allocator_type = Alloc;
-  using key_type = Key;
-  using size_type = size_t;
-  using difference_type = ptrdiff_t;
-
-  using slot_policy = SlotPolicy;
-  using slot_type = typename slot_policy::slot_type;
-  using value_type = typename slot_policy::value_type;
-  using init_type = typename slot_policy::mutable_value_type;
-  using pointer = value_type *;
-  using const_pointer = const value_type *;
-  using reference = value_type &;
-  using const_reference = const value_type &;
-
-  using value_compare =
-      absl::conditional_t<IsMap,
-                          map_value_compare<original_key_compare, value_type>,
-                          original_key_compare>;
-  using is_map_container = std::integral_constant<bool, IsMap>;
-
-  // For the given lookup key type, returns whether we can have multiple
-  // equivalent keys in the btree. If this is a multi-container, then we can.
-  // Otherwise, we can have multiple equivalent keys only if all of the
-  // following conditions are met:
-  // - The comparator is transparent.
-  // - The lookup key type is not the same as key_type.
-  // - The comparator is not a StringBtreeDefault{Less,Greater} comparator
-  //   that we know has the same equivalence classes for all lookup types.
-  template <typename LookupKey>
-  constexpr static bool can_have_multiple_equivalent_keys() {
-    return IsMulti || (IsTransparent<key_compare>::value &&
-                       !std::is_same<LookupKey, Key>::value &&
-                       !kIsKeyCompareStringAdapted);
-  }
-
-  enum {
-    kTargetNodeSize = TargetNodeSize,
-
-    // Upper bound for the available space for slots. This is largest for leaf
-    // nodes, which have overhead of at least a pointer + 4 bytes (for storing
-    // 3 field_types and an enum).
-    kNodeSlotSpace = TargetNodeSize - /*minimum overhead=*/(sizeof(void *) + 4),
-  };
-
-  // This is an integral type large enough to hold as many slots as will fit a
-  // node of TargetNodeSize bytes.
-  using node_count_type =
-      absl::conditional_t<(kNodeSlotSpace / sizeof(slot_type) >
-                           (std::numeric_limits<uint8_t>::max)()),
-                          uint16_t, uint8_t>;  // NOLINT
-};
-
-// An adapter class that converts a lower-bound compare into an upper-bound
-// compare. Note: there is no need to make a version of this adapter specialized
-// for key-compare-to functors because the upper-bound (the first value greater
-// than the input) is never an exact match.
-template <typename Compare>
-struct upper_bound_adapter {
-  explicit upper_bound_adapter(const Compare &c) : comp(c) {}
-  template <typename K1, typename K2>
-  bool operator()(const K1 &a, const K2 &b) const {
-    // Returns true when a is not greater than b.
-    return !compare_internal::compare_result_as_less_than(comp(b, a));
-  }
-
- private:
-  Compare comp;
-};
-
-enum class MatchKind : uint8_t { kEq, kNe };
-
-template <typename V, bool IsCompareTo>
-struct SearchResult {
-  V value;
-  MatchKind match;
-
-  static constexpr bool HasMatch() { return true; }
-  bool IsEq() const { return match == MatchKind::kEq; }
-};
-
-// When we don't use CompareTo, `match` is not present.
-// This ensures that callers can't use it accidentally when it provides no
-// useful information.
-template <typename V>
-struct SearchResult<V, false> {
-  SearchResult() {}
-  explicit SearchResult(V v) : value(v) {}
-  SearchResult(V v, MatchKind /*match*/) : value(v) {}
-
-  V value;
-
-  static constexpr bool HasMatch() { return false; }
-  static constexpr bool IsEq() { return false; }
-};
-
-// A node in the btree holding. The same node type is used for both internal
-// and leaf nodes in the btree, though the nodes are allocated in such a way
-// that the children array is only valid in internal nodes.
-template <typename Params>
-class btree_node {
-  using is_key_compare_to = typename Params::is_key_compare_to;
-  using field_type = typename Params::node_count_type;
-  using allocator_type = typename Params::allocator_type;
-  using slot_type = typename Params::slot_type;
-  using original_key_compare = typename Params::original_key_compare;
-
- public:
-  using params_type = Params;
-  using key_type = typename Params::key_type;
-  using value_type = typename Params::value_type;
-  using pointer = typename Params::pointer;
-  using const_pointer = typename Params::const_pointer;
-  using reference = typename Params::reference;
-  using const_reference = typename Params::const_reference;
-  using key_compare = typename Params::key_compare;
-  using size_type = typename Params::size_type;
-  using difference_type = typename Params::difference_type;
-
-  // Btree decides whether to use linear node search as follows:
-  //   - If the comparator expresses a preference, use that.
-  //   - If the key expresses a preference, use that.
-  //   - If the key is arithmetic and the comparator is std::less or
-  //     std::greater, choose linear.
-  //   - Otherwise, choose binary.
-  // TODO(ezb): Might make sense to add condition(s) based on node-size.
-  using use_linear_search = std::integral_constant<
-      bool, has_linear_node_search_preference<original_key_compare>::value
-                ? prefers_linear_node_search<original_key_compare>::value
-            : has_linear_node_search_preference<key_type>::value
-                ? prefers_linear_node_search<key_type>::value
-                : std::is_arithmetic<key_type>::value &&
-                      (std::is_same<std::less<key_type>,
-                                    original_key_compare>::value ||
-                       std::is_same<std::greater<key_type>,
-                                    original_key_compare>::value)>;
-
-  // This class is organized by absl::container_internal::Layout as if it had
-  // the following structure:
-  //   // A pointer to the node's parent.
-  //   btree_node *parent;
-  //
-  //   // When ABSL_BTREE_ENABLE_GENERATIONS is defined, we also have a
-  //   // generation integer in order to check that when iterators are
-  //   // used, they haven't been invalidated already. Only the generation on
-  //   // the root is used, but we have one on each node because whether a node
-  //   // is root or not can change.
-  //   uint32_t generation;
-  //
-  //   // The position of the node in the node's parent.
-  //   field_type position;
-  //   // The index of the first populated value in `values`.
-  //   // TODO(ezb): right now, `start` is always 0. Update insertion/merge
-  //   // logic to allow for floating storage within nodes.
-  //   field_type start;
-  //   // The index after the last populated value in `values`. Currently, this
-  //   // is the same as the count of values.
-  //   field_type finish;
-  //   // The maximum number of values the node can hold. This is an integer in
-  //   // [1, kNodeSlots] for root leaf nodes, kNodeSlots for non-root leaf
-  //   // nodes, and kInternalNodeMaxCount (as a sentinel value) for internal
-  //   // nodes (even though there are still kNodeSlots values in the node).
-  //   // TODO(ezb): make max_count use only 4 bits and record log2(capacity)
-  //   // to free extra bits for is_root, etc.
-  //   field_type max_count;
-  //
-  //   // The array of values. The capacity is `max_count` for leaf nodes and
-  //   // kNodeSlots for internal nodes. Only the values in
-  //   // [start, finish) have been initialized and are valid.
-  //   slot_type values[max_count];
-  //
-  //   // The array of child pointers. The keys in children[i] are all less
-  //   // than key(i). The keys in children[i + 1] are all greater than key(i).
-  //   // There are 0 children for leaf nodes and kNodeSlots + 1 children for
-  //   // internal nodes.
-  //   btree_node *children[kNodeSlots + 1];
-  //
-  // This class is only constructed by EmptyNodeType. Normally, pointers to the
-  // layout above are allocated, cast to btree_node*, and de-allocated within
-  // the btree implementation.
-  ~btree_node() = default;
-  btree_node(btree_node const &) = delete;
-  btree_node &operator=(btree_node const &) = delete;
-
- protected:
-  btree_node() = default;
-
- private:
-  using layout_type =
-      absl::container_internal::Layout<btree_node *, uint32_t, field_type,
-                                       slot_type, btree_node *>;
-  constexpr static size_type SizeWithNSlots(size_type n) {
-    return layout_type(
-               /*parent*/ 1,
-               /*generation*/ BtreeGenerationsEnabled() ? 1 : 0,
-               /*position, start, finish, max_count*/ 4,
-               /*slots*/ n,
-               /*children*/ 0)
-        .AllocSize();
-  }
-  // A lower bound for the overhead of fields other than slots in a leaf node.
-  constexpr static size_type MinimumOverhead() {
-    return SizeWithNSlots(1) - sizeof(slot_type);
-  }
-
-  // Compute how many values we can fit onto a leaf node taking into account
-  // padding.
-  constexpr static size_type NodeTargetSlots(const size_type begin,
-                                             const size_type end) {
-    return begin == end ? begin
-           : SizeWithNSlots((begin + end) / 2 + 1) >
-                   params_type::kTargetNodeSize
-               ? NodeTargetSlots(begin, (begin + end) / 2)
-               : NodeTargetSlots((begin + end) / 2 + 1, end);
-  }
-
-  constexpr static size_type kTargetNodeSize = params_type::kTargetNodeSize;
-  constexpr static size_type kNodeTargetSlots =
-      NodeTargetSlots(0, kTargetNodeSize);
-
-  // We need a minimum of 3 slots per internal node in order to perform
-  // splitting (1 value for the two nodes involved in the split and 1 value
-  // propagated to the parent as the delimiter for the split). For performance
-  // reasons, we don't allow 3 slots-per-node due to bad worst case occupancy of
-  // 1/3 (for a node, not a b-tree).
-  constexpr static size_type kMinNodeSlots = 4;
-
-  constexpr static size_type kNodeSlots =
-      kNodeTargetSlots >= kMinNodeSlots ? kNodeTargetSlots : kMinNodeSlots;
-
-  // The node is internal (i.e. is not a leaf node) if and only if `max_count`
-  // has this value.
-  constexpr static field_type kInternalNodeMaxCount = 0;
-
-  constexpr static layout_type Layout(const size_type slot_count,
-                                      const size_type child_count) {
-    return layout_type(
-        /*parent*/ 1,
-        /*generation*/ BtreeGenerationsEnabled() ? 1 : 0,
-        /*position, start, finish, max_count*/ 4,
-        /*slots*/ slot_count,
-        /*children*/ child_count);
-  }
-  // Leaves can have less than kNodeSlots values.
-  constexpr static layout_type LeafLayout(
-      const size_type slot_count = kNodeSlots) {
-    return Layout(slot_count, 0);
-  }
-  constexpr static layout_type InternalLayout() {
-    return Layout(kNodeSlots, kNodeSlots + 1);
-  }
-  constexpr static size_type LeafSize(const size_type slot_count = kNodeSlots) {
-    return LeafLayout(slot_count).AllocSize();
-  }
-  constexpr static size_type InternalSize() {
-    return InternalLayout().AllocSize();
-  }
-
-  constexpr static size_type Alignment() {
-    static_assert(LeafLayout(1).Alignment() == InternalLayout().Alignment(),
-                  "Alignment of all nodes must be equal.");
-    return InternalLayout().Alignment();
-  }
-
-  // N is the index of the type in the Layout definition.
-  // ElementType<N> is the Nth type in the Layout definition.
-  template <size_type N>
-  inline typename layout_type::template ElementType<N> *GetField() {
-    // We assert that we don't read from values that aren't there.
-    assert(N < 4 || is_internal());
-    return InternalLayout().template Pointer<N>(reinterpret_cast<char *>(this));
-  }
-  template <size_type N>
-  inline const typename layout_type::template ElementType<N> *GetField() const {
-    assert(N < 4 || is_internal());
-    return InternalLayout().template Pointer<N>(
-        reinterpret_cast<const char *>(this));
-  }
-  void set_parent(btree_node *p) { *GetField<0>() = p; }
-  field_type &mutable_finish() { return GetField<2>()[2]; }
-  slot_type *slot(size_type i) { return &GetField<3>()[i]; }
-  slot_type *start_slot() { return slot(start()); }
-  slot_type *finish_slot() { return slot(finish()); }
-  const slot_type *slot(size_type i) const { return &GetField<3>()[i]; }
-  void set_position(field_type v) { GetField<2>()[0] = v; }
-  void set_start(field_type v) { GetField<2>()[1] = v; }
-  void set_finish(field_type v) { GetField<2>()[2] = v; }
-  // This method is only called by the node init methods.
-  void set_max_count(field_type v) { GetField<2>()[3] = v; }
-
- public:
-  // Whether this is a leaf node or not. This value doesn't change after the
-  // node is created.
-  bool is_leaf() const { return GetField<2>()[3] != kInternalNodeMaxCount; }
-  // Whether this is an internal node or not. This value doesn't change after
-  // the node is created.
-  bool is_internal() const { return !is_leaf(); }
-
-  // Getter for the position of this node in its parent.
-  field_type position() const { return GetField<2>()[0]; }
-
-  // Getter for the offset of the first value in the `values` array.
-  field_type start() const {
-    // TODO(ezb): when floating storage is implemented, return GetField<2>()[1];
-    assert(GetField<2>()[1] == 0);
-    return 0;
-  }
-
-  // Getter for the offset after the last value in the `values` array.
-  field_type finish() const { return GetField<2>()[2]; }
-
-  // Getters for the number of values stored in this node.
-  field_type count() const {
-    assert(finish() >= start());
-    return finish() - start();
-  }
-  field_type max_count() const {
-    // Internal nodes have max_count==kInternalNodeMaxCount.
-    // Leaf nodes have max_count in [1, kNodeSlots].
-    const field_type max_count = GetField<2>()[3];
-    return max_count == field_type{kInternalNodeMaxCount}
-               ? field_type{kNodeSlots}
-               : max_count;
-  }
-
-  // Getter for the parent of this node.
-  btree_node *parent() const { return *GetField<0>(); }
-  // Getter for whether the node is the root of the tree. The parent of the
-  // root of the tree is the leftmost node in the tree which is guaranteed to
-  // be a leaf.
-  bool is_root() const { return parent()->is_leaf(); }
-  void make_root() {
-    assert(parent()->is_root());
-    set_generation(parent()->generation());
-    set_parent(parent()->parent());
-  }
-
-  // Gets the root node's generation integer, which is the one used by the tree.
-  uint32_t *get_root_generation() const {
-    assert(BtreeGenerationsEnabled());
-    const btree_node *curr = this;
-    for (; !curr->is_root(); curr = curr->parent()) continue;
-    return const_cast<uint32_t *>(&curr->GetField<1>()[0]);
-  }
-
-  // Returns the generation for iterator validation.
-  uint32_t generation() const {
-    return BtreeGenerationsEnabled() ? *get_root_generation() : 0;
-  }
-  // Updates generation. Should only be called on a root node or during node
-  // initialization.
-  void set_generation(uint32_t generation) {
-    if (BtreeGenerationsEnabled()) GetField<1>()[0] = generation;
-  }
-  // Updates the generation. We do this whenever the node is mutated.
-  void next_generation() {
-    if (BtreeGenerationsEnabled()) ++*get_root_generation();
-  }
-
-  // Getters for the key/value at position i in the node.
-  const key_type &key(size_type i) const { return params_type::key(slot(i)); }
-  reference value(size_type i) { return params_type::element(slot(i)); }
-  const_reference value(size_type i) const {
-    return params_type::element(slot(i));
-  }
-
-  // Getters/setter for the child at position i in the node.
-  btree_node *child(field_type i) const { return GetField<4>()[i]; }
-  btree_node *start_child() const { return child(start()); }
-  btree_node *&mutable_child(field_type i) { return GetField<4>()[i]; }
-  void clear_child(field_type i) {
-    absl::container_internal::SanitizerPoisonObject(&mutable_child(i));
-  }
-  void set_child_noupdate_position(field_type i, btree_node *c) {
-    absl::container_internal::SanitizerUnpoisonObject(&mutable_child(i));
-    mutable_child(i) = c;
-  }
-  void set_child(field_type i, btree_node *c) {
-    set_child_noupdate_position(i, c);
-    c->set_position(i);
-  }
-  void init_child(field_type i, btree_node *c) {
-    set_child(i, c);
-    c->set_parent(this);
-  }
-
-  // Returns the position of the first value whose key is not less than k.
-  template <typename K>
-  SearchResult<size_type, is_key_compare_to::value> lower_bound(
-      const K &k, const key_compare &comp) const {
-    return use_linear_search::value ? linear_search(k, comp)
-                                    : binary_search(k, comp);
-  }
-  // Returns the position of the first value whose key is greater than k.
-  template <typename K>
-  size_type upper_bound(const K &k, const key_compare &comp) const {
-    auto upper_compare = upper_bound_adapter<key_compare>(comp);
-    return use_linear_search::value ? linear_search(k, upper_compare).value
-                                    : binary_search(k, upper_compare).value;
-  }
-
-  template <typename K, typename Compare>
-  SearchResult<size_type, btree_is_key_compare_to<Compare, key_type>::value>
-  linear_search(const K &k, const Compare &comp) const {
-    return linear_search_impl(k, start(), finish(), comp,
-                              btree_is_key_compare_to<Compare, key_type>());
-  }
-
-  template <typename K, typename Compare>
-  SearchResult<size_type, btree_is_key_compare_to<Compare, key_type>::value>
-  binary_search(const K &k, const Compare &comp) const {
-    return binary_search_impl(k, start(), finish(), comp,
-                              btree_is_key_compare_to<Compare, key_type>());
-  }
-
-  // Returns the position of the first value whose key is not less than k using
-  // linear search performed using plain compare.
-  template <typename K, typename Compare>
-  SearchResult<size_type, false> linear_search_impl(
-      const K &k, size_type s, const size_type e, const Compare &comp,
-      std::false_type /* IsCompareTo */) const {
-    while (s < e) {
-      if (!comp(key(s), k)) {
-        break;
-      }
-      ++s;
-    }
-    return SearchResult<size_type, false>{s};
-  }
-
-  // Returns the position of the first value whose key is not less than k using
-  // linear search performed using compare-to.
-  template <typename K, typename Compare>
-  SearchResult<size_type, true> linear_search_impl(
-      const K &k, size_type s, const size_type e, const Compare &comp,
-      std::true_type /* IsCompareTo */) const {
-    while (s < e) {
-      const absl::weak_ordering c = comp(key(s), k);
-      if (c == 0) {
-        return {s, MatchKind::kEq};
-      } else if (c > 0) {
-        break;
-      }
-      ++s;
-    }
-    return {s, MatchKind::kNe};
-  }
-
-  // Returns the position of the first value whose key is not less than k using
-  // binary search performed using plain compare.
-  template <typename K, typename Compare>
-  SearchResult<size_type, false> binary_search_impl(
-      const K &k, size_type s, size_type e, const Compare &comp,
-      std::false_type /* IsCompareTo */) const {
-    while (s != e) {
-      const size_type mid = (s + e) >> 1;
-      if (comp(key(mid), k)) {
-        s = mid + 1;
-      } else {
-        e = mid;
-      }
-    }
-    return SearchResult<size_type, false>{s};
-  }
-
-  // Returns the position of the first value whose key is not less than k using
-  // binary search performed using compare-to.
-  template <typename K, typename CompareTo>
-  SearchResult<size_type, true> binary_search_impl(
-      const K &k, size_type s, size_type e, const CompareTo &comp,
-      std::true_type /* IsCompareTo */) const {
-    if (params_type::template can_have_multiple_equivalent_keys<K>()) {
-      MatchKind exact_match = MatchKind::kNe;
-      while (s != e) {
-        const size_type mid = (s + e) >> 1;
-        const absl::weak_ordering c = comp(key(mid), k);
-        if (c < 0) {
-          s = mid + 1;
-        } else {
-          e = mid;
-          if (c == 0) {
-            // Need to return the first value whose key is not less than k,
-            // which requires continuing the binary search if there could be
-            // multiple equivalent keys.
-            exact_match = MatchKind::kEq;
-          }
-        }
-      }
-      return {s, exact_match};
-    } else {  // Can't have multiple equivalent keys.
-      while (s != e) {
-        const size_type mid = (s + e) >> 1;
-        const absl::weak_ordering c = comp(key(mid), k);
-        if (c < 0) {
-          s = mid + 1;
-        } else if (c > 0) {
-          e = mid;
-        } else {
-          return {mid, MatchKind::kEq};
-        }
-      }
-      return {s, MatchKind::kNe};
-    }
-  }
-
-  // Returns whether key i is ordered correctly with respect to the other keys
-  // in the node. The motivation here is to detect comparators that violate
-  // transitivity. Note: we only do comparisons of keys on this node rather than
-  // the whole tree so that this is constant time.
-  template <typename Compare>
-  bool is_ordered_correctly(field_type i, const Compare &comp) const {
-    if (std::is_base_of<BtreeTestOnlyCheckedCompareOptOutBase,
-                        Compare>::value ||
-        params_type::kIsKeyCompareStringAdapted) {
-      return true;
-    }
-
-    const auto compare = [&](field_type a, field_type b) {
-      const absl::weak_ordering cmp =
-          compare_internal::do_three_way_comparison(comp, key(a), key(b));
-      return cmp < 0 ? -1 : cmp > 0 ? 1 : 0;
-    };
-    int cmp = -1;
-    constexpr bool kCanHaveEquivKeys =
-        params_type::template can_have_multiple_equivalent_keys<key_type>();
-    for (field_type j = start(); j < finish(); ++j) {
-      if (j == i) {
-        if (cmp > 0) return false;
-        continue;
-      }
-      int new_cmp = compare(j, i);
-      if (new_cmp < cmp || (!kCanHaveEquivKeys && new_cmp == 0)) return false;
-      cmp = new_cmp;
-    }
-    return true;
-  }
-
-  // Emplaces a value at position i, shifting all existing values and
-  // children at positions >= i to the right by 1.
-  template <typename... Args>
-  void emplace_value(field_type i, allocator_type *alloc, Args &&...args);
-
-  // Removes the values at positions [i, i + to_erase), shifting all existing
-  // values and children after that range to the left by to_erase. Clears all
-  // children between [i, i + to_erase).
-  void remove_values(field_type i, field_type to_erase, allocator_type *alloc);
-
-  // Rebalances a node with its right sibling.
-  void rebalance_right_to_left(field_type to_move, btree_node *right,
-                               allocator_type *alloc);
-  void rebalance_left_to_right(field_type to_move, btree_node *right,
-                               allocator_type *alloc);
-
-  // Splits a node, moving a portion of the node's values to its right sibling.
-  void split(int insert_position, btree_node *dest, allocator_type *alloc);
-
-  // Merges a node with its right sibling, moving all of the values and the
-  // delimiting key in the parent node onto itself, and deleting the src node.
-  void merge(btree_node *src, allocator_type *alloc);
-
-  // Node allocation/deletion routines.
-  void init_leaf(field_type position, field_type max_count,
-                 btree_node *parent) {
-    set_generation(0);
-    set_parent(parent);
-    set_position(position);
-    set_start(0);
-    set_finish(0);
-    set_max_count(max_count);
-    absl::container_internal::SanitizerPoisonMemoryRegion(
-        start_slot(), max_count * sizeof(slot_type));
-  }
-  void init_internal(field_type position, btree_node *parent) {
-    init_leaf(position, kNodeSlots, parent);
-    // Set `max_count` to a sentinel value to indicate that this node is
-    // internal.
-    set_max_count(kInternalNodeMaxCount);
-    absl::container_internal::SanitizerPoisonMemoryRegion(
-        &mutable_child(start()), (kNodeSlots + 1) * sizeof(btree_node *));
-  }
-
-  static void deallocate(const size_type size, btree_node *node,
-                         allocator_type *alloc) {
-    absl::container_internal::SanitizerUnpoisonMemoryRegion(node, size);
-    absl::container_internal::Deallocate<Alignment()>(alloc, node, size);
-  }
-
-  // Deletes a node and all of its children.
-  static void clear_and_delete(btree_node *node, allocator_type *alloc);
-
- private:
-  template <typename... Args>
-  void value_init(const field_type i, allocator_type *alloc, Args &&...args) {
-    next_generation();
-    absl::container_internal::SanitizerUnpoisonObject(slot(i));
-    params_type::construct(alloc, slot(i), std::forward<Args>(args)...);
-  }
-  void value_destroy(const field_type i, allocator_type *alloc) {
-    next_generation();
-    params_type::destroy(alloc, slot(i));
-    absl::container_internal::SanitizerPoisonObject(slot(i));
-  }
-  void value_destroy_n(const field_type i, const field_type n,
-                       allocator_type *alloc) {
-    next_generation();
-    for (slot_type *s = slot(i), *end = slot(i + n); s != end; ++s) {
-      params_type::destroy(alloc, s);
-      absl::container_internal::SanitizerPoisonObject(s);
-    }
-  }
-
-  static void transfer(slot_type *dest, slot_type *src, allocator_type *alloc) {
-    absl::container_internal::SanitizerUnpoisonObject(dest);
-    params_type::transfer(alloc, dest, src);
-    absl::container_internal::SanitizerPoisonObject(src);
-  }
-
-  // Transfers value from slot `src_i` in `src_node` to slot `dest_i` in `this`.
-  void transfer(const size_type dest_i, const size_type src_i,
-                btree_node *src_node, allocator_type *alloc) {
-    next_generation();
-    transfer(slot(dest_i), src_node->slot(src_i), alloc);
-  }
-
-  // Transfers `n` values starting at value `src_i` in `src_node` into the
-  // values starting at value `dest_i` in `this`.
-  void transfer_n(const size_type n, const size_type dest_i,
-                  const size_type src_i, btree_node *src_node,
-                  allocator_type *alloc) {
-    next_generation();
-    for (slot_type *src = src_node->slot(src_i), *end = src + n,
-                   *dest = slot(dest_i);
-         src != end; ++src, ++dest) {
-      transfer(dest, src, alloc);
-    }
-  }
-
-  // Same as above, except that we start at the end and work our way to the
-  // beginning.
-  void transfer_n_backward(const size_type n, const size_type dest_i,
-                           const size_type src_i, btree_node *src_node,
-                           allocator_type *alloc) {
-    next_generation();
-    for (slot_type *src = src_node->slot(src_i + n), *end = src - n,
-                   *dest = slot(dest_i + n);
-         src != end; --src, --dest) {
-      // If we modified the loop index calculations above to avoid the -1s here,
-      // it would result in UB in the computation of `end` (and possibly `src`
-      // as well, if n == 0), since slot() is effectively an array index and it
-      // is UB to compute the address of any out-of-bounds array element except
-      // for one-past-the-end.
-      transfer(dest - 1, src - 1, alloc);
-    }
-  }
-
-  template <typename P>
-  friend class btree;
-  template <typename N, typename R, typename P>
-  friend class btree_iterator;
-  friend class BtreeNodePeer;
-  friend struct btree_access;
-};
-
-template <typename Node>
-bool AreNodesFromSameContainer(const Node *node_a, const Node *node_b) {
-  // If either node is null, then give up on checking whether they're from the
-  // same container. (If exactly one is null, then we'll trigger the
-  // default-constructed assert in Equals.)
-  if (node_a == nullptr || node_b == nullptr) return true;
-  while (!node_a->is_root()) node_a = node_a->parent();
-  while (!node_b->is_root()) node_b = node_b->parent();
-  return node_a == node_b;
-}
-
-class btree_iterator_generation_info_enabled {
- public:
-  explicit btree_iterator_generation_info_enabled(uint32_t g)
-      : generation_(g) {}
-
-  // Updates the generation. For use internally right before we return an
-  // iterator to the user.
-  template <typename Node>
-  void update_generation(const Node *node) {
-    if (node != nullptr) generation_ = node->generation();
-  }
-  uint32_t generation() const { return generation_; }
-
-  template <typename Node>
-  void assert_valid_generation(const Node *node) const {
-    if (node != nullptr && node->generation() != generation_) {
-      ABSL_INTERNAL_LOG(
-          FATAL,
-          "Attempting to use an invalidated iterator. The corresponding b-tree "
-          "container has been mutated since this iterator was constructed.");
-    }
-  }
-
- private:
-  // Used to check that the iterator hasn't been invalidated.
-  uint32_t generation_;
-};
-
-class btree_iterator_generation_info_disabled {
- public:
-  explicit btree_iterator_generation_info_disabled(uint32_t) {}
-  static void update_generation(const void *) {}
-  static uint32_t generation() { return 0; }
-  static void assert_valid_generation(const void *) {}
-};
-
-#ifdef ABSL_BTREE_ENABLE_GENERATIONS
-using btree_iterator_generation_info = btree_iterator_generation_info_enabled;
-#else
-using btree_iterator_generation_info = btree_iterator_generation_info_disabled;
-#endif
-
-template <typename Node, typename Reference, typename Pointer>
-class btree_iterator : private btree_iterator_generation_info {
-  using field_type = typename Node::field_type;
-  using key_type = typename Node::key_type;
-  using size_type = typename Node::size_type;
-  using params_type = typename Node::params_type;
-  using is_map_container = typename params_type::is_map_container;
-
-  using node_type = Node;
-  using normal_node = typename std::remove_const<Node>::type;
-  using const_node = const Node;
-  using normal_pointer = typename params_type::pointer;
-  using normal_reference = typename params_type::reference;
-  using const_pointer = typename params_type::const_pointer;
-  using const_reference = typename params_type::const_reference;
-  using slot_type = typename params_type::slot_type;
-
-  // In sets, all iterators are const.
-  using iterator = absl::conditional_t<
-      is_map_container::value,
-      btree_iterator<normal_node, normal_reference, normal_pointer>,
-      btree_iterator<normal_node, const_reference, const_pointer>>;
-  using const_iterator =
-      btree_iterator<const_node, const_reference, const_pointer>;
-
- public:
-  // These aliases are public for std::iterator_traits.
-  using difference_type = typename Node::difference_type;
-  using value_type = typename params_type::value_type;
-  using pointer = Pointer;
-  using reference = Reference;
-  using iterator_category = std::bidirectional_iterator_tag;
-
-  btree_iterator() : btree_iterator(nullptr, -1) {}
-  explicit btree_iterator(Node *n) : btree_iterator(n, n->start()) {}
-  btree_iterator(Node *n, int p)
-      : btree_iterator_generation_info(n != nullptr ? n->generation()
-                                                    : ~uint32_t{}),
-        node_(n),
-        position_(p) {}
-
-  // NOTE: this SFINAE allows for implicit conversions from iterator to
-  // const_iterator, but it specifically avoids hiding the copy constructor so
-  // that the trivial one will be used when possible.
-  template <typename N, typename R, typename P,
-            absl::enable_if_t<
-                std::is_same<btree_iterator<N, R, P>, iterator>::value &&
-                    std::is_same<btree_iterator, const_iterator>::value,
-                int> = 0>
-  btree_iterator(const btree_iterator<N, R, P> other)  // NOLINT
-      : btree_iterator_generation_info(other),
-        node_(other.node_),
-        position_(other.position_) {}
-
-  bool operator==(const iterator &other) const {
-    return Equals(other);
-  }
-  bool operator==(const const_iterator &other) const {
-    return Equals(other);
-  }
-  bool operator!=(const iterator &other) const {
-    return !Equals(other);
-  }
-  bool operator!=(const const_iterator &other) const {
-    return !Equals(other);
-  }
-
-  // Returns n such that n calls to ++other yields *this.
-  // Precondition: n exists.
-  difference_type operator-(const_iterator other) const {
-    if (node_ == other.node_) {
-      if (node_->is_leaf()) return position_ - other.position_;
-      if (position_ == other.position_) return 0;
-    }
-    return distance_slow(other);
-  }
-
-  // Accessors for the key/value the iterator is pointing at.
-  reference operator*() const {
-    ABSL_HARDENING_ASSERT(node_ != nullptr);
-    assert_valid_generation(node_);
-    ABSL_HARDENING_ASSERT(position_ >= node_->start());
-    if (position_ >= node_->finish()) {
-      ABSL_HARDENING_ASSERT(!IsEndIterator() && "Dereferencing end() iterator");
-      ABSL_HARDENING_ASSERT(position_ < node_->finish());
-    }
-    return node_->value(static_cast<field_type>(position_));
-  }
-  pointer operator->() const { return &operator*(); }
-
-  btree_iterator &operator++() {
-    increment();
-    return *this;
-  }
-  btree_iterator &operator--() {
-    decrement();
-    return *this;
-  }
-  btree_iterator operator++(int) {
-    btree_iterator tmp = *this;
-    ++*this;
-    return tmp;
-  }
-  btree_iterator operator--(int) {
-    btree_iterator tmp = *this;
-    --*this;
-    return tmp;
-  }
-
- private:
-  friend iterator;
-  friend const_iterator;
-  template <typename Params>
-  friend class btree;
-  template <typename Tree>
-  friend class btree_container;
-  template <typename Tree>
-  friend class btree_set_container;
-  template <typename Tree>
-  friend class btree_map_container;
-  template <typename Tree>
-  friend class btree_multiset_container;
-  template <typename TreeType, typename CheckerType>
-  friend class base_checker;
-  friend struct btree_access;
-
-  // This SFINAE allows explicit conversions from const_iterator to
-  // iterator, but also avoids hiding the copy constructor.
-  // NOTE: the const_cast is safe because this constructor is only called by
-  // non-const methods and the container owns the nodes.
-  template <typename N, typename R, typename P,
-            absl::enable_if_t<
-                std::is_same<btree_iterator<N, R, P>, const_iterator>::value &&
-                    std::is_same<btree_iterator, iterator>::value,
-                int> = 0>
-  explicit btree_iterator(const btree_iterator<N, R, P> other)
-      : btree_iterator_generation_info(other.generation()),
-        node_(const_cast<node_type *>(other.node_)),
-        position_(other.position_) {}
-
-  bool Equals(const const_iterator other) const {
-    ABSL_HARDENING_ASSERT(((node_ == nullptr && other.node_ == nullptr) ||
-                           (node_ != nullptr && other.node_ != nullptr)) &&
-                          "Comparing default-constructed iterator with "
-                          "non-default-constructed iterator.");
-    // Note: we use assert instead of ABSL_HARDENING_ASSERT here because this
-    // changes the complexity of Equals from O(1) to O(log(N) + log(M)) where
-    // N/M are sizes of the containers containing node_/other.node_.
-    assert(AreNodesFromSameContainer(node_, other.node_) &&
-           "Comparing iterators from different containers.");
-    assert_valid_generation(node_);
-    other.assert_valid_generation(other.node_);
-    return node_ == other.node_ && position_ == other.position_;
-  }
-
-  bool IsEndIterator() const {
-    if (position_ != node_->finish()) return false;
-    node_type *node = node_;
-    while (!node->is_root()) {
-      if (node->position() != node->parent()->finish()) return false;
-      node = node->parent();
-    }
-    return true;
-  }
-
-  // Returns n such that n calls to ++other yields *this.
-  // Precondition: n exists && (this->node_ != other.node_ ||
-  // !this->node_->is_leaf() || this->position_ != other.position_).
-  difference_type distance_slow(const_iterator other) const;
-
-  // Increment/decrement the iterator.
-  void increment() {
-    assert_valid_generation(node_);
-    if (node_->is_leaf() && ++position_ < node_->finish()) {
-      return;
-    }
-    increment_slow();
-  }
-  void increment_slow();
-
-  void decrement() {
-    assert_valid_generation(node_);
-    if (node_->is_leaf() && --position_ >= node_->start()) {
-      return;
-    }
-    decrement_slow();
-  }
-  void decrement_slow();
-
-  const key_type &key() const {
-    return node_->key(static_cast<size_type>(position_));
-  }
-  decltype(std::declval<Node *>()->slot(0)) slot() {
-    return node_->slot(static_cast<size_type>(position_));
-  }
-
-  void update_generation() {
-    btree_iterator_generation_info::update_generation(node_);
-  }
-
-  // The node in the tree the iterator is pointing at.
-  Node *node_;
-  // The position within the node of the tree the iterator is pointing at.
-  // NOTE: this is an int rather than a field_type because iterators can point
-  // to invalid positions (such as -1) in certain circumstances.
-  int position_;
-};
-
-template <typename Params>
-class btree {
-  using node_type = btree_node<Params>;
-  using is_key_compare_to = typename Params::is_key_compare_to;
-  using field_type = typename node_type::field_type;
-
-  // We use a static empty node for the root/leftmost/rightmost of empty btrees
-  // in order to avoid branching in begin()/end().
-  struct EmptyNodeType : node_type {
-    using field_type = typename node_type::field_type;
-    node_type *parent;
-#ifdef ABSL_BTREE_ENABLE_GENERATIONS
-    uint32_t generation = 0;
-#endif
-    field_type position = 0;
-    field_type start = 0;
-    field_type finish = 0;
-    // max_count must be != kInternalNodeMaxCount (so that this node is regarded
-    // as a leaf node). max_count() is never called when the tree is empty.
-    field_type max_count = node_type::kInternalNodeMaxCount + 1;
-
-    constexpr EmptyNodeType() : parent(this) {}
-  };
-
-  static node_type *EmptyNode() {
-    alignas(node_type::Alignment()) static constexpr EmptyNodeType empty_node;
-    return const_cast<EmptyNodeType *>(&empty_node);
-  }
-
-  enum : uint32_t {
-    kNodeSlots = node_type::kNodeSlots,
-    kMinNodeValues = kNodeSlots / 2,
-  };
-
-  struct node_stats {
-    using size_type = typename Params::size_type;
-
-    node_stats(size_type l, size_type i) : leaf_nodes(l), internal_nodes(i) {}
-
-    node_stats &operator+=(const node_stats &other) {
-      leaf_nodes += other.leaf_nodes;
-      internal_nodes += other.internal_nodes;
-      return *this;
-    }
-
-    size_type leaf_nodes;
-    size_type internal_nodes;
-  };
-
- public:
-  using key_type = typename Params::key_type;
-  using value_type = typename Params::value_type;
-  using size_type = typename Params::size_type;
-  using difference_type = typename Params::difference_type;
-  using key_compare = typename Params::key_compare;
-  using original_key_compare = typename Params::original_key_compare;
-  using value_compare = typename Params::value_compare;
-  using allocator_type = typename Params::allocator_type;
-  using reference = typename Params::reference;
-  using const_reference = typename Params::const_reference;
-  using pointer = typename Params::pointer;
-  using const_pointer = typename Params::const_pointer;
-  using iterator =
-      typename btree_iterator<node_type, reference, pointer>::iterator;
-  using const_iterator = typename iterator::const_iterator;
-  using reverse_iterator = std::reverse_iterator<iterator>;
-  using const_reverse_iterator = std::reverse_iterator<const_iterator>;
-  using node_handle_type = node_handle<Params, Params, allocator_type>;
-
-  // Internal types made public for use by btree_container types.
-  using params_type = Params;
-  using slot_type = typename Params::slot_type;
-
- private:
-  // Copies or moves (depending on the template parameter) the values in
-  // other into this btree in their order in other. This btree must be empty
-  // before this method is called. This method is used in copy construction,
-  // copy assignment, and move assignment.
-  template <typename Btree>
-  void copy_or_move_values_in_order(Btree &other);
-
-  // Validates that various assumptions/requirements are true at compile time.
-  constexpr static bool static_assert_validation();
-
- public:
-  btree(const key_compare &comp, const allocator_type &alloc)
-      : root_(EmptyNode()), rightmost_(comp, alloc, EmptyNode()), size_(0) {}
-
-  btree(const btree &other) : btree(other, other.allocator()) {}
-  btree(const btree &other, const allocator_type &alloc)
-      : btree(other.key_comp(), alloc) {
-    copy_or_move_values_in_order(other);
-  }
-  btree(btree &&other) noexcept
-      : root_(std::exchange(other.root_, EmptyNode())),
-        rightmost_(std::move(other.rightmost_)),
-        size_(std::exchange(other.size_, 0u)) {
-    other.mutable_rightmost() = EmptyNode();
-  }
-  btree(btree &&other, const allocator_type &alloc)
-      : btree(other.key_comp(), alloc) {
-    if (alloc == other.allocator()) {
-      swap(other);
-    } else {
-      // Move values from `other` one at a time when allocators are different.
-      copy_or_move_values_in_order(other);
-    }
-  }
-
-  ~btree() {
-    // Put static_asserts in destructor to avoid triggering them before the type
-    // is complete.
-    static_assert(static_assert_validation(), "This call must be elided.");
-    clear();
-  }
-
-  // Assign the contents of other to *this.
-  btree &operator=(const btree &other);
-  btree &operator=(btree &&other) noexcept;
-
-  iterator begin() { return iterator(leftmost()); }
-  const_iterator begin() const { return const_iterator(leftmost()); }
-  iterator end() { return iterator(rightmost(), rightmost()->finish()); }
-  const_iterator end() const {
-    return const_iterator(rightmost(), rightmost()->finish());
-  }
-  reverse_iterator rbegin() { return reverse_iterator(end()); }
-  const_reverse_iterator rbegin() const {
-    return const_reverse_iterator(end());
-  }
-  reverse_iterator rend() { return reverse_iterator(begin()); }
-  const_reverse_iterator rend() const {
-    return const_reverse_iterator(begin());
-  }
-
-  // Finds the first element whose key is not less than `key`.
-  template <typename K>
-  iterator lower_bound(const K &key) {
-    return internal_end(internal_lower_bound(key).value);
-  }
-  template <typename K>
-  const_iterator lower_bound(const K &key) const {
-    return internal_end(internal_lower_bound(key).value);
-  }
-
-  // Finds the first element whose key is not less than `key` and also returns
-  // whether that element is equal to `key`.
-  template <typename K>
-  std::pair<iterator, bool> lower_bound_equal(const K &key) const;
-
-  // Finds the first element whose key is greater than `key`.
-  template <typename K>
-  iterator upper_bound(const K &key) {
-    return internal_end(internal_upper_bound(key));
-  }
-  template <typename K>
-  const_iterator upper_bound(const K &key) const {
-    return internal_end(internal_upper_bound(key));
-  }
-
-  // Finds the range of values which compare equal to key. The first member of
-  // the returned pair is equal to lower_bound(key). The second member of the
-  // pair is equal to upper_bound(key).
-  template <typename K>
-  std::pair<iterator, iterator> equal_range(const K &key);
-  template <typename K>
-  std::pair<const_iterator, const_iterator> equal_range(const K &key) const {
-    return const_cast<btree *>(this)->equal_range(key);
-  }
-
-  // Inserts a value into the btree only if it does not already exist. The
-  // boolean return value indicates whether insertion succeeded or failed.
-  // Requirement: if `key` already exists in the btree, does not consume `args`.
-  // Requirement: `key` is never referenced after consuming `args`.
-  template <typename K, typename... Args>
-  std::pair<iterator, bool> insert_unique(const K &key, Args &&...args);
-
-  // Inserts with hint. Checks to see if the value should be placed immediately
-  // before `position` in the tree. If so, then the insertion will take
-  // amortized constant time. If not, the insertion will take amortized
-  // logarithmic time as if a call to insert_unique() were made.
-  // Requirement: if `key` already exists in the btree, does not consume `args`.
-  // Requirement: `key` is never referenced after consuming `args`.
-  template <typename K, typename... Args>
-  std::pair<iterator, bool> insert_hint_unique(iterator position, const K &key,
-                                               Args &&...args);
-
-  // Insert a range of values into the btree.
-  // Note: the first overload avoids constructing a value_type if the key
-  // already exists in the btree.
-  template <typename InputIterator,
-            typename = decltype(std::declval<const key_compare &>()(
-                params_type::key(*std::declval<InputIterator>()),
-                std::declval<const key_type &>()))>
-  void insert_iterator_unique(InputIterator b, InputIterator e, int);
-  // We need the second overload for cases in which we need to construct a
-  // value_type in order to compare it with the keys already in the btree.
-  template <typename InputIterator>
-  void insert_iterator_unique(InputIterator b, InputIterator e, char);
-
-  // Inserts a value into the btree.
-  template <typename ValueType>
-  iterator insert_multi(const key_type &key, ValueType &&v);
-
-  // Inserts a value into the btree.
-  template <typename ValueType>
-  iterator insert_multi(ValueType &&v) {
-    return insert_multi(params_type::key(v), std::forward<ValueType>(v));
-  }
-
-  // Insert with hint. Check to see if the value should be placed immediately
-  // before position in the tree. If it does, then the insertion will take
-  // amortized constant time. If not, the insertion will take amortized
-  // logarithmic time as if a call to insert_multi(v) were made.
-  template <typename ValueType>
-  iterator insert_hint_multi(iterator position, ValueType &&v);
-
-  // Insert a range of values into the btree.
-  template <typename InputIterator>
-  void insert_iterator_multi(InputIterator b,
-                             InputIterator e);
-
-  // Erase the specified iterator from the btree. The iterator must be valid
-  // (i.e. not equal to end()).  Return an iterator pointing to the node after
-  // the one that was erased (or end() if none exists).
-  // Requirement: does not read the value at `*iter`.
-  iterator erase(iterator iter);
-
-  // Erases range. Returns the number of keys erased and an iterator pointing
-  // to the element after the last erased element.
-  std::pair<size_type, iterator> erase_range(iterator begin, iterator end);
-
-  // Finds an element with key equivalent to `key` or returns `end()` if `key`
-  // is not present.
-  template <typename K>
-  iterator find(const K &key) {
-    return internal_end(internal_find(key));
-  }
-  template <typename K>
-  const_iterator find(const K &key) const {
-    return internal_end(internal_find(key));
-  }
-
-  // Clear the btree, deleting all of the values it contains.
-  void clear();
-
-  // Swaps the contents of `this` and `other`.
-  void swap(btree &other);
-
-  const key_compare &key_comp() const noexcept {
-    return rightmost_.template get<0>();
-  }
-  template <typename K1, typename K2>
-  bool compare_keys(const K1 &a, const K2 &b) const {
-    return compare_internal::compare_result_as_less_than(key_comp()(a, b));
-  }
-
-  value_compare value_comp() const {
-    return value_compare(original_key_compare(key_comp()));
-  }
-
-  // Verifies the structure of the btree.
-  void verify() const;
-
-  // Size routines.
-  size_type size() const { return size_; }
-  size_type max_size() const { return (std::numeric_limits<size_type>::max)(); }
-  bool empty() const { return size_ == 0; }
-
-  // The height of the btree. An empty tree will have height 0.
-  size_type height() const {
-    size_type h = 0;
-    if (!empty()) {
-      // Count the length of the chain from the leftmost node up to the
-      // root. We actually count from the root back around to the level below
-      // the root, but the calculation is the same because of the circularity
-      // of that traversal.
-      const node_type *n = root();
-      do {
-        ++h;
-        n = n->parent();
-      } while (n != root());
-    }
-    return h;
-  }
-
-  // The number of internal, leaf and total nodes used by the btree.
-  size_type leaf_nodes() const { return internal_stats(root()).leaf_nodes; }
-  size_type internal_nodes() const {
-    return internal_stats(root()).internal_nodes;
-  }
-  size_type nodes() const {
-    node_stats stats = internal_stats(root());
-    return stats.leaf_nodes + stats.internal_nodes;
-  }
-
-  // The total number of bytes used by the btree.
-  // TODO(b/169338300): update to support node_btree_*.
-  size_type bytes_used() const {
-    node_stats stats = internal_stats(root());
-    if (stats.leaf_nodes == 1 && stats.internal_nodes == 0) {
-      return sizeof(*this) + node_type::LeafSize(root()->max_count());
-    } else {
-      return sizeof(*this) + stats.leaf_nodes * node_type::LeafSize() +
-             stats.internal_nodes * node_type::InternalSize();
-    }
-  }
-
-  // The average number of bytes used per value stored in the btree assuming
-  // random insertion order.
-  static double average_bytes_per_value() {
-    // The expected number of values per node with random insertion order is the
-    // average of the maximum and minimum numbers of values per node.
-    const double expected_values_per_node = (kNodeSlots + kMinNodeValues) / 2.0;
-    return node_type::LeafSize() / expected_values_per_node;
-  }
-
-  // The fullness of the btree. Computed as the number of elements in the btree
-  // divided by the maximum number of elements a tree with the current number
-  // of nodes could hold. A value of 1 indicates perfect space
-  // utilization. Smaller values indicate space wastage.
-  // Returns 0 for empty trees.
-  double fullness() const {
-    if (empty()) return 0.0;
-    return static_cast<double>(size()) / (nodes() * kNodeSlots);
-  }
-  // The overhead of the btree structure in bytes per node. Computed as the
-  // total number of bytes used by the btree minus the number of bytes used for
-  // storing elements divided by the number of elements.
-  // Returns 0 for empty trees.
-  double overhead() const {
-    if (empty()) return 0.0;
-    return (bytes_used() - size() * sizeof(value_type)) /
-           static_cast<double>(size());
-  }
-
-  // The allocator used by the btree.
-  allocator_type get_allocator() const { return allocator(); }
-
- private:
-  friend struct btree_access;
-
-  // Internal accessor routines.
-  node_type *root() { return root_; }
-  const node_type *root() const { return root_; }
-  node_type *&mutable_root() noexcept { return root_; }
-  node_type *rightmost() { return rightmost_.template get<2>(); }
-  const node_type *rightmost() const { return rightmost_.template get<2>(); }
-  node_type *&mutable_rightmost() noexcept {
-    return rightmost_.template get<2>();
-  }
-  key_compare *mutable_key_comp() noexcept {
-    return &rightmost_.template get<0>();
-  }
-
-  // The leftmost node is stored as the parent of the root node.
-  node_type *leftmost() { return root()->parent(); }
-  const node_type *leftmost() const { return root()->parent(); }
-
-  // Allocator routines.
-  allocator_type *mutable_allocator() noexcept {
-    return &rightmost_.template get<1>();
-  }
-  const allocator_type &allocator() const noexcept {
-    return rightmost_.template get<1>();
-  }
-
-  // Allocates a correctly aligned node of at least size bytes using the
-  // allocator.
-  node_type *allocate(size_type size) {
-    return reinterpret_cast<node_type *>(
-        absl::container_internal::Allocate<node_type::Alignment()>(
-            mutable_allocator(), size));
-  }
-
-  // Node creation/deletion routines.
-  node_type *new_internal_node(field_type position, node_type *parent) {
-    node_type *n = allocate(node_type::InternalSize());
-    n->init_internal(position, parent);
-    return n;
-  }
-  node_type *new_leaf_node(field_type position, node_type *parent) {
-    node_type *n = allocate(node_type::LeafSize());
-    n->init_leaf(position, kNodeSlots, parent);
-    return n;
-  }
-  node_type *new_leaf_root_node(field_type max_count) {
-    node_type *n = allocate(node_type::LeafSize(max_count));
-    n->init_leaf(/*position=*/0, max_count, /*parent=*/n);
-    return n;
-  }
-
-  // Deletion helper routines.
-  iterator rebalance_after_delete(iterator iter);
-
-  // Rebalances or splits the node iter points to.
-  void rebalance_or_split(iterator *iter);
-
-  // Merges the values of left, right and the delimiting key on their parent
-  // onto left, removing the delimiting key and deleting right.
-  void merge_nodes(node_type *left, node_type *right);
-
-  // Tries to merge node with its left or right sibling, and failing that,
-  // rebalance with its left or right sibling. Returns true if a merge
-  // occurred, at which point it is no longer valid to access node. Returns
-  // false if no merging took place.
-  bool try_merge_or_rebalance(iterator *iter);
-
-  // Tries to shrink the height of the tree by 1.
-  void try_shrink();
-
-  iterator internal_end(iterator iter) {
-    return iter.node_ != nullptr ? iter : end();
-  }
-  const_iterator internal_end(const_iterator iter) const {
-    return iter.node_ != nullptr ? iter : end();
-  }
-
-  // Emplaces a value into the btree immediately before iter. Requires that
-  // key(v) <= iter.key() and (--iter).key() <= key(v).
-  template <typename... Args>
-  iterator internal_emplace(iterator iter, Args &&...args);
-
-  // Returns an iterator pointing to the first value >= the value "iter" is
-  // pointing at. Note that "iter" might be pointing to an invalid location such
-  // as iter.position_ == iter.node_->finish(). This routine simply moves iter
-  // up in the tree to a valid location. Requires: iter.node_ is non-null.
-  template <typename IterType>
-  static IterType internal_last(IterType iter);
-
-  // Returns an iterator pointing to the leaf position at which key would
-  // reside in the tree, unless there is an exact match - in which case, the
-  // result may not be on a leaf. When there's a three-way comparator, we can
-  // return whether there was an exact match. This allows the caller to avoid a
-  // subsequent comparison to determine if an exact match was made, which is
-  // important for keys with expensive comparison, such as strings.
-  template <typename K>
-  SearchResult<iterator, is_key_compare_to::value> internal_locate(
-      const K &key) const;
-
-  // Internal routine which implements lower_bound().
-  template <typename K>
-  SearchResult<iterator, is_key_compare_to::value> internal_lower_bound(
-      const K &key) const;
-
-  // Internal routine which implements upper_bound().
-  template <typename K>
-  iterator internal_upper_bound(const K &key) const;
-
-  // Internal routine which implements find().
-  template <typename K>
-  iterator internal_find(const K &key) const;
-
-  // Verifies the tree structure of node.
-  size_type internal_verify(const node_type *node, const key_type *lo,
-                            const key_type *hi) const;
-
-  node_stats internal_stats(const node_type *node) const {
-    // The root can be a static empty node.
-    if (node == nullptr || (node == root() && empty())) {
-      return node_stats(0, 0);
-    }
-    if (node->is_leaf()) {
-      return node_stats(1, 0);
-    }
-    node_stats res(0, 1);
-    for (int i = node->start(); i <= node->finish(); ++i) {
-      res += internal_stats(node->child(i));
-    }
-    return res;
-  }
-
-  node_type *root_;
-
-  // A pointer to the rightmost node. Note that the leftmost node is stored as
-  // the root's parent. We use compressed tuple in order to save space because
-  // key_compare and allocator_type are usually empty.
-  absl::container_internal::CompressedTuple<key_compare, allocator_type,
-                                            node_type *>
-      rightmost_;
-
-  // Number of values.
-  size_type size_;
-};
-
-////
-// btree_node methods
-template <typename P>
-template <typename... Args>
-inline void btree_node<P>::emplace_value(const field_type i,
-                                         allocator_type *alloc,
-                                         Args &&...args) {
-  assert(i >= start());
-  assert(i <= finish());
-  // Shift old values to create space for new value and then construct it in
-  // place.
-  if (i < finish()) {
-    transfer_n_backward(finish() - i, /*dest_i=*/i + 1, /*src_i=*/i, this,
-                        alloc);
-  }
-  value_init(static_cast<field_type>(i), alloc, std::forward<Args>(args)...);
-  set_finish(finish() + 1);
-
-  if (is_internal() && finish() > i + 1) {
-    for (field_type j = finish(); j > i + 1; --j) {
-      set_child(j, child(j - 1));
-    }
-    clear_child(i + 1);
-  }
-}
-
-template <typename P>
-inline void btree_node<P>::remove_values(const field_type i,
-                                         const field_type to_erase,
-                                         allocator_type *alloc) {
-  // Transfer values after the removed range into their new places.
-  value_destroy_n(i, to_erase, alloc);
-  const field_type orig_finish = finish();
-  const field_type src_i = i + to_erase;
-  transfer_n(orig_finish - src_i, i, src_i, this, alloc);
-
-  if (is_internal()) {
-    // Delete all children between begin and end.
-    for (field_type j = 0; j < to_erase; ++j) {
-      clear_and_delete(child(i + j + 1), alloc);
-    }
-    // Rotate children after end into new positions.
-    for (field_type j = i + to_erase + 1; j <= orig_finish; ++j) {
-      set_child(j - to_erase, child(j));
-      clear_child(j);
-    }
-  }
-  set_finish(orig_finish - to_erase);
-}
-
-template <typename P>
-void btree_node<P>::rebalance_right_to_left(field_type to_move,
-                                            btree_node *right,
-                                            allocator_type *alloc) {
-  assert(parent() == right->parent());
-  assert(position() + 1 == right->position());
-  assert(right->count() >= count());
-  assert(to_move >= 1);
-  assert(to_move <= right->count());
-
-  // 1) Move the delimiting value in the parent to the left node.
-  transfer(finish(), position(), parent(), alloc);
-
-  // 2) Move the (to_move - 1) values from the right node to the left node.
-  transfer_n(to_move - 1, finish() + 1, right->start(), right, alloc);
-
-  // 3) Move the new delimiting value to the parent from the right node.
-  parent()->transfer(position(), right->start() + to_move - 1, right, alloc);
-
-  // 4) Shift the values in the right node to their correct positions.
-  right->transfer_n(right->count() - to_move, right->start(),
-                    right->start() + to_move, right, alloc);
-
-  if (is_internal()) {
-    // Move the child pointers from the right to the left node.
-    for (field_type i = 0; i < to_move; ++i) {
-      init_child(finish() + i + 1, right->child(i));
-    }
-    for (field_type i = right->start(); i <= right->finish() - to_move; ++i) {
-      assert(i + to_move <= right->max_count());
-      right->init_child(i, right->child(i + to_move));
-      right->clear_child(i + to_move);
-    }
-  }
-
-  // Fixup `finish` on the left and right nodes.
-  set_finish(finish() + to_move);
-  right->set_finish(right->finish() - to_move);
-}
-
-template <typename P>
-void btree_node<P>::rebalance_left_to_right(field_type to_move,
-                                            btree_node *right,
-                                            allocator_type *alloc) {
-  assert(parent() == right->parent());
-  assert(position() + 1 == right->position());
-  assert(count() >= right->count());
-  assert(to_move >= 1);
-  assert(to_move <= count());
-
-  // Values in the right node are shifted to the right to make room for the
-  // new to_move values. Then, the delimiting value in the parent and the
-  // other (to_move - 1) values in the left node are moved into the right node.
-  // Lastly, a new delimiting value is moved from the left node into the
-  // parent, and the remaining empty left node entries are destroyed.
-
-  // 1) Shift existing values in the right node to their correct positions.
-  right->transfer_n_backward(right->count(), right->start() + to_move,
-                             right->start(), right, alloc);
-
-  // 2) Move the delimiting value in the parent to the right node.
-  right->transfer(right->start() + to_move - 1, position(), parent(), alloc);
-
-  // 3) Move the (to_move - 1) values from the left node to the right node.
-  right->transfer_n(to_move - 1, right->start(), finish() - (to_move - 1), this,
-                    alloc);
-
-  // 4) Move the new delimiting value to the parent from the left node.
-  parent()->transfer(position(), finish() - to_move, this, alloc);
-
-  if (is_internal()) {
-    // Move the child pointers from the left to the right node.
-    for (field_type i = right->finish() + 1; i > right->start(); --i) {
-      right->init_child(i - 1 + to_move, right->child(i - 1));
-      right->clear_child(i - 1);
-    }
-    for (field_type i = 1; i <= to_move; ++i) {
-      right->init_child(i - 1, child(finish() - to_move + i));
-      clear_child(finish() - to_move + i);
-    }
-  }
-
-  // Fixup the counts on the left and right nodes.
-  set_finish(finish() - to_move);
-  right->set_finish(right->finish() + to_move);
-}
-
-template <typename P>
-void btree_node<P>::split(const int insert_position, btree_node *dest,
-                          allocator_type *alloc) {
-  assert(dest->count() == 0);
-  assert(max_count() == kNodeSlots);
-  assert(position() + 1 == dest->position());
-  assert(parent() == dest->parent());
-
-  // We bias the split based on the position being inserted. If we're
-  // inserting at the beginning of the left node then bias the split to put
-  // more values on the right node. If we're inserting at the end of the
-  // right node then bias the split to put more values on the left node.
-  if (insert_position == start()) {
-    dest->set_finish(dest->start() + finish() - 1);
-  } else if (insert_position == kNodeSlots) {
-    dest->set_finish(dest->start());
-  } else {
-    dest->set_finish(dest->start() + count() / 2);
-  }
-  set_finish(finish() - dest->count());
-  assert(count() >= 1);
-
-  // Move values from the left sibling to the right sibling.
-  dest->transfer_n(dest->count(), dest->start(), finish(), this, alloc);
-
-  // The split key is the largest value in the left sibling.
-  --mutable_finish();
-  parent()->emplace_value(position(), alloc, finish_slot());
-  value_destroy(finish(), alloc);
-  parent()->set_child_noupdate_position(position() + 1, dest);
-
-  if (is_internal()) {
-    for (field_type i = dest->start(), j = finish() + 1; i <= dest->finish();
-         ++i, ++j) {
-      assert(child(j) != nullptr);
-      dest->init_child(i, child(j));
-      clear_child(j);
-    }
-  }
-}
-
-template <typename P>
-void btree_node<P>::merge(btree_node *src, allocator_type *alloc) {
-  assert(parent() == src->parent());
-  assert(position() + 1 == src->position());
-
-  // Move the delimiting value to the left node.
-  value_init(finish(), alloc, parent()->slot(position()));
-
-  // Move the values from the right to the left node.
-  transfer_n(src->count(), finish() + 1, src->start(), src, alloc);
-
-  if (is_internal()) {
-    // Move the child pointers from the right to the left node.
-    for (field_type i = src->start(), j = finish() + 1; i <= src->finish();
-         ++i, ++j) {
-      init_child(j, src->child(i));
-      src->clear_child(i);
-    }
-  }
-
-  // Fixup `finish` on the src and dest nodes.
-  set_finish(start() + 1 + count() + src->count());
-  src->set_finish(src->start());
-
-  // Remove the value on the parent node and delete the src node.
-  parent()->remove_values(position(), /*to_erase=*/1, alloc);
-}
-
-template <typename P>
-void btree_node<P>::clear_and_delete(btree_node *node, allocator_type *alloc) {
-  if (node->is_leaf()) {
-    node->value_destroy_n(node->start(), node->count(), alloc);
-    deallocate(LeafSize(node->max_count()), node, alloc);
-    return;
-  }
-  if (node->count() == 0) {
-    deallocate(InternalSize(), node, alloc);
-    return;
-  }
-
-  // The parent of the root of the subtree we are deleting.
-  btree_node *delete_root_parent = node->parent();
-
-  // Navigate to the leftmost leaf under node, and then delete upwards.
-  while (node->is_internal()) node = node->start_child();
-#ifdef ABSL_BTREE_ENABLE_GENERATIONS
-  // When generations are enabled, we delete the leftmost leaf last in case it's
-  // the parent of the root and we need to check whether it's a leaf before we
-  // can update the root's generation.
-  // TODO(ezb): if we change btree_node::is_root to check a bool inside the node
-  // instead of checking whether the parent is a leaf, we can remove this logic.
-  btree_node *leftmost_leaf = node;
-#endif
-  // Use `size_type` because `pos` needs to be able to hold `kNodeSlots+1`,
-  // which isn't guaranteed to be a valid `field_type`.
-  size_type pos = node->position();
-  btree_node *parent = node->parent();
-  for (;;) {
-    // In each iteration of the next loop, we delete one leaf node and go right.
-    assert(pos <= parent->finish());
-    do {
-      node = parent->child(static_cast<field_type>(pos));
-      if (node->is_internal()) {
-        // Navigate to the leftmost leaf under node.
-        while (node->is_internal()) node = node->start_child();
-        pos = node->position();
-        parent = node->parent();
-      }
-      node->value_destroy_n(node->start(), node->count(), alloc);
-#ifdef ABSL_BTREE_ENABLE_GENERATIONS
-      if (leftmost_leaf != node)
-#endif
-        deallocate(LeafSize(node->max_count()), node, alloc);
-      ++pos;
-    } while (pos <= parent->finish());
-
-    // Once we've deleted all children of parent, delete parent and go up/right.
-    assert(pos > parent->finish());
-    do {
-      node = parent;
-      pos = node->position();
-      parent = node->parent();
-      node->value_destroy_n(node->start(), node->count(), alloc);
-      deallocate(InternalSize(), node, alloc);
-      if (parent == delete_root_parent) {
-#ifdef ABSL_BTREE_ENABLE_GENERATIONS
-        deallocate(LeafSize(leftmost_leaf->max_count()), leftmost_leaf, alloc);
-#endif
-        return;
-      }
-      ++pos;
-    } while (pos > parent->finish());
-  }
-}
-
-////
-// btree_iterator methods
-
-// Note: the implementation here is based on btree_node::clear_and_delete.
-template <typename N, typename R, typename P>
-auto btree_iterator<N, R, P>::distance_slow(const_iterator other) const
-    -> difference_type {
-  const_iterator begin = other;
-  const_iterator end = *this;
-  assert(begin.node_ != end.node_ || !begin.node_->is_leaf() ||
-         begin.position_ != end.position_);
-
-  const node_type *node = begin.node_;
-  // We need to compensate for double counting if begin.node_ is a leaf node.
-  difference_type count = node->is_leaf() ? -begin.position_ : 0;
-
-  // First navigate to the leftmost leaf node past begin.
-  if (node->is_internal()) {
-    ++count;
-    node = node->child(begin.position_ + 1);
-  }
-  while (node->is_internal()) node = node->start_child();
-
-  // Use `size_type` because `pos` needs to be able to hold `kNodeSlots+1`,
-  // which isn't guaranteed to be a valid `field_type`.
-  size_type pos = node->position();
-  const node_type *parent = node->parent();
-  for (;;) {
-    // In each iteration of the next loop, we count one leaf node and go right.
-    assert(pos <= parent->finish());
-    do {
-      node = parent->child(static_cast<field_type>(pos));
-      if (node->is_internal()) {
-        // Navigate to the leftmost leaf under node.
-        while (node->is_internal()) node = node->start_child();
-        pos = node->position();
-        parent = node->parent();
-      }
-      if (node == end.node_) return count + end.position_;
-      if (parent == end.node_ && pos == static_cast<size_type>(end.position_))
-        return count + node->count();
-      // +1 is for the next internal node value.
-      count += node->count() + 1;
-      ++pos;
-    } while (pos <= parent->finish());
-
-    // Once we've counted all children of parent, go up/right.
-    assert(pos > parent->finish());
-    do {
-      node = parent;
-      pos = node->position();
-      parent = node->parent();
-      // -1 because we counted the value at end and shouldn't.
-      if (parent == end.node_ && pos == static_cast<size_type>(end.position_))
-        return count - 1;
-      ++pos;
-    } while (pos > parent->finish());
-  }
-}
-
-template <typename N, typename R, typename P>
-void btree_iterator<N, R, P>::increment_slow() {
-  if (node_->is_leaf()) {
-    assert(position_ >= node_->finish());
-    btree_iterator save(*this);
-    while (position_ == node_->finish() && !node_->is_root()) {
-      assert(node_->parent()->child(node_->position()) == node_);
-      position_ = node_->position();
-      node_ = node_->parent();
-    }
-    // TODO(ezb): assert we aren't incrementing end() instead of handling.
-    if (position_ == node_->finish()) {
-      *this = save;
-    }
-  } else {
-    assert(position_ < node_->finish());
-    node_ = node_->child(static_cast<field_type>(position_ + 1));
-    while (node_->is_internal()) {
-      node_ = node_->start_child();
-    }
-    position_ = node_->start();
-  }
-}
-
-template <typename N, typename R, typename P>
-void btree_iterator<N, R, P>::decrement_slow() {
-  if (node_->is_leaf()) {
-    assert(position_ <= -1);
-    btree_iterator save(*this);
-    while (position_ < node_->start() && !node_->is_root()) {
-      assert(node_->parent()->child(node_->position()) == node_);
-      position_ = node_->position() - 1;
-      node_ = node_->parent();
-    }
-    // TODO(ezb): assert we aren't decrementing begin() instead of handling.
-    if (position_ < node_->start()) {
-      *this = save;
-    }
-  } else {
-    assert(position_ >= node_->start());
-    node_ = node_->child(static_cast<field_type>(position_));
-    while (node_->is_internal()) {
-      node_ = node_->child(node_->finish());
-    }
-    position_ = node_->finish() - 1;
-  }
-}
-
-////
-// btree methods
-template <typename P>
-template <typename Btree>
-void btree<P>::copy_or_move_values_in_order(Btree &other) {
-  static_assert(std::is_same<btree, Btree>::value ||
-                    std::is_same<const btree, Btree>::value,
-                "Btree type must be same or const.");
-  assert(empty());
-
-  // We can avoid key comparisons because we know the order of the
-  // values is the same order we'll store them in.
-  auto iter = other.begin();
-  if (iter == other.end()) return;
-  insert_multi(iter.slot());
-  ++iter;
-  for (; iter != other.end(); ++iter) {
-    // If the btree is not empty, we can just insert the new value at the end
-    // of the tree.
-    internal_emplace(end(), iter.slot());
-  }
-}
-
-template <typename P>
-constexpr bool btree<P>::static_assert_validation() {
-  static_assert(std::is_nothrow_copy_constructible<key_compare>::value,
-                "Key comparison must be nothrow copy constructible");
-  static_assert(std::is_nothrow_copy_constructible<allocator_type>::value,
-                "Allocator must be nothrow copy constructible");
-  static_assert(std::is_trivially_copyable<iterator>::value,
-                "iterator not trivially copyable.");
-
-  // Note: We assert that kTargetValues, which is computed from
-  // Params::kTargetNodeSize, must fit the node_type::field_type.
-  static_assert(
-      kNodeSlots < (1 << (8 * sizeof(typename node_type::field_type))),
-      "target node size too large");
-
-  // Verify that key_compare returns an absl::{weak,strong}_ordering or bool.
-  static_assert(
-      compare_has_valid_result_type<key_compare, key_type>(),
-      "key comparison function must return absl::{weak,strong}_ordering or "
-      "bool.");
-
-  // Test the assumption made in setting kNodeSlotSpace.
-  static_assert(node_type::MinimumOverhead() >= sizeof(void *) + 4,
-                "node space assumption incorrect");
-
-  return true;
-}
-
-template <typename P>
-template <typename K>
-auto btree<P>::lower_bound_equal(const K &key) const
-    -> std::pair<iterator, bool> {
-  const SearchResult<iterator, is_key_compare_to::value> res =
-      internal_lower_bound(key);
-  const iterator lower = iterator(internal_end(res.value));
-  const bool equal = res.HasMatch()
-                         ? res.IsEq()
-                         : lower != end() && !compare_keys(key, lower.key());
-  return {lower, equal};
-}
-
-template <typename P>
-template <typename K>
-auto btree<P>::equal_range(const K &key) -> std::pair<iterator, iterator> {
-  const std::pair<iterator, bool> lower_and_equal = lower_bound_equal(key);
-  const iterator lower = lower_and_equal.first;
-  if (!lower_and_equal.second) {
-    return {lower, lower};
-  }
-
-  const iterator next = std::next(lower);
-  if (!params_type::template can_have_multiple_equivalent_keys<K>()) {
-    // The next iterator after lower must point to a key greater than `key`.
-    // Note: if this assert fails, then it may indicate that the comparator does
-    // not meet the equivalence requirements for Compare
-    // (see https://en.cppreference.com/w/cpp/named_req/Compare).
-    assert(next == end() || compare_keys(key, next.key()));
-    return {lower, next};
-  }
-  // Try once more to avoid the call to upper_bound() if there's only one
-  // equivalent key. This should prevent all calls to upper_bound() in cases of
-  // unique-containers with heterogeneous comparators in which all comparison
-  // operators have the same equivalence classes.
-  if (next == end() || compare_keys(key, next.key())) return {lower, next};
-
-  // In this case, we need to call upper_bound() to avoid worst case O(N)
-  // behavior if we were to iterate over equal keys.
-  return {lower, upper_bound(key)};
-}
-
-template <typename P>
-template <typename K, typename... Args>
-auto btree<P>::insert_unique(const K &key, Args &&...args)
-    -> std::pair<iterator, bool> {
-  if (empty()) {
-    mutable_root() = mutable_rightmost() = new_leaf_root_node(1);
-  }
-
-  SearchResult<iterator, is_key_compare_to::value> res = internal_locate(key);
-  iterator iter = res.value;
-
-  if (res.HasMatch()) {
-    if (res.IsEq()) {
-      // The key already exists in the tree, do nothing.
-      return {iter, false};
-    }
-  } else {
-    iterator last = internal_last(iter);
-    if (last.node_ && !compare_keys(key, last.key())) {
-      // The key already exists in the tree, do nothing.
-      return {last, false};
-    }
-  }
-  return {internal_emplace(iter, std::forward<Args>(args)...), true};
-}
-
-template <typename P>
-template <typename K, typename... Args>
-inline auto btree<P>::insert_hint_unique(iterator position, const K &key,
-                                         Args &&...args)
-    -> std::pair<iterator, bool> {
-  if (!empty()) {
-    if (position == end() || compare_keys(key, position.key())) {
-      if (position == begin() || compare_keys(std::prev(position).key(), key)) {
-        // prev.key() < key < position.key()
-        return {internal_emplace(position, std::forward<Args>(args)...), true};
-      }
-    } else if (compare_keys(position.key(), key)) {
-      ++position;
-      if (position == end() || compare_keys(key, position.key())) {
-        // {original `position`}.key() < key < {current `position`}.key()
-        return {internal_emplace(position, std::forward<Args>(args)...), true};
-      }
-    } else {
-      // position.key() == key
-      return {position, false};
-    }
-  }
-  return insert_unique(key, std::forward<Args>(args)...);
-}
-
-template <typename P>
-template <typename InputIterator, typename>
-void btree<P>::insert_iterator_unique(InputIterator b, InputIterator e, int) {
-  for (; b != e; ++b) {
-    insert_hint_unique(end(), params_type::key(*b), *b);
-  }
-}
-
-template <typename P>
-template <typename InputIterator>
-void btree<P>::insert_iterator_unique(InputIterator b, InputIterator e, char) {
-  for (; b != e; ++b) {
-    // Use a node handle to manage a temp slot.
-    auto node_handle =
-        CommonAccess::Construct<node_handle_type>(get_allocator(), *b);
-    slot_type *slot = CommonAccess::GetSlot(node_handle);
-    insert_hint_unique(end(), params_type::key(slot), slot);
-  }
-}
-
-template <typename P>
-template <typename ValueType>
-auto btree<P>::insert_multi(const key_type &key, ValueType &&v) -> iterator {
-  if (empty()) {
-    mutable_root() = mutable_rightmost() = new_leaf_root_node(1);
-  }
-
-  iterator iter = internal_upper_bound(key);
-  if (iter.node_ == nullptr) {
-    iter = end();
-  }
-  return internal_emplace(iter, std::forward<ValueType>(v));
-}
-
-template <typename P>
-template <typename ValueType>
-auto btree<P>::insert_hint_multi(iterator position, ValueType &&v) -> iterator {
-  if (!empty()) {
-    const key_type &key = params_type::key(v);
-    if (position == end() || !compare_keys(position.key(), key)) {
-      if (position == begin() ||
-          !compare_keys(key, std::prev(position).key())) {
-        // prev.key() <= key <= position.key()
-        return internal_emplace(position, std::forward<ValueType>(v));
-      }
-    } else {
-      ++position;
-      if (position == end() || !compare_keys(position.key(), key)) {
-        // {original `position`}.key() < key < {current `position`}.key()
-        return internal_emplace(position, std::forward<ValueType>(v));
-      }
-    }
-  }
-  return insert_multi(std::forward<ValueType>(v));
-}
-
-template <typename P>
-template <typename InputIterator>
-void btree<P>::insert_iterator_multi(InputIterator b, InputIterator e) {
-  for (; b != e; ++b) {
-    insert_hint_multi(end(), *b);
-  }
-}
-
-template <typename P>
-auto btree<P>::operator=(const btree &other) -> btree & {
-  if (this != &other) {
-    clear();
-
-    *mutable_key_comp() = other.key_comp();
-    if (absl::allocator_traits<
-            allocator_type>::propagate_on_container_copy_assignment::value) {
-      *mutable_allocator() = other.allocator();
-    }
-
-    copy_or_move_values_in_order(other);
-  }
-  return *this;
-}
-
-template <typename P>
-auto btree<P>::operator=(btree &&other) noexcept -> btree & {
-  if (this != &other) {
-    clear();
-
-    using std::swap;
-    if (absl::allocator_traits<
-            allocator_type>::propagate_on_container_move_assignment::value) {
-      swap(root_, other.root_);
-      // Note: `rightmost_` also contains the allocator and the key comparator.
-      swap(rightmost_, other.rightmost_);
-      swap(size_, other.size_);
-    } else {
-      if (allocator() == other.allocator()) {
-        swap(mutable_root(), other.mutable_root());
-        swap(*mutable_key_comp(), *other.mutable_key_comp());
-        swap(mutable_rightmost(), other.mutable_rightmost());
-        swap(size_, other.size_);
-      } else {
-        // We aren't allowed to propagate the allocator and the allocator is
-        // different so we can't take over its memory. We must move each element
-        // individually. We need both `other` and `this` to have `other`s key
-        // comparator while moving the values so we can't swap the key
-        // comparators.
-        *mutable_key_comp() = other.key_comp();
-        copy_or_move_values_in_order(other);
-      }
-    }
-  }
-  return *this;
-}
-
-template <typename P>
-auto btree<P>::erase(iterator iter) -> iterator {
-  iter.node_->value_destroy(static_cast<field_type>(iter.position_),
-                            mutable_allocator());
-  iter.update_generation();
-
-  const bool internal_delete = iter.node_->is_internal();
-  if (internal_delete) {
-    // Deletion of a value on an internal node. First, transfer the largest
-    // value from our left child here, then erase/rebalance from that position.
-    // We can get to the largest value from our left child by decrementing iter.
-    iterator internal_iter(iter);
-    --iter;
-    assert(iter.node_->is_leaf());
-    internal_iter.node_->transfer(
-        static_cast<size_type>(internal_iter.position_),
-        static_cast<size_type>(iter.position_), iter.node_,
-        mutable_allocator());
-  } else {
-    // Shift values after erased position in leaf. In the internal case, we
-    // don't need to do this because the leaf position is the end of the node.
-    const field_type transfer_from =
-        static_cast<field_type>(iter.position_ + 1);
-    const field_type num_to_transfer = iter.node_->finish() - transfer_from;
-    iter.node_->transfer_n(num_to_transfer,
-                           static_cast<size_type>(iter.position_),
-                           transfer_from, iter.node_, mutable_allocator());
-  }
-  // Update node finish and container size.
-  iter.node_->set_finish(iter.node_->finish() - 1);
-  --size_;
-
-  // We want to return the next value after the one we just erased. If we
-  // erased from an internal node (internal_delete == true), then the next
-  // value is ++(++iter). If we erased from a leaf node (internal_delete ==
-  // false) then the next value is ++iter. Note that ++iter may point to an
-  // internal node and the value in the internal node may move to a leaf node
-  // (iter.node_) when rebalancing is performed at the leaf level.
-
-  iterator res = rebalance_after_delete(iter);
-
-  // If we erased from an internal node, advance the iterator.
-  if (internal_delete) {
-    ++res;
-  }
-  return res;
-}
-
-template <typename P>
-auto btree<P>::rebalance_after_delete(iterator iter) -> iterator {
-  // Merge/rebalance as we walk back up the tree.
-  iterator res(iter);
-  bool first_iteration = true;
-  for (;;) {
-    if (iter.node_ == root()) {
-      try_shrink();
-      if (empty()) {
-        return end();
-      }
-      break;
-    }
-    if (iter.node_->count() >= kMinNodeValues) {
-      break;
-    }
-    bool merged = try_merge_or_rebalance(&iter);
-    // On the first iteration, we should update `res` with `iter` because `res`
-    // may have been invalidated.
-    if (first_iteration) {
-      res = iter;
-      first_iteration = false;
-    }
-    if (!merged) {
-      break;
-    }
-    iter.position_ = iter.node_->position();
-    iter.node_ = iter.node_->parent();
-  }
-  res.update_generation();
-
-  // Adjust our return value. If we're pointing at the end of a node, advance
-  // the iterator.
-  if (res.position_ == res.node_->finish()) {
-    res.position_ = res.node_->finish() - 1;
-    ++res;
-  }
-
-  return res;
-}
-
-// Note: we tried implementing this more efficiently by erasing all of the
-// elements in [begin, end) at once and then doing rebalancing once at the end
-// (rather than interleaving deletion and rebalancing), but that adds a lot of
-// complexity, which seems to outweigh the performance win.
-template <typename P>
-auto btree<P>::erase_range(iterator begin, iterator end)
-    -> std::pair<size_type, iterator> {
-  size_type count = static_cast<size_type>(end - begin);
-  assert(count >= 0);
-
-  if (count == 0) {
-    return {0, begin};
-  }
-
-  if (static_cast<size_type>(count) == size_) {
-    clear();
-    return {count, this->end()};
-  }
-
-  if (begin.node_ == end.node_) {
-    assert(end.position_ > begin.position_);
-    begin.node_->remove_values(
-        static_cast<field_type>(begin.position_),
-        static_cast<field_type>(end.position_ - begin.position_),
-        mutable_allocator());
-    size_ -= count;
-    return {count, rebalance_after_delete(begin)};
-  }
-
-  const size_type target_size = size_ - count;
-  while (size_ > target_size) {
-    if (begin.node_->is_leaf()) {
-      const size_type remaining_to_erase = size_ - target_size;
-      const size_type remaining_in_node =
-          static_cast<size_type>(begin.node_->finish() - begin.position_);
-      const field_type to_erase = static_cast<field_type>(
-          (std::min)(remaining_to_erase, remaining_in_node));
-      begin.node_->remove_values(static_cast<field_type>(begin.position_),
-                                 to_erase, mutable_allocator());
-      size_ -= to_erase;
-      begin = rebalance_after_delete(begin);
-    } else {
-      begin = erase(begin);
-    }
-  }
-  begin.update_generation();
-  return {count, begin};
-}
-
-template <typename P>
-void btree<P>::clear() {
-  if (!empty()) {
-    node_type::clear_and_delete(root(), mutable_allocator());
-  }
-  mutable_root() = mutable_rightmost() = EmptyNode();
-  size_ = 0;
-}
-
-template <typename P>
-void btree<P>::swap(btree &other) {
-  using std::swap;
-  if (absl::allocator_traits<
-          allocator_type>::propagate_on_container_swap::value) {
-    // Note: `rightmost_` also contains the allocator and the key comparator.
-    swap(rightmost_, other.rightmost_);
-  } else {
-    // It's undefined behavior if the allocators are unequal here.
-    assert(allocator() == other.allocator());
-    swap(mutable_rightmost(), other.mutable_rightmost());
-    swap(*mutable_key_comp(), *other.mutable_key_comp());
-  }
-  swap(mutable_root(), other.mutable_root());
-  swap(size_, other.size_);
-}
-
-template <typename P>
-void btree<P>::verify() const {
-  assert(root() != nullptr);
-  assert(leftmost() != nullptr);
-  assert(rightmost() != nullptr);
-  assert(empty() || size() == internal_verify(root(), nullptr, nullptr));
-  assert(leftmost() == (++const_iterator(root(), -1)).node_);
-  assert(rightmost() == (--const_iterator(root(), root()->finish())).node_);
-  assert(leftmost()->is_leaf());
-  assert(rightmost()->is_leaf());
-}
-
-template <typename P>
-void btree<P>::rebalance_or_split(iterator *iter) {
-  node_type *&node = iter->node_;
-  int &insert_position = iter->position_;
-  assert(node->count() == node->max_count());
-  assert(kNodeSlots == node->max_count());
-
-  // First try to make room on the node by rebalancing.
-  node_type *parent = node->parent();
-  if (node != root()) {
-    if (node->position() > parent->start()) {
-      // Try rebalancing with our left sibling.
-      node_type *left = parent->child(node->position() - 1);
-      assert(left->max_count() == kNodeSlots);
-      if (left->count() < kNodeSlots) {
-        // We bias rebalancing based on the position being inserted. If we're
-        // inserting at the end of the right node then we bias rebalancing to
-        // fill up the left node.
-        field_type to_move =
-            (kNodeSlots - left->count()) /
-            (1 + (static_cast<field_type>(insert_position) < kNodeSlots));
-        to_move = (std::max)(field_type{1}, to_move);
-
-        if (static_cast<field_type>(insert_position) - to_move >=
-                node->start() ||
-            left->count() + to_move < kNodeSlots) {
-          left->rebalance_right_to_left(to_move, node, mutable_allocator());
-
-          assert(node->max_count() - node->count() == to_move);
-          insert_position = static_cast<int>(
-              static_cast<field_type>(insert_position) - to_move);
-          if (insert_position < node->start()) {
-            insert_position = insert_position + left->count() + 1;
-            node = left;
-          }
-
-          assert(node->count() < node->max_count());
-          return;
-        }
-      }
-    }
-
-    if (node->position() < parent->finish()) {
-      // Try rebalancing with our right sibling.
-      node_type *right = parent->child(node->position() + 1);
-      assert(right->max_count() == kNodeSlots);
-      if (right->count() < kNodeSlots) {
-        // We bias rebalancing based on the position being inserted. If we're
-        // inserting at the beginning of the left node then we bias rebalancing
-        // to fill up the right node.
-        field_type to_move = (kNodeSlots - right->count()) /
-                             (1 + (insert_position > node->start()));
-        to_move = (std::max)(field_type{1}, to_move);
-
-        if (static_cast<field_type>(insert_position) <=
-                node->finish() - to_move ||
-            right->count() + to_move < kNodeSlots) {
-          node->rebalance_left_to_right(to_move, right, mutable_allocator());
-
-          if (insert_position > node->finish()) {
-            insert_position = insert_position - node->count() - 1;
-            node = right;
-          }
-
-          assert(node->count() < node->max_count());
-          return;
-        }
-      }
-    }
-
-    // Rebalancing failed, make sure there is room on the parent node for a new
-    // value.
-    assert(parent->max_count() == kNodeSlots);
-    if (parent->count() == kNodeSlots) {
-      iterator parent_iter(parent, node->position());
-      rebalance_or_split(&parent_iter);
-      parent = node->parent();
-    }
-  } else {
-    // Rebalancing not possible because this is the root node.
-    // Create a new root node and set the current root node as the child of the
-    // new root.
-    parent = new_internal_node(/*position=*/0, parent);
-    parent->set_generation(root()->generation());
-    parent->init_child(parent->start(), node);
-    mutable_root() = parent;
-    // If the former root was a leaf node, then it's now the rightmost node.
-    assert(parent->start_child()->is_internal() ||
-           parent->start_child() == rightmost());
-  }
-
-  // Split the node.
-  node_type *split_node;
-  if (node->is_leaf()) {
-    split_node = new_leaf_node(node->position() + 1, parent);
-    node->split(insert_position, split_node, mutable_allocator());
-    if (rightmost() == node) mutable_rightmost() = split_node;
-  } else {
-    split_node = new_internal_node(node->position() + 1, parent);
-    node->split(insert_position, split_node, mutable_allocator());
-  }
-
-  if (insert_position > node->finish()) {
-    insert_position = insert_position - node->count() - 1;
-    node = split_node;
-  }
-}
-
-template <typename P>
-void btree<P>::merge_nodes(node_type *left, node_type *right) {
-  left->merge(right, mutable_allocator());
-  if (rightmost() == right) mutable_rightmost() = left;
-}
-
-template <typename P>
-bool btree<P>::try_merge_or_rebalance(iterator *iter) {
-  node_type *parent = iter->node_->parent();
-  if (iter->node_->position() > parent->start()) {
-    // Try merging with our left sibling.
-    node_type *left = parent->child(iter->node_->position() - 1);
-    assert(left->max_count() == kNodeSlots);
-    if (1U + left->count() + iter->node_->count() <= kNodeSlots) {
-      iter->position_ += 1 + left->count();
-      merge_nodes(left, iter->node_);
-      iter->node_ = left;
-      return true;
-    }
-  }
-  if (iter->node_->position() < parent->finish()) {
-    // Try merging with our right sibling.
-    node_type *right = parent->child(iter->node_->position() + 1);
-    assert(right->max_count() == kNodeSlots);
-    if (1U + iter->node_->count() + right->count() <= kNodeSlots) {
-      merge_nodes(iter->node_, right);
-      return true;
-    }
-    // Try rebalancing with our right sibling. We don't perform rebalancing if
-    // we deleted the first element from iter->node_ and the node is not
-    // empty. This is a small optimization for the common pattern of deleting
-    // from the front of the tree.
-    if (right->count() > kMinNodeValues &&
-        (iter->node_->count() == 0 || iter->position_ > iter->node_->start())) {
-      field_type to_move = (right->count() - iter->node_->count()) / 2;
-      to_move =
-          (std::min)(to_move, static_cast<field_type>(right->count() - 1));
-      iter->node_->rebalance_right_to_left(to_move, right, mutable_allocator());
-      return false;
-    }
-  }
-  if (iter->node_->position() > parent->start()) {
-    // Try rebalancing with our left sibling. We don't perform rebalancing if
-    // we deleted the last element from iter->node_ and the node is not
-    // empty. This is a small optimization for the common pattern of deleting
-    // from the back of the tree.
-    node_type *left = parent->child(iter->node_->position() - 1);
-    if (left->count() > kMinNodeValues &&
-        (iter->node_->count() == 0 ||
-         iter->position_ < iter->node_->finish())) {
-      field_type to_move = (left->count() - iter->node_->count()) / 2;
-      to_move = (std::min)(to_move, static_cast<field_type>(left->count() - 1));
-      left->rebalance_left_to_right(to_move, iter->node_, mutable_allocator());
-      iter->position_ += to_move;
-      return false;
-    }
-  }
-  return false;
-}
-
-template <typename P>
-void btree<P>::try_shrink() {
-  node_type *orig_root = root();
-  if (orig_root->count() > 0) {
-    return;
-  }
-  // Deleted the last item on the root node, shrink the height of the tree.
-  if (orig_root->is_leaf()) {
-    assert(size() == 0);
-    mutable_root() = mutable_rightmost() = EmptyNode();
-  } else {
-    node_type *child = orig_root->start_child();
-    child->make_root();
-    mutable_root() = child;
-  }
-  node_type::clear_and_delete(orig_root, mutable_allocator());
-}
-
-template <typename P>
-template <typename IterType>
-inline IterType btree<P>::internal_last(IterType iter) {
-  assert(iter.node_ != nullptr);
-  while (iter.position_ == iter.node_->finish()) {
-    iter.position_ = iter.node_->position();
-    iter.node_ = iter.node_->parent();
-    if (iter.node_->is_leaf()) {
-      iter.node_ = nullptr;
-      break;
-    }
-  }
-  iter.update_generation();
-  return iter;
-}
-
-template <typename P>
-template <typename... Args>
-inline auto btree<P>::internal_emplace(iterator iter, Args &&...args)
-    -> iterator {
-  if (iter.node_->is_internal()) {
-    // We can't insert on an internal node. Instead, we'll insert after the
-    // previous value which is guaranteed to be on a leaf node.
-    --iter;
-    ++iter.position_;
-  }
-  const field_type max_count = iter.node_->max_count();
-  allocator_type *alloc = mutable_allocator();
-
-  const auto transfer_and_delete = [&](node_type *old_node,
-                                       node_type *new_node) {
-    new_node->transfer_n(old_node->count(), new_node->start(),
-                         old_node->start(), old_node, alloc);
-    new_node->set_finish(old_node->finish());
-    old_node->set_finish(old_node->start());
-    new_node->set_generation(old_node->generation());
-    node_type::clear_and_delete(old_node, alloc);
-  };
-  const auto replace_leaf_root_node = [&](field_type new_node_size) {
-    assert(iter.node_ == root());
-    node_type *old_root = iter.node_;
-    node_type *new_root = iter.node_ = new_leaf_root_node(new_node_size);
-    transfer_and_delete(old_root, new_root);
-    mutable_root() = mutable_rightmost() = new_root;
-  };
-
-  bool replaced_node = false;
-  if (iter.node_->count() == max_count) {
-    // Make room in the leaf for the new item.
-    if (max_count < kNodeSlots) {
-      // Insertion into the root where the root is smaller than the full node
-      // size. Simply grow the size of the root node.
-      replace_leaf_root_node(static_cast<field_type>(
-          (std::min)(static_cast<int>(kNodeSlots), 2 * max_count)));
-      replaced_node = true;
-    } else {
-      rebalance_or_split(&iter);
-    }
-  }
-  (void)replaced_node;
-#if defined(ABSL_HAVE_ADDRESS_SANITIZER) || \
-    defined(ABSL_HAVE_HWADDRESS_SANITIZER)
-  if (!replaced_node) {
-    assert(iter.node_->is_leaf());
-    if (iter.node_->is_root()) {
-      replace_leaf_root_node(max_count);
-    } else {
-      node_type *old_node = iter.node_;
-      const bool was_rightmost = rightmost() == old_node;
-      const bool was_leftmost = leftmost() == old_node;
-      node_type *parent = old_node->parent();
-      const field_type position = old_node->position();
-      node_type *new_node = iter.node_ = new_leaf_node(position, parent);
-      parent->set_child_noupdate_position(position, new_node);
-      transfer_and_delete(old_node, new_node);
-      if (was_rightmost) mutable_rightmost() = new_node;
-      // The leftmost node is stored as the parent of the root node.
-      if (was_leftmost) root()->set_parent(new_node);
-    }
-  }
-#endif
-  iter.node_->emplace_value(static_cast<field_type>(iter.position_), alloc,
-                            std::forward<Args>(args)...);
-  assert(
-      iter.node_->is_ordered_correctly(static_cast<field_type>(iter.position_),
-                                       original_key_compare(key_comp())) &&
-      "If this assert fails, then either (1) the comparator may violate "
-      "transitivity, i.e. comp(a,b) && comp(b,c) -> comp(a,c) (see "
-      "https://en.cppreference.com/w/cpp/named_req/Compare), or (2) a "
-      "key may have been mutated after it was inserted into the tree.");
-  ++size_;
-  iter.update_generation();
-  return iter;
-}
-
-template <typename P>
-template <typename K>
-inline auto btree<P>::internal_locate(const K &key) const
-    -> SearchResult<iterator, is_key_compare_to::value> {
-  iterator iter(const_cast<node_type *>(root()));
-  for (;;) {
-    SearchResult<size_type, is_key_compare_to::value> res =
-        iter.node_->lower_bound(key, key_comp());
-    iter.position_ = static_cast<int>(res.value);
-    if (res.IsEq()) {
-      return {iter, MatchKind::kEq};
-    }
-    // Note: in the non-key-compare-to case, we don't need to walk all the way
-    // down the tree if the keys are equal, but determining equality would
-    // require doing an extra comparison on each node on the way down, and we
-    // will need to go all the way to the leaf node in the expected case.
-    if (iter.node_->is_leaf()) {
-      break;
-    }
-    iter.node_ = iter.node_->child(static_cast<field_type>(iter.position_));
-  }
-  // Note: in the non-key-compare-to case, the key may actually be equivalent
-  // here (and the MatchKind::kNe is ignored).
-  return {iter, MatchKind::kNe};
-}
-
-template <typename P>
-template <typename K>
-auto btree<P>::internal_lower_bound(const K &key) const
-    -> SearchResult<iterator, is_key_compare_to::value> {
-  if (!params_type::template can_have_multiple_equivalent_keys<K>()) {
-    SearchResult<iterator, is_key_compare_to::value> ret = internal_locate(key);
-    ret.value = internal_last(ret.value);
-    return ret;
-  }
-  iterator iter(const_cast<node_type *>(root()));
-  SearchResult<size_type, is_key_compare_to::value> res;
-  bool seen_eq = false;
-  for (;;) {
-    res = iter.node_->lower_bound(key, key_comp());
-    iter.position_ = static_cast<int>(res.value);
-    if (iter.node_->is_leaf()) {
-      break;
-    }
-    seen_eq = seen_eq || res.IsEq();
-    iter.node_ = iter.node_->child(static_cast<field_type>(iter.position_));
-  }
-  if (res.IsEq()) return {iter, MatchKind::kEq};
-  return {internal_last(iter), seen_eq ? MatchKind::kEq : MatchKind::kNe};
-}
-
-template <typename P>
-template <typename K>
-auto btree<P>::internal_upper_bound(const K &key) const -> iterator {
-  iterator iter(const_cast<node_type *>(root()));
-  for (;;) {
-    iter.position_ = static_cast<int>(iter.node_->upper_bound(key, key_comp()));
-    if (iter.node_->is_leaf()) {
-      break;
-    }
-    iter.node_ = iter.node_->child(static_cast<field_type>(iter.position_));
-  }
-  return internal_last(iter);
-}
-
-template <typename P>
-template <typename K>
-auto btree<P>::internal_find(const K &key) const -> iterator {
-  SearchResult<iterator, is_key_compare_to::value> res = internal_locate(key);
-  if (res.HasMatch()) {
-    if (res.IsEq()) {
-      return res.value;
-    }
-  } else {
-    const iterator iter = internal_last(res.value);
-    if (iter.node_ != nullptr && !compare_keys(key, iter.key())) {
-      return iter;
-    }
-  }
-  return {nullptr, 0};
-}
-
-template <typename P>
-typename btree<P>::size_type btree<P>::internal_verify(
-    const node_type *node, const key_type *lo, const key_type *hi) const {
-  assert(node->count() > 0);
-  assert(node->count() <= node->max_count());
-  if (lo) {
-    assert(!compare_keys(node->key(node->start()), *lo));
-  }
-  if (hi) {
-    assert(!compare_keys(*hi, node->key(node->finish() - 1)));
-  }
-  for (int i = node->start() + 1; i < node->finish(); ++i) {
-    assert(!compare_keys(node->key(i), node->key(i - 1)));
-  }
-  size_type count = node->count();
-  if (node->is_internal()) {
-    for (field_type i = node->start(); i <= node->finish(); ++i) {
-      assert(node->child(i) != nullptr);
-      assert(node->child(i)->parent() == node);
-      assert(node->child(i)->position() == i);
-      count += internal_verify(node->child(i),
-                               i == node->start() ? lo : &node->key(i - 1),
-                               i == node->finish() ? hi : &node->key(i));
-    }
-  }
-  return count;
-}
-
-struct btree_access {
-  template <typename BtreeContainer, typename Pred>
-  static auto erase_if(BtreeContainer &container, Pred pred) ->
-      typename BtreeContainer::size_type {
-    const auto initial_size = container.size();
-    auto &tree = container.tree_;
-    auto *alloc = tree.mutable_allocator();
-    for (auto it = container.begin(); it != container.end();) {
-      if (!pred(*it)) {
-        ++it;
-        continue;
-      }
-      auto *node = it.node_;
-      if (node->is_internal()) {
-        // Handle internal nodes normally.
-        it = container.erase(it);
-        continue;
-      }
-      // If this is a leaf node, then we do all the erases from this node
-      // at once before doing rebalancing.
-
-      // The current position to transfer slots to.
-      int to_pos = it.position_;
-      node->value_destroy(it.position_, alloc);
-      while (++it.position_ < node->finish()) {
-        it.update_generation();
-        if (pred(*it)) {
-          node->value_destroy(it.position_, alloc);
-        } else {
-          node->transfer(node->slot(to_pos++), node->slot(it.position_), alloc);
-        }
-      }
-      const int num_deleted = node->finish() - to_pos;
-      tree.size_ -= num_deleted;
-      node->set_finish(to_pos);
-      it.position_ = to_pos;
-      it = tree.rebalance_after_delete(it);
-    }
-    return initial_size - container.size();
-  }
-};
-
-#undef ABSL_BTREE_ENABLE_GENERATIONS
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_BTREE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/btree_container.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/btree_container.h
deleted file mode 100644
index a68ce445..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/btree_container.h
+++ /dev/null
@@ -1,763 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_BTREE_CONTAINER_H_
-#define ABSL_CONTAINER_INTERNAL_BTREE_CONTAINER_H_
-
-#include <algorithm>
-#include <initializer_list>
-#include <iterator>
-#include <utility>
-
-#include "absl/base/attributes.h"
-#include "absl/base/internal/throw_delegate.h"
-#include "absl/container/internal/btree.h"  // IWYU pragma: export
-#include "absl/container/internal/common.h"
-#include "absl/memory/memory.h"
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-// A common base class for btree_set, btree_map, btree_multiset, and
-// btree_multimap.
-template <typename Tree>
-class btree_container {
-  using params_type = typename Tree::params_type;
-
- protected:
-  // Alias used for heterogeneous lookup functions.
-  // `key_arg<K>` evaluates to `K` when the functors are transparent and to
-  // `key_type` otherwise. It permits template argument deduction on `K` for the
-  // transparent case.
-  template <class K>
-  using key_arg =
-      typename KeyArg<params_type::kIsKeyCompareTransparent>::template type<
-          K, typename Tree::key_type>;
-
- public:
-  using key_type = typename Tree::key_type;
-  using value_type = typename Tree::value_type;
-  using size_type = typename Tree::size_type;
-  using difference_type = typename Tree::difference_type;
-  using key_compare = typename Tree::original_key_compare;
-  using value_compare = typename Tree::value_compare;
-  using allocator_type = typename Tree::allocator_type;
-  using reference = typename Tree::reference;
-  using const_reference = typename Tree::const_reference;
-  using pointer = typename Tree::pointer;
-  using const_pointer = typename Tree::const_pointer;
-  using iterator = typename Tree::iterator;
-  using const_iterator = typename Tree::const_iterator;
-  using reverse_iterator = typename Tree::reverse_iterator;
-  using const_reverse_iterator = typename Tree::const_reverse_iterator;
-  using node_type = typename Tree::node_handle_type;
-
-  struct extract_and_get_next_return_type {
-    node_type node;
-    iterator next;
-  };
-
-  // Constructors/assignments.
-  btree_container() : tree_(key_compare(), allocator_type()) {}
-  explicit btree_container(const key_compare &comp,
-                           const allocator_type &alloc = allocator_type())
-      : tree_(comp, alloc) {}
-  explicit btree_container(const allocator_type &alloc)
-      : tree_(key_compare(), alloc) {}
-
-  btree_container(const btree_container &other)
-      : btree_container(other, absl::allocator_traits<allocator_type>::
-                                   select_on_container_copy_construction(
-                                       other.get_allocator())) {}
-  btree_container(const btree_container &other, const allocator_type &alloc)
-      : tree_(other.tree_, alloc) {}
-
-  btree_container(btree_container &&other) noexcept(
-      std::is_nothrow_move_constructible<Tree>::value) = default;
-  btree_container(btree_container &&other, const allocator_type &alloc)
-      : tree_(std::move(other.tree_), alloc) {}
-
-  btree_container &operator=(const btree_container &other) = default;
-  btree_container &operator=(btree_container &&other) noexcept(
-      std::is_nothrow_move_assignable<Tree>::value) = default;
-
-  // Iterator routines.
-  iterator begin() ABSL_ATTRIBUTE_LIFETIME_BOUND { return tree_.begin(); }
-  const_iterator begin() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.begin();
-  }
-  const_iterator cbegin() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.begin();
-  }
-  iterator end() ABSL_ATTRIBUTE_LIFETIME_BOUND { return tree_.end(); }
-  const_iterator end() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.end();
-  }
-  const_iterator cend() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.end();
-  }
-  reverse_iterator rbegin() ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.rbegin();
-  }
-  const_reverse_iterator rbegin() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.rbegin();
-  }
-  const_reverse_iterator crbegin() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.rbegin();
-  }
-  reverse_iterator rend() ABSL_ATTRIBUTE_LIFETIME_BOUND { return tree_.rend(); }
-  const_reverse_iterator rend() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.rend();
-  }
-  const_reverse_iterator crend() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.rend();
-  }
-
-  // Lookup routines.
-  template <typename K = key_type>
-  size_type count(const key_arg<K> &key) const {
-    auto equal_range = this->equal_range(key);
-    return equal_range.second - equal_range.first;
-  }
-  template <typename K = key_type>
-  iterator find(const key_arg<K> &key) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.find(key);
-  }
-  template <typename K = key_type>
-  const_iterator find(const key_arg<K> &key) const
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.find(key);
-  }
-  template <typename K = key_type>
-  bool contains(const key_arg<K> &key) const {
-    return find(key) != end();
-  }
-  template <typename K = key_type>
-  iterator lower_bound(const key_arg<K> &key) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.lower_bound(key);
-  }
-  template <typename K = key_type>
-  const_iterator lower_bound(const key_arg<K> &key) const
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.lower_bound(key);
-  }
-  template <typename K = key_type>
-  iterator upper_bound(const key_arg<K> &key) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.upper_bound(key);
-  }
-  template <typename K = key_type>
-  const_iterator upper_bound(const key_arg<K> &key) const
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.upper_bound(key);
-  }
-  template <typename K = key_type>
-  std::pair<iterator, iterator> equal_range(const key_arg<K> &key)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.equal_range(key);
-  }
-  template <typename K = key_type>
-  std::pair<const_iterator, const_iterator> equal_range(
-      const key_arg<K> &key) const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.equal_range(key);
-  }
-
-  // Deletion routines. Note that there is also a deletion routine that is
-  // specific to btree_set_container/btree_multiset_container.
-
-  // Erase the specified iterator from the btree. The iterator must be valid
-  // (i.e. not equal to end()).  Return an iterator pointing to the node after
-  // the one that was erased (or end() if none exists).
-  iterator erase(const_iterator iter) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.erase(iterator(iter));
-  }
-  iterator erase(iterator iter) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.erase(iter);
-  }
-  iterator erase(const_iterator first,
-                 const_iterator last) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return tree_.erase_range(iterator(first), iterator(last)).second;
-  }
-  template <typename K = key_type>
-  size_type erase(const key_arg<K> &key) {
-    auto equal_range = this->equal_range(key);
-    return tree_.erase_range(equal_range.first, equal_range.second).first;
-  }
-
-  // Extract routines.
-  extract_and_get_next_return_type extract_and_get_next(const_iterator position)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    // Use Construct instead of Transfer because the rebalancing code will
-    // destroy the slot later.
-    // Note: we rely on erase() taking place after Construct().
-    return {CommonAccess::Construct<node_type>(get_allocator(),
-                                               iterator(position).slot()),
-            erase(position)};
-  }
-  node_type extract(iterator position) {
-    // Use Construct instead of Transfer because the rebalancing code will
-    // destroy the slot later.
-    auto node =
-        CommonAccess::Construct<node_type>(get_allocator(), position.slot());
-    erase(position);
-    return node;
-  }
-  node_type extract(const_iterator position) {
-    return extract(iterator(position));
-  }
-
-  // Utility routines.
-  ABSL_ATTRIBUTE_REINITIALIZES void clear() { tree_.clear(); }
-  void swap(btree_container &other) { tree_.swap(other.tree_); }
-  void verify() const { tree_.verify(); }
-
-  // Size routines.
-  size_type size() const { return tree_.size(); }
-  size_type max_size() const { return tree_.max_size(); }
-  bool empty() const { return tree_.empty(); }
-
-  friend bool operator==(const btree_container &x, const btree_container &y) {
-    if (x.size() != y.size()) return false;
-    return std::equal(x.begin(), x.end(), y.begin());
-  }
-
-  friend bool operator!=(const btree_container &x, const btree_container &y) {
-    return !(x == y);
-  }
-
-  friend bool operator<(const btree_container &x, const btree_container &y) {
-    return std::lexicographical_compare(x.begin(), x.end(), y.begin(), y.end());
-  }
-
-  friend bool operator>(const btree_container &x, const btree_container &y) {
-    return y < x;
-  }
-
-  friend bool operator<=(const btree_container &x, const btree_container &y) {
-    return !(y < x);
-  }
-
-  friend bool operator>=(const btree_container &x, const btree_container &y) {
-    return !(x < y);
-  }
-
-  // The allocator used by the btree.
-  allocator_type get_allocator() const { return tree_.get_allocator(); }
-
-  // The key comparator used by the btree.
-  key_compare key_comp() const { return key_compare(tree_.key_comp()); }
-  value_compare value_comp() const { return tree_.value_comp(); }
-
-  // Support absl::Hash.
-  template <typename State>
-  friend State AbslHashValue(State h, const btree_container &b) {
-    for (const auto &v : b) {
-      h = State::combine(std::move(h), v);
-    }
-    return State::combine(std::move(h), b.size());
-  }
-
- protected:
-  friend struct btree_access;
-  Tree tree_;
-};
-
-// A common base class for btree_set and btree_map.
-template <typename Tree>
-class btree_set_container : public btree_container<Tree> {
-  using super_type = btree_container<Tree>;
-  using params_type = typename Tree::params_type;
-  using init_type = typename params_type::init_type;
-  using is_key_compare_to = typename params_type::is_key_compare_to;
-  friend class BtreeNodePeer;
-
- protected:
-  template <class K>
-  using key_arg = typename super_type::template key_arg<K>;
-
- public:
-  using key_type = typename Tree::key_type;
-  using value_type = typename Tree::value_type;
-  using size_type = typename Tree::size_type;
-  using key_compare = typename Tree::original_key_compare;
-  using allocator_type = typename Tree::allocator_type;
-  using iterator = typename Tree::iterator;
-  using const_iterator = typename Tree::const_iterator;
-  using node_type = typename super_type::node_type;
-  using insert_return_type = InsertReturnType<iterator, node_type>;
-
-  // Inherit constructors.
-  using super_type::super_type;
-  btree_set_container() {}
-
-  // Range constructors.
-  template <class InputIterator>
-  btree_set_container(InputIterator b, InputIterator e,
-                      const key_compare &comp = key_compare(),
-                      const allocator_type &alloc = allocator_type())
-      : super_type(comp, alloc) {
-    insert(b, e);
-  }
-  template <class InputIterator>
-  btree_set_container(InputIterator b, InputIterator e,
-                      const allocator_type &alloc)
-      : btree_set_container(b, e, key_compare(), alloc) {}
-
-  // Initializer list constructors.
-  btree_set_container(std::initializer_list<init_type> init,
-                      const key_compare &comp = key_compare(),
-                      const allocator_type &alloc = allocator_type())
-      : btree_set_container(init.begin(), init.end(), comp, alloc) {}
-  btree_set_container(std::initializer_list<init_type> init,
-                      const allocator_type &alloc)
-      : btree_set_container(init.begin(), init.end(), alloc) {}
-
-  // Insertion routines.
-  std::pair<iterator, bool> insert(const value_type &v)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return this->tree_.insert_unique(params_type::key(v), v);
-  }
-  std::pair<iterator, bool> insert(value_type &&v)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return this->tree_.insert_unique(params_type::key(v), std::move(v));
-  }
-  template <typename... Args>
-  std::pair<iterator, bool> emplace(Args &&...args)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    // Use a node handle to manage a temp slot.
-    auto node = CommonAccess::Construct<node_type>(this->get_allocator(),
-                                                   std::forward<Args>(args)...);
-    auto *slot = CommonAccess::GetSlot(node);
-    return this->tree_.insert_unique(params_type::key(slot), slot);
-  }
-  iterator insert(const_iterator hint,
-                  const value_type &v) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return this->tree_
-        .insert_hint_unique(iterator(hint), params_type::key(v), v)
-        .first;
-  }
-  iterator insert(const_iterator hint,
-                  value_type &&v) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return this->tree_
-        .insert_hint_unique(iterator(hint), params_type::key(v), std::move(v))
-        .first;
-  }
-  template <typename... Args>
-  iterator emplace_hint(const_iterator hint,
-                        Args &&...args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    // Use a node handle to manage a temp slot.
-    auto node = CommonAccess::Construct<node_type>(this->get_allocator(),
-                                                   std::forward<Args>(args)...);
-    auto *slot = CommonAccess::GetSlot(node);
-    return this->tree_
-        .insert_hint_unique(iterator(hint), params_type::key(slot), slot)
-        .first;
-  }
-  template <typename InputIterator>
-  void insert(InputIterator b, InputIterator e) {
-    this->tree_.insert_iterator_unique(b, e, 0);
-  }
-  void insert(std::initializer_list<init_type> init) {
-    this->tree_.insert_iterator_unique(init.begin(), init.end(), 0);
-  }
-  insert_return_type insert(node_type &&node) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    if (!node) return {this->end(), false, node_type()};
-    std::pair<iterator, bool> res =
-        this->tree_.insert_unique(params_type::key(CommonAccess::GetSlot(node)),
-                                  CommonAccess::GetSlot(node));
-    if (res.second) {
-      CommonAccess::Destroy(&node);
-      return {res.first, true, node_type()};
-    } else {
-      return {res.first, false, std::move(node)};
-    }
-  }
-  iterator insert(const_iterator hint,
-                  node_type &&node) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    if (!node) return this->end();
-    std::pair<iterator, bool> res = this->tree_.insert_hint_unique(
-        iterator(hint), params_type::key(CommonAccess::GetSlot(node)),
-        CommonAccess::GetSlot(node));
-    if (res.second) CommonAccess::Destroy(&node);
-    return res.first;
-  }
-
-  // Node extraction routines.
-  template <typename K = key_type>
-  node_type extract(const key_arg<K> &key) {
-    const std::pair<iterator, bool> lower_and_equal =
-        this->tree_.lower_bound_equal(key);
-    return lower_and_equal.second ? extract(lower_and_equal.first)
-                                  : node_type();
-  }
-  using super_type::extract;
-
-  // Merge routines.
-  // Moves elements from `src` into `this`. If the element already exists in
-  // `this`, it is left unmodified in `src`.
-  template <
-      typename T,
-      typename absl::enable_if_t<
-          absl::conjunction<
-              std::is_same<value_type, typename T::value_type>,
-              std::is_same<allocator_type, typename T::allocator_type>,
-              std::is_same<typename params_type::is_map_container,
-                           typename T::params_type::is_map_container>>::value,
-          int> = 0>
-  void merge(btree_container<T> &src) {  // NOLINT
-    for (auto src_it = src.begin(); src_it != src.end();) {
-      if (insert(std::move(params_type::element(src_it.slot()))).second) {
-        src_it = src.erase(src_it);
-      } else {
-        ++src_it;
-      }
-    }
-  }
-
-  template <
-      typename T,
-      typename absl::enable_if_t<
-          absl::conjunction<
-              std::is_same<value_type, typename T::value_type>,
-              std::is_same<allocator_type, typename T::allocator_type>,
-              std::is_same<typename params_type::is_map_container,
-                           typename T::params_type::is_map_container>>::value,
-          int> = 0>
-  void merge(btree_container<T> &&src) {
-    merge(src);
-  }
-};
-
-// Base class for btree_map.
-template <typename Tree>
-class btree_map_container : public btree_set_container<Tree> {
-  using super_type = btree_set_container<Tree>;
-  using params_type = typename Tree::params_type;
-  friend class BtreeNodePeer;
-
- private:
-  template <class K>
-  using key_arg = typename super_type::template key_arg<K>;
-
- public:
-  using key_type = typename Tree::key_type;
-  using mapped_type = typename params_type::mapped_type;
-  using value_type = typename Tree::value_type;
-  using key_compare = typename Tree::original_key_compare;
-  using allocator_type = typename Tree::allocator_type;
-  using iterator = typename Tree::iterator;
-  using const_iterator = typename Tree::const_iterator;
-
-  // Inherit constructors.
-  using super_type::super_type;
-  btree_map_container() {}
-
-  // Insertion routines.
-  // Note: the nullptr template arguments and extra `const M&` overloads allow
-  // for supporting bitfield arguments.
-  template <typename K = key_type, class M>
-  std::pair<iterator, bool> insert_or_assign(const key_arg<K> &k, const M &obj)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign_impl(k, obj);
-  }
-  template <typename K = key_type, class M, K * = nullptr>
-  std::pair<iterator, bool> insert_or_assign(key_arg<K> &&k, const M &obj)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign_impl(std::forward<K>(k), obj);
-  }
-  template <typename K = key_type, class M, M * = nullptr>
-  std::pair<iterator, bool> insert_or_assign(const key_arg<K> &k, M &&obj)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign_impl(k, std::forward<M>(obj));
-  }
-  template <typename K = key_type, class M, K * = nullptr, M * = nullptr>
-  std::pair<iterator, bool> insert_or_assign(key_arg<K> &&k, M &&obj)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign_impl(std::forward<K>(k), std::forward<M>(obj));
-  }
-  template <typename K = key_type, class M>
-  iterator insert_or_assign(const_iterator hint, const key_arg<K> &k,
-                            const M &obj) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign_hint_impl(hint, k, obj);
-  }
-  template <typename K = key_type, class M, K * = nullptr>
-  iterator insert_or_assign(const_iterator hint, key_arg<K> &&k,
-                            const M &obj) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign_hint_impl(hint, std::forward<K>(k), obj);
-  }
-  template <typename K = key_type, class M, M * = nullptr>
-  iterator insert_or_assign(const_iterator hint, const key_arg<K> &k,
-                            M &&obj) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign_hint_impl(hint, k, std::forward<M>(obj));
-  }
-  template <typename K = key_type, class M, K * = nullptr, M * = nullptr>
-  iterator insert_or_assign(const_iterator hint, key_arg<K> &&k,
-                            M &&obj) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign_hint_impl(hint, std::forward<K>(k),
-                                      std::forward<M>(obj));
-  }
-
-  template <typename K = key_type, typename... Args,
-            typename absl::enable_if_t<
-                !std::is_convertible<K, const_iterator>::value, int> = 0>
-  std::pair<iterator, bool> try_emplace(const key_arg<K> &k, Args &&...args)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return try_emplace_impl(k, std::forward<Args>(args)...);
-  }
-  template <typename K = key_type, typename... Args,
-            typename absl::enable_if_t<
-                !std::is_convertible<K, const_iterator>::value, int> = 0>
-  std::pair<iterator, bool> try_emplace(key_arg<K> &&k, Args &&...args)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return try_emplace_impl(std::forward<K>(k), std::forward<Args>(args)...);
-  }
-  template <typename K = key_type, typename... Args>
-  iterator try_emplace(const_iterator hint, const key_arg<K> &k,
-                       Args &&...args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return try_emplace_hint_impl(hint, k, std::forward<Args>(args)...);
-  }
-  template <typename K = key_type, typename... Args>
-  iterator try_emplace(const_iterator hint, key_arg<K> &&k,
-                       Args &&...args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return try_emplace_hint_impl(hint, std::forward<K>(k),
-                                 std::forward<Args>(args)...);
-  }
-
-  template <typename K = key_type>
-  mapped_type &operator[](const key_arg<K> &k) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return try_emplace(k).first->second;
-  }
-  template <typename K = key_type>
-  mapped_type &operator[](key_arg<K> &&k) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return try_emplace(std::forward<K>(k)).first->second;
-  }
-
-  template <typename K = key_type>
-  mapped_type &at(const key_arg<K> &key) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    auto it = this->find(key);
-    if (it == this->end())
-      base_internal::ThrowStdOutOfRange("absl::btree_map::at");
-    return it->second;
-  }
-  template <typename K = key_type>
-  const mapped_type &at(const key_arg<K> &key) const
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    auto it = this->find(key);
-    if (it == this->end())
-      base_internal::ThrowStdOutOfRange("absl::btree_map::at");
-    return it->second;
-  }
-
- private:
-  // Note: when we call `std::forward<M>(obj)` twice, it's safe because
-  // insert_unique/insert_hint_unique are guaranteed to not consume `obj` when
-  // `ret.second` is false.
-  template <class K, class M>
-  std::pair<iterator, bool> insert_or_assign_impl(K &&k, M &&obj) {
-    const std::pair<iterator, bool> ret =
-        this->tree_.insert_unique(k, std::forward<K>(k), std::forward<M>(obj));
-    if (!ret.second) ret.first->second = std::forward<M>(obj);
-    return ret;
-  }
-  template <class K, class M>
-  iterator insert_or_assign_hint_impl(const_iterator hint, K &&k, M &&obj) {
-    const std::pair<iterator, bool> ret = this->tree_.insert_hint_unique(
-        iterator(hint), k, std::forward<K>(k), std::forward<M>(obj));
-    if (!ret.second) ret.first->second = std::forward<M>(obj);
-    return ret.first;
-  }
-
-  template <class K, class... Args>
-  std::pair<iterator, bool> try_emplace_impl(K &&k, Args &&... args) {
-    return this->tree_.insert_unique(
-        k, std::piecewise_construct, std::forward_as_tuple(std::forward<K>(k)),
-        std::forward_as_tuple(std::forward<Args>(args)...));
-  }
-  template <class K, class... Args>
-  iterator try_emplace_hint_impl(const_iterator hint, K &&k, Args &&... args) {
-    return this->tree_
-        .insert_hint_unique(iterator(hint), k, std::piecewise_construct,
-                            std::forward_as_tuple(std::forward<K>(k)),
-                            std::forward_as_tuple(std::forward<Args>(args)...))
-        .first;
-  }
-};
-
-// A common base class for btree_multiset and btree_multimap.
-template <typename Tree>
-class btree_multiset_container : public btree_container<Tree> {
-  using super_type = btree_container<Tree>;
-  using params_type = typename Tree::params_type;
-  using init_type = typename params_type::init_type;
-  using is_key_compare_to = typename params_type::is_key_compare_to;
-  friend class BtreeNodePeer;
-
-  template <class K>
-  using key_arg = typename super_type::template key_arg<K>;
-
- public:
-  using key_type = typename Tree::key_type;
-  using value_type = typename Tree::value_type;
-  using size_type = typename Tree::size_type;
-  using key_compare = typename Tree::original_key_compare;
-  using allocator_type = typename Tree::allocator_type;
-  using iterator = typename Tree::iterator;
-  using const_iterator = typename Tree::const_iterator;
-  using node_type = typename super_type::node_type;
-
-  // Inherit constructors.
-  using super_type::super_type;
-  btree_multiset_container() {}
-
-  // Range constructors.
-  template <class InputIterator>
-  btree_multiset_container(InputIterator b, InputIterator e,
-                           const key_compare &comp = key_compare(),
-                           const allocator_type &alloc = allocator_type())
-      : super_type(comp, alloc) {
-    insert(b, e);
-  }
-  template <class InputIterator>
-  btree_multiset_container(InputIterator b, InputIterator e,
-                           const allocator_type &alloc)
-      : btree_multiset_container(b, e, key_compare(), alloc) {}
-
-  // Initializer list constructors.
-  btree_multiset_container(std::initializer_list<init_type> init,
-                           const key_compare &comp = key_compare(),
-                           const allocator_type &alloc = allocator_type())
-      : btree_multiset_container(init.begin(), init.end(), comp, alloc) {}
-  btree_multiset_container(std::initializer_list<init_type> init,
-                           const allocator_type &alloc)
-      : btree_multiset_container(init.begin(), init.end(), alloc) {}
-
-  // Insertion routines.
-  iterator insert(const value_type &v) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return this->tree_.insert_multi(v);
-  }
-  iterator insert(value_type &&v) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return this->tree_.insert_multi(std::move(v));
-  }
-  iterator insert(const_iterator hint,
-                  const value_type &v) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return this->tree_.insert_hint_multi(iterator(hint), v);
-  }
-  iterator insert(const_iterator hint,
-                  value_type &&v) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return this->tree_.insert_hint_multi(iterator(hint), std::move(v));
-  }
-  template <typename InputIterator>
-  void insert(InputIterator b, InputIterator e) {
-    this->tree_.insert_iterator_multi(b, e);
-  }
-  void insert(std::initializer_list<init_type> init) {
-    this->tree_.insert_iterator_multi(init.begin(), init.end());
-  }
-  template <typename... Args>
-  iterator emplace(Args &&...args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    // Use a node handle to manage a temp slot.
-    auto node = CommonAccess::Construct<node_type>(this->get_allocator(),
-                                                   std::forward<Args>(args)...);
-    return this->tree_.insert_multi(CommonAccess::GetSlot(node));
-  }
-  template <typename... Args>
-  iterator emplace_hint(const_iterator hint,
-                        Args &&...args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    // Use a node handle to manage a temp slot.
-    auto node = CommonAccess::Construct<node_type>(this->get_allocator(),
-                                                   std::forward<Args>(args)...);
-    return this->tree_.insert_hint_multi(iterator(hint),
-                                         CommonAccess::GetSlot(node));
-  }
-  iterator insert(node_type &&node) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    if (!node) return this->end();
-    iterator res =
-        this->tree_.insert_multi(params_type::key(CommonAccess::GetSlot(node)),
-                                 CommonAccess::GetSlot(node));
-    CommonAccess::Destroy(&node);
-    return res;
-  }
-  iterator insert(const_iterator hint,
-                  node_type &&node) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    if (!node) return this->end();
-    iterator res = this->tree_.insert_hint_multi(
-        iterator(hint),
-        std::move(params_type::element(CommonAccess::GetSlot(node))));
-    CommonAccess::Destroy(&node);
-    return res;
-  }
-
-  // Node extraction routines.
-  template <typename K = key_type>
-  node_type extract(const key_arg<K> &key) {
-    const std::pair<iterator, bool> lower_and_equal =
-        this->tree_.lower_bound_equal(key);
-    return lower_and_equal.second ? extract(lower_and_equal.first)
-                                  : node_type();
-  }
-  using super_type::extract;
-
-  // Merge routines.
-  // Moves all elements from `src` into `this`.
-  template <
-      typename T,
-      typename absl::enable_if_t<
-          absl::conjunction<
-              std::is_same<value_type, typename T::value_type>,
-              std::is_same<allocator_type, typename T::allocator_type>,
-              std::is_same<typename params_type::is_map_container,
-                           typename T::params_type::is_map_container>>::value,
-          int> = 0>
-  void merge(btree_container<T> &src) {  // NOLINT
-    for (auto src_it = src.begin(), end = src.end(); src_it != end; ++src_it) {
-      insert(std::move(params_type::element(src_it.slot())));
-    }
-    src.clear();
-  }
-
-  template <
-      typename T,
-      typename absl::enable_if_t<
-          absl::conjunction<
-              std::is_same<value_type, typename T::value_type>,
-              std::is_same<allocator_type, typename T::allocator_type>,
-              std::is_same<typename params_type::is_map_container,
-                           typename T::params_type::is_map_container>>::value,
-          int> = 0>
-  void merge(btree_container<T> &&src) {
-    merge(src);
-  }
-};
-
-// A base class for btree_multimap.
-template <typename Tree>
-class btree_multimap_container : public btree_multiset_container<Tree> {
-  using super_type = btree_multiset_container<Tree>;
-  using params_type = typename Tree::params_type;
-  friend class BtreeNodePeer;
-
- public:
-  using mapped_type = typename params_type::mapped_type;
-
-  // Inherit constructors.
-  using super_type::super_type;
-  btree_multimap_container() {}
-};
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_BTREE_CONTAINER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/common.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/common.h
deleted file mode 100644
index 9239bb4d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/common.h
+++ /dev/null
@@ -1,207 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_COMMON_H_
-#define ABSL_CONTAINER_INTERNAL_COMMON_H_
-
-#include <cassert>
-#include <type_traits>
-
-#include "absl/meta/type_traits.h"
-#include "absl/types/optional.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-template <class, class = void>
-struct IsTransparent : std::false_type {};
-template <class T>
-struct IsTransparent<T, absl::void_t<typename T::is_transparent>>
-    : std::true_type {};
-
-template <bool is_transparent>
-struct KeyArg {
-  // Transparent. Forward `K`.
-  template <typename K, typename key_type>
-  using type = K;
-};
-
-template <>
-struct KeyArg<false> {
-  // Not transparent. Always use `key_type`.
-  template <typename K, typename key_type>
-  using type = key_type;
-};
-
-// The node_handle concept from C++17.
-// We specialize node_handle for sets and maps. node_handle_base holds the
-// common API of both.
-template <typename PolicyTraits, typename Alloc>
-class node_handle_base {
- protected:
-  using slot_type = typename PolicyTraits::slot_type;
-
- public:
-  using allocator_type = Alloc;
-
-  constexpr node_handle_base() = default;
-  node_handle_base(node_handle_base&& other) noexcept {
-    *this = std::move(other);
-  }
-  ~node_handle_base() { destroy(); }
-  node_handle_base& operator=(node_handle_base&& other) noexcept {
-    destroy();
-    if (!other.empty()) {
-      alloc_ = other.alloc_;
-      PolicyTraits::transfer(alloc(), slot(), other.slot());
-      other.reset();
-    }
-    return *this;
-  }
-
-  bool empty() const noexcept { return !alloc_; }
-  explicit operator bool() const noexcept { return !empty(); }
-  allocator_type get_allocator() const { return *alloc_; }
-
- protected:
-  friend struct CommonAccess;
-
-  struct transfer_tag_t {};
-  node_handle_base(transfer_tag_t, const allocator_type& a, slot_type* s)
-      : alloc_(a) {
-    PolicyTraits::transfer(alloc(), slot(), s);
-  }
-
-  struct construct_tag_t {};
-  template <typename... Args>
-  node_handle_base(construct_tag_t, const allocator_type& a, Args&&... args)
-      : alloc_(a) {
-    PolicyTraits::construct(alloc(), slot(), std::forward<Args>(args)...);
-  }
-
-  void destroy() {
-    if (!empty()) {
-      PolicyTraits::destroy(alloc(), slot());
-      reset();
-    }
-  }
-
-  void reset() {
-    assert(alloc_.has_value());
-    alloc_ = absl::nullopt;
-  }
-
-  slot_type* slot() const {
-    assert(!empty());
-    return reinterpret_cast<slot_type*>(std::addressof(slot_space_));
-  }
-  allocator_type* alloc() { return std::addressof(*alloc_); }
-
- private:
-  absl::optional<allocator_type> alloc_ = {};
-  alignas(slot_type) mutable unsigned char slot_space_[sizeof(slot_type)] = {};
-};
-
-// For sets.
-template <typename Policy, typename PolicyTraits, typename Alloc,
-          typename = void>
-class node_handle : public node_handle_base<PolicyTraits, Alloc> {
-  using Base = node_handle_base<PolicyTraits, Alloc>;
-
- public:
-  using value_type = typename PolicyTraits::value_type;
-
-  constexpr node_handle() {}
-
-  value_type& value() const { return PolicyTraits::element(this->slot()); }
-
- private:
-  friend struct CommonAccess;
-
-  using Base::Base;
-};
-
-// For maps.
-template <typename Policy, typename PolicyTraits, typename Alloc>
-class node_handle<Policy, PolicyTraits, Alloc,
-                  absl::void_t<typename Policy::mapped_type>>
-    : public node_handle_base<PolicyTraits, Alloc> {
-  using Base = node_handle_base<PolicyTraits, Alloc>;
-  using slot_type = typename PolicyTraits::slot_type;
-
- public:
-  using key_type = typename Policy::key_type;
-  using mapped_type = typename Policy::mapped_type;
-
-  constexpr node_handle() {}
-
-  // When C++17 is available, we can use std::launder to provide mutable
-  // access to the key. Otherwise, we provide const access.
-  auto key() const
-      -> decltype(PolicyTraits::mutable_key(std::declval<slot_type*>())) {
-    return PolicyTraits::mutable_key(this->slot());
-  }
-
-  mapped_type& mapped() const {
-    return PolicyTraits::value(&PolicyTraits::element(this->slot()));
-  }
-
- private:
-  friend struct CommonAccess;
-
-  using Base::Base;
-};
-
-// Provide access to non-public node-handle functions.
-struct CommonAccess {
-  template <typename Node>
-  static auto GetSlot(const Node& node) -> decltype(node.slot()) {
-    return node.slot();
-  }
-
-  template <typename Node>
-  static void Destroy(Node* node) {
-    node->destroy();
-  }
-
-  template <typename Node>
-  static void Reset(Node* node) {
-    node->reset();
-  }
-
-  template <typename T, typename... Args>
-  static T Transfer(Args&&... args) {
-    return T(typename T::transfer_tag_t{}, std::forward<Args>(args)...);
-  }
-
-  template <typename T, typename... Args>
-  static T Construct(Args&&... args) {
-    return T(typename T::construct_tag_t{}, std::forward<Args>(args)...);
-  }
-};
-
-// Implement the insert_return_type<> concept of C++17.
-template <class Iterator, class NodeType>
-struct InsertReturnType {
-  Iterator position;
-  bool inserted;
-  NodeType node;
-};
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_COMMON_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/common_policy_traits.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/common_policy_traits.h
deleted file mode 100644
index c521f612..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/common_policy_traits.h
+++ /dev/null
@@ -1,143 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_COMMON_POLICY_TRAITS_H_
-#define ABSL_CONTAINER_INTERNAL_COMMON_POLICY_TRAITS_H_
-
-#include <cstddef>
-#include <cstring>
-#include <memory>
-#include <new>
-#include <type_traits>
-#include <utility>
-
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-// Defines how slots are initialized/destroyed/moved.
-template <class Policy, class = void>
-struct common_policy_traits {
-  // The actual object stored in the container.
-  using slot_type = typename Policy::slot_type;
-  using reference = decltype(Policy::element(std::declval<slot_type*>()));
-  using value_type = typename std::remove_reference<reference>::type;
-
-  // PRECONDITION: `slot` is UNINITIALIZED
-  // POSTCONDITION: `slot` is INITIALIZED
-  template <class Alloc, class... Args>
-  static void construct(Alloc* alloc, slot_type* slot, Args&&... args) {
-    Policy::construct(alloc, slot, std::forward<Args>(args)...);
-  }
-
-  // PRECONDITION: `slot` is INITIALIZED
-  // POSTCONDITION: `slot` is UNINITIALIZED
-  // Returns std::true_type in case destroy is trivial.
-  template <class Alloc>
-  static auto destroy(Alloc* alloc, slot_type* slot) {
-    return Policy::destroy(alloc, slot);
-  }
-
-  // Transfers the `old_slot` to `new_slot`. Any memory allocated by the
-  // allocator inside `old_slot` to `new_slot` can be transferred.
-  //
-  // OPTIONAL: defaults to:
-  //
-  //     clone(new_slot, std::move(*old_slot));
-  //     destroy(old_slot);
-  //
-  // PRECONDITION: `new_slot` is UNINITIALIZED and `old_slot` is INITIALIZED
-  // POSTCONDITION: `new_slot` is INITIALIZED and `old_slot` is
-  //                UNINITIALIZED
-  template <class Alloc>
-  static void transfer(Alloc* alloc, slot_type* new_slot, slot_type* old_slot) {
-    transfer_impl(alloc, new_slot, old_slot, Rank2{});
-  }
-
-  // PRECONDITION: `slot` is INITIALIZED
-  // POSTCONDITION: `slot` is INITIALIZED
-  // Note: we use remove_const_t so that the two overloads have different args
-  // in the case of sets with explicitly const value_types.
-  template <class P = Policy>
-  static auto element(absl::remove_const_t<slot_type>* slot)
-      -> decltype(P::element(slot)) {
-    return P::element(slot);
-  }
-  template <class P = Policy>
-  static auto element(const slot_type* slot) -> decltype(P::element(slot)) {
-    return P::element(slot);
-  }
-
-  static constexpr bool transfer_uses_memcpy() {
-    return std::is_same<decltype(transfer_impl<std::allocator<char>>(
-                            nullptr, nullptr, nullptr, Rank2{})),
-                        std::true_type>::value;
-  }
-
-  // Returns true if destroy is trivial and can be omitted.
-  template <class Alloc>
-  static constexpr bool destroy_is_trivial() {
-    return std::is_same<decltype(destroy<Alloc>(nullptr, nullptr)),
-                        std::true_type>::value;
-  }
-
- private:
-  // Use go/ranked-overloads for dispatching.
-  struct Rank0 {};
-  struct Rank1 : Rank0 {};
-  struct Rank2 : Rank1 {};
-
-  // Use auto -> decltype as an enabler.
-  // P::transfer returns std::true_type if transfer uses memcpy (e.g. in
-  // node_slot_policy).
-  template <class Alloc, class P = Policy>
-  static auto transfer_impl(Alloc* alloc, slot_type* new_slot,
-                            slot_type* old_slot,
-                            Rank2) -> decltype(P::transfer(alloc, new_slot,
-                                                           old_slot)) {
-    return P::transfer(alloc, new_slot, old_slot);
-  }
-#if defined(__cpp_lib_launder) && __cpp_lib_launder >= 201606
-  // This overload returns true_type for the trait below.
-  // The conditional_t is to make the enabler type dependent.
-  template <class Alloc,
-            typename = std::enable_if_t<absl::is_trivially_relocatable<
-                std::conditional_t<false, Alloc, value_type>>::value>>
-  static std::true_type transfer_impl(Alloc*, slot_type* new_slot,
-                                      slot_type* old_slot, Rank1) {
-    // TODO(b/247130232): remove casts after fixing warnings.
-    // TODO(b/251814870): remove casts after fixing warnings.
-    std::memcpy(
-        static_cast<void*>(std::launder(
-            const_cast<std::remove_const_t<value_type>*>(&element(new_slot)))),
-        static_cast<const void*>(&element(old_slot)), sizeof(value_type));
-    return {};
-  }
-#endif
-
-  template <class Alloc>
-  static void transfer_impl(Alloc* alloc, slot_type* new_slot,
-                            slot_type* old_slot, Rank0) {
-    construct(alloc, new_slot, std::move(element(old_slot)));
-    destroy(alloc, old_slot);
-  }
-};
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_COMMON_POLICY_TRAITS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/compressed_tuple.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/compressed_tuple.h
deleted file mode 100644
index f05a1fdc..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/compressed_tuple.h
+++ /dev/null
@@ -1,271 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Helper class to perform the Empty Base Optimization.
-// Ts can contain classes and non-classes, empty or not. For the ones that
-// are empty classes, we perform the optimization. If all types in Ts are empty
-// classes, then CompressedTuple<Ts...> is itself an empty class.
-//
-// To access the members, use member get<N>() function.
-//
-// Eg:
-//   absl::container_internal::CompressedTuple<int, T1, T2, T3> value(7, t1, t2,
-//                                                                    t3);
-//   assert(value.get<0>() == 7);
-//   T1& t1 = value.get<1>();
-//   const T2& t2 = value.get<2>();
-//   ...
-//
-// https://en.cppreference.com/w/cpp/language/ebo
-
-#ifndef ABSL_CONTAINER_INTERNAL_COMPRESSED_TUPLE_H_
-#define ABSL_CONTAINER_INTERNAL_COMPRESSED_TUPLE_H_
-
-#include <initializer_list>
-#include <tuple>
-#include <type_traits>
-#include <utility>
-
-#include "absl/utility/utility.h"
-
-#if defined(_MSC_VER) && !defined(__NVCC__)
-// We need to mark these classes with this declspec to ensure that
-// CompressedTuple happens.
-#define ABSL_INTERNAL_COMPRESSED_TUPLE_DECLSPEC __declspec(empty_bases)
-#else
-#define ABSL_INTERNAL_COMPRESSED_TUPLE_DECLSPEC
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-template <typename... Ts>
-class CompressedTuple;
-
-namespace internal_compressed_tuple {
-
-template <typename D, size_t I>
-struct Elem;
-template <typename... B, size_t I>
-struct Elem<CompressedTuple<B...>, I>
-    : std::tuple_element<I, std::tuple<B...>> {};
-template <typename D, size_t I>
-using ElemT = typename Elem<D, I>::type;
-
-// We can't use EBCO on other CompressedTuples because that would mean that we
-// derive from multiple Storage<> instantiations with the same I parameter,
-// and potentially from multiple identical Storage<> instantiations.  So anytime
-// we use type inheritance rather than encapsulation, we mark
-// CompressedTupleImpl, to make this easy to detect.
-struct uses_inheritance {};
-
-template <typename T>
-constexpr bool ShouldUseBase() {
-  return std::is_class<T>::value && std::is_empty<T>::value &&
-         !std::is_final<T>::value &&
-         !std::is_base_of<uses_inheritance, T>::value;
-}
-
-// The storage class provides two specializations:
-//  - For empty classes, it stores T as a base class.
-//  - For everything else, it stores T as a member.
-template <typename T, size_t I, bool UseBase = ShouldUseBase<T>()>
-struct Storage {
-  T value;
-  constexpr Storage() = default;
-  template <typename V>
-  explicit constexpr Storage(absl::in_place_t, V&& v)
-      : value(std::forward<V>(v)) {}
-  constexpr const T& get() const& { return value; }
-  T& get() & { return value; }
-  constexpr const T&& get() const&& { return std::move(*this).value; }
-  T&& get() && { return std::move(*this).value; }
-};
-
-template <typename T, size_t I>
-struct ABSL_INTERNAL_COMPRESSED_TUPLE_DECLSPEC Storage<T, I, true> : T {
-  constexpr Storage() = default;
-
-  template <typename V>
-  explicit constexpr Storage(absl::in_place_t, V&& v) : T(std::forward<V>(v)) {}
-
-  constexpr const T& get() const& { return *this; }
-  T& get() & { return *this; }
-  constexpr const T&& get() const&& { return std::move(*this); }
-  T&& get() && { return std::move(*this); }
-};
-
-template <typename D, typename I, bool ShouldAnyUseBase>
-struct ABSL_INTERNAL_COMPRESSED_TUPLE_DECLSPEC CompressedTupleImpl;
-
-template <typename... Ts, size_t... I, bool ShouldAnyUseBase>
-struct ABSL_INTERNAL_COMPRESSED_TUPLE_DECLSPEC CompressedTupleImpl<
-    CompressedTuple<Ts...>, absl::index_sequence<I...>, ShouldAnyUseBase>
-    // We use the dummy identity function through std::integral_constant to
-    // convince MSVC of accepting and expanding I in that context. Without it
-    // you would get:
-    //   error C3548: 'I': parameter pack cannot be used in this context
-    : uses_inheritance,
-      Storage<Ts, std::integral_constant<size_t, I>::value>... {
-  constexpr CompressedTupleImpl() = default;
-  template <typename... Vs>
-  explicit constexpr CompressedTupleImpl(absl::in_place_t, Vs&&... args)
-      : Storage<Ts, I>(absl::in_place, std::forward<Vs>(args))... {}
-  friend CompressedTuple<Ts...>;
-};
-
-template <typename... Ts, size_t... I>
-struct ABSL_INTERNAL_COMPRESSED_TUPLE_DECLSPEC CompressedTupleImpl<
-    CompressedTuple<Ts...>, absl::index_sequence<I...>, false>
-    // We use the dummy identity function as above...
-    : Storage<Ts, std::integral_constant<size_t, I>::value, false>... {
-  constexpr CompressedTupleImpl() = default;
-  template <typename... Vs>
-  explicit constexpr CompressedTupleImpl(absl::in_place_t, Vs&&... args)
-      : Storage<Ts, I, false>(absl::in_place, std::forward<Vs>(args))... {}
-  friend CompressedTuple<Ts...>;
-};
-
-std::false_type Or(std::initializer_list<std::false_type>);
-std::true_type Or(std::initializer_list<bool>);
-
-// MSVC requires this to be done separately rather than within the declaration
-// of CompressedTuple below.
-template <typename... Ts>
-constexpr bool ShouldAnyUseBase() {
-  return decltype(
-      Or({std::integral_constant<bool, ShouldUseBase<Ts>()>()...})){};
-}
-
-template <typename T, typename V>
-using TupleElementMoveConstructible =
-    typename std::conditional<std::is_reference<T>::value,
-                              std::is_convertible<V, T>,
-                              std::is_constructible<T, V&&>>::type;
-
-template <bool SizeMatches, class T, class... Vs>
-struct TupleMoveConstructible : std::false_type {};
-
-template <class... Ts, class... Vs>
-struct TupleMoveConstructible<true, CompressedTuple<Ts...>, Vs...>
-    : std::integral_constant<
-          bool, absl::conjunction<
-                    TupleElementMoveConstructible<Ts, Vs&&>...>::value> {};
-
-template <typename T>
-struct compressed_tuple_size;
-
-template <typename... Es>
-struct compressed_tuple_size<CompressedTuple<Es...>>
-    : public std::integral_constant<std::size_t, sizeof...(Es)> {};
-
-template <class T, class... Vs>
-struct TupleItemsMoveConstructible
-    : std::integral_constant<
-          bool, TupleMoveConstructible<compressed_tuple_size<T>::value ==
-                                           sizeof...(Vs),
-                                       T, Vs...>::value> {};
-
-}  // namespace internal_compressed_tuple
-
-// Helper class to perform the Empty Base Class Optimization.
-// Ts can contain classes and non-classes, empty or not. For the ones that
-// are empty classes, we perform the CompressedTuple. If all types in Ts are
-// empty classes, then CompressedTuple<Ts...> is itself an empty class.  (This
-// does not apply when one or more of those empty classes is itself an empty
-// CompressedTuple.)
-//
-// To access the members, use member .get<N>() function.
-//
-// Eg:
-//   absl::container_internal::CompressedTuple<int, T1, T2, T3> value(7, t1, t2,
-//                                                                    t3);
-//   assert(value.get<0>() == 7);
-//   T1& t1 = value.get<1>();
-//   const T2& t2 = value.get<2>();
-//   ...
-//
-// https://en.cppreference.com/w/cpp/language/ebo
-template <typename... Ts>
-class ABSL_INTERNAL_COMPRESSED_TUPLE_DECLSPEC CompressedTuple
-    : private internal_compressed_tuple::CompressedTupleImpl<
-          CompressedTuple<Ts...>, absl::index_sequence_for<Ts...>,
-          internal_compressed_tuple::ShouldAnyUseBase<Ts...>()> {
- private:
-  template <int I>
-  using ElemT = internal_compressed_tuple::ElemT<CompressedTuple, I>;
-
-  template <int I>
-  using StorageT = internal_compressed_tuple::Storage<ElemT<I>, I>;
-
- public:
-  // There seems to be a bug in MSVC dealing in which using '=default' here will
-  // cause the compiler to ignore the body of other constructors. The work-
-  // around is to explicitly implement the default constructor.
-#if defined(_MSC_VER)
-  constexpr CompressedTuple() : CompressedTuple::CompressedTupleImpl() {}
-#else
-  constexpr CompressedTuple() = default;
-#endif
-  explicit constexpr CompressedTuple(const Ts&... base)
-      : CompressedTuple::CompressedTupleImpl(absl::in_place, base...) {}
-
-  template <typename First, typename... Vs,
-            absl::enable_if_t<
-                absl::conjunction<
-                    // Ensure we are not hiding default copy/move constructors.
-                    absl::negation<std::is_same<void(CompressedTuple),
-                                                void(absl::decay_t<First>)>>,
-                    internal_compressed_tuple::TupleItemsMoveConstructible<
-                        CompressedTuple<Ts...>, First, Vs...>>::value,
-                bool> = true>
-  explicit constexpr CompressedTuple(First&& first, Vs&&... base)
-      : CompressedTuple::CompressedTupleImpl(absl::in_place,
-                                             std::forward<First>(first),
-                                             std::forward<Vs>(base)...) {}
-
-  template <int I>
-  ElemT<I>& get() & {
-    return StorageT<I>::get();
-  }
-
-  template <int I>
-  constexpr const ElemT<I>& get() const& {
-    return StorageT<I>::get();
-  }
-
-  template <int I>
-  ElemT<I>&& get() && {
-    return std::move(*this).StorageT<I>::get();
-  }
-
-  template <int I>
-  constexpr const ElemT<I>&& get() const&& {
-    return std::move(*this).StorageT<I>::get();
-  }
-};
-
-// Explicit specialization for a zero-element tuple
-// (needed to avoid ambiguous overloads for the default constructor).
-template <>
-class ABSL_INTERNAL_COMPRESSED_TUPLE_DECLSPEC CompressedTuple<> {};
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#undef ABSL_INTERNAL_COMPRESSED_TUPLE_DECLSPEC
-
-#endif  // ABSL_CONTAINER_INTERNAL_COMPRESSED_TUPLE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/container_memory.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/container_memory.h
deleted file mode 100644
index ba8e08a2..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/container_memory.h
+++ /dev/null
@@ -1,491 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_CONTAINER_MEMORY_H_
-#define ABSL_CONTAINER_INTERNAL_CONTAINER_MEMORY_H_
-
-#include <cassert>
-#include <cstddef>
-#include <cstring>
-#include <memory>
-#include <new>
-#include <tuple>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/config.h"
-#include "absl/memory/memory.h"
-#include "absl/meta/type_traits.h"
-#include "absl/utility/utility.h"
-
-#ifdef ABSL_HAVE_ADDRESS_SANITIZER
-#include <sanitizer/asan_interface.h>
-#endif
-
-#ifdef ABSL_HAVE_MEMORY_SANITIZER
-#include <sanitizer/msan_interface.h>
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-template <size_t Alignment>
-struct alignas(Alignment) AlignedType {};
-
-// Allocates at least n bytes aligned to the specified alignment.
-// Alignment must be a power of 2. It must be positive.
-//
-// Note that many allocators don't honor alignment requirements above certain
-// threshold (usually either alignof(std::max_align_t) or alignof(void*)).
-// Allocate() doesn't apply alignment corrections. If the underlying allocator
-// returns insufficiently alignment pointer, that's what you are going to get.
-template <size_t Alignment, class Alloc>
-void* Allocate(Alloc* alloc, size_t n) {
-  static_assert(Alignment > 0, "");
-  assert(n && "n must be positive");
-  using M = AlignedType<Alignment>;
-  using A = typename absl::allocator_traits<Alloc>::template rebind_alloc<M>;
-  using AT = typename absl::allocator_traits<Alloc>::template rebind_traits<M>;
-  // On macOS, "mem_alloc" is a #define with one argument defined in
-  // rpc/types.h, so we can't name the variable "mem_alloc" and initialize it
-  // with the "foo(bar)" syntax.
-  A my_mem_alloc(*alloc);
-  void* p = AT::allocate(my_mem_alloc, (n + sizeof(M) - 1) / sizeof(M));
-  assert(reinterpret_cast<uintptr_t>(p) % Alignment == 0 &&
-         "allocator does not respect alignment");
-  return p;
-}
-
-// Returns true if the destruction of the value with given Allocator will be
-// trivial.
-template <class Allocator, class ValueType>
-constexpr auto IsDestructionTrivial() {
-  constexpr bool result =
-      std::is_trivially_destructible<ValueType>::value &&
-      std::is_same<typename absl::allocator_traits<
-                       Allocator>::template rebind_alloc<char>,
-                   std::allocator<char>>::value;
-  return std::integral_constant<bool, result>();
-}
-
-// The pointer must have been previously obtained by calling
-// Allocate<Alignment>(alloc, n).
-template <size_t Alignment, class Alloc>
-void Deallocate(Alloc* alloc, void* p, size_t n) {
-  static_assert(Alignment > 0, "");
-  assert(n && "n must be positive");
-  using M = AlignedType<Alignment>;
-  using A = typename absl::allocator_traits<Alloc>::template rebind_alloc<M>;
-  using AT = typename absl::allocator_traits<Alloc>::template rebind_traits<M>;
-  // On macOS, "mem_alloc" is a #define with one argument defined in
-  // rpc/types.h, so we can't name the variable "mem_alloc" and initialize it
-  // with the "foo(bar)" syntax.
-  A my_mem_alloc(*alloc);
-  AT::deallocate(my_mem_alloc, static_cast<M*>(p),
-                 (n + sizeof(M) - 1) / sizeof(M));
-}
-
-namespace memory_internal {
-
-// Constructs T into uninitialized storage pointed by `ptr` using the args
-// specified in the tuple.
-template <class Alloc, class T, class Tuple, size_t... I>
-void ConstructFromTupleImpl(Alloc* alloc, T* ptr, Tuple&& t,
-                            absl::index_sequence<I...>) {
-  absl::allocator_traits<Alloc>::construct(
-      *alloc, ptr, std::get<I>(std::forward<Tuple>(t))...);
-}
-
-template <class T, class F>
-struct WithConstructedImplF {
-  template <class... Args>
-  decltype(std::declval<F>()(std::declval<T>())) operator()(
-      Args&&... args) const {
-    return std::forward<F>(f)(T(std::forward<Args>(args)...));
-  }
-  F&& f;
-};
-
-template <class T, class Tuple, size_t... Is, class F>
-decltype(std::declval<F>()(std::declval<T>())) WithConstructedImpl(
-    Tuple&& t, absl::index_sequence<Is...>, F&& f) {
-  return WithConstructedImplF<T, F>{std::forward<F>(f)}(
-      std::get<Is>(std::forward<Tuple>(t))...);
-}
-
-template <class T, size_t... Is>
-auto TupleRefImpl(T&& t, absl::index_sequence<Is...>)
-    -> decltype(std::forward_as_tuple(std::get<Is>(std::forward<T>(t))...)) {
-  return std::forward_as_tuple(std::get<Is>(std::forward<T>(t))...);
-}
-
-// Returns a tuple of references to the elements of the input tuple. T must be a
-// tuple.
-template <class T>
-auto TupleRef(T&& t) -> decltype(TupleRefImpl(
-    std::forward<T>(t),
-    absl::make_index_sequence<
-        std::tuple_size<typename std::decay<T>::type>::value>())) {
-  return TupleRefImpl(
-      std::forward<T>(t),
-      absl::make_index_sequence<
-          std::tuple_size<typename std::decay<T>::type>::value>());
-}
-
-template <class F, class K, class V>
-decltype(std::declval<F>()(std::declval<const K&>(), std::piecewise_construct,
-                           std::declval<std::tuple<K>>(), std::declval<V>()))
-DecomposePairImpl(F&& f, std::pair<std::tuple<K>, V> p) {
-  const auto& key = std::get<0>(p.first);
-  return std::forward<F>(f)(key, std::piecewise_construct, std::move(p.first),
-                            std::move(p.second));
-}
-
-}  // namespace memory_internal
-
-// Constructs T into uninitialized storage pointed by `ptr` using the args
-// specified in the tuple.
-template <class Alloc, class T, class Tuple>
-void ConstructFromTuple(Alloc* alloc, T* ptr, Tuple&& t) {
-  memory_internal::ConstructFromTupleImpl(
-      alloc, ptr, std::forward<Tuple>(t),
-      absl::make_index_sequence<
-          std::tuple_size<typename std::decay<Tuple>::type>::value>());
-}
-
-// Constructs T using the args specified in the tuple and calls F with the
-// constructed value.
-template <class T, class Tuple, class F>
-decltype(std::declval<F>()(std::declval<T>())) WithConstructed(Tuple&& t,
-                                                               F&& f) {
-  return memory_internal::WithConstructedImpl<T>(
-      std::forward<Tuple>(t),
-      absl::make_index_sequence<
-          std::tuple_size<typename std::decay<Tuple>::type>::value>(),
-      std::forward<F>(f));
-}
-
-// Given arguments of an std::pair's constructor, PairArgs() returns a pair of
-// tuples with references to the passed arguments. The tuples contain
-// constructor arguments for the first and the second elements of the pair.
-//
-// The following two snippets are equivalent.
-//
-// 1. std::pair<F, S> p(args...);
-//
-// 2. auto a = PairArgs(args...);
-//    std::pair<F, S> p(std::piecewise_construct,
-//                      std::move(a.first), std::move(a.second));
-inline std::pair<std::tuple<>, std::tuple<>> PairArgs() { return {}; }
-template <class F, class S>
-std::pair<std::tuple<F&&>, std::tuple<S&&>> PairArgs(F&& f, S&& s) {
-  return {std::piecewise_construct, std::forward_as_tuple(std::forward<F>(f)),
-          std::forward_as_tuple(std::forward<S>(s))};
-}
-template <class F, class S>
-std::pair<std::tuple<const F&>, std::tuple<const S&>> PairArgs(
-    const std::pair<F, S>& p) {
-  return PairArgs(p.first, p.second);
-}
-template <class F, class S>
-std::pair<std::tuple<F&&>, std::tuple<S&&>> PairArgs(std::pair<F, S>&& p) {
-  return PairArgs(std::forward<F>(p.first), std::forward<S>(p.second));
-}
-template <class F, class S>
-auto PairArgs(std::piecewise_construct_t, F&& f, S&& s)
-    -> decltype(std::make_pair(memory_internal::TupleRef(std::forward<F>(f)),
-                               memory_internal::TupleRef(std::forward<S>(s)))) {
-  return std::make_pair(memory_internal::TupleRef(std::forward<F>(f)),
-                        memory_internal::TupleRef(std::forward<S>(s)));
-}
-
-// A helper function for implementing apply() in map policies.
-template <class F, class... Args>
-auto DecomposePair(F&& f, Args&&... args)
-    -> decltype(memory_internal::DecomposePairImpl(
-        std::forward<F>(f), PairArgs(std::forward<Args>(args)...))) {
-  return memory_internal::DecomposePairImpl(
-      std::forward<F>(f), PairArgs(std::forward<Args>(args)...));
-}
-
-// A helper function for implementing apply() in set policies.
-template <class F, class Arg>
-decltype(std::declval<F>()(std::declval<const Arg&>(), std::declval<Arg>()))
-DecomposeValue(F&& f, Arg&& arg) {
-  const auto& key = arg;
-  return std::forward<F>(f)(key, std::forward<Arg>(arg));
-}
-
-// Helper functions for asan and msan.
-inline void SanitizerPoisonMemoryRegion(const void* m, size_t s) {
-#ifdef ABSL_HAVE_ADDRESS_SANITIZER
-  ASAN_POISON_MEMORY_REGION(m, s);
-#endif
-#ifdef ABSL_HAVE_MEMORY_SANITIZER
-  __msan_poison(m, s);
-#endif
-  (void)m;
-  (void)s;
-}
-
-inline void SanitizerUnpoisonMemoryRegion(const void* m, size_t s) {
-#ifdef ABSL_HAVE_ADDRESS_SANITIZER
-  ASAN_UNPOISON_MEMORY_REGION(m, s);
-#endif
-#ifdef ABSL_HAVE_MEMORY_SANITIZER
-  __msan_unpoison(m, s);
-#endif
-  (void)m;
-  (void)s;
-}
-
-template <typename T>
-inline void SanitizerPoisonObject(const T* object) {
-  SanitizerPoisonMemoryRegion(object, sizeof(T));
-}
-
-template <typename T>
-inline void SanitizerUnpoisonObject(const T* object) {
-  SanitizerUnpoisonMemoryRegion(object, sizeof(T));
-}
-
-namespace memory_internal {
-
-// If Pair is a standard-layout type, OffsetOf<Pair>::kFirst and
-// OffsetOf<Pair>::kSecond are equivalent to offsetof(Pair, first) and
-// offsetof(Pair, second) respectively. Otherwise they are -1.
-//
-// The purpose of OffsetOf is to avoid calling offsetof() on non-standard-layout
-// type, which is non-portable.
-template <class Pair, class = std::true_type>
-struct OffsetOf {
-  static constexpr size_t kFirst = static_cast<size_t>(-1);
-  static constexpr size_t kSecond = static_cast<size_t>(-1);
-};
-
-template <class Pair>
-struct OffsetOf<Pair, typename std::is_standard_layout<Pair>::type> {
-  static constexpr size_t kFirst = offsetof(Pair, first);
-  static constexpr size_t kSecond = offsetof(Pair, second);
-};
-
-template <class K, class V>
-struct IsLayoutCompatible {
- private:
-  struct Pair {
-    K first;
-    V second;
-  };
-
-  // Is P layout-compatible with Pair?
-  template <class P>
-  static constexpr bool LayoutCompatible() {
-    return std::is_standard_layout<P>() && sizeof(P) == sizeof(Pair) &&
-           alignof(P) == alignof(Pair) &&
-           memory_internal::OffsetOf<P>::kFirst ==
-               memory_internal::OffsetOf<Pair>::kFirst &&
-           memory_internal::OffsetOf<P>::kSecond ==
-               memory_internal::OffsetOf<Pair>::kSecond;
-  }
-
- public:
-  // Whether pair<const K, V> and pair<K, V> are layout-compatible. If they are,
-  // then it is safe to store them in a union and read from either.
-  static constexpr bool value = std::is_standard_layout<K>() &&
-                                std::is_standard_layout<Pair>() &&
-                                memory_internal::OffsetOf<Pair>::kFirst == 0 &&
-                                LayoutCompatible<std::pair<K, V>>() &&
-                                LayoutCompatible<std::pair<const K, V>>();
-};
-
-}  // namespace memory_internal
-
-// The internal storage type for key-value containers like flat_hash_map.
-//
-// It is convenient for the value_type of a flat_hash_map<K, V> to be
-// pair<const K, V>; the "const K" prevents accidental modification of the key
-// when dealing with the reference returned from find() and similar methods.
-// However, this creates other problems; we want to be able to emplace(K, V)
-// efficiently with move operations, and similarly be able to move a
-// pair<K, V> in insert().
-//
-// The solution is this union, which aliases the const and non-const versions
-// of the pair. This also allows flat_hash_map<const K, V> to work, even though
-// that has the same efficiency issues with move in emplace() and insert() -
-// but people do it anyway.
-//
-// If kMutableKeys is false, only the value member can be accessed.
-//
-// If kMutableKeys is true, key can be accessed through all slots while value
-// and mutable_value must be accessed only via INITIALIZED slots. Slots are
-// created and destroyed via mutable_value so that the key can be moved later.
-//
-// Accessing one of the union fields while the other is active is safe as
-// long as they are layout-compatible, which is guaranteed by the definition of
-// kMutableKeys. For C++11, the relevant section of the standard is
-// https://timsong-cpp.github.io/cppwp/n3337/class.mem#19 (9.2.19)
-template <class K, class V>
-union map_slot_type {
-  map_slot_type() {}
-  ~map_slot_type() = delete;
-  using value_type = std::pair<const K, V>;
-  using mutable_value_type =
-      std::pair<absl::remove_const_t<K>, absl::remove_const_t<V>>;
-
-  value_type value;
-  mutable_value_type mutable_value;
-  absl::remove_const_t<K> key;
-};
-
-template <class K, class V>
-struct map_slot_policy {
-  using slot_type = map_slot_type<K, V>;
-  using value_type = std::pair<const K, V>;
-  using mutable_value_type =
-      std::pair<absl::remove_const_t<K>, absl::remove_const_t<V>>;
-
- private:
-  static void emplace(slot_type* slot) {
-    // The construction of union doesn't do anything at runtime but it allows us
-    // to access its members without violating aliasing rules.
-    new (slot) slot_type;
-  }
-  // If pair<const K, V> and pair<K, V> are layout-compatible, we can accept one
-  // or the other via slot_type. We are also free to access the key via
-  // slot_type::key in this case.
-  using kMutableKeys = memory_internal::IsLayoutCompatible<K, V>;
-
- public:
-  static value_type& element(slot_type* slot) { return slot->value; }
-  static const value_type& element(const slot_type* slot) {
-    return slot->value;
-  }
-
-  // When C++17 is available, we can use std::launder to provide mutable
-  // access to the key for use in node handle.
-#if defined(__cpp_lib_launder) && __cpp_lib_launder >= 201606
-  static K& mutable_key(slot_type* slot) {
-    // Still check for kMutableKeys so that we can avoid calling std::launder
-    // unless necessary because it can interfere with optimizations.
-    return kMutableKeys::value ? slot->key
-                               : *std::launder(const_cast<K*>(
-                                     std::addressof(slot->value.first)));
-  }
-#else  // !(defined(__cpp_lib_launder) && __cpp_lib_launder >= 201606)
-  static const K& mutable_key(slot_type* slot) { return key(slot); }
-#endif
-
-  static const K& key(const slot_type* slot) {
-    return kMutableKeys::value ? slot->key : slot->value.first;
-  }
-
-  template <class Allocator, class... Args>
-  static void construct(Allocator* alloc, slot_type* slot, Args&&... args) {
-    emplace(slot);
-    if (kMutableKeys::value) {
-      absl::allocator_traits<Allocator>::construct(*alloc, &slot->mutable_value,
-                                                   std::forward<Args>(args)...);
-    } else {
-      absl::allocator_traits<Allocator>::construct(*alloc, &slot->value,
-                                                   std::forward<Args>(args)...);
-    }
-  }
-
-  // Construct this slot by moving from another slot.
-  template <class Allocator>
-  static void construct(Allocator* alloc, slot_type* slot, slot_type* other) {
-    emplace(slot);
-    if (kMutableKeys::value) {
-      absl::allocator_traits<Allocator>::construct(
-          *alloc, &slot->mutable_value, std::move(other->mutable_value));
-    } else {
-      absl::allocator_traits<Allocator>::construct(*alloc, &slot->value,
-                                                   std::move(other->value));
-    }
-  }
-
-  // Construct this slot by copying from another slot.
-  template <class Allocator>
-  static void construct(Allocator* alloc, slot_type* slot,
-                        const slot_type* other) {
-    emplace(slot);
-    absl::allocator_traits<Allocator>::construct(*alloc, &slot->value,
-                                                 other->value);
-  }
-
-  template <class Allocator>
-  static auto destroy(Allocator* alloc, slot_type* slot) {
-    if (kMutableKeys::value) {
-      absl::allocator_traits<Allocator>::destroy(*alloc, &slot->mutable_value);
-    } else {
-      absl::allocator_traits<Allocator>::destroy(*alloc, &slot->value);
-    }
-    return IsDestructionTrivial<Allocator, value_type>();
-  }
-
-  template <class Allocator>
-  static auto transfer(Allocator* alloc, slot_type* new_slot,
-                       slot_type* old_slot) {
-    auto is_relocatable =
-        typename absl::is_trivially_relocatable<value_type>::type();
-
-    emplace(new_slot);
-#if defined(__cpp_lib_launder) && __cpp_lib_launder >= 201606
-    if (is_relocatable) {
-      // TODO(b/247130232,b/251814870): remove casts after fixing warnings.
-      std::memcpy(static_cast<void*>(std::launder(&new_slot->value)),
-                  static_cast<const void*>(&old_slot->value),
-                  sizeof(value_type));
-      return is_relocatable;
-    }
-#endif
-
-    if (kMutableKeys::value) {
-      absl::allocator_traits<Allocator>::construct(
-          *alloc, &new_slot->mutable_value, std::move(old_slot->mutable_value));
-    } else {
-      absl::allocator_traits<Allocator>::construct(*alloc, &new_slot->value,
-                                                   std::move(old_slot->value));
-    }
-    destroy(alloc, old_slot);
-    return is_relocatable;
-  }
-};
-
-// Type erased function for computing hash of the slot.
-using HashSlotFn = size_t (*)(const void* hash_fn, void* slot);
-
-// Type erased function to apply `Fn` to data inside of the `slot`.
-// The data is expected to have type `T`.
-template <class Fn, class T>
-size_t TypeErasedApplyToSlotFn(const void* fn, void* slot) {
-  const auto* f = static_cast<const Fn*>(fn);
-  return (*f)(*static_cast<const T*>(slot));
-}
-
-// Type erased function to apply `Fn` to data inside of the `*slot_ptr`.
-// The data is expected to have type `T`.
-template <class Fn, class T>
-size_t TypeErasedDerefAndApplyToSlotFn(const void* fn, void* slot_ptr) {
-  const auto* f = static_cast<const Fn*>(fn);
-  const T* slot = *static_cast<const T**>(slot_ptr);
-  return (*f)(*slot);
-}
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_CONTAINER_MEMORY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_function_defaults.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_function_defaults.h
deleted file mode 100644
index 0f07bcfe..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_function_defaults.h
+++ /dev/null
@@ -1,276 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Define the default Hash and Eq functions for SwissTable containers.
-//
-// std::hash<T> and std::equal_to<T> are not appropriate hash and equal
-// functions for SwissTable containers. There are two reasons for this.
-//
-// SwissTable containers are power of 2 sized containers:
-//
-// This means they use the lower bits of the hash value to find the slot for
-// each entry. The typical hash function for integral types is the identity.
-// This is a very weak hash function for SwissTable and any power of 2 sized
-// hashtable implementation which will lead to excessive collisions. For
-// SwissTable we use murmur3 style mixing to reduce collisions to a minimum.
-//
-// SwissTable containers support heterogeneous lookup:
-//
-// In order to make heterogeneous lookup work, hash and equal functions must be
-// polymorphic. At the same time they have to satisfy the same requirements the
-// C++ standard imposes on hash functions and equality operators. That is:
-//
-//   if hash_default_eq<T>(a, b) returns true for any a and b of type T, then
-//   hash_default_hash<T>(a) must equal hash_default_hash<T>(b)
-//
-// For SwissTable containers this requirement is relaxed to allow a and b of
-// any and possibly different types. Note that like the standard the hash and
-// equal functions are still bound to T. This is important because some type U
-// can be hashed by/tested for equality differently depending on T. A notable
-// example is `const char*`. `const char*` is treated as a c-style string when
-// the hash function is hash<std::string> but as a pointer when the hash
-// function is hash<void*>.
-//
-#ifndef ABSL_CONTAINER_INTERNAL_HASH_FUNCTION_DEFAULTS_H_
-#define ABSL_CONTAINER_INTERNAL_HASH_FUNCTION_DEFAULTS_H_
-
-#include <cstddef>
-#include <functional>
-#include <memory>
-#include <string>
-#include <type_traits>
-
-#include "absl/base/config.h"
-#include "absl/container/internal/common.h"
-#include "absl/hash/hash.h"
-#include "absl/meta/type_traits.h"
-#include "absl/strings/cord.h"
-#include "absl/strings/string_view.h"
-
-#ifdef ABSL_HAVE_STD_STRING_VIEW
-#include <string_view>
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-// The hash of an object of type T is computed by using absl::Hash.
-template <class T, class E = void>
-struct HashEq {
-  using Hash = absl::Hash<T>;
-  using Eq = std::equal_to<T>;
-};
-
-struct StringHash {
-  using is_transparent = void;
-
-  size_t operator()(absl::string_view v) const {
-    return absl::Hash<absl::string_view>{}(v);
-  }
-  size_t operator()(const absl::Cord& v) const {
-    return absl::Hash<absl::Cord>{}(v);
-  }
-};
-
-struct StringEq {
-  using is_transparent = void;
-  bool operator()(absl::string_view lhs, absl::string_view rhs) const {
-    return lhs == rhs;
-  }
-  bool operator()(const absl::Cord& lhs, const absl::Cord& rhs) const {
-    return lhs == rhs;
-  }
-  bool operator()(const absl::Cord& lhs, absl::string_view rhs) const {
-    return lhs == rhs;
-  }
-  bool operator()(absl::string_view lhs, const absl::Cord& rhs) const {
-    return lhs == rhs;
-  }
-};
-
-// Supports heterogeneous lookup for string-like elements.
-struct StringHashEq {
-  using Hash = StringHash;
-  using Eq = StringEq;
-};
-
-template <>
-struct HashEq<std::string> : StringHashEq {};
-template <>
-struct HashEq<absl::string_view> : StringHashEq {};
-template <>
-struct HashEq<absl::Cord> : StringHashEq {};
-
-#ifdef ABSL_HAVE_STD_STRING_VIEW
-
-template <typename TChar>
-struct BasicStringHash {
-  using is_transparent = void;
-
-  size_t operator()(std::basic_string_view<TChar> v) const {
-    return absl::Hash<std::basic_string_view<TChar>>{}(v);
-  }
-};
-
-template <typename TChar>
-struct BasicStringEq {
-  using is_transparent = void;
-  bool operator()(std::basic_string_view<TChar> lhs,
-                  std::basic_string_view<TChar> rhs) const {
-    return lhs == rhs;
-  }
-};
-
-// Supports heterogeneous lookup for w/u16/u32 string + string_view + char*.
-template <typename TChar>
-struct BasicStringHashEq {
-  using Hash = BasicStringHash<TChar>;
-  using Eq = BasicStringEq<TChar>;
-};
-
-template <>
-struct HashEq<std::wstring> : BasicStringHashEq<wchar_t> {};
-template <>
-struct HashEq<std::wstring_view> : BasicStringHashEq<wchar_t> {};
-template <>
-struct HashEq<std::u16string> : BasicStringHashEq<char16_t> {};
-template <>
-struct HashEq<std::u16string_view> : BasicStringHashEq<char16_t> {};
-template <>
-struct HashEq<std::u32string> : BasicStringHashEq<char32_t> {};
-template <>
-struct HashEq<std::u32string_view> : BasicStringHashEq<char32_t> {};
-
-#endif  // ABSL_HAVE_STD_STRING_VIEW
-
-// Supports heterogeneous lookup for pointers and smart pointers.
-template <class T>
-struct HashEq<T*> {
-  struct Hash {
-    using is_transparent = void;
-    template <class U>
-    size_t operator()(const U& ptr) const {
-      return absl::Hash<const T*>{}(HashEq::ToPtr(ptr));
-    }
-  };
-  struct Eq {
-    using is_transparent = void;
-    template <class A, class B>
-    bool operator()(const A& a, const B& b) const {
-      return HashEq::ToPtr(a) == HashEq::ToPtr(b);
-    }
-  };
-
- private:
-  static const T* ToPtr(const T* ptr) { return ptr; }
-  template <class U, class D>
-  static const T* ToPtr(const std::unique_ptr<U, D>& ptr) {
-    return ptr.get();
-  }
-  template <class U>
-  static const T* ToPtr(const std::shared_ptr<U>& ptr) {
-    return ptr.get();
-  }
-};
-
-template <class T, class D>
-struct HashEq<std::unique_ptr<T, D>> : HashEq<T*> {};
-template <class T>
-struct HashEq<std::shared_ptr<T>> : HashEq<T*> {};
-
-template <typename T, typename E = void>
-struct HasAbslContainerHash : std::false_type {};
-
-template <typename T>
-struct HasAbslContainerHash<T, absl::void_t<typename T::absl_container_hash>>
-    : std::true_type {};
-
-template <typename T, typename E = void>
-struct HasAbslContainerEq : std::false_type {};
-
-template <typename T>
-struct HasAbslContainerEq<T, absl::void_t<typename T::absl_container_eq>>
-    : std::true_type {};
-
-template <typename T, typename E = void>
-struct AbslContainerEq {
-  using type = std::equal_to<>;
-};
-
-template <typename T>
-struct AbslContainerEq<
-    T, typename std::enable_if_t<HasAbslContainerEq<T>::value>> {
-  using type = typename T::absl_container_eq;
-};
-
-template <typename T, typename E = void>
-struct AbslContainerHash {
-  using type = void;
-};
-
-template <typename T>
-struct AbslContainerHash<
-    T, typename std::enable_if_t<HasAbslContainerHash<T>::value>> {
-  using type = typename T::absl_container_hash;
-};
-
-// HashEq specialization for user types that provide `absl_container_hash` and
-// (optionally) `absl_container_eq`. This specialization allows user types to
-// provide heterogeneous lookup without requiring to explicitly specify Hash/Eq
-// type arguments in unordered Abseil containers.
-//
-// Both `absl_container_hash` and `absl_container_eq` should be transparent
-// (have inner is_transparent type). While there is no technical reason to
-// restrict to transparent-only types, there is also no feasible use case when
-// it shouldn't be transparent - it is easier to relax the requirement later if
-// such a case arises rather than restricting it.
-//
-// If type provides only `absl_container_hash` then `eq` part will be
-// `std::equal_to<void>`.
-//
-// User types are not allowed to provide only a `Eq` part as there is no
-// feasible use case for this behavior - if Hash should be a default one then Eq
-// should be an equivalent to the `std::equal_to<T>`.
-template <typename T>
-struct HashEq<T, typename std::enable_if_t<HasAbslContainerHash<T>::value>> {
-  using Hash = typename AbslContainerHash<T>::type;
-  using Eq = typename AbslContainerEq<T>::type;
-  static_assert(IsTransparent<Hash>::value,
-                "absl_container_hash must be transparent. To achieve it add a "
-                "`using is_transparent = void;` clause to this type.");
-  static_assert(IsTransparent<Eq>::value,
-                "absl_container_eq must be transparent. To achieve it add a "
-                "`using is_transparent = void;` clause to this type.");
-};
-
-// This header's visibility is restricted.  If you need to access the default
-// hasher please use the container's ::hasher alias instead.
-//
-// Example: typename Hash = typename absl::flat_hash_map<K, V>::hasher
-template <class T>
-using hash_default_hash = typename container_internal::HashEq<T>::Hash;
-
-// This header's visibility is restricted.  If you need to access the default
-// key equal please use the container's ::key_equal alias instead.
-//
-// Example: typename Eq = typename absl::flat_hash_map<K, V, Hash>::key_equal
-template <class T>
-using hash_default_eq = typename container_internal::HashEq<T>::Eq;
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_HASH_FUNCTION_DEFAULTS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_generator_testing.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_generator_testing.h
deleted file mode 100644
index f1f555a5..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_generator_testing.h
+++ /dev/null
@@ -1,182 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Generates random values for testing. Specialized only for the few types we
-// care about.
-
-#ifndef ABSL_CONTAINER_INTERNAL_HASH_GENERATOR_TESTING_H_
-#define ABSL_CONTAINER_INTERNAL_HASH_GENERATOR_TESTING_H_
-
-#include <stdint.h>
-
-#include <algorithm>
-#include <cassert>
-#include <iosfwd>
-#include <random>
-#include <tuple>
-#include <type_traits>
-#include <utility>
-#include <vector>
-
-#include "absl/container/internal/hash_policy_testing.h"
-#include "absl/memory/memory.h"
-#include "absl/meta/type_traits.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-namespace hash_internal {
-namespace generator_internal {
-
-template <class Container, class = void>
-struct IsMap : std::false_type {};
-
-template <class Map>
-struct IsMap<Map, absl::void_t<typename Map::mapped_type>> : std::true_type {};
-
-}  // namespace generator_internal
-
-std::mt19937_64* GetSharedRng();
-
-enum Enum {
-  kEnumEmpty,
-  kEnumDeleted,
-};
-
-enum class EnumClass : uint64_t {
-  kEmpty,
-  kDeleted,
-};
-
-inline std::ostream& operator<<(std::ostream& o, const EnumClass& ec) {
-  return o << static_cast<uint64_t>(ec);
-}
-
-template <class T, class E = void>
-struct Generator;
-
-template <class T>
-struct Generator<T, typename std::enable_if<std::is_integral<T>::value>::type> {
-  T operator()() const {
-    std::uniform_int_distribution<T> dist;
-    return dist(*GetSharedRng());
-  }
-};
-
-template <>
-struct Generator<Enum> {
-  Enum operator()() const {
-    std::uniform_int_distribution<typename std::underlying_type<Enum>::type>
-        dist;
-    while (true) {
-      auto variate = dist(*GetSharedRng());
-      if (variate != kEnumEmpty && variate != kEnumDeleted)
-        return static_cast<Enum>(variate);
-    }
-  }
-};
-
-template <>
-struct Generator<EnumClass> {
-  EnumClass operator()() const {
-    std::uniform_int_distribution<
-        typename std::underlying_type<EnumClass>::type>
-        dist;
-    while (true) {
-      EnumClass variate = static_cast<EnumClass>(dist(*GetSharedRng()));
-      if (variate != EnumClass::kEmpty && variate != EnumClass::kDeleted)
-        return static_cast<EnumClass>(variate);
-    }
-  }
-};
-
-template <>
-struct Generator<std::string> {
-  std::string operator()() const;
-};
-
-template <>
-struct Generator<absl::string_view> {
-  absl::string_view operator()() const;
-};
-
-template <>
-struct Generator<NonStandardLayout> {
-  NonStandardLayout operator()() const {
-    return NonStandardLayout(Generator<std::string>()());
-  }
-};
-
-template <class K, class V>
-struct Generator<std::pair<K, V>> {
-  std::pair<K, V> operator()() const {
-    return std::pair<K, V>(Generator<typename std::decay<K>::type>()(),
-                           Generator<typename std::decay<V>::type>()());
-  }
-};
-
-template <class... Ts>
-struct Generator<std::tuple<Ts...>> {
-  std::tuple<Ts...> operator()() const {
-    return std::tuple<Ts...>(Generator<typename std::decay<Ts>::type>()()...);
-  }
-};
-
-template <class T>
-struct Generator<std::unique_ptr<T>> {
-  std::unique_ptr<T> operator()() const {
-    return absl::make_unique<T>(Generator<T>()());
-  }
-};
-
-template <class U>
-struct Generator<U, absl::void_t<decltype(std::declval<U&>().key()),
-                                decltype(std::declval<U&>().value())>>
-    : Generator<std::pair<
-          typename std::decay<decltype(std::declval<U&>().key())>::type,
-          typename std::decay<decltype(std::declval<U&>().value())>::type>> {};
-
-template <class Container>
-using GeneratedType = decltype(
-    std::declval<const Generator<
-        typename std::conditional<generator_internal::IsMap<Container>::value,
-                                  typename Container::value_type,
-                                  typename Container::key_type>::type>&>()());
-
-// Naive wrapper that performs a linear search of previous values.
-// Beware this is O(SQR), which is reasonable for smaller kMaxValues.
-template <class T, size_t kMaxValues = 64, class E = void>
-struct UniqueGenerator {
-  Generator<T, E> gen;
-  std::vector<T> values;
-
-  T operator()() {
-    assert(values.size() < kMaxValues);
-    for (;;) {
-      T value = gen();
-      if (std::find(values.begin(), values.end(), value) == values.end()) {
-        values.push_back(value);
-        return value;
-      }
-    }
-  }
-};
-
-}  // namespace hash_internal
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_HASH_GENERATOR_TESTING_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_policy_testing.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_policy_testing.h
deleted file mode 100644
index 66bb12ec..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_policy_testing.h
+++ /dev/null
@@ -1,183 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Utilities to help tests verify that hash tables properly handle stateful
-// allocators and hash functions.
-
-#ifndef ABSL_CONTAINER_INTERNAL_HASH_POLICY_TESTING_H_
-#define ABSL_CONTAINER_INTERNAL_HASH_POLICY_TESTING_H_
-
-#include <cstdlib>
-#include <limits>
-#include <memory>
-#include <ostream>
-#include <type_traits>
-#include <utility>
-#include <vector>
-
-#include "absl/hash/hash.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-namespace hash_testing_internal {
-
-template <class Derived>
-struct WithId {
-  WithId() : id_(next_id<Derived>()) {}
-  WithId(const WithId& that) : id_(that.id_) {}
-  WithId(WithId&& that) : id_(that.id_) { that.id_ = 0; }
-  WithId& operator=(const WithId& that) {
-    id_ = that.id_;
-    return *this;
-  }
-  WithId& operator=(WithId&& that) {
-    id_ = that.id_;
-    that.id_ = 0;
-    return *this;
-  }
-
-  size_t id() const { return id_; }
-
-  friend bool operator==(const WithId& a, const WithId& b) {
-    return a.id_ == b.id_;
-  }
-  friend bool operator!=(const WithId& a, const WithId& b) { return !(a == b); }
-
- protected:
-  explicit WithId(size_t id) : id_(id) {}
-
- private:
-  size_t id_;
-
-  template <class T>
-  static size_t next_id() {
-    // 0 is reserved for moved from state.
-    static size_t gId = 1;
-    return gId++;
-  }
-};
-
-}  // namespace hash_testing_internal
-
-struct NonStandardLayout {
-  NonStandardLayout() {}
-  explicit NonStandardLayout(std::string s) : value(std::move(s)) {}
-  virtual ~NonStandardLayout() {}
-
-  friend bool operator==(const NonStandardLayout& a,
-                         const NonStandardLayout& b) {
-    return a.value == b.value;
-  }
-  friend bool operator!=(const NonStandardLayout& a,
-                         const NonStandardLayout& b) {
-    return a.value != b.value;
-  }
-
-  template <typename H>
-  friend H AbslHashValue(H h, const NonStandardLayout& v) {
-    return H::combine(std::move(h), v.value);
-  }
-
-  std::string value;
-};
-
-struct StatefulTestingHash
-    : absl::container_internal::hash_testing_internal::WithId<
-          StatefulTestingHash> {
-  template <class T>
-  size_t operator()(const T& t) const {
-    return absl::Hash<T>{}(t);
-  }
-};
-
-struct StatefulTestingEqual
-    : absl::container_internal::hash_testing_internal::WithId<
-          StatefulTestingEqual> {
-  template <class T, class U>
-  bool operator()(const T& t, const U& u) const {
-    return t == u;
-  }
-};
-
-// It is expected that Alloc() == Alloc() for all allocators so we cannot use
-// WithId base. We need to explicitly assign ids.
-template <class T = int>
-struct Alloc : std::allocator<T> {
-  using propagate_on_container_swap = std::true_type;
-
-  // Using old paradigm for this to ensure compatibility.
-  explicit Alloc(size_t id = 0) : id_(id) {}
-
-  Alloc(const Alloc&) = default;
-  Alloc& operator=(const Alloc&) = default;
-
-  template <class U>
-  Alloc(const Alloc<U>& that) : std::allocator<T>(that), id_(that.id()) {}
-
-  template <class U>
-  struct rebind {
-    using other = Alloc<U>;
-  };
-
-  size_t id() const { return id_; }
-
-  friend bool operator==(const Alloc& a, const Alloc& b) {
-    return a.id_ == b.id_;
-  }
-  friend bool operator!=(const Alloc& a, const Alloc& b) { return !(a == b); }
-
- private:
-  size_t id_ = (std::numeric_limits<size_t>::max)();
-};
-
-template <class Map>
-auto items(const Map& m) -> std::vector<
-    std::pair<typename Map::key_type, typename Map::mapped_type>> {
-  using std::get;
-  std::vector<std::pair<typename Map::key_type, typename Map::mapped_type>> res;
-  res.reserve(m.size());
-  for (const auto& v : m) res.emplace_back(get<0>(v), get<1>(v));
-  return res;
-}
-
-template <class Set>
-auto keys(const Set& s)
-    -> std::vector<typename std::decay<typename Set::key_type>::type> {
-  std::vector<typename std::decay<typename Set::key_type>::type> res;
-  res.reserve(s.size());
-  for (const auto& v : s) res.emplace_back(v);
-  return res;
-}
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-// ABSL_UNORDERED_SUPPORTS_ALLOC_CTORS is false for glibcxx versions
-// where the unordered containers are missing certain constructors that
-// take allocator arguments. This test is defined ad-hoc for the platforms
-// we care about (notably Crosstool 17) because libstdcxx's useless
-// versioning scheme precludes a more principled solution.
-// From GCC-4.9 Changelog: (src: https://gcc.gnu.org/gcc-4.9/changes.html)
-// "the unordered associative containers in <unordered_map> and <unordered_set>
-// meet the allocator-aware container requirements;"
-#if defined(__GLIBCXX__) && __GLIBCXX__ <= 20140425
-#define ABSL_UNORDERED_SUPPORTS_ALLOC_CTORS 0
-#else
-#define ABSL_UNORDERED_SUPPORTS_ALLOC_CTORS 1
-#endif
-
-#endif  // ABSL_CONTAINER_INTERNAL_HASH_POLICY_TESTING_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_policy_traits.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_policy_traits.h
deleted file mode 100644
index ec08794a..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/hash_policy_traits.h
+++ /dev/null
@@ -1,207 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_HASH_POLICY_TRAITS_H_
-#define ABSL_CONTAINER_INTERNAL_HASH_POLICY_TRAITS_H_
-
-#include <cstddef>
-#include <memory>
-#include <new>
-#include <type_traits>
-#include <utility>
-
-#include "absl/container/internal/common_policy_traits.h"
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-// Defines how slots are initialized/destroyed/moved.
-template <class Policy, class = void>
-struct hash_policy_traits : common_policy_traits<Policy> {
-  // The type of the keys stored in the hashtable.
-  using key_type = typename Policy::key_type;
-
- private:
-  struct ReturnKey {
-    // When C++17 is available, we can use std::launder to provide mutable
-    // access to the key for use in node handle.
-#if defined(__cpp_lib_launder) && __cpp_lib_launder >= 201606
-    template <class Key,
-              absl::enable_if_t<std::is_lvalue_reference<Key>::value, int> = 0>
-    static key_type& Impl(Key&& k, int) {
-      return *std::launder(
-          const_cast<key_type*>(std::addressof(std::forward<Key>(k))));
-    }
-#endif
-
-    template <class Key>
-    static Key Impl(Key&& k, char) {
-      return std::forward<Key>(k);
-    }
-
-    // When Key=T&, we forward the lvalue reference.
-    // When Key=T, we return by value to avoid a dangling reference.
-    // eg, for string_hash_map.
-    template <class Key, class... Args>
-    auto operator()(Key&& k, const Args&...) const
-        -> decltype(Impl(std::forward<Key>(k), 0)) {
-      return Impl(std::forward<Key>(k), 0);
-    }
-  };
-
-  template <class P = Policy, class = void>
-  struct ConstantIteratorsImpl : std::false_type {};
-
-  template <class P>
-  struct ConstantIteratorsImpl<P, absl::void_t<typename P::constant_iterators>>
-      : P::constant_iterators {};
-
- public:
-  // The actual object stored in the hash table.
-  using slot_type = typename Policy::slot_type;
-
-  // The argument type for insertions into the hashtable. This is different
-  // from value_type for increased performance. See initializer_list constructor
-  // and insert() member functions for more details.
-  using init_type = typename Policy::init_type;
-
-  using reference = decltype(Policy::element(std::declval<slot_type*>()));
-  using pointer = typename std::remove_reference<reference>::type*;
-  using value_type = typename std::remove_reference<reference>::type;
-
-  // Policies can set this variable to tell raw_hash_set that all iterators
-  // should be constant, even `iterator`. This is useful for set-like
-  // containers.
-  // Defaults to false if not provided by the policy.
-  using constant_iterators = ConstantIteratorsImpl<>;
-
-  // Returns the amount of memory owned by `slot`, exclusive of `sizeof(*slot)`.
-  //
-  // If `slot` is nullptr, returns the constant amount of memory owned by any
-  // full slot or -1 if slots own variable amounts of memory.
-  //
-  // PRECONDITION: `slot` is INITIALIZED or nullptr
-  template <class P = Policy>
-  static size_t space_used(const slot_type* slot) {
-    return P::space_used(slot);
-  }
-
-  // Provides generalized access to the key for elements, both for elements in
-  // the table and for elements that have not yet been inserted (or even
-  // constructed).  We would like an API that allows us to say: `key(args...)`
-  // but we cannot do that for all cases, so we use this more general API that
-  // can be used for many things, including the following:
-  //
-  //   - Given an element in a table, get its key.
-  //   - Given an element initializer, get its key.
-  //   - Given `emplace()` arguments, get the element key.
-  //
-  // Implementations of this must adhere to a very strict technical
-  // specification around aliasing and consuming arguments:
-  //
-  // Let `value_type` be the result type of `element()` without ref- and
-  // cv-qualifiers. The first argument is a functor, the rest are constructor
-  // arguments for `value_type`. Returns `std::forward<F>(f)(k, xs...)`, where
-  // `k` is the element key, and `xs...` are the new constructor arguments for
-  // `value_type`. It's allowed for `k` to alias `xs...`, and for both to alias
-  // `ts...`. The key won't be touched once `xs...` are used to construct an
-  // element; `ts...` won't be touched at all, which allows `apply()` to consume
-  // any rvalues among them.
-  //
-  // If `value_type` is constructible from `Ts&&...`, `Policy::apply()` must not
-  // trigger a hard compile error unless it originates from `f`. In other words,
-  // `Policy::apply()` must be SFINAE-friendly. If `value_type` is not
-  // constructible from `Ts&&...`, either SFINAE or a hard compile error is OK.
-  //
-  // If `Ts...` is `[cv] value_type[&]` or `[cv] init_type[&]`,
-  // `Policy::apply()` must work. A compile error is not allowed, SFINAE or not.
-  template <class F, class... Ts, class P = Policy>
-  static auto apply(F&& f, Ts&&... ts)
-      -> decltype(P::apply(std::forward<F>(f), std::forward<Ts>(ts)...)) {
-    return P::apply(std::forward<F>(f), std::forward<Ts>(ts)...);
-  }
-
-  // Returns the "key" portion of the slot.
-  // Used for node handle manipulation.
-  template <class P = Policy>
-  static auto mutable_key(slot_type* slot)
-      -> decltype(P::apply(ReturnKey(), hash_policy_traits::element(slot))) {
-    return P::apply(ReturnKey(), hash_policy_traits::element(slot));
-  }
-
-  // Returns the "value" (as opposed to the "key") portion of the element. Used
-  // by maps to implement `operator[]`, `at()` and `insert_or_assign()`.
-  template <class T, class P = Policy>
-  static auto value(T* elem) -> decltype(P::value(elem)) {
-    return P::value(elem);
-  }
-
-  using HashSlotFn = size_t (*)(const void* hash_fn, void* slot);
-
-  template <class Hash>
-  static constexpr HashSlotFn get_hash_slot_fn() {
-// get_hash_slot_fn may return nullptr to signal that non type erased function
-// should be used. GCC warns against comparing function address with nullptr.
-#if defined(__GNUC__) && !defined(__clang__)
-#pragma GCC diagnostic push
-// silent error: the address of * will never be NULL [-Werror=address]
-#pragma GCC diagnostic ignored "-Waddress"
-#endif
-    return Policy::template get_hash_slot_fn<Hash>() == nullptr
-               ? &hash_slot_fn_non_type_erased<Hash>
-               : Policy::template get_hash_slot_fn<Hash>();
-#if defined(__GNUC__) && !defined(__clang__)
-#pragma GCC diagnostic pop
-#endif
-  }
-
-  // Whether small object optimization is enabled. False by default.
-  static constexpr bool soo_enabled() { return soo_enabled_impl(Rank1{}); }
-
- private:
-  template <class Hash>
-  struct HashElement {
-    template <class K, class... Args>
-    size_t operator()(const K& key, Args&&...) const {
-      return h(key);
-    }
-    const Hash& h;
-  };
-
-  template <class Hash>
-  static size_t hash_slot_fn_non_type_erased(const void* hash_fn, void* slot) {
-    return Policy::apply(HashElement<Hash>{*static_cast<const Hash*>(hash_fn)},
-                         Policy::element(static_cast<slot_type*>(slot)));
-  }
-
-  // Use go/ranked-overloads for dispatching. Rank1 is preferred.
-  struct Rank0 {};
-  struct Rank1 : Rank0 {};
-
-  // Use auto -> decltype as an enabler.
-  template <class P = Policy>
-  static constexpr auto soo_enabled_impl(Rank1) -> decltype(P::soo_enabled()) {
-    return P::soo_enabled();
-  }
-
-  static constexpr bool soo_enabled_impl(Rank0) { return false; }
-};
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_HASH_POLICY_TRAITS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/hashtable_debug.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/hashtable_debug.h
deleted file mode 100644
index c79c1a98..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/hashtable_debug.h
+++ /dev/null
@@ -1,102 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// This library provides APIs to debug the probing behavior of hash tables.
-//
-// In general, the probing behavior is a black box for users and only the
-// side effects can be measured in the form of performance differences.
-// These APIs give a glimpse on the actual behavior of the probing algorithms in
-// these hashtables given a specified hash function and a set of elements.
-//
-// The probe count distribution can be used to assess the quality of the hash
-// function for that particular hash table. Note that a hash function that
-// performs well in one hash table implementation does not necessarily performs
-// well in a different one.
-//
-// This library supports std::unordered_{set,map}, dense_hash_{set,map} and
-// absl::{flat,node,string}_hash_{set,map}.
-
-#ifndef ABSL_CONTAINER_INTERNAL_HASHTABLE_DEBUG_H_
-#define ABSL_CONTAINER_INTERNAL_HASHTABLE_DEBUG_H_
-
-#include <cstddef>
-#include <algorithm>
-#include <type_traits>
-#include <vector>
-
-#include "absl/container/internal/hashtable_debug_hooks.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-// Returns the number of probes required to lookup `key`.  Returns 0 for a
-// search with no collisions.  Higher values mean more hash collisions occurred;
-// however, the exact meaning of this number varies according to the container
-// type.
-template <typename C>
-size_t GetHashtableDebugNumProbes(
-    const C& c, const typename C::key_type& key) {
-  return absl::container_internal::hashtable_debug_internal::
-      HashtableDebugAccess<C>::GetNumProbes(c, key);
-}
-
-// Gets a histogram of the number of probes for each elements in the container.
-// The sum of all the values in the vector is equal to container.size().
-template <typename C>
-std::vector<size_t> GetHashtableDebugNumProbesHistogram(const C& container) {
-  std::vector<size_t> v;
-  for (auto it = container.begin(); it != container.end(); ++it) {
-    size_t num_probes = GetHashtableDebugNumProbes(
-        container,
-        absl::container_internal::hashtable_debug_internal::GetKey<C>(*it, 0));
-    v.resize((std::max)(v.size(), num_probes + 1));
-    v[num_probes]++;
-  }
-  return v;
-}
-
-struct HashtableDebugProbeSummary {
-  size_t total_elements;
-  size_t total_num_probes;
-  double mean;
-};
-
-// Gets a summary of the probe count distribution for the elements in the
-// container.
-template <typename C>
-HashtableDebugProbeSummary GetHashtableDebugProbeSummary(const C& container) {
-  auto probes = GetHashtableDebugNumProbesHistogram(container);
-  HashtableDebugProbeSummary summary = {};
-  for (size_t i = 0; i < probes.size(); ++i) {
-    summary.total_elements += probes[i];
-    summary.total_num_probes += probes[i] * i;
-  }
-  summary.mean = 1.0 * summary.total_num_probes / summary.total_elements;
-  return summary;
-}
-
-// Returns the number of bytes requested from the allocator by the container
-// and not freed.
-template <typename C>
-size_t AllocatedByteSize(const C& c) {
-  return absl::container_internal::hashtable_debug_internal::
-      HashtableDebugAccess<C>::AllocatedByteSize(c);
-}
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_HASHTABLE_DEBUG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/hashtable_debug_hooks.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/hashtable_debug_hooks.h
deleted file mode 100644
index 3e9ea595..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/hashtable_debug_hooks.h
+++ /dev/null
@@ -1,85 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Provides the internal API for hashtable_debug.h.
-
-#ifndef ABSL_CONTAINER_INTERNAL_HASHTABLE_DEBUG_HOOKS_H_
-#define ABSL_CONTAINER_INTERNAL_HASHTABLE_DEBUG_HOOKS_H_
-
-#include <cstddef>
-
-#include <algorithm>
-#include <type_traits>
-#include <vector>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-namespace hashtable_debug_internal {
-
-// If it is a map, call get<0>().
-using std::get;
-template <typename T, typename = typename T::mapped_type>
-auto GetKey(const typename T::value_type& pair, int) -> decltype(get<0>(pair)) {
-  return get<0>(pair);
-}
-
-// If it is not a map, return the value directly.
-template <typename T>
-const typename T::key_type& GetKey(const typename T::key_type& key, char) {
-  return key;
-}
-
-// Containers should specialize this to provide debug information for that
-// container.
-template <class Container, typename Enabler = void>
-struct HashtableDebugAccess {
-  // Returns the number of probes required to find `key` in `c`.  The "number of
-  // probes" is a concept that can vary by container.  Implementations should
-  // return 0 when `key` was found in the minimum number of operations and
-  // should increment the result for each non-trivial operation required to find
-  // `key`.
-  //
-  // The default implementation uses the bucket api from the standard and thus
-  // works for `std::unordered_*` containers.
-  static size_t GetNumProbes(const Container& c,
-                             const typename Container::key_type& key) {
-    if (!c.bucket_count()) return {};
-    size_t num_probes = 0;
-    size_t bucket = c.bucket(key);
-    for (auto it = c.begin(bucket), e = c.end(bucket);; ++it, ++num_probes) {
-      if (it == e) return num_probes;
-      if (c.key_eq()(key, GetKey<Container>(*it, 0))) return num_probes;
-    }
-  }
-
-  // Returns the number of bytes requested from the allocator by the container
-  // and not freed.
-  //
-  // static size_t AllocatedByteSize(const Container& c);
-
-  // Returns a tight lower bound for AllocatedByteSize(c) where `c` is of type
-  // `Container` and `c.size()` is equal to `num_elements`.
-  //
-  // static size_t LowerBoundAllocatedByteSize(size_t num_elements);
-};
-
-}  // namespace hashtable_debug_internal
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_HASHTABLE_DEBUG_HOOKS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/hashtablez_sampler.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/hashtablez_sampler.h
deleted file mode 100644
index e41ee2d7..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/hashtablez_sampler.h
+++ /dev/null
@@ -1,257 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: hashtablez_sampler.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines the API for a low level library to sample hashtables
-// and collect runtime statistics about them.
-//
-// `HashtablezSampler` controls the lifecycle of `HashtablezInfo` objects which
-// store information about a single sample.
-//
-// `Record*` methods store information into samples.
-// `Sample()` and `Unsample()` make use of a single global sampler with
-// properties controlled by the flags hashtablez_enabled,
-// hashtablez_sample_rate, and hashtablez_max_samples.
-//
-// WARNING
-//
-// Using this sampling API may cause sampled Swiss tables to use the global
-// allocator (operator `new`) in addition to any custom allocator.  If you
-// are using a table in an unusual circumstance where allocation or calling a
-// linux syscall is unacceptable, this could interfere.
-//
-// This utility is internal-only. Use at your own risk.
-
-#ifndef ABSL_CONTAINER_INTERNAL_HASHTABLEZ_SAMPLER_H_
-#define ABSL_CONTAINER_INTERNAL_HASHTABLEZ_SAMPLER_H_
-
-#include <atomic>
-#include <functional>
-#include <memory>
-#include <vector>
-
-#include "absl/base/config.h"
-#include "absl/base/internal/per_thread_tls.h"
-#include "absl/base/optimization.h"
-#include "absl/profiling/internal/sample_recorder.h"
-#include "absl/synchronization/mutex.h"
-#include "absl/utility/utility.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-// Stores information about a sampled hashtable.  All mutations to this *must*
-// be made through `Record*` functions below.  All reads from this *must* only
-// occur in the callback to `HashtablezSampler::Iterate`.
-struct HashtablezInfo : public profiling_internal::Sample<HashtablezInfo> {
-  // Constructs the object but does not fill in any fields.
-  HashtablezInfo();
-  ~HashtablezInfo();
-  HashtablezInfo(const HashtablezInfo&) = delete;
-  HashtablezInfo& operator=(const HashtablezInfo&) = delete;
-
-  // Puts the object into a clean state, fills in the logically `const` members,
-  // blocking for any readers that are currently sampling the object.
-  void PrepareForSampling(int64_t stride, size_t inline_element_size_value)
-      ABSL_EXCLUSIVE_LOCKS_REQUIRED(init_mu);
-
-  // These fields are mutated by the various Record* APIs and need to be
-  // thread-safe.
-  std::atomic<size_t> capacity;
-  std::atomic<size_t> size;
-  std::atomic<size_t> num_erases;
-  std::atomic<size_t> num_rehashes;
-  std::atomic<size_t> max_probe_length;
-  std::atomic<size_t> total_probe_length;
-  std::atomic<size_t> hashes_bitwise_or;
-  std::atomic<size_t> hashes_bitwise_and;
-  std::atomic<size_t> hashes_bitwise_xor;
-  std::atomic<size_t> max_reserve;
-
-  // All of the fields below are set by `PrepareForSampling`, they must not be
-  // mutated in `Record*` functions.  They are logically `const` in that sense.
-  // These are guarded by init_mu, but that is not externalized to clients,
-  // which can read them only during `SampleRecorder::Iterate` which will hold
-  // the lock.
-  static constexpr int kMaxStackDepth = 64;
-  absl::Time create_time;
-  int32_t depth;
-  void* stack[kMaxStackDepth];
-  size_t inline_element_size;  // How big is the slot?
-};
-
-void RecordRehashSlow(HashtablezInfo* info, size_t total_probe_length);
-
-void RecordReservationSlow(HashtablezInfo* info, size_t target_capacity);
-
-void RecordClearedReservationSlow(HashtablezInfo* info);
-
-void RecordStorageChangedSlow(HashtablezInfo* info, size_t size,
-                              size_t capacity);
-
-void RecordInsertSlow(HashtablezInfo* info, size_t hash,
-                      size_t distance_from_desired);
-
-void RecordEraseSlow(HashtablezInfo* info);
-
-struct SamplingState {
-  int64_t next_sample;
-  // When we make a sampling decision, we record that distance so we can weight
-  // each sample.
-  int64_t sample_stride;
-};
-
-HashtablezInfo* SampleSlow(SamplingState& next_sample,
-                           size_t inline_element_size);
-void UnsampleSlow(HashtablezInfo* info);
-
-#if defined(ABSL_INTERNAL_HASHTABLEZ_SAMPLE)
-#error ABSL_INTERNAL_HASHTABLEZ_SAMPLE cannot be directly set
-#endif  // defined(ABSL_INTERNAL_HASHTABLEZ_SAMPLE)
-
-#if defined(ABSL_INTERNAL_HASHTABLEZ_SAMPLE)
-class HashtablezInfoHandle {
- public:
-  explicit HashtablezInfoHandle() : info_(nullptr) {}
-  explicit HashtablezInfoHandle(HashtablezInfo* info) : info_(info) {}
-
-  // We do not have a destructor. Caller is responsible for calling Unregister
-  // before destroying the handle.
-  void Unregister() {
-    if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
-    UnsampleSlow(info_);
-  }
-
-  inline bool IsSampled() const { return ABSL_PREDICT_FALSE(info_ != nullptr); }
-
-  inline void RecordStorageChanged(size_t size, size_t capacity) {
-    if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
-    RecordStorageChangedSlow(info_, size, capacity);
-  }
-
-  inline void RecordRehash(size_t total_probe_length) {
-    if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
-    RecordRehashSlow(info_, total_probe_length);
-  }
-
-  inline void RecordReservation(size_t target_capacity) {
-    if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
-    RecordReservationSlow(info_, target_capacity);
-  }
-
-  inline void RecordClearedReservation() {
-    if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
-    RecordClearedReservationSlow(info_);
-  }
-
-  inline void RecordInsert(size_t hash, size_t distance_from_desired) {
-    if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
-    RecordInsertSlow(info_, hash, distance_from_desired);
-  }
-
-  inline void RecordErase() {
-    if (ABSL_PREDICT_TRUE(info_ == nullptr)) return;
-    RecordEraseSlow(info_);
-  }
-
-  friend inline void swap(HashtablezInfoHandle& lhs,
-                          HashtablezInfoHandle& rhs) {
-    std::swap(lhs.info_, rhs.info_);
-  }
-
- private:
-  friend class HashtablezInfoHandlePeer;
-  HashtablezInfo* info_;
-};
-#else
-// Ensure that when Hashtablez is turned off at compile time, HashtablezInfo can
-// be removed by the linker, in order to reduce the binary size.
-class HashtablezInfoHandle {
- public:
-  explicit HashtablezInfoHandle() = default;
-  explicit HashtablezInfoHandle(std::nullptr_t) {}
-
-  inline void Unregister() {}
-  inline bool IsSampled() const { return false; }
-  inline void RecordStorageChanged(size_t /*size*/, size_t /*capacity*/) {}
-  inline void RecordRehash(size_t /*total_probe_length*/) {}
-  inline void RecordReservation(size_t /*target_capacity*/) {}
-  inline void RecordClearedReservation() {}
-  inline void RecordInsert(size_t /*hash*/, size_t /*distance_from_desired*/) {}
-  inline void RecordErase() {}
-
-  friend inline void swap(HashtablezInfoHandle& /*lhs*/,
-                          HashtablezInfoHandle& /*rhs*/) {}
-};
-#endif  // defined(ABSL_INTERNAL_HASHTABLEZ_SAMPLE)
-
-#if defined(ABSL_INTERNAL_HASHTABLEZ_SAMPLE)
-extern ABSL_PER_THREAD_TLS_KEYWORD SamplingState global_next_sample;
-#endif  // defined(ABSL_INTERNAL_HASHTABLEZ_SAMPLE)
-
-// Returns an RAII sampling handle that manages registration and unregistation
-// with the global sampler.
-inline HashtablezInfoHandle Sample(
-    size_t inline_element_size ABSL_ATTRIBUTE_UNUSED) {
-#if defined(ABSL_INTERNAL_HASHTABLEZ_SAMPLE)
-  if (ABSL_PREDICT_TRUE(--global_next_sample.next_sample > 0)) {
-    return HashtablezInfoHandle(nullptr);
-  }
-  return HashtablezInfoHandle(
-      SampleSlow(global_next_sample, inline_element_size));
-#else
-  return HashtablezInfoHandle(nullptr);
-#endif  // !ABSL_PER_THREAD_TLS
-}
-
-using HashtablezSampler =
-    ::absl::profiling_internal::SampleRecorder<HashtablezInfo>;
-
-// Returns a global Sampler.
-HashtablezSampler& GlobalHashtablezSampler();
-
-using HashtablezConfigListener = void (*)();
-void SetHashtablezConfigListener(HashtablezConfigListener l);
-
-// Enables or disables sampling for Swiss tables.
-bool IsHashtablezEnabled();
-void SetHashtablezEnabled(bool enabled);
-void SetHashtablezEnabledInternal(bool enabled);
-
-// Sets the rate at which Swiss tables will be sampled.
-int32_t GetHashtablezSampleParameter();
-void SetHashtablezSampleParameter(int32_t rate);
-void SetHashtablezSampleParameterInternal(int32_t rate);
-
-// Sets a soft max for the number of samples that will be kept.
-size_t GetHashtablezMaxSamples();
-void SetHashtablezMaxSamples(size_t max);
-void SetHashtablezMaxSamplesInternal(size_t max);
-
-// Configuration override.
-// This allows process-wide sampling without depending on order of
-// initialization of static storage duration objects.
-// The definition of this constant is weak, which allows us to inject a
-// different value for it at link time.
-extern "C" bool ABSL_INTERNAL_C_SYMBOL(AbslContainerInternalSampleEverything)();
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_HASHTABLEZ_SAMPLER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/inlined_vector.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/inlined_vector.h
deleted file mode 100644
index a1575328..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/inlined_vector.h
+++ /dev/null
@@ -1,1114 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_INLINED_VECTOR_H_
-#define ABSL_CONTAINER_INTERNAL_INLINED_VECTOR_H_
-
-#include <algorithm>
-#include <cstddef>
-#include <cstring>
-#include <iterator>
-#include <limits>
-#include <memory>
-#include <new>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/macros.h"
-#include "absl/container/internal/compressed_tuple.h"
-#include "absl/memory/memory.h"
-#include "absl/meta/type_traits.h"
-#include "absl/types/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace inlined_vector_internal {
-
-// GCC does not deal very well with the below code
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Warray-bounds"
-#endif
-
-template <typename A>
-using AllocatorTraits = std::allocator_traits<A>;
-template <typename A>
-using ValueType = typename AllocatorTraits<A>::value_type;
-template <typename A>
-using SizeType = typename AllocatorTraits<A>::size_type;
-template <typename A>
-using Pointer = typename AllocatorTraits<A>::pointer;
-template <typename A>
-using ConstPointer = typename AllocatorTraits<A>::const_pointer;
-template <typename A>
-using SizeType = typename AllocatorTraits<A>::size_type;
-template <typename A>
-using DifferenceType = typename AllocatorTraits<A>::difference_type;
-template <typename A>
-using Reference = ValueType<A>&;
-template <typename A>
-using ConstReference = const ValueType<A>&;
-template <typename A>
-using Iterator = Pointer<A>;
-template <typename A>
-using ConstIterator = ConstPointer<A>;
-template <typename A>
-using ReverseIterator = typename std::reverse_iterator<Iterator<A>>;
-template <typename A>
-using ConstReverseIterator = typename std::reverse_iterator<ConstIterator<A>>;
-template <typename A>
-using MoveIterator = typename std::move_iterator<Iterator<A>>;
-
-template <typename Iterator>
-using IsAtLeastForwardIterator = std::is_convertible<
-    typename std::iterator_traits<Iterator>::iterator_category,
-    std::forward_iterator_tag>;
-
-template <typename A>
-using IsMoveAssignOk = std::is_move_assignable<ValueType<A>>;
-template <typename A>
-using IsSwapOk = absl::type_traits_internal::IsSwappable<ValueType<A>>;
-
-template <typename T>
-struct TypeIdentity {
-  using type = T;
-};
-
-// Used for function arguments in template functions to prevent ADL by forcing
-// callers to explicitly specify the template parameter.
-template <typename T>
-using NoTypeDeduction = typename TypeIdentity<T>::type;
-
-template <typename A, bool IsTriviallyDestructible =
-                          absl::is_trivially_destructible<ValueType<A>>::value>
-struct DestroyAdapter;
-
-template <typename A>
-struct DestroyAdapter<A, /* IsTriviallyDestructible */ false> {
-  static void DestroyElements(A& allocator, Pointer<A> destroy_first,
-                              SizeType<A> destroy_size) {
-    for (SizeType<A> i = destroy_size; i != 0;) {
-      --i;
-      AllocatorTraits<A>::destroy(allocator, destroy_first + i);
-    }
-  }
-};
-
-template <typename A>
-struct DestroyAdapter<A, /* IsTriviallyDestructible */ true> {
-  static void DestroyElements(A& allocator, Pointer<A> destroy_first,
-                              SizeType<A> destroy_size) {
-    static_cast<void>(allocator);
-    static_cast<void>(destroy_first);
-    static_cast<void>(destroy_size);
-  }
-};
-
-template <typename A>
-struct Allocation {
-  Pointer<A> data = nullptr;
-  SizeType<A> capacity = 0;
-};
-
-template <typename A,
-          bool IsOverAligned =
-              (alignof(ValueType<A>) > ABSL_INTERNAL_DEFAULT_NEW_ALIGNMENT)>
-struct MallocAdapter {
-  static Allocation<A> Allocate(A& allocator, SizeType<A> requested_capacity) {
-    return {AllocatorTraits<A>::allocate(allocator, requested_capacity),
-            requested_capacity};
-  }
-
-  static void Deallocate(A& allocator, Pointer<A> pointer,
-                         SizeType<A> capacity) {
-    AllocatorTraits<A>::deallocate(allocator, pointer, capacity);
-  }
-};
-
-template <typename A, typename ValueAdapter>
-void ConstructElements(NoTypeDeduction<A>& allocator,
-                       Pointer<A> construct_first, ValueAdapter& values,
-                       SizeType<A> construct_size) {
-  for (SizeType<A> i = 0; i < construct_size; ++i) {
-    ABSL_INTERNAL_TRY { values.ConstructNext(allocator, construct_first + i); }
-    ABSL_INTERNAL_CATCH_ANY {
-      DestroyAdapter<A>::DestroyElements(allocator, construct_first, i);
-      ABSL_INTERNAL_RETHROW;
-    }
-  }
-}
-
-template <typename A, typename ValueAdapter>
-void AssignElements(Pointer<A> assign_first, ValueAdapter& values,
-                    SizeType<A> assign_size) {
-  for (SizeType<A> i = 0; i < assign_size; ++i) {
-    values.AssignNext(assign_first + i);
-  }
-}
-
-template <typename A>
-struct StorageView {
-  Pointer<A> data;
-  SizeType<A> size;
-  SizeType<A> capacity;
-};
-
-template <typename A, typename Iterator>
-class IteratorValueAdapter {
- public:
-  explicit IteratorValueAdapter(const Iterator& it) : it_(it) {}
-
-  void ConstructNext(A& allocator, Pointer<A> construct_at) {
-    AllocatorTraits<A>::construct(allocator, construct_at, *it_);
-    ++it_;
-  }
-
-  void AssignNext(Pointer<A> assign_at) {
-    *assign_at = *it_;
-    ++it_;
-  }
-
- private:
-  Iterator it_;
-};
-
-template <typename A>
-class CopyValueAdapter {
- public:
-  explicit CopyValueAdapter(ConstPointer<A> p) : ptr_(p) {}
-
-  void ConstructNext(A& allocator, Pointer<A> construct_at) {
-    AllocatorTraits<A>::construct(allocator, construct_at, *ptr_);
-  }
-
-  void AssignNext(Pointer<A> assign_at) { *assign_at = *ptr_; }
-
- private:
-  ConstPointer<A> ptr_;
-};
-
-template <typename A>
-class DefaultValueAdapter {
- public:
-  explicit DefaultValueAdapter() {}
-
-  void ConstructNext(A& allocator, Pointer<A> construct_at) {
-    AllocatorTraits<A>::construct(allocator, construct_at);
-  }
-
-  void AssignNext(Pointer<A> assign_at) { *assign_at = ValueType<A>(); }
-};
-
-template <typename A>
-class AllocationTransaction {
- public:
-  explicit AllocationTransaction(A& allocator)
-      : allocator_data_(allocator, nullptr), capacity_(0) {}
-
-  ~AllocationTransaction() {
-    if (DidAllocate()) {
-      MallocAdapter<A>::Deallocate(GetAllocator(), GetData(), GetCapacity());
-    }
-  }
-
-  AllocationTransaction(const AllocationTransaction&) = delete;
-  void operator=(const AllocationTransaction&) = delete;
-
-  A& GetAllocator() { return allocator_data_.template get<0>(); }
-  Pointer<A>& GetData() { return allocator_data_.template get<1>(); }
-  SizeType<A>& GetCapacity() { return capacity_; }
-
-  bool DidAllocate() { return GetData() != nullptr; }
-
-  Pointer<A> Allocate(SizeType<A> requested_capacity) {
-    Allocation<A> result =
-        MallocAdapter<A>::Allocate(GetAllocator(), requested_capacity);
-    GetData() = result.data;
-    GetCapacity() = result.capacity;
-    return result.data;
-  }
-
-  ABSL_MUST_USE_RESULT Allocation<A> Release() && {
-    Allocation<A> result = {GetData(), GetCapacity()};
-    Reset();
-    return result;
-  }
-
- private:
-  void Reset() {
-    GetData() = nullptr;
-    GetCapacity() = 0;
-  }
-
-  container_internal::CompressedTuple<A, Pointer<A>> allocator_data_;
-  SizeType<A> capacity_;
-};
-
-template <typename A>
-class ConstructionTransaction {
- public:
-  explicit ConstructionTransaction(A& allocator)
-      : allocator_data_(allocator, nullptr), size_(0) {}
-
-  ~ConstructionTransaction() {
-    if (DidConstruct()) {
-      DestroyAdapter<A>::DestroyElements(GetAllocator(), GetData(), GetSize());
-    }
-  }
-
-  ConstructionTransaction(const ConstructionTransaction&) = delete;
-  void operator=(const ConstructionTransaction&) = delete;
-
-  A& GetAllocator() { return allocator_data_.template get<0>(); }
-  Pointer<A>& GetData() { return allocator_data_.template get<1>(); }
-  SizeType<A>& GetSize() { return size_; }
-
-  bool DidConstruct() { return GetData() != nullptr; }
-  template <typename ValueAdapter>
-  void Construct(Pointer<A> data, ValueAdapter& values, SizeType<A> size) {
-    ConstructElements<A>(GetAllocator(), data, values, size);
-    GetData() = data;
-    GetSize() = size;
-  }
-  void Commit() && {
-    GetData() = nullptr;
-    GetSize() = 0;
-  }
-
- private:
-  container_internal::CompressedTuple<A, Pointer<A>> allocator_data_;
-  SizeType<A> size_;
-};
-
-template <typename T, size_t N, typename A>
-class Storage {
- public:
-  struct MemcpyPolicy {};
-  struct ElementwiseAssignPolicy {};
-  struct ElementwiseSwapPolicy {};
-  struct ElementwiseConstructPolicy {};
-
-  using MoveAssignmentPolicy = absl::conditional_t<
-      // Fast path: if the value type can be trivially move assigned and
-      // destroyed, and we know the allocator doesn't do anything fancy, then
-      // it's safe for us to simply adopt the contents of the storage for
-      // `other` and remove its own reference to them. It's as if we had
-      // individually move-assigned each value and then destroyed the original.
-      absl::conjunction<absl::is_trivially_move_assignable<ValueType<A>>,
-                        absl::is_trivially_destructible<ValueType<A>>,
-                        std::is_same<A, std::allocator<ValueType<A>>>>::value,
-      MemcpyPolicy,
-      // Otherwise we use move assignment if possible. If not, we simulate
-      // move assignment using move construction.
-      //
-      // Note that this is in contrast to e.g. std::vector and std::optional,
-      // which are themselves not move-assignable when their contained type is
-      // not.
-      absl::conditional_t<IsMoveAssignOk<A>::value, ElementwiseAssignPolicy,
-                          ElementwiseConstructPolicy>>;
-
-  // The policy to be used specifically when swapping inlined elements.
-  using SwapInlinedElementsPolicy = absl::conditional_t<
-      // Fast path: if the value type can be trivially relocated, and we
-      // know the allocator doesn't do anything fancy, then it's safe for us
-      // to simply swap the bytes in the inline storage. It's as if we had
-      // relocated the first vector's elements into temporary storage,
-      // relocated the second's elements into the (now-empty) first's,
-      // and then relocated from temporary storage into the second.
-      absl::conjunction<absl::is_trivially_relocatable<ValueType<A>>,
-                        std::is_same<A, std::allocator<ValueType<A>>>>::value,
-      MemcpyPolicy,
-      absl::conditional_t<IsSwapOk<A>::value, ElementwiseSwapPolicy,
-                          ElementwiseConstructPolicy>>;
-
-  static SizeType<A> NextCapacity(SizeType<A> current_capacity) {
-    return current_capacity * 2;
-  }
-
-  static SizeType<A> ComputeCapacity(SizeType<A> current_capacity,
-                                     SizeType<A> requested_capacity) {
-    return (std::max)(NextCapacity(current_capacity), requested_capacity);
-  }
-
-  // ---------------------------------------------------------------------------
-  // Storage Constructors and Destructor
-  // ---------------------------------------------------------------------------
-
-  Storage() : metadata_(A(), /* size and is_allocated */ 0u) {}
-
-  explicit Storage(const A& allocator)
-      : metadata_(allocator, /* size and is_allocated */ 0u) {}
-
-  ~Storage() {
-    // Fast path: if we are empty and not allocated, there's nothing to do.
-    if (GetSizeAndIsAllocated() == 0) {
-      return;
-    }
-
-    // Fast path: if no destructors need to be run and we know the allocator
-    // doesn't do anything fancy, then all we need to do is deallocate (and
-    // maybe not even that).
-    if (absl::is_trivially_destructible<ValueType<A>>::value &&
-        std::is_same<A, std::allocator<ValueType<A>>>::value) {
-      DeallocateIfAllocated();
-      return;
-    }
-
-    DestroyContents();
-  }
-
-  // ---------------------------------------------------------------------------
-  // Storage Member Accessors
-  // ---------------------------------------------------------------------------
-
-  SizeType<A>& GetSizeAndIsAllocated() { return metadata_.template get<1>(); }
-
-  const SizeType<A>& GetSizeAndIsAllocated() const {
-    return metadata_.template get<1>();
-  }
-
-  SizeType<A> GetSize() const { return GetSizeAndIsAllocated() >> 1; }
-
-  bool GetIsAllocated() const { return GetSizeAndIsAllocated() & 1; }
-
-  Pointer<A> GetAllocatedData() {
-    // GCC 12 has a false-positive -Wmaybe-uninitialized warning here.
-#if ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(12, 0)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
-#endif
-    return data_.allocated.allocated_data;
-#if ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(12, 0)
-#pragma GCC diagnostic pop
-#endif
-  }
-
-  ConstPointer<A> GetAllocatedData() const {
-    return data_.allocated.allocated_data;
-  }
-
-  // ABSL_ATTRIBUTE_NO_SANITIZE_CFI is used because the memory pointed to may be
-  // uninitialized, a common pattern in allocate()+construct() APIs.
-  // https://clang.llvm.org/docs/ControlFlowIntegrity.html#bad-cast-checking
-  // NOTE: When this was written, LLVM documentation did not explicitly
-  // mention that casting `char*` and using `reinterpret_cast` qualifies
-  // as a bad cast.
-  ABSL_ATTRIBUTE_NO_SANITIZE_CFI Pointer<A> GetInlinedData() {
-    return reinterpret_cast<Pointer<A>>(data_.inlined.inlined_data);
-  }
-
-  ABSL_ATTRIBUTE_NO_SANITIZE_CFI ConstPointer<A> GetInlinedData() const {
-    return reinterpret_cast<ConstPointer<A>>(data_.inlined.inlined_data);
-  }
-
-  SizeType<A> GetAllocatedCapacity() const {
-    return data_.allocated.allocated_capacity;
-  }
-
-  SizeType<A> GetInlinedCapacity() const {
-    return static_cast<SizeType<A>>(kOptimalInlinedSize);
-  }
-
-  StorageView<A> MakeStorageView() {
-    return GetIsAllocated() ? StorageView<A>{GetAllocatedData(), GetSize(),
-                                             GetAllocatedCapacity()}
-                            : StorageView<A>{GetInlinedData(), GetSize(),
-                                             GetInlinedCapacity()};
-  }
-
-  A& GetAllocator() { return metadata_.template get<0>(); }
-
-  const A& GetAllocator() const { return metadata_.template get<0>(); }
-
-  // ---------------------------------------------------------------------------
-  // Storage Member Mutators
-  // ---------------------------------------------------------------------------
-
-  ABSL_ATTRIBUTE_NOINLINE void InitFrom(const Storage& other);
-
-  template <typename ValueAdapter>
-  void Initialize(ValueAdapter values, SizeType<A> new_size);
-
-  template <typename ValueAdapter>
-  void Assign(ValueAdapter values, SizeType<A> new_size);
-
-  template <typename ValueAdapter>
-  void Resize(ValueAdapter values, SizeType<A> new_size);
-
-  template <typename ValueAdapter>
-  Iterator<A> Insert(ConstIterator<A> pos, ValueAdapter values,
-                     SizeType<A> insert_count);
-
-  template <typename... Args>
-  Reference<A> EmplaceBack(Args&&... args);
-
-  Iterator<A> Erase(ConstIterator<A> from, ConstIterator<A> to);
-
-  void Reserve(SizeType<A> requested_capacity);
-
-  void ShrinkToFit();
-
-  void Swap(Storage* other_storage_ptr);
-
-  void SetIsAllocated() {
-    GetSizeAndIsAllocated() |= static_cast<SizeType<A>>(1);
-  }
-
-  void UnsetIsAllocated() {
-    GetSizeAndIsAllocated() &= ((std::numeric_limits<SizeType<A>>::max)() - 1);
-  }
-
-  void SetSize(SizeType<A> size) {
-    GetSizeAndIsAllocated() =
-        (size << 1) | static_cast<SizeType<A>>(GetIsAllocated());
-  }
-
-  void SetAllocatedSize(SizeType<A> size) {
-    GetSizeAndIsAllocated() = (size << 1) | static_cast<SizeType<A>>(1);
-  }
-
-  void SetInlinedSize(SizeType<A> size) {
-    GetSizeAndIsAllocated() = size << static_cast<SizeType<A>>(1);
-  }
-
-  void AddSize(SizeType<A> count) {
-    GetSizeAndIsAllocated() += count << static_cast<SizeType<A>>(1);
-  }
-
-  void SubtractSize(SizeType<A> count) {
-    ABSL_HARDENING_ASSERT(count <= GetSize());
-
-    GetSizeAndIsAllocated() -= count << static_cast<SizeType<A>>(1);
-  }
-
-  void SetAllocation(Allocation<A> allocation) {
-    data_.allocated.allocated_data = allocation.data;
-    data_.allocated.allocated_capacity = allocation.capacity;
-  }
-
-  void MemcpyFrom(const Storage& other_storage) {
-    // Assumption check: it doesn't make sense to memcpy inlined elements unless
-    // we know the allocator doesn't do anything fancy, and one of the following
-    // holds:
-    //
-    //  *  The elements are trivially relocatable.
-    //
-    //  *  It's possible to trivially assign the elements and then destroy the
-    //     source.
-    //
-    //  *  It's possible to trivially copy construct/assign the elements.
-    //
-    {
-      using V = ValueType<A>;
-      ABSL_HARDENING_ASSERT(
-          other_storage.GetIsAllocated() ||
-          (std::is_same<A, std::allocator<V>>::value &&
-           (
-               // First case above
-               absl::is_trivially_relocatable<V>::value ||
-               // Second case above
-               (absl::is_trivially_move_assignable<V>::value &&
-                absl::is_trivially_destructible<V>::value) ||
-               // Third case above
-               (absl::is_trivially_copy_constructible<V>::value ||
-                absl::is_trivially_copy_assignable<V>::value))));
-    }
-
-    GetSizeAndIsAllocated() = other_storage.GetSizeAndIsAllocated();
-    data_ = other_storage.data_;
-  }
-
-  void DeallocateIfAllocated() {
-    if (GetIsAllocated()) {
-      MallocAdapter<A>::Deallocate(GetAllocator(), GetAllocatedData(),
-                                   GetAllocatedCapacity());
-    }
-  }
-
- private:
-  ABSL_ATTRIBUTE_NOINLINE void DestroyContents();
-
-  using Metadata = container_internal::CompressedTuple<A, SizeType<A>>;
-
-  struct Allocated {
-    Pointer<A> allocated_data;
-    SizeType<A> allocated_capacity;
-  };
-
-  // `kOptimalInlinedSize` is an automatically adjusted inlined capacity of the
-  // `InlinedVector`. Sometimes, it is possible to increase the capacity (from
-  // the user requested `N`) without increasing the size of the `InlinedVector`.
-  static constexpr size_t kOptimalInlinedSize =
-      (std::max)(N, sizeof(Allocated) / sizeof(ValueType<A>));
-
-  struct Inlined {
-    alignas(ValueType<A>) char inlined_data[sizeof(
-        ValueType<A>[kOptimalInlinedSize])];
-  };
-
-  union Data {
-    Allocated allocated;
-    Inlined inlined;
-  };
-
-  void SwapN(ElementwiseSwapPolicy, Storage* other, SizeType<A> n);
-  void SwapN(ElementwiseConstructPolicy, Storage* other, SizeType<A> n);
-
-  void SwapInlinedElements(MemcpyPolicy, Storage* other);
-  template <typename NotMemcpyPolicy>
-  void SwapInlinedElements(NotMemcpyPolicy, Storage* other);
-
-  template <typename... Args>
-  ABSL_ATTRIBUTE_NOINLINE Reference<A> EmplaceBackSlow(Args&&... args);
-
-  Metadata metadata_;
-  Data data_;
-};
-
-template <typename T, size_t N, typename A>
-void Storage<T, N, A>::DestroyContents() {
-  Pointer<A> data = GetIsAllocated() ? GetAllocatedData() : GetInlinedData();
-  DestroyAdapter<A>::DestroyElements(GetAllocator(), data, GetSize());
-  DeallocateIfAllocated();
-}
-
-template <typename T, size_t N, typename A>
-void Storage<T, N, A>::InitFrom(const Storage& other) {
-  const SizeType<A> n = other.GetSize();
-  ABSL_HARDENING_ASSERT(n > 0);  // Empty sources handled handled in caller.
-  ConstPointer<A> src;
-  Pointer<A> dst;
-  if (!other.GetIsAllocated()) {
-    dst = GetInlinedData();
-    src = other.GetInlinedData();
-  } else {
-    // Because this is only called from the `InlinedVector` constructors, it's
-    // safe to take on the allocation with size `0`. If `ConstructElements(...)`
-    // throws, deallocation will be automatically handled by `~Storage()`.
-    SizeType<A> requested_capacity = ComputeCapacity(GetInlinedCapacity(), n);
-    Allocation<A> allocation =
-        MallocAdapter<A>::Allocate(GetAllocator(), requested_capacity);
-    SetAllocation(allocation);
-    dst = allocation.data;
-    src = other.GetAllocatedData();
-  }
-
-  // Fast path: if the value type is trivially copy constructible and we know
-  // the allocator doesn't do anything fancy, then we know it is legal for us to
-  // simply memcpy the other vector's elements.
-  if (absl::is_trivially_copy_constructible<ValueType<A>>::value &&
-      std::is_same<A, std::allocator<ValueType<A>>>::value) {
-    std::memcpy(reinterpret_cast<char*>(dst),
-                reinterpret_cast<const char*>(src), n * sizeof(ValueType<A>));
-  } else {
-    auto values = IteratorValueAdapter<A, ConstPointer<A>>(src);
-    ConstructElements<A>(GetAllocator(), dst, values, n);
-  }
-
-  GetSizeAndIsAllocated() = other.GetSizeAndIsAllocated();
-}
-
-template <typename T, size_t N, typename A>
-template <typename ValueAdapter>
-auto Storage<T, N, A>::Initialize(ValueAdapter values,
-                                  SizeType<A> new_size) -> void {
-  // Only callable from constructors!
-  ABSL_HARDENING_ASSERT(!GetIsAllocated());
-  ABSL_HARDENING_ASSERT(GetSize() == 0);
-
-  Pointer<A> construct_data;
-  if (new_size > GetInlinedCapacity()) {
-    // Because this is only called from the `InlinedVector` constructors, it's
-    // safe to take on the allocation with size `0`. If `ConstructElements(...)`
-    // throws, deallocation will be automatically handled by `~Storage()`.
-    SizeType<A> requested_capacity =
-        ComputeCapacity(GetInlinedCapacity(), new_size);
-    Allocation<A> allocation =
-        MallocAdapter<A>::Allocate(GetAllocator(), requested_capacity);
-    construct_data = allocation.data;
-    SetAllocation(allocation);
-    SetIsAllocated();
-  } else {
-    construct_data = GetInlinedData();
-  }
-
-  ConstructElements<A>(GetAllocator(), construct_data, values, new_size);
-
-  // Since the initial size was guaranteed to be `0` and the allocated bit is
-  // already correct for either case, *adding* `new_size` gives us the correct
-  // result faster than setting it directly.
-  AddSize(new_size);
-}
-
-template <typename T, size_t N, typename A>
-template <typename ValueAdapter>
-auto Storage<T, N, A>::Assign(ValueAdapter values,
-                              SizeType<A> new_size) -> void {
-  StorageView<A> storage_view = MakeStorageView();
-
-  AllocationTransaction<A> allocation_tx(GetAllocator());
-
-  absl::Span<ValueType<A>> assign_loop;
-  absl::Span<ValueType<A>> construct_loop;
-  absl::Span<ValueType<A>> destroy_loop;
-
-  if (new_size > storage_view.capacity) {
-    SizeType<A> requested_capacity =
-        ComputeCapacity(storage_view.capacity, new_size);
-    construct_loop = {allocation_tx.Allocate(requested_capacity), new_size};
-    destroy_loop = {storage_view.data, storage_view.size};
-  } else if (new_size > storage_view.size) {
-    assign_loop = {storage_view.data, storage_view.size};
-    construct_loop = {storage_view.data + storage_view.size,
-                      new_size - storage_view.size};
-  } else {
-    assign_loop = {storage_view.data, new_size};
-    destroy_loop = {storage_view.data + new_size, storage_view.size - new_size};
-  }
-
-  AssignElements<A>(assign_loop.data(), values, assign_loop.size());
-
-  ConstructElements<A>(GetAllocator(), construct_loop.data(), values,
-                       construct_loop.size());
-
-  DestroyAdapter<A>::DestroyElements(GetAllocator(), destroy_loop.data(),
-                                     destroy_loop.size());
-
-  if (allocation_tx.DidAllocate()) {
-    DeallocateIfAllocated();
-    SetAllocation(std::move(allocation_tx).Release());
-    SetIsAllocated();
-  }
-
-  SetSize(new_size);
-}
-
-template <typename T, size_t N, typename A>
-template <typename ValueAdapter>
-auto Storage<T, N, A>::Resize(ValueAdapter values,
-                              SizeType<A> new_size) -> void {
-  StorageView<A> storage_view = MakeStorageView();
-  Pointer<A> const base = storage_view.data;
-  const SizeType<A> size = storage_view.size;
-  A& alloc = GetAllocator();
-  if (new_size <= size) {
-    // Destroy extra old elements.
-    DestroyAdapter<A>::DestroyElements(alloc, base + new_size, size - new_size);
-  } else if (new_size <= storage_view.capacity) {
-    // Construct new elements in place.
-    ConstructElements<A>(alloc, base + size, values, new_size - size);
-  } else {
-    // Steps:
-    //  a. Allocate new backing store.
-    //  b. Construct new elements in new backing store.
-    //  c. Move existing elements from old backing store to new backing store.
-    //  d. Destroy all elements in old backing store.
-    // Use transactional wrappers for the first two steps so we can roll
-    // back if necessary due to exceptions.
-    AllocationTransaction<A> allocation_tx(alloc);
-    SizeType<A> requested_capacity =
-        ComputeCapacity(storage_view.capacity, new_size);
-    Pointer<A> new_data = allocation_tx.Allocate(requested_capacity);
-
-    ConstructionTransaction<A> construction_tx(alloc);
-    construction_tx.Construct(new_data + size, values, new_size - size);
-
-    IteratorValueAdapter<A, MoveIterator<A>> move_values(
-        (MoveIterator<A>(base)));
-    ConstructElements<A>(alloc, new_data, move_values, size);
-
-    DestroyAdapter<A>::DestroyElements(alloc, base, size);
-    std::move(construction_tx).Commit();
-    DeallocateIfAllocated();
-    SetAllocation(std::move(allocation_tx).Release());
-    SetIsAllocated();
-  }
-  SetSize(new_size);
-}
-
-template <typename T, size_t N, typename A>
-template <typename ValueAdapter>
-auto Storage<T, N, A>::Insert(ConstIterator<A> pos, ValueAdapter values,
-                              SizeType<A> insert_count) -> Iterator<A> {
-  StorageView<A> storage_view = MakeStorageView();
-
-  auto insert_index = static_cast<SizeType<A>>(
-      std::distance(ConstIterator<A>(storage_view.data), pos));
-  SizeType<A> insert_end_index = insert_index + insert_count;
-  SizeType<A> new_size = storage_view.size + insert_count;
-
-  if (new_size > storage_view.capacity) {
-    AllocationTransaction<A> allocation_tx(GetAllocator());
-    ConstructionTransaction<A> construction_tx(GetAllocator());
-    ConstructionTransaction<A> move_construction_tx(GetAllocator());
-
-    IteratorValueAdapter<A, MoveIterator<A>> move_values(
-        MoveIterator<A>(storage_view.data));
-
-    SizeType<A> requested_capacity =
-        ComputeCapacity(storage_view.capacity, new_size);
-    Pointer<A> new_data = allocation_tx.Allocate(requested_capacity);
-
-    construction_tx.Construct(new_data + insert_index, values, insert_count);
-
-    move_construction_tx.Construct(new_data, move_values, insert_index);
-
-    ConstructElements<A>(GetAllocator(), new_data + insert_end_index,
-                         move_values, storage_view.size - insert_index);
-
-    DestroyAdapter<A>::DestroyElements(GetAllocator(), storage_view.data,
-                                       storage_view.size);
-
-    std::move(construction_tx).Commit();
-    std::move(move_construction_tx).Commit();
-    DeallocateIfAllocated();
-    SetAllocation(std::move(allocation_tx).Release());
-
-    SetAllocatedSize(new_size);
-    return Iterator<A>(new_data + insert_index);
-  } else {
-    SizeType<A> move_construction_destination_index =
-        (std::max)(insert_end_index, storage_view.size);
-
-    ConstructionTransaction<A> move_construction_tx(GetAllocator());
-
-    IteratorValueAdapter<A, MoveIterator<A>> move_construction_values(
-        MoveIterator<A>(storage_view.data +
-                        (move_construction_destination_index - insert_count)));
-    absl::Span<ValueType<A>> move_construction = {
-        storage_view.data + move_construction_destination_index,
-        new_size - move_construction_destination_index};
-
-    Pointer<A> move_assignment_values = storage_view.data + insert_index;
-    absl::Span<ValueType<A>> move_assignment = {
-        storage_view.data + insert_end_index,
-        move_construction_destination_index - insert_end_index};
-
-    absl::Span<ValueType<A>> insert_assignment = {move_assignment_values,
-                                                  move_construction.size()};
-
-    absl::Span<ValueType<A>> insert_construction = {
-        insert_assignment.data() + insert_assignment.size(),
-        insert_count - insert_assignment.size()};
-
-    move_construction_tx.Construct(move_construction.data(),
-                                   move_construction_values,
-                                   move_construction.size());
-
-    for (Pointer<A>
-             destination = move_assignment.data() + move_assignment.size(),
-             last_destination = move_assignment.data(),
-             source = move_assignment_values + move_assignment.size();
-         ;) {
-      --destination;
-      --source;
-      if (destination < last_destination) break;
-      *destination = std::move(*source);
-    }
-
-    AssignElements<A>(insert_assignment.data(), values,
-                      insert_assignment.size());
-
-    ConstructElements<A>(GetAllocator(), insert_construction.data(), values,
-                         insert_construction.size());
-
-    std::move(move_construction_tx).Commit();
-
-    AddSize(insert_count);
-    return Iterator<A>(storage_view.data + insert_index);
-  }
-}
-
-template <typename T, size_t N, typename A>
-template <typename... Args>
-auto Storage<T, N, A>::EmplaceBack(Args&&... args) -> Reference<A> {
-  StorageView<A> storage_view = MakeStorageView();
-  const SizeType<A> n = storage_view.size;
-  if (ABSL_PREDICT_TRUE(n != storage_view.capacity)) {
-    // Fast path; new element fits.
-    Pointer<A> last_ptr = storage_view.data + n;
-    AllocatorTraits<A>::construct(GetAllocator(), last_ptr,
-                                  std::forward<Args>(args)...);
-    AddSize(1);
-    return *last_ptr;
-  }
-  // TODO(b/173712035): Annotate with musttail attribute to prevent regression.
-  return EmplaceBackSlow(std::forward<Args>(args)...);
-}
-
-template <typename T, size_t N, typename A>
-template <typename... Args>
-auto Storage<T, N, A>::EmplaceBackSlow(Args&&... args) -> Reference<A> {
-  StorageView<A> storage_view = MakeStorageView();
-  AllocationTransaction<A> allocation_tx(GetAllocator());
-  IteratorValueAdapter<A, MoveIterator<A>> move_values(
-      MoveIterator<A>(storage_view.data));
-  SizeType<A> requested_capacity = NextCapacity(storage_view.capacity);
-  Pointer<A> construct_data = allocation_tx.Allocate(requested_capacity);
-  Pointer<A> last_ptr = construct_data + storage_view.size;
-
-  // Construct new element.
-  AllocatorTraits<A>::construct(GetAllocator(), last_ptr,
-                                std::forward<Args>(args)...);
-  // Move elements from old backing store to new backing store.
-  ABSL_INTERNAL_TRY {
-    ConstructElements<A>(GetAllocator(), allocation_tx.GetData(), move_values,
-                         storage_view.size);
-  }
-  ABSL_INTERNAL_CATCH_ANY {
-    AllocatorTraits<A>::destroy(GetAllocator(), last_ptr);
-    ABSL_INTERNAL_RETHROW;
-  }
-  // Destroy elements in old backing store.
-  DestroyAdapter<A>::DestroyElements(GetAllocator(), storage_view.data,
-                                     storage_view.size);
-
-  DeallocateIfAllocated();
-  SetAllocation(std::move(allocation_tx).Release());
-  SetIsAllocated();
-  AddSize(1);
-  return *last_ptr;
-}
-
-template <typename T, size_t N, typename A>
-auto Storage<T, N, A>::Erase(ConstIterator<A> from,
-                             ConstIterator<A> to) -> Iterator<A> {
-  StorageView<A> storage_view = MakeStorageView();
-
-  auto erase_size = static_cast<SizeType<A>>(std::distance(from, to));
-  auto erase_index = static_cast<SizeType<A>>(
-      std::distance(ConstIterator<A>(storage_view.data), from));
-  SizeType<A> erase_end_index = erase_index + erase_size;
-
-  // Fast path: if the value type is trivially relocatable and we know
-  // the allocator doesn't do anything fancy, then we know it is legal for us to
-  // simply destroy the elements in the "erasure window" (which cannot throw)
-  // and then memcpy downward to close the window.
-  if (absl::is_trivially_relocatable<ValueType<A>>::value &&
-      std::is_nothrow_destructible<ValueType<A>>::value &&
-      std::is_same<A, std::allocator<ValueType<A>>>::value) {
-    DestroyAdapter<A>::DestroyElements(
-        GetAllocator(), storage_view.data + erase_index, erase_size);
-    std::memmove(
-        reinterpret_cast<char*>(storage_view.data + erase_index),
-        reinterpret_cast<const char*>(storage_view.data + erase_end_index),
-        (storage_view.size - erase_end_index) * sizeof(ValueType<A>));
-  } else {
-    IteratorValueAdapter<A, MoveIterator<A>> move_values(
-        MoveIterator<A>(storage_view.data + erase_end_index));
-
-    AssignElements<A>(storage_view.data + erase_index, move_values,
-                      storage_view.size - erase_end_index);
-
-    DestroyAdapter<A>::DestroyElements(
-        GetAllocator(), storage_view.data + (storage_view.size - erase_size),
-        erase_size);
-  }
-  SubtractSize(erase_size);
-  return Iterator<A>(storage_view.data + erase_index);
-}
-
-template <typename T, size_t N, typename A>
-auto Storage<T, N, A>::Reserve(SizeType<A> requested_capacity) -> void {
-  StorageView<A> storage_view = MakeStorageView();
-
-  if (ABSL_PREDICT_FALSE(requested_capacity <= storage_view.capacity)) return;
-
-  AllocationTransaction<A> allocation_tx(GetAllocator());
-
-  IteratorValueAdapter<A, MoveIterator<A>> move_values(
-      MoveIterator<A>(storage_view.data));
-
-  SizeType<A> new_requested_capacity =
-      ComputeCapacity(storage_view.capacity, requested_capacity);
-  Pointer<A> new_data = allocation_tx.Allocate(new_requested_capacity);
-
-  ConstructElements<A>(GetAllocator(), new_data, move_values,
-                       storage_view.size);
-
-  DestroyAdapter<A>::DestroyElements(GetAllocator(), storage_view.data,
-                                     storage_view.size);
-
-  DeallocateIfAllocated();
-  SetAllocation(std::move(allocation_tx).Release());
-  SetIsAllocated();
-}
-
-template <typename T, size_t N, typename A>
-auto Storage<T, N, A>::ShrinkToFit() -> void {
-  // May only be called on allocated instances!
-  ABSL_HARDENING_ASSERT(GetIsAllocated());
-
-  StorageView<A> storage_view{GetAllocatedData(), GetSize(),
-                              GetAllocatedCapacity()};
-
-  if (ABSL_PREDICT_FALSE(storage_view.size == storage_view.capacity)) return;
-
-  AllocationTransaction<A> allocation_tx(GetAllocator());
-
-  IteratorValueAdapter<A, MoveIterator<A>> move_values(
-      MoveIterator<A>(storage_view.data));
-
-  Pointer<A> construct_data;
-  if (storage_view.size > GetInlinedCapacity()) {
-    SizeType<A> requested_capacity = storage_view.size;
-    construct_data = allocation_tx.Allocate(requested_capacity);
-    if (allocation_tx.GetCapacity() >= storage_view.capacity) {
-      // Already using the smallest available heap allocation.
-      return;
-    }
-  } else {
-    construct_data = GetInlinedData();
-  }
-
-  ABSL_INTERNAL_TRY {
-    ConstructElements<A>(GetAllocator(), construct_data, move_values,
-                         storage_view.size);
-  }
-  ABSL_INTERNAL_CATCH_ANY {
-    SetAllocation({storage_view.data, storage_view.capacity});
-    ABSL_INTERNAL_RETHROW;
-  }
-
-  DestroyAdapter<A>::DestroyElements(GetAllocator(), storage_view.data,
-                                     storage_view.size);
-
-  MallocAdapter<A>::Deallocate(GetAllocator(), storage_view.data,
-                               storage_view.capacity);
-
-  if (allocation_tx.DidAllocate()) {
-    SetAllocation(std::move(allocation_tx).Release());
-  } else {
-    UnsetIsAllocated();
-  }
-}
-
-template <typename T, size_t N, typename A>
-auto Storage<T, N, A>::Swap(Storage* other_storage_ptr) -> void {
-  using std::swap;
-  ABSL_HARDENING_ASSERT(this != other_storage_ptr);
-
-  if (GetIsAllocated() && other_storage_ptr->GetIsAllocated()) {
-    swap(data_.allocated, other_storage_ptr->data_.allocated);
-  } else if (!GetIsAllocated() && !other_storage_ptr->GetIsAllocated()) {
-    SwapInlinedElements(SwapInlinedElementsPolicy{}, other_storage_ptr);
-  } else {
-    Storage* allocated_ptr = this;
-    Storage* inlined_ptr = other_storage_ptr;
-    if (!allocated_ptr->GetIsAllocated()) swap(allocated_ptr, inlined_ptr);
-
-    StorageView<A> allocated_storage_view{
-        allocated_ptr->GetAllocatedData(), allocated_ptr->GetSize(),
-        allocated_ptr->GetAllocatedCapacity()};
-
-    IteratorValueAdapter<A, MoveIterator<A>> move_values(
-        MoveIterator<A>(inlined_ptr->GetInlinedData()));
-
-    ABSL_INTERNAL_TRY {
-      ConstructElements<A>(inlined_ptr->GetAllocator(),
-                           allocated_ptr->GetInlinedData(), move_values,
-                           inlined_ptr->GetSize());
-    }
-    ABSL_INTERNAL_CATCH_ANY {
-      allocated_ptr->SetAllocation(Allocation<A>{
-          allocated_storage_view.data, allocated_storage_view.capacity});
-      ABSL_INTERNAL_RETHROW;
-    }
-
-    DestroyAdapter<A>::DestroyElements(inlined_ptr->GetAllocator(),
-                                       inlined_ptr->GetInlinedData(),
-                                       inlined_ptr->GetSize());
-
-    inlined_ptr->SetAllocation(Allocation<A>{allocated_storage_view.data,
-                                             allocated_storage_view.capacity});
-  }
-
-  swap(GetSizeAndIsAllocated(), other_storage_ptr->GetSizeAndIsAllocated());
-  swap(GetAllocator(), other_storage_ptr->GetAllocator());
-}
-
-template <typename T, size_t N, typename A>
-void Storage<T, N, A>::SwapN(ElementwiseSwapPolicy, Storage* other,
-                             SizeType<A> n) {
-  std::swap_ranges(GetInlinedData(), GetInlinedData() + n,
-                   other->GetInlinedData());
-}
-
-template <typename T, size_t N, typename A>
-void Storage<T, N, A>::SwapN(ElementwiseConstructPolicy, Storage* other,
-                             SizeType<A> n) {
-  Pointer<A> a = GetInlinedData();
-  Pointer<A> b = other->GetInlinedData();
-  // see note on allocators in `SwapInlinedElements`.
-  A& allocator_a = GetAllocator();
-  A& allocator_b = other->GetAllocator();
-  for (SizeType<A> i = 0; i < n; ++i, ++a, ++b) {
-    ValueType<A> tmp(std::move(*a));
-
-    AllocatorTraits<A>::destroy(allocator_a, a);
-    AllocatorTraits<A>::construct(allocator_b, a, std::move(*b));
-
-    AllocatorTraits<A>::destroy(allocator_b, b);
-    AllocatorTraits<A>::construct(allocator_a, b, std::move(tmp));
-  }
-}
-
-template <typename T, size_t N, typename A>
-void Storage<T, N, A>::SwapInlinedElements(MemcpyPolicy, Storage* other) {
-  Data tmp = data_;
-  data_ = other->data_;
-  other->data_ = tmp;
-}
-
-template <typename T, size_t N, typename A>
-template <typename NotMemcpyPolicy>
-void Storage<T, N, A>::SwapInlinedElements(NotMemcpyPolicy policy,
-                                           Storage* other) {
-  // Note: `destroy` needs to use pre-swap allocator while `construct` -
-  // post-swap allocator. Allocators will be swapped later on outside of
-  // `SwapInlinedElements`.
-  Storage* small_ptr = this;
-  Storage* large_ptr = other;
-  if (small_ptr->GetSize() > large_ptr->GetSize()) {
-    std::swap(small_ptr, large_ptr);
-  }
-
-  auto small_size = small_ptr->GetSize();
-  auto diff = large_ptr->GetSize() - small_size;
-  SwapN(policy, other, small_size);
-
-  IteratorValueAdapter<A, MoveIterator<A>> move_values(
-      MoveIterator<A>(large_ptr->GetInlinedData() + small_size));
-
-  ConstructElements<A>(large_ptr->GetAllocator(),
-                       small_ptr->GetInlinedData() + small_size, move_values,
-                       diff);
-
-  DestroyAdapter<A>::DestroyElements(large_ptr->GetAllocator(),
-                                     large_ptr->GetInlinedData() + small_size,
-                                     diff);
-}
-
-// End ignore "array-bounds"
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic pop
-#endif
-
-}  // namespace inlined_vector_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_INLINED_VECTOR_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/layout.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/layout.h
deleted file mode 100644
index 1bf739cc..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/layout.h
+++ /dev/null
@@ -1,728 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-//                           MOTIVATION AND TUTORIAL
-//
-// If you want to put in a single heap allocation N doubles followed by M ints,
-// it's easy if N and M are known at compile time.
-//
-//   struct S {
-//     double a[N];
-//     int b[M];
-//   };
-//
-//   S* p = new S;
-//
-// But what if N and M are known only in run time? Class template Layout to the
-// rescue! It's a portable generalization of the technique known as struct hack.
-//
-//   // This object will tell us everything we need to know about the memory
-//   // layout of double[N] followed by int[M]. It's structurally identical to
-//   // size_t[2] that stores N and M. It's very cheap to create.
-//   const Layout<double, int> layout(N, M);
-//
-//   // Allocate enough memory for both arrays. `AllocSize()` tells us how much
-//   // memory is needed. We are free to use any allocation function we want as
-//   // long as it returns aligned memory.
-//   std::unique_ptr<unsigned char[]> p(new unsigned char[layout.AllocSize()]);
-//
-//   // Obtain the pointer to the array of doubles.
-//   // Equivalent to `reinterpret_cast<double*>(p.get())`.
-//   //
-//   // We could have written layout.Pointer<0>(p) instead. If all the types are
-//   // unique you can use either form, but if some types are repeated you must
-//   // use the index form.
-//   double* a = layout.Pointer<double>(p.get());
-//
-//   // Obtain the pointer to the array of ints.
-//   // Equivalent to `reinterpret_cast<int*>(p.get() + N * 8)`.
-//   int* b = layout.Pointer<int>(p);
-//
-// If we are unable to specify sizes of all fields, we can pass as many sizes as
-// we can to `Partial()`. In return, it'll allow us to access the fields whose
-// locations and sizes can be computed from the provided information.
-// `Partial()` comes in handy when the array sizes are embedded into the
-// allocation.
-//
-//   // size_t[0] containing N, size_t[1] containing M, double[N], int[M].
-//   using L = Layout<size_t, size_t, double, int>;
-//
-//   unsigned char* Allocate(size_t n, size_t m) {
-//     const L layout(1, 1, n, m);
-//     unsigned char* p = new unsigned char[layout.AllocSize()];
-//     *layout.Pointer<0>(p) = n;
-//     *layout.Pointer<1>(p) = m;
-//     return p;
-//   }
-//
-//   void Use(unsigned char* p) {
-//     // First, extract N and M.
-//     // Specify that the first array has only one element. Using `prefix` we
-//     // can access the first two arrays but not more.
-//     constexpr auto prefix = L::Partial(1);
-//     size_t n = *prefix.Pointer<0>(p);
-//     size_t m = *prefix.Pointer<1>(p);
-//
-//     // Now we can get pointers to the payload.
-//     const L layout(1, 1, n, m);
-//     double* a = layout.Pointer<double>(p);
-//     int* b = layout.Pointer<int>(p);
-//   }
-//
-// The layout we used above combines fixed-size with dynamically-sized fields.
-// This is quite common. Layout is optimized for this use case and generates
-// optimal code. All computations that can be performed at compile time are
-// indeed performed at compile time.
-//
-// Efficiency tip: The order of fields matters. In `Layout<T1, ..., TN>` try to
-// ensure that `alignof(T1) >= ... >= alignof(TN)`. This way you'll have no
-// padding in between arrays.
-//
-// You can manually override the alignment of an array by wrapping the type in
-// `Aligned<T, N>`. `Layout<..., Aligned<T, N>, ...>` has exactly the same API
-// and behavior as `Layout<..., T, ...>` except that the first element of the
-// array of `T` is aligned to `N` (the rest of the elements follow without
-// padding). `N` cannot be less than `alignof(T)`.
-//
-// `AllocSize()` and `Pointer()` are the most basic methods for dealing with
-// memory layouts. Check out the reference or code below to discover more.
-//
-//                            EXAMPLE
-//
-//   // Immutable move-only string with sizeof equal to sizeof(void*). The
-//   // string size and the characters are kept in the same heap allocation.
-//   class CompactString {
-//    public:
-//     CompactString(const char* s = "") {
-//       const size_t size = strlen(s);
-//       // size_t[1] followed by char[size + 1].
-//       const L layout(1, size + 1);
-//       p_.reset(new unsigned char[layout.AllocSize()]);
-//       // If running under ASAN, mark the padding bytes, if any, to catch
-//       // memory errors.
-//       layout.PoisonPadding(p_.get());
-//       // Store the size in the allocation.
-//       *layout.Pointer<size_t>(p_.get()) = size;
-//       // Store the characters in the allocation.
-//       memcpy(layout.Pointer<char>(p_.get()), s, size + 1);
-//     }
-//
-//     size_t size() const {
-//       // Equivalent to reinterpret_cast<size_t&>(*p).
-//       return *L::Partial().Pointer<size_t>(p_.get());
-//     }
-//
-//     const char* c_str() const {
-//       // Equivalent to reinterpret_cast<char*>(p.get() + sizeof(size_t)).
-//       // The argument in Partial(1) specifies that we have size_t[1] in front
-//       // of the characters.
-//       return L::Partial(1).Pointer<char>(p_.get());
-//     }
-//
-//    private:
-//     // Our heap allocation contains a size_t followed by an array of chars.
-//     using L = Layout<size_t, char>;
-//     std::unique_ptr<unsigned char[]> p_;
-//   };
-//
-//   int main() {
-//     CompactString s = "hello";
-//     assert(s.size() == 5);
-//     assert(strcmp(s.c_str(), "hello") == 0);
-//   }
-//
-//                               DOCUMENTATION
-//
-// The interface exported by this file consists of:
-// - class `Layout<>` and its public members.
-// - The public members of class `internal_layout::LayoutImpl<>`. That class
-//   isn't intended to be used directly, and its name and template parameter
-//   list are internal implementation details, but the class itself provides
-//   most of the functionality in this file. See comments on its members for
-//   detailed documentation.
-//
-// `Layout<T1,... Tn>::Partial(count1,..., countm)` (where `m` <= `n`) returns a
-// `LayoutImpl<>` object. `Layout<T1,..., Tn> layout(count1,..., countn)`
-// creates a `Layout` object, which exposes the same functionality by inheriting
-// from `LayoutImpl<>`.
-
-#ifndef ABSL_CONTAINER_INTERNAL_LAYOUT_H_
-#define ABSL_CONTAINER_INTERNAL_LAYOUT_H_
-
-#include <assert.h>
-#include <stddef.h>
-#include <stdint.h>
-
-#include <ostream>
-#include <string>
-#include <tuple>
-#include <type_traits>
-#include <typeinfo>
-#include <utility>
-
-#include "absl/base/config.h"
-#include "absl/debugging/internal/demangle.h"
-#include "absl/meta/type_traits.h"
-#include "absl/strings/str_cat.h"
-#include "absl/types/span.h"
-#include "absl/utility/utility.h"
-
-#ifdef ABSL_HAVE_ADDRESS_SANITIZER
-#include <sanitizer/asan_interface.h>
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-// A type wrapper that instructs `Layout` to use the specific alignment for the
-// array. `Layout<..., Aligned<T, N>, ...>` has exactly the same API
-// and behavior as `Layout<..., T, ...>` except that the first element of the
-// array of `T` is aligned to `N` (the rest of the elements follow without
-// padding).
-//
-// Requires: `N >= alignof(T)` and `N` is a power of 2.
-template <class T, size_t N>
-struct Aligned;
-
-namespace internal_layout {
-
-template <class T>
-struct NotAligned {};
-
-template <class T, size_t N>
-struct NotAligned<const Aligned<T, N>> {
-  static_assert(sizeof(T) == 0, "Aligned<T, N> cannot be const-qualified");
-};
-
-template <size_t>
-using IntToSize = size_t;
-
-template <class>
-using TypeToSize = size_t;
-
-template <class T>
-struct Type : NotAligned<T> {
-  using type = T;
-};
-
-template <class T, size_t N>
-struct Type<Aligned<T, N>> {
-  using type = T;
-};
-
-template <class T>
-struct SizeOf : NotAligned<T>, std::integral_constant<size_t, sizeof(T)> {};
-
-template <class T, size_t N>
-struct SizeOf<Aligned<T, N>> : std::integral_constant<size_t, sizeof(T)> {};
-
-// Note: workaround for https://gcc.gnu.org/PR88115
-template <class T>
-struct AlignOf : NotAligned<T> {
-  static constexpr size_t value = alignof(T);
-};
-
-template <class T, size_t N>
-struct AlignOf<Aligned<T, N>> {
-  static_assert(N % alignof(T) == 0,
-                "Custom alignment can't be lower than the type's alignment");
-  static constexpr size_t value = N;
-};
-
-// Does `Ts...` contain `T`?
-template <class T, class... Ts>
-using Contains = absl::disjunction<std::is_same<T, Ts>...>;
-
-template <class From, class To>
-using CopyConst =
-    typename std::conditional<std::is_const<From>::value, const To, To>::type;
-
-// Note: We're not qualifying this with absl:: because it doesn't compile under
-// MSVC.
-template <class T>
-using SliceType = Span<T>;
-
-// This namespace contains no types. It prevents functions defined in it from
-// being found by ADL.
-namespace adl_barrier {
-
-template <class Needle, class... Ts>
-constexpr size_t Find(Needle, Needle, Ts...) {
-  static_assert(!Contains<Needle, Ts...>(), "Duplicate element type");
-  return 0;
-}
-
-template <class Needle, class T, class... Ts>
-constexpr size_t Find(Needle, T, Ts...) {
-  return adl_barrier::Find(Needle(), Ts()...) + 1;
-}
-
-constexpr bool IsPow2(size_t n) { return !(n & (n - 1)); }
-
-// Returns `q * m` for the smallest `q` such that `q * m >= n`.
-// Requires: `m` is a power of two. It's enforced by IsLegalElementType below.
-constexpr size_t Align(size_t n, size_t m) { return (n + m - 1) & ~(m - 1); }
-
-constexpr size_t Min(size_t a, size_t b) { return b < a ? b : a; }
-
-constexpr size_t Max(size_t a) { return a; }
-
-template <class... Ts>
-constexpr size_t Max(size_t a, size_t b, Ts... rest) {
-  return adl_barrier::Max(b < a ? a : b, rest...);
-}
-
-template <class T>
-std::string TypeName() {
-  std::string out;
-#if ABSL_INTERNAL_HAS_RTTI
-  absl::StrAppend(&out, "<",
-                  absl::debugging_internal::DemangleString(typeid(T).name()),
-                  ">");
-#endif
-  return out;
-}
-
-}  // namespace adl_barrier
-
-template <bool C>
-using EnableIf = typename std::enable_if<C, int>::type;
-
-// Can `T` be a template argument of `Layout`?
-template <class T>
-using IsLegalElementType = std::integral_constant<
-    bool, !std::is_reference<T>::value && !std::is_volatile<T>::value &&
-              !std::is_reference<typename Type<T>::type>::value &&
-              !std::is_volatile<typename Type<T>::type>::value &&
-              adl_barrier::IsPow2(AlignOf<T>::value)>;
-
-template <class Elements, class SizeSeq, class OffsetSeq>
-class LayoutImpl;
-
-// Public base class of `Layout` and the result type of `Layout::Partial()`.
-//
-// `Elements...` contains all template arguments of `Layout` that created this
-// instance.
-//
-// `SizeSeq...` is `[0, NumSizes)` where `NumSizes` is the number of arguments
-// passed to `Layout::Partial()` or `Layout::Layout()`.
-//
-// `OffsetSeq...` is `[0, NumOffsets)` where `NumOffsets` is
-// `Min(sizeof...(Elements), NumSizes + 1)` (the number of arrays for which we
-// can compute offsets).
-template <class... Elements, size_t... SizeSeq, size_t... OffsetSeq>
-class LayoutImpl<std::tuple<Elements...>, absl::index_sequence<SizeSeq...>,
-                 absl::index_sequence<OffsetSeq...>> {
- private:
-  static_assert(sizeof...(Elements) > 0, "At least one field is required");
-  static_assert(absl::conjunction<IsLegalElementType<Elements>...>::value,
-                "Invalid element type (see IsLegalElementType)");
-
-  enum {
-    NumTypes = sizeof...(Elements),
-    NumSizes = sizeof...(SizeSeq),
-    NumOffsets = sizeof...(OffsetSeq),
-  };
-
-  // These are guaranteed by `Layout`.
-  static_assert(NumOffsets == adl_barrier::Min(NumTypes, NumSizes + 1),
-                "Internal error");
-  static_assert(NumTypes > 0, "Internal error");
-
-  // Returns the index of `T` in `Elements...`. Results in a compilation error
-  // if `Elements...` doesn't contain exactly one instance of `T`.
-  template <class T>
-  static constexpr size_t ElementIndex() {
-    static_assert(Contains<Type<T>, Type<typename Type<Elements>::type>...>(),
-                  "Type not found");
-    return adl_barrier::Find(Type<T>(),
-                             Type<typename Type<Elements>::type>()...);
-  }
-
-  template <size_t N>
-  using ElementAlignment =
-      AlignOf<typename std::tuple_element<N, std::tuple<Elements...>>::type>;
-
- public:
-  // Element types of all arrays packed in a tuple.
-  using ElementTypes = std::tuple<typename Type<Elements>::type...>;
-
-  // Element type of the Nth array.
-  template <size_t N>
-  using ElementType = typename std::tuple_element<N, ElementTypes>::type;
-
-  constexpr explicit LayoutImpl(IntToSize<SizeSeq>... sizes)
-      : size_{sizes...} {}
-
-  // Alignment of the layout, equal to the strictest alignment of all elements.
-  // All pointers passed to the methods of layout must be aligned to this value.
-  static constexpr size_t Alignment() {
-    return adl_barrier::Max(AlignOf<Elements>::value...);
-  }
-
-  // Offset in bytes of the Nth array.
-  //
-  //   // int[3], 4 bytes of padding, double[4].
-  //   Layout<int, double> x(3, 4);
-  //   assert(x.Offset<0>() == 0);   // The ints starts from 0.
-  //   assert(x.Offset<1>() == 16);  // The doubles starts from 16.
-  //
-  // Requires: `N <= NumSizes && N < sizeof...(Ts)`.
-  template <size_t N, EnableIf<N == 0> = 0>
-  constexpr size_t Offset() const {
-    return 0;
-  }
-
-  template <size_t N, EnableIf<N != 0> = 0>
-  constexpr size_t Offset() const {
-    static_assert(N < NumOffsets, "Index out of bounds");
-    return adl_barrier::Align(
-        Offset<N - 1>() + SizeOf<ElementType<N - 1>>::value * size_[N - 1],
-        ElementAlignment<N>::value);
-  }
-
-  // Offset in bytes of the array with the specified element type. There must
-  // be exactly one such array and its zero-based index must be at most
-  // `NumSizes`.
-  //
-  //   // int[3], 4 bytes of padding, double[4].
-  //   Layout<int, double> x(3, 4);
-  //   assert(x.Offset<int>() == 0);      // The ints starts from 0.
-  //   assert(x.Offset<double>() == 16);  // The doubles starts from 16.
-  template <class T>
-  constexpr size_t Offset() const {
-    return Offset<ElementIndex<T>()>();
-  }
-
-  // Offsets in bytes of all arrays for which the offsets are known.
-  constexpr std::array<size_t, NumOffsets> Offsets() const {
-    return {{Offset<OffsetSeq>()...}};
-  }
-
-  // The number of elements in the Nth array. This is the Nth argument of
-  // `Layout::Partial()` or `Layout::Layout()` (zero-based).
-  //
-  //   // int[3], 4 bytes of padding, double[4].
-  //   Layout<int, double> x(3, 4);
-  //   assert(x.Size<0>() == 3);
-  //   assert(x.Size<1>() == 4);
-  //
-  // Requires: `N < NumSizes`.
-  template <size_t N>
-  constexpr size_t Size() const {
-    static_assert(N < NumSizes, "Index out of bounds");
-    return size_[N];
-  }
-
-  // The number of elements in the array with the specified element type.
-  // There must be exactly one such array and its zero-based index must be
-  // at most `NumSizes`.
-  //
-  //   // int[3], 4 bytes of padding, double[4].
-  //   Layout<int, double> x(3, 4);
-  //   assert(x.Size<int>() == 3);
-  //   assert(x.Size<double>() == 4);
-  template <class T>
-  constexpr size_t Size() const {
-    return Size<ElementIndex<T>()>();
-  }
-
-  // The number of elements of all arrays for which they are known.
-  constexpr std::array<size_t, NumSizes> Sizes() const {
-    return {{Size<SizeSeq>()...}};
-  }
-
-  // Pointer to the beginning of the Nth array.
-  //
-  // `Char` must be `[const] [signed|unsigned] char`.
-  //
-  //   // int[3], 4 bytes of padding, double[4].
-  //   Layout<int, double> x(3, 4);
-  //   unsigned char* p = new unsigned char[x.AllocSize()];
-  //   int* ints = x.Pointer<0>(p);
-  //   double* doubles = x.Pointer<1>(p);
-  //
-  // Requires: `N <= NumSizes && N < sizeof...(Ts)`.
-  // Requires: `p` is aligned to `Alignment()`.
-  template <size_t N, class Char>
-  CopyConst<Char, ElementType<N>>* Pointer(Char* p) const {
-    using C = typename std::remove_const<Char>::type;
-    static_assert(
-        std::is_same<C, char>() || std::is_same<C, unsigned char>() ||
-            std::is_same<C, signed char>(),
-        "The argument must be a pointer to [const] [signed|unsigned] char");
-    constexpr size_t alignment = Alignment();
-    (void)alignment;
-    assert(reinterpret_cast<uintptr_t>(p) % alignment == 0);
-    return reinterpret_cast<CopyConst<Char, ElementType<N>>*>(p + Offset<N>());
-  }
-
-  // Pointer to the beginning of the array with the specified element type.
-  // There must be exactly one such array and its zero-based index must be at
-  // most `NumSizes`.
-  //
-  // `Char` must be `[const] [signed|unsigned] char`.
-  //
-  //   // int[3], 4 bytes of padding, double[4].
-  //   Layout<int, double> x(3, 4);
-  //   unsigned char* p = new unsigned char[x.AllocSize()];
-  //   int* ints = x.Pointer<int>(p);
-  //   double* doubles = x.Pointer<double>(p);
-  //
-  // Requires: `p` is aligned to `Alignment()`.
-  template <class T, class Char>
-  CopyConst<Char, T>* Pointer(Char* p) const {
-    return Pointer<ElementIndex<T>()>(p);
-  }
-
-  // Pointers to all arrays for which pointers are known.
-  //
-  // `Char` must be `[const] [signed|unsigned] char`.
-  //
-  //   // int[3], 4 bytes of padding, double[4].
-  //   Layout<int, double> x(3, 4);
-  //   unsigned char* p = new unsigned char[x.AllocSize()];
-  //
-  //   int* ints;
-  //   double* doubles;
-  //   std::tie(ints, doubles) = x.Pointers(p);
-  //
-  // Requires: `p` is aligned to `Alignment()`.
-  //
-  // Note: We're not using ElementType alias here because it does not compile
-  // under MSVC.
-  template <class Char>
-  std::tuple<CopyConst<
-      Char, typename std::tuple_element<OffsetSeq, ElementTypes>::type>*...>
-  Pointers(Char* p) const {
-    return std::tuple<CopyConst<Char, ElementType<OffsetSeq>>*...>(
-        Pointer<OffsetSeq>(p)...);
-  }
-
-  // The Nth array.
-  //
-  // `Char` must be `[const] [signed|unsigned] char`.
-  //
-  //   // int[3], 4 bytes of padding, double[4].
-  //   Layout<int, double> x(3, 4);
-  //   unsigned char* p = new unsigned char[x.AllocSize()];
-  //   Span<int> ints = x.Slice<0>(p);
-  //   Span<double> doubles = x.Slice<1>(p);
-  //
-  // Requires: `N < NumSizes`.
-  // Requires: `p` is aligned to `Alignment()`.
-  template <size_t N, class Char>
-  SliceType<CopyConst<Char, ElementType<N>>> Slice(Char* p) const {
-    return SliceType<CopyConst<Char, ElementType<N>>>(Pointer<N>(p), Size<N>());
-  }
-
-  // The array with the specified element type. There must be exactly one
-  // such array and its zero-based index must be less than `NumSizes`.
-  //
-  // `Char` must be `[const] [signed|unsigned] char`.
-  //
-  //   // int[3], 4 bytes of padding, double[4].
-  //   Layout<int, double> x(3, 4);
-  //   unsigned char* p = new unsigned char[x.AllocSize()];
-  //   Span<int> ints = x.Slice<int>(p);
-  //   Span<double> doubles = x.Slice<double>(p);
-  //
-  // Requires: `p` is aligned to `Alignment()`.
-  template <class T, class Char>
-  SliceType<CopyConst<Char, T>> Slice(Char* p) const {
-    return Slice<ElementIndex<T>()>(p);
-  }
-
-  // All arrays with known sizes.
-  //
-  // `Char` must be `[const] [signed|unsigned] char`.
-  //
-  //   // int[3], 4 bytes of padding, double[4].
-  //   Layout<int, double> x(3, 4);
-  //   unsigned char* p = new unsigned char[x.AllocSize()];
-  //
-  //   Span<int> ints;
-  //   Span<double> doubles;
-  //   std::tie(ints, doubles) = x.Slices(p);
-  //
-  // Requires: `p` is aligned to `Alignment()`.
-  //
-  // Note: We're not using ElementType alias here because it does not compile
-  // under MSVC.
-  template <class Char>
-  std::tuple<SliceType<CopyConst<
-      Char, typename std::tuple_element<SizeSeq, ElementTypes>::type>>...>
-  Slices(Char* p) const {
-    // Workaround for https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63875 (fixed
-    // in 6.1).
-    (void)p;
-    return std::tuple<SliceType<CopyConst<Char, ElementType<SizeSeq>>>...>(
-        Slice<SizeSeq>(p)...);
-  }
-
-  // The size of the allocation that fits all arrays.
-  //
-  //   // int[3], 4 bytes of padding, double[4].
-  //   Layout<int, double> x(3, 4);
-  //   unsigned char* p = new unsigned char[x.AllocSize()];  // 48 bytes
-  //
-  // Requires: `NumSizes == sizeof...(Ts)`.
-  constexpr size_t AllocSize() const {
-    static_assert(NumTypes == NumSizes, "You must specify sizes of all fields");
-    return Offset<NumTypes - 1>() +
-        SizeOf<ElementType<NumTypes - 1>>::value * size_[NumTypes - 1];
-  }
-
-  // If built with --config=asan, poisons padding bytes (if any) in the
-  // allocation. The pointer must point to a memory block at least
-  // `AllocSize()` bytes in length.
-  //
-  // `Char` must be `[const] [signed|unsigned] char`.
-  //
-  // Requires: `p` is aligned to `Alignment()`.
-  template <class Char, size_t N = NumOffsets - 1, EnableIf<N == 0> = 0>
-  void PoisonPadding(const Char* p) const {
-    Pointer<0>(p);  // verify the requirements on `Char` and `p`
-  }
-
-  template <class Char, size_t N = NumOffsets - 1, EnableIf<N != 0> = 0>
-  void PoisonPadding(const Char* p) const {
-    static_assert(N < NumOffsets, "Index out of bounds");
-    (void)p;
-#ifdef ABSL_HAVE_ADDRESS_SANITIZER
-    PoisonPadding<Char, N - 1>(p);
-    // The `if` is an optimization. It doesn't affect the observable behaviour.
-    if (ElementAlignment<N - 1>::value % ElementAlignment<N>::value) {
-      size_t start =
-          Offset<N - 1>() + SizeOf<ElementType<N - 1>>::value * size_[N - 1];
-      ASAN_POISON_MEMORY_REGION(p + start, Offset<N>() - start);
-    }
-#endif
-  }
-
-  // Human-readable description of the memory layout. Useful for debugging.
-  // Slow.
-  //
-  //   // char[5], 3 bytes of padding, int[3], 4 bytes of padding, followed
-  //   // by an unknown number of doubles.
-  //   auto x = Layout<char, int, double>::Partial(5, 3);
-  //   assert(x.DebugString() ==
-  //          "@0<char>(1)[5]; @8<int>(4)[3]; @24<double>(8)");
-  //
-  // Each field is in the following format: @offset<type>(sizeof)[size] (<type>
-  // may be missing depending on the target platform). For example,
-  // @8<int>(4)[3] means that at offset 8 we have an array of ints, where each
-  // int is 4 bytes, and we have 3 of those ints. The size of the last field may
-  // be missing (as in the example above). Only fields with known offsets are
-  // described. Type names may differ across platforms: one compiler might
-  // produce "unsigned*" where another produces "unsigned int *".
-  std::string DebugString() const {
-    const auto offsets = Offsets();
-    const size_t sizes[] = {SizeOf<ElementType<OffsetSeq>>::value...};
-    const std::string types[] = {
-        adl_barrier::TypeName<ElementType<OffsetSeq>>()...};
-    std::string res = absl::StrCat("@0", types[0], "(", sizes[0], ")");
-    for (size_t i = 0; i != NumOffsets - 1; ++i) {
-      absl::StrAppend(&res, "[", size_[i], "]; @", offsets[i + 1], types[i + 1],
-                      "(", sizes[i + 1], ")");
-    }
-    // NumSizes is a constant that may be zero. Some compilers cannot see that
-    // inside the if statement "size_[NumSizes - 1]" must be valid.
-    int last = static_cast<int>(NumSizes) - 1;
-    if (NumTypes == NumSizes && last >= 0) {
-      absl::StrAppend(&res, "[", size_[last], "]");
-    }
-    return res;
-  }
-
- private:
-  // Arguments of `Layout::Partial()` or `Layout::Layout()`.
-  size_t size_[NumSizes > 0 ? NumSizes : 1];
-};
-
-template <size_t NumSizes, class... Ts>
-using LayoutType = LayoutImpl<
-    std::tuple<Ts...>, absl::make_index_sequence<NumSizes>,
-    absl::make_index_sequence<adl_barrier::Min(sizeof...(Ts), NumSizes + 1)>>;
-
-}  // namespace internal_layout
-
-// Descriptor of arrays of various types and sizes laid out in memory one after
-// another. See the top of the file for documentation.
-//
-// Check out the public API of internal_layout::LayoutImpl above. The type is
-// internal to the library but its methods are public, and they are inherited
-// by `Layout`.
-template <class... Ts>
-class Layout : public internal_layout::LayoutType<sizeof...(Ts), Ts...> {
- public:
-  static_assert(sizeof...(Ts) > 0, "At least one field is required");
-  static_assert(
-      absl::conjunction<internal_layout::IsLegalElementType<Ts>...>::value,
-      "Invalid element type (see IsLegalElementType)");
-
-  // The result type of `Partial()` with `NumSizes` arguments.
-  template <size_t NumSizes>
-  using PartialType = internal_layout::LayoutType<NumSizes, Ts...>;
-
-  // `Layout` knows the element types of the arrays we want to lay out in
-  // memory but not the number of elements in each array.
-  // `Partial(size1, ..., sizeN)` allows us to specify the latter. The
-  // resulting immutable object can be used to obtain pointers to the
-  // individual arrays.
-  //
-  // It's allowed to pass fewer array sizes than the number of arrays. E.g.,
-  // if all you need is to the offset of the second array, you only need to
-  // pass one argument -- the number of elements in the first array.
-  //
-  //   // int[3] followed by 4 bytes of padding and an unknown number of
-  //   // doubles.
-  //   auto x = Layout<int, double>::Partial(3);
-  //   // doubles start at byte 16.
-  //   assert(x.Offset<1>() == 16);
-  //
-  // If you know the number of elements in all arrays, you can still call
-  // `Partial()` but it's more convenient to use the constructor of `Layout`.
-  //
-  //   Layout<int, double> x(3, 5);
-  //
-  // Note: The sizes of the arrays must be specified in number of elements,
-  // not in bytes.
-  //
-  // Requires: `sizeof...(Sizes) <= sizeof...(Ts)`.
-  // Requires: all arguments are convertible to `size_t`.
-  template <class... Sizes>
-  static constexpr PartialType<sizeof...(Sizes)> Partial(Sizes&&... sizes) {
-    static_assert(sizeof...(Sizes) <= sizeof...(Ts), "");
-    return PartialType<sizeof...(Sizes)>(std::forward<Sizes>(sizes)...);
-  }
-
-  // Creates a layout with the sizes of all arrays specified. If you know
-  // only the sizes of the first N arrays (where N can be zero), you can use
-  // `Partial()` defined above. The constructor is essentially equivalent to
-  // calling `Partial()` and passing in all array sizes; the constructor is
-  // provided as a convenient abbreviation.
-  //
-  // Note: The sizes of the arrays must be specified in number of elements,
-  // not in bytes.
-  constexpr explicit Layout(internal_layout::TypeToSize<Ts>... sizes)
-      : internal_layout::LayoutType<sizeof...(Ts), Ts...>(sizes...) {}
-};
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_LAYOUT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/node_slot_policy.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/node_slot_policy.h
deleted file mode 100644
index 3f1874d4..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/node_slot_policy.h
+++ /dev/null
@@ -1,95 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Adapts a policy for nodes.
-//
-// The node policy should model:
-//
-// struct Policy {
-//   // Returns a new node allocated and constructed using the allocator, using
-//   // the specified arguments.
-//   template <class Alloc, class... Args>
-//   value_type* new_element(Alloc* alloc, Args&&... args) const;
-//
-//   // Destroys and deallocates node using the allocator.
-//   template <class Alloc>
-//   void delete_element(Alloc* alloc, value_type* node) const;
-// };
-//
-// It may also optionally define `value()` and `apply()`. For documentation on
-// these, see hash_policy_traits.h.
-
-#ifndef ABSL_CONTAINER_INTERNAL_NODE_SLOT_POLICY_H_
-#define ABSL_CONTAINER_INTERNAL_NODE_SLOT_POLICY_H_
-
-#include <cassert>
-#include <cstddef>
-#include <memory>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-template <class Reference, class Policy>
-struct node_slot_policy {
-  static_assert(std::is_lvalue_reference<Reference>::value, "");
-
-  using slot_type = typename std::remove_cv<
-      typename std::remove_reference<Reference>::type>::type*;
-
-  template <class Alloc, class... Args>
-  static void construct(Alloc* alloc, slot_type* slot, Args&&... args) {
-    *slot = Policy::new_element(alloc, std::forward<Args>(args)...);
-  }
-
-  template <class Alloc>
-  static void destroy(Alloc* alloc, slot_type* slot) {
-    Policy::delete_element(alloc, *slot);
-  }
-
-  // Returns true_type to indicate that transfer can use memcpy.
-  template <class Alloc>
-  static std::true_type transfer(Alloc*, slot_type* new_slot,
-                                 slot_type* old_slot) {
-    *new_slot = *old_slot;
-    return {};
-  }
-
-  static size_t space_used(const slot_type* slot) {
-    if (slot == nullptr) return Policy::element_space_used(nullptr);
-    return Policy::element_space_used(*slot);
-  }
-
-  static Reference element(slot_type* slot) { return **slot; }
-
-  template <class T, class P = Policy>
-  static auto value(T* elem) -> decltype(P::value(elem)) {
-    return P::value(elem);
-  }
-
-  template <class... Ts, class P = Policy>
-  static auto apply(Ts&&... ts) -> decltype(P::apply(std::forward<Ts>(ts)...)) {
-    return P::apply(std::forward<Ts>(ts)...);
-  }
-};
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_NODE_SLOT_POLICY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/raw_hash_map.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/raw_hash_map.h
deleted file mode 100644
index 464bf23b..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/raw_hash_map.h
+++ /dev/null
@@ -1,226 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_RAW_HASH_MAP_H_
-#define ABSL_CONTAINER_INTERNAL_RAW_HASH_MAP_H_
-
-#include <tuple>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/internal/throw_delegate.h"
-#include "absl/container/internal/container_memory.h"
-#include "absl/container/internal/raw_hash_set.h"  // IWYU pragma: export
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-template <class Policy, class Hash, class Eq, class Alloc>
-class raw_hash_map : public raw_hash_set<Policy, Hash, Eq, Alloc> {
-  // P is Policy. It's passed as a template argument to support maps that have
-  // incomplete types as values, as in unordered_map<K, IncompleteType>.
-  // MappedReference<> may be a non-reference type.
-  template <class P>
-  using MappedReference = decltype(P::value(
-      std::addressof(std::declval<typename raw_hash_map::reference>())));
-
-  // MappedConstReference<> may be a non-reference type.
-  template <class P>
-  using MappedConstReference = decltype(P::value(
-      std::addressof(std::declval<typename raw_hash_map::const_reference>())));
-
-  using KeyArgImpl =
-      KeyArg<IsTransparent<Eq>::value && IsTransparent<Hash>::value>;
-
- public:
-  using key_type = typename Policy::key_type;
-  using mapped_type = typename Policy::mapped_type;
-  template <class K>
-  using key_arg = typename KeyArgImpl::template type<K, key_type>;
-
-  static_assert(!std::is_reference<key_type>::value, "");
-
-  // TODO(b/187807849): Evaluate whether to support reference mapped_type and
-  // remove this assertion if/when it is supported.
-  static_assert(!std::is_reference<mapped_type>::value, "");
-
-  using iterator = typename raw_hash_map::raw_hash_set::iterator;
-  using const_iterator = typename raw_hash_map::raw_hash_set::const_iterator;
-
-  raw_hash_map() {}
-  using raw_hash_map::raw_hash_set::raw_hash_set;
-
-  // The last two template parameters ensure that both arguments are rvalues
-  // (lvalue arguments are handled by the overloads below). This is necessary
-  // for supporting bitfield arguments.
-  //
-  //   union { int n : 1; };
-  //   flat_hash_map<int, int> m;
-  //   m.insert_or_assign(n, n);
-  template <class K = key_type, class V = mapped_type, K* = nullptr,
-            V* = nullptr>
-  std::pair<iterator, bool> insert_or_assign(key_arg<K>&& k, V&& v)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign_impl(std::forward<K>(k), std::forward<V>(v));
-  }
-
-  template <class K = key_type, class V = mapped_type, K* = nullptr>
-  std::pair<iterator, bool> insert_or_assign(key_arg<K>&& k, const V& v)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign_impl(std::forward<K>(k), v);
-  }
-
-  template <class K = key_type, class V = mapped_type, V* = nullptr>
-  std::pair<iterator, bool> insert_or_assign(const key_arg<K>& k, V&& v)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign_impl(k, std::forward<V>(v));
-  }
-
-  template <class K = key_type, class V = mapped_type>
-  std::pair<iterator, bool> insert_or_assign(const key_arg<K>& k, const V& v)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign_impl(k, v);
-  }
-
-  template <class K = key_type, class V = mapped_type, K* = nullptr,
-            V* = nullptr>
-  iterator insert_or_assign(const_iterator, key_arg<K>&& k,
-                            V&& v) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign(std::forward<K>(k), std::forward<V>(v)).first;
-  }
-
-  template <class K = key_type, class V = mapped_type, K* = nullptr>
-  iterator insert_or_assign(const_iterator, key_arg<K>&& k,
-                            const V& v) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign(std::forward<K>(k), v).first;
-  }
-
-  template <class K = key_type, class V = mapped_type, V* = nullptr>
-  iterator insert_or_assign(const_iterator, const key_arg<K>& k,
-                            V&& v) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign(k, std::forward<V>(v)).first;
-  }
-
-  template <class K = key_type, class V = mapped_type>
-  iterator insert_or_assign(const_iterator, const key_arg<K>& k,
-                            const V& v) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert_or_assign(k, v).first;
-  }
-
-  // All `try_emplace()` overloads make the same guarantees regarding rvalue
-  // arguments as `std::unordered_map::try_emplace()`, namely that these
-  // functions will not move from rvalue arguments if insertions do not happen.
-  template <class K = key_type, class... Args,
-            typename std::enable_if<
-                !std::is_convertible<K, const_iterator>::value, int>::type = 0,
-            K* = nullptr>
-  std::pair<iterator, bool> try_emplace(key_arg<K>&& k, Args&&... args)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return try_emplace_impl(std::forward<K>(k), std::forward<Args>(args)...);
-  }
-
-  template <class K = key_type, class... Args,
-            typename std::enable_if<
-                !std::is_convertible<K, const_iterator>::value, int>::type = 0>
-  std::pair<iterator, bool> try_emplace(const key_arg<K>& k, Args&&... args)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return try_emplace_impl(k, std::forward<Args>(args)...);
-  }
-
-  template <class K = key_type, class... Args, K* = nullptr>
-  iterator try_emplace(const_iterator, key_arg<K>&& k,
-                       Args&&... args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return try_emplace(std::forward<K>(k), std::forward<Args>(args)...).first;
-  }
-
-  template <class K = key_type, class... Args>
-  iterator try_emplace(const_iterator, const key_arg<K>& k,
-                       Args&&... args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return try_emplace(k, std::forward<Args>(args)...).first;
-  }
-
-  template <class K = key_type, class P = Policy>
-  MappedReference<P> at(const key_arg<K>& key) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    auto it = this->find(key);
-    if (it == this->end()) {
-      base_internal::ThrowStdOutOfRange(
-          "absl::container_internal::raw_hash_map<>::at");
-    }
-    return Policy::value(&*it);
-  }
-
-  template <class K = key_type, class P = Policy>
-  MappedConstReference<P> at(const key_arg<K>& key) const
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    auto it = this->find(key);
-    if (it == this->end()) {
-      base_internal::ThrowStdOutOfRange(
-          "absl::container_internal::raw_hash_map<>::at");
-    }
-    return Policy::value(&*it);
-  }
-
-  template <class K = key_type, class P = Policy, K* = nullptr>
-  MappedReference<P> operator[](key_arg<K>&& key)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    // It is safe to use unchecked_deref here because try_emplace
-    // will always return an iterator pointing to a valid item in the table,
-    // since it inserts if nothing is found for the given key.
-    return Policy::value(
-        &this->unchecked_deref(try_emplace(std::forward<K>(key)).first));
-  }
-
-  template <class K = key_type, class P = Policy>
-  MappedReference<P> operator[](const key_arg<K>& key)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    // It is safe to use unchecked_deref here because try_emplace
-    // will always return an iterator pointing to a valid item in the table,
-    // since it inserts if nothing is found for the given key.
-    return Policy::value(&this->unchecked_deref(try_emplace(key).first));
-  }
-
- private:
-  template <class K, class V>
-  std::pair<iterator, bool> insert_or_assign_impl(K&& k, V&& v)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    auto res = this->find_or_prepare_insert(k);
-    if (res.second) {
-      this->emplace_at(res.first, std::forward<K>(k), std::forward<V>(v));
-    } else {
-      Policy::value(&*res.first) = std::forward<V>(v);
-    }
-    return res;
-  }
-
-  template <class K = key_type, class... Args>
-  std::pair<iterator, bool> try_emplace_impl(K&& k, Args&&... args)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    auto res = this->find_or_prepare_insert(k);
-    if (res.second) {
-      this->emplace_at(res.first, std::piecewise_construct,
-                       std::forward_as_tuple(std::forward<K>(k)),
-                       std::forward_as_tuple(std::forward<Args>(args)...));
-    }
-    return res;
-  }
-};
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_RAW_HASH_MAP_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/raw_hash_set.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/raw_hash_set.h
deleted file mode 100644
index 258458b0..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/raw_hash_set.h
+++ /dev/null
@@ -1,3904 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// An open-addressing
-// hashtable with quadratic probing.
-//
-// This is a low level hashtable on top of which different interfaces can be
-// implemented, like flat_hash_set, node_hash_set, string_hash_set, etc.
-//
-// The table interface is similar to that of std::unordered_set. Notable
-// differences are that most member functions support heterogeneous keys when
-// BOTH the hash and eq functions are marked as transparent. They do so by
-// providing a typedef called `is_transparent`.
-//
-// When heterogeneous lookup is enabled, functions that take key_type act as if
-// they have an overload set like:
-//
-//   iterator find(const key_type& key);
-//   template <class K>
-//   iterator find(const K& key);
-//
-//   size_type erase(const key_type& key);
-//   template <class K>
-//   size_type erase(const K& key);
-//
-//   std::pair<iterator, iterator> equal_range(const key_type& key);
-//   template <class K>
-//   std::pair<iterator, iterator> equal_range(const K& key);
-//
-// When heterogeneous lookup is disabled, only the explicit `key_type` overloads
-// exist.
-//
-// find() also supports passing the hash explicitly:
-//
-//   iterator find(const key_type& key, size_t hash);
-//   template <class U>
-//   iterator find(const U& key, size_t hash);
-//
-// In addition the pointer to element and iterator stability guarantees are
-// weaker: all iterators and pointers are invalidated after a new element is
-// inserted.
-//
-// IMPLEMENTATION DETAILS
-//
-// # Table Layout
-//
-// A raw_hash_set's backing array consists of control bytes followed by slots
-// that may or may not contain objects.
-//
-// The layout of the backing array, for `capacity` slots, is thus, as a
-// pseudo-struct:
-//
-//   struct BackingArray {
-//     // Sampling handler. This field isn't present when the sampling is
-//     // disabled or this allocation hasn't been selected for sampling.
-//     HashtablezInfoHandle infoz_;
-//     // The number of elements we can insert before growing the capacity.
-//     size_t growth_left;
-//     // Control bytes for the "real" slots.
-//     ctrl_t ctrl[capacity];
-//     // Always `ctrl_t::kSentinel`. This is used by iterators to find when to
-//     // stop and serves no other purpose.
-//     ctrl_t sentinel;
-//     // A copy of the first `kWidth - 1` elements of `ctrl`. This is used so
-//     // that if a probe sequence picks a value near the end of `ctrl`,
-//     // `Group` will have valid control bytes to look at.
-//     ctrl_t clones[kWidth - 1];
-//     // The actual slot data.
-//     slot_type slots[capacity];
-//   };
-//
-// The length of this array is computed by `RawHashSetLayout::alloc_size` below.
-//
-// Control bytes (`ctrl_t`) are bytes (collected into groups of a
-// platform-specific size) that define the state of the corresponding slot in
-// the slot array. Group manipulation is tightly optimized to be as efficient
-// as possible: SSE and friends on x86, clever bit operations on other arches.
-//
-//      Group 1         Group 2        Group 3
-// +---------------+---------------+---------------+
-// | | | | | | | | | | | | | | | | | | | | | | | | |
-// +---------------+---------------+---------------+
-//
-// Each control byte is either a special value for empty slots, deleted slots
-// (sometimes called *tombstones*), and a special end-of-table marker used by
-// iterators, or, if occupied, seven bits (H2) from the hash of the value in the
-// corresponding slot.
-//
-// Storing control bytes in a separate array also has beneficial cache effects,
-// since more logical slots will fit into a cache line.
-//
-// # Small Object Optimization (SOO)
-//
-// When the size/alignment of the value_type and the capacity of the table are
-// small, we enable small object optimization and store the values inline in
-// the raw_hash_set object. This optimization allows us to avoid
-// allocation/deallocation as well as cache/dTLB misses.
-//
-// # Hashing
-//
-// We compute two separate hashes, `H1` and `H2`, from the hash of an object.
-// `H1(hash(x))` is an index into `slots`, and essentially the starting point
-// for the probe sequence. `H2(hash(x))` is a 7-bit value used to filter out
-// objects that cannot possibly be the one we are looking for.
-//
-// # Table operations.
-//
-// The key operations are `insert`, `find`, and `erase`.
-//
-// Since `insert` and `erase` are implemented in terms of `find`, we describe
-// `find` first. To `find` a value `x`, we compute `hash(x)`. From
-// `H1(hash(x))` and the capacity, we construct a `probe_seq` that visits every
-// group of slots in some interesting order.
-//
-// We now walk through these indices. At each index, we select the entire group
-// starting with that index and extract potential candidates: occupied slots
-// with a control byte equal to `H2(hash(x))`. If we find an empty slot in the
-// group, we stop and return an error. Each candidate slot `y` is compared with
-// `x`; if `x == y`, we are done and return `&y`; otherwise we continue to the
-// next probe index. Tombstones effectively behave like full slots that never
-// match the value we're looking for.
-//
-// The `H2` bits ensure when we compare a slot to an object with `==`, we are
-// likely to have actually found the object.  That is, the chance is low that
-// `==` is called and returns `false`.  Thus, when we search for an object, we
-// are unlikely to call `==` many times.  This likelyhood can be analyzed as
-// follows (assuming that H2 is a random enough hash function).
-//
-// Let's assume that there are `k` "wrong" objects that must be examined in a
-// probe sequence.  For example, when doing a `find` on an object that is in the
-// table, `k` is the number of objects between the start of the probe sequence
-// and the final found object (not including the final found object).  The
-// expected number of objects with an H2 match is then `k/128`.  Measurements
-// and analysis indicate that even at high load factors, `k` is less than 32,
-// meaning that the number of "false positive" comparisons we must perform is
-// less than 1/8 per `find`.
-
-// `insert` is implemented in terms of `unchecked_insert`, which inserts a
-// value presumed to not be in the table (violating this requirement will cause
-// the table to behave erratically). Given `x` and its hash `hash(x)`, to insert
-// it, we construct a `probe_seq` once again, and use it to find the first
-// group with an unoccupied (empty *or* deleted) slot. We place `x` into the
-// first such slot in the group and mark it as full with `x`'s H2.
-//
-// To `insert`, we compose `unchecked_insert` with `find`. We compute `h(x)` and
-// perform a `find` to see if it's already present; if it is, we're done. If
-// it's not, we may decide the table is getting overcrowded (i.e. the load
-// factor is greater than 7/8 for big tables; `is_small()` tables use a max load
-// factor of 1); in this case, we allocate a bigger array, `unchecked_insert`
-// each element of the table into the new array (we know that no insertion here
-// will insert an already-present value), and discard the old backing array. At
-// this point, we may `unchecked_insert` the value `x`.
-//
-// Below, `unchecked_insert` is partly implemented by `prepare_insert`, which
-// presents a viable, initialized slot pointee to the caller.
-//
-// `erase` is implemented in terms of `erase_at`, which takes an index to a
-// slot. Given an offset, we simply create a tombstone and destroy its contents.
-// If we can prove that the slot would not appear in a probe sequence, we can
-// make the slot as empty, instead. We can prove this by observing that if a
-// group has any empty slots, it has never been full (assuming we never create
-// an empty slot in a group with no empties, which this heuristic guarantees we
-// never do) and find would stop at this group anyways (since it does not probe
-// beyond groups with empties).
-//
-// `erase` is `erase_at` composed with `find`: if we
-// have a value `x`, we can perform a `find`, and then `erase_at` the resulting
-// slot.
-//
-// To iterate, we simply traverse the array, skipping empty and deleted slots
-// and stopping when we hit a `kSentinel`.
-
-#ifndef ABSL_CONTAINER_INTERNAL_RAW_HASH_SET_H_
-#define ABSL_CONTAINER_INTERNAL_RAW_HASH_SET_H_
-
-#include <algorithm>
-#include <cassert>
-#include <cmath>
-#include <cstddef>
-#include <cstdint>
-#include <cstring>
-#include <initializer_list>
-#include <iterator>
-#include <limits>
-#include <memory>
-#include <tuple>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/internal/endian.h"
-#include "absl/base/internal/raw_logging.h"
-#include "absl/base/macros.h"
-#include "absl/base/optimization.h"
-#include "absl/base/options.h"
-#include "absl/base/port.h"
-#include "absl/base/prefetch.h"
-#include "absl/container/internal/common.h"  // IWYU pragma: export // for node_handle
-#include "absl/container/internal/compressed_tuple.h"
-#include "absl/container/internal/container_memory.h"
-#include "absl/container/internal/hash_policy_traits.h"
-#include "absl/container/internal/hashtable_debug_hooks.h"
-#include "absl/container/internal/hashtablez_sampler.h"
-#include "absl/memory/memory.h"
-#include "absl/meta/type_traits.h"
-#include "absl/numeric/bits.h"
-#include "absl/utility/utility.h"
-
-#ifdef ABSL_INTERNAL_HAVE_SSE2
-#include <emmintrin.h>
-#endif
-
-#ifdef ABSL_INTERNAL_HAVE_SSSE3
-#include <tmmintrin.h>
-#endif
-
-#ifdef _MSC_VER
-#include <intrin.h>
-#endif
-
-#ifdef ABSL_INTERNAL_HAVE_ARM_NEON
-#include <arm_neon.h>
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-#ifdef ABSL_SWISSTABLE_ENABLE_GENERATIONS
-#error ABSL_SWISSTABLE_ENABLE_GENERATIONS cannot be directly set
-#elif defined(ABSL_HAVE_ADDRESS_SANITIZER) || \
-    defined(ABSL_HAVE_HWADDRESS_SANITIZER) || \
-    defined(ABSL_HAVE_MEMORY_SANITIZER)
-// When compiled in sanitizer mode, we add generation integers to the backing
-// array and iterators. In the backing array, we store the generation between
-// the control bytes and the slots. When iterators are dereferenced, we assert
-// that the container has not been mutated in a way that could cause iterator
-// invalidation since the iterator was initialized.
-#define ABSL_SWISSTABLE_ENABLE_GENERATIONS
-#endif
-
-// We use uint8_t so we don't need to worry about padding.
-using GenerationType = uint8_t;
-
-// A sentinel value for empty generations. Using 0 makes it easy to constexpr
-// initialize an array of this value.
-constexpr GenerationType SentinelEmptyGeneration() { return 0; }
-
-constexpr GenerationType NextGeneration(GenerationType generation) {
-  return ++generation == SentinelEmptyGeneration() ? ++generation : generation;
-}
-
-#ifdef ABSL_SWISSTABLE_ENABLE_GENERATIONS
-constexpr bool SwisstableGenerationsEnabled() { return true; }
-constexpr size_t NumGenerationBytes() { return sizeof(GenerationType); }
-#else
-constexpr bool SwisstableGenerationsEnabled() { return false; }
-constexpr size_t NumGenerationBytes() { return 0; }
-#endif
-
-template <typename AllocType>
-void SwapAlloc(AllocType& lhs, AllocType& rhs,
-               std::true_type /* propagate_on_container_swap */) {
-  using std::swap;
-  swap(lhs, rhs);
-}
-template <typename AllocType>
-void SwapAlloc(AllocType& lhs, AllocType& rhs,
-               std::false_type /* propagate_on_container_swap */) {
-  (void)lhs;
-  (void)rhs;
-  assert(lhs == rhs &&
-         "It's UB to call swap with unequal non-propagating allocators.");
-}
-
-template <typename AllocType>
-void CopyAlloc(AllocType& lhs, AllocType& rhs,
-               std::true_type /* propagate_alloc */) {
-  lhs = rhs;
-}
-template <typename AllocType>
-void CopyAlloc(AllocType&, AllocType&, std::false_type /* propagate_alloc */) {}
-
-// The state for a probe sequence.
-//
-// Currently, the sequence is a triangular progression of the form
-//
-//   p(i) := Width * (i^2 + i)/2 + hash (mod mask + 1)
-//
-// The use of `Width` ensures that each probe step does not overlap groups;
-// the sequence effectively outputs the addresses of *groups* (although not
-// necessarily aligned to any boundary). The `Group` machinery allows us
-// to check an entire group with minimal branching.
-//
-// Wrapping around at `mask + 1` is important, but not for the obvious reason.
-// As described above, the first few entries of the control byte array
-// are mirrored at the end of the array, which `Group` will find and use
-// for selecting candidates. However, when those candidates' slots are
-// actually inspected, there are no corresponding slots for the cloned bytes,
-// so we need to make sure we've treated those offsets as "wrapping around".
-//
-// It turns out that this probe sequence visits every group exactly once if the
-// number of groups is a power of two, since (i^2+i)/2 is a bijection in
-// Z/(2^m). See https://en.wikipedia.org/wiki/Quadratic_probing
-template <size_t Width>
-class probe_seq {
- public:
-  // Creates a new probe sequence using `hash` as the initial value of the
-  // sequence and `mask` (usually the capacity of the table) as the mask to
-  // apply to each value in the progression.
-  probe_seq(size_t hash, size_t mask) {
-    assert(((mask + 1) & mask) == 0 && "not a mask");
-    mask_ = mask;
-    offset_ = hash & mask_;
-  }
-
-  // The offset within the table, i.e., the value `p(i)` above.
-  size_t offset() const { return offset_; }
-  size_t offset(size_t i) const { return (offset_ + i) & mask_; }
-
-  void next() {
-    index_ += Width;
-    offset_ += index_;
-    offset_ &= mask_;
-  }
-  // 0-based probe index, a multiple of `Width`.
-  size_t index() const { return index_; }
-
- private:
-  size_t mask_;
-  size_t offset_;
-  size_t index_ = 0;
-};
-
-template <class ContainerKey, class Hash, class Eq>
-struct RequireUsableKey {
-  template <class PassedKey, class... Args>
-  std::pair<
-      decltype(std::declval<const Hash&>()(std::declval<const PassedKey&>())),
-      decltype(std::declval<const Eq&>()(std::declval<const ContainerKey&>(),
-                                         std::declval<const PassedKey&>()))>*
-  operator()(const PassedKey&, const Args&...) const;
-};
-
-template <class E, class Policy, class Hash, class Eq, class... Ts>
-struct IsDecomposable : std::false_type {};
-
-template <class Policy, class Hash, class Eq, class... Ts>
-struct IsDecomposable<
-    absl::void_t<decltype(Policy::apply(
-        RequireUsableKey<typename Policy::key_type, Hash, Eq>(),
-        std::declval<Ts>()...))>,
-    Policy, Hash, Eq, Ts...> : std::true_type {};
-
-// TODO(alkis): Switch to std::is_nothrow_swappable when gcc/clang supports it.
-template <class T>
-constexpr bool IsNoThrowSwappable(std::true_type = {} /* is_swappable */) {
-  using std::swap;
-  return noexcept(swap(std::declval<T&>(), std::declval<T&>()));
-}
-template <class T>
-constexpr bool IsNoThrowSwappable(std::false_type /* is_swappable */) {
-  return false;
-}
-
-template <typename T>
-uint32_t TrailingZeros(T x) {
-  ABSL_ASSUME(x != 0);
-  return static_cast<uint32_t>(countr_zero(x));
-}
-
-// 8 bytes bitmask with most significant bit set for every byte.
-constexpr uint64_t kMsbs8Bytes = 0x8080808080808080ULL;
-
-// An abstract bitmask, such as that emitted by a SIMD instruction.
-//
-// Specifically, this type implements a simple bitset whose representation is
-// controlled by `SignificantBits` and `Shift`. `SignificantBits` is the number
-// of abstract bits in the bitset, while `Shift` is the log-base-two of the
-// width of an abstract bit in the representation.
-// This mask provides operations for any number of real bits set in an abstract
-// bit. To add iteration on top of that, implementation must guarantee no more
-// than the most significant real bit is set in a set abstract bit.
-template <class T, int SignificantBits, int Shift = 0>
-class NonIterableBitMask {
- public:
-  explicit NonIterableBitMask(T mask) : mask_(mask) {}
-
-  explicit operator bool() const { return this->mask_ != 0; }
-
-  // Returns the index of the lowest *abstract* bit set in `self`.
-  uint32_t LowestBitSet() const {
-    return container_internal::TrailingZeros(mask_) >> Shift;
-  }
-
-  // Returns the index of the highest *abstract* bit set in `self`.
-  uint32_t HighestBitSet() const {
-    return static_cast<uint32_t>((bit_width(mask_) - 1) >> Shift);
-  }
-
-  // Returns the number of trailing zero *abstract* bits.
-  uint32_t TrailingZeros() const {
-    return container_internal::TrailingZeros(mask_) >> Shift;
-  }
-
-  // Returns the number of leading zero *abstract* bits.
-  uint32_t LeadingZeros() const {
-    constexpr int total_significant_bits = SignificantBits << Shift;
-    constexpr int extra_bits = sizeof(T) * 8 - total_significant_bits;
-    return static_cast<uint32_t>(
-               countl_zero(static_cast<T>(mask_ << extra_bits))) >>
-           Shift;
-  }
-
-  T mask_;
-};
-
-// Mask that can be iterable
-//
-// For example, when `SignificantBits` is 16 and `Shift` is zero, this is just
-// an ordinary 16-bit bitset occupying the low 16 bits of `mask`. When
-// `SignificantBits` is 8 and `Shift` is 3, abstract bits are represented as
-// the bytes `0x00` and `0x80`, and it occupies all 64 bits of the bitmask.
-// If NullifyBitsOnIteration is true (only allowed for Shift == 3),
-// non zero abstract bit is allowed to have additional bits
-// (e.g., `0xff`, `0x83` and `0x9c` are ok, but `0x6f` is not).
-//
-// For example:
-//   for (int i : BitMask<uint32_t, 16>(0b101)) -> yields 0, 2
-//   for (int i : BitMask<uint64_t, 8, 3>(0x0000000080800000)) -> yields 2, 3
-template <class T, int SignificantBits, int Shift = 0,
-          bool NullifyBitsOnIteration = false>
-class BitMask : public NonIterableBitMask<T, SignificantBits, Shift> {
-  using Base = NonIterableBitMask<T, SignificantBits, Shift>;
-  static_assert(std::is_unsigned<T>::value, "");
-  static_assert(Shift == 0 || Shift == 3, "");
-  static_assert(!NullifyBitsOnIteration || Shift == 3, "");
-
- public:
-  explicit BitMask(T mask) : Base(mask) {
-    if (Shift == 3 && !NullifyBitsOnIteration) {
-      assert(this->mask_ == (this->mask_ & kMsbs8Bytes));
-    }
-  }
-  // BitMask is an iterator over the indices of its abstract bits.
-  using value_type = int;
-  using iterator = BitMask;
-  using const_iterator = BitMask;
-
-  BitMask& operator++() {
-    if (Shift == 3 && NullifyBitsOnIteration) {
-      this->mask_ &= kMsbs8Bytes;
-    }
-    this->mask_ &= (this->mask_ - 1);
-    return *this;
-  }
-
-  uint32_t operator*() const { return Base::LowestBitSet(); }
-
-  BitMask begin() const { return *this; }
-  BitMask end() const { return BitMask(0); }
-
- private:
-  friend bool operator==(const BitMask& a, const BitMask& b) {
-    return a.mask_ == b.mask_;
-  }
-  friend bool operator!=(const BitMask& a, const BitMask& b) {
-    return a.mask_ != b.mask_;
-  }
-};
-
-using h2_t = uint8_t;
-
-// The values here are selected for maximum performance. See the static asserts
-// below for details.
-
-// A `ctrl_t` is a single control byte, which can have one of four
-// states: empty, deleted, full (which has an associated seven-bit h2_t value)
-// and the sentinel. They have the following bit patterns:
-//
-//      empty: 1 0 0 0 0 0 0 0
-//    deleted: 1 1 1 1 1 1 1 0
-//       full: 0 h h h h h h h  // h represents the hash bits.
-//   sentinel: 1 1 1 1 1 1 1 1
-//
-// These values are specifically tuned for SSE-flavored SIMD.
-// The static_asserts below detail the source of these choices.
-//
-// We use an enum class so that when strict aliasing is enabled, the compiler
-// knows ctrl_t doesn't alias other types.
-enum class ctrl_t : int8_t {
-  kEmpty = -128,   // 0b10000000
-  kDeleted = -2,   // 0b11111110
-  kSentinel = -1,  // 0b11111111
-};
-static_assert(
-    (static_cast<int8_t>(ctrl_t::kEmpty) &
-     static_cast<int8_t>(ctrl_t::kDeleted) &
-     static_cast<int8_t>(ctrl_t::kSentinel) & 0x80) != 0,
-    "Special markers need to have the MSB to make checking for them efficient");
-static_assert(
-    ctrl_t::kEmpty < ctrl_t::kSentinel && ctrl_t::kDeleted < ctrl_t::kSentinel,
-    "ctrl_t::kEmpty and ctrl_t::kDeleted must be smaller than "
-    "ctrl_t::kSentinel to make the SIMD test of IsEmptyOrDeleted() efficient");
-static_assert(
-    ctrl_t::kSentinel == static_cast<ctrl_t>(-1),
-    "ctrl_t::kSentinel must be -1 to elide loading it from memory into SIMD "
-    "registers (pcmpeqd xmm, xmm)");
-static_assert(ctrl_t::kEmpty == static_cast<ctrl_t>(-128),
-              "ctrl_t::kEmpty must be -128 to make the SIMD check for its "
-              "existence efficient (psignb xmm, xmm)");
-static_assert(
-    (~static_cast<int8_t>(ctrl_t::kEmpty) &
-     ~static_cast<int8_t>(ctrl_t::kDeleted) &
-     static_cast<int8_t>(ctrl_t::kSentinel) & 0x7F) != 0,
-    "ctrl_t::kEmpty and ctrl_t::kDeleted must share an unset bit that is not "
-    "shared by ctrl_t::kSentinel to make the scalar test for "
-    "MaskEmptyOrDeleted() efficient");
-static_assert(ctrl_t::kDeleted == static_cast<ctrl_t>(-2),
-              "ctrl_t::kDeleted must be -2 to make the implementation of "
-              "ConvertSpecialToEmptyAndFullToDeleted efficient");
-
-// See definition comment for why this is size 32.
-ABSL_DLL extern const ctrl_t kEmptyGroup[32];
-
-// Returns a pointer to a control byte group that can be used by empty tables.
-inline ctrl_t* EmptyGroup() {
-  // Const must be cast away here; no uses of this function will actually write
-  // to it because it is only used for empty tables.
-  return const_cast<ctrl_t*>(kEmptyGroup + 16);
-}
-
-// For use in SOO iterators.
-// TODO(b/289225379): we could potentially get rid of this by adding an is_soo
-// bit in iterators. This would add branches but reduce cache misses.
-ABSL_DLL extern const ctrl_t kSooControl[17];
-
-// Returns a pointer to a full byte followed by a sentinel byte.
-inline ctrl_t* SooControl() {
-  // Const must be cast away here; no uses of this function will actually write
-  // to it because it is only used for SOO iterators.
-  return const_cast<ctrl_t*>(kSooControl);
-}
-// Whether ctrl is from the SooControl array.
-inline bool IsSooControl(const ctrl_t* ctrl) { return ctrl == SooControl(); }
-
-// Returns a pointer to a generation to use for an empty hashtable.
-GenerationType* EmptyGeneration();
-
-// Returns whether `generation` is a generation for an empty hashtable that
-// could be returned by EmptyGeneration().
-inline bool IsEmptyGeneration(const GenerationType* generation) {
-  return *generation == SentinelEmptyGeneration();
-}
-
-// Mixes a randomly generated per-process seed with `hash` and `ctrl` to
-// randomize insertion order within groups.
-bool ShouldInsertBackwardsForDebug(size_t capacity, size_t hash,
-                                   const ctrl_t* ctrl);
-
-// Returns insert position for the given mask.
-// We want to add entropy even when ASLR is not enabled.
-// In debug build we will randomly insert in either the front or back of
-// the group.
-// TODO(kfm,sbenza): revisit after we do unconditional mixing
-template <class Mask>
-ABSL_ATTRIBUTE_ALWAYS_INLINE inline auto GetInsertionOffset(
-    Mask mask, ABSL_ATTRIBUTE_UNUSED size_t capacity,
-    ABSL_ATTRIBUTE_UNUSED size_t hash,
-    ABSL_ATTRIBUTE_UNUSED const ctrl_t* ctrl) {
-#if defined(NDEBUG)
-  return mask.LowestBitSet();
-#else
-  return ShouldInsertBackwardsForDebug(capacity, hash, ctrl)
-             ? mask.HighestBitSet()
-             : mask.LowestBitSet();
-#endif
-}
-
-// Returns a per-table, hash salt, which changes on resize. This gets mixed into
-// H1 to randomize iteration order per-table.
-//
-// The seed consists of the ctrl_ pointer, which adds enough entropy to ensure
-// non-determinism of iteration order in most cases.
-inline size_t PerTableSalt(const ctrl_t* ctrl) {
-  // The low bits of the pointer have little or no entropy because of
-  // alignment. We shift the pointer to try to use higher entropy bits. A
-  // good number seems to be 12 bits, because that aligns with page size.
-  return reinterpret_cast<uintptr_t>(ctrl) >> 12;
-}
-// Extracts the H1 portion of a hash: 57 bits mixed with a per-table salt.
-inline size_t H1(size_t hash, const ctrl_t* ctrl) {
-  return (hash >> 7) ^ PerTableSalt(ctrl);
-}
-
-// Extracts the H2 portion of a hash: the 7 bits not used for H1.
-//
-// These are used as an occupied control byte.
-inline h2_t H2(size_t hash) { return hash & 0x7F; }
-
-// Helpers for checking the state of a control byte.
-inline bool IsEmpty(ctrl_t c) { return c == ctrl_t::kEmpty; }
-inline bool IsFull(ctrl_t c) { return c >= static_cast<ctrl_t>(0); }
-inline bool IsDeleted(ctrl_t c) { return c == ctrl_t::kDeleted; }
-inline bool IsEmptyOrDeleted(ctrl_t c) { return c < ctrl_t::kSentinel; }
-
-#ifdef ABSL_INTERNAL_HAVE_SSE2
-// Quick reference guide for intrinsics used below:
-//
-// * __m128i: An XMM (128-bit) word.
-//
-// * _mm_setzero_si128: Returns a zero vector.
-// * _mm_set1_epi8:     Returns a vector with the same i8 in each lane.
-//
-// * _mm_subs_epi8:    Saturating-subtracts two i8 vectors.
-// * _mm_and_si128:    Ands two i128s together.
-// * _mm_or_si128:     Ors two i128s together.
-// * _mm_andnot_si128: And-nots two i128s together.
-//
-// * _mm_cmpeq_epi8: Component-wise compares two i8 vectors for equality,
-//                   filling each lane with 0x00 or 0xff.
-// * _mm_cmpgt_epi8: Same as above, but using > rather than ==.
-//
-// * _mm_loadu_si128:  Performs an unaligned load of an i128.
-// * _mm_storeu_si128: Performs an unaligned store of an i128.
-//
-// * _mm_sign_epi8:     Retains, negates, or zeroes each i8 lane of the first
-//                      argument if the corresponding lane of the second
-//                      argument is positive, negative, or zero, respectively.
-// * _mm_movemask_epi8: Selects the sign bit out of each i8 lane and produces a
-//                      bitmask consisting of those bits.
-// * _mm_shuffle_epi8:  Selects i8s from the first argument, using the low
-//                      four bits of each i8 lane in the second argument as
-//                      indices.
-
-// https://github.com/abseil/abseil-cpp/issues/209
-// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87853
-// _mm_cmpgt_epi8 is broken under GCC with -funsigned-char
-// Work around this by using the portable implementation of Group
-// when using -funsigned-char under GCC.
-inline __m128i _mm_cmpgt_epi8_fixed(__m128i a, __m128i b) {
-#if defined(__GNUC__) && !defined(__clang__)
-  if (std::is_unsigned<char>::value) {
-    const __m128i mask = _mm_set1_epi8(0x80);
-    const __m128i diff = _mm_subs_epi8(b, a);
-    return _mm_cmpeq_epi8(_mm_and_si128(diff, mask), mask);
-  }
-#endif
-  return _mm_cmpgt_epi8(a, b);
-}
-
-struct GroupSse2Impl {
-  static constexpr size_t kWidth = 16;  // the number of slots per group
-
-  explicit GroupSse2Impl(const ctrl_t* pos) {
-    ctrl = _mm_loadu_si128(reinterpret_cast<const __m128i*>(pos));
-  }
-
-  // Returns a bitmask representing the positions of slots that match hash.
-  BitMask<uint16_t, kWidth> Match(h2_t hash) const {
-    auto match = _mm_set1_epi8(static_cast<char>(hash));
-    BitMask<uint16_t, kWidth> result = BitMask<uint16_t, kWidth>(0);
-    result = BitMask<uint16_t, kWidth>(
-        static_cast<uint16_t>(_mm_movemask_epi8(_mm_cmpeq_epi8(match, ctrl))));
-    return result;
-  }
-
-  // Returns a bitmask representing the positions of empty slots.
-  NonIterableBitMask<uint16_t, kWidth> MaskEmpty() const {
-#ifdef ABSL_INTERNAL_HAVE_SSSE3
-    // This only works because ctrl_t::kEmpty is -128.
-    return NonIterableBitMask<uint16_t, kWidth>(
-        static_cast<uint16_t>(_mm_movemask_epi8(_mm_sign_epi8(ctrl, ctrl))));
-#else
-    auto match = _mm_set1_epi8(static_cast<char>(ctrl_t::kEmpty));
-    return NonIterableBitMask<uint16_t, kWidth>(
-        static_cast<uint16_t>(_mm_movemask_epi8(_mm_cmpeq_epi8(match, ctrl))));
-#endif
-  }
-
-  // Returns a bitmask representing the positions of full slots.
-  // Note: for `is_small()` tables group may contain the "same" slot twice:
-  // original and mirrored.
-  BitMask<uint16_t, kWidth> MaskFull() const {
-    return BitMask<uint16_t, kWidth>(
-        static_cast<uint16_t>(_mm_movemask_epi8(ctrl) ^ 0xffff));
-  }
-
-  // Returns a bitmask representing the positions of non full slots.
-  // Note: this includes: kEmpty, kDeleted, kSentinel.
-  // It is useful in contexts when kSentinel is not present.
-  auto MaskNonFull() const {
-    return BitMask<uint16_t, kWidth>(
-        static_cast<uint16_t>(_mm_movemask_epi8(ctrl)));
-  }
-
-  // Returns a bitmask representing the positions of empty or deleted slots.
-  NonIterableBitMask<uint16_t, kWidth> MaskEmptyOrDeleted() const {
-    auto special = _mm_set1_epi8(static_cast<char>(ctrl_t::kSentinel));
-    return NonIterableBitMask<uint16_t, kWidth>(static_cast<uint16_t>(
-        _mm_movemask_epi8(_mm_cmpgt_epi8_fixed(special, ctrl))));
-  }
-
-  // Returns the number of trailing empty or deleted elements in the group.
-  uint32_t CountLeadingEmptyOrDeleted() const {
-    auto special = _mm_set1_epi8(static_cast<char>(ctrl_t::kSentinel));
-    return TrailingZeros(static_cast<uint32_t>(
-        _mm_movemask_epi8(_mm_cmpgt_epi8_fixed(special, ctrl)) + 1));
-  }
-
-  void ConvertSpecialToEmptyAndFullToDeleted(ctrl_t* dst) const {
-    auto msbs = _mm_set1_epi8(static_cast<char>(-128));
-    auto x126 = _mm_set1_epi8(126);
-#ifdef ABSL_INTERNAL_HAVE_SSSE3
-    auto res = _mm_or_si128(_mm_shuffle_epi8(x126, ctrl), msbs);
-#else
-    auto zero = _mm_setzero_si128();
-    auto special_mask = _mm_cmpgt_epi8_fixed(zero, ctrl);
-    auto res = _mm_or_si128(msbs, _mm_andnot_si128(special_mask, x126));
-#endif
-    _mm_storeu_si128(reinterpret_cast<__m128i*>(dst), res);
-  }
-
-  __m128i ctrl;
-};
-#endif  // ABSL_INTERNAL_RAW_HASH_SET_HAVE_SSE2
-
-#if defined(ABSL_INTERNAL_HAVE_ARM_NEON) && defined(ABSL_IS_LITTLE_ENDIAN)
-struct GroupAArch64Impl {
-  static constexpr size_t kWidth = 8;
-
-  explicit GroupAArch64Impl(const ctrl_t* pos) {
-    ctrl = vld1_u8(reinterpret_cast<const uint8_t*>(pos));
-  }
-
-  auto Match(h2_t hash) const {
-    uint8x8_t dup = vdup_n_u8(hash);
-    auto mask = vceq_u8(ctrl, dup);
-    return BitMask<uint64_t, kWidth, /*Shift=*/3,
-                   /*NullifyBitsOnIteration=*/true>(
-        vget_lane_u64(vreinterpret_u64_u8(mask), 0));
-  }
-
-  NonIterableBitMask<uint64_t, kWidth, 3> MaskEmpty() const {
-    uint64_t mask =
-        vget_lane_u64(vreinterpret_u64_u8(vceq_s8(
-                          vdup_n_s8(static_cast<int8_t>(ctrl_t::kEmpty)),
-                          vreinterpret_s8_u8(ctrl))),
-                      0);
-    return NonIterableBitMask<uint64_t, kWidth, 3>(mask);
-  }
-
-  // Returns a bitmask representing the positions of full slots.
-  // Note: for `is_small()` tables group may contain the "same" slot twice:
-  // original and mirrored.
-  auto MaskFull() const {
-    uint64_t mask = vget_lane_u64(
-        vreinterpret_u64_u8(vcge_s8(vreinterpret_s8_u8(ctrl),
-                                    vdup_n_s8(static_cast<int8_t>(0)))),
-        0);
-    return BitMask<uint64_t, kWidth, /*Shift=*/3,
-                   /*NullifyBitsOnIteration=*/true>(mask);
-  }
-
-  // Returns a bitmask representing the positions of non full slots.
-  // Note: this includes: kEmpty, kDeleted, kSentinel.
-  // It is useful in contexts when kSentinel is not present.
-  auto MaskNonFull() const {
-    uint64_t mask = vget_lane_u64(
-        vreinterpret_u64_u8(vclt_s8(vreinterpret_s8_u8(ctrl),
-                                    vdup_n_s8(static_cast<int8_t>(0)))),
-        0);
-    return BitMask<uint64_t, kWidth, /*Shift=*/3,
-                   /*NullifyBitsOnIteration=*/true>(mask);
-  }
-
-  NonIterableBitMask<uint64_t, kWidth, 3> MaskEmptyOrDeleted() const {
-    uint64_t mask =
-        vget_lane_u64(vreinterpret_u64_u8(vcgt_s8(
-                          vdup_n_s8(static_cast<int8_t>(ctrl_t::kSentinel)),
-                          vreinterpret_s8_u8(ctrl))),
-                      0);
-    return NonIterableBitMask<uint64_t, kWidth, 3>(mask);
-  }
-
-  uint32_t CountLeadingEmptyOrDeleted() const {
-    uint64_t mask =
-        vget_lane_u64(vreinterpret_u64_u8(vcle_s8(
-                          vdup_n_s8(static_cast<int8_t>(ctrl_t::kSentinel)),
-                          vreinterpret_s8_u8(ctrl))),
-                      0);
-    // Similar to MaskEmptyorDeleted() but we invert the logic to invert the
-    // produced bitfield. We then count number of trailing zeros.
-    // Clang and GCC optimize countr_zero to rbit+clz without any check for 0,
-    // so we should be fine.
-    return static_cast<uint32_t>(countr_zero(mask)) >> 3;
-  }
-
-  void ConvertSpecialToEmptyAndFullToDeleted(ctrl_t* dst) const {
-    uint64_t mask = vget_lane_u64(vreinterpret_u64_u8(ctrl), 0);
-    constexpr uint64_t slsbs = 0x0202020202020202ULL;
-    constexpr uint64_t midbs = 0x7e7e7e7e7e7e7e7eULL;
-    auto x = slsbs & (mask >> 6);
-    auto res = (x + midbs) | kMsbs8Bytes;
-    little_endian::Store64(dst, res);
-  }
-
-  uint8x8_t ctrl;
-};
-#endif  // ABSL_INTERNAL_HAVE_ARM_NEON && ABSL_IS_LITTLE_ENDIAN
-
-struct GroupPortableImpl {
-  static constexpr size_t kWidth = 8;
-
-  explicit GroupPortableImpl(const ctrl_t* pos)
-      : ctrl(little_endian::Load64(pos)) {}
-
-  BitMask<uint64_t, kWidth, 3> Match(h2_t hash) const {
-    // For the technique, see:
-    // http://graphics.stanford.edu/~seander/bithacks.html##ValueInWord
-    // (Determine if a word has a byte equal to n).
-    //
-    // Caveat: there are false positives but:
-    // - they only occur if there is a real match
-    // - they never occur on ctrl_t::kEmpty, ctrl_t::kDeleted, ctrl_t::kSentinel
-    // - they will be handled gracefully by subsequent checks in code
-    //
-    // Example:
-    //   v = 0x1716151413121110
-    //   hash = 0x12
-    //   retval = (v - lsbs) & ~v & msbs = 0x0000000080800000
-    constexpr uint64_t lsbs = 0x0101010101010101ULL;
-    auto x = ctrl ^ (lsbs * hash);
-    return BitMask<uint64_t, kWidth, 3>((x - lsbs) & ~x & kMsbs8Bytes);
-  }
-
-  NonIterableBitMask<uint64_t, kWidth, 3> MaskEmpty() const {
-    return NonIterableBitMask<uint64_t, kWidth, 3>((ctrl & ~(ctrl << 6)) &
-                                                   kMsbs8Bytes);
-  }
-
-  // Returns a bitmask representing the positions of full slots.
-  // Note: for `is_small()` tables group may contain the "same" slot twice:
-  // original and mirrored.
-  BitMask<uint64_t, kWidth, 3> MaskFull() const {
-    return BitMask<uint64_t, kWidth, 3>((ctrl ^ kMsbs8Bytes) & kMsbs8Bytes);
-  }
-
-  // Returns a bitmask representing the positions of non full slots.
-  // Note: this includes: kEmpty, kDeleted, kSentinel.
-  // It is useful in contexts when kSentinel is not present.
-  auto MaskNonFull() const {
-    return BitMask<uint64_t, kWidth, 3>(ctrl  & kMsbs8Bytes);
-  }
-
-  NonIterableBitMask<uint64_t, kWidth, 3> MaskEmptyOrDeleted() const {
-    return NonIterableBitMask<uint64_t, kWidth, 3>((ctrl & ~(ctrl << 7)) &
-                                                   kMsbs8Bytes);
-  }
-
-  uint32_t CountLeadingEmptyOrDeleted() const {
-    // ctrl | ~(ctrl >> 7) will have the lowest bit set to zero for kEmpty and
-    // kDeleted. We lower all other bits and count number of trailing zeros.
-    constexpr uint64_t bits = 0x0101010101010101ULL;
-    return static_cast<uint32_t>(countr_zero((ctrl | ~(ctrl >> 7)) & bits) >>
-                                 3);
-  }
-
-  void ConvertSpecialToEmptyAndFullToDeleted(ctrl_t* dst) const {
-    constexpr uint64_t lsbs = 0x0101010101010101ULL;
-    auto x = ctrl & kMsbs8Bytes;
-    auto res = (~x + (x >> 7)) & ~lsbs;
-    little_endian::Store64(dst, res);
-  }
-
-  uint64_t ctrl;
-};
-
-#ifdef ABSL_INTERNAL_HAVE_SSE2
-using Group = GroupSse2Impl;
-using GroupFullEmptyOrDeleted = GroupSse2Impl;
-#elif defined(ABSL_INTERNAL_HAVE_ARM_NEON) && defined(ABSL_IS_LITTLE_ENDIAN)
-using Group = GroupAArch64Impl;
-// For Aarch64, we use the portable implementation for counting and masking
-// full, empty or deleted group elements. This is to avoid the latency of moving
-// between data GPRs and Neon registers when it does not provide a benefit.
-// Using Neon is profitable when we call Match(), but is not when we don't,
-// which is the case when we do *EmptyOrDeleted and MaskFull operations.
-// It is difficult to make a similar approach beneficial on other architectures
-// such as x86 since they have much lower GPR <-> vector register transfer
-// latency and 16-wide Groups.
-using GroupFullEmptyOrDeleted = GroupPortableImpl;
-#else
-using Group = GroupPortableImpl;
-using GroupFullEmptyOrDeleted = GroupPortableImpl;
-#endif
-
-// When there is an insertion with no reserved growth, we rehash with
-// probability `min(1, RehashProbabilityConstant() / capacity())`. Using a
-// constant divided by capacity ensures that inserting N elements is still O(N)
-// in the average case. Using the constant 16 means that we expect to rehash ~8
-// times more often than when generations are disabled. We are adding expected
-// rehash_probability * #insertions/capacity_growth = 16/capacity * ((7/8 -
-// 7/16) * capacity)/capacity_growth = ~7 extra rehashes per capacity growth.
-inline size_t RehashProbabilityConstant() { return 16; }
-
-class CommonFieldsGenerationInfoEnabled {
-  // A sentinel value for reserved_growth_ indicating that we just ran out of
-  // reserved growth on the last insertion. When reserve is called and then
-  // insertions take place, reserved_growth_'s state machine is N, ..., 1,
-  // kReservedGrowthJustRanOut, 0.
-  static constexpr size_t kReservedGrowthJustRanOut =
-      (std::numeric_limits<size_t>::max)();
-
- public:
-  CommonFieldsGenerationInfoEnabled() = default;
-  CommonFieldsGenerationInfoEnabled(CommonFieldsGenerationInfoEnabled&& that)
-      : reserved_growth_(that.reserved_growth_),
-        reservation_size_(that.reservation_size_),
-        generation_(that.generation_) {
-    that.reserved_growth_ = 0;
-    that.reservation_size_ = 0;
-    that.generation_ = EmptyGeneration();
-  }
-  CommonFieldsGenerationInfoEnabled& operator=(
-      CommonFieldsGenerationInfoEnabled&&) = default;
-
-  // Whether we should rehash on insert in order to detect bugs of using invalid
-  // references. We rehash on the first insertion after reserved_growth_ reaches
-  // 0 after a call to reserve. We also do a rehash with low probability
-  // whenever reserved_growth_ is zero.
-  bool should_rehash_for_bug_detection_on_insert(const ctrl_t* ctrl,
-                                                 size_t capacity) const;
-  // Similar to above, except that we don't depend on reserved_growth_.
-  bool should_rehash_for_bug_detection_on_move(const ctrl_t* ctrl,
-                                               size_t capacity) const;
-  void maybe_increment_generation_on_insert() {
-    if (reserved_growth_ == kReservedGrowthJustRanOut) reserved_growth_ = 0;
-
-    if (reserved_growth_ > 0) {
-      if (--reserved_growth_ == 0) reserved_growth_ = kReservedGrowthJustRanOut;
-    } else {
-      increment_generation();
-    }
-  }
-  void increment_generation() { *generation_ = NextGeneration(*generation_); }
-  void reset_reserved_growth(size_t reservation, size_t size) {
-    reserved_growth_ = reservation - size;
-  }
-  size_t reserved_growth() const { return reserved_growth_; }
-  void set_reserved_growth(size_t r) { reserved_growth_ = r; }
-  size_t reservation_size() const { return reservation_size_; }
-  void set_reservation_size(size_t r) { reservation_size_ = r; }
-  GenerationType generation() const { return *generation_; }
-  void set_generation(GenerationType g) { *generation_ = g; }
-  GenerationType* generation_ptr() const { return generation_; }
-  void set_generation_ptr(GenerationType* g) { generation_ = g; }
-
- private:
-  // The number of insertions remaining that are guaranteed to not rehash due to
-  // a prior call to reserve. Note: we store reserved growth in addition to
-  // reservation size because calls to erase() decrease size_ but don't decrease
-  // reserved growth.
-  size_t reserved_growth_ = 0;
-  // The maximum argument to reserve() since the container was cleared. We need
-  // to keep track of this, in addition to reserved growth, because we reset
-  // reserved growth to this when erase(begin(), end()) is called.
-  size_t reservation_size_ = 0;
-  // Pointer to the generation counter, which is used to validate iterators and
-  // is stored in the backing array between the control bytes and the slots.
-  // Note that we can't store the generation inside the container itself and
-  // keep a pointer to the container in the iterators because iterators must
-  // remain valid when the container is moved.
-  // Note: we could derive this pointer from the control pointer, but it makes
-  // the code more complicated, and there's a benefit in having the sizes of
-  // raw_hash_set in sanitizer mode and non-sanitizer mode a bit more different,
-  // which is that tests are less likely to rely on the size remaining the same.
-  GenerationType* generation_ = EmptyGeneration();
-};
-
-class CommonFieldsGenerationInfoDisabled {
- public:
-  CommonFieldsGenerationInfoDisabled() = default;
-  CommonFieldsGenerationInfoDisabled(CommonFieldsGenerationInfoDisabled&&) =
-      default;
-  CommonFieldsGenerationInfoDisabled& operator=(
-      CommonFieldsGenerationInfoDisabled&&) = default;
-
-  bool should_rehash_for_bug_detection_on_insert(const ctrl_t*, size_t) const {
-    return false;
-  }
-  bool should_rehash_for_bug_detection_on_move(const ctrl_t*, size_t) const {
-    return false;
-  }
-  void maybe_increment_generation_on_insert() {}
-  void increment_generation() {}
-  void reset_reserved_growth(size_t, size_t) {}
-  size_t reserved_growth() const { return 0; }
-  void set_reserved_growth(size_t) {}
-  size_t reservation_size() const { return 0; }
-  void set_reservation_size(size_t) {}
-  GenerationType generation() const { return 0; }
-  void set_generation(GenerationType) {}
-  GenerationType* generation_ptr() const { return nullptr; }
-  void set_generation_ptr(GenerationType*) {}
-};
-
-class HashSetIteratorGenerationInfoEnabled {
- public:
-  HashSetIteratorGenerationInfoEnabled() = default;
-  explicit HashSetIteratorGenerationInfoEnabled(
-      const GenerationType* generation_ptr)
-      : generation_ptr_(generation_ptr), generation_(*generation_ptr) {}
-
-  GenerationType generation() const { return generation_; }
-  void reset_generation() { generation_ = *generation_ptr_; }
-  const GenerationType* generation_ptr() const { return generation_ptr_; }
-  void set_generation_ptr(const GenerationType* ptr) { generation_ptr_ = ptr; }
-
- private:
-  const GenerationType* generation_ptr_ = EmptyGeneration();
-  GenerationType generation_ = *generation_ptr_;
-};
-
-class HashSetIteratorGenerationInfoDisabled {
- public:
-  HashSetIteratorGenerationInfoDisabled() = default;
-  explicit HashSetIteratorGenerationInfoDisabled(const GenerationType*) {}
-
-  GenerationType generation() const { return 0; }
-  void reset_generation() {}
-  const GenerationType* generation_ptr() const { return nullptr; }
-  void set_generation_ptr(const GenerationType*) {}
-};
-
-#ifdef ABSL_SWISSTABLE_ENABLE_GENERATIONS
-using CommonFieldsGenerationInfo = CommonFieldsGenerationInfoEnabled;
-using HashSetIteratorGenerationInfo = HashSetIteratorGenerationInfoEnabled;
-#else
-using CommonFieldsGenerationInfo = CommonFieldsGenerationInfoDisabled;
-using HashSetIteratorGenerationInfo = HashSetIteratorGenerationInfoDisabled;
-#endif
-
-// Returns whether `n` is a valid capacity (i.e., number of slots).
-//
-// A valid capacity is a non-zero integer `2^m - 1`.
-inline bool IsValidCapacity(size_t n) { return ((n + 1) & n) == 0 && n > 0; }
-
-// Returns the number of "cloned control bytes".
-//
-// This is the number of control bytes that are present both at the beginning
-// of the control byte array and at the end, such that we can create a
-// `Group::kWidth`-width probe window starting from any control byte.
-constexpr size_t NumClonedBytes() { return Group::kWidth - 1; }
-
-// Returns the number of control bytes including cloned.
-constexpr size_t NumControlBytes(size_t capacity) {
-  return capacity + 1 + NumClonedBytes();
-}
-
-// Computes the offset from the start of the backing allocation of control.
-// infoz and growth_left are stored at the beginning of the backing array.
-inline static size_t ControlOffset(bool has_infoz) {
-  return (has_infoz ? sizeof(HashtablezInfoHandle) : 0) + sizeof(size_t);
-}
-
-// Helper class for computing offsets and allocation size of hash set fields.
-class RawHashSetLayout {
- public:
-  explicit RawHashSetLayout(size_t capacity, size_t slot_align, bool has_infoz)
-      : capacity_(capacity),
-        control_offset_(ControlOffset(has_infoz)),
-        generation_offset_(control_offset_ + NumControlBytes(capacity)),
-        slot_offset_(
-            (generation_offset_ + NumGenerationBytes() + slot_align - 1) &
-            (~slot_align + 1)) {
-    assert(IsValidCapacity(capacity));
-  }
-
-  // Returns the capacity of a table.
-  size_t capacity() const { return capacity_; }
-
-  // Returns precomputed offset from the start of the backing allocation of
-  // control.
-  size_t control_offset() const { return control_offset_; }
-
-  // Given the capacity of a table, computes the offset (from the start of the
-  // backing allocation) of the generation counter (if it exists).
-  size_t generation_offset() const { return generation_offset_; }
-
-  // Given the capacity of a table, computes the offset (from the start of the
-  // backing allocation) at which the slots begin.
-  size_t slot_offset() const { return slot_offset_; }
-
-  // Given the capacity of a table, computes the total size of the backing
-  // array.
-  size_t alloc_size(size_t slot_size) const {
-    return slot_offset_ + capacity_ * slot_size;
-  }
-
- private:
-  size_t capacity_;
-  size_t control_offset_;
-  size_t generation_offset_;
-  size_t slot_offset_;
-};
-
-// We only allow a maximum of 1 SOO element, which makes the implementation
-// much simpler. Complications with multiple SOO elements include:
-// - Satisfying the guarantee that erasing one element doesn't invalidate
-//   iterators to other elements means we would probably need actual SOO
-//   control bytes.
-// - In order to prevent user code from depending on iteration order for small
-//   tables, we would need to randomize the iteration order somehow.
-constexpr size_t SooCapacity() { return 1; }
-// Sentinel type to indicate SOO CommonFields construction.
-struct soo_tag_t {};
-// Sentinel type to indicate SOO CommonFields construction with full size.
-struct full_soo_tag_t {};
-
-// This allows us to work around an uninitialized memory warning when
-// constructing begin() iterators in empty hashtables.
-union MaybeInitializedPtr {
-  void* get() const {
-    // Suppress erroneous uninitialized memory errors on GCC. GCC thinks that
-    // the call to slot_array() in find_or_prepare_insert() is reading
-    // uninitialized memory, but slot_array is only called there when the table
-    // is non-empty and this memory is initialized when the table is non-empty.
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
-#pragma GCC diagnostic ignored "-Wuninitialized"
-#endif
-    return p;
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic pop
-#endif
-  }
-  void set(void* ptr) { p = ptr; }
-
-  void* p;
-};
-
-struct HeapPtrs {
-  HeapPtrs() = default;
-  explicit HeapPtrs(ctrl_t* c) : control(c) {}
-
-  // The control bytes (and, also, a pointer near to the base of the backing
-  // array).
-  //
-  // This contains `capacity + 1 + NumClonedBytes()` entries, even
-  // when the table is empty (hence EmptyGroup).
-  //
-  // Note that growth_left is stored immediately before this pointer.
-  // May be uninitialized for SOO tables.
-  ctrl_t* control;
-
-  // The beginning of the slots, located at `SlotOffset()` bytes after
-  // `control`. May be uninitialized for empty tables.
-  // Note: we can't use `slots` because Qt defines "slots" as a macro.
-  MaybeInitializedPtr slot_array;
-};
-
-// Manages the backing array pointers or the SOO slot. When raw_hash_set::is_soo
-// is true, the SOO slot is stored in `soo_data`. Otherwise, we use `heap`.
-union HeapOrSoo {
-  HeapOrSoo() = default;
-  explicit HeapOrSoo(ctrl_t* c) : heap(c) {}
-
-  HeapPtrs heap;
-  unsigned char soo_data[sizeof(HeapPtrs)];
-};
-
-// CommonFields hold the fields in raw_hash_set that do not depend
-// on template parameters. This allows us to conveniently pass all
-// of this state to helper functions as a single argument.
-class CommonFields : public CommonFieldsGenerationInfo {
- public:
-  CommonFields() : capacity_(0), size_(0), heap_or_soo_(EmptyGroup()) {}
-  explicit CommonFields(soo_tag_t) : capacity_(SooCapacity()), size_(0) {}
-  explicit CommonFields(full_soo_tag_t)
-      : capacity_(SooCapacity()), size_(size_t{1} << HasInfozShift()) {}
-
-  // Not copyable
-  CommonFields(const CommonFields&) = delete;
-  CommonFields& operator=(const CommonFields&) = delete;
-
-  // Movable
-  CommonFields(CommonFields&& that) = default;
-  CommonFields& operator=(CommonFields&&) = default;
-
-  template <bool kSooEnabled>
-  static CommonFields CreateDefault() {
-    return kSooEnabled ? CommonFields{soo_tag_t{}} : CommonFields{};
-  }
-
-  // The inline data for SOO is written on top of control_/slots_.
-  const void* soo_data() const { return heap_or_soo_.soo_data; }
-  void* soo_data() { return heap_or_soo_.soo_data; }
-
-  HeapOrSoo heap_or_soo() const { return heap_or_soo_; }
-  const HeapOrSoo& heap_or_soo_ref() const { return heap_or_soo_; }
-
-  ctrl_t* control() const { return heap_or_soo_.heap.control; }
-  void set_control(ctrl_t* c) { heap_or_soo_.heap.control = c; }
-  void* backing_array_start() const {
-    // growth_left (and maybe infoz) is stored before control bytes.
-    assert(reinterpret_cast<uintptr_t>(control()) % alignof(size_t) == 0);
-    return control() - ControlOffset(has_infoz());
-  }
-
-  // Note: we can't use slots() because Qt defines "slots" as a macro.
-  void* slot_array() const { return heap_or_soo_.heap.slot_array.get(); }
-  MaybeInitializedPtr slots_union() const {
-    // Suppress erroneous uninitialized memory errors on GCC.
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
-#endif
-    return heap_or_soo_.heap.slot_array;
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic pop
-#endif
-  }
-  void set_slots(void* s) { heap_or_soo_.heap.slot_array.set(s); }
-
-  // The number of filled slots.
-  size_t size() const { return size_ >> HasInfozShift(); }
-  void set_size(size_t s) {
-    size_ = (s << HasInfozShift()) | (size_ & HasInfozMask());
-  }
-  void set_empty_soo() {
-    AssertInSooMode();
-    size_ = 0;
-  }
-  void set_full_soo() {
-    AssertInSooMode();
-    size_ = size_t{1} << HasInfozShift();
-  }
-  void increment_size() {
-    assert(size() < capacity());
-    size_ += size_t{1} << HasInfozShift();
-  }
-  void decrement_size() {
-    assert(size() > 0);
-    size_ -= size_t{1} << HasInfozShift();
-  }
-
-  // The total number of available slots.
-  size_t capacity() const { return capacity_; }
-  void set_capacity(size_t c) {
-    assert(c == 0 || IsValidCapacity(c));
-    capacity_ = c;
-  }
-
-  // The number of slots we can still fill without needing to rehash.
-  // This is stored in the heap allocation before the control bytes.
-  // TODO(b/289225379): experiment with moving growth_left back inline to
-  // increase room for SOO.
-  size_t growth_left() const {
-    const size_t* gl_ptr = reinterpret_cast<size_t*>(control()) - 1;
-    assert(reinterpret_cast<uintptr_t>(gl_ptr) % alignof(size_t) == 0);
-    return *gl_ptr;
-  }
-  void set_growth_left(size_t gl) {
-    size_t* gl_ptr = reinterpret_cast<size_t*>(control()) - 1;
-    assert(reinterpret_cast<uintptr_t>(gl_ptr) % alignof(size_t) == 0);
-    *gl_ptr = gl;
-  }
-
-  bool has_infoz() const {
-    return ABSL_PREDICT_FALSE((size_ & HasInfozMask()) != 0);
-  }
-  void set_has_infoz(bool has_infoz) {
-    size_ = (size() << HasInfozShift()) | static_cast<size_t>(has_infoz);
-  }
-
-  HashtablezInfoHandle infoz() {
-    return has_infoz()
-               ? *reinterpret_cast<HashtablezInfoHandle*>(backing_array_start())
-               : HashtablezInfoHandle();
-  }
-  void set_infoz(HashtablezInfoHandle infoz) {
-    assert(has_infoz());
-    *reinterpret_cast<HashtablezInfoHandle*>(backing_array_start()) = infoz;
-  }
-
-  bool should_rehash_for_bug_detection_on_insert() const {
-    return CommonFieldsGenerationInfo::
-        should_rehash_for_bug_detection_on_insert(control(), capacity());
-  }
-  bool should_rehash_for_bug_detection_on_move() const {
-    return CommonFieldsGenerationInfo::
-        should_rehash_for_bug_detection_on_move(control(), capacity());
-  }
-  void reset_reserved_growth(size_t reservation) {
-    CommonFieldsGenerationInfo::reset_reserved_growth(reservation, size());
-  }
-
-  // The size of the backing array allocation.
-  size_t alloc_size(size_t slot_size, size_t slot_align) const {
-    return RawHashSetLayout(capacity(), slot_align, has_infoz())
-        .alloc_size(slot_size);
-  }
-
-  // Move fields other than heap_or_soo_.
-  void move_non_heap_or_soo_fields(CommonFields& that) {
-    static_cast<CommonFieldsGenerationInfo&>(*this) =
-        std::move(static_cast<CommonFieldsGenerationInfo&>(that));
-    capacity_ = that.capacity_;
-    size_ = that.size_;
-  }
-
-  // Returns the number of control bytes set to kDeleted. For testing only.
-  size_t TombstonesCount() const {
-    return static_cast<size_t>(
-        std::count(control(), control() + capacity(), ctrl_t::kDeleted));
-  }
-
- private:
-  // We store the has_infoz bit in the lowest bit of size_.
-  static constexpr size_t HasInfozShift() { return 1; }
-  static constexpr size_t HasInfozMask() {
-    return (size_t{1} << HasInfozShift()) - 1;
-  }
-
-  // We can't assert that SOO is enabled because we don't have SooEnabled(), but
-  // we assert what we can.
-  void AssertInSooMode() const {
-    assert(capacity() == SooCapacity());
-    assert(!has_infoz());
-  }
-
-  // The number of slots in the backing array. This is always 2^N-1 for an
-  // integer N. NOTE: we tried experimenting with compressing the capacity and
-  // storing it together with size_: (a) using 6 bits to store the corresponding
-  // power (N in 2^N-1), and (b) storing 2^N as the most significant bit of
-  // size_ and storing size in the low bits. Both of these experiments were
-  // regressions, presumably because we need capacity to do find operations.
-  size_t capacity_;
-
-  // The size and also has one bit that stores whether we have infoz.
-  // TODO(b/289225379): we could put size_ into HeapOrSoo and make capacity_
-  // encode the size in SOO case. We would be making size()/capacity() more
-  // expensive in order to have more SOO space.
-  size_t size_;
-
-  // Either the control/slots pointers or the SOO slot.
-  HeapOrSoo heap_or_soo_;
-};
-
-template <class Policy, class Hash, class Eq, class Alloc>
-class raw_hash_set;
-
-// Returns the next valid capacity after `n`.
-inline size_t NextCapacity(size_t n) {
-  assert(IsValidCapacity(n) || n == 0);
-  return n * 2 + 1;
-}
-
-// Applies the following mapping to every byte in the control array:
-//   * kDeleted -> kEmpty
-//   * kEmpty -> kEmpty
-//   * _ -> kDeleted
-// PRECONDITION:
-//   IsValidCapacity(capacity)
-//   ctrl[capacity] == ctrl_t::kSentinel
-//   ctrl[i] != ctrl_t::kSentinel for all i < capacity
-void ConvertDeletedToEmptyAndFullToDeleted(ctrl_t* ctrl, size_t capacity);
-
-// Converts `n` into the next valid capacity, per `IsValidCapacity`.
-inline size_t NormalizeCapacity(size_t n) {
-  return n ? ~size_t{} >> countl_zero(n) : 1;
-}
-
-// General notes on capacity/growth methods below:
-// - We use 7/8th as maximum load factor. For 16-wide groups, that gives an
-//   average of two empty slots per group.
-// - For (capacity+1) >= Group::kWidth, growth is 7/8*capacity.
-// - For (capacity+1) < Group::kWidth, growth == capacity. In this case, we
-//   never need to probe (the whole table fits in one group) so we don't need a
-//   load factor less than 1.
-
-// Given `capacity`, applies the load factor; i.e., it returns the maximum
-// number of values we should put into the table before a resizing rehash.
-inline size_t CapacityToGrowth(size_t capacity) {
-  assert(IsValidCapacity(capacity));
-  // `capacity*7/8`
-  if (Group::kWidth == 8 && capacity == 7) {
-    // x-x/8 does not work when x==7.
-    return 6;
-  }
-  return capacity - capacity / 8;
-}
-
-// Given `growth`, "unapplies" the load factor to find how large the capacity
-// should be to stay within the load factor.
-//
-// This might not be a valid capacity and `NormalizeCapacity()` should be
-// called on this.
-inline size_t GrowthToLowerboundCapacity(size_t growth) {
-  // `growth*8/7`
-  if (Group::kWidth == 8 && growth == 7) {
-    // x+(x-1)/7 does not work when x==7.
-    return 8;
-  }
-  return growth + static_cast<size_t>((static_cast<int64_t>(growth) - 1) / 7);
-}
-
-template <class InputIter>
-size_t SelectBucketCountForIterRange(InputIter first, InputIter last,
-                                     size_t bucket_count) {
-  if (bucket_count != 0) {
-    return bucket_count;
-  }
-  using InputIterCategory =
-      typename std::iterator_traits<InputIter>::iterator_category;
-  if (std::is_base_of<std::random_access_iterator_tag,
-                      InputIterCategory>::value) {
-    return GrowthToLowerboundCapacity(
-        static_cast<size_t>(std::distance(first, last)));
-  }
-  return 0;
-}
-
-constexpr bool SwisstableDebugEnabled() {
-#if defined(ABSL_SWISSTABLE_ENABLE_GENERATIONS) || \
-    ABSL_OPTION_HARDENED == 1 || !defined(NDEBUG)
-  return true;
-#else
-  return false;
-#endif
-}
-
-inline void AssertIsFull(const ctrl_t* ctrl, GenerationType generation,
-                         const GenerationType* generation_ptr,
-                         const char* operation) {
-  if (!SwisstableDebugEnabled()) return;
-  // `SwisstableDebugEnabled()` is also true for release builds with hardening
-  // enabled. To minimize their impact in those builds:
-  // - use `ABSL_PREDICT_FALSE()` to provide a compiler hint for code layout
-  // - use `ABSL_RAW_LOG()` with a format string to reduce code size and improve
-  //   the chances that the hot paths will be inlined.
-  if (ABSL_PREDICT_FALSE(ctrl == nullptr)) {
-    ABSL_RAW_LOG(FATAL, "%s called on end() iterator.", operation);
-  }
-  if (ABSL_PREDICT_FALSE(ctrl == EmptyGroup())) {
-    ABSL_RAW_LOG(FATAL, "%s called on default-constructed iterator.",
-                 operation);
-  }
-  if (SwisstableGenerationsEnabled()) {
-    if (ABSL_PREDICT_FALSE(generation != *generation_ptr)) {
-      ABSL_RAW_LOG(FATAL,
-                   "%s called on invalid iterator. The table could have "
-                   "rehashed or moved since this iterator was initialized.",
-                   operation);
-    }
-    if (ABSL_PREDICT_FALSE(!IsFull(*ctrl))) {
-      ABSL_RAW_LOG(
-          FATAL,
-          "%s called on invalid iterator. The element was likely erased.",
-          operation);
-    }
-  } else {
-    if (ABSL_PREDICT_FALSE(!IsFull(*ctrl))) {
-      ABSL_RAW_LOG(
-          FATAL,
-          "%s called on invalid iterator. The element might have been erased "
-          "or the table might have rehashed. Consider running with "
-          "--config=asan to diagnose rehashing issues.",
-          operation);
-    }
-  }
-}
-
-// Note that for comparisons, null/end iterators are valid.
-inline void AssertIsValidForComparison(const ctrl_t* ctrl,
-                                       GenerationType generation,
-                                       const GenerationType* generation_ptr) {
-  if (!SwisstableDebugEnabled()) return;
-  const bool ctrl_is_valid_for_comparison =
-      ctrl == nullptr || ctrl == EmptyGroup() || IsFull(*ctrl);
-  if (SwisstableGenerationsEnabled()) {
-    if (ABSL_PREDICT_FALSE(generation != *generation_ptr)) {
-      ABSL_RAW_LOG(FATAL,
-                   "Invalid iterator comparison. The table could have rehashed "
-                   "or moved since this iterator was initialized.");
-    }
-    if (ABSL_PREDICT_FALSE(!ctrl_is_valid_for_comparison)) {
-      ABSL_RAW_LOG(
-          FATAL, "Invalid iterator comparison. The element was likely erased.");
-    }
-  } else {
-    ABSL_HARDENING_ASSERT(
-        ctrl_is_valid_for_comparison &&
-        "Invalid iterator comparison. The element might have been erased or "
-        "the table might have rehashed. Consider running with --config=asan to "
-        "diagnose rehashing issues.");
-  }
-}
-
-// If the two iterators come from the same container, then their pointers will
-// interleave such that ctrl_a <= ctrl_b < slot_a <= slot_b or vice/versa.
-// Note: we take slots by reference so that it's not UB if they're uninitialized
-// as long as we don't read them (when ctrl is null).
-inline bool AreItersFromSameContainer(const ctrl_t* ctrl_a,
-                                      const ctrl_t* ctrl_b,
-                                      const void* const& slot_a,
-                                      const void* const& slot_b) {
-  // If either control byte is null, then we can't tell.
-  if (ctrl_a == nullptr || ctrl_b == nullptr) return true;
-  const bool a_is_soo = IsSooControl(ctrl_a);
-  if (a_is_soo != IsSooControl(ctrl_b)) return false;
-  if (a_is_soo) return slot_a == slot_b;
-
-  const void* low_slot = slot_a;
-  const void* hi_slot = slot_b;
-  if (ctrl_a > ctrl_b) {
-    std::swap(ctrl_a, ctrl_b);
-    std::swap(low_slot, hi_slot);
-  }
-  return ctrl_b < low_slot && low_slot <= hi_slot;
-}
-
-// Asserts that two iterators come from the same container.
-// Note: we take slots by reference so that it's not UB if they're uninitialized
-// as long as we don't read them (when ctrl is null).
-inline void AssertSameContainer(const ctrl_t* ctrl_a, const ctrl_t* ctrl_b,
-                                const void* const& slot_a,
-                                const void* const& slot_b,
-                                const GenerationType* generation_ptr_a,
-                                const GenerationType* generation_ptr_b) {
-  if (!SwisstableDebugEnabled()) return;
-  // `SwisstableDebugEnabled()` is also true for release builds with hardening
-  // enabled. To minimize their impact in those builds:
-  // - use `ABSL_PREDICT_FALSE()` to provide a compiler hint for code layout
-  // - use `ABSL_RAW_LOG()` with a format string to reduce code size and improve
-  //   the chances that the hot paths will be inlined.
-
-  // fail_if(is_invalid, message) crashes when is_invalid is true and provides
-  // an error message based on `message`.
-  const auto fail_if = [](bool is_invalid, const char* message) {
-    if (ABSL_PREDICT_FALSE(is_invalid)) {
-      ABSL_RAW_LOG(FATAL, "Invalid iterator comparison. %s", message);
-    }
-  };
-
-  const bool a_is_default = ctrl_a == EmptyGroup();
-  const bool b_is_default = ctrl_b == EmptyGroup();
-  if (a_is_default && b_is_default) return;
-  fail_if(a_is_default != b_is_default,
-          "Comparing default-constructed hashtable iterator with a "
-          "non-default-constructed hashtable iterator.");
-
-  if (SwisstableGenerationsEnabled()) {
-    if (ABSL_PREDICT_TRUE(generation_ptr_a == generation_ptr_b)) return;
-    // Users don't need to know whether the tables are SOO so don't mention SOO
-    // in the debug message.
-    const bool a_is_soo = IsSooControl(ctrl_a);
-    const bool b_is_soo = IsSooControl(ctrl_b);
-    fail_if(a_is_soo != b_is_soo || (a_is_soo && b_is_soo),
-            "Comparing iterators from different hashtables.");
-
-    const bool a_is_empty = IsEmptyGeneration(generation_ptr_a);
-    const bool b_is_empty = IsEmptyGeneration(generation_ptr_b);
-    fail_if(a_is_empty != b_is_empty,
-            "Comparing an iterator from an empty hashtable with an iterator "
-            "from a non-empty hashtable.");
-    fail_if(a_is_empty && b_is_empty,
-            "Comparing iterators from different empty hashtables.");
-
-    const bool a_is_end = ctrl_a == nullptr;
-    const bool b_is_end = ctrl_b == nullptr;
-    fail_if(a_is_end || b_is_end,
-            "Comparing iterator with an end() iterator from a different "
-            "hashtable.");
-    fail_if(true, "Comparing non-end() iterators from different hashtables.");
-  } else {
-    ABSL_HARDENING_ASSERT(
-        AreItersFromSameContainer(ctrl_a, ctrl_b, slot_a, slot_b) &&
-        "Invalid iterator comparison. The iterators may be from different "
-        "containers or the container might have rehashed or moved. Consider "
-        "running with --config=asan to diagnose issues.");
-  }
-}
-
-struct FindInfo {
-  size_t offset;
-  size_t probe_length;
-};
-
-// Whether a table is "small". A small table fits entirely into a probing
-// group, i.e., has a capacity < `Group::kWidth`.
-//
-// In small mode we are able to use the whole capacity. The extra control
-// bytes give us at least one "empty" control byte to stop the iteration.
-// This is important to make 1 a valid capacity.
-//
-// In small mode only the first `capacity` control bytes after the sentinel
-// are valid. The rest contain dummy ctrl_t::kEmpty values that do not
-// represent a real slot. This is important to take into account on
-// `find_first_non_full()`, where we never try
-// `ShouldInsertBackwards()` for small tables.
-inline bool is_small(size_t capacity) { return capacity < Group::kWidth - 1; }
-
-// Whether a table fits entirely into a probing group.
-// Arbitrary order of elements in such tables is correct.
-inline bool is_single_group(size_t capacity) {
-  return capacity <= Group::kWidth;
-}
-
-// Begins a probing operation on `common.control`, using `hash`.
-inline probe_seq<Group::kWidth> probe(const ctrl_t* ctrl, const size_t capacity,
-                                      size_t hash) {
-  return probe_seq<Group::kWidth>(H1(hash, ctrl), capacity);
-}
-inline probe_seq<Group::kWidth> probe(const CommonFields& common, size_t hash) {
-  return probe(common.control(), common.capacity(), hash);
-}
-
-// Probes an array of control bits using a probe sequence derived from `hash`,
-// and returns the offset corresponding to the first deleted or empty slot.
-//
-// Behavior when the entire table is full is undefined.
-//
-// NOTE: this function must work with tables having both empty and deleted
-// slots in the same group. Such tables appear during `erase()`.
-template <typename = void>
-inline FindInfo find_first_non_full(const CommonFields& common, size_t hash) {
-  auto seq = probe(common, hash);
-  const ctrl_t* ctrl = common.control();
-  while (true) {
-    GroupFullEmptyOrDeleted g{ctrl + seq.offset()};
-    auto mask = g.MaskEmptyOrDeleted();
-    if (mask) {
-      return {
-          seq.offset(GetInsertionOffset(mask, common.capacity(), hash, ctrl)),
-          seq.index()};
-    }
-    seq.next();
-    assert(seq.index() <= common.capacity() && "full table!");
-  }
-}
-
-// Extern template for inline function keep possibility of inlining.
-// When compiler decided to not inline, no symbols will be added to the
-// corresponding translation unit.
-extern template FindInfo find_first_non_full(const CommonFields&, size_t);
-
-// Non-inlined version of find_first_non_full for use in less
-// performance critical routines.
-FindInfo find_first_non_full_outofline(const CommonFields&, size_t);
-
-inline void ResetGrowthLeft(CommonFields& common) {
-  common.set_growth_left(CapacityToGrowth(common.capacity()) - common.size());
-}
-
-// Sets `ctrl` to `{kEmpty, kSentinel, ..., kEmpty}`, marking the entire
-// array as marked as empty.
-inline void ResetCtrl(CommonFields& common, size_t slot_size) {
-  const size_t capacity = common.capacity();
-  ctrl_t* ctrl = common.control();
-  std::memset(ctrl, static_cast<int8_t>(ctrl_t::kEmpty),
-              capacity + 1 + NumClonedBytes());
-  ctrl[capacity] = ctrl_t::kSentinel;
-  SanitizerPoisonMemoryRegion(common.slot_array(), slot_size * capacity);
-}
-
-// Sets sanitizer poisoning for slot corresponding to control byte being set.
-inline void DoSanitizeOnSetCtrl(const CommonFields& c, size_t i, ctrl_t h,
-                                size_t slot_size) {
-  assert(i < c.capacity());
-  auto* slot_i = static_cast<const char*>(c.slot_array()) + i * slot_size;
-  if (IsFull(h)) {
-    SanitizerUnpoisonMemoryRegion(slot_i, slot_size);
-  } else {
-    SanitizerPoisonMemoryRegion(slot_i, slot_size);
-  }
-}
-
-// Sets `ctrl[i]` to `h`.
-//
-// Unlike setting it directly, this function will perform bounds checks and
-// mirror the value to the cloned tail if necessary.
-inline void SetCtrl(const CommonFields& c, size_t i, ctrl_t h,
-                    size_t slot_size) {
-  DoSanitizeOnSetCtrl(c, i, h, slot_size);
-  ctrl_t* ctrl = c.control();
-  ctrl[i] = h;
-  ctrl[((i - NumClonedBytes()) & c.capacity()) +
-       (NumClonedBytes() & c.capacity())] = h;
-}
-// Overload for setting to an occupied `h2_t` rather than a special `ctrl_t`.
-inline void SetCtrl(const CommonFields& c, size_t i, h2_t h, size_t slot_size) {
-  SetCtrl(c, i, static_cast<ctrl_t>(h), slot_size);
-}
-
-// Like SetCtrl, but in a single group table, we can save some operations when
-// setting the cloned control byte.
-inline void SetCtrlInSingleGroupTable(const CommonFields& c, size_t i, ctrl_t h,
-                                      size_t slot_size) {
-  assert(is_single_group(c.capacity()));
-  DoSanitizeOnSetCtrl(c, i, h, slot_size);
-  ctrl_t* ctrl = c.control();
-  ctrl[i] = h;
-  ctrl[i + c.capacity() + 1] = h;
-}
-// Overload for setting to an occupied `h2_t` rather than a special `ctrl_t`.
-inline void SetCtrlInSingleGroupTable(const CommonFields& c, size_t i, h2_t h,
-                                      size_t slot_size) {
-  SetCtrlInSingleGroupTable(c, i, static_cast<ctrl_t>(h), slot_size);
-}
-
-// growth_left (which is a size_t) is stored with the backing array.
-constexpr size_t BackingArrayAlignment(size_t align_of_slot) {
-  return (std::max)(align_of_slot, alignof(size_t));
-}
-
-// Returns the address of the ith slot in slots where each slot occupies
-// slot_size.
-inline void* SlotAddress(void* slot_array, size_t slot, size_t slot_size) {
-  return reinterpret_cast<void*>(reinterpret_cast<char*>(slot_array) +
-                                 (slot * slot_size));
-}
-
-// Iterates over all full slots and calls `cb(const ctrl_t*, SlotType*)`.
-// NOTE: no erasure from this table allowed during Callback call.
-template <class SlotType, class Callback>
-ABSL_ATTRIBUTE_ALWAYS_INLINE inline void IterateOverFullSlots(
-    const CommonFields& c, SlotType* slot, Callback cb) {
-  const size_t cap = c.capacity();
-  const ctrl_t* ctrl = c.control();
-  if (is_small(cap)) {
-    // Mirrored/cloned control bytes in small table are also located in the
-    // first group (starting from position 0). We are taking group from position
-    // `capacity` in order to avoid duplicates.
-
-    // Small tables capacity fits into portable group, where
-    // GroupPortableImpl::MaskFull is more efficient for the
-    // capacity <= GroupPortableImpl::kWidth.
-    assert(cap <= GroupPortableImpl::kWidth &&
-           "unexpectedly large small capacity");
-    static_assert(Group::kWidth >= GroupPortableImpl::kWidth,
-                  "unexpected group width");
-    // Group starts from kSentinel slot, so indices in the mask will
-    // be increased by 1.
-    const auto mask = GroupPortableImpl(ctrl + cap).MaskFull();
-    --ctrl;
-    --slot;
-    for (uint32_t i : mask) {
-      cb(ctrl + i, slot + i);
-    }
-    return;
-  }
-  size_t remaining = c.size();
-  while (remaining != 0) {
-    for (uint32_t i : GroupFullEmptyOrDeleted(ctrl).MaskFull()) {
-      cb(ctrl + i, slot + i);
-      --remaining;
-    }
-    slot += Group::kWidth;
-    ctrl += Group::kWidth;
-  }
-}
-
-// Helper class to perform resize of the hash set.
-//
-// It contains special optimizations for small group resizes.
-// See GrowIntoSingleGroupShuffleControlBytes for details.
-class HashSetResizeHelper {
- public:
-  explicit HashSetResizeHelper(CommonFields& c, bool was_soo, bool had_soo_slot)
-      : old_capacity_(c.capacity()),
-        had_infoz_(c.has_infoz()),
-        was_soo_(was_soo),
-        had_soo_slot_(had_soo_slot) {}
-
-  // Optimized for small groups version of `find_first_non_full`.
-  // Beneficial only right after calling `raw_hash_set::resize`.
-  // It is safe to call in case capacity is big or was not changed, but there
-  // will be no performance benefit.
-  // It has implicit assumption that `resize` will call
-  // `GrowSizeIntoSingleGroup*` in case `IsGrowingIntoSingleGroupApplicable`.
-  // Falls back to `find_first_non_full` in case of big groups, so it is
-  // safe to use after `rehash_and_grow_if_necessary`.
-  static FindInfo FindFirstNonFullAfterResize(const CommonFields& c,
-                                              size_t old_capacity,
-                                              size_t hash) {
-    if (!IsGrowingIntoSingleGroupApplicable(old_capacity, c.capacity())) {
-      return find_first_non_full(c, hash);
-    }
-    // Find a location for the new element non-deterministically.
-    // Note that any position is correct.
-    // It will located at `half_old_capacity` or one of the other
-    // empty slots with approximately 50% probability each.
-    size_t offset = probe(c, hash).offset();
-
-    // Note that we intentionally use unsigned int underflow.
-    if (offset - (old_capacity + 1) >= old_capacity) {
-      // Offset fall on kSentinel or into the mostly occupied first half.
-      offset = old_capacity / 2;
-    }
-    assert(IsEmpty(c.control()[offset]));
-    return FindInfo{offset, 0};
-  }
-
-  HeapOrSoo& old_heap_or_soo() { return old_heap_or_soo_; }
-  void* old_soo_data() { return old_heap_or_soo_.soo_data; }
-  ctrl_t* old_ctrl() const {
-    assert(!was_soo_);
-    return old_heap_or_soo_.heap.control;
-  }
-  void* old_slots() const {
-    assert(!was_soo_);
-    return old_heap_or_soo_.heap.slot_array.get();
-  }
-  size_t old_capacity() const { return old_capacity_; }
-
-  // Returns the index of the SOO slot when growing from SOO to non-SOO in a
-  // single group. See also InitControlBytesAfterSoo(). It's important to use
-  // index 1 so that when resizing from capacity 1 to 3, we can still have
-  // random iteration order between the first two inserted elements.
-  // I.e. it allows inserting the second element at either index 0 or 2.
-  static size_t SooSlotIndex() { return 1; }
-
-  // Allocates a backing array for the hashtable.
-  // Reads `capacity` and updates all other fields based on the result of
-  // the allocation.
-  //
-  // It also may do the folowing actions:
-  // 1. initialize control bytes
-  // 2. initialize slots
-  // 3. deallocate old slots.
-  //
-  // We are bundling a lot of functionality
-  // in one ABSL_ATTRIBUTE_NOINLINE function in order to minimize binary code
-  // duplication in raw_hash_set<>::resize.
-  //
-  // `c.capacity()` must be nonzero.
-  // POSTCONDITIONS:
-  //  1. CommonFields is initialized.
-  //
-  //  if IsGrowingIntoSingleGroupApplicable && TransferUsesMemcpy
-  //    Both control bytes and slots are fully initialized.
-  //    old_slots are deallocated.
-  //    infoz.RecordRehash is called.
-  //
-  //  if IsGrowingIntoSingleGroupApplicable && !TransferUsesMemcpy
-  //    Control bytes are fully initialized.
-  //    infoz.RecordRehash is called.
-  //    GrowSizeIntoSingleGroup must be called to finish slots initialization.
-  //
-  //  if !IsGrowingIntoSingleGroupApplicable
-  //    Control bytes are initialized to empty table via ResetCtrl.
-  //    raw_hash_set<>::resize must insert elements regularly.
-  //    infoz.RecordRehash is called if old_capacity == 0.
-  //
-  //  Returns IsGrowingIntoSingleGroupApplicable result to avoid recomputation.
-  template <typename Alloc, size_t SizeOfSlot, bool TransferUsesMemcpy,
-            size_t AlignOfSlot>
-  ABSL_ATTRIBUTE_NOINLINE bool InitializeSlots(CommonFields& c, Alloc alloc,
-                                               ctrl_t soo_slot_h2) {
-    assert(c.capacity());
-    // Folks with custom allocators often make unwarranted assumptions about the
-    // behavior of their classes vis-a-vis trivial destructability and what
-    // calls they will or won't make.  Avoid sampling for people with custom
-    // allocators to get us out of this mess.  This is not a hard guarantee but
-    // a workaround while we plan the exact guarantee we want to provide.
-    const size_t sample_size =
-        (std::is_same<Alloc, std::allocator<char>>::value &&
-         (was_soo_ || old_capacity_ == 0))
-            ? SizeOfSlot
-            : 0;
-    // TODO(b/289225379): when SOO is enabled, we should still sample on first
-    // insertion and if Sample is non-null, then we should force a heap
-    // allocation. Note that we'll also have to force capacity of 3 so that
-    // is_soo() still works.
-    HashtablezInfoHandle infoz =
-        sample_size > 0 ? Sample(sample_size) : c.infoz();
-
-    const bool has_infoz = infoz.IsSampled();
-    RawHashSetLayout layout(c.capacity(), AlignOfSlot, has_infoz);
-    char* mem = static_cast<char*>(Allocate<BackingArrayAlignment(AlignOfSlot)>(
-        &alloc, layout.alloc_size(SizeOfSlot)));
-    const GenerationType old_generation = c.generation();
-    c.set_generation_ptr(
-        reinterpret_cast<GenerationType*>(mem + layout.generation_offset()));
-    c.set_generation(NextGeneration(old_generation));
-    c.set_control(reinterpret_cast<ctrl_t*>(mem + layout.control_offset()));
-    c.set_slots(mem + layout.slot_offset());
-    ResetGrowthLeft(c);
-
-    const bool grow_single_group =
-        IsGrowingIntoSingleGroupApplicable(old_capacity_, layout.capacity());
-    if (was_soo_ && grow_single_group) {
-      InitControlBytesAfterSoo(c.control(), soo_slot_h2, layout.capacity());
-      if (TransferUsesMemcpy && had_soo_slot_) {
-        TransferSlotAfterSoo(c, SizeOfSlot);
-      }
-    } else if (old_capacity_ != 0 && grow_single_group) {
-      if (TransferUsesMemcpy) {
-        GrowSizeIntoSingleGroupTransferable(c, SizeOfSlot);
-        DeallocateOld<AlignOfSlot>(alloc, SizeOfSlot);
-      } else {
-        GrowIntoSingleGroupShuffleControlBytes(c.control(), layout.capacity());
-      }
-    } else {
-      ResetCtrl(c, SizeOfSlot);
-    }
-
-    c.set_has_infoz(has_infoz);
-    if (has_infoz) {
-      infoz.RecordStorageChanged(c.size(), layout.capacity());
-      if (was_soo_ || grow_single_group || old_capacity_ == 0) {
-        infoz.RecordRehash(0);
-      }
-      c.set_infoz(infoz);
-    }
-    return grow_single_group;
-  }
-
-  // Relocates slots into new single group consistent with
-  // GrowIntoSingleGroupShuffleControlBytes.
-  //
-  // PRECONDITIONS:
-  // 1. GrowIntoSingleGroupShuffleControlBytes was already called.
-  template <class PolicyTraits, class Alloc>
-  void GrowSizeIntoSingleGroup(CommonFields& c, Alloc& alloc_ref) {
-    assert(old_capacity_ < Group::kWidth / 2);
-    assert(IsGrowingIntoSingleGroupApplicable(old_capacity_, c.capacity()));
-    using slot_type = typename PolicyTraits::slot_type;
-    assert(is_single_group(c.capacity()));
-
-    auto* new_slots = reinterpret_cast<slot_type*>(c.slot_array());
-    auto* old_slots_ptr = reinterpret_cast<slot_type*>(old_slots());
-
-    size_t shuffle_bit = old_capacity_ / 2 + 1;
-    for (size_t i = 0; i < old_capacity_; ++i) {
-      if (IsFull(old_ctrl()[i])) {
-        size_t new_i = i ^ shuffle_bit;
-        SanitizerUnpoisonMemoryRegion(new_slots + new_i, sizeof(slot_type));
-        PolicyTraits::transfer(&alloc_ref, new_slots + new_i,
-                               old_slots_ptr + i);
-      }
-    }
-    PoisonSingleGroupEmptySlots(c, sizeof(slot_type));
-  }
-
-  // Deallocates old backing array.
-  template <size_t AlignOfSlot, class CharAlloc>
-  void DeallocateOld(CharAlloc alloc_ref, size_t slot_size) {
-    SanitizerUnpoisonMemoryRegion(old_slots(), slot_size * old_capacity_);
-    auto layout = RawHashSetLayout(old_capacity_, AlignOfSlot, had_infoz_);
-    Deallocate<BackingArrayAlignment(AlignOfSlot)>(
-        &alloc_ref, old_ctrl() - layout.control_offset(),
-        layout.alloc_size(slot_size));
-  }
-
- private:
-  // Returns true if `GrowSizeIntoSingleGroup` can be used for resizing.
-  static bool IsGrowingIntoSingleGroupApplicable(size_t old_capacity,
-                                                 size_t new_capacity) {
-    // NOTE that `old_capacity < new_capacity` in order to have
-    // `old_capacity < Group::kWidth / 2` to make faster copies of 8 bytes.
-    return is_single_group(new_capacity) && old_capacity < new_capacity;
-  }
-
-  // Relocates control bytes and slots into new single group for
-  // transferable objects.
-  // Must be called only if IsGrowingIntoSingleGroupApplicable returned true.
-  void GrowSizeIntoSingleGroupTransferable(CommonFields& c, size_t slot_size);
-
-  // If there was an SOO slot and slots are transferable, transfers the SOO slot
-  // into the new heap allocation. Must be called only if
-  // IsGrowingIntoSingleGroupApplicable returned true.
-  void TransferSlotAfterSoo(CommonFields& c, size_t slot_size);
-
-  // Shuffle control bits deterministically to the next capacity.
-  // Returns offset for newly added element with given hash.
-  //
-  // PRECONDITIONs:
-  // 1. new_ctrl is allocated for new_capacity,
-  //    but not initialized.
-  // 2. new_capacity is a single group.
-  //
-  // All elements are transferred into the first `old_capacity + 1` positions
-  // of the new_ctrl. Elements are rotated by `old_capacity_ / 2 + 1` positions
-  // in order to change an order and keep it non deterministic.
-  // Although rotation itself deterministic, position of the new added element
-  // will be based on `H1` and is not deterministic.
-  //
-  // Examples:
-  // S = kSentinel, E = kEmpty
-  //
-  // old_ctrl = SEEEEEEEE...
-  // new_ctrl = ESEEEEEEE...
-  //
-  // old_ctrl = 0SEEEEEEE...
-  // new_ctrl = E0ESE0EEE...
-  //
-  // old_ctrl = 012S012EEEEEEEEE...
-  // new_ctrl = 2E01EEES2E01EEE...
-  //
-  // old_ctrl = 0123456S0123456EEEEEEEEEEE...
-  // new_ctrl = 456E0123EEEEEES456E0123EEE...
-  void GrowIntoSingleGroupShuffleControlBytes(ctrl_t* new_ctrl,
-                                              size_t new_capacity) const;
-
-  // If the table was SOO, initializes new control bytes. `h2` is the control
-  // byte corresponding to the full slot. Must be called only if
-  // IsGrowingIntoSingleGroupApplicable returned true.
-  // Requires: `had_soo_slot_ || h2 == ctrl_t::kEmpty`.
-  void InitControlBytesAfterSoo(ctrl_t* new_ctrl, ctrl_t h2,
-                                size_t new_capacity);
-
-  // Shuffle trivially transferable slots in the way consistent with
-  // GrowIntoSingleGroupShuffleControlBytes.
-  //
-  // PRECONDITIONs:
-  // 1. old_capacity must be non-zero.
-  // 2. new_ctrl is fully initialized using
-  //    GrowIntoSingleGroupShuffleControlBytes.
-  // 3. new_slots is allocated and *not* poisoned.
-  //
-  // POSTCONDITIONS:
-  // 1. new_slots are transferred from old_slots_ consistent with
-  //    GrowIntoSingleGroupShuffleControlBytes.
-  // 2. Empty new_slots are *not* poisoned.
-  void GrowIntoSingleGroupShuffleTransferableSlots(void* new_slots,
-                                                   size_t slot_size) const;
-
-  // Poison empty slots that were transferred using the deterministic algorithm
-  // described above.
-  // PRECONDITIONs:
-  // 1. new_ctrl is fully initialized using
-  //    GrowIntoSingleGroupShuffleControlBytes.
-  // 2. new_slots is fully initialized consistent with
-  //    GrowIntoSingleGroupShuffleControlBytes.
-  void PoisonSingleGroupEmptySlots(CommonFields& c, size_t slot_size) const {
-    // poison non full items
-    for (size_t i = 0; i < c.capacity(); ++i) {
-      if (!IsFull(c.control()[i])) {
-        SanitizerPoisonMemoryRegion(SlotAddress(c.slot_array(), i, slot_size),
-                                    slot_size);
-      }
-    }
-  }
-
-  HeapOrSoo old_heap_or_soo_;
-  size_t old_capacity_;
-  bool had_infoz_;
-  bool was_soo_;
-  bool had_soo_slot_;
-};
-
-inline void PrepareInsertCommon(CommonFields& common) {
-  common.increment_size();
-  common.maybe_increment_generation_on_insert();
-}
-
-// Like prepare_insert, but for the case of inserting into a full SOO table.
-size_t PrepareInsertAfterSoo(size_t hash, size_t slot_size,
-                             CommonFields& common);
-
-// PolicyFunctions bundles together some information for a particular
-// raw_hash_set<T, ...> instantiation. This information is passed to
-// type-erased functions that want to do small amounts of type-specific
-// work.
-struct PolicyFunctions {
-  size_t slot_size;
-
-  // Returns the hash of the pointed-to slot.
-  size_t (*hash_slot)(const void* hash_fn, void* slot);
-
-  // Transfer the contents of src_slot to dst_slot.
-  void (*transfer)(void* set, void* dst_slot, void* src_slot);
-
-  // Deallocate the backing store from common.
-  void (*dealloc)(CommonFields& common, const PolicyFunctions& policy);
-};
-
-// ClearBackingArray clears the backing array, either modifying it in place,
-// or creating a new one based on the value of "reuse".
-// REQUIRES: c.capacity > 0
-void ClearBackingArray(CommonFields& c, const PolicyFunctions& policy,
-                       bool reuse, bool soo_enabled);
-
-// Type-erased version of raw_hash_set::erase_meta_only.
-void EraseMetaOnly(CommonFields& c, size_t index, size_t slot_size);
-
-// Function to place in PolicyFunctions::dealloc for raw_hash_sets
-// that are using std::allocator. This allows us to share the same
-// function body for raw_hash_set instantiations that have the
-// same slot alignment.
-template <size_t AlignOfSlot>
-ABSL_ATTRIBUTE_NOINLINE void DeallocateStandard(CommonFields& common,
-                                                const PolicyFunctions& policy) {
-  // Unpoison before returning the memory to the allocator.
-  SanitizerUnpoisonMemoryRegion(common.slot_array(),
-                                policy.slot_size * common.capacity());
-
-  std::allocator<char> alloc;
-  common.infoz().Unregister();
-  Deallocate<BackingArrayAlignment(AlignOfSlot)>(
-      &alloc, common.backing_array_start(),
-      common.alloc_size(policy.slot_size, AlignOfSlot));
-}
-
-// For trivially relocatable types we use memcpy directly. This allows us to
-// share the same function body for raw_hash_set instantiations that have the
-// same slot size as long as they are relocatable.
-template <size_t SizeOfSlot>
-ABSL_ATTRIBUTE_NOINLINE void TransferRelocatable(void*, void* dst, void* src) {
-  memcpy(dst, src, SizeOfSlot);
-}
-
-// Type-erased version of raw_hash_set::drop_deletes_without_resize.
-void DropDeletesWithoutResize(CommonFields& common, const void* hash_fn,
-                              const PolicyFunctions& policy, void* tmp_space);
-
-// A SwissTable.
-//
-// Policy: a policy defines how to perform different operations on
-// the slots of the hashtable (see hash_policy_traits.h for the full interface
-// of policy).
-//
-// Hash: a (possibly polymorphic) functor that hashes keys of the hashtable. The
-// functor should accept a key and return size_t as hash. For best performance
-// it is important that the hash function provides high entropy across all bits
-// of the hash.
-//
-// Eq: a (possibly polymorphic) functor that compares two keys for equality. It
-// should accept two (of possibly different type) keys and return a bool: true
-// if they are equal, false if they are not. If two keys compare equal, then
-// their hash values as defined by Hash MUST be equal.
-//
-// Allocator: an Allocator
-// [https://en.cppreference.com/w/cpp/named_req/Allocator] with which
-// the storage of the hashtable will be allocated and the elements will be
-// constructed and destroyed.
-template <class Policy, class Hash, class Eq, class Alloc>
-class raw_hash_set {
-  using PolicyTraits = hash_policy_traits<Policy>;
-  using KeyArgImpl =
-      KeyArg<IsTransparent<Eq>::value && IsTransparent<Hash>::value>;
-
- public:
-  using init_type = typename PolicyTraits::init_type;
-  using key_type = typename PolicyTraits::key_type;
-  // TODO(sbenza): Hide slot_type as it is an implementation detail. Needs user
-  // code fixes!
-  using slot_type = typename PolicyTraits::slot_type;
-  using allocator_type = Alloc;
-  using size_type = size_t;
-  using difference_type = ptrdiff_t;
-  using hasher = Hash;
-  using key_equal = Eq;
-  using policy_type = Policy;
-  using value_type = typename PolicyTraits::value_type;
-  using reference = value_type&;
-  using const_reference = const value_type&;
-  using pointer = typename absl::allocator_traits<
-      allocator_type>::template rebind_traits<value_type>::pointer;
-  using const_pointer = typename absl::allocator_traits<
-      allocator_type>::template rebind_traits<value_type>::const_pointer;
-
-  // Alias used for heterogeneous lookup functions.
-  // `key_arg<K>` evaluates to `K` when the functors are transparent and to
-  // `key_type` otherwise. It permits template argument deduction on `K` for the
-  // transparent case.
-  template <class K>
-  using key_arg = typename KeyArgImpl::template type<K, key_type>;
-
- private:
-  // TODO(b/289225379): we could add extra SOO space inside raw_hash_set
-  // after CommonFields to allow inlining larger slot_types (e.g. std::string),
-  // but it's a bit complicated if we want to support incomplete mapped_type in
-  // flat_hash_map. We could potentially do this for flat_hash_set and for an
-  // allowlist of `mapped_type`s of flat_hash_map that includes e.g. arithmetic
-  // types, strings, cords, and pairs/tuples of allowlisted types.
-  constexpr static bool SooEnabled() {
-    return PolicyTraits::soo_enabled() &&
-           sizeof(slot_type) <= sizeof(HeapOrSoo) &&
-           alignof(slot_type) <= alignof(HeapOrSoo);
-  }
-
-  // Whether `size` fits in the SOO capacity of this table.
-  bool fits_in_soo(size_t size) const {
-    return SooEnabled() && size <= SooCapacity();
-  }
-  // Whether this table is in SOO mode or non-SOO mode.
-  bool is_soo() const { return fits_in_soo(capacity()); }
-  bool is_full_soo() const { return is_soo() && !empty(); }
-
-  // Give an early error when key_type is not hashable/eq.
-  auto KeyTypeCanBeHashed(const Hash& h, const key_type& k) -> decltype(h(k));
-  auto KeyTypeCanBeEq(const Eq& eq, const key_type& k) -> decltype(eq(k, k));
-
-  using AllocTraits = absl::allocator_traits<allocator_type>;
-  using SlotAlloc = typename absl::allocator_traits<
-      allocator_type>::template rebind_alloc<slot_type>;
-  // People are often sloppy with the exact type of their allocator (sometimes
-  // it has an extra const or is missing the pair, but rebinds made it work
-  // anyway).
-  using CharAlloc =
-      typename absl::allocator_traits<Alloc>::template rebind_alloc<char>;
-  using SlotAllocTraits = typename absl::allocator_traits<
-      allocator_type>::template rebind_traits<slot_type>;
-
-  static_assert(std::is_lvalue_reference<reference>::value,
-                "Policy::element() must return a reference");
-
-  template <typename T>
-  struct SameAsElementReference
-      : std::is_same<typename std::remove_cv<
-                         typename std::remove_reference<reference>::type>::type,
-                     typename std::remove_cv<
-                         typename std::remove_reference<T>::type>::type> {};
-
-  // An enabler for insert(T&&): T must be convertible to init_type or be the
-  // same as [cv] value_type [ref].
-  // Note: we separate SameAsElementReference into its own type to avoid using
-  // reference unless we need to. MSVC doesn't seem to like it in some
-  // cases.
-  template <class T>
-  using RequiresInsertable = typename std::enable_if<
-      absl::disjunction<std::is_convertible<T, init_type>,
-                        SameAsElementReference<T>>::value,
-      int>::type;
-
-  // RequiresNotInit is a workaround for gcc prior to 7.1.
-  // See https://godbolt.org/g/Y4xsUh.
-  template <class T>
-  using RequiresNotInit =
-      typename std::enable_if<!std::is_same<T, init_type>::value, int>::type;
-
-  template <class... Ts>
-  using IsDecomposable = IsDecomposable<void, PolicyTraits, Hash, Eq, Ts...>;
-
- public:
-  static_assert(std::is_same<pointer, value_type*>::value,
-                "Allocators with custom pointer types are not supported");
-  static_assert(std::is_same<const_pointer, const value_type*>::value,
-                "Allocators with custom pointer types are not supported");
-
-  class iterator : private HashSetIteratorGenerationInfo {
-    friend class raw_hash_set;
-
-   public:
-    using iterator_category = std::forward_iterator_tag;
-    using value_type = typename raw_hash_set::value_type;
-    using reference =
-        absl::conditional_t<PolicyTraits::constant_iterators::value,
-                            const value_type&, value_type&>;
-    using pointer = absl::remove_reference_t<reference>*;
-    using difference_type = typename raw_hash_set::difference_type;
-
-    iterator() {}
-
-    // PRECONDITION: not an end() iterator.
-    reference operator*() const {
-      AssertIsFull(ctrl_, generation(), generation_ptr(), "operator*()");
-      return unchecked_deref();
-    }
-
-    // PRECONDITION: not an end() iterator.
-    pointer operator->() const {
-      AssertIsFull(ctrl_, generation(), generation_ptr(), "operator->");
-      return &operator*();
-    }
-
-    // PRECONDITION: not an end() iterator.
-    iterator& operator++() {
-      AssertIsFull(ctrl_, generation(), generation_ptr(), "operator++");
-      ++ctrl_;
-      ++slot_;
-      skip_empty_or_deleted();
-      if (ABSL_PREDICT_FALSE(*ctrl_ == ctrl_t::kSentinel)) ctrl_ = nullptr;
-      return *this;
-    }
-    // PRECONDITION: not an end() iterator.
-    iterator operator++(int) {
-      auto tmp = *this;
-      ++*this;
-      return tmp;
-    }
-
-    friend bool operator==(const iterator& a, const iterator& b) {
-      AssertIsValidForComparison(a.ctrl_, a.generation(), a.generation_ptr());
-      AssertIsValidForComparison(b.ctrl_, b.generation(), b.generation_ptr());
-      AssertSameContainer(a.ctrl_, b.ctrl_, a.slot_, b.slot_,
-                          a.generation_ptr(), b.generation_ptr());
-      return a.ctrl_ == b.ctrl_;
-    }
-    friend bool operator!=(const iterator& a, const iterator& b) {
-      return !(a == b);
-    }
-
-   private:
-    iterator(ctrl_t* ctrl, slot_type* slot,
-             const GenerationType* generation_ptr)
-        : HashSetIteratorGenerationInfo(generation_ptr),
-          ctrl_(ctrl),
-          slot_(slot) {
-      // This assumption helps the compiler know that any non-end iterator is
-      // not equal to any end iterator.
-      ABSL_ASSUME(ctrl != nullptr);
-    }
-    // This constructor is used in begin() to avoid an MSan
-    // use-of-uninitialized-value error. Delegating from this constructor to
-    // the previous one doesn't avoid the error.
-    iterator(ctrl_t* ctrl, MaybeInitializedPtr slot,
-             const GenerationType* generation_ptr)
-        : HashSetIteratorGenerationInfo(generation_ptr),
-          ctrl_(ctrl),
-          slot_(to_slot(slot.get())) {
-      // This assumption helps the compiler know that any non-end iterator is
-      // not equal to any end iterator.
-      ABSL_ASSUME(ctrl != nullptr);
-    }
-    // For end() iterators.
-    explicit iterator(const GenerationType* generation_ptr)
-        : HashSetIteratorGenerationInfo(generation_ptr), ctrl_(nullptr) {}
-
-    // Fixes up `ctrl_` to point to a full or sentinel by advancing `ctrl_` and
-    // `slot_` until they reach one.
-    void skip_empty_or_deleted() {
-      while (IsEmptyOrDeleted(*ctrl_)) {
-        uint32_t shift =
-            GroupFullEmptyOrDeleted{ctrl_}.CountLeadingEmptyOrDeleted();
-        ctrl_ += shift;
-        slot_ += shift;
-      }
-    }
-
-    ctrl_t* control() const { return ctrl_; }
-    slot_type* slot() const { return slot_; }
-
-    // We use EmptyGroup() for default-constructed iterators so that they can
-    // be distinguished from end iterators, which have nullptr ctrl_.
-    ctrl_t* ctrl_ = EmptyGroup();
-    // To avoid uninitialized member warnings, put slot_ in an anonymous union.
-    // The member is not initialized on singleton and end iterators.
-    union {
-      slot_type* slot_;
-    };
-
-    // An equality check which skips ABSL Hardening iterator invalidation
-    // checks.
-    // Should be used when the lifetimes of the iterators are well-enough
-    // understood to prove that they cannot be invalid.
-    bool unchecked_equals(const iterator& b) { return ctrl_ == b.control(); }
-
-    // Dereferences the iterator without ABSL Hardening iterator invalidation
-    // checks.
-    reference unchecked_deref() const { return PolicyTraits::element(slot_); }
-  };
-
-  class const_iterator {
-    friend class raw_hash_set;
-    template <class Container, typename Enabler>
-    friend struct absl::container_internal::hashtable_debug_internal::
-        HashtableDebugAccess;
-
-   public:
-    using iterator_category = typename iterator::iterator_category;
-    using value_type = typename raw_hash_set::value_type;
-    using reference = typename raw_hash_set::const_reference;
-    using pointer = typename raw_hash_set::const_pointer;
-    using difference_type = typename raw_hash_set::difference_type;
-
-    const_iterator() = default;
-    // Implicit construction from iterator.
-    const_iterator(iterator i) : inner_(std::move(i)) {}  // NOLINT
-
-    reference operator*() const { return *inner_; }
-    pointer operator->() const { return inner_.operator->(); }
-
-    const_iterator& operator++() {
-      ++inner_;
-      return *this;
-    }
-    const_iterator operator++(int) { return inner_++; }
-
-    friend bool operator==(const const_iterator& a, const const_iterator& b) {
-      return a.inner_ == b.inner_;
-    }
-    friend bool operator!=(const const_iterator& a, const const_iterator& b) {
-      return !(a == b);
-    }
-
-   private:
-    const_iterator(const ctrl_t* ctrl, const slot_type* slot,
-                   const GenerationType* gen)
-        : inner_(const_cast<ctrl_t*>(ctrl), const_cast<slot_type*>(slot), gen) {
-    }
-    ctrl_t* control() const { return inner_.control(); }
-    slot_type* slot() const { return inner_.slot(); }
-
-    iterator inner_;
-
-    bool unchecked_equals(const const_iterator& b) {
-      return inner_.unchecked_equals(b.inner_);
-    }
-  };
-
-  using node_type = node_handle<Policy, hash_policy_traits<Policy>, Alloc>;
-  using insert_return_type = InsertReturnType<iterator, node_type>;
-
-  // Note: can't use `= default` due to non-default noexcept (causes
-  // problems for some compilers). NOLINTNEXTLINE
-  raw_hash_set() noexcept(
-      std::is_nothrow_default_constructible<hasher>::value &&
-      std::is_nothrow_default_constructible<key_equal>::value &&
-      std::is_nothrow_default_constructible<allocator_type>::value) {}
-
-  ABSL_ATTRIBUTE_NOINLINE explicit raw_hash_set(
-      size_t bucket_count, const hasher& hash = hasher(),
-      const key_equal& eq = key_equal(),
-      const allocator_type& alloc = allocator_type())
-      : settings_(CommonFields::CreateDefault<SooEnabled()>(), hash, eq,
-                  alloc) {
-    if (bucket_count > (SooEnabled() ? SooCapacity() : 0)) {
-      resize(NormalizeCapacity(bucket_count));
-    }
-  }
-
-  raw_hash_set(size_t bucket_count, const hasher& hash,
-               const allocator_type& alloc)
-      : raw_hash_set(bucket_count, hash, key_equal(), alloc) {}
-
-  raw_hash_set(size_t bucket_count, const allocator_type& alloc)
-      : raw_hash_set(bucket_count, hasher(), key_equal(), alloc) {}
-
-  explicit raw_hash_set(const allocator_type& alloc)
-      : raw_hash_set(0, hasher(), key_equal(), alloc) {}
-
-  template <class InputIter>
-  raw_hash_set(InputIter first, InputIter last, size_t bucket_count = 0,
-               const hasher& hash = hasher(), const key_equal& eq = key_equal(),
-               const allocator_type& alloc = allocator_type())
-      : raw_hash_set(SelectBucketCountForIterRange(first, last, bucket_count),
-                     hash, eq, alloc) {
-    insert(first, last);
-  }
-
-  template <class InputIter>
-  raw_hash_set(InputIter first, InputIter last, size_t bucket_count,
-               const hasher& hash, const allocator_type& alloc)
-      : raw_hash_set(first, last, bucket_count, hash, key_equal(), alloc) {}
-
-  template <class InputIter>
-  raw_hash_set(InputIter first, InputIter last, size_t bucket_count,
-               const allocator_type& alloc)
-      : raw_hash_set(first, last, bucket_count, hasher(), key_equal(), alloc) {}
-
-  template <class InputIter>
-  raw_hash_set(InputIter first, InputIter last, const allocator_type& alloc)
-      : raw_hash_set(first, last, 0, hasher(), key_equal(), alloc) {}
-
-  // Instead of accepting std::initializer_list<value_type> as the first
-  // argument like std::unordered_set<value_type> does, we have two overloads
-  // that accept std::initializer_list<T> and std::initializer_list<init_type>.
-  // This is advantageous for performance.
-  //
-  //   // Turns {"abc", "def"} into std::initializer_list<std::string>, then
-  //   // copies the strings into the set.
-  //   std::unordered_set<std::string> s = {"abc", "def"};
-  //
-  //   // Turns {"abc", "def"} into std::initializer_list<const char*>, then
-  //   // copies the strings into the set.
-  //   absl::flat_hash_set<std::string> s = {"abc", "def"};
-  //
-  // The same trick is used in insert().
-  //
-  // The enabler is necessary to prevent this constructor from triggering where
-  // the copy constructor is meant to be called.
-  //
-  //   absl::flat_hash_set<int> a, b{a};
-  //
-  // RequiresNotInit<T> is a workaround for gcc prior to 7.1.
-  template <class T, RequiresNotInit<T> = 0, RequiresInsertable<T> = 0>
-  raw_hash_set(std::initializer_list<T> init, size_t bucket_count = 0,
-               const hasher& hash = hasher(), const key_equal& eq = key_equal(),
-               const allocator_type& alloc = allocator_type())
-      : raw_hash_set(init.begin(), init.end(), bucket_count, hash, eq, alloc) {}
-
-  raw_hash_set(std::initializer_list<init_type> init, size_t bucket_count = 0,
-               const hasher& hash = hasher(), const key_equal& eq = key_equal(),
-               const allocator_type& alloc = allocator_type())
-      : raw_hash_set(init.begin(), init.end(), bucket_count, hash, eq, alloc) {}
-
-  template <class T, RequiresNotInit<T> = 0, RequiresInsertable<T> = 0>
-  raw_hash_set(std::initializer_list<T> init, size_t bucket_count,
-               const hasher& hash, const allocator_type& alloc)
-      : raw_hash_set(init, bucket_count, hash, key_equal(), alloc) {}
-
-  raw_hash_set(std::initializer_list<init_type> init, size_t bucket_count,
-               const hasher& hash, const allocator_type& alloc)
-      : raw_hash_set(init, bucket_count, hash, key_equal(), alloc) {}
-
-  template <class T, RequiresNotInit<T> = 0, RequiresInsertable<T> = 0>
-  raw_hash_set(std::initializer_list<T> init, size_t bucket_count,
-               const allocator_type& alloc)
-      : raw_hash_set(init, bucket_count, hasher(), key_equal(), alloc) {}
-
-  raw_hash_set(std::initializer_list<init_type> init, size_t bucket_count,
-               const allocator_type& alloc)
-      : raw_hash_set(init, bucket_count, hasher(), key_equal(), alloc) {}
-
-  template <class T, RequiresNotInit<T> = 0, RequiresInsertable<T> = 0>
-  raw_hash_set(std::initializer_list<T> init, const allocator_type& alloc)
-      : raw_hash_set(init, 0, hasher(), key_equal(), alloc) {}
-
-  raw_hash_set(std::initializer_list<init_type> init,
-               const allocator_type& alloc)
-      : raw_hash_set(init, 0, hasher(), key_equal(), alloc) {}
-
-  raw_hash_set(const raw_hash_set& that)
-      : raw_hash_set(that, AllocTraits::select_on_container_copy_construction(
-                               that.alloc_ref())) {}
-
-  raw_hash_set(const raw_hash_set& that, const allocator_type& a)
-      : raw_hash_set(GrowthToLowerboundCapacity(that.size()), that.hash_ref(),
-                     that.eq_ref(), a) {
-    const size_t size = that.size();
-    if (size == 0) {
-      return;
-    }
-    // We don't use `that.is_soo()` here because `that` can have non-SOO
-    // capacity but have a size that fits into SOO capacity.
-    if (fits_in_soo(size)) {
-      assert(size == 1);
-      common().set_full_soo();
-      emplace_at(soo_iterator(), *that.begin());
-      return;
-    }
-    assert(!that.is_soo());
-    const size_t cap = capacity();
-    // Note about single group tables:
-    // 1. It is correct to have any order of elements.
-    // 2. Order has to be non deterministic.
-    // 3. We are assigning elements with arbitrary `shift` starting from
-    //    `capacity + shift` position.
-    // 4. `shift` must be coprime with `capacity + 1` in order to be able to use
-    //     modular arithmetic to traverse all positions, instead if cycling
-    //     through a subset of positions. Odd numbers are coprime with any
-    //     `capacity + 1` (2^N).
-    size_t offset = cap;
-    const size_t shift =
-        is_single_group(cap) ? (PerTableSalt(control()) | 1) : 0;
-    IterateOverFullSlots(
-        that.common(), that.slot_array(),
-        [&](const ctrl_t* that_ctrl,
-            slot_type* that_slot) ABSL_ATTRIBUTE_ALWAYS_INLINE {
-          if (shift == 0) {
-            // Big tables case. Position must be searched via probing.
-            // The table is guaranteed to be empty, so we can do faster than
-            // a full `insert`.
-            const size_t hash = PolicyTraits::apply(
-                HashElement{hash_ref()}, PolicyTraits::element(that_slot));
-            FindInfo target = find_first_non_full_outofline(common(), hash);
-            infoz().RecordInsert(hash, target.probe_length);
-            offset = target.offset;
-          } else {
-            // Small tables case. Next position is computed via shift.
-            offset = (offset + shift) & cap;
-          }
-          const h2_t h2 = static_cast<h2_t>(*that_ctrl);
-          assert(  // We rely that hash is not changed for small tables.
-              H2(PolicyTraits::apply(HashElement{hash_ref()},
-                                     PolicyTraits::element(that_slot))) == h2 &&
-              "hash function value changed unexpectedly during the copy");
-          SetCtrl(common(), offset, h2, sizeof(slot_type));
-          emplace_at(iterator_at(offset), PolicyTraits::element(that_slot));
-          common().maybe_increment_generation_on_insert();
-        });
-    if (shift != 0) {
-      // On small table copy we do not record individual inserts.
-      // RecordInsert requires hash, but it is unknown for small tables.
-      infoz().RecordStorageChanged(size, cap);
-    }
-    common().set_size(size);
-    set_growth_left(growth_left() - size);
-  }
-
-  ABSL_ATTRIBUTE_NOINLINE raw_hash_set(raw_hash_set&& that) noexcept(
-      std::is_nothrow_copy_constructible<hasher>::value &&
-      std::is_nothrow_copy_constructible<key_equal>::value &&
-      std::is_nothrow_copy_constructible<allocator_type>::value)
-      :  // Hash, equality and allocator are copied instead of moved because
-         // `that` must be left valid. If Hash is std::function<Key>, moving it
-         // would create a nullptr functor that cannot be called.
-         // TODO(b/296061262): move instead of copying hash/eq/alloc.
-         // Note: we avoid using exchange for better generated code.
-        settings_(PolicyTraits::transfer_uses_memcpy() || !that.is_full_soo()
-                      ? std::move(that.common())
-                      : CommonFields{full_soo_tag_t{}},
-                  that.hash_ref(), that.eq_ref(), that.alloc_ref()) {
-    if (!PolicyTraits::transfer_uses_memcpy() && that.is_full_soo()) {
-      transfer(soo_slot(), that.soo_slot());
-    }
-    that.common() = CommonFields::CreateDefault<SooEnabled()>();
-    maybe_increment_generation_or_rehash_on_move();
-  }
-
-  raw_hash_set(raw_hash_set&& that, const allocator_type& a)
-      : settings_(CommonFields::CreateDefault<SooEnabled()>(), that.hash_ref(),
-                  that.eq_ref(), a) {
-    if (a == that.alloc_ref()) {
-      swap_common(that);
-      maybe_increment_generation_or_rehash_on_move();
-    } else {
-      move_elements_allocs_unequal(std::move(that));
-    }
-  }
-
-  raw_hash_set& operator=(const raw_hash_set& that) {
-    if (ABSL_PREDICT_FALSE(this == &that)) return *this;
-    constexpr bool propagate_alloc =
-        AllocTraits::propagate_on_container_copy_assignment::value;
-    // TODO(ezb): maybe avoid allocating a new backing array if this->capacity()
-    // is an exact match for that.size(). If this->capacity() is too big, then
-    // it would make iteration very slow to reuse the allocation. Maybe we can
-    // do the same heuristic as clear() and reuse if it's small enough.
-    raw_hash_set tmp(that, propagate_alloc ? that.alloc_ref() : alloc_ref());
-    // NOLINTNEXTLINE: not returning *this for performance.
-    return assign_impl<propagate_alloc>(std::move(tmp));
-  }
-
-  raw_hash_set& operator=(raw_hash_set&& that) noexcept(
-      absl::allocator_traits<allocator_type>::is_always_equal::value &&
-      std::is_nothrow_move_assignable<hasher>::value &&
-      std::is_nothrow_move_assignable<key_equal>::value) {
-    // TODO(sbenza): We should only use the operations from the noexcept clause
-    // to make sure we actually adhere to that contract.
-    // NOLINTNEXTLINE: not returning *this for performance.
-    return move_assign(
-        std::move(that),
-        typename AllocTraits::propagate_on_container_move_assignment());
-  }
-
-  ~raw_hash_set() { destructor_impl(); }
-
-  iterator begin() ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    if (ABSL_PREDICT_FALSE(empty())) return end();
-    if (is_soo()) return soo_iterator();
-    iterator it = {control(), common().slots_union(),
-                   common().generation_ptr()};
-    it.skip_empty_or_deleted();
-    assert(IsFull(*it.control()));
-    return it;
-  }
-  iterator end() ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return iterator(common().generation_ptr());
-  }
-
-  const_iterator begin() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return const_cast<raw_hash_set*>(this)->begin();
-  }
-  const_iterator end() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return iterator(common().generation_ptr());
-  }
-  const_iterator cbegin() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return begin();
-  }
-  const_iterator cend() const ABSL_ATTRIBUTE_LIFETIME_BOUND { return end(); }
-
-  bool empty() const { return !size(); }
-  size_t size() const { return common().size(); }
-  size_t capacity() const {
-    const size_t cap = common().capacity();
-    // Use local variables because compiler complains about using functions in
-    // assume.
-    ABSL_ATTRIBUTE_UNUSED static constexpr bool kEnabled = SooEnabled();
-    ABSL_ATTRIBUTE_UNUSED static constexpr size_t kCapacity = SooCapacity();
-    ABSL_ASSUME(!kEnabled || cap >= kCapacity);
-    return cap;
-  }
-  size_t max_size() const { return (std::numeric_limits<size_t>::max)(); }
-
-  ABSL_ATTRIBUTE_REINITIALIZES void clear() {
-    // Iterating over this container is O(bucket_count()). When bucket_count()
-    // is much greater than size(), iteration becomes prohibitively expensive.
-    // For clear() it is more important to reuse the allocated array when the
-    // container is small because allocation takes comparatively long time
-    // compared to destruction of the elements of the container. So we pick the
-    // largest bucket_count() threshold for which iteration is still fast and
-    // past that we simply deallocate the array.
-    const size_t cap = capacity();
-    if (cap == 0) {
-      // Already guaranteed to be empty; so nothing to do.
-    } else if (is_soo()) {
-      if (!empty()) destroy(soo_slot());
-      common().set_empty_soo();
-    } else {
-      destroy_slots();
-      ClearBackingArray(common(), GetPolicyFunctions(), /*reuse=*/cap < 128,
-                        SooEnabled());
-    }
-    common().set_reserved_growth(0);
-    common().set_reservation_size(0);
-  }
-
-  // This overload kicks in when the argument is an rvalue of insertable and
-  // decomposable type other than init_type.
-  //
-  //   flat_hash_map<std::string, int> m;
-  //   m.insert(std::make_pair("abc", 42));
-  // TODO(cheshire): A type alias T2 is introduced as a workaround for the nvcc
-  // bug.
-  template <class T, RequiresInsertable<T> = 0, class T2 = T,
-            typename std::enable_if<IsDecomposable<T2>::value, int>::type = 0,
-            T* = nullptr>
-  std::pair<iterator, bool> insert(T&& value) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return emplace(std::forward<T>(value));
-  }
-
-  // This overload kicks in when the argument is a bitfield or an lvalue of
-  // insertable and decomposable type.
-  //
-  //   union { int n : 1; };
-  //   flat_hash_set<int> s;
-  //   s.insert(n);
-  //
-  //   flat_hash_set<std::string> s;
-  //   const char* p = "hello";
-  //   s.insert(p);
-  //
-  template <
-      class T, RequiresInsertable<const T&> = 0,
-      typename std::enable_if<IsDecomposable<const T&>::value, int>::type = 0>
-  std::pair<iterator, bool> insert(const T& value)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return emplace(value);
-  }
-
-  // This overload kicks in when the argument is an rvalue of init_type. Its
-  // purpose is to handle brace-init-list arguments.
-  //
-  //   flat_hash_map<std::string, int> s;
-  //   s.insert({"abc", 42});
-  std::pair<iterator, bool> insert(init_type&& value)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return emplace(std::move(value));
-  }
-
-  // TODO(cheshire): A type alias T2 is introduced as a workaround for the nvcc
-  // bug.
-  template <class T, RequiresInsertable<T> = 0, class T2 = T,
-            typename std::enable_if<IsDecomposable<T2>::value, int>::type = 0,
-            T* = nullptr>
-  iterator insert(const_iterator, T&& value) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert(std::forward<T>(value)).first;
-  }
-
-  template <
-      class T, RequiresInsertable<const T&> = 0,
-      typename std::enable_if<IsDecomposable<const T&>::value, int>::type = 0>
-  iterator insert(const_iterator,
-                  const T& value) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert(value).first;
-  }
-
-  iterator insert(const_iterator,
-                  init_type&& value) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return insert(std::move(value)).first;
-  }
-
-  template <class InputIt>
-  void insert(InputIt first, InputIt last) {
-    for (; first != last; ++first) emplace(*first);
-  }
-
-  template <class T, RequiresNotInit<T> = 0, RequiresInsertable<const T&> = 0>
-  void insert(std::initializer_list<T> ilist) {
-    insert(ilist.begin(), ilist.end());
-  }
-
-  void insert(std::initializer_list<init_type> ilist) {
-    insert(ilist.begin(), ilist.end());
-  }
-
-  insert_return_type insert(node_type&& node) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    if (!node) return {end(), false, node_type()};
-    const auto& elem = PolicyTraits::element(CommonAccess::GetSlot(node));
-    auto res = PolicyTraits::apply(
-        InsertSlot<false>{*this, std::move(*CommonAccess::GetSlot(node))},
-        elem);
-    if (res.second) {
-      CommonAccess::Reset(&node);
-      return {res.first, true, node_type()};
-    } else {
-      return {res.first, false, std::move(node)};
-    }
-  }
-
-  iterator insert(const_iterator,
-                  node_type&& node) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    auto res = insert(std::move(node));
-    node = std::move(res.node);
-    return res.position;
-  }
-
-  // This overload kicks in if we can deduce the key from args. This enables us
-  // to avoid constructing value_type if an entry with the same key already
-  // exists.
-  //
-  // For example:
-  //
-  //   flat_hash_map<std::string, std::string> m = {{"abc", "def"}};
-  //   // Creates no std::string copies and makes no heap allocations.
-  //   m.emplace("abc", "xyz");
-  template <class... Args, typename std::enable_if<
-                               IsDecomposable<Args...>::value, int>::type = 0>
-  std::pair<iterator, bool> emplace(Args&&... args)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return PolicyTraits::apply(EmplaceDecomposable{*this},
-                               std::forward<Args>(args)...);
-  }
-
-  // This overload kicks in if we cannot deduce the key from args. It constructs
-  // value_type unconditionally and then either moves it into the table or
-  // destroys.
-  template <class... Args, typename std::enable_if<
-                               !IsDecomposable<Args...>::value, int>::type = 0>
-  std::pair<iterator, bool> emplace(Args&&... args)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    alignas(slot_type) unsigned char raw[sizeof(slot_type)];
-    slot_type* slot = to_slot(&raw);
-
-    construct(slot, std::forward<Args>(args)...);
-    const auto& elem = PolicyTraits::element(slot);
-    return PolicyTraits::apply(InsertSlot<true>{*this, std::move(*slot)}, elem);
-  }
-
-  template <class... Args>
-  iterator emplace_hint(const_iterator,
-                        Args&&... args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return emplace(std::forward<Args>(args)...).first;
-  }
-
-  // Extension API: support for lazy emplace.
-  //
-  // Looks up key in the table. If found, returns the iterator to the element.
-  // Otherwise calls `f` with one argument of type `raw_hash_set::constructor`,
-  // and returns an iterator to the new element.
-  //
-  // `f` must abide by several restrictions:
-  //  - it MUST call `raw_hash_set::constructor` with arguments as if a
-  //    `raw_hash_set::value_type` is constructed,
-  //  - it MUST NOT access the container before the call to
-  //    `raw_hash_set::constructor`, and
-  //  - it MUST NOT erase the lazily emplaced element.
-  // Doing any of these is undefined behavior.
-  //
-  // For example:
-  //
-  //   std::unordered_set<ArenaString> s;
-  //   // Makes ArenaStr even if "abc" is in the map.
-  //   s.insert(ArenaString(&arena, "abc"));
-  //
-  //   flat_hash_set<ArenaStr> s;
-  //   // Makes ArenaStr only if "abc" is not in the map.
-  //   s.lazy_emplace("abc", [&](const constructor& ctor) {
-  //     ctor(&arena, "abc");
-  //   });
-  //
-  // WARNING: This API is currently experimental. If there is a way to implement
-  // the same thing with the rest of the API, prefer that.
-  class constructor {
-    friend class raw_hash_set;
-
-   public:
-    template <class... Args>
-    void operator()(Args&&... args) const {
-      assert(*slot_);
-      PolicyTraits::construct(alloc_, *slot_, std::forward<Args>(args)...);
-      *slot_ = nullptr;
-    }
-
-   private:
-    constructor(allocator_type* a, slot_type** slot) : alloc_(a), slot_(slot) {}
-
-    allocator_type* alloc_;
-    slot_type** slot_;
-  };
-
-  template <class K = key_type, class F>
-  iterator lazy_emplace(const key_arg<K>& key,
-                        F&& f) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    auto res = find_or_prepare_insert(key);
-    if (res.second) {
-      slot_type* slot = res.first.slot();
-      std::forward<F>(f)(constructor(&alloc_ref(), &slot));
-      assert(!slot);
-    }
-    return res.first;
-  }
-
-  // Extension API: support for heterogeneous keys.
-  //
-  //   std::unordered_set<std::string> s;
-  //   // Turns "abc" into std::string.
-  //   s.erase("abc");
-  //
-  //   flat_hash_set<std::string> s;
-  //   // Uses "abc" directly without copying it into std::string.
-  //   s.erase("abc");
-  template <class K = key_type>
-  size_type erase(const key_arg<K>& key) {
-    auto it = find(key);
-    if (it == end()) return 0;
-    erase(it);
-    return 1;
-  }
-
-  // Erases the element pointed to by `it`.  Unlike `std::unordered_set::erase`,
-  // this method returns void to reduce algorithmic complexity to O(1).  The
-  // iterator is invalidated, so any increment should be done before calling
-  // erase.  In order to erase while iterating across a map, use the following
-  // idiom (which also works for standard containers):
-  //
-  // for (auto it = m.begin(), end = m.end(); it != end;) {
-  //   // `erase()` will invalidate `it`, so advance `it` first.
-  //   auto copy_it = it++;
-  //   if (<pred>) {
-  //     m.erase(copy_it);
-  //   }
-  // }
-  void erase(const_iterator cit) { erase(cit.inner_); }
-
-  // This overload is necessary because otherwise erase<K>(const K&) would be
-  // a better match if non-const iterator is passed as an argument.
-  void erase(iterator it) {
-    AssertIsFull(it.control(), it.generation(), it.generation_ptr(), "erase()");
-    destroy(it.slot());
-    if (is_soo()) {
-      common().set_empty_soo();
-    } else {
-      erase_meta_only(it);
-    }
-  }
-
-  iterator erase(const_iterator first,
-                 const_iterator last) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    // We check for empty first because ClearBackingArray requires that
-    // capacity() > 0 as a precondition.
-    if (empty()) return end();
-    if (first == last) return last.inner_;
-    if (is_soo()) {
-      destroy(soo_slot());
-      common().set_empty_soo();
-      return end();
-    }
-    if (first == begin() && last == end()) {
-      // TODO(ezb): we access control bytes in destroy_slots so it could make
-      // sense to combine destroy_slots and ClearBackingArray to avoid cache
-      // misses when the table is large. Note that we also do this in clear().
-      destroy_slots();
-      ClearBackingArray(common(), GetPolicyFunctions(), /*reuse=*/true,
-                        SooEnabled());
-      common().set_reserved_growth(common().reservation_size());
-      return end();
-    }
-    while (first != last) {
-      erase(first++);
-    }
-    return last.inner_;
-  }
-
-  // Moves elements from `src` into `this`.
-  // If the element already exists in `this`, it is left unmodified in `src`.
-  template <typename H, typename E>
-  void merge(raw_hash_set<Policy, H, E, Alloc>& src) {  // NOLINT
-    assert(this != &src);
-    // Returns whether insertion took place.
-    const auto insert_slot = [this](slot_type* src_slot) {
-      return PolicyTraits::apply(InsertSlot<false>{*this, std::move(*src_slot)},
-                                 PolicyTraits::element(src_slot))
-          .second;
-    };
-
-    if (src.is_soo()) {
-      if (src.empty()) return;
-      if (insert_slot(src.soo_slot())) src.common().set_empty_soo();
-      return;
-    }
-    for (auto it = src.begin(), e = src.end(); it != e;) {
-      auto next = std::next(it);
-      if (insert_slot(it.slot())) src.erase_meta_only(it);
-      it = next;
-    }
-  }
-
-  template <typename H, typename E>
-  void merge(raw_hash_set<Policy, H, E, Alloc>&& src) {
-    merge(src);
-  }
-
-  node_type extract(const_iterator position) {
-    AssertIsFull(position.control(), position.inner_.generation(),
-                 position.inner_.generation_ptr(), "extract()");
-    auto node = CommonAccess::Transfer<node_type>(alloc_ref(), position.slot());
-    if (is_soo()) {
-      common().set_empty_soo();
-    } else {
-      erase_meta_only(position);
-    }
-    return node;
-  }
-
-  template <
-      class K = key_type,
-      typename std::enable_if<!std::is_same<K, iterator>::value, int>::type = 0>
-  node_type extract(const key_arg<K>& key) {
-    auto it = find(key);
-    return it == end() ? node_type() : extract(const_iterator{it});
-  }
-
-  void swap(raw_hash_set& that) noexcept(
-      IsNoThrowSwappable<hasher>() && IsNoThrowSwappable<key_equal>() &&
-      IsNoThrowSwappable<allocator_type>(
-          typename AllocTraits::propagate_on_container_swap{})) {
-    using std::swap;
-    swap_common(that);
-    swap(hash_ref(), that.hash_ref());
-    swap(eq_ref(), that.eq_ref());
-    SwapAlloc(alloc_ref(), that.alloc_ref(),
-              typename AllocTraits::propagate_on_container_swap{});
-  }
-
-  void rehash(size_t n) {
-    const size_t cap = capacity();
-    if (n == 0) {
-      if (cap == 0 || is_soo()) return;
-      if (empty()) {
-        ClearBackingArray(common(), GetPolicyFunctions(), /*reuse=*/false,
-                          SooEnabled());
-        return;
-      }
-      if (SooEnabled() && size() <= SooCapacity()) {
-        alignas(slot_type) unsigned char slot_space[sizeof(slot_type)];
-        slot_type* tmp_slot = to_slot(slot_space);
-        transfer(tmp_slot, begin().slot());
-        ClearBackingArray(common(), GetPolicyFunctions(), /*reuse=*/false,
-                          SooEnabled());
-        transfer(soo_slot(), tmp_slot);
-        common().set_full_soo();
-        return;
-      }
-    }
-
-    // bitor is a faster way of doing `max` here. We will round up to the next
-    // power-of-2-minus-1, so bitor is good enough.
-    auto m = NormalizeCapacity(n | GrowthToLowerboundCapacity(size()));
-    // n == 0 unconditionally rehashes as per the standard.
-    if (n == 0 || m > cap) {
-      resize(m);
-
-      // This is after resize, to ensure that we have completed the allocation
-      // and have potentially sampled the hashtable.
-      infoz().RecordReservation(n);
-    }
-  }
-
-  void reserve(size_t n) {
-    const size_t max_size_before_growth =
-        is_soo() ? SooCapacity() : size() + growth_left();
-    if (n > max_size_before_growth) {
-      size_t m = GrowthToLowerboundCapacity(n);
-      resize(NormalizeCapacity(m));
-
-      // This is after resize, to ensure that we have completed the allocation
-      // and have potentially sampled the hashtable.
-      infoz().RecordReservation(n);
-    }
-    common().reset_reserved_growth(n);
-    common().set_reservation_size(n);
-  }
-
-  // Extension API: support for heterogeneous keys.
-  //
-  //   std::unordered_set<std::string> s;
-  //   // Turns "abc" into std::string.
-  //   s.count("abc");
-  //
-  //   ch_set<std::string> s;
-  //   // Uses "abc" directly without copying it into std::string.
-  //   s.count("abc");
-  template <class K = key_type>
-  size_t count(const key_arg<K>& key) const {
-    return find(key) == end() ? 0 : 1;
-  }
-
-  // Issues CPU prefetch instructions for the memory needed to find or insert
-  // a key.  Like all lookup functions, this support heterogeneous keys.
-  //
-  // NOTE: This is a very low level operation and should not be used without
-  // specific benchmarks indicating its importance.
-  template <class K = key_type>
-  void prefetch(const key_arg<K>& key) const {
-    if (SooEnabled() ? is_soo() : capacity() == 0) return;
-    (void)key;
-    // Avoid probing if we won't be able to prefetch the addresses received.
-#ifdef ABSL_HAVE_PREFETCH
-    prefetch_heap_block();
-    auto seq = probe(common(), hash_ref()(key));
-    PrefetchToLocalCache(control() + seq.offset());
-    PrefetchToLocalCache(slot_array() + seq.offset());
-#endif  // ABSL_HAVE_PREFETCH
-  }
-
-  // The API of find() has two extensions.
-  //
-  // 1. The hash can be passed by the user. It must be equal to the hash of the
-  // key.
-  //
-  // 2. The type of the key argument doesn't have to be key_type. This is so
-  // called heterogeneous key support.
-  template <class K = key_type>
-  iterator find(const key_arg<K>& key,
-                size_t hash) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    if (is_soo()) return find_soo(key);
-    return find_non_soo(key, hash);
-  }
-  template <class K = key_type>
-  iterator find(const key_arg<K>& key) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    if (is_soo()) return find_soo(key);
-    prefetch_heap_block();
-    return find_non_soo(key, hash_ref()(key));
-  }
-
-  template <class K = key_type>
-  const_iterator find(const key_arg<K>& key,
-                      size_t hash) const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return const_cast<raw_hash_set*>(this)->find(key, hash);
-  }
-  template <class K = key_type>
-  const_iterator find(const key_arg<K>& key) const
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return const_cast<raw_hash_set*>(this)->find(key);
-  }
-
-  template <class K = key_type>
-  bool contains(const key_arg<K>& key) const {
-    // Here neither the iterator returned by `find()` nor `end()` can be invalid
-    // outside of potential thread-safety issues.
-    // `find()`'s return value is constructed, used, and then destructed
-    // all in this context.
-    return !find(key).unchecked_equals(end());
-  }
-
-  template <class K = key_type>
-  std::pair<iterator, iterator> equal_range(const key_arg<K>& key)
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    auto it = find(key);
-    if (it != end()) return {it, std::next(it)};
-    return {it, it};
-  }
-  template <class K = key_type>
-  std::pair<const_iterator, const_iterator> equal_range(
-      const key_arg<K>& key) const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    auto it = find(key);
-    if (it != end()) return {it, std::next(it)};
-    return {it, it};
-  }
-
-  size_t bucket_count() const { return capacity(); }
-  float load_factor() const {
-    return capacity() ? static_cast<double>(size()) / capacity() : 0.0;
-  }
-  float max_load_factor() const { return 1.0f; }
-  void max_load_factor(float) {
-    // Does nothing.
-  }
-
-  hasher hash_function() const { return hash_ref(); }
-  key_equal key_eq() const { return eq_ref(); }
-  allocator_type get_allocator() const { return alloc_ref(); }
-
-  friend bool operator==(const raw_hash_set& a, const raw_hash_set& b) {
-    if (a.size() != b.size()) return false;
-    const raw_hash_set* outer = &a;
-    const raw_hash_set* inner = &b;
-    if (outer->capacity() > inner->capacity()) std::swap(outer, inner);
-    for (const value_type& elem : *outer) {
-      auto it = PolicyTraits::apply(FindElement{*inner}, elem);
-      if (it == inner->end() || !(*it == elem)) return false;
-    }
-    return true;
-  }
-
-  friend bool operator!=(const raw_hash_set& a, const raw_hash_set& b) {
-    return !(a == b);
-  }
-
-  template <typename H>
-  friend typename std::enable_if<H::template is_hashable<value_type>::value,
-                                 H>::type
-  AbslHashValue(H h, const raw_hash_set& s) {
-    return H::combine(H::combine_unordered(std::move(h), s.begin(), s.end()),
-                      s.size());
-  }
-
-  friend void swap(raw_hash_set& a,
-                   raw_hash_set& b) noexcept(noexcept(a.swap(b))) {
-    a.swap(b);
-  }
-
- private:
-  template <class Container, typename Enabler>
-  friend struct absl::container_internal::hashtable_debug_internal::
-      HashtableDebugAccess;
-
-  struct FindElement {
-    template <class K, class... Args>
-    const_iterator operator()(const K& key, Args&&...) const {
-      return s.find(key);
-    }
-    const raw_hash_set& s;
-  };
-
-  struct HashElement {
-    template <class K, class... Args>
-    size_t operator()(const K& key, Args&&...) const {
-      return h(key);
-    }
-    const hasher& h;
-  };
-
-  template <class K1>
-  struct EqualElement {
-    template <class K2, class... Args>
-    bool operator()(const K2& lhs, Args&&...) const {
-      return eq(lhs, rhs);
-    }
-    const K1& rhs;
-    const key_equal& eq;
-  };
-
-  struct EmplaceDecomposable {
-    template <class K, class... Args>
-    std::pair<iterator, bool> operator()(const K& key, Args&&... args) const {
-      auto res = s.find_or_prepare_insert(key);
-      if (res.second) {
-        s.emplace_at(res.first, std::forward<Args>(args)...);
-      }
-      return res;
-    }
-    raw_hash_set& s;
-  };
-
-  template <bool do_destroy>
-  struct InsertSlot {
-    template <class K, class... Args>
-    std::pair<iterator, bool> operator()(const K& key, Args&&...) && {
-      auto res = s.find_or_prepare_insert(key);
-      if (res.second) {
-        s.transfer(res.first.slot(), &slot);
-      } else if (do_destroy) {
-        s.destroy(&slot);
-      }
-      return res;
-    }
-    raw_hash_set& s;
-    // Constructed slot. Either moved into place or destroyed.
-    slot_type&& slot;
-  };
-
-  // TODO(b/303305702): re-enable reentrant validation.
-  template <typename... Args>
-  inline void construct(slot_type* slot, Args&&... args) {
-    PolicyTraits::construct(&alloc_ref(), slot, std::forward<Args>(args)...);
-  }
-  inline void destroy(slot_type* slot) {
-    PolicyTraits::destroy(&alloc_ref(), slot);
-  }
-  inline void transfer(slot_type* to, slot_type* from) {
-    PolicyTraits::transfer(&alloc_ref(), to, from);
-  }
-
-  // TODO(b/289225379): consider having a helper class that has the impls for
-  // SOO functionality.
-  template <class K = key_type>
-  iterator find_soo(const key_arg<K>& key) {
-    assert(is_soo());
-    return empty() || !PolicyTraits::apply(EqualElement<K>{key, eq_ref()},
-                                           PolicyTraits::element(soo_slot()))
-               ? end()
-               : soo_iterator();
-  }
-
-  template <class K = key_type>
-  iterator find_non_soo(const key_arg<K>& key, size_t hash) {
-    assert(!is_soo());
-    auto seq = probe(common(), hash);
-    const ctrl_t* ctrl = control();
-    while (true) {
-      Group g{ctrl + seq.offset()};
-      for (uint32_t i : g.Match(H2(hash))) {
-        if (ABSL_PREDICT_TRUE(PolicyTraits::apply(
-                EqualElement<K>{key, eq_ref()},
-                PolicyTraits::element(slot_array() + seq.offset(i)))))
-          return iterator_at(seq.offset(i));
-      }
-      if (ABSL_PREDICT_TRUE(g.MaskEmpty())) return end();
-      seq.next();
-      assert(seq.index() <= capacity() && "full table!");
-    }
-  }
-
-  inline void destroy_slots() {
-    assert(!is_soo());
-    if (PolicyTraits::template destroy_is_trivial<Alloc>()) return;
-    IterateOverFullSlots(
-        common(), slot_array(),
-        [&](const ctrl_t*, slot_type* slot)
-            ABSL_ATTRIBUTE_ALWAYS_INLINE { this->destroy(slot); });
-  }
-
-  inline void dealloc() {
-    assert(capacity() != 0);
-    // Unpoison before returning the memory to the allocator.
-    SanitizerUnpoisonMemoryRegion(slot_array(), sizeof(slot_type) * capacity());
-    infoz().Unregister();
-    Deallocate<BackingArrayAlignment(alignof(slot_type))>(
-        &alloc_ref(), common().backing_array_start(),
-        common().alloc_size(sizeof(slot_type), alignof(slot_type)));
-  }
-
-  inline void destructor_impl() {
-    if (capacity() == 0) return;
-    if (is_soo()) {
-      if (!empty()) destroy(soo_slot());
-      return;
-    }
-    destroy_slots();
-    dealloc();
-  }
-
-  // Erases, but does not destroy, the value pointed to by `it`.
-  //
-  // This merely updates the pertinent control byte. This can be used in
-  // conjunction with Policy::transfer to move the object to another place.
-  void erase_meta_only(const_iterator it) {
-    assert(!is_soo());
-    EraseMetaOnly(common(), static_cast<size_t>(it.control() - control()),
-                  sizeof(slot_type));
-  }
-
-  size_t hash_of(slot_type* slot) const {
-    return PolicyTraits::apply(HashElement{hash_ref()},
-                               PolicyTraits::element(slot));
-  }
-
-  // Resizes table to the new capacity and move all elements to the new
-  // positions accordingly.
-  //
-  // Note that for better performance instead of
-  // find_first_non_full(common(), hash),
-  // HashSetResizeHelper::FindFirstNonFullAfterResize(
-  //    common(), old_capacity, hash)
-  // can be called right after `resize`.
-  ABSL_ATTRIBUTE_NOINLINE void resize(size_t new_capacity) {
-    assert(IsValidCapacity(new_capacity));
-    assert(!fits_in_soo(new_capacity));
-    const bool was_soo = is_soo();
-    const bool had_soo_slot = was_soo && !empty();
-    const ctrl_t soo_slot_h2 =
-        had_soo_slot ? static_cast<ctrl_t>(H2(hash_of(soo_slot())))
-                     : ctrl_t::kEmpty;
-    HashSetResizeHelper resize_helper(common(), was_soo, had_soo_slot);
-    // Initialize HashSetResizeHelper::old_heap_or_soo_. We can't do this in
-    // HashSetResizeHelper constructor because it can't transfer slots when
-    // transfer_uses_memcpy is false.
-    // TODO(b/289225379): try to handle more of the SOO cases inside
-    // InitializeSlots. See comment on cl/555990034 snapshot #63.
-    if (PolicyTraits::transfer_uses_memcpy() || !had_soo_slot) {
-      resize_helper.old_heap_or_soo() = common().heap_or_soo();
-    } else {
-      transfer(to_slot(resize_helper.old_soo_data()), soo_slot());
-    }
-    common().set_capacity(new_capacity);
-    // Note that `InitializeSlots` does different number initialization steps
-    // depending on the values of `transfer_uses_memcpy` and capacities.
-    // Refer to the comment in `InitializeSlots` for more details.
-    const bool grow_single_group =
-        resize_helper.InitializeSlots<CharAlloc, sizeof(slot_type),
-                                      PolicyTraits::transfer_uses_memcpy(),
-                                      alignof(slot_type)>(
-            common(), CharAlloc(alloc_ref()), soo_slot_h2);
-
-    // In the SooEnabled() case, capacity is never 0 so we don't check.
-    if (!SooEnabled() && resize_helper.old_capacity() == 0) {
-      // InitializeSlots did all the work including infoz().RecordRehash().
-      return;
-    }
-    assert(resize_helper.old_capacity() > 0);
-    // Nothing more to do in this case.
-    if (was_soo && !had_soo_slot) return;
-
-    slot_type* new_slots = slot_array();
-    if (grow_single_group) {
-      if (PolicyTraits::transfer_uses_memcpy()) {
-        // InitializeSlots did all the work.
-        return;
-      }
-      if (was_soo) {
-        transfer(new_slots + resize_helper.SooSlotIndex(),
-                 to_slot(resize_helper.old_soo_data()));
-        return;
-      } else {
-        // We want GrowSizeIntoSingleGroup to be called here in order to make
-        // InitializeSlots not depend on PolicyTraits.
-        resize_helper.GrowSizeIntoSingleGroup<PolicyTraits>(common(),
-                                                            alloc_ref());
-      }
-    } else {
-      // InitializeSlots prepares control bytes to correspond to empty table.
-      const auto insert_slot = [&](slot_type* slot) {
-        size_t hash = PolicyTraits::apply(HashElement{hash_ref()},
-                                          PolicyTraits::element(slot));
-        auto target = find_first_non_full(common(), hash);
-        SetCtrl(common(), target.offset, H2(hash), sizeof(slot_type));
-        transfer(new_slots + target.offset, slot);
-        return target.probe_length;
-      };
-      if (was_soo) {
-        insert_slot(to_slot(resize_helper.old_soo_data()));
-        return;
-      } else {
-        auto* old_slots =
-            reinterpret_cast<slot_type*>(resize_helper.old_slots());
-        size_t total_probe_length = 0;
-        for (size_t i = 0; i != resize_helper.old_capacity(); ++i) {
-          if (IsFull(resize_helper.old_ctrl()[i])) {
-            total_probe_length += insert_slot(old_slots + i);
-          }
-        }
-        infoz().RecordRehash(total_probe_length);
-      }
-    }
-    resize_helper.DeallocateOld<alignof(slot_type)>(CharAlloc(alloc_ref()),
-                                                    sizeof(slot_type));
-  }
-
-  // Prunes control bytes to remove as many tombstones as possible.
-  //
-  // See the comment on `rehash_and_grow_if_necessary()`.
-  inline void drop_deletes_without_resize() {
-    // Stack-allocate space for swapping elements.
-    alignas(slot_type) unsigned char tmp[sizeof(slot_type)];
-    DropDeletesWithoutResize(common(), &hash_ref(), GetPolicyFunctions(), tmp);
-  }
-
-  // Called whenever the table *might* need to conditionally grow.
-  //
-  // This function is an optimization opportunity to perform a rehash even when
-  // growth is unnecessary, because vacating tombstones is beneficial for
-  // performance in the long-run.
-  void rehash_and_grow_if_necessary() {
-    const size_t cap = capacity();
-    if (cap > Group::kWidth &&
-        // Do these calculations in 64-bit to avoid overflow.
-        size() * uint64_t{32} <= cap * uint64_t{25}) {
-      // Squash DELETED without growing if there is enough capacity.
-      //
-      // Rehash in place if the current size is <= 25/32 of capacity.
-      // Rationale for such a high factor: 1) drop_deletes_without_resize() is
-      // faster than resize, and 2) it takes quite a bit of work to add
-      // tombstones.  In the worst case, seems to take approximately 4
-      // insert/erase pairs to create a single tombstone and so if we are
-      // rehashing because of tombstones, we can afford to rehash-in-place as
-      // long as we are reclaiming at least 1/8 the capacity without doing more
-      // than 2X the work.  (Where "work" is defined to be size() for rehashing
-      // or rehashing in place, and 1 for an insert or erase.)  But rehashing in
-      // place is faster per operation than inserting or even doubling the size
-      // of the table, so we actually afford to reclaim even less space from a
-      // resize-in-place.  The decision is to rehash in place if we can reclaim
-      // at about 1/8th of the usable capacity (specifically 3/28 of the
-      // capacity) which means that the total cost of rehashing will be a small
-      // fraction of the total work.
-      //
-      // Here is output of an experiment using the BM_CacheInSteadyState
-      // benchmark running the old case (where we rehash-in-place only if we can
-      // reclaim at least 7/16*capacity) vs. this code (which rehashes in place
-      // if we can recover 3/32*capacity).
-      //
-      // Note that although in the worst-case number of rehashes jumped up from
-      // 15 to 190, but the number of operations per second is almost the same.
-      //
-      // Abridged output of running BM_CacheInSteadyState benchmark from
-      // raw_hash_set_benchmark.   N is the number of insert/erase operations.
-      //
-      //      | OLD (recover >= 7/16        | NEW (recover >= 3/32)
-      // size |    N/s LoadFactor NRehashes |    N/s LoadFactor NRehashes
-      //  448 | 145284       0.44        18 | 140118       0.44        19
-      //  493 | 152546       0.24        11 | 151417       0.48        28
-      //  538 | 151439       0.26        11 | 151152       0.53        38
-      //  583 | 151765       0.28        11 | 150572       0.57        50
-      //  628 | 150241       0.31        11 | 150853       0.61        66
-      //  672 | 149602       0.33        12 | 150110       0.66        90
-      //  717 | 149998       0.35        12 | 149531       0.70       129
-      //  762 | 149836       0.37        13 | 148559       0.74       190
-      //  807 | 149736       0.39        14 | 151107       0.39        14
-      //  852 | 150204       0.42        15 | 151019       0.42        15
-      drop_deletes_without_resize();
-    } else {
-      // Otherwise grow the container.
-      resize(NextCapacity(cap));
-    }
-  }
-
-  // Casting directly from e.g. char* to slot_type* can cause compilation errors
-  // on objective-C. This function converts to void* first, avoiding the issue.
-  static slot_type* to_slot(void* buf) {
-    return reinterpret_cast<slot_type*>(buf);
-  }
-
-  // Requires that lhs does not have a full SOO slot.
-  static void move_common(bool that_is_full_soo, allocator_type& rhs_alloc,
-                          CommonFields& lhs, CommonFields&& rhs) {
-    if (PolicyTraits::transfer_uses_memcpy() || !that_is_full_soo) {
-      lhs = std::move(rhs);
-    } else {
-      lhs.move_non_heap_or_soo_fields(rhs);
-      // TODO(b/303305702): add reentrancy guard.
-      PolicyTraits::transfer(&rhs_alloc, to_slot(lhs.soo_data()),
-                             to_slot(rhs.soo_data()));
-    }
-  }
-
-  // Swaps common fields making sure to avoid memcpy'ing a full SOO slot if we
-  // aren't allowed to do so.
-  void swap_common(raw_hash_set& that) {
-    using std::swap;
-    if (PolicyTraits::transfer_uses_memcpy()) {
-      swap(common(), that.common());
-      return;
-    }
-    CommonFields tmp = CommonFields::CreateDefault<SooEnabled()>();
-    const bool that_is_full_soo = that.is_full_soo();
-    move_common(that_is_full_soo, that.alloc_ref(), tmp,
-                std::move(that.common()));
-    move_common(is_full_soo(), alloc_ref(), that.common(), std::move(common()));
-    move_common(that_is_full_soo, that.alloc_ref(), common(), std::move(tmp));
-  }
-
-  void maybe_increment_generation_or_rehash_on_move() {
-    if (!SwisstableGenerationsEnabled() || capacity() == 0 || is_soo()) {
-      return;
-    }
-    common().increment_generation();
-    if (!empty() && common().should_rehash_for_bug_detection_on_move()) {
-      resize(capacity());
-    }
-  }
-
-  template<bool propagate_alloc>
-  raw_hash_set& assign_impl(raw_hash_set&& that) {
-    // We don't bother checking for this/that aliasing. We just need to avoid
-    // breaking the invariants in that case.
-    destructor_impl();
-    move_common(that.is_full_soo(), that.alloc_ref(), common(),
-                std::move(that.common()));
-    // TODO(b/296061262): move instead of copying hash/eq/alloc.
-    hash_ref() = that.hash_ref();
-    eq_ref() = that.eq_ref();
-    CopyAlloc(alloc_ref(), that.alloc_ref(),
-              std::integral_constant<bool, propagate_alloc>());
-    that.common() = CommonFields::CreateDefault<SooEnabled()>();
-    maybe_increment_generation_or_rehash_on_move();
-    return *this;
-  }
-
-  raw_hash_set& move_elements_allocs_unequal(raw_hash_set&& that) {
-    const size_t size = that.size();
-    if (size == 0) return *this;
-    reserve(size);
-    for (iterator it = that.begin(); it != that.end(); ++it) {
-      insert(std::move(PolicyTraits::element(it.slot())));
-      that.destroy(it.slot());
-    }
-    if (!that.is_soo()) that.dealloc();
-    that.common() = CommonFields::CreateDefault<SooEnabled()>();
-    maybe_increment_generation_or_rehash_on_move();
-    return *this;
-  }
-
-  raw_hash_set& move_assign(raw_hash_set&& that,
-                            std::true_type /*propagate_alloc*/) {
-    return assign_impl<true>(std::move(that));
-  }
-  raw_hash_set& move_assign(raw_hash_set&& that,
-                            std::false_type /*propagate_alloc*/) {
-    if (alloc_ref() == that.alloc_ref()) {
-      return assign_impl<false>(std::move(that));
-    }
-    // Aliasing can't happen here because allocs would compare equal above.
-    assert(this != &that);
-    destructor_impl();
-    // We can't take over that's memory so we need to move each element.
-    // While moving elements, this should have that's hash/eq so copy hash/eq
-    // before moving elements.
-    // TODO(b/296061262): move instead of copying hash/eq.
-    hash_ref() = that.hash_ref();
-    eq_ref() = that.eq_ref();
-    return move_elements_allocs_unequal(std::move(that));
-  }
-
- protected:
-  // Attempts to find `key` in the table; if it isn't found, returns an iterator
-  // where the value can be inserted into, with the control byte already set to
-  // `key`'s H2. Returns a bool indicating whether an insertion can take place.
-  template <class K>
-  std::pair<iterator, bool> find_or_prepare_insert(const K& key) {
-    if (is_soo()) {
-      if (empty()) {
-        common().set_full_soo();
-        return {soo_iterator(), true};
-      }
-      if (PolicyTraits::apply(EqualElement<K>{key, eq_ref()},
-                              PolicyTraits::element(soo_slot()))) {
-        return {soo_iterator(), false};
-      }
-      resize(NextCapacity(SooCapacity()));
-      const size_t index =
-          PrepareInsertAfterSoo(hash_ref()(key), sizeof(slot_type), common());
-      return {iterator_at(index), true};
-    }
-    prefetch_heap_block();
-    auto hash = hash_ref()(key);
-    auto seq = probe(common(), hash);
-    const ctrl_t* ctrl = control();
-    while (true) {
-      Group g{ctrl + seq.offset()};
-      for (uint32_t i : g.Match(H2(hash))) {
-        if (ABSL_PREDICT_TRUE(PolicyTraits::apply(
-                EqualElement<K>{key, eq_ref()},
-                PolicyTraits::element(slot_array() + seq.offset(i)))))
-          return {iterator_at(seq.offset(i)), false};
-      }
-      if (ABSL_PREDICT_TRUE(g.MaskEmpty())) break;
-      seq.next();
-      assert(seq.index() <= capacity() && "full table!");
-    }
-    return {iterator_at(prepare_insert(hash)), true};
-  }
-
-  // Given the hash of a value not currently in the table, finds the next
-  // viable slot index to insert it at.
-  //
-  // REQUIRES: At least one non-full slot available.
-  size_t prepare_insert(size_t hash) ABSL_ATTRIBUTE_NOINLINE {
-    assert(!is_soo());
-    const bool rehash_for_bug_detection =
-        common().should_rehash_for_bug_detection_on_insert();
-    if (rehash_for_bug_detection) {
-      // Move to a different heap allocation in order to detect bugs.
-      const size_t cap = capacity();
-      resize(growth_left() > 0 ? cap : NextCapacity(cap));
-    }
-    FindInfo target;
-    if (!rehash_for_bug_detection && ABSL_PREDICT_FALSE(growth_left() == 0)) {
-      const size_t old_capacity = capacity();
-      rehash_and_grow_if_necessary();
-      // NOTE: It is safe to use `FindFirstNonFullAfterResize` after
-      // `rehash_and_grow_if_necessary`, whether capacity changes or not.
-      // `rehash_and_grow_if_necessary` may *not* call `resize`
-      // and perform `drop_deletes_without_resize` instead. But this
-      // could happen only on big tables and will not change capacity.
-      // For big tables `FindFirstNonFullAfterResize` will always
-      // fallback to normal `find_first_non_full`.
-      target = HashSetResizeHelper::FindFirstNonFullAfterResize(
-          common(), old_capacity, hash);
-    } else {
-      target = find_first_non_full(common(), hash);
-    }
-    PrepareInsertCommon(common());
-    set_growth_left(growth_left() - IsEmpty(control()[target.offset]));
-    SetCtrl(common(), target.offset, H2(hash), sizeof(slot_type));
-    infoz().RecordInsert(hash, target.probe_length);
-    return target.offset;
-  }
-
-  // Constructs the value in the space pointed by the iterator. This only works
-  // after an unsuccessful find_or_prepare_insert() and before any other
-  // modifications happen in the raw_hash_set.
-  //
-  // PRECONDITION: iter was returned from find_or_prepare_insert(k), where k is
-  // the key decomposed from `forward<Args>(args)...`, and the bool returned by
-  // find_or_prepare_insert(k) was true.
-  // POSTCONDITION: *m.iterator_at(i) == value_type(forward<Args>(args)...).
-  template <class... Args>
-  void emplace_at(iterator iter, Args&&... args) {
-    construct(iter.slot(), std::forward<Args>(args)...);
-
-    assert(PolicyTraits::apply(FindElement{*this}, *iter) == iter &&
-           "constructed value does not match the lookup key");
-  }
-
-  iterator iterator_at(size_t i) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return {control() + i, slot_array() + i, common().generation_ptr()};
-  }
-  const_iterator iterator_at(size_t i) const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return const_cast<raw_hash_set*>(this)->iterator_at(i);
-  }
-
-  reference unchecked_deref(iterator it) { return it.unchecked_deref(); }
-
- private:
-  friend struct RawHashSetTestOnlyAccess;
-
-  // The number of slots we can still fill without needing to rehash.
-  //
-  // This is stored separately due to tombstones: we do not include tombstones
-  // in the growth capacity, because we'd like to rehash when the table is
-  // otherwise filled with tombstones: otherwise, probe sequences might get
-  // unacceptably long without triggering a rehash. Callers can also force a
-  // rehash via the standard `rehash(0)`, which will recompute this value as a
-  // side-effect.
-  //
-  // See `CapacityToGrowth()`.
-  size_t growth_left() const {
-    assert(!is_soo());
-    return common().growth_left();
-  }
-  void set_growth_left(size_t gl) {
-    assert(!is_soo());
-    return common().set_growth_left(gl);
-  }
-
-  // Prefetch the heap-allocated memory region to resolve potential TLB and
-  // cache misses. This is intended to overlap with execution of calculating the
-  // hash for a key.
-  void prefetch_heap_block() const {
-    if (is_soo()) return;
-#if ABSL_HAVE_BUILTIN(__builtin_prefetch) || defined(__GNUC__)
-    __builtin_prefetch(control(), 0, 1);
-#endif
-  }
-
-  CommonFields& common() { return settings_.template get<0>(); }
-  const CommonFields& common() const { return settings_.template get<0>(); }
-
-  ctrl_t* control() const {
-    assert(!is_soo());
-    return common().control();
-  }
-  slot_type* slot_array() const {
-    assert(!is_soo());
-    return static_cast<slot_type*>(common().slot_array());
-  }
-  slot_type* soo_slot() {
-    assert(is_soo());
-    return static_cast<slot_type*>(common().soo_data());
-  }
-  const slot_type* soo_slot() const {
-    return reinterpret_cast<raw_hash_set*>(this)->soo_slot();
-  }
-  iterator soo_iterator() {
-    return {SooControl(), soo_slot(), common().generation_ptr()};
-  }
-  const_iterator soo_iterator() const {
-    return reinterpret_cast<raw_hash_set*>(this)->soo_iterator();
-  }
-  HashtablezInfoHandle infoz() {
-    assert(!is_soo());
-    return common().infoz();
-  }
-
-  hasher& hash_ref() { return settings_.template get<1>(); }
-  const hasher& hash_ref() const { return settings_.template get<1>(); }
-  key_equal& eq_ref() { return settings_.template get<2>(); }
-  const key_equal& eq_ref() const { return settings_.template get<2>(); }
-  allocator_type& alloc_ref() { return settings_.template get<3>(); }
-  const allocator_type& alloc_ref() const {
-    return settings_.template get<3>();
-  }
-
-  static void transfer_slot_fn(void* set, void* dst, void* src) {
-    auto* h = static_cast<raw_hash_set*>(set);
-    h->transfer(static_cast<slot_type*>(dst), static_cast<slot_type*>(src));
-  }
-  // Note: dealloc_fn will only be used if we have a non-standard allocator.
-  static void dealloc_fn(CommonFields& common, const PolicyFunctions&) {
-    auto* set = reinterpret_cast<raw_hash_set*>(&common);
-
-    // Unpoison before returning the memory to the allocator.
-    SanitizerUnpoisonMemoryRegion(common.slot_array(),
-                                  sizeof(slot_type) * common.capacity());
-
-    common.infoz().Unregister();
-    Deallocate<BackingArrayAlignment(alignof(slot_type))>(
-        &set->alloc_ref(), common.backing_array_start(),
-        common.alloc_size(sizeof(slot_type), alignof(slot_type)));
-  }
-
-  static const PolicyFunctions& GetPolicyFunctions() {
-    static constexpr PolicyFunctions value = {
-        sizeof(slot_type),
-        PolicyTraits::template get_hash_slot_fn<hasher>(),
-        PolicyTraits::transfer_uses_memcpy()
-            ? TransferRelocatable<sizeof(slot_type)>
-            : &raw_hash_set::transfer_slot_fn,
-        (std::is_same<SlotAlloc, std::allocator<slot_type>>::value
-             ? &DeallocateStandard<alignof(slot_type)>
-             : &raw_hash_set::dealloc_fn),
-    };
-    return value;
-  }
-
-  // Bundle together CommonFields plus other objects which might be empty.
-  // CompressedTuple will ensure that sizeof is not affected by any of the empty
-  // fields that occur after CommonFields.
-  absl::container_internal::CompressedTuple<CommonFields, hasher, key_equal,
-                                            allocator_type>
-      settings_{CommonFields::CreateDefault<SooEnabled()>(), hasher{},
-                key_equal{}, allocator_type{}};
-};
-
-// Erases all elements that satisfy the predicate `pred` from the container `c`.
-template <typename P, typename H, typename E, typename A, typename Predicate>
-typename raw_hash_set<P, H, E, A>::size_type EraseIf(
-    Predicate& pred, raw_hash_set<P, H, E, A>* c) {
-  const auto initial_size = c->size();
-  for (auto it = c->begin(), last = c->end(); it != last;) {
-    if (pred(*it)) {
-      c->erase(it++);
-    } else {
-      ++it;
-    }
-  }
-  return initial_size - c->size();
-}
-
-namespace hashtable_debug_internal {
-template <typename Set>
-struct HashtableDebugAccess<Set, absl::void_t<typename Set::raw_hash_set>> {
-  using Traits = typename Set::PolicyTraits;
-  using Slot = typename Traits::slot_type;
-
-  static size_t GetNumProbes(const Set& set,
-                             const typename Set::key_type& key) {
-    if (set.is_soo()) return 0;
-    size_t num_probes = 0;
-    size_t hash = set.hash_ref()(key);
-    auto seq = probe(set.common(), hash);
-    const ctrl_t* ctrl = set.control();
-    while (true) {
-      container_internal::Group g{ctrl + seq.offset()};
-      for (uint32_t i : g.Match(container_internal::H2(hash))) {
-        if (Traits::apply(
-                typename Set::template EqualElement<typename Set::key_type>{
-                    key, set.eq_ref()},
-                Traits::element(set.slot_array() + seq.offset(i))))
-          return num_probes;
-        ++num_probes;
-      }
-      if (g.MaskEmpty()) return num_probes;
-      seq.next();
-      ++num_probes;
-    }
-  }
-
-  static size_t AllocatedByteSize(const Set& c) {
-    size_t capacity = c.capacity();
-    if (capacity == 0) return 0;
-    size_t m =
-        c.is_soo() ? 0 : c.common().alloc_size(sizeof(Slot), alignof(Slot));
-
-    size_t per_slot = Traits::space_used(static_cast<const Slot*>(nullptr));
-    if (per_slot != ~size_t{}) {
-      m += per_slot * c.size();
-    } else {
-      for (auto it = c.begin(); it != c.end(); ++it) {
-        m += Traits::space_used(it.slot());
-      }
-    }
-    return m;
-  }
-};
-
-}  // namespace hashtable_debug_internal
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#undef ABSL_SWISSTABLE_ENABLE_GENERATIONS
-
-#endif  // ABSL_CONTAINER_INTERNAL_RAW_HASH_SET_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/test_allocator.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/test_allocator.h
deleted file mode 100644
index 8e365a3c..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/test_allocator.h
+++ /dev/null
@@ -1,387 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_TEST_ALLOCATOR_H_
-#define ABSL_CONTAINER_INTERNAL_TEST_ALLOCATOR_H_
-
-#include <cassert>
-#include <cstddef>
-#include <cstdint>
-#include <memory>
-#include <type_traits>
-
-#include "gtest/gtest.h"
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-// This is a stateful allocator, but the state lives outside of the
-// allocator (in whatever test is using the allocator). This is odd
-// but helps in tests where the allocator is propagated into nested
-// containers - that chain of allocators uses the same state and is
-// thus easier to query for aggregate allocation information.
-template <typename T>
-class CountingAllocator {
- public:
-  using Allocator = std::allocator<T>;
-  using AllocatorTraits = std::allocator_traits<Allocator>;
-  using value_type = typename AllocatorTraits::value_type;
-  using pointer = typename AllocatorTraits::pointer;
-  using const_pointer = typename AllocatorTraits::const_pointer;
-  using size_type = typename AllocatorTraits::size_type;
-  using difference_type = typename AllocatorTraits::difference_type;
-
-  CountingAllocator() = default;
-  explicit CountingAllocator(int64_t* bytes_used) : bytes_used_(bytes_used) {}
-  CountingAllocator(int64_t* bytes_used, int64_t* instance_count)
-      : bytes_used_(bytes_used), instance_count_(instance_count) {}
-
-  template <typename U>
-  CountingAllocator(const CountingAllocator<U>& x)
-      : bytes_used_(x.bytes_used_), instance_count_(x.instance_count_) {}
-
-  pointer allocate(
-      size_type n,
-      typename AllocatorTraits::const_void_pointer hint = nullptr) {
-    Allocator allocator;
-    pointer ptr = AllocatorTraits::allocate(allocator, n, hint);
-    if (bytes_used_ != nullptr) {
-      *bytes_used_ += n * sizeof(T);
-    }
-    return ptr;
-  }
-
-  void deallocate(pointer p, size_type n) {
-    Allocator allocator;
-    AllocatorTraits::deallocate(allocator, p, n);
-    if (bytes_used_ != nullptr) {
-      *bytes_used_ -= n * sizeof(T);
-    }
-  }
-
-  template <typename U, typename... Args>
-  void construct(U* p, Args&&... args) {
-    Allocator allocator;
-    AllocatorTraits::construct(allocator, p, std::forward<Args>(args)...);
-    if (instance_count_ != nullptr) {
-      *instance_count_ += 1;
-    }
-  }
-
-  template <typename U>
-  void destroy(U* p) {
-    Allocator allocator;
-    // Ignore GCC warning bug.
-#if ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(12, 0)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wuse-after-free"
-#endif
-    AllocatorTraits::destroy(allocator, p);
-#if ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(12, 0)
-#pragma GCC diagnostic pop
-#endif
-    if (instance_count_ != nullptr) {
-      *instance_count_ -= 1;
-    }
-  }
-
-  template <typename U>
-  class rebind {
-   public:
-    using other = CountingAllocator<U>;
-  };
-
-  friend bool operator==(const CountingAllocator& a,
-                         const CountingAllocator& b) {
-    return a.bytes_used_ == b.bytes_used_ &&
-           a.instance_count_ == b.instance_count_;
-  }
-
-  friend bool operator!=(const CountingAllocator& a,
-                         const CountingAllocator& b) {
-    return !(a == b);
-  }
-
-  int64_t* bytes_used_ = nullptr;
-  int64_t* instance_count_ = nullptr;
-};
-
-template <typename T>
-struct CopyAssignPropagatingCountingAlloc : public CountingAllocator<T> {
-  using propagate_on_container_copy_assignment = std::true_type;
-
-  using Base = CountingAllocator<T>;
-  using Base::Base;
-
-  template <typename U>
-  explicit CopyAssignPropagatingCountingAlloc(
-      const CopyAssignPropagatingCountingAlloc<U>& other)
-      : Base(other.bytes_used_, other.instance_count_) {}
-
-  template <typename U>
-  struct rebind {
-    using other = CopyAssignPropagatingCountingAlloc<U>;
-  };
-};
-
-template <typename T>
-struct MoveAssignPropagatingCountingAlloc : public CountingAllocator<T> {
-  using propagate_on_container_move_assignment = std::true_type;
-
-  using Base = CountingAllocator<T>;
-  using Base::Base;
-
-  template <typename U>
-  explicit MoveAssignPropagatingCountingAlloc(
-      const MoveAssignPropagatingCountingAlloc<U>& other)
-      : Base(other.bytes_used_, other.instance_count_) {}
-
-  template <typename U>
-  struct rebind {
-    using other = MoveAssignPropagatingCountingAlloc<U>;
-  };
-};
-
-template <typename T>
-struct SwapPropagatingCountingAlloc : public CountingAllocator<T> {
-  using propagate_on_container_swap = std::true_type;
-
-  using Base = CountingAllocator<T>;
-  using Base::Base;
-
-  template <typename U>
-  explicit SwapPropagatingCountingAlloc(
-      const SwapPropagatingCountingAlloc<U>& other)
-      : Base(other.bytes_used_, other.instance_count_) {}
-
-  template <typename U>
-  struct rebind {
-    using other = SwapPropagatingCountingAlloc<U>;
-  };
-};
-
-// Tries to allocate memory at the minimum alignment even when the default
-// allocator uses a higher alignment.
-template <typename T>
-struct MinimumAlignmentAlloc : std::allocator<T> {
-  MinimumAlignmentAlloc() = default;
-
-  template <typename U>
-  explicit MinimumAlignmentAlloc(const MinimumAlignmentAlloc<U>& /*other*/) {}
-
-  template <class U>
-  struct rebind {
-    using other = MinimumAlignmentAlloc<U>;
-  };
-
-  T* allocate(size_t n) {
-    T* ptr = std::allocator<T>::allocate(n + 1);
-    char* cptr = reinterpret_cast<char*>(ptr);
-    cptr += alignof(T);
-    return reinterpret_cast<T*>(cptr);
-  }
-
-  void deallocate(T* ptr, size_t n) {
-    char* cptr = reinterpret_cast<char*>(ptr);
-    cptr -= alignof(T);
-    std::allocator<T>::deallocate(reinterpret_cast<T*>(cptr), n + 1);
-  }
-};
-
-inline bool IsAssertEnabled() {
-  // Use an assert with side-effects to figure out if they are actually enabled.
-  bool assert_enabled = false;
-  assert([&]() {  // NOLINT
-    assert_enabled = true;
-    return true;
-  }());
-  return assert_enabled;
-}
-
-template <template <class Alloc> class Container>
-void TestCopyAssignAllocPropagation() {
-  int64_t bytes1 = 0, instances1 = 0, bytes2 = 0, instances2 = 0;
-  CopyAssignPropagatingCountingAlloc<int> allocator1(&bytes1, &instances1);
-  CopyAssignPropagatingCountingAlloc<int> allocator2(&bytes2, &instances2);
-
-  // Test propagating allocator_type.
-  {
-    Container<CopyAssignPropagatingCountingAlloc<int>> c1(allocator1);
-    Container<CopyAssignPropagatingCountingAlloc<int>> c2(allocator2);
-
-    for (int i = 0; i < 100; ++i) c1.insert(i);
-
-    EXPECT_NE(c2.get_allocator(), allocator1);
-    EXPECT_EQ(instances1, 100);
-    EXPECT_EQ(instances2, 0);
-
-    c2 = c1;
-
-    EXPECT_EQ(c2.get_allocator(), allocator1);
-    EXPECT_EQ(instances1, 200);
-    EXPECT_EQ(instances2, 0);
-  }
-  // Test non-propagating allocator_type with different allocators.
-  {
-    Container<CountingAllocator<int>> c1(allocator1), c2(allocator2);
-
-    for (int i = 0; i < 100; ++i) c1.insert(i);
-
-    EXPECT_EQ(c2.get_allocator(), allocator2);
-    EXPECT_EQ(instances1, 100);
-    EXPECT_EQ(instances2, 0);
-
-    c2 = c1;
-
-    EXPECT_EQ(c2.get_allocator(), allocator2);
-    EXPECT_EQ(instances1, 100);
-    EXPECT_EQ(instances2, 100);
-  }
-  EXPECT_EQ(bytes1, 0);
-  EXPECT_EQ(instances1, 0);
-  EXPECT_EQ(bytes2, 0);
-  EXPECT_EQ(instances2, 0);
-}
-
-template <template <class Alloc> class Container>
-void TestMoveAssignAllocPropagation() {
-  int64_t bytes1 = 0, instances1 = 0, bytes2 = 0, instances2 = 0;
-  MoveAssignPropagatingCountingAlloc<int> allocator1(&bytes1, &instances1);
-  MoveAssignPropagatingCountingAlloc<int> allocator2(&bytes2, &instances2);
-
-  // Test propagating allocator_type.
-  {
-    Container<MoveAssignPropagatingCountingAlloc<int>> c1(allocator1);
-    Container<MoveAssignPropagatingCountingAlloc<int>> c2(allocator2);
-
-    for (int i = 0; i < 100; ++i) c1.insert(i);
-
-    EXPECT_NE(c2.get_allocator(), allocator1);
-    EXPECT_EQ(instances1, 100);
-    EXPECT_EQ(instances2, 0);
-
-    c2 = std::move(c1);
-
-    EXPECT_EQ(c2.get_allocator(), allocator1);
-    EXPECT_EQ(instances1, 100);
-    EXPECT_EQ(instances2, 0);
-  }
-  // Test non-propagating allocator_type with equal allocators.
-  {
-    Container<CountingAllocator<int>> c1(allocator1), c2(allocator1);
-
-    for (int i = 0; i < 100; ++i) c1.insert(i);
-
-    EXPECT_EQ(c2.get_allocator(), allocator1);
-    EXPECT_EQ(instances1, 100);
-    EXPECT_EQ(instances2, 0);
-
-    c2 = std::move(c1);
-
-    EXPECT_EQ(c2.get_allocator(), allocator1);
-    EXPECT_EQ(instances1, 100);
-    EXPECT_EQ(instances2, 0);
-  }
-  // Test non-propagating allocator_type with different allocators.
-  {
-    Container<CountingAllocator<int>> c1(allocator1), c2(allocator2);
-
-    for (int i = 0; i < 100; ++i) c1.insert(i);
-
-    EXPECT_NE(c2.get_allocator(), allocator1);
-    EXPECT_EQ(instances1, 100);
-    EXPECT_EQ(instances2, 0);
-
-    c2 = std::move(c1);
-
-    EXPECT_EQ(c2.get_allocator(), allocator2);
-    EXPECT_LE(instances1, 100);  // The values in c1 may or may not have been
-                                 // destroyed at this point.
-    EXPECT_EQ(instances2, 100);
-  }
-  EXPECT_EQ(bytes1, 0);
-  EXPECT_EQ(instances1, 0);
-  EXPECT_EQ(bytes2, 0);
-  EXPECT_EQ(instances2, 0);
-}
-
-template <template <class Alloc> class Container>
-void TestSwapAllocPropagation() {
-  int64_t bytes1 = 0, instances1 = 0, bytes2 = 0, instances2 = 0;
-  SwapPropagatingCountingAlloc<int> allocator1(&bytes1, &instances1);
-  SwapPropagatingCountingAlloc<int> allocator2(&bytes2, &instances2);
-
-  // Test propagating allocator_type.
-  {
-    Container<SwapPropagatingCountingAlloc<int>> c1(allocator1), c2(allocator2);
-
-    for (int i = 0; i < 100; ++i) c1.insert(i);
-
-    EXPECT_NE(c2.get_allocator(), allocator1);
-    EXPECT_EQ(instances1, 100);
-    EXPECT_EQ(instances2, 0);
-
-    c2.swap(c1);
-
-    EXPECT_EQ(c2.get_allocator(), allocator1);
-    EXPECT_EQ(instances1, 100);
-    EXPECT_EQ(instances2, 0);
-  }
-  // Test non-propagating allocator_type with equal allocators.
-  {
-    Container<CountingAllocator<int>> c1(allocator1), c2(allocator1);
-
-    for (int i = 0; i < 100; ++i) c1.insert(i);
-
-    EXPECT_EQ(c2.get_allocator(), allocator1);
-    EXPECT_EQ(instances1, 100);
-    EXPECT_EQ(instances2, 0);
-
-    c2.swap(c1);
-
-    EXPECT_EQ(c2.get_allocator(), allocator1);
-    EXPECT_EQ(instances1, 100);
-    EXPECT_EQ(instances2, 0);
-  }
-  // Test non-propagating allocator_type with different allocators.
-  {
-    Container<CountingAllocator<int>> c1(allocator1), c2(allocator2);
-
-    for (int i = 0; i < 100; ++i) c1.insert(i);
-
-    EXPECT_NE(c1.get_allocator(), c2.get_allocator());
-    if (IsAssertEnabled()) {
-      EXPECT_DEATH_IF_SUPPORTED(c2.swap(c1), "");
-    }
-  }
-  EXPECT_EQ(bytes1, 0);
-  EXPECT_EQ(instances1, 0);
-  EXPECT_EQ(bytes2, 0);
-  EXPECT_EQ(instances2, 0);
-}
-
-template <template <class Alloc> class Container>
-void TestAllocPropagation() {
-  TestCopyAssignAllocPropagation<Container>();
-  TestMoveAssignAllocPropagation<Container>();
-  TestSwapAllocPropagation<Container>();
-}
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_TEST_ALLOCATOR_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/test_instance_tracker.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/test_instance_tracker.h
deleted file mode 100644
index 5ff6fd71..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/test_instance_tracker.h
+++ /dev/null
@@ -1,274 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_TEST_INSTANCE_TRACKER_H_
-#define ABSL_CONTAINER_INTERNAL_TEST_INSTANCE_TRACKER_H_
-
-#include <cstdlib>
-#include <ostream>
-
-#include "absl/types/compare.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace test_internal {
-
-// A type that counts number of occurrences of the type, the live occurrences of
-// the type, as well as the number of copies, moves, swaps, and comparisons that
-// have occurred on the type. This is used as a base class for the copyable,
-// copyable+movable, and movable types below that are used in actual tests. Use
-// InstanceTracker in tests to track the number of instances.
-class BaseCountedInstance {
- public:
-  explicit BaseCountedInstance(int x) : value_(x) {
-    ++num_instances_;
-    ++num_live_instances_;
-  }
-  BaseCountedInstance(const BaseCountedInstance& x)
-      : value_(x.value_), is_live_(x.is_live_) {
-    ++num_instances_;
-    if (is_live_) ++num_live_instances_;
-    ++num_copies_;
-  }
-  BaseCountedInstance(BaseCountedInstance&& x)
-      : value_(x.value_), is_live_(x.is_live_) {
-    x.is_live_ = false;
-    ++num_instances_;
-    ++num_moves_;
-  }
-  ~BaseCountedInstance() {
-    --num_instances_;
-    if (is_live_) --num_live_instances_;
-  }
-
-  BaseCountedInstance& operator=(const BaseCountedInstance& x) {
-    value_ = x.value_;
-    if (is_live_) --num_live_instances_;
-    is_live_ = x.is_live_;
-    if (is_live_) ++num_live_instances_;
-    ++num_copies_;
-    return *this;
-  }
-  BaseCountedInstance& operator=(BaseCountedInstance&& x) {
-    value_ = x.value_;
-    if (is_live_) --num_live_instances_;
-    is_live_ = x.is_live_;
-    x.is_live_ = false;
-    ++num_moves_;
-    return *this;
-  }
-
-  bool operator==(const BaseCountedInstance& x) const {
-    ++num_comparisons_;
-    return value_ == x.value_;
-  }
-
-  bool operator!=(const BaseCountedInstance& x) const {
-    ++num_comparisons_;
-    return value_ != x.value_;
-  }
-
-  bool operator<(const BaseCountedInstance& x) const {
-    ++num_comparisons_;
-    return value_ < x.value_;
-  }
-
-  bool operator>(const BaseCountedInstance& x) const {
-    ++num_comparisons_;
-    return value_ > x.value_;
-  }
-
-  bool operator<=(const BaseCountedInstance& x) const {
-    ++num_comparisons_;
-    return value_ <= x.value_;
-  }
-
-  bool operator>=(const BaseCountedInstance& x) const {
-    ++num_comparisons_;
-    return value_ >= x.value_;
-  }
-
-  absl::weak_ordering compare(const BaseCountedInstance& x) const {
-    ++num_comparisons_;
-    return value_ < x.value_
-               ? absl::weak_ordering::less
-               : value_ == x.value_ ? absl::weak_ordering::equivalent
-                                    : absl::weak_ordering::greater;
-  }
-
-  int value() const {
-    if (!is_live_) std::abort();
-    return value_;
-  }
-
-  friend std::ostream& operator<<(std::ostream& o,
-                                  const BaseCountedInstance& v) {
-    return o << "[value:" << v.value() << "]";
-  }
-
-  // Implementation of efficient swap() that counts swaps.
-  static void SwapImpl(
-      BaseCountedInstance& lhs,    // NOLINT(runtime/references)
-      BaseCountedInstance& rhs) {  // NOLINT(runtime/references)
-    using std::swap;
-    swap(lhs.value_, rhs.value_);
-    swap(lhs.is_live_, rhs.is_live_);
-    ++BaseCountedInstance::num_swaps_;
-  }
-
- private:
-  friend class InstanceTracker;
-
-  int value_;
-
-  // Indicates if the value is live, ie it hasn't been moved away from.
-  bool is_live_ = true;
-
-  // Number of instances.
-  static int num_instances_;
-
-  // Number of live instances (those that have not been moved away from.)
-  static int num_live_instances_;
-
-  // Number of times that BaseCountedInstance objects were moved.
-  static int num_moves_;
-
-  // Number of times that BaseCountedInstance objects were copied.
-  static int num_copies_;
-
-  // Number of times that BaseCountedInstance objects were swapped.
-  static int num_swaps_;
-
-  // Number of times that BaseCountedInstance objects were compared.
-  static int num_comparisons_;
-};
-
-// Helper to track the BaseCountedInstance instance counters. Expects that the
-// number of instances and live_instances are the same when it is constructed
-// and when it is destructed.
-class InstanceTracker {
- public:
-  InstanceTracker()
-      : start_instances_(BaseCountedInstance::num_instances_),
-        start_live_instances_(BaseCountedInstance::num_live_instances_) {
-    ResetCopiesMovesSwaps();
-  }
-  ~InstanceTracker() {
-    if (instances() != 0) std::abort();
-    if (live_instances() != 0) std::abort();
-  }
-
-  // Returns the number of BaseCountedInstance instances both containing valid
-  // values and those moved away from compared to when the InstanceTracker was
-  // constructed
-  int instances() const {
-    return BaseCountedInstance::num_instances_ - start_instances_;
-  }
-
-  // Returns the number of live BaseCountedInstance instances compared to when
-  // the InstanceTracker was constructed
-  int live_instances() const {
-    return BaseCountedInstance::num_live_instances_ - start_live_instances_;
-  }
-
-  // Returns the number of moves on BaseCountedInstance objects since
-  // construction or since the last call to ResetCopiesMovesSwaps().
-  int moves() const { return BaseCountedInstance::num_moves_ - start_moves_; }
-
-  // Returns the number of copies on BaseCountedInstance objects since
-  // construction or the last call to ResetCopiesMovesSwaps().
-  int copies() const {
-    return BaseCountedInstance::num_copies_ - start_copies_;
-  }
-
-  // Returns the number of swaps on BaseCountedInstance objects since
-  // construction or the last call to ResetCopiesMovesSwaps().
-  int swaps() const { return BaseCountedInstance::num_swaps_ - start_swaps_; }
-
-  // Returns the number of comparisons on BaseCountedInstance objects since
-  // construction or the last call to ResetCopiesMovesSwaps().
-  int comparisons() const {
-    return BaseCountedInstance::num_comparisons_ - start_comparisons_;
-  }
-
-  // Resets the base values for moves, copies, comparisons, and swaps to the
-  // current values, so that subsequent Get*() calls for moves, copies,
-  // comparisons, and swaps will compare to the situation at the point of this
-  // call.
-  void ResetCopiesMovesSwaps() {
-    start_moves_ = BaseCountedInstance::num_moves_;
-    start_copies_ = BaseCountedInstance::num_copies_;
-    start_swaps_ = BaseCountedInstance::num_swaps_;
-    start_comparisons_ = BaseCountedInstance::num_comparisons_;
-  }
-
- private:
-  int start_instances_;
-  int start_live_instances_;
-  int start_moves_;
-  int start_copies_;
-  int start_swaps_;
-  int start_comparisons_;
-};
-
-// Copyable, not movable.
-class CopyableOnlyInstance : public BaseCountedInstance {
- public:
-  explicit CopyableOnlyInstance(int x) : BaseCountedInstance(x) {}
-  CopyableOnlyInstance(const CopyableOnlyInstance& rhs) = default;
-  CopyableOnlyInstance& operator=(const CopyableOnlyInstance& rhs) = default;
-
-  friend void swap(CopyableOnlyInstance& lhs, CopyableOnlyInstance& rhs) {
-    BaseCountedInstance::SwapImpl(lhs, rhs);
-  }
-
-  static bool supports_move() { return false; }
-};
-
-// Copyable and movable.
-class CopyableMovableInstance : public BaseCountedInstance {
- public:
-  explicit CopyableMovableInstance(int x) : BaseCountedInstance(x) {}
-  CopyableMovableInstance(const CopyableMovableInstance& rhs) = default;
-  CopyableMovableInstance(CopyableMovableInstance&& rhs) = default;
-  CopyableMovableInstance& operator=(const CopyableMovableInstance& rhs) =
-      default;
-  CopyableMovableInstance& operator=(CopyableMovableInstance&& rhs) = default;
-
-  friend void swap(CopyableMovableInstance& lhs, CopyableMovableInstance& rhs) {
-    BaseCountedInstance::SwapImpl(lhs, rhs);
-  }
-
-  static bool supports_move() { return true; }
-};
-
-// Only movable, not default-constructible.
-class MovableOnlyInstance : public BaseCountedInstance {
- public:
-  explicit MovableOnlyInstance(int x) : BaseCountedInstance(x) {}
-  MovableOnlyInstance(MovableOnlyInstance&& other) = default;
-  MovableOnlyInstance& operator=(MovableOnlyInstance&& other) = default;
-
-  friend void swap(MovableOnlyInstance& lhs, MovableOnlyInstance& rhs) {
-    BaseCountedInstance::SwapImpl(lhs, rhs);
-  }
-
-  static bool supports_move() { return true; }
-};
-
-}  // namespace test_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_TEST_INSTANCE_TRACKER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/tracked.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/tracked.h
deleted file mode 100644
index 29f5829f..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/tracked.h
+++ /dev/null
@@ -1,83 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_TRACKED_H_
-#define ABSL_CONTAINER_INTERNAL_TRACKED_H_
-
-#include <stddef.h>
-
-#include <memory>
-#include <utility>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-// A class that tracks its copies and moves so that it can be queried in tests.
-template <class T>
-class Tracked {
- public:
-  Tracked() {}
-  // NOLINTNEXTLINE(runtime/explicit)
-  Tracked(const T& val) : val_(val) {}
-  Tracked(const Tracked& that)
-      : val_(that.val_),
-        num_moves_(that.num_moves_),
-        num_copies_(that.num_copies_) {
-    ++(*num_copies_);
-  }
-  Tracked(Tracked&& that)
-      : val_(std::move(that.val_)),
-        num_moves_(std::move(that.num_moves_)),
-        num_copies_(std::move(that.num_copies_)) {
-    ++(*num_moves_);
-  }
-  Tracked& operator=(const Tracked& that) {
-    val_ = that.val_;
-    num_moves_ = that.num_moves_;
-    num_copies_ = that.num_copies_;
-    ++(*num_copies_);
-  }
-  Tracked& operator=(Tracked&& that) {
-    val_ = std::move(that.val_);
-    num_moves_ = std::move(that.num_moves_);
-    num_copies_ = std::move(that.num_copies_);
-    ++(*num_moves_);
-  }
-
-  const T& val() const { return val_; }
-
-  friend bool operator==(const Tracked& a, const Tracked& b) {
-    return a.val_ == b.val_;
-  }
-  friend bool operator!=(const Tracked& a, const Tracked& b) {
-    return !(a == b);
-  }
-
-  size_t num_copies() { return *num_copies_; }
-  size_t num_moves() { return *num_moves_; }
-
- private:
-  T val_;
-  std::shared_ptr<size_t> num_moves_ = std::make_shared<size_t>(0);
-  std::shared_ptr<size_t> num_copies_ = std::make_shared<size_t>(0);
-};
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_TRACKED_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_constructor_test.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_constructor_test.h
deleted file mode 100644
index 7e84dc25..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_constructor_test.h
+++ /dev/null
@@ -1,494 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_UNORDERED_MAP_CONSTRUCTOR_TEST_H_
-#define ABSL_CONTAINER_INTERNAL_UNORDERED_MAP_CONSTRUCTOR_TEST_H_
-
-#include <algorithm>
-#include <unordered_map>
-#include <vector>
-
-#include "gmock/gmock.h"
-#include "gtest/gtest.h"
-#include "absl/container/internal/hash_generator_testing.h"
-#include "absl/container/internal/hash_policy_testing.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-template <class UnordMap>
-class ConstructorTest : public ::testing::Test {};
-
-TYPED_TEST_SUITE_P(ConstructorTest);
-
-TYPED_TEST_P(ConstructorTest, NoArgs) {
-  TypeParam m;
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(m, ::testing::UnorderedElementsAre());
-}
-
-TYPED_TEST_P(ConstructorTest, BucketCount) {
-  TypeParam m(123);
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(m, ::testing::UnorderedElementsAre());
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, BucketCountHash) {
-  using H = typename TypeParam::hasher;
-  H hasher;
-  TypeParam m(123, hasher);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(m, ::testing::UnorderedElementsAre());
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, BucketCountHashEqual) {
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  H hasher;
-  E equal;
-  TypeParam m(123, hasher, equal);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_EQ(m.key_eq(), equal);
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(m, ::testing::UnorderedElementsAre());
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, BucketCountHashEqualAlloc) {
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  TypeParam m(123, hasher, equal, alloc);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_EQ(m.key_eq(), equal);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(m, ::testing::UnorderedElementsAre());
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-template <typename T>
-struct is_std_unordered_map : std::false_type {};
-
-template <typename... T>
-struct is_std_unordered_map<std::unordered_map<T...>> : std::true_type {};
-
-#if defined(UNORDERED_MAP_CXX14) || defined(UNORDERED_MAP_CXX17)
-using has_cxx14_std_apis = std::true_type;
-#else
-using has_cxx14_std_apis = std::false_type;
-#endif
-
-template <typename T>
-using expect_cxx14_apis =
-    absl::disjunction<absl::negation<is_std_unordered_map<T>>,
-                      has_cxx14_std_apis>;
-
-template <typename TypeParam>
-void BucketCountAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void BucketCountAllocTest(std::true_type) {
-  using A = typename TypeParam::allocator_type;
-  A alloc(0);
-  TypeParam m(123, alloc);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(m, ::testing::UnorderedElementsAre());
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, BucketCountAlloc) {
-  BucketCountAllocTest<TypeParam>(expect_cxx14_apis<TypeParam>());
-}
-
-template <typename TypeParam>
-void BucketCountHashAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void BucketCountHashAllocTest(std::true_type) {
-  using H = typename TypeParam::hasher;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  A alloc(0);
-  TypeParam m(123, hasher, alloc);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(m, ::testing::UnorderedElementsAre());
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, BucketCountHashAlloc) {
-  BucketCountHashAllocTest<TypeParam>(expect_cxx14_apis<TypeParam>());
-}
-
-#if ABSL_UNORDERED_SUPPORTS_ALLOC_CTORS
-using has_alloc_std_constructors = std::true_type;
-#else
-using has_alloc_std_constructors = std::false_type;
-#endif
-
-template <typename T>
-using expect_alloc_constructors =
-    absl::disjunction<absl::negation<is_std_unordered_map<T>>,
-                      has_alloc_std_constructors>;
-
-template <typename TypeParam>
-void AllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void AllocTest(std::true_type) {
-  using A = typename TypeParam::allocator_type;
-  A alloc(0);
-  TypeParam m(alloc);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(m, ::testing::UnorderedElementsAre());
-}
-
-TYPED_TEST_P(ConstructorTest, Alloc) {
-  AllocTest<TypeParam>(expect_alloc_constructors<TypeParam>());
-}
-
-TYPED_TEST_P(ConstructorTest, InputIteratorBucketHashEqualAlloc) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::UniqueGenerator<T>());
-  TypeParam m(values.begin(), values.end(), 123, hasher, equal, alloc);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_EQ(m.key_eq(), equal);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_THAT(items(m), ::testing::UnorderedElementsAreArray(values));
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-template <typename TypeParam>
-void InputIteratorBucketAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void InputIteratorBucketAllocTest(std::true_type) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using A = typename TypeParam::allocator_type;
-  A alloc(0);
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::UniqueGenerator<T>());
-  TypeParam m(values.begin(), values.end(), 123, alloc);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_THAT(items(m), ::testing::UnorderedElementsAreArray(values));
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, InputIteratorBucketAlloc) {
-  InputIteratorBucketAllocTest<TypeParam>(expect_cxx14_apis<TypeParam>());
-}
-
-template <typename TypeParam>
-void InputIteratorBucketHashAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void InputIteratorBucketHashAllocTest(std::true_type) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  A alloc(0);
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::UniqueGenerator<T>());
-  TypeParam m(values.begin(), values.end(), 123, hasher, alloc);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_THAT(items(m), ::testing::UnorderedElementsAreArray(values));
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, InputIteratorBucketHashAlloc) {
-  InputIteratorBucketHashAllocTest<TypeParam>(expect_cxx14_apis<TypeParam>());
-}
-
-TYPED_TEST_P(ConstructorTest, CopyConstructor) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  hash_internal::UniqueGenerator<T> gen;
-  TypeParam m(123, hasher, equal, alloc);
-  for (size_t i = 0; i != 10; ++i) m.insert(gen());
-  TypeParam n(m);
-  EXPECT_EQ(m.hash_function(), n.hash_function());
-  EXPECT_EQ(m.key_eq(), n.key_eq());
-  EXPECT_EQ(m.get_allocator(), n.get_allocator());
-  EXPECT_EQ(m, n);
-}
-
-template <typename TypeParam>
-void CopyConstructorAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void CopyConstructorAllocTest(std::true_type) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  hash_internal::UniqueGenerator<T> gen;
-  TypeParam m(123, hasher, equal, alloc);
-  for (size_t i = 0; i != 10; ++i) m.insert(gen());
-  TypeParam n(m, A(11));
-  EXPECT_EQ(m.hash_function(), n.hash_function());
-  EXPECT_EQ(m.key_eq(), n.key_eq());
-  EXPECT_NE(m.get_allocator(), n.get_allocator());
-  EXPECT_EQ(m, n);
-}
-
-TYPED_TEST_P(ConstructorTest, CopyConstructorAlloc) {
-  CopyConstructorAllocTest<TypeParam>(expect_alloc_constructors<TypeParam>());
-}
-
-// TODO(alkis): Test non-propagating allocators on copy constructors.
-
-TYPED_TEST_P(ConstructorTest, MoveConstructor) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  hash_internal::UniqueGenerator<T> gen;
-  TypeParam m(123, hasher, equal, alloc);
-  for (size_t i = 0; i != 10; ++i) m.insert(gen());
-  TypeParam t(m);
-  TypeParam n(std::move(t));
-  EXPECT_EQ(m.hash_function(), n.hash_function());
-  EXPECT_EQ(m.key_eq(), n.key_eq());
-  EXPECT_EQ(m.get_allocator(), n.get_allocator());
-  EXPECT_EQ(m, n);
-}
-
-template <typename TypeParam>
-void MoveConstructorAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void MoveConstructorAllocTest(std::true_type) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  hash_internal::UniqueGenerator<T> gen;
-  TypeParam m(123, hasher, equal, alloc);
-  for (size_t i = 0; i != 10; ++i) m.insert(gen());
-  TypeParam t(m);
-  TypeParam n(std::move(t), A(1));
-  EXPECT_EQ(m.hash_function(), n.hash_function());
-  EXPECT_EQ(m.key_eq(), n.key_eq());
-  EXPECT_NE(m.get_allocator(), n.get_allocator());
-  EXPECT_EQ(m, n);
-}
-
-TYPED_TEST_P(ConstructorTest, MoveConstructorAlloc) {
-  MoveConstructorAllocTest<TypeParam>(expect_alloc_constructors<TypeParam>());
-}
-
-// TODO(alkis): Test non-propagating allocators on move constructors.
-
-TYPED_TEST_P(ConstructorTest, InitializerListBucketHashEqualAlloc) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  hash_internal::UniqueGenerator<T> gen;
-  std::initializer_list<T> values = {gen(), gen(), gen(), gen(), gen()};
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  TypeParam m(values, 123, hasher, equal, alloc);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_EQ(m.key_eq(), equal);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_THAT(items(m), ::testing::UnorderedElementsAreArray(values));
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-template <typename TypeParam>
-void InitializerListBucketAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void InitializerListBucketAllocTest(std::true_type) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using A = typename TypeParam::allocator_type;
-  hash_internal::UniqueGenerator<T> gen;
-  std::initializer_list<T> values = {gen(), gen(), gen(), gen(), gen()};
-  A alloc(0);
-  TypeParam m(values, 123, alloc);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_THAT(items(m), ::testing::UnorderedElementsAreArray(values));
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, InitializerListBucketAlloc) {
-  InitializerListBucketAllocTest<TypeParam>(expect_cxx14_apis<TypeParam>());
-}
-
-template <typename TypeParam>
-void InitializerListBucketHashAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void InitializerListBucketHashAllocTest(std::true_type) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  A alloc(0);
-  hash_internal::UniqueGenerator<T> gen;
-  std::initializer_list<T> values = {gen(), gen(), gen(), gen(), gen()};
-  TypeParam m(values, 123, hasher, alloc);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_THAT(items(m), ::testing::UnorderedElementsAreArray(values));
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, InitializerListBucketHashAlloc) {
-  InitializerListBucketHashAllocTest<TypeParam>(expect_cxx14_apis<TypeParam>());
-}
-
-TYPED_TEST_P(ConstructorTest, Assignment) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  hash_internal::UniqueGenerator<T> gen;
-  TypeParam m({gen(), gen(), gen()}, 123, hasher, equal, alloc);
-  TypeParam n;
-  n = m;
-  EXPECT_EQ(m.hash_function(), n.hash_function());
-  EXPECT_EQ(m.key_eq(), n.key_eq());
-  EXPECT_EQ(m, n);
-}
-
-// TODO(alkis): Test [non-]propagating allocators on move/copy assignments
-// (it depends on traits).
-
-TYPED_TEST_P(ConstructorTest, MoveAssignment) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  hash_internal::UniqueGenerator<T> gen;
-  TypeParam m({gen(), gen(), gen()}, 123, hasher, equal, alloc);
-  TypeParam t(m);
-  TypeParam n;
-  n = std::move(t);
-  EXPECT_EQ(m.hash_function(), n.hash_function());
-  EXPECT_EQ(m.key_eq(), n.key_eq());
-  EXPECT_EQ(m, n);
-}
-
-TYPED_TEST_P(ConstructorTest, AssignmentFromInitializerList) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  hash_internal::UniqueGenerator<T> gen;
-  std::initializer_list<T> values = {gen(), gen(), gen(), gen(), gen()};
-  TypeParam m;
-  m = values;
-  EXPECT_THAT(items(m), ::testing::UnorderedElementsAreArray(values));
-}
-
-TYPED_TEST_P(ConstructorTest, AssignmentOverwritesExisting) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  hash_internal::UniqueGenerator<T> gen;
-  TypeParam m({gen(), gen(), gen()});
-  TypeParam n({gen()});
-  n = m;
-  EXPECT_EQ(m, n);
-}
-
-TYPED_TEST_P(ConstructorTest, MoveAssignmentOverwritesExisting) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  hash_internal::UniqueGenerator<T> gen;
-  TypeParam m({gen(), gen(), gen()});
-  TypeParam t(m);
-  TypeParam n({gen()});
-  n = std::move(t);
-  EXPECT_EQ(m, n);
-}
-
-TYPED_TEST_P(ConstructorTest, AssignmentFromInitializerListOverwritesExisting) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  hash_internal::UniqueGenerator<T> gen;
-  std::initializer_list<T> values = {gen(), gen(), gen(), gen(), gen()};
-  TypeParam m;
-  m = values;
-  EXPECT_THAT(items(m), ::testing::UnorderedElementsAreArray(values));
-}
-
-TYPED_TEST_P(ConstructorTest, AssignmentOnSelf) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  hash_internal::UniqueGenerator<T> gen;
-  std::initializer_list<T> values = {gen(), gen(), gen(), gen(), gen()};
-  TypeParam m(values);
-  m = *&m;  // Avoid -Wself-assign
-  EXPECT_THAT(items(m), ::testing::UnorderedElementsAreArray(values));
-}
-
-// We cannot test self move as standard states that it leaves standard
-// containers in unspecified state (and in practice in causes memory-leak
-// according to heap-checker!).
-
-REGISTER_TYPED_TEST_SUITE_P(
-    ConstructorTest, NoArgs, BucketCount, BucketCountHash, BucketCountHashEqual,
-    BucketCountHashEqualAlloc, BucketCountAlloc, BucketCountHashAlloc, Alloc,
-    InputIteratorBucketHashEqualAlloc, InputIteratorBucketAlloc,
-    InputIteratorBucketHashAlloc, CopyConstructor, CopyConstructorAlloc,
-    MoveConstructor, MoveConstructorAlloc, InitializerListBucketHashEqualAlloc,
-    InitializerListBucketAlloc, InitializerListBucketHashAlloc, Assignment,
-    MoveAssignment, AssignmentFromInitializerList, AssignmentOverwritesExisting,
-    MoveAssignmentOverwritesExisting,
-    AssignmentFromInitializerListOverwritesExisting, AssignmentOnSelf);
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_UNORDERED_MAP_CONSTRUCTOR_TEST_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_lookup_test.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_lookup_test.h
deleted file mode 100644
index 3713cd9a..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_lookup_test.h
+++ /dev/null
@@ -1,117 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_UNORDERED_MAP_LOOKUP_TEST_H_
-#define ABSL_CONTAINER_INTERNAL_UNORDERED_MAP_LOOKUP_TEST_H_
-
-#include "gmock/gmock.h"
-#include "gtest/gtest.h"
-#include "absl/container/internal/hash_generator_testing.h"
-#include "absl/container/internal/hash_policy_testing.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-template <class UnordMap>
-class LookupTest : public ::testing::Test {};
-
-TYPED_TEST_SUITE_P(LookupTest);
-
-TYPED_TEST_P(LookupTest, At) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m(values.begin(), values.end());
-  for (const auto& p : values) {
-    const auto& val = m.at(p.first);
-    EXPECT_EQ(p.second, val) << ::testing::PrintToString(p.first);
-  }
-}
-
-TYPED_TEST_P(LookupTest, OperatorBracket) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using V = typename TypeParam::mapped_type;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m;
-  for (const auto& p : values) {
-    auto& val = m[p.first];
-    EXPECT_EQ(V(), val) << ::testing::PrintToString(p.first);
-    val = p.second;
-  }
-  for (const auto& p : values)
-    EXPECT_EQ(p.second, m[p.first]) << ::testing::PrintToString(p.first);
-}
-
-TYPED_TEST_P(LookupTest, Count) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m;
-  for (const auto& p : values)
-    EXPECT_EQ(0, m.count(p.first)) << ::testing::PrintToString(p.first);
-  m.insert(values.begin(), values.end());
-  for (const auto& p : values)
-    EXPECT_EQ(1, m.count(p.first)) << ::testing::PrintToString(p.first);
-}
-
-TYPED_TEST_P(LookupTest, Find) {
-  using std::get;
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m;
-  for (const auto& p : values)
-    EXPECT_TRUE(m.end() == m.find(p.first))
-        << ::testing::PrintToString(p.first);
-  m.insert(values.begin(), values.end());
-  for (const auto& p : values) {
-    auto it = m.find(p.first);
-    EXPECT_TRUE(m.end() != it) << ::testing::PrintToString(p.first);
-    EXPECT_EQ(p.second, get<1>(*it)) << ::testing::PrintToString(p.first);
-  }
-}
-
-TYPED_TEST_P(LookupTest, EqualRange) {
-  using std::get;
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m;
-  for (const auto& p : values) {
-    auto r = m.equal_range(p.first);
-    ASSERT_EQ(0, std::distance(r.first, r.second));
-  }
-  m.insert(values.begin(), values.end());
-  for (const auto& p : values) {
-    auto r = m.equal_range(p.first);
-    ASSERT_EQ(1, std::distance(r.first, r.second));
-    EXPECT_EQ(p.second, get<1>(*r.first)) << ::testing::PrintToString(p.first);
-  }
-}
-
-REGISTER_TYPED_TEST_SUITE_P(LookupTest, At, OperatorBracket, Count, Find,
-                            EqualRange);
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_UNORDERED_MAP_LOOKUP_TEST_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_members_test.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_members_test.h
deleted file mode 100644
index 7d48cdb8..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_members_test.h
+++ /dev/null
@@ -1,87 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_UNORDERED_MAP_MEMBERS_TEST_H_
-#define ABSL_CONTAINER_INTERNAL_UNORDERED_MAP_MEMBERS_TEST_H_
-
-#include <type_traits>
-#include "gmock/gmock.h"
-#include "gtest/gtest.h"
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-template <class UnordMap>
-class MembersTest : public ::testing::Test {};
-
-TYPED_TEST_SUITE_P(MembersTest);
-
-template <typename T>
-void UseType() {}
-
-TYPED_TEST_P(MembersTest, Typedefs) {
-  EXPECT_TRUE((std::is_same<std::pair<const typename TypeParam::key_type,
-                                      typename TypeParam::mapped_type>,
-                            typename TypeParam::value_type>()));
-  EXPECT_TRUE((absl::conjunction<
-               absl::negation<std::is_signed<typename TypeParam::size_type>>,
-               std::is_integral<typename TypeParam::size_type>>()));
-  EXPECT_TRUE((absl::conjunction<
-               std::is_signed<typename TypeParam::difference_type>,
-               std::is_integral<typename TypeParam::difference_type>>()));
-  EXPECT_TRUE((std::is_convertible<
-               decltype(std::declval<const typename TypeParam::hasher&>()(
-                   std::declval<const typename TypeParam::key_type&>())),
-               size_t>()));
-  EXPECT_TRUE((std::is_convertible<
-               decltype(std::declval<const typename TypeParam::key_equal&>()(
-                   std::declval<const typename TypeParam::key_type&>(),
-                   std::declval<const typename TypeParam::key_type&>())),
-               bool>()));
-  EXPECT_TRUE((std::is_same<typename TypeParam::allocator_type::value_type,
-                            typename TypeParam::value_type>()));
-  EXPECT_TRUE((std::is_same<typename TypeParam::value_type&,
-                            typename TypeParam::reference>()));
-  EXPECT_TRUE((std::is_same<const typename TypeParam::value_type&,
-                            typename TypeParam::const_reference>()));
-  EXPECT_TRUE((std::is_same<typename std::allocator_traits<
-                                typename TypeParam::allocator_type>::pointer,
-                            typename TypeParam::pointer>()));
-  EXPECT_TRUE(
-      (std::is_same<typename std::allocator_traits<
-                        typename TypeParam::allocator_type>::const_pointer,
-                    typename TypeParam::const_pointer>()));
-}
-
-TYPED_TEST_P(MembersTest, SimpleFunctions) {
-  EXPECT_GT(TypeParam().max_size(), 0);
-}
-
-TYPED_TEST_P(MembersTest, BeginEnd) {
-  TypeParam t = {typename TypeParam::value_type{}};
-  EXPECT_EQ(t.begin(), t.cbegin());
-  EXPECT_EQ(t.end(), t.cend());
-  EXPECT_NE(t.begin(), t.end());
-  EXPECT_NE(t.cbegin(), t.cend());
-}
-
-REGISTER_TYPED_TEST_SUITE_P(MembersTest, Typedefs, SimpleFunctions, BeginEnd);
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_UNORDERED_MAP_MEMBERS_TEST_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_modifiers_test.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_modifiers_test.h
deleted file mode 100644
index 4d9ab30f..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_map_modifiers_test.h
+++ /dev/null
@@ -1,352 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_UNORDERED_MAP_MODIFIERS_TEST_H_
-#define ABSL_CONTAINER_INTERNAL_UNORDERED_MAP_MODIFIERS_TEST_H_
-
-#include <memory>
-
-#include "gmock/gmock.h"
-#include "gtest/gtest.h"
-#include "absl/container/internal/hash_generator_testing.h"
-#include "absl/container/internal/hash_policy_testing.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-template <class UnordMap>
-class ModifiersTest : public ::testing::Test {};
-
-TYPED_TEST_SUITE_P(ModifiersTest);
-
-TYPED_TEST_P(ModifiersTest, Clear) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m(values.begin(), values.end());
-  ASSERT_THAT(items(m), ::testing::UnorderedElementsAreArray(values));
-  m.clear();
-  EXPECT_THAT(items(m), ::testing::UnorderedElementsAre());
-  EXPECT_TRUE(m.empty());
-}
-
-TYPED_TEST_P(ModifiersTest, Insert) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using V = typename TypeParam::mapped_type;
-  T val = hash_internal::Generator<T>()();
-  TypeParam m;
-  auto p = m.insert(val);
-  EXPECT_TRUE(p.second);
-  EXPECT_EQ(val, *p.first);
-  T val2 = {val.first, hash_internal::Generator<V>()()};
-  p = m.insert(val2);
-  EXPECT_FALSE(p.second);
-  EXPECT_EQ(val, *p.first);
-}
-
-TYPED_TEST_P(ModifiersTest, InsertHint) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using V = typename TypeParam::mapped_type;
-  T val = hash_internal::Generator<T>()();
-  TypeParam m;
-  auto it = m.insert(m.end(), val);
-  EXPECT_TRUE(it != m.end());
-  EXPECT_EQ(val, *it);
-  T val2 = {val.first, hash_internal::Generator<V>()()};
-  it = m.insert(it, val2);
-  EXPECT_TRUE(it != m.end());
-  EXPECT_EQ(val, *it);
-}
-
-TYPED_TEST_P(ModifiersTest, InsertRange) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m;
-  m.insert(values.begin(), values.end());
-  ASSERT_THAT(items(m), ::testing::UnorderedElementsAreArray(values));
-}
-
-TYPED_TEST_P(ModifiersTest, InsertWithinCapacity) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using V = typename TypeParam::mapped_type;
-  T val = hash_internal::Generator<T>()();
-  TypeParam m;
-  m.reserve(10);
-  const size_t original_capacity = m.bucket_count();
-  m.insert(val);
-  EXPECT_EQ(m.bucket_count(), original_capacity);
-  T val2 = {val.first, hash_internal::Generator<V>()()};
-  m.insert(val2);
-  EXPECT_EQ(m.bucket_count(), original_capacity);
-}
-
-TYPED_TEST_P(ModifiersTest, InsertRangeWithinCapacity) {
-#if !defined(__GLIBCXX__)
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> base_values;
-  std::generate_n(std::back_inserter(base_values), 10,
-                  hash_internal::Generator<T>());
-  std::vector<T> values;
-  while (values.size() != 100) {
-    std::copy_n(base_values.begin(), 10, std::back_inserter(values));
-  }
-  TypeParam m;
-  m.reserve(10);
-  const size_t original_capacity = m.bucket_count();
-  m.insert(values.begin(), values.end());
-  EXPECT_EQ(m.bucket_count(), original_capacity);
-#endif
-}
-
-TYPED_TEST_P(ModifiersTest, InsertOrAssign) {
-#ifdef UNORDERED_MAP_CXX17
-  using std::get;
-  using K = typename TypeParam::key_type;
-  using V = typename TypeParam::mapped_type;
-  K k = hash_internal::Generator<K>()();
-  V val = hash_internal::Generator<V>()();
-  TypeParam m;
-  auto p = m.insert_or_assign(k, val);
-  EXPECT_TRUE(p.second);
-  EXPECT_EQ(k, get<0>(*p.first));
-  EXPECT_EQ(val, get<1>(*p.first));
-  V val2 = hash_internal::Generator<V>()();
-  p = m.insert_or_assign(k, val2);
-  EXPECT_FALSE(p.second);
-  EXPECT_EQ(k, get<0>(*p.first));
-  EXPECT_EQ(val2, get<1>(*p.first));
-#endif
-}
-
-TYPED_TEST_P(ModifiersTest, InsertOrAssignHint) {
-#ifdef UNORDERED_MAP_CXX17
-  using std::get;
-  using K = typename TypeParam::key_type;
-  using V = typename TypeParam::mapped_type;
-  K k = hash_internal::Generator<K>()();
-  V val = hash_internal::Generator<V>()();
-  TypeParam m;
-  auto it = m.insert_or_assign(m.end(), k, val);
-  EXPECT_TRUE(it != m.end());
-  EXPECT_EQ(k, get<0>(*it));
-  EXPECT_EQ(val, get<1>(*it));
-  V val2 = hash_internal::Generator<V>()();
-  it = m.insert_or_assign(it, k, val2);
-  EXPECT_EQ(k, get<0>(*it));
-  EXPECT_EQ(val2, get<1>(*it));
-#endif
-}
-
-TYPED_TEST_P(ModifiersTest, Emplace) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using V = typename TypeParam::mapped_type;
-  T val = hash_internal::Generator<T>()();
-  TypeParam m;
-  // TODO(alkis): We need a way to run emplace in a more meaningful way. Perhaps
-  // with test traits/policy.
-  auto p = m.emplace(val);
-  EXPECT_TRUE(p.second);
-  EXPECT_EQ(val, *p.first);
-  T val2 = {val.first, hash_internal::Generator<V>()()};
-  p = m.emplace(val2);
-  EXPECT_FALSE(p.second);
-  EXPECT_EQ(val, *p.first);
-}
-
-TYPED_TEST_P(ModifiersTest, EmplaceHint) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using V = typename TypeParam::mapped_type;
-  T val = hash_internal::Generator<T>()();
-  TypeParam m;
-  // TODO(alkis): We need a way to run emplace in a more meaningful way. Perhaps
-  // with test traits/policy.
-  auto it = m.emplace_hint(m.end(), val);
-  EXPECT_EQ(val, *it);
-  T val2 = {val.first, hash_internal::Generator<V>()()};
-  it = m.emplace_hint(it, val2);
-  EXPECT_EQ(val, *it);
-}
-
-TYPED_TEST_P(ModifiersTest, TryEmplace) {
-#ifdef UNORDERED_MAP_CXX17
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using V = typename TypeParam::mapped_type;
-  T val = hash_internal::Generator<T>()();
-  TypeParam m;
-  // TODO(alkis): We need a way to run emplace in a more meaningful way. Perhaps
-  // with test traits/policy.
-  auto p = m.try_emplace(val.first, val.second);
-  EXPECT_TRUE(p.second);
-  EXPECT_EQ(val, *p.first);
-  T val2 = {val.first, hash_internal::Generator<V>()()};
-  p = m.try_emplace(val2.first, val2.second);
-  EXPECT_FALSE(p.second);
-  EXPECT_EQ(val, *p.first);
-#endif
-}
-
-TYPED_TEST_P(ModifiersTest, TryEmplaceHint) {
-#ifdef UNORDERED_MAP_CXX17
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using V = typename TypeParam::mapped_type;
-  T val = hash_internal::Generator<T>()();
-  TypeParam m;
-  // TODO(alkis): We need a way to run emplace in a more meaningful way. Perhaps
-  // with test traits/policy.
-  auto it = m.try_emplace(m.end(), val.first, val.second);
-  EXPECT_EQ(val, *it);
-  T val2 = {val.first, hash_internal::Generator<V>()()};
-  it = m.try_emplace(it, val2.first, val2.second);
-  EXPECT_EQ(val, *it);
-#endif
-}
-
-template <class V>
-using IfNotVoid = typename std::enable_if<!std::is_void<V>::value, V>::type;
-
-// In openmap we chose not to return the iterator from erase because that's
-// more expensive. As such we adapt erase to return an iterator here.
-struct EraseFirst {
-  template <class Map>
-  auto operator()(Map* m, int) const
-      -> IfNotVoid<decltype(m->erase(m->begin()))> {
-    return m->erase(m->begin());
-  }
-  template <class Map>
-  typename Map::iterator operator()(Map* m, ...) const {
-    auto it = m->begin();
-    m->erase(it++);
-    return it;
-  }
-};
-
-TYPED_TEST_P(ModifiersTest, Erase) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using std::get;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m(values.begin(), values.end());
-  ASSERT_THAT(items(m), ::testing::UnorderedElementsAreArray(values));
-  auto& first = *m.begin();
-  std::vector<T> values2;
-  for (const auto& val : values)
-    if (get<0>(val) != get<0>(first)) values2.push_back(val);
-  auto it = EraseFirst()(&m, 0);
-  ASSERT_TRUE(it != m.end());
-  EXPECT_EQ(1, std::count(values2.begin(), values2.end(), *it));
-  EXPECT_THAT(items(m), ::testing::UnorderedElementsAreArray(values2.begin(),
-                                                             values2.end()));
-}
-
-TYPED_TEST_P(ModifiersTest, EraseRange) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m(values.begin(), values.end());
-  ASSERT_THAT(items(m), ::testing::UnorderedElementsAreArray(values));
-  auto it = m.erase(m.begin(), m.end());
-  EXPECT_THAT(items(m), ::testing::UnorderedElementsAre());
-  EXPECT_TRUE(it == m.end());
-}
-
-TYPED_TEST_P(ModifiersTest, EraseKey) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m(values.begin(), values.end());
-  ASSERT_THAT(items(m), ::testing::UnorderedElementsAreArray(values));
-  EXPECT_EQ(1, m.erase(values[0].first));
-  EXPECT_EQ(0, std::count(m.begin(), m.end(), values[0]));
-  EXPECT_THAT(items(m), ::testing::UnorderedElementsAreArray(values.begin() + 1,
-                                                             values.end()));
-}
-
-TYPED_TEST_P(ModifiersTest, Swap) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> v1;
-  std::vector<T> v2;
-  std::generate_n(std::back_inserter(v1), 5, hash_internal::Generator<T>());
-  std::generate_n(std::back_inserter(v2), 5, hash_internal::Generator<T>());
-  TypeParam m1(v1.begin(), v1.end());
-  TypeParam m2(v2.begin(), v2.end());
-  EXPECT_THAT(items(m1), ::testing::UnorderedElementsAreArray(v1));
-  EXPECT_THAT(items(m2), ::testing::UnorderedElementsAreArray(v2));
-  m1.swap(m2);
-  EXPECT_THAT(items(m1), ::testing::UnorderedElementsAreArray(v2));
-  EXPECT_THAT(items(m2), ::testing::UnorderedElementsAreArray(v1));
-}
-
-// TODO(alkis): Write tests for extract.
-// TODO(alkis): Write tests for merge.
-
-REGISTER_TYPED_TEST_SUITE_P(ModifiersTest, Clear, Insert, InsertHint,
-                            InsertRange, InsertWithinCapacity,
-                            InsertRangeWithinCapacity, InsertOrAssign,
-                            InsertOrAssignHint, Emplace, EmplaceHint,
-                            TryEmplace, TryEmplaceHint, Erase, EraseRange,
-                            EraseKey, Swap);
-
-template <typename Type>
-struct is_unique_ptr : std::false_type {};
-
-template <typename Type>
-struct is_unique_ptr<std::unique_ptr<Type>> : std::true_type {};
-
-template <class UnordMap>
-class UniquePtrModifiersTest : public ::testing::Test {
- protected:
-  UniquePtrModifiersTest() {
-    static_assert(is_unique_ptr<typename UnordMap::mapped_type>::value,
-                  "UniquePtrModifiersTyest may only be called with a "
-                  "std::unique_ptr value type.");
-  }
-};
-
-GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(UniquePtrModifiersTest);
-
-TYPED_TEST_SUITE_P(UniquePtrModifiersTest);
-
-// Test that we do not move from rvalue arguments if an insertion does not
-// happen.
-TYPED_TEST_P(UniquePtrModifiersTest, TryEmplace) {
-#ifdef UNORDERED_MAP_CXX17
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using V = typename TypeParam::mapped_type;
-  T val = hash_internal::Generator<T>()();
-  TypeParam m;
-  auto p = m.try_emplace(val.first, std::move(val.second));
-  EXPECT_TRUE(p.second);
-  // A moved from std::unique_ptr is guaranteed to be nullptr.
-  EXPECT_EQ(val.second, nullptr);
-  T val2 = {val.first, hash_internal::Generator<V>()()};
-  p = m.try_emplace(val2.first, std::move(val2.second));
-  EXPECT_FALSE(p.second);
-  EXPECT_NE(val2.second, nullptr);
-#endif
-}
-
-REGISTER_TYPED_TEST_SUITE_P(UniquePtrModifiersTest, TryEmplace);
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_UNORDERED_MAP_MODIFIERS_TEST_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_constructor_test.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_constructor_test.h
deleted file mode 100644
index af1116e6..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_constructor_test.h
+++ /dev/null
@@ -1,496 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_UNORDERED_SET_CONSTRUCTOR_TEST_H_
-#define ABSL_CONTAINER_INTERNAL_UNORDERED_SET_CONSTRUCTOR_TEST_H_
-
-#include <algorithm>
-#include <unordered_set>
-#include <vector>
-
-#include "gmock/gmock.h"
-#include "gtest/gtest.h"
-#include "absl/container/internal/hash_generator_testing.h"
-#include "absl/container/internal/hash_policy_testing.h"
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-template <class UnordMap>
-class ConstructorTest : public ::testing::Test {};
-
-TYPED_TEST_SUITE_P(ConstructorTest);
-
-TYPED_TEST_P(ConstructorTest, NoArgs) {
-  TypeParam m;
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAre());
-}
-
-TYPED_TEST_P(ConstructorTest, BucketCount) {
-  TypeParam m(123);
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAre());
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, BucketCountHash) {
-  using H = typename TypeParam::hasher;
-  H hasher;
-  TypeParam m(123, hasher);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAre());
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, BucketCountHashEqual) {
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  H hasher;
-  E equal;
-  TypeParam m(123, hasher, equal);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_EQ(m.key_eq(), equal);
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAre());
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, BucketCountHashEqualAlloc) {
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  TypeParam m(123, hasher, equal, alloc);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_EQ(m.key_eq(), equal);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAre());
-  EXPECT_GE(m.bucket_count(), 123);
-
-  const auto& cm = m;
-  EXPECT_EQ(cm.hash_function(), hasher);
-  EXPECT_EQ(cm.key_eq(), equal);
-  EXPECT_EQ(cm.get_allocator(), alloc);
-  EXPECT_TRUE(cm.empty());
-  EXPECT_THAT(keys(cm), ::testing::UnorderedElementsAre());
-  EXPECT_GE(cm.bucket_count(), 123);
-}
-
-template <typename T>
-struct is_std_unordered_set : std::false_type {};
-
-template <typename... T>
-struct is_std_unordered_set<std::unordered_set<T...>> : std::true_type {};
-
-#if defined(UNORDERED_SET_CXX14) || defined(UNORDERED_SET_CXX17)
-using has_cxx14_std_apis = std::true_type;
-#else
-using has_cxx14_std_apis = std::false_type;
-#endif
-
-template <typename T>
-using expect_cxx14_apis =
-    absl::disjunction<absl::negation<is_std_unordered_set<T>>,
-                      has_cxx14_std_apis>;
-
-template <typename TypeParam>
-void BucketCountAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void BucketCountAllocTest(std::true_type) {
-  using A = typename TypeParam::allocator_type;
-  A alloc(0);
-  TypeParam m(123, alloc);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAre());
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, BucketCountAlloc) {
-  BucketCountAllocTest<TypeParam>(expect_cxx14_apis<TypeParam>());
-}
-
-template <typename TypeParam>
-void BucketCountHashAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void BucketCountHashAllocTest(std::true_type) {
-  using H = typename TypeParam::hasher;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  A alloc(0);
-  TypeParam m(123, hasher, alloc);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAre());
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, BucketCountHashAlloc) {
-  BucketCountHashAllocTest<TypeParam>(expect_cxx14_apis<TypeParam>());
-}
-
-#if ABSL_UNORDERED_SUPPORTS_ALLOC_CTORS
-using has_alloc_std_constructors = std::true_type;
-#else
-using has_alloc_std_constructors = std::false_type;
-#endif
-
-template <typename T>
-using expect_alloc_constructors =
-    absl::disjunction<absl::negation<is_std_unordered_set<T>>,
-                      has_alloc_std_constructors>;
-
-template <typename TypeParam>
-void AllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void AllocTest(std::true_type) {
-  using A = typename TypeParam::allocator_type;
-  A alloc(0);
-  TypeParam m(alloc);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_TRUE(m.empty());
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAre());
-}
-
-TYPED_TEST_P(ConstructorTest, Alloc) {
-  AllocTest<TypeParam>(expect_alloc_constructors<TypeParam>());
-}
-
-TYPED_TEST_P(ConstructorTest, InputIteratorBucketHashEqualAlloc) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  std::vector<T> values;
-  for (size_t i = 0; i != 10; ++i)
-    values.push_back(hash_internal::Generator<T>()());
-  TypeParam m(values.begin(), values.end(), 123, hasher, equal, alloc);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_EQ(m.key_eq(), equal);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values));
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-template <typename TypeParam>
-void InputIteratorBucketAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void InputIteratorBucketAllocTest(std::true_type) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using A = typename TypeParam::allocator_type;
-  A alloc(0);
-  std::vector<T> values;
-  for (size_t i = 0; i != 10; ++i)
-    values.push_back(hash_internal::Generator<T>()());
-  TypeParam m(values.begin(), values.end(), 123, alloc);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values));
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, InputIteratorBucketAlloc) {
-  InputIteratorBucketAllocTest<TypeParam>(expect_cxx14_apis<TypeParam>());
-}
-
-template <typename TypeParam>
-void InputIteratorBucketHashAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void InputIteratorBucketHashAllocTest(std::true_type) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  A alloc(0);
-  std::vector<T> values;
-  for (size_t i = 0; i != 10; ++i)
-    values.push_back(hash_internal::Generator<T>()());
-  TypeParam m(values.begin(), values.end(), 123, hasher, alloc);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values));
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, InputIteratorBucketHashAlloc) {
-  InputIteratorBucketHashAllocTest<TypeParam>(expect_cxx14_apis<TypeParam>());
-}
-
-TYPED_TEST_P(ConstructorTest, CopyConstructor) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  TypeParam m(123, hasher, equal, alloc);
-  for (size_t i = 0; i != 10; ++i) m.insert(hash_internal::Generator<T>()());
-  TypeParam n(m);
-  EXPECT_EQ(m.hash_function(), n.hash_function());
-  EXPECT_EQ(m.key_eq(), n.key_eq());
-  EXPECT_EQ(m.get_allocator(), n.get_allocator());
-  EXPECT_EQ(m, n);
-  EXPECT_NE(TypeParam(0, hasher, equal, alloc), n);
-}
-
-template <typename TypeParam>
-void CopyConstructorAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void CopyConstructorAllocTest(std::true_type) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  TypeParam m(123, hasher, equal, alloc);
-  for (size_t i = 0; i != 10; ++i) m.insert(hash_internal::Generator<T>()());
-  TypeParam n(m, A(11));
-  EXPECT_EQ(m.hash_function(), n.hash_function());
-  EXPECT_EQ(m.key_eq(), n.key_eq());
-  EXPECT_NE(m.get_allocator(), n.get_allocator());
-  EXPECT_EQ(m, n);
-}
-
-TYPED_TEST_P(ConstructorTest, CopyConstructorAlloc) {
-  CopyConstructorAllocTest<TypeParam>(expect_alloc_constructors<TypeParam>());
-}
-
-// TODO(alkis): Test non-propagating allocators on copy constructors.
-
-TYPED_TEST_P(ConstructorTest, MoveConstructor) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  TypeParam m(123, hasher, equal, alloc);
-  for (size_t i = 0; i != 10; ++i) m.insert(hash_internal::Generator<T>()());
-  TypeParam t(m);
-  TypeParam n(std::move(t));
-  EXPECT_EQ(m.hash_function(), n.hash_function());
-  EXPECT_EQ(m.key_eq(), n.key_eq());
-  EXPECT_EQ(m.get_allocator(), n.get_allocator());
-  EXPECT_EQ(m, n);
-}
-
-template <typename TypeParam>
-void MoveConstructorAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void MoveConstructorAllocTest(std::true_type) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  TypeParam m(123, hasher, equal, alloc);
-  for (size_t i = 0; i != 10; ++i) m.insert(hash_internal::Generator<T>()());
-  TypeParam t(m);
-  TypeParam n(std::move(t), A(1));
-  EXPECT_EQ(m.hash_function(), n.hash_function());
-  EXPECT_EQ(m.key_eq(), n.key_eq());
-  EXPECT_NE(m.get_allocator(), n.get_allocator());
-  EXPECT_EQ(m, n);
-}
-
-TYPED_TEST_P(ConstructorTest, MoveConstructorAlloc) {
-  MoveConstructorAllocTest<TypeParam>(expect_alloc_constructors<TypeParam>());
-}
-
-// TODO(alkis): Test non-propagating allocators on move constructors.
-
-TYPED_TEST_P(ConstructorTest, InitializerListBucketHashEqualAlloc) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  hash_internal::Generator<T> gen;
-  std::initializer_list<T> values = {gen(), gen(), gen(), gen(), gen()};
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  TypeParam m(values, 123, hasher, equal, alloc);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_EQ(m.key_eq(), equal);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values));
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-template <typename TypeParam>
-void InitializerListBucketAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void InitializerListBucketAllocTest(std::true_type) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using A = typename TypeParam::allocator_type;
-  hash_internal::Generator<T> gen;
-  std::initializer_list<T> values = {gen(), gen(), gen(), gen(), gen()};
-  A alloc(0);
-  TypeParam m(values, 123, alloc);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values));
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, InitializerListBucketAlloc) {
-  InitializerListBucketAllocTest<TypeParam>(expect_cxx14_apis<TypeParam>());
-}
-
-template <typename TypeParam>
-void InitializerListBucketHashAllocTest(std::false_type) {}
-
-template <typename TypeParam>
-void InitializerListBucketHashAllocTest(std::true_type) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  A alloc(0);
-  hash_internal::Generator<T> gen;
-  std::initializer_list<T> values = {gen(), gen(), gen(), gen(), gen()};
-  TypeParam m(values, 123, hasher, alloc);
-  EXPECT_EQ(m.hash_function(), hasher);
-  EXPECT_EQ(m.get_allocator(), alloc);
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values));
-  EXPECT_GE(m.bucket_count(), 123);
-}
-
-TYPED_TEST_P(ConstructorTest, InitializerListBucketHashAlloc) {
-  InitializerListBucketHashAllocTest<TypeParam>(expect_cxx14_apis<TypeParam>());
-}
-
-TYPED_TEST_P(ConstructorTest, CopyAssignment) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  hash_internal::Generator<T> gen;
-  TypeParam m({gen(), gen(), gen()}, 123, hasher, equal, alloc);
-  TypeParam n;
-  n = m;
-  EXPECT_EQ(m.hash_function(), n.hash_function());
-  EXPECT_EQ(m.key_eq(), n.key_eq());
-  EXPECT_EQ(m, n);
-}
-
-// TODO(alkis): Test [non-]propagating allocators on move/copy assignments
-// (it depends on traits).
-
-TYPED_TEST_P(ConstructorTest, MoveAssignment) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  using H = typename TypeParam::hasher;
-  using E = typename TypeParam::key_equal;
-  using A = typename TypeParam::allocator_type;
-  H hasher;
-  E equal;
-  A alloc(0);
-  hash_internal::Generator<T> gen;
-  TypeParam m({gen(), gen(), gen()}, 123, hasher, equal, alloc);
-  TypeParam t(m);
-  TypeParam n;
-  n = std::move(t);
-  EXPECT_EQ(m.hash_function(), n.hash_function());
-  EXPECT_EQ(m.key_eq(), n.key_eq());
-  EXPECT_EQ(m, n);
-}
-
-TYPED_TEST_P(ConstructorTest, AssignmentFromInitializerList) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  hash_internal::Generator<T> gen;
-  std::initializer_list<T> values = {gen(), gen(), gen(), gen(), gen()};
-  TypeParam m;
-  m = values;
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values));
-}
-
-TYPED_TEST_P(ConstructorTest, AssignmentOverwritesExisting) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  hash_internal::Generator<T> gen;
-  TypeParam m({gen(), gen(), gen()});
-  TypeParam n({gen()});
-  n = m;
-  EXPECT_EQ(m, n);
-}
-
-TYPED_TEST_P(ConstructorTest, MoveAssignmentOverwritesExisting) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  hash_internal::Generator<T> gen;
-  TypeParam m({gen(), gen(), gen()});
-  TypeParam t(m);
-  TypeParam n({gen()});
-  n = std::move(t);
-  EXPECT_EQ(m, n);
-}
-
-TYPED_TEST_P(ConstructorTest, AssignmentFromInitializerListOverwritesExisting) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  hash_internal::Generator<T> gen;
-  std::initializer_list<T> values = {gen(), gen(), gen(), gen(), gen()};
-  TypeParam m;
-  m = values;
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values));
-}
-
-TYPED_TEST_P(ConstructorTest, AssignmentOnSelf) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  hash_internal::Generator<T> gen;
-  std::initializer_list<T> values = {gen(), gen(), gen(), gen(), gen()};
-  TypeParam m(values);
-  m = *&m;  // Avoid -Wself-assign.
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values));
-}
-
-REGISTER_TYPED_TEST_SUITE_P(
-    ConstructorTest, NoArgs, BucketCount, BucketCountHash, BucketCountHashEqual,
-    BucketCountHashEqualAlloc, BucketCountAlloc, BucketCountHashAlloc, Alloc,
-    InputIteratorBucketHashEqualAlloc, InputIteratorBucketAlloc,
-    InputIteratorBucketHashAlloc, CopyConstructor, CopyConstructorAlloc,
-    MoveConstructor, MoveConstructorAlloc, InitializerListBucketHashEqualAlloc,
-    InitializerListBucketAlloc, InitializerListBucketHashAlloc, CopyAssignment,
-    MoveAssignment, AssignmentFromInitializerList, AssignmentOverwritesExisting,
-    MoveAssignmentOverwritesExisting,
-    AssignmentFromInitializerListOverwritesExisting, AssignmentOnSelf);
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_UNORDERED_SET_CONSTRUCTOR_TEST_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_lookup_test.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_lookup_test.h
deleted file mode 100644
index b35f766e..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_lookup_test.h
+++ /dev/null
@@ -1,91 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_UNORDERED_SET_LOOKUP_TEST_H_
-#define ABSL_CONTAINER_INTERNAL_UNORDERED_SET_LOOKUP_TEST_H_
-
-#include "gmock/gmock.h"
-#include "gtest/gtest.h"
-#include "absl/container/internal/hash_generator_testing.h"
-#include "absl/container/internal/hash_policy_testing.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-template <class UnordSet>
-class LookupTest : public ::testing::Test {};
-
-TYPED_TEST_SUITE_P(LookupTest);
-
-TYPED_TEST_P(LookupTest, Count) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m;
-  for (const auto& v : values)
-    EXPECT_EQ(0, m.count(v)) << ::testing::PrintToString(v);
-  m.insert(values.begin(), values.end());
-  for (const auto& v : values)
-    EXPECT_EQ(1, m.count(v)) << ::testing::PrintToString(v);
-}
-
-TYPED_TEST_P(LookupTest, Find) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m;
-  for (const auto& v : values)
-    EXPECT_TRUE(m.end() == m.find(v)) << ::testing::PrintToString(v);
-  m.insert(values.begin(), values.end());
-  for (const auto& v : values) {
-    typename TypeParam::iterator it = m.find(v);
-    static_assert(std::is_same<const typename TypeParam::value_type&,
-                               decltype(*it)>::value,
-                  "");
-    static_assert(std::is_same<const typename TypeParam::value_type*,
-                               decltype(it.operator->())>::value,
-                  "");
-    EXPECT_TRUE(m.end() != it) << ::testing::PrintToString(v);
-    EXPECT_EQ(v, *it) << ::testing::PrintToString(v);
-  }
-}
-
-TYPED_TEST_P(LookupTest, EqualRange) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m;
-  for (const auto& v : values) {
-    auto r = m.equal_range(v);
-    ASSERT_EQ(0, std::distance(r.first, r.second));
-  }
-  m.insert(values.begin(), values.end());
-  for (const auto& v : values) {
-    auto r = m.equal_range(v);
-    ASSERT_EQ(1, std::distance(r.first, r.second));
-    EXPECT_EQ(v, *r.first);
-  }
-}
-
-REGISTER_TYPED_TEST_SUITE_P(LookupTest, Count, Find, EqualRange);
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_UNORDERED_SET_LOOKUP_TEST_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_members_test.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_members_test.h
deleted file mode 100644
index 4c5e104a..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_members_test.h
+++ /dev/null
@@ -1,86 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_UNORDERED_SET_MEMBERS_TEST_H_
-#define ABSL_CONTAINER_INTERNAL_UNORDERED_SET_MEMBERS_TEST_H_
-
-#include <type_traits>
-#include "gmock/gmock.h"
-#include "gtest/gtest.h"
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-template <class UnordSet>
-class MembersTest : public ::testing::Test {};
-
-TYPED_TEST_SUITE_P(MembersTest);
-
-template <typename T>
-void UseType() {}
-
-TYPED_TEST_P(MembersTest, Typedefs) {
-  EXPECT_TRUE((std::is_same<typename TypeParam::key_type,
-                            typename TypeParam::value_type>()));
-  EXPECT_TRUE((absl::conjunction<
-               absl::negation<std::is_signed<typename TypeParam::size_type>>,
-               std::is_integral<typename TypeParam::size_type>>()));
-  EXPECT_TRUE((absl::conjunction<
-               std::is_signed<typename TypeParam::difference_type>,
-               std::is_integral<typename TypeParam::difference_type>>()));
-  EXPECT_TRUE((std::is_convertible<
-               decltype(std::declval<const typename TypeParam::hasher&>()(
-                   std::declval<const typename TypeParam::key_type&>())),
-               size_t>()));
-  EXPECT_TRUE((std::is_convertible<
-               decltype(std::declval<const typename TypeParam::key_equal&>()(
-                   std::declval<const typename TypeParam::key_type&>(),
-                   std::declval<const typename TypeParam::key_type&>())),
-               bool>()));
-  EXPECT_TRUE((std::is_same<typename TypeParam::allocator_type::value_type,
-                            typename TypeParam::value_type>()));
-  EXPECT_TRUE((std::is_same<typename TypeParam::value_type&,
-                            typename TypeParam::reference>()));
-  EXPECT_TRUE((std::is_same<const typename TypeParam::value_type&,
-                            typename TypeParam::const_reference>()));
-  EXPECT_TRUE((std::is_same<typename std::allocator_traits<
-                                typename TypeParam::allocator_type>::pointer,
-                            typename TypeParam::pointer>()));
-  EXPECT_TRUE(
-      (std::is_same<typename std::allocator_traits<
-                        typename TypeParam::allocator_type>::const_pointer,
-                    typename TypeParam::const_pointer>()));
-}
-
-TYPED_TEST_P(MembersTest, SimpleFunctions) {
-  EXPECT_GT(TypeParam().max_size(), 0);
-}
-
-TYPED_TEST_P(MembersTest, BeginEnd) {
-  TypeParam t = {typename TypeParam::value_type{}};
-  EXPECT_EQ(t.begin(), t.cbegin());
-  EXPECT_EQ(t.end(), t.cend());
-  EXPECT_NE(t.begin(), t.end());
-  EXPECT_NE(t.cbegin(), t.cend());
-}
-
-REGISTER_TYPED_TEST_SUITE_P(MembersTest, Typedefs, SimpleFunctions, BeginEnd);
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_UNORDERED_SET_MEMBERS_TEST_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_modifiers_test.h b/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_modifiers_test.h
deleted file mode 100644
index d8864bb2..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/internal/unordered_set_modifiers_test.h
+++ /dev/null
@@ -1,221 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CONTAINER_INTERNAL_UNORDERED_SET_MODIFIERS_TEST_H_
-#define ABSL_CONTAINER_INTERNAL_UNORDERED_SET_MODIFIERS_TEST_H_
-
-#include "gmock/gmock.h"
-#include "gtest/gtest.h"
-#include "absl/container/internal/hash_generator_testing.h"
-#include "absl/container/internal/hash_policy_testing.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-
-template <class UnordSet>
-class ModifiersTest : public ::testing::Test {};
-
-TYPED_TEST_SUITE_P(ModifiersTest);
-
-TYPED_TEST_P(ModifiersTest, Clear) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m(values.begin(), values.end());
-  ASSERT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values));
-  m.clear();
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAre());
-  EXPECT_TRUE(m.empty());
-}
-
-TYPED_TEST_P(ModifiersTest, Insert) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  T val = hash_internal::Generator<T>()();
-  TypeParam m;
-  auto p = m.insert(val);
-  EXPECT_TRUE(p.second);
-  EXPECT_EQ(val, *p.first);
-  p = m.insert(val);
-  EXPECT_FALSE(p.second);
-}
-
-TYPED_TEST_P(ModifiersTest, InsertHint) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  T val = hash_internal::Generator<T>()();
-  TypeParam m;
-  auto it = m.insert(m.end(), val);
-  EXPECT_TRUE(it != m.end());
-  EXPECT_EQ(val, *it);
-  it = m.insert(it, val);
-  EXPECT_TRUE(it != m.end());
-  EXPECT_EQ(val, *it);
-}
-
-TYPED_TEST_P(ModifiersTest, InsertRange) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m;
-  m.insert(values.begin(), values.end());
-  ASSERT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values));
-}
-
-TYPED_TEST_P(ModifiersTest, InsertWithinCapacity) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  T val = hash_internal::Generator<T>()();
-  TypeParam m;
-  m.reserve(10);
-  const size_t original_capacity = m.bucket_count();
-  m.insert(val);
-  EXPECT_EQ(m.bucket_count(), original_capacity);
-  m.insert(val);
-  EXPECT_EQ(m.bucket_count(), original_capacity);
-}
-
-TYPED_TEST_P(ModifiersTest, InsertRangeWithinCapacity) {
-#if !defined(__GLIBCXX__)
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> base_values;
-  std::generate_n(std::back_inserter(base_values), 10,
-                  hash_internal::Generator<T>());
-  std::vector<T> values;
-  while (values.size() != 100) {
-    values.insert(values.end(), base_values.begin(), base_values.end());
-  }
-  TypeParam m;
-  m.reserve(10);
-  const size_t original_capacity = m.bucket_count();
-  m.insert(values.begin(), values.end());
-  EXPECT_EQ(m.bucket_count(), original_capacity);
-#endif
-}
-
-TYPED_TEST_P(ModifiersTest, Emplace) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  T val = hash_internal::Generator<T>()();
-  TypeParam m;
-  // TODO(alkis): We need a way to run emplace in a more meaningful way. Perhaps
-  // with test traits/policy.
-  auto p = m.emplace(val);
-  EXPECT_TRUE(p.second);
-  EXPECT_EQ(val, *p.first);
-  p = m.emplace(val);
-  EXPECT_FALSE(p.second);
-  EXPECT_EQ(val, *p.first);
-}
-
-TYPED_TEST_P(ModifiersTest, EmplaceHint) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  T val = hash_internal::Generator<T>()();
-  TypeParam m;
-  // TODO(alkis): We need a way to run emplace in a more meaningful way. Perhaps
-  // with test traits/policy.
-  auto it = m.emplace_hint(m.end(), val);
-  EXPECT_EQ(val, *it);
-  it = m.emplace_hint(it, val);
-  EXPECT_EQ(val, *it);
-}
-
-template <class V>
-using IfNotVoid = typename std::enable_if<!std::is_void<V>::value, V>::type;
-
-// In openmap we chose not to return the iterator from erase because that's
-// more expensive. As such we adapt erase to return an iterator here.
-struct EraseFirst {
-  template <class Map>
-  auto operator()(Map* m, int) const
-      -> IfNotVoid<decltype(m->erase(m->begin()))> {
-    return m->erase(m->begin());
-  }
-  template <class Map>
-  typename Map::iterator operator()(Map* m, ...) const {
-    auto it = m->begin();
-    m->erase(it++);
-    return it;
-  }
-};
-
-TYPED_TEST_P(ModifiersTest, Erase) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m(values.begin(), values.end());
-  ASSERT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values));
-  std::vector<T> values2;
-  for (const auto& val : values)
-    if (val != *m.begin()) values2.push_back(val);
-  auto it = EraseFirst()(&m, 0);
-  ASSERT_TRUE(it != m.end());
-  EXPECT_EQ(1, std::count(values2.begin(), values2.end(), *it));
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values2.begin(),
-                                                            values2.end()));
-}
-
-TYPED_TEST_P(ModifiersTest, EraseRange) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m(values.begin(), values.end());
-  ASSERT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values));
-  auto it = m.erase(m.begin(), m.end());
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAre());
-  EXPECT_TRUE(it == m.end());
-}
-
-TYPED_TEST_P(ModifiersTest, EraseKey) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> values;
-  std::generate_n(std::back_inserter(values), 10,
-                  hash_internal::Generator<T>());
-  TypeParam m(values.begin(), values.end());
-  ASSERT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values));
-  EXPECT_EQ(1, m.erase(values[0]));
-  EXPECT_EQ(0, std::count(m.begin(), m.end(), values[0]));
-  EXPECT_THAT(keys(m), ::testing::UnorderedElementsAreArray(values.begin() + 1,
-                                                            values.end()));
-}
-
-TYPED_TEST_P(ModifiersTest, Swap) {
-  using T = hash_internal::GeneratedType<TypeParam>;
-  std::vector<T> v1;
-  std::vector<T> v2;
-  std::generate_n(std::back_inserter(v1), 5, hash_internal::Generator<T>());
-  std::generate_n(std::back_inserter(v2), 5, hash_internal::Generator<T>());
-  TypeParam m1(v1.begin(), v1.end());
-  TypeParam m2(v2.begin(), v2.end());
-  EXPECT_THAT(keys(m1), ::testing::UnorderedElementsAreArray(v1));
-  EXPECT_THAT(keys(m2), ::testing::UnorderedElementsAreArray(v2));
-  m1.swap(m2);
-  EXPECT_THAT(keys(m1), ::testing::UnorderedElementsAreArray(v2));
-  EXPECT_THAT(keys(m2), ::testing::UnorderedElementsAreArray(v1));
-}
-
-// TODO(alkis): Write tests for extract.
-// TODO(alkis): Write tests for merge.
-
-REGISTER_TYPED_TEST_SUITE_P(ModifiersTest, Clear, Insert, InsertHint,
-                            InsertRange, InsertWithinCapacity,
-                            InsertRangeWithinCapacity, Emplace, EmplaceHint,
-                            Erase, EraseRange, EraseKey, Swap);
-
-}  // namespace container_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_INTERNAL_UNORDERED_SET_MODIFIERS_TEST_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/node_hash_map.h b/projects/llm_framework/include/abseil-cpp/absl/container/node_hash_map.h
deleted file mode 100644
index cb41543c..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/node_hash_map.h
+++ /dev/null
@@ -1,629 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: node_hash_map.h
-// -----------------------------------------------------------------------------
-//
-// An `absl::node_hash_map<K, V>` is an unordered associative container of
-// unique keys and associated values designed to be a more efficient replacement
-// for `std::unordered_map`. Like `unordered_map`, search, insertion, and
-// deletion of map elements can be done as an `O(1)` operation. However,
-// `node_hash_map` (and other unordered associative containers known as the
-// collection of Abseil "Swiss tables") contain other optimizations that result
-// in both memory and computation advantages.
-//
-// In most cases, your default choice for a hash map should be a map of type
-// `flat_hash_map`. However, if you need pointer stability and cannot store
-// a `flat_hash_map` with `unique_ptr` elements, a `node_hash_map` may be a
-// valid alternative. As well, if you are migrating your code from using
-// `std::unordered_map`, a `node_hash_map` provides a more straightforward
-// migration, because it guarantees pointer stability. Consider migrating to
-// `node_hash_map` and perhaps converting to a more efficient `flat_hash_map`
-// upon further review.
-
-#ifndef ABSL_CONTAINER_NODE_HASH_MAP_H_
-#define ABSL_CONTAINER_NODE_HASH_MAP_H_
-
-#include <cstddef>
-#include <tuple>
-#include <type_traits>
-#include <utility>
-
-#include "absl/algorithm/container.h"
-#include "absl/base/macros.h"
-#include "absl/container/internal/container_memory.h"
-#include "absl/container/internal/hash_function_defaults.h"  // IWYU pragma: export
-#include "absl/container/internal/node_slot_policy.h"
-#include "absl/container/internal/raw_hash_map.h"  // IWYU pragma: export
-#include "absl/memory/memory.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-template <class Key, class Value>
-class NodeHashMapPolicy;
-}  // namespace container_internal
-
-// -----------------------------------------------------------------------------
-// absl::node_hash_map
-// -----------------------------------------------------------------------------
-//
-// An `absl::node_hash_map<K, V>` is an unordered associative container which
-// has been optimized for both speed and memory footprint in most common use
-// cases. Its interface is similar to that of `std::unordered_map<K, V>` with
-// the following notable differences:
-//
-// * Supports heterogeneous lookup, through `find()`, `operator[]()` and
-//   `insert()`, provided that the map is provided a compatible heterogeneous
-//   hashing function and equality operator. See below for details.
-// * Contains a `capacity()` member function indicating the number of element
-//   slots (open, deleted, and empty) within the hash map.
-// * Returns `void` from the `erase(iterator)` overload.
-//
-// By default, `node_hash_map` uses the `absl::Hash` hashing framework.
-// All fundamental and Abseil types that support the `absl::Hash` framework have
-// a compatible equality operator for comparing insertions into `node_hash_map`.
-// If your type is not yet supported by the `absl::Hash` framework, see
-// absl/hash/hash.h for information on extending Abseil hashing to user-defined
-// types.
-//
-// Using `absl::node_hash_map` at interface boundaries in dynamically loaded
-// libraries (e.g. .dll, .so) is unsupported due to way `absl::Hash` values may
-// be randomized across dynamically loaded libraries.
-//
-// To achieve heterogeneous lookup for custom types either `Hash` and `Eq` type
-// parameters can be used or `T` should have public inner types
-// `absl_container_hash` and (optionally) `absl_container_eq`. In either case,
-// `typename Hash::is_transparent` and `typename Eq::is_transparent` should be
-// well-formed. Both types are basically functors:
-// * `Hash` should support `size_t operator()(U val) const` that returns a hash
-// for the given `val`.
-// * `Eq` should support `bool operator()(U lhs, V rhs) const` that returns true
-// if `lhs` is equal to `rhs`.
-//
-// In most cases `T` needs only to provide the `absl_container_hash`. In this
-// case `std::equal_to<void>` will be used instead of `eq` part.
-//
-// Example:
-//
-//   // Create a node hash map of three strings (that map to strings)
-//   absl::node_hash_map<std::string, std::string> ducks =
-//     {{"a", "huey"}, {"b", "dewey"}, {"c", "louie"}};
-//
-//  // Insert a new element into the node hash map
-//  ducks.insert({"d", "donald"}};
-//
-//  // Force a rehash of the node hash map
-//  ducks.rehash(0);
-//
-//  // Find the element with the key "b"
-//  std::string search_key = "b";
-//  auto result = ducks.find(search_key);
-//  if (result != ducks.end()) {
-//    std::cout << "Result: " << result->second << std::endl;
-//  }
-template <class Key, class Value,
-          class Hash = absl::container_internal::hash_default_hash<Key>,
-          class Eq = absl::container_internal::hash_default_eq<Key>,
-          class Alloc = std::allocator<std::pair<const Key, Value>>>
-class node_hash_map
-    : public absl::container_internal::raw_hash_map<
-          absl::container_internal::NodeHashMapPolicy<Key, Value>, Hash, Eq,
-          Alloc> {
-  using Base = typename node_hash_map::raw_hash_map;
-
- public:
-  // Constructors and Assignment Operators
-  //
-  // A node_hash_map supports the same overload set as `std::unordered_map`
-  // for construction and assignment:
-  //
-  // *  Default constructor
-  //
-  //    // No allocation for the table's elements is made.
-  //    absl::node_hash_map<int, std::string> map1;
-  //
-  // * Initializer List constructor
-  //
-  //   absl::node_hash_map<int, std::string> map2 =
-  //       {{1, "huey"}, {2, "dewey"}, {3, "louie"},};
-  //
-  // * Copy constructor
-  //
-  //   absl::node_hash_map<int, std::string> map3(map2);
-  //
-  // * Copy assignment operator
-  //
-  //  // Hash functor and Comparator are copied as well
-  //  absl::node_hash_map<int, std::string> map4;
-  //  map4 = map3;
-  //
-  // * Move constructor
-  //
-  //   // Move is guaranteed efficient
-  //   absl::node_hash_map<int, std::string> map5(std::move(map4));
-  //
-  // * Move assignment operator
-  //
-  //   // May be efficient if allocators are compatible
-  //   absl::node_hash_map<int, std::string> map6;
-  //   map6 = std::move(map5);
-  //
-  // * Range constructor
-  //
-  //   std::vector<std::pair<int, std::string>> v = {{1, "a"}, {2, "b"}};
-  //   absl::node_hash_map<int, std::string> map7(v.begin(), v.end());
-  node_hash_map() {}
-  using Base::Base;
-
-  // node_hash_map::begin()
-  //
-  // Returns an iterator to the beginning of the `node_hash_map`.
-  using Base::begin;
-
-  // node_hash_map::cbegin()
-  //
-  // Returns a const iterator to the beginning of the `node_hash_map`.
-  using Base::cbegin;
-
-  // node_hash_map::cend()
-  //
-  // Returns a const iterator to the end of the `node_hash_map`.
-  using Base::cend;
-
-  // node_hash_map::end()
-  //
-  // Returns an iterator to the end of the `node_hash_map`.
-  using Base::end;
-
-  // node_hash_map::capacity()
-  //
-  // Returns the number of element slots (assigned, deleted, and empty)
-  // available within the `node_hash_map`.
-  //
-  // NOTE: this member function is particular to `absl::node_hash_map` and is
-  // not provided in the `std::unordered_map` API.
-  using Base::capacity;
-
-  // node_hash_map::empty()
-  //
-  // Returns whether or not the `node_hash_map` is empty.
-  using Base::empty;
-
-  // node_hash_map::max_size()
-  //
-  // Returns the largest theoretical possible number of elements within a
-  // `node_hash_map` under current memory constraints. This value can be thought
-  // of as the largest value of `std::distance(begin(), end())` for a
-  // `node_hash_map<K, V>`.
-  using Base::max_size;
-
-  // node_hash_map::size()
-  //
-  // Returns the number of elements currently within the `node_hash_map`.
-  using Base::size;
-
-  // node_hash_map::clear()
-  //
-  // Removes all elements from the `node_hash_map`. Invalidates any references,
-  // pointers, or iterators referring to contained elements.
-  //
-  // NOTE: this operation may shrink the underlying buffer. To avoid shrinking
-  // the underlying buffer call `erase(begin(), end())`.
-  using Base::clear;
-
-  // node_hash_map::erase()
-  //
-  // Erases elements within the `node_hash_map`. Erasing does not trigger a
-  // rehash. Overloads are listed below.
-  //
-  // void erase(const_iterator pos):
-  //
-  //   Erases the element at `position` of the `node_hash_map`, returning
-  //   `void`.
-  //
-  //   NOTE: this return behavior is different than that of STL containers in
-  //   general and `std::unordered_map` in particular.
-  //
-  // iterator erase(const_iterator first, const_iterator last):
-  //
-  //   Erases the elements in the open interval [`first`, `last`), returning an
-  //   iterator pointing to `last`. The special case of calling
-  //   `erase(begin(), end())` resets the reserved growth such that if
-  //   `reserve(N)` has previously been called and there has been no intervening
-  //   call to `clear()`, then after calling `erase(begin(), end())`, it is safe
-  //   to assume that inserting N elements will not cause a rehash.
-  //
-  // size_type erase(const key_type& key):
-  //
-  //   Erases the element with the matching key, if it exists, returning the
-  //   number of elements erased (0 or 1).
-  using Base::erase;
-
-  // node_hash_map::insert()
-  //
-  // Inserts an element of the specified value into the `node_hash_map`,
-  // returning an iterator pointing to the newly inserted element, provided that
-  // an element with the given key does not already exist. If rehashing occurs
-  // due to the insertion, all iterators are invalidated. Overloads are listed
-  // below.
-  //
-  // std::pair<iterator,bool> insert(const init_type& value):
-  //
-  //   Inserts a value into the `node_hash_map`. Returns a pair consisting of an
-  //   iterator to the inserted element (or to the element that prevented the
-  //   insertion) and a `bool` denoting whether the insertion took place.
-  //
-  // std::pair<iterator,bool> insert(T&& value):
-  // std::pair<iterator,bool> insert(init_type&& value):
-  //
-  //   Inserts a moveable value into the `node_hash_map`. Returns a `std::pair`
-  //   consisting of an iterator to the inserted element (or to the element that
-  //   prevented the insertion) and a `bool` denoting whether the insertion took
-  //   place.
-  //
-  // iterator insert(const_iterator hint, const init_type& value):
-  // iterator insert(const_iterator hint, T&& value):
-  // iterator insert(const_iterator hint, init_type&& value);
-  //
-  //   Inserts a value, using the position of `hint` as a non-binding suggestion
-  //   for where to begin the insertion search. Returns an iterator to the
-  //   inserted element, or to the existing element that prevented the
-  //   insertion.
-  //
-  // void insert(InputIterator first, InputIterator last):
-  //
-  //   Inserts a range of values [`first`, `last`).
-  //
-  //   NOTE: Although the STL does not specify which element may be inserted if
-  //   multiple keys compare equivalently, for `node_hash_map` we guarantee the
-  //   first match is inserted.
-  //
-  // void insert(std::initializer_list<init_type> ilist):
-  //
-  //   Inserts the elements within the initializer list `ilist`.
-  //
-  //   NOTE: Although the STL does not specify which element may be inserted if
-  //   multiple keys compare equivalently within the initializer list, for
-  //   `node_hash_map` we guarantee the first match is inserted.
-  using Base::insert;
-
-  // node_hash_map::insert_or_assign()
-  //
-  // Inserts an element of the specified value into the `node_hash_map` provided
-  // that a value with the given key does not already exist, or replaces it with
-  // the element value if a key for that value already exists, returning an
-  // iterator pointing to the newly inserted element. If rehashing occurs due to
-  // the insertion, all iterators are invalidated. Overloads are listed
-  // below.
-  //
-  // std::pair<iterator, bool> insert_or_assign(const init_type& k, T&& obj):
-  // std::pair<iterator, bool> insert_or_assign(init_type&& k, T&& obj):
-  //
-  //   Inserts/Assigns (or moves) the element of the specified key into the
-  //   `node_hash_map`.
-  //
-  // iterator insert_or_assign(const_iterator hint,
-  //                           const init_type& k, T&& obj):
-  // iterator insert_or_assign(const_iterator hint, init_type&& k, T&& obj):
-  //
-  //   Inserts/Assigns (or moves) the element of the specified key into the
-  //   `node_hash_map` using the position of `hint` as a non-binding suggestion
-  //   for where to begin the insertion search.
-  using Base::insert_or_assign;
-
-  // node_hash_map::emplace()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `node_hash_map`, provided that no element with the given key
-  // already exists.
-  //
-  // The element may be constructed even if there already is an element with the
-  // key in the container, in which case the newly constructed element will be
-  // destroyed immediately. Prefer `try_emplace()` unless your key is not
-  // copyable or moveable.
-  //
-  // If rehashing occurs due to the insertion, all iterators are invalidated.
-  using Base::emplace;
-
-  // node_hash_map::emplace_hint()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `node_hash_map`, using the position of `hint` as a non-binding
-  // suggestion for where to begin the insertion search, and only inserts
-  // provided that no element with the given key already exists.
-  //
-  // The element may be constructed even if there already is an element with the
-  // key in the container, in which case the newly constructed element will be
-  // destroyed immediately. Prefer `try_emplace()` unless your key is not
-  // copyable or moveable.
-  //
-  // If rehashing occurs due to the insertion, all iterators are invalidated.
-  using Base::emplace_hint;
-
-  // node_hash_map::try_emplace()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `node_hash_map`, provided that no element with the given key
-  // already exists. Unlike `emplace()`, if an element with the given key
-  // already exists, we guarantee that no element is constructed.
-  //
-  // If rehashing occurs due to the insertion, all iterators are invalidated.
-  // Overloads are listed below.
-  //
-  //   std::pair<iterator, bool> try_emplace(const key_type& k, Args&&... args):
-  //   std::pair<iterator, bool> try_emplace(key_type&& k, Args&&... args):
-  //
-  // Inserts (via copy or move) the element of the specified key into the
-  // `node_hash_map`.
-  //
-  //   iterator try_emplace(const_iterator hint,
-  //                        const key_type& k, Args&&... args):
-  //   iterator try_emplace(const_iterator hint, key_type&& k, Args&&... args):
-  //
-  // Inserts (via copy or move) the element of the specified key into the
-  // `node_hash_map` using the position of `hint` as a non-binding suggestion
-  // for where to begin the insertion search.
-  //
-  // All `try_emplace()` overloads make the same guarantees regarding rvalue
-  // arguments as `std::unordered_map::try_emplace()`, namely that these
-  // functions will not move from rvalue arguments if insertions do not happen.
-  using Base::try_emplace;
-
-  // node_hash_map::extract()
-  //
-  // Extracts the indicated element, erasing it in the process, and returns it
-  // as a C++17-compatible node handle. Overloads are listed below.
-  //
-  // node_type extract(const_iterator position):
-  //
-  //   Extracts the key,value pair of the element at the indicated position and
-  //   returns a node handle owning that extracted data.
-  //
-  // node_type extract(const key_type& x):
-  //
-  //   Extracts the key,value pair of the element with a key matching the passed
-  //   key value and returns a node handle owning that extracted data. If the
-  //   `node_hash_map` does not contain an element with a matching key, this
-  //   function returns an empty node handle.
-  //
-  // NOTE: when compiled in an earlier version of C++ than C++17,
-  // `node_type::key()` returns a const reference to the key instead of a
-  // mutable reference. We cannot safely return a mutable reference without
-  // std::launder (which is not available before C++17).
-  using Base::extract;
-
-  // node_hash_map::merge()
-  //
-  // Extracts elements from a given `source` node hash map into this
-  // `node_hash_map`. If the destination `node_hash_map` already contains an
-  // element with an equivalent key, that element is not extracted.
-  using Base::merge;
-
-  // node_hash_map::swap(node_hash_map& other)
-  //
-  // Exchanges the contents of this `node_hash_map` with those of the `other`
-  // node hash map, avoiding invocation of any move, copy, or swap operations on
-  // individual elements.
-  //
-  // All iterators and references on the `node_hash_map` remain valid, excepting
-  // for the past-the-end iterator, which is invalidated.
-  //
-  // `swap()` requires that the node hash map's hashing and key equivalence
-  // functions be Swappable, and are exchanged using unqualified calls to
-  // non-member `swap()`. If the map's allocator has
-  // `std::allocator_traits<allocator_type>::propagate_on_container_swap::value`
-  // set to `true`, the allocators are also exchanged using an unqualified call
-  // to non-member `swap()`; otherwise, the allocators are not swapped.
-  using Base::swap;
-
-  // node_hash_map::rehash(count)
-  //
-  // Rehashes the `node_hash_map`, setting the number of slots to be at least
-  // the passed value. If the new number of slots increases the load factor more
-  // than the current maximum load factor
-  // (`count` < `size()` / `max_load_factor()`), then the new number of slots
-  // will be at least `size()` / `max_load_factor()`.
-  //
-  // To force a rehash, pass rehash(0).
-  using Base::rehash;
-
-  // node_hash_map::reserve(count)
-  //
-  // Sets the number of slots in the `node_hash_map` to the number needed to
-  // accommodate at least `count` total elements without exceeding the current
-  // maximum load factor, and may rehash the container if needed.
-  using Base::reserve;
-
-  // node_hash_map::at()
-  //
-  // Returns a reference to the mapped value of the element with key equivalent
-  // to the passed key.
-  using Base::at;
-
-  // node_hash_map::contains()
-  //
-  // Determines whether an element with a key comparing equal to the given `key`
-  // exists within the `node_hash_map`, returning `true` if so or `false`
-  // otherwise.
-  using Base::contains;
-
-  // node_hash_map::count(const Key& key) const
-  //
-  // Returns the number of elements with a key comparing equal to the given
-  // `key` within the `node_hash_map`. note that this function will return
-  // either `1` or `0` since duplicate keys are not allowed within a
-  // `node_hash_map`.
-  using Base::count;
-
-  // node_hash_map::equal_range()
-  //
-  // Returns a closed range [first, last], defined by a `std::pair` of two
-  // iterators, containing all elements with the passed key in the
-  // `node_hash_map`.
-  using Base::equal_range;
-
-  // node_hash_map::find()
-  //
-  // Finds an element with the passed `key` within the `node_hash_map`.
-  using Base::find;
-
-  // node_hash_map::operator[]()
-  //
-  // Returns a reference to the value mapped to the passed key within the
-  // `node_hash_map`, performing an `insert()` if the key does not already
-  // exist. If an insertion occurs and results in a rehashing of the container,
-  // all iterators are invalidated. Otherwise iterators are not affected and
-  // references are not invalidated. Overloads are listed below.
-  //
-  // T& operator[](const Key& key):
-  //
-  //   Inserts an init_type object constructed in-place if the element with the
-  //   given key does not exist.
-  //
-  // T& operator[](Key&& key):
-  //
-  //   Inserts an init_type object constructed in-place provided that an element
-  //   with the given key does not exist.
-  using Base::operator[];
-
-  // node_hash_map::bucket_count()
-  //
-  // Returns the number of "buckets" within the `node_hash_map`.
-  using Base::bucket_count;
-
-  // node_hash_map::load_factor()
-  //
-  // Returns the current load factor of the `node_hash_map` (the average number
-  // of slots occupied with a value within the hash map).
-  using Base::load_factor;
-
-  // node_hash_map::max_load_factor()
-  //
-  // Manages the maximum load factor of the `node_hash_map`. Overloads are
-  // listed below.
-  //
-  // float node_hash_map::max_load_factor()
-  //
-  //   Returns the current maximum load factor of the `node_hash_map`.
-  //
-  // void node_hash_map::max_load_factor(float ml)
-  //
-  //   Sets the maximum load factor of the `node_hash_map` to the passed value.
-  //
-  //   NOTE: This overload is provided only for API compatibility with the STL;
-  //   `node_hash_map` will ignore any set load factor and manage its rehashing
-  //   internally as an implementation detail.
-  using Base::max_load_factor;
-
-  // node_hash_map::get_allocator()
-  //
-  // Returns the allocator function associated with this `node_hash_map`.
-  using Base::get_allocator;
-
-  // node_hash_map::hash_function()
-  //
-  // Returns the hashing function used to hash the keys within this
-  // `node_hash_map`.
-  using Base::hash_function;
-
-  // node_hash_map::key_eq()
-  //
-  // Returns the function used for comparing keys equality.
-  using Base::key_eq;
-};
-
-// erase_if(node_hash_map<>, Pred)
-//
-// Erases all elements that satisfy the predicate `pred` from the container `c`.
-// Returns the number of erased elements.
-template <typename K, typename V, typename H, typename E, typename A,
-          typename Predicate>
-typename node_hash_map<K, V, H, E, A>::size_type erase_if(
-    node_hash_map<K, V, H, E, A>& c, Predicate pred) {
-  return container_internal::EraseIf(pred, &c);
-}
-
-namespace container_internal {
-
-template <class Key, class Value>
-class NodeHashMapPolicy
-    : public absl::container_internal::node_slot_policy<
-          std::pair<const Key, Value>&, NodeHashMapPolicy<Key, Value>> {
-  using value_type = std::pair<const Key, Value>;
-
- public:
-  using key_type = Key;
-  using mapped_type = Value;
-  using init_type = std::pair</*non const*/ key_type, mapped_type>;
-
-  template <class Allocator, class... Args>
-  static value_type* new_element(Allocator* alloc, Args&&... args) {
-    using PairAlloc = typename absl::allocator_traits<
-        Allocator>::template rebind_alloc<value_type>;
-    PairAlloc pair_alloc(*alloc);
-    value_type* res =
-        absl::allocator_traits<PairAlloc>::allocate(pair_alloc, 1);
-    absl::allocator_traits<PairAlloc>::construct(pair_alloc, res,
-                                                 std::forward<Args>(args)...);
-    return res;
-  }
-
-  template <class Allocator>
-  static void delete_element(Allocator* alloc, value_type* pair) {
-    using PairAlloc = typename absl::allocator_traits<
-        Allocator>::template rebind_alloc<value_type>;
-    PairAlloc pair_alloc(*alloc);
-    absl::allocator_traits<PairAlloc>::destroy(pair_alloc, pair);
-    absl::allocator_traits<PairAlloc>::deallocate(pair_alloc, pair, 1);
-  }
-
-  template <class F, class... Args>
-  static decltype(absl::container_internal::DecomposePair(
-      std::declval<F>(), std::declval<Args>()...))
-  apply(F&& f, Args&&... args) {
-    return absl::container_internal::DecomposePair(std::forward<F>(f),
-                                                   std::forward<Args>(args)...);
-  }
-
-  static size_t element_space_used(const value_type*) {
-    return sizeof(value_type);
-  }
-
-  static Value& value(value_type* elem) { return elem->second; }
-  static const Value& value(const value_type* elem) { return elem->second; }
-
-  template <class Hash>
-  static constexpr HashSlotFn get_hash_slot_fn() {
-    return memory_internal::IsLayoutCompatible<Key, Value>::value
-               ? &TypeErasedDerefAndApplyToSlotFn<Hash, Key>
-               : nullptr;
-  }
-};
-}  // namespace container_internal
-
-namespace container_algorithm_internal {
-
-// Specialization of trait in absl/algorithm/container.h
-template <class Key, class T, class Hash, class KeyEqual, class Allocator>
-struct IsUnorderedContainer<
-    absl::node_hash_map<Key, T, Hash, KeyEqual, Allocator>> : std::true_type {};
-
-}  // namespace container_algorithm_internal
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_NODE_HASH_MAP_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/container/node_hash_set.h b/projects/llm_framework/include/abseil-cpp/absl/container/node_hash_set.h
deleted file mode 100644
index 8cc4b624..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/container/node_hash_set.h
+++ /dev/null
@@ -1,524 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: node_hash_set.h
-// -----------------------------------------------------------------------------
-//
-// An `absl::node_hash_set<T>` is an unordered associative container designed to
-// be a more efficient replacement for `std::unordered_set`. Like
-// `unordered_set`, search, insertion, and deletion of set elements can be done
-// as an `O(1)` operation. However, `node_hash_set` (and other unordered
-// associative containers known as the collection of Abseil "Swiss tables")
-// contain other optimizations that result in both memory and computation
-// advantages.
-//
-// In most cases, your default choice for a hash table should be a map of type
-// `flat_hash_map` or a set of type `flat_hash_set`. However, if you need
-// pointer stability, a `node_hash_set` should be your preferred choice. As
-// well, if you are migrating your code from using `std::unordered_set`, a
-// `node_hash_set` should be an easy migration. Consider migrating to
-// `node_hash_set` and perhaps converting to a more efficient `flat_hash_set`
-// upon further review.
-
-#ifndef ABSL_CONTAINER_NODE_HASH_SET_H_
-#define ABSL_CONTAINER_NODE_HASH_SET_H_
-
-#include <cstddef>
-#include <type_traits>
-
-#include "absl/algorithm/container.h"
-#include "absl/base/macros.h"
-#include "absl/container/internal/container_memory.h"
-#include "absl/container/internal/hash_function_defaults.h"  // IWYU pragma: export
-#include "absl/container/internal/node_slot_policy.h"
-#include "absl/container/internal/raw_hash_set.h"  // IWYU pragma: export
-#include "absl/memory/memory.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace container_internal {
-template <typename T>
-struct NodeHashSetPolicy;
-}  // namespace container_internal
-
-// -----------------------------------------------------------------------------
-// absl::node_hash_set
-// -----------------------------------------------------------------------------
-//
-// An `absl::node_hash_set<T>` is an unordered associative container which
-// has been optimized for both speed and memory footprint in most common use
-// cases. Its interface is similar to that of `std::unordered_set<T>` with the
-// following notable differences:
-//
-// * Supports heterogeneous lookup, through `find()`, `operator[]()` and
-//   `insert()`, provided that the set is provided a compatible heterogeneous
-//   hashing function and equality operator. See below for details.
-// * Contains a `capacity()` member function indicating the number of element
-//   slots (open, deleted, and empty) within the hash set.
-// * Returns `void` from the `erase(iterator)` overload.
-//
-// By default, `node_hash_set` uses the `absl::Hash` hashing framework.
-// All fundamental and Abseil types that support the `absl::Hash` framework have
-// a compatible equality operator for comparing insertions into `node_hash_set`.
-// If your type is not yet supported by the `absl::Hash` framework, see
-// absl/hash/hash.h for information on extending Abseil hashing to user-defined
-// types.
-//
-// Using `absl::node_hash_set` at interface boundaries in dynamically loaded
-// libraries (e.g. .dll, .so) is unsupported due to way `absl::Hash` values may
-// be randomized across dynamically loaded libraries.
-//
-// To achieve heterogeneous lookup for custom types either `Hash` and `Eq` type
-// parameters can be used or `T` should have public inner types
-// `absl_container_hash` and (optionally) `absl_container_eq`. In either case,
-// `typename Hash::is_transparent` and `typename Eq::is_transparent` should be
-// well-formed. Both types are basically functors:
-// * `Hash` should support `size_t operator()(U val) const` that returns a hash
-// for the given `val`.
-// * `Eq` should support `bool operator()(U lhs, V rhs) const` that returns true
-// if `lhs` is equal to `rhs`.
-//
-// In most cases `T` needs only to provide the `absl_container_hash`. In this
-// case `std::equal_to<void>` will be used instead of `eq` part.
-//
-// Example:
-//
-//   // Create a node hash set of three strings
-//   absl::node_hash_set<std::string> ducks =
-//     {"huey", "dewey", "louie"};
-//
-//  // Insert a new element into the node hash set
-//  ducks.insert("donald");
-//
-//  // Force a rehash of the node hash set
-//  ducks.rehash(0);
-//
-//  // See if "dewey" is present
-//  if (ducks.contains("dewey")) {
-//    std::cout << "We found dewey!" << std::endl;
-//  }
-template <class T, class Hash = absl::container_internal::hash_default_hash<T>,
-          class Eq = absl::container_internal::hash_default_eq<T>,
-          class Alloc = std::allocator<T>>
-class node_hash_set
-    : public absl::container_internal::raw_hash_set<
-          absl::container_internal::NodeHashSetPolicy<T>, Hash, Eq, Alloc> {
-  using Base = typename node_hash_set::raw_hash_set;
-
- public:
-  // Constructors and Assignment Operators
-  //
-  // A node_hash_set supports the same overload set as `std::unordered_set`
-  // for construction and assignment:
-  //
-  // *  Default constructor
-  //
-  //    // No allocation for the table's elements is made.
-  //    absl::node_hash_set<std::string> set1;
-  //
-  // * Initializer List constructor
-  //
-  //   absl::node_hash_set<std::string> set2 =
-  //       {{"huey"}, {"dewey"}, {"louie"}};
-  //
-  // * Copy constructor
-  //
-  //   absl::node_hash_set<std::string> set3(set2);
-  //
-  // * Copy assignment operator
-  //
-  //  // Hash functor and Comparator are copied as well
-  //  absl::node_hash_set<std::string> set4;
-  //  set4 = set3;
-  //
-  // * Move constructor
-  //
-  //   // Move is guaranteed efficient
-  //   absl::node_hash_set<std::string> set5(std::move(set4));
-  //
-  // * Move assignment operator
-  //
-  //   // May be efficient if allocators are compatible
-  //   absl::node_hash_set<std::string> set6;
-  //   set6 = std::move(set5);
-  //
-  // * Range constructor
-  //
-  //   std::vector<std::string> v = {"a", "b"};
-  //   absl::node_hash_set<std::string> set7(v.begin(), v.end());
-  node_hash_set() {}
-  using Base::Base;
-
-  // node_hash_set::begin()
-  //
-  // Returns an iterator to the beginning of the `node_hash_set`.
-  using Base::begin;
-
-  // node_hash_set::cbegin()
-  //
-  // Returns a const iterator to the beginning of the `node_hash_set`.
-  using Base::cbegin;
-
-  // node_hash_set::cend()
-  //
-  // Returns a const iterator to the end of the `node_hash_set`.
-  using Base::cend;
-
-  // node_hash_set::end()
-  //
-  // Returns an iterator to the end of the `node_hash_set`.
-  using Base::end;
-
-  // node_hash_set::capacity()
-  //
-  // Returns the number of element slots (assigned, deleted, and empty)
-  // available within the `node_hash_set`.
-  //
-  // NOTE: this member function is particular to `absl::node_hash_set` and is
-  // not provided in the `std::unordered_set` API.
-  using Base::capacity;
-
-  // node_hash_set::empty()
-  //
-  // Returns whether or not the `node_hash_set` is empty.
-  using Base::empty;
-
-  // node_hash_set::max_size()
-  //
-  // Returns the largest theoretical possible number of elements within a
-  // `node_hash_set` under current memory constraints. This value can be thought
-  // of the largest value of `std::distance(begin(), end())` for a
-  // `node_hash_set<T>`.
-  using Base::max_size;
-
-  // node_hash_set::size()
-  //
-  // Returns the number of elements currently within the `node_hash_set`.
-  using Base::size;
-
-  // node_hash_set::clear()
-  //
-  // Removes all elements from the `node_hash_set`. Invalidates any references,
-  // pointers, or iterators referring to contained elements.
-  //
-  // NOTE: this operation may shrink the underlying buffer. To avoid shrinking
-  // the underlying buffer call `erase(begin(), end())`.
-  using Base::clear;
-
-  // node_hash_set::erase()
-  //
-  // Erases elements within the `node_hash_set`. Erasing does not trigger a
-  // rehash. Overloads are listed below.
-  //
-  // void erase(const_iterator pos):
-  //
-  //   Erases the element at `position` of the `node_hash_set`, returning
-  //   `void`.
-  //
-  //   NOTE: this return behavior is different than that of STL containers in
-  //   general and `std::unordered_set` in particular.
-  //
-  // iterator erase(const_iterator first, const_iterator last):
-  //
-  //   Erases the elements in the open interval [`first`, `last`), returning an
-  //   iterator pointing to `last`. The special case of calling
-  //   `erase(begin(), end())` resets the reserved growth such that if
-  //   `reserve(N)` has previously been called and there has been no intervening
-  //   call to `clear()`, then after calling `erase(begin(), end())`, it is safe
-  //   to assume that inserting N elements will not cause a rehash.
-  //
-  // size_type erase(const key_type& key):
-  //
-  //   Erases the element with the matching key, if it exists, returning the
-  //   number of elements erased (0 or 1).
-  using Base::erase;
-
-  // node_hash_set::insert()
-  //
-  // Inserts an element of the specified value into the `node_hash_set`,
-  // returning an iterator pointing to the newly inserted element, provided that
-  // an element with the given key does not already exist. If rehashing occurs
-  // due to the insertion, all iterators are invalidated. Overloads are listed
-  // below.
-  //
-  // std::pair<iterator,bool> insert(const T& value):
-  //
-  //   Inserts a value into the `node_hash_set`. Returns a pair consisting of an
-  //   iterator to the inserted element (or to the element that prevented the
-  //   insertion) and a bool denoting whether the insertion took place.
-  //
-  // std::pair<iterator,bool> insert(T&& value):
-  //
-  //   Inserts a moveable value into the `node_hash_set`. Returns a pair
-  //   consisting of an iterator to the inserted element (or to the element that
-  //   prevented the insertion) and a bool denoting whether the insertion took
-  //   place.
-  //
-  // iterator insert(const_iterator hint, const T& value):
-  // iterator insert(const_iterator hint, T&& value):
-  //
-  //   Inserts a value, using the position of `hint` as a non-binding suggestion
-  //   for where to begin the insertion search. Returns an iterator to the
-  //   inserted element, or to the existing element that prevented the
-  //   insertion.
-  //
-  // void insert(InputIterator first, InputIterator last):
-  //
-  //   Inserts a range of values [`first`, `last`).
-  //
-  //   NOTE: Although the STL does not specify which element may be inserted if
-  //   multiple keys compare equivalently, for `node_hash_set` we guarantee the
-  //   first match is inserted.
-  //
-  // void insert(std::initializer_list<T> ilist):
-  //
-  //   Inserts the elements within the initializer list `ilist`.
-  //
-  //   NOTE: Although the STL does not specify which element may be inserted if
-  //   multiple keys compare equivalently within the initializer list, for
-  //   `node_hash_set` we guarantee the first match is inserted.
-  using Base::insert;
-
-  // node_hash_set::emplace()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `node_hash_set`, provided that no element with the given key
-  // already exists.
-  //
-  // The element may be constructed even if there already is an element with the
-  // key in the container, in which case the newly constructed element will be
-  // destroyed immediately.
-  //
-  // If rehashing occurs due to the insertion, all iterators are invalidated.
-  using Base::emplace;
-
-  // node_hash_set::emplace_hint()
-  //
-  // Inserts an element of the specified value by constructing it in-place
-  // within the `node_hash_set`, using the position of `hint` as a non-binding
-  // suggestion for where to begin the insertion search, and only inserts
-  // provided that no element with the given key already exists.
-  //
-  // The element may be constructed even if there already is an element with the
-  // key in the container, in which case the newly constructed element will be
-  // destroyed immediately.
-  //
-  // If rehashing occurs due to the insertion, all iterators are invalidated.
-  using Base::emplace_hint;
-
-  // node_hash_set::extract()
-  //
-  // Extracts the indicated element, erasing it in the process, and returns it
-  // as a C++17-compatible node handle. Overloads are listed below.
-  //
-  // node_type extract(const_iterator position):
-  //
-  //   Extracts the element at the indicated position and returns a node handle
-  //   owning that extracted data.
-  //
-  // node_type extract(const key_type& x):
-  //
-  //   Extracts the element with the key matching the passed key value and
-  //   returns a node handle owning that extracted data. If the `node_hash_set`
-  //   does not contain an element with a matching key, this function returns an
-  // empty node handle.
-  using Base::extract;
-
-  // node_hash_set::merge()
-  //
-  // Extracts elements from a given `source` node hash set into this
-  // `node_hash_set`. If the destination `node_hash_set` already contains an
-  // element with an equivalent key, that element is not extracted.
-  using Base::merge;
-
-  // node_hash_set::swap(node_hash_set& other)
-  //
-  // Exchanges the contents of this `node_hash_set` with those of the `other`
-  // node hash set, avoiding invocation of any move, copy, or swap operations on
-  // individual elements.
-  //
-  // All iterators and references on the `node_hash_set` remain valid, excepting
-  // for the past-the-end iterator, which is invalidated.
-  //
-  // `swap()` requires that the node hash set's hashing and key equivalence
-  // functions be Swappable, and are exchanged using unqualified calls to
-  // non-member `swap()`. If the set's allocator has
-  // `std::allocator_traits<allocator_type>::propagate_on_container_swap::value`
-  // set to `true`, the allocators are also exchanged using an unqualified call
-  // to non-member `swap()`; otherwise, the allocators are not swapped.
-  using Base::swap;
-
-  // node_hash_set::rehash(count)
-  //
-  // Rehashes the `node_hash_set`, setting the number of slots to be at least
-  // the passed value. If the new number of slots increases the load factor more
-  // than the current maximum load factor
-  // (`count` < `size()` / `max_load_factor()`), then the new number of slots
-  // will be at least `size()` / `max_load_factor()`.
-  //
-  // To force a rehash, pass rehash(0).
-  //
-  // NOTE: unlike behavior in `std::unordered_set`, references are also
-  // invalidated upon a `rehash()`.
-  using Base::rehash;
-
-  // node_hash_set::reserve(count)
-  //
-  // Sets the number of slots in the `node_hash_set` to the number needed to
-  // accommodate at least `count` total elements without exceeding the current
-  // maximum load factor, and may rehash the container if needed.
-  using Base::reserve;
-
-  // node_hash_set::contains()
-  //
-  // Determines whether an element comparing equal to the given `key` exists
-  // within the `node_hash_set`, returning `true` if so or `false` otherwise.
-  using Base::contains;
-
-  // node_hash_set::count(const Key& key) const
-  //
-  // Returns the number of elements comparing equal to the given `key` within
-  // the `node_hash_set`. note that this function will return either `1` or `0`
-  // since duplicate elements are not allowed within a `node_hash_set`.
-  using Base::count;
-
-  // node_hash_set::equal_range()
-  //
-  // Returns a closed range [first, last], defined by a `std::pair` of two
-  // iterators, containing all elements with the passed key in the
-  // `node_hash_set`.
-  using Base::equal_range;
-
-  // node_hash_set::find()
-  //
-  // Finds an element with the passed `key` within the `node_hash_set`.
-  using Base::find;
-
-  // node_hash_set::bucket_count()
-  //
-  // Returns the number of "buckets" within the `node_hash_set`. Note that
-  // because a node hash set contains all elements within its internal storage,
-  // this value simply equals the current capacity of the `node_hash_set`.
-  using Base::bucket_count;
-
-  // node_hash_set::load_factor()
-  //
-  // Returns the current load factor of the `node_hash_set` (the average number
-  // of slots occupied with a value within the hash set).
-  using Base::load_factor;
-
-  // node_hash_set::max_load_factor()
-  //
-  // Manages the maximum load factor of the `node_hash_set`. Overloads are
-  // listed below.
-  //
-  // float node_hash_set::max_load_factor()
-  //
-  //   Returns the current maximum load factor of the `node_hash_set`.
-  //
-  // void node_hash_set::max_load_factor(float ml)
-  //
-  //   Sets the maximum load factor of the `node_hash_set` to the passed value.
-  //
-  //   NOTE: This overload is provided only for API compatibility with the STL;
-  //   `node_hash_set` will ignore any set load factor and manage its rehashing
-  //   internally as an implementation detail.
-  using Base::max_load_factor;
-
-  // node_hash_set::get_allocator()
-  //
-  // Returns the allocator function associated with this `node_hash_set`.
-  using Base::get_allocator;
-
-  // node_hash_set::hash_function()
-  //
-  // Returns the hashing function used to hash the keys within this
-  // `node_hash_set`.
-  using Base::hash_function;
-
-  // node_hash_set::key_eq()
-  //
-  // Returns the function used for comparing keys equality.
-  using Base::key_eq;
-};
-
-// erase_if(node_hash_set<>, Pred)
-//
-// Erases all elements that satisfy the predicate `pred` from the container `c`.
-// Returns the number of erased elements.
-template <typename T, typename H, typename E, typename A, typename Predicate>
-typename node_hash_set<T, H, E, A>::size_type erase_if(
-    node_hash_set<T, H, E, A>& c, Predicate pred) {
-  return container_internal::EraseIf(pred, &c);
-}
-
-namespace container_internal {
-
-template <class T>
-struct NodeHashSetPolicy
-    : absl::container_internal::node_slot_policy<T&, NodeHashSetPolicy<T>> {
-  using key_type = T;
-  using init_type = T;
-  using constant_iterators = std::true_type;
-
-  template <class Allocator, class... Args>
-  static T* new_element(Allocator* alloc, Args&&... args) {
-    using ValueAlloc =
-        typename absl::allocator_traits<Allocator>::template rebind_alloc<T>;
-    ValueAlloc value_alloc(*alloc);
-    T* res = absl::allocator_traits<ValueAlloc>::allocate(value_alloc, 1);
-    absl::allocator_traits<ValueAlloc>::construct(value_alloc, res,
-                                                  std::forward<Args>(args)...);
-    return res;
-  }
-
-  template <class Allocator>
-  static void delete_element(Allocator* alloc, T* elem) {
-    using ValueAlloc =
-        typename absl::allocator_traits<Allocator>::template rebind_alloc<T>;
-    ValueAlloc value_alloc(*alloc);
-    absl::allocator_traits<ValueAlloc>::destroy(value_alloc, elem);
-    absl::allocator_traits<ValueAlloc>::deallocate(value_alloc, elem, 1);
-  }
-
-  template <class F, class... Args>
-  static decltype(absl::container_internal::DecomposeValue(
-      std::declval<F>(), std::declval<Args>()...))
-  apply(F&& f, Args&&... args) {
-    return absl::container_internal::DecomposeValue(
-        std::forward<F>(f), std::forward<Args>(args)...);
-  }
-
-  static size_t element_space_used(const T*) { return sizeof(T); }
-
-  template <class Hash>
-  static constexpr HashSlotFn get_hash_slot_fn() {
-    return &TypeErasedDerefAndApplyToSlotFn<Hash, T>;
-  }
-};
-}  // namespace container_internal
-
-namespace container_algorithm_internal {
-
-// Specialization of trait in absl/algorithm/container.h
-template <class Key, class Hash, class KeyEqual, class Allocator>
-struct IsUnorderedContainer<absl::node_hash_set<Key, Hash, KeyEqual, Allocator>>
-    : std::true_type {};
-
-}  // namespace container_algorithm_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CONTAINER_NODE_HASH_SET_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/crc/crc32c.h b/projects/llm_framework/include/abseil-cpp/absl/crc/crc32c.h
deleted file mode 100644
index 362861e4..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/crc/crc32c.h
+++ /dev/null
@@ -1,190 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: crc32c.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines the API for computing CRC32C values as checksums
-// for arbitrary sequences of bytes provided as a string buffer.
-//
-// The API includes the basic functions for computing such CRC32C values and
-// some utility functions for performing more efficient mathematical
-// computations using an existing checksum.
-#ifndef ABSL_CRC_CRC32C_H_
-#define ABSL_CRC_CRC32C_H_
-
-#include <cstdint>
-#include <ostream>
-
-#include "absl/crc/internal/crc32c_inline.h"
-#include "absl/strings/str_format.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-//-----------------------------------------------------------------------------
-// crc32c_t
-//-----------------------------------------------------------------------------
-
-// `crc32c_t` defines a strongly-typed integer for holding a CRC32C value.
-//
-// Some operators are intentionally omitted. Only equality operators are defined
-// so that `crc32c_t` can be directly compared. Methods for putting `crc32c_t`
-// directly into a set are omitted because this is bug-prone due to checksum
-// collisions. Use an explicit conversion to the `uint32_t` space for operations
-// that treat `crc32c_t` as an integer.
-class crc32c_t final {
- public:
-  crc32c_t() = default;
-  constexpr explicit crc32c_t(uint32_t crc) : crc_(crc) {}
-
-  crc32c_t(const crc32c_t&) = default;
-  crc32c_t& operator=(const crc32c_t&) = default;
-
-  explicit operator uint32_t() const { return crc_; }
-
-  friend bool operator==(crc32c_t lhs, crc32c_t rhs) {
-    return static_cast<uint32_t>(lhs) == static_cast<uint32_t>(rhs);
-  }
-
-  friend bool operator!=(crc32c_t lhs, crc32c_t rhs) { return !(lhs == rhs); }
-
-  template <typename Sink>
-  friend void AbslStringify(Sink& sink, crc32c_t crc) {
-    absl::Format(&sink, "%08x", static_cast<uint32_t>(crc));
-  }
-
- private:
-  uint32_t crc_;
-};
-
-
-namespace crc_internal {
-// Non-inline code path for `absl::ExtendCrc32c()`. Do not call directly.
-// Call `absl::ExtendCrc32c()` (defined below) instead.
-crc32c_t ExtendCrc32cInternal(crc32c_t initial_crc,
-                              absl::string_view buf_to_add);
-}  // namespace crc_internal
-
-// -----------------------------------------------------------------------------
-// CRC32C Computation Functions
-// -----------------------------------------------------------------------------
-
-// ComputeCrc32c()
-//
-// Returns the CRC32C value of the provided string.
-crc32c_t ComputeCrc32c(absl::string_view buf);
-
-// ExtendCrc32c()
-//
-// Computes a CRC32C value from an `initial_crc` CRC32C value including the
-// `buf_to_add` bytes of an additional buffer. Using this function is more
-// efficient than computing a CRC32C value for the combined buffer from
-// scratch.
-//
-// Note: `ExtendCrc32c` with an initial_crc of 0 is equivalent to
-// `ComputeCrc32c`.
-//
-// This operation has a runtime cost of O(`buf_to_add.size()`)
-inline crc32c_t ExtendCrc32c(crc32c_t initial_crc,
-                             absl::string_view buf_to_add) {
-  // Approximately 75% of calls have size <= 64.
-  if (buf_to_add.size() <= 64) {
-    uint32_t crc = static_cast<uint32_t>(initial_crc);
-    if (crc_internal::ExtendCrc32cInline(&crc, buf_to_add.data(),
-                                         buf_to_add.size())) {
-      return crc32c_t{crc};
-    }
-  }
-  return crc_internal::ExtendCrc32cInternal(initial_crc, buf_to_add);
-}
-
-// ExtendCrc32cByZeroes()
-//
-// Computes a CRC32C value for a buffer with an `initial_crc` CRC32C value,
-// where `length` bytes with a value of 0 are appended to the buffer. Using this
-// function is more efficient than computing a CRC32C value for the combined
-// buffer from scratch.
-//
-// This operation has a runtime cost of O(log(`length`))
-crc32c_t ExtendCrc32cByZeroes(crc32c_t initial_crc, size_t length);
-
-// MemcpyCrc32c()
-//
-// Copies `src` to `dest` using `memcpy()` semantics, returning the CRC32C
-// value of the copied buffer.
-//
-// Using `MemcpyCrc32c()` is potentially faster than performing the `memcpy()`
-// and `ComputeCrc32c()` operations separately.
-crc32c_t MemcpyCrc32c(void* dest, const void* src, size_t count,
-                      crc32c_t initial_crc = crc32c_t{0});
-
-// -----------------------------------------------------------------------------
-// CRC32C Arithmetic Functions
-// -----------------------------------------------------------------------------
-
-// The following functions perform arithmetic on CRC32C values, which are
-// generally more efficient than recalculating any given result's CRC32C value.
-
-// ConcatCrc32c()
-//
-// Calculates the CRC32C value of two buffers with known CRC32C values
-// concatenated together.
-//
-// Given a buffer with CRC32C value `crc1` and a buffer with
-// CRC32C value `crc2` and length, `crc2_length`, returns the CRC32C value of
-// the concatenation of these two buffers.
-//
-// This operation has a runtime cost of O(log(`crc2_length`)).
-crc32c_t ConcatCrc32c(crc32c_t crc1, crc32c_t crc2, size_t crc2_length);
-
-// RemoveCrc32cPrefix()
-//
-// Calculates the CRC32C value of an existing buffer with a series of bytes
-// (the prefix) removed from the beginning of that buffer.
-//
-// Given the CRC32C value of an existing buffer, `full_string_crc`; The CRC32C
-// value of a prefix of that buffer, `prefix_crc`; and the length of the buffer
-// with the prefix removed, `remaining_string_length` , return the CRC32C
-// value of the buffer with the prefix removed.
-//
-// This operation has a runtime cost of O(log(`remaining_string_length`)).
-crc32c_t RemoveCrc32cPrefix(crc32c_t prefix_crc, crc32c_t full_string_crc,
-                            size_t remaining_string_length);
-// RemoveCrc32cSuffix()
-//
-// Calculates the CRC32C value of an existing buffer with a series of bytes
-// (the suffix) removed from the end of that buffer.
-//
-// Given a CRC32C value of an existing buffer `full_string_crc`, the CRC32C
-// value of the suffix to remove `suffix_crc`, and the length of that suffix
-// `suffix_len`, returns the CRC32C value of the buffer with suffix removed.
-//
-// This operation has a runtime cost of O(log(`suffix_len`))
-crc32c_t RemoveCrc32cSuffix(crc32c_t full_string_crc, crc32c_t suffix_crc,
-                            size_t suffix_length);
-
-// operator<<
-//
-// Streams the CRC32C value `crc` to the stream `os`.
-inline std::ostream& operator<<(std::ostream& os, crc32c_t crc) {
-  return os << absl::StreamFormat("%08x", static_cast<uint32_t>(crc));
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CRC_CRC32C_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/cpu_detect.h b/projects/llm_framework/include/abseil-cpp/absl/crc/internal/cpu_detect.h
deleted file mode 100644
index 01e19590..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/cpu_detect.h
+++ /dev/null
@@ -1,63 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CRC_INTERNAL_CPU_DETECT_H_
-#define ABSL_CRC_INTERNAL_CPU_DETECT_H_
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace crc_internal {
-
-// Enumeration of architectures that we have special-case tuning parameters for.
-// This set may change over time.
-enum class CpuType {
-  kUnknown,
-  kIntelHaswell,
-  kAmdRome,
-  kAmdNaples,
-  kAmdMilan,
-  kAmdGenoa,
-  kAmdRyzenV3000,
-  kIntelCascadelakeXeon,
-  kIntelSkylakeXeon,
-  kIntelBroadwell,
-  kIntelSkylake,
-  kIntelIvybridge,
-  kIntelSandybridge,
-  kIntelWestmere,
-  kArmNeoverseN1,
-  kArmNeoverseV1,
-  kAmpereSiryn,
-  kArmNeoverseN2,
-  kArmNeoverseV2
-};
-
-// Returns the type of host CPU this code is running on.  Returns kUnknown if
-// the host CPU is of unknown type, or if detection otherwise fails.
-CpuType GetCpuType();
-
-// Returns whether the host CPU supports the CPU features needed for our
-// accelerated implementations. The CpuTypes enumerated above apart from
-// kUnknown support the required features. On unknown CPUs, we can use
-// this to see if it's safe to use hardware acceleration, though without any
-// tuning.
-bool SupportsArmCRC32PMULL();
-
-}  // namespace crc_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CRC_INTERNAL_CPU_DETECT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc.h b/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc.h
deleted file mode 100644
index 4efdd032..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc.h
+++ /dev/null
@@ -1,83 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CRC_INTERNAL_CRC_H_
-#define ABSL_CRC_INTERNAL_CRC_H_
-
-#include <cstdint>
-
-#include "absl/base/config.h"
-
-// This class implements CRCs (aka Rabin Fingerprints).
-// Treats the input as a polynomial with coefficients in Z(2),
-// and finds the remainder when divided by an primitive polynomial
-// of the appropriate length.
-
-// A polynomial is represented by the bit pattern formed by its coefficients,
-// but with the highest order bit not stored.
-// The highest degree coefficient is stored in the lowest numbered bit
-// in the lowest addressed byte.   Thus, in what follows, the highest degree
-// coefficient that is stored is in the low order bit of "lo" or "*lo".
-
-// Hardware acceleration is used when available.
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace crc_internal {
-
-class CRC {
- public:
-  virtual ~CRC();
-
-  // If "*crc" is the CRC of bytestring A, place the CRC of
-  // the bytestring formed from the concatenation of A and the "length"
-  // bytes at "bytes" into "*crc".
-  virtual void Extend(uint32_t* crc, const void* bytes,
-                      size_t length) const = 0;
-
-  // Equivalent to Extend(crc, bytes, length) where "bytes"
-  // points to an array of "length" zero bytes.
-  virtual void ExtendByZeroes(uint32_t* crc, size_t length) const = 0;
-
-  // Inverse operation of ExtendByZeroes.  If `crc` is the CRC value of a string
-  // ending in `length` zero bytes, this returns a CRC value of that string
-  // with those zero bytes removed.
-  virtual void UnextendByZeroes(uint32_t* crc, size_t length) const = 0;
-
-  // Apply a non-linear transformation to "*crc" so that
-  // it is safe to CRC the result with the same polynomial without
-  // any reduction of error-detection ability in the outer CRC.
-  // Unscramble() performs the inverse transformation.
-  // It is strongly recommended that CRCs be scrambled before storage or
-  // transmission, and unscrambled at the other end before further manipulation.
-  virtual void Scramble(uint32_t* crc) const = 0;
-  virtual void Unscramble(uint32_t* crc) const = 0;
-
-  // Crc32c() returns the singleton implementation of CRC for the CRC32C
-  // polynomial.  Returns a handle that MUST NOT be destroyed with delete.
-  static CRC* Crc32c();
-
- protected:
-  CRC();  // Clients may not call constructor; use Crc32c() instead.
-
- private:
-  CRC(const CRC&) = delete;
-  CRC& operator=(const CRC&) = delete;
-};
-
-}  // namespace crc_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CRC_INTERNAL_CRC_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc32_x86_arm_combined_simd.h b/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc32_x86_arm_combined_simd.h
deleted file mode 100644
index 1938a9b2..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc32_x86_arm_combined_simd.h
+++ /dev/null
@@ -1,299 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CRC_INTERNAL_CRC32_X86_ARM_COMBINED_SIMD_H_
-#define ABSL_CRC_INTERNAL_CRC32_X86_ARM_COMBINED_SIMD_H_
-
-#include <cstdint>
-
-#include "absl/base/config.h"
-
-// -------------------------------------------------------------------------
-// Many x86 and ARM machines have CRC acceleration hardware.
-// We can do a faster version of Extend() on such machines.
-// We define a translation layer for both x86 and ARM for the ease of use and
-// most performance gains.
-
-// This implementation requires 64-bit CRC instructions (part of SSE 4.2) and
-// PCLMULQDQ instructions. 32-bit builds with SSE 4.2 do exist, so the
-// __x86_64__ condition is necessary.
-#if defined(__x86_64__) && defined(__SSE4_2__) && defined(__PCLMUL__)
-
-#include <x86intrin.h>
-#define ABSL_CRC_INTERNAL_HAVE_X86_SIMD
-
-#elif defined(_MSC_VER) && !defined(__clang__) && defined(__AVX__) && \
-    defined(_M_AMD64)
-
-// MSVC AVX (/arch:AVX) implies SSE 4.2 and PCLMULQDQ.
-#include <intrin.h>
-#define ABSL_CRC_INTERNAL_HAVE_X86_SIMD
-
-#elif defined(__aarch64__) && defined(__LITTLE_ENDIAN__) &&                 \
-    defined(__ARM_FEATURE_CRC32) && defined(ABSL_INTERNAL_HAVE_ARM_NEON) && \
-    defined(__ARM_FEATURE_CRYPTO)
-
-#include <arm_acle.h>
-#include <arm_neon.h>
-#define ABSL_CRC_INTERNAL_HAVE_ARM_SIMD
-
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace crc_internal {
-
-#if defined(ABSL_CRC_INTERNAL_HAVE_ARM_SIMD) || \
-    defined(ABSL_CRC_INTERNAL_HAVE_X86_SIMD)
-
-#if defined(ABSL_CRC_INTERNAL_HAVE_ARM_SIMD)
-using V128 = uint64x2_t;
-#else
-// Note: Do not use __m128i_u, it is not portable.
-// Use V128_LoadU() perform an unaligned load from __m128i*.
-using V128 = __m128i;
-#endif
-
-// Starting with the initial value in |crc|, accumulates a CRC32 value for
-// unsigned integers of different sizes.
-uint32_t CRC32_u8(uint32_t crc, uint8_t v);
-
-uint32_t CRC32_u16(uint32_t crc, uint16_t v);
-
-uint32_t CRC32_u32(uint32_t crc, uint32_t v);
-
-uint32_t CRC32_u64(uint32_t crc, uint64_t v);
-
-// Loads 128 bits of integer data. |src| must be 16-byte aligned.
-V128 V128_Load(const V128* src);
-
-// Load 128 bits of integer data. |src| does not need to be aligned.
-V128 V128_LoadU(const V128* src);
-
-// Store 128 bits of integer data. |src| must be 16-byte aligned.
-void V128_Store(V128* dst, V128 data);
-
-// Polynomially multiplies the high 64 bits of |l| and |r|.
-V128 V128_PMulHi(const V128 l, const V128 r);
-
-// Polynomially multiplies the low 64 bits of |l| and |r|.
-V128 V128_PMulLow(const V128 l, const V128 r);
-
-// Polynomially multiplies the low 64 bits of |r| and high 64 bits of |l|.
-V128 V128_PMul01(const V128 l, const V128 r);
-
-// Polynomially multiplies the low 64 bits of |l| and high 64 bits of |r|.
-V128 V128_PMul10(const V128 l, const V128 r);
-
-// Produces a XOR operation of |l| and |r|.
-V128 V128_Xor(const V128 l, const V128 r);
-
-// Produces an AND operation of |l| and |r|.
-V128 V128_And(const V128 l, const V128 r);
-
-// Sets two 64 bit integers to one 128 bit vector. The order is reverse.
-// dst[63:0] := |r|
-// dst[127:64] := |l|
-V128 V128_From2x64(const uint64_t l, const uint64_t r);
-
-// Shift |l| right by |imm| bytes while shifting in zeros.
-template <int imm>
-V128 V128_ShiftRight(const V128 l);
-
-// Extracts a 32-bit integer from |l|, selected with |imm|.
-template <int imm>
-int V128_Extract32(const V128 l);
-
-// Extracts a 64-bit integer from |l|, selected with |imm|.
-template <int imm>
-uint64_t V128_Extract64(const V128 l);
-
-// Extracts the low 64 bits from V128.
-int64_t V128_Low64(const V128 l);
-
-// Left-shifts packed 64-bit integers in l by r.
-V128 V128_ShiftLeft64(const V128 l, const V128 r);
-
-#endif
-
-#if defined(ABSL_CRC_INTERNAL_HAVE_X86_SIMD)
-
-inline uint32_t CRC32_u8(uint32_t crc, uint8_t v) {
-  return _mm_crc32_u8(crc, v);
-}
-
-inline uint32_t CRC32_u16(uint32_t crc, uint16_t v) {
-  return _mm_crc32_u16(crc, v);
-}
-
-inline uint32_t CRC32_u32(uint32_t crc, uint32_t v) {
-  return _mm_crc32_u32(crc, v);
-}
-
-inline uint32_t CRC32_u64(uint32_t crc, uint64_t v) {
-  return static_cast<uint32_t>(_mm_crc32_u64(crc, v));
-}
-
-inline V128 V128_Load(const V128* src) { return _mm_load_si128(src); }
-
-inline V128 V128_LoadU(const V128* src) { return _mm_loadu_si128(src); }
-
-inline void V128_Store(V128* dst, V128 data) { _mm_store_si128(dst, data); }
-
-inline V128 V128_PMulHi(const V128 l, const V128 r) {
-  return _mm_clmulepi64_si128(l, r, 0x11);
-}
-
-inline V128 V128_PMulLow(const V128 l, const V128 r) {
-  return _mm_clmulepi64_si128(l, r, 0x00);
-}
-
-inline V128 V128_PMul01(const V128 l, const V128 r) {
-  return _mm_clmulepi64_si128(l, r, 0x01);
-}
-
-inline V128 V128_PMul10(const V128 l, const V128 r) {
-  return _mm_clmulepi64_si128(l, r, 0x10);
-}
-
-inline V128 V128_Xor(const V128 l, const V128 r) { return _mm_xor_si128(l, r); }
-
-inline V128 V128_And(const V128 l, const V128 r) { return _mm_and_si128(l, r); }
-
-inline V128 V128_From2x64(const uint64_t l, const uint64_t r) {
-  return _mm_set_epi64x(static_cast<int64_t>(l), static_cast<int64_t>(r));
-}
-
-template <int imm>
-inline V128 V128_ShiftRight(const V128 l) {
-  return _mm_srli_si128(l, imm);
-}
-
-template <int imm>
-inline int V128_Extract32(const V128 l) {
-  return _mm_extract_epi32(l, imm);
-}
-
-template <int imm>
-inline uint64_t V128_Extract64(const V128 l) {
-  return static_cast<uint64_t>(_mm_extract_epi64(l, imm));
-}
-
-inline int64_t V128_Low64(const V128 l) { return _mm_cvtsi128_si64(l); }
-
-inline V128 V128_ShiftLeft64(const V128 l, const V128 r) {
-  return _mm_sll_epi64(l, r);
-}
-
-#elif defined(ABSL_CRC_INTERNAL_HAVE_ARM_SIMD)
-
-inline uint32_t CRC32_u8(uint32_t crc, uint8_t v) { return __crc32cb(crc, v); }
-
-inline uint32_t CRC32_u16(uint32_t crc, uint16_t v) {
-  return __crc32ch(crc, v);
-}
-
-inline uint32_t CRC32_u32(uint32_t crc, uint32_t v) {
-  return __crc32cw(crc, v);
-}
-
-inline uint32_t CRC32_u64(uint32_t crc, uint64_t v) {
-  return __crc32cd(crc, v);
-}
-
-inline V128 V128_Load(const V128* src) {
-  return vld1q_u64(reinterpret_cast<const uint64_t*>(src));
-}
-
-inline V128 V128_LoadU(const V128* src) {
-  return vld1q_u64(reinterpret_cast<const uint64_t*>(src));
-}
-
-inline void V128_Store(V128* dst, V128 data) {
-  vst1q_u64(reinterpret_cast<uint64_t*>(dst), data);
-}
-
-// Using inline assembly as clang does not generate the pmull2 instruction and
-// performance drops by 15-20%.
-// TODO(b/193678732): Investigate why there is a slight performance hit when
-// using intrinsics instead of inline assembly.
-inline V128 V128_PMulHi(const V128 l, const V128 r) {
-  uint64x2_t res;
-  __asm__ __volatile__("pmull2 %0.1q, %1.2d, %2.2d \n\t"
-                       : "=w"(res)
-                       : "w"(l), "w"(r));
-  return res;
-}
-
-// TODO(b/193678732): Investigate why the compiler decides to move the constant
-// loop multiplicands from GPR to Neon registers every loop iteration.
-inline V128 V128_PMulLow(const V128 l, const V128 r) {
-  uint64x2_t res;
-  __asm__ __volatile__("pmull %0.1q, %1.1d, %2.1d \n\t"
-                       : "=w"(res)
-                       : "w"(l), "w"(r));
-  return res;
-}
-
-inline V128 V128_PMul01(const V128 l, const V128 r) {
-  return reinterpret_cast<V128>(vmull_p64(
-      reinterpret_cast<poly64_t>(vget_high_p64(vreinterpretq_p64_u64(l))),
-      reinterpret_cast<poly64_t>(vget_low_p64(vreinterpretq_p64_u64(r)))));
-}
-
-inline V128 V128_PMul10(const V128 l, const V128 r) {
-  return reinterpret_cast<V128>(vmull_p64(
-      reinterpret_cast<poly64_t>(vget_low_p64(vreinterpretq_p64_u64(l))),
-      reinterpret_cast<poly64_t>(vget_high_p64(vreinterpretq_p64_u64(r)))));
-}
-
-inline V128 V128_Xor(const V128 l, const V128 r) { return veorq_u64(l, r); }
-
-inline V128 V128_And(const V128 l, const V128 r) { return vandq_u64(l, r); }
-
-inline V128 V128_From2x64(const uint64_t l, const uint64_t r) {
-  return vcombine_u64(vcreate_u64(r), vcreate_u64(l));
-}
-
-template <int imm>
-inline V128 V128_ShiftRight(const V128 l) {
-  return vreinterpretq_u64_s8(
-      vextq_s8(vreinterpretq_s8_u64(l), vdupq_n_s8(0), imm));
-}
-
-template <int imm>
-inline int V128_Extract32(const V128 l) {
-  return vgetq_lane_s32(vreinterpretq_s32_u64(l), imm);
-}
-
-template <int imm>
-inline uint64_t V128_Extract64(const V128 l) {
-  return vgetq_lane_u64(l, imm);
-}
-
-inline int64_t V128_Low64(const V128 l) {
-  return vgetq_lane_s64(vreinterpretq_s64_u64(l), 0);
-}
-
-inline V128 V128_ShiftLeft64(const V128 l, const V128 r) {
-  return vshlq_u64(l, vreinterpretq_s64_u64(r));
-}
-
-#endif
-
-}  // namespace crc_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CRC_INTERNAL_CRC32_X86_ARM_COMBINED_SIMD_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc32c.h b/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc32c.h
deleted file mode 100644
index 34027c55..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc32c.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CRC_INTERNAL_CRC32C_H_
-#define ABSL_CRC_INTERNAL_CRC32C_H_
-
-#include "absl/base/config.h"
-#include "absl/crc/crc32c.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace crc_internal {
-
-// Modifies a CRC32 value by removing `length` bytes with a value of 0 from
-// the end of the string.
-//
-// This is the inverse operation of ExtendCrc32cByZeroes().
-//
-// This operation has a runtime cost of O(log(`length`))
-//
-// Internal implementation detail, exposed for testing only.
-crc32c_t UnextendCrc32cByZeroes(crc32c_t initial_crc, size_t length);
-
-}  // namespace crc_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CRC_INTERNAL_CRC32C_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc32c_inline.h b/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc32c_inline.h
deleted file mode 100644
index 6236c10b..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc32c_inline.h
+++ /dev/null
@@ -1,72 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CRC_INTERNAL_CRC32C_INLINE_H_
-#define ABSL_CRC_INTERNAL_CRC32C_INLINE_H_
-
-#include <cstdint>
-
-#include "absl/base/config.h"
-#include "absl/base/internal/endian.h"
-#include "absl/crc/internal/crc32_x86_arm_combined_simd.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace crc_internal {
-
-// CRC32C implementation optimized for small inputs.
-// Either computes crc and return true, or if there is
-// no hardware support does nothing and returns false.
-inline bool ExtendCrc32cInline(uint32_t* crc, const char* p, size_t n) {
-#if defined(ABSL_CRC_INTERNAL_HAVE_ARM_SIMD) || \
-    defined(ABSL_CRC_INTERNAL_HAVE_X86_SIMD)
-  constexpr uint32_t kCrc32Xor = 0xffffffffU;
-  *crc ^= kCrc32Xor;
-  if (n & 1) {
-    *crc = CRC32_u8(*crc, static_cast<uint8_t>(*p));
-    n--;
-    p++;
-  }
-  if (n & 2) {
-    *crc = CRC32_u16(*crc, absl::little_endian::Load16(p));
-    n -= 2;
-    p += 2;
-  }
-  if (n & 4) {
-    *crc = CRC32_u32(*crc, absl::little_endian::Load32(p));
-    n -= 4;
-    p += 4;
-  }
-  while (n) {
-    *crc = CRC32_u64(*crc, absl::little_endian::Load64(p));
-    n -= 8;
-    p += 8;
-  }
-  *crc ^= kCrc32Xor;
-  return true;
-#else
-  // No hardware support, signal the need to fallback.
-  static_cast<void>(crc);
-  static_cast<void>(p);
-  static_cast<void>(n);
-  return false;
-#endif  // defined(ABSL_CRC_INTERNAL_HAVE_ARM_SIMD) ||
-        // defined(ABSL_CRC_INTERNAL_HAVE_X86_SIMD)
-}
-
-}  // namespace crc_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CRC_INTERNAL_CRC32C_INLINE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc_cord_state.h b/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc_cord_state.h
deleted file mode 100644
index fbbb8c00..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc_cord_state.h
+++ /dev/null
@@ -1,159 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CRC_INTERNAL_CRC_CORD_STATE_H_
-#define ABSL_CRC_INTERNAL_CRC_CORD_STATE_H_
-
-#include <atomic>
-#include <cstddef>
-#include <deque>
-
-#include "absl/base/config.h"
-#include "absl/crc/crc32c.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace crc_internal {
-
-// CrcCordState is a copy-on-write class that holds the chunked CRC32C data
-// that allows CrcCord to perform efficient substring operations. CrcCordState
-// is used as a member variable in CrcCord. When a CrcCord is converted to a
-// Cord, the CrcCordState is shallow-copied into the root node of the Cord. If
-// the converted Cord is modified outside of CrcCord, the CrcCordState is
-// discarded from the Cord. If the Cord is converted back to a CrcCord, and the
-// Cord is still carrying the CrcCordState in its root node, the CrcCord can
-// re-use the CrcCordState, making the construction of the CrcCord cheap.
-//
-// CrcCordState does not try to encapsulate the CRC32C state (CrcCord requires
-// knowledge of how CrcCordState represents the CRC32C state). It does
-// encapsulate the copy-on-write nature of the state.
-class CrcCordState {
- public:
-  // Constructors.
-  CrcCordState();
-  CrcCordState(const CrcCordState&);
-  CrcCordState(CrcCordState&&);
-
-  // Destructor. Atomically unreferences the data.
-  ~CrcCordState();
-
-  // Copy and move operators.
-  CrcCordState& operator=(const CrcCordState&);
-  CrcCordState& operator=(CrcCordState&&);
-
-  // A (length, crc) pair.
-  struct PrefixCrc {
-    PrefixCrc() = default;
-    PrefixCrc(size_t length_arg, absl::crc32c_t crc_arg)
-        : length(length_arg), crc(crc_arg) {}
-
-    size_t length = 0;
-
-    // TODO(absl-team): Memory stomping often zeros out memory. If this struct
-    // gets overwritten, we could end up with {0, 0}, which is the correct CRC
-    // for a string of length 0. Consider storing a scrambled value and
-    // unscrambling it before verifying it.
-    absl::crc32c_t crc = absl::crc32c_t{0};
-  };
-
-  // The representation of the chunked CRC32C data.
-  struct Rep {
-    // `removed_prefix` is the crc and length of any prefix that has been
-    // removed from the Cord (for example, by calling
-    // `CrcCord::RemovePrefix()`). To get the checksum of any prefix of the
-    // cord, this value must be subtracted from `prefix_crc`. See `Checksum()`
-    // for an example.
-    //
-    // CrcCordState is said to be "normalized" if removed_prefix.length == 0.
-    PrefixCrc removed_prefix;
-
-    // A deque of (length, crc) pairs, representing length and crc of a prefix
-    // of the Cord, before removed_prefix has been subtracted. The lengths of
-    // the prefixes are stored in increasing order. If the Cord is not empty,
-    // the last value in deque is the contains the CRC32C of the entire Cord
-    // when removed_prefix is subtracted from it.
-    std::deque<PrefixCrc> prefix_crc;
-  };
-
-  // Returns a reference to the representation of the chunked CRC32C data.
-  const Rep& rep() const { return refcounted_rep_->rep; }
-
-  // Returns a mutable reference to the representation of the chunked CRC32C
-  // data. Calling this function will copy the data if another instance also
-  // holds a reference to the data, so it is important to call rep() instead if
-  // the data may not be mutated.
-  Rep* mutable_rep() {
-    if (refcounted_rep_->count.load(std::memory_order_acquire) != 1) {
-      RefcountedRep* copy = new RefcountedRep;
-      copy->rep = refcounted_rep_->rep;
-      Unref(refcounted_rep_);
-      refcounted_rep_ = copy;
-    }
-    return &refcounted_rep_->rep;
-  }
-
-  // Returns the CRC32C of the entire Cord.
-  absl::crc32c_t Checksum() const;
-
-  // Returns true if the chunked CRC32C cached is normalized.
-  bool IsNormalized() const { return rep().removed_prefix.length == 0; }
-
-  // Normalizes the chunked CRC32C checksum cache by subtracting any removed
-  // prefix from the chunks.
-  void Normalize();
-
-  // Returns the number of cached chunks.
-  size_t NumChunks() const { return rep().prefix_crc.size(); }
-
-  // Helper that returns the (length, crc) of the `n`-th cached chunked.
-  PrefixCrc NormalizedPrefixCrcAtNthChunk(size_t n) const;
-
-  // Poisons all chunks to so that Checksum() will likely be incorrect with high
-  // probability.
-  void Poison();
-
- private:
-  struct RefcountedRep {
-    std::atomic<int32_t> count{1};
-    Rep rep;
-  };
-
-  // Adds a reference to the shared global empty `RefcountedRep`, and returns a
-  // pointer to the `RefcountedRep`. This is an optimization to avoid unneeded
-  // allocations when the allocation is unlikely to ever be used. The returned
-  // pointer can be `Unref()`ed when it is no longer needed.  Since the returned
-  // instance will always have a reference counter greater than 1, attempts to
-  // modify it (by calling `mutable_rep()`) will create a new unshared copy.
-  static RefcountedRep* RefSharedEmptyRep();
-
-  static void Ref(RefcountedRep* r) {
-    assert(r != nullptr);
-    r->count.fetch_add(1, std::memory_order_relaxed);
-  }
-
-  static void Unref(RefcountedRep* r) {
-    assert(r != nullptr);
-    if (r->count.fetch_sub(1, std::memory_order_acq_rel) == 1) {
-      delete r;
-    }
-  }
-
-  RefcountedRep* refcounted_rep_;
-};
-
-}  // namespace crc_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CRC_INTERNAL_CRC_CORD_STATE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc_internal.h b/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc_internal.h
deleted file mode 100644
index 4d3582d9..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc_internal.h
+++ /dev/null
@@ -1,177 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CRC_INTERNAL_CRC_INTERNAL_H_
-#define ABSL_CRC_INTERNAL_CRC_INTERNAL_H_
-
-#include <cstdint>
-#include <memory>
-#include <vector>
-
-#include "absl/base/internal/raw_logging.h"
-#include "absl/crc/internal/crc.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-namespace crc_internal {
-
-// Prefetch constants used in some Extend() implementations
-constexpr int kPrefetchHorizon = ABSL_CACHELINE_SIZE * 4;  // Prefetch this far
-// Shorter prefetch distance for smaller buffers
-constexpr int kPrefetchHorizonMedium = ABSL_CACHELINE_SIZE * 1;
-static_assert(kPrefetchHorizon >= 64, "CRCPrefetchHorizon less than loop len");
-
-// We require the Scramble() function:
-//  - to be reversible (Unscramble() must exist)
-//  - to be non-linear in the polynomial's Galois field (so the CRC of a
-//    scrambled CRC is not linearly affected by the scrambled CRC, even if
-//    using the same polynomial)
-//  - not to be its own inverse.  Preferably, if X=Scramble^N(X) and N!=0, then
-//    N is large.
-//  - to be fast.
-//  - not to change once defined.
-// We introduce non-linearity in two ways:
-//     Addition of a constant.
-//         - The carries introduce non-linearity; we use bits of an irrational
-//           (phi) to make it unlikely that we introduce no carries.
-//     Rotate by a constant number of bits.
-//         - We use floor(degree/2)+1, which does not divide the degree, and
-//           splits the bits nearly evenly, which makes it less likely the
-//           halves will be the same or one will be all zeroes.
-// We do both things to improve the chances of non-linearity in the face of
-// bit patterns with low numbers of bits set, while still being fast.
-// Below is the constant that we add.  The bits are the first 128 bits of the
-// fractional part of phi, with a 1 ored into the bottom bit to maximize the
-// cycle length of repeated adds.
-constexpr uint64_t kScrambleHi = (static_cast<uint64_t>(0x4f1bbcdcU) << 32) |
-                                 static_cast<uint64_t>(0xbfa53e0aU);
-constexpr uint64_t kScrambleLo = (static_cast<uint64_t>(0xf9ce6030U) << 32) |
-                                 static_cast<uint64_t>(0x2e76e41bU);
-
-class CRCImpl : public CRC {  // Implementation of the abstract class CRC
- public:
-  using Uint32By256 = uint32_t[256];
-
-  CRCImpl() = default;
-  ~CRCImpl() override = default;
-
-  // The internal version of CRC::New().
-  static CRCImpl* NewInternal();
-
-  // Fill in a table for updating a CRC by one word of 'word_size' bytes
-  // [last_lo, last_hi] contains the answer if the last bit in the word
-  // is set.
-  static void FillWordTable(uint32_t poly, uint32_t last, int word_size,
-                            Uint32By256* t);
-
-  // Build the table for extending by zeroes, returning the number of entries.
-  // For a in {1, 2, ..., ZEROES_BASE-1}, b in {0, 1, 2, 3, ...},
-  // entry j=a-1+(ZEROES_BASE-1)*b
-  // contains a polynomial Pi such that multiplying
-  // a CRC by Pi mod P, where P is the CRC polynomial, is equivalent to
-  // appending a*2**(ZEROES_BASE_LG*b) zero bytes to the original string.
-  static int FillZeroesTable(uint32_t poly, Uint32By256* t);
-
-  virtual void InitTables() = 0;
-
- private:
-  CRCImpl(const CRCImpl&) = delete;
-  CRCImpl& operator=(const CRCImpl&) = delete;
-};
-
-// This is the 32-bit implementation.  It handles all sizes from 8 to 32.
-class CRC32 : public CRCImpl {
- public:
-  CRC32() = default;
-  ~CRC32() override = default;
-
-  void Extend(uint32_t* crc, const void* bytes, size_t length) const override;
-  void ExtendByZeroes(uint32_t* crc, size_t length) const override;
-  void Scramble(uint32_t* crc) const override;
-  void Unscramble(uint32_t* crc) const override;
-  void UnextendByZeroes(uint32_t* crc, size_t length) const override;
-
-  void InitTables() override;
-
- private:
-  // Common implementation guts for ExtendByZeroes and UnextendByZeroes().
-  //
-  // zeroes_table is a table as returned by FillZeroesTable(), containing
-  // polynomials representing CRCs of strings-of-zeros of various lengths,
-  // and which can be combined by polynomial multiplication.  poly_table is
-  // a table of CRC byte extension values.  These tables are determined by
-  // the generator polynomial.
-  //
-  // These will be set to reverse_zeroes_ and reverse_table0_ for Unextend, and
-  // CRC32::zeroes_ and CRC32::table0_ for Extend.
-  static void ExtendByZeroesImpl(uint32_t* crc, size_t length,
-                                 const uint32_t zeroes_table[256],
-                                 const uint32_t poly_table[256]);
-
-  uint32_t table0_[256];  // table of byte extensions
-  uint32_t zeroes_[256];  // table of zero extensions
-
-  // table of 4-byte extensions shifted by 12 bytes of zeroes
-  uint32_t table_[4][256];
-
-  // Reverse lookup tables, using the alternate polynomial used by
-  // UnextendByZeroes().
-  uint32_t reverse_table0_[256];  // table of reverse byte extensions
-  uint32_t reverse_zeroes_[256];  // table of reverse zero extensions
-
-  CRC32(const CRC32&) = delete;
-  CRC32& operator=(const CRC32&) = delete;
-};
-
-// Helpers
-
-// Return a bit mask containing len 1-bits.
-// Requires 0 < len <= sizeof(T)
-template <typename T>
-T MaskOfLength(int len) {
-  // shift 2 by len-1 rather than 1 by len because shifts of wordsize
-  // are undefined.
-  return (T(2) << (len - 1)) - 1;
-}
-
-// Rotate low-order "width" bits of "in" right by "r" bits,
-// setting other bits in word to arbitrary values.
-template <typename T>
-T RotateRight(T in, int width, int r) {
-  return (in << (width - r)) | ((in >> r) & MaskOfLength<T>(width - r));
-}
-
-// RoundUp<N>(p) returns the lowest address >= p aligned to an N-byte
-// boundary.  Requires that N is a power of 2.
-template <int alignment>
-const uint8_t* RoundUp(const uint8_t* p) {
-  static_assert((alignment & (alignment - 1)) == 0, "alignment is not 2^n");
-  constexpr uintptr_t mask = alignment - 1;
-  const uintptr_t as_uintptr = reinterpret_cast<uintptr_t>(p);
-  return reinterpret_cast<const uint8_t*>((as_uintptr + mask) & ~mask);
-}
-
-// Return a newly created CRC32AcceleratedX86ARMCombined if we can use Intel's
-// or ARM's CRC acceleration for a given polynomial.  Return nullptr otherwise.
-CRCImpl* TryNewCRC32AcceleratedX86ARMCombined();
-
-// Return all possible hardware accelerated implementations. For testing only.
-std::vector<std::unique_ptr<CRCImpl>> NewCRC32AcceleratedX86ARMCombinedAll();
-
-}  // namespace crc_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CRC_INTERNAL_CRC_INTERNAL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc_memcpy.h b/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc_memcpy.h
deleted file mode 100644
index a0fed65a..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/crc_memcpy.h
+++ /dev/null
@@ -1,122 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CRC_INTERNAL_CRC_MEMCPY_H_
-#define ABSL_CRC_INTERNAL_CRC_MEMCPY_H_
-
-#include <cstddef>
-#include <memory>
-
-#include "absl/base/config.h"
-#include "absl/crc/crc32c.h"
-#include "absl/crc/internal/crc32_x86_arm_combined_simd.h"
-
-// Defined if the class AcceleratedCrcMemcpyEngine exists.
-// TODO(b/299127771): Consider relaxing the pclmul requirement once the other
-// intrinsics are conditionally compiled without it.
-#if defined(ABSL_CRC_INTERNAL_HAVE_X86_SIMD)
-#define ABSL_INTERNAL_HAVE_X86_64_ACCELERATED_CRC_MEMCPY_ENGINE 1
-#elif defined(ABSL_CRC_INTERNAL_HAVE_ARM_SIMD)
-#define ABSL_INTERNAL_HAVE_ARM_ACCELERATED_CRC_MEMCPY_ENGINE 1
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace crc_internal {
-
-class CrcMemcpyEngine {
- public:
-  virtual ~CrcMemcpyEngine() = default;
-
-  virtual crc32c_t Compute(void* __restrict dst, const void* __restrict src,
-                           std::size_t length, crc32c_t initial_crc) const = 0;
-
- protected:
-  CrcMemcpyEngine() = default;
-};
-
-class CrcMemcpy {
- public:
-  static crc32c_t CrcAndCopy(void* __restrict dst, const void* __restrict src,
-                             std::size_t length,
-                             crc32c_t initial_crc = crc32c_t{0},
-                             bool non_temporal = false) {
-    static const ArchSpecificEngines engines = GetArchSpecificEngines();
-    auto* engine = non_temporal ? engines.non_temporal : engines.temporal;
-    return engine->Compute(dst, src, length, initial_crc);
-  }
-
-  // For testing only: get an architecture-specific engine for tests.
-  static std::unique_ptr<CrcMemcpyEngine> GetTestEngine(int vector,
-                                                        int integer);
-
- private:
-  struct ArchSpecificEngines {
-    CrcMemcpyEngine* temporal;
-    CrcMemcpyEngine* non_temporal;
-  };
-
-  static ArchSpecificEngines GetArchSpecificEngines();
-};
-
-// Fallback CRC-memcpy engine.
-class FallbackCrcMemcpyEngine : public CrcMemcpyEngine {
- public:
-  FallbackCrcMemcpyEngine() = default;
-  FallbackCrcMemcpyEngine(const FallbackCrcMemcpyEngine&) = delete;
-  FallbackCrcMemcpyEngine operator=(const FallbackCrcMemcpyEngine&) = delete;
-
-  crc32c_t Compute(void* __restrict dst, const void* __restrict src,
-                   std::size_t length, crc32c_t initial_crc) const override;
-};
-
-// CRC Non-Temporal-Memcpy engine.
-class CrcNonTemporalMemcpyEngine : public CrcMemcpyEngine {
- public:
-  CrcNonTemporalMemcpyEngine() = default;
-  CrcNonTemporalMemcpyEngine(const CrcNonTemporalMemcpyEngine&) = delete;
-  CrcNonTemporalMemcpyEngine operator=(const CrcNonTemporalMemcpyEngine&) =
-      delete;
-
-  crc32c_t Compute(void* __restrict dst, const void* __restrict src,
-                   std::size_t length, crc32c_t initial_crc) const override;
-};
-
-// CRC Non-Temporal-Memcpy AVX engine.
-class CrcNonTemporalMemcpyAVXEngine : public CrcMemcpyEngine {
- public:
-  CrcNonTemporalMemcpyAVXEngine() = default;
-  CrcNonTemporalMemcpyAVXEngine(const CrcNonTemporalMemcpyAVXEngine&) = delete;
-  CrcNonTemporalMemcpyAVXEngine operator=(
-      const CrcNonTemporalMemcpyAVXEngine&) = delete;
-
-  crc32c_t Compute(void* __restrict dst, const void* __restrict src,
-                   std::size_t length, crc32c_t initial_crc) const override;
-};
-
-// Copy source to destination and return the CRC32C of the data copied.  If an
-// accelerated version is available, use the accelerated version, otherwise use
-// the generic fallback version.
-inline crc32c_t Crc32CAndCopy(void* __restrict dst, const void* __restrict src,
-                              std::size_t length,
-                              crc32c_t initial_crc = crc32c_t{0},
-                              bool non_temporal = false) {
-  return CrcMemcpy::CrcAndCopy(dst, src, length, initial_crc, non_temporal);
-}
-
-}  // namespace crc_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CRC_INTERNAL_CRC_MEMCPY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/non_temporal_arm_intrinsics.h b/projects/llm_framework/include/abseil-cpp/absl/crc/internal/non_temporal_arm_intrinsics.h
deleted file mode 100644
index 9e5ccfc4..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/non_temporal_arm_intrinsics.h
+++ /dev/null
@@ -1,79 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CRC_INTERNAL_NON_TEMPORAL_ARM_INTRINSICS_H_
-#define ABSL_CRC_INTERNAL_NON_TEMPORAL_ARM_INTRINSICS_H_
-
-#include "absl/base/config.h"
-
-#ifdef __aarch64__
-#include <arm_neon.h>
-
-typedef int64x2_t __m128i; /* 128-bit vector containing integers */
-#define vreinterpretq_m128i_s32(x) vreinterpretq_s64_s32(x)
-#define vreinterpretq_s64_m128i(x) (x)
-
-// Guarantees that every preceding store is globally visible before any
-// subsequent store.
-// https://msdn.microsoft.com/en-us/library/5h2w73d1%28v=vs.90%29.aspx
-static inline __attribute__((always_inline)) void _mm_sfence(void) {
-  __sync_synchronize();
-}
-
-// Load 128-bits of integer data from unaligned memory into dst. This intrinsic
-// may perform better than _mm_loadu_si128 when the data crosses a cache line
-// boundary.
-//
-//   dst[127:0] := MEM[mem_addr+127:mem_addr]
-//
-// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_lddqu_si128
-#define _mm_lddqu_si128 _mm_loadu_si128
-
-// Loads 128-bit value. :
-// https://msdn.microsoft.com/zh-cn/library/f4k12ae8(v=vs.90).aspx
-static inline __attribute__((always_inline)) __m128i _mm_loadu_si128(
-    const __m128i *p) {
-  return vreinterpretq_m128i_s32(vld1q_s32((const int32_t *)p));
-}
-
-// Stores the data in a to the address p without polluting the caches.  If the
-// cache line containing address p is already in the cache, the cache will be
-// updated.
-// https://msdn.microsoft.com/en-us/library/ba08y07y%28v=vs.90%29.aspx
-static inline __attribute__((always_inline)) void _mm_stream_si128(__m128i *p,
-                                                                   __m128i a) {
-#if ABSL_HAVE_BUILTIN(__builtin_nontemporal_store)
-  __builtin_nontemporal_store(a, p);
-#else
-  vst1q_s64((int64_t *)p, vreinterpretq_s64_m128i(a));
-#endif
-}
-
-// Sets the 16 signed 8-bit integer values.
-// https://msdn.microsoft.com/en-us/library/x0cx8zd3(v=vs.90).aspx
-static inline __attribute__((always_inline)) __m128i _mm_set_epi8(
-    signed char b15, signed char b14, signed char b13, signed char b12,
-    signed char b11, signed char b10, signed char b9, signed char b8,
-    signed char b7, signed char b6, signed char b5, signed char b4,
-    signed char b3, signed char b2, signed char b1, signed char b0) {
-  int8_t __attribute__((aligned(16)))
-  data[16] = {(int8_t)b0,  (int8_t)b1,  (int8_t)b2,  (int8_t)b3,
-              (int8_t)b4,  (int8_t)b5,  (int8_t)b6,  (int8_t)b7,
-              (int8_t)b8,  (int8_t)b9,  (int8_t)b10, (int8_t)b11,
-              (int8_t)b12, (int8_t)b13, (int8_t)b14, (int8_t)b15};
-  return (__m128i)vld1q_s8(data);
-}
-#endif  // __aarch64__
-
-#endif  // ABSL_CRC_INTERNAL_NON_TEMPORAL_ARM_INTRINSICS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/non_temporal_memcpy.h b/projects/llm_framework/include/abseil-cpp/absl/crc/internal/non_temporal_memcpy.h
deleted file mode 100644
index b3d94bad..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/crc/internal/non_temporal_memcpy.h
+++ /dev/null
@@ -1,180 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_CRC_INTERNAL_NON_TEMPORAL_MEMCPY_H_
-#define ABSL_CRC_INTERNAL_NON_TEMPORAL_MEMCPY_H_
-
-#ifdef _MSC_VER
-#include <intrin.h>
-#endif
-
-#ifdef __SSE__
-#include <xmmintrin.h>
-#endif
-
-#ifdef __SSE2__
-#include <emmintrin.h>
-#endif
-
-#ifdef __SSE3__
-#include <pmmintrin.h>
-#endif
-
-#ifdef __AVX__
-#include <immintrin.h>
-#endif
-
-#ifdef __aarch64__
-#include "absl/crc/internal/non_temporal_arm_intrinsics.h"
-#endif
-
-#include <algorithm>
-#include <cassert>
-#include <cstdint>
-#include <cstring>
-
-#include "absl/base/config.h"
-#include "absl/base/optimization.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace crc_internal {
-
-// This non-temporal memcpy does regular load and non-temporal store memory
-// copy. It is compatible to both 16-byte aligned and unaligned addresses. If
-// data at the destination is not immediately accessed, using non-temporal
-// memcpy can save 1 DRAM load of the destination cacheline.
-constexpr size_t kCacheLineSize = ABSL_CACHELINE_SIZE;
-
-// If the objects overlap, the behavior is undefined.
-inline void *non_temporal_store_memcpy(void *__restrict dst,
-                                       const void *__restrict src, size_t len) {
-#if defined(__SSE3__) || defined(__aarch64__) || \
-    (defined(_MSC_VER) && defined(__AVX__))
-  // This implementation requires SSE3.
-  // MSVC cannot target SSE3 directly, but when MSVC targets AVX,
-  // SSE3 support is implied.
-  uint8_t *d = reinterpret_cast<uint8_t *>(dst);
-  const uint8_t *s = reinterpret_cast<const uint8_t *>(src);
-
-  // memcpy() the misaligned header. At the end of this if block, <d> is
-  // aligned to a 64-byte cacheline boundary or <len> == 0.
-  if (reinterpret_cast<uintptr_t>(d) & (kCacheLineSize - 1)) {
-    uintptr_t bytes_before_alignment_boundary =
-        kCacheLineSize -
-        (reinterpret_cast<uintptr_t>(d) & (kCacheLineSize - 1));
-    size_t header_len = (std::min)(bytes_before_alignment_boundary, len);
-    assert(bytes_before_alignment_boundary < kCacheLineSize);
-    memcpy(d, s, header_len);
-    d += header_len;
-    s += header_len;
-    len -= header_len;
-  }
-
-  if (len >= kCacheLineSize) {
-    _mm_sfence();
-    __m128i *dst_cacheline = reinterpret_cast<__m128i *>(d);
-    const __m128i *src_cacheline = reinterpret_cast<const __m128i *>(s);
-    constexpr int kOpsPerCacheLine = kCacheLineSize / sizeof(__m128i);
-    size_t loops = len / kCacheLineSize;
-
-    while (len >= kCacheLineSize) {
-      __m128i temp1, temp2, temp3, temp4;
-      temp1 = _mm_lddqu_si128(src_cacheline + 0);
-      temp2 = _mm_lddqu_si128(src_cacheline + 1);
-      temp3 = _mm_lddqu_si128(src_cacheline + 2);
-      temp4 = _mm_lddqu_si128(src_cacheline + 3);
-      _mm_stream_si128(dst_cacheline + 0, temp1);
-      _mm_stream_si128(dst_cacheline + 1, temp2);
-      _mm_stream_si128(dst_cacheline + 2, temp3);
-      _mm_stream_si128(dst_cacheline + 3, temp4);
-      src_cacheline += kOpsPerCacheLine;
-      dst_cacheline += kOpsPerCacheLine;
-      len -= kCacheLineSize;
-    }
-    d += loops * kCacheLineSize;
-    s += loops * kCacheLineSize;
-    _mm_sfence();
-  }
-
-  // memcpy the tail.
-  if (len) {
-    memcpy(d, s, len);
-  }
-  return dst;
-#else
-  // Fallback to regular memcpy.
-  return memcpy(dst, src, len);
-#endif  // __SSE3__ || __aarch64__ || (_MSC_VER && __AVX__)
-}
-
-inline void *non_temporal_store_memcpy_avx(void *__restrict dst,
-                                           const void *__restrict src,
-                                           size_t len) {
-#ifdef __AVX__
-  uint8_t *d = reinterpret_cast<uint8_t *>(dst);
-  const uint8_t *s = reinterpret_cast<const uint8_t *>(src);
-
-  // memcpy() the misaligned header. At the end of this if block, <d> is
-  // aligned to a 64-byte cacheline boundary or <len> == 0.
-  if (reinterpret_cast<uintptr_t>(d) & (kCacheLineSize - 1)) {
-    uintptr_t bytes_before_alignment_boundary =
-        kCacheLineSize -
-        (reinterpret_cast<uintptr_t>(d) & (kCacheLineSize - 1));
-    size_t header_len = (std::min)(bytes_before_alignment_boundary, len);
-    assert(bytes_before_alignment_boundary < kCacheLineSize);
-    memcpy(d, s, header_len);
-    d += header_len;
-    s += header_len;
-    len -= header_len;
-  }
-
-  if (len >= kCacheLineSize) {
-    _mm_sfence();
-    __m256i *dst_cacheline = reinterpret_cast<__m256i *>(d);
-    const __m256i *src_cacheline = reinterpret_cast<const __m256i *>(s);
-    constexpr int kOpsPerCacheLine = kCacheLineSize / sizeof(__m256i);
-    size_t loops = len / kCacheLineSize;
-
-    while (len >= kCacheLineSize) {
-      __m256i temp1, temp2;
-      temp1 = _mm256_lddqu_si256(src_cacheline + 0);
-      temp2 = _mm256_lddqu_si256(src_cacheline + 1);
-      _mm256_stream_si256(dst_cacheline + 0, temp1);
-      _mm256_stream_si256(dst_cacheline + 1, temp2);
-      src_cacheline += kOpsPerCacheLine;
-      dst_cacheline += kOpsPerCacheLine;
-      len -= kCacheLineSize;
-    }
-    d += loops * kCacheLineSize;
-    s += loops * kCacheLineSize;
-    _mm_sfence();
-  }
-
-  // memcpy the tail.
-  if (len) {
-    memcpy(d, s, len);
-  }
-  return dst;
-#else
-  // Fallback to regular memcpy when AVX is not available.
-  return memcpy(dst, src, len);
-#endif  // __AVX__
-}
-
-}  // namespace crc_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_CRC_INTERNAL_NON_TEMPORAL_MEMCPY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/debugging/failure_signal_handler.h b/projects/llm_framework/include/abseil-cpp/absl/debugging/failure_signal_handler.h
deleted file mode 100644
index 4117facb..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/debugging/failure_signal_handler.h
+++ /dev/null
@@ -1,121 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: failure_signal_handler.h
-// -----------------------------------------------------------------------------
-//
-// This file configures the Abseil *failure signal handler* to capture and dump
-// useful debugging information (such as a stacktrace) upon program failure.
-//
-// To use the failure signal handler, call `absl::InstallFailureSignalHandler()`
-// very early in your program, usually in the first few lines of main():
-//
-// int main(int argc, char** argv) {
-//   // Initialize the symbolizer to get a human-readable stack trace
-//   absl::InitializeSymbolizer(argv[0]);
-//
-//   absl::FailureSignalHandlerOptions options;
-//   absl::InstallFailureSignalHandler(options);
-//   DoSomethingInteresting();
-//   return 0;
-// }
-//
-// Any program that raises a fatal signal (such as `SIGSEGV`, `SIGILL`,
-// `SIGFPE`, `SIGABRT`, `SIGTERM`, `SIGBUS`, and `SIGTRAP`) will call the
-// installed failure signal handler and provide debugging information to stderr.
-//
-// Note that you should *not* install the Abseil failure signal handler more
-// than once. You may, of course, have another (non-Abseil) failure signal
-// handler installed (which would be triggered if Abseil's failure signal
-// handler sets `call_previous_handler` to `true`).
-
-#ifndef ABSL_DEBUGGING_FAILURE_SIGNAL_HANDLER_H_
-#define ABSL_DEBUGGING_FAILURE_SIGNAL_HANDLER_H_
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// FailureSignalHandlerOptions
-//
-// Struct for holding `absl::InstallFailureSignalHandler()` configuration
-// options.
-struct FailureSignalHandlerOptions {
-  // If true, try to symbolize the stacktrace emitted on failure, provided that
-  // you have initialized a symbolizer for that purpose. (See symbolize.h for
-  // more information.)
-  bool symbolize_stacktrace = true;
-
-  // If true, try to run signal handlers on an alternate stack (if supported on
-  // the given platform). An alternate stack is useful for program crashes due
-  // to a stack overflow; by running on a alternate stack, the signal handler
-  // may run even when normal stack space has been exhausted. The downside of
-  // using an alternate stack is that extra memory for the alternate stack needs
-  // to be pre-allocated.
-  bool use_alternate_stack = true;
-
-  // If positive, indicates the number of seconds after which the failure signal
-  // handler is invoked to abort the program. Setting such an alarm is useful in
-  // cases where the failure signal handler itself may become hung or
-  // deadlocked.
-  int alarm_on_failure_secs = 3;
-
-  // If true, call the previously registered signal handler for the signal that
-  // was received (if one was registered) after the existing signal handler
-  // runs. This mechanism can be used to chain signal handlers together.
-  //
-  // If false, the signal is raised to the default handler for that signal
-  // (which normally terminates the program).
-  //
-  // IMPORTANT: If true, the chained fatal signal handlers must not try to
-  // recover from the fatal signal. Instead, they should terminate the program
-  // via some mechanism, like raising the default handler for the signal, or by
-  // calling `_exit()`. Note that the failure signal handler may put parts of
-  // the Abseil library into a state from which they cannot recover.
-  bool call_previous_handler = false;
-
-  // If non-null, indicates a pointer to a callback function that will be called
-  // upon failure, with a string argument containing failure data. This function
-  // may be used as a hook to write failure data to a secondary location, such
-  // as a log file. This function will also be called with null data, as a hint
-  // to flush any buffered data before the program may be terminated. Consider
-  // flushing any buffered data in all calls to this function.
-  //
-  // Since this function runs within a signal handler, it should be
-  // async-signal-safe if possible.
-  // See http://man7.org/linux/man-pages/man7/signal-safety.7.html
-  void (*writerfn)(const char*) = nullptr;
-};
-
-// InstallFailureSignalHandler()
-//
-// Installs a signal handler for the common failure signals `SIGSEGV`, `SIGILL`,
-// `SIGFPE`, `SIGABRT`, `SIGTERM`, `SIGBUG`, and `SIGTRAP` (provided they exist
-// on the given platform). The failure signal handler dumps program failure data
-// useful for debugging in an unspecified format to stderr. This data may
-// include the program counter, a stacktrace, and register information on some
-// systems; do not rely on an exact format for the output, as it is subject to
-// change.
-void InstallFailureSignalHandler(const FailureSignalHandlerOptions& options);
-
-namespace debugging_internal {
-const char* FailureSignalToString(int signo);
-}  // namespace debugging_internal
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_DEBUGGING_FAILURE_SIGNAL_HANDLER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/address_is_readable.h b/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/address_is_readable.h
deleted file mode 100644
index 4bbaf4d6..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/address_is_readable.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_DEBUGGING_INTERNAL_ADDRESS_IS_READABLE_H_
-#define ABSL_DEBUGGING_INTERNAL_ADDRESS_IS_READABLE_H_
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace debugging_internal {
-
-// Return whether the byte at *addr is readable, without faulting.
-// Save and restores errno.
-bool AddressIsReadable(const void *addr);
-
-}  // namespace debugging_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_DEBUGGING_INTERNAL_ADDRESS_IS_READABLE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/demangle.h b/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/demangle.h
deleted file mode 100644
index 146d1150..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/demangle.h
+++ /dev/null
@@ -1,73 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_DEBUGGING_INTERNAL_DEMANGLE_H_
-#define ABSL_DEBUGGING_INTERNAL_DEMANGLE_H_
-
-#include <string>
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace debugging_internal {
-
-// Demangle `mangled`.  On success, return true and write the
-// demangled symbol name to `out`.  Otherwise, return false.
-// `out` is modified even if demangling is unsuccessful.
-//
-// This function provides an alternative to libstdc++'s abi::__cxa_demangle,
-// which is not async signal safe (it uses malloc internally).  It's intended to
-// be used in async signal handlers to symbolize stack traces.
-//
-// Note that this demangler doesn't support full demangling.  More
-// specifically, it doesn't print types of function parameters and
-// types of template arguments.  It just skips them.  However, it's
-// still very useful to extract basic information such as class,
-// function, constructor, destructor, and operator names.
-//
-// See the implementation note in demangle.cc if you are interested.
-//
-// Example:
-//
-// | Mangled Name  | Demangle    | DemangleString
-// |---------------|-------------|-----------------------
-// | _Z1fv         | f()         | f()
-// | _Z1fi         | f()         | f(int)
-// | _Z3foo3bar    | foo()       | foo(bar)
-// | _Z1fIiEvi     | f<>()       | void f<int>(int)
-// | _ZN1N1fE      | N::f        | N::f
-// | _ZN3Foo3BarEv | Foo::Bar()  | Foo::Bar()
-// | _Zrm1XS_"     | operator%() | operator%(X, X)
-// | _ZN3FooC1Ev   | Foo::Foo()  | Foo::Foo()
-// | _Z1fSs        | f()         | f(std::basic_string<char,
-// |               |             |   std::char_traits<char>,
-// |               |             |   std::allocator<char> >)
-//
-// See the unit test for more examples.
-//
-// Note: we might want to write demanglers for ABIs other than Itanium
-// C++ ABI in the future.
-bool Demangle(const char* mangled, char* out, size_t out_size);
-
-// A wrapper around `abi::__cxa_demangle()`.  On success, returns the demangled
-// name.  On failure, returns the input mangled name.
-//
-// This function is not async-signal-safe.
-std::string DemangleString(const char* mangled);
-
-}  // namespace debugging_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_DEBUGGING_INTERNAL_DEMANGLE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/elf_mem_image.h b/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/elf_mem_image.h
deleted file mode 100644
index e7fe6ab0..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/elf_mem_image.h
+++ /dev/null
@@ -1,140 +0,0 @@
-/*
- * Copyright 2017 The Abseil Authors.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      https://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-// Allow dynamic symbol lookup for in-memory Elf images.
-
-#ifndef ABSL_DEBUGGING_INTERNAL_ELF_MEM_IMAGE_H_
-#define ABSL_DEBUGGING_INTERNAL_ELF_MEM_IMAGE_H_
-
-// Including this will define the __GLIBC__ macro if glibc is being
-// used.
-#include <climits>
-
-#include "absl/base/config.h"
-
-// Maybe one day we can rewrite this file not to require the elf
-// symbol extensions in glibc, but for right now we need them.
-#ifdef ABSL_HAVE_ELF_MEM_IMAGE
-#error ABSL_HAVE_ELF_MEM_IMAGE cannot be directly set
-#endif
-
-#if defined(__ELF__) && !defined(__OpenBSD__) && !defined(__QNX__) && \
-    !defined(__native_client__) && !defined(__asmjs__) &&             \
-    !defined(__wasm__) && !defined(__HAIKU__) && !defined(__sun) &&   \
-    !defined(__VXWORKS__) && !defined(__hexagon__)
-#define ABSL_HAVE_ELF_MEM_IMAGE 1
-#endif
-
-#ifdef ABSL_HAVE_ELF_MEM_IMAGE
-
-#include <link.h>  // for ElfW
-
-#if defined(__FreeBSD__) && !defined(ElfW)
-#define ElfW(x) __ElfN(x)
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace debugging_internal {
-
-// An in-memory ELF image (may not exist on disk).
-class ElfMemImage {
- private:
-  // Sentinel: there could never be an elf image at &kInvalidBaseSentinel.
-  static const int kInvalidBaseSentinel;
-
- public:
-  // Sentinel: there could never be an elf image at this address.
-  static constexpr const void *const kInvalidBase =
-    static_cast<const void*>(&kInvalidBaseSentinel);
-
-  // Information about a single vdso symbol.
-  // All pointers are into .dynsym, .dynstr, or .text of the VDSO.
-  // Do not free() them or modify through them.
-  struct SymbolInfo {
-    const char      *name;      // E.g. "__vdso_getcpu"
-    const char      *version;   // E.g. "LINUX_2.6", could be ""
-                                // for unversioned symbol.
-    const void      *address;   // Relocated symbol address.
-    const ElfW(Sym) *symbol;    // Symbol in the dynamic symbol table.
-  };
-
-  // Supports iteration over all dynamic symbols.
-  class SymbolIterator {
-   public:
-    friend class ElfMemImage;
-    const SymbolInfo *operator->() const;
-    const SymbolInfo &operator*() const;
-    SymbolIterator& operator++();
-    bool operator!=(const SymbolIterator &rhs) const;
-    bool operator==(const SymbolIterator &rhs) const;
-   private:
-    SymbolIterator(const void *const image, int index);
-    void Update(int incr);
-    SymbolInfo info_;
-    int index_;
-    const void *const image_;
-  };
-
-
-  explicit ElfMemImage(const void *base);
-  void                 Init(const void *base);
-  bool                 IsPresent() const { return ehdr_ != nullptr; }
-  const ElfW(Phdr)*    GetPhdr(int index) const;
-  const ElfW(Sym)*     GetDynsym(int index) const;
-  const ElfW(Versym)*  GetVersym(int index) const;
-  const ElfW(Verdef)*  GetVerdef(int index) const;
-  const ElfW(Verdaux)* GetVerdefAux(const ElfW(Verdef) *verdef) const;
-  const char*          GetDynstr(ElfW(Word) offset) const;
-  const void*          GetSymAddr(const ElfW(Sym) *sym) const;
-  const char*          GetVerstr(ElfW(Word) offset) const;
-  int                  GetNumSymbols() const;
-
-  SymbolIterator begin() const;
-  SymbolIterator end() const;
-
-  // Look up versioned dynamic symbol in the image.
-  // Returns false if image is not present, or doesn't contain given
-  // symbol/version/type combination.
-  // If info_out is non-null, additional details are filled in.
-  bool LookupSymbol(const char *name, const char *version,
-                    int symbol_type, SymbolInfo *info_out) const;
-
-  // Find info about symbol (if any) which overlaps given address.
-  // Returns true if symbol was found; false if image isn't present
-  // or doesn't have a symbol overlapping given address.
-  // If info_out is non-null, additional details are filled in.
-  bool LookupSymbolByAddress(const void *address, SymbolInfo *info_out) const;
-
- private:
-  const ElfW(Ehdr) *ehdr_;
-  const ElfW(Sym) *dynsym_;
-  const ElfW(Versym) *versym_;
-  const ElfW(Verdef) *verdef_;
-  const ElfW(Word) *hash_;
-  const char *dynstr_;
-  size_t strsize_;
-  size_t verdefnum_;
-  ElfW(Addr) link_base_;     // Link-time base (p_vaddr of first PT_LOAD).
-};
-
-}  // namespace debugging_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_HAVE_ELF_MEM_IMAGE
-
-#endif  // ABSL_DEBUGGING_INTERNAL_ELF_MEM_IMAGE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/examine_stack.h b/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/examine_stack.h
deleted file mode 100644
index 190af87f..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/examine_stack.h
+++ /dev/null
@@ -1,64 +0,0 @@
-//
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_DEBUGGING_INTERNAL_EXAMINE_STACK_H_
-#define ABSL_DEBUGGING_INTERNAL_EXAMINE_STACK_H_
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace debugging_internal {
-
-// Type of function used for printing in stack trace dumping, etc.
-// We avoid closures to keep things simple.
-typedef void OutputWriter(const char*, void*);
-
-// RegisterDebugStackTraceHook() allows to register a single routine
-// `hook` that is called each time DumpStackTrace() is called.
-// `hook` may be called from a signal handler.
-typedef void (*SymbolizeUrlEmitter)(void* const stack[], int depth,
-                                    OutputWriter* writer, void* writer_arg);
-
-// Registration of SymbolizeUrlEmitter for use inside of a signal handler.
-// This is inherently unsafe and must be signal safe code.
-void RegisterDebugStackTraceHook(SymbolizeUrlEmitter hook);
-SymbolizeUrlEmitter GetDebugStackTraceHook();
-
-// Returns the program counter from signal context, or nullptr if
-// unknown. `vuc` is a ucontext_t*. We use void* to avoid the use of
-// ucontext_t on non-POSIX systems.
-void* GetProgramCounter(void* const vuc);
-
-// Uses `writer` to dump the program counter, stack trace, and stack
-// frame sizes.
-void DumpPCAndFrameSizesAndStackTrace(void* const pc, void* const stack[],
-                                      int frame_sizes[], int depth,
-                                      int min_dropped_frames,
-                                      bool symbolize_stacktrace,
-                                      OutputWriter* writer, void* writer_arg);
-
-// Dump current stack trace omitting the topmost `min_dropped_frames` stack
-// frames.
-void DumpStackTrace(int min_dropped_frames, int max_num_frames,
-                    bool symbolize_stacktrace, OutputWriter* writer,
-                    void* writer_arg);
-
-}  // namespace debugging_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_DEBUGGING_INTERNAL_EXAMINE_STACK_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/stack_consumption.h b/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/stack_consumption.h
deleted file mode 100644
index f41b64c3..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/stack_consumption.h
+++ /dev/null
@@ -1,50 +0,0 @@
-//
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-// Helper function for measuring stack consumption of signal handlers.
-
-#ifndef ABSL_DEBUGGING_INTERNAL_STACK_CONSUMPTION_H_
-#define ABSL_DEBUGGING_INTERNAL_STACK_CONSUMPTION_H_
-
-#include "absl/base/config.h"
-
-// The code in this module is not portable.
-// Use this feature test macro to detect its availability.
-#ifdef ABSL_INTERNAL_HAVE_DEBUGGING_STACK_CONSUMPTION
-#error ABSL_INTERNAL_HAVE_DEBUGGING_STACK_CONSUMPTION cannot be set directly
-#elif !defined(__APPLE__) && !defined(_WIN32) &&                     \
-    (defined(__i386__) || defined(__x86_64__) || defined(__ppc__) || \
-     defined(__aarch64__) || defined(__riscv))
-#define ABSL_INTERNAL_HAVE_DEBUGGING_STACK_CONSUMPTION 1
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace debugging_internal {
-
-// Returns the stack consumption in bytes for the code exercised by
-// signal_handler.  To measure stack consumption, signal_handler is registered
-// as a signal handler, so the code that it exercises must be async-signal
-// safe.  The argument of signal_handler is an implementation detail of signal
-// handlers and should ignored by the code for signal_handler.  Use global
-// variables to pass information between your test code and signal_handler.
-int GetSignalHandlerStackConsumption(void (*signal_handler)(int));
-
-}  // namespace debugging_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_INTERNAL_HAVE_DEBUGGING_STACK_CONSUMPTION
-
-#endif  // ABSL_DEBUGGING_INTERNAL_STACK_CONSUMPTION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/stacktrace_config.h b/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/stacktrace_config.h
deleted file mode 100644
index 3929b1b7..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/stacktrace_config.h
+++ /dev/null
@@ -1,88 +0,0 @@
-/*
- * Copyright 2017 The Abseil Authors.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      https://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
-
- * Defines ABSL_STACKTRACE_INL_HEADER to the *-inl.h containing
- * actual unwinder implementation.
- * This header is "private" to stacktrace.cc.
- * DO NOT include it into any other files.
-*/
-#ifndef ABSL_DEBUGGING_INTERNAL_STACKTRACE_CONFIG_H_
-#define ABSL_DEBUGGING_INTERNAL_STACKTRACE_CONFIG_H_
-
-#include "absl/base/config.h"
-
-#if defined(ABSL_STACKTRACE_INL_HEADER)
-#error ABSL_STACKTRACE_INL_HEADER cannot be directly set
-
-#elif defined(_WIN32)
-#define ABSL_STACKTRACE_INL_HEADER \
-    "absl/debugging/internal/stacktrace_win32-inl.inc"
-
-#elif defined(__APPLE__)
-#ifdef ABSL_HAVE_THREAD_LOCAL
-// Thread local support required for UnwindImpl.
-#define ABSL_STACKTRACE_INL_HEADER \
-  "absl/debugging/internal/stacktrace_generic-inl.inc"
-#endif  // defined(ABSL_HAVE_THREAD_LOCAL)
-
-// Emscripten stacktraces rely on JS. Do not use them in standalone mode.
-#elif defined(__EMSCRIPTEN__) && !defined(STANDALONE_WASM)
-#define ABSL_STACKTRACE_INL_HEADER \
-  "absl/debugging/internal/stacktrace_emscripten-inl.inc"
-
-#elif defined(__linux__) && !defined(__ANDROID__)
-
-#if defined(NO_FRAME_POINTER) && \
-    (defined(__i386__) || defined(__x86_64__) || defined(__aarch64__))
-// Note: The libunwind-based implementation is not available to open-source
-// users.
-#define ABSL_STACKTRACE_INL_HEADER \
-  "absl/debugging/internal/stacktrace_libunwind-inl.inc"
-#define STACKTRACE_USES_LIBUNWIND 1
-#elif defined(NO_FRAME_POINTER) && defined(__has_include)
-#if __has_include(<execinfo.h>)
-// Note: When using glibc this may require -funwind-tables to function properly.
-#define ABSL_STACKTRACE_INL_HEADER \
-  "absl/debugging/internal/stacktrace_generic-inl.inc"
-#endif  // __has_include(<execinfo.h>)
-#elif defined(__i386__) || defined(__x86_64__)
-#define ABSL_STACKTRACE_INL_HEADER \
-  "absl/debugging/internal/stacktrace_x86-inl.inc"
-#elif defined(__ppc__) || defined(__PPC__)
-#define ABSL_STACKTRACE_INL_HEADER \
-  "absl/debugging/internal/stacktrace_powerpc-inl.inc"
-#elif defined(__aarch64__)
-#define ABSL_STACKTRACE_INL_HEADER \
-  "absl/debugging/internal/stacktrace_aarch64-inl.inc"
-#elif defined(__riscv)
-#define ABSL_STACKTRACE_INL_HEADER \
-  "absl/debugging/internal/stacktrace_riscv-inl.inc"
-#elif defined(__has_include)
-#if __has_include(<execinfo.h>)
-// Note: When using glibc this may require -funwind-tables to function properly.
-#define ABSL_STACKTRACE_INL_HEADER \
-  "absl/debugging/internal/stacktrace_generic-inl.inc"
-#endif  // __has_include(<execinfo.h>)
-#endif  // defined(__has_include)
-
-#endif  // defined(__linux__) && !defined(__ANDROID__)
-
-// Fallback to the empty implementation.
-#if !defined(ABSL_STACKTRACE_INL_HEADER)
-#define ABSL_STACKTRACE_INL_HEADER \
-  "absl/debugging/internal/stacktrace_unimplemented-inl.inc"
-#endif
-
-#endif  // ABSL_DEBUGGING_INTERNAL_STACKTRACE_CONFIG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/symbolize.h b/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/symbolize.h
deleted file mode 100644
index 5593fde6..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/symbolize.h
+++ /dev/null
@@ -1,153 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-// This file contains internal parts of the Abseil symbolizer.
-// Do not depend on the anything in this file, it may change at anytime.
-
-#ifndef ABSL_DEBUGGING_INTERNAL_SYMBOLIZE_H_
-#define ABSL_DEBUGGING_INTERNAL_SYMBOLIZE_H_
-
-#ifdef __cplusplus
-
-#include <cstddef>
-#include <cstdint>
-
-#include "absl/base/config.h"
-#include "absl/strings/string_view.h"
-
-#ifdef ABSL_INTERNAL_HAVE_ELF_SYMBOLIZE
-#error ABSL_INTERNAL_HAVE_ELF_SYMBOLIZE cannot be directly set
-#elif defined(__ELF__) && defined(__GLIBC__) && !defined(__native_client__) \
-      && !defined(__asmjs__) && !defined(__wasm__)
-#define ABSL_INTERNAL_HAVE_ELF_SYMBOLIZE 1
-
-#include <elf.h>
-#include <link.h>  // For ElfW() macro.
-#include <functional>
-#include <string>
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace debugging_internal {
-
-// Iterates over all sections, invoking callback on each with the section name
-// and the section header.
-//
-// Returns true on success; otherwise returns false in case of errors.
-//
-// This is not async-signal-safe.
-bool ForEachSection(int fd,
-                    const std::function<bool(absl::string_view name,
-                                             const ElfW(Shdr) &)>& callback);
-
-// Gets the section header for the given name, if it exists. Returns true on
-// success. Otherwise, returns false.
-bool GetSectionHeaderByName(int fd, const char *name, size_t name_len,
-                            ElfW(Shdr) *out);
-
-}  // namespace debugging_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_INTERNAL_HAVE_ELF_SYMBOLIZE
-
-#ifdef ABSL_INTERNAL_HAVE_DARWIN_SYMBOLIZE
-#error ABSL_INTERNAL_HAVE_DARWIN_SYMBOLIZE cannot be directly set
-#elif defined(__APPLE__)
-#define ABSL_INTERNAL_HAVE_DARWIN_SYMBOLIZE 1
-#endif
-
-#ifdef ABSL_INTERNAL_HAVE_EMSCRIPTEN_SYMBOLIZE
-#error ABSL_INTERNAL_HAVE_EMSCRIPTEN_SYMBOLIZE cannot be directly set
-#elif defined(__EMSCRIPTEN__)
-#define ABSL_INTERNAL_HAVE_EMSCRIPTEN_SYMBOLIZE 1
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace debugging_internal {
-
-struct SymbolDecoratorArgs {
-  // The program counter we are getting symbolic name for.
-  const void *pc;
-  // 0 for main executable, load address for shared libraries.
-  ptrdiff_t relocation;
-  // Read-only file descriptor for ELF image covering "pc",
-  // or -1 if no such ELF image exists in /proc/self/maps.
-  int fd;
-  // Output buffer, size.
-  // Note: the buffer may not be empty -- default symbolizer may have already
-  // produced some output, and earlier decorators may have adorned it in
-  // some way. You are free to replace or augment the contents (within the
-  // symbol_buf_size limit).
-  char *const symbol_buf;
-  size_t symbol_buf_size;
-  // Temporary scratch space, size.
-  // Use that space in preference to allocating your own stack buffer to
-  // conserve stack.
-  char *const tmp_buf;
-  size_t tmp_buf_size;
-  // User-provided argument
-  void* arg;
-};
-using SymbolDecorator = void (*)(const SymbolDecoratorArgs *);
-
-// Installs a function-pointer as a decorator. Returns a value less than zero
-// if the system cannot install the decorator. Otherwise, returns a unique
-// identifier corresponding to the decorator. This identifier can be used to
-// uninstall the decorator - See RemoveSymbolDecorator() below.
-int InstallSymbolDecorator(SymbolDecorator decorator, void* arg);
-
-// Removes a previously installed function-pointer decorator. Parameter "ticket"
-// is the return-value from calling InstallSymbolDecorator().
-bool RemoveSymbolDecorator(int ticket);
-
-// Remove all installed decorators.  Returns true if successful, false if
-// symbolization is currently in progress.
-bool RemoveAllSymbolDecorators();
-
-// Registers an address range to a file mapping.
-//
-// Preconditions:
-//   start <= end
-//   filename != nullptr
-//
-// Returns true if the file was successfully registered.
-bool RegisterFileMappingHint(const void* start, const void* end,
-                             uint64_t offset, const char* filename);
-
-// Looks up the file mapping registered by RegisterFileMappingHint for an
-// address range. If there is one, the file name is stored in *filename and
-// *start and *end are modified to reflect the registered mapping. Returns
-// whether any hint was found.
-bool GetFileMappingHint(const void** start, const void** end, uint64_t* offset,
-                        const char** filename);
-
-}  // namespace debugging_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // __cplusplus
-
-#include <stdbool.h>
-
-#ifdef __cplusplus
-extern "C"
-#endif  // __cplusplus
-
-    bool
-    AbslInternalGetFileMappingHint(const void** start, const void** end,
-                                   uint64_t* offset, const char** filename);
-
-#endif  // ABSL_DEBUGGING_INTERNAL_SYMBOLIZE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/vdso_support.h b/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/vdso_support.h
deleted file mode 100644
index 6562c6c2..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/debugging/internal/vdso_support.h
+++ /dev/null
@@ -1,158 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-// Allow dynamic symbol lookup in the kernel VDSO page.
-//
-// VDSO stands for "Virtual Dynamic Shared Object" -- a page of
-// executable code, which looks like a shared library, but doesn't
-// necessarily exist anywhere on disk, and which gets mmap()ed into
-// every process by kernels which support VDSO, such as 2.6.x for 32-bit
-// executables, and 2.6.24 and above for 64-bit executables.
-//
-// More details could be found here:
-// http://www.trilithium.com/johan/2005/08/linux-gate/
-//
-// VDSOSupport -- a class representing kernel VDSO (if present).
-//
-// Example usage:
-//  VDSOSupport vdso;
-//  VDSOSupport::SymbolInfo info;
-//  typedef (*FN)(unsigned *, void *, void *);
-//  FN fn = nullptr;
-//  if (vdso.LookupSymbol("__vdso_getcpu", "LINUX_2.6", STT_FUNC, &info)) {
-//     fn = reinterpret_cast<FN>(info.address);
-//  }
-
-#ifndef ABSL_DEBUGGING_INTERNAL_VDSO_SUPPORT_H_
-#define ABSL_DEBUGGING_INTERNAL_VDSO_SUPPORT_H_
-
-#include <atomic>
-
-#include "absl/base/attributes.h"
-#include "absl/debugging/internal/elf_mem_image.h"
-
-#ifdef ABSL_HAVE_ELF_MEM_IMAGE
-
-#ifdef ABSL_HAVE_VDSO_SUPPORT
-#error ABSL_HAVE_VDSO_SUPPORT cannot be directly set
-#else
-#define ABSL_HAVE_VDSO_SUPPORT 1
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace debugging_internal {
-
-// NOTE: this class may be used from within tcmalloc, and can not
-// use any memory allocation routines.
-class VDSOSupport {
- public:
-  VDSOSupport();
-
-  typedef ElfMemImage::SymbolInfo SymbolInfo;
-  typedef ElfMemImage::SymbolIterator SymbolIterator;
-
-  // On PowerPC64 VDSO symbols can either be of type STT_FUNC or STT_NOTYPE
-  // depending on how the kernel is built.  The kernel is normally built with
-  // STT_NOTYPE type VDSO symbols.  Let's make things simpler first by using a
-  // compile-time constant.
-#ifdef __powerpc64__
-  enum { kVDSOSymbolType = STT_NOTYPE };
-#else
-  enum { kVDSOSymbolType = STT_FUNC };
-#endif
-
-  // Answers whether we have a vdso at all.
-  bool IsPresent() const { return image_.IsPresent(); }
-
-  // Allow to iterate over all VDSO symbols.
-  SymbolIterator begin() const { return image_.begin(); }
-  SymbolIterator end() const { return image_.end(); }
-
-  // Look up versioned dynamic symbol in the kernel VDSO.
-  // Returns false if VDSO is not present, or doesn't contain given
-  // symbol/version/type combination.
-  // If info_out != nullptr, additional details are filled in.
-  bool LookupSymbol(const char *name, const char *version,
-                    int symbol_type, SymbolInfo *info_out) const;
-
-  // Find info about symbol (if any) which overlaps given address.
-  // Returns true if symbol was found; false if VDSO isn't present
-  // or doesn't have a symbol overlapping given address.
-  // If info_out != nullptr, additional details are filled in.
-  bool LookupSymbolByAddress(const void *address, SymbolInfo *info_out) const;
-
-  // Used only for testing. Replace real VDSO base with a mock.
-  // Returns previous value of vdso_base_. After you are done testing,
-  // you are expected to call SetBase() with previous value, in order to
-  // reset state to the way it was.
-  const void *SetBase(const void *s);
-
-  // Computes vdso_base_ and returns it. Should be called as early as
-  // possible; before any thread creation, chroot or setuid.
-  static const void *Init();
-
- private:
-  // image_ represents VDSO ELF image in memory.
-  // image_.ehdr_ == nullptr implies there is no VDSO.
-  ElfMemImage image_;
-
-  // Cached value of auxv AT_SYSINFO_EHDR, computed once.
-  // This is a tri-state:
-  //   kInvalidBase   => value hasn't been determined yet.
-  //              0   => there is no VDSO.
-  //           else   => vma of VDSO Elf{32,64}_Ehdr.
-  //
-  // When testing with mock VDSO, low bit is set.
-  // The low bit is always available because vdso_base_ is
-  // page-aligned.
-  static std::atomic<const void *> vdso_base_;
-
-  // NOLINT on 'long' because these routines mimic kernel api.
-  // The 'cache' parameter may be used by some versions of the kernel,
-  // and should be nullptr or point to a static buffer containing at
-  // least two 'long's.
-  static long InitAndGetCPU(unsigned *cpu, void *cache,     // NOLINT 'long'.
-                            void *unused);
-  static long GetCPUViaSyscall(unsigned *cpu, void *cache,  // NOLINT 'long'.
-                               void *unused);
-  typedef long (*GetCpuFn)(unsigned *cpu, void *cache,      // NOLINT 'long'.
-                           void *unused);
-
-  // This function pointer may point to InitAndGetCPU,
-  // GetCPUViaSyscall, or __vdso_getcpu at different stages of initialization.
-  ABSL_CONST_INIT static std::atomic<GetCpuFn> getcpu_fn_;
-
-  friend int GetCPU(void);  // Needs access to getcpu_fn_.
-
-  VDSOSupport(const VDSOSupport&) = delete;
-  VDSOSupport& operator=(const VDSOSupport&) = delete;
-};
-
-// Same as sched_getcpu() on later glibc versions.
-// Return current CPU, using (fast) __vdso_getcpu@LINUX_2.6 if present,
-// otherwise use syscall(SYS_getcpu,...).
-// May return -1 with errno == ENOSYS if the kernel doesn't
-// support SYS_getcpu.
-int GetCPU();
-
-}  // namespace debugging_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_HAVE_ELF_MEM_IMAGE
-
-#endif  // ABSL_DEBUGGING_INTERNAL_VDSO_SUPPORT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/debugging/leak_check.h b/projects/llm_framework/include/abseil-cpp/absl/debugging/leak_check.h
deleted file mode 100644
index 6bd79406..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/debugging/leak_check.h
+++ /dev/null
@@ -1,150 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: leak_check.h
-// -----------------------------------------------------------------------------
-//
-// This file contains functions that affect leak checking behavior within
-// targets built with the LeakSanitizer (LSan), a memory leak detector that is
-// integrated within the AddressSanitizer (ASan) as an additional component, or
-// which can be used standalone. LSan and ASan are included (or can be provided)
-// as additional components for most compilers such as Clang, gcc and MSVC.
-// Note: this leak checking API is not yet supported in MSVC.
-// Leak checking is enabled by default in all ASan builds.
-//
-// https://clang.llvm.org/docs/LeakSanitizer.html
-// https://github.com/google/sanitizers/wiki/AddressSanitizerLeakSanitizer
-//
-// GCC and Clang both automatically enable LeakSanitizer when AddressSanitizer
-// is enabled. To use the mode, simply pass `-fsanitize=address` to both the
-// compiler and linker. An example Bazel command could be
-//
-//   $ bazel test --copt=-fsanitize=address --linkopt=-fsanitize=address ...
-//
-// GCC and Clang auto support a standalone LeakSanitizer mode (a mode which does
-// not also use AddressSanitizer). To use the mode, simply pass
-// `-fsanitize=leak` to both the compiler and linker. Since GCC does not
-// currently provide a way of detecting this mode at compile-time, GCC users
-// must also pass -DLEAK_SANITIZER to the compiler. An example Bazel command
-// could be
-//
-//   $ bazel test --copt=-DLEAK_SANITIZER --copt=-fsanitize=leak
-//     --linkopt=-fsanitize=leak ...
-//
-// -----------------------------------------------------------------------------
-#ifndef ABSL_DEBUGGING_LEAK_CHECK_H_
-#define ABSL_DEBUGGING_LEAK_CHECK_H_
-
-#include <cstddef>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// HaveLeakSanitizer()
-//
-// Returns true if a leak-checking sanitizer (either ASan or standalone LSan) is
-// currently built into this target.
-bool HaveLeakSanitizer();
-
-// LeakCheckerIsActive()
-//
-// Returns true if a leak-checking sanitizer (either ASan or standalone LSan) is
-// currently built into this target and is turned on.
-bool LeakCheckerIsActive();
-
-// DoIgnoreLeak()
-//
-// Implements `IgnoreLeak()` below. This function should usually
-// not be called directly; calling `IgnoreLeak()` is preferred.
-void DoIgnoreLeak(const void* ptr);
-
-// IgnoreLeak()
-//
-// Instruct the leak sanitizer to ignore leak warnings on the object referenced
-// by the passed pointer, as well as all heap objects transitively referenced
-// by it. The passed object pointer can point to either the beginning of the
-// object or anywhere within it.
-//
-// Example:
-//
-//   static T* obj = IgnoreLeak(new T(...));
-//
-// If the passed `ptr` does not point to an actively allocated object at the
-// time `IgnoreLeak()` is called, the call is a no-op; if it is actively
-// allocated, leak sanitizer will assume this object is referenced even if
-// there is no actual reference in user memory.
-//
-template <typename T>
-T* IgnoreLeak(T* ptr) {
-  DoIgnoreLeak(ptr);
-  return ptr;
-}
-
-// FindAndReportLeaks()
-//
-// If any leaks are detected, prints a leak report and returns true.  This
-// function may be called repeatedly, and does not affect end-of-process leak
-// checking.
-//
-// Example:
-// if (FindAndReportLeaks()) {
-//   ... diagnostic already printed. Exit with failure code.
-//   exit(1)
-// }
-bool FindAndReportLeaks();
-
-// LeakCheckDisabler
-//
-// This helper class indicates that any heap allocations done in the code block
-// covered by the scoped object, which should be allocated on the stack, will
-// not be reported as leaks. Leak check disabling will occur within the code
-// block and any nested function calls within the code block.
-//
-// Example:
-//
-//   void Foo() {
-//     LeakCheckDisabler disabler;
-//     ... code that allocates objects whose leaks should be ignored ...
-//   }
-//
-// REQUIRES: Destructor runs in same thread as constructor
-class LeakCheckDisabler {
- public:
-  LeakCheckDisabler();
-  LeakCheckDisabler(const LeakCheckDisabler&) = delete;
-  LeakCheckDisabler& operator=(const LeakCheckDisabler&) = delete;
-  ~LeakCheckDisabler();
-};
-
-// RegisterLivePointers()
-//
-// Registers `ptr[0,size-1]` as pointers to memory that is still actively being
-// referenced and for which leak checking should be ignored. This function is
-// useful if you store pointers in mapped memory, for memory ranges that we know
-// are correct but for which normal analysis would flag as leaked code.
-void RegisterLivePointers(const void* ptr, size_t size);
-
-// UnRegisterLivePointers()
-//
-// Deregisters the pointers previously marked as active in
-// `RegisterLivePointers()`, enabling leak checking of those pointers.
-void UnRegisterLivePointers(const void* ptr, size_t size);
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_DEBUGGING_LEAK_CHECK_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/debugging/stacktrace.h b/projects/llm_framework/include/abseil-cpp/absl/debugging/stacktrace.h
deleted file mode 100644
index 0ec0ffda..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/debugging/stacktrace.h
+++ /dev/null
@@ -1,231 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: stacktrace.h
-// -----------------------------------------------------------------------------
-//
-// This file contains routines to extract the current stack trace and associated
-// stack frames. These functions are thread-safe and async-signal-safe.
-//
-// Note that stack trace functionality is platform dependent and requires
-// additional support from the compiler/build system in most cases. (That is,
-// this functionality generally only works on platforms/builds that have been
-// specifically configured to support it.)
-//
-// Note: stack traces in Abseil that do not utilize a symbolizer will result in
-// frames consisting of function addresses rather than human-readable function
-// names. (See symbolize.h for information on symbolizing these values.)
-
-#ifndef ABSL_DEBUGGING_STACKTRACE_H_
-#define ABSL_DEBUGGING_STACKTRACE_H_
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// GetStackFrames()
-//
-// Records program counter values for up to `max_depth` frames, skipping the
-// most recent `skip_count` stack frames, stores their corresponding values
-// and sizes in `results` and `sizes` buffers, and returns the number of frames
-// stored. (Note that the frame generated for the `absl::GetStackFrames()`
-// routine itself is also skipped.)
-//
-// Example:
-//
-//      main() { foo(); }
-//      foo() { bar(); }
-//      bar() {
-//        void* result[10];
-//        int sizes[10];
-//        int depth = absl::GetStackFrames(result, sizes, 10, 1);
-//      }
-//
-// The current stack frame would consist of three function calls: `bar()`,
-// `foo()`, and then `main()`; however, since the `GetStackFrames()` call sets
-// `skip_count` to `1`, it will skip the frame for `bar()`, the most recently
-// invoked function call. It will therefore return 2 and fill `result` with
-// program counters within the following functions:
-//
-//      result[0]       foo()
-//      result[1]       main()
-//
-// (Note: in practice, a few more entries after `main()` may be added to account
-// for startup processes.)
-//
-// Corresponding stack frame sizes will also be recorded:
-//
-//    sizes[0]       16
-//    sizes[1]       16
-//
-// (Stack frame sizes of `16` above are just for illustration purposes.)
-//
-// Stack frame sizes of 0 or less indicate that those frame sizes couldn't
-// be identified.
-//
-// This routine may return fewer stack frame entries than are
-// available. Also note that `result` and `sizes` must both be non-null.
-extern int GetStackFrames(void** result, int* sizes, int max_depth,
-                          int skip_count);
-
-// GetStackFramesWithContext()
-//
-// Records program counter values obtained from a signal handler. Records
-// program counter values for up to `max_depth` frames, skipping the most recent
-// `skip_count` stack frames, stores their corresponding values and sizes in
-// `results` and `sizes` buffers, and returns the number of frames stored. (Note
-// that the frame generated for the `absl::GetStackFramesWithContext()` routine
-// itself is also skipped.)
-//
-// The `uc` parameter, if non-null, should be a pointer to a `ucontext_t` value
-// passed to a signal handler registered via the `sa_sigaction` field of a
-// `sigaction` struct. (See
-// http://man7.org/linux/man-pages/man2/sigaction.2.html.) The `uc` value may
-// help a stack unwinder to provide a better stack trace under certain
-// conditions. `uc` may safely be null.
-//
-// The `min_dropped_frames` output parameter, if non-null, points to the
-// location to note any dropped stack frames, if any, due to buffer limitations
-// or other reasons. (This value will be set to `0` if no frames were dropped.)
-// The number of total stack frames is guaranteed to be >= skip_count +
-// max_depth + *min_dropped_frames.
-extern int GetStackFramesWithContext(void** result, int* sizes, int max_depth,
-                                     int skip_count, const void* uc,
-                                     int* min_dropped_frames);
-
-// GetStackTrace()
-//
-// Records program counter values for up to `max_depth` frames, skipping the
-// most recent `skip_count` stack frames, stores their corresponding values
-// in `results`, and returns the number of frames
-// stored. Note that this function is similar to `absl::GetStackFrames()`
-// except that it returns the stack trace only, and not stack frame sizes.
-//
-// Example:
-//
-//      main() { foo(); }
-//      foo() { bar(); }
-//      bar() {
-//        void* result[10];
-//        int depth = absl::GetStackTrace(result, 10, 1);
-//      }
-//
-// This produces:
-//
-//      result[0]       foo
-//      result[1]       main
-//           ....       ...
-//
-// `result` must not be null.
-extern int GetStackTrace(void** result, int max_depth, int skip_count);
-
-// GetStackTraceWithContext()
-//
-// Records program counter values obtained from a signal handler. Records
-// program counter values for up to `max_depth` frames, skipping the most recent
-// `skip_count` stack frames, stores their corresponding values in `results`,
-// and returns the number of frames stored. (Note that the frame generated for
-// the `absl::GetStackFramesWithContext()` routine itself is also skipped.)
-//
-// The `uc` parameter, if non-null, should be a pointer to a `ucontext_t` value
-// passed to a signal handler registered via the `sa_sigaction` field of a
-// `sigaction` struct. (See
-// http://man7.org/linux/man-pages/man2/sigaction.2.html.) The `uc` value may
-// help a stack unwinder to provide a better stack trace under certain
-// conditions. `uc` may safely be null.
-//
-// The `min_dropped_frames` output parameter, if non-null, points to the
-// location to note any dropped stack frames, if any, due to buffer limitations
-// or other reasons. (This value will be set to `0` if no frames were dropped.)
-// The number of total stack frames is guaranteed to be >= skip_count +
-// max_depth + *min_dropped_frames.
-extern int GetStackTraceWithContext(void** result, int max_depth,
-                                    int skip_count, const void* uc,
-                                    int* min_dropped_frames);
-
-// SetStackUnwinder()
-//
-// Provides a custom function for unwinding stack frames that will be used in
-// place of the default stack unwinder when invoking the static
-// GetStack{Frames,Trace}{,WithContext}() functions above.
-//
-// The arguments passed to the unwinder function will match the
-// arguments passed to `absl::GetStackFramesWithContext()` except that sizes
-// will be non-null iff the caller is interested in frame sizes.
-//
-// If unwinder is set to null, we revert to the default stack-tracing behavior.
-//
-// *****************************************************************************
-// WARNING
-// *****************************************************************************
-//
-// absl::SetStackUnwinder is not suitable for general purpose use.  It is
-// provided for custom runtimes.
-// Some things to watch out for when calling `absl::SetStackUnwinder()`:
-//
-// (a) The unwinder may be called from within signal handlers and
-// therefore must be async-signal-safe.
-//
-// (b) Even after a custom stack unwinder has been unregistered, other
-// threads may still be in the process of using that unwinder.
-// Therefore do not clean up any state that may be needed by an old
-// unwinder.
-// *****************************************************************************
-extern void SetStackUnwinder(int (*unwinder)(void** pcs, int* sizes,
-                                             int max_depth, int skip_count,
-                                             const void* uc,
-                                             int* min_dropped_frames));
-
-// DefaultStackUnwinder()
-//
-// Records program counter values of up to `max_depth` frames, skipping the most
-// recent `skip_count` stack frames, and stores their corresponding values in
-// `pcs`. (Note that the frame generated for this call itself is also skipped.)
-// This function acts as a generic stack-unwinder; prefer usage of the more
-// specific `GetStack{Trace,Frames}{,WithContext}()` functions above.
-//
-// If you have set your own stack unwinder (with the `SetStackUnwinder()`
-// function above, you can still get the default stack unwinder by calling
-// `DefaultStackUnwinder()`, which will ignore any previously set stack unwinder
-// and use the default one instead.
-//
-// Because this function is generic, only `pcs` is guaranteed to be non-null
-// upon return. It is legal for `sizes`, `uc`, and `min_dropped_frames` to all
-// be null when called.
-//
-// The semantics are the same as the corresponding `GetStack*()` function in the
-// case where `absl::SetStackUnwinder()` was never called. Equivalents are:
-//
-//                       null sizes         |        non-nullptr sizes
-//             |==========================================================|
-//     null uc | GetStackTrace()            | GetStackFrames()            |
-// non-null uc | GetStackTraceWithContext() | GetStackFramesWithContext() |
-//             |==========================================================|
-extern int DefaultStackUnwinder(void** pcs, int* sizes, int max_depth,
-                                int skip_count, const void* uc,
-                                int* min_dropped_frames);
-
-namespace debugging_internal {
-// Returns true for platforms which are expected to have functioning stack trace
-// implementations. Intended to be used for tests which want to exclude
-// verification of logic known to be broken because stack traces are not
-// working.
-extern bool StackTraceWorksForTest();
-}  // namespace debugging_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_DEBUGGING_STACKTRACE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/debugging/symbolize.h b/projects/llm_framework/include/abseil-cpp/absl/debugging/symbolize.h
deleted file mode 100644
index 43d93a86..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/debugging/symbolize.h
+++ /dev/null
@@ -1,99 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: symbolize.h
-// -----------------------------------------------------------------------------
-//
-// This file configures the Abseil symbolizer for use in converting instruction
-// pointer addresses (program counters) into human-readable names (function
-// calls, etc.) within Abseil code.
-//
-// The symbolizer may be invoked from several sources:
-//
-//   * Implicitly, through the installation of an Abseil failure signal handler.
-//     (See failure_signal_handler.h for more information.)
-//   * By calling `Symbolize()` directly on a program counter you obtain through
-//     `absl::GetStackTrace()` or `absl::GetStackFrames()`. (See stacktrace.h
-//     for more information.
-//   * By calling `Symbolize()` directly on a program counter you obtain through
-//     other means (which would be platform-dependent).
-//
-// In all of the above cases, the symbolizer must first be initialized before
-// any program counter values can be symbolized. If you are installing a failure
-// signal handler, initialize the symbolizer before you do so.
-//
-// Example:
-//
-//   int main(int argc, char** argv) {
-//     // Initialize the Symbolizer before installing the failure signal handler
-//     absl::InitializeSymbolizer(argv[0]);
-//
-//     // Now you may install the failure signal handler
-//     absl::FailureSignalHandlerOptions options;
-//     absl::InstallFailureSignalHandler(options);
-//
-//     // Start running your main program
-//     ...
-//     return 0;
-//  }
-//
-#ifndef ABSL_DEBUGGING_SYMBOLIZE_H_
-#define ABSL_DEBUGGING_SYMBOLIZE_H_
-
-#include "absl/debugging/internal/symbolize.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// InitializeSymbolizer()
-//
-// Initializes the program counter symbolizer, given the path of the program
-// (typically obtained through `main()`s `argv[0]`). The Abseil symbolizer
-// allows you to read program counters (instruction pointer values) using their
-// human-readable names within output such as stack traces.
-//
-// Example:
-//
-// int main(int argc, char *argv[]) {
-//   absl::InitializeSymbolizer(argv[0]);
-//   // Now you can use the symbolizer
-// }
-void InitializeSymbolizer(const char* argv0);
-//
-// Symbolize()
-//
-// Symbolizes a program counter (instruction pointer value) `pc` and, on
-// success, writes the name to `out`. The symbol name is demangled, if possible.
-// Note that the symbolized name may be truncated and will be NUL-terminated.
-// Demangling is supported for symbols generated by GCC 3.x or newer). Returns
-// `false` on failure.
-//
-// Example:
-//
-//   // Print a program counter and its symbol name.
-//   static void DumpPCAndSymbol(void *pc) {
-//     char tmp[1024];
-//     const char *symbol = "(unknown)";
-//     if (absl::Symbolize(pc, tmp, sizeof(tmp))) {
-//       symbol = tmp;
-//     }
-//     absl::PrintF("%p  %s\n", pc, symbol);
-//  }
-bool Symbolize(const void *pc, char *out, int out_size);
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_DEBUGGING_SYMBOLIZE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/commandlineflag.h b/projects/llm_framework/include/abseil-cpp/absl/flags/commandlineflag.h
deleted file mode 100644
index 26ec0e7d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/commandlineflag.h
+++ /dev/null
@@ -1,211 +0,0 @@
-//
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: commandlineflag.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines the `CommandLineFlag`, which acts as a type-erased
-// handle for accessing metadata about the Abseil Flag in question.
-//
-// Because an actual Abseil flag is of an unspecified type, you should not
-// manipulate or interact directly with objects of that type. Instead, use the
-// CommandLineFlag type as an intermediary.
-#ifndef ABSL_FLAGS_COMMANDLINEFLAG_H_
-#define ABSL_FLAGS_COMMANDLINEFLAG_H_
-
-#include <memory>
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/base/internal/fast_type_id.h"
-#include "absl/flags/internal/commandlineflag.h"
-#include "absl/strings/string_view.h"
-#include "absl/types/optional.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace flags_internal {
-class PrivateHandleAccessor;
-}  // namespace flags_internal
-
-// CommandLineFlag
-//
-// This type acts as a type-erased handle for an instance of an Abseil Flag and
-// holds reflection information pertaining to that flag. Use CommandLineFlag to
-// access a flag's name, location, help string etc.
-//
-// To obtain an absl::CommandLineFlag, invoke `absl::FindCommandLineFlag()`
-// passing it the flag name string.
-//
-// Example:
-//
-//   // Obtain reflection handle for a flag named "flagname".
-//   const absl::CommandLineFlag* my_flag_data =
-//        absl::FindCommandLineFlag("flagname");
-//
-//   // Now you can get flag info from that reflection handle.
-//   std::string flag_location = my_flag_data->Filename();
-//   ...
-
-// These are only used as constexpr global objects.
-// They do not use a virtual destructor to simplify their implementation.
-// They are not destroyed except at program exit, so leaks do not matter.
-#if defined(__GNUC__) && !defined(__clang__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wnon-virtual-dtor"
-#endif
-class CommandLineFlag {
- public:
-  constexpr CommandLineFlag() = default;
-
-  // Not copyable/assignable.
-  CommandLineFlag(const CommandLineFlag&) = delete;
-  CommandLineFlag& operator=(const CommandLineFlag&) = delete;
-
-  // absl::CommandLineFlag::IsOfType()
-  //
-  // Return true iff flag has type T.
-  template <typename T>
-  inline bool IsOfType() const {
-    return TypeId() == base_internal::FastTypeId<T>();
-  }
-
-  // absl::CommandLineFlag::TryGet()
-  //
-  // Attempts to retrieve the flag value. Returns value on success,
-  // absl::nullopt otherwise.
-  template <typename T>
-  absl::optional<T> TryGet() const {
-    if (IsRetired() || !IsOfType<T>()) {
-      return absl::nullopt;
-    }
-
-    // Implementation notes:
-    //
-    // We are wrapping a union around the value of `T` to serve three purposes:
-    //
-    //  1. `U.value` has correct size and alignment for a value of type `T`
-    //  2. The `U.value` constructor is not invoked since U's constructor does
-    //     not do it explicitly.
-    //  3. The `U.value` destructor is invoked since U's destructor does it
-    //     explicitly. This makes `U` a kind of RAII wrapper around non default
-    //     constructible value of T, which is destructed when we leave the
-    //     scope. We do need to destroy U.value, which is constructed by
-    //     CommandLineFlag::Read even though we left it in a moved-from state
-    //     after std::move.
-    //
-    // All of this serves to avoid requiring `T` being default constructible.
-    union U {
-      T value;
-      U() {}
-      ~U() { value.~T(); }
-    };
-    U u;
-
-    Read(&u.value);
-    // allow retired flags to be "read", so we can report invalid access.
-    if (IsRetired()) {
-      return absl::nullopt;
-    }
-    return std::move(u.value);
-  }
-
-  // absl::CommandLineFlag::Name()
-  //
-  // Returns name of this flag.
-  virtual absl::string_view Name() const = 0;
-
-  // absl::CommandLineFlag::Filename()
-  //
-  // Returns name of the file where this flag is defined.
-  virtual std::string Filename() const = 0;
-
-  // absl::CommandLineFlag::Help()
-  //
-  // Returns help message associated with this flag.
-  virtual std::string Help() const = 0;
-
-  // absl::CommandLineFlag::IsRetired()
-  //
-  // Returns true iff this object corresponds to retired flag.
-  virtual bool IsRetired() const;
-
-  // absl::CommandLineFlag::DefaultValue()
-  //
-  // Returns the default value for this flag.
-  virtual std::string DefaultValue() const = 0;
-
-  // absl::CommandLineFlag::CurrentValue()
-  //
-  // Returns the current value for this flag.
-  virtual std::string CurrentValue() const = 0;
-
-  // absl::CommandLineFlag::ParseFrom()
-  //
-  // Sets the value of the flag based on specified string `value`. If the flag
-  // was successfully set to new value, it returns true. Otherwise, sets `error`
-  // to indicate the error, leaves the flag unchanged, and returns false.
-  bool ParseFrom(absl::string_view value, std::string* error);
-
- protected:
-  ~CommandLineFlag() = default;
-
- private:
-  friend class flags_internal::PrivateHandleAccessor;
-
-  // Sets the value of the flag based on specified string `value`. If the flag
-  // was successfully set to new value, it returns true. Otherwise, sets `error`
-  // to indicate the error, leaves the flag unchanged, and returns false. There
-  // are three ways to set the flag's value:
-  //  * Update the current flag value
-  //  * Update the flag's default value
-  //  * Update the current flag value if it was never set before
-  // The mode is selected based on `set_mode` parameter.
-  virtual bool ParseFrom(absl::string_view value,
-                         flags_internal::FlagSettingMode set_mode,
-                         flags_internal::ValueSource source,
-                         std::string& error) = 0;
-
-  // Returns id of the flag's value type.
-  virtual flags_internal::FlagFastTypeId TypeId() const = 0;
-
-  // Interface to save flag to some persistent state. Returns current flag state
-  // or nullptr if flag does not support saving and restoring a state.
-  virtual std::unique_ptr<flags_internal::FlagStateInterface> SaveState() = 0;
-
-  // Copy-construct a new value of the flag's type in a memory referenced by
-  // the dst based on the current flag's value.
-  virtual void Read(void* dst) const = 0;
-
-  // To be deleted. Used to return true if flag's current value originated from
-  // command line.
-  virtual bool IsSpecifiedOnCommandLine() const = 0;
-
-  // Validates supplied value using validator or parseflag routine
-  virtual bool ValidateInputValue(absl::string_view value) const = 0;
-
-  // Checks that flags default value can be converted to string and back to the
-  // flag's value type.
-  virtual void CheckDefaultValueParsingRoundtrip() const = 0;
-};
-#if defined(__GNUC__) && !defined(__clang__)
-#pragma GCC diagnostic pop
-#endif
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FLAGS_COMMANDLINEFLAG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/config.h b/projects/llm_framework/include/abseil-cpp/absl/flags/config.h
deleted file mode 100644
index 14c4235b..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/config.h
+++ /dev/null
@@ -1,68 +0,0 @@
-//
-//  Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_FLAGS_CONFIG_H_
-#define ABSL_FLAGS_CONFIG_H_
-
-// Determine if we should strip string literals from the Flag objects.
-// By default we strip string literals on mobile platforms.
-#if !defined(ABSL_FLAGS_STRIP_NAMES)
-
-#if defined(__ANDROID__)
-#define ABSL_FLAGS_STRIP_NAMES 1
-
-#elif defined(__APPLE__)
-#include <TargetConditionals.h>
-#if defined(TARGET_OS_IPHONE) && TARGET_OS_IPHONE
-#define ABSL_FLAGS_STRIP_NAMES 1
-#elif defined(TARGET_OS_EMBEDDED) && TARGET_OS_EMBEDDED
-#define ABSL_FLAGS_STRIP_NAMES 1
-#endif  // TARGET_OS_*
-#endif
-
-#endif  // !defined(ABSL_FLAGS_STRIP_NAMES)
-
-#if !defined(ABSL_FLAGS_STRIP_NAMES)
-// If ABSL_FLAGS_STRIP_NAMES wasn't set on the command line or above,
-// the default is not to strip.
-#define ABSL_FLAGS_STRIP_NAMES 0
-#endif
-
-#if !defined(ABSL_FLAGS_STRIP_HELP)
-// By default, if we strip names, we also strip help.
-#define ABSL_FLAGS_STRIP_HELP ABSL_FLAGS_STRIP_NAMES
-#endif
-
-// These macros represent the "source of truth" for the list of supported
-// built-in types.
-#define ABSL_FLAGS_INTERNAL_BUILTIN_TYPES(A) \
-  A(bool, bool)                              \
-  A(short, short)                            \
-  A(unsigned short, unsigned_short)          \
-  A(int, int)                                \
-  A(unsigned int, unsigned_int)              \
-  A(long, long)                              \
-  A(unsigned long, unsigned_long)            \
-  A(long long, long_long)                    \
-  A(unsigned long long, unsigned_long_long)  \
-  A(double, double)                          \
-  A(float, float)
-
-#define ABSL_FLAGS_INTERNAL_SUPPORTED_TYPES(A) \
-  ABSL_FLAGS_INTERNAL_BUILTIN_TYPES(A)         \
-  A(std::string, std_string)                   \
-  A(std::vector<std::string>, std_vector_of_string)
-
-#endif  // ABSL_FLAGS_CONFIG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/declare.h b/projects/llm_framework/include/abseil-cpp/absl/flags/declare.h
deleted file mode 100644
index 8d2a856e..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/declare.h
+++ /dev/null
@@ -1,68 +0,0 @@
-//
-//  Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: declare.h
-// -----------------------------------------------------------------------------
-//
-// This file defines the ABSL_DECLARE_FLAG macro, allowing you to declare an
-// `absl::Flag` for use within a translation unit. You should place this
-// declaration within the header file associated with the .cc file that defines
-// and owns the `Flag`.
-
-#ifndef ABSL_FLAGS_DECLARE_H_
-#define ABSL_FLAGS_DECLARE_H_
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace flags_internal {
-
-// absl::Flag<T> represents a flag of type 'T' created by ABSL_FLAG.
-template <typename T>
-class Flag;
-
-}  // namespace flags_internal
-
-// Flag
-//
-// Forward declaration of the `absl::Flag` type for use in defining the macro.
-template <typename T>
-using Flag = flags_internal::Flag<T>;
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-// ABSL_DECLARE_FLAG()
-//
-// This macro is a convenience for declaring use of an `absl::Flag` within a
-// translation unit. This macro should be used within a header file to
-// declare usage of the flag within any .cc file including that header file.
-//
-// The ABSL_DECLARE_FLAG(type, name) macro expands to:
-//
-//   extern absl::Flag<type> FLAGS_name;
-#define ABSL_DECLARE_FLAG(type, name) ABSL_DECLARE_FLAG_INTERNAL(type, name)
-
-// Internal implementation of ABSL_DECLARE_FLAG to allow macro expansion of its
-// arguments. Clients must use ABSL_DECLARE_FLAG instead.
-#define ABSL_DECLARE_FLAG_INTERNAL(type, name)               \
-  extern absl::Flag<type> FLAGS_##name;                      \
-  namespace absl /* block flags in namespaces */ {}          \
-  /* second redeclaration is to allow applying attributes */ \
-  extern absl::Flag<type> FLAGS_##name
-
-#endif  // ABSL_FLAGS_DECLARE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/flag.h b/projects/llm_framework/include/abseil-cpp/absl/flags/flag.h
deleted file mode 100644
index a8e0e932..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/flag.h
+++ /dev/null
@@ -1,303 +0,0 @@
-//
-//  Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: flag.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines the `absl::Flag<T>` type for holding command-line
-// flag data, and abstractions to create, get and set such flag data.
-//
-// It is important to note that this type is **unspecified** (an implementation
-// detail) and you do not construct or manipulate actual `absl::Flag<T>`
-// instances. Instead, you define and declare flags using the
-// `ABSL_FLAG()` and `ABSL_DECLARE_FLAG()` macros, and get and set flag values
-// using the `absl::GetFlag()` and `absl::SetFlag()` functions.
-
-#ifndef ABSL_FLAGS_FLAG_H_
-#define ABSL_FLAGS_FLAG_H_
-
-#include <cstdint>
-#include <string>
-#include <type_traits>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/optimization.h"
-#include "absl/flags/commandlineflag.h"
-#include "absl/flags/config.h"
-#include "absl/flags/internal/flag.h"
-#include "absl/flags/internal/registry.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// Flag
-//
-// An `absl::Flag` holds a command-line flag value, providing a runtime
-// parameter to a binary. Such flags should be defined in the global namespace
-// and (preferably) in the module containing the binary's `main()` function.
-//
-// You should not construct and cannot use the `absl::Flag` type directly;
-// instead, you should declare flags using the `ABSL_DECLARE_FLAG()` macro
-// within a header file, and define your flag using `ABSL_FLAG()` within your
-// header's associated `.cc` file. Such flags will be named `FLAGS_name`.
-//
-// Example:
-//
-//    .h file
-//
-//      // Declares usage of a flag named "FLAGS_count"
-//      ABSL_DECLARE_FLAG(int, count);
-//
-//    .cc file
-//
-//      // Defines a flag named "FLAGS_count" with a default `int` value of 0.
-//      ABSL_FLAG(int, count, 0, "Count of items to process");
-//
-// No public methods of `absl::Flag<T>` are part of the Abseil Flags API.
-//
-// For type support of Abseil Flags, see the marshalling.h header file, which
-// discusses supported standard types, optional flags, and additional Abseil
-// type support.
-
-template <typename T>
-using Flag = flags_internal::Flag<T>;
-
-// GetFlag()
-//
-// Returns the value (of type `T`) of an `absl::Flag<T>` instance, by value. Do
-// not construct an `absl::Flag<T>` directly and call `absl::GetFlag()`;
-// instead, refer to flag's constructed variable name (e.g. `FLAGS_name`).
-// Because this function returns by value and not by reference, it is
-// thread-safe, but note that the operation may be expensive; as a result, avoid
-// `absl::GetFlag()` within any tight loops.
-//
-// Example:
-//
-//   // FLAGS_count is a Flag of type `int`
-//   int my_count = absl::GetFlag(FLAGS_count);
-//
-//   // FLAGS_firstname is a Flag of type `std::string`
-//   std::string first_name = absl::GetFlag(FLAGS_firstname);
-template <typename T>
-ABSL_MUST_USE_RESULT T GetFlag(const absl::Flag<T>& flag) {
-  return flags_internal::FlagImplPeer::InvokeGet<T>(flag);
-}
-
-// SetFlag()
-//
-// Sets the value of an `absl::Flag` to the value `v`. Do not construct an
-// `absl::Flag<T>` directly and call `absl::SetFlag()`; instead, use the
-// flag's variable name (e.g. `FLAGS_name`). This function is
-// thread-safe, but is potentially expensive. Avoid setting flags in general,
-// but especially within performance-critical code.
-template <typename T>
-void SetFlag(absl::Flag<T>* flag, const T& v) {
-  flags_internal::FlagImplPeer::InvokeSet(*flag, v);
-}
-
-// Overload of `SetFlag()` to allow callers to pass in a value that is
-// convertible to `T`. E.g., use this overload to pass a "const char*" when `T`
-// is `std::string`.
-template <typename T, typename V>
-void SetFlag(absl::Flag<T>* flag, const V& v) {
-  T value(v);
-  flags_internal::FlagImplPeer::InvokeSet(*flag, value);
-}
-
-// GetFlagReflectionHandle()
-//
-// Returns the reflection handle corresponding to specified Abseil Flag
-// instance. Use this handle to access flag's reflection information, like name,
-// location, default value etc.
-//
-// Example:
-//
-//   std::string = absl::GetFlagReflectionHandle(FLAGS_count).DefaultValue();
-
-template <typename T>
-const CommandLineFlag& GetFlagReflectionHandle(const absl::Flag<T>& f) {
-  return flags_internal::FlagImplPeer::InvokeReflect(f);
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-
-// ABSL_FLAG()
-//
-// This macro defines an `absl::Flag<T>` instance of a specified type `T`:
-//
-//   ABSL_FLAG(T, name, default_value, help);
-//
-// where:
-//
-//   * `T` is a supported flag type (see the list of types in `marshalling.h`),
-//   * `name` designates the name of the flag (as a global variable
-//     `FLAGS_name`),
-//   * `default_value` is an expression holding the default value for this flag
-//     (which must be implicitly convertible to `T`),
-//   * `help` is the help text, which can also be an expression.
-//
-// This macro expands to a flag named 'FLAGS_name' of type 'T':
-//
-//   absl::Flag<T> FLAGS_name = ...;
-//
-// Note that all such instances are created as global variables.
-//
-// For `ABSL_FLAG()` values that you wish to expose to other translation units,
-// it is recommended to define those flags within the `.cc` file associated with
-// the header where the flag is declared.
-//
-// Note: do not construct objects of type `absl::Flag<T>` directly. Only use the
-// `ABSL_FLAG()` macro for such construction.
-#define ABSL_FLAG(Type, name, default_value, help) \
-  ABSL_FLAG_IMPL(Type, name, default_value, help)
-
-// ABSL_FLAG().OnUpdate()
-//
-// Defines a flag of type `T` with a callback attached:
-//
-//   ABSL_FLAG(T, name, default_value, help).OnUpdate(callback);
-//
-// `callback` should be convertible to `void (*)()`.
-//
-// After any setting of the flag value, the callback will be called at least
-// once. A rapid sequence of changes may be merged together into the same
-// callback. No concurrent calls to the callback will be made for the same
-// flag. Callbacks are allowed to read the current value of the flag but must
-// not mutate that flag.
-//
-// The update mechanism guarantees "eventual consistency"; if the callback
-// derives an auxiliary data structure from the flag value, it is guaranteed
-// that eventually the flag value and the derived data structure will be
-// consistent.
-//
-// Note: ABSL_FLAG.OnUpdate() does not have a public definition. Hence, this
-// comment serves as its API documentation.
-
-// -----------------------------------------------------------------------------
-// Implementation details below this section
-// -----------------------------------------------------------------------------
-
-// ABSL_FLAG_IMPL macro definition conditional on ABSL_FLAGS_STRIP_NAMES
-#define ABSL_FLAG_IMPL_FLAG_PTR(flag) flag
-#define ABSL_FLAG_IMPL_HELP_ARG(name)                      \
-  absl::flags_internal::HelpArg<AbslFlagHelpGenFor##name>( \
-      FLAGS_help_storage_##name)
-#define ABSL_FLAG_IMPL_DEFAULT_ARG(Type, name) \
-  absl::flags_internal::DefaultArg<Type, AbslFlagDefaultGenFor##name>(0)
-
-#if ABSL_FLAGS_STRIP_NAMES
-#define ABSL_FLAG_IMPL_FLAGNAME(txt) ""
-#define ABSL_FLAG_IMPL_FILENAME() ""
-#define ABSL_FLAG_IMPL_REGISTRAR(T, flag)                                      \
-  absl::flags_internal::FlagRegistrar<T, false>(ABSL_FLAG_IMPL_FLAG_PTR(flag), \
-                                                nullptr)
-#else
-#define ABSL_FLAG_IMPL_FLAGNAME(txt) txt
-#define ABSL_FLAG_IMPL_FILENAME() __FILE__
-#define ABSL_FLAG_IMPL_REGISTRAR(T, flag)                                     \
-  absl::flags_internal::FlagRegistrar<T, true>(ABSL_FLAG_IMPL_FLAG_PTR(flag), \
-                                               __FILE__)
-#endif
-
-// ABSL_FLAG_IMPL macro definition conditional on ABSL_FLAGS_STRIP_HELP
-
-#if ABSL_FLAGS_STRIP_HELP
-#define ABSL_FLAG_IMPL_FLAGHELP(txt) absl::flags_internal::kStrippedFlagHelp
-#else
-#define ABSL_FLAG_IMPL_FLAGHELP(txt) txt
-#endif
-
-// AbslFlagHelpGenFor##name is used to encapsulate both immediate (method Const)
-// and lazy (method NonConst) evaluation of help message expression. We choose
-// between the two via the call to HelpArg in absl::Flag instantiation below.
-// If help message expression is constexpr evaluable compiler will optimize
-// away this whole struct.
-// TODO(rogeeff): place these generated structs into local namespace and apply
-// ABSL_INTERNAL_UNIQUE_SHORT_NAME.
-// TODO(rogeeff): Apply __attribute__((nodebug)) to FLAGS_help_storage_##name
-#define ABSL_FLAG_IMPL_DECLARE_HELP_WRAPPER(name, txt)                       \
-  struct AbslFlagHelpGenFor##name {                                          \
-    /* The expression is run in the caller as part of the   */               \
-    /* default value argument. That keeps temporaries alive */               \
-    /* long enough for NonConst to work correctly.          */               \
-    static constexpr absl::string_view Value(                                \
-        absl::string_view absl_flag_help = ABSL_FLAG_IMPL_FLAGHELP(txt)) {   \
-      return absl_flag_help;                                                 \
-    }                                                                        \
-    static std::string NonConst() { return std::string(Value()); }           \
-  };                                                                         \
-  constexpr auto FLAGS_help_storage_##name ABSL_INTERNAL_UNIQUE_SMALL_NAME() \
-      ABSL_ATTRIBUTE_SECTION_VARIABLE(flags_help_cold) =                     \
-          absl::flags_internal::HelpStringAsArray<AbslFlagHelpGenFor##name>( \
-              0);
-
-#define ABSL_FLAG_IMPL_DECLARE_DEF_VAL_WRAPPER(name, Type, default_value)     \
-  struct AbslFlagDefaultGenFor##name {                                        \
-    Type value = absl::flags_internal::InitDefaultValue<Type>(default_value); \
-    static void Gen(void* absl_flag_default_loc) {                            \
-      new (absl_flag_default_loc) Type(AbslFlagDefaultGenFor##name{}.value);  \
-    }                                                                         \
-  };
-
-// ABSL_FLAG_IMPL
-//
-// Note: Name of registrar object is not arbitrary. It is used to "grab"
-// global name for FLAGS_no<flag_name> symbol, thus preventing the possibility
-// of defining two flags with names foo and nofoo.
-#define ABSL_FLAG_IMPL(Type, name, default_value, help)                       \
-  extern ::absl::Flag<Type> FLAGS_##name;                                     \
-  namespace absl /* block flags in namespaces */ {}                           \
-  ABSL_FLAG_IMPL_DECLARE_DEF_VAL_WRAPPER(name, Type, default_value)           \
-  ABSL_FLAG_IMPL_DECLARE_HELP_WRAPPER(name, help)                             \
-  ABSL_CONST_INIT absl::Flag<Type> FLAGS_##name{                              \
-      ABSL_FLAG_IMPL_FLAGNAME(#name), ABSL_FLAG_IMPL_FILENAME(),              \
-      ABSL_FLAG_IMPL_HELP_ARG(name), ABSL_FLAG_IMPL_DEFAULT_ARG(Type, name)}; \
-  extern absl::flags_internal::FlagRegistrarEmpty FLAGS_no##name;             \
-  absl::flags_internal::FlagRegistrarEmpty FLAGS_no##name =                   \
-      ABSL_FLAG_IMPL_REGISTRAR(Type, FLAGS_##name)
-
-// ABSL_RETIRED_FLAG
-//
-// Designates the flag (which is usually pre-existing) as "retired." A retired
-// flag is a flag that is now unused by the program, but may still be passed on
-// the command line, usually by production scripts. A retired flag is ignored
-// and code can't access it at runtime.
-//
-// This macro registers a retired flag with given name and type, with a name
-// identical to the name of the original flag you are retiring. The retired
-// flag's type can change over time, so that you can retire code to support a
-// custom flag type.
-//
-// This macro has the same signature as `ABSL_FLAG`. To retire a flag, simply
-// replace an `ABSL_FLAG` definition with `ABSL_RETIRED_FLAG`, leaving the
-// arguments unchanged (unless of course you actually want to retire the flag
-// type at this time as well).
-//
-// `default_value` is only used as a double check on the type. `explanation` is
-// unused.
-// TODO(rogeeff): replace RETIRED_FLAGS with FLAGS once forward declarations of
-// retired flags are cleaned up.
-#define ABSL_RETIRED_FLAG(type, name, default_value, explanation)      \
-  static absl::flags_internal::RetiredFlag<type> RETIRED_FLAGS_##name; \
-  ABSL_ATTRIBUTE_UNUSED static const auto RETIRED_FLAGS_REG_##name =   \
-      (RETIRED_FLAGS_##name.Retire(#name),                             \
-       ::absl::flags_internal::FlagRegistrarEmpty{})
-
-#endif  // ABSL_FLAGS_FLAG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/commandlineflag.h b/projects/llm_framework/include/abseil-cpp/absl/flags/internal/commandlineflag.h
deleted file mode 100644
index ebfe81ba..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/commandlineflag.h
+++ /dev/null
@@ -1,68 +0,0 @@
-//
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_FLAGS_INTERNAL_COMMANDLINEFLAG_H_
-#define ABSL_FLAGS_INTERNAL_COMMANDLINEFLAG_H_
-
-#include "absl/base/config.h"
-#include "absl/base/internal/fast_type_id.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace flags_internal {
-
-// An alias for flag fast type id. This value identifies the flag value type
-// similarly to typeid(T), without relying on RTTI being available. In most
-// cases this id is enough to uniquely identify the flag's value type. In a few
-// cases we'll have to resort to using actual RTTI implementation if it is
-// available.
-using FlagFastTypeId = absl::base_internal::FastTypeIdType;
-
-// Options that control SetCommandLineOptionWithMode.
-enum FlagSettingMode {
-  // update the flag's value unconditionally (can call this multiple times).
-  SET_FLAGS_VALUE,
-  // update the flag's value, but *only if* it has not yet been updated
-  // with SET_FLAGS_VALUE, SET_FLAG_IF_DEFAULT, or "FLAGS_xxx = nondef".
-  SET_FLAG_IF_DEFAULT,
-  // set the flag's default value to this.  If the flag has not been updated
-  // yet (via SET_FLAGS_VALUE, SET_FLAG_IF_DEFAULT, or "FLAGS_xxx = nondef")
-  // change the flag's current value to the new default value as well.
-  SET_FLAGS_DEFAULT
-};
-
-// Options that control ParseFrom: Source of a value.
-enum ValueSource {
-  // Flag is being set by value specified on a command line.
-  kCommandLine,
-  // Flag is being set by value specified in the code.
-  kProgrammaticChange,
-};
-
-// Handle to FlagState objects. Specific flag state objects will restore state
-// of a flag produced this flag state from method CommandLineFlag::SaveState().
-class FlagStateInterface {
- public:
-  virtual ~FlagStateInterface();
-
-  // Restores the flag originated this object to the saved state.
-  virtual void Restore() const = 0;
-};
-
-}  // namespace flags_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FLAGS_INTERNAL_COMMANDLINEFLAG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/flag.h b/projects/llm_framework/include/abseil-cpp/absl/flags/internal/flag.h
deleted file mode 100644
index 602531b1..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/flag.h
+++ /dev/null
@@ -1,805 +0,0 @@
-//
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_FLAGS_INTERNAL_FLAG_H_
-#define ABSL_FLAGS_INTERNAL_FLAG_H_
-
-#include <stddef.h>
-#include <stdint.h>
-
-#include <atomic>
-#include <cstring>
-#include <memory>
-#include <string>
-#include <type_traits>
-#include <typeinfo>
-
-#include "absl/base/attributes.h"
-#include "absl/base/call_once.h"
-#include "absl/base/casts.h"
-#include "absl/base/config.h"
-#include "absl/base/optimization.h"
-#include "absl/base/thread_annotations.h"
-#include "absl/flags/commandlineflag.h"
-#include "absl/flags/config.h"
-#include "absl/flags/internal/commandlineflag.h"
-#include "absl/flags/internal/registry.h"
-#include "absl/flags/internal/sequence_lock.h"
-#include "absl/flags/marshalling.h"
-#include "absl/meta/type_traits.h"
-#include "absl/strings/string_view.h"
-#include "absl/synchronization/mutex.h"
-#include "absl/utility/utility.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-///////////////////////////////////////////////////////////////////////////////
-// Forward declaration of absl::Flag<T> public API.
-namespace flags_internal {
-template <typename T>
-class Flag;
-}  // namespace flags_internal
-
-template <typename T>
-using Flag = flags_internal::Flag<T>;
-
-template <typename T>
-ABSL_MUST_USE_RESULT T GetFlag(const absl::Flag<T>& flag);
-
-template <typename T>
-void SetFlag(absl::Flag<T>* flag, const T& v);
-
-template <typename T, typename V>
-void SetFlag(absl::Flag<T>* flag, const V& v);
-
-template <typename U>
-const CommandLineFlag& GetFlagReflectionHandle(const absl::Flag<U>& f);
-
-///////////////////////////////////////////////////////////////////////////////
-// Flag value type operations, eg., parsing, copying, etc. are provided
-// by function specific to that type with a signature matching FlagOpFn.
-
-namespace flags_internal {
-
-enum class FlagOp {
-  kAlloc,
-  kDelete,
-  kCopy,
-  kCopyConstruct,
-  kSizeof,
-  kFastTypeId,
-  kRuntimeTypeId,
-  kParse,
-  kUnparse,
-  kValueOffset,
-};
-using FlagOpFn = void* (*)(FlagOp, const void*, void*, void*);
-
-// Forward declaration for Flag value specific operations.
-template <typename T>
-void* FlagOps(FlagOp op, const void* v1, void* v2, void* v3);
-
-// Allocate aligned memory for a flag value.
-inline void* Alloc(FlagOpFn op) {
-  return op(FlagOp::kAlloc, nullptr, nullptr, nullptr);
-}
-// Deletes memory interpreting obj as flag value type pointer.
-inline void Delete(FlagOpFn op, void* obj) {
-  op(FlagOp::kDelete, nullptr, obj, nullptr);
-}
-// Copies src to dst interpreting as flag value type pointers.
-inline void Copy(FlagOpFn op, const void* src, void* dst) {
-  op(FlagOp::kCopy, src, dst, nullptr);
-}
-// Construct a copy of flag value in a location pointed by dst
-// based on src - pointer to the flag's value.
-inline void CopyConstruct(FlagOpFn op, const void* src, void* dst) {
-  op(FlagOp::kCopyConstruct, src, dst, nullptr);
-}
-// Makes a copy of flag value pointed by obj.
-inline void* Clone(FlagOpFn op, const void* obj) {
-  void* res = flags_internal::Alloc(op);
-  flags_internal::CopyConstruct(op, obj, res);
-  return res;
-}
-// Returns true if parsing of input text is successful.
-inline bool Parse(FlagOpFn op, absl::string_view text, void* dst,
-                  std::string* error) {
-  return op(FlagOp::kParse, &text, dst, error) != nullptr;
-}
-// Returns string representing supplied value.
-inline std::string Unparse(FlagOpFn op, const void* val) {
-  std::string result;
-  op(FlagOp::kUnparse, val, &result, nullptr);
-  return result;
-}
-// Returns size of flag value type.
-inline size_t Sizeof(FlagOpFn op) {
-  // This sequence of casts reverses the sequence from
-  // `flags_internal::FlagOps()`
-  return static_cast<size_t>(reinterpret_cast<intptr_t>(
-      op(FlagOp::kSizeof, nullptr, nullptr, nullptr)));
-}
-// Returns fast type id corresponding to the value type.
-inline FlagFastTypeId FastTypeId(FlagOpFn op) {
-  return reinterpret_cast<FlagFastTypeId>(
-      op(FlagOp::kFastTypeId, nullptr, nullptr, nullptr));
-}
-// Returns fast type id corresponding to the value type.
-inline const std::type_info* RuntimeTypeId(FlagOpFn op) {
-  return reinterpret_cast<const std::type_info*>(
-      op(FlagOp::kRuntimeTypeId, nullptr, nullptr, nullptr));
-}
-// Returns offset of the field value_ from the field impl_ inside of
-// absl::Flag<T> data. Given FlagImpl pointer p you can get the
-// location of the corresponding value as:
-//      reinterpret_cast<char*>(p) + ValueOffset().
-inline ptrdiff_t ValueOffset(FlagOpFn op) {
-  // This sequence of casts reverses the sequence from
-  // `flags_internal::FlagOps()`
-  return static_cast<ptrdiff_t>(reinterpret_cast<intptr_t>(
-      op(FlagOp::kValueOffset, nullptr, nullptr, nullptr)));
-}
-
-// Returns an address of RTTI's typeid(T).
-template <typename T>
-inline const std::type_info* GenRuntimeTypeId() {
-#ifdef ABSL_INTERNAL_HAS_RTTI
-  return &typeid(T);
-#else
-  return nullptr;
-#endif
-}
-
-///////////////////////////////////////////////////////////////////////////////
-// Flag help auxiliary structs.
-
-// This is help argument for absl::Flag encapsulating the string literal pointer
-// or pointer to function generating it as well as enum descriminating two
-// cases.
-using HelpGenFunc = std::string (*)();
-
-template <size_t N>
-struct FixedCharArray {
-  char value[N];
-
-  template <size_t... I>
-  static constexpr FixedCharArray<N> FromLiteralString(
-      absl::string_view str, absl::index_sequence<I...>) {
-    return (void)str, FixedCharArray<N>({{str[I]..., '\0'}});
-  }
-};
-
-template <typename Gen, size_t N = Gen::Value().size()>
-constexpr FixedCharArray<N + 1> HelpStringAsArray(int) {
-  return FixedCharArray<N + 1>::FromLiteralString(
-      Gen::Value(), absl::make_index_sequence<N>{});
-}
-
-template <typename Gen>
-constexpr std::false_type HelpStringAsArray(char) {
-  return std::false_type{};
-}
-
-union FlagHelpMsg {
-  constexpr explicit FlagHelpMsg(const char* help_msg) : literal(help_msg) {}
-  constexpr explicit FlagHelpMsg(HelpGenFunc help_gen) : gen_func(help_gen) {}
-
-  const char* literal;
-  HelpGenFunc gen_func;
-};
-
-enum class FlagHelpKind : uint8_t { kLiteral = 0, kGenFunc = 1 };
-
-struct FlagHelpArg {
-  FlagHelpMsg source;
-  FlagHelpKind kind;
-};
-
-extern const char kStrippedFlagHelp[];
-
-// These two HelpArg overloads allows us to select at compile time one of two
-// way to pass Help argument to absl::Flag. We'll be passing
-// AbslFlagHelpGenFor##name as Gen and integer 0 as a single argument to prefer
-// first overload if possible. If help message is evaluatable on constexpr
-// context We'll be able to make FixedCharArray out of it and we'll choose first
-// overload. In this case the help message expression is immediately evaluated
-// and is used to construct the absl::Flag. No additional code is generated by
-// ABSL_FLAG Otherwise SFINAE kicks in and first overload is dropped from the
-// consideration, in which case the second overload will be used. The second
-// overload does not attempt to evaluate the help message expression
-// immediately and instead delays the evaluation by returning the function
-// pointer (&T::NonConst) generating the help message when necessary. This is
-// evaluatable in constexpr context, but the cost is an extra function being
-// generated in the ABSL_FLAG code.
-template <typename Gen, size_t N>
-constexpr FlagHelpArg HelpArg(const FixedCharArray<N>& value) {
-  return {FlagHelpMsg(value.value), FlagHelpKind::kLiteral};
-}
-
-template <typename Gen>
-constexpr FlagHelpArg HelpArg(std::false_type) {
-  return {FlagHelpMsg(&Gen::NonConst), FlagHelpKind::kGenFunc};
-}
-
-///////////////////////////////////////////////////////////////////////////////
-// Flag default value auxiliary structs.
-
-// Signature for the function generating the initial flag value (usually
-// based on default value supplied in flag's definition)
-using FlagDfltGenFunc = void (*)(void*);
-
-union FlagDefaultSrc {
-  constexpr explicit FlagDefaultSrc(FlagDfltGenFunc gen_func_arg)
-      : gen_func(gen_func_arg) {}
-
-#define ABSL_FLAGS_INTERNAL_DFLT_FOR_TYPE(T, name) \
-  T name##_value;                                  \
-  constexpr explicit FlagDefaultSrc(T value) : name##_value(value) {}  // NOLINT
-  ABSL_FLAGS_INTERNAL_BUILTIN_TYPES(ABSL_FLAGS_INTERNAL_DFLT_FOR_TYPE)
-#undef ABSL_FLAGS_INTERNAL_DFLT_FOR_TYPE
-
-  void* dynamic_value;
-  FlagDfltGenFunc gen_func;
-};
-
-enum class FlagDefaultKind : uint8_t {
-  kDynamicValue = 0,
-  kGenFunc = 1,
-  kOneWord = 2  // for default values UP to one word in size
-};
-
-struct FlagDefaultArg {
-  FlagDefaultSrc source;
-  FlagDefaultKind kind;
-};
-
-// This struct and corresponding overload to InitDefaultValue are used to
-// facilitate usage of {} as default value in ABSL_FLAG macro.
-// TODO(rogeeff): Fix handling types with explicit constructors.
-struct EmptyBraces {};
-
-template <typename T>
-constexpr T InitDefaultValue(T t) {
-  return t;
-}
-
-template <typename T>
-constexpr T InitDefaultValue(EmptyBraces) {
-  return T{};
-}
-
-template <typename ValueT, typename GenT,
-          typename std::enable_if<std::is_integral<ValueT>::value, int>::type =
-              ((void)GenT{}, 0)>
-constexpr FlagDefaultArg DefaultArg(int) {
-  return {FlagDefaultSrc(GenT{}.value), FlagDefaultKind::kOneWord};
-}
-
-template <typename ValueT, typename GenT>
-constexpr FlagDefaultArg DefaultArg(char) {
-  return {FlagDefaultSrc(&GenT::Gen), FlagDefaultKind::kGenFunc};
-}
-
-///////////////////////////////////////////////////////////////////////////////
-// Flag current value auxiliary structs.
-
-constexpr int64_t UninitializedFlagValue() {
-  return static_cast<int64_t>(0xababababababababll);
-}
-
-template <typename T>
-using FlagUseValueAndInitBitStorage =
-    std::integral_constant<bool, std::is_trivially_copyable<T>::value &&
-                                     std::is_default_constructible<T>::value &&
-                                     (sizeof(T) < 8)>;
-
-template <typename T>
-using FlagUseOneWordStorage =
-    std::integral_constant<bool, std::is_trivially_copyable<T>::value &&
-                                     (sizeof(T) <= 8)>;
-
-template <class T>
-using FlagUseSequenceLockStorage =
-    std::integral_constant<bool, std::is_trivially_copyable<T>::value &&
-                                     (sizeof(T) > 8)>;
-
-enum class FlagValueStorageKind : uint8_t {
-  kValueAndInitBit = 0,
-  kOneWordAtomic = 1,
-  kSequenceLocked = 2,
-  kAlignedBuffer = 3,
-};
-
-template <typename T>
-static constexpr FlagValueStorageKind StorageKind() {
-  return FlagUseValueAndInitBitStorage<T>::value
-             ? FlagValueStorageKind::kValueAndInitBit
-         : FlagUseOneWordStorage<T>::value
-             ? FlagValueStorageKind::kOneWordAtomic
-         : FlagUseSequenceLockStorage<T>::value
-             ? FlagValueStorageKind::kSequenceLocked
-             : FlagValueStorageKind::kAlignedBuffer;
-}
-
-struct FlagOneWordValue {
-  constexpr explicit FlagOneWordValue(int64_t v) : value(v) {}
-  std::atomic<int64_t> value;
-};
-
-template <typename T>
-struct alignas(8) FlagValueAndInitBit {
-  T value;
-  // Use an int instead of a bool to guarantee that a non-zero value has
-  // a bit set.
-  uint8_t init;
-};
-
-template <typename T,
-          FlagValueStorageKind Kind = flags_internal::StorageKind<T>()>
-struct FlagValue;
-
-template <typename T>
-struct FlagValue<T, FlagValueStorageKind::kValueAndInitBit> : FlagOneWordValue {
-  constexpr FlagValue() : FlagOneWordValue(0) {}
-  bool Get(const SequenceLock&, T& dst) const {
-    int64_t storage = value.load(std::memory_order_acquire);
-    if (ABSL_PREDICT_FALSE(storage == 0)) {
-      return false;
-    }
-    dst = absl::bit_cast<FlagValueAndInitBit<T>>(storage).value;
-    return true;
-  }
-};
-
-template <typename T>
-struct FlagValue<T, FlagValueStorageKind::kOneWordAtomic> : FlagOneWordValue {
-  constexpr FlagValue() : FlagOneWordValue(UninitializedFlagValue()) {}
-  bool Get(const SequenceLock&, T& dst) const {
-    int64_t one_word_val = value.load(std::memory_order_acquire);
-    if (ABSL_PREDICT_FALSE(one_word_val == UninitializedFlagValue())) {
-      return false;
-    }
-    std::memcpy(&dst, static_cast<const void*>(&one_word_val), sizeof(T));
-    return true;
-  }
-};
-
-template <typename T>
-struct FlagValue<T, FlagValueStorageKind::kSequenceLocked> {
-  bool Get(const SequenceLock& lock, T& dst) const {
-    return lock.TryRead(&dst, value_words, sizeof(T));
-  }
-
-  static constexpr int kNumWords =
-      flags_internal::AlignUp(sizeof(T), sizeof(uint64_t)) / sizeof(uint64_t);
-
-  alignas(T) alignas(
-      std::atomic<uint64_t>) std::atomic<uint64_t> value_words[kNumWords];
-};
-
-template <typename T>
-struct FlagValue<T, FlagValueStorageKind::kAlignedBuffer> {
-  bool Get(const SequenceLock&, T&) const { return false; }
-
-  alignas(T) char value[sizeof(T)];
-};
-
-///////////////////////////////////////////////////////////////////////////////
-// Flag callback auxiliary structs.
-
-// Signature for the mutation callback used by watched Flags
-// The callback is noexcept.
-// TODO(rogeeff): add noexcept after C++17 support is added.
-using FlagCallbackFunc = void (*)();
-
-struct FlagCallback {
-  FlagCallbackFunc func;
-  absl::Mutex guard;  // Guard for concurrent callback invocations.
-};
-
-///////////////////////////////////////////////////////////////////////////////
-// Flag implementation, which does not depend on flag value type.
-// The class encapsulates the Flag's data and access to it.
-
-struct DynValueDeleter {
-  explicit DynValueDeleter(FlagOpFn op_arg = nullptr);
-  void operator()(void* ptr) const;
-
-  FlagOpFn op;
-};
-
-class FlagState;
-
-// These are only used as constexpr global objects.
-// They do not use a virtual destructor to simplify their implementation.
-// They are not destroyed except at program exit, so leaks do not matter.
-#if defined(__GNUC__) && !defined(__clang__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wnon-virtual-dtor"
-#endif
-class FlagImpl final : public CommandLineFlag {
- public:
-  constexpr FlagImpl(const char* name, const char* filename, FlagOpFn op,
-                     FlagHelpArg help, FlagValueStorageKind value_kind,
-                     FlagDefaultArg default_arg)
-      : name_(name),
-        filename_(filename),
-        op_(op),
-        help_(help.source),
-        help_source_kind_(static_cast<uint8_t>(help.kind)),
-        value_storage_kind_(static_cast<uint8_t>(value_kind)),
-        def_kind_(static_cast<uint8_t>(default_arg.kind)),
-        modified_(false),
-        on_command_line_(false),
-        callback_(nullptr),
-        default_value_(default_arg.source),
-        data_guard_{} {}
-
-  // Constant access methods
-  int64_t ReadOneWord() const ABSL_LOCKS_EXCLUDED(*DataGuard());
-  bool ReadOneBool() const ABSL_LOCKS_EXCLUDED(*DataGuard());
-  void Read(void* dst) const override ABSL_LOCKS_EXCLUDED(*DataGuard());
-  void Read(bool* value) const ABSL_LOCKS_EXCLUDED(*DataGuard()) {
-    *value = ReadOneBool();
-  }
-  template <typename T,
-            absl::enable_if_t<flags_internal::StorageKind<T>() ==
-                                  FlagValueStorageKind::kOneWordAtomic,
-                              int> = 0>
-  void Read(T* value) const ABSL_LOCKS_EXCLUDED(*DataGuard()) {
-    int64_t v = ReadOneWord();
-    std::memcpy(value, static_cast<const void*>(&v), sizeof(T));
-  }
-  template <typename T,
-            typename std::enable_if<flags_internal::StorageKind<T>() ==
-                                        FlagValueStorageKind::kValueAndInitBit,
-                                    int>::type = 0>
-  void Read(T* value) const ABSL_LOCKS_EXCLUDED(*DataGuard()) {
-    *value = absl::bit_cast<FlagValueAndInitBit<T>>(ReadOneWord()).value;
-  }
-
-  // Mutating access methods
-  void Write(const void* src) ABSL_LOCKS_EXCLUDED(*DataGuard());
-
-  // Interfaces to operate on callbacks.
-  void SetCallback(const FlagCallbackFunc mutation_callback)
-      ABSL_LOCKS_EXCLUDED(*DataGuard());
-  void InvokeCallback() const ABSL_EXCLUSIVE_LOCKS_REQUIRED(*DataGuard());
-
-  // Used in read/write operations to validate source/target has correct type.
-  // For example if flag is declared as absl::Flag<int> FLAGS_foo, a call to
-  // absl::GetFlag(FLAGS_foo) validates that the type of FLAGS_foo is indeed
-  // int. To do that we pass the "assumed" type id (which is deduced from type
-  // int) as an argument `type_id`, which is in turn is validated against the
-  // type id stored in flag object by flag definition statement.
-  void AssertValidType(FlagFastTypeId type_id,
-                       const std::type_info* (*gen_rtti)()) const;
-
- private:
-  template <typename T>
-  friend class Flag;
-  friend class FlagState;
-
-  // Ensures that `data_guard_` is initialized and returns it.
-  absl::Mutex* DataGuard() const
-      ABSL_LOCK_RETURNED(reinterpret_cast<absl::Mutex*>(data_guard_));
-  // Returns heap allocated value of type T initialized with default value.
-  std::unique_ptr<void, DynValueDeleter> MakeInitValue() const
-      ABSL_EXCLUSIVE_LOCKS_REQUIRED(*DataGuard());
-  // Flag initialization called via absl::call_once.
-  void Init();
-
-  // Offset value access methods. One per storage kind. These methods to not
-  // respect const correctness, so be very carefull using them.
-
-  // This is a shared helper routine which encapsulates most of the magic. Since
-  // it is only used inside the three routines below, which are defined in
-  // flag.cc, we can define it in that file as well.
-  template <typename StorageT>
-  StorageT* OffsetValue() const;
-  // This is an accessor for a value stored in an aligned buffer storage
-  // used for non-trivially-copyable data types.
-  // Returns a mutable pointer to the start of a buffer.
-  void* AlignedBufferValue() const;
-
-  // The same as above, but used for sequencelock-protected storage.
-  std::atomic<uint64_t>* AtomicBufferValue() const;
-
-  // This is an accessor for a value stored as one word atomic. Returns a
-  // mutable reference to an atomic value.
-  std::atomic<int64_t>& OneWordValue() const;
-
-  // Attempts to parse supplied `value` string. If parsing is successful,
-  // returns new value. Otherwise returns nullptr.
-  std::unique_ptr<void, DynValueDeleter> TryParse(absl::string_view value,
-                                                  std::string& err) const
-      ABSL_EXCLUSIVE_LOCKS_REQUIRED(*DataGuard());
-  // Stores the flag value based on the pointer to the source.
-  void StoreValue(const void* src) ABSL_EXCLUSIVE_LOCKS_REQUIRED(*DataGuard());
-
-  // Copy the flag data, protected by `seq_lock_` into `dst`.
-  //
-  // REQUIRES: ValueStorageKind() == kSequenceLocked.
-  void ReadSequenceLockedData(void* dst) const
-      ABSL_LOCKS_EXCLUDED(*DataGuard());
-
-  FlagHelpKind HelpSourceKind() const {
-    return static_cast<FlagHelpKind>(help_source_kind_);
-  }
-  FlagValueStorageKind ValueStorageKind() const {
-    return static_cast<FlagValueStorageKind>(value_storage_kind_);
-  }
-  FlagDefaultKind DefaultKind() const
-      ABSL_EXCLUSIVE_LOCKS_REQUIRED(*DataGuard()) {
-    return static_cast<FlagDefaultKind>(def_kind_);
-  }
-
-  // CommandLineFlag interface implementation
-  absl::string_view Name() const override;
-  std::string Filename() const override;
-  std::string Help() const override;
-  FlagFastTypeId TypeId() const override;
-  bool IsSpecifiedOnCommandLine() const override
-      ABSL_LOCKS_EXCLUDED(*DataGuard());
-  std::string DefaultValue() const override ABSL_LOCKS_EXCLUDED(*DataGuard());
-  std::string CurrentValue() const override ABSL_LOCKS_EXCLUDED(*DataGuard());
-  bool ValidateInputValue(absl::string_view value) const override
-      ABSL_LOCKS_EXCLUDED(*DataGuard());
-  void CheckDefaultValueParsingRoundtrip() const override
-      ABSL_LOCKS_EXCLUDED(*DataGuard());
-
-  int64_t ModificationCount() const ABSL_EXCLUSIVE_LOCKS_REQUIRED(*DataGuard());
-
-  // Interfaces to save and restore flags to/from persistent state.
-  // Returns current flag state or nullptr if flag does not support
-  // saving and restoring a state.
-  std::unique_ptr<FlagStateInterface> SaveState() override
-      ABSL_LOCKS_EXCLUDED(*DataGuard());
-
-  // Restores the flag state to the supplied state object. If there is
-  // nothing to restore returns false. Otherwise returns true.
-  bool RestoreState(const FlagState& flag_state)
-      ABSL_LOCKS_EXCLUDED(*DataGuard());
-
-  bool ParseFrom(absl::string_view value, FlagSettingMode set_mode,
-                 ValueSource source, std::string& error) override
-      ABSL_LOCKS_EXCLUDED(*DataGuard());
-
-  // Immutable flag's state.
-
-  // Flags name passed to ABSL_FLAG as second arg.
-  const char* const name_;
-  // The file name where ABSL_FLAG resides.
-  const char* const filename_;
-  // Type-specific operations "vtable".
-  const FlagOpFn op_;
-  // Help message literal or function to generate it.
-  const FlagHelpMsg help_;
-  // Indicates if help message was supplied as literal or generator func.
-  const uint8_t help_source_kind_ : 1;
-  // Kind of storage this flag is using for the flag's value.
-  const uint8_t value_storage_kind_ : 2;
-
-  uint8_t : 0;  // The bytes containing the const bitfields must not be
-                // shared with bytes containing the mutable bitfields.
-
-  // Mutable flag's state (guarded by `data_guard_`).
-
-  // def_kind_ is not guard by DataGuard() since it is accessed in Init without
-  // locks.
-  uint8_t def_kind_ : 2;
-  // Has this flag's value been modified?
-  bool modified_ : 1 ABSL_GUARDED_BY(*DataGuard());
-  // Has this flag been specified on command line.
-  bool on_command_line_ : 1 ABSL_GUARDED_BY(*DataGuard());
-
-  // Unique tag for absl::call_once call to initialize this flag.
-  absl::once_flag init_control_;
-
-  // Sequence lock / mutation counter.
-  flags_internal::SequenceLock seq_lock_;
-
-  // Optional flag's callback and absl::Mutex to guard the invocations.
-  FlagCallback* callback_ ABSL_GUARDED_BY(*DataGuard());
-  // Either a pointer to the function generating the default value based on the
-  // value specified in ABSL_FLAG or pointer to the dynamically set default
-  // value via SetCommandLineOptionWithMode. def_kind_ is used to distinguish
-  // these two cases.
-  FlagDefaultSrc default_value_;
-
-  // This is reserved space for an absl::Mutex to guard flag data. It will be
-  // initialized in FlagImpl::Init via placement new.
-  // We can't use "absl::Mutex data_guard_", since this class is not literal.
-  // We do not want to use "absl::Mutex* data_guard_", since this would require
-  // heap allocation during initialization, which is both slows program startup
-  // and can fail. Using reserved space + placement new allows us to avoid both
-  // problems.
-  alignas(absl::Mutex) mutable char data_guard_[sizeof(absl::Mutex)];
-};
-#if defined(__GNUC__) && !defined(__clang__)
-#pragma GCC diagnostic pop
-#endif
-
-///////////////////////////////////////////////////////////////////////////////
-// The Flag object parameterized by the flag's value type. This class implements
-// flag reflection handle interface.
-
-template <typename T>
-class Flag {
- public:
-  constexpr Flag(const char* name, const char* filename, FlagHelpArg help,
-                 const FlagDefaultArg default_arg)
-      : impl_(name, filename, &FlagOps<T>, help,
-              flags_internal::StorageKind<T>(), default_arg),
-        value_() {}
-
-  // CommandLineFlag interface
-  absl::string_view Name() const { return impl_.Name(); }
-  std::string Filename() const { return impl_.Filename(); }
-  std::string Help() const { return impl_.Help(); }
-  // Do not use. To be removed.
-  bool IsSpecifiedOnCommandLine() const {
-    return impl_.IsSpecifiedOnCommandLine();
-  }
-  std::string DefaultValue() const { return impl_.DefaultValue(); }
-  std::string CurrentValue() const { return impl_.CurrentValue(); }
-
- private:
-  template <typename, bool>
-  friend class FlagRegistrar;
-  friend class FlagImplPeer;
-
-  T Get() const {
-    // See implementation notes in CommandLineFlag::Get().
-    union U {
-      T value;
-      U() {}
-      ~U() { value.~T(); }
-    };
-    U u;
-
-#if !defined(NDEBUG)
-    impl_.AssertValidType(base_internal::FastTypeId<T>(), &GenRuntimeTypeId<T>);
-#endif
-
-    if (ABSL_PREDICT_FALSE(!value_.Get(impl_.seq_lock_, u.value))) {
-      impl_.Read(&u.value);
-    }
-    return std::move(u.value);
-  }
-  void Set(const T& v) {
-    impl_.AssertValidType(base_internal::FastTypeId<T>(), &GenRuntimeTypeId<T>);
-    impl_.Write(&v);
-  }
-
-  // Access to the reflection.
-  const CommandLineFlag& Reflect() const { return impl_; }
-
-  // Flag's data
-  // The implementation depends on value_ field to be placed exactly after the
-  // impl_ field, so that impl_ can figure out the offset to the value and
-  // access it.
-  FlagImpl impl_;
-  FlagValue<T> value_;
-};
-
-///////////////////////////////////////////////////////////////////////////////
-// Trampoline for friend access
-
-class FlagImplPeer {
- public:
-  template <typename T, typename FlagType>
-  static T InvokeGet(const FlagType& flag) {
-    return flag.Get();
-  }
-  template <typename FlagType, typename T>
-  static void InvokeSet(FlagType& flag, const T& v) {
-    flag.Set(v);
-  }
-  template <typename FlagType>
-  static const CommandLineFlag& InvokeReflect(const FlagType& f) {
-    return f.Reflect();
-  }
-};
-
-///////////////////////////////////////////////////////////////////////////////
-// Implementation of Flag value specific operations routine.
-template <typename T>
-void* FlagOps(FlagOp op, const void* v1, void* v2, void* v3) {
-  switch (op) {
-    case FlagOp::kAlloc: {
-      std::allocator<T> alloc;
-      return std::allocator_traits<std::allocator<T>>::allocate(alloc, 1);
-    }
-    case FlagOp::kDelete: {
-      T* p = static_cast<T*>(v2);
-      p->~T();
-      std::allocator<T> alloc;
-      std::allocator_traits<std::allocator<T>>::deallocate(alloc, p, 1);
-      return nullptr;
-    }
-    case FlagOp::kCopy:
-      *static_cast<T*>(v2) = *static_cast<const T*>(v1);
-      return nullptr;
-    case FlagOp::kCopyConstruct:
-      new (v2) T(*static_cast<const T*>(v1));
-      return nullptr;
-    case FlagOp::kSizeof:
-      return reinterpret_cast<void*>(static_cast<uintptr_t>(sizeof(T)));
-    case FlagOp::kFastTypeId:
-      return const_cast<void*>(base_internal::FastTypeId<T>());
-    case FlagOp::kRuntimeTypeId:
-      return const_cast<std::type_info*>(GenRuntimeTypeId<T>());
-    case FlagOp::kParse: {
-      // Initialize the temporary instance of type T based on current value in
-      // destination (which is going to be flag's default value).
-      T temp(*static_cast<T*>(v2));
-      if (!absl::ParseFlag<T>(*static_cast<const absl::string_view*>(v1), &temp,
-                              static_cast<std::string*>(v3))) {
-        return nullptr;
-      }
-      *static_cast<T*>(v2) = std::move(temp);
-      return v2;
-    }
-    case FlagOp::kUnparse:
-      *static_cast<std::string*>(v2) =
-          absl::UnparseFlag<T>(*static_cast<const T*>(v1));
-      return nullptr;
-    case FlagOp::kValueOffset: {
-      // Round sizeof(FlagImp) to a multiple of alignof(FlagValue<T>) to get the
-      // offset of the data.
-      size_t round_to = alignof(FlagValue<T>);
-      size_t offset = (sizeof(FlagImpl) + round_to - 1) / round_to * round_to;
-      return reinterpret_cast<void*>(offset);
-    }
-  }
-  return nullptr;
-}
-
-///////////////////////////////////////////////////////////////////////////////
-// This class facilitates Flag object registration and tail expression-based
-// flag definition, for example:
-// ABSL_FLAG(int, foo, 42, "Foo help").OnUpdate(NotifyFooWatcher);
-struct FlagRegistrarEmpty {};
-template <typename T, bool do_register>
-class FlagRegistrar {
- public:
-  constexpr explicit FlagRegistrar(Flag<T>& flag, const char* filename)
-      : flag_(flag) {
-    if (do_register)
-      flags_internal::RegisterCommandLineFlag(flag_.impl_, filename);
-  }
-
-  FlagRegistrar OnUpdate(FlagCallbackFunc cb) && {
-    flag_.impl_.SetCallback(cb);
-    return *this;
-  }
-
-  // Make the registrar "die" gracefully as an empty struct on a line where
-  // registration happens. Registrar objects are intended to live only as
-  // temporary.
-  constexpr operator FlagRegistrarEmpty() const { return {}; }  // NOLINT
-
- private:
-  Flag<T>& flag_;  // Flag being registered (not owned).
-};
-
-}  // namespace flags_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FLAGS_INTERNAL_FLAG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/parse.h b/projects/llm_framework/include/abseil-cpp/absl/flags/internal/parse.h
deleted file mode 100644
index 10c531b8..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/parse.h
+++ /dev/null
@@ -1,70 +0,0 @@
-//
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_FLAGS_INTERNAL_PARSE_H_
-#define ABSL_FLAGS_INTERNAL_PARSE_H_
-
-#include <iostream>
-#include <ostream>
-#include <string>
-#include <vector>
-
-#include "absl/base/config.h"
-#include "absl/flags/declare.h"
-#include "absl/flags/internal/usage.h"
-#include "absl/strings/string_view.h"
-
-ABSL_DECLARE_FLAG(std::vector<std::string>, flagfile);
-ABSL_DECLARE_FLAG(std::vector<std::string>, fromenv);
-ABSL_DECLARE_FLAG(std::vector<std::string>, tryfromenv);
-ABSL_DECLARE_FLAG(std::vector<std::string>, undefok);
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace flags_internal {
-
-enum class UsageFlagsAction { kHandleUsage, kIgnoreUsage };
-enum class OnUndefinedFlag {
-  kIgnoreUndefined,
-  kReportUndefined,
-  kAbortIfUndefined
-};
-
-// This is not a public interface. This interface exists to expose the ability
-// to change help output stream in case of parsing errors. This is used by
-// internal unit tests to validate expected outputs.
-// When this was written, `EXPECT_EXIT` only supported matchers on stderr,
-// but not on stdout.
-std::vector<char*> ParseCommandLineImpl(
-    int argc, char* argv[], UsageFlagsAction usage_flag_action,
-    OnUndefinedFlag undef_flag_action,
-    std::ostream& error_help_output = std::cout);
-
-// --------------------------------------------------------------------
-// Inspect original command line
-
-// Returns true if flag with specified name was either present on the original
-// command line or specified in flag file present on the original command line.
-bool WasPresentOnCommandLine(absl::string_view flag_name);
-
-// Return existing flags similar to the parameter, in order to help in case of
-// misspellings.
-std::vector<std::string> GetMisspellingHints(absl::string_view flag);
-
-}  // namespace flags_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FLAGS_INTERNAL_PARSE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/path_util.h b/projects/llm_framework/include/abseil-cpp/absl/flags/internal/path_util.h
deleted file mode 100644
index a6594d33..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/path_util.h
+++ /dev/null
@@ -1,62 +0,0 @@
-//
-//  Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_FLAGS_INTERNAL_PATH_UTIL_H_
-#define ABSL_FLAGS_INTERNAL_PATH_UTIL_H_
-
-#include "absl/base/config.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace flags_internal {
-
-// A portable interface that returns the basename of the filename passed as an
-// argument. It is similar to basename(3)
-// <https://linux.die.net/man/3/basename>.
-// For example:
-//     flags_internal::Basename("a/b/prog/file.cc")
-// returns "file.cc"
-//     flags_internal::Basename("file.cc")
-// returns "file.cc"
-inline absl::string_view Basename(absl::string_view filename) {
-  auto last_slash_pos = filename.find_last_of("/\\");
-
-  return last_slash_pos == absl::string_view::npos
-             ? filename
-             : filename.substr(last_slash_pos + 1);
-}
-
-// A portable interface that returns the directory name of the filename
-// passed as an argument, including the trailing slash.
-// Returns the empty string if a slash is not found in the input file name.
-// For example:
-//      flags_internal::Package("a/b/prog/file.cc")
-// returns "a/b/prog/"
-//      flags_internal::Package("file.cc")
-// returns ""
-inline absl::string_view Package(absl::string_view filename) {
-  auto last_slash_pos = filename.find_last_of("/\\");
-
-  return last_slash_pos == absl::string_view::npos
-             ? absl::string_view()
-             : filename.substr(0, last_slash_pos + 1);
-}
-
-}  // namespace flags_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FLAGS_INTERNAL_PATH_UTIL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/private_handle_accessor.h b/projects/llm_framework/include/abseil-cpp/absl/flags/internal/private_handle_accessor.h
deleted file mode 100644
index c64435cd..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/private_handle_accessor.h
+++ /dev/null
@@ -1,61 +0,0 @@
-//
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_FLAGS_INTERNAL_PRIVATE_HANDLE_ACCESSOR_H_
-#define ABSL_FLAGS_INTERNAL_PRIVATE_HANDLE_ACCESSOR_H_
-
-#include <memory>
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/flags/commandlineflag.h"
-#include "absl/flags/internal/commandlineflag.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace flags_internal {
-
-// This class serves as a trampoline to access private methods of
-// CommandLineFlag. This class is intended for use exclusively internally inside
-// of the Abseil Flags implementation.
-class PrivateHandleAccessor {
- public:
-  // Access to CommandLineFlag::TypeId.
-  static FlagFastTypeId TypeId(const CommandLineFlag& flag);
-
-  // Access to CommandLineFlag::SaveState.
-  static std::unique_ptr<FlagStateInterface> SaveState(CommandLineFlag& flag);
-
-  // Access to CommandLineFlag::IsSpecifiedOnCommandLine.
-  static bool IsSpecifiedOnCommandLine(const CommandLineFlag& flag);
-
-  // Access to CommandLineFlag::ValidateInputValue.
-  static bool ValidateInputValue(const CommandLineFlag& flag,
-                                 absl::string_view value);
-
-  // Access to CommandLineFlag::CheckDefaultValueParsingRoundtrip.
-  static void CheckDefaultValueParsingRoundtrip(const CommandLineFlag& flag);
-
-  static bool ParseFrom(CommandLineFlag& flag, absl::string_view value,
-                        flags_internal::FlagSettingMode set_mode,
-                        flags_internal::ValueSource source, std::string& error);
-};
-
-}  // namespace flags_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FLAGS_INTERNAL_PRIVATE_HANDLE_ACCESSOR_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/program_name.h b/projects/llm_framework/include/abseil-cpp/absl/flags/internal/program_name.h
deleted file mode 100644
index b99b94fe..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/program_name.h
+++ /dev/null
@@ -1,50 +0,0 @@
-//
-//  Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_FLAGS_INTERNAL_PROGRAM_NAME_H_
-#define ABSL_FLAGS_INTERNAL_PROGRAM_NAME_H_
-
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/strings/string_view.h"
-
-// --------------------------------------------------------------------
-// Program name
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace flags_internal {
-
-// Returns program invocation name or "UNKNOWN" if `SetProgramInvocationName()`
-// is never called. At the moment this is always set to argv[0] as part of
-// library initialization.
-std::string ProgramInvocationName();
-
-// Returns base name for program invocation name. For example, if
-//   ProgramInvocationName() == "a/b/mybinary"
-// then
-//   ShortProgramInvocationName() == "mybinary"
-std::string ShortProgramInvocationName();
-
-// Sets program invocation name to a new value. Should only be called once
-// during program initialization, before any threads are spawned.
-void SetProgramInvocationName(absl::string_view prog_name_str);
-
-}  // namespace flags_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FLAGS_INTERNAL_PROGRAM_NAME_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/registry.h b/projects/llm_framework/include/abseil-cpp/absl/flags/internal/registry.h
deleted file mode 100644
index 4b68c85f..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/registry.h
+++ /dev/null
@@ -1,97 +0,0 @@
-//
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_FLAGS_INTERNAL_REGISTRY_H_
-#define ABSL_FLAGS_INTERNAL_REGISTRY_H_
-
-#include <functional>
-
-#include "absl/base/config.h"
-#include "absl/flags/commandlineflag.h"
-#include "absl/flags/internal/commandlineflag.h"
-#include "absl/strings/string_view.h"
-
-// --------------------------------------------------------------------
-// Global flags registry API.
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace flags_internal {
-
-// Executes specified visitor for each non-retired flag in the registry. While
-// callback are executed, the registry is locked and can't be changed.
-void ForEachFlag(std::function<void(CommandLineFlag&)> visitor);
-
-//-----------------------------------------------------------------------------
-
-bool RegisterCommandLineFlag(CommandLineFlag&, const char* filename);
-
-void FinalizeRegistry();
-
-//-----------------------------------------------------------------------------
-// Retired registrations:
-//
-// Retired flag registrations are treated specially. A 'retired' flag is
-// provided only for compatibility with automated invocations that still
-// name it.  A 'retired' flag:
-//   - is not bound to a C++ FLAGS_ reference.
-//   - has a type and a value, but that value is intentionally inaccessible.
-//   - does not appear in --help messages.
-//   - is fully supported by _all_ flag parsing routines.
-//   - consumes args normally, and complains about type mismatches in its
-//     argument.
-//   - emits a complaint but does not die (e.g. LOG(ERROR)) if it is
-//     accessed by name through the flags API for parsing or otherwise.
-//
-// The registrations for a flag happen in an unspecified order as the
-// initializers for the namespace-scope objects of a program are run.
-// Any number of weak registrations for a flag can weakly define the flag.
-// One non-weak registration will upgrade the flag from weak to non-weak.
-// Further weak registrations of a non-weak flag are ignored.
-//
-// This mechanism is designed to support moving dead flags into a
-// 'graveyard' library.  An example migration:
-//
-//   0: Remove references to this FLAGS_flagname in the C++ codebase.
-//   1: Register as 'retired' in old_lib.
-//   2: Make old_lib depend on graveyard.
-//   3: Add a redundant 'retired' registration to graveyard.
-//   4: Remove the old_lib 'retired' registration.
-//   5: Eventually delete the graveyard registration entirely.
-//
-
-// Retire flag with name "name" and type indicated by ops.
-void Retire(const char* name, FlagFastTypeId type_id, char* buf);
-
-constexpr size_t kRetiredFlagObjSize = 3 * sizeof(void*);
-constexpr size_t kRetiredFlagObjAlignment = alignof(void*);
-
-// Registered a retired flag with name 'flag_name' and type 'T'.
-template <typename T>
-class RetiredFlag {
- public:
-  void Retire(const char* flag_name) {
-    flags_internal::Retire(flag_name, base_internal::FastTypeId<T>(), buf_);
-  }
-
- private:
-  alignas(kRetiredFlagObjAlignment) char buf_[kRetiredFlagObjSize];
-};
-
-}  // namespace flags_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FLAGS_INTERNAL_REGISTRY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/sequence_lock.h b/projects/llm_framework/include/abseil-cpp/absl/flags/internal/sequence_lock.h
deleted file mode 100644
index 36318ab9..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/sequence_lock.h
+++ /dev/null
@@ -1,187 +0,0 @@
-//
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_FLAGS_INTERNAL_SEQUENCE_LOCK_H_
-#define ABSL_FLAGS_INTERNAL_SEQUENCE_LOCK_H_
-
-#include <stddef.h>
-#include <stdint.h>
-
-#include <atomic>
-#include <cassert>
-#include <cstring>
-
-#include "absl/base/optimization.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace flags_internal {
-
-// Align 'x' up to the nearest 'align' bytes.
-inline constexpr size_t AlignUp(size_t x, size_t align) {
-  return align * ((x + align - 1) / align);
-}
-
-// A SequenceLock implements lock-free reads. A sequence counter is incremented
-// before and after each write, and readers access the counter before and after
-// accessing the protected data. If the counter is verified to not change during
-// the access, and the sequence counter value was even, then the reader knows
-// that the read was race-free and valid. Otherwise, the reader must fall back
-// to a Mutex-based code path.
-//
-// This particular SequenceLock starts in an "uninitialized" state in which
-// TryRead() returns false. It must be enabled by calling MarkInitialized().
-// This serves as a marker that the associated flag value has not yet been
-// initialized and a slow path needs to be taken.
-//
-// The memory reads and writes protected by this lock must use the provided
-// `TryRead()` and `Write()` functions. These functions behave similarly to
-// `memcpy()`, with one oddity: the protected data must be an array of
-// `std::atomic<uint64>`. This is to comply with the C++ standard, which
-// considers data races on non-atomic objects to be undefined behavior. See "Can
-// Seqlocks Get Along With Programming Language Memory Models?"[1] by Hans J.
-// Boehm for more details.
-//
-// [1] https://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf
-class SequenceLock {
- public:
-  constexpr SequenceLock() : lock_(kUninitialized) {}
-
-  // Mark that this lock is ready for use.
-  void MarkInitialized() {
-    assert(lock_.load(std::memory_order_relaxed) == kUninitialized);
-    lock_.store(0, std::memory_order_release);
-  }
-
-  // Copy "size" bytes of data from "src" to "dst", protected as a read-side
-  // critical section of the sequence lock.
-  //
-  // Unlike traditional sequence lock implementations which loop until getting a
-  // clean read, this implementation returns false in the case of concurrent
-  // calls to `Write`. In such a case, the caller should fall back to a
-  // locking-based slow path.
-  //
-  // Returns false if the sequence lock was not yet marked as initialized.
-  //
-  // NOTE: If this returns false, "dst" may be overwritten with undefined
-  // (potentially uninitialized) data.
-  bool TryRead(void* dst, const std::atomic<uint64_t>* src, size_t size) const {
-    // Acquire barrier ensures that no loads done by f() are reordered
-    // above the first load of the sequence counter.
-    int64_t seq_before = lock_.load(std::memory_order_acquire);
-    if (ABSL_PREDICT_FALSE(seq_before & 1) == 1) return false;
-    RelaxedCopyFromAtomic(dst, src, size);
-    // Another acquire fence ensures that the load of 'lock_' below is
-    // strictly ordered after the RelaxedCopyToAtomic call above.
-    std::atomic_thread_fence(std::memory_order_acquire);
-    int64_t seq_after = lock_.load(std::memory_order_relaxed);
-    return ABSL_PREDICT_TRUE(seq_before == seq_after);
-  }
-
-  // Copy "size" bytes from "src" to "dst" as a write-side critical section
-  // of the sequence lock. Any concurrent readers will be forced to retry
-  // until they get a read that does not conflict with this write.
-  //
-  // This call must be externally synchronized against other calls to Write,
-  // but may proceed concurrently with reads.
-  void Write(std::atomic<uint64_t>* dst, const void* src, size_t size) {
-    // We can use relaxed instructions to increment the counter since we
-    // are extenally synchronized. The std::atomic_thread_fence below
-    // ensures that the counter updates don't get interleaved with the
-    // copy to the data.
-    int64_t orig_seq = lock_.load(std::memory_order_relaxed);
-    assert((orig_seq & 1) == 0);  // Must be initially unlocked.
-    lock_.store(orig_seq + 1, std::memory_order_relaxed);
-
-    // We put a release fence between update to lock_ and writes to shared data.
-    // Thus all stores to shared data are effectively release operations and
-    // update to lock_ above cannot be re-ordered past any of them. Note that
-    // this barrier is not for the fetch_add above.  A release barrier for the
-    // fetch_add would be before it, not after.
-    std::atomic_thread_fence(std::memory_order_release);
-    RelaxedCopyToAtomic(dst, src, size);
-    // "Release" semantics ensure that none of the writes done by
-    // RelaxedCopyToAtomic() can be reordered after the following modification.
-    lock_.store(orig_seq + 2, std::memory_order_release);
-  }
-
-  // Return the number of times that Write() has been called.
-  //
-  // REQUIRES: This must be externally synchronized against concurrent calls to
-  // `Write()` or `IncrementModificationCount()`.
-  // REQUIRES: `MarkInitialized()` must have been previously called.
-  int64_t ModificationCount() const {
-    int64_t val = lock_.load(std::memory_order_relaxed);
-    assert(val != kUninitialized && (val & 1) == 0);
-    return val / 2;
-  }
-
-  // REQUIRES: This must be externally synchronized against concurrent calls to
-  // `Write()` or `ModificationCount()`.
-  // REQUIRES: `MarkInitialized()` must have been previously called.
-  void IncrementModificationCount() {
-    int64_t val = lock_.load(std::memory_order_relaxed);
-    assert(val != kUninitialized);
-    lock_.store(val + 2, std::memory_order_relaxed);
-  }
-
- private:
-  // Perform the equivalent of "memcpy(dst, src, size)", but using relaxed
-  // atomics.
-  static void RelaxedCopyFromAtomic(void* dst, const std::atomic<uint64_t>* src,
-                                    size_t size) {
-    char* dst_byte = static_cast<char*>(dst);
-    while (size >= sizeof(uint64_t)) {
-      uint64_t word = src->load(std::memory_order_relaxed);
-      std::memcpy(dst_byte, &word, sizeof(word));
-      dst_byte += sizeof(word);
-      src++;
-      size -= sizeof(word);
-    }
-    if (size > 0) {
-      uint64_t word = src->load(std::memory_order_relaxed);
-      std::memcpy(dst_byte, &word, size);
-    }
-  }
-
-  // Perform the equivalent of "memcpy(dst, src, size)", but using relaxed
-  // atomics.
-  static void RelaxedCopyToAtomic(std::atomic<uint64_t>* dst, const void* src,
-                                  size_t size) {
-    const char* src_byte = static_cast<const char*>(src);
-    while (size >= sizeof(uint64_t)) {
-      uint64_t word;
-      std::memcpy(&word, src_byte, sizeof(word));
-      dst->store(word, std::memory_order_relaxed);
-      src_byte += sizeof(word);
-      dst++;
-      size -= sizeof(word);
-    }
-    if (size > 0) {
-      uint64_t word = 0;
-      std::memcpy(&word, src_byte, size);
-      dst->store(word, std::memory_order_relaxed);
-    }
-  }
-
-  static constexpr int64_t kUninitialized = -1;
-  std::atomic<int64_t> lock_;
-};
-
-}  // namespace flags_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FLAGS_INTERNAL_SEQUENCE_LOCK_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/usage.h b/projects/llm_framework/include/abseil-cpp/absl/flags/internal/usage.h
deleted file mode 100644
index a96cbf38..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/internal/usage.h
+++ /dev/null
@@ -1,106 +0,0 @@
-//
-//  Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_FLAGS_INTERNAL_USAGE_H_
-#define ABSL_FLAGS_INTERNAL_USAGE_H_
-
-#include <iosfwd>
-#include <ostream>
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/flags/commandlineflag.h"
-#include "absl/strings/string_view.h"
-
-// --------------------------------------------------------------------
-// Usage reporting interfaces
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace flags_internal {
-
-// The format to report the help messages in.
-enum class HelpFormat {
-  kHumanReadable,
-};
-
-// The kind of usage help requested.
-enum class HelpMode {
-  kNone,
-  kImportant,
-  kShort,
-  kFull,
-  kPackage,
-  kMatch,
-  kVersion,
-  kOnlyCheckArgs
-};
-
-// Streams the help message describing `flag` to `out`.
-// The default value for `flag` is included in the output.
-void FlagHelp(std::ostream& out, const CommandLineFlag& flag,
-              HelpFormat format = HelpFormat::kHumanReadable);
-
-// Produces the help messages for all flags matching the filter. A flag matches
-// the filter if it is defined in a file with a filename which includes
-// filter string as a substring. You can use '/' and '.' to restrict the
-// matching to a specific file names. For example:
-//   FlagsHelp(out, "/path/to/file.");
-// restricts help to only flags which resides in files named like:
-//  .../path/to/file.<ext>
-// for any extension 'ext'. If the filter is empty this function produces help
-// messages for all flags.
-void FlagsHelp(std::ostream& out, absl::string_view filter,
-               HelpFormat format, absl::string_view program_usage_message);
-
-// --------------------------------------------------------------------
-
-// If any of the 'usage' related command line flags (listed on the bottom of
-// this file) has been set this routine produces corresponding help message in
-// the specified output stream and returns HelpMode that was handled. Otherwise
-// it returns HelpMode::kNone.
-HelpMode HandleUsageFlags(std::ostream& out,
-                          absl::string_view program_usage_message);
-
-// --------------------------------------------------------------------
-// Encapsulates the logic of exiting the binary depending on handled help mode.
-
-void MaybeExit(HelpMode mode);
-
-// --------------------------------------------------------------------
-// Globals representing usage reporting flags
-
-// Returns substring to filter help output (--help=substr argument)
-std::string GetFlagsHelpMatchSubstr();
-// Returns the requested help mode.
-HelpMode GetFlagsHelpMode();
-// Returns the requested help format.
-HelpFormat GetFlagsHelpFormat();
-
-// These are corresponding setters to the attributes above.
-void SetFlagsHelpMatchSubstr(absl::string_view);
-void SetFlagsHelpMode(HelpMode);
-void SetFlagsHelpFormat(HelpFormat);
-
-// Deduces usage flags from the input argument in a form --name=value or
-// --name. argument is already split into name and value before we call this
-// function.
-bool DeduceUsageFlags(absl::string_view name, absl::string_view value);
-
-}  // namespace flags_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FLAGS_INTERNAL_USAGE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/marshalling.h b/projects/llm_framework/include/abseil-cpp/absl/flags/marshalling.h
deleted file mode 100644
index 301213a9..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/marshalling.h
+++ /dev/null
@@ -1,361 +0,0 @@
-//
-//  Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: marshalling.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines the API for extending Abseil flag support to
-// custom types, and defines the set of overloads for fundamental types.
-//
-// Out of the box, the Abseil flags library supports the following types:
-//
-// * `bool`
-// * `int16_t`
-// * `uint16_t`
-// * `int32_t`
-// * `uint32_t`
-// * `int64_t`
-// * `uint64_t`
-// * `float`
-// * `double`
-// * `std::string`
-// * `std::vector<std::string>`
-// * `std::optional<T>`
-// * `absl::LogSeverity` (provided natively for layering reasons)
-//
-// Note that support for integral types is implemented using overloads for
-// variable-width fundamental types (`short`, `int`, `long`, etc.). However,
-// you should prefer the fixed-width integral types (`int32_t`, `uint64_t`,
-// etc.) we've noted above within flag definitions.
-//
-// In addition, several Abseil libraries provide their own custom support for
-// Abseil flags. Documentation for these formats is provided in the type's
-// `AbslParseFlag()` definition.
-//
-// The Abseil time library provides the following support for civil time values:
-//
-// * `absl::CivilSecond`
-// * `absl::CivilMinute`
-// * `absl::CivilHour`
-// * `absl::CivilDay`
-// * `absl::CivilMonth`
-// * `absl::CivilYear`
-//
-// and also provides support for the following absolute time values:
-//
-// * `absl::Duration`
-// * `absl::Time`
-//
-// Additional support for Abseil types will be noted here as it is added.
-//
-// You can also provide your own custom flags by adding overloads for
-// `AbslParseFlag()` and `AbslUnparseFlag()` to your type definitions. (See
-// below.)
-//
-// -----------------------------------------------------------------------------
-// Optional Flags
-// -----------------------------------------------------------------------------
-//
-// The Abseil flags library supports flags of type `std::optional<T>` where
-// `T` is a type of one of the supported flags. We refer to this flag type as
-// an "optional flag." An optional flag is either "valueless", holding no value
-// of type `T` (indicating that the flag has not been set) or a value of type
-// `T`. The valueless state in C++ code is represented by a value of
-// `std::nullopt` for the optional flag.
-//
-// Using `std::nullopt` as an optional flag's default value allows you to check
-// whether such a flag was ever specified on the command line:
-//
-//   if (absl::GetFlag(FLAGS_foo).has_value()) {
-//     // flag was set on command line
-//   } else {
-//     // flag was not passed on command line
-//   }
-//
-// Using an optional flag in this manner avoids common workarounds for
-// indicating such an unset flag (such as using sentinel values to indicate this
-// state).
-//
-// An optional flag also allows a developer to pass a flag in an "unset"
-// valueless state on the command line, allowing the flag to later be set in
-// binary logic. An optional flag's valueless state is indicated by the special
-// notation of passing the value as an empty string through the syntax `--flag=`
-// or `--flag ""`.
-//
-//   $ binary_with_optional --flag_in_unset_state=
-//   $ binary_with_optional --flag_in_unset_state ""
-//
-// Note: as a result of the above syntax requirements, an optional flag cannot
-// be set to a `T` of any value which unparses to the empty string.
-//
-// -----------------------------------------------------------------------------
-// Adding Type Support for Abseil Flags
-// -----------------------------------------------------------------------------
-//
-// To add support for your user-defined type, add overloads of `AbslParseFlag()`
-// and `AbslUnparseFlag()` as free (non-member) functions to your type. If `T`
-// is a class type, these functions can be friend function definitions. These
-// overloads must be added to the same namespace where the type is defined, so
-// that they can be discovered by Argument-Dependent Lookup (ADL).
-//
-// Example:
-//
-//   namespace foo {
-//
-//   enum OutputMode { kPlainText, kHtml };
-//
-//   // AbslParseFlag converts from a string to OutputMode.
-//   // Must be in same namespace as OutputMode.
-//
-//   // Parses an OutputMode from the command line flag value `text`. Returns
-//   // `true` and sets `*mode` on success; returns `false` and sets `*error`
-//   // on failure.
-//   bool AbslParseFlag(absl::string_view text,
-//                      OutputMode* mode,
-//                      std::string* error) {
-//     if (text == "plaintext") {
-//       *mode = kPlainText;
-//       return true;
-//     }
-//     if (text == "html") {
-//       *mode = kHtml;
-//      return true;
-//     }
-//     *error = "unknown value for enumeration";
-//     return false;
-//  }
-//
-//  // AbslUnparseFlag converts from an OutputMode to a string.
-//  // Must be in same namespace as OutputMode.
-//
-//  // Returns a textual flag value corresponding to the OutputMode `mode`.
-//  std::string AbslUnparseFlag(OutputMode mode) {
-//    switch (mode) {
-//      case kPlainText: return "plaintext";
-//      case kHtml: return "html";
-//    }
-//    return absl::StrCat(mode);
-//  }
-//
-// Notice that neither `AbslParseFlag()` nor `AbslUnparseFlag()` are class
-// members, but free functions. `AbslParseFlag/AbslUnparseFlag()` overloads
-// for a type should only be declared in the same file and namespace as said
-// type. The proper `AbslParseFlag/AbslUnparseFlag()` implementations for a
-// given type will be discovered via Argument-Dependent Lookup (ADL).
-//
-// `AbslParseFlag()` may need, in turn, to parse simpler constituent types
-// using `absl::ParseFlag()`. For example, a custom struct `MyFlagType`
-// consisting of a `std::pair<int, std::string>` would add an `AbslParseFlag()`
-// overload for its `MyFlagType` like so:
-//
-// Example:
-//
-//   namespace my_flag_type {
-//
-//   struct MyFlagType {
-//     std::pair<int, std::string> my_flag_data;
-//   };
-//
-//   bool AbslParseFlag(absl::string_view text, MyFlagType* flag,
-//                      std::string* err);
-//
-//   std::string AbslUnparseFlag(const MyFlagType&);
-//
-//   // Within the implementation, `AbslParseFlag()` will, in turn invoke
-//   // `absl::ParseFlag()` on its constituent `int` and `std::string` types
-//   // (which have built-in Abseil flag support).
-//
-//   bool AbslParseFlag(absl::string_view text, MyFlagType* flag,
-//                      std::string* err) {
-//     std::pair<absl::string_view, absl::string_view> tokens =
-//         absl::StrSplit(text, ',');
-//     if (!absl::ParseFlag(tokens.first, &flag->my_flag_data.first, err))
-//         return false;
-//     if (!absl::ParseFlag(tokens.second, &flag->my_flag_data.second, err))
-//         return false;
-//     return true;
-//   }
-//
-//   // Similarly, for unparsing, we can simply invoke `absl::UnparseFlag()` on
-//   // the constituent types.
-//   std::string AbslUnparseFlag(const MyFlagType& flag) {
-//     return absl::StrCat(absl::UnparseFlag(flag.my_flag_data.first),
-//                         ",",
-//                         absl::UnparseFlag(flag.my_flag_data.second));
-//   }
-#ifndef ABSL_FLAGS_MARSHALLING_H_
-#define ABSL_FLAGS_MARSHALLING_H_
-
-#include "absl/base/config.h"
-#include "absl/numeric/int128.h"
-
-#if defined(ABSL_HAVE_STD_OPTIONAL) && !defined(ABSL_USES_STD_OPTIONAL)
-#include <optional>
-#endif
-#include <string>
-#include <vector>
-
-#include "absl/strings/string_view.h"
-#include "absl/types/optional.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// Forward declaration to be used inside composable flag parse/unparse
-// implementations
-template <typename T>
-inline bool ParseFlag(absl::string_view input, T* dst, std::string* error);
-template <typename T>
-inline std::string UnparseFlag(const T& v);
-
-namespace flags_internal {
-
-// Overloads of `AbslParseFlag()` and `AbslUnparseFlag()` for fundamental types.
-bool AbslParseFlag(absl::string_view, bool*, std::string*);
-bool AbslParseFlag(absl::string_view, short*, std::string*);           // NOLINT
-bool AbslParseFlag(absl::string_view, unsigned short*, std::string*);  // NOLINT
-bool AbslParseFlag(absl::string_view, int*, std::string*);             // NOLINT
-bool AbslParseFlag(absl::string_view, unsigned int*, std::string*);    // NOLINT
-bool AbslParseFlag(absl::string_view, long*, std::string*);            // NOLINT
-bool AbslParseFlag(absl::string_view, unsigned long*, std::string*);   // NOLINT
-bool AbslParseFlag(absl::string_view, long long*, std::string*);       // NOLINT
-bool AbslParseFlag(absl::string_view, unsigned long long*,             // NOLINT
-                   std::string*);
-bool AbslParseFlag(absl::string_view, absl::int128*, std::string*);    // NOLINT
-bool AbslParseFlag(absl::string_view, absl::uint128*, std::string*);   // NOLINT
-bool AbslParseFlag(absl::string_view, float*, std::string*);
-bool AbslParseFlag(absl::string_view, double*, std::string*);
-bool AbslParseFlag(absl::string_view, std::string*, std::string*);
-bool AbslParseFlag(absl::string_view, std::vector<std::string>*, std::string*);
-
-template <typename T>
-bool AbslParseFlag(absl::string_view text, absl::optional<T>* f,
-                   std::string* err) {
-  if (text.empty()) {
-    *f = absl::nullopt;
-    return true;
-  }
-  T value;
-  if (!absl::ParseFlag(text, &value, err)) return false;
-
-  *f = std::move(value);
-  return true;
-}
-
-#if defined(ABSL_HAVE_STD_OPTIONAL) && !defined(ABSL_USES_STD_OPTIONAL)
-template <typename T>
-bool AbslParseFlag(absl::string_view text, std::optional<T>* f,
-                   std::string* err) {
-  if (text.empty()) {
-    *f = std::nullopt;
-    return true;
-  }
-  T value;
-  if (!absl::ParseFlag(text, &value, err)) return false;
-
-  *f = std::move(value);
-  return true;
-}
-#endif
-
-template <typename T>
-bool InvokeParseFlag(absl::string_view input, T* dst, std::string* err) {
-  // Comment on next line provides a good compiler error message if T
-  // does not have AbslParseFlag(absl::string_view, T*, std::string*).
-  return AbslParseFlag(input, dst, err);  // Is T missing AbslParseFlag?
-}
-
-// Strings and std:: containers do not have the same overload resolution
-// considerations as fundamental types. Naming these 'AbslUnparseFlag' means we
-// can avoid the need for additional specializations of Unparse (below).
-std::string AbslUnparseFlag(absl::string_view v);
-std::string AbslUnparseFlag(const std::vector<std::string>&);
-
-template <typename T>
-std::string AbslUnparseFlag(const absl::optional<T>& f) {
-  return f.has_value() ? absl::UnparseFlag(*f) : "";
-}
-
-#if defined(ABSL_HAVE_STD_OPTIONAL) && !defined(ABSL_USES_STD_OPTIONAL)
-template <typename T>
-std::string AbslUnparseFlag(const std::optional<T>& f) {
-  return f.has_value() ? absl::UnparseFlag(*f) : "";
-}
-#endif
-
-template <typename T>
-std::string Unparse(const T& v) {
-  // Comment on next line provides a good compiler error message if T does not
-  // have UnparseFlag.
-  return AbslUnparseFlag(v);  // Is T missing AbslUnparseFlag?
-}
-
-// Overloads for builtin types.
-std::string Unparse(bool v);
-std::string Unparse(short v);               // NOLINT
-std::string Unparse(unsigned short v);      // NOLINT
-std::string Unparse(int v);                 // NOLINT
-std::string Unparse(unsigned int v);        // NOLINT
-std::string Unparse(long v);                // NOLINT
-std::string Unparse(unsigned long v);       // NOLINT
-std::string Unparse(long long v);           // NOLINT
-std::string Unparse(unsigned long long v);  // NOLINT
-std::string Unparse(absl::int128 v);
-std::string Unparse(absl::uint128 v);
-std::string Unparse(float v);
-std::string Unparse(double v);
-
-}  // namespace flags_internal
-
-// ParseFlag()
-//
-// Parses a string value into a flag value of type `T`. Do not add overloads of
-// this function for your type directly; instead, add an `AbslParseFlag()`
-// free function as documented above.
-//
-// Some implementations of `AbslParseFlag()` for types which consist of other,
-// constituent types which already have Abseil flag support, may need to call
-// `absl::ParseFlag()` on those consituent string values. (See above.)
-template <typename T>
-inline bool ParseFlag(absl::string_view input, T* dst, std::string* error) {
-  return flags_internal::InvokeParseFlag(input, dst, error);
-}
-
-// UnparseFlag()
-//
-// Unparses a flag value of type `T` into a string value. Do not add overloads
-// of this function for your type directly; instead, add an `AbslUnparseFlag()`
-// free function as documented above.
-//
-// Some implementations of `AbslUnparseFlag()` for types which consist of other,
-// constituent types which already have Abseil flag support, may want to call
-// `absl::UnparseFlag()` on those constituent types. (See above.)
-template <typename T>
-inline std::string UnparseFlag(const T& v) {
-  return flags_internal::Unparse(v);
-}
-
-// Overloads for `absl::LogSeverity` can't (easily) appear alongside that type's
-// definition because it is layered below flags.  See proper documentation in
-// base/log_severity.h.
-enum class LogSeverity : int;
-bool AbslParseFlag(absl::string_view, absl::LogSeverity*, std::string*);
-std::string AbslUnparseFlag(absl::LogSeverity);
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FLAGS_MARSHALLING_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/parse.h b/projects/llm_framework/include/abseil-cpp/absl/flags/parse.h
deleted file mode 100644
index f2a5cb10..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/parse.h
+++ /dev/null
@@ -1,130 +0,0 @@
-//
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: parse.h
-// -----------------------------------------------------------------------------
-//
-// This file defines the main parsing function for Abseil flags:
-// `absl::ParseCommandLine()`.
-
-#ifndef ABSL_FLAGS_PARSE_H_
-#define ABSL_FLAGS_PARSE_H_
-
-#include <string>
-#include <vector>
-
-#include "absl/base/config.h"
-#include "absl/flags/internal/parse.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// This type represent information about an unrecognized flag in the command
-// line.
-struct UnrecognizedFlag {
-  enum Source { kFromArgv, kFromFlagfile };
-
-  explicit UnrecognizedFlag(Source s, absl::string_view f)
-      : source(s), flag_name(f) {}
-  // This field indicates where we found this flag: on the original command line
-  // or read in some flag file.
-  Source source;
-  // Name of the flag we did not recognize in --flag_name=value or --flag_name.
-  std::string flag_name;
-};
-
-inline bool operator==(const UnrecognizedFlag& lhs,
-                       const UnrecognizedFlag& rhs) {
-  return lhs.source == rhs.source && lhs.flag_name == rhs.flag_name;
-}
-
-namespace flags_internal {
-
-HelpMode ParseAbseilFlagsOnlyImpl(
-    int argc, char* argv[], std::vector<char*>& positional_args,
-    std::vector<UnrecognizedFlag>& unrecognized_flags,
-    UsageFlagsAction usage_flag_action);
-
-}  // namespace flags_internal
-
-// ParseAbseilFlagsOnly()
-//
-// Parses a list of command-line arguments, passed in the `argc` and `argv[]`
-// parameters, into a set of Abseil Flag values, returning any unparsed
-// arguments in `positional_args` and `unrecognized_flags` output parameters.
-//
-// This function classifies all the arguments (including content of the
-// flagfiles, if any) into one of the following groups:
-//
-//   * arguments specified as "--flag=value" or "--flag value" that match
-//     registered or built-in Abseil Flags. These are "Abseil Flag arguments."
-//   * arguments specified as "--flag" that are unrecognized as Abseil Flags
-//   * arguments that are not specified as "--flag" are positional arguments
-//   * arguments that follow the flag-terminating delimiter (`--`) are also
-//     treated as positional arguments regardless of their syntax.
-//
-// All of the deduced Abseil Flag arguments are then parsed into their
-// corresponding flag values. If any syntax errors are found in these arguments,
-// the binary exits with code 1.
-//
-// This function also handles Abseil Flags built-in usage flags (e.g. --help)
-// if any were present on the command line.
-//
-// All the remaining positional arguments including original program name
-// (argv[0]) are are returned in the `positional_args` output parameter.
-//
-// All unrecognized flags that are not otherwise ignored are returned in the
-// `unrecognized_flags` output parameter. Note that the special `undefok`
-// flag allows you to specify flags which can be safely ignored; `undefok`
-// specifies these flags as a comma-separated list. Any unrecognized flags
-// that appear within `undefok` will therefore be ignored and not included in
-// the `unrecognized_flag` output parameter.
-//
-void ParseAbseilFlagsOnly(int argc, char* argv[],
-                          std::vector<char*>& positional_args,
-                          std::vector<UnrecognizedFlag>& unrecognized_flags);
-
-// ReportUnrecognizedFlags()
-//
-// Reports an error to `stderr` for all non-ignored unrecognized flags in
-// the provided `unrecognized_flags` list.
-void ReportUnrecognizedFlags(
-    const std::vector<UnrecognizedFlag>& unrecognized_flags);
-
-// ParseCommandLine()
-//
-// First parses Abseil Flags only from the command line according to the
-// description in `ParseAbseilFlagsOnly`. In addition this function handles
-// unrecognized and usage flags.
-//
-// If any unrecognized flags are located they are reported using
-// `ReportUnrecognizedFlags`.
-//
-// If any errors detected during command line parsing, this routine reports a
-// usage message and aborts the program.
-//
-// If any built-in usage flags were specified on the command line (e.g.
-// `--help`), this function reports help messages and then gracefully exits the
-// program.
-//
-// This function returns all the remaining positional arguments collected by
-// `ParseAbseilFlagsOnly`.
-std::vector<char*> ParseCommandLine(int argc, char* argv[]);
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FLAGS_PARSE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/reflection.h b/projects/llm_framework/include/abseil-cpp/absl/flags/reflection.h
deleted file mode 100644
index e6baf5de..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/reflection.h
+++ /dev/null
@@ -1,90 +0,0 @@
-//
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: reflection.h
-// -----------------------------------------------------------------------------
-//
-// This file defines the routines to access and operate on an Abseil Flag's
-// reflection handle.
-
-#ifndef ABSL_FLAGS_REFLECTION_H_
-#define ABSL_FLAGS_REFLECTION_H_
-
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/container/flat_hash_map.h"
-#include "absl/flags/commandlineflag.h"
-#include "absl/flags/internal/commandlineflag.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace flags_internal {
-class FlagSaverImpl;
-}  // namespace flags_internal
-
-// FindCommandLineFlag()
-//
-// Returns the reflection handle of an Abseil flag of the specified name, or
-// `nullptr` if not found. This function will emit a warning if the name of a
-// 'retired' flag is specified.
-absl::CommandLineFlag* FindCommandLineFlag(absl::string_view name);
-
-// Returns current state of the Flags registry in a form of mapping from flag
-// name to a flag reflection handle.
-absl::flat_hash_map<absl::string_view, absl::CommandLineFlag*> GetAllFlags();
-
-//------------------------------------------------------------------------------
-// FlagSaver
-//------------------------------------------------------------------------------
-//
-// A FlagSaver object stores the state of flags in the scope where the FlagSaver
-// is defined, allowing modification of those flags within that scope and
-// automatic restoration of the flags to their previous state upon leaving the
-// scope.
-//
-// A FlagSaver can be used within tests to temporarily change the test
-// environment and restore the test case to its previous state.
-//
-// Example:
-//
-//   void MyFunc() {
-//    absl::FlagSaver fs;
-//    ...
-//    absl::SetFlag(&FLAGS_myFlag, otherValue);
-//    ...
-//  } // scope of FlagSaver left, flags return to previous state
-//
-// This class is thread-safe.
-
-class FlagSaver {
- public:
-  FlagSaver();
-  ~FlagSaver();
-
-  FlagSaver(const FlagSaver&) = delete;
-  void operator=(const FlagSaver&) = delete;
-
- private:
-  flags_internal::FlagSaverImpl* impl_;
-};
-
-//-----------------------------------------------------------------------------
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FLAGS_REFLECTION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/usage.h b/projects/llm_framework/include/abseil-cpp/absl/flags/usage.h
deleted file mode 100644
index ad12ab7a..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/usage.h
+++ /dev/null
@@ -1,43 +0,0 @@
-//
-//  Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_FLAGS_USAGE_H_
-#define ABSL_FLAGS_USAGE_H_
-
-#include "absl/base/config.h"
-#include "absl/strings/string_view.h"
-
-// --------------------------------------------------------------------
-// Usage reporting interfaces
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// Sets the "usage" message to be used by help reporting routines.
-// For example:
-//  absl::SetProgramUsageMessage(
-//      absl::StrCat("This program does nothing.  Sample usage:\n", argv[0],
-//                   " <uselessarg1> <uselessarg2>"));
-// Do not include commandline flags in the usage: we do that for you!
-// Note: Calling SetProgramUsageMessage twice will trigger a call to std::exit.
-void SetProgramUsageMessage(absl::string_view new_usage_message);
-
-// Returns the usage message set by SetProgramUsageMessage().
-absl::string_view ProgramUsageMessage();
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FLAGS_USAGE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/flags/usage_config.h b/projects/llm_framework/include/abseil-cpp/absl/flags/usage_config.h
deleted file mode 100644
index ded70300..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/flags/usage_config.h
+++ /dev/null
@@ -1,135 +0,0 @@
-//
-//  Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: usage_config.h
-// -----------------------------------------------------------------------------
-//
-// This file defines the main usage reporting configuration interfaces and
-// documents Abseil's supported built-in usage flags. If these flags are found
-// when parsing a command-line, Abseil will exit the program and display
-// appropriate help messages.
-#ifndef ABSL_FLAGS_USAGE_CONFIG_H_
-#define ABSL_FLAGS_USAGE_CONFIG_H_
-
-#include <functional>
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/strings/string_view.h"
-
-// -----------------------------------------------------------------------------
-// Built-in Usage Flags
-// -----------------------------------------------------------------------------
-//
-// Abseil supports the following built-in usage flags. When passed, these flags
-// exit the program and :
-//
-// * --help
-//     Shows help on important flags for this binary
-// * --helpfull
-//     Shows help on all flags
-// * --helpshort
-//     Shows help on only the main module for this program
-// * --helppackage
-//     Shows help on all modules in the main package
-// * --version
-//     Shows the version and build info for this binary and exits
-// * --only_check_args
-//     Exits after checking all flags
-// * --helpon
-//     Shows help on the modules named by this flag value
-// * --helpmatch
-//     Shows help on modules whose name contains the specified substring
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-namespace flags_internal {
-using FlagKindFilter = std::function<bool (absl::string_view)>;
-}  // namespace flags_internal
-
-// FlagsUsageConfig
-//
-// This structure contains the collection of callbacks for changing the behavior
-// of the usage reporting routines in Abseil Flags.
-struct FlagsUsageConfig {
-  // Returns true if flags defined in the given source code file should be
-  // reported with --helpshort flag. For example, if the file
-  // "path/to/my/code.cc" defines the flag "--my_flag", and
-  // contains_helpshort_flags("path/to/my/code.cc") returns true, invoking the
-  // program with --helpshort will include information about --my_flag in the
-  // program output.
-  flags_internal::FlagKindFilter contains_helpshort_flags;
-
-  // Returns true if flags defined in the filename should be reported with
-  // --help flag. For example, if the file
-  // "path/to/my/code.cc" defines the flag "--my_flag", and
-  // contains_help_flags("path/to/my/code.cc") returns true, invoking the
-  // program with --help will include information about --my_flag in the
-  // program output.
-  flags_internal::FlagKindFilter contains_help_flags;
-
-  // Returns true if flags defined in the filename should be reported with
-  // --helppackage flag. For example, if the file
-  // "path/to/my/code.cc" defines the flag "--my_flag", and
-  // contains_helppackage_flags("path/to/my/code.cc") returns true, invoking the
-  // program with --helppackage will include information about --my_flag in the
-  // program output.
-  flags_internal::FlagKindFilter contains_helppackage_flags;
-
-  // Generates string containing program version. This is the string reported
-  // when user specifies --version in a command line.
-  std::function<std::string()> version_string;
-
-  // Normalizes the filename specific to the build system/filesystem used. This
-  // routine is used when we report the information about the flag definition
-  // location. For instance, if your build resides at some location you do not
-  // want to expose in the usage output, you can trim it to show only relevant
-  // part.
-  // For example:
-  //   normalize_filename("/my_company/some_long_path/src/project/file.cc")
-  // might produce
-  //   "project/file.cc".
-  std::function<std::string(absl::string_view)> normalize_filename;
-};
-
-// SetFlagsUsageConfig()
-//
-// Sets the usage reporting configuration callbacks. If any of the callbacks are
-// not set in usage_config instance, then the default value of the callback is
-// used.
-void SetFlagsUsageConfig(FlagsUsageConfig usage_config);
-
-namespace flags_internal {
-
-FlagsUsageConfig GetUsageConfig();
-
-void ReportUsageError(absl::string_view msg, bool is_fatal);
-
-}  // namespace flags_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-extern "C" {
-
-// Additional report of fatal usage error message before we std::exit. Error is
-// fatal if is_fatal argument to ReportUsageError is true.
-void ABSL_INTERNAL_C_SYMBOL(AbslInternalReportFatalUsageError)(
-    absl::string_view);
-
-}  // extern "C"
-
-#endif  // ABSL_FLAGS_USAGE_CONFIG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/functional/any_invocable.h b/projects/llm_framework/include/abseil-cpp/absl/functional/any_invocable.h
deleted file mode 100644
index 68d88253..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/functional/any_invocable.h
+++ /dev/null
@@ -1,324 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: any_invocable.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines an `absl::AnyInvocable` type that assumes ownership
-// and wraps an object of an invocable type. (Invocable types adhere to the
-// concept specified in https://en.cppreference.com/w/cpp/concepts/invocable.)
-//
-// In general, prefer `absl::AnyInvocable` when you need a type-erased
-// function parameter that needs to take ownership of the type.
-//
-// NOTE: `absl::AnyInvocable` is similar to the C++23 `std::move_only_function`
-// abstraction, but has a slightly different API and is not designed to be a
-// drop-in replacement or C++11-compatible backfill of that type.
-//
-// Credits to Matt Calabrese (https://github.com/mattcalabrese) for the original
-// implementation.
-
-#ifndef ABSL_FUNCTIONAL_ANY_INVOCABLE_H_
-#define ABSL_FUNCTIONAL_ANY_INVOCABLE_H_
-
-#include <cstddef>
-#include <initializer_list>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/config.h"
-#include "absl/functional/internal/any_invocable.h"
-#include "absl/meta/type_traits.h"
-#include "absl/utility/utility.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// absl::AnyInvocable
-//
-// `absl::AnyInvocable` is a functional wrapper type, like `std::function`, that
-// assumes ownership of an invocable object. Unlike `std::function`, an
-// `absl::AnyInvocable` is more type-safe and provides the following additional
-// benefits:
-//
-// * Properly adheres to const correctness of the underlying type
-// * Is move-only so avoids concurrency problems with copied invocables and
-//   unnecessary copies in general.
-// * Supports reference qualifiers allowing it to perform unique actions (noted
-//   below).
-//
-// `absl::AnyInvocable` is a template, and an `absl::AnyInvocable` instantiation
-// may wrap any invocable object with a compatible function signature, e.g.
-// having arguments and return types convertible to types matching the
-// `absl::AnyInvocable` signature, and also matching any stated reference
-// qualifiers, as long as that type is moveable. It therefore provides broad
-// type erasure for functional objects.
-//
-// An `absl::AnyInvocable` is typically used as a type-erased function parameter
-// for accepting various functional objects:
-//
-// // Define a function taking an AnyInvocable parameter.
-// void my_func(absl::AnyInvocable<int()> f) {
-//   ...
-// };
-//
-// // That function can accept any invocable type:
-//
-// // Accept a function reference. We don't need to move a reference.
-// int func1() { return 0; };
-// my_func(func1);
-//
-// // Accept a lambda. We use std::move here because otherwise my_func would
-// // copy the lambda.
-// auto lambda = []() { return 0; };
-// my_func(std::move(lambda));
-//
-// // Accept a function pointer. We don't need to move a function pointer.
-// func2 = &func1;
-// my_func(func2);
-//
-// // Accept an std::function by moving it. Note that the lambda is copyable
-// // (satisfying std::function requirements) and moveable (satisfying
-// // absl::AnyInvocable requirements).
-// std::function<int()> func6 = []() { return 0; };
-// my_func(std::move(func6));
-//
-// `AnyInvocable` also properly respects `const` qualifiers, reference
-// qualifiers, and the `noexcept` specification (only in C++ 17 and beyond) as
-// part of the user-specified function type (e.g.
-// `AnyInvocable<void()&& const noexcept>`). These qualifiers will be applied to
-// the `AnyInvocable` object's `operator()`, and the underlying invocable must
-// be compatible with those qualifiers.
-//
-// Comparison of const and non-const function types:
-//
-//   // Store a closure inside of `func` with the function type `int()`.
-//   // Note that we have made `func` itself `const`.
-//   const AnyInvocable<int()> func = [](){ return 0; };
-//
-//   func();  // Compile-error: the passed type `int()` isn't `const`.
-//
-//   // Store a closure inside of `const_func` with the function type
-//   // `int() const`.
-//   // Note that we have also made `const_func` itself `const`.
-//   const AnyInvocable<int() const> const_func = [](){ return 0; };
-//
-//   const_func();  // Fine: `int() const` is `const`.
-//
-// In the above example, the call `func()` would have compiled if
-// `std::function` were used even though the types are not const compatible.
-// This is a bug, and using `absl::AnyInvocable` properly detects that bug.
-//
-// In addition to affecting the signature of `operator()`, the `const` and
-// reference qualifiers of the function type also appropriately constrain which
-// kinds of invocable objects you are allowed to place into the `AnyInvocable`
-// instance. If you specify a function type that is const-qualified, then
-// anything that you attempt to put into the `AnyInvocable` must be callable on
-// a `const` instance of that type.
-//
-// Constraint example:
-//
-//   // Fine because the lambda is callable when `const`.
-//   AnyInvocable<int() const> func = [=](){ return 0; };
-//
-//   // This is a compile-error because the lambda isn't callable when `const`.
-//   AnyInvocable<int() const> error = [=]() mutable { return 0; };
-//
-// An `&&` qualifier can be used to express that an `absl::AnyInvocable`
-// instance should be invoked at most once:
-//
-//   // Invokes `continuation` with the logical result of an operation when
-//   // that operation completes (common in asynchronous code).
-//   void CallOnCompletion(AnyInvocable<void(int)&&> continuation) {
-//     int result_of_foo = foo();
-//
-//     // `std::move` is required because the `operator()` of `continuation` is
-//     // rvalue-reference qualified.
-//     std::move(continuation)(result_of_foo);
-//   }
-//
-// Attempting to call `absl::AnyInvocable` multiple times in such a case
-// results in undefined behavior.
-template <class Sig>
-class AnyInvocable : private internal_any_invocable::Impl<Sig> {
- private:
-  static_assert(
-      std::is_function<Sig>::value,
-      "The template argument of AnyInvocable must be a function type.");
-
-  using Impl = internal_any_invocable::Impl<Sig>;
-
- public:
-  // The return type of Sig
-  using result_type = typename Impl::result_type;
-
-  // Constructors
-
-  // Constructs the `AnyInvocable` in an empty state.
-  AnyInvocable() noexcept = default;
-  AnyInvocable(std::nullptr_t) noexcept {}  // NOLINT
-
-  // Constructs the `AnyInvocable` from an existing `AnyInvocable` by a move.
-  // Note that `f` is not guaranteed to be empty after move-construction,
-  // although it may be.
-  AnyInvocable(AnyInvocable&& /*f*/) noexcept = default;
-
-  // Constructs an `AnyInvocable` from an invocable object.
-  //
-  // Upon construction, `*this` is only empty if `f` is a function pointer or
-  // member pointer type and is null, or if `f` is an `AnyInvocable` that is
-  // empty.
-  template <class F, typename = absl::enable_if_t<
-                         internal_any_invocable::CanConvert<Sig, F>::value>>
-  AnyInvocable(F&& f)  // NOLINT
-      : Impl(internal_any_invocable::ConversionConstruct(),
-             std::forward<F>(f)) {}
-
-  // Constructs an `AnyInvocable` that holds an invocable object of type `T`,
-  // which is constructed in-place from the given arguments.
-  //
-  // Example:
-  //
-  //   AnyInvocable<int(int)> func(
-  //       absl::in_place_type<PossiblyImmovableType>, arg1, arg2);
-  //
-  template <class T, class... Args,
-            typename = absl::enable_if_t<
-                internal_any_invocable::CanEmplace<Sig, T, Args...>::value>>
-  explicit AnyInvocable(absl::in_place_type_t<T>, Args&&... args)
-      : Impl(absl::in_place_type<absl::decay_t<T>>,
-             std::forward<Args>(args)...) {
-    static_assert(std::is_same<T, absl::decay_t<T>>::value,
-                  "The explicit template argument of in_place_type is required "
-                  "to be an unqualified object type.");
-  }
-
-  // Overload of the above constructor to support list-initialization.
-  template <class T, class U, class... Args,
-            typename = absl::enable_if_t<internal_any_invocable::CanEmplace<
-                Sig, T, std::initializer_list<U>&, Args...>::value>>
-  explicit AnyInvocable(absl::in_place_type_t<T>,
-                        std::initializer_list<U> ilist, Args&&... args)
-      : Impl(absl::in_place_type<absl::decay_t<T>>, ilist,
-             std::forward<Args>(args)...) {
-    static_assert(std::is_same<T, absl::decay_t<T>>::value,
-                  "The explicit template argument of in_place_type is required "
-                  "to be an unqualified object type.");
-  }
-
-  // Assignment Operators
-
-  // Assigns an `AnyInvocable` through move-assignment.
-  // Note that `f` is not guaranteed to be empty after move-assignment
-  // although it may be.
-  AnyInvocable& operator=(AnyInvocable&& /*f*/) noexcept = default;
-
-  // Assigns an `AnyInvocable` from a nullptr, clearing the `AnyInvocable`. If
-  // not empty, destroys the target, putting `*this` into an empty state.
-  AnyInvocable& operator=(std::nullptr_t) noexcept {
-    this->Clear();
-    return *this;
-  }
-
-  // Assigns an `AnyInvocable` from an existing `AnyInvocable` instance.
-  //
-  // Upon assignment, `*this` is only empty if `f` is a function pointer or
-  // member pointer type and is null, or if `f` is an `AnyInvocable` that is
-  // empty.
-  template <class F, typename = absl::enable_if_t<
-                         internal_any_invocable::CanAssign<Sig, F>::value>>
-  AnyInvocable& operator=(F&& f) {
-    *this = AnyInvocable(std::forward<F>(f));
-    return *this;
-  }
-
-  // Assigns an `AnyInvocable` from a reference to an invocable object.
-  // Upon assignment, stores a reference to the invocable object in the
-  // `AnyInvocable` instance.
-  template <
-      class F,
-      typename = absl::enable_if_t<
-          internal_any_invocable::CanAssignReferenceWrapper<Sig, F>::value>>
-  AnyInvocable& operator=(std::reference_wrapper<F> f) noexcept {
-    *this = AnyInvocable(f);
-    return *this;
-  }
-
-  // Destructor
-
-  // If not empty, destroys the target.
-  ~AnyInvocable() = default;
-
-  // absl::AnyInvocable::swap()
-  //
-  // Exchanges the targets of `*this` and `other`.
-  void swap(AnyInvocable& other) noexcept { std::swap(*this, other); }
-
-  // absl::AnyInvocable::operator bool()
-  //
-  // Returns `true` if `*this` is not empty.
-  //
-  // WARNING: An `AnyInvocable` that wraps an empty `std::function` is not
-  // itself empty. This behavior is consistent with the standard equivalent
-  // `std::move_only_function`.
-  //
-  // In other words:
-  //   std::function<void()> f;  // empty
-  //   absl::AnyInvocable<void()> a = std::move(f);  // not empty
-  explicit operator bool() const noexcept { return this->HasValue(); }
-
-  // Invokes the target object of `*this`. `*this` must not be empty.
-  //
-  // Note: The signature of this function call operator is the same as the
-  //       template parameter `Sig`.
-  using Impl::operator();
-
-  // Equality operators
-
-  // Returns `true` if `*this` is empty.
-  friend bool operator==(const AnyInvocable& f, std::nullptr_t) noexcept {
-    return !f.HasValue();
-  }
-
-  // Returns `true` if `*this` is empty.
-  friend bool operator==(std::nullptr_t, const AnyInvocable& f) noexcept {
-    return !f.HasValue();
-  }
-
-  // Returns `false` if `*this` is empty.
-  friend bool operator!=(const AnyInvocable& f, std::nullptr_t) noexcept {
-    return f.HasValue();
-  }
-
-  // Returns `false` if `*this` is empty.
-  friend bool operator!=(std::nullptr_t, const AnyInvocable& f) noexcept {
-    return f.HasValue();
-  }
-
-  // swap()
-  //
-  // Exchanges the targets of `f1` and `f2`.
-  friend void swap(AnyInvocable& f1, AnyInvocable& f2) noexcept { f1.swap(f2); }
-
- private:
-  // Friending other instantiations is necessary for conversions.
-  template <bool /*SigIsNoexcept*/, class /*ReturnType*/, class... /*P*/>
-  friend class internal_any_invocable::CoreImpl;
-};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FUNCTIONAL_ANY_INVOCABLE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/functional/bind_front.h b/projects/llm_framework/include/abseil-cpp/absl/functional/bind_front.h
deleted file mode 100644
index 885f24b8..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/functional/bind_front.h
+++ /dev/null
@@ -1,194 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: bind_front.h
-// -----------------------------------------------------------------------------
-//
-// `absl::bind_front()` returns a functor by binding a number of arguments to
-// the front of a provided (usually more generic) functor. Unlike `std::bind`,
-// it does not require the use of argument placeholders. The simpler syntax of
-// `absl::bind_front()` allows you to avoid known misuses with `std::bind()`.
-//
-// `absl::bind_front()` is meant as a drop-in replacement for C++20's upcoming
-// `std::bind_front()`, which similarly resolves these issues with
-// `std::bind()`. Both `bind_front()` alternatives, unlike `std::bind()`, allow
-// partial function application. (See
-// https://en.wikipedia.org/wiki/Partial_application).
-
-#ifndef ABSL_FUNCTIONAL_BIND_FRONT_H_
-#define ABSL_FUNCTIONAL_BIND_FRONT_H_
-
-#if defined(__cpp_lib_bind_front) && __cpp_lib_bind_front >= 201907L
-#include <functional>  // For std::bind_front.
-#endif  // defined(__cpp_lib_bind_front) && __cpp_lib_bind_front >= 201907L
-
-#include <utility>
-
-#include "absl/functional/internal/front_binder.h"
-#include "absl/utility/utility.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// bind_front()
-//
-// Binds the first N arguments of an invocable object and stores them by value.
-//
-// Like `std::bind()`, `absl::bind_front()` is implicitly convertible to
-// `std::function`.  In particular, it may be used as a simpler replacement for
-// `std::bind()` in most cases, as it does not require placeholders to be
-// specified. More importantly, it provides more reliable correctness guarantees
-// than `std::bind()`; while `std::bind()` will silently ignore passing more
-// parameters than expected, for example, `absl::bind_front()` will report such
-// mis-uses as errors. In C++20, `absl::bind_front` is replaced by
-// `std::bind_front`.
-//
-// absl::bind_front(a...) can be seen as storing the results of
-// std::make_tuple(a...).
-//
-// Example: Binding a free function.
-//
-//   int Minus(int a, int b) { return a - b; }
-//
-//   assert(absl::bind_front(Minus)(3, 2) == 3 - 2);
-//   assert(absl::bind_front(Minus, 3)(2) == 3 - 2);
-//   assert(absl::bind_front(Minus, 3, 2)() == 3 - 2);
-//
-// Example: Binding a member function.
-//
-//   struct Math {
-//     int Double(int a) const { return 2 * a; }
-//   };
-//
-//   Math math;
-//
-//   assert(absl::bind_front(&Math::Double)(&math, 3) == 2 * 3);
-//   // Stores a pointer to math inside the functor.
-//   assert(absl::bind_front(&Math::Double, &math)(3) == 2 * 3);
-//   // Stores a copy of math inside the functor.
-//   assert(absl::bind_front(&Math::Double, math)(3) == 2 * 3);
-//   // Stores std::unique_ptr<Math> inside the functor.
-//   assert(absl::bind_front(&Math::Double,
-//                           std::unique_ptr<Math>(new Math))(3) == 2 * 3);
-//
-// Example: Using `absl::bind_front()`, instead of `std::bind()`, with
-//          `std::function`.
-//
-//   class FileReader {
-//    public:
-//     void ReadFileAsync(const std::string& filename, std::string* content,
-//                        const std::function<void()>& done) {
-//       // Calls Executor::Schedule(std::function<void()>).
-//       Executor::DefaultExecutor()->Schedule(
-//           absl::bind_front(&FileReader::BlockingRead, this,
-//                            filename, content, done));
-//     }
-//
-//    private:
-//     void BlockingRead(const std::string& filename, std::string* content,
-//                       const std::function<void()>& done) {
-//       CHECK_OK(file::GetContents(filename, content, {}));
-//       done();
-//     }
-//   };
-//
-// `absl::bind_front()` stores bound arguments explicitly using the type passed
-// rather than implicitly based on the type accepted by its functor.
-//
-// Example: Binding arguments explicitly.
-//
-//   void LogStringView(absl::string_view sv) {
-//     LOG(INFO) << sv;
-//   }
-//
-//   Executor* e = Executor::DefaultExecutor();
-//   std::string s = "hello";
-//   absl::string_view sv = s;
-//
-//   // absl::bind_front(LogStringView, arg) makes a copy of arg and stores it.
-//   e->Schedule(absl::bind_front(LogStringView, sv)); // ERROR: dangling
-//                                                     // string_view.
-//
-//   e->Schedule(absl::bind_front(LogStringView, s));  // OK: stores a copy of
-//                                                     // s.
-//
-// To store some of the arguments passed to `absl::bind_front()` by reference,
-//  use std::ref()` and `std::cref()`.
-//
-// Example: Storing some of the bound arguments by reference.
-//
-//   class Service {
-//    public:
-//     void Serve(const Request& req, std::function<void()>* done) {
-//       // The request protocol buffer won't be deleted until done is called.
-//       // It's safe to store a reference to it inside the functor.
-//       Executor::DefaultExecutor()->Schedule(
-//           absl::bind_front(&Service::BlockingServe, this, std::cref(req),
-//           done));
-//     }
-//
-//    private:
-//     void BlockingServe(const Request& req, std::function<void()>* done);
-//   };
-//
-// Example: Storing bound arguments by reference.
-//
-//   void Print(const std::string& a, const std::string& b) {
-//     std::cerr << a << b;
-//   }
-//
-//   std::string hi = "Hello, ";
-//   std::vector<std::string> names = {"Chuk", "Gek"};
-//   // Doesn't copy hi.
-//   for_each(names.begin(), names.end(),
-//            absl::bind_front(Print, std::ref(hi)));
-//
-//   // DO NOT DO THIS: the functor may outlive "hi", resulting in
-//   // dangling references.
-//   foo->DoInFuture(absl::bind_front(Print, std::ref(hi), "Guest"));  // BAD!
-//   auto f = absl::bind_front(Print, std::ref(hi), "Guest"); // BAD!
-//
-// Example: Storing reference-like types.
-//
-//   void Print(absl::string_view a, const std::string& b) {
-//     std::cerr << a << b;
-//   }
-//
-//   std::string hi = "Hello, ";
-//   // Copies "hi".
-//   absl::bind_front(Print, hi)("Chuk");
-//
-//   // Compile error: std::reference_wrapper<const string> is not implicitly
-//   // convertible to string_view.
-//   // absl::bind_front(Print, std::cref(hi))("Chuk");
-//
-//   // Doesn't copy "hi".
-//   absl::bind_front(Print, absl::string_view(hi))("Chuk");
-//
-#if defined(__cpp_lib_bind_front) && __cpp_lib_bind_front >= 201907L
-using std::bind_front;
-#else   // defined(__cpp_lib_bind_front) && __cpp_lib_bind_front >= 201907L
-template <class F, class... BoundArgs>
-constexpr functional_internal::bind_front_t<F, BoundArgs...> bind_front(
-    F&& func, BoundArgs&&... args) {
-  return functional_internal::bind_front_t<F, BoundArgs...>(
-      absl::in_place, std::forward<F>(func), std::forward<BoundArgs>(args)...);
-}
-#endif  // defined(__cpp_lib_bind_front) && __cpp_lib_bind_front >= 201907L
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FUNCTIONAL_BIND_FRONT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/functional/function_ref.h b/projects/llm_framework/include/abseil-cpp/absl/functional/function_ref.h
deleted file mode 100644
index 96cece55..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/functional/function_ref.h
+++ /dev/null
@@ -1,151 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: function_ref.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines the `absl::FunctionRef` type for holding a
-// non-owning reference to an object of any invocable type. This function
-// reference is typically most useful as a type-erased argument type for
-// accepting function types that neither take ownership nor copy the type; using
-// the reference type in this case avoids a copy and an allocation. Best
-// practices of other non-owning reference-like objects (such as
-// `absl::string_view`) apply here.
-//
-//  An `absl::FunctionRef` is similar in usage to a `std::function` but has the
-//  following differences:
-//
-//  * It doesn't own the underlying object.
-//  * It doesn't have a null or empty state.
-//  * It never performs deep copies or allocations.
-//  * It's much faster and cheaper to construct.
-//  * It's trivially copyable and destructable.
-//
-// Generally, `absl::FunctionRef` should not be used as a return value, data
-// member, or to initialize a `std::function`. Such usages will often lead to
-// problematic lifetime issues. Once you convert something to an
-// `absl::FunctionRef` you cannot make a deep copy later.
-//
-// This class is suitable for use wherever a "const std::function<>&"
-// would be used without making a copy. ForEach functions and other versions of
-// the visitor pattern are a good example of when this class should be used.
-//
-// This class is trivial to copy and should be passed by value.
-#ifndef ABSL_FUNCTIONAL_FUNCTION_REF_H_
-#define ABSL_FUNCTIONAL_FUNCTION_REF_H_
-
-#include <cassert>
-#include <functional>
-#include <type_traits>
-
-#include "absl/base/attributes.h"
-#include "absl/functional/internal/function_ref.h"
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// FunctionRef
-//
-// Dummy class declaration to allow the partial specialization based on function
-// types below.
-template <typename T>
-class FunctionRef;
-
-// FunctionRef
-//
-// An `absl::FunctionRef` is a lightweight wrapper to any invocable object with
-// a compatible signature. Generally, an `absl::FunctionRef` should only be used
-// as an argument type and should be preferred as an argument over a const
-// reference to a `std::function`. `absl::FunctionRef` itself does not allocate,
-// although the wrapped invocable may.
-//
-// Example:
-//
-//   // The following function takes a function callback by const reference
-//   bool Visitor(const std::function<void(my_proto&,
-//                                         absl::string_view)>& callback);
-//
-//   // Assuming that the function is not stored or otherwise copied, it can be
-//   // replaced by an `absl::FunctionRef`:
-//   bool Visitor(absl::FunctionRef<void(my_proto&, absl::string_view)>
-//                  callback);
-//
-// Note: the assignment operator within an `absl::FunctionRef` is intentionally
-// deleted to prevent misuse; because the `absl::FunctionRef` does not own the
-// underlying type, assignment likely indicates misuse.
-template <typename R, typename... Args>
-class FunctionRef<R(Args...)> {
- private:
-  // Used to disable constructors for objects that are not compatible with the
-  // signature of this FunctionRef.
-  template <typename F,
-            typename FR = absl::base_internal::invoke_result_t<F, Args&&...>>
-  using EnableIfCompatible =
-      typename std::enable_if<std::is_void<R>::value ||
-                              std::is_convertible<FR, R>::value>::type;
-
- public:
-  // Constructs a FunctionRef from any invocable type.
-  template <typename F, typename = EnableIfCompatible<const F&>>
-  // NOLINTNEXTLINE(runtime/explicit)
-  FunctionRef(const F& f ABSL_ATTRIBUTE_LIFETIME_BOUND)
-      : invoker_(&absl::functional_internal::InvokeObject<F, R, Args...>) {
-    absl::functional_internal::AssertNonNull(f);
-    ptr_.obj = &f;
-  }
-
-  // Overload for function pointers. This eliminates a level of indirection that
-  // would happen if the above overload was used (it lets us store the pointer
-  // instead of a pointer to a pointer).
-  //
-  // This overload is also used for references to functions, since references to
-  // functions can decay to function pointers implicitly.
-  template <
-      typename F, typename = EnableIfCompatible<F*>,
-      absl::functional_internal::EnableIf<absl::is_function<F>::value> = 0>
-  FunctionRef(F* f)  // NOLINT(runtime/explicit)
-      : invoker_(&absl::functional_internal::InvokeFunction<F*, R, Args...>) {
-    assert(f != nullptr);
-    ptr_.fun = reinterpret_cast<decltype(ptr_.fun)>(f);
-  }
-
-  // To help prevent subtle lifetime bugs, FunctionRef is not assignable.
-  // Typically, it should only be used as an argument type.
-  FunctionRef& operator=(const FunctionRef& rhs) = delete;
-  FunctionRef(const FunctionRef& rhs) = default;
-
-  // Call the underlying object.
-  R operator()(Args... args) const {
-    return invoker_(ptr_, std::forward<Args>(args)...);
-  }
-
- private:
-  absl::functional_internal::VoidPtr ptr_;
-  absl::functional_internal::Invoker<R, Args...> invoker_;
-};
-
-// Allow const qualified function signatures. Since FunctionRef requires
-// constness anyway we can just make this a no-op.
-template <typename R, typename... Args>
-class FunctionRef<R(Args...) const> : public FunctionRef<R(Args...)> {
- public:
-  using FunctionRef<R(Args...)>::FunctionRef;
-};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FUNCTIONAL_FUNCTION_REF_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/functional/internal/any_invocable.h b/projects/llm_framework/include/abseil-cpp/absl/functional/internal/any_invocable.h
deleted file mode 100644
index b04436d1..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/functional/internal/any_invocable.h
+++ /dev/null
@@ -1,891 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Implementation details for `absl::AnyInvocable`
-
-#ifndef ABSL_FUNCTIONAL_INTERNAL_ANY_INVOCABLE_H_
-#define ABSL_FUNCTIONAL_INTERNAL_ANY_INVOCABLE_H_
-
-////////////////////////////////////////////////////////////////////////////////
-//                                                                            //
-// This implementation of the proposed `any_invocable` uses an approach that  //
-// chooses between local storage and remote storage for the contained target  //
-// object based on the target object's size, alignment requirements, and      //
-// whether or not it has a nothrow move constructor. Additional optimizations //
-// are performed when the object is a trivially copyable type [basic.types].  //
-//                                                                            //
-// There are three datamembers per `AnyInvocable` instance                    //
-//                                                                            //
-// 1) A union containing either                                               //
-//        - A pointer to the target object referred to via a void*, or        //
-//        - the target object, emplaced into a raw char buffer                //
-//                                                                            //
-// 2) A function pointer to a "manager" function operation that takes a       //
-//    discriminator and logically branches to either perform a move operation //
-//    or destroy operation based on that discriminator.                       //
-//                                                                            //
-// 3) A function pointer to an "invoker" function operation that invokes the  //
-//    target object, directly returning the result.                           //
-//                                                                            //
-// When in the logically empty state, the manager function is an empty        //
-// function and the invoker function is one that would be undefined-behavior  //
-// to call.                                                                   //
-//                                                                            //
-// An additional optimization is performed when converting from one           //
-// AnyInvocable to another where only the noexcept specification and/or the   //
-// cv/ref qualifiers of the function type differ. In these cases, the         //
-// conversion works by "moving the guts", similar to if they were the same    //
-// exact type, as opposed to having to perform an additional layer of         //
-// wrapping through remote storage.                                           //
-//                                                                            //
-////////////////////////////////////////////////////////////////////////////////
-
-// IWYU pragma: private, include "absl/functional/any_invocable.h"
-
-#include <cassert>
-#include <cstddef>
-#include <cstring>
-#include <exception>
-#include <functional>
-#include <initializer_list>
-#include <memory>
-#include <new>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/config.h"
-#include "absl/base/internal/invoke.h"
-#include "absl/base/macros.h"
-#include "absl/base/optimization.h"
-#include "absl/meta/type_traits.h"
-#include "absl/utility/utility.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// Helper macro used to prevent spelling `noexcept` in language versions older
-// than C++17, where it is not part of the type system, in order to avoid
-// compilation failures and internal compiler errors.
-#if ABSL_INTERNAL_CPLUSPLUS_LANG >= 201703L
-#define ABSL_INTERNAL_NOEXCEPT_SPEC(noex) noexcept(noex)
-#else
-#define ABSL_INTERNAL_NOEXCEPT_SPEC(noex)
-#endif
-
-// Defined in functional/any_invocable.h
-template <class Sig>
-class AnyInvocable;
-
-namespace internal_any_invocable {
-
-// Constants relating to the small-object-storage for AnyInvocable
-enum StorageProperty : std::size_t {
-  kAlignment = alignof(std::max_align_t),  // The alignment of the storage
-  kStorageSize = sizeof(void*) * 2         // The size of the storage
-};
-
-////////////////////////////////////////////////////////////////////////////////
-//
-// A metafunction for checking if a type is an AnyInvocable instantiation.
-// This is used during conversion operations.
-template <class T>
-struct IsAnyInvocable : std::false_type {};
-
-template <class Sig>
-struct IsAnyInvocable<AnyInvocable<Sig>> : std::true_type {};
-//
-////////////////////////////////////////////////////////////////////////////////
-
-// A type trait that tells us whether or not a target function type should be
-// stored locally in the small object optimization storage
-template <class T>
-using IsStoredLocally = std::integral_constant<
-    bool, sizeof(T) <= kStorageSize && alignof(T) <= kAlignment &&
-              kAlignment % alignof(T) == 0 &&
-              std::is_nothrow_move_constructible<T>::value>;
-
-// An implementation of std::remove_cvref_t of C++20.
-template <class T>
-using RemoveCVRef =
-    typename std::remove_cv<typename std::remove_reference<T>::type>::type;
-
-////////////////////////////////////////////////////////////////////////////////
-//
-// An implementation of the C++ standard INVOKE<R> pseudo-macro, operation is
-// equivalent to std::invoke except that it forces an implicit conversion to the
-// specified return type. If "R" is void, the function is executed and the
-// return value is simply ignored.
-template <class ReturnType, class F, class... P,
-          typename = absl::enable_if_t<std::is_void<ReturnType>::value>>
-void InvokeR(F&& f, P&&... args) {
-  absl::base_internal::invoke(std::forward<F>(f), std::forward<P>(args)...);
-}
-
-template <class ReturnType, class F, class... P,
-          absl::enable_if_t<!std::is_void<ReturnType>::value, int> = 0>
-ReturnType InvokeR(F&& f, P&&... args) {
-  // GCC 12 has a false-positive -Wmaybe-uninitialized warning here.
-#if ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(12, 0)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
-#endif
-  return absl::base_internal::invoke(std::forward<F>(f),
-                                     std::forward<P>(args)...);
-#if ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(12, 0)
-#pragma GCC diagnostic pop
-#endif
-}
-
-//
-////////////////////////////////////////////////////////////////////////////////
-
-////////////////////////////////////////////////////////////////////////////////
-///
-// A metafunction that takes a "T" corresponding to a parameter type of the
-// user's specified function type, and yields the parameter type to use for the
-// type-erased invoker. In order to prevent observable moves, this must be
-// either a reference or, if the type is trivial, the original parameter type
-// itself. Since the parameter type may be incomplete at the point that this
-// metafunction is used, we can only do this optimization for scalar types
-// rather than for any trivial type.
-template <typename T>
-T ForwardImpl(std::true_type);
-
-template <typename T>
-T&& ForwardImpl(std::false_type);
-
-// NOTE: We deliberately use an intermediate struct instead of a direct alias,
-// as a workaround for b/206991861 on MSVC versions < 1924.
-template <class T>
-struct ForwardedParameter {
-  using type = decltype((
-      ForwardImpl<T>)(std::integral_constant<bool,
-                                             std::is_scalar<T>::value>()));
-};
-
-template <class T>
-using ForwardedParameterType = typename ForwardedParameter<T>::type;
-//
-////////////////////////////////////////////////////////////////////////////////
-
-// A discriminator when calling the "manager" function that describes operation
-// type-erased operation should be invoked.
-//
-// "relocate_from_to" specifies that the manager should perform a move.
-//
-// "dispose" specifies that the manager should perform a destroy.
-enum class FunctionToCall : bool { relocate_from_to, dispose };
-
-// The portion of `AnyInvocable` state that contains either a pointer to the
-// target object or the object itself in local storage
-union TypeErasedState {
-  struct {
-    // A pointer to the type-erased object when remotely stored
-    void* target;
-    // The size of the object for `RemoteManagerTrivial`
-    std::size_t size;
-  } remote;
-
-  // Local-storage for the type-erased object when small and trivial enough
-  alignas(kAlignment) char storage[kStorageSize];
-};
-
-// A typed accessor for the object in `TypeErasedState` storage
-template <class T>
-T& ObjectInLocalStorage(TypeErasedState* const state) {
-  // We launder here because the storage may be reused with the same type.
-#if defined(__cpp_lib_launder) && __cpp_lib_launder >= 201606L
-  return *std::launder(reinterpret_cast<T*>(&state->storage));
-#elif ABSL_HAVE_BUILTIN(__builtin_launder)
-  return *__builtin_launder(reinterpret_cast<T*>(&state->storage));
-#else
-
-  // When `std::launder` or equivalent are not available, we rely on undefined
-  // behavior, which works as intended on Abseil's officially supported
-  // platforms as of Q2 2022.
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#endif
-  return *reinterpret_cast<T*>(&state->storage);
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic pop
-#endif
-
-#endif
-}
-
-// The type for functions issuing lifetime-related operations: move and dispose
-// A pointer to such a function is contained in each `AnyInvocable` instance.
-// NOTE: When specifying `FunctionToCall::`dispose, the same state must be
-// passed as both "from" and "to".
-using ManagerType = void(FunctionToCall /*operation*/,
-                         TypeErasedState* /*from*/, TypeErasedState* /*to*/)
-    ABSL_INTERNAL_NOEXCEPT_SPEC(true);
-
-// The type for functions issuing the actual invocation of the object
-// A pointer to such a function is contained in each AnyInvocable instance.
-template <bool SigIsNoexcept, class ReturnType, class... P>
-using InvokerType = ReturnType(TypeErasedState*, ForwardedParameterType<P>...)
-    ABSL_INTERNAL_NOEXCEPT_SPEC(SigIsNoexcept);
-
-// The manager that is used when AnyInvocable is empty
-inline void EmptyManager(FunctionToCall /*operation*/,
-                         TypeErasedState* /*from*/,
-                         TypeErasedState* /*to*/) noexcept {}
-
-// The manager that is used when a target function is in local storage and is
-// a trivially copyable type.
-inline void LocalManagerTrivial(FunctionToCall /*operation*/,
-                                TypeErasedState* const from,
-                                TypeErasedState* const to) noexcept {
-  // This single statement without branching handles both possible operations.
-  //
-  // For FunctionToCall::dispose, "from" and "to" point to the same state, and
-  // so this assignment logically would do nothing.
-  //
-  // Note: Correctness here relies on http://wg21.link/p0593, which has only
-  // become standard in C++20, though implementations do not break it in
-  // practice for earlier versions of C++.
-  //
-  // The correct way to do this without that paper is to first placement-new a
-  // default-constructed T in "to->storage" prior to the memmove, but doing so
-  // requires a different function to be created for each T that is stored
-  // locally, which can cause unnecessary bloat and be less cache friendly.
-  *to = *from;
-
-  // Note: Because the type is trivially copyable, the destructor does not need
-  // to be called ("trivially copyable" requires a trivial destructor).
-}
-
-// The manager that is used when a target function is in local storage and is
-// not a trivially copyable type.
-template <class T>
-void LocalManagerNontrivial(FunctionToCall operation,
-                            TypeErasedState* const from,
-                            TypeErasedState* const to) noexcept {
-  static_assert(IsStoredLocally<T>::value,
-                "Local storage must only be used for supported types.");
-  static_assert(!std::is_trivially_copyable<T>::value,
-                "Locally stored types must be trivially copyable.");
-
-  T& from_object = (ObjectInLocalStorage<T>)(from);
-
-  switch (operation) {
-    case FunctionToCall::relocate_from_to:
-      // NOTE: Requires that the left-hand operand is already empty.
-      ::new (static_cast<void*>(&to->storage)) T(std::move(from_object));
-      ABSL_FALLTHROUGH_INTENDED;
-    case FunctionToCall::dispose:
-      from_object.~T();  // Must not throw. // NOLINT
-      return;
-  }
-  ABSL_UNREACHABLE();
-}
-
-// The invoker that is used when a target function is in local storage
-// Note: QualTRef here is the target function type along with cv and reference
-// qualifiers that must be used when calling the function.
-template <bool SigIsNoexcept, class ReturnType, class QualTRef, class... P>
-ReturnType LocalInvoker(
-    TypeErasedState* const state,
-    ForwardedParameterType<P>... args) noexcept(SigIsNoexcept) {
-  using RawT = RemoveCVRef<QualTRef>;
-  static_assert(
-      IsStoredLocally<RawT>::value,
-      "Target object must be in local storage in order to be invoked from it.");
-
-  auto& f = (ObjectInLocalStorage<RawT>)(state);
-  return (InvokeR<ReturnType>)(static_cast<QualTRef>(f),
-                               static_cast<ForwardedParameterType<P>>(args)...);
-}
-
-// The manager that is used when a target function is in remote storage and it
-// has a trivial destructor
-inline void RemoteManagerTrivial(FunctionToCall operation,
-                                 TypeErasedState* const from,
-                                 TypeErasedState* const to) noexcept {
-  switch (operation) {
-    case FunctionToCall::relocate_from_to:
-      // NOTE: Requires that the left-hand operand is already empty.
-      to->remote = from->remote;
-      return;
-    case FunctionToCall::dispose:
-#if defined(__cpp_sized_deallocation)
-      ::operator delete(from->remote.target, from->remote.size);
-#else   // __cpp_sized_deallocation
-      ::operator delete(from->remote.target);
-#endif  // __cpp_sized_deallocation
-      return;
-  }
-  ABSL_UNREACHABLE();
-}
-
-// The manager that is used when a target function is in remote storage and the
-// destructor of the type is not trivial
-template <class T>
-void RemoteManagerNontrivial(FunctionToCall operation,
-                             TypeErasedState* const from,
-                             TypeErasedState* const to) noexcept {
-  static_assert(!IsStoredLocally<T>::value,
-                "Remote storage must only be used for types that do not "
-                "qualify for local storage.");
-
-  switch (operation) {
-    case FunctionToCall::relocate_from_to:
-      // NOTE: Requires that the left-hand operand is already empty.
-      to->remote.target = from->remote.target;
-      return;
-    case FunctionToCall::dispose:
-      ::delete static_cast<T*>(from->remote.target);  // Must not throw.
-      return;
-  }
-  ABSL_UNREACHABLE();
-}
-
-// The invoker that is used when a target function is in remote storage
-template <bool SigIsNoexcept, class ReturnType, class QualTRef, class... P>
-ReturnType RemoteInvoker(
-    TypeErasedState* const state,
-    ForwardedParameterType<P>... args) noexcept(SigIsNoexcept) {
-  using RawT = RemoveCVRef<QualTRef>;
-  static_assert(!IsStoredLocally<RawT>::value,
-                "Target object must be in remote storage in order to be "
-                "invoked from it.");
-
-  auto& f = *static_cast<RawT*>(state->remote.target);
-  return (InvokeR<ReturnType>)(static_cast<QualTRef>(f),
-                               static_cast<ForwardedParameterType<P>>(args)...);
-}
-
-////////////////////////////////////////////////////////////////////////////////
-//
-// A metafunction that checks if a type T is an instantiation of
-// absl::in_place_type_t (needed for constructor constraints of AnyInvocable).
-template <class T>
-struct IsInPlaceType : std::false_type {};
-
-template <class T>
-struct IsInPlaceType<absl::in_place_type_t<T>> : std::true_type {};
-//
-////////////////////////////////////////////////////////////////////////////////
-
-// A constructor name-tag used with CoreImpl (below) to request the
-// conversion-constructor. QualDecayedTRef is the decayed-type of the object to
-// wrap, along with the cv and reference qualifiers that must be applied when
-// performing an invocation of the wrapped object.
-template <class QualDecayedTRef>
-struct TypedConversionConstruct {};
-
-// A helper base class for all core operations of AnyInvocable. Most notably,
-// this class creates the function call operator and constraint-checkers so that
-// the top-level class does not have to be a series of partial specializations.
-//
-// Note: This definition exists (as opposed to being a declaration) so that if
-// the user of the top-level template accidentally passes a template argument
-// that is not a function type, they will get a static_assert in AnyInvocable's
-// class body rather than an error stating that Impl is not defined.
-template <class Sig>
-class Impl {};  // Note: This is partially-specialized later.
-
-// A std::unique_ptr deleter that deletes memory allocated via ::operator new.
-#if defined(__cpp_sized_deallocation)
-class TrivialDeleter {
- public:
-  explicit TrivialDeleter(std::size_t size) : size_(size) {}
-
-  void operator()(void* target) const {
-    ::operator delete(target, size_);
-  }
-
- private:
-  std::size_t size_;
-};
-#else   // __cpp_sized_deallocation
-class TrivialDeleter {
- public:
-  explicit TrivialDeleter(std::size_t) {}
-
-  void operator()(void* target) const { ::operator delete(target); }
-};
-#endif  // __cpp_sized_deallocation
-
-template <bool SigIsNoexcept, class ReturnType, class... P>
-class CoreImpl;
-
-constexpr bool IsCompatibleConversion(void*, void*) { return false; }
-template <bool NoExceptSrc, bool NoExceptDest, class... T>
-constexpr bool IsCompatibleConversion(CoreImpl<NoExceptSrc, T...>*,
-                                      CoreImpl<NoExceptDest, T...>*) {
-  return !NoExceptDest || NoExceptSrc;
-}
-
-// A helper base class for all core operations of AnyInvocable that do not
-// depend on the cv/ref qualifiers of the function type.
-template <bool SigIsNoexcept, class ReturnType, class... P>
-class CoreImpl {
- public:
-  using result_type = ReturnType;
-
-  CoreImpl() noexcept : manager_(EmptyManager), invoker_(nullptr) {}
-
-  enum class TargetType {
-    kPointer,
-    kCompatibleAnyInvocable,
-    kIncompatibleAnyInvocable,
-    kOther,
-  };
-
-  // Note: QualDecayedTRef here includes the cv-ref qualifiers associated with
-  // the invocation of the Invocable. The unqualified type is the target object
-  // type to be stored.
-  template <class QualDecayedTRef, class F>
-  explicit CoreImpl(TypedConversionConstruct<QualDecayedTRef>, F&& f) {
-    using DecayedT = RemoveCVRef<QualDecayedTRef>;
-
-    constexpr TargetType kTargetType =
-        (std::is_pointer<DecayedT>::value ||
-         std::is_member_pointer<DecayedT>::value)
-            ? TargetType::kPointer
-        : IsCompatibleAnyInvocable<DecayedT>::value
-            ? TargetType::kCompatibleAnyInvocable
-        : IsAnyInvocable<DecayedT>::value
-            ? TargetType::kIncompatibleAnyInvocable
-            : TargetType::kOther;
-    // NOTE: We only use integers instead of enums as template parameters in
-    // order to work around a bug on C++14 under MSVC 2017.
-    // See b/236131881.
-    Initialize<kTargetType, QualDecayedTRef>(std::forward<F>(f));
-  }
-
-  // Note: QualTRef here includes the cv-ref qualifiers associated with the
-  // invocation of the Invocable. The unqualified type is the target object
-  // type to be stored.
-  template <class QualTRef, class... Args>
-  explicit CoreImpl(absl::in_place_type_t<QualTRef>, Args&&... args) {
-    InitializeStorage<QualTRef>(std::forward<Args>(args)...);
-  }
-
-  CoreImpl(CoreImpl&& other) noexcept {
-    other.manager_(FunctionToCall::relocate_from_to, &other.state_, &state_);
-    manager_ = other.manager_;
-    invoker_ = other.invoker_;
-    other.manager_ = EmptyManager;
-    other.invoker_ = nullptr;
-  }
-
-  CoreImpl& operator=(CoreImpl&& other) noexcept {
-    // Put the left-hand operand in an empty state.
-    //
-    // Note: A full reset that leaves us with an object that has its invariants
-    // intact is necessary in order to handle self-move. This is required by
-    // types that are used with certain operations of the standard library, such
-    // as the default definition of std::swap when both operands target the same
-    // object.
-    Clear();
-
-    // Perform the actual move/destroy operation on the target function.
-    other.manager_(FunctionToCall::relocate_from_to, &other.state_, &state_);
-    manager_ = other.manager_;
-    invoker_ = other.invoker_;
-    other.manager_ = EmptyManager;
-    other.invoker_ = nullptr;
-
-    return *this;
-  }
-
-  ~CoreImpl() { manager_(FunctionToCall::dispose, &state_, &state_); }
-
-  // Check whether or not the AnyInvocable is in the empty state.
-  bool HasValue() const { return invoker_ != nullptr; }
-
-  // Effects: Puts the object into its empty state.
-  void Clear() {
-    manager_(FunctionToCall::dispose, &state_, &state_);
-    manager_ = EmptyManager;
-    invoker_ = nullptr;
-  }
-
-  template <TargetType target_type, class QualDecayedTRef, class F,
-            absl::enable_if_t<target_type == TargetType::kPointer, int> = 0>
-  void Initialize(F&& f) {
-// This condition handles types that decay into pointers, which includes
-// function references. Since function references cannot be null, GCC warns
-// against comparing their decayed form with nullptr.
-// Since this is template-heavy code, we prefer to disable these warnings
-// locally instead of adding yet another overload of this function.
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wpragmas"
-#pragma GCC diagnostic ignored "-Waddress"
-#pragma GCC diagnostic ignored "-Wnonnull-compare"
-#endif
-    if (static_cast<RemoveCVRef<QualDecayedTRef>>(f) == nullptr) {
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic pop
-#endif
-      manager_ = EmptyManager;
-      invoker_ = nullptr;
-      return;
-    }
-    InitializeStorage<QualDecayedTRef>(std::forward<F>(f));
-  }
-
-  template <TargetType target_type, class QualDecayedTRef, class F,
-            absl::enable_if_t<
-                target_type == TargetType::kCompatibleAnyInvocable, int> = 0>
-  void Initialize(F&& f) {
-    // In this case we can "steal the guts" of the other AnyInvocable.
-    f.manager_(FunctionToCall::relocate_from_to, &f.state_, &state_);
-    manager_ = f.manager_;
-    invoker_ = f.invoker_;
-
-    f.manager_ = EmptyManager;
-    f.invoker_ = nullptr;
-  }
-
-  template <TargetType target_type, class QualDecayedTRef, class F,
-            absl::enable_if_t<
-                target_type == TargetType::kIncompatibleAnyInvocable, int> = 0>
-  void Initialize(F&& f) {
-    if (f.HasValue()) {
-      InitializeStorage<QualDecayedTRef>(std::forward<F>(f));
-    } else {
-      manager_ = EmptyManager;
-      invoker_ = nullptr;
-    }
-  }
-
-  template <TargetType target_type, class QualDecayedTRef, class F,
-            typename = absl::enable_if_t<target_type == TargetType::kOther>>
-  void Initialize(F&& f) {
-    InitializeStorage<QualDecayedTRef>(std::forward<F>(f));
-  }
-
-  // Use local (inline) storage for applicable target object types.
-  template <class QualTRef, class... Args,
-            typename = absl::enable_if_t<
-                IsStoredLocally<RemoveCVRef<QualTRef>>::value>>
-  void InitializeStorage(Args&&... args) {
-    using RawT = RemoveCVRef<QualTRef>;
-    ::new (static_cast<void*>(&state_.storage))
-        RawT(std::forward<Args>(args)...);
-
-    invoker_ = LocalInvoker<SigIsNoexcept, ReturnType, QualTRef, P...>;
-    // We can simplify our manager if we know the type is trivially copyable.
-    InitializeLocalManager<RawT>();
-  }
-
-  // Use remote storage for target objects that cannot be stored locally.
-  template <class QualTRef, class... Args,
-            absl::enable_if_t<!IsStoredLocally<RemoveCVRef<QualTRef>>::value,
-                              int> = 0>
-  void InitializeStorage(Args&&... args) {
-    InitializeRemoteManager<RemoveCVRef<QualTRef>>(std::forward<Args>(args)...);
-    // This is set after everything else in case an exception is thrown in an
-    // earlier step of the initialization.
-    invoker_ = RemoteInvoker<SigIsNoexcept, ReturnType, QualTRef, P...>;
-  }
-
-  template <class T,
-            typename = absl::enable_if_t<std::is_trivially_copyable<T>::value>>
-  void InitializeLocalManager() {
-    manager_ = LocalManagerTrivial;
-  }
-
-  template <class T,
-            absl::enable_if_t<!std::is_trivially_copyable<T>::value, int> = 0>
-  void InitializeLocalManager() {
-    manager_ = LocalManagerNontrivial<T>;
-  }
-
-  template <class T>
-  using HasTrivialRemoteStorage =
-      std::integral_constant<bool, std::is_trivially_destructible<T>::value &&
-                                       alignof(T) <=
-                                           ABSL_INTERNAL_DEFAULT_NEW_ALIGNMENT>;
-
-  template <class T, class... Args,
-            typename = absl::enable_if_t<HasTrivialRemoteStorage<T>::value>>
-  void InitializeRemoteManager(Args&&... args) {
-    // unique_ptr is used for exception-safety in case construction throws.
-    std::unique_ptr<void, TrivialDeleter> uninitialized_target(
-        ::operator new(sizeof(T)), TrivialDeleter(sizeof(T)));
-    ::new (uninitialized_target.get()) T(std::forward<Args>(args)...);
-    state_.remote.target = uninitialized_target.release();
-    state_.remote.size = sizeof(T);
-    manager_ = RemoteManagerTrivial;
-  }
-
-  template <class T, class... Args,
-            absl::enable_if_t<!HasTrivialRemoteStorage<T>::value, int> = 0>
-  void InitializeRemoteManager(Args&&... args) {
-    state_.remote.target = ::new T(std::forward<Args>(args)...);
-    manager_ = RemoteManagerNontrivial<T>;
-  }
-
-  //////////////////////////////////////////////////////////////////////////////
-  //
-  // Type trait to determine if the template argument is an AnyInvocable whose
-  // function type is compatible enough with ours such that we can
-  // "move the guts" out of it when moving, rather than having to place a new
-  // object into remote storage.
-
-  template <typename Other>
-  struct IsCompatibleAnyInvocable {
-    static constexpr bool value = false;
-  };
-
-  template <typename Sig>
-  struct IsCompatibleAnyInvocable<AnyInvocable<Sig>> {
-    static constexpr bool value =
-        (IsCompatibleConversion)(static_cast<
-                                     typename AnyInvocable<Sig>::CoreImpl*>(
-                                     nullptr),
-                                 static_cast<CoreImpl*>(nullptr));
-  };
-
-  //
-  //////////////////////////////////////////////////////////////////////////////
-
-  TypeErasedState state_;
-  ManagerType* manager_;
-  InvokerType<SigIsNoexcept, ReturnType, P...>* invoker_;
-};
-
-// A constructor name-tag used with Impl to request the
-// conversion-constructor
-struct ConversionConstruct {};
-
-////////////////////////////////////////////////////////////////////////////////
-//
-// A metafunction that is normally an identity metafunction except that when
-// given a std::reference_wrapper<T>, it yields T&. This is necessary because
-// currently std::reference_wrapper's operator() is not conditionally noexcept,
-// so when checking if such an Invocable is nothrow-invocable, we must pull out
-// the underlying type.
-template <class T>
-struct UnwrapStdReferenceWrapperImpl {
-  using type = T;
-};
-
-template <class T>
-struct UnwrapStdReferenceWrapperImpl<std::reference_wrapper<T>> {
-  using type = T&;
-};
-
-template <class T>
-using UnwrapStdReferenceWrapper =
-    typename UnwrapStdReferenceWrapperImpl<T>::type;
-//
-////////////////////////////////////////////////////////////////////////////////
-
-// An alias that always yields std::true_type (used with constraints) where
-// substitution failures happen when forming the template arguments.
-template <class... T>
-using TrueAlias =
-    std::integral_constant<bool, sizeof(absl::void_t<T...>*) != 0>;
-
-/*SFINAE constraints for the conversion-constructor.*/
-template <class Sig, class F,
-          class = absl::enable_if_t<
-              !std::is_same<RemoveCVRef<F>, AnyInvocable<Sig>>::value>>
-using CanConvert = TrueAlias<
-    absl::enable_if_t<!IsInPlaceType<RemoveCVRef<F>>::value>,
-    absl::enable_if_t<Impl<Sig>::template CallIsValid<F>::value>,
-    absl::enable_if_t<
-        Impl<Sig>::template CallIsNoexceptIfSigIsNoexcept<F>::value>,
-    absl::enable_if_t<std::is_constructible<absl::decay_t<F>, F>::value>>;
-
-/*SFINAE constraints for the std::in_place constructors.*/
-template <class Sig, class F, class... Args>
-using CanEmplace = TrueAlias<
-    absl::enable_if_t<Impl<Sig>::template CallIsValid<F>::value>,
-    absl::enable_if_t<
-        Impl<Sig>::template CallIsNoexceptIfSigIsNoexcept<F>::value>,
-    absl::enable_if_t<std::is_constructible<absl::decay_t<F>, Args...>::value>>;
-
-/*SFINAE constraints for the conversion-assign operator.*/
-template <class Sig, class F,
-          class = absl::enable_if_t<
-              !std::is_same<RemoveCVRef<F>, AnyInvocable<Sig>>::value>>
-using CanAssign = TrueAlias<
-    absl::enable_if_t<Impl<Sig>::template CallIsValid<F>::value>,
-    absl::enable_if_t<
-        Impl<Sig>::template CallIsNoexceptIfSigIsNoexcept<F>::value>,
-    absl::enable_if_t<std::is_constructible<absl::decay_t<F>, F>::value>>;
-
-/*SFINAE constraints for the reference-wrapper conversion-assign operator.*/
-template <class Sig, class F>
-using CanAssignReferenceWrapper = TrueAlias<
-    absl::enable_if_t<
-        Impl<Sig>::template CallIsValid<std::reference_wrapper<F>>::value>,
-    absl::enable_if_t<Impl<Sig>::template CallIsNoexceptIfSigIsNoexcept<
-        std::reference_wrapper<F>>::value>>;
-
-////////////////////////////////////////////////////////////////////////////////
-//
-// The constraint for checking whether or not a call meets the noexcept
-// callability requirements. This is a preprocessor macro because specifying it
-// this way as opposed to a disjunction/branch can improve the user-side error
-// messages and avoids an instantiation of std::is_nothrow_invocable_r in the
-// cases where the user did not specify a noexcept function type.
-//
-#define ABSL_INTERNAL_ANY_INVOCABLE_NOEXCEPT_CONSTRAINT(inv_quals, noex) \
-  ABSL_INTERNAL_ANY_INVOCABLE_NOEXCEPT_CONSTRAINT_##noex(inv_quals)
-
-// The disjunction below is because we can't rely on std::is_nothrow_invocable_r
-// to give the right result when ReturnType is non-moveable in toolchains that
-// don't treat non-moveable result types correctly. For example this was the
-// case in libc++ before commit c3a24882 (2022-05).
-#define ABSL_INTERNAL_ANY_INVOCABLE_NOEXCEPT_CONSTRAINT_true(inv_quals)      \
-  absl::enable_if_t<absl::disjunction<                                       \
-      std::is_nothrow_invocable_r<                                           \
-          ReturnType, UnwrapStdReferenceWrapper<absl::decay_t<F>> inv_quals, \
-          P...>,                                                             \
-      std::conjunction<                                                      \
-          std::is_nothrow_invocable<                                         \
-              UnwrapStdReferenceWrapper<absl::decay_t<F>> inv_quals, P...>,  \
-          std::is_same<                                                      \
-              ReturnType,                                                    \
-              absl::base_internal::invoke_result_t<                          \
-                  UnwrapStdReferenceWrapper<absl::decay_t<F>> inv_quals,     \
-                  P...>>>>::value>
-
-#define ABSL_INTERNAL_ANY_INVOCABLE_NOEXCEPT_CONSTRAINT_false(inv_quals)
-//
-////////////////////////////////////////////////////////////////////////////////
-
-// A macro to generate partial specializations of Impl with the different
-// combinations of supported cv/reference qualifiers and noexcept specifier.
-//
-// Here, `cv` are the cv-qualifiers if any, `ref` is the ref-qualifier if any,
-// inv_quals is the reference type to be used when invoking the target, and
-// noex is "true" if the function type is noexcept, or false if it is not.
-//
-// The CallIsValid condition is more complicated than simply using
-// absl::base_internal::is_invocable_r because we can't rely on it to give the
-// right result when ReturnType is non-moveable in toolchains that don't treat
-// non-moveable result types correctly. For example this was the case in libc++
-// before commit c3a24882 (2022-05).
-#define ABSL_INTERNAL_ANY_INVOCABLE_IMPL_(cv, ref, inv_quals, noex)            \
-  template <class ReturnType, class... P>                                      \
-  class Impl<ReturnType(P...) cv ref ABSL_INTERNAL_NOEXCEPT_SPEC(noex)>        \
-      : public CoreImpl<noex, ReturnType, P...> {                              \
-   public:                                                                     \
-    /*The base class, which contains the datamembers and core operations*/     \
-    using Core = CoreImpl<noex, ReturnType, P...>;                             \
-                                                                               \
-    /*SFINAE constraint to check if F is invocable with the proper signature*/ \
-    template <class F>                                                         \
-    using CallIsValid = TrueAlias<absl::enable_if_t<absl::disjunction<         \
-        absl::base_internal::is_invocable_r<ReturnType,                        \
-                                            absl::decay_t<F> inv_quals, P...>, \
-        std::is_same<ReturnType,                                               \
-                     absl::base_internal::invoke_result_t<                     \
-                         absl::decay_t<F> inv_quals, P...>>>::value>>;         \
-                                                                               \
-    /*SFINAE constraint to check if F is nothrow-invocable when necessary*/    \
-    template <class F>                                                         \
-    using CallIsNoexceptIfSigIsNoexcept =                                      \
-        TrueAlias<ABSL_INTERNAL_ANY_INVOCABLE_NOEXCEPT_CONSTRAINT(inv_quals,   \
-                                                                  noex)>;      \
-                                                                               \
-    /*Put the AnyInvocable into an empty state.*/                              \
-    Impl() = default;                                                          \
-                                                                               \
-    /*The implementation of a conversion-constructor from "f*/                 \
-    /*This forwards to Core, attaching inv_quals so that the base class*/      \
-    /*knows how to properly type-erase the invocation.*/                       \
-    template <class F>                                                         \
-    explicit Impl(ConversionConstruct, F&& f)                                  \
-        : Core(TypedConversionConstruct<                                       \
-                   typename std::decay<F>::type inv_quals>(),                  \
-               std::forward<F>(f)) {}                                          \
-                                                                               \
-    /*Forward along the in-place construction parameters.*/                    \
-    template <class T, class... Args>                                          \
-    explicit Impl(absl::in_place_type_t<T>, Args&&... args)                    \
-        : Core(absl::in_place_type<absl::decay_t<T> inv_quals>,                \
-               std::forward<Args>(args)...) {}                                 \
-                                                                               \
-    /*Raises a fatal error when the AnyInvocable is invoked after a move*/     \
-    static ReturnType InvokedAfterMove(                                        \
-      TypeErasedState*,                                                        \
-      ForwardedParameterType<P>...) noexcept(noex) {                           \
-      ABSL_HARDENING_ASSERT(false && "AnyInvocable use-after-move");           \
-      std::terminate();                                                        \
-    }                                                                          \
-                                                                               \
-    InvokerType<noex, ReturnType, P...>* ExtractInvoker() cv {                 \
-      using QualifiedTestType = int cv ref;                                    \
-      auto* invoker = this->invoker_;                                          \
-      if (!std::is_const<QualifiedTestType>::value &&                          \
-          std::is_rvalue_reference<QualifiedTestType>::value) {                \
-        ABSL_ASSERT([this]() {                                                 \
-          /* We checked that this isn't const above, so const_cast is safe */  \
-          const_cast<Impl*>(this)->invoker_ = InvokedAfterMove;                \
-          return this->HasValue();                                             \
-        }());                                                                  \
-      }                                                                        \
-      return invoker;                                                          \
-    }                                                                          \
-                                                                               \
-    /*The actual invocation operation with the proper signature*/              \
-    ReturnType operator()(P... args) cv ref noexcept(noex) {                   \
-      assert(this->invoker_ != nullptr);                                       \
-      return this->ExtractInvoker()(                                           \
-          const_cast<TypeErasedState*>(&this->state_),                         \
-          static_cast<ForwardedParameterType<P>>(args)...);                    \
-    }                                                                          \
-  }
-
-// Define the `noexcept(true)` specialization only for C++17 and beyond, when
-// `noexcept` is part of the type system.
-#if ABSL_INTERNAL_CPLUSPLUS_LANG >= 201703L
-// A convenience macro that defines specializations for the noexcept(true) and
-// noexcept(false) forms, given the other properties.
-#define ABSL_INTERNAL_ANY_INVOCABLE_IMPL(cv, ref, inv_quals)    \
-  ABSL_INTERNAL_ANY_INVOCABLE_IMPL_(cv, ref, inv_quals, false); \
-  ABSL_INTERNAL_ANY_INVOCABLE_IMPL_(cv, ref, inv_quals, true)
-#else
-#define ABSL_INTERNAL_ANY_INVOCABLE_IMPL(cv, ref, inv_quals) \
-  ABSL_INTERNAL_ANY_INVOCABLE_IMPL_(cv, ref, inv_quals, false)
-#endif
-
-// Non-ref-qualified partial specializations
-ABSL_INTERNAL_ANY_INVOCABLE_IMPL(, , &);
-ABSL_INTERNAL_ANY_INVOCABLE_IMPL(const, , const&);
-
-// Lvalue-ref-qualified partial specializations
-ABSL_INTERNAL_ANY_INVOCABLE_IMPL(, &, &);
-ABSL_INTERNAL_ANY_INVOCABLE_IMPL(const, &, const&);
-
-// Rvalue-ref-qualified partial specializations
-ABSL_INTERNAL_ANY_INVOCABLE_IMPL(, &&, &&);
-ABSL_INTERNAL_ANY_INVOCABLE_IMPL(const, &&, const&&);
-
-// Undef the detail-only macros.
-#undef ABSL_INTERNAL_ANY_INVOCABLE_IMPL
-#undef ABSL_INTERNAL_ANY_INVOCABLE_IMPL_
-#undef ABSL_INTERNAL_ANY_INVOCABLE_NOEXCEPT_CONSTRAINT_false
-#undef ABSL_INTERNAL_ANY_INVOCABLE_NOEXCEPT_CONSTRAINT_true
-#undef ABSL_INTERNAL_ANY_INVOCABLE_NOEXCEPT_CONSTRAINT
-#undef ABSL_INTERNAL_NOEXCEPT_SPEC
-
-}  // namespace internal_any_invocable
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FUNCTIONAL_INTERNAL_ANY_INVOCABLE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/functional/internal/front_binder.h b/projects/llm_framework/include/abseil-cpp/absl/functional/internal/front_binder.h
deleted file mode 100644
index 44a54928..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/functional/internal/front_binder.h
+++ /dev/null
@@ -1,95 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-// Implementation details for `absl::bind_front()`.
-
-#ifndef ABSL_FUNCTIONAL_INTERNAL_FRONT_BINDER_H_
-#define ABSL_FUNCTIONAL_INTERNAL_FRONT_BINDER_H_
-
-#include <cstddef>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/internal/invoke.h"
-#include "absl/container/internal/compressed_tuple.h"
-#include "absl/meta/type_traits.h"
-#include "absl/utility/utility.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace functional_internal {
-
-// Invoke the method, expanding the tuple of bound arguments.
-template <class R, class Tuple, size_t... Idx, class... Args>
-R Apply(Tuple&& bound, absl::index_sequence<Idx...>, Args&&... free) {
-  return base_internal::invoke(
-      std::forward<Tuple>(bound).template get<Idx>()...,
-      std::forward<Args>(free)...);
-}
-
-template <class F, class... BoundArgs>
-class FrontBinder {
-  using BoundArgsT = absl::container_internal::CompressedTuple<F, BoundArgs...>;
-  using Idx = absl::make_index_sequence<sizeof...(BoundArgs) + 1>;
-
-  BoundArgsT bound_args_;
-
- public:
-  template <class... Ts>
-  constexpr explicit FrontBinder(absl::in_place_t, Ts&&... ts)
-      : bound_args_(std::forward<Ts>(ts)...) {}
-
-  template <class... FreeArgs, class R = base_internal::invoke_result_t<
-                                   F&, BoundArgs&..., FreeArgs&&...>>
-  R operator()(FreeArgs&&... free_args) & {
-    return functional_internal::Apply<R>(bound_args_, Idx(),
-                                         std::forward<FreeArgs>(free_args)...);
-  }
-
-  template <class... FreeArgs,
-            class R = base_internal::invoke_result_t<
-                const F&, const BoundArgs&..., FreeArgs&&...>>
-  R operator()(FreeArgs&&... free_args) const& {
-    return functional_internal::Apply<R>(bound_args_, Idx(),
-                                         std::forward<FreeArgs>(free_args)...);
-  }
-
-  template <class... FreeArgs, class R = base_internal::invoke_result_t<
-                                   F&&, BoundArgs&&..., FreeArgs&&...>>
-  R operator()(FreeArgs&&... free_args) && {
-    // This overload is called when *this is an rvalue. If some of the bound
-    // arguments are stored by value or rvalue reference, we move them.
-    return functional_internal::Apply<R>(std::move(bound_args_), Idx(),
-                                         std::forward<FreeArgs>(free_args)...);
-  }
-
-  template <class... FreeArgs,
-            class R = base_internal::invoke_result_t<
-                const F&&, const BoundArgs&&..., FreeArgs&&...>>
-  R operator()(FreeArgs&&... free_args) const&& {
-    // This overload is called when *this is an rvalue. If some of the bound
-    // arguments are stored by value or rvalue reference, we move them.
-    return functional_internal::Apply<R>(std::move(bound_args_), Idx(),
-                                         std::forward<FreeArgs>(free_args)...);
-  }
-};
-
-template <class F, class... BoundArgs>
-using bind_front_t = FrontBinder<decay_t<F>, absl::decay_t<BoundArgs>...>;
-
-}  // namespace functional_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FUNCTIONAL_INTERNAL_FRONT_BINDER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/functional/internal/function_ref.h b/projects/llm_framework/include/abseil-cpp/absl/functional/internal/function_ref.h
deleted file mode 100644
index 1cd34a3c..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/functional/internal/function_ref.h
+++ /dev/null
@@ -1,116 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_FUNCTIONAL_INTERNAL_FUNCTION_REF_H_
-#define ABSL_FUNCTIONAL_INTERNAL_FUNCTION_REF_H_
-
-#include <cassert>
-#include <functional>
-#include <type_traits>
-
-#include "absl/base/internal/invoke.h"
-#include "absl/functional/any_invocable.h"
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace functional_internal {
-
-// Like a void* that can handle function pointers as well. The standard does not
-// allow function pointers to round-trip through void*, but void(*)() is fine.
-//
-// Note: It's important that this class remains trivial and is the same size as
-// a pointer, since this allows the compiler to perform tail-call optimizations
-// when the underlying function is a callable object with a matching signature.
-union VoidPtr {
-  const void* obj;
-  void (*fun)();
-};
-
-// Chooses the best type for passing T as an argument.
-// Attempt to be close to SystemV AMD64 ABI. Objects with trivial copy ctor are
-// passed by value.
-template <typename T,
-          bool IsLValueReference = std::is_lvalue_reference<T>::value>
-struct PassByValue : std::false_type {};
-
-template <typename T>
-struct PassByValue<T, /*IsLValueReference=*/false>
-    : std::integral_constant<bool,
-                             absl::is_trivially_copy_constructible<T>::value &&
-                                 absl::is_trivially_copy_assignable<
-                                     typename std::remove_cv<T>::type>::value &&
-                                 std::is_trivially_destructible<T>::value &&
-                                 sizeof(T) <= 2 * sizeof(void*)> {};
-
-template <typename T>
-struct ForwardT : std::conditional<PassByValue<T>::value, T, T&&> {};
-
-// An Invoker takes a pointer to the type-erased invokable object, followed by
-// the arguments that the invokable object expects.
-//
-// Note: The order of arguments here is an optimization, since member functions
-// have an implicit "this" pointer as their first argument, putting VoidPtr
-// first allows the compiler to perform tail-call optimization in many cases.
-template <typename R, typename... Args>
-using Invoker = R (*)(VoidPtr, typename ForwardT<Args>::type...);
-
-//
-// InvokeObject and InvokeFunction provide static "Invoke" functions that can be
-// used as Invokers for objects or functions respectively.
-//
-// static_cast<R> handles the case the return type is void.
-template <typename Obj, typename R, typename... Args>
-R InvokeObject(VoidPtr ptr, typename ForwardT<Args>::type... args) {
-  auto o = static_cast<const Obj*>(ptr.obj);
-  return static_cast<R>(
-      absl::base_internal::invoke(*o, std::forward<Args>(args)...));
-}
-
-template <typename Fun, typename R, typename... Args>
-R InvokeFunction(VoidPtr ptr, typename ForwardT<Args>::type... args) {
-  auto f = reinterpret_cast<Fun>(ptr.fun);
-  return static_cast<R>(
-      absl::base_internal::invoke(f, std::forward<Args>(args)...));
-}
-
-template <typename Sig>
-void AssertNonNull(const std::function<Sig>& f) {
-  assert(f != nullptr);
-  (void)f;
-}
-
-template <typename Sig>
-void AssertNonNull(const AnyInvocable<Sig>& f) {
-  assert(f != nullptr);
-  (void)f;
-}
-
-template <typename F>
-void AssertNonNull(const F&) {}
-
-template <typename F, typename C>
-void AssertNonNull(F C::*f) {
-  assert(f != nullptr);
-  (void)f;
-}
-
-template <bool C>
-using EnableIf = typename ::std::enable_if<C, int>::type;
-
-}  // namespace functional_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FUNCTIONAL_INTERNAL_FUNCTION_REF_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/functional/overload.h b/projects/llm_framework/include/abseil-cpp/absl/functional/overload.h
deleted file mode 100644
index 732795fc..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/functional/overload.h
+++ /dev/null
@@ -1,82 +0,0 @@
-// Copyright 2023 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: overload.h
-// -----------------------------------------------------------------------------
-//
-// `absl::Overload()` returns a functor that provides overloads based on the
-// functors passed to it.
-// Before using this function, consider whether named function overloads would
-// be a better design.
-// One use case for this is locally defining visitors for `std::visit` inside a
-// function using lambdas.
-
-// Example: Using  `absl::Overload` to define a visitor for `std::variant`.
-//
-// std::variant<int, std::string, double> v(int{1});
-//
-// assert(std::visit(absl::Overload(
-//                        [](int) -> absl::string_view { return "int"; },
-//                        [](const std::string&) -> absl::string_view {
-//                          return "string";
-//                        },
-//                        [](double) -> absl::string_view { return "double"; }),
-//                     v) == "int");
-//
-// One of the lambda may specify overload for several types via generic lambda.
-//
-// absl::variant<std::string, int32_t, int64_t> v(int32_t{1});
-// assert(std::visit(absl::Overload(
-//     [](const std::string& s) { return s.size(); },
-//     [](const auto& s) { return sizeof(s); }), v) == 4);
-//
-// Note: absl::Overload requires C++17.
-
-#ifndef ABSL_FUNCTIONAL_OVERLOAD_H_
-#define ABSL_FUNCTIONAL_OVERLOAD_H_
-
-#include "absl/base/config.h"
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-#if defined(ABSL_INTERNAL_CPLUSPLUS_LANG) && \
-    ABSL_INTERNAL_CPLUSPLUS_LANG >= 201703L
-
-template <int&... ExplicitArgumentBarrier, typename... T>
-auto Overload(T&&... ts) {
-  struct OverloadImpl : absl::remove_cvref_t<T>... {
-    using absl::remove_cvref_t<T>::operator()...;
-  };
-  return OverloadImpl{std::forward<T>(ts)...};
-}
-#else
-namespace functional_internal {
-template <typename T>
-constexpr bool kDependentFalse = false;
-}
-
-template <typename Dependent = int, typename... T>
-auto Overload(T&&...) {
-  static_assert(functional_internal::kDependentFalse<Dependent>,
-                "Overload is only usable with C++17 or above.");
-}
-
-#endif
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_FUNCTIONAL_OVERLOAD_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/hash/hash.h b/projects/llm_framework/include/abseil-cpp/absl/hash/hash.h
deleted file mode 100644
index 470cca48..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/hash/hash.h
+++ /dev/null
@@ -1,424 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: hash.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines the Abseil `hash` library and the Abseil hashing
-// framework. This framework consists of the following:
-//
-//   * The `absl::Hash` functor, which is used to invoke the hasher within the
-//     Abseil hashing framework. `absl::Hash<T>` supports most basic types and
-//     a number of Abseil types out of the box.
-//   * `AbslHashValue`, an extension point that allows you to extend types to
-//     support Abseil hashing without requiring you to define a hashing
-//     algorithm.
-//   * `HashState`, a type-erased class which implements the manipulation of the
-//     hash state (H) itself; contains member functions `combine()`,
-//     `combine_contiguous()`, and `combine_unordered()`; and which you can use
-//     to contribute to an existing hash state when hashing your types.
-//
-// Unlike `std::hash` or other hashing frameworks, the Abseil hashing framework
-// provides most of its utility by abstracting away the hash algorithm (and its
-// implementation) entirely. Instead, a type invokes the Abseil hashing
-// framework by simply combining its state with the state of known, hashable
-// types. Hashing of that combined state is separately done by `absl::Hash`.
-//
-// One should assume that a hash algorithm is chosen randomly at the start of
-// each process.  E.g., `absl::Hash<int>{}(9)` in one process and
-// `absl::Hash<int>{}(9)` in another process are likely to differ.
-//
-// `absl::Hash` may also produce different values from different dynamically
-// loaded libraries. For this reason, `absl::Hash` values must never cross
-// boundaries in dynamically loaded libraries (including when used in types like
-// hash containers.)
-//
-// `absl::Hash` is intended to strongly mix input bits with a target of passing
-// an [Avalanche Test](https://en.wikipedia.org/wiki/Avalanche_effect).
-//
-// Example:
-//
-//   // Suppose we have a class `Circle` for which we want to add hashing:
-//   class Circle {
-//    public:
-//     ...
-//    private:
-//     std::pair<int, int> center_;
-//     int radius_;
-//   };
-//
-//   // To add hashing support to `Circle`, we simply need to add a free
-//   // (non-member) function `AbslHashValue()`, and return the combined hash
-//   // state of the existing hash state and the class state. You can add such a
-//   // free function using a friend declaration within the body of the class:
-//   class Circle {
-//    public:
-//     ...
-//     template <typename H>
-//     friend H AbslHashValue(H h, const Circle& c) {
-//       return H::combine(std::move(h), c.center_, c.radius_);
-//     }
-//     ...
-//   };
-//
-// For more information, see Adding Type Support to `absl::Hash` below.
-//
-#ifndef ABSL_HASH_HASH_H_
-#define ABSL_HASH_HASH_H_
-
-#include <tuple>
-#include <utility>
-
-#include "absl/functional/function_ref.h"
-#include "absl/hash/internal/hash.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// -----------------------------------------------------------------------------
-// `absl::Hash`
-// -----------------------------------------------------------------------------
-//
-// `absl::Hash<T>` is a convenient general-purpose hash functor for any type `T`
-// satisfying any of the following conditions (in order):
-//
-//  * T is an arithmetic or pointer type
-//  * T defines an overload for `AbslHashValue(H, const T&)` for an arbitrary
-//    hash state `H`.
-//  - T defines a specialization of `std::hash<T>`
-//
-// `absl::Hash` intrinsically supports the following types:
-//
-//   * All integral types (including bool)
-//   * All enum types
-//   * All floating-point types (although hashing them is discouraged)
-//   * All pointer types, including nullptr_t
-//   * std::pair<T1, T2>, if T1 and T2 are hashable
-//   * std::tuple<Ts...>, if all the Ts... are hashable
-//   * std::unique_ptr and std::shared_ptr
-//   * All string-like types including:
-//     * absl::Cord
-//     * std::string (as well as any instance of std::basic_string that
-//       uses one of {char, wchar_t, char16_t, char32_t} and its associated
-//       std::char_traits)
-//     * std::string_view (as well as any instance of std::basic_string_view
-//       that uses one of {char, wchar_t, char16_t, char32_t} and its associated
-//       std::char_traits)
-//  * All the standard sequence containers (provided the elements are hashable)
-//  * All the standard associative containers (provided the elements are
-//    hashable)
-//  * absl types such as the following:
-//    * absl::string_view
-//    * absl::uint128
-//    * absl::Time, absl::Duration, and absl::TimeZone
-//  * absl containers (provided the elements are hashable) such as the
-//    following:
-//    * absl::flat_hash_set, absl::node_hash_set, absl::btree_set
-//    * absl::flat_hash_map, absl::node_hash_map, absl::btree_map
-//    * absl::btree_multiset, absl::btree_multimap
-//    * absl::InlinedVector
-//    * absl::FixedArray
-//
-// When absl::Hash is used to hash an unordered container with a custom hash
-// functor, the elements are hashed using default absl::Hash semantics, not
-// the custom hash functor.  This is consistent with the behavior of
-// operator==() on unordered containers, which compares elements pairwise with
-// operator==() rather than the custom equality functor.  It is usually a
-// mistake to use either operator==() or absl::Hash on unordered collections
-// that use functors incompatible with operator==() equality.
-//
-// Note: the list above is not meant to be exhaustive. Additional type support
-// may be added, in which case the above list will be updated.
-//
-// -----------------------------------------------------------------------------
-// absl::Hash Invocation Evaluation
-// -----------------------------------------------------------------------------
-//
-// When invoked, `absl::Hash<T>` searches for supplied hash functions in the
-// following order:
-//
-//   * Natively supported types out of the box (see above)
-//   * Types for which an `AbslHashValue()` overload is provided (such as
-//     user-defined types). See "Adding Type Support to `absl::Hash`" below.
-//   * Types which define a `std::hash<T>` specialization
-//
-// The fallback to legacy hash functions exists mainly for backwards
-// compatibility. If you have a choice, prefer defining an `AbslHashValue`
-// overload instead of specializing any legacy hash functors.
-//
-// -----------------------------------------------------------------------------
-// The Hash State Concept, and using `HashState` for Type Erasure
-// -----------------------------------------------------------------------------
-//
-// The `absl::Hash` framework relies on the Concept of a "hash state." Such a
-// hash state is used in several places:
-//
-// * Within existing implementations of `absl::Hash<T>` to store the hashed
-//   state of an object. Note that it is up to the implementation how it stores
-//   such state. A hash table, for example, may mix the state to produce an
-//   integer value; a testing framework may simply hold a vector of that state.
-// * Within implementations of `AbslHashValue()` used to extend user-defined
-//   types. (See "Adding Type Support to absl::Hash" below.)
-// * Inside a `HashState`, providing type erasure for the concept of a hash
-//   state, which you can use to extend the `absl::Hash` framework for types
-//   that are otherwise difficult to extend using `AbslHashValue()`. (See the
-//   `HashState` class below.)
-//
-// The "hash state" concept contains three member functions for mixing hash
-// state:
-//
-// * `H::combine(state, values...)`
-//
-//   Combines an arbitrary number of values into a hash state, returning the
-//   updated state. Note that the existing hash state is move-only and must be
-//   passed by value.
-//
-//   Each of the value types T must be hashable by H.
-//
-//   NOTE:
-//
-//     state = H::combine(std::move(state), value1, value2, value3);
-//
-//   must be guaranteed to produce the same hash expansion as
-//
-//     state = H::combine(std::move(state), value1);
-//     state = H::combine(std::move(state), value2);
-//     state = H::combine(std::move(state), value3);
-//
-// * `H::combine_contiguous(state, data, size)`
-//
-//    Combines a contiguous array of `size` elements into a hash state,
-//    returning the updated state. Note that the existing hash state is
-//    move-only and must be passed by value.
-//
-//    NOTE:
-//
-//      state = H::combine_contiguous(std::move(state), data, size);
-//
-//    need NOT be guaranteed to produce the same hash expansion as a loop
-//    (it may perform internal optimizations). If you need this guarantee, use a
-//    loop instead.
-//
-// * `H::combine_unordered(state, begin, end)`
-//
-//    Combines a set of elements denoted by an iterator pair into a hash
-//    state, returning the updated state.  Note that the existing hash
-//    state is move-only and must be passed by value.
-//
-//    Unlike the other two methods, the hashing is order-independent.
-//    This can be used to hash unordered collections.
-//
-// -----------------------------------------------------------------------------
-// Adding Type Support to `absl::Hash`
-// -----------------------------------------------------------------------------
-//
-// To add support for your user-defined type, add a proper `AbslHashValue()`
-// overload as a free (non-member) function. The overload will take an
-// existing hash state and should combine that state with state from the type.
-//
-// Example:
-//
-//   template <typename H>
-//   H AbslHashValue(H state, const MyType& v) {
-//     return H::combine(std::move(state), v.field1, ..., v.fieldN);
-//   }
-//
-// where `(field1, ..., fieldN)` are the members you would use on your
-// `operator==` to define equality.
-//
-// Notice that `AbslHashValue` is not a class member, but an ordinary function.
-// An `AbslHashValue` overload for a type should only be declared in the same
-// file and namespace as said type. The proper `AbslHashValue` implementation
-// for a given type will be discovered via ADL.
-//
-// Note: unlike `std::hash', `absl::Hash` should never be specialized. It must
-// only be extended by adding `AbslHashValue()` overloads.
-//
-template <typename T>
-using Hash = absl::hash_internal::Hash<T>;
-
-// HashOf
-//
-// absl::HashOf() is a helper that generates a hash from the values of its
-// arguments.  It dispatches to absl::Hash directly, as follows:
-//  * HashOf(t) == absl::Hash<T>{}(t)
-//  * HashOf(a, b, c) == HashOf(std::make_tuple(a, b, c))
-//
-// HashOf(a1, a2, ...) == HashOf(b1, b2, ...) is guaranteed when
-//  * The argument lists have pairwise identical C++ types
-//  * a1 == b1 && a2 == b2 && ...
-//
-// The requirement that the arguments match in both type and value is critical.
-// It means that `a == b` does not necessarily imply `HashOf(a) == HashOf(b)` if
-// `a` and `b` have different types. For example, `HashOf(2) != HashOf(2.0)`.
-template <int&... ExplicitArgumentBarrier, typename... Types>
-size_t HashOf(const Types&... values) {
-  auto tuple = std::tie(values...);
-  return absl::Hash<decltype(tuple)>{}(tuple);
-}
-
-// HashState
-//
-// A type erased version of the hash state concept, for use in user-defined
-// `AbslHashValue` implementations that can't use templates (such as PImpl
-// classes, virtual functions, etc.). The type erasure adds overhead so it
-// should be avoided unless necessary.
-//
-// Note: This wrapper will only erase calls to
-//     combine_contiguous(H, const unsigned char*, size_t)
-//     RunCombineUnordered(H, CombinerF)
-//
-// All other calls will be handled internally and will not invoke overloads
-// provided by the wrapped class.
-//
-// Users of this class should still define a template `AbslHashValue` function,
-// but can use `absl::HashState::Create(&state)` to erase the type of the hash
-// state and dispatch to their private hashing logic.
-//
-// This state can be used like any other hash state. In particular, you can call
-// `HashState::combine()` and `HashState::combine_contiguous()` on it.
-//
-// Example:
-//
-//   class Interface {
-//    public:
-//     template <typename H>
-//     friend H AbslHashValue(H state, const Interface& value) {
-//       state = H::combine(std::move(state), std::type_index(typeid(*this)));
-//       value.HashValue(absl::HashState::Create(&state));
-//       return state;
-//     }
-//    private:
-//     virtual void HashValue(absl::HashState state) const = 0;
-//   };
-//
-//   class Impl : Interface {
-//    private:
-//     void HashValue(absl::HashState state) const override {
-//       absl::HashState::combine(std::move(state), v1_, v2_);
-//     }
-//     int v1_;
-//     std::string v2_;
-//   };
-class HashState : public hash_internal::HashStateBase<HashState> {
- public:
-  // HashState::Create()
-  //
-  // Create a new `HashState` instance that wraps `state`. All calls to
-  // `combine()` and `combine_contiguous()` on the new instance will be
-  // redirected to the original `state` object. The `state` object must outlive
-  // the `HashState` instance.
-  template <typename T>
-  static HashState Create(T* state) {
-    HashState s;
-    s.Init(state);
-    return s;
-  }
-
-  HashState(const HashState&) = delete;
-  HashState& operator=(const HashState&) = delete;
-  HashState(HashState&&) = default;
-  HashState& operator=(HashState&&) = default;
-
-  // HashState::combine()
-  //
-  // Combines an arbitrary number of values into a hash state, returning the
-  // updated state.
-  using HashState::HashStateBase::combine;
-
-  // HashState::combine_contiguous()
-  //
-  // Combines a contiguous array of `size` elements into a hash state, returning
-  // the updated state.
-  static HashState combine_contiguous(HashState hash_state,
-                                      const unsigned char* first, size_t size) {
-    hash_state.combine_contiguous_(hash_state.state_, first, size);
-    return hash_state;
-  }
-  using HashState::HashStateBase::combine_contiguous;
-
- private:
-  HashState() = default;
-
-  friend class HashState::HashStateBase;
-
-  template <typename T>
-  static void CombineContiguousImpl(void* p, const unsigned char* first,
-                                    size_t size) {
-    T& state = *static_cast<T*>(p);
-    state = T::combine_contiguous(std::move(state), first, size);
-  }
-
-  template <typename T>
-  void Init(T* state) {
-    state_ = state;
-    combine_contiguous_ = &CombineContiguousImpl<T>;
-    run_combine_unordered_ = &RunCombineUnorderedImpl<T>;
-  }
-
-  template <typename HS>
-  struct CombineUnorderedInvoker {
-    template <typename T, typename ConsumerT>
-    void operator()(T inner_state, ConsumerT inner_cb) {
-      f(HashState::Create(&inner_state),
-        [&](HashState& inner_erased) { inner_cb(inner_erased.Real<T>()); });
-    }
-
-    absl::FunctionRef<void(HS, absl::FunctionRef<void(HS&)>)> f;
-  };
-
-  template <typename T>
-  static HashState RunCombineUnorderedImpl(
-      HashState state,
-      absl::FunctionRef<void(HashState, absl::FunctionRef<void(HashState&)>)>
-          f) {
-    // Note that this implementation assumes that inner_state and outer_state
-    // are the same type.  This isn't true in the SpyHash case, but SpyHash
-    // types are move-convertible to each other, so this still works.
-    T& real_state = state.Real<T>();
-    real_state = T::RunCombineUnordered(
-        std::move(real_state), CombineUnorderedInvoker<HashState>{f});
-    return state;
-  }
-
-  template <typename CombinerT>
-  static HashState RunCombineUnordered(HashState state, CombinerT combiner) {
-    auto* run = state.run_combine_unordered_;
-    return run(std::move(state), std::ref(combiner));
-  }
-
-  // Do not erase an already erased state.
-  void Init(HashState* state) {
-    state_ = state->state_;
-    combine_contiguous_ = state->combine_contiguous_;
-    run_combine_unordered_ = state->run_combine_unordered_;
-  }
-
-  template <typename T>
-  T& Real() {
-    return *static_cast<T*>(state_);
-  }
-
-  void* state_;
-  void (*combine_contiguous_)(void*, const unsigned char*, size_t);
-  HashState (*run_combine_unordered_)(
-      HashState state,
-      absl::FunctionRef<void(HashState, absl::FunctionRef<void(HashState&)>)>);
-};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_HASH_HASH_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/hash/hash_testing.h b/projects/llm_framework/include/abseil-cpp/absl/hash/hash_testing.h
deleted file mode 100644
index 1e1c5741..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/hash/hash_testing.h
+++ /dev/null
@@ -1,378 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_HASH_HASH_TESTING_H_
-#define ABSL_HASH_HASH_TESTING_H_
-
-#include <initializer_list>
-#include <tuple>
-#include <type_traits>
-#include <vector>
-
-#include "gmock/gmock.h"
-#include "gtest/gtest.h"
-#include "absl/hash/internal/spy_hash_state.h"
-#include "absl/meta/type_traits.h"
-#include "absl/strings/str_cat.h"
-#include "absl/types/variant.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// Run the absl::Hash algorithm over all the elements passed in and verify that
-// their hash expansion is congruent with their `==` operator.
-//
-// It is used in conjunction with EXPECT_TRUE. Failures will output information
-// on what requirement failed and on which objects.
-//
-// Users should pass a collection of types as either an initializer list or a
-// container of cases.
-//
-//   EXPECT_TRUE(absl::VerifyTypeImplementsAbslHashCorrectly(
-//       {v1, v2, ..., vN}));
-//
-//   std::vector<MyType> cases;
-//   // Fill cases...
-//   EXPECT_TRUE(absl::VerifyTypeImplementsAbslHashCorrectly(cases));
-//
-// Users can pass a variety of types for testing heterogeneous lookup with
-// `std::make_tuple`:
-//
-//   EXPECT_TRUE(absl::VerifyTypeImplementsAbslHashCorrectly(
-//       std::make_tuple(v1, v2, ..., vN)));
-//
-//
-// Ideally, the values passed should provide enough coverage of the `==`
-// operator and the AbslHashValue implementations.
-// For dynamically sized types, the empty state should usually be included in
-// the values.
-//
-// The function accepts an optional comparator function, in case that `==` is
-// not enough for the values provided.
-//
-// Usage:
-//
-//   EXPECT_TRUE(absl::VerifyTypeImplementsAbslHashCorrectly(
-//       std::make_tuple(v1, v2, ..., vN), MyCustomEq{}));
-//
-// It checks the following requirements:
-//   1. The expansion for a value is deterministic.
-//   2. For any two objects `a` and `b` in the sequence, if `a == b` evaluates
-//      to true, then their hash expansion must be equal.
-//   3. If `a == b` evaluates to false their hash expansion must be unequal.
-//   4. If `a == b` evaluates to false neither hash expansion can be a
-//      suffix of the other.
-//   5. AbslHashValue overloads should not be called by the user. They are only
-//      meant to be called by the framework. Users should call H::combine() and
-//      H::combine_contiguous().
-//   6. No moved-from instance of the hash state is used in the implementation
-//      of AbslHashValue.
-//
-// The values do not have to have the same type. This can be useful for
-// equivalent types that support heterogeneous lookup.
-//
-// A possible reason for breaking (2) is combining state in the hash expansion
-// that was not used in `==`.
-// For example:
-//
-// struct Bad2 {
-//   int a, b;
-//   template <typename H>
-//   friend H AbslHashValue(H state, Bad2 x) {
-//     // Uses a and b.
-//     return H::combine(std::move(state), x.a, x.b);
-//   }
-//   friend bool operator==(Bad2 x, Bad2 y) {
-//     // Only uses a.
-//     return x.a == y.a;
-//   }
-// };
-//
-// As for (3), breaking this usually means that there is state being passed to
-// the `==` operator that is not used in the hash expansion.
-// For example:
-//
-// struct Bad3 {
-//   int a, b;
-//   template <typename H>
-//   friend H AbslHashValue(H state, Bad3 x) {
-//     // Only uses a.
-//     return H::combine(std::move(state), x.a);
-//   }
-//   friend bool operator==(Bad3 x, Bad3 y) {
-//     // Uses a and b.
-//     return x.a == y.a && x.b == y.b;
-//   }
-// };
-//
-// Finally, a common way to break 4 is by combining dynamic ranges without
-// combining the size of the range.
-// For example:
-//
-// struct Bad4 {
-//   int *p, size;
-//   template <typename H>
-//   friend H AbslHashValue(H state, Bad4 x) {
-//     return H::combine_contiguous(std::move(state), x.p, x.p + x.size);
-//   }
-//   friend bool operator==(Bad4 x, Bad4 y) {
-//    // Compare two ranges for equality. C++14 code can instead use std::equal.
-//     return absl::equal(x.p, x.p + x.size, y.p, y.p + y.size);
-//   }
-// };
-//
-// An easy solution to this is to combine the size after combining the range,
-// like so:
-// template <typename H>
-// friend H AbslHashValue(H state, Bad4 x) {
-//   return H::combine(
-//       H::combine_contiguous(std::move(state), x.p, x.p + x.size), x.size);
-// }
-//
-template <int&... ExplicitBarrier, typename Container>
-ABSL_MUST_USE_RESULT testing::AssertionResult
-VerifyTypeImplementsAbslHashCorrectly(const Container& values);
-
-template <int&... ExplicitBarrier, typename Container, typename Eq>
-ABSL_MUST_USE_RESULT testing::AssertionResult
-VerifyTypeImplementsAbslHashCorrectly(const Container& values, Eq equals);
-
-template <int&..., typename T>
-ABSL_MUST_USE_RESULT testing::AssertionResult
-VerifyTypeImplementsAbslHashCorrectly(std::initializer_list<T> values);
-
-template <int&..., typename T, typename Eq>
-ABSL_MUST_USE_RESULT testing::AssertionResult
-VerifyTypeImplementsAbslHashCorrectly(std::initializer_list<T> values,
-                                      Eq equals);
-
-namespace hash_internal {
-
-struct PrintVisitor {
-  size_t index;
-  template <typename T>
-  std::string operator()(const T* value) const {
-    return absl::StrCat("#", index, "(", testing::PrintToString(*value), ")");
-  }
-};
-
-template <typename Eq>
-struct EqVisitor {
-  Eq eq;
-  template <typename T, typename U>
-  bool operator()(const T* t, const U* u) const {
-    return eq(*t, *u);
-  }
-};
-
-struct ExpandVisitor {
-  template <typename T>
-  SpyHashState operator()(const T* value) const {
-    return SpyHashState::combine(SpyHashState(), *value);
-  }
-};
-
-template <typename Container, typename Eq>
-ABSL_MUST_USE_RESULT testing::AssertionResult
-VerifyTypeImplementsAbslHashCorrectly(const Container& values, Eq equals) {
-  using V = typename Container::value_type;
-
-  struct Info {
-    const V& value;
-    size_t index;
-    std::string ToString() const {
-      return absl::visit(PrintVisitor{index}, value);
-    }
-    SpyHashState expand() const { return absl::visit(ExpandVisitor{}, value); }
-  };
-
-  using EqClass = std::vector<Info>;
-  std::vector<EqClass> classes;
-
-  // Gather the values in equivalence classes.
-  size_t i = 0;
-  for (const auto& value : values) {
-    EqClass* c = nullptr;
-    for (auto& eqclass : classes) {
-      if (absl::visit(EqVisitor<Eq>{equals}, value, eqclass[0].value)) {
-        c = &eqclass;
-        break;
-      }
-    }
-    if (c == nullptr) {
-      classes.emplace_back();
-      c = &classes.back();
-    }
-    c->push_back({value, i});
-    ++i;
-
-    // Verify potential errors captured by SpyHashState.
-    if (auto error = c->back().expand().error()) {
-      return testing::AssertionFailure() << *error;
-    }
-  }
-
-  if (classes.size() < 2) {
-    return testing::AssertionFailure()
-           << "At least two equivalence classes are expected.";
-  }
-
-  // We assume that equality is correctly implemented.
-  // Now we verify that AbslHashValue is also correctly implemented.
-
-  for (const auto& c : classes) {
-    // All elements of the equivalence class must have the same hash
-    // expansion.
-    const SpyHashState expected = c[0].expand();
-    for (const Info& v : c) {
-      if (v.expand() != v.expand()) {
-        return testing::AssertionFailure()
-               << "Hash expansion for " << v.ToString()
-               << " is non-deterministic.";
-      }
-      if (v.expand() != expected) {
-        return testing::AssertionFailure()
-               << "Values " << c[0].ToString() << " and " << v.ToString()
-               << " evaluate as equal but have an unequal hash expansion.";
-      }
-    }
-
-    // Elements from other classes must have different hash expansion.
-    for (const auto& c2 : classes) {
-      if (&c == &c2) continue;
-      const SpyHashState c2_hash = c2[0].expand();
-      switch (SpyHashState::Compare(expected, c2_hash)) {
-        case SpyHashState::CompareResult::kEqual:
-          return testing::AssertionFailure()
-                 << "Values " << c[0].ToString() << " and " << c2[0].ToString()
-                 << " evaluate as unequal but have an equal hash expansion.";
-        case SpyHashState::CompareResult::kBSuffixA:
-          return testing::AssertionFailure()
-                 << "Hash expansion of " << c2[0].ToString()
-                 << " is a suffix of the hash expansion of " << c[0].ToString()
-                 << ".";
-        case SpyHashState::CompareResult::kASuffixB:
-          return testing::AssertionFailure()
-                 << "Hash expansion of " << c[0].ToString()
-                 << " is a suffix of the hash expansion of " << c2[0].ToString()
-                 << ".";
-        case SpyHashState::CompareResult::kUnequal:
-          break;
-      }
-    }
-  }
-  return testing::AssertionSuccess();
-}
-
-template <typename... T>
-struct TypeSet {
-  template <typename U, bool = disjunction<std::is_same<T, U>...>::value>
-  struct Insert {
-    using type = TypeSet<U, T...>;
-  };
-  template <typename U>
-  struct Insert<U, true> {
-    using type = TypeSet;
-  };
-
-  template <template <typename...> class C>
-  using apply = C<T...>;
-};
-
-template <typename... T>
-struct MakeTypeSet : TypeSet<> {};
-template <typename T, typename... Ts>
-struct MakeTypeSet<T, Ts...> : MakeTypeSet<Ts...>::template Insert<T>::type {};
-
-template <typename... T>
-using VariantForTypes = typename MakeTypeSet<
-    const typename std::decay<T>::type*...>::template apply<absl::variant>;
-
-template <typename Container>
-struct ContainerAsVector {
-  using V = absl::variant<const typename Container::value_type*>;
-  using Out = std::vector<V>;
-
-  static Out Do(const Container& values) {
-    Out out;
-    for (const auto& v : values) out.push_back(&v);
-    return out;
-  }
-};
-
-template <typename... T>
-struct ContainerAsVector<std::tuple<T...>> {
-  using V = VariantForTypes<T...>;
-  using Out = std::vector<V>;
-
-  template <size_t... I>
-  static Out DoImpl(const std::tuple<T...>& tuple, absl::index_sequence<I...>) {
-    return Out{&std::get<I>(tuple)...};
-  }
-
-  static Out Do(const std::tuple<T...>& values) {
-    return DoImpl(values, absl::index_sequence_for<T...>());
-  }
-};
-
-template <>
-struct ContainerAsVector<std::tuple<>> {
-  static std::vector<VariantForTypes<int>> Do(std::tuple<>) { return {}; }
-};
-
-struct DefaultEquals {
-  template <typename T, typename U>
-  bool operator()(const T& t, const U& u) const {
-    return t == u;
-  }
-};
-
-}  // namespace hash_internal
-
-template <int&..., typename Container>
-ABSL_MUST_USE_RESULT testing::AssertionResult
-VerifyTypeImplementsAbslHashCorrectly(const Container& values) {
-  return hash_internal::VerifyTypeImplementsAbslHashCorrectly(
-      hash_internal::ContainerAsVector<Container>::Do(values),
-      hash_internal::DefaultEquals{});
-}
-
-template <int&..., typename Container, typename Eq>
-ABSL_MUST_USE_RESULT testing::AssertionResult
-VerifyTypeImplementsAbslHashCorrectly(const Container& values, Eq equals) {
-  return hash_internal::VerifyTypeImplementsAbslHashCorrectly(
-      hash_internal::ContainerAsVector<Container>::Do(values), equals);
-}
-
-template <int&..., typename T>
-ABSL_MUST_USE_RESULT testing::AssertionResult
-VerifyTypeImplementsAbslHashCorrectly(std::initializer_list<T> values) {
-  return hash_internal::VerifyTypeImplementsAbslHashCorrectly(
-      hash_internal::ContainerAsVector<std::initializer_list<T>>::Do(values),
-      hash_internal::DefaultEquals{});
-}
-
-template <int&..., typename T, typename Eq>
-ABSL_MUST_USE_RESULT testing::AssertionResult
-VerifyTypeImplementsAbslHashCorrectly(std::initializer_list<T> values,
-                                      Eq equals) {
-  return hash_internal::VerifyTypeImplementsAbslHashCorrectly(
-      hash_internal::ContainerAsVector<std::initializer_list<T>>::Do(values),
-      equals);
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_HASH_HASH_TESTING_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/hash/internal/city.h b/projects/llm_framework/include/abseil-cpp/absl/hash/internal/city.h
deleted file mode 100644
index 393da0b9..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/hash/internal/city.h
+++ /dev/null
@@ -1,78 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// https://code.google.com/p/cityhash/
-//
-// This file provides a few functions for hashing strings.  All of them are
-// high-quality functions in the sense that they pass standard tests such
-// as Austin Appleby's SMHasher.  They are also fast.
-//
-// For 64-bit x86 code, on short strings, we don't know of anything faster than
-// CityHash64 that is of comparable quality.  We believe our nearest competitor
-// is Murmur3.  For 64-bit x86 code, CityHash64 is an excellent choice for hash
-// tables and most other hashing (excluding cryptography).
-//
-// For 32-bit x86 code, we don't know of anything faster than CityHash32 that
-// is of comparable quality.  We believe our nearest competitor is Murmur3A.
-// (On 64-bit CPUs, it is typically faster to use the other CityHash variants.)
-//
-// Functions in the CityHash family are not suitable for cryptography.
-//
-// Please see CityHash's README file for more details on our performance
-// measurements and so on.
-//
-// WARNING: This code has been only lightly tested on big-endian platforms!
-// It is known to work well on little-endian platforms that have a small penalty
-// for unaligned reads, such as current Intel and AMD moderate-to-high-end CPUs.
-// It should work on all 32-bit and 64-bit platforms that allow unaligned reads;
-// bug reports are welcome.
-//
-// By the way, for some hash functions, given strings a and b, the hash
-// of a+b is easily derived from the hashes of a and b.  This property
-// doesn't hold for any hash functions in this file.
-
-#ifndef ABSL_HASH_INTERNAL_CITY_H_
-#define ABSL_HASH_INTERNAL_CITY_H_
-
-#include <stdint.h>
-#include <stdlib.h>  // for size_t.
-
-#include <utility>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace hash_internal {
-
-// Hash function for a byte array.
-uint64_t CityHash64(const char *s, size_t len);
-
-// Hash function for a byte array.  For convenience, a 64-bit seed is also
-// hashed into the result.
-uint64_t CityHash64WithSeed(const char *s, size_t len, uint64_t seed);
-
-// Hash function for a byte array.  For convenience, two seeds are also
-// hashed into the result.
-uint64_t CityHash64WithSeeds(const char *s, size_t len, uint64_t seed0,
-                             uint64_t seed1);
-
-// Hash function for a byte array.  Most useful in 32-bit binaries.
-uint32_t CityHash32(const char *s, size_t len);
-
-}  // namespace hash_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_HASH_INTERNAL_CITY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/hash/internal/hash.h b/projects/llm_framework/include/abseil-cpp/absl/hash/internal/hash.h
deleted file mode 100644
index b7d89b01..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/hash/internal/hash.h
+++ /dev/null
@@ -1,1383 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: hash.h
-// -----------------------------------------------------------------------------
-//
-#ifndef ABSL_HASH_INTERNAL_HASH_H_
-#define ABSL_HASH_INTERNAL_HASH_H_
-
-#ifdef __APPLE__
-#include <Availability.h>
-#include <TargetConditionals.h>
-#endif
-
-#include "absl/base/config.h"
-
-// For feature testing and determining which headers can be included.
-#if ABSL_INTERNAL_CPLUSPLUS_LANG >= 202002L
-#include <version>
-#else
-#include <ciso646>
-#endif
-
-#include <algorithm>
-#include <array>
-#include <bitset>
-#include <cmath>
-#include <cstddef>
-#include <cstring>
-#include <deque>
-#include <forward_list>
-#include <functional>
-#include <iterator>
-#include <limits>
-#include <list>
-#include <map>
-#include <memory>
-#include <set>
-#include <string>
-#include <tuple>
-#include <type_traits>
-#include <unordered_map>
-#include <unordered_set>
-#include <utility>
-#include <vector>
-
-#include "absl/base/internal/unaligned_access.h"
-#include "absl/base/port.h"
-#include "absl/container/fixed_array.h"
-#include "absl/hash/internal/city.h"
-#include "absl/hash/internal/low_level_hash.h"
-#include "absl/meta/type_traits.h"
-#include "absl/numeric/bits.h"
-#include "absl/numeric/int128.h"
-#include "absl/strings/string_view.h"
-#include "absl/types/optional.h"
-#include "absl/types/variant.h"
-#include "absl/utility/utility.h"
-
-#if defined(__cpp_lib_filesystem) && __cpp_lib_filesystem >= 201703L && \
-    !defined(_LIBCPP_HAS_NO_FILESYSTEM_LIBRARY)
-#include <filesystem>  // NOLINT
-#endif
-
-#ifdef ABSL_HAVE_STD_STRING_VIEW
-#include <string_view>
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-class HashState;
-
-namespace hash_internal {
-
-// Internal detail: Large buffers are hashed in smaller chunks.  This function
-// returns the size of these chunks.
-constexpr size_t PiecewiseChunkSize() { return 1024; }
-
-// PiecewiseCombiner
-//
-// PiecewiseCombiner is an internal-only helper class for hashing a piecewise
-// buffer of `char` or `unsigned char` as though it were contiguous.  This class
-// provides two methods:
-//
-//   H add_buffer(state, data, size)
-//   H finalize(state)
-//
-// `add_buffer` can be called zero or more times, followed by a single call to
-// `finalize`.  This will produce the same hash expansion as concatenating each
-// buffer piece into a single contiguous buffer, and passing this to
-// `H::combine_contiguous`.
-//
-//  Example usage:
-//    PiecewiseCombiner combiner;
-//    for (const auto& piece : pieces) {
-//      state = combiner.add_buffer(std::move(state), piece.data, piece.size);
-//    }
-//    return combiner.finalize(std::move(state));
-class PiecewiseCombiner {
- public:
-  PiecewiseCombiner() : position_(0) {}
-  PiecewiseCombiner(const PiecewiseCombiner&) = delete;
-  PiecewiseCombiner& operator=(const PiecewiseCombiner&) = delete;
-
-  // PiecewiseCombiner::add_buffer()
-  //
-  // Appends the given range of bytes to the sequence to be hashed, which may
-  // modify the provided hash state.
-  template <typename H>
-  H add_buffer(H state, const unsigned char* data, size_t size);
-  template <typename H>
-  H add_buffer(H state, const char* data, size_t size) {
-    return add_buffer(std::move(state),
-                      reinterpret_cast<const unsigned char*>(data), size);
-  }
-
-  // PiecewiseCombiner::finalize()
-  //
-  // Finishes combining the hash sequence, which may may modify the provided
-  // hash state.
-  //
-  // Once finalize() is called, add_buffer() may no longer be called. The
-  // resulting hash state will be the same as if the pieces passed to
-  // add_buffer() were concatenated into a single flat buffer, and then provided
-  // to H::combine_contiguous().
-  template <typename H>
-  H finalize(H state);
-
- private:
-  unsigned char buf_[PiecewiseChunkSize()];
-  size_t position_;
-};
-
-// is_hashable()
-//
-// Trait class which returns true if T is hashable by the absl::Hash framework.
-// Used for the AbslHashValue implementations for composite types below.
-template <typename T>
-struct is_hashable;
-
-// HashStateBase
-//
-// An internal implementation detail that contains common implementation details
-// for all of the "hash state objects" objects generated by Abseil.  This is not
-// a public API; users should not create classes that inherit from this.
-//
-// A hash state object is the template argument `H` passed to `AbslHashValue`.
-// It represents an intermediate state in the computation of an unspecified hash
-// algorithm. `HashStateBase` provides a CRTP style base class for hash state
-// implementations. Developers adding type support for `absl::Hash` should not
-// rely on any parts of the state object other than the following member
-// functions:
-//
-//   * HashStateBase::combine()
-//   * HashStateBase::combine_contiguous()
-//   * HashStateBase::combine_unordered()
-//
-// A derived hash state class of type `H` must provide a public member function
-// with a signature similar to the following:
-//
-//    `static H combine_contiguous(H state, const unsigned char*, size_t)`.
-//
-// It must also provide a private template method named RunCombineUnordered.
-//
-// A "consumer" is a 1-arg functor returning void.  Its argument is a reference
-// to an inner hash state object, and it may be called multiple times.  When
-// called, the functor consumes the entropy from the provided state object,
-// and resets that object to its empty state.
-//
-// A "combiner" is a stateless 2-arg functor returning void.  Its arguments are
-// an inner hash state object and an ElementStateConsumer functor.  A combiner
-// uses the provided inner hash state object to hash each element of the
-// container, passing the inner hash state object to the consumer after hashing
-// each element.
-//
-// Given these definitions, a derived hash state class of type H
-// must provide a private template method with a signature similar to the
-// following:
-//
-//    `template <typename CombinerT>`
-//    `static H RunCombineUnordered(H outer_state, CombinerT combiner)`
-//
-// This function is responsible for constructing the inner state object and
-// providing a consumer to the combiner.  It uses side effects of the consumer
-// and combiner to mix the state of each element in an order-independent manner,
-// and uses this to return an updated value of `outer_state`.
-//
-// This inside-out approach generates efficient object code in the normal case,
-// but allows us to use stack storage to implement the absl::HashState type
-// erasure mechanism (avoiding heap allocations while hashing).
-//
-// `HashStateBase` will provide a complete implementation for a hash state
-// object in terms of these two methods.
-//
-// Example:
-//
-//   // Use CRTP to define your derived class.
-//   struct MyHashState : HashStateBase<MyHashState> {
-//       static H combine_contiguous(H state, const unsigned char*, size_t);
-//       using MyHashState::HashStateBase::combine;
-//       using MyHashState::HashStateBase::combine_contiguous;
-//       using MyHashState::HashStateBase::combine_unordered;
-//     private:
-//       template <typename CombinerT>
-//       static H RunCombineUnordered(H state, CombinerT combiner);
-//   };
-template <typename H>
-class HashStateBase {
- public:
-  // HashStateBase::combine()
-  //
-  // Combines an arbitrary number of values into a hash state, returning the
-  // updated state.
-  //
-  // Each of the value types `T` must be separately hashable by the Abseil
-  // hashing framework.
-  //
-  // NOTE:
-  //
-  //   state = H::combine(std::move(state), value1, value2, value3);
-  //
-  // is guaranteed to produce the same hash expansion as:
-  //
-  //   state = H::combine(std::move(state), value1);
-  //   state = H::combine(std::move(state), value2);
-  //   state = H::combine(std::move(state), value3);
-  template <typename T, typename... Ts>
-  static H combine(H state, const T& value, const Ts&... values);
-  static H combine(H state) { return state; }
-
-  // HashStateBase::combine_contiguous()
-  //
-  // Combines a contiguous array of `size` elements into a hash state, returning
-  // the updated state.
-  //
-  // NOTE:
-  //
-  //   state = H::combine_contiguous(std::move(state), data, size);
-  //
-  // is NOT guaranteed to produce the same hash expansion as a for-loop (it may
-  // perform internal optimizations).  If you need this guarantee, use the
-  // for-loop instead.
-  template <typename T>
-  static H combine_contiguous(H state, const T* data, size_t size);
-
-  template <typename I>
-  static H combine_unordered(H state, I begin, I end);
-
-  using AbslInternalPiecewiseCombiner = PiecewiseCombiner;
-
-  template <typename T>
-  using is_hashable = absl::hash_internal::is_hashable<T>;
-
- private:
-  // Common implementation of the iteration step of a "combiner", as described
-  // above.
-  template <typename I>
-  struct CombineUnorderedCallback {
-    I begin;
-    I end;
-
-    template <typename InnerH, typename ElementStateConsumer>
-    void operator()(InnerH inner_state, ElementStateConsumer cb) {
-      for (; begin != end; ++begin) {
-        inner_state = H::combine(std::move(inner_state), *begin);
-        cb(inner_state);
-      }
-    }
-  };
-};
-
-// is_uniquely_represented
-//
-// `is_uniquely_represented<T>` is a trait class that indicates whether `T`
-// is uniquely represented.
-//
-// A type is "uniquely represented" if two equal values of that type are
-// guaranteed to have the same bytes in their underlying storage. In other
-// words, if `a == b`, then `memcmp(&a, &b, sizeof(T))` is guaranteed to be
-// zero. This property cannot be detected automatically, so this trait is false
-// by default, but can be specialized by types that wish to assert that they are
-// uniquely represented. This makes them eligible for certain optimizations.
-//
-// If you have any doubt whatsoever, do not specialize this template.
-// The default is completely safe, and merely disables some optimizations
-// that will not matter for most types. Specializing this template,
-// on the other hand, can be very hazardous.
-//
-// To be uniquely represented, a type must not have multiple ways of
-// representing the same value; for example, float and double are not
-// uniquely represented, because they have distinct representations for
-// +0 and -0. Furthermore, the type's byte representation must consist
-// solely of user-controlled data, with no padding bits and no compiler-
-// controlled data such as vptrs or sanitizer metadata. This is usually
-// very difficult to guarantee, because in most cases the compiler can
-// insert data and padding bits at its own discretion.
-//
-// If you specialize this template for a type `T`, you must do so in the file
-// that defines that type (or in this file). If you define that specialization
-// anywhere else, `is_uniquely_represented<T>` could have different meanings
-// in different places.
-//
-// The Enable parameter is meaningless; it is provided as a convenience,
-// to support certain SFINAE techniques when defining specializations.
-template <typename T, typename Enable = void>
-struct is_uniquely_represented : std::false_type {};
-
-// is_uniquely_represented<unsigned char>
-//
-// unsigned char is a synonym for "byte", so it is guaranteed to be
-// uniquely represented.
-template <>
-struct is_uniquely_represented<unsigned char> : std::true_type {};
-
-// is_uniquely_represented for non-standard integral types
-//
-// Integral types other than bool should be uniquely represented on any
-// platform that this will plausibly be ported to.
-template <typename Integral>
-struct is_uniquely_represented<
-    Integral, typename std::enable_if<std::is_integral<Integral>::value>::type>
-    : std::true_type {};
-
-// is_uniquely_represented<bool>
-//
-//
-template <>
-struct is_uniquely_represented<bool> : std::false_type {};
-
-// hash_bytes()
-//
-// Convenience function that combines `hash_state` with the byte representation
-// of `value`.
-template <typename H, typename T>
-H hash_bytes(H hash_state, const T& value) {
-  const unsigned char* start = reinterpret_cast<const unsigned char*>(&value);
-  return H::combine_contiguous(std::move(hash_state), start, sizeof(value));
-}
-
-// -----------------------------------------------------------------------------
-// AbslHashValue for Basic Types
-// -----------------------------------------------------------------------------
-
-// Note: Default `AbslHashValue` implementations live in `hash_internal`. This
-// allows us to block lexical scope lookup when doing an unqualified call to
-// `AbslHashValue` below. User-defined implementations of `AbslHashValue` can
-// only be found via ADL.
-
-// AbslHashValue() for hashing bool values
-//
-// We use SFINAE to ensure that this overload only accepts bool, not types that
-// are convertible to bool.
-template <typename H, typename B>
-typename std::enable_if<std::is_same<B, bool>::value, H>::type AbslHashValue(
-    H hash_state, B value) {
-  return H::combine(std::move(hash_state),
-                    static_cast<unsigned char>(value ? 1 : 0));
-}
-
-// AbslHashValue() for hashing enum values
-template <typename H, typename Enum>
-typename std::enable_if<std::is_enum<Enum>::value, H>::type AbslHashValue(
-    H hash_state, Enum e) {
-  // In practice, we could almost certainly just invoke hash_bytes directly,
-  // but it's possible that a sanitizer might one day want to
-  // store data in the unused bits of an enum. To avoid that risk, we
-  // convert to the underlying type before hashing. Hopefully this will get
-  // optimized away; if not, we can reopen discussion with c-toolchain-team.
-  return H::combine(std::move(hash_state),
-                    static_cast<typename std::underlying_type<Enum>::type>(e));
-}
-// AbslHashValue() for hashing floating-point values
-template <typename H, typename Float>
-typename std::enable_if<std::is_same<Float, float>::value ||
-                            std::is_same<Float, double>::value,
-                        H>::type
-AbslHashValue(H hash_state, Float value) {
-  return hash_internal::hash_bytes(std::move(hash_state),
-                                   value == 0 ? 0 : value);
-}
-
-// Long double has the property that it might have extra unused bytes in it.
-// For example, in x86 sizeof(long double)==16 but it only really uses 80-bits
-// of it. This means we can't use hash_bytes on a long double and have to
-// convert it to something else first.
-template <typename H, typename LongDouble>
-typename std::enable_if<std::is_same<LongDouble, long double>::value, H>::type
-AbslHashValue(H hash_state, LongDouble value) {
-  const int category = std::fpclassify(value);
-  switch (category) {
-    case FP_INFINITE:
-      // Add the sign bit to differentiate between +Inf and -Inf
-      hash_state = H::combine(std::move(hash_state), std::signbit(value));
-      break;
-
-    case FP_NAN:
-    case FP_ZERO:
-    default:
-      // Category is enough for these.
-      break;
-
-    case FP_NORMAL:
-    case FP_SUBNORMAL:
-      // We can't convert `value` directly to double because this would have
-      // undefined behavior if the value is out of range.
-      // std::frexp gives us a value in the range (-1, -.5] or [.5, 1) that is
-      // guaranteed to be in range for `double`. The truncation is
-      // implementation defined, but that works as long as it is deterministic.
-      int exp;
-      auto mantissa = static_cast<double>(std::frexp(value, &exp));
-      hash_state = H::combine(std::move(hash_state), mantissa, exp);
-  }
-
-  return H::combine(std::move(hash_state), category);
-}
-
-// Without this overload, an array decays to a pointer and we hash that, which
-// is not likely to be what the caller intended.
-template <typename H, typename T, size_t N>
-H AbslHashValue(H hash_state, T (&)[N]) {
-  static_assert(
-      sizeof(T) == -1,
-      "Hashing C arrays is not allowed. For string literals, wrap the literal "
-      "in absl::string_view(). To hash the array contents, use "
-      "absl::MakeSpan() or make the array an std::array. To hash the array "
-      "address, use &array[0].");
-  return hash_state;
-}
-
-// AbslHashValue() for hashing pointers
-template <typename H, typename T>
-std::enable_if_t<std::is_pointer<T>::value, H> AbslHashValue(H hash_state,
-                                                             T ptr) {
-  auto v = reinterpret_cast<uintptr_t>(ptr);
-  // Due to alignment, pointers tend to have low bits as zero, and the next few
-  // bits follow a pattern since they are also multiples of some base value.
-  // Mixing the pointer twice helps prevent stuck low bits for certain alignment
-  // values.
-  return H::combine(std::move(hash_state), v, v);
-}
-
-// AbslHashValue() for hashing nullptr_t
-template <typename H>
-H AbslHashValue(H hash_state, std::nullptr_t) {
-  return H::combine(std::move(hash_state), static_cast<void*>(nullptr));
-}
-
-// AbslHashValue() for hashing pointers-to-member
-template <typename H, typename T, typename C>
-H AbslHashValue(H hash_state, T C::*ptr) {
-  auto salient_ptm_size = [](std::size_t n) -> std::size_t {
-#if defined(_MSC_VER)
-    // Pointers-to-member-function on MSVC consist of one pointer plus 0, 1, 2,
-    // or 3 ints. In 64-bit mode, they are 8-byte aligned and thus can contain
-    // padding (namely when they have 1 or 3 ints). The value below is a lower
-    // bound on the number of salient, non-padding bytes that we use for
-    // hashing.
-    if (alignof(T C::*) == alignof(int)) {
-      // No padding when all subobjects have the same size as the total
-      // alignment. This happens in 32-bit mode.
-      return n;
-    } else {
-      // Padding for 1 int (size 16) or 3 ints (size 24).
-      // With 2 ints, the size is 16 with no padding, which we pessimize.
-      return n == 24 ? 20 : n == 16 ? 12 : n;
-    }
-#else
-  // On other platforms, we assume that pointers-to-members do not have
-  // padding.
-#ifdef __cpp_lib_has_unique_object_representations
-    static_assert(std::has_unique_object_representations<T C::*>::value);
-#endif  // __cpp_lib_has_unique_object_representations
-    return n;
-#endif
-  };
-  return H::combine_contiguous(std::move(hash_state),
-                               reinterpret_cast<unsigned char*>(&ptr),
-                               salient_ptm_size(sizeof ptr));
-}
-
-// -----------------------------------------------------------------------------
-// AbslHashValue for Composite Types
-// -----------------------------------------------------------------------------
-
-// AbslHashValue() for hashing pairs
-template <typename H, typename T1, typename T2>
-typename std::enable_if<is_hashable<T1>::value && is_hashable<T2>::value,
-                        H>::type
-AbslHashValue(H hash_state, const std::pair<T1, T2>& p) {
-  return H::combine(std::move(hash_state), p.first, p.second);
-}
-
-// hash_tuple()
-//
-// Helper function for hashing a tuple. The third argument should
-// be an index_sequence running from 0 to tuple_size<Tuple> - 1.
-template <typename H, typename Tuple, size_t... Is>
-H hash_tuple(H hash_state, const Tuple& t, absl::index_sequence<Is...>) {
-  return H::combine(std::move(hash_state), std::get<Is>(t)...);
-}
-
-// AbslHashValue for hashing tuples
-template <typename H, typename... Ts>
-#if defined(_MSC_VER)
-// This SFINAE gets MSVC confused under some conditions. Let's just disable it
-// for now.
-H
-#else   // _MSC_VER
-typename std::enable_if<absl::conjunction<is_hashable<Ts>...>::value, H>::type
-#endif  // _MSC_VER
-AbslHashValue(H hash_state, const std::tuple<Ts...>& t) {
-  return hash_internal::hash_tuple(std::move(hash_state), t,
-                                   absl::make_index_sequence<sizeof...(Ts)>());
-}
-
-// -----------------------------------------------------------------------------
-// AbslHashValue for Pointers
-// -----------------------------------------------------------------------------
-
-// AbslHashValue for hashing unique_ptr
-template <typename H, typename T, typename D>
-H AbslHashValue(H hash_state, const std::unique_ptr<T, D>& ptr) {
-  return H::combine(std::move(hash_state), ptr.get());
-}
-
-// AbslHashValue for hashing shared_ptr
-template <typename H, typename T>
-H AbslHashValue(H hash_state, const std::shared_ptr<T>& ptr) {
-  return H::combine(std::move(hash_state), ptr.get());
-}
-
-// -----------------------------------------------------------------------------
-// AbslHashValue for String-Like Types
-// -----------------------------------------------------------------------------
-
-// AbslHashValue for hashing strings
-//
-// All the string-like types supported here provide the same hash expansion for
-// the same character sequence. These types are:
-//
-//  - `absl::Cord`
-//  - `std::string` (and std::basic_string<T, std::char_traits<T>, A> for
-//      any allocator A and any T in {char, wchar_t, char16_t, char32_t})
-//  - `absl::string_view`, `std::string_view`, `std::wstring_view`,
-//    `std::u16string_view`, and `std::u32_string_view`.
-//
-// For simplicity, we currently support only strings built on `char`, `wchar_t`,
-// `char16_t`, or `char32_t`. This support may be broadened, if necessary, but
-// with some caution - this overload would misbehave in cases where the traits'
-// `eq()` member isn't equivalent to `==` on the underlying character type.
-template <typename H>
-H AbslHashValue(H hash_state, absl::string_view str) {
-  return H::combine(
-      H::combine_contiguous(std::move(hash_state), str.data(), str.size()),
-      str.size());
-}
-
-// Support std::wstring, std::u16string and std::u32string.
-template <typename Char, typename Alloc, typename H,
-          typename = absl::enable_if_t<std::is_same<Char, wchar_t>::value ||
-                                       std::is_same<Char, char16_t>::value ||
-                                       std::is_same<Char, char32_t>::value>>
-H AbslHashValue(
-    H hash_state,
-    const std::basic_string<Char, std::char_traits<Char>, Alloc>& str) {
-  return H::combine(
-      H::combine_contiguous(std::move(hash_state), str.data(), str.size()),
-      str.size());
-}
-
-#ifdef ABSL_HAVE_STD_STRING_VIEW
-
-// Support std::wstring_view, std::u16string_view and std::u32string_view.
-template <typename Char, typename H,
-          typename = absl::enable_if_t<std::is_same<Char, wchar_t>::value ||
-                                       std::is_same<Char, char16_t>::value ||
-                                       std::is_same<Char, char32_t>::value>>
-H AbslHashValue(H hash_state, std::basic_string_view<Char> str) {
-  return H::combine(
-      H::combine_contiguous(std::move(hash_state), str.data(), str.size()),
-      str.size());
-}
-
-#endif  // ABSL_HAVE_STD_STRING_VIEW
-
-#if defined(__cpp_lib_filesystem) && __cpp_lib_filesystem >= 201703L && \
-    !defined(_LIBCPP_HAS_NO_FILESYSTEM_LIBRARY) && \
-    (!defined(__ENVIRONMENT_IPHONE_OS_VERSION_MIN_REQUIRED__) ||        \
-     __ENVIRONMENT_IPHONE_OS_VERSION_MIN_REQUIRED__ >= 130000)
-
-#define ABSL_INTERNAL_STD_FILESYSTEM_PATH_HASH_AVAILABLE 1
-
-// Support std::filesystem::path. The SFINAE is required because some string
-// types are implicitly convertible to std::filesystem::path.
-template <typename Path, typename H,
-          typename = absl::enable_if_t<
-              std::is_same_v<Path, std::filesystem::path>>>
-H AbslHashValue(H hash_state, const Path& path) {
-  // This is implemented by deferring to the standard library to compute the
-  // hash.  The standard library requires that for two paths, `p1 == p2`, then
-  // `hash_value(p1) == hash_value(p2)`. `AbslHashValue` has the same
-  // requirement. Since `operator==` does platform specific matching, deferring
-  // to the standard library is the simplest approach.
-  return H::combine(std::move(hash_state), std::filesystem::hash_value(path));
-}
-
-#endif  // ABSL_INTERNAL_STD_FILESYSTEM_PATH_HASH_AVAILABLE
-
-// -----------------------------------------------------------------------------
-// AbslHashValue for Sequence Containers
-// -----------------------------------------------------------------------------
-
-// AbslHashValue for hashing std::array
-template <typename H, typename T, size_t N>
-typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue(
-    H hash_state, const std::array<T, N>& array) {
-  return H::combine_contiguous(std::move(hash_state), array.data(),
-                               array.size());
-}
-
-// AbslHashValue for hashing std::deque
-template <typename H, typename T, typename Allocator>
-typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue(
-    H hash_state, const std::deque<T, Allocator>& deque) {
-  // TODO(gromer): investigate a more efficient implementation taking
-  // advantage of the chunk structure.
-  for (const auto& t : deque) {
-    hash_state = H::combine(std::move(hash_state), t);
-  }
-  return H::combine(std::move(hash_state), deque.size());
-}
-
-// AbslHashValue for hashing std::forward_list
-template <typename H, typename T, typename Allocator>
-typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue(
-    H hash_state, const std::forward_list<T, Allocator>& list) {
-  size_t size = 0;
-  for (const T& t : list) {
-    hash_state = H::combine(std::move(hash_state), t);
-    ++size;
-  }
-  return H::combine(std::move(hash_state), size);
-}
-
-// AbslHashValue for hashing std::list
-template <typename H, typename T, typename Allocator>
-typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue(
-    H hash_state, const std::list<T, Allocator>& list) {
-  for (const auto& t : list) {
-    hash_state = H::combine(std::move(hash_state), t);
-  }
-  return H::combine(std::move(hash_state), list.size());
-}
-
-// AbslHashValue for hashing std::vector
-//
-// Do not use this for vector<bool> on platforms that have a working
-// implementation of std::hash. It does not have a .data(), and a fallback for
-// std::hash<> is most likely faster.
-template <typename H, typename T, typename Allocator>
-typename std::enable_if<is_hashable<T>::value && !std::is_same<T, bool>::value,
-                        H>::type
-AbslHashValue(H hash_state, const std::vector<T, Allocator>& vector) {
-  return H::combine(H::combine_contiguous(std::move(hash_state), vector.data(),
-                                          vector.size()),
-                    vector.size());
-}
-
-// AbslHashValue special cases for hashing std::vector<bool>
-
-#if defined(ABSL_IS_BIG_ENDIAN) && \
-    (defined(__GLIBCXX__) || defined(__GLIBCPP__))
-
-// std::hash in libstdc++ does not work correctly with vector<bool> on Big
-// Endian platforms therefore we need to implement a custom AbslHashValue for
-// it. More details on the bug:
-// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102531
-template <typename H, typename T, typename Allocator>
-typename std::enable_if<is_hashable<T>::value && std::is_same<T, bool>::value,
-                        H>::type
-AbslHashValue(H hash_state, const std::vector<T, Allocator>& vector) {
-  typename H::AbslInternalPiecewiseCombiner combiner;
-  for (const auto& i : vector) {
-    unsigned char c = static_cast<unsigned char>(i);
-    hash_state = combiner.add_buffer(std::move(hash_state), &c, sizeof(c));
-  }
-  return H::combine(combiner.finalize(std::move(hash_state)), vector.size());
-}
-#else
-// When not working around the libstdc++ bug above, we still have to contend
-// with the fact that std::hash<vector<bool>> is often poor quality, hashing
-// directly on the internal words and on no other state.  On these platforms,
-// vector<bool>{1, 1} and vector<bool>{1, 1, 0} hash to the same value.
-//
-// Mixing in the size (as we do in our other vector<> implementations) on top
-// of the library-provided hash implementation avoids this QOI issue.
-template <typename H, typename T, typename Allocator>
-typename std::enable_if<is_hashable<T>::value && std::is_same<T, bool>::value,
-                        H>::type
-AbslHashValue(H hash_state, const std::vector<T, Allocator>& vector) {
-  return H::combine(std::move(hash_state),
-                    std::hash<std::vector<T, Allocator>>{}(vector),
-                    vector.size());
-}
-#endif
-
-// -----------------------------------------------------------------------------
-// AbslHashValue for Ordered Associative Containers
-// -----------------------------------------------------------------------------
-
-// AbslHashValue for hashing std::map
-template <typename H, typename Key, typename T, typename Compare,
-          typename Allocator>
-typename std::enable_if<is_hashable<Key>::value && is_hashable<T>::value,
-                        H>::type
-AbslHashValue(H hash_state, const std::map<Key, T, Compare, Allocator>& map) {
-  for (const auto& t : map) {
-    hash_state = H::combine(std::move(hash_state), t);
-  }
-  return H::combine(std::move(hash_state), map.size());
-}
-
-// AbslHashValue for hashing std::multimap
-template <typename H, typename Key, typename T, typename Compare,
-          typename Allocator>
-typename std::enable_if<is_hashable<Key>::value && is_hashable<T>::value,
-                        H>::type
-AbslHashValue(H hash_state,
-              const std::multimap<Key, T, Compare, Allocator>& map) {
-  for (const auto& t : map) {
-    hash_state = H::combine(std::move(hash_state), t);
-  }
-  return H::combine(std::move(hash_state), map.size());
-}
-
-// AbslHashValue for hashing std::set
-template <typename H, typename Key, typename Compare, typename Allocator>
-typename std::enable_if<is_hashable<Key>::value, H>::type AbslHashValue(
-    H hash_state, const std::set<Key, Compare, Allocator>& set) {
-  for (const auto& t : set) {
-    hash_state = H::combine(std::move(hash_state), t);
-  }
-  return H::combine(std::move(hash_state), set.size());
-}
-
-// AbslHashValue for hashing std::multiset
-template <typename H, typename Key, typename Compare, typename Allocator>
-typename std::enable_if<is_hashable<Key>::value, H>::type AbslHashValue(
-    H hash_state, const std::multiset<Key, Compare, Allocator>& set) {
-  for (const auto& t : set) {
-    hash_state = H::combine(std::move(hash_state), t);
-  }
-  return H::combine(std::move(hash_state), set.size());
-}
-
-// -----------------------------------------------------------------------------
-// AbslHashValue for Unordered Associative Containers
-// -----------------------------------------------------------------------------
-
-// AbslHashValue for hashing std::unordered_set
-template <typename H, typename Key, typename Hash, typename KeyEqual,
-          typename Alloc>
-typename std::enable_if<is_hashable<Key>::value, H>::type AbslHashValue(
-    H hash_state, const std::unordered_set<Key, Hash, KeyEqual, Alloc>& s) {
-  return H::combine(
-      H::combine_unordered(std::move(hash_state), s.begin(), s.end()),
-      s.size());
-}
-
-// AbslHashValue for hashing std::unordered_multiset
-template <typename H, typename Key, typename Hash, typename KeyEqual,
-          typename Alloc>
-typename std::enable_if<is_hashable<Key>::value, H>::type AbslHashValue(
-    H hash_state,
-    const std::unordered_multiset<Key, Hash, KeyEqual, Alloc>& s) {
-  return H::combine(
-      H::combine_unordered(std::move(hash_state), s.begin(), s.end()),
-      s.size());
-}
-
-// AbslHashValue for hashing std::unordered_set
-template <typename H, typename Key, typename T, typename Hash,
-          typename KeyEqual, typename Alloc>
-typename std::enable_if<is_hashable<Key>::value && is_hashable<T>::value,
-                        H>::type
-AbslHashValue(H hash_state,
-              const std::unordered_map<Key, T, Hash, KeyEqual, Alloc>& s) {
-  return H::combine(
-      H::combine_unordered(std::move(hash_state), s.begin(), s.end()),
-      s.size());
-}
-
-// AbslHashValue for hashing std::unordered_multiset
-template <typename H, typename Key, typename T, typename Hash,
-          typename KeyEqual, typename Alloc>
-typename std::enable_if<is_hashable<Key>::value && is_hashable<T>::value,
-                        H>::type
-AbslHashValue(H hash_state,
-              const std::unordered_multimap<Key, T, Hash, KeyEqual, Alloc>& s) {
-  return H::combine(
-      H::combine_unordered(std::move(hash_state), s.begin(), s.end()),
-      s.size());
-}
-
-// -----------------------------------------------------------------------------
-// AbslHashValue for Wrapper Types
-// -----------------------------------------------------------------------------
-
-// AbslHashValue for hashing std::reference_wrapper
-template <typename H, typename T>
-typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue(
-    H hash_state, std::reference_wrapper<T> opt) {
-  return H::combine(std::move(hash_state), opt.get());
-}
-
-// AbslHashValue for hashing absl::optional
-template <typename H, typename T>
-typename std::enable_if<is_hashable<T>::value, H>::type AbslHashValue(
-    H hash_state, const absl::optional<T>& opt) {
-  if (opt) hash_state = H::combine(std::move(hash_state), *opt);
-  return H::combine(std::move(hash_state), opt.has_value());
-}
-
-// VariantVisitor
-template <typename H>
-struct VariantVisitor {
-  H&& hash_state;
-  template <typename T>
-  H operator()(const T& t) const {
-    return H::combine(std::move(hash_state), t);
-  }
-};
-
-// AbslHashValue for hashing absl::variant
-template <typename H, typename... T>
-typename std::enable_if<conjunction<is_hashable<T>...>::value, H>::type
-AbslHashValue(H hash_state, const absl::variant<T...>& v) {
-  if (!v.valueless_by_exception()) {
-    hash_state = absl::visit(VariantVisitor<H>{std::move(hash_state)}, v);
-  }
-  return H::combine(std::move(hash_state), v.index());
-}
-
-// -----------------------------------------------------------------------------
-// AbslHashValue for Other Types
-// -----------------------------------------------------------------------------
-
-// AbslHashValue for hashing std::bitset is not defined on Little Endian
-// platforms, for the same reason as for vector<bool> (see std::vector above):
-// It does not expose the raw bytes, and a fallback to std::hash<> is most
-// likely faster.
-
-#if defined(ABSL_IS_BIG_ENDIAN) && \
-    (defined(__GLIBCXX__) || defined(__GLIBCPP__))
-// AbslHashValue for hashing std::bitset
-//
-// std::hash in libstdc++ does not work correctly with std::bitset on Big Endian
-// platforms therefore we need to implement a custom AbslHashValue for it. More
-// details on the bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102531
-template <typename H, size_t N>
-H AbslHashValue(H hash_state, const std::bitset<N>& set) {
-  typename H::AbslInternalPiecewiseCombiner combiner;
-  for (size_t i = 0; i < N; i++) {
-    unsigned char c = static_cast<unsigned char>(set[i]);
-    hash_state = combiner.add_buffer(std::move(hash_state), &c, sizeof(c));
-  }
-  return H::combine(combiner.finalize(std::move(hash_state)), N);
-}
-#endif
-
-// -----------------------------------------------------------------------------
-
-// hash_range_or_bytes()
-//
-// Mixes all values in the range [data, data+size) into the hash state.
-// This overload accepts only uniquely-represented types, and hashes them by
-// hashing the entire range of bytes.
-template <typename H, typename T>
-typename std::enable_if<is_uniquely_represented<T>::value, H>::type
-hash_range_or_bytes(H hash_state, const T* data, size_t size) {
-  const auto* bytes = reinterpret_cast<const unsigned char*>(data);
-  return H::combine_contiguous(std::move(hash_state), bytes, sizeof(T) * size);
-}
-
-// hash_range_or_bytes()
-template <typename H, typename T>
-typename std::enable_if<!is_uniquely_represented<T>::value, H>::type
-hash_range_or_bytes(H hash_state, const T* data, size_t size) {
-  for (const auto end = data + size; data < end; ++data) {
-    hash_state = H::combine(std::move(hash_state), *data);
-  }
-  return hash_state;
-}
-
-#if defined(ABSL_INTERNAL_LEGACY_HASH_NAMESPACE) && \
-    ABSL_META_INTERNAL_STD_HASH_SFINAE_FRIENDLY_
-#define ABSL_HASH_INTERNAL_SUPPORT_LEGACY_HASH_ 1
-#else
-#define ABSL_HASH_INTERNAL_SUPPORT_LEGACY_HASH_ 0
-#endif
-
-// HashSelect
-//
-// Type trait to select the appropriate hash implementation to use.
-// HashSelect::type<T> will give the proper hash implementation, to be invoked
-// as:
-//   HashSelect::type<T>::Invoke(state, value)
-// Also, HashSelect::type<T>::value is a boolean equal to `true` if there is a
-// valid `Invoke` function. Types that are not hashable will have a ::value of
-// `false`.
-struct HashSelect {
- private:
-  struct State : HashStateBase<State> {
-    static State combine_contiguous(State hash_state, const unsigned char*,
-                                    size_t);
-    using State::HashStateBase::combine_contiguous;
-  };
-
-  struct UniquelyRepresentedProbe {
-    template <typename H, typename T>
-    static auto Invoke(H state, const T& value)
-        -> absl::enable_if_t<is_uniquely_represented<T>::value, H> {
-      return hash_internal::hash_bytes(std::move(state), value);
-    }
-  };
-
-  struct HashValueProbe {
-    template <typename H, typename T>
-    static auto Invoke(H state, const T& value) -> absl::enable_if_t<
-        std::is_same<H,
-                     decltype(AbslHashValue(std::move(state), value))>::value,
-        H> {
-      return AbslHashValue(std::move(state), value);
-    }
-  };
-
-  struct LegacyHashProbe {
-#if ABSL_HASH_INTERNAL_SUPPORT_LEGACY_HASH_
-    template <typename H, typename T>
-    static auto Invoke(H state, const T& value) -> absl::enable_if_t<
-        std::is_convertible<
-            decltype(ABSL_INTERNAL_LEGACY_HASH_NAMESPACE::hash<T>()(value)),
-            size_t>::value,
-        H> {
-      return hash_internal::hash_bytes(
-          std::move(state),
-          ABSL_INTERNAL_LEGACY_HASH_NAMESPACE::hash<T>{}(value));
-    }
-#endif  // ABSL_HASH_INTERNAL_SUPPORT_LEGACY_HASH_
-  };
-
-  struct StdHashProbe {
-    template <typename H, typename T>
-    static auto Invoke(H state, const T& value)
-        -> absl::enable_if_t<type_traits_internal::IsHashable<T>::value, H> {
-      return hash_internal::hash_bytes(std::move(state), std::hash<T>{}(value));
-    }
-  };
-
-  template <typename Hash, typename T>
-  struct Probe : Hash {
-   private:
-    template <typename H, typename = decltype(H::Invoke(
-                              std::declval<State>(), std::declval<const T&>()))>
-    static std::true_type Test(int);
-    template <typename U>
-    static std::false_type Test(char);
-
-   public:
-    static constexpr bool value = decltype(Test<Hash>(0))::value;
-  };
-
- public:
-  // Probe each implementation in order.
-  // disjunction provides short circuiting wrt instantiation.
-  template <typename T>
-  using Apply = absl::disjunction<         //
-      Probe<UniquelyRepresentedProbe, T>,  //
-      Probe<HashValueProbe, T>,            //
-      Probe<LegacyHashProbe, T>,           //
-      Probe<StdHashProbe, T>,              //
-      std::false_type>;
-};
-
-template <typename T>
-struct is_hashable
-    : std::integral_constant<bool, HashSelect::template Apply<T>::value> {};
-
-// MixingHashState
-class ABSL_DLL MixingHashState : public HashStateBase<MixingHashState> {
-  // absl::uint128 is not an alias or a thin wrapper around the intrinsic.
-  // We use the intrinsic when available to improve performance.
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-  using uint128 = __uint128_t;
-#else   // ABSL_HAVE_INTRINSIC_INT128
-  using uint128 = absl::uint128;
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-
-  static constexpr uint64_t kMul =
-  sizeof(size_t) == 4 ? uint64_t{0xcc9e2d51}
-                      : uint64_t{0x9ddfea08eb382d69};
-
-  template <typename T>
-  using IntegralFastPath =
-      conjunction<std::is_integral<T>, is_uniquely_represented<T>>;
-
- public:
-  // Move only
-  MixingHashState(MixingHashState&&) = default;
-  MixingHashState& operator=(MixingHashState&&) = default;
-
-  // MixingHashState::combine_contiguous()
-  //
-  // Fundamental base case for hash recursion: mixes the given range of bytes
-  // into the hash state.
-  static MixingHashState combine_contiguous(MixingHashState hash_state,
-                                            const unsigned char* first,
-                                            size_t size) {
-    return MixingHashState(
-        CombineContiguousImpl(hash_state.state_, first, size,
-                              std::integral_constant<int, sizeof(size_t)>{}));
-  }
-  using MixingHashState::HashStateBase::combine_contiguous;
-
-  // MixingHashState::hash()
-  //
-  // For performance reasons in non-opt mode, we specialize this for
-  // integral types.
-  // Otherwise we would be instantiating and calling dozens of functions for
-  // something that is just one multiplication and a couple xor's.
-  // The result should be the same as running the whole algorithm, but faster.
-  template <typename T, absl::enable_if_t<IntegralFastPath<T>::value, int> = 0>
-  static size_t hash(T value) {
-    return static_cast<size_t>(
-        Mix(Seed(), static_cast<std::make_unsigned_t<T>>(value)));
-  }
-
-  // Overload of MixingHashState::hash()
-  template <typename T, absl::enable_if_t<!IntegralFastPath<T>::value, int> = 0>
-  static size_t hash(const T& value) {
-    return static_cast<size_t>(combine(MixingHashState{}, value).state_);
-  }
-
- private:
-  // Invoked only once for a given argument; that plus the fact that this is
-  // move-only ensures that there is only one non-moved-from object.
-  MixingHashState() : state_(Seed()) {}
-
-  friend class MixingHashState::HashStateBase;
-
-  template <typename CombinerT>
-  static MixingHashState RunCombineUnordered(MixingHashState state,
-                                             CombinerT combiner) {
-    uint64_t unordered_state = 0;
-    combiner(MixingHashState{}, [&](MixingHashState& inner_state) {
-      // Add the hash state of the element to the running total, but mix the
-      // carry bit back into the low bit.  This in intended to avoid losing
-      // entropy to overflow, especially when unordered_multisets contain
-      // multiple copies of the same value.
-      auto element_state = inner_state.state_;
-      unordered_state += element_state;
-      if (unordered_state < element_state) {
-        ++unordered_state;
-      }
-      inner_state = MixingHashState{};
-    });
-    return MixingHashState::combine(std::move(state), unordered_state);
-  }
-
-  // Allow the HashState type-erasure implementation to invoke
-  // RunCombinedUnordered() directly.
-  friend class absl::HashState;
-
-  // Workaround for MSVC bug.
-  // We make the type copyable to fix the calling convention, even though we
-  // never actually copy it. Keep it private to not affect the public API of the
-  // type.
-  MixingHashState(const MixingHashState&) = default;
-
-  explicit MixingHashState(uint64_t state) : state_(state) {}
-
-  // Implementation of the base case for combine_contiguous where we actually
-  // mix the bytes into the state.
-  // Dispatch to different implementations of the combine_contiguous depending
-  // on the value of `sizeof(size_t)`.
-  static uint64_t CombineContiguousImpl(uint64_t state,
-                                        const unsigned char* first, size_t len,
-                                        std::integral_constant<int, 4>
-                                        /* sizeof_size_t */);
-  static uint64_t CombineContiguousImpl(uint64_t state,
-                                        const unsigned char* first, size_t len,
-                                        std::integral_constant<int, 8>
-                                        /* sizeof_size_t */);
-
-  // Slow dispatch path for calls to CombineContiguousImpl with a size argument
-  // larger than PiecewiseChunkSize().  Has the same effect as calling
-  // CombineContiguousImpl() repeatedly with the chunk stride size.
-  static uint64_t CombineLargeContiguousImpl32(uint64_t state,
-                                               const unsigned char* first,
-                                               size_t len);
-  static uint64_t CombineLargeContiguousImpl64(uint64_t state,
-                                               const unsigned char* first,
-                                               size_t len);
-
-  // Reads 9 to 16 bytes from p.
-  // The least significant 8 bytes are in .first, the rest (zero padded) bytes
-  // are in .second.
-  static std::pair<uint64_t, uint64_t> Read9To16(const unsigned char* p,
-                                                 size_t len) {
-    uint64_t low_mem = absl::base_internal::UnalignedLoad64(p);
-    uint64_t high_mem = absl::base_internal::UnalignedLoad64(p + len - 8);
-#ifdef ABSL_IS_LITTLE_ENDIAN
-    uint64_t most_significant = high_mem;
-    uint64_t least_significant = low_mem;
-#else
-    uint64_t most_significant = low_mem;
-    uint64_t least_significant = high_mem;
-#endif
-    return {least_significant, most_significant};
-  }
-
-  // Reads 4 to 8 bytes from p. Zero pads to fill uint64_t.
-  static uint64_t Read4To8(const unsigned char* p, size_t len) {
-    uint32_t low_mem = absl::base_internal::UnalignedLoad32(p);
-    uint32_t high_mem = absl::base_internal::UnalignedLoad32(p + len - 4);
-#ifdef ABSL_IS_LITTLE_ENDIAN
-    uint32_t most_significant = high_mem;
-    uint32_t least_significant = low_mem;
-#else
-    uint32_t most_significant = low_mem;
-    uint32_t least_significant = high_mem;
-#endif
-    return (static_cast<uint64_t>(most_significant) << (len - 4) * 8) |
-           least_significant;
-  }
-
-  // Reads 1 to 3 bytes from p. Zero pads to fill uint32_t.
-  static uint32_t Read1To3(const unsigned char* p, size_t len) {
-    // The trick used by this implementation is to avoid branches if possible.
-    unsigned char mem0 = p[0];
-    unsigned char mem1 = p[len / 2];
-    unsigned char mem2 = p[len - 1];
-#ifdef ABSL_IS_LITTLE_ENDIAN
-    unsigned char significant2 = mem2;
-    unsigned char significant1 = mem1;
-    unsigned char significant0 = mem0;
-#else
-    unsigned char significant2 = mem0;
-    unsigned char significant1 = len == 2 ? mem0 : mem1;
-    unsigned char significant0 = mem2;
-#endif
-    return static_cast<uint32_t>(significant0 |                     //
-                                 (significant1 << (len / 2 * 8)) |  //
-                                 (significant2 << ((len - 1) * 8)));
-  }
-
-  ABSL_ATTRIBUTE_ALWAYS_INLINE static uint64_t Mix(uint64_t state, uint64_t v) {
-    // Though the 128-bit product on AArch64 needs two instructions, it is
-    // still a good balance between speed and hash quality.
-    using MultType =
-        absl::conditional_t<sizeof(size_t) == 4, uint64_t, uint128>;
-    // We do the addition in 64-bit space to make sure the 128-bit
-    // multiplication is fast. If we were to do it as MultType the compiler has
-    // to assume that the high word is non-zero and needs to perform 2
-    // multiplications instead of one.
-    MultType m = state + v;
-    m *= kMul;
-    return static_cast<uint64_t>(m ^ (m >> (sizeof(m) * 8 / 2)));
-  }
-
-  // An extern to avoid bloat on a direct call to LowLevelHash() with fixed
-  // values for both the seed and salt parameters.
-  static uint64_t LowLevelHashImpl(const unsigned char* data, size_t len);
-
-  ABSL_ATTRIBUTE_ALWAYS_INLINE static uint64_t Hash64(const unsigned char* data,
-                                                      size_t len) {
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-    return LowLevelHashImpl(data, len);
-#else
-    return hash_internal::CityHash64(reinterpret_cast<const char*>(data), len);
-#endif
-  }
-
-  // Seed()
-  //
-  // A non-deterministic seed.
-  //
-  // The current purpose of this seed is to generate non-deterministic results
-  // and prevent having users depend on the particular hash values.
-  // It is not meant as a security feature right now, but it leaves the door
-  // open to upgrade it to a true per-process random seed. A true random seed
-  // costs more and we don't need to pay for that right now.
-  //
-  // On platforms with ASLR, we take advantage of it to make a per-process
-  // random value.
-  // See https://en.wikipedia.org/wiki/Address_space_layout_randomization
-  //
-  // On other platforms this is still going to be non-deterministic but most
-  // probably per-build and not per-process.
-  ABSL_ATTRIBUTE_ALWAYS_INLINE static uint64_t Seed() {
-#if (!defined(__clang__) || __clang_major__ > 11) && \
-    (!defined(__apple_build_version__) ||            \
-     __apple_build_version__ >= 19558921)  // Xcode 12
-    return static_cast<uint64_t>(reinterpret_cast<uintptr_t>(&kSeed));
-#else
-    // Workaround the absence of
-    // https://github.com/llvm/llvm-project/commit/bc15bf66dcca76cc06fe71fca35b74dc4d521021.
-    return static_cast<uint64_t>(reinterpret_cast<uintptr_t>(kSeed));
-#endif
-  }
-  static const void* const kSeed;
-
-  uint64_t state_;
-};
-
-// MixingHashState::CombineContiguousImpl()
-inline uint64_t MixingHashState::CombineContiguousImpl(
-    uint64_t state, const unsigned char* first, size_t len,
-    std::integral_constant<int, 4> /* sizeof_size_t */) {
-  // For large values we use CityHash, for small ones we just use a
-  // multiplicative hash.
-  uint64_t v;
-  if (len > 8) {
-    if (ABSL_PREDICT_FALSE(len > PiecewiseChunkSize())) {
-      return CombineLargeContiguousImpl32(state, first, len);
-    }
-    v = hash_internal::CityHash32(reinterpret_cast<const char*>(first), len);
-  } else if (len >= 4) {
-    v = Read4To8(first, len);
-  } else if (len > 0) {
-    v = Read1To3(first, len);
-  } else {
-    // Empty ranges have no effect.
-    return state;
-  }
-  return Mix(state, v);
-}
-
-// Overload of MixingHashState::CombineContiguousImpl()
-inline uint64_t MixingHashState::CombineContiguousImpl(
-    uint64_t state, const unsigned char* first, size_t len,
-    std::integral_constant<int, 8> /* sizeof_size_t */) {
-  // For large values we use LowLevelHash or CityHash depending on the platform,
-  // for small ones we just use a multiplicative hash.
-  uint64_t v;
-  if (len > 16) {
-    if (ABSL_PREDICT_FALSE(len > PiecewiseChunkSize())) {
-      return CombineLargeContiguousImpl64(state, first, len);
-    }
-    v = Hash64(first, len);
-  } else if (len > 8) {
-    // This hash function was constructed by the ML-driven algorithm discovery
-    // using reinforcement learning. We fed the agent lots of inputs from
-    // microbenchmarks, SMHasher, low hamming distance from generated inputs and
-    // picked up the one that was good on micro and macrobenchmarks.
-    auto p = Read9To16(first, len);
-    uint64_t lo = p.first;
-    uint64_t hi = p.second;
-    // Rotation by 53 was found to be most often useful when discovering these
-    // hashing algorithms with ML techniques.
-    lo = absl::rotr(lo, 53);
-    state += kMul;
-    lo += state;
-    state ^= hi;
-    uint128 m = state;
-    m *= lo;
-    return static_cast<uint64_t>(m ^ (m >> 64));
-  } else if (len >= 4) {
-    v = Read4To8(first, len);
-  } else if (len > 0) {
-    v = Read1To3(first, len);
-  } else {
-    // Empty ranges have no effect.
-    return state;
-  }
-  return Mix(state, v);
-}
-
-struct AggregateBarrier {};
-
-// HashImpl
-
-// Add a private base class to make sure this type is not an aggregate.
-// Aggregates can be aggregate initialized even if the default constructor is
-// deleted.
-struct PoisonedHash : private AggregateBarrier {
-  PoisonedHash() = delete;
-  PoisonedHash(const PoisonedHash&) = delete;
-  PoisonedHash& operator=(const PoisonedHash&) = delete;
-};
-
-template <typename T>
-struct HashImpl {
-  size_t operator()(const T& value) const {
-    return MixingHashState::hash(value);
-  }
-};
-
-template <typename T>
-struct Hash
-    : absl::conditional_t<is_hashable<T>::value, HashImpl<T>, PoisonedHash> {};
-
-template <typename H>
-template <typename T, typename... Ts>
-H HashStateBase<H>::combine(H state, const T& value, const Ts&... values) {
-  return H::combine(hash_internal::HashSelect::template Apply<T>::Invoke(
-                        std::move(state), value),
-                    values...);
-}
-
-// HashStateBase::combine_contiguous()
-template <typename H>
-template <typename T>
-H HashStateBase<H>::combine_contiguous(H state, const T* data, size_t size) {
-  return hash_internal::hash_range_or_bytes(std::move(state), data, size);
-}
-
-// HashStateBase::combine_unordered()
-template <typename H>
-template <typename I>
-H HashStateBase<H>::combine_unordered(H state, I begin, I end) {
-  return H::RunCombineUnordered(std::move(state),
-                                CombineUnorderedCallback<I>{begin, end});
-}
-
-// HashStateBase::PiecewiseCombiner::add_buffer()
-template <typename H>
-H PiecewiseCombiner::add_buffer(H state, const unsigned char* data,
-                                size_t size) {
-  if (position_ + size < PiecewiseChunkSize()) {
-    // This partial chunk does not fill our existing buffer
-    memcpy(buf_ + position_, data, size);
-    position_ += size;
-    return state;
-  }
-
-  // If the buffer is partially filled we need to complete the buffer
-  // and hash it.
-  if (position_ != 0) {
-    const size_t bytes_needed = PiecewiseChunkSize() - position_;
-    memcpy(buf_ + position_, data, bytes_needed);
-    state = H::combine_contiguous(std::move(state), buf_, PiecewiseChunkSize());
-    data += bytes_needed;
-    size -= bytes_needed;
-  }
-
-  // Hash whatever chunks we can without copying
-  while (size >= PiecewiseChunkSize()) {
-    state = H::combine_contiguous(std::move(state), data, PiecewiseChunkSize());
-    data += PiecewiseChunkSize();
-    size -= PiecewiseChunkSize();
-  }
-  // Fill the buffer with the remainder
-  memcpy(buf_, data, size);
-  position_ = size;
-  return state;
-}
-
-// HashStateBase::PiecewiseCombiner::finalize()
-template <typename H>
-H PiecewiseCombiner::finalize(H state) {
-  // Hash the remainder left in the buffer, which may be empty
-  return H::combine_contiguous(std::move(state), buf_, position_);
-}
-
-}  // namespace hash_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_HASH_INTERNAL_HASH_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/hash/internal/hash_test.h b/projects/llm_framework/include/abseil-cpp/absl/hash/internal/hash_test.h
deleted file mode 100644
index 9963dc0b..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/hash/internal/hash_test.h
+++ /dev/null
@@ -1,87 +0,0 @@
-// Copyright 2023 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-// Common code shared between absl/hash/hash_test.cc and
-// absl/hash/hash_instantiated_test.cc.
-
-#ifndef ABSL_HASH_INTERNAL_HASH_TEST_H_
-#define ABSL_HASH_INTERNAL_HASH_TEST_H_
-
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/config.h"
-#include "absl/hash/hash.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace hash_test_internal {
-
-// Utility wrapper of T for the purposes of testing the `AbslHash` type erasure
-// mechanism.  `TypeErasedValue<T>` can be constructed with a `T`, and can
-// be compared and hashed.  However, all hashing goes through the hashing
-// type-erasure framework.
-template <typename T>
-class TypeErasedValue {
- public:
-  TypeErasedValue() = default;
-  TypeErasedValue(const TypeErasedValue&) = default;
-  TypeErasedValue(TypeErasedValue&&) = default;
-  explicit TypeErasedValue(const T& n) : n_(n) {}
-
-  template <typename H>
-  friend H AbslHashValue(H hash_state, const TypeErasedValue& v) {
-    v.HashValue(absl::HashState::Create(&hash_state));
-    return hash_state;
-  }
-
-  void HashValue(absl::HashState state) const {
-    absl::HashState::combine(std::move(state), n_);
-  }
-
-  bool operator==(const TypeErasedValue& rhs) const { return n_ == rhs.n_; }
-  bool operator!=(const TypeErasedValue& rhs) const { return !(*this == rhs); }
-
- private:
-  T n_;
-};
-
-// A TypeErasedValue refinement, for containers.  It exposes the wrapped
-// `value_type` and is constructible from an initializer list.
-template <typename T>
-class TypeErasedContainer : public TypeErasedValue<T> {
- public:
-  using value_type = typename T::value_type;
-  TypeErasedContainer() = default;
-  TypeErasedContainer(const TypeErasedContainer&) = default;
-  TypeErasedContainer(TypeErasedContainer&&) = default;
-  explicit TypeErasedContainer(const T& n) : TypeErasedValue<T>(n) {}
-  TypeErasedContainer(std::initializer_list<value_type> init_list)
-      : TypeErasedContainer(T(init_list.begin(), init_list.end())) {}
-  // one-argument constructor of value type T, to appease older toolchains that
-  // get confused by one-element initializer lists in some contexts
-  explicit TypeErasedContainer(const value_type& v)
-      : TypeErasedContainer(T(&v, &v + 1)) {}
-};
-
-// Helper trait to verify if T is hashable. We use absl::Hash's poison status to
-// detect it.
-template <typename T>
-using is_hashable = std::is_default_constructible<absl::Hash<T>>;
-
-}  // namespace hash_test_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_HASH_INTERNAL_HASH_TEST_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/hash/internal/low_level_hash.h b/projects/llm_framework/include/abseil-cpp/absl/hash/internal/low_level_hash.h
deleted file mode 100644
index d460e351..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/hash/internal/low_level_hash.h
+++ /dev/null
@@ -1,54 +0,0 @@
-// Copyright 2020 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// This file provides the Google-internal implementation of LowLevelHash.
-//
-// LowLevelHash is a fast hash function for hash tables, the fastest we've
-// currently (late 2020) found that passes the SMHasher tests. The algorithm
-// relies on intrinsic 128-bit multiplication for speed. This is not meant to be
-// secure - just fast.
-//
-// It is closely based on a version of wyhash, but does not maintain or
-// guarantee future compatibility with it.
-
-#ifndef ABSL_HASH_INTERNAL_LOW_LEVEL_HASH_H_
-#define ABSL_HASH_INTERNAL_LOW_LEVEL_HASH_H_
-
-#include <stdint.h>
-#include <stdlib.h>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace hash_internal {
-
-// Hash function for a byte array. A 64-bit seed and a set of five 64-bit
-// integers are hashed into the result.
-//
-// To allow all hashable types (including string_view and Span) to depend on
-// this algorithm, we keep the API low-level, with as few dependencies as
-// possible.
-uint64_t LowLevelHash(const void* data, size_t len, uint64_t seed,
-                      const uint64_t salt[5]);
-
-// Same as above except the length must be greater than 16.
-uint64_t LowLevelHashLenGt16(const void* data, size_t len, uint64_t seed,
-                             const uint64_t salt[5]);
-
-}  // namespace hash_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_HASH_INTERNAL_LOW_LEVEL_HASH_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/hash/internal/spy_hash_state.h b/projects/llm_framework/include/abseil-cpp/absl/hash/internal/spy_hash_state.h
deleted file mode 100644
index 357c301c..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/hash/internal/spy_hash_state.h
+++ /dev/null
@@ -1,266 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_HASH_INTERNAL_SPY_HASH_STATE_H_
-#define ABSL_HASH_INTERNAL_SPY_HASH_STATE_H_
-
-#include <algorithm>
-#include <ostream>
-#include <string>
-#include <vector>
-
-#include "absl/hash/hash.h"
-#include "absl/strings/match.h"
-#include "absl/strings/str_format.h"
-#include "absl/strings/str_join.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace hash_internal {
-
-// SpyHashState is an implementation of the HashState API that simply
-// accumulates all input bytes in an internal buffer. This makes it useful
-// for testing AbslHashValue overloads (so long as they are templated on the
-// HashState parameter), since it can report the exact hash representation
-// that the AbslHashValue overload produces.
-//
-// Sample usage:
-// EXPECT_EQ(SpyHashState::combine(SpyHashState(), foo),
-//           SpyHashState::combine(SpyHashState(), bar));
-template <typename T>
-class SpyHashStateImpl : public HashStateBase<SpyHashStateImpl<T>> {
- public:
-  SpyHashStateImpl() : error_(std::make_shared<absl::optional<std::string>>()) {
-    static_assert(std::is_void<T>::value, "");
-  }
-
-  // Move-only
-  SpyHashStateImpl(const SpyHashStateImpl&) = delete;
-  SpyHashStateImpl& operator=(const SpyHashStateImpl&) = delete;
-
-  SpyHashStateImpl(SpyHashStateImpl&& other) noexcept {
-    *this = std::move(other);
-  }
-
-  SpyHashStateImpl& operator=(SpyHashStateImpl&& other) noexcept {
-    hash_representation_ = std::move(other.hash_representation_);
-    error_ = other.error_;
-    moved_from_ = other.moved_from_;
-    other.moved_from_ = true;
-    return *this;
-  }
-
-  template <typename U>
-  SpyHashStateImpl(SpyHashStateImpl<U>&& other) {  // NOLINT
-    hash_representation_ = std::move(other.hash_representation_);
-    error_ = other.error_;
-    moved_from_ = other.moved_from_;
-    other.moved_from_ = true;
-  }
-
-  template <typename A, typename... Args>
-  static SpyHashStateImpl combine(SpyHashStateImpl s, const A& a,
-                                  const Args&... args) {
-    // Pass an instance of SpyHashStateImpl<A> when trying to combine `A`. This
-    // allows us to test that the user only uses this instance for combine calls
-    // and does not call AbslHashValue directly.
-    // See AbslHashValue implementation at the bottom.
-    s = SpyHashStateImpl<A>::HashStateBase::combine(std::move(s), a);
-    return SpyHashStateImpl::combine(std::move(s), args...);
-  }
-  static SpyHashStateImpl combine(SpyHashStateImpl s) {
-    if (direct_absl_hash_value_error_) {
-      *s.error_ = "AbslHashValue should not be invoked directly.";
-    } else if (s.moved_from_) {
-      *s.error_ = "Used moved-from instance of the hash state object.";
-    }
-    return s;
-  }
-
-  static void SetDirectAbslHashValueError() {
-    direct_absl_hash_value_error_ = true;
-  }
-
-  // Two SpyHashStateImpl objects are equal if they hold equal hash
-  // representations.
-  friend bool operator==(const SpyHashStateImpl& lhs,
-                         const SpyHashStateImpl& rhs) {
-    return lhs.hash_representation_ == rhs.hash_representation_;
-  }
-
-  friend bool operator!=(const SpyHashStateImpl& lhs,
-                         const SpyHashStateImpl& rhs) {
-    return !(lhs == rhs);
-  }
-
-  enum class CompareResult {
-    kEqual,
-    kASuffixB,
-    kBSuffixA,
-    kUnequal,
-  };
-
-  static CompareResult Compare(const SpyHashStateImpl& a,
-                               const SpyHashStateImpl& b) {
-    const std::string a_flat = absl::StrJoin(a.hash_representation_, "");
-    const std::string b_flat = absl::StrJoin(b.hash_representation_, "");
-    if (a_flat == b_flat) return CompareResult::kEqual;
-    if (absl::EndsWith(a_flat, b_flat)) return CompareResult::kBSuffixA;
-    if (absl::EndsWith(b_flat, a_flat)) return CompareResult::kASuffixB;
-    return CompareResult::kUnequal;
-  }
-
-  // operator<< prints the hash representation as a hex and ASCII dump, to
-  // facilitate debugging.
-  friend std::ostream& operator<<(std::ostream& out,
-                                  const SpyHashStateImpl& hash_state) {
-    out << "[\n";
-    for (auto& s : hash_state.hash_representation_) {
-      size_t offset = 0;
-      for (char c : s) {
-        if (offset % 16 == 0) {
-          out << absl::StreamFormat("\n0x%04x: ", offset);
-        }
-        if (offset % 2 == 0) {
-          out << " ";
-        }
-        out << absl::StreamFormat("%02x", c);
-        ++offset;
-      }
-      out << "\n";
-    }
-    return out << "]";
-  }
-
-  // The base case of the combine recursion, which writes raw bytes into the
-  // internal buffer.
-  static SpyHashStateImpl combine_contiguous(SpyHashStateImpl hash_state,
-                                             const unsigned char* begin,
-                                             size_t size) {
-    const size_t large_chunk_stride = PiecewiseChunkSize();
-    // Combining a large contiguous buffer must have the same effect as
-    // doing it piecewise by the stride length, followed by the (possibly
-    // empty) remainder.
-    while (size > large_chunk_stride) {
-      hash_state = SpyHashStateImpl::combine_contiguous(
-          std::move(hash_state), begin, large_chunk_stride);
-      begin += large_chunk_stride;
-      size -= large_chunk_stride;
-    }
-
-    if (size > 0) {
-      hash_state.hash_representation_.emplace_back(
-          reinterpret_cast<const char*>(begin), size);
-    }
-    return hash_state;
-  }
-
-  using SpyHashStateImpl::HashStateBase::combine_contiguous;
-
-  template <typename CombinerT>
-  static SpyHashStateImpl RunCombineUnordered(SpyHashStateImpl state,
-                                              CombinerT combiner) {
-    UnorderedCombinerCallback cb;
-
-    combiner(SpyHashStateImpl<void>{}, std::ref(cb));
-
-    std::sort(cb.element_hash_representations.begin(),
-              cb.element_hash_representations.end());
-    state.hash_representation_.insert(state.hash_representation_.end(),
-                                      cb.element_hash_representations.begin(),
-                                      cb.element_hash_representations.end());
-    if (cb.error && cb.error->has_value()) {
-      state.error_ = std::move(cb.error);
-    }
-    return state;
-  }
-
-  absl::optional<std::string> error() const {
-    if (moved_from_) {
-      return "Returned a moved-from instance of the hash state object.";
-    }
-    return *error_;
-  }
-
- private:
-  template <typename U>
-  friend class SpyHashStateImpl;
-
-  struct UnorderedCombinerCallback {
-    std::vector<std::string> element_hash_representations;
-    std::shared_ptr<absl::optional<std::string>> error;
-
-    // The inner spy can have a different type.
-    template <typename U>
-    void operator()(SpyHashStateImpl<U>& inner) {
-      element_hash_representations.push_back(
-          absl::StrJoin(inner.hash_representation_, ""));
-      if (inner.error_->has_value()) {
-        error = std::move(inner.error_);
-      }
-      inner = SpyHashStateImpl<void>{};
-    }
-  };
-
-  // This is true if SpyHashStateImpl<T> has been passed to a call of
-  // AbslHashValue with the wrong type. This detects that the user called
-  // AbslHashValue directly (because the hash state type does not match).
-  static bool direct_absl_hash_value_error_;
-
-  std::vector<std::string> hash_representation_;
-  // This is a shared_ptr because we want all instances of the particular
-  // SpyHashState run to share the field. This way we can set the error for
-  // use-after-move and all the copies will see it.
-  std::shared_ptr<absl::optional<std::string>> error_;
-  bool moved_from_ = false;
-};
-
-template <typename T>
-bool SpyHashStateImpl<T>::direct_absl_hash_value_error_;
-
-template <bool& B>
-struct OdrUse {
-  constexpr OdrUse() {}
-  bool& b = B;
-};
-
-template <void (*)()>
-struct RunOnStartup {
-  static bool run;
-  static constexpr OdrUse<run> kOdrUse{};
-};
-
-template <void (*f)()>
-bool RunOnStartup<f>::run = (f(), true);
-
-template <
-    typename T, typename U,
-    // Only trigger for when (T != U),
-    typename = absl::enable_if_t<!std::is_same<T, U>::value>,
-    // This statement works in two ways:
-    //  - First, it instantiates RunOnStartup and forces the initialization of
-    //    `run`, which set the global variable.
-    //  - Second, it triggers a SFINAE error disabling the overload to prevent
-    //    compile time errors. If we didn't disable the overload we would get
-    //    ambiguous overload errors, which we don't want.
-    int = RunOnStartup<SpyHashStateImpl<T>::SetDirectAbslHashValueError>::run>
-void AbslHashValue(SpyHashStateImpl<T>, const U&);
-
-using SpyHashState = SpyHashStateImpl<void>;
-
-}  // namespace hash_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_HASH_INTERNAL_SPY_HASH_STATE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/absl_check.h b/projects/llm_framework/include/abseil-cpp/absl/log/absl_check.h
deleted file mode 100644
index 1bb43bd3..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/absl_check.h
+++ /dev/null
@@ -1,117 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/absl_check.h
-// -----------------------------------------------------------------------------
-//
-// This header declares a family of `ABSL_CHECK` macros as alternative spellings
-// for `CHECK` macros in `check.h`.
-//
-// Except for those whose names begin with `ABSL_DCHECK`, these macros are not
-// controlled by `NDEBUG` (cf. `assert`), so the check will be executed
-// regardless of compilation mode. `ABSL_CHECK` and friends are thus useful for
-// confirming invariants in situations where continuing to run would be worse
-// than terminating, e.g., due to risk of data corruption or security
-// compromise.  It is also more robust and portable to deliberately terminate
-// at a particular place with a useful message and backtrace than to assume some
-// ultimately unspecified and unreliable crashing behavior (such as a
-// "segmentation fault").
-//
-// For full documentation of each macro, see comments in `check.h`, which has an
-// identical set of macros without the ABSL_* prefix.
-
-#ifndef ABSL_LOG_ABSL_CHECK_H_
-#define ABSL_LOG_ABSL_CHECK_H_
-
-#include "absl/log/internal/check_impl.h"
-
-#define ABSL_CHECK(condition) \
-  ABSL_LOG_INTERNAL_CHECK_IMPL((condition), #condition)
-#define ABSL_QCHECK(condition) \
-  ABSL_LOG_INTERNAL_QCHECK_IMPL((condition), #condition)
-#define ABSL_PCHECK(condition) \
-  ABSL_LOG_INTERNAL_PCHECK_IMPL((condition), #condition)
-#define ABSL_DCHECK(condition) \
-  ABSL_LOG_INTERNAL_DCHECK_IMPL((condition), #condition)
-
-#define ABSL_CHECK_EQ(val1, val2) \
-  ABSL_LOG_INTERNAL_CHECK_EQ_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_CHECK_NE(val1, val2) \
-  ABSL_LOG_INTERNAL_CHECK_NE_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_CHECK_LE(val1, val2) \
-  ABSL_LOG_INTERNAL_CHECK_LE_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_CHECK_LT(val1, val2) \
-  ABSL_LOG_INTERNAL_CHECK_LT_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_CHECK_GE(val1, val2) \
-  ABSL_LOG_INTERNAL_CHECK_GE_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_CHECK_GT(val1, val2) \
-  ABSL_LOG_INTERNAL_CHECK_GT_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_QCHECK_EQ(val1, val2) \
-  ABSL_LOG_INTERNAL_QCHECK_EQ_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_QCHECK_NE(val1, val2) \
-  ABSL_LOG_INTERNAL_QCHECK_NE_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_QCHECK_LE(val1, val2) \
-  ABSL_LOG_INTERNAL_QCHECK_LE_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_QCHECK_LT(val1, val2) \
-  ABSL_LOG_INTERNAL_QCHECK_LT_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_QCHECK_GE(val1, val2) \
-  ABSL_LOG_INTERNAL_QCHECK_GE_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_QCHECK_GT(val1, val2) \
-  ABSL_LOG_INTERNAL_QCHECK_GT_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_DCHECK_EQ(val1, val2) \
-  ABSL_LOG_INTERNAL_DCHECK_EQ_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_DCHECK_NE(val1, val2) \
-  ABSL_LOG_INTERNAL_DCHECK_NE_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_DCHECK_LE(val1, val2) \
-  ABSL_LOG_INTERNAL_DCHECK_LE_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_DCHECK_LT(val1, val2) \
-  ABSL_LOG_INTERNAL_DCHECK_LT_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_DCHECK_GE(val1, val2) \
-  ABSL_LOG_INTERNAL_DCHECK_GE_IMPL((val1), #val1, (val2), #val2)
-#define ABSL_DCHECK_GT(val1, val2) \
-  ABSL_LOG_INTERNAL_DCHECK_GT_IMPL((val1), #val1, (val2), #val2)
-
-#define ABSL_CHECK_OK(status) ABSL_LOG_INTERNAL_CHECK_OK_IMPL((status), #status)
-#define ABSL_QCHECK_OK(status) \
-  ABSL_LOG_INTERNAL_QCHECK_OK_IMPL((status), #status)
-#define ABSL_DCHECK_OK(status) \
-  ABSL_LOG_INTERNAL_DCHECK_OK_IMPL((status), #status)
-
-#define ABSL_CHECK_STREQ(s1, s2) \
-  ABSL_LOG_INTERNAL_CHECK_STREQ_IMPL((s1), #s1, (s2), #s2)
-#define ABSL_CHECK_STRNE(s1, s2) \
-  ABSL_LOG_INTERNAL_CHECK_STRNE_IMPL((s1), #s1, (s2), #s2)
-#define ABSL_CHECK_STRCASEEQ(s1, s2) \
-  ABSL_LOG_INTERNAL_CHECK_STRCASEEQ_IMPL((s1), #s1, (s2), #s2)
-#define ABSL_CHECK_STRCASENE(s1, s2) \
-  ABSL_LOG_INTERNAL_CHECK_STRCASENE_IMPL((s1), #s1, (s2), #s2)
-#define ABSL_QCHECK_STREQ(s1, s2) \
-  ABSL_LOG_INTERNAL_QCHECK_STREQ_IMPL((s1), #s1, (s2), #s2)
-#define ABSL_QCHECK_STRNE(s1, s2) \
-  ABSL_LOG_INTERNAL_QCHECK_STRNE_IMPL((s1), #s1, (s2), #s2)
-#define ABSL_QCHECK_STRCASEEQ(s1, s2) \
-  ABSL_LOG_INTERNAL_QCHECK_STRCASEEQ_IMPL((s1), #s1, (s2), #s2)
-#define ABSL_QCHECK_STRCASENE(s1, s2) \
-  ABSL_LOG_INTERNAL_QCHECK_STRCASENE_IMPL((s1), #s1, (s2), #s2)
-#define ABSL_DCHECK_STREQ(s1, s2) \
-  ABSL_LOG_INTERNAL_DCHECK_STREQ_IMPL((s1), #s1, (s2), #s2)
-#define ABSL_DCHECK_STRNE(s1, s2) \
-  ABSL_LOG_INTERNAL_DCHECK_STRNE_IMPL((s1), #s1, (s2), #s2)
-#define ABSL_DCHECK_STRCASEEQ(s1, s2) \
-  ABSL_LOG_INTERNAL_DCHECK_STRCASEEQ_IMPL((s1), #s1, (s2), #s2)
-#define ABSL_DCHECK_STRCASENE(s1, s2) \
-  ABSL_LOG_INTERNAL_DCHECK_STRCASENE_IMPL((s1), #s1, (s2), #s2)
-
-#endif  // ABSL_LOG_ABSL_CHECK_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/absl_log.h b/projects/llm_framework/include/abseil-cpp/absl/log/absl_log.h
deleted file mode 100644
index 0fd9ae3f..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/absl_log.h
+++ /dev/null
@@ -1,115 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/absl_log.h
-// -----------------------------------------------------------------------------
-//
-// This header declares a family of `ABSL_LOG` macros as alternative spellings
-// for macros in `log.h`.
-//
-// Basic invocation looks like this:
-//
-//   ABSL_LOG(INFO) << "Found " << num_cookies << " cookies";
-//
-// Most `ABSL_LOG` macros take a severity level argument. The severity levels
-// are `INFO`, `WARNING`, `ERROR`, and `FATAL`.
-//
-// For full documentation, see comments in `log.h`, which includes full
-// reference documentation on use of the equivalent `LOG` macro and has an
-// identical set of macros without the ABSL_* prefix.
-
-#ifndef ABSL_LOG_ABSL_LOG_H_
-#define ABSL_LOG_ABSL_LOG_H_
-
-#include "absl/log/internal/log_impl.h"
-
-#define ABSL_LOG(severity) ABSL_LOG_INTERNAL_LOG_IMPL(_##severity)
-#define ABSL_PLOG(severity) ABSL_LOG_INTERNAL_PLOG_IMPL(_##severity)
-#define ABSL_DLOG(severity) ABSL_LOG_INTERNAL_DLOG_IMPL(_##severity)
-
-#define ABSL_VLOG(verbose_level) ABSL_LOG_INTERNAL_VLOG_IMPL(verbose_level)
-#define ABSL_DVLOG(verbose_level) ABSL_LOG_INTERNAL_DVLOG_IMPL(verbose_level)
-
-#define ABSL_LOG_IF(severity, condition) \
-  ABSL_LOG_INTERNAL_LOG_IF_IMPL(_##severity, condition)
-#define ABSL_PLOG_IF(severity, condition) \
-  ABSL_LOG_INTERNAL_PLOG_IF_IMPL(_##severity, condition)
-#define ABSL_DLOG_IF(severity, condition) \
-  ABSL_LOG_INTERNAL_DLOG_IF_IMPL(_##severity, condition)
-
-#define ABSL_LOG_EVERY_N(severity, n) \
-  ABSL_LOG_INTERNAL_LOG_EVERY_N_IMPL(_##severity, n)
-#define ABSL_LOG_FIRST_N(severity, n) \
-  ABSL_LOG_INTERNAL_LOG_FIRST_N_IMPL(_##severity, n)
-#define ABSL_LOG_EVERY_POW_2(severity) \
-  ABSL_LOG_INTERNAL_LOG_EVERY_POW_2_IMPL(_##severity)
-#define ABSL_LOG_EVERY_N_SEC(severity, n_seconds) \
-  ABSL_LOG_INTERNAL_LOG_EVERY_N_SEC_IMPL(_##severity, n_seconds)
-
-#define ABSL_PLOG_EVERY_N(severity, n) \
-  ABSL_LOG_INTERNAL_PLOG_EVERY_N_IMPL(_##severity, n)
-#define ABSL_PLOG_FIRST_N(severity, n) \
-  ABSL_LOG_INTERNAL_PLOG_FIRST_N_IMPL(_##severity, n)
-#define ABSL_PLOG_EVERY_POW_2(severity) \
-  ABSL_LOG_INTERNAL_PLOG_EVERY_POW_2_IMPL(_##severity)
-#define ABSL_PLOG_EVERY_N_SEC(severity, n_seconds) \
-  ABSL_LOG_INTERNAL_PLOG_EVERY_N_SEC_IMPL(_##severity, n_seconds)
-
-#define ABSL_DLOG_EVERY_N(severity, n) \
-  ABSL_LOG_INTERNAL_DLOG_EVERY_N_IMPL(_##severity, n)
-#define ABSL_DLOG_FIRST_N(severity, n) \
-  ABSL_LOG_INTERNAL_DLOG_FIRST_N_IMPL(_##severity, n)
-#define ABSL_DLOG_EVERY_POW_2(severity) \
-  ABSL_LOG_INTERNAL_DLOG_EVERY_POW_2_IMPL(_##severity)
-#define ABSL_DLOG_EVERY_N_SEC(severity, n_seconds) \
-  ABSL_LOG_INTERNAL_DLOG_EVERY_N_SEC_IMPL(_##severity, n_seconds)
-
-#define ABSL_VLOG_EVERY_N(verbose_level, n) \
-  ABSL_LOG_INTERNAL_VLOG_EVERY_N_IMPL(verbose_level, n)
-#define ABSL_VLOG_FIRST_N(verbose_level, n) \
-  ABSL_LOG_INTERNAL_VLOG_FIRST_N_IMPL(verbose_level, n)
-#define ABSL_VLOG_EVERY_POW_2(verbose_level, n) \
-  ABSL_LOG_INTERNAL_VLOG_EVERY_POW_2_IMPL(verbose_level, n)
-#define ABSL_VLOG_EVERY_N_SEC(verbose_level, n) \
-  ABSL_LOG_INTERNAL_VLOG_EVERY_N_SEC_IMPL(verbose_level, n)
-
-#define ABSL_LOG_IF_EVERY_N(severity, condition, n) \
-  ABSL_LOG_INTERNAL_LOG_IF_EVERY_N_IMPL(_##severity, condition, n)
-#define ABSL_LOG_IF_FIRST_N(severity, condition, n) \
-  ABSL_LOG_INTERNAL_LOG_IF_FIRST_N_IMPL(_##severity, condition, n)
-#define ABSL_LOG_IF_EVERY_POW_2(severity, condition) \
-  ABSL_LOG_INTERNAL_LOG_IF_EVERY_POW_2_IMPL(_##severity, condition)
-#define ABSL_LOG_IF_EVERY_N_SEC(severity, condition, n_seconds) \
-  ABSL_LOG_INTERNAL_LOG_IF_EVERY_N_SEC_IMPL(_##severity, condition, n_seconds)
-
-#define ABSL_PLOG_IF_EVERY_N(severity, condition, n) \
-  ABSL_LOG_INTERNAL_PLOG_IF_EVERY_N_IMPL(_##severity, condition, n)
-#define ABSL_PLOG_IF_FIRST_N(severity, condition, n) \
-  ABSL_LOG_INTERNAL_PLOG_IF_FIRST_N_IMPL(_##severity, condition, n)
-#define ABSL_PLOG_IF_EVERY_POW_2(severity, condition) \
-  ABSL_LOG_INTERNAL_PLOG_IF_EVERY_POW_2_IMPL(_##severity, condition)
-#define ABSL_PLOG_IF_EVERY_N_SEC(severity, condition, n_seconds) \
-  ABSL_LOG_INTERNAL_PLOG_IF_EVERY_N_SEC_IMPL(_##severity, condition, n_seconds)
-
-#define ABSL_DLOG_IF_EVERY_N(severity, condition, n) \
-  ABSL_LOG_INTERNAL_DLOG_IF_EVERY_N_IMPL(_##severity, condition, n)
-#define ABSL_DLOG_IF_FIRST_N(severity, condition, n) \
-  ABSL_LOG_INTERNAL_DLOG_IF_FIRST_N_IMPL(_##severity, condition, n)
-#define ABSL_DLOG_IF_EVERY_POW_2(severity, condition) \
-  ABSL_LOG_INTERNAL_DLOG_IF_EVERY_POW_2_IMPL(_##severity, condition)
-#define ABSL_DLOG_IF_EVERY_N_SEC(severity, condition, n_seconds) \
-  ABSL_LOG_INTERNAL_DLOG_IF_EVERY_N_SEC_IMPL(_##severity, condition, n_seconds)
-
-#endif  // ABSL_LOG_ABSL_LOG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/absl_vlog_is_on.h b/projects/llm_framework/include/abseil-cpp/absl/log/absl_vlog_is_on.h
deleted file mode 100644
index 29096b48..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/absl_vlog_is_on.h
+++ /dev/null
@@ -1,93 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/absl_vlog_is_on.h
-// -----------------------------------------------------------------------------
-//
-// This header defines the `ABSL_VLOG_IS_ON()` macro that controls the
-// variable-verbosity conditional logging.
-//
-// It's used by `VLOG` in log.h, or it can also be used directly like this:
-//
-//   if (ABSL_VLOG_IS_ON(2)) {
-//     foo_server.RecomputeStatisticsExpensive();
-//     LOG(INFO) << foo_server.LastStatisticsAsString();
-//   }
-//
-// Each source file has an effective verbosity level that's a non-negative
-// integer computed from the `--vmodule` and `--v` flags.
-// `ABSL_VLOG_IS_ON(n)` is true, and `VLOG(n)` logs, if that effective verbosity
-// level is greater than or equal to `n`.
-//
-// `--vmodule` takes a comma-delimited list of key=value pairs.  Each key is a
-// pattern matched against filenames, and the values give the effective severity
-// level applied to matching files.  '?' and '*' characters in patterns are
-// interpreted as single-character and zero-or-more-character wildcards.
-// Patterns including a slash character are matched against full pathnames,
-// while those without are matched against basenames only.  One suffix (i.e. the
-// last . and everything after it) is stripped from each filename prior to
-// matching, as is the special suffix "-inl".
-//
-// Files are matched against globs in `--vmodule` in order, and the first match
-// determines the verbosity level.
-//
-// Files which do not match any pattern in `--vmodule` use the value of `--v` as
-// their effective verbosity level.  The default is 0.
-//
-// SetVLOGLevel helper function is provided to do limited dynamic control over
-// V-logging by appending to `--vmodule`. Because these go at the beginning of
-// the list, they take priority over any globs previously added.
-//
-// Resetting --vmodule will override all previous modifications to `--vmodule`,
-// including via SetVLOGLevel.
-
-#ifndef ABSL_LOG_ABSL_VLOG_IS_ON_H_
-#define ABSL_LOG_ABSL_VLOG_IS_ON_H_
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/log/internal/vlog_config.h"  // IWYU pragma: export
-#include "absl/strings/string_view.h"
-
-// IWYU pragma: private, include "absl/log/log.h"
-
-// This is expanded at the callsite to allow the compiler to optimize
-// always-false cases out of the build.
-// An ABSL_MAX_VLOG_VERBOSITY of 2 means that VLOG(3) and above should never
-// log.
-#ifdef ABSL_MAX_VLOG_VERBOSITY
-#define ABSL_LOG_INTERNAL_MAX_LOG_VERBOSITY_CHECK(x) \
-  ((x) <= ABSL_MAX_VLOG_VERBOSITY)&&
-#else
-#define ABSL_LOG_INTERNAL_MAX_LOG_VERBOSITY_CHECK(x)
-#endif
-
-// Each ABSL_VLOG_IS_ON call site gets its own VLogSite that registers with the
-// global linked list of sites to asynchronously update its verbosity level on
-// changes to --v or --vmodule. The verbosity can also be set by manually
-// calling SetVLOGLevel.
-//
-// ABSL_VLOG_IS_ON is not async signal safe, but it is guaranteed not to
-// allocate new memory.
-#define ABSL_VLOG_IS_ON(verbose_level)                                     \
-  (ABSL_LOG_INTERNAL_MAX_LOG_VERBOSITY_CHECK(verbose_level)[]()            \
-       ->::absl::log_internal::VLogSite *                                  \
-   {                                                                       \
-     ABSL_CONST_INIT static ::absl::log_internal::VLogSite site(__FILE__); \
-     return &site;                                                         \
-   }()                                                                     \
-       ->IsEnabled(verbose_level))
-
-#endif  // ABSL_LOG_ABSL_VLOG_IS_ON_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/check.h b/projects/llm_framework/include/abseil-cpp/absl/log/check.h
deleted file mode 100644
index 50f633dd..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/check.h
+++ /dev/null
@@ -1,209 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/check.h
-// -----------------------------------------------------------------------------
-//
-// This header declares a family of `CHECK` macros.
-//
-// `CHECK` macros terminate the program with a fatal error if the specified
-// condition is not true.
-//
-// Except for those whose names begin with `DCHECK`, these macros are not
-// controlled by `NDEBUG` (cf. `assert`), so the check will be executed
-// regardless of compilation mode. `CHECK` and friends are thus useful for
-// confirming invariants in situations where continuing to run would be worse
-// than terminating, e.g., due to risk of data corruption or security
-// compromise.  It is also more robust and portable to deliberately terminate
-// at a particular place with a useful message and backtrace than to assume some
-// ultimately unspecified and unreliable crashing behavior (such as a
-// "segmentation fault").
-
-#ifndef ABSL_LOG_CHECK_H_
-#define ABSL_LOG_CHECK_H_
-
-#include "absl/log/internal/check_impl.h"
-#include "absl/log/internal/check_op.h"     // IWYU pragma: export
-#include "absl/log/internal/conditions.h"   // IWYU pragma: export
-#include "absl/log/internal/log_message.h"  // IWYU pragma: export
-#include "absl/log/internal/strip.h"        // IWYU pragma: export
-
-// CHECK()
-//
-// `CHECK` terminates the program with a fatal error if `condition` is not true.
-//
-// The message may include additional information such as stack traces, when
-// available.
-//
-// Example:
-//
-//   CHECK(!cheese.empty()) << "Out of Cheese";
-//
-// Might produce a message like:
-//
-//   Check failed: !cheese.empty() Out of Cheese
-#define CHECK(condition) ABSL_LOG_INTERNAL_CHECK_IMPL((condition), #condition)
-
-// QCHECK()
-//
-// `QCHECK` behaves like `CHECK` but does not print a full stack trace and does
-// not run registered error handlers (as `QFATAL`).  It is useful when the
-// problem is definitely unrelated to program flow, e.g. when validating user
-// input.
-#define QCHECK(condition) ABSL_LOG_INTERNAL_QCHECK_IMPL((condition), #condition)
-
-// PCHECK()
-//
-// `PCHECK` behaves like `CHECK` but appends a description of the current state
-// of `errno` to the failure message.
-//
-// Example:
-//
-//   int fd = open("/var/empty/missing", O_RDONLY);
-//   PCHECK(fd != -1) << "posix is difficult";
-//
-// Might produce a message like:
-//
-//   Check failed: fd != -1 posix is difficult: No such file or directory [2]
-#define PCHECK(condition) ABSL_LOG_INTERNAL_PCHECK_IMPL((condition), #condition)
-
-// DCHECK()
-//
-// `DCHECK` behaves like `CHECK` in debug mode and does nothing otherwise (as
-// `DLOG`).  Unlike with `CHECK` (but as with `assert`), it is not safe to rely
-// on evaluation of `condition`: when `NDEBUG` is enabled, DCHECK does not
-// evaluate the condition.
-#define DCHECK(condition) ABSL_LOG_INTERNAL_DCHECK_IMPL((condition), #condition)
-
-// `CHECK_EQ` and friends are syntactic sugar for `CHECK(x == y)` that
-// automatically output the expression being tested and the evaluated values on
-// either side.
-//
-// Example:
-//
-//   int x = 3, y = 5;
-//   CHECK_EQ(2 * x, y) << "oops!";
-//
-// Might produce a message like:
-//
-//   Check failed: 2 * x == y (6 vs. 5) oops!
-//
-// The values must implement the appropriate comparison operator as well as
-// `operator<<(std::ostream&, ...)`.  Care is taken to ensure that each
-// argument is evaluated exactly once, and that anything which is legal to pass
-// as a function argument is legal here.  In particular, the arguments may be
-// temporary expressions which will end up being destroyed at the end of the
-// statement,
-//
-// Example:
-//
-//   CHECK_EQ(std::string("abc")[1], 'b');
-//
-// WARNING: Passing `NULL` as an argument to `CHECK_EQ` and similar macros does
-// not compile.  Use `nullptr` instead.
-#define CHECK_EQ(val1, val2) \
-  ABSL_LOG_INTERNAL_CHECK_EQ_IMPL((val1), #val1, (val2), #val2)
-#define CHECK_NE(val1, val2) \
-  ABSL_LOG_INTERNAL_CHECK_NE_IMPL((val1), #val1, (val2), #val2)
-#define CHECK_LE(val1, val2) \
-  ABSL_LOG_INTERNAL_CHECK_LE_IMPL((val1), #val1, (val2), #val2)
-#define CHECK_LT(val1, val2) \
-  ABSL_LOG_INTERNAL_CHECK_LT_IMPL((val1), #val1, (val2), #val2)
-#define CHECK_GE(val1, val2) \
-  ABSL_LOG_INTERNAL_CHECK_GE_IMPL((val1), #val1, (val2), #val2)
-#define CHECK_GT(val1, val2) \
-  ABSL_LOG_INTERNAL_CHECK_GT_IMPL((val1), #val1, (val2), #val2)
-#define QCHECK_EQ(val1, val2) \
-  ABSL_LOG_INTERNAL_QCHECK_EQ_IMPL((val1), #val1, (val2), #val2)
-#define QCHECK_NE(val1, val2) \
-  ABSL_LOG_INTERNAL_QCHECK_NE_IMPL((val1), #val1, (val2), #val2)
-#define QCHECK_LE(val1, val2) \
-  ABSL_LOG_INTERNAL_QCHECK_LE_IMPL((val1), #val1, (val2), #val2)
-#define QCHECK_LT(val1, val2) \
-  ABSL_LOG_INTERNAL_QCHECK_LT_IMPL((val1), #val1, (val2), #val2)
-#define QCHECK_GE(val1, val2) \
-  ABSL_LOG_INTERNAL_QCHECK_GE_IMPL((val1), #val1, (val2), #val2)
-#define QCHECK_GT(val1, val2) \
-  ABSL_LOG_INTERNAL_QCHECK_GT_IMPL((val1), #val1, (val2), #val2)
-#define DCHECK_EQ(val1, val2) \
-  ABSL_LOG_INTERNAL_DCHECK_EQ_IMPL((val1), #val1, (val2), #val2)
-#define DCHECK_NE(val1, val2) \
-  ABSL_LOG_INTERNAL_DCHECK_NE_IMPL((val1), #val1, (val2), #val2)
-#define DCHECK_LE(val1, val2) \
-  ABSL_LOG_INTERNAL_DCHECK_LE_IMPL((val1), #val1, (val2), #val2)
-#define DCHECK_LT(val1, val2) \
-  ABSL_LOG_INTERNAL_DCHECK_LT_IMPL((val1), #val1, (val2), #val2)
-#define DCHECK_GE(val1, val2) \
-  ABSL_LOG_INTERNAL_DCHECK_GE_IMPL((val1), #val1, (val2), #val2)
-#define DCHECK_GT(val1, val2) \
-  ABSL_LOG_INTERNAL_DCHECK_GT_IMPL((val1), #val1, (val2), #val2)
-
-// `CHECK_OK` and friends validate that the provided `absl::Status` or
-// `absl::StatusOr<T>` is OK.  If it isn't, they print a failure message that
-// includes the actual status and terminate the program.
-//
-// As with all `DCHECK` variants, `DCHECK_OK` has no effect (not even
-// evaluating its argument) if `NDEBUG` is enabled.
-//
-// Example:
-//
-//   CHECK_OK(FunctionReturnsStatus(x, y, z)) << "oops!";
-//
-// Might produce a message like:
-//
-//   Check failed: FunctionReturnsStatus(x, y, z) is OK (ABORTED: timeout) oops!
-#define CHECK_OK(status) ABSL_LOG_INTERNAL_CHECK_OK_IMPL((status), #status)
-#define QCHECK_OK(status) ABSL_LOG_INTERNAL_QCHECK_OK_IMPL((status), #status)
-#define DCHECK_OK(status) ABSL_LOG_INTERNAL_DCHECK_OK_IMPL((status), #status)
-
-// `CHECK_STREQ` and friends provide `CHECK_EQ` functionality for C strings,
-// i.e., null-terminated char arrays.  The `CASE` versions are case-insensitive.
-//
-// Example:
-//
-//   CHECK_STREQ(argv[0], "./skynet");
-//
-// Note that both arguments may be temporary strings which are destroyed by the
-// compiler at the end of the current full expression.
-//
-// Example:
-//
-//   CHECK_STREQ(Foo().c_str(), Bar().c_str());
-#define CHECK_STREQ(s1, s2) \
-  ABSL_LOG_INTERNAL_CHECK_STREQ_IMPL((s1), #s1, (s2), #s2)
-#define CHECK_STRNE(s1, s2) \
-  ABSL_LOG_INTERNAL_CHECK_STRNE_IMPL((s1), #s1, (s2), #s2)
-#define CHECK_STRCASEEQ(s1, s2) \
-  ABSL_LOG_INTERNAL_CHECK_STRCASEEQ_IMPL((s1), #s1, (s2), #s2)
-#define CHECK_STRCASENE(s1, s2) \
-  ABSL_LOG_INTERNAL_CHECK_STRCASENE_IMPL((s1), #s1, (s2), #s2)
-#define QCHECK_STREQ(s1, s2) \
-  ABSL_LOG_INTERNAL_QCHECK_STREQ_IMPL((s1), #s1, (s2), #s2)
-#define QCHECK_STRNE(s1, s2) \
-  ABSL_LOG_INTERNAL_QCHECK_STRNE_IMPL((s1), #s1, (s2), #s2)
-#define QCHECK_STRCASEEQ(s1, s2) \
-  ABSL_LOG_INTERNAL_QCHECK_STRCASEEQ_IMPL((s1), #s1, (s2), #s2)
-#define QCHECK_STRCASENE(s1, s2) \
-  ABSL_LOG_INTERNAL_QCHECK_STRCASENE_IMPL((s1), #s1, (s2), #s2)
-#define DCHECK_STREQ(s1, s2) \
-  ABSL_LOG_INTERNAL_DCHECK_STREQ_IMPL((s1), #s1, (s2), #s2)
-#define DCHECK_STRNE(s1, s2) \
-  ABSL_LOG_INTERNAL_DCHECK_STRNE_IMPL((s1), #s1, (s2), #s2)
-#define DCHECK_STRCASEEQ(s1, s2) \
-  ABSL_LOG_INTERNAL_DCHECK_STRCASEEQ_IMPL((s1), #s1, (s2), #s2)
-#define DCHECK_STRCASENE(s1, s2) \
-  ABSL_LOG_INTERNAL_DCHECK_STRCASENE_IMPL((s1), #s1, (s2), #s2)
-
-#endif  // ABSL_LOG_CHECK_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/die_if_null.h b/projects/llm_framework/include/abseil-cpp/absl/log/die_if_null.h
deleted file mode 100644
index 127a9ac8..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/die_if_null.h
+++ /dev/null
@@ -1,76 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/die_if_null.h
-// -----------------------------------------------------------------------------
-//
-// This header declares macro `ABSL_DIE_IF_NULL`.
-
-#ifndef ABSL_LOG_DIE_IF_NULL_H_
-#define ABSL_LOG_DIE_IF_NULL_H_
-
-#include <stdint.h>
-
-#include <utility>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/optimization.h"
-
-// ABSL_DIE_IF_NULL()
-//
-// `ABSL_DIE_IF_NULL` behaves as `CHECK_NE` against `nullptr` but *also*
-// "returns" its argument.  It is useful in initializers where statements (like
-// `CHECK_NE`) can't be used.  Outside initializers, prefer `CHECK` or
-// `CHECK_NE`. `ABSL_DIE_IF_NULL` works for both raw pointers and (compatible)
-// smart pointers including `std::unique_ptr` and `std::shared_ptr`; more
-// generally, it works for any type that can be compared to nullptr_t.  For
-// types that aren't raw pointers, `ABSL_DIE_IF_NULL` returns a reference to
-// its argument, preserving the value category. Example:
-//
-//   Foo() : bar_(ABSL_DIE_IF_NULL(MethodReturningUniquePtr())) {}
-//
-// Use `CHECK(ptr)` or `CHECK(ptr != nullptr)` if the returned pointer is
-// unused.
-#define ABSL_DIE_IF_NULL(val) \
-  ::absl::log_internal::DieIfNull(__FILE__, __LINE__, #val, (val))
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-
-// Crashes the process after logging `exprtext` annotated at the `file` and
-// `line` location. Called when `ABSL_DIE_IF_NULL` fails. Calling this function
-// generates less code than its implementation would if inlined, for a slight
-// code size reduction each time `ABSL_DIE_IF_NULL` is called.
-ABSL_ATTRIBUTE_NORETURN ABSL_ATTRIBUTE_NOINLINE void DieBecauseNull(
-    const char* file, int line, const char* exprtext);
-
-// Helper for `ABSL_DIE_IF_NULL`.
-template <typename T>
-ABSL_MUST_USE_RESULT T DieIfNull(const char* file, int line,
-                                 const char* exprtext, T&& t) {
-  if (ABSL_PREDICT_FALSE(t == nullptr)) {
-    // Call a non-inline helper function for a small code size improvement.
-    DieBecauseNull(file, line, exprtext);
-  }
-  return std::forward<T>(t);
-}
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_DIE_IF_NULL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/flags.h b/projects/llm_framework/include/abseil-cpp/absl/log/flags.h
deleted file mode 100644
index 146cfdd6..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/flags.h
+++ /dev/null
@@ -1,43 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/flags.h
-// -----------------------------------------------------------------------------
-//
-
-#ifndef ABSL_LOG_FLAGS_H_
-#define ABSL_LOG_FLAGS_H_
-
-// The Abseil Logging library supports the following command line flags to
-// configure logging behavior at runtime:
-//
-// --stderrthreshold=<value>
-// Log messages at or above this threshold level are copied to stderr.
-//
-// --minloglevel=<value>
-// Messages logged at a lower level than this are discarded and don't actually
-// get logged anywhere.
-//
-// --log_backtrace_at=<file:linenum>
-// Emit a backtrace (stack trace) when logging at file:linenum.
-//
-// To use these commandline flags, the //absl/log:flags library must be
-// explicitly linked, and absl::ParseCommandLine() must be called before the
-// call to absl::InitializeLog().
-//
-// To configure the Log library programmatically, use the interfaces defined in
-// absl/log/globals.h.
-
-#endif  // ABSL_LOG_FLAGS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/globals.h b/projects/llm_framework/include/abseil-cpp/absl/log/globals.h
deleted file mode 100644
index b36e47e6..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/globals.h
+++ /dev/null
@@ -1,218 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/globals.h
-// -----------------------------------------------------------------------------
-//
-// This header declares global logging library configuration knobs.
-
-#ifndef ABSL_LOG_GLOBALS_H_
-#define ABSL_LOG_GLOBALS_H_
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/log_severity.h"
-#include "absl/log/internal/vlog_config.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-//------------------------------------------------------------------------------
-//  Minimum Log Level
-//------------------------------------------------------------------------------
-//
-// Messages logged at or above this severity are directed to all registered log
-// sinks or skipped otherwise. This parameter can also be modified using
-// command line flag --minloglevel.
-// See absl/base/log_severity.h for descriptions of severity levels.
-
-// MinLogLevel()
-//
-// Returns the value of the Minimum Log Level parameter.
-// This function is async-signal-safe.
-ABSL_MUST_USE_RESULT absl::LogSeverityAtLeast MinLogLevel();
-
-// SetMinLogLevel()
-//
-// Updates the value of Minimum Log Level parameter.
-// This function is async-signal-safe.
-void SetMinLogLevel(absl::LogSeverityAtLeast severity);
-
-namespace log_internal {
-
-// ScopedMinLogLevel
-//
-// RAII type used to temporarily update the Min Log Level parameter.
-class ScopedMinLogLevel final {
- public:
-  explicit ScopedMinLogLevel(absl::LogSeverityAtLeast severity);
-  ScopedMinLogLevel(const ScopedMinLogLevel&) = delete;
-  ScopedMinLogLevel& operator=(const ScopedMinLogLevel&) = delete;
-  ~ScopedMinLogLevel();
-
- private:
-  absl::LogSeverityAtLeast saved_severity_;
-};
-
-}  // namespace log_internal
-
-//------------------------------------------------------------------------------
-// Stderr Threshold
-//------------------------------------------------------------------------------
-//
-// Messages logged at or above this level are directed to stderr in
-// addition to other registered log sinks. This parameter can also be modified
-// using command line flag --stderrthreshold.
-// See absl/base/log_severity.h for descriptions of severity levels.
-
-// StderrThreshold()
-//
-// Returns the value of the Stderr Threshold parameter.
-// This function is async-signal-safe.
-ABSL_MUST_USE_RESULT absl::LogSeverityAtLeast StderrThreshold();
-
-// SetStderrThreshold()
-//
-// Updates the Stderr Threshold parameter.
-// This function is async-signal-safe.
-void SetStderrThreshold(absl::LogSeverityAtLeast severity);
-inline void SetStderrThreshold(absl::LogSeverity severity) {
-  absl::SetStderrThreshold(static_cast<absl::LogSeverityAtLeast>(severity));
-}
-
-// ScopedStderrThreshold
-//
-// RAII type used to temporarily update the Stderr Threshold parameter.
-class ScopedStderrThreshold final {
- public:
-  explicit ScopedStderrThreshold(absl::LogSeverityAtLeast severity);
-  ScopedStderrThreshold(const ScopedStderrThreshold&) = delete;
-  ScopedStderrThreshold& operator=(const ScopedStderrThreshold&) = delete;
-  ~ScopedStderrThreshold();
-
- private:
-  absl::LogSeverityAtLeast saved_severity_;
-};
-
-//------------------------------------------------------------------------------
-// Log Backtrace At
-//------------------------------------------------------------------------------
-//
-// Users can request an existing `LOG` statement, specified by file and line
-// number, to also include a backtrace when logged.
-
-// ShouldLogBacktraceAt()
-//
-// Returns true if we should log a backtrace at the specified location.
-namespace log_internal {
-ABSL_MUST_USE_RESULT bool ShouldLogBacktraceAt(absl::string_view file,
-                                               int line);
-}  // namespace log_internal
-
-// SetLogBacktraceLocation()
-//
-// Sets the location the backtrace should be logged at.  If the specified
-// location isn't a `LOG` statement, the effect will be the same as
-// `ClearLogBacktraceLocation` (but less efficient).
-void SetLogBacktraceLocation(absl::string_view file, int line);
-
-// ClearLogBacktraceLocation()
-//
-// Clears the set location so that backtraces will no longer be logged at it.
-void ClearLogBacktraceLocation();
-
-//------------------------------------------------------------------------------
-// Prepend Log Prefix
-//------------------------------------------------------------------------------
-//
-// This option tells the logging library that every logged message
-// should include the prefix (severity, date, time, PID, etc.)
-
-// ShouldPrependLogPrefix()
-//
-// Returns the value of the Prepend Log Prefix option.
-// This function is async-signal-safe.
-ABSL_MUST_USE_RESULT bool ShouldPrependLogPrefix();
-
-// EnableLogPrefix()
-//
-// Updates the value of the Prepend Log Prefix option.
-// This function is async-signal-safe.
-void EnableLogPrefix(bool on_off);
-
-//------------------------------------------------------------------------------
-// Set Global VLOG Level
-//------------------------------------------------------------------------------
-//
-// Sets the global `(ABSL_)VLOG(_IS_ON)` level to `log_level`.  This level is
-// applied to any sites whose filename doesn't match any `module_pattern`.
-// Returns the prior value.
-inline int SetGlobalVLogLevel(int log_level) {
-  return absl::log_internal::UpdateGlobalVLogLevel(log_level);
-}
-
-//------------------------------------------------------------------------------
-// Set VLOG Level
-//------------------------------------------------------------------------------
-//
-// Sets `(ABSL_)VLOG(_IS_ON)` level for `module_pattern` to `log_level`.  This
-// allows programmatic control of what is normally set by the --vmodule flag.
-// Returns the level that previously applied to `module_pattern`.
-inline int SetVLogLevel(absl::string_view module_pattern, int log_level) {
-  return absl::log_internal::PrependVModule(module_pattern, log_level);
-}
-
-//------------------------------------------------------------------------------
-// Configure Android Native Log Tag
-//------------------------------------------------------------------------------
-//
-// The logging library forwards to the Android system log API when built for
-// Android.  That API takes a string "tag" value in addition to a message and
-// severity level.  The tag is used to identify the source of messages and to
-// filter them.  This library uses the tag "native" by default.
-
-// SetAndroidNativeTag()
-//
-// Stores a copy of the string pointed to by `tag` and uses it as the Android
-// logging tag thereafter. `tag` must not be null.
-// This function must not be called more than once!
-void SetAndroidNativeTag(const char* tag);
-
-namespace log_internal {
-// GetAndroidNativeTag()
-//
-// Returns the configured Android logging tag.
-const char* GetAndroidNativeTag();
-}  // namespace log_internal
-
-namespace log_internal {
-
-using LoggingGlobalsListener = void (*)();
-void SetLoggingGlobalsListener(LoggingGlobalsListener l);
-
-// Internal implementation for the setter routines. These are used
-// to break circular dependencies between flags and globals. Each "Raw"
-// routine corresponds to the non-"Raw" counterpart and used to set the
-// configuration parameter directly without calling back to the listener.
-void RawSetMinLogLevel(absl::LogSeverityAtLeast severity);
-void RawSetStderrThreshold(absl::LogSeverityAtLeast severity);
-void RawEnableLogPrefix(bool on_off);
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_GLOBALS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/initialize.h b/projects/llm_framework/include/abseil-cpp/absl/log/initialize.h
deleted file mode 100644
index f600eb60..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/initialize.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/initialize.h
-// -----------------------------------------------------------------------------
-//
-// This header declares the Abseil Log initialization routine InitializeLog().
-
-#ifndef ABSL_LOG_INITIALIZE_H_
-#define ABSL_LOG_INITIALIZE_H_
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// InitializeLog()
-//
-// Initializes the Abseil logging library.
-//
-// Before this function is called, all log messages are directed only to stderr.
-// After initialization is finished, log messages are directed to all registered
-// `LogSink`s.
-//
-// It is an error to call this function twice.
-//
-// There is no corresponding function to shut down the logging library.
-void InitializeLog();
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INITIALIZE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/append_truncated.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/append_truncated.h
deleted file mode 100644
index f0e7912c..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/append_truncated.h
+++ /dev/null
@@ -1,47 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_LOG_INTERNAL_APPEND_TRUNCATED_H_
-#define ABSL_LOG_INTERNAL_APPEND_TRUNCATED_H_
-
-#include <cstddef>
-#include <cstring>
-
-#include "absl/base/config.h"
-#include "absl/strings/string_view.h"
-#include "absl/types/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-// Copies into `dst` as many bytes of `src` as will fit, then truncates the
-// copied bytes from the front of `dst` and returns the number of bytes written.
-inline size_t AppendTruncated(absl::string_view src, absl::Span<char> &dst) {
-  if (src.size() > dst.size()) src = src.substr(0, dst.size());
-  memcpy(dst.data(), src.data(), src.size());
-  dst.remove_prefix(src.size());
-  return src.size();
-}
-// Likewise, but `n` copies of `c`.
-inline size_t AppendTruncated(char c, size_t n, absl::Span<char> &dst) {
-  if (n > dst.size()) n = dst.size();
-  memset(dst.data(), c, n);
-  dst.remove_prefix(n);
-  return n;
-}
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_APPEND_TRUNCATED_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/check_impl.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/check_impl.h
deleted file mode 100644
index 00f25f80..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/check_impl.h
+++ /dev/null
@@ -1,150 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_LOG_INTERNAL_CHECK_IMPL_H_
-#define ABSL_LOG_INTERNAL_CHECK_IMPL_H_
-
-#include "absl/base/optimization.h"
-#include "absl/log/internal/check_op.h"
-#include "absl/log/internal/conditions.h"
-#include "absl/log/internal/log_message.h"
-#include "absl/log/internal/strip.h"
-
-// CHECK
-#define ABSL_LOG_INTERNAL_CHECK_IMPL(condition, condition_text)       \
-  ABSL_LOG_INTERNAL_CONDITION_FATAL(STATELESS,                        \
-                                    ABSL_PREDICT_FALSE(!(condition))) \
-  ABSL_LOG_INTERNAL_CHECK(condition_text).InternalStream()
-
-#define ABSL_LOG_INTERNAL_QCHECK_IMPL(condition, condition_text)       \
-  ABSL_LOG_INTERNAL_CONDITION_QFATAL(STATELESS,                        \
-                                     ABSL_PREDICT_FALSE(!(condition))) \
-  ABSL_LOG_INTERNAL_QCHECK(condition_text).InternalStream()
-
-#define ABSL_LOG_INTERNAL_PCHECK_IMPL(condition, condition_text) \
-  ABSL_LOG_INTERNAL_CHECK_IMPL(condition, condition_text).WithPerror()
-
-#ifndef NDEBUG
-#define ABSL_LOG_INTERNAL_DCHECK_IMPL(condition, condition_text) \
-  ABSL_LOG_INTERNAL_CHECK_IMPL(condition, condition_text)
-#else
-#define ABSL_LOG_INTERNAL_DCHECK_IMPL(condition, condition_text) \
-  ABSL_LOG_INTERNAL_CHECK_IMPL(true || (condition), "true")
-#endif
-
-// CHECK_EQ
-#define ABSL_LOG_INTERNAL_CHECK_EQ_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_CHECK_OP(Check_EQ, ==, val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_CHECK_NE_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_CHECK_OP(Check_NE, !=, val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_CHECK_LE_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_CHECK_OP(Check_LE, <=, val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_CHECK_LT_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_CHECK_OP(Check_LT, <, val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_CHECK_GE_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_CHECK_OP(Check_GE, >=, val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_CHECK_GT_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_CHECK_OP(Check_GT, >, val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_QCHECK_EQ_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_QCHECK_OP(Check_EQ, ==, val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_QCHECK_NE_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_QCHECK_OP(Check_NE, !=, val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_QCHECK_LE_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_QCHECK_OP(Check_LE, <=, val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_QCHECK_LT_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_QCHECK_OP(Check_LT, <, val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_QCHECK_GE_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_QCHECK_OP(Check_GE, >=, val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_QCHECK_GT_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_QCHECK_OP(Check_GT, >, val1, val1_text, val2, val2_text)
-#ifndef NDEBUG
-#define ABSL_LOG_INTERNAL_DCHECK_EQ_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_CHECK_EQ_IMPL(val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_DCHECK_NE_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_CHECK_NE_IMPL(val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_DCHECK_LE_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_CHECK_LE_IMPL(val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_DCHECK_LT_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_CHECK_LT_IMPL(val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_DCHECK_GE_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_CHECK_GE_IMPL(val1, val1_text, val2, val2_text)
-#define ABSL_LOG_INTERNAL_DCHECK_GT_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_CHECK_GT_IMPL(val1, val1_text, val2, val2_text)
-#else  // ndef NDEBUG
-#define ABSL_LOG_INTERNAL_DCHECK_EQ_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_DCHECK_NOP(val1, val2)
-#define ABSL_LOG_INTERNAL_DCHECK_NE_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_DCHECK_NOP(val1, val2)
-#define ABSL_LOG_INTERNAL_DCHECK_LE_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_DCHECK_NOP(val1, val2)
-#define ABSL_LOG_INTERNAL_DCHECK_LT_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_DCHECK_NOP(val1, val2)
-#define ABSL_LOG_INTERNAL_DCHECK_GE_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_DCHECK_NOP(val1, val2)
-#define ABSL_LOG_INTERNAL_DCHECK_GT_IMPL(val1, val1_text, val2, val2_text) \
-  ABSL_LOG_INTERNAL_DCHECK_NOP(val1, val2)
-#endif  // def NDEBUG
-
-// CHECK_OK
-#define ABSL_LOG_INTERNAL_CHECK_OK_IMPL(status, status_text) \
-  ABSL_LOG_INTERNAL_CHECK_OK(status, status_text)
-#define ABSL_LOG_INTERNAL_QCHECK_OK_IMPL(status, status_text) \
-  ABSL_LOG_INTERNAL_QCHECK_OK(status, status_text)
-#ifndef NDEBUG
-#define ABSL_LOG_INTERNAL_DCHECK_OK_IMPL(status, status_text) \
-  ABSL_LOG_INTERNAL_CHECK_OK(status, status_text)
-#else
-#define ABSL_LOG_INTERNAL_DCHECK_OK_IMPL(status, status_text) \
-  ABSL_LOG_INTERNAL_DCHECK_NOP(status, nullptr)
-#endif
-
-// CHECK_STREQ
-#define ABSL_LOG_INTERNAL_CHECK_STREQ_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_CHECK_STROP(strcmp, ==, true, s1, s1_text, s2, s2_text)
-#define ABSL_LOG_INTERNAL_CHECK_STRNE_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_CHECK_STROP(strcmp, !=, false, s1, s1_text, s2, s2_text)
-#define ABSL_LOG_INTERNAL_CHECK_STRCASEEQ_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_CHECK_STROP(strcasecmp, ==, true, s1, s1_text, s2, s2_text)
-#define ABSL_LOG_INTERNAL_CHECK_STRCASENE_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_CHECK_STROP(strcasecmp, !=, false, s1, s1_text, s2, s2_text)
-#define ABSL_LOG_INTERNAL_QCHECK_STREQ_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_QCHECK_STROP(strcmp, ==, true, s1, s1_text, s2, s2_text)
-#define ABSL_LOG_INTERNAL_QCHECK_STRNE_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_QCHECK_STROP(strcmp, !=, false, s1, s1_text, s2, s2_text)
-#define ABSL_LOG_INTERNAL_QCHECK_STRCASEEQ_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_QCHECK_STROP(strcasecmp, ==, true, s1, s1_text, s2, s2_text)
-#define ABSL_LOG_INTERNAL_QCHECK_STRCASENE_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_QCHECK_STROP(strcasecmp, !=, false, s1, s1_text, s2,  \
-                                 s2_text)
-#ifndef NDEBUG
-#define ABSL_LOG_INTERNAL_DCHECK_STREQ_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_CHECK_STREQ_IMPL(s1, s1_text, s2, s2_text)
-#define ABSL_LOG_INTERNAL_DCHECK_STRCASEEQ_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_CHECK_STRCASEEQ_IMPL(s1, s1_text, s2, s2_text)
-#define ABSL_LOG_INTERNAL_DCHECK_STRNE_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_CHECK_STRNE_IMPL(s1, s1_text, s2, s2_text)
-#define ABSL_LOG_INTERNAL_DCHECK_STRCASENE_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_CHECK_STRCASENE_IMPL(s1, s1_text, s2, s2_text)
-#else  // ndef NDEBUG
-#define ABSL_LOG_INTERNAL_DCHECK_STREQ_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_DCHECK_NOP(s1, s2)
-#define ABSL_LOG_INTERNAL_DCHECK_STRCASEEQ_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_DCHECK_NOP(s1, s2)
-#define ABSL_LOG_INTERNAL_DCHECK_STRNE_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_DCHECK_NOP(s1, s2)
-#define ABSL_LOG_INTERNAL_DCHECK_STRCASENE_IMPL(s1, s1_text, s2, s2_text) \
-  ABSL_LOG_INTERNAL_DCHECK_NOP(s1, s2)
-#endif  // def NDEBUG
-
-#endif  // ABSL_LOG_INTERNAL_CHECK_IMPL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/check_op.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/check_op.h
deleted file mode 100644
index 11f0f407..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/check_op.h
+++ /dev/null
@@ -1,420 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/internal/check_op.h
-// -----------------------------------------------------------------------------
-//
-// This file declares helpers routines and macros used to implement `CHECK`
-// macros.
-
-#ifndef ABSL_LOG_INTERNAL_CHECK_OP_H_
-#define ABSL_LOG_INTERNAL_CHECK_OP_H_
-
-#include <stdint.h>
-
-#include <ostream>
-#include <sstream>
-#include <string>
-#include <utility>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/optimization.h"
-#include "absl/log/internal/nullguard.h"
-#include "absl/log/internal/nullstream.h"
-#include "absl/log/internal/strip.h"
-
-// `ABSL_LOG_INTERNAL_STRIP_STRING_LITERAL` wraps string literals that
-// should be stripped when `ABSL_MIN_LOG_LEVEL` exceeds `kFatal`.
-#ifdef ABSL_MIN_LOG_LEVEL
-#define ABSL_LOG_INTERNAL_STRIP_STRING_LITERAL(literal)         \
-  (::absl::LogSeverity::kFatal >=                               \
-           static_cast<::absl::LogSeverity>(ABSL_MIN_LOG_LEVEL) \
-       ? (literal)                                              \
-       : "")
-#else
-#define ABSL_LOG_INTERNAL_STRIP_STRING_LITERAL(literal) (literal)
-#endif
-
-#ifdef NDEBUG
-// `NDEBUG` is defined, so `DCHECK_EQ(x, y)` and so on do nothing.  However, we
-// still want the compiler to parse `x` and `y`, because we don't want to lose
-// potentially useful errors and warnings.
-#define ABSL_LOG_INTERNAL_DCHECK_NOP(x, y)   \
-  while (false && ((void)(x), (void)(y), 0)) \
-  ::absl::log_internal::NullStream().InternalStream()
-#endif
-
-#define ABSL_LOG_INTERNAL_CHECK_OP(name, op, val1, val1_text, val2, val2_text) \
-  while (                                                                      \
-      ::std::string* absl_log_internal_check_op_result ABSL_ATTRIBUTE_UNUSED = \
-          ::absl::log_internal::name##Impl(                                    \
-              ::absl::log_internal::GetReferenceableValue(val1),               \
-              ::absl::log_internal::GetReferenceableValue(val2),               \
-              ABSL_LOG_INTERNAL_STRIP_STRING_LITERAL(val1_text                 \
-                                                     " " #op " " val2_text)))  \
-    ABSL_LOG_INTERNAL_CONDITION_FATAL(STATELESS, true)                         \
-  ABSL_LOG_INTERNAL_CHECK(*absl_log_internal_check_op_result).InternalStream()
-#define ABSL_LOG_INTERNAL_QCHECK_OP(name, op, val1, val1_text, val2, \
-                                    val2_text)                       \
-  while (::std::string* absl_log_internal_qcheck_op_result =         \
-             ::absl::log_internal::name##Impl(                       \
-                 ::absl::log_internal::GetReferenceableValue(val1),  \
-                 ::absl::log_internal::GetReferenceableValue(val2),  \
-                 ABSL_LOG_INTERNAL_STRIP_STRING_LITERAL(             \
-                     val1_text " " #op " " val2_text)))              \
-    ABSL_LOG_INTERNAL_CONDITION_QFATAL(STATELESS, true)              \
-  ABSL_LOG_INTERNAL_QCHECK(*absl_log_internal_qcheck_op_result).InternalStream()
-#define ABSL_LOG_INTERNAL_CHECK_STROP(func, op, expected, s1, s1_text, s2,     \
-                                      s2_text)                                 \
-  while (::std::string* absl_log_internal_check_strop_result =                 \
-             ::absl::log_internal::Check##func##expected##Impl(                \
-                 (s1), (s2),                                                   \
-                 ABSL_LOG_INTERNAL_STRIP_STRING_LITERAL(s1_text " " #op        \
-                                                                " " s2_text))) \
-    ABSL_LOG_INTERNAL_CONDITION_FATAL(STATELESS, true)                         \
-  ABSL_LOG_INTERNAL_CHECK(*absl_log_internal_check_strop_result)               \
-      .InternalStream()
-#define ABSL_LOG_INTERNAL_QCHECK_STROP(func, op, expected, s1, s1_text, s2,    \
-                                       s2_text)                                \
-  while (::std::string* absl_log_internal_qcheck_strop_result =                \
-             ::absl::log_internal::Check##func##expected##Impl(                \
-                 (s1), (s2),                                                   \
-                 ABSL_LOG_INTERNAL_STRIP_STRING_LITERAL(s1_text " " #op        \
-                                                                " " s2_text))) \
-    ABSL_LOG_INTERNAL_CONDITION_QFATAL(STATELESS, true)                        \
-  ABSL_LOG_INTERNAL_QCHECK(*absl_log_internal_qcheck_strop_result)             \
-      .InternalStream()
-// This one is tricky:
-// * We must evaluate `val` exactly once, yet we need to do two things with it:
-//   evaluate `.ok()` and (sometimes) `.ToString()`.
-// * `val` might be an `absl::Status` or some `absl::StatusOr<T>`.
-// * `val` might be e.g. `ATemporary().GetStatus()`, which may return a
-//   reference to a member of `ATemporary` that is only valid until the end of
-//   the full expression.
-// * We don't want this file to depend on `absl::Status` `#include`s or linkage,
-//   nor do we want to move the definition to status and introduce a dependency
-//   in the other direction.  We can be assured that callers must already have a
-//   `Status` and the necessary `#include`s and linkage.
-// * Callsites should be small and fast (at least when `val.ok()`): one branch,
-//   minimal stack footprint.
-//   * In particular, the string concat stuff should be out-of-line and emitted
-//     in only one TU to save linker input size
-// * We want the `val.ok()` check inline so static analyzers and optimizers can
-//   see it.
-// * As usual, no braces so we can stream into the expansion with `operator<<`.
-// * Also as usual, it must expand to a single (partial) statement with no
-//   ambiguous-else problems.
-// * When stripped by `ABSL_MIN_LOG_LEVEL`, we must discard the `<expr> is OK`
-//   string literal and abort without doing any streaming.  We don't need to
-//   strip the call to stringify the non-ok `Status` as long as we don't log it;
-//   dropping the `Status`'s message text is out of scope.
-#define ABSL_LOG_INTERNAL_CHECK_OK(val, val_text)                        \
-  for (::std::pair<const ::absl::Status*, ::std::string*>                \
-           absl_log_internal_check_ok_goo;                               \
-       absl_log_internal_check_ok_goo.first =                            \
-           ::absl::log_internal::AsStatus(val),                          \
-       absl_log_internal_check_ok_goo.second =                           \
-           ABSL_PREDICT_TRUE(absl_log_internal_check_ok_goo.first->ok()) \
-               ? nullptr                                                 \
-               : ::absl::status_internal::MakeCheckFailString(           \
-                     absl_log_internal_check_ok_goo.first,               \
-                     ABSL_LOG_INTERNAL_STRIP_STRING_LITERAL(val_text     \
-                                                            " is OK")),  \
-       !ABSL_PREDICT_TRUE(absl_log_internal_check_ok_goo.first->ok());)  \
-    ABSL_LOG_INTERNAL_CONDITION_FATAL(STATELESS, true)                   \
-  ABSL_LOG_INTERNAL_CHECK(*absl_log_internal_check_ok_goo.second)        \
-      .InternalStream()
-#define ABSL_LOG_INTERNAL_QCHECK_OK(val, val_text)                        \
-  for (::std::pair<const ::absl::Status*, ::std::string*>                 \
-           absl_log_internal_qcheck_ok_goo;                               \
-       absl_log_internal_qcheck_ok_goo.first =                            \
-           ::absl::log_internal::AsStatus(val),                           \
-       absl_log_internal_qcheck_ok_goo.second =                           \
-           ABSL_PREDICT_TRUE(absl_log_internal_qcheck_ok_goo.first->ok()) \
-               ? nullptr                                                  \
-               : ::absl::status_internal::MakeCheckFailString(            \
-                     absl_log_internal_qcheck_ok_goo.first,               \
-                     ABSL_LOG_INTERNAL_STRIP_STRING_LITERAL(val_text      \
-                                                            " is OK")),   \
-       !ABSL_PREDICT_TRUE(absl_log_internal_qcheck_ok_goo.first->ok());)  \
-    ABSL_LOG_INTERNAL_CONDITION_QFATAL(STATELESS, true)                   \
-  ABSL_LOG_INTERNAL_QCHECK(*absl_log_internal_qcheck_ok_goo.second)       \
-      .InternalStream()
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-class Status;
-template <typename T>
-class StatusOr;
-
-namespace status_internal {
-ABSL_ATTRIBUTE_PURE_FUNCTION std::string* MakeCheckFailString(
-    const absl::Status* status, const char* prefix);
-}  // namespace status_internal
-
-namespace log_internal {
-
-// Convert a Status or a StatusOr to its underlying status value.
-//
-// (This implementation does not require a dep on absl::Status to work.)
-inline const absl::Status* AsStatus(const absl::Status& s) { return &s; }
-template <typename T>
-const absl::Status* AsStatus(const absl::StatusOr<T>& s) {
-  return &s.status();
-}
-
-// A helper class for formatting `expr (V1 vs. V2)` in a `CHECK_XX` statement.
-// See `MakeCheckOpString` for sample usage.
-class CheckOpMessageBuilder final {
- public:
-  // Inserts `exprtext` and ` (` to the stream.
-  explicit CheckOpMessageBuilder(const char* exprtext);
-  ~CheckOpMessageBuilder() = default;
-  // For inserting the first variable.
-  std::ostream& ForVar1() { return stream_; }
-  // For inserting the second variable (adds an intermediate ` vs. `).
-  std::ostream& ForVar2();
-  // Get the result (inserts the closing `)`).
-  std::string* NewString();
-
- private:
-  std::ostringstream stream_;
-};
-
-// This formats a value for a failing `CHECK_XX` statement.  Ordinarily, it uses
-// the definition for `operator<<`, with a few special cases below.
-template <typename T>
-inline void MakeCheckOpValueString(std::ostream& os, const T& v) {
-  os << log_internal::NullGuard<T>::Guard(v);
-}
-
-// Overloads for char types provide readable values for unprintable characters.
-void MakeCheckOpValueString(std::ostream& os, char v);
-void MakeCheckOpValueString(std::ostream& os, signed char v);
-void MakeCheckOpValueString(std::ostream& os, unsigned char v);
-void MakeCheckOpValueString(std::ostream& os, const void* p);
-
-namespace detect_specialization {
-
-// MakeCheckOpString is being specialized for every T and U pair that is being
-// passed to the CHECK_op macros. However, there is a lot of redundancy in these
-// specializations that creates unnecessary library and binary bloat.
-// The number of instantiations tends to be O(n^2) because we have two
-// independent inputs. This technique works by reducing `n`.
-//
-// Most user-defined types being passed to CHECK_op end up being printed as a
-// builtin type. For example, enums tend to be implicitly converted to its
-// underlying type when calling operator<<, and pointers are printed with the
-// `const void*` overload.
-// To reduce the number of instantiations we coerce these values before calling
-// MakeCheckOpString instead of inside it.
-//
-// To detect if this coercion is needed, we duplicate all the relevant
-// operator<< overloads as specified in the standard, just in a different
-// namespace. If the call to `stream << value` becomes ambiguous, it means that
-// one of these overloads is the one selected by overload resolution. We then
-// do overload resolution again just with our overload set to see which one gets
-// selected. That tells us which type to coerce to.
-// If the augmented call was not ambiguous, it means that none of these were
-// selected and we can't coerce the input.
-//
-// As a secondary step to reduce code duplication, we promote integral types to
-// their 64-bit variant. This does not change the printed value, but reduces the
-// number of instantiations even further. Promoting an integer is very cheap at
-// the call site.
-int64_t operator<<(std::ostream&, short value);           // NOLINT
-int64_t operator<<(std::ostream&, unsigned short value);  // NOLINT
-int64_t operator<<(std::ostream&, int value);
-int64_t operator<<(std::ostream&, unsigned int value);
-int64_t operator<<(std::ostream&, long value);                 // NOLINT
-uint64_t operator<<(std::ostream&, unsigned long value);       // NOLINT
-int64_t operator<<(std::ostream&, long long value);            // NOLINT
-uint64_t operator<<(std::ostream&, unsigned long long value);  // NOLINT
-float operator<<(std::ostream&, float value);
-double operator<<(std::ostream&, double value);
-long double operator<<(std::ostream&, long double value);
-bool operator<<(std::ostream&, bool value);
-const void* operator<<(std::ostream&, const void* value);
-const void* operator<<(std::ostream&, std::nullptr_t);
-
-// These `char` overloads are specified like this in the standard, so we have to
-// write them exactly the same to ensure the call is ambiguous.
-// If we wrote it in a different way (eg taking std::ostream instead of the
-// template) then one call might have a higher rank than the other and it would
-// not be ambiguous.
-template <typename Traits>
-char operator<<(std::basic_ostream<char, Traits>&, char);
-template <typename Traits>
-signed char operator<<(std::basic_ostream<char, Traits>&, signed char);
-template <typename Traits>
-unsigned char operator<<(std::basic_ostream<char, Traits>&, unsigned char);
-template <typename Traits>
-const char* operator<<(std::basic_ostream<char, Traits>&, const char*);
-template <typename Traits>
-const signed char* operator<<(std::basic_ostream<char, Traits>&,
-                              const signed char*);
-template <typename Traits>
-const unsigned char* operator<<(std::basic_ostream<char, Traits>&,
-                                const unsigned char*);
-
-// This overload triggers when the call is not ambiguous.
-// It means that T is being printed with some overload not on this list.
-// We keep the value as `const T&`.
-template <typename T, typename = decltype(std::declval<std::ostream&>()
-                                          << std::declval<const T&>())>
-const T& Detect(int);
-
-// This overload triggers when the call is ambiguous.
-// It means that T is either one from this list or printed as one from this
-// list. Eg an enum that decays to `int` for printing.
-// We ask the overload set to give us the type we want to convert it to.
-template <typename T>
-decltype(detect_specialization::operator<<(std::declval<std::ostream&>(),
-                                           std::declval<const T&>()))
-Detect(char);
-
-}  // namespace detect_specialization
-
-template <typename T>
-using CheckOpStreamType = decltype(detect_specialization::Detect<T>(0));
-
-// Build the error message string.  Specify no inlining for code size.
-template <typename T1, typename T2>
-ABSL_ATTRIBUTE_RETURNS_NONNULL std::string* MakeCheckOpString(
-    T1 v1, T2 v2, const char* exprtext) ABSL_ATTRIBUTE_NOINLINE;
-
-template <typename T1, typename T2>
-std::string* MakeCheckOpString(T1 v1, T2 v2, const char* exprtext) {
-  CheckOpMessageBuilder comb(exprtext);
-  MakeCheckOpValueString(comb.ForVar1(), v1);
-  MakeCheckOpValueString(comb.ForVar2(), v2);
-  return comb.NewString();
-}
-
-// Add a few commonly used instantiations as extern to reduce size of objects
-// files.
-#define ABSL_LOG_INTERNAL_DEFINE_MAKE_CHECK_OP_STRING_EXTERN(x) \
-  extern template std::string* MakeCheckOpString(x, x, const char*)
-ABSL_LOG_INTERNAL_DEFINE_MAKE_CHECK_OP_STRING_EXTERN(bool);
-ABSL_LOG_INTERNAL_DEFINE_MAKE_CHECK_OP_STRING_EXTERN(int64_t);
-ABSL_LOG_INTERNAL_DEFINE_MAKE_CHECK_OP_STRING_EXTERN(uint64_t);
-ABSL_LOG_INTERNAL_DEFINE_MAKE_CHECK_OP_STRING_EXTERN(float);
-ABSL_LOG_INTERNAL_DEFINE_MAKE_CHECK_OP_STRING_EXTERN(double);
-ABSL_LOG_INTERNAL_DEFINE_MAKE_CHECK_OP_STRING_EXTERN(char);
-ABSL_LOG_INTERNAL_DEFINE_MAKE_CHECK_OP_STRING_EXTERN(unsigned char);
-ABSL_LOG_INTERNAL_DEFINE_MAKE_CHECK_OP_STRING_EXTERN(const std::string&);
-ABSL_LOG_INTERNAL_DEFINE_MAKE_CHECK_OP_STRING_EXTERN(const absl::string_view&);
-ABSL_LOG_INTERNAL_DEFINE_MAKE_CHECK_OP_STRING_EXTERN(const char*);
-ABSL_LOG_INTERNAL_DEFINE_MAKE_CHECK_OP_STRING_EXTERN(const signed char*);
-ABSL_LOG_INTERNAL_DEFINE_MAKE_CHECK_OP_STRING_EXTERN(const unsigned char*);
-ABSL_LOG_INTERNAL_DEFINE_MAKE_CHECK_OP_STRING_EXTERN(const void*);
-#undef ABSL_LOG_INTERNAL_DEFINE_MAKE_CHECK_OP_STRING_EXTERN
-
-// `ABSL_LOG_INTERNAL_CHECK_OP_IMPL_RESULT` skips formatting the Check_OP result
-// string iff `ABSL_MIN_LOG_LEVEL` exceeds `kFatal`, instead returning an empty
-// string.
-#ifdef ABSL_MIN_LOG_LEVEL
-#define ABSL_LOG_INTERNAL_CHECK_OP_IMPL_RESULT(U1, U2, v1, v2, exprtext) \
-  ((::absl::LogSeverity::kFatal >=                                       \
-    static_cast<::absl::LogSeverity>(ABSL_MIN_LOG_LEVEL))                \
-       ? MakeCheckOpString<U1, U2>(v1, v2, exprtext)                     \
-       : new std::string())
-#else
-#define ABSL_LOG_INTERNAL_CHECK_OP_IMPL_RESULT(U1, U2, v1, v2, exprtext) \
-  MakeCheckOpString<U1, U2>(v1, v2, exprtext)
-#endif
-
-// Helper functions for `ABSL_LOG_INTERNAL_CHECK_OP` macro family.  The
-// `(int, int)` override works around the issue that the compiler will not
-// instantiate the template version of the function on values of unnamed enum
-// type.
-#define ABSL_LOG_INTERNAL_CHECK_OP_IMPL(name, op)                        \
-  template <typename T1, typename T2>                                    \
-  inline constexpr ::std::string* name##Impl(const T1& v1, const T2& v2, \
-                                             const char* exprtext) {     \
-    using U1 = CheckOpStreamType<T1>;                                    \
-    using U2 = CheckOpStreamType<T2>;                                    \
-    return ABSL_PREDICT_TRUE(v1 op v2)                                   \
-               ? nullptr                                                 \
-               : ABSL_LOG_INTERNAL_CHECK_OP_IMPL_RESULT(U1, U2, v1, v2,  \
-                                                        exprtext);       \
-  }                                                                      \
-  inline constexpr ::std::string* name##Impl(int v1, int v2,             \
-                                             const char* exprtext) {     \
-    return name##Impl<int, int>(v1, v2, exprtext);                       \
-  }
-
-ABSL_LOG_INTERNAL_CHECK_OP_IMPL(Check_EQ, ==)
-ABSL_LOG_INTERNAL_CHECK_OP_IMPL(Check_NE, !=)
-ABSL_LOG_INTERNAL_CHECK_OP_IMPL(Check_LE, <=)
-ABSL_LOG_INTERNAL_CHECK_OP_IMPL(Check_LT, <)
-ABSL_LOG_INTERNAL_CHECK_OP_IMPL(Check_GE, >=)
-ABSL_LOG_INTERNAL_CHECK_OP_IMPL(Check_GT, >)
-#undef ABSL_LOG_INTERNAL_CHECK_OP_IMPL_RESULT
-#undef ABSL_LOG_INTERNAL_CHECK_OP_IMPL
-
-std::string* CheckstrcmptrueImpl(const char* s1, const char* s2,
-                                 const char* exprtext);
-std::string* CheckstrcmpfalseImpl(const char* s1, const char* s2,
-                                  const char* exprtext);
-std::string* CheckstrcasecmptrueImpl(const char* s1, const char* s2,
-                                     const char* exprtext);
-std::string* CheckstrcasecmpfalseImpl(const char* s1, const char* s2,
-                                      const char* exprtext);
-
-// `CHECK_EQ` and friends want to pass their arguments by reference, however
-// this winds up exposing lots of cases where people have defined and
-// initialized static const data members but never declared them (i.e. in a .cc
-// file), meaning they are not referenceable.  This function avoids that problem
-// for integers (the most common cases) by overloading for every primitive
-// integer type, even the ones we discourage, and returning them by value.
-template <typename T>
-inline constexpr const T& GetReferenceableValue(const T& t) {
-  return t;
-}
-inline constexpr char GetReferenceableValue(char t) { return t; }
-inline constexpr unsigned char GetReferenceableValue(unsigned char t) {
-  return t;
-}
-inline constexpr signed char GetReferenceableValue(signed char t) { return t; }
-inline constexpr short GetReferenceableValue(short t) { return t; }  // NOLINT
-inline constexpr unsigned short GetReferenceableValue(               // NOLINT
-    unsigned short t) {                                              // NOLINT
-  return t;
-}
-inline constexpr int GetReferenceableValue(int t) { return t; }
-inline constexpr unsigned int GetReferenceableValue(unsigned int t) {
-  return t;
-}
-inline constexpr long GetReferenceableValue(long t) { return t; }  // NOLINT
-inline constexpr unsigned long GetReferenceableValue(              // NOLINT
-    unsigned long t) {                                             // NOLINT
-  return t;
-}
-inline constexpr long long GetReferenceableValue(long long t) {  // NOLINT
-  return t;
-}
-inline constexpr unsigned long long GetReferenceableValue(  // NOLINT
-    unsigned long long t) {                                 // NOLINT
-  return t;
-}
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_CHECK_OP_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/conditions.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/conditions.h
deleted file mode 100644
index 645f3c23..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/conditions.h
+++ /dev/null
@@ -1,239 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/internal/conditions.h
-// -----------------------------------------------------------------------------
-//
-// This file contains implementation of conditional log statements, like LOG_IF
-// including all the ABSL_LOG_INTERNAL_..._CONDITION_... macros and
-// various condition classes like LogEveryNState.
-
-#ifndef ABSL_LOG_INTERNAL_CONDITIONS_H_
-#define ABSL_LOG_INTERNAL_CONDITIONS_H_
-
-#if defined(_WIN32) || defined(__hexagon__)
-#include <cstdlib>
-#else
-#include <unistd.h>
-#endif
-#include <stdlib.h>
-
-#include <atomic>
-#include <cstdint>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/log/internal/voidify.h"
-
-// `ABSL_LOG_INTERNAL_CONDITION` prefixes another macro that expands to a
-// temporary `LogMessage` instantiation followed by zero or more streamed
-// expressions.  This definition is tricky to read correctly.  It evaluates to
-// either
-//
-//   (void)0;
-//
-// or
-//
-//   ::absl::log_internal::Voidify() &&
-//       ::absl::log_internal::LogMessage(...) << "the user's message";
-//
-// If the condition is evaluable at compile time, as is often the case, it
-// compiles away to just one side or the other.
-//
-// Although this is not used anywhere a statement (e.g. `if`) could not go,
-// the ternary expression does a better job avoiding spurious diagnostics
-// (dangling else, missing switch case) and preserving noreturn semantics (e.g.
-// on `LOG(FATAL)`) without requiring braces.
-//
-// The `switch` ensures that this expansion is the beginning of a statement (as
-// opposed to an expression) and prevents shenanigans like
-// `AFunction(LOG(INFO))` and `decltype(LOG(INFO))`.  The apparently-redundant
-// `default` case makes the condition more amenable to Clang dataflow analysis.
-#define ABSL_LOG_INTERNAL_STATELESS_CONDITION(condition) \
-  switch (0)                                             \
-  case 0:                                                \
-  default:                                               \
-    !(condition) ? (void)0 : ::absl::log_internal::Voidify()&&
-
-// `ABSL_LOG_INTERNAL_STATEFUL_CONDITION` applies a condition like
-// `ABSL_LOG_INTERNAL_STATELESS_CONDITION` but adds to that a series of variable
-// declarations, including a local static object which stores the state needed
-// to implement the stateful macros like `LOG_EVERY_N`.
-//
-// `for`-loops are used to declare scoped variables without braces (to permit
-// streaming into the macro's expansion) and without the dangling-`else`
-// problems/diagnostics that come with `if`.
-//
-// Two more variables are declared in separate `for`-loops:
-//
-// * `COUNTER` implements a streamable token whose value when streamed is the
-//   number of times execution has passed through the macro.
-// * A boolean flag is used to prevent any of the `for`-loops from ever actually
-//   looping.
-#define ABSL_LOG_INTERNAL_STATEFUL_CONDITION(condition)             \
-  for (bool absl_log_internal_stateful_condition_do_log(condition); \
-       absl_log_internal_stateful_condition_do_log;                 \
-       absl_log_internal_stateful_condition_do_log = false)         \
-  ABSL_LOG_INTERNAL_STATEFUL_CONDITION_IMPL
-#define ABSL_LOG_INTERNAL_STATEFUL_CONDITION_IMPL(kind, ...)              \
-  for (static ::absl::log_internal::Log##kind##State                      \
-           absl_log_internal_stateful_condition_state;                    \
-       absl_log_internal_stateful_condition_do_log &&                     \
-       absl_log_internal_stateful_condition_state.ShouldLog(__VA_ARGS__); \
-       absl_log_internal_stateful_condition_do_log = false)               \
-    for (const uint32_t COUNTER ABSL_ATTRIBUTE_UNUSED =                   \
-             absl_log_internal_stateful_condition_state.counter();        \
-         absl_log_internal_stateful_condition_do_log;                     \
-         absl_log_internal_stateful_condition_do_log = false)
-
-// `ABSL_LOG_INTERNAL_CONDITION_*` serve to combine any conditions from the
-// macro (e.g. `LOG_IF` or `VLOG`) with inherent conditions (e.g.
-// `ABSL_MIN_LOG_LEVEL`) into a single boolean expression.  We could chain
-// ternary operators instead, however some versions of Clang sometimes issue
-// spurious diagnostics after such expressions due to a control flow analysis
-// bug.
-#ifdef ABSL_MIN_LOG_LEVEL
-#define ABSL_LOG_INTERNAL_CONDITION_INFO(type, condition) \
-  ABSL_LOG_INTERNAL_##type##_CONDITION(                   \
-      (condition) && ::absl::LogSeverity::kInfo >=        \
-                         static_cast<::absl::LogSeverity>(ABSL_MIN_LOG_LEVEL))
-#define ABSL_LOG_INTERNAL_CONDITION_WARNING(type, condition) \
-  ABSL_LOG_INTERNAL_##type##_CONDITION(                      \
-      (condition) && ::absl::LogSeverity::kWarning >=        \
-                         static_cast<::absl::LogSeverity>(ABSL_MIN_LOG_LEVEL))
-#define ABSL_LOG_INTERNAL_CONDITION_ERROR(type, condition) \
-  ABSL_LOG_INTERNAL_##type##_CONDITION(                    \
-      (condition) && ::absl::LogSeverity::kError >=        \
-                         static_cast<::absl::LogSeverity>(ABSL_MIN_LOG_LEVEL))
-// NOTE: Use ternary operators instead of short-circuiting to mitigate
-// https://bugs.llvm.org/show_bug.cgi?id=51928.
-#define ABSL_LOG_INTERNAL_CONDITION_FATAL(type, condition)                 \
-  ABSL_LOG_INTERNAL_##type##_CONDITION(                                    \
-      ((condition)                                                         \
-           ? (::absl::LogSeverity::kFatal >=                               \
-                      static_cast<::absl::LogSeverity>(ABSL_MIN_LOG_LEVEL) \
-                  ? true                                                   \
-                  : (::absl::log_internal::AbortQuietly(), false))         \
-           : false))
-// NOTE: Use ternary operators instead of short-circuiting to mitigate
-// https://bugs.llvm.org/show_bug.cgi?id=51928.
-#define ABSL_LOG_INTERNAL_CONDITION_QFATAL(type, condition)                \
-  ABSL_LOG_INTERNAL_##type##_CONDITION(                                    \
-      ((condition)                                                         \
-           ? (::absl::LogSeverity::kFatal >=                               \
-                      static_cast<::absl::LogSeverity>(ABSL_MIN_LOG_LEVEL) \
-                  ? true                                                   \
-                  : (::absl::log_internal::ExitQuietly(), false))          \
-           : false))
-#define ABSL_LOG_INTERNAL_CONDITION_DFATAL(type, condition)             \
-  ABSL_LOG_INTERNAL_##type##_CONDITION(                                 \
-      (ABSL_ASSUME(absl::kLogDebugFatal == absl::LogSeverity::kError || \
-                   absl::kLogDebugFatal == absl::LogSeverity::kFatal),  \
-       (condition) &&                                                   \
-           (::absl::kLogDebugFatal >=                                   \
-                static_cast<::absl::LogSeverity>(ABSL_MIN_LOG_LEVEL) || \
-            (::absl::kLogDebugFatal == ::absl::LogSeverity::kFatal &&   \
-             (::absl::log_internal::AbortQuietly(), false)))))
-
-#define ABSL_LOG_INTERNAL_CONDITION_LEVEL(severity)                            \
-  for (int absl_log_internal_severity_loop = 1;                                \
-       absl_log_internal_severity_loop; absl_log_internal_severity_loop = 0)   \
-    for (const absl::LogSeverity absl_log_internal_severity =                  \
-             ::absl::NormalizeLogSeverity(severity);                           \
-         absl_log_internal_severity_loop; absl_log_internal_severity_loop = 0) \
-  ABSL_LOG_INTERNAL_CONDITION_LEVEL_IMPL
-#define ABSL_LOG_INTERNAL_CONDITION_LEVEL_IMPL(type, condition)          \
-  ABSL_LOG_INTERNAL_##type##_CONDITION((                                  \
-      (condition) &&                                                     \
-          (absl_log_internal_severity >=                                 \
-               static_cast<::absl::LogSeverity>(ABSL_MIN_LOG_LEVEL) ||   \
-           (absl_log_internal_severity == ::absl::LogSeverity::kFatal && \
-            (::absl::log_internal::AbortQuietly(), false)))))
-#else  // ndef ABSL_MIN_LOG_LEVEL
-#define ABSL_LOG_INTERNAL_CONDITION_INFO(type, condition) \
-  ABSL_LOG_INTERNAL_##type##_CONDITION(condition)
-#define ABSL_LOG_INTERNAL_CONDITION_WARNING(type, condition) \
-  ABSL_LOG_INTERNAL_##type##_CONDITION(condition)
-#define ABSL_LOG_INTERNAL_CONDITION_ERROR(type, condition) \
-  ABSL_LOG_INTERNAL_##type##_CONDITION(condition)
-#define ABSL_LOG_INTERNAL_CONDITION_FATAL(type, condition) \
-  ABSL_LOG_INTERNAL_##type##_CONDITION(condition)
-#define ABSL_LOG_INTERNAL_CONDITION_QFATAL(type, condition) \
-  ABSL_LOG_INTERNAL_##type##_CONDITION(condition)
-#define ABSL_LOG_INTERNAL_CONDITION_DFATAL(type, condition) \
-  ABSL_LOG_INTERNAL_##type##_CONDITION(condition)
-#define ABSL_LOG_INTERNAL_CONDITION_LEVEL(severity)                            \
-  for (int absl_log_internal_severity_loop = 1;                                \
-       absl_log_internal_severity_loop; absl_log_internal_severity_loop = 0)   \
-    for (const absl::LogSeverity absl_log_internal_severity =                  \
-             ::absl::NormalizeLogSeverity(severity);                           \
-         absl_log_internal_severity_loop; absl_log_internal_severity_loop = 0) \
-  ABSL_LOG_INTERNAL_CONDITION_LEVEL_IMPL
-#define ABSL_LOG_INTERNAL_CONDITION_LEVEL_IMPL(type, condition) \
-  ABSL_LOG_INTERNAL_##type##_CONDITION(condition)
-#endif  // ndef ABSL_MIN_LOG_LEVEL
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-
-// Stateful condition class name should be "Log" + name + "State".
-class LogEveryNState final {
- public:
-  bool ShouldLog(int n);
-  uint32_t counter() { return counter_.load(std::memory_order_relaxed); }
-
- private:
-  std::atomic<uint32_t> counter_{0};
-};
-
-class LogFirstNState final {
- public:
-  bool ShouldLog(int n);
-  uint32_t counter() { return counter_.load(std::memory_order_relaxed); }
-
- private:
-  std::atomic<uint32_t> counter_{0};
-};
-
-class LogEveryPow2State final {
- public:
-  bool ShouldLog();
-  uint32_t counter() { return counter_.load(std::memory_order_relaxed); }
-
- private:
-  std::atomic<uint32_t> counter_{0};
-};
-
-class LogEveryNSecState final {
- public:
-  bool ShouldLog(double seconds);
-  uint32_t counter() { return counter_.load(std::memory_order_relaxed); }
-
- private:
-  std::atomic<uint32_t> counter_{0};
-  // Cycle count according to CycleClock that we should next log at.
-  std::atomic<int64_t> next_log_time_cycles_{0};
-};
-
-// Helper routines to abort the application quietly
-
-ABSL_ATTRIBUTE_NORETURN inline void AbortQuietly() { abort(); }
-ABSL_ATTRIBUTE_NORETURN inline void ExitQuietly() { _exit(1); }
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_CONDITIONS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/config.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/config.h
deleted file mode 100644
index 379e9ab9..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/config.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/internal/config.h
-// -----------------------------------------------------------------------------
-//
-
-#ifndef ABSL_LOG_INTERNAL_CONFIG_H_
-#define ABSL_LOG_INTERNAL_CONFIG_H_
-
-#include "absl/base/config.h"
-
-#ifdef _WIN32
-#include <cstdint>
-#else
-#include <sys/types.h>
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-
-#ifdef _WIN32
-using Tid = uint32_t;
-#else
-using Tid = pid_t;
-#endif
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_CONFIG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/flags.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/flags.h
deleted file mode 100644
index c4539785..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/flags.h
+++ /dev/null
@@ -1,59 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/log_flags.h
-// -----------------------------------------------------------------------------
-//
-// This header declares set of flags which can be used to configure Abseil
-// Logging library behaviour at runtime.
-
-#ifndef ABSL_LOG_INTERNAL_FLAGS_H_
-#define ABSL_LOG_INTERNAL_FLAGS_H_
-
-#include <string>
-
-#include "absl/flags/declare.h"
-
-// !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
-// These flags should not be used in C++ code to access logging library
-// configuration knobs. Use interfaces defined in absl/log/globals.h
-// instead. It is still ok to use these flags on a command line.
-// !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
-
-// Log messages at this severity or above are sent to stderr in *addition* to
-// `LogSink`s.  Defaults to `ERROR`.  See log_severity.h for numeric values of
-// severity levels.
-ABSL_DECLARE_FLAG(int, stderrthreshold);
-
-// Log messages at this severity or above are logged; others are discarded.
-// Defaults to `INFO`, i.e. log all severities.  See log_severity.h for numeric
-// values of severity levels.
-ABSL_DECLARE_FLAG(int, minloglevel);
-
-// If specified in the form file:linenum, any messages logged from a matching
-// location will also include a backtrace.
-ABSL_DECLARE_FLAG(std::string, log_backtrace_at);
-
-// If true, the log prefix (severity, date, time, PID, etc.) is prepended to
-// each message logged. Defaults to true.
-ABSL_DECLARE_FLAG(bool, log_prefix);
-
-// Global log verbosity level. Default is 0.
-ABSL_DECLARE_FLAG(int, v);
-
-// Per-module log verbosity level. By default is empty and is unused.
-ABSL_DECLARE_FLAG(std::string, vmodule);
-
-#endif  // ABSL_LOG_INTERNAL_FLAGS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/fnmatch.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/fnmatch.h
deleted file mode 100644
index 4ea147cc..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/fnmatch.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// Copyright 2023 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_LOG_INTERNAL_FNMATCH_H_
-#define ABSL_LOG_INTERNAL_FNMATCH_H_
-
-#include "absl/base/config.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-// Like POSIX `fnmatch`, but:
-// * accepts `string_view`
-// * does not allocate any dynamic memory
-// * only supports * and ? wildcards and not bracket expressions [...]
-// * wildcards may match /
-// * no backslash-escaping
-bool FNMatch(absl::string_view pattern, absl::string_view str);
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_FNMATCH_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/globals.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/globals.h
deleted file mode 100644
index 27bc0d09..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/globals.h
+++ /dev/null
@@ -1,101 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/internal/globals.h
-// -----------------------------------------------------------------------------
-//
-// This header file contains various global objects and static helper routines
-// use in logging implementation.
-
-#ifndef ABSL_LOG_INTERNAL_GLOBALS_H_
-#define ABSL_LOG_INTERNAL_GLOBALS_H_
-
-#include "absl/base/config.h"
-#include "absl/base/log_severity.h"
-#include "absl/strings/string_view.h"
-#include "absl/time/time.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-
-// IsInitialized returns true if the logging library is initialized.
-// This function is async-signal-safe
-bool IsInitialized();
-
-// SetLoggingInitialized is called once after logging initialization is done.
-void SetInitialized();
-
-// Unconditionally write a `message` to stderr. If `severity` exceeds kInfo
-// we also flush the stderr stream.
-void WriteToStderr(absl::string_view message, absl::LogSeverity severity);
-
-// Set the TimeZone used for human-friendly times (for example, the log message
-// prefix) printed by the logging library. This may only be called once.
-void SetTimeZone(absl::TimeZone tz);
-
-// Returns the TimeZone used for human-friendly times (for example, the log
-// message prefix) printed by the logging library Returns nullptr prior to
-// initialization.
-const absl::TimeZone* TimeZone();
-
-// Returns true if stack traces emitted by the logging library should be
-// symbolized. This function is async-signal-safe.
-bool ShouldSymbolizeLogStackTrace();
-
-// Enables or disables symbolization of stack traces emitted by the
-// logging library. This function is async-signal-safe.
-void EnableSymbolizeLogStackTrace(bool on_off);
-
-// Returns the maximum number of frames that appear in stack traces
-// emitted by the logging library. This function is async-signal-safe.
-int MaxFramesInLogStackTrace();
-
-// Sets the maximum number of frames that appear in stack traces emitted by
-// the logging library. This function is async-signal-safe.
-void SetMaxFramesInLogStackTrace(int max_num_frames);
-
-// Determines whether we exit the program for a LOG(DFATAL) message in
-// debug mode.  It does this by skipping the call to Fail/FailQuietly.
-// This is intended for testing only.
-//
-// This can have some effects on LOG(FATAL) as well. Failure messages
-// are always allocated (rather than sharing a buffer), the crash
-// reason is not recorded, the "gwq" status message is not updated,
-// and the stack trace is not recorded.  The LOG(FATAL) *will* still
-// exit the program. Since this function is used only in testing,
-// these differences are acceptable.
-//
-// Additionally, LOG(LEVEL(FATAL)) is indistinguishable from LOG(DFATAL) and
-// will not terminate the program if SetExitOnDFatal(false) has been called.
-bool ExitOnDFatal();
-
-// SetExitOnDFatal() sets the ExitOnDFatal() status
-void SetExitOnDFatal(bool on_off);
-
-// Determines if the logging library should suppress logging of stacktraces in
-// the `SIGABRT` handler, typically because we just logged a stacktrace as part
-// of `LOG(FATAL)` and are about to send ourselves a `SIGABRT` to end the
-// program.
-bool SuppressSigabortTrace();
-
-// Sets the SuppressSigabortTrace() status and returns the previous state.
-bool SetSuppressSigabortTrace(bool on_off);
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_GLOBALS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/log_format.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/log_format.h
deleted file mode 100644
index 95a45edf..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/log_format.h
+++ /dev/null
@@ -1,78 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/internal/log_format.h
-// -----------------------------------------------------------------------------
-//
-// This file declares routines implementing formatting of log message and log
-// prefix.
-
-#ifndef ABSL_LOG_INTERNAL_LOG_FORMAT_H_
-#define ABSL_LOG_INTERNAL_LOG_FORMAT_H_
-
-#include <stddef.h>
-
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/base/log_severity.h"
-#include "absl/log/internal/config.h"
-#include "absl/strings/string_view.h"
-#include "absl/time/civil_time.h"
-#include "absl/time/time.h"
-#include "absl/types/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-
-enum class PrefixFormat {
-  kNotRaw,
-  kRaw,
-};
-
-// Formats log message based on provided data.
-std::string FormatLogMessage(absl::LogSeverity severity,
-                             absl::CivilSecond civil_second,
-                             absl::Duration subsecond, log_internal::Tid tid,
-                             absl::string_view basename, int line,
-                             PrefixFormat format, absl::string_view message);
-
-// Formats various entry metadata into a text string meant for use as a
-// prefix on a log message string.  Writes into `buf`, advances `buf` to point
-// at the remainder of the buffer (i.e. past any written bytes), and returns the
-// number of bytes written.
-//
-// In addition to calling `buf->remove_prefix()` (or the equivalent), this
-// function may also do `buf->remove_suffix(buf->size())` in cases where no more
-// bytes (i.e. no message data) should be written into the buffer.  For example,
-// if the prefix ought to be:
-//   I0926 09:00:00.000000 1234567 foo.cc:123]
-// `buf` is too small, the function might fill the whole buffer:
-//   I0926 09:00:00.000000 1234
-// (note the apparrently incorrect thread ID), or it might write less:
-//   I0926 09:00:00.000000
-// In this case, it might also empty `buf` prior to returning to prevent
-// message data from being written into the space where a reader would expect to
-// see a thread ID.
-size_t FormatLogPrefix(absl::LogSeverity severity, absl::Time timestamp,
-                       log_internal::Tid tid, absl::string_view basename,
-                       int line, PrefixFormat format, absl::Span<char>& buf);
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_LOG_FORMAT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/log_impl.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/log_impl.h
deleted file mode 100644
index 99de6dbb..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/log_impl.h
+++ /dev/null
@@ -1,282 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_LOG_INTERNAL_LOG_IMPL_H_
-#define ABSL_LOG_INTERNAL_LOG_IMPL_H_
-
-#include "absl/log/absl_vlog_is_on.h"
-#include "absl/log/internal/conditions.h"
-#include "absl/log/internal/log_message.h"
-#include "absl/log/internal/strip.h"
-
-// ABSL_LOG()
-#define ABSL_LOG_INTERNAL_LOG_IMPL(severity)             \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATELESS, true) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-// ABSL_PLOG()
-#define ABSL_LOG_INTERNAL_PLOG_IMPL(severity)              \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATELESS, true)   \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream() \
-          .WithPerror()
-
-// ABSL_DLOG()
-#ifndef NDEBUG
-#define ABSL_LOG_INTERNAL_DLOG_IMPL(severity)            \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATELESS, true) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-#else
-#define ABSL_LOG_INTERNAL_DLOG_IMPL(severity)             \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATELESS, false) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-#endif
-
-// The `switch` ensures that this expansion is the begnning of a statement (as
-// opposed to an expression). The use of both `case 0` and `default` is to
-// suppress a compiler warning.
-#define ABSL_LOG_INTERNAL_VLOG_IMPL(verbose_level)                         \
-  switch (const int absl_logging_internal_verbose_level = (verbose_level)) \
-  case 0:                                                                  \
-  default:                                                                 \
-    ABSL_LOG_INTERNAL_LOG_IF_IMPL(                                         \
-        _INFO, ABSL_VLOG_IS_ON(absl_logging_internal_verbose_level))       \
-        .WithVerbosity(absl_logging_internal_verbose_level)
-
-#ifndef NDEBUG
-#define ABSL_LOG_INTERNAL_DVLOG_IMPL(verbose_level)                        \
-  switch (const int absl_logging_internal_verbose_level = (verbose_level)) \
-  case 0:                                                                  \
-  default:                                                                 \
-    ABSL_LOG_INTERNAL_LOG_IF_IMPL(                                         \
-        _INFO, ABSL_VLOG_IS_ON(absl_logging_internal_verbose_level))       \
-        .WithVerbosity(absl_logging_internal_verbose_level)
-#else
-#define ABSL_LOG_INTERNAL_DVLOG_IMPL(verbose_level)                           \
-  switch (const int absl_logging_internal_verbose_level = (verbose_level))    \
-  case 0:                                                                     \
-  default:                                                                    \
-    ABSL_LOG_INTERNAL_LOG_IF_IMPL(                                            \
-        _INFO, false && ABSL_VLOG_IS_ON(absl_logging_internal_verbose_level)) \
-        .WithVerbosity(absl_logging_internal_verbose_level)
-#endif
-
-#define ABSL_LOG_INTERNAL_LOG_IF_IMPL(severity, condition)    \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATELESS, condition) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-#define ABSL_LOG_INTERNAL_PLOG_IF_IMPL(severity, condition)   \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATELESS, condition) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()    \
-          .WithPerror()
-
-#ifndef NDEBUG
-#define ABSL_LOG_INTERNAL_DLOG_IF_IMPL(severity, condition)   \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATELESS, condition) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-#else
-#define ABSL_LOG_INTERNAL_DLOG_IF_IMPL(severity, condition)              \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATELESS, false && (condition)) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-#endif
-
-// ABSL_LOG_EVERY_N
-#define ABSL_LOG_INTERNAL_LOG_EVERY_N_IMPL(severity, n)            \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, true)(EveryN, n) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-// ABSL_LOG_FIRST_N
-#define ABSL_LOG_INTERNAL_LOG_FIRST_N_IMPL(severity, n)            \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, true)(FirstN, n) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-// ABSL_LOG_EVERY_POW_2
-#define ABSL_LOG_INTERNAL_LOG_EVERY_POW_2_IMPL(severity)           \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, true)(EveryPow2) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-// ABSL_LOG_EVERY_N_SEC
-#define ABSL_LOG_INTERNAL_LOG_EVERY_N_SEC_IMPL(severity, n_seconds)           \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, true)(EveryNSec, n_seconds) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_PLOG_EVERY_N_IMPL(severity, n)           \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, true)(EveryN, n) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()         \
-          .WithPerror()
-
-#define ABSL_LOG_INTERNAL_PLOG_FIRST_N_IMPL(severity, n)           \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, true)(FirstN, n) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()         \
-          .WithPerror()
-
-#define ABSL_LOG_INTERNAL_PLOG_EVERY_POW_2_IMPL(severity)          \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, true)(EveryPow2) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()         \
-          .WithPerror()
-
-#define ABSL_LOG_INTERNAL_PLOG_EVERY_N_SEC_IMPL(severity, n_seconds)          \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, true)(EveryNSec, n_seconds) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()                    \
-          .WithPerror()
-
-#ifndef NDEBUG
-#define ABSL_LOG_INTERNAL_DLOG_EVERY_N_IMPL(severity, n) \
-  ABSL_LOG_INTERNAL_CONDITION_INFO(STATEFUL, true)       \
-  (EveryN, n) ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_DLOG_FIRST_N_IMPL(severity, n) \
-  ABSL_LOG_INTERNAL_CONDITION_INFO(STATEFUL, true)       \
-  (FirstN, n) ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_DLOG_EVERY_POW_2_IMPL(severity) \
-  ABSL_LOG_INTERNAL_CONDITION_INFO(STATEFUL, true)        \
-  (EveryPow2) ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_DLOG_EVERY_N_SEC_IMPL(severity, n_seconds) \
-  ABSL_LOG_INTERNAL_CONDITION_INFO(STATEFUL, true)                   \
-  (EveryNSec, n_seconds) ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#else  // def NDEBUG
-#define ABSL_LOG_INTERNAL_DLOG_EVERY_N_IMPL(severity, n) \
-  ABSL_LOG_INTERNAL_CONDITION_INFO(STATEFUL, false)      \
-  (EveryN, n) ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_DLOG_FIRST_N_IMPL(severity, n) \
-  ABSL_LOG_INTERNAL_CONDITION_INFO(STATEFUL, false)      \
-  (FirstN, n) ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_DLOG_EVERY_POW_2_IMPL(severity) \
-  ABSL_LOG_INTERNAL_CONDITION_INFO(STATEFUL, false)       \
-  (EveryPow2) ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_DLOG_EVERY_N_SEC_IMPL(severity, n_seconds) \
-  ABSL_LOG_INTERNAL_CONDITION_INFO(STATEFUL, false)                  \
-  (EveryNSec, n_seconds) ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-#endif  // def NDEBUG
-
-#define ABSL_LOG_INTERNAL_VLOG_EVERY_N_IMPL(verbose_level, n)                \
-  switch (const int absl_logging_internal_verbose_level = (verbose_level))   \
-  case 0:                                                                    \
-  default:                                                                   \
-    ABSL_LOG_INTERNAL_CONDITION_INFO(                                        \
-        STATEFUL, ABSL_VLOG_IS_ON(absl_logging_internal_verbose_level))      \
-  (EveryN, n) ABSL_LOGGING_INTERNAL_LOG_INFO.InternalStream().WithVerbosity( \
-      absl_logging_internal_verbose_level)
-
-#define ABSL_LOG_INTERNAL_VLOG_FIRST_N_IMPL(verbose_level, n)                \
-  switch (const int absl_logging_internal_verbose_level = (verbose_level))   \
-  case 0:                                                                    \
-  default:                                                                   \
-    ABSL_LOG_INTERNAL_CONDITION_INFO(                                        \
-        STATEFUL, ABSL_VLOG_IS_ON(absl_logging_internal_verbose_level))      \
-  (FirstN, n) ABSL_LOGGING_INTERNAL_LOG_INFO.InternalStream().WithVerbosity( \
-      absl_logging_internal_verbose_level)
-
-#define ABSL_LOG_INTERNAL_VLOG_EVERY_POW_2_IMPL(verbose_level)               \
-  switch (const int absl_logging_internal_verbose_level = (verbose_level))   \
-  case 0:                                                                    \
-  default:                                                                   \
-    ABSL_LOG_INTERNAL_CONDITION_INFO(                                        \
-        STATEFUL, ABSL_VLOG_IS_ON(absl_logging_internal_verbose_level))      \
-  (EveryPow2) ABSL_LOGGING_INTERNAL_LOG_INFO.InternalStream().WithVerbosity( \
-      absl_logging_internal_verbose_level)
-
-#define ABSL_LOG_INTERNAL_VLOG_EVERY_N_SEC_IMPL(verbose_level, n_seconds)  \
-  switch (const int absl_logging_internal_verbose_level = (verbose_level)) \
-  case 0:                                                                  \
-  default:                                                                 \
-    ABSL_LOG_INTERNAL_CONDITION_INFO(                                      \
-        STATEFUL, ABSL_VLOG_IS_ON(absl_logging_internal_verbose_level))    \
-  (EveryNSec, n_seconds) ABSL_LOGGING_INTERNAL_LOG_INFO.InternalStream()   \
-      .WithVerbosity(absl_logging_internal_verbose_level)
-
-#define ABSL_LOG_INTERNAL_LOG_IF_EVERY_N_IMPL(severity, condition, n)   \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, condition)(EveryN, n) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_LOG_IF_FIRST_N_IMPL(severity, condition, n)   \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, condition)(FirstN, n) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_LOG_IF_EVERY_POW_2_IMPL(severity, condition)  \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, condition)(EveryPow2) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_LOG_IF_EVERY_N_SEC_IMPL(severity, condition,  \
-                                                  n_seconds)            \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, condition)(EveryNSec, \
-                                                             n_seconds) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_PLOG_IF_EVERY_N_IMPL(severity, condition, n)  \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, condition)(EveryN, n) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()              \
-          .WithPerror()
-
-#define ABSL_LOG_INTERNAL_PLOG_IF_FIRST_N_IMPL(severity, condition, n)  \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, condition)(FirstN, n) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()              \
-          .WithPerror()
-
-#define ABSL_LOG_INTERNAL_PLOG_IF_EVERY_POW_2_IMPL(severity, condition) \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, condition)(EveryPow2) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()              \
-          .WithPerror()
-
-#define ABSL_LOG_INTERNAL_PLOG_IF_EVERY_N_SEC_IMPL(severity, condition, \
-                                                   n_seconds)           \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, condition)(EveryNSec, \
-                                                             n_seconds) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()              \
-          .WithPerror()
-
-#ifndef NDEBUG
-#define ABSL_LOG_INTERNAL_DLOG_IF_EVERY_N_IMPL(severity, condition, n)  \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, condition)(EveryN, n) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_DLOG_IF_FIRST_N_IMPL(severity, condition, n)  \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, condition)(FirstN, n) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_DLOG_IF_EVERY_POW_2_IMPL(severity, condition) \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, condition)(EveryPow2) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_DLOG_IF_EVERY_N_SEC_IMPL(severity, condition, \
-                                                   n_seconds)           \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, condition)(EveryNSec, \
-                                                             n_seconds) \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#else  // def NDEBUG
-#define ABSL_LOG_INTERNAL_DLOG_IF_EVERY_N_IMPL(severity, condition, n)   \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, false && (condition))( \
-      EveryN, n) ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_DLOG_IF_FIRST_N_IMPL(severity, condition, n)   \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, false && (condition))( \
-      FirstN, n) ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_DLOG_IF_EVERY_POW_2_IMPL(severity, condition)  \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, false && (condition))( \
-      EveryPow2) ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-
-#define ABSL_LOG_INTERNAL_DLOG_IF_EVERY_N_SEC_IMPL(severity, condition,  \
-                                                   n_seconds)            \
-  ABSL_LOG_INTERNAL_CONDITION##severity(STATEFUL, false && (condition))( \
-      EveryNSec, n_seconds)                                              \
-      ABSL_LOGGING_INTERNAL_LOG##severity.InternalStream()
-#endif  // def NDEBUG
-
-#endif  // ABSL_LOG_INTERNAL_LOG_IMPL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/log_message.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/log_message.h
deleted file mode 100644
index 4ecb8a14..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/log_message.h
+++ /dev/null
@@ -1,375 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/internal/log_message.h
-// -----------------------------------------------------------------------------
-//
-// This file declares `class absl::log_internal::LogMessage`. This class more or
-// less represents a particular log message. LOG/CHECK macros create a
-// temporary instance of `LogMessage` and then stream values to it.  At the end
-// of the LOG/CHECK statement, LogMessage instance goes out of scope and
-// `~LogMessage` directs the message to the registered log sinks.
-// Heap-allocation of `LogMessage` is unsupported.  Construction outside of a
-// `LOG` macro is unsupported.
-
-#ifndef ABSL_LOG_INTERNAL_LOG_MESSAGE_H_
-#define ABSL_LOG_INTERNAL_LOG_MESSAGE_H_
-
-#include <ios>
-#include <memory>
-#include <ostream>
-#include <streambuf>
-#include <string>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/internal/errno_saver.h"
-#include "absl/base/log_severity.h"
-#include "absl/log/internal/nullguard.h"
-#include "absl/log/log_entry.h"
-#include "absl/log/log_sink.h"
-#include "absl/strings/has_absl_stringify.h"
-#include "absl/strings/string_view.h"
-#include "absl/time/time.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-constexpr int kLogMessageBufferSize = 15000;
-
-class LogMessage {
- public:
-  struct InfoTag {};
-  struct WarningTag {};
-  struct ErrorTag {};
-
-  // Used for `LOG`.
-  LogMessage(const char* file, int line,
-             absl::LogSeverity severity) ABSL_ATTRIBUTE_COLD;
-  // These constructors are slightly smaller/faster to call; the severity is
-  // curried into the function pointer.
-  LogMessage(const char* file, int line,
-             InfoTag) ABSL_ATTRIBUTE_COLD ABSL_ATTRIBUTE_NOINLINE;
-  LogMessage(const char* file, int line,
-             WarningTag) ABSL_ATTRIBUTE_COLD ABSL_ATTRIBUTE_NOINLINE;
-  LogMessage(const char* file, int line,
-             ErrorTag) ABSL_ATTRIBUTE_COLD ABSL_ATTRIBUTE_NOINLINE;
-  LogMessage(const LogMessage&) = delete;
-  LogMessage& operator=(const LogMessage&) = delete;
-  ~LogMessage() ABSL_ATTRIBUTE_COLD;
-
-  // Overrides the location inferred from the callsite.  The string pointed to
-  // by `file` must be valid until the end of the statement.
-  LogMessage& AtLocation(absl::string_view file, int line);
-  // Omits the prefix from this line.  The prefix includes metadata about the
-  // logged data such as source code location and timestamp.
-  LogMessage& NoPrefix();
-  // Sets the verbosity field of the logged message as if it was logged by
-  // `VLOG(verbose_level)`.  Unlike `VLOG`, this method does not affect
-  // evaluation of the statement when the specified `verbose_level` has been
-  // disabled.  The only effect is on `absl::LogSink` implementations which
-  // make use of the `absl::LogSink::verbosity()` value.  The value
-  // `absl::LogEntry::kNoVerbosityLevel` can be specified to mark the message
-  // not verbose.
-  LogMessage& WithVerbosity(int verbose_level);
-  // Uses the specified timestamp instead of one collected in the constructor.
-  LogMessage& WithTimestamp(absl::Time timestamp);
-  // Uses the specified thread ID instead of one collected in the constructor.
-  LogMessage& WithThreadID(absl::LogEntry::tid_t tid);
-  // Copies all metadata (but no data) from the specified `absl::LogEntry`.
-  LogMessage& WithMetadataFrom(const absl::LogEntry& entry);
-  // Appends to the logged message a colon, a space, a textual description of
-  // the current value of `errno` (as by strerror(3)), and the numerical value
-  // of `errno`.
-  LogMessage& WithPerror();
-  // Sends this message to `*sink` in addition to whatever other sinks it would
-  // otherwise have been sent to.  `sink` must not be null.
-  LogMessage& ToSinkAlso(absl::LogSink* sink);
-  // Sends this message to `*sink` and no others.  `sink` must not be null.
-  LogMessage& ToSinkOnly(absl::LogSink* sink);
-
-  // Don't call this method from outside this library.
-  LogMessage& InternalStream() { return *this; }
-
-  // By-value overloads for small, common types let us overlook common failures
-  // to define globals and static data members (i.e. in a .cc file).
-  // clang-format off
-  // The CUDA toolchain cannot handle these <<<'s:
-  LogMessage& operator<<(char v) { return operator<< <char>(v); }
-  LogMessage& operator<<(signed char v) { return operator<< <signed char>(v); }
-  LogMessage& operator<<(unsigned char v) {
-    return operator<< <unsigned char>(v);
-  }
-  LogMessage& operator<<(signed short v) {  // NOLINT
-    return operator<< <signed short>(v);  // NOLINT
-  }
-  LogMessage& operator<<(signed int v) { return operator<< <signed int>(v); }
-  LogMessage& operator<<(signed long v) {  // NOLINT
-    return operator<< <signed long>(v);  // NOLINT
-  }
-  LogMessage& operator<<(signed long long v) {  // NOLINT
-    return operator<< <signed long long>(v);  // NOLINT
-  }
-  LogMessage& operator<<(unsigned short v) {  // NOLINT
-    return operator<< <unsigned short>(v);  // NOLINT
-  }
-  LogMessage& operator<<(unsigned int v) {
-    return operator<< <unsigned int>(v);
-  }
-  LogMessage& operator<<(unsigned long v) {  // NOLINT
-    return operator<< <unsigned long>(v);  // NOLINT
-  }
-  LogMessage& operator<<(unsigned long long v) {  // NOLINT
-    return operator<< <unsigned long long>(v);  // NOLINT
-  }
-  LogMessage& operator<<(void* v) { return operator<< <void*>(v); }
-  LogMessage& operator<<(const void* v) { return operator<< <const void*>(v); }
-  LogMessage& operator<<(float v) { return operator<< <float>(v); }
-  LogMessage& operator<<(double v) { return operator<< <double>(v); }
-  LogMessage& operator<<(bool v) { return operator<< <bool>(v); }
-  // clang-format on
-
-  // These overloads are more efficient since no `ostream` is involved.
-  LogMessage& operator<<(const std::string& v);
-  LogMessage& operator<<(absl::string_view v);
-
-  // Handle stream manipulators e.g. std::endl.
-  LogMessage& operator<<(std::ostream& (*m)(std::ostream& os));
-  LogMessage& operator<<(std::ios_base& (*m)(std::ios_base& os));
-
-  // Literal strings.  This allows us to record C string literals as literals in
-  // the logging.proto.Value.
-  //
-  // Allow this overload to be inlined to prevent generating instantiations of
-  // this template for every value of `SIZE` encountered in each source code
-  // file. That significantly increases linker input sizes. Inlining is cheap
-  // because the argument to this overload is almost always a string literal so
-  // the call to `strlen` can be replaced at compile time. The overload for
-  // `char[]` below should not be inlined. The compiler typically does not have
-  // the string at compile time and cannot replace the call to `strlen` so
-  // inlining it increases the binary size. See the discussion on
-  // cl/107527369.
-  template <int SIZE>
-  LogMessage& operator<<(const char (&buf)[SIZE]);
-
-  // This prevents non-const `char[]` arrays from looking like literals.
-  template <int SIZE>
-  LogMessage& operator<<(char (&buf)[SIZE]) ABSL_ATTRIBUTE_NOINLINE;
-
-  // Types that support `AbslStringify()` are serialized that way.
-  template <typename T,
-            typename std::enable_if<absl::HasAbslStringify<T>::value,
-                                    int>::type = 0>
-  LogMessage& operator<<(const T& v) ABSL_ATTRIBUTE_NOINLINE;
-
-  // Types that don't support `AbslStringify()` but do support streaming into a
-  // `std::ostream&` are serialized that way.
-  template <typename T,
-            typename std::enable_if<!absl::HasAbslStringify<T>::value,
-                                    int>::type = 0>
-  LogMessage& operator<<(const T& v) ABSL_ATTRIBUTE_NOINLINE;
-
-  // Note: We explicitly do not support `operator<<` for non-const references
-  // because it breaks logging of non-integer bitfield types (i.e., enums).
-
- protected:
-  // Call `abort()` or similar to perform `LOG(FATAL)` crash.  It is assumed
-  // that the caller has already generated and written the trace as appropriate.
-  ABSL_ATTRIBUTE_NORETURN static void FailWithoutStackTrace();
-
-  // Similar to `FailWithoutStackTrace()`, but without `abort()`.  Terminates
-  // the process with an error exit code.
-  ABSL_ATTRIBUTE_NORETURN static void FailQuietly();
-
-  // Dispatches the completed `absl::LogEntry` to applicable `absl::LogSink`s.
-  // This might as well be inlined into `~LogMessage` except that
-  // `~LogMessageFatal` needs to call it early.
-  void Flush();
-
-  // After this is called, failures are done as quiet as possible for this log
-  // message.
-  void SetFailQuietly();
-
- private:
-  struct LogMessageData;  // Opaque type containing message state
-  friend class AsLiteralImpl;
-  friend class StringifySink;
-
-  // This streambuf writes directly into the structured logging buffer so that
-  // arbitrary types can be encoded as string data (using
-  // `operator<<(std::ostream &, ...)` without any extra allocation or copying.
-  // Space is reserved before the data to store the length field, which is
-  // filled in by `~OstreamView`.
-  class OstreamView final : public std::streambuf {
-   public:
-    explicit OstreamView(LogMessageData& message_data);
-    ~OstreamView() override;
-    OstreamView(const OstreamView&) = delete;
-    OstreamView& operator=(const OstreamView&) = delete;
-    std::ostream& stream();
-
-   private:
-    LogMessageData& data_;
-    absl::Span<char> encoded_remaining_copy_;
-    absl::Span<char> message_start_;
-    absl::Span<char> string_start_;
-  };
-
-  enum class StringType {
-    kLiteral,
-    kNotLiteral,
-  };
-  template <StringType str_type>
-  void CopyToEncodedBuffer(absl::string_view str) ABSL_ATTRIBUTE_NOINLINE;
-  template <StringType str_type>
-  void CopyToEncodedBuffer(char ch, size_t num) ABSL_ATTRIBUTE_NOINLINE;
-
-  // Returns `true` if the message is fatal or enabled debug-fatal.
-  bool IsFatal() const;
-
-  // Records some tombstone-type data in anticipation of `Die`.
-  void PrepareToDie();
-  void Die();
-
-  void SendToLog();
-
-  // Checks `FLAGS_log_backtrace_at` and appends a backtrace if appropriate.
-  void LogBacktraceIfNeeded();
-
-  // This should be the first data member so that its initializer captures errno
-  // before any other initializers alter it (e.g. with calls to new) and so that
-  // no other destructors run afterward an alter it (e.g. with calls to delete).
-  absl::base_internal::ErrnoSaver errno_saver_;
-
-  // We keep the data in a separate struct so that each instance of `LogMessage`
-  // uses less stack space.
-  std::unique_ptr<LogMessageData> data_;
-};
-
-// Helper class so that `AbslStringify()` can modify the LogMessage.
-class StringifySink final {
- public:
-  explicit StringifySink(LogMessage& message) : message_(message) {}
-
-  void Append(size_t count, char ch) {
-    message_.CopyToEncodedBuffer<LogMessage::StringType::kNotLiteral>(ch,
-                                                                      count);
-  }
-
-  void Append(absl::string_view v) {
-    message_.CopyToEncodedBuffer<LogMessage::StringType::kNotLiteral>(v);
-  }
-
-  // For types that implement `AbslStringify` using `absl::Format()`.
-  friend void AbslFormatFlush(StringifySink* sink, absl::string_view v) {
-    sink->Append(v);
-  }
-
- private:
-  LogMessage& message_;
-};
-
-// Note: the following is declared `ABSL_ATTRIBUTE_NOINLINE`
-template <typename T,
-          typename std::enable_if<absl::HasAbslStringify<T>::value, int>::type>
-LogMessage& LogMessage::operator<<(const T& v) {
-  StringifySink sink(*this);
-  // Replace with public API.
-  AbslStringify(sink, v);
-  return *this;
-}
-
-// Note: the following is declared `ABSL_ATTRIBUTE_NOINLINE`
-template <typename T,
-          typename std::enable_if<!absl::HasAbslStringify<T>::value, int>::type>
-LogMessage& LogMessage::operator<<(const T& v) {
-  OstreamView view(*data_);
-  view.stream() << log_internal::NullGuard<T>().Guard(v);
-  return *this;
-}
-
-template <int SIZE>
-LogMessage& LogMessage::operator<<(const char (&buf)[SIZE]) {
-  CopyToEncodedBuffer<StringType::kLiteral>(buf);
-  return *this;
-}
-
-// Note: the following is declared `ABSL_ATTRIBUTE_NOINLINE`
-template <int SIZE>
-LogMessage& LogMessage::operator<<(char (&buf)[SIZE]) {
-  CopyToEncodedBuffer<StringType::kNotLiteral>(buf);
-  return *this;
-}
-// We instantiate these specializations in the library's TU to save space in
-// other TUs.  Since the template is marked `ABSL_ATTRIBUTE_NOINLINE` we will be
-// emitting a function call either way.
-extern template LogMessage& LogMessage::operator<<(const char& v);
-extern template LogMessage& LogMessage::operator<<(const signed char& v);
-extern template LogMessage& LogMessage::operator<<(const unsigned char& v);
-extern template LogMessage& LogMessage::operator<<(const short& v);  // NOLINT
-extern template LogMessage& LogMessage::operator<<(
-    const unsigned short& v);  // NOLINT
-extern template LogMessage& LogMessage::operator<<(const int& v);
-extern template LogMessage& LogMessage::operator<<(
-    const unsigned int& v);                                         // NOLINT
-extern template LogMessage& LogMessage::operator<<(const long& v);  // NOLINT
-extern template LogMessage& LogMessage::operator<<(
-    const unsigned long& v);  // NOLINT
-extern template LogMessage& LogMessage::operator<<(
-    const long long& v);  // NOLINT
-extern template LogMessage& LogMessage::operator<<(
-    const unsigned long long& v);  // NOLINT
-extern template LogMessage& LogMessage::operator<<(void* const& v);
-extern template LogMessage& LogMessage::operator<<(const void* const& v);
-extern template LogMessage& LogMessage::operator<<(const float& v);
-extern template LogMessage& LogMessage::operator<<(const double& v);
-extern template LogMessage& LogMessage::operator<<(const bool& v);
-
-extern template void LogMessage::CopyToEncodedBuffer<
-    LogMessage::StringType::kLiteral>(absl::string_view str);
-extern template void LogMessage::CopyToEncodedBuffer<
-    LogMessage::StringType::kNotLiteral>(absl::string_view str);
-extern template void
-LogMessage::CopyToEncodedBuffer<LogMessage::StringType::kLiteral>(char ch,
-                                                                  size_t num);
-extern template void LogMessage::CopyToEncodedBuffer<
-    LogMessage::StringType::kNotLiteral>(char ch, size_t num);
-
-// `LogMessageFatal` ensures the process will exit in failure after logging this
-// message.
-class LogMessageFatal final : public LogMessage {
- public:
-  LogMessageFatal(const char* file, int line) ABSL_ATTRIBUTE_COLD;
-  LogMessageFatal(const char* file, int line,
-                  absl::string_view failure_msg) ABSL_ATTRIBUTE_COLD;
-  ABSL_ATTRIBUTE_NORETURN ~LogMessageFatal();
-};
-
-class LogMessageQuietlyFatal final : public LogMessage {
- public:
-  LogMessageQuietlyFatal(const char* file, int line) ABSL_ATTRIBUTE_COLD;
-  LogMessageQuietlyFatal(const char* file, int line,
-                         absl::string_view failure_msg) ABSL_ATTRIBUTE_COLD;
-  ABSL_ATTRIBUTE_NORETURN ~LogMessageQuietlyFatal();
-};
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-extern "C" ABSL_ATTRIBUTE_WEAK void ABSL_INTERNAL_C_SYMBOL(
-    AbslInternalOnFatalLogMessage)(const absl::LogEntry&);
-
-#endif  // ABSL_LOG_INTERNAL_LOG_MESSAGE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/log_sink_set.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/log_sink_set.h
deleted file mode 100644
index 88ab073b..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/log_sink_set.h
+++ /dev/null
@@ -1,54 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/internal/log_sink_set.h
-// -----------------------------------------------------------------------------
-
-#ifndef ABSL_LOG_INTERNAL_LOG_SINK_SET_H_
-#define ABSL_LOG_INTERNAL_LOG_SINK_SET_H_
-
-#include "absl/base/config.h"
-#include "absl/log/log_entry.h"
-#include "absl/log/log_sink.h"
-#include "absl/types/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-
-// Returns true if a globally-registered `LogSink`'s `Send()` is currently
-// being invoked on this thread.
-bool ThreadIsLoggingToLogSink();
-
-// This function may log to two sets of sinks:
-//
-// * If `extra_sinks_only` is true, it will dispatch only to `extra_sinks`.
-//   `LogMessage::ToSinkAlso` and `LogMessage::ToSinkOnly` are used to attach
-//    extra sinks to the entry.
-// * Otherwise it will also log to the global sinks set. This set is managed
-//   by `absl::AddLogSink` and `absl::RemoveLogSink`.
-void LogToSinks(const absl::LogEntry& entry,
-                absl::Span<absl::LogSink*> extra_sinks, bool extra_sinks_only);
-
-// Implementation for operations with log sink set.
-void AddLogSink(absl::LogSink* sink);
-void RemoveLogSink(absl::LogSink* sink);
-void FlushLogSinks();
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_LOG_SINK_SET_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/nullguard.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/nullguard.h
deleted file mode 100644
index 623943c5..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/nullguard.h
+++ /dev/null
@@ -1,88 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/internal/nullguard.h
-// -----------------------------------------------------------------------------
-//
-// NullGuard exists such that NullGuard<T>::Guard(v) returns v, unless passed a
-// nullptr_t, or a null char* or const char*, in which case it returns "(null)".
-// This allows streaming NullGuard<T>::Guard(v) to an output stream without
-// hitting undefined behavior for null values.
-
-#ifndef ABSL_LOG_INTERNAL_NULLGUARD_H_
-#define ABSL_LOG_INTERNAL_NULLGUARD_H_
-
-#include <array>
-#include <cstddef>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-
-ABSL_CONST_INIT ABSL_DLL extern const std::array<char, 7> kCharNull;
-ABSL_CONST_INIT ABSL_DLL extern const std::array<signed char, 7>
-    kSignedCharNull;
-ABSL_CONST_INIT ABSL_DLL extern const std::array<unsigned char, 7>
-    kUnsignedCharNull;
-
-template <typename T>
-struct NullGuard final {
-  static const T& Guard(const T& v) { return v; }
-};
-template <>
-struct NullGuard<char*> final {
-  static const char* Guard(const char* v) { return v ? v : kCharNull.data(); }
-};
-template <>
-struct NullGuard<const char*> final {
-  static const char* Guard(const char* v) { return v ? v : kCharNull.data(); }
-};
-template <>
-struct NullGuard<signed char*> final {
-  static const signed char* Guard(const signed char* v) {
-    return v ? v : kSignedCharNull.data();
-  }
-};
-template <>
-struct NullGuard<const signed char*> final {
-  static const signed char* Guard(const signed char* v) {
-    return v ? v : kSignedCharNull.data();
-  }
-};
-template <>
-struct NullGuard<unsigned char*> final {
-  static const unsigned char* Guard(const unsigned char* v) {
-    return v ? v : kUnsignedCharNull.data();
-  }
-};
-template <>
-struct NullGuard<const unsigned char*> final {
-  static const unsigned char* Guard(const unsigned char* v) {
-    return v ? v : kUnsignedCharNull.data();
-  }
-};
-template <>
-struct NullGuard<std::nullptr_t> final {
-  static const char* Guard(const std::nullptr_t&) { return kCharNull.data(); }
-};
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_NULLGUARD_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/nullstream.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/nullstream.h
deleted file mode 100644
index 9266852e..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/nullstream.h
+++ /dev/null
@@ -1,136 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/internal/nullstream.h
-// -----------------------------------------------------------------------------
-//
-// Classes `NullStream`, `NullStreamMaybeFatal ` and `NullStreamFatal`
-// implement a subset of the `LogMessage` API and are used instead when logging
-// of messages has been disabled.
-
-#ifndef ABSL_LOG_INTERNAL_NULLSTREAM_H_
-#define ABSL_LOG_INTERNAL_NULLSTREAM_H_
-
-#ifdef _WIN32
-#include <cstdlib>
-#else
-#include <unistd.h>
-#endif
-#include <ios>
-#include <ostream>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/log_severity.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-
-// A `NullStream` implements the API of `LogMessage` (a few methods and
-// `operator<<`) but does nothing.  All methods are defined inline so the
-// compiler can eliminate the whole instance and discard anything that's
-// streamed in.
-class NullStream {
- public:
-  NullStream& AtLocation(absl::string_view, int) { return *this; }
-  template <typename SourceLocationType>
-  NullStream& AtLocation(SourceLocationType) {
-    return *this;
-  }
-  NullStream& NoPrefix() { return *this; }
-  NullStream& WithVerbosity(int) { return *this; }
-  template <typename TimeType>
-  NullStream& WithTimestamp(TimeType) {
-    return *this;
-  }
-  template <typename Tid>
-  NullStream& WithThreadID(Tid) {
-    return *this;
-  }
-  template <typename LogEntryType>
-  NullStream& WithMetadataFrom(const LogEntryType&) {
-    return *this;
-  }
-  NullStream& WithPerror() { return *this; }
-  template <typename LogSinkType>
-  NullStream& ToSinkAlso(LogSinkType*) {
-    return *this;
-  }
-  template <typename LogSinkType>
-  NullStream& ToSinkOnly(LogSinkType*) {
-    return *this;
-  }
-  template <typename LogSinkType>
-  NullStream& OutputToSink(LogSinkType*, bool) {
-    return *this;
-  }
-  NullStream& InternalStream() { return *this; }
-};
-template <typename T>
-inline NullStream& operator<<(NullStream& str, const T&) {
-  return str;
-}
-inline NullStream& operator<<(NullStream& str,
-                              std::ostream& (*)(std::ostream& os)) {
-  return str;
-}
-inline NullStream& operator<<(NullStream& str,
-                              std::ios_base& (*)(std::ios_base& os)) {
-  return str;
-}
-
-// `NullStreamMaybeFatal` implements the process termination semantics of
-// `LogMessage`, which is used for `DFATAL` severity and expression-defined
-// severity e.g. `LOG(LEVEL(HowBadIsIt()))`.  Like `LogMessage`, it terminates
-// the process when destroyed if the passed-in severity equals `FATAL`.
-class NullStreamMaybeFatal final : public NullStream {
- public:
-  explicit NullStreamMaybeFatal(absl::LogSeverity severity)
-      : fatal_(severity == absl::LogSeverity::kFatal) {}
-  ~NullStreamMaybeFatal() {
-    if (fatal_) {
-      _exit(1);
-    }
-  }
-
- private:
-  bool fatal_;
-};
-
-// `NullStreamFatal` implements the process termination semantics of
-// `LogMessageFatal`, which means it always terminates the process.  `DFATAL`
-// and expression-defined severity use `NullStreamMaybeFatal` above.
-class NullStreamFatal final : public NullStream {
- public:
-  NullStreamFatal() = default;
-  // ABSL_ATTRIBUTE_NORETURN doesn't seem to work on destructors with msvc, so
-  // disable msvc's warning about the d'tor never returning.
-#if defined(_MSC_VER) && !defined(__clang__)
-#pragma warning(push)
-#pragma warning(disable : 4722)
-#endif
-  ABSL_ATTRIBUTE_NORETURN ~NullStreamFatal() { _exit(1); }
-#if defined(_MSC_VER) && !defined(__clang__)
-#pragma warning(pop)
-#endif
-};
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_GLOBALS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/proto.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/proto.h
deleted file mode 100644
index c8d14acc..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/proto.h
+++ /dev/null
@@ -1,288 +0,0 @@
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-// -----------------------------------------------------------------------------
-// File: internal/proto.h
-// -----------------------------------------------------------------------------
-//
-// Declares functions for serializing and deserializing data to and from memory
-// buffers in protocol buffer wire format.  This library takes no steps to
-// ensure that the encoded data matches with any message specification.
-
-#ifndef ABSL_LOG_INTERNAL_PROTO_H_
-#define ABSL_LOG_INTERNAL_PROTO_H_
-
-#include <cstddef>
-#include <cstdint>
-#include <limits>
-
-#include "absl/base/attributes.h"
-#include "absl/base/casts.h"
-#include "absl/base/config.h"
-#include "absl/strings/string_view.h"
-#include "absl/types/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-
-// absl::Span<char> represents a view into the available space in a mutable
-// buffer during encoding.  Encoding functions shrink the span as they go so
-// that the same view can be passed to a series of Encode functions.  If the
-// data do not fit, nothing is encoded, the view is set to size zero (so that
-// all subsequent encode calls fail), and false is returned.  Otherwise true is
-// returned.
-
-// In particular, attempting to encode a series of data into an insufficient
-// buffer has consistent and efficient behavior without any caller-side error
-// checking.  Individual values will be encoded in their entirety or not at all
-// (unless one of the `Truncate` functions is used).  Once a value is omitted
-// because it does not fit, no subsequent values will be encoded to preserve
-// ordering; the decoded sequence will be a prefix of the original sequence.
-
-// There are two ways to encode a message-typed field:
-//
-// * Construct its contents in a separate buffer and use `EncodeBytes` to copy
-//   it into the primary buffer with type, tag, and length.
-// * Use `EncodeMessageStart` to write type and tag fields and reserve space for
-//   the length field, then encode the contents directly into the buffer, then
-//   use `EncodeMessageLength` to write the actual length into the reserved
-//   bytes.  This works fine if the actual length takes fewer bytes to encode
-//   than were reserved, although you don't get your extra bytes back.
-//   This approach will always produce a valid encoding, but your protocol may
-//   require that the whole message field by omitted if the buffer is too small
-//   to contain all desired subfields.  In this case, operate on a copy of the
-//   buffer view and assign back only if everything fit, i.e. if the last
-//   `Encode` call returned true.
-
-// Encodes the specified integer as a varint field and returns true if it fits.
-// Used for int32_t, int64_t, uint32_t, uint64_t, bool, and enum field types.
-// Consumes up to kMaxVarintSize * 2 bytes (20).
-bool EncodeVarint(uint64_t tag, uint64_t value, absl::Span<char> *buf);
-inline bool EncodeVarint(uint64_t tag, int64_t value, absl::Span<char> *buf) {
-  return EncodeVarint(tag, static_cast<uint64_t>(value), buf);
-}
-inline bool EncodeVarint(uint64_t tag, uint32_t value, absl::Span<char> *buf) {
-  return EncodeVarint(tag, static_cast<uint64_t>(value), buf);
-}
-inline bool EncodeVarint(uint64_t tag, int32_t value, absl::Span<char> *buf) {
-  return EncodeVarint(tag, static_cast<uint64_t>(value), buf);
-}
-
-// Encodes the specified integer as a varint field using ZigZag encoding and
-// returns true if it fits.
-// Used for sint32 and sint64 field types.
-// Consumes up to kMaxVarintSize * 2 bytes (20).
-inline bool EncodeVarintZigZag(uint64_t tag, int64_t value,
-                               absl::Span<char> *buf) {
-  if (value < 0)
-    return EncodeVarint(tag, 2 * static_cast<uint64_t>(-(value + 1)) + 1, buf);
-  return EncodeVarint(tag, 2 * static_cast<uint64_t>(value), buf);
-}
-
-// Encodes the specified integer as a 64-bit field and returns true if it fits.
-// Used for fixed64 and sfixed64 field types.
-// Consumes up to kMaxVarintSize + 8 bytes (18).
-bool Encode64Bit(uint64_t tag, uint64_t value, absl::Span<char> *buf);
-inline bool Encode64Bit(uint64_t tag, int64_t value, absl::Span<char> *buf) {
-  return Encode64Bit(tag, static_cast<uint64_t>(value), buf);
-}
-inline bool Encode64Bit(uint64_t tag, uint32_t value, absl::Span<char> *buf) {
-  return Encode64Bit(tag, static_cast<uint64_t>(value), buf);
-}
-inline bool Encode64Bit(uint64_t tag, int32_t value, absl::Span<char> *buf) {
-  return Encode64Bit(tag, static_cast<uint64_t>(value), buf);
-}
-
-// Encodes the specified double as a 64-bit field and returns true if it fits.
-// Used for double field type.
-// Consumes up to kMaxVarintSize + 8 bytes (18).
-inline bool EncodeDouble(uint64_t tag, double value, absl::Span<char> *buf) {
-  return Encode64Bit(tag, absl::bit_cast<uint64_t>(value), buf);
-}
-
-// Encodes the specified integer as a 32-bit field and returns true if it fits.
-// Used for fixed32 and sfixed32 field types.
-// Consumes up to kMaxVarintSize + 4 bytes (14).
-bool Encode32Bit(uint64_t tag, uint32_t value, absl::Span<char> *buf);
-inline bool Encode32Bit(uint64_t tag, int32_t value, absl::Span<char> *buf) {
-  return Encode32Bit(tag, static_cast<uint32_t>(value), buf);
-}
-
-// Encodes the specified float as a 32-bit field and returns true if it fits.
-// Used for float field type.
-// Consumes up to kMaxVarintSize + 4 bytes (14).
-inline bool EncodeFloat(uint64_t tag, float value, absl::Span<char> *buf) {
-  return Encode32Bit(tag, absl::bit_cast<uint32_t>(value), buf);
-}
-
-// Encodes the specified bytes as a length-delimited field and returns true if
-// they fit.
-// Used for string, bytes, message, and packed-repeated field type.
-// Consumes up to kMaxVarintSize * 2 + value.size() bytes (20 + value.size()).
-bool EncodeBytes(uint64_t tag, absl::Span<const char> value,
-                 absl::Span<char> *buf);
-
-// Encodes as many of the specified bytes as will fit as a length-delimited
-// field and returns true as long as the field header (`tag_type` and `length`)
-// fits.
-// Used for string, bytes, message, and packed-repeated field type.
-// Consumes up to kMaxVarintSize * 2 + value.size() bytes (20 + value.size()).
-bool EncodeBytesTruncate(uint64_t tag, absl::Span<const char> value,
-                         absl::Span<char> *buf);
-
-// Encodes the specified string as a length-delimited field and returns true if
-// it fits.
-// Used for string, bytes, message, and packed-repeated field type.
-// Consumes up to kMaxVarintSize * 2 + value.size() bytes (20 + value.size()).
-inline bool EncodeString(uint64_t tag, absl::string_view value,
-                         absl::Span<char> *buf) {
-  return EncodeBytes(tag, value, buf);
-}
-
-// Encodes as much of the specified string as will fit as a length-delimited
-// field and returns true as long as the field header (`tag_type` and `length`)
-// fits.
-// Used for string, bytes, message, and packed-repeated field type.
-// Consumes up to kMaxVarintSize * 2 + value.size() bytes (20 + value.size()).
-inline bool EncodeStringTruncate(uint64_t tag, absl::string_view value,
-                                 absl::Span<char> *buf) {
-  return EncodeBytesTruncate(tag, value, buf);
-}
-
-// Encodes the header for a length-delimited field containing up to `max_size`
-// bytes or the number remaining in the buffer, whichever is less.  If the
-// header fits, a non-nullptr `Span` is returned; this must be passed to
-// `EncodeMessageLength` after all contents are encoded to finalize the length
-// field.  If the header does not fit, a nullptr `Span` is returned which is
-// safe to pass to `EncodeMessageLength` but need not be.
-// Used for string, bytes, message, and packed-repeated field type.
-// Consumes up to kMaxVarintSize * 2 bytes (20).
-ABSL_MUST_USE_RESULT absl::Span<char> EncodeMessageStart(uint64_t tag,
-                                                         uint64_t max_size,
-                                                         absl::Span<char> *buf);
-
-// Finalizes the length field in `msg` so that it encompasses all data encoded
-// since the call to `EncodeMessageStart` which returned `msg`.  Does nothing if
-// `msg` is a `nullptr` `Span`.
-void EncodeMessageLength(absl::Span<char> msg, const absl::Span<char> *buf);
-
-enum class WireType : uint64_t {
-  kVarint = 0,
-  k64Bit = 1,
-  kLengthDelimited = 2,
-  k32Bit = 5,
-};
-
-constexpr size_t VarintSize(uint64_t value) {
-  return value < 128 ? 1 : 1 + VarintSize(value >> 7);
-}
-constexpr size_t MinVarintSize() {
-  return VarintSize((std::numeric_limits<uint64_t>::min)());
-}
-constexpr size_t MaxVarintSize() {
-  return VarintSize((std::numeric_limits<uint64_t>::max)());
-}
-
-constexpr uint64_t MaxVarintForSize(size_t size) {
-  return size >= 10 ? (std::numeric_limits<uint64_t>::max)()
-                    : (static_cast<uint64_t>(1) << size * 7) - 1;
-}
-
-// `BufferSizeFor` returns a number of bytes guaranteed to be sufficient to
-// store encoded fields of the specified WireTypes regardless of tag numbers and
-// data values.  This only makes sense for `WireType::kLengthDelimited` if you
-// add in the length of the contents yourself, e.g. for string and bytes fields
-// by adding the lengths of any encoded strings to the return value or for
-// submessage fields by enumerating the fields you may encode into their
-// contents.
-constexpr size_t BufferSizeFor() { return 0; }
-template <typename... T>
-constexpr size_t BufferSizeFor(WireType type, T... tail) {
-  // tag_type + data + ...
-  return MaxVarintSize() +
-         (type == WireType::kVarint ? MaxVarintSize() :              //
-              type == WireType::k64Bit ? 8 :                         //
-                  type == WireType::k32Bit ? 4 : MaxVarintSize()) +  //
-         BufferSizeFor(tail...);
-}
-
-// absl::Span<const char> represents a view into the un-processed space in a
-// buffer during decoding.  Decoding functions shrink the span as they go so
-// that the same view can be decoded iteratively until all data are processed.
-// In general, if the buffer is exhausted but additional bytes are expected by
-// the decoder, it will return values as if the additional bytes were zeros.
-// Length-delimited fields are an exception - if the encoded length field
-// indicates more data bytes than are available in the buffer, the `bytes_value`
-// and `string_value` accessors will return truncated views.
-
-class ProtoField final {
- public:
-  // Consumes bytes from `data` and returns true if there were any bytes to
-  // decode.
-  bool DecodeFrom(absl::Span<const char> *data);
-  uint64_t tag() const { return tag_; }
-  WireType type() const { return type_; }
-
-  // These value accessors will return nonsense if the data were not encoded in
-  // the corresponding wiretype from the corresponding C++ (or other language)
-  // type.
-
-  double double_value() const { return absl::bit_cast<double>(value_); }
-  float float_value() const {
-    return absl::bit_cast<float>(static_cast<uint32_t>(value_));
-  }
-  int32_t int32_value() const { return static_cast<int32_t>(value_); }
-  int64_t int64_value() const { return static_cast<int64_t>(value_); }
-  int32_t sint32_value() const {
-    if (value_ % 2) return static_cast<int32_t>(0 - ((value_ - 1) / 2) - 1);
-    return static_cast<int32_t>(value_ / 2);
-  }
-  int64_t sint64_value() const {
-    if (value_ % 2) return 0 - ((value_ - 1) / 2) - 1;
-    return value_ / 2;
-  }
-  uint32_t uint32_value() const { return static_cast<uint32_t>(value_); }
-  uint64_t uint64_value() const { return value_; }
-  bool bool_value() const { return value_ != 0; }
-  // To decode an enum, call int32_value() and cast to the appropriate type.
-  // Note that the official C++ proto compiler treats enum fields with values
-  // that do not correspond to a defined enumerator as unknown fields.
-
-  // To decode fields within a submessage field, call
-  // `DecodeNextField(field.BytesValue())`.
-  absl::Span<const char> bytes_value() const { return data_; }
-  absl::string_view string_value() const {
-    const auto data = bytes_value();
-    return absl::string_view(data.data(), data.size());
-  }
-  // Returns the encoded length of a length-delimited field.  This equals
-  // `bytes_value().size()` except when the latter has been truncated due to
-  // buffer underrun.
-  uint64_t encoded_length() const { return value_; }
-
- private:
-  uint64_t tag_;
-  WireType type_;
-  // For `kTypeVarint`, `kType64Bit`, and `kType32Bit`, holds the decoded value.
-  // For `kTypeLengthDelimited`, holds the decoded length.
-  uint64_t value_;
-  absl::Span<const char> data_;
-};
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_PROTO_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/strip.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/strip.h
deleted file mode 100644
index f8d27869..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/strip.h
+++ /dev/null
@@ -1,72 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/internal/strip.h
-// -----------------------------------------------------------------------------
-//
-
-#ifndef ABSL_LOG_INTERNAL_STRIP_H_
-#define ABSL_LOG_INTERNAL_STRIP_H_
-
-#include "absl/base/log_severity.h"
-#include "absl/log/internal/log_message.h"
-#include "absl/log/internal/nullstream.h"
-
-// `ABSL_LOGGING_INTERNAL_LOG_*` evaluates to a temporary `LogMessage` object or
-// to a related object with a compatible API but different behavior.  This set
-// of defines comes in three flavors: vanilla, plus two variants that strip some
-// logging in subtly different ways for subtly different reasons (see below).
-#if defined(STRIP_LOG) && STRIP_LOG
-#define ABSL_LOGGING_INTERNAL_LOG_INFO ::absl::log_internal::NullStream()
-#define ABSL_LOGGING_INTERNAL_LOG_WARNING ::absl::log_internal::NullStream()
-#define ABSL_LOGGING_INTERNAL_LOG_ERROR ::absl::log_internal::NullStream()
-#define ABSL_LOGGING_INTERNAL_LOG_FATAL ::absl::log_internal::NullStreamFatal()
-#define ABSL_LOGGING_INTERNAL_LOG_QFATAL ::absl::log_internal::NullStreamFatal()
-#define ABSL_LOGGING_INTERNAL_LOG_DFATAL \
-  ::absl::log_internal::NullStreamMaybeFatal(::absl::kLogDebugFatal)
-#define ABSL_LOGGING_INTERNAL_LOG_LEVEL(severity) \
-  ::absl::log_internal::NullStreamMaybeFatal(absl_log_internal_severity)
-#define ABSL_LOG_INTERNAL_CHECK(failure_message) ABSL_LOGGING_INTERNAL_LOG_FATAL
-#define ABSL_LOG_INTERNAL_QCHECK(failure_message) \
-  ABSL_LOGGING_INTERNAL_LOG_QFATAL
-#else  // !defined(STRIP_LOG) || !STRIP_LOG
-#define ABSL_LOGGING_INTERNAL_LOG_INFO \
-  ::absl::log_internal::LogMessage(    \
-      __FILE__, __LINE__, ::absl::log_internal::LogMessage::InfoTag{})
-#define ABSL_LOGGING_INTERNAL_LOG_WARNING \
-  ::absl::log_internal::LogMessage(       \
-      __FILE__, __LINE__, ::absl::log_internal::LogMessage::WarningTag{})
-#define ABSL_LOGGING_INTERNAL_LOG_ERROR \
-  ::absl::log_internal::LogMessage(     \
-      __FILE__, __LINE__, ::absl::log_internal::LogMessage::ErrorTag{})
-#define ABSL_LOGGING_INTERNAL_LOG_FATAL \
-  ::absl::log_internal::LogMessageFatal(__FILE__, __LINE__)
-#define ABSL_LOGGING_INTERNAL_LOG_QFATAL \
-  ::absl::log_internal::LogMessageQuietlyFatal(__FILE__, __LINE__)
-#define ABSL_LOGGING_INTERNAL_LOG_DFATAL \
-  ::absl::log_internal::LogMessage(__FILE__, __LINE__, ::absl::kLogDebugFatal)
-#define ABSL_LOGGING_INTERNAL_LOG_LEVEL(severity)      \
-  ::absl::log_internal::LogMessage(__FILE__, __LINE__, \
-                                   absl_log_internal_severity)
-// These special cases dispatch to special-case constructors that allow us to
-// avoid an extra function call and shrink non-LTO binaries by a percent or so.
-#define ABSL_LOG_INTERNAL_CHECK(failure_message) \
-  ::absl::log_internal::LogMessageFatal(__FILE__, __LINE__, failure_message)
-#define ABSL_LOG_INTERNAL_QCHECK(failure_message)                  \
-  ::absl::log_internal::LogMessageQuietlyFatal(__FILE__, __LINE__, \
-                                               failure_message)
-#endif  // !defined(STRIP_LOG) || !STRIP_LOG
-
-#endif  // ABSL_LOG_INTERNAL_STRIP_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/structured.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/structured.h
deleted file mode 100644
index 5223dbc3..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/structured.h
+++ /dev/null
@@ -1,58 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/internal/structured.h
-// -----------------------------------------------------------------------------
-
-#ifndef ABSL_LOG_INTERNAL_STRUCTURED_H_
-#define ABSL_LOG_INTERNAL_STRUCTURED_H_
-
-#include <ostream>
-
-#include "absl/base/config.h"
-#include "absl/log/internal/log_message.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-
-class ABSL_MUST_USE_RESULT AsLiteralImpl final {
- public:
-  explicit AsLiteralImpl(absl::string_view str) : str_(str) {}
-  AsLiteralImpl(const AsLiteralImpl&) = default;
-  AsLiteralImpl& operator=(const AsLiteralImpl&) = default;
-
- private:
-  absl::string_view str_;
-
-  friend std::ostream& operator<<(std::ostream& os, AsLiteralImpl as_literal) {
-    return os << as_literal.str_;
-  }
-  void AddToMessage(log_internal::LogMessage& m) {
-    m.CopyToEncodedBuffer<log_internal::LogMessage::StringType::kLiteral>(str_);
-  }
-  friend log_internal::LogMessage& operator<<(log_internal::LogMessage& m,
-                                              AsLiteralImpl as_literal) {
-    as_literal.AddToMessage(m);
-    return m;
-  }
-};
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_STRUCTURED_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/test_actions.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/test_actions.h
deleted file mode 100644
index 649a0505..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/test_actions.h
+++ /dev/null
@@ -1,90 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/internal/test_actions.h
-// -----------------------------------------------------------------------------
-//
-// This file declares Googletest's actions used in the Abseil Logging library
-// unit tests.
-
-#ifndef ABSL_LOG_INTERNAL_TEST_ACTIONS_H_
-#define ABSL_LOG_INTERNAL_TEST_ACTIONS_H_
-
-#include <iostream>
-#include <ostream>
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/base/log_severity.h"
-#include "absl/log/log_entry.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-
-// These actions are used by the child process in a death test.
-//
-// Expectations set in the child cannot cause test failure in the parent
-// directly.  Instead, the child can use these actions with
-// `EXPECT_CALL`/`WillOnce` and `ON_CALL`/`WillByDefault` (for unexpected calls)
-// to write messages to stderr that the parent can match against.
-struct WriteToStderr final {
-  explicit WriteToStderr(absl::string_view m) : message(m) {}
-  std::string message;
-
-  template <typename... Args>
-  void operator()(const Args&...) const {
-    std::cerr << message << std::endl;
-  }
-};
-
-struct WriteToStderrWithFilename final {
-  explicit WriteToStderrWithFilename(absl::string_view m) : message(m) {}
-
-  std::string message;
-
-  void operator()(const absl::LogEntry& entry) const;
-};
-
-struct WriteEntryToStderr final {
-  explicit WriteEntryToStderr(absl::string_view m) : message(m) {}
-
-  std::string message = "";
-
-  void operator()(const absl::LogEntry& entry) const;
-  void operator()(absl::LogSeverity, absl::string_view,
-                  absl::string_view) const;
-};
-
-// See the documentation for `DeathTestValidateExpectations` above.
-// `DeathTestExpectedLogging` should be used once in a given death test, and the
-// applicable severity level is the one that should be passed to
-// `DeathTestValidateExpectations`.
-inline WriteEntryToStderr DeathTestExpectedLogging() {
-  return WriteEntryToStderr{"Mock received expected entry:"};
-}
-
-// `DeathTestUnexpectedLogging` should be used zero or more times to mark
-// messages that should not hit the logs as the process dies.
-inline WriteEntryToStderr DeathTestUnexpectedLogging() {
-  return WriteEntryToStderr{"Mock received unexpected entry:"};
-}
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_TEST_ACTIONS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/test_helpers.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/test_helpers.h
deleted file mode 100644
index 714bc7bd..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/test_helpers.h
+++ /dev/null
@@ -1,71 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/internal/test_helpers.h
-// -----------------------------------------------------------------------------
-//
-// This file declares testing helpers for the logging library.
-
-#ifndef ABSL_LOG_INTERNAL_TEST_HELPERS_H_
-#define ABSL_LOG_INTERNAL_TEST_HELPERS_H_
-
-#include "gtest/gtest.h"
-#include "absl/base/config.h"
-#include "absl/base/log_severity.h"
-#include "absl/log/globals.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-
-// `ABSL_MIN_LOG_LEVEL` can't be used directly since it is not always defined.
-constexpr auto kAbslMinLogLevel =
-#ifdef ABSL_MIN_LOG_LEVEL
-    static_cast<absl::LogSeverityAtLeast>(ABSL_MIN_LOG_LEVEL);
-#else
-    absl::LogSeverityAtLeast::kInfo;
-#endif
-
-// Returns false if the specified severity level is disabled by
-// `ABSL_MIN_LOG_LEVEL` or `absl::MinLogLevel()`.
-bool LoggingEnabledAt(absl::LogSeverity severity);
-
-// -----------------------------------------------------------------------------
-// Googletest Death Test Predicates
-// -----------------------------------------------------------------------------
-
-#if GTEST_HAS_DEATH_TEST
-
-bool DiedOfFatal(int exit_status);
-bool DiedOfQFatal(int exit_status);
-
-#endif
-
-// -----------------------------------------------------------------------------
-// Helper for Log initialization in test
-// -----------------------------------------------------------------------------
-
-class LogTestEnvironment : public ::testing::Environment {
- public:
-  ~LogTestEnvironment() override = default;
-
-  void SetUp() override;
-};
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_TEST_HELPERS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/test_matchers.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/test_matchers.h
deleted file mode 100644
index fc653a91..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/test_matchers.h
+++ /dev/null
@@ -1,94 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/internal/test_matchers.h
-// -----------------------------------------------------------------------------
-//
-// This file declares Googletest's matchers used in the Abseil Logging library
-// unit tests.
-
-#ifndef ABSL_LOG_INTERNAL_TEST_MATCHERS_H_
-#define ABSL_LOG_INTERNAL_TEST_MATCHERS_H_
-
-#include <iosfwd>
-#include <sstream>
-#include <string>
-
-#include "gmock/gmock.h"
-#include "gtest/gtest.h"
-#include "absl/base/config.h"
-#include "absl/base/log_severity.h"
-#include "absl/log/internal/test_helpers.h"
-#include "absl/log/log_entry.h"
-#include "absl/strings/string_view.h"
-#include "absl/time/time.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-// In some configurations, Googletest's string matchers (e.g.
-// `::testing::EndsWith`) need help to match `absl::string_view`.
-::testing::Matcher<absl::string_view> AsString(
-    const ::testing::Matcher<const std::string&>& str_matcher);
-
-// These matchers correspond to the components of `absl::LogEntry`.
-::testing::Matcher<const absl::LogEntry&> SourceFilename(
-    const ::testing::Matcher<absl::string_view>& source_filename);
-::testing::Matcher<const absl::LogEntry&> SourceBasename(
-    const ::testing::Matcher<absl::string_view>& source_basename);
-// Be careful with this one; multi-line statements using `__LINE__` evaluate
-// differently on different platforms.  In particular, the MSVC implementation
-// of `EXPECT_DEATH` returns the line number of the macro expansion to all lines
-// within the code block that's expected to die.
-::testing::Matcher<const absl::LogEntry&> SourceLine(
-    const ::testing::Matcher<int>& source_line);
-::testing::Matcher<const absl::LogEntry&> Prefix(
-    const ::testing::Matcher<bool>& prefix);
-::testing::Matcher<const absl::LogEntry&> LogSeverity(
-    const ::testing::Matcher<absl::LogSeverity>& log_severity);
-::testing::Matcher<const absl::LogEntry&> Timestamp(
-    const ::testing::Matcher<absl::Time>& timestamp);
-// Matches if the `LogEntry`'s timestamp falls after the instantiation of this
-// matcher and before its execution, as is normal when used with EXPECT_CALL.
-::testing::Matcher<const absl::LogEntry&> TimestampInMatchWindow();
-::testing::Matcher<const absl::LogEntry&> ThreadID(
-    const ::testing::Matcher<absl::LogEntry::tid_t>&);
-::testing::Matcher<const absl::LogEntry&> TextMessageWithPrefixAndNewline(
-    const ::testing::Matcher<absl::string_view>&
-        text_message_with_prefix_and_newline);
-::testing::Matcher<const absl::LogEntry&> TextMessageWithPrefix(
-    const ::testing::Matcher<absl::string_view>& text_message_with_prefix);
-::testing::Matcher<const absl::LogEntry&> TextMessage(
-    const ::testing::Matcher<absl::string_view>& text_message);
-::testing::Matcher<const absl::LogEntry&> TextPrefix(
-    const ::testing::Matcher<absl::string_view>& text_prefix);
-::testing::Matcher<const absl::LogEntry&> Verbosity(
-    const ::testing::Matcher<int>& verbosity);
-::testing::Matcher<const absl::LogEntry&> Stacktrace(
-    const ::testing::Matcher<absl::string_view>& stacktrace);
-// Behaves as `Eq(stream.str())`, but produces better failure messages.
-::testing::Matcher<absl::string_view> MatchesOstream(
-    const std::ostringstream& stream);
-::testing::Matcher<const std::string&> DeathTestValidateExpectations();
-
-::testing::Matcher<const absl::LogEntry&> RawEncodedMessage(
-    const ::testing::Matcher<absl::string_view>& raw_encoded_message);
-#define ENCODED_MESSAGE(message_matcher) ::testing::_
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_TEST_MATCHERS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/vlog_config.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/vlog_config.h
deleted file mode 100644
index b6e322c4..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/vlog_config.h
+++ /dev/null
@@ -1,163 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// vlog_config.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines `VLogSite`, a public primitive that represents
-// a callsite for the `VLOG` family of macros and related libraries.
-// It also declares and defines multiple internal utilities used to implement
-// `VLOG`, such as `VLogSiteManager`.
-
-#ifndef ABSL_LOG_INTERNAL_VLOG_CONFIG_H_
-#define ABSL_LOG_INTERNAL_VLOG_CONFIG_H_
-
-// IWYU pragma: private, include "absl/log/log.h"
-
-#include <atomic>
-#include <cstdint>
-#include <functional>
-#include <limits>
-#include <type_traits>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/optimization.h"
-#include "absl/base/thread_annotations.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-
-class SyntheticBinary;
-class VLogSite;
-
-int RegisterAndInitialize(VLogSite* v);
-void UpdateVLogSites();
-constexpr int kUseFlag = (std::numeric_limits<int16_t>::min)();
-
-// Represents a unique callsite for a `VLOG()` or `VLOG_IS_ON()` call.
-//
-// Libraries that provide `VLOG`-like functionality should use this to
-// efficiently handle --vmodule.
-//
-// VLogSite objects must not be destroyed until the program exits. Doing so will
-// probably yield nasty segfaults in VLogSiteManager::UpdateLogSites(). The
-// recommendation is to make all such objects function-local statics.
-class VLogSite final {
- public:
-  // `f` must not be destroyed until the program exits.
-  explicit constexpr VLogSite(const char* f)
-      : file_(f), v_(kUninitialized), next_(nullptr) {}
-  VLogSite(const VLogSite&) = delete;
-  VLogSite& operator=(const VLogSite&) = delete;
-
-  // Inlining the function yields a ~3x performance improvement at the cost of a
-  // 1.5x code size increase at the call site.
-  // Takes locks but does not allocate memory.
-  ABSL_ATTRIBUTE_ALWAYS_INLINE
-  bool IsEnabled(int level) {
-    int stale_v = v_.load(std::memory_order_relaxed);
-    if (ABSL_PREDICT_TRUE(level > stale_v)) {
-      return false;
-    }
-
-    // We put everything other than the fast path, i.e. vlogging is initialized
-    // but not on, behind an out-of-line function to reduce code size.
-    // "level" is almost always a call-site constant, so we can save a bit
-    // of code space by special-casing for a few common levels.
-#if ABSL_HAVE_BUILTIN(__builtin_constant_p) || defined(__GNUC__)
-    if (__builtin_constant_p(level)) {
-      if (level == 0) return SlowIsEnabled0(stale_v);
-      if (level == 1) return SlowIsEnabled1(stale_v);
-      if (level == 2) return SlowIsEnabled2(stale_v);
-      if (level == 3) return SlowIsEnabled3(stale_v);
-      if (level == 4) return SlowIsEnabled4(stale_v);
-      if (level == 5) return SlowIsEnabled5(stale_v);
-    }
-#endif
-    return SlowIsEnabled(stale_v, level);
-  }
-
- private:
-  friend int log_internal::RegisterAndInitialize(VLogSite* v);
-  friend void log_internal::UpdateVLogSites();
-  friend class log_internal::SyntheticBinary;
-  static constexpr int kUninitialized = (std::numeric_limits<int>::max)();
-
-  // SlowIsEnabled performs slower checks to determine whether a log site is
-  // enabled. Because it is expected to be called somewhat rarely
-  // (comparatively), it is not inlined to save on code size.
-  //
-  // Prerequisites to calling SlowIsEnabled:
-  //   1) stale_v is uninitialized OR
-  //   2) stale_v is initialized and >= level (meaning we must log).
-  // Takes locks but does not allocate memory.
-  ABSL_ATTRIBUTE_NOINLINE
-  bool SlowIsEnabled(int stale_v, int level);
-  ABSL_ATTRIBUTE_NOINLINE bool SlowIsEnabled0(int stale_v);
-  ABSL_ATTRIBUTE_NOINLINE bool SlowIsEnabled1(int stale_v);
-  ABSL_ATTRIBUTE_NOINLINE bool SlowIsEnabled2(int stale_v);
-  ABSL_ATTRIBUTE_NOINLINE bool SlowIsEnabled3(int stale_v);
-  ABSL_ATTRIBUTE_NOINLINE bool SlowIsEnabled4(int stale_v);
-  ABSL_ATTRIBUTE_NOINLINE bool SlowIsEnabled5(int stale_v);
-
-  // This object is too size-sensitive to use absl::string_view.
-  const char* const file_;
-  std::atomic<int> v_;
-  std::atomic<VLogSite*> next_;
-};
-static_assert(std::is_trivially_destructible<VLogSite>::value,
-              "VLogSite must be trivially destructible");
-
-// Returns the current verbose log level of `file`.
-// Does not allocate memory.
-int VLogLevel(absl::string_view file);
-
-// Registers a site `v` to get updated as `vmodule` and `v` change.  Also
-// initializes the site based on their current values, and returns that result.
-// Does not allocate memory.
-int RegisterAndInitialize(VLogSite* v);
-
-// Allocates memory.
-void UpdateVLogSites();
-
-// Completely overwrites the saved value of `vmodule`.
-// Allocates memory.
-void UpdateVModule(absl::string_view vmodule);
-
-// Updates the global verbosity level to `v` and returns the prior value.
-// Allocates memory.
-int UpdateGlobalVLogLevel(int v);
-
-// Atomically prepends `module_pattern=log_level` to the start of vmodule.
-// Returns the prior value for `module_pattern` if there was an exact match and
-// `global_v` otherwise.
-// Allocates memory.
-int PrependVModule(absl::string_view module_pattern, int log_level);
-
-// Registers `on_update` to be called whenever `v` or `vmodule` change.
-// Allocates memory.
-void OnVLogVerbosityUpdate(std::function<void()> cb);
-
-// Does not allocate memory.
-VLogSite* SetVModuleListHeadForTestOnly(VLogSite* v);
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_VLOG_CONFIG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/internal/voidify.h b/projects/llm_framework/include/abseil-cpp/absl/log/internal/voidify.h
deleted file mode 100644
index 8f62da20..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/internal/voidify.h
+++ /dev/null
@@ -1,44 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/internal/voidify.h
-// -----------------------------------------------------------------------------
-//
-// This class is used to explicitly ignore values in the conditional logging
-// macros. This avoids compiler warnings like "value computed is not used" and
-// "statement has no effect".
-
-#ifndef ABSL_LOG_INTERNAL_VOIDIFY_H_
-#define ABSL_LOG_INTERNAL_VOIDIFY_H_
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace log_internal {
-
-class Voidify final {
- public:
-  // This has to be an operator with a precedence lower than << but higher than
-  // ?:
-  template <typename T>
-  void operator&&(const T&) const&& {}
-};
-
-}  // namespace log_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_INTERNAL_VOIDIFY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/log.h b/projects/llm_framework/include/abseil-cpp/absl/log/log.h
deleted file mode 100644
index b721b087..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/log.h
+++ /dev/null
@@ -1,361 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/log.h
-// -----------------------------------------------------------------------------
-//
-// This header declares a family of LOG macros.
-//
-// Basic invocation looks like this:
-//
-//   LOG(INFO) << "Found " << num_cookies << " cookies";
-//
-// Most `LOG` macros take a severity level argument.  The severity levels are
-// `INFO`, `WARNING`, `ERROR`, and `FATAL`.  They are defined
-// in absl/base/log_severity.h.
-// * The `FATAL` severity level terminates the program with a stack trace after
-//   logging its message.  Error handlers registered with `RunOnFailure`
-//   (process_state.h) are run, but exit handlers registered with `atexit(3)`
-//   are not.
-// * The `QFATAL` pseudo-severity level is equivalent to `FATAL` but triggers
-//   quieter termination messages, e.g. without a full stack trace, and skips
-//   running registered error handlers.
-// * The `DFATAL` pseudo-severity level is defined as `FATAL` in debug mode and
-//   as `ERROR` otherwise.
-// Some preprocessor shenanigans are used to ensure that e.g. `LOG(INFO)` has
-// the same meaning even if a local symbol or preprocessor macro named `INFO` is
-// defined.  To specify a severity level using an expression instead of a
-// literal, use `LEVEL(expr)`.
-// Example:
-//
-//   LOG(LEVEL(stale ? absl::LogSeverity::kWarning : absl::LogSeverity::kInfo))
-//       << "Cookies are " << days << " days old";
-
-// `LOG` macros evaluate to an unterminated statement.  The value at the end of
-// the statement supports some chainable methods:
-//
-//   * .AtLocation(absl::string_view file, int line)
-//     .AtLocation(absl::SourceLocation loc)
-//     Overrides the location inferred from the callsite.  The string pointed to
-//     by `file` must be valid until the end of the statement.
-//   * .NoPrefix()
-//     Omits the prefix from this line.  The prefix includes metadata about the
-//     logged data such as source code location and timestamp.
-//   * .WithVerbosity(int verbose_level)
-//     Sets the verbosity field of the logged message as if it was logged by
-//     `VLOG(verbose_level)`.  Unlike `VLOG`, this method does not affect
-//     evaluation of the statement when the specified `verbose_level` has been
-//     disabled.  The only effect is on `LogSink` implementations which make use
-//     of the `absl::LogSink::verbosity()` value.  The value
-//     `absl::LogEntry::kNoVerbosityLevel` can be specified to mark the message
-//     not verbose.
-//   * .WithTimestamp(absl::Time timestamp)
-//     Uses the specified timestamp instead of one collected at the time of
-//     execution.
-//   * .WithThreadID(absl::LogEntry::tid_t tid)
-//     Uses the specified thread ID instead of one collected at the time of
-//     execution.
-//   * .WithMetadataFrom(const absl::LogEntry &entry)
-//     Copies all metadata (but no data) from the specified `absl::LogEntry`.
-//     This can be used to change the severity of a message, but it has some
-//     limitations:
-//     * `ABSL_MIN_LOG_LEVEL` is evaluated against the severity passed into
-//       `LOG` (or the implicit `FATAL` level of `CHECK`).
-//     * `LOG(FATAL)` and `CHECK` terminate the process unconditionally, even if
-//       the severity is changed later.
-//     `.WithMetadataFrom(entry)` should almost always be used in combination
-//     with `LOG(LEVEL(entry.log_severity()))`.
-//   * .WithPerror()
-//     Appends to the logged message a colon, a space, a textual description of
-//     the current value of `errno` (as by `strerror(3)`), and the numerical
-//     value of `errno`.
-//   * .ToSinkAlso(absl::LogSink* sink)
-//     Sends this message to `*sink` in addition to whatever other sinks it
-//     would otherwise have been sent to.  `sink` must not be null.
-//   * .ToSinkOnly(absl::LogSink* sink)
-//     Sends this message to `*sink` and no others.  `sink` must not be null.
-//
-// No interfaces in this header are async-signal-safe; their use in signal
-// handlers is unsupported and may deadlock your program or eat your lunch.
-//
-// Many logging statements are inherently conditional.  For example,
-// `LOG_IF(INFO, !foo)` does nothing if `foo` is true.  Even seemingly
-// unconditional statements like `LOG(INFO)` might be disabled at
-// compile-time to minimize binary size or for security reasons.
-//
-// * Except for the condition in a `CHECK` or `QCHECK` statement, programs must
-//   not rely on evaluation of expressions anywhere in logging statements for
-//   correctness.  For example, this is ok:
-//
-//     CHECK((fp = fopen("config.ini", "r")) != nullptr);
-//
-//   But this is probably not ok:
-//
-//     LOG(INFO) << "Server status: " << StartServerAndReturnStatusString();
-//
-//   The example below is bad too; the `i++` in the `LOG_IF` condition might
-//   not be evaluated, resulting in an infinite loop:
-//
-//     for (int i = 0; i < 1000000;)
-//       LOG_IF(INFO, i++ % 1000 == 0) << "Still working...";
-//
-// * Except where otherwise noted, conditions which cause a statement not to log
-//   also cause expressions not to be evaluated.  Programs may rely on this for
-//   performance reasons, e.g. by streaming the result of an expensive function
-//   call into a `DLOG` or `LOG_EVERY_N` statement.
-// * Care has been taken to ensure that expressions are parsed by the compiler
-//   even if they are never evaluated.  This means that syntax errors will be
-//   caught and variables will be considered used for the purposes of
-//   unused-variable diagnostics.  For example, this statement won't compile
-//   even if `INFO`-level logging has been compiled out:
-//
-//     int number_of_cakes = 40;
-//     LOG(INFO) << "Number of cakes: " << number_of_cake;  // Note the typo!
-//
-//   Similarly, this won't produce unused-variable compiler diagnostics even
-//   if `INFO`-level logging is compiled out:
-//
-//     {
-//       char fox_line1[] = "Hatee-hatee-hatee-ho!";
-//       LOG_IF(ERROR, false) << "The fox says " << fox_line1;
-//       char fox_line2[] = "A-oo-oo-oo-ooo!";
-//       LOG(INFO) << "The fox also says " << fox_line2;
-//     }
-//
-//   This error-checking is not perfect; for example, symbols that have been
-//   declared but not defined may not produce link errors if used in logging
-//   statements that compile away.
-//
-// Expressions streamed into these macros are formatted using `operator<<` just
-// as they would be if streamed into a `std::ostream`, however it should be
-// noted that their actual type is unspecified.
-//
-// To implement a custom formatting operator for a type you own, there are two
-// options: `AbslStringify()` or `std::ostream& operator<<(std::ostream&, ...)`.
-// It is recommended that users make their types loggable through
-// `AbslStringify()` as it is a universal stringification extension that also
-// enables `absl::StrFormat` and `absl::StrCat` support. If both
-// `AbslStringify()` and `std::ostream& operator<<(std::ostream&, ...)` are
-// defined, `AbslStringify()` will be used.
-//
-// To use the `AbslStringify()` API, define a friend function template in your
-// type's namespace with the following signature:
-//
-//   template <typename Sink>
-//   void AbslStringify(Sink& sink, const UserDefinedType& value);
-//
-// `Sink` has the same interface as `absl::FormatSink`, but without
-// `PutPaddedString()`.
-//
-// Example:
-//
-//   struct Point {
-//     template <typename Sink>
-//     friend void AbslStringify(Sink& sink, const Point& p) {
-//       absl::Format(&sink, "(%v, %v)", p.x, p.y);
-//     }
-//
-//     int x;
-//     int y;
-//   };
-//
-// To use `std::ostream& operator<<(std::ostream&, ...)`, define
-// `std::ostream& operator<<(std::ostream&, ...)` in your type's namespace (for
-// ADL) just as you would to stream it to `std::cout`.
-//
-// Currently `AbslStringify()` ignores output manipulators but this is not
-// guaranteed behavior and may be subject to change in the future. If you would
-// like guaranteed behavior regarding output manipulators, please use
-// `std::ostream& operator<<(std::ostream&, ...)` to make custom types loggable
-// instead.
-//
-// Those macros that support streaming honor output manipulators and `fmtflag`
-// changes that output data (e.g. `std::ends`) or control formatting of data
-// (e.g. `std::hex` and `std::fixed`), however flushing such a stream is
-// ignored.  The message produced by a log statement is sent to registered
-// `absl::LogSink` instances at the end of the statement; those sinks are
-// responsible for their own flushing (e.g. to disk) semantics.
-//
-// Flag settings are not carried over from one `LOG` statement to the next; this
-// is a bit different than e.g. `std::cout`:
-//
-//   LOG(INFO) << std::hex << 0xdeadbeef;  // logs "0xdeadbeef"
-//   LOG(INFO) << 0xdeadbeef;              // logs "3735928559"
-
-#ifndef ABSL_LOG_LOG_H_
-#define ABSL_LOG_LOG_H_
-
-#include "absl/log/internal/log_impl.h"
-#include "absl/log/vlog_is_on.h"  // IWYU pragma: export
-
-// LOG()
-//
-// `LOG` takes a single argument which is a severity level.  Data streamed in
-// comprise the logged message.
-// Example:
-//
-//   LOG(INFO) << "Found " << num_cookies << " cookies";
-#define LOG(severity) ABSL_LOG_INTERNAL_LOG_IMPL(_##severity)
-
-// PLOG()
-//
-// `PLOG` behaves like `LOG` except that a description of the current state of
-// `errno` is appended to the streamed message.
-#define PLOG(severity) ABSL_LOG_INTERNAL_PLOG_IMPL(_##severity)
-
-// DLOG()
-//
-// `DLOG` behaves like `LOG` in debug mode (i.e. `#ifndef NDEBUG`).  Otherwise
-// it compiles away and does nothing.  Note that `DLOG(FATAL)` does not
-// terminate the program if `NDEBUG` is defined.
-#define DLOG(severity) ABSL_LOG_INTERNAL_DLOG_IMPL(_##severity)
-
-// `VLOG` uses numeric levels to provide verbose logging that can configured at
-// runtime, including at a per-module level.  `VLOG` statements are logged at
-// `INFO` severity if they are logged at all; the numeric levels are on a
-// different scale than the proper severity levels.  Positive levels are
-// disabled by default.  Negative levels should not be used.
-// Example:
-//
-//   VLOG(1) << "I print when you run the program with --v=1 or higher";
-//   VLOG(2) << "I print when you run the program with --v=2 or higher";
-//
-// See vlog_is_on.h for further documentation, including the usage of the
-// --vmodule flag to log at different levels in different source files.
-#define VLOG(severity) ABSL_LOG_INTERNAL_VLOG_IMPL(severity)
-
-// `DVLOG` behaves like `VLOG` in debug mode (i.e. `#ifndef NDEBUG`).
-// Otherwise, it compiles away and does nothing.
-#define DVLOG(severity) ABSL_LOG_INTERNAL_DVLOG_IMPL(severity)
-
-// `LOG_IF` and friends add a second argument which specifies a condition.  If
-// the condition is false, nothing is logged.
-// Example:
-//
-//   LOG_IF(INFO, num_cookies > 10) << "Got lots of cookies";
-//
-// There is no `VLOG_IF` because the order of evaluation of the arguments is
-// ambiguous and the alternate spelling with an `if`-statement is trivial.
-#define LOG_IF(severity, condition) \
-  ABSL_LOG_INTERNAL_LOG_IF_IMPL(_##severity, condition)
-#define PLOG_IF(severity, condition) \
-  ABSL_LOG_INTERNAL_PLOG_IF_IMPL(_##severity, condition)
-#define DLOG_IF(severity, condition) \
-  ABSL_LOG_INTERNAL_DLOG_IF_IMPL(_##severity, condition)
-
-// LOG_EVERY_N
-//
-// An instance of `LOG_EVERY_N` increments a hidden zero-initialized counter
-// every time execution passes through it and logs the specified message when
-// the counter's value is a multiple of `n`, doing nothing otherwise.  Each
-// instance has its own counter.  The counter's value can be logged by streaming
-// the symbol `COUNTER`.  `LOG_EVERY_N` is thread-safe.
-// Example:
-//
-//   LOG_EVERY_N(WARNING, 1000) << "Got a packet with a bad CRC (" << COUNTER
-//                              << " total)";
-#define LOG_EVERY_N(severity, n) \
-  ABSL_LOG_INTERNAL_LOG_EVERY_N_IMPL(_##severity, n)
-
-// LOG_FIRST_N
-//
-// `LOG_FIRST_N` behaves like `LOG_EVERY_N` except that the specified message is
-// logged when the counter's value is less than `n`.  `LOG_FIRST_N` is
-// thread-safe.
-#define LOG_FIRST_N(severity, n) \
-  ABSL_LOG_INTERNAL_LOG_FIRST_N_IMPL(_##severity, n)
-
-// LOG_EVERY_POW_2
-//
-// `LOG_EVERY_POW_2` behaves like `LOG_EVERY_N` except that the specified
-// message is logged when the counter's value is a power of 2.
-// `LOG_EVERY_POW_2` is thread-safe.
-#define LOG_EVERY_POW_2(severity) \
-  ABSL_LOG_INTERNAL_LOG_EVERY_POW_2_IMPL(_##severity)
-
-// LOG_EVERY_N_SEC
-//
-// An instance of `LOG_EVERY_N_SEC` uses a hidden state variable to log the
-// specified message at most once every `n_seconds`.  A hidden counter of
-// executions (whether a message is logged or not) is also maintained and can be
-// logged by streaming the symbol `COUNTER`.  `LOG_EVERY_N_SEC` is thread-safe.
-// Example:
-//
-//   LOG_EVERY_N_SEC(INFO, 2.5) << "Got " << COUNTER << " cookies so far";
-#define LOG_EVERY_N_SEC(severity, n_seconds) \
-  ABSL_LOG_INTERNAL_LOG_EVERY_N_SEC_IMPL(_##severity, n_seconds)
-
-#define PLOG_EVERY_N(severity, n) \
-  ABSL_LOG_INTERNAL_PLOG_EVERY_N_IMPL(_##severity, n)
-#define PLOG_FIRST_N(severity, n) \
-  ABSL_LOG_INTERNAL_PLOG_FIRST_N_IMPL(_##severity, n)
-#define PLOG_EVERY_POW_2(severity) \
-  ABSL_LOG_INTERNAL_PLOG_EVERY_POW_2_IMPL(_##severity)
-#define PLOG_EVERY_N_SEC(severity, n_seconds) \
-  ABSL_LOG_INTERNAL_PLOG_EVERY_N_SEC_IMPL(_##severity, n_seconds)
-
-#define DLOG_EVERY_N(severity, n) \
-  ABSL_LOG_INTERNAL_DLOG_EVERY_N_IMPL(_##severity, n)
-#define DLOG_FIRST_N(severity, n) \
-  ABSL_LOG_INTERNAL_DLOG_FIRST_N_IMPL(_##severity, n)
-#define DLOG_EVERY_POW_2(severity) \
-  ABSL_LOG_INTERNAL_DLOG_EVERY_POW_2_IMPL(_##severity)
-#define DLOG_EVERY_N_SEC(severity, n_seconds) \
-  ABSL_LOG_INTERNAL_DLOG_EVERY_N_SEC_IMPL(_##severity, n_seconds)
-
-#define VLOG_EVERY_N(severity, n) \
-  ABSL_LOG_INTERNAL_VLOG_EVERY_N_IMPL(severity, n)
-#define VLOG_FIRST_N(severity, n) \
-  ABSL_LOG_INTERNAL_VLOG_FIRST_N_IMPL(severity, n)
-#define VLOG_EVERY_POW_2(severity) \
-  ABSL_LOG_INTERNAL_VLOG_EVERY_POW_2_IMPL(severity)
-#define VLOG_EVERY_N_SEC(severity, n_seconds) \
-  ABSL_LOG_INTERNAL_VLOG_EVERY_N_SEC_IMPL(severity, n_seconds)
-
-// `LOG_IF_EVERY_N` and friends behave as the corresponding `LOG_EVERY_N`
-// but neither increment a counter nor log a message if condition is false (as
-// `LOG_IF`).
-// Example:
-//
-//   LOG_IF_EVERY_N(INFO, (size > 1024), 10) << "Got the " << COUNTER
-//                                           << "th big cookie";
-#define LOG_IF_EVERY_N(severity, condition, n) \
-  ABSL_LOG_INTERNAL_LOG_IF_EVERY_N_IMPL(_##severity, condition, n)
-#define LOG_IF_FIRST_N(severity, condition, n) \
-  ABSL_LOG_INTERNAL_LOG_IF_FIRST_N_IMPL(_##severity, condition, n)
-#define LOG_IF_EVERY_POW_2(severity, condition) \
-  ABSL_LOG_INTERNAL_LOG_IF_EVERY_POW_2_IMPL(_##severity, condition)
-#define LOG_IF_EVERY_N_SEC(severity, condition, n_seconds) \
-  ABSL_LOG_INTERNAL_LOG_IF_EVERY_N_SEC_IMPL(_##severity, condition, n_seconds)
-
-#define PLOG_IF_EVERY_N(severity, condition, n) \
-  ABSL_LOG_INTERNAL_PLOG_IF_EVERY_N_IMPL(_##severity, condition, n)
-#define PLOG_IF_FIRST_N(severity, condition, n) \
-  ABSL_LOG_INTERNAL_PLOG_IF_FIRST_N_IMPL(_##severity, condition, n)
-#define PLOG_IF_EVERY_POW_2(severity, condition) \
-  ABSL_LOG_INTERNAL_PLOG_IF_EVERY_POW_2_IMPL(_##severity, condition)
-#define PLOG_IF_EVERY_N_SEC(severity, condition, n_seconds) \
-  ABSL_LOG_INTERNAL_PLOG_IF_EVERY_N_SEC_IMPL(_##severity, condition, n_seconds)
-
-#define DLOG_IF_EVERY_N(severity, condition, n) \
-  ABSL_LOG_INTERNAL_DLOG_IF_EVERY_N_IMPL(_##severity, condition, n)
-#define DLOG_IF_FIRST_N(severity, condition, n) \
-  ABSL_LOG_INTERNAL_DLOG_IF_FIRST_N_IMPL(_##severity, condition, n)
-#define DLOG_IF_EVERY_POW_2(severity, condition) \
-  ABSL_LOG_INTERNAL_DLOG_IF_EVERY_POW_2_IMPL(_##severity, condition)
-#define DLOG_IF_EVERY_N_SEC(severity, condition, n_seconds) \
-  ABSL_LOG_INTERNAL_DLOG_IF_EVERY_N_SEC_IMPL(_##severity, condition, n_seconds)
-
-#endif  // ABSL_LOG_LOG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/log_entry.h b/projects/llm_framework/include/abseil-cpp/absl/log/log_entry.h
deleted file mode 100644
index 7a55dfe2..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/log_entry.h
+++ /dev/null
@@ -1,221 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/log_entry.h
-// -----------------------------------------------------------------------------
-//
-// This header declares `class absl::LogEntry`, which represents a log record as
-// passed to `LogSink::Send`. Data returned by pointer or by reference or by
-// `absl::string_view` must be copied if they are needed after the lifetime of
-// the `absl::LogEntry`.
-
-#ifndef ABSL_LOG_LOG_ENTRY_H_
-#define ABSL_LOG_LOG_ENTRY_H_
-
-#include <cstddef>
-#include <string>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/log_severity.h"
-#include "absl/log/internal/config.h"
-#include "absl/strings/string_view.h"
-#include "absl/time/time.h"
-#include "absl/types/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-namespace log_internal {
-// Test only friend.
-class LogEntryTestPeer;
-class LogMessage;
-}  // namespace log_internal
-
-// LogEntry
-//
-// Represents a single entry in a log, i.e., one `LOG` statement or failed
-// `CHECK`.
-//
-// `LogEntry` is thread-compatible.
-class LogEntry final {
- public:
-  using tid_t = log_internal::Tid;
-
-  // For non-verbose log entries, `verbosity()` returns `kNoVerbosityLevel`.
-  static constexpr int kNoVerbosityLevel = -1;
-  static constexpr int kNoVerboseLevel = -1;  // TO BE removed
-
-  // Pass `LogEntry` by reference, and do not store it as its state does not
-  // outlive the call to `LogSink::Send()`.
-  LogEntry(const LogEntry&) = delete;
-  LogEntry& operator=(const LogEntry&) = delete;
-
-  // Source file and line where the log message occurred.  Taken from `__FILE__`
-  // and `__LINE__` unless overridden by `LOG(...).AtLocation(...)`.
-  //
-  // Take special care not to use the values returned by `source_filename()` and
-  // `source_basename()` after the lifetime of the entry.  This is always
-  // incorrect, but it will often work in practice because they usually point
-  // into a statically allocated character array obtained from `__FILE__`.
-  // Statements like `LOG(INFO).AtLocation(std::string(...), ...)` will expose
-  // the bug.  If you need the data later, you must copy them.
-  absl::string_view source_filename() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return full_filename_;
-  }
-  absl::string_view source_basename() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return base_filename_;
-  }
-  int source_line() const { return line_; }
-
-  // LogEntry::prefix()
-  //
-  // True unless the metadata prefix was suppressed once by
-  // `LOG(...).NoPrefix()` or globally by `absl::EnableLogPrefix(false)`.
-  // Implies `text_message_with_prefix() == text_message()`.
-  bool prefix() const { return prefix_; }
-
-  // LogEntry::log_severity()
-  //
-  // Returns this entry's severity.  For `LOG`, taken from the first argument;
-  // for `CHECK`, always `absl::LogSeverity::kFatal`.
-  absl::LogSeverity log_severity() const { return severity_; }
-
-  // LogEntry::verbosity()
-  //
-  // Returns this entry's verbosity, or `kNoVerbosityLevel` for a non-verbose
-  // entry. Taken from the argument to `VLOG` or from
-  // `LOG(...).WithVerbosity(...)`.
-  int verbosity() const { return verbose_level_; }
-
-  // LogEntry::timestamp()
-  //
-  // Returns the time at which this entry was written.  Captured during
-  // evaluation of `LOG`, but can be overridden by
-  // `LOG(...).WithTimestamp(...)`.
-  //
-  // Take care not to rely on timestamps increasing monotonically, or even to
-  // rely on timestamps having any particular relationship with reality (since
-  // they can be overridden).
-  absl::Time timestamp() const { return timestamp_; }
-
-  // LogEntry::tid()
-  //
-  // Returns the ID of the thread that wrote this entry.  Captured during
-  // evaluation of `LOG`, but can be overridden by `LOG(...).WithThreadID(...)`.
-  //
-  // Take care not to *rely* on reported thread IDs as they can be overridden as
-  // specified above.
-  tid_t tid() const { return tid_; }
-
-  // Text-formatted version of the log message.  An underlying buffer holds
-  // these contiguous data:
-  //
-  // * A prefix formed by formatting metadata (timestamp, filename, line number,
-  //   etc.)
-  //   The prefix may be empty - see `LogEntry::prefix()` - and may rarely be
-  //   truncated if the metadata are very long.
-  // * The streamed data
-  //   The data may be empty if nothing was streamed, or may be truncated to fit
-  //   the buffer.
-  // * A newline
-  // * A nul terminator
-  //
-  // The newline and nul terminator will be present even if the prefix and/or
-  // data are truncated.
-  //
-  // These methods give access to the most commonly useful substrings of the
-  // buffer's contents.  Other combinations can be obtained with substring
-  // arithmetic.
-  //
-  // The buffer does not outlive the entry; if you need the data later, you must
-  // copy them.
-  absl::string_view text_message_with_prefix_and_newline() const
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return absl::string_view(
-        text_message_with_prefix_and_newline_and_nul_.data(),
-        text_message_with_prefix_and_newline_and_nul_.size() - 1);
-  }
-  absl::string_view text_message_with_prefix() const
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return absl::string_view(
-        text_message_with_prefix_and_newline_and_nul_.data(),
-        text_message_with_prefix_and_newline_and_nul_.size() - 2);
-  }
-  absl::string_view text_message_with_newline() const
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return absl::string_view(
-        text_message_with_prefix_and_newline_and_nul_.data() + prefix_len_,
-        text_message_with_prefix_and_newline_and_nul_.size() - prefix_len_ - 1);
-  }
-  absl::string_view text_message() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return absl::string_view(
-        text_message_with_prefix_and_newline_and_nul_.data() + prefix_len_,
-        text_message_with_prefix_and_newline_and_nul_.size() - prefix_len_ - 2);
-  }
-  const char* text_message_with_prefix_and_newline_c_str() const
-      ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return text_message_with_prefix_and_newline_and_nul_.data();
-  }
-
-  // Returns a serialized protobuf holding the operands streamed into this
-  // log message.  The message definition is not yet published.
-  //
-  // The buffer does not outlive the entry; if you need the data later, you must
-  // copy them.
-  absl::string_view encoded_message() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return encoding_;
-  }
-
-  // LogEntry::stacktrace()
-  //
-  // Optional stacktrace, e.g. for `FATAL` logs and failed `CHECK`s.
-  //
-  // Fatal entries are dispatched to each sink twice: first with all data and
-  // metadata but no stacktrace, and then with the stacktrace.  This is done
-  // because stacktrace collection is sometimes slow and fallible, and it's
-  // critical to log enough information to diagnose the failure even if the
-  // stacktrace collection hangs.
-  //
-  // The buffer does not outlive the entry; if you need the data later, you must
-  // copy them.
-  absl::string_view stacktrace() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return stacktrace_;
-  }
-
- private:
-  LogEntry() = default;
-
-  absl::string_view full_filename_;
-  absl::string_view base_filename_;
-  int line_;
-  bool prefix_;
-  absl::LogSeverity severity_;
-  int verbose_level_;  // >=0 for `VLOG`, etc.; otherwise `kNoVerbosityLevel`.
-  absl::Time timestamp_;
-  tid_t tid_;
-  absl::Span<const char> text_message_with_prefix_and_newline_and_nul_;
-  size_t prefix_len_;
-  absl::string_view encoding_;
-  std::string stacktrace_;
-
-  friend class log_internal::LogEntryTestPeer;
-  friend class log_internal::LogMessage;
-};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_LOG_ENTRY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/log_sink.h b/projects/llm_framework/include/abseil-cpp/absl/log/log_sink.h
deleted file mode 100644
index 2910070d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/log_sink.h
+++ /dev/null
@@ -1,71 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/log_sink.h
-// -----------------------------------------------------------------------------
-//
-// This header declares the interface class `absl::LogSink`.
-
-#ifndef ABSL_LOG_LOG_SINK_H_
-#define ABSL_LOG_LOG_SINK_H_
-
-#include "absl/base/config.h"
-#include "absl/log/log_entry.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// absl::LogSink
-//
-// `absl::LogSink` is an interface which can be extended to intercept and
-// process particular messages (with `LOG.ToSinkOnly()` or
-// `LOG.ToSinkAlso()`) or all messages (if registered with
-// `absl::AddLogSink`).  Implementations must not take any locks that might be
-// held by the `LOG` caller.
-class LogSink {
- public:
-  virtual ~LogSink() = default;
-
-  // LogSink::Send()
-  //
-  // `Send` is called synchronously during the log statement.  `Send` must be
-  // thread-safe.
-  //
-  // It is safe to use `LOG` within an implementation of `Send`.  `ToSinkOnly`
-  // and `ToSinkAlso` are safe in general but can be used to create an infinite
-  // loop if you try.
-  virtual void Send(const absl::LogEntry& entry) = 0;
-
-  // LogSink::Flush()
-  //
-  // Sinks that buffer messages should override this method to flush the buffer
-  // and return.  `Flush` must be thread-safe.
-  virtual void Flush() {}
-
- protected:
-  LogSink() = default;
-  // Implementations may be copyable and/or movable.
-  LogSink(const LogSink&) = default;
-  LogSink& operator=(const LogSink&) = default;
-
- private:
-  // https://lld.llvm.org/missingkeyfunction.html#missing-key-function
-  virtual void KeyFunction() const final;  // NOLINT(readability/inheritance)
-};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_LOG_SINK_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/log_sink_registry.h b/projects/llm_framework/include/abseil-cpp/absl/log/log_sink_registry.h
deleted file mode 100644
index bf76ccee..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/log_sink_registry.h
+++ /dev/null
@@ -1,61 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/log_sink_registry.h
-// -----------------------------------------------------------------------------
-//
-// This header declares APIs to operate on global set of registered log sinks.
-
-#ifndef ABSL_LOG_LOG_SINK_REGISTRY_H_
-#define ABSL_LOG_LOG_SINK_REGISTRY_H_
-
-#include "absl/base/config.h"
-#include "absl/log/internal/log_sink_set.h"
-#include "absl/log/log_sink.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// AddLogSink(), RemoveLogSink()
-//
-// Adds or removes a `absl::LogSink` as a consumer of logging data.
-//
-// These functions are thread-safe.
-//
-// It is an error to attempt to add a sink that's already registered or to
-// attempt to remove one that isn't.
-//
-// To avoid unbounded recursion, dispatch to registered `absl::LogSink`s is
-// disabled per-thread while running the `Send()` method of registered
-// `absl::LogSink`s.  Affected messages are dispatched to a special internal
-// sink instead which writes them to `stderr`.
-//
-// Do not call these inside `absl::LogSink::Send`.
-inline void AddLogSink(absl::LogSink* sink) { log_internal::AddLogSink(sink); }
-inline void RemoveLogSink(absl::LogSink* sink) {
-  log_internal::RemoveLogSink(sink);
-}
-
-// FlushLogSinks()
-//
-// Calls `absl::LogSink::Flush` on all registered sinks.
-//
-// Do not call this inside `absl::LogSink::Send`.
-inline void FlushLogSinks() { log_internal::FlushLogSinks(); }
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_LOG_SINK_REGISTRY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/log_streamer.h b/projects/llm_framework/include/abseil-cpp/absl/log/log_streamer.h
deleted file mode 100644
index 4ed2435d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/log_streamer.h
+++ /dev/null
@@ -1,181 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/log_streamer.h
-// -----------------------------------------------------------------------------
-//
-// This header declares the class `LogStreamer` and convenience functions to
-// construct LogStreamer objects with different associated log severity levels.
-
-#ifndef ABSL_LOG_LOG_STREAMER_H_
-#define ABSL_LOG_LOG_STREAMER_H_
-
-#include <ios>
-#include <memory>
-#include <ostream>
-#include <string>
-#include <utility>
-
-#include "absl/base/config.h"
-#include "absl/base/log_severity.h"
-#include "absl/log/absl_log.h"
-#include "absl/strings/internal/ostringstream.h"
-#include "absl/strings/string_view.h"
-#include "absl/types/optional.h"
-#include "absl/utility/utility.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// LogStreamer
-//
-// Although you can stream into `LOG(INFO)`, you can't pass it into a function
-// that takes a `std::ostream` parameter. `LogStreamer::stream()` provides a
-// `std::ostream` that buffers everything that's streamed in.  The buffer's
-// contents are logged as if by `LOG` when the `LogStreamer` is destroyed.
-// If nothing is streamed in, an empty message is logged.  If the specified
-// severity is `absl::LogSeverity::kFatal`, the program will be terminated when
-// the `LogStreamer` is destroyed regardless of whether any data were streamed
-// in.
-//
-// Factory functions corresponding to the `absl::LogSeverity` enumerators
-// are provided for convenience; if the desired severity is variable, invoke the
-// constructor directly.
-//
-// LogStreamer is movable, but not copyable.
-//
-// Examples:
-//
-//   ShaveYakAndWriteToStream(
-//       yak, absl::LogInfoStreamer(__FILE__, __LINE__).stream());
-//
-//   {
-//     // This logs a single line containing data streamed by all three function
-//     // calls.
-//     absl::LogStreamer streamer(absl::LogSeverity::kInfo, __FILE__, __LINE__);
-//     ShaveYakAndWriteToStream(yak1, streamer.stream());
-//     streamer.stream() << " ";
-//     ShaveYakAndWriteToStream(yak2, streamer.stream());
-//     streamer.stream() << " ";
-//     ShaveYakAndWriteToStreamPointer(yak3, &streamer.stream());
-//   }
-class LogStreamer final {
- public:
-  // LogStreamer::LogStreamer()
-  //
-  // Creates a LogStreamer with a given `severity` that will log a message
-  // attributed to the given `file` and `line`.
-  explicit LogStreamer(absl::LogSeverity severity, absl::string_view file,
-                       int line)
-      : severity_(severity),
-        line_(line),
-        file_(file),
-        stream_(absl::in_place, &buf_) {
-    // To match `LOG`'s defaults:
-    stream_->setf(std::ios_base::showbase | std::ios_base::boolalpha);
-  }
-
-  // A moved-from `absl::LogStreamer` does not `LOG` when destroyed,
-  // and a program that streams into one has undefined behavior.
-  LogStreamer(LogStreamer&& that) noexcept
-      : severity_(that.severity_),
-        line_(that.line_),
-        file_(std::move(that.file_)),
-        buf_(std::move(that.buf_)),
-        stream_(std::move(that.stream_)) {
-    if (stream_.has_value()) stream_->str(&buf_);
-    that.stream_.reset();
-  }
-  LogStreamer& operator=(LogStreamer&& that) {
-    ABSL_LOG_IF(LEVEL(severity_), stream_).AtLocation(file_, line_) << buf_;
-    severity_ = that.severity_;
-    file_ = std::move(that.file_);
-    line_ = that.line_;
-    buf_ = std::move(that.buf_);
-    stream_ = std::move(that.stream_);
-    if (stream_.has_value()) stream_->str(&buf_);
-    that.stream_.reset();
-    return *this;
-  }
-
-  // LogStreamer::~LogStreamer()
-  //
-  // Logs this LogStreamer's buffered content as if by LOG.
-  ~LogStreamer() {
-    ABSL_LOG_IF(LEVEL(severity_), stream_.has_value()).AtLocation(file_, line_)
-        << buf_;
-  }
-
-  // LogStreamer::stream()
-  //
-  // Returns the `std::ostream` to use to write into this LogStreamer' internal
-  // buffer.
-  std::ostream& stream() { return *stream_; }
-
- private:
-  absl::LogSeverity severity_;
-  int line_;
-  std::string file_;
-  std::string buf_;
-  // A disengaged `stream_` indicates a moved-from `LogStreamer` that should not
-  // `LOG` upon destruction.
-  absl::optional<absl::strings_internal::OStringStream> stream_;
-};
-
-// LogInfoStreamer()
-//
-// Returns a LogStreamer that writes at level LogSeverity::kInfo.
-inline LogStreamer LogInfoStreamer(absl::string_view file, int line) {
-  return absl::LogStreamer(absl::LogSeverity::kInfo, file, line);
-}
-
-// LogWarningStreamer()
-//
-// Returns a LogStreamer that writes at level LogSeverity::kWarning.
-inline LogStreamer LogWarningStreamer(absl::string_view file, int line) {
-  return absl::LogStreamer(absl::LogSeverity::kWarning, file, line);
-}
-
-// LogErrorStreamer()
-//
-// Returns a LogStreamer that writes at level LogSeverity::kError.
-inline LogStreamer LogErrorStreamer(absl::string_view file, int line) {
-  return absl::LogStreamer(absl::LogSeverity::kError, file, line);
-}
-
-// LogFatalStreamer()
-//
-// Returns a LogStreamer that writes at level LogSeverity::kFatal.
-//
-// The program will be terminated when this `LogStreamer` is destroyed,
-// regardless of whether any data were streamed in.
-inline LogStreamer LogFatalStreamer(absl::string_view file, int line) {
-  return absl::LogStreamer(absl::LogSeverity::kFatal, file, line);
-}
-
-// LogDebugFatalStreamer()
-//
-// Returns a LogStreamer that writes at level LogSeverity::kLogDebugFatal.
-//
-// In debug mode, the program will be terminated when this `LogStreamer` is
-// destroyed, regardless of whether any data were streamed in.
-inline LogStreamer LogDebugFatalStreamer(absl::string_view file, int line) {
-  return absl::LogStreamer(absl::kLogDebugFatal, file, line);
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_LOG_STREAMER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/scoped_mock_log.h b/projects/llm_framework/include/abseil-cpp/absl/log/scoped_mock_log.h
deleted file mode 100644
index 399e604d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/scoped_mock_log.h
+++ /dev/null
@@ -1,197 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/scoped_mock_log.h
-// -----------------------------------------------------------------------------
-//
-// This header declares `class absl::ScopedMockLog`, for use in testing.
-
-#ifndef ABSL_LOG_SCOPED_MOCK_LOG_H_
-#define ABSL_LOG_SCOPED_MOCK_LOG_H_
-
-#include <atomic>
-#include <string>
-
-#include "gmock/gmock.h"
-#include "absl/base/config.h"
-#include "absl/base/log_severity.h"
-#include "absl/log/log_entry.h"
-#include "absl/log/log_sink.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// MockLogDefault
-//
-// Controls how ScopedMockLog responds to unexpected calls by default.
-enum class MockLogDefault { kIgnoreUnexpected, kDisallowUnexpected };
-
-// ScopedMockLog
-//
-// ScopedMockLog is a LogSink that intercepts LOG() messages issued during its
-// lifespan.
-//
-// Using this together with GoogleTest, it's easy to test how a piece of code
-// calls LOG(). The typical usage, noting the distinction between
-// "uninteresting" and "unexpected", looks like this:
-//
-//   using ::testing::_;
-//   using ::testing::AnyNumber;
-//   using ::testing::EndsWith;
-//   using ::testing::kDoNotCaptureLogsYet;
-//   using ::testing::Lt;
-//
-//   TEST(FooTest, LogsCorrectly) {
-//     // Simple robust setup, ignores unexpected logs.
-//     absl::ScopedMockLog log;
-//
-//     // We expect the WARNING "Something bad!" exactly twice.
-//     EXPECT_CALL(log, Log(absl::LogSeverity::kWarning, _, "Something bad!"))
-//         .Times(2);
-//
-//     // But we want no messages from foo.cc.
-//     EXPECT_CALL(log, Log(_, EndsWith("/foo.cc"), _)).Times(0);
-//
-//     log.StartCapturingLogs();  // Call this after done setting expectations.
-//     Foo();  // Exercises the code under test.
-//   }
-//
-//   TEST(BarTest, LogsExactlyCorrectly) {
-//     // Strict checking, fails for unexpected logs.
-//     absl::ScopedMockLog log(absl::MockLogDefault::kDisallowUnexpected);
-//
-//     // ... but ignore low severity messages
-//     EXPECT_CALL(log, Log(Lt(absl::LogSeverity::kWarning), _, _))
-//         .Times(AnyNumber());
-//
-//     // We expect the ERROR "Something bad!" exactly once.
-//     EXPECT_CALL(log, Log(absl::LogSeverity::kError, EndsWith("/foo.cc"),
-//                 "Something bad!"))
-//         .Times(1);
-//
-//     log.StartCapturingLogs();  // Call this after done setting expectations.
-//     Bar();  // Exercises the code under test.
-//    }
-//
-// Note that in a multi-threaded environment, all LOG() messages from a single
-// thread will be handled in sequence, but that cannot be guaranteed for
-// messages from different threads. In fact, if the same or multiple
-// expectations are matched on two threads concurrently, their actions will be
-// executed concurrently as well and may interleave.
-class ScopedMockLog final {
- public:
-  // ScopedMockLog::ScopedMockLog()
-  //
-  // Sets up the log and adds default expectations.
-  explicit ScopedMockLog(
-      MockLogDefault default_exp = MockLogDefault::kIgnoreUnexpected);
-  ScopedMockLog(const ScopedMockLog&) = delete;
-  ScopedMockLog& operator=(const ScopedMockLog&) = delete;
-
-  // ScopedMockLog::~ScopedMockLog()
-  //
-  // Stops intercepting logs and destroys this ScopedMockLog.
-  ~ScopedMockLog();
-
-  // ScopedMockLog::StartCapturingLogs()
-  //
-  // Starts log capturing if the object isn't already doing so. Otherwise
-  // crashes.
-  //
-  // Usually this method is called in the same thread that created this
-  // ScopedMockLog. It is the user's responsibility to not call this method if
-  // another thread may be calling it or StopCapturingLogs() at the same time.
-  // It is undefined behavior to add expectations while capturing logs is
-  // enabled.
-  void StartCapturingLogs();
-
-  // ScopedMockLog::StopCapturingLogs()
-  //
-  // Stops log capturing if the object is capturing logs. Otherwise crashes.
-  //
-  // Usually this method is called in the same thread that created this object.
-  // It is the user's responsibility to not call this method if another thread
-  // may be calling it or StartCapturingLogs() at the same time.
-  //
-  // It is UB to add expectations, while capturing logs is enabled.
-  void StopCapturingLogs();
-
-  // ScopedMockLog::UseAsLocalSink()
-  //
-  // Each `ScopedMockLog` is implemented with an `absl::LogSink`; this method
-  // returns a reference to that sink (e.g. for use with
-  // `LOG(...).ToSinkOnly()`) and marks the `ScopedMockLog` as having been used
-  // even if `StartCapturingLogs` is never called.
-  absl::LogSink& UseAsLocalSink();
-
-  // Implements the mock method:
-  //
-  //   void Log(LogSeverity severity, absl::string_view file_path,
-  //            absl::string_view message);
-  //
-  // The second argument to Log() is the full path of the source file in
-  // which the LOG() was issued.
-  //
-  // This is a shorthand form, which should be used by most users. Use the
-  // `Send` mock only if you want to add expectations for other log message
-  // attributes.
-  MOCK_METHOD(void, Log,
-              (absl::LogSeverity severity, const std::string& file_path,
-               const std::string& message));
-
-  // Implements the mock method:
-  //
-  //   void Send(const absl::LogEntry& entry);
-  //
-  // This is the most generic form of mock that can be specified. Use this mock
-  // only if you want to add expectations for log message attributes different
-  // from the log message text, log message path and log message severity.
-  //
-  // If no expectations are specified for this mock, the default action is to
-  // forward the call to the `Log` mock.
-  MOCK_METHOD(void, Send, (const absl::LogEntry&));
-
-  // Implements the mock method:
-  //
-  //   void Flush();
-  //
-  // Use this mock only if you want to add expectations for log flush calls.
-  MOCK_METHOD(void, Flush, ());
-
- private:
-  class ForwardingSink final : public absl::LogSink {
-   public:
-    explicit ForwardingSink(ScopedMockLog* sml) : sml_(sml) {}
-    ForwardingSink(const ForwardingSink&) = delete;
-    ForwardingSink& operator=(const ForwardingSink&) = delete;
-    void Send(const absl::LogEntry& entry) override { sml_->Send(entry); }
-    void Flush() override { sml_->Flush(); }
-
-   private:
-    ScopedMockLog* sml_;
-  };
-
-  ForwardingSink sink_;
-  bool is_capturing_logs_;
-  // Until C++20, the default constructor leaves the underlying value wrapped in
-  // std::atomic uninitialized, so all constructors should be sure to initialize
-  // is_triggered_.
-  std::atomic<bool> is_triggered_;
-};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_SCOPED_MOCK_LOG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/structured.h b/projects/llm_framework/include/abseil-cpp/absl/log/structured.h
deleted file mode 100644
index 9ad69fbd..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/structured.h
+++ /dev/null
@@ -1,70 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/structured.h
-// -----------------------------------------------------------------------------
-//
-// This header declares APIs supporting structured logging, allowing log
-// statements to be more easily parsed, especially by automated processes.
-//
-// When structured logging is in use, data streamed into a `LOG` statement are
-// encoded as `Value` fields in a `logging.proto.Event` protocol buffer message.
-// The individual data are exposed programmatically to `LogSink`s and to the
-// user via some log reading tools which are able to query the structured data
-// more usefully than would be possible if each message was a single opaque
-// string.  These helpers allow user code to add additional structure to the
-// data they stream.
-
-#ifndef ABSL_LOG_STRUCTURED_H_
-#define ABSL_LOG_STRUCTURED_H_
-
-#include <ostream>
-
-#include "absl/base/config.h"
-#include "absl/log/internal/structured.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// LogAsLiteral()
-//
-// Annotates its argument as a string literal so that structured logging
-// captures it as a `literal` field instead of a `str` field (the default).
-// This does not affect the text representation, only the structure.
-//
-// Streaming `LogAsLiteral(s)` into a `std::ostream` behaves just like streaming
-// `s` directly.
-//
-// Using `LogAsLiteral()` is occasionally appropriate and useful when proxying
-// data logged from another system or another language.  For example:
-//
-//   void Logger::LogString(absl::string_view str, absl::LogSeverity severity,
-//                          const char *file, int line) {
-//     LOG(LEVEL(severity)).AtLocation(file, line) << str;
-//   }
-//   void Logger::LogStringLiteral(absl::string_view str,
-//                                 absl::LogSeverity severity, const char *file,
-//                                 int line) {
-//     LOG(LEVEL(severity)).AtLocation(file, line) << absl::LogAsLiteral(str);
-//   }
-inline log_internal::AsLiteralImpl LogAsLiteral(absl::string_view s) {
-  return log_internal::AsLiteralImpl(s);
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_LOG_STRUCTURED_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/log/vlog_is_on.h b/projects/llm_framework/include/abseil-cpp/absl/log/vlog_is_on.h
deleted file mode 100644
index 78986513..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/log/vlog_is_on.h
+++ /dev/null
@@ -1,72 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: log/vlog_is_on.h
-// -----------------------------------------------------------------------------
-//
-// This header defines the `VLOG_IS_ON()` macro that controls the
-// variable-verbosity conditional logging.
-//
-// It's used by `VLOG` in log.h, or it can also be used directly like this:
-//
-//   if (VLOG_IS_ON(2)) {
-//     foo_server.RecomputeStatisticsExpensive();
-//     LOG(INFO) << foo_server.LastStatisticsAsString();
-//   }
-//
-// Each source file has an effective verbosity level that's a non-negative
-// integer computed from the `--vmodule` and `--v` flags.
-// `VLOG_IS_ON(n)` is true, and `VLOG(n)` logs, if that effective verbosity
-// level is greater than or equal to `n`.
-//
-// `--vmodule` takes a comma-delimited list of key=value pairs.  Each key is a
-// pattern matched against filenames, and the values give the effective severity
-// level applied to matching files.  '?' and '*' characters in patterns are
-// interpreted as single-character and zero-or-more-character wildcards.
-// Patterns including a slash character are matched against full pathnames,
-// while those without are matched against basenames only.  One suffix (i.e. the
-// last . and everything after it) is stripped from each filename prior to
-// matching, as is the special suffix "-inl".
-//
-// Files are matched against globs in `--vmodule` in order, and the first match
-// determines the verbosity level.
-//
-// Files which do not match any pattern in `--vmodule` use the value of `--v` as
-// their effective verbosity level.  The default is 0.
-//
-// SetVLOGLevel helper function is provided to do limited dynamic control over
-// V-logging by appending to `--vmodule`. Because these go at the beginning of
-// the list, they take priority over any globs previously added.
-//
-// Resetting --vmodule will override all previous modifications to `--vmodule`,
-// including via SetVLOGLevel.
-
-#ifndef ABSL_LOG_VLOG_IS_ON_H_
-#define ABSL_LOG_VLOG_IS_ON_H_
-
-#include "absl/log/absl_vlog_is_on.h"  // IWYU pragma: export
-
-// IWYU pragma: private, include "absl/log/log.h"
-
-// Each VLOG_IS_ON call site gets its own VLogSite that registers with the
-// global linked list of sites to asynchronously update its verbosity level on
-// changes to --v or --vmodule. The verbosity can also be set by manually
-// calling SetVLOGLevel.
-//
-// VLOG_IS_ON is not async signal safe, but it is guaranteed not to allocate
-// new memory.
-#define VLOG_IS_ON(verbose_level) ABSL_VLOG_IS_ON(verbose_level)
-
-#endif  // ABSL_LOG_VLOG_IS_ON_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/memory/memory.h b/projects/llm_framework/include/abseil-cpp/absl/memory/memory.h
deleted file mode 100644
index 3508135c..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/memory/memory.h
+++ /dev/null
@@ -1,278 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: memory.h
-// -----------------------------------------------------------------------------
-//
-// This header file contains utility functions for managing the creation and
-// conversion of smart pointers. This file is an extension to the C++
-// standard <memory> library header file.
-
-#ifndef ABSL_MEMORY_MEMORY_H_
-#define ABSL_MEMORY_MEMORY_H_
-
-#include <cstddef>
-#include <limits>
-#include <memory>
-#include <new>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/macros.h"
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// -----------------------------------------------------------------------------
-// Function Template: WrapUnique()
-// -----------------------------------------------------------------------------
-//
-// Adopts ownership from a raw pointer and transfers it to the returned
-// `std::unique_ptr`, whose type is deduced. Because of this deduction, *do not*
-// specify the template type `T` when calling `WrapUnique`.
-//
-// Example:
-//   X* NewX(int, int);
-//   auto x = WrapUnique(NewX(1, 2));  // 'x' is std::unique_ptr<X>.
-//
-// Do not call WrapUnique with an explicit type, as in
-// `WrapUnique<X>(NewX(1, 2))`.  The purpose of WrapUnique is to automatically
-// deduce the pointer type. If you wish to make the type explicit, just use
-// `std::unique_ptr` directly.
-//
-//   auto x = std::unique_ptr<X>(NewX(1, 2));
-//                  - or -
-//   std::unique_ptr<X> x(NewX(1, 2));
-//
-// While `absl::WrapUnique` is useful for capturing the output of a raw
-// pointer factory, prefer 'absl::make_unique<T>(args...)' over
-// 'absl::WrapUnique(new T(args...))'.
-//
-//   auto x = WrapUnique(new X(1, 2));  // works, but nonideal.
-//   auto x = make_unique<X>(1, 2);     // safer, standard, avoids raw 'new'.
-//
-// Note that `absl::WrapUnique(p)` is valid only if `delete p` is a valid
-// expression. In particular, `absl::WrapUnique()` cannot wrap pointers to
-// arrays, functions or void, and it must not be used to capture pointers
-// obtained from array-new expressions (even though that would compile!).
-template <typename T>
-std::unique_ptr<T> WrapUnique(T* ptr) {
-  static_assert(!std::is_array<T>::value, "array types are unsupported");
-  static_assert(std::is_object<T>::value, "non-object types are unsupported");
-  return std::unique_ptr<T>(ptr);
-}
-
-// -----------------------------------------------------------------------------
-// Function Template: make_unique<T>()
-// -----------------------------------------------------------------------------
-//
-// Creates a `std::unique_ptr<>`, while avoiding issues creating temporaries
-// during the construction process. `absl::make_unique<>` also avoids redundant
-// type declarations, by avoiding the need to explicitly use the `new` operator.
-//
-// https://en.cppreference.com/w/cpp/memory/unique_ptr/make_unique
-//
-// For more background on why `std::unique_ptr<T>(new T(a,b))` is problematic,
-// see Herb Sutter's explanation on
-// (Exception-Safe Function Calls)[https://herbsutter.com/gotw/_102/].
-// (In general, reviewers should treat `new T(a,b)` with scrutiny.)
-//
-// Historical note: Abseil once provided a C++11 compatible implementation of
-// the C++14's `std::make_unique`. Now that C++11 support has been sunsetted,
-// `absl::make_unique` simply uses the STL-provided implementation. New code
-// should use `std::make_unique`.
-using std::make_unique;
-
-// -----------------------------------------------------------------------------
-// Function Template: RawPtr()
-// -----------------------------------------------------------------------------
-//
-// Extracts the raw pointer from a pointer-like value `ptr`. `absl::RawPtr` is
-// useful within templates that need to handle a complement of raw pointers,
-// `std::nullptr_t`, and smart pointers.
-template <typename T>
-auto RawPtr(T&& ptr) -> decltype(std::addressof(*ptr)) {
-  // ptr is a forwarding reference to support Ts with non-const operators.
-  return (ptr != nullptr) ? std::addressof(*ptr) : nullptr;
-}
-inline std::nullptr_t RawPtr(std::nullptr_t) { return nullptr; }
-
-// -----------------------------------------------------------------------------
-// Function Template: ShareUniquePtr()
-// -----------------------------------------------------------------------------
-//
-// Adopts a `std::unique_ptr` rvalue and returns a `std::shared_ptr` of deduced
-// type. Ownership (if any) of the held value is transferred to the returned
-// shared pointer.
-//
-// Example:
-//
-//     auto up = absl::make_unique<int>(10);
-//     auto sp = absl::ShareUniquePtr(std::move(up));  // shared_ptr<int>
-//     CHECK_EQ(*sp, 10);
-//     CHECK(up == nullptr);
-//
-// Note that this conversion is correct even when T is an array type, and more
-// generally it works for *any* deleter of the `unique_ptr` (single-object
-// deleter, array deleter, or any custom deleter), since the deleter is adopted
-// by the shared pointer as well. The deleter is copied (unless it is a
-// reference).
-//
-// Implements the resolution of [LWG 2415](http://wg21.link/lwg2415), by which a
-// null shared pointer does not attempt to call the deleter.
-template <typename T, typename D>
-std::shared_ptr<T> ShareUniquePtr(std::unique_ptr<T, D>&& ptr) {
-  return ptr ? std::shared_ptr<T>(std::move(ptr)) : std::shared_ptr<T>();
-}
-
-// -----------------------------------------------------------------------------
-// Function Template: WeakenPtr()
-// -----------------------------------------------------------------------------
-//
-// Creates a weak pointer associated with a given shared pointer. The returned
-// value is a `std::weak_ptr` of deduced type.
-//
-// Example:
-//
-//    auto sp = std::make_shared<int>(10);
-//    auto wp = absl::WeakenPtr(sp);
-//    CHECK_EQ(sp.get(), wp.lock().get());
-//    sp.reset();
-//    CHECK(wp.lock() == nullptr);
-//
-template <typename T>
-std::weak_ptr<T> WeakenPtr(const std::shared_ptr<T>& ptr) {
-  return std::weak_ptr<T>(ptr);
-}
-
-// -----------------------------------------------------------------------------
-// Class Template: pointer_traits
-// -----------------------------------------------------------------------------
-//
-// Historical note: Abseil once provided an implementation of
-// `std::pointer_traits` for platforms that had not yet provided it. Those
-// platforms are no longer supported. New code should simply use
-// `std::pointer_traits`.
-using std::pointer_traits;
-
-// -----------------------------------------------------------------------------
-// Class Template: allocator_traits
-// -----------------------------------------------------------------------------
-//
-// Historical note: Abseil once provided an implementation of
-// `std::allocator_traits` for platforms that had not yet provided it. Those
-// platforms are no longer supported. New code should simply use
-// `std::allocator_traits`.
-using std::allocator_traits;
-
-namespace memory_internal {
-
-// ExtractOr<E, O, D>::type evaluates to E<O> if possible. Otherwise, D.
-template <template <typename> class Extract, typename Obj, typename Default,
-          typename>
-struct ExtractOr {
-  using type = Default;
-};
-
-template <template <typename> class Extract, typename Obj, typename Default>
-struct ExtractOr<Extract, Obj, Default, void_t<Extract<Obj>>> {
-  using type = Extract<Obj>;
-};
-
-template <template <typename> class Extract, typename Obj, typename Default>
-using ExtractOrT = typename ExtractOr<Extract, Obj, Default, void>::type;
-
-// This template alias transforms Alloc::is_nothrow into a metafunction with
-// Alloc as a parameter so it can be used with ExtractOrT<>.
-template <typename Alloc>
-using GetIsNothrow = typename Alloc::is_nothrow;
-
-}  // namespace memory_internal
-
-// ABSL_ALLOCATOR_NOTHROW is a build time configuration macro for user to
-// specify whether the default allocation function can throw or never throws.
-// If the allocation function never throws, user should define it to a non-zero
-// value (e.g. via `-DABSL_ALLOCATOR_NOTHROW`).
-// If the allocation function can throw, user should leave it undefined or
-// define it to zero.
-//
-// allocator_is_nothrow<Alloc> is a traits class that derives from
-// Alloc::is_nothrow if present, otherwise std::false_type. It's specialized
-// for Alloc = std::allocator<T> for any type T according to the state of
-// ABSL_ALLOCATOR_NOTHROW.
-//
-// default_allocator_is_nothrow is a class that derives from std::true_type
-// when the default allocator (global operator new) never throws, and
-// std::false_type when it can throw. It is a convenience shorthand for writing
-// allocator_is_nothrow<std::allocator<T>> (T can be any type).
-// NOTE: allocator_is_nothrow<std::allocator<T>> is guaranteed to derive from
-// the same type for all T, because users should specialize neither
-// allocator_is_nothrow nor std::allocator.
-template <typename Alloc>
-struct allocator_is_nothrow
-    : memory_internal::ExtractOrT<memory_internal::GetIsNothrow, Alloc,
-                                  std::false_type> {};
-
-#if defined(ABSL_ALLOCATOR_NOTHROW) && ABSL_ALLOCATOR_NOTHROW
-template <typename T>
-struct allocator_is_nothrow<std::allocator<T>> : std::true_type {};
-struct default_allocator_is_nothrow : std::true_type {};
-#else
-struct default_allocator_is_nothrow : std::false_type {};
-#endif
-
-namespace memory_internal {
-template <typename Allocator, typename Iterator, typename... Args>
-void ConstructRange(Allocator& alloc, Iterator first, Iterator last,
-                    const Args&... args) {
-  for (Iterator cur = first; cur != last; ++cur) {
-    ABSL_INTERNAL_TRY {
-      std::allocator_traits<Allocator>::construct(alloc, std::addressof(*cur),
-                                                  args...);
-    }
-    ABSL_INTERNAL_CATCH_ANY {
-      while (cur != first) {
-        --cur;
-        std::allocator_traits<Allocator>::destroy(alloc, std::addressof(*cur));
-      }
-      ABSL_INTERNAL_RETHROW;
-    }
-  }
-}
-
-template <typename Allocator, typename Iterator, typename InputIterator>
-void CopyRange(Allocator& alloc, Iterator destination, InputIterator first,
-               InputIterator last) {
-  for (Iterator cur = destination; first != last;
-       static_cast<void>(++cur), static_cast<void>(++first)) {
-    ABSL_INTERNAL_TRY {
-      std::allocator_traits<Allocator>::construct(alloc, std::addressof(*cur),
-                                                  *first);
-    }
-    ABSL_INTERNAL_CATCH_ANY {
-      while (cur != destination) {
-        --cur;
-        std::allocator_traits<Allocator>::destroy(alloc, std::addressof(*cur));
-      }
-      ABSL_INTERNAL_RETHROW;
-    }
-  }
-}
-}  // namespace memory_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_MEMORY_MEMORY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/meta/type_traits.h b/projects/llm_framework/include/abseil-cpp/absl/meta/type_traits.h
deleted file mode 100644
index ed5e6080..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/meta/type_traits.h
+++ /dev/null
@@ -1,580 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// type_traits.h
-// -----------------------------------------------------------------------------
-//
-// This file contains C++11-compatible versions of standard <type_traits> API
-// functions for determining the characteristics of types. Such traits can
-// support type inference, classification, and transformation, as well as
-// make it easier to write templates based on generic type behavior.
-//
-// See https://en.cppreference.com/w/cpp/header/type_traits
-//
-// WARNING: use of many of the constructs in this header will count as "complex
-// template metaprogramming", so before proceeding, please carefully consider
-// https://google.github.io/styleguide/cppguide.html#Template_metaprogramming
-//
-// WARNING: using template metaprogramming to detect or depend on API
-// features is brittle and not guaranteed. Neither the standard library nor
-// Abseil provides any guarantee that APIs are stable in the face of template
-// metaprogramming. Use with caution.
-#ifndef ABSL_META_TYPE_TRAITS_H_
-#define ABSL_META_TYPE_TRAITS_H_
-
-#include <cstddef>
-#include <functional>
-#include <type_traits>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-
-// Defines the default alignment. `__STDCPP_DEFAULT_NEW_ALIGNMENT__` is a C++17
-// feature.
-#if defined(__STDCPP_DEFAULT_NEW_ALIGNMENT__)
-#define ABSL_INTERNAL_DEFAULT_NEW_ALIGNMENT __STDCPP_DEFAULT_NEW_ALIGNMENT__
-#else  // defined(__STDCPP_DEFAULT_NEW_ALIGNMENT__)
-#define ABSL_INTERNAL_DEFAULT_NEW_ALIGNMENT alignof(std::max_align_t)
-#endif  // defined(__STDCPP_DEFAULT_NEW_ALIGNMENT__)
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-namespace type_traits_internal {
-
-template <typename... Ts>
-struct VoidTImpl {
-  using type = void;
-};
-
-////////////////////////////////
-// Library Fundamentals V2 TS //
-////////////////////////////////
-
-// NOTE: The `is_detected` family of templates here differ from the library
-// fundamentals specification in that for library fundamentals, `Op<Args...>` is
-// evaluated as soon as the type `is_detected<Op, Args...>` undergoes
-// substitution, regardless of whether or not the `::value` is accessed. That
-// is inconsistent with all other standard traits and prevents lazy evaluation
-// in larger contexts (such as if the `is_detected` check is a trailing argument
-// of a `conjunction`. This implementation opts to instead be lazy in the same
-// way that the standard traits are (this "defect" of the detection idiom
-// specifications has been reported).
-
-template <class Enabler, template <class...> class Op, class... Args>
-struct is_detected_impl {
-  using type = std::false_type;
-};
-
-template <template <class...> class Op, class... Args>
-struct is_detected_impl<typename VoidTImpl<Op<Args...>>::type, Op, Args...> {
-  using type = std::true_type;
-};
-
-template <template <class...> class Op, class... Args>
-struct is_detected : is_detected_impl<void, Op, Args...>::type {};
-
-template <class Enabler, class To, template <class...> class Op, class... Args>
-struct is_detected_convertible_impl {
-  using type = std::false_type;
-};
-
-template <class To, template <class...> class Op, class... Args>
-struct is_detected_convertible_impl<
-    typename std::enable_if<std::is_convertible<Op<Args...>, To>::value>::type,
-    To, Op, Args...> {
-  using type = std::true_type;
-};
-
-template <class To, template <class...> class Op, class... Args>
-struct is_detected_convertible
-    : is_detected_convertible_impl<void, To, Op, Args...>::type {};
-
-}  // namespace type_traits_internal
-
-// void_t()
-//
-// Ignores the type of any its arguments and returns `void`. In general, this
-// metafunction allows you to create a general case that maps to `void` while
-// allowing specializations that map to specific types.
-//
-// This metafunction is designed to be a drop-in replacement for the C++17
-// `std::void_t` metafunction.
-//
-// NOTE: `absl::void_t` does not use the standard-specified implementation so
-// that it can remain compatible with gcc < 5.1. This can introduce slightly
-// different behavior, such as when ordering partial specializations.
-template <typename... Ts>
-using void_t = typename type_traits_internal::VoidTImpl<Ts...>::type;
-
-// conjunction
-//
-// Performs a compile-time logical AND operation on the passed types (which
-// must have  `::value` members convertible to `bool`. Short-circuits if it
-// encounters any `false` members (and does not compare the `::value` members
-// of any remaining arguments).
-//
-// This metafunction is designed to be a drop-in replacement for the C++17
-// `std::conjunction` metafunction.
-template <typename... Ts>
-struct conjunction : std::true_type {};
-
-template <typename T, typename... Ts>
-struct conjunction<T, Ts...>
-    : std::conditional<T::value, conjunction<Ts...>, T>::type {};
-
-template <typename T>
-struct conjunction<T> : T {};
-
-// disjunction
-//
-// Performs a compile-time logical OR operation on the passed types (which
-// must have  `::value` members convertible to `bool`. Short-circuits if it
-// encounters any `true` members (and does not compare the `::value` members
-// of any remaining arguments).
-//
-// This metafunction is designed to be a drop-in replacement for the C++17
-// `std::disjunction` metafunction.
-template <typename... Ts>
-struct disjunction : std::false_type {};
-
-template <typename T, typename... Ts>
-struct disjunction<T, Ts...>
-    : std::conditional<T::value, T, disjunction<Ts...>>::type {};
-
-template <typename T>
-struct disjunction<T> : T {};
-
-// negation
-//
-// Performs a compile-time logical NOT operation on the passed type (which
-// must have  `::value` members convertible to `bool`.
-//
-// This metafunction is designed to be a drop-in replacement for the C++17
-// `std::negation` metafunction.
-template <typename T>
-struct negation : std::integral_constant<bool, !T::value> {};
-
-// is_function()
-//
-// Determines whether the passed type `T` is a function type.
-//
-// This metafunction is designed to be a drop-in replacement for the C++11
-// `std::is_function()` metafunction for platforms that have incomplete C++11
-// support (such as libstdc++ 4.x).
-//
-// This metafunction works because appending `const` to a type does nothing to
-// function types and reference types (and forms a const-qualified type
-// otherwise).
-template <typename T>
-struct is_function
-    : std::integral_constant<
-          bool, !(std::is_reference<T>::value ||
-                  std::is_const<typename std::add_const<T>::type>::value)> {};
-
-// is_copy_assignable()
-// is_move_assignable()
-// is_trivially_destructible()
-// is_trivially_default_constructible()
-// is_trivially_move_constructible()
-// is_trivially_copy_constructible()
-// is_trivially_move_assignable()
-// is_trivially_copy_assignable()
-//
-// Historical note: Abseil once provided implementations of these type traits
-// for platforms that lacked full support. New code should prefer to use the
-// std variants.
-//
-// See the documentation for the STL <type_traits> header for more information:
-// https://en.cppreference.com/w/cpp/header/type_traits
-using std::is_copy_assignable;
-using std::is_move_assignable;
-using std::is_trivially_copy_assignable;
-using std::is_trivially_copy_constructible;
-using std::is_trivially_default_constructible;
-using std::is_trivially_destructible;
-using std::is_trivially_move_assignable;
-using std::is_trivially_move_constructible;
-
-#if defined(__cpp_lib_remove_cvref) && __cpp_lib_remove_cvref >= 201711L
-template <typename T>
-using remove_cvref = std::remove_cvref<T>;
-
-template <typename T>
-using remove_cvref_t = typename std::remove_cvref<T>::type;
-#else
-// remove_cvref()
-//
-// C++11 compatible implementation of std::remove_cvref which was added in
-// C++20.
-template <typename T>
-struct remove_cvref {
-  using type =
-      typename std::remove_cv<typename std::remove_reference<T>::type>::type;
-};
-
-template <typename T>
-using remove_cvref_t = typename remove_cvref<T>::type;
-#endif
-
-// -----------------------------------------------------------------------------
-// C++14 "_t" trait aliases
-// -----------------------------------------------------------------------------
-
-template <typename T>
-using remove_cv_t = typename std::remove_cv<T>::type;
-
-template <typename T>
-using remove_const_t = typename std::remove_const<T>::type;
-
-template <typename T>
-using remove_volatile_t = typename std::remove_volatile<T>::type;
-
-template <typename T>
-using add_cv_t = typename std::add_cv<T>::type;
-
-template <typename T>
-using add_const_t = typename std::add_const<T>::type;
-
-template <typename T>
-using add_volatile_t = typename std::add_volatile<T>::type;
-
-template <typename T>
-using remove_reference_t = typename std::remove_reference<T>::type;
-
-template <typename T>
-using add_lvalue_reference_t = typename std::add_lvalue_reference<T>::type;
-
-template <typename T>
-using add_rvalue_reference_t = typename std::add_rvalue_reference<T>::type;
-
-template <typename T>
-using remove_pointer_t = typename std::remove_pointer<T>::type;
-
-template <typename T>
-using add_pointer_t = typename std::add_pointer<T>::type;
-
-template <typename T>
-using make_signed_t = typename std::make_signed<T>::type;
-
-template <typename T>
-using make_unsigned_t = typename std::make_unsigned<T>::type;
-
-template <typename T>
-using remove_extent_t = typename std::remove_extent<T>::type;
-
-template <typename T>
-using remove_all_extents_t = typename std::remove_all_extents<T>::type;
-
-ABSL_INTERNAL_DISABLE_DEPRECATED_DECLARATION_WARNING
-namespace type_traits_internal {
-// This trick to retrieve a default alignment is necessary for our
-// implementation of aligned_storage_t to be consistent with any
-// implementation of std::aligned_storage.
-template <size_t Len, typename T = std::aligned_storage<Len>>
-struct default_alignment_of_aligned_storage;
-
-template <size_t Len, size_t Align>
-struct default_alignment_of_aligned_storage<
-    Len, std::aligned_storage<Len, Align>> {
-  static constexpr size_t value = Align;
-};
-}  // namespace type_traits_internal
-
-// TODO(b/260219225): std::aligned_storage(_t) is deprecated in C++23.
-template <size_t Len, size_t Align = type_traits_internal::
-                          default_alignment_of_aligned_storage<Len>::value>
-using aligned_storage_t = typename std::aligned_storage<Len, Align>::type;
-ABSL_INTERNAL_RESTORE_DEPRECATED_DECLARATION_WARNING
-
-template <typename T>
-using decay_t = typename std::decay<T>::type;
-
-template <bool B, typename T = void>
-using enable_if_t = typename std::enable_if<B, T>::type;
-
-template <bool B, typename T, typename F>
-using conditional_t = typename std::conditional<B, T, F>::type;
-
-template <typename... T>
-using common_type_t = typename std::common_type<T...>::type;
-
-template <typename T>
-using underlying_type_t = typename std::underlying_type<T>::type;
-
-namespace type_traits_internal {
-
-#if (defined(__cpp_lib_is_invocable) && __cpp_lib_is_invocable >= 201703L) || \
-    (defined(_MSVC_LANG) && _MSVC_LANG >= 201703L)
-// std::result_of is deprecated (C++17) or removed (C++20)
-template <typename>
-struct result_of;
-template <typename F, typename... Args>
-struct result_of<F(Args...)> : std::invoke_result<F, Args...> {};
-#else
-template <typename F>
-using result_of = std::result_of<F>;
-#endif
-
-}  // namespace type_traits_internal
-
-template <typename F>
-using result_of_t = typename type_traits_internal::result_of<F>::type;
-
-namespace type_traits_internal {
-// In MSVC we can't probe std::hash or stdext::hash because it triggers a
-// static_assert instead of failing substitution. Libc++ prior to 4.0
-// also used a static_assert.
-//
-#if defined(_MSC_VER) || (defined(_LIBCPP_VERSION) && \
-                          _LIBCPP_VERSION < 4000 && _LIBCPP_STD_VER > 11)
-#define ABSL_META_INTERNAL_STD_HASH_SFINAE_FRIENDLY_ 0
-#else
-#define ABSL_META_INTERNAL_STD_HASH_SFINAE_FRIENDLY_ 1
-#endif
-
-#if !ABSL_META_INTERNAL_STD_HASH_SFINAE_FRIENDLY_
-template <typename Key, typename = size_t>
-struct IsHashable : std::true_type {};
-#else   // ABSL_META_INTERNAL_STD_HASH_SFINAE_FRIENDLY_
-template <typename Key, typename = void>
-struct IsHashable : std::false_type {};
-
-template <typename Key>
-struct IsHashable<
-    Key,
-    absl::enable_if_t<std::is_convertible<
-        decltype(std::declval<std::hash<Key>&>()(std::declval<Key const&>())),
-        std::size_t>::value>> : std::true_type {};
-#endif  // !ABSL_META_INTERNAL_STD_HASH_SFINAE_FRIENDLY_
-
-struct AssertHashEnabledHelper {
- private:
-  static void Sink(...) {}
-  struct NAT {};
-
-  template <class Key>
-  static auto GetReturnType(int)
-      -> decltype(std::declval<std::hash<Key>>()(std::declval<Key const&>()));
-  template <class Key>
-  static NAT GetReturnType(...);
-
-  template <class Key>
-  static std::nullptr_t DoIt() {
-    static_assert(IsHashable<Key>::value,
-                  "std::hash<Key> does not provide a call operator");
-    static_assert(
-        std::is_default_constructible<std::hash<Key>>::value,
-        "std::hash<Key> must be default constructible when it is enabled");
-    static_assert(
-        std::is_copy_constructible<std::hash<Key>>::value,
-        "std::hash<Key> must be copy constructible when it is enabled");
-    static_assert(absl::is_copy_assignable<std::hash<Key>>::value,
-                  "std::hash<Key> must be copy assignable when it is enabled");
-    // is_destructible is unchecked as it's implied by each of the
-    // is_constructible checks.
-    using ReturnType = decltype(GetReturnType<Key>(0));
-    static_assert(std::is_same<ReturnType, NAT>::value ||
-                      std::is_same<ReturnType, size_t>::value,
-                  "std::hash<Key> must return size_t");
-    return nullptr;
-  }
-
-  template <class... Ts>
-  friend void AssertHashEnabled();
-};
-
-template <class... Ts>
-inline void AssertHashEnabled() {
-  using Helper = AssertHashEnabledHelper;
-  Helper::Sink(Helper::DoIt<Ts>()...);
-}
-
-}  // namespace type_traits_internal
-
-// An internal namespace that is required to implement the C++17 swap traits.
-// It is not further nested in type_traits_internal to avoid long symbol names.
-namespace swap_internal {
-
-// Necessary for the traits.
-using std::swap;
-
-// This declaration prevents global `swap` and `absl::swap` overloads from being
-// considered unless ADL picks them up.
-void swap();
-
-template <class T>
-using IsSwappableImpl = decltype(swap(std::declval<T&>(), std::declval<T&>()));
-
-// NOTE: This dance with the default template parameter is for MSVC.
-template <class T,
-          class IsNoexcept = std::integral_constant<
-              bool, noexcept(swap(std::declval<T&>(), std::declval<T&>()))>>
-using IsNothrowSwappableImpl = typename std::enable_if<IsNoexcept::value>::type;
-
-// IsSwappable
-//
-// Determines whether the standard swap idiom is a valid expression for
-// arguments of type `T`.
-template <class T>
-struct IsSwappable
-    : absl::type_traits_internal::is_detected<IsSwappableImpl, T> {};
-
-// IsNothrowSwappable
-//
-// Determines whether the standard swap idiom is a valid expression for
-// arguments of type `T` and is noexcept.
-template <class T>
-struct IsNothrowSwappable
-    : absl::type_traits_internal::is_detected<IsNothrowSwappableImpl, T> {};
-
-// Swap()
-//
-// Performs the swap idiom from a namespace where valid candidates may only be
-// found in `std` or via ADL.
-template <class T, absl::enable_if_t<IsSwappable<T>::value, int> = 0>
-void Swap(T& lhs, T& rhs) noexcept(IsNothrowSwappable<T>::value) {
-  swap(lhs, rhs);
-}
-
-// StdSwapIsUnconstrained
-//
-// Some standard library implementations are broken in that they do not
-// constrain `std::swap`. This will effectively tell us if we are dealing with
-// one of those implementations.
-using StdSwapIsUnconstrained = IsSwappable<void()>;
-
-}  // namespace swap_internal
-
-namespace type_traits_internal {
-
-// Make the swap-related traits/function accessible from this namespace.
-using swap_internal::IsNothrowSwappable;
-using swap_internal::IsSwappable;
-using swap_internal::StdSwapIsUnconstrained;
-using swap_internal::Swap;
-
-}  // namespace type_traits_internal
-
-// absl::is_trivially_relocatable<T>
-//
-// Detects whether a type is known to be "trivially relocatable" -- meaning it
-// can be relocated from one place to another as if by memcpy/memmove.
-// This implies that its object representation doesn't depend on its address,
-// and also none of its special member functions do anything strange.
-//
-// This trait is conservative. If it's true then the type is definitely
-// trivially relocatable, but if it's false then the type may or may not be. For
-// example, std::vector<int> is trivially relocatable on every known STL
-// implementation, but absl::is_trivially_relocatable<std::vector<int>> remains
-// false.
-//
-// Example:
-//
-// if constexpr (absl::is_trivially_relocatable<T>::value) {
-//   memcpy(new_location, old_location, sizeof(T));
-// } else {
-//   new(new_location) T(std::move(*old_location));
-//   old_location->~T();
-// }
-//
-// Upstream documentation:
-//
-// https://clang.llvm.org/docs/LanguageExtensions.html#:~:text=__is_trivially_relocatable
-
-// If the compiler offers a builtin that tells us the answer, we can use that.
-// This covers all of the cases in the fallback below, plus types that opt in
-// using e.g. [[clang::trivial_abi]].
-//
-// Clang on Windows has the builtin, but it falsely claims types with a
-// user-provided destructor are trivial (http://b/275003464). So we opt out
-// there.
-//
-// TODO(b/275003464): remove the opt-out once the bug is fixed.
-//
-// Starting with Xcode 15, the Apple compiler will falsely say a type
-// with a user-provided move constructor is trivially relocatable
-// (b/324278148). We will opt out without a version check, due to
-// the fluidity of Apple versions.
-//
-// TODO(b/324278148): If all versions we use have the bug fixed, then
-// remove the condition.
-//
-// Clang on all platforms fails to detect that a type with a user-provided
-// move-assignment operator is not trivially relocatable. So in fact we
-// opt out of Clang altogether, for now.
-//
-// TODO(b/325479096): Remove the opt-out once Clang's behavior is fixed.
-//
-// According to https://github.com/abseil/abseil-cpp/issues/1479, this does not
-// work with NVCC either.
-#if ABSL_HAVE_BUILTIN(__is_trivially_relocatable) && \
-    (defined(__cpp_impl_trivially_relocatable) ||    \
-     (!defined(__clang__) && !defined(__APPLE__) && !defined(__NVCC__)))
-template <class T>
-struct is_trivially_relocatable
-    : std::integral_constant<bool, __is_trivially_relocatable(T)> {};
-#else
-// Otherwise we use a fallback that detects only those types we can feasibly
-// detect. Any type that is trivially copyable is by definition trivially
-// relocatable.
-template <class T>
-struct is_trivially_relocatable : std::is_trivially_copyable<T> {};
-#endif
-
-// absl::is_constant_evaluated()
-//
-// Detects whether the function call occurs within a constant-evaluated context.
-// Returns true if the evaluation of the call occurs within the evaluation of an
-// expression or conversion that is manifestly constant-evaluated; otherwise
-// returns false.
-//
-// This function is implemented in terms of `std::is_constant_evaluated` for
-// c++20 and up. For older c++ versions, the function is implemented in terms
-// of `__builtin_is_constant_evaluated` if available, otherwise the function
-// will fail to compile.
-//
-// Applications can inspect `ABSL_HAVE_CONSTANT_EVALUATED` at compile time
-// to check if this function is supported.
-//
-// Example:
-//
-// constexpr MyClass::MyClass(int param) {
-// #ifdef ABSL_HAVE_CONSTANT_EVALUATED
-//   if (!absl::is_constant_evaluated()) {
-//     ABSL_LOG(INFO) << "MyClass(" << param << ")";
-//   }
-// #endif  // ABSL_HAVE_CONSTANT_EVALUATED
-// }
-//
-// Upstream documentation:
-//
-// http://en.cppreference.com/w/cpp/types/is_constant_evaluated
-// http://gcc.gnu.org/onlinedocs/gcc/Other-Builtins.html#:~:text=__builtin_is_constant_evaluated
-//
-#if defined(ABSL_HAVE_CONSTANT_EVALUATED)
-constexpr bool is_constant_evaluated() noexcept {
-#ifdef __cpp_lib_is_constant_evaluated
-  return std::is_constant_evaluated();
-#elif ABSL_HAVE_BUILTIN(__builtin_is_constant_evaluated)
-  return __builtin_is_constant_evaluated();
-#endif
-}
-#endif  // ABSL_HAVE_CONSTANT_EVALUATED
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_META_TYPE_TRAITS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/numeric/bits.h b/projects/llm_framework/include/abseil-cpp/absl/numeric/bits.h
deleted file mode 100644
index c76454c8..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/numeric/bits.h
+++ /dev/null
@@ -1,196 +0,0 @@
-// Copyright 2020 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: bits.h
-// -----------------------------------------------------------------------------
-//
-// This file contains implementations of C++20's bitwise math functions, as
-// defined by:
-//
-// P0553R4:
-//  http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p0553r4.html
-// P0556R3:
-//  http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0556r3.html
-// P1355R2:
-//  http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1355r2.html
-// P1956R1:
-//  http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p1956r1.pdf
-//
-// When using a standard library that implements these functions, we use the
-// standard library's implementation.
-
-#ifndef ABSL_NUMERIC_BITS_H_
-#define ABSL_NUMERIC_BITS_H_
-
-#include <cstdint>
-#include <limits>
-#include <type_traits>
-
-#include "absl/base/config.h"
-
-#if ABSL_INTERNAL_CPLUSPLUS_LANG >= 202002L
-#include <bit>
-#endif
-
-#include "absl/base/attributes.h"
-#include "absl/numeric/internal/bits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// https://github.com/llvm/llvm-project/issues/64544
-// libc++ had the wrong signature for std::rotl and std::rotr
-// prior to libc++ 18.0.
-//
-#if (defined(__cpp_lib_bitops) && __cpp_lib_bitops >= 201907L) &&     \
-    (!defined(_LIBCPP_VERSION) || _LIBCPP_VERSION >= 180000)
-using std::rotl;
-using std::rotr;
-
-#else
-
-// Rotating functions
-template <class T>
-ABSL_MUST_USE_RESULT constexpr
-    typename std::enable_if<std::is_unsigned<T>::value, T>::type
-    rotl(T x, int s) noexcept {
-  return numeric_internal::RotateLeft(x, s);
-}
-
-template <class T>
-ABSL_MUST_USE_RESULT constexpr
-    typename std::enable_if<std::is_unsigned<T>::value, T>::type
-    rotr(T x, int s) noexcept {
-  return numeric_internal::RotateRight(x, s);
-}
-
-#endif
-
-// https://github.com/llvm/llvm-project/issues/64544
-// libc++ had the wrong signature for std::rotl and std::rotr
-// prior to libc++ 18.0.
-//
-#if (defined(__cpp_lib_bitops) && __cpp_lib_bitops >= 201907L)
-
-using std::countl_one;
-using std::countl_zero;
-using std::countr_one;
-using std::countr_zero;
-using std::popcount;
-
-#else
-
-// Counting functions
-//
-// While these functions are typically constexpr, on some platforms, they may
-// not be marked as constexpr due to constraints of the compiler/available
-// intrinsics.
-template <class T>
-ABSL_INTERNAL_CONSTEXPR_CLZ inline
-    typename std::enable_if<std::is_unsigned<T>::value, int>::type
-    countl_zero(T x) noexcept {
-  return numeric_internal::CountLeadingZeroes(x);
-}
-
-template <class T>
-ABSL_INTERNAL_CONSTEXPR_CLZ inline
-    typename std::enable_if<std::is_unsigned<T>::value, int>::type
-    countl_one(T x) noexcept {
-  // Avoid integer promotion to a wider type
-  return countl_zero(static_cast<T>(~x));
-}
-
-template <class T>
-ABSL_INTERNAL_CONSTEXPR_CTZ inline
-    typename std::enable_if<std::is_unsigned<T>::value, int>::type
-    countr_zero(T x) noexcept {
-  return numeric_internal::CountTrailingZeroes(x);
-}
-
-template <class T>
-ABSL_INTERNAL_CONSTEXPR_CTZ inline
-    typename std::enable_if<std::is_unsigned<T>::value, int>::type
-    countr_one(T x) noexcept {
-  // Avoid integer promotion to a wider type
-  return countr_zero(static_cast<T>(~x));
-}
-
-template <class T>
-ABSL_INTERNAL_CONSTEXPR_POPCOUNT inline
-    typename std::enable_if<std::is_unsigned<T>::value, int>::type
-    popcount(T x) noexcept {
-  return numeric_internal::Popcount(x);
-}
-
-#endif
-
-#if (defined(__cpp_lib_int_pow2) && __cpp_lib_int_pow2 >= 202002L)
-
-using std::bit_ceil;
-using std::bit_floor;
-using std::bit_width;
-using std::has_single_bit;
-
-#else
-
-// Returns: true if x is an integral power of two; false otherwise.
-template <class T>
-constexpr inline typename std::enable_if<std::is_unsigned<T>::value, bool>::type
-has_single_bit(T x) noexcept {
-  return x != 0 && (x & (x - 1)) == 0;
-}
-
-// Returns: If x == 0, 0; otherwise one plus the base-2 logarithm of x, with any
-// fractional part discarded.
-template <class T>
-ABSL_INTERNAL_CONSTEXPR_CLZ inline
-    typename std::enable_if<std::is_unsigned<T>::value, int>::type
-    bit_width(T x) noexcept {
-  return std::numeric_limits<T>::digits - countl_zero(x);
-}
-
-// Returns: If x == 0, 0; otherwise the maximal value y such that
-// has_single_bit(y) is true and y <= x.
-template <class T>
-ABSL_INTERNAL_CONSTEXPR_CLZ inline
-    typename std::enable_if<std::is_unsigned<T>::value, T>::type
-    bit_floor(T x) noexcept {
-  return x == 0 ? 0 : T{1} << (bit_width(x) - 1);
-}
-
-// Returns: N, where N is the smallest power of 2 greater than or equal to x.
-//
-// Preconditions: N is representable as a value of type T.
-template <class T>
-ABSL_INTERNAL_CONSTEXPR_CLZ inline
-    typename std::enable_if<std::is_unsigned<T>::value, T>::type
-    bit_ceil(T x) {
-  // If T is narrower than unsigned, T{1} << bit_width will be promoted.  We
-  // want to force it to wraparound so that bit_ceil of an invalid value are not
-  // core constant expressions.
-  //
-  // BitCeilNonPowerOf2 triggers an overflow in constexpr contexts if we would
-  // undergo promotion to unsigned but not fit the result into T without
-  // truncation.
-  return has_single_bit(x) ? T{1} << (bit_width(x) - 1)
-                           : numeric_internal::BitCeilNonPowerOf2(x);
-}
-
-#endif
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_NUMERIC_BITS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/numeric/int128.h b/projects/llm_framework/include/abseil-cpp/absl/numeric/int128.h
deleted file mode 100644
index a17d0e1b..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/numeric/int128.h
+++ /dev/null
@@ -1,1157 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: int128.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines 128-bit integer types, `uint128` and `int128`.
-//
-// TODO(absl-team): This module is inconsistent as many inline `uint128` methods
-// are defined in this file, while many inline `int128` methods are defined in
-// the `int128_*_intrinsic.inc` files.
-
-#ifndef ABSL_NUMERIC_INT128_H_
-#define ABSL_NUMERIC_INT128_H_
-
-#include <cassert>
-#include <cmath>
-#include <cstdint>
-#include <cstring>
-#include <iosfwd>
-#include <limits>
-#include <string>
-#include <utility>
-
-#include "absl/base/config.h"
-#include "absl/base/macros.h"
-#include "absl/base/port.h"
-
-#if defined(_MSC_VER)
-// In very old versions of MSVC and when the /Zc:wchar_t flag is off, wchar_t is
-// a typedef for unsigned short.  Otherwise wchar_t is mapped to the __wchar_t
-// builtin type.  We need to make sure not to define operator wchar_t()
-// alongside operator unsigned short() in these instances.
-#define ABSL_INTERNAL_WCHAR_T __wchar_t
-#if defined(_M_X64) && !defined(_M_ARM64EC)
-#include <intrin.h>
-#pragma intrinsic(_umul128)
-#endif  // defined(_M_X64)
-#else   // defined(_MSC_VER)
-#define ABSL_INTERNAL_WCHAR_T wchar_t
-#endif  // defined(_MSC_VER)
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-class int128;
-
-// uint128
-//
-// An unsigned 128-bit integer type. The API is meant to mimic an intrinsic type
-// as closely as is practical, including exhibiting undefined behavior in
-// analogous cases (e.g. division by zero). This type is intended to be a
-// drop-in replacement once C++ supports an intrinsic `uint128_t` type; when
-// that occurs, existing well-behaved uses of `uint128` will continue to work
-// using that new type.
-//
-// Note: code written with this type will continue to compile once `uint128_t`
-// is introduced, provided the replacement helper functions
-// `Uint128(Low|High)64()` and `MakeUint128()` are made.
-//
-// A `uint128` supports the following:
-//
-//   * Implicit construction from integral types
-//   * Explicit conversion to integral types
-//
-// Additionally, if your compiler supports `__int128`, `uint128` is
-// interoperable with that type. (Abseil checks for this compatibility through
-// the `ABSL_HAVE_INTRINSIC_INT128` macro.)
-//
-// However, a `uint128` differs from intrinsic integral types in the following
-// ways:
-//
-//   * Errors on implicit conversions that do not preserve value (such as
-//     loss of precision when converting to float values).
-//   * Requires explicit construction from and conversion to floating point
-//     types.
-//   * Conversion to integral types requires an explicit static_cast() to
-//     mimic use of the `-Wnarrowing` compiler flag.
-//   * The alignment requirement of `uint128` may differ from that of an
-//     intrinsic 128-bit integer type depending on platform and build
-//     configuration.
-//
-// Example:
-//
-//     float y = absl::Uint128Max();  // Error. uint128 cannot be implicitly
-//                                    // converted to float.
-//
-//     absl::uint128 v;
-//     uint64_t i = v;                         // Error
-//     uint64_t i = static_cast<uint64_t>(v);  // OK
-//
-class
-#if defined(ABSL_HAVE_INTRINSIC_INT128)
-    alignas(unsigned __int128)
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-        uint128 {
- public:
-  uint128() = default;
-
-  // Constructors from arithmetic types
-  constexpr uint128(int v);                 // NOLINT(runtime/explicit)
-  constexpr uint128(unsigned int v);        // NOLINT(runtime/explicit)
-  constexpr uint128(long v);                // NOLINT(runtime/int)
-  constexpr uint128(unsigned long v);       // NOLINT(runtime/int)
-  constexpr uint128(long long v);           // NOLINT(runtime/int)
-  constexpr uint128(unsigned long long v);  // NOLINT(runtime/int)
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-  constexpr uint128(__int128 v);           // NOLINT(runtime/explicit)
-  constexpr uint128(unsigned __int128 v);  // NOLINT(runtime/explicit)
-#endif                                     // ABSL_HAVE_INTRINSIC_INT128
-  constexpr uint128(int128 v);             // NOLINT(runtime/explicit)
-  explicit uint128(float v);
-  explicit uint128(double v);
-  explicit uint128(long double v);
-
-  // Assignment operators from arithmetic types
-  uint128& operator=(int v);
-  uint128& operator=(unsigned int v);
-  uint128& operator=(long v);                // NOLINT(runtime/int)
-  uint128& operator=(unsigned long v);       // NOLINT(runtime/int)
-  uint128& operator=(long long v);           // NOLINT(runtime/int)
-  uint128& operator=(unsigned long long v);  // NOLINT(runtime/int)
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-  uint128& operator=(__int128 v);
-  uint128& operator=(unsigned __int128 v);
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-  uint128& operator=(int128 v);
-
-  // Conversion operators to other arithmetic types
-  constexpr explicit operator bool() const;
-  constexpr explicit operator char() const;
-  constexpr explicit operator signed char() const;
-  constexpr explicit operator unsigned char() const;
-  constexpr explicit operator char16_t() const;
-  constexpr explicit operator char32_t() const;
-  constexpr explicit operator ABSL_INTERNAL_WCHAR_T() const;
-  constexpr explicit operator short() const;  // NOLINT(runtime/int)
-  // NOLINTNEXTLINE(runtime/int)
-  constexpr explicit operator unsigned short() const;
-  constexpr explicit operator int() const;
-  constexpr explicit operator unsigned int() const;
-  constexpr explicit operator long() const;  // NOLINT(runtime/int)
-  // NOLINTNEXTLINE(runtime/int)
-  constexpr explicit operator unsigned long() const;
-  // NOLINTNEXTLINE(runtime/int)
-  constexpr explicit operator long long() const;
-  // NOLINTNEXTLINE(runtime/int)
-  constexpr explicit operator unsigned long long() const;
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-  constexpr explicit operator __int128() const;
-  constexpr explicit operator unsigned __int128() const;
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-  explicit operator float() const;
-  explicit operator double() const;
-  explicit operator long double() const;
-
-  // Trivial copy constructor, assignment operator and destructor.
-
-  // Arithmetic operators.
-  uint128& operator+=(uint128 other);
-  uint128& operator-=(uint128 other);
-  uint128& operator*=(uint128 other);
-  // Long division/modulo for uint128.
-  uint128& operator/=(uint128 other);
-  uint128& operator%=(uint128 other);
-  uint128 operator++(int);
-  uint128 operator--(int);
-  uint128& operator<<=(int);
-  uint128& operator>>=(int);
-  uint128& operator&=(uint128 other);
-  uint128& operator|=(uint128 other);
-  uint128& operator^=(uint128 other);
-  uint128& operator++();
-  uint128& operator--();
-
-  // Uint128Low64()
-  //
-  // Returns the lower 64-bit value of a `uint128` value.
-  friend constexpr uint64_t Uint128Low64(uint128 v);
-
-  // Uint128High64()
-  //
-  // Returns the higher 64-bit value of a `uint128` value.
-  friend constexpr uint64_t Uint128High64(uint128 v);
-
-  // MakeUInt128()
-  //
-  // Constructs a `uint128` numeric value from two 64-bit unsigned integers.
-  // Note that this factory function is the only way to construct a `uint128`
-  // from integer values greater than 2^64.
-  //
-  // Example:
-  //
-  //   absl::uint128 big = absl::MakeUint128(1, 0);
-  friend constexpr uint128 MakeUint128(uint64_t high, uint64_t low);
-
-  // Uint128Max()
-  //
-  // Returns the highest value for a 128-bit unsigned integer.
-  friend constexpr uint128 Uint128Max();
-
-  // Support for absl::Hash.
-  template <typename H>
-  friend H AbslHashValue(H h, uint128 v) {
-    return H::combine(std::move(h), Uint128High64(v), Uint128Low64(v));
-  }
-
-  // Support for absl::StrCat() etc.
-  template <typename Sink>
-  friend void AbslStringify(Sink& sink, uint128 v) {
-    sink.Append(v.ToString());
-  }
-
- private:
-  constexpr uint128(uint64_t high, uint64_t low);
-
-  std::string ToString() const;
-
-  // TODO(strel) Update implementation to use __int128 once all users of
-  // uint128 are fixed to not depend on alignof(uint128) == 8. Also add
-  // alignas(16) to class definition to keep alignment consistent across
-  // platforms.
-#if defined(ABSL_IS_LITTLE_ENDIAN)
-  uint64_t lo_;
-  uint64_t hi_;
-#elif defined(ABSL_IS_BIG_ENDIAN)
-  uint64_t hi_;
-  uint64_t lo_;
-#else  // byte order
-#error "Unsupported byte order: must be little-endian or big-endian."
-#endif  // byte order
-};
-
-// allow uint128 to be logged
-std::ostream& operator<<(std::ostream& os, uint128 v);
-
-// TODO(strel) add operator>>(std::istream&, uint128)
-
-constexpr uint128 Uint128Max() {
-  return uint128((std::numeric_limits<uint64_t>::max)(),
-                 (std::numeric_limits<uint64_t>::max)());
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-// Specialized numeric_limits for uint128.
-namespace std {
-template <>
-class numeric_limits<absl::uint128> {
- public:
-  static constexpr bool is_specialized = true;
-  static constexpr bool is_signed = false;
-  static constexpr bool is_integer = true;
-  static constexpr bool is_exact = true;
-  static constexpr bool has_infinity = false;
-  static constexpr bool has_quiet_NaN = false;
-  static constexpr bool has_signaling_NaN = false;
-  static constexpr float_denorm_style has_denorm = denorm_absent;
-  static constexpr bool has_denorm_loss = false;
-  static constexpr float_round_style round_style = round_toward_zero;
-  static constexpr bool is_iec559 = false;
-  static constexpr bool is_bounded = true;
-  static constexpr bool is_modulo = true;
-  static constexpr int digits = 128;
-  static constexpr int digits10 = 38;
-  static constexpr int max_digits10 = 0;
-  static constexpr int radix = 2;
-  static constexpr int min_exponent = 0;
-  static constexpr int min_exponent10 = 0;
-  static constexpr int max_exponent = 0;
-  static constexpr int max_exponent10 = 0;
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-  static constexpr bool traps = numeric_limits<unsigned __int128>::traps;
-#else   // ABSL_HAVE_INTRINSIC_INT128
-  static constexpr bool traps = numeric_limits<uint64_t>::traps;
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-  static constexpr bool tinyness_before = false;
-
-  static constexpr absl::uint128(min)() { return 0; }
-  static constexpr absl::uint128 lowest() { return 0; }
-  static constexpr absl::uint128(max)() { return absl::Uint128Max(); }
-  static constexpr absl::uint128 epsilon() { return 0; }
-  static constexpr absl::uint128 round_error() { return 0; }
-  static constexpr absl::uint128 infinity() { return 0; }
-  static constexpr absl::uint128 quiet_NaN() { return 0; }
-  static constexpr absl::uint128 signaling_NaN() { return 0; }
-  static constexpr absl::uint128 denorm_min() { return 0; }
-};
-}  // namespace std
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// int128
-//
-// A signed 128-bit integer type. The API is meant to mimic an intrinsic
-// integral type as closely as is practical, including exhibiting undefined
-// behavior in analogous cases (e.g. division by zero).
-//
-// An `int128` supports the following:
-//
-//   * Implicit construction from integral types
-//   * Explicit conversion to integral types
-//
-// However, an `int128` differs from intrinsic integral types in the following
-// ways:
-//
-//   * It is not implicitly convertible to other integral types.
-//   * Requires explicit construction from and conversion to floating point
-//     types.
-
-// Additionally, if your compiler supports `__int128`, `int128` is
-// interoperable with that type. (Abseil checks for this compatibility through
-// the `ABSL_HAVE_INTRINSIC_INT128` macro.)
-//
-// The design goal for `int128` is that it will be compatible with a future
-// `int128_t`, if that type becomes a part of the standard.
-//
-// Example:
-//
-//     float y = absl::int128(17);  // Error. int128 cannot be implicitly
-//                                  // converted to float.
-//
-//     absl::int128 v;
-//     int64_t i = v;                        // Error
-//     int64_t i = static_cast<int64_t>(v);  // OK
-//
-class int128 {
- public:
-  int128() = default;
-
-  // Constructors from arithmetic types
-  constexpr int128(int v);                 // NOLINT(runtime/explicit)
-  constexpr int128(unsigned int v);        // NOLINT(runtime/explicit)
-  constexpr int128(long v);                // NOLINT(runtime/int)
-  constexpr int128(unsigned long v);       // NOLINT(runtime/int)
-  constexpr int128(long long v);           // NOLINT(runtime/int)
-  constexpr int128(unsigned long long v);  // NOLINT(runtime/int)
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-  constexpr int128(__int128 v);  // NOLINT(runtime/explicit)
-  constexpr explicit int128(unsigned __int128 v);
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-  constexpr explicit int128(uint128 v);
-  explicit int128(float v);
-  explicit int128(double v);
-  explicit int128(long double v);
-
-  // Assignment operators from arithmetic types
-  int128& operator=(int v);
-  int128& operator=(unsigned int v);
-  int128& operator=(long v);                // NOLINT(runtime/int)
-  int128& operator=(unsigned long v);       // NOLINT(runtime/int)
-  int128& operator=(long long v);           // NOLINT(runtime/int)
-  int128& operator=(unsigned long long v);  // NOLINT(runtime/int)
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-  int128& operator=(__int128 v);
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-
-  // Conversion operators to other arithmetic types
-  constexpr explicit operator bool() const;
-  constexpr explicit operator char() const;
-  constexpr explicit operator signed char() const;
-  constexpr explicit operator unsigned char() const;
-  constexpr explicit operator char16_t() const;
-  constexpr explicit operator char32_t() const;
-  constexpr explicit operator ABSL_INTERNAL_WCHAR_T() const;
-  constexpr explicit operator short() const;  // NOLINT(runtime/int)
-  // NOLINTNEXTLINE(runtime/int)
-  constexpr explicit operator unsigned short() const;
-  constexpr explicit operator int() const;
-  constexpr explicit operator unsigned int() const;
-  constexpr explicit operator long() const;  // NOLINT(runtime/int)
-  // NOLINTNEXTLINE(runtime/int)
-  constexpr explicit operator unsigned long() const;
-  // NOLINTNEXTLINE(runtime/int)
-  constexpr explicit operator long long() const;
-  // NOLINTNEXTLINE(runtime/int)
-  constexpr explicit operator unsigned long long() const;
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-  constexpr explicit operator __int128() const;
-  constexpr explicit operator unsigned __int128() const;
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-  explicit operator float() const;
-  explicit operator double() const;
-  explicit operator long double() const;
-
-  // Trivial copy constructor, assignment operator and destructor.
-
-  // Arithmetic operators
-  int128& operator+=(int128 other);
-  int128& operator-=(int128 other);
-  int128& operator*=(int128 other);
-  int128& operator/=(int128 other);
-  int128& operator%=(int128 other);
-  int128 operator++(int);  // postfix increment: i++
-  int128 operator--(int);  // postfix decrement: i--
-  int128& operator++();    // prefix increment:  ++i
-  int128& operator--();    // prefix decrement:  --i
-  int128& operator&=(int128 other);
-  int128& operator|=(int128 other);
-  int128& operator^=(int128 other);
-  int128& operator<<=(int amount);
-  int128& operator>>=(int amount);
-
-  // Int128Low64()
-  //
-  // Returns the lower 64-bit value of a `int128` value.
-  friend constexpr uint64_t Int128Low64(int128 v);
-
-  // Int128High64()
-  //
-  // Returns the higher 64-bit value of a `int128` value.
-  friend constexpr int64_t Int128High64(int128 v);
-
-  // MakeInt128()
-  //
-  // Constructs a `int128` numeric value from two 64-bit integers. Note that
-  // signedness is conveyed in the upper `high` value.
-  //
-  //   (absl::int128(1) << 64) * high + low
-  //
-  // Note that this factory function is the only way to construct a `int128`
-  // from integer values greater than 2^64 or less than -2^64.
-  //
-  // Example:
-  //
-  //   absl::int128 big = absl::MakeInt128(1, 0);
-  //   absl::int128 big_n = absl::MakeInt128(-1, 0);
-  friend constexpr int128 MakeInt128(int64_t high, uint64_t low);
-
-  // Int128Max()
-  //
-  // Returns the maximum value for a 128-bit signed integer.
-  friend constexpr int128 Int128Max();
-
-  // Int128Min()
-  //
-  // Returns the minimum value for a 128-bit signed integer.
-  friend constexpr int128 Int128Min();
-
-  // Support for absl::Hash.
-  template <typename H>
-  friend H AbslHashValue(H h, int128 v) {
-    return H::combine(std::move(h), Int128High64(v), Int128Low64(v));
-  }
-
-  // Support for absl::StrCat() etc.
-  template <typename Sink>
-  friend void AbslStringify(Sink& sink, int128 v) {
-    sink.Append(v.ToString());
-  }
-
- private:
-  constexpr int128(int64_t high, uint64_t low);
-
-  std::string ToString() const;
-
-#if defined(ABSL_HAVE_INTRINSIC_INT128)
-  __int128 v_;
-#else  // ABSL_HAVE_INTRINSIC_INT128
-#if defined(ABSL_IS_LITTLE_ENDIAN)
-  uint64_t lo_;
-  int64_t hi_;
-#elif defined(ABSL_IS_BIG_ENDIAN)
-  int64_t hi_;
-  uint64_t lo_;
-#else  // byte order
-#error "Unsupported byte order: must be little-endian or big-endian."
-#endif  // byte order
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-};
-
-std::ostream& operator<<(std::ostream& os, int128 v);
-
-// TODO(absl-team) add operator>>(std::istream&, int128)
-
-constexpr int128 Int128Max() {
-  return int128((std::numeric_limits<int64_t>::max)(),
-                (std::numeric_limits<uint64_t>::max)());
-}
-
-constexpr int128 Int128Min() {
-  return int128((std::numeric_limits<int64_t>::min)(), 0);
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-// Specialized numeric_limits for int128.
-namespace std {
-template <>
-class numeric_limits<absl::int128> {
- public:
-  static constexpr bool is_specialized = true;
-  static constexpr bool is_signed = true;
-  static constexpr bool is_integer = true;
-  static constexpr bool is_exact = true;
-  static constexpr bool has_infinity = false;
-  static constexpr bool has_quiet_NaN = false;
-  static constexpr bool has_signaling_NaN = false;
-  static constexpr float_denorm_style has_denorm = denorm_absent;
-  static constexpr bool has_denorm_loss = false;
-  static constexpr float_round_style round_style = round_toward_zero;
-  static constexpr bool is_iec559 = false;
-  static constexpr bool is_bounded = true;
-  static constexpr bool is_modulo = false;
-  static constexpr int digits = 127;
-  static constexpr int digits10 = 38;
-  static constexpr int max_digits10 = 0;
-  static constexpr int radix = 2;
-  static constexpr int min_exponent = 0;
-  static constexpr int min_exponent10 = 0;
-  static constexpr int max_exponent = 0;
-  static constexpr int max_exponent10 = 0;
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-  static constexpr bool traps = numeric_limits<__int128>::traps;
-#else   // ABSL_HAVE_INTRINSIC_INT128
-  static constexpr bool traps = numeric_limits<uint64_t>::traps;
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-  static constexpr bool tinyness_before = false;
-
-  static constexpr absl::int128(min)() { return absl::Int128Min(); }
-  static constexpr absl::int128 lowest() { return absl::Int128Min(); }
-  static constexpr absl::int128(max)() { return absl::Int128Max(); }
-  static constexpr absl::int128 epsilon() { return 0; }
-  static constexpr absl::int128 round_error() { return 0; }
-  static constexpr absl::int128 infinity() { return 0; }
-  static constexpr absl::int128 quiet_NaN() { return 0; }
-  static constexpr absl::int128 signaling_NaN() { return 0; }
-  static constexpr absl::int128 denorm_min() { return 0; }
-};
-}  // namespace std
-
-// --------------------------------------------------------------------------
-//                      Implementation details follow
-// --------------------------------------------------------------------------
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-constexpr uint128 MakeUint128(uint64_t high, uint64_t low) {
-  return uint128(high, low);
-}
-
-// Assignment from integer types.
-
-inline uint128& uint128::operator=(int v) { return *this = uint128(v); }
-
-inline uint128& uint128::operator=(unsigned int v) {
-  return *this = uint128(v);
-}
-
-inline uint128& uint128::operator=(long v) {  // NOLINT(runtime/int)
-  return *this = uint128(v);
-}
-
-// NOLINTNEXTLINE(runtime/int)
-inline uint128& uint128::operator=(unsigned long v) {
-  return *this = uint128(v);
-}
-
-// NOLINTNEXTLINE(runtime/int)
-inline uint128& uint128::operator=(long long v) { return *this = uint128(v); }
-
-// NOLINTNEXTLINE(runtime/int)
-inline uint128& uint128::operator=(unsigned long long v) {
-  return *this = uint128(v);
-}
-
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-inline uint128& uint128::operator=(__int128 v) { return *this = uint128(v); }
-
-inline uint128& uint128::operator=(unsigned __int128 v) {
-  return *this = uint128(v);
-}
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-
-inline uint128& uint128::operator=(int128 v) { return *this = uint128(v); }
-
-// Arithmetic operators.
-
-constexpr uint128 operator<<(uint128 lhs, int amount);
-constexpr uint128 operator>>(uint128 lhs, int amount);
-constexpr uint128 operator+(uint128 lhs, uint128 rhs);
-constexpr uint128 operator-(uint128 lhs, uint128 rhs);
-uint128 operator*(uint128 lhs, uint128 rhs);
-uint128 operator/(uint128 lhs, uint128 rhs);
-uint128 operator%(uint128 lhs, uint128 rhs);
-
-inline uint128& uint128::operator<<=(int amount) {
-  *this = *this << amount;
-  return *this;
-}
-
-inline uint128& uint128::operator>>=(int amount) {
-  *this = *this >> amount;
-  return *this;
-}
-
-inline uint128& uint128::operator+=(uint128 other) {
-  *this = *this + other;
-  return *this;
-}
-
-inline uint128& uint128::operator-=(uint128 other) {
-  *this = *this - other;
-  return *this;
-}
-
-inline uint128& uint128::operator*=(uint128 other) {
-  *this = *this * other;
-  return *this;
-}
-
-inline uint128& uint128::operator/=(uint128 other) {
-  *this = *this / other;
-  return *this;
-}
-
-inline uint128& uint128::operator%=(uint128 other) {
-  *this = *this % other;
-  return *this;
-}
-
-constexpr uint64_t Uint128Low64(uint128 v) { return v.lo_; }
-
-constexpr uint64_t Uint128High64(uint128 v) { return v.hi_; }
-
-// Constructors from integer types.
-
-#if defined(ABSL_IS_LITTLE_ENDIAN)
-
-constexpr uint128::uint128(uint64_t high, uint64_t low) : lo_{low}, hi_{high} {}
-
-constexpr uint128::uint128(int v)
-    : lo_{static_cast<uint64_t>(v)},
-      hi_{v < 0 ? (std::numeric_limits<uint64_t>::max)() : 0} {}
-constexpr uint128::uint128(long v)  // NOLINT(runtime/int)
-    : lo_{static_cast<uint64_t>(v)},
-      hi_{v < 0 ? (std::numeric_limits<uint64_t>::max)() : 0} {}
-constexpr uint128::uint128(long long v)  // NOLINT(runtime/int)
-    : lo_{static_cast<uint64_t>(v)},
-      hi_{v < 0 ? (std::numeric_limits<uint64_t>::max)() : 0} {}
-
-constexpr uint128::uint128(unsigned int v) : lo_{v}, hi_{0} {}
-// NOLINTNEXTLINE(runtime/int)
-constexpr uint128::uint128(unsigned long v) : lo_{v}, hi_{0} {}
-// NOLINTNEXTLINE(runtime/int)
-constexpr uint128::uint128(unsigned long long v) : lo_{v}, hi_{0} {}
-
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-constexpr uint128::uint128(__int128 v)
-    : lo_{static_cast<uint64_t>(v & ~uint64_t{0})},
-      hi_{static_cast<uint64_t>(static_cast<unsigned __int128>(v) >> 64)} {}
-constexpr uint128::uint128(unsigned __int128 v)
-    : lo_{static_cast<uint64_t>(v & ~uint64_t{0})},
-      hi_{static_cast<uint64_t>(v >> 64)} {}
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-
-constexpr uint128::uint128(int128 v)
-    : lo_{Int128Low64(v)}, hi_{static_cast<uint64_t>(Int128High64(v))} {}
-
-#elif defined(ABSL_IS_BIG_ENDIAN)
-
-constexpr uint128::uint128(uint64_t high, uint64_t low) : hi_{high}, lo_{low} {}
-
-constexpr uint128::uint128(int v)
-    : hi_{v < 0 ? (std::numeric_limits<uint64_t>::max)() : 0},
-      lo_{static_cast<uint64_t>(v)} {}
-constexpr uint128::uint128(long v)  // NOLINT(runtime/int)
-    : hi_{v < 0 ? (std::numeric_limits<uint64_t>::max)() : 0},
-      lo_{static_cast<uint64_t>(v)} {}
-constexpr uint128::uint128(long long v)  // NOLINT(runtime/int)
-    : hi_{v < 0 ? (std::numeric_limits<uint64_t>::max)() : 0},
-      lo_{static_cast<uint64_t>(v)} {}
-
-constexpr uint128::uint128(unsigned int v) : hi_{0}, lo_{v} {}
-// NOLINTNEXTLINE(runtime/int)
-constexpr uint128::uint128(unsigned long v) : hi_{0}, lo_{v} {}
-// NOLINTNEXTLINE(runtime/int)
-constexpr uint128::uint128(unsigned long long v) : hi_{0}, lo_{v} {}
-
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-constexpr uint128::uint128(__int128 v)
-    : hi_{static_cast<uint64_t>(static_cast<unsigned __int128>(v) >> 64)},
-      lo_{static_cast<uint64_t>(v & ~uint64_t{0})} {}
-constexpr uint128::uint128(unsigned __int128 v)
-    : hi_{static_cast<uint64_t>(v >> 64)},
-      lo_{static_cast<uint64_t>(v & ~uint64_t{0})} {}
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-
-constexpr uint128::uint128(int128 v)
-    : hi_{static_cast<uint64_t>(Int128High64(v))}, lo_{Int128Low64(v)} {}
-
-#else  // byte order
-#error "Unsupported byte order: must be little-endian or big-endian."
-#endif  // byte order
-
-// Conversion operators to integer types.
-
-constexpr uint128::operator bool() const { return lo_ || hi_; }
-
-constexpr uint128::operator char() const { return static_cast<char>(lo_); }
-
-constexpr uint128::operator signed char() const {
-  return static_cast<signed char>(lo_);
-}
-
-constexpr uint128::operator unsigned char() const {
-  return static_cast<unsigned char>(lo_);
-}
-
-constexpr uint128::operator char16_t() const {
-  return static_cast<char16_t>(lo_);
-}
-
-constexpr uint128::operator char32_t() const {
-  return static_cast<char32_t>(lo_);
-}
-
-constexpr uint128::operator ABSL_INTERNAL_WCHAR_T() const {
-  return static_cast<ABSL_INTERNAL_WCHAR_T>(lo_);
-}
-
-// NOLINTNEXTLINE(runtime/int)
-constexpr uint128::operator short() const { return static_cast<short>(lo_); }
-
-constexpr uint128::operator unsigned short() const {  // NOLINT(runtime/int)
-  return static_cast<unsigned short>(lo_);            // NOLINT(runtime/int)
-}
-
-constexpr uint128::operator int() const { return static_cast<int>(lo_); }
-
-constexpr uint128::operator unsigned int() const {
-  return static_cast<unsigned int>(lo_);
-}
-
-// NOLINTNEXTLINE(runtime/int)
-constexpr uint128::operator long() const { return static_cast<long>(lo_); }
-
-constexpr uint128::operator unsigned long() const {  // NOLINT(runtime/int)
-  return static_cast<unsigned long>(lo_);            // NOLINT(runtime/int)
-}
-
-constexpr uint128::operator long long() const {  // NOLINT(runtime/int)
-  return static_cast<long long>(lo_);            // NOLINT(runtime/int)
-}
-
-constexpr uint128::operator unsigned long long() const {  // NOLINT(runtime/int)
-  return static_cast<unsigned long long>(lo_);            // NOLINT(runtime/int)
-}
-
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-constexpr uint128::operator __int128() const {
-  return (static_cast<__int128>(hi_) << 64) + lo_;
-}
-
-constexpr uint128::operator unsigned __int128() const {
-  return (static_cast<unsigned __int128>(hi_) << 64) + lo_;
-}
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-
-// Conversion operators to floating point types.
-
-inline uint128::operator float() const {
-  return static_cast<float>(lo_) + std::ldexp(static_cast<float>(hi_), 64);
-}
-
-inline uint128::operator double() const {
-  return static_cast<double>(lo_) + std::ldexp(static_cast<double>(hi_), 64);
-}
-
-inline uint128::operator long double() const {
-  return static_cast<long double>(lo_) +
-         std::ldexp(static_cast<long double>(hi_), 64);
-}
-
-// Comparison operators.
-
-constexpr bool operator==(uint128 lhs, uint128 rhs) {
-#if defined(ABSL_HAVE_INTRINSIC_INT128)
-  return static_cast<unsigned __int128>(lhs) ==
-         static_cast<unsigned __int128>(rhs);
-#else
-  return (Uint128Low64(lhs) == Uint128Low64(rhs) &&
-          Uint128High64(lhs) == Uint128High64(rhs));
-#endif
-}
-
-constexpr bool operator!=(uint128 lhs, uint128 rhs) { return !(lhs == rhs); }
-
-constexpr bool operator<(uint128 lhs, uint128 rhs) {
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-  return static_cast<unsigned __int128>(lhs) <
-         static_cast<unsigned __int128>(rhs);
-#else
-  return (Uint128High64(lhs) == Uint128High64(rhs))
-             ? (Uint128Low64(lhs) < Uint128Low64(rhs))
-             : (Uint128High64(lhs) < Uint128High64(rhs));
-#endif
-}
-
-constexpr bool operator>(uint128 lhs, uint128 rhs) { return rhs < lhs; }
-
-constexpr bool operator<=(uint128 lhs, uint128 rhs) { return !(rhs < lhs); }
-
-constexpr bool operator>=(uint128 lhs, uint128 rhs) { return !(lhs < rhs); }
-
-// Unary operators.
-
-constexpr inline uint128 operator+(uint128 val) { return val; }
-
-constexpr inline int128 operator+(int128 val) { return val; }
-
-constexpr uint128 operator-(uint128 val) {
-#if defined(ABSL_HAVE_INTRINSIC_INT128)
-  return -static_cast<unsigned __int128>(val);
-#else
-  return MakeUint128(
-      ~Uint128High64(val) + static_cast<unsigned long>(Uint128Low64(val) == 0),
-      ~Uint128Low64(val) + 1);
-#endif
-}
-
-constexpr inline bool operator!(uint128 val) {
-#if defined(ABSL_HAVE_INTRINSIC_INT128)
-  return !static_cast<unsigned __int128>(val);
-#else
-  return !Uint128High64(val) && !Uint128Low64(val);
-#endif
-}
-
-// Logical operators.
-
-constexpr inline uint128 operator~(uint128 val) {
-#if defined(ABSL_HAVE_INTRINSIC_INT128)
-  return ~static_cast<unsigned __int128>(val);
-#else
-  return MakeUint128(~Uint128High64(val), ~Uint128Low64(val));
-#endif
-}
-
-constexpr inline uint128 operator|(uint128 lhs, uint128 rhs) {
-#if defined(ABSL_HAVE_INTRINSIC_INT128)
-  return static_cast<unsigned __int128>(lhs) |
-         static_cast<unsigned __int128>(rhs);
-#else
-  return MakeUint128(Uint128High64(lhs) | Uint128High64(rhs),
-                     Uint128Low64(lhs) | Uint128Low64(rhs));
-#endif
-}
-
-constexpr inline uint128 operator&(uint128 lhs, uint128 rhs) {
-#if defined(ABSL_HAVE_INTRINSIC_INT128)
-  return static_cast<unsigned __int128>(lhs) &
-         static_cast<unsigned __int128>(rhs);
-#else
-  return MakeUint128(Uint128High64(lhs) & Uint128High64(rhs),
-                     Uint128Low64(lhs) & Uint128Low64(rhs));
-#endif
-}
-
-constexpr inline uint128 operator^(uint128 lhs, uint128 rhs) {
-#if defined(ABSL_HAVE_INTRINSIC_INT128)
-  return static_cast<unsigned __int128>(lhs) ^
-         static_cast<unsigned __int128>(rhs);
-#else
-  return MakeUint128(Uint128High64(lhs) ^ Uint128High64(rhs),
-                     Uint128Low64(lhs) ^ Uint128Low64(rhs));
-#endif
-}
-
-inline uint128& uint128::operator|=(uint128 other) {
-  *this = *this | other;
-  return *this;
-}
-
-inline uint128& uint128::operator&=(uint128 other) {
-  *this = *this & other;
-  return *this;
-}
-
-inline uint128& uint128::operator^=(uint128 other) {
-  *this = *this ^ other;
-  return *this;
-}
-
-// Arithmetic operators.
-
-constexpr uint128 operator<<(uint128 lhs, int amount) {
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-  return static_cast<unsigned __int128>(lhs) << amount;
-#else
-  // uint64_t shifts of >= 64 are undefined, so we will need some
-  // special-casing.
-  return amount >= 64  ? MakeUint128(Uint128Low64(lhs) << (amount - 64), 0)
-         : amount == 0 ? lhs
-                       : MakeUint128((Uint128High64(lhs) << amount) |
-                                         (Uint128Low64(lhs) >> (64 - amount)),
-                                     Uint128Low64(lhs) << amount);
-#endif
-}
-
-constexpr uint128 operator>>(uint128 lhs, int amount) {
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-  return static_cast<unsigned __int128>(lhs) >> amount;
-#else
-  // uint64_t shifts of >= 64 are undefined, so we will need some
-  // special-casing.
-  return amount >= 64  ? MakeUint128(0, Uint128High64(lhs) >> (amount - 64))
-         : amount == 0 ? lhs
-                       : MakeUint128(Uint128High64(lhs) >> amount,
-                                     (Uint128Low64(lhs) >> amount) |
-                                         (Uint128High64(lhs) << (64 - amount)));
-#endif
-}
-
-#if !defined(ABSL_HAVE_INTRINSIC_INT128)
-namespace int128_internal {
-constexpr uint128 AddResult(uint128 result, uint128 lhs) {
-  // check for carry
-  return (Uint128Low64(result) < Uint128Low64(lhs))
-             ? MakeUint128(Uint128High64(result) + 1, Uint128Low64(result))
-             : result;
-}
-}  // namespace int128_internal
-#endif
-
-constexpr uint128 operator+(uint128 lhs, uint128 rhs) {
-#if defined(ABSL_HAVE_INTRINSIC_INT128)
-  return static_cast<unsigned __int128>(lhs) +
-         static_cast<unsigned __int128>(rhs);
-#else
-  return int128_internal::AddResult(
-      MakeUint128(Uint128High64(lhs) + Uint128High64(rhs),
-                  Uint128Low64(lhs) + Uint128Low64(rhs)),
-      lhs);
-#endif
-}
-
-#if !defined(ABSL_HAVE_INTRINSIC_INT128)
-namespace int128_internal {
-constexpr uint128 SubstructResult(uint128 result, uint128 lhs, uint128 rhs) {
-  // check for carry
-  return (Uint128Low64(lhs) < Uint128Low64(rhs))
-             ? MakeUint128(Uint128High64(result) - 1, Uint128Low64(result))
-             : result;
-}
-}  // namespace int128_internal
-#endif
-
-constexpr uint128 operator-(uint128 lhs, uint128 rhs) {
-#if defined(ABSL_HAVE_INTRINSIC_INT128)
-  return static_cast<unsigned __int128>(lhs) -
-         static_cast<unsigned __int128>(rhs);
-#else
-  return int128_internal::SubstructResult(
-      MakeUint128(Uint128High64(lhs) - Uint128High64(rhs),
-                  Uint128Low64(lhs) - Uint128Low64(rhs)),
-      lhs, rhs);
-#endif
-}
-
-inline uint128 operator*(uint128 lhs, uint128 rhs) {
-#if defined(ABSL_HAVE_INTRINSIC_INT128)
-  // TODO(strel) Remove once alignment issues are resolved and unsigned __int128
-  // can be used for uint128 storage.
-  return static_cast<unsigned __int128>(lhs) *
-         static_cast<unsigned __int128>(rhs);
-#elif defined(_MSC_VER) && defined(_M_X64) && !defined(_M_ARM64EC)
-  uint64_t carry;
-  uint64_t low = _umul128(Uint128Low64(lhs), Uint128Low64(rhs), &carry);
-  return MakeUint128(Uint128Low64(lhs) * Uint128High64(rhs) +
-                         Uint128High64(lhs) * Uint128Low64(rhs) + carry,
-                     low);
-#else   // ABSL_HAVE_INTRINSIC128
-  uint64_t a32 = Uint128Low64(lhs) >> 32;
-  uint64_t a00 = Uint128Low64(lhs) & 0xffffffff;
-  uint64_t b32 = Uint128Low64(rhs) >> 32;
-  uint64_t b00 = Uint128Low64(rhs) & 0xffffffff;
-  uint128 result =
-      MakeUint128(Uint128High64(lhs) * Uint128Low64(rhs) +
-                      Uint128Low64(lhs) * Uint128High64(rhs) + a32 * b32,
-                  a00 * b00);
-  result += uint128(a32 * b00) << 32;
-  result += uint128(a00 * b32) << 32;
-  return result;
-#endif  // ABSL_HAVE_INTRINSIC128
-}
-
-#if defined(ABSL_HAVE_INTRINSIC_INT128)
-inline uint128 operator/(uint128 lhs, uint128 rhs) {
-  return static_cast<unsigned __int128>(lhs) /
-         static_cast<unsigned __int128>(rhs);
-}
-
-inline uint128 operator%(uint128 lhs, uint128 rhs) {
-  return static_cast<unsigned __int128>(lhs) %
-         static_cast<unsigned __int128>(rhs);
-}
-#endif
-
-// Increment/decrement operators.
-
-inline uint128 uint128::operator++(int) {
-  uint128 tmp(*this);
-  *this += 1;
-  return tmp;
-}
-
-inline uint128 uint128::operator--(int) {
-  uint128 tmp(*this);
-  *this -= 1;
-  return tmp;
-}
-
-inline uint128& uint128::operator++() {
-  *this += 1;
-  return *this;
-}
-
-inline uint128& uint128::operator--() {
-  *this -= 1;
-  return *this;
-}
-
-constexpr int128 MakeInt128(int64_t high, uint64_t low) {
-  return int128(high, low);
-}
-
-// Assignment from integer types.
-inline int128& int128::operator=(int v) { return *this = int128(v); }
-
-inline int128& int128::operator=(unsigned int v) { return *this = int128(v); }
-
-inline int128& int128::operator=(long v) {  // NOLINT(runtime/int)
-  return *this = int128(v);
-}
-
-// NOLINTNEXTLINE(runtime/int)
-inline int128& int128::operator=(unsigned long v) { return *this = int128(v); }
-
-// NOLINTNEXTLINE(runtime/int)
-inline int128& int128::operator=(long long v) { return *this = int128(v); }
-
-// NOLINTNEXTLINE(runtime/int)
-inline int128& int128::operator=(unsigned long long v) {
-  return *this = int128(v);
-}
-
-// Arithmetic operators.
-constexpr int128 operator-(int128 v);
-constexpr int128 operator+(int128 lhs, int128 rhs);
-constexpr int128 operator-(int128 lhs, int128 rhs);
-int128 operator*(int128 lhs, int128 rhs);
-int128 operator/(int128 lhs, int128 rhs);
-int128 operator%(int128 lhs, int128 rhs);
-constexpr int128 operator|(int128 lhs, int128 rhs);
-constexpr int128 operator&(int128 lhs, int128 rhs);
-constexpr int128 operator^(int128 lhs, int128 rhs);
-constexpr int128 operator<<(int128 lhs, int amount);
-constexpr int128 operator>>(int128 lhs, int amount);
-
-inline int128& int128::operator+=(int128 other) {
-  *this = *this + other;
-  return *this;
-}
-
-inline int128& int128::operator-=(int128 other) {
-  *this = *this - other;
-  return *this;
-}
-
-inline int128& int128::operator*=(int128 other) {
-  *this = *this * other;
-  return *this;
-}
-
-inline int128& int128::operator/=(int128 other) {
-  *this = *this / other;
-  return *this;
-}
-
-inline int128& int128::operator%=(int128 other) {
-  *this = *this % other;
-  return *this;
-}
-
-inline int128& int128::operator|=(int128 other) {
-  *this = *this | other;
-  return *this;
-}
-
-inline int128& int128::operator&=(int128 other) {
-  *this = *this & other;
-  return *this;
-}
-
-inline int128& int128::operator^=(int128 other) {
-  *this = *this ^ other;
-  return *this;
-}
-
-inline int128& int128::operator<<=(int amount) {
-  *this = *this << amount;
-  return *this;
-}
-
-inline int128& int128::operator>>=(int amount) {
-  *this = *this >> amount;
-  return *this;
-}
-
-// Forward declaration for comparison operators.
-constexpr bool operator!=(int128 lhs, int128 rhs);
-
-namespace int128_internal {
-
-// Casts from unsigned to signed while preserving the underlying binary
-// representation.
-constexpr int64_t BitCastToSigned(uint64_t v) {
-  // Casting an unsigned integer to a signed integer of the same
-  // width is implementation defined behavior if the source value would not fit
-  // in the destination type. We step around it with a roundtrip bitwise not
-  // operation to make sure this function remains constexpr. Clang, GCC, and
-  // MSVC optimize this to a no-op on x86-64.
-  return v & (uint64_t{1} << 63) ? ~static_cast<int64_t>(~v)
-                                 : static_cast<int64_t>(v);
-}
-
-}  // namespace int128_internal
-
-#if defined(ABSL_HAVE_INTRINSIC_INT128)
-#include "absl/numeric/int128_have_intrinsic.inc"  // IWYU pragma: export
-#else  // ABSL_HAVE_INTRINSIC_INT128
-#include "absl/numeric/int128_no_intrinsic.inc"  // IWYU pragma: export
-#endif  // ABSL_HAVE_INTRINSIC_INT128
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#undef ABSL_INTERNAL_WCHAR_T
-
-#endif  // ABSL_NUMERIC_INT128_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/numeric/internal/bits.h b/projects/llm_framework/include/abseil-cpp/absl/numeric/internal/bits.h
deleted file mode 100644
index bfef06bc..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/numeric/internal/bits.h
+++ /dev/null
@@ -1,358 +0,0 @@
-// Copyright 2020 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_NUMERIC_INTERNAL_BITS_H_
-#define ABSL_NUMERIC_INTERNAL_BITS_H_
-
-#include <cstdint>
-#include <limits>
-#include <type_traits>
-
-// Clang on Windows has __builtin_clzll; otherwise we need to use the
-// windows intrinsic functions.
-#if defined(_MSC_VER) && !defined(__clang__)
-#include <intrin.h>
-#endif
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-
-#if defined(__GNUC__) && !defined(__clang__)
-// GCC
-#define ABSL_NUMERIC_INTERNAL_HAVE_BUILTIN_OR_GCC(x) 1
-#else
-#define ABSL_NUMERIC_INTERNAL_HAVE_BUILTIN_OR_GCC(x) ABSL_HAVE_BUILTIN(x)
-#endif
-
-#if ABSL_NUMERIC_INTERNAL_HAVE_BUILTIN_OR_GCC(__builtin_popcountl) && \
-    ABSL_NUMERIC_INTERNAL_HAVE_BUILTIN_OR_GCC(__builtin_popcountll)
-#define ABSL_INTERNAL_CONSTEXPR_POPCOUNT constexpr
-#define ABSL_INTERNAL_HAS_CONSTEXPR_POPCOUNT 1
-#else
-#define ABSL_INTERNAL_CONSTEXPR_POPCOUNT
-#define ABSL_INTERNAL_HAS_CONSTEXPR_POPCOUNT 0
-#endif
-
-#if ABSL_NUMERIC_INTERNAL_HAVE_BUILTIN_OR_GCC(__builtin_clz) && \
-    ABSL_NUMERIC_INTERNAL_HAVE_BUILTIN_OR_GCC(__builtin_clzll)
-#define ABSL_INTERNAL_CONSTEXPR_CLZ constexpr
-#define ABSL_INTERNAL_HAS_CONSTEXPR_CLZ 1
-#else
-#define ABSL_INTERNAL_CONSTEXPR_CLZ
-#define ABSL_INTERNAL_HAS_CONSTEXPR_CLZ 0
-#endif
-
-#if ABSL_NUMERIC_INTERNAL_HAVE_BUILTIN_OR_GCC(__builtin_ctz) && \
-    ABSL_NUMERIC_INTERNAL_HAVE_BUILTIN_OR_GCC(__builtin_ctzll)
-#define ABSL_INTERNAL_CONSTEXPR_CTZ constexpr
-#define ABSL_INTERNAL_HAS_CONSTEXPR_CTZ 1
-#else
-#define ABSL_INTERNAL_CONSTEXPR_CTZ
-#define ABSL_INTERNAL_HAS_CONSTEXPR_CTZ 0
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace numeric_internal {
-
-constexpr bool IsPowerOf2(unsigned int x) noexcept {
-  return x != 0 && (x & (x - 1)) == 0;
-}
-
-template <class T>
-ABSL_MUST_USE_RESULT ABSL_ATTRIBUTE_ALWAYS_INLINE constexpr T RotateRight(
-    T x, int s) noexcept {
-  static_assert(std::is_unsigned<T>::value, "T must be unsigned");
-  static_assert(IsPowerOf2(std::numeric_limits<T>::digits),
-                "T must have a power-of-2 size");
-
-  return static_cast<T>(x >> (s & (std::numeric_limits<T>::digits - 1))) |
-         static_cast<T>(x << ((-s) & (std::numeric_limits<T>::digits - 1)));
-}
-
-template <class T>
-ABSL_MUST_USE_RESULT ABSL_ATTRIBUTE_ALWAYS_INLINE constexpr T RotateLeft(
-    T x, int s) noexcept {
-  static_assert(std::is_unsigned<T>::value, "T must be unsigned");
-  static_assert(IsPowerOf2(std::numeric_limits<T>::digits),
-                "T must have a power-of-2 size");
-
-  return static_cast<T>(x << (s & (std::numeric_limits<T>::digits - 1))) |
-         static_cast<T>(x >> ((-s) & (std::numeric_limits<T>::digits - 1)));
-}
-
-ABSL_ATTRIBUTE_ALWAYS_INLINE ABSL_INTERNAL_CONSTEXPR_POPCOUNT inline int
-Popcount32(uint32_t x) noexcept {
-#if ABSL_NUMERIC_INTERNAL_HAVE_BUILTIN_OR_GCC(__builtin_popcount)
-  static_assert(sizeof(unsigned int) == sizeof(x),
-                "__builtin_popcount does not take 32-bit arg");
-  return __builtin_popcount(x);
-#else
-  x -= ((x >> 1) & 0x55555555);
-  x = ((x >> 2) & 0x33333333) + (x & 0x33333333);
-  return static_cast<int>((((x + (x >> 4)) & 0xF0F0F0F) * 0x1010101) >> 24);
-#endif
-}
-
-ABSL_ATTRIBUTE_ALWAYS_INLINE ABSL_INTERNAL_CONSTEXPR_POPCOUNT inline int
-Popcount64(uint64_t x) noexcept {
-#if ABSL_NUMERIC_INTERNAL_HAVE_BUILTIN_OR_GCC(__builtin_popcountll)
-  static_assert(sizeof(unsigned long long) == sizeof(x),  // NOLINT(runtime/int)
-                "__builtin_popcount does not take 64-bit arg");
-  return __builtin_popcountll(x);
-#else
-  x -= (x >> 1) & 0x5555555555555555ULL;
-  x = ((x >> 2) & 0x3333333333333333ULL) + (x & 0x3333333333333333ULL);
-  return static_cast<int>(
-      (((x + (x >> 4)) & 0xF0F0F0F0F0F0F0FULL) * 0x101010101010101ULL) >> 56);
-#endif
-}
-
-template <class T>
-ABSL_ATTRIBUTE_ALWAYS_INLINE ABSL_INTERNAL_CONSTEXPR_POPCOUNT inline int
-Popcount(T x) noexcept {
-  static_assert(std::is_unsigned<T>::value, "T must be unsigned");
-  static_assert(IsPowerOf2(std::numeric_limits<T>::digits),
-                "T must have a power-of-2 size");
-  static_assert(sizeof(x) <= sizeof(uint64_t), "T is too large");
-  return sizeof(x) <= sizeof(uint32_t) ? Popcount32(x) : Popcount64(x);
-}
-
-ABSL_ATTRIBUTE_ALWAYS_INLINE ABSL_INTERNAL_CONSTEXPR_CLZ inline int
-CountLeadingZeroes32(uint32_t x) {
-#if ABSL_NUMERIC_INTERNAL_HAVE_BUILTIN_OR_GCC(__builtin_clz)
-  // Use __builtin_clz, which uses the following instructions:
-  //  x86: bsr, lzcnt
-  //  ARM64: clz
-  //  PPC: cntlzd
-
-  static_assert(sizeof(unsigned int) == sizeof(x),
-                "__builtin_clz does not take 32-bit arg");
-  // Handle 0 as a special case because __builtin_clz(0) is undefined.
-  return x == 0 ? 32 : __builtin_clz(x);
-#elif defined(_MSC_VER) && !defined(__clang__)
-  unsigned long result = 0;  // NOLINT(runtime/int)
-  if (_BitScanReverse(&result, x)) {
-    return 31 - result;
-  }
-  return 32;
-#else
-  int zeroes = 28;
-  if (x >> 16) {
-    zeroes -= 16;
-    x >>= 16;
-  }
-  if (x >> 8) {
-    zeroes -= 8;
-    x >>= 8;
-  }
-  if (x >> 4) {
-    zeroes -= 4;
-    x >>= 4;
-  }
-  return "\4\3\2\2\1\1\1\1\0\0\0\0\0\0\0"[x] + zeroes;
-#endif
-}
-
-ABSL_ATTRIBUTE_ALWAYS_INLINE ABSL_INTERNAL_CONSTEXPR_CLZ inline int
-CountLeadingZeroes16(uint16_t x) {
-#if ABSL_HAVE_BUILTIN(__builtin_clzs)
-  static_assert(sizeof(unsigned short) == sizeof(x),  // NOLINT(runtime/int)
-                "__builtin_clzs does not take 16-bit arg");
-  return x == 0 ? 16 : __builtin_clzs(x);
-#else
-  return CountLeadingZeroes32(x) - 16;
-#endif
-}
-
-ABSL_ATTRIBUTE_ALWAYS_INLINE ABSL_INTERNAL_CONSTEXPR_CLZ inline int
-CountLeadingZeroes64(uint64_t x) {
-#if ABSL_NUMERIC_INTERNAL_HAVE_BUILTIN_OR_GCC(__builtin_clzll)
-  // Use __builtin_clzll, which uses the following instructions:
-  //  x86: bsr, lzcnt
-  //  ARM64: clz
-  //  PPC: cntlzd
-  static_assert(sizeof(unsigned long long) == sizeof(x),  // NOLINT(runtime/int)
-                "__builtin_clzll does not take 64-bit arg");
-
-  // Handle 0 as a special case because __builtin_clzll(0) is undefined.
-  return x == 0 ? 64 : __builtin_clzll(x);
-#elif defined(_MSC_VER) && !defined(__clang__) && \
-    (defined(_M_X64) || defined(_M_ARM64))
-  // MSVC does not have __buitin_clzll. Use _BitScanReverse64.
-  unsigned long result = 0;  // NOLINT(runtime/int)
-  if (_BitScanReverse64(&result, x)) {
-    return 63 - result;
-  }
-  return 64;
-#elif defined(_MSC_VER) && !defined(__clang__)
-  // MSVC does not have __buitin_clzll. Compose two calls to _BitScanReverse
-  unsigned long result = 0;  // NOLINT(runtime/int)
-  if ((x >> 32) &&
-      _BitScanReverse(&result, static_cast<unsigned long>(x >> 32))) {
-    return 31 - result;
-  }
-  if (_BitScanReverse(&result, static_cast<unsigned long>(x))) {
-    return 63 - result;
-  }
-  return 64;
-#else
-  int zeroes = 60;
-  if (x >> 32) {
-    zeroes -= 32;
-    x >>= 32;
-  }
-  if (x >> 16) {
-    zeroes -= 16;
-    x >>= 16;
-  }
-  if (x >> 8) {
-    zeroes -= 8;
-    x >>= 8;
-  }
-  if (x >> 4) {
-    zeroes -= 4;
-    x >>= 4;
-  }
-  return "\4\3\2\2\1\1\1\1\0\0\0\0\0\0\0"[x] + zeroes;
-#endif
-}
-
-template <typename T>
-ABSL_ATTRIBUTE_ALWAYS_INLINE ABSL_INTERNAL_CONSTEXPR_CLZ inline int
-CountLeadingZeroes(T x) {
-  static_assert(std::is_unsigned<T>::value, "T must be unsigned");
-  static_assert(IsPowerOf2(std::numeric_limits<T>::digits),
-                "T must have a power-of-2 size");
-  static_assert(sizeof(T) <= sizeof(uint64_t), "T too large");
-  return sizeof(T) <= sizeof(uint16_t)
-             ? CountLeadingZeroes16(static_cast<uint16_t>(x)) -
-                   (std::numeric_limits<uint16_t>::digits -
-                    std::numeric_limits<T>::digits)
-             : (sizeof(T) <= sizeof(uint32_t)
-                    ? CountLeadingZeroes32(static_cast<uint32_t>(x)) -
-                          (std::numeric_limits<uint32_t>::digits -
-                           std::numeric_limits<T>::digits)
-                    : CountLeadingZeroes64(x));
-}
-
-ABSL_ATTRIBUTE_ALWAYS_INLINE ABSL_INTERNAL_CONSTEXPR_CTZ inline int
-CountTrailingZeroesNonzero32(uint32_t x) {
-#if ABSL_NUMERIC_INTERNAL_HAVE_BUILTIN_OR_GCC(__builtin_ctz)
-  static_assert(sizeof(unsigned int) == sizeof(x),
-                "__builtin_ctz does not take 32-bit arg");
-  return __builtin_ctz(x);
-#elif defined(_MSC_VER) && !defined(__clang__)
-  unsigned long result = 0;  // NOLINT(runtime/int)
-  _BitScanForward(&result, x);
-  return result;
-#else
-  int c = 31;
-  x &= ~x + 1;
-  if (x & 0x0000FFFF) c -= 16;
-  if (x & 0x00FF00FF) c -= 8;
-  if (x & 0x0F0F0F0F) c -= 4;
-  if (x & 0x33333333) c -= 2;
-  if (x & 0x55555555) c -= 1;
-  return c;
-#endif
-}
-
-ABSL_ATTRIBUTE_ALWAYS_INLINE ABSL_INTERNAL_CONSTEXPR_CTZ inline int
-CountTrailingZeroesNonzero64(uint64_t x) {
-#if ABSL_NUMERIC_INTERNAL_HAVE_BUILTIN_OR_GCC(__builtin_ctzll)
-  static_assert(sizeof(unsigned long long) == sizeof(x),  // NOLINT(runtime/int)
-                "__builtin_ctzll does not take 64-bit arg");
-  return __builtin_ctzll(x);
-#elif defined(_MSC_VER) && !defined(__clang__) && \
-    (defined(_M_X64) || defined(_M_ARM64))
-  unsigned long result = 0;  // NOLINT(runtime/int)
-  _BitScanForward64(&result, x);
-  return result;
-#elif defined(_MSC_VER) && !defined(__clang__)
-  unsigned long result = 0;  // NOLINT(runtime/int)
-  if (static_cast<uint32_t>(x) == 0) {
-    _BitScanForward(&result, static_cast<unsigned long>(x >> 32));
-    return result + 32;
-  }
-  _BitScanForward(&result, static_cast<unsigned long>(x));
-  return result;
-#else
-  int c = 63;
-  x &= ~x + 1;
-  if (x & 0x00000000FFFFFFFF) c -= 32;
-  if (x & 0x0000FFFF0000FFFF) c -= 16;
-  if (x & 0x00FF00FF00FF00FF) c -= 8;
-  if (x & 0x0F0F0F0F0F0F0F0F) c -= 4;
-  if (x & 0x3333333333333333) c -= 2;
-  if (x & 0x5555555555555555) c -= 1;
-  return c;
-#endif
-}
-
-ABSL_ATTRIBUTE_ALWAYS_INLINE ABSL_INTERNAL_CONSTEXPR_CTZ inline int
-CountTrailingZeroesNonzero16(uint16_t x) {
-#if ABSL_HAVE_BUILTIN(__builtin_ctzs)
-  static_assert(sizeof(unsigned short) == sizeof(x),  // NOLINT(runtime/int)
-                "__builtin_ctzs does not take 16-bit arg");
-  return __builtin_ctzs(x);
-#else
-  return CountTrailingZeroesNonzero32(x);
-#endif
-}
-
-template <class T>
-ABSL_ATTRIBUTE_ALWAYS_INLINE ABSL_INTERNAL_CONSTEXPR_CTZ inline int
-CountTrailingZeroes(T x) noexcept {
-  static_assert(std::is_unsigned<T>::value, "T must be unsigned");
-  static_assert(IsPowerOf2(std::numeric_limits<T>::digits),
-                "T must have a power-of-2 size");
-  static_assert(sizeof(T) <= sizeof(uint64_t), "T too large");
-  return x == 0 ? std::numeric_limits<T>::digits
-                : (sizeof(T) <= sizeof(uint16_t)
-                       ? CountTrailingZeroesNonzero16(static_cast<uint16_t>(x))
-                       : (sizeof(T) <= sizeof(uint32_t)
-                              ? CountTrailingZeroesNonzero32(
-                                    static_cast<uint32_t>(x))
-                              : CountTrailingZeroesNonzero64(x)));
-}
-
-// If T is narrower than unsigned, T{1} << bit_width will be promoted.  We
-// want to force it to wraparound so that bit_ceil of an invalid value are not
-// core constant expressions.
-template <class T>
-ABSL_ATTRIBUTE_ALWAYS_INLINE ABSL_INTERNAL_CONSTEXPR_CLZ inline
-    typename std::enable_if<std::is_unsigned<T>::value, T>::type
-    BitCeilPromotionHelper(T x, T promotion) {
-  return (T{1} << (x + promotion)) >> promotion;
-}
-
-template <class T>
-ABSL_ATTRIBUTE_ALWAYS_INLINE ABSL_INTERNAL_CONSTEXPR_CLZ inline
-    typename std::enable_if<std::is_unsigned<T>::value, T>::type
-    BitCeilNonPowerOf2(T x) {
-  // If T is narrower than unsigned, it undergoes promotion to unsigned when we
-  // shift.  We calculate the number of bits added by the wider type.
-  return BitCeilPromotionHelper(
-      static_cast<T>(std::numeric_limits<T>::digits - CountLeadingZeroes(x)),
-      T{sizeof(T) >= sizeof(unsigned) ? 0
-                                      : std::numeric_limits<unsigned>::digits -
-                                            std::numeric_limits<T>::digits});
-}
-
-}  // namespace numeric_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_NUMERIC_INTERNAL_BITS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/numeric/internal/representation.h b/projects/llm_framework/include/abseil-cpp/absl/numeric/internal/representation.h
deleted file mode 100644
index 82d332fd..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/numeric/internal/representation.h
+++ /dev/null
@@ -1,55 +0,0 @@
-// Copyright 2021 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_NUMERIC_INTERNAL_REPRESENTATION_H_
-#define ABSL_NUMERIC_INTERNAL_REPRESENTATION_H_
-
-#include <limits>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace numeric_internal {
-
-// Returns true iff long double is represented as a pair of doubles added
-// together.
-inline constexpr bool IsDoubleDouble() {
-  // A double-double value always has exactly twice the precision of a double
-  // value--one double carries the high digits and one double carries the low
-  // digits. This property is not shared with any other common floating-point
-  // representation, so this test won't trigger false positives. For reference,
-  // this table gives the number of bits of precision of each common
-  // floating-point representation:
-  //
-  //                type     precision
-  //         IEEE single          24 b
-  //         IEEE double          53
-  //     x86 long double          64
-  //       double-double         106
-  //      IEEE quadruple         113
-  //
-  // Note in particular that a quadruple-precision float has greater precision
-  // than a double-double float despite taking up the same amount of memory; the
-  // quad has more of its bits allocated to the mantissa than the double-double
-  // has.
-  return std::numeric_limits<long double>::digits ==
-         2 * std::numeric_limits<double>::digits;
-}
-
-}  // namespace numeric_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_NUMERIC_INTERNAL_REPRESENTATION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/profiling/internal/exponential_biased.h b/projects/llm_framework/include/abseil-cpp/absl/profiling/internal/exponential_biased.h
deleted file mode 100644
index d31f7782..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/profiling/internal/exponential_biased.h
+++ /dev/null
@@ -1,130 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_PROFILING_INTERNAL_EXPONENTIAL_BIASED_H_
-#define ABSL_PROFILING_INTERNAL_EXPONENTIAL_BIASED_H_
-
-#include <stdint.h>
-
-#include "absl/base/config.h"
-#include "absl/base/macros.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace profiling_internal {
-
-// ExponentialBiased provides a small and fast random number generator for a
-// rounded exponential distribution. This generator manages very little state,
-// and imposes no synchronization overhead. This makes it useful in specialized
-// scenarios requiring minimum overhead, such as stride based periodic sampling.
-//
-// ExponentialBiased provides two closely related functions, GetSkipCount() and
-// GetStride(), both returning a rounded integer defining a number of events
-// required before some event with a given mean probability occurs.
-//
-// The distribution is useful to generate a random wait time or some periodic
-// event with a given mean probability. For example, if an action is supposed to
-// happen on average once every 'N' events, then we can get a random 'stride'
-// counting down how long before the event to happen. For example, if we'd want
-// to sample one in every 1000 'Frobber' calls, our code could look like this:
-//
-//   Frobber::Frobber() {
-//     stride_ = exponential_biased_.GetStride(1000);
-//   }
-//
-//   void Frobber::Frob(int arg) {
-//     if (--stride == 0) {
-//       SampleFrob(arg);
-//       stride_ = exponential_biased_.GetStride(1000);
-//     }
-//     ...
-//   }
-//
-// The rounding of the return value creates a bias, especially for smaller means
-// where the distribution of the fraction is not evenly distributed. We correct
-// this bias by tracking the fraction we rounded up or down on each iteration,
-// effectively tracking the distance between the cumulative value, and the
-// rounded cumulative value. For example, given a mean of 2:
-//
-//   raw = 1.63076, cumulative = 1.63076, rounded = 2, bias = -0.36923
-//   raw = 0.14624, cumulative = 1.77701, rounded = 2, bias =  0.14624
-//   raw = 4.93194, cumulative = 6.70895, rounded = 7, bias = -0.06805
-//   raw = 0.24206, cumulative = 6.95101, rounded = 7, bias =  0.24206
-//   etc...
-//
-// Adjusting with rounding bias is relatively trivial:
-//
-//    double value = bias_ + exponential_distribution(mean)();
-//    double rounded_value = std::rint(value);
-//    bias_ = value - rounded_value;
-//    return rounded_value;
-//
-// This class is thread-compatible.
-class ExponentialBiased {
- public:
-  // The number of bits set by NextRandom.
-  static constexpr int kPrngNumBits = 48;
-
-  // `GetSkipCount()` returns the number of events to skip before some chosen
-  // event happens. For example, randomly tossing a coin, we will on average
-  // throw heads once before we get tails. We can simulate random coin tosses
-  // using GetSkipCount() as:
-  //
-  //   ExponentialBiased eb;
-  //   for (...) {
-  //     int number_of_heads_before_tail = eb.GetSkipCount(1);
-  //     for (int flips = 0; flips < number_of_heads_before_tail; ++flips) {
-  //       printf("head...");
-  //     }
-  //     printf("tail\n");
-  //   }
-  //
-  int64_t GetSkipCount(int64_t mean);
-
-  // GetStride() returns the number of events required for a specific event to
-  // happen. See the class comments for a usage example. `GetStride()` is
-  // equivalent to `GetSkipCount(mean - 1) + 1`. When to use `GetStride()` or
-  // `GetSkipCount()` depends mostly on what best fits the use case.
-  int64_t GetStride(int64_t mean);
-
-  // Computes a random number in the range [0, 1<<(kPrngNumBits+1) - 1]
-  //
-  // This is public to enable testing.
-  static uint64_t NextRandom(uint64_t rnd);
-
- private:
-  void Initialize();
-
-  uint64_t rng_{0};
-  double bias_{0};
-  bool initialized_{false};
-};
-
-// Returns the next prng value.
-// pRNG is: aX+b mod c with a = 0x5DEECE66D, b =  0xB, c = 1<<48
-// This is the lrand64 generator.
-inline uint64_t ExponentialBiased::NextRandom(uint64_t rnd) {
-  const uint64_t prng_mult = uint64_t{0x5DEECE66D};
-  const uint64_t prng_add = 0xB;
-  const uint64_t prng_mod_power = 48;
-  const uint64_t prng_mod_mask =
-      ~((~static_cast<uint64_t>(0)) << prng_mod_power);
-  return (prng_mult * rnd + prng_add) & prng_mod_mask;
-}
-
-}  // namespace profiling_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_PROFILING_INTERNAL_EXPONENTIAL_BIASED_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/profiling/internal/periodic_sampler.h b/projects/llm_framework/include/abseil-cpp/absl/profiling/internal/periodic_sampler.h
deleted file mode 100644
index 54f0af45..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/profiling/internal/periodic_sampler.h
+++ /dev/null
@@ -1,211 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_PROFILING_INTERNAL_PERIODIC_SAMPLER_H_
-#define ABSL_PROFILING_INTERNAL_PERIODIC_SAMPLER_H_
-
-#include <stdint.h>
-
-#include <atomic>
-
-#include "absl/base/optimization.h"
-#include "absl/profiling/internal/exponential_biased.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace profiling_internal {
-
-// PeriodicSamplerBase provides the basic period sampler implementation.
-//
-// This is the base class for the templated PeriodicSampler class, which holds
-// a global std::atomic value identified by a user defined tag, such that
-// each specific PeriodSampler implementation holds its own global period.
-//
-// PeriodicSamplerBase is thread-compatible except where stated otherwise.
-class PeriodicSamplerBase {
- public:
-  // PeriodicSamplerBase is trivial / copyable / movable / destructible.
-  PeriodicSamplerBase() = default;
-  PeriodicSamplerBase(PeriodicSamplerBase&&) = default;
-  PeriodicSamplerBase(const PeriodicSamplerBase&) = default;
-
-  // Returns true roughly once every `period` calls. This is established by a
-  // randomly picked `stride` that is counted down on each call to `Sample`.
-  // This stride is picked such that the probability of `Sample()` returning
-  // true is 1 in `period`.
-  inline bool Sample() noexcept;
-
-  // The below methods are intended for optimized use cases where the
-  // size of the inlined fast path code is highly important. Applications
-  // should use the `Sample()` method unless they have proof that their
-  // specific use case requires the optimizations offered by these methods.
-  //
-  // An example of such a use case is SwissTable sampling. All sampling checks
-  // are in inlined SwissTable methods, and the number of call sites is huge.
-  // In this case, the inlined code size added to each translation unit calling
-  // SwissTable methods is non-trivial.
-  //
-  // The `SubtleMaybeSample()` function spuriously returns true even if the
-  // function should not be sampled, applications MUST match each call to
-  // 'SubtleMaybeSample()' returning true with a `SubtleConfirmSample()` call,
-  // and use the result of the latter as the sampling decision.
-  // In other words: the code should logically be equivalent to:
-  //
-  //    if (SubtleMaybeSample() && SubtleConfirmSample()) {
-  //      // Sample this call
-  //    }
-  //
-  // In the 'inline-size' optimized case, the `SubtleConfirmSample()` call can
-  // be placed out of line, for example, the typical use case looks as follows:
-  //
-  //   // --- frobber.h -----------
-  //   void FrobberSampled();
-  //
-  //   inline void FrobberImpl() {
-  //     // ...
-  //   }
-  //
-  //   inline void Frobber() {
-  //     if (ABSL_PREDICT_FALSE(sampler.SubtleMaybeSample())) {
-  //       FrobberSampled();
-  //     } else {
-  //       FrobberImpl();
-  //     }
-  //   }
-  //
-  //   // --- frobber.cc -----------
-  //   void FrobberSampled() {
-  //     if (!sampler.SubtleConfirmSample())) {
-  //       // Spurious false positive
-  //       FrobberImpl();
-  //       return;
-  //     }
-  //
-  //     // Sampled execution
-  //     // ...
-  //   }
-  inline bool SubtleMaybeSample() noexcept;
-  bool SubtleConfirmSample() noexcept;
-
- protected:
-  // We explicitly don't use a virtual destructor as this class is never
-  // virtually destroyed, and it keeps the class trivial, which avoids TLS
-  // prologue and epilogue code for our TLS instances.
-  ~PeriodicSamplerBase() = default;
-
-  // Returns the next stride for our sampler.
-  // This function is virtual for testing purposes only.
-  virtual int64_t GetExponentialBiased(int period) noexcept;
-
- private:
-  // Returns the current period of this sampler. Thread-safe.
-  virtual int period() const noexcept = 0;
-
-  // Keep and decrement stride_ as an unsigned integer, but compare the value
-  // to zero casted as a signed int. clang and msvc do not create optimum code
-  // if we use signed for the combined decrement and sign comparison.
-  //
-  // Below 3 alternative options, all compiles generate the best code
-  // using the unsigned increment <---> signed int comparison option.
-  //
-  // Option 1:
-  //   int64_t stride_;
-  //   if (ABSL_PREDICT_TRUE(++stride_ < 0)) { ... }
-  //
-  //   GCC   x64 (OK) : https://gcc.godbolt.org/z/R5MzzA
-  //   GCC   ppc (OK) : https://gcc.godbolt.org/z/z7NZAt
-  //   Clang x64 (BAD): https://gcc.godbolt.org/z/t4gPsd
-  //   ICC   x64 (OK) : https://gcc.godbolt.org/z/rE6s8W
-  //   MSVC  x64 (OK) : https://gcc.godbolt.org/z/ARMXqS
-  //
-  // Option 2:
-  //   int64_t stride_ = 0;
-  //   if (ABSL_PREDICT_TRUE(--stride_ >= 0)) { ... }
-  //
-  //   GCC   x64 (OK) : https://gcc.godbolt.org/z/jSQxYK
-  //   GCC   ppc (OK) : https://gcc.godbolt.org/z/VJdYaA
-  //   Clang x64 (BAD): https://gcc.godbolt.org/z/Xm4NjX
-  //   ICC   x64 (OK) : https://gcc.godbolt.org/z/4snaFd
-  //   MSVC  x64 (BAD): https://gcc.godbolt.org/z/BgnEKE
-  //
-  // Option 3:
-  //   uint64_t stride_;
-  //   if (ABSL_PREDICT_TRUE(static_cast<int64_t>(++stride_) < 0)) { ... }
-  //
-  //   GCC   x64 (OK) : https://gcc.godbolt.org/z/bFbfPy
-  //   GCC   ppc (OK) : https://gcc.godbolt.org/z/S9KkUE
-  //   Clang x64 (OK) : https://gcc.godbolt.org/z/UYzRb4
-  //   ICC   x64 (OK) : https://gcc.godbolt.org/z/ptTNfD
-  //   MSVC  x64 (OK) : https://gcc.godbolt.org/z/76j4-5
-  uint64_t stride_ = 0;
-  absl::profiling_internal::ExponentialBiased rng_;
-};
-
-inline bool PeriodicSamplerBase::SubtleMaybeSample() noexcept {
-  // See comments on `stride_` for the unsigned increment / signed compare.
-  if (ABSL_PREDICT_TRUE(static_cast<int64_t>(++stride_) < 0)) {
-    return false;
-  }
-  return true;
-}
-
-inline bool PeriodicSamplerBase::Sample() noexcept {
-  return ABSL_PREDICT_FALSE(SubtleMaybeSample()) ? SubtleConfirmSample()
-                                                 : false;
-}
-
-// PeriodicSampler is a concreted periodic sampler implementation.
-// The user provided Tag identifies the implementation, and is required to
-// isolate the global state of this instance from other instances.
-//
-// Typical use case:
-//
-//   struct HashTablezTag {};
-//   thread_local PeriodicSampler sampler;
-//
-//   void HashTableSamplingLogic(...) {
-//     if (sampler.Sample()) {
-//       HashTableSlowSamplePath(...);
-//     }
-//   }
-//
-template <typename Tag, int default_period = 0>
-class PeriodicSampler final : public PeriodicSamplerBase {
- public:
-  ~PeriodicSampler() = default;
-
-  int period() const noexcept final {
-    return period_.load(std::memory_order_relaxed);
-  }
-
-  // Sets the global period for this sampler. Thread-safe.
-  // Setting a period of 0 disables the sampler, i.e., every call to Sample()
-  // will return false. Setting a period of 1 puts the sampler in 'always on'
-  // mode, i.e., every call to Sample() returns true.
-  static void SetGlobalPeriod(int period) {
-    period_.store(period, std::memory_order_relaxed);
-  }
-
- private:
-  static std::atomic<int> period_;
-};
-
-template <typename Tag, int default_period>
-std::atomic<int> PeriodicSampler<Tag, default_period>::period_(default_period);
-
-}  // namespace profiling_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_PROFILING_INTERNAL_PERIODIC_SAMPLER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/profiling/internal/sample_recorder.h b/projects/llm_framework/include/abseil-cpp/absl/profiling/internal/sample_recorder.h
deleted file mode 100644
index 371f6c47..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/profiling/internal/sample_recorder.h
+++ /dev/null
@@ -1,253 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: sample_recorder.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines a lock-free linked list for recording samples
-// collected from a random/stochastic process.
-//
-// This utility is internal-only. Use at your own risk.
-
-#ifndef ABSL_PROFILING_INTERNAL_SAMPLE_RECORDER_H_
-#define ABSL_PROFILING_INTERNAL_SAMPLE_RECORDER_H_
-
-#include <atomic>
-#include <cstddef>
-#include <functional>
-
-#include "absl/base/config.h"
-#include "absl/base/thread_annotations.h"
-#include "absl/synchronization/mutex.h"
-#include "absl/time/time.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace profiling_internal {
-
-// Sample<T> that has members required for linking samples in the linked list of
-// samples maintained by the SampleRecorder.  Type T defines the sampled data.
-template <typename T>
-struct Sample {
-  // Guards the ability to restore the sample to a pristine state.  This
-  // prevents races with sampling and resurrecting an object.
-  absl::Mutex init_mu;
-  T* next = nullptr;
-  T* dead ABSL_GUARDED_BY(init_mu) = nullptr;
-  int64_t weight;  // How many sampling events were required to sample this one.
-};
-
-// Holds samples and their associated stack traces with a soft limit of
-// `SetHashtablezMaxSamples()`.
-//
-// Thread safe.
-template <typename T>
-class SampleRecorder {
- public:
-  SampleRecorder();
-  ~SampleRecorder();
-
-  // Registers for sampling.  Returns an opaque registration info.
-  template <typename... Targs>
-  T* Register(Targs&&... args);
-
-  // Unregisters the sample.
-  void Unregister(T* sample);
-
-  // The dispose callback will be called on all samples the moment they are
-  // being unregistered. Only affects samples that are unregistered after the
-  // callback has been set.
-  // Returns the previous callback.
-  using DisposeCallback = void (*)(const T&);
-  DisposeCallback SetDisposeCallback(DisposeCallback f);
-
-  // Iterates over all the registered `StackInfo`s.  Returning the number of
-  // samples that have been dropped.
-  int64_t Iterate(const std::function<void(const T& stack)>& f);
-
-  size_t GetMaxSamples() const;
-  void SetMaxSamples(size_t max);
-
- private:
-  void PushNew(T* sample);
-  void PushDead(T* sample);
-  template <typename... Targs>
-  T* PopDead(Targs... args);
-
-  std::atomic<size_t> dropped_samples_;
-  std::atomic<size_t> size_estimate_;
-  std::atomic<size_t> max_samples_{1 << 20};
-
-  // Intrusive lock free linked lists for tracking samples.
-  //
-  // `all_` records all samples (they are never removed from this list) and is
-  // terminated with a `nullptr`.
-  //
-  // `graveyard_.dead` is a circular linked list.  When it is empty,
-  // `graveyard_.dead == &graveyard`.  The list is circular so that
-  // every item on it (even the last) has a non-null dead pointer.  This allows
-  // `Iterate` to determine if a given sample is live or dead using only
-  // information on the sample itself.
-  //
-  // For example, nodes [A, B, C, D, E] with [A, C, E] alive and [B, D] dead
-  // looks like this (G is the Graveyard):
-  //
-  //           +---+    +---+    +---+    +---+    +---+
-  //    all -->| A |--->| B |--->| C |--->| D |--->| E |
-  //           |   |    |   |    |   |    |   |    |   |
-  //   +---+   |   | +->|   |-+  |   | +->|   |-+  |   |
-  //   | G |   +---+ |  +---+ |  +---+ |  +---+ |  +---+
-  //   |   |         |        |        |        |
-  //   |   | --------+        +--------+        |
-  //   +---+                                    |
-  //     ^                                      |
-  //     +--------------------------------------+
-  //
-  std::atomic<T*> all_;
-  T graveyard_;
-
-  std::atomic<DisposeCallback> dispose_;
-};
-
-template <typename T>
-typename SampleRecorder<T>::DisposeCallback
-SampleRecorder<T>::SetDisposeCallback(DisposeCallback f) {
-  return dispose_.exchange(f, std::memory_order_relaxed);
-}
-
-template <typename T>
-SampleRecorder<T>::SampleRecorder()
-    : dropped_samples_(0), size_estimate_(0), all_(nullptr), dispose_(nullptr) {
-  absl::MutexLock l(&graveyard_.init_mu);
-  graveyard_.dead = &graveyard_;
-}
-
-template <typename T>
-SampleRecorder<T>::~SampleRecorder() {
-  T* s = all_.load(std::memory_order_acquire);
-  while (s != nullptr) {
-    T* next = s->next;
-    delete s;
-    s = next;
-  }
-}
-
-template <typename T>
-void SampleRecorder<T>::PushNew(T* sample) {
-  sample->next = all_.load(std::memory_order_relaxed);
-  while (!all_.compare_exchange_weak(sample->next, sample,
-                                     std::memory_order_release,
-                                     std::memory_order_relaxed)) {
-  }
-}
-
-template <typename T>
-void SampleRecorder<T>::PushDead(T* sample) {
-  if (auto* dispose = dispose_.load(std::memory_order_relaxed)) {
-    dispose(*sample);
-  }
-
-  absl::MutexLock graveyard_lock(&graveyard_.init_mu);
-  absl::MutexLock sample_lock(&sample->init_mu);
-  sample->dead = graveyard_.dead;
-  graveyard_.dead = sample;
-}
-
-template <typename T>
-template <typename... Targs>
-T* SampleRecorder<T>::PopDead(Targs... args) {
-  absl::MutexLock graveyard_lock(&graveyard_.init_mu);
-
-  // The list is circular, so eventually it collapses down to
-  //   graveyard_.dead == &graveyard_
-  // when it is empty.
-  T* sample = graveyard_.dead;
-  if (sample == &graveyard_) return nullptr;
-
-  absl::MutexLock sample_lock(&sample->init_mu);
-  graveyard_.dead = sample->dead;
-  sample->dead = nullptr;
-  sample->PrepareForSampling(std::forward<Targs>(args)...);
-  return sample;
-}
-
-template <typename T>
-template <typename... Targs>
-T* SampleRecorder<T>::Register(Targs&&... args) {
-  size_t size = size_estimate_.fetch_add(1, std::memory_order_relaxed);
-  if (size > max_samples_.load(std::memory_order_relaxed)) {
-    size_estimate_.fetch_sub(1, std::memory_order_relaxed);
-    dropped_samples_.fetch_add(1, std::memory_order_relaxed);
-    return nullptr;
-  }
-
-  T* sample = PopDead(args...);
-  if (sample == nullptr) {
-    // Resurrection failed.  Hire a new warlock.
-    sample = new T();
-    {
-      absl::MutexLock sample_lock(&sample->init_mu);
-      // If flag initialization happens to occur (perhaps in another thread)
-      // while in this block, it will lock `graveyard_` which is usually always
-      // locked before any sample. This will appear as a lock inversion.
-      // However, this code is run exactly once per sample, and this sample
-      // cannot be accessed until after it is returned from this method.  This
-      // means that this lock state can never be recreated, so we can safely
-      // inform the deadlock detector to ignore it.
-      sample->init_mu.ForgetDeadlockInfo();
-      sample->PrepareForSampling(std::forward<Targs>(args)...);
-    }
-    PushNew(sample);
-  }
-
-  return sample;
-}
-
-template <typename T>
-void SampleRecorder<T>::Unregister(T* sample) {
-  PushDead(sample);
-  size_estimate_.fetch_sub(1, std::memory_order_relaxed);
-}
-
-template <typename T>
-int64_t SampleRecorder<T>::Iterate(
-    const std::function<void(const T& stack)>& f) {
-  T* s = all_.load(std::memory_order_acquire);
-  while (s != nullptr) {
-    absl::MutexLock l(&s->init_mu);
-    if (s->dead == nullptr) {
-      f(*s);
-    }
-    s = s->next;
-  }
-
-  return dropped_samples_.load(std::memory_order_relaxed);
-}
-
-template <typename T>
-void SampleRecorder<T>::SetMaxSamples(size_t max) {
-  max_samples_.store(max, std::memory_order_release);
-}
-
-template <typename T>
-size_t SampleRecorder<T>::GetMaxSamples() const {
-  return max_samples_.load(std::memory_order_acquire);
-}
-
-}  // namespace profiling_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_PROFILING_INTERNAL_SAMPLE_RECORDER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/bernoulli_distribution.h b/projects/llm_framework/include/abseil-cpp/absl/random/bernoulli_distribution.h
deleted file mode 100644
index d81b6ae6..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/bernoulli_distribution.h
+++ /dev/null
@@ -1,200 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_BERNOULLI_DISTRIBUTION_H_
-#define ABSL_RANDOM_BERNOULLI_DISTRIBUTION_H_
-
-#include <cstdint>
-#include <istream>
-#include <limits>
-
-#include "absl/base/optimization.h"
-#include "absl/random/internal/fast_uniform_bits.h"
-#include "absl/random/internal/iostream_state_saver.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// absl::bernoulli_distribution is a drop in replacement for
-// std::bernoulli_distribution. It guarantees that (given a perfect
-// UniformRandomBitGenerator) the acceptance probability is *exactly* equal to
-// the given double.
-//
-// The implementation assumes that double is IEEE754
-class bernoulli_distribution {
- public:
-  using result_type = bool;
-
-  class param_type {
-   public:
-    using distribution_type = bernoulli_distribution;
-
-    explicit param_type(double p = 0.5) : prob_(p) {
-      assert(p >= 0.0 && p <= 1.0);
-    }
-
-    double p() const { return prob_; }
-
-    friend bool operator==(const param_type& p1, const param_type& p2) {
-      return p1.p() == p2.p();
-    }
-    friend bool operator!=(const param_type& p1, const param_type& p2) {
-      return p1.p() != p2.p();
-    }
-
-   private:
-    double prob_;
-  };
-
-  bernoulli_distribution() : bernoulli_distribution(0.5) {}
-
-  explicit bernoulli_distribution(double p) : param_(p) {}
-
-  explicit bernoulli_distribution(param_type p) : param_(p) {}
-
-  // no-op
-  void reset() {}
-
-  template <typename URBG>
-  bool operator()(URBG& g) {  // NOLINT(runtime/references)
-    return Generate(param_.p(), g);
-  }
-
-  template <typename URBG>
-  bool operator()(URBG& g,  // NOLINT(runtime/references)
-                  const param_type& param) {
-    return Generate(param.p(), g);
-  }
-
-  param_type param() const { return param_; }
-  void param(const param_type& param) { param_ = param; }
-
-  double p() const { return param_.p(); }
-
-  result_type(min)() const { return false; }
-  result_type(max)() const { return true; }
-
-  friend bool operator==(const bernoulli_distribution& d1,
-                         const bernoulli_distribution& d2) {
-    return d1.param_ == d2.param_;
-  }
-
-  friend bool operator!=(const bernoulli_distribution& d1,
-                         const bernoulli_distribution& d2) {
-    return d1.param_ != d2.param_;
-  }
-
- private:
-  static constexpr uint64_t kP32 = static_cast<uint64_t>(1) << 32;
-
-  template <typename URBG>
-  static bool Generate(double p, URBG& g);  // NOLINT(runtime/references)
-
-  param_type param_;
-};
-
-template <typename CharT, typename Traits>
-std::basic_ostream<CharT, Traits>& operator<<(
-    std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
-    const bernoulli_distribution& x) {
-  auto saver = random_internal::make_ostream_state_saver(os);
-  os.precision(random_internal::stream_precision_helper<double>::kPrecision);
-  os << x.p();
-  return os;
-}
-
-template <typename CharT, typename Traits>
-std::basic_istream<CharT, Traits>& operator>>(
-    std::basic_istream<CharT, Traits>& is,  // NOLINT(runtime/references)
-    bernoulli_distribution& x) {            // NOLINT(runtime/references)
-  auto saver = random_internal::make_istream_state_saver(is);
-  auto p = random_internal::read_floating_point<double>(is);
-  if (!is.fail()) {
-    x.param(bernoulli_distribution::param_type(p));
-  }
-  return is;
-}
-
-template <typename URBG>
-bool bernoulli_distribution::Generate(double p,
-                                      URBG& g) {  // NOLINT(runtime/references)
-  random_internal::FastUniformBits<uint32_t> fast_u32;
-
-  while (true) {
-    // There are two aspects of the definition of `c` below that are worth
-    // commenting on.  First, because `p` is in the range [0, 1], `c` is in the
-    // range [0, 2^32] which does not fit in a uint32_t and therefore requires
-    // 64 bits.
-    //
-    // Second, `c` is constructed by first casting explicitly to a signed
-    // integer and then casting explicitly to an unsigned integer of the same
-    // size.  This is done because the hardware conversion instructions produce
-    // signed integers from double; if taken as a uint64_t the conversion would
-    // be wrong for doubles greater than 2^63 (not relevant in this use-case).
-    // If converted directly to an unsigned integer, the compiler would end up
-    // emitting code to handle such large values that are not relevant due to
-    // the known bounds on `c`.  To avoid these extra instructions this
-    // implementation converts first to the signed type and then convert to
-    // unsigned (which is a no-op).
-    const uint64_t c = static_cast<uint64_t>(static_cast<int64_t>(p * kP32));
-    const uint32_t v = fast_u32(g);
-    // FAST PATH: this path fails with probability 1/2^32.  Note that simply
-    // returning v <= c would approximate P very well (up to an absolute error
-    // of 1/2^32); the slow path (taken in that range of possible error, in the
-    // case of equality) eliminates the remaining error.
-    if (ABSL_PREDICT_TRUE(v != c)) return v < c;
-
-    // It is guaranteed that `q` is strictly less than 1, because if `q` were
-    // greater than or equal to 1, the same would be true for `p`. Certainly `p`
-    // cannot be greater than 1, and if `p == 1`, then the fast path would
-    // necessary have been taken already.
-    const double q = static_cast<double>(c) / kP32;
-
-    // The probability of acceptance on the fast path is `q` and so the
-    // probability of acceptance here should be `p - q`.
-    //
-    // Note that `q` is obtained from `p` via some shifts and conversions, the
-    // upshot of which is that `q` is simply `p` with some of the
-    // least-significant bits of its mantissa set to zero. This means that the
-    // difference `p - q` will not have any rounding errors. To see why, pretend
-    // that double has 10 bits of resolution and q is obtained from `p` in such
-    // a way that the 4 least-significant bits of its mantissa are set to zero.
-    // For example:
-    //   p   = 1.1100111011 * 2^-1
-    //   q   = 1.1100110000 * 2^-1
-    // p - q = 1.011        * 2^-8
-    // The difference `p - q` has exactly the nonzero mantissa bits that were
-    // "lost" in `q` producing a number which is certainly representable in a
-    // double.
-    const double left = p - q;
-
-    // By construction, the probability of being on this slow path is 1/2^32, so
-    // P(accept in slow path) = P(accept| in slow path) * P(slow path),
-    // which means the probability of acceptance here is `1 / (left * kP32)`:
-    const double here = left * kP32;
-
-    // The simplest way to compute the result of this trial is to repeat the
-    // whole algorithm with the new probability. This terminates because even
-    // given  arbitrarily unfriendly "random" bits, each iteration either
-    // multiplies a tiny probability by 2^32 (if c == 0) or strips off some
-    // number of nonzero mantissa bits. That process is bounded.
-    if (here == 0) return false;
-    p = here;
-  }
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_BERNOULLI_DISTRIBUTION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/beta_distribution.h b/projects/llm_framework/include/abseil-cpp/absl/random/beta_distribution.h
deleted file mode 100644
index c154066f..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/beta_distribution.h
+++ /dev/null
@@ -1,427 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_BETA_DISTRIBUTION_H_
-#define ABSL_RANDOM_BETA_DISTRIBUTION_H_
-
-#include <cassert>
-#include <cmath>
-#include <istream>
-#include <limits>
-#include <ostream>
-#include <type_traits>
-
-#include "absl/meta/type_traits.h"
-#include "absl/random/internal/fast_uniform_bits.h"
-#include "absl/random/internal/fastmath.h"
-#include "absl/random/internal/generate_real.h"
-#include "absl/random/internal/iostream_state_saver.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// absl::beta_distribution:
-// Generate a floating-point variate conforming to a Beta distribution:
-//   pdf(x) \propto x^(alpha-1) * (1-x)^(beta-1),
-// where the params alpha and beta are both strictly positive real values.
-//
-// The support is the open interval (0, 1), but the return value might be equal
-// to 0 or 1, due to numerical errors when alpha and beta are very different.
-//
-// Usage note: One usage is that alpha and beta are counts of number of
-// successes and failures. When the total number of trials are large, consider
-// approximating a beta distribution with a Gaussian distribution with the same
-// mean and variance. One could use the skewness, which depends only on the
-// smaller of alpha and beta when the number of trials are sufficiently large,
-// to quantify how far a beta distribution is from the normal distribution.
-template <typename RealType = double>
-class beta_distribution {
- public:
-  using result_type = RealType;
-
-  class param_type {
-   public:
-    using distribution_type = beta_distribution;
-
-    explicit param_type(result_type alpha, result_type beta)
-        : alpha_(alpha), beta_(beta) {
-      assert(alpha >= 0);
-      assert(beta >= 0);
-      assert(alpha <= (std::numeric_limits<result_type>::max)());
-      assert(beta <= (std::numeric_limits<result_type>::max)());
-      if (alpha == 0 || beta == 0) {
-        method_ = DEGENERATE_SMALL;
-        x_ = (alpha >= beta) ? 1 : 0;
-        return;
-      }
-      // a_ = min(beta, alpha), b_ = max(beta, alpha).
-      if (beta < alpha) {
-        inverted_ = true;
-        a_ = beta;
-        b_ = alpha;
-      } else {
-        inverted_ = false;
-        a_ = alpha;
-        b_ = beta;
-      }
-      if (a_ <= 1 && b_ >= ThresholdForLargeA()) {
-        method_ = DEGENERATE_SMALL;
-        x_ = inverted_ ? result_type(1) : result_type(0);
-        return;
-      }
-      // For threshold values, see also:
-      // Evaluation of Beta Generation Algorithms, Ying-Chao Hung, et. al.
-      // February, 2009.
-      if ((b_ < 1.0 && a_ + b_ <= 1.2) || a_ <= ThresholdForSmallA()) {
-        // Choose Joehnk over Cheng when it's faster or when Cheng encounters
-        // numerical issues.
-        method_ = JOEHNK;
-        a_ = result_type(1) / alpha_;
-        b_ = result_type(1) / beta_;
-        if (std::isinf(a_) || std::isinf(b_)) {
-          method_ = DEGENERATE_SMALL;
-          x_ = inverted_ ? result_type(1) : result_type(0);
-        }
-        return;
-      }
-      if (a_ >= ThresholdForLargeA()) {
-        method_ = DEGENERATE_LARGE;
-        // Note: on PPC for long double, evaluating
-        // `std::numeric_limits::max() / ThresholdForLargeA` results in NaN.
-        result_type r = a_ / b_;
-        x_ = (inverted_ ? result_type(1) : r) / (1 + r);
-        return;
-      }
-      x_ = a_ + b_;
-      log_x_ = std::log(x_);
-      if (a_ <= 1) {
-        method_ = CHENG_BA;
-        y_ = result_type(1) / a_;
-        gamma_ = a_ + a_;
-        return;
-      }
-      method_ = CHENG_BB;
-      result_type r = (a_ - 1) / (b_ - 1);
-      y_ = std::sqrt((1 + r) / (b_ * r * 2 - r + 1));
-      gamma_ = a_ + result_type(1) / y_;
-    }
-
-    result_type alpha() const { return alpha_; }
-    result_type beta() const { return beta_; }
-
-    friend bool operator==(const param_type& a, const param_type& b) {
-      return a.alpha_ == b.alpha_ && a.beta_ == b.beta_;
-    }
-
-    friend bool operator!=(const param_type& a, const param_type& b) {
-      return !(a == b);
-    }
-
-   private:
-    friend class beta_distribution;
-
-#ifdef _MSC_VER
-    // MSVC does not have constexpr implementations for std::log and std::exp
-    // so they are computed at runtime.
-#define ABSL_RANDOM_INTERNAL_LOG_EXP_CONSTEXPR
-#else
-#define ABSL_RANDOM_INTERNAL_LOG_EXP_CONSTEXPR constexpr
-#endif
-
-    // The threshold for whether std::exp(1/a) is finite.
-    // Note that this value is quite large, and a smaller a_ is NOT abnormal.
-    static ABSL_RANDOM_INTERNAL_LOG_EXP_CONSTEXPR result_type
-    ThresholdForSmallA() {
-      return result_type(1) /
-             std::log((std::numeric_limits<result_type>::max)());
-    }
-
-    // The threshold for whether a * std::log(a) is finite.
-    static ABSL_RANDOM_INTERNAL_LOG_EXP_CONSTEXPR result_type
-    ThresholdForLargeA() {
-      return std::exp(
-          std::log((std::numeric_limits<result_type>::max)()) -
-          std::log(std::log((std::numeric_limits<result_type>::max)())) -
-          ThresholdPadding());
-    }
-
-#undef ABSL_RANDOM_INTERNAL_LOG_EXP_CONSTEXPR
-
-    // Pad the threshold for large A for long double on PPC. This is done via a
-    // template specialization below.
-    static constexpr result_type ThresholdPadding() { return 0; }
-
-    enum Method {
-      JOEHNK,    // Uses algorithm Joehnk
-      CHENG_BA,  // Uses algorithm BA in Cheng
-      CHENG_BB,  // Uses algorithm BB in Cheng
-
-      // Note: See also:
-      //   Hung et al. Evaluation of beta generation algorithms. Communications
-      //   in Statistics-Simulation and Computation 38.4 (2009): 750-770.
-      // especially:
-      //   Zechner, Heinz, and Ernst Stadlober. Generating beta variates via
-      //   patchwork rejection. Computing 50.1 (1993): 1-18.
-
-      DEGENERATE_SMALL,  // a_ is abnormally small.
-      DEGENERATE_LARGE,  // a_ is abnormally large.
-    };
-
-    result_type alpha_;
-    result_type beta_;
-
-    result_type a_;  // the smaller of {alpha, beta}, or 1.0/alpha_ in JOEHNK
-    result_type b_;  // the larger of {alpha, beta}, or 1.0/beta_ in JOEHNK
-    result_type x_;  // alpha + beta, or the result in degenerate cases
-    result_type log_x_;  // log(x_)
-    result_type y_;      // "beta" in Cheng
-    result_type gamma_;  // "gamma" in Cheng
-
-    Method method_;
-
-    // Placing this last for optimal alignment.
-    // Whether alpha_ != a_, i.e. true iff alpha_ > beta_.
-    bool inverted_;
-
-    static_assert(std::is_floating_point<RealType>::value,
-                  "Class-template absl::beta_distribution<> must be "
-                  "parameterized using a floating-point type.");
-  };
-
-  beta_distribution() : beta_distribution(1) {}
-
-  explicit beta_distribution(result_type alpha, result_type beta = 1)
-      : param_(alpha, beta) {}
-
-  explicit beta_distribution(const param_type& p) : param_(p) {}
-
-  void reset() {}
-
-  // Generating functions
-  template <typename URBG>
-  result_type operator()(URBG& g) {  // NOLINT(runtime/references)
-    return (*this)(g, param_);
-  }
-
-  template <typename URBG>
-  result_type operator()(URBG& g,  // NOLINT(runtime/references)
-                         const param_type& p);
-
-  param_type param() const { return param_; }
-  void param(const param_type& p) { param_ = p; }
-
-  result_type(min)() const { return 0; }
-  result_type(max)() const { return 1; }
-
-  result_type alpha() const { return param_.alpha(); }
-  result_type beta() const { return param_.beta(); }
-
-  friend bool operator==(const beta_distribution& a,
-                         const beta_distribution& b) {
-    return a.param_ == b.param_;
-  }
-  friend bool operator!=(const beta_distribution& a,
-                         const beta_distribution& b) {
-    return a.param_ != b.param_;
-  }
-
- private:
-  template <typename URBG>
-  result_type AlgorithmJoehnk(URBG& g,  // NOLINT(runtime/references)
-                              const param_type& p);
-
-  template <typename URBG>
-  result_type AlgorithmCheng(URBG& g,  // NOLINT(runtime/references)
-                             const param_type& p);
-
-  template <typename URBG>
-  result_type DegenerateCase(URBG& g,  // NOLINT(runtime/references)
-                             const param_type& p) {
-    if (p.method_ == param_type::DEGENERATE_SMALL && p.alpha_ == p.beta_) {
-      // Returns 0 or 1 with equal probability.
-      random_internal::FastUniformBits<uint8_t> fast_u8;
-      return static_cast<result_type>((fast_u8(g) & 0x10) !=
-                                      0);  // pick any single bit.
-    }
-    return p.x_;
-  }
-
-  param_type param_;
-  random_internal::FastUniformBits<uint64_t> fast_u64_;
-};
-
-#if defined(__powerpc64__) || defined(__PPC64__) || defined(__powerpc__) || \
-    defined(__ppc__) || defined(__PPC__)
-// PPC needs a more stringent boundary for long double.
-template <>
-constexpr long double
-beta_distribution<long double>::param_type::ThresholdPadding() {
-  return 10;
-}
-#endif
-
-template <typename RealType>
-template <typename URBG>
-typename beta_distribution<RealType>::result_type
-beta_distribution<RealType>::AlgorithmJoehnk(
-    URBG& g,  // NOLINT(runtime/references)
-    const param_type& p) {
-  using random_internal::GeneratePositiveTag;
-  using random_internal::GenerateRealFromBits;
-  using real_type =
-      absl::conditional_t<std::is_same<RealType, float>::value, float, double>;
-
-  // Based on Joehnk, M. D. Erzeugung von betaverteilten und gammaverteilten
-  // Zufallszahlen. Metrika 8.1 (1964): 5-15.
-  // This method is described in Knuth, Vol 2 (Third Edition), pp 134.
-
-  result_type u, v, x, y, z;
-  for (;;) {
-    u = GenerateRealFromBits<real_type, GeneratePositiveTag, false>(
-        fast_u64_(g));
-    v = GenerateRealFromBits<real_type, GeneratePositiveTag, false>(
-        fast_u64_(g));
-
-    // Direct method. std::pow is slow for float, so rely on the optimizer to
-    // remove the std::pow() path for that case.
-    if (!std::is_same<float, result_type>::value) {
-      x = std::pow(u, p.a_);
-      y = std::pow(v, p.b_);
-      z = x + y;
-      if (z > 1) {
-        // Reject if and only if `x + y > 1.0`
-        continue;
-      }
-      if (z > 0) {
-        // When both alpha and beta are small, x and y are both close to 0, so
-        // divide by (x+y) directly may result in nan.
-        return x / z;
-      }
-    }
-
-    // Log transform.
-    // x = log( pow(u, p.a_) ), y = log( pow(v, p.b_) )
-    // since u, v <= 1.0,  x, y < 0.
-    x = std::log(u) * p.a_;
-    y = std::log(v) * p.b_;
-    if (!std::isfinite(x) || !std::isfinite(y)) {
-      continue;
-    }
-    // z = log( pow(u, a) + pow(v, b) )
-    z = x > y ? (x + std::log(1 + std::exp(y - x)))
-              : (y + std::log(1 + std::exp(x - y)));
-    // Reject iff log(x+y) > 0.
-    if (z > 0) {
-      continue;
-    }
-    return std::exp(x - z);
-  }
-}
-
-template <typename RealType>
-template <typename URBG>
-typename beta_distribution<RealType>::result_type
-beta_distribution<RealType>::AlgorithmCheng(
-    URBG& g,  // NOLINT(runtime/references)
-    const param_type& p) {
-  using random_internal::GeneratePositiveTag;
-  using random_internal::GenerateRealFromBits;
-  using real_type =
-      absl::conditional_t<std::is_same<RealType, float>::value, float, double>;
-
-  // Based on Cheng, Russell CH. Generating beta variates with nonintegral
-  // shape parameters. Communications of the ACM 21.4 (1978): 317-322.
-  // (https://dl.acm.org/citation.cfm?id=359482).
-  static constexpr result_type kLogFour =
-      result_type(1.3862943611198906188344642429163531361);  // log(4)
-  static constexpr result_type kS =
-      result_type(2.6094379124341003746007593332261876);  // 1+log(5)
-
-  const bool use_algorithm_ba = (p.method_ == param_type::CHENG_BA);
-  result_type u1, u2, v, w, z, r, s, t, bw_inv, lhs;
-  for (;;) {
-    u1 = GenerateRealFromBits<real_type, GeneratePositiveTag, false>(
-        fast_u64_(g));
-    u2 = GenerateRealFromBits<real_type, GeneratePositiveTag, false>(
-        fast_u64_(g));
-    v = p.y_ * std::log(u1 / (1 - u1));
-    w = p.a_ * std::exp(v);
-    bw_inv = result_type(1) / (p.b_ + w);
-    r = p.gamma_ * v - kLogFour;
-    s = p.a_ + r - w;
-    z = u1 * u1 * u2;
-    if (!use_algorithm_ba && s + kS >= 5 * z) {
-      break;
-    }
-    t = std::log(z);
-    if (!use_algorithm_ba && s >= t) {
-      break;
-    }
-    lhs = p.x_ * (p.log_x_ + std::log(bw_inv)) + r;
-    if (lhs >= t) {
-      break;
-    }
-  }
-  return p.inverted_ ? (1 - w * bw_inv) : w * bw_inv;
-}
-
-template <typename RealType>
-template <typename URBG>
-typename beta_distribution<RealType>::result_type
-beta_distribution<RealType>::operator()(URBG& g,  // NOLINT(runtime/references)
-                                        const param_type& p) {
-  switch (p.method_) {
-    case param_type::JOEHNK:
-      return AlgorithmJoehnk(g, p);
-    case param_type::CHENG_BA:
-      ABSL_FALLTHROUGH_INTENDED;
-    case param_type::CHENG_BB:
-      return AlgorithmCheng(g, p);
-    default:
-      return DegenerateCase(g, p);
-  }
-}
-
-template <typename CharT, typename Traits, typename RealType>
-std::basic_ostream<CharT, Traits>& operator<<(
-    std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
-    const beta_distribution<RealType>& x) {
-  auto saver = random_internal::make_ostream_state_saver(os);
-  os.precision(random_internal::stream_precision_helper<RealType>::kPrecision);
-  os << x.alpha() << os.fill() << x.beta();
-  return os;
-}
-
-template <typename CharT, typename Traits, typename RealType>
-std::basic_istream<CharT, Traits>& operator>>(
-    std::basic_istream<CharT, Traits>& is,  // NOLINT(runtime/references)
-    beta_distribution<RealType>& x) {       // NOLINT(runtime/references)
-  using result_type = typename beta_distribution<RealType>::result_type;
-  using param_type = typename beta_distribution<RealType>::param_type;
-  result_type alpha, beta;
-
-  auto saver = random_internal::make_istream_state_saver(is);
-  alpha = random_internal::read_floating_point<result_type>(is);
-  if (is.fail()) return is;
-  beta = random_internal::read_floating_point<result_type>(is);
-  if (!is.fail()) {
-    x.param(param_type(alpha, beta));
-  }
-  return is;
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_BETA_DISTRIBUTION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/bit_gen_ref.h b/projects/llm_framework/include/abseil-cpp/absl/random/bit_gen_ref.h
deleted file mode 100644
index 2160a8dd..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/bit_gen_ref.h
+++ /dev/null
@@ -1,186 +0,0 @@
-//
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: bit_gen_ref.h
-// -----------------------------------------------------------------------------
-//
-// This header defines a bit generator "reference" class, for use in interfaces
-// that take both Abseil (e.g. `absl::BitGen`) and standard library (e.g.
-// `std::mt19937`) bit generators.
-
-#ifndef ABSL_RANDOM_BIT_GEN_REF_H_
-#define ABSL_RANDOM_BIT_GEN_REF_H_
-
-#include <limits>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/attributes.h"
-#include "absl/base/internal/fast_type_id.h"
-#include "absl/base/macros.h"
-#include "absl/meta/type_traits.h"
-#include "absl/random/internal/distribution_caller.h"
-#include "absl/random/internal/fast_uniform_bits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-template <typename URBG, typename = void, typename = void, typename = void>
-struct is_urbg : std::false_type {};
-
-template <typename URBG>
-struct is_urbg<
-    URBG,
-    absl::enable_if_t<std::is_same<
-        typename URBG::result_type,
-        typename std::decay<decltype((URBG::min)())>::type>::value>,
-    absl::enable_if_t<std::is_same<
-        typename URBG::result_type,
-        typename std::decay<decltype((URBG::max)())>::type>::value>,
-    absl::enable_if_t<std::is_same<
-        typename URBG::result_type,
-        typename std::decay<decltype(std::declval<URBG>()())>::type>::value>>
-    : std::true_type {};
-
-template <typename>
-struct DistributionCaller;
-class MockHelpers;
-
-}  // namespace random_internal
-
-// -----------------------------------------------------------------------------
-// absl::BitGenRef
-// -----------------------------------------------------------------------------
-//
-// `absl::BitGenRef` is a type-erasing class that provides a generator-agnostic
-// non-owning "reference" interface for use in place of any specific uniform
-// random bit generator (URBG). This class may be used for both Abseil
-// (e.g. `absl::BitGen`, `absl::InsecureBitGen`) and Standard library (e.g
-// `std::mt19937`, `std::minstd_rand`) bit generators.
-//
-// Like other reference classes, `absl::BitGenRef` does not own the
-// underlying bit generator, and the underlying instance must outlive the
-// `absl::BitGenRef`.
-//
-// `absl::BitGenRef` is particularly useful when used with an
-// `absl::MockingBitGen` to test specific paths in functions which use random
-// values.
-//
-// Example:
-//    void TakesBitGenRef(absl::BitGenRef gen) {
-//      int x = absl::Uniform<int>(gen, 0, 1000);
-//    }
-//
-class BitGenRef {
-  // SFINAE to detect whether the URBG type includes a member matching
-  // bool InvokeMock(base_internal::FastTypeIdType, void*, void*).
-  //
-  // These live inside BitGenRef so that they have friend access
-  // to MockingBitGen. (see similar methods in DistributionCaller).
-  template <template <class...> class Trait, class AlwaysVoid, class... Args>
-  struct detector : std::false_type {};
-  template <template <class...> class Trait, class... Args>
-  struct detector<Trait, absl::void_t<Trait<Args...>>, Args...>
-      : std::true_type {};
-
-  template <class T>
-  using invoke_mock_t = decltype(std::declval<T*>()->InvokeMock(
-      std::declval<base_internal::FastTypeIdType>(), std::declval<void*>(),
-      std::declval<void*>()));
-
-  template <typename T>
-  using HasInvokeMock = typename detector<invoke_mock_t, void, T>::type;
-
- public:
-  BitGenRef(const BitGenRef&) = default;
-  BitGenRef(BitGenRef&&) = default;
-  BitGenRef& operator=(const BitGenRef&) = default;
-  BitGenRef& operator=(BitGenRef&&) = default;
-
-  template <typename URBG, typename absl::enable_if_t<
-                               (!std::is_same<URBG, BitGenRef>::value &&
-                                random_internal::is_urbg<URBG>::value &&
-                                !HasInvokeMock<URBG>::value)>* = nullptr>
-  BitGenRef(URBG& gen ABSL_ATTRIBUTE_LIFETIME_BOUND)  // NOLINT
-      : t_erased_gen_ptr_(reinterpret_cast<uintptr_t>(&gen)),
-        mock_call_(NotAMock),
-        generate_impl_fn_(ImplFn<URBG>) {}
-
-  template <typename URBG,
-            typename absl::enable_if_t<(!std::is_same<URBG, BitGenRef>::value &&
-                                        random_internal::is_urbg<URBG>::value &&
-                                        HasInvokeMock<URBG>::value)>* = nullptr>
-  BitGenRef(URBG& gen ABSL_ATTRIBUTE_LIFETIME_BOUND)  // NOLINT
-      : t_erased_gen_ptr_(reinterpret_cast<uintptr_t>(&gen)),
-        mock_call_(&MockCall<URBG>),
-        generate_impl_fn_(ImplFn<URBG>) {}
-
-  using result_type = uint64_t;
-
-  static constexpr result_type(min)() {
-    return (std::numeric_limits<result_type>::min)();
-  }
-
-  static constexpr result_type(max)() {
-    return (std::numeric_limits<result_type>::max)();
-  }
-
-  result_type operator()() { return generate_impl_fn_(t_erased_gen_ptr_); }
-
- private:
-  using impl_fn = result_type (*)(uintptr_t);
-  using mock_call_fn = bool (*)(uintptr_t, base_internal::FastTypeIdType, void*,
-                                void*);
-
-  template <typename URBG>
-  static result_type ImplFn(uintptr_t ptr) {
-    // Ensure that the return values from operator() fill the entire
-    // range promised by result_type, min() and max().
-    absl::random_internal::FastUniformBits<result_type> fast_uniform_bits;
-    return fast_uniform_bits(*reinterpret_cast<URBG*>(ptr));
-  }
-
-  // Get a type-erased InvokeMock pointer.
-  template <typename URBG>
-  static bool MockCall(uintptr_t gen_ptr, base_internal::FastTypeIdType type,
-                       void* result, void* arg_tuple) {
-    return reinterpret_cast<URBG*>(gen_ptr)->InvokeMock(type, result,
-                                                        arg_tuple);
-  }
-  static bool NotAMock(uintptr_t, base_internal::FastTypeIdType, void*, void*) {
-    return false;
-  }
-
-  inline bool InvokeMock(base_internal::FastTypeIdType type, void* args_tuple,
-                         void* result) {
-    if (mock_call_ == NotAMock) return false;  // avoids an indirect call.
-    return mock_call_(t_erased_gen_ptr_, type, args_tuple, result);
-  }
-
-  uintptr_t t_erased_gen_ptr_;
-  mock_call_fn mock_call_;
-  impl_fn generate_impl_fn_;
-
-  template <typename>
-  friend struct ::absl::random_internal::DistributionCaller;  // for InvokeMock
-  friend class ::absl::random_internal::MockHelpers;          // for InvokeMock
-};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_BIT_GEN_REF_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/discrete_distribution.h b/projects/llm_framework/include/abseil-cpp/absl/random/discrete_distribution.h
deleted file mode 100644
index 171aa11a..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/discrete_distribution.h
+++ /dev/null
@@ -1,247 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_DISCRETE_DISTRIBUTION_H_
-#define ABSL_RANDOM_DISCRETE_DISTRIBUTION_H_
-
-#include <cassert>
-#include <cmath>
-#include <istream>
-#include <limits>
-#include <numeric>
-#include <type_traits>
-#include <utility>
-#include <vector>
-
-#include "absl/random/bernoulli_distribution.h"
-#include "absl/random/internal/iostream_state_saver.h"
-#include "absl/random/uniform_int_distribution.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// absl::discrete_distribution
-//
-// A discrete distribution produces random integers i, where 0 <= i < n
-// distributed according to the discrete probability function:
-//
-//     P(i|p0,...,pn−1)=pi
-//
-// This class is an implementation of discrete_distribution (see
-// [rand.dist.samp.discrete]).
-//
-// The algorithm used is Walker's Aliasing algorithm, described in Knuth, Vol 2.
-// absl::discrete_distribution takes O(N) time to precompute the probabilities
-// (where N is the number of possible outcomes in the distribution) at
-// construction, and then takes O(1) time for each variate generation.  Many
-// other implementations also take O(N) time to construct an ordered sequence of
-// partial sums, plus O(log N) time per variate to binary search.
-//
-template <typename IntType = int>
-class discrete_distribution {
- public:
-  using result_type = IntType;
-
-  class param_type {
-   public:
-    using distribution_type = discrete_distribution;
-
-    param_type() { init(); }
-
-    template <typename InputIterator>
-    explicit param_type(InputIterator begin, InputIterator end)
-        : p_(begin, end) {
-      init();
-    }
-
-    explicit param_type(std::initializer_list<double> weights) : p_(weights) {
-      init();
-    }
-
-    template <class UnaryOperation>
-    explicit param_type(size_t nw, double xmin, double xmax,
-                        UnaryOperation fw) {
-      if (nw > 0) {
-        p_.reserve(nw);
-        double delta = (xmax - xmin) / static_cast<double>(nw);
-        assert(delta > 0);
-        double t = delta * 0.5;
-        for (size_t i = 0; i < nw; ++i) {
-          p_.push_back(fw(xmin + i * delta + t));
-        }
-      }
-      init();
-    }
-
-    const std::vector<double>& probabilities() const { return p_; }
-    size_t n() const { return p_.size() - 1; }
-
-    friend bool operator==(const param_type& a, const param_type& b) {
-      return a.probabilities() == b.probabilities();
-    }
-
-    friend bool operator!=(const param_type& a, const param_type& b) {
-      return !(a == b);
-    }
-
-   private:
-    friend class discrete_distribution;
-
-    void init();
-
-    std::vector<double> p_;                     // normalized probabilities
-    std::vector<std::pair<double, size_t>> q_;  // (acceptance, alternate) pairs
-
-    static_assert(std::is_integral<result_type>::value,
-                  "Class-template absl::discrete_distribution<> must be "
-                  "parameterized using an integral type.");
-  };
-
-  discrete_distribution() : param_() {}
-
-  explicit discrete_distribution(const param_type& p) : param_(p) {}
-
-  template <typename InputIterator>
-  explicit discrete_distribution(InputIterator begin, InputIterator end)
-      : param_(begin, end) {}
-
-  explicit discrete_distribution(std::initializer_list<double> weights)
-      : param_(weights) {}
-
-  template <class UnaryOperation>
-  explicit discrete_distribution(size_t nw, double xmin, double xmax,
-                                 UnaryOperation fw)
-      : param_(nw, xmin, xmax, std::move(fw)) {}
-
-  void reset() {}
-
-  // generating functions
-  template <typename URBG>
-  result_type operator()(URBG& g) {  // NOLINT(runtime/references)
-    return (*this)(g, param_);
-  }
-
-  template <typename URBG>
-  result_type operator()(URBG& g,  // NOLINT(runtime/references)
-                         const param_type& p);
-
-  const param_type& param() const { return param_; }
-  void param(const param_type& p) { param_ = p; }
-
-  result_type(min)() const { return 0; }
-  result_type(max)() const {
-    return static_cast<result_type>(param_.n());
-  }  // inclusive
-
-  // NOTE [rand.dist.sample.discrete] returns a std::vector<double> not a
-  // const std::vector<double>&.
-  const std::vector<double>& probabilities() const {
-    return param_.probabilities();
-  }
-
-  friend bool operator==(const discrete_distribution& a,
-                         const discrete_distribution& b) {
-    return a.param_ == b.param_;
-  }
-  friend bool operator!=(const discrete_distribution& a,
-                         const discrete_distribution& b) {
-    return a.param_ != b.param_;
-  }
-
- private:
-  param_type param_;
-};
-
-// --------------------------------------------------------------------------
-// Implementation details only below
-// --------------------------------------------------------------------------
-
-namespace random_internal {
-
-// Using the vector `*probabilities`, whose values are the weights or
-// probabilities of an element being selected, constructs the proportional
-// probabilities used by the discrete distribution.  `*probabilities` will be
-// scaled, if necessary, so that its entries sum to a value sufficiently close
-// to 1.0.
-std::vector<std::pair<double, size_t>> InitDiscreteDistribution(
-    std::vector<double>* probabilities);
-
-}  // namespace random_internal
-
-template <typename IntType>
-void discrete_distribution<IntType>::param_type::init() {
-  if (p_.empty()) {
-    p_.push_back(1.0);
-    q_.emplace_back(1.0, 0);
-  } else {
-    assert(n() <= (std::numeric_limits<IntType>::max)());
-    q_ = random_internal::InitDiscreteDistribution(&p_);
-  }
-}
-
-template <typename IntType>
-template <typename URBG>
-typename discrete_distribution<IntType>::result_type
-discrete_distribution<IntType>::operator()(
-    URBG& g,  // NOLINT(runtime/references)
-    const param_type& p) {
-  const auto idx = absl::uniform_int_distribution<result_type>(0, p.n())(g);
-  const auto& q = p.q_[idx];
-  const bool selected = absl::bernoulli_distribution(q.first)(g);
-  return selected ? idx : static_cast<result_type>(q.second);
-}
-
-template <typename CharT, typename Traits, typename IntType>
-std::basic_ostream<CharT, Traits>& operator<<(
-    std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
-    const discrete_distribution<IntType>& x) {
-  auto saver = random_internal::make_ostream_state_saver(os);
-  const auto& probabilities = x.param().probabilities();
-  os << probabilities.size();
-
-  os.precision(random_internal::stream_precision_helper<double>::kPrecision);
-  for (const auto& p : probabilities) {
-    os << os.fill() << p;
-  }
-  return os;
-}
-
-template <typename CharT, typename Traits, typename IntType>
-std::basic_istream<CharT, Traits>& operator>>(
-    std::basic_istream<CharT, Traits>& is,  // NOLINT(runtime/references)
-    discrete_distribution<IntType>& x) {    // NOLINT(runtime/references)
-  using param_type = typename discrete_distribution<IntType>::param_type;
-  auto saver = random_internal::make_istream_state_saver(is);
-
-  size_t n;
-  std::vector<double> p;
-
-  is >> n;
-  if (is.fail()) return is;
-  if (n > 0) {
-    p.reserve(n);
-    for (IntType i = 0; i < n && !is.fail(); ++i) {
-      auto tmp = random_internal::read_floating_point<double>(is);
-      if (is.fail()) return is;
-      p.push_back(tmp);
-    }
-  }
-  x.param(param_type(p.begin(), p.end()));
-  return is;
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_DISCRETE_DISTRIBUTION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/distributions.h b/projects/llm_framework/include/abseil-cpp/absl/random/distributions.h
deleted file mode 100644
index 4e3b332e..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/distributions.h
+++ /dev/null
@@ -1,452 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: distributions.h
-// -----------------------------------------------------------------------------
-//
-// This header defines functions representing distributions, which you use in
-// combination with an Abseil random bit generator to produce random values
-// according to the rules of that distribution.
-//
-// The Abseil random library defines the following distributions within this
-// file:
-//
-//   * `absl::Uniform` for uniform (constant) distributions having constant
-//     probability
-//   * `absl::Bernoulli` for discrete distributions having exactly two outcomes
-//   * `absl::Beta` for continuous distributions parameterized through two
-//     free parameters
-//   * `absl::Exponential` for discrete distributions of events occurring
-//     continuously and independently at a constant average rate
-//   * `absl::Gaussian` (also known as "normal distributions") for continuous
-//     distributions using an associated quadratic function
-//   * `absl::LogUniform` for continuous uniform distributions where the log
-//     to the given base of all values is uniform
-//   * `absl::Poisson` for discrete probability distributions that express the
-//     probability of a given number of events occurring within a fixed interval
-//   * `absl::Zipf` for discrete probability distributions commonly used for
-//     modelling of rare events
-//
-// Prefer use of these distribution function classes over manual construction of
-// your own distribution classes, as it allows library maintainers greater
-// flexibility to change the underlying implementation in the future.
-
-#ifndef ABSL_RANDOM_DISTRIBUTIONS_H_
-#define ABSL_RANDOM_DISTRIBUTIONS_H_
-
-#include <algorithm>
-#include <cmath>
-#include <limits>
-#include <random>
-#include <type_traits>
-
-#include "absl/base/internal/inline_variable.h"
-#include "absl/random/bernoulli_distribution.h"
-#include "absl/random/beta_distribution.h"
-#include "absl/random/exponential_distribution.h"
-#include "absl/random/gaussian_distribution.h"
-#include "absl/random/internal/distribution_caller.h"  // IWYU pragma: export
-#include "absl/random/internal/uniform_helper.h"  // IWYU pragma: export
-#include "absl/random/log_uniform_int_distribution.h"
-#include "absl/random/poisson_distribution.h"
-#include "absl/random/uniform_int_distribution.h"
-#include "absl/random/uniform_real_distribution.h"
-#include "absl/random/zipf_distribution.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-ABSL_INTERNAL_INLINE_CONSTEXPR(IntervalClosedClosedTag, IntervalClosedClosed,
-                               {});
-ABSL_INTERNAL_INLINE_CONSTEXPR(IntervalClosedClosedTag, IntervalClosed, {});
-ABSL_INTERNAL_INLINE_CONSTEXPR(IntervalClosedOpenTag, IntervalClosedOpen, {});
-ABSL_INTERNAL_INLINE_CONSTEXPR(IntervalOpenOpenTag, IntervalOpenOpen, {});
-ABSL_INTERNAL_INLINE_CONSTEXPR(IntervalOpenOpenTag, IntervalOpen, {});
-ABSL_INTERNAL_INLINE_CONSTEXPR(IntervalOpenClosedTag, IntervalOpenClosed, {});
-
-// -----------------------------------------------------------------------------
-// absl::Uniform<T>(tag, bitgen, lo, hi)
-// -----------------------------------------------------------------------------
-//
-// `absl::Uniform()` produces random values of type `T` uniformly distributed in
-// a defined interval {lo, hi}. The interval `tag` defines the type of interval
-// which should be one of the following possible values:
-//
-//   * `absl::IntervalOpenOpen`
-//   * `absl::IntervalOpenClosed`
-//   * `absl::IntervalClosedOpen`
-//   * `absl::IntervalClosedClosed`
-//
-// where "open" refers to an exclusive value (excluded) from the output, while
-// "closed" refers to an inclusive value (included) from the output.
-//
-// In the absence of an explicit return type `T`, `absl::Uniform()` will deduce
-// the return type based on the provided endpoint arguments {A lo, B hi}.
-// Given these endpoints, one of {A, B} will be chosen as the return type, if
-// a type can be implicitly converted into the other in a lossless way. The
-// lack of any such implicit conversion between {A, B} will produce a
-// compile-time error
-//
-// See https://en.wikipedia.org/wiki/Uniform_distribution_(continuous)
-//
-// Example:
-//
-//   absl::BitGen bitgen;
-//
-//   // Produce a random float value between 0.0 and 1.0, inclusive
-//   auto x = absl::Uniform(absl::IntervalClosedClosed, bitgen, 0.0f, 1.0f);
-//
-//   // The most common interval of `absl::IntervalClosedOpen` is available by
-//   // default:
-//
-//   auto x = absl::Uniform(bitgen, 0.0f, 1.0f);
-//
-//   // Return-types are typically inferred from the arguments, however callers
-//   // can optionally provide an explicit return-type to the template.
-//
-//   auto x = absl::Uniform<float>(bitgen, 0, 1);
-//
-template <typename R = void, typename TagType, typename URBG>
-typename absl::enable_if_t<!std::is_same<R, void>::value, R>  //
-Uniform(TagType tag,
-        URBG&& urbg,  // NOLINT(runtime/references)
-        R lo, R hi) {
-  using gen_t = absl::decay_t<URBG>;
-  using distribution_t = random_internal::UniformDistributionWrapper<R>;
-
-  auto a = random_internal::uniform_lower_bound(tag, lo, hi);
-  auto b = random_internal::uniform_upper_bound(tag, lo, hi);
-  if (!random_internal::is_uniform_range_valid(a, b)) return lo;
-
-  return random_internal::DistributionCaller<gen_t>::template Call<
-      distribution_t>(&urbg, tag, lo, hi);
-}
-
-// absl::Uniform<T>(bitgen, lo, hi)
-//
-// Overload of `Uniform()` using the default closed-open interval of [lo, hi),
-// and returning values of type `T`
-template <typename R = void, typename URBG>
-typename absl::enable_if_t<!std::is_same<R, void>::value, R>  //
-Uniform(URBG&& urbg,  // NOLINT(runtime/references)
-        R lo, R hi) {
-  using gen_t = absl::decay_t<URBG>;
-  using distribution_t = random_internal::UniformDistributionWrapper<R>;
-  constexpr auto tag = absl::IntervalClosedOpen;
-
-  auto a = random_internal::uniform_lower_bound(tag, lo, hi);
-  auto b = random_internal::uniform_upper_bound(tag, lo, hi);
-  if (!random_internal::is_uniform_range_valid(a, b)) return lo;
-
-  return random_internal::DistributionCaller<gen_t>::template Call<
-      distribution_t>(&urbg, lo, hi);
-}
-
-// absl::Uniform(tag, bitgen, lo, hi)
-//
-// Overload of `Uniform()` using different (but compatible) lo, hi types. Note
-// that a compile-error will result if the return type cannot be deduced
-// correctly from the passed types.
-template <typename R = void, typename TagType, typename URBG, typename A,
-          typename B>
-typename absl::enable_if_t<std::is_same<R, void>::value,
-                           random_internal::uniform_inferred_return_t<A, B>>
-Uniform(TagType tag,
-        URBG&& urbg,  // NOLINT(runtime/references)
-        A lo, B hi) {
-  using gen_t = absl::decay_t<URBG>;
-  using return_t = typename random_internal::uniform_inferred_return_t<A, B>;
-  using distribution_t = random_internal::UniformDistributionWrapper<return_t>;
-
-  auto a = random_internal::uniform_lower_bound<return_t>(tag, lo, hi);
-  auto b = random_internal::uniform_upper_bound<return_t>(tag, lo, hi);
-  if (!random_internal::is_uniform_range_valid(a, b)) return lo;
-
-  return random_internal::DistributionCaller<gen_t>::template Call<
-      distribution_t>(&urbg, tag, static_cast<return_t>(lo),
-                                static_cast<return_t>(hi));
-}
-
-// absl::Uniform(bitgen, lo, hi)
-//
-// Overload of `Uniform()` using different (but compatible) lo, hi types and the
-// default closed-open interval of [lo, hi). Note that a compile-error will
-// result if the return type cannot be deduced correctly from the passed types.
-template <typename R = void, typename URBG, typename A, typename B>
-typename absl::enable_if_t<std::is_same<R, void>::value,
-                           random_internal::uniform_inferred_return_t<A, B>>
-Uniform(URBG&& urbg,  // NOLINT(runtime/references)
-        A lo, B hi) {
-  using gen_t = absl::decay_t<URBG>;
-  using return_t = typename random_internal::uniform_inferred_return_t<A, B>;
-  using distribution_t = random_internal::UniformDistributionWrapper<return_t>;
-
-  constexpr auto tag = absl::IntervalClosedOpen;
-  auto a = random_internal::uniform_lower_bound<return_t>(tag, lo, hi);
-  auto b = random_internal::uniform_upper_bound<return_t>(tag, lo, hi);
-  if (!random_internal::is_uniform_range_valid(a, b)) return lo;
-
-  return random_internal::DistributionCaller<gen_t>::template Call<
-      distribution_t>(&urbg, static_cast<return_t>(lo),
-                                static_cast<return_t>(hi));
-}
-
-// absl::Uniform<unsigned T>(bitgen)
-//
-// Overload of Uniform() using the minimum and maximum values of a given type
-// `T` (which must be unsigned), returning a value of type `unsigned T`
-template <typename R, typename URBG>
-typename absl::enable_if_t<!std::is_signed<R>::value, R>  //
-Uniform(URBG&& urbg) {  // NOLINT(runtime/references)
-  using gen_t = absl::decay_t<URBG>;
-  using distribution_t = random_internal::UniformDistributionWrapper<R>;
-
-  return random_internal::DistributionCaller<gen_t>::template Call<
-      distribution_t>(&urbg);
-}
-
-// -----------------------------------------------------------------------------
-// absl::Bernoulli(bitgen, p)
-// -----------------------------------------------------------------------------
-//
-// `absl::Bernoulli` produces a random boolean value, with probability `p`
-// (where 0.0 <= p <= 1.0) equaling `true`.
-//
-// Prefer `absl::Bernoulli` to produce boolean values over other alternatives
-// such as comparing an `absl::Uniform()` value to a specific output.
-//
-// See https://en.wikipedia.org/wiki/Bernoulli_distribution
-//
-// Example:
-//
-//   absl::BitGen bitgen;
-//   ...
-//   if (absl::Bernoulli(bitgen, 1.0/3721.0)) {
-//     std::cout << "Asteroid field navigation successful.";
-//   }
-//
-template <typename URBG>
-bool Bernoulli(URBG&& urbg,  // NOLINT(runtime/references)
-               double p) {
-  using gen_t = absl::decay_t<URBG>;
-  using distribution_t = absl::bernoulli_distribution;
-
-  return random_internal::DistributionCaller<gen_t>::template Call<
-      distribution_t>(&urbg, p);
-}
-
-// -----------------------------------------------------------------------------
-// absl::Beta<T>(bitgen, alpha, beta)
-// -----------------------------------------------------------------------------
-//
-// `absl::Beta` produces a floating point number distributed in the closed
-// interval [0,1] and parameterized by two values `alpha` and `beta` as per a
-// Beta distribution. `T` must be a floating point type, but may be inferred
-// from the types of `alpha` and `beta`.
-//
-// See https://en.wikipedia.org/wiki/Beta_distribution.
-//
-// Example:
-//
-//   absl::BitGen bitgen;
-//   ...
-//   double sample = absl::Beta(bitgen, 3.0, 2.0);
-//
-template <typename RealType, typename URBG>
-RealType Beta(URBG&& urbg,  // NOLINT(runtime/references)
-              RealType alpha, RealType beta) {
-  static_assert(
-      std::is_floating_point<RealType>::value,
-      "Template-argument 'RealType' must be a floating-point type, in "
-      "absl::Beta<RealType, URBG>(...)");
-
-  using gen_t = absl::decay_t<URBG>;
-  using distribution_t = typename absl::beta_distribution<RealType>;
-
-  return random_internal::DistributionCaller<gen_t>::template Call<
-      distribution_t>(&urbg, alpha, beta);
-}
-
-// -----------------------------------------------------------------------------
-// absl::Exponential<T>(bitgen, lambda = 1)
-// -----------------------------------------------------------------------------
-//
-// `absl::Exponential` produces a floating point number representing the
-// distance (time) between two consecutive events in a point process of events
-// occurring continuously and independently at a constant average rate. `T` must
-// be a floating point type, but may be inferred from the type of `lambda`.
-//
-// See https://en.wikipedia.org/wiki/Exponential_distribution.
-//
-// Example:
-//
-//   absl::BitGen bitgen;
-//   ...
-//   double call_length = absl::Exponential(bitgen, 7.0);
-//
-template <typename RealType, typename URBG>
-RealType Exponential(URBG&& urbg,  // NOLINT(runtime/references)
-                     RealType lambda = 1) {
-  static_assert(
-      std::is_floating_point<RealType>::value,
-      "Template-argument 'RealType' must be a floating-point type, in "
-      "absl::Exponential<RealType, URBG>(...)");
-
-  using gen_t = absl::decay_t<URBG>;
-  using distribution_t = typename absl::exponential_distribution<RealType>;
-
-  return random_internal::DistributionCaller<gen_t>::template Call<
-      distribution_t>(&urbg, lambda);
-}
-
-// -----------------------------------------------------------------------------
-// absl::Gaussian<T>(bitgen, mean = 0, stddev = 1)
-// -----------------------------------------------------------------------------
-//
-// `absl::Gaussian` produces a floating point number selected from the Gaussian
-// (ie. "Normal") distribution. `T` must be a floating point type, but may be
-// inferred from the types of `mean` and `stddev`.
-//
-// See https://en.wikipedia.org/wiki/Normal_distribution
-//
-// Example:
-//
-//   absl::BitGen bitgen;
-//   ...
-//   double giraffe_height = absl::Gaussian(bitgen, 16.3, 3.3);
-//
-template <typename RealType, typename URBG>
-RealType Gaussian(URBG&& urbg,  // NOLINT(runtime/references)
-                  RealType mean = 0, RealType stddev = 1) {
-  static_assert(
-      std::is_floating_point<RealType>::value,
-      "Template-argument 'RealType' must be a floating-point type, in "
-      "absl::Gaussian<RealType, URBG>(...)");
-
-  using gen_t = absl::decay_t<URBG>;
-  using distribution_t = typename absl::gaussian_distribution<RealType>;
-
-  return random_internal::DistributionCaller<gen_t>::template Call<
-      distribution_t>(&urbg, mean, stddev);
-}
-
-// -----------------------------------------------------------------------------
-// absl::LogUniform<T>(bitgen, lo, hi, base = 2)
-// -----------------------------------------------------------------------------
-//
-// `absl::LogUniform` produces random values distributed where the log to a
-// given base of all values is uniform in a closed interval [lo, hi]. `T` must
-// be an integral type, but may be inferred from the types of `lo` and `hi`.
-//
-// I.e., `LogUniform(0, n, b)` is uniformly distributed across buckets
-// [0], [1, b-1], [b, b^2-1] .. [b^(k-1), (b^k)-1] .. [b^floor(log(n, b)), n]
-// and is uniformly distributed within each bucket.
-//
-// The resulting probability density is inversely related to bucket size, though
-// values in the final bucket may be more likely than previous values. (In the
-// extreme case where n = b^i the final value will be tied with zero as the most
-// probable result.
-//
-// If `lo` is nonzero then this distribution is shifted to the desired interval,
-// so LogUniform(lo, hi, b) is equivalent to LogUniform(0, hi-lo, b)+lo.
-//
-// See https://en.wikipedia.org/wiki/Log-normal_distribution
-//
-// Example:
-//
-//   absl::BitGen bitgen;
-//   ...
-//   int v = absl::LogUniform(bitgen, 0, 1000);
-//
-template <typename IntType, typename URBG>
-IntType LogUniform(URBG&& urbg,  // NOLINT(runtime/references)
-                   IntType lo, IntType hi, IntType base = 2) {
-  static_assert(random_internal::IsIntegral<IntType>::value,
-                "Template-argument 'IntType' must be an integral type, in "
-                "absl::LogUniform<IntType, URBG>(...)");
-
-  using gen_t = absl::decay_t<URBG>;
-  using distribution_t = typename absl::log_uniform_int_distribution<IntType>;
-
-  return random_internal::DistributionCaller<gen_t>::template Call<
-      distribution_t>(&urbg, lo, hi, base);
-}
-
-// -----------------------------------------------------------------------------
-// absl::Poisson<T>(bitgen, mean = 1)
-// -----------------------------------------------------------------------------
-//
-// `absl::Poisson` produces discrete probabilities for a given number of events
-// occurring within a fixed interval within the closed interval [0, max]. `T`
-// must be an integral type.
-//
-// See https://en.wikipedia.org/wiki/Poisson_distribution
-//
-// Example:
-//
-//   absl::BitGen bitgen;
-//   ...
-//   int requests_per_minute = absl::Poisson<int>(bitgen, 3.2);
-//
-template <typename IntType, typename URBG>
-IntType Poisson(URBG&& urbg,  // NOLINT(runtime/references)
-                double mean = 1.0) {
-  static_assert(random_internal::IsIntegral<IntType>::value,
-                "Template-argument 'IntType' must be an integral type, in "
-                "absl::Poisson<IntType, URBG>(...)");
-
-  using gen_t = absl::decay_t<URBG>;
-  using distribution_t = typename absl::poisson_distribution<IntType>;
-
-  return random_internal::DistributionCaller<gen_t>::template Call<
-      distribution_t>(&urbg, mean);
-}
-
-// -----------------------------------------------------------------------------
-// absl::Zipf<T>(bitgen, hi = max, q = 2, v = 1)
-// -----------------------------------------------------------------------------
-//
-// `absl::Zipf` produces discrete probabilities commonly used for modelling of
-// rare events over the closed interval [0, hi]. The parameters `v` and `q`
-// determine the skew of the distribution. `T`  must be an integral type, but
-// may be inferred from the type of `hi`.
-//
-// See http://mathworld.wolfram.com/ZipfDistribution.html
-//
-// Example:
-//
-//   absl::BitGen bitgen;
-//   ...
-//   int term_rank = absl::Zipf<int>(bitgen);
-//
-template <typename IntType, typename URBG>
-IntType Zipf(URBG&& urbg,  // NOLINT(runtime/references)
-             IntType hi = (std::numeric_limits<IntType>::max)(), double q = 2.0,
-             double v = 1.0) {
-  static_assert(random_internal::IsIntegral<IntType>::value,
-                "Template-argument 'IntType' must be an integral type, in "
-                "absl::Zipf<IntType, URBG>(...)");
-
-  using gen_t = absl::decay_t<URBG>;
-  using distribution_t = typename absl::zipf_distribution<IntType>;
-
-  return random_internal::DistributionCaller<gen_t>::template Call<
-      distribution_t>(&urbg, hi, q, v);
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_DISTRIBUTIONS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/exponential_distribution.h b/projects/llm_framework/include/abseil-cpp/absl/random/exponential_distribution.h
deleted file mode 100644
index b5caf8a1..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/exponential_distribution.h
+++ /dev/null
@@ -1,165 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_EXPONENTIAL_DISTRIBUTION_H_
-#define ABSL_RANDOM_EXPONENTIAL_DISTRIBUTION_H_
-
-#include <cassert>
-#include <cmath>
-#include <istream>
-#include <limits>
-#include <type_traits>
-
-#include "absl/meta/type_traits.h"
-#include "absl/random/internal/fast_uniform_bits.h"
-#include "absl/random/internal/generate_real.h"
-#include "absl/random/internal/iostream_state_saver.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// absl::exponential_distribution:
-// Generates a number conforming to an exponential distribution and is
-// equivalent to the standard [rand.dist.pois.exp] distribution.
-template <typename RealType = double>
-class exponential_distribution {
- public:
-  using result_type = RealType;
-
-  class param_type {
-   public:
-    using distribution_type = exponential_distribution;
-
-    explicit param_type(result_type lambda = 1) : lambda_(lambda) {
-      assert(lambda > 0);
-      neg_inv_lambda_ = -result_type(1) / lambda_;
-    }
-
-    result_type lambda() const { return lambda_; }
-
-    friend bool operator==(const param_type& a, const param_type& b) {
-      return a.lambda_ == b.lambda_;
-    }
-
-    friend bool operator!=(const param_type& a, const param_type& b) {
-      return !(a == b);
-    }
-
-   private:
-    friend class exponential_distribution;
-
-    result_type lambda_;
-    result_type neg_inv_lambda_;
-
-    static_assert(
-        std::is_floating_point<RealType>::value,
-        "Class-template absl::exponential_distribution<> must be parameterized "
-        "using a floating-point type.");
-  };
-
-  exponential_distribution() : exponential_distribution(1) {}
-
-  explicit exponential_distribution(result_type lambda) : param_(lambda) {}
-
-  explicit exponential_distribution(const param_type& p) : param_(p) {}
-
-  void reset() {}
-
-  // Generating functions
-  template <typename URBG>
-  result_type operator()(URBG& g) {  // NOLINT(runtime/references)
-    return (*this)(g, param_);
-  }
-
-  template <typename URBG>
-  result_type operator()(URBG& g,  // NOLINT(runtime/references)
-                         const param_type& p);
-
-  param_type param() const { return param_; }
-  void param(const param_type& p) { param_ = p; }
-
-  result_type(min)() const { return 0; }
-  result_type(max)() const {
-    return std::numeric_limits<result_type>::infinity();
-  }
-
-  result_type lambda() const { return param_.lambda(); }
-
-  friend bool operator==(const exponential_distribution& a,
-                         const exponential_distribution& b) {
-    return a.param_ == b.param_;
-  }
-  friend bool operator!=(const exponential_distribution& a,
-                         const exponential_distribution& b) {
-    return a.param_ != b.param_;
-  }
-
- private:
-  param_type param_;
-  random_internal::FastUniformBits<uint64_t> fast_u64_;
-};
-
-// --------------------------------------------------------------------------
-// Implementation details follow
-// --------------------------------------------------------------------------
-
-template <typename RealType>
-template <typename URBG>
-typename exponential_distribution<RealType>::result_type
-exponential_distribution<RealType>::operator()(
-    URBG& g,  // NOLINT(runtime/references)
-    const param_type& p) {
-  using random_internal::GenerateNegativeTag;
-  using random_internal::GenerateRealFromBits;
-  using real_type =
-      absl::conditional_t<std::is_same<RealType, float>::value, float, double>;
-
-  const result_type u = GenerateRealFromBits<real_type, GenerateNegativeTag,
-                                             false>(fast_u64_(g));  // U(-1, 0)
-
-  // log1p(-x) is mathematically equivalent to log(1 - x) but has more
-  // accuracy for x near zero.
-  return p.neg_inv_lambda_ * std::log1p(u);
-}
-
-template <typename CharT, typename Traits, typename RealType>
-std::basic_ostream<CharT, Traits>& operator<<(
-    std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
-    const exponential_distribution<RealType>& x) {
-  auto saver = random_internal::make_ostream_state_saver(os);
-  os.precision(random_internal::stream_precision_helper<RealType>::kPrecision);
-  os << x.lambda();
-  return os;
-}
-
-template <typename CharT, typename Traits, typename RealType>
-std::basic_istream<CharT, Traits>& operator>>(
-    std::basic_istream<CharT, Traits>& is,    // NOLINT(runtime/references)
-    exponential_distribution<RealType>& x) {  // NOLINT(runtime/references)
-  using result_type = typename exponential_distribution<RealType>::result_type;
-  using param_type = typename exponential_distribution<RealType>::param_type;
-  result_type lambda;
-
-  auto saver = random_internal::make_istream_state_saver(is);
-  lambda = random_internal::read_floating_point<result_type>(is);
-  if (!is.fail()) {
-    x.param(param_type(lambda));
-  }
-  return is;
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_EXPONENTIAL_DISTRIBUTION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/gaussian_distribution.h b/projects/llm_framework/include/abseil-cpp/absl/random/gaussian_distribution.h
deleted file mode 100644
index 4b07a5c0..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/gaussian_distribution.h
+++ /dev/null
@@ -1,275 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_GAUSSIAN_DISTRIBUTION_H_
-#define ABSL_RANDOM_GAUSSIAN_DISTRIBUTION_H_
-
-// absl::gaussian_distribution implements the Ziggurat algorithm
-// for generating random gaussian numbers.
-//
-// Implementation based on "The Ziggurat Method for Generating Random Variables"
-// by George Marsaglia and Wai Wan Tsang: http://www.jstatsoft.org/v05/i08/
-//
-
-#include <cmath>
-#include <cstdint>
-#include <istream>
-#include <limits>
-#include <type_traits>
-
-#include "absl/base/config.h"
-#include "absl/random/internal/fast_uniform_bits.h"
-#include "absl/random/internal/generate_real.h"
-#include "absl/random/internal/iostream_state_saver.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// absl::gaussian_distribution_base implements the underlying ziggurat algorithm
-// using the ziggurat tables generated by the gaussian_distribution_gentables
-// binary.
-//
-// The specific algorithm has some of the improvements suggested by the
-// 2005 paper, "An Improved Ziggurat Method to Generate Normal Random Samples",
-// Jurgen A Doornik.  (https://www.doornik.com/research/ziggurat.pdf)
-class ABSL_DLL gaussian_distribution_base {
- public:
-  template <typename URBG>
-  inline double zignor(URBG& g);  // NOLINT(runtime/references)
-
- private:
-  friend class TableGenerator;
-
-  template <typename URBG>
-  inline double zignor_fallback(URBG& g,  // NOLINT(runtime/references)
-                                bool neg);
-
-  // Constants used for the gaussian distribution.
-  static constexpr double kR = 3.442619855899;  // Start of the tail.
-  static constexpr double kRInv = 0.29047645161474317;  // ~= (1.0 / kR) .
-  static constexpr double kV = 9.91256303526217e-3;
-  static constexpr uint64_t kMask = 0x07f;
-
-  // The ziggurat tables store the pdf(f) and inverse-pdf(x) for equal-area
-  // points on one-half of the normal distribution, where the pdf function,
-  // pdf = e ^ (-1/2 *x^2), assumes that the mean = 0 & stddev = 1.
-  //
-  // These tables are just over 2kb in size; larger tables might improve the
-  // distributions, but also lead to more cache pollution.
-  //
-  // x = {3.71308, 3.44261, 3.22308, ..., 0}
-  // f = {0.00101, 0.00266, 0.00554, ..., 1}
-  struct Tables {
-    double x[kMask + 2];
-    double f[kMask + 2];
-  };
-  static const Tables zg_;
-  random_internal::FastUniformBits<uint64_t> fast_u64_;
-};
-
-}  // namespace random_internal
-
-// absl::gaussian_distribution:
-// Generates a number conforming to a Gaussian distribution.
-template <typename RealType = double>
-class gaussian_distribution : random_internal::gaussian_distribution_base {
- public:
-  using result_type = RealType;
-
-  class param_type {
-   public:
-    using distribution_type = gaussian_distribution;
-
-    explicit param_type(result_type mean = 0, result_type stddev = 1)
-        : mean_(mean), stddev_(stddev) {}
-
-    // Returns the mean distribution parameter.  The mean specifies the location
-    // of the peak.  The default value is 0.0.
-    result_type mean() const { return mean_; }
-
-    // Returns the deviation distribution parameter.  The default value is 1.0.
-    result_type stddev() const { return stddev_; }
-
-    friend bool operator==(const param_type& a, const param_type& b) {
-      return a.mean_ == b.mean_ && a.stddev_ == b.stddev_;
-    }
-
-    friend bool operator!=(const param_type& a, const param_type& b) {
-      return !(a == b);
-    }
-
-   private:
-    result_type mean_;
-    result_type stddev_;
-
-    static_assert(
-        std::is_floating_point<RealType>::value,
-        "Class-template absl::gaussian_distribution<> must be parameterized "
-        "using a floating-point type.");
-  };
-
-  gaussian_distribution() : gaussian_distribution(0) {}
-
-  explicit gaussian_distribution(result_type mean, result_type stddev = 1)
-      : param_(mean, stddev) {}
-
-  explicit gaussian_distribution(const param_type& p) : param_(p) {}
-
-  void reset() {}
-
-  // Generating functions
-  template <typename URBG>
-  result_type operator()(URBG& g) {  // NOLINT(runtime/references)
-    return (*this)(g, param_);
-  }
-
-  template <typename URBG>
-  result_type operator()(URBG& g,  // NOLINT(runtime/references)
-                         const param_type& p);
-
-  param_type param() const { return param_; }
-  void param(const param_type& p) { param_ = p; }
-
-  result_type(min)() const {
-    return -std::numeric_limits<result_type>::infinity();
-  }
-  result_type(max)() const {
-    return std::numeric_limits<result_type>::infinity();
-  }
-
-  result_type mean() const { return param_.mean(); }
-  result_type stddev() const { return param_.stddev(); }
-
-  friend bool operator==(const gaussian_distribution& a,
-                         const gaussian_distribution& b) {
-    return a.param_ == b.param_;
-  }
-  friend bool operator!=(const gaussian_distribution& a,
-                         const gaussian_distribution& b) {
-    return a.param_ != b.param_;
-  }
-
- private:
-  param_type param_;
-};
-
-// --------------------------------------------------------------------------
-// Implementation details only below
-// --------------------------------------------------------------------------
-
-template <typename RealType>
-template <typename URBG>
-typename gaussian_distribution<RealType>::result_type
-gaussian_distribution<RealType>::operator()(
-    URBG& g,  // NOLINT(runtime/references)
-    const param_type& p) {
-  return p.mean() + p.stddev() * static_cast<result_type>(zignor(g));
-}
-
-template <typename CharT, typename Traits, typename RealType>
-std::basic_ostream<CharT, Traits>& operator<<(
-    std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
-    const gaussian_distribution<RealType>& x) {
-  auto saver = random_internal::make_ostream_state_saver(os);
-  os.precision(random_internal::stream_precision_helper<RealType>::kPrecision);
-  os << x.mean() << os.fill() << x.stddev();
-  return os;
-}
-
-template <typename CharT, typename Traits, typename RealType>
-std::basic_istream<CharT, Traits>& operator>>(
-    std::basic_istream<CharT, Traits>& is,  // NOLINT(runtime/references)
-    gaussian_distribution<RealType>& x) {   // NOLINT(runtime/references)
-  using result_type = typename gaussian_distribution<RealType>::result_type;
-  using param_type = typename gaussian_distribution<RealType>::param_type;
-
-  auto saver = random_internal::make_istream_state_saver(is);
-  auto mean = random_internal::read_floating_point<result_type>(is);
-  if (is.fail()) return is;
-  auto stddev = random_internal::read_floating_point<result_type>(is);
-  if (!is.fail()) {
-    x.param(param_type(mean, stddev));
-  }
-  return is;
-}
-
-namespace random_internal {
-
-template <typename URBG>
-inline double gaussian_distribution_base::zignor_fallback(URBG& g, bool neg) {
-  using random_internal::GeneratePositiveTag;
-  using random_internal::GenerateRealFromBits;
-
-  // This fallback path happens approximately 0.05% of the time.
-  double x, y;
-  do {
-    // kRInv = 1/r, U(0, 1)
-    x = kRInv *
-        std::log(GenerateRealFromBits<double, GeneratePositiveTag, false>(
-            fast_u64_(g)));
-    y = -std::log(
-        GenerateRealFromBits<double, GeneratePositiveTag, false>(fast_u64_(g)));
-  } while ((y + y) < (x * x));
-  return neg ? (x - kR) : (kR - x);
-}
-
-template <typename URBG>
-inline double gaussian_distribution_base::zignor(
-    URBG& g) {  // NOLINT(runtime/references)
-  using random_internal::GeneratePositiveTag;
-  using random_internal::GenerateRealFromBits;
-  using random_internal::GenerateSignedTag;
-
-  while (true) {
-    // We use a single uint64_t to generate both a double and a strip.
-    // These bits are unused when the generated double is > 1/2^5.
-    // This may introduce some bias from the duplicated low bits of small
-    // values (those smaller than 1/2^5, which all end up on the left tail).
-    uint64_t bits = fast_u64_(g);
-    int i = static_cast<int>(bits & kMask);  // pick a random strip
-    double j = GenerateRealFromBits<double, GenerateSignedTag, false>(
-        bits);  // U(-1, 1)
-    const double x = j * zg_.x[i];
-
-    // Retangular box. Handles >97% of all cases.
-    // For any given box, this handles between 75% and 99% of values.
-    // Equivalent to U(01) < (x[i+1] / x[i]), and when i == 0, ~93.5%
-    if (std::abs(x) < zg_.x[i + 1]) {
-      return x;
-    }
-
-    // i == 0: Base box. Sample using a ratio of uniforms.
-    if (i == 0) {
-      // This path happens about 0.05% of the time.
-      return zignor_fallback(g, j < 0);
-    }
-
-    // i > 0: Wedge samples using precomputed values.
-    double v = GenerateRealFromBits<double, GeneratePositiveTag, false>(
-        fast_u64_(g));  // U(0, 1)
-    if ((zg_.f[i + 1] + v * (zg_.f[i] - zg_.f[i + 1])) <
-        std::exp(-0.5 * x * x)) {
-      return x;
-    }
-
-    // The wedge was missed; reject the value and try again.
-  }
-}
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_GAUSSIAN_DISTRIBUTION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/chi_square.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/chi_square.h
deleted file mode 100644
index 07f4fbe5..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/chi_square.h
+++ /dev/null
@@ -1,89 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_CHI_SQUARE_H_
-#define ABSL_RANDOM_INTERNAL_CHI_SQUARE_H_
-
-// The chi-square statistic.
-//
-// Useful for evaluating if `D` independent random variables are behaving as
-// expected, or if two distributions are similar.  (`D` is the degrees of
-// freedom).
-//
-// Each bucket should have an expected count of 10 or more for the chi square to
-// be meaningful.
-
-#include <cassert>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-constexpr const char kChiSquared[] = "chi-squared";
-
-// Returns the measured chi square value, using a single expected value.  This
-// assumes that the values in [begin, end) are uniformly distributed.
-template <typename Iterator>
-double ChiSquareWithExpected(Iterator begin, Iterator end, double expected) {
-  // Compute the sum and the number of buckets.
-  assert(expected >= 10);  // require at least 10 samples per bucket.
-  double chi_square = 0;
-  for (auto it = begin; it != end; it++) {
-    double d = static_cast<double>(*it) - expected;
-    chi_square += d * d;
-  }
-  chi_square = chi_square / expected;
-  return chi_square;
-}
-
-// Returns the measured chi square value, taking the actual value of each bucket
-// from the first set of iterators, and the expected value of each bucket from
-// the second set of iterators.
-template <typename Iterator, typename Expected>
-double ChiSquare(Iterator it, Iterator end, Expected eit, Expected eend) {
-  double chi_square = 0;
-  for (; it != end && eit != eend; ++it, ++eit) {
-    if (*it > 0) {
-      assert(*eit > 0);
-    }
-    double e = static_cast<double>(*eit);
-    double d = static_cast<double>(*it - *eit);
-    if (d != 0) {
-      assert(e > 0);
-      chi_square += (d * d) / e;
-    }
-  }
-  assert(it == end && eit == eend);
-  return chi_square;
-}
-
-// ======================================================================
-// The following methods can be used for an arbitrary significance level.
-//
-
-// Calculates critical chi-square values to produce the given p-value using a
-// bisection search for a value within epsilon, relying on the monotonicity of
-// ChiSquarePValue().
-double ChiSquareValue(int dof, double p);
-
-// Calculates the p-value (probability) of a given chi-square value.
-double ChiSquarePValue(double chi_square, int dof);
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_CHI_SQUARE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/distribution_caller.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/distribution_caller.h
deleted file mode 100644
index 0f162a4e..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/distribution_caller.h
+++ /dev/null
@@ -1,95 +0,0 @@
-//
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_RANDOM_INTERNAL_DISTRIBUTION_CALLER_H_
-#define ABSL_RANDOM_INTERNAL_DISTRIBUTION_CALLER_H_
-
-#include <utility>
-#include <type_traits>
-
-#include "absl/base/config.h"
-#include "absl/base/internal/fast_type_id.h"
-#include "absl/utility/utility.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// DistributionCaller provides an opportunity to overload the general
-// mechanism for calling a distribution, allowing for mock-RNG classes
-// to intercept such calls.
-template <typename URBG>
-struct DistributionCaller {
-  static_assert(!std::is_pointer<URBG>::value,
-                "You must pass a reference, not a pointer.");
-  // SFINAE to detect whether the URBG type includes a member matching
-  // bool InvokeMock(base_internal::FastTypeIdType, void*, void*).
-  //
-  // These live inside BitGenRef so that they have friend access
-  // to MockingBitGen. (see similar methods in DistributionCaller).
-  template <template <class...> class Trait, class AlwaysVoid, class... Args>
-  struct detector : std::false_type {};
-  template <template <class...> class Trait, class... Args>
-  struct detector<Trait, absl::void_t<Trait<Args...>>, Args...>
-      : std::true_type {};
-
-  template <class T>
-  using invoke_mock_t = decltype(std::declval<T*>()->InvokeMock(
-      std::declval<::absl::base_internal::FastTypeIdType>(),
-      std::declval<void*>(), std::declval<void*>()));
-
-  using HasInvokeMock = typename detector<invoke_mock_t, void, URBG>::type;
-
-  // Default implementation of distribution caller.
-  template <typename DistrT, typename... Args>
-  static typename DistrT::result_type Impl(std::false_type, URBG* urbg,
-                                           Args&&... args) {
-    DistrT dist(std::forward<Args>(args)...);
-    return dist(*urbg);
-  }
-
-  // Mock implementation of distribution caller.
-  // The underlying KeyT must match the KeyT constructed by MockOverloadSet.
-  template <typename DistrT, typename... Args>
-  static typename DistrT::result_type Impl(std::true_type, URBG* urbg,
-                                           Args&&... args) {
-    using ResultT = typename DistrT::result_type;
-    using ArgTupleT = std::tuple<absl::decay_t<Args>...>;
-    using KeyT = ResultT(DistrT, ArgTupleT);
-
-    ArgTupleT arg_tuple(std::forward<Args>(args)...);
-    ResultT result;
-    if (!urbg->InvokeMock(::absl::base_internal::FastTypeId<KeyT>(), &arg_tuple,
-                          &result)) {
-      auto dist = absl::make_from_tuple<DistrT>(arg_tuple);
-      result = dist(*urbg);
-    }
-    return result;
-  }
-
-  // Default implementation of distribution caller.
-  template <typename DistrT, typename... Args>
-  static typename DistrT::result_type Call(URBG* urbg, Args&&... args) {
-    return Impl<DistrT, Args...>(HasInvokeMock{}, urbg,
-                                 std::forward<Args>(args)...);
-  }
-};
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_DISTRIBUTION_CALLER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/distribution_test_util.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/distribution_test_util.h
deleted file mode 100644
index 6d94cf6c..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/distribution_test_util.h
+++ /dev/null
@@ -1,113 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_DISTRIBUTION_TEST_UTIL_H_
-#define ABSL_RANDOM_INTERNAL_DISTRIBUTION_TEST_UTIL_H_
-
-#include <cstddef>
-#include <iostream>
-#include <vector>
-
-#include "absl/strings/string_view.h"
-#include "absl/types/span.h"
-
-// NOTE: The functions in this file are test only, and are should not be used in
-// non-test code.
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// http://webspace.ship.edu/pgmarr/Geo441/Lectures/Lec%205%20-%20Normality%20Testing.pdf
-
-// Compute the 1st to 4th standard moments:
-// mean, variance, skewness, and kurtosis.
-// http://www.itl.nist.gov/div898/handbook/eda/section3/eda35b.htm
-struct DistributionMoments {
-  size_t n = 0;
-  double mean = 0.0;
-  double variance = 0.0;
-  double skewness = 0.0;
-  double kurtosis = 0.0;
-};
-DistributionMoments ComputeDistributionMoments(
-    absl::Span<const double> data_points);
-
-std::ostream& operator<<(std::ostream& os, const DistributionMoments& moments);
-
-// Computes the Z-score for a set of data with the given distribution moments
-// compared against `expected_mean`.
-double ZScore(double expected_mean, const DistributionMoments& moments);
-
-// Returns the probability of success required for a single trial to ensure that
-// after `num_trials` trials, the probability of at least one failure is no more
-// than `p_fail`.
-double RequiredSuccessProbability(double p_fail, int num_trials);
-
-// Computes the maximum distance from the mean tolerable, for Z-Tests that are
-// expected to pass with `acceptance_probability`. Will terminate if the
-// resulting tolerance is zero (due to passing in 0.0 for
-// `acceptance_probability` or rounding errors).
-//
-// For example,
-// MaxErrorTolerance(0.001) = 0.0
-// MaxErrorTolerance(0.5) = ~0.47
-// MaxErrorTolerance(1.0) = inf
-double MaxErrorTolerance(double acceptance_probability);
-
-// Approximation to inverse of the Error Function in double precision.
-// (http://people.maths.ox.ac.uk/gilesm/files/gems_erfinv.pdf)
-double erfinv(double x);
-
-// Beta(p, q) = Gamma(p) * Gamma(q) / Gamma(p+q)
-double beta(double p, double q);
-
-// The inverse of the normal survival function.
-double InverseNormalSurvival(double x);
-
-// Returns whether actual is "near" expected, based on the bound.
-bool Near(absl::string_view msg, double actual, double expected, double bound);
-
-// Implements the incomplete regularized beta function, AS63, BETAIN.
-//    https://www.jstor.org/stable/2346797
-//
-// BetaIncomplete(x, p, q), where
-//   `x` is the value of the upper limit
-//   `p` is beta parameter p, `q` is beta parameter q.
-//
-// NOTE: This is a test-only function which is only accurate to within, at most,
-// 1e-13 of the actual value.
-//
-double BetaIncomplete(double x, double p, double q);
-
-// Implements the inverse of the incomplete regularized beta function, AS109,
-// XINBTA.
-//   https://www.jstor.org/stable/2346798
-//   https://www.jstor.org/stable/2346887
-//
-// BetaIncompleteInv(p, q, beta, alhpa)
-//   `p` is beta parameter p, `q` is beta parameter q.
-//   `alpha` is the value of the lower tail area.
-//
-// NOTE: This is a test-only function and, when successful, is only accurate to
-// within ~1e-6 of the actual value; there are some cases where it diverges from
-// the actual value by much more than that.  The function uses Newton's method,
-// and thus the runtime is highly variable.
-double BetaIncompleteInv(double p, double q, double alpha);
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_DISTRIBUTION_TEST_UTIL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/explicit_seed_seq.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/explicit_seed_seq.h
deleted file mode 100644
index 25f79153..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/explicit_seed_seq.h
+++ /dev/null
@@ -1,92 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_EXPLICIT_SEED_SEQ_H_
-#define ABSL_RANDOM_INTERNAL_EXPLICIT_SEED_SEQ_H_
-
-#include <algorithm>
-#include <cstddef>
-#include <cstdint>
-#include <initializer_list>
-#include <iterator>
-#include <vector>
-
-#include "absl/base/config.h"
-#include "absl/base/internal/endian.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// This class conforms to the C++ Standard "Seed Sequence" concept
-// [rand.req.seedseq].
-//
-// An "ExplicitSeedSeq" is meant to provide a conformant interface for
-// forwarding pre-computed seed material to the constructor of a class
-// conforming to the "Uniform Random Bit Generator" concept. This class makes no
-// attempt to mutate the state provided by its constructor, and returns it
-// directly via ExplicitSeedSeq::generate().
-//
-// If this class is asked to generate more seed material than was provided to
-// the constructor, then the remaining bytes will be filled with deterministic,
-// nonrandom data.
-class ExplicitSeedSeq {
- public:
-  using result_type = uint32_t;
-
-  ExplicitSeedSeq() : state_() {}
-
-  // Copy and move both allowed.
-  ExplicitSeedSeq(const ExplicitSeedSeq& other) = default;
-  ExplicitSeedSeq& operator=(const ExplicitSeedSeq& other) = default;
-  ExplicitSeedSeq(ExplicitSeedSeq&& other) = default;
-  ExplicitSeedSeq& operator=(ExplicitSeedSeq&& other) = default;
-
-  template <typename Iterator>
-  ExplicitSeedSeq(Iterator begin, Iterator end) {
-    for (auto it = begin; it != end; it++) {
-      state_.push_back(*it & 0xffffffff);
-    }
-  }
-
-  template <typename T>
-  ExplicitSeedSeq(std::initializer_list<T> il)
-      : ExplicitSeedSeq(il.begin(), il.end()) {}
-
-  size_t size() const { return state_.size(); }
-
-  template <typename OutIterator>
-  void param(OutIterator out) const {
-    std::copy(std::begin(state_), std::end(state_), out);
-  }
-
-  template <typename OutIterator>
-  void generate(OutIterator begin, OutIterator end) {
-    for (size_t index = 0; begin != end; begin++) {
-      *begin = state_.empty() ? 0 : state_[index++];
-      if (index >= state_.size()) {
-        index = 0;
-      }
-    }
-  }
-
- protected:
-  std::vector<uint32_t> state_;
-};
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_EXPLICIT_SEED_SEQ_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/fast_uniform_bits.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/fast_uniform_bits.h
deleted file mode 100644
index 83ee5c0f..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/fast_uniform_bits.h
+++ /dev/null
@@ -1,271 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_FAST_UNIFORM_BITS_H_
-#define ABSL_RANDOM_INTERNAL_FAST_UNIFORM_BITS_H_
-
-#include <cstddef>
-#include <cstdint>
-#include <limits>
-#include <type_traits>
-
-#include "absl/base/config.h"
-#include "absl/meta/type_traits.h"
-#include "absl/random/internal/traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-// Returns true if the input value is zero or a power of two. Useful for
-// determining if the range of output values in a URBG
-template <typename UIntType>
-constexpr bool IsPowerOfTwoOrZero(UIntType n) {
-  return (n == 0) || ((n & (n - 1)) == 0);
-}
-
-// Computes the length of the range of values producible by the URBG, or returns
-// zero if that would encompass the entire range of representable values in
-// URBG::result_type.
-template <typename URBG>
-constexpr typename URBG::result_type RangeSize() {
-  using result_type = typename URBG::result_type;
-  static_assert((URBG::max)() != (URBG::min)(), "URBG range cannot be 0.");
-  return ((URBG::max)() == (std::numeric_limits<result_type>::max)() &&
-          (URBG::min)() == std::numeric_limits<result_type>::lowest())
-             ? result_type{0}
-             : ((URBG::max)() - (URBG::min)() + result_type{1});
-}
-
-// Computes the floor of the log. (i.e., std::floor(std::log2(N));
-template <typename UIntType>
-constexpr UIntType IntegerLog2(UIntType n) {
-  return (n <= 1) ? 0 : 1 + IntegerLog2(n >> 1);
-}
-
-// Returns the number of bits of randomness returned through
-// `PowerOfTwoVariate(urbg)`.
-template <typename URBG>
-constexpr size_t NumBits() {
-  return static_cast<size_t>(
-      RangeSize<URBG>() == 0
-          ? std::numeric_limits<typename URBG::result_type>::digits
-          : IntegerLog2(RangeSize<URBG>()));
-}
-
-// Given a shift value `n`, constructs a mask with exactly the low `n` bits set.
-// If `n == 0`, all bits are set.
-template <typename UIntType>
-constexpr UIntType MaskFromShift(size_t n) {
-  return ((n % std::numeric_limits<UIntType>::digits) == 0)
-             ? ~UIntType{0}
-             : (UIntType{1} << n) - UIntType{1};
-}
-
-// Tags used to dispatch FastUniformBits::generate to the simple or more complex
-// entropy extraction algorithm.
-struct SimplifiedLoopTag {};
-struct RejectionLoopTag {};
-
-// FastUniformBits implements a fast path to acquire uniform independent bits
-// from a type which conforms to the [rand.req.urbg] concept.
-// Parameterized by:
-//  `UIntType`: the result (output) type
-//
-// The std::independent_bits_engine [rand.adapt.ibits] adaptor can be
-// instantiated from an existing generator through a copy or a move. It does
-// not, however, facilitate the production of pseudorandom bits from an un-owned
-// generator that will outlive the std::independent_bits_engine instance.
-template <typename UIntType = uint64_t>
-class FastUniformBits {
- public:
-  using result_type = UIntType;
-
-  static constexpr result_type(min)() { return 0; }
-  static constexpr result_type(max)() {
-    return (std::numeric_limits<result_type>::max)();
-  }
-
-  template <typename URBG>
-  result_type operator()(URBG& g);  // NOLINT(runtime/references)
-
- private:
-  static_assert(IsUnsigned<UIntType>::value,
-                "Class-template FastUniformBits<> must be parameterized using "
-                "an unsigned type.");
-
-  // Generate() generates a random value, dispatched on whether
-  // the underlying URBG must use rejection sampling to generate a value,
-  // or whether a simplified loop will suffice.
-  template <typename URBG>
-  result_type Generate(URBG& g,  // NOLINT(runtime/references)
-                       SimplifiedLoopTag);
-
-  template <typename URBG>
-  result_type Generate(URBG& g,  // NOLINT(runtime/references)
-                       RejectionLoopTag);
-};
-
-template <typename UIntType>
-template <typename URBG>
-typename FastUniformBits<UIntType>::result_type
-FastUniformBits<UIntType>::operator()(URBG& g) {  // NOLINT(runtime/references)
-  // kRangeMask is the mask used when sampling variates from the URBG when the
-  // width of the URBG range is not a power of 2.
-  // Y = (2 ^ kRange) - 1
-  static_assert((URBG::max)() > (URBG::min)(),
-                "URBG::max and URBG::min may not be equal.");
-
-  using tag = absl::conditional_t<IsPowerOfTwoOrZero(RangeSize<URBG>()),
-                                  SimplifiedLoopTag, RejectionLoopTag>;
-  return Generate(g, tag{});
-}
-
-template <typename UIntType>
-template <typename URBG>
-typename FastUniformBits<UIntType>::result_type
-FastUniformBits<UIntType>::Generate(URBG& g,  // NOLINT(runtime/references)
-                                    SimplifiedLoopTag) {
-  // The simplified version of FastUniformBits works only on URBGs that have
-  // a range that is a power of 2. In this case we simply loop and shift without
-  // attempting to balance the bits across calls.
-  static_assert(IsPowerOfTwoOrZero(RangeSize<URBG>()),
-                "incorrect Generate tag for URBG instance");
-
-  static constexpr size_t kResultBits =
-      std::numeric_limits<result_type>::digits;
-  static constexpr size_t kUrbgBits = NumBits<URBG>();
-  static constexpr size_t kIters =
-      (kResultBits / kUrbgBits) + (kResultBits % kUrbgBits != 0);
-  static constexpr size_t kShift = (kIters == 1) ? 0 : kUrbgBits;
-  static constexpr auto kMin = (URBG::min)();
-
-  result_type r = static_cast<result_type>(g() - kMin);
-  for (size_t n = 1; n < kIters; ++n) {
-    r = static_cast<result_type>(r << kShift) +
-        static_cast<result_type>(g() - kMin);
-  }
-  return r;
-}
-
-template <typename UIntType>
-template <typename URBG>
-typename FastUniformBits<UIntType>::result_type
-FastUniformBits<UIntType>::Generate(URBG& g,  // NOLINT(runtime/references)
-                                    RejectionLoopTag) {
-  static_assert(!IsPowerOfTwoOrZero(RangeSize<URBG>()),
-                "incorrect Generate tag for URBG instance");
-  using urbg_result_type = typename URBG::result_type;
-
-  // See [rand.adapt.ibits] for more details on the constants calculated below.
-  //
-  // It is preferable to use roughly the same number of bits from each generator
-  // call, however this is only possible when the number of bits provided by the
-  // URBG is a divisor of the number of bits in `result_type`. In all other
-  // cases, the number of bits used cannot always be the same, but it can be
-  // guaranteed to be off by at most 1. Thus we run two loops, one with a
-  // smaller bit-width size (`kSmallWidth`) and one with a larger width size
-  // (satisfying `kLargeWidth == kSmallWidth + 1`). The loops are run
-  // `kSmallIters` and `kLargeIters` times respectively such
-  // that
-  //
-  //    `kResultBits == kSmallIters * kSmallBits
-  //                    + kLargeIters * kLargeBits`
-  //
-  // where `kResultBits` is the total number of bits in `result_type`.
-  //
-  static constexpr size_t kResultBits =
-      std::numeric_limits<result_type>::digits;                      // w
-  static constexpr urbg_result_type kUrbgRange = RangeSize<URBG>();  // R
-  static constexpr size_t kUrbgBits = NumBits<URBG>();               // m
-
-  // compute the initial estimate of the bits used.
-  // [rand.adapt.ibits] 2 (c)
-  static constexpr size_t kA =  // ceil(w/m)
-      (kResultBits / kUrbgBits) + ((kResultBits % kUrbgBits) != 0);  // n'
-
-  static constexpr size_t kABits = kResultBits / kA;  // w0'
-  static constexpr urbg_result_type kARejection =
-      ((kUrbgRange >> kABits) << kABits);  // y0'
-
-  // refine the selection to reduce the rejection frequency.
-  static constexpr size_t kTotalIters =
-      ((kUrbgRange - kARejection) <= (kARejection / kA)) ? kA : (kA + 1);  // n
-
-  // [rand.adapt.ibits] 2 (b)
-  static constexpr size_t kSmallIters =
-      kTotalIters - (kResultBits % kTotalIters);                   // n0
-  static constexpr size_t kSmallBits = kResultBits / kTotalIters;  // w0
-  static constexpr urbg_result_type kSmallRejection =
-      ((kUrbgRange >> kSmallBits) << kSmallBits);  // y0
-
-  static constexpr size_t kLargeBits = kSmallBits + 1;  // w0+1
-  static constexpr urbg_result_type kLargeRejection =
-      ((kUrbgRange >> kLargeBits) << kLargeBits);  // y1
-
-  //
-  // Because `kLargeBits == kSmallBits + 1`, it follows that
-  //
-  //     `kResultBits == kSmallIters * kSmallBits + kLargeIters`
-  //
-  // and therefore
-  //
-  //     `kLargeIters == kTotalWidth % kSmallWidth`
-  //
-  // Intuitively, each iteration with the large width accounts for one unit
-  // of the remainder when `kTotalWidth` is divided by `kSmallWidth`. As
-  // mentioned above, if the URBG width is a divisor of `kTotalWidth`, then
-  // there would be no need for any large iterations (i.e., one loop would
-  // suffice), and indeed, in this case, `kLargeIters` would be zero.
-  static_assert(kResultBits == kSmallIters * kSmallBits +
-                                   (kTotalIters - kSmallIters) * kLargeBits,
-                "Error in looping constant calculations.");
-
-  // The small shift is essentially small bits, but due to the potential
-  // of generating a smaller result_type from a larger urbg type, the actual
-  // shift might be 0.
-  static constexpr size_t kSmallShift = kSmallBits % kResultBits;
-  static constexpr auto kSmallMask =
-      MaskFromShift<urbg_result_type>(kSmallShift);
-  static constexpr size_t kLargeShift = kLargeBits % kResultBits;
-  static constexpr auto kLargeMask =
-      MaskFromShift<urbg_result_type>(kLargeShift);
-
-  static constexpr auto kMin = (URBG::min)();
-
-  result_type s = 0;
-  for (size_t n = 0; n < kSmallIters; ++n) {
-    urbg_result_type v;
-    do {
-      v = g() - kMin;
-    } while (v >= kSmallRejection);
-
-    s = (s << kSmallShift) + static_cast<result_type>(v & kSmallMask);
-  }
-
-  for (size_t n = kSmallIters; n < kTotalIters; ++n) {
-    urbg_result_type v;
-    do {
-      v = g() - kMin;
-    } while (v >= kLargeRejection);
-
-    s = (s << kLargeShift) + static_cast<result_type>(v & kLargeMask);
-  }
-  return s;
-}
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_FAST_UNIFORM_BITS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/fastmath.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/fastmath.h
deleted file mode 100644
index 963b7690..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/fastmath.h
+++ /dev/null
@@ -1,57 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_FASTMATH_H_
-#define ABSL_RANDOM_INTERNAL_FASTMATH_H_
-
-// This file contains fast math functions (bitwise ops as well as some others)
-// which are implementation details of various absl random number distributions.
-
-#include <cassert>
-#include <cmath>
-#include <cstdint>
-
-#include "absl/numeric/bits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// Compute log2(n) using integer operations.
-// While std::log2 is more accurate than std::log(n) / std::log(2), for
-// very large numbers--those close to std::numeric_limits<uint64_t>::max() - 2,
-// for instance--std::log2 rounds up rather than down, which introduces
-// definite skew in the results.
-inline int IntLog2Floor(uint64_t n) {
-  return (n <= 1) ? 0 : (63 - countl_zero(n));
-}
-inline int IntLog2Ceil(uint64_t n) {
-  return (n <= 1) ? 0 : (64 - countl_zero(n - 1));
-}
-
-inline double StirlingLogFactorial(double n) {
-  assert(n >= 1);
-  // Using Stirling's approximation.
-  constexpr double kLog2PI = 1.83787706640934548356;
-  const double logn = std::log(n);
-  const double ninv = 1.0 / static_cast<double>(n);
-  return n * logn - n + 0.5 * (kLog2PI + logn) + (1.0 / 12.0) * ninv -
-         (1.0 / 360.0) * ninv * ninv * ninv;
-}
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_FASTMATH_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/generate_real.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/generate_real.h
deleted file mode 100644
index 9a6f4005..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/generate_real.h
+++ /dev/null
@@ -1,144 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_GENERATE_REAL_H_
-#define ABSL_RANDOM_INTERNAL_GENERATE_REAL_H_
-
-// This file contains some implementation details which are used by one or more
-// of the absl random number distributions.
-
-#include <cstdint>
-#include <cstring>
-#include <limits>
-#include <type_traits>
-
-#include "absl/meta/type_traits.h"
-#include "absl/numeric/bits.h"
-#include "absl/random/internal/fastmath.h"
-#include "absl/random/internal/traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// Tristate tag types controlling the output of GenerateRealFromBits.
-struct GeneratePositiveTag {};
-struct GenerateNegativeTag {};
-struct GenerateSignedTag {};
-
-// GenerateRealFromBits generates a single real value from a single 64-bit
-// `bits` with template fields controlling the output.
-//
-// The `SignedTag` parameter controls whether positive, negative,
-// or either signed/unsigned may be returned.
-//   When SignedTag == GeneratePositiveTag, range is U(0, 1)
-//   When SignedTag == GenerateNegativeTag, range is U(-1, 0)
-//   When SignedTag == GenerateSignedTag, range is U(-1, 1)
-//
-// When the `IncludeZero` parameter is true, the function may return 0 for some
-// inputs, otherwise it never returns 0.
-//
-// When a value in U(0,1) is required, use:
-//   GenerateRealFromBits<double, PositiveValueT, true>;
-//
-// When a value in U(-1,1) is required, use:
-//   GenerateRealFromBits<double, SignedValueT, false>;
-//
-//   This generates more distinct values than the mathematical equivalent
-//   `U(0, 1) * 2.0 - 1.0`.
-//
-// Scaling the result by powers of 2 (and avoiding a multiply) is also possible:
-//   GenerateRealFromBits<double>(..., -1);  => U(0, 0.5)
-//   GenerateRealFromBits<double>(..., 1);   => U(0, 2)
-//
-template <typename RealType,  // Real type, either float or double.
-          typename SignedTag = GeneratePositiveTag,  // Whether a positive,
-                                                     // negative, or signed
-                                                     // value is generated.
-          bool IncludeZero = true>
-inline RealType GenerateRealFromBits(uint64_t bits, int exp_bias = 0) {
-  using real_type = RealType;
-  using uint_type = absl::conditional_t<std::is_same<real_type, float>::value,
-                                        uint32_t, uint64_t>;
-
-  static_assert(
-      (std::is_same<double, real_type>::value ||
-       std::is_same<float, real_type>::value),
-      "GenerateRealFromBits must be parameterized by either float or double.");
-
-  static_assert(sizeof(uint_type) == sizeof(real_type),
-                "Mismatched unsigned and real types.");
-
-  static_assert((std::numeric_limits<real_type>::is_iec559 &&
-                 std::numeric_limits<real_type>::radix == 2),
-                "RealType representation is not IEEE 754 binary.");
-
-  static_assert((std::is_same<SignedTag, GeneratePositiveTag>::value ||
-                 std::is_same<SignedTag, GenerateNegativeTag>::value ||
-                 std::is_same<SignedTag, GenerateSignedTag>::value),
-                "");
-
-  static constexpr int kExp = std::numeric_limits<real_type>::digits - 1;
-  static constexpr uint_type kMask = (static_cast<uint_type>(1) << kExp) - 1u;
-  static constexpr int kUintBits = sizeof(uint_type) * 8;
-
-  int exp = exp_bias + int{std::numeric_limits<real_type>::max_exponent - 2};
-
-  // Determine the sign bit.
-  // Depending on the SignedTag, this may use the left-most bit
-  // or it may be a constant value.
-  uint_type sign = std::is_same<SignedTag, GenerateNegativeTag>::value
-                       ? (static_cast<uint_type>(1) << (kUintBits - 1))
-                       : 0;
-  if (std::is_same<SignedTag, GenerateSignedTag>::value) {
-    if (std::is_same<uint_type, uint64_t>::value) {
-      sign = bits & uint64_t{0x8000000000000000};
-    }
-    if (std::is_same<uint_type, uint32_t>::value) {
-      const uint64_t tmp = bits & uint64_t{0x8000000000000000};
-      sign = static_cast<uint32_t>(tmp >> 32);
-    }
-    // adjust the bits and the exponent to account for removing
-    // the leading bit.
-    bits = bits & uint64_t{0x7FFFFFFFFFFFFFFF};
-    exp++;
-  }
-  if (IncludeZero) {
-    if (bits == 0u) return 0;
-  }
-
-  // Number of leading zeros is mapped to the exponent: 2^-clz
-  // bits is 0..01xxxxxx. After shifting, we're left with 1xxx...0..0
-  int clz = countl_zero(bits);
-  bits <<= (IncludeZero ? clz : (clz & 63));  // remove 0-bits.
-  exp -= clz;                                 // set the exponent.
-  bits >>= (63 - kExp);
-
-  // Construct the 32-bit or 64-bit IEEE 754 floating-point value from
-  // the individual fields: sign, exp, mantissa(bits).
-  uint_type val = sign | (static_cast<uint_type>(exp) << kExp) |
-                  (static_cast<uint_type>(bits) & kMask);
-
-  // bit_cast to the output-type
-  real_type result;
-  memcpy(static_cast<void*>(&result), static_cast<const void*>(&val),
-         sizeof(result));
-  return result;
-}
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_GENERATE_REAL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/iostream_state_saver.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/iostream_state_saver.h
deleted file mode 100644
index e6e242ee..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/iostream_state_saver.h
+++ /dev/null
@@ -1,245 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_IOSTREAM_STATE_SAVER_H_
-#define ABSL_RANDOM_INTERNAL_IOSTREAM_STATE_SAVER_H_
-
-#include <cmath>
-#include <iostream>
-#include <limits>
-#include <type_traits>
-
-#include "absl/meta/type_traits.h"
-#include "absl/numeric/int128.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// The null_state_saver does nothing.
-template <typename T>
-class null_state_saver {
- public:
-  using stream_type = T;
-  using flags_type = std::ios_base::fmtflags;
-
-  null_state_saver(T&, flags_type) {}
-  ~null_state_saver() {}
-};
-
-// ostream_state_saver is a RAII object to save and restore the common
-// basic_ostream flags used when implementing `operator <<()` on any of
-// the absl random distributions.
-template <typename OStream>
-class ostream_state_saver {
- public:
-  using ostream_type = OStream;
-  using flags_type = std::ios_base::fmtflags;
-  using fill_type = typename ostream_type::char_type;
-  using precision_type = std::streamsize;
-
-  ostream_state_saver(ostream_type& os,  // NOLINT(runtime/references)
-                      flags_type flags, fill_type fill)
-      : os_(os),
-        flags_(os.flags(flags)),
-        fill_(os.fill(fill)),
-        precision_(os.precision()) {
-    // Save state in initialized variables.
-  }
-
-  ~ostream_state_saver() {
-    // Restore saved state.
-    os_.precision(precision_);
-    os_.fill(fill_);
-    os_.flags(flags_);
-  }
-
- private:
-  ostream_type& os_;
-  const flags_type flags_;
-  const fill_type fill_;
-  const precision_type precision_;
-};
-
-#if defined(__NDK_MAJOR__) && __NDK_MAJOR__ < 16
-#define ABSL_RANDOM_INTERNAL_IOSTREAM_HEXFLOAT 1
-#else
-#define ABSL_RANDOM_INTERNAL_IOSTREAM_HEXFLOAT 0
-#endif
-
-template <typename CharT, typename Traits>
-ostream_state_saver<std::basic_ostream<CharT, Traits>> make_ostream_state_saver(
-    std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
-    std::ios_base::fmtflags flags = std::ios_base::dec | std::ios_base::left |
-#if ABSL_RANDOM_INTERNAL_IOSTREAM_HEXFLOAT
-                                    std::ios_base::fixed |
-#endif
-                                    std::ios_base::scientific) {
-  using result_type = ostream_state_saver<std::basic_ostream<CharT, Traits>>;
-  return result_type(os, flags, os.widen(' '));
-}
-
-template <typename T>
-typename absl::enable_if_t<!std::is_base_of<std::ios_base, T>::value,
-                           null_state_saver<T>>
-make_ostream_state_saver(T& is,  // NOLINT(runtime/references)
-                         std::ios_base::fmtflags flags = std::ios_base::dec) {
-  std::cerr << "null_state_saver";
-  using result_type = null_state_saver<T>;
-  return result_type(is, flags);
-}
-
-// stream_precision_helper<type>::kPrecision returns the base 10 precision
-// required to stream and reconstruct a real type exact binary value through
-// a binary->decimal->binary transition.
-template <typename T>
-struct stream_precision_helper {
-  // max_digits10 may be 0 on MSVC; if so, use digits10 + 3.
-  static constexpr int kPrecision =
-      (std::numeric_limits<T>::max_digits10 > std::numeric_limits<T>::digits10)
-          ? std::numeric_limits<T>::max_digits10
-          : (std::numeric_limits<T>::digits10 + 3);
-};
-
-template <>
-struct stream_precision_helper<float> {
-  static constexpr int kPrecision = 9;
-};
-template <>
-struct stream_precision_helper<double> {
-  static constexpr int kPrecision = 17;
-};
-template <>
-struct stream_precision_helper<long double> {
-  static constexpr int kPrecision = 36;  // assuming fp128
-};
-
-// istream_state_saver is a RAII object to save and restore the common
-// std::basic_istream<> flags used when implementing `operator >>()` on any of
-// the absl random distributions.
-template <typename IStream>
-class istream_state_saver {
- public:
-  using istream_type = IStream;
-  using flags_type = std::ios_base::fmtflags;
-
-  istream_state_saver(istream_type& is,  // NOLINT(runtime/references)
-                      flags_type flags)
-      : is_(is), flags_(is.flags(flags)) {}
-
-  ~istream_state_saver() { is_.flags(flags_); }
-
- private:
-  istream_type& is_;
-  flags_type flags_;
-};
-
-template <typename CharT, typename Traits>
-istream_state_saver<std::basic_istream<CharT, Traits>> make_istream_state_saver(
-    std::basic_istream<CharT, Traits>& is,  // NOLINT(runtime/references)
-    std::ios_base::fmtflags flags = std::ios_base::dec |
-                                    std::ios_base::scientific |
-                                    std::ios_base::skipws) {
-  using result_type = istream_state_saver<std::basic_istream<CharT, Traits>>;
-  return result_type(is, flags);
-}
-
-template <typename T>
-typename absl::enable_if_t<!std::is_base_of<std::ios_base, T>::value,
-                           null_state_saver<T>>
-make_istream_state_saver(T& is,  // NOLINT(runtime/references)
-                         std::ios_base::fmtflags flags = std::ios_base::dec) {
-  using result_type = null_state_saver<T>;
-  return result_type(is, flags);
-}
-
-// stream_format_type<T> is a helper struct to convert types which
-// basic_iostream cannot output as decimal numbers into types which
-// basic_iostream can output as decimal numbers. Specifically:
-// * signed/unsigned char-width types are converted to int.
-// * TODO(lar): __int128 => uint128, except there is no operator << yet.
-//
-template <typename T>
-struct stream_format_type
-    : public std::conditional<(sizeof(T) == sizeof(char)), int, T> {};
-
-// stream_u128_helper allows us to write out either absl::uint128 or
-// __uint128_t types in the same way, which enables their use as internal
-// state of PRNG engines.
-template <typename T>
-struct stream_u128_helper;
-
-template <>
-struct stream_u128_helper<absl::uint128> {
-  template <typename IStream>
-  inline absl::uint128 read(IStream& in) {
-    uint64_t h = 0;
-    uint64_t l = 0;
-    in >> h >> l;
-    return absl::MakeUint128(h, l);
-  }
-
-  template <typename OStream>
-  inline void write(absl::uint128 val, OStream& out) {
-    uint64_t h = absl::Uint128High64(val);
-    uint64_t l = absl::Uint128Low64(val);
-    out << h << out.fill() << l;
-  }
-};
-
-#ifdef ABSL_HAVE_INTRINSIC_INT128
-template <>
-struct stream_u128_helper<__uint128_t> {
-  template <typename IStream>
-  inline __uint128_t read(IStream& in) {
-    uint64_t h = 0;
-    uint64_t l = 0;
-    in >> h >> l;
-    return (static_cast<__uint128_t>(h) << 64) | l;
-  }
-
-  template <typename OStream>
-  inline void write(__uint128_t val, OStream& out) {
-    uint64_t h = static_cast<uint64_t>(val >> 64u);
-    uint64_t l = static_cast<uint64_t>(val);
-    out << h << out.fill() << l;
-  }
-};
-#endif
-
-template <typename FloatType, typename IStream>
-inline FloatType read_floating_point(IStream& is) {
-  static_assert(std::is_floating_point<FloatType>::value, "");
-  FloatType dest;
-  is >> dest;
-  // Parsing a double value may report a subnormal value as an error
-  // despite being able to represent it.
-  // See https://stackoverflow.com/q/52410931/3286653
-  // It may also report an underflow when parsing DOUBLE_MIN as an
-  // ERANGE error, as the parsed value may be smaller than DOUBLE_MIN
-  // and rounded up.
-  // See: https://stackoverflow.com/q/42005462
-  if (is.fail() &&
-      (std::fabs(dest) == (std::numeric_limits<FloatType>::min)() ||
-       std::fpclassify(dest) == FP_SUBNORMAL)) {
-    is.clear(is.rdstate() & (~std::ios_base::failbit));
-  }
-  return dest;
-}
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_IOSTREAM_STATE_SAVER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/mock_helpers.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/mock_helpers.h
deleted file mode 100644
index a7a97bfc..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/mock_helpers.h
+++ /dev/null
@@ -1,135 +0,0 @@
-//
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_MOCK_HELPERS_H_
-#define ABSL_RANDOM_INTERNAL_MOCK_HELPERS_H_
-
-#include <tuple>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/internal/fast_type_id.h"
-#include "absl/types/optional.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// MockHelpers works in conjunction with MockOverloadSet, MockingBitGen, and
-// BitGenRef to enable the mocking capability for absl distribution functions.
-//
-// MockingBitGen registers mocks based on the typeid of a mock signature, KeyT,
-// which is used to generate a unique id.
-//
-// KeyT is a signature of the form:
-//   result_type(discriminator_type, std::tuple<args...>)
-// The mocked function signature will be composed from KeyT as:
-//   result_type(args...)
-//
-class MockHelpers {
-  using IdType = ::absl::base_internal::FastTypeIdType;
-
-  // Given a key signature type used to index the mock, extract the components.
-  // KeyT is expected to have the form:
-  //   result_type(discriminator_type, arg_tuple_type)
-  template <typename KeyT>
-  struct KeySignature;
-
-  template <typename ResultT, typename DiscriminatorT, typename ArgTupleT>
-  struct KeySignature<ResultT(DiscriminatorT, ArgTupleT)> {
-    using result_type = ResultT;
-    using discriminator_type = DiscriminatorT;
-    using arg_tuple_type = ArgTupleT;
-  };
-
-  // Detector for InvokeMock.
-  template <class T>
-  using invoke_mock_t = decltype(std::declval<T*>()->InvokeMock(
-      std::declval<IdType>(), std::declval<void*>(), std::declval<void*>()));
-
-  // Empty implementation of InvokeMock.
-  template <typename KeyT, typename ReturnT, typename ArgTupleT, typename URBG,
-            typename... Args>
-  static absl::optional<ReturnT> InvokeMockImpl(char, URBG*, Args&&...) {
-    return absl::nullopt;
-  }
-
-  // Non-empty implementation of InvokeMock.
-  template <typename KeyT, typename ReturnT, typename ArgTupleT, typename URBG,
-            typename = invoke_mock_t<URBG>, typename... Args>
-  static absl::optional<ReturnT> InvokeMockImpl(int, URBG* urbg,
-                                                Args&&... args) {
-    ArgTupleT arg_tuple(std::forward<Args>(args)...);
-    ReturnT result;
-    if (urbg->InvokeMock(::absl::base_internal::FastTypeId<KeyT>(), &arg_tuple,
-                         &result)) {
-      return result;
-    }
-    return absl::nullopt;
-  }
-
- public:
-  // InvokeMock is private; this provides access for some specialized use cases.
-  template <typename URBG>
-  static inline bool PrivateInvokeMock(URBG* urbg, IdType type,
-                                       void* args_tuple, void* result) {
-    return urbg->InvokeMock(type, args_tuple, result);
-  }
-
-  // Invoke a mock for the KeyT (may or may not be a signature).
-  //
-  // KeyT is used to generate a typeid-based lookup key for the mock.
-  // KeyT is a signature of the form:
-  //   result_type(discriminator_type, std::tuple<args...>)
-  // The mocked function signature will be composed from KeyT as:
-  //   result_type(args...)
-  //
-  // An instance of arg_tuple_type must be constructable from Args..., since
-  // the underlying mechanism requires a pointer to an argument tuple.
-  template <typename KeyT, typename URBG, typename... Args>
-  static auto MaybeInvokeMock(URBG* urbg, Args&&... args)
-      -> absl::optional<typename KeySignature<KeyT>::result_type> {
-    // Use function overloading to dispatch to the implementation since
-    // more modern patterns (e.g. require + constexpr) are not supported in all
-    // compiler configurations.
-    return InvokeMockImpl<KeyT, typename KeySignature<KeyT>::result_type,
-                          typename KeySignature<KeyT>::arg_tuple_type, URBG>(
-        0, urbg, std::forward<Args>(args)...);
-  }
-
-  // Acquire a mock for the KeyT (may or may not be a signature).
-  //
-  // KeyT is used to generate a typeid-based lookup for the mock.
-  // KeyT is a signature of the form:
-  //   result_type(discriminator_type, std::tuple<args...>)
-  // The mocked function signature will be composed from KeyT as:
-  //   result_type(args...)
-  template <typename KeyT, typename MockURBG>
-  static auto MockFor(MockURBG& m)
-      -> decltype(m.template RegisterMock<
-                  typename KeySignature<KeyT>::result_type,
-                  typename KeySignature<KeyT>::arg_tuple_type>(
-          m, std::declval<IdType>())) {
-    return m.template RegisterMock<typename KeySignature<KeyT>::result_type,
-                                   typename KeySignature<KeyT>::arg_tuple_type>(
-        m, ::absl::base_internal::FastTypeId<KeyT>());
-  }
-};
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_MOCK_HELPERS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/mock_overload_set.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/mock_overload_set.h
deleted file mode 100644
index 0d9c6c12..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/mock_overload_set.h
+++ /dev/null
@@ -1,100 +0,0 @@
-//
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_MOCK_OVERLOAD_SET_H_
-#define ABSL_RANDOM_INTERNAL_MOCK_OVERLOAD_SET_H_
-
-#include <type_traits>
-
-#include "gmock/gmock.h"
-#include "absl/random/internal/mock_helpers.h"
-#include "absl/random/mocking_bit_gen.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-template <typename DistrT, typename Fn>
-struct MockSingleOverload;
-
-// MockSingleOverload
-//
-// MockSingleOverload hooks in to gMock's `ON_CALL` and `EXPECT_CALL` macros.
-// EXPECT_CALL(mock_single_overload, Call(...))` will expand to a call to
-// `mock_single_overload.gmock_Call(...)`. Because expectations are stored on
-// the MockingBitGen (an argument passed inside `Call(...)`), this forwards to
-// arguments to MockingBitGen::Register.
-//
-// The underlying KeyT must match the KeyT constructed by DistributionCaller.
-template <typename DistrT, typename Ret, typename... Args>
-struct MockSingleOverload<DistrT, Ret(MockingBitGen&, Args...)> {
-  static_assert(std::is_same<typename DistrT::result_type, Ret>::value,
-                "Overload signature must have return type matching the "
-                "distribution result_type.");
-  using KeyT = Ret(DistrT, std::tuple<Args...>);
-
-  template <typename MockURBG>
-  auto gmock_Call(MockURBG& gen, const ::testing::Matcher<Args>&... matchers)
-      -> decltype(MockHelpers::MockFor<KeyT>(gen).gmock_Call(matchers...)) {
-    static_assert(std::is_base_of<MockingBitGen, MockURBG>::value,
-                  "Mocking requires an absl::MockingBitGen");
-    return MockHelpers::MockFor<KeyT>(gen).gmock_Call(matchers...);
-  }
-};
-
-template <typename DistrT, typename Ret, typename Arg, typename... Args>
-struct MockSingleOverload<DistrT, Ret(Arg, MockingBitGen&, Args...)> {
-  static_assert(std::is_same<typename DistrT::result_type, Ret>::value,
-                "Overload signature must have return type matching the "
-                "distribution result_type.");
-  using KeyT = Ret(DistrT, std::tuple<Arg, Args...>);
-
-  template <typename MockURBG>
-  auto gmock_Call(const ::testing::Matcher<Arg>& matcher, MockURBG& gen,
-                  const ::testing::Matcher<Args>&... matchers)
-      -> decltype(MockHelpers::MockFor<KeyT>(gen).gmock_Call(matcher,
-                                                             matchers...)) {
-    static_assert(std::is_base_of<MockingBitGen, MockURBG>::value,
-                  "Mocking requires an absl::MockingBitGen");
-    return MockHelpers::MockFor<KeyT>(gen).gmock_Call(matcher, matchers...);
-  }
-};
-
-// MockOverloadSet
-//
-// MockOverloadSet takes a distribution and a collection of signatures and
-// performs overload resolution amongst all the overloads. This makes
-// `EXPECT_CALL(mock_overload_set, Call(...))` expand and do overload resolution
-// correctly.
-template <typename DistrT, typename... Signatures>
-struct MockOverloadSet;
-
-template <typename DistrT, typename Sig>
-struct MockOverloadSet<DistrT, Sig> : public MockSingleOverload<DistrT, Sig> {
-  using MockSingleOverload<DistrT, Sig>::gmock_Call;
-};
-
-template <typename DistrT, typename FirstSig, typename... Rest>
-struct MockOverloadSet<DistrT, FirstSig, Rest...>
-    : public MockSingleOverload<DistrT, FirstSig>,
-      public MockOverloadSet<DistrT, Rest...> {
-  using MockSingleOverload<DistrT, FirstSig>::gmock_Call;
-  using MockOverloadSet<DistrT, Rest...>::gmock_Call;
-};
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-#endif  // ABSL_RANDOM_INTERNAL_MOCK_OVERLOAD_SET_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/nanobenchmark.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/nanobenchmark.h
deleted file mode 100644
index a5097ba2..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/nanobenchmark.h
+++ /dev/null
@@ -1,172 +0,0 @@
-// Copyright 2017 Google Inc. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_NANOBENCHMARK_H_
-#define ABSL_RANDOM_INTERNAL_NANOBENCHMARK_H_
-
-// Benchmarks functions of a single integer argument with realistic branch
-// prediction hit rates. Uses a robust estimator to summarize the measurements.
-// The precision is about 0.2%.
-//
-// Examples: see nanobenchmark_test.cc.
-//
-// Background: Microbenchmarks such as http://github.com/google/benchmark
-// can measure elapsed times on the order of a microsecond. Shorter functions
-// are typically measured by repeating them thousands of times and dividing
-// the total elapsed time by this count. Unfortunately, repetition (especially
-// with the same input parameter!) influences the runtime. In time-critical
-// code, it is reasonable to expect warm instruction/data caches and TLBs,
-// but a perfect record of which branches will be taken is unrealistic.
-// Unless the application also repeatedly invokes the measured function with
-// the same parameter, the benchmark is measuring something very different -
-// a best-case result, almost as if the parameter were made a compile-time
-// constant. This may lead to erroneous conclusions about branch-heavy
-// algorithms outperforming branch-free alternatives.
-//
-// Our approach differs in three ways. Adding fences to the timer functions
-// reduces variability due to instruction reordering, improving the timer
-// resolution to about 40 CPU cycles. However, shorter functions must still
-// be invoked repeatedly. For more realistic branch prediction performance,
-// we vary the input parameter according to a user-specified distribution.
-// Thus, instead of VaryInputs(Measure(Repeat(func))), we change the
-// loop nesting to Measure(Repeat(VaryInputs(func))). We also estimate the
-// central tendency of the measurement samples with the "half sample mode",
-// which is more robust to outliers and skewed data than the mean or median.
-
-// NOTE: for compatibility with multiple translation units compiled with
-// distinct flags, avoid #including headers that define functions.
-
-#include <stddef.h>
-#include <stdint.h>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal_nanobenchmark {
-
-// Input influencing the function being measured (e.g. number of bytes to copy).
-using FuncInput = size_t;
-
-// "Proof of work" returned by Func to ensure the compiler does not elide it.
-using FuncOutput = uint64_t;
-
-// Function to measure: either 1) a captureless lambda or function with two
-// arguments or 2) a lambda with capture, in which case the first argument
-// is reserved for use by MeasureClosure.
-using Func = FuncOutput (*)(const void*, FuncInput);
-
-// Internal parameters that determine precision/resolution/measuring time.
-struct Params {
-  // For measuring timer overhead/resolution. Used in a nested loop =>
-  // quadratic time, acceptable because we know timer overhead is "low".
-  // constexpr because this is used to define array bounds.
-  static constexpr size_t kTimerSamples = 256;
-
-  // Best-case precision, expressed as a divisor of the timer resolution.
-  // Larger => more calls to Func and higher precision.
-  size_t precision_divisor = 1024;
-
-  // Ratio between full and subset input distribution sizes. Cannot be less
-  // than 2; larger values increase measurement time but more faithfully
-  // model the given input distribution.
-  size_t subset_ratio = 2;
-
-  // Together with the estimated Func duration, determines how many times to
-  // call Func before checking the sample variability. Larger values increase
-  // measurement time, memory/cache use and precision.
-  double seconds_per_eval = 4E-3;
-
-  // The minimum number of samples before estimating the central tendency.
-  size_t min_samples_per_eval = 7;
-
-  // The mode is better than median for estimating the central tendency of
-  // skewed/fat-tailed distributions, but it requires sufficient samples
-  // relative to the width of half-ranges.
-  size_t min_mode_samples = 64;
-
-  // Maximum permissible variability (= median absolute deviation / center).
-  double target_rel_mad = 0.002;
-
-  // Abort after this many evals without reaching target_rel_mad. This
-  // prevents infinite loops.
-  size_t max_evals = 9;
-
-  // Retry the measure loop up to this many times.
-  size_t max_measure_retries = 2;
-
-  // Whether to print additional statistics to stdout.
-  bool verbose = true;
-};
-
-// Measurement result for each unique input.
-struct Result {
-  FuncInput input;
-
-  // Robust estimate (mode or median) of duration.
-  float ticks;
-
-  // Measure of variability (median absolute deviation relative to "ticks").
-  float variability;
-};
-
-// Ensures the thread is running on the specified cpu, and no others.
-// Reduces noise due to desynchronized socket RDTSC and context switches.
-// If "cpu" is negative, pin to the currently running core.
-void PinThreadToCPU(const int cpu = -1);
-
-// Returns tick rate, useful for converting measurements to seconds. Invariant
-// means the tick counter frequency is independent of CPU throttling or sleep.
-// This call may be expensive, callers should cache the result.
-double InvariantTicksPerSecond();
-
-// Precisely measures the number of ticks elapsed when calling "func" with the
-// given inputs, shuffled to ensure realistic branch prediction hit rates.
-//
-// "func" returns a 'proof of work' to ensure its computations are not elided.
-// "arg" is passed to Func, or reserved for internal use by MeasureClosure.
-// "inputs" is an array of "num_inputs" (not necessarily unique) arguments to
-//   "func". The values should be chosen to maximize coverage of "func". This
-//   represents a distribution, so a value's frequency should reflect its
-//   probability in the real application. Order does not matter; for example, a
-//   uniform distribution over [0, 4) could be represented as {3,0,2,1}.
-// Returns how many Result were written to "results": one per unique input, or
-//   zero if the measurement failed (an error message goes to stderr).
-size_t Measure(const Func func, const void* arg, const FuncInput* inputs,
-               const size_t num_inputs, Result* results,
-               const Params& p = Params());
-
-// Calls operator() of the given closure (lambda function).
-template <class Closure>
-static FuncOutput CallClosure(const void* f, const FuncInput input) {
-  return (*reinterpret_cast<const Closure*>(f))(input);
-}
-
-// Same as Measure, except "closure" is typically a lambda function of
-// FuncInput -> FuncOutput with a capture list.
-template <class Closure>
-static inline size_t MeasureClosure(const Closure& closure,
-                                    const FuncInput* inputs,
-                                    const size_t num_inputs, Result* results,
-                                    const Params& p = Params()) {
-  return Measure(reinterpret_cast<Func>(&CallClosure<Closure>),
-                 reinterpret_cast<const void*>(&closure), inputs, num_inputs,
-                 results, p);
-}
-
-}  // namespace random_internal_nanobenchmark
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_NANOBENCHMARK_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/nonsecure_base.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/nonsecure_base.h
deleted file mode 100644
index c3b80335..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/nonsecure_base.h
+++ /dev/null
@@ -1,161 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_NONSECURE_BASE_H_
-#define ABSL_RANDOM_INTERNAL_NONSECURE_BASE_H_
-
-#include <algorithm>
-#include <cstdint>
-#include <iterator>
-#include <type_traits>
-#include <utility>
-#include <vector>
-
-#include "absl/base/macros.h"
-#include "absl/container/inlined_vector.h"
-#include "absl/meta/type_traits.h"
-#include "absl/random/internal/pool_urbg.h"
-#include "absl/random/internal/salted_seed_seq.h"
-#include "absl/random/internal/seed_material.h"
-#include "absl/types/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// RandenPoolSeedSeq is a custom seed sequence type where generate() fills the
-// provided buffer via the RandenPool entropy source.
-class RandenPoolSeedSeq {
- private:
-  struct ContiguousTag {};
-  struct BufferTag {};
-
-  // Generate random unsigned values directly into the buffer.
-  template <typename Contiguous>
-  void generate_impl(ContiguousTag, Contiguous begin, Contiguous end) {
-    const size_t n = static_cast<size_t>(std::distance(begin, end));
-    auto* a = &(*begin);
-    RandenPool<uint8_t>::Fill(
-        absl::MakeSpan(reinterpret_cast<uint8_t*>(a), sizeof(*a) * n));
-  }
-
-  // Construct a buffer of size n and fill it with values, then copy
-  // those values into the seed iterators.
-  template <typename RandomAccessIterator>
-  void generate_impl(BufferTag, RandomAccessIterator begin,
-                     RandomAccessIterator end) {
-    const size_t n = std::distance(begin, end);
-    absl::InlinedVector<uint32_t, 8> data(n, 0);
-    RandenPool<uint32_t>::Fill(absl::MakeSpan(data.begin(), data.end()));
-    std::copy(std::begin(data), std::end(data), begin);
-  }
-
- public:
-  using result_type = uint32_t;
-
-  size_t size() { return 0; }
-
-  template <typename OutIterator>
-  void param(OutIterator) const {}
-
-  template <typename RandomAccessIterator>
-  void generate(RandomAccessIterator begin, RandomAccessIterator end) {
-    // RandomAccessIterator must be assignable from uint32_t
-    if (begin != end) {
-      using U = typename std::iterator_traits<RandomAccessIterator>::value_type;
-      // ContiguousTag indicates the common case of a known contiguous buffer,
-      // which allows directly filling the buffer. In C++20,
-      // std::contiguous_iterator_tag provides a mechanism for testing this
-      // capability, however until Abseil's support requirements allow us to
-      // assume C++20, limit checks to a few common cases.
-      using TagType = absl::conditional_t<
-          (std::is_pointer<RandomAccessIterator>::value ||
-           std::is_same<RandomAccessIterator,
-                        typename std::vector<U>::iterator>::value),
-          ContiguousTag, BufferTag>;
-
-      generate_impl(TagType{}, begin, end);
-    }
-  }
-};
-
-// Each instance of NonsecureURBGBase<URBG> will be seeded by variates produced
-// by a thread-unique URBG-instance.
-template <typename URBG, typename Seeder = RandenPoolSeedSeq>
-class NonsecureURBGBase {
- public:
-  using result_type = typename URBG::result_type;
-
-  // Default constructor
-  NonsecureURBGBase() : urbg_(ConstructURBG()) {}
-
-  // Copy disallowed, move allowed.
-  NonsecureURBGBase(const NonsecureURBGBase&) = delete;
-  NonsecureURBGBase& operator=(const NonsecureURBGBase&) = delete;
-  NonsecureURBGBase(NonsecureURBGBase&&) = default;
-  NonsecureURBGBase& operator=(NonsecureURBGBase&&) = default;
-
-  // Constructor using a seed
-  template <class SSeq, typename = typename absl::enable_if_t<
-                            !std::is_same<SSeq, NonsecureURBGBase>::value>>
-  explicit NonsecureURBGBase(SSeq&& seq)
-      : urbg_(ConstructURBG(std::forward<SSeq>(seq))) {}
-
-  // Note: on MSVC, min() or max() can be interpreted as MIN() or MAX(), so we
-  // enclose min() or max() in parens as (min)() and (max)().
-  // Additionally, clang-format requires no space before this construction.
-
-  // NonsecureURBGBase::min()
-  static constexpr result_type(min)() { return (URBG::min)(); }
-
-  // NonsecureURBGBase::max()
-  static constexpr result_type(max)() { return (URBG::max)(); }
-
-  // NonsecureURBGBase::operator()()
-  result_type operator()() { return urbg_(); }
-
-  // NonsecureURBGBase::discard()
-  void discard(unsigned long long values) {  // NOLINT(runtime/int)
-    urbg_.discard(values);
-  }
-
-  bool operator==(const NonsecureURBGBase& other) const {
-    return urbg_ == other.urbg_;
-  }
-
-  bool operator!=(const NonsecureURBGBase& other) const {
-    return !(urbg_ == other.urbg_);
-  }
-
- private:
-  static URBG ConstructURBG() {
-    Seeder seeder;
-    return URBG(seeder);
-  }
-
-  template <typename SSeq>
-  static URBG ConstructURBG(SSeq&& seq) {  // NOLINT(runtime/references)
-    auto salted_seq =
-        random_internal::MakeSaltedSeedSeq(std::forward<SSeq>(seq));
-    return URBG(salted_seq);
-  }
-
-  URBG urbg_;
-};
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_NONSECURE_BASE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/pcg_engine.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/pcg_engine.h
deleted file mode 100644
index e1f4ef33..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/pcg_engine.h
+++ /dev/null
@@ -1,287 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_PCG_ENGINE_H_
-#define ABSL_RANDOM_INTERNAL_PCG_ENGINE_H_
-
-#include <type_traits>
-
-#include "absl/base/config.h"
-#include "absl/meta/type_traits.h"
-#include "absl/numeric/bits.h"
-#include "absl/numeric/int128.h"
-#include "absl/random/internal/fastmath.h"
-#include "absl/random/internal/iostream_state_saver.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// pcg_engine is a simplified implementation of Melissa O'Neil's PCG engine in
-// C++.  PCG combines a linear congruential generator (LCG) with output state
-// mixing functions to generate each random variate.  pcg_engine supports only a
-// single sequence (oneseq), and does not support streams.
-//
-// pcg_engine is parameterized by two types:
-//   Params, which provides the multiplier and increment values;
-//   Mix, which mixes the state into the result.
-//
-template <typename Params, typename Mix>
-class pcg_engine {
-  static_assert(std::is_same<typename Params::state_type,
-                             typename Mix::state_type>::value,
-                "Class-template absl::pcg_engine must be parameterized by "
-                "Params and Mix with identical state_type");
-
-  static_assert(std::is_unsigned<typename Mix::result_type>::value,
-                "Class-template absl::pcg_engine must be parameterized by "
-                "an unsigned Mix::result_type");
-
-  using params_type = Params;
-  using mix_type = Mix;
-  using state_type = typename Mix::state_type;
-
- public:
-  // C++11 URBG interface:
-  using result_type = typename Mix::result_type;
-
-  static constexpr result_type(min)() {
-    return (std::numeric_limits<result_type>::min)();
-  }
-
-  static constexpr result_type(max)() {
-    return (std::numeric_limits<result_type>::max)();
-  }
-
-  explicit pcg_engine(uint64_t seed_value = 0) { seed(seed_value); }
-
-  template <class SeedSequence,
-            typename = typename absl::enable_if_t<
-                !std::is_same<SeedSequence, pcg_engine>::value>>
-  explicit pcg_engine(SeedSequence&& seq) {
-    seed(seq);
-  }
-
-  pcg_engine(const pcg_engine&) = default;
-  pcg_engine& operator=(const pcg_engine&) = default;
-  pcg_engine(pcg_engine&&) = default;
-  pcg_engine& operator=(pcg_engine&&) = default;
-
-  result_type operator()() {
-    // Advance the LCG state, always using the new value to generate the output.
-    state_ = lcg(state_);
-    return Mix{}(state_);
-  }
-
-  void seed(uint64_t seed_value = 0) {
-    state_type tmp = seed_value;
-    state_ = lcg(tmp + Params::increment());
-  }
-
-  template <class SeedSequence>
-  typename absl::enable_if_t<
-      !std::is_convertible<SeedSequence, uint64_t>::value, void>
-  seed(SeedSequence&& seq) {
-    reseed(seq);
-  }
-
-  void discard(uint64_t count) { state_ = advance(state_, count); }
-
-  bool operator==(const pcg_engine& other) const {
-    return state_ == other.state_;
-  }
-
-  bool operator!=(const pcg_engine& other) const { return !(*this == other); }
-
-  template <class CharT, class Traits>
-  friend typename absl::enable_if_t<(sizeof(state_type) == 16),
-                                    std::basic_ostream<CharT, Traits>&>
-  operator<<(
-      std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
-      const pcg_engine& engine) {
-    auto saver = random_internal::make_ostream_state_saver(os);
-    random_internal::stream_u128_helper<state_type> helper;
-    helper.write(pcg_engine::params_type::multiplier(), os);
-    os << os.fill();
-    helper.write(pcg_engine::params_type::increment(), os);
-    os << os.fill();
-    helper.write(engine.state_, os);
-    return os;
-  }
-
-  template <class CharT, class Traits>
-  friend typename absl::enable_if_t<(sizeof(state_type) <= 8),
-                                    std::basic_ostream<CharT, Traits>&>
-  operator<<(
-      std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
-      const pcg_engine& engine) {
-    auto saver = random_internal::make_ostream_state_saver(os);
-    os << pcg_engine::params_type::multiplier() << os.fill();
-    os << pcg_engine::params_type::increment() << os.fill();
-    os << engine.state_;
-    return os;
-  }
-
-  template <class CharT, class Traits>
-  friend typename absl::enable_if_t<(sizeof(state_type) == 16),
-                                    std::basic_istream<CharT, Traits>&>
-  operator>>(
-      std::basic_istream<CharT, Traits>& is,  // NOLINT(runtime/references)
-      pcg_engine& engine) {                   // NOLINT(runtime/references)
-    random_internal::stream_u128_helper<state_type> helper;
-    auto mult = helper.read(is);
-    auto inc = helper.read(is);
-    auto tmp = helper.read(is);
-    if (mult != pcg_engine::params_type::multiplier() ||
-        inc != pcg_engine::params_type::increment()) {
-      // signal failure by setting the failbit.
-      is.setstate(is.rdstate() | std::ios_base::failbit);
-    }
-    if (!is.fail()) {
-      engine.state_ = tmp;
-    }
-    return is;
-  }
-
-  template <class CharT, class Traits>
-  friend typename absl::enable_if_t<(sizeof(state_type) <= 8),
-                                    std::basic_istream<CharT, Traits>&>
-  operator>>(
-      std::basic_istream<CharT, Traits>& is,  // NOLINT(runtime/references)
-      pcg_engine& engine) {                   // NOLINT(runtime/references)
-    state_type mult{}, inc{}, tmp{};
-    is >> mult >> inc >> tmp;
-    if (mult != pcg_engine::params_type::multiplier() ||
-        inc != pcg_engine::params_type::increment()) {
-      // signal failure by setting the failbit.
-      is.setstate(is.rdstate() | std::ios_base::failbit);
-    }
-    if (!is.fail()) {
-      engine.state_ = tmp;
-    }
-    return is;
-  }
-
- private:
-  state_type state_;
-
-  // Returns the linear-congruential generator next state.
-  static inline constexpr state_type lcg(state_type s) {
-    return s * Params::multiplier() + Params::increment();
-  }
-
-  // Returns the linear-congruential arbitrary seek state.
-  inline state_type advance(state_type s, uint64_t n) const {
-    state_type mult = Params::multiplier();
-    state_type inc = Params::increment();
-    state_type m = 1;
-    state_type i = 0;
-    while (n > 0) {
-      if (n & 1) {
-        m *= mult;
-        i = i * mult + inc;
-      }
-      inc = (mult + 1) * inc;
-      mult *= mult;
-      n >>= 1;
-    }
-    return m * s + i;
-  }
-
-  template <class SeedSequence>
-  void reseed(SeedSequence& seq) {
-    using sequence_result_type = typename SeedSequence::result_type;
-    constexpr size_t kBufferSize =
-        sizeof(state_type) / sizeof(sequence_result_type);
-    sequence_result_type buffer[kBufferSize];
-    seq.generate(std::begin(buffer), std::end(buffer));
-    // Convert the seed output to a single state value.
-    state_type tmp = buffer[0];
-    for (size_t i = 1; i < kBufferSize; i++) {
-      tmp <<= (sizeof(sequence_result_type) * 8);
-      tmp |= buffer[i];
-    }
-    state_ = lcg(tmp + params_type::increment());
-  }
-};
-
-// Parameterized implementation of the PCG 128-bit oneseq state.
-// This provides state_type, multiplier, and increment for pcg_engine.
-template <uint64_t kMultA, uint64_t kMultB, uint64_t kIncA, uint64_t kIncB>
-class pcg128_params {
- public:
-  using state_type = absl::uint128;
-  static inline constexpr state_type multiplier() {
-    return absl::MakeUint128(kMultA, kMultB);
-  }
-  static inline constexpr state_type increment() {
-    return absl::MakeUint128(kIncA, kIncB);
-  }
-};
-
-// Implementation of the PCG xsl_rr_128_64 128-bit mixing function, which
-// accepts an input of state_type and mixes it into an output of result_type.
-struct pcg_xsl_rr_128_64 {
-  using state_type = absl::uint128;
-  using result_type = uint64_t;
-
-  inline uint64_t operator()(state_type state) {
-    // This is equivalent to the xsl_rr_128_64 mixing function.
-    uint64_t rotate = static_cast<uint64_t>(state >> 122u);
-    state ^= state >> 64;
-    uint64_t s = static_cast<uint64_t>(state);
-    return rotr(s, static_cast<int>(rotate));
-  }
-};
-
-// Parameterized implementation of the PCG 64-bit oneseq state.
-// This provides state_type, multiplier, and increment for pcg_engine.
-template <uint64_t kMult, uint64_t kInc>
-class pcg64_params {
- public:
-  using state_type = uint64_t;
-  static inline constexpr state_type multiplier() { return kMult; }
-  static inline constexpr state_type increment() { return kInc; }
-};
-
-// Implementation of the PCG xsh_rr_64_32 64-bit mixing function, which accepts
-// an input of state_type and mixes it into an output of result_type.
-struct pcg_xsh_rr_64_32 {
-  using state_type = uint64_t;
-  using result_type = uint32_t;
-  inline uint32_t operator()(uint64_t state) {
-    return rotr(static_cast<uint32_t>(((state >> 18) ^ state) >> 27),
-                state >> 59);
-  }
-};
-
-// Stable pcg_engine implementations:
-// This is a 64-bit generator using 128-bits of state.
-// The output sequence is equivalent to Melissa O'Neil's pcg64_oneseq.
-using pcg64_2018_engine = pcg_engine<
-    random_internal::pcg128_params<0x2360ed051fc65da4ull, 0x4385df649fccf645ull,
-                                   0x5851f42d4c957f2d, 0x14057b7ef767814f>,
-    random_internal::pcg_xsl_rr_128_64>;
-
-// This is a 32-bit generator using 64-bits of state.
-// This is equivalent to Melissa O'Neil's pcg32_oneseq.
-using pcg32_2018_engine = pcg_engine<
-    random_internal::pcg64_params<0x5851f42d4c957f2dull, 0x14057b7ef767814full>,
-    random_internal::pcg_xsh_rr_64_32>;
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_PCG_ENGINE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/platform.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/platform.h
deleted file mode 100644
index d779f481..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/platform.h
+++ /dev/null
@@ -1,171 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_PLATFORM_H_
-#define ABSL_RANDOM_INTERNAL_PLATFORM_H_
-
-// HERMETIC NOTE: The randen_hwaes target must not introduce duplicate
-// symbols from arbitrary system and other headers, since it may be built
-// with different flags from other targets, using different levels of
-// optimization, potentially introducing ODR violations.
-
-// -----------------------------------------------------------------------------
-// Platform Feature Checks
-// -----------------------------------------------------------------------------
-
-// Currently supported operating systems and associated preprocessor
-// symbols:
-//
-//   Linux and Linux-derived           __linux__
-//   Android                           __ANDROID__ (implies __linux__)
-//   Linux (non-Android)               __linux__ && !__ANDROID__
-//   Darwin (macOS and iOS)            __APPLE__
-//   Akaros (http://akaros.org)        __ros__
-//   Windows                           _WIN32
-//   NaCL                              __native_client__
-//   AsmJS                             __asmjs__
-//   WebAssembly                       __wasm__
-//   Fuchsia                           __Fuchsia__
-//
-// Note that since Android defines both __ANDROID__ and __linux__, one
-// may probe for either Linux or Android by simply testing for __linux__.
-//
-// NOTE: For __APPLE__ platforms, we use #include <TargetConditionals.h>
-// to distinguish os variants.
-//
-// http://nadeausoftware.com/articles/2012/01/c_c_tip_how_use_compiler_predefined_macros_detect_operating_system
-
-#if defined(__APPLE__)
-#include <TargetConditionals.h>
-#endif
-
-// -----------------------------------------------------------------------------
-// Architecture Checks
-// -----------------------------------------------------------------------------
-
-// These preprocessor directives are trying to determine CPU architecture,
-// including necessary headers to support hardware AES.
-//
-// ABSL_ARCH_{X86/PPC/ARM} macros determine the platform.
-#if defined(__x86_64__) || defined(__x86_64) || defined(_M_AMD64) || \
-    defined(_M_X64)
-#define ABSL_ARCH_X86_64
-#elif defined(__i386) || defined(_M_IX86)
-#define ABSL_ARCH_X86_32
-#elif defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64)
-#define ABSL_ARCH_AARCH64
-#elif defined(__arm__) || defined(__ARMEL__) || defined(_M_ARM)
-#define ABSL_ARCH_ARM
-#elif defined(__powerpc64__) || defined(__PPC64__) || defined(__powerpc__) || \
-    defined(__ppc__) || defined(__PPC__)
-#define ABSL_ARCH_PPC
-#else
-// Unsupported architecture.
-//  * https://sourceforge.net/p/predef/wiki/Architectures/
-//  * https://msdn.microsoft.com/en-us/library/b0084kay.aspx
-//  * for gcc, clang: "echo | gcc -E -dM -"
-#endif
-
-// -----------------------------------------------------------------------------
-// Attribute Checks
-// -----------------------------------------------------------------------------
-
-// ABSL_RANDOM_INTERNAL_RESTRICT annotates whether pointers may be considered
-// to be unaliased.
-#if defined(__clang__) || defined(__GNUC__)
-#define ABSL_RANDOM_INTERNAL_RESTRICT __restrict__
-#elif defined(_MSC_VER)
-#define ABSL_RANDOM_INTERNAL_RESTRICT __restrict
-#else
-#define ABSL_RANDOM_INTERNAL_RESTRICT
-#endif
-
-// ABSL_HAVE_ACCELERATED_AES indicates whether the currently active compiler
-// flags (e.g. -maes) allow using hardware accelerated AES instructions, which
-// implies us assuming that the target platform supports them.
-#define ABSL_HAVE_ACCELERATED_AES 0
-
-#if defined(ABSL_ARCH_X86_64)
-
-#if defined(__AES__) || defined(__AVX__)
-#undef ABSL_HAVE_ACCELERATED_AES
-#define ABSL_HAVE_ACCELERATED_AES 1
-#endif
-
-#elif defined(ABSL_ARCH_PPC)
-
-// Rely on VSX and CRYPTO extensions for vcipher on PowerPC.
-#if (defined(__VEC__) || defined(__ALTIVEC__)) && defined(__VSX__) && \
-    defined(__CRYPTO__)
-#undef ABSL_HAVE_ACCELERATED_AES
-#define ABSL_HAVE_ACCELERATED_AES 1
-#endif
-
-#elif defined(ABSL_ARCH_ARM) || defined(ABSL_ARCH_AARCH64)
-
-// http://infocenter.arm.com/help/topic/com.arm.doc.ihi0053c/IHI0053C_acle_2_0.pdf
-// Rely on NEON+CRYPTO extensions for ARM.
-#if defined(__ARM_NEON) && defined(__ARM_FEATURE_CRYPTO)
-#undef ABSL_HAVE_ACCELERATED_AES
-#define ABSL_HAVE_ACCELERATED_AES 1
-#endif
-
-#endif
-
-// NaCl does not allow AES.
-#if defined(__native_client__)
-#undef ABSL_HAVE_ACCELERATED_AES
-#define ABSL_HAVE_ACCELERATED_AES 0
-#endif
-
-// ABSL_RANDOM_INTERNAL_AES_DISPATCH indicates whether the currently active
-// platform has, or should use run-time dispatch for selecting the
-// accelerated Randen implementation.
-#define ABSL_RANDOM_INTERNAL_AES_DISPATCH 0
-
-#if defined(ABSL_ARCH_X86_64)
-// Dispatch is available on x86_64
-#undef ABSL_RANDOM_INTERNAL_AES_DISPATCH
-#define ABSL_RANDOM_INTERNAL_AES_DISPATCH 1
-#elif defined(__linux__) && defined(ABSL_ARCH_PPC)
-// Or when running linux PPC
-#undef ABSL_RANDOM_INTERNAL_AES_DISPATCH
-#define ABSL_RANDOM_INTERNAL_AES_DISPATCH 1
-#elif defined(__linux__) && defined(ABSL_ARCH_AARCH64)
-// Or when running linux AArch64
-#undef ABSL_RANDOM_INTERNAL_AES_DISPATCH
-#define ABSL_RANDOM_INTERNAL_AES_DISPATCH 1
-#elif defined(__linux__) && defined(ABSL_ARCH_ARM) && (__ARM_ARCH >= 8)
-// Or when running linux ARM v8 or higher.
-// (This captures a lot of Android configurations.)
-#undef ABSL_RANDOM_INTERNAL_AES_DISPATCH
-#define ABSL_RANDOM_INTERNAL_AES_DISPATCH 1
-#endif
-
-// NaCl does not allow dispatch.
-#if defined(__native_client__)
-#undef ABSL_RANDOM_INTERNAL_AES_DISPATCH
-#define ABSL_RANDOM_INTERNAL_AES_DISPATCH 0
-#endif
-
-// iOS does not support dispatch, even on x86, since applications
-// should be bundled as fat binaries, with a different build tailored for
-// each specific supported platform/architecture.
-#if (defined(TARGET_OS_IPHONE) && TARGET_OS_IPHONE) || \
-    (defined(TARGET_OS_IPHONE_SIMULATOR) && TARGET_OS_IPHONE_SIMULATOR)
-#undef ABSL_RANDOM_INTERNAL_AES_DISPATCH
-#define ABSL_RANDOM_INTERNAL_AES_DISPATCH 0
-#endif
-
-#endif  // ABSL_RANDOM_INTERNAL_PLATFORM_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/pool_urbg.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/pool_urbg.h
deleted file mode 100644
index 05721929..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/pool_urbg.h
+++ /dev/null
@@ -1,131 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_POOL_URBG_H_
-#define ABSL_RANDOM_INTERNAL_POOL_URBG_H_
-
-#include <cinttypes>
-#include <limits>
-
-#include "absl/random/internal/traits.h"
-#include "absl/types/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// RandenPool is a thread-safe random number generator [random.req.urbg] that
-// uses an underlying pool of Randen generators to generate values.  Each thread
-// has affinity to one instance of the underlying pool generators.  Concurrent
-// access is guarded by a spin-lock.
-template <typename T>
-class RandenPool {
- public:
-  using result_type = T;
-  static_assert(std::is_unsigned<result_type>::value,
-                "RandenPool template argument must be a built-in unsigned "
-                "integer type");
-
-  static constexpr result_type(min)() {
-    return (std::numeric_limits<result_type>::min)();
-  }
-
-  static constexpr result_type(max)() {
-    return (std::numeric_limits<result_type>::max)();
-  }
-
-  RandenPool() {}
-
-  // Returns a single value.
-  inline result_type operator()() { return Generate(); }
-
-  // Fill data with random values.
-  static void Fill(absl::Span<result_type> data);
-
- protected:
-  // Generate returns a single value.
-  static result_type Generate();
-};
-
-extern template class RandenPool<uint8_t>;
-extern template class RandenPool<uint16_t>;
-extern template class RandenPool<uint32_t>;
-extern template class RandenPool<uint64_t>;
-
-// PoolURBG uses an underlying pool of random generators to implement a
-// thread-compatible [random.req.urbg] interface with an internal cache of
-// values.
-template <typename T, size_t kBufferSize>
-class PoolURBG {
-  // Inheritance to access the protected static members of RandenPool.
-  using unsigned_type = typename make_unsigned_bits<T>::type;
-  using PoolType = RandenPool<unsigned_type>;
-  using SpanType = absl::Span<unsigned_type>;
-
-  static constexpr size_t kInitialBuffer = kBufferSize + 1;
-  static constexpr size_t kHalfBuffer = kBufferSize / 2;
-
- public:
-  using result_type = T;
-
-  static_assert(std::is_unsigned<result_type>::value,
-                "PoolURBG must be parameterized by an unsigned integer type");
-
-  static_assert(kBufferSize > 1,
-                "PoolURBG must be parameterized by a buffer-size > 1");
-
-  static_assert(kBufferSize <= 256,
-                "PoolURBG must be parameterized by a buffer-size <= 256");
-
-  static constexpr result_type(min)() {
-    return (std::numeric_limits<result_type>::min)();
-  }
-
-  static constexpr result_type(max)() {
-    return (std::numeric_limits<result_type>::max)();
-  }
-
-  PoolURBG() : next_(kInitialBuffer) {}
-
-  // copy-constructor does not copy cache.
-  PoolURBG(const PoolURBG&) : next_(kInitialBuffer) {}
-  const PoolURBG& operator=(const PoolURBG&) {
-    next_ = kInitialBuffer;
-    return *this;
-  }
-
-  // move-constructor does move cache.
-  PoolURBG(PoolURBG&&) = default;
-  PoolURBG& operator=(PoolURBG&&) = default;
-
-  inline result_type operator()() {
-    if (next_ >= kBufferSize) {
-      next_ = (kBufferSize > 2 && next_ > kBufferSize) ? kHalfBuffer : 0;
-      PoolType::Fill(SpanType(reinterpret_cast<unsigned_type*>(state_ + next_),
-                              kBufferSize - next_));
-    }
-    return state_[next_++];
-  }
-
- private:
-  // Buffer size.
-  size_t next_;  // index within state_
-  result_type state_[kBufferSize];
-};
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_POOL_URBG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen.h
deleted file mode 100644
index 9ff4a7a5..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen.h
+++ /dev/null
@@ -1,96 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_RANDEN_H_
-#define ABSL_RANDOM_INTERNAL_RANDEN_H_
-
-#include <cstddef>
-
-#include "absl/random/internal/platform.h"
-#include "absl/random/internal/randen_hwaes.h"
-#include "absl/random/internal/randen_slow.h"
-#include "absl/random/internal/randen_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// RANDen = RANDom generator or beetroots in Swiss High German.
-// 'Strong' (well-distributed, unpredictable, backtracking-resistant) random
-// generator, faster in some benchmarks than std::mt19937_64 and pcg64_c32.
-//
-// Randen implements the basic state manipulation methods.
-class Randen {
- public:
-  static constexpr size_t kStateBytes = RandenTraits::kStateBytes;
-  static constexpr size_t kCapacityBytes = RandenTraits::kCapacityBytes;
-  static constexpr size_t kSeedBytes = RandenTraits::kSeedBytes;
-
-  ~Randen() = default;
-
-  Randen();
-
-  // Generate updates the randen sponge. The outer portion of the sponge
-  // (kCapacityBytes .. kStateBytes) may be consumed as PRNG state.
-  // REQUIRES: state points to kStateBytes of state.
-  inline void Generate(void* state) const {
-#if ABSL_RANDOM_INTERNAL_AES_DISPATCH
-    // HW AES Dispatch.
-    if (has_crypto_) {
-      RandenHwAes::Generate(keys_, state);
-    } else {
-      RandenSlow::Generate(keys_, state);
-    }
-#elif ABSL_HAVE_ACCELERATED_AES
-    // HW AES is enabled.
-    RandenHwAes::Generate(keys_, state);
-#else
-    // HW AES is disabled.
-    RandenSlow::Generate(keys_, state);
-#endif
-  }
-
-  // Absorb incorporates additional seed material into the randen sponge.  After
-  // absorb returns, Generate must be called before the state may be consumed.
-  // REQUIRES: seed points to kSeedBytes of seed.
-  // REQUIRES: state points to kStateBytes of state.
-  inline void Absorb(const void* seed, void* state) const {
-#if ABSL_RANDOM_INTERNAL_AES_DISPATCH
-    // HW AES Dispatch.
-    if (has_crypto_) {
-      RandenHwAes::Absorb(seed, state);
-    } else {
-      RandenSlow::Absorb(seed, state);
-    }
-#elif ABSL_HAVE_ACCELERATED_AES
-    // HW AES is enabled.
-    RandenHwAes::Absorb(seed, state);
-#else
-    // HW AES is disabled.
-    RandenSlow::Absorb(seed, state);
-#endif
-  }
-
- private:
-  const void* keys_;
-#if ABSL_RANDOM_INTERNAL_AES_DISPATCH
-  bool has_crypto_;
-#endif
-};
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_RANDEN_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_detect.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_detect.h
deleted file mode 100644
index f283f432..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_detect.h
+++ /dev/null
@@ -1,33 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_RANDEN_DETECT_H_
-#define ABSL_RANDOM_INTERNAL_RANDEN_DETECT_H_
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// Returns whether the current CPU supports RandenHwAes implementation.
-// This typically involves supporting cryptographic extensions on whichever
-// platform is currently running.
-bool CPUSupportsRandenHwAes();
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_RANDEN_DETECT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_engine.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_engine.h
deleted file mode 100644
index fe2d9f6c..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_engine.h
+++ /dev/null
@@ -1,264 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_RANDEN_ENGINE_H_
-#define ABSL_RANDOM_INTERNAL_RANDEN_ENGINE_H_
-
-#include <algorithm>
-#include <cinttypes>
-#include <cstdlib>
-#include <iostream>
-#include <iterator>
-#include <limits>
-#include <type_traits>
-
-#include "absl/base/internal/endian.h"
-#include "absl/meta/type_traits.h"
-#include "absl/random/internal/iostream_state_saver.h"
-#include "absl/random/internal/randen.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// Deterministic pseudorandom byte generator with backtracking resistance
-// (leaking the state does not compromise prior outputs). Based on Reverie
-// (see "A Robust and Sponge-Like PRNG with Improved Efficiency") instantiated
-// with an improved Simpira-like permutation.
-// Returns values of type "T" (must be a built-in unsigned integer type).
-//
-// RANDen = RANDom generator or beetroots in Swiss High German.
-// 'Strong' (well-distributed, unpredictable, backtracking-resistant) random
-// generator, faster in some benchmarks than std::mt19937_64 and pcg64_c32.
-template <typename T>
-class alignas(8) randen_engine {
- public:
-  // C++11 URBG interface:
-  using result_type = T;
-  static_assert(std::is_unsigned<result_type>::value,
-                "randen_engine template argument must be a built-in unsigned "
-                "integer type");
-
-  static constexpr result_type(min)() {
-    return (std::numeric_limits<result_type>::min)();
-  }
-
-  static constexpr result_type(max)() {
-    return (std::numeric_limits<result_type>::max)();
-  }
-
-  randen_engine() : randen_engine(0) {}
-  explicit randen_engine(result_type seed_value) { seed(seed_value); }
-
-  template <class SeedSequence,
-            typename = typename absl::enable_if_t<
-                !std::is_same<SeedSequence, randen_engine>::value>>
-  explicit randen_engine(SeedSequence&& seq) {
-    seed(seq);
-  }
-
-  // alignment requirements dictate custom copy and move constructors.
-  randen_engine(const randen_engine& other)
-      : next_(other.next_), impl_(other.impl_) {
-    std::memcpy(state(), other.state(), kStateSizeT * sizeof(result_type));
-  }
-  randen_engine& operator=(const randen_engine& other) {
-    next_ = other.next_;
-    impl_ = other.impl_;
-    std::memcpy(state(), other.state(), kStateSizeT * sizeof(result_type));
-    return *this;
-  }
-
-  // Returns random bits from the buffer in units of result_type.
-  result_type operator()() {
-    // Refill the buffer if needed (unlikely).
-    auto* begin = state();
-    if (next_ >= kStateSizeT) {
-      next_ = kCapacityT;
-      impl_.Generate(begin);
-    }
-    return little_endian::ToHost(begin[next_++]);
-  }
-
-  template <class SeedSequence>
-  typename absl::enable_if_t<
-      !std::is_convertible<SeedSequence, result_type>::value>
-  seed(SeedSequence&& seq) {
-    // Zeroes the state.
-    seed();
-    reseed(seq);
-  }
-
-  void seed(result_type seed_value = 0) {
-    next_ = kStateSizeT;
-    // Zeroes the inner state and fills the outer state with seed_value to
-    // mimic the behaviour of reseed
-    auto* begin = state();
-    std::fill(begin, begin + kCapacityT, 0);
-    std::fill(begin + kCapacityT, begin + kStateSizeT, seed_value);
-  }
-
-  // Inserts entropy into (part of) the state. Calling this periodically with
-  // sufficient entropy ensures prediction resistance (attackers cannot predict
-  // future outputs even if state is compromised).
-  template <class SeedSequence>
-  void reseed(SeedSequence& seq) {
-    using sequence_result_type = typename SeedSequence::result_type;
-    static_assert(sizeof(sequence_result_type) == 4,
-                  "SeedSequence::result_type must be 32-bit");
-    constexpr size_t kBufferSize =
-        Randen::kSeedBytes / sizeof(sequence_result_type);
-    alignas(16) sequence_result_type buffer[kBufferSize];
-
-    // Randen::Absorb XORs the seed into state, which is then mixed by a call
-    // to Randen::Generate. Seeding with only the provided entropy is preferred
-    // to using an arbitrary generate() call, so use [rand.req.seed_seq]
-    // size as a proxy for the number of entropy units that can be generated
-    // without relying on seed sequence mixing...
-    const size_t entropy_size = seq.size();
-    if (entropy_size < kBufferSize) {
-      // ... and only request that many values, or 256-bits, when unspecified.
-      const size_t requested_entropy = (entropy_size == 0) ? 8u : entropy_size;
-      std::fill(buffer + requested_entropy, buffer + kBufferSize, 0);
-      seq.generate(buffer, buffer + requested_entropy);
-#ifdef ABSL_IS_BIG_ENDIAN
-      // Randen expects the seed buffer to be in Little Endian; reverse it on
-      // Big Endian platforms.
-      for (sequence_result_type& e : buffer) {
-        e = absl::little_endian::FromHost(e);
-      }
-#endif
-      // The Randen paper suggests preferentially initializing even-numbered
-      // 128-bit vectors of the randen state (there are 16 such vectors).
-      // The seed data is merged into the state offset by 128-bits, which
-      // implies preferring seed bytes [16..31, ..., 208..223]. Since the
-      // buffer is 32-bit values, we swap the corresponding buffer positions in
-      // 128-bit chunks.
-      size_t dst = kBufferSize;
-      while (dst > 7) {
-        // leave the odd bucket as-is.
-        dst -= 4;
-        size_t src = dst >> 1;
-        // swap 128-bits into the even bucket
-        std::swap(buffer[--dst], buffer[--src]);
-        std::swap(buffer[--dst], buffer[--src]);
-        std::swap(buffer[--dst], buffer[--src]);
-        std::swap(buffer[--dst], buffer[--src]);
-      }
-    } else {
-      seq.generate(buffer, buffer + kBufferSize);
-    }
-    impl_.Absorb(buffer, state());
-
-    // Generate will be called when operator() is called
-    next_ = kStateSizeT;
-  }
-
-  void discard(uint64_t count) {
-    uint64_t step = std::min<uint64_t>(kStateSizeT - next_, count);
-    count -= step;
-
-    constexpr uint64_t kRateT = kStateSizeT - kCapacityT;
-    auto* begin = state();
-    while (count > 0) {
-      next_ = kCapacityT;
-      impl_.Generate(*reinterpret_cast<result_type(*)[kStateSizeT]>(begin));
-      step = std::min<uint64_t>(kRateT, count);
-      count -= step;
-    }
-    next_ += step;
-  }
-
-  bool operator==(const randen_engine& other) const {
-    const auto* begin = state();
-    return next_ == other.next_ &&
-           std::equal(begin, begin + kStateSizeT, other.state());
-  }
-
-  bool operator!=(const randen_engine& other) const {
-    return !(*this == other);
-  }
-
-  template <class CharT, class Traits>
-  friend std::basic_ostream<CharT, Traits>& operator<<(
-      std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
-      const randen_engine<T>& engine) {       // NOLINT(runtime/references)
-    using numeric_type =
-        typename random_internal::stream_format_type<result_type>::type;
-    auto saver = random_internal::make_ostream_state_saver(os);
-    auto* it = engine.state();
-    for (auto* end = it + kStateSizeT; it < end; ++it) {
-      // In the case that `elem` is `uint8_t`, it must be cast to something
-      // larger so that it prints as an integer rather than a character. For
-      // simplicity, apply the cast all circumstances.
-      os << static_cast<numeric_type>(little_endian::FromHost(*it))
-         << os.fill();
-    }
-    os << engine.next_;
-    return os;
-  }
-
-  template <class CharT, class Traits>
-  friend std::basic_istream<CharT, Traits>& operator>>(
-      std::basic_istream<CharT, Traits>& is,  // NOLINT(runtime/references)
-      randen_engine<T>& engine) {             // NOLINT(runtime/references)
-    using numeric_type =
-        typename random_internal::stream_format_type<result_type>::type;
-    result_type state[kStateSizeT];
-    size_t next;
-    for (auto& elem : state) {
-      // It is not possible to read uint8_t from wide streams, so it is
-      // necessary to read a wider type and then cast it to uint8_t.
-      numeric_type value;
-      is >> value;
-      elem = little_endian::ToHost(static_cast<result_type>(value));
-    }
-    is >> next;
-    if (is.fail()) {
-      return is;
-    }
-    std::memcpy(engine.state(), state, sizeof(state));
-    engine.next_ = next;
-    return is;
-  }
-
- private:
-  static constexpr size_t kStateSizeT =
-      Randen::kStateBytes / sizeof(result_type);
-  static constexpr size_t kCapacityT =
-      Randen::kCapacityBytes / sizeof(result_type);
-
-  // Returns the state array pointer, which is aligned to 16 bytes.
-  // The first kCapacityT are the `inner' sponge; the remainder are available.
-  result_type* state() {
-    return reinterpret_cast<result_type*>(
-        (reinterpret_cast<uintptr_t>(&raw_state_) & 0xf) ? (raw_state_ + 8)
-                                                         : raw_state_);
-  }
-  const result_type* state() const {
-    return const_cast<randen_engine*>(this)->state();
-  }
-
-  // raw state array, manually aligned in state(). This overallocates
-  // by 8 bytes since C++ does not guarantee extended heap alignment.
-  alignas(8) char raw_state_[Randen::kStateBytes + 8];
-  size_t next_;  // index within state()
-  Randen impl_;
-};
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_RANDEN_ENGINE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_hwaes.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_hwaes.h
deleted file mode 100644
index 71a7f69f..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_hwaes.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_RANDEN_HWAES_H_
-#define ABSL_RANDOM_INTERNAL_RANDEN_HWAES_H_
-
-#include "absl/base/config.h"
-
-// HERMETIC NOTE: The randen_hwaes target must not introduce duplicate
-// symbols from arbitrary system and other headers, since it may be built
-// with different flags from other targets, using different levels of
-// optimization, potentially introducing ODR violations.
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// RANDen = RANDom generator or beetroots in Swiss High German.
-// 'Strong' (well-distributed, unpredictable, backtracking-resistant) random
-// generator, faster in some benchmarks than std::mt19937_64 and pcg64_c32.
-//
-// RandenHwAes implements the basic state manipulation methods.
-class RandenHwAes {
- public:
-  static void Generate(const void* keys, void* state_void);
-  static void Absorb(const void* seed_void, void* state_void);
-  static const void* GetKeys();
-};
-
-// HasRandenHwAesImplementation returns true when there is an accelerated
-// implementation, and false otherwise.  If there is no implementation,
-// then attempting to use it will abort the program.
-bool HasRandenHwAesImplementation();
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_RANDEN_HWAES_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_slow.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_slow.h
deleted file mode 100644
index 532c3a89..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_slow.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_RANDEN_SLOW_H_
-#define ABSL_RANDOM_INTERNAL_RANDEN_SLOW_H_
-
-#include <cstddef>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// RANDen = RANDom generator or beetroots in Swiss High German.
-// RandenSlow implements the basic state manipulation methods for
-// architectures lacking AES hardware acceleration intrinsics.
-class RandenSlow {
- public:
-  static void Generate(const void* keys, void* state_void);
-  static void Absorb(const void* seed_void, void* state_void);
-  static const void* GetKeys();
-};
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_RANDEN_SLOW_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_traits.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_traits.h
deleted file mode 100644
index 120022c9..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/randen_traits.h
+++ /dev/null
@@ -1,88 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_RANDEN_TRAITS_H_
-#define ABSL_RANDOM_INTERNAL_RANDEN_TRAITS_H_
-
-// HERMETIC NOTE: The randen_hwaes target must not introduce duplicate
-// symbols from arbitrary system and other headers, since it may be built
-// with different flags from other targets, using different levels of
-// optimization, potentially introducing ODR violations.
-
-#include <cstddef>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// RANDen = RANDom generator or beetroots in Swiss High German.
-// 'Strong' (well-distributed, unpredictable, backtracking-resistant) random
-// generator, faster in some benchmarks than std::mt19937_64 and pcg64_c32.
-//
-// High-level summary:
-// 1) Reverie (see "A Robust and Sponge-Like PRNG with Improved Efficiency") is
-//    a sponge-like random generator that requires a cryptographic permutation.
-//    It improves upon "Provably Robust Sponge-Based PRNGs and KDFs" by
-//    achieving backtracking resistance with only one Permute() per buffer.
-//
-// 2) "Simpira v2: A Family of Efficient Permutations Using the AES Round
-//    Function" constructs up to 1024-bit permutations using an improved
-//    Generalized Feistel network with 2-round AES-128 functions. This Feistel
-//    block shuffle achieves diffusion faster and is less vulnerable to
-//    sliced-biclique attacks than the Type-2 cyclic shuffle.
-//
-// 3) "Improving the Generalized Feistel" and "New criterion for diffusion
-//    property" extends the same kind of improved Feistel block shuffle to 16
-//    branches, which enables a 2048-bit permutation.
-//
-// Combine these three ideas and also change Simpira's subround keys from
-// structured/low-entropy counters to digits of Pi (or other random source).
-
-// RandenTraits contains the basic algorithm traits, such as the size of the
-// state, seed, sponge, etc.
-struct RandenTraits {
-  // Size of the entire sponge / state for the randen PRNG.
-  static constexpr size_t kStateBytes = 256;  // 2048-bit
-
-  // Size of the 'inner' (inaccessible) part of the sponge. Larger values would
-  // require more frequent calls to RandenGenerate.
-  static constexpr size_t kCapacityBytes = 16;  // 128-bit
-
-  // Size of the default seed consumed by the sponge.
-  static constexpr size_t kSeedBytes = kStateBytes - kCapacityBytes;
-
-  // Assuming 128-bit blocks, the number of blocks in the state.
-  // Largest size for which security proofs are known.
-  static constexpr size_t kFeistelBlocks = 16;
-
-  // Ensures SPRP security and two full subblock diffusions.
-  // Must be > 4 * log2(kFeistelBlocks).
-  static constexpr size_t kFeistelRounds = 16 + 1;
-
-  // Size of the key. A 128-bit key block is used for every-other
-  // feistel block (Type-2 generalized Feistel network) in each round.
-  static constexpr size_t kKeyBytes = 16 * kFeistelRounds * kFeistelBlocks / 2;
-};
-
-// Randen key arrays. In randen_round_keys.cc
-extern const unsigned char kRandenRoundKeys[RandenTraits::kKeyBytes];
-extern const unsigned char kRandenRoundKeysBE[RandenTraits::kKeyBytes];
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_RANDEN_TRAITS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/salted_seed_seq.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/salted_seed_seq.h
deleted file mode 100644
index 06291865..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/salted_seed_seq.h
+++ /dev/null
@@ -1,165 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_SALTED_SEED_SEQ_H_
-#define ABSL_RANDOM_INTERNAL_SALTED_SEED_SEQ_H_
-
-#include <cstdint>
-#include <cstdlib>
-#include <initializer_list>
-#include <iterator>
-#include <memory>
-#include <type_traits>
-#include <utility>
-#include <vector>
-
-#include "absl/container/inlined_vector.h"
-#include "absl/meta/type_traits.h"
-#include "absl/random/internal/seed_material.h"
-#include "absl/types/optional.h"
-#include "absl/types/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// This class conforms to the C++ Standard "Seed Sequence" concept
-// [rand.req.seedseq].
-//
-// A `SaltedSeedSeq` is meant to wrap an existing seed sequence and modify
-// generated sequence by mixing with extra entropy. This entropy may be
-// build-dependent or process-dependent. The implementation may change to be
-// have either or both kinds of entropy. If salt is not available sequence is
-// not modified.
-template <typename SSeq>
-class SaltedSeedSeq {
- public:
-  using inner_sequence_type = SSeq;
-  using result_type = typename SSeq::result_type;
-
-  SaltedSeedSeq() : seq_(absl::make_unique<SSeq>()) {}
-
-  template <typename Iterator>
-  SaltedSeedSeq(Iterator begin, Iterator end)
-      : seq_(absl::make_unique<SSeq>(begin, end)) {}
-
-  template <typename T>
-  SaltedSeedSeq(std::initializer_list<T> il)
-      : SaltedSeedSeq(il.begin(), il.end()) {}
-
-  SaltedSeedSeq(const SaltedSeedSeq&) = delete;
-  SaltedSeedSeq& operator=(const SaltedSeedSeq&) = delete;
-
-  SaltedSeedSeq(SaltedSeedSeq&&) = default;
-  SaltedSeedSeq& operator=(SaltedSeedSeq&&) = default;
-
-  template <typename RandomAccessIterator>
-  void generate(RandomAccessIterator begin, RandomAccessIterator end) {
-    using U = typename std::iterator_traits<RandomAccessIterator>::value_type;
-
-    // The common case is that generate is called with ContiguousIterators
-    // to uint arrays. Such contiguous memory regions may be optimized,
-    // which we detect here.
-    using TagType = absl::conditional_t<
-        (std::is_same<U, uint32_t>::value &&
-         (std::is_pointer<RandomAccessIterator>::value ||
-          std::is_same<RandomAccessIterator,
-                       typename std::vector<U>::iterator>::value)),
-        ContiguousAndUint32Tag, DefaultTag>;
-    if (begin != end) {
-      generate_impl(TagType{}, begin, end, std::distance(begin, end));
-    }
-  }
-
-  template <typename OutIterator>
-  void param(OutIterator out) const {
-    seq_->param(out);
-  }
-
-  size_t size() const { return seq_->size(); }
-
- private:
-  struct ContiguousAndUint32Tag {};
-  struct DefaultTag {};
-
-  // Generate which requires the iterators are contiguous pointers to uint32_t.
-  // Fills the initial seed buffer the underlying SSeq::generate() call,
-  // then mixes in the salt material.
-  template <typename Contiguous>
-  void generate_impl(ContiguousAndUint32Tag, Contiguous begin, Contiguous end,
-                     size_t n) {
-    seq_->generate(begin, end);
-    const uint32_t salt = absl::random_internal::GetSaltMaterial().value_or(0);
-    auto span = absl::Span<uint32_t>(&*begin, n);
-    MixIntoSeedMaterial(absl::MakeConstSpan(&salt, 1), span);
-  }
-
-  // The uncommon case for generate is that it is called with iterators over
-  // some other buffer type which is assignable from a 32-bit value. In this
-  // case we allocate a temporary 32-bit buffer and then copy-assign back
-  // to the initial inputs.
-  template <typename RandomAccessIterator>
-  void generate_impl(DefaultTag, RandomAccessIterator begin,
-                     RandomAccessIterator, size_t n) {
-    // Allocates a seed buffer of `n` elements, generates the seed, then
-    // copies the result into the `out` iterator.
-    absl::InlinedVector<uint32_t, 8> data(n, 0);
-    generate_impl(ContiguousAndUint32Tag{}, data.begin(), data.end(), n);
-    std::copy(data.begin(), data.end(), begin);
-  }
-
-  // Because [rand.req.seedseq] is not required to be copy-constructible,
-  // copy-assignable nor movable, we wrap it with unique pointer to be able
-  // to move SaltedSeedSeq.
-  std::unique_ptr<SSeq> seq_;
-};
-
-// is_salted_seed_seq indicates whether the type is a SaltedSeedSeq.
-template <typename T, typename = void>
-struct is_salted_seed_seq : public std::false_type {};
-
-template <typename T>
-struct is_salted_seed_seq<
-    T, typename std::enable_if<std::is_same<
-           T, SaltedSeedSeq<typename T::inner_sequence_type>>::value>::type>
-    : public std::true_type {};
-
-// MakeSaltedSeedSeq returns a salted variant of the seed sequence.
-// When provided with an existing SaltedSeedSeq, returns the input parameter,
-// otherwise constructs a new SaltedSeedSeq which embodies the original
-// non-salted seed parameters.
-template <
-    typename SSeq,  //
-    typename EnableIf = absl::enable_if_t<is_salted_seed_seq<SSeq>::value>>
-SSeq MakeSaltedSeedSeq(SSeq&& seq) {
-  return SSeq(std::forward<SSeq>(seq));
-}
-
-template <
-    typename SSeq,  //
-    typename EnableIf = absl::enable_if_t<!is_salted_seed_seq<SSeq>::value>>
-SaltedSeedSeq<typename std::decay<SSeq>::type> MakeSaltedSeedSeq(SSeq&& seq) {
-  using sseq_type = typename std::decay<SSeq>::type;
-  using result_type = typename sseq_type::result_type;
-
-  absl::InlinedVector<result_type, 8> data;
-  seq.param(std::back_inserter(data));
-  return SaltedSeedSeq<sseq_type>(data.begin(), data.end());
-}
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_SALTED_SEED_SEQ_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/seed_material.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/seed_material.h
deleted file mode 100644
index 4be10e92..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/seed_material.h
+++ /dev/null
@@ -1,104 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_SEED_MATERIAL_H_
-#define ABSL_RANDOM_INTERNAL_SEED_MATERIAL_H_
-
-#include <cassert>
-#include <cstdint>
-#include <cstdlib>
-#include <string>
-#include <vector>
-
-#include "absl/base/attributes.h"
-#include "absl/random/internal/fast_uniform_bits.h"
-#include "absl/types/optional.h"
-#include "absl/types/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// Returns the number of 32-bit blocks needed to contain the given number of
-// bits.
-constexpr size_t SeedBitsToBlocks(size_t seed_size) {
-  return (seed_size + 31) / 32;
-}
-
-// Amount of entropy (measured in bits) used to instantiate a Seed Sequence,
-// with which to create a URBG.
-constexpr size_t kEntropyBitsNeeded = 256;
-
-// Amount of entropy (measured in 32-bit blocks) used to instantiate a Seed
-// Sequence, with which to create a URBG.
-constexpr size_t kEntropyBlocksNeeded =
-    random_internal::SeedBitsToBlocks(kEntropyBitsNeeded);
-
-static_assert(kEntropyBlocksNeeded > 0,
-              "Entropy used to seed URBGs must be nonzero.");
-
-// Attempts to fill a span of uint32_t-values using an OS-provided source of
-// true entropy (eg. /dev/urandom) into an array of uint32_t blocks of data. The
-// resulting array may be used to initialize an instance of a class conforming
-// to the C++ Standard "Seed Sequence" concept [rand.req.seedseq].
-//
-// If values.data() == nullptr, the behavior is undefined.
-ABSL_MUST_USE_RESULT
-bool ReadSeedMaterialFromOSEntropy(absl::Span<uint32_t> values);
-
-// Attempts to fill a span of uint32_t-values using variates generated by an
-// existing instance of a class conforming to the C++ Standard "Uniform Random
-// Bit Generator" concept [rand.req.urng]. The resulting data may be used to
-// initialize an instance of a class conforming to the C++ Standard
-// "Seed Sequence" concept [rand.req.seedseq].
-//
-// If urbg == nullptr or values.data() == nullptr, the behavior is undefined.
-template <typename URBG>
-ABSL_MUST_USE_RESULT bool ReadSeedMaterialFromURBG(
-    URBG* urbg, absl::Span<uint32_t> values) {
-  random_internal::FastUniformBits<uint32_t> distr;
-
-  assert(urbg != nullptr && values.data() != nullptr);
-  if (urbg == nullptr || values.data() == nullptr) {
-    return false;
-  }
-
-  for (uint32_t& seed_value : values) {
-    seed_value = distr(*urbg);
-  }
-  return true;
-}
-
-// Mixes given sequence of values with into given sequence of seed material.
-// Time complexity of this function is O(sequence.size() *
-// seed_material.size()).
-//
-// Algorithm is based on code available at
-// https://gist.github.com/imneme/540829265469e673d045
-// by Melissa O'Neill.
-void MixIntoSeedMaterial(absl::Span<const uint32_t> sequence,
-                         absl::Span<uint32_t> seed_material);
-
-// Returns salt value.
-//
-// Salt is obtained only once and stored in static variable.
-//
-// May return empty value if optaining the salt was not possible.
-absl::optional<uint32_t> GetSaltMaterial();
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_SEED_MATERIAL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/sequence_urbg.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/sequence_urbg.h
deleted file mode 100644
index bc96a12c..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/sequence_urbg.h
+++ /dev/null
@@ -1,60 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_SEQUENCE_URBG_H_
-#define ABSL_RANDOM_INTERNAL_SEQUENCE_URBG_H_
-
-#include <cstdint>
-#include <cstring>
-#include <limits>
-#include <type_traits>
-#include <vector>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// `sequence_urbg` is a simple random number generator which meets the
-// requirements of [rand.req.urbg], and is solely for testing absl
-// distributions.
-class sequence_urbg {
- public:
-  using result_type = uint64_t;
-
-  static constexpr result_type(min)() {
-    return (std::numeric_limits<result_type>::min)();
-  }
-  static constexpr result_type(max)() {
-    return (std::numeric_limits<result_type>::max)();
-  }
-
-  sequence_urbg(std::initializer_list<result_type> data) : i_(0), data_(data) {}
-  void reset() { i_ = 0; }
-
-  result_type operator()() { return data_[i_++ % data_.size()]; }
-
-  size_t invocations() const { return i_; }
-
- private:
-  size_t i_;
-  std::vector<result_type> data_;
-};
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_SEQUENCE_URBG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/traits.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/traits.h
deleted file mode 100644
index f874a0f7..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/traits.h
+++ /dev/null
@@ -1,149 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_TRAITS_H_
-#define ABSL_RANDOM_INTERNAL_TRAITS_H_
-
-#include <cstdint>
-#include <limits>
-#include <type_traits>
-
-#include "absl/base/config.h"
-#include "absl/numeric/bits.h"
-#include "absl/numeric/int128.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// random_internal::is_widening_convertible<A, B>
-//
-// Returns whether a type A is widening-convertible to a type B.
-//
-// A is widening-convertible to B means:
-//   A a = <any number>;
-//   B b = a;
-//   A c = b;
-//   EXPECT_EQ(a, c);
-template <typename A, typename B>
-class is_widening_convertible {
-  // As long as there are enough bits in the exact part of a number:
-  // - unsigned can fit in float, signed, unsigned
-  // - signed can fit in float, signed
-  // - float can fit in float
-  // So we define rank to be:
-  // - rank(float) -> 2
-  // - rank(signed) -> 1
-  // - rank(unsigned) -> 0
-  template <class T>
-  static constexpr int rank() {
-    return !std::numeric_limits<T>::is_integer +
-           std::numeric_limits<T>::is_signed;
-  }
-
- public:
-  // If an arithmetic-type B can represent at least as many digits as a type A,
-  // and B belongs to a rank no lower than A, then A can be safely represented
-  // by B through a widening-conversion.
-  static constexpr bool value =
-      std::numeric_limits<A>::digits <= std::numeric_limits<B>::digits &&
-      rank<A>() <= rank<B>();
-};
-
-template <typename T>
-struct IsIntegral : std::is_integral<T> {};
-template <>
-struct IsIntegral<absl::int128> : std::true_type {};
-template <>
-struct IsIntegral<absl::uint128> : std::true_type {};
-
-template <typename T>
-struct MakeUnsigned : std::make_unsigned<T> {};
-template <>
-struct MakeUnsigned<absl::int128> {
-  using type = absl::uint128;
-};
-template <>
-struct MakeUnsigned<absl::uint128> {
-  using type = absl::uint128;
-};
-
-template <typename T>
-struct IsUnsigned : std::is_unsigned<T> {};
-template <>
-struct IsUnsigned<absl::int128> : std::false_type {};
-template <>
-struct IsUnsigned<absl::uint128> : std::true_type {};
-
-// unsigned_bits<N>::type returns the unsigned int type with the indicated
-// number of bits.
-template <size_t N>
-struct unsigned_bits;
-
-template <>
-struct unsigned_bits<8> {
-  using type = uint8_t;
-};
-template <>
-struct unsigned_bits<16> {
-  using type = uint16_t;
-};
-template <>
-struct unsigned_bits<32> {
-  using type = uint32_t;
-};
-template <>
-struct unsigned_bits<64> {
-  using type = uint64_t;
-};
-
-template <>
-struct unsigned_bits<128> {
-  using type = absl::uint128;
-};
-
-// 256-bit wrapper for wide multiplications.
-struct U256 {
-  uint128 hi;
-  uint128 lo;
-};
-template <>
-struct unsigned_bits<256> {
-  using type = U256;
-};
-
-template <typename IntType>
-struct make_unsigned_bits {
-  using type = typename unsigned_bits<
-      std::numeric_limits<typename MakeUnsigned<IntType>::type>::digits>::type;
-};
-
-template <typename T>
-int BitWidth(T v) {
-  // Workaround for bit_width not supporting int128.
-  // Don't hardcode `64` to make sure this code does not trigger compiler
-  // warnings in smaller types.
-  constexpr int half_bits = sizeof(T) * 8 / 2;
-  if (sizeof(T) == 16 && (v >> half_bits) != 0) {
-    return bit_width(static_cast<uint64_t>(v >> half_bits)) + half_bits;
-  } else {
-    return bit_width(static_cast<uint64_t>(v));
-  }
-}
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_TRAITS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/uniform_helper.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/uniform_helper.h
deleted file mode 100644
index db737e13..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/uniform_helper.h
+++ /dev/null
@@ -1,244 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-#ifndef ABSL_RANDOM_INTERNAL_UNIFORM_HELPER_H_
-#define ABSL_RANDOM_INTERNAL_UNIFORM_HELPER_H_
-
-#include <cmath>
-#include <limits>
-#include <type_traits>
-
-#include "absl/base/config.h"
-#include "absl/meta/type_traits.h"
-#include "absl/random/internal/traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-template <typename IntType>
-class uniform_int_distribution;
-
-template <typename RealType>
-class uniform_real_distribution;
-
-// Interval tag types which specify whether the interval is open or closed
-// on either boundary.
-
-namespace random_internal {
-template <typename T>
-struct TagTypeCompare {};
-
-template <typename T>
-constexpr bool operator==(TagTypeCompare<T>, TagTypeCompare<T>) {
-  // Tags are mono-states. They always compare equal.
-  return true;
-}
-template <typename T>
-constexpr bool operator!=(TagTypeCompare<T>, TagTypeCompare<T>) {
-  return false;
-}
-
-}  // namespace random_internal
-
-struct IntervalClosedClosedTag
-    : public random_internal::TagTypeCompare<IntervalClosedClosedTag> {};
-struct IntervalClosedOpenTag
-    : public random_internal::TagTypeCompare<IntervalClosedOpenTag> {};
-struct IntervalOpenClosedTag
-    : public random_internal::TagTypeCompare<IntervalOpenClosedTag> {};
-struct IntervalOpenOpenTag
-    : public random_internal::TagTypeCompare<IntervalOpenOpenTag> {};
-
-namespace random_internal {
-
-// In the absence of an explicitly provided return-type, the template
-// "uniform_inferred_return_t<A, B>" is used to derive a suitable type, based on
-// the data-types of the endpoint-arguments {A lo, B hi}.
-//
-// Given endpoints {A lo, B hi}, one of {A, B} will be chosen as the
-// return-type, if one type can be implicitly converted into the other, in a
-// lossless way. The template "is_widening_convertible" implements the
-// compile-time logic for deciding if such a conversion is possible.
-//
-// If no such conversion between {A, B} exists, then the overload for
-// absl::Uniform() will be discarded, and the call will be ill-formed.
-// Return-type for absl::Uniform() when the return-type is inferred.
-template <typename A, typename B>
-using uniform_inferred_return_t =
-    absl::enable_if_t<absl::disjunction<is_widening_convertible<A, B>,
-                                        is_widening_convertible<B, A>>::value,
-                      typename std::conditional<
-                          is_widening_convertible<A, B>::value, B, A>::type>;
-
-// The functions
-//    uniform_lower_bound(tag, a, b)
-// and
-//    uniform_upper_bound(tag, a, b)
-// are used as implementation-details for absl::Uniform().
-//
-// Conceptually,
-//    [a, b] == [uniform_lower_bound(IntervalClosedClosed, a, b),
-//               uniform_upper_bound(IntervalClosedClosed, a, b)]
-//    (a, b) == [uniform_lower_bound(IntervalOpenOpen, a, b),
-//               uniform_upper_bound(IntervalOpenOpen, a, b)]
-//    [a, b) == [uniform_lower_bound(IntervalClosedOpen, a, b),
-//               uniform_upper_bound(IntervalClosedOpen, a, b)]
-//    (a, b] == [uniform_lower_bound(IntervalOpenClosed, a, b),
-//               uniform_upper_bound(IntervalOpenClosed, a, b)]
-//
-template <typename IntType, typename Tag>
-typename absl::enable_if_t<
-    absl::conjunction<
-        IsIntegral<IntType>,
-        absl::disjunction<std::is_same<Tag, IntervalOpenClosedTag>,
-                          std::is_same<Tag, IntervalOpenOpenTag>>>::value,
-    IntType>
-uniform_lower_bound(Tag, IntType a, IntType) {
-  return a < (std::numeric_limits<IntType>::max)() ? (a + 1) : a;
-}
-
-template <typename FloatType, typename Tag>
-typename absl::enable_if_t<
-    absl::conjunction<
-        std::is_floating_point<FloatType>,
-        absl::disjunction<std::is_same<Tag, IntervalOpenClosedTag>,
-                          std::is_same<Tag, IntervalOpenOpenTag>>>::value,
-    FloatType>
-uniform_lower_bound(Tag, FloatType a, FloatType b) {
-  return std::nextafter(a, b);
-}
-
-template <typename NumType, typename Tag>
-typename absl::enable_if_t<
-    absl::disjunction<std::is_same<Tag, IntervalClosedClosedTag>,
-                      std::is_same<Tag, IntervalClosedOpenTag>>::value,
-    NumType>
-uniform_lower_bound(Tag, NumType a, NumType) {
-  return a;
-}
-
-template <typename IntType, typename Tag>
-typename absl::enable_if_t<
-    absl::conjunction<
-        IsIntegral<IntType>,
-        absl::disjunction<std::is_same<Tag, IntervalClosedOpenTag>,
-                          std::is_same<Tag, IntervalOpenOpenTag>>>::value,
-    IntType>
-uniform_upper_bound(Tag, IntType, IntType b) {
-  return b > (std::numeric_limits<IntType>::min)() ? (b - 1) : b;
-}
-
-template <typename FloatType, typename Tag>
-typename absl::enable_if_t<
-    absl::conjunction<
-        std::is_floating_point<FloatType>,
-        absl::disjunction<std::is_same<Tag, IntervalClosedOpenTag>,
-                          std::is_same<Tag, IntervalOpenOpenTag>>>::value,
-    FloatType>
-uniform_upper_bound(Tag, FloatType, FloatType b) {
-  return b;
-}
-
-template <typename IntType, typename Tag>
-typename absl::enable_if_t<
-    absl::conjunction<
-        IsIntegral<IntType>,
-        absl::disjunction<std::is_same<Tag, IntervalClosedClosedTag>,
-                          std::is_same<Tag, IntervalOpenClosedTag>>>::value,
-    IntType>
-uniform_upper_bound(Tag, IntType, IntType b) {
-  return b;
-}
-
-template <typename FloatType, typename Tag>
-typename absl::enable_if_t<
-    absl::conjunction<
-        std::is_floating_point<FloatType>,
-        absl::disjunction<std::is_same<Tag, IntervalClosedClosedTag>,
-                          std::is_same<Tag, IntervalOpenClosedTag>>>::value,
-    FloatType>
-uniform_upper_bound(Tag, FloatType, FloatType b) {
-  return std::nextafter(b, (std::numeric_limits<FloatType>::max)());
-}
-
-// Returns whether the bounds are valid for the underlying distribution.
-// Inputs must have already been resolved via uniform_*_bound calls.
-//
-// The c++ standard constraints in [rand.dist.uni.int] are listed as:
-//    requires: lo <= hi.
-//
-// In the uniform_int_distrubtion, {lo, hi} are closed, closed. Thus:
-// [0, 0] is legal.
-// [0, 0) is not legal, but [0, 1) is, which translates to [0, 0].
-// (0, 1) is not legal, but (0, 2) is, which translates to [1, 1].
-// (0, 0] is not legal, but (0, 1] is, which translates to [1, 1].
-//
-// The c++ standard constraints in [rand.dist.uni.real] are listed as:
-//    requires: lo <= hi.
-//    requires: (hi - lo) <= numeric_limits<T>::max()
-//
-// In the uniform_real_distribution, {lo, hi} are closed, open, Thus:
-// [0, 0] is legal, which is [0, 0+epsilon).
-// [0, 0) is legal.
-// (0, 0) is not legal, but (0-epsilon, 0+epsilon) is.
-// (0, 0] is not legal, but (0, 0+epsilon] is.
-//
-template <typename FloatType>
-absl::enable_if_t<std::is_floating_point<FloatType>::value, bool>
-is_uniform_range_valid(FloatType a, FloatType b) {
-  return a <= b && std::isfinite(b - a);
-}
-
-template <typename IntType>
-absl::enable_if_t<IsIntegral<IntType>::value, bool>
-is_uniform_range_valid(IntType a, IntType b) {
-  return a <= b;
-}
-
-// UniformDistribution selects either absl::uniform_int_distribution
-// or absl::uniform_real_distribution depending on the NumType parameter.
-template <typename NumType>
-using UniformDistribution =
-    typename std::conditional<IsIntegral<NumType>::value,
-                              absl::uniform_int_distribution<NumType>,
-                              absl::uniform_real_distribution<NumType>>::type;
-
-// UniformDistributionWrapper is used as the underlying distribution type
-// by the absl::Uniform template function. It selects the proper Abseil
-// uniform distribution and provides constructor overloads that match the
-// expected parameter order as well as adjusting distribution bounds based
-// on the tag.
-template <typename NumType>
-struct UniformDistributionWrapper : public UniformDistribution<NumType> {
-  template <typename TagType>
-  explicit UniformDistributionWrapper(TagType, NumType lo, NumType hi)
-      : UniformDistribution<NumType>(
-            uniform_lower_bound<NumType>(TagType{}, lo, hi),
-            uniform_upper_bound<NumType>(TagType{}, lo, hi)) {}
-
-  explicit UniformDistributionWrapper(NumType lo, NumType hi)
-      : UniformDistribution<NumType>(
-            uniform_lower_bound<NumType>(IntervalClosedOpenTag(), lo, hi),
-            uniform_upper_bound<NumType>(IntervalClosedOpenTag(), lo, hi)) {}
-
-  explicit UniformDistributionWrapper()
-      : UniformDistribution<NumType>(std::numeric_limits<NumType>::lowest(),
-                                     (std::numeric_limits<NumType>::max)()) {}
-};
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_UNIFORM_HELPER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/internal/wide_multiply.h b/projects/llm_framework/include/abseil-cpp/absl/random/internal/wide_multiply.h
deleted file mode 100644
index 891e3630..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/internal/wide_multiply.h
+++ /dev/null
@@ -1,96 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_WIDE_MULTIPLY_H_
-#define ABSL_RANDOM_INTERNAL_WIDE_MULTIPLY_H_
-
-#include <cstdint>
-#include <limits>
-#include <type_traits>
-
-#if (defined(_WIN32) || defined(_WIN64)) && defined(_M_IA64)
-#include <intrin.h>  // NOLINT(build/include_order)
-#pragma intrinsic(_umul128)
-#define ABSL_INTERNAL_USE_UMUL128 1
-#endif
-
-#include "absl/base/config.h"
-#include "absl/numeric/bits.h"
-#include "absl/numeric/int128.h"
-#include "absl/random/internal/traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace random_internal {
-
-// wide_multiply<T> multiplies two N-bit values to a 2N-bit result.
-template <typename UIntType>
-struct wide_multiply {
-  static constexpr size_t kN = std::numeric_limits<UIntType>::digits;
-  using input_type = UIntType;
-  using result_type = typename random_internal::unsigned_bits<kN * 2>::type;
-
-  static result_type multiply(input_type a, input_type b) {
-    return static_cast<result_type>(a) * b;
-  }
-
-  static input_type hi(result_type r) {
-    return static_cast<input_type>(r >> kN);
-  }
-  static input_type lo(result_type r) { return static_cast<input_type>(r); }
-
-  static_assert(std::is_unsigned<UIntType>::value,
-                "Class-template wide_multiply<> argument must be unsigned.");
-};
-
-// MultiplyU128ToU256 multiplies two 128-bit values to a 256-bit value.
-inline U256 MultiplyU128ToU256(uint128 a, uint128 b) {
-  const uint128 a00 = static_cast<uint64_t>(a);
-  const uint128 a64 = a >> 64;
-  const uint128 b00 = static_cast<uint64_t>(b);
-  const uint128 b64 = b >> 64;
-
-  const uint128 c00 = a00 * b00;
-  const uint128 c64a = a00 * b64;
-  const uint128 c64b = a64 * b00;
-  const uint128 c128 = a64 * b64;
-
-  const uint64_t carry =
-      static_cast<uint64_t>(((c00 >> 64) + static_cast<uint64_t>(c64a) +
-                             static_cast<uint64_t>(c64b)) >>
-                            64);
-
-  return {c128 + (c64a >> 64) + (c64b >> 64) + carry,
-          c00 + (c64a << 64) + (c64b << 64)};
-}
-
-
-template <>
-struct wide_multiply<uint128> {
-  using input_type = uint128;
-  using result_type = U256;
-
-  static result_type multiply(input_type a, input_type b) {
-    return MultiplyU128ToU256(a, b);
-  }
-
-  static input_type hi(result_type r) { return r.hi; }
-  static input_type lo(result_type r) { return r.lo; }
-};
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_WIDE_MULTIPLY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/log_uniform_int_distribution.h b/projects/llm_framework/include/abseil-cpp/absl/random/log_uniform_int_distribution.h
deleted file mode 100644
index 4afff8f6..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/log_uniform_int_distribution.h
+++ /dev/null
@@ -1,256 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_LOG_UNIFORM_INT_DISTRIBUTION_H_
-#define ABSL_RANDOM_LOG_UNIFORM_INT_DISTRIBUTION_H_
-
-#include <algorithm>
-#include <cassert>
-#include <cmath>
-#include <istream>
-#include <limits>
-#include <ostream>
-#include <type_traits>
-
-#include "absl/numeric/bits.h"
-#include "absl/random/internal/fastmath.h"
-#include "absl/random/internal/generate_real.h"
-#include "absl/random/internal/iostream_state_saver.h"
-#include "absl/random/internal/traits.h"
-#include "absl/random/uniform_int_distribution.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// log_uniform_int_distribution:
-//
-// Returns a random variate R in range [min, max] such that
-// floor(log(R-min, base)) is uniformly distributed.
-// We ensure uniformity by discretization using the
-// boundary sets [0, 1, base, base * base, ... min(base*n, max)]
-//
-template <typename IntType = int>
-class log_uniform_int_distribution {
- private:
-  using unsigned_type =
-      typename random_internal::make_unsigned_bits<IntType>::type;
-
- public:
-  using result_type = IntType;
-
-  class param_type {
-   public:
-    using distribution_type = log_uniform_int_distribution;
-
-    explicit param_type(
-        result_type min = 0,
-        result_type max = (std::numeric_limits<result_type>::max)(),
-        result_type base = 2)
-        : min_(min),
-          max_(max),
-          base_(base),
-          range_(static_cast<unsigned_type>(max_) -
-                 static_cast<unsigned_type>(min_)),
-          log_range_(0) {
-      assert(max_ >= min_);
-      assert(base_ > 1);
-
-      if (base_ == 2) {
-        // Determine where the first set bit is on range(), giving a log2(range)
-        // value which can be used to construct bounds.
-        log_range_ = (std::min)(random_internal::BitWidth(range()),
-                                std::numeric_limits<unsigned_type>::digits);
-      } else {
-        // NOTE: Computing the logN(x) introduces error from 2 sources:
-        // 1. Conversion of int to double loses precision for values >=
-        // 2^53, which may cause some log() computations to operate on
-        // different values.
-        // 2. The error introduced by the division will cause the result
-        // to differ from the expected value.
-        //
-        // Thus a result which should equal K may equal K +/- epsilon,
-        // which can eliminate some values depending on where the bounds fall.
-        const double inv_log_base = 1.0 / std::log(static_cast<double>(base_));
-        const double log_range = std::log(static_cast<double>(range()) + 0.5);
-        log_range_ = static_cast<int>(std::ceil(inv_log_base * log_range));
-      }
-    }
-
-    result_type(min)() const { return min_; }
-    result_type(max)() const { return max_; }
-    result_type base() const { return base_; }
-
-    friend bool operator==(const param_type& a, const param_type& b) {
-      return a.min_ == b.min_ && a.max_ == b.max_ && a.base_ == b.base_;
-    }
-
-    friend bool operator!=(const param_type& a, const param_type& b) {
-      return !(a == b);
-    }
-
-   private:
-    friend class log_uniform_int_distribution;
-
-    int log_range() const { return log_range_; }
-    unsigned_type range() const { return range_; }
-
-    result_type min_;
-    result_type max_;
-    result_type base_;
-    unsigned_type range_;  // max - min
-    int log_range_;        // ceil(logN(range_))
-
-    static_assert(random_internal::IsIntegral<IntType>::value,
-                  "Class-template absl::log_uniform_int_distribution<> must be "
-                  "parameterized using an integral type.");
-  };
-
-  log_uniform_int_distribution() : log_uniform_int_distribution(0) {}
-
-  explicit log_uniform_int_distribution(
-      result_type min,
-      result_type max = (std::numeric_limits<result_type>::max)(),
-      result_type base = 2)
-      : param_(min, max, base) {}
-
-  explicit log_uniform_int_distribution(const param_type& p) : param_(p) {}
-
-  void reset() {}
-
-  // generating functions
-  template <typename URBG>
-  result_type operator()(URBG& g) {  // NOLINT(runtime/references)
-    return (*this)(g, param_);
-  }
-
-  template <typename URBG>
-  result_type operator()(URBG& g,  // NOLINT(runtime/references)
-                         const param_type& p) {
-    return static_cast<result_type>((p.min)() + Generate(g, p));
-  }
-
-  result_type(min)() const { return (param_.min)(); }
-  result_type(max)() const { return (param_.max)(); }
-  result_type base() const { return param_.base(); }
-
-  param_type param() const { return param_; }
-  void param(const param_type& p) { param_ = p; }
-
-  friend bool operator==(const log_uniform_int_distribution& a,
-                         const log_uniform_int_distribution& b) {
-    return a.param_ == b.param_;
-  }
-  friend bool operator!=(const log_uniform_int_distribution& a,
-                         const log_uniform_int_distribution& b) {
-    return a.param_ != b.param_;
-  }
-
- private:
-  // Returns a log-uniform variate in the range [0, p.range()]. The caller
-  // should add min() to shift the result to the correct range.
-  template <typename URNG>
-  unsigned_type Generate(URNG& g,  // NOLINT(runtime/references)
-                         const param_type& p);
-
-  param_type param_;
-};
-
-template <typename IntType>
-template <typename URBG>
-typename log_uniform_int_distribution<IntType>::unsigned_type
-log_uniform_int_distribution<IntType>::Generate(
-    URBG& g,  // NOLINT(runtime/references)
-    const param_type& p) {
-  // sample e over [0, log_range]. Map the results of e to this:
-  // 0 => 0
-  // 1 => [1, b-1]
-  // 2 => [b, (b^2)-1]
-  // n => [b^(n-1)..(b^n)-1]
-  const int e = absl::uniform_int_distribution<int>(0, p.log_range())(g);
-  if (e == 0) {
-    return 0;
-  }
-  const int d = e - 1;
-
-  unsigned_type base_e, top_e;
-  if (p.base() == 2) {
-    base_e = static_cast<unsigned_type>(1) << d;
-
-    top_e = (e >= std::numeric_limits<unsigned_type>::digits)
-                ? (std::numeric_limits<unsigned_type>::max)()
-                : (static_cast<unsigned_type>(1) << e) - 1;
-  } else {
-    const double r = std::pow(static_cast<double>(p.base()), d);
-    const double s = (r * static_cast<double>(p.base())) - 1.0;
-
-    base_e =
-        (r > static_cast<double>((std::numeric_limits<unsigned_type>::max)()))
-            ? (std::numeric_limits<unsigned_type>::max)()
-            : static_cast<unsigned_type>(r);
-
-    top_e =
-        (s > static_cast<double>((std::numeric_limits<unsigned_type>::max)()))
-            ? (std::numeric_limits<unsigned_type>::max)()
-            : static_cast<unsigned_type>(s);
-  }
-
-  const unsigned_type lo = (base_e >= p.range()) ? p.range() : base_e;
-  const unsigned_type hi = (top_e >= p.range()) ? p.range() : top_e;
-
-  // choose uniformly over [lo, hi]
-  return absl::uniform_int_distribution<result_type>(
-      static_cast<result_type>(lo), static_cast<result_type>(hi))(g);
-}
-
-template <typename CharT, typename Traits, typename IntType>
-std::basic_ostream<CharT, Traits>& operator<<(
-    std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
-    const log_uniform_int_distribution<IntType>& x) {
-  using stream_type =
-      typename random_internal::stream_format_type<IntType>::type;
-  auto saver = random_internal::make_ostream_state_saver(os);
-  os << static_cast<stream_type>((x.min)()) << os.fill()
-     << static_cast<stream_type>((x.max)()) << os.fill()
-     << static_cast<stream_type>(x.base());
-  return os;
-}
-
-template <typename CharT, typename Traits, typename IntType>
-std::basic_istream<CharT, Traits>& operator>>(
-    std::basic_istream<CharT, Traits>& is,       // NOLINT(runtime/references)
-    log_uniform_int_distribution<IntType>& x) {  // NOLINT(runtime/references)
-  using param_type = typename log_uniform_int_distribution<IntType>::param_type;
-  using result_type =
-      typename log_uniform_int_distribution<IntType>::result_type;
-  using stream_type =
-      typename random_internal::stream_format_type<IntType>::type;
-
-  stream_type min;
-  stream_type max;
-  stream_type base;
-
-  auto saver = random_internal::make_istream_state_saver(is);
-  is >> min >> max >> base;
-  if (!is.fail()) {
-    x.param(param_type(static_cast<result_type>(min),
-                       static_cast<result_type>(max),
-                       static_cast<result_type>(base)));
-  }
-  return is;
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_LOG_UNIFORM_INT_DISTRIBUTION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/mock_distributions.h b/projects/llm_framework/include/abseil-cpp/absl/random/mock_distributions.h
deleted file mode 100644
index 764ab370..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/mock_distributions.h
+++ /dev/null
@@ -1,266 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: mock_distributions.h
-// -----------------------------------------------------------------------------
-//
-// This file contains mock distribution functions for use alongside an
-// `absl::MockingBitGen` object within the Googletest testing framework. Such
-// mocks are useful to provide deterministic values as return values within
-// (otherwise random) Abseil distribution functions.
-//
-// The return type of each function is a mock expectation object which
-// is used to set the match result.
-//
-// More information about the Googletest testing framework is available at
-// https://github.com/google/googletest
-//
-// EXPECT_CALL and ON_CALL need to be made within the same DLL component as
-// the call to absl::Uniform and related methods, otherwise mocking will fail
-// since the  underlying implementation creates a type-specific pointer which
-// will be distinct across different DLL boundaries.
-//
-// Example:
-//
-//   absl::MockingBitGen mock;
-//   EXPECT_CALL(absl::MockUniform<int>(), Call(mock, 1, 1000))
-//     .WillRepeatedly(testing::ReturnRoundRobin({20, 40}));
-//
-//   EXPECT_EQ(absl::Uniform<int>(gen, 1, 1000), 20);
-//   EXPECT_EQ(absl::Uniform<int>(gen, 1, 1000), 40);
-//   EXPECT_EQ(absl::Uniform<int>(gen, 1, 1000), 20);
-//   EXPECT_EQ(absl::Uniform<int>(gen, 1, 1000), 40);
-
-#ifndef ABSL_RANDOM_MOCK_DISTRIBUTIONS_H_
-#define ABSL_RANDOM_MOCK_DISTRIBUTIONS_H_
-
-#include <limits>
-#include <type_traits>
-#include <utility>
-
-#include "gmock/gmock.h"
-#include "gtest/gtest.h"
-#include "absl/meta/type_traits.h"
-#include "absl/random/distributions.h"
-#include "absl/random/internal/mock_overload_set.h"
-#include "absl/random/mocking_bit_gen.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// -----------------------------------------------------------------------------
-// absl::MockUniform
-// -----------------------------------------------------------------------------
-//
-// Matches calls to absl::Uniform.
-//
-// `absl::MockUniform` is a class template used in conjunction with Googletest's
-// `ON_CALL()` and `EXPECT_CALL()` macros. To use it, default-construct an
-// instance of it inside `ON_CALL()` or `EXPECT_CALL()`, and use `Call(...)` the
-// same way one would define mocks on a Googletest `MockFunction()`.
-//
-// Example:
-//
-//  absl::MockingBitGen mock;
-//  EXPECT_CALL(absl::MockUniform<uint32_t>(), Call(mock))
-//     .WillOnce(Return(123456));
-//  auto x = absl::Uniform<uint32_t>(mock);
-//  assert(x == 123456)
-//
-template <typename R>
-using MockUniform = random_internal::MockOverloadSet<
-    random_internal::UniformDistributionWrapper<R>,
-    R(IntervalClosedOpenTag, MockingBitGen&, R, R),
-    R(IntervalClosedClosedTag, MockingBitGen&, R, R),
-    R(IntervalOpenOpenTag, MockingBitGen&, R, R),
-    R(IntervalOpenClosedTag, MockingBitGen&, R, R), R(MockingBitGen&, R, R),
-    R(MockingBitGen&)>;
-
-// -----------------------------------------------------------------------------
-// absl::MockBernoulli
-// -----------------------------------------------------------------------------
-//
-// Matches calls to absl::Bernoulli.
-//
-// `absl::MockBernoulli` is a class used in conjunction with Googletest's
-// `ON_CALL()` and `EXPECT_CALL()` macros. To use it, default-construct an
-// instance of it inside `ON_CALL()` or `EXPECT_CALL()`, and use `Call(...)` the
-// same way one would define mocks on a Googletest `MockFunction()`.
-//
-// Example:
-//
-//  absl::MockingBitGen mock;
-//  EXPECT_CALL(absl::MockBernoulli(), Call(mock, testing::_))
-//     .WillOnce(Return(false));
-//  assert(absl::Bernoulli(mock, 0.5) == false);
-//
-using MockBernoulli =
-    random_internal::MockOverloadSet<absl::bernoulli_distribution,
-                                     bool(MockingBitGen&, double)>;
-
-// -----------------------------------------------------------------------------
-// absl::MockBeta
-// -----------------------------------------------------------------------------
-//
-// Matches calls to absl::Beta.
-//
-// `absl::MockBeta` is a class used in conjunction with Googletest's `ON_CALL()`
-// and `EXPECT_CALL()` macros. To use it, default-construct an instance of it
-// inside `ON_CALL()` or `EXPECT_CALL()`, and use `Call(...)` the same way one
-// would define mocks on a Googletest `MockFunction()`.
-//
-// Example:
-//
-//  absl::MockingBitGen mock;
-//  EXPECT_CALL(absl::MockBeta(), Call(mock, 3.0, 2.0))
-//     .WillOnce(Return(0.567));
-//  auto x = absl::Beta<double>(mock, 3.0, 2.0);
-//  assert(x == 0.567);
-//
-template <typename RealType>
-using MockBeta =
-    random_internal::MockOverloadSet<absl::beta_distribution<RealType>,
-                                     RealType(MockingBitGen&, RealType,
-                                              RealType)>;
-
-// -----------------------------------------------------------------------------
-// absl::MockExponential
-// -----------------------------------------------------------------------------
-//
-// Matches calls to absl::Exponential.
-//
-// `absl::MockExponential` is a class template used in conjunction with
-// Googletest's `ON_CALL()` and `EXPECT_CALL()` macros. To use it,
-// default-construct an instance of it inside `ON_CALL()` or `EXPECT_CALL()`,
-// and use `Call(...)` the same way one would define mocks on a
-// Googletest `MockFunction()`.
-//
-// Example:
-//
-//  absl::MockingBitGen mock;
-//  EXPECT_CALL(absl::MockExponential<double>(), Call(mock, 0.5))
-//     .WillOnce(Return(12.3456789));
-//  auto x = absl::Exponential<double>(mock, 0.5);
-//  assert(x == 12.3456789)
-//
-template <typename RealType>
-using MockExponential =
-    random_internal::MockOverloadSet<absl::exponential_distribution<RealType>,
-                                     RealType(MockingBitGen&, RealType)>;
-
-// -----------------------------------------------------------------------------
-// absl::MockGaussian
-// -----------------------------------------------------------------------------
-//
-// Matches calls to absl::Gaussian.
-//
-// `absl::MockGaussian` is a class template used in conjunction with
-// Googletest's `ON_CALL()` and `EXPECT_CALL()` macros. To use it,
-// default-construct an instance of it inside `ON_CALL()` or `EXPECT_CALL()`,
-// and use `Call(...)` the same way one would define mocks on a
-// Googletest `MockFunction()`.
-//
-// Example:
-//
-//  absl::MockingBitGen mock;
-//  EXPECT_CALL(absl::MockGaussian<double>(), Call(mock, 16.3, 3.3))
-//     .WillOnce(Return(12.3456789));
-//  auto x = absl::Gaussian<double>(mock, 16.3, 3.3);
-//  assert(x == 12.3456789)
-//
-template <typename RealType>
-using MockGaussian =
-    random_internal::MockOverloadSet<absl::gaussian_distribution<RealType>,
-                                     RealType(MockingBitGen&, RealType,
-                                              RealType)>;
-
-// -----------------------------------------------------------------------------
-// absl::MockLogUniform
-// -----------------------------------------------------------------------------
-//
-// Matches calls to absl::LogUniform.
-//
-// `absl::MockLogUniform` is a class template used in conjunction with
-// Googletest's `ON_CALL()` and `EXPECT_CALL()` macros. To use it,
-// default-construct an instance of it inside `ON_CALL()` or `EXPECT_CALL()`,
-// and use `Call(...)` the same way one would define mocks on a
-// Googletest `MockFunction()`.
-//
-// Example:
-//
-//  absl::MockingBitGen mock;
-//  EXPECT_CALL(absl::MockLogUniform<int>(), Call(mock, 10, 10000, 10))
-//     .WillOnce(Return(1221));
-//  auto x = absl::LogUniform<int>(mock, 10, 10000, 10);
-//  assert(x == 1221)
-//
-template <typename IntType>
-using MockLogUniform = random_internal::MockOverloadSet<
-    absl::log_uniform_int_distribution<IntType>,
-    IntType(MockingBitGen&, IntType, IntType, IntType)>;
-
-// -----------------------------------------------------------------------------
-// absl::MockPoisson
-// -----------------------------------------------------------------------------
-//
-// Matches calls to absl::Poisson.
-//
-// `absl::MockPoisson` is a class template used in conjunction with Googletest's
-// `ON_CALL()` and `EXPECT_CALL()` macros. To use it, default-construct an
-// instance of it inside `ON_CALL()` or `EXPECT_CALL()`, and use `Call(...)` the
-// same way one would define mocks on a Googletest `MockFunction()`.
-//
-// Example:
-//
-//  absl::MockingBitGen mock;
-//  EXPECT_CALL(absl::MockPoisson<int>(), Call(mock, 2.0))
-//     .WillOnce(Return(1221));
-//  auto x = absl::Poisson<int>(mock, 2.0);
-//  assert(x == 1221)
-//
-template <typename IntType>
-using MockPoisson =
-    random_internal::MockOverloadSet<absl::poisson_distribution<IntType>,
-                                     IntType(MockingBitGen&, double)>;
-
-// -----------------------------------------------------------------------------
-// absl::MockZipf
-// -----------------------------------------------------------------------------
-//
-// Matches calls to absl::Zipf.
-//
-// `absl::MockZipf` is a class template used in conjunction with Googletest's
-// `ON_CALL()` and `EXPECT_CALL()` macros. To use it, default-construct an
-// instance of it inside `ON_CALL()` or `EXPECT_CALL()`, and use `Call(...)` the
-// same way one would define mocks on a Googletest `MockFunction()`.
-//
-// Example:
-//
-//  absl::MockingBitGen mock;
-//  EXPECT_CALL(absl::MockZipf<int>(), Call(mock, 1000000, 2.0, 1.0))
-//     .WillOnce(Return(1221));
-//  auto x = absl::Zipf<int>(mock, 1000000, 2.0, 1.0);
-//  assert(x == 1221)
-//
-template <typename IntType>
-using MockZipf =
-    random_internal::MockOverloadSet<absl::zipf_distribution<IntType>,
-                                     IntType(MockingBitGen&, IntType, double,
-                                             double)>;
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_MOCK_DISTRIBUTIONS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/mocking_bit_gen.h b/projects/llm_framework/include/abseil-cpp/absl/random/mocking_bit_gen.h
deleted file mode 100644
index 89fa5a47..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/mocking_bit_gen.h
+++ /dev/null
@@ -1,240 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// mocking_bit_gen.h
-// -----------------------------------------------------------------------------
-//
-// This file includes an `absl::MockingBitGen` class to use as a mock within the
-// Googletest testing framework. Such a mock is useful to provide deterministic
-// values as return values within (otherwise random) Abseil distribution
-// functions. Such determinism within a mock is useful within testing frameworks
-// to test otherwise indeterminate APIs.
-//
-// More information about the Googletest testing framework is available at
-// https://github.com/google/googletest
-
-#ifndef ABSL_RANDOM_MOCKING_BIT_GEN_H_
-#define ABSL_RANDOM_MOCKING_BIT_GEN_H_
-
-#include <iterator>
-#include <limits>
-#include <memory>
-#include <tuple>
-#include <type_traits>
-#include <utility>
-
-#include "gmock/gmock.h"
-#include "gtest/gtest.h"
-#include "absl/base/internal/fast_type_id.h"
-#include "absl/container/flat_hash_map.h"
-#include "absl/meta/type_traits.h"
-#include "absl/random/distributions.h"
-#include "absl/random/internal/distribution_caller.h"
-#include "absl/random/random.h"
-#include "absl/strings/str_cat.h"
-#include "absl/strings/str_join.h"
-#include "absl/types/span.h"
-#include "absl/types/variant.h"
-#include "absl/utility/utility.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-namespace random_internal {
-template <typename>
-struct DistributionCaller;
-class MockHelpers;
-
-}  // namespace random_internal
-class BitGenRef;
-
-// MockingBitGen
-//
-// `absl::MockingBitGen` is a mock Uniform Random Bit Generator (URBG) class
-// which can act in place of an `absl::BitGen` URBG within tests using the
-// Googletest testing framework.
-//
-// Usage:
-//
-// Use an `absl::MockingBitGen` along with a mock distribution object (within
-// mock_distributions.h) inside Googletest constructs such as ON_CALL(),
-// EXPECT_TRUE(), etc. to produce deterministic results conforming to the
-// distribution's API contract.
-//
-// Example:
-//
-//  // Mock a call to an `absl::Bernoulli` distribution using Googletest
-//   absl::MockingBitGen bitgen;
-//
-//   ON_CALL(absl::MockBernoulli(), Call(bitgen, 0.5))
-//       .WillByDefault(testing::Return(true));
-//   EXPECT_TRUE(absl::Bernoulli(bitgen, 0.5));
-//
-//  // Mock a call to an `absl::Uniform` distribution within Googletest
-//  absl::MockingBitGen bitgen;
-//
-//   ON_CALL(absl::MockUniform<int>(), Call(bitgen, testing::_, testing::_))
-//       .WillByDefault([] (int low, int high) {
-//           return low + (high - low) / 2;
-//       });
-//
-//   EXPECT_EQ(absl::Uniform<int>(gen, 0, 10), 5);
-//   EXPECT_EQ(absl::Uniform<int>(gen, 30, 40), 35);
-//
-// At this time, only mock distributions supplied within the Abseil random
-// library are officially supported.
-//
-// EXPECT_CALL and ON_CALL need to be made within the same DLL component as
-// the call to absl::Uniform and related methods, otherwise mocking will fail
-// since the  underlying implementation creates a type-specific pointer which
-// will be distinct across different DLL boundaries.
-//
-class MockingBitGen {
- public:
-  MockingBitGen() = default;
-  ~MockingBitGen() = default;
-
-  // URBG interface
-  using result_type = absl::BitGen::result_type;
-
-  static constexpr result_type(min)() { return (absl::BitGen::min)(); }
-  static constexpr result_type(max)() { return (absl::BitGen::max)(); }
-  result_type operator()() { return gen_(); }
-
- private:
-  // GetMockFnType returns the testing::MockFunction for a result and tuple.
-  // This method only exists for type deduction and is otherwise unimplemented.
-  template <typename ResultT, typename... Args>
-  static auto GetMockFnType(ResultT, std::tuple<Args...>)
-      -> ::testing::MockFunction<ResultT(Args...)>;
-
-  // MockFnCaller is a helper method for use with absl::apply to
-  // apply an ArgTupleT to a compatible MockFunction.
-  // NOTE: MockFnCaller is essentially equivalent to the lambda:
-  // [fn](auto... args) { return fn->Call(std::move(args)...)}
-  // however that fails to build on some supported platforms.
-  template <typename MockFnType, typename ResultT, typename Tuple>
-  struct MockFnCaller;
-
-  // specialization for std::tuple.
-  template <typename MockFnType, typename ResultT, typename... Args>
-  struct MockFnCaller<MockFnType, ResultT, std::tuple<Args...>> {
-    MockFnType* fn;
-    inline ResultT operator()(Args... args) {
-      return fn->Call(std::move(args)...);
-    }
-  };
-
-  // FunctionHolder owns a particular ::testing::MockFunction associated with
-  // a mocked type signature, and implement the type-erased Apply call, which
-  // applies type-erased arguments to the mock.
-  class FunctionHolder {
-   public:
-    virtual ~FunctionHolder() = default;
-
-    // Call is a dispatch function which converts the
-    // generic type-erased parameters into a specific mock invocation call.
-    virtual void Apply(/*ArgTupleT*/ void* args_tuple,
-                       /*ResultT*/ void* result) = 0;
-  };
-
-  template <typename MockFnType, typename ResultT, typename ArgTupleT>
-  class FunctionHolderImpl final : public FunctionHolder {
-   public:
-    void Apply(void* args_tuple, void* result) override {
-      // Requires tuple_args to point to a ArgTupleT, which is a
-      // std::tuple<Args...> used to invoke the mock function. Requires result
-      // to point to a ResultT, which is the result of the call.
-      *static_cast<ResultT*>(result) =
-          absl::apply(MockFnCaller<MockFnType, ResultT, ArgTupleT>{&mock_fn_},
-                      *static_cast<ArgTupleT*>(args_tuple));
-    }
-
-    MockFnType mock_fn_;
-  };
-
-  // MockingBitGen::RegisterMock
-  //
-  // RegisterMock<ResultT, ArgTupleT>(FastTypeIdType) is the main extension
-  // point for extending the MockingBitGen framework. It provides a mechanism to
-  // install a mock expectation for a function like ResultT(Args...) keyed by
-  // type_idex onto the MockingBitGen context. The key is that the type_index
-  // used to register must match the type index used to call the mock.
-  //
-  // The returned MockFunction<...> type can be used to setup additional
-  // distribution parameters of the expectation.
-  template <typename ResultT, typename ArgTupleT, typename SelfT>
-  auto RegisterMock(SelfT&, base_internal::FastTypeIdType type)
-      -> decltype(GetMockFnType(std::declval<ResultT>(),
-                                std::declval<ArgTupleT>()))& {
-    using MockFnType = decltype(GetMockFnType(std::declval<ResultT>(),
-                                              std::declval<ArgTupleT>()));
-
-    using WrappedFnType = absl::conditional_t<
-        std::is_same<SelfT, ::testing::NiceMock<absl::MockingBitGen>>::value,
-        ::testing::NiceMock<MockFnType>,
-        absl::conditional_t<
-            std::is_same<SelfT,
-                         ::testing::NaggyMock<absl::MockingBitGen>>::value,
-            ::testing::NaggyMock<MockFnType>,
-            absl::conditional_t<
-                std::is_same<SelfT,
-                             ::testing::StrictMock<absl::MockingBitGen>>::value,
-                ::testing::StrictMock<MockFnType>, MockFnType>>>;
-
-    using ImplT = FunctionHolderImpl<WrappedFnType, ResultT, ArgTupleT>;
-    auto& mock = mocks_[type];
-    if (!mock) {
-      mock = absl::make_unique<ImplT>();
-    }
-    return static_cast<ImplT*>(mock.get())->mock_fn_;
-  }
-
-  // MockingBitGen::InvokeMock
-  //
-  // InvokeMock(FastTypeIdType, args, result) is the entrypoint for invoking
-  // mocks registered on MockingBitGen.
-  //
-  // When no mocks are registered on the provided FastTypeIdType, returns false.
-  // Otherwise attempts to invoke the mock function ResultT(Args...) that
-  // was previously registered via the type_index.
-  // Requires tuple_args to point to a ArgTupleT, which is a std::tuple<Args...>
-  // used to invoke the mock function.
-  // Requires result to point to a ResultT, which is the result of the call.
-  inline bool InvokeMock(base_internal::FastTypeIdType type, void* args_tuple,
-                         void* result) {
-    // Trigger a mock, if there exists one that matches `param`.
-    auto it = mocks_.find(type);
-    if (it == mocks_.end()) return false;
-    it->second->Apply(args_tuple, result);
-    return true;
-  }
-
-  absl::flat_hash_map<base_internal::FastTypeIdType,
-                      std::unique_ptr<FunctionHolder>>
-      mocks_;
-  absl::BitGen gen_;
-
-  template <typename>
-  friend struct ::absl::random_internal::DistributionCaller;  // for InvokeMock
-  friend class ::absl::BitGenRef;                             // for InvokeMock
-  friend class ::absl::random_internal::MockHelpers;  // for RegisterMock,
-                                                      // InvokeMock
-};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_MOCKING_BIT_GEN_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/poisson_distribution.h b/projects/llm_framework/include/abseil-cpp/absl/random/poisson_distribution.h
deleted file mode 100644
index f4573082..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/poisson_distribution.h
+++ /dev/null
@@ -1,261 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_POISSON_DISTRIBUTION_H_
-#define ABSL_RANDOM_POISSON_DISTRIBUTION_H_
-
-#include <cassert>
-#include <cmath>
-#include <istream>
-#include <limits>
-#include <ostream>
-#include <type_traits>
-
-#include "absl/random/internal/fast_uniform_bits.h"
-#include "absl/random/internal/fastmath.h"
-#include "absl/random/internal/generate_real.h"
-#include "absl/random/internal/iostream_state_saver.h"
-#include "absl/random/internal/traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// absl::poisson_distribution:
-// Generates discrete variates conforming to a Poisson distribution.
-//   p(n) = (mean^n / n!) exp(-mean)
-//
-// Depending on the parameter, the distribution selects one of the following
-// algorithms:
-// * The standard algorithm, attributed to Knuth, extended using a split method
-// for larger values
-// * The "Ratio of Uniforms as a convenient method for sampling from classical
-// discrete distributions", Stadlober, 1989.
-// http://www.sciencedirect.com/science/article/pii/0377042790903495
-//
-// NOTE: param_type.mean() is a double, which permits values larger than
-// poisson_distribution<IntType>::max(), however this should be avoided and
-// the distribution results are limited to the max() value.
-//
-// The goals of this implementation are to provide good performance while still
-// beig thread-safe: This limits the implementation to not using lgamma provided
-// by <math.h>.
-//
-template <typename IntType = int>
-class poisson_distribution {
- public:
-  using result_type = IntType;
-
-  class param_type {
-   public:
-    using distribution_type = poisson_distribution;
-    explicit param_type(double mean = 1.0);
-
-    double mean() const { return mean_; }
-
-    friend bool operator==(const param_type& a, const param_type& b) {
-      return a.mean_ == b.mean_;
-    }
-
-    friend bool operator!=(const param_type& a, const param_type& b) {
-      return !(a == b);
-    }
-
-   private:
-    friend class poisson_distribution;
-
-    double mean_;
-    double emu_;  // e ^ -mean_
-    double lmu_;  // ln(mean_)
-    double s_;
-    double log_k_;
-    int split_;
-
-    static_assert(random_internal::IsIntegral<IntType>::value,
-                  "Class-template absl::poisson_distribution<> must be "
-                  "parameterized using an integral type.");
-  };
-
-  poisson_distribution() : poisson_distribution(1.0) {}
-
-  explicit poisson_distribution(double mean) : param_(mean) {}
-
-  explicit poisson_distribution(const param_type& p) : param_(p) {}
-
-  void reset() {}
-
-  // generating functions
-  template <typename URBG>
-  result_type operator()(URBG& g) {  // NOLINT(runtime/references)
-    return (*this)(g, param_);
-  }
-
-  template <typename URBG>
-  result_type operator()(URBG& g,  // NOLINT(runtime/references)
-                         const param_type& p);
-
-  param_type param() const { return param_; }
-  void param(const param_type& p) { param_ = p; }
-
-  result_type(min)() const { return 0; }
-  result_type(max)() const { return (std::numeric_limits<result_type>::max)(); }
-
-  double mean() const { return param_.mean(); }
-
-  friend bool operator==(const poisson_distribution& a,
-                         const poisson_distribution& b) {
-    return a.param_ == b.param_;
-  }
-  friend bool operator!=(const poisson_distribution& a,
-                         const poisson_distribution& b) {
-    return a.param_ != b.param_;
-  }
-
- private:
-  param_type param_;
-  random_internal::FastUniformBits<uint64_t> fast_u64_;
-};
-
-// -----------------------------------------------------------------------------
-// Implementation details follow
-// -----------------------------------------------------------------------------
-
-template <typename IntType>
-poisson_distribution<IntType>::param_type::param_type(double mean)
-    : mean_(mean), split_(0) {
-  assert(mean >= 0);
-  assert(mean <=
-         static_cast<double>((std::numeric_limits<result_type>::max)()));
-  // As a defensive measure, avoid large values of the mean.  The rejection
-  // algorithm used does not support very large values well.  It my be worth
-  // changing algorithms to better deal with these cases.
-  assert(mean <= 1e10);
-  if (mean_ < 10) {
-    // For small lambda, use the knuth method.
-    split_ = 1;
-    emu_ = std::exp(-mean_);
-  } else if (mean_ <= 50) {
-    // Use split-knuth method.
-    split_ = 1 + static_cast<int>(mean_ / 10.0);
-    emu_ = std::exp(-mean_ / static_cast<double>(split_));
-  } else {
-    // Use ratio of uniforms method.
-    constexpr double k2E = 0.7357588823428846;
-    constexpr double kSA = 0.4494580810294493;
-
-    lmu_ = std::log(mean_);
-    double a = mean_ + 0.5;
-    s_ = kSA + std::sqrt(k2E * a);
-    const double mode = std::ceil(mean_) - 1;
-    log_k_ = lmu_ * mode - absl::random_internal::StirlingLogFactorial(mode);
-  }
-}
-
-template <typename IntType>
-template <typename URBG>
-typename poisson_distribution<IntType>::result_type
-poisson_distribution<IntType>::operator()(
-    URBG& g,  // NOLINT(runtime/references)
-    const param_type& p) {
-  using random_internal::GeneratePositiveTag;
-  using random_internal::GenerateRealFromBits;
-  using random_internal::GenerateSignedTag;
-
-  if (p.split_ != 0) {
-    // Use Knuth's algorithm with range splitting to avoid floating-point
-    // errors. Knuth's algorithm is: Ui is a sequence of uniform variates on
-    // (0,1); return the number of variates required for product(Ui) <
-    // exp(-lambda).
-    //
-    // The expected number of variates required for Knuth's method can be
-    // computed as follows:
-    // The expected value of U is 0.5, so solving for 0.5^n < exp(-lambda) gives
-    // the expected number of uniform variates
-    // required for a given lambda, which is:
-    //  lambda = [2, 5,  9, 10, 11, 12, 13, 14, 15, 16, 17]
-    //  n      = [3, 8, 13, 15, 16, 18, 19, 21, 22, 24, 25]
-    //
-    result_type n = 0;
-    for (int split = p.split_; split > 0; --split) {
-      double r = 1.0;
-      do {
-        r *= GenerateRealFromBits<double, GeneratePositiveTag, true>(
-            fast_u64_(g));  // U(-1, 0)
-        ++n;
-      } while (r > p.emu_);
-      --n;
-    }
-    return n;
-  }
-
-  // Use ratio of uniforms method.
-  //
-  // Let u ~ Uniform(0, 1), v ~ Uniform(-1, 1),
-  //     a = lambda + 1/2,
-  //     s = 1.5 - sqrt(3/e) + sqrt(2(lambda + 1/2)/e),
-  //     x = s * v/u + a.
-  // P(floor(x) = k | u^2 < f(floor(x))/k), where
-  // f(m) = lambda^m exp(-lambda)/ m!, for 0 <= m, and f(m) = 0 otherwise,
-  // and k = max(f).
-  const double a = p.mean_ + 0.5;
-  for (;;) {
-    const double u = GenerateRealFromBits<double, GeneratePositiveTag, false>(
-        fast_u64_(g));  // U(0, 1)
-    const double v = GenerateRealFromBits<double, GenerateSignedTag, false>(
-        fast_u64_(g));  // U(-1, 1)
-
-    const double x = std::floor(p.s_ * v / u + a);
-    if (x < 0) continue;  // f(negative) = 0
-    const double rhs = x * p.lmu_;
-    // clang-format off
-    double s = (x <= 1.0) ? 0.0
-             : (x == 2.0) ? 0.693147180559945
-             : absl::random_internal::StirlingLogFactorial(x);
-    // clang-format on
-    const double lhs = 2.0 * std::log(u) + p.log_k_ + s;
-    if (lhs < rhs) {
-      return x > static_cast<double>((max)())
-                 ? (max)()
-                 : static_cast<result_type>(x);  // f(x)/k >= u^2
-    }
-  }
-}
-
-template <typename CharT, typename Traits, typename IntType>
-std::basic_ostream<CharT, Traits>& operator<<(
-    std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
-    const poisson_distribution<IntType>& x) {
-  auto saver = random_internal::make_ostream_state_saver(os);
-  os.precision(random_internal::stream_precision_helper<double>::kPrecision);
-  os << x.mean();
-  return os;
-}
-
-template <typename CharT, typename Traits, typename IntType>
-std::basic_istream<CharT, Traits>& operator>>(
-    std::basic_istream<CharT, Traits>& is,  // NOLINT(runtime/references)
-    poisson_distribution<IntType>& x) {     // NOLINT(runtime/references)
-  using param_type = typename poisson_distribution<IntType>::param_type;
-
-  auto saver = random_internal::make_istream_state_saver(is);
-  double mean = random_internal::read_floating_point<double>(is);
-  if (!is.fail()) {
-    x.param(param_type(mean));
-  }
-  return is;
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_POISSON_DISTRIBUTION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/random.h b/projects/llm_framework/include/abseil-cpp/absl/random/random.h
deleted file mode 100644
index 76720867..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/random.h
+++ /dev/null
@@ -1,189 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: random.h
-// -----------------------------------------------------------------------------
-//
-// This header defines the recommended Uniform Random Bit Generator (URBG)
-// types for use within the Abseil Random library. These types are not
-// suitable for security-related use-cases, but should suffice for most other
-// uses of generating random values.
-//
-// The Abseil random library provides the following URBG types:
-//
-//   * BitGen, a good general-purpose bit generator, optimized for generating
-//     random (but not cryptographically secure) values
-//   * InsecureBitGen, a slightly faster, though less random, bit generator, for
-//     cases where the existing BitGen is a drag on performance.
-
-#ifndef ABSL_RANDOM_RANDOM_H_
-#define ABSL_RANDOM_RANDOM_H_
-
-#include <random>
-
-#include "absl/random/distributions.h"  // IWYU pragma: export
-#include "absl/random/internal/nonsecure_base.h"  // IWYU pragma: export
-#include "absl/random/internal/pcg_engine.h"  // IWYU pragma: export
-#include "absl/random/internal/pool_urbg.h"
-#include "absl/random/internal/randen_engine.h"
-#include "absl/random/seed_sequences.h"  // IWYU pragma: export
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// -----------------------------------------------------------------------------
-// absl::BitGen
-// -----------------------------------------------------------------------------
-//
-// `absl::BitGen` is a general-purpose random bit generator for generating
-// random values for use within the Abseil random library. Typically, you use a
-// bit generator in combination with a distribution to provide random values.
-//
-// Example:
-//
-//   // Create an absl::BitGen. There is no need to seed this bit generator.
-//   absl::BitGen gen;
-//
-//   // Generate an integer value in the closed interval [1,6]
-//   int die_roll = absl::uniform_int_distribution<int>(1, 6)(gen);
-//
-// `absl::BitGen` is seeded by default with non-deterministic data to produce
-// different sequences of random values across different instances, including
-// different binary invocations. This behavior is different than the standard
-// library bit generators, which use golden values as their seeds. Default
-// construction intentionally provides no stability guarantees, to avoid
-// accidental dependence on such a property.
-//
-// `absl::BitGen` may be constructed with an optional seed sequence type,
-// conforming to [rand.req.seed_seq], which will be mixed with additional
-// non-deterministic data as detailed below.
-//
-// Example:
-//
-//  // Create an absl::BitGen using an std::seed_seq seed sequence
-//  std::seed_seq seq{1,2,3};
-//  absl::BitGen gen_with_seed(seq);
-//
-//  // Generate an integer value in the closed interval [1,6]
-//  int die_roll2 = absl::uniform_int_distribution<int>(1, 6)(gen_with_seed);
-//
-// Constructing two `absl::BitGen`s with the same seed sequence in the same
-// process will produce the same sequence of variates, but need not do so across
-// multiple processes even if they're executing the same binary.
-//
-// `absl::BitGen` meets the requirements of the Uniform Random Bit Generator
-// (URBG) concept as per the C++17 standard [rand.req.urng] though differs
-// slightly with [rand.req.eng]. Like its standard library equivalents (e.g.
-// `std::mersenne_twister_engine`) `absl::BitGen` is not cryptographically
-// secure.
-//
-// This type has been optimized to perform better than Mersenne Twister
-// (https://en.wikipedia.org/wiki/Mersenne_Twister) and many other complex URBG
-// types on modern x86, ARM, and PPC architectures.
-//
-// This type is thread-compatible, but not thread-safe.
-
-// ---------------------------------------------------------------------------
-// absl::BitGen member functions
-// ---------------------------------------------------------------------------
-
-// absl::BitGen::operator()()
-//
-// Calls the BitGen, returning a generated value.
-
-// absl::BitGen::min()
-//
-// Returns the smallest possible value from this bit generator.
-
-// absl::BitGen::max()
-//
-// Returns the largest possible value from this bit generator.
-
-// absl::BitGen::discard(num)
-//
-// Advances the internal state of this bit generator by `num` times, and
-// discards the intermediate results.
-// ---------------------------------------------------------------------------
-
-using BitGen = random_internal::NonsecureURBGBase<
-    random_internal::randen_engine<uint64_t>>;
-
-// -----------------------------------------------------------------------------
-// absl::InsecureBitGen
-// -----------------------------------------------------------------------------
-//
-// `absl::InsecureBitGen` is an efficient random bit generator for generating
-// random values, recommended only for performance-sensitive use cases where
-// `absl::BitGen` is not satisfactory when compute-bounded by bit generation
-// costs.
-//
-// Example:
-//
-//   // Create an absl::InsecureBitGen
-//   absl::InsecureBitGen gen;
-//   for (size_t i = 0; i < 1000000; i++) {
-//
-//     // Generate a bunch of random values from some complex distribution
-//     auto my_rnd = some_distribution(gen, 1, 1000);
-//   }
-//
-// Like `absl::BitGen`, `absl::InsecureBitGen` is seeded by default with
-// non-deterministic data to produce different sequences of random values across
-// different instances, including different binary invocations. (This behavior
-// is different than the standard library bit generators, which use golden
-// values as their seeds.)
-//
-// `absl::InsecureBitGen` may be constructed with an optional seed sequence
-// type, conforming to [rand.req.seed_seq], which will be mixed with additional
-// non-deterministic data, as detailed in the `absl::BitGen` comment.
-//
-// `absl::InsecureBitGen` meets the requirements of the Uniform Random Bit
-// Generator (URBG) concept as per the C++17 standard [rand.req.urng] though
-// its implementation differs slightly with [rand.req.eng]. Like its standard
-// library equivalents (e.g. `std::mersenne_twister_engine`)
-// `absl::InsecureBitGen` is not cryptographically secure.
-//
-// Prefer `absl::BitGen` over `absl::InsecureBitGen` as the general type is
-// often fast enough for the vast majority of applications.
-
-using InsecureBitGen =
-    random_internal::NonsecureURBGBase<random_internal::pcg64_2018_engine>;
-
-// ---------------------------------------------------------------------------
-// absl::InsecureBitGen member functions
-// ---------------------------------------------------------------------------
-
-// absl::InsecureBitGen::operator()()
-//
-// Calls the InsecureBitGen, returning a generated value.
-
-// absl::InsecureBitGen::min()
-//
-// Returns the smallest possible value from this bit generator.
-
-// absl::InsecureBitGen::max()
-//
-// Returns the largest possible value from this bit generator.
-
-// absl::InsecureBitGen::discard(num)
-//
-// Advances the internal state of this bit generator by `num` times, and
-// discards the intermediate results.
-// ---------------------------------------------------------------------------
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_RANDOM_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/seed_gen_exception.h b/projects/llm_framework/include/abseil-cpp/absl/random/seed_gen_exception.h
deleted file mode 100644
index 53539005..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/seed_gen_exception.h
+++ /dev/null
@@ -1,55 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: seed_gen_exception.h
-// -----------------------------------------------------------------------------
-//
-// This header defines an exception class which may be thrown if unpredictable
-// events prevent the derivation of suitable seed-material for constructing a
-// bit generator conforming to [rand.req.urng] (eg. entropy cannot be read from
-// /dev/urandom on a Unix-based system).
-//
-// Note: if exceptions are disabled, `std::terminate()` is called instead.
-
-#ifndef ABSL_RANDOM_SEED_GEN_EXCEPTION_H_
-#define ABSL_RANDOM_SEED_GEN_EXCEPTION_H_
-
-#include <exception>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-//------------------------------------------------------------------------------
-// SeedGenException
-//------------------------------------------------------------------------------
-class SeedGenException : public std::exception {
- public:
-  SeedGenException() = default;
-  ~SeedGenException() override;
-  const char* what() const noexcept override;
-};
-
-namespace random_internal {
-
-// throw delegator
-[[noreturn]] void ThrowSeedGenException();
-
-}  // namespace random_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_SEED_GEN_EXCEPTION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/seed_sequences.h b/projects/llm_framework/include/abseil-cpp/absl/random/seed_sequences.h
deleted file mode 100644
index 33970be5..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/seed_sequences.h
+++ /dev/null
@@ -1,113 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: seed_sequences.h
-// -----------------------------------------------------------------------------
-//
-// This header contains utilities for creating and working with seed sequences
-// conforming to [rand.req.seedseq]. In general, direct construction of seed
-// sequences is discouraged, but use-cases for construction of identical bit
-// generators (using the same seed sequence) may be helpful (e.g. replaying a
-// simulation whose state is derived from variates of a bit generator).
-
-#ifndef ABSL_RANDOM_SEED_SEQUENCES_H_
-#define ABSL_RANDOM_SEED_SEQUENCES_H_
-
-#include <iterator>
-#include <random>
-
-#include "absl/base/config.h"
-#include "absl/base/nullability.h"
-#include "absl/random/internal/salted_seed_seq.h"
-#include "absl/random/internal/seed_material.h"
-#include "absl/random/seed_gen_exception.h"
-#include "absl/strings/string_view.h"
-#include "absl/types/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// -----------------------------------------------------------------------------
-// absl::SeedSeq
-// -----------------------------------------------------------------------------
-//
-// `absl::SeedSeq` constructs a seed sequence according to [rand.req.seedseq]
-// for use within bit generators. `absl::SeedSeq`, unlike `std::seed_seq`
-// additionally salts the generated seeds with extra implementation-defined
-// entropy. For that reason, you can use `absl::SeedSeq` in combination with
-// standard library bit generators (e.g. `std::mt19937`) to introduce
-// non-determinism in your seeds.
-//
-// Example:
-//
-//   absl::SeedSeq my_seed_seq({a, b, c});
-//   std::mt19937 my_bitgen(my_seed_seq);
-//
-using SeedSeq = random_internal::SaltedSeedSeq<std::seed_seq>;
-
-// -----------------------------------------------------------------------------
-// absl::CreateSeedSeqFrom(bitgen*)
-// -----------------------------------------------------------------------------
-//
-// Constructs a seed sequence conforming to [rand.req.seedseq] using variates
-// produced by a provided bit generator.
-//
-// You should generally avoid direct construction of seed sequences, but
-// use-cases for reuse of a seed sequence to construct identical bit generators
-// may be helpful (eg. replaying a simulation whose state is derived from bit
-// generator values).
-//
-// If bitgen == nullptr, then behavior is undefined.
-//
-// Example:
-//
-//   absl::BitGen my_bitgen;
-//   auto seed_seq = absl::CreateSeedSeqFrom(&my_bitgen);
-//   absl::BitGen new_engine(seed_seq); // derived from my_bitgen, but not
-//                                      // correlated.
-//
-template <typename URBG>
-SeedSeq CreateSeedSeqFrom(URBG* urbg) {
-  SeedSeq::result_type
-      seed_material[random_internal::kEntropyBlocksNeeded];
-
-  if (!random_internal::ReadSeedMaterialFromURBG(
-          urbg, absl::MakeSpan(seed_material))) {
-    random_internal::ThrowSeedGenException();
-  }
-  return SeedSeq(std::begin(seed_material), std::end(seed_material));
-}
-
-// -----------------------------------------------------------------------------
-// absl::MakeSeedSeq()
-// -----------------------------------------------------------------------------
-//
-// Constructs an `absl::SeedSeq` salting the generated values using
-// implementation-defined entropy. The returned sequence can be used to create
-// equivalent bit generators correlated using this sequence.
-//
-// Example:
-//
-//   auto my_seed_seq = absl::MakeSeedSeq();
-//   std::mt19937 rng1(my_seed_seq);
-//   std::mt19937 rng2(my_seed_seq);
-//   EXPECT_EQ(rng1(), rng2());
-//
-SeedSeq MakeSeedSeq();
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_SEED_SEQUENCES_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/uniform_int_distribution.h b/projects/llm_framework/include/abseil-cpp/absl/random/uniform_int_distribution.h
deleted file mode 100644
index fae80252..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/uniform_int_distribution.h
+++ /dev/null
@@ -1,275 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: uniform_int_distribution.h
-// -----------------------------------------------------------------------------
-//
-// This header defines a class for representing a uniform integer distribution
-// over the closed (inclusive) interval [a,b]. You use this distribution in
-// combination with an Abseil random bit generator to produce random values
-// according to the rules of the distribution.
-//
-// `absl::uniform_int_distribution` is a drop-in replacement for the C++11
-// `std::uniform_int_distribution` [rand.dist.uni.int] but is considerably
-// faster than the libstdc++ implementation.
-
-#ifndef ABSL_RANDOM_UNIFORM_INT_DISTRIBUTION_H_
-#define ABSL_RANDOM_UNIFORM_INT_DISTRIBUTION_H_
-
-#include <cassert>
-#include <istream>
-#include <limits>
-#include <type_traits>
-
-#include "absl/base/optimization.h"
-#include "absl/random/internal/fast_uniform_bits.h"
-#include "absl/random/internal/iostream_state_saver.h"
-#include "absl/random/internal/traits.h"
-#include "absl/random/internal/wide_multiply.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// absl::uniform_int_distribution<T>
-//
-// This distribution produces random integer values uniformly distributed in the
-// closed (inclusive) interval [a, b].
-//
-// Example:
-//
-//   absl::BitGen gen;
-//
-//   // Use the distribution to produce a value between 1 and 6, inclusive.
-//   int die_roll = absl::uniform_int_distribution<int>(1, 6)(gen);
-//
-template <typename IntType = int>
-class uniform_int_distribution {
- private:
-  using unsigned_type =
-      typename random_internal::make_unsigned_bits<IntType>::type;
-
- public:
-  using result_type = IntType;
-
-  class param_type {
-   public:
-    using distribution_type = uniform_int_distribution;
-
-    explicit param_type(
-        result_type lo = 0,
-        result_type hi = (std::numeric_limits<result_type>::max)())
-        : lo_(lo),
-          range_(static_cast<unsigned_type>(hi) -
-                 static_cast<unsigned_type>(lo)) {
-      // [rand.dist.uni.int] precondition 2
-      assert(lo <= hi);
-    }
-
-    result_type a() const { return lo_; }
-    result_type b() const {
-      return static_cast<result_type>(static_cast<unsigned_type>(lo_) + range_);
-    }
-
-    friend bool operator==(const param_type& a, const param_type& b) {
-      return a.lo_ == b.lo_ && a.range_ == b.range_;
-    }
-
-    friend bool operator!=(const param_type& a, const param_type& b) {
-      return !(a == b);
-    }
-
-   private:
-    friend class uniform_int_distribution;
-    unsigned_type range() const { return range_; }
-
-    result_type lo_;
-    unsigned_type range_;
-
-    static_assert(random_internal::IsIntegral<result_type>::value,
-                  "Class-template absl::uniform_int_distribution<> must be "
-                  "parameterized using an integral type.");
-  };  // param_type
-
-  uniform_int_distribution() : uniform_int_distribution(0) {}
-
-  explicit uniform_int_distribution(
-      result_type lo,
-      result_type hi = (std::numeric_limits<result_type>::max)())
-      : param_(lo, hi) {}
-
-  explicit uniform_int_distribution(const param_type& param) : param_(param) {}
-
-  // uniform_int_distribution<T>::reset()
-  //
-  // Resets the uniform int distribution. Note that this function has no effect
-  // because the distribution already produces independent values.
-  void reset() {}
-
-  template <typename URBG>
-  result_type operator()(URBG& gen) {  // NOLINT(runtime/references)
-    return (*this)(gen, param());
-  }
-
-  template <typename URBG>
-  result_type operator()(
-      URBG& gen, const param_type& param) {  // NOLINT(runtime/references)
-    return static_cast<result_type>(param.a() + Generate(gen, param.range()));
-  }
-
-  result_type a() const { return param_.a(); }
-  result_type b() const { return param_.b(); }
-
-  param_type param() const { return param_; }
-  void param(const param_type& params) { param_ = params; }
-
-  result_type(min)() const { return a(); }
-  result_type(max)() const { return b(); }
-
-  friend bool operator==(const uniform_int_distribution& a,
-                         const uniform_int_distribution& b) {
-    return a.param_ == b.param_;
-  }
-  friend bool operator!=(const uniform_int_distribution& a,
-                         const uniform_int_distribution& b) {
-    return !(a == b);
-  }
-
- private:
-  // Generates a value in the *closed* interval [0, R]
-  template <typename URBG>
-  unsigned_type Generate(URBG& g,  // NOLINT(runtime/references)
-                         unsigned_type R);
-  param_type param_;
-};
-
-// -----------------------------------------------------------------------------
-// Implementation details follow
-// -----------------------------------------------------------------------------
-template <typename CharT, typename Traits, typename IntType>
-std::basic_ostream<CharT, Traits>& operator<<(
-    std::basic_ostream<CharT, Traits>& os,
-    const uniform_int_distribution<IntType>& x) {
-  using stream_type =
-      typename random_internal::stream_format_type<IntType>::type;
-  auto saver = random_internal::make_ostream_state_saver(os);
-  os << static_cast<stream_type>(x.a()) << os.fill()
-     << static_cast<stream_type>(x.b());
-  return os;
-}
-
-template <typename CharT, typename Traits, typename IntType>
-std::basic_istream<CharT, Traits>& operator>>(
-    std::basic_istream<CharT, Traits>& is,
-    uniform_int_distribution<IntType>& x) {
-  using param_type = typename uniform_int_distribution<IntType>::param_type;
-  using result_type = typename uniform_int_distribution<IntType>::result_type;
-  using stream_type =
-      typename random_internal::stream_format_type<IntType>::type;
-
-  stream_type a;
-  stream_type b;
-
-  auto saver = random_internal::make_istream_state_saver(is);
-  is >> a >> b;
-  if (!is.fail()) {
-    x.param(
-        param_type(static_cast<result_type>(a), static_cast<result_type>(b)));
-  }
-  return is;
-}
-
-template <typename IntType>
-template <typename URBG>
-typename random_internal::make_unsigned_bits<IntType>::type
-uniform_int_distribution<IntType>::Generate(
-    URBG& g,  // NOLINT(runtime/references)
-    typename random_internal::make_unsigned_bits<IntType>::type R) {
-  random_internal::FastUniformBits<unsigned_type> fast_bits;
-  unsigned_type bits = fast_bits(g);
-  const unsigned_type Lim = R + 1;
-  if ((R & Lim) == 0) {
-    // If the interval's length is a power of two range, just take the low bits.
-    return bits & R;
-  }
-
-  // Generates a uniform variate on [0, Lim) using fixed-point multiplication.
-  // The above fast-path guarantees that Lim is representable in unsigned_type.
-  //
-  // Algorithm adapted from
-  // http://lemire.me/blog/2016/06/30/fast-random-shuffling/, with added
-  // explanation.
-  //
-  // The algorithm creates a uniform variate `bits` in the interval [0, 2^N),
-  // and treats it as the fractional part of a fixed-point real value in [0, 1),
-  // multiplied by 2^N.  For example, 0.25 would be represented as 2^(N - 2),
-  // because 2^N * 0.25 == 2^(N - 2).
-  //
-  // Next, `bits` and `Lim` are multiplied with a wide-multiply to bring the
-  // value into the range [0, Lim).  The integral part (the high word of the
-  // multiplication result) is then very nearly the desired result.  However,
-  // this is not quite accurate; viewing the multiplication result as one
-  // double-width integer, the resulting values for the sample are mapped as
-  // follows:
-  //
-  // If the result lies in this interval:       Return this value:
-  //        [0, 2^N)                                    0
-  //        [2^N, 2 * 2^N)                              1
-  //        ...                                         ...
-  //        [K * 2^N, (K + 1) * 2^N)                    K
-  //        ...                                         ...
-  //        [(Lim - 1) * 2^N, Lim * 2^N)                Lim - 1
-  //
-  // While all of these intervals have the same size, the result of `bits * Lim`
-  // must be a multiple of `Lim`, and not all of these intervals contain the
-  // same number of multiples of `Lim`.  In particular, some contain
-  // `F = floor(2^N / Lim)` and some contain `F + 1 = ceil(2^N / Lim)`.  This
-  // difference produces a small nonuniformity, which is corrected by applying
-  // rejection sampling to one of the values in the "larger intervals" (i.e.,
-  // the intervals containing `F + 1` multiples of `Lim`.
-  //
-  // An interval contains `F + 1` multiples of `Lim` if and only if its smallest
-  // value modulo 2^N is less than `2^N % Lim`.  The unique value satisfying
-  // this property is used as the one for rejection.  That is, a value of
-  // `bits * Lim` is rejected if `(bit * Lim) % 2^N < (2^N % Lim)`.
-
-  using helper = random_internal::wide_multiply<unsigned_type>;
-  auto product = helper::multiply(bits, Lim);
-
-  // Two optimizations here:
-  // * Rejection occurs with some probability less than 1/2, and for reasonable
-  //   ranges considerably less (in particular, less than 1/(F+1)), so
-  //   ABSL_PREDICT_FALSE is apt.
-  // * `Lim` is an overestimate of `threshold`, and doesn't require a divide.
-  if (ABSL_PREDICT_FALSE(helper::lo(product) < Lim)) {
-    // This quantity is exactly equal to `2^N % Lim`, but does not require high
-    // precision calculations: `2^N % Lim` is congruent to `(2^N - Lim) % Lim`.
-    // Ideally this could be expressed simply as `-X` rather than `2^N - X`, but
-    // for types smaller than int, this calculation is incorrect due to integer
-    // promotion rules.
-    const unsigned_type threshold =
-        ((std::numeric_limits<unsigned_type>::max)() - Lim + 1) % Lim;
-    while (helper::lo(product) < threshold) {
-      bits = fast_bits(g);
-      product = helper::multiply(bits, Lim);
-    }
-  }
-
-  return helper::hi(product);
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_UNIFORM_INT_DISTRIBUTION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/uniform_real_distribution.h b/projects/llm_framework/include/abseil-cpp/absl/random/uniform_real_distribution.h
deleted file mode 100644
index 19683341..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/uniform_real_distribution.h
+++ /dev/null
@@ -1,202 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: uniform_real_distribution.h
-// -----------------------------------------------------------------------------
-//
-// This header defines a class for representing a uniform floating-point
-// distribution over a half-open interval [a,b). You use this distribution in
-// combination with an Abseil random bit generator to produce random values
-// according to the rules of the distribution.
-//
-// `absl::uniform_real_distribution` is a drop-in replacement for the C++11
-// `std::uniform_real_distribution` [rand.dist.uni.real] but is considerably
-// faster than the libstdc++ implementation.
-//
-// Note: the standard-library version may occasionally return `1.0` when
-// default-initialized. See https://bugs.llvm.org//show_bug.cgi?id=18767
-// `absl::uniform_real_distribution` does not exhibit this behavior.
-
-#ifndef ABSL_RANDOM_UNIFORM_REAL_DISTRIBUTION_H_
-#define ABSL_RANDOM_UNIFORM_REAL_DISTRIBUTION_H_
-
-#include <cassert>
-#include <cmath>
-#include <cstdint>
-#include <istream>
-#include <limits>
-#include <type_traits>
-
-#include "absl/meta/type_traits.h"
-#include "absl/random/internal/fast_uniform_bits.h"
-#include "absl/random/internal/generate_real.h"
-#include "absl/random/internal/iostream_state_saver.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// absl::uniform_real_distribution<T>
-//
-// This distribution produces random floating-point values uniformly distributed
-// over the half-open interval [a, b).
-//
-// Example:
-//
-//   absl::BitGen gen;
-//
-//   // Use the distribution to produce a value between 0.0 (inclusive)
-//   // and 1.0 (exclusive).
-//   double value = absl::uniform_real_distribution<double>(0, 1)(gen);
-//
-template <typename RealType = double>
-class uniform_real_distribution {
- public:
-  using result_type = RealType;
-
-  class param_type {
-   public:
-    using distribution_type = uniform_real_distribution;
-
-    explicit param_type(result_type lo = 0, result_type hi = 1)
-        : lo_(lo), hi_(hi), range_(hi - lo) {
-      // [rand.dist.uni.real] preconditions 2 & 3
-      assert(lo <= hi);
-
-      // NOTE: For integral types, we can promote the range to an unsigned type,
-      // which gives full width of the range. However for real (fp) types, this
-      // is not possible, so value generation cannot use the full range of the
-      // real type.
-      assert(range_ <= (std::numeric_limits<result_type>::max)());
-    }
-
-    result_type a() const { return lo_; }
-    result_type b() const { return hi_; }
-
-    friend bool operator==(const param_type& a, const param_type& b) {
-      return a.lo_ == b.lo_ && a.hi_ == b.hi_;
-    }
-
-    friend bool operator!=(const param_type& a, const param_type& b) {
-      return !(a == b);
-    }
-
-   private:
-    friend class uniform_real_distribution;
-    result_type lo_, hi_, range_;
-
-    static_assert(std::is_floating_point<RealType>::value,
-                  "Class-template absl::uniform_real_distribution<> must be "
-                  "parameterized using a floating-point type.");
-  };
-
-  uniform_real_distribution() : uniform_real_distribution(0) {}
-
-  explicit uniform_real_distribution(result_type lo, result_type hi = 1)
-      : param_(lo, hi) {}
-
-  explicit uniform_real_distribution(const param_type& param) : param_(param) {}
-
-  // uniform_real_distribution<T>::reset()
-  //
-  // Resets the uniform real distribution. Note that this function has no effect
-  // because the distribution already produces independent values.
-  void reset() {}
-
-  template <typename URBG>
-  result_type operator()(URBG& gen) {  // NOLINT(runtime/references)
-    return operator()(gen, param_);
-  }
-
-  template <typename URBG>
-  result_type operator()(URBG& gen,  // NOLINT(runtime/references)
-                         const param_type& p);
-
-  result_type a() const { return param_.a(); }
-  result_type b() const { return param_.b(); }
-
-  param_type param() const { return param_; }
-  void param(const param_type& params) { param_ = params; }
-
-  result_type(min)() const { return a(); }
-  result_type(max)() const { return b(); }
-
-  friend bool operator==(const uniform_real_distribution& a,
-                         const uniform_real_distribution& b) {
-    return a.param_ == b.param_;
-  }
-  friend bool operator!=(const uniform_real_distribution& a,
-                         const uniform_real_distribution& b) {
-    return a.param_ != b.param_;
-  }
-
- private:
-  param_type param_;
-  random_internal::FastUniformBits<uint64_t> fast_u64_;
-};
-
-// -----------------------------------------------------------------------------
-// Implementation details follow
-// -----------------------------------------------------------------------------
-template <typename RealType>
-template <typename URBG>
-typename uniform_real_distribution<RealType>::result_type
-uniform_real_distribution<RealType>::operator()(
-    URBG& gen, const param_type& p) {  // NOLINT(runtime/references)
-  using random_internal::GeneratePositiveTag;
-  using random_internal::GenerateRealFromBits;
-  using real_type =
-      absl::conditional_t<std::is_same<RealType, float>::value, float, double>;
-
-  while (true) {
-    const result_type sample =
-        GenerateRealFromBits<real_type, GeneratePositiveTag, true>(
-            fast_u64_(gen));
-    const result_type res = p.a() + (sample * p.range_);
-    if (res < p.b() || p.range_ <= 0 || !std::isfinite(p.range_)) {
-      return res;
-    }
-    // else sample rejected, try again.
-  }
-}
-
-template <typename CharT, typename Traits, typename RealType>
-std::basic_ostream<CharT, Traits>& operator<<(
-    std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
-    const uniform_real_distribution<RealType>& x) {
-  auto saver = random_internal::make_ostream_state_saver(os);
-  os.precision(random_internal::stream_precision_helper<RealType>::kPrecision);
-  os << x.a() << os.fill() << x.b();
-  return os;
-}
-
-template <typename CharT, typename Traits, typename RealType>
-std::basic_istream<CharT, Traits>& operator>>(
-    std::basic_istream<CharT, Traits>& is,     // NOLINT(runtime/references)
-    uniform_real_distribution<RealType>& x) {  // NOLINT(runtime/references)
-  using param_type = typename uniform_real_distribution<RealType>::param_type;
-  using result_type = typename uniform_real_distribution<RealType>::result_type;
-  auto saver = random_internal::make_istream_state_saver(is);
-  auto a = random_internal::read_floating_point<result_type>(is);
-  if (is.fail()) return is;
-  auto b = random_internal::read_floating_point<result_type>(is);
-  if (!is.fail()) {
-    x.param(param_type(a, b));
-  }
-  return is;
-}
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_UNIFORM_REAL_DISTRIBUTION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/random/zipf_distribution.h b/projects/llm_framework/include/abseil-cpp/absl/random/zipf_distribution.h
deleted file mode 100644
index 03497b1b..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/random/zipf_distribution.h
+++ /dev/null
@@ -1,272 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_ZIPF_DISTRIBUTION_H_
-#define ABSL_RANDOM_ZIPF_DISTRIBUTION_H_
-
-#include <cassert>
-#include <cmath>
-#include <istream>
-#include <limits>
-#include <ostream>
-#include <type_traits>
-
-#include "absl/random/internal/iostream_state_saver.h"
-#include "absl/random/internal/traits.h"
-#include "absl/random/uniform_real_distribution.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// absl::zipf_distribution produces random integer-values in the range [0, k],
-// distributed according to the unnormalized discrete probability function:
-//
-//  P(x) = (v + x) ^ -q
-//
-// The parameter `v` must be greater than 0 and the parameter `q` must be
-// greater than 1. If either of these parameters take invalid values then the
-// behavior is undefined.
-//
-// IntType is the result_type generated by the generator. It must be of integral
-// type; a static_assert ensures this is the case.
-//
-// The implementation is based on W.Hormann, G.Derflinger:
-//
-// "Rejection-Inversion to Generate Variates from Monotone Discrete
-// Distributions"
-//
-// http://eeyore.wu-wien.ac.at/papers/96-04-04.wh-der.ps.gz
-//
-template <typename IntType = int>
-class zipf_distribution {
- public:
-  using result_type = IntType;
-
-  class param_type {
-   public:
-    using distribution_type = zipf_distribution;
-
-    // Preconditions: k > 0, v > 0, q > 1
-    // The precondidtions are validated when NDEBUG is not defined via
-    // a pair of assert() directives.
-    // If NDEBUG is defined and either or both of these parameters take invalid
-    // values, the behavior of the class is undefined.
-    explicit param_type(result_type k = (std::numeric_limits<IntType>::max)(),
-                        double q = 2.0, double v = 1.0);
-
-    result_type k() const { return k_; }
-    double q() const { return q_; }
-    double v() const { return v_; }
-
-    friend bool operator==(const param_type& a, const param_type& b) {
-      return a.k_ == b.k_ && a.q_ == b.q_ && a.v_ == b.v_;
-    }
-    friend bool operator!=(const param_type& a, const param_type& b) {
-      return !(a == b);
-    }
-
-   private:
-    friend class zipf_distribution;
-    inline double h(double x) const;
-    inline double hinv(double x) const;
-    inline double compute_s() const;
-    inline double pow_negative_q(double x) const;
-
-    // Parameters here are exactly the same as the parameters of Algorithm ZRI
-    // in the paper.
-    IntType k_;
-    double q_;
-    double v_;
-
-    double one_minus_q_;  // 1-q
-    double s_;
-    double one_minus_q_inv_;  // 1 / 1-q
-    double hxm_;              // h(k + 0.5)
-    double hx0_minus_hxm_;    // h(x0) - h(k + 0.5)
-
-    static_assert(random_internal::IsIntegral<IntType>::value,
-                  "Class-template absl::zipf_distribution<> must be "
-                  "parameterized using an integral type.");
-  };
-
-  zipf_distribution()
-      : zipf_distribution((std::numeric_limits<IntType>::max)()) {}
-
-  explicit zipf_distribution(result_type k, double q = 2.0, double v = 1.0)
-      : param_(k, q, v) {}
-
-  explicit zipf_distribution(const param_type& p) : param_(p) {}
-
-  void reset() {}
-
-  template <typename URBG>
-  result_type operator()(URBG& g) {  // NOLINT(runtime/references)
-    return (*this)(g, param_);
-  }
-
-  template <typename URBG>
-  result_type operator()(URBG& g,  // NOLINT(runtime/references)
-                         const param_type& p);
-
-  result_type k() const { return param_.k(); }
-  double q() const { return param_.q(); }
-  double v() const { return param_.v(); }
-
-  param_type param() const { return param_; }
-  void param(const param_type& p) { param_ = p; }
-
-  result_type(min)() const { return 0; }
-  result_type(max)() const { return k(); }
-
-  friend bool operator==(const zipf_distribution& a,
-                         const zipf_distribution& b) {
-    return a.param_ == b.param_;
-  }
-  friend bool operator!=(const zipf_distribution& a,
-                         const zipf_distribution& b) {
-    return a.param_ != b.param_;
-  }
-
- private:
-  param_type param_;
-};
-
-// --------------------------------------------------------------------------
-// Implementation details follow
-// --------------------------------------------------------------------------
-
-template <typename IntType>
-zipf_distribution<IntType>::param_type::param_type(
-    typename zipf_distribution<IntType>::result_type k, double q, double v)
-    : k_(k), q_(q), v_(v), one_minus_q_(1 - q) {
-  assert(q > 1);
-  assert(v > 0);
-  assert(k > 0);
-  one_minus_q_inv_ = 1 / one_minus_q_;
-
-  // Setup for the ZRI algorithm (pg 17 of the paper).
-  // Compute: h(i max) => h(k + 0.5)
-  constexpr double kMax = 18446744073709549568.0;
-  double kd = static_cast<double>(k);
-  // TODO(absl-team): Determine if this check is needed, and if so, add a test
-  // that fails for k > kMax
-  if (kd > kMax) {
-    // Ensure that our maximum value is capped to a value which will
-    // round-trip back through double.
-    kd = kMax;
-  }
-  hxm_ = h(kd + 0.5);
-
-  // Compute: h(0)
-  const bool use_precomputed = (v == 1.0 && q == 2.0);
-  const double h0x5 = use_precomputed ? (-1.0 / 1.5)  // exp(-log(1.5))
-                                      : h(0.5);
-  const double elogv_q = (v_ == 1.0) ? 1 : pow_negative_q(v_);
-
-  // h(0) = h(0.5) - exp(log(v) * -q)
-  hx0_minus_hxm_ = (h0x5 - elogv_q) - hxm_;
-
-  // And s
-  s_ = use_precomputed ? 0.46153846153846123 : compute_s();
-}
-
-template <typename IntType>
-double zipf_distribution<IntType>::param_type::h(double x) const {
-  // std::exp(one_minus_q_ * std::log(v_ + x)) * one_minus_q_inv_;
-  x += v_;
-  return (one_minus_q_ == -1.0)
-             ? (-1.0 / x)  // -exp(-log(x))
-             : (std::exp(std::log(x) * one_minus_q_) * one_minus_q_inv_);
-}
-
-template <typename IntType>
-double zipf_distribution<IntType>::param_type::hinv(double x) const {
-  // std::exp(one_minus_q_inv_ * std::log(one_minus_q_ * x)) - v_;
-  return -v_ + ((one_minus_q_ == -1.0)
-                    ? (-1.0 / x)  // exp(-log(-x))
-                    : std::exp(one_minus_q_inv_ * std::log(one_minus_q_ * x)));
-}
-
-template <typename IntType>
-double zipf_distribution<IntType>::param_type::compute_s() const {
-  // 1 - hinv(h(1.5) - std::exp(std::log(v_ + 1) * -q_));
-  return 1.0 - hinv(h(1.5) - pow_negative_q(v_ + 1.0));
-}
-
-template <typename IntType>
-double zipf_distribution<IntType>::param_type::pow_negative_q(double x) const {
-  // std::exp(std::log(x) * -q_);
-  return q_ == 2.0 ? (1.0 / (x * x)) : std::exp(std::log(x) * -q_);
-}
-
-template <typename IntType>
-template <typename URBG>
-typename zipf_distribution<IntType>::result_type
-zipf_distribution<IntType>::operator()(
-    URBG& g, const param_type& p) {  // NOLINT(runtime/references)
-  absl::uniform_real_distribution<double> uniform_double;
-  double k;
-  for (;;) {
-    const double v = uniform_double(g);
-    const double u = p.hxm_ + v * p.hx0_minus_hxm_;
-    const double x = p.hinv(u);
-    k = rint(x);              // std::floor(x + 0.5);
-    if (k > static_cast<double>(p.k())) continue;  // reject k > max_k
-    if (k - x <= p.s_) break;
-    const double h = p.h(k + 0.5);
-    const double r = p.pow_negative_q(p.v_ + k);
-    if (u >= h - r) break;
-  }
-  IntType ki = static_cast<IntType>(k);
-  assert(ki <= p.k_);
-  return ki;
-}
-
-template <typename CharT, typename Traits, typename IntType>
-std::basic_ostream<CharT, Traits>& operator<<(
-    std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
-    const zipf_distribution<IntType>& x) {
-  using stream_type =
-      typename random_internal::stream_format_type<IntType>::type;
-  auto saver = random_internal::make_ostream_state_saver(os);
-  os.precision(random_internal::stream_precision_helper<double>::kPrecision);
-  os << static_cast<stream_type>(x.k()) << os.fill() << x.q() << os.fill()
-     << x.v();
-  return os;
-}
-
-template <typename CharT, typename Traits, typename IntType>
-std::basic_istream<CharT, Traits>& operator>>(
-    std::basic_istream<CharT, Traits>& is,  // NOLINT(runtime/references)
-    zipf_distribution<IntType>& x) {        // NOLINT(runtime/references)
-  using result_type = typename zipf_distribution<IntType>::result_type;
-  using param_type = typename zipf_distribution<IntType>::param_type;
-  using stream_type =
-      typename random_internal::stream_format_type<IntType>::type;
-  stream_type k;
-  double q;
-  double v;
-
-  auto saver = random_internal::make_istream_state_saver(is);
-  is >> k >> q >> v;
-  if (!is.fail()) {
-    x.param(param_type(static_cast<result_type>(k), q, v));
-  }
-  return is;
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_ZIPF_DISTRIBUTION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/status/internal/status_internal.h b/projects/llm_framework/include/abseil-cpp/absl/status/internal/status_internal.h
deleted file mode 100644
index c9f43832..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/status/internal/status_internal.h
+++ /dev/null
@@ -1,132 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-#ifndef ABSL_STATUS_INTERNAL_STATUS_INTERNAL_H_
-#define ABSL_STATUS_INTERNAL_STATUS_INTERNAL_H_
-
-#include <atomic>
-#include <cstdint>
-#include <memory>
-#include <string>
-#include <utility>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/nullability.h"
-#include "absl/container/inlined_vector.h"
-#include "absl/strings/cord.h"
-#include "absl/strings/string_view.h"
-#include "absl/types/optional.h"
-
-#ifndef SWIG
-// Disabled for SWIG as it doesn't parse attributes correctly.
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-// Returned Status objects may not be ignored. Codesearch doesn't handle ifdefs
-// as part of a class definitions (b/6995610), so we use a forward declaration.
-//
-// TODO(b/176172494): ABSL_MUST_USE_RESULT should expand to the more strict
-// [[nodiscard]]. For now, just use [[nodiscard]] directly when it is available.
-#if ABSL_HAVE_CPP_ATTRIBUTE(nodiscard)
-class [[nodiscard]] ABSL_ATTRIBUTE_TRIVIAL_ABI Status;
-#else
-class ABSL_MUST_USE_RESULT ABSL_ATTRIBUTE_TRIVIAL_ABI Status;
-#endif
-ABSL_NAMESPACE_END
-}  // namespace absl
-#endif  // !SWIG
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-enum class StatusCode : int;
-enum class StatusToStringMode : int;
-
-namespace status_internal {
-
-// Container for status payloads.
-struct Payload {
-  std::string type_url;
-  absl::Cord payload;
-};
-
-using Payloads = absl::InlinedVector<Payload, 1>;
-
-// Reference-counted representation of Status data.
-class StatusRep {
- public:
-  StatusRep(absl::StatusCode code_arg, absl::string_view message_arg,
-            std::unique_ptr<status_internal::Payloads> payloads_arg)
-      : ref_(int32_t{1}),
-        code_(code_arg),
-        message_(message_arg),
-        payloads_(std::move(payloads_arg)) {}
-
-  absl::StatusCode code() const { return code_; }
-  const std::string& message() const { return message_; }
-
-  // Ref and unref are const to allow access through a const pointer, and are
-  // used during copying operations.
-  void Ref() const { ref_.fetch_add(1, std::memory_order_relaxed); }
-  void Unref() const;
-
-  // Payload methods correspond to the same methods in absl::Status.
-  absl::optional<absl::Cord> GetPayload(absl::string_view type_url) const;
-  void SetPayload(absl::string_view type_url, absl::Cord payload);
-  struct EraseResult {
-    bool erased;
-    uintptr_t new_rep;
-  };
-  EraseResult ErasePayload(absl::string_view type_url);
-  void ForEachPayload(
-      absl::FunctionRef<void(absl::string_view, const absl::Cord&)> visitor)
-      const;
-
-  std::string ToString(StatusToStringMode mode) const;
-
-  bool operator==(const StatusRep& other) const;
-  bool operator!=(const StatusRep& other) const { return !(*this == other); }
-
-  // Returns an equivalent heap allocated StatusRep with refcount 1.
-  //
-  // `this` is not safe to be used after calling as it may have been deleted.
-  absl::Nonnull<StatusRep*> CloneAndUnref() const;
-
- private:
-  mutable std::atomic<int32_t> ref_;
-  absl::StatusCode code_;
-
-  // As an internal implementation detail, we guarantee that if status.message()
-  // is non-empty, then the resulting string_view is null terminated.
-  // This is required to implement 'StatusMessageAsCStr(...)'
-  std::string message_;
-  std::unique_ptr<status_internal::Payloads> payloads_;
-};
-
-absl::StatusCode MapToLocalCode(int value);
-
-// Returns a pointer to a newly-allocated string with the given `prefix`,
-// suitable for output as an error message in assertion/`CHECK()` failures.
-//
-// This is an internal implementation detail for Abseil logging.
-ABSL_ATTRIBUTE_PURE_FUNCTION
-absl::Nonnull<std::string*> MakeCheckFailString(
-    absl::Nonnull<const absl::Status*> status,
-    absl::Nonnull<const char*> prefix);
-
-}  // namespace status_internal
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STATUS_INTERNAL_STATUS_INTERNAL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/status/internal/statusor_internal.h b/projects/llm_framework/include/abseil-cpp/absl/status/internal/statusor_internal.h
deleted file mode 100644
index 5be94903..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/status/internal/statusor_internal.h
+++ /dev/null
@@ -1,435 +0,0 @@
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-#ifndef ABSL_STATUS_INTERNAL_STATUSOR_INTERNAL_H_
-#define ABSL_STATUS_INTERNAL_STATUSOR_INTERNAL_H_
-
-#include <cstdint>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/attributes.h"
-#include "absl/base/nullability.h"
-#include "absl/meta/type_traits.h"
-#include "absl/status/status.h"
-#include "absl/strings/string_view.h"
-#include "absl/utility/utility.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-template <typename T>
-class ABSL_MUST_USE_RESULT StatusOr;
-
-namespace internal_statusor {
-
-// Detects whether `U` has conversion operator to `StatusOr<T>`, i.e. `operator
-// StatusOr<T>()`.
-template <typename T, typename U, typename = void>
-struct HasConversionOperatorToStatusOr : std::false_type {};
-
-template <typename T, typename U>
-void test(char (*)[sizeof(std::declval<U>().operator absl::StatusOr<T>())]);
-
-template <typename T, typename U>
-struct HasConversionOperatorToStatusOr<T, U, decltype(test<T, U>(0))>
-    : std::true_type {};
-
-// Detects whether `T` is constructible or convertible from `StatusOr<U>`.
-template <typename T, typename U>
-using IsConstructibleOrConvertibleFromStatusOr =
-    absl::disjunction<std::is_constructible<T, StatusOr<U>&>,
-                      std::is_constructible<T, const StatusOr<U>&>,
-                      std::is_constructible<T, StatusOr<U>&&>,
-                      std::is_constructible<T, const StatusOr<U>&&>,
-                      std::is_convertible<StatusOr<U>&, T>,
-                      std::is_convertible<const StatusOr<U>&, T>,
-                      std::is_convertible<StatusOr<U>&&, T>,
-                      std::is_convertible<const StatusOr<U>&&, T>>;
-
-// Detects whether `T` is constructible or convertible or assignable from
-// `StatusOr<U>`.
-template <typename T, typename U>
-using IsConstructibleOrConvertibleOrAssignableFromStatusOr =
-    absl::disjunction<IsConstructibleOrConvertibleFromStatusOr<T, U>,
-                      std::is_assignable<T&, StatusOr<U>&>,
-                      std::is_assignable<T&, const StatusOr<U>&>,
-                      std::is_assignable<T&, StatusOr<U>&&>,
-                      std::is_assignable<T&, const StatusOr<U>&&>>;
-
-// Detects whether direct initializing `StatusOr<T>` from `U` is ambiguous, i.e.
-// when `U` is `StatusOr<V>` and `T` is constructible or convertible from `V`.
-template <typename T, typename U>
-struct IsDirectInitializationAmbiguous
-    : public absl::conditional_t<
-          std::is_same<absl::remove_cvref_t<U>, U>::value, std::false_type,
-          IsDirectInitializationAmbiguous<T, absl::remove_cvref_t<U>>> {};
-
-template <typename T, typename V>
-struct IsDirectInitializationAmbiguous<T, absl::StatusOr<V>>
-    : public IsConstructibleOrConvertibleFromStatusOr<T, V> {};
-
-// Checks against the constraints of the direction initialization, i.e. when
-// `StatusOr<T>::StatusOr(U&&)` should participate in overload resolution.
-template <typename T, typename U>
-using IsDirectInitializationValid = absl::disjunction<
-    // Short circuits if T is basically U.
-    std::is_same<T, absl::remove_cvref_t<U>>,
-    absl::negation<absl::disjunction<
-        std::is_same<absl::StatusOr<T>, absl::remove_cvref_t<U>>,
-        std::is_same<absl::Status, absl::remove_cvref_t<U>>,
-        std::is_same<absl::in_place_t, absl::remove_cvref_t<U>>,
-        IsDirectInitializationAmbiguous<T, U>>>>;
-
-// This trait detects whether `StatusOr<T>::operator=(U&&)` is ambiguous, which
-// is equivalent to whether all the following conditions are met:
-// 1. `U` is `StatusOr<V>`.
-// 2. `T` is constructible and assignable from `V`.
-// 3. `T` is constructible and assignable from `U` (i.e. `StatusOr<V>`).
-// For example, the following code is considered ambiguous:
-// (`T` is `bool`, `U` is `StatusOr<bool>`, `V` is `bool`)
-//   StatusOr<bool> s1 = true;  // s1.ok() && s1.ValueOrDie() == true
-//   StatusOr<bool> s2 = false;  // s2.ok() && s2.ValueOrDie() == false
-//   s1 = s2;  // ambiguous, `s1 = s2.ValueOrDie()` or `s1 = bool(s2)`?
-template <typename T, typename U>
-struct IsForwardingAssignmentAmbiguous
-    : public absl::conditional_t<
-          std::is_same<absl::remove_cvref_t<U>, U>::value, std::false_type,
-          IsForwardingAssignmentAmbiguous<T, absl::remove_cvref_t<U>>> {};
-
-template <typename T, typename U>
-struct IsForwardingAssignmentAmbiguous<T, absl::StatusOr<U>>
-    : public IsConstructibleOrConvertibleOrAssignableFromStatusOr<T, U> {};
-
-// Checks against the constraints of the forwarding assignment, i.e. whether
-// `StatusOr<T>::operator(U&&)` should participate in overload resolution.
-template <typename T, typename U>
-using IsForwardingAssignmentValid = absl::disjunction<
-    // Short circuits if T is basically U.
-    std::is_same<T, absl::remove_cvref_t<U>>,
-    absl::negation<absl::disjunction<
-        std::is_same<absl::StatusOr<T>, absl::remove_cvref_t<U>>,
-        std::is_same<absl::Status, absl::remove_cvref_t<U>>,
-        std::is_same<absl::in_place_t, absl::remove_cvref_t<U>>,
-        IsForwardingAssignmentAmbiguous<T, U>>>>;
-
-class Helper {
- public:
-  // Move type-agnostic error handling to the .cc.
-  static void HandleInvalidStatusCtorArg(absl::Nonnull<Status*>);
-  ABSL_ATTRIBUTE_NORETURN static void Crash(const absl::Status& status);
-};
-
-// Construct an instance of T in `p` through placement new, passing Args... to
-// the constructor.
-// This abstraction is here mostly for the gcc performance fix.
-template <typename T, typename... Args>
-ABSL_ATTRIBUTE_NONNULL(1)
-void PlacementNew(absl::Nonnull<void*> p, Args&&... args) {
-  new (p) T(std::forward<Args>(args)...);
-}
-
-// Helper base class to hold the data and all operations.
-// We move all this to a base class to allow mixing with the appropriate
-// TraitsBase specialization.
-template <typename T>
-class StatusOrData {
-  template <typename U>
-  friend class StatusOrData;
-
- public:
-  StatusOrData() = delete;
-
-  StatusOrData(const StatusOrData& other) {
-    if (other.ok()) {
-      MakeValue(other.data_);
-      MakeStatus();
-    } else {
-      MakeStatus(other.status_);
-    }
-  }
-
-  StatusOrData(StatusOrData&& other) noexcept {
-    if (other.ok()) {
-      MakeValue(std::move(other.data_));
-      MakeStatus();
-    } else {
-      MakeStatus(std::move(other.status_));
-    }
-  }
-
-  template <typename U>
-  explicit StatusOrData(const StatusOrData<U>& other) {
-    if (other.ok()) {
-      MakeValue(other.data_);
-      MakeStatus();
-    } else {
-      MakeStatus(other.status_);
-    }
-  }
-
-  template <typename U>
-  explicit StatusOrData(StatusOrData<U>&& other) {
-    if (other.ok()) {
-      MakeValue(std::move(other.data_));
-      MakeStatus();
-    } else {
-      MakeStatus(std::move(other.status_));
-    }
-  }
-
-  template <typename... Args>
-  explicit StatusOrData(absl::in_place_t, Args&&... args)
-      : data_(std::forward<Args>(args)...) {
-    MakeStatus();
-  }
-
-  explicit StatusOrData(const T& value) : data_(value) {
-    MakeStatus();
-  }
-  explicit StatusOrData(T&& value) : data_(std::move(value)) {
-    MakeStatus();
-  }
-
-  template <typename U,
-            absl::enable_if_t<std::is_constructible<absl::Status, U&&>::value,
-                              int> = 0>
-  explicit StatusOrData(U&& v) : status_(std::forward<U>(v)) {
-    EnsureNotOk();
-  }
-
-  StatusOrData& operator=(const StatusOrData& other) {
-    if (this == &other) return *this;
-    if (other.ok())
-      Assign(other.data_);
-    else
-      AssignStatus(other.status_);
-    return *this;
-  }
-
-  StatusOrData& operator=(StatusOrData&& other) {
-    if (this == &other) return *this;
-    if (other.ok())
-      Assign(std::move(other.data_));
-    else
-      AssignStatus(std::move(other.status_));
-    return *this;
-  }
-
-  ~StatusOrData() {
-    if (ok()) {
-      status_.~Status();
-      data_.~T();
-    } else {
-      status_.~Status();
-    }
-  }
-
-  template <typename U>
-  void Assign(U&& value) {
-    if (ok()) {
-      data_ = std::forward<U>(value);
-    } else {
-      MakeValue(std::forward<U>(value));
-      status_ = OkStatus();
-    }
-  }
-
-  template <typename U>
-  void AssignStatus(U&& v) {
-    Clear();
-    status_ = static_cast<absl::Status>(std::forward<U>(v));
-    EnsureNotOk();
-  }
-
-  bool ok() const { return status_.ok(); }
-
- protected:
-  // status_ will always be active after the constructor.
-  // We make it a union to be able to initialize exactly how we need without
-  // waste.
-  // Eg. in the copy constructor we use the default constructor of Status in
-  // the ok() path to avoid an extra Ref call.
-  union {
-    Status status_;
-  };
-
-  // data_ is active iff status_.ok()==true
-  struct Dummy {};
-  union {
-    // When T is const, we need some non-const object we can cast to void* for
-    // the placement new. dummy_ is that object.
-    Dummy dummy_;
-    T data_;
-  };
-
-  void Clear() {
-    if (ok()) data_.~T();
-  }
-
-  void EnsureOk() const {
-    if (ABSL_PREDICT_FALSE(!ok())) Helper::Crash(status_);
-  }
-
-  void EnsureNotOk() {
-    if (ABSL_PREDICT_FALSE(ok())) Helper::HandleInvalidStatusCtorArg(&status_);
-  }
-
-  // Construct the value (ie. data_) through placement new with the passed
-  // argument.
-  template <typename... Arg>
-  void MakeValue(Arg&&... arg) {
-    internal_statusor::PlacementNew<T>(&dummy_, std::forward<Arg>(arg)...);
-  }
-
-  // Construct the status (ie. status_) through placement new with the passed
-  // argument.
-  template <typename... Args>
-  void MakeStatus(Args&&... args) {
-    internal_statusor::PlacementNew<Status>(&status_,
-                                            std::forward<Args>(args)...);
-  }
-};
-
-// Helper base classes to allow implicitly deleted constructors and assignment
-// operators in `StatusOr`. For example, `CopyCtorBase` will explicitly delete
-// the copy constructor when T is not copy constructible and `StatusOr` will
-// inherit that behavior implicitly.
-template <typename T, bool = std::is_copy_constructible<T>::value>
-struct CopyCtorBase {
-  CopyCtorBase() = default;
-  CopyCtorBase(const CopyCtorBase&) = default;
-  CopyCtorBase(CopyCtorBase&&) = default;
-  CopyCtorBase& operator=(const CopyCtorBase&) = default;
-  CopyCtorBase& operator=(CopyCtorBase&&) = default;
-};
-
-template <typename T>
-struct CopyCtorBase<T, false> {
-  CopyCtorBase() = default;
-  CopyCtorBase(const CopyCtorBase&) = delete;
-  CopyCtorBase(CopyCtorBase&&) = default;
-  CopyCtorBase& operator=(const CopyCtorBase&) = default;
-  CopyCtorBase& operator=(CopyCtorBase&&) = default;
-};
-
-template <typename T, bool = std::is_move_constructible<T>::value>
-struct MoveCtorBase {
-  MoveCtorBase() = default;
-  MoveCtorBase(const MoveCtorBase&) = default;
-  MoveCtorBase(MoveCtorBase&&) = default;
-  MoveCtorBase& operator=(const MoveCtorBase&) = default;
-  MoveCtorBase& operator=(MoveCtorBase&&) = default;
-};
-
-template <typename T>
-struct MoveCtorBase<T, false> {
-  MoveCtorBase() = default;
-  MoveCtorBase(const MoveCtorBase&) = default;
-  MoveCtorBase(MoveCtorBase&&) = delete;
-  MoveCtorBase& operator=(const MoveCtorBase&) = default;
-  MoveCtorBase& operator=(MoveCtorBase&&) = default;
-};
-
-template <typename T, bool = std::is_copy_constructible<T>::value&&
-                          std::is_copy_assignable<T>::value>
-struct CopyAssignBase {
-  CopyAssignBase() = default;
-  CopyAssignBase(const CopyAssignBase&) = default;
-  CopyAssignBase(CopyAssignBase&&) = default;
-  CopyAssignBase& operator=(const CopyAssignBase&) = default;
-  CopyAssignBase& operator=(CopyAssignBase&&) = default;
-};
-
-template <typename T>
-struct CopyAssignBase<T, false> {
-  CopyAssignBase() = default;
-  CopyAssignBase(const CopyAssignBase&) = default;
-  CopyAssignBase(CopyAssignBase&&) = default;
-  CopyAssignBase& operator=(const CopyAssignBase&) = delete;
-  CopyAssignBase& operator=(CopyAssignBase&&) = default;
-};
-
-template <typename T, bool = std::is_move_constructible<T>::value&&
-                          std::is_move_assignable<T>::value>
-struct MoveAssignBase {
-  MoveAssignBase() = default;
-  MoveAssignBase(const MoveAssignBase&) = default;
-  MoveAssignBase(MoveAssignBase&&) = default;
-  MoveAssignBase& operator=(const MoveAssignBase&) = default;
-  MoveAssignBase& operator=(MoveAssignBase&&) = default;
-};
-
-template <typename T>
-struct MoveAssignBase<T, false> {
-  MoveAssignBase() = default;
-  MoveAssignBase(const MoveAssignBase&) = default;
-  MoveAssignBase(MoveAssignBase&&) = default;
-  MoveAssignBase& operator=(const MoveAssignBase&) = default;
-  MoveAssignBase& operator=(MoveAssignBase&&) = delete;
-};
-
-ABSL_ATTRIBUTE_NORETURN void ThrowBadStatusOrAccess(absl::Status status);
-
-// Used to introduce jitter into the output of printing functions for
-// `StatusOr` (i.e. `AbslStringify` and `operator<<`).
-class StringifyRandom {
-  enum BracesType {
-    kBareParens = 0,
-    kSpaceParens,
-    kBareBrackets,
-    kSpaceBrackets,
-  };
-
-  // Returns a random `BracesType` determined once per binary load.
-  static BracesType RandomBraces() {
-    static const BracesType kRandomBraces = static_cast<BracesType>(
-        (reinterpret_cast<uintptr_t>(&kRandomBraces) >> 4) % 4);
-    return kRandomBraces;
-  }
-
- public:
-  static inline absl::string_view OpenBrackets() {
-    switch (RandomBraces()) {
-      case kBareParens:
-        return "(";
-      case kSpaceParens:
-        return "( ";
-      case kBareBrackets:
-        return "[";
-      case kSpaceBrackets:
-        return "[ ";
-    }
-    return "(";
-  }
-
-  static inline absl::string_view CloseBrackets() {
-    switch (RandomBraces()) {
-      case kBareParens:
-        return ")";
-      case kSpaceParens:
-        return " )";
-      case kBareBrackets:
-        return "]";
-      case kSpaceBrackets:
-        return " ]";
-    }
-    return ")";
-  }
-};
-
-}  // namespace internal_statusor
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STATUS_INTERNAL_STATUSOR_INTERNAL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/status/status.h b/projects/llm_framework/include/abseil-cpp/absl/status/status.h
deleted file mode 100644
index 6cfe49f2..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/status/status.h
+++ /dev/null
@@ -1,943 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: status.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines the Abseil `status` library, consisting of:
-//
-//   * An `absl::Status` class for holding error handling information
-//   * A set of canonical `absl::StatusCode` error codes, and associated
-//     utilities for generating and propagating status codes.
-//   * A set of helper functions for creating status codes and checking their
-//     values
-//
-// Within Google, `absl::Status` is the primary mechanism for communicating
-// errors in C++, and is used to represent error state in both in-process
-// library calls as well as RPC calls. Some of these errors may be recoverable,
-// but others may not. Most functions that can produce a recoverable error
-// should be designed to return an `absl::Status` (or `absl::StatusOr`).
-//
-// Example:
-//
-// absl::Status myFunction(absl::string_view fname, ...) {
-//   ...
-//   // encounter error
-//   if (error condition) {
-//     return absl::InvalidArgumentError("bad mode");
-//   }
-//   // else, return OK
-//   return absl::OkStatus();
-// }
-//
-// An `absl::Status` is designed to either return "OK" or one of a number of
-// different error codes, corresponding to typical error conditions.
-// In almost all cases, when using `absl::Status` you should use the canonical
-// error codes (of type `absl::StatusCode`) enumerated in this header file.
-// These canonical codes are understood across the codebase and will be
-// accepted across all API and RPC boundaries.
-#ifndef ABSL_STATUS_STATUS_H_
-#define ABSL_STATUS_STATUS_H_
-
-#include <cassert>
-#include <cstdint>
-#include <ostream>
-#include <string>
-#include <utility>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/macros.h"
-#include "absl/base/nullability.h"
-#include "absl/base/optimization.h"
-#include "absl/functional/function_ref.h"
-#include "absl/status/internal/status_internal.h"
-#include "absl/strings/cord.h"
-#include "absl/strings/string_view.h"
-#include "absl/types/optional.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// absl::StatusCode
-//
-// An `absl::StatusCode` is an enumerated type indicating either no error ("OK")
-// or an error condition. In most cases, an `absl::Status` indicates a
-// recoverable error, and the purpose of signalling an error is to indicate what
-// action to take in response to that error. These error codes map to the proto
-// RPC error codes indicated in https://cloud.google.com/apis/design/errors.
-//
-// The errors listed below are the canonical errors associated with
-// `absl::Status` and are used throughout the codebase. As a result, these
-// error codes are somewhat generic.
-//
-// In general, try to return the most specific error that applies if more than
-// one error may pertain. For example, prefer `kOutOfRange` over
-// `kFailedPrecondition` if both codes apply. Similarly prefer `kNotFound` or
-// `kAlreadyExists` over `kFailedPrecondition`.
-//
-// Because these errors may cross RPC boundaries, these codes are tied to the
-// `google.rpc.Code` definitions within
-// https://github.com/googleapis/googleapis/blob/master/google/rpc/code.proto
-// The string value of these RPC codes is denoted within each enum below.
-//
-// If your error handling code requires more context, you can attach payloads
-// to your status. See `absl::Status::SetPayload()` and
-// `absl::Status::GetPayload()` below.
-enum class StatusCode : int {
-  // StatusCode::kOk
-  //
-  // kOK (gRPC code "OK") does not indicate an error; this value is returned on
-  // success. It is typical to check for this value before proceeding on any
-  // given call across an API or RPC boundary. To check this value, use the
-  // `absl::Status::ok()` member function rather than inspecting the raw code.
-  kOk = 0,
-
-  // StatusCode::kCancelled
-  //
-  // kCancelled (gRPC code "CANCELLED") indicates the operation was cancelled,
-  // typically by the caller.
-  kCancelled = 1,
-
-  // StatusCode::kUnknown
-  //
-  // kUnknown (gRPC code "UNKNOWN") indicates an unknown error occurred. In
-  // general, more specific errors should be raised, if possible. Errors raised
-  // by APIs that do not return enough error information may be converted to
-  // this error.
-  kUnknown = 2,
-
-  // StatusCode::kInvalidArgument
-  //
-  // kInvalidArgument (gRPC code "INVALID_ARGUMENT") indicates the caller
-  // specified an invalid argument, such as a malformed filename. Note that use
-  // of such errors should be narrowly limited to indicate the invalid nature of
-  // the arguments themselves. Errors with validly formed arguments that may
-  // cause errors with the state of the receiving system should be denoted with
-  // `kFailedPrecondition` instead.
-  kInvalidArgument = 3,
-
-  // StatusCode::kDeadlineExceeded
-  //
-  // kDeadlineExceeded (gRPC code "DEADLINE_EXCEEDED") indicates a deadline
-  // expired before the operation could complete. For operations that may change
-  // state within a system, this error may be returned even if the operation has
-  // completed successfully. For example, a successful response from a server
-  // could have been delayed long enough for the deadline to expire.
-  kDeadlineExceeded = 4,
-
-  // StatusCode::kNotFound
-  //
-  // kNotFound (gRPC code "NOT_FOUND") indicates some requested entity (such as
-  // a file or directory) was not found.
-  //
-  // `kNotFound` is useful if a request should be denied for an entire class of
-  // users, such as during a gradual feature rollout or undocumented allow list.
-  // If a request should be denied for specific sets of users, such as through
-  // user-based access control, use `kPermissionDenied` instead.
-  kNotFound = 5,
-
-  // StatusCode::kAlreadyExists
-  //
-  // kAlreadyExists (gRPC code "ALREADY_EXISTS") indicates that the entity a
-  // caller attempted to create (such as a file or directory) is already
-  // present.
-  kAlreadyExists = 6,
-
-  // StatusCode::kPermissionDenied
-  //
-  // kPermissionDenied (gRPC code "PERMISSION_DENIED") indicates that the caller
-  // does not have permission to execute the specified operation. Note that this
-  // error is different than an error due to an *un*authenticated user. This
-  // error code does not imply the request is valid or the requested entity
-  // exists or satisfies any other pre-conditions.
-  //
-  // `kPermissionDenied` must not be used for rejections caused by exhausting
-  // some resource. Instead, use `kResourceExhausted` for those errors.
-  // `kPermissionDenied` must not be used if the caller cannot be identified.
-  // Instead, use `kUnauthenticated` for those errors.
-  kPermissionDenied = 7,
-
-  // StatusCode::kResourceExhausted
-  //
-  // kResourceExhausted (gRPC code "RESOURCE_EXHAUSTED") indicates some resource
-  // has been exhausted, perhaps a per-user quota, or perhaps the entire file
-  // system is out of space.
-  kResourceExhausted = 8,
-
-  // StatusCode::kFailedPrecondition
-  //
-  // kFailedPrecondition (gRPC code "FAILED_PRECONDITION") indicates that the
-  // operation was rejected because the system is not in a state required for
-  // the operation's execution. For example, a directory to be deleted may be
-  // non-empty, an "rmdir" operation is applied to a non-directory, etc.
-  //
-  // Some guidelines that may help a service implementer in deciding between
-  // `kFailedPrecondition`, `kAborted`, and `kUnavailable`:
-  //
-  //  (a) Use `kUnavailable` if the client can retry just the failing call.
-  //  (b) Use `kAborted` if the client should retry at a higher transaction
-  //      level (such as when a client-specified test-and-set fails, indicating
-  //      the client should restart a read-modify-write sequence).
-  //  (c) Use `kFailedPrecondition` if the client should not retry until
-  //      the system state has been explicitly fixed. For example, if a "rmdir"
-  //      fails because the directory is non-empty, `kFailedPrecondition`
-  //      should be returned since the client should not retry unless
-  //      the files are deleted from the directory.
-  kFailedPrecondition = 9,
-
-  // StatusCode::kAborted
-  //
-  // kAborted (gRPC code "ABORTED") indicates the operation was aborted,
-  // typically due to a concurrency issue such as a sequencer check failure or a
-  // failed transaction.
-  //
-  // See the guidelines above for deciding between `kFailedPrecondition`,
-  // `kAborted`, and `kUnavailable`.
-  kAborted = 10,
-
-  // StatusCode::kOutOfRange
-  //
-  // kOutOfRange (gRPC code "OUT_OF_RANGE") indicates the operation was
-  // attempted past the valid range, such as seeking or reading past an
-  // end-of-file.
-  //
-  // Unlike `kInvalidArgument`, this error indicates a problem that may
-  // be fixed if the system state changes. For example, a 32-bit file
-  // system will generate `kInvalidArgument` if asked to read at an
-  // offset that is not in the range [0,2^32-1], but it will generate
-  // `kOutOfRange` if asked to read from an offset past the current
-  // file size.
-  //
-  // There is a fair bit of overlap between `kFailedPrecondition` and
-  // `kOutOfRange`.  We recommend using `kOutOfRange` (the more specific
-  // error) when it applies so that callers who are iterating through
-  // a space can easily look for an `kOutOfRange` error to detect when
-  // they are done.
-  kOutOfRange = 11,
-
-  // StatusCode::kUnimplemented
-  //
-  // kUnimplemented (gRPC code "UNIMPLEMENTED") indicates the operation is not
-  // implemented or supported in this service. In this case, the operation
-  // should not be re-attempted.
-  kUnimplemented = 12,
-
-  // StatusCode::kInternal
-  //
-  // kInternal (gRPC code "INTERNAL") indicates an internal error has occurred
-  // and some invariants expected by the underlying system have not been
-  // satisfied. This error code is reserved for serious errors.
-  kInternal = 13,
-
-  // StatusCode::kUnavailable
-  //
-  // kUnavailable (gRPC code "UNAVAILABLE") indicates the service is currently
-  // unavailable and that this is most likely a transient condition. An error
-  // such as this can be corrected by retrying with a backoff scheme. Note that
-  // it is not always safe to retry non-idempotent operations.
-  //
-  // See the guidelines above for deciding between `kFailedPrecondition`,
-  // `kAborted`, and `kUnavailable`.
-  kUnavailable = 14,
-
-  // StatusCode::kDataLoss
-  //
-  // kDataLoss (gRPC code "DATA_LOSS") indicates that unrecoverable data loss or
-  // corruption has occurred. As this error is serious, proper alerting should
-  // be attached to errors such as this.
-  kDataLoss = 15,
-
-  // StatusCode::kUnauthenticated
-  //
-  // kUnauthenticated (gRPC code "UNAUTHENTICATED") indicates that the request
-  // does not have valid authentication credentials for the operation. Correct
-  // the authentication and try again.
-  kUnauthenticated = 16,
-
-  // StatusCode::DoNotUseReservedForFutureExpansionUseDefaultInSwitchInstead_
-  //
-  // NOTE: this error code entry should not be used and you should not rely on
-  // its value, which may change.
-  //
-  // The purpose of this enumerated value is to force people who handle status
-  // codes with `switch()` statements to *not* simply enumerate all possible
-  // values, but instead provide a "default:" case. Providing such a default
-  // case ensures that code will compile when new codes are added.
-  kDoNotUseReservedForFutureExpansionUseDefaultInSwitchInstead_ = 20
-};
-
-// StatusCodeToString()
-//
-// Returns the name for the status code, or "" if it is an unknown value.
-std::string StatusCodeToString(StatusCode code);
-
-// operator<<
-//
-// Streams StatusCodeToString(code) to `os`.
-std::ostream& operator<<(std::ostream& os, StatusCode code);
-
-// absl::StatusToStringMode
-//
-// An `absl::StatusToStringMode` is an enumerated type indicating how
-// `absl::Status::ToString()` should construct the output string for a non-ok
-// status.
-enum class StatusToStringMode : int {
-  // ToString will not contain any extra data (such as payloads). It will only
-  // contain the error code and message, if any.
-  kWithNoExtraData = 0,
-  // ToString will contain the payloads.
-  kWithPayload = 1 << 0,
-  // ToString will include all the extra data this Status has.
-  kWithEverything = ~kWithNoExtraData,
-  // Default mode used by ToString. Its exact value might change in the future.
-  kDefault = kWithPayload,
-};
-
-// absl::StatusToStringMode is specified as a bitmask type, which means the
-// following operations must be provided:
-inline constexpr StatusToStringMode operator&(StatusToStringMode lhs,
-                                              StatusToStringMode rhs) {
-  return static_cast<StatusToStringMode>(static_cast<int>(lhs) &
-                                         static_cast<int>(rhs));
-}
-inline constexpr StatusToStringMode operator|(StatusToStringMode lhs,
-                                              StatusToStringMode rhs) {
-  return static_cast<StatusToStringMode>(static_cast<int>(lhs) |
-                                         static_cast<int>(rhs));
-}
-inline constexpr StatusToStringMode operator^(StatusToStringMode lhs,
-                                              StatusToStringMode rhs) {
-  return static_cast<StatusToStringMode>(static_cast<int>(lhs) ^
-                                         static_cast<int>(rhs));
-}
-inline constexpr StatusToStringMode operator~(StatusToStringMode arg) {
-  return static_cast<StatusToStringMode>(~static_cast<int>(arg));
-}
-inline StatusToStringMode& operator&=(StatusToStringMode& lhs,
-                                      StatusToStringMode rhs) {
-  lhs = lhs & rhs;
-  return lhs;
-}
-inline StatusToStringMode& operator|=(StatusToStringMode& lhs,
-                                      StatusToStringMode rhs) {
-  lhs = lhs | rhs;
-  return lhs;
-}
-inline StatusToStringMode& operator^=(StatusToStringMode& lhs,
-                                      StatusToStringMode rhs) {
-  lhs = lhs ^ rhs;
-  return lhs;
-}
-
-// absl::Status
-//
-// The `absl::Status` class is generally used to gracefully handle errors
-// across API boundaries (and in particular across RPC boundaries). Some of
-// these errors may be recoverable, but others may not. Most
-// functions which can produce a recoverable error should be designed to return
-// either an `absl::Status` (or the similar `absl::StatusOr<T>`, which holds
-// either an object of type `T` or an error).
-//
-// API developers should construct their functions to return `absl::OkStatus()`
-// upon success, or an `absl::StatusCode` upon another type of error (e.g
-// an `absl::StatusCode::kInvalidArgument` error). The API provides convenience
-// functions to construct each status code.
-//
-// Example:
-//
-// absl::Status myFunction(absl::string_view fname, ...) {
-//   ...
-//   // encounter error
-//   if (error condition) {
-//     // Construct an absl::StatusCode::kInvalidArgument error
-//     return absl::InvalidArgumentError("bad mode");
-//   }
-//   // else, return OK
-//   return absl::OkStatus();
-// }
-//
-// Users handling status error codes should prefer checking for an OK status
-// using the `ok()` member function. Handling multiple error codes may justify
-// use of switch statement, but only check for error codes you know how to
-// handle; do not try to exhaustively match against all canonical error codes.
-// Errors that cannot be handled should be logged and/or propagated for higher
-// levels to deal with. If you do use a switch statement, make sure that you
-// also provide a `default:` switch case, so that code does not break as other
-// canonical codes are added to the API.
-//
-// Example:
-//
-//   absl::Status result = DoSomething();
-//   if (!result.ok()) {
-//     LOG(ERROR) << result;
-//   }
-//
-//   // Provide a default if switching on multiple error codes
-//   switch (result.code()) {
-//     // The user hasn't authenticated. Ask them to reauth
-//     case absl::StatusCode::kUnauthenticated:
-//       DoReAuth();
-//       break;
-//     // The user does not have permission. Log an error.
-//     case absl::StatusCode::kPermissionDenied:
-//       LOG(ERROR) << result;
-//       break;
-//     // Propagate the error otherwise.
-//     default:
-//       return true;
-//   }
-//
-// An `absl::Status` can optionally include a payload with more information
-// about the error. Typically, this payload serves one of several purposes:
-//
-//   * It may provide more fine-grained semantic information about the error to
-//     facilitate actionable remedies.
-//   * It may provide human-readable contextual information that is more
-//     appropriate to display to an end user.
-//
-// Example:
-//
-//   absl::Status result = DoSomething();
-//   // Inform user to retry after 30 seconds
-//   // See more error details in googleapis/google/rpc/error_details.proto
-//   if (absl::IsResourceExhausted(result)) {
-//     google::rpc::RetryInfo info;
-//     info.retry_delay().seconds() = 30;
-//     // Payloads require a unique key (a URL to ensure no collisions with
-//     // other payloads), and an `absl::Cord` to hold the encoded data.
-//     absl::string_view url = "type.googleapis.com/google.rpc.RetryInfo";
-//     result.SetPayload(url, info.SerializeAsCord());
-//     return result;
-//   }
-//
-// For documentation see https://abseil.io/docs/cpp/guides/status.
-//
-// Returned Status objects may not be ignored. status_internal.h has a forward
-// declaration of the form
-// class ABSL_MUST_USE_RESULT Status;
-class ABSL_ATTRIBUTE_TRIVIAL_ABI Status final {
- public:
-  // Constructors
-
-  // This default constructor creates an OK status with no message or payload.
-  // Avoid this constructor and prefer explicit construction of an OK status
-  // with `absl::OkStatus()`.
-  Status();
-
-  // Creates a status in the canonical error space with the specified
-  // `absl::StatusCode` and error message.  If `code == absl::StatusCode::kOk`,  // NOLINT
-  // `msg` is ignored and an object identical to an OK status is constructed.
-  //
-  // The `msg` string must be in UTF-8. The implementation may complain (e.g.,  // NOLINT
-  // by printing a warning) if it is not.
-  Status(absl::StatusCode code, absl::string_view msg);
-
-  Status(const Status&);
-  Status& operator=(const Status& x);
-
-  // Move operators
-
-  // The moved-from state is valid but unspecified.
-  Status(Status&&) noexcept;
-  Status& operator=(Status&&) noexcept;
-
-  ~Status();
-
-  // Status::Update()
-  //
-  // Updates the existing status with `new_status` provided that `this->ok()`.
-  // If the existing status already contains a non-OK error, this update has no
-  // effect and preserves the current data. Note that this behavior may change
-  // in the future to augment a current non-ok status with additional
-  // information about `new_status`.
-  //
-  // `Update()` provides a convenient way of keeping track of the first error
-  // encountered.
-  //
-  // Example:
-  //   // Instead of "if (overall_status.ok()) overall_status = new_status"
-  //   overall_status.Update(new_status);
-  //
-  void Update(const Status& new_status);
-  void Update(Status&& new_status);
-
-  // Status::ok()
-  //
-  // Returns `true` if `this->code()` == `absl::StatusCode::kOk`,
-  // indicating the absence of an error.
-  // Prefer checking for an OK status using this member function.
-  ABSL_MUST_USE_RESULT bool ok() const;
-
-  // Status::code()
-  //
-  // Returns the canonical error code of type `absl::StatusCode` of this status.
-  absl::StatusCode code() const;
-
-  // Status::raw_code()
-  //
-  // Returns a raw (canonical) error code corresponding to the enum value of
-  // `google.rpc.Code` definitions within
-  // https://github.com/googleapis/googleapis/blob/master/google/rpc/code.proto.
-  // These values could be out of the range of canonical `absl::StatusCode`
-  // enum values.
-  //
-  // NOTE: This function should only be called when converting to an associated
-  // wire format. Use `Status::code()` for error handling.
-  int raw_code() const;
-
-  // Status::message()
-  //
-  // Returns the error message associated with this error code, if available.
-  // Note that this message rarely describes the error code.  It is not unusual
-  // for the error message to be the empty string. As a result, prefer
-  // `operator<<` or `Status::ToString()` for debug logging.
-  absl::string_view message() const;
-
-  friend bool operator==(const Status&, const Status&);
-  friend bool operator!=(const Status&, const Status&);
-
-  // Status::ToString()
-  //
-  // Returns a string based on the `mode`. By default, it returns combination of
-  // the error code name, the message and any associated payload messages. This
-  // string is designed simply to be human readable and its exact format should
-  // not be load bearing. Do not depend on the exact format of the result of
-  // `ToString()` which is subject to change.
-  //
-  // The printed code name and the message are generally substrings of the
-  // result, and the payloads to be printed use the status payload printer
-  // mechanism (which is internal).
-  std::string ToString(
-      StatusToStringMode mode = StatusToStringMode::kDefault) const;
-
-  // Support `absl::StrCat`, `absl::StrFormat`, etc.
-  template <typename Sink>
-  friend void AbslStringify(Sink& sink, const Status& status) {
-    sink.Append(status.ToString(StatusToStringMode::kWithEverything));
-  }
-
-  // Status::IgnoreError()
-  //
-  // Ignores any errors. This method does nothing except potentially suppress
-  // complaints from any tools that are checking that errors are not dropped on
-  // the floor.
-  void IgnoreError() const;
-
-  // swap()
-  //
-  // Swap the contents of one status with another.
-  friend void swap(Status& a, Status& b) noexcept;
-
-  //----------------------------------------------------------------------------
-  // Payload Management APIs
-  //----------------------------------------------------------------------------
-
-  // A payload may be attached to a status to provide additional context to an
-  // error that may not be satisfied by an existing `absl::StatusCode`.
-  // Typically, this payload serves one of several purposes:
-  //
-  //   * It may provide more fine-grained semantic information about the error
-  //     to facilitate actionable remedies.
-  //   * It may provide human-readable contextual information that is more
-  //     appropriate to display to an end user.
-  //
-  // A payload consists of a [key,value] pair, where the key is a string
-  // referring to a unique "type URL" and the value is an object of type
-  // `absl::Cord` to hold the contextual data.
-  //
-  // The "type URL" should be unique and follow the format of a URL
-  // (https://en.wikipedia.org/wiki/URL) and, ideally, provide some
-  // documentation or schema on how to interpret its associated data. For
-  // example, the default type URL for a protobuf message type is
-  // "type.googleapis.com/packagename.messagename". Other custom wire formats
-  // should define the format of type URL in a similar practice so as to
-  // minimize the chance of conflict between type URLs.
-  // Users should ensure that the type URL can be mapped to a concrete
-  // C++ type if they want to deserialize the payload and read it effectively.
-  //
-  // To attach a payload to a status object, call `Status::SetPayload()`,
-  // passing it the type URL and an `absl::Cord` of associated data. Similarly,
-  // to extract the payload from a status, call `Status::GetPayload()`. You
-  // may attach multiple payloads (with differing type URLs) to any given
-  // status object, provided that the status is currently exhibiting an error
-  // code (i.e. is not OK).
-
-  // Status::GetPayload()
-  //
-  // Gets the payload of a status given its unique `type_url` key, if present.
-  absl::optional<absl::Cord> GetPayload(absl::string_view type_url) const;
-
-  // Status::SetPayload()
-  //
-  // Sets the payload for a non-ok status using a `type_url` key, overwriting
-  // any existing payload for that `type_url`.
-  //
-  // NOTE: This function does nothing if the Status is ok.
-  void SetPayload(absl::string_view type_url, absl::Cord payload);
-
-  // Status::ErasePayload()
-  //
-  // Erases the payload corresponding to the `type_url` key.  Returns `true` if
-  // the payload was present.
-  bool ErasePayload(absl::string_view type_url);
-
-  // Status::ForEachPayload()
-  //
-  // Iterates over the stored payloads and calls the
-  // `visitor(type_key, payload)` callable for each one.
-  //
-  // NOTE: The order of calls to `visitor()` is not specified and may change at
-  // any time.
-  //
-  // NOTE: Any mutation on the same 'absl::Status' object during visitation is
-  // forbidden and could result in undefined behavior.
-  void ForEachPayload(
-      absl::FunctionRef<void(absl::string_view, const absl::Cord&)> visitor)
-      const;
-
- private:
-  friend Status CancelledError();
-
-  // Creates a status in the canonical error space with the specified
-  // code, and an empty error message.
-  explicit Status(absl::StatusCode code);
-
-  // Underlying constructor for status from a rep_.
-  explicit Status(uintptr_t rep) : rep_(rep) {}
-
-  static void Ref(uintptr_t rep);
-  static void Unref(uintptr_t rep);
-
-  // REQUIRES: !ok()
-  // Ensures rep is not inlined or shared with any other Status.
-  static absl::Nonnull<status_internal::StatusRep*> PrepareToModify(
-      uintptr_t rep);
-
-  // MSVC 14.0 limitation requires the const.
-  static constexpr const char kMovedFromString[] =
-      "Status accessed after move.";
-
-  static absl::Nonnull<const std::string*> EmptyString();
-  static absl::Nonnull<const std::string*> MovedFromString();
-
-  // Returns whether rep contains an inlined representation.
-  // See rep_ for details.
-  static constexpr bool IsInlined(uintptr_t rep);
-
-  // Indicates whether this Status was the rhs of a move operation. See rep_
-  // for details.
-  static constexpr bool IsMovedFrom(uintptr_t rep);
-  static constexpr uintptr_t MovedFromRep();
-
-  // Convert between error::Code and the inlined uintptr_t representation used
-  // by rep_. See rep_ for details.
-  static constexpr uintptr_t CodeToInlinedRep(absl::StatusCode code);
-  static constexpr absl::StatusCode InlinedRepToCode(uintptr_t rep);
-
-  // Converts between StatusRep* and the external uintptr_t representation used
-  // by rep_. See rep_ for details.
-  static uintptr_t PointerToRep(status_internal::StatusRep* r);
-  static absl::Nonnull<const status_internal::StatusRep*> RepToPointer(
-      uintptr_t r);
-
-  static std::string ToStringSlow(uintptr_t rep, StatusToStringMode mode);
-
-  // Status supports two different representations.
-  //  - When the low bit is set it is an inlined representation.
-  //    It uses the canonical error space, no message or payload.
-  //    The error code is (rep_ >> 2).
-  //    The (rep_ & 2) bit is the "moved from" indicator, used in IsMovedFrom().
-  //  - When the low bit is off it is an external representation.
-  //    In this case all the data comes from a heap allocated Rep object.
-  //    rep_ is a status_internal::StatusRep* pointer to that structure.
-  uintptr_t rep_;
-
-  friend class status_internal::StatusRep;
-};
-
-// OkStatus()
-//
-// Returns an OK status, equivalent to a default constructed instance. Prefer
-// usage of `absl::OkStatus()` when constructing such an OK status.
-Status OkStatus();
-
-// operator<<()
-//
-// Prints a human-readable representation of `x` to `os`.
-std::ostream& operator<<(std::ostream& os, const Status& x);
-
-// IsAborted()
-// IsAlreadyExists()
-// IsCancelled()
-// IsDataLoss()
-// IsDeadlineExceeded()
-// IsFailedPrecondition()
-// IsInternal()
-// IsInvalidArgument()
-// IsNotFound()
-// IsOutOfRange()
-// IsPermissionDenied()
-// IsResourceExhausted()
-// IsUnauthenticated()
-// IsUnavailable()
-// IsUnimplemented()
-// IsUnknown()
-//
-// These convenience functions return `true` if a given status matches the
-// `absl::StatusCode` error code of its associated function.
-ABSL_MUST_USE_RESULT bool IsAborted(const Status& status);
-ABSL_MUST_USE_RESULT bool IsAlreadyExists(const Status& status);
-ABSL_MUST_USE_RESULT bool IsCancelled(const Status& status);
-ABSL_MUST_USE_RESULT bool IsDataLoss(const Status& status);
-ABSL_MUST_USE_RESULT bool IsDeadlineExceeded(const Status& status);
-ABSL_MUST_USE_RESULT bool IsFailedPrecondition(const Status& status);
-ABSL_MUST_USE_RESULT bool IsInternal(const Status& status);
-ABSL_MUST_USE_RESULT bool IsInvalidArgument(const Status& status);
-ABSL_MUST_USE_RESULT bool IsNotFound(const Status& status);
-ABSL_MUST_USE_RESULT bool IsOutOfRange(const Status& status);
-ABSL_MUST_USE_RESULT bool IsPermissionDenied(const Status& status);
-ABSL_MUST_USE_RESULT bool IsResourceExhausted(const Status& status);
-ABSL_MUST_USE_RESULT bool IsUnauthenticated(const Status& status);
-ABSL_MUST_USE_RESULT bool IsUnavailable(const Status& status);
-ABSL_MUST_USE_RESULT bool IsUnimplemented(const Status& status);
-ABSL_MUST_USE_RESULT bool IsUnknown(const Status& status);
-
-// AbortedError()
-// AlreadyExistsError()
-// CancelledError()
-// DataLossError()
-// DeadlineExceededError()
-// FailedPreconditionError()
-// InternalError()
-// InvalidArgumentError()
-// NotFoundError()
-// OutOfRangeError()
-// PermissionDeniedError()
-// ResourceExhaustedError()
-// UnauthenticatedError()
-// UnavailableError()
-// UnimplementedError()
-// UnknownError()
-//
-// These convenience functions create an `absl::Status` object with an error
-// code as indicated by the associated function name, using the error message
-// passed in `message`.
-Status AbortedError(absl::string_view message);
-Status AlreadyExistsError(absl::string_view message);
-Status CancelledError(absl::string_view message);
-Status DataLossError(absl::string_view message);
-Status DeadlineExceededError(absl::string_view message);
-Status FailedPreconditionError(absl::string_view message);
-Status InternalError(absl::string_view message);
-Status InvalidArgumentError(absl::string_view message);
-Status NotFoundError(absl::string_view message);
-Status OutOfRangeError(absl::string_view message);
-Status PermissionDeniedError(absl::string_view message);
-Status ResourceExhaustedError(absl::string_view message);
-Status UnauthenticatedError(absl::string_view message);
-Status UnavailableError(absl::string_view message);
-Status UnimplementedError(absl::string_view message);
-Status UnknownError(absl::string_view message);
-
-// ErrnoToStatusCode()
-//
-// Returns the StatusCode for `error_number`, which should be an `errno` value.
-// See https://en.cppreference.com/w/cpp/error/errno_macros and similar
-// references.
-absl::StatusCode ErrnoToStatusCode(int error_number);
-
-// ErrnoToStatus()
-//
-// Convenience function that creates a `absl::Status` using an `error_number`,
-// which should be an `errno` value.
-Status ErrnoToStatus(int error_number, absl::string_view message);
-
-//------------------------------------------------------------------------------
-// Implementation details follow
-//------------------------------------------------------------------------------
-
-inline Status::Status() : Status(absl::StatusCode::kOk) {}
-
-inline Status::Status(absl::StatusCode code) : Status(CodeToInlinedRep(code)) {}
-
-inline Status::Status(const Status& x) : Status(x.rep_) { Ref(rep_); }
-
-inline Status& Status::operator=(const Status& x) {
-  uintptr_t old_rep = rep_;
-  if (x.rep_ != old_rep) {
-    Ref(x.rep_);
-    rep_ = x.rep_;
-    Unref(old_rep);
-  }
-  return *this;
-}
-
-inline Status::Status(Status&& x) noexcept : Status(x.rep_) {
-  x.rep_ = MovedFromRep();
-}
-
-inline Status& Status::operator=(Status&& x) noexcept {
-  uintptr_t old_rep = rep_;
-  if (x.rep_ != old_rep) {
-    rep_ = x.rep_;
-    x.rep_ = MovedFromRep();
-    Unref(old_rep);
-  }
-  return *this;
-}
-
-inline void Status::Update(const Status& new_status) {
-  if (ok()) {
-    *this = new_status;
-  }
-}
-
-inline void Status::Update(Status&& new_status) {
-  if (ok()) {
-    *this = std::move(new_status);
-  }
-}
-
-inline Status::~Status() { Unref(rep_); }
-
-inline bool Status::ok() const {
-  return rep_ == CodeToInlinedRep(absl::StatusCode::kOk);
-}
-
-inline absl::StatusCode Status::code() const {
-  return status_internal::MapToLocalCode(raw_code());
-}
-
-inline int Status::raw_code() const {
-  if (IsInlined(rep_)) return static_cast<int>(InlinedRepToCode(rep_));
-  return static_cast<int>(RepToPointer(rep_)->code());
-}
-
-inline absl::string_view Status::message() const {
-  return !IsInlined(rep_)
-             ? RepToPointer(rep_)->message()
-             : (IsMovedFrom(rep_) ? absl::string_view(kMovedFromString)
-                                  : absl::string_view());
-}
-
-inline bool operator==(const Status& lhs, const Status& rhs) {
-  if (lhs.rep_ == rhs.rep_) return true;
-  if (Status::IsInlined(lhs.rep_)) return false;
-  if (Status::IsInlined(rhs.rep_)) return false;
-  return *Status::RepToPointer(lhs.rep_) == *Status::RepToPointer(rhs.rep_);
-}
-
-inline bool operator!=(const Status& lhs, const Status& rhs) {
-  return !(lhs == rhs);
-}
-
-inline std::string Status::ToString(StatusToStringMode mode) const {
-  return ok() ? "OK" : ToStringSlow(rep_, mode);
-}
-
-inline void Status::IgnoreError() const {
-  // no-op
-}
-
-inline void swap(absl::Status& a, absl::Status& b) noexcept {
-  using std::swap;
-  swap(a.rep_, b.rep_);
-}
-
-inline absl::optional<absl::Cord> Status::GetPayload(
-    absl::string_view type_url) const {
-  if (IsInlined(rep_)) return absl::nullopt;
-  return RepToPointer(rep_)->GetPayload(type_url);
-}
-
-inline void Status::SetPayload(absl::string_view type_url, absl::Cord payload) {
-  if (ok()) return;
-  status_internal::StatusRep* rep = PrepareToModify(rep_);
-  rep->SetPayload(type_url, std::move(payload));
-  rep_ = PointerToRep(rep);
-}
-
-inline bool Status::ErasePayload(absl::string_view type_url) {
-  if (IsInlined(rep_)) return false;
-  status_internal::StatusRep* rep = PrepareToModify(rep_);
-  auto res = rep->ErasePayload(type_url);
-  rep_ = res.new_rep;
-  return res.erased;
-}
-
-inline void Status::ForEachPayload(
-    absl::FunctionRef<void(absl::string_view, const absl::Cord&)> visitor)
-    const {
-  if (IsInlined(rep_)) return;
-  RepToPointer(rep_)->ForEachPayload(visitor);
-}
-
-constexpr bool Status::IsInlined(uintptr_t rep) { return (rep & 1) != 0; }
-
-constexpr bool Status::IsMovedFrom(uintptr_t rep) { return (rep & 2) != 0; }
-
-constexpr uintptr_t Status::CodeToInlinedRep(absl::StatusCode code) {
-  return (static_cast<uintptr_t>(code) << 2) + 1;
-}
-
-constexpr absl::StatusCode Status::InlinedRepToCode(uintptr_t rep) {
-  ABSL_ASSERT(IsInlined(rep));
-  return static_cast<absl::StatusCode>(rep >> 2);
-}
-
-constexpr uintptr_t Status::MovedFromRep() {
-  return CodeToInlinedRep(absl::StatusCode::kInternal) | 2;
-}
-
-inline absl::Nonnull<const status_internal::StatusRep*> Status::RepToPointer(
-    uintptr_t rep) {
-  assert(!IsInlined(rep));
-  return reinterpret_cast<const status_internal::StatusRep*>(rep);
-}
-
-inline uintptr_t Status::PointerToRep(
-    absl::Nonnull<status_internal::StatusRep*> rep) {
-  return reinterpret_cast<uintptr_t>(rep);
-}
-
-inline void Status::Ref(uintptr_t rep) {
-  if (!IsInlined(rep)) RepToPointer(rep)->Ref();
-}
-
-inline void Status::Unref(uintptr_t rep) {
-  if (!IsInlined(rep)) RepToPointer(rep)->Unref();
-}
-
-inline Status OkStatus() { return Status(); }
-
-// Creates a `Status` object with the `absl::StatusCode::kCancelled` error code
-// and an empty message. It is provided only for efficiency, given that
-// message-less kCancelled errors are common in the infrastructure.
-inline Status CancelledError() { return Status(absl::StatusCode::kCancelled); }
-
-// Retrieves a message's status as a null terminated C string. The lifetime of
-// this string is tied to the lifetime of the status object itself.
-//
-// If the status's message is empty, the empty string is returned.
-//
-// StatusMessageAsCStr exists for C support. Use `status.message()` in C++.
-absl::Nonnull<const char*> StatusMessageAsCStr(
-    const Status& status ABSL_ATTRIBUTE_LIFETIME_BOUND);
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STATUS_STATUS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/status/status_payload_printer.h b/projects/llm_framework/include/abseil-cpp/absl/status/status_payload_printer.h
deleted file mode 100644
index f22255e1..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/status/status_payload_printer.h
+++ /dev/null
@@ -1,52 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-#ifndef ABSL_STATUS_STATUS_PAYLOAD_PRINTER_H_
-#define ABSL_STATUS_STATUS_PAYLOAD_PRINTER_H_
-
-#include <string>
-
-#include "absl/base/nullability.h"
-#include "absl/strings/cord.h"
-#include "absl/strings/string_view.h"
-#include "absl/types/optional.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace status_internal {
-
-// By default, `Status::ToString` and `operator<<(Status)` print a payload by
-// dumping the type URL and the raw bytes. To help debugging, we provide an
-// extension point, which is a global printer function that can be set by users
-// to specify how to print payloads. The function takes the type URL and the
-// payload as input, and should return a valid human-readable string on success
-// or `absl::nullopt` on failure (in which case it falls back to the default
-// approach of printing the raw bytes).
-// NOTE: This is an internal API and the design is subject to change in the
-// future in a non-backward-compatible way. Since it's only meant for debugging
-// purpose, you should not rely on it in any critical logic.
-using StatusPayloadPrinter = absl::Nullable<absl::optional<std::string> (*)(
-    absl::string_view, const absl::Cord&)>;
-
-// Sets the global payload printer. Only one printer should be set per process.
-// If multiple printers are set, it's undefined which one will be used.
-void SetStatusPayloadPrinter(StatusPayloadPrinter);
-
-// Returns the global payload printer if previously set, otherwise `nullptr`.
-StatusPayloadPrinter GetStatusPayloadPrinter();
-
-}  // namespace status_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STATUS_STATUS_PAYLOAD_PRINTER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/status/statusor.h b/projects/llm_framework/include/abseil-cpp/absl/status/statusor.h
deleted file mode 100644
index cd35e5b4..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/status/statusor.h
+++ /dev/null
@@ -1,830 +0,0 @@
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: statusor.h
-// -----------------------------------------------------------------------------
-//
-// An `absl::StatusOr<T>` represents a union of an `absl::Status` object
-// and an object of type `T`. The `absl::StatusOr<T>` will either contain an
-// object of type `T` (indicating a successful operation), or an error (of type
-// `absl::Status`) explaining why such a value is not present.
-//
-// In general, check the success of an operation returning an
-// `absl::StatusOr<T>` like you would an `absl::Status` by using the `ok()`
-// member function.
-//
-// Example:
-//
-//   StatusOr<Foo> result = Calculation();
-//   if (result.ok()) {
-//     result->DoSomethingCool();
-//   } else {
-//     LOG(ERROR) << result.status();
-//   }
-#ifndef ABSL_STATUS_STATUSOR_H_
-#define ABSL_STATUS_STATUSOR_H_
-
-#include <exception>
-#include <initializer_list>
-#include <new>
-#include <ostream>
-#include <string>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/attributes.h"
-#include "absl/base/nullability.h"
-#include "absl/base/call_once.h"
-#include "absl/meta/type_traits.h"
-#include "absl/status/internal/statusor_internal.h"
-#include "absl/status/status.h"
-#include "absl/strings/has_absl_stringify.h"
-#include "absl/strings/has_ostream_operator.h"
-#include "absl/strings/str_format.h"
-#include "absl/types/variant.h"
-#include "absl/utility/utility.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// BadStatusOrAccess
-//
-// This class defines the type of object to throw (if exceptions are enabled),
-// when accessing the value of an `absl::StatusOr<T>` object that does not
-// contain a value. This behavior is analogous to that of
-// `std::bad_optional_access` in the case of accessing an invalid
-// `std::optional` value.
-//
-// Example:
-//
-// try {
-//   absl::StatusOr<int> v = FetchInt();
-//   DoWork(v.value());  // Accessing value() when not "OK" may throw
-// } catch (absl::BadStatusOrAccess& ex) {
-//   LOG(ERROR) << ex.status();
-// }
-class BadStatusOrAccess : public std::exception {
- public:
-  explicit BadStatusOrAccess(absl::Status status);
-  ~BadStatusOrAccess() override = default;
-
-  BadStatusOrAccess(const BadStatusOrAccess& other);
-  BadStatusOrAccess& operator=(const BadStatusOrAccess& other);
-  BadStatusOrAccess(BadStatusOrAccess&& other);
-  BadStatusOrAccess& operator=(BadStatusOrAccess&& other);
-
-  // BadStatusOrAccess::what()
-  //
-  // Returns the associated explanatory string of the `absl::StatusOr<T>`
-  // object's error code. This function contains information about the failing
-  // status, but its exact formatting may change and should not be depended on.
-  //
-  // The pointer of this string is guaranteed to be valid until any non-const
-  // function is invoked on the exception object.
-  absl::Nonnull<const char*> what() const noexcept override;
-
-  // BadStatusOrAccess::status()
-  //
-  // Returns the associated `absl::Status` of the `absl::StatusOr<T>` object's
-  // error.
-  const absl::Status& status() const;
-
- private:
-  void InitWhat() const;
-
-  absl::Status status_;
-  mutable absl::once_flag init_what_;
-  mutable std::string what_;
-};
-
-// Returned StatusOr objects may not be ignored.
-template <typename T>
-#if ABSL_HAVE_CPP_ATTRIBUTE(nodiscard)
-// TODO(b/176172494): ABSL_MUST_USE_RESULT should expand to the more strict
-// [[nodiscard]]. For now, just use [[nodiscard]] directly when it is available.
-class [[nodiscard]] StatusOr;
-#else
-class ABSL_MUST_USE_RESULT StatusOr;
-#endif  // ABSL_HAVE_CPP_ATTRIBUTE(nodiscard)
-
-// absl::StatusOr<T>
-//
-// The `absl::StatusOr<T>` class template is a union of an `absl::Status` object
-// and an object of type `T`. The `absl::StatusOr<T>` models an object that is
-// either a usable object, or an error (of type `absl::Status`) explaining why
-// such an object is not present. An `absl::StatusOr<T>` is typically the return
-// value of a function which may fail.
-//
-// An `absl::StatusOr<T>` can never hold an "OK" status (an
-// `absl::StatusCode::kOk` value); instead, the presence of an object of type
-// `T` indicates success. Instead of checking for a `kOk` value, use the
-// `absl::StatusOr<T>::ok()` member function. (It is for this reason, and code
-// readability, that using the `ok()` function is preferred for `absl::Status`
-// as well.)
-//
-// Example:
-//
-//   StatusOr<Foo> result = DoBigCalculationThatCouldFail();
-//   if (result.ok()) {
-//     result->DoSomethingCool();
-//   } else {
-//     LOG(ERROR) << result.status();
-//   }
-//
-// Accessing the object held by an `absl::StatusOr<T>` should be performed via
-// `operator*` or `operator->`, after a call to `ok()` confirms that the
-// `absl::StatusOr<T>` holds an object of type `T`:
-//
-// Example:
-//
-//   absl::StatusOr<int> i = GetCount();
-//   if (i.ok()) {
-//     updated_total += *i;
-//   }
-//
-// NOTE: using `absl::StatusOr<T>::value()` when no valid value is present will
-// throw an exception if exceptions are enabled or terminate the process when
-// exceptions are not enabled.
-//
-// Example:
-//
-//   StatusOr<Foo> result = DoBigCalculationThatCouldFail();
-//   const Foo& foo = result.value();    // Crash/exception if no value present
-//   foo.DoSomethingCool();
-//
-// A `absl::StatusOr<T*>` can be constructed from a null pointer like any other
-// pointer value, and the result will be that `ok()` returns `true` and
-// `value()` returns `nullptr`. Checking the value of pointer in an
-// `absl::StatusOr<T*>` generally requires a bit more care, to ensure both that
-// a value is present and that value is not null:
-//
-//  StatusOr<std::unique_ptr<Foo>> result = FooFactory::MakeNewFoo(arg);
-//  if (!result.ok()) {
-//    LOG(ERROR) << result.status();
-//  } else if (*result == nullptr) {
-//    LOG(ERROR) << "Unexpected null pointer";
-//  } else {
-//    (*result)->DoSomethingCool();
-//  }
-//
-// Example factory implementation returning StatusOr<T>:
-//
-//  StatusOr<Foo> FooFactory::MakeFoo(int arg) {
-//    if (arg <= 0) {
-//      return absl::Status(absl::StatusCode::kInvalidArgument,
-//                          "Arg must be positive");
-//    }
-//    return Foo(arg);
-//  }
-template <typename T>
-class StatusOr : private internal_statusor::StatusOrData<T>,
-                 private internal_statusor::CopyCtorBase<T>,
-                 private internal_statusor::MoveCtorBase<T>,
-                 private internal_statusor::CopyAssignBase<T>,
-                 private internal_statusor::MoveAssignBase<T> {
-  template <typename U>
-  friend class StatusOr;
-
-  typedef internal_statusor::StatusOrData<T> Base;
-
- public:
-  // StatusOr<T>::value_type
-  //
-  // This instance data provides a generic `value_type` member for use within
-  // generic programming. This usage is analogous to that of
-  // `optional::value_type` in the case of `std::optional`.
-  typedef T value_type;
-
-  // Constructors
-
-  // Constructs a new `absl::StatusOr` with an `absl::StatusCode::kUnknown`
-  // status. This constructor is marked 'explicit' to prevent usages in return
-  // values such as 'return {};', under the misconception that
-  // `absl::StatusOr<std::vector<int>>` will be initialized with an empty
-  // vector, instead of an `absl::StatusCode::kUnknown` error code.
-  explicit StatusOr();
-
-  // `StatusOr<T>` is copy constructible if `T` is copy constructible.
-  StatusOr(const StatusOr&) = default;
-  // `StatusOr<T>` is copy assignable if `T` is copy constructible and copy
-  // assignable.
-  StatusOr& operator=(const StatusOr&) = default;
-
-  // `StatusOr<T>` is move constructible if `T` is move constructible.
-  StatusOr(StatusOr&&) = default;
-  // `StatusOr<T>` is moveAssignable if `T` is move constructible and move
-  // assignable.
-  StatusOr& operator=(StatusOr&&) = default;
-
-  // Converting Constructors
-
-  // Constructs a new `absl::StatusOr<T>` from an `absl::StatusOr<U>`, when `T`
-  // is constructible from `U`. To avoid ambiguity, these constructors are
-  // disabled if `T` is also constructible from `StatusOr<U>.`. This constructor
-  // is explicit if and only if the corresponding construction of `T` from `U`
-  // is explicit. (This constructor inherits its explicitness from the
-  // underlying constructor.)
-  template <
-      typename U,
-      absl::enable_if_t<
-          absl::conjunction<
-              absl::negation<std::is_same<T, U>>,
-              std::is_constructible<T, const U&>,
-              std::is_convertible<const U&, T>,
-              absl::negation<
-                  internal_statusor::IsConstructibleOrConvertibleFromStatusOr<
-                      T, U>>>::value,
-          int> = 0>
-  StatusOr(const StatusOr<U>& other)  // NOLINT
-      : Base(static_cast<const typename StatusOr<U>::Base&>(other)) {}
-  template <
-      typename U,
-      absl::enable_if_t<
-          absl::conjunction<
-              absl::negation<std::is_same<T, U>>,
-              std::is_constructible<T, const U&>,
-              absl::negation<std::is_convertible<const U&, T>>,
-              absl::negation<
-                  internal_statusor::IsConstructibleOrConvertibleFromStatusOr<
-                      T, U>>>::value,
-          int> = 0>
-  explicit StatusOr(const StatusOr<U>& other)
-      : Base(static_cast<const typename StatusOr<U>::Base&>(other)) {}
-
-  template <
-      typename U,
-      absl::enable_if_t<
-          absl::conjunction<
-              absl::negation<std::is_same<T, U>>, std::is_constructible<T, U&&>,
-              std::is_convertible<U&&, T>,
-              absl::negation<
-                  internal_statusor::IsConstructibleOrConvertibleFromStatusOr<
-                      T, U>>>::value,
-          int> = 0>
-  StatusOr(StatusOr<U>&& other)  // NOLINT
-      : Base(static_cast<typename StatusOr<U>::Base&&>(other)) {}
-  template <
-      typename U,
-      absl::enable_if_t<
-          absl::conjunction<
-              absl::negation<std::is_same<T, U>>, std::is_constructible<T, U&&>,
-              absl::negation<std::is_convertible<U&&, T>>,
-              absl::negation<
-                  internal_statusor::IsConstructibleOrConvertibleFromStatusOr<
-                      T, U>>>::value,
-          int> = 0>
-  explicit StatusOr(StatusOr<U>&& other)
-      : Base(static_cast<typename StatusOr<U>::Base&&>(other)) {}
-
-  // Converting Assignment Operators
-
-  // Creates an `absl::StatusOr<T>` through assignment from an
-  // `absl::StatusOr<U>` when:
-  //
-  //   * Both `absl::StatusOr<T>` and `absl::StatusOr<U>` are OK by assigning
-  //     `U` to `T` directly.
-  //   * `absl::StatusOr<T>` is OK and `absl::StatusOr<U>` contains an error
-  //      code by destroying `absl::StatusOr<T>`'s value and assigning from
-  //      `absl::StatusOr<U>'
-  //   * `absl::StatusOr<T>` contains an error code and `absl::StatusOr<U>` is
-  //      OK by directly initializing `T` from `U`.
-  //   * Both `absl::StatusOr<T>` and `absl::StatusOr<U>` contain an error
-  //     code by assigning the `Status` in `absl::StatusOr<U>` to
-  //     `absl::StatusOr<T>`
-  //
-  // These overloads only apply if `absl::StatusOr<T>` is constructible and
-  // assignable from `absl::StatusOr<U>` and `StatusOr<T>` cannot be directly
-  // assigned from `StatusOr<U>`.
-  template <
-      typename U,
-      absl::enable_if_t<
-          absl::conjunction<
-              absl::negation<std::is_same<T, U>>,
-              std::is_constructible<T, const U&>,
-              std::is_assignable<T, const U&>,
-              absl::negation<
-                  internal_statusor::
-                      IsConstructibleOrConvertibleOrAssignableFromStatusOr<
-                          T, U>>>::value,
-          int> = 0>
-  StatusOr& operator=(const StatusOr<U>& other) {
-    this->Assign(other);
-    return *this;
-  }
-  template <
-      typename U,
-      absl::enable_if_t<
-          absl::conjunction<
-              absl::negation<std::is_same<T, U>>, std::is_constructible<T, U&&>,
-              std::is_assignable<T, U&&>,
-              absl::negation<
-                  internal_statusor::
-                      IsConstructibleOrConvertibleOrAssignableFromStatusOr<
-                          T, U>>>::value,
-          int> = 0>
-  StatusOr& operator=(StatusOr<U>&& other) {
-    this->Assign(std::move(other));
-    return *this;
-  }
-
-  // Constructs a new `absl::StatusOr<T>` with a non-ok status. After calling
-  // this constructor, `this->ok()` will be `false` and calls to `value()` will
-  // crash, or produce an exception if exceptions are enabled.
-  //
-  // The constructor also takes any type `U` that is convertible to
-  // `absl::Status`. This constructor is explicit if an only if `U` is not of
-  // type `absl::Status` and the conversion from `U` to `Status` is explicit.
-  //
-  // REQUIRES: !Status(std::forward<U>(v)).ok(). This requirement is DCHECKed.
-  // In optimized builds, passing absl::OkStatus() here will have the effect
-  // of passing absl::StatusCode::kInternal as a fallback.
-  template <
-      typename U = absl::Status,
-      absl::enable_if_t<
-          absl::conjunction<
-              std::is_convertible<U&&, absl::Status>,
-              std::is_constructible<absl::Status, U&&>,
-              absl::negation<std::is_same<absl::decay_t<U>, absl::StatusOr<T>>>,
-              absl::negation<std::is_same<absl::decay_t<U>, T>>,
-              absl::negation<std::is_same<absl::decay_t<U>, absl::in_place_t>>,
-              absl::negation<internal_statusor::HasConversionOperatorToStatusOr<
-                  T, U&&>>>::value,
-          int> = 0>
-  StatusOr(U&& v) : Base(std::forward<U>(v)) {}
-
-  template <
-      typename U = absl::Status,
-      absl::enable_if_t<
-          absl::conjunction<
-              absl::negation<std::is_convertible<U&&, absl::Status>>,
-              std::is_constructible<absl::Status, U&&>,
-              absl::negation<std::is_same<absl::decay_t<U>, absl::StatusOr<T>>>,
-              absl::negation<std::is_same<absl::decay_t<U>, T>>,
-              absl::negation<std::is_same<absl::decay_t<U>, absl::in_place_t>>,
-              absl::negation<internal_statusor::HasConversionOperatorToStatusOr<
-                  T, U&&>>>::value,
-          int> = 0>
-  explicit StatusOr(U&& v) : Base(std::forward<U>(v)) {}
-
-  template <
-      typename U = absl::Status,
-      absl::enable_if_t<
-          absl::conjunction<
-              std::is_convertible<U&&, absl::Status>,
-              std::is_constructible<absl::Status, U&&>,
-              absl::negation<std::is_same<absl::decay_t<U>, absl::StatusOr<T>>>,
-              absl::negation<std::is_same<absl::decay_t<U>, T>>,
-              absl::negation<std::is_same<absl::decay_t<U>, absl::in_place_t>>,
-              absl::negation<internal_statusor::HasConversionOperatorToStatusOr<
-                  T, U&&>>>::value,
-          int> = 0>
-  StatusOr& operator=(U&& v) {
-    this->AssignStatus(std::forward<U>(v));
-    return *this;
-  }
-
-  // Perfect-forwarding value assignment operator.
-
-  // If `*this` contains a `T` value before the call, the contained value is
-  // assigned from `std::forward<U>(v)`; Otherwise, it is directly-initialized
-  // from `std::forward<U>(v)`.
-  // This function does not participate in overload unless:
-  // 1. `std::is_constructible_v<T, U>` is true,
-  // 2. `std::is_assignable_v<T&, U>` is true.
-  // 3. `std::is_same_v<StatusOr<T>, std::remove_cvref_t<U>>` is false.
-  // 4. Assigning `U` to `T` is not ambiguous:
-  //  If `U` is `StatusOr<V>` and `T` is constructible and assignable from
-  //  both `StatusOr<V>` and `V`, the assignment is considered bug-prone and
-  //  ambiguous thus will fail to compile. For example:
-  //    StatusOr<bool> s1 = true;  // s1.ok() && *s1 == true
-  //    StatusOr<bool> s2 = false;  // s2.ok() && *s2 == false
-  //    s1 = s2;  // ambiguous, `s1 = *s2` or `s1 = bool(s2)`?
-  template <
-      typename U = T,
-      typename = typename std::enable_if<absl::conjunction<
-          std::is_constructible<T, U&&>, std::is_assignable<T&, U&&>,
-          absl::disjunction<
-              std::is_same<absl::remove_cvref_t<U>, T>,
-              absl::conjunction<
-                  absl::negation<std::is_convertible<U&&, absl::Status>>,
-                  absl::negation<internal_statusor::
-                                     HasConversionOperatorToStatusOr<T, U&&>>>>,
-          internal_statusor::IsForwardingAssignmentValid<T, U&&>>::value>::type>
-  StatusOr& operator=(U&& v) {
-    this->Assign(std::forward<U>(v));
-    return *this;
-  }
-
-  // Constructs the inner value `T` in-place using the provided args, using the
-  // `T(args...)` constructor.
-  template <typename... Args>
-  explicit StatusOr(absl::in_place_t, Args&&... args);
-  template <typename U, typename... Args>
-  explicit StatusOr(absl::in_place_t, std::initializer_list<U> ilist,
-                    Args&&... args);
-
-  // Constructs the inner value `T` in-place using the provided args, using the
-  // `T(U)` (direct-initialization) constructor. This constructor is only valid
-  // if `T` can be constructed from a `U`. Can accept move or copy constructors.
-  //
-  // This constructor is explicit if `U` is not convertible to `T`. To avoid
-  // ambiguity, this constructor is disabled if `U` is a `StatusOr<J>`, where
-  // `J` is convertible to `T`.
-  template <
-      typename U = T,
-      absl::enable_if_t<
-          absl::conjunction<
-              internal_statusor::IsDirectInitializationValid<T, U&&>,
-              std::is_constructible<T, U&&>, std::is_convertible<U&&, T>,
-              absl::disjunction<
-                  std::is_same<absl::remove_cvref_t<U>, T>,
-                  absl::conjunction<
-                      absl::negation<std::is_convertible<U&&, absl::Status>>,
-                      absl::negation<
-                          internal_statusor::HasConversionOperatorToStatusOr<
-                              T, U&&>>>>>::value,
-          int> = 0>
-  StatusOr(U&& u)  // NOLINT
-      : StatusOr(absl::in_place, std::forward<U>(u)) {}
-
-  template <
-      typename U = T,
-      absl::enable_if_t<
-          absl::conjunction<
-              internal_statusor::IsDirectInitializationValid<T, U&&>,
-              absl::disjunction<
-                  std::is_same<absl::remove_cvref_t<U>, T>,
-                  absl::conjunction<
-                      absl::negation<std::is_constructible<absl::Status, U&&>>,
-                      absl::negation<
-                          internal_statusor::HasConversionOperatorToStatusOr<
-                              T, U&&>>>>,
-              std::is_constructible<T, U&&>,
-              absl::negation<std::is_convertible<U&&, T>>>::value,
-          int> = 0>
-  explicit StatusOr(U&& u)  // NOLINT
-      : StatusOr(absl::in_place, std::forward<U>(u)) {}
-
-  // StatusOr<T>::ok()
-  //
-  // Returns whether or not this `absl::StatusOr<T>` holds a `T` value. This
-  // member function is analogous to `absl::Status::ok()` and should be used
-  // similarly to check the status of return values.
-  //
-  // Example:
-  //
-  // StatusOr<Foo> result = DoBigCalculationThatCouldFail();
-  // if (result.ok()) {
-  //    // Handle result
-  // else {
-  //    // Handle error
-  // }
-  ABSL_MUST_USE_RESULT bool ok() const { return this->status_.ok(); }
-
-  // StatusOr<T>::status()
-  //
-  // Returns a reference to the current `absl::Status` contained within the
-  // `absl::StatusOr<T>`. If `absl::StatusOr<T>` contains a `T`, then this
-  // function returns `absl::OkStatus()`.
-  const Status& status() const&;
-  Status status() &&;
-
-  // StatusOr<T>::value()
-  //
-  // Returns a reference to the held value if `this->ok()`. Otherwise, throws
-  // `absl::BadStatusOrAccess` if exceptions are enabled, or is guaranteed to
-  // terminate the process if exceptions are disabled.
-  //
-  // If you have already checked the status using `this->ok()`, you probably
-  // want to use `operator*()` or `operator->()` to access the value instead of
-  // `value`.
-  //
-  // Note: for value types that are cheap to copy, prefer simple code:
-  //
-  //   T value = statusor.value();
-  //
-  // Otherwise, if the value type is expensive to copy, but can be left
-  // in the StatusOr, simply assign to a reference:
-  //
-  //   T& value = statusor.value();  // or `const T&`
-  //
-  // Otherwise, if the value type supports an efficient move, it can be
-  // used as follows:
-  //
-  //   T value = std::move(statusor).value();
-  //
-  // The `std::move` on statusor instead of on the whole expression enables
-  // warnings about possible uses of the statusor object after the move.
-  const T& value() const& ABSL_ATTRIBUTE_LIFETIME_BOUND;
-  T& value() & ABSL_ATTRIBUTE_LIFETIME_BOUND;
-  const T&& value() const&& ABSL_ATTRIBUTE_LIFETIME_BOUND;
-  T&& value() && ABSL_ATTRIBUTE_LIFETIME_BOUND;
-
-  // StatusOr<T>:: operator*()
-  //
-  // Returns a reference to the current value.
-  //
-  // REQUIRES: `this->ok() == true`, otherwise the behavior is undefined.
-  //
-  // Use `this->ok()` to verify that there is a current value within the
-  // `absl::StatusOr<T>`. Alternatively, see the `value()` member function for a
-  // similar API that guarantees crashing or throwing an exception if there is
-  // no current value.
-  const T& operator*() const& ABSL_ATTRIBUTE_LIFETIME_BOUND;
-  T& operator*() & ABSL_ATTRIBUTE_LIFETIME_BOUND;
-  const T&& operator*() const&& ABSL_ATTRIBUTE_LIFETIME_BOUND;
-  T&& operator*() && ABSL_ATTRIBUTE_LIFETIME_BOUND;
-
-  // StatusOr<T>::operator->()
-  //
-  // Returns a pointer to the current value.
-  //
-  // REQUIRES: `this->ok() == true`, otherwise the behavior is undefined.
-  //
-  // Use `this->ok()` to verify that there is a current value.
-  const T* operator->() const ABSL_ATTRIBUTE_LIFETIME_BOUND;
-  T* operator->() ABSL_ATTRIBUTE_LIFETIME_BOUND;
-
-  // StatusOr<T>::value_or()
-  //
-  // Returns the current value if `this->ok() == true`. Otherwise constructs a
-  // value using the provided `default_value`.
-  //
-  // Unlike `value`, this function returns by value, copying the current value
-  // if necessary. If the value type supports an efficient move, it can be used
-  // as follows:
-  //
-  //   T value = std::move(statusor).value_or(def);
-  //
-  // Unlike with `value`, calling `std::move()` on the result of `value_or` will
-  // still trigger a copy.
-  template <typename U>
-  T value_or(U&& default_value) const&;
-  template <typename U>
-  T value_or(U&& default_value) &&;
-
-  // StatusOr<T>::IgnoreError()
-  //
-  // Ignores any errors. This method does nothing except potentially suppress
-  // complaints from any tools that are checking that errors are not dropped on
-  // the floor.
-  void IgnoreError() const;
-
-  // StatusOr<T>::emplace()
-  //
-  // Reconstructs the inner value T in-place using the provided args, using the
-  // T(args...) constructor. Returns reference to the reconstructed `T`.
-  template <typename... Args>
-  T& emplace(Args&&... args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    if (ok()) {
-      this->Clear();
-      this->MakeValue(std::forward<Args>(args)...);
-    } else {
-      this->MakeValue(std::forward<Args>(args)...);
-      this->status_ = absl::OkStatus();
-    }
-    return this->data_;
-  }
-
-  template <
-      typename U, typename... Args,
-      absl::enable_if_t<
-          std::is_constructible<T, std::initializer_list<U>&, Args&&...>::value,
-          int> = 0>
-  T& emplace(std::initializer_list<U> ilist,
-             Args&&... args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    if (ok()) {
-      this->Clear();
-      this->MakeValue(ilist, std::forward<Args>(args)...);
-    } else {
-      this->MakeValue(ilist, std::forward<Args>(args)...);
-      this->status_ = absl::OkStatus();
-    }
-    return this->data_;
-  }
-
-  // StatusOr<T>::AssignStatus()
-  //
-  // Sets the status of `absl::StatusOr<T>` to the given non-ok status value.
-  //
-  // NOTE: We recommend using the constructor and `operator=` where possible.
-  // This method is intended for use in generic programming, to enable setting
-  // the status of a `StatusOr<T>` when `T` may be `Status`. In that case, the
-  // constructor and `operator=` would assign into the inner value of type
-  // `Status`, rather than status of the `StatusOr` (b/280392796).
-  //
-  // REQUIRES: !Status(std::forward<U>(v)).ok(). This requirement is DCHECKed.
-  // In optimized builds, passing absl::OkStatus() here will have the effect
-  // of passing absl::StatusCode::kInternal as a fallback.
-  using internal_statusor::StatusOrData<T>::AssignStatus;
-
- private:
-  using internal_statusor::StatusOrData<T>::Assign;
-  template <typename U>
-  void Assign(const absl::StatusOr<U>& other);
-  template <typename U>
-  void Assign(absl::StatusOr<U>&& other);
-};
-
-// operator==()
-//
-// This operator checks the equality of two `absl::StatusOr<T>` objects.
-template <typename T>
-bool operator==(const StatusOr<T>& lhs, const StatusOr<T>& rhs) {
-  if (lhs.ok() && rhs.ok()) return *lhs == *rhs;
-  return lhs.status() == rhs.status();
-}
-
-// operator!=()
-//
-// This operator checks the inequality of two `absl::StatusOr<T>` objects.
-template <typename T>
-bool operator!=(const StatusOr<T>& lhs, const StatusOr<T>& rhs) {
-  return !(lhs == rhs);
-}
-
-// Prints the `value` or the status in brackets to `os`.
-//
-// Requires `T` supports `operator<<`.  Do not rely on the output format which
-// may change without notice.
-template <typename T, typename std::enable_if<
-                          absl::HasOstreamOperator<T>::value, int>::type = 0>
-std::ostream& operator<<(std::ostream& os, const StatusOr<T>& status_or) {
-  if (status_or.ok()) {
-    os << status_or.value();
-  } else {
-    os << internal_statusor::StringifyRandom::OpenBrackets()
-       << status_or.status()
-       << internal_statusor::StringifyRandom::CloseBrackets();
-  }
-  return os;
-}
-
-// As above, but supports `StrCat`, `StrFormat`, etc.
-//
-// Requires `T` has `AbslStringify`.  Do not rely on the output format which
-// may change without notice.
-template <
-    typename Sink, typename T,
-    typename std::enable_if<absl::HasAbslStringify<T>::value, int>::type = 0>
-void AbslStringify(Sink& sink, const StatusOr<T>& status_or) {
-  if (status_or.ok()) {
-    absl::Format(&sink, "%v", status_or.value());
-  } else {
-    absl::Format(&sink, "%s%v%s",
-                 internal_statusor::StringifyRandom::OpenBrackets(),
-                 status_or.status(),
-                 internal_statusor::StringifyRandom::CloseBrackets());
-  }
-}
-
-//------------------------------------------------------------------------------
-// Implementation details for StatusOr<T>
-//------------------------------------------------------------------------------
-
-// TODO(sbenza): avoid the string here completely.
-template <typename T>
-StatusOr<T>::StatusOr() : Base(Status(absl::StatusCode::kUnknown, "")) {}
-
-template <typename T>
-template <typename U>
-inline void StatusOr<T>::Assign(const StatusOr<U>& other) {
-  if (other.ok()) {
-    this->Assign(*other);
-  } else {
-    this->AssignStatus(other.status());
-  }
-}
-
-template <typename T>
-template <typename U>
-inline void StatusOr<T>::Assign(StatusOr<U>&& other) {
-  if (other.ok()) {
-    this->Assign(*std::move(other));
-  } else {
-    this->AssignStatus(std::move(other).status());
-  }
-}
-template <typename T>
-template <typename... Args>
-StatusOr<T>::StatusOr(absl::in_place_t, Args&&... args)
-    : Base(absl::in_place, std::forward<Args>(args)...) {}
-
-template <typename T>
-template <typename U, typename... Args>
-StatusOr<T>::StatusOr(absl::in_place_t, std::initializer_list<U> ilist,
-                      Args&&... args)
-    : Base(absl::in_place, ilist, std::forward<Args>(args)...) {}
-
-template <typename T>
-const Status& StatusOr<T>::status() const& {
-  return this->status_;
-}
-template <typename T>
-Status StatusOr<T>::status() && {
-  return ok() ? OkStatus() : std::move(this->status_);
-}
-
-template <typename T>
-const T& StatusOr<T>::value() const& {
-  if (!this->ok()) internal_statusor::ThrowBadStatusOrAccess(this->status_);
-  return this->data_;
-}
-
-template <typename T>
-T& StatusOr<T>::value() & {
-  if (!this->ok()) internal_statusor::ThrowBadStatusOrAccess(this->status_);
-  return this->data_;
-}
-
-template <typename T>
-const T&& StatusOr<T>::value() const&& {
-  if (!this->ok()) {
-    internal_statusor::ThrowBadStatusOrAccess(std::move(this->status_));
-  }
-  return std::move(this->data_);
-}
-
-template <typename T>
-T&& StatusOr<T>::value() && {
-  if (!this->ok()) {
-    internal_statusor::ThrowBadStatusOrAccess(std::move(this->status_));
-  }
-  return std::move(this->data_);
-}
-
-template <typename T>
-const T& StatusOr<T>::operator*() const& {
-  this->EnsureOk();
-  return this->data_;
-}
-
-template <typename T>
-T& StatusOr<T>::operator*() & {
-  this->EnsureOk();
-  return this->data_;
-}
-
-template <typename T>
-const T&& StatusOr<T>::operator*() const&& {
-  this->EnsureOk();
-  return std::move(this->data_);
-}
-
-template <typename T>
-T&& StatusOr<T>::operator*() && {
-  this->EnsureOk();
-  return std::move(this->data_);
-}
-
-template <typename T>
-absl::Nonnull<const T*> StatusOr<T>::operator->() const {
-  this->EnsureOk();
-  return &this->data_;
-}
-
-template <typename T>
-absl::Nonnull<T*> StatusOr<T>::operator->() {
-  this->EnsureOk();
-  return &this->data_;
-}
-
-template <typename T>
-template <typename U>
-T StatusOr<T>::value_or(U&& default_value) const& {
-  if (ok()) {
-    return this->data_;
-  }
-  return std::forward<U>(default_value);
-}
-
-template <typename T>
-template <typename U>
-T StatusOr<T>::value_or(U&& default_value) && {
-  if (ok()) {
-    return std::move(this->data_);
-  }
-  return std::forward<U>(default_value);
-}
-
-template <typename T>
-void StatusOr<T>::IgnoreError() const {
-  // no-op
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STATUS_STATUSOR_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/ascii.h b/projects/llm_framework/include/abseil-cpp/absl/strings/ascii.h
deleted file mode 100644
index c238f4de..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/ascii.h
+++ /dev/null
@@ -1,244 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: ascii.h
-// -----------------------------------------------------------------------------
-//
-// This package contains functions operating on characters and strings
-// restricted to standard ASCII. These include character classification
-// functions analogous to those found in the ANSI C Standard Library <ctype.h>
-// header file.
-//
-// C++ implementations provide <ctype.h> functionality based on their
-// C environment locale. In general, reliance on such a locale is not ideal, as
-// the locale standard is problematic (and may not return invariant information
-// for the same character set, for example). These `ascii_*()` functions are
-// hard-wired for standard ASCII, much faster, and guaranteed to behave
-// consistently.  They will never be overloaded, nor will their function
-// signature change.
-//
-// `ascii_isalnum()`, `ascii_isalpha()`, `ascii_isascii()`, `ascii_isblank()`,
-// `ascii_iscntrl()`, `ascii_isdigit()`, `ascii_isgraph()`, `ascii_islower()`,
-// `ascii_isprint()`, `ascii_ispunct()`, `ascii_isspace()`, `ascii_isupper()`,
-// `ascii_isxdigit()`
-//   Analogous to the <ctype.h> functions with similar names, these
-//   functions take an unsigned char and return a bool, based on whether the
-//   character matches the condition specified.
-//
-//   If the input character has a numerical value greater than 127, these
-//   functions return `false`.
-//
-// `ascii_tolower()`, `ascii_toupper()`
-//   Analogous to the <ctype.h> functions with similar names, these functions
-//   take an unsigned char and return a char.
-//
-//   If the input character is not an ASCII {lower,upper}-case letter (including
-//   numerical values greater than 127) then the functions return the same value
-//   as the input character.
-
-#ifndef ABSL_STRINGS_ASCII_H_
-#define ABSL_STRINGS_ASCII_H_
-
-#include <algorithm>
-#include <cstddef>
-#include <string>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/nullability.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace ascii_internal {
-
-// Declaration for an array of bitfields holding character information.
-ABSL_DLL extern const unsigned char kPropertyBits[256];
-
-// Declaration for the array of characters to upper-case characters.
-ABSL_DLL extern const char kToUpper[256];
-
-// Declaration for the array of characters to lower-case characters.
-ABSL_DLL extern const char kToLower[256];
-
-}  // namespace ascii_internal
-
-// ascii_isalpha()
-//
-// Determines whether the given character is an alphabetic character.
-inline bool ascii_isalpha(unsigned char c) {
-  return (ascii_internal::kPropertyBits[c] & 0x01) != 0;
-}
-
-// ascii_isalnum()
-//
-// Determines whether the given character is an alphanumeric character.
-inline bool ascii_isalnum(unsigned char c) {
-  return (ascii_internal::kPropertyBits[c] & 0x04) != 0;
-}
-
-// ascii_isspace()
-//
-// Determines whether the given character is a whitespace character (space,
-// tab, vertical tab, formfeed, linefeed, or carriage return).
-inline bool ascii_isspace(unsigned char c) {
-  return (ascii_internal::kPropertyBits[c] & 0x08) != 0;
-}
-
-// ascii_ispunct()
-//
-// Determines whether the given character is a punctuation character.
-inline bool ascii_ispunct(unsigned char c) {
-  return (ascii_internal::kPropertyBits[c] & 0x10) != 0;
-}
-
-// ascii_isblank()
-//
-// Determines whether the given character is a blank character (tab or space).
-inline bool ascii_isblank(unsigned char c) {
-  return (ascii_internal::kPropertyBits[c] & 0x20) != 0;
-}
-
-// ascii_iscntrl()
-//
-// Determines whether the given character is a control character.
-inline bool ascii_iscntrl(unsigned char c) {
-  return (ascii_internal::kPropertyBits[c] & 0x40) != 0;
-}
-
-// ascii_isxdigit()
-//
-// Determines whether the given character can be represented as a hexadecimal
-// digit character (i.e. {0-9} or {A-F}).
-inline bool ascii_isxdigit(unsigned char c) {
-  return (ascii_internal::kPropertyBits[c] & 0x80) != 0;
-}
-
-// ascii_isdigit()
-//
-// Determines whether the given character can be represented as a decimal
-// digit character (i.e. {0-9}).
-inline bool ascii_isdigit(unsigned char c) { return c >= '0' && c <= '9'; }
-
-// ascii_isprint()
-//
-// Determines whether the given character is printable, including spaces.
-inline bool ascii_isprint(unsigned char c) { return c >= 32 && c < 127; }
-
-// ascii_isgraph()
-//
-// Determines whether the given character has a graphical representation.
-inline bool ascii_isgraph(unsigned char c) { return c > 32 && c < 127; }
-
-// ascii_isupper()
-//
-// Determines whether the given character is uppercase.
-inline bool ascii_isupper(unsigned char c) { return c >= 'A' && c <= 'Z'; }
-
-// ascii_islower()
-//
-// Determines whether the given character is lowercase.
-inline bool ascii_islower(unsigned char c) { return c >= 'a' && c <= 'z'; }
-
-// ascii_isascii()
-//
-// Determines whether the given character is ASCII.
-inline bool ascii_isascii(unsigned char c) { return c < 128; }
-
-// ascii_tolower()
-//
-// Returns an ASCII character, converting to lowercase if uppercase is
-// passed. Note that character values > 127 are simply returned.
-inline char ascii_tolower(unsigned char c) {
-  return ascii_internal::kToLower[c];
-}
-
-// Converts the characters in `s` to lowercase, changing the contents of `s`.
-void AsciiStrToLower(absl::Nonnull<std::string*> s);
-
-// Creates a lowercase string from a given absl::string_view.
-ABSL_MUST_USE_RESULT inline std::string AsciiStrToLower(absl::string_view s) {
-  std::string result(s);
-  absl::AsciiStrToLower(&result);
-  return result;
-}
-
-// ascii_toupper()
-//
-// Returns the ASCII character, converting to upper-case if lower-case is
-// passed. Note that characters values > 127 are simply returned.
-inline char ascii_toupper(unsigned char c) {
-  return ascii_internal::kToUpper[c];
-}
-
-// Converts the characters in `s` to uppercase, changing the contents of `s`.
-void AsciiStrToUpper(absl::Nonnull<std::string*> s);
-
-// Creates an uppercase string from a given absl::string_view.
-ABSL_MUST_USE_RESULT inline std::string AsciiStrToUpper(absl::string_view s) {
-  std::string result(s);
-  absl::AsciiStrToUpper(&result);
-  return result;
-}
-
-// Returns absl::string_view with whitespace stripped from the beginning of the
-// given string_view.
-ABSL_MUST_USE_RESULT inline absl::string_view StripLeadingAsciiWhitespace(
-    absl::string_view str) {
-  auto it = std::find_if_not(str.begin(), str.end(), absl::ascii_isspace);
-  return str.substr(static_cast<size_t>(it - str.begin()));
-}
-
-// Strips in place whitespace from the beginning of the given string.
-inline void StripLeadingAsciiWhitespace(absl::Nonnull<std::string*> str) {
-  auto it = std::find_if_not(str->begin(), str->end(), absl::ascii_isspace);
-  str->erase(str->begin(), it);
-}
-
-// Returns absl::string_view with whitespace stripped from the end of the given
-// string_view.
-ABSL_MUST_USE_RESULT inline absl::string_view StripTrailingAsciiWhitespace(
-    absl::string_view str) {
-  auto it = std::find_if_not(str.rbegin(), str.rend(), absl::ascii_isspace);
-  return str.substr(0, static_cast<size_t>(str.rend() - it));
-}
-
-// Strips in place whitespace from the end of the given string
-inline void StripTrailingAsciiWhitespace(absl::Nonnull<std::string*> str) {
-  auto it = std::find_if_not(str->rbegin(), str->rend(), absl::ascii_isspace);
-  str->erase(static_cast<size_t>(str->rend() - it));
-}
-
-// Returns absl::string_view with whitespace stripped from both ends of the
-// given string_view.
-ABSL_MUST_USE_RESULT inline absl::string_view StripAsciiWhitespace(
-    absl::string_view str) {
-  return StripTrailingAsciiWhitespace(StripLeadingAsciiWhitespace(str));
-}
-
-// Strips in place whitespace from both ends of the given string
-inline void StripAsciiWhitespace(absl::Nonnull<std::string*> str) {
-  StripTrailingAsciiWhitespace(str);
-  StripLeadingAsciiWhitespace(str);
-}
-
-// Removes leading, trailing, and consecutive internal whitespace.
-void RemoveExtraAsciiWhitespace(absl::Nonnull<std::string*> str);
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_ASCII_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/charconv.h b/projects/llm_framework/include/abseil-cpp/absl/strings/charconv.h
deleted file mode 100644
index be250902..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/charconv.h
+++ /dev/null
@@ -1,123 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_CHARCONV_H_
-#define ABSL_STRINGS_CHARCONV_H_
-
-#include <system_error>  // NOLINT(build/c++11)
-
-#include "absl/base/config.h"
-#include "absl/base/nullability.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// Workalike compatibility version of std::chars_format from C++17.
-//
-// This is an bitfield enumerator which can be passed to absl::from_chars to
-// configure the string-to-float conversion.
-enum class chars_format {
-  scientific = 1,
-  fixed = 2,
-  hex = 4,
-  general = fixed | scientific,
-};
-
-// The return result of a string-to-number conversion.
-//
-// `ec` will be set to `invalid_argument` if a well-formed number was not found
-// at the start of the input range, `result_out_of_range` if a well-formed
-// number was found, but it was out of the representable range of the requested
-// type, or to std::errc() otherwise.
-//
-// If a well-formed number was found, `ptr` is set to one past the sequence of
-// characters that were successfully parsed.  If none was found, `ptr` is set
-// to the `first` argument to from_chars.
-struct from_chars_result {
-  absl::Nonnull<const char*> ptr;
-  std::errc ec;
-};
-
-// Workalike compatibility version of std::from_chars from C++17.  Currently
-// this only supports the `double` and `float` types.
-//
-// This interface incorporates the proposed resolutions for library issues
-// DR 3080 and DR 3081.  If these are adopted with different wording,
-// Abseil's behavior will change to match the standard.  (The behavior most
-// likely to change is for DR 3081, which says what `value` will be set to in
-// the case of overflow and underflow.  Code that wants to avoid possible
-// breaking changes in this area should not depend on `value` when the returned
-// from_chars_result indicates a range error.)
-//
-// Searches the range [first, last) for the longest matching pattern beginning
-// at `first` that represents a floating point number.  If one is found, store
-// the result in `value`.
-//
-// The matching pattern format is almost the same as that of strtod(), except
-// that (1) C locale is not respected, (2) an initial '+' character in the
-// input range will never be matched, and (3) leading whitespaces are not
-// ignored.
-//
-// If `fmt` is set, it must be one of the enumerator values of the chars_format.
-// (This is despite the fact that chars_format is a bitmask type.)  If set to
-// `scientific`, a matching number must contain an exponent.  If set to `fixed`,
-// then an exponent will never match.  (For example, the string "1e5" will be
-// parsed as "1".)  If set to `hex`, then a hexadecimal float is parsed in the
-// format that strtod() accepts, except that a "0x" prefix is NOT matched.
-// (In particular, in `hex` mode, the input "0xff" results in the largest
-// matching pattern "0".)
-absl::from_chars_result from_chars(absl::Nonnull<const char*> first,
-                                   absl::Nonnull<const char*> last,
-                                   double& value,  // NOLINT
-                                   chars_format fmt = chars_format::general);
-
-absl::from_chars_result from_chars(absl::Nonnull<const char*> first,
-                                   absl::Nonnull<const char*> last,
-                                   float& value,  // NOLINT
-                                   chars_format fmt = chars_format::general);
-
-// std::chars_format is specified as a bitmask type, which means the following
-// operations must be provided:
-inline constexpr chars_format operator&(chars_format lhs, chars_format rhs) {
-  return static_cast<chars_format>(static_cast<int>(lhs) &
-                                   static_cast<int>(rhs));
-}
-inline constexpr chars_format operator|(chars_format lhs, chars_format rhs) {
-  return static_cast<chars_format>(static_cast<int>(lhs) |
-                                   static_cast<int>(rhs));
-}
-inline constexpr chars_format operator^(chars_format lhs, chars_format rhs) {
-  return static_cast<chars_format>(static_cast<int>(lhs) ^
-                                   static_cast<int>(rhs));
-}
-inline constexpr chars_format operator~(chars_format arg) {
-  return static_cast<chars_format>(~static_cast<int>(arg));
-}
-inline chars_format& operator&=(chars_format& lhs, chars_format rhs) {
-  lhs = lhs & rhs;
-  return lhs;
-}
-inline chars_format& operator|=(chars_format& lhs, chars_format rhs) {
-  lhs = lhs | rhs;
-  return lhs;
-}
-inline chars_format& operator^=(chars_format& lhs, chars_format rhs) {
-  lhs = lhs ^ rhs;
-  return lhs;
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_CHARCONV_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/charset.h b/projects/llm_framework/include/abseil-cpp/absl/strings/charset.h
deleted file mode 100644
index ff4e81a4..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/charset.h
+++ /dev/null
@@ -1,164 +0,0 @@
-// Copyright 2022 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: charset.h
-// -----------------------------------------------------------------------------
-//
-// This file contains absl::CharSet, a fast, bit-vector set of 8-bit unsigned
-// characters.
-//
-// Instances can be initialized as constexpr constants. For example:
-//
-//   constexpr absl::CharSet kJustX = absl::CharSet::Char('x');
-//   constexpr absl::CharSet kMySymbols = absl::CharSet("$@!");
-//   constexpr absl::CharSet kLetters = absl::CharSet::Range('a', 'z');
-//
-// Multiple instances can be combined that still forms a constexpr expression.
-// For example:
-//
-//   constexpr absl::CharSet kLettersAndNumbers =
-//       absl::CharSet::Range('a', 'z') | absl::CharSet::Range('0', '9');
-//
-// Several pre-defined character classes are available that mirror the methods
-// from <cctype>. For example:
-//
-//   constexpr absl::CharSet kLettersAndWhitespace =
-//       absl::CharSet::AsciiAlphabet() | absl::CharSet::AsciiWhitespace();
-//
-// To check membership, use the .contains method, e.g.
-//
-//   absl::CharSet hex_letters("abcdef");
-//   hex_letters.contains('a');  // true
-//   hex_letters.contains('g');  // false
-
-#ifndef ABSL_STRINGS_CHARSET_H_
-#define ABSL_STRINGS_CHARSET_H_
-
-#include <cstddef>
-#include <cstdint>
-#include <cstring>
-
-#include "absl/base/macros.h"
-#include "absl/base/port.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-
-class CharSet {
- public:
-  constexpr CharSet() : m_() {}
-
-  // Initializes with a given string_view.
-  constexpr explicit CharSet(absl::string_view str) : m_() {
-    for (char c : str) {
-      SetChar(static_cast<unsigned char>(c));
-    }
-  }
-
-  constexpr bool contains(char c) const {
-    return ((m_[static_cast<unsigned char>(c) / 64] >>
-             (static_cast<unsigned char>(c) % 64)) &
-            0x1) == 0x1;
-  }
-
-  constexpr bool empty() const {
-    for (uint64_t c : m_) {
-      if (c != 0) return false;
-    }
-    return true;
-  }
-
-  // Containing only a single specified char.
-  static constexpr CharSet Char(char x) {
-    return CharSet(CharMaskForWord(x, 0), CharMaskForWord(x, 1),
-                   CharMaskForWord(x, 2), CharMaskForWord(x, 3));
-  }
-
-  // Containing all the chars in the closed interval [lo,hi].
-  static constexpr CharSet Range(char lo, char hi) {
-    return CharSet(RangeForWord(lo, hi, 0), RangeForWord(lo, hi, 1),
-                   RangeForWord(lo, hi, 2), RangeForWord(lo, hi, 3));
-  }
-
-  friend constexpr CharSet operator&(const CharSet& a, const CharSet& b) {
-    return CharSet(a.m_[0] & b.m_[0], a.m_[1] & b.m_[1], a.m_[2] & b.m_[2],
-                   a.m_[3] & b.m_[3]);
-  }
-
-  friend constexpr CharSet operator|(const CharSet& a, const CharSet& b) {
-    return CharSet(a.m_[0] | b.m_[0], a.m_[1] | b.m_[1], a.m_[2] | b.m_[2],
-                   a.m_[3] | b.m_[3]);
-  }
-
-  friend constexpr CharSet operator~(const CharSet& a) {
-    return CharSet(~a.m_[0], ~a.m_[1], ~a.m_[2], ~a.m_[3]);
-  }
-
-  // Mirrors the char-classifying predicates in <cctype>.
-  static constexpr CharSet AsciiUppercase() { return CharSet::Range('A', 'Z'); }
-  static constexpr CharSet AsciiLowercase() { return CharSet::Range('a', 'z'); }
-  static constexpr CharSet AsciiDigits() { return CharSet::Range('0', '9'); }
-  static constexpr CharSet AsciiAlphabet() {
-    return AsciiLowercase() | AsciiUppercase();
-  }
-  static constexpr CharSet AsciiAlphanumerics() {
-    return AsciiDigits() | AsciiAlphabet();
-  }
-  static constexpr CharSet AsciiHexDigits() {
-    return AsciiDigits() | CharSet::Range('A', 'F') | CharSet::Range('a', 'f');
-  }
-  static constexpr CharSet AsciiPrintable() {
-    return CharSet::Range(0x20, 0x7e);
-  }
-  static constexpr CharSet AsciiWhitespace() { return CharSet("\t\n\v\f\r "); }
-  static constexpr CharSet AsciiPunctuation() {
-    return AsciiPrintable() & ~AsciiWhitespace() & ~AsciiAlphanumerics();
-  }
-
- private:
-  constexpr CharSet(uint64_t b0, uint64_t b1, uint64_t b2, uint64_t b3)
-      : m_{b0, b1, b2, b3} {}
-
-  static constexpr uint64_t RangeForWord(char lo, char hi, uint64_t word) {
-    return OpenRangeFromZeroForWord(static_cast<unsigned char>(hi) + 1, word) &
-           ~OpenRangeFromZeroForWord(static_cast<unsigned char>(lo), word);
-  }
-
-  // All the chars in the specified word of the range [0, upper).
-  static constexpr uint64_t OpenRangeFromZeroForWord(uint64_t upper,
-                                                     uint64_t word) {
-    return (upper <= 64 * word) ? 0
-           : (upper >= 64 * (word + 1))
-               ? ~static_cast<uint64_t>(0)
-               : (~static_cast<uint64_t>(0) >> (64 - upper % 64));
-  }
-
-  static constexpr uint64_t CharMaskForWord(char x, uint64_t word) {
-    return (static_cast<unsigned char>(x) / 64 == word)
-               ? (static_cast<uint64_t>(1)
-                  << (static_cast<unsigned char>(x) % 64))
-               : 0;
-  }
-
-  constexpr void SetChar(unsigned char c) {
-    m_[c / 64] |= static_cast<uint64_t>(1) << (c % 64);
-  }
-
-  uint64_t m_[4];
-};
-
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_CHARSET_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/cord.h b/projects/llm_framework/include/abseil-cpp/absl/strings/cord.h
deleted file mode 100644
index 2583aa8a..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/cord.h
+++ /dev/null
@@ -1,1707 +0,0 @@
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: cord.h
-// -----------------------------------------------------------------------------
-//
-// This file defines the `absl::Cord` data structure and operations on that data
-// structure. A Cord is a string-like sequence of characters optimized for
-// specific use cases. Unlike a `std::string`, which stores an array of
-// contiguous characters, Cord data is stored in a structure consisting of
-// separate, reference-counted "chunks."
-//
-// Because a Cord consists of these chunks, data can be added to or removed from
-// a Cord during its lifetime. Chunks may also be shared between Cords. Unlike a
-// `std::string`, a Cord can therefore accommodate data that changes over its
-// lifetime, though it's not quite "mutable"; it can change only in the
-// attachment, detachment, or rearrangement of chunks of its constituent data.
-//
-// A Cord provides some benefit over `std::string` under the following (albeit
-// narrow) circumstances:
-//
-//   * Cord data is designed to grow and shrink over a Cord's lifetime. Cord
-//     provides efficient insertions and deletions at the start and end of the
-//     character sequences, avoiding copies in those cases. Static data should
-//     generally be stored as strings.
-//   * External memory consisting of string-like data can be directly added to
-//     a Cord without requiring copies or allocations.
-//   * Cord data may be shared and copied cheaply. Cord provides a copy-on-write
-//     implementation and cheap sub-Cord operations. Copying a Cord is an O(1)
-//     operation.
-//
-// As a consequence to the above, Cord data is generally large. Small data
-// should generally use strings, as construction of a Cord requires some
-// overhead. Small Cords (<= 15 bytes) are represented inline, but most small
-// Cords are expected to grow over their lifetimes.
-//
-// Note that because a Cord is made up of separate chunked data, random access
-// to character data within a Cord is slower than within a `std::string`.
-//
-// Thread Safety
-//
-// Cord has the same thread-safety properties as many other types like
-// std::string, std::vector<>, int, etc -- it is thread-compatible. In
-// particular, if threads do not call non-const methods, then it is safe to call
-// const methods without synchronization. Copying a Cord produces a new instance
-// that can be used concurrently with the original in arbitrary ways.
-
-#ifndef ABSL_STRINGS_CORD_H_
-#define ABSL_STRINGS_CORD_H_
-
-#include <algorithm>
-#include <cstddef>
-#include <cstdint>
-#include <cstring>
-#include <iosfwd>
-#include <iterator>
-#include <string>
-#include <type_traits>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/internal/endian.h"
-#include "absl/base/internal/per_thread_tls.h"
-#include "absl/base/macros.h"
-#include "absl/base/nullability.h"
-#include "absl/base/port.h"
-#include "absl/container/inlined_vector.h"
-#include "absl/crc/internal/crc_cord_state.h"
-#include "absl/functional/function_ref.h"
-#include "absl/meta/type_traits.h"
-#include "absl/strings/cord_analysis.h"
-#include "absl/strings/cord_buffer.h"
-#include "absl/strings/internal/cord_data_edge.h"
-#include "absl/strings/internal/cord_internal.h"
-#include "absl/strings/internal/cord_rep_btree.h"
-#include "absl/strings/internal/cord_rep_btree_reader.h"
-#include "absl/strings/internal/cord_rep_crc.h"
-#include "absl/strings/internal/cordz_functions.h"
-#include "absl/strings/internal/cordz_info.h"
-#include "absl/strings/internal/cordz_statistics.h"
-#include "absl/strings/internal/cordz_update_scope.h"
-#include "absl/strings/internal/cordz_update_tracker.h"
-#include "absl/strings/internal/resize_uninitialized.h"
-#include "absl/strings/internal/string_constant.h"
-#include "absl/strings/string_view.h"
-#include "absl/types/optional.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-class Cord;
-class CordTestPeer;
-template <typename Releaser>
-Cord MakeCordFromExternal(absl::string_view, Releaser&&);
-void CopyCordToString(const Cord& src, absl::Nonnull<std::string*> dst);
-
-// Cord memory accounting modes
-enum class CordMemoryAccounting {
-  // Counts the *approximate* number of bytes held in full or in part by this
-  // Cord (which may not remain the same between invocations). Cords that share
-  // memory could each be "charged" independently for the same shared memory.
-  // See also comment on `kTotalMorePrecise` on internally shared memory.
-  kTotal,
-
-  // Counts the *approximate* number of bytes held in full or in part by this
-  // Cord for the distinct memory held by this cord. This option is similar
-  // to `kTotal`, except that if the cord has multiple references to the same
-  // memory, that memory is only counted once.
-  //
-  // For example:
-  //   absl::Cord cord;
-  //   cord.Append(some_other_cord);
-  //   cord.Append(some_other_cord);
-  //   // Counts `some_other_cord` twice:
-  //   cord.EstimatedMemoryUsage(kTotal);
-  //   // Counts `some_other_cord` once:
-  //   cord.EstimatedMemoryUsage(kTotalMorePrecise);
-  //
-  // The `kTotalMorePrecise` number is more expensive to compute as it requires
-  // deduplicating all memory references. Applications should prefer to use
-  // `kFairShare` or `kTotal` unless they really need a more precise estimate
-  // on "how much memory is potentially held / kept alive by this cord?"
-  kTotalMorePrecise,
-
-  // Counts the *approximate* number of bytes held in full or in part by this
-  // Cord weighted by the sharing ratio of that data. For example, if some data
-  // edge is shared by 4 different Cords, then each cord is attributed 1/4th of
-  // the total memory usage as a 'fair share' of the total memory usage.
-  kFairShare,
-};
-
-// Cord
-//
-// A Cord is a sequence of characters, designed to be more efficient than a
-// `std::string` in certain circumstances: namely, large string data that needs
-// to change over its lifetime or shared, especially when such data is shared
-// across API boundaries.
-//
-// A Cord stores its character data in a structure that allows efficient prepend
-// and append operations. This makes a Cord useful for large string data sent
-// over in a wire format that may need to be prepended or appended at some point
-// during the data exchange (e.g. HTTP, protocol buffers). For example, a
-// Cord is useful for storing an HTTP request, and prepending an HTTP header to
-// such a request.
-//
-// Cords should not be used for storing general string data, however. They
-// require overhead to construct and are slower than strings for random access.
-//
-// The Cord API provides the following common API operations:
-//
-// * Create or assign Cords out of existing string data, memory, or other Cords
-// * Append and prepend data to an existing Cord
-// * Create new Sub-Cords from existing Cord data
-// * Swap Cord data and compare Cord equality
-// * Write out Cord data by constructing a `std::string`
-//
-// Additionally, the API provides iterator utilities to iterate through Cord
-// data via chunks or character bytes.
-//
-class Cord {
- private:
-  template <typename T>
-  using EnableIfString =
-      absl::enable_if_t<std::is_same<T, std::string>::value, int>;
-
- public:
-  // Cord::Cord() Constructors.
-
-  // Creates an empty Cord.
-  constexpr Cord() noexcept;
-
-  // Creates a Cord from an existing Cord. Cord is copyable and efficiently
-  // movable. The moved-from state is valid but unspecified.
-  Cord(const Cord& src);
-  Cord(Cord&& src) noexcept;
-  Cord& operator=(const Cord& x);
-  Cord& operator=(Cord&& x) noexcept;
-
-  // Creates a Cord from a `src` string. This constructor is marked explicit to
-  // prevent implicit Cord constructions from arguments convertible to an
-  // `absl::string_view`.
-  explicit Cord(absl::string_view src);
-  Cord& operator=(absl::string_view src);
-
-  // Creates a Cord from a `std::string&&` rvalue. These constructors are
-  // templated to avoid ambiguities for types that are convertible to both
-  // `absl::string_view` and `std::string`, such as `const char*`.
-  template <typename T, EnableIfString<T> = 0>
-  explicit Cord(T&& src);
-  template <typename T, EnableIfString<T> = 0>
-  Cord& operator=(T&& src);
-
-  // Cord::~Cord()
-  //
-  // Destructs the Cord.
-  ~Cord() {
-    if (contents_.is_tree()) DestroyCordSlow();
-  }
-
-  // MakeCordFromExternal()
-  //
-  // Creates a Cord that takes ownership of external string memory. The
-  // contents of `data` are not copied to the Cord; instead, the external
-  // memory is added to the Cord and reference-counted. This data may not be
-  // changed for the life of the Cord, though it may be prepended or appended
-  // to.
-  //
-  // `MakeCordFromExternal()` takes a callable "releaser" that is invoked when
-  // the reference count for `data` reaches zero. As noted above, this data must
-  // remain live until the releaser is invoked. The callable releaser also must:
-  //
-  //   * be move constructible
-  //   * support `void operator()(absl::string_view) const` or `void operator()`
-  //
-  // Example:
-  //
-  // Cord MakeCord(BlockPool* pool) {
-  //   Block* block = pool->NewBlock();
-  //   FillBlock(block);
-  //   return absl::MakeCordFromExternal(
-  //       block->ToStringView(),
-  //       [pool, block](absl::string_view v) {
-  //         pool->FreeBlock(block, v);
-  //       });
-  // }
-  //
-  // WARNING: Because a Cord can be reference-counted, it's likely a bug if your
-  // releaser doesn't do anything. For example, consider the following:
-  //
-  // void Foo(const char* buffer, int len) {
-  //   auto c = absl::MakeCordFromExternal(absl::string_view(buffer, len),
-  //                                       [](absl::string_view) {});
-  //
-  //   // BUG: If Bar() copies its cord for any reason, including keeping a
-  //   // substring of it, the lifetime of buffer might be extended beyond
-  //   // when Foo() returns.
-  //   Bar(c);
-  // }
-  template <typename Releaser>
-  friend Cord MakeCordFromExternal(absl::string_view data, Releaser&& releaser);
-
-  // Cord::Clear()
-  //
-  // Releases the Cord data. Any nodes that share data with other Cords, if
-  // applicable, will have their reference counts reduced by 1.
-  ABSL_ATTRIBUTE_REINITIALIZES void Clear();
-
-  // Cord::Append()
-  //
-  // Appends data to the Cord, which may come from another Cord or other string
-  // data.
-  void Append(const Cord& src);
-  void Append(Cord&& src);
-  void Append(absl::string_view src);
-  template <typename T, EnableIfString<T> = 0>
-  void Append(T&& src);
-
-  // Appends `buffer` to this cord, unless `buffer` has a zero length in which
-  // case this method has no effect on this cord instance.
-  // This method is guaranteed to consume `buffer`.
-  void Append(CordBuffer buffer);
-
-  // Returns a CordBuffer, re-using potential existing capacity in this cord.
-  //
-  // Cord instances may have additional unused capacity in the last (or first)
-  // nodes of the underlying tree to facilitate amortized growth. This method
-  // allows applications to explicitly use this spare capacity if available,
-  // or create a new CordBuffer instance otherwise.
-  // If this cord has a final non-shared node with at least `min_capacity`
-  // available, then this method will return that buffer including its data
-  // contents. I.e.; the returned buffer will have a non-zero length, and
-  // a capacity of at least `buffer.length + min_capacity`. Otherwise, this
-  // method will return `CordBuffer::CreateWithDefaultLimit(capacity)`.
-  //
-  // Below an example of using GetAppendBuffer. Notice that in this example we
-  // use `GetAppendBuffer()` only on the first iteration. As we know nothing
-  // about any initial extra capacity in `cord`, we may be able to use the extra
-  // capacity. But as we add new buffers with fully utilized contents after that
-  // we avoid calling `GetAppendBuffer()` on subsequent iterations: while this
-  // works fine, it results in an unnecessary inspection of cord contents:
-  //
-  //   void AppendRandomDataToCord(absl::Cord &cord, size_t n) {
-  //     bool first = true;
-  //     while (n > 0) {
-  //       CordBuffer buffer = first ? cord.GetAppendBuffer(n)
-  //                                 : CordBuffer::CreateWithDefaultLimit(n);
-  //       absl::Span<char> data = buffer.available_up_to(n);
-  //       FillRandomValues(data.data(), data.size());
-  //       buffer.IncreaseLengthBy(data.size());
-  //       cord.Append(std::move(buffer));
-  //       n -= data.size();
-  //       first = false;
-  //     }
-  //   }
-  CordBuffer GetAppendBuffer(size_t capacity, size_t min_capacity = 16);
-
-  // Returns a CordBuffer, re-using potential existing capacity in this cord.
-  //
-  // This function is identical to `GetAppendBuffer`, except that in the case
-  // where a new `CordBuffer` is allocated, it is allocated using the provided
-  // custom limit instead of the default limit. `GetAppendBuffer` will default
-  // to `CordBuffer::CreateWithDefaultLimit(capacity)` whereas this method
-  // will default to `CordBuffer::CreateWithCustomLimit(block_size, capacity)`.
-  // This method is equivalent to `GetAppendBuffer` if `block_size` is zero.
-  // See the documentation for `CreateWithCustomLimit` for more details on the
-  // restrictions and legal values for `block_size`.
-  CordBuffer GetCustomAppendBuffer(size_t block_size, size_t capacity,
-                                   size_t min_capacity = 16);
-
-  // Cord::Prepend()
-  //
-  // Prepends data to the Cord, which may come from another Cord or other string
-  // data.
-  void Prepend(const Cord& src);
-  void Prepend(absl::string_view src);
-  template <typename T, EnableIfString<T> = 0>
-  void Prepend(T&& src);
-
-  // Prepends `buffer` to this cord, unless `buffer` has a zero length in which
-  // case this method has no effect on this cord instance.
-  // This method is guaranteed to consume `buffer`.
-  void Prepend(CordBuffer buffer);
-
-  // Cord::RemovePrefix()
-  //
-  // Removes the first `n` bytes of a Cord.
-  void RemovePrefix(size_t n);
-  void RemoveSuffix(size_t n);
-
-  // Cord::Subcord()
-  //
-  // Returns a new Cord representing the subrange [pos, pos + new_size) of
-  // *this. If pos >= size(), the result is empty(). If
-  // (pos + new_size) >= size(), the result is the subrange [pos, size()).
-  Cord Subcord(size_t pos, size_t new_size) const;
-
-  // Cord::swap()
-  //
-  // Swaps the contents of the Cord with `other`.
-  void swap(Cord& other) noexcept;
-
-  // swap()
-  //
-  // Swaps the contents of two Cords.
-  friend void swap(Cord& x, Cord& y) noexcept { x.swap(y); }
-
-  // Cord::size()
-  //
-  // Returns the size of the Cord.
-  size_t size() const;
-
-  // Cord::empty()
-  //
-  // Determines whether the given Cord is empty, returning `true` if so.
-  bool empty() const;
-
-  // Cord::EstimatedMemoryUsage()
-  //
-  // Returns the *approximate* number of bytes held by this cord.
-  // See CordMemoryAccounting for more information on the accounting method.
-  size_t EstimatedMemoryUsage(CordMemoryAccounting accounting_method =
-                                  CordMemoryAccounting::kTotal) const;
-
-  // Cord::Compare()
-  //
-  // Compares 'this' Cord with rhs. This function and its relatives treat Cords
-  // as sequences of unsigned bytes. The comparison is a straightforward
-  // lexicographic comparison. `Cord::Compare()` returns values as follows:
-  //
-  //   -1  'this' Cord is smaller
-  //    0  two Cords are equal
-  //    1  'this' Cord is larger
-  int Compare(absl::string_view rhs) const;
-  int Compare(const Cord& rhs) const;
-
-  // Cord::StartsWith()
-  //
-  // Determines whether the Cord starts with the passed string data `rhs`.
-  bool StartsWith(const Cord& rhs) const;
-  bool StartsWith(absl::string_view rhs) const;
-
-  // Cord::EndsWith()
-  //
-  // Determines whether the Cord ends with the passed string data `rhs`.
-  bool EndsWith(absl::string_view rhs) const;
-  bool EndsWith(const Cord& rhs) const;
-
-  // Cord::Contains()
-  //
-  // Determines whether the Cord contains the passed string data `rhs`.
-  bool Contains(absl::string_view rhs) const;
-  bool Contains(const Cord& rhs) const;
-
-  // Cord::operator std::string()
-  //
-  // Converts a Cord into a `std::string()`. This operator is marked explicit to
-  // prevent unintended Cord usage in functions that take a string.
-  explicit operator std::string() const;
-
-  // CopyCordToString()
-  //
-  // Copies the contents of a `src` Cord into a `*dst` string.
-  //
-  // This function optimizes the case of reusing the destination string since it
-  // can reuse previously allocated capacity. However, this function does not
-  // guarantee that pointers previously returned by `dst->data()` remain valid
-  // even if `*dst` had enough capacity to hold `src`. If `*dst` is a new
-  // object, prefer to simply use the conversion operator to `std::string`.
-  friend void CopyCordToString(const Cord& src,
-                               absl::Nonnull<std::string*> dst);
-
-  class CharIterator;
-
-  //----------------------------------------------------------------------------
-  // Cord::ChunkIterator
-  //----------------------------------------------------------------------------
-  //
-  // A `Cord::ChunkIterator` allows iteration over the constituent chunks of its
-  // Cord. Such iteration allows you to perform non-const operations on the data
-  // of a Cord without modifying it.
-  //
-  // Generally, you do not instantiate a `Cord::ChunkIterator` directly;
-  // instead, you create one implicitly through use of the `Cord::Chunks()`
-  // member function.
-  //
-  // The `Cord::ChunkIterator` has the following properties:
-  //
-  //   * The iterator is invalidated after any non-const operation on the
-  //     Cord object over which it iterates.
-  //   * The `string_view` returned by dereferencing a valid, non-`end()`
-  //     iterator is guaranteed to be non-empty.
-  //   * Two `ChunkIterator` objects can be compared equal if and only if they
-  //     remain valid and iterate over the same Cord.
-  //   * The iterator in this case is a proxy iterator; the `string_view`
-  //     returned by the iterator does not live inside the Cord, and its
-  //     lifetime is limited to the lifetime of the iterator itself. To help
-  //     prevent lifetime issues, `ChunkIterator::reference` is not a true
-  //     reference type and is equivalent to `value_type`.
-  //   * The iterator keeps state that can grow for Cords that contain many
-  //     nodes and are imbalanced due to sharing. Prefer to pass this type by
-  //     const reference instead of by value.
-  class ChunkIterator {
-   public:
-    using iterator_category = std::input_iterator_tag;
-    using value_type = absl::string_view;
-    using difference_type = ptrdiff_t;
-    using pointer = absl::Nonnull<const value_type*>;
-    using reference = value_type;
-
-    ChunkIterator() = default;
-
-    ChunkIterator& operator++();
-    ChunkIterator operator++(int);
-    bool operator==(const ChunkIterator& other) const;
-    bool operator!=(const ChunkIterator& other) const;
-    reference operator*() const;
-    pointer operator->() const;
-
-    friend class Cord;
-    friend class CharIterator;
-
-   private:
-    using CordRep = absl::cord_internal::CordRep;
-    using CordRepBtree = absl::cord_internal::CordRepBtree;
-    using CordRepBtreeReader = absl::cord_internal::CordRepBtreeReader;
-
-    // Constructs a `begin()` iterator from `tree`.
-    explicit ChunkIterator(absl::Nonnull<cord_internal::CordRep*> tree);
-
-    // Constructs a `begin()` iterator from `cord`.
-    explicit ChunkIterator(absl::Nonnull<const Cord*> cord);
-
-    // Initializes this instance from a tree. Invoked by constructors.
-    void InitTree(absl::Nonnull<cord_internal::CordRep*> tree);
-
-    // Removes `n` bytes from `current_chunk_`. Expects `n` to be smaller than
-    // `current_chunk_.size()`.
-    void RemoveChunkPrefix(size_t n);
-    Cord AdvanceAndReadBytes(size_t n);
-    void AdvanceBytes(size_t n);
-
-    // Btree specific operator++
-    ChunkIterator& AdvanceBtree();
-    void AdvanceBytesBtree(size_t n);
-
-    // A view into bytes of the current `CordRep`. It may only be a view to a
-    // suffix of bytes if this is being used by `CharIterator`.
-    absl::string_view current_chunk_;
-    // The current leaf, or `nullptr` if the iterator points to short data.
-    // If the current chunk is a substring node, current_leaf_ points to the
-    // underlying flat or external node.
-    absl::Nullable<absl::cord_internal::CordRep*> current_leaf_ = nullptr;
-    // The number of bytes left in the `Cord` over which we are iterating.
-    size_t bytes_remaining_ = 0;
-
-    // Cord reader for cord btrees. Empty if not traversing a btree.
-    CordRepBtreeReader btree_reader_;
-  };
-
-  // Cord::chunk_begin()
-  //
-  // Returns an iterator to the first chunk of the `Cord`.
-  //
-  // Generally, prefer using `Cord::Chunks()` within a range-based for loop for
-  // iterating over the chunks of a Cord. This method may be useful for getting
-  // a `ChunkIterator` where range-based for-loops are not useful.
-  //
-  // Example:
-  //
-  //   absl::Cord::ChunkIterator FindAsChunk(const absl::Cord& c,
-  //                                         absl::string_view s) {
-  //     return std::find(c.chunk_begin(), c.chunk_end(), s);
-  //   }
-  ChunkIterator chunk_begin() const ABSL_ATTRIBUTE_LIFETIME_BOUND;
-
-  // Cord::chunk_end()
-  //
-  // Returns an iterator one increment past the last chunk of the `Cord`.
-  //
-  // Generally, prefer using `Cord::Chunks()` within a range-based for loop for
-  // iterating over the chunks of a Cord. This method may be useful for getting
-  // a `ChunkIterator` where range-based for-loops may not be available.
-  ChunkIterator chunk_end() const ABSL_ATTRIBUTE_LIFETIME_BOUND;
-
-  //----------------------------------------------------------------------------
-  // Cord::ChunkRange
-  //----------------------------------------------------------------------------
-  //
-  // `ChunkRange` is a helper class for iterating over the chunks of the `Cord`,
-  // producing an iterator which can be used within a range-based for loop.
-  // Construction of a `ChunkRange` will return an iterator pointing to the
-  // first chunk of the Cord. Generally, do not construct a `ChunkRange`
-  // directly; instead, prefer to use the `Cord::Chunks()` method.
-  //
-  // Implementation note: `ChunkRange` is simply a convenience wrapper over
-  // `Cord::chunk_begin()` and `Cord::chunk_end()`.
-  class ChunkRange {
-   public:
-    // Fulfill minimum c++ container requirements [container.requirements]
-    // These (partial) container type definitions allow ChunkRange to be used
-    // in various utilities expecting a subset of [container.requirements].
-    // For example, the below enables using `::testing::ElementsAre(...)`
-    using value_type = absl::string_view;
-    using reference = value_type&;
-    using const_reference = const value_type&;
-    using iterator = ChunkIterator;
-    using const_iterator = ChunkIterator;
-
-    explicit ChunkRange(absl::Nonnull<const Cord*> cord) : cord_(cord) {}
-
-    ChunkIterator begin() const;
-    ChunkIterator end() const;
-
-   private:
-    absl::Nonnull<const Cord*> cord_;
-  };
-
-  // Cord::Chunks()
-  //
-  // Returns a `Cord::ChunkRange` for iterating over the chunks of a `Cord` with
-  // a range-based for-loop. For most iteration tasks on a Cord, use
-  // `Cord::Chunks()` to retrieve this iterator.
-  //
-  // Example:
-  //
-  //   void ProcessChunks(const Cord& cord) {
-  //     for (absl::string_view chunk : cord.Chunks()) { ... }
-  //   }
-  //
-  // Note that the ordinary caveats of temporary lifetime extension apply:
-  //
-  //   void Process() {
-  //     for (absl::string_view chunk : CordFactory().Chunks()) {
-  //       // The temporary Cord returned by CordFactory has been destroyed!
-  //     }
-  //   }
-  ChunkRange Chunks() const ABSL_ATTRIBUTE_LIFETIME_BOUND;
-
-  //----------------------------------------------------------------------------
-  // Cord::CharIterator
-  //----------------------------------------------------------------------------
-  //
-  // A `Cord::CharIterator` allows iteration over the constituent characters of
-  // a `Cord`.
-  //
-  // Generally, you do not instantiate a `Cord::CharIterator` directly; instead,
-  // you create one implicitly through use of the `Cord::Chars()` member
-  // function.
-  //
-  // A `Cord::CharIterator` has the following properties:
-  //
-  //   * The iterator is invalidated after any non-const operation on the
-  //     Cord object over which it iterates.
-  //   * Two `CharIterator` objects can be compared equal if and only if they
-  //     remain valid and iterate over the same Cord.
-  //   * The iterator keeps state that can grow for Cords that contain many
-  //     nodes and are imbalanced due to sharing. Prefer to pass this type by
-  //     const reference instead of by value.
-  //   * This type cannot act as a forward iterator because a `Cord` can reuse
-  //     sections of memory. This fact violates the requirement for forward
-  //     iterators to compare equal if dereferencing them returns the same
-  //     object.
-  class CharIterator {
-   public:
-    using iterator_category = std::input_iterator_tag;
-    using value_type = char;
-    using difference_type = ptrdiff_t;
-    using pointer = absl::Nonnull<const char*>;
-    using reference = const char&;
-
-    CharIterator() = default;
-
-    CharIterator& operator++();
-    CharIterator operator++(int);
-    bool operator==(const CharIterator& other) const;
-    bool operator!=(const CharIterator& other) const;
-    reference operator*() const;
-    pointer operator->() const;
-
-    friend Cord;
-
-   private:
-    explicit CharIterator(absl::Nonnull<const Cord*> cord)
-        : chunk_iterator_(cord) {}
-
-    ChunkIterator chunk_iterator_;
-  };
-
-  // Cord::AdvanceAndRead()
-  //
-  // Advances the `Cord::CharIterator` by `n_bytes` and returns the bytes
-  // advanced as a separate `Cord`. `n_bytes` must be less than or equal to the
-  // number of bytes within the Cord; otherwise, behavior is undefined. It is
-  // valid to pass `char_end()` and `0`.
-  static Cord AdvanceAndRead(absl::Nonnull<CharIterator*> it, size_t n_bytes);
-
-  // Cord::Advance()
-  //
-  // Advances the `Cord::CharIterator` by `n_bytes`. `n_bytes` must be less than
-  // or equal to the number of bytes remaining within the Cord; otherwise,
-  // behavior is undefined. It is valid to pass `char_end()` and `0`.
-  static void Advance(absl::Nonnull<CharIterator*> it, size_t n_bytes);
-
-  // Cord::ChunkRemaining()
-  //
-  // Returns the longest contiguous view starting at the iterator's position.
-  //
-  // `it` must be dereferenceable.
-  static absl::string_view ChunkRemaining(const CharIterator& it);
-
-  // Cord::char_begin()
-  //
-  // Returns an iterator to the first character of the `Cord`.
-  //
-  // Generally, prefer using `Cord::Chars()` within a range-based for loop for
-  // iterating over the chunks of a Cord. This method may be useful for getting
-  // a `CharIterator` where range-based for-loops may not be available.
-  CharIterator char_begin() const ABSL_ATTRIBUTE_LIFETIME_BOUND;
-
-  // Cord::char_end()
-  //
-  // Returns an iterator to one past the last character of the `Cord`.
-  //
-  // Generally, prefer using `Cord::Chars()` within a range-based for loop for
-  // iterating over the chunks of a Cord. This method may be useful for getting
-  // a `CharIterator` where range-based for-loops are not useful.
-  CharIterator char_end() const ABSL_ATTRIBUTE_LIFETIME_BOUND;
-
-  // Cord::CharRange
-  //
-  // `CharRange` is a helper class for iterating over the characters of a
-  // producing an iterator which can be used within a range-based for loop.
-  // Construction of a `CharRange` will return an iterator pointing to the first
-  // character of the Cord. Generally, do not construct a `CharRange` directly;
-  // instead, prefer to use the `Cord::Chars()` method shown below.
-  //
-  // Implementation note: `CharRange` is simply a convenience wrapper over
-  // `Cord::char_begin()` and `Cord::char_end()`.
-  class CharRange {
-   public:
-    // Fulfill minimum c++ container requirements [container.requirements]
-    // These (partial) container type definitions allow CharRange to be used
-    // in various utilities expecting a subset of [container.requirements].
-    // For example, the below enables using `::testing::ElementsAre(...)`
-    using value_type = char;
-    using reference = value_type&;
-    using const_reference = const value_type&;
-    using iterator = CharIterator;
-    using const_iterator = CharIterator;
-
-    explicit CharRange(absl::Nonnull<const Cord*> cord) : cord_(cord) {}
-
-    CharIterator begin() const;
-    CharIterator end() const;
-
-   private:
-    absl::Nonnull<const Cord*> cord_;
-  };
-
-  // Cord::Chars()
-  //
-  // Returns a `Cord::CharRange` for iterating over the characters of a `Cord`
-  // with a range-based for-loop. For most character-based iteration tasks on a
-  // Cord, use `Cord::Chars()` to retrieve this iterator.
-  //
-  // Example:
-  //
-  //   void ProcessCord(const Cord& cord) {
-  //     for (char c : cord.Chars()) { ... }
-  //   }
-  //
-  // Note that the ordinary caveats of temporary lifetime extension apply:
-  //
-  //   void Process() {
-  //     for (char c : CordFactory().Chars()) {
-  //       // The temporary Cord returned by CordFactory has been destroyed!
-  //     }
-  //   }
-  CharRange Chars() const ABSL_ATTRIBUTE_LIFETIME_BOUND;
-
-  // Cord::operator[]
-  //
-  // Gets the "i"th character of the Cord and returns it, provided that
-  // 0 <= i < Cord.size().
-  //
-  // NOTE: This routine is reasonably efficient. It is roughly
-  // logarithmic based on the number of chunks that make up the cord. Still,
-  // if you need to iterate over the contents of a cord, you should
-  // use a CharIterator/ChunkIterator rather than call operator[] or Get()
-  // repeatedly in a loop.
-  char operator[](size_t i) const;
-
-  // Cord::TryFlat()
-  //
-  // If this cord's representation is a single flat array, returns a
-  // string_view referencing that array.  Otherwise returns nullopt.
-  absl::optional<absl::string_view> TryFlat() const
-      ABSL_ATTRIBUTE_LIFETIME_BOUND;
-
-  // Cord::Flatten()
-  //
-  // Flattens the cord into a single array and returns a view of the data.
-  //
-  // If the cord was already flat, the contents are not modified.
-  absl::string_view Flatten() ABSL_ATTRIBUTE_LIFETIME_BOUND;
-
-  // Cord::Find()
-  //
-  // Returns an iterator to the first occurrance of the substring `needle`.
-  //
-  // If the substring `needle` does not occur, `Cord::char_end()` is returned.
-  CharIterator Find(absl::string_view needle) const;
-  CharIterator Find(const absl::Cord& needle) const;
-
-  // Supports absl::Cord as a sink object for absl::Format().
-  friend void AbslFormatFlush(absl::Nonnull<absl::Cord*> cord,
-                              absl::string_view part) {
-    cord->Append(part);
-  }
-
-  // Support automatic stringification with absl::StrCat and absl::StrFormat.
-  template <typename Sink>
-  friend void AbslStringify(Sink& sink, const absl::Cord& cord) {
-    for (absl::string_view chunk : cord.Chunks()) {
-      sink.Append(chunk);
-    }
-  }
-
-  // Cord::SetExpectedChecksum()
-  //
-  // Stores a checksum value with this non-empty cord instance, for later
-  // retrieval.
-  //
-  // The expected checksum is a number stored out-of-band, alongside the data.
-  // It is preserved across copies and assignments, but any mutations to a cord
-  // will cause it to lose its expected checksum.
-  //
-  // The expected checksum is not part of a Cord's value, and does not affect
-  // operations such as equality or hashing.
-  //
-  // This field is intended to store a CRC32C checksum for later validation, to
-  // help support end-to-end checksum workflows.  However, the Cord API itself
-  // does no CRC validation, and assigns no meaning to this number.
-  //
-  // This call has no effect if this cord is empty.
-  void SetExpectedChecksum(uint32_t crc);
-
-  // Returns this cord's expected checksum, if it has one.  Otherwise, returns
-  // nullopt.
-  absl::optional<uint32_t> ExpectedChecksum() const;
-
-  template <typename H>
-  friend H AbslHashValue(H hash_state, const absl::Cord& c) {
-    absl::optional<absl::string_view> maybe_flat = c.TryFlat();
-    if (maybe_flat.has_value()) {
-      return H::combine(std::move(hash_state), *maybe_flat);
-    }
-    return c.HashFragmented(std::move(hash_state));
-  }
-
-  // Create a Cord with the contents of StringConstant<T>::value.
-  // No allocations will be done and no data will be copied.
-  // This is an INTERNAL API and subject to change or removal. This API can only
-  // be used by spelling absl::strings_internal::MakeStringConstant, which is
-  // also an internal API.
-  template <typename T>
-  // NOLINTNEXTLINE(google-explicit-constructor)
-  constexpr Cord(strings_internal::StringConstant<T>);
-
- private:
-  using CordRep = absl::cord_internal::CordRep;
-  using CordRepFlat = absl::cord_internal::CordRepFlat;
-  using CordzInfo = cord_internal::CordzInfo;
-  using CordzUpdateScope = cord_internal::CordzUpdateScope;
-  using CordzUpdateTracker = cord_internal::CordzUpdateTracker;
-  using InlineData = cord_internal::InlineData;
-  using MethodIdentifier = CordzUpdateTracker::MethodIdentifier;
-
-  // Creates a cord instance with `method` representing the originating
-  // public API call causing the cord to be created.
-  explicit Cord(absl::string_view src, MethodIdentifier method);
-
-  friend class CordTestPeer;
-  friend bool operator==(const Cord& lhs, const Cord& rhs);
-  friend bool operator==(const Cord& lhs, absl::string_view rhs);
-
-  friend absl::Nullable<const CordzInfo*> GetCordzInfoForTesting(
-      const Cord& cord);
-
-  // Calls the provided function once for each cord chunk, in order.  Unlike
-  // Chunks(), this API will not allocate memory.
-  void ForEachChunk(absl::FunctionRef<void(absl::string_view)>) const;
-
-  // Allocates new contiguous storage for the contents of the cord. This is
-  // called by Flatten() when the cord was not already flat.
-  absl::string_view FlattenSlowPath();
-
-  // Actual cord contents are hidden inside the following simple
-  // class so that we can isolate the bulk of cord.cc from changes
-  // to the representation.
-  //
-  // InlineRep holds either a tree pointer, or an array of kMaxInline bytes.
-  class InlineRep {
-   public:
-    static constexpr unsigned char kMaxInline = cord_internal::kMaxInline;
-    static_assert(kMaxInline >= sizeof(absl::cord_internal::CordRep*), "");
-
-    constexpr InlineRep() : data_() {}
-    explicit InlineRep(InlineData::DefaultInitType init) : data_(init) {}
-    InlineRep(const InlineRep& src);
-    InlineRep(InlineRep&& src);
-    InlineRep& operator=(const InlineRep& src);
-    InlineRep& operator=(InlineRep&& src) noexcept;
-
-    explicit constexpr InlineRep(absl::string_view sv,
-                                 absl::Nullable<CordRep*> rep);
-
-    void Swap(absl::Nonnull<InlineRep*> rhs);
-    size_t size() const;
-    // Returns nullptr if holding pointer
-    absl::Nullable<const char*> data() const;
-    // Discards pointer, if any
-    void set_data(absl::Nonnull<const char*> data, size_t n);
-    absl::Nonnull<char*> set_data(size_t n);  // Write data to the result
-    // Returns nullptr if holding bytes
-    absl::Nullable<absl::cord_internal::CordRep*> tree() const;
-    absl::Nonnull<absl::cord_internal::CordRep*> as_tree() const;
-    absl::Nonnull<const char*> as_chars() const;
-    // Returns non-null iff was holding a pointer
-    absl::Nullable<absl::cord_internal::CordRep*> clear();
-    // Converts to pointer if necessary.
-    void reduce_size(size_t n);    // REQUIRES: holding data
-    void remove_prefix(size_t n);  // REQUIRES: holding data
-    void AppendArray(absl::string_view src, MethodIdentifier method);
-    absl::string_view FindFlatStartPiece() const;
-
-    // Creates a CordRepFlat instance from the current inlined data with `extra'
-    // bytes of desired additional capacity.
-    absl::Nonnull<CordRepFlat*> MakeFlatWithExtraCapacity(size_t extra);
-
-    // Sets the tree value for this instance. `rep` must not be null.
-    // Requires the current instance to hold a tree, and a lock to be held on
-    // any CordzInfo referenced by this instance. The latter is enforced through
-    // the CordzUpdateScope argument. If the current instance is sampled, then
-    // the CordzInfo instance is updated to reference the new `rep` value.
-    void SetTree(absl::Nonnull<CordRep*> rep, const CordzUpdateScope& scope);
-
-    // Identical to SetTree(), except that `rep` is allowed to be null, in
-    // which case the current instance is reset to an empty value.
-    void SetTreeOrEmpty(absl::Nullable<CordRep*> rep,
-                        const CordzUpdateScope& scope);
-
-    // Sets the tree value for this instance, and randomly samples this cord.
-    // This function disregards existing contents in `data_`, and should be
-    // called when a Cord is 'promoted' from an 'uninitialized' or 'inlined'
-    // value to a non-inlined (tree / ring) value.
-    void EmplaceTree(absl::Nonnull<CordRep*> rep, MethodIdentifier method);
-
-    // Identical to EmplaceTree, except that it copies the parent stack from
-    // the provided `parent` data if the parent is sampled.
-    void EmplaceTree(absl::Nonnull<CordRep*> rep, const InlineData& parent,
-                     MethodIdentifier method);
-
-    // Commits the change of a newly created, or updated `rep` root value into
-    // this cord. `old_rep` indicates the old (inlined or tree) value of the
-    // cord, and determines if the commit invokes SetTree() or EmplaceTree().
-    void CommitTree(absl::Nullable<const CordRep*> old_rep,
-                    absl::Nonnull<CordRep*> rep, const CordzUpdateScope& scope,
-                    MethodIdentifier method);
-
-    void AppendTreeToInlined(absl::Nonnull<CordRep*> tree,
-                             MethodIdentifier method);
-    void AppendTreeToTree(absl::Nonnull<CordRep*> tree,
-                          MethodIdentifier method);
-    void AppendTree(absl::Nonnull<CordRep*> tree, MethodIdentifier method);
-    void PrependTreeToInlined(absl::Nonnull<CordRep*> tree,
-                              MethodIdentifier method);
-    void PrependTreeToTree(absl::Nonnull<CordRep*> tree,
-                           MethodIdentifier method);
-    void PrependTree(absl::Nonnull<CordRep*> tree, MethodIdentifier method);
-
-    bool IsSame(const InlineRep& other) const { return data_ == other.data_; }
-
-    void CopyTo(absl::Nonnull<std::string*> dst) const {
-      // memcpy is much faster when operating on a known size. On most supported
-      // platforms, the small string optimization is large enough that resizing
-      // to 15 bytes does not cause a memory allocation.
-      absl::strings_internal::STLStringResizeUninitialized(dst, kMaxInline);
-      data_.copy_max_inline_to(&(*dst)[0]);
-      // erase is faster than resize because the logic for memory allocation is
-      // not needed.
-      dst->erase(inline_size());
-    }
-
-    // Copies the inline contents into `dst`. Assumes the cord is not empty.
-    void CopyToArray(absl::Nonnull<char*> dst) const;
-
-    bool is_tree() const { return data_.is_tree(); }
-
-    // Returns true if the Cord is being profiled by cordz.
-    bool is_profiled() const { return data_.is_tree() && data_.is_profiled(); }
-
-    // Returns the available inlined capacity, or 0 if is_tree() == true.
-    size_t remaining_inline_capacity() const {
-      return data_.is_tree() ? 0 : kMaxInline - data_.inline_size();
-    }
-
-    // Returns the profiled CordzInfo, or nullptr if not sampled.
-    absl::Nullable<absl::cord_internal::CordzInfo*> cordz_info() const {
-      return data_.cordz_info();
-    }
-
-    // Sets the profiled CordzInfo.
-    void set_cordz_info(absl::Nonnull<cord_internal::CordzInfo*> cordz_info) {
-      assert(cordz_info != nullptr);
-      data_.set_cordz_info(cordz_info);
-    }
-
-    // Resets the current cordz_info to null / empty.
-    void clear_cordz_info() { data_.clear_cordz_info(); }
-
-   private:
-    friend class Cord;
-
-    void AssignSlow(const InlineRep& src);
-    // Unrefs the tree and stops profiling.
-    void UnrefTree();
-
-    void ResetToEmpty() { data_ = {}; }
-
-    void set_inline_size(size_t size) { data_.set_inline_size(size); }
-    size_t inline_size() const { return data_.inline_size(); }
-
-    // Empty cords that carry a checksum have a CordRepCrc node with a null
-    // child node. The code can avoid lots of special cases where it would
-    // otherwise transition from tree to inline storage if we just remove the
-    // CordRepCrc node before mutations. Must never be called inside a
-    // CordzUpdateScope since it untracks the cordz info.
-    void MaybeRemoveEmptyCrcNode();
-
-    cord_internal::InlineData data_;
-  };
-  InlineRep contents_;
-
-  // Helper for GetFlat() and TryFlat().
-  static bool GetFlatAux(absl::Nonnull<absl::cord_internal::CordRep*> rep,
-                         absl::Nonnull<absl::string_view*> fragment);
-
-  // Helper for ForEachChunk().
-  static void ForEachChunkAux(
-      absl::Nonnull<absl::cord_internal::CordRep*> rep,
-      absl::FunctionRef<void(absl::string_view)> callback);
-
-  // The destructor for non-empty Cords.
-  void DestroyCordSlow();
-
-  // Out-of-line implementation of slower parts of logic.
-  void CopyToArraySlowPath(absl::Nonnull<char*> dst) const;
-  int CompareSlowPath(absl::string_view rhs, size_t compared_size,
-                      size_t size_to_compare) const;
-  int CompareSlowPath(const Cord& rhs, size_t compared_size,
-                      size_t size_to_compare) const;
-  bool EqualsImpl(absl::string_view rhs, size_t size_to_compare) const;
-  bool EqualsImpl(const Cord& rhs, size_t size_to_compare) const;
-  int CompareImpl(const Cord& rhs) const;
-
-  template <typename ResultType, typename RHS>
-  friend ResultType GenericCompare(const Cord& lhs, const RHS& rhs,
-                                   size_t size_to_compare);
-  static absl::string_view GetFirstChunk(const Cord& c);
-  static absl::string_view GetFirstChunk(absl::string_view sv);
-
-  // Returns a new reference to contents_.tree(), or steals an existing
-  // reference if called on an rvalue.
-  absl::Nonnull<absl::cord_internal::CordRep*> TakeRep() const&;
-  absl::Nonnull<absl::cord_internal::CordRep*> TakeRep() &&;
-
-  // Helper for Append().
-  template <typename C>
-  void AppendImpl(C&& src);
-
-  // Appends / Prepends `src` to this instance, using precise sizing.
-  // This method does explicitly not attempt to use any spare capacity
-  // in any pending last added private owned flat.
-  // Requires `src` to be <= kMaxFlatLength.
-  void AppendPrecise(absl::string_view src, MethodIdentifier method);
-  void PrependPrecise(absl::string_view src, MethodIdentifier method);
-
-  CordBuffer GetAppendBufferSlowPath(size_t block_size, size_t capacity,
-                                     size_t min_capacity);
-
-  // Prepends the provided data to this instance. `method` contains the public
-  // API method for this action which is tracked for Cordz sampling purposes.
-  void PrependArray(absl::string_view src, MethodIdentifier method);
-
-  // Assigns the value in 'src' to this instance, 'stealing' its contents.
-  // Requires src.length() > kMaxBytesToCopy.
-  Cord& AssignLargeString(std::string&& src);
-
-  // Helper for AbslHashValue().
-  template <typename H>
-  H HashFragmented(H hash_state) const {
-    typename H::AbslInternalPiecewiseCombiner combiner;
-    ForEachChunk([&combiner, &hash_state](absl::string_view chunk) {
-      hash_state = combiner.add_buffer(std::move(hash_state), chunk.data(),
-                                       chunk.size());
-    });
-    return H::combine(combiner.finalize(std::move(hash_state)), size());
-  }
-
-  friend class CrcCord;
-  void SetCrcCordState(crc_internal::CrcCordState state);
-  absl::Nullable<const crc_internal::CrcCordState*> MaybeGetCrcCordState()
-      const;
-
-  CharIterator FindImpl(CharIterator it, absl::string_view needle) const;
-};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// allow a Cord to be logged
-extern std::ostream& operator<<(std::ostream& out, const Cord& cord);
-
-// ------------------------------------------------------------------
-// Internal details follow.  Clients should ignore.
-
-namespace cord_internal {
-
-// Does non-template-specific `CordRepExternal` initialization.
-// Requires `data` to be non-empty.
-void InitializeCordRepExternal(absl::string_view data,
-                               absl::Nonnull<CordRepExternal*> rep);
-
-// Creates a new `CordRep` that owns `data` and `releaser` and returns a pointer
-// to it. Requires `data` to be non-empty.
-template <typename Releaser>
-// NOLINTNEXTLINE - suppress clang-tidy raw pointer return.
-absl::Nonnull<CordRep*> NewExternalRep(absl::string_view data,
-                                       Releaser&& releaser) {
-  assert(!data.empty());
-  using ReleaserType = absl::decay_t<Releaser>;
-  CordRepExternal* rep = new CordRepExternalImpl<ReleaserType>(
-      std::forward<Releaser>(releaser), 0);
-  InitializeCordRepExternal(data, rep);
-  return rep;
-}
-
-// Overload for function reference types that dispatches using a function
-// pointer because there are no `alignof()` or `sizeof()` a function reference.
-// NOLINTNEXTLINE - suppress clang-tidy raw pointer return.
-inline absl::Nonnull<CordRep*> NewExternalRep(absl::string_view data,
-                               void (&releaser)(absl::string_view)) {
-  return NewExternalRep(data, &releaser);
-}
-
-}  // namespace cord_internal
-
-template <typename Releaser>
-Cord MakeCordFromExternal(absl::string_view data, Releaser&& releaser) {
-  Cord cord;
-  if (ABSL_PREDICT_TRUE(!data.empty())) {
-    cord.contents_.EmplaceTree(::absl::cord_internal::NewExternalRep(
-                                   data, std::forward<Releaser>(releaser)),
-                               Cord::MethodIdentifier::kMakeCordFromExternal);
-  } else {
-    using ReleaserType = absl::decay_t<Releaser>;
-    cord_internal::InvokeReleaser(
-        cord_internal::Rank1{}, ReleaserType(std::forward<Releaser>(releaser)),
-        data);
-  }
-  return cord;
-}
-
-constexpr Cord::InlineRep::InlineRep(absl::string_view sv,
-                                     absl::Nullable<CordRep*> rep)
-    : data_(sv, rep) {}
-
-inline Cord::InlineRep::InlineRep(const Cord::InlineRep& src)
-    : data_(InlineData::kDefaultInit) {
-  if (CordRep* tree = src.tree()) {
-    EmplaceTree(CordRep::Ref(tree), src.data_,
-                CordzUpdateTracker::kConstructorCord);
-  } else {
-    data_ = src.data_;
-  }
-}
-
-inline Cord::InlineRep::InlineRep(Cord::InlineRep&& src) : data_(src.data_) {
-  src.ResetToEmpty();
-}
-
-inline Cord::InlineRep& Cord::InlineRep::operator=(const Cord::InlineRep& src) {
-  if (this == &src) {
-    return *this;
-  }
-  if (!is_tree() && !src.is_tree()) {
-    data_ = src.data_;
-    return *this;
-  }
-  AssignSlow(src);
-  return *this;
-}
-
-inline Cord::InlineRep& Cord::InlineRep::operator=(
-    Cord::InlineRep&& src) noexcept {
-  if (is_tree()) {
-    UnrefTree();
-  }
-  data_ = src.data_;
-  src.ResetToEmpty();
-  return *this;
-}
-
-inline void Cord::InlineRep::Swap(absl::Nonnull<Cord::InlineRep*> rhs) {
-  if (rhs == this) {
-    return;
-  }
-  using std::swap;
-  swap(data_, rhs->data_);
-}
-
-inline absl::Nullable<const char*> Cord::InlineRep::data() const {
-  return is_tree() ? nullptr : data_.as_chars();
-}
-
-inline absl::Nonnull<const char*> Cord::InlineRep::as_chars() const {
-  assert(!data_.is_tree());
-  return data_.as_chars();
-}
-
-inline absl::Nonnull<absl::cord_internal::CordRep*> Cord::InlineRep::as_tree()
-    const {
-  assert(data_.is_tree());
-  return data_.as_tree();
-}
-
-inline absl::Nullable<absl::cord_internal::CordRep*> Cord::InlineRep::tree()
-    const {
-  if (is_tree()) {
-    return as_tree();
-  } else {
-    return nullptr;
-  }
-}
-
-inline size_t Cord::InlineRep::size() const {
-  return is_tree() ? as_tree()->length : inline_size();
-}
-
-inline absl::Nonnull<cord_internal::CordRepFlat*>
-Cord::InlineRep::MakeFlatWithExtraCapacity(size_t extra) {
-  static_assert(cord_internal::kMinFlatLength >= sizeof(data_), "");
-  size_t len = data_.inline_size();
-  auto* result = CordRepFlat::New(len + extra);
-  result->length = len;
-  data_.copy_max_inline_to(result->Data());
-  return result;
-}
-
-inline void Cord::InlineRep::EmplaceTree(absl::Nonnull<CordRep*> rep,
-                                         MethodIdentifier method) {
-  assert(rep);
-  data_.make_tree(rep);
-  CordzInfo::MaybeTrackCord(data_, method);
-}
-
-inline void Cord::InlineRep::EmplaceTree(absl::Nonnull<CordRep*> rep,
-                                         const InlineData& parent,
-                                         MethodIdentifier method) {
-  data_.make_tree(rep);
-  CordzInfo::MaybeTrackCord(data_, parent, method);
-}
-
-inline void Cord::InlineRep::SetTree(absl::Nonnull<CordRep*> rep,
-                                     const CordzUpdateScope& scope) {
-  assert(rep);
-  assert(data_.is_tree());
-  data_.set_tree(rep);
-  scope.SetCordRep(rep);
-}
-
-inline void Cord::InlineRep::SetTreeOrEmpty(absl::Nullable<CordRep*> rep,
-                                            const CordzUpdateScope& scope) {
-  assert(data_.is_tree());
-  if (rep) {
-    data_.set_tree(rep);
-  } else {
-    data_ = {};
-  }
-  scope.SetCordRep(rep);
-}
-
-inline void Cord::InlineRep::CommitTree(absl::Nullable<const CordRep*> old_rep,
-                                        absl::Nonnull<CordRep*> rep,
-                                        const CordzUpdateScope& scope,
-                                        MethodIdentifier method) {
-  if (old_rep) {
-    SetTree(rep, scope);
-  } else {
-    EmplaceTree(rep, method);
-  }
-}
-
-inline absl::Nullable<absl::cord_internal::CordRep*> Cord::InlineRep::clear() {
-  if (is_tree()) {
-    CordzInfo::MaybeUntrackCord(cordz_info());
-  }
-  absl::cord_internal::CordRep* result = tree();
-  ResetToEmpty();
-  return result;
-}
-
-inline void Cord::InlineRep::CopyToArray(absl::Nonnull<char*> dst) const {
-  assert(!is_tree());
-  size_t n = inline_size();
-  assert(n != 0);
-  cord_internal::SmallMemmove(dst, data_.as_chars(), n);
-}
-
-inline void Cord::InlineRep::MaybeRemoveEmptyCrcNode() {
-  CordRep* rep = tree();
-  if (rep == nullptr || ABSL_PREDICT_TRUE(rep->length > 0)) {
-    return;
-  }
-  assert(rep->IsCrc());
-  assert(rep->crc()->child == nullptr);
-  CordzInfo::MaybeUntrackCord(cordz_info());
-  CordRep::Unref(rep);
-  ResetToEmpty();
-}
-
-constexpr inline Cord::Cord() noexcept {}
-
-inline Cord::Cord(absl::string_view src)
-    : Cord(src, CordzUpdateTracker::kConstructorString) {}
-
-template <typename T>
-constexpr Cord::Cord(strings_internal::StringConstant<T>)
-    : contents_(strings_internal::StringConstant<T>::value,
-                strings_internal::StringConstant<T>::value.size() <=
-                        cord_internal::kMaxInline
-                    ? nullptr
-                    : &cord_internal::ConstInitExternalStorage<
-                          strings_internal::StringConstant<T>>::value) {}
-
-inline Cord& Cord::operator=(const Cord& x) {
-  contents_ = x.contents_;
-  return *this;
-}
-
-template <typename T, Cord::EnableIfString<T>>
-Cord& Cord::operator=(T&& src) {
-  if (src.size() <= cord_internal::kMaxBytesToCopy) {
-    return operator=(absl::string_view(src));
-  } else {
-    return AssignLargeString(std::forward<T>(src));
-  }
-}
-
-inline Cord::Cord(const Cord& src) : contents_(src.contents_) {}
-
-inline Cord::Cord(Cord&& src) noexcept : contents_(std::move(src.contents_)) {}
-
-inline void Cord::swap(Cord& other) noexcept {
-  contents_.Swap(&other.contents_);
-}
-
-inline Cord& Cord::operator=(Cord&& x) noexcept {
-  contents_ = std::move(x.contents_);
-  return *this;
-}
-
-extern template Cord::Cord(std::string&& src);
-
-inline size_t Cord::size() const {
-  // Length is 1st field in str.rep_
-  return contents_.size();
-}
-
-inline bool Cord::empty() const { return size() == 0; }
-
-inline size_t Cord::EstimatedMemoryUsage(
-    CordMemoryAccounting accounting_method) const {
-  size_t result = sizeof(Cord);
-  if (const absl::cord_internal::CordRep* rep = contents_.tree()) {
-    switch (accounting_method) {
-      case CordMemoryAccounting::kFairShare:
-        result += cord_internal::GetEstimatedFairShareMemoryUsage(rep);
-        break;
-      case CordMemoryAccounting::kTotalMorePrecise:
-        result += cord_internal::GetMorePreciseMemoryUsage(rep);
-        break;
-      case CordMemoryAccounting::kTotal:
-        result += cord_internal::GetEstimatedMemoryUsage(rep);
-        break;
-    }
-  }
-  return result;
-}
-
-inline absl::optional<absl::string_view> Cord::TryFlat() const {
-  absl::cord_internal::CordRep* rep = contents_.tree();
-  if (rep == nullptr) {
-    return absl::string_view(contents_.data(), contents_.size());
-  }
-  absl::string_view fragment;
-  if (GetFlatAux(rep, &fragment)) {
-    return fragment;
-  }
-  return absl::nullopt;
-}
-
-inline absl::string_view Cord::Flatten() {
-  absl::cord_internal::CordRep* rep = contents_.tree();
-  if (rep == nullptr) {
-    return absl::string_view(contents_.data(), contents_.size());
-  } else {
-    absl::string_view already_flat_contents;
-    if (GetFlatAux(rep, &already_flat_contents)) {
-      return already_flat_contents;
-    }
-  }
-  return FlattenSlowPath();
-}
-
-inline void Cord::Append(absl::string_view src) {
-  contents_.AppendArray(src, CordzUpdateTracker::kAppendString);
-}
-
-inline void Cord::Prepend(absl::string_view src) {
-  PrependArray(src, CordzUpdateTracker::kPrependString);
-}
-
-inline void Cord::Append(CordBuffer buffer) {
-  if (ABSL_PREDICT_FALSE(buffer.length() == 0)) return;
-  absl::string_view short_value;
-  if (CordRep* rep = buffer.ConsumeValue(short_value)) {
-    contents_.AppendTree(rep, CordzUpdateTracker::kAppendCordBuffer);
-  } else {
-    AppendPrecise(short_value, CordzUpdateTracker::kAppendCordBuffer);
-  }
-}
-
-inline void Cord::Prepend(CordBuffer buffer) {
-  if (ABSL_PREDICT_FALSE(buffer.length() == 0)) return;
-  absl::string_view short_value;
-  if (CordRep* rep = buffer.ConsumeValue(short_value)) {
-    contents_.PrependTree(rep, CordzUpdateTracker::kPrependCordBuffer);
-  } else {
-    PrependPrecise(short_value, CordzUpdateTracker::kPrependCordBuffer);
-  }
-}
-
-inline CordBuffer Cord::GetAppendBuffer(size_t capacity, size_t min_capacity) {
-  if (empty()) return CordBuffer::CreateWithDefaultLimit(capacity);
-  return GetAppendBufferSlowPath(0, capacity, min_capacity);
-}
-
-inline CordBuffer Cord::GetCustomAppendBuffer(size_t block_size,
-                                              size_t capacity,
-                                              size_t min_capacity) {
-  if (empty()) {
-    return block_size ? CordBuffer::CreateWithCustomLimit(block_size, capacity)
-                      : CordBuffer::CreateWithDefaultLimit(capacity);
-  }
-  return GetAppendBufferSlowPath(block_size, capacity, min_capacity);
-}
-
-extern template void Cord::Append(std::string&& src);
-extern template void Cord::Prepend(std::string&& src);
-
-inline int Cord::Compare(const Cord& rhs) const {
-  if (!contents_.is_tree() && !rhs.contents_.is_tree()) {
-    return contents_.data_.Compare(rhs.contents_.data_);
-  }
-
-  return CompareImpl(rhs);
-}
-
-// Does 'this' cord start/end with rhs
-inline bool Cord::StartsWith(const Cord& rhs) const {
-  if (contents_.IsSame(rhs.contents_)) return true;
-  size_t rhs_size = rhs.size();
-  if (size() < rhs_size) return false;
-  return EqualsImpl(rhs, rhs_size);
-}
-
-inline bool Cord::StartsWith(absl::string_view rhs) const {
-  size_t rhs_size = rhs.size();
-  if (size() < rhs_size) return false;
-  return EqualsImpl(rhs, rhs_size);
-}
-
-inline void Cord::ChunkIterator::InitTree(
-    absl::Nonnull<cord_internal::CordRep*> tree) {
-  tree = cord_internal::SkipCrcNode(tree);
-  if (tree->tag == cord_internal::BTREE) {
-    current_chunk_ = btree_reader_.Init(tree->btree());
-  } else {
-    current_leaf_ = tree;
-    current_chunk_ = cord_internal::EdgeData(tree);
-  }
-}
-
-inline Cord::ChunkIterator::ChunkIterator(
-    absl::Nonnull<cord_internal::CordRep*> tree) {
-  bytes_remaining_ = tree->length;
-  InitTree(tree);
-}
-
-inline Cord::ChunkIterator::ChunkIterator(absl::Nonnull<const Cord*> cord) {
-  if (CordRep* tree = cord->contents_.tree()) {
-    bytes_remaining_ = tree->length;
-    if (ABSL_PREDICT_TRUE(bytes_remaining_ != 0)) {
-      InitTree(tree);
-    } else {
-      current_chunk_ = {};
-    }
-  } else {
-    bytes_remaining_ = cord->contents_.inline_size();
-    current_chunk_ = {cord->contents_.data(), bytes_remaining_};
-  }
-}
-
-inline Cord::ChunkIterator& Cord::ChunkIterator::AdvanceBtree() {
-  current_chunk_ = btree_reader_.Next();
-  return *this;
-}
-
-inline void Cord::ChunkIterator::AdvanceBytesBtree(size_t n) {
-  assert(n >= current_chunk_.size());
-  bytes_remaining_ -= n;
-  if (bytes_remaining_) {
-    if (n == current_chunk_.size()) {
-      current_chunk_ = btree_reader_.Next();
-    } else {
-      size_t offset = btree_reader_.length() - bytes_remaining_;
-      current_chunk_ = btree_reader_.Seek(offset);
-    }
-  } else {
-    current_chunk_ = {};
-  }
-}
-
-inline Cord::ChunkIterator& Cord::ChunkIterator::operator++() {
-  ABSL_HARDENING_ASSERT(bytes_remaining_ > 0 &&
-                        "Attempted to iterate past `end()`");
-  assert(bytes_remaining_ >= current_chunk_.size());
-  bytes_remaining_ -= current_chunk_.size();
-  if (bytes_remaining_ > 0) {
-    if (btree_reader_) {
-      return AdvanceBtree();
-    } else {
-      assert(!current_chunk_.empty());  // Called on invalid iterator.
-    }
-    current_chunk_ = {};
-  }
-  return *this;
-}
-
-inline Cord::ChunkIterator Cord::ChunkIterator::operator++(int) {
-  ChunkIterator tmp(*this);
-  operator++();
-  return tmp;
-}
-
-inline bool Cord::ChunkIterator::operator==(const ChunkIterator& other) const {
-  return bytes_remaining_ == other.bytes_remaining_;
-}
-
-inline bool Cord::ChunkIterator::operator!=(const ChunkIterator& other) const {
-  return !(*this == other);
-}
-
-inline Cord::ChunkIterator::reference Cord::ChunkIterator::operator*() const {
-  ABSL_HARDENING_ASSERT(bytes_remaining_ != 0);
-  return current_chunk_;
-}
-
-inline Cord::ChunkIterator::pointer Cord::ChunkIterator::operator->() const {
-  ABSL_HARDENING_ASSERT(bytes_remaining_ != 0);
-  return &current_chunk_;
-}
-
-inline void Cord::ChunkIterator::RemoveChunkPrefix(size_t n) {
-  assert(n < current_chunk_.size());
-  current_chunk_.remove_prefix(n);
-  bytes_remaining_ -= n;
-}
-
-inline void Cord::ChunkIterator::AdvanceBytes(size_t n) {
-  assert(bytes_remaining_ >= n);
-  if (ABSL_PREDICT_TRUE(n < current_chunk_.size())) {
-    RemoveChunkPrefix(n);
-  } else if (n != 0) {
-    if (btree_reader_) {
-      AdvanceBytesBtree(n);
-    } else {
-      bytes_remaining_ = 0;
-    }
-  }
-}
-
-inline Cord::ChunkIterator Cord::chunk_begin() const {
-  return ChunkIterator(this);
-}
-
-inline Cord::ChunkIterator Cord::chunk_end() const { return ChunkIterator(); }
-
-inline Cord::ChunkIterator Cord::ChunkRange::begin() const {
-  return cord_->chunk_begin();
-}
-
-inline Cord::ChunkIterator Cord::ChunkRange::end() const {
-  return cord_->chunk_end();
-}
-
-inline Cord::ChunkRange Cord::Chunks() const { return ChunkRange(this); }
-
-inline Cord::CharIterator& Cord::CharIterator::operator++() {
-  if (ABSL_PREDICT_TRUE(chunk_iterator_->size() > 1)) {
-    chunk_iterator_.RemoveChunkPrefix(1);
-  } else {
-    ++chunk_iterator_;
-  }
-  return *this;
-}
-
-inline Cord::CharIterator Cord::CharIterator::operator++(int) {
-  CharIterator tmp(*this);
-  operator++();
-  return tmp;
-}
-
-inline bool Cord::CharIterator::operator==(const CharIterator& other) const {
-  return chunk_iterator_ == other.chunk_iterator_;
-}
-
-inline bool Cord::CharIterator::operator!=(const CharIterator& other) const {
-  return !(*this == other);
-}
-
-inline Cord::CharIterator::reference Cord::CharIterator::operator*() const {
-  return *chunk_iterator_->data();
-}
-
-inline Cord::CharIterator::pointer Cord::CharIterator::operator->() const {
-  return chunk_iterator_->data();
-}
-
-inline Cord Cord::AdvanceAndRead(absl::Nonnull<CharIterator*> it,
-                                 size_t n_bytes) {
-  assert(it != nullptr);
-  return it->chunk_iterator_.AdvanceAndReadBytes(n_bytes);
-}
-
-inline void Cord::Advance(absl::Nonnull<CharIterator*> it, size_t n_bytes) {
-  assert(it != nullptr);
-  it->chunk_iterator_.AdvanceBytes(n_bytes);
-}
-
-inline absl::string_view Cord::ChunkRemaining(const CharIterator& it) {
-  return *it.chunk_iterator_;
-}
-
-inline Cord::CharIterator Cord::char_begin() const {
-  return CharIterator(this);
-}
-
-inline Cord::CharIterator Cord::char_end() const { return CharIterator(); }
-
-inline Cord::CharIterator Cord::CharRange::begin() const {
-  return cord_->char_begin();
-}
-
-inline Cord::CharIterator Cord::CharRange::end() const {
-  return cord_->char_end();
-}
-
-inline Cord::CharRange Cord::Chars() const { return CharRange(this); }
-
-inline void Cord::ForEachChunk(
-    absl::FunctionRef<void(absl::string_view)> callback) const {
-  absl::cord_internal::CordRep* rep = contents_.tree();
-  if (rep == nullptr) {
-    callback(absl::string_view(contents_.data(), contents_.size()));
-  } else {
-    ForEachChunkAux(rep, callback);
-  }
-}
-
-// Nonmember Cord-to-Cord relational operators.
-inline bool operator==(const Cord& lhs, const Cord& rhs) {
-  if (lhs.contents_.IsSame(rhs.contents_)) return true;
-  size_t rhs_size = rhs.size();
-  if (lhs.size() != rhs_size) return false;
-  return lhs.EqualsImpl(rhs, rhs_size);
-}
-
-inline bool operator!=(const Cord& x, const Cord& y) { return !(x == y); }
-inline bool operator<(const Cord& x, const Cord& y) { return x.Compare(y) < 0; }
-inline bool operator>(const Cord& x, const Cord& y) { return x.Compare(y) > 0; }
-inline bool operator<=(const Cord& x, const Cord& y) {
-  return x.Compare(y) <= 0;
-}
-inline bool operator>=(const Cord& x, const Cord& y) {
-  return x.Compare(y) >= 0;
-}
-
-// Nonmember Cord-to-absl::string_view relational operators.
-//
-// Due to implicit conversions, these also enable comparisons of Cord with
-// std::string and const char*.
-inline bool operator==(const Cord& lhs, absl::string_view rhs) {
-  size_t lhs_size = lhs.size();
-  size_t rhs_size = rhs.size();
-  if (lhs_size != rhs_size) return false;
-  return lhs.EqualsImpl(rhs, rhs_size);
-}
-
-inline bool operator==(absl::string_view x, const Cord& y) { return y == x; }
-inline bool operator!=(const Cord& x, absl::string_view y) { return !(x == y); }
-inline bool operator!=(absl::string_view x, const Cord& y) { return !(x == y); }
-inline bool operator<(const Cord& x, absl::string_view y) {
-  return x.Compare(y) < 0;
-}
-inline bool operator<(absl::string_view x, const Cord& y) {
-  return y.Compare(x) > 0;
-}
-inline bool operator>(const Cord& x, absl::string_view y) { return y < x; }
-inline bool operator>(absl::string_view x, const Cord& y) { return y < x; }
-inline bool operator<=(const Cord& x, absl::string_view y) { return !(y < x); }
-inline bool operator<=(absl::string_view x, const Cord& y) { return !(y < x); }
-inline bool operator>=(const Cord& x, absl::string_view y) { return !(x < y); }
-inline bool operator>=(absl::string_view x, const Cord& y) { return !(x < y); }
-
-// Some internals exposed to test code.
-namespace strings_internal {
-class CordTestAccess {
- public:
-  static size_t FlatOverhead();
-  static size_t MaxFlatLength();
-  static size_t SizeofCordRepExternal();
-  static size_t SizeofCordRepSubstring();
-  static size_t FlatTagToLength(uint8_t tag);
-  static uint8_t LengthToTag(size_t s);
-};
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_CORD_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/cord_analysis.h b/projects/llm_framework/include/abseil-cpp/absl/strings/cord_analysis.h
deleted file mode 100644
index f8ce3489..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/cord_analysis.h
+++ /dev/null
@@ -1,63 +0,0 @@
-// Copyright 2021 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_CORD_ANALYSIS_H_
-#define ABSL_STRINGS_CORD_ANALYSIS_H_
-
-#include <cstddef>
-#include <cstdint>
-
-#include "absl/base/config.h"
-#include "absl/base/nullability.h"
-#include "absl/strings/internal/cord_internal.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// Returns the *approximate* number of bytes held in full or in part by this
-// Cord (which may not remain the same between invocations). Cords that share
-// memory could each be "charged" independently for the same shared memory.
-size_t GetEstimatedMemoryUsage(absl::Nonnull<const CordRep*> rep);
-
-// Returns the *approximate* number of bytes held in full or in part by this
-// Cord for the distinct memory held by this cord. This is similar to
-// `GetEstimatedMemoryUsage()`, except that if the cord has multiple references
-// to the same memory, that memory is only counted once.
-//
-// For example:
-//   absl::Cord cord;
-//   cord.append(some_other_cord);
-//   cord.append(some_other_cord);
-//    // Calls GetEstimatedMemoryUsage() and counts `other_cord` twice:
-//   cord.EstimatedMemoryUsage(kTotal);
-//    // Calls GetMorePreciseMemoryUsage() and counts `other_cord` once:
-//   cord.EstimatedMemoryUsage(kTotalMorePrecise);
-//
-// This is more expensive than `GetEstimatedMemoryUsage()` as it requires
-// deduplicating all memory references.
-size_t GetMorePreciseMemoryUsage(absl::Nonnull<const CordRep*> rep);
-
-// Returns the *approximate* number of bytes held in full or in part by this
-// CordRep weighted by the sharing ratio of that data. For example, if some data
-// edge is shared by 4 different Cords, then each cord is attribute 1/4th of
-// the total memory usage as a 'fair share' of the total memory usage.
-size_t GetEstimatedFairShareMemoryUsage(absl::Nonnull<const CordRep*> rep);
-
-}  // namespace cord_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-
-#endif  // ABSL_STRINGS_CORD_ANALYSIS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/cord_buffer.h b/projects/llm_framework/include/abseil-cpp/absl/strings/cord_buffer.h
deleted file mode 100644
index bc0e4e45..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/cord_buffer.h
+++ /dev/null
@@ -1,572 +0,0 @@
-// Copyright 2021 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: cord_buffer.h
-// -----------------------------------------------------------------------------
-//
-// This file defines an `absl::CordBuffer` data structure to hold data for
-// eventual inclusion within an existing `Cord` data structure. Cord buffers are
-// useful for building large Cords that may require custom allocation of its
-// associated memory.
-//
-#ifndef ABSL_STRINGS_CORD_BUFFER_H_
-#define ABSL_STRINGS_CORD_BUFFER_H_
-
-#include <algorithm>
-#include <cassert>
-#include <cstddef>
-#include <cstdint>
-#include <memory>
-#include <utility>
-
-#include "absl/base/config.h"
-#include "absl/base/macros.h"
-#include "absl/numeric/bits.h"
-#include "absl/strings/internal/cord_internal.h"
-#include "absl/strings/internal/cord_rep_flat.h"
-#include "absl/types/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-class Cord;
-class CordBufferTestPeer;
-
-// CordBuffer
-//
-// CordBuffer manages memory buffers for purposes such as zero-copy APIs as well
-// as applications building cords with large data requiring granular control
-// over the allocation and size of cord data. For example, a function creating
-// a cord of random data could use a CordBuffer as follows:
-//
-//   absl::Cord CreateRandomCord(size_t length) {
-//     absl::Cord cord;
-//     while (length > 0) {
-//       CordBuffer buffer = CordBuffer::CreateWithDefaultLimit(length);
-//       absl::Span<char> data = buffer.available_up_to(length);
-//       FillRandomValues(data.data(), data.size());
-//       buffer.IncreaseLengthBy(data.size());
-//       cord.Append(std::move(buffer));
-//       length -= data.size();
-//     }
-//     return cord;
-//   }
-//
-// CordBuffer instances are by default limited to a capacity of `kDefaultLimit`
-// bytes. `kDefaultLimit` is currently just under 4KiB, but this default may
-// change in the future and/or for specific architectures. The default limit is
-// aimed to provide a good trade-off between performance and memory overhead.
-// Smaller buffers typically incur more compute cost while larger buffers are
-// more CPU efficient but create significant memory overhead because of such
-// allocations being less granular. Using larger buffers may also increase the
-// risk of memory fragmentation.
-//
-// Applications create a buffer using one of the `CreateWithDefaultLimit()` or
-// `CreateWithCustomLimit()` methods. The returned instance will have a non-zero
-// capacity and a zero length. Applications use the `data()` method to set the
-// contents of the managed memory, and once done filling the buffer, use the
-// `IncreaseLengthBy()` or 'SetLength()' method to specify the length of the
-// initialized data before adding the buffer to a Cord.
-//
-// The `CreateWithCustomLimit()` method is intended for applications needing
-// larger buffers than the default memory limit, allowing the allocation of up
-// to a capacity of `kCustomLimit` bytes minus some minimum internal overhead.
-// The usage of `CreateWithCustomLimit()` should be limited to only those use
-// cases where the distribution of the input is relatively well known, and/or
-// where the trade-off between the efficiency gains outweigh the risk of memory
-// fragmentation. See the documentation for `CreateWithCustomLimit()` for more
-// information on using larger custom limits.
-//
-// The capacity of a `CordBuffer` returned by one of the `Create` methods may
-// be larger than the requested capacity due to rounding, alignment and
-// granularity of the memory allocator. Applications should use the `capacity`
-// method to obtain the effective capacity of the returned instance as
-// demonstrated in the provided example above.
-//
-// CordBuffer is a move-only class. All references into the managed memory are
-// invalidated when an instance is moved into either another CordBuffer instance
-// or a Cord. Writing to a location obtained by a previous call to `data()`
-// after an instance was moved will lead to undefined behavior.
-//
-// A `moved from` CordBuffer instance will have a valid, but empty state.
-// CordBuffer is thread compatible.
-class CordBuffer {
- public:
-  // kDefaultLimit
-  //
-  // Default capacity limits of allocated CordBuffers.
-  // See the class comments for more information on allocation limits.
-  static constexpr size_t kDefaultLimit = cord_internal::kMaxFlatLength;
-
-  // kCustomLimit
-  //
-  // Maximum size for CreateWithCustomLimit() allocated buffers.
-  // Note that the effective capacity may be slightly less
-  // because of internal overhead of internal cord buffers.
-  static constexpr size_t kCustomLimit = 64U << 10;
-
-  // Constructors, Destructors and Assignment Operators
-
-  // Creates an empty CordBuffer.
-  CordBuffer() = default;
-
-  // Destroys this CordBuffer instance and, if not empty, releases any memory
-  // managed by this instance, invalidating previously returned references.
-  ~CordBuffer();
-
-  // CordBuffer is move-only
-  CordBuffer(CordBuffer&& rhs) noexcept;
-  CordBuffer& operator=(CordBuffer&&) noexcept;
-  CordBuffer(const CordBuffer&) = delete;
-  CordBuffer& operator=(const CordBuffer&) = delete;
-
-  // CordBuffer::MaximumPayload()
-  //
-  // Returns the guaranteed maximum payload for a CordBuffer returned by the
-  // `CreateWithDefaultLimit()` method. While small, each internal buffer inside
-  // a Cord incurs an overhead to manage the length, type and reference count
-  // for the buffer managed inside the cord tree. Applications can use this
-  // method to get approximate number of buffers required for a given byte
-  // size, etc.
-  //
-  // For example:
-  //   const size_t payload = absl::CordBuffer::MaximumPayload();
-  //   const size_t buffer_count = (total_size + payload - 1) / payload;
-  //   buffers.reserve(buffer_count);
-  static constexpr size_t MaximumPayload();
-
-  // Overload to the above `MaximumPayload()` except that it returns the
-  // maximum payload for a CordBuffer returned by the `CreateWithCustomLimit()`
-  // method given the provided `block_size`.
-  static constexpr size_t MaximumPayload(size_t block_size);
-
-  // CordBuffer::CreateWithDefaultLimit()
-  //
-  // Creates a CordBuffer instance of the desired `capacity`, capped at the
-  // default limit `kDefaultLimit`. The returned buffer has a guaranteed
-  // capacity of at least `min(kDefaultLimit, capacity)`. See the class comments
-  // for more information on buffer capacities and intended usage.
-  static CordBuffer CreateWithDefaultLimit(size_t capacity);
-
-  // CordBuffer::CreateWithCustomLimit()
-  //
-  // Creates a CordBuffer instance of the desired `capacity` rounded to an
-  // appropriate power of 2 size less than, or equal to `block_size`.
-  // Requires `block_size` to be a power of 2.
-  //
-  // If `capacity` is less than or equal to `kDefaultLimit`, then this method
-  // behaves identical to `CreateWithDefaultLimit`, which means that the caller
-  // is guaranteed to get a buffer of at least the requested capacity.
-  //
-  // If `capacity` is greater than or equal to `block_size`, then this method
-  // returns a buffer with an `allocated size` of `block_size` bytes. Otherwise,
-  // this methods returns a buffer with a suitable smaller power of 2 block size
-  // to satisfy the request. The actual size depends on a number of factors, and
-  // is typically (but not necessarily) the highest or second highest power of 2
-  // value less than or equal to `capacity`.
-  //
-  // The 'allocated size' includes a small amount of overhead required for
-  // internal state, which is currently 13 bytes on 64-bit platforms. For
-  // example: a buffer created with `block_size` and `capacity' set to 8KiB
-  // will have an allocated size of 8KiB, and an effective internal `capacity`
-  // of 8KiB - 13 = 8179 bytes.
-  //
-  // To demonstrate this in practice, let's assume we want to read data from
-  // somewhat larger files using approximately 64KiB buffers:
-  //
-  //   absl::Cord ReadFromFile(int fd, size_t n) {
-  //     absl::Cord cord;
-  //     while (n > 0) {
-  //       CordBuffer buffer = CordBuffer::CreateWithCustomLimit(64 << 10, n);
-  //       absl::Span<char> data = buffer.available_up_to(n);
-  //       ReadFileDataOrDie(fd, data.data(), data.size());
-  //       buffer.IncreaseLengthBy(data.size());
-  //       cord.Append(std::move(buffer));
-  //       n -= data.size();
-  //     }
-  //     return cord;
-  //   }
-  //
-  // If we'd use this function to read a file of 659KiB, we may get the
-  // following pattern of allocated cord buffer sizes:
-  //
-  //   CreateWithCustomLimit(64KiB, 674816) --> ~64KiB (65523)
-  //   CreateWithCustomLimit(64KiB, 674816) --> ~64KiB (65523)
-  //   ...
-  //   CreateWithCustomLimit(64KiB,  19586) --> ~16KiB (16371)
-  //   CreateWithCustomLimit(64KiB,   3215) -->   3215 (at least 3215)
-  //
-  // The reason the method returns a 16K buffer instead of a roughly 19K buffer
-  // is to reduce memory overhead and fragmentation risks. Using carefully
-  // chosen power of 2 values reduces the entropy of allocated memory sizes.
-  //
-  // Additionally, let's assume we'd use the above function on files that are
-  // generally smaller than 64K. If we'd use 'precise' sized buffers for such
-  // files, than we'd get a very wide distribution of allocated memory sizes
-  // rounded to 4K page sizes, and we'd end up with a lot of unused capacity.
-  //
-  // In general, application should only use custom sizes if the data they are
-  // consuming or storing is expected to be many times the chosen block size,
-  // and be based on objective data and performance metrics. For example, a
-  // compress function may work faster and consume less CPU when using larger
-  // buffers. Such an application should pick a size offering a reasonable
-  // trade-off between expected data size, compute savings with larger buffers,
-  // and the cost or fragmentation effect of larger buffers.
-  // Applications must pick a reasonable spot on that curve, and make sure their
-  // data meets their expectations in size distributions such as "mostly large".
-  static CordBuffer CreateWithCustomLimit(size_t block_size, size_t capacity);
-
-  // CordBuffer::available()
-  //
-  // Returns the span delineating the available capacity in this buffer
-  // which is defined as `{ data() + length(), capacity() - length() }`.
-  absl::Span<char> available();
-
-  // CordBuffer::available_up_to()
-  //
-  // Returns the span delineating the available capacity in this buffer limited
-  // to `size` bytes. This is equivalent to `available().subspan(0, size)`.
-  absl::Span<char> available_up_to(size_t size);
-
-  // CordBuffer::data()
-  //
-  // Returns a non-null reference to the data managed by this instance.
-  // Applications are allowed to write up to `capacity` bytes of instance data.
-  // CordBuffer data is uninitialized by default. Reading data from an instance
-  // that has not yet been initialized will lead to undefined behavior.
-  char* data();
-  const char* data() const;
-
-  // CordBuffer::length()
-  //
-  // Returns the length of this instance. The default length of a CordBuffer is
-  // 0, indicating an 'empty' CordBuffer. Applications must specify the length
-  // of the data in a CordBuffer before adding it to a Cord.
-  size_t length() const;
-
-  // CordBuffer::capacity()
-  //
-  // Returns the capacity of this instance. All instances have a non-zero
-  // capacity: default and `moved from` instances have a small internal buffer.
-  size_t capacity() const;
-
-  // CordBuffer::IncreaseLengthBy()
-  //
-  // Increases the length of this buffer by the specified 'n' bytes.
-  // Applications must make sure all data in this buffer up to the new length
-  // has been initialized before adding a CordBuffer to a Cord: failure to do so
-  // will lead to undefined behavior.  Requires `length() + n <= capacity()`.
-  // Typically, applications will use 'available_up_to()` to get a span of the
-  // desired capacity, and use `span.size()` to increase the length as in:
-  //   absl::Span<char> span = buffer.available_up_to(desired);
-  //   buffer.IncreaseLengthBy(span.size());
-  //   memcpy(span.data(), src, span.size());
-  //   etc...
-  void IncreaseLengthBy(size_t n);
-
-  // CordBuffer::SetLength()
-  //
-  // Sets the data length of this instance. Applications must make sure all data
-  // of the specified length has been initialized before adding a CordBuffer to
-  // a Cord: failure to do so will lead to undefined behavior.
-  // Setting the length to a small value or zero does not release any memory
-  // held by this CordBuffer instance. Requires `length <= capacity()`.
-  // Applications should preferably use the `IncreaseLengthBy()` method above
-  // in combination with the 'available()` or `available_up_to()` methods.
-  void SetLength(size_t length);
-
- private:
-  // Make sure we don't accidentally over promise.
-  static_assert(kCustomLimit <= cord_internal::kMaxLargeFlatSize, "");
-
-  // Assume the cost of an 'uprounded' allocation to CeilPow2(size) versus
-  // the cost of allocating at least 1 extra flat <= 4KB:
-  // - Flat overhead = 13 bytes
-  // - Btree amortized cost / node =~ 13 bytes
-  // - 64 byte granularity of tcmalloc at 4K =~ 32 byte average
-  // CPU cost and efficiency requires we should at least 'save' something by
-  // splitting, as a poor man's measure, we say the slop needs to be
-  // at least double the cost offset to make it worth splitting: ~128 bytes.
-  static constexpr size_t kMaxPageSlop = 128;
-
-  // Overhead for allocation a flat.
-  static constexpr size_t kOverhead = cord_internal::kFlatOverhead;
-
-  using CordRepFlat = cord_internal::CordRepFlat;
-
-  // `Rep` is the internal data representation of a CordBuffer. The internal
-  // representation has an internal small size optimization similar to
-  // std::string (SSO).
-  struct Rep {
-    // Inline SSO size of a CordBuffer
-    static constexpr size_t kInlineCapacity = sizeof(intptr_t) * 2 - 1;
-
-    // Creates a default instance with kInlineCapacity.
-    Rep() : short_rep{} {}
-
-    // Creates an instance managing an allocated non zero CordRep.
-    explicit Rep(cord_internal::CordRepFlat* rep) : long_rep{rep} {
-      assert(rep != nullptr);
-    }
-
-    // Returns true if this instance manages the SSO internal buffer.
-    bool is_short() const {
-      constexpr size_t offset = offsetof(Short, raw_size);
-      return (reinterpret_cast<const char*>(this)[offset] & 1) != 0;
-    }
-
-    // Returns the available area of the internal SSO data
-    absl::Span<char> short_available() {
-      const size_t length = short_length();
-      return absl::Span<char>(short_rep.data + length,
-                              kInlineCapacity - length);
-    }
-
-    // Returns the available area of the internal SSO data
-    absl::Span<char> long_available() const {
-      assert(!is_short());
-      const size_t length = long_rep.rep->length;
-      return absl::Span<char>(long_rep.rep->Data() + length,
-                              long_rep.rep->Capacity() - length);
-    }
-
-    // Returns the length of the internal SSO data.
-    size_t short_length() const {
-      assert(is_short());
-      return static_cast<size_t>(short_rep.raw_size >> 1);
-    }
-
-    // Sets the length of the internal SSO data.
-    // Disregards any previously set CordRep instance.
-    void set_short_length(size_t length) {
-      short_rep.raw_size = static_cast<char>((length << 1) + 1);
-    }
-
-    // Adds `n` to the current short length.
-    void add_short_length(size_t n) {
-      assert(is_short());
-      short_rep.raw_size += static_cast<char>(n << 1);
-    }
-
-    // Returns reference to the internal SSO data buffer.
-    char* data() {
-      assert(is_short());
-      return short_rep.data;
-    }
-    const char* data() const {
-      assert(is_short());
-      return short_rep.data;
-    }
-
-    // Returns a pointer the external CordRep managed by this instance.
-    cord_internal::CordRepFlat* rep() const {
-      assert(!is_short());
-      return long_rep.rep;
-    }
-
-    // The internal representation takes advantage of the fact that allocated
-    // memory is always on an even address, and uses the least significant bit
-    // of the first or last byte (depending on endianness) as the inline size
-    // indicator overlapping with the least significant byte of the CordRep*.
-#if defined(ABSL_IS_BIG_ENDIAN)
-    struct Long {
-      explicit Long(cord_internal::CordRepFlat* rep_arg) : rep(rep_arg) {}
-      void* padding;
-      cord_internal::CordRepFlat* rep;
-    };
-    struct Short {
-      char data[sizeof(Long) - 1];
-      char raw_size = 1;
-    };
-#else
-    struct Long {
-      explicit Long(cord_internal::CordRepFlat* rep_arg) : rep(rep_arg) {}
-      cord_internal::CordRepFlat* rep;
-      void* padding;
-    };
-    struct Short {
-      char raw_size = 1;
-      char data[sizeof(Long) - 1];
-    };
-#endif
-
-    union {
-      Long long_rep;
-      Short short_rep;
-    };
-  };
-
-  // Power2 functions
-  static bool IsPow2(size_t size) { return absl::has_single_bit(size); }
-  static size_t Log2Floor(size_t size) {
-    return static_cast<size_t>(absl::bit_width(size) - 1);
-  }
-  static size_t Log2Ceil(size_t size) {
-    return static_cast<size_t>(absl::bit_width(size - 1));
-  }
-
-  // Implementation of `CreateWithCustomLimit()`.
-  // This implementation allows for future memory allocation hints to
-  // be passed down into the CordRepFlat allocation function.
-  template <typename... AllocationHints>
-  static CordBuffer CreateWithCustomLimitImpl(size_t block_size,
-                                              size_t capacity,
-                                              AllocationHints... hints);
-
-  // Consumes the value contained in this instance and resets the instance.
-  // This method returns a non-null Cordrep* if the current instances manages a
-  // CordRep*, and resets the instance to an empty SSO instance. If the current
-  // instance is an SSO instance, then this method returns nullptr and sets
-  // `short_value` to the inlined data value. In either case, the current
-  // instance length is reset to zero.
-  // This method is intended to be used by Cord internal functions only.
-  cord_internal::CordRep* ConsumeValue(absl::string_view& short_value) {
-    cord_internal::CordRep* rep = nullptr;
-    if (rep_.is_short()) {
-      short_value = absl::string_view(rep_.data(), rep_.short_length());
-    } else {
-      rep = rep_.rep();
-    }
-    rep_.set_short_length(0);
-    return rep;
-  }
-
-  // Internal constructor.
-  explicit CordBuffer(cord_internal::CordRepFlat* rep) : rep_(rep) {
-    assert(rep != nullptr);
-  }
-
-  Rep rep_;
-
-  friend class Cord;
-  friend class CordBufferTestPeer;
-};
-
-inline constexpr size_t CordBuffer::MaximumPayload() {
-  return cord_internal::kMaxFlatLength;
-}
-
-inline constexpr size_t CordBuffer::MaximumPayload(size_t block_size) {
-  return (std::min)(kCustomLimit, block_size) - cord_internal::kFlatOverhead;
-}
-
-inline CordBuffer CordBuffer::CreateWithDefaultLimit(size_t capacity) {
-  if (capacity > Rep::kInlineCapacity) {
-    auto* rep = cord_internal::CordRepFlat::New(capacity);
-    rep->length = 0;
-    return CordBuffer(rep);
-  }
-  return CordBuffer();
-}
-
-template <typename... AllocationHints>
-inline CordBuffer CordBuffer::CreateWithCustomLimitImpl(
-    size_t block_size, size_t capacity, AllocationHints... hints) {
-  assert(IsPow2(block_size));
-  capacity = (std::min)(capacity, kCustomLimit);
-  block_size = (std::min)(block_size, kCustomLimit);
-  if (capacity + kOverhead >= block_size) {
-    capacity = block_size;
-  } else if (capacity <= kDefaultLimit) {
-    capacity = capacity + kOverhead;
-  } else if (!IsPow2(capacity)) {
-    // Check if rounded up to next power 2 is a good enough fit
-    // with limited waste making it an acceptable direct fit.
-    const size_t rounded_up = size_t{1} << Log2Ceil(capacity);
-    const size_t slop = rounded_up - capacity;
-    if (slop >= kOverhead && slop <= kMaxPageSlop + kOverhead) {
-      capacity = rounded_up;
-    } else {
-      // Round down to highest power of 2 <= capacity.
-      // Consider a more aggressive step down if that may reduce the
-      // risk of fragmentation where 'people are holding it wrong'.
-      const size_t rounded_down = size_t{1} << Log2Floor(capacity);
-      capacity = rounded_down;
-    }
-  }
-  const size_t length = capacity - kOverhead;
-  auto* rep = CordRepFlat::New(CordRepFlat::Large(), length, hints...);
-  rep->length = 0;
-  return CordBuffer(rep);
-}
-
-inline CordBuffer CordBuffer::CreateWithCustomLimit(size_t block_size,
-                                                    size_t capacity) {
-  return CreateWithCustomLimitImpl(block_size, capacity);
-}
-
-inline CordBuffer::~CordBuffer() {
-  if (!rep_.is_short()) {
-    cord_internal::CordRepFlat::Delete(rep_.rep());
-  }
-}
-
-inline CordBuffer::CordBuffer(CordBuffer&& rhs) noexcept : rep_(rhs.rep_) {
-  rhs.rep_.set_short_length(0);
-}
-
-inline CordBuffer& CordBuffer::operator=(CordBuffer&& rhs) noexcept {
-  if (!rep_.is_short()) cord_internal::CordRepFlat::Delete(rep_.rep());
-  rep_ = rhs.rep_;
-  rhs.rep_.set_short_length(0);
-  return *this;
-}
-
-inline absl::Span<char> CordBuffer::available() {
-  return rep_.is_short() ? rep_.short_available() : rep_.long_available();
-}
-
-inline absl::Span<char> CordBuffer::available_up_to(size_t size) {
-  return available().subspan(0, size);
-}
-
-inline char* CordBuffer::data() {
-  return rep_.is_short() ? rep_.data() : rep_.rep()->Data();
-}
-
-inline const char* CordBuffer::data() const {
-  return rep_.is_short() ? rep_.data() : rep_.rep()->Data();
-}
-
-inline size_t CordBuffer::capacity() const {
-  return rep_.is_short() ? Rep::kInlineCapacity : rep_.rep()->Capacity();
-}
-
-inline size_t CordBuffer::length() const {
-  return rep_.is_short() ? rep_.short_length() : rep_.rep()->length;
-}
-
-inline void CordBuffer::SetLength(size_t length) {
-  ABSL_HARDENING_ASSERT(length <= capacity());
-  if (rep_.is_short()) {
-    rep_.set_short_length(length);
-  } else {
-    rep_.rep()->length = length;
-  }
-}
-
-inline void CordBuffer::IncreaseLengthBy(size_t n) {
-  ABSL_HARDENING_ASSERT(n <= capacity() && length() + n <= capacity());
-  if (rep_.is_short()) {
-    rep_.add_short_length(n);
-  } else {
-    rep_.rep()->length += n;
-  }
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_CORD_BUFFER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/cord_test_helpers.h b/projects/llm_framework/include/abseil-cpp/absl/strings/cord_test_helpers.h
deleted file mode 100644
index ca52240a..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/cord_test_helpers.h
+++ /dev/null
@@ -1,122 +0,0 @@
-//
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_STRINGS_CORD_TEST_HELPERS_H_
-#define ABSL_STRINGS_CORD_TEST_HELPERS_H_
-
-#include <cstdint>
-#include <iostream>
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/strings/cord.h"
-#include "absl/strings/internal/cord_internal.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// Cord sizes relevant for testing
-enum class TestCordSize {
-  // An empty value
-  kEmpty = 0,
-
-  // An inlined string value
-  kInlined = cord_internal::kMaxInline / 2 + 1,
-
-  // 'Well known' SSO lengths (excluding terminating zero).
-  // libstdcxx has a maximum SSO of 15, libc++ has a maximum SSO of 22.
-  kStringSso1 = 15,
-  kStringSso2 = 22,
-
-  // A string value which is too large to fit in inlined data, but small enough
-  // such that Cord prefers copying the value if possible, i.e.: not stealing
-  // std::string inputs, or referencing existing CordReps on Append, etc.
-  kSmall = cord_internal::kMaxBytesToCopy / 2 + 1,
-
-  // A string value large enough that Cord prefers to reference or steal from
-  // existing inputs rather than copying contents of the input.
-  kMedium = cord_internal::kMaxFlatLength / 2 + 1,
-
-  // A string value large enough to cause it to be stored in multiple flats.
-  kLarge = cord_internal::kMaxFlatLength * 4
-};
-
-// To string helper
-inline absl::string_view ToString(TestCordSize size) {
-  switch (size) {
-    case TestCordSize::kEmpty:
-      return "Empty";
-    case TestCordSize::kInlined:
-      return "Inlined";
-    case TestCordSize::kSmall:
-      return "Small";
-    case TestCordSize::kStringSso1:
-      return "StringSso1";
-    case TestCordSize::kStringSso2:
-      return "StringSso2";
-    case TestCordSize::kMedium:
-      return "Medium";
-    case TestCordSize::kLarge:
-      return "Large";
-  }
-  return "???";
-}
-
-// Returns the length matching the specified size
-inline size_t Length(TestCordSize size) { return static_cast<size_t>(size); }
-
-// Stream output helper
-inline std::ostream& operator<<(std::ostream& stream, TestCordSize size) {
-  return stream << ToString(size);
-}
-
-// Creates a multi-segment Cord from an iterable container of strings.  The
-// resulting Cord is guaranteed to have one segment for every string in the
-// container.  This allows code to be unit tested with multi-segment Cord
-// inputs.
-//
-// Example:
-//
-//   absl::Cord c = absl::MakeFragmentedCord({"A ", "fragmented ", "Cord"});
-//   EXPECT_FALSE(c.GetFlat(&unused));
-//
-// The mechanism by which this Cord is created is an implementation detail.  Any
-// implementation that produces a multi-segment Cord may produce a flat Cord in
-// the future as new optimizations are added to the Cord class.
-// MakeFragmentedCord will, however, always be updated to return a multi-segment
-// Cord.
-template <typename Container>
-Cord MakeFragmentedCord(const Container& c) {
-  Cord result;
-  for (const auto& s : c) {
-    auto* external = new std::string(s);
-    Cord tmp = absl::MakeCordFromExternal(
-        *external, [external](absl::string_view) { delete external; });
-    tmp.Prepend(result);
-    result = tmp;
-  }
-  return result;
-}
-
-inline Cord MakeFragmentedCord(std::initializer_list<absl::string_view> list) {
-  return MakeFragmentedCord<std::initializer_list<absl::string_view>>(list);
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_CORD_TEST_HELPERS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/cordz_test_helpers.h b/projects/llm_framework/include/abseil-cpp/absl/strings/cordz_test_helpers.h
deleted file mode 100644
index 619f13c2..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/cordz_test_helpers.h
+++ /dev/null
@@ -1,153 +0,0 @@
-// Copyright 2021 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_CORDZ_TEST_HELPERS_H_
-#define ABSL_STRINGS_CORDZ_TEST_HELPERS_H_
-
-#include <utility>
-
-#include "gmock/gmock.h"
-#include "gtest/gtest.h"
-#include "absl/base/config.h"
-#include "absl/base/macros.h"
-#include "absl/base/nullability.h"
-#include "absl/strings/cord.h"
-#include "absl/strings/internal/cord_internal.h"
-#include "absl/strings/internal/cordz_info.h"
-#include "absl/strings/internal/cordz_sample_token.h"
-#include "absl/strings/internal/cordz_statistics.h"
-#include "absl/strings/internal/cordz_update_tracker.h"
-#include "absl/strings/str_cat.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// Returns the CordzInfo for the cord, or nullptr if the cord is not sampled.
-inline absl::Nullable<const cord_internal::CordzInfo*> GetCordzInfoForTesting(
-    const Cord& cord) {
-  if (!cord.contents_.is_tree()) return nullptr;
-  return cord.contents_.cordz_info();
-}
-
-// Returns true if the provided cordz_info is in the list of sampled cords.
-inline bool CordzInfoIsListed(
-    absl::Nonnull<const cord_internal::CordzInfo*> cordz_info,
-    cord_internal::CordzSampleToken token = {}) {
-  for (const cord_internal::CordzInfo& info : token) {
-    if (cordz_info == &info) return true;
-  }
-  return false;
-}
-
-// Matcher on Cord that verifies all of:
-// - the cord is sampled
-// - the CordzInfo of the cord is listed / discoverable.
-// - the reported CordzStatistics match the cord's actual properties
-// - the cord has an (initial) UpdateTracker count of 1 for `method`
-MATCHER_P(HasValidCordzInfoOf, method, "CordzInfo matches cord") {
-  const cord_internal::CordzInfo* cord_info = GetCordzInfoForTesting(arg);
-  if (cord_info == nullptr) {
-    *result_listener << "cord is not sampled";
-    return false;
-  }
-  if (!CordzInfoIsListed(cord_info)) {
-    *result_listener << "cord is sampled, but not listed";
-    return false;
-  }
-  cord_internal::CordzStatistics stat = cord_info->GetCordzStatistics();
-  if (stat.size != arg.size()) {
-    *result_listener << "cordz size " << stat.size
-                     << " does not match cord size " << arg.size();
-    return false;
-  }
-  if (stat.update_tracker.Value(method) != 1) {
-    *result_listener << "Expected method count 1 for " << method << ", found "
-                     << stat.update_tracker.Value(method);
-    return false;
-  }
-  return true;
-}
-
-// Matcher on Cord that verifies that the cord is sampled and that the CordzInfo
-// update tracker has 'method' with a call count of 'n'
-MATCHER_P2(CordzMethodCountEq, method, n,
-           absl::StrCat("CordzInfo method count equals ", n)) {
-  const cord_internal::CordzInfo* cord_info = GetCordzInfoForTesting(arg);
-  if (cord_info == nullptr) {
-    *result_listener << "cord is not sampled";
-    return false;
-  }
-  cord_internal::CordzStatistics stat = cord_info->GetCordzStatistics();
-  if (stat.update_tracker.Value(method) != n) {
-    *result_listener << "Expected method count " << n << " for " << method
-                     << ", found " << stat.update_tracker.Value(method);
-    return false;
-  }
-  return true;
-}
-
-// Cordz will only update with a new rate once the previously scheduled event
-// has fired. When we disable Cordz, a long delay takes place where we won't
-// consider profiling new Cords. CordzSampleIntervalHelper will burn through
-// that interval and allow for testing that assumes that the average sampling
-// interval is a particular value.
-class CordzSamplingIntervalHelper {
- public:
-  explicit CordzSamplingIntervalHelper(int32_t interval)
-      : orig_mean_interval_(absl::cord_internal::get_cordz_mean_interval()) {
-    absl::cord_internal::set_cordz_mean_interval(interval);
-    absl::cord_internal::cordz_set_next_sample_for_testing(interval);
-  }
-
-  ~CordzSamplingIntervalHelper() {
-    absl::cord_internal::set_cordz_mean_interval(orig_mean_interval_);
-    absl::cord_internal::cordz_set_next_sample_for_testing(orig_mean_interval_);
-  }
-
- private:
-  int32_t orig_mean_interval_;
-};
-
-// Wrapper struct managing a small CordRep `rep`
-struct TestCordRep {
-  absl::Nonnull<cord_internal::CordRepFlat*> rep;
-
-  TestCordRep() {
-    rep = cord_internal::CordRepFlat::New(100);
-    rep->length = 100;
-    memset(rep->Data(), 1, 100);
-  }
-  ~TestCordRep() { cord_internal::CordRep::Unref(rep); }
-};
-
-// Wrapper struct managing a small CordRep `rep`, and
-// an InlineData `data` initialized with that CordRep.
-struct TestCordData {
-  TestCordRep rep;
-  cord_internal::InlineData data{rep.rep};
-};
-
-// Creates a Cord that is not sampled
-template <typename... Args>
-Cord UnsampledCord(Args... args) {
-  CordzSamplingIntervalHelper never(9999);
-  Cord cord(std::forward<Args>(args)...);
-  ABSL_ASSERT(GetCordzInfoForTesting(cord) == nullptr);
-  return cord;
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_CORDZ_TEST_HELPERS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/escaping.h b/projects/llm_framework/include/abseil-cpp/absl/strings/escaping.h
deleted file mode 100644
index 3f34fbfc..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/escaping.h
+++ /dev/null
@@ -1,186 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: escaping.h
-// -----------------------------------------------------------------------------
-//
-// This header file contains string utilities involved in escaping and
-// unescaping strings in various ways.
-
-#ifndef ABSL_STRINGS_ESCAPING_H_
-#define ABSL_STRINGS_ESCAPING_H_
-
-#include <cstddef>
-#include <string>
-#include <vector>
-
-#include "absl/base/attributes.h"
-#include "absl/base/macros.h"
-#include "absl/base/nullability.h"
-#include "absl/strings/ascii.h"
-#include "absl/strings/str_join.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// CUnescape()
-//
-// Unescapes a `source` string and copies it into `dest`, rewriting C-style
-// escape sequences (https://en.cppreference.com/w/cpp/language/escape) into
-// their proper code point equivalents, returning `true` if successful.
-//
-// The following unescape sequences can be handled:
-//
-//   * ASCII escape sequences ('\n','\r','\\', etc.) to their ASCII equivalents
-//   * Octal escape sequences ('\nnn') to byte nnn. The unescaped value must
-//     resolve to a single byte or an error will occur. E.g. values greater than
-//     0xff will produce an error.
-//   * Hexadecimal escape sequences ('\xnn') to byte nn. While an arbitrary
-//     number of following digits are allowed, the unescaped value must resolve
-//     to a single byte or an error will occur. E.g. '\x0045' is equivalent to
-//     '\x45', but '\x1234' will produce an error.
-//   * Unicode escape sequences ('\unnnn' for exactly four hex digits or
-//     '\Unnnnnnnn' for exactly eight hex digits, which will be encoded in
-//     UTF-8. (E.g., `\u2019` unescapes to the three bytes 0xE2, 0x80, and
-//     0x99).
-//
-// If any errors are encountered, this function returns `false`, leaving the
-// `dest` output parameter in an unspecified state, and stores the first
-// encountered error in `error`. To disable error reporting, set `error` to
-// `nullptr` or use the overload with no error reporting below.
-//
-// Example:
-//
-//   std::string s = "foo\\rbar\\nbaz\\t";
-//   std::string unescaped_s;
-//   if (!absl::CUnescape(s, &unescaped_s)) {
-//     ...
-//   }
-//   EXPECT_EQ(unescaped_s, "foo\rbar\nbaz\t");
-bool CUnescape(absl::string_view source, absl::Nonnull<std::string*> dest,
-               absl::Nullable<std::string*> error);
-
-// Overload of `CUnescape()` with no error reporting.
-inline bool CUnescape(absl::string_view source,
-                      absl::Nonnull<std::string*> dest) {
-  return CUnescape(source, dest, nullptr);
-}
-
-// CEscape()
-//
-// Escapes a 'src' string using C-style escapes sequences
-// (https://en.cppreference.com/w/cpp/language/escape), escaping other
-// non-printable/non-whitespace bytes as octal sequences (e.g. "\377").
-//
-// Example:
-//
-//   std::string s = "foo\rbar\tbaz\010\011\012\013\014\x0d\n";
-//   std::string escaped_s = absl::CEscape(s);
-//   EXPECT_EQ(escaped_s, "foo\\rbar\\tbaz\\010\\t\\n\\013\\014\\r\\n");
-std::string CEscape(absl::string_view src);
-
-// CHexEscape()
-//
-// Escapes a 'src' string using C-style escape sequences, escaping
-// other non-printable/non-whitespace bytes as hexadecimal sequences (e.g.
-// "\xFF").
-//
-// Example:
-//
-//   std::string s = "foo\rbar\tbaz\010\011\012\013\014\x0d\n";
-//   std::string escaped_s = absl::CHexEscape(s);
-//   EXPECT_EQ(escaped_s, "foo\\rbar\\tbaz\\x08\\t\\n\\x0b\\x0c\\r\\n");
-std::string CHexEscape(absl::string_view src);
-
-// Utf8SafeCEscape()
-//
-// Escapes a 'src' string using C-style escape sequences, escaping bytes as
-// octal sequences, and passing through UTF-8 characters without conversion.
-// I.e., when encountering any bytes with their high bit set, this function
-// will not escape those values, whether or not they are valid UTF-8.
-std::string Utf8SafeCEscape(absl::string_view src);
-
-// Utf8SafeCHexEscape()
-//
-// Escapes a 'src' string using C-style escape sequences, escaping bytes as
-// hexadecimal sequences, and passing through UTF-8 characters without
-// conversion.
-std::string Utf8SafeCHexEscape(absl::string_view src);
-
-// Base64Escape()
-//
-// Encodes a `src` string into a base64-encoded 'dest' string with padding
-// characters. This function conforms with RFC 4648 section 4 (base64) and RFC
-// 2045.
-void Base64Escape(absl::string_view src, absl::Nonnull<std::string*> dest);
-std::string Base64Escape(absl::string_view src);
-
-// WebSafeBase64Escape()
-//
-// Encodes a `src` string into a base64 string, like Base64Escape() does, but
-// outputs '-' instead of '+' and '_' instead of '/', and does not pad 'dest'.
-// This function conforms with RFC 4648 section 5 (base64url).
-void WebSafeBase64Escape(absl::string_view src,
-                         absl::Nonnull<std::string*> dest);
-std::string WebSafeBase64Escape(absl::string_view src);
-
-// Base64Unescape()
-//
-// Converts a `src` string encoded in Base64 (RFC 4648 section 4) to its binary
-// equivalent, writing it to a `dest` buffer, returning `true` on success. If
-// `src` contains invalid characters, `dest` is cleared and returns `false`.
-// If padding is included (note that `Base64Escape()` does produce it), it must
-// be correct. In the padding, '=' and '.' are treated identically.
-bool Base64Unescape(absl::string_view src, absl::Nonnull<std::string*> dest);
-
-// WebSafeBase64Unescape()
-//
-// Converts a `src` string encoded in "web safe" Base64 (RFC 4648 section 5) to
-// its binary equivalent, writing it to a `dest` buffer. If `src` contains
-// invalid characters, `dest` is cleared and returns `false`. If padding is
-// included (note that `WebSafeBase64Escape()` does not produce it), it must be
-// correct. In the padding, '=' and '.' are treated identically.
-bool WebSafeBase64Unescape(absl::string_view src,
-                           absl::Nonnull<std::string*> dest);
-
-// HexStringToBytes()
-//
-// Converts the hexadecimal encoded data in `hex` into raw bytes in the `bytes`
-// output string.  If `hex` does not consist of valid hexadecimal data, this
-// function returns false and leaves `bytes` in an unspecified state. Returns
-// true on success.
-ABSL_MUST_USE_RESULT bool HexStringToBytes(absl::string_view hex,
-                                           absl::Nonnull<std::string*> bytes);
-
-// HexStringToBytes()
-//
-// Converts an ASCII hex string into bytes, returning binary data of length
-// `from.size()/2`. The input must be valid hexadecimal data, otherwise the
-// return value is unspecified.
-ABSL_DEPRECATED("Use the HexStringToBytes() that returns a bool")
-std::string HexStringToBytes(absl::string_view from);
-
-// BytesToHexString()
-//
-// Converts binary data into an ASCII text string, returning a string of size
-// `2*from.size()`.
-std::string BytesToHexString(absl::string_view from);
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_ESCAPING_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/has_absl_stringify.h b/projects/llm_framework/include/abseil-cpp/absl/strings/has_absl_stringify.h
deleted file mode 100644
index 274a7865..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/has_absl_stringify.h
+++ /dev/null
@@ -1,63 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_HAS_ABSL_STRINGIFY_H_
-#define ABSL_STRINGS_HAS_ABSL_STRINGIFY_H_
-
-#include <type_traits>
-#include <utility>
-
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-namespace strings_internal {
-
-// This is an empty class not intended to be used. It exists so that
-// `HasAbslStringify` can reference a universal class rather than needing to be
-// copied for each new sink.
-class UnimplementedSink {
- public:
-  void Append(size_t count, char ch);
-
-  void Append(string_view v);
-
-  // Support `absl::Format(&sink, format, args...)`.
-  friend void AbslFormatFlush(UnimplementedSink* sink, absl::string_view v);
-};
-
-}  // namespace strings_internal
-
-// `HasAbslStringify<T>` detects if type `T` supports the `AbslStringify()`
-// customization point (see
-// https://abseil.io/docs/cpp/guides/format#abslstringify for the
-// documentation).
-//
-// Note that there are types that can be `StrCat`-ed that do not use the
-// `AbslStringify` customization point (for example, `int`).
-
-template <typename T, typename = void>
-struct HasAbslStringify : std::false_type {};
-
-template <typename T>
-struct HasAbslStringify<
-    T, std::enable_if_t<std::is_void<decltype(AbslStringify(
-           std::declval<strings_internal::UnimplementedSink&>(),
-           std::declval<const T&>()))>::value>> : std::true_type {};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_HAS_ABSL_STRINGIFY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/has_ostream_operator.h b/projects/llm_framework/include/abseil-cpp/absl/strings/has_ostream_operator.h
deleted file mode 100644
index 156ffc74..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/has_ostream_operator.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Copyright 2023 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_HAS_OSTREAM_OPERATOR_H_
-#define ABSL_STRINGS_HAS_OSTREAM_OPERATOR_H_
-
-#include <ostream>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// Detects if type `T` supports streaming to `std::ostream`s with `operator<<`.
-
-template <typename T, typename = void>
-struct HasOstreamOperator : std::false_type {};
-
-template <typename T>
-struct HasOstreamOperator<
-    T, std::enable_if_t<std::is_same<
-           std::ostream&, decltype(std::declval<std::ostream&>()
-                                   << std::declval<const T&>())>::value>>
-    : std::true_type {};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_HAS_OSTREAM_OPERATOR_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/charconv_bigint.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/charconv_bigint.h
deleted file mode 100644
index 5c0c375d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/charconv_bigint.h
+++ /dev/null
@@ -1,423 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CHARCONV_BIGINT_H_
-#define ABSL_STRINGS_INTERNAL_CHARCONV_BIGINT_H_
-
-#include <algorithm>
-#include <cstdint>
-#include <iostream>
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/strings/ascii.h"
-#include "absl/strings/internal/charconv_parse.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace strings_internal {
-
-// The largest power that 5 that can be raised to, and still fit in a uint32_t.
-constexpr int kMaxSmallPowerOfFive = 13;
-// The largest power that 10 that can be raised to, and still fit in a uint32_t.
-constexpr int kMaxSmallPowerOfTen = 9;
-
-ABSL_DLL extern const uint32_t
-    kFiveToNth[kMaxSmallPowerOfFive + 1];
-ABSL_DLL extern const uint32_t kTenToNth[kMaxSmallPowerOfTen + 1];
-
-// Large, fixed-width unsigned integer.
-//
-// Exact rounding for decimal-to-binary floating point conversion requires very
-// large integer math, but a design goal of absl::from_chars is to avoid
-// allocating memory.  The integer precision needed for decimal-to-binary
-// conversions is large but bounded, so a huge fixed-width integer class
-// suffices.
-//
-// This is an intentionally limited big integer class.  Only needed operations
-// are implemented.  All storage lives in an array data member, and all
-// arithmetic is done in-place, to avoid requiring separate storage for operand
-// and result.
-//
-// This is an internal class.  Some methods live in the .cc file, and are
-// instantiated only for the values of max_words we need.
-template <int max_words>
-class BigUnsigned {
- public:
-  static_assert(max_words == 4 || max_words == 84,
-                "unsupported max_words value");
-
-  BigUnsigned() : size_(0), words_{} {}
-  explicit constexpr BigUnsigned(uint64_t v)
-      : size_((v >> 32) ? 2 : v ? 1 : 0),
-        words_{static_cast<uint32_t>(v & 0xffffffffu),
-               static_cast<uint32_t>(v >> 32)} {}
-
-  // Constructs a BigUnsigned from the given string_view containing a decimal
-  // value.  If the input string is not a decimal integer, constructs a 0
-  // instead.
-  explicit BigUnsigned(absl::string_view sv) : size_(0), words_{} {
-    // Check for valid input, returning a 0 otherwise.  This is reasonable
-    // behavior only because this constructor is for unit tests.
-    if (std::find_if_not(sv.begin(), sv.end(), ascii_isdigit) != sv.end() ||
-        sv.empty()) {
-      return;
-    }
-    int exponent_adjust =
-        ReadDigits(sv.data(), sv.data() + sv.size(), Digits10() + 1);
-    if (exponent_adjust > 0) {
-      MultiplyByTenToTheNth(exponent_adjust);
-    }
-  }
-
-  // Loads the mantissa value of a previously-parsed float.
-  //
-  // Returns the associated decimal exponent.  The value of the parsed float is
-  // exactly *this * 10**exponent.
-  int ReadFloatMantissa(const ParsedFloat& fp, int significant_digits);
-
-  // Returns the number of decimal digits of precision this type provides.  All
-  // numbers with this many decimal digits or fewer are representable by this
-  // type.
-  //
-  // Analogous to std::numeric_limits<BigUnsigned>::digits10.
-  static constexpr int Digits10() {
-    // 9975007/1035508 is very slightly less than log10(2**32).
-    return static_cast<uint64_t>(max_words) * 9975007 / 1035508;
-  }
-
-  // Shifts left by the given number of bits.
-  void ShiftLeft(int count) {
-    if (count > 0) {
-      const int word_shift = count / 32;
-      if (word_shift >= max_words) {
-        SetToZero();
-        return;
-      }
-      size_ = (std::min)(size_ + word_shift, max_words);
-      count %= 32;
-      if (count == 0) {
-        std::copy_backward(words_, words_ + size_ - word_shift, words_ + size_);
-      } else {
-        for (int i = (std::min)(size_, max_words - 1); i > word_shift; --i) {
-          words_[i] = (words_[i - word_shift] << count) |
-                      (words_[i - word_shift - 1] >> (32 - count));
-        }
-        words_[word_shift] = words_[0] << count;
-        // Grow size_ if necessary.
-        if (size_ < max_words && words_[size_]) {
-          ++size_;
-        }
-      }
-      std::fill_n(words_, word_shift, 0u);
-    }
-  }
-
-
-  // Multiplies by v in-place.
-  void MultiplyBy(uint32_t v) {
-    if (size_ == 0 || v == 1) {
-      return;
-    }
-    if (v == 0) {
-      SetToZero();
-      return;
-    }
-    const uint64_t factor = v;
-    uint64_t window = 0;
-    for (int i = 0; i < size_; ++i) {
-      window += factor * words_[i];
-      words_[i] = window & 0xffffffff;
-      window >>= 32;
-    }
-    // If carry bits remain and there's space for them, grow size_.
-    if (window && size_ < max_words) {
-      words_[size_] = window & 0xffffffff;
-      ++size_;
-    }
-  }
-
-  void MultiplyBy(uint64_t v) {
-    uint32_t words[2];
-    words[0] = static_cast<uint32_t>(v);
-    words[1] = static_cast<uint32_t>(v >> 32);
-    if (words[1] == 0) {
-      MultiplyBy(words[0]);
-    } else {
-      MultiplyBy(2, words);
-    }
-  }
-
-  // Multiplies in place by 5 to the power of n.  n must be non-negative.
-  void MultiplyByFiveToTheNth(int n) {
-    while (n >= kMaxSmallPowerOfFive) {
-      MultiplyBy(kFiveToNth[kMaxSmallPowerOfFive]);
-      n -= kMaxSmallPowerOfFive;
-    }
-    if (n > 0) {
-      MultiplyBy(kFiveToNth[n]);
-    }
-  }
-
-  // Multiplies in place by 10 to the power of n.  n must be non-negative.
-  void MultiplyByTenToTheNth(int n) {
-    if (n > kMaxSmallPowerOfTen) {
-      // For large n, raise to a power of 5, then shift left by the same amount.
-      // (10**n == 5**n * 2**n.)  This requires fewer multiplications overall.
-      MultiplyByFiveToTheNth(n);
-      ShiftLeft(n);
-    } else if (n > 0) {
-      // We can do this more quickly for very small N by using a single
-      // multiplication.
-      MultiplyBy(kTenToNth[n]);
-    }
-  }
-
-  // Returns the value of 5**n, for non-negative n.  This implementation uses
-  // a lookup table, and is faster then seeding a BigUnsigned with 1 and calling
-  // MultiplyByFiveToTheNth().
-  static BigUnsigned FiveToTheNth(int n);
-
-  // Multiplies by another BigUnsigned, in-place.
-  template <int M>
-  void MultiplyBy(const BigUnsigned<M>& other) {
-    MultiplyBy(other.size(), other.words());
-  }
-
-  void SetToZero() {
-    std::fill_n(words_, size_, 0u);
-    size_ = 0;
-  }
-
-  // Returns the value of the nth word of this BigUnsigned.  This is
-  // range-checked, and returns 0 on out-of-bounds accesses.
-  uint32_t GetWord(int index) const {
-    if (index < 0 || index >= size_) {
-      return 0;
-    }
-    return words_[index];
-  }
-
-  // Returns this integer as a decimal string.  This is not used in the decimal-
-  // to-binary conversion; it is intended to aid in testing.
-  std::string ToString() const;
-
-  int size() const { return size_; }
-  const uint32_t* words() const { return words_; }
-
- private:
-  // Reads the number between [begin, end), possibly containing a decimal point,
-  // into this BigUnsigned.
-  //
-  // Callers are required to ensure [begin, end) contains a valid number, with
-  // one or more decimal digits and at most one decimal point.  This routine
-  // will behave unpredictably if these preconditions are not met.
-  //
-  // Only the first `significant_digits` digits are read.  Digits beyond this
-  // limit are "sticky": If the final significant digit is 0 or 5, and if any
-  // dropped digit is nonzero, then that final significant digit is adjusted up
-  // to 1 or 6.  This adjustment allows for precise rounding.
-  //
-  // Returns `exponent_adjustment`, a power-of-ten exponent adjustment to
-  // account for the decimal point and for dropped significant digits.  After
-  // this function returns,
-  //   actual_value_of_parsed_string ~= *this * 10**exponent_adjustment.
-  int ReadDigits(const char* begin, const char* end, int significant_digits);
-
-  // Performs a step of big integer multiplication.  This computes the full
-  // (64-bit-wide) values that should be added at the given index (step), and
-  // adds to that location in-place.
-  //
-  // Because our math all occurs in place, we must multiply starting from the
-  // highest word working downward.  (This is a bit more expensive due to the
-  // extra carries involved.)
-  //
-  // This must be called in steps, for each word to be calculated, starting from
-  // the high end and working down to 0.  The first value of `step` should be
-  //   `std::min(original_size + other.size_ - 2, max_words - 1)`.
-  // The reason for this expression is that multiplying the i'th word from one
-  // multiplicand and the j'th word of another multiplicand creates a
-  // two-word-wide value to be stored at the (i+j)'th element.  The highest
-  // word indices we will access are `original_size - 1` from this object, and
-  // `other.size_ - 1` from our operand.  Therefore,
-  // `original_size + other.size_ - 2` is the first step we should calculate,
-  // but limited on an upper bound by max_words.
-
-  // Working from high-to-low ensures that we do not overwrite the portions of
-  // the initial value of *this which are still needed for later steps.
-  //
-  // Once called with step == 0, *this contains the result of the
-  // multiplication.
-  //
-  // `original_size` is the size_ of *this before the first call to
-  // MultiplyStep().  `other_words` and `other_size` are the contents of our
-  // operand.  `step` is the step to perform, as described above.
-  void MultiplyStep(int original_size, const uint32_t* other_words,
-                    int other_size, int step);
-
-  void MultiplyBy(int other_size, const uint32_t* other_words) {
-    const int original_size = size_;
-    const int first_step =
-        (std::min)(original_size + other_size - 2, max_words - 1);
-    for (int step = first_step; step >= 0; --step) {
-      MultiplyStep(original_size, other_words, other_size, step);
-    }
-  }
-
-  // Adds a 32-bit value to the index'th word, with carry.
-  void AddWithCarry(int index, uint32_t value) {
-    if (value) {
-      while (index < max_words && value > 0) {
-        words_[index] += value;
-        // carry if we overflowed in this word:
-        if (value > words_[index]) {
-          value = 1;
-          ++index;
-        } else {
-          value = 0;
-        }
-      }
-      size_ = (std::min)(max_words, (std::max)(index + 1, size_));
-    }
-  }
-
-  void AddWithCarry(int index, uint64_t value) {
-    if (value && index < max_words) {
-      uint32_t high = value >> 32;
-      uint32_t low = value & 0xffffffff;
-      words_[index] += low;
-      if (words_[index] < low) {
-        ++high;
-        if (high == 0) {
-          // Carry from the low word caused our high word to overflow.
-          // Short circuit here to do the right thing.
-          AddWithCarry(index + 2, static_cast<uint32_t>(1));
-          return;
-        }
-      }
-      if (high > 0) {
-        AddWithCarry(index + 1, high);
-      } else {
-        // Normally 32-bit AddWithCarry() sets size_, but since we don't call
-        // it when `high` is 0, do it ourselves here.
-        size_ = (std::min)(max_words, (std::max)(index + 1, size_));
-      }
-    }
-  }
-
-  // Divide this in place by a constant divisor.  Returns the remainder of the
-  // division.
-  template <uint32_t divisor>
-  uint32_t DivMod() {
-    uint64_t accumulator = 0;
-    for (int i = size_ - 1; i >= 0; --i) {
-      accumulator <<= 32;
-      accumulator += words_[i];
-      // accumulator / divisor will never overflow an int32_t in this loop
-      words_[i] = static_cast<uint32_t>(accumulator / divisor);
-      accumulator = accumulator % divisor;
-    }
-    while (size_ > 0 && words_[size_ - 1] == 0) {
-      --size_;
-    }
-    return static_cast<uint32_t>(accumulator);
-  }
-
-  // The number of elements in words_ that may carry significant values.
-  // All elements beyond this point are 0.
-  //
-  // When size_ is 0, this BigUnsigned stores the value 0.
-  // When size_ is nonzero, is *not* guaranteed that words_[size_ - 1] is
-  // nonzero.  This can occur due to overflow truncation.
-  // In particular, x.size_ != y.size_ does *not* imply x != y.
-  int size_;
-  uint32_t words_[max_words];
-};
-
-// Compares two big integer instances.
-//
-// Returns -1 if lhs < rhs, 0 if lhs == rhs, and 1 if lhs > rhs.
-template <int N, int M>
-int Compare(const BigUnsigned<N>& lhs, const BigUnsigned<M>& rhs) {
-  int limit = (std::max)(lhs.size(), rhs.size());
-  for (int i = limit - 1; i >= 0; --i) {
-    const uint32_t lhs_word = lhs.GetWord(i);
-    const uint32_t rhs_word = rhs.GetWord(i);
-    if (lhs_word < rhs_word) {
-      return -1;
-    } else if (lhs_word > rhs_word) {
-      return 1;
-    }
-  }
-  return 0;
-}
-
-template <int N, int M>
-bool operator==(const BigUnsigned<N>& lhs, const BigUnsigned<M>& rhs) {
-  int limit = (std::max)(lhs.size(), rhs.size());
-  for (int i = 0; i < limit; ++i) {
-    if (lhs.GetWord(i) != rhs.GetWord(i)) {
-      return false;
-    }
-  }
-  return true;
-}
-
-template <int N, int M>
-bool operator!=(const BigUnsigned<N>& lhs, const BigUnsigned<M>& rhs) {
-  return !(lhs == rhs);
-}
-
-template <int N, int M>
-bool operator<(const BigUnsigned<N>& lhs, const BigUnsigned<M>& rhs) {
-  return Compare(lhs, rhs) == -1;
-}
-
-template <int N, int M>
-bool operator>(const BigUnsigned<N>& lhs, const BigUnsigned<M>& rhs) {
-  return rhs < lhs;
-}
-template <int N, int M>
-bool operator<=(const BigUnsigned<N>& lhs, const BigUnsigned<M>& rhs) {
-  return !(rhs < lhs);
-}
-template <int N, int M>
-bool operator>=(const BigUnsigned<N>& lhs, const BigUnsigned<M>& rhs) {
-  return !(lhs < rhs);
-}
-
-// Output operator for BigUnsigned, for testing purposes only.
-template <int N>
-std::ostream& operator<<(std::ostream& os, const BigUnsigned<N>& num) {
-  return os << num.ToString();
-}
-
-// Explicit instantiation declarations for the sizes of BigUnsigned that we
-// are using.
-//
-// For now, the choices of 4 and 84 are arbitrary; 4 is a small value that is
-// still bigger than an int128, and 84 is a large value we will want to use
-// in the from_chars implementation.
-//
-// Comments justifying the use of 84 belong in the from_chars implementation,
-// and will be added in a follow-up CL.
-extern template class BigUnsigned<4>;
-extern template class BigUnsigned<84>;
-
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CHARCONV_BIGINT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/charconv_parse.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/charconv_parse.h
deleted file mode 100644
index 505998b5..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/charconv_parse.h
+++ /dev/null
@@ -1,99 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CHARCONV_PARSE_H_
-#define ABSL_STRINGS_INTERNAL_CHARCONV_PARSE_H_
-
-#include <cstdint>
-
-#include "absl/base/config.h"
-#include "absl/strings/charconv.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace strings_internal {
-
-// Enum indicating whether a parsed float is a number or special value.
-enum class FloatType { kNumber, kInfinity, kNan };
-
-// The decomposed parts of a parsed `float` or `double`.
-struct ParsedFloat {
-  // Representation of the parsed mantissa, with the decimal point adjusted to
-  // make it an integer.
-  //
-  // During decimal scanning, this contains 19 significant digits worth of
-  // mantissa value.  If digits beyond this point are found, they
-  // are truncated, and if any of these dropped digits are nonzero, then
-  // `mantissa` is inexact, and the full mantissa is stored in [subrange_begin,
-  // subrange_end).
-  //
-  // During hexadecimal scanning, this contains 15 significant hex digits worth
-  // of mantissa value.  Digits beyond this point are sticky -- they are
-  // truncated, but if any dropped digits are nonzero, the low bit of mantissa
-  // will be set.  (This allows for precise rounding, and avoids the need
-  // to store the full mantissa in [subrange_begin, subrange_end).)
-  uint64_t mantissa = 0;
-
-  // Floating point expontent.  This reflects any decimal point adjustments and
-  // any truncated digits from the mantissa.  The absolute value of the parsed
-  // number is represented by mantissa * (base ** exponent), where base==10 for
-  // decimal floats, and base==2 for hexadecimal floats.
-  int exponent = 0;
-
-  // The literal exponent value scanned from the input, or 0 if none was
-  // present.  This does not reflect any adjustments applied to mantissa.
-  int literal_exponent = 0;
-
-  // The type of number scanned.
-  FloatType type = FloatType::kNumber;
-
-  // When non-null, [subrange_begin, subrange_end) marks a range of characters
-  // that require further processing.  The meaning is dependent on float type.
-  // If type == kNumber and this is set, this is a "wide input": the input
-  // mantissa contained more than 19 digits.  The range contains the full
-  // mantissa.  It plus `literal_exponent` need to be examined to find the best
-  // floating point match.
-  // If type == kNan and this is set, the range marks the contents of a
-  // matched parenthesized character region after the NaN.
-  const char* subrange_begin = nullptr;
-  const char* subrange_end = nullptr;
-
-  // One-past-the-end of the successfully parsed region, or nullptr if no
-  // matching pattern was found.
-  const char* end = nullptr;
-};
-
-// Read the floating point number in the provided range, and populate
-// ParsedFloat accordingly.
-//
-// format_flags is a bitmask value specifying what patterns this API will match.
-// `scientific` and `fixed`  are honored per std::from_chars rules
-// ([utility.from.chars], C++17): if exactly one of these bits is set, then an
-// exponent is required, or dislallowed, respectively.
-//
-// Template parameter `base` must be either 10 or 16.  For base 16, a "0x" is
-// *not* consumed.  The `hex` bit from format_flags is ignored by ParseFloat.
-template <int base>
-ParsedFloat ParseFloat(const char* begin, const char* end,
-                       absl::chars_format format_flags);
-
-extern template ParsedFloat ParseFloat<10>(const char* begin, const char* end,
-                                           absl::chars_format format_flags);
-extern template ParsedFloat ParseFloat<16>(const char* begin, const char* end,
-                                           absl::chars_format format_flags);
-
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-#endif  // ABSL_STRINGS_INTERNAL_CHARCONV_PARSE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_data_edge.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_data_edge.h
deleted file mode 100644
index e18b33e1..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_data_edge.h
+++ /dev/null
@@ -1,63 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CORD_DATA_EDGE_H_
-#define ABSL_STRINGS_INTERNAL_CORD_DATA_EDGE_H_
-
-#include <cassert>
-#include <cstddef>
-
-#include "absl/base/config.h"
-#include "absl/strings/internal/cord_internal.h"
-#include "absl/strings/internal/cord_rep_flat.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// Returns true if the provided rep is a FLAT, EXTERNAL or a SUBSTRING node
-// holding a FLAT or EXTERNAL child rep. Requires `rep != nullptr`.
-inline bool IsDataEdge(const CordRep* edge) {
-  assert(edge != nullptr);
-
-  // The fast path is that `edge` is an EXTERNAL or FLAT node, making the below
-  // if a single, well predicted branch. We then repeat the FLAT or EXTERNAL
-  // check in the slow path of the SUBSTRING check to optimize for the hot path.
-  if (edge->tag == EXTERNAL || edge->tag >= FLAT) return true;
-  if (edge->tag == SUBSTRING) edge = edge->substring()->child;
-  return edge->tag == EXTERNAL || edge->tag >= FLAT;
-}
-
-// Returns the `absl::string_view` data reference for the provided data edge.
-// Requires 'IsDataEdge(edge) == true`.
-inline absl::string_view EdgeData(const CordRep* edge) {
-  assert(IsDataEdge(edge));
-
-  size_t offset = 0;
-  const size_t length = edge->length;
-  if (edge->IsSubstring()) {
-    offset = edge->substring()->start;
-    edge = edge->substring()->child;
-  }
-  return edge->tag >= FLAT
-             ? absl::string_view{edge->flat()->Data() + offset, length}
-             : absl::string_view{edge->external()->base + offset, length};
-}
-
-}  // namespace cord_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CORD_DATA_EDGE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_internal.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_internal.h
deleted file mode 100644
index a52deea1..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_internal.h
+++ /dev/null
@@ -1,912 +0,0 @@
-// Copyright 2021 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CORD_INTERNAL_H_
-#define ABSL_STRINGS_INTERNAL_CORD_INTERNAL_H_
-
-#include <atomic>
-#include <cassert>
-#include <cstddef>
-#include <cstdint>
-#include <type_traits>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/internal/endian.h"
-#include "absl/base/internal/invoke.h"
-#include "absl/base/optimization.h"
-#include "absl/container/internal/compressed_tuple.h"
-#include "absl/container/internal/container_memory.h"
-#include "absl/meta/type_traits.h"
-#include "absl/strings/string_view.h"
-
-// We can only add poisoning if we can detect consteval executions.
-#if defined(ABSL_HAVE_CONSTANT_EVALUATED) && \
-    (defined(ABSL_HAVE_ADDRESS_SANITIZER) || \
-     defined(ABSL_HAVE_MEMORY_SANITIZER))
-#define ABSL_INTERNAL_CORD_HAVE_SANITIZER 1
-#endif
-
-#define ABSL_CORD_INTERNAL_NO_SANITIZE \
-  ABSL_ATTRIBUTE_NO_SANITIZE_ADDRESS ABSL_ATTRIBUTE_NO_SANITIZE_MEMORY
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// The overhead of a vtable is too much for Cord, so we roll our own subclasses
-// using only a single byte to differentiate classes from each other - the "tag"
-// byte.  Define the subclasses first so we can provide downcasting helper
-// functions in the base class.
-struct CordRep;
-struct CordRepConcat;
-struct CordRepExternal;
-struct CordRepFlat;
-struct CordRepSubstring;
-struct CordRepCrc;
-class CordRepBtree;
-
-class CordzInfo;
-
-// Default feature enable states for cord ring buffers
-enum CordFeatureDefaults { kCordShallowSubcordsDefault = false };
-
-extern std::atomic<bool> shallow_subcords_enabled;
-
-inline void enable_shallow_subcords(bool enable) {
-  shallow_subcords_enabled.store(enable, std::memory_order_relaxed);
-}
-
-enum Constants {
-  // The inlined size to use with absl::InlinedVector.
-  //
-  // Note: The InlinedVectors in this file (and in cord.h) do not need to use
-  // the same value for their inlined size. The fact that they do is historical.
-  // It may be desirable for each to use a different inlined size optimized for
-  // that InlinedVector's usage.
-  //
-  // TODO(jgm): Benchmark to see if there's a more optimal value than 47 for
-  // the inlined vector size (47 exists for backward compatibility).
-  kInlinedVectorSize = 47,
-
-  // Prefer copying blocks of at most this size, otherwise reference count.
-  kMaxBytesToCopy = 511
-};
-
-// Emits a fatal error "Unexpected node type: xyz" and aborts the program.
-ABSL_ATTRIBUTE_NORETURN void LogFatalNodeType(CordRep* rep);
-
-// Fast implementation of memmove for up to 15 bytes. This implementation is
-// safe for overlapping regions. If nullify_tail is true, the destination is
-// padded with '\0' up to 15 bytes.
-template <bool nullify_tail = false>
-inline void SmallMemmove(char* dst, const char* src, size_t n) {
-  if (n >= 8) {
-    assert(n <= 15);
-    uint64_t buf1;
-    uint64_t buf2;
-    memcpy(&buf1, src, 8);
-    memcpy(&buf2, src + n - 8, 8);
-    if (nullify_tail) {
-      memset(dst + 7, 0, 8);
-    }
-    // GCC 12 has a false-positive -Wstringop-overflow warning here.
-#if ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(12, 0)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wstringop-overflow"
-#endif
-    memcpy(dst, &buf1, 8);
-    memcpy(dst + n - 8, &buf2, 8);
-#if ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(12, 0)
-#pragma GCC diagnostic pop
-#endif
-  } else if (n >= 4) {
-    uint32_t buf1;
-    uint32_t buf2;
-    memcpy(&buf1, src, 4);
-    memcpy(&buf2, src + n - 4, 4);
-    if (nullify_tail) {
-      memset(dst + 4, 0, 4);
-      memset(dst + 7, 0, 8);
-    }
-    memcpy(dst, &buf1, 4);
-    memcpy(dst + n - 4, &buf2, 4);
-  } else {
-    if (n != 0) {
-      dst[0] = src[0];
-      dst[n / 2] = src[n / 2];
-      dst[n - 1] = src[n - 1];
-    }
-    if (nullify_tail) {
-      memset(dst + 7, 0, 8);
-      memset(dst + n, 0, 8);
-    }
-  }
-}
-
-// Compact class for tracking the reference count and state flags for CordRep
-// instances.  Data is stored in an atomic int32_t for compactness and speed.
-class RefcountAndFlags {
- public:
-  constexpr RefcountAndFlags() : count_{kRefIncrement} {}
-  struct Immortal {};
-  explicit constexpr RefcountAndFlags(Immortal) : count_(kImmortalFlag) {}
-
-  // Increments the reference count. Imposes no memory ordering.
-  inline void Increment() {
-    count_.fetch_add(kRefIncrement, std::memory_order_relaxed);
-  }
-
-  // Asserts that the current refcount is greater than 0. If the refcount is
-  // greater than 1, decrements the reference count.
-  //
-  // Returns false if there are no references outstanding; true otherwise.
-  // Inserts barriers to ensure that state written before this method returns
-  // false will be visible to a thread that just observed this method returning
-  // false.  Always returns false when the immortal bit is set.
-  inline bool Decrement() {
-    int32_t refcount = count_.load(std::memory_order_acquire);
-    assert(refcount > 0 || refcount & kImmortalFlag);
-    return refcount != kRefIncrement &&
-           count_.fetch_sub(kRefIncrement, std::memory_order_acq_rel) !=
-               kRefIncrement;
-  }
-
-  // Same as Decrement but expect that refcount is greater than 1.
-  inline bool DecrementExpectHighRefcount() {
-    int32_t refcount =
-        count_.fetch_sub(kRefIncrement, std::memory_order_acq_rel);
-    assert(refcount > 0 || refcount & kImmortalFlag);
-    return refcount != kRefIncrement;
-  }
-
-  // Returns the current reference count using acquire semantics.
-  inline size_t Get() const {
-    return static_cast<size_t>(count_.load(std::memory_order_acquire) >>
-                               kNumFlags);
-  }
-
-  // Returns whether the atomic integer is 1.
-  // If the reference count is used in the conventional way, a
-  // reference count of 1 implies that the current thread owns the
-  // reference and no other thread shares it.
-  // This call performs the test for a reference count of one, and
-  // performs the memory barrier needed for the owning thread
-  // to act on the object, knowing that it has exclusive access to the
-  // object. Always returns false when the immortal bit is set.
-  inline bool IsOne() {
-    return count_.load(std::memory_order_acquire) == kRefIncrement;
-  }
-
-  bool IsImmortal() const {
-    return (count_.load(std::memory_order_relaxed) & kImmortalFlag) != 0;
-  }
-
- private:
-  // We reserve the bottom bit for flag.
-  // kImmortalBit indicates that this entity should never be collected; it is
-  // used for the StringConstant constructor to avoid collecting immutable
-  // constant cords.
-  enum Flags {
-    kNumFlags = 1,
-
-    kImmortalFlag = 0x1,
-    kRefIncrement = (1 << kNumFlags),
-  };
-
-  std::atomic<int32_t> count_;
-};
-
-// Various representations that we allow
-enum CordRepKind {
-  UNUSED_0 = 0,
-  SUBSTRING = 1,
-  CRC = 2,
-  BTREE = 3,
-  UNUSED_4 = 4,
-  EXTERNAL = 5,
-
-  // We have different tags for different sized flat arrays,
-  // starting with FLAT, and limited to MAX_FLAT_TAG. The below values map to an
-  // allocated range of 32 bytes to 256 KB. The current granularity is:
-  // - 8 byte granularity for flat sizes in [32 - 512]
-  // - 64 byte granularity for flat sizes in (512 - 8KiB]
-  // - 4KiB byte granularity for flat sizes in (8KiB, 256 KiB]
-  // If a new tag is needed in the future, then 'FLAT' and 'MAX_FLAT_TAG' should
-  // be adjusted as well as the Tag <---> Size mapping logic so that FLAT still
-  // represents the minimum flat allocation size. (32 bytes as of now).
-  FLAT = 6,
-  MAX_FLAT_TAG = 248
-};
-
-// There are various locations where we want to check if some rep is a 'plain'
-// data edge, i.e. an external or flat rep. By having FLAT == EXTERNAL + 1, we
-// can perform this check in a single branch as 'tag >= EXTERNAL'
-// Note that we can leave this optimization to the compiler. The compiler will
-// DTRT when it sees a condition like `tag == EXTERNAL || tag >= FLAT`.
-static_assert(FLAT == EXTERNAL + 1, "EXTERNAL and FLAT not consecutive");
-
-struct CordRep {
-  // Result from an `extract edge` operation. Contains the (possibly changed)
-  // tree node as well as the extracted edge, or {tree, nullptr} if no edge
-  // could be extracted.
-  // On success, the returned `tree` value is null if `extracted` was the only
-  // data edge inside the tree, a data edge if there were only two data edges in
-  // the tree, or the (possibly new / smaller) remaining tree with the extracted
-  // data edge removed.
-  struct ExtractResult {
-    CordRep* tree;
-    CordRep* extracted;
-  };
-
-  CordRep() = default;
-  constexpr CordRep(RefcountAndFlags::Immortal immortal, size_t l)
-      : length(l), refcount(immortal), tag(EXTERNAL), storage{} {}
-
-  // The following three fields have to be less than 32 bytes since
-  // that is the smallest supported flat node size. Some code optimizations rely
-  // on the specific layout of these fields. Notably: the non-trivial field
-  // `refcount` being preceded by `length`, and being tailed by POD data
-  // members only.
-  // # LINT.IfChange
-  size_t length;
-  RefcountAndFlags refcount;
-  // If tag < FLAT, it represents CordRepKind and indicates the type of node.
-  // Otherwise, the node type is CordRepFlat and the tag is the encoded size.
-  uint8_t tag;
-
-  // `storage` provides two main purposes:
-  // - the starting point for FlatCordRep.Data() [flexible-array-member]
-  // - 3 bytes of additional storage for use by derived classes.
-  // The latter is used by CordrepConcat and CordRepBtree. CordRepConcat stores
-  // a 'depth' value in storage[0], and the (future) CordRepBtree class stores
-  // `height`, `begin` and `end` in the 3 entries. Otherwise we would need to
-  // allocate room for these in the derived class, as not all compilers reuse
-  // padding space from the base class (clang and gcc do, MSVC does not, etc)
-  uint8_t storage[3];
-  // # LINT.ThenChange(cord_rep_btree.h:copy_raw)
-
-  // Returns true if this instance's tag matches the requested type.
-  constexpr bool IsSubstring() const { return tag == SUBSTRING; }
-  constexpr bool IsCrc() const { return tag == CRC; }
-  constexpr bool IsExternal() const { return tag == EXTERNAL; }
-  constexpr bool IsFlat() const { return tag >= FLAT; }
-  constexpr bool IsBtree() const { return tag == BTREE; }
-
-  inline CordRepSubstring* substring();
-  inline const CordRepSubstring* substring() const;
-  inline CordRepCrc* crc();
-  inline const CordRepCrc* crc() const;
-  inline CordRepExternal* external();
-  inline const CordRepExternal* external() const;
-  inline CordRepFlat* flat();
-  inline const CordRepFlat* flat() const;
-  inline CordRepBtree* btree();
-  inline const CordRepBtree* btree() const;
-
-  // --------------------------------------------------------------------
-  // Memory management
-
-  // Destroys the provided `rep`.
-  static void Destroy(CordRep* rep);
-
-  // Increments the reference count of `rep`.
-  // Requires `rep` to be a non-null pointer value.
-  static inline CordRep* Ref(CordRep* rep);
-
-  // Decrements the reference count of `rep`. Destroys rep if count reaches
-  // zero. Requires `rep` to be a non-null pointer value.
-  static inline void Unref(CordRep* rep);
-};
-
-struct CordRepSubstring : public CordRep {
-  size_t start;  // Starting offset of substring in child
-  CordRep* child;
-
-  // Creates a substring on `child`, adopting a reference on `child`.
-  // Requires `child` to be either a flat or external node, and `pos` and `n` to
-  // form a non-empty partial sub range of `'child`, i.e.:
-  // `n > 0 && n < length && n + pos <= length`
-  static inline CordRepSubstring* Create(CordRep* child, size_t pos, size_t n);
-
-  // Creates a substring of `rep`. Does not adopt a reference on `rep`.
-  // Requires `IsDataEdge(rep) && n > 0 && pos + n <= rep->length`.
-  // If `n == rep->length` then this method returns `CordRep::Ref(rep)`
-  // If `rep` is a substring of a flat or external node, then this method will
-  // return a new substring of that flat or external node with `pos` adjusted
-  // with the original `start` position.
-  static inline CordRep* Substring(CordRep* rep, size_t pos, size_t n);
-};
-
-// Type for function pointer that will invoke the releaser function and also
-// delete the `CordRepExternalImpl` corresponding to the passed in
-// `CordRepExternal`.
-using ExternalReleaserInvoker = void (*)(CordRepExternal*);
-
-// External CordReps are allocated together with a type erased releaser. The
-// releaser is stored in the memory directly following the CordRepExternal.
-struct CordRepExternal : public CordRep {
-  CordRepExternal() = default;
-  explicit constexpr CordRepExternal(absl::string_view str)
-      : CordRep(RefcountAndFlags::Immortal{}, str.size()),
-        base(str.data()),
-        releaser_invoker(nullptr) {}
-
-  const char* base;
-  // Pointer to function that knows how to call and destroy the releaser.
-  ExternalReleaserInvoker releaser_invoker;
-
-  // Deletes (releases) the external rep.
-  // Requires rep != nullptr and rep->IsExternal()
-  static void Delete(CordRep* rep);
-};
-
-// Use go/ranked-overloads for dispatching.
-struct Rank0 {};
-struct Rank1 : Rank0 {};
-
-template <typename Releaser, typename = ::absl::base_internal::invoke_result_t<
-                                 Releaser, absl::string_view>>
-void InvokeReleaser(Rank1, Releaser&& releaser, absl::string_view data) {
-  ::absl::base_internal::invoke(std::forward<Releaser>(releaser), data);
-}
-
-template <typename Releaser,
-          typename = ::absl::base_internal::invoke_result_t<Releaser>>
-void InvokeReleaser(Rank0, Releaser&& releaser, absl::string_view) {
-  ::absl::base_internal::invoke(std::forward<Releaser>(releaser));
-}
-
-// We use CompressedTuple so that we can benefit from EBCO.
-template <typename Releaser>
-struct CordRepExternalImpl
-    : public CordRepExternal,
-      public ::absl::container_internal::CompressedTuple<Releaser> {
-  // The extra int arg is so that we can avoid interfering with copy/move
-  // constructors while still benefitting from perfect forwarding.
-  template <typename T>
-  CordRepExternalImpl(T&& releaser, int)
-      : CordRepExternalImpl::CompressedTuple(std::forward<T>(releaser)) {
-    this->releaser_invoker = &Release;
-  }
-
-  ~CordRepExternalImpl() {
-    InvokeReleaser(Rank1{}, std::move(this->template get<0>()),
-                   absl::string_view(base, length));
-  }
-
-  static void Release(CordRepExternal* rep) {
-    delete static_cast<CordRepExternalImpl*>(rep);
-  }
-};
-
-inline CordRepSubstring* CordRepSubstring::Create(CordRep* child, size_t pos,
-                                                  size_t n) {
-  assert(child != nullptr);
-  assert(n > 0);
-  assert(n < child->length);
-  assert(pos < child->length);
-  assert(n <= child->length - pos);
-
-  // TODO(b/217376272): Harden internal logic.
-  // Move to strategical places inside the Cord logic and make this an assert.
-  if (ABSL_PREDICT_FALSE(!(child->IsExternal() || child->IsFlat()))) {
-    LogFatalNodeType(child);
-  }
-
-  CordRepSubstring* rep = new CordRepSubstring();
-  rep->length = n;
-  rep->tag = SUBSTRING;
-  rep->start = pos;
-  rep->child = child;
-  return rep;
-}
-
-inline CordRep* CordRepSubstring::Substring(CordRep* rep, size_t pos,
-                                            size_t n) {
-  assert(rep != nullptr);
-  assert(n != 0);
-  assert(pos < rep->length);
-  assert(n <= rep->length - pos);
-  if (n == rep->length) return CordRep::Ref(rep);
-  if (rep->IsSubstring()) {
-    pos += rep->substring()->start;
-    rep = rep->substring()->child;
-  }
-  CordRepSubstring* substr = new CordRepSubstring();
-  substr->length = n;
-  substr->tag = SUBSTRING;
-  substr->start = pos;
-  substr->child = CordRep::Ref(rep);
-  return substr;
-}
-
-inline void CordRepExternal::Delete(CordRep* rep) {
-  assert(rep != nullptr && rep->IsExternal());
-  auto* rep_external = static_cast<CordRepExternal*>(rep);
-  assert(rep_external->releaser_invoker != nullptr);
-  rep_external->releaser_invoker(rep_external);
-}
-
-template <typename Str>
-struct ConstInitExternalStorage {
-  ABSL_CONST_INIT static CordRepExternal value;
-};
-
-template <typename Str>
-ABSL_CONST_INIT CordRepExternal
-    ConstInitExternalStorage<Str>::value(Str::value);
-
-enum {
-  kMaxInline = 15,
-};
-
-constexpr char GetOrNull(absl::string_view data, size_t pos) {
-  return pos < data.size() ? data[pos] : '\0';
-}
-
-// We store cordz_info as 64 bit pointer value in little endian format. This
-// guarantees that the least significant byte of cordz_info matches the first
-// byte of the inline data representation in `data`, which holds the inlined
-// size or the 'is_tree' bit.
-using cordz_info_t = int64_t;
-
-// Assert that the `cordz_info` pointer value perfectly overlaps the last half
-// of `data` and can hold a pointer value.
-static_assert(sizeof(cordz_info_t) * 2 == kMaxInline + 1, "");
-static_assert(sizeof(cordz_info_t) >= sizeof(intptr_t), "");
-
-// LittleEndianByte() creates a little endian representation of 'value', i.e.:
-// a little endian value where the first byte in the host's representation
-// holds 'value`, with all other bytes being 0.
-static constexpr cordz_info_t LittleEndianByte(unsigned char value) {
-#if defined(ABSL_IS_BIG_ENDIAN)
-  return static_cast<cordz_info_t>(value) << ((sizeof(cordz_info_t) - 1) * 8);
-#else
-  return value;
-#endif
-}
-
-class InlineData {
- public:
-  // DefaultInitType forces the use of the default initialization constructor.
-  enum DefaultInitType { kDefaultInit };
-
-  // kNullCordzInfo holds the little endian representation of intptr_t(1)
-  // This is the 'null' / initial value of 'cordz_info'. The null value
-  // is specifically big endian 1 as with 64-bit pointers, the last
-  // byte of cordz_info overlaps with the last byte holding the tag.
-  static constexpr cordz_info_t kNullCordzInfo = LittleEndianByte(1);
-
-  // kTagOffset contains the offset of the control byte / tag. This constant is
-  // intended mostly for debugging purposes: do not remove this constant as it
-  // is actively inspected and used by gdb pretty printing code.
-  static constexpr size_t kTagOffset = 0;
-
-  // Implement `~InlineData()` conditionally: we only need this destructor to
-  // unpoison poisoned instances under *SAN, and it will only compile correctly
-  // if the current compiler supports `absl::is_constant_evaluated()`.
-#ifdef ABSL_INTERNAL_CORD_HAVE_SANITIZER
-  ~InlineData() noexcept { unpoison(); }
-#endif
-
-  constexpr InlineData() noexcept { poison_this(); }
-
-  explicit InlineData(DefaultInitType) noexcept : rep_(kDefaultInit) {
-    poison_this();
-  }
-
-  explicit InlineData(CordRep* rep) noexcept : rep_(rep) {
-    ABSL_ASSERT(rep != nullptr);
-  }
-
-  // Explicit constexpr constructor to create a constexpr InlineData
-  // value. Creates an inlined SSO value if `rep` is null, otherwise
-  // creates a tree instance value.
-  constexpr InlineData(absl::string_view sv, CordRep* rep) noexcept
-      : rep_(rep ? Rep(rep) : Rep(sv)) {
-    poison();
-  }
-
-  constexpr InlineData(const InlineData& rhs) noexcept;
-  InlineData& operator=(const InlineData& rhs) noexcept;
-  friend void swap(InlineData& lhs, InlineData& rhs) noexcept;
-
-  friend bool operator==(const InlineData& lhs, const InlineData& rhs) {
-#ifdef ABSL_INTERNAL_CORD_HAVE_SANITIZER
-    const Rep l = lhs.rep_.SanitizerSafeCopy();
-    const Rep r = rhs.rep_.SanitizerSafeCopy();
-    return memcmp(&l, &r, sizeof(l)) == 0;
-#else
-    return memcmp(&lhs, &rhs, sizeof(lhs)) == 0;
-#endif
-  }
-  friend bool operator!=(const InlineData& lhs, const InlineData& rhs) {
-    return !operator==(lhs, rhs);
-  }
-
-  // Poisons the unused inlined SSO data if the current instance
-  // is inlined, else un-poisons the entire instance.
-  constexpr void poison();
-
-  // Un-poisons this instance.
-  constexpr void unpoison();
-
-  // Poisons the current instance. This is used on default initialization.
-  constexpr void poison_this();
-
-  // Returns true if the current instance is empty.
-  // The 'empty value' is an inlined data value of zero length.
-  bool is_empty() const { return rep_.tag() == 0; }
-
-  // Returns true if the current instance holds a tree value.
-  bool is_tree() const { return (rep_.tag() & 1) != 0; }
-
-  // Returns true if the current instance holds a cordz_info value.
-  // Requires the current instance to hold a tree value.
-  bool is_profiled() const {
-    assert(is_tree());
-    return rep_.cordz_info() != kNullCordzInfo;
-  }
-
-  // Returns true if either of the provided instances hold a cordz_info value.
-  // This method is more efficient than the equivalent `data1.is_profiled() ||
-  // data2.is_profiled()`. Requires both arguments to hold a tree.
-  static bool is_either_profiled(const InlineData& data1,
-                                 const InlineData& data2) {
-    assert(data1.is_tree() && data2.is_tree());
-    return (data1.rep_.cordz_info() | data2.rep_.cordz_info()) !=
-           kNullCordzInfo;
-  }
-
-  // Returns the cordz_info sampling instance for this instance, or nullptr
-  // if the current instance is not sampled and does not have CordzInfo data.
-  // Requires the current instance to hold a tree value.
-  CordzInfo* cordz_info() const {
-    assert(is_tree());
-    intptr_t info = static_cast<intptr_t>(absl::little_endian::ToHost64(
-        static_cast<uint64_t>(rep_.cordz_info())));
-    assert(info & 1);
-    return reinterpret_cast<CordzInfo*>(info - 1);
-  }
-
-  // Sets the current cordz_info sampling instance for this instance, or nullptr
-  // if the current instance is not sampled and does not have CordzInfo data.
-  // Requires the current instance to hold a tree value.
-  void set_cordz_info(CordzInfo* cordz_info) {
-    assert(is_tree());
-    uintptr_t info = reinterpret_cast<uintptr_t>(cordz_info) | 1;
-    rep_.set_cordz_info(
-        static_cast<cordz_info_t>(absl::little_endian::FromHost64(info)));
-  }
-
-  // Resets the current cordz_info to null / empty.
-  void clear_cordz_info() {
-    assert(is_tree());
-    rep_.set_cordz_info(kNullCordzInfo);
-  }
-
-  // Returns a read only pointer to the character data inside this instance.
-  // Requires the current instance to hold inline data.
-  const char* as_chars() const {
-    assert(!is_tree());
-    return rep_.as_chars();
-  }
-
-  // Returns a mutable pointer to the character data inside this instance.
-  // Should be used for 'write only' operations setting an inlined value.
-  // Applications can set the value of inlined data either before or after
-  // setting the inlined size, i.e., both of the below are valid:
-  //
-  //   // Set inlined data and inline size
-  //   memcpy(data_.as_chars(), data, size);
-  //   data_.set_inline_size(size);
-  //
-  //   // Set inlined size and inline data
-  //   data_.set_inline_size(size);
-  //   memcpy(data_.as_chars(), data, size);
-  //
-  // It's an error to read from the returned pointer without a preceding write
-  // if the current instance does not hold inline data, i.e.: is_tree() == true.
-  char* as_chars() { return rep_.as_chars(); }
-
-  // Returns the tree value of this value.
-  // Requires the current instance to hold a tree value.
-  CordRep* as_tree() const {
-    assert(is_tree());
-    return rep_.tree();
-  }
-
-  void set_inline_data(const char* data, size_t n) {
-    ABSL_ASSERT(n <= kMaxInline);
-    unpoison();
-    rep_.set_tag(static_cast<int8_t>(n << 1));
-    SmallMemmove<true>(rep_.as_chars(), data, n);
-    poison();
-  }
-
-  void copy_max_inline_to(char* dst) const {
-    assert(!is_tree());
-    memcpy(dst, rep_.SanitizerSafeCopy().as_chars(), kMaxInline);
-  }
-
-  // Initialize this instance to holding the tree value `rep`,
-  // initializing the cordz_info to null, i.e.: 'not profiled'.
-  void make_tree(CordRep* rep) {
-    unpoison();
-    rep_.make_tree(rep);
-  }
-
-  // Set the tree value of this instance to 'rep`.
-  // Requires the current instance to already hold a tree value.
-  // Does not affect the value of cordz_info.
-  void set_tree(CordRep* rep) {
-    assert(is_tree());
-    rep_.set_tree(rep);
-  }
-
-  // Returns the size of the inlined character data inside this instance.
-  // Requires the current instance to hold inline data.
-  size_t inline_size() const { return rep_.inline_size(); }
-
-  // Sets the size of the inlined character data inside this instance.
-  // Requires `size` to be <= kMaxInline.
-  // See the documentation on 'as_chars()' for more information and examples.
-  void set_inline_size(size_t size) {
-    unpoison();
-    rep_.set_inline_size(size);
-    poison();
-  }
-
-  // Compares 'this' inlined data  with rhs. The comparison is a straightforward
-  // lexicographic comparison. `Compare()` returns values as follows:
-  //
-  //   -1  'this' InlineData instance is smaller
-  //    0  the InlineData instances are equal
-  //    1  'this' InlineData instance larger
-  int Compare(const InlineData& rhs) const {
-    return Compare(rep_.SanitizerSafeCopy(), rhs.rep_.SanitizerSafeCopy());
-  }
-
- private:
-  struct Rep {
-    // See cordz_info_t for forced alignment and size of `cordz_info` details.
-    struct AsTree {
-      explicit constexpr AsTree(absl::cord_internal::CordRep* tree)
-          : rep(tree) {}
-      cordz_info_t cordz_info = kNullCordzInfo;
-      absl::cord_internal::CordRep* rep;
-    };
-
-    explicit Rep(DefaultInitType) {}
-    constexpr Rep() : data{0} {}
-    constexpr Rep(const Rep&) = default;
-    constexpr Rep& operator=(const Rep&) = default;
-
-    explicit constexpr Rep(CordRep* rep) : as_tree(rep) {}
-
-    explicit constexpr Rep(absl::string_view chars)
-        : data{static_cast<char>((chars.size() << 1)),
-               GetOrNull(chars, 0),
-               GetOrNull(chars, 1),
-               GetOrNull(chars, 2),
-               GetOrNull(chars, 3),
-               GetOrNull(chars, 4),
-               GetOrNull(chars, 5),
-               GetOrNull(chars, 6),
-               GetOrNull(chars, 7),
-               GetOrNull(chars, 8),
-               GetOrNull(chars, 9),
-               GetOrNull(chars, 10),
-               GetOrNull(chars, 11),
-               GetOrNull(chars, 12),
-               GetOrNull(chars, 13),
-               GetOrNull(chars, 14)} {}
-
-    // Disable sanitizer as we must always be able to read `tag`.
-    ABSL_CORD_INTERNAL_NO_SANITIZE
-    int8_t tag() const { return reinterpret_cast<const int8_t*>(this)[0]; }
-    void set_tag(int8_t rhs) { reinterpret_cast<int8_t*>(this)[0] = rhs; }
-
-    char* as_chars() { return data + 1; }
-    const char* as_chars() const { return data + 1; }
-
-    bool is_tree() const { return (tag() & 1) != 0; }
-
-    size_t inline_size() const {
-      ABSL_ASSERT(!is_tree());
-      return static_cast<size_t>(tag()) >> 1;
-    }
-
-    void set_inline_size(size_t size) {
-      ABSL_ASSERT(size <= kMaxInline);
-      set_tag(static_cast<int8_t>(size << 1));
-    }
-
-    CordRep* tree() const { return as_tree.rep; }
-    void set_tree(CordRep* rhs) { as_tree.rep = rhs; }
-
-    cordz_info_t cordz_info() const { return as_tree.cordz_info; }
-    void set_cordz_info(cordz_info_t rhs) { as_tree.cordz_info = rhs; }
-
-    void make_tree(CordRep* tree) {
-      as_tree.rep = tree;
-      as_tree.cordz_info = kNullCordzInfo;
-    }
-
-#ifdef ABSL_INTERNAL_CORD_HAVE_SANITIZER
-    constexpr Rep SanitizerSafeCopy() const {
-      if (!absl::is_constant_evaluated()) {
-        Rep res;
-        if (is_tree()) {
-          res = *this;
-        } else {
-          res.set_tag(tag());
-          memcpy(res.as_chars(), as_chars(), inline_size());
-        }
-        return res;
-      } else {
-        return *this;
-      }
-    }
-#else
-    constexpr const Rep& SanitizerSafeCopy() const { return *this; }
-#endif
-
-    // If the data has length <= kMaxInline, we store it in `data`, and
-    // store the size in the first char of `data` shifted left + 1.
-    // Else we store it in a tree and store a pointer to that tree in
-    // `as_tree.rep` with a tagged pointer to make `tag() & 1` non zero.
-    union {
-      char data[kMaxInline + 1];
-      AsTree as_tree;
-    };
-
-    // TODO(b/145829486): see swap(InlineData, InlineData) for more info.
-    inline void SwapValue(Rep rhs, Rep& refrhs) {
-      memcpy(&refrhs, this, sizeof(*this));
-      memcpy(this, &rhs, sizeof(*this));
-    }
-  };
-
-  // Private implementation of `Compare()`
-  static inline int Compare(const Rep& lhs, const Rep& rhs) {
-    uint64_t x, y;
-    memcpy(&x, lhs.as_chars(), sizeof(x));
-    memcpy(&y, rhs.as_chars(), sizeof(y));
-    if (x == y) {
-      memcpy(&x, lhs.as_chars() + 7, sizeof(x));
-      memcpy(&y, rhs.as_chars() + 7, sizeof(y));
-      if (x == y) {
-        if (lhs.inline_size() == rhs.inline_size()) return 0;
-        return lhs.inline_size() < rhs.inline_size() ? -1 : 1;
-      }
-    }
-    x = absl::big_endian::FromHost64(x);
-    y = absl::big_endian::FromHost64(y);
-    return x < y ? -1 : 1;
-  }
-
-  Rep rep_;
-};
-
-static_assert(sizeof(InlineData) == kMaxInline + 1, "");
-
-#ifdef ABSL_INTERNAL_CORD_HAVE_SANITIZER
-
-constexpr InlineData::InlineData(const InlineData& rhs) noexcept
-    : rep_(rhs.rep_.SanitizerSafeCopy()) {
-  poison();
-}
-
-inline InlineData& InlineData::operator=(const InlineData& rhs) noexcept {
-  unpoison();
-  rep_ = rhs.rep_.SanitizerSafeCopy();
-  poison();
-  return *this;
-}
-
-constexpr void InlineData::poison_this() {
-  if (!absl::is_constant_evaluated()) {
-    container_internal::SanitizerPoisonObject(this);
-  }
-}
-
-constexpr void InlineData::unpoison() {
-  if (!absl::is_constant_evaluated()) {
-    container_internal::SanitizerUnpoisonObject(this);
-  }
-}
-
-constexpr void InlineData::poison() {
-  if (!absl::is_constant_evaluated()) {
-    if (is_tree()) {
-      container_internal::SanitizerUnpoisonObject(this);
-    } else if (const size_t size = inline_size()) {
-      if (size < kMaxInline) {
-        const char* end = rep_.as_chars() + size;
-        container_internal::SanitizerPoisonMemoryRegion(end, kMaxInline - size);
-      }
-    } else {
-      container_internal::SanitizerPoisonObject(this);
-    }
-  }
-}
-
-#else  // ABSL_INTERNAL_CORD_HAVE_SANITIZER
-
-constexpr InlineData::InlineData(const InlineData&) noexcept = default;
-inline InlineData& InlineData::operator=(const InlineData&) noexcept = default;
-
-constexpr void InlineData::poison_this() {}
-constexpr void InlineData::unpoison() {}
-constexpr void InlineData::poison() {}
-
-#endif  // ABSL_INTERNAL_CORD_HAVE_SANITIZER
-
-inline CordRepSubstring* CordRep::substring() {
-  assert(IsSubstring());
-  return static_cast<CordRepSubstring*>(this);
-}
-
-inline const CordRepSubstring* CordRep::substring() const {
-  assert(IsSubstring());
-  return static_cast<const CordRepSubstring*>(this);
-}
-
-inline CordRepExternal* CordRep::external() {
-  assert(IsExternal());
-  return static_cast<CordRepExternal*>(this);
-}
-
-inline const CordRepExternal* CordRep::external() const {
-  assert(IsExternal());
-  return static_cast<const CordRepExternal*>(this);
-}
-
-inline CordRep* CordRep::Ref(CordRep* rep) {
-  // ABSL_ASSUME is a workaround for
-  // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105585
-  ABSL_ASSUME(rep != nullptr);
-  rep->refcount.Increment();
-  return rep;
-}
-
-inline void CordRep::Unref(CordRep* rep) {
-  assert(rep != nullptr);
-  // Expect refcount to be 0. Avoiding the cost of an atomic decrement should
-  // typically outweigh the cost of an extra branch checking for ref == 1.
-  if (ABSL_PREDICT_FALSE(!rep->refcount.DecrementExpectHighRefcount())) {
-    Destroy(rep);
-  }
-}
-
-inline void swap(InlineData& lhs, InlineData& rhs) noexcept {
-  lhs.unpoison();
-  rhs.unpoison();
-  // TODO(b/145829486): `std::swap(lhs.rep_, rhs.rep_)` results in bad codegen
-  // on clang, spilling the temporary swap value on the stack. Since `Rep` is
-  // trivial, we can make clang DTRT by calling a hand-rolled `SwapValue` where
-  // we pass `rhs` both by value (register allocated) and by reference. The IR
-  // then folds and inlines correctly into an optimized swap without spill.
-  lhs.rep_.SwapValue(rhs.rep_, rhs.rep_);
-  rhs.poison();
-  lhs.poison();
-}
-
-}  // namespace cord_internal
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-#endif  // ABSL_STRINGS_INTERNAL_CORD_INTERNAL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_btree.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_btree.h
deleted file mode 100644
index be94b62e..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_btree.h
+++ /dev/null
@@ -1,944 +0,0 @@
-// Copyright 2021 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CORD_REP_BTREE_H_
-#define ABSL_STRINGS_INTERNAL_CORD_REP_BTREE_H_
-
-#include <cassert>
-#include <cstdint>
-#include <iosfwd>
-
-#include "absl/base/config.h"
-#include "absl/base/internal/raw_logging.h"
-#include "absl/base/optimization.h"
-#include "absl/strings/internal/cord_data_edge.h"
-#include "absl/strings/internal/cord_internal.h"
-#include "absl/strings/internal/cord_rep_flat.h"
-#include "absl/strings/string_view.h"
-#include "absl/types/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// `SetCordBtreeExhaustiveValidation()` can be set to force exhaustive
-// validation in debug assertions, and code that calls `IsValid()`
-// explicitly. By default, assertions should be relatively cheap and
-// AssertValid() can easily lead to O(n^2) complexity as recursive / full tree
-// validation is O(n).
-void SetCordBtreeExhaustiveValidation(bool do_exaustive_validation);
-bool IsCordBtreeExhaustiveValidationEnabled();
-
-class CordRepBtreeNavigator;
-
-// CordRepBtree is as the name implies a btree implementation of a Cordrep tree.
-// Data is stored at the leaf level only, non leaf nodes contain down pointers
-// only. Allowed types of data edges are FLAT, EXTERNAL and SUBSTRINGs of FLAT
-// or EXTERNAL nodes. The implementation allows for data to be added to either
-// end of the tree only, it does not provide any 'insert' logic. This has the
-// benefit that we can expect good fill ratios: all nodes except the outer
-// 'legs' will have 100% fill ratios for trees built using Append/Prepend
-// methods. Merged trees will typically have a fill ratio well above 50% as in a
-// similar fashion, one side of the merged tree will typically have a 100% fill
-// ratio, and the 'open' end will average 50%. All operations are O(log(n)) or
-// better, and the tree never needs balancing.
-//
-// All methods accepting a CordRep* or CordRepBtree* adopt a reference on that
-// input unless explicitly stated otherwise. All functions returning a CordRep*
-// or CordRepBtree* instance transfer a reference back to the caller.
-// Simplified, callers both 'donate' and 'consume' a reference count on each
-// call, simplifying the API. An example of building a tree:
-//
-//   CordRepBtree* tree = CordRepBtree::Create(MakeFlat("Hello"));
-//   tree = CordRepBtree::Append(tree, MakeFlat("world"));
-//
-// In the above example, all inputs are consumed, making each call affecting
-// `tree` reference count neutral. The returned `tree` value can be different
-// from the input if the input is shared with other threads, or if the tree
-// grows in height, but callers typically never have to concern themselves with
-// that and trust that all methods DTRT at all times.
-class CordRepBtree : public CordRep {
- public:
-  // EdgeType identifies `front` and `back` enum values.
-  // Various implementations in CordRepBtree such as `Add` and `Edge` are
-  // generic and templated on operating on either of the boundary edges.
-  // For more information on the possible edges contained in a CordRepBtree
-  // instance see the documentation for `edges_`.
-  enum class EdgeType { kFront, kBack };
-
-  // Convenience constants into `EdgeType`
-  static constexpr EdgeType kFront = EdgeType::kFront;
-  static constexpr EdgeType kBack = EdgeType::kBack;
-
-  // Maximum number of edges: based on experiments and performance data, we can
-  // pick suitable values resulting in optimum cacheline aligned values. The
-  // preferred values are based on 64-bit systems where we aim to align this
-  // class onto 64 bytes, i.e.:  6 = 64 bytes, 14 = 128 bytes, etc.
-  // TODO(b/192061034): experiment with alternative sizes.
-  static constexpr size_t kMaxCapacity = 6;
-
-  // Reasonable maximum height of the btree. We can expect a fill ratio of at
-  // least 50%: trees are always expanded at the front or back. Concatenating
-  // trees will then typically fold at the top most node, where the lower nodes
-  // are at least at capacity on one side of joined inputs. At a lower fill
-  // rate of 4 edges per node, we have capacity for ~16 million leaf nodes.
-  // We will fail / abort if an application ever exceeds this height, which
-  // should be extremely rare (near impossible) and be an indication of an
-  // application error: we do not assume it reasonable for any application to
-  // operate correctly with such monster trees.
-  // Another compelling reason for the number `12` is that any contextual stack
-  // required for navigation or insertion requires 12 words and 12 bytes, which
-  // fits inside 2 cache lines with some room to spare, and is reasonable as a
-  // local stack variable compared to Cord's current near 400 bytes stack use.
-  // The maximum `height` value of a node is then `kMaxDepth - 1` as node height
-  // values start with a value of 0 for leaf nodes.
-  static constexpr size_t kMaxDepth = 12;
-  // See comments on height() for why this is an int and not a size_t.
-  static constexpr int kMaxHeight = static_cast<int>(kMaxDepth - 1);
-
-  // `Action` defines the action for unwinding changes done at the btree's leaf
-  // level that need to be propagated up to the parent node(s). Each operation
-  // on a node has an effect / action defined as follows:
-  // - kSelf
-  //   The operation (add / update, etc) was performed directly on the node as
-  //   the node is private to the current thread (i.e.: not shared directly or
-  //   indirectly through a refcount > 1). Changes can be propagated directly to
-  //   all parent nodes as all parent nodes are also then private to the current
-  //   thread.
-  // - kCopied
-  //   The operation (add / update, etc) was performed on a copy of the original
-  //   node, as the node is (potentially) directly or indirectly shared with
-  //   other threads. Changes need to be propagated into the parent nodes where
-  //   the old down pointer must be unreffed and replaced with this new copy.
-  //   Such changes to parent nodes may themselves require a copy if the parent
-  //   node is also shared. A kCopied action can propagate all the way to the
-  //   top node where we then must unref the `tree` input provided by the
-  //   caller, and return the new copy.
-  // - kPopped
-  //   The operation (typically add) could not be satisfied due to insufficient
-  //   capacity in the targeted node, and a new 'leg' was created that needs to
-  //   be added into the parent node. For example, adding a FLAT inside a leaf
-  //   node that is at capacity will create a new leaf node containing that
-  //   FLAT, that needs to be 'popped' up the btree. Such 'pop' actions can
-  //   cascade up the tree if parent nodes are also at capacity. A 'Popped'
-  //   action propagating all the way to the top of the tree will result in
-  //   the tree becoming one level higher than the current tree through a final
-  //   `CordRepBtree::New(tree, popped)` call, resulting in a new top node
-  //   referencing the old tree and the new (fully popped upwards) 'leg'.
-  enum Action { kSelf, kCopied, kPopped };
-
-  // Result of an operation on a node. See the `Action` enum for details.
-  struct OpResult {
-    CordRepBtree* tree;
-    Action action;
-  };
-
-  // Return value of the CopyPrefix and CopySuffix methods which can
-  // return a node or data edge at any height inside the tree.
-  // A height of 0 defines the lowest (leaf) node, a height of -1 identifies
-  // `edge` as being a plain data node: EXTERNAL / FLAT or SUBSTRING thereof.
-  struct CopyResult {
-    CordRep* edge;
-    int height;
-  };
-
-  // Logical position inside a node:
-  // - index: index of the edge.
-  // - n: size or offset value depending on context.
-  struct Position {
-    size_t index;
-    size_t n;
-  };
-
-  // Creates a btree from the given input. Adopts a ref of `rep`.
-  // If the input `rep` is itself a btree, i.e., `IsBtree()`, then this
-  // function immediately returns `rep->btree()`. If the input is a valid data
-  // edge (see IsDataEdge()), then a new leaf node is returned containing `rep`
-  // as the sole data edge. Else, the input is assumed to be a (legacy) concat
-  // tree, and the input is consumed and transformed into a btree().
-  static CordRepBtree* Create(CordRep* rep);
-
-  // Destroys the provided tree. Should only be called by cord internal API's,
-  // typically after a ref_count.Decrement() on the last reference count.
-  static void Destroy(CordRepBtree* tree);
-
-  // Destruction
-  static void Delete(CordRepBtree* tree) { delete tree; }
-
-  // Use CordRep::Unref() as we overload for absl::Span<CordRep* const>.
-  using CordRep::Unref;
-
-  // Unrefs all edges in `edges` which are assumed to be 'likely one'.
-  static void Unref(absl::Span<CordRep* const> edges);
-
-  // Appends / Prepends an existing CordRep instance to this tree.
-  // The below methods accept three types of input:
-  // 1) `rep` is a data node (See `IsDataNode` for valid data edges).
-  // `rep` is appended or prepended to this tree 'as is'.
-  // 2) `rep` is a BTREE.
-  // `rep` is merged into `tree` respecting the Append/Prepend order.
-  // 3) `rep` is some other (legacy) type.
-  // `rep` is converted in place and added to `tree`
-  // Requires `tree` and `rep` to be not null.
-  static CordRepBtree* Append(CordRepBtree* tree, CordRep* rep);
-  static CordRepBtree* Prepend(CordRepBtree* tree, CordRep* rep);
-
-  // Append/Prepend the data in `data` to this tree.
-  // The `extra` parameter defines how much extra capacity should be allocated
-  // for any additional FLAT being allocated. This is an optimization hint from
-  // the caller. For example, a caller may need to add 2 string_views of data
-  // "abc" and "defghi" which are not consecutive. The caller can in this case
-  // invoke `AddData(tree, "abc", 6)`, and any newly added flat is allocated
-  // where possible with at least 6 bytes of extra capacity beyond `length`.
-  // This helps avoiding data getting fragmented over multiple flats.
-  // There is no limit on the size of `data`. If `data` can not be stored inside
-  // a single flat, then the function will iteratively add flats until all data
-  // has been consumed and appended or prepended to the tree.
-  static CordRepBtree* Append(CordRepBtree* tree, string_view data,
-                              size_t extra = 0);
-  static CordRepBtree* Prepend(CordRepBtree* tree, string_view data,
-                               size_t extra = 0);
-
-  // Returns a new tree, containing `n` bytes of data from this instance
-  // starting at offset `offset`. Where possible, the returned tree shares
-  // (re-uses) data edges and nodes with this instance to minimize the
-  // combined memory footprint of both trees.
-  // Requires `offset + n <= length`. Returns `nullptr` if `n` is zero.
-  CordRep* SubTree(size_t offset, size_t n);
-
-  // Removes `n` trailing bytes from `tree`, and returns the resulting tree
-  // or data edge. Returns `tree` if n is zero, and nullptr if n == length.
-  // This function is logically identical to:
-  //   result = tree->SubTree(0, tree->length - n);
-  //   Unref(tree);
-  //   return result;
-  // However, the actual implementation will as much as possible perform 'in
-  // place' modifications on the tree on all nodes and edges that are mutable.
-  // For example, in a fully privately owned tree with the last edge being a
-  // flat of length 12, RemoveSuffix(1) will simply set the length of that data
-  // edge to 11, and reduce the length of all nodes on the edge path by 1.
-  static CordRep* RemoveSuffix(CordRepBtree* tree, size_t n);
-
-  // Returns the character at the given offset.
-  char GetCharacter(size_t offset) const;
-
-  // Returns true if this node holds a single data edge, and if so, sets
-  // `fragment` to reference the contained data. `fragment` is an optional
-  // output parameter and allowed to be null.
-  bool IsFlat(absl::string_view* fragment) const;
-
-  // Returns true if the data of `n` bytes starting at offset `offset`
-  // is contained in a single data edge, and if so, sets fragment to reference
-  // the contained data. `fragment` is an optional output parameter and allowed
-  // to be null.
-  bool IsFlat(size_t offset, size_t n, absl::string_view* fragment) const;
-
-  // Returns a span (mutable range of bytes) of up to `size` bytes into the
-  // last FLAT data edge inside this tree under the following conditions:
-  // - none of the nodes down into the FLAT node are shared.
-  // - the last data edge in this tree is a non-shared FLAT.
-  // - the referenced FLAT has additional capacity available.
-  // If all these conditions are met, a non-empty span is returned, and the
-  // length of the flat node and involved tree nodes have been increased by
-  // `span.length()`. The caller is responsible for immediately assigning values
-  // to all uninitialized data reference by the returned span.
-  // Requires `this->refcount.IsOne()`: this function forces the caller to do
-  // this fast path check on the top level node, as this is the most commonly
-  // shared node of a cord tree.
-  Span<char> GetAppendBuffer(size_t size);
-
-  // Extracts the right-most data edge from this tree iff:
-  // - the tree and all internal edges to the right-most node are not shared.
-  // - the right-most node is a FLAT node and not shared.
-  // - the right-most node has at least the desired extra capacity.
-  //
-  // Returns {tree, nullptr} if any of the above conditions are not met.
-  // This method effectively removes data from the tree. The intent of this
-  // method is to allow applications appending small string data to use
-  // pre-existing capacity, and add the modified rep back to the tree.
-  //
-  // Simplified such code would look similar to this:
-  //   void MyTreeBuilder::Append(string_view data) {
-  //     ExtractResult result = CordRepBtree::ExtractAppendBuffer(tree_, 1);
-  //     if (CordRep* rep = result.extracted) {
-  //       size_t available = rep->Capacity() - rep->length;
-  //       size_t n = std::min(data.size(), n);
-  //       memcpy(rep->Data(), data.data(), n);
-  //       rep->length += n;
-  //       data.remove_prefix(n);
-  //       if (!result.tree->IsBtree()) {
-  //         tree_ = CordRepBtree::Create(result.tree);
-  //       }
-  //       tree_ = CordRepBtree::Append(tree_, rep);
-  //     }
-  //     ...
-  //     // Remaining edge in `result.tree`.
-  //   }
-  static ExtractResult ExtractAppendBuffer(CordRepBtree* tree,
-                                           size_t extra_capacity = 1);
-
-  // Returns the `height` of the tree. The height of a tree is limited to
-  // kMaxHeight. `height` is implemented as an `int` as in some places we
-  // use negative (-1) values for 'data edges'.
-  int height() const { return static_cast<int>(storage[0]); }
-
-  // Properties: begin, back, end, front/back boundary indexes.
-  size_t begin() const { return static_cast<size_t>(storage[1]); }
-  size_t back() const { return static_cast<size_t>(storage[2]) - 1; }
-  size_t end() const { return static_cast<size_t>(storage[2]); }
-  size_t index(EdgeType edge) const {
-    return edge == kFront ? begin() : back();
-  }
-
-  // Properties: size and capacity.
-  // `capacity` contains the current capacity of this instance, where
-  // `kMaxCapacity` contains the maximum capacity of a btree node.
-  // For now, `capacity` and `kMaxCapacity` return the same value, but this may
-  // change in the future if we see benefit in dynamically sizing 'small' nodes
-  // to 'large' nodes for large data trees.
-  size_t size() const { return end() - begin(); }
-  size_t capacity() const { return kMaxCapacity; }
-
-  // Edge access
-  inline CordRep* Edge(size_t index) const;
-  inline CordRep* Edge(EdgeType edge_type) const;
-  inline absl::Span<CordRep* const> Edges() const;
-  inline absl::Span<CordRep* const> Edges(size_t begin, size_t end) const;
-
-  // Returns reference to the data edge at `index`.
-  // Requires this instance to be a leaf node, and `index` to be valid index.
-  inline absl::string_view Data(size_t index) const;
-
-  // Diagnostics: returns true if `tree` is valid and internally consistent.
-  // If `shallow` is false, then the provided top level node and all child nodes
-  // below it are recursively checked. If `shallow` is true, only the provided
-  // node in `tree` and the cumulative length, type and height of the direct
-  // child nodes of `tree` are checked. The value of `shallow` is ignored if the
-  // internal `cord_btree_exhaustive_validation` diagnostics variable is true,
-  // in which case the performed validations works as if `shallow` were false.
-  // This function is intended for debugging and testing purposes only.
-  static bool IsValid(const CordRepBtree* tree, bool shallow = false);
-
-  // Diagnostics: asserts that the provided tree is valid.
-  // `AssertValid()` performs a shallow validation by default. `shallow` can be
-  // set to false in which case an exhaustive validation is performed. This
-  // function is implemented in terms of calling `IsValid()` and asserting the
-  // return value to be true. See `IsValid()` for more information.
-  // This function is intended for debugging and testing purposes only.
-  static CordRepBtree* AssertValid(CordRepBtree* tree, bool shallow = true);
-  static const CordRepBtree* AssertValid(const CordRepBtree* tree,
-                                         bool shallow = true);
-
-  // Diagnostics: dump the contents of this tree to `stream`.
-  // This function is intended for debugging and testing purposes only.
-  static void Dump(const CordRep* rep, std::ostream& stream);
-  static void Dump(const CordRep* rep, absl::string_view label,
-                   std::ostream& stream);
-  static void Dump(const CordRep* rep, absl::string_view label,
-                   bool include_contents, std::ostream& stream);
-
-  // Adds the edge `edge` to this node if possible. `owned` indicates if the
-  // current node is potentially shared or not with other threads. Returns:
-  // - {kSelf, <this>}
-  //   The edge was directly added to this node.
-  // - {kCopied, <node>}
-  //   The edge was added to a copy of this node.
-  // - {kPopped, New(edge, height())}
-  //   A new leg with the edge was created as this node has no extra capacity.
-  template <EdgeType edge_type>
-  inline OpResult AddEdge(bool owned, CordRep* edge, size_t delta);
-
-  // Replaces the front or back edge with the provided new edge. Returns:
-  // - {kSelf, <this>}
-  //   The edge was directly set in this node. The old edge is unreffed.
-  // - {kCopied, <node>}
-  //   A copy of this node was created with the new edge value.
-  // In both cases, the function adopts a reference on `edge`.
-  template <EdgeType edge_type>
-  OpResult SetEdge(bool owned, CordRep* edge, size_t delta);
-
-  // Creates a new empty node at the specified height.
-  static CordRepBtree* New(int height = 0);
-
-  // Creates a new node containing `rep`, with the height being computed
-  // automatically based on the type of `rep`.
-  static CordRepBtree* New(CordRep* rep);
-
-  // Creates a new node containing both `front` and `back` at height
-  // `front.height() + 1`. Requires `back.height() == front.height()`.
-  static CordRepBtree* New(CordRepBtree* front, CordRepBtree* back);
-
-  // Creates a fully balanced tree from the provided tree by rebuilding a new
-  // tree from all data edges in the input. This function is automatically
-  // invoked internally when the tree exceeds the maximum height.
-  static CordRepBtree* Rebuild(CordRepBtree* tree);
-
- private:
-  CordRepBtree() = default;
-  ~CordRepBtree() = default;
-
-  // Initializes the main properties `tag`, `begin`, `end`, `height`.
-  inline void InitInstance(int height, size_t begin = 0, size_t end = 0);
-
-  // Direct property access begin / end
-  void set_begin(size_t begin) { storage[1] = static_cast<uint8_t>(begin); }
-  void set_end(size_t end) { storage[2] = static_cast<uint8_t>(end); }
-
-  // Decreases the value of `begin` by `n`, and returns the new value. Notice
-  // how this returns the new value unlike atomic::fetch_add which returns the
-  // old value. This is because this is used to prepend edges at 'begin - 1'.
-  size_t sub_fetch_begin(size_t n) {
-    storage[1] -= static_cast<uint8_t>(n);
-    return storage[1];
-  }
-
-  // Increases the value of `end` by `n`, and returns the previous value. This
-  // function is typically used to append edges at 'end'.
-  size_t fetch_add_end(size_t n) {
-    const uint8_t current = storage[2];
-    storage[2] = static_cast<uint8_t>(current + n);
-    return current;
-  }
-
-  // Returns the index of the last edge starting on, or before `offset`, with
-  // `n` containing the relative offset of `offset` inside that edge.
-  // Requires `offset` < length.
-  Position IndexOf(size_t offset) const;
-
-  // Returns the index of the last edge starting before `offset`, with `n`
-  // containing the relative offset of `offset` inside that edge.
-  // This function is useful to find the edges for some span of bytes ending at
-  // `offset` (i.e., `n` bytes). For example:
-  //
-  //   Position pos = IndexBefore(n)
-  //   edges = Edges(begin(), pos.index)     // All full edges (may be empty)
-  //   last = Sub(Edge(pos.index), 0, pos.n) // Last partial edge (may be empty)
-  //
-  // Requires 0 < `offset` <= length.
-  Position IndexBefore(size_t offset) const;
-
-  // Returns the index of the edge ending at (or on) length `length`, and the
-  // number of bytes inside that edge up to `length`. For example, if we have a
-  // Node with 2 edges, one of 10 and one of 20 long, then IndexOfLength(27)
-  // will return {1, 17}, and IndexOfLength(10) will return {0, 10}.
-  Position IndexOfLength(size_t n) const;
-
-  // Identical to the above function except starting from the position `front`.
-  // This function is equivalent to `IndexBefore(front.n + offset)`, with
-  // the difference that this function is optimized to start at `front.index`.
-  Position IndexBefore(Position front, size_t offset) const;
-
-  // Returns the index of the edge directly beyond the edge containing offset
-  // `offset`, with `n` containing the distance of that edge from `offset`.
-  // This function is useful for iteratively finding suffix nodes and remaining
-  // partial bytes in left-most suffix nodes as for example in CopySuffix.
-  // Requires `offset` < length.
-  Position IndexBeyond(size_t offset) const;
-
-  // Creates a new leaf node containing as much data as possible from `data`.
-  // The data is added either forwards or reversed depending on `edge_type`.
-  // Callers must check the length of the returned node to determine if all data
-  // was copied or not.
-  // See the `Append/Prepend` function for the meaning and purpose of `extra`.
-  template <EdgeType edge_type>
-  static CordRepBtree* NewLeaf(absl::string_view data, size_t extra);
-
-  // Creates a raw copy of this Btree node with the specified length, copying
-  // all properties, but without adding any references to existing edges.
-  CordRepBtree* CopyRaw(size_t new_length) const;
-
-  // Creates a full copy of this Btree node, adding a reference on all edges.
-  CordRepBtree* Copy() const;
-
-  // Creates a partial copy of this Btree node, copying all edges up to `end`,
-  // adding a reference on each copied edge, and sets the length of the newly
-  // created copy to `new_length`.
-  CordRepBtree* CopyBeginTo(size_t end, size_t new_length) const;
-
-  // Returns a tree containing the edges [tree->begin(), end) and length
-  // of `new_length`. This method consumes a reference on the provided
-  // tree, and logically performs the following operation:
-  //   result = tree->CopyBeginTo(end, new_length);
-  //   CordRep::Unref(tree);
-  //   return result;
-  static CordRepBtree* ConsumeBeginTo(CordRepBtree* tree, size_t end,
-                                      size_t new_length);
-
-  // Creates a partial copy of this Btree node, copying all edges starting at
-  // `begin`, adding a reference on each copied edge, and sets the length of
-  // the newly created copy to `new_length`.
-  CordRepBtree* CopyToEndFrom(size_t begin, size_t new_length) const;
-
-  // Extracts and returns the front edge from the provided tree.
-  // This method consumes a reference on the provided tree, and logically
-  // performs the following operation:
-  //   edge = CordRep::Ref(tree->Edge(kFront));
-  //   CordRep::Unref(tree);
-  //   return edge;
-  static CordRep* ExtractFront(CordRepBtree* tree);
-
-  // Returns a tree containing the result of appending `right` to `left`.
-  static CordRepBtree* MergeTrees(CordRepBtree* left, CordRepBtree* right);
-
-  // Fallback functions for `Create()`, `Append()` and `Prepend()` which
-  // deal with legacy / non conforming input, i.e.: CONCAT trees.
-  static CordRepBtree* CreateSlow(CordRep* rep);
-  static CordRepBtree* AppendSlow(CordRepBtree*, CordRep* rep);
-  static CordRepBtree* PrependSlow(CordRepBtree*, CordRep* rep);
-
-  // Recursively rebuilds `tree` into `stack`. If 'consume` is set to true, the
-  // function will consume a reference on `tree`. `stack` is a null terminated
-  // array containing the new tree's state, with the current leaf node at
-  // stack[0], and parent nodes above that, or null for 'top of tree'.
-  static void Rebuild(CordRepBtree** stack, CordRepBtree* tree, bool consume);
-
-  // Aligns existing edges to start at index 0, to allow for a new edge to be
-  // added to the back of the current edges.
-  inline void AlignBegin();
-
-  // Aligns existing edges to end at `capacity`, to allow for a new edge to be
-  // added in front of the current edges.
-  inline void AlignEnd();
-
-  // Adds the provided edge to this node.
-  // Requires this node to have capacity for the edge. Realigns / moves
-  // existing edges as needed to prepend or append the new edge.
-  template <EdgeType edge_type>
-  inline void Add(CordRep* rep);
-
-  // Adds the provided edges to this node.
-  // Requires this node to have capacity for the edges. Realigns / moves
-  // existing edges as needed to prepend or append the new edges.
-  template <EdgeType edge_type>
-  inline void Add(absl::Span<CordRep* const>);
-
-  // Adds data from `data` to this node until either all data has been consumed,
-  // or there is no more capacity for additional flat nodes inside this node.
-  // Requires the current node to be a leaf node, data to be non empty, and the
-  // current node to have capacity for at least one more data edge.
-  // Returns any remaining data from `data` that was not added, which is
-  // depending on the edge type (front / back) either the remaining prefix of
-  // suffix of the input.
-  // See the `Append/Prepend` function for the meaning and purpose of `extra`.
-  template <EdgeType edge_type>
-  absl::string_view AddData(absl::string_view data, size_t extra);
-
-  // Replace the front or back edge with the provided value.
-  // Adopts a reference on `edge` and unrefs the old edge.
-  template <EdgeType edge_type>
-  inline void SetEdge(CordRep* edge);
-
-  // Returns a partial copy of the current tree containing the first `n` bytes
-  // of data. `CopyResult` contains both the resulting edge and its height. The
-  // resulting tree may be less high than the current tree, or even be a single
-  // matching data edge if `allow_folding` is set to true.
-  // For example, if `n == 1`, then the result will be the single data edge, and
-  // height will be set to -1 (one below the owning leaf node). If n == 0, this
-  // function returns null. Requires `n <= length`
-  CopyResult CopyPrefix(size_t n, bool allow_folding = true);
-
-  // Returns a partial copy of the current tree containing all data starting
-  // after `offset`. `CopyResult` contains both the resulting edge and its
-  // height. The resulting tree may be less high than the current tree, or even
-  // be a single matching data edge. For example, if `n == length - 1`, then the
-  // result will be a single data edge, and height will be set to -1 (one below
-  // the owning leaf node).
-  // Requires `offset < length`
-  CopyResult CopySuffix(size_t offset);
-
-  // Returns a OpResult value of {this, kSelf} or {Copy(), kCopied}
-  // depending on the value of `owned`.
-  inline OpResult ToOpResult(bool owned);
-
-  // Adds `rep` to the specified tree, returning the modified tree.
-  template <EdgeType edge_type>
-  static CordRepBtree* AddCordRep(CordRepBtree* tree, CordRep* rep);
-
-  // Adds `data` to the specified tree, returning the modified tree.
-  // See the `Append/Prepend` function for the meaning and purpose of `extra`.
-  template <EdgeType edge_type>
-  static CordRepBtree* AddData(CordRepBtree* tree, absl::string_view data,
-                               size_t extra = 0);
-
-  // Merges `src` into `dst` with `src` being added either before (kFront) or
-  // after (kBack) `dst`. Requires the height of `dst` to be greater than or
-  // equal to the height of `src`.
-  template <EdgeType edge_type>
-  static CordRepBtree* Merge(CordRepBtree* dst, CordRepBtree* src);
-
-  // Fallback version of GetAppendBuffer for large trees: GetAppendBuffer()
-  // implements an inlined version for trees of limited height (3 levels),
-  // GetAppendBufferSlow implements the logic for large trees.
-  Span<char> GetAppendBufferSlow(size_t size);
-
-  // `edges_` contains all edges starting from this instance.
-  // These are explicitly `child` edges only, a cord btree (or any cord tree in
-  // that respect) does not store `parent` pointers anywhere: multiple trees /
-  // parents can reference the same shared child edge. The type of these edges
-  // depends on the height of the node. `Leaf nodes` (height == 0) contain `data
-  // edges` (external or flat nodes, or sub-strings thereof). All other nodes
-  // (height > 0) contain pointers to BTREE nodes with a height of `height - 1`.
-  CordRep* edges_[kMaxCapacity];
-
-  friend class CordRepBtreeTestPeer;
-  friend class CordRepBtreeNavigator;
-};
-
-inline CordRepBtree* CordRep::btree() {
-  assert(IsBtree());
-  return static_cast<CordRepBtree*>(this);
-}
-
-inline const CordRepBtree* CordRep::btree() const {
-  assert(IsBtree());
-  return static_cast<const CordRepBtree*>(this);
-}
-
-inline void CordRepBtree::InitInstance(int height, size_t begin, size_t end) {
-  tag = BTREE;
-  storage[0] = static_cast<uint8_t>(height);
-  storage[1] = static_cast<uint8_t>(begin);
-  storage[2] = static_cast<uint8_t>(end);
-}
-
-inline CordRep* CordRepBtree::Edge(size_t index) const {
-  assert(index >= begin());
-  assert(index < end());
-  return edges_[index];
-}
-
-inline CordRep* CordRepBtree::Edge(EdgeType edge_type) const {
-  return edges_[edge_type == kFront ? begin() : back()];
-}
-
-inline absl::Span<CordRep* const> CordRepBtree::Edges() const {
-  return {edges_ + begin(), size()};
-}
-
-inline absl::Span<CordRep* const> CordRepBtree::Edges(size_t begin,
-                                                      size_t end) const {
-  assert(begin <= end);
-  assert(begin >= this->begin());
-  assert(end <= this->end());
-  return {edges_ + begin, static_cast<size_t>(end - begin)};
-}
-
-inline absl::string_view CordRepBtree::Data(size_t index) const {
-  assert(height() == 0);
-  return EdgeData(Edge(index));
-}
-
-inline CordRepBtree* CordRepBtree::New(int height) {
-  CordRepBtree* tree = new CordRepBtree;
-  tree->length = 0;
-  tree->InitInstance(height);
-  return tree;
-}
-
-inline CordRepBtree* CordRepBtree::New(CordRep* rep) {
-  CordRepBtree* tree = new CordRepBtree;
-  int height = rep->IsBtree() ? rep->btree()->height() + 1 : 0;
-  tree->length = rep->length;
-  tree->InitInstance(height, /*begin=*/0, /*end=*/1);
-  tree->edges_[0] = rep;
-  return tree;
-}
-
-inline CordRepBtree* CordRepBtree::New(CordRepBtree* front,
-                                       CordRepBtree* back) {
-  assert(front->height() == back->height());
-  CordRepBtree* tree = new CordRepBtree;
-  tree->length = front->length + back->length;
-  tree->InitInstance(front->height() + 1, /*begin=*/0, /*end=*/2);
-  tree->edges_[0] = front;
-  tree->edges_[1] = back;
-  return tree;
-}
-
-inline void CordRepBtree::Unref(absl::Span<CordRep* const> edges) {
-  for (CordRep* edge : edges) {
-    if (ABSL_PREDICT_FALSE(!edge->refcount.Decrement())) {
-      CordRep::Destroy(edge);
-    }
-  }
-}
-
-inline CordRepBtree* CordRepBtree::CopyRaw(size_t new_length) const {
-  CordRepBtree* tree = new CordRepBtree;
-
-  // `length` and `refcount` are the first members of `CordRepBtree`.
-  // We initialize `length` using the given length, have `refcount` be set to
-  // ref = 1 through its default constructor, and copy all data beyond
-  // 'refcount' which starts with `tag` using a single memcpy: all contents
-  // except `refcount` is trivially copyable, and the compiler does not
-  // efficiently coalesce member-wise copy of these members.
-  // See https://gcc.godbolt.org/z/qY8zsca6z
-  // # LINT.IfChange(copy_raw)
-  tree->length = new_length;
-  uint8_t* dst = &tree->tag;
-  const uint8_t* src = &tag;
-  const ptrdiff_t offset = src - reinterpret_cast<const uint8_t*>(this);
-  memcpy(dst, src, sizeof(CordRepBtree) - static_cast<size_t>(offset));
-  return tree;
-  // # LINT.ThenChange()
-}
-
-inline CordRepBtree* CordRepBtree::Copy() const {
-  CordRepBtree* tree = CopyRaw(length);
-  for (CordRep* rep : Edges()) CordRep::Ref(rep);
-  return tree;
-}
-
-inline CordRepBtree* CordRepBtree::CopyToEndFrom(size_t begin,
-                                                 size_t new_length) const {
-  assert(begin >= this->begin());
-  assert(begin <= this->end());
-  CordRepBtree* tree = CopyRaw(new_length);
-  tree->set_begin(begin);
-  for (CordRep* edge : tree->Edges()) CordRep::Ref(edge);
-  return tree;
-}
-
-inline CordRepBtree* CordRepBtree::CopyBeginTo(size_t end,
-                                               size_t new_length) const {
-  assert(end <= capacity());
-  assert(end >= this->begin());
-  CordRepBtree* tree = CopyRaw(new_length);
-  tree->set_end(end);
-  for (CordRep* edge : tree->Edges()) CordRep::Ref(edge);
-  return tree;
-}
-
-inline void CordRepBtree::AlignBegin() {
-  // The below code itself does not need to be fast as typically we have
-  // mono-directional append/prepend calls, and `begin` / `end` are typically
-  // adjusted no more than once. But we want to avoid potential register clobber
-  // effects, making the compiler emit register save/store/spills, and minimize
-  // the size of code.
-  const size_t delta = begin();
-  if (ABSL_PREDICT_FALSE(delta != 0)) {
-    const size_t new_end = end() - delta;
-    set_begin(0);
-    set_end(new_end);
-    // TODO(mvels): we can write this using 2 loads / 2 stores depending on
-    // total size for the kMaxCapacity = 6 case. I.e., we can branch (switch) on
-    // size, and then do overlapping load/store of up to 4 pointers (inlined as
-    // XMM, YMM or ZMM load/store) and up to 2 pointers (XMM / YMM), which is a)
-    // compact and b) not clobbering any registers.
-    ABSL_ASSUME(new_end <= kMaxCapacity);
-#ifdef __clang__
-#pragma unroll 1
-#endif
-    for (size_t i = 0; i < new_end; ++i) {
-      edges_[i] = edges_[i + delta];
-    }
-  }
-}
-
-inline void CordRepBtree::AlignEnd() {
-  // See comments in `AlignBegin` for motivation on the hand-rolled for loops.
-  const size_t delta = capacity() - end();
-  if (delta != 0) {
-    const size_t new_begin = begin() + delta;
-    const size_t new_end = end() + delta;
-    set_begin(new_begin);
-    set_end(new_end);
-    ABSL_ASSUME(new_end <= kMaxCapacity);
-#ifdef __clang__
-#pragma unroll 1
-#endif
-    for (size_t i = new_end - 1; i >= new_begin; --i) {
-      edges_[i] = edges_[i - delta];
-    }
-  }
-}
-
-template <>
-inline void CordRepBtree::Add<CordRepBtree::kBack>(CordRep* rep) {
-  AlignBegin();
-  edges_[fetch_add_end(1)] = rep;
-}
-
-template <>
-inline void CordRepBtree::Add<CordRepBtree::kBack>(
-    absl::Span<CordRep* const> edges) {
-  AlignBegin();
-  size_t new_end = end();
-  for (CordRep* edge : edges) edges_[new_end++] = edge;
-  set_end(new_end);
-}
-
-template <>
-inline void CordRepBtree::Add<CordRepBtree::kFront>(CordRep* rep) {
-  AlignEnd();
-  edges_[sub_fetch_begin(1)] = rep;
-}
-
-template <>
-inline void CordRepBtree::Add<CordRepBtree::kFront>(
-    absl::Span<CordRep* const> edges) {
-  AlignEnd();
-  size_t new_begin = begin() - edges.size();
-  set_begin(new_begin);
-  for (CordRep* edge : edges) edges_[new_begin++] = edge;
-}
-
-template <CordRepBtree::EdgeType edge_type>
-inline void CordRepBtree::SetEdge(CordRep* edge) {
-  const int idx = edge_type == kFront ? begin() : back();
-  CordRep::Unref(edges_[idx]);
-  edges_[idx] = edge;
-}
-
-inline CordRepBtree::OpResult CordRepBtree::ToOpResult(bool owned) {
-  return owned ? OpResult{this, kSelf} : OpResult{Copy(), kCopied};
-}
-
-inline CordRepBtree::Position CordRepBtree::IndexOf(size_t offset) const {
-  assert(offset < length);
-  size_t index = begin();
-  while (offset >= edges_[index]->length) offset -= edges_[index++]->length;
-  return {index, offset};
-}
-
-inline CordRepBtree::Position CordRepBtree::IndexBefore(size_t offset) const {
-  assert(offset > 0);
-  assert(offset <= length);
-  size_t index = begin();
-  while (offset > edges_[index]->length) offset -= edges_[index++]->length;
-  return {index, offset};
-}
-
-inline CordRepBtree::Position CordRepBtree::IndexBefore(Position front,
-                                                        size_t offset) const {
-  size_t index = front.index;
-  offset = offset + front.n;
-  while (offset > edges_[index]->length) offset -= edges_[index++]->length;
-  return {index, offset};
-}
-
-inline CordRepBtree::Position CordRepBtree::IndexOfLength(size_t n) const {
-  assert(n <= length);
-  size_t index = back();
-  size_t strip = length - n;
-  while (strip >= edges_[index]->length) strip -= edges_[index--]->length;
-  return {index, edges_[index]->length - strip};
-}
-
-inline CordRepBtree::Position CordRepBtree::IndexBeyond(
-    const size_t offset) const {
-  // We need to find the edge which `starting offset` is beyond (>=)`offset`.
-  // For this we can't use the `offset -= length` logic of IndexOf. Instead, we
-  // track the offset of the `current edge` in `off`, which we increase as we
-  // iterate over the edges until we find the matching edge.
-  size_t off = 0;
-  size_t index = begin();
-  while (offset > off) off += edges_[index++]->length;
-  return {index, off - offset};
-}
-
-inline CordRepBtree* CordRepBtree::Create(CordRep* rep) {
-  if (IsDataEdge(rep)) return New(rep);
-  return CreateSlow(rep);
-}
-
-inline Span<char> CordRepBtree::GetAppendBuffer(size_t size) {
-  assert(refcount.IsOne());
-  CordRepBtree* tree = this;
-  const int height = this->height();
-  CordRepBtree* n1 = tree;
-  CordRepBtree* n2 = tree;
-  CordRepBtree* n3 = tree;
-  switch (height) {
-    case 3:
-      tree = tree->Edge(kBack)->btree();
-      if (!tree->refcount.IsOne()) return {};
-      n2 = tree;
-      ABSL_FALLTHROUGH_INTENDED;
-    case 2:
-      tree = tree->Edge(kBack)->btree();
-      if (!tree->refcount.IsOne()) return {};
-      n1 = tree;
-      ABSL_FALLTHROUGH_INTENDED;
-    case 1:
-      tree = tree->Edge(kBack)->btree();
-      if (!tree->refcount.IsOne()) return {};
-      ABSL_FALLTHROUGH_INTENDED;
-    case 0:
-      CordRep* edge = tree->Edge(kBack);
-      if (!edge->refcount.IsOne()) return {};
-      if (edge->tag < FLAT) return {};
-      size_t avail = edge->flat()->Capacity() - edge->length;
-      if (avail == 0) return {};
-      size_t delta = (std::min)(size, avail);
-      Span<char> span = {edge->flat()->Data() + edge->length, delta};
-      edge->length += delta;
-      switch (height) {
-        case 3:
-          n3->length += delta;
-          ABSL_FALLTHROUGH_INTENDED;
-        case 2:
-          n2->length += delta;
-          ABSL_FALLTHROUGH_INTENDED;
-        case 1:
-          n1->length += delta;
-          ABSL_FALLTHROUGH_INTENDED;
-        case 0:
-          tree->length += delta;
-          return span;
-      }
-      break;
-  }
-  return GetAppendBufferSlow(size);
-}
-
-extern template CordRepBtree* CordRepBtree::AddCordRep<CordRepBtree::kBack>(
-    CordRepBtree* tree, CordRep* rep);
-
-extern template CordRepBtree* CordRepBtree::AddCordRep<CordRepBtree::kFront>(
-    CordRepBtree* tree, CordRep* rep);
-
-inline CordRepBtree* CordRepBtree::Append(CordRepBtree* tree, CordRep* rep) {
-  if (ABSL_PREDICT_TRUE(IsDataEdge(rep))) {
-    return CordRepBtree::AddCordRep<kBack>(tree, rep);
-  }
-  return AppendSlow(tree, rep);
-}
-
-inline CordRepBtree* CordRepBtree::Prepend(CordRepBtree* tree, CordRep* rep) {
-  if (ABSL_PREDICT_TRUE(IsDataEdge(rep))) {
-    return CordRepBtree::AddCordRep<kFront>(tree, rep);
-  }
-  return PrependSlow(tree, rep);
-}
-
-#ifdef NDEBUG
-
-inline CordRepBtree* CordRepBtree::AssertValid(CordRepBtree* tree,
-                                               bool /* shallow */) {
-  return tree;
-}
-
-inline const CordRepBtree* CordRepBtree::AssertValid(const CordRepBtree* tree,
-                                                     bool /* shallow */) {
-  return tree;
-}
-
-#endif
-
-}  // namespace cord_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CORD_REP_BTREE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_btree_navigator.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_btree_navigator.h
deleted file mode 100644
index 3d581c87..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_btree_navigator.h
+++ /dev/null
@@ -1,267 +0,0 @@
-// Copyright 2021 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CORD_REP_BTREE_NAVIGATOR_H_
-#define ABSL_STRINGS_INTERNAL_CORD_REP_BTREE_NAVIGATOR_H_
-
-#include <cassert>
-#include <iostream>
-
-#include "absl/strings/internal/cord_internal.h"
-#include "absl/strings/internal/cord_rep_btree.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// CordRepBtreeNavigator is a bi-directional navigator allowing callers to
-// navigate all the (leaf) data edges in a CordRepBtree instance.
-//
-// A CordRepBtreeNavigator instance is by default empty. Callers initialize a
-// navigator instance by calling one of `InitFirst()`, `InitLast()` or
-// `InitOffset()`, which establishes a current position. Callers can then
-// navigate using the `Next`, `Previous`, `Skip` and `Seek` methods.
-//
-// The navigator instance does not take or adopt a reference on the provided
-// `tree` on any of the initialization calls. Callers are responsible for
-// guaranteeing the lifecycle of the provided tree. A navigator instance can
-// be reset to the empty state by calling `Reset`.
-//
-// A navigator only keeps positional state on the 'current data edge', it does
-// explicitly not keep any 'offset' state. The class does accept and return
-// offsets in the `Read()`, `Skip()` and 'Seek()` methods as these would
-// otherwise put a big burden on callers. Callers are expected to maintain
-// (returned) offset info if they require such granular state.
-class CordRepBtreeNavigator {
- public:
-  // The logical position as returned by the Seek() and Skip() functions.
-  // Returns the current leaf edge for the desired seek or skip position and
-  // the offset of that position inside that edge.
-  struct Position {
-    CordRep* edge;
-    size_t offset;
-  };
-
-  // The read result as returned by the Read() function.
-  // `tree` contains the resulting tree which is identical to the result
-  // of calling CordRepBtree::SubTree(...) on the tree being navigated.
-  // `n` contains the number of bytes used from the last navigated to
-  // edge of the tree.
-  struct ReadResult {
-    CordRep* tree;
-    size_t n;
-  };
-
-  // Returns true if this instance is not empty.
-  explicit operator bool() const;
-
-  // Returns the tree for this instance or nullptr if empty.
-  CordRepBtree* btree() const;
-
-  // Returns the data edge of the current position.
-  // Requires this instance to not be empty.
-  CordRep* Current() const;
-
-  // Resets this navigator to `tree`, returning the first data edge in the tree.
-  CordRep* InitFirst(CordRepBtree* tree);
-
-  // Resets this navigator to `tree`, returning the last data edge in the tree.
-  CordRep* InitLast(CordRepBtree* tree);
-
-  // Resets this navigator to `tree` returning the data edge at position
-  // `offset` and the relative offset of `offset` into that data edge.
-  // Returns `Position.edge = nullptr` if the provided offset is greater
-  // than or equal to the length of the tree, in which case the state of
-  // the navigator instance remains unchanged.
-  Position InitOffset(CordRepBtree* tree, size_t offset);
-
-  // Navigates to the next data edge.
-  // Returns the next data edge or nullptr if there is no next data edge, in
-  // which case the current position remains unchanged.
-  CordRep* Next();
-
-  // Navigates to the previous data edge.
-  // Returns the previous data edge or nullptr if there is no previous data
-  // edge, in which case the current position remains unchanged.
-  CordRep* Previous();
-
-  // Navigates to the data edge at position `offset`. Returns the navigated to
-  // data edge in `Position.edge` and the relative offset of `offset` into that
-  // data edge in `Position.offset`. Returns `Position.edge = nullptr` if the
-  // provide offset is greater than or equal to the tree's length.
-  Position Seek(size_t offset);
-
-  // Reads `n` bytes of data starting at offset `edge_offset` of the current
-  // data edge, and returns the result in `ReadResult.tree`. `ReadResult.n`
-  // contains the 'bytes used` from the last / current data edge in the tree.
-  // This allows users that mix regular navigation (using string views) and
-  // 'read into cord' navigation to keep track of the current state, and which
-  // bytes have been consumed from a navigator.
-  // This function returns `ReadResult.tree = nullptr` if the requested length
-  // exceeds the length of the tree starting at the current data edge.
-  ReadResult Read(size_t edge_offset, size_t n);
-
-  // Skips `n` bytes forward from the current data edge, returning the navigated
-  // to data edge in `Position.edge` and `Position.offset` containing the offset
-  // inside that data edge. Note that the state of the navigator is left
-  // unchanged if `n` is smaller than the length of the current data edge.
-  Position Skip(size_t n);
-
-  // Resets this instance to the default / empty state.
-  void Reset();
-
- private:
-  // Slow path for Next() if Next() reached the end of a leaf node. Backtracks
-  // up the stack until it finds a node that has a 'next' position available,
-  // and then does a 'front dive' towards the next leaf node.
-  CordRep* NextUp();
-
-  // Slow path for Previous() if Previous() reached the beginning of a leaf
-  // node. Backtracks up the stack until it finds a node that has a 'previous'
-  // position available, and then does a 'back dive' towards the previous leaf
-  // node.
-  CordRep* PreviousUp();
-
-  // Generic implementation of InitFirst() and InitLast().
-  template <CordRepBtree::EdgeType edge_type>
-  CordRep* Init(CordRepBtree* tree);
-
-  // `height_` contains the height of the current tree, or -1 if empty.
-  int height_ = -1;
-
-  // `index_` and `node_` contain the navigation state as the 'path' to the
-  // current data edge which is at `node_[0]->Edge(index_[0])`. The contents
-  // of these are undefined until the instance is initialized (`height_ >= 0`).
-  uint8_t index_[CordRepBtree::kMaxDepth];
-  CordRepBtree* node_[CordRepBtree::kMaxDepth];
-};
-
-// Returns true if this instance is not empty.
-inline CordRepBtreeNavigator::operator bool() const { return height_ >= 0; }
-
-inline CordRepBtree* CordRepBtreeNavigator::btree() const {
-  return height_ >= 0 ? node_[height_] : nullptr;
-}
-
-inline CordRep* CordRepBtreeNavigator::Current() const {
-  assert(height_ >= 0);
-  return node_[0]->Edge(index_[0]);
-}
-
-inline void CordRepBtreeNavigator::Reset() { height_ = -1; }
-
-inline CordRep* CordRepBtreeNavigator::InitFirst(CordRepBtree* tree) {
-  return Init<CordRepBtree::kFront>(tree);
-}
-
-inline CordRep* CordRepBtreeNavigator::InitLast(CordRepBtree* tree) {
-  return Init<CordRepBtree::kBack>(tree);
-}
-
-template <CordRepBtree::EdgeType edge_type>
-inline CordRep* CordRepBtreeNavigator::Init(CordRepBtree* tree) {
-  assert(tree != nullptr);
-  assert(tree->size() > 0);
-  assert(tree->height() <= CordRepBtree::kMaxHeight);
-  int height = height_ = tree->height();
-  size_t index = tree->index(edge_type);
-  node_[height] = tree;
-  index_[height] = static_cast<uint8_t>(index);
-  while (--height >= 0) {
-    tree = tree->Edge(index)->btree();
-    node_[height] = tree;
-    index = tree->index(edge_type);
-    index_[height] = static_cast<uint8_t>(index);
-  }
-  return node_[0]->Edge(index);
-}
-
-inline CordRepBtreeNavigator::Position CordRepBtreeNavigator::Seek(
-    size_t offset) {
-  assert(btree() != nullptr);
-  int height = height_;
-  CordRepBtree* edge = node_[height];
-  if (ABSL_PREDICT_FALSE(offset >= edge->length)) return {nullptr, 0};
-  CordRepBtree::Position index = edge->IndexOf(offset);
-  index_[height] = static_cast<uint8_t>(index.index);
-  while (--height >= 0) {
-    edge = edge->Edge(index.index)->btree();
-    node_[height] = edge;
-    index = edge->IndexOf(index.n);
-    index_[height] = static_cast<uint8_t>(index.index);
-  }
-  return {edge->Edge(index.index), index.n};
-}
-
-inline CordRepBtreeNavigator::Position CordRepBtreeNavigator::InitOffset(
-    CordRepBtree* tree, size_t offset) {
-  assert(tree != nullptr);
-  assert(tree->height() <= CordRepBtree::kMaxHeight);
-  if (ABSL_PREDICT_FALSE(offset >= tree->length)) return {nullptr, 0};
-  height_ = tree->height();
-  node_[height_] = tree;
-  return Seek(offset);
-}
-
-inline CordRep* CordRepBtreeNavigator::Next() {
-  CordRepBtree* edge = node_[0];
-  return index_[0] == edge->back() ? NextUp() : edge->Edge(++index_[0]);
-}
-
-inline CordRep* CordRepBtreeNavigator::Previous() {
-  CordRepBtree* edge = node_[0];
-  return index_[0] == edge->begin() ? PreviousUp() : edge->Edge(--index_[0]);
-}
-
-inline CordRep* CordRepBtreeNavigator::NextUp() {
-  assert(index_[0] == node_[0]->back());
-  CordRepBtree* edge;
-  size_t index;
-  int height = 0;
-  do {
-    if (++height > height_) return nullptr;
-    edge = node_[height];
-    index = index_[height] + 1;
-  } while (index == edge->end());
-  index_[height] = static_cast<uint8_t>(index);
-  do {
-    node_[--height] = edge = edge->Edge(index)->btree();
-    index_[height] = static_cast<uint8_t>(index = edge->begin());
-  } while (height > 0);
-  return edge->Edge(index);
-}
-
-inline CordRep* CordRepBtreeNavigator::PreviousUp() {
-  assert(index_[0] == node_[0]->begin());
-  CordRepBtree* edge;
-  size_t index;
-  int height = 0;
-  do {
-    if (++height > height_) return nullptr;
-    edge = node_[height];
-    index = index_[height];
-  } while (index == edge->begin());
-  index_[height] = static_cast<uint8_t>(--index);
-  do {
-    node_[--height] = edge = edge->Edge(index)->btree();
-    index_[height] = static_cast<uint8_t>(index = edge->back());
-  } while (height > 0);
-  return edge->Edge(index);
-}
-
-}  // namespace cord_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CORD_REP_BTREE_NAVIGATOR_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_btree_reader.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_btree_reader.h
deleted file mode 100644
index 8db8f8dd..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_btree_reader.h
+++ /dev/null
@@ -1,212 +0,0 @@
-// Copyright 2021 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CORD_REP_BTREE_READER_H_
-#define ABSL_STRINGS_INTERNAL_CORD_REP_BTREE_READER_H_
-
-#include <cassert>
-
-#include "absl/base/config.h"
-#include "absl/strings/internal/cord_data_edge.h"
-#include "absl/strings/internal/cord_internal.h"
-#include "absl/strings/internal/cord_rep_btree.h"
-#include "absl/strings/internal/cord_rep_btree_navigator.h"
-#include "absl/strings/internal/cord_rep_flat.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// CordRepBtreeReader implements logic to iterate over cord btrees.
-// References to the underlying data are returned as absl::string_view values.
-// The most typical use case is a forward only iteration over tree data.
-// The class also provides `Skip()`, `Seek()` and `Read()` methods similar to
-// CordRepBtreeNavigator that allow more advanced navigation.
-//
-// Example: iterate over all data inside a cord btree:
-//
-//   CordRepBtreeReader reader;
-//   for (string_view sv = reader.Init(tree); !sv.Empty(); sv = sv.Next()) {
-//     DoSomethingWithDataIn(sv);
-//   }
-//
-// All navigation methods always return the next 'chunk' of data. The class
-// assumes that all data is directly 'consumed' by the caller. For example:
-// invoking `Skip()` will skip the desired number of bytes, and directly
-// read and return the next chunk of data directly after the skipped bytes.
-//
-// Example: iterate over all data inside a btree skipping the first 100 bytes:
-//
-//   CordRepBtreeReader reader;
-//   absl::string_view sv = reader.Init(tree);
-//   if (sv.length() > 100) {
-//     sv.RemovePrefix(100);
-//   } else {
-//     sv = reader.Skip(100 - sv.length());
-//   }
-//   while (!sv.empty()) {
-//     DoSomethingWithDataIn(sv);
-//     absl::string_view sv = reader.Next();
-//   }
-//
-// It is important to notice that `remaining` is based on the end position of
-// the last data edge returned to the caller, not the cumulative data returned
-// to the caller which can be less in cases of skipping or seeking over data.
-//
-// For example, consider a cord btree with five data edges: "abc", "def", "ghi",
-// "jkl" and "mno":
-//
-//   absl::string_view sv;
-//   CordRepBtreeReader reader;
-//
-//   sv = reader.Init(tree); // sv = "abc", remaining = 12
-//   sv = reader.Skip(4);    // sv = "hi",  remaining = 6
-//   sv = reader.Skip(2);    // sv = "l",   remaining = 3
-//   sv = reader.Next();     // sv = "mno", remaining = 0
-//   sv = reader.Seek(1);    // sv = "bc", remaining = 12
-//
-class CordRepBtreeReader {
- public:
-  using ReadResult = CordRepBtreeNavigator::ReadResult;
-  using Position = CordRepBtreeNavigator::Position;
-
-  // Returns true if this instance is not empty.
-  explicit operator bool() const { return navigator_.btree() != nullptr; }
-
-  // Returns the tree referenced by this instance or nullptr if empty.
-  CordRepBtree* btree() const { return navigator_.btree(); }
-
-  // Returns the current data edge inside the referenced btree.
-  // Requires that the current instance is not empty.
-  CordRep* node() const { return navigator_.Current(); }
-
-  // Returns the length of the referenced tree.
-  // Requires that the current instance is not empty.
-  size_t length() const;
-
-  // Returns the number of remaining bytes available for iteration, which is the
-  // number of bytes directly following the end of the last chunk returned.
-  // This value will be zero if we iterated over the last edge in the bound
-  // tree, in which case any call to Next() or Skip() will return an empty
-  // string_view reflecting the EOF state.
-  // Note that a call to `Seek()` resets `remaining` to a value based on the
-  // end position of the chunk returned by that call.
-  size_t remaining() const { return remaining_; }
-
-  // Resets this instance to an empty value.
-  void Reset() { navigator_.Reset(); }
-
-  // Initializes this instance with `tree`. `tree` must not be null.
-  // Returns a reference to the first data edge of the provided tree.
-  absl::string_view Init(CordRepBtree* tree);
-
-  // Navigates to and returns the next data edge of the referenced tree.
-  // Returns an empty string_view if an attempt is made to read beyond the end
-  // of the tree, i.e.: if `remaining()` is zero indicating an EOF condition.
-  // Requires that the current instance is not empty.
-  absl::string_view Next();
-
-  // Skips the provided amount of bytes and returns a reference to the data
-  // directly following the skipped bytes.
-  absl::string_view Skip(size_t skip);
-
-  // Reads `n` bytes into `tree`.
-  // If `chunk_size` is zero, starts reading at the next data edge. If
-  // `chunk_size` is non zero, the read starts at the last `chunk_size` bytes of
-  // the last returned data edge. Effectively, this means that the read starts
-  // at offset `consumed() - chunk_size`.
-  // Requires that `chunk_size` is less than or equal to the length of the
-  // last returned data edge. The purpose of `chunk_size` is to simplify code
-  // partially consuming a returned chunk and wanting to include the remaining
-  // bytes in the Read call. For example, the below code will read 1000 bytes of
-  // data into a cord tree if the first chunk starts with "big:":
-  //
-  //   CordRepBtreeReader reader;
-  //   absl::string_view sv = reader.Init(tree);
-  //   if (absl::StartsWith(sv, "big:")) {
-  //     CordRepBtree tree;
-  //     sv = reader.Read(1000, sv.size() - 4 /* "big:" */, &tree);
-  //   }
-  //
-  // This method will return an empty string view if all remaining data was
-  // read. If `n` exceeded the amount of remaining data this function will
-  // return an empty string view and `tree` will be set to nullptr.
-  // In both cases, `consumed` will be set to `length`.
-  absl::string_view Read(size_t n, size_t chunk_size, CordRep*& tree);
-
-  // Navigates to the chunk at offset `offset`.
-  // Returns a reference into the navigated to chunk, adjusted for the relative
-  // position of `offset` into that chunk. For example, calling `Seek(13)` on a
-  // cord tree containing 2 chunks of 10 and 20 bytes respectively will return
-  // a string view into the second chunk starting at offset 3 with a size of 17.
-  // Returns an empty string view if `offset` is equal to or greater than the
-  // length of the referenced tree.
-  absl::string_view Seek(size_t offset);
-
- private:
-  size_t remaining_ = 0;
-  CordRepBtreeNavigator navigator_;
-};
-
-inline size_t CordRepBtreeReader::length() const {
-  assert(btree() != nullptr);
-  return btree()->length;
-}
-
-inline absl::string_view CordRepBtreeReader::Init(CordRepBtree* tree) {
-  assert(tree != nullptr);
-  const CordRep* edge = navigator_.InitFirst(tree);
-  remaining_ = tree->length - edge->length;
-  return EdgeData(edge);
-}
-
-inline absl::string_view CordRepBtreeReader::Next() {
-  if (remaining_ == 0) return {};
-  const CordRep* edge = navigator_.Next();
-  assert(edge != nullptr);
-  remaining_ -= edge->length;
-  return EdgeData(edge);
-}
-
-inline absl::string_view CordRepBtreeReader::Skip(size_t skip) {
-  // As we are always positioned on the last 'consumed' edge, we
-  // need to skip the current edge as well as `skip`.
-  const size_t edge_length = navigator_.Current()->length;
-  CordRepBtreeNavigator::Position pos = navigator_.Skip(skip + edge_length);
-  if (ABSL_PREDICT_FALSE(pos.edge == nullptr)) {
-    remaining_ = 0;
-    return {};
-  }
-  // The combined length of all edges skipped before `pos.edge` is `skip -
-  // pos.offset`, all of which are 'consumed', as well as the current edge.
-  remaining_ -= skip - pos.offset + pos.edge->length;
-  return EdgeData(pos.edge).substr(pos.offset);
-}
-
-inline absl::string_view CordRepBtreeReader::Seek(size_t offset) {
-  const CordRepBtreeNavigator::Position pos = navigator_.Seek(offset);
-  if (ABSL_PREDICT_FALSE(pos.edge == nullptr)) {
-    remaining_ = 0;
-    return {};
-  }
-  absl::string_view chunk = EdgeData(pos.edge).substr(pos.offset);
-  remaining_ = length() - offset - chunk.length();
-  return chunk;
-}
-
-}  // namespace cord_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CORD_REP_BTREE_READER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_consume.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_consume.h
deleted file mode 100644
index bece1874..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_consume.h
+++ /dev/null
@@ -1,47 +0,0 @@
-// Copyright 2021 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CORD_REP_CONSUME_H_
-#define ABSL_STRINGS_INTERNAL_CORD_REP_CONSUME_H_
-
-#include <functional>
-
-#include "absl/functional/function_ref.h"
-#include "absl/strings/internal/cord_internal.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// Consume() and ReverseConsume() consume CONCAT based trees and invoke the
-// provided functor with the contained nodes in the proper forward or reverse
-// order, which is used to convert CONCAT trees into other tree or cord data.
-// All CONCAT and SUBSTRING nodes are processed internally. The 'offset`
-// parameter of the functor is non-zero for any nodes below SUBSTRING nodes.
-// It's up to the caller to form these back into SUBSTRING nodes or otherwise
-// store offset / prefix information. These functions are intended to be used
-// only for migration / transitional code where due to factors such as ODR
-// violations, we can not 100% guarantee that all code respects 'new format'
-// settings and flags, so we need to be able to parse old data on the fly until
-// all old code is deprecated / no longer the default format.
-void Consume(CordRep* rep,
-             FunctionRef<void(CordRep*, size_t, size_t)> consume_fn);
-void ReverseConsume(CordRep* rep,
-                    FunctionRef<void(CordRep*, size_t, size_t)> consume_fn);
-
-}  // namespace cord_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CORD_REP_CONSUME_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_crc.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_crc.h
deleted file mode 100644
index 379d7a60..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_crc.h
+++ /dev/null
@@ -1,103 +0,0 @@
-// Copyright 2021 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CORD_REP_CRC_H_
-#define ABSL_STRINGS_INTERNAL_CORD_REP_CRC_H_
-
-#include <cassert>
-#include <cstdint>
-
-#include "absl/base/config.h"
-#include "absl/base/optimization.h"
-#include "absl/crc/internal/crc_cord_state.h"
-#include "absl/strings/internal/cord_internal.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// CordRepCrc is a CordRep node intended only to appear at the top level of a
-// cord tree.  It associates an "expected CRC" with the contained data, to allow
-// for easy passage of checksum data in Cord data flows.
-//
-// From Cord's perspective, the crc value has no semantics; any validation of
-// the contained checksum is the user's responsibility.
-struct CordRepCrc : public CordRep {
-  CordRep* child;
-  absl::crc_internal::CrcCordState crc_cord_state;
-
-  // Consumes `child` and returns a CordRepCrc prefixed tree containing `child`.
-  // If the specified `child` is itself a CordRepCrc node, then this method
-  // either replaces the existing node, or directly updates the crc state in it
-  // depending on the node being shared or not, i.e.: refcount.IsOne().
-  // `child` must only be null if the Cord is empty. Never returns null.
-  static CordRepCrc* New(CordRep* child, crc_internal::CrcCordState state);
-
-  // Destroys (deletes) the provided node. `node` must not be null.
-  static void Destroy(CordRepCrc* node);
-};
-
-// Consumes `rep` and returns a CordRep* with any outer CordRepCrc wrapper
-// removed.  This is usually a no-op (returning `rep`), but this will remove and
-// unref an outer CordRepCrc node.
-inline CordRep* RemoveCrcNode(CordRep* rep) {
-  assert(rep != nullptr);
-  if (ABSL_PREDICT_FALSE(rep->IsCrc())) {
-    CordRep* child = rep->crc()->child;
-    if (rep->refcount.IsOne()) {
-      delete rep->crc();
-    } else {
-      CordRep::Ref(child);
-      CordRep::Unref(rep);
-    }
-    return child;
-  }
-  return rep;
-}
-
-// Returns `rep` if it is not a CordRepCrc node, or its child if it is.
-// Does not consume or create a reference on `rep` or the returned value.
-inline CordRep* SkipCrcNode(CordRep* rep) {
-  assert(rep != nullptr);
-  if (ABSL_PREDICT_FALSE(rep->IsCrc())) {
-    return rep->crc()->child;
-  } else {
-    return rep;
-  }
-}
-
-inline const CordRep* SkipCrcNode(const CordRep* rep) {
-  assert(rep != nullptr);
-  if (ABSL_PREDICT_FALSE(rep->IsCrc())) {
-    return rep->crc()->child;
-  } else {
-    return rep;
-  }
-}
-
-inline CordRepCrc* CordRep::crc() {
-  assert(IsCrc());
-  return static_cast<CordRepCrc*>(this);
-}
-
-inline const CordRepCrc* CordRep::crc() const {
-  assert(IsCrc());
-  return static_cast<const CordRepCrc*>(this);
-}
-
-}  // namespace cord_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CORD_REP_CRC_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_flat.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_flat.h
deleted file mode 100644
index 27c4b21e..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_flat.h
+++ /dev/null
@@ -1,195 +0,0 @@
-// Copyright 2020 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CORD_REP_FLAT_H_
-#define ABSL_STRINGS_INTERNAL_CORD_REP_FLAT_H_
-
-#include <cassert>
-#include <cstddef>
-#include <cstdint>
-#include <memory>
-
-#include "absl/base/config.h"
-#include "absl/base/macros.h"
-#include "absl/strings/internal/cord_internal.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// Note: all constants below are never ODR used and internal to cord, we define
-// these as static constexpr to avoid 'in struct' definition and usage clutter.
-
-// Largest and smallest flat node lengths we are willing to allocate
-// Flat allocation size is stored in tag, which currently can encode sizes up
-// to 4K, encoded as multiple of either 8 or 32 bytes.
-// If we allow for larger sizes, we need to change this to 8/64, 16/128, etc.
-// kMinFlatSize is bounded by tag needing to be at least FLAT * 8 bytes, and
-// ideally a 'nice' size aligning with allocation and cacheline sizes like 32.
-// kMaxFlatSize is bounded by the size resulting in a computed tag no greater
-// than MAX_FLAT_TAG. MAX_FLAT_TAG provides for additional 'high' tag values.
-static constexpr size_t kFlatOverhead = offsetof(CordRep, storage);
-static constexpr size_t kMinFlatSize = 32;
-static constexpr size_t kMaxFlatSize = 4096;
-static constexpr size_t kMaxFlatLength = kMaxFlatSize - kFlatOverhead;
-static constexpr size_t kMinFlatLength = kMinFlatSize - kFlatOverhead;
-static constexpr size_t kMaxLargeFlatSize = 256 * 1024;
-static constexpr size_t kMaxLargeFlatLength = kMaxLargeFlatSize - kFlatOverhead;
-
-// kTagBase should make the Size <--> Tag computation resilient
-// against changes to the value of FLAT when we add a new tag..
-static constexpr uint8_t kTagBase = FLAT - 4;
-
-// Converts the provided rounded size to the corresponding tag
-constexpr uint8_t AllocatedSizeToTagUnchecked(size_t size) {
-  return static_cast<uint8_t>(size <= 512 ? kTagBase + size / 8
-                              : size <= 8192
-                                  ? kTagBase + 512 / 8 + size / 64 - 512 / 64
-                                  : kTagBase + 512 / 8 + ((8192 - 512) / 64) +
-                                        size / 4096 - 8192 / 4096);
-}
-
-// Converts the provided tag to the corresponding allocated size
-constexpr size_t TagToAllocatedSize(uint8_t tag) {
-  return (tag <= kTagBase + 512 / 8) ? tag * 8 - kTagBase * 8
-         : (tag <= kTagBase + (512 / 8) + ((8192 - 512) / 64))
-             ? 512 + tag * 64 - kTagBase * 64 - 512 / 8 * 64
-             : 8192 + tag * 4096 - kTagBase * 4096 -
-                   ((512 / 8) + ((8192 - 512) / 64)) * 4096;
-}
-
-static_assert(AllocatedSizeToTagUnchecked(kMinFlatSize) == FLAT, "");
-static_assert(AllocatedSizeToTagUnchecked(kMaxLargeFlatSize) == MAX_FLAT_TAG,
-              "");
-
-// RoundUp logically performs `((n + m - 1) / m) * m` to round up to the nearest
-// multiple of `m`, optimized for the invariant that `m` is a power of 2.
-constexpr size_t RoundUp(size_t n, size_t m) {
-  return (n + m - 1) & (0 - m);
-}
-
-// Returns the size to the nearest equal or larger value that can be
-// expressed exactly as a tag value.
-inline size_t RoundUpForTag(size_t size) {
-  return RoundUp(size, (size <= 512) ? 8 : (size <= 8192 ? 64 : 4096));
-}
-
-// Converts the allocated size to a tag, rounding down if the size
-// does not exactly match a 'tag expressible' size value. The result is
-// undefined if the size exceeds the maximum size that can be encoded in
-// a tag, i.e., if size is larger than TagToAllocatedSize(<max tag>).
-inline uint8_t AllocatedSizeToTag(size_t size) {
-  const uint8_t tag = AllocatedSizeToTagUnchecked(size);
-  assert(tag <= MAX_FLAT_TAG);
-  return tag;
-}
-
-// Converts the provided tag to the corresponding available data length
-constexpr size_t TagToLength(uint8_t tag) {
-  return TagToAllocatedSize(tag) - kFlatOverhead;
-}
-
-// Enforce that kMaxFlatSize maps to a well-known exact tag value.
-static_assert(TagToAllocatedSize(MAX_FLAT_TAG) == kMaxLargeFlatSize,
-              "Bad tag logic");
-
-struct CordRepFlat : public CordRep {
-  // Tag for explicit 'large flat' allocation
-  struct Large {};
-
-  // Creates a new flat node.
-  template <size_t max_flat_size, typename... Args>
-  static CordRepFlat* NewImpl(size_t len, Args... args ABSL_ATTRIBUTE_UNUSED) {
-    if (len <= kMinFlatLength) {
-      len = kMinFlatLength;
-    } else if (len > max_flat_size - kFlatOverhead) {
-      len = max_flat_size - kFlatOverhead;
-    }
-
-    // Round size up so it matches a size we can exactly express in a tag.
-    const size_t size = RoundUpForTag(len + kFlatOverhead);
-    void* const raw_rep = ::operator new(size);
-    // GCC 13 has a false-positive -Wstringop-overflow warning here.
-    #if ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(13, 0)
-    #pragma GCC diagnostic push
-    #pragma GCC diagnostic ignored "-Wstringop-overflow"
-    #endif
-    CordRepFlat* rep = new (raw_rep) CordRepFlat();
-    rep->tag = AllocatedSizeToTag(size);
-    #if ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(13, 0)
-    #pragma GCC diagnostic pop
-    #endif
-    return rep;
-  }
-
-  static CordRepFlat* New(size_t len) { return NewImpl<kMaxFlatSize>(len); }
-
-  static CordRepFlat* New(Large, size_t len) {
-    return NewImpl<kMaxLargeFlatSize>(len);
-  }
-
-  // Deletes a CordRepFlat instance created previously through a call to New().
-  // Flat CordReps are allocated and constructed with raw ::operator new and
-  // placement new, and must be destructed and deallocated accordingly.
-  static void Delete(CordRep*rep) {
-    assert(rep->tag >= FLAT && rep->tag <= MAX_FLAT_TAG);
-
-#if defined(__cpp_sized_deallocation)
-    size_t size = TagToAllocatedSize(rep->tag);
-    rep->~CordRep();
-    ::operator delete(rep, size);
-#else
-    rep->~CordRep();
-    ::operator delete(rep);
-#endif
-  }
-
-  // Create a CordRepFlat containing `data`, with an optional additional
-  // extra capacity of up to `extra` bytes. Requires that `data.size()`
-  // is less than kMaxFlatLength.
-  static CordRepFlat* Create(absl::string_view data, size_t extra = 0) {
-    assert(data.size() <= kMaxFlatLength);
-    CordRepFlat* flat = New(data.size() + (std::min)(extra, kMaxFlatLength));
-    memcpy(flat->Data(), data.data(), data.size());
-    flat->length = data.size();
-    return flat;
-  }
-
-  // Returns a pointer to the data inside this flat rep.
-  char* Data() { return reinterpret_cast<char*>(storage); }
-  const char* Data() const { return reinterpret_cast<const char*>(storage); }
-
-  // Returns the maximum capacity (payload size) of this instance.
-  size_t Capacity() const { return TagToLength(tag); }
-
-  // Returns the allocated size (payload + overhead) of this instance.
-  size_t AllocatedSize() const { return TagToAllocatedSize(tag); }
-};
-
-// Now that CordRepFlat is defined, we can define CordRep's helper casts:
-inline CordRepFlat* CordRep::flat() {
-  assert(tag >= FLAT && tag <= MAX_FLAT_TAG);
-  return reinterpret_cast<CordRepFlat*>(this);
-}
-
-inline const CordRepFlat* CordRep::flat() const {
-  assert(tag >= FLAT && tag <= MAX_FLAT_TAG);
-  return reinterpret_cast<const CordRepFlat*>(this);
-}
-
-}  // namespace cord_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CORD_REP_FLAT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_test_util.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_test_util.h
deleted file mode 100644
index 18a0a195..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cord_rep_test_util.h
+++ /dev/null
@@ -1,205 +0,0 @@
-// Copyright 2021 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CORD_REP_TEST_UTIL_H_
-#define ABSL_STRINGS_INTERNAL_CORD_REP_TEST_UTIL_H_
-
-#include <cassert>
-#include <memory>
-#include <random>
-#include <string>
-#include <vector>
-
-#include "absl/base/config.h"
-#include "absl/base/internal/raw_logging.h"
-#include "absl/strings/internal/cord_internal.h"
-#include "absl/strings/internal/cord_rep_btree.h"
-#include "absl/strings/internal/cord_rep_flat.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cordrep_testing {
-
-inline cord_internal::CordRepSubstring* MakeSubstring(
-    size_t start, size_t len, cord_internal::CordRep* rep) {
-  auto* sub = new cord_internal::CordRepSubstring;
-  sub->tag = cord_internal::SUBSTRING;
-  sub->start = start;
-  sub->length = len <= 0 ? rep->length - start + len : len;
-  sub->child = rep;
-  return sub;
-}
-
-inline cord_internal::CordRepFlat* MakeFlat(absl::string_view value) {
-  assert(value.length() <= cord_internal::kMaxFlatLength);
-  auto* flat = cord_internal::CordRepFlat::New(value.length());
-  flat->length = value.length();
-  memcpy(flat->Data(), value.data(), value.length());
-  return flat;
-}
-
-// Creates an external node for testing
-inline cord_internal::CordRepExternal* MakeExternal(absl::string_view s) {
-  struct Rep : public cord_internal::CordRepExternal {
-    std::string s;
-    explicit Rep(absl::string_view sv) : s(sv) {
-      this->tag = cord_internal::EXTERNAL;
-      this->base = s.data();
-      this->length = s.length();
-      this->releaser_invoker = [](cord_internal::CordRepExternal* self) {
-        delete static_cast<Rep*>(self);
-      };
-    }
-  };
-  return new Rep(s);
-}
-
-inline std::string CreateRandomString(size_t n) {
-  absl::string_view data =
-      "abcdefghijklmnopqrstuvwxyz"
-      "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
-      "0123456789~!@#$%^&*()_+=-<>?:\"{}[]|";
-  std::minstd_rand rnd;
-  std::uniform_int_distribution<size_t> dist(0, data.size() - 1);
-  std::string s(n, ' ');
-  for (size_t i = 0; i < n; ++i) {
-    s[i] = data[dist(rnd)];
-  }
-  return s;
-}
-
-// Creates an array of flats from the provided string, chopping
-// the provided string up into flats of size `chunk_size` characters
-// resulting in roughly `data.size() / chunk_size` total flats.
-inline std::vector<cord_internal::CordRep*> CreateFlatsFromString(
-    absl::string_view data, size_t chunk_size) {
-  assert(chunk_size > 0);
-  std::vector<cord_internal::CordRep*> flats;
-  for (absl::string_view s = data; !s.empty(); s.remove_prefix(chunk_size)) {
-    flats.push_back(MakeFlat(s.substr(0, chunk_size)));
-  }
-  return flats;
-}
-
-inline cord_internal::CordRepBtree* CordRepBtreeFromFlats(
-    absl::Span<cord_internal::CordRep* const> flats) {
-  assert(!flats.empty());
-  auto* node = cord_internal::CordRepBtree::Create(flats[0]);
-  for (size_t i = 1; i < flats.size(); ++i) {
-    node = cord_internal::CordRepBtree::Append(node, flats[i]);
-  }
-  return node;
-}
-
-template <typename Fn>
-inline void CordVisitReps(cord_internal::CordRep* rep, Fn&& fn) {
-  fn(rep);
-  while (rep->tag == cord_internal::SUBSTRING) {
-    rep = rep->substring()->child;
-    fn(rep);
-  }
-  if (rep->tag == cord_internal::BTREE) {
-    for (cord_internal::CordRep* edge : rep->btree()->Edges()) {
-      CordVisitReps(edge, fn);
-    }
-  }
-}
-
-template <typename Predicate>
-inline std::vector<cord_internal::CordRep*> CordCollectRepsIf(
-    Predicate&& predicate, cord_internal::CordRep* rep) {
-  std::vector<cord_internal::CordRep*> reps;
-  CordVisitReps(rep, [&reps, &predicate](cord_internal::CordRep* rep) {
-    if (predicate(rep)) reps.push_back(rep);
-  });
-  return reps;
-}
-
-inline std::vector<cord_internal::CordRep*> CordCollectReps(
-    cord_internal::CordRep* rep) {
-  std::vector<cord_internal::CordRep*> reps;
-  auto fn = [&reps](cord_internal::CordRep* rep) { reps.push_back(rep); };
-  CordVisitReps(rep, fn);
-  return reps;
-}
-
-inline void CordToString(cord_internal::CordRep* rep, std::string& s) {
-  size_t offset = 0;
-  size_t length = rep->length;
-  while (rep->tag == cord_internal::SUBSTRING) {
-    offset += rep->substring()->start;
-    rep = rep->substring()->child;
-  }
-  if (rep->tag == cord_internal::BTREE) {
-    for (cord_internal::CordRep* edge : rep->btree()->Edges()) {
-      CordToString(edge, s);
-    }
-  } else if (rep->tag >= cord_internal::FLAT) {
-    s.append(rep->flat()->Data() + offset, length);
-  } else if (rep->tag == cord_internal::EXTERNAL) {
-    s.append(rep->external()->base + offset, length);
-  } else {
-    ABSL_RAW_LOG(FATAL, "Unsupported tag %d", rep->tag);
-  }
-}
-
-inline std::string CordToString(cord_internal::CordRep* rep) {
-  std::string s;
-  s.reserve(rep->length);
-  CordToString(rep, s);
-  return s;
-}
-
-// RAII Helper class to automatically unref reps on destruction.
-class AutoUnref {
- public:
-  ~AutoUnref() {
-    for (CordRep* rep : unrefs_) CordRep::Unref(rep);
-  }
-
-  // Adds `rep` to the list of reps to be unreffed at destruction.
-  template <typename CordRepType>
-  CordRepType* Add(CordRepType* rep) {
-    unrefs_.push_back(rep);
-    return rep;
-  }
-
-  // Increments the reference count of `rep` by one, and adds it to
-  // the list of reps to be unreffed at destruction.
-  template <typename CordRepType>
-  CordRepType* Ref(CordRepType* rep) {
-    unrefs_.push_back(CordRep::Ref(rep));
-    return rep;
-  }
-
-  // Increments the reference count of `rep` by one if `condition` is true,
-  // and adds it to the list of reps to be unreffed at destruction.
-  template <typename CordRepType>
-  CordRepType* RefIf(bool condition, CordRepType* rep) {
-    if (condition) unrefs_.push_back(CordRep::Ref(rep));
-    return rep;
-  }
-
- private:
-  using CordRep = absl::cord_internal::CordRep;
-
-  std::vector<CordRep*> unrefs_;
-};
-
-}  // namespace cordrep_testing
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CORD_REP_TEST_UTIL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_functions.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_functions.h
deleted file mode 100644
index ed108bf1..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_functions.h
+++ /dev/null
@@ -1,77 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CORDZ_FUNCTIONS_H_
-#define ABSL_STRINGS_INTERNAL_CORDZ_FUNCTIONS_H_
-
-#include <stdint.h>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/optimization.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// Returns the current sample rate. This represents the average interval
-// between samples.
-int32_t get_cordz_mean_interval();
-
-// Sets the sample rate with the average interval between samples.
-void set_cordz_mean_interval(int32_t mean_interval);
-
-// Cordz is only enabled on Linux with thread_local support.
-#if defined(ABSL_INTERNAL_CORDZ_ENABLED)
-#error ABSL_INTERNAL_CORDZ_ENABLED cannot be set directly
-#elif defined(__linux__) && defined(ABSL_HAVE_THREAD_LOCAL)
-#define ABSL_INTERNAL_CORDZ_ENABLED 1
-#endif
-
-#ifdef ABSL_INTERNAL_CORDZ_ENABLED
-
-// cordz_next_sample is the number of events until the next sample event. If
-// the value is 1 or less, the code will check on the next event if cordz is
-// enabled, and if so, will sample the Cord. cordz is only enabled when we can
-// use thread locals.
-ABSL_CONST_INIT extern thread_local int64_t cordz_next_sample;
-
-// Determines if the next sample should be profiled. If it is, the value pointed
-// at by next_sample will be set with the interval until the next sample.
-bool cordz_should_profile_slow();
-
-// Returns true if the next cord should be sampled.
-inline bool cordz_should_profile() {
-  if (ABSL_PREDICT_TRUE(cordz_next_sample > 1)) {
-    cordz_next_sample--;
-    return false;
-  }
-  return cordz_should_profile_slow();
-}
-
-// Sets the interval until the next sample (for testing only)
-void cordz_set_next_sample_for_testing(int64_t next_sample);
-
-#else  // ABSL_INTERNAL_CORDZ_ENABLED
-
-inline bool cordz_should_profile() { return false; }
-inline void cordz_set_next_sample_for_testing(int64_t) {}
-
-#endif  // ABSL_INTERNAL_CORDZ_ENABLED
-
-}  // namespace cord_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CORDZ_FUNCTIONS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_handle.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_handle.h
deleted file mode 100644
index 08e3f0d3..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_handle.h
+++ /dev/null
@@ -1,98 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CORDZ_HANDLE_H_
-#define ABSL_STRINGS_INTERNAL_CORDZ_HANDLE_H_
-
-#include <atomic>
-#include <vector>
-
-#include "absl/base/config.h"
-#include "absl/base/internal/raw_logging.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// This base class allows multiple types of object (CordzInfo and
-// CordzSampleToken) to exist simultaneously on the delete queue (pointed to by
-// global_dq_tail and traversed using dq_prev_ and dq_next_). The
-// delete queue guarantees that once a profiler creates a CordzSampleToken and
-// has gained visibility into a CordzInfo object, that CordzInfo object will not
-// be deleted prematurely. This allows the profiler to inspect all CordzInfo
-// objects that are alive without needing to hold a global lock.
-class ABSL_DLL CordzHandle {
- public:
-  CordzHandle() : CordzHandle(false) {}
-
-  bool is_snapshot() const { return is_snapshot_; }
-
-  // Returns true if this instance is safe to be deleted because it is either a
-  // snapshot, which is always safe to delete, or not included in the global
-  // delete queue and thus not included in any snapshot.
-  // Callers are responsible for making sure this instance can not be newly
-  // discovered by other threads. For example, CordzInfo instances first de-list
-  // themselves from the global CordzInfo list before determining if they are
-  // safe to be deleted directly.
-  // If SafeToDelete returns false, callers MUST use the Delete() method to
-  // safely queue CordzHandle instances for deletion.
-  bool SafeToDelete() const;
-
-  // Deletes the provided instance, or puts it on the delete queue to be deleted
-  // once there are no more sample tokens (snapshot) instances potentially
-  // referencing the instance. `handle` should not be null.
-  static void Delete(CordzHandle* handle);
-
-  // Returns the current entries in the delete queue in LIFO order.
-  static std::vector<const CordzHandle*> DiagnosticsGetDeleteQueue();
-
-  // Returns true if the provided handle is nullptr or guarded by this handle.
-  // Since the CordzSnapshot token is itself a CordzHandle, this method will
-  // allow tests to check if that token is keeping an arbitrary CordzHandle
-  // alive.
-  bool DiagnosticsHandleIsSafeToInspect(const CordzHandle* handle) const;
-
-  // Returns the current entries in the delete queue, in LIFO order, that are
-  // protected by this. CordzHandle objects are only placed on the delete queue
-  // after CordzHandle::Delete is called with them as an argument. Only
-  // CordzHandle objects that are not also CordzSnapshot objects will be
-  // included in the return vector. For each of the handles in the return
-  // vector, the earliest that their memory can be freed is when this
-  // CordzSnapshot object is deleted.
-  std::vector<const CordzHandle*> DiagnosticsGetSafeToInspectDeletedHandles();
-
- protected:
-  explicit CordzHandle(bool is_snapshot);
-  virtual ~CordzHandle();
-
- private:
-  const bool is_snapshot_;
-
-  // dq_prev_ and dq_next_ require the global queue mutex to be held.
-  // Unfortunately we can't use thread annotations such that the thread safety
-  // analysis understands that queue_ and global_queue_ are one and the same.
-  CordzHandle* dq_prev_  = nullptr;
-  CordzHandle* dq_next_ = nullptr;
-};
-
-class CordzSnapshot : public CordzHandle {
- public:
-  CordzSnapshot() : CordzHandle(true) {}
-};
-
-}  // namespace cord_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CORDZ_HANDLE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_info.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_info.h
deleted file mode 100644
index 17eaa91c..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_info.h
+++ /dev/null
@@ -1,298 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CORDZ_INFO_H_
-#define ABSL_STRINGS_INTERNAL_CORDZ_INFO_H_
-
-#include <atomic>
-#include <cstdint>
-#include <functional>
-
-#include "absl/base/config.h"
-#include "absl/base/internal/raw_logging.h"
-#include "absl/base/internal/spinlock.h"
-#include "absl/base/thread_annotations.h"
-#include "absl/strings/internal/cord_internal.h"
-#include "absl/strings/internal/cordz_functions.h"
-#include "absl/strings/internal/cordz_handle.h"
-#include "absl/strings/internal/cordz_statistics.h"
-#include "absl/strings/internal/cordz_update_tracker.h"
-#include "absl/synchronization/mutex.h"
-#include "absl/types/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// CordzInfo tracks a profiled Cord. Each of these objects can be in two places.
-// If a Cord is alive, the CordzInfo will be in the global_cordz_infos map, and
-// can also be retrieved via the linked list starting with
-// global_cordz_infos_head and continued via the cordz_info_next() method. When
-// a Cord has reached the end of its lifespan, the CordzInfo object will be
-// migrated out of the global_cordz_infos list and the global_cordz_infos_map,
-// and will either be deleted or appended to the global_delete_queue. If it is
-// placed on the global_delete_queue, the CordzInfo object will be cleaned in
-// the destructor of a CordzSampleToken object.
-class ABSL_LOCKABLE CordzInfo : public CordzHandle {
- public:
-  using MethodIdentifier = CordzUpdateTracker::MethodIdentifier;
-
-  // TrackCord creates a CordzInfo instance which tracks important metrics of
-  // a sampled cord, and stores the created CordzInfo instance into `cord'. All
-  // CordzInfo instances are placed in a global list which is used to discover
-  // and snapshot all actively tracked cords. Callers are responsible for
-  // calling UntrackCord() before the tracked Cord instance is deleted, or to
-  // stop tracking the sampled Cord. Callers are also responsible for guarding
-  // changes to the 'tree' value of a Cord (InlineData.tree) through the Lock()
-  // and Unlock() calls. Any change resulting in a new tree value for the cord
-  // requires a call to SetCordRep() before the old tree has been unreffed
-  // and/or deleted. `method` identifies the Cord public API method initiating
-  // the cord to be sampled.
-  // Requires `cord` to hold a tree, and `cord.cordz_info()` to be null.
-  static void TrackCord(InlineData& cord, MethodIdentifier method);
-
-  // Identical to TrackCord(), except that this function fills the
-  // `parent_stack` and `parent_method` properties of the returned CordzInfo
-  // instance from the provided `src` instance if `src` is sampled.
-  // This function should be used for sampling 'copy constructed' and 'copy
-  // assigned' cords. This function allows 'cord` to be already sampled, in
-  // which case the CordzInfo will be newly created from `src`.
-  static void TrackCord(InlineData& cord, const InlineData& src,
-                        MethodIdentifier method);
-
-  // Maybe sample the cord identified by 'cord' for method 'method'.
-  // Uses `cordz_should_profile` to randomly pick cords to be sampled, and if
-  // so, invokes `TrackCord` to start sampling `cord`.
-  static void MaybeTrackCord(InlineData& cord, MethodIdentifier method);
-
-  // Maybe sample the cord identified by 'cord' for method 'method'.
-  // `src` identifies a 'parent' cord which is assigned to `cord`, typically the
-  // input cord for a copy constructor, or an assign method such as `operator=`
-  // `cord` will be sampled if (and only if) `src` is sampled.
-  // If `cord` is currently being sampled and `src` is not being sampled, then
-  // this function will stop sampling the cord and reset the cord's cordz_info.
-  //
-  // Previously this function defined that `cord` will be sampled if either
-  // `src` is sampled, or if `cord` is randomly picked for sampling. However,
-  // this can cause issues, as there may be paths where some cord is assigned an
-  // indirect copy of it's own value. As such a 'string of copies' would then
-  // remain sampled (`src.is_profiled`), then assigning such a cord back to
-  // 'itself' creates a cycle where the cord will converge to 'always sampled`.
-  //
-  // For example:
-  //
-  //   Cord x;
-  //   for (...) {
-  //     // Copy ctor --> y.is_profiled := x.is_profiled | random(...)
-  //     Cord y = x;
-  //     ...
-  //     // Assign x = y --> x.is_profiled = y.is_profiled | random(...)
-  //     //              ==> x.is_profiled |= random(...)
-  //     //              ==> x converges to 'always profiled'
-  //     x = y;
-  //   }
-  static void MaybeTrackCord(InlineData& cord, const InlineData& src,
-                             MethodIdentifier method);
-
-  // Stops tracking changes for a sampled cord, and deletes the provided info.
-  // This function must be called before the sampled cord instance is deleted,
-  // and before the root cordrep of the sampled cord is unreffed.
-  // This function may extend the lifetime of the cordrep in cases where the
-  // CordInfo instance is being held by a concurrent collection thread.
-  void Untrack();
-
-  // Invokes UntrackCord() on `info` if `info` is not null.
-  static void MaybeUntrackCord(CordzInfo* info);
-
-  CordzInfo() = delete;
-  CordzInfo(const CordzInfo&) = delete;
-  CordzInfo& operator=(const CordzInfo&) = delete;
-
-  // Retrieves the oldest existing CordzInfo.
-  static CordzInfo* Head(const CordzSnapshot& snapshot)
-      ABSL_NO_THREAD_SAFETY_ANALYSIS;
-
-  // Retrieves the next oldest existing CordzInfo older than 'this' instance.
-  CordzInfo* Next(const CordzSnapshot& snapshot) const
-      ABSL_NO_THREAD_SAFETY_ANALYSIS;
-
-  // Locks this instance for the update identified by `method`.
-  // Increases the count for `method` in `update_tracker`.
-  void Lock(MethodIdentifier method) ABSL_EXCLUSIVE_LOCK_FUNCTION(mutex_);
-
-  // Unlocks this instance. If the contained `rep` has been set to null
-  // indicating the Cord has been cleared or is otherwise no longer sampled,
-  // then this method will delete this CordzInfo instance.
-  void Unlock() ABSL_UNLOCK_FUNCTION(mutex_);
-
-  // Asserts that this CordzInfo instance is locked.
-  void AssertHeld() ABSL_ASSERT_EXCLUSIVE_LOCK(mutex_);
-
-  // Updates the `rep` property of this instance. This methods is invoked by
-  // Cord logic each time the root node of a sampled Cord changes, and before
-  // the old root reference count is deleted. This guarantees that collection
-  // code can always safely take a reference on the tracked cord.
-  // Requires a lock to be held through the `Lock()` method.
-  // TODO(b/117940323): annotate with ABSL_EXCLUSIVE_LOCKS_REQUIRED once all
-  // Cord code is in a state where this can be proven true by the compiler.
-  void SetCordRep(CordRep* rep);
-
-  // Returns the current `rep` property of this instance with a reference
-  // added, or null if this instance represents a cord that has since been
-  // deleted or untracked.
-  CordRep* RefCordRep() const ABSL_LOCKS_EXCLUDED(mutex_);
-
-  // Returns the current value of `rep_` for testing purposes only.
-  CordRep* GetCordRepForTesting() const ABSL_NO_THREAD_SAFETY_ANALYSIS {
-    return rep_;
-  }
-
-  // Sets the current value of `rep_` for testing purposes only.
-  void SetCordRepForTesting(CordRep* rep) ABSL_NO_THREAD_SAFETY_ANALYSIS {
-    rep_ = rep;
-  }
-
-  // Returns the stack trace for where the cord was first sampled. Cords are
-  // potentially sampled when they promote from an inlined cord to a tree or
-  // ring representation, which is not necessarily the location where the cord
-  // was first created. Some cords are created as inlined cords, and only as
-  // data is added do they become a non-inlined cord. However, typically the
-  // location represents reasonably well where the cord is 'created'.
-  absl::Span<void* const> GetStack() const;
-
-  // Returns the stack trace for a sampled cord's 'parent stack trace'. This
-  // value may be set if the cord is sampled (promoted) after being created
-  // from, or being assigned the value of an existing (sampled) cord.
-  absl::Span<void* const> GetParentStack() const;
-
-  // Retrieves the CordzStatistics associated with this Cord. The statistics
-  // are only updated when a Cord goes through a mutation, such as an Append
-  // or RemovePrefix.
-  CordzStatistics GetCordzStatistics() const;
-
- private:
-  using SpinLock = absl::base_internal::SpinLock;
-  using SpinLockHolder = ::absl::base_internal::SpinLockHolder;
-
-  // Global cordz info list. CordzInfo stores a pointer to the global list
-  // instance to harden against ODR violations.
-  struct List {
-    constexpr explicit List(absl::ConstInitType)
-        : mutex(absl::kConstInit,
-                absl::base_internal::SCHEDULE_COOPERATIVE_AND_KERNEL) {}
-
-    SpinLock mutex;
-    std::atomic<CordzInfo*> head ABSL_GUARDED_BY(mutex){nullptr};
-  };
-
-  static constexpr size_t kMaxStackDepth = 64;
-
-  explicit CordzInfo(CordRep* rep, const CordzInfo* src,
-                     MethodIdentifier method);
-  ~CordzInfo() override;
-
-  // Sets `rep_` without holding a lock.
-  void UnsafeSetCordRep(CordRep* rep) ABSL_NO_THREAD_SAFETY_ANALYSIS;
-
-  void Track();
-
-  // Returns the parent method from `src`, which is either `parent_method_` or
-  // `method_` depending on `parent_method_` being kUnknown.
-  // Returns kUnknown if `src` is null.
-  static MethodIdentifier GetParentMethod(const CordzInfo* src);
-
-  // Fills the provided stack from `src`, copying either `parent_stack_` or
-  // `stack_` depending on `parent_stack_` being empty, returning the size of
-  // the parent stack.
-  // Returns 0 if `src` is null.
-  static size_t FillParentStack(const CordzInfo* src, void** stack);
-
-  void ODRCheck() const {
-#ifndef NDEBUG
-    ABSL_RAW_CHECK(list_ == &global_list_, "ODR violation in Cord");
-#endif
-  }
-
-  // Non-inlined implementation of `MaybeTrackCord`, which is executed if
-  // either `src` is sampled or `cord` is sampled, and either untracks or
-  // tracks `cord` as documented per `MaybeTrackCord`.
-  static void MaybeTrackCordImpl(InlineData& cord, const InlineData& src,
-                                 MethodIdentifier method);
-
-  ABSL_CONST_INIT static List global_list_;
-  List* const list_ = &global_list_;
-
-  // ci_prev_ and ci_next_ require the global list mutex to be held.
-  // Unfortunately we can't use thread annotations such that the thread safety
-  // analysis understands that list_ and global_list_ are one and the same.
-  std::atomic<CordzInfo*> ci_prev_{nullptr};
-  std::atomic<CordzInfo*> ci_next_{nullptr};
-
-  mutable absl::Mutex mutex_;
-  CordRep* rep_ ABSL_GUARDED_BY(mutex_);
-
-  void* stack_[kMaxStackDepth];
-  void* parent_stack_[kMaxStackDepth];
-  const size_t stack_depth_;
-  const size_t parent_stack_depth_;
-  const MethodIdentifier method_;
-  const MethodIdentifier parent_method_;
-  CordzUpdateTracker update_tracker_;
-  const absl::Time create_time_;
-};
-
-inline ABSL_ATTRIBUTE_ALWAYS_INLINE void CordzInfo::MaybeTrackCord(
-    InlineData& cord, MethodIdentifier method) {
-  if (ABSL_PREDICT_FALSE(cordz_should_profile())) {
-    TrackCord(cord, method);
-  }
-}
-
-inline ABSL_ATTRIBUTE_ALWAYS_INLINE void CordzInfo::MaybeTrackCord(
-    InlineData& cord, const InlineData& src, MethodIdentifier method) {
-  if (ABSL_PREDICT_FALSE(InlineData::is_either_profiled(cord, src))) {
-    MaybeTrackCordImpl(cord, src, method);
-  }
-}
-
-inline ABSL_ATTRIBUTE_ALWAYS_INLINE void CordzInfo::MaybeUntrackCord(
-    CordzInfo* info) {
-  if (ABSL_PREDICT_FALSE(info)) {
-    info->Untrack();
-  }
-}
-
-inline void CordzInfo::AssertHeld() ABSL_ASSERT_EXCLUSIVE_LOCK(mutex_) {
-#ifndef NDEBUG
-  mutex_.AssertHeld();
-#endif
-}
-
-inline void CordzInfo::SetCordRep(CordRep* rep) {
-  AssertHeld();
-  rep_ = rep;
-}
-
-inline void CordzInfo::UnsafeSetCordRep(CordRep* rep) { rep_ = rep; }
-
-inline CordRep* CordzInfo::RefCordRep() const ABSL_LOCKS_EXCLUDED(mutex_) {
-  MutexLock lock(&mutex_);
-  return rep_ ? CordRep::Ref(rep_) : nullptr;
-}
-
-}  // namespace cord_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CORDZ_INFO_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_sample_token.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_sample_token.h
deleted file mode 100644
index 2a86bc3b..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_sample_token.h
+++ /dev/null
@@ -1,97 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#include "absl/base/config.h"
-#include "absl/strings/internal/cordz_handle.h"
-#include "absl/strings/internal/cordz_info.h"
-
-#ifndef ABSL_STRINGS_INTERNAL_CORDZ_SAMPLE_TOKEN_H_
-#define ABSL_STRINGS_INTERNAL_CORDZ_SAMPLE_TOKEN_H_
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// The existence of a CordzSampleToken guarantees that a reader can traverse the
-// global_cordz_infos_head linked-list without needing to hold a mutex. When a
-// CordzSampleToken exists, all CordzInfo objects that would be destroyed are
-// instead appended to a deletion queue. When the CordzSampleToken is destroyed,
-// it will also clean up any of these CordzInfo objects.
-//
-// E.g., ST are CordzSampleToken objects and CH are CordzHandle objects.
-//   ST1 <- CH1 <- CH2 <- ST2 <- CH3 <- global_delete_queue_tail
-//
-// This list tracks that CH1 and CH2 were created after ST1, so the thread
-// holding ST1 might have a reference to CH1, CH2, ST2, and CH3. However, ST2
-// was created later, so the thread holding the ST2 token cannot have a
-// reference to ST1, CH1, or CH2. If ST1 is cleaned up first, that thread will
-// delete ST1, CH1, and CH2. If instead ST2 is cleaned up first, that thread
-// will only delete ST2.
-//
-// If ST1 is cleaned up first, the new list will be:
-//   ST2 <- CH3 <- global_delete_queue_tail
-//
-// If ST2 is cleaned up first, the new list will be:
-//   ST1 <- CH1 <- CH2 <- CH3 <- global_delete_queue_tail
-//
-// All new CordzHandle objects are appended to the list, so if a new thread
-// comes along before either ST1 or ST2 are cleaned up, the new list will be:
-//   ST1 <- CH1 <- CH2 <- ST2 <- CH3 <- ST3 <- global_delete_queue_tail
-//
-// A thread must hold the global_delete_queue_mu mutex whenever it's altering
-// this list.
-//
-// It is safe for thread that holds a CordzSampleToken to read
-// global_cordz_infos at any time since the objects it is able to retrieve will
-// not be deleted while the CordzSampleToken exists.
-class CordzSampleToken : public CordzSnapshot {
- public:
-  class Iterator {
-   public:
-    using iterator_category = std::input_iterator_tag;
-    using value_type = const CordzInfo&;
-    using difference_type = ptrdiff_t;
-    using pointer = const CordzInfo*;
-    using reference = value_type;
-
-    Iterator() = default;
-
-    Iterator& operator++();
-    Iterator operator++(int);
-    friend bool operator==(const Iterator& lhs, const Iterator& rhs);
-    friend bool operator!=(const Iterator& lhs, const Iterator& rhs);
-    reference operator*() const;
-    pointer operator->() const;
-
-   private:
-    friend class CordzSampleToken;
-    explicit Iterator(const CordzSampleToken* token);
-
-    const CordzSampleToken* token_ = nullptr;
-    pointer current_ = nullptr;
-  };
-
-  CordzSampleToken() = default;
-  CordzSampleToken(const CordzSampleToken&) = delete;
-  CordzSampleToken& operator=(const CordzSampleToken&) = delete;
-
-  Iterator begin() { return Iterator(this); }
-  Iterator end() { return Iterator(); }
-};
-
-}  // namespace cord_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CORDZ_SAMPLE_TOKEN_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_statistics.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_statistics.h
deleted file mode 100644
index 9f558df4..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_statistics.h
+++ /dev/null
@@ -1,88 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CORDZ_STATISTICS_H_
-#define ABSL_STRINGS_INTERNAL_CORDZ_STATISTICS_H_
-
-#include <cstdint>
-
-#include "absl/base/config.h"
-#include "absl/strings/internal/cordz_update_tracker.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// CordzStatistics captures some meta information about a Cord's shape.
-struct CordzStatistics {
-  using MethodIdentifier = CordzUpdateTracker::MethodIdentifier;
-
-  // Node counts information
-  struct NodeCounts {
-    size_t flat = 0;       // #flats
-    size_t flat_64 = 0;    // #flats up to 64 bytes
-    size_t flat_128 = 0;   // #flats up to 128 bytes
-    size_t flat_256 = 0;   // #flats up to 256 bytes
-    size_t flat_512 = 0;   // #flats up to 512 bytes
-    size_t flat_1k = 0;    // #flats up to 1K bytes
-    size_t external = 0;   // #external reps
-    size_t substring = 0;  // #substring reps
-    size_t concat = 0;     // #concat reps
-    size_t ring = 0;       // #ring buffer reps
-    size_t btree = 0;      // #btree reps
-    size_t crc = 0;        // #crc reps
-  };
-
-  // The size of the cord in bytes. This matches the result of Cord::size().
-  size_t size = 0;
-
-  // The estimated memory used by the sampled cord. This value matches the
-  // value as reported by Cord::EstimatedMemoryUsage().
-  // A value of 0 implies the property has not been recorded.
-  size_t estimated_memory_usage = 0;
-
-  // The effective memory used by the sampled cord, inversely weighted by the
-  // effective indegree of each allocated node. This is a representation of the
-  // fair share of memory usage that should be attributed to the sampled cord.
-  // This value is more useful for cases where one or more nodes are referenced
-  // by multiple Cord instances, and for cases where a Cord includes the same
-  // node multiple times (either directly or indirectly).
-  // A value of 0 implies the property has not been recorded.
-  size_t estimated_fair_share_memory_usage = 0;
-
-  // The total number of nodes referenced by this cord.
-  // For ring buffer Cords, this includes the 'ring buffer' node.
-  // For btree Cords, this includes all 'CordRepBtree' tree nodes as well as all
-  // the substring, flat and external nodes referenced by the tree.
-  // A value of 0 implies the property has not been recorded.
-  size_t node_count = 0;
-
-  // Detailed node counts per type
-  NodeCounts node_counts;
-
-  // The cord method responsible for sampling the cord.
-  MethodIdentifier method = MethodIdentifier::kUnknown;
-
-  // The cord method responsible for sampling the parent cord if applicable.
-  MethodIdentifier parent_method = MethodIdentifier::kUnknown;
-
-  // Update tracker tracking invocation count per cord method.
-  CordzUpdateTracker update_tracker;
-};
-
-}  // namespace cord_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CORDZ_STATISTICS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_update_scope.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_update_scope.h
deleted file mode 100644
index 57ba75de..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_update_scope.h
+++ /dev/null
@@ -1,71 +0,0 @@
-// Copyright 2021 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CORDZ_UPDATE_SCOPE_H_
-#define ABSL_STRINGS_INTERNAL_CORDZ_UPDATE_SCOPE_H_
-
-#include "absl/base/config.h"
-#include "absl/base/optimization.h"
-#include "absl/base/thread_annotations.h"
-#include "absl/strings/internal/cord_internal.h"
-#include "absl/strings/internal/cordz_info.h"
-#include "absl/strings/internal/cordz_update_tracker.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// CordzUpdateScope scopes an update to the provided CordzInfo.
-// The class invokes `info->Lock(method)` and `info->Unlock()` to guard
-// cordrep updates. This class does nothing if `info` is null.
-// See also the 'Lock`, `Unlock` and `SetCordRep` methods in `CordzInfo`.
-class ABSL_SCOPED_LOCKABLE CordzUpdateScope {
- public:
-  CordzUpdateScope(CordzInfo* info, CordzUpdateTracker::MethodIdentifier method)
-      ABSL_EXCLUSIVE_LOCK_FUNCTION(info)
-      : info_(info) {
-    if (ABSL_PREDICT_FALSE(info_)) {
-      info->Lock(method);
-    }
-  }
-
-  // CordzUpdateScope can not be copied or assigned to.
-  CordzUpdateScope(CordzUpdateScope&& rhs) = delete;
-  CordzUpdateScope(const CordzUpdateScope&) = delete;
-  CordzUpdateScope& operator=(CordzUpdateScope&& rhs) = delete;
-  CordzUpdateScope& operator=(const CordzUpdateScope&) = delete;
-
-  ~CordzUpdateScope() ABSL_UNLOCK_FUNCTION() {
-    if (ABSL_PREDICT_FALSE(info_)) {
-      info_->Unlock();
-    }
-  }
-
-  void SetCordRep(CordRep* rep) const {
-    if (ABSL_PREDICT_FALSE(info_)) {
-      info_->SetCordRep(rep);
-    }
-  }
-
-  CordzInfo* info() const { return info_; }
-
- private:
-  CordzInfo* info_;
-};
-
-}  // namespace cord_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CORDZ_UPDATE_SCOPE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_update_tracker.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_update_tracker.h
deleted file mode 100644
index c5170662..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/cordz_update_tracker.h
+++ /dev/null
@@ -1,123 +0,0 @@
-// Copyright 2021 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_CORDZ_UPDATE_TRACKER_H_
-#define ABSL_STRINGS_INTERNAL_CORDZ_UPDATE_TRACKER_H_
-
-#include <atomic>
-#include <cstdint>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace cord_internal {
-
-// CordzUpdateTracker tracks counters for Cord update methods.
-//
-// The purpose of CordzUpdateTracker is to track the number of calls to methods
-// updating Cord data for sampled cords. The class internally uses 'lossy'
-// atomic operations: Cord is thread-compatible, so there is no need to
-// synchronize updates. However, Cordz collection threads may call 'Value()' at
-// any point, so the class needs to provide thread safe access.
-//
-// This class is thread-safe. But as per above comments, all non-const methods
-// should be used single-threaded only: updates are thread-safe but lossy.
-class CordzUpdateTracker {
- public:
-  // Tracked update methods.
-  enum MethodIdentifier {
-    kUnknown,
-    kAppendCord,
-    kAppendCordBuffer,
-    kAppendExternalMemory,
-    kAppendString,
-    kAssignCord,
-    kAssignString,
-    kClear,
-    kConstructorCord,
-    kConstructorString,
-    kCordReader,
-    kFlatten,
-    kGetAppendBuffer,
-    kGetAppendRegion,
-    kMakeCordFromExternal,
-    kMoveAppendCord,
-    kMoveAssignCord,
-    kMovePrependCord,
-    kPrependCord,
-    kPrependCordBuffer,
-    kPrependString,
-    kRemovePrefix,
-    kRemoveSuffix,
-    kSetExpectedChecksum,
-    kSubCord,
-
-    // kNumMethods defines the number of entries: must be the last entry.
-    kNumMethods,
-  };
-
-  // Constructs a new instance. All counters are zero-initialized.
-  constexpr CordzUpdateTracker() noexcept : values_{} {}
-
-  // Copy constructs a new instance.
-  CordzUpdateTracker(const CordzUpdateTracker& rhs) noexcept { *this = rhs; }
-
-  // Assigns the provided value to this instance.
-  CordzUpdateTracker& operator=(const CordzUpdateTracker& rhs) noexcept {
-    for (int i = 0; i < kNumMethods; ++i) {
-      values_[i].store(rhs.values_[i].load(std::memory_order_relaxed),
-                       std::memory_order_relaxed);
-    }
-    return *this;
-  }
-
-  // Returns the value for the specified method.
-  int64_t Value(MethodIdentifier method) const {
-    return values_[method].load(std::memory_order_relaxed);
-  }
-
-  // Increases the value for the specified method by `n`
-  void LossyAdd(MethodIdentifier method, int64_t n = 1) {
-    auto& value = values_[method];
-    value.store(value.load(std::memory_order_relaxed) + n,
-                std::memory_order_relaxed);
-  }
-
-  // Adds all the values from `src` to this instance
-  void LossyAdd(const CordzUpdateTracker& src) {
-    for (int i = 0; i < kNumMethods; ++i) {
-      MethodIdentifier method = static_cast<MethodIdentifier>(i);
-      if (int64_t value = src.Value(method)) {
-        LossyAdd(method, value);
-      }
-    }
-  }
-
- private:
-  // Until C++20 std::atomic is not constexpr default-constructible, so we need
-  // a wrapper for this class to be constexpr constructible.
-  class Counter : public std::atomic<int64_t> {
-   public:
-    constexpr Counter() noexcept : std::atomic<int64_t>(0) {}
-  };
-
-  Counter values_[kNumMethods];
-};
-
-}  // namespace cord_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_CORDZ_UPDATE_TRACKER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/damerau_levenshtein_distance.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/damerau_levenshtein_distance.h
deleted file mode 100644
index 7a4bd644..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/damerau_levenshtein_distance.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_DAMERAU_LEVENSHTEIN_DISTANCE_H_
-#define ABSL_STRINGS_INTERNAL_DAMERAU_LEVENSHTEIN_DISTANCE_H_
-
-#include <cstdint>
-
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace strings_internal {
-// Calculate DamerauLevenshtein distance between two strings.
-// When the distance is larger than cutoff, the code just returns cutoff + 1.
-uint8_t CappedDamerauLevenshteinDistance(absl::string_view s1,
-                                         absl::string_view s2, uint8_t cutoff);
-
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_DAMERAU_LEVENSHTEIN_DISTANCE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/escaping.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/escaping.h
deleted file mode 100644
index 2186f778..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/escaping.h
+++ /dev/null
@@ -1,57 +0,0 @@
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_ESCAPING_H_
-#define ABSL_STRINGS_INTERNAL_ESCAPING_H_
-
-#include <cassert>
-
-#include "absl/strings/internal/resize_uninitialized.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace strings_internal {
-
-ABSL_CONST_INIT extern const char kBase64Chars[];
-ABSL_CONST_INIT extern const char kWebSafeBase64Chars[];
-
-// Calculates the length of a Base64 encoding (RFC 4648) of a string of length
-// `input_len`, with or without padding per `do_padding`. Note that 'web-safe'
-// encoding (section 5 of the RFC) does not change this length.
-size_t CalculateBase64EscapedLenInternal(size_t input_len, bool do_padding);
-
-// Base64-encodes `src` using the alphabet provided in `base64` (which
-// determines whether to do web-safe encoding or not) and writes the result to
-// `dest`. If `do_padding` is true, `dest` is padded with '=' chars until its
-// length is a multiple of 3. Returns the length of `dest`.
-size_t Base64EscapeInternal(const unsigned char* src, size_t szsrc, char* dest,
-                            size_t szdest, const char* base64, bool do_padding);
-template <typename String>
-void Base64EscapeInternal(const unsigned char* src, size_t szsrc, String* dest,
-                          bool do_padding, const char* base64_chars) {
-  const size_t calc_escaped_size =
-      CalculateBase64EscapedLenInternal(szsrc, do_padding);
-  STLStringResizeUninitialized(dest, calc_escaped_size);
-
-  const size_t escaped_len = Base64EscapeInternal(
-      src, szsrc, &(*dest)[0], dest->size(), base64_chars, do_padding);
-  assert(calc_escaped_size == escaped_len);
-  dest->erase(escaped_len);
-}
-
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_ESCAPING_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/escaping_test_common.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/escaping_test_common.h
deleted file mode 100644
index 7b18017a..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/escaping_test_common.h
+++ /dev/null
@@ -1,133 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// This test contains common things needed by both escaping_test.cc and
-// escaping_benchmark.cc.
-
-#ifndef ABSL_STRINGS_INTERNAL_ESCAPING_TEST_COMMON_H_
-#define ABSL_STRINGS_INTERNAL_ESCAPING_TEST_COMMON_H_
-
-#include <array>
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace strings_internal {
-
-struct base64_testcase {
-  absl::string_view plaintext;
-  absl::string_view cyphertext;
-};
-
-inline const std::array<base64_testcase, 5>& base64_strings() {
-  static const std::array<base64_testcase, 5> testcase{{
-      // Some google quotes
-      // Cyphertext created with "uuencode (GNU sharutils) 4.6.3"
-      // (Note that we're testing the websafe encoding, though, so if
-      // you add messages, be sure to run "tr -- '+/' '-_'" on the output)
-      { "I was always good at math and science, and I never realized "
-        "that was unusual or somehow undesirable. So one of the things "
-        "I care a lot about is helping to remove that stigma, "
-        "to show girls that you can be feminine, you can like the things "
-        "that girls like, but you can also be really good at technology. "
-        "You can be really good at building things."
-        " - Marissa Meyer, Newsweek, 2010-12-22" "\n",
-
-        "SSB3YXMgYWx3YXlzIGdvb2QgYXQgbWF0aCBhbmQgc2NpZW5jZSwgYW5kIEkg"
-        "bmV2ZXIgcmVhbGl6ZWQgdGhhdCB3YXMgdW51c3VhbCBvciBzb21laG93IHVu"
-        "ZGVzaXJhYmxlLiBTbyBvbmUgb2YgdGhlIHRoaW5ncyBJIGNhcmUgYSBsb3Qg"
-        "YWJvdXQgaXMgaGVscGluZyB0byByZW1vdmUgdGhhdCBzdGlnbWEsIHRvIHNo"
-        "b3cgZ2lybHMgdGhhdCB5b3UgY2FuIGJlIGZlbWluaW5lLCB5b3UgY2FuIGxp"
-        "a2UgdGhlIHRoaW5ncyB0aGF0IGdpcmxzIGxpa2UsIGJ1dCB5b3UgY2FuIGFs"
-        "c28gYmUgcmVhbGx5IGdvb2QgYXQgdGVjaG5vbG9neS4gWW91IGNhbiBiZSBy"
-        "ZWFsbHkgZ29vZCBhdCBidWlsZGluZyB0aGluZ3MuIC0gTWFyaXNzYSBNZXll"
-        "ciwgTmV3c3dlZWssIDIwMTAtMTItMjIK" },
-
-      { "Typical first year for a new cluster: "
-        "~0.5 overheating "
-        "~1 PDU failure "
-        "~1 rack-move "
-        "~1 network rewiring "
-        "~20 rack failures "
-        "~5 racks go wonky "
-        "~8 network maintenances "
-        "~12 router reloads "
-        "~3 router failures "
-        "~dozens of minor 30-second blips for dns "
-        "~1000 individual machine failures "
-        "~thousands of hard drive failures "
-        "slow disks, bad memory, misconfigured machines, flaky machines, etc."
-        " - Jeff Dean, The Joys of Real Hardware" "\n",
-
-        "VHlwaWNhbCBmaXJzdCB5ZWFyIGZvciBhIG5ldyBjbHVzdGVyOiB-MC41IG92"
-        "ZXJoZWF0aW5nIH4xIFBEVSBmYWlsdXJlIH4xIHJhY2stbW92ZSB-MSBuZXR3"
-        "b3JrIHJld2lyaW5nIH4yMCByYWNrIGZhaWx1cmVzIH41IHJhY2tzIGdvIHdv"
-        "bmt5IH44IG5ldHdvcmsgbWFpbnRlbmFuY2VzIH4xMiByb3V0ZXIgcmVsb2Fk"
-        "cyB-MyByb3V0ZXIgZmFpbHVyZXMgfmRvemVucyBvZiBtaW5vciAzMC1zZWNv"
-        "bmQgYmxpcHMgZm9yIGRucyB-MTAwMCBpbmRpdmlkdWFsIG1hY2hpbmUgZmFp"
-        "bHVyZXMgfnRob3VzYW5kcyBvZiBoYXJkIGRyaXZlIGZhaWx1cmVzIHNsb3cg"
-        "ZGlza3MsIGJhZCBtZW1vcnksIG1pc2NvbmZpZ3VyZWQgbWFjaGluZXMsIGZs"
-        "YWt5IG1hY2hpbmVzLCBldGMuIC0gSmVmZiBEZWFuLCBUaGUgSm95cyBvZiBS"
-        "ZWFsIEhhcmR3YXJlCg" },
-
-      { "I'm the head of the webspam team at Google.  "
-        "That means that if you type your name into Google and get porn back, "
-        "it's my fault. Unless you're a porn star, in which case porn is a "
-        "completely reasonable response."
-        " - Matt Cutts, Google Plus" "\n",
-
-        "SSdtIHRoZSBoZWFkIG9mIHRoZSB3ZWJzcGFtIHRlYW0gYXQgR29vZ2xlLiAg"
-        "VGhhdCBtZWFucyB0aGF0IGlmIHlvdSB0eXBlIHlvdXIgbmFtZSBpbnRvIEdv"
-        "b2dsZSBhbmQgZ2V0IHBvcm4gYmFjaywgaXQncyBteSBmYXVsdC4gVW5sZXNz"
-        "IHlvdSdyZSBhIHBvcm4gc3RhciwgaW4gd2hpY2ggY2FzZSBwb3JuIGlzIGEg"
-        "Y29tcGxldGVseSByZWFzb25hYmxlIHJlc3BvbnNlLiAtIE1hdHQgQ3V0dHMs"
-        "IEdvb2dsZSBQbHVzCg" },
-
-      { "It will still be a long time before machines approach human "
-        "intelligence. "
-        "But luckily, machines don't actually have to be intelligent; "
-        "they just have to fake it. Access to a wealth of information, "
-        "combined with a rudimentary decision-making capacity, "
-        "can often be almost as useful. Of course, the results are better yet "
-        "when coupled with intelligence. A reference librarian with access to "
-        "a good search engine is a formidable tool."
-        " - Craig Silverstein, Siemens Pictures of the Future, Spring 2004"
-        "\n",
-
-        "SXQgd2lsbCBzdGlsbCBiZSBhIGxvbmcgdGltZSBiZWZvcmUgbWFjaGluZXMg"
-        "YXBwcm9hY2ggaHVtYW4gaW50ZWxsaWdlbmNlLiBCdXQgbHVja2lseSwgbWFj"
-        "aGluZXMgZG9uJ3QgYWN0dWFsbHkgaGF2ZSB0byBiZSBpbnRlbGxpZ2VudDsg"
-        "dGhleSBqdXN0IGhhdmUgdG8gZmFrZSBpdC4gQWNjZXNzIHRvIGEgd2VhbHRo"
-        "IG9mIGluZm9ybWF0aW9uLCBjb21iaW5lZCB3aXRoIGEgcnVkaW1lbnRhcnkg"
-        "ZGVjaXNpb24tbWFraW5nIGNhcGFjaXR5LCBjYW4gb2Z0ZW4gYmUgYWxtb3N0"
-        "IGFzIHVzZWZ1bC4gT2YgY291cnNlLCB0aGUgcmVzdWx0cyBhcmUgYmV0dGVy"
-        "IHlldCB3aGVuIGNvdXBsZWQgd2l0aCBpbnRlbGxpZ2VuY2UuIEEgcmVmZXJl"
-        "bmNlIGxpYnJhcmlhbiB3aXRoIGFjY2VzcyB0byBhIGdvb2Qgc2VhcmNoIGVu"
-        "Z2luZSBpcyBhIGZvcm1pZGFibGUgdG9vbC4gLSBDcmFpZyBTaWx2ZXJzdGVp"
-        "biwgU2llbWVucyBQaWN0dXJlcyBvZiB0aGUgRnV0dXJlLCBTcHJpbmcgMjAw"
-        "NAo" },
-
-      // Degenerate edge case
-      { "",
-        "" },
-  }};
-
-  return testcase;
-}
-
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_ESCAPING_TEST_COMMON_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/has_absl_stringify.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/has_absl_stringify.h
deleted file mode 100644
index f82cfe26..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/has_absl_stringify.h
+++ /dev/null
@@ -1,44 +0,0 @@
-// Copyright 2024 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_HAS_ABSL_STRINGIFY_H_
-#define ABSL_STRINGS_INTERNAL_HAS_ABSL_STRINGIFY_H_
-
-#include "absl/strings/has_absl_stringify.h"
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-namespace strings_internal {
-
-// This exists to fix a circular dependency problem with the GoogleTest release.
-// GoogleTest referenced this internal file and this internal trait.  Since
-// simultaneous releases are not possible since once release must reference
-// another, we will temporarily add this back.
-// https://github.com/google/googletest/blob/v1.14.x/googletest/include/gtest/gtest-printers.h#L119
-//
-// This file can be deleted after the next Abseil and GoogleTest release.
-//
-// https://github.com/google/googletest/pull/4368#issuecomment-1717699895
-// https://github.com/google/googletest/pull/4368#issuecomment-1717699895
-using ::absl::HasAbslStringify;
-
-}  // namespace strings_internal
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_HAS_ABSL_STRINGIFY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/memutil.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/memutil.h
deleted file mode 100644
index b5911a01..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/memutil.h
+++ /dev/null
@@ -1,40 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_STRINGS_INTERNAL_MEMUTIL_H_
-#define ABSL_STRINGS_INTERNAL_MEMUTIL_H_
-
-#include <cstddef>
-#include <cstring>
-
-#include "absl/base/port.h"  // disable some warnings on Windows
-#include "absl/strings/ascii.h"  // for absl::ascii_tolower
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace strings_internal {
-
-// Performs a byte-by-byte comparison of `len` bytes of the strings `s1` and
-// `s2`, ignoring the case of the characters. It returns an integer less than,
-// equal to, or greater than zero if `s1` is found, respectively, to be less
-// than, to match, or be greater than `s2`.
-int memcasecmp(const char* s1, const char* s2, size_t len);
-
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_MEMUTIL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/numbers_test_common.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/numbers_test_common.h
deleted file mode 100644
index eaa88a88..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/numbers_test_common.h
+++ /dev/null
@@ -1,184 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// This file contains common things needed by numbers_test.cc,
-// numbers_legacy_test.cc and numbers_benchmark.cc.
-
-#ifndef ABSL_STRINGS_INTERNAL_NUMBERS_TEST_COMMON_H_
-#define ABSL_STRINGS_INTERNAL_NUMBERS_TEST_COMMON_H_
-
-#include <array>
-#include <cstdint>
-#include <limits>
-#include <string>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace strings_internal {
-
-template <typename IntType>
-inline bool Itoa(IntType value, int base, std::string* destination) {
-  destination->clear();
-  if (base <= 1 || base > 36) {
-    return false;
-  }
-
-  if (value == 0) {
-    destination->push_back('0');
-    return true;
-  }
-
-  bool negative = value < 0;
-  while (value != 0) {
-    const IntType next_value = value / base;
-    // Can't use std::abs here because of problems when IntType is unsigned.
-    int remainder =
-        static_cast<int>(value > next_value * base ? value - next_value * base
-                                                   : next_value * base - value);
-    char c = remainder < 10 ? '0' + remainder : 'A' + remainder - 10;
-    destination->insert(0, 1, c);
-    value = next_value;
-  }
-
-  if (negative) {
-    destination->insert(0, 1, '-');
-  }
-  return true;
-}
-
-struct uint32_test_case {
-  const char* str;
-  bool expect_ok;
-  int base;  // base to pass to the conversion function
-  uint32_t expected;
-};
-
-inline const std::array<uint32_test_case, 27>& strtouint32_test_cases() {
-  static const std::array<uint32_test_case, 27> test_cases{{
-      {"0xffffffff", true, 16, (std::numeric_limits<uint32_t>::max)()},
-      {"0x34234324", true, 16, 0x34234324},
-      {"34234324", true, 16, 0x34234324},
-      {"0", true, 16, 0},
-      {" \t\n 0xffffffff", true, 16, (std::numeric_limits<uint32_t>::max)()},
-      {" \f\v 46", true, 10, 46},  // must accept weird whitespace
-      {" \t\n 72717222", true, 8, 072717222},
-      {" \t\n 072717222", true, 8, 072717222},
-      {" \t\n 072717228", false, 8, 07271722},
-      {"0", true, 0, 0},
-
-      // Base-10 version.
-      {"34234324", true, 0, 34234324},
-      {"4294967295", true, 0, (std::numeric_limits<uint32_t>::max)()},
-      {"34234324 \n\t", true, 10, 34234324},
-
-      // Unusual base
-      {"0", true, 3, 0},
-      {"2", true, 3, 2},
-      {"11", true, 3, 4},
-
-      // Invalid uints.
-      {"", false, 0, 0},
-      {"  ", false, 0, 0},
-      {"abc", false, 0, 0},  // would be valid hex, but prefix is missing
-      {"34234324a", false, 0, 34234324},
-      {"34234.3", false, 0, 34234},
-      {"-1", false, 0, 0},
-      {"   -123", false, 0, 0},
-      {" \t\n -123", false, 0, 0},
-
-      // Out of bounds.
-      {"4294967296", false, 0, (std::numeric_limits<uint32_t>::max)()},
-      {"0x100000000", false, 0, (std::numeric_limits<uint32_t>::max)()},
-      {nullptr, false, 0, 0},
-  }};
-  return test_cases;
-}
-
-struct uint64_test_case {
-  const char* str;
-  bool expect_ok;
-  int base;
-  uint64_t expected;
-};
-
-inline const std::array<uint64_test_case, 34>& strtouint64_test_cases() {
-  static const std::array<uint64_test_case, 34> test_cases{{
-      {"0x3423432448783446", true, 16, int64_t{0x3423432448783446}},
-      {"3423432448783446", true, 16, int64_t{0x3423432448783446}},
-
-      {"0", true, 16, 0},
-      {"000", true, 0, 0},
-      {"0", true, 0, 0},
-      {" \t\n 0xffffffffffffffff", true, 16,
-       (std::numeric_limits<uint64_t>::max)()},
-
-      {"012345670123456701234", true, 8, int64_t{012345670123456701234}},
-      {"12345670123456701234", true, 8, int64_t{012345670123456701234}},
-
-      {"12845670123456701234", false, 8, 0},
-
-      // Base-10 version.
-      {"34234324487834466", true, 0, int64_t{34234324487834466}},
-
-      {" \t\n 18446744073709551615", true, 0,
-       (std::numeric_limits<uint64_t>::max)()},
-
-      {"34234324487834466 \n\t ", true, 0, int64_t{34234324487834466}},
-
-      {" \f\v 46", true, 10, 46},  // must accept weird whitespace
-
-      // Unusual base
-      {"0", true, 3, 0},
-      {"2", true, 3, 2},
-      {"11", true, 3, 4},
-
-      {"0", true, 0, 0},
-
-      // Invalid uints.
-      {"", false, 0, 0},
-      {"  ", false, 0, 0},
-      {"abc", false, 0, 0},
-      {"34234324487834466a", false, 0, 0},
-      {"34234487834466.3", false, 0, 0},
-      {"-1", false, 0, 0},
-      {"   -123", false, 0, 0},
-      {" \t\n -123", false, 0, 0},
-
-      // Out of bounds.
-      {"18446744073709551616", false, 10, 0},
-      {"18446744073709551616", false, 0, 0},
-      {"0x10000000000000000", false, 16,
-       (std::numeric_limits<uint64_t>::max)()},
-      {"0X10000000000000000", false, 16,
-       (std::numeric_limits<uint64_t>::max)()},  // 0X versus 0x.
-      {"0x10000000000000000", false, 0, (std::numeric_limits<uint64_t>::max)()},
-      {"0X10000000000000000", false, 0,
-       (std::numeric_limits<uint64_t>::max)()},  // 0X versus 0x.
-
-      {"0x1234", true, 16, 0x1234},
-
-      // Base-10 string version.
-      {"1234", true, 0, 1234},
-      {nullptr, false, 0, 0},
-  }};
-  return test_cases;
-}
-
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_NUMBERS_TEST_COMMON_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/ostringstream.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/ostringstream.h
deleted file mode 100644
index c0e237db..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/ostringstream.h
+++ /dev/null
@@ -1,114 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_OSTRINGSTREAM_H_
-#define ABSL_STRINGS_INTERNAL_OSTRINGSTREAM_H_
-
-#include <cassert>
-#include <ios>
-#include <ostream>
-#include <streambuf>
-#include <string>
-#include <utility>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace strings_internal {
-
-// The same as std::ostringstream but appends to a user-specified std::string,
-// and is faster. It is ~70% faster to create, ~50% faster to write to, and
-// completely free to extract the result std::string.
-//
-//   std::string s;
-//   OStringStream strm(&s);
-//   strm << 42 << ' ' << 3.14;  // appends to `s`
-//
-// The stream object doesn't have to be named. Starting from C++11 operator<<
-// works with rvalues of std::ostream.
-//
-//   std::string s;
-//   OStringStream(&s) << 42 << ' ' << 3.14;  // appends to `s`
-//
-// OStringStream is faster to create than std::ostringstream but it's still
-// relatively slow. Avoid creating multiple streams where a single stream will
-// do.
-//
-// Creates unnecessary instances of OStringStream: slow.
-//
-//   std::string s;
-//   OStringStream(&s) << 42;
-//   OStringStream(&s) << ' ';
-//   OStringStream(&s) << 3.14;
-//
-// Creates a single instance of OStringStream and reuses it: fast.
-//
-//   std::string s;
-//   OStringStream strm(&s);
-//   strm << 42;
-//   strm << ' ';
-//   strm << 3.14;
-//
-// Note: flush() has no effect. No reason to call it.
-class OStringStream final : public std::ostream {
- public:
-  // The argument can be null, in which case you'll need to call str(p) with a
-  // non-null argument before you can write to the stream.
-  //
-  // The destructor of OStringStream doesn't use the std::string. It's OK to
-  // destroy the std::string before the stream.
-  explicit OStringStream(std::string* str)
-      : std::ostream(&buf_), buf_(str) {}
-  OStringStream(OStringStream&& that)
-      : std::ostream(std::move(static_cast<std::ostream&>(that))),
-        buf_(that.buf_) {
-    rdbuf(&buf_);
-  }
-  OStringStream& operator=(OStringStream&& that) {
-    std::ostream::operator=(std::move(static_cast<std::ostream&>(that)));
-    buf_ = that.buf_;
-    rdbuf(&buf_);
-    return *this;
-  }
-
-  std::string* str() { return buf_.str(); }
-  const std::string* str() const { return buf_.str(); }
-  void str(std::string* str) { buf_.str(str); }
-
- private:
-  class Streambuf final : public std::streambuf {
-   public:
-    explicit Streambuf(std::string* str) : str_(str) {}
-    Streambuf(const Streambuf&) = default;
-    Streambuf& operator=(const Streambuf&) = default;
-
-    std::string* str() { return str_; }
-    const std::string* str() const { return str_; }
-    void str(std::string* str) { str_ = str; }
-
-   protected:
-    int_type overflow(int c) override;
-    std::streamsize xsputn(const char* s, std::streamsize n) override;
-
-   private:
-    std::string* str_;
-  } buf_;
-};
-
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_OSTRINGSTREAM_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/pow10_helper.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/pow10_helper.h
deleted file mode 100644
index c37c2c3f..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/pow10_helper.h
+++ /dev/null
@@ -1,40 +0,0 @@
-//
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// This test helper library contains a table of powers of 10, to guarantee
-// precise values are computed across the full range of doubles. We can't rely
-// on the pow() function, because not all standard libraries ship a version
-// that is precise.
-#ifndef ABSL_STRINGS_INTERNAL_POW10_HELPER_H_
-#define ABSL_STRINGS_INTERNAL_POW10_HELPER_H_
-
-#include <vector>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace strings_internal {
-
-// Computes the precise value of 10^exp. (I.e. the nearest representable
-// double to the exact value, rounding to nearest-even in the (single) case of
-// being exactly halfway between.)
-double Pow10(int exp);
-
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_POW10_HELPER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/resize_uninitialized.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/resize_uninitialized.h
deleted file mode 100644
index 49859dcc..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/resize_uninitialized.h
+++ /dev/null
@@ -1,119 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_STRINGS_INTERNAL_RESIZE_UNINITIALIZED_H_
-#define ABSL_STRINGS_INTERNAL_RESIZE_UNINITIALIZED_H_
-
-#include <algorithm>
-#include <string>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/port.h"
-#include "absl/meta/type_traits.h"  //  for void_t
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace strings_internal {
-
-// In this type trait, we look for a __resize_default_init member function, and
-// we use it if available, otherwise, we use resize. We provide HasMember to
-// indicate whether __resize_default_init is present.
-template <typename string_type, typename = void>
-struct ResizeUninitializedTraits {
-  using HasMember = std::false_type;
-  static void Resize(string_type* s, size_t new_size) { s->resize(new_size); }
-};
-
-// __resize_default_init is provided by libc++ >= 8.0
-template <typename string_type>
-struct ResizeUninitializedTraits<
-    string_type, absl::void_t<decltype(std::declval<string_type&>()
-                                           .__resize_default_init(237))> > {
-  using HasMember = std::true_type;
-  static void Resize(string_type* s, size_t new_size) {
-    s->__resize_default_init(new_size);
-  }
-};
-
-// Returns true if the std::string implementation supports a resize where
-// the new characters added to the std::string are left untouched.
-//
-// (A better name might be "STLStringSupportsUninitializedResize", alluding to
-// the previous function.)
-template <typename string_type>
-inline constexpr bool STLStringSupportsNontrashingResize(string_type*) {
-  return ResizeUninitializedTraits<string_type>::HasMember::value;
-}
-
-// Like str->resize(new_size), except any new characters added to "*str" as a
-// result of resizing may be left uninitialized, rather than being filled with
-// '0' bytes. Typically used when code is then going to overwrite the backing
-// store of the std::string with known data.
-template <typename string_type, typename = void>
-inline void STLStringResizeUninitialized(string_type* s, size_t new_size) {
-  ResizeUninitializedTraits<string_type>::Resize(s, new_size);
-}
-
-// Used to ensure exponential growth so that the amortized complexity of
-// increasing the string size by a small amount is O(1), in contrast to
-// O(str->size()) in the case of precise growth.
-template <typename string_type>
-void STLStringReserveAmortized(string_type* s, size_t new_size) {
-  const size_t cap = s->capacity();
-  if (new_size > cap) {
-    // Make sure to always grow by at least a factor of 2x.
-    s->reserve((std::max)(new_size, 2 * cap));
-  }
-}
-
-// In this type trait, we look for an __append_default_init member function, and
-// we use it if available, otherwise, we use append.
-template <typename string_type, typename = void>
-struct AppendUninitializedTraits {
-  static void Append(string_type* s, size_t n) {
-    s->append(n, typename string_type::value_type());
-  }
-};
-
-template <typename string_type>
-struct AppendUninitializedTraits<
-    string_type, absl::void_t<decltype(std::declval<string_type&>()
-                                           .__append_default_init(237))> > {
-  static void Append(string_type* s, size_t n) {
-    s->__append_default_init(n);
-  }
-};
-
-// Like STLStringResizeUninitialized(str, new_size), except guaranteed to use
-// exponential growth so that the amortized complexity of increasing the string
-// size by a small amount is O(1), in contrast to O(str->size()) in the case of
-// precise growth.
-template <typename string_type>
-void STLStringResizeUninitializedAmortized(string_type* s, size_t new_size) {
-  const size_t size = s->size();
-  if (new_size > size) {
-    AppendUninitializedTraits<string_type>::Append(s, new_size - size);
-  } else {
-    s->erase(new_size);
-  }
-}
-
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_RESIZE_UNINITIALIZED_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/stl_type_traits.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/stl_type_traits.h
deleted file mode 100644
index e50468b0..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/stl_type_traits.h
+++ /dev/null
@@ -1,248 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-// The file provides the IsStrictlyBaseOfAndConvertibleToSTLContainer type
-// trait metafunction to assist in working with the _GLIBCXX_DEBUG debug
-// wrappers of STL containers.
-//
-// DO NOT INCLUDE THIS FILE DIRECTLY. Use this file by including
-// absl/strings/str_split.h.
-//
-// IWYU pragma: private, include "absl/strings/str_split.h"
-
-#ifndef ABSL_STRINGS_INTERNAL_STL_TYPE_TRAITS_H_
-#define ABSL_STRINGS_INTERNAL_STL_TYPE_TRAITS_H_
-
-#include <array>
-#include <bitset>
-#include <deque>
-#include <forward_list>
-#include <list>
-#include <map>
-#include <set>
-#include <type_traits>
-#include <unordered_map>
-#include <unordered_set>
-#include <vector>
-
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace strings_internal {
-
-template <typename C, template <typename...> class T>
-struct IsSpecializationImpl : std::false_type {};
-template <template <typename...> class T, typename... Args>
-struct IsSpecializationImpl<T<Args...>, T> : std::true_type {};
-template <typename C, template <typename...> class T>
-using IsSpecialization = IsSpecializationImpl<absl::decay_t<C>, T>;
-
-template <typename C>
-struct IsArrayImpl : std::false_type {};
-template <template <typename, size_t> class A, typename T, size_t N>
-struct IsArrayImpl<A<T, N>> : std::is_same<A<T, N>, std::array<T, N>> {};
-template <typename C>
-using IsArray = IsArrayImpl<absl::decay_t<C>>;
-
-template <typename C>
-struct IsBitsetImpl : std::false_type {};
-template <template <size_t> class B, size_t N>
-struct IsBitsetImpl<B<N>> : std::is_same<B<N>, std::bitset<N>> {};
-template <typename C>
-using IsBitset = IsBitsetImpl<absl::decay_t<C>>;
-
-template <typename C>
-struct IsSTLContainer
-    : absl::disjunction<
-          IsArray<C>, IsBitset<C>, IsSpecialization<C, std::deque>,
-          IsSpecialization<C, std::forward_list>,
-          IsSpecialization<C, std::list>, IsSpecialization<C, std::map>,
-          IsSpecialization<C, std::multimap>, IsSpecialization<C, std::set>,
-          IsSpecialization<C, std::multiset>,
-          IsSpecialization<C, std::unordered_map>,
-          IsSpecialization<C, std::unordered_multimap>,
-          IsSpecialization<C, std::unordered_set>,
-          IsSpecialization<C, std::unordered_multiset>,
-          IsSpecialization<C, std::vector>> {};
-
-template <typename C, template <typename...> class T, typename = void>
-struct IsBaseOfSpecializationImpl : std::false_type {};
-// IsBaseOfSpecializationImpl needs multiple partial specializations to SFINAE
-// on the existence of container dependent types and plug them into the STL
-// template.
-template <typename C, template <typename, typename> class T>
-struct IsBaseOfSpecializationImpl<
-    C, T, absl::void_t<typename C::value_type, typename C::allocator_type>>
-    : std::is_base_of<C,
-                      T<typename C::value_type, typename C::allocator_type>> {};
-template <typename C, template <typename, typename, typename> class T>
-struct IsBaseOfSpecializationImpl<
-    C, T,
-    absl::void_t<typename C::key_type, typename C::key_compare,
-                 typename C::allocator_type>>
-    : std::is_base_of<C, T<typename C::key_type, typename C::key_compare,
-                           typename C::allocator_type>> {};
-template <typename C, template <typename, typename, typename, typename> class T>
-struct IsBaseOfSpecializationImpl<
-    C, T,
-    absl::void_t<typename C::key_type, typename C::mapped_type,
-                 typename C::key_compare, typename C::allocator_type>>
-    : std::is_base_of<C,
-                      T<typename C::key_type, typename C::mapped_type,
-                        typename C::key_compare, typename C::allocator_type>> {
-};
-template <typename C, template <typename, typename, typename, typename> class T>
-struct IsBaseOfSpecializationImpl<
-    C, T,
-    absl::void_t<typename C::key_type, typename C::hasher,
-                 typename C::key_equal, typename C::allocator_type>>
-    : std::is_base_of<C, T<typename C::key_type, typename C::hasher,
-                           typename C::key_equal, typename C::allocator_type>> {
-};
-template <typename C,
-          template <typename, typename, typename, typename, typename> class T>
-struct IsBaseOfSpecializationImpl<
-    C, T,
-    absl::void_t<typename C::key_type, typename C::mapped_type,
-                 typename C::hasher, typename C::key_equal,
-                 typename C::allocator_type>>
-    : std::is_base_of<C, T<typename C::key_type, typename C::mapped_type,
-                           typename C::hasher, typename C::key_equal,
-                           typename C::allocator_type>> {};
-template <typename C, template <typename...> class T>
-using IsBaseOfSpecialization = IsBaseOfSpecializationImpl<absl::decay_t<C>, T>;
-
-template <typename C>
-struct IsBaseOfArrayImpl : std::false_type {};
-template <template <typename, size_t> class A, typename T, size_t N>
-struct IsBaseOfArrayImpl<A<T, N>> : std::is_base_of<A<T, N>, std::array<T, N>> {
-};
-template <typename C>
-using IsBaseOfArray = IsBaseOfArrayImpl<absl::decay_t<C>>;
-
-template <typename C>
-struct IsBaseOfBitsetImpl : std::false_type {};
-template <template <size_t> class B, size_t N>
-struct IsBaseOfBitsetImpl<B<N>> : std::is_base_of<B<N>, std::bitset<N>> {};
-template <typename C>
-using IsBaseOfBitset = IsBaseOfBitsetImpl<absl::decay_t<C>>;
-
-template <typename C>
-struct IsBaseOfSTLContainer
-    : absl::disjunction<IsBaseOfArray<C>, IsBaseOfBitset<C>,
-                        IsBaseOfSpecialization<C, std::deque>,
-                        IsBaseOfSpecialization<C, std::forward_list>,
-                        IsBaseOfSpecialization<C, std::list>,
-                        IsBaseOfSpecialization<C, std::map>,
-                        IsBaseOfSpecialization<C, std::multimap>,
-                        IsBaseOfSpecialization<C, std::set>,
-                        IsBaseOfSpecialization<C, std::multiset>,
-                        IsBaseOfSpecialization<C, std::unordered_map>,
-                        IsBaseOfSpecialization<C, std::unordered_multimap>,
-                        IsBaseOfSpecialization<C, std::unordered_set>,
-                        IsBaseOfSpecialization<C, std::unordered_multiset>,
-                        IsBaseOfSpecialization<C, std::vector>> {};
-
-template <typename C, template <typename...> class T, typename = void>
-struct IsConvertibleToSpecializationImpl : std::false_type {};
-// IsConvertibleToSpecializationImpl needs multiple partial specializations to
-// SFINAE on the existence of container dependent types and plug them into the
-// STL template.
-template <typename C, template <typename, typename> class T>
-struct IsConvertibleToSpecializationImpl<
-    C, T, absl::void_t<typename C::value_type, typename C::allocator_type>>
-    : std::is_convertible<
-          C, T<typename C::value_type, typename C::allocator_type>> {};
-template <typename C, template <typename, typename, typename> class T>
-struct IsConvertibleToSpecializationImpl<
-    C, T,
-    absl::void_t<typename C::key_type, typename C::key_compare,
-                 typename C::allocator_type>>
-    : std::is_convertible<C, T<typename C::key_type, typename C::key_compare,
-                               typename C::allocator_type>> {};
-template <typename C, template <typename, typename, typename, typename> class T>
-struct IsConvertibleToSpecializationImpl<
-    C, T,
-    absl::void_t<typename C::key_type, typename C::mapped_type,
-                 typename C::key_compare, typename C::allocator_type>>
-    : std::is_convertible<
-          C, T<typename C::key_type, typename C::mapped_type,
-               typename C::key_compare, typename C::allocator_type>> {};
-template <typename C, template <typename, typename, typename, typename> class T>
-struct IsConvertibleToSpecializationImpl<
-    C, T,
-    absl::void_t<typename C::key_type, typename C::hasher,
-                 typename C::key_equal, typename C::allocator_type>>
-    : std::is_convertible<
-          C, T<typename C::key_type, typename C::hasher, typename C::key_equal,
-               typename C::allocator_type>> {};
-template <typename C,
-          template <typename, typename, typename, typename, typename> class T>
-struct IsConvertibleToSpecializationImpl<
-    C, T,
-    absl::void_t<typename C::key_type, typename C::mapped_type,
-                 typename C::hasher, typename C::key_equal,
-                 typename C::allocator_type>>
-    : std::is_convertible<C, T<typename C::key_type, typename C::mapped_type,
-                               typename C::hasher, typename C::key_equal,
-                               typename C::allocator_type>> {};
-template <typename C, template <typename...> class T>
-using IsConvertibleToSpecialization =
-    IsConvertibleToSpecializationImpl<absl::decay_t<C>, T>;
-
-template <typename C>
-struct IsConvertibleToArrayImpl : std::false_type {};
-template <template <typename, size_t> class A, typename T, size_t N>
-struct IsConvertibleToArrayImpl<A<T, N>>
-    : std::is_convertible<A<T, N>, std::array<T, N>> {};
-template <typename C>
-using IsConvertibleToArray = IsConvertibleToArrayImpl<absl::decay_t<C>>;
-
-template <typename C>
-struct IsConvertibleToBitsetImpl : std::false_type {};
-template <template <size_t> class B, size_t N>
-struct IsConvertibleToBitsetImpl<B<N>>
-    : std::is_convertible<B<N>, std::bitset<N>> {};
-template <typename C>
-using IsConvertibleToBitset = IsConvertibleToBitsetImpl<absl::decay_t<C>>;
-
-template <typename C>
-struct IsConvertibleToSTLContainer
-    : absl::disjunction<
-          IsConvertibleToArray<C>, IsConvertibleToBitset<C>,
-          IsConvertibleToSpecialization<C, std::deque>,
-          IsConvertibleToSpecialization<C, std::forward_list>,
-          IsConvertibleToSpecialization<C, std::list>,
-          IsConvertibleToSpecialization<C, std::map>,
-          IsConvertibleToSpecialization<C, std::multimap>,
-          IsConvertibleToSpecialization<C, std::set>,
-          IsConvertibleToSpecialization<C, std::multiset>,
-          IsConvertibleToSpecialization<C, std::unordered_map>,
-          IsConvertibleToSpecialization<C, std::unordered_multimap>,
-          IsConvertibleToSpecialization<C, std::unordered_set>,
-          IsConvertibleToSpecialization<C, std::unordered_multiset>,
-          IsConvertibleToSpecialization<C, std::vector>> {};
-
-template <typename C>
-struct IsStrictlyBaseOfAndConvertibleToSTLContainer
-    : absl::conjunction<absl::negation<IsSTLContainer<C>>,
-                        IsBaseOfSTLContainer<C>,
-                        IsConvertibleToSTLContainer<C>> {};
-
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-#endif  // ABSL_STRINGS_INTERNAL_STL_TYPE_TRAITS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/arg.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/arg.h
deleted file mode 100644
index 309161d5..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/arg.h
+++ /dev/null
@@ -1,671 +0,0 @@
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_STR_FORMAT_ARG_H_
-#define ABSL_STRINGS_INTERNAL_STR_FORMAT_ARG_H_
-
-#include <string.h>
-#include <wchar.h>
-
-#include <algorithm>
-#include <cstddef>
-#include <cstdint>
-#include <cstdio>
-#include <limits>
-#include <memory>
-#include <sstream>
-#include <string>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/config.h"
-#include "absl/base/optimization.h"
-#include "absl/meta/type_traits.h"
-#include "absl/numeric/int128.h"
-#include "absl/strings/has_absl_stringify.h"
-#include "absl/strings/internal/str_format/extension.h"
-#include "absl/strings/string_view.h"
-
-#if defined(ABSL_HAVE_STD_STRING_VIEW)
-#include <string_view>
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-class Cord;
-class FormatCountCapture;
-class FormatSink;
-
-template <absl::FormatConversionCharSet C>
-struct FormatConvertResult;
-class FormatConversionSpec;
-
-namespace str_format_internal {
-
-template <FormatConversionCharSet C>
-struct ArgConvertResult {
-  bool value;
-};
-
-using IntegralConvertResult = ArgConvertResult<FormatConversionCharSetUnion(
-    FormatConversionCharSetInternal::c,
-    FormatConversionCharSetInternal::kNumeric,
-    FormatConversionCharSetInternal::kStar,
-    FormatConversionCharSetInternal::v)>;
-using FloatingConvertResult = ArgConvertResult<FormatConversionCharSetUnion(
-    FormatConversionCharSetInternal::kFloating,
-    FormatConversionCharSetInternal::v)>;
-using CharConvertResult = ArgConvertResult<FormatConversionCharSetUnion(
-    FormatConversionCharSetInternal::c,
-    FormatConversionCharSetInternal::kNumeric,
-    FormatConversionCharSetInternal::kStar)>;
-
-template <typename T, typename = void>
-struct HasUserDefinedConvert : std::false_type {};
-
-template <typename T>
-struct HasUserDefinedConvert<T, void_t<decltype(AbslFormatConvert(
-                                    std::declval<const T&>(),
-                                    std::declval<const FormatConversionSpec&>(),
-                                    std::declval<FormatSink*>()))>>
-    : std::true_type {};
-
-// These declarations prevent ADL lookup from continuing in absl namespaces,
-// we are deliberately using these as ADL hooks and want them to consider
-// non-absl namespaces only.
-void AbslFormatConvert();
-void AbslStringify();
-
-template <typename T>
-bool ConvertIntArg(T v, FormatConversionSpecImpl conv, FormatSinkImpl* sink);
-
-// Forward declarations of internal `ConvertIntArg` function template
-// instantiations are here to avoid including the template body in the headers
-// and instantiating it in large numbers of translation units. Explicit
-// instantiations can be found in "absl/strings/internal/str_format/arg.cc"
-extern template bool ConvertIntArg<char>(char v, FormatConversionSpecImpl conv,
-                                         FormatSinkImpl* sink);
-extern template bool ConvertIntArg<signed char>(signed char v,
-                                                FormatConversionSpecImpl conv,
-                                                FormatSinkImpl* sink);
-extern template bool ConvertIntArg<unsigned char>(unsigned char v,
-                                                  FormatConversionSpecImpl conv,
-                                                  FormatSinkImpl* sink);
-extern template bool ConvertIntArg<wchar_t>(wchar_t v,
-                                            FormatConversionSpecImpl conv,
-                                            FormatSinkImpl* sink);
-extern template bool ConvertIntArg<short>(short v,  // NOLINT
-                                          FormatConversionSpecImpl conv,
-                                          FormatSinkImpl* sink);
-extern template bool ConvertIntArg<unsigned short>(   // NOLINT
-    unsigned short v, FormatConversionSpecImpl conv,  // NOLINT
-    FormatSinkImpl* sink);
-extern template bool ConvertIntArg<int>(int v, FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-extern template bool ConvertIntArg<unsigned int>(unsigned int v,
-                                                 FormatConversionSpecImpl conv,
-                                                 FormatSinkImpl* sink);
-extern template bool ConvertIntArg<long>(                           // NOLINT
-    long v, FormatConversionSpecImpl conv, FormatSinkImpl* sink);   // NOLINT
-extern template bool ConvertIntArg<unsigned long>(unsigned long v,  // NOLINT
-                                                  FormatConversionSpecImpl conv,
-                                                  FormatSinkImpl* sink);
-extern template bool ConvertIntArg<long long>(long long v,  // NOLINT
-                                              FormatConversionSpecImpl conv,
-                                              FormatSinkImpl* sink);
-extern template bool ConvertIntArg<unsigned long long>(   // NOLINT
-    unsigned long long v, FormatConversionSpecImpl conv,  // NOLINT
-    FormatSinkImpl* sink);
-
-template <typename T>
-auto FormatConvertImpl(const T& v, FormatConversionSpecImpl conv,
-                       FormatSinkImpl* sink)
-    -> decltype(AbslFormatConvert(v,
-                                  std::declval<const FormatConversionSpec&>(),
-                                  std::declval<FormatSink*>())) {
-  using FormatConversionSpecT =
-      absl::enable_if_t<sizeof(const T& (*)()) != 0, FormatConversionSpec>;
-  using FormatSinkT =
-      absl::enable_if_t<sizeof(const T& (*)()) != 0, FormatSink>;
-  auto fcs = conv.Wrap<FormatConversionSpecT>();
-  auto fs = sink->Wrap<FormatSinkT>();
-  return AbslFormatConvert(v, fcs, &fs);
-}
-
-template <typename T>
-auto FormatConvertImpl(const T& v, FormatConversionSpecImpl conv,
-                       FormatSinkImpl* sink)
-    -> std::enable_if_t<std::is_enum<T>::value &&
-                            std::is_void<decltype(AbslStringify(
-                                std::declval<FormatSink&>(), v))>::value,
-                        IntegralConvertResult> {
-  if (conv.conversion_char() == FormatConversionCharInternal::v) {
-    using FormatSinkT =
-        absl::enable_if_t<sizeof(const T& (*)()) != 0, FormatSink>;
-    auto fs = sink->Wrap<FormatSinkT>();
-    AbslStringify(fs, v);
-    return {true};
-  } else {
-    return {ConvertIntArg(
-        static_cast<typename std::underlying_type<T>::type>(v), conv, sink)};
-  }
-}
-
-template <typename T>
-auto FormatConvertImpl(const T& v, FormatConversionSpecImpl,
-                       FormatSinkImpl* sink)
-    -> std::enable_if_t<!std::is_enum<T>::value &&
-                            !std::is_same<T, absl::Cord>::value &&
-                            std::is_void<decltype(AbslStringify(
-                                std::declval<FormatSink&>(), v))>::value,
-                        ArgConvertResult<FormatConversionCharSetInternal::v>> {
-  using FormatSinkT =
-      absl::enable_if_t<sizeof(const T& (*)()) != 0, FormatSink>;
-  auto fs = sink->Wrap<FormatSinkT>();
-  AbslStringify(fs, v);
-  return {true};
-}
-
-template <typename T>
-class StreamedWrapper;
-
-// If 'v' can be converted (in the printf sense) according to 'conv',
-// then convert it, appending to `sink` and return `true`.
-// Otherwise fail and return `false`.
-
-// AbslFormatConvert(v, conv, sink) is intended to be found by ADL on 'v'
-// as an extension mechanism. These FormatConvertImpl functions are the default
-// implementations.
-// The ADL search is augmented via the 'Sink*' parameter, which also
-// serves as a disambiguator to reject possible unintended 'AbslFormatConvert'
-// functions in the namespaces associated with 'v'.
-
-// Raw pointers.
-struct VoidPtr {
-  VoidPtr() = default;
-  template <typename T,
-            decltype(reinterpret_cast<uintptr_t>(std::declval<T*>())) = 0>
-  VoidPtr(T* ptr)  // NOLINT
-      : value(ptr ? reinterpret_cast<uintptr_t>(ptr) : 0) {}
-  uintptr_t value;
-};
-
-template <FormatConversionCharSet C>
-constexpr FormatConversionCharSet ExtractCharSet(FormatConvertResult<C>) {
-  return C;
-}
-
-template <FormatConversionCharSet C>
-constexpr FormatConversionCharSet ExtractCharSet(ArgConvertResult<C>) {
-  return C;
-}
-
-ArgConvertResult<FormatConversionCharSetInternal::p> FormatConvertImpl(
-    VoidPtr v, FormatConversionSpecImpl conv, FormatSinkImpl* sink);
-
-// Strings.
-using StringConvertResult = ArgConvertResult<FormatConversionCharSetUnion(
-    FormatConversionCharSetInternal::s,
-    FormatConversionCharSetInternal::v)>;
-StringConvertResult FormatConvertImpl(const std::string& v,
-                                      FormatConversionSpecImpl conv,
-                                      FormatSinkImpl* sink);
-StringConvertResult FormatConvertImpl(const std::wstring& v,
-                                      FormatConversionSpecImpl conv,
-                                      FormatSinkImpl* sink);
-StringConvertResult FormatConvertImpl(string_view v,
-                                      FormatConversionSpecImpl conv,
-                                      FormatSinkImpl* sink);
-#if defined(ABSL_HAVE_STD_STRING_VIEW)
-StringConvertResult FormatConvertImpl(std::wstring_view v,
-                                      FormatConversionSpecImpl conv,
-                                      FormatSinkImpl* sink);
-#if !defined(ABSL_USES_STD_STRING_VIEW)
-inline StringConvertResult FormatConvertImpl(std::string_view v,
-                                             FormatConversionSpecImpl conv,
-                                             FormatSinkImpl* sink) {
-  return FormatConvertImpl(absl::string_view(v.data(), v.size()), conv, sink);
-}
-#endif  // !ABSL_USES_STD_STRING_VIEW
-#endif  // ABSL_HAVE_STD_STRING_VIEW
-
-using StringPtrConvertResult = ArgConvertResult<FormatConversionCharSetUnion(
-    FormatConversionCharSetInternal::s,
-    FormatConversionCharSetInternal::p)>;
-StringPtrConvertResult FormatConvertImpl(const char* v,
-                                         FormatConversionSpecImpl conv,
-                                         FormatSinkImpl* sink);
-StringPtrConvertResult FormatConvertImpl(const wchar_t* v,
-                                         FormatConversionSpecImpl conv,
-                                         FormatSinkImpl* sink);
-// This overload is needed to disambiguate, since `nullptr` could match either
-// of the other overloads equally well.
-StringPtrConvertResult FormatConvertImpl(std::nullptr_t,
-                                         FormatConversionSpecImpl conv,
-                                         FormatSinkImpl* sink);
-
-template <class AbslCord, typename std::enable_if<std::is_same<
-                              AbslCord, absl::Cord>::value>::type* = nullptr>
-StringConvertResult FormatConvertImpl(const AbslCord& value,
-                                      FormatConversionSpecImpl conv,
-                                      FormatSinkImpl* sink) {
-  bool is_left = conv.has_left_flag();
-  size_t space_remaining = 0;
-
-  int width = conv.width();
-  if (width >= 0) space_remaining = static_cast<size_t>(width);
-
-  size_t to_write = value.size();
-
-  int precision = conv.precision();
-  if (precision >= 0)
-    to_write = (std::min)(to_write, static_cast<size_t>(precision));
-
-  space_remaining = Excess(to_write, space_remaining);
-
-  if (space_remaining > 0 && !is_left) sink->Append(space_remaining, ' ');
-
-  for (string_view piece : value.Chunks()) {
-    if (piece.size() > to_write) {
-      piece.remove_suffix(piece.size() - to_write);
-      to_write = 0;
-    } else {
-      to_write -= piece.size();
-    }
-    sink->Append(piece);
-    if (to_write == 0) {
-      break;
-    }
-  }
-
-  if (space_remaining > 0 && is_left) sink->Append(space_remaining, ' ');
-  return {true};
-}
-
-bool ConvertBoolArg(bool v, FormatSinkImpl* sink);
-
-// Floats.
-FloatingConvertResult FormatConvertImpl(float v, FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-FloatingConvertResult FormatConvertImpl(double v, FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-FloatingConvertResult FormatConvertImpl(long double v,
-                                        FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-
-// Chars.
-CharConvertResult FormatConvertImpl(char v, FormatConversionSpecImpl conv,
-                                    FormatSinkImpl* sink);
-CharConvertResult FormatConvertImpl(wchar_t v,
-                                    FormatConversionSpecImpl conv,
-                                    FormatSinkImpl* sink);
-
-// Ints.
-IntegralConvertResult FormatConvertImpl(signed char v,
-                                        FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-IntegralConvertResult FormatConvertImpl(unsigned char v,
-                                        FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-IntegralConvertResult FormatConvertImpl(short v,  // NOLINT
-                                        FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-IntegralConvertResult FormatConvertImpl(unsigned short v,  // NOLINT
-                                        FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-IntegralConvertResult FormatConvertImpl(int v, FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-IntegralConvertResult FormatConvertImpl(unsigned v,
-                                        FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-IntegralConvertResult FormatConvertImpl(long v,  // NOLINT
-                                        FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-IntegralConvertResult FormatConvertImpl(unsigned long v,  // NOLINT
-                                        FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-IntegralConvertResult FormatConvertImpl(long long v,  // NOLINT
-                                        FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-IntegralConvertResult FormatConvertImpl(unsigned long long v,  // NOLINT
-                                        FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-IntegralConvertResult FormatConvertImpl(int128 v, FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-IntegralConvertResult FormatConvertImpl(uint128 v,
-                                        FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink);
-
-// This function needs to be a template due to ambiguity regarding type
-// conversions.
-template <typename T, enable_if_t<std::is_same<T, bool>::value, int> = 0>
-IntegralConvertResult FormatConvertImpl(T v, FormatConversionSpecImpl conv,
-                                        FormatSinkImpl* sink) {
-  if (conv.conversion_char() == FormatConversionCharInternal::v) {
-    return {ConvertBoolArg(v, sink)};
-  }
-
-  return FormatConvertImpl(static_cast<int>(v), conv, sink);
-}
-
-// We provide this function to help the checker, but it is never defined.
-// FormatArgImpl will use the underlying Convert functions instead.
-template <typename T>
-typename std::enable_if<std::is_enum<T>::value &&
-                            !HasUserDefinedConvert<T>::value &&
-                            !HasAbslStringify<T>::value,
-                        IntegralConvertResult>::type
-FormatConvertImpl(T v, FormatConversionSpecImpl conv, FormatSinkImpl* sink);
-
-template <typename T>
-StringConvertResult FormatConvertImpl(const StreamedWrapper<T>& v,
-                                      FormatConversionSpecImpl conv,
-                                      FormatSinkImpl* out) {
-  std::ostringstream oss;
-  oss << v.v_;
-  if (!oss) return {false};
-  return str_format_internal::FormatConvertImpl(oss.str(), conv, out);
-}
-
-// Use templates and dependent types to delay evaluation of the function
-// until after FormatCountCapture is fully defined.
-struct FormatCountCaptureHelper {
-  template <class T = int>
-  static ArgConvertResult<FormatConversionCharSetInternal::n> ConvertHelper(
-      const FormatCountCapture& v, FormatConversionSpecImpl conv,
-      FormatSinkImpl* sink) {
-    const absl::enable_if_t<sizeof(T) != 0, FormatCountCapture>& v2 = v;
-
-    if (conv.conversion_char() !=
-        str_format_internal::FormatConversionCharInternal::n) {
-      return {false};
-    }
-    *v2.p_ = static_cast<int>(sink->size());
-    return {true};
-  }
-};
-
-template <class T = int>
-ArgConvertResult<FormatConversionCharSetInternal::n> FormatConvertImpl(
-    const FormatCountCapture& v, FormatConversionSpecImpl conv,
-    FormatSinkImpl* sink) {
-  return FormatCountCaptureHelper::ConvertHelper(v, conv, sink);
-}
-
-// Helper friend struct to hide implementation details from the public API of
-// FormatArgImpl.
-struct FormatArgImplFriend {
-  template <typename Arg>
-  static bool ToInt(Arg arg, int* out) {
-    // A value initialized FormatConversionSpecImpl has a `none` conv, which
-    // tells the dispatcher to run the `int` conversion.
-    return arg.dispatcher_(arg.data_, {}, out);
-  }
-
-  template <typename Arg>
-  static bool Convert(Arg arg, FormatConversionSpecImpl conv,
-                      FormatSinkImpl* out) {
-    return arg.dispatcher_(arg.data_, conv, out);
-  }
-
-  template <typename Arg>
-  static typename Arg::Dispatcher GetVTablePtrForTest(Arg arg) {
-    return arg.dispatcher_;
-  }
-};
-
-template <typename Arg>
-constexpr FormatConversionCharSet ArgumentToConv() {
-  using ConvResult = decltype(str_format_internal::FormatConvertImpl(
-      std::declval<const Arg&>(),
-      std::declval<const FormatConversionSpecImpl&>(),
-      std::declval<FormatSinkImpl*>()));
-  return absl::str_format_internal::ExtractCharSet(ConvResult{});
-}
-
-// A type-erased handle to a format argument.
-class FormatArgImpl {
- private:
-  enum { kInlinedSpace = 8 };
-
-  using VoidPtr = str_format_internal::VoidPtr;
-
-  union Data {
-    const void* ptr;
-    const volatile void* volatile_ptr;
-    char buf[kInlinedSpace];
-  };
-
-  using Dispatcher = bool (*)(Data, FormatConversionSpecImpl, void* out);
-
-  template <typename T>
-  struct store_by_value
-      : std::integral_constant<bool, (sizeof(T) <= kInlinedSpace) &&
-                                         (std::is_integral<T>::value ||
-                                          std::is_floating_point<T>::value ||
-                                          std::is_pointer<T>::value ||
-                                          std::is_same<VoidPtr, T>::value)> {};
-
-  enum StoragePolicy { ByPointer, ByVolatilePointer, ByValue };
-  template <typename T>
-  struct storage_policy
-      : std::integral_constant<StoragePolicy,
-                               (std::is_volatile<T>::value
-                                    ? ByVolatilePointer
-                                    : (store_by_value<T>::value ? ByValue
-                                                                : ByPointer))> {
-  };
-
-  // To reduce the number of vtables we will decay values before hand.
-  // Anything with a user-defined Convert will get its own vtable.
-  // For everything else:
-  //   - Decay char* and char arrays into `const char*`
-  //   - Decay wchar_t* and wchar_t arrays into `const wchar_t*`
-  //   - Decay any other pointer to `const void*`
-  //   - Decay all enums to the integral promotion of their underlying type.
-  //   - Decay function pointers to void*.
-  template <typename T, typename = void>
-  struct DecayType {
-    static constexpr bool kHasUserDefined =
-        str_format_internal::HasUserDefinedConvert<T>::value ||
-        HasAbslStringify<T>::value;
-    using type = typename std::conditional<
-        !kHasUserDefined && std::is_convertible<T, const char*>::value,
-        const char*,
-        typename std::conditional<
-            !kHasUserDefined && std::is_convertible<T, const wchar_t*>::value,
-            const wchar_t*,
-            typename std::conditional<
-                !kHasUserDefined && std::is_convertible<T, VoidPtr>::value,
-                VoidPtr,
-                const T&>::type>::type>::type;
-  };
-  template <typename T>
-  struct DecayType<
-      T, typename std::enable_if<
-             !str_format_internal::HasUserDefinedConvert<T>::value &&
-             !HasAbslStringify<T>::value && std::is_enum<T>::value>::type> {
-    using type = decltype(+typename std::underlying_type<T>::type());
-  };
-
- public:
-  template <typename T>
-  explicit FormatArgImpl(const T& value) {
-    using D = typename DecayType<T>::type;
-    static_assert(
-        std::is_same<D, const T&>::value || storage_policy<D>::value == ByValue,
-        "Decayed types must be stored by value");
-    Init(static_cast<D>(value));
-  }
-
- private:
-  friend struct str_format_internal::FormatArgImplFriend;
-  template <typename T, StoragePolicy = storage_policy<T>::value>
-  struct Manager;
-
-  template <typename T>
-  struct Manager<T, ByPointer> {
-    static Data SetValue(const T& value) {
-      Data data;
-      data.ptr = std::addressof(value);
-      return data;
-    }
-
-    static const T& Value(Data arg) { return *static_cast<const T*>(arg.ptr); }
-  };
-
-  template <typename T>
-  struct Manager<T, ByVolatilePointer> {
-    static Data SetValue(const T& value) {
-      Data data;
-      data.volatile_ptr = &value;
-      return data;
-    }
-
-    static const T& Value(Data arg) {
-      return *static_cast<const T*>(arg.volatile_ptr);
-    }
-  };
-
-  template <typename T>
-  struct Manager<T, ByValue> {
-    static Data SetValue(const T& value) {
-      Data data;
-      memcpy(data.buf, &value, sizeof(value));
-      return data;
-    }
-
-    static T Value(Data arg) {
-      T value;
-      memcpy(&value, arg.buf, sizeof(T));
-      return value;
-    }
-  };
-
-  template <typename T>
-  void Init(const T& value) {
-    data_ = Manager<T>::SetValue(value);
-    dispatcher_ = &Dispatch<T>;
-  }
-
-  template <typename T>
-  static int ToIntVal(const T& val) {
-    using CommonType = typename std::conditional<std::is_signed<T>::value,
-                                                 int64_t, uint64_t>::type;
-    if (static_cast<CommonType>(val) >
-        static_cast<CommonType>((std::numeric_limits<int>::max)())) {
-      return (std::numeric_limits<int>::max)();
-    } else if (std::is_signed<T>::value &&
-               static_cast<CommonType>(val) <
-                   static_cast<CommonType>((std::numeric_limits<int>::min)())) {
-      return (std::numeric_limits<int>::min)();
-    }
-    return static_cast<int>(val);
-  }
-
-  template <typename T>
-  static bool ToInt(Data arg, int* out, std::true_type /* is_integral */,
-                    std::false_type) {
-    *out = ToIntVal(Manager<T>::Value(arg));
-    return true;
-  }
-
-  template <typename T>
-  static bool ToInt(Data arg, int* out, std::false_type,
-                    std::true_type /* is_enum */) {
-    *out = ToIntVal(static_cast<typename std::underlying_type<T>::type>(
-        Manager<T>::Value(arg)));
-    return true;
-  }
-
-  template <typename T>
-  static bool ToInt(Data, int*, std::false_type, std::false_type) {
-    return false;
-  }
-
-  template <typename T>
-  static bool Dispatch(Data arg, FormatConversionSpecImpl spec, void* out) {
-    // A `none` conv indicates that we want the `int` conversion.
-    if (ABSL_PREDICT_FALSE(spec.conversion_char() ==
-                           FormatConversionCharInternal::kNone)) {
-      return ToInt<T>(arg, static_cast<int*>(out), std::is_integral<T>(),
-                      std::is_enum<T>());
-    }
-    if (ABSL_PREDICT_FALSE(!Contains(ArgumentToConv<T>(),
-                                     spec.conversion_char()))) {
-      return false;
-    }
-    return str_format_internal::FormatConvertImpl(
-               Manager<T>::Value(arg), spec,
-               static_cast<FormatSinkImpl*>(out))
-        .value;
-  }
-
-  Data data_;
-  Dispatcher dispatcher_;
-};
-
-#define ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(T, E)                     \
-  E template bool FormatArgImpl::Dispatch<T>(Data, FormatConversionSpecImpl, \
-                                             void*)
-
-#define ABSL_INTERNAL_FORMAT_DISPATCH_OVERLOADS_EXPAND_NO_WSTRING_VIEW_(...)   \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(str_format_internal::VoidPtr,     \
-                                             __VA_ARGS__);                     \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(bool, __VA_ARGS__);               \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(char, __VA_ARGS__);               \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(signed char, __VA_ARGS__);        \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(unsigned char, __VA_ARGS__);      \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(short, __VA_ARGS__); /* NOLINT */ \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(unsigned short,      /* NOLINT */ \
-                                             __VA_ARGS__);                     \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(int, __VA_ARGS__);                \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(unsigned int, __VA_ARGS__);       \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(long, __VA_ARGS__); /* NOLINT */  \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(unsigned long,      /* NOLINT */  \
-                                             __VA_ARGS__);                     \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(long long, /* NOLINT */           \
-                                             __VA_ARGS__);                     \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(unsigned long long, /* NOLINT */  \
-                                             __VA_ARGS__);                     \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(int128, __VA_ARGS__);             \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(uint128, __VA_ARGS__);            \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(float, __VA_ARGS__);              \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(double, __VA_ARGS__);             \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(long double, __VA_ARGS__);        \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(const char*, __VA_ARGS__);        \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(std::string, __VA_ARGS__);        \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(string_view, __VA_ARGS__);        \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(const wchar_t*, __VA_ARGS__);     \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(std::wstring, __VA_ARGS__)
-
-#if defined(ABSL_HAVE_STD_STRING_VIEW)
-#define ABSL_INTERNAL_FORMAT_DISPATCH_OVERLOADS_EXPAND_(...)       \
-  ABSL_INTERNAL_FORMAT_DISPATCH_OVERLOADS_EXPAND_NO_WSTRING_VIEW_( \
-      __VA_ARGS__);                                                \
-  ABSL_INTERNAL_FORMAT_DISPATCH_INSTANTIATE_(std::wstring_view, __VA_ARGS__)
-#else
-#define ABSL_INTERNAL_FORMAT_DISPATCH_OVERLOADS_EXPAND_(...) \
-  ABSL_INTERNAL_FORMAT_DISPATCH_OVERLOADS_EXPAND_NO_WSTRING_VIEW_(__VA_ARGS__)
-#endif
-
-ABSL_INTERNAL_FORMAT_DISPATCH_OVERLOADS_EXPAND_(extern);
-
-
-}  // namespace str_format_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_STR_FORMAT_ARG_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/bind.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/bind.h
deleted file mode 100644
index 120bc355..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/bind.h
+++ /dev/null
@@ -1,237 +0,0 @@
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_STR_FORMAT_BIND_H_
-#define ABSL_STRINGS_INTERNAL_STR_FORMAT_BIND_H_
-
-#include <cassert>
-#include <cstdio>
-#include <ostream>
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/container/inlined_vector.h"
-#include "absl/strings/internal/str_format/arg.h"
-#include "absl/strings/internal/str_format/checker.h"
-#include "absl/strings/internal/str_format/constexpr_parser.h"
-#include "absl/strings/internal/str_format/extension.h"
-#include "absl/strings/internal/str_format/parser.h"
-#include "absl/strings/string_view.h"
-#include "absl/types/span.h"
-#include "absl/utility/utility.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-class UntypedFormatSpec;
-
-namespace str_format_internal {
-
-class BoundConversion : public FormatConversionSpecImpl {
- public:
-  const FormatArgImpl* arg() const { return arg_; }
-  void set_arg(const FormatArgImpl* a) { arg_ = a; }
-
- private:
-  const FormatArgImpl* arg_;
-};
-
-// This is the type-erased class that the implementation uses.
-class UntypedFormatSpecImpl {
- public:
-  UntypedFormatSpecImpl() = delete;
-
-  explicit UntypedFormatSpecImpl(string_view s)
-      : data_(s.data()), size_(s.size()) {}
-  explicit UntypedFormatSpecImpl(
-      const str_format_internal::ParsedFormatBase* pc)
-      : data_(pc), size_(~size_t{}) {}
-
-  bool has_parsed_conversion() const { return size_ == ~size_t{}; }
-
-  string_view str() const {
-    assert(!has_parsed_conversion());
-    return string_view(static_cast<const char*>(data_), size_);
-  }
-  const str_format_internal::ParsedFormatBase* parsed_conversion() const {
-    assert(has_parsed_conversion());
-    return static_cast<const str_format_internal::ParsedFormatBase*>(data_);
-  }
-
-  template <typename T>
-  static const UntypedFormatSpecImpl& Extract(const T& s) {
-    return s.spec_;
-  }
-
- private:
-  const void* data_;
-  size_t size_;
-};
-
-template <typename T, FormatConversionCharSet...>
-struct MakeDependent {
-  using type = T;
-};
-
-// Implicitly convertible from `const char*`, `string_view`, and the
-// `ExtendedParsedFormat` type. This abstraction allows all format functions to
-// operate on any without providing too many overloads.
-template <FormatConversionCharSet... Args>
-class FormatSpecTemplate
-    : public MakeDependent<UntypedFormatSpec, Args...>::type {
-  using Base = typename MakeDependent<UntypedFormatSpec, Args...>::type;
-
-  template <bool res>
-  struct ErrorMaker {
-    constexpr bool operator()(int) const { return res; }
-  };
-
-  template <int i, int j>
-  static constexpr bool CheckArity(ErrorMaker<true> SpecifierCount = {},
-                                   ErrorMaker<i == j> ParametersPassed = {}) {
-    static_assert(SpecifierCount(i) == ParametersPassed(j),
-                  "Number of arguments passed must match the number of "
-                  "conversion specifiers.");
-    return true;
-  }
-
-  template <FormatConversionCharSet specified, FormatConversionCharSet passed,
-            int arg>
-  static constexpr bool CheckMatch(
-      ErrorMaker<Contains(specified, passed)> MismatchedArgumentNumber = {}) {
-    static_assert(MismatchedArgumentNumber(arg),
-                  "Passed argument must match specified format.");
-    return true;
-  }
-
-  template <FormatConversionCharSet... C, size_t... I>
-  static bool CheckMatches(absl::index_sequence<I...>) {
-    bool res[] = {true, CheckMatch<Args, C, I + 1>()...};
-    (void)res;
-    return true;
-  }
-
- public:
-#ifdef ABSL_INTERNAL_ENABLE_FORMAT_CHECKER
-
-  // Honeypot overload for when the string is not constexpr.
-  // We use the 'unavailable' attribute to give a better compiler error than
-  // just 'method is deleted'.
-  FormatSpecTemplate(...)  // NOLINT
-      __attribute__((unavailable("Format string is not constexpr.")));
-
-  // Honeypot overload for when the format is constexpr and invalid.
-  // We use the 'unavailable' attribute to give a better compiler error than
-  // just 'method is deleted'.
-  // To avoid checking the format twice, we just check that the format is
-  // constexpr. If it is valid, then the overload below will kick in.
-  // We add the template here to make this overload have lower priority.
-  template <typename = void>
-  FormatSpecTemplate(const char* s)  // NOLINT
-      __attribute__((
-          enable_if(str_format_internal::EnsureConstexpr(s), "constexpr trap"),
-          unavailable(
-              "Format specified does not match the arguments passed.")));
-
-  template <typename T = void>
-  FormatSpecTemplate(string_view s)  // NOLINT
-      __attribute__((enable_if(str_format_internal::EnsureConstexpr(s),
-                               "constexpr trap")))
-      : Base("to avoid noise in the compiler error") {
-    static_assert(sizeof(T*) == 0,
-                  "Format specified does not match the arguments passed.");
-  }
-
-  // Good format overload.
-  FormatSpecTemplate(const char* s)  // NOLINT
-      __attribute__((enable_if(ValidFormatImpl<Args...>(s), "bad format trap")))
-      : Base(s) {}
-
-  FormatSpecTemplate(string_view s)  // NOLINT
-      __attribute__((enable_if(ValidFormatImpl<Args...>(s), "bad format trap")))
-      : Base(s) {}
-
-#else  // ABSL_INTERNAL_ENABLE_FORMAT_CHECKER
-
-  FormatSpecTemplate(const char* s) : Base(s) {}  // NOLINT
-  FormatSpecTemplate(string_view s) : Base(s) {}  // NOLINT
-
-#endif  // ABSL_INTERNAL_ENABLE_FORMAT_CHECKER
-
-  template <FormatConversionCharSet... C>
-  FormatSpecTemplate(const ExtendedParsedFormat<C...>& pc)  // NOLINT
-      : Base(&pc) {
-    CheckArity<sizeof...(C), sizeof...(Args)>();
-    CheckMatches<C...>(absl::make_index_sequence<sizeof...(C)>{});
-  }
-};
-
-class Streamable {
- public:
-  Streamable(const UntypedFormatSpecImpl& format,
-             absl::Span<const FormatArgImpl> args)
-      : format_(format), args_(args.begin(), args.end()) {}
-
-  std::ostream& Print(std::ostream& os) const;
-
-  friend std::ostream& operator<<(std::ostream& os, const Streamable& l) {
-    return l.Print(os);
-  }
-
- private:
-  const UntypedFormatSpecImpl& format_;
-  absl::InlinedVector<FormatArgImpl, 4> args_;
-};
-
-// for testing
-std::string Summarize(UntypedFormatSpecImpl format,
-                      absl::Span<const FormatArgImpl> args);
-bool BindWithPack(const UnboundConversion* props,
-                  absl::Span<const FormatArgImpl> pack, BoundConversion* bound);
-
-bool FormatUntyped(FormatRawSinkImpl raw_sink, UntypedFormatSpecImpl format,
-                   absl::Span<const FormatArgImpl> args);
-
-std::string& AppendPack(std::string* out, UntypedFormatSpecImpl format,
-                        absl::Span<const FormatArgImpl> args);
-
-std::string FormatPack(UntypedFormatSpecImpl format,
-                       absl::Span<const FormatArgImpl> args);
-
-int FprintF(std::FILE* output, UntypedFormatSpecImpl format,
-            absl::Span<const FormatArgImpl> args);
-int SnprintF(char* output, size_t size, UntypedFormatSpecImpl format,
-             absl::Span<const FormatArgImpl> args);
-
-// Returned by Streamed(v). Converts via '%s' to the std::string created
-// by std::ostream << v.
-template <typename T>
-class StreamedWrapper {
- public:
-  explicit StreamedWrapper(const T& v) : v_(v) {}
-
- private:
-  template <typename S>
-  friend ArgConvertResult<FormatConversionCharSetUnion(
-      FormatConversionCharSetInternal::s, FormatConversionCharSetInternal::v)>
-  FormatConvertImpl(const StreamedWrapper<S>& v, FormatConversionSpecImpl conv,
-                    FormatSinkImpl* out);
-  const T& v_;
-};
-
-}  // namespace str_format_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_STR_FORMAT_BIND_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/checker.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/checker.h
deleted file mode 100644
index eab6ab9d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/checker.h
+++ /dev/null
@@ -1,100 +0,0 @@
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_STR_FORMAT_CHECKER_H_
-#define ABSL_STRINGS_INTERNAL_STR_FORMAT_CHECKER_H_
-
-#include <algorithm>
-
-#include "absl/base/attributes.h"
-#include "absl/strings/internal/str_format/arg.h"
-#include "absl/strings/internal/str_format/constexpr_parser.h"
-#include "absl/strings/internal/str_format/extension.h"
-
-// Compile time check support for entry points.
-
-#ifndef ABSL_INTERNAL_ENABLE_FORMAT_CHECKER
-// We disable format checker under vscode intellisense compilation.
-// See https://github.com/microsoft/vscode-cpptools/issues/3683 for
-// more details.
-#if ABSL_HAVE_ATTRIBUTE(enable_if) && !defined(__native_client__) && \
-    !defined(__INTELLISENSE__)
-#define ABSL_INTERNAL_ENABLE_FORMAT_CHECKER 1
-#endif  // ABSL_HAVE_ATTRIBUTE(enable_if) && !defined(__native_client__) &&
-        // !defined(__INTELLISENSE__)
-#endif  // ABSL_INTERNAL_ENABLE_FORMAT_CHECKER
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace str_format_internal {
-
-#ifdef ABSL_INTERNAL_ENABLE_FORMAT_CHECKER
-
-template <FormatConversionCharSet... C>
-constexpr bool ValidFormatImpl(string_view format) {
-  int next_arg = 0;
-  const char* p = format.data();
-  const char* const end = p + format.size();
-  constexpr FormatConversionCharSet
-      kAllowedConvs[(std::max)(sizeof...(C), size_t{1})] = {C...};
-  bool used[(std::max)(sizeof...(C), size_t{1})]{};
-  constexpr int kNumArgs = sizeof...(C);
-  while (p != end) {
-    while (p != end && *p != '%') ++p;
-    if (p == end) {
-      break;
-    }
-    if (p + 1 >= end) return false;
-    if (p[1] == '%') {
-      // %%
-      p += 2;
-      continue;
-    }
-
-    UnboundConversion conv(absl::kConstInit);
-    p = ConsumeUnboundConversion(p + 1, end, &conv, &next_arg);
-    if (p == nullptr) return false;
-    if (conv.arg_position <= 0 || conv.arg_position > kNumArgs) {
-      return false;
-    }
-    if (!Contains(kAllowedConvs[conv.arg_position - 1], conv.conv)) {
-      return false;
-    }
-    used[conv.arg_position - 1] = true;
-    for (auto extra : {conv.width, conv.precision}) {
-      if (extra.is_from_arg()) {
-        int pos = extra.get_from_arg();
-        if (pos <= 0 || pos > kNumArgs) return false;
-        used[pos - 1] = true;
-        if (!Contains(kAllowedConvs[pos - 1], '*')) {
-          return false;
-        }
-      }
-    }
-  }
-  if (sizeof...(C) != 0) {
-    for (bool b : used) {
-      if (!b) return false;
-    }
-  }
-  return true;
-}
-
-#endif  // ABSL_INTERNAL_ENABLE_FORMAT_CHECKER
-
-}  // namespace str_format_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_STR_FORMAT_CHECKER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/constexpr_parser.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/constexpr_parser.h
deleted file mode 100644
index 8f593870..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/constexpr_parser.h
+++ /dev/null
@@ -1,357 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_STR_FORMAT_CONSTEXPR_PARSER_H_
-#define ABSL_STRINGS_INTERNAL_STR_FORMAT_CONSTEXPR_PARSER_H_
-
-#include <cassert>
-#include <cstdint>
-#include <cstdio>
-#include <limits>
-
-#include "absl/base/config.h"
-#include "absl/base/const_init.h"
-#include "absl/base/optimization.h"
-#include "absl/strings/internal/str_format/extension.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace str_format_internal {
-
-// The analyzed properties of a single specified conversion.
-struct UnboundConversion {
-  // This is a user defined default constructor on purpose to skip the
-  // initialization of parts of the object that are not necessary.
-  UnboundConversion() {}  // NOLINT
-
-  // This constructor is provided for the static checker. We don't want to do
-  // the unnecessary initialization in the normal case.
-  explicit constexpr UnboundConversion(absl::ConstInitType)
-      : arg_position{}, width{}, precision{} {}
-
-  class InputValue {
-   public:
-    constexpr void set_value(int value) {
-      assert(value >= 0);
-      value_ = value;
-    }
-    constexpr int value() const { return value_; }
-
-    // Marks the value as "from arg". aka the '*' format.
-    // Requires `value >= 1`.
-    // When set, is_from_arg() return true and get_from_arg() returns the
-    // original value.
-    // `value()`'s return value is unspecified in this state.
-    constexpr void set_from_arg(int value) {
-      assert(value > 0);
-      value_ = -value - 1;
-    }
-    constexpr bool is_from_arg() const { return value_ < -1; }
-    constexpr int get_from_arg() const {
-      assert(is_from_arg());
-      return -value_ - 1;
-    }
-
-   private:
-    int value_ = -1;
-  };
-
-  // No need to initialize. It will always be set in the parser.
-  int arg_position;
-
-  InputValue width;
-  InputValue precision;
-
-  Flags flags = Flags::kBasic;
-  LengthMod length_mod = LengthMod::none;
-  FormatConversionChar conv = FormatConversionCharInternal::kNone;
-};
-
-// Helper tag class for the table below.
-// It allows fast `char -> ConversionChar/LengthMod/Flags` checking and
-// conversions.
-class ConvTag {
- public:
-  constexpr ConvTag(FormatConversionChar conversion_char)  // NOLINT
-      : tag_(static_cast<uint8_t>(conversion_char)) {}
-  constexpr ConvTag(LengthMod length_mod)  // NOLINT
-      : tag_(0x80 | static_cast<uint8_t>(length_mod)) {}
-  constexpr ConvTag(Flags flags)  // NOLINT
-      : tag_(0xc0 | static_cast<uint8_t>(flags)) {}
-  constexpr ConvTag() : tag_(0xFF) {}
-
-  constexpr bool is_conv() const { return (tag_ & 0x80) == 0; }
-  constexpr bool is_length() const { return (tag_ & 0xC0) == 0x80; }
-  constexpr bool is_flags() const { return (tag_ & 0xE0) == 0xC0; }
-
-  constexpr FormatConversionChar as_conv() const {
-    assert(is_conv());
-    assert(!is_length());
-    assert(!is_flags());
-    return static_cast<FormatConversionChar>(tag_);
-  }
-  constexpr LengthMod as_length() const {
-    assert(!is_conv());
-    assert(is_length());
-    assert(!is_flags());
-    return static_cast<LengthMod>(tag_ & 0x3F);
-  }
-  constexpr Flags as_flags() const {
-    assert(!is_conv());
-    assert(!is_length());
-    assert(is_flags());
-    return static_cast<Flags>(tag_ & 0x1F);
-  }
-
- private:
-  uint8_t tag_;
-};
-
-struct ConvTagHolder {
-  using CC = FormatConversionCharInternal;
-  using LM = LengthMod;
-
-  // Abbreviations to fit in the table below.
-  static constexpr auto kFSign = Flags::kSignCol;
-  static constexpr auto kFAlt = Flags::kAlt;
-  static constexpr auto kFPos = Flags::kShowPos;
-  static constexpr auto kFLeft = Flags::kLeft;
-  static constexpr auto kFZero = Flags::kZero;
-
-  static constexpr ConvTag value[256] = {
-      {},     {},    {},    {},    {},    {},     {},    {},     // 00-07
-      {},     {},    {},    {},    {},    {},     {},    {},     // 08-0f
-      {},     {},    {},    {},    {},    {},     {},    {},     // 10-17
-      {},     {},    {},    {},    {},    {},     {},    {},     // 18-1f
-      kFSign, {},    {},    kFAlt, {},    {},     {},    {},     //  !"#$%&'
-      {},     {},    {},    kFPos, {},    kFLeft, {},    {},     // ()*+,-./
-      kFZero, {},    {},    {},    {},    {},     {},    {},     // 01234567
-      {},     {},    {},    {},    {},    {},     {},    {},     // 89:;<=>?
-      {},     CC::A, {},    {},    {},    CC::E,  CC::F, CC::G,  // @ABCDEFG
-      {},     {},    {},    {},    LM::L, {},     {},    {},     // HIJKLMNO
-      {},     {},    {},    {},    {},    {},     {},    {},     // PQRSTUVW
-      CC::X,  {},    {},    {},    {},    {},     {},    {},     // XYZ[\]^_
-      {},     CC::a, {},    CC::c, CC::d, CC::e,  CC::f, CC::g,  // `abcdefg
-      LM::h,  CC::i, LM::j, {},    LM::l, {},     CC::n, CC::o,  // hijklmno
-      CC::p,  LM::q, {},    CC::s, LM::t, CC::u,  CC::v, {},     // pqrstuvw
-      CC::x,  {},    LM::z, {},    {},    {},     {},    {},     // xyz{|}!
-      {},     {},    {},    {},    {},    {},     {},    {},     // 80-87
-      {},     {},    {},    {},    {},    {},     {},    {},     // 88-8f
-      {},     {},    {},    {},    {},    {},     {},    {},     // 90-97
-      {},     {},    {},    {},    {},    {},     {},    {},     // 98-9f
-      {},     {},    {},    {},    {},    {},     {},    {},     // a0-a7
-      {},     {},    {},    {},    {},    {},     {},    {},     // a8-af
-      {},     {},    {},    {},    {},    {},     {},    {},     // b0-b7
-      {},     {},    {},    {},    {},    {},     {},    {},     // b8-bf
-      {},     {},    {},    {},    {},    {},     {},    {},     // c0-c7
-      {},     {},    {},    {},    {},    {},     {},    {},     // c8-cf
-      {},     {},    {},    {},    {},    {},     {},    {},     // d0-d7
-      {},     {},    {},    {},    {},    {},     {},    {},     // d8-df
-      {},     {},    {},    {},    {},    {},     {},    {},     // e0-e7
-      {},     {},    {},    {},    {},    {},     {},    {},     // e8-ef
-      {},     {},    {},    {},    {},    {},     {},    {},     // f0-f7
-      {},     {},    {},    {},    {},    {},     {},    {},     // f8-ff
-  };
-};
-
-// Keep a single table for all the conversion chars and length modifiers.
-constexpr ConvTag GetTagForChar(char c) {
-  return ConvTagHolder::value[static_cast<unsigned char>(c)];
-}
-
-constexpr bool CheckFastPathSetting(const UnboundConversion& conv) {
-  bool width_precision_needed =
-      conv.width.value() >= 0 || conv.precision.value() >= 0;
-  if (width_precision_needed && conv.flags == Flags::kBasic) {
-#if defined(__clang__)
-    // Some compilers complain about this in constexpr even when not executed,
-    // so only enable the error dump in clang.
-    fprintf(stderr,
-            "basic=%d left=%d show_pos=%d sign_col=%d alt=%d zero=%d "
-            "width=%d precision=%d\n",
-            conv.flags == Flags::kBasic ? 1 : 0,
-            FlagsContains(conv.flags, Flags::kLeft) ? 1 : 0,
-            FlagsContains(conv.flags, Flags::kShowPos) ? 1 : 0,
-            FlagsContains(conv.flags, Flags::kSignCol) ? 1 : 0,
-            FlagsContains(conv.flags, Flags::kAlt) ? 1 : 0,
-            FlagsContains(conv.flags, Flags::kZero) ? 1 : 0, conv.width.value(),
-            conv.precision.value());
-#endif  // defined(__clang__)
-    return false;
-  }
-  return true;
-}
-
-constexpr int ParseDigits(char& c, const char*& pos, const char* const end) {
-  int digits = c - '0';
-  // We do not want to overflow `digits` so we consume at most digits10
-  // digits. If there are more digits the parsing will fail later on when the
-  // digit doesn't match the expected characters.
-  int num_digits = std::numeric_limits<int>::digits10;
-  for (;;) {
-    if (ABSL_PREDICT_FALSE(pos == end)) break;
-    c = *pos++;
-    if ('0' > c || c > '9') break;
-    --num_digits;
-    if (ABSL_PREDICT_FALSE(!num_digits)) break;
-    digits = 10 * digits + c - '0';
-  }
-  return digits;
-}
-
-template <bool is_positional>
-constexpr const char* ConsumeConversion(const char* pos, const char* const end,
-                                        UnboundConversion* conv,
-                                        int* next_arg) {
-  const char* const original_pos = pos;
-  char c = 0;
-  // Read the next char into `c` and update `pos`. Returns false if there are
-  // no more chars to read.
-#define ABSL_FORMAT_PARSER_INTERNAL_GET_CHAR()          \
-  do {                                                  \
-    if (ABSL_PREDICT_FALSE(pos == end)) return nullptr; \
-    c = *pos++;                                         \
-  } while (0)
-
-  if (is_positional) {
-    ABSL_FORMAT_PARSER_INTERNAL_GET_CHAR();
-    if (ABSL_PREDICT_FALSE(c < '1' || c > '9')) return nullptr;
-    conv->arg_position = ParseDigits(c, pos, end);
-    assert(conv->arg_position > 0);
-    if (ABSL_PREDICT_FALSE(c != '$')) return nullptr;
-  }
-
-  ABSL_FORMAT_PARSER_INTERNAL_GET_CHAR();
-
-  // We should start with the basic flag on.
-  assert(conv->flags == Flags::kBasic);
-
-  // Any non alpha character makes this conversion not basic.
-  // This includes flags (-+ #0), width (1-9, *) or precision (.).
-  // All conversion characters and length modifiers are alpha characters.
-  if (c < 'A') {
-    while (c <= '0') {
-      auto tag = GetTagForChar(c);
-      if (tag.is_flags()) {
-        conv->flags = conv->flags | tag.as_flags();
-        ABSL_FORMAT_PARSER_INTERNAL_GET_CHAR();
-      } else {
-        break;
-      }
-    }
-
-    if (c <= '9') {
-      if (c >= '0') {
-        int maybe_width = ParseDigits(c, pos, end);
-        if (!is_positional && c == '$') {
-          if (ABSL_PREDICT_FALSE(*next_arg != 0)) return nullptr;
-          // Positional conversion.
-          *next_arg = -1;
-          return ConsumeConversion<true>(original_pos, end, conv, next_arg);
-        }
-        conv->flags = conv->flags | Flags::kNonBasic;
-        conv->width.set_value(maybe_width);
-      } else if (c == '*') {
-        conv->flags = conv->flags | Flags::kNonBasic;
-        ABSL_FORMAT_PARSER_INTERNAL_GET_CHAR();
-        if (is_positional) {
-          if (ABSL_PREDICT_FALSE(c < '1' || c > '9')) return nullptr;
-          conv->width.set_from_arg(ParseDigits(c, pos, end));
-          if (ABSL_PREDICT_FALSE(c != '$')) return nullptr;
-          ABSL_FORMAT_PARSER_INTERNAL_GET_CHAR();
-        } else {
-          conv->width.set_from_arg(++*next_arg);
-        }
-      }
-    }
-
-    if (c == '.') {
-      conv->flags = conv->flags | Flags::kNonBasic;
-      ABSL_FORMAT_PARSER_INTERNAL_GET_CHAR();
-      if ('0' <= c && c <= '9') {
-        conv->precision.set_value(ParseDigits(c, pos, end));
-      } else if (c == '*') {
-        ABSL_FORMAT_PARSER_INTERNAL_GET_CHAR();
-        if (is_positional) {
-          if (ABSL_PREDICT_FALSE(c < '1' || c > '9')) return nullptr;
-          conv->precision.set_from_arg(ParseDigits(c, pos, end));
-          if (c != '$') return nullptr;
-          ABSL_FORMAT_PARSER_INTERNAL_GET_CHAR();
-        } else {
-          conv->precision.set_from_arg(++*next_arg);
-        }
-      } else {
-        conv->precision.set_value(0);
-      }
-    }
-  }
-
-  auto tag = GetTagForChar(c);
-
-  if (ABSL_PREDICT_FALSE(c == 'v' && conv->flags != Flags::kBasic)) {
-    return nullptr;
-  }
-
-  if (ABSL_PREDICT_FALSE(!tag.is_conv())) {
-    if (ABSL_PREDICT_FALSE(!tag.is_length())) return nullptr;
-
-    // It is a length modifier.
-    LengthMod length_mod = tag.as_length();
-    ABSL_FORMAT_PARSER_INTERNAL_GET_CHAR();
-    if (c == 'h' && length_mod == LengthMod::h) {
-      conv->length_mod = LengthMod::hh;
-      ABSL_FORMAT_PARSER_INTERNAL_GET_CHAR();
-    } else if (c == 'l' && length_mod == LengthMod::l) {
-      conv->length_mod = LengthMod::ll;
-      ABSL_FORMAT_PARSER_INTERNAL_GET_CHAR();
-    } else {
-      conv->length_mod = length_mod;
-    }
-    tag = GetTagForChar(c);
-
-    if (ABSL_PREDICT_FALSE(c == 'v')) return nullptr;
-    if (ABSL_PREDICT_FALSE(!tag.is_conv())) return nullptr;
-
-    // `wchar_t` args are marked non-basic so `Bind()` will copy the length mod.
-    if (conv->length_mod == LengthMod::l && c == 'c') {
-      conv->flags = conv->flags | Flags::kNonBasic;
-    }
-  }
-#undef ABSL_FORMAT_PARSER_INTERNAL_GET_CHAR
-
-  assert(CheckFastPathSetting(*conv));
-  (void)(&CheckFastPathSetting);
-
-  conv->conv = tag.as_conv();
-  if (!is_positional) conv->arg_position = ++*next_arg;
-  return pos;
-}
-
-// Consume conversion spec prefix (not including '%') of [p, end) if valid.
-// Examples of valid specs would be e.g.: "s", "d", "-12.6f".
-// If valid, it returns the first character following the conversion spec,
-// and the spec part is broken down and returned in 'conv'.
-// If invalid, returns nullptr.
-constexpr const char* ConsumeUnboundConversion(const char* p, const char* end,
-                                               UnboundConversion* conv,
-                                               int* next_arg) {
-  if (*next_arg < 0) return ConsumeConversion<true>(p, end, conv, next_arg);
-  return ConsumeConversion<false>(p, end, conv, next_arg);
-}
-
-}  // namespace str_format_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_STR_FORMAT_CONSTEXPR_PARSER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/extension.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/extension.h
deleted file mode 100644
index 173284c6..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/extension.h
+++ /dev/null
@@ -1,456 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-#ifndef ABSL_STRINGS_INTERNAL_STR_FORMAT_EXTENSION_H_
-#define ABSL_STRINGS_INTERNAL_STR_FORMAT_EXTENSION_H_
-
-
-#include <cstddef>
-#include <cstdint>
-#include <cstring>
-#include <ostream>
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/strings/internal/str_format/output.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-enum class FormatConversionChar : uint8_t;
-enum class FormatConversionCharSet : uint64_t;
-enum class LengthMod : std::uint8_t { h, hh, l, ll, L, j, z, t, q, none };
-
-namespace str_format_internal {
-
-class FormatRawSinkImpl {
- public:
-  // Implicitly convert from any type that provides the hook function as
-  // described above.
-  template <typename T, decltype(str_format_internal::InvokeFlush(
-                            std::declval<T*>(), string_view()))* = nullptr>
-  FormatRawSinkImpl(T* raw)  // NOLINT
-      : sink_(raw), write_(&FormatRawSinkImpl::Flush<T>) {}
-
-  void Write(string_view s) { write_(sink_, s); }
-
-  template <typename T>
-  static FormatRawSinkImpl Extract(T s) {
-    return s.sink_;
-  }
-
- private:
-  template <typename T>
-  static void Flush(void* r, string_view s) {
-    str_format_internal::InvokeFlush(static_cast<T*>(r), s);
-  }
-
-  void* sink_;
-  void (*write_)(void*, string_view);
-};
-
-// An abstraction to which conversions write their string data.
-class FormatSinkImpl {
- public:
-  explicit FormatSinkImpl(FormatRawSinkImpl raw) : raw_(raw) {}
-
-  ~FormatSinkImpl() { Flush(); }
-
-  void Flush() {
-    raw_.Write(string_view(buf_, static_cast<size_t>(pos_ - buf_)));
-    pos_ = buf_;
-  }
-
-  void Append(size_t n, char c) {
-    if (n == 0) return;
-    size_ += n;
-    auto raw_append = [&](size_t count) {
-      memset(pos_, c, count);
-      pos_ += count;
-    };
-    while (n > Avail()) {
-      n -= Avail();
-      if (Avail() > 0) {
-        raw_append(Avail());
-      }
-      Flush();
-    }
-    raw_append(n);
-  }
-
-  void Append(string_view v) {
-    size_t n = v.size();
-    if (n == 0) return;
-    size_ += n;
-    if (n >= Avail()) {
-      Flush();
-      raw_.Write(v);
-      return;
-    }
-    memcpy(pos_, v.data(), n);
-    pos_ += n;
-  }
-
-  size_t size() const { return size_; }
-
-  // Put 'v' to 'sink' with specified width, precision, and left flag.
-  bool PutPaddedString(string_view v, int width, int precision, bool left);
-
-  template <typename T>
-  T Wrap() {
-    return T(this);
-  }
-
-  template <typename T>
-  static FormatSinkImpl* Extract(T* s) {
-    return s->sink_;
-  }
-
- private:
-  size_t Avail() const {
-    return static_cast<size_t>(buf_ + sizeof(buf_) - pos_);
-  }
-
-  FormatRawSinkImpl raw_;
-  size_t size_ = 0;
-  char* pos_ = buf_;
-  char buf_[1024];
-};
-
-enum class Flags : uint8_t {
-  kBasic = 0,
-  kLeft = 1 << 0,
-  kShowPos = 1 << 1,
-  kSignCol = 1 << 2,
-  kAlt = 1 << 3,
-  kZero = 1 << 4,
-  // This is not a real flag. It just exists to turn off kBasic when no other
-  // flags are set. This is for when width/precision are specified, or a length
-  // modifier affects the behavior ("%lc").
-  kNonBasic = 1 << 5,
-};
-
-constexpr Flags operator|(Flags a, Flags b) {
-  return static_cast<Flags>(static_cast<uint8_t>(a) | static_cast<uint8_t>(b));
-}
-
-constexpr bool FlagsContains(Flags haystack, Flags needle) {
-  return (static_cast<uint8_t>(haystack) & static_cast<uint8_t>(needle)) ==
-         static_cast<uint8_t>(needle);
-}
-
-std::string FlagsToString(Flags v);
-
-inline std::ostream& operator<<(std::ostream& os, Flags v) {
-  return os << FlagsToString(v);
-}
-
-// clang-format off
-#define ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(X_VAL, X_SEP) \
-  /* text */ \
-  X_VAL(c) X_SEP X_VAL(s) X_SEP \
-  /* ints */ \
-  X_VAL(d) X_SEP X_VAL(i) X_SEP X_VAL(o) X_SEP \
-  X_VAL(u) X_SEP X_VAL(x) X_SEP X_VAL(X) X_SEP \
-  /* floats */ \
-  X_VAL(f) X_SEP X_VAL(F) X_SEP X_VAL(e) X_SEP X_VAL(E) X_SEP \
-  X_VAL(g) X_SEP X_VAL(G) X_SEP X_VAL(a) X_SEP X_VAL(A) X_SEP \
-  /* misc */ \
-  X_VAL(n) X_SEP X_VAL(p) X_SEP X_VAL(v)
-// clang-format on
-
-// This type should not be referenced, it exists only to provide labels
-// internally that match the values declared in FormatConversionChar in
-// str_format.h. This is meant to allow internal libraries to use the same
-// declared interface type as the public interface
-// (absl::StrFormatConversionChar) while keeping the definition in a public
-// header.
-// Internal libraries should use the form
-// `FormatConversionCharInternal::c`, `FormatConversionCharInternal::kNone` for
-// comparisons.  Use in switch statements is not recommended due to a bug in how
-// gcc 4.9 -Wswitch handles declared but undefined enums.
-struct FormatConversionCharInternal {
-  FormatConversionCharInternal() = delete;
-
- private:
-  // clang-format off
-  enum class Enum : uint8_t {
-    c, s,                    // text
-    d, i, o, u, x, X,        // int
-    f, F, e, E, g, G, a, A,  // float
-    n, p, v,                    // misc
-    kNone
-  };
-  // clang-format on
- public:
-#define ABSL_INTERNAL_X_VAL(id)              \
-  static constexpr FormatConversionChar id = \
-      static_cast<FormatConversionChar>(Enum::id);
-  ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(ABSL_INTERNAL_X_VAL, )
-#undef ABSL_INTERNAL_X_VAL
-  static constexpr FormatConversionChar kNone =
-      static_cast<FormatConversionChar>(Enum::kNone);
-};
-// clang-format on
-
-inline FormatConversionChar FormatConversionCharFromChar(char c) {
-  switch (c) {
-#define ABSL_INTERNAL_X_VAL(id) \
-  case #id[0]:                  \
-    return FormatConversionCharInternal::id;
-    ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(ABSL_INTERNAL_X_VAL, )
-#undef ABSL_INTERNAL_X_VAL
-  }
-  return FormatConversionCharInternal::kNone;
-}
-
-inline bool FormatConversionCharIsUpper(FormatConversionChar c) {
-  if (c == FormatConversionCharInternal::X ||
-      c == FormatConversionCharInternal::F ||
-      c == FormatConversionCharInternal::E ||
-      c == FormatConversionCharInternal::G ||
-      c == FormatConversionCharInternal::A) {
-    return true;
-  } else {
-    return false;
-  }
-}
-
-inline bool FormatConversionCharIsFloat(FormatConversionChar c) {
-  if (c == FormatConversionCharInternal::a ||
-      c == FormatConversionCharInternal::e ||
-      c == FormatConversionCharInternal::f ||
-      c == FormatConversionCharInternal::g ||
-      c == FormatConversionCharInternal::A ||
-      c == FormatConversionCharInternal::E ||
-      c == FormatConversionCharInternal::F ||
-      c == FormatConversionCharInternal::G) {
-    return true;
-  } else {
-    return false;
-  }
-}
-
-inline char FormatConversionCharToChar(FormatConversionChar c) {
-  if (c == FormatConversionCharInternal::kNone) {
-    return '\0';
-
-#define ABSL_INTERNAL_X_VAL(e)                       \
-  } else if (c == FormatConversionCharInternal::e) { \
-    return #e[0];
-#define ABSL_INTERNAL_X_SEP
-  ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(ABSL_INTERNAL_X_VAL,
-                                         ABSL_INTERNAL_X_SEP)
-  } else {
-    return '\0';
-  }
-
-#undef ABSL_INTERNAL_X_VAL
-#undef ABSL_INTERNAL_X_SEP
-}
-
-// The associated char.
-inline std::ostream& operator<<(std::ostream& os, FormatConversionChar v) {
-  char c = FormatConversionCharToChar(v);
-  if (!c) c = '?';
-  return os << c;
-}
-
-struct FormatConversionSpecImplFriend;
-
-class FormatConversionSpecImpl {
- public:
-  // Width and precision are not specified, no flags are set.
-  bool is_basic() const { return flags_ == Flags::kBasic; }
-  bool has_left_flag() const { return FlagsContains(flags_, Flags::kLeft); }
-  bool has_show_pos_flag() const {
-    return FlagsContains(flags_, Flags::kShowPos);
-  }
-  bool has_sign_col_flag() const {
-    return FlagsContains(flags_, Flags::kSignCol);
-  }
-  bool has_alt_flag() const { return FlagsContains(flags_, Flags::kAlt); }
-  bool has_zero_flag() const { return FlagsContains(flags_, Flags::kZero); }
-
-  LengthMod length_mod() const { return length_mod_; }
-
-  FormatConversionChar conversion_char() const {
-    // Keep this field first in the struct . It generates better code when
-    // accessing it when ConversionSpec is passed by value in registers.
-    static_assert(offsetof(FormatConversionSpecImpl, conv_) == 0, "");
-    return conv_;
-  }
-
-  void set_conversion_char(FormatConversionChar c) { conv_ = c; }
-
-  // Returns the specified width. If width is unspecfied, it returns a negative
-  // value.
-  int width() const { return width_; }
-  // Returns the specified precision. If precision is unspecfied, it returns a
-  // negative value.
-  int precision() const { return precision_; }
-
-  template <typename T>
-  T Wrap() {
-    return T(*this);
-  }
-
- private:
-  friend struct str_format_internal::FormatConversionSpecImplFriend;
-  FormatConversionChar conv_ = FormatConversionCharInternal::kNone;
-  Flags flags_;
-  LengthMod length_mod_ = LengthMod::none;
-  int width_;
-  int precision_;
-};
-
-struct FormatConversionSpecImplFriend final {
-  static void SetFlags(Flags f, FormatConversionSpecImpl* conv) {
-    conv->flags_ = f;
-  }
-  static void SetLengthMod(LengthMod l, FormatConversionSpecImpl* conv) {
-    conv->length_mod_ = l;
-  }
-  static void SetConversionChar(FormatConversionChar c,
-                                FormatConversionSpecImpl* conv) {
-    conv->conv_ = c;
-  }
-  static void SetWidth(int w, FormatConversionSpecImpl* conv) {
-    conv->width_ = w;
-  }
-  static void SetPrecision(int p, FormatConversionSpecImpl* conv) {
-    conv->precision_ = p;
-  }
-  static std::string FlagsToString(const FormatConversionSpecImpl& spec) {
-    return str_format_internal::FlagsToString(spec.flags_);
-  }
-};
-
-// Type safe OR operator.
-// We need this for two reasons:
-//  1. operator| on enums makes them decay to integers and the result is an
-//     integer. We need the result to stay as an enum.
-//  2. We use "enum class" which would not work even if we accepted the decay.
-constexpr FormatConversionCharSet FormatConversionCharSetUnion(
-    FormatConversionCharSet a) {
-  return a;
-}
-
-template <typename... CharSet>
-constexpr FormatConversionCharSet FormatConversionCharSetUnion(
-    FormatConversionCharSet a, CharSet... rest) {
-  return static_cast<FormatConversionCharSet>(
-      static_cast<uint64_t>(a) |
-      static_cast<uint64_t>(FormatConversionCharSetUnion(rest...)));
-}
-
-constexpr uint64_t FormatConversionCharToConvInt(FormatConversionChar c) {
-  return uint64_t{1} << (1 + static_cast<uint8_t>(c));
-}
-
-constexpr uint64_t FormatConversionCharToConvInt(char conv) {
-  return
-#define ABSL_INTERNAL_CHAR_SET_CASE(c)                                 \
-  conv == #c[0]                                                        \
-      ? FormatConversionCharToConvInt(FormatConversionCharInternal::c) \
-      :
-      ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(ABSL_INTERNAL_CHAR_SET_CASE, )
-#undef ABSL_INTERNAL_CHAR_SET_CASE
-                  conv == '*'
-          ? 1
-          : 0;
-}
-
-constexpr FormatConversionCharSet FormatConversionCharToConvValue(char conv) {
-  return static_cast<FormatConversionCharSet>(
-      FormatConversionCharToConvInt(conv));
-}
-
-struct FormatConversionCharSetInternal {
-#define ABSL_INTERNAL_CHAR_SET_CASE(c)         \
-  static constexpr FormatConversionCharSet c = \
-      FormatConversionCharToConvValue(#c[0]);
-  ABSL_INTERNAL_CONVERSION_CHARS_EXPAND_(ABSL_INTERNAL_CHAR_SET_CASE, )
-#undef ABSL_INTERNAL_CHAR_SET_CASE
-
-  // Used for width/precision '*' specification.
-  static constexpr FormatConversionCharSet kStar =
-      FormatConversionCharToConvValue('*');
-
-  static constexpr FormatConversionCharSet kIntegral =
-      FormatConversionCharSetUnion(d, i, u, o, x, X);
-  static constexpr FormatConversionCharSet kFloating =
-      FormatConversionCharSetUnion(a, e, f, g, A, E, F, G);
-  static constexpr FormatConversionCharSet kNumeric =
-      FormatConversionCharSetUnion(kIntegral, kFloating);
-  static constexpr FormatConversionCharSet kPointer = p;
-};
-
-// Type safe OR operator.
-// We need this for two reasons:
-//  1. operator| on enums makes them decay to integers and the result is an
-//     integer. We need the result to stay as an enum.
-//  2. We use "enum class" which would not work even if we accepted the decay.
-constexpr FormatConversionCharSet operator|(FormatConversionCharSet a,
-                                            FormatConversionCharSet b) {
-  return FormatConversionCharSetUnion(a, b);
-}
-
-// Overloaded conversion functions to support absl::ParsedFormat.
-// Get a conversion with a single character in it.
-constexpr FormatConversionCharSet ToFormatConversionCharSet(char c) {
-  return static_cast<FormatConversionCharSet>(
-      FormatConversionCharToConvValue(c));
-}
-
-// Get a conversion with a single character in it.
-constexpr FormatConversionCharSet ToFormatConversionCharSet(
-    FormatConversionCharSet c) {
-  return c;
-}
-
-template <typename T>
-void ToFormatConversionCharSet(T) = delete;
-
-// Checks whether `c` exists in `set`.
-constexpr bool Contains(FormatConversionCharSet set, char c) {
-  return (static_cast<uint64_t>(set) &
-          static_cast<uint64_t>(FormatConversionCharToConvValue(c))) != 0;
-}
-
-// Checks whether all the characters in `c` are contained in `set`
-constexpr bool Contains(FormatConversionCharSet set,
-                        FormatConversionCharSet c) {
-  return (static_cast<uint64_t>(set) & static_cast<uint64_t>(c)) ==
-         static_cast<uint64_t>(c);
-}
-
-// Checks whether all the characters in `c` are contained in `set`
-constexpr bool Contains(FormatConversionCharSet set, FormatConversionChar c) {
-  return (static_cast<uint64_t>(set) & FormatConversionCharToConvInt(c)) != 0;
-}
-
-// Return capacity - used, clipped to a minimum of 0.
-inline size_t Excess(size_t used, size_t capacity) {
-  return used < capacity ? capacity - used : 0;
-}
-
-}  // namespace str_format_internal
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_STR_FORMAT_EXTENSION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/float_conversion.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/float_conversion.h
deleted file mode 100644
index 71100e71..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/float_conversion.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_STR_FORMAT_FLOAT_CONVERSION_H_
-#define ABSL_STRINGS_INTERNAL_STR_FORMAT_FLOAT_CONVERSION_H_
-
-#include "absl/strings/internal/str_format/extension.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace str_format_internal {
-
-bool ConvertFloatImpl(float v, const FormatConversionSpecImpl &conv,
-                      FormatSinkImpl *sink);
-
-bool ConvertFloatImpl(double v, const FormatConversionSpecImpl &conv,
-                      FormatSinkImpl *sink);
-
-bool ConvertFloatImpl(long double v, const FormatConversionSpecImpl &conv,
-                      FormatSinkImpl *sink);
-
-}  // namespace str_format_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_STR_FORMAT_FLOAT_CONVERSION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/output.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/output.h
deleted file mode 100644
index 15e751ab..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/output.h
+++ /dev/null
@@ -1,97 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Output extension hooks for the Format library.
-// `internal::InvokeFlush` calls the appropriate flush function for the
-// specified output argument.
-// `BufferRawSink` is a simple output sink for a char buffer. Used by SnprintF.
-// `FILERawSink` is a std::FILE* based sink. Used by PrintF and FprintF.
-
-#ifndef ABSL_STRINGS_INTERNAL_STR_FORMAT_OUTPUT_H_
-#define ABSL_STRINGS_INTERNAL_STR_FORMAT_OUTPUT_H_
-
-#include <cstdio>
-#include <ios>
-#include <ostream>
-#include <string>
-
-#include "absl/base/port.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace str_format_internal {
-
-// RawSink implementation that writes into a char* buffer.
-// It will not overflow the buffer, but will keep the total count of chars
-// that would have been written.
-class BufferRawSink {
- public:
-  BufferRawSink(char* buffer, size_t size) : buffer_(buffer), size_(size) {}
-
-  size_t total_written() const { return total_written_; }
-  void Write(string_view v);
-
- private:
-  char* buffer_;
-  size_t size_;
-  size_t total_written_ = 0;
-};
-
-// RawSink implementation that writes into a FILE*.
-// It keeps track of the total number of bytes written and any error encountered
-// during the writes.
-class FILERawSink {
- public:
-  explicit FILERawSink(std::FILE* output) : output_(output) {}
-
-  void Write(string_view v);
-
-  size_t count() const { return count_; }
-  int error() const { return error_; }
-
- private:
-  std::FILE* output_;
-  int error_ = 0;
-  size_t count_ = 0;
-};
-
-// Provide RawSink integration with common types from the STL.
-inline void AbslFormatFlush(std::string* out, string_view s) {
-  out->append(s.data(), s.size());
-}
-inline void AbslFormatFlush(std::ostream* out, string_view s) {
-  out->write(s.data(), static_cast<std::streamsize>(s.size()));
-}
-
-inline void AbslFormatFlush(FILERawSink* sink, string_view v) {
-  sink->Write(v);
-}
-
-inline void AbslFormatFlush(BufferRawSink* sink, string_view v) {
-  sink->Write(v);
-}
-
-// This is a SFINAE to get a better compiler error message when the type
-// is not supported.
-template <typename T>
-auto InvokeFlush(T* out, string_view s) -> decltype(AbslFormatFlush(out, s)) {
-  AbslFormatFlush(out, s);
-}
-
-}  // namespace str_format_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_STR_FORMAT_OUTPUT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/parser.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/parser.h
deleted file mode 100644
index b1d6d5fd..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_format/parser.h
+++ /dev/null
@@ -1,269 +0,0 @@
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_STR_FORMAT_PARSER_H_
-#define ABSL_STRINGS_INTERNAL_STR_FORMAT_PARSER_H_
-
-#include <stddef.h>
-#include <stdlib.h>
-
-#include <cassert>
-#include <cstring>
-#include <initializer_list>
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "absl/base/config.h"
-#include "absl/base/optimization.h"
-#include "absl/strings/internal/str_format/checker.h"
-#include "absl/strings/internal/str_format/constexpr_parser.h"
-#include "absl/strings/internal/str_format/extension.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace str_format_internal {
-
-std::string LengthModToString(LengthMod v);
-
-const char* ConsumeUnboundConversionNoInline(const char* p, const char* end,
-                                             UnboundConversion* conv,
-                                             int* next_arg);
-
-// Parse the format string provided in 'src' and pass the identified items into
-// 'consumer'.
-// Text runs will be passed by calling
-//   Consumer::Append(string_view);
-// ConversionItems will be passed by calling
-//   Consumer::ConvertOne(UnboundConversion, string_view);
-// In the case of ConvertOne, the string_view that is passed is the
-// portion of the format string corresponding to the conversion, not including
-// the leading %. On success, it returns true. On failure, it stops and returns
-// false.
-template <typename Consumer>
-bool ParseFormatString(string_view src, Consumer consumer) {
-  int next_arg = 0;
-  const char* p = src.data();
-  const char* const end = p + src.size();
-  while (p != end) {
-    const char* percent =
-        static_cast<const char*>(memchr(p, '%', static_cast<size_t>(end - p)));
-    if (!percent) {
-      // We found the last substring.
-      return consumer.Append(string_view(p, static_cast<size_t>(end - p)));
-    }
-    // We found a percent, so push the text run then process the percent.
-    if (ABSL_PREDICT_FALSE(!consumer.Append(
-            string_view(p, static_cast<size_t>(percent - p))))) {
-      return false;
-    }
-    if (ABSL_PREDICT_FALSE(percent + 1 >= end)) return false;
-
-    auto tag = GetTagForChar(percent[1]);
-    if (tag.is_conv()) {
-      if (ABSL_PREDICT_FALSE(next_arg < 0)) {
-        // This indicates an error in the format string.
-        // The only way to get `next_arg < 0` here is to have a positional
-        // argument first which sets next_arg to -1 and then a non-positional
-        // argument.
-        return false;
-      }
-      p = percent + 2;
-
-      // Keep this case separate from the one below.
-      // ConvertOne is more efficient when the compiler can see that the `basic`
-      // flag is set.
-      UnboundConversion conv;
-      conv.conv = tag.as_conv();
-      conv.arg_position = ++next_arg;
-      if (ABSL_PREDICT_FALSE(
-              !consumer.ConvertOne(conv, string_view(percent + 1, 1)))) {
-        return false;
-      }
-    } else if (percent[1] != '%') {
-      UnboundConversion conv;
-      p = ConsumeUnboundConversionNoInline(percent + 1, end, &conv, &next_arg);
-      if (ABSL_PREDICT_FALSE(p == nullptr)) return false;
-      if (ABSL_PREDICT_FALSE(!consumer.ConvertOne(
-              conv, string_view(percent + 1,
-                                static_cast<size_t>(p - (percent + 1)))))) {
-        return false;
-      }
-    } else {
-      if (ABSL_PREDICT_FALSE(!consumer.Append("%"))) return false;
-      p = percent + 2;
-      continue;
-    }
-  }
-  return true;
-}
-
-// Always returns true, or fails to compile in a constexpr context if s does not
-// point to a constexpr char array.
-constexpr bool EnsureConstexpr(string_view s) {
-  return s.empty() || s[0] == s[0];
-}
-
-class ParsedFormatBase {
- public:
-  explicit ParsedFormatBase(
-      string_view format, bool allow_ignored,
-      std::initializer_list<FormatConversionCharSet> convs);
-
-  ParsedFormatBase(const ParsedFormatBase& other) { *this = other; }
-
-  ParsedFormatBase(ParsedFormatBase&& other) { *this = std::move(other); }
-
-  ParsedFormatBase& operator=(const ParsedFormatBase& other) {
-    if (this == &other) return *this;
-    has_error_ = other.has_error_;
-    items_ = other.items_;
-    size_t text_size = items_.empty() ? 0 : items_.back().text_end;
-    data_.reset(new char[text_size]);
-    memcpy(data_.get(), other.data_.get(), text_size);
-    return *this;
-  }
-
-  ParsedFormatBase& operator=(ParsedFormatBase&& other) {
-    if (this == &other) return *this;
-    has_error_ = other.has_error_;
-    data_ = std::move(other.data_);
-    items_ = std::move(other.items_);
-    // Reset the vector to make sure the invariants hold.
-    other.items_.clear();
-    return *this;
-  }
-
-  template <typename Consumer>
-  bool ProcessFormat(Consumer consumer) const {
-    const char* const base = data_.get();
-    string_view text(base, 0);
-    for (const auto& item : items_) {
-      const char* const end = text.data() + text.size();
-      text =
-          string_view(end, static_cast<size_t>((base + item.text_end) - end));
-      if (item.is_conversion) {
-        if (!consumer.ConvertOne(item.conv, text)) return false;
-      } else {
-        if (!consumer.Append(text)) return false;
-      }
-    }
-    return !has_error_;
-  }
-
-  bool has_error() const { return has_error_; }
-
- private:
-  // Returns whether the conversions match and if !allow_ignored it verifies
-  // that all conversions are used by the format.
-  bool MatchesConversions(
-      bool allow_ignored,
-      std::initializer_list<FormatConversionCharSet> convs) const;
-
-  struct ParsedFormatConsumer;
-
-  struct ConversionItem {
-    bool is_conversion;
-    // Points to the past-the-end location of this element in the data_ array.
-    size_t text_end;
-    UnboundConversion conv;
-  };
-
-  bool has_error_;
-  std::unique_ptr<char[]> data_;
-  std::vector<ConversionItem> items_;
-};
-
-
-// A value type representing a preparsed format.  These can be created, copied
-// around, and reused to speed up formatting loops.
-// The user must specify through the template arguments the conversion
-// characters used in the format. This will be checked at compile time.
-//
-// This class uses Conv enum values to specify each argument.
-// This allows for more flexibility as you can specify multiple possible
-// conversion characters for each argument.
-// ParsedFormat<char...> is a simplified alias for when the user only
-// needs to specify a single conversion character for each argument.
-//
-// Example:
-//   // Extended format supports multiple characters per argument:
-//   using MyFormat = ExtendedParsedFormat<Conv::d | Conv::x>;
-//   MyFormat GetFormat(bool use_hex) {
-//     if (use_hex) return MyFormat("foo %x bar");
-//     return MyFormat("foo %d bar");
-//   }
-//   // 'format' can be used with any value that supports 'd' and 'x',
-//   // like `int`.
-//   auto format = GetFormat(use_hex);
-//   value = StringF(format, i);
-//
-// This class also supports runtime format checking with the ::New() and
-// ::NewAllowIgnored() factory functions.
-// This is the only API that allows the user to pass a runtime specified format
-// string. These factory functions will return NULL if the format does not match
-// the conversions requested by the user.
-template <FormatConversionCharSet... C>
-class ExtendedParsedFormat : public str_format_internal::ParsedFormatBase {
- public:
-  explicit ExtendedParsedFormat(string_view format)
-#ifdef ABSL_INTERNAL_ENABLE_FORMAT_CHECKER
-      __attribute__((
-          enable_if(str_format_internal::EnsureConstexpr(format),
-                    "Format string is not constexpr."),
-          enable_if(str_format_internal::ValidFormatImpl<C...>(format),
-                    "Format specified does not match the template arguments.")))
-#endif  // ABSL_INTERNAL_ENABLE_FORMAT_CHECKER
-      : ExtendedParsedFormat(format, false) {
-  }
-
-  // ExtendedParsedFormat factory function.
-  // The user still has to specify the conversion characters, but they will not
-  // be checked at compile time. Instead, it will be checked at runtime.
-  // This delays the checking to runtime, but allows the user to pass
-  // dynamically sourced formats.
-  // It returns NULL if the format does not match the conversion characters.
-  // The user is responsible for checking the return value before using it.
-  //
-  // The 'New' variant will check that all the specified arguments are being
-  // consumed by the format and return NULL if any argument is being ignored.
-  // The 'NewAllowIgnored' variant will not verify this and will allow formats
-  // that ignore arguments.
-  static std::unique_ptr<ExtendedParsedFormat> New(string_view format) {
-    return New(format, false);
-  }
-  static std::unique_ptr<ExtendedParsedFormat> NewAllowIgnored(
-      string_view format) {
-    return New(format, true);
-  }
-
- private:
-  static std::unique_ptr<ExtendedParsedFormat> New(string_view format,
-                                                   bool allow_ignored) {
-    std::unique_ptr<ExtendedParsedFormat> conv(
-        new ExtendedParsedFormat(format, allow_ignored));
-    if (conv->has_error()) return nullptr;
-    return conv;
-  }
-
-  ExtendedParsedFormat(string_view s, bool allow_ignored)
-      : ParsedFormatBase(s, allow_ignored, {C...}) {}
-};
-}  // namespace str_format_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_STR_FORMAT_PARSER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_join_internal.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_join_internal.h
deleted file mode 100644
index 96d41b68..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_join_internal.h
+++ /dev/null
@@ -1,325 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-// This file declares INTERNAL parts of the Join API that are inlined/templated
-// or otherwise need to be available at compile time. The main abstractions
-// defined in this file are:
-//
-//   - A handful of default Formatters
-//   - JoinAlgorithm() overloads
-//   - JoinRange() overloads
-//   - JoinTuple()
-//
-// DO NOT INCLUDE THIS FILE DIRECTLY. Use this file by including
-// absl/strings/str_join.h
-//
-// IWYU pragma: private, include "absl/strings/str_join.h"
-
-#ifndef ABSL_STRINGS_INTERNAL_STR_JOIN_INTERNAL_H_
-#define ABSL_STRINGS_INTERNAL_STR_JOIN_INTERNAL_H_
-
-#include <cstdint>
-#include <cstring>
-#include <iterator>
-#include <limits>
-#include <memory>
-#include <string>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/internal/raw_logging.h"
-#include "absl/strings/internal/ostringstream.h"
-#include "absl/strings/internal/resize_uninitialized.h"
-#include "absl/strings/str_cat.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace strings_internal {
-
-//
-// Formatter objects
-//
-// The following are implementation classes for standard Formatter objects. The
-// factory functions that users will call to create and use these formatters are
-// defined and documented in strings/join.h.
-//
-
-// The default formatter. Converts alpha-numeric types to strings.
-struct AlphaNumFormatterImpl {
-  // This template is needed in order to support passing in a dereferenced
-  // vector<bool>::iterator
-  template <typename T>
-  void operator()(std::string* out, const T& t) const {
-    StrAppend(out, AlphaNum(t));
-  }
-
-  void operator()(std::string* out, const AlphaNum& t) const {
-    StrAppend(out, t);
-  }
-};
-
-// A type that's used to overload the JoinAlgorithm() function (defined below)
-// for ranges that do not require additional formatting (e.g., a range of
-// strings).
-
-struct NoFormatter : public AlphaNumFormatterImpl {};
-
-// Formats types to strings using the << operator.
-class StreamFormatterImpl {
- public:
-  // The method isn't const because it mutates state. Making it const will
-  // render StreamFormatterImpl thread-hostile.
-  template <typename T>
-  void operator()(std::string* out, const T& t) {
-    // The stream is created lazily to avoid paying the relatively high cost
-    // of its construction when joining an empty range.
-    if (strm_) {
-      strm_->clear();  // clear the bad, fail and eof bits in case they were set
-      strm_->str(out);
-    } else {
-      strm_.reset(new strings_internal::OStringStream(out));
-    }
-    *strm_ << t;
-  }
-
- private:
-  std::unique_ptr<strings_internal::OStringStream> strm_;
-};
-
-// Formats a std::pair<>. The 'first' member is formatted using f1_ and the
-// 'second' member is formatted using f2_. sep_ is the separator.
-template <typename F1, typename F2>
-class PairFormatterImpl {
- public:
-  PairFormatterImpl(F1 f1, absl::string_view sep, F2 f2)
-      : f1_(std::move(f1)), sep_(sep), f2_(std::move(f2)) {}
-
-  template <typename T>
-  void operator()(std::string* out, const T& p) {
-    f1_(out, p.first);
-    out->append(sep_);
-    f2_(out, p.second);
-  }
-
-  template <typename T>
-  void operator()(std::string* out, const T& p) const {
-    f1_(out, p.first);
-    out->append(sep_);
-    f2_(out, p.second);
-  }
-
- private:
-  F1 f1_;
-  std::string sep_;
-  F2 f2_;
-};
-
-// Wraps another formatter and dereferences the argument to operator() then
-// passes the dereferenced argument to the wrapped formatter. This can be
-// useful, for example, to join a std::vector<int*>.
-template <typename Formatter>
-class DereferenceFormatterImpl {
- public:
-  DereferenceFormatterImpl() : f_() {}
-  explicit DereferenceFormatterImpl(Formatter&& f)
-      : f_(std::forward<Formatter>(f)) {}
-
-  template <typename T>
-  void operator()(std::string* out, const T& t) {
-    f_(out, *t);
-  }
-
-  template <typename T>
-  void operator()(std::string* out, const T& t) const {
-    f_(out, *t);
-  }
-
- private:
-  Formatter f_;
-};
-
-// DefaultFormatter<T> is a traits class that selects a default Formatter to use
-// for the given type T. The ::Type member names the Formatter to use. This is
-// used by the strings::Join() functions that do NOT take a Formatter argument,
-// in which case a default Formatter must be chosen.
-//
-// AlphaNumFormatterImpl is the default in the base template, followed by
-// specializations for other types.
-template <typename ValueType>
-struct DefaultFormatter {
-  typedef AlphaNumFormatterImpl Type;
-};
-template <>
-struct DefaultFormatter<const char*> {
-  typedef AlphaNumFormatterImpl Type;
-};
-template <>
-struct DefaultFormatter<char*> {
-  typedef AlphaNumFormatterImpl Type;
-};
-template <>
-struct DefaultFormatter<std::string> {
-  typedef NoFormatter Type;
-};
-template <>
-struct DefaultFormatter<absl::string_view> {
-  typedef NoFormatter Type;
-};
-template <typename ValueType>
-struct DefaultFormatter<ValueType*> {
-  typedef DereferenceFormatterImpl<typename DefaultFormatter<ValueType>::Type>
-      Type;
-};
-
-template <typename ValueType>
-struct DefaultFormatter<std::unique_ptr<ValueType>>
-    : public DefaultFormatter<ValueType*> {};
-
-//
-// JoinAlgorithm() functions
-//
-
-// The main joining algorithm. This simply joins the elements in the given
-// iterator range, each separated by the given separator, into an output string,
-// and formats each element using the provided Formatter object.
-template <typename Iterator, typename Formatter>
-std::string JoinAlgorithm(Iterator start, Iterator end, absl::string_view s,
-                          Formatter&& f) {
-  std::string result;
-  absl::string_view sep("");
-  for (Iterator it = start; it != end; ++it) {
-    result.append(sep.data(), sep.size());
-    f(&result, *it);
-    sep = s;
-  }
-  return result;
-}
-
-// A joining algorithm that's optimized for a forward iterator range of
-// string-like objects that do not need any additional formatting. This is to
-// optimize the common case of joining, say, a std::vector<string> or a
-// std::vector<absl::string_view>.
-//
-// This is an overload of the previous JoinAlgorithm() function. Here the
-// Formatter argument is of type NoFormatter. Since NoFormatter is an internal
-// type, this overload is only invoked when strings::Join() is called with a
-// range of string-like objects (e.g., std::string, absl::string_view), and an
-// explicit Formatter argument was NOT specified.
-//
-// The optimization is that the needed space will be reserved in the output
-// string to avoid the need to resize while appending. To do this, the iterator
-// range will be traversed twice: once to calculate the total needed size, and
-// then again to copy the elements and delimiters to the output string.
-template <typename Iterator,
-          typename = typename std::enable_if<std::is_convertible<
-              typename std::iterator_traits<Iterator>::iterator_category,
-              std::forward_iterator_tag>::value>::type>
-std::string JoinAlgorithm(Iterator start, Iterator end, absl::string_view s,
-                          NoFormatter) {
-  std::string result;
-  if (start != end) {
-    // Sums size
-    auto&& start_value = *start;
-    // Use uint64_t to prevent size_t overflow. We assume it is not possible for
-    // in memory strings to overflow a uint64_t.
-    uint64_t result_size = start_value.size();
-    for (Iterator it = start; ++it != end;) {
-      result_size += s.size();
-      result_size += (*it).size();
-    }
-
-    if (result_size > 0) {
-      constexpr uint64_t kMaxSize =
-          uint64_t{(std::numeric_limits<size_t>::max)()};
-      ABSL_INTERNAL_CHECK(result_size <= kMaxSize, "size_t overflow");
-      STLStringResizeUninitialized(&result, static_cast<size_t>(result_size));
-
-      // Joins strings
-      char* result_buf = &*result.begin();
-
-      memcpy(result_buf, start_value.data(), start_value.size());
-      result_buf += start_value.size();
-      for (Iterator it = start; ++it != end;) {
-        memcpy(result_buf, s.data(), s.size());
-        result_buf += s.size();
-        auto&& value = *it;
-        memcpy(result_buf, value.data(), value.size());
-        result_buf += value.size();
-      }
-    }
-  }
-
-  return result;
-}
-
-// JoinTupleLoop implements a loop over the elements of a std::tuple, which
-// are heterogeneous. The primary template matches the tuple interior case. It
-// continues the iteration after appending a separator (for nonzero indices)
-// and formatting an element of the tuple. The specialization for the I=N case
-// matches the end-of-tuple, and terminates the iteration.
-template <size_t I, size_t N>
-struct JoinTupleLoop {
-  template <typename Tup, typename Formatter>
-  void operator()(std::string* out, const Tup& tup, absl::string_view sep,
-                  Formatter&& fmt) {
-    if (I > 0) out->append(sep.data(), sep.size());
-    fmt(out, std::get<I>(tup));
-    JoinTupleLoop<I + 1, N>()(out, tup, sep, fmt);
-  }
-};
-template <size_t N>
-struct JoinTupleLoop<N, N> {
-  template <typename Tup, typename Formatter>
-  void operator()(std::string*, const Tup&, absl::string_view, Formatter&&) {}
-};
-
-template <typename... T, typename Formatter>
-std::string JoinAlgorithm(const std::tuple<T...>& tup, absl::string_view sep,
-                          Formatter&& fmt) {
-  std::string result;
-  JoinTupleLoop<0, sizeof...(T)>()(&result, tup, sep, fmt);
-  return result;
-}
-
-template <typename Iterator>
-std::string JoinRange(Iterator first, Iterator last,
-                      absl::string_view separator) {
-  // No formatter was explicitly given, so a default must be chosen.
-  typedef typename std::iterator_traits<Iterator>::value_type ValueType;
-  typedef typename DefaultFormatter<ValueType>::Type Formatter;
-  return JoinAlgorithm(first, last, separator, Formatter());
-}
-
-template <typename Range, typename Formatter>
-std::string JoinRange(const Range& range, absl::string_view separator,
-                      Formatter&& fmt) {
-  using std::begin;
-  using std::end;
-  return JoinAlgorithm(begin(range), end(range), separator, fmt);
-}
-
-template <typename Range>
-std::string JoinRange(const Range& range, absl::string_view separator) {
-  using std::begin;
-  using std::end;
-  return JoinRange(begin(range), end(range), separator);
-}
-
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_STR_JOIN_INTERNAL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_split_internal.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_split_internal.h
deleted file mode 100644
index 11ea96f2..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/str_split_internal.h
+++ /dev/null
@@ -1,485 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-// This file declares INTERNAL parts of the Split API that are inline/templated
-// or otherwise need to be available at compile time. The main abstractions
-// defined in here are
-//
-//   - ConvertibleToStringView
-//   - SplitIterator<>
-//   - Splitter<>
-//
-// DO NOT INCLUDE THIS FILE DIRECTLY. Use this file by including
-// absl/strings/str_split.h.
-//
-// IWYU pragma: private, include "absl/strings/str_split.h"
-
-#ifndef ABSL_STRINGS_INTERNAL_STR_SPLIT_INTERNAL_H_
-#define ABSL_STRINGS_INTERNAL_STR_SPLIT_INTERNAL_H_
-
-#include <array>
-#include <cstddef>
-#include <initializer_list>
-#include <iterator>
-#include <tuple>
-#include <type_traits>
-#include <utility>
-#include <vector>
-
-#include "absl/base/macros.h"
-#include "absl/base/port.h"
-#include "absl/meta/type_traits.h"
-#include "absl/strings/string_view.h"
-
-#ifdef _GLIBCXX_DEBUG
-#include "absl/strings/internal/stl_type_traits.h"
-#endif  // _GLIBCXX_DEBUG
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace strings_internal {
-
-// This class is implicitly constructible from everything that absl::string_view
-// is implicitly constructible from, except for rvalue strings.  This means it
-// can be used as a function parameter in places where passing a temporary
-// string might cause memory lifetime issues.
-class ConvertibleToStringView {
- public:
-  ConvertibleToStringView(const char* s)  // NOLINT(runtime/explicit)
-      : value_(s) {}
-  ConvertibleToStringView(char* s) : value_(s) {}  // NOLINT(runtime/explicit)
-  ConvertibleToStringView(absl::string_view s)     // NOLINT(runtime/explicit)
-      : value_(s) {}
-  ConvertibleToStringView(const std::string& s)  // NOLINT(runtime/explicit)
-      : value_(s) {}
-
-  // Disable conversion from rvalue strings.
-  ConvertibleToStringView(std::string&& s) = delete;
-  ConvertibleToStringView(const std::string&& s) = delete;
-
-  absl::string_view value() const { return value_; }
-
- private:
-  absl::string_view value_;
-};
-
-// An iterator that enumerates the parts of a string from a Splitter. The text
-// to be split, the Delimiter, and the Predicate are all taken from the given
-// Splitter object. Iterators may only be compared if they refer to the same
-// Splitter instance.
-//
-// This class is NOT part of the public splitting API.
-template <typename Splitter>
-class SplitIterator {
- public:
-  using iterator_category = std::input_iterator_tag;
-  using value_type = absl::string_view;
-  using difference_type = ptrdiff_t;
-  using pointer = const value_type*;
-  using reference = const value_type&;
-
-  enum State { kInitState, kLastState, kEndState };
-  SplitIterator(State state, const Splitter* splitter)
-      : pos_(0),
-        state_(state),
-        splitter_(splitter),
-        delimiter_(splitter->delimiter()),
-        predicate_(splitter->predicate()) {
-    // Hack to maintain backward compatibility. This one block makes it so an
-    // empty absl::string_view whose .data() happens to be nullptr behaves
-    // *differently* from an otherwise empty absl::string_view whose .data() is
-    // not nullptr. This is an undesirable difference in general, but this
-    // behavior is maintained to avoid breaking existing code that happens to
-    // depend on this old behavior/bug. Perhaps it will be fixed one day. The
-    // difference in behavior is as follows:
-    //   Split(absl::string_view(""), '-');  // {""}
-    //   Split(absl::string_view(), '-');    // {}
-    if (splitter_->text().data() == nullptr) {
-      state_ = kEndState;
-      pos_ = splitter_->text().size();
-      return;
-    }
-
-    if (state_ == kEndState) {
-      pos_ = splitter_->text().size();
-    } else {
-      ++(*this);
-    }
-  }
-
-  bool at_end() const { return state_ == kEndState; }
-
-  reference operator*() const { return curr_; }
-  pointer operator->() const { return &curr_; }
-
-  SplitIterator& operator++() {
-    do {
-      if (state_ == kLastState) {
-        state_ = kEndState;
-        return *this;
-      }
-      const absl::string_view text = splitter_->text();
-      const absl::string_view d = delimiter_.Find(text, pos_);
-      if (d.data() == text.data() + text.size()) state_ = kLastState;
-      curr_ = text.substr(pos_,
-                          static_cast<size_t>(d.data() - (text.data() + pos_)));
-      pos_ += curr_.size() + d.size();
-    } while (!predicate_(curr_));
-    return *this;
-  }
-
-  SplitIterator operator++(int) {
-    SplitIterator old(*this);
-    ++(*this);
-    return old;
-  }
-
-  friend bool operator==(const SplitIterator& a, const SplitIterator& b) {
-    return a.state_ == b.state_ && a.pos_ == b.pos_;
-  }
-
-  friend bool operator!=(const SplitIterator& a, const SplitIterator& b) {
-    return !(a == b);
-  }
-
- private:
-  size_t pos_;
-  State state_;
-  absl::string_view curr_;
-  const Splitter* splitter_;
-  typename Splitter::DelimiterType delimiter_;
-  typename Splitter::PredicateType predicate_;
-};
-
-// HasMappedType<T>::value is true iff there exists a type T::mapped_type.
-template <typename T, typename = void>
-struct HasMappedType : std::false_type {};
-template <typename T>
-struct HasMappedType<T, absl::void_t<typename T::mapped_type>>
-    : std::true_type {};
-
-// HasValueType<T>::value is true iff there exists a type T::value_type.
-template <typename T, typename = void>
-struct HasValueType : std::false_type {};
-template <typename T>
-struct HasValueType<T, absl::void_t<typename T::value_type>> : std::true_type {
-};
-
-// HasConstIterator<T>::value is true iff there exists a type T::const_iterator.
-template <typename T, typename = void>
-struct HasConstIterator : std::false_type {};
-template <typename T>
-struct HasConstIterator<T, absl::void_t<typename T::const_iterator>>
-    : std::true_type {};
-
-// HasEmplace<T>::value is true iff there exists a method T::emplace().
-template <typename T, typename = void>
-struct HasEmplace : std::false_type {};
-template <typename T>
-struct HasEmplace<T, absl::void_t<decltype(std::declval<T>().emplace())>>
-    : std::true_type {};
-
-// IsInitializerList<T>::value is true iff T is an std::initializer_list. More
-// details below in Splitter<> where this is used.
-std::false_type IsInitializerListDispatch(...);  // default: No
-template <typename T>
-std::true_type IsInitializerListDispatch(std::initializer_list<T>*);
-template <typename T>
-struct IsInitializerList
-    : decltype(IsInitializerListDispatch(static_cast<T*>(nullptr))) {};
-
-// A SplitterIsConvertibleTo<C>::type alias exists iff the specified condition
-// is true for type 'C'.
-//
-// Restricts conversion to container-like types (by testing for the presence of
-// a const_iterator member type) and also to disable conversion to an
-// std::initializer_list (which also has a const_iterator). Otherwise, code
-// compiled in C++11 will get an error due to ambiguous conversion paths (in
-// C++11 std::vector<T>::operator= is overloaded to take either a std::vector<T>
-// or an std::initializer_list<T>).
-
-template <typename C, bool has_value_type, bool has_mapped_type>
-struct SplitterIsConvertibleToImpl : std::false_type {};
-
-template <typename C>
-struct SplitterIsConvertibleToImpl<C, true, false>
-    : std::is_constructible<typename C::value_type, absl::string_view> {};
-
-template <typename C>
-struct SplitterIsConvertibleToImpl<C, true, true>
-    : absl::conjunction<
-          std::is_constructible<typename C::key_type, absl::string_view>,
-          std::is_constructible<typename C::mapped_type, absl::string_view>> {};
-
-template <typename C>
-struct SplitterIsConvertibleTo
-    : SplitterIsConvertibleToImpl<
-          C,
-#ifdef _GLIBCXX_DEBUG
-          !IsStrictlyBaseOfAndConvertibleToSTLContainer<C>::value &&
-#endif  // _GLIBCXX_DEBUG
-              !IsInitializerList<
-                  typename std::remove_reference<C>::type>::value &&
-              HasValueType<C>::value && HasConstIterator<C>::value,
-          HasMappedType<C>::value> {
-};
-
-template <typename StringType, typename Container, typename = void>
-struct ShouldUseLifetimeBound : std::false_type {};
-
-template <typename StringType, typename Container>
-struct ShouldUseLifetimeBound<
-    StringType, Container,
-    std::enable_if_t<
-        std::is_same<StringType, std::string>::value &&
-        std::is_same<typename Container::value_type, absl::string_view>::value>>
-    : std::true_type {};
-
-template <typename StringType, typename First, typename Second>
-using ShouldUseLifetimeBoundForPair = std::integral_constant<
-    bool, std::is_same<StringType, std::string>::value &&
-              (std::is_same<First, absl::string_view>::value ||
-               std::is_same<Second, absl::string_view>::value)>;
-
-
-// This class implements the range that is returned by absl::StrSplit(). This
-// class has templated conversion operators that allow it to be implicitly
-// converted to a variety of types that the caller may have specified on the
-// left-hand side of an assignment.
-//
-// The main interface for interacting with this class is through its implicit
-// conversion operators. However, this class may also be used like a container
-// in that it has .begin() and .end() member functions. It may also be used
-// within a range-for loop.
-//
-// Output containers can be collections of any type that is constructible from
-// an absl::string_view.
-//
-// An Predicate functor may be supplied. This predicate will be used to filter
-// the split strings: only strings for which the predicate returns true will be
-// kept. A Predicate object is any unary functor that takes an absl::string_view
-// and returns bool.
-//
-// The StringType parameter can be either string_view or string, depending on
-// whether the Splitter refers to a string stored elsewhere, or if the string
-// resides inside the Splitter itself.
-template <typename Delimiter, typename Predicate, typename StringType>
-class Splitter {
- public:
-  using DelimiterType = Delimiter;
-  using PredicateType = Predicate;
-  using const_iterator = strings_internal::SplitIterator<Splitter>;
-  using value_type = typename std::iterator_traits<const_iterator>::value_type;
-
-  Splitter(StringType input_text, Delimiter d, Predicate p)
-      : text_(std::move(input_text)),
-        delimiter_(std::move(d)),
-        predicate_(std::move(p)) {}
-
-  absl::string_view text() const { return text_; }
-  const Delimiter& delimiter() const { return delimiter_; }
-  const Predicate& predicate() const { return predicate_; }
-
-  // Range functions that iterate the split substrings as absl::string_view
-  // objects. These methods enable a Splitter to be used in a range-based for
-  // loop.
-  const_iterator begin() const { return {const_iterator::kInitState, this}; }
-  const_iterator end() const { return {const_iterator::kEndState, this}; }
-
-  // An implicit conversion operator that is restricted to only those containers
-  // that the splitter is convertible to.
-  template <
-      typename Container,
-      std::enable_if_t<ShouldUseLifetimeBound<StringType, Container>::value &&
-                           SplitterIsConvertibleTo<Container>::value,
-                       std::nullptr_t> = nullptr>
-  // NOLINTNEXTLINE(google-explicit-constructor)
-  operator Container() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return ConvertToContainer<Container, typename Container::value_type,
-                              HasMappedType<Container>::value>()(*this);
-  }
-
-  template <
-      typename Container,
-      std::enable_if_t<!ShouldUseLifetimeBound<StringType, Container>::value &&
-                           SplitterIsConvertibleTo<Container>::value,
-                       std::nullptr_t> = nullptr>
-  // NOLINTNEXTLINE(google-explicit-constructor)
-  operator Container() const {
-    return ConvertToContainer<Container, typename Container::value_type,
-                              HasMappedType<Container>::value>()(*this);
-  }
-
-  // Returns a pair with its .first and .second members set to the first two
-  // strings returned by the begin() iterator. Either/both of .first and .second
-  // will be constructed with empty strings if the iterator doesn't have a
-  // corresponding value.
-  template <typename First, typename Second,
-            std::enable_if_t<
-                ShouldUseLifetimeBoundForPair<StringType, First, Second>::value,
-                std::nullptr_t> = nullptr>
-  // NOLINTNEXTLINE(google-explicit-constructor)
-  operator std::pair<First, Second>() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return ConvertToPair<First, Second>();
-  }
-
-  template <typename First, typename Second,
-            std::enable_if_t<!ShouldUseLifetimeBoundForPair<StringType, First,
-                                                            Second>::value,
-                             std::nullptr_t> = nullptr>
-  // NOLINTNEXTLINE(google-explicit-constructor)
-  operator std::pair<First, Second>() const {
-    return ConvertToPair<First, Second>();
-  }
-
- private:
-  template <typename First, typename Second>
-  std::pair<First, Second> ConvertToPair() const {
-    absl::string_view first, second;
-    auto it = begin();
-    if (it != end()) {
-      first = *it;
-      if (++it != end()) {
-        second = *it;
-      }
-    }
-    return {First(first), Second(second)};
-  }
-
-  // ConvertToContainer is a functor converting a Splitter to the requested
-  // Container of ValueType. It is specialized below to optimize splitting to
-  // certain combinations of Container and ValueType.
-  //
-  // This base template handles the generic case of storing the split results in
-  // the requested non-map-like container and converting the split substrings to
-  // the requested type.
-  template <typename Container, typename ValueType, bool is_map = false>
-  struct ConvertToContainer {
-    Container operator()(const Splitter& splitter) const {
-      Container c;
-      auto it = std::inserter(c, c.end());
-      for (const auto& sp : splitter) {
-        *it++ = ValueType(sp);
-      }
-      return c;
-    }
-  };
-
-  // Partial specialization for a std::vector<absl::string_view>.
-  //
-  // Optimized for the common case of splitting to a
-  // std::vector<absl::string_view>. In this case we first split the results to
-  // a small array of absl::string_view on the stack, to reduce reallocations.
-  template <typename A>
-  struct ConvertToContainer<std::vector<absl::string_view, A>,
-                            absl::string_view, false> {
-    std::vector<absl::string_view, A> operator()(
-        const Splitter& splitter) const {
-      struct raw_view {
-        const char* data;
-        size_t size;
-        operator absl::string_view() const {  // NOLINT(runtime/explicit)
-          return {data, size};
-        }
-      };
-      std::vector<absl::string_view, A> v;
-      std::array<raw_view, 16> ar;
-      for (auto it = splitter.begin(); !it.at_end();) {
-        size_t index = 0;
-        do {
-          ar[index].data = it->data();
-          ar[index].size = it->size();
-          ++it;
-        } while (++index != ar.size() && !it.at_end());
-        // We static_cast index to a signed type to work around overzealous
-        // compiler warnings about signedness.
-        v.insert(v.end(), ar.begin(),
-                 ar.begin() + static_cast<ptrdiff_t>(index));
-      }
-      return v;
-    }
-  };
-
-  // Partial specialization for a std::vector<std::string>.
-  //
-  // Optimized for the common case of splitting to a std::vector<std::string>.
-  // In this case we first split the results to a std::vector<absl::string_view>
-  // so the returned std::vector<std::string> can have space reserved to avoid
-  // std::string moves.
-  template <typename A>
-  struct ConvertToContainer<std::vector<std::string, A>, std::string, false> {
-    std::vector<std::string, A> operator()(const Splitter& splitter) const {
-      const std::vector<absl::string_view> v = splitter;
-      return std::vector<std::string, A>(v.begin(), v.end());
-    }
-  };
-
-  // Partial specialization for containers of pairs (e.g., maps).
-  //
-  // The algorithm is to insert a new pair into the map for each even-numbered
-  // item, with the even-numbered item as the key with a default-constructed
-  // value. Each odd-numbered item will then be assigned to the last pair's
-  // value.
-  template <typename Container, typename First, typename Second>
-  struct ConvertToContainer<Container, std::pair<const First, Second>, true> {
-    using iterator = typename Container::iterator;
-
-    Container operator()(const Splitter& splitter) const {
-      Container m;
-      iterator it;
-      bool insert = true;
-      for (const absl::string_view sv : splitter) {
-        if (insert) {
-          it = InsertOrEmplace(&m, sv);
-        } else {
-          it->second = Second(sv);
-        }
-        insert = !insert;
-      }
-      return m;
-    }
-
-    // Inserts the key and an empty value into the map, returning an iterator to
-    // the inserted item. We use emplace() if available, otherwise insert().
-    template <typename M>
-    static absl::enable_if_t<HasEmplace<M>::value, iterator> InsertOrEmplace(
-        M* m, absl::string_view key) {
-      // Use piecewise_construct to support old versions of gcc in which pair
-      // constructor can't otherwise construct string from string_view.
-      return ToIter(m->emplace(std::piecewise_construct, std::make_tuple(key),
-                               std::tuple<>()));
-    }
-    template <typename M>
-    static absl::enable_if_t<!HasEmplace<M>::value, iterator> InsertOrEmplace(
-        M* m, absl::string_view key) {
-      return ToIter(m->insert(std::make_pair(First(key), Second(""))));
-    }
-
-    static iterator ToIter(std::pair<iterator, bool> pair) {
-      return pair.first;
-    }
-    static iterator ToIter(iterator iter) { return iter; }
-  };
-
-  StringType text_;
-  Delimiter delimiter_;
-  Predicate predicate_;
-};
-
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_STR_SPLIT_INTERNAL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/string_constant.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/string_constant.h
deleted file mode 100644
index f68b17d7..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/string_constant.h
+++ /dev/null
@@ -1,72 +0,0 @@
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_STRING_CONSTANT_H_
-#define ABSL_STRINGS_INTERNAL_STRING_CONSTANT_H_
-
-#include "absl/meta/type_traits.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace strings_internal {
-
-// StringConstant<T> represents a compile time string constant.
-// It can be accessed via its `absl::string_view value` static member.
-// It is guaranteed that the `string_view` returned has constant `.data()`,
-// constant `.size()` and constant `value[i]` for all `0 <= i < .size()`
-//
-// The `T` is an opaque type. It is guaranteed that different string constants
-// will have different values of `T`. This allows users to associate the string
-// constant with other static state at compile time.
-//
-// Instances should be made using the `MakeStringConstant()` factory function
-// below.
-template <typename T>
-struct StringConstant {
- private:
-  static constexpr bool TryConstexprEval(absl::string_view view) {
-    return view.empty() || 2 * view[0] != 1;
-  }
-
- public:
-  static constexpr absl::string_view value = T{}();
-  constexpr absl::string_view operator()() const { return value; }
-
-  // Check to be sure `view` points to constant data.
-  // Otherwise, it can't be constant evaluated.
-  static_assert(TryConstexprEval(value),
-                "The input string_view must point to constant data.");
-};
-
-#ifdef ABSL_INTERNAL_NEED_REDUNDANT_CONSTEXPR_DECL
-template <typename T>
-constexpr absl::string_view StringConstant<T>::value;
-#endif
-
-// Factory function for `StringConstant` instances.
-// It supports callables that have a constexpr default constructor and a
-// constexpr operator().
-// It must return an `absl::string_view` or `const char*` pointing to constant
-// data. This is validated at compile time.
-template <typename T>
-constexpr StringConstant<T> MakeStringConstant(T) {
-  return {};
-}
-
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_STRING_CONSTANT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/stringify_sink.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/stringify_sink.h
deleted file mode 100644
index fc3747bb..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/stringify_sink.h
+++ /dev/null
@@ -1,57 +0,0 @@
-// Copyright 2022 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_STRINGS_INTERNAL_STRINGIFY_SINK_H_
-#define ABSL_STRINGS_INTERNAL_STRINGIFY_SINK_H_
-
-#include <string>
-#include <type_traits>
-#include <utility>
-
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-namespace strings_internal {
-class StringifySink {
- public:
-  void Append(size_t count, char ch);
-
-  void Append(string_view v);
-
-  // Support `absl::Format(&sink, format, args...)`.
-  friend void AbslFormatFlush(StringifySink* sink, absl::string_view v) {
-    sink->Append(v);
-  }
-
- private:
-  template <typename T>
-  friend string_view ExtractStringification(StringifySink& sink, const T& v);
-
-  std::string buffer_;
-};
-
-template <typename T>
-string_view ExtractStringification(StringifySink& sink, const T& v) {
-  AbslStringify(sink, v);
-  return sink.buffer_;
-}
-
-}  // namespace strings_internal
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_STRINGIFY_SINK_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/utf8.h b/projects/llm_framework/include/abseil-cpp/absl/strings/internal/utf8.h
deleted file mode 100644
index 32fb1093..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/internal/utf8.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// UTF8 utilities, implemented to reduce dependencies.
-
-#ifndef ABSL_STRINGS_INTERNAL_UTF8_H_
-#define ABSL_STRINGS_INTERNAL_UTF8_H_
-
-#include <cstddef>
-#include <cstdint>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace strings_internal {
-
-// For Unicode code points 0 through 0x10FFFF, EncodeUTF8Char writes
-// out the UTF-8 encoding into buffer, and returns the number of chars
-// it wrote.
-//
-// As described in https://tools.ietf.org/html/rfc3629#section-3 , the encodings
-// are:
-//    00 -     7F : 0xxxxxxx
-//    80 -    7FF : 110xxxxx 10xxxxxx
-//   800 -   FFFF : 1110xxxx 10xxxxxx 10xxxxxx
-// 10000 - 10FFFF : 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
-//
-// Values greater than 0x10FFFF are not supported and may or may not write
-// characters into buffer, however never will more than kMaxEncodedUTF8Size
-// bytes be written, regardless of the value of utf8_char.
-enum { kMaxEncodedUTF8Size = 4 };
-size_t EncodeUTF8Char(char *buffer, char32_t utf8_char);
-
-}  // namespace strings_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_INTERNAL_UTF8_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/match.h b/projects/llm_framework/include/abseil-cpp/absl/strings/match.h
deleted file mode 100644
index 1eeafbbf..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/match.h
+++ /dev/null
@@ -1,119 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: match.h
-// -----------------------------------------------------------------------------
-//
-// This file contains simple utilities for performing string matching checks.
-// All of these function parameters are specified as `absl::string_view`,
-// meaning that these functions can accept `std::string`, `absl::string_view` or
-// NUL-terminated C-style strings.
-//
-// Examples:
-//   std::string s = "foo";
-//   absl::string_view sv = "f";
-//   assert(absl::StrContains(s, sv));
-//
-// Note: The order of parameters in these functions is designed to mimic the
-// order an equivalent member function would exhibit;
-// e.g. `s.Contains(x)` ==> `absl::StrContains(s, x).
-#ifndef ABSL_STRINGS_MATCH_H_
-#define ABSL_STRINGS_MATCH_H_
-
-#include <cstring>
-
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// StrContains()
-//
-// Returns whether a given string `haystack` contains the substring `needle`.
-inline bool StrContains(absl::string_view haystack,
-                        absl::string_view needle) noexcept {
-  return haystack.find(needle, 0) != haystack.npos;
-}
-
-inline bool StrContains(absl::string_view haystack, char needle) noexcept {
-  return haystack.find(needle) != haystack.npos;
-}
-
-// StartsWith()
-//
-// Returns whether a given string `text` begins with `prefix`.
-inline bool StartsWith(absl::string_view text,
-                       absl::string_view prefix) noexcept {
-  return prefix.empty() ||
-         (text.size() >= prefix.size() &&
-          memcmp(text.data(), prefix.data(), prefix.size()) == 0);
-}
-
-// EndsWith()
-//
-// Returns whether a given string `text` ends with `suffix`.
-inline bool EndsWith(absl::string_view text,
-                     absl::string_view suffix) noexcept {
-  return suffix.empty() ||
-         (text.size() >= suffix.size() &&
-          memcmp(text.data() + (text.size() - suffix.size()), suffix.data(),
-                 suffix.size()) == 0);
-}
-// StrContainsIgnoreCase()
-//
-// Returns whether a given ASCII string `haystack` contains the ASCII substring
-// `needle`, ignoring case in the comparison.
-bool StrContainsIgnoreCase(absl::string_view haystack,
-                           absl::string_view needle) noexcept;
-
-bool StrContainsIgnoreCase(absl::string_view haystack,
-                           char needle) noexcept;
-
-// EqualsIgnoreCase()
-//
-// Returns whether given ASCII strings `piece1` and `piece2` are equal, ignoring
-// case in the comparison.
-bool EqualsIgnoreCase(absl::string_view piece1,
-                      absl::string_view piece2) noexcept;
-
-// StartsWithIgnoreCase()
-//
-// Returns whether a given ASCII string `text` starts with `prefix`,
-// ignoring case in the comparison.
-bool StartsWithIgnoreCase(absl::string_view text,
-                          absl::string_view prefix) noexcept;
-
-// EndsWithIgnoreCase()
-//
-// Returns whether a given ASCII string `text` ends with `suffix`, ignoring
-// case in the comparison.
-bool EndsWithIgnoreCase(absl::string_view text,
-                        absl::string_view suffix) noexcept;
-
-// Yields the longest prefix in common between both input strings.
-// Pointer-wise, the returned result is a subset of input "a".
-absl::string_view FindLongestCommonPrefix(absl::string_view a,
-                                          absl::string_view b);
-
-// Yields the longest suffix in common between both input strings.
-// Pointer-wise, the returned result is a subset of input "a".
-absl::string_view FindLongestCommonSuffix(absl::string_view a,
-                                          absl::string_view b);
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_MATCH_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/numbers.h b/projects/llm_framework/include/abseil-cpp/absl/strings/numbers.h
deleted file mode 100644
index ad4e66b6..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/numbers.h
+++ /dev/null
@@ -1,456 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: numbers.h
-// -----------------------------------------------------------------------------
-//
-// This package contains functions for converting strings to numbers. For
-// converting numbers to strings, use `StrCat()` or `StrAppend()` in str_cat.h,
-// which automatically detect and convert most number values appropriately.
-
-#ifndef ABSL_STRINGS_NUMBERS_H_
-#define ABSL_STRINGS_NUMBERS_H_
-
-#ifdef __SSSE3__
-#include <tmmintrin.h>
-#endif
-
-#ifdef _MSC_VER
-#include <intrin.h>
-#endif
-
-#include <cstddef>
-#include <cstdint>
-#include <cstdlib>
-#include <cstring>
-#include <ctime>
-#include <limits>
-#include <string>
-#include <type_traits>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/internal/endian.h"
-#include "absl/base/macros.h"
-#include "absl/base/nullability.h"
-#include "absl/base/optimization.h"
-#include "absl/base/port.h"
-#include "absl/numeric/bits.h"
-#include "absl/numeric/int128.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// SimpleAtoi()
-//
-// Converts the given string (optionally followed or preceded by ASCII
-// whitespace) into an integer value, returning `true` if successful. The string
-// must reflect a base-10 integer whose value falls within the range of the
-// integer type (optionally preceded by a `+` or `-`). If any errors are
-// encountered, this function returns `false`, leaving `out` in an unspecified
-// state.
-template <typename int_type>
-ABSL_MUST_USE_RESULT bool SimpleAtoi(absl::string_view str,
-                                     absl::Nonnull<int_type*> out);
-
-// SimpleAtof()
-//
-// Converts the given string (optionally followed or preceded by ASCII
-// whitespace) into a float, which may be rounded on overflow or underflow,
-// returning `true` if successful.
-// See https://en.cppreference.com/w/c/string/byte/strtof for details about the
-// allowed formats for `str`, except SimpleAtof() is locale-independent and will
-// always use the "C" locale. If any errors are encountered, this function
-// returns `false`, leaving `out` in an unspecified state.
-ABSL_MUST_USE_RESULT bool SimpleAtof(absl::string_view str,
-                                     absl::Nonnull<float*> out);
-
-// SimpleAtod()
-//
-// Converts the given string (optionally followed or preceded by ASCII
-// whitespace) into a double, which may be rounded on overflow or underflow,
-// returning `true` if successful.
-// See https://en.cppreference.com/w/c/string/byte/strtof for details about the
-// allowed formats for `str`, except SimpleAtod is locale-independent and will
-// always use the "C" locale. If any errors are encountered, this function
-// returns `false`, leaving `out` in an unspecified state.
-ABSL_MUST_USE_RESULT bool SimpleAtod(absl::string_view str,
-                                     absl::Nonnull<double*> out);
-
-// SimpleAtob()
-//
-// Converts the given string into a boolean, returning `true` if successful.
-// The following case-insensitive strings are interpreted as boolean `true`:
-// "true", "t", "yes", "y", "1". The following case-insensitive strings
-// are interpreted as boolean `false`: "false", "f", "no", "n", "0". If any
-// errors are encountered, this function returns `false`, leaving `out` in an
-// unspecified state.
-ABSL_MUST_USE_RESULT bool SimpleAtob(absl::string_view str,
-                                     absl::Nonnull<bool*> out);
-
-// SimpleHexAtoi()
-//
-// Converts a hexadecimal string (optionally followed or preceded by ASCII
-// whitespace) to an integer, returning `true` if successful. Only valid base-16
-// hexadecimal integers whose value falls within the range of the integer type
-// (optionally preceded by a `+` or `-`) can be converted. A valid hexadecimal
-// value may include both upper and lowercase character symbols, and may
-// optionally include a leading "0x" (or "0X") number prefix, which is ignored
-// by this function. If any errors are encountered, this function returns
-// `false`, leaving `out` in an unspecified state.
-template <typename int_type>
-ABSL_MUST_USE_RESULT bool SimpleHexAtoi(absl::string_view str,
-                                        absl::Nonnull<int_type*> out);
-
-// Overloads of SimpleHexAtoi() for 128 bit integers.
-ABSL_MUST_USE_RESULT inline bool SimpleHexAtoi(
-    absl::string_view str, absl::Nonnull<absl::int128*> out);
-ABSL_MUST_USE_RESULT inline bool SimpleHexAtoi(
-    absl::string_view str, absl::Nonnull<absl::uint128*> out);
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-// End of public API.  Implementation details follow.
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace numbers_internal {
-
-// Digit conversion.
-ABSL_DLL extern const char kHexChar[17];  // 0123456789abcdef
-ABSL_DLL extern const char
-    kHexTable[513];  // 000102030405060708090a0b0c0d0e0f1011...
-
-// Writes a two-character representation of 'i' to 'buf'. 'i' must be in the
-// range 0 <= i < 100, and buf must have space for two characters. Example:
-//   char buf[2];
-//   PutTwoDigits(42, buf);
-//   // buf[0] == '4'
-//   // buf[1] == '2'
-void PutTwoDigits(uint32_t i, absl::Nonnull<char*> buf);
-
-// safe_strto?() functions for implementing SimpleAtoi()
-
-bool safe_strto32_base(absl::string_view text, absl::Nonnull<int32_t*> value,
-                       int base);
-bool safe_strto64_base(absl::string_view text, absl::Nonnull<int64_t*> value,
-                       int base);
-bool safe_strto128_base(absl::string_view text,
-                        absl::Nonnull<absl::int128*> value, int base);
-bool safe_strtou32_base(absl::string_view text, absl::Nonnull<uint32_t*> value,
-                        int base);
-bool safe_strtou64_base(absl::string_view text, absl::Nonnull<uint64_t*> value,
-                        int base);
-bool safe_strtou128_base(absl::string_view text,
-                         absl::Nonnull<absl::uint128*> value, int base);
-
-static const int kFastToBufferSize = 32;
-static const int kSixDigitsToBufferSize = 16;
-
-template <class T>
-std::enable_if_t<!std::is_unsigned<T>::value, bool> IsNegative(const T& v) {
-  return v < T();
-}
-
-template <class T>
-std::enable_if_t<std::is_unsigned<T>::value, std::false_type> IsNegative(
-    const T&) {
-  // The integer is unsigned, so return a compile-time constant.
-  // This can help the optimizer avoid having to prove bool to be false later.
-  return std::false_type();
-}
-
-template <class T>
-std::enable_if_t<std::is_unsigned<std::decay_t<T>>::value, T&&>
-UnsignedAbsoluteValue(T&& v ABSL_ATTRIBUTE_LIFETIME_BOUND) {
-  // The value is unsigned; just return the original.
-  return std::forward<T>(v);
-}
-
-template <class T>
-ABSL_ATTRIBUTE_CONST_FUNCTION
-    std::enable_if_t<!std::is_unsigned<T>::value, std::make_unsigned_t<T>>
-    UnsignedAbsoluteValue(T v) {
-  using U = std::make_unsigned_t<T>;
-  return IsNegative(v) ? U() - static_cast<U>(v) : static_cast<U>(v);
-}
-
-// Returns the number of base-10 digits in the given number.
-// Note that this strictly counts digits. It does not count the sign.
-// The `initial_digits` parameter is the starting point, which is normally equal
-// to 1 because the number of digits in 0 is 1 (a special case).
-// However, callers may e.g. wish to change it to 2 to account for the sign.
-template <typename T>
-std::enable_if_t<std::is_unsigned<T>::value, uint32_t> Base10Digits(
-    T v, const uint32_t initial_digits = 1) {
-  uint32_t r = initial_digits;
-  // If code size becomes an issue, the 'if' stage can be removed for a minor
-  // performance loss.
-  for (;;) {
-    if (ABSL_PREDICT_TRUE(v < 10 * 10)) {
-      r += (v >= 10);
-      break;
-    }
-    if (ABSL_PREDICT_TRUE(v < 1000 * 10)) {
-      r += (v >= 1000) + 2;
-      break;
-    }
-    if (ABSL_PREDICT_TRUE(v < 100000 * 10)) {
-      r += (v >= 100000) + 4;
-      break;
-    }
-    r += 6;
-    v = static_cast<T>(v / 1000000);
-  }
-  return r;
-}
-
-template <typename T>
-std::enable_if_t<std::is_signed<T>::value, uint32_t> Base10Digits(
-    T v, uint32_t r = 1) {
-  // Branchlessly add 1 to account for a minus sign.
-  r += static_cast<uint32_t>(IsNegative(v));
-  return Base10Digits(UnsignedAbsoluteValue(v), r);
-}
-
-// These functions return the number of base-10 digits, but multiplied by -1 if
-// the input itself is negative. This is handy and efficient for later usage,
-// since the bitwise complement of the result becomes equal to the number of
-// characters required.
-ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative(
-    signed char v);
-ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative(
-    unsigned char v);
-ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative(
-    short v);  // NOLINT
-ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative(
-    unsigned short v);  // NOLINT
-ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative(int v);
-ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative(
-    unsigned int v);
-ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative(
-    long v);  // NOLINT
-ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative(
-    unsigned long v);  // NOLINT
-ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative(
-    long long v);  // NOLINT
-ABSL_ATTRIBUTE_CONST_FUNCTION int GetNumDigitsOrNegativeIfNegative(
-    unsigned long long v);  // NOLINT
-
-// Helper function for fast formatting of floating-point values.
-// The result is the same as printf's "%g", a.k.a. "%.6g"; that is, six
-// significant digits are returned, trailing zeros are removed, and numbers
-// outside the range 0.0001-999999 are output using scientific notation
-// (1.23456e+06). This routine is heavily optimized.
-// Required buffer size is `kSixDigitsToBufferSize`.
-size_t SixDigitsToBuffer(double d, absl::Nonnull<char*> buffer);
-
-// All of these functions take an output buffer
-// as an argument and return a pointer to the last byte they wrote, which is the
-// terminating '\0'. At most `kFastToBufferSize` bytes are written.
-absl::Nonnull<char*> FastIntToBuffer(int32_t i, absl::Nonnull<char*> buffer);
-absl::Nonnull<char*> FastIntToBuffer(uint32_t i, absl::Nonnull<char*> buffer);
-absl::Nonnull<char*> FastIntToBuffer(int64_t i, absl::Nonnull<char*> buffer);
-absl::Nonnull<char*> FastIntToBuffer(uint64_t i, absl::Nonnull<char*> buffer);
-
-// For enums and integer types that are not an exact match for the types above,
-// use templates to call the appropriate one of the four overloads above.
-template <typename int_type>
-absl::Nonnull<char*> FastIntToBuffer(int_type i, absl::Nonnull<char*> buffer) {
-  static_assert(sizeof(i) <= 64 / 8,
-                "FastIntToBuffer works only with 64-bit-or-less integers.");
-  // TODO(jorg): This signed-ness check is used because it works correctly
-  // with enums, and it also serves to check that int_type is not a pointer.
-  // If one day something like std::is_signed<enum E> works, switch to it.
-  // These conditions are constexpr bools to suppress MSVC warning C4127.
-  constexpr bool kIsSigned = static_cast<int_type>(1) - 2 < 0;
-  constexpr bool kUse64Bit = sizeof(i) > 32 / 8;
-  if (kIsSigned) {
-    if (kUse64Bit) {
-      return FastIntToBuffer(static_cast<int64_t>(i), buffer);
-    } else {
-      return FastIntToBuffer(static_cast<int32_t>(i), buffer);
-    }
-  } else {
-    if (kUse64Bit) {
-      return FastIntToBuffer(static_cast<uint64_t>(i), buffer);
-    } else {
-      return FastIntToBuffer(static_cast<uint32_t>(i), buffer);
-    }
-  }
-}
-
-// These functions do NOT add any null-terminator.
-// They return a pointer to the beginning of the written string.
-// The digit counts provided must *exactly* match the number of base-10 digits
-// in the number, or the behavior is undefined.
-// (i.e. do NOT count the minus sign, or over- or under-count the digits.)
-absl::Nonnull<char*> FastIntToBufferBackward(int32_t i,
-                                             absl::Nonnull<char*> buffer_end,
-                                             uint32_t exact_digit_count);
-absl::Nonnull<char*> FastIntToBufferBackward(uint32_t i,
-                                             absl::Nonnull<char*> buffer_end,
-                                             uint32_t exact_digit_count);
-absl::Nonnull<char*> FastIntToBufferBackward(int64_t i,
-                                             absl::Nonnull<char*> buffer_end,
-                                             uint32_t exact_digit_count);
-absl::Nonnull<char*> FastIntToBufferBackward(uint64_t i,
-                                             absl::Nonnull<char*> buffer_end,
-                                             uint32_t exact_digit_count);
-
-// For enums and integer types that are not an exact match for the types above,
-// use templates to call the appropriate one of the four overloads above.
-template <typename int_type>
-absl::Nonnull<char*> FastIntToBufferBackward(int_type i,
-                                             absl::Nonnull<char*> buffer_end,
-                                             uint32_t exact_digit_count) {
-  static_assert(
-      sizeof(i) <= 64 / 8,
-      "FastIntToBufferBackward works only with 64-bit-or-less integers.");
-  // This signed-ness check is used because it works correctly
-  // with enums, and it also serves to check that int_type is not a pointer.
-  // If one day something like std::is_signed<enum E> works, switch to it.
-  // These conditions are constexpr bools to suppress MSVC warning C4127.
-  constexpr bool kIsSigned = static_cast<int_type>(1) - 2 < 0;
-  constexpr bool kUse64Bit = sizeof(i) > 32 / 8;
-  if (kIsSigned) {
-    if (kUse64Bit) {
-      return FastIntToBufferBackward(static_cast<int64_t>(i), buffer_end,
-                                     exact_digit_count);
-    } else {
-      return FastIntToBufferBackward(static_cast<int32_t>(i), buffer_end,
-                                     exact_digit_count);
-    }
-  } else {
-    if (kUse64Bit) {
-      return FastIntToBufferBackward(static_cast<uint64_t>(i), buffer_end,
-                                     exact_digit_count);
-    } else {
-      return FastIntToBufferBackward(static_cast<uint32_t>(i), buffer_end,
-                                     exact_digit_count);
-    }
-  }
-}
-
-// Implementation of SimpleAtoi, generalized to support arbitrary base (used
-// with base different from 10 elsewhere in Abseil implementation).
-template <typename int_type>
-ABSL_MUST_USE_RESULT bool safe_strtoi_base(absl::string_view s,
-                                           absl::Nonnull<int_type*> out,
-                                           int base) {
-  static_assert(sizeof(*out) == 4 || sizeof(*out) == 8,
-                "SimpleAtoi works only with 32-bit or 64-bit integers.");
-  static_assert(!std::is_floating_point<int_type>::value,
-                "Use SimpleAtof or SimpleAtod instead.");
-  bool parsed;
-  // TODO(jorg): This signed-ness check is used because it works correctly
-  // with enums, and it also serves to check that int_type is not a pointer.
-  // If one day something like std::is_signed<enum E> works, switch to it.
-  // These conditions are constexpr bools to suppress MSVC warning C4127.
-  constexpr bool kIsSigned = static_cast<int_type>(1) - 2 < 0;
-  constexpr bool kUse64Bit = sizeof(*out) == 64 / 8;
-  if (kIsSigned) {
-    if (kUse64Bit) {
-      int64_t val;
-      parsed = numbers_internal::safe_strto64_base(s, &val, base);
-      *out = static_cast<int_type>(val);
-    } else {
-      int32_t val;
-      parsed = numbers_internal::safe_strto32_base(s, &val, base);
-      *out = static_cast<int_type>(val);
-    }
-  } else {
-    if (kUse64Bit) {
-      uint64_t val;
-      parsed = numbers_internal::safe_strtou64_base(s, &val, base);
-      *out = static_cast<int_type>(val);
-    } else {
-      uint32_t val;
-      parsed = numbers_internal::safe_strtou32_base(s, &val, base);
-      *out = static_cast<int_type>(val);
-    }
-  }
-  return parsed;
-}
-
-// FastHexToBufferZeroPad16()
-//
-// Outputs `val` into `out` as if by `snprintf(out, 17, "%016x", val)` but
-// without the terminating null character. Thus `out` must be of length >= 16.
-// Returns the number of non-pad digits of the output (it can never be zero
-// since 0 has one digit).
-inline size_t FastHexToBufferZeroPad16(uint64_t val, absl::Nonnull<char*> out) {
-#ifdef ABSL_INTERNAL_HAVE_SSSE3
-  uint64_t be = absl::big_endian::FromHost64(val);
-  const auto kNibbleMask = _mm_set1_epi8(0xf);
-  const auto kHexDigits = _mm_setr_epi8('0', '1', '2', '3', '4', '5', '6', '7',
-                                        '8', '9', 'a', 'b', 'c', 'd', 'e', 'f');
-  auto v = _mm_loadl_epi64(reinterpret_cast<__m128i*>(&be));  // load lo dword
-  auto v4 = _mm_srli_epi64(v, 4);                            // shift 4 right
-  auto il = _mm_unpacklo_epi8(v4, v);                        // interleave bytes
-  auto m = _mm_and_si128(il, kNibbleMask);                   // mask out nibbles
-  auto hexchars = _mm_shuffle_epi8(kHexDigits, m);           // hex chars
-  _mm_storeu_si128(reinterpret_cast<__m128i*>(out), hexchars);
-#else
-  for (int i = 0; i < 8; ++i) {
-    auto byte = (val >> (56 - 8 * i)) & 0xFF;
-    auto* hex = &absl::numbers_internal::kHexTable[byte * 2];
-    std::memcpy(out + 2 * i, hex, 2);
-  }
-#endif
-  // | 0x1 so that even 0 has 1 digit.
-  return 16 - static_cast<size_t>(countl_zero(val | 0x1) / 4);
-}
-
-}  // namespace numbers_internal
-
-template <typename int_type>
-ABSL_MUST_USE_RESULT bool SimpleAtoi(absl::string_view str,
-                                     absl::Nonnull<int_type*> out) {
-  return numbers_internal::safe_strtoi_base(str, out, 10);
-}
-
-ABSL_MUST_USE_RESULT inline bool SimpleAtoi(absl::string_view str,
-                                            absl::Nonnull<absl::int128*> out) {
-  return numbers_internal::safe_strto128_base(str, out, 10);
-}
-
-ABSL_MUST_USE_RESULT inline bool SimpleAtoi(absl::string_view str,
-                                            absl::Nonnull<absl::uint128*> out) {
-  return numbers_internal::safe_strtou128_base(str, out, 10);
-}
-
-template <typename int_type>
-ABSL_MUST_USE_RESULT bool SimpleHexAtoi(absl::string_view str,
-                                        absl::Nonnull<int_type*> out) {
-  return numbers_internal::safe_strtoi_base(str, out, 16);
-}
-
-ABSL_MUST_USE_RESULT inline bool SimpleHexAtoi(
-    absl::string_view str, absl::Nonnull<absl::int128*> out) {
-  return numbers_internal::safe_strto128_base(str, out, 16);
-}
-
-ABSL_MUST_USE_RESULT inline bool SimpleHexAtoi(
-    absl::string_view str, absl::Nonnull<absl::uint128*> out) {
-  return numbers_internal::safe_strtou128_base(str, out, 16);
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_NUMBERS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/str_cat.h b/projects/llm_framework/include/abseil-cpp/absl/strings/str_cat.h
deleted file mode 100644
index 33355bfd..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/str_cat.h
+++ /dev/null
@@ -1,636 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: str_cat.h
-// -----------------------------------------------------------------------------
-//
-// This package contains functions for efficiently concatenating and appending
-// strings: `StrCat()` and `StrAppend()`. Most of the work within these routines
-// is actually handled through use of a special AlphaNum type, which was
-// designed to be used as a parameter type that efficiently manages conversion
-// to strings and avoids copies in the above operations.
-//
-// Any routine accepting either a string or a number may accept `AlphaNum`.
-// The basic idea is that by accepting a `const AlphaNum &` as an argument
-// to your function, your callers will automagically convert bools, integers,
-// and floating point values to strings for you.
-//
-// NOTE: Use of `AlphaNum` outside of the //absl/strings package is unsupported
-// except for the specific case of function parameters of type `AlphaNum` or
-// `const AlphaNum &`. In particular, instantiating `AlphaNum` directly as a
-// stack variable is not supported.
-//
-// Conversion from 8-bit values is not accepted because, if it were, then an
-// attempt to pass ':' instead of ":" might result in a 58 ending up in your
-// result.
-//
-// Bools convert to "0" or "1". Pointers to types other than `char *` are not
-// valid inputs. No output is generated for null `char *` pointers.
-//
-// Floating point numbers are formatted with six-digit precision, which is
-// the default for "std::cout <<" or printf "%g" (the same as "%.6g").
-//
-// You can convert to hexadecimal output rather than decimal output using the
-// `Hex` type contained here. To do so, pass `Hex(my_int)` as a parameter to
-// `StrCat()` or `StrAppend()`. You may specify a minimum hex field width using
-// a `PadSpec` enum.
-//
-// User-defined types can be formatted with the `AbslStringify()` customization
-// point. The API relies on detecting an overload in the user-defined type's
-// namespace of a free (non-member) `AbslStringify()` function as a definition
-// (typically declared as a friend and implemented in-line.
-// with the following signature:
-//
-// class MyClass { ... };
-//
-// template <typename Sink>
-// void AbslStringify(Sink& sink, const MyClass& value);
-//
-// An `AbslStringify()` overload for a type should only be declared in the same
-// file and namespace as said type.
-//
-// Note that `AbslStringify()` also supports use with `absl::StrFormat()` and
-// `absl::Substitute()`.
-//
-// Example:
-//
-// struct Point {
-//   // To add formatting support to `Point`, we simply need to add a free
-//   // (non-member) function `AbslStringify()`. This method specifies how
-//   // Point should be printed when absl::StrCat() is called on it. You can add
-//   // such a free function using a friend declaration within the body of the
-//   // class. The sink parameter is a templated type to avoid requiring
-//   // dependencies.
-//   template <typename Sink> friend void AbslStringify(Sink&
-//   sink, const Point& p) {
-//     absl::Format(&sink, "(%v, %v)", p.x, p.y);
-//   }
-//
-//   int x;
-//   int y;
-// };
-// -----------------------------------------------------------------------------
-
-#ifndef ABSL_STRINGS_STR_CAT_H_
-#define ABSL_STRINGS_STR_CAT_H_
-
-#include <algorithm>
-#include <array>
-#include <cassert>
-#include <cstddef>
-#include <cstdint>
-#include <cstring>
-#include <initializer_list>
-#include <string>
-#include <type_traits>
-#include <utility>
-#include <vector>
-
-#include "absl/base/attributes.h"
-#include "absl/base/nullability.h"
-#include "absl/base/port.h"
-#include "absl/meta/type_traits.h"
-#include "absl/strings/has_absl_stringify.h"
-#include "absl/strings/internal/resize_uninitialized.h"
-#include "absl/strings/internal/stringify_sink.h"
-#include "absl/strings/numbers.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-namespace strings_internal {
-// AlphaNumBuffer allows a way to pass a string to StrCat without having to do
-// memory allocation.  It is simply a pair of a fixed-size character array, and
-// a size.  Please don't use outside of absl, yet.
-template <size_t max_size>
-struct AlphaNumBuffer {
-  std::array<char, max_size> data;
-  size_t size;
-};
-
-}  // namespace strings_internal
-
-// Enum that specifies the number of significant digits to return in a `Hex` or
-// `Dec` conversion and fill character to use. A `kZeroPad2` value, for example,
-// would produce hexadecimal strings such as "0a","0f" and a 'kSpacePad5' value
-// would produce hexadecimal strings such as "    a","    f".
-enum PadSpec : uint8_t {
-  kNoPad = 1,
-  kZeroPad2,
-  kZeroPad3,
-  kZeroPad4,
-  kZeroPad5,
-  kZeroPad6,
-  kZeroPad7,
-  kZeroPad8,
-  kZeroPad9,
-  kZeroPad10,
-  kZeroPad11,
-  kZeroPad12,
-  kZeroPad13,
-  kZeroPad14,
-  kZeroPad15,
-  kZeroPad16,
-  kZeroPad17,
-  kZeroPad18,
-  kZeroPad19,
-  kZeroPad20,
-
-  kSpacePad2 = kZeroPad2 + 64,
-  kSpacePad3,
-  kSpacePad4,
-  kSpacePad5,
-  kSpacePad6,
-  kSpacePad7,
-  kSpacePad8,
-  kSpacePad9,
-  kSpacePad10,
-  kSpacePad11,
-  kSpacePad12,
-  kSpacePad13,
-  kSpacePad14,
-  kSpacePad15,
-  kSpacePad16,
-  kSpacePad17,
-  kSpacePad18,
-  kSpacePad19,
-  kSpacePad20,
-};
-
-// -----------------------------------------------------------------------------
-// Hex
-// -----------------------------------------------------------------------------
-//
-// `Hex` stores a set of hexadecimal string conversion parameters for use
-// within `AlphaNum` string conversions.
-struct Hex {
-  uint64_t value;
-  uint8_t width;
-  char fill;
-
-  template <typename Int>
-  explicit Hex(
-      Int v, PadSpec spec = absl::kNoPad,
-      typename std::enable_if<sizeof(Int) == 1 &&
-                              !std::is_pointer<Int>::value>::type* = nullptr)
-      : Hex(spec, static_cast<uint8_t>(v)) {}
-  template <typename Int>
-  explicit Hex(
-      Int v, PadSpec spec = absl::kNoPad,
-      typename std::enable_if<sizeof(Int) == 2 &&
-                              !std::is_pointer<Int>::value>::type* = nullptr)
-      : Hex(spec, static_cast<uint16_t>(v)) {}
-  template <typename Int>
-  explicit Hex(
-      Int v, PadSpec spec = absl::kNoPad,
-      typename std::enable_if<sizeof(Int) == 4 &&
-                              !std::is_pointer<Int>::value>::type* = nullptr)
-      : Hex(spec, static_cast<uint32_t>(v)) {}
-  template <typename Int>
-  explicit Hex(
-      Int v, PadSpec spec = absl::kNoPad,
-      typename std::enable_if<sizeof(Int) == 8 &&
-                              !std::is_pointer<Int>::value>::type* = nullptr)
-      : Hex(spec, static_cast<uint64_t>(v)) {}
-  template <typename Pointee>
-  explicit Hex(absl::Nullable<Pointee*> v, PadSpec spec = absl::kNoPad)
-      : Hex(spec, reinterpret_cast<uintptr_t>(v)) {}
-
-  template <typename S>
-  friend void AbslStringify(S& sink, Hex hex) {
-    static_assert(
-        numbers_internal::kFastToBufferSize >= 32,
-        "This function only works when output buffer >= 32 bytes long");
-    char buffer[numbers_internal::kFastToBufferSize];
-    char* const end = &buffer[numbers_internal::kFastToBufferSize];
-    auto real_width =
-        absl::numbers_internal::FastHexToBufferZeroPad16(hex.value, end - 16);
-    if (real_width >= hex.width) {
-      sink.Append(absl::string_view(end - real_width, real_width));
-    } else {
-      // Pad first 16 chars because FastHexToBufferZeroPad16 pads only to 16 and
-      // max pad width can be up to 20.
-      std::memset(end - 32, hex.fill, 16);
-      // Patch up everything else up to the real_width.
-      std::memset(end - real_width - 16, hex.fill, 16);
-      sink.Append(absl::string_view(end - hex.width, hex.width));
-    }
-  }
-
- private:
-  Hex(PadSpec spec, uint64_t v)
-      : value(v),
-        width(spec == absl::kNoPad
-                  ? 1
-                  : spec >= absl::kSpacePad2 ? spec - absl::kSpacePad2 + 2
-                                             : spec - absl::kZeroPad2 + 2),
-        fill(spec >= absl::kSpacePad2 ? ' ' : '0') {}
-};
-
-// -----------------------------------------------------------------------------
-// Dec
-// -----------------------------------------------------------------------------
-//
-// `Dec` stores a set of decimal string conversion parameters for use
-// within `AlphaNum` string conversions.  Dec is slower than the default
-// integer conversion, so use it only if you need padding.
-struct Dec {
-  uint64_t value;
-  uint8_t width;
-  char fill;
-  bool neg;
-
-  template <typename Int>
-  explicit Dec(Int v, PadSpec spec = absl::kNoPad,
-               typename std::enable_if<(sizeof(Int) <= 8)>::type* = nullptr)
-      : value(v >= 0 ? static_cast<uint64_t>(v)
-                     : uint64_t{0} - static_cast<uint64_t>(v)),
-        width(spec == absl::kNoPad       ? 1
-              : spec >= absl::kSpacePad2 ? spec - absl::kSpacePad2 + 2
-                                         : spec - absl::kZeroPad2 + 2),
-        fill(spec >= absl::kSpacePad2 ? ' ' : '0'),
-        neg(v < 0) {}
-
-  template <typename S>
-  friend void AbslStringify(S& sink, Dec dec) {
-    assert(dec.width <= numbers_internal::kFastToBufferSize);
-    char buffer[numbers_internal::kFastToBufferSize];
-    char* const end = &buffer[numbers_internal::kFastToBufferSize];
-    char* const minfill = end - dec.width;
-    char* writer = end;
-    uint64_t val = dec.value;
-    while (val > 9) {
-      *--writer = '0' + (val % 10);
-      val /= 10;
-    }
-    *--writer = '0' + static_cast<char>(val);
-    if (dec.neg) *--writer = '-';
-
-    ptrdiff_t fillers = writer - minfill;
-    if (fillers > 0) {
-      // Tricky: if the fill character is ' ', then it's <fill><+/-><digits>
-      // But...: if the fill character is '0', then it's <+/-><fill><digits>
-      bool add_sign_again = false;
-      if (dec.neg && dec.fill == '0') {  // If filling with '0',
-        ++writer;                    // ignore the sign we just added
-        add_sign_again = true;       // and re-add the sign later.
-      }
-      writer -= fillers;
-      std::fill_n(writer, fillers, dec.fill);
-      if (add_sign_again) *--writer = '-';
-    }
-
-    sink.Append(absl::string_view(writer, static_cast<size_t>(end - writer)));
-  }
-};
-
-// -----------------------------------------------------------------------------
-// AlphaNum
-// -----------------------------------------------------------------------------
-//
-// The `AlphaNum` class acts as the main parameter type for `StrCat()` and
-// `StrAppend()`, providing efficient conversion of numeric, boolean, decimal,
-// and hexadecimal values (through the `Dec` and `Hex` types) into strings.
-// `AlphaNum` should only be used as a function parameter. Do not instantiate
-//  `AlphaNum` directly as a stack variable.
-
-class AlphaNum {
- public:
-  // No bool ctor -- bools convert to an integral type.
-  // A bool ctor would also convert incoming pointers (bletch).
-
-  // Prevent brace initialization
-  template <typename T>
-  AlphaNum(std::initializer_list<T>) = delete;  // NOLINT(runtime/explicit)
-
-  AlphaNum(int x)  // NOLINT(runtime/explicit)
-      : piece_(digits_, static_cast<size_t>(
-                            numbers_internal::FastIntToBuffer(x, digits_) -
-                            &digits_[0])) {}
-  AlphaNum(unsigned int x)  // NOLINT(runtime/explicit)
-      : piece_(digits_, static_cast<size_t>(
-                            numbers_internal::FastIntToBuffer(x, digits_) -
-                            &digits_[0])) {}
-  AlphaNum(long x)  // NOLINT(*)
-      : piece_(digits_, static_cast<size_t>(
-                            numbers_internal::FastIntToBuffer(x, digits_) -
-                            &digits_[0])) {}
-  AlphaNum(unsigned long x)  // NOLINT(*)
-      : piece_(digits_, static_cast<size_t>(
-                            numbers_internal::FastIntToBuffer(x, digits_) -
-                            &digits_[0])) {}
-  AlphaNum(long long x)  // NOLINT(*)
-      : piece_(digits_, static_cast<size_t>(
-                            numbers_internal::FastIntToBuffer(x, digits_) -
-                            &digits_[0])) {}
-  AlphaNum(unsigned long long x)  // NOLINT(*)
-      : piece_(digits_, static_cast<size_t>(
-                            numbers_internal::FastIntToBuffer(x, digits_) -
-                            &digits_[0])) {}
-
-  AlphaNum(float f)  // NOLINT(runtime/explicit)
-      : piece_(digits_, numbers_internal::SixDigitsToBuffer(f, digits_)) {}
-  AlphaNum(double f)  // NOLINT(runtime/explicit)
-      : piece_(digits_, numbers_internal::SixDigitsToBuffer(f, digits_)) {}
-
-  template <size_t size>
-  AlphaNum(  // NOLINT(runtime/explicit)
-      const strings_internal::AlphaNumBuffer<size>& buf
-          ABSL_ATTRIBUTE_LIFETIME_BOUND)
-      : piece_(&buf.data[0], buf.size) {}
-
-  AlphaNum(absl::Nullable<const char*> c_str  // NOLINT(runtime/explicit)
-               ABSL_ATTRIBUTE_LIFETIME_BOUND)
-      : piece_(NullSafeStringView(c_str)) {}
-  AlphaNum(absl::string_view pc  // NOLINT(runtime/explicit)
-               ABSL_ATTRIBUTE_LIFETIME_BOUND)
-      : piece_(pc) {}
-
-  template <typename T, typename = typename std::enable_if<
-                            HasAbslStringify<T>::value>::type>
-  AlphaNum(  // NOLINT(runtime/explicit)
-      const T& v ABSL_ATTRIBUTE_LIFETIME_BOUND,
-      strings_internal::StringifySink&& sink ABSL_ATTRIBUTE_LIFETIME_BOUND = {})
-      : piece_(strings_internal::ExtractStringification(sink, v)) {}
-
-  template <typename Allocator>
-  AlphaNum(  // NOLINT(runtime/explicit)
-      const std::basic_string<char, std::char_traits<char>, Allocator>& str
-          ABSL_ATTRIBUTE_LIFETIME_BOUND)
-      : piece_(str) {}
-
-  // Use string literals ":" instead of character literals ':'.
-  AlphaNum(char c) = delete;  // NOLINT(runtime/explicit)
-
-  AlphaNum(const AlphaNum&) = delete;
-  AlphaNum& operator=(const AlphaNum&) = delete;
-
-  absl::string_view::size_type size() const { return piece_.size(); }
-  absl::Nullable<const char*> data() const { return piece_.data(); }
-  absl::string_view Piece() const { return piece_; }
-
-  // Match unscoped enums.  Use integral promotion so that a `char`-backed
-  // enum becomes a wider integral type AlphaNum will accept.
-  template <typename T,
-            typename = typename std::enable_if<
-                std::is_enum<T>{} && std::is_convertible<T, int>{} &&
-                !HasAbslStringify<T>::value>::type>
-  AlphaNum(T e)  // NOLINT(runtime/explicit)
-      : AlphaNum(+e) {}
-
-  // This overload matches scoped enums.  We must explicitly cast to the
-  // underlying type, but use integral promotion for the same reason as above.
-  template <typename T,
-            typename std::enable_if<std::is_enum<T>{} &&
-                                        !std::is_convertible<T, int>{} &&
-                                        !HasAbslStringify<T>::value,
-                                    char*>::type = nullptr>
-  AlphaNum(T e)  // NOLINT(runtime/explicit)
-      : AlphaNum(+static_cast<typename std::underlying_type<T>::type>(e)) {}
-
-  // vector<bool>::reference and const_reference require special help to
-  // convert to `AlphaNum` because it requires two user defined conversions.
-  template <
-      typename T,
-      typename std::enable_if<
-          std::is_class<T>::value &&
-          (std::is_same<T, std::vector<bool>::reference>::value ||
-           std::is_same<T, std::vector<bool>::const_reference>::value)>::type* =
-          nullptr>
-  AlphaNum(T e) : AlphaNum(static_cast<bool>(e)) {}  // NOLINT(runtime/explicit)
-
- private:
-  absl::string_view piece_;
-  char digits_[numbers_internal::kFastToBufferSize];
-};
-
-// -----------------------------------------------------------------------------
-// StrCat()
-// -----------------------------------------------------------------------------
-//
-// Merges given strings or numbers, using no delimiter(s), returning the merged
-// result as a string.
-//
-// `StrCat()` is designed to be the fastest possible way to construct a string
-// out of a mix of raw C strings, string_views, strings, bool values,
-// and numeric values.
-//
-// Don't use `StrCat()` for user-visible strings. The localization process
-// works poorly on strings built up out of fragments.
-//
-// For clarity and performance, don't use `StrCat()` when appending to a
-// string. Use `StrAppend()` instead. In particular, avoid using any of these
-// (anti-)patterns:
-//
-//   str.append(StrCat(...))
-//   str += StrCat(...)
-//   str = StrCat(str, ...)
-//
-// The last case is the worst, with a potential to change a loop
-// from a linear time operation with O(1) dynamic allocations into a
-// quadratic time operation with O(n) dynamic allocations.
-//
-// See `StrAppend()` below for more information.
-
-namespace strings_internal {
-
-// Do not call directly - this is not part of the public API.
-std::string CatPieces(std::initializer_list<absl::string_view> pieces);
-void AppendPieces(absl::Nonnull<std::string*> dest,
-                  std::initializer_list<absl::string_view> pieces);
-
-void STLStringAppendUninitializedAmortized(std::string* dest, size_t to_append);
-
-// `SingleArgStrCat` overloads take built-in `int`, `long` and `long long` types
-// (signed / unsigned) to avoid ambiguity on the call side. If we used int32_t
-// and int64_t, then at least one of the three (`int` / `long` / `long long`)
-// would have been ambiguous when passed to `SingleArgStrCat`.
-std::string SingleArgStrCat(int x);
-std::string SingleArgStrCat(unsigned int x);
-std::string SingleArgStrCat(long x);                // NOLINT
-std::string SingleArgStrCat(unsigned long x);       // NOLINT
-std::string SingleArgStrCat(long long x);           // NOLINT
-std::string SingleArgStrCat(unsigned long long x);  // NOLINT
-std::string SingleArgStrCat(float x);
-std::string SingleArgStrCat(double x);
-
-// `SingleArgStrAppend` overloads are defined here for the same reasons as with
-// `SingleArgStrCat` above.
-void SingleArgStrAppend(std::string& str, int x);
-void SingleArgStrAppend(std::string& str, unsigned int x);
-void SingleArgStrAppend(std::string& str, long x);                // NOLINT
-void SingleArgStrAppend(std::string& str, unsigned long x);       // NOLINT
-void SingleArgStrAppend(std::string& str, long long x);           // NOLINT
-void SingleArgStrAppend(std::string& str, unsigned long long x);  // NOLINT
-
-template <typename T,
-          typename = std::enable_if_t<std::is_arithmetic<T>::value &&
-                                      !std::is_same<T, char>::value &&
-                                      !std::is_same<T, bool>::value>>
-using EnableIfFastCase = T;
-
-}  // namespace strings_internal
-
-ABSL_MUST_USE_RESULT inline std::string StrCat() { return std::string(); }
-
-template <typename T>
-ABSL_MUST_USE_RESULT inline std::string StrCat(
-    strings_internal::EnableIfFastCase<T> a) {
-  return strings_internal::SingleArgStrCat(a);
-}
-ABSL_MUST_USE_RESULT inline std::string StrCat(const AlphaNum& a) {
-  return std::string(a.data(), a.size());
-}
-
-ABSL_MUST_USE_RESULT std::string StrCat(const AlphaNum& a, const AlphaNum& b);
-ABSL_MUST_USE_RESULT std::string StrCat(const AlphaNum& a, const AlphaNum& b,
-                                        const AlphaNum& c);
-ABSL_MUST_USE_RESULT std::string StrCat(const AlphaNum& a, const AlphaNum& b,
-                                        const AlphaNum& c, const AlphaNum& d);
-
-// Support 5 or more arguments
-template <typename... AV>
-ABSL_MUST_USE_RESULT inline std::string StrCat(
-    const AlphaNum& a, const AlphaNum& b, const AlphaNum& c, const AlphaNum& d,
-    const AlphaNum& e, const AV&... args) {
-  return strings_internal::CatPieces(
-      {a.Piece(), b.Piece(), c.Piece(), d.Piece(), e.Piece(),
-       static_cast<const AlphaNum&>(args).Piece()...});
-}
-
-// -----------------------------------------------------------------------------
-// StrAppend()
-// -----------------------------------------------------------------------------
-//
-// Appends a string or set of strings to an existing string, in a similar
-// fashion to `StrCat()`.
-//
-// WARNING: `StrAppend(&str, a, b, c, ...)` requires that none of the
-// a, b, c, parameters be a reference into str. For speed, `StrAppend()` does
-// not try to check each of its input arguments to be sure that they are not
-// a subset of the string being appended to. That is, while this will work:
-//
-//   std::string s = "foo";
-//   s += s;
-//
-// This output is undefined:
-//
-//   std::string s = "foo";
-//   StrAppend(&s, s);
-//
-// This output is undefined as well, since `absl::string_view` does not own its
-// data:
-//
-//   std::string s = "foobar";
-//   absl::string_view p = s;
-//   StrAppend(&s, p);
-
-inline void StrAppend(absl::Nonnull<std::string*>) {}
-void StrAppend(absl::Nonnull<std::string*> dest, const AlphaNum& a);
-void StrAppend(absl::Nonnull<std::string*> dest, const AlphaNum& a,
-               const AlphaNum& b);
-void StrAppend(absl::Nonnull<std::string*> dest, const AlphaNum& a,
-               const AlphaNum& b, const AlphaNum& c);
-void StrAppend(absl::Nonnull<std::string*> dest, const AlphaNum& a,
-               const AlphaNum& b, const AlphaNum& c, const AlphaNum& d);
-
-// Support 5 or more arguments
-template <typename... AV>
-inline void StrAppend(absl::Nonnull<std::string*> dest, const AlphaNum& a,
-                      const AlphaNum& b, const AlphaNum& c, const AlphaNum& d,
-                      const AlphaNum& e, const AV&... args) {
-  strings_internal::AppendPieces(
-      dest, {a.Piece(), b.Piece(), c.Piece(), d.Piece(), e.Piece(),
-             static_cast<const AlphaNum&>(args).Piece()...});
-}
-
-template <class String, class T>
-std::enable_if_t<
-    std::is_integral<absl::strings_internal::EnableIfFastCase<T>>::value, void>
-StrAppend(absl::Nonnull<String*> result, T i) {
-  return absl::strings_internal::SingleArgStrAppend(*result, i);
-}
-
-// This overload is only selected if all the parameters are numbers that can be
-// handled quickly.
-// Later we can look into how we can extend this to more general argument
-// mixtures without bloating codegen too much, or copying unnecessarily.
-template <typename String, typename... T>
-std::enable_if_t<
-    (sizeof...(T) > 1),
-    std::common_type_t<std::conditional_t<
-        true, void, absl::strings_internal::EnableIfFastCase<T>>...>>
-StrAppend(absl::Nonnull<String*> str, T... args) {
-  // Do not add unnecessary variables, logic, or even "free" lambdas here.
-  // They can add overhead for the compiler and/or at run time.
-  // Furthermore, assume this function will be inlined.
-  // This function is carefully tailored to be able to be largely optimized away
-  // so that it becomes near-equivalent to the caller handling each argument
-  // individually while minimizing register pressure, so that the compiler
-  // can inline it with minimal overhead.
-
-  // First, calculate the total length, so we can perform just a single resize.
-  // Save all the lengths for later.
-  size_t total_length = 0;
-  const ptrdiff_t lengths[] = {
-      absl::numbers_internal::GetNumDigitsOrNegativeIfNegative(args)...};
-  for (const ptrdiff_t possibly_negative_length : lengths) {
-    // Lengths are negative for negative numbers. Keep them for later use, but
-    // take their absolute values for calculating total lengths;
-    total_length += possibly_negative_length < 0
-                        ? static_cast<size_t>(-possibly_negative_length)
-                        : static_cast<size_t>(possibly_negative_length);
-  }
-
-  // Now reserve space for all the arguments.
-  const size_t old_size = str->size();
-  absl::strings_internal::STLStringAppendUninitializedAmortized(str,
-                                                                total_length);
-
-  // Finally, output each argument one-by-one, from left to right.
-  size_t i = 0;  // The current argument we're processing
-  ptrdiff_t n;   // The length of the current argument
-  typename String::pointer pos = &(*str)[old_size];
-  using SomeTrivialEmptyType = std::false_type;
-  const SomeTrivialEmptyType dummy;
-  // Ugly code due to the lack of C++17 fold expressions
-  const SomeTrivialEmptyType dummies[] = {
-      (/* Comma expressions are poor man's C++17 fold expression for C++14 */
-       (void)(n = lengths[i]),
-       (void)(n < 0 ? (void)(*pos++ = '-'), (n = ~n) : 0),
-       (void)absl::numbers_internal::FastIntToBufferBackward(
-           absl::numbers_internal::UnsignedAbsoluteValue(std::move(args)),
-           pos += n, static_cast<uint32_t>(n)),
-       (void)++i, dummy)...};
-  (void)dummies;  // Remove & migrate to fold expressions in C++17
-}
-
-// Helper function for the future StrCat default floating-point format, %.6g
-// This is fast.
-inline strings_internal::AlphaNumBuffer<
-    numbers_internal::kSixDigitsToBufferSize>
-SixDigits(double d) {
-  strings_internal::AlphaNumBuffer<numbers_internal::kSixDigitsToBufferSize>
-      result;
-  result.size = numbers_internal::SixDigitsToBuffer(d, &result.data[0]);
-  return result;
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_STR_CAT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/str_format.h b/projects/llm_framework/include/abseil-cpp/absl/strings/str_format.h
deleted file mode 100644
index 66b6af58..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/str_format.h
+++ /dev/null
@@ -1,887 +0,0 @@
-//
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: str_format.h
-// -----------------------------------------------------------------------------
-//
-// The `str_format` library is a typesafe replacement for the family of
-// `printf()` string formatting routines within the `<cstdio>` standard library
-// header. Like the `printf` family, `str_format` uses a "format string" to
-// perform argument substitutions based on types. See the `FormatSpec` section
-// below for format string documentation.
-//
-// Example:
-//
-//   std::string s = absl::StrFormat(
-//                      "%s %s You have $%d!", "Hello", name, dollars);
-//
-// The library consists of the following basic utilities:
-//
-//   * `absl::StrFormat()`, a type-safe replacement for `std::sprintf()`, to
-//     write a format string to a `string` value.
-//   * `absl::StrAppendFormat()` to append a format string to a `string`
-//   * `absl::StreamFormat()` to more efficiently write a format string to a
-//     stream, such as`std::cout`.
-//   * `absl::PrintF()`, `absl::FPrintF()` and `absl::SNPrintF()` as
-//     drop-in replacements for `std::printf()`, `std::fprintf()` and
-//     `std::snprintf()`.
-//
-//     Note: An `absl::SPrintF()` drop-in replacement is not supported as it
-//     is generally unsafe due to buffer overflows. Use `absl::StrFormat` which
-//     returns the string as output instead of expecting a pre-allocated buffer.
-//
-// Additionally, you can provide a format string (and its associated arguments)
-// using one of the following abstractions:
-//
-//   * A `FormatSpec` class template fully encapsulates a format string and its
-//     type arguments and is usually provided to `str_format` functions as a
-//     variadic argument of type `FormatSpec<Arg...>`. The `FormatSpec<Args...>`
-//     template is evaluated at compile-time, providing type safety.
-//   * A `ParsedFormat` instance, which encapsulates a specific, pre-compiled
-//     format string for a specific set of type(s), and which can be passed
-//     between API boundaries. (The `FormatSpec` type should not be used
-//     directly except as an argument type for wrapper functions.)
-//
-// The `str_format` library provides the ability to output its format strings to
-// arbitrary sink types:
-//
-//   * A generic `Format()` function to write outputs to arbitrary sink types,
-//     which must implement a `FormatRawSink` interface.
-//
-//   * A `FormatUntyped()` function that is similar to `Format()` except it is
-//     loosely typed. `FormatUntyped()` is not a template and does not perform
-//     any compile-time checking of the format string; instead, it returns a
-//     boolean from a runtime check.
-//
-// In addition, the `str_format` library provides extension points for
-// augmenting formatting to new types.  See "StrFormat Extensions" below.
-
-#ifndef ABSL_STRINGS_STR_FORMAT_H_
-#define ABSL_STRINGS_STR_FORMAT_H_
-
-#include <cstdint>
-#include <cstdio>
-#include <string>
-#include <type_traits>
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/base/nullability.h"
-#include "absl/strings/internal/str_format/arg.h"  // IWYU pragma: export
-#include "absl/strings/internal/str_format/bind.h"  // IWYU pragma: export
-#include "absl/strings/internal/str_format/checker.h"  // IWYU pragma: export
-#include "absl/strings/internal/str_format/extension.h"  // IWYU pragma: export
-#include "absl/strings/internal/str_format/parser.h"  // IWYU pragma: export
-#include "absl/strings/string_view.h"
-#include "absl/types/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// UntypedFormatSpec
-//
-// A type-erased class that can be used directly within untyped API entry
-// points. An `UntypedFormatSpec` is specifically used as an argument to
-// `FormatUntyped()`.
-//
-// Example:
-//
-//   absl::UntypedFormatSpec format("%d");
-//   std::string out;
-//   CHECK(absl::FormatUntyped(&out, format, {absl::FormatArg(1)}));
-class UntypedFormatSpec {
- public:
-  UntypedFormatSpec() = delete;
-  UntypedFormatSpec(const UntypedFormatSpec&) = delete;
-  UntypedFormatSpec& operator=(const UntypedFormatSpec&) = delete;
-
-  explicit UntypedFormatSpec(string_view s) : spec_(s) {}
-
- protected:
-  explicit UntypedFormatSpec(
-      absl::Nonnull<const str_format_internal::ParsedFormatBase*> pc)
-      : spec_(pc) {}
-
- private:
-  friend str_format_internal::UntypedFormatSpecImpl;
-  str_format_internal::UntypedFormatSpecImpl spec_;
-};
-
-// FormatStreamed()
-//
-// Takes a streamable argument and returns an object that can print it
-// with '%s'. Allows printing of types that have an `operator<<` but no
-// intrinsic type support within `StrFormat()` itself.
-//
-// Example:
-//
-//   absl::StrFormat("%s", absl::FormatStreamed(obj));
-template <typename T>
-str_format_internal::StreamedWrapper<T> FormatStreamed(const T& v) {
-  return str_format_internal::StreamedWrapper<T>(v);
-}
-
-// FormatCountCapture
-//
-// This class provides a way to safely wrap `StrFormat()` captures of `%n`
-// conversions, which denote the number of characters written by a formatting
-// operation to this point, into an integer value.
-//
-// This wrapper is designed to allow safe usage of `%n` within `StrFormat(); in
-// the `printf()` family of functions, `%n` is not safe to use, as the `int *`
-// buffer can be used to capture arbitrary data.
-//
-// Example:
-//
-//   int n = 0;
-//   std::string s = absl::StrFormat("%s%d%n", "hello", 123,
-//                       absl::FormatCountCapture(&n));
-//   EXPECT_EQ(8, n);
-class FormatCountCapture {
- public:
-  explicit FormatCountCapture(absl::Nonnull<int*> p) : p_(p) {}
-
- private:
-  // FormatCountCaptureHelper is used to define FormatConvertImpl() for this
-  // class.
-  friend struct str_format_internal::FormatCountCaptureHelper;
-  // Unused() is here because of the false positive from -Wunused-private-field
-  // p_ is used in the templated function of the friend FormatCountCaptureHelper
-  // class.
-  absl::Nonnull<int*> Unused() { return p_; }
-  absl::Nonnull<int*> p_;
-};
-
-// FormatSpec
-//
-// The `FormatSpec` type defines the makeup of a format string within the
-// `str_format` library. It is a variadic class template that is evaluated at
-// compile-time, according to the format string and arguments that are passed to
-// it.
-//
-// You should not need to manipulate this type directly. You should only name it
-// if you are writing wrapper functions which accept format arguments that will
-// be provided unmodified to functions in this library. Such a wrapper function
-// might be a class method that provides format arguments and/or internally uses
-// the result of formatting.
-//
-// For a `FormatSpec` to be valid at compile-time, it must be provided as
-// either:
-//
-// * A `constexpr` literal or `absl::string_view`, which is how it most often
-//   used.
-// * A `ParsedFormat` instantiation, which ensures the format string is
-//   valid before use. (See below.)
-//
-// Example:
-//
-//   // Provided as a string literal.
-//   absl::StrFormat("Welcome to %s, Number %d!", "The Village", 6);
-//
-//   // Provided as a constexpr absl::string_view.
-//   constexpr absl::string_view formatString = "Welcome to %s, Number %d!";
-//   absl::StrFormat(formatString, "The Village", 6);
-//
-//   // Provided as a pre-compiled ParsedFormat object.
-//   // Note that this example is useful only for illustration purposes.
-//   absl::ParsedFormat<'s', 'd'> formatString("Welcome to %s, Number %d!");
-//   absl::StrFormat(formatString, "TheVillage", 6);
-//
-// A format string generally follows the POSIX syntax as used within the POSIX
-// `printf` specification. (Exceptions are noted below.)
-//
-// (See http://pubs.opengroup.org/onlinepubs/9699919799/functions/fprintf.html)
-//
-// In specific, the `FormatSpec` supports the following type specifiers:
-//   * `c` for characters
-//   * `s` for strings
-//   * `d` or `i` for integers
-//   * `o` for unsigned integer conversions into octal
-//   * `x` or `X` for unsigned integer conversions into hex
-//   * `u` for unsigned integers
-//   * `f` or `F` for floating point values into decimal notation
-//   * `e` or `E` for floating point values into exponential notation
-//   * `a` or `A` for floating point values into hex exponential notation
-//   * `g` or `G` for floating point values into decimal or exponential
-//     notation based on their precision
-//   * `p` for pointer address values
-//   * `n` for the special case of writing out the number of characters
-//     written to this point. The resulting value must be captured within an
-//     `absl::FormatCountCapture` type.
-//   * `v` for values using the default format for a deduced type. These deduced
-//     types include many of the primitive types denoted here as well as
-//     user-defined types containing the proper extensions. (See below for more
-//     information.)
-//
-// Implementation-defined behavior:
-//   * A null pointer provided to "%s" or "%p" is output as "(nil)".
-//   * A non-null pointer provided to "%p" is output in hex as if by %#x or
-//     %#lx.
-//
-// NOTE: `o`, `x\X` and `u` will convert signed values to their unsigned
-// counterpart before formatting.
-//
-// Examples:
-//     "%c", 'a'                -> "a"
-//     "%c", 32                 -> " "
-//     "%s", "C"                -> "C"
-//     "%s", std::string("C++") -> "C++"
-//     "%d", -10                -> "-10"
-//     "%o", 10                 -> "12"
-//     "%x", 16                 -> "10"
-//     "%f", 123456789          -> "123456789.000000"
-//     "%e", .01                -> "1.00000e-2"
-//     "%a", -3.0               -> "-0x1.8p+1"
-//     "%g", .01                -> "1e-2"
-//     "%p", (void*)&value      -> "0x7ffdeb6ad2a4"
-//
-//     int n = 0;
-//     std::string s = absl::StrFormat(
-//         "%s%d%n", "hello", 123, absl::FormatCountCapture(&n));
-//     EXPECT_EQ(8, n);
-//
-// NOTE: the `v` specifier (for "value") is a type specifier not present in the
-// POSIX specification. %v will format values according to their deduced type.
-// `v` uses `d` for signed integer values, `u` for unsigned integer values, `g`
-// for floating point values, and formats boolean values as "true"/"false"
-// (instead of 1 or 0 for booleans formatted using d). `const char*` is not
-// supported; please use `std::string` and `string_view`. `char` is also not
-// supported due to ambiguity of the type. This specifier does not support
-// modifiers.
-//
-// The `FormatSpec` intrinsically supports all of these fundamental C++ types:
-//
-// *   Characters: `char`, `signed char`, `unsigned char`, `wchar_t`
-// *   Integers: `int`, `short`, `unsigned short`, `unsigned`, `long`,
-//         `unsigned long`, `long long`, `unsigned long long`
-// *   Enums: printed as their underlying integral value
-// *   Floating-point: `float`, `double`, `long double`
-//
-// However, in the `str_format` library, a format conversion specifies a broader
-// C++ conceptual category instead of an exact type. For example, `%s` binds to
-// any string-like argument, so `std::string`, `std::wstring`,
-// `absl::string_view`, `const char*`, and `const wchar_t*` are all accepted.
-// Likewise, `%d` accepts any integer-like argument, etc.
-
-template <typename... Args>
-using FormatSpec = str_format_internal::FormatSpecTemplate<
-    str_format_internal::ArgumentToConv<Args>()...>;
-
-// ParsedFormat
-//
-// A `ParsedFormat` is a class template representing a preparsed `FormatSpec`,
-// with template arguments specifying the conversion characters used within the
-// format string. Such characters must be valid format type specifiers, and
-// these type specifiers are checked at compile-time.
-//
-// Instances of `ParsedFormat` can be created, copied, and reused to speed up
-// formatting loops. A `ParsedFormat` may either be constructed statically, or
-// dynamically through its `New()` factory function, which only constructs a
-// runtime object if the format is valid at that time.
-//
-// Example:
-//
-//   // Verified at compile time.
-//   absl::ParsedFormat<'s', 'd'> format_string("Welcome to %s, Number %d!");
-//   absl::StrFormat(format_string, "TheVillage", 6);
-//
-//   // Verified at runtime.
-//   auto format_runtime = absl::ParsedFormat<'d'>::New(format_string);
-//   if (format_runtime) {
-//     value = absl::StrFormat(*format_runtime, i);
-//   } else {
-//     ... error case ...
-//   }
-
-#if defined(__cpp_nontype_template_parameter_auto)
-// If C++17 is available, an 'extended' format is also allowed that can specify
-// multiple conversion characters per format argument, using a combination of
-// `absl::FormatConversionCharSet` enum values (logically a set union)
-//  via the `|` operator. (Single character-based arguments are still accepted,
-// but cannot be combined). Some common conversions also have predefined enum
-// values, such as `absl::FormatConversionCharSet::kIntegral`.
-//
-// Example:
-//   // Extended format supports multiple conversion characters per argument,
-//   // specified via a combination of `FormatConversionCharSet` enums.
-//   using MyFormat = absl::ParsedFormat<absl::FormatConversionCharSet::d |
-//                                       absl::FormatConversionCharSet::x>;
-//   MyFormat GetFormat(bool use_hex) {
-//     if (use_hex) return MyFormat("foo %x bar");
-//     return MyFormat("foo %d bar");
-//   }
-//   // `format` can be used with any value that supports 'd' and 'x',
-//   // like `int`.
-//   auto format = GetFormat(use_hex);
-//   value = StringF(format, i);
-template <auto... Conv>
-using ParsedFormat = absl::str_format_internal::ExtendedParsedFormat<
-    absl::str_format_internal::ToFormatConversionCharSet(Conv)...>;
-#else
-template <char... Conv>
-using ParsedFormat = str_format_internal::ExtendedParsedFormat<
-    absl::str_format_internal::ToFormatConversionCharSet(Conv)...>;
-#endif  // defined(__cpp_nontype_template_parameter_auto)
-
-// StrFormat()
-//
-// Returns a `string` given a `printf()`-style format string and zero or more
-// additional arguments. Use it as you would `sprintf()`. `StrFormat()` is the
-// primary formatting function within the `str_format` library, and should be
-// used in most cases where you need type-safe conversion of types into
-// formatted strings.
-//
-// The format string generally consists of ordinary character data along with
-// one or more format conversion specifiers (denoted by the `%` character).
-// Ordinary character data is returned unchanged into the result string, while
-// each conversion specification performs a type substitution from
-// `StrFormat()`'s other arguments. See the comments for `FormatSpec` for full
-// information on the makeup of this format string.
-//
-// Example:
-//
-//   std::string s = absl::StrFormat(
-//       "Welcome to %s, Number %d!", "The Village", 6);
-//   EXPECT_EQ("Welcome to The Village, Number 6!", s);
-//
-// Returns an empty string in case of error.
-template <typename... Args>
-ABSL_MUST_USE_RESULT std::string StrFormat(const FormatSpec<Args...>& format,
-                                           const Args&... args) {
-  return str_format_internal::FormatPack(
-      str_format_internal::UntypedFormatSpecImpl::Extract(format),
-      {str_format_internal::FormatArgImpl(args)...});
-}
-
-// StrAppendFormat()
-//
-// Appends to a `dst` string given a format string, and zero or more additional
-// arguments, returning `*dst` as a convenience for chaining purposes. Appends
-// nothing in case of error (but possibly alters its capacity).
-//
-// Example:
-//
-//   std::string orig("For example PI is approximately ");
-//   std::cout << StrAppendFormat(&orig, "%12.6f", 3.14);
-template <typename... Args>
-std::string& StrAppendFormat(absl::Nonnull<std::string*> dst,
-                             const FormatSpec<Args...>& format,
-                             const Args&... args) {
-  return str_format_internal::AppendPack(
-      dst, str_format_internal::UntypedFormatSpecImpl::Extract(format),
-      {str_format_internal::FormatArgImpl(args)...});
-}
-
-// StreamFormat()
-//
-// Writes to an output stream given a format string and zero or more arguments,
-// generally in a manner that is more efficient than streaming the result of
-// `absl::StrFormat()`. The returned object must be streamed before the full
-// expression ends.
-//
-// Example:
-//
-//   std::cout << StreamFormat("%12.6f", 3.14);
-template <typename... Args>
-ABSL_MUST_USE_RESULT str_format_internal::Streamable StreamFormat(
-    const FormatSpec<Args...>& format, const Args&... args) {
-  return str_format_internal::Streamable(
-      str_format_internal::UntypedFormatSpecImpl::Extract(format),
-      {str_format_internal::FormatArgImpl(args)...});
-}
-
-// PrintF()
-//
-// Writes to stdout given a format string and zero or more arguments. This
-// function is functionally equivalent to `std::printf()` (and type-safe);
-// prefer `absl::PrintF()` over `std::printf()`.
-//
-// Example:
-//
-//   std::string_view s = "Ulaanbaatar";
-//   absl::PrintF("The capital of Mongolia is %s", s);
-//
-//   Outputs: "The capital of Mongolia is Ulaanbaatar"
-//
-template <typename... Args>
-int PrintF(const FormatSpec<Args...>& format, const Args&... args) {
-  return str_format_internal::FprintF(
-      stdout, str_format_internal::UntypedFormatSpecImpl::Extract(format),
-      {str_format_internal::FormatArgImpl(args)...});
-}
-
-// FPrintF()
-//
-// Writes to a file given a format string and zero or more arguments. This
-// function is functionally equivalent to `std::fprintf()` (and type-safe);
-// prefer `absl::FPrintF()` over `std::fprintf()`.
-//
-// Example:
-//
-//   std::string_view s = "Ulaanbaatar";
-//   absl::FPrintF(stdout, "The capital of Mongolia is %s", s);
-//
-//   Outputs: "The capital of Mongolia is Ulaanbaatar"
-//
-template <typename... Args>
-int FPrintF(absl::Nonnull<std::FILE*> output, const FormatSpec<Args...>& format,
-            const Args&... args) {
-  return str_format_internal::FprintF(
-      output, str_format_internal::UntypedFormatSpecImpl::Extract(format),
-      {str_format_internal::FormatArgImpl(args)...});
-}
-
-// SNPrintF()
-//
-// Writes to a sized buffer given a format string and zero or more arguments.
-// This function is functionally equivalent to `std::snprintf()` (and
-// type-safe); prefer `absl::SNPrintF()` over `std::snprintf()`.
-//
-// In particular, a successful call to `absl::SNPrintF()` writes at most `size`
-// bytes of the formatted output to `output`, including a NUL-terminator, and
-// returns the number of bytes that would have been written if truncation did
-// not occur. In the event of an error, a negative value is returned and `errno`
-// is set.
-//
-// Example:
-//
-//   std::string_view s = "Ulaanbaatar";
-//   char output[128];
-//   absl::SNPrintF(output, sizeof(output),
-//                  "The capital of Mongolia is %s", s);
-//
-//   Post-condition: output == "The capital of Mongolia is Ulaanbaatar"
-//
-template <typename... Args>
-int SNPrintF(absl::Nonnull<char*> output, std::size_t size,
-             const FormatSpec<Args...>& format, const Args&... args) {
-  return str_format_internal::SnprintF(
-      output, size, str_format_internal::UntypedFormatSpecImpl::Extract(format),
-      {str_format_internal::FormatArgImpl(args)...});
-}
-
-// -----------------------------------------------------------------------------
-// Custom Output Formatting Functions
-// -----------------------------------------------------------------------------
-
-// FormatRawSink
-//
-// FormatRawSink is a type erased wrapper around arbitrary sink objects
-// specifically used as an argument to `Format()`.
-//
-// All the object has to do define an overload of `AbslFormatFlush()` for the
-// sink, usually by adding a ADL-based free function in the same namespace as
-// the sink:
-//
-//   void AbslFormatFlush(MySink* dest, absl::string_view part);
-//
-// where `dest` is the pointer passed to `absl::Format()`. The function should
-// append `part` to `dest`.
-//
-// FormatRawSink does not own the passed sink object. The passed object must
-// outlive the FormatRawSink.
-class FormatRawSink {
- public:
-  // Implicitly convert from any type that provides the hook function as
-  // described above.
-  template <typename T,
-            typename = typename std::enable_if<std::is_constructible<
-                str_format_internal::FormatRawSinkImpl, T*>::value>::type>
-  FormatRawSink(absl::Nonnull<T*> raw)  // NOLINT
-      : sink_(raw) {}
-
- private:
-  friend str_format_internal::FormatRawSinkImpl;
-  str_format_internal::FormatRawSinkImpl sink_;
-};
-
-// Format()
-//
-// Writes a formatted string to an arbitrary sink object (implementing the
-// `absl::FormatRawSink` interface), using a format string and zero or more
-// additional arguments.
-//
-// By default, `std::string`, `std::ostream`, and `absl::Cord` are supported as
-// destination objects. If a `std::string` is used the formatted string is
-// appended to it.
-//
-// `absl::Format()` is a generic version of `absl::StrAppendFormat()`, for
-// custom sinks. The format string, like format strings for `StrFormat()`, is
-// checked at compile-time.
-//
-// On failure, this function returns `false` and the state of the sink is
-// unspecified.
-template <typename... Args>
-bool Format(FormatRawSink raw_sink, const FormatSpec<Args...>& format,
-            const Args&... args) {
-  return str_format_internal::FormatUntyped(
-      str_format_internal::FormatRawSinkImpl::Extract(raw_sink),
-      str_format_internal::UntypedFormatSpecImpl::Extract(format),
-      {str_format_internal::FormatArgImpl(args)...});
-}
-
-// FormatArg
-//
-// A type-erased handle to a format argument specifically used as an argument to
-// `FormatUntyped()`. You may construct `FormatArg` by passing
-// reference-to-const of any printable type. `FormatArg` is both copyable and
-// assignable. The source data must outlive the `FormatArg` instance. See
-// example below.
-//
-using FormatArg = str_format_internal::FormatArgImpl;
-
-// FormatUntyped()
-//
-// Writes a formatted string to an arbitrary sink object (implementing the
-// `absl::FormatRawSink` interface), using an `UntypedFormatSpec` and zero or
-// more additional arguments.
-//
-// This function acts as the most generic formatting function in the
-// `str_format` library. The caller provides a raw sink, an unchecked format
-// string, and (usually) a runtime specified list of arguments; no compile-time
-// checking of formatting is performed within this function. As a result, a
-// caller should check the return value to verify that no error occurred.
-// On failure, this function returns `false` and the state of the sink is
-// unspecified.
-//
-// The arguments are provided in an `absl::Span<const absl::FormatArg>`.
-// Each `absl::FormatArg` object binds to a single argument and keeps a
-// reference to it. The values used to create the `FormatArg` objects must
-// outlive this function call.
-//
-// Example:
-//
-//   std::optional<std::string> FormatDynamic(
-//       const std::string& in_format,
-//       const vector<std::string>& in_args) {
-//     std::string out;
-//     std::vector<absl::FormatArg> args;
-//     for (const auto& v : in_args) {
-//       // It is important that 'v' is a reference to the objects in in_args.
-//       // The values we pass to FormatArg must outlive the call to
-//       // FormatUntyped.
-//       args.emplace_back(v);
-//     }
-//     absl::UntypedFormatSpec format(in_format);
-//     if (!absl::FormatUntyped(&out, format, args)) {
-//       return std::nullopt;
-//     }
-//     return std::move(out);
-//   }
-//
-ABSL_MUST_USE_RESULT inline bool FormatUntyped(
-    FormatRawSink raw_sink, const UntypedFormatSpec& format,
-    absl::Span<const FormatArg> args) {
-  return str_format_internal::FormatUntyped(
-      str_format_internal::FormatRawSinkImpl::Extract(raw_sink),
-      str_format_internal::UntypedFormatSpecImpl::Extract(format), args);
-}
-
-//------------------------------------------------------------------------------
-// StrFormat Extensions
-//------------------------------------------------------------------------------
-//
-// AbslStringify()
-//
-// A simpler customization API for formatting user-defined types using
-// absl::StrFormat(). The API relies on detecting an overload in the
-// user-defined type's namespace of a free (non-member) `AbslStringify()`
-// function as a friend definition with the following signature:
-//
-// template <typename Sink>
-// void AbslStringify(Sink& sink, const X& value);
-//
-// An `AbslStringify()` overload for a type should only be declared in the same
-// file and namespace as said type.
-//
-// Note that unlike with AbslFormatConvert(), AbslStringify() does not allow
-// customization of allowed conversion characters. AbslStringify() uses `%v` as
-// the underlying conversion specififer. Additionally, AbslStringify() supports
-// use with absl::StrCat while AbslFormatConvert() does not.
-//
-// Example:
-//
-// struct Point {
-//   // To add formatting support to `Point`, we simply need to add a free
-//   // (non-member) function `AbslStringify()`. This method prints in the
-//   // request format using the underlying `%v` specifier. You can add such a
-//   // free function using a friend declaration within the body of the class.
-//   // The sink parameter is a templated type to avoid requiring dependencies.
-//   template <typename Sink>
-//   friend void AbslStringify(Sink& sink, const Point& p) {
-//     absl::Format(&sink, "(%v, %v)", p.x, p.y);
-//   }
-//
-//   int x;
-//   int y;
-// };
-//
-// AbslFormatConvert()
-//
-// The StrFormat library provides a customization API for formatting
-// user-defined types using absl::StrFormat(). The API relies on detecting an
-// overload in the user-defined type's namespace of a free (non-member)
-// `AbslFormatConvert()` function, usually as a friend definition with the
-// following signature:
-//
-// absl::FormatConvertResult<...> AbslFormatConvert(
-//     const X& value,
-//     const absl::FormatConversionSpec& spec,
-//     absl::FormatSink *sink);
-//
-// An `AbslFormatConvert()` overload for a type should only be declared in the
-// same file and namespace as said type.
-//
-// The abstractions within this definition include:
-//
-// * An `absl::FormatConversionSpec` to specify the fields to pull from a
-//   user-defined type's format string
-// * An `absl::FormatSink` to hold the converted string data during the
-//   conversion process.
-// * An `absl::FormatConvertResult` to hold the status of the returned
-//   formatting operation
-//
-// The return type encodes all the conversion characters that your
-// AbslFormatConvert() routine accepts.  The return value should be {true}.
-// A return value of {false} will result in `StrFormat()` returning
-// an empty string.  This result will be propagated to the result of
-// `FormatUntyped`.
-//
-// Example:
-//
-// struct Point {
-//   // To add formatting support to `Point`, we simply need to add a free
-//   // (non-member) function `AbslFormatConvert()`.  This method interprets
-//   // `spec` to print in the request format. The allowed conversion characters
-//   // can be restricted via the type of the result, in this example
-//   // string and integral formatting are allowed (but not, for instance
-//   // floating point characters like "%f").  You can add such a free function
-//   // using a friend declaration within the body of the class:
-//   friend absl::FormatConvertResult<absl::FormatConversionCharSet::kString |
-//                                    absl::FormatConversionCharSet::kIntegral>
-//   AbslFormatConvert(const Point& p, const absl::FormatConversionSpec& spec,
-//                     absl::FormatSink* s) {
-//     if (spec.conversion_char() == absl::FormatConversionChar::s) {
-//       absl::Format(s, "x=%vy=%v", p.x, p.y);
-//     } else {
-//       absl::Format(s, "%v,%v", p.x, p.y);
-//     }
-//     return {true};
-//   }
-//
-//   int x;
-//   int y;
-// };
-
-// clang-format off
-
-// FormatConversionChar
-//
-// Specifies the formatting character provided in the format string
-// passed to `StrFormat()`.
-enum class FormatConversionChar : uint8_t {
-  c, s,                    // text
-  d, i, o, u, x, X,        // int
-  f, F, e, E, g, G, a, A,  // float
-  n, p, v                  // misc
-};
-// clang-format on
-
-// FormatConversionSpec
-//
-// Specifies modifications to the conversion of the format string, through use
-// of one or more format flags in the source format string.
-class FormatConversionSpec {
- public:
-  // FormatConversionSpec::is_basic()
-  //
-  // Indicates that width and precision are not specified, and no additional
-  // flags are set for this conversion character in the format string.
-  bool is_basic() const { return impl_.is_basic(); }
-
-  // FormatConversionSpec::has_left_flag()
-  //
-  // Indicates whether the result should be left justified for this conversion
-  // character in the format string. This flag is set through use of a '-'
-  // character in the format string. E.g. "%-s"
-  bool has_left_flag() const { return impl_.has_left_flag(); }
-
-  // FormatConversionSpec::has_show_pos_flag()
-  //
-  // Indicates whether a sign column is prepended to the result for this
-  // conversion character in the format string, even if the result is positive.
-  // This flag is set through use of a '+' character in the format string.
-  // E.g. "%+d"
-  bool has_show_pos_flag() const { return impl_.has_show_pos_flag(); }
-
-  // FormatConversionSpec::has_sign_col_flag()
-  //
-  // Indicates whether a mandatory sign column is added to the result for this
-  // conversion character. This flag is set through use of a space character
-  // (' ') in the format string. E.g. "% i"
-  bool has_sign_col_flag() const { return impl_.has_sign_col_flag(); }
-
-  // FormatConversionSpec::has_alt_flag()
-  //
-  // Indicates whether an "alternate" format is applied to the result for this
-  // conversion character. Alternative forms depend on the type of conversion
-  // character, and unallowed alternatives are undefined. This flag is set
-  // through use of a '#' character in the format string. E.g. "%#h"
-  bool has_alt_flag() const { return impl_.has_alt_flag(); }
-
-  // FormatConversionSpec::has_zero_flag()
-  //
-  // Indicates whether zeroes should be prepended to the result for this
-  // conversion character instead of spaces. This flag is set through use of the
-  // '0' character in the format string. E.g. "%0f"
-  bool has_zero_flag() const { return impl_.has_zero_flag(); }
-
-  // FormatConversionSpec::conversion_char()
-  //
-  // Returns the underlying conversion character.
-  FormatConversionChar conversion_char() const {
-    return impl_.conversion_char();
-  }
-
-  // FormatConversionSpec::width()
-  //
-  // Returns the specified width (indicated through use of a non-zero integer
-  // value or '*' character) of the conversion character. If width is
-  // unspecified, it returns a negative value.
-  int width() const { return impl_.width(); }
-
-  // FormatConversionSpec::precision()
-  //
-  // Returns the specified precision (through use of the '.' character followed
-  // by a non-zero integer value or '*' character) of the conversion character.
-  // If precision is unspecified, it returns a negative value.
-  int precision() const { return impl_.precision(); }
-
- private:
-  explicit FormatConversionSpec(
-      str_format_internal::FormatConversionSpecImpl impl)
-      : impl_(impl) {}
-
-  friend str_format_internal::FormatConversionSpecImpl;
-
-  absl::str_format_internal::FormatConversionSpecImpl impl_;
-};
-
-// Type safe OR operator for FormatConversionCharSet to allow accepting multiple
-// conversion chars in custom format converters.
-constexpr FormatConversionCharSet operator|(FormatConversionCharSet a,
-                                            FormatConversionCharSet b) {
-  return static_cast<FormatConversionCharSet>(static_cast<uint64_t>(a) |
-                                              static_cast<uint64_t>(b));
-}
-
-// FormatConversionCharSet
-//
-// Specifies the _accepted_ conversion types as a template parameter to
-// FormatConvertResult for custom implementations of `AbslFormatConvert`.
-// Note the helper predefined alias definitions (kIntegral, etc.) below.
-enum class FormatConversionCharSet : uint64_t {
-  // text
-  c = str_format_internal::FormatConversionCharToConvInt('c'),
-  s = str_format_internal::FormatConversionCharToConvInt('s'),
-  // integer
-  d = str_format_internal::FormatConversionCharToConvInt('d'),
-  i = str_format_internal::FormatConversionCharToConvInt('i'),
-  o = str_format_internal::FormatConversionCharToConvInt('o'),
-  u = str_format_internal::FormatConversionCharToConvInt('u'),
-  x = str_format_internal::FormatConversionCharToConvInt('x'),
-  X = str_format_internal::FormatConversionCharToConvInt('X'),
-  // Float
-  f = str_format_internal::FormatConversionCharToConvInt('f'),
-  F = str_format_internal::FormatConversionCharToConvInt('F'),
-  e = str_format_internal::FormatConversionCharToConvInt('e'),
-  E = str_format_internal::FormatConversionCharToConvInt('E'),
-  g = str_format_internal::FormatConversionCharToConvInt('g'),
-  G = str_format_internal::FormatConversionCharToConvInt('G'),
-  a = str_format_internal::FormatConversionCharToConvInt('a'),
-  A = str_format_internal::FormatConversionCharToConvInt('A'),
-  // misc
-  n = str_format_internal::FormatConversionCharToConvInt('n'),
-  p = str_format_internal::FormatConversionCharToConvInt('p'),
-  v = str_format_internal::FormatConversionCharToConvInt('v'),
-
-  // Used for width/precision '*' specification.
-  kStar = static_cast<uint64_t>(
-      absl::str_format_internal::FormatConversionCharSetInternal::kStar),
-  // Some predefined values:
-  kIntegral = d | i | u | o | x | X,
-  kFloating = a | e | f | g | A | E | F | G,
-  kNumeric = kIntegral | kFloating,
-  kString = s,
-  kPointer = p,
-};
-
-// FormatSink
-//
-// A format sink is a generic abstraction to which conversions may write their
-// formatted string data. `absl::FormatConvert()` uses this sink to write its
-// formatted string.
-//
-class FormatSink {
- public:
-  // FormatSink::Append()
-  //
-  // Appends `count` copies of `ch` to the format sink.
-  void Append(size_t count, char ch) { sink_->Append(count, ch); }
-
-  // Overload of FormatSink::Append() for appending the characters of a string
-  // view to a format sink.
-  void Append(string_view v) { sink_->Append(v); }
-
-  // FormatSink::PutPaddedString()
-  //
-  // Appends `precision` number of bytes of `v` to the format sink. If this is
-  // less than `width`, spaces will be appended first (if `left` is false), or
-  // after (if `left` is true) to ensure the total amount appended is
-  // at least `width`.
-  bool PutPaddedString(string_view v, int width, int precision, bool left) {
-    return sink_->PutPaddedString(v, width, precision, left);
-  }
-
-  // Support `absl::Format(&sink, format, args...)`.
-  friend void AbslFormatFlush(absl::Nonnull<FormatSink*> sink,
-                              absl::string_view v) {
-    sink->Append(v);
-  }
-
- private:
-  friend str_format_internal::FormatSinkImpl;
-  explicit FormatSink(absl::Nonnull<str_format_internal::FormatSinkImpl*> s)
-      : sink_(s) {}
-  absl::Nonnull<str_format_internal::FormatSinkImpl*> sink_;
-};
-
-// FormatConvertResult
-//
-// Indicates whether a call to AbslFormatConvert() was successful.
-// This return type informs the StrFormat extension framework (through
-// ADL but using the return type) of what conversion characters are supported.
-// It is strongly discouraged to return {false}, as this will result in an
-// empty string in StrFormat.
-template <FormatConversionCharSet C>
-struct FormatConvertResult {
-  bool value;
-};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_STR_FORMAT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/str_join.h b/projects/llm_framework/include/abseil-cpp/absl/strings/str_join.h
deleted file mode 100644
index 6a92c0f5..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/str_join.h
+++ /dev/null
@@ -1,287 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: str_join.h
-// -----------------------------------------------------------------------------
-//
-// This header file contains functions for joining a range of elements and
-// returning the result as a std::string. StrJoin operations are specified by
-// passing a range, a separator string to use between the elements joined, and
-// an optional Formatter responsible for converting each argument in the range
-// to a string. If omitted, a default `AlphaNumFormatter()` is called on the
-// elements to be joined, using the same formatting that `absl::StrCat()` uses.
-// This package defines a number of default formatters, and you can define your
-// own implementations.
-//
-// Ranges are specified by passing a container with `std::begin()` and
-// `std::end()` iterators, container-specific `begin()` and `end()` iterators, a
-// brace-initialized `std::initializer_list`, or a `std::tuple` of heterogeneous
-// objects. The separator string is specified as an `absl::string_view`.
-//
-// Because the default formatter uses the `absl::AlphaNum` class,
-// `absl::StrJoin()`, like `absl::StrCat()`, will work out-of-the-box on
-// collections of strings, ints, floats, doubles, etc.
-//
-// Example:
-//
-//   std::vector<std::string> v = {"foo", "bar", "baz"};
-//   std::string s = absl::StrJoin(v, "-");
-//   EXPECT_EQ("foo-bar-baz", s);
-//
-// See comments on the `absl::StrJoin()` function for more examples.
-
-#ifndef ABSL_STRINGS_STR_JOIN_H_
-#define ABSL_STRINGS_STR_JOIN_H_
-
-#include <cstdio>
-#include <cstring>
-#include <initializer_list>
-#include <iterator>
-#include <string>
-#include <tuple>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/macros.h"
-#include "absl/strings/internal/str_join_internal.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// -----------------------------------------------------------------------------
-// Concept: Formatter
-// -----------------------------------------------------------------------------
-//
-// A Formatter is a function object that is responsible for formatting its
-// argument as a string and appending it to a given output std::string.
-// Formatters may be implemented as function objects, lambdas, or normal
-// functions. You may provide your own Formatter to enable `absl::StrJoin()` to
-// work with arbitrary types.
-//
-// The following is an example of a custom Formatter that uses
-// `absl::FormatDuration` to join a list of `absl::Duration`s.
-//
-//   std::vector<absl::Duration> v = {absl::Seconds(1), absl::Milliseconds(10)};
-//   std::string s =
-//       absl::StrJoin(v, ", ", [](std::string* out, absl::Duration dur) {
-//         absl::StrAppend(out, absl::FormatDuration(dur));
-//       });
-//   EXPECT_EQ(s, "1s, 10ms");
-//
-// The following standard formatters are provided within this file:
-//
-// - `AlphaNumFormatter()` (the default)
-// - `StreamFormatter()`
-// - `PairFormatter()`
-// - `DereferenceFormatter()`
-
-// AlphaNumFormatter()
-//
-// Default formatter used if none is specified. Uses `absl::AlphaNum` to convert
-// numeric arguments to strings.
-inline strings_internal::AlphaNumFormatterImpl AlphaNumFormatter() {
-  return strings_internal::AlphaNumFormatterImpl();
-}
-
-// StreamFormatter()
-//
-// Formats its argument using the << operator.
-inline strings_internal::StreamFormatterImpl StreamFormatter() {
-  return strings_internal::StreamFormatterImpl();
-}
-
-// Function Template: PairFormatter(Formatter, absl::string_view, Formatter)
-//
-// Formats a `std::pair` by putting a given separator between the pair's
-// `.first` and `.second` members. This formatter allows you to specify
-// custom Formatters for both the first and second member of each pair.
-template <typename FirstFormatter, typename SecondFormatter>
-inline strings_internal::PairFormatterImpl<FirstFormatter, SecondFormatter>
-PairFormatter(FirstFormatter f1, absl::string_view sep, SecondFormatter f2) {
-  return strings_internal::PairFormatterImpl<FirstFormatter, SecondFormatter>(
-      std::move(f1), sep, std::move(f2));
-}
-
-// Function overload of PairFormatter() for using a default
-// `AlphaNumFormatter()` for each Formatter in the pair.
-inline strings_internal::PairFormatterImpl<
-    strings_internal::AlphaNumFormatterImpl,
-    strings_internal::AlphaNumFormatterImpl>
-PairFormatter(absl::string_view sep) {
-  return PairFormatter(AlphaNumFormatter(), sep, AlphaNumFormatter());
-}
-
-// Function Template: DereferenceFormatter(Formatter)
-//
-// Formats its argument by dereferencing it and then applying the given
-// formatter. This formatter is useful for formatting a container of
-// pointer-to-T. This pattern often shows up when joining repeated fields in
-// protocol buffers.
-template <typename Formatter>
-strings_internal::DereferenceFormatterImpl<Formatter> DereferenceFormatter(
-    Formatter&& f) {
-  return strings_internal::DereferenceFormatterImpl<Formatter>(
-      std::forward<Formatter>(f));
-}
-
-// Function overload of `DereferenceFormatter()` for using a default
-// `AlphaNumFormatter()`.
-inline strings_internal::DereferenceFormatterImpl<
-    strings_internal::AlphaNumFormatterImpl>
-DereferenceFormatter() {
-  return strings_internal::DereferenceFormatterImpl<
-      strings_internal::AlphaNumFormatterImpl>(AlphaNumFormatter());
-}
-
-// -----------------------------------------------------------------------------
-// StrJoin()
-// -----------------------------------------------------------------------------
-//
-// Joins a range of elements and returns the result as a std::string.
-// `absl::StrJoin()` takes a range, a separator string to use between the
-// elements joined, and an optional Formatter responsible for converting each
-// argument in the range to a string.
-//
-// If omitted, the default `AlphaNumFormatter()` is called on the elements to be
-// joined.
-//
-// Example 1:
-//   // Joins a collection of strings. This pattern also works with a collection
-//   // of `absl::string_view` or even `const char*`.
-//   std::vector<std::string> v = {"foo", "bar", "baz"};
-//   std::string s = absl::StrJoin(v, "-");
-//   EXPECT_EQ(s, "foo-bar-baz");
-//
-// Example 2:
-//   // Joins the values in the given `std::initializer_list<>` specified using
-//   // brace initialization. This pattern also works with an initializer_list
-//   // of ints or `absl::string_view` -- any `AlphaNum`-compatible type.
-//   std::string s = absl::StrJoin({"foo", "bar", "baz"}, "-");
-//   EXPECT_EQs, "foo-bar-baz");
-//
-// Example 3:
-//   // Joins a collection of ints. This pattern also works with floats,
-//   // doubles, int64s -- any `StrCat()`-compatible type.
-//   std::vector<int> v = {1, 2, 3, -4};
-//   std::string s = absl::StrJoin(v, "-");
-//   EXPECT_EQ(s, "1-2-3--4");
-//
-// Example 4:
-//   // Joins a collection of pointer-to-int. By default, pointers are
-//   // dereferenced and the pointee is formatted using the default format for
-//   // that type; such dereferencing occurs for all levels of indirection, so
-//   // this pattern works just as well for `std::vector<int**>` as for
-//   // `std::vector<int*>`.
-//   int x = 1, y = 2, z = 3;
-//   std::vector<int*> v = {&x, &y, &z};
-//   std::string s = absl::StrJoin(v, "-");
-//   EXPECT_EQ(s, "1-2-3");
-//
-// Example 5:
-//   // Dereferencing of `std::unique_ptr<>` is also supported:
-//   std::vector<std::unique_ptr<int>> v
-//   v.emplace_back(new int(1));
-//   v.emplace_back(new int(2));
-//   v.emplace_back(new int(3));
-//   std::string s = absl::StrJoin(v, "-");
-//   EXPECT_EQ(s, "1-2-3");
-//
-// Example 6:
-//   // Joins a `std::map`, with each key-value pair separated by an equals
-//   // sign. This pattern would also work with, say, a
-//   // `std::vector<std::pair<>>`.
-//   std::map<std::string, int> m = {
-//       {"a", 1},
-//       {"b", 2},
-//       {"c", 3}};
-//   std::string s = absl::StrJoin(m, ",", absl::PairFormatter("="));
-//   EXPECT_EQ(s, "a=1,b=2,c=3");
-//
-// Example 7:
-//   // These examples show how `absl::StrJoin()` handles a few common edge
-//   // cases:
-//   std::vector<std::string> v_empty;
-//   EXPECT_EQ(absl::StrJoin(v_empty, "-"), "");
-//
-//   std::vector<std::string> v_one_item = {"foo"};
-//   EXPECT_EQ(absl::StrJoin(v_one_item, "-"), "foo");
-//
-//   std::vector<std::string> v_empty_string = {""};
-//   EXPECT_EQ(absl::StrJoin(v_empty_string, "-"), "");
-//
-//   std::vector<std::string> v_one_item_empty_string = {"a", ""};
-//   EXPECT_EQ(absl::StrJoin(v_one_item_empty_string, "-"), "a-");
-//
-//   std::vector<std::string> v_two_empty_string = {"", ""};
-//   EXPECT_EQ(absl::StrJoin(v_two_empty_string, "-"), "-");
-//
-// Example 8:
-//   // Joins a `std::tuple<T...>` of heterogeneous types, converting each to
-//   // a std::string using the `absl::AlphaNum` class.
-//   std::string s = absl::StrJoin(std::make_tuple(123, "abc", 0.456), "-");
-//   EXPECT_EQ(s, "123-abc-0.456");
-
-template <typename Iterator, typename Formatter>
-std::string StrJoin(Iterator start, Iterator end, absl::string_view sep,
-                    Formatter&& fmt) {
-  return strings_internal::JoinAlgorithm(start, end, sep, fmt);
-}
-
-template <typename Range, typename Formatter>
-std::string StrJoin(const Range& range, absl::string_view separator,
-                    Formatter&& fmt) {
-  return strings_internal::JoinRange(range, separator, fmt);
-}
-
-template <typename T, typename Formatter>
-std::string StrJoin(std::initializer_list<T> il, absl::string_view separator,
-                    Formatter&& fmt) {
-  return strings_internal::JoinRange(il, separator, fmt);
-}
-
-template <typename... T, typename Formatter>
-std::string StrJoin(const std::tuple<T...>& value, absl::string_view separator,
-                    Formatter&& fmt) {
-  return strings_internal::JoinAlgorithm(value, separator, fmt);
-}
-
-template <typename Iterator>
-std::string StrJoin(Iterator start, Iterator end, absl::string_view separator) {
-  return strings_internal::JoinRange(start, end, separator);
-}
-
-template <typename Range>
-std::string StrJoin(const Range& range, absl::string_view separator) {
-  return strings_internal::JoinRange(range, separator);
-}
-
-template <typename T>
-std::string StrJoin(std::initializer_list<T> il,
-                    absl::string_view separator) {
-  return strings_internal::JoinRange(il, separator);
-}
-
-template <typename... T>
-std::string StrJoin(const std::tuple<T...>& value,
-                    absl::string_view separator) {
-  return strings_internal::JoinAlgorithm(value, separator, AlphaNumFormatter());
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_STR_JOIN_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/str_replace.h b/projects/llm_framework/include/abseil-cpp/absl/strings/str_replace.h
deleted file mode 100644
index e77ced3e..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/str_replace.h
+++ /dev/null
@@ -1,222 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: str_replace.h
-// -----------------------------------------------------------------------------
-//
-// This file defines `absl::StrReplaceAll()`, a general-purpose string
-// replacement function designed for large, arbitrary text substitutions,
-// especially on strings which you are receiving from some other system for
-// further processing (e.g. processing regular expressions, escaping HTML
-// entities, etc.). `StrReplaceAll` is designed to be efficient even when only
-// one substitution is being performed, or when substitution is rare.
-//
-// If the string being modified is known at compile-time, and the substitutions
-// vary, `absl::Substitute()` may be a better choice.
-//
-// Example:
-//
-// std::string html_escaped = absl::StrReplaceAll(user_input, {
-//                                                {"&", "&amp;"},
-//                                                {"<", "&lt;"},
-//                                                {">", "&gt;"},
-//                                                {"\"", "&quot;"},
-//                                                {"'", "&#39;"}});
-#ifndef ABSL_STRINGS_STR_REPLACE_H_
-#define ABSL_STRINGS_STR_REPLACE_H_
-
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "absl/base/attributes.h"
-#include "absl/base/nullability.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// StrReplaceAll()
-//
-// Replaces character sequences within a given string with replacements provided
-// within an initializer list of key/value pairs. Candidate replacements are
-// considered in order as they occur within the string, with earlier matches
-// taking precedence, and longer matches taking precedence for candidates
-// starting at the same position in the string. Once a substitution is made, the
-// replaced text is not considered for any further substitutions.
-//
-// Example:
-//
-//   std::string s = absl::StrReplaceAll(
-//       "$who bought $count #Noun. Thanks $who!",
-//       {{"$count", absl::StrCat(5)},
-//        {"$who", "Bob"},
-//        {"#Noun", "Apples"}});
-//   EXPECT_EQ("Bob bought 5 Apples. Thanks Bob!", s);
-ABSL_MUST_USE_RESULT std::string StrReplaceAll(
-    absl::string_view s,
-    std::initializer_list<std::pair<absl::string_view, absl::string_view>>
-        replacements);
-
-// Overload of `StrReplaceAll()` to accept a container of key/value replacement
-// pairs (typically either an associative map or a `std::vector` of `std::pair`
-// elements). A vector of pairs is generally more efficient.
-//
-// Examples:
-//
-//   std::map<const absl::string_view, const absl::string_view> replacements;
-//   replacements["$who"] = "Bob";
-//   replacements["$count"] = "5";
-//   replacements["#Noun"] = "Apples";
-//   std::string s = absl::StrReplaceAll(
-//       "$who bought $count #Noun. Thanks $who!",
-//       replacements);
-//   EXPECT_EQ("Bob bought 5 Apples. Thanks Bob!", s);
-//
-//   // A std::vector of std::pair elements can be more efficient.
-//   std::vector<std::pair<const absl::string_view, std::string>> replacements;
-//   replacements.push_back({"&", "&amp;"});
-//   replacements.push_back({"<", "&lt;"});
-//   replacements.push_back({">", "&gt;"});
-//   std::string s = absl::StrReplaceAll("if (ptr < &foo)",
-//                                  replacements);
-//   EXPECT_EQ("if (ptr &lt; &amp;foo)", s);
-template <typename StrToStrMapping>
-std::string StrReplaceAll(absl::string_view s,
-                          const StrToStrMapping& replacements);
-
-// Overload of `StrReplaceAll()` to replace character sequences within a given
-// output string *in place* with replacements provided within an initializer
-// list of key/value pairs, returning the number of substitutions that occurred.
-//
-// Example:
-//
-//   std::string s = std::string("$who bought $count #Noun. Thanks $who!");
-//   int count;
-//   count = absl::StrReplaceAll({{"$count", absl::StrCat(5)},
-//                               {"$who", "Bob"},
-//                               {"#Noun", "Apples"}}, &s);
-//  EXPECT_EQ(count, 4);
-//  EXPECT_EQ("Bob bought 5 Apples. Thanks Bob!", s);
-int StrReplaceAll(
-    std::initializer_list<std::pair<absl::string_view, absl::string_view>>
-        replacements,
-    absl::Nonnull<std::string*> target);
-
-// Overload of `StrReplaceAll()` to replace patterns within a given output
-// string *in place* with replacements provided within a container of key/value
-// pairs.
-//
-// Example:
-//
-//   std::string s = std::string("if (ptr < &foo)");
-//   int count = absl::StrReplaceAll({{"&", "&amp;"},
-//                                    {"<", "&lt;"},
-//                                    {">", "&gt;"}}, &s);
-//  EXPECT_EQ(count, 2);
-//  EXPECT_EQ("if (ptr &lt; &amp;foo)", s);
-template <typename StrToStrMapping>
-int StrReplaceAll(const StrToStrMapping& replacements,
-                  absl::Nonnull<std::string*> target);
-
-// Implementation details only, past this point.
-namespace strings_internal {
-
-struct ViableSubstitution {
-  absl::string_view old;
-  absl::string_view replacement;
-  size_t offset;
-
-  ViableSubstitution(absl::string_view old_str,
-                     absl::string_view replacement_str, size_t offset_val)
-      : old(old_str), replacement(replacement_str), offset(offset_val) {}
-
-  // One substitution occurs "before" another (takes priority) if either
-  // it has the lowest offset, or it has the same offset but a larger size.
-  bool OccursBefore(const ViableSubstitution& y) const {
-    if (offset != y.offset) return offset < y.offset;
-    return old.size() > y.old.size();
-  }
-};
-
-// Build a vector of ViableSubstitutions based on the given list of
-// replacements. subs can be implemented as a priority_queue. However, it turns
-// out that most callers have small enough a list of substitutions that the
-// overhead of such a queue isn't worth it.
-template <typename StrToStrMapping>
-std::vector<ViableSubstitution> FindSubstitutions(
-    absl::string_view s, const StrToStrMapping& replacements) {
-  std::vector<ViableSubstitution> subs;
-  subs.reserve(replacements.size());
-
-  for (const auto& rep : replacements) {
-    using std::get;
-    absl::string_view old(get<0>(rep));
-
-    size_t pos = s.find(old);
-    if (pos == s.npos) continue;
-
-    // Ignore attempts to replace "". This condition is almost never true,
-    // but above condition is frequently true. That's why we test for this
-    // now and not before.
-    if (old.empty()) continue;
-
-    subs.emplace_back(old, get<1>(rep), pos);
-
-    // Insertion sort to ensure the last ViableSubstitution comes before
-    // all the others.
-    size_t index = subs.size();
-    while (--index && subs[index - 1].OccursBefore(subs[index])) {
-      std::swap(subs[index], subs[index - 1]);
-    }
-  }
-  return subs;
-}
-
-int ApplySubstitutions(absl::string_view s,
-                       absl::Nonnull<std::vector<ViableSubstitution>*> subs_ptr,
-                       absl::Nonnull<std::string*> result_ptr);
-
-}  // namespace strings_internal
-
-template <typename StrToStrMapping>
-std::string StrReplaceAll(absl::string_view s,
-                          const StrToStrMapping& replacements) {
-  auto subs = strings_internal::FindSubstitutions(s, replacements);
-  std::string result;
-  result.reserve(s.size());
-  strings_internal::ApplySubstitutions(s, &subs, &result);
-  return result;
-}
-
-template <typename StrToStrMapping>
-int StrReplaceAll(const StrToStrMapping& replacements,
-                  absl::Nonnull<std::string*> target) {
-  auto subs = strings_internal::FindSubstitutions(*target, replacements);
-  if (subs.empty()) return 0;
-
-  std::string result;
-  result.reserve(target->size());
-  int substitutions =
-      strings_internal::ApplySubstitutions(*target, &subs, &result);
-  target->swap(result);
-  return substitutions;
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_STR_REPLACE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/str_split.h b/projects/llm_framework/include/abseil-cpp/absl/strings/str_split.h
deleted file mode 100644
index 49bff7b2..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/str_split.h
+++ /dev/null
@@ -1,565 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: str_split.h
-// -----------------------------------------------------------------------------
-//
-// This file contains functions for splitting strings. It defines the main
-// `StrSplit()` function, several delimiters for determining the boundaries on
-// which to split the string, and predicates for filtering delimited results.
-// `StrSplit()` adapts the returned collection to the type specified by the
-// caller.
-//
-// Example:
-//
-//   // Splits the given string on commas. Returns the results in a
-//   // vector of strings.
-//   std::vector<std::string> v = absl::StrSplit("a,b,c", ',');
-//   // Can also use ","
-//   // v[0] == "a", v[1] == "b", v[2] == "c"
-//
-// See StrSplit() below for more information.
-#ifndef ABSL_STRINGS_STR_SPLIT_H_
-#define ABSL_STRINGS_STR_SPLIT_H_
-
-#include <algorithm>
-#include <cstddef>
-#include <map>
-#include <set>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "absl/base/internal/raw_logging.h"
-#include "absl/base/macros.h"
-#include "absl/strings/internal/str_split_internal.h"
-#include "absl/strings/string_view.h"
-#include "absl/strings/strip.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-//------------------------------------------------------------------------------
-// Delimiters
-//------------------------------------------------------------------------------
-//
-// `StrSplit()` uses delimiters to define the boundaries between elements in the
-// provided input. Several `Delimiter` types are defined below. If a string
-// (`const char*`, `std::string`, or `absl::string_view`) is passed in place of
-// an explicit `Delimiter` object, `StrSplit()` treats it the same way as if it
-// were passed a `ByString` delimiter.
-//
-// A `Delimiter` is an object with a `Find()` function that knows how to find
-// the first occurrence of itself in a given `absl::string_view`.
-//
-// The following `Delimiter` types are available for use within `StrSplit()`:
-//
-//   - `ByString` (default for string arguments)
-//   - `ByChar` (default for a char argument)
-//   - `ByAnyChar`
-//   - `ByLength`
-//   - `MaxSplits`
-//
-// A Delimiter's `Find()` member function will be passed an input `text` that is
-// to be split and a position (`pos`) to begin searching for the next delimiter
-// in `text`. The returned absl::string_view should refer to the next occurrence
-// (after `pos`) of the represented delimiter; this returned absl::string_view
-// represents the next location where the input `text` should be broken.
-//
-// The returned absl::string_view may be zero-length if the Delimiter does not
-// represent a part of the string (e.g., a fixed-length delimiter). If no
-// delimiter is found in the input `text`, a zero-length absl::string_view
-// referring to `text.end()` should be returned (e.g.,
-// `text.substr(text.size())`). It is important that the returned
-// absl::string_view always be within the bounds of the input `text` given as an
-// argument--it must not refer to a string that is physically located outside of
-// the given string.
-//
-// The following example is a simple Delimiter object that is created with a
-// single char and will look for that char in the text passed to the `Find()`
-// function:
-//
-//   struct SimpleDelimiter {
-//     const char c_;
-//     explicit SimpleDelimiter(char c) : c_(c) {}
-//     absl::string_view Find(absl::string_view text, size_t pos) {
-//       auto found = text.find(c_, pos);
-//       if (found == absl::string_view::npos)
-//         return text.substr(text.size());
-//
-//       return text.substr(found, 1);
-//     }
-//   };
-
-// ByString
-//
-// A sub-string delimiter. If `StrSplit()` is passed a string in place of a
-// `Delimiter` object, the string will be implicitly converted into a
-// `ByString` delimiter.
-//
-// Example:
-//
-//   // Because a string literal is converted to an `absl::ByString`,
-//   // the following two splits are equivalent.
-//
-//   std::vector<std::string> v1 = absl::StrSplit("a, b, c", ", ");
-//
-//   using absl::ByString;
-//   std::vector<std::string> v2 = absl::StrSplit("a, b, c",
-//                                                ByString(", "));
-//   // v[0] == "a", v[1] == "b", v[2] == "c"
-class ByString {
- public:
-  explicit ByString(absl::string_view sp);
-  absl::string_view Find(absl::string_view text, size_t pos) const;
-
- private:
-  const std::string delimiter_;
-};
-
-// ByAsciiWhitespace
-//
-// A sub-string delimiter that splits by ASCII whitespace
-// (space, tab, vertical tab, formfeed, linefeed, or carriage return).
-// Note: you probably want to use absl::SkipEmpty() as well!
-//
-// This class is equivalent to ByAnyChar with ASCII whitespace chars.
-//
-// Example:
-//
-//   std::vector<std::string> v = absl::StrSplit(
-//       "a b\tc\n  d  \n", absl::ByAsciiWhitespace(), absl::SkipEmpty());
-//   // v[0] == "a", v[1] == "b", v[2] == "c", v[3] == "d"
-class ByAsciiWhitespace {
- public:
-  absl::string_view Find(absl::string_view text, size_t pos) const;
-};
-
-// ByChar
-//
-// A single character delimiter. `ByChar` is functionally equivalent to a
-// 1-char string within a `ByString` delimiter, but slightly more efficient.
-//
-// Example:
-//
-//   // Because a char literal is converted to a absl::ByChar,
-//   // the following two splits are equivalent.
-//   std::vector<std::string> v1 = absl::StrSplit("a,b,c", ',');
-//   using absl::ByChar;
-//   std::vector<std::string> v2 = absl::StrSplit("a,b,c", ByChar(','));
-//   // v[0] == "a", v[1] == "b", v[2] == "c"
-//
-// `ByChar` is also the default delimiter if a single character is given
-// as the delimiter to `StrSplit()`. For example, the following calls are
-// equivalent:
-//
-//   std::vector<std::string> v = absl::StrSplit("a-b", '-');
-//
-//   using absl::ByChar;
-//   std::vector<std::string> v = absl::StrSplit("a-b", ByChar('-'));
-//
-class ByChar {
- public:
-  explicit ByChar(char c) : c_(c) {}
-  absl::string_view Find(absl::string_view text, size_t pos) const;
-
- private:
-  char c_;
-};
-
-// ByAnyChar
-//
-// A delimiter that will match any of the given byte-sized characters within
-// its provided string.
-//
-// Note: this delimiter works with single-byte string data, but does not work
-// with variable-width encodings, such as UTF-8.
-//
-// Example:
-//
-//   using absl::ByAnyChar;
-//   std::vector<std::string> v = absl::StrSplit("a,b=c", ByAnyChar(",="));
-//   // v[0] == "a", v[1] == "b", v[2] == "c"
-//
-// If `ByAnyChar` is given the empty string, it behaves exactly like
-// `ByString` and matches each individual character in the input string.
-//
-class ByAnyChar {
- public:
-  explicit ByAnyChar(absl::string_view sp);
-  absl::string_view Find(absl::string_view text, size_t pos) const;
-
- private:
-  const std::string delimiters_;
-};
-
-// ByLength
-//
-// A delimiter for splitting into equal-length strings. The length argument to
-// the constructor must be greater than 0.
-//
-// Note: this delimiter works with single-byte string data, but does not work
-// with variable-width encodings, such as UTF-8.
-//
-// Example:
-//
-//   using absl::ByLength;
-//   std::vector<std::string> v = absl::StrSplit("123456789", ByLength(3));
-
-//   // v[0] == "123", v[1] == "456", v[2] == "789"
-//
-// Note that the string does not have to be a multiple of the fixed split
-// length. In such a case, the last substring will be shorter.
-//
-//   using absl::ByLength;
-//   std::vector<std::string> v = absl::StrSplit("12345", ByLength(2));
-//
-//   // v[0] == "12", v[1] == "34", v[2] == "5"
-class ByLength {
- public:
-  explicit ByLength(ptrdiff_t length);
-  absl::string_view Find(absl::string_view text, size_t pos) const;
-
- private:
-  const ptrdiff_t length_;
-};
-
-namespace strings_internal {
-
-// A traits-like metafunction for selecting the default Delimiter object type
-// for a particular Delimiter type. The base case simply exposes type Delimiter
-// itself as the delimiter's Type. However, there are specializations for
-// string-like objects that map them to the ByString delimiter object.
-// This allows functions like absl::StrSplit() and absl::MaxSplits() to accept
-// string-like objects (e.g., ',') as delimiter arguments but they will be
-// treated as if a ByString delimiter was given.
-template <typename Delimiter>
-struct SelectDelimiter {
-  using type = Delimiter;
-};
-
-template <>
-struct SelectDelimiter<char> {
-  using type = ByChar;
-};
-template <>
-struct SelectDelimiter<char*> {
-  using type = ByString;
-};
-template <>
-struct SelectDelimiter<const char*> {
-  using type = ByString;
-};
-template <>
-struct SelectDelimiter<absl::string_view> {
-  using type = ByString;
-};
-template <>
-struct SelectDelimiter<std::string> {
-  using type = ByString;
-};
-
-// Wraps another delimiter and sets a max number of matches for that delimiter.
-template <typename Delimiter>
-class MaxSplitsImpl {
- public:
-  MaxSplitsImpl(Delimiter delimiter, int limit)
-      : delimiter_(delimiter), limit_(limit), count_(0) {}
-  absl::string_view Find(absl::string_view text, size_t pos) {
-    if (count_++ == limit_) {
-      return absl::string_view(text.data() + text.size(),
-                               0);  // No more matches.
-    }
-    return delimiter_.Find(text, pos);
-  }
-
- private:
-  Delimiter delimiter_;
-  const int limit_;
-  int count_;
-};
-
-}  // namespace strings_internal
-
-// MaxSplits()
-//
-// A delimiter that limits the number of matches which can occur to the passed
-// `limit`. The last element in the returned collection will contain all
-// remaining unsplit pieces, which may contain instances of the delimiter.
-// The collection will contain at most `limit` + 1 elements.
-// Example:
-//
-//   using absl::MaxSplits;
-//   std::vector<std::string> v = absl::StrSplit("a,b,c", MaxSplits(',', 1));
-//
-//   // v[0] == "a", v[1] == "b,c"
-template <typename Delimiter>
-inline strings_internal::MaxSplitsImpl<
-    typename strings_internal::SelectDelimiter<Delimiter>::type>
-MaxSplits(Delimiter delimiter, int limit) {
-  typedef
-      typename strings_internal::SelectDelimiter<Delimiter>::type DelimiterType;
-  return strings_internal::MaxSplitsImpl<DelimiterType>(
-      DelimiterType(delimiter), limit);
-}
-
-//------------------------------------------------------------------------------
-// Predicates
-//------------------------------------------------------------------------------
-//
-// Predicates filter the results of a `StrSplit()` by determining whether or not
-// a resultant element is included in the result set. A predicate may be passed
-// as an optional third argument to the `StrSplit()` function.
-//
-// Predicates are unary functions (or functors) that take a single
-// `absl::string_view` argument and return a bool indicating whether the
-// argument should be included (`true`) or excluded (`false`).
-//
-// Predicates are useful when filtering out empty substrings. By default, empty
-// substrings may be returned by `StrSplit()`, which is similar to the way split
-// functions work in other programming languages.
-
-// AllowEmpty()
-//
-// Always returns `true`, indicating that all strings--including empty
-// strings--should be included in the split output. This predicate is not
-// strictly needed because this is the default behavior of `StrSplit()`;
-// however, it might be useful at some call sites to make the intent explicit.
-//
-// Example:
-//
-//  std::vector<std::string> v = absl::StrSplit(" a , ,,b,", ',', AllowEmpty());
-//
-//  // v[0] == " a ", v[1] == " ", v[2] == "", v[3] = "b", v[4] == ""
-struct AllowEmpty {
-  bool operator()(absl::string_view) const { return true; }
-};
-
-// SkipEmpty()
-//
-// Returns `false` if the given `absl::string_view` is empty, indicating that
-// `StrSplit()` should omit the empty string.
-//
-// Example:
-//
-//   std::vector<std::string> v = absl::StrSplit(",a,,b,", ',', SkipEmpty());
-//
-//   // v[0] == "a", v[1] == "b"
-//
-// Note: `SkipEmpty()` does not consider a string containing only whitespace
-// to be empty. To skip such whitespace as well, use the `SkipWhitespace()`
-// predicate.
-struct SkipEmpty {
-  bool operator()(absl::string_view sp) const { return !sp.empty(); }
-};
-
-// SkipWhitespace()
-//
-// Returns `false` if the given `absl::string_view` is empty *or* contains only
-// whitespace, indicating that `StrSplit()` should omit the string.
-//
-// Example:
-//
-//   std::vector<std::string> v = absl::StrSplit(" a , ,,b,",
-//                                               ',', SkipWhitespace());
-//   // v[0] == " a ", v[1] == "b"
-//
-//   // SkipEmpty() would return whitespace elements
-//   std::vector<std::string> v = absl::StrSplit(" a , ,,b,", ',', SkipEmpty());
-//   // v[0] == " a ", v[1] == " ", v[2] == "b"
-struct SkipWhitespace {
-  bool operator()(absl::string_view sp) const {
-    sp = absl::StripAsciiWhitespace(sp);
-    return !sp.empty();
-  }
-};
-
-template <typename T>
-using EnableSplitIfString =
-    typename std::enable_if<std::is_same<T, std::string>::value ||
-                            std::is_same<T, const std::string>::value,
-                            int>::type;
-
-//------------------------------------------------------------------------------
-//                                  StrSplit()
-//------------------------------------------------------------------------------
-
-// StrSplit()
-//
-// Splits a given string based on the provided `Delimiter` object, returning the
-// elements within the type specified by the caller. Optionally, you may pass a
-// `Predicate` to `StrSplit()` indicating whether to include or exclude the
-// resulting element within the final result set. (See the overviews for
-// Delimiters and Predicates above.)
-//
-// Example:
-//
-//   std::vector<std::string> v = absl::StrSplit("a,b,c,d", ',');
-//   // v[0] == "a", v[1] == "b", v[2] == "c", v[3] == "d"
-//
-// You can also provide an explicit `Delimiter` object:
-//
-// Example:
-//
-//   using absl::ByAnyChar;
-//   std::vector<std::string> v = absl::StrSplit("a,b=c", ByAnyChar(",="));
-//   // v[0] == "a", v[1] == "b", v[2] == "c"
-//
-// See above for more information on delimiters.
-//
-// By default, empty strings are included in the result set. You can optionally
-// include a third `Predicate` argument to apply a test for whether the
-// resultant element should be included in the result set:
-//
-// Example:
-//
-//   std::vector<std::string> v = absl::StrSplit(" a , ,,b,",
-//                                               ',', SkipWhitespace());
-//   // v[0] == " a ", v[1] == "b"
-//
-// See above for more information on predicates.
-//
-//------------------------------------------------------------------------------
-// StrSplit() Return Types
-//------------------------------------------------------------------------------
-//
-// The `StrSplit()` function adapts the returned collection to the collection
-// specified by the caller (e.g. `std::vector` above). The returned collections
-// may contain `std::string`, `absl::string_view` (in which case the original
-// string being split must ensure that it outlives the collection), or any
-// object that can be explicitly created from an `absl::string_view`. This
-// behavior works for:
-//
-// 1) All standard STL containers including `std::vector`, `std::list`,
-//    `std::deque`, `std::set`,`std::multiset`, 'std::map`, and `std::multimap`
-// 2) `std::pair` (which is not actually a container). See below.
-//
-// Example:
-//
-//   // The results are returned as `absl::string_view` objects. Note that we
-//   // have to ensure that the input string outlives any results.
-//   std::vector<absl::string_view> v = absl::StrSplit("a,b,c", ',');
-//
-//   // Stores results in a std::set<std::string>, which also performs
-//   // de-duplication and orders the elements in ascending order.
-//   std::set<std::string> a = absl::StrSplit("b,a,c,a,b", ',');
-//   // a[0] == "a", a[1] == "b", a[2] == "c"
-//
-//   // `StrSplit()` can be used within a range-based for loop, in which case
-//   // each element will be of type `absl::string_view`.
-//   std::vector<std::string> v;
-//   for (const auto sv : absl::StrSplit("a,b,c", ',')) {
-//     if (sv != "b") v.emplace_back(sv);
-//   }
-//   // v[0] == "a", v[1] == "c"
-//
-//   // Stores results in a map. The map implementation assumes that the input
-//   // is provided as a series of key/value pairs. For example, the 0th element
-//   // resulting from the split will be stored as a key to the 1st element. If
-//   // an odd number of elements are resolved, the last element is paired with
-//   // a default-constructed value (e.g., empty string).
-//   std::map<std::string, std::string> m = absl::StrSplit("a,b,c", ',');
-//   // m["a"] == "b", m["c"] == ""     // last component value equals ""
-//
-// Splitting to `std::pair` is an interesting case because it can hold only two
-// elements and is not a collection type. When splitting to a `std::pair` the
-// first two split strings become the `std::pair` `.first` and `.second`
-// members, respectively. The remaining split substrings are discarded. If there
-// are less than two split substrings, the empty string is used for the
-// corresponding `std::pair` member.
-//
-// Example:
-//
-//   // Stores first two split strings as the members in a std::pair.
-//   std::pair<std::string, std::string> p = absl::StrSplit("a,b,c", ',');
-//   // p.first == "a", p.second == "b"       // "c" is omitted.
-//
-// The `StrSplit()` function can be used multiple times to perform more
-// complicated splitting logic, such as intelligently parsing key-value pairs.
-//
-// Example:
-//
-//   // The input string "a=b=c,d=e,f=,g" becomes
-//   // { "a" => "b=c", "d" => "e", "f" => "", "g" => "" }
-//   std::map<std::string, std::string> m;
-//   for (absl::string_view sp : absl::StrSplit("a=b=c,d=e,f=,g", ',')) {
-//     m.insert(absl::StrSplit(sp, absl::MaxSplits('=', 1)));
-//   }
-//   EXPECT_EQ("b=c", m.find("a")->second);
-//   EXPECT_EQ("e", m.find("d")->second);
-//   EXPECT_EQ("", m.find("f")->second);
-//   EXPECT_EQ("", m.find("g")->second);
-//
-// WARNING: Due to a legacy bug that is maintained for backward compatibility,
-// splitting the following empty string_views produces different results:
-//
-//   absl::StrSplit(absl::string_view(""), '-');  // {""}
-//   absl::StrSplit(absl::string_view(), '-');    // {}, but should be {""}
-//
-// Try not to depend on this distinction because the bug may one day be fixed.
-template <typename Delimiter>
-strings_internal::Splitter<
-    typename strings_internal::SelectDelimiter<Delimiter>::type, AllowEmpty,
-    absl::string_view>
-StrSplit(strings_internal::ConvertibleToStringView text, Delimiter d) {
-  using DelimiterType =
-      typename strings_internal::SelectDelimiter<Delimiter>::type;
-  return strings_internal::Splitter<DelimiterType, AllowEmpty,
-                                    absl::string_view>(
-      text.value(), DelimiterType(d), AllowEmpty());
-}
-
-template <typename Delimiter, typename StringType,
-          EnableSplitIfString<StringType> = 0>
-strings_internal::Splitter<
-    typename strings_internal::SelectDelimiter<Delimiter>::type, AllowEmpty,
-    std::string>
-StrSplit(StringType&& text, Delimiter d) {
-  using DelimiterType =
-      typename strings_internal::SelectDelimiter<Delimiter>::type;
-  return strings_internal::Splitter<DelimiterType, AllowEmpty, std::string>(
-      std::move(text), DelimiterType(d), AllowEmpty());
-}
-
-template <typename Delimiter, typename Predicate>
-strings_internal::Splitter<
-    typename strings_internal::SelectDelimiter<Delimiter>::type, Predicate,
-    absl::string_view>
-StrSplit(strings_internal::ConvertibleToStringView text, Delimiter d,
-         Predicate p) {
-  using DelimiterType =
-      typename strings_internal::SelectDelimiter<Delimiter>::type;
-  return strings_internal::Splitter<DelimiterType, Predicate,
-                                    absl::string_view>(
-      text.value(), DelimiterType(d), std::move(p));
-}
-
-template <typename Delimiter, typename Predicate, typename StringType,
-          EnableSplitIfString<StringType> = 0>
-strings_internal::Splitter<
-    typename strings_internal::SelectDelimiter<Delimiter>::type, Predicate,
-    std::string>
-StrSplit(StringType&& text, Delimiter d, Predicate p) {
-  using DelimiterType =
-      typename strings_internal::SelectDelimiter<Delimiter>::type;
-  return strings_internal::Splitter<DelimiterType, Predicate, std::string>(
-      std::move(text), DelimiterType(d), std::move(p));
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_STR_SPLIT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/string_view.h b/projects/llm_framework/include/abseil-cpp/absl/strings/string_view.h
deleted file mode 100644
index b393c6fc..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/string_view.h
+++ /dev/null
@@ -1,769 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: string_view.h
-// -----------------------------------------------------------------------------
-//
-// This file contains the definition of the `absl::string_view` class. A
-// `string_view` points to a contiguous span of characters, often part or all of
-// another `std::string`, double-quoted string literal, character array, or even
-// another `string_view`.
-//
-// This `absl::string_view` abstraction is designed to be a drop-in
-// replacement for the C++17 `std::string_view` abstraction.
-#ifndef ABSL_STRINGS_STRING_VIEW_H_
-#define ABSL_STRINGS_STRING_VIEW_H_
-
-#include <algorithm>
-#include <cassert>
-#include <cstddef>
-#include <cstring>
-#include <iosfwd>
-#include <iterator>
-#include <limits>
-#include <string>
-
-#include "absl/base/attributes.h"
-#include "absl/base/nullability.h"
-#include "absl/base/config.h"
-#include "absl/base/internal/throw_delegate.h"
-#include "absl/base/macros.h"
-#include "absl/base/optimization.h"
-#include "absl/base/port.h"
-
-#ifdef ABSL_USES_STD_STRING_VIEW
-
-#include <string_view>  // IWYU pragma: export
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-using string_view = std::string_view;
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#else  // ABSL_USES_STD_STRING_VIEW
-
-#if ABSL_HAVE_BUILTIN(__builtin_memcmp) ||        \
-    (defined(__GNUC__) && !defined(__clang__)) || \
-    (defined(_MSC_VER) && _MSC_VER >= 1928)
-#define ABSL_INTERNAL_STRING_VIEW_MEMCMP __builtin_memcmp
-#else  // ABSL_HAVE_BUILTIN(__builtin_memcmp)
-#define ABSL_INTERNAL_STRING_VIEW_MEMCMP memcmp
-#endif  // ABSL_HAVE_BUILTIN(__builtin_memcmp)
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// absl::string_view
-//
-// A `string_view` provides a lightweight view into the string data provided by
-// a `std::string`, double-quoted string literal, character array, or even
-// another `string_view`. A `string_view` does *not* own the string to which it
-// points, and that data cannot be modified through the view.
-//
-// You can use `string_view` as a function or method parameter anywhere a
-// parameter can receive a double-quoted string literal, `const char*`,
-// `std::string`, or another `absl::string_view` argument with no need to copy
-// the string data. Systematic use of `string_view` within function arguments
-// reduces data copies and `strlen()` calls.
-//
-// Because of its small size, prefer passing `string_view` by value:
-//
-//   void MyFunction(absl::string_view arg);
-//
-// If circumstances require, you may also pass one by const reference:
-//
-//   void MyFunction(const absl::string_view& arg);  // not preferred
-//
-// Passing by value generates slightly smaller code for many architectures.
-//
-// In either case, the source data of the `string_view` must outlive the
-// `string_view` itself.
-//
-// A `string_view` is also suitable for local variables if you know that the
-// lifetime of the underlying object is longer than the lifetime of your
-// `string_view` variable. However, beware of binding a `string_view` to a
-// temporary value:
-//
-//   // BAD use of string_view: lifetime problem
-//   absl::string_view sv = obj.ReturnAString();
-//
-//   // GOOD use of string_view: str outlives sv
-//   std::string str = obj.ReturnAString();
-//   absl::string_view sv = str;
-//
-// Due to lifetime issues, a `string_view` is sometimes a poor choice for a
-// return value and usually a poor choice for a data member. If you do use a
-// `string_view` this way, it is your responsibility to ensure that the object
-// pointed to by the `string_view` outlives the `string_view`.
-//
-// A `string_view` may represent a whole string or just part of a string. For
-// example, when splitting a string, `std::vector<absl::string_view>` is a
-// natural data type for the output.
-//
-// For another example, a Cord is a non-contiguous, potentially very
-// long string-like object.  The Cord class has an interface that iteratively
-// provides string_view objects that point to the successive pieces of a Cord
-// object.
-//
-// When constructed from a source which is NUL-terminated, the `string_view`
-// itself will not include the NUL-terminator unless a specific size (including
-// the NUL) is passed to the constructor. As a result, common idioms that work
-// on NUL-terminated strings do not work on `string_view` objects. If you write
-// code that scans a `string_view`, you must check its length rather than test
-// for nul, for example. Note, however, that nuls may still be embedded within
-// a `string_view` explicitly.
-//
-// You may create a null `string_view` in two ways:
-//
-//   absl::string_view sv;
-//   absl::string_view sv(nullptr, 0);
-//
-// For the above, `sv.data() == nullptr`, `sv.length() == 0`, and
-// `sv.empty() == true`. Also, if you create a `string_view` with a non-null
-// pointer then `sv.data() != nullptr`. Thus, you can use `string_view()` to
-// signal an undefined value that is different from other `string_view` values
-// in a similar fashion to how `const char* p1 = nullptr;` is different from
-// `const char* p2 = "";`. However, in practice, it is not recommended to rely
-// on this behavior.
-//
-// Be careful not to confuse a null `string_view` with an empty one. A null
-// `string_view` is an empty `string_view`, but some empty `string_view`s are
-// not null. Prefer checking for emptiness over checking for null.
-//
-// There are many ways to create an empty string_view:
-//
-//   const char* nullcp = nullptr;
-//   // string_view.size() will return 0 in all cases.
-//   absl::string_view();
-//   absl::string_view(nullcp, 0);
-//   absl::string_view("");
-//   absl::string_view("", 0);
-//   absl::string_view("abcdef", 0);
-//   absl::string_view("abcdef" + 6, 0);
-//
-// All empty `string_view` objects whether null or not, are equal:
-//
-//   absl::string_view() == absl::string_view("", 0)
-//   absl::string_view(nullptr, 0) == absl::string_view("abcdef"+6, 0)
-class string_view {
- public:
-  using traits_type = std::char_traits<char>;
-  using value_type = char;
-  using pointer = absl::Nullable<char*>;
-  using const_pointer = absl::Nullable<const char*>;
-  using reference = char&;
-  using const_reference = const char&;
-  using const_iterator = absl::Nullable<const char*>;
-  using iterator = const_iterator;
-  using const_reverse_iterator = std::reverse_iterator<const_iterator>;
-  using reverse_iterator = const_reverse_iterator;
-  using size_type = size_t;
-  using difference_type = std::ptrdiff_t;
-
-  static constexpr size_type npos = static_cast<size_type>(-1);
-
-  // Null `string_view` constructor
-  constexpr string_view() noexcept : ptr_(nullptr), length_(0) {}
-
-  // Implicit constructors
-
-  template <typename Allocator>
-  string_view(  // NOLINT(runtime/explicit)
-      const std::basic_string<char, std::char_traits<char>, Allocator>& str
-          ABSL_ATTRIBUTE_LIFETIME_BOUND) noexcept
-      // This is implemented in terms of `string_view(p, n)` so `str.size()`
-      // doesn't need to be reevaluated after `ptr_` is set.
-      // The length check is also skipped since it is unnecessary and causes
-      // code bloat.
-      : string_view(str.data(), str.size(), SkipCheckLengthTag{}) {}
-
-  // Implicit constructor of a `string_view` from NUL-terminated `str`. When
-  // accepting possibly null strings, use `absl::NullSafeStringView(str)`
-  // instead (see below).
-  // The length check is skipped since it is unnecessary and causes code bloat.
-  constexpr string_view(  // NOLINT(runtime/explicit)
-      absl::Nonnull<const char*> str)
-      : ptr_(str), length_(str ? StrlenInternal(str) : 0) {}
-
-  // Implicit constructor of a `string_view` from a `const char*` and length.
-  constexpr string_view(absl::Nullable<const char*> data, size_type len)
-      : ptr_(data), length_(CheckLengthInternal(len)) {}
-
-  // NOTE: Harmlessly omitted to work around gdb bug.
-  //   constexpr string_view(const string_view&) noexcept = default;
-  //   string_view& operator=(const string_view&) noexcept = default;
-
-  // Iterators
-
-  // string_view::begin()
-  //
-  // Returns an iterator pointing to the first character at the beginning of the
-  // `string_view`, or `end()` if the `string_view` is empty.
-  constexpr const_iterator begin() const noexcept { return ptr_; }
-
-  // string_view::end()
-  //
-  // Returns an iterator pointing just beyond the last character at the end of
-  // the `string_view`. This iterator acts as a placeholder; attempting to
-  // access it results in undefined behavior.
-  constexpr const_iterator end() const noexcept { return ptr_ + length_; }
-
-  // string_view::cbegin()
-  //
-  // Returns a const iterator pointing to the first character at the beginning
-  // of the `string_view`, or `end()` if the `string_view` is empty.
-  constexpr const_iterator cbegin() const noexcept { return begin(); }
-
-  // string_view::cend()
-  //
-  // Returns a const iterator pointing just beyond the last character at the end
-  // of the `string_view`. This pointer acts as a placeholder; attempting to
-  // access its element results in undefined behavior.
-  constexpr const_iterator cend() const noexcept { return end(); }
-
-  // string_view::rbegin()
-  //
-  // Returns a reverse iterator pointing to the last character at the end of the
-  // `string_view`, or `rend()` if the `string_view` is empty.
-  const_reverse_iterator rbegin() const noexcept {
-    return const_reverse_iterator(end());
-  }
-
-  // string_view::rend()
-  //
-  // Returns a reverse iterator pointing just before the first character at the
-  // beginning of the `string_view`. This pointer acts as a placeholder;
-  // attempting to access its element results in undefined behavior.
-  const_reverse_iterator rend() const noexcept {
-    return const_reverse_iterator(begin());
-  }
-
-  // string_view::crbegin()
-  //
-  // Returns a const reverse iterator pointing to the last character at the end
-  // of the `string_view`, or `crend()` if the `string_view` is empty.
-  const_reverse_iterator crbegin() const noexcept { return rbegin(); }
-
-  // string_view::crend()
-  //
-  // Returns a const reverse iterator pointing just before the first character
-  // at the beginning of the `string_view`. This pointer acts as a placeholder;
-  // attempting to access its element results in undefined behavior.
-  const_reverse_iterator crend() const noexcept { return rend(); }
-
-  // Capacity Utilities
-
-  // string_view::size()
-  //
-  // Returns the number of characters in the `string_view`.
-  constexpr size_type size() const noexcept { return length_; }
-
-  // string_view::length()
-  //
-  // Returns the number of characters in the `string_view`. Alias for `size()`.
-  constexpr size_type length() const noexcept { return size(); }
-
-  // string_view::max_size()
-  //
-  // Returns the maximum number of characters the `string_view` can hold.
-  constexpr size_type max_size() const noexcept { return kMaxSize; }
-
-  // string_view::empty()
-  //
-  // Checks if the `string_view` is empty (refers to no characters).
-  constexpr bool empty() const noexcept { return length_ == 0; }
-
-  // string_view::operator[]
-  //
-  // Returns the ith element of the `string_view` using the array operator.
-  // Note that this operator does not perform any bounds checking.
-  constexpr const_reference operator[](size_type i) const {
-    return ABSL_HARDENING_ASSERT(i < size()), ptr_[i];
-  }
-
-  // string_view::at()
-  //
-  // Returns the ith element of the `string_view`. Bounds checking is performed,
-  // and an exception of type `std::out_of_range` will be thrown on invalid
-  // access.
-  constexpr const_reference at(size_type i) const {
-    return ABSL_PREDICT_TRUE(i < size())
-               ? ptr_[i]
-               : ((void)base_internal::ThrowStdOutOfRange(
-                      "absl::string_view::at"),
-                  ptr_[i]);
-  }
-
-  // string_view::front()
-  //
-  // Returns the first element of a `string_view`.
-  constexpr const_reference front() const {
-    return ABSL_HARDENING_ASSERT(!empty()), ptr_[0];
-  }
-
-  // string_view::back()
-  //
-  // Returns the last element of a `string_view`.
-  constexpr const_reference back() const {
-    return ABSL_HARDENING_ASSERT(!empty()), ptr_[size() - 1];
-  }
-
-  // string_view::data()
-  //
-  // Returns a pointer to the underlying character array (which is of course
-  // stored elsewhere). Note that `string_view::data()` may contain embedded nul
-  // characters, but the returned buffer may or may not be NUL-terminated;
-  // therefore, do not pass `data()` to a routine that expects a NUL-terminated
-  // string.
-  constexpr const_pointer data() const noexcept { return ptr_; }
-
-  // Modifiers
-
-  // string_view::remove_prefix()
-  //
-  // Removes the first `n` characters from the `string_view`. Note that the
-  // underlying string is not changed, only the view.
-  constexpr void remove_prefix(size_type n) {
-    ABSL_HARDENING_ASSERT(n <= length_);
-    ptr_ += n;
-    length_ -= n;
-  }
-
-  // string_view::remove_suffix()
-  //
-  // Removes the last `n` characters from the `string_view`. Note that the
-  // underlying string is not changed, only the view.
-  constexpr void remove_suffix(size_type n) {
-    ABSL_HARDENING_ASSERT(n <= length_);
-    length_ -= n;
-  }
-
-  // string_view::swap()
-  //
-  // Swaps this `string_view` with another `string_view`.
-  constexpr void swap(string_view& s) noexcept {
-    auto t = *this;
-    *this = s;
-    s = t;
-  }
-
-  // Explicit conversion operators
-
-  // Converts to `std::basic_string`.
-  template <typename A>
-  explicit operator std::basic_string<char, traits_type, A>() const {
-    if (!data()) return {};
-    return std::basic_string<char, traits_type, A>(data(), size());
-  }
-
-  // string_view::copy()
-  //
-  // Copies the contents of the `string_view` at offset `pos` and length `n`
-  // into `buf`.
-  size_type copy(char* buf, size_type n, size_type pos = 0) const {
-    if (ABSL_PREDICT_FALSE(pos > length_)) {
-      base_internal::ThrowStdOutOfRange("absl::string_view::copy");
-    }
-    size_type rlen = (std::min)(length_ - pos, n);
-    if (rlen > 0) {
-      const char* start = ptr_ + pos;
-      traits_type::copy(buf, start, rlen);
-    }
-    return rlen;
-  }
-
-  // string_view::substr()
-  //
-  // Returns a "substring" of the `string_view` (at offset `pos` and length
-  // `n`) as another string_view. This function throws `std::out_of_bounds` if
-  // `pos > size`.
-  // Use absl::ClippedSubstr if you need a truncating substr operation.
-  constexpr string_view substr(size_type pos = 0, size_type n = npos) const {
-    return ABSL_PREDICT_FALSE(pos > length_)
-               ? (base_internal::ThrowStdOutOfRange(
-                      "absl::string_view::substr"),
-                  string_view())
-               : string_view(ptr_ + pos, Min(n, length_ - pos));
-  }
-
-  // string_view::compare()
-  //
-  // Performs a lexicographical comparison between this `string_view` and
-  // another `string_view` `x`, returning a negative value if `*this` is less
-  // than `x`, 0 if `*this` is equal to `x`, and a positive value if `*this`
-  // is greater than `x`.
-  constexpr int compare(string_view x) const noexcept {
-    return CompareImpl(length_, x.length_,
-                       Min(length_, x.length_) == 0
-                           ? 0
-                           : ABSL_INTERNAL_STRING_VIEW_MEMCMP(
-                                 ptr_, x.ptr_, Min(length_, x.length_)));
-  }
-
-  // Overload of `string_view::compare()` for comparing a substring of the
-  // 'string_view` and another `absl::string_view`.
-  constexpr int compare(size_type pos1, size_type count1, string_view v) const {
-    return substr(pos1, count1).compare(v);
-  }
-
-  // Overload of `string_view::compare()` for comparing a substring of the
-  // `string_view` and a substring of another `absl::string_view`.
-  constexpr int compare(size_type pos1, size_type count1, string_view v,
-                        size_type pos2, size_type count2) const {
-    return substr(pos1, count1).compare(v.substr(pos2, count2));
-  }
-
-  // Overload of `string_view::compare()` for comparing a `string_view` and a
-  // a different C-style string `s`.
-  constexpr int compare(absl::Nonnull<const char*> s) const {
-    return compare(string_view(s));
-  }
-
-  // Overload of `string_view::compare()` for comparing a substring of the
-  // `string_view` and a different string C-style string `s`.
-  constexpr int compare(size_type pos1, size_type count1,
-                        absl::Nonnull<const char*> s) const {
-    return substr(pos1, count1).compare(string_view(s));
-  }
-
-  // Overload of `string_view::compare()` for comparing a substring of the
-  // `string_view` and a substring of a different C-style string `s`.
-  constexpr int compare(size_type pos1, size_type count1,
-                        absl::Nonnull<const char*> s, size_type count2) const {
-    return substr(pos1, count1).compare(string_view(s, count2));
-  }
-
-  // Find Utilities
-
-  // string_view::find()
-  //
-  // Finds the first occurrence of the substring `s` within the `string_view`,
-  // returning the position of the first character's match, or `npos` if no
-  // match was found.
-  size_type find(string_view s, size_type pos = 0) const noexcept;
-
-  // Overload of `string_view::find()` for finding the given character `c`
-  // within the `string_view`.
-  size_type find(char c, size_type pos = 0) const noexcept;
-
-  // Overload of `string_view::find()` for finding a substring of a different
-  // C-style string `s` within the `string_view`.
-  size_type find(absl::Nonnull<const char*> s, size_type pos,
-                 size_type count) const {
-    return find(string_view(s, count), pos);
-  }
-
-  // Overload of `string_view::find()` for finding a different C-style string
-  // `s` within the `string_view`.
-  size_type find(absl::Nonnull<const char *> s, size_type pos = 0) const {
-    return find(string_view(s), pos);
-  }
-
-  // string_view::rfind()
-  //
-  // Finds the last occurrence of a substring `s` within the `string_view`,
-  // returning the position of the first character's match, or `npos` if no
-  // match was found.
-  size_type rfind(string_view s, size_type pos = npos) const noexcept;
-
-  // Overload of `string_view::rfind()` for finding the last given character `c`
-  // within the `string_view`.
-  size_type rfind(char c, size_type pos = npos) const noexcept;
-
-  // Overload of `string_view::rfind()` for finding a substring of a different
-  // C-style string `s` within the `string_view`.
-  size_type rfind(absl::Nonnull<const char*> s, size_type pos,
-                  size_type count) const {
-    return rfind(string_view(s, count), pos);
-  }
-
-  // Overload of `string_view::rfind()` for finding a different C-style string
-  // `s` within the `string_view`.
-  size_type rfind(absl::Nonnull<const char*> s, size_type pos = npos) const {
-    return rfind(string_view(s), pos);
-  }
-
-  // string_view::find_first_of()
-  //
-  // Finds the first occurrence of any of the characters in `s` within the
-  // `string_view`, returning the start position of the match, or `npos` if no
-  // match was found.
-  size_type find_first_of(string_view s, size_type pos = 0) const noexcept;
-
-  // Overload of `string_view::find_first_of()` for finding a character `c`
-  // within the `string_view`.
-  size_type find_first_of(char c, size_type pos = 0) const noexcept {
-    return find(c, pos);
-  }
-
-  // Overload of `string_view::find_first_of()` for finding a substring of a
-  // different C-style string `s` within the `string_view`.
-  size_type find_first_of(absl::Nonnull<const char*> s, size_type pos,
-                          size_type count) const {
-    return find_first_of(string_view(s, count), pos);
-  }
-
-  // Overload of `string_view::find_first_of()` for finding a different C-style
-  // string `s` within the `string_view`.
-  size_type find_first_of(absl::Nonnull<const char*> s,
-                          size_type pos = 0) const {
-    return find_first_of(string_view(s), pos);
-  }
-
-  // string_view::find_last_of()
-  //
-  // Finds the last occurrence of any of the characters in `s` within the
-  // `string_view`, returning the start position of the match, or `npos` if no
-  // match was found.
-  size_type find_last_of(string_view s, size_type pos = npos) const noexcept;
-
-  // Overload of `string_view::find_last_of()` for finding a character `c`
-  // within the `string_view`.
-  size_type find_last_of(char c, size_type pos = npos) const noexcept {
-    return rfind(c, pos);
-  }
-
-  // Overload of `string_view::find_last_of()` for finding a substring of a
-  // different C-style string `s` within the `string_view`.
-  size_type find_last_of(absl::Nonnull<const char*> s, size_type pos,
-                         size_type count) const {
-    return find_last_of(string_view(s, count), pos);
-  }
-
-  // Overload of `string_view::find_last_of()` for finding a different C-style
-  // string `s` within the `string_view`.
-  size_type find_last_of(absl::Nonnull<const char*> s,
-                         size_type pos = npos) const {
-    return find_last_of(string_view(s), pos);
-  }
-
-  // string_view::find_first_not_of()
-  //
-  // Finds the first occurrence of any of the characters not in `s` within the
-  // `string_view`, returning the start position of the first non-match, or
-  // `npos` if no non-match was found.
-  size_type find_first_not_of(string_view s, size_type pos = 0) const noexcept;
-
-  // Overload of `string_view::find_first_not_of()` for finding a character
-  // that is not `c` within the `string_view`.
-  size_type find_first_not_of(char c, size_type pos = 0) const noexcept;
-
-  // Overload of `string_view::find_first_not_of()` for finding a substring of a
-  // different C-style string `s` within the `string_view`.
-  size_type find_first_not_of(absl::Nonnull<const char*> s, size_type pos,
-                              size_type count) const {
-    return find_first_not_of(string_view(s, count), pos);
-  }
-
-  // Overload of `string_view::find_first_not_of()` for finding a different
-  // C-style string `s` within the `string_view`.
-  size_type find_first_not_of(absl::Nonnull<const char*> s,
-                              size_type pos = 0) const {
-    return find_first_not_of(string_view(s), pos);
-  }
-
-  // string_view::find_last_not_of()
-  //
-  // Finds the last occurrence of any of the characters not in `s` within the
-  // `string_view`, returning the start position of the last non-match, or
-  // `npos` if no non-match was found.
-  size_type find_last_not_of(string_view s,
-                             size_type pos = npos) const noexcept;
-
-  // Overload of `string_view::find_last_not_of()` for finding a character
-  // that is not `c` within the `string_view`.
-  size_type find_last_not_of(char c, size_type pos = npos) const noexcept;
-
-  // Overload of `string_view::find_last_not_of()` for finding a substring of a
-  // different C-style string `s` within the `string_view`.
-  size_type find_last_not_of(absl::Nonnull<const char*> s, size_type pos,
-                             size_type count) const {
-    return find_last_not_of(string_view(s, count), pos);
-  }
-
-  // Overload of `string_view::find_last_not_of()` for finding a different
-  // C-style string `s` within the `string_view`.
-  size_type find_last_not_of(absl::Nonnull<const char*> s,
-                             size_type pos = npos) const {
-    return find_last_not_of(string_view(s), pos);
-  }
-
-#if ABSL_INTERNAL_CPLUSPLUS_LANG >= 202002L
-  // string_view::starts_with()
-  //
-  // Returns true if the `string_view` starts with the prefix `s`.
-  //
-  // This method only exists when targeting at least C++20.
-  // If support for C++ prior to C++20 is required, use `absl::StartsWith()`
-  // from `//absl/strings/match.h` for compatibility.
-  constexpr bool starts_with(string_view s) const noexcept {
-    return s.empty() ||
-           (size() >= s.size() &&
-            ABSL_INTERNAL_STRING_VIEW_MEMCMP(data(), s.data(), s.size()) == 0);
-  }
-
-  // Overload of `string_view::starts_with()` that returns true if `c` is the
-  // first character of the `string_view`.
-  constexpr bool starts_with(char c) const noexcept {
-    return !empty() && front() == c;
-  }
-
-  // Overload of `string_view::starts_with()` that returns true if the
-  // `string_view` starts with the C-style prefix `s`.
-  constexpr bool starts_with(const char* s) const {
-    return starts_with(string_view(s));
-  }
-
-  // string_view::ends_with()
-  //
-  // Returns true if the `string_view` ends with the suffix `s`.
-  //
-  // This method only exists when targeting at least C++20.
-  // If support for C++ prior to C++20 is required, use `absl::EndsWith()`
-  // from `//absl/strings/match.h` for compatibility.
-  constexpr bool ends_with(string_view s) const noexcept {
-    return s.empty() || (size() >= s.size() && ABSL_INTERNAL_STRING_VIEW_MEMCMP(
-                                                   data() + (size() - s.size()),
-                                                   s.data(), s.size()) == 0);
-  }
-
-  // Overload of `string_view::ends_with()` that returns true if `c` is the
-  // last character of the `string_view`.
-  constexpr bool ends_with(char c) const noexcept {
-    return !empty() && back() == c;
-  }
-
-  // Overload of `string_view::ends_with()` that returns true if the
-  // `string_view` ends with the C-style suffix `s`.
-  constexpr bool ends_with(const char* s) const {
-    return ends_with(string_view(s));
-  }
-#endif  // ABSL_INTERNAL_CPLUSPLUS_LANG >= 202002L
-
- private:
-  // The constructor from std::string delegates to this constructor.
-  // See the comment on that constructor for the rationale.
-  struct SkipCheckLengthTag {};
-  string_view(absl::Nullable<const char*> data, size_type len,
-              SkipCheckLengthTag) noexcept
-      : ptr_(data), length_(len) {}
-
-  static constexpr size_type kMaxSize =
-      (std::numeric_limits<difference_type>::max)();
-
-  static constexpr size_type CheckLengthInternal(size_type len) {
-    return ABSL_HARDENING_ASSERT(len <= kMaxSize), len;
-  }
-
-  static constexpr size_type StrlenInternal(absl::Nonnull<const char*> str) {
-#if defined(_MSC_VER) && !defined(__clang__)
-    // MSVC 2017+ can evaluate this at compile-time.
-    const char* begin = str;
-    while (*str != '\0') ++str;
-    return str - begin;
-#elif ABSL_HAVE_BUILTIN(__builtin_strlen) || \
-    (defined(__GNUC__) && !defined(__clang__))
-    // GCC has __builtin_strlen according to
-    // https://gcc.gnu.org/onlinedocs/gcc-4.7.0/gcc/Other-Builtins.html, but
-    // ABSL_HAVE_BUILTIN doesn't detect that, so we use the extra checks above.
-    // __builtin_strlen is constexpr.
-    return __builtin_strlen(str);
-#else
-    return str ? strlen(str) : 0;
-#endif
-  }
-
-  static constexpr size_t Min(size_type length_a, size_type length_b) {
-    return length_a < length_b ? length_a : length_b;
-  }
-
-  static constexpr int CompareImpl(size_type length_a, size_type length_b,
-                                   int compare_result) {
-    return compare_result == 0 ? static_cast<int>(length_a > length_b) -
-                                     static_cast<int>(length_a < length_b)
-                               : (compare_result < 0 ? -1 : 1);
-  }
-
-  absl::Nullable<const char*> ptr_;
-  size_type length_;
-};
-
-// This large function is defined inline so that in a fairly common case where
-// one of the arguments is a literal, the compiler can elide a lot of the
-// following comparisons.
-constexpr bool operator==(string_view x, string_view y) noexcept {
-  return x.size() == y.size() &&
-         (x.empty() ||
-          ABSL_INTERNAL_STRING_VIEW_MEMCMP(x.data(), y.data(), x.size()) == 0);
-}
-
-constexpr bool operator!=(string_view x, string_view y) noexcept {
-  return !(x == y);
-}
-
-constexpr bool operator<(string_view x, string_view y) noexcept {
-  return x.compare(y) < 0;
-}
-
-constexpr bool operator>(string_view x, string_view y) noexcept {
-  return y < x;
-}
-
-constexpr bool operator<=(string_view x, string_view y) noexcept {
-  return !(y < x);
-}
-
-constexpr bool operator>=(string_view x, string_view y) noexcept {
-  return !(x < y);
-}
-
-// IO Insertion Operator
-std::ostream& operator<<(std::ostream& o, string_view piece);
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#undef ABSL_INTERNAL_STRING_VIEW_MEMCMP
-
-#endif  // ABSL_USES_STD_STRING_VIEW
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// ClippedSubstr()
-//
-// Like `s.substr(pos, n)`, but clips `pos` to an upper bound of `s.size()`.
-// Provided because std::string_view::substr throws if `pos > size()`
-inline string_view ClippedSubstr(string_view s, size_t pos,
-                                 size_t n = string_view::npos) {
-  pos = (std::min)(pos, static_cast<size_t>(s.size()));
-  return s.substr(pos, n);
-}
-
-// NullSafeStringView()
-//
-// Creates an `absl::string_view` from a pointer `p` even if it's null-valued.
-// This function should be used where an `absl::string_view` can be created from
-// a possibly-null pointer.
-constexpr string_view NullSafeStringView(absl::Nullable<const char*> p) {
-  return p ? string_view(p) : string_view();
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_STRING_VIEW_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/strip.h b/projects/llm_framework/include/abseil-cpp/absl/strings/strip.h
deleted file mode 100644
index e3cda5ba..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/strip.h
+++ /dev/null
@@ -1,96 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: strip.h
-// -----------------------------------------------------------------------------
-//
-// This file contains various functions for stripping substrings from a string.
-#ifndef ABSL_STRINGS_STRIP_H_
-#define ABSL_STRINGS_STRIP_H_
-
-#include <cstddef>
-#include <string>
-
-#include "absl/base/macros.h"
-#include "absl/base/nullability.h"
-#include "absl/strings/ascii.h"
-#include "absl/strings/match.h"
-#include "absl/strings/string_view.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// ConsumePrefix()
-//
-// Strips the `expected` prefix, if found, from the start of `str`.
-// If the operation succeeded, `true` is returned.  If not, `false`
-// is returned and `str` is not modified.
-//
-// Example:
-//
-//   absl::string_view input("abc");
-//   EXPECT_TRUE(absl::ConsumePrefix(&input, "a"));
-//   EXPECT_EQ(input, "bc");
-inline bool ConsumePrefix(absl::Nonnull<absl::string_view*> str,
-                          absl::string_view expected) {
-  if (!absl::StartsWith(*str, expected)) return false;
-  str->remove_prefix(expected.size());
-  return true;
-}
-// ConsumeSuffix()
-//
-// Strips the `expected` suffix, if found, from the end of `str`.
-// If the operation succeeded, `true` is returned.  If not, `false`
-// is returned and `str` is not modified.
-//
-// Example:
-//
-//   absl::string_view input("abcdef");
-//   EXPECT_TRUE(absl::ConsumeSuffix(&input, "def"));
-//   EXPECT_EQ(input, "abc");
-inline bool ConsumeSuffix(absl::Nonnull<absl::string_view*> str,
-                          absl::string_view expected) {
-  if (!absl::EndsWith(*str, expected)) return false;
-  str->remove_suffix(expected.size());
-  return true;
-}
-
-// StripPrefix()
-//
-// Returns a view into the input string `str` with the given `prefix` removed,
-// but leaving the original string intact. If the prefix does not match at the
-// start of the string, returns the original string instead.
-ABSL_MUST_USE_RESULT inline absl::string_view StripPrefix(
-    absl::string_view str, absl::string_view prefix) {
-  if (absl::StartsWith(str, prefix)) str.remove_prefix(prefix.size());
-  return str;
-}
-
-// StripSuffix()
-//
-// Returns a view into the input string `str` with the given `suffix` removed,
-// but leaving the original string intact. If the suffix does not match at the
-// end of the string, returns the original string instead.
-ABSL_MUST_USE_RESULT inline absl::string_view StripSuffix(
-    absl::string_view str, absl::string_view suffix) {
-  if (absl::EndsWith(str, suffix)) str.remove_suffix(suffix.size());
-  return str;
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_STRIP_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/strings/substitute.h b/projects/llm_framework/include/abseil-cpp/absl/strings/substitute.h
deleted file mode 100644
index 6c7cba4b..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/strings/substitute.h
+++ /dev/null
@@ -1,767 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: substitute.h
-// -----------------------------------------------------------------------------
-//
-// This package contains functions for efficiently performing string
-// substitutions using a format string with positional notation:
-// `Substitute()` and `SubstituteAndAppend()`.
-//
-// Unlike printf-style format specifiers, `Substitute()` functions do not need
-// to specify the type of the substitution arguments. Supported arguments
-// following the format string, such as strings, string_views, ints,
-// floats, and bools, are automatically converted to strings during the
-// substitution process. (See below for a full list of supported types.)
-//
-// `Substitute()` does not allow you to specify *how* to format a value, beyond
-// the default conversion to string. For example, you cannot format an integer
-// in hex.
-//
-// The format string uses positional identifiers indicated by a dollar sign ($)
-// and single digit positional ids to indicate which substitution arguments to
-// use at that location within the format string.
-//
-// A '$$' sequence in the format string causes a literal '$' character to be
-// output.
-//
-// Example 1:
-//   std::string s = Substitute("$1 purchased $0 $2 for $$10. Thanks $1!",
-//                              5, "Bob", "Apples");
-//   EXPECT_EQ("Bob purchased 5 Apples for $10. Thanks Bob!", s);
-//
-// Example 2:
-//   std::string s = "Hi. ";
-//   SubstituteAndAppend(&s, "My name is $0 and I am $1 years old.", "Bob", 5);
-//   EXPECT_EQ("Hi. My name is Bob and I am 5 years old.", s);
-//
-// Supported types:
-//   * absl::string_view, std::string, const char* (null is equivalent to "")
-//   * int32_t, int64_t, uint32_t, uint64_t
-//   * float, double
-//   * bool (Printed as "true" or "false")
-//   * pointer types other than char* (Printed as "0x<lower case hex string>",
-//     except that null is printed as "NULL")
-//   * user-defined types via the `AbslStringify()` customization point. See the
-//     documentation for `absl::StrCat` for an explanation on how to use this.
-//
-// If an invalid format string is provided, Substitute returns an empty string
-// and SubstituteAndAppend does not change the provided output string.
-// A format string is invalid if it:
-//   * ends in an unescaped $ character,
-//     e.g. "Hello $", or
-//   * calls for a position argument which is not provided,
-//     e.g. Substitute("Hello $2", "world"), or
-//   * specifies a non-digit, non-$ character after an unescaped $ character,
-//     e.g. "Hello $f".
-// In debug mode, i.e. #ifndef NDEBUG, such errors terminate the program.
-
-#ifndef ABSL_STRINGS_SUBSTITUTE_H_
-#define ABSL_STRINGS_SUBSTITUTE_H_
-
-#include <cstring>
-#include <string>
-#include <type_traits>
-#include <vector>
-
-#include "absl/base/macros.h"
-#include "absl/base/nullability.h"
-#include "absl/base/port.h"
-#include "absl/strings/ascii.h"
-#include "absl/strings/escaping.h"
-#include "absl/strings/internal/stringify_sink.h"
-#include "absl/strings/numbers.h"
-#include "absl/strings/str_cat.h"
-#include "absl/strings/str_split.h"
-#include "absl/strings/string_view.h"
-#include "absl/strings/strip.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace substitute_internal {
-
-// Arg
-//
-// This class provides an argument type for `absl::Substitute()` and
-// `absl::SubstituteAndAppend()`. `Arg` handles implicit conversion of various
-// types to a string. (`Arg` is very similar to the `AlphaNum` class in
-// `StrCat()`.)
-//
-// This class has implicit constructors.
-class Arg {
- public:
-  // Overloads for string-y things
-  //
-  // Explicitly overload `const char*` so the compiler doesn't cast to `bool`.
-  Arg(absl::Nullable<const char*> value)  // NOLINT(google-explicit-constructor)
-      : piece_(absl::NullSafeStringView(value)) {}
-  template <typename Allocator>
-  Arg(  // NOLINT
-      const std::basic_string<char, std::char_traits<char>, Allocator>&
-          value) noexcept
-      : piece_(value) {}
-  Arg(absl::string_view value)  // NOLINT(google-explicit-constructor)
-      : piece_(value) {}
-
-  // Overloads for primitives
-  //
-  // No overloads are available for signed and unsigned char because if people
-  // are explicitly declaring their chars as signed or unsigned then they are
-  // probably using them as 8-bit integers and would probably prefer an integer
-  // representation. However, we can't really know, so we make the caller decide
-  // what to do.
-  Arg(char value)  // NOLINT(google-explicit-constructor)
-      : piece_(scratch_, 1) {
-    scratch_[0] = value;
-  }
-  Arg(short value)  // NOLINT(*)
-      : piece_(scratch_,
-               static_cast<size_t>(
-                   numbers_internal::FastIntToBuffer(value, scratch_) -
-                   scratch_)) {}
-  Arg(unsigned short value)  // NOLINT(*)
-      : piece_(scratch_,
-               static_cast<size_t>(
-                   numbers_internal::FastIntToBuffer(value, scratch_) -
-                   scratch_)) {}
-  Arg(int value)  // NOLINT(google-explicit-constructor)
-      : piece_(scratch_,
-               static_cast<size_t>(
-                   numbers_internal::FastIntToBuffer(value, scratch_) -
-                   scratch_)) {}
-  Arg(unsigned int value)  // NOLINT(google-explicit-constructor)
-      : piece_(scratch_,
-               static_cast<size_t>(
-                   numbers_internal::FastIntToBuffer(value, scratch_) -
-                   scratch_)) {}
-  Arg(long value)  // NOLINT(*)
-      : piece_(scratch_,
-               static_cast<size_t>(
-                   numbers_internal::FastIntToBuffer(value, scratch_) -
-                   scratch_)) {}
-  Arg(unsigned long value)  // NOLINT(*)
-      : piece_(scratch_,
-               static_cast<size_t>(
-                   numbers_internal::FastIntToBuffer(value, scratch_) -
-                   scratch_)) {}
-  Arg(long long value)  // NOLINT(*)
-      : piece_(scratch_,
-               static_cast<size_t>(
-                   numbers_internal::FastIntToBuffer(value, scratch_) -
-                   scratch_)) {}
-  Arg(unsigned long long value)  // NOLINT(*)
-      : piece_(scratch_,
-               static_cast<size_t>(
-                   numbers_internal::FastIntToBuffer(value, scratch_) -
-                   scratch_)) {}
-  Arg(float value)  // NOLINT(google-explicit-constructor)
-      : piece_(scratch_, numbers_internal::SixDigitsToBuffer(value, scratch_)) {
-  }
-  Arg(double value)  // NOLINT(google-explicit-constructor)
-      : piece_(scratch_, numbers_internal::SixDigitsToBuffer(value, scratch_)) {
-  }
-  Arg(bool value)  // NOLINT(google-explicit-constructor)
-      : piece_(value ? "true" : "false") {}
-
-  template <typename T, typename = typename std::enable_if<
-                            HasAbslStringify<T>::value>::type>
-  Arg(  // NOLINT(google-explicit-constructor)
-      const T& v, strings_internal::StringifySink&& sink = {})
-      : piece_(strings_internal::ExtractStringification(sink, v)) {}
-
-  Arg(Hex hex);  // NOLINT(google-explicit-constructor)
-  Arg(Dec dec);  // NOLINT(google-explicit-constructor)
-
-  // vector<bool>::reference and const_reference require special help to convert
-  // to `Arg` because it requires two user defined conversions.
-  template <typename T,
-            absl::enable_if_t<
-                std::is_class<T>::value &&
-                (std::is_same<T, std::vector<bool>::reference>::value ||
-                 std::is_same<T, std::vector<bool>::const_reference>::value)>* =
-                nullptr>
-  Arg(T value)  // NOLINT(google-explicit-constructor)
-      : Arg(static_cast<bool>(value)) {}
-
-  // `void*` values, with the exception of `char*`, are printed as
-  // "0x<hex value>". However, in the case of `nullptr`, "NULL" is printed.
-  Arg(  // NOLINT(google-explicit-constructor)
-      absl::Nullable<const void*> value);
-
-  // Normal enums are already handled by the integer formatters.
-  // This overload matches only scoped enums.
-  template <typename T,
-            typename = typename std::enable_if<
-                std::is_enum<T>{} && !std::is_convertible<T, int>{} &&
-                !HasAbslStringify<T>::value>::type>
-  Arg(T value)  // NOLINT(google-explicit-constructor)
-      : Arg(static_cast<typename std::underlying_type<T>::type>(value)) {}
-
-  Arg(const Arg&) = delete;
-  Arg& operator=(const Arg&) = delete;
-
-  absl::string_view piece() const { return piece_; }
-
- private:
-  absl::string_view piece_;
-  char scratch_[numbers_internal::kFastToBufferSize];
-};
-
-// Internal helper function. Don't call this from outside this implementation.
-// This interface may change without notice.
-void SubstituteAndAppendArray(
-    absl::Nonnull<std::string*> output, absl::string_view format,
-    absl::Nullable<const absl::string_view*> args_array, size_t num_args);
-
-#if defined(ABSL_BAD_CALL_IF)
-constexpr int CalculateOneBit(absl::Nonnull<const char*> format) {
-  // Returns:
-  // * 2^N for '$N' when N is in [0-9]
-  // * 0 for correct '$' escaping: '$$'.
-  // * -1 otherwise.
-  return (*format < '0' || *format > '9') ? (*format == '$' ? 0 : -1)
-                                          : (1 << (*format - '0'));
-}
-
-constexpr const char* SkipNumber(absl::Nonnull<const char*> format) {
-  return !*format ? format : (format + 1);
-}
-
-constexpr int PlaceholderBitmask(absl::Nonnull<const char*> format) {
-  return !*format
-             ? 0
-             : *format != '$' ? PlaceholderBitmask(format + 1)
-                              : (CalculateOneBit(format + 1) |
-                                 PlaceholderBitmask(SkipNumber(format + 1)));
-}
-#endif  // ABSL_BAD_CALL_IF
-
-}  // namespace substitute_internal
-
-//
-// PUBLIC API
-//
-
-// SubstituteAndAppend()
-//
-// Substitutes variables into a given format string and appends to a given
-// output string. See file comments above for usage.
-//
-// The declarations of `SubstituteAndAppend()` below consist of overloads
-// for passing 0 to 10 arguments, respectively.
-//
-// NOTE: A zero-argument `SubstituteAndAppend()` may be used within variadic
-// templates to allow a variable number of arguments.
-//
-// Example:
-//  template <typename... Args>
-//  void VarMsg(std::string* boilerplate, absl::string_view format,
-//      const Args&... args) {
-//    absl::SubstituteAndAppend(boilerplate, format, args...);
-//  }
-//
-inline void SubstituteAndAppend(absl::Nonnull<std::string*> output,
-                                absl::string_view format) {
-  substitute_internal::SubstituteAndAppendArray(output, format, nullptr, 0);
-}
-
-inline void SubstituteAndAppend(absl::Nonnull<std::string*> output,
-                                absl::string_view format,
-                                const substitute_internal::Arg& a0) {
-  const absl::string_view args[] = {a0.piece()};
-  substitute_internal::SubstituteAndAppendArray(output, format, args,
-                                                ABSL_ARRAYSIZE(args));
-}
-
-inline void SubstituteAndAppend(absl::Nonnull<std::string*> output,
-                                absl::string_view format,
-                                const substitute_internal::Arg& a0,
-                                const substitute_internal::Arg& a1) {
-  const absl::string_view args[] = {a0.piece(), a1.piece()};
-  substitute_internal::SubstituteAndAppendArray(output, format, args,
-                                                ABSL_ARRAYSIZE(args));
-}
-
-inline void SubstituteAndAppend(absl::Nonnull<std::string*> output,
-                                absl::string_view format,
-                                const substitute_internal::Arg& a0,
-                                const substitute_internal::Arg& a1,
-                                const substitute_internal::Arg& a2) {
-  const absl::string_view args[] = {a0.piece(), a1.piece(), a2.piece()};
-  substitute_internal::SubstituteAndAppendArray(output, format, args,
-                                                ABSL_ARRAYSIZE(args));
-}
-
-inline void SubstituteAndAppend(absl::Nonnull<std::string*> output,
-                                absl::string_view format,
-                                const substitute_internal::Arg& a0,
-                                const substitute_internal::Arg& a1,
-                                const substitute_internal::Arg& a2,
-                                const substitute_internal::Arg& a3) {
-  const absl::string_view args[] = {a0.piece(), a1.piece(), a2.piece(),
-                                    a3.piece()};
-  substitute_internal::SubstituteAndAppendArray(output, format, args,
-                                                ABSL_ARRAYSIZE(args));
-}
-
-inline void SubstituteAndAppend(absl::Nonnull<std::string*> output,
-                                absl::string_view format,
-                                const substitute_internal::Arg& a0,
-                                const substitute_internal::Arg& a1,
-                                const substitute_internal::Arg& a2,
-                                const substitute_internal::Arg& a3,
-                                const substitute_internal::Arg& a4) {
-  const absl::string_view args[] = {a0.piece(), a1.piece(), a2.piece(),
-                                    a3.piece(), a4.piece()};
-  substitute_internal::SubstituteAndAppendArray(output, format, args,
-                                                ABSL_ARRAYSIZE(args));
-}
-
-inline void SubstituteAndAppend(
-    absl::Nonnull<std::string*> output, absl::string_view format,
-    const substitute_internal::Arg& a0, const substitute_internal::Arg& a1,
-    const substitute_internal::Arg& a2, const substitute_internal::Arg& a3,
-    const substitute_internal::Arg& a4, const substitute_internal::Arg& a5) {
-  const absl::string_view args[] = {a0.piece(), a1.piece(), a2.piece(),
-                                    a3.piece(), a4.piece(), a5.piece()};
-  substitute_internal::SubstituteAndAppendArray(output, format, args,
-                                                ABSL_ARRAYSIZE(args));
-}
-
-inline void SubstituteAndAppend(
-    absl::Nonnull<std::string*> output, absl::string_view format,
-    const substitute_internal::Arg& a0, const substitute_internal::Arg& a1,
-    const substitute_internal::Arg& a2, const substitute_internal::Arg& a3,
-    const substitute_internal::Arg& a4, const substitute_internal::Arg& a5,
-    const substitute_internal::Arg& a6) {
-  const absl::string_view args[] = {a0.piece(), a1.piece(), a2.piece(),
-                                    a3.piece(), a4.piece(), a5.piece(),
-                                    a6.piece()};
-  substitute_internal::SubstituteAndAppendArray(output, format, args,
-                                                ABSL_ARRAYSIZE(args));
-}
-
-inline void SubstituteAndAppend(
-    absl::Nonnull<std::string*> output, absl::string_view format,
-    const substitute_internal::Arg& a0, const substitute_internal::Arg& a1,
-    const substitute_internal::Arg& a2, const substitute_internal::Arg& a3,
-    const substitute_internal::Arg& a4, const substitute_internal::Arg& a5,
-    const substitute_internal::Arg& a6, const substitute_internal::Arg& a7) {
-  const absl::string_view args[] = {a0.piece(), a1.piece(), a2.piece(),
-                                    a3.piece(), a4.piece(), a5.piece(),
-                                    a6.piece(), a7.piece()};
-  substitute_internal::SubstituteAndAppendArray(output, format, args,
-                                                ABSL_ARRAYSIZE(args));
-}
-
-inline void SubstituteAndAppend(
-    absl::Nonnull<std::string*> output, absl::string_view format,
-    const substitute_internal::Arg& a0, const substitute_internal::Arg& a1,
-    const substitute_internal::Arg& a2, const substitute_internal::Arg& a3,
-    const substitute_internal::Arg& a4, const substitute_internal::Arg& a5,
-    const substitute_internal::Arg& a6, const substitute_internal::Arg& a7,
-    const substitute_internal::Arg& a8) {
-  const absl::string_view args[] = {a0.piece(), a1.piece(), a2.piece(),
-                                    a3.piece(), a4.piece(), a5.piece(),
-                                    a6.piece(), a7.piece(), a8.piece()};
-  substitute_internal::SubstituteAndAppendArray(output, format, args,
-                                                ABSL_ARRAYSIZE(args));
-}
-
-inline void SubstituteAndAppend(
-    absl::Nonnull<std::string*> output, absl::string_view format,
-    const substitute_internal::Arg& a0, const substitute_internal::Arg& a1,
-    const substitute_internal::Arg& a2, const substitute_internal::Arg& a3,
-    const substitute_internal::Arg& a4, const substitute_internal::Arg& a5,
-    const substitute_internal::Arg& a6, const substitute_internal::Arg& a7,
-    const substitute_internal::Arg& a8, const substitute_internal::Arg& a9) {
-  const absl::string_view args[] = {
-      a0.piece(), a1.piece(), a2.piece(), a3.piece(), a4.piece(),
-      a5.piece(), a6.piece(), a7.piece(), a8.piece(), a9.piece()};
-  substitute_internal::SubstituteAndAppendArray(output, format, args,
-                                                ABSL_ARRAYSIZE(args));
-}
-
-#if defined(ABSL_BAD_CALL_IF)
-// This body of functions catches cases where the number of placeholders
-// doesn't match the number of data arguments.
-void SubstituteAndAppend(absl::Nonnull<std::string*> output,
-                         absl::Nonnull<const char*> format)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 0,
-        "There were no substitution arguments "
-        "but this format string either has a $[0-9] in it or contains "
-        "an unescaped $ character (use $$ instead)");
-
-void SubstituteAndAppend(absl::Nonnull<std::string*> output,
-                         absl::Nonnull<const char*> format,
-                         const substitute_internal::Arg& a0)
-    ABSL_BAD_CALL_IF(substitute_internal::PlaceholderBitmask(format) != 1,
-                     "There was 1 substitution argument given, but "
-                     "this format string is missing its $0, contains "
-                     "one of $1-$9, or contains an unescaped $ character (use "
-                     "$$ instead)");
-
-void SubstituteAndAppend(absl::Nonnull<std::string*> output,
-                         absl::Nonnull<const char*> format,
-                         const substitute_internal::Arg& a0,
-                         const substitute_internal::Arg& a1)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 3,
-        "There were 2 substitution arguments given, but this format string is "
-        "missing its $0/$1, contains one of $2-$9, or contains an "
-        "unescaped $ character (use $$ instead)");
-
-void SubstituteAndAppend(absl::Nonnull<std::string*> output,
-                         absl::Nonnull<const char*> format,
-                         const substitute_internal::Arg& a0,
-                         const substitute_internal::Arg& a1,
-                         const substitute_internal::Arg& a2)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 7,
-        "There were 3 substitution arguments given, but "
-        "this format string is missing its $0/$1/$2, contains one of "
-        "$3-$9, or contains an unescaped $ character (use $$ instead)");
-
-void SubstituteAndAppend(absl::Nonnull<std::string*> output,
-                         absl::Nonnull<const char*> format,
-                         const substitute_internal::Arg& a0,
-                         const substitute_internal::Arg& a1,
-                         const substitute_internal::Arg& a2,
-                         const substitute_internal::Arg& a3)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 15,
-        "There were 4 substitution arguments given, but "
-        "this format string is missing its $0-$3, contains one of "
-        "$4-$9, or contains an unescaped $ character (use $$ instead)");
-
-void SubstituteAndAppend(absl::Nonnull<std::string*> output,
-                         absl::Nonnull<const char*> format,
-                         const substitute_internal::Arg& a0,
-                         const substitute_internal::Arg& a1,
-                         const substitute_internal::Arg& a2,
-                         const substitute_internal::Arg& a3,
-                         const substitute_internal::Arg& a4)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 31,
-        "There were 5 substitution arguments given, but "
-        "this format string is missing its $0-$4, contains one of "
-        "$5-$9, or contains an unescaped $ character (use $$ instead)");
-
-void SubstituteAndAppend(
-    absl::Nonnull<std::string*> output, absl::Nonnull<const char*> format,
-    const substitute_internal::Arg& a0, const substitute_internal::Arg& a1,
-    const substitute_internal::Arg& a2, const substitute_internal::Arg& a3,
-    const substitute_internal::Arg& a4, const substitute_internal::Arg& a5)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 63,
-        "There were 6 substitution arguments given, but "
-        "this format string is missing its $0-$5, contains one of "
-        "$6-$9, or contains an unescaped $ character (use $$ instead)");
-
-void SubstituteAndAppend(
-    absl::Nonnull<std::string*> output, absl::Nonnull<const char*> format,
-    const substitute_internal::Arg& a0, const substitute_internal::Arg& a1,
-    const substitute_internal::Arg& a2, const substitute_internal::Arg& a3,
-    const substitute_internal::Arg& a4, const substitute_internal::Arg& a5,
-    const substitute_internal::Arg& a6)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 127,
-        "There were 7 substitution arguments given, but "
-        "this format string is missing its $0-$6, contains one of "
-        "$7-$9, or contains an unescaped $ character (use $$ instead)");
-
-void SubstituteAndAppend(
-    absl::Nonnull<std::string*> output, absl::Nonnull<const char*> format,
-    const substitute_internal::Arg& a0, const substitute_internal::Arg& a1,
-    const substitute_internal::Arg& a2, const substitute_internal::Arg& a3,
-    const substitute_internal::Arg& a4, const substitute_internal::Arg& a5,
-    const substitute_internal::Arg& a6, const substitute_internal::Arg& a7)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 255,
-        "There were 8 substitution arguments given, but "
-        "this format string is missing its $0-$7, contains one of "
-        "$8-$9, or contains an unescaped $ character (use $$ instead)");
-
-void SubstituteAndAppend(
-    absl::Nonnull<std::string*> output, absl::Nonnull<const char*> format,
-    const substitute_internal::Arg& a0, const substitute_internal::Arg& a1,
-    const substitute_internal::Arg& a2, const substitute_internal::Arg& a3,
-    const substitute_internal::Arg& a4, const substitute_internal::Arg& a5,
-    const substitute_internal::Arg& a6, const substitute_internal::Arg& a7,
-    const substitute_internal::Arg& a8)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 511,
-        "There were 9 substitution arguments given, but "
-        "this format string is missing its $0-$8, contains a $9, or "
-        "contains an unescaped $ character (use $$ instead)");
-
-void SubstituteAndAppend(
-    absl::Nonnull<std::string*> output, absl::Nonnull<const char*> format,
-    const substitute_internal::Arg& a0, const substitute_internal::Arg& a1,
-    const substitute_internal::Arg& a2, const substitute_internal::Arg& a3,
-    const substitute_internal::Arg& a4, const substitute_internal::Arg& a5,
-    const substitute_internal::Arg& a6, const substitute_internal::Arg& a7,
-    const substitute_internal::Arg& a8, const substitute_internal::Arg& a9)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 1023,
-        "There were 10 substitution arguments given, but this "
-        "format string either doesn't contain all of $0 through $9 or "
-        "contains an unescaped $ character (use $$ instead)");
-#endif  // ABSL_BAD_CALL_IF
-
-// Substitute()
-//
-// Substitutes variables into a given format string. See file comments above
-// for usage.
-//
-// The declarations of `Substitute()` below consist of overloads for passing 0
-// to 10 arguments, respectively.
-//
-// NOTE: A zero-argument `Substitute()` may be used within variadic templates to
-// allow a variable number of arguments.
-//
-// Example:
-//  template <typename... Args>
-//  void VarMsg(absl::string_view format, const Args&... args) {
-//    std::string s = absl::Substitute(format, args...);
-
-ABSL_MUST_USE_RESULT inline std::string Substitute(absl::string_view format) {
-  std::string result;
-  SubstituteAndAppend(&result, format);
-  return result;
-}
-
-ABSL_MUST_USE_RESULT inline std::string Substitute(
-    absl::string_view format, const substitute_internal::Arg& a0) {
-  std::string result;
-  SubstituteAndAppend(&result, format, a0);
-  return result;
-}
-
-ABSL_MUST_USE_RESULT inline std::string Substitute(
-    absl::string_view format, const substitute_internal::Arg& a0,
-    const substitute_internal::Arg& a1) {
-  std::string result;
-  SubstituteAndAppend(&result, format, a0, a1);
-  return result;
-}
-
-ABSL_MUST_USE_RESULT inline std::string Substitute(
-    absl::string_view format, const substitute_internal::Arg& a0,
-    const substitute_internal::Arg& a1, const substitute_internal::Arg& a2) {
-  std::string result;
-  SubstituteAndAppend(&result, format, a0, a1, a2);
-  return result;
-}
-
-ABSL_MUST_USE_RESULT inline std::string Substitute(
-    absl::string_view format, const substitute_internal::Arg& a0,
-    const substitute_internal::Arg& a1, const substitute_internal::Arg& a2,
-    const substitute_internal::Arg& a3) {
-  std::string result;
-  SubstituteAndAppend(&result, format, a0, a1, a2, a3);
-  return result;
-}
-
-ABSL_MUST_USE_RESULT inline std::string Substitute(
-    absl::string_view format, const substitute_internal::Arg& a0,
-    const substitute_internal::Arg& a1, const substitute_internal::Arg& a2,
-    const substitute_internal::Arg& a3, const substitute_internal::Arg& a4) {
-  std::string result;
-  SubstituteAndAppend(&result, format, a0, a1, a2, a3, a4);
-  return result;
-}
-
-ABSL_MUST_USE_RESULT inline std::string Substitute(
-    absl::string_view format, const substitute_internal::Arg& a0,
-    const substitute_internal::Arg& a1, const substitute_internal::Arg& a2,
-    const substitute_internal::Arg& a3, const substitute_internal::Arg& a4,
-    const substitute_internal::Arg& a5) {
-  std::string result;
-  SubstituteAndAppend(&result, format, a0, a1, a2, a3, a4, a5);
-  return result;
-}
-
-ABSL_MUST_USE_RESULT inline std::string Substitute(
-    absl::string_view format, const substitute_internal::Arg& a0,
-    const substitute_internal::Arg& a1, const substitute_internal::Arg& a2,
-    const substitute_internal::Arg& a3, const substitute_internal::Arg& a4,
-    const substitute_internal::Arg& a5, const substitute_internal::Arg& a6) {
-  std::string result;
-  SubstituteAndAppend(&result, format, a0, a1, a2, a3, a4, a5, a6);
-  return result;
-}
-
-ABSL_MUST_USE_RESULT inline std::string Substitute(
-    absl::string_view format, const substitute_internal::Arg& a0,
-    const substitute_internal::Arg& a1, const substitute_internal::Arg& a2,
-    const substitute_internal::Arg& a3, const substitute_internal::Arg& a4,
-    const substitute_internal::Arg& a5, const substitute_internal::Arg& a6,
-    const substitute_internal::Arg& a7) {
-  std::string result;
-  SubstituteAndAppend(&result, format, a0, a1, a2, a3, a4, a5, a6, a7);
-  return result;
-}
-
-ABSL_MUST_USE_RESULT inline std::string Substitute(
-    absl::string_view format, const substitute_internal::Arg& a0,
-    const substitute_internal::Arg& a1, const substitute_internal::Arg& a2,
-    const substitute_internal::Arg& a3, const substitute_internal::Arg& a4,
-    const substitute_internal::Arg& a5, const substitute_internal::Arg& a6,
-    const substitute_internal::Arg& a7, const substitute_internal::Arg& a8) {
-  std::string result;
-  SubstituteAndAppend(&result, format, a0, a1, a2, a3, a4, a5, a6, a7, a8);
-  return result;
-}
-
-ABSL_MUST_USE_RESULT inline std::string Substitute(
-    absl::string_view format, const substitute_internal::Arg& a0,
-    const substitute_internal::Arg& a1, const substitute_internal::Arg& a2,
-    const substitute_internal::Arg& a3, const substitute_internal::Arg& a4,
-    const substitute_internal::Arg& a5, const substitute_internal::Arg& a6,
-    const substitute_internal::Arg& a7, const substitute_internal::Arg& a8,
-    const substitute_internal::Arg& a9) {
-  std::string result;
-  SubstituteAndAppend(&result, format, a0, a1, a2, a3, a4, a5, a6, a7, a8, a9);
-  return result;
-}
-
-#if defined(ABSL_BAD_CALL_IF)
-// This body of functions catches cases where the number of placeholders
-// doesn't match the number of data arguments.
-std::string Substitute(absl::Nonnull<const char*> format)
-    ABSL_BAD_CALL_IF(substitute_internal::PlaceholderBitmask(format) != 0,
-                     "There were no substitution arguments "
-                     "but this format string either has a $[0-9] in it or "
-                     "contains an unescaped $ character (use $$ instead)");
-
-std::string Substitute(absl::Nonnull<const char*> format,
-                       const substitute_internal::Arg& a0)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 1,
-        "There was 1 substitution argument given, but "
-        "this format string is missing its $0, contains one of $1-$9, "
-        "or contains an unescaped $ character (use $$ instead)");
-
-std::string Substitute(absl::Nonnull<const char*> format,
-                       const substitute_internal::Arg& a0,
-                       const substitute_internal::Arg& a1)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 3,
-        "There were 2 substitution arguments given, but "
-        "this format string is missing its $0/$1, contains one of "
-        "$2-$9, or contains an unescaped $ character (use $$ instead)");
-
-std::string Substitute(absl::Nonnull<const char*> format,
-                       const substitute_internal::Arg& a0,
-                       const substitute_internal::Arg& a1,
-                       const substitute_internal::Arg& a2)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 7,
-        "There were 3 substitution arguments given, but "
-        "this format string is missing its $0/$1/$2, contains one of "
-        "$3-$9, or contains an unescaped $ character (use $$ instead)");
-
-std::string Substitute(absl::Nonnull<const char*> format,
-                       const substitute_internal::Arg& a0,
-                       const substitute_internal::Arg& a1,
-                       const substitute_internal::Arg& a2,
-                       const substitute_internal::Arg& a3)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 15,
-        "There were 4 substitution arguments given, but "
-        "this format string is missing its $0-$3, contains one of "
-        "$4-$9, or contains an unescaped $ character (use $$ instead)");
-
-std::string Substitute(absl::Nonnull<const char*> format,
-                       const substitute_internal::Arg& a0,
-                       const substitute_internal::Arg& a1,
-                       const substitute_internal::Arg& a2,
-                       const substitute_internal::Arg& a3,
-                       const substitute_internal::Arg& a4)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 31,
-        "There were 5 substitution arguments given, but "
-        "this format string is missing its $0-$4, contains one of "
-        "$5-$9, or contains an unescaped $ character (use $$ instead)");
-
-std::string Substitute(absl::Nonnull<const char*> format,
-                       const substitute_internal::Arg& a0,
-                       const substitute_internal::Arg& a1,
-                       const substitute_internal::Arg& a2,
-                       const substitute_internal::Arg& a3,
-                       const substitute_internal::Arg& a4,
-                       const substitute_internal::Arg& a5)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 63,
-        "There were 6 substitution arguments given, but "
-        "this format string is missing its $0-$5, contains one of "
-        "$6-$9, or contains an unescaped $ character (use $$ instead)");
-
-std::string Substitute(
-    absl::Nonnull<const char*> format, const substitute_internal::Arg& a0,
-    const substitute_internal::Arg& a1, const substitute_internal::Arg& a2,
-    const substitute_internal::Arg& a3, const substitute_internal::Arg& a4,
-    const substitute_internal::Arg& a5, const substitute_internal::Arg& a6)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 127,
-        "There were 7 substitution arguments given, but "
-        "this format string is missing its $0-$6, contains one of "
-        "$7-$9, or contains an unescaped $ character (use $$ instead)");
-
-std::string Substitute(
-    absl::Nonnull<const char*> format, const substitute_internal::Arg& a0,
-    const substitute_internal::Arg& a1, const substitute_internal::Arg& a2,
-    const substitute_internal::Arg& a3, const substitute_internal::Arg& a4,
-    const substitute_internal::Arg& a5, const substitute_internal::Arg& a6,
-    const substitute_internal::Arg& a7)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 255,
-        "There were 8 substitution arguments given, but "
-        "this format string is missing its $0-$7, contains one of "
-        "$8-$9, or contains an unescaped $ character (use $$ instead)");
-
-std::string Substitute(
-    absl::Nonnull<const char*> format, const substitute_internal::Arg& a0,
-    const substitute_internal::Arg& a1, const substitute_internal::Arg& a2,
-    const substitute_internal::Arg& a3, const substitute_internal::Arg& a4,
-    const substitute_internal::Arg& a5, const substitute_internal::Arg& a6,
-    const substitute_internal::Arg& a7, const substitute_internal::Arg& a8)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 511,
-        "There were 9 substitution arguments given, but "
-        "this format string is missing its $0-$8, contains a $9, or "
-        "contains an unescaped $ character (use $$ instead)");
-
-std::string Substitute(
-    absl::Nonnull<const char*> format, const substitute_internal::Arg& a0,
-    const substitute_internal::Arg& a1, const substitute_internal::Arg& a2,
-    const substitute_internal::Arg& a3, const substitute_internal::Arg& a4,
-    const substitute_internal::Arg& a5, const substitute_internal::Arg& a6,
-    const substitute_internal::Arg& a7, const substitute_internal::Arg& a8,
-    const substitute_internal::Arg& a9)
-    ABSL_BAD_CALL_IF(
-        substitute_internal::PlaceholderBitmask(format) != 1023,
-        "There were 10 substitution arguments given, but this "
-        "format string either doesn't contain all of $0 through $9 or "
-        "contains an unescaped $ character (use $$ instead)");
-#endif  // ABSL_BAD_CALL_IF
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_STRINGS_SUBSTITUTE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/barrier.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/barrier.h
deleted file mode 100644
index d8e75440..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/barrier.h
+++ /dev/null
@@ -1,79 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// barrier.h
-// -----------------------------------------------------------------------------
-
-#ifndef ABSL_SYNCHRONIZATION_BARRIER_H_
-#define ABSL_SYNCHRONIZATION_BARRIER_H_
-
-#include "absl/base/thread_annotations.h"
-#include "absl/synchronization/mutex.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// Barrier
-//
-// This class creates a barrier which blocks threads until a prespecified
-// threshold of threads (`num_threads`) utilizes the barrier. A thread utilizes
-// the `Barrier` by calling `Block()` on the barrier, which will block that
-// thread; no call to `Block()` will return until `num_threads` threads have
-// called it.
-//
-// Exactly one call to `Block()` will return `true`, which is then responsible
-// for destroying the barrier; because stack allocation will cause the barrier
-// to be deleted when it is out of scope, barriers should not be stack
-// allocated.
-//
-// Example:
-//
-//   // Main thread creates a `Barrier`:
-//   barrier = new Barrier(num_threads);
-//
-//   // Each participating thread could then call:
-//   if (barrier->Block()) delete barrier;  // Exactly one call to `Block()`
-//                                          // returns `true`; that call
-//                                          // deletes the barrier.
-class Barrier {
- public:
-  // `num_threads` is the number of threads that will participate in the barrier
-  explicit Barrier(int num_threads)
-      : num_to_block_(num_threads), num_to_exit_(num_threads) {}
-
-  Barrier(const Barrier&) = delete;
-  Barrier& operator=(const Barrier&) = delete;
-
-  // Barrier::Block()
-  //
-  // Blocks the current thread, and returns only when the `num_threads`
-  // threshold of threads utilizing this barrier has been reached. `Block()`
-  // returns `true` for precisely one caller, which may then destroy the
-  // barrier.
-  //
-  // Memory ordering: For any threads X and Y, any action taken by X
-  // before X calls `Block()` will be visible to Y after Y returns from
-  // `Block()`.
-  bool Block();
-
- private:
-  Mutex lock_;
-  int num_to_block_ ABSL_GUARDED_BY(lock_);
-  int num_to_exit_ ABSL_GUARDED_BY(lock_);
-};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-#endif  // ABSL_SYNCHRONIZATION_BARRIER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/blocking_counter.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/blocking_counter.h
deleted file mode 100644
index 1908fdb1..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/blocking_counter.h
+++ /dev/null
@@ -1,101 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// blocking_counter.h
-// -----------------------------------------------------------------------------
-
-#ifndef ABSL_SYNCHRONIZATION_BLOCKING_COUNTER_H_
-#define ABSL_SYNCHRONIZATION_BLOCKING_COUNTER_H_
-
-#include <atomic>
-
-#include "absl/base/thread_annotations.h"
-#include "absl/synchronization/mutex.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// BlockingCounter
-//
-// This class allows a thread to block for a pre-specified number of actions.
-// `BlockingCounter` maintains a single non-negative abstract integer "count"
-// with an initial value `initial_count`. A thread can then call `Wait()` on
-// this blocking counter to block until the specified number of events occur;
-// worker threads then call 'DecrementCount()` on the counter upon completion of
-// their work. Once the counter's internal "count" reaches zero, the blocked
-// thread unblocks.
-//
-// A `BlockingCounter` requires the following:
-//     - its `initial_count` is non-negative.
-//     - the number of calls to `DecrementCount()` on it is at most
-//       `initial_count`.
-//     - `Wait()` is called at most once on it.
-//
-// Given the above requirements, a `BlockingCounter` provides the following
-// guarantees:
-//     - Once its internal "count" reaches zero, no legal action on the object
-//       can further change the value of "count".
-//     - When `Wait()` returns, it is legal to destroy the `BlockingCounter`.
-//     - When `Wait()` returns, the number of calls to `DecrementCount()` on
-//       this blocking counter exactly equals `initial_count`.
-//
-// Example:
-//     BlockingCounter bcount(N);         // there are N items of work
-//     ... Allow worker threads to start.
-//     ... On completing each work item, workers do:
-//     ... bcount.DecrementCount();      // an item of work has been completed
-//
-//     bcount.Wait();                    // wait for all work to be complete
-//
-class BlockingCounter {
- public:
-  explicit BlockingCounter(int initial_count);
-
-  BlockingCounter(const BlockingCounter&) = delete;
-  BlockingCounter& operator=(const BlockingCounter&) = delete;
-
-  // BlockingCounter::DecrementCount()
-  //
-  // Decrements the counter's "count" by one, and return "count == 0". This
-  // function requires that "count != 0" when it is called.
-  //
-  // Memory ordering: For any threads X and Y, any action taken by X
-  // before it calls `DecrementCount()` is visible to thread Y after
-  // Y's call to `DecrementCount()`, provided Y's call returns `true`.
-  bool DecrementCount();
-
-  // BlockingCounter::Wait()
-  //
-  // Blocks until the counter reaches zero. This function may be called at most
-  // once. On return, `DecrementCount()` will have been called "initial_count"
-  // times and the blocking counter may be destroyed.
-  //
-  // Memory ordering: For any threads X and Y, any action taken by X
-  // before X calls `DecrementCount()` is visible to Y after Y returns
-  // from `Wait()`.
-  void Wait();
-
- private:
-  Mutex lock_;
-  std::atomic<int> count_;
-  int num_waiting_ ABSL_GUARDED_BY(lock_);
-  bool done_ ABSL_GUARDED_BY(lock_);
-};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_SYNCHRONIZATION_BLOCKING_COUNTER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/create_thread_identity.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/create_thread_identity.h
deleted file mode 100644
index 4cfde091..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/create_thread_identity.h
+++ /dev/null
@@ -1,56 +0,0 @@
-/*
- * Copyright 2017 The Abseil Authors.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *      https://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-// Interface for getting the current ThreadIdentity, creating one if necessary.
-// See thread_identity.h.
-//
-// This file is separate from thread_identity.h because creating a new
-// ThreadIdentity requires slightly higher level libraries (per_thread_sem
-// and low_level_alloc) than accessing an existing one.  This separation allows
-// us to have a smaller //absl/base:base.
-
-#ifndef ABSL_SYNCHRONIZATION_INTERNAL_CREATE_THREAD_IDENTITY_H_
-#define ABSL_SYNCHRONIZATION_INTERNAL_CREATE_THREAD_IDENTITY_H_
-
-#include "absl/base/internal/thread_identity.h"
-#include "absl/base/port.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace synchronization_internal {
-
-// Allocates and attaches a ThreadIdentity object for the calling thread.
-// For private use only.
-base_internal::ThreadIdentity* CreateThreadIdentity();
-
-// Returns the ThreadIdentity object representing the calling thread; guaranteed
-// to be unique for its lifetime.  The returned object will remain valid for the
-// program's lifetime; although it may be re-assigned to a subsequent thread.
-// If one does not exist for the calling thread, allocate it now.
-inline base_internal::ThreadIdentity* GetOrCreateCurrentThreadIdentity() {
-  base_internal::ThreadIdentity* identity =
-      base_internal::CurrentThreadIdentityIfPresent();
-  if (ABSL_PREDICT_FALSE(identity == nullptr)) {
-    return CreateThreadIdentity();
-  }
-  return identity;
-}
-
-}  // namespace synchronization_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_SYNCHRONIZATION_INTERNAL_CREATE_THREAD_IDENTITY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/futex.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/futex.h
deleted file mode 100644
index 573c01b7..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/futex.h
+++ /dev/null
@@ -1,177 +0,0 @@
-// Copyright 2020 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-#ifndef ABSL_SYNCHRONIZATION_INTERNAL_FUTEX_H_
-#define ABSL_SYNCHRONIZATION_INTERNAL_FUTEX_H_
-
-#include "absl/base/config.h"
-
-#ifndef _WIN32
-#include <sys/time.h>
-#include <unistd.h>
-#endif
-
-#ifdef __linux__
-#include <linux/futex.h>
-#include <sys/syscall.h>
-#endif
-
-#include <errno.h>
-#include <stdio.h>
-#include <time.h>
-
-#include <atomic>
-#include <cstdint>
-#include <limits>
-
-#include "absl/base/optimization.h"
-#include "absl/synchronization/internal/kernel_timeout.h"
-
-#ifdef ABSL_INTERNAL_HAVE_FUTEX
-#error ABSL_INTERNAL_HAVE_FUTEX may not be set on the command line
-#elif defined(__BIONIC__)
-// Bionic supports all the futex operations we need even when some of the futex
-// definitions are missing.
-#define ABSL_INTERNAL_HAVE_FUTEX
-#elif defined(__linux__) && defined(FUTEX_CLOCK_REALTIME)
-// FUTEX_CLOCK_REALTIME requires Linux >= 2.6.28.
-#define ABSL_INTERNAL_HAVE_FUTEX
-#endif
-
-#ifdef ABSL_INTERNAL_HAVE_FUTEX
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace synchronization_internal {
-
-// Some Android headers are missing these definitions even though they
-// support these futex operations.
-#ifdef __BIONIC__
-#ifndef SYS_futex
-#define SYS_futex __NR_futex
-#endif
-#ifndef FUTEX_WAIT_BITSET
-#define FUTEX_WAIT_BITSET 9
-#endif
-#ifndef FUTEX_PRIVATE_FLAG
-#define FUTEX_PRIVATE_FLAG 128
-#endif
-#ifndef FUTEX_CLOCK_REALTIME
-#define FUTEX_CLOCK_REALTIME 256
-#endif
-#ifndef FUTEX_BITSET_MATCH_ANY
-#define FUTEX_BITSET_MATCH_ANY 0xFFFFFFFF
-#endif
-#endif
-
-#if defined(__NR_futex_time64) && !defined(SYS_futex_time64)
-#define SYS_futex_time64 __NR_futex_time64
-#endif
-
-#if defined(SYS_futex_time64) && !defined(SYS_futex)
-#define SYS_futex SYS_futex_time64
-using FutexTimespec = struct timespec;
-#else
-// Some libc implementations have switched to an unconditional 64-bit `time_t`
-// definition. This means that `struct timespec` may not match the layout
-// expected by the kernel ABI on 32-bit platforms. So we define the
-// FutexTimespec that matches the kernel timespec definition. It should be safe
-// to use this struct for 64-bit userspace builds too, since it will use another
-// SYS_futex kernel call with 64-bit tv_sec inside timespec.
-struct FutexTimespec {
-  long tv_sec;   // NOLINT
-  long tv_nsec;  // NOLINT
-};
-#endif
-
-class FutexImpl {
- public:
-  // Atomically check that `*v == val`, and if it is, then sleep until the until
-  // woken by `Wake()`.
-  static int Wait(std::atomic<int32_t>* v, int32_t val) {
-    return WaitAbsoluteTimeout(v, val, nullptr);
-  }
-
-  // Atomically check that `*v == val`, and if it is, then sleep until
-  // CLOCK_REALTIME reaches `*abs_timeout`, or until woken by `Wake()`.
-  static int WaitAbsoluteTimeout(std::atomic<int32_t>* v, int32_t val,
-                                 const struct timespec* abs_timeout) {
-    FutexTimespec ts;
-    // https://locklessinc.com/articles/futex_cheat_sheet/
-    // Unlike FUTEX_WAIT, FUTEX_WAIT_BITSET uses absolute time.
-    auto err = syscall(
-        SYS_futex, reinterpret_cast<int32_t*>(v),
-        FUTEX_WAIT_BITSET | FUTEX_PRIVATE_FLAG | FUTEX_CLOCK_REALTIME, val,
-        ToFutexTimespec(abs_timeout, &ts), nullptr, FUTEX_BITSET_MATCH_ANY);
-    if (err != 0) {
-      return -errno;
-    }
-    return 0;
-  }
-
-  // Atomically check that `*v == val`, and if it is, then sleep until
-  // `*rel_timeout` has elapsed, or until woken by `Wake()`.
-  static int WaitRelativeTimeout(std::atomic<int32_t>* v, int32_t val,
-                                 const struct timespec* rel_timeout) {
-    FutexTimespec ts;
-    // Atomically check that the futex value is still 0, and if it
-    // is, sleep until abs_timeout or until woken by FUTEX_WAKE.
-    auto err =
-        syscall(SYS_futex, reinterpret_cast<int32_t*>(v), FUTEX_PRIVATE_FLAG,
-                val, ToFutexTimespec(rel_timeout, &ts));
-    if (err != 0) {
-      return -errno;
-    }
-    return 0;
-  }
-
-  // Wakes at most `count` waiters that have entered the sleep state on `v`.
-  static int Wake(std::atomic<int32_t>* v, int32_t count) {
-    auto err = syscall(SYS_futex, reinterpret_cast<int32_t*>(v),
-                       FUTEX_WAKE | FUTEX_PRIVATE_FLAG, count);
-    if (ABSL_PREDICT_FALSE(err < 0)) {
-      return -errno;
-    }
-    return 0;
-  }
-
- private:
-  static FutexTimespec* ToFutexTimespec(const struct timespec* userspace_ts,
-                                        FutexTimespec* futex_ts) {
-    if (userspace_ts == nullptr) {
-      return nullptr;
-    }
-
-    using FutexSeconds = decltype(futex_ts->tv_sec);
-    using FutexNanoseconds = decltype(futex_ts->tv_nsec);
-
-    constexpr auto kMaxSeconds{(std::numeric_limits<FutexSeconds>::max)()};
-    if (userspace_ts->tv_sec > kMaxSeconds) {
-      futex_ts->tv_sec = kMaxSeconds;
-    } else {
-      futex_ts->tv_sec = static_cast<FutexSeconds>(userspace_ts->tv_sec);
-    }
-    futex_ts->tv_nsec = static_cast<FutexNanoseconds>(userspace_ts->tv_nsec);
-    return futex_ts;
-  }
-};
-
-class Futex : public FutexImpl {};
-
-}  // namespace synchronization_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_INTERNAL_HAVE_FUTEX
-
-#endif  // ABSL_SYNCHRONIZATION_INTERNAL_FUTEX_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/futex_waiter.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/futex_waiter.h
deleted file mode 100644
index 11dfa93b..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/futex_waiter.h
+++ /dev/null
@@ -1,63 +0,0 @@
-// Copyright 2023 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_SYNCHRONIZATION_INTERNAL_FUTEX_WAITER_H_
-#define ABSL_SYNCHRONIZATION_INTERNAL_FUTEX_WAITER_H_
-
-#include <atomic>
-#include <cstdint>
-
-#include "absl/base/config.h"
-#include "absl/synchronization/internal/kernel_timeout.h"
-#include "absl/synchronization/internal/futex.h"
-#include "absl/synchronization/internal/waiter_base.h"
-
-#ifdef ABSL_INTERNAL_HAVE_FUTEX
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace synchronization_internal {
-
-#define ABSL_INTERNAL_HAVE_FUTEX_WAITER 1
-
-class FutexWaiter : public WaiterCrtp<FutexWaiter> {
- public:
-  FutexWaiter() : futex_(0) {}
-
-  bool Wait(KernelTimeout t);
-  void Post();
-  void Poke();
-
-  static constexpr char kName[] = "FutexWaiter";
-
- private:
-  // Atomically check that `*v == val`, and if it is, then sleep until the
-  // timeout `t` has been reached, or until woken by `Wake()`.
-  static int WaitUntil(std::atomic<int32_t>* v, int32_t val,
-                       KernelTimeout t);
-
-  // Futexes are defined by specification to be 32-bits.
-  // Thus std::atomic<int32_t> must be just an int32_t with lockfree methods.
-  std::atomic<int32_t> futex_;
-  static_assert(sizeof(int32_t) == sizeof(futex_), "Wrong size for futex");
-};
-
-}  // namespace synchronization_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_INTERNAL_HAVE_FUTEX
-
-#endif  // ABSL_SYNCHRONIZATION_INTERNAL_FUTEX_WAITER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/graphcycles.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/graphcycles.h
deleted file mode 100644
index ceba33e4..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/graphcycles.h
+++ /dev/null
@@ -1,141 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_SYNCHRONIZATION_INTERNAL_GRAPHCYCLES_H_
-#define ABSL_SYNCHRONIZATION_INTERNAL_GRAPHCYCLES_H_
-
-// GraphCycles detects the introduction of a cycle into a directed
-// graph that is being built up incrementally.
-//
-// Nodes are identified by small integers.  It is not possible to
-// record multiple edges with the same (source, destination) pair;
-// requests to add an edge where one already exists are silently
-// ignored.
-//
-// It is also not possible to introduce a cycle; an attempt to insert
-// an edge that would introduce a cycle fails and returns false.
-//
-// GraphCycles uses no internal locking; calls into it should be
-// serialized externally.
-
-// Performance considerations:
-//   Works well on sparse graphs, poorly on dense graphs.
-//   Extra information is maintained incrementally to detect cycles quickly.
-//   InsertEdge() is very fast when the edge already exists, and reasonably fast
-//   otherwise.
-//   FindPath() is linear in the size of the graph.
-// The current implementation uses O(|V|+|E|) space.
-
-#include <cstdint>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace synchronization_internal {
-
-// Opaque identifier for a graph node.
-struct GraphId {
-  uint64_t handle;
-
-  bool operator==(const GraphId& x) const { return handle == x.handle; }
-  bool operator!=(const GraphId& x) const { return handle != x.handle; }
-};
-
-// Return an invalid graph id that will never be assigned by GraphCycles.
-inline GraphId InvalidGraphId() {
-  return GraphId{0};
-}
-
-class GraphCycles {
- public:
-  GraphCycles();
-  ~GraphCycles();
-
-  // Return the id to use for ptr, assigning one if necessary.
-  // Subsequent calls with the same ptr value will return the same id
-  // until Remove().
-  GraphId GetId(void* ptr);
-
-  // Remove "ptr" from the graph.  Its corresponding node and all
-  // edges to and from it are removed.
-  void RemoveNode(void* ptr);
-
-  // Return the pointer associated with id, or nullptr if id is not
-  // currently in the graph.
-  void* Ptr(GraphId id);
-
-  // Attempt to insert an edge from source_node to dest_node.  If the
-  // edge would introduce a cycle, return false without making any
-  // changes. Otherwise add the edge and return true.
-  bool InsertEdge(GraphId source_node, GraphId dest_node);
-
-  // Remove any edge that exists from source_node to dest_node.
-  void RemoveEdge(GraphId source_node, GraphId dest_node);
-
-  // Return whether node exists in the graph.
-  bool HasNode(GraphId node);
-
-  // Return whether there is an edge directly from source_node to dest_node.
-  bool HasEdge(GraphId source_node, GraphId dest_node) const;
-
-  // Return whether dest_node is reachable from source_node
-  // by following edges.
-  bool IsReachable(GraphId source_node, GraphId dest_node) const;
-
-  // Find a path from "source" to "dest".  If such a path exists,
-  // place the nodes on the path in the array path[], and return
-  // the number of nodes on the path.  If the path is longer than
-  // max_path_len nodes, only the first max_path_len nodes are placed
-  // in path[].  The client should compare the return value with
-  // max_path_len" to see when this occurs.  If no path exists, return
-  // 0.  Any valid path stored in path[] will start with "source" and
-  // end with "dest".  There is no guarantee that the path is the
-  // shortest, but no node will appear twice in the path, except the
-  // source and destination node if they are identical; therefore, the
-  // return value is at most one greater than the number of nodes in
-  // the graph.
-  int FindPath(GraphId source, GraphId dest, int max_path_len,
-               GraphId path[]) const;
-
-  // Update the stack trace recorded for id with the current stack
-  // trace if the last time it was updated had a smaller priority
-  // than the priority passed on this call.
-  //
-  // *get_stack_trace is called to get the stack trace.
-  void UpdateStackTrace(GraphId id, int priority,
-                        int (*get_stack_trace)(void**, int));
-
-  // Set *ptr to the beginning of the array that holds the recorded
-  // stack trace for id and return the depth of the stack trace.
-  int GetStackTrace(GraphId id, void*** ptr);
-
-  // Check internal invariants. Crashes on failure, returns true on success.
-  // Expensive: should only be called from graphcycles_test.cc.
-  bool CheckInvariants() const;
-
-  // ----------------------------------------------------
-  struct Rep;
- private:
-  Rep *rep_;      // opaque representation
-  GraphCycles(const GraphCycles&) = delete;
-  GraphCycles& operator=(const GraphCycles&) = delete;
-};
-
-}  // namespace synchronization_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/kernel_timeout.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/kernel_timeout.h
deleted file mode 100644
index 06404a75..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/kernel_timeout.h
+++ /dev/null
@@ -1,178 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_SYNCHRONIZATION_INTERNAL_KERNEL_TIMEOUT_H_
-#define ABSL_SYNCHRONIZATION_INTERNAL_KERNEL_TIMEOUT_H_
-
-#ifndef _WIN32
-#include <sys/types.h>
-#endif
-
-#include <algorithm>
-#include <chrono>  // NOLINT(build/c++11)
-#include <cstdint>
-#include <ctime>
-#include <limits>
-
-#include "absl/base/config.h"
-#include "absl/base/internal/raw_logging.h"
-#include "absl/time/clock.h"
-#include "absl/time/time.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace synchronization_internal {
-
-// An optional timeout, with nanosecond granularity.
-//
-// This is a private low-level API for use by a handful of low-level
-// components. Higher-level components should build APIs based on
-// absl::Time and absl::Duration.
-class KernelTimeout {
- public:
-  // Construct an absolute timeout that should expire at `t`.
-  explicit KernelTimeout(absl::Time t);
-
-  // Construct a relative timeout that should expire after `d`.
-  explicit KernelTimeout(absl::Duration d);
-
-  // Infinite timeout.
-  constexpr KernelTimeout() : rep_(kNoTimeout) {}
-
-  // A more explicit factory for those who prefer it.
-  // Equivalent to `KernelTimeout()`.
-  static constexpr KernelTimeout Never() { return KernelTimeout(); }
-
-  // Returns true if there is a timeout that will eventually expire.
-  // Returns false if the timeout is infinite.
-  bool has_timeout() const { return rep_ != kNoTimeout; }
-
-  // If `has_timeout()` is true, returns true if the timeout was provided as an
-  // `absl::Time`. The return value is undefined if `has_timeout()` is false
-  // because all indefinite timeouts are equivalent.
-  bool is_absolute_timeout() const { return (rep_ & 1) == 0; }
-
-  // If `has_timeout()` is true, returns true if the timeout was provided as an
-  // `absl::Duration`. The return value is undefined if `has_timeout()` is false
-  // because all indefinite timeouts are equivalent.
-  bool is_relative_timeout() const { return (rep_ & 1) == 1; }
-
-  // Convert to `struct timespec` for interfaces that expect an absolute
-  // timeout. If !has_timeout() or is_relative_timeout(), attempts to convert to
-  // a reasonable absolute timeout, but callers should to test has_timeout() and
-  // is_relative_timeout() and prefer to use a more appropriate interface.
-  struct timespec MakeAbsTimespec() const;
-
-  // Convert to `struct timespec` for interfaces that expect a relative
-  // timeout. If !has_timeout() or is_absolute_timeout(), attempts to convert to
-  // a reasonable relative timeout, but callers should to test has_timeout() and
-  // is_absolute_timeout() and prefer to use a more appropriate interface. Since
-  // the return value is a relative duration, it should be recomputed by calling
-  // this method in the case of a spurious wakeup.
-  struct timespec MakeRelativeTimespec() const;
-
-#ifndef _WIN32
-  // Convert to `struct timespec` for interfaces that expect an absolute timeout
-  // on a specific clock `c`. This is similar to `MakeAbsTimespec()`, but
-  // callers usually want to use this method with `CLOCK_MONOTONIC` when
-  // relative timeouts are requested, and when the appropriate interface expects
-  // an absolute timeout relative to a specific clock (for example,
-  // pthread_cond_clockwait() or sem_clockwait()). If !has_timeout(), attempts
-  // to convert to a reasonable absolute timeout, but callers should to test
-  // has_timeout() prefer to use a more appropriate interface.
-  struct timespec MakeClockAbsoluteTimespec(clockid_t c) const;
-#endif
-
-  // Convert to unix epoch nanos for interfaces that expect an absolute timeout
-  // in nanoseconds. If !has_timeout() or is_relative_timeout(), attempts to
-  // convert to a reasonable absolute timeout, but callers should to test
-  // has_timeout() and is_relative_timeout() and prefer to use a more
-  // appropriate interface.
-  int64_t MakeAbsNanos() const;
-
-  // Converts to milliseconds from now, or INFINITE when
-  // !has_timeout(). For use by SleepConditionVariableSRW on
-  // Windows. Callers should recognize that the return value is a
-  // relative duration (it should be recomputed by calling this method
-  // in the case of a spurious wakeup).
-  // This header file may be included transitively by public header files,
-  // so we define our own DWORD and INFINITE instead of getting them from
-  // <intsafe.h> and <WinBase.h>.
-  typedef unsigned long DWord;  // NOLINT
-  DWord InMillisecondsFromNow() const;
-
-  // Convert to std::chrono::time_point for interfaces that expect an absolute
-  // timeout, like std::condition_variable::wait_until(). If !has_timeout() or
-  // is_relative_timeout(), attempts to convert to a reasonable absolute
-  // timeout, but callers should test has_timeout() and is_relative_timeout()
-  // and prefer to use a more appropriate interface.
-  std::chrono::time_point<std::chrono::system_clock> ToChronoTimePoint() const;
-
-  // Convert to std::chrono::time_point for interfaces that expect a relative
-  // timeout, like std::condition_variable::wait_for(). If !has_timeout() or
-  // is_absolute_timeout(), attempts to convert to a reasonable relative
-  // timeout, but callers should test has_timeout() and is_absolute_timeout()
-  // and prefer to use a more appropriate interface. Since the return value is a
-  // relative duration, it should be recomputed by calling this method in the
-  // case of a spurious wakeup.
-  std::chrono::nanoseconds ToChronoDuration() const;
-
-  // Returns true if steady (aka monotonic) clocks are supported by the system.
-  // This method exists because go/btm requires synchronized clocks, and
-  // thus requires we use the system (aka walltime) clock.
-  static constexpr bool SupportsSteadyClock() { return true; }
-
- private:
-  // Returns the current time, expressed as a count of nanoseconds since the
-  // epoch used by an arbitrary clock. The implementation tries to use a steady
-  // (monotonic) clock if one is available.
-  static int64_t SteadyClockNow();
-
-  // Internal representation.
-  //   - If the value is kNoTimeout, then the timeout is infinite, and
-  //     has_timeout() will return true.
-  //   - If the low bit is 0, then the high 63 bits is the number of nanoseconds
-  //     after the unix epoch.
-  //   - If the low bit is 1, then the high 63 bits is the number of nanoseconds
-  //     after the epoch used by SteadyClockNow().
-  //
-  // In all cases the time is stored as an absolute time, the only difference is
-  // the clock epoch. The use of absolute times is important since in the case
-  // of a relative timeout with a spurious wakeup, the program would have to
-  // restart the wait, and thus needs a way of recomputing the remaining time.
-  uint64_t rep_;
-
-  // Returns the number of nanoseconds stored in the internal representation.
-  // When combined with the clock epoch indicated by the low bit (which is
-  // accessed through is_absolute_timeout() and is_relative_timeout()), the
-  // return value is used to compute when the timeout should occur.
-  int64_t RawAbsNanos() const { return static_cast<int64_t>(rep_ >> 1); }
-
-  // Converts to nanoseconds from now. Since the return value is a relative
-  // duration, it should be recomputed by calling this method in the case of a
-  // spurious wakeup.
-  int64_t InNanosecondsFromNow() const;
-
-  // A value that represents no timeout (or an infinite timeout).
-  static constexpr uint64_t kNoTimeout = (std::numeric_limits<uint64_t>::max)();
-
-  // The maximum value that can be stored in the high 63 bits.
-  static constexpr int64_t kMaxNanos = (std::numeric_limits<int64_t>::max)();
-};
-
-}  // namespace synchronization_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_SYNCHRONIZATION_INTERNAL_KERNEL_TIMEOUT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/per_thread_sem.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/per_thread_sem.h
deleted file mode 100644
index 144ab3cd..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/per_thread_sem.h
+++ /dev/null
@@ -1,119 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-// PerThreadSem is a low-level synchronization primitive controlling the
-// runnability of a single thread, used internally by Mutex and CondVar.
-//
-// This is NOT a general-purpose synchronization mechanism, and should not be
-// used directly by applications.  Applications should use Mutex and CondVar.
-//
-// The semantics of PerThreadSem are the same as that of a counting semaphore.
-// Each thread maintains an abstract "count" value associated with its identity.
-
-#ifndef ABSL_SYNCHRONIZATION_INTERNAL_PER_THREAD_SEM_H_
-#define ABSL_SYNCHRONIZATION_INTERNAL_PER_THREAD_SEM_H_
-
-#include <atomic>
-
-#include "absl/base/internal/thread_identity.h"
-#include "absl/synchronization/internal/create_thread_identity.h"
-#include "absl/synchronization/internal/kernel_timeout.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-class Mutex;
-
-namespace synchronization_internal {
-
-class PerThreadSem {
- public:
-  PerThreadSem() = delete;
-  PerThreadSem(const PerThreadSem&) = delete;
-  PerThreadSem& operator=(const PerThreadSem&) = delete;
-
-  // Routine invoked periodically (once a second) by a background thread.
-  // Has no effect on user-visible state.
-  static void Tick(base_internal::ThreadIdentity* identity);
-
-  // ---------------------------------------------------------------------------
-  // Routines used by autosizing threadpools to detect when threads are
-  // blocked.  Each thread has a counter pointer, initially zero.  If non-zero,
-  // the implementation atomically increments the counter when it blocks on a
-  // semaphore, a decrements it again when it wakes.  This allows a threadpool
-  // to keep track of how many of its threads are blocked.
-  // SetThreadBlockedCounter() should be used only by threadpool
-  // implementations.  GetThreadBlockedCounter() should be used by modules that
-  // block threads; if the pointer returned is non-zero, the location should be
-  // incremented before the thread blocks, and decremented after it wakes.
-  static void SetThreadBlockedCounter(std::atomic<int> *counter);
-  static std::atomic<int> *GetThreadBlockedCounter();
-
- private:
-  // Create the PerThreadSem associated with "identity".  Initializes count=0.
-  // REQUIRES: May only be called by ThreadIdentity.
-  static inline void Init(base_internal::ThreadIdentity* identity);
-
-  // Increments "identity"'s count.
-  static inline void Post(base_internal::ThreadIdentity* identity);
-
-  // Waits until either our count > 0 or t has expired.
-  // If count > 0, decrements count and returns true.  Otherwise returns false.
-  // !t.has_timeout() => Wait(t) will return true.
-  static inline bool Wait(KernelTimeout t);
-
-  // Permitted callers.
-  friend class PerThreadSemTest;
-  friend class absl::Mutex;
-  friend void OneTimeInitThreadIdentity(absl::base_internal::ThreadIdentity*);
-};
-
-}  // namespace synchronization_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-// In some build configurations we pass --detect-odr-violations to the
-// gold linker.  This causes it to flag weak symbol overrides as ODR
-// violations.  Because ODR only applies to C++ and not C,
-// --detect-odr-violations ignores symbols not mangled with C++ names.
-// By changing our extension points to be extern "C", we dodge this
-// check.
-extern "C" {
-void ABSL_INTERNAL_C_SYMBOL(AbslInternalPerThreadSemInit)(
-    absl::base_internal::ThreadIdentity* identity);
-void ABSL_INTERNAL_C_SYMBOL(AbslInternalPerThreadSemPost)(
-    absl::base_internal::ThreadIdentity* identity);
-bool ABSL_INTERNAL_C_SYMBOL(AbslInternalPerThreadSemWait)(
-    absl::synchronization_internal::KernelTimeout t);
-void ABSL_INTERNAL_C_SYMBOL(AbslInternalPerThreadSemPoke)(
-    absl::base_internal::ThreadIdentity* identity);
-}  // extern "C"
-
-void absl::synchronization_internal::PerThreadSem::Init(
-    absl::base_internal::ThreadIdentity* identity) {
-  ABSL_INTERNAL_C_SYMBOL(AbslInternalPerThreadSemInit)(identity);
-}
-
-void absl::synchronization_internal::PerThreadSem::Post(
-    absl::base_internal::ThreadIdentity* identity) {
-  ABSL_INTERNAL_C_SYMBOL(AbslInternalPerThreadSemPost)(identity);
-}
-
-bool absl::synchronization_internal::PerThreadSem::Wait(
-    absl::synchronization_internal::KernelTimeout t) {
-  return ABSL_INTERNAL_C_SYMBOL(AbslInternalPerThreadSemWait)(t);
-}
-
-#endif  // ABSL_SYNCHRONIZATION_INTERNAL_PER_THREAD_SEM_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/pthread_waiter.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/pthread_waiter.h
deleted file mode 100644
index 23db76ad..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/pthread_waiter.h
+++ /dev/null
@@ -1,60 +0,0 @@
-// Copyright 2023 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_SYNCHRONIZATION_INTERNAL_PTHREAD_WAITER_H_
-#define ABSL_SYNCHRONIZATION_INTERNAL_PTHREAD_WAITER_H_
-
-#if !defined(_WIN32) && !defined(__MINGW32__)
-#include <pthread.h>
-
-#include "absl/base/config.h"
-#include "absl/synchronization/internal/kernel_timeout.h"
-#include "absl/synchronization/internal/waiter_base.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace synchronization_internal {
-
-#define ABSL_INTERNAL_HAVE_PTHREAD_WAITER 1
-
-class PthreadWaiter : public WaiterCrtp<PthreadWaiter> {
- public:
-  PthreadWaiter();
-
-  bool Wait(KernelTimeout t);
-  void Post();
-  void Poke();
-
-  static constexpr char kName[] = "PthreadWaiter";
-
- private:
-  int TimedWait(KernelTimeout t);
-
-  // REQUIRES: mu_ must be held.
-  void InternalCondVarPoke();
-
-  pthread_mutex_t mu_;
-  pthread_cond_t cv_;
-  int waiter_count_;
-  int wakeup_count_;  // Unclaimed wakeups.
-};
-
-}  // namespace synchronization_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // !defined(_WIN32) && !defined(__MINGW32__)
-
-#endif  // ABSL_SYNCHRONIZATION_INTERNAL_PTHREAD_WAITER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/sem_waiter.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/sem_waiter.h
deleted file mode 100644
index c22746f9..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/sem_waiter.h
+++ /dev/null
@@ -1,65 +0,0 @@
-// Copyright 2023 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_SYNCHRONIZATION_INTERNAL_SEM_WAITER_H_
-#define ABSL_SYNCHRONIZATION_INTERNAL_SEM_WAITER_H_
-
-#include "absl/base/config.h"
-
-#ifdef ABSL_HAVE_SEMAPHORE_H
-#include <semaphore.h>
-
-#include <atomic>
-#include <cstdint>
-
-#include "absl/base/internal/thread_identity.h"
-#include "absl/synchronization/internal/futex.h"
-#include "absl/synchronization/internal/kernel_timeout.h"
-#include "absl/synchronization/internal/waiter_base.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace synchronization_internal {
-
-#define ABSL_INTERNAL_HAVE_SEM_WAITER 1
-
-class SemWaiter : public WaiterCrtp<SemWaiter> {
- public:
-  SemWaiter();
-
-  bool Wait(KernelTimeout t);
-  void Post();
-  void Poke();
-
-  static constexpr char kName[] = "SemWaiter";
-
- private:
-  int TimedWait(KernelTimeout t);
-
-  sem_t sem_;
-
-  // This seems superfluous, but for Poke() we need to cause spurious
-  // wakeups on the semaphore. Hence we can't actually use the
-  // semaphore's count.
-  std::atomic<int> wakeups_;
-};
-
-}  // namespace synchronization_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_HAVE_SEMAPHORE_H
-
-#endif  // ABSL_SYNCHRONIZATION_INTERNAL_SEM_WAITER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/stdcpp_waiter.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/stdcpp_waiter.h
deleted file mode 100644
index e592a27b..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/stdcpp_waiter.h
+++ /dev/null
@@ -1,56 +0,0 @@
-// Copyright 2023 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_SYNCHRONIZATION_INTERNAL_STDCPP_WAITER_H_
-#define ABSL_SYNCHRONIZATION_INTERNAL_STDCPP_WAITER_H_
-
-#include <condition_variable>  // NOLINT(build/c++11)
-#include <mutex>  // NOLINT(build/c++11)
-
-#include "absl/base/config.h"
-#include "absl/synchronization/internal/kernel_timeout.h"
-#include "absl/synchronization/internal/waiter_base.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace synchronization_internal {
-
-#define ABSL_INTERNAL_HAVE_STDCPP_WAITER 1
-
-class StdcppWaiter : public WaiterCrtp<StdcppWaiter> {
- public:
-  StdcppWaiter();
-
-  bool Wait(KernelTimeout t);
-  void Post();
-  void Poke();
-
-  static constexpr char kName[] = "StdcppWaiter";
-
- private:
-  // REQUIRES: mu_ must be held.
-  void InternalCondVarPoke();
-
-  std::mutex mu_;
-  std::condition_variable cv_;
-  int waiter_count_;
-  int wakeup_count_;  // Unclaimed wakeups.
-};
-
-}  // namespace synchronization_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_SYNCHRONIZATION_INTERNAL_STDCPP_WAITER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/thread_pool.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/thread_pool.h
deleted file mode 100644
index 5eb0bb60..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/thread_pool.h
+++ /dev/null
@@ -1,96 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_SYNCHRONIZATION_INTERNAL_THREAD_POOL_H_
-#define ABSL_SYNCHRONIZATION_INTERNAL_THREAD_POOL_H_
-
-#include <cassert>
-#include <cstddef>
-#include <functional>
-#include <queue>
-#include <thread>  // NOLINT(build/c++11)
-#include <utility>
-#include <vector>
-
-#include "absl/base/thread_annotations.h"
-#include "absl/functional/any_invocable.h"
-#include "absl/synchronization/mutex.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace synchronization_internal {
-
-// A simple ThreadPool implementation for tests.
-class ThreadPool {
- public:
-  explicit ThreadPool(int num_threads) {
-    threads_.reserve(num_threads);
-    for (int i = 0; i < num_threads; ++i) {
-      threads_.push_back(std::thread(&ThreadPool::WorkLoop, this));
-    }
-  }
-
-  ThreadPool(const ThreadPool &) = delete;
-  ThreadPool &operator=(const ThreadPool &) = delete;
-
-  ~ThreadPool() {
-    {
-      absl::MutexLock l(&mu_);
-      for (size_t i = 0; i < threads_.size(); i++) {
-        queue_.push(nullptr);  // Shutdown signal.
-      }
-    }
-    for (auto &t : threads_) {
-      t.join();
-    }
-  }
-
-  // Schedule a function to be run on a ThreadPool thread immediately.
-  void Schedule(absl::AnyInvocable<void()> func) {
-    assert(func != nullptr);
-    absl::MutexLock l(&mu_);
-    queue_.push(std::move(func));
-  }
-
- private:
-  bool WorkAvailable() const ABSL_EXCLUSIVE_LOCKS_REQUIRED(mu_) {
-    return !queue_.empty();
-  }
-
-  void WorkLoop() {
-    while (true) {
-      absl::AnyInvocable<void()> func;
-      {
-        absl::MutexLock l(&mu_);
-        mu_.Await(absl::Condition(this, &ThreadPool::WorkAvailable));
-        func = std::move(queue_.front());
-        queue_.pop();
-      }
-      if (func == nullptr) {  // Shutdown signal.
-        break;
-      }
-      func();
-    }
-  }
-
-  absl::Mutex mu_;
-  std::queue<absl::AnyInvocable<void()>> queue_ ABSL_GUARDED_BY(mu_);
-  std::vector<std::thread> threads_;
-};
-
-}  // namespace synchronization_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_SYNCHRONIZATION_INTERNAL_THREAD_POOL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/waiter.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/waiter.h
deleted file mode 100644
index 6ba204be..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/waiter.h
+++ /dev/null
@@ -1,69 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_SYNCHRONIZATION_INTERNAL_WAITER_H_
-#define ABSL_SYNCHRONIZATION_INTERNAL_WAITER_H_
-
-#include "absl/base/config.h"
-#include "absl/synchronization/internal/futex_waiter.h"
-#include "absl/synchronization/internal/pthread_waiter.h"
-#include "absl/synchronization/internal/sem_waiter.h"
-#include "absl/synchronization/internal/stdcpp_waiter.h"
-#include "absl/synchronization/internal/win32_waiter.h"
-
-// May be chosen at compile time via -DABSL_FORCE_WAITER_MODE=<index>
-#define ABSL_WAITER_MODE_FUTEX 0
-#define ABSL_WAITER_MODE_SEM 1
-#define ABSL_WAITER_MODE_CONDVAR 2
-#define ABSL_WAITER_MODE_WIN32 3
-#define ABSL_WAITER_MODE_STDCPP 4
-
-#if defined(ABSL_FORCE_WAITER_MODE)
-#define ABSL_WAITER_MODE ABSL_FORCE_WAITER_MODE
-#elif defined(ABSL_INTERNAL_HAVE_WIN32_WAITER)
-#define ABSL_WAITER_MODE ABSL_WAITER_MODE_WIN32
-#elif defined(ABSL_INTERNAL_HAVE_FUTEX_WAITER)
-#define ABSL_WAITER_MODE ABSL_WAITER_MODE_FUTEX
-#elif defined(ABSL_INTERNAL_HAVE_SEM_WAITER)
-#define ABSL_WAITER_MODE ABSL_WAITER_MODE_SEM
-#elif defined(ABSL_INTERNAL_HAVE_PTHREAD_WAITER)
-#define ABSL_WAITER_MODE ABSL_WAITER_MODE_CONDVAR
-#elif defined(ABSL_INTERNAL_HAVE_STDCPP_WAITER)
-#define ABSL_WAITER_MODE ABSL_WAITER_MODE_STDCPP
-#else
-#error ABSL_WAITER_MODE is undefined
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace synchronization_internal {
-
-#if ABSL_WAITER_MODE == ABSL_WAITER_MODE_FUTEX
-using Waiter = FutexWaiter;
-#elif ABSL_WAITER_MODE == ABSL_WAITER_MODE_SEM
-using Waiter = SemWaiter;
-#elif ABSL_WAITER_MODE == ABSL_WAITER_MODE_CONDVAR
-using Waiter = PthreadWaiter;
-#elif ABSL_WAITER_MODE == ABSL_WAITER_MODE_WIN32
-using Waiter = Win32Waiter;
-#elif ABSL_WAITER_MODE == ABSL_WAITER_MODE_STDCPP
-using Waiter = StdcppWaiter;
-#endif
-
-}  // namespace synchronization_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_SYNCHRONIZATION_INTERNAL_WAITER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/waiter_base.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/waiter_base.h
deleted file mode 100644
index cf175481..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/waiter_base.h
+++ /dev/null
@@ -1,90 +0,0 @@
-// Copyright 2023 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_SYNCHRONIZATION_INTERNAL_WAITER_BASE_H_
-#define ABSL_SYNCHRONIZATION_INTERNAL_WAITER_BASE_H_
-
-#include "absl/base/config.h"
-#include "absl/base/internal/thread_identity.h"
-#include "absl/synchronization/internal/kernel_timeout.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace synchronization_internal {
-
-// `Waiter` is a platform specific semaphore implementation that `PerThreadSem`
-// waits on to implement blocking in `absl::Mutex`.  Implementations should
-// inherit from `WaiterCrtp` and must implement `Wait()`, `Post()`, and `Poke()`
-// as described in `WaiterBase`.  `waiter.h` selects the implementation and uses
-// static-dispatch for performance.
-class WaiterBase {
- public:
-  WaiterBase() = default;
-
-  // Not copyable or movable
-  WaiterBase(const WaiterBase&) = delete;
-  WaiterBase& operator=(const WaiterBase&) = delete;
-
-  // Blocks the calling thread until a matching call to `Post()` or
-  // `t` has passed. Returns `true` if woken (`Post()` called),
-  // `false` on timeout.
-  //
-  // bool Wait(KernelTimeout t);
-
-  // Restart the caller of `Wait()` as with a normal semaphore.
-  //
-  // void Post();
-
-  // If anyone is waiting, wake them up temporarily and cause them to
-  // call `MaybeBecomeIdle()`. They will then return to waiting for a
-  // `Post()` or timeout.
-  //
-  // void Poke();
-
-  // Returns the name of this implementation. Used only for debugging.
-  //
-  // static constexpr char kName[];
-
-  // How many periods to remain idle before releasing resources
-#ifndef ABSL_HAVE_THREAD_SANITIZER
-  static constexpr int kIdlePeriods = 60;
-#else
-  // Memory consumption under ThreadSanitizer is a serious concern,
-  // so we release resources sooner. The value of 1 leads to 1 to 2 second
-  // delay before marking a thread as idle.
-  static constexpr int kIdlePeriods = 1;
-#endif
-
- protected:
-  static void MaybeBecomeIdle();
-};
-
-template <typename T>
-class WaiterCrtp : public WaiterBase {
- public:
-  // Returns the Waiter associated with the identity.
-  static T* GetWaiter(base_internal::ThreadIdentity* identity) {
-    static_assert(
-        sizeof(T) <= sizeof(base_internal::ThreadIdentity::WaiterState),
-        "Insufficient space for Waiter");
-    return reinterpret_cast<T*>(identity->waiter_state.data);
-  }
-};
-
-}  // namespace synchronization_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_SYNCHRONIZATION_INTERNAL_WAITER_BASE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/win32_waiter.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/win32_waiter.h
deleted file mode 100644
index fdab264e..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/internal/win32_waiter.h
+++ /dev/null
@@ -1,72 +0,0 @@
-// Copyright 2023 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-
-#ifndef ABSL_SYNCHRONIZATION_INTERNAL_WIN32_WAITER_H_
-#define ABSL_SYNCHRONIZATION_INTERNAL_WIN32_WAITER_H_
-
-#ifdef _WIN32
-#include <sdkddkver.h>
-#endif
-
-#if defined(_WIN32) && !defined(__MINGW32__) && \
-    _WIN32_WINNT >= _WIN32_WINNT_VISTA
-
-#include "absl/base/config.h"
-#include "absl/synchronization/internal/kernel_timeout.h"
-#include "absl/synchronization/internal/waiter_base.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace synchronization_internal {
-
-#define ABSL_INTERNAL_HAVE_WIN32_WAITER 1
-
-class Win32Waiter : public WaiterCrtp<Win32Waiter> {
- public:
-  Win32Waiter();
-
-  bool Wait(KernelTimeout t);
-  void Post();
-  void Poke();
-
-  static constexpr char kName[] = "Win32Waiter";
-
- private:
-  // WinHelper - Used to define utilities for accessing the lock and
-  // condition variable storage once the types are complete.
-  class WinHelper;
-
-  // REQUIRES: WinHelper::GetLock(this) must be held.
-  void InternalCondVarPoke();
-
-  // We can't include Windows.h in our headers, so we use aligned character
-  // buffers to define the storage of SRWLOCK and CONDITION_VARIABLE.
-  // SRW locks and condition variables do not need to be explicitly destroyed.
-  // https://docs.microsoft.com/en-us/windows/win32/api/synchapi/nf-synchapi-initializesrwlock
-  // https://stackoverflow.com/questions/28975958/why-does-windows-have-no-deleteconditionvariable-function-to-go-together-with
-  alignas(void*) unsigned char mu_storage_[sizeof(void*)];
-  alignas(void*) unsigned char cv_storage_[sizeof(void*)];
-  int waiter_count_;
-  int wakeup_count_;
-};
-
-}  // namespace synchronization_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // defined(_WIN32) && !defined(__MINGW32__) &&
-        // _WIN32_WINNT >= _WIN32_WINNT_VISTA
-
-#endif  // ABSL_SYNCHRONIZATION_INTERNAL_WIN32_WAITER_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/mutex.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/mutex.h
deleted file mode 100644
index d53a22bb..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/mutex.h
+++ /dev/null
@@ -1,1218 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// mutex.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines a `Mutex` -- a mutually exclusive lock -- and the
-// most common type of synchronization primitive for facilitating locks on
-// shared resources. A mutex is used to prevent multiple threads from accessing
-// and/or writing to a shared resource concurrently.
-//
-// Unlike a `std::mutex`, the Abseil `Mutex` provides the following additional
-// features:
-//   * Conditional predicates intrinsic to the `Mutex` object
-//   * Shared/reader locks, in addition to standard exclusive/writer locks
-//   * Deadlock detection and debug support.
-//
-// The following helper classes are also defined within this file:
-//
-//  MutexLock - An RAII wrapper to acquire and release a `Mutex` for exclusive/
-//              write access within the current scope.
-//
-//  ReaderMutexLock
-//            - An RAII wrapper to acquire and release a `Mutex` for shared/read
-//              access within the current scope.
-//
-//  WriterMutexLock
-//            - Effectively an alias for `MutexLock` above, designed for use in
-//              distinguishing reader and writer locks within code.
-//
-// In addition to simple mutex locks, this file also defines ways to perform
-// locking under certain conditions.
-//
-//  Condition - (Preferred) Used to wait for a particular predicate that
-//              depends on state protected by the `Mutex` to become true.
-//  CondVar   - A lower-level variant of `Condition` that relies on
-//              application code to explicitly signal the `CondVar` when
-//              a condition has been met.
-//
-// See below for more information on using `Condition` or `CondVar`.
-//
-// Mutexes and mutex behavior can be quite complicated. The information within
-// this header file is limited, as a result. Please consult the Mutex guide for
-// more complete information and examples.
-
-#ifndef ABSL_SYNCHRONIZATION_MUTEX_H_
-#define ABSL_SYNCHRONIZATION_MUTEX_H_
-
-#include <atomic>
-#include <cstdint>
-#include <cstring>
-#include <iterator>
-#include <string>
-
-#include "absl/base/attributes.h"
-#include "absl/base/const_init.h"
-#include "absl/base/internal/identity.h"
-#include "absl/base/internal/low_level_alloc.h"
-#include "absl/base/internal/thread_identity.h"
-#include "absl/base/internal/tsan_mutex_interface.h"
-#include "absl/base/port.h"
-#include "absl/base/thread_annotations.h"
-#include "absl/synchronization/internal/kernel_timeout.h"
-#include "absl/synchronization/internal/per_thread_sem.h"
-#include "absl/time/time.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-class Condition;
-struct SynchWaitParams;
-
-// -----------------------------------------------------------------------------
-// Mutex
-// -----------------------------------------------------------------------------
-//
-// A `Mutex` is a non-reentrant (aka non-recursive) Mutually Exclusive lock
-// on some resource, typically a variable or data structure with associated
-// invariants. Proper usage of mutexes prevents concurrent access by different
-// threads to the same resource.
-//
-// A `Mutex` has two basic operations: `Mutex::Lock()` and `Mutex::Unlock()`.
-// The `Lock()` operation *acquires* a `Mutex` (in a state known as an
-// *exclusive* -- or *write* -- lock), and the `Unlock()` operation *releases* a
-// Mutex. During the span of time between the Lock() and Unlock() operations,
-// a mutex is said to be *held*. By design, all mutexes support exclusive/write
-// locks, as this is the most common way to use a mutex.
-//
-// Mutex operations are only allowed under certain conditions; otherwise an
-// operation is "invalid", and disallowed by the API. The conditions concern
-// both the current state of the mutex and the identity of the threads that
-// are performing the operations.
-//
-// The `Mutex` state machine for basic lock/unlock operations is quite simple:
-//
-// |                | Lock()                 | Unlock() |
-// |----------------+------------------------+----------|
-// | Free           | Exclusive              | invalid  |
-// | Exclusive      | blocks, then exclusive | Free     |
-//
-// The full conditions are as follows.
-//
-// * Calls to `Unlock()` require that the mutex be held, and must be made in the
-//   same thread that performed the corresponding `Lock()` operation which
-//   acquired the mutex; otherwise the call is invalid.
-//
-// * The mutex being non-reentrant (or non-recursive) means that a call to
-//   `Lock()` or `TryLock()` must not be made in a thread that already holds the
-//   mutex; such a call is invalid.
-//
-// * In other words, the state of being "held" has both a temporal component
-//   (from `Lock()` until `Unlock()`) as well as a thread identity component:
-//   the mutex is held *by a particular thread*.
-//
-// An "invalid" operation has undefined behavior. The `Mutex` implementation
-// is allowed to do anything on an invalid call, including, but not limited to,
-// crashing with a useful error message, silently succeeding, or corrupting
-// data structures. In debug mode, the implementation may crash with a useful
-// error message.
-//
-// `Mutex` is not guaranteed to be "fair" in prioritizing waiting threads; it
-// is, however, approximately fair over long periods, and starvation-free for
-// threads at the same priority.
-//
-// The lock/unlock primitives are now annotated with lock annotations
-// defined in (base/thread_annotations.h). When writing multi-threaded code,
-// you should use lock annotations whenever possible to document your lock
-// synchronization policy. Besides acting as documentation, these annotations
-// also help compilers or static analysis tools to identify and warn about
-// issues that could potentially result in race conditions and deadlocks.
-//
-// For more information about the lock annotations, please see
-// [Thread Safety
-// Analysis](http://clang.llvm.org/docs/ThreadSafetyAnalysis.html) in the Clang
-// documentation.
-//
-// See also `MutexLock`, below, for scoped `Mutex` acquisition.
-
-class ABSL_LOCKABLE Mutex {
- public:
-  // Creates a `Mutex` that is not held by anyone. This constructor is
-  // typically used for Mutexes allocated on the heap or the stack.
-  //
-  // To create `Mutex` instances with static storage duration
-  // (e.g. a namespace-scoped or global variable), see
-  // `Mutex::Mutex(absl::kConstInit)` below instead.
-  Mutex();
-
-  // Creates a mutex with static storage duration.  A global variable
-  // constructed this way avoids the lifetime issues that can occur on program
-  // startup and shutdown.  (See absl/base/const_init.h.)
-  //
-  // For Mutexes allocated on the heap and stack, instead use the default
-  // constructor, which can interact more fully with the thread sanitizer.
-  //
-  // Example usage:
-  //   namespace foo {
-  //   ABSL_CONST_INIT absl::Mutex mu(absl::kConstInit);
-  //   }
-  explicit constexpr Mutex(absl::ConstInitType);
-
-  ~Mutex();
-
-  // Mutex::Lock()
-  //
-  // Blocks the calling thread, if necessary, until this `Mutex` is free, and
-  // then acquires it exclusively. (This lock is also known as a "write lock.")
-  void Lock() ABSL_EXCLUSIVE_LOCK_FUNCTION();
-
-  // Mutex::Unlock()
-  //
-  // Releases this `Mutex` and returns it from the exclusive/write state to the
-  // free state. Calling thread must hold the `Mutex` exclusively.
-  void Unlock() ABSL_UNLOCK_FUNCTION();
-
-  // Mutex::TryLock()
-  //
-  // If the mutex can be acquired without blocking, does so exclusively and
-  // returns `true`. Otherwise, returns `false`. Returns `true` with high
-  // probability if the `Mutex` was free.
-  bool TryLock() ABSL_EXCLUSIVE_TRYLOCK_FUNCTION(true);
-
-  // Mutex::AssertHeld()
-  //
-  // Require that the mutex be held exclusively (write mode) by this thread.
-  //
-  // If the mutex is not currently held by this thread, this function may report
-  // an error (typically by crashing with a diagnostic) or it may do nothing.
-  // This function is intended only as a tool to assist debugging; it doesn't
-  // guarantee correctness.
-  void AssertHeld() const ABSL_ASSERT_EXCLUSIVE_LOCK();
-
-  // ---------------------------------------------------------------------------
-  // Reader-Writer Locking
-  // ---------------------------------------------------------------------------
-
-  // A Mutex can also be used as a starvation-free reader-writer lock.
-  // Neither read-locks nor write-locks are reentrant/recursive to avoid
-  // potential client programming errors.
-  //
-  // The Mutex API provides `Writer*()` aliases for the existing `Lock()`,
-  // `Unlock()` and `TryLock()` methods for use within applications mixing
-  // reader/writer locks. Using `Reader*()` and `Writer*()` operations in this
-  // manner can make locking behavior clearer when mixing read and write modes.
-  //
-  // Introducing reader locks necessarily complicates the `Mutex` state
-  // machine somewhat. The table below illustrates the allowed state transitions
-  // of a mutex in such cases. Note that ReaderLock() may block even if the lock
-  // is held in shared mode; this occurs when another thread is blocked on a
-  // call to WriterLock().
-  //
-  // ---------------------------------------------------------------------------
-  //     Operation: WriterLock() Unlock()  ReaderLock()           ReaderUnlock()
-  // ---------------------------------------------------------------------------
-  // State
-  // ---------------------------------------------------------------------------
-  // Free           Exclusive    invalid   Shared(1)              invalid
-  // Shared(1)      blocks       invalid   Shared(2) or blocks    Free
-  // Shared(n) n>1  blocks       invalid   Shared(n+1) or blocks  Shared(n-1)
-  // Exclusive      blocks       Free      blocks                 invalid
-  // ---------------------------------------------------------------------------
-  //
-  // In comments below, "shared" refers to a state of Shared(n) for any n > 0.
-
-  // Mutex::ReaderLock()
-  //
-  // Blocks the calling thread, if necessary, until this `Mutex` is either free,
-  // or in shared mode, and then acquires a share of it. Note that
-  // `ReaderLock()` will block if some other thread has an exclusive/writer lock
-  // on the mutex.
-
-  void ReaderLock() ABSL_SHARED_LOCK_FUNCTION();
-
-  // Mutex::ReaderUnlock()
-  //
-  // Releases a read share of this `Mutex`. `ReaderUnlock` may return a mutex to
-  // the free state if this thread holds the last reader lock on the mutex. Note
-  // that you cannot call `ReaderUnlock()` on a mutex held in write mode.
-  void ReaderUnlock() ABSL_UNLOCK_FUNCTION();
-
-  // Mutex::ReaderTryLock()
-  //
-  // If the mutex can be acquired without blocking, acquires this mutex for
-  // shared access and returns `true`. Otherwise, returns `false`. Returns
-  // `true` with high probability if the `Mutex` was free or shared.
-  bool ReaderTryLock() ABSL_SHARED_TRYLOCK_FUNCTION(true);
-
-  // Mutex::AssertReaderHeld()
-  //
-  // Require that the mutex be held at least in shared mode (read mode) by this
-  // thread.
-  //
-  // If the mutex is not currently held by this thread, this function may report
-  // an error (typically by crashing with a diagnostic) or it may do nothing.
-  // This function is intended only as a tool to assist debugging; it doesn't
-  // guarantee correctness.
-  void AssertReaderHeld() const ABSL_ASSERT_SHARED_LOCK();
-
-  // Mutex::WriterLock()
-  // Mutex::WriterUnlock()
-  // Mutex::WriterTryLock()
-  //
-  // Aliases for `Mutex::Lock()`, `Mutex::Unlock()`, and `Mutex::TryLock()`.
-  //
-  // These methods may be used (along with the complementary `Reader*()`
-  // methods) to distinguish simple exclusive `Mutex` usage (`Lock()`,
-  // etc.) from reader/writer lock usage.
-  void WriterLock() ABSL_EXCLUSIVE_LOCK_FUNCTION() { this->Lock(); }
-
-  void WriterUnlock() ABSL_UNLOCK_FUNCTION() { this->Unlock(); }
-
-  bool WriterTryLock() ABSL_EXCLUSIVE_TRYLOCK_FUNCTION(true) {
-    return this->TryLock();
-  }
-
-  // ---------------------------------------------------------------------------
-  // Conditional Critical Regions
-  // ---------------------------------------------------------------------------
-
-  // Conditional usage of a `Mutex` can occur using two distinct paradigms:
-  //
-  //   * Use of `Mutex` member functions with `Condition` objects.
-  //   * Use of the separate `CondVar` abstraction.
-  //
-  // In general, prefer use of `Condition` and the `Mutex` member functions
-  // listed below over `CondVar`. When there are multiple threads waiting on
-  // distinctly different conditions, however, a battery of `CondVar`s may be
-  // more efficient. This section discusses use of `Condition` objects.
-  //
-  // `Mutex` contains member functions for performing lock operations only under
-  // certain conditions, of class `Condition`. For correctness, the `Condition`
-  // must return a boolean that is a pure function, only of state protected by
-  // the `Mutex`. The condition must be invariant w.r.t. environmental state
-  // such as thread, cpu id, or time, and must be `noexcept`. The condition will
-  // always be invoked with the mutex held in at least read mode, so you should
-  // not block it for long periods or sleep it on a timer.
-  //
-  // Since a condition must not depend directly on the current time, use
-  // `*WithTimeout()` member function variants to make your condition
-  // effectively true after a given duration, or `*WithDeadline()` variants to
-  // make your condition effectively true after a given time.
-  //
-  // The condition function should have no side-effects aside from debug
-  // logging; as a special exception, the function may acquire other mutexes
-  // provided it releases all those that it acquires.  (This exception was
-  // required to allow logging.)
-
-  // Mutex::Await()
-  //
-  // Unlocks this `Mutex` and blocks until simultaneously both `cond` is `true`
-  // and this `Mutex` can be reacquired, then reacquires this `Mutex` in the
-  // same mode in which it was previously held. If the condition is initially
-  // `true`, `Await()` *may* skip the release/re-acquire step.
-  //
-  // `Await()` requires that this thread holds this `Mutex` in some mode.
-  void Await(const Condition& cond) {
-    AwaitCommon(cond, synchronization_internal::KernelTimeout::Never());
-  }
-
-  // Mutex::LockWhen()
-  // Mutex::ReaderLockWhen()
-  // Mutex::WriterLockWhen()
-  //
-  // Blocks until simultaneously both `cond` is `true` and this `Mutex` can
-  // be acquired, then atomically acquires this `Mutex`. `LockWhen()` is
-  // logically equivalent to `*Lock(); Await();` though they may have different
-  // performance characteristics.
-  void LockWhen(const Condition& cond) ABSL_EXCLUSIVE_LOCK_FUNCTION() {
-    LockWhenCommon(cond, synchronization_internal::KernelTimeout::Never(),
-                   true);
-  }
-
-  void ReaderLockWhen(const Condition& cond) ABSL_SHARED_LOCK_FUNCTION() {
-    LockWhenCommon(cond, synchronization_internal::KernelTimeout::Never(),
-                   false);
-  }
-
-  void WriterLockWhen(const Condition& cond) ABSL_EXCLUSIVE_LOCK_FUNCTION() {
-    this->LockWhen(cond);
-  }
-
-  // ---------------------------------------------------------------------------
-  // Mutex Variants with Timeouts/Deadlines
-  // ---------------------------------------------------------------------------
-
-  // Mutex::AwaitWithTimeout()
-  // Mutex::AwaitWithDeadline()
-  //
-  // Unlocks this `Mutex` and blocks until simultaneously:
-  //   - either `cond` is true or the {timeout has expired, deadline has passed}
-  //     and
-  //   - this `Mutex` can be reacquired,
-  // then reacquire this `Mutex` in the same mode in which it was previously
-  // held, returning `true` iff `cond` is `true` on return.
-  //
-  // If the condition is initially `true`, the implementation *may* skip the
-  // release/re-acquire step and return immediately.
-  //
-  // Deadlines in the past are equivalent to an immediate deadline.
-  // Negative timeouts are equivalent to a zero timeout.
-  //
-  // This method requires that this thread holds this `Mutex` in some mode.
-  bool AwaitWithTimeout(const Condition& cond, absl::Duration timeout) {
-    return AwaitCommon(cond, synchronization_internal::KernelTimeout{timeout});
-  }
-
-  bool AwaitWithDeadline(const Condition& cond, absl::Time deadline) {
-    return AwaitCommon(cond, synchronization_internal::KernelTimeout{deadline});
-  }
-
-  // Mutex::LockWhenWithTimeout()
-  // Mutex::ReaderLockWhenWithTimeout()
-  // Mutex::WriterLockWhenWithTimeout()
-  //
-  // Blocks until simultaneously both:
-  //   - either `cond` is `true` or the timeout has expired, and
-  //   - this `Mutex` can be acquired,
-  // then atomically acquires this `Mutex`, returning `true` iff `cond` is
-  // `true` on return.
-  //
-  // Negative timeouts are equivalent to a zero timeout.
-  bool LockWhenWithTimeout(const Condition& cond, absl::Duration timeout)
-      ABSL_EXCLUSIVE_LOCK_FUNCTION() {
-    return LockWhenCommon(
-        cond, synchronization_internal::KernelTimeout{timeout}, true);
-  }
-  bool ReaderLockWhenWithTimeout(const Condition& cond, absl::Duration timeout)
-      ABSL_SHARED_LOCK_FUNCTION() {
-    return LockWhenCommon(
-        cond, synchronization_internal::KernelTimeout{timeout}, false);
-  }
-  bool WriterLockWhenWithTimeout(const Condition& cond, absl::Duration timeout)
-      ABSL_EXCLUSIVE_LOCK_FUNCTION() {
-    return this->LockWhenWithTimeout(cond, timeout);
-  }
-
-  // Mutex::LockWhenWithDeadline()
-  // Mutex::ReaderLockWhenWithDeadline()
-  // Mutex::WriterLockWhenWithDeadline()
-  //
-  // Blocks until simultaneously both:
-  //   - either `cond` is `true` or the deadline has been passed, and
-  //   - this `Mutex` can be acquired,
-  // then atomically acquires this Mutex, returning `true` iff `cond` is `true`
-  // on return.
-  //
-  // Deadlines in the past are equivalent to an immediate deadline.
-  bool LockWhenWithDeadline(const Condition& cond, absl::Time deadline)
-      ABSL_EXCLUSIVE_LOCK_FUNCTION() {
-    return LockWhenCommon(
-        cond, synchronization_internal::KernelTimeout{deadline}, true);
-  }
-  bool ReaderLockWhenWithDeadline(const Condition& cond, absl::Time deadline)
-      ABSL_SHARED_LOCK_FUNCTION() {
-    return LockWhenCommon(
-        cond, synchronization_internal::KernelTimeout{deadline}, false);
-  }
-  bool WriterLockWhenWithDeadline(const Condition& cond, absl::Time deadline)
-      ABSL_EXCLUSIVE_LOCK_FUNCTION() {
-    return this->LockWhenWithDeadline(cond, deadline);
-  }
-
-  // ---------------------------------------------------------------------------
-  // Debug Support: Invariant Checking, Deadlock Detection, Logging.
-  // ---------------------------------------------------------------------------
-
-  // Mutex::EnableInvariantDebugging()
-  //
-  // If `invariant`!=null and if invariant debugging has been enabled globally,
-  // cause `(*invariant)(arg)` to be called at moments when the invariant for
-  // this `Mutex` should hold (for example: just after acquire, just before
-  // release).
-  //
-  // The routine `invariant` should have no side-effects since it is not
-  // guaranteed how many times it will be called; it should check the invariant
-  // and crash if it does not hold. Enabling global invariant debugging may
-  // substantially reduce `Mutex` performance; it should be set only for
-  // non-production runs.  Optimization options may also disable invariant
-  // checks.
-  void EnableInvariantDebugging(void (*invariant)(void*), void* arg);
-
-  // Mutex::EnableDebugLog()
-  //
-  // Cause all subsequent uses of this `Mutex` to be logged via
-  // `ABSL_RAW_LOG(INFO)`. Log entries are tagged with `name` if no previous
-  // call to `EnableInvariantDebugging()` or `EnableDebugLog()` has been made.
-  //
-  // Note: This method substantially reduces `Mutex` performance.
-  void EnableDebugLog(const char* name);
-
-  // Deadlock detection
-
-  // Mutex::ForgetDeadlockInfo()
-  //
-  // Forget any deadlock-detection information previously gathered
-  // about this `Mutex`. Call this method in debug mode when the lock ordering
-  // of a `Mutex` changes.
-  void ForgetDeadlockInfo();
-
-  // Mutex::AssertNotHeld()
-  //
-  // Return immediately if this thread does not hold this `Mutex` in any
-  // mode; otherwise, may report an error (typically by crashing with a
-  // diagnostic), or may return immediately.
-  //
-  // Currently this check is performed only if all of:
-  //    - in debug mode
-  //    - SetMutexDeadlockDetectionMode() has been set to kReport or kAbort
-  //    - number of locks concurrently held by this thread is not large.
-  // are true.
-  void AssertNotHeld() const;
-
-  // Special cases.
-
-  // A `MuHow` is a constant that indicates how a lock should be acquired.
-  // Internal implementation detail.  Clients should ignore.
-  typedef const struct MuHowS* MuHow;
-
-  // Mutex::InternalAttemptToUseMutexInFatalSignalHandler()
-  //
-  // Causes the `Mutex` implementation to prepare itself for re-entry caused by
-  // future use of `Mutex` within a fatal signal handler. This method is
-  // intended for use only for last-ditch attempts to log crash information.
-  // It does not guarantee that attempts to use Mutexes within the handler will
-  // not deadlock; it merely makes other faults less likely.
-  //
-  // WARNING:  This routine must be invoked from a signal handler, and the
-  // signal handler must either loop forever or terminate the process.
-  // Attempts to return from (or `longjmp` out of) the signal handler once this
-  // call has been made may cause arbitrary program behaviour including
-  // crashes and deadlocks.
-  static void InternalAttemptToUseMutexInFatalSignalHandler();
-
- private:
-  std::atomic<intptr_t> mu_;  // The Mutex state.
-
-  // Post()/Wait() versus associated PerThreadSem; in class for required
-  // friendship with PerThreadSem.
-  static void IncrementSynchSem(Mutex* mu, base_internal::PerThreadSynch* w);
-  static bool DecrementSynchSem(Mutex* mu, base_internal::PerThreadSynch* w,
-                                synchronization_internal::KernelTimeout t);
-
-  // slow path acquire
-  void LockSlowLoop(SynchWaitParams* waitp, int flags);
-  // wrappers around LockSlowLoop()
-  bool LockSlowWithDeadline(MuHow how, const Condition* cond,
-                            synchronization_internal::KernelTimeout t,
-                            int flags);
-  void LockSlow(MuHow how, const Condition* cond,
-                int flags) ABSL_ATTRIBUTE_COLD;
-  // slow path release
-  void UnlockSlow(SynchWaitParams* waitp) ABSL_ATTRIBUTE_COLD;
-  // TryLock slow path.
-  bool TryLockSlow();
-  // ReaderTryLock slow path.
-  bool ReaderTryLockSlow();
-  // Common code between Await() and AwaitWithTimeout/Deadline()
-  bool AwaitCommon(const Condition& cond,
-                   synchronization_internal::KernelTimeout t);
-  bool LockWhenCommon(const Condition& cond,
-                      synchronization_internal::KernelTimeout t, bool write);
-  // Attempt to remove thread s from queue.
-  void TryRemove(base_internal::PerThreadSynch* s);
-  // Block a thread on mutex.
-  void Block(base_internal::PerThreadSynch* s);
-  // Wake a thread; return successor.
-  base_internal::PerThreadSynch* Wakeup(base_internal::PerThreadSynch* w);
-  void Dtor();
-
-  friend class CondVar;   // for access to Trans()/Fer().
-  void Trans(MuHow how);  // used for CondVar->Mutex transfer
-  void Fer(
-      base_internal::PerThreadSynch* w);  // used for CondVar->Mutex transfer
-
-  // Catch the error of writing Mutex when intending MutexLock.
-  explicit Mutex(const volatile Mutex* /*ignored*/) {}
-
-  Mutex(const Mutex&) = delete;
-  Mutex& operator=(const Mutex&) = delete;
-};
-
-// -----------------------------------------------------------------------------
-// Mutex RAII Wrappers
-// -----------------------------------------------------------------------------
-
-// MutexLock
-//
-// `MutexLock` is a helper class, which acquires and releases a `Mutex` via
-// RAII.
-//
-// Example:
-//
-// Class Foo {
-//  public:
-//   Foo::Bar* Baz() {
-//     MutexLock lock(&mu_);
-//     ...
-//     return bar;
-//   }
-//
-// private:
-//   Mutex mu_;
-// };
-class ABSL_SCOPED_LOCKABLE MutexLock {
- public:
-  // Constructors
-
-  // Calls `mu->Lock()` and returns when that call returns. That is, `*mu` is
-  // guaranteed to be locked when this object is constructed. Requires that
-  // `mu` be dereferenceable.
-  explicit MutexLock(Mutex* mu) ABSL_EXCLUSIVE_LOCK_FUNCTION(mu) : mu_(mu) {
-    this->mu_->Lock();
-  }
-
-  // Like above, but calls `mu->LockWhen(cond)` instead. That is, in addition to
-  // the above, the condition given by `cond` is also guaranteed to hold when
-  // this object is constructed.
-  explicit MutexLock(Mutex* mu, const Condition& cond)
-      ABSL_EXCLUSIVE_LOCK_FUNCTION(mu)
-      : mu_(mu) {
-    this->mu_->LockWhen(cond);
-  }
-
-  MutexLock(const MutexLock&) = delete;  // NOLINT(runtime/mutex)
-  MutexLock(MutexLock&&) = delete;       // NOLINT(runtime/mutex)
-  MutexLock& operator=(const MutexLock&) = delete;
-  MutexLock& operator=(MutexLock&&) = delete;
-
-  ~MutexLock() ABSL_UNLOCK_FUNCTION() { this->mu_->Unlock(); }
-
- private:
-  Mutex* const mu_;
-};
-
-// ReaderMutexLock
-//
-// The `ReaderMutexLock` is a helper class, like `MutexLock`, which acquires and
-// releases a shared lock on a `Mutex` via RAII.
-class ABSL_SCOPED_LOCKABLE ReaderMutexLock {
- public:
-  explicit ReaderMutexLock(Mutex* mu) ABSL_SHARED_LOCK_FUNCTION(mu) : mu_(mu) {
-    mu->ReaderLock();
-  }
-
-  explicit ReaderMutexLock(Mutex* mu, const Condition& cond)
-      ABSL_SHARED_LOCK_FUNCTION(mu)
-      : mu_(mu) {
-    mu->ReaderLockWhen(cond);
-  }
-
-  ReaderMutexLock(const ReaderMutexLock&) = delete;
-  ReaderMutexLock(ReaderMutexLock&&) = delete;
-  ReaderMutexLock& operator=(const ReaderMutexLock&) = delete;
-  ReaderMutexLock& operator=(ReaderMutexLock&&) = delete;
-
-  ~ReaderMutexLock() ABSL_UNLOCK_FUNCTION() { this->mu_->ReaderUnlock(); }
-
- private:
-  Mutex* const mu_;
-};
-
-// WriterMutexLock
-//
-// The `WriterMutexLock` is a helper class, like `MutexLock`, which acquires and
-// releases a write (exclusive) lock on a `Mutex` via RAII.
-class ABSL_SCOPED_LOCKABLE WriterMutexLock {
- public:
-  explicit WriterMutexLock(Mutex* mu) ABSL_EXCLUSIVE_LOCK_FUNCTION(mu)
-      : mu_(mu) {
-    mu->WriterLock();
-  }
-
-  explicit WriterMutexLock(Mutex* mu, const Condition& cond)
-      ABSL_EXCLUSIVE_LOCK_FUNCTION(mu)
-      : mu_(mu) {
-    mu->WriterLockWhen(cond);
-  }
-
-  WriterMutexLock(const WriterMutexLock&) = delete;
-  WriterMutexLock(WriterMutexLock&&) = delete;
-  WriterMutexLock& operator=(const WriterMutexLock&) = delete;
-  WriterMutexLock& operator=(WriterMutexLock&&) = delete;
-
-  ~WriterMutexLock() ABSL_UNLOCK_FUNCTION() { this->mu_->WriterUnlock(); }
-
- private:
-  Mutex* const mu_;
-};
-
-// -----------------------------------------------------------------------------
-// Condition
-// -----------------------------------------------------------------------------
-//
-// `Mutex` contains a number of member functions which take a `Condition` as an
-// argument; clients can wait for conditions to become `true` before attempting
-// to acquire the mutex. These sections are known as "condition critical"
-// sections. To use a `Condition`, you simply need to construct it, and use
-// within an appropriate `Mutex` member function; everything else in the
-// `Condition` class is an implementation detail.
-//
-// A `Condition` is specified as a function pointer which returns a boolean.
-// `Condition` functions should be pure functions -- their results should depend
-// only on passed arguments, should not consult any external state (such as
-// clocks), and should have no side-effects, aside from debug logging. Any
-// objects that the function may access should be limited to those which are
-// constant while the mutex is blocked on the condition (e.g. a stack variable),
-// or objects of state protected explicitly by the mutex.
-//
-// No matter which construction is used for `Condition`, the underlying
-// function pointer / functor / callable must not throw any
-// exceptions. Correctness of `Mutex` / `Condition` is not guaranteed in
-// the face of a throwing `Condition`. (When Abseil is allowed to depend
-// on C++17, these function pointers will be explicitly marked
-// `noexcept`; until then this requirement cannot be enforced in the
-// type system.)
-//
-// Note: to use a `Condition`, you need only construct it and pass it to a
-// suitable `Mutex' member function, such as `Mutex::Await()`, or to the
-// constructor of one of the scope guard classes.
-//
-// Example using LockWhen/Unlock:
-//
-//   // assume count_ is not internal reference count
-//   int count_ ABSL_GUARDED_BY(mu_);
-//   Condition count_is_zero(+[](int *count) { return *count == 0; }, &count_);
-//
-//   mu_.LockWhen(count_is_zero);
-//   // ...
-//   mu_.Unlock();
-//
-// Example using a scope guard:
-//
-//   {
-//     MutexLock lock(&mu_, count_is_zero);
-//     // ...
-//   }
-//
-// When multiple threads are waiting on exactly the same condition, make sure
-// that they are constructed with the same parameters (same pointer to function
-// + arg, or same pointer to object + method), so that the mutex implementation
-// can avoid redundantly evaluating the same condition for each thread.
-class Condition {
- public:
-  // A Condition that returns the result of "(*func)(arg)"
-  Condition(bool (*func)(void*), void* arg);
-
-  // Templated version for people who are averse to casts.
-  //
-  // To use a lambda, prepend it with unary plus, which converts the lambda
-  // into a function pointer:
-  //     Condition(+[](T* t) { return ...; }, arg).
-  //
-  // Note: lambdas in this case must contain no bound variables.
-  //
-  // See class comment for performance advice.
-  template <typename T>
-  Condition(bool (*func)(T*), T* arg);
-
-  // Same as above, but allows for cases where `arg` comes from a pointer that
-  // is convertible to the function parameter type `T*` but not an exact match.
-  //
-  // For example, the argument might be `X*` but the function takes `const X*`,
-  // or the argument might be `Derived*` while the function takes `Base*`, and
-  // so on for cases where the argument pointer can be implicitly converted.
-  //
-  // Implementation notes: This constructor overload is required in addition to
-  // the one above to allow deduction of `T` from `arg` for cases such as where
-  // a function template is passed as `func`. Also, the dummy `typename = void`
-  // template parameter exists just to work around a MSVC mangling bug.
-  template <typename T, typename = void>
-  Condition(bool (*func)(T*),
-            typename absl::internal::type_identity<T>::type* arg);
-
-  // Templated version for invoking a method that returns a `bool`.
-  //
-  // `Condition(object, &Class::Method)` constructs a `Condition` that evaluates
-  // `object->Method()`.
-  //
-  // Implementation Note: `absl::internal::type_identity` is used to allow
-  // methods to come from base classes. A simpler signature like
-  // `Condition(T*, bool (T::*)())` does not suffice.
-  template <typename T>
-  Condition(T* object,
-            bool (absl::internal::type_identity<T>::type::*method)());
-
-  // Same as above, for const members
-  template <typename T>
-  Condition(const T* object,
-            bool (absl::internal::type_identity<T>::type::*method)() const);
-
-  // A Condition that returns the value of `*cond`
-  explicit Condition(const bool* cond);
-
-  // Templated version for invoking a functor that returns a `bool`.
-  // This approach accepts pointers to non-mutable lambdas, `std::function`,
-  // the result of` std::bind` and user-defined functors that define
-  // `bool F::operator()() const`.
-  //
-  // Example:
-  //
-  //   auto reached = [this, current]() {
-  //     mu_.AssertReaderHeld();                // For annotalysis.
-  //     return processed_ >= current;
-  //   };
-  //   mu_.Await(Condition(&reached));
-  //
-  // NOTE: never use "mu_.AssertHeld()" instead of "mu_.AssertReaderHeld()" in
-  // the lambda as it may be called when the mutex is being unlocked from a
-  // scope holding only a reader lock, which will make the assertion not
-  // fulfilled and crash the binary.
-
-  // See class comment for performance advice. In particular, if there
-  // might be more than one waiter for the same condition, make sure
-  // that all waiters construct the condition with the same pointers.
-
-  // Implementation note: The second template parameter ensures that this
-  // constructor doesn't participate in overload resolution if T doesn't have
-  // `bool operator() const`.
-  template <typename T, typename E = decltype(static_cast<bool (T::*)() const>(
-                            &T::operator()))>
-  explicit Condition(const T* obj)
-      : Condition(obj, static_cast<bool (T::*)() const>(&T::operator())) {}
-
-  // A Condition that always returns `true`.
-  // kTrue is only useful in a narrow set of circumstances, mostly when
-  // it's passed conditionally. For example:
-  //
-  //   mu.LockWhen(some_flag ? kTrue : SomeOtherCondition);
-  //
-  // Note: {LockWhen,Await}With{Deadline,Timeout} methods with kTrue condition
-  // don't return immediately when the timeout happens, they still block until
-  // the Mutex becomes available. The return value of these methods does
-  // not indicate if the timeout was reached; rather it indicates whether or
-  // not the condition is true.
-  ABSL_CONST_INIT static const Condition kTrue;
-
-  // Evaluates the condition.
-  bool Eval() const;
-
-  // Returns `true` if the two conditions are guaranteed to return the same
-  // value if evaluated at the same time, `false` if the evaluation *may* return
-  // different results.
-  //
-  // Two `Condition` values are guaranteed equal if both their `func` and `arg`
-  // components are the same. A null pointer is equivalent to a `true`
-  // condition.
-  static bool GuaranteedEqual(const Condition* a, const Condition* b);
-
- private:
-  // Sizing an allocation for a method pointer can be subtle. In the Itanium
-  // specifications, a method pointer has a predictable, uniform size. On the
-  // other hand, MSVC ABI, method pointer sizes vary based on the
-  // inheritance of the class. Specifically, method pointers from classes with
-  // multiple inheritance are bigger than those of classes with single
-  // inheritance. Other variations also exist.
-
-#ifndef _MSC_VER
-  // Allocation for a function pointer or method pointer.
-  // The {0} initializer ensures that all unused bytes of this buffer are
-  // always zeroed out.  This is necessary, because GuaranteedEqual() compares
-  // all of the bytes, unaware of which bytes are relevant to a given `eval_`.
-  using MethodPtr = bool (Condition::*)();
-  char callback_[sizeof(MethodPtr)] = {0};
-#else
-  // It is well known that the larget MSVC pointer-to-member is 24 bytes. This
-  // may be the largest known pointer-to-member of any platform. For this
-  // reason we will allocate 24 bytes for MSVC platform toolchains.
-  char callback_[24] = {0};
-#endif
-
-  // Function with which to evaluate callbacks and/or arguments.
-  bool (*eval_)(const Condition*) = nullptr;
-
-  // Either an argument for a function call or an object for a method call.
-  void* arg_ = nullptr;
-
-  // Various functions eval_ can point to:
-  static bool CallVoidPtrFunction(const Condition*);
-  template <typename T>
-  static bool CastAndCallFunction(const Condition* c);
-  template <typename T, typename ConditionMethodPtr>
-  static bool CastAndCallMethod(const Condition* c);
-
-  // Helper methods for storing, validating, and reading callback arguments.
-  template <typename T>
-  inline void StoreCallback(T callback) {
-    static_assert(
-        sizeof(callback) <= sizeof(callback_),
-        "An overlarge pointer was passed as a callback to Condition.");
-    std::memcpy(callback_, &callback, sizeof(callback));
-  }
-
-  template <typename T>
-  inline void ReadCallback(T* callback) const {
-    std::memcpy(callback, callback_, sizeof(*callback));
-  }
-
-  static bool AlwaysTrue(const Condition*) { return true; }
-
-  // Used only to create kTrue.
-  constexpr Condition() : eval_(AlwaysTrue), arg_(nullptr) {}
-};
-
-// -----------------------------------------------------------------------------
-// CondVar
-// -----------------------------------------------------------------------------
-//
-// A condition variable, reflecting state evaluated separately outside of the
-// `Mutex` object, which can be signaled to wake callers.
-// This class is not normally needed; use `Mutex` member functions such as
-// `Mutex::Await()` and intrinsic `Condition` abstractions. In rare cases
-// with many threads and many conditions, `CondVar` may be faster.
-//
-// The implementation may deliver signals to any condition variable at
-// any time, even when no call to `Signal()` or `SignalAll()` is made; as a
-// result, upon being awoken, you must check the logical condition you have
-// been waiting upon.
-//
-// Examples:
-//
-// Usage for a thread waiting for some condition C protected by mutex mu:
-//       mu.Lock();
-//       while (!C) { cv->Wait(&mu); }        // releases and reacquires mu
-//       //  C holds; process data
-//       mu.Unlock();
-//
-// Usage to wake T is:
-//       mu.Lock();
-//       // process data, possibly establishing C
-//       if (C) { cv->Signal(); }
-//       mu.Unlock();
-//
-// If C may be useful to more than one waiter, use `SignalAll()` instead of
-// `Signal()`.
-//
-// With this implementation it is efficient to use `Signal()/SignalAll()` inside
-// the locked region; this usage can make reasoning about your program easier.
-//
-class CondVar {
- public:
-  // A `CondVar` allocated on the heap or on the stack can use the this
-  // constructor.
-  CondVar();
-
-  // CondVar::Wait()
-  //
-  // Atomically releases a `Mutex` and blocks on this condition variable.
-  // Waits until awakened by a call to `Signal()` or `SignalAll()` (or a
-  // spurious wakeup), then reacquires the `Mutex` and returns.
-  //
-  // Requires and ensures that the current thread holds the `Mutex`.
-  void Wait(Mutex* mu) {
-    WaitCommon(mu, synchronization_internal::KernelTimeout::Never());
-  }
-
-  // CondVar::WaitWithTimeout()
-  //
-  // Atomically releases a `Mutex` and blocks on this condition variable.
-  // Waits until awakened by a call to `Signal()` or `SignalAll()` (or a
-  // spurious wakeup), or until the timeout has expired, then reacquires
-  // the `Mutex` and returns.
-  //
-  // Returns true if the timeout has expired without this `CondVar`
-  // being signalled in any manner. If both the timeout has expired
-  // and this `CondVar` has been signalled, the implementation is free
-  // to return `true` or `false`.
-  //
-  // Requires and ensures that the current thread holds the `Mutex`.
-  bool WaitWithTimeout(Mutex* mu, absl::Duration timeout) {
-    return WaitCommon(mu, synchronization_internal::KernelTimeout(timeout));
-  }
-
-  // CondVar::WaitWithDeadline()
-  //
-  // Atomically releases a `Mutex` and blocks on this condition variable.
-  // Waits until awakened by a call to `Signal()` or `SignalAll()` (or a
-  // spurious wakeup), or until the deadline has passed, then reacquires
-  // the `Mutex` and returns.
-  //
-  // Deadlines in the past are equivalent to an immediate deadline.
-  //
-  // Returns true if the deadline has passed without this `CondVar`
-  // being signalled in any manner. If both the deadline has passed
-  // and this `CondVar` has been signalled, the implementation is free
-  // to return `true` or `false`.
-  //
-  // Requires and ensures that the current thread holds the `Mutex`.
-  bool WaitWithDeadline(Mutex* mu, absl::Time deadline) {
-    return WaitCommon(mu, synchronization_internal::KernelTimeout(deadline));
-  }
-
-  // CondVar::Signal()
-  //
-  // Signal this `CondVar`; wake at least one waiter if one exists.
-  void Signal();
-
-  // CondVar::SignalAll()
-  //
-  // Signal this `CondVar`; wake all waiters.
-  void SignalAll();
-
-  // CondVar::EnableDebugLog()
-  //
-  // Causes all subsequent uses of this `CondVar` to be logged via
-  // `ABSL_RAW_LOG(INFO)`. Log entries are tagged with `name` if `name != 0`.
-  // Note: this method substantially reduces `CondVar` performance.
-  void EnableDebugLog(const char* name);
-
- private:
-  bool WaitCommon(Mutex* mutex, synchronization_internal::KernelTimeout t);
-  void Remove(base_internal::PerThreadSynch* s);
-  std::atomic<intptr_t> cv_;  // Condition variable state.
-  CondVar(const CondVar&) = delete;
-  CondVar& operator=(const CondVar&) = delete;
-};
-
-// Variants of MutexLock.
-//
-// If you find yourself using one of these, consider instead using
-// Mutex::Unlock() and/or if-statements for clarity.
-
-// MutexLockMaybe
-//
-// MutexLockMaybe is like MutexLock, but is a no-op when mu is null.
-class ABSL_SCOPED_LOCKABLE MutexLockMaybe {
- public:
-  explicit MutexLockMaybe(Mutex* mu) ABSL_EXCLUSIVE_LOCK_FUNCTION(mu)
-      : mu_(mu) {
-    if (this->mu_ != nullptr) {
-      this->mu_->Lock();
-    }
-  }
-
-  explicit MutexLockMaybe(Mutex* mu, const Condition& cond)
-      ABSL_EXCLUSIVE_LOCK_FUNCTION(mu)
-      : mu_(mu) {
-    if (this->mu_ != nullptr) {
-      this->mu_->LockWhen(cond);
-    }
-  }
-
-  ~MutexLockMaybe() ABSL_UNLOCK_FUNCTION() {
-    if (this->mu_ != nullptr) {
-      this->mu_->Unlock();
-    }
-  }
-
- private:
-  Mutex* const mu_;
-  MutexLockMaybe(const MutexLockMaybe&) = delete;
-  MutexLockMaybe(MutexLockMaybe&&) = delete;
-  MutexLockMaybe& operator=(const MutexLockMaybe&) = delete;
-  MutexLockMaybe& operator=(MutexLockMaybe&&) = delete;
-};
-
-// ReleasableMutexLock
-//
-// ReleasableMutexLock is like MutexLock, but permits `Release()` of its
-// mutex before destruction. `Release()` may be called at most once.
-class ABSL_SCOPED_LOCKABLE ReleasableMutexLock {
- public:
-  explicit ReleasableMutexLock(Mutex* mu) ABSL_EXCLUSIVE_LOCK_FUNCTION(mu)
-      : mu_(mu) {
-    this->mu_->Lock();
-  }
-
-  explicit ReleasableMutexLock(Mutex* mu, const Condition& cond)
-      ABSL_EXCLUSIVE_LOCK_FUNCTION(mu)
-      : mu_(mu) {
-    this->mu_->LockWhen(cond);
-  }
-
-  ~ReleasableMutexLock() ABSL_UNLOCK_FUNCTION() {
-    if (this->mu_ != nullptr) {
-      this->mu_->Unlock();
-    }
-  }
-
-  void Release() ABSL_UNLOCK_FUNCTION();
-
- private:
-  Mutex* mu_;
-  ReleasableMutexLock(const ReleasableMutexLock&) = delete;
-  ReleasableMutexLock(ReleasableMutexLock&&) = delete;
-  ReleasableMutexLock& operator=(const ReleasableMutexLock&) = delete;
-  ReleasableMutexLock& operator=(ReleasableMutexLock&&) = delete;
-};
-
-inline Mutex::Mutex() : mu_(0) {
-  ABSL_TSAN_MUTEX_CREATE(this, __tsan_mutex_not_static);
-}
-
-inline constexpr Mutex::Mutex(absl::ConstInitType) : mu_(0) {}
-
-#if !defined(__APPLE__) && !defined(ABSL_BUILD_DLL)
-ABSL_ATTRIBUTE_ALWAYS_INLINE
-inline Mutex::~Mutex() { Dtor(); }
-#endif
-
-#if defined(NDEBUG) && !defined(ABSL_HAVE_THREAD_SANITIZER)
-// Use default (empty) destructor in release build for performance reasons.
-// We need to mark both Dtor and ~Mutex as always inline for inconsistent
-// builds that use both NDEBUG and !NDEBUG with dynamic libraries. In these
-// cases we want the empty functions to dissolve entirely rather than being
-// exported from dynamic libraries and potentially override the non-empty ones.
-ABSL_ATTRIBUTE_ALWAYS_INLINE
-inline void Mutex::Dtor() {}
-#endif
-
-inline CondVar::CondVar() : cv_(0) {}
-
-// static
-template <typename T, typename ConditionMethodPtr>
-bool Condition::CastAndCallMethod(const Condition* c) {
-  T* object = static_cast<T*>(c->arg_);
-  ConditionMethodPtr condition_method_pointer;
-  c->ReadCallback(&condition_method_pointer);
-  return (object->*condition_method_pointer)();
-}
-
-// static
-template <typename T>
-bool Condition::CastAndCallFunction(const Condition* c) {
-  bool (*function)(T*);
-  c->ReadCallback(&function);
-  T* argument = static_cast<T*>(c->arg_);
-  return (*function)(argument);
-}
-
-template <typename T>
-inline Condition::Condition(bool (*func)(T*), T* arg)
-    : eval_(&CastAndCallFunction<T>),
-      arg_(const_cast<void*>(static_cast<const void*>(arg))) {
-  static_assert(sizeof(&func) <= sizeof(callback_),
-                "An overlarge function pointer was passed to Condition.");
-  StoreCallback(func);
-}
-
-template <typename T, typename>
-inline Condition::Condition(
-    bool (*func)(T*), typename absl::internal::type_identity<T>::type* arg)
-    // Just delegate to the overload above.
-    : Condition(func, arg) {}
-
-template <typename T>
-inline Condition::Condition(
-    T* object, bool (absl::internal::type_identity<T>::type::*method)())
-    : eval_(&CastAndCallMethod<T, decltype(method)>), arg_(object) {
-  static_assert(sizeof(&method) <= sizeof(callback_),
-                "An overlarge method pointer was passed to Condition.");
-  StoreCallback(method);
-}
-
-template <typename T>
-inline Condition::Condition(
-    const T* object,
-    bool (absl::internal::type_identity<T>::type::*method)() const)
-    : eval_(&CastAndCallMethod<const T, decltype(method)>),
-      arg_(reinterpret_cast<void*>(const_cast<T*>(object))) {
-  StoreCallback(method);
-}
-
-// Register hooks for profiling support.
-//
-// The function pointer registered here will be called whenever a mutex is
-// contended.  The callback is given the cycles for which waiting happened (as
-// measured by //absl/base/internal/cycleclock.h, and which may not
-// be real "cycle" counts.)
-//
-// There is no ordering guarantee between when the hook is registered and when
-// callbacks will begin.  Only a single profiler can be installed in a running
-// binary; if this function is called a second time with a different function
-// pointer, the value is ignored (and will cause an assertion failure in debug
-// mode.)
-void RegisterMutexProfiler(void (*fn)(int64_t wait_cycles));
-
-// Register a hook for Mutex tracing.
-//
-// The function pointer registered here will be called whenever a mutex is
-// contended.  The callback is given an opaque handle to the contended mutex,
-// an event name, and the number of wait cycles (as measured by
-// //absl/base/internal/cycleclock.h, and which may not be real
-// "cycle" counts.)
-//
-// The only event name currently sent is "slow release".
-//
-// This has the same ordering and single-use limitations as
-// RegisterMutexProfiler() above.
-void RegisterMutexTracer(void (*fn)(const char* msg, const void* obj,
-                                    int64_t wait_cycles));
-
-// Register a hook for CondVar tracing.
-//
-// The function pointer registered here will be called here on various CondVar
-// events.  The callback is given an opaque handle to the CondVar object and
-// a string identifying the event.  This is thread-safe, but only a single
-// tracer can be registered.
-//
-// Events that can be sent are "Wait", "Unwait", "Signal wakeup", and
-// "SignalAll wakeup".
-//
-// This has the same ordering and single-use limitations as
-// RegisterMutexProfiler() above.
-void RegisterCondVarTracer(void (*fn)(const char* msg, const void* cv));
-
-// EnableMutexInvariantDebugging()
-//
-// Enable or disable global support for Mutex invariant debugging.  If enabled,
-// then invariant predicates can be registered per-Mutex for debug checking.
-// See Mutex::EnableInvariantDebugging().
-void EnableMutexInvariantDebugging(bool enabled);
-
-// When in debug mode, and when the feature has been enabled globally, the
-// implementation will keep track of lock ordering and complain (or optionally
-// crash) if a cycle is detected in the acquired-before graph.
-
-// Possible modes of operation for the deadlock detector in debug mode.
-enum class OnDeadlockCycle {
-  kIgnore,  // Neither report on nor attempt to track cycles in lock ordering
-  kReport,  // Report lock cycles to stderr when detected
-  kAbort,   // Report lock cycles to stderr when detected, then abort
-};
-
-// SetMutexDeadlockDetectionMode()
-//
-// Enable or disable global support for detection of potential deadlocks
-// due to Mutex lock ordering inversions.  When set to 'kIgnore', tracking of
-// lock ordering is disabled.  Otherwise, in debug builds, a lock ordering graph
-// will be maintained internally, and detected cycles will be reported in
-// the manner chosen here.
-void SetMutexDeadlockDetectionMode(OnDeadlockCycle mode);
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-// In some build configurations we pass --detect-odr-violations to the
-// gold linker.  This causes it to flag weak symbol overrides as ODR
-// violations.  Because ODR only applies to C++ and not C,
-// --detect-odr-violations ignores symbols not mangled with C++ names.
-// By changing our extension points to be extern "C", we dodge this
-// check.
-extern "C" {
-void ABSL_INTERNAL_C_SYMBOL(AbslInternalMutexYield)();
-}  // extern "C"
-
-#endif  // ABSL_SYNCHRONIZATION_MUTEX_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/synchronization/notification.h b/projects/llm_framework/include/abseil-cpp/absl/synchronization/notification.h
deleted file mode 100644
index 8986d9a4..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/synchronization/notification.h
+++ /dev/null
@@ -1,123 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// notification.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines a `Notification` abstraction, which allows threads
-// to receive notification of a single occurrence of a single event.
-//
-// The `Notification` object maintains a private boolean "notified" state that
-// transitions to `true` at most once. The `Notification` class provides the
-// following primary member functions:
-//   * `HasBeenNotified()` to query its state
-//   * `WaitForNotification*()` to have threads wait until the "notified" state
-//      is `true`.
-//   * `Notify()` to set the notification's "notified" state to `true` and
-//     notify all waiting threads that the event has occurred.
-//     This method may only be called once.
-//
-// Note that while `Notify()` may only be called once, it is perfectly valid to
-// call any of the `WaitForNotification*()` methods multiple times, from
-// multiple threads -- even after the notification's "notified" state has been
-// set -- in which case those methods will immediately return.
-//
-// Note that the lifetime of a `Notification` requires careful consideration;
-// it might not be safe to destroy a notification after calling `Notify()` since
-// it is still legal for other threads to call `WaitForNotification*()` methods
-// on the notification. However, observers responding to a "notified" state of
-// `true` can safely delete the notification without interfering with the call
-// to `Notify()` in the other thread.
-//
-// Memory ordering: For any threads X and Y, if X calls `Notify()`, then any
-// action taken by X before it calls `Notify()` is visible to thread Y after:
-//  * Y returns from `WaitForNotification()`, or
-//  * Y receives a `true` return value from either `HasBeenNotified()` or
-//    `WaitForNotificationWithTimeout()`.
-
-#ifndef ABSL_SYNCHRONIZATION_NOTIFICATION_H_
-#define ABSL_SYNCHRONIZATION_NOTIFICATION_H_
-
-#include <atomic>
-
-#include "absl/base/attributes.h"
-#include "absl/synchronization/mutex.h"
-#include "absl/time/time.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// -----------------------------------------------------------------------------
-// Notification
-// -----------------------------------------------------------------------------
-class Notification {
- public:
-  // Initializes the "notified" state to unnotified.
-  Notification() : notified_yet_(false) {}
-  explicit Notification(bool prenotify) : notified_yet_(prenotify) {}
-  Notification(const Notification&) = delete;
-  Notification& operator=(const Notification&) = delete;
-  ~Notification();
-
-  // Notification::HasBeenNotified()
-  //
-  // Returns the value of the notification's internal "notified" state.
-  ABSL_MUST_USE_RESULT bool HasBeenNotified() const {
-    return HasBeenNotifiedInternal(&this->notified_yet_);
-  }
-
-  // Notification::WaitForNotification()
-  //
-  // Blocks the calling thread until the notification's "notified" state is
-  // `true`. Note that if `Notify()` has been previously called on this
-  // notification, this function will immediately return.
-  void WaitForNotification() const;
-
-  // Notification::WaitForNotificationWithTimeout()
-  //
-  // Blocks until either the notification's "notified" state is `true` (which
-  // may occur immediately) or the timeout has elapsed, returning the value of
-  // its "notified" state in either case.
-  bool WaitForNotificationWithTimeout(absl::Duration timeout) const;
-
-  // Notification::WaitForNotificationWithDeadline()
-  //
-  // Blocks until either the notification's "notified" state is `true` (which
-  // may occur immediately) or the deadline has expired, returning the value of
-  // its "notified" state in either case.
-  bool WaitForNotificationWithDeadline(absl::Time deadline) const;
-
-  // Notification::Notify()
-  //
-  // Sets the "notified" state of this notification to `true` and wakes waiting
-  // threads. Note: do not call `Notify()` multiple times on the same
-  // `Notification`; calling `Notify()` more than once on the same notification
-  // results in undefined behavior.
-  void Notify();
-
- private:
-  static inline bool HasBeenNotifiedInternal(
-      const std::atomic<bool>* notified_yet) {
-    return notified_yet->load(std::memory_order_acquire);
-  }
-
-  mutable Mutex mutex_;
-  std::atomic<bool> notified_yet_;  // written under mutex_
-};
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_SYNCHRONIZATION_NOTIFICATION_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/time/civil_time.h b/projects/llm_framework/include/abseil-cpp/absl/time/civil_time.h
deleted file mode 100644
index 3e904a11..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/time/civil_time.h
+++ /dev/null
@@ -1,589 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: civil_time.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines abstractions for computing with "civil time".
-// The term "civil time" refers to the legally recognized human-scale time
-// that is represented by the six fields `YYYY-MM-DD hh:mm:ss`. A "date"
-// is perhaps the most common example of a civil time (represented here as
-// an `absl::CivilDay`).
-//
-// Modern-day civil time follows the Gregorian Calendar and is a
-// time-zone-independent concept: a civil time of "2015-06-01 12:00:00", for
-// example, is not tied to a time zone. Put another way, a civil time does not
-// map to a unique point in time; a civil time must be mapped to an absolute
-// time *through* a time zone.
-//
-// Because a civil time is what most people think of as "time," it is common to
-// map absolute times to civil times to present to users.
-//
-// Time zones define the relationship between absolute and civil times. Given an
-// absolute or civil time and a time zone, you can compute the other time:
-//
-//   Civil Time = F(Absolute Time, Time Zone)
-//   Absolute Time = G(Civil Time, Time Zone)
-//
-// The Abseil time library allows you to construct such civil times from
-// absolute times; consult time.h for such functionality.
-//
-// This library provides six classes for constructing civil-time objects, and
-// provides several helper functions for rounding, iterating, and performing
-// arithmetic on civil-time objects, while avoiding complications like
-// daylight-saving time (DST):
-//
-//   * `absl::CivilSecond`
-//   * `absl::CivilMinute`
-//   * `absl::CivilHour`
-//   * `absl::CivilDay`
-//   * `absl::CivilMonth`
-//   * `absl::CivilYear`
-//
-// Example:
-//
-//   // Construct a civil-time object for a specific day
-//   const absl::CivilDay cd(1969, 07, 20);
-//
-//   // Construct a civil-time object for a specific second
-//   const absl::CivilSecond cd(2018, 8, 1, 12, 0, 1);
-//
-// Note: In C++14 and later, this library is usable in a constexpr context.
-//
-// Example:
-//
-//   // Valid in C++14
-//   constexpr absl::CivilDay cd(1969, 07, 20);
-
-#ifndef ABSL_TIME_CIVIL_TIME_H_
-#define ABSL_TIME_CIVIL_TIME_H_
-
-#include <iosfwd>
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/strings/string_view.h"
-#include "absl/time/internal/cctz/include/cctz/civil_time.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-namespace time_internal {
-struct second_tag : cctz::detail::second_tag {};
-struct minute_tag : second_tag, cctz::detail::minute_tag {};
-struct hour_tag : minute_tag, cctz::detail::hour_tag {};
-struct day_tag : hour_tag, cctz::detail::day_tag {};
-struct month_tag : day_tag, cctz::detail::month_tag {};
-struct year_tag : month_tag, cctz::detail::year_tag {};
-}  // namespace time_internal
-
-// -----------------------------------------------------------------------------
-// CivilSecond, CivilMinute, CivilHour, CivilDay, CivilMonth, CivilYear
-// -----------------------------------------------------------------------------
-//
-// Each of these civil-time types is a simple value type with the same
-// interface for construction and the same six accessors for each of the civil
-// time fields (year, month, day, hour, minute, and second, aka YMDHMS). These
-// classes differ only in their alignment, which is indicated by the type name
-// and specifies the field on which arithmetic operates.
-//
-// CONSTRUCTION
-//
-// Each of the civil-time types can be constructed in two ways: by directly
-// passing to the constructor up to six integers representing the YMDHMS fields,
-// or by copying the YMDHMS fields from a differently aligned civil-time type.
-// Omitted fields are assigned their minimum valid value. Hours, minutes, and
-// seconds will be set to 0, month and day will be set to 1. Since there is no
-// minimum year, the default is 1970.
-//
-// Examples:
-//
-//   absl::CivilDay default_value;               // 1970-01-01 00:00:00
-//
-//   absl::CivilDay a(2015, 2, 3);               // 2015-02-03 00:00:00
-//   absl::CivilDay b(2015, 2, 3, 4, 5, 6);      // 2015-02-03 00:00:00
-//   absl::CivilDay c(2015);                     // 2015-01-01 00:00:00
-//
-//   absl::CivilSecond ss(2015, 2, 3, 4, 5, 6);  // 2015-02-03 04:05:06
-//   absl::CivilMinute mm(ss);                   // 2015-02-03 04:05:00
-//   absl::CivilHour hh(mm);                     // 2015-02-03 04:00:00
-//   absl::CivilDay d(hh);                       // 2015-02-03 00:00:00
-//   absl::CivilMonth m(d);                      // 2015-02-01 00:00:00
-//   absl::CivilYear y(m);                       // 2015-01-01 00:00:00
-//
-//   m = absl::CivilMonth(y);                    // 2015-01-01 00:00:00
-//   d = absl::CivilDay(m);                      // 2015-01-01 00:00:00
-//   hh = absl::CivilHour(d);                    // 2015-01-01 00:00:00
-//   mm = absl::CivilMinute(hh);                 // 2015-01-01 00:00:00
-//   ss = absl::CivilSecond(mm);                 // 2015-01-01 00:00:00
-//
-// Each civil-time class is aligned to the civil-time field indicated in the
-// class's name after normalization. Alignment is performed by setting all the
-// inferior fields to their minimum valid value (as described above). The
-// following are examples of how each of the six types would align the fields
-// representing November 22, 2015 at 12:34:56 in the afternoon. (Note: the
-// string format used here is not important; it's just a shorthand way of
-// showing the six YMDHMS fields.)
-//
-//   absl::CivilSecond   : 2015-11-22 12:34:56
-//   absl::CivilMinute   : 2015-11-22 12:34:00
-//   absl::CivilHour     : 2015-11-22 12:00:00
-//   absl::CivilDay      : 2015-11-22 00:00:00
-//   absl::CivilMonth    : 2015-11-01 00:00:00
-//   absl::CivilYear     : 2015-01-01 00:00:00
-//
-// Each civil-time type performs arithmetic on the field to which it is
-// aligned. This means that adding 1 to an absl::CivilDay increments the day
-// field (normalizing as necessary), and subtracting 7 from an absl::CivilMonth
-// operates on the month field (normalizing as necessary). All arithmetic
-// produces a valid civil time. Difference requires two similarly aligned
-// civil-time objects and returns the scalar answer in units of the objects'
-// alignment. For example, the difference between two absl::CivilHour objects
-// will give an answer in units of civil hours.
-//
-// ALIGNMENT CONVERSION
-//
-// The alignment of a civil-time object cannot change, but the object may be
-// used to construct a new object with a different alignment. This is referred
-// to as "realigning". When realigning to a type with the same or more
-// precision (e.g., absl::CivilDay -> absl::CivilSecond), the conversion may be
-// performed implicitly since no information is lost. However, if information
-// could be discarded (e.g., CivilSecond -> CivilDay), the conversion must
-// be explicit at the call site.
-//
-// Examples:
-//
-//   void UseDay(absl::CivilDay day);
-//
-//   absl::CivilSecond cs;
-//   UseDay(cs);                  // Won't compile because data may be discarded
-//   UseDay(absl::CivilDay(cs));  // OK: explicit conversion
-//
-//   absl::CivilDay cd;
-//   UseDay(cd);                  // OK: no conversion needed
-//
-//   absl::CivilMonth cm;
-//   UseDay(cm);                  // OK: implicit conversion to absl::CivilDay
-//
-// NORMALIZATION
-//
-// Normalization takes invalid values and adjusts them to produce valid values.
-// Within the civil-time library, integer arguments passed to the Civil*
-// constructors may be out-of-range, in which case they are normalized by
-// carrying overflow into a field of courser granularity to produce valid
-// civil-time objects. This normalization enables natural arithmetic on
-// constructor arguments without worrying about the field's range.
-//
-// Examples:
-//
-//   // Out-of-range; normalized to 2016-11-01
-//   absl::CivilDay d(2016, 10, 32);
-//   // Out-of-range, negative: normalized to 2016-10-30T23
-//   absl::CivilHour h1(2016, 10, 31, -1);
-//   // Normalization is cumulative: normalized to 2016-10-30T23
-//   absl::CivilHour h2(2016, 10, 32, -25);
-//
-// Note: If normalization is undesired, you can signal an error by comparing
-// the constructor arguments to the normalized values returned by the YMDHMS
-// properties.
-//
-// COMPARISON
-//
-// Comparison between civil-time objects considers all six YMDHMS fields,
-// regardless of the type's alignment. Comparison between differently aligned
-// civil-time types is allowed.
-//
-// Examples:
-//
-//   absl::CivilDay feb_3(2015, 2, 3);  // 2015-02-03 00:00:00
-//   absl::CivilDay mar_4(2015, 3, 4);  // 2015-03-04 00:00:00
-//   // feb_3 < mar_4
-//   // absl::CivilYear(feb_3) == absl::CivilYear(mar_4)
-//
-//   absl::CivilSecond feb_3_noon(2015, 2, 3, 12, 0, 0);  // 2015-02-03 12:00:00
-//   // feb_3 < feb_3_noon
-//   // feb_3 == absl::CivilDay(feb_3_noon)
-//
-//   // Iterates all the days of February 2015.
-//   for (absl::CivilDay d(2015, 2, 1); d < absl::CivilMonth(2015, 3); ++d) {
-//     // ...
-//   }
-//
-// ARITHMETIC
-//
-// Civil-time types support natural arithmetic operators such as addition,
-// subtraction, and difference. Arithmetic operates on the civil-time field
-// indicated in the type's name. Difference operators require arguments with
-// the same alignment and return the answer in units of the alignment.
-//
-// Example:
-//
-//   absl::CivilDay a(2015, 2, 3);
-//   ++a;                              // 2015-02-04 00:00:00
-//   --a;                              // 2015-02-03 00:00:00
-//   absl::CivilDay b = a + 1;         // 2015-02-04 00:00:00
-//   absl::CivilDay c = 1 + b;         // 2015-02-05 00:00:00
-//   int n = c - a;                    // n = 2 (civil days)
-//   int m = c - absl::CivilMonth(c);  // Won't compile: different types.
-//
-// ACCESSORS
-//
-// Each civil-time type has accessors for all six of the civil-time fields:
-// year, month, day, hour, minute, and second.
-//
-// civil_year_t year()
-// int          month()
-// int          day()
-// int          hour()
-// int          minute()
-// int          second()
-//
-// Recall that fields inferior to the type's alignment will be set to their
-// minimum valid value.
-//
-// Example:
-//
-//   absl::CivilDay d(2015, 6, 28);
-//   // d.year() == 2015
-//   // d.month() == 6
-//   // d.day() == 28
-//   // d.hour() == 0
-//   // d.minute() == 0
-//   // d.second() == 0
-//
-// CASE STUDY: Adding a month to January 31.
-//
-// One of the classic questions that arises when considering a civil time
-// library (or a date library or a date/time library) is this:
-//   "What is the result of adding a month to January 31?"
-// This is an interesting question because it is unclear what is meant by a
-// "month", and several different answers are possible, depending on context:
-//
-//   1. March 3 (or 2 if a leap year), if "add a month" means to add a month to
-//      the current month, and adjust the date to overflow the extra days into
-//      March. In this case the result of "February 31" would be normalized as
-//      within the civil-time library.
-//   2. February 28 (or 29 if a leap year), if "add a month" means to add a
-//      month, and adjust the date while holding the resulting month constant.
-//      In this case, the result of "February 31" would be truncated to the last
-//      day in February.
-//   3. An error. The caller may get some error, an exception, an invalid date
-//      object, or perhaps return `false`. This may make sense because there is
-//      no single unambiguously correct answer to the question.
-//
-// Practically speaking, any answer that is not what the programmer intended
-// is the wrong answer.
-//
-// The Abseil time library avoids this problem by making it impossible to
-// ask ambiguous questions. All civil-time objects are aligned to a particular
-// civil-field boundary (such as aligned to a year, month, day, hour, minute,
-// or second), and arithmetic operates on the field to which the object is
-// aligned. This means that in order to "add a month" the object must first be
-// aligned to a month boundary, which is equivalent to the first day of that
-// month.
-//
-// Of course, there are ways to compute an answer the question at hand using
-// this Abseil time library, but they require the programmer to be explicit
-// about the answer they expect. To illustrate, let's see how to compute all
-// three of the above possible answers to the question of "Jan 31 plus 1
-// month":
-//
-// Example:
-//
-//   const absl::CivilDay d(2015, 1, 31);
-//
-//   // Answer 1:
-//   // Add 1 to the month field in the constructor, and rely on normalization.
-//   const auto normalized = absl::CivilDay(d.year(), d.month() + 1, d.day());
-//   // normalized == 2015-03-03 (aka Feb 31)
-//
-//   // Answer 2:
-//   // Add 1 to month field, capping to the end of next month.
-//   const auto next_month = absl::CivilMonth(d) + 1;
-//   const auto last_day_of_next_month = absl::CivilDay(next_month + 1) - 1;
-//   const auto capped = std::min(normalized, last_day_of_next_month);
-//   // capped == 2015-02-28
-//
-//   // Answer 3:
-//   // Signal an error if the normalized answer is not in next month.
-//   if (absl::CivilMonth(normalized) != next_month) {
-//     // error, month overflow
-//   }
-//
-using CivilSecond =
-    time_internal::cctz::detail::civil_time<time_internal::second_tag>;
-using CivilMinute =
-    time_internal::cctz::detail::civil_time<time_internal::minute_tag>;
-using CivilHour =
-    time_internal::cctz::detail::civil_time<time_internal::hour_tag>;
-using CivilDay =
-    time_internal::cctz::detail::civil_time<time_internal::day_tag>;
-using CivilMonth =
-    time_internal::cctz::detail::civil_time<time_internal::month_tag>;
-using CivilYear =
-    time_internal::cctz::detail::civil_time<time_internal::year_tag>;
-
-// civil_year_t
-//
-// Type alias of a civil-time year value. This type is guaranteed to (at least)
-// support any year value supported by `time_t`.
-//
-// Example:
-//
-//   absl::CivilSecond cs = ...;
-//   absl::civil_year_t y = cs.year();
-//   cs = absl::CivilSecond(y, 1, 1, 0, 0, 0);  // CivilSecond(CivilYear(cs))
-//
-using civil_year_t = time_internal::cctz::year_t;
-
-// civil_diff_t
-//
-// Type alias of the difference between two civil-time values.
-// This type is used to indicate arguments that are not
-// normalized (such as parameters to the civil-time constructors), the results
-// of civil-time subtraction, or the operand to civil-time addition.
-//
-// Example:
-//
-//   absl::civil_diff_t n_sec = cs1 - cs2;             // cs1 == cs2 + n_sec;
-//
-using civil_diff_t = time_internal::cctz::diff_t;
-
-// Weekday::monday, Weekday::tuesday, Weekday::wednesday, Weekday::thursday,
-// Weekday::friday, Weekday::saturday, Weekday::sunday
-//
-// The Weekday enum class represents the civil-time concept of a "weekday" with
-// members for all days of the week.
-//
-//   absl::Weekday wd = absl::Weekday::thursday;
-//
-using Weekday = time_internal::cctz::weekday;
-
-// GetWeekday()
-//
-// Returns the absl::Weekday for the given (realigned) civil-time value.
-//
-// Example:
-//
-//   absl::CivilDay a(2015, 8, 13);
-//   absl::Weekday wd = absl::GetWeekday(a);  // wd == absl::Weekday::thursday
-//
-inline Weekday GetWeekday(CivilSecond cs) {
-  return time_internal::cctz::get_weekday(cs);
-}
-
-// NextWeekday()
-// PrevWeekday()
-//
-// Returns the absl::CivilDay that strictly follows or precedes a given
-// absl::CivilDay, and that falls on the given absl::Weekday.
-//
-// Example, given the following month:
-//
-//       August 2015
-//   Su Mo Tu We Th Fr Sa
-//                      1
-//    2  3  4  5  6  7  8
-//    9 10 11 12 13 14 15
-//   16 17 18 19 20 21 22
-//   23 24 25 26 27 28 29
-//   30 31
-//
-//   absl::CivilDay a(2015, 8, 13);
-//   // absl::GetWeekday(a) == absl::Weekday::thursday
-//   absl::CivilDay b = absl::NextWeekday(a, absl::Weekday::thursday);
-//   // b = 2015-08-20
-//   absl::CivilDay c = absl::PrevWeekday(a, absl::Weekday::thursday);
-//   // c = 2015-08-06
-//
-//   absl::CivilDay d = ...
-//   // Gets the following Thursday if d is not already Thursday
-//   absl::CivilDay thurs1 = absl::NextWeekday(d - 1, absl::Weekday::thursday);
-//   // Gets the previous Thursday if d is not already Thursday
-//   absl::CivilDay thurs2 = absl::PrevWeekday(d + 1, absl::Weekday::thursday);
-//
-inline CivilDay NextWeekday(CivilDay cd, Weekday wd) {
-  return CivilDay(time_internal::cctz::next_weekday(cd, wd));
-}
-inline CivilDay PrevWeekday(CivilDay cd, Weekday wd) {
-  return CivilDay(time_internal::cctz::prev_weekday(cd, wd));
-}
-
-// GetYearDay()
-//
-// Returns the day-of-year for the given (realigned) civil-time value.
-//
-// Example:
-//
-//   absl::CivilDay a(2015, 1, 1);
-//   int yd_jan_1 = absl::GetYearDay(a);   // yd_jan_1 = 1
-//   absl::CivilDay b(2015, 12, 31);
-//   int yd_dec_31 = absl::GetYearDay(b);  // yd_dec_31 = 365
-//
-inline int GetYearDay(CivilSecond cs) {
-  return time_internal::cctz::get_yearday(cs);
-}
-
-// FormatCivilTime()
-//
-// Formats the given civil-time value into a string value of the following
-// format:
-//
-//  Type        | Format
-//  ---------------------------------
-//  CivilSecond | YYYY-MM-DDTHH:MM:SS
-//  CivilMinute | YYYY-MM-DDTHH:MM
-//  CivilHour   | YYYY-MM-DDTHH
-//  CivilDay    | YYYY-MM-DD
-//  CivilMonth  | YYYY-MM
-//  CivilYear   | YYYY
-//
-// Example:
-//
-//   absl::CivilDay d = absl::CivilDay(1969, 7, 20);
-//   std::string day_string = absl::FormatCivilTime(d);  // "1969-07-20"
-//
-std::string FormatCivilTime(CivilSecond c);
-std::string FormatCivilTime(CivilMinute c);
-std::string FormatCivilTime(CivilHour c);
-std::string FormatCivilTime(CivilDay c);
-std::string FormatCivilTime(CivilMonth c);
-std::string FormatCivilTime(CivilYear c);
-
-// Support for StrFormat(), StrCat(), etc
-template <typename Sink>
-void AbslStringify(Sink& sink, CivilSecond c) {
-  sink.Append(FormatCivilTime(c));
-}
-template <typename Sink>
-void AbslStringify(Sink& sink, CivilMinute c) {
-  sink.Append(FormatCivilTime(c));
-}
-template <typename Sink>
-void AbslStringify(Sink& sink, CivilHour c) {
-  sink.Append(FormatCivilTime(c));
-}
-template <typename Sink>
-void AbslStringify(Sink& sink, CivilDay c) {
-  sink.Append(FormatCivilTime(c));
-}
-template <typename Sink>
-void AbslStringify(Sink& sink, CivilMonth c) {
-  sink.Append(FormatCivilTime(c));
-}
-template <typename Sink>
-void AbslStringify(Sink& sink, CivilYear c) {
-  sink.Append(FormatCivilTime(c));
-}
-
-// absl::ParseCivilTime()
-//
-// Parses a civil-time value from the specified `absl::string_view` into the
-// passed output parameter. Returns `true` upon successful parsing.
-//
-// The expected form of the input string is as follows:
-//
-//  Type        | Format
-//  ---------------------------------
-//  CivilSecond | YYYY-MM-DDTHH:MM:SS
-//  CivilMinute | YYYY-MM-DDTHH:MM
-//  CivilHour   | YYYY-MM-DDTHH
-//  CivilDay    | YYYY-MM-DD
-//  CivilMonth  | YYYY-MM
-//  CivilYear   | YYYY
-//
-// Example:
-//
-//   absl::CivilDay d;
-//   bool ok = absl::ParseCivilTime("2018-01-02", &d); // OK
-//
-// Note that parsing will fail if the string's format does not match the
-// expected type exactly. `ParseLenientCivilTime()` below is more lenient.
-//
-bool ParseCivilTime(absl::string_view s, CivilSecond* c);
-bool ParseCivilTime(absl::string_view s, CivilMinute* c);
-bool ParseCivilTime(absl::string_view s, CivilHour* c);
-bool ParseCivilTime(absl::string_view s, CivilDay* c);
-bool ParseCivilTime(absl::string_view s, CivilMonth* c);
-bool ParseCivilTime(absl::string_view s, CivilYear* c);
-
-// ParseLenientCivilTime()
-//
-// Parses any of the formats accepted by `absl::ParseCivilTime()`, but is more
-// lenient if the format of the string does not exactly match the associated
-// type.
-//
-// Example:
-//
-//   absl::CivilDay d;
-//   bool ok = absl::ParseLenientCivilTime("1969-07-20", &d); // OK
-//   ok = absl::ParseLenientCivilTime("1969-07-20T10", &d);   // OK: T10 floored
-//   ok = absl::ParseLenientCivilTime("1969-07", &d);   // OK: day defaults to 1
-//
-bool ParseLenientCivilTime(absl::string_view s, CivilSecond* c);
-bool ParseLenientCivilTime(absl::string_view s, CivilMinute* c);
-bool ParseLenientCivilTime(absl::string_view s, CivilHour* c);
-bool ParseLenientCivilTime(absl::string_view s, CivilDay* c);
-bool ParseLenientCivilTime(absl::string_view s, CivilMonth* c);
-bool ParseLenientCivilTime(absl::string_view s, CivilYear* c);
-
-namespace time_internal {  // For functions found via ADL on civil-time tags.
-
-// Streaming Operators
-//
-// Each civil-time type may be sent to an output stream using operator<<().
-// The result matches the string produced by `FormatCivilTime()`.
-//
-// Example:
-//
-//   absl::CivilDay d = absl::CivilDay(1969, 7, 20);
-//   std::cout << "Date is: " << d << "\n";
-//
-std::ostream& operator<<(std::ostream& os, CivilYear y);
-std::ostream& operator<<(std::ostream& os, CivilMonth m);
-std::ostream& operator<<(std::ostream& os, CivilDay d);
-std::ostream& operator<<(std::ostream& os, CivilHour h);
-std::ostream& operator<<(std::ostream& os, CivilMinute m);
-std::ostream& operator<<(std::ostream& os, CivilSecond s);
-
-// AbslParseFlag()
-//
-// Parses the command-line flag string representation `s` into a civil-time
-// value. Flags must be specified in a format that is valid for
-// `absl::ParseLenientCivilTime()`.
-bool AbslParseFlag(absl::string_view s, CivilSecond* c, std::string* error);
-bool AbslParseFlag(absl::string_view s, CivilMinute* c, std::string* error);
-bool AbslParseFlag(absl::string_view s, CivilHour* c, std::string* error);
-bool AbslParseFlag(absl::string_view s, CivilDay* c, std::string* error);
-bool AbslParseFlag(absl::string_view s, CivilMonth* c, std::string* error);
-bool AbslParseFlag(absl::string_view s, CivilYear* c, std::string* error);
-
-// AbslUnparseFlag()
-//
-// Unparses a civil-time value into a command-line string representation using
-// the format specified by `absl::ParseCivilTime()`.
-std::string AbslUnparseFlag(CivilSecond c);
-std::string AbslUnparseFlag(CivilMinute c);
-std::string AbslUnparseFlag(CivilHour c);
-std::string AbslUnparseFlag(CivilDay c);
-std::string AbslUnparseFlag(CivilMonth c);
-std::string AbslUnparseFlag(CivilYear c);
-
-}  // namespace time_internal
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TIME_CIVIL_TIME_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/time/clock.h b/projects/llm_framework/include/abseil-cpp/absl/time/clock.h
deleted file mode 100644
index 41d2cf27..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/time/clock.h
+++ /dev/null
@@ -1,78 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: clock.h
-// -----------------------------------------------------------------------------
-//
-// This header file contains utility functions for working with the system-wide
-// realtime clock. For descriptions of the main time abstractions used within
-// this header file, consult the time.h header file.
-#ifndef ABSL_TIME_CLOCK_H_
-#define ABSL_TIME_CLOCK_H_
-
-#include <cstdint>
-
-#include "absl/base/config.h"
-#include "absl/base/macros.h"
-#include "absl/time/time.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// Now()
-//
-// Returns the current time, expressed as an `absl::Time` absolute time value.
-absl::Time Now();
-
-// GetCurrentTimeNanos()
-//
-// Returns the current time, expressed as a count of nanoseconds since the Unix
-// Epoch (https://en.wikipedia.org/wiki/Unix_time). Prefer `absl::Now()` instead
-// for all but the most performance-sensitive cases (i.e. when you are calling
-// this function hundreds of thousands of times per second).
-int64_t GetCurrentTimeNanos();
-
-// SleepFor()
-//
-// Sleeps for the specified duration, expressed as an `absl::Duration`.
-//
-// Notes:
-// * Signal interruptions will not reduce the sleep duration.
-// * Returns immediately when passed a nonpositive duration.
-void SleepFor(absl::Duration duration);
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-// -----------------------------------------------------------------------------
-// Implementation Details
-// -----------------------------------------------------------------------------
-
-// In some build configurations we pass --detect-odr-violations to the
-// gold linker.  This causes it to flag weak symbol overrides as ODR
-// violations.  Because ODR only applies to C++ and not C,
-// --detect-odr-violations ignores symbols not mangled with C++ names.
-// By changing our extension points to be extern "C", we dodge this
-// check.
-extern "C" {
-ABSL_DLL void ABSL_INTERNAL_C_SYMBOL(AbslInternalSleepFor)(
-    absl::Duration duration);
-}  // extern "C"
-
-inline void absl::SleepFor(absl::Duration duration) {
-  ABSL_INTERNAL_C_SYMBOL(AbslInternalSleepFor)(duration);
-}
-
-#endif  // ABSL_TIME_CLOCK_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/civil_time.h b/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/civil_time.h
deleted file mode 100644
index d47ff86f..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/civil_time.h
+++ /dev/null
@@ -1,332 +0,0 @@
-// Copyright 2016 Google Inc. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//   https://www.apache.org/licenses/LICENSE-2.0
-//
-//   Unless required by applicable law or agreed to in writing, software
-//   distributed under the License is distributed on an "AS IS" BASIS,
-//   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-//   See the License for the specific language governing permissions and
-//   limitations under the License.
-
-#ifndef ABSL_TIME_INTERNAL_CCTZ_CIVIL_TIME_H_
-#define ABSL_TIME_INTERNAL_CCTZ_CIVIL_TIME_H_
-
-#include "absl/base/config.h"
-#include "absl/time/internal/cctz/include/cctz/civil_time_detail.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace time_internal {
-namespace cctz {
-
-// The term "civil time" refers to the legally recognized human-scale time
-// that is represented by the six fields YYYY-MM-DD hh:mm:ss. Modern-day civil
-// time follows the Gregorian Calendar and is a time-zone-independent concept.
-// A "date" is perhaps the most common example of a civil time (represented in
-// this library as cctz::civil_day). This library provides six classes and a
-// handful of functions that help with rounding, iterating, and arithmetic on
-// civil times while avoiding complications like daylight-saving time (DST).
-//
-// The following six classes form the core of this civil-time library:
-//
-//   * civil_second
-//   * civil_minute
-//   * civil_hour
-//   * civil_day
-//   * civil_month
-//   * civil_year
-//
-// Each class is a simple value type with the same interface for construction
-// and the same six accessors for each of the civil fields (year, month, day,
-// hour, minute, and second, aka YMDHMS). These classes differ only in their
-// alignment, which is indicated by the type name and specifies the field on
-// which arithmetic operates.
-//
-// Each class can be constructed by passing up to six optional integer
-// arguments representing the YMDHMS fields (in that order) to the
-// constructor. Omitted fields are assigned their minimum valid value. Hours,
-// minutes, and seconds will be set to 0, month and day will be set to 1, and
-// since there is no minimum valid year, it will be set to 1970. So, a
-// default-constructed civil-time object will have YMDHMS fields representing
-// "1970-01-01 00:00:00". Fields that are out-of-range are normalized (e.g.,
-// October 32 -> November 1) so that all civil-time objects represent valid
-// values.
-//
-// Each civil-time class is aligned to the civil-time field indicated in the
-// class's name after normalization. Alignment is performed by setting all the
-// inferior fields to their minimum valid value (as described above). The
-// following are examples of how each of the six types would align the fields
-// representing November 22, 2015 at 12:34:56 in the afternoon. (Note: the
-// string format used here is not important; it's just a shorthand way of
-// showing the six YMDHMS fields.)
-//
-//   civil_second  2015-11-22 12:34:56
-//   civil_minute  2015-11-22 12:34:00
-//   civil_hour    2015-11-22 12:00:00
-//   civil_day     2015-11-22 00:00:00
-//   civil_month   2015-11-01 00:00:00
-//   civil_year    2015-01-01 00:00:00
-//
-// Each civil-time type performs arithmetic on the field to which it is
-// aligned. This means that adding 1 to a civil_day increments the day field
-// (normalizing as necessary), and subtracting 7 from a civil_month operates
-// on the month field (normalizing as necessary). All arithmetic produces a
-// valid civil time. Difference requires two similarly aligned civil-time
-// objects and returns the scalar answer in units of the objects' alignment.
-// For example, the difference between two civil_hour objects will give an
-// answer in units of civil hours.
-//
-// In addition to the six civil-time types just described, there are
-// a handful of helper functions and algorithms for performing common
-// calculations. These are described below.
-//
-// Note: In C++14 and later, this library is usable in a constexpr context.
-//
-// CONSTRUCTION:
-//
-// Each of the civil-time types can be constructed in two ways: by directly
-// passing to the constructor up to six (optional) integers representing the
-// YMDHMS fields, or by copying the YMDHMS fields from a differently aligned
-// civil-time type.
-//
-//   civil_day default_value;  // 1970-01-01 00:00:00
-//
-//   civil_day a(2015, 2, 3);           // 2015-02-03 00:00:00
-//   civil_day b(2015, 2, 3, 4, 5, 6);  // 2015-02-03 00:00:00
-//   civil_day c(2015);                 // 2015-01-01 00:00:00
-//
-//   civil_second ss(2015, 2, 3, 4, 5, 6);  // 2015-02-03 04:05:06
-//   civil_minute mm(ss);                   // 2015-02-03 04:05:00
-//   civil_hour hh(mm);                     // 2015-02-03 04:00:00
-//   civil_day d(hh);                       // 2015-02-03 00:00:00
-//   civil_month m(d);                      // 2015-02-01 00:00:00
-//   civil_year y(m);                       // 2015-01-01 00:00:00
-//
-//   m = civil_month(y);     // 2015-01-01 00:00:00
-//   d = civil_day(m);       // 2015-01-01 00:00:00
-//   hh = civil_hour(d);     // 2015-01-01 00:00:00
-//   mm = civil_minute(hh);  // 2015-01-01 00:00:00
-//   ss = civil_second(mm);  // 2015-01-01 00:00:00
-//
-// ALIGNMENT CONVERSION:
-//
-// The alignment of a civil-time object cannot change, but the object may be
-// used to construct a new object with a different alignment. This is referred
-// to as "realigning". When realigning to a type with the same or more
-// precision (e.g., civil_day -> civil_second), the conversion may be
-// performed implicitly since no information is lost. However, if information
-// could be discarded (e.g., civil_second -> civil_day), the conversion must
-// be explicit at the call site.
-//
-//   void fun(const civil_day& day);
-//
-//   civil_second cs;
-//   fun(cs);  // Won't compile because data may be discarded
-//   fun(civil_day(cs));  // OK: explicit conversion
-//
-//   civil_day cd;
-//   fun(cd);  // OK: no conversion needed
-//
-//   civil_month cm;
-//   fun(cm);  // OK: implicit conversion to civil_day
-//
-// NORMALIZATION:
-//
-// Integer arguments passed to the constructor may be out-of-range, in which
-// case they are normalized to produce a valid civil-time object. This enables
-// natural arithmetic on constructor arguments without worrying about the
-// field's range. Normalization guarantees that there are no invalid
-// civil-time objects.
-//
-//   civil_day d(2016, 10, 32);  // Out-of-range day; normalized to 2016-11-01
-//
-// Note: If normalization is undesired, you can signal an error by comparing
-// the constructor arguments to the normalized values returned by the YMDHMS
-// properties.
-//
-// PROPERTIES:
-//
-// All civil-time types have accessors for all six of the civil-time fields:
-// year, month, day, hour, minute, and second. Recall that fields inferior to
-// the type's alignment will be set to their minimum valid value.
-//
-//   civil_day d(2015, 6, 28);
-//   // d.year() == 2015
-//   // d.month() == 6
-//   // d.day() == 28
-//   // d.hour() == 0
-//   // d.minute() == 0
-//   // d.second() == 0
-//
-// COMPARISON:
-//
-// Comparison always considers all six YMDHMS fields, regardless of the type's
-// alignment. Comparison between differently aligned civil-time types is
-// allowed.
-//
-//   civil_day feb_3(2015, 2, 3);  // 2015-02-03 00:00:00
-//   civil_day mar_4(2015, 3, 4);  // 2015-03-04 00:00:00
-//   // feb_3 < mar_4
-//   // civil_year(feb_3) == civil_year(mar_4)
-//
-//   civil_second feb_3_noon(2015, 2, 3, 12, 0, 0);  // 2015-02-03 12:00:00
-//   // feb_3 < feb_3_noon
-//   // feb_3 == civil_day(feb_3_noon)
-//
-//   // Iterates all the days of February 2015.
-//   for (civil_day d(2015, 2, 1); d < civil_month(2015, 3); ++d) {
-//     // ...
-//   }
-//
-// STREAMING:
-//
-// Each civil-time type may be sent to an output stream using operator<<().
-// The output format follows the pattern "YYYY-MM-DDThh:mm:ss" where fields
-// inferior to the type's alignment are omitted.
-//
-//   civil_second cs(2015, 2, 3, 4, 5, 6);
-//   std::cout << cs << "\n";  // Outputs: 2015-02-03T04:05:06
-//
-//   civil_day cd(cs);
-//   std::cout << cd << "\n";  // Outputs: 2015-02-03
-//
-//   civil_year cy(cs);
-//   std::cout << cy << "\n";  // Outputs: 2015
-//
-// ARITHMETIC:
-//
-// Civil-time types support natural arithmetic operators such as addition,
-// subtraction, and difference. Arithmetic operates on the civil-time field
-// indicated in the type's name. Difference requires arguments with the same
-// alignment and returns the answer in units of the alignment.
-//
-//   civil_day a(2015, 2, 3);
-//   ++a;                         // 2015-02-04 00:00:00
-//   --a;                         // 2015-02-03 00:00:00
-//   civil_day b = a + 1;         // 2015-02-04 00:00:00
-//   civil_day c = 1 + b;         // 2015-02-05 00:00:00
-//   int n = c - a;               // n = 2 (civil days)
-//   int m = c - civil_month(c);  // Won't compile: different types.
-//
-// EXAMPLE: Adding a month to January 31.
-//
-// One of the classic questions that arises when considering a civil-time
-// library (or a date library or a date/time library) is this: "What happens
-// when you add a month to January 31?" This is an interesting question
-// because there could be a number of possible answers:
-//
-//   1. March 3 (or 2 if a leap year). This may make sense if the operation
-//      wants the equivalent of February 31.
-//   2. February 28 (or 29 if a leap year). This may make sense if the operation
-//      wants the last day of January to go to the last day of February.
-//   3. Error. The caller may get some error, an exception, an invalid date
-//      object, or maybe false is returned. This may make sense because there is
-//      no single unambiguously correct answer to the question.
-//
-// Practically speaking, any answer that is not what the programmer intended
-// is the wrong answer.
-//
-// This civil-time library avoids the problem by making it impossible to ask
-// ambiguous questions. All civil-time objects are aligned to a particular
-// civil-field boundary (such as aligned to a year, month, day, hour, minute,
-// or second), and arithmetic operates on the field to which the object is
-// aligned. This means that in order to "add a month" the object must first be
-// aligned to a month boundary, which is equivalent to the first day of that
-// month.
-//
-// Of course, there are ways to compute an answer the question at hand using
-// this civil-time library, but they require the programmer to be explicit
-// about the answer they expect. To illustrate, let's see how to compute all
-// three of the above possible answers to the question of "Jan 31 plus 1
-// month":
-//
-//   const civil_day d(2015, 1, 31);
-//
-//   // Answer 1:
-//   // Add 1 to the month field in the constructor, and rely on normalization.
-//   const auto ans_normalized = civil_day(d.year(), d.month() + 1, d.day());
-//   // ans_normalized == 2015-03-03 (aka Feb 31)
-//
-//   // Answer 2:
-//   // Add 1 to month field, capping to the end of next month.
-//   const auto next_month = civil_month(d) + 1;
-//   const auto last_day_of_next_month = civil_day(next_month + 1) - 1;
-//   const auto ans_capped = std::min(ans_normalized, last_day_of_next_month);
-//   // ans_capped == 2015-02-28
-//
-//   // Answer 3:
-//   // Signal an error if the normalized answer is not in next month.
-//   if (civil_month(ans_normalized) != next_month) {
-//     // error, month overflow
-//   }
-//
-using civil_year = detail::civil_year;
-using civil_month = detail::civil_month;
-using civil_day = detail::civil_day;
-using civil_hour = detail::civil_hour;
-using civil_minute = detail::civil_minute;
-using civil_second = detail::civil_second;
-
-// An enum class with members monday, tuesday, wednesday, thursday, friday,
-// saturday, and sunday. These enum values may be sent to an output stream
-// using operator<<(). The result is the full weekday name in English with a
-// leading capital letter.
-//
-//   weekday wd = weekday::thursday;
-//   std::cout << wd << "\n";  // Outputs: Thursday
-//
-using detail::weekday;
-
-// Returns the weekday for the given civil-time value.
-//
-//   civil_day a(2015, 8, 13);
-//   weekday wd = get_weekday(a);  // wd == weekday::thursday
-//
-using detail::get_weekday;
-
-// Returns the civil_day that strictly follows or precedes the given
-// civil_day, and that falls on the given weekday.
-//
-// For example, given:
-//
-//     August 2015
-// Su Mo Tu We Th Fr Sa
-//                    1
-//  2  3  4  5  6  7  8
-//  9 10 11 12 13 14 15
-// 16 17 18 19 20 21 22
-// 23 24 25 26 27 28 29
-// 30 31
-//
-//   civil_day a(2015, 8, 13);  // get_weekday(a) == weekday::thursday
-//   civil_day b = next_weekday(a, weekday::thursday);  // b = 2015-08-20
-//   civil_day c = prev_weekday(a, weekday::thursday);  // c = 2015-08-06
-//
-//   civil_day d = ...
-//   // Gets the following Thursday if d is not already Thursday
-//   civil_day thurs1 = next_weekday(d - 1, weekday::thursday);
-//   // Gets the previous Thursday if d is not already Thursday
-//   civil_day thurs2 = prev_weekday(d + 1, weekday::thursday);
-//
-using detail::next_weekday;
-using detail::prev_weekday;
-
-// Returns the day-of-year for the given civil-time value.
-//
-//   civil_day a(2015, 1, 1);
-//   int yd_jan_1 = get_yearday(a);   // yd_jan_1 = 1
-//   civil_day b(2015, 12, 31);
-//   int yd_dec_31 = get_yearday(b);  // yd_dec_31 = 365
-//
-using detail::get_yearday;
-
-}  // namespace cctz
-}  // namespace time_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TIME_INTERNAL_CCTZ_CIVIL_TIME_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/civil_time_detail.h b/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/civil_time_detail.h
deleted file mode 100644
index 2b0aed56..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/civil_time_detail.h
+++ /dev/null
@@ -1,632 +0,0 @@
-// Copyright 2016 Google Inc. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//   https://www.apache.org/licenses/LICENSE-2.0
-//
-//   Unless required by applicable law or agreed to in writing, software
-//   distributed under the License is distributed on an "AS IS" BASIS,
-//   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-//   See the License for the specific language governing permissions and
-//   limitations under the License.
-
-#ifndef ABSL_TIME_INTERNAL_CCTZ_CIVIL_TIME_DETAIL_H_
-#define ABSL_TIME_INTERNAL_CCTZ_CIVIL_TIME_DETAIL_H_
-
-#include <cstdint>
-#include <limits>
-#include <ostream>
-#include <type_traits>
-
-#include "absl/base/config.h"
-
-// Disable constexpr support unless we are in C++14 mode.
-#if __cpp_constexpr >= 201304 || (defined(_MSC_VER) && _MSC_VER >= 1910)
-#define CONSTEXPR_D constexpr  // data
-#define CONSTEXPR_F constexpr  // function
-#define CONSTEXPR_M constexpr  // member
-#else
-#define CONSTEXPR_D const
-#define CONSTEXPR_F inline
-#define CONSTEXPR_M
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace time_internal {
-namespace cctz {
-
-// Support years that at least span the range of 64-bit time_t values.
-using year_t = std::int_fast64_t;
-
-// Type alias that indicates an argument is not normalized (e.g., the
-// constructor parameters and operands/results of addition/subtraction).
-using diff_t = std::int_fast64_t;
-
-namespace detail {
-
-// Type aliases that indicate normalized argument values.
-using month_t = std::int_fast8_t;   // [1:12]
-using day_t = std::int_fast8_t;     // [1:31]
-using hour_t = std::int_fast8_t;    // [0:23]
-using minute_t = std::int_fast8_t;  // [0:59]
-using second_t = std::int_fast8_t;  // [0:59]
-
-// Normalized civil-time fields: Y-M-D HH:MM:SS.
-struct fields {
-  CONSTEXPR_M fields(year_t year, month_t month, day_t day, hour_t hour,
-                     minute_t minute, second_t second)
-      : y(year), m(month), d(day), hh(hour), mm(minute), ss(second) {}
-  std::int_least64_t y;
-  std::int_least8_t m;
-  std::int_least8_t d;
-  std::int_least8_t hh;
-  std::int_least8_t mm;
-  std::int_least8_t ss;
-};
-
-struct second_tag {};
-struct minute_tag : second_tag {};
-struct hour_tag : minute_tag {};
-struct day_tag : hour_tag {};
-struct month_tag : day_tag {};
-struct year_tag : month_tag {};
-
-////////////////////////////////////////////////////////////////////////
-
-// Field normalization (without avoidable overflow).
-
-namespace impl {
-
-CONSTEXPR_F bool is_leap_year(year_t y) noexcept {
-  return y % 4 == 0 && (y % 100 != 0 || y % 400 == 0);
-}
-CONSTEXPR_F int year_index(year_t y, month_t m) noexcept {
-  const int yi = static_cast<int>((y + (m > 2)) % 400);
-  return yi < 0 ? yi + 400 : yi;
-}
-CONSTEXPR_F int days_per_century(int yi) noexcept {
-  return 36524 + (yi == 0 || yi > 300);
-}
-CONSTEXPR_F int days_per_4years(int yi) noexcept {
-  return 1460 + (yi == 0 || yi > 300 || (yi - 1) % 100 < 96);
-}
-CONSTEXPR_F int days_per_year(year_t y, month_t m) noexcept {
-  return is_leap_year(y + (m > 2)) ? 366 : 365;
-}
-CONSTEXPR_F int days_per_month(year_t y, month_t m) noexcept {
-  CONSTEXPR_D int k_days_per_month[1 + 12] = {
-      -1, 31, 28, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31  // non leap year
-  };
-  return k_days_per_month[m] + (m == 2 && is_leap_year(y));
-}
-
-CONSTEXPR_F fields n_day(year_t y, month_t m, diff_t d, diff_t cd, hour_t hh,
-                         minute_t mm, second_t ss) noexcept {
-  year_t ey = y % 400;
-  const year_t oey = ey;
-  ey += (cd / 146097) * 400;
-  cd %= 146097;
-  if (cd < 0) {
-    ey -= 400;
-    cd += 146097;
-  }
-  ey += (d / 146097) * 400;
-  d = d % 146097 + cd;
-  if (d > 0) {
-    if (d > 146097) {
-      ey += 400;
-      d -= 146097;
-    }
-  } else {
-    if (d > -365) {
-      // We often hit the previous year when stepping a civil time backwards,
-      // so special case it to avoid counting up by 100/4/1-year chunks.
-      ey -= 1;
-      d += days_per_year(ey, m);
-    } else {
-      ey -= 400;
-      d += 146097;
-    }
-  }
-  if (d > 365) {
-    int yi = year_index(ey, m);  // Index into Gregorian 400 year cycle.
-    for (;;) {
-      int n = days_per_century(yi);
-      if (d <= n) break;
-      d -= n;
-      ey += 100;
-      yi += 100;
-      if (yi >= 400) yi -= 400;
-    }
-    for (;;) {
-      int n = days_per_4years(yi);
-      if (d <= n) break;
-      d -= n;
-      ey += 4;
-      yi += 4;
-      if (yi >= 400) yi -= 400;
-    }
-    for (;;) {
-      int n = days_per_year(ey, m);
-      if (d <= n) break;
-      d -= n;
-      ++ey;
-    }
-  }
-  if (d > 28) {
-    for (;;) {
-      int n = days_per_month(ey, m);
-      if (d <= n) break;
-      d -= n;
-      if (++m > 12) {
-        ++ey;
-        m = 1;
-      }
-    }
-  }
-  return fields(y + (ey - oey), m, static_cast<day_t>(d), hh, mm, ss);
-}
-CONSTEXPR_F fields n_mon(year_t y, diff_t m, diff_t d, diff_t cd, hour_t hh,
-                         minute_t mm, second_t ss) noexcept {
-  if (m != 12) {
-    y += m / 12;
-    m %= 12;
-    if (m <= 0) {
-      y -= 1;
-      m += 12;
-    }
-  }
-  return n_day(y, static_cast<month_t>(m), d, cd, hh, mm, ss);
-}
-CONSTEXPR_F fields n_hour(year_t y, diff_t m, diff_t d, diff_t cd, diff_t hh,
-                          minute_t mm, second_t ss) noexcept {
-  cd += hh / 24;
-  hh %= 24;
-  if (hh < 0) {
-    cd -= 1;
-    hh += 24;
-  }
-  return n_mon(y, m, d, cd, static_cast<hour_t>(hh), mm, ss);
-}
-CONSTEXPR_F fields n_min(year_t y, diff_t m, diff_t d, diff_t hh, diff_t ch,
-                         diff_t mm, second_t ss) noexcept {
-  ch += mm / 60;
-  mm %= 60;
-  if (mm < 0) {
-    ch -= 1;
-    mm += 60;
-  }
-  return n_hour(y, m, d, hh / 24 + ch / 24, hh % 24 + ch % 24,
-                static_cast<minute_t>(mm), ss);
-}
-CONSTEXPR_F fields n_sec(year_t y, diff_t m, diff_t d, diff_t hh, diff_t mm,
-                         diff_t ss) noexcept {
-  // Optimization for when (non-constexpr) fields are already normalized.
-  if (0 <= ss && ss < 60) {
-    const second_t nss = static_cast<second_t>(ss);
-    if (0 <= mm && mm < 60) {
-      const minute_t nmm = static_cast<minute_t>(mm);
-      if (0 <= hh && hh < 24) {
-        const hour_t nhh = static_cast<hour_t>(hh);
-        if (1 <= d && d <= 28 && 1 <= m && m <= 12) {
-          const day_t nd = static_cast<day_t>(d);
-          const month_t nm = static_cast<month_t>(m);
-          return fields(y, nm, nd, nhh, nmm, nss);
-        }
-        return n_mon(y, m, d, 0, nhh, nmm, nss);
-      }
-      return n_hour(y, m, d, hh / 24, hh % 24, nmm, nss);
-    }
-    return n_min(y, m, d, hh, mm / 60, mm % 60, nss);
-  }
-  diff_t cm = ss / 60;
-  ss %= 60;
-  if (ss < 0) {
-    cm -= 1;
-    ss += 60;
-  }
-  return n_min(y, m, d, hh, mm / 60 + cm / 60, mm % 60 + cm % 60,
-               static_cast<second_t>(ss));
-}
-
-}  // namespace impl
-
-////////////////////////////////////////////////////////////////////////
-
-// Increments the indicated (normalized) field by "n".
-CONSTEXPR_F fields step(second_tag, fields f, diff_t n) noexcept {
-  return impl::n_sec(f.y, f.m, f.d, f.hh, f.mm + n / 60, f.ss + n % 60);
-}
-CONSTEXPR_F fields step(minute_tag, fields f, diff_t n) noexcept {
-  return impl::n_min(f.y, f.m, f.d, f.hh + n / 60, 0, f.mm + n % 60, f.ss);
-}
-CONSTEXPR_F fields step(hour_tag, fields f, diff_t n) noexcept {
-  return impl::n_hour(f.y, f.m, f.d + n / 24, 0, f.hh + n % 24, f.mm, f.ss);
-}
-CONSTEXPR_F fields step(day_tag, fields f, diff_t n) noexcept {
-  return impl::n_day(f.y, f.m, f.d, n, f.hh, f.mm, f.ss);
-}
-CONSTEXPR_F fields step(month_tag, fields f, diff_t n) noexcept {
-  return impl::n_mon(f.y + n / 12, f.m + n % 12, f.d, 0, f.hh, f.mm, f.ss);
-}
-CONSTEXPR_F fields step(year_tag, fields f, diff_t n) noexcept {
-  return fields(f.y + n, f.m, f.d, f.hh, f.mm, f.ss);
-}
-
-////////////////////////////////////////////////////////////////////////
-
-namespace impl {
-
-// Returns (v * f + a) but avoiding intermediate overflow when possible.
-CONSTEXPR_F diff_t scale_add(diff_t v, diff_t f, diff_t a) noexcept {
-  return (v < 0) ? ((v + 1) * f + a) - f : ((v - 1) * f + a) + f;
-}
-
-// Map a (normalized) Y/M/D to the number of days before/after 1970-01-01.
-// Probably overflows for years outside [-292277022656:292277026595].
-CONSTEXPR_F diff_t ymd_ord(year_t y, month_t m, day_t d) noexcept {
-  const diff_t eyear = (m <= 2) ? y - 1 : y;
-  const diff_t era = (eyear >= 0 ? eyear : eyear - 399) / 400;
-  const diff_t yoe = eyear - era * 400;
-  const diff_t doy = (153 * (m + (m > 2 ? -3 : 9)) + 2) / 5 + d - 1;
-  const diff_t doe = yoe * 365 + yoe / 4 - yoe / 100 + doy;
-  return era * 146097 + doe - 719468;
-}
-
-// Returns the difference in days between two normalized Y-M-D tuples.
-// ymd_ord() will encounter integer overflow given extreme year values,
-// yet the difference between two such extreme values may actually be
-// small, so we take a little care to avoid overflow when possible by
-// exploiting the 146097-day cycle.
-CONSTEXPR_F diff_t day_difference(year_t y1, month_t m1, day_t d1, year_t y2,
-                                  month_t m2, day_t d2) noexcept {
-  const diff_t a_c4_off = y1 % 400;
-  const diff_t b_c4_off = y2 % 400;
-  diff_t c4_diff = (y1 - a_c4_off) - (y2 - b_c4_off);
-  diff_t delta = ymd_ord(a_c4_off, m1, d1) - ymd_ord(b_c4_off, m2, d2);
-  if (c4_diff > 0 && delta < 0) {
-    delta += 2 * 146097;
-    c4_diff -= 2 * 400;
-  } else if (c4_diff < 0 && delta > 0) {
-    delta -= 2 * 146097;
-    c4_diff += 2 * 400;
-  }
-  return (c4_diff / 400 * 146097) + delta;
-}
-
-}  // namespace impl
-
-// Returns the difference between fields structs using the indicated unit.
-CONSTEXPR_F diff_t difference(year_tag, fields f1, fields f2) noexcept {
-  return f1.y - f2.y;
-}
-CONSTEXPR_F diff_t difference(month_tag, fields f1, fields f2) noexcept {
-  return impl::scale_add(difference(year_tag{}, f1, f2), 12, (f1.m - f2.m));
-}
-CONSTEXPR_F diff_t difference(day_tag, fields f1, fields f2) noexcept {
-  return impl::day_difference(f1.y, f1.m, f1.d, f2.y, f2.m, f2.d);
-}
-CONSTEXPR_F diff_t difference(hour_tag, fields f1, fields f2) noexcept {
-  return impl::scale_add(difference(day_tag{}, f1, f2), 24, (f1.hh - f2.hh));
-}
-CONSTEXPR_F diff_t difference(minute_tag, fields f1, fields f2) noexcept {
-  return impl::scale_add(difference(hour_tag{}, f1, f2), 60, (f1.mm - f2.mm));
-}
-CONSTEXPR_F diff_t difference(second_tag, fields f1, fields f2) noexcept {
-  return impl::scale_add(difference(minute_tag{}, f1, f2), 60, f1.ss - f2.ss);
-}
-
-////////////////////////////////////////////////////////////////////////
-
-// Aligns the (normalized) fields struct to the indicated field.
-CONSTEXPR_F fields align(second_tag, fields f) noexcept { return f; }
-CONSTEXPR_F fields align(minute_tag, fields f) noexcept {
-  return fields{f.y, f.m, f.d, f.hh, f.mm, 0};
-}
-CONSTEXPR_F fields align(hour_tag, fields f) noexcept {
-  return fields{f.y, f.m, f.d, f.hh, 0, 0};
-}
-CONSTEXPR_F fields align(day_tag, fields f) noexcept {
-  return fields{f.y, f.m, f.d, 0, 0, 0};
-}
-CONSTEXPR_F fields align(month_tag, fields f) noexcept {
-  return fields{f.y, f.m, 1, 0, 0, 0};
-}
-CONSTEXPR_F fields align(year_tag, fields f) noexcept {
-  return fields{f.y, 1, 1, 0, 0, 0};
-}
-
-////////////////////////////////////////////////////////////////////////
-
-namespace impl {
-
-template <typename H>
-H AbslHashValueImpl(second_tag, H h, fields f) {
-  return H::combine(std::move(h), f.y, f.m, f.d, f.hh, f.mm, f.ss);
-}
-template <typename H>
-H AbslHashValueImpl(minute_tag, H h, fields f) {
-  return H::combine(std::move(h), f.y, f.m, f.d, f.hh, f.mm);
-}
-template <typename H>
-H AbslHashValueImpl(hour_tag, H h, fields f) {
-  return H::combine(std::move(h), f.y, f.m, f.d, f.hh);
-}
-template <typename H>
-H AbslHashValueImpl(day_tag, H h, fields f) {
-  return H::combine(std::move(h), f.y, f.m, f.d);
-}
-template <typename H>
-H AbslHashValueImpl(month_tag, H h, fields f) {
-  return H::combine(std::move(h), f.y, f.m);
-}
-template <typename H>
-H AbslHashValueImpl(year_tag, H h, fields f) {
-  return H::combine(std::move(h), f.y);
-}
-
-}  // namespace impl
-
-////////////////////////////////////////////////////////////////////////
-
-template <typename T>
-class civil_time {
- public:
-  explicit CONSTEXPR_M civil_time(year_t y, diff_t m = 1, diff_t d = 1,
-                                  diff_t hh = 0, diff_t mm = 0,
-                                  diff_t ss = 0) noexcept
-      : civil_time(impl::n_sec(y, m, d, hh, mm, ss)) {}
-
-  CONSTEXPR_M civil_time() noexcept : f_{1970, 1, 1, 0, 0, 0} {}
-  civil_time(const civil_time&) = default;
-  civil_time& operator=(const civil_time&) = default;
-
-  // Conversion between civil times of different alignment. Conversion to
-  // a more precise alignment is allowed implicitly (e.g., day -> hour),
-  // but conversion where information is discarded must be explicit
-  // (e.g., second -> minute).
-  template <typename U, typename S>
-  using preserves_data =
-      typename std::enable_if<std::is_base_of<U, S>::value>::type;
-  template <typename U>
-  CONSTEXPR_M civil_time(const civil_time<U>& ct,
-                         preserves_data<T, U>* = nullptr) noexcept
-      : civil_time(ct.f_) {}
-  template <typename U>
-  explicit CONSTEXPR_M civil_time(const civil_time<U>& ct,
-                                  preserves_data<U, T>* = nullptr) noexcept
-      : civil_time(ct.f_) {}
-
-  // Factories for the maximum/minimum representable civil_time.
-  static CONSTEXPR_F auto(max)() -> civil_time {
-    const auto max_year = (std::numeric_limits<std::int_least64_t>::max)();
-    return civil_time(max_year, 12, 31, 23, 59, 59);
-  }
-  static CONSTEXPR_F auto(min)() -> civil_time {
-    const auto min_year = (std::numeric_limits<std::int_least64_t>::min)();
-    return civil_time(min_year, 1, 1, 0, 0, 0);
-  }
-
-  // Field accessors.  Note: All but year() return an int.
-  CONSTEXPR_M year_t year() const noexcept { return f_.y; }
-  CONSTEXPR_M int month() const noexcept { return f_.m; }
-  CONSTEXPR_M int day() const noexcept { return f_.d; }
-  CONSTEXPR_M int hour() const noexcept { return f_.hh; }
-  CONSTEXPR_M int minute() const noexcept { return f_.mm; }
-  CONSTEXPR_M int second() const noexcept { return f_.ss; }
-
-  // Assigning arithmetic.
-  CONSTEXPR_M civil_time& operator+=(diff_t n) noexcept {
-    return *this = *this + n;
-  }
-  CONSTEXPR_M civil_time& operator-=(diff_t n) noexcept {
-    return *this = *this - n;
-  }
-  CONSTEXPR_M civil_time& operator++() noexcept { return *this += 1; }
-  CONSTEXPR_M civil_time operator++(int) noexcept {
-    const civil_time a = *this;
-    ++*this;
-    return a;
-  }
-  CONSTEXPR_M civil_time& operator--() noexcept { return *this -= 1; }
-  CONSTEXPR_M civil_time operator--(int) noexcept {
-    const civil_time a = *this;
-    --*this;
-    return a;
-  }
-
-  // Binary arithmetic operators.
-  friend CONSTEXPR_F civil_time operator+(civil_time a, diff_t n) noexcept {
-    return civil_time(step(T{}, a.f_, n));
-  }
-  friend CONSTEXPR_F civil_time operator+(diff_t n, civil_time a) noexcept {
-    return a + n;
-  }
-  friend CONSTEXPR_F civil_time operator-(civil_time a, diff_t n) noexcept {
-    return n != (std::numeric_limits<diff_t>::min)()
-               ? civil_time(step(T{}, a.f_, -n))
-               : civil_time(step(T{}, step(T{}, a.f_, -(n + 1)), 1));
-  }
-  friend CONSTEXPR_F diff_t operator-(civil_time lhs, civil_time rhs) noexcept {
-    return difference(T{}, lhs.f_, rhs.f_);
-  }
-
-  template <typename H>
-  friend H AbslHashValue(H h, civil_time a) {
-    return impl::AbslHashValueImpl(T{}, std::move(h), a.f_);
-  }
-
- private:
-  // All instantiations of this template are allowed to call the following
-  // private constructor and access the private fields member.
-  template <typename U>
-  friend class civil_time;
-
-  // The designated constructor that all others eventually call.
-  explicit CONSTEXPR_M civil_time(fields f) noexcept : f_(align(T{}, f)) {}
-
-  fields f_;
-};
-
-// Disallows difference between differently aligned types.
-// auto n = civil_day(...) - civil_hour(...);  // would be confusing.
-template <typename T, typename U>
-CONSTEXPR_F diff_t operator-(civil_time<T>, civil_time<U>) = delete;
-
-using civil_year = civil_time<year_tag>;
-using civil_month = civil_time<month_tag>;
-using civil_day = civil_time<day_tag>;
-using civil_hour = civil_time<hour_tag>;
-using civil_minute = civil_time<minute_tag>;
-using civil_second = civil_time<second_tag>;
-
-////////////////////////////////////////////////////////////////////////
-
-// Relational operators that work with differently aligned objects.
-// Always compares all six fields.
-template <typename T1, typename T2>
-CONSTEXPR_F bool operator<(const civil_time<T1>& lhs,
-                           const civil_time<T2>& rhs) noexcept {
-  return (
-      lhs.year() < rhs.year() ||
-      (lhs.year() == rhs.year() &&
-       (lhs.month() < rhs.month() ||
-        (lhs.month() == rhs.month() &&
-         (lhs.day() < rhs.day() || (lhs.day() == rhs.day() &&
-                                    (lhs.hour() < rhs.hour() ||
-                                     (lhs.hour() == rhs.hour() &&
-                                      (lhs.minute() < rhs.minute() ||
-                                       (lhs.minute() == rhs.minute() &&
-                                        (lhs.second() < rhs.second())))))))))));
-}
-template <typename T1, typename T2>
-CONSTEXPR_F bool operator<=(const civil_time<T1>& lhs,
-                            const civil_time<T2>& rhs) noexcept {
-  return !(rhs < lhs);
-}
-template <typename T1, typename T2>
-CONSTEXPR_F bool operator>=(const civil_time<T1>& lhs,
-                            const civil_time<T2>& rhs) noexcept {
-  return !(lhs < rhs);
-}
-template <typename T1, typename T2>
-CONSTEXPR_F bool operator>(const civil_time<T1>& lhs,
-                           const civil_time<T2>& rhs) noexcept {
-  return rhs < lhs;
-}
-template <typename T1, typename T2>
-CONSTEXPR_F bool operator==(const civil_time<T1>& lhs,
-                            const civil_time<T2>& rhs) noexcept {
-  return lhs.year() == rhs.year() && lhs.month() == rhs.month() &&
-         lhs.day() == rhs.day() && lhs.hour() == rhs.hour() &&
-         lhs.minute() == rhs.minute() && lhs.second() == rhs.second();
-}
-template <typename T1, typename T2>
-CONSTEXPR_F bool operator!=(const civil_time<T1>& lhs,
-                            const civil_time<T2>& rhs) noexcept {
-  return !(lhs == rhs);
-}
-
-////////////////////////////////////////////////////////////////////////
-
-enum class weekday {
-  monday,
-  tuesday,
-  wednesday,
-  thursday,
-  friday,
-  saturday,
-  sunday,
-};
-
-CONSTEXPR_F weekday get_weekday(const civil_second& cs) noexcept {
-  CONSTEXPR_D weekday k_weekday_by_mon_off[13] = {
-      weekday::monday,    weekday::tuesday,  weekday::wednesday,
-      weekday::thursday,  weekday::friday,   weekday::saturday,
-      weekday::sunday,    weekday::monday,   weekday::tuesday,
-      weekday::wednesday, weekday::thursday, weekday::friday,
-      weekday::saturday,
-  };
-  CONSTEXPR_D int k_weekday_offsets[1 + 12] = {
-      -1, 0, 3, 2, 5, 0, 3, 5, 1, 4, 6, 2, 4,
-  };
-  year_t wd = 2400 + (cs.year() % 400) - (cs.month() < 3);
-  wd += wd / 4 - wd / 100 + wd / 400;
-  wd += k_weekday_offsets[cs.month()] + cs.day();
-  return k_weekday_by_mon_off[wd % 7 + 6];
-}
-
-////////////////////////////////////////////////////////////////////////
-
-CONSTEXPR_F civil_day next_weekday(civil_day cd, weekday wd) noexcept {
-  CONSTEXPR_D weekday k_weekdays_forw[14] = {
-      weekday::monday,    weekday::tuesday,  weekday::wednesday,
-      weekday::thursday,  weekday::friday,   weekday::saturday,
-      weekday::sunday,    weekday::monday,   weekday::tuesday,
-      weekday::wednesday, weekday::thursday, weekday::friday,
-      weekday::saturday,  weekday::sunday,
-  };
-  weekday base = get_weekday(cd);
-  for (int i = 0;; ++i) {
-    if (base == k_weekdays_forw[i]) {
-      for (int j = i + 1;; ++j) {
-        if (wd == k_weekdays_forw[j]) {
-          return cd + (j - i);
-        }
-      }
-    }
-  }
-}
-
-CONSTEXPR_F civil_day prev_weekday(civil_day cd, weekday wd) noexcept {
-  CONSTEXPR_D weekday k_weekdays_back[14] = {
-      weekday::sunday,   weekday::saturday,  weekday::friday,
-      weekday::thursday, weekday::wednesday, weekday::tuesday,
-      weekday::monday,   weekday::sunday,    weekday::saturday,
-      weekday::friday,   weekday::thursday,  weekday::wednesday,
-      weekday::tuesday,  weekday::monday,
-  };
-  weekday base = get_weekday(cd);
-  for (int i = 0;; ++i) {
-    if (base == k_weekdays_back[i]) {
-      for (int j = i + 1;; ++j) {
-        if (wd == k_weekdays_back[j]) {
-          return cd - (j - i);
-        }
-      }
-    }
-  }
-}
-
-CONSTEXPR_F int get_yearday(const civil_second& cs) noexcept {
-  CONSTEXPR_D int k_month_offsets[1 + 12] = {
-      -1, 0, 31, 59, 90, 120, 151, 181, 212, 243, 273, 304, 334,
-  };
-  const int feb29 = (cs.month() > 2 && impl::is_leap_year(cs.year()));
-  return k_month_offsets[cs.month()] + feb29 + cs.day();
-}
-
-////////////////////////////////////////////////////////////////////////
-
-std::ostream& operator<<(std::ostream& os, const civil_year& y);
-std::ostream& operator<<(std::ostream& os, const civil_month& m);
-std::ostream& operator<<(std::ostream& os, const civil_day& d);
-std::ostream& operator<<(std::ostream& os, const civil_hour& h);
-std::ostream& operator<<(std::ostream& os, const civil_minute& m);
-std::ostream& operator<<(std::ostream& os, const civil_second& s);
-std::ostream& operator<<(std::ostream& os, weekday wd);
-
-}  // namespace detail
-}  // namespace cctz
-}  // namespace time_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#undef CONSTEXPR_M
-#undef CONSTEXPR_F
-#undef CONSTEXPR_D
-
-#endif  // ABSL_TIME_INTERNAL_CCTZ_CIVIL_TIME_DETAIL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/time_zone.h b/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/time_zone.h
deleted file mode 100644
index b2b0cf6f..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/time_zone.h
+++ /dev/null
@@ -1,460 +0,0 @@
-// Copyright 2016 Google Inc. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//   https://www.apache.org/licenses/LICENSE-2.0
-//
-//   Unless required by applicable law or agreed to in writing, software
-//   distributed under the License is distributed on an "AS IS" BASIS,
-//   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-//   See the License for the specific language governing permissions and
-//   limitations under the License.
-
-// A library for translating between absolute times (represented by
-// std::chrono::time_points of the std::chrono::system_clock) and civil
-// times (represented by cctz::civil_second) using the rules defined by
-// a time zone (cctz::time_zone).
-
-#ifndef ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_H_
-#define ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_H_
-
-#include <chrono>
-#include <cstdint>
-#include <limits>
-#include <ratio>  // NOLINT: We use std::ratio in this header
-#include <string>
-#include <utility>
-
-#include "absl/base/config.h"
-#include "absl/time/internal/cctz/include/cctz/civil_time.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace time_internal {
-namespace cctz {
-
-// Convenience aliases. Not intended as public API points.
-template <typename D>
-using time_point = std::chrono::time_point<std::chrono::system_clock, D>;
-using seconds = std::chrono::duration<std::int_fast64_t>;
-using sys_seconds = seconds;  // Deprecated.  Use cctz::seconds instead.
-
-namespace detail {
-template <typename D>
-std::pair<time_point<seconds>, D> split_seconds(const time_point<D>& tp);
-std::pair<time_point<seconds>, seconds> split_seconds(
-    const time_point<seconds>& tp);
-}  // namespace detail
-
-// cctz::time_zone is an opaque, small, value-type class representing a
-// geo-political region within which particular rules are used for mapping
-// between absolute and civil times. Time zones are named using the TZ
-// identifiers from the IANA Time Zone Database, such as "America/Los_Angeles"
-// or "Australia/Sydney". Time zones are created from factory functions such
-// as load_time_zone(). Note: strings like "PST" and "EDT" are not valid TZ
-// identifiers.
-//
-// Example:
-//   cctz::time_zone utc = cctz::utc_time_zone();
-//   cctz::time_zone pst = cctz::fixed_time_zone(std::chrono::hours(-8));
-//   cctz::time_zone loc = cctz::local_time_zone();
-//   cctz::time_zone lax;
-//   if (!cctz::load_time_zone("America/Los_Angeles", &lax)) { ... }
-//
-// See also:
-// - http://www.iana.org/time-zones
-// - https://en.wikipedia.org/wiki/Zoneinfo
-class time_zone {
- public:
-  time_zone() : time_zone(nullptr) {}  // Equivalent to UTC
-  time_zone(const time_zone&) = default;
-  time_zone& operator=(const time_zone&) = default;
-
-  std::string name() const;
-
-  // An absolute_lookup represents the civil time (cctz::civil_second) within
-  // this time_zone at the given absolute time (time_point). There are
-  // additionally a few other fields that may be useful when working with
-  // older APIs, such as std::tm.
-  //
-  // Example:
-  //   const cctz::time_zone tz = ...
-  //   const auto tp = std::chrono::system_clock::now();
-  //   const cctz::time_zone::absolute_lookup al = tz.lookup(tp);
-  struct absolute_lookup {
-    civil_second cs;
-    // Note: The following fields exist for backward compatibility with older
-    // APIs. Accessing these fields directly is a sign of imprudent logic in
-    // the calling code. Modern time-related code should only access this data
-    // indirectly by way of cctz::format().
-    int offset;        // civil seconds east of UTC
-    bool is_dst;       // is offset non-standard?
-    const char* abbr;  // time-zone abbreviation (e.g., "PST")
-  };
-  absolute_lookup lookup(const time_point<seconds>& tp) const;
-  template <typename D>
-  absolute_lookup lookup(const time_point<D>& tp) const {
-    return lookup(detail::split_seconds(tp).first);
-  }
-
-  // A civil_lookup represents the absolute time(s) (time_point) that
-  // correspond to the given civil time (cctz::civil_second) within this
-  // time_zone. Usually the given civil time represents a unique instant
-  // in time, in which case the conversion is unambiguous. However,
-  // within this time zone, the given civil time may be skipped (e.g.,
-  // during a positive UTC offset shift), or repeated (e.g., during a
-  // negative UTC offset shift). To account for these possibilities,
-  // civil_lookup is richer than just a single time_point.
-  //
-  // In all cases the civil_lookup::kind enum will indicate the nature
-  // of the given civil-time argument, and the pre, trans, and post
-  // members will give the absolute time answers using the pre-transition
-  // offset, the transition point itself, and the post-transition offset,
-  // respectively (all three times are equal if kind == UNIQUE). If any
-  // of these three absolute times is outside the representable range of a
-  // time_point<seconds> the field is set to its maximum/minimum value.
-  //
-  // Example:
-  //   cctz::time_zone lax;
-  //   if (!cctz::load_time_zone("America/Los_Angeles", &lax)) { ... }
-  //
-  //   // A unique civil time.
-  //   auto jan01 = lax.lookup(cctz::civil_second(2011, 1, 1, 0, 0, 0));
-  //   // jan01.kind == cctz::time_zone::civil_lookup::UNIQUE
-  //   // jan01.pre    is 2011/01/01 00:00:00 -0800
-  //   // jan01.trans  is 2011/01/01 00:00:00 -0800
-  //   // jan01.post   is 2011/01/01 00:00:00 -0800
-  //
-  //   // A Spring DST transition, when there is a gap in civil time.
-  //   auto mar13 = lax.lookup(cctz::civil_second(2011, 3, 13, 2, 15, 0));
-  //   // mar13.kind == cctz::time_zone::civil_lookup::SKIPPED
-  //   // mar13.pre   is 2011/03/13 03:15:00 -0700
-  //   // mar13.trans is 2011/03/13 03:00:00 -0700
-  //   // mar13.post  is 2011/03/13 01:15:00 -0800
-  //
-  //   // A Fall DST transition, when civil times are repeated.
-  //   auto nov06 = lax.lookup(cctz::civil_second(2011, 11, 6, 1, 15, 0));
-  //   // nov06.kind == cctz::time_zone::civil_lookup::REPEATED
-  //   // nov06.pre   is 2011/11/06 01:15:00 -0700
-  //   // nov06.trans is 2011/11/06 01:00:00 -0800
-  //   // nov06.post  is 2011/11/06 01:15:00 -0800
-  struct civil_lookup {
-    enum civil_kind {
-      UNIQUE,    // the civil time was singular (pre == trans == post)
-      SKIPPED,   // the civil time did not exist (pre >= trans > post)
-      REPEATED,  // the civil time was ambiguous (pre < trans <= post)
-    } kind;
-    time_point<seconds> pre;    // uses the pre-transition offset
-    time_point<seconds> trans;  // instant of civil-offset change
-    time_point<seconds> post;   // uses the post-transition offset
-  };
-  civil_lookup lookup(const civil_second& cs) const;
-
-  // Finds the time of the next/previous offset change in this time zone.
-  //
-  // By definition, next_transition(tp, &trans) returns false when tp has
-  // its maximum value, and prev_transition(tp, &trans) returns false
-  // when tp has its minimum value. If the zone has no transitions, the
-  // result will also be false no matter what the argument.
-  //
-  // Otherwise, when tp has its minimum value, next_transition(tp, &trans)
-  // returns true and sets trans to the first recorded transition. Chains
-  // of calls to next_transition()/prev_transition() will eventually return
-  // false, but it is unspecified exactly when next_transition(tp, &trans)
-  // jumps to false, or what time is set by prev_transition(tp, &trans) for
-  // a very distant tp.
-  //
-  // Note: Enumeration of time-zone transitions is for informational purposes
-  // only. Modern time-related code should not care about when offset changes
-  // occur.
-  //
-  // Example:
-  //   cctz::time_zone nyc;
-  //   if (!cctz::load_time_zone("America/New_York", &nyc)) { ... }
-  //   const auto now = std::chrono::system_clock::now();
-  //   auto tp = cctz::time_point<cctz::seconds>::min();
-  //   cctz::time_zone::civil_transition trans;
-  //   while (tp <= now && nyc.next_transition(tp, &trans)) {
-  //     // transition: trans.from -> trans.to
-  //     tp = nyc.lookup(trans.to).trans;
-  //   }
-  struct civil_transition {
-    civil_second from;  // the civil time we jump from
-    civil_second to;    // the civil time we jump to
-  };
-  bool next_transition(const time_point<seconds>& tp,
-                       civil_transition* trans) const;
-  template <typename D>
-  bool next_transition(const time_point<D>& tp, civil_transition* trans) const {
-    return next_transition(detail::split_seconds(tp).first, trans);
-  }
-  bool prev_transition(const time_point<seconds>& tp,
-                       civil_transition* trans) const;
-  template <typename D>
-  bool prev_transition(const time_point<D>& tp, civil_transition* trans) const {
-    return prev_transition(detail::split_seconds(tp).first, trans);
-  }
-
-  // version() and description() provide additional information about the
-  // time zone. The content of each of the returned strings is unspecified,
-  // however, when the IANA Time Zone Database is the underlying data source
-  // the version() string will be in the familar form (e.g, "2018e") or
-  // empty when unavailable.
-  //
-  // Note: These functions are for informational or testing purposes only.
-  std::string version() const;  // empty when unknown
-  std::string description() const;
-
-  // Relational operators.
-  friend bool operator==(time_zone lhs, time_zone rhs) {
-    return &lhs.effective_impl() == &rhs.effective_impl();
-  }
-  friend bool operator!=(time_zone lhs, time_zone rhs) { return !(lhs == rhs); }
-
-  template <typename H>
-  friend H AbslHashValue(H h, time_zone tz) {
-    return H::combine(std::move(h), &tz.effective_impl());
-  }
-
-  class Impl;
-
- private:
-  explicit time_zone(const Impl* impl) : impl_(impl) {}
-  const Impl& effective_impl() const;  // handles implicit UTC
-  const Impl* impl_;
-};
-
-// Loads the named time zone. May perform I/O on the initial load.
-// If the name is invalid, or some other kind of error occurs, returns
-// false and "*tz" is set to the UTC time zone.
-bool load_time_zone(const std::string& name, time_zone* tz);
-
-// Returns a time_zone representing UTC. Cannot fail.
-time_zone utc_time_zone();
-
-// Returns a time zone that is a fixed offset (seconds east) from UTC.
-// Note: If the absolute value of the offset is greater than 24 hours
-// you'll get UTC (i.e., zero offset) instead.
-time_zone fixed_time_zone(const seconds& offset);
-
-// Returns a time zone representing the local time zone. Falls back to UTC.
-// Note: local_time_zone.name() may only be something like "localtime".
-time_zone local_time_zone();
-
-// Returns the civil time (cctz::civil_second) within the given time zone at
-// the given absolute time (time_point). Since the additional fields provided
-// by the time_zone::absolute_lookup struct should rarely be needed in modern
-// code, this convert() function is simpler and should be preferred.
-template <typename D>
-inline civil_second convert(const time_point<D>& tp, const time_zone& tz) {
-  return tz.lookup(tp).cs;
-}
-
-// Returns the absolute time (time_point) that corresponds to the given civil
-// time within the given time zone. If the civil time is not unique (i.e., if
-// it was either repeated or non-existent), then the returned time_point is
-// the best estimate that preserves relative order. That is, this function
-// guarantees that if cs1 < cs2, then convert(cs1, tz) <= convert(cs2, tz).
-inline time_point<seconds> convert(const civil_second& cs,
-                                   const time_zone& tz) {
-  const time_zone::civil_lookup cl = tz.lookup(cs);
-  if (cl.kind == time_zone::civil_lookup::SKIPPED) return cl.trans;
-  return cl.pre;
-}
-
-namespace detail {
-using femtoseconds = std::chrono::duration<std::int_fast64_t, std::femto>;
-std::string format(const std::string&, const time_point<seconds>&,
-                   const femtoseconds&, const time_zone&);
-bool parse(const std::string&, const std::string&, const time_zone&,
-           time_point<seconds>*, femtoseconds*, std::string* err = nullptr);
-template <typename Rep, std::intmax_t Denom>
-bool join_seconds(
-    const time_point<seconds>& sec, const femtoseconds& fs,
-    time_point<std::chrono::duration<Rep, std::ratio<1, Denom>>>* tpp);
-template <typename Rep, std::intmax_t Num>
-bool join_seconds(
-    const time_point<seconds>& sec, const femtoseconds& fs,
-    time_point<std::chrono::duration<Rep, std::ratio<Num, 1>>>* tpp);
-template <typename Rep>
-bool join_seconds(
-    const time_point<seconds>& sec, const femtoseconds& fs,
-    time_point<std::chrono::duration<Rep, std::ratio<1, 1>>>* tpp);
-bool join_seconds(const time_point<seconds>& sec, const femtoseconds&,
-                  time_point<seconds>* tpp);
-}  // namespace detail
-
-// Formats the given time_point in the given cctz::time_zone according to
-// the provided format string. Uses strftime()-like formatting options,
-// with the following extensions:
-//
-//   - %Ez  - RFC3339-compatible numeric UTC offset (+hh:mm or -hh:mm)
-//   - %E*z - Full-resolution numeric UTC offset (+hh:mm:ss or -hh:mm:ss)
-//   - %E#S - Seconds with # digits of fractional precision
-//   - %E*S - Seconds with full fractional precision (a literal '*')
-//   - %E#f - Fractional seconds with # digits of precision
-//   - %E*f - Fractional seconds with full precision (a literal '*')
-//   - %E4Y - Four-character years (-999 ... -001, 0000, 0001 ... 9999)
-//   - %ET  - The RFC3339 "date-time" separator "T"
-//
-// Note that %E0S behaves like %S, and %E0f produces no characters. In
-// contrast %E*f always produces at least one digit, which may be '0'.
-//
-// Note that %Y produces as many characters as it takes to fully render the
-// year. A year outside of [-999:9999] when formatted with %E4Y will produce
-// more than four characters, just like %Y.
-//
-// Tip: Format strings should include the UTC offset (e.g., %z, %Ez, or %E*z)
-// so that the resulting string uniquely identifies an absolute time.
-//
-// Example:
-//   cctz::time_zone lax;
-//   if (!cctz::load_time_zone("America/Los_Angeles", &lax)) { ... }
-//   auto tp = cctz::convert(cctz::civil_second(2013, 1, 2, 3, 4, 5), lax);
-//   std::string f = cctz::format("%H:%M:%S", tp, lax);  // "03:04:05"
-//   f = cctz::format("%H:%M:%E3S", tp, lax);            // "03:04:05.000"
-template <typename D>
-inline std::string format(const std::string& fmt, const time_point<D>& tp,
-                          const time_zone& tz) {
-  const auto p = detail::split_seconds(tp);
-  const auto n = std::chrono::duration_cast<detail::femtoseconds>(p.second);
-  return detail::format(fmt, p.first, n, tz);
-}
-
-// Parses an input string according to the provided format string and
-// returns the corresponding time_point. Uses strftime()-like formatting
-// options, with the same extensions as cctz::format(), but with the
-// exceptions that %E#S is interpreted as %E*S, and %E#f as %E*f. %Ez
-// and %E*z also accept the same inputs, which (along with %z) includes
-// 'z' and 'Z' as synonyms for +00:00.  %ET accepts either 'T' or 't'.
-//
-// %Y consumes as many numeric characters as it can, so the matching data
-// should always be terminated with a non-numeric. %E4Y always consumes
-// exactly four characters, including any sign.
-//
-// Unspecified fields are taken from the default date and time of ...
-//
-//   "1970-01-01 00:00:00.0 +0000"
-//
-// For example, parsing a string of "15:45" (%H:%M) will return a time_point
-// that represents "1970-01-01 15:45:00.0 +0000".
-//
-// Note that parse() returns time instants, so it makes most sense to parse
-// fully-specified date/time strings that include a UTC offset (%z, %Ez, or
-// %E*z).
-//
-// Note also that parse() only heeds the fields year, month, day, hour,
-// minute, (fractional) second, and UTC offset. Other fields, like weekday (%a
-// or %A), while parsed for syntactic validity, are ignored in the conversion.
-//
-// Date and time fields that are out-of-range will be treated as errors rather
-// than normalizing them like cctz::civil_second() would do. For example, it
-// is an error to parse the date "Oct 32, 2013" because 32 is out of range.
-//
-// A second of ":60" is normalized to ":00" of the following minute with
-// fractional seconds discarded. The following table shows how the given
-// seconds and subseconds will be parsed:
-//
-//   "59.x" -> 59.x  // exact
-//   "60.x" -> 00.0  // normalized
-//   "00.x" -> 00.x  // exact
-//
-// Errors are indicated by returning false.
-//
-// Example:
-//   const cctz::time_zone tz = ...
-//   std::chrono::system_clock::time_point tp;
-//   if (cctz::parse("%Y-%m-%d", "2015-10-09", tz, &tp)) {
-//     ...
-//   }
-template <typename D>
-inline bool parse(const std::string& fmt, const std::string& input,
-                  const time_zone& tz, time_point<D>* tpp) {
-  time_point<seconds> sec;
-  detail::femtoseconds fs;
-  return detail::parse(fmt, input, tz, &sec, &fs) &&
-         detail::join_seconds(sec, fs, tpp);
-}
-
-namespace detail {
-
-// Split a time_point<D> into a time_point<seconds> and a D subseconds.
-// Undefined behavior if time_point<seconds> is not of sufficient range.
-// Note that this means it is UB to call cctz::time_zone::lookup(tp) or
-// cctz::format(fmt, tp, tz) with a time_point that is outside the range
-// of a 64-bit std::time_t.
-template <typename D>
-std::pair<time_point<seconds>, D> split_seconds(const time_point<D>& tp) {
-  auto sec = std::chrono::time_point_cast<seconds>(tp);
-  auto sub = tp - sec;
-  if (sub.count() < 0) {
-    sec -= seconds(1);
-    sub += seconds(1);
-  }
-  return {sec, std::chrono::duration_cast<D>(sub)};
-}
-
-inline std::pair<time_point<seconds>, seconds> split_seconds(
-    const time_point<seconds>& tp) {
-  return {tp, seconds::zero()};
-}
-
-// Join a time_point<seconds> and femto subseconds into a time_point<D>.
-// Floors to the resolution of time_point<D>. Returns false if time_point<D>
-// is not of sufficient range.
-template <typename Rep, std::intmax_t Denom>
-bool join_seconds(
-    const time_point<seconds>& sec, const femtoseconds& fs,
-    time_point<std::chrono::duration<Rep, std::ratio<1, Denom>>>* tpp) {
-  using D = std::chrono::duration<Rep, std::ratio<1, Denom>>;
-  // TODO(#199): Return false if result unrepresentable as a time_point<D>.
-  *tpp = std::chrono::time_point_cast<D>(sec);
-  *tpp += std::chrono::duration_cast<D>(fs);
-  return true;
-}
-
-template <typename Rep, std::intmax_t Num>
-bool join_seconds(
-    const time_point<seconds>& sec, const femtoseconds&,
-    time_point<std::chrono::duration<Rep, std::ratio<Num, 1>>>* tpp) {
-  using D = std::chrono::duration<Rep, std::ratio<Num, 1>>;
-  auto count = sec.time_since_epoch().count();
-  if (count >= 0 || count % Num == 0) {
-    count /= Num;
-  } else {
-    count /= Num;
-    count -= 1;
-  }
-  if (count > (std::numeric_limits<Rep>::max)()) return false;
-  if (count < (std::numeric_limits<Rep>::min)()) return false;
-  *tpp = time_point<D>() + D{static_cast<Rep>(count)};
-  return true;
-}
-
-template <typename Rep>
-bool join_seconds(
-    const time_point<seconds>& sec, const femtoseconds&,
-    time_point<std::chrono::duration<Rep, std::ratio<1, 1>>>* tpp) {
-  using D = std::chrono::duration<Rep, std::ratio<1, 1>>;
-  auto count = sec.time_since_epoch().count();
-  if (count > (std::numeric_limits<Rep>::max)()) return false;
-  if (count < (std::numeric_limits<Rep>::min)()) return false;
-  *tpp = time_point<D>() + D{static_cast<Rep>(count)};
-  return true;
-}
-
-inline bool join_seconds(const time_point<seconds>& sec, const femtoseconds&,
-                         time_point<seconds>* tpp) {
-  *tpp = sec;
-  return true;
-}
-
-}  // namespace detail
-}  // namespace cctz
-}  // namespace time_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/zone_info_source.h b/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/zone_info_source.h
deleted file mode 100644
index 012eb4ec..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/include/cctz/zone_info_source.h
+++ /dev/null
@@ -1,102 +0,0 @@
-// Copyright 2016 Google Inc. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//   https://www.apache.org/licenses/LICENSE-2.0
-//
-//   Unless required by applicable law or agreed to in writing, software
-//   distributed under the License is distributed on an "AS IS" BASIS,
-//   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-//   See the License for the specific language governing permissions and
-//   limitations under the License.
-
-#ifndef ABSL_TIME_INTERNAL_CCTZ_ZONE_INFO_SOURCE_H_
-#define ABSL_TIME_INTERNAL_CCTZ_ZONE_INFO_SOURCE_H_
-
-#include <cstddef>
-#include <functional>
-#include <memory>
-#include <string>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace time_internal {
-namespace cctz {
-
-// A stdio-like interface for providing zoneinfo data for a particular zone.
-class ZoneInfoSource {
- public:
-  virtual ~ZoneInfoSource();
-
-  virtual std::size_t Read(void* ptr, std::size_t size) = 0;  // like fread()
-  virtual int Skip(std::size_t offset) = 0;                   // like fseek()
-
-  // Until the zoneinfo data supports versioning information, we provide
-  // a way for a ZoneInfoSource to indicate it out-of-band.  The default
-  // implementation returns an empty string.
-  virtual std::string Version() const;
-};
-
-}  // namespace cctz
-}  // namespace time_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace time_internal {
-namespace cctz_extension {
-
-// A function-pointer type for a factory that returns a ZoneInfoSource
-// given the name of a time zone and a fallback factory.  Returns null
-// when the data for the named zone cannot be found.
-using ZoneInfoSourceFactory =
-    std::unique_ptr<absl::time_internal::cctz::ZoneInfoSource> (*)(
-        const std::string&,
-        const std::function<std::unique_ptr<
-            absl::time_internal::cctz::ZoneInfoSource>(const std::string&)>&);
-
-// The user can control the mapping of zone names to zoneinfo data by
-// providing a definition for cctz_extension::zone_info_source_factory.
-// For example, given functions my_factory() and my_other_factory() that
-// can return a ZoneInfoSource for a named zone, we could inject them into
-// cctz::load_time_zone() with:
-//
-//   namespace cctz_extension {
-//   namespace {
-//   std::unique_ptr<cctz::ZoneInfoSource> CustomFactory(
-//       const std::string& name,
-//       const std::function<std::unique_ptr<cctz::ZoneInfoSource>(
-//           const std::string& name)>& fallback_factory) {
-//     if (auto zip = my_factory(name)) return zip;
-//     if (auto zip = fallback_factory(name)) return zip;
-//     if (auto zip = my_other_factory(name)) return zip;
-//     return nullptr;
-//   }
-//   }  // namespace
-//   ZoneInfoSourceFactory zone_info_source_factory = CustomFactory;
-//   }  // namespace cctz_extension
-//
-// This might be used, say, to use zoneinfo data embedded in the program,
-// or read from a (possibly compressed) file archive, or both.
-//
-// cctz_extension::zone_info_source_factory() will be called:
-//   (1) from the same thread as the cctz::load_time_zone() call,
-//   (2) only once for any zone name, and
-//   (3) serially (i.e., no concurrent execution).
-//
-// The fallback factory obtains zoneinfo data by reading files in ${TZDIR},
-// and it is used automatically when no zone_info_source_factory definition
-// is linked into the program.
-extern ZoneInfoSourceFactory zone_info_source_factory;
-
-}  // namespace cctz_extension
-}  // namespace time_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TIME_INTERNAL_CCTZ_ZONE_INFO_SOURCE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_fixed.h b/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_fixed.h
deleted file mode 100644
index e74a0bbd..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_fixed.h
+++ /dev/null
@@ -1,52 +0,0 @@
-// Copyright 2016 Google Inc. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//   https://www.apache.org/licenses/LICENSE-2.0
-//
-//   Unless required by applicable law or agreed to in writing, software
-//   distributed under the License is distributed on an "AS IS" BASIS,
-//   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-//   See the License for the specific language governing permissions and
-//   limitations under the License.
-
-#ifndef ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_FIXED_H_
-#define ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_FIXED_H_
-
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/time/internal/cctz/include/cctz/time_zone.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace time_internal {
-namespace cctz {
-
-// Helper functions for dealing with the names and abbreviations
-// of time zones that are a fixed offset (seconds east) from UTC.
-// FixedOffsetFromName() extracts the offset from a valid fixed-offset
-// name, while FixedOffsetToName() and FixedOffsetToAbbr() generate
-// the canonical zone name and abbreviation respectively for the given
-// offset.
-//
-// A fixed-offset name looks like "Fixed/UTC<+-><hours>:<mins>:<secs>".
-// Its abbreviation is of the form "UTC(<+->H?H(MM(SS)?)?)?" where the
-// optional pieces are omitted when their values are zero.  (Note that
-// the sign is the opposite of that used in a POSIX TZ specification.)
-//
-// Note: FixedOffsetFromName() fails on syntax errors or when the parsed
-// offset exceeds 24 hours.  FixedOffsetToName() and FixedOffsetToAbbr()
-// both produce "UTC" when the argument offset exceeds 24 hours.
-bool FixedOffsetFromName(const std::string& name, seconds* offset);
-std::string FixedOffsetToName(const seconds& offset);
-std::string FixedOffsetToAbbr(const seconds& offset);
-
-}  // namespace cctz
-}  // namespace time_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_FIXED_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_if.h b/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_if.h
deleted file mode 100644
index bec9beb5..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_if.h
+++ /dev/null
@@ -1,80 +0,0 @@
-// Copyright 2016 Google Inc. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//   https://www.apache.org/licenses/LICENSE-2.0
-//
-//   Unless required by applicable law or agreed to in writing, software
-//   distributed under the License is distributed on an "AS IS" BASIS,
-//   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-//   See the License for the specific language governing permissions and
-//   limitations under the License.
-
-#ifndef ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_IF_H_
-#define ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_IF_H_
-
-#include <chrono>
-#include <cstdint>
-#include <memory>
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/time/internal/cctz/include/cctz/civil_time.h"
-#include "absl/time/internal/cctz/include/cctz/time_zone.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace time_internal {
-namespace cctz {
-
-// A simple interface used to hide time-zone complexities from time_zone::Impl.
-// Subclasses implement the functions for civil-time conversions in the zone.
-class TimeZoneIf {
- public:
-  // Factory functions for TimeZoneIf implementations.
-  static std::unique_ptr<TimeZoneIf> UTC();  // never fails
-  static std::unique_ptr<TimeZoneIf> Make(const std::string& name);
-
-  virtual ~TimeZoneIf();
-
-  virtual time_zone::absolute_lookup BreakTime(
-      const time_point<seconds>& tp) const = 0;
-  virtual time_zone::civil_lookup MakeTime(const civil_second& cs) const = 0;
-
-  virtual bool NextTransition(const time_point<seconds>& tp,
-                              time_zone::civil_transition* trans) const = 0;
-  virtual bool PrevTransition(const time_point<seconds>& tp,
-                              time_zone::civil_transition* trans) const = 0;
-
-  virtual std::string Version() const = 0;
-  virtual std::string Description() const = 0;
-
- protected:
-  TimeZoneIf() = default;
-  TimeZoneIf(const TimeZoneIf&) = delete;
-  TimeZoneIf& operator=(const TimeZoneIf&) = delete;
-};
-
-// Convert between time_point<seconds> and a count of seconds since the
-// Unix epoch.  We assume that the std::chrono::system_clock and the
-// Unix clock are second aligned, and that the results are representable.
-// (That is, that they share an epoch, which is required since C++20.)
-inline std::int_fast64_t ToUnixSeconds(const time_point<seconds>& tp) {
-  return (tp - std::chrono::time_point_cast<seconds>(
-                   std::chrono::system_clock::from_time_t(0)))
-      .count();
-}
-inline time_point<seconds> FromUnixSeconds(std::int_fast64_t t) {
-  return std::chrono::time_point_cast<seconds>(
-             std::chrono::system_clock::from_time_t(0)) +
-         seconds(t);
-}
-
-}  // namespace cctz
-}  // namespace time_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_IF_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_impl.h b/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_impl.h
deleted file mode 100644
index 8308a3b4..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_impl.h
+++ /dev/null
@@ -1,97 +0,0 @@
-// Copyright 2016 Google Inc. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//   https://www.apache.org/licenses/LICENSE-2.0
-//
-//   Unless required by applicable law or agreed to in writing, software
-//   distributed under the License is distributed on an "AS IS" BASIS,
-//   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-//   See the License for the specific language governing permissions and
-//   limitations under the License.
-
-#ifndef ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_IMPL_H_
-#define ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_IMPL_H_
-
-#include <memory>
-#include <string>
-
-#include "absl/base/config.h"
-#include "absl/time/internal/cctz/include/cctz/civil_time.h"
-#include "absl/time/internal/cctz/include/cctz/time_zone.h"
-#include "time_zone_if.h"
-#include "time_zone_info.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace time_internal {
-namespace cctz {
-
-// time_zone::Impl is the internal object referenced by a cctz::time_zone.
-class time_zone::Impl {
- public:
-  // The UTC time zone. Also used for other time zones that fail to load.
-  static time_zone UTC();
-
-  // Load a named time zone. Returns false if the name is invalid, or if
-  // some other kind of error occurs. Note that loading "UTC" never fails.
-  static bool LoadTimeZone(const std::string& name, time_zone* tz);
-
-  // Clears the map of cached time zones.  Primarily for use in benchmarks
-  // that gauge the performance of loading/parsing the time-zone data.
-  static void ClearTimeZoneMapTestOnly();
-
-  // The primary key is the time-zone ID (e.g., "America/New_York").
-  const std::string& Name() const {
-    // TODO: It would nice if the zoneinfo data included the zone name.
-    return name_;
-  }
-
-  // Breaks a time_point down to civil-time components in this time zone.
-  time_zone::absolute_lookup BreakTime(const time_point<seconds>& tp) const {
-    return zone_->BreakTime(tp);
-  }
-
-  // Converts the civil-time components in this time zone into a time_point.
-  // That is, the opposite of BreakTime(). The requested civil time may be
-  // ambiguous or illegal due to a change of UTC offset.
-  time_zone::civil_lookup MakeTime(const civil_second& cs) const {
-    return zone_->MakeTime(cs);
-  }
-
-  // Finds the time of the next/previous offset change in this time zone.
-  bool NextTransition(const time_point<seconds>& tp,
-                      time_zone::civil_transition* trans) const {
-    return zone_->NextTransition(tp, trans);
-  }
-  bool PrevTransition(const time_point<seconds>& tp,
-                      time_zone::civil_transition* trans) const {
-    return zone_->PrevTransition(tp, trans);
-  }
-
-  // Returns an implementation-defined version string for this time zone.
-  std::string Version() const { return zone_->Version(); }
-
-  // Returns an implementation-defined description of this time zone.
-  std::string Description() const { return zone_->Description(); }
-
- private:
-  Impl();
-  explicit Impl(const std::string& name);
-  Impl(const Impl&) = delete;
-  Impl& operator=(const Impl&) = delete;
-
-  static const Impl* UTCImpl();
-
-  const std::string name_;
-  std::unique_ptr<TimeZoneIf> zone_;
-};
-
-}  // namespace cctz
-}  // namespace time_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_IMPL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_info.h b/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_info.h
deleted file mode 100644
index 689df6f9..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_info.h
+++ /dev/null
@@ -1,128 +0,0 @@
-// Copyright 2016 Google Inc. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//   https://www.apache.org/licenses/LICENSE-2.0
-//
-//   Unless required by applicable law or agreed to in writing, software
-//   distributed under the License is distributed on an "AS IS" BASIS,
-//   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-//   See the License for the specific language governing permissions and
-//   limitations under the License.
-
-#ifndef ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_INFO_H_
-#define ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_INFO_H_
-
-#include <atomic>
-#include <cstddef>
-#include <cstdint>
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "absl/base/config.h"
-#include "absl/time/internal/cctz/include/cctz/civil_time.h"
-#include "absl/time/internal/cctz/include/cctz/time_zone.h"
-#include "absl/time/internal/cctz/include/cctz/zone_info_source.h"
-#include "time_zone_if.h"
-#include "tzfile.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace time_internal {
-namespace cctz {
-
-// A transition to a new UTC offset.
-struct Transition {
-  std::int_least64_t unix_time;   // the instant of this transition
-  std::uint_least8_t type_index;  // index of the transition type
-  civil_second civil_sec;         // local civil time of transition
-  civil_second prev_civil_sec;    // local civil time one second earlier
-
-  struct ByUnixTime {
-    inline bool operator()(const Transition& lhs, const Transition& rhs) const {
-      return lhs.unix_time < rhs.unix_time;
-    }
-  };
-  struct ByCivilTime {
-    inline bool operator()(const Transition& lhs, const Transition& rhs) const {
-      return lhs.civil_sec < rhs.civil_sec;
-    }
-  };
-};
-
-// The characteristics of a particular transition.
-struct TransitionType {
-  std::int_least32_t utc_offset;  // the new prevailing UTC offset
-  civil_second civil_max;         // max convertible civil time for offset
-  civil_second civil_min;         // min convertible civil time for offset
-  bool is_dst;                    // did we move into daylight-saving time
-  std::uint_least8_t abbr_index;  // index of the new abbreviation
-};
-
-// A time zone backed by the IANA Time Zone Database (zoneinfo).
-class TimeZoneInfo : public TimeZoneIf {
- public:
-  // Factories.
-  static std::unique_ptr<TimeZoneInfo> UTC();  // never fails
-  static std::unique_ptr<TimeZoneInfo> Make(const std::string& name);
-
-  // TimeZoneIf implementations.
-  time_zone::absolute_lookup BreakTime(
-      const time_point<seconds>& tp) const override;
-  time_zone::civil_lookup MakeTime(const civil_second& cs) const override;
-  bool NextTransition(const time_point<seconds>& tp,
-                      time_zone::civil_transition* trans) const override;
-  bool PrevTransition(const time_point<seconds>& tp,
-                      time_zone::civil_transition* trans) const override;
-  std::string Version() const override;
-  std::string Description() const override;
-
- private:
-  TimeZoneInfo() = default;
-  TimeZoneInfo(const TimeZoneInfo&) = delete;
-  TimeZoneInfo& operator=(const TimeZoneInfo&) = delete;
-
-  bool GetTransitionType(std::int_fast32_t utc_offset, bool is_dst,
-                         const std::string& abbr, std::uint_least8_t* index);
-  bool EquivTransitions(std::uint_fast8_t tt1_index,
-                        std::uint_fast8_t tt2_index) const;
-  bool ExtendTransitions();
-
-  bool ResetToBuiltinUTC(const seconds& offset);
-  bool Load(const std::string& name);
-  bool Load(ZoneInfoSource* zip);
-
-  // Helpers for BreakTime() and MakeTime().
-  time_zone::absolute_lookup LocalTime(std::int_fast64_t unix_time,
-                                       const TransitionType& tt) const;
-  time_zone::absolute_lookup LocalTime(std::int_fast64_t unix_time,
-                                       const Transition& tr) const;
-  time_zone::civil_lookup TimeLocal(const civil_second& cs,
-                                    year_t c4_shift) const;
-
-  std::vector<Transition> transitions_;  // ordered by unix_time and civil_sec
-  std::vector<TransitionType> transition_types_;  // distinct transition types
-  std::uint_fast8_t default_transition_type_;     // for before first transition
-  std::string abbreviations_;  // all the NUL-terminated abbreviations
-
-  std::string version_;      // the tzdata version if available
-  std::string future_spec_;  // for after the last zic transition
-  bool extended_;            // future_spec_ was used to generate transitions
-  year_t last_year_;         // the final year of the generated transitions
-
-  // We remember the transitions found during the last BreakTime() and
-  // MakeTime() calls. If the next request is for the same transition we
-  // will avoid re-searching.
-  mutable std::atomic<std::size_t> local_time_hint_ = {};  // BreakTime() hint
-  mutable std::atomic<std::size_t> time_local_hint_ = {};  // MakeTime() hint
-};
-
-}  // namespace cctz
-}  // namespace time_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_INFO_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_libc.h b/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_libc.h
deleted file mode 100644
index ae210737..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_libc.h
+++ /dev/null
@@ -1,60 +0,0 @@
-// Copyright 2016 Google Inc. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//   https://www.apache.org/licenses/LICENSE-2.0
-//
-//   Unless required by applicable law or agreed to in writing, software
-//   distributed under the License is distributed on an "AS IS" BASIS,
-//   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-//   See the License for the specific language governing permissions and
-//   limitations under the License.
-
-#ifndef ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_LIBC_H_
-#define ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_LIBC_H_
-
-#include <memory>
-#include <string>
-
-#include "absl/base/config.h"
-#include "time_zone_if.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace time_internal {
-namespace cctz {
-
-// A time zone backed by gmtime_r(3), localtime_r(3), and mktime(3),
-// and which therefore only supports UTC and the local time zone.
-class TimeZoneLibC : public TimeZoneIf {
- public:
-  // Factory.
-  static std::unique_ptr<TimeZoneLibC> Make(const std::string& name);
-
-  // TimeZoneIf implementations.
-  time_zone::absolute_lookup BreakTime(
-      const time_point<seconds>& tp) const override;
-  time_zone::civil_lookup MakeTime(const civil_second& cs) const override;
-  bool NextTransition(const time_point<seconds>& tp,
-                      time_zone::civil_transition* trans) const override;
-  bool PrevTransition(const time_point<seconds>& tp,
-                      time_zone::civil_transition* trans) const override;
-  std::string Version() const override;
-  std::string Description() const override;
-
- private:
-  explicit TimeZoneLibC(const std::string& name);
-  TimeZoneLibC(const TimeZoneLibC&) = delete;
-  TimeZoneLibC& operator=(const TimeZoneLibC&) = delete;
-
-  const bool local_;  // localtime or UTC
-};
-
-}  // namespace cctz
-}  // namespace time_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_LIBC_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_posix.h b/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_posix.h
deleted file mode 100644
index 7fd2b9ec..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/time_zone_posix.h
+++ /dev/null
@@ -1,132 +0,0 @@
-// Copyright 2016 Google Inc. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//   https://www.apache.org/licenses/LICENSE-2.0
-//
-//   Unless required by applicable law or agreed to in writing, software
-//   distributed under the License is distributed on an "AS IS" BASIS,
-//   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-//   See the License for the specific language governing permissions and
-//   limitations under the License.
-
-// Parsing of a POSIX zone spec as described in the TZ part of section 8.3 in
-// http://pubs.opengroup.org/onlinepubs/009695399/basedefs/xbd_chap08.html.
-//
-// The current POSIX spec for America/Los_Angeles is "PST8PDT,M3.2.0,M11.1.0",
-// which would be broken down as ...
-//
-//   PosixTimeZone {
-//     std_abbr = "PST"
-//     std_offset = -28800
-//     dst_abbr = "PDT"
-//     dst_offset = -25200
-//     dst_start = PosixTransition {
-//       date {
-//         m {
-//           month = 3
-//           week = 2
-//           weekday = 0
-//         }
-//       }
-//       time {
-//         offset = 7200
-//       }
-//     }
-//     dst_end = PosixTransition {
-//       date {
-//         m {
-//           month = 11
-//           week = 1
-//           weekday = 0
-//         }
-//       }
-//       time {
-//         offset = 7200
-//       }
-//     }
-//   }
-
-#ifndef ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_POSIX_H_
-#define ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_POSIX_H_
-
-#include <cstdint>
-#include <string>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace time_internal {
-namespace cctz {
-
-// The date/time of the transition. The date is specified as either:
-// (J) the Nth day of the year (1 <= N <= 365), excluding leap days, or
-// (N) the Nth day of the year (0 <= N <= 365), including leap days, or
-// (M) the Nth weekday of a month (e.g., the 2nd Sunday in March).
-// The time, specified as a day offset, identifies the particular moment
-// of the transition, and may be negative or >= 24h, and in which case
-// it would take us to another day, and perhaps week, or even month.
-struct PosixTransition {
-  enum DateFormat { J, N, M };
-
-  struct Date {
-    struct NonLeapDay {
-      std::int_fast16_t day;  // day of non-leap year [1:365]
-    };
-    struct Day {
-      std::int_fast16_t day;  // day of year [0:365]
-    };
-    struct MonthWeekWeekday {
-      std::int_fast8_t month;    // month of year [1:12]
-      std::int_fast8_t week;     // week of month [1:5] (5==last)
-      std::int_fast8_t weekday;  // 0==Sun, ..., 6=Sat
-    };
-
-    DateFormat fmt;
-
-    union {
-      NonLeapDay j;
-      Day n;
-      MonthWeekWeekday m;
-    };
-  };
-
-  struct Time {
-    std::int_fast32_t offset;  // seconds before/after 00:00:00
-  };
-
-  Date date;
-  Time time;
-};
-
-// The entirety of a POSIX-string specified time-zone rule. The standard
-// abbreviation and offset are always given. If the time zone includes
-// daylight saving, then the daylight abbreviation is non-empty and the
-// remaining fields are also valid. Note that the start/end transitions
-// are not ordered---in the southern hemisphere the transition to end
-// daylight time occurs first in any particular year.
-struct PosixTimeZone {
-  std::string std_abbr;
-  std::int_fast32_t std_offset;
-
-  std::string dst_abbr;
-  std::int_fast32_t dst_offset;
-  PosixTransition dst_start;
-  PosixTransition dst_end;
-};
-
-// Breaks down a POSIX time-zone specification into its constituent pieces,
-// filling in any missing values (DST offset, or start/end transition times)
-// with the standard-defined defaults. Returns false if the specification
-// could not be parsed (although some fields of *res may have been altered).
-bool ParsePosixSpec(const std::string& spec, PosixTimeZone* res);
-
-}  // namespace cctz
-}  // namespace time_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TIME_INTERNAL_CCTZ_TIME_ZONE_POSIX_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/tzfile.h b/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/tzfile.h
deleted file mode 100644
index 2be3bb8d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/time/internal/cctz/src/tzfile.h
+++ /dev/null
@@ -1,118 +0,0 @@
-/* Layout and location of TZif files.  */
-
-#ifndef TZFILE_H
-
-#define TZFILE_H
-
-/*
-** This file is in the public domain, so clarified as of
-** 1996-06-05 by Arthur David Olson.
-*/
-
-/*
-** This header is for use ONLY with the time conversion code.
-** There is no guarantee that it will remain unchanged,
-** or that it will remain at all.
-** Do NOT copy it to any system include directory.
-** Thank you!
-*/
-
-/*
-** Information about time zone files.
-*/
-
-#ifndef TZDEFRULES
-#define TZDEFRULES "posixrules"
-#endif /* !defined TZDEFRULES */
-
-/* See Internet RFC 8536 for more details about the following format.  */
-
-/*
-** Each file begins with. . .
-*/
-
-#define TZ_MAGIC "TZif"
-
-struct tzhead {
-  char tzh_magic[4];      /* TZ_MAGIC */
-  char tzh_version[1];    /* '\0' or '2'-'4' as of 2021 */
-  char tzh_reserved[15];  /* reserved; must be zero */
-  char tzh_ttisutcnt[4];  /* coded number of trans. time flags */
-  char tzh_ttisstdcnt[4]; /* coded number of trans. time flags */
-  char tzh_leapcnt[4];    /* coded number of leap seconds */
-  char tzh_timecnt[4];    /* coded number of transition times */
-  char tzh_typecnt[4];    /* coded number of local time types */
-  char tzh_charcnt[4];    /* coded number of abbr. chars */
-};
-
-/*
-** . . .followed by. . .
-**
-**	tzh_timecnt (char [4])s		coded transition times a la time(2)
-**	tzh_timecnt (unsigned char)s	types of local time starting at above
-**	tzh_typecnt repetitions of
-**		one (char [4])		coded UT offset in seconds
-**		one (unsigned char)	used to set tm_isdst
-**		one (unsigned char)	that's an abbreviation list index
-**	tzh_charcnt (char)s		'\0'-terminated zone abbreviations
-**	tzh_leapcnt repetitions of
-**		one (char [4])		coded leap second transition times
-**		one (char [4])		total correction after above
-**	tzh_ttisstdcnt (char)s		indexed by type; if 1, transition
-**					time is standard time, if 0,
-**					transition time is local (wall clock)
-**					time; if absent, transition times are
-**					assumed to be local time
-**	tzh_ttisutcnt (char)s		indexed by type; if 1, transition
-**					time is UT, if 0, transition time is
-**					local time; if absent, transition
-**					times are assumed to be local time.
-**					When this is 1, the corresponding
-**					std/wall indicator must also be 1.
-*/
-
-/*
-** If tzh_version is '2' or greater, the above is followed by a second instance
-** of tzhead and a second instance of the data in which each coded transition
-** time uses 8 rather than 4 chars,
-** then a POSIX-TZ-environment-variable-style string for use in handling
-** instants after the last transition time stored in the file
-** (with nothing between the newlines if there is no POSIX.1-2017
-** representation for such instants).
-**
-** If tz_version is '3' or greater, the above is extended as follows.
-** First, the TZ string's hour offset may range from -167
-** through 167 as compared to the POSIX-required 0 through 24.
-** Second, its DST start time may be January 1 at 00:00 and its stop
-** time December 31 at 24:00 plus the difference between DST and
-** standard time, indicating DST all year.
-*/
-
-/*
-** In the current implementation, "tzset()" refuses to deal with files that
-** exceed any of the limits below.
-*/
-
-#ifndef TZ_MAX_TIMES
-/* This must be at least 242 for Europe/London with 'zic -b fat'.  */
-#define TZ_MAX_TIMES 2000
-#endif /* !defined TZ_MAX_TIMES */
-
-#ifndef TZ_MAX_TYPES
-/* This must be at least 18 for Europe/Vilnius with 'zic -b fat'.  */
-#define TZ_MAX_TYPES 256 /* Limited by what (unsigned char)'s can hold */
-#endif                   /* !defined TZ_MAX_TYPES */
-
-#ifndef TZ_MAX_CHARS
-/* This must be at least 40 for America/Anchorage.  */
-#define TZ_MAX_CHARS 50 /* Maximum number of abbreviation characters */
-                        /* (limited by what unsigned chars can hold) */
-#endif                  /* !defined TZ_MAX_CHARS */
-
-#ifndef TZ_MAX_LEAPS
-/* This must be at least 27 for leap seconds from 1972 through mid-2023.
-   There's a plan to discontinue leap seconds by 2035.  */
-#define TZ_MAX_LEAPS 50 /* Maximum number of leap second corrections */
-#endif                  /* !defined TZ_MAX_LEAPS */
-
-#endif /* !defined TZFILE_H */
diff --git a/projects/llm_framework/include/abseil-cpp/absl/time/internal/test_util.h b/projects/llm_framework/include/abseil-cpp/absl/time/internal/test_util.h
deleted file mode 100644
index 5c4bf1f6..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/time/internal/test_util.h
+++ /dev/null
@@ -1,33 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_TIME_INTERNAL_TEST_UTIL_H_
-#define ABSL_TIME_INTERNAL_TEST_UTIL_H_
-
-#include <string>
-
-#include "absl/time/time.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace time_internal {
-
-// Loads the named timezone, but dies on any failure.
-absl::TimeZone LoadTimeZone(const std::string& name);
-
-}  // namespace time_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TIME_INTERNAL_TEST_UTIL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/time/time.h b/projects/llm_framework/include/abseil-cpp/absl/time/time.h
deleted file mode 100644
index d367ace2..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/time/time.h
+++ /dev/null
@@ -1,1811 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// File: time.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines abstractions for computing with absolute points
-// in time, durations of time, and formatting and parsing time within a given
-// time zone. The following abstractions are defined:
-//
-//  * `absl::Time` defines an absolute, specific instance in time
-//  * `absl::Duration` defines a signed, fixed-length span of time
-//  * `absl::TimeZone` defines geopolitical time zone regions (as collected
-//     within the IANA Time Zone database (https://www.iana.org/time-zones)).
-//
-// Note: Absolute times are distinct from civil times, which refer to the
-// human-scale time commonly represented by `YYYY-MM-DD hh:mm:ss`. The mapping
-// between absolute and civil times can be specified by use of time zones
-// (`absl::TimeZone` within this API). That is:
-//
-//   Civil Time = F(Absolute Time, Time Zone)
-//   Absolute Time = G(Civil Time, Time Zone)
-//
-// See civil_time.h for abstractions related to constructing and manipulating
-// civil time.
-//
-// Example:
-//
-//   absl::TimeZone nyc;
-//   // LoadTimeZone() may fail so it's always better to check for success.
-//   if (!absl::LoadTimeZone("America/New_York", &nyc)) {
-//      // handle error case
-//   }
-//
-//   // My flight leaves NYC on Jan 2, 2017 at 03:04:05
-//   absl::CivilSecond cs(2017, 1, 2, 3, 4, 5);
-//   absl::Time takeoff = absl::FromCivil(cs, nyc);
-//
-//   absl::Duration flight_duration = absl::Hours(21) + absl::Minutes(35);
-//   absl::Time landing = takeoff + flight_duration;
-//
-//   absl::TimeZone syd;
-//   if (!absl::LoadTimeZone("Australia/Sydney", &syd)) {
-//      // handle error case
-//   }
-//   std::string s = absl::FormatTime(
-//       "My flight will land in Sydney on %Y-%m-%d at %H:%M:%S",
-//       landing, syd);
-
-#ifndef ABSL_TIME_TIME_H_
-#define ABSL_TIME_TIME_H_
-
-#if !defined(_MSC_VER)
-#include <sys/time.h>
-#else
-// We don't include `winsock2.h` because it drags in `windows.h` and friends,
-// and they define conflicting macros like OPAQUE, ERROR, and more. This has the
-// potential to break Abseil users.
-//
-// Instead we only forward declare `timeval` and require Windows users include
-// `winsock2.h` themselves. This is both inconsistent and troublesome, but so is
-// including 'windows.h' so we are picking the lesser of two evils here.
-struct timeval;
-#endif
-#include <chrono>  // NOLINT(build/c++11)
-#include <cmath>
-#include <cstdint>
-#include <ctime>
-#include <limits>
-#include <ostream>
-#include <string>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/config.h"
-#include "absl/base/macros.h"
-#include "absl/strings/string_view.h"
-#include "absl/time/civil_time.h"
-#include "absl/time/internal/cctz/include/cctz/time_zone.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-class Duration;  // Defined below
-class Time;      // Defined below
-class TimeZone;  // Defined below
-
-namespace time_internal {
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time FromUnixDuration(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration ToUnixDuration(Time t);
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr int64_t GetRepHi(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr uint32_t GetRepLo(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration MakeDuration(int64_t hi,
-                                                              uint32_t lo);
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration MakeDuration(int64_t hi,
-                                                              int64_t lo);
-ABSL_ATTRIBUTE_CONST_FUNCTION inline Duration MakePosDoubleDuration(double n);
-constexpr int64_t kTicksPerNanosecond = 4;
-constexpr int64_t kTicksPerSecond = 1000 * 1000 * 1000 * kTicksPerNanosecond;
-template <std::intmax_t N>
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration FromInt64(int64_t v,
-                                                           std::ratio<1, N>);
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration FromInt64(int64_t v,
-                                                           std::ratio<60>);
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration FromInt64(int64_t v,
-                                                           std::ratio<3600>);
-template <typename T>
-using EnableIfIntegral = typename std::enable_if<
-    std::is_integral<T>::value || std::is_enum<T>::value, int>::type;
-template <typename T>
-using EnableIfFloat =
-    typename std::enable_if<std::is_floating_point<T>::value, int>::type;
-}  // namespace time_internal
-
-// Duration
-//
-// The `absl::Duration` class represents a signed, fixed-length amount of time.
-// A `Duration` is generated using a unit-specific factory function, or is
-// the result of subtracting one `absl::Time` from another. Durations behave
-// like unit-safe integers and they support all the natural integer-like
-// arithmetic operations. Arithmetic overflows and saturates at +/- infinity.
-// `Duration` should be passed by value rather than const reference.
-//
-// Factory functions `Nanoseconds()`, `Microseconds()`, `Milliseconds()`,
-// `Seconds()`, `Minutes()`, `Hours()` and `InfiniteDuration()` allow for
-// creation of constexpr `Duration` values
-//
-// Examples:
-//
-//   constexpr absl::Duration ten_ns = absl::Nanoseconds(10);
-//   constexpr absl::Duration min = absl::Minutes(1);
-//   constexpr absl::Duration hour = absl::Hours(1);
-//   absl::Duration dur = 60 * min;  // dur == hour
-//   absl::Duration half_sec = absl::Milliseconds(500);
-//   absl::Duration quarter_sec = 0.25 * absl::Seconds(1);
-//
-// `Duration` values can be easily converted to an integral number of units
-// using the division operator.
-//
-// Example:
-//
-//   constexpr absl::Duration dur = absl::Milliseconds(1500);
-//   int64_t ns = dur / absl::Nanoseconds(1);   // ns == 1500000000
-//   int64_t ms = dur / absl::Milliseconds(1);  // ms == 1500
-//   int64_t sec = dur / absl::Seconds(1);    // sec == 1 (subseconds truncated)
-//   int64_t min = dur / absl::Minutes(1);    // min == 0
-//
-// See the `IDivDuration()` and `FDivDuration()` functions below for details on
-// how to access the fractional parts of the quotient.
-//
-// Alternatively, conversions can be performed using helpers such as
-// `ToInt64Microseconds()` and `ToDoubleSeconds()`.
-class Duration {
- public:
-  // Value semantics.
-  constexpr Duration() : rep_hi_(0), rep_lo_(0) {}  // zero-length duration
-
-  // Copyable.
-#if !defined(__clang__) && defined(_MSC_VER) && _MSC_VER < 1930
-  // Explicitly defining the constexpr copy constructor avoids an MSVC bug.
-  constexpr Duration(const Duration& d)
-      : rep_hi_(d.rep_hi_), rep_lo_(d.rep_lo_) {}
-#else
-  constexpr Duration(const Duration& d) = default;
-#endif
-  Duration& operator=(const Duration& d) = default;
-
-  // Compound assignment operators.
-  Duration& operator+=(Duration d);
-  Duration& operator-=(Duration d);
-  Duration& operator*=(int64_t r);
-  Duration& operator*=(double r);
-  Duration& operator/=(int64_t r);
-  Duration& operator/=(double r);
-  Duration& operator%=(Duration rhs);
-
-  // Overloads that forward to either the int64_t or double overloads above.
-  // Integer operands must be representable as int64_t. Integer division is
-  // truncating, so values less than the resolution will be returned as zero.
-  // Floating-point multiplication and division is rounding (halfway cases
-  // rounding away from zero), so values less than the resolution may be
-  // returned as either the resolution or zero.  In particular, `d / 2.0`
-  // can produce `d` when it is the resolution and "even".
-  template <typename T, time_internal::EnableIfIntegral<T> = 0>
-  Duration& operator*=(T r) {
-    int64_t x = r;
-    return *this *= x;
-  }
-
-  template <typename T, time_internal::EnableIfIntegral<T> = 0>
-  Duration& operator/=(T r) {
-    int64_t x = r;
-    return *this /= x;
-  }
-
-  template <typename T, time_internal::EnableIfFloat<T> = 0>
-  Duration& operator*=(T r) {
-    double x = r;
-    return *this *= x;
-  }
-
-  template <typename T, time_internal::EnableIfFloat<T> = 0>
-  Duration& operator/=(T r) {
-    double x = r;
-    return *this /= x;
-  }
-
-  template <typename H>
-  friend H AbslHashValue(H h, Duration d) {
-    return H::combine(std::move(h), d.rep_hi_.Get(), d.rep_lo_);
-  }
-
- private:
-  friend constexpr int64_t time_internal::GetRepHi(Duration d);
-  friend constexpr uint32_t time_internal::GetRepLo(Duration d);
-  friend constexpr Duration time_internal::MakeDuration(int64_t hi,
-                                                        uint32_t lo);
-  constexpr Duration(int64_t hi, uint32_t lo) : rep_hi_(hi), rep_lo_(lo) {}
-
-  // We store `rep_hi_` 4-byte rather than 8-byte aligned to avoid 4 bytes of
-  // tail padding.
-  class HiRep {
-   public:
-    // Default constructor default-initializes `hi_`, which has the same
-    // semantics as default-initializing an `int64_t` (undetermined value).
-    HiRep() = default;
-
-    HiRep(const HiRep&) = default;
-    HiRep& operator=(const HiRep&) = default;
-
-    explicit constexpr HiRep(const int64_t value)
-        :  // C++17 forbids default-initialization in constexpr contexts. We can
-           // remove this in C++20.
-#if defined(ABSL_IS_BIG_ENDIAN) && ABSL_IS_BIG_ENDIAN
-          hi_(0),
-          lo_(0)
-#else
-          lo_(0),
-          hi_(0)
-#endif
-    {
-      *this = value;
-    }
-
-    constexpr int64_t Get() const {
-      const uint64_t unsigned_value =
-          (static_cast<uint64_t>(hi_) << 32) | static_cast<uint64_t>(lo_);
-      // `static_cast<int64_t>(unsigned_value)` is implementation-defined
-      // before c++20. On all supported platforms the behaviour is that mandated
-      // by c++20, i.e. "If the destination type is signed, [...] the result is
-      // the unique value of the destination type equal to the source value
-      // modulo 2^n, where n is the number of bits used to represent the
-      // destination type."
-      static_assert(
-          (static_cast<int64_t>((std::numeric_limits<uint64_t>::max)()) ==
-           int64_t{-1}) &&
-              (static_cast<int64_t>(static_cast<uint64_t>(
-                                        (std::numeric_limits<int64_t>::max)()) +
-                                    1) ==
-               (std::numeric_limits<int64_t>::min)()),
-          "static_cast<int64_t>(uint64_t) does not have c++20 semantics");
-      return static_cast<int64_t>(unsigned_value);
-    }
-
-    constexpr HiRep& operator=(const int64_t value) {
-      // "If the destination type is unsigned, the resulting value is the
-      // smallest unsigned value equal to the source value modulo 2^n
-      // where `n` is the number of bits used to represent the destination
-      // type".
-      const auto unsigned_value = static_cast<uint64_t>(value);
-      hi_ = static_cast<uint32_t>(unsigned_value >> 32);
-      lo_ = static_cast<uint32_t>(unsigned_value);
-      return *this;
-    }
-
-   private:
-    // Notes:
-    //  - Ideally we would use a `char[]` and `std::bitcast`, but the latter
-    //    does not exist (and is not constexpr in `absl`) before c++20.
-    //  - Order is optimized depending on endianness so that the compiler can
-    //    turn `Get()` (resp. `operator=()`) into a single 8-byte load (resp.
-    //    store).
-#if defined(ABSL_IS_BIG_ENDIAN) && ABSL_IS_BIG_ENDIAN
-    uint32_t hi_;
-    uint32_t lo_;
-#else
-    uint32_t lo_;
-    uint32_t hi_;
-#endif
-  };
-  HiRep rep_hi_;
-  uint32_t rep_lo_;
-};
-
-// Relational Operators
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr bool operator<(Duration lhs,
-                                                       Duration rhs);
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr bool operator>(Duration lhs,
-                                                       Duration rhs) {
-  return rhs < lhs;
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr bool operator>=(Duration lhs,
-                                                        Duration rhs) {
-  return !(lhs < rhs);
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr bool operator<=(Duration lhs,
-                                                        Duration rhs) {
-  return !(rhs < lhs);
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr bool operator==(Duration lhs,
-                                                        Duration rhs);
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr bool operator!=(Duration lhs,
-                                                        Duration rhs) {
-  return !(lhs == rhs);
-}
-
-// Additive Operators
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration operator-(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION inline Duration operator+(Duration lhs,
-                                                        Duration rhs) {
-  return lhs += rhs;
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION inline Duration operator-(Duration lhs,
-                                                        Duration rhs) {
-  return lhs -= rhs;
-}
-
-// IDivDuration()
-//
-// Divides a numerator `Duration` by a denominator `Duration`, returning the
-// quotient and remainder. The remainder always has the same sign as the
-// numerator. The returned quotient and remainder respect the identity:
-//
-//   numerator = denominator * quotient + remainder
-//
-// Returned quotients are capped to the range of `int64_t`, with the difference
-// spilling into the remainder to uphold the above identity. This means that the
-// remainder returned could differ from the remainder returned by
-// `Duration::operator%` for huge quotients.
-//
-// See also the notes on `InfiniteDuration()` below regarding the behavior of
-// division involving zero and infinite durations.
-//
-// Example:
-//
-//   constexpr absl::Duration a =
-//       absl::Seconds(std::numeric_limits<int64_t>::max());  // big
-//   constexpr absl::Duration b = absl::Nanoseconds(1);       // small
-//
-//   absl::Duration rem = a % b;
-//   // rem == absl::ZeroDuration()
-//
-//   // Here, q would overflow int64_t, so rem accounts for the difference.
-//   int64_t q = absl::IDivDuration(a, b, &rem);
-//   // q == std::numeric_limits<int64_t>::max(), rem == a - b * q
-int64_t IDivDuration(Duration num, Duration den, Duration* rem);
-
-// FDivDuration()
-//
-// Divides a `Duration` numerator into a fractional number of units of a
-// `Duration` denominator.
-//
-// See also the notes on `InfiniteDuration()` below regarding the behavior of
-// division involving zero and infinite durations.
-//
-// Example:
-//
-//   double d = absl::FDivDuration(absl::Milliseconds(1500), absl::Seconds(1));
-//   // d == 1.5
-ABSL_ATTRIBUTE_CONST_FUNCTION double FDivDuration(Duration num, Duration den);
-
-// Multiplicative Operators
-// Integer operands must be representable as int64_t.
-template <typename T>
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration operator*(Duration lhs, T rhs) {
-  return lhs *= rhs;
-}
-template <typename T>
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration operator*(T lhs, Duration rhs) {
-  return rhs *= lhs;
-}
-template <typename T>
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration operator/(Duration lhs, T rhs) {
-  return lhs /= rhs;
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION inline int64_t operator/(Duration lhs,
-                                                       Duration rhs) {
-  return IDivDuration(lhs, rhs,
-                      &lhs);  // trunc towards zero
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION inline Duration operator%(Duration lhs,
-                                                        Duration rhs) {
-  return lhs %= rhs;
-}
-
-// ZeroDuration()
-//
-// Returns a zero-length duration. This function behaves just like the default
-// constructor, but the name helps make the semantics clear at call sites.
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration ZeroDuration() {
-  return Duration();
-}
-
-// AbsDuration()
-//
-// Returns the absolute value of a duration.
-ABSL_ATTRIBUTE_CONST_FUNCTION inline Duration AbsDuration(Duration d) {
-  return (d < ZeroDuration()) ? -d : d;
-}
-
-// Trunc()
-//
-// Truncates a duration (toward zero) to a multiple of a non-zero unit.
-//
-// Example:
-//
-//   absl::Duration d = absl::Nanoseconds(123456789);
-//   absl::Duration a = absl::Trunc(d, absl::Microseconds(1));  // 123456us
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration Trunc(Duration d, Duration unit);
-
-// Floor()
-//
-// Floors a duration using the passed duration unit to its largest value not
-// greater than the duration.
-//
-// Example:
-//
-//   absl::Duration d = absl::Nanoseconds(123456789);
-//   absl::Duration b = absl::Floor(d, absl::Microseconds(1));  // 123456us
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration Floor(Duration d, Duration unit);
-
-// Ceil()
-//
-// Returns the ceiling of a duration using the passed duration unit to its
-// smallest value not less than the duration.
-//
-// Example:
-//
-//   absl::Duration d = absl::Nanoseconds(123456789);
-//   absl::Duration c = absl::Ceil(d, absl::Microseconds(1));   // 123457us
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration Ceil(Duration d, Duration unit);
-
-// InfiniteDuration()
-//
-// Returns an infinite `Duration`.  To get a `Duration` representing negative
-// infinity, use `-InfiniteDuration()`.
-//
-// Duration arithmetic overflows to +/- infinity and saturates. In general,
-// arithmetic with `Duration` infinities is similar to IEEE 754 infinities
-// except where IEEE 754 NaN would be involved, in which case +/-
-// `InfiniteDuration()` is used in place of a "nan" Duration.
-//
-// Examples:
-//
-//   constexpr absl::Duration inf = absl::InfiniteDuration();
-//   const absl::Duration d = ... any finite duration ...
-//
-//   inf == inf + inf
-//   inf == inf + d
-//   inf == inf - inf
-//   -inf == d - inf
-//
-//   inf == d * 1e100
-//   inf == inf / 2
-//   0 == d / inf
-//   INT64_MAX == inf / d
-//
-//   d < inf
-//   -inf < d
-//
-//   // Division by zero returns infinity, or INT64_MIN/MAX where appropriate.
-//   inf == d / 0
-//   INT64_MAX == d / absl::ZeroDuration()
-//
-// The examples involving the `/` operator above also apply to `IDivDuration()`
-// and `FDivDuration()`.
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration InfiniteDuration();
-
-// Nanoseconds()
-// Microseconds()
-// Milliseconds()
-// Seconds()
-// Minutes()
-// Hours()
-//
-// Factory functions for constructing `Duration` values from an integral number
-// of the unit indicated by the factory function's name. The number must be
-// representable as int64_t.
-//
-// NOTE: no "Days()" factory function exists because "a day" is ambiguous.
-// Civil days are not always 24 hours long, and a 24-hour duration often does
-// not correspond with a civil day. If a 24-hour duration is needed, use
-// `absl::Hours(24)`. If you actually want a civil day, use absl::CivilDay
-// from civil_time.h.
-//
-// Example:
-//
-//   absl::Duration a = absl::Seconds(60);
-//   absl::Duration b = absl::Minutes(1);  // b == a
-template <typename T, time_internal::EnableIfIntegral<T> = 0>
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration Nanoseconds(T n) {
-  return time_internal::FromInt64(n, std::nano{});
-}
-template <typename T, time_internal::EnableIfIntegral<T> = 0>
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration Microseconds(T n) {
-  return time_internal::FromInt64(n, std::micro{});
-}
-template <typename T, time_internal::EnableIfIntegral<T> = 0>
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration Milliseconds(T n) {
-  return time_internal::FromInt64(n, std::milli{});
-}
-template <typename T, time_internal::EnableIfIntegral<T> = 0>
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration Seconds(T n) {
-  return time_internal::FromInt64(n, std::ratio<1>{});
-}
-template <typename T, time_internal::EnableIfIntegral<T> = 0>
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration Minutes(T n) {
-  return time_internal::FromInt64(n, std::ratio<60>{});
-}
-template <typename T, time_internal::EnableIfIntegral<T> = 0>
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration Hours(T n) {
-  return time_internal::FromInt64(n, std::ratio<3600>{});
-}
-
-// Factory overloads for constructing `Duration` values from a floating-point
-// number of the unit indicated by the factory function's name. These functions
-// exist for convenience, but they are not as efficient as the integral
-// factories, which should be preferred.
-//
-// Example:
-//
-//   auto a = absl::Seconds(1.5);        // OK
-//   auto b = absl::Milliseconds(1500);  // BETTER
-template <typename T, time_internal::EnableIfFloat<T> = 0>
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration Nanoseconds(T n) {
-  return n * Nanoseconds(1);
-}
-template <typename T, time_internal::EnableIfFloat<T> = 0>
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration Microseconds(T n) {
-  return n * Microseconds(1);
-}
-template <typename T, time_internal::EnableIfFloat<T> = 0>
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration Milliseconds(T n) {
-  return n * Milliseconds(1);
-}
-template <typename T, time_internal::EnableIfFloat<T> = 0>
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration Seconds(T n) {
-  if (n >= 0) {  // Note: `NaN >= 0` is false.
-    if (n >= static_cast<T>((std::numeric_limits<int64_t>::max)())) {
-      return InfiniteDuration();
-    }
-    return time_internal::MakePosDoubleDuration(n);
-  } else {
-    if (std::isnan(n))
-      return std::signbit(n) ? -InfiniteDuration() : InfiniteDuration();
-    if (n <= (std::numeric_limits<int64_t>::min)()) return -InfiniteDuration();
-    return -time_internal::MakePosDoubleDuration(-n);
-  }
-}
-template <typename T, time_internal::EnableIfFloat<T> = 0>
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration Minutes(T n) {
-  return n * Minutes(1);
-}
-template <typename T, time_internal::EnableIfFloat<T> = 0>
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration Hours(T n) {
-  return n * Hours(1);
-}
-
-// ToInt64Nanoseconds()
-// ToInt64Microseconds()
-// ToInt64Milliseconds()
-// ToInt64Seconds()
-// ToInt64Minutes()
-// ToInt64Hours()
-//
-// Helper functions that convert a Duration to an integral count of the
-// indicated unit. These return the same results as the `IDivDuration()`
-// function, though they usually do so more efficiently; see the
-// documentation of `IDivDuration()` for details about overflow, etc.
-//
-// Example:
-//
-//   absl::Duration d = absl::Milliseconds(1500);
-//   int64_t isec = absl::ToInt64Seconds(d);  // isec == 1
-ABSL_ATTRIBUTE_CONST_FUNCTION int64_t ToInt64Nanoseconds(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION int64_t ToInt64Microseconds(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION int64_t ToInt64Milliseconds(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION int64_t ToInt64Seconds(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION int64_t ToInt64Minutes(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION int64_t ToInt64Hours(Duration d);
-
-// ToDoubleNanoseconds()
-// ToDoubleMicroseconds()
-// ToDoubleMilliseconds()
-// ToDoubleSeconds()
-// ToDoubleMinutes()
-// ToDoubleHours()
-//
-// Helper functions that convert a Duration to a floating point count of the
-// indicated unit. These functions are shorthand for the `FDivDuration()`
-// function above; see its documentation for details about overflow, etc.
-//
-// Example:
-//
-//   absl::Duration d = absl::Milliseconds(1500);
-//   double dsec = absl::ToDoubleSeconds(d);  // dsec == 1.5
-ABSL_ATTRIBUTE_CONST_FUNCTION double ToDoubleNanoseconds(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION double ToDoubleMicroseconds(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION double ToDoubleMilliseconds(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION double ToDoubleSeconds(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION double ToDoubleMinutes(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION double ToDoubleHours(Duration d);
-
-// FromChrono()
-//
-// Converts any of the pre-defined std::chrono durations to an absl::Duration.
-//
-// Example:
-//
-//   std::chrono::milliseconds ms(123);
-//   absl::Duration d = absl::FromChrono(ms);
-ABSL_ATTRIBUTE_PURE_FUNCTION constexpr Duration FromChrono(
-    const std::chrono::nanoseconds& d);
-ABSL_ATTRIBUTE_PURE_FUNCTION constexpr Duration FromChrono(
-    const std::chrono::microseconds& d);
-ABSL_ATTRIBUTE_PURE_FUNCTION constexpr Duration FromChrono(
-    const std::chrono::milliseconds& d);
-ABSL_ATTRIBUTE_PURE_FUNCTION constexpr Duration FromChrono(
-    const std::chrono::seconds& d);
-ABSL_ATTRIBUTE_PURE_FUNCTION constexpr Duration FromChrono(
-    const std::chrono::minutes& d);
-ABSL_ATTRIBUTE_PURE_FUNCTION constexpr Duration FromChrono(
-    const std::chrono::hours& d);
-
-// ToChronoNanoseconds()
-// ToChronoMicroseconds()
-// ToChronoMilliseconds()
-// ToChronoSeconds()
-// ToChronoMinutes()
-// ToChronoHours()
-//
-// Converts an absl::Duration to any of the pre-defined std::chrono durations.
-// If overflow would occur, the returned value will saturate at the min/max
-// chrono duration value instead.
-//
-// Example:
-//
-//   absl::Duration d = absl::Microseconds(123);
-//   auto x = absl::ToChronoMicroseconds(d);
-//   auto y = absl::ToChronoNanoseconds(d);  // x == y
-//   auto z = absl::ToChronoSeconds(absl::InfiniteDuration());
-//   // z == std::chrono::seconds::max()
-ABSL_ATTRIBUTE_CONST_FUNCTION std::chrono::nanoseconds ToChronoNanoseconds(
-    Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION std::chrono::microseconds ToChronoMicroseconds(
-    Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION std::chrono::milliseconds ToChronoMilliseconds(
-    Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION std::chrono::seconds ToChronoSeconds(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION std::chrono::minutes ToChronoMinutes(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION std::chrono::hours ToChronoHours(Duration d);
-
-// FormatDuration()
-//
-// Returns a string representing the duration in the form "72h3m0.5s".
-// Returns "inf" or "-inf" for +/- `InfiniteDuration()`.
-ABSL_ATTRIBUTE_CONST_FUNCTION std::string FormatDuration(Duration d);
-
-// Output stream operator.
-inline std::ostream& operator<<(std::ostream& os, Duration d) {
-  return os << FormatDuration(d);
-}
-
-// Support for StrFormat(), StrCat() etc.
-template <typename Sink>
-void AbslStringify(Sink& sink, Duration d) {
-  sink.Append(FormatDuration(d));
-}
-
-// ParseDuration()
-//
-// Parses a duration string consisting of a possibly signed sequence of
-// decimal numbers, each with an optional fractional part and a unit
-// suffix.  The valid suffixes are "ns", "us" "ms", "s", "m", and "h".
-// Simple examples include "300ms", "-1.5h", and "2h45m".  Parses "0" as
-// `ZeroDuration()`. Parses "inf" and "-inf" as +/- `InfiniteDuration()`.
-bool ParseDuration(absl::string_view dur_string, Duration* d);
-
-// AbslParseFlag()
-//
-// Parses a command-line flag string representation `text` into a Duration
-// value. Duration flags must be specified in a format that is valid input for
-// `absl::ParseDuration()`.
-bool AbslParseFlag(absl::string_view text, Duration* dst, std::string* error);
-
-
-// AbslUnparseFlag()
-//
-// Unparses a Duration value into a command-line string representation using
-// the format specified by `absl::ParseDuration()`.
-std::string AbslUnparseFlag(Duration d);
-
-ABSL_DEPRECATED("Use AbslParseFlag() instead.")
-bool ParseFlag(const std::string& text, Duration* dst, std::string* error);
-ABSL_DEPRECATED("Use AbslUnparseFlag() instead.")
-std::string UnparseFlag(Duration d);
-
-// Time
-//
-// An `absl::Time` represents a specific instant in time. Arithmetic operators
-// are provided for naturally expressing time calculations. Instances are
-// created using `absl::Now()` and the `absl::From*()` factory functions that
-// accept the gamut of other time representations. Formatting and parsing
-// functions are provided for conversion to and from strings.  `absl::Time`
-// should be passed by value rather than const reference.
-//
-// `absl::Time` assumes there are 60 seconds in a minute, which means the
-// underlying time scales must be "smeared" to eliminate leap seconds.
-// See https://developers.google.com/time/smear.
-//
-// Even though `absl::Time` supports a wide range of timestamps, exercise
-// caution when using values in the distant past. `absl::Time` uses the
-// Proleptic Gregorian calendar, which extends the Gregorian calendar backward
-// to dates before its introduction in 1582.
-// See https://en.wikipedia.org/wiki/Proleptic_Gregorian_calendar
-// for more information. Use the ICU calendar classes to convert a date in
-// some other calendar (http://userguide.icu-project.org/datetime/calendar).
-//
-// Similarly, standardized time zones are a reasonably recent innovation, with
-// the Greenwich prime meridian being established in 1884. The TZ database
-// itself does not profess accurate offsets for timestamps prior to 1970. The
-// breakdown of future timestamps is subject to the whim of regional
-// governments.
-//
-// The `absl::Time` class represents an instant in time as a count of clock
-// ticks of some granularity (resolution) from some starting point (epoch).
-//
-// `absl::Time` uses a resolution that is high enough to avoid loss in
-// precision, and a range that is wide enough to avoid overflow, when
-// converting between tick counts in most Google time scales (i.e., resolution
-// of at least one nanosecond, and range +/-100 billion years).  Conversions
-// between the time scales are performed by truncating (towards negative
-// infinity) to the nearest representable point.
-//
-// Examples:
-//
-//   absl::Time t1 = ...;
-//   absl::Time t2 = t1 + absl::Minutes(2);
-//   absl::Duration d = t2 - t1;  // == absl::Minutes(2)
-//
-class Time {
- public:
-  // Value semantics.
-
-  // Returns the Unix epoch.  However, those reading your code may not know
-  // or expect the Unix epoch as the default value, so make your code more
-  // readable by explicitly initializing all instances before use.
-  //
-  // Example:
-  //   absl::Time t = absl::UnixEpoch();
-  //   absl::Time t = absl::Now();
-  //   absl::Time t = absl::TimeFromTimeval(tv);
-  //   absl::Time t = absl::InfinitePast();
-  constexpr Time() = default;
-
-  // Copyable.
-  constexpr Time(const Time& t) = default;
-  Time& operator=(const Time& t) = default;
-
-  // Assignment operators.
-  Time& operator+=(Duration d) {
-    rep_ += d;
-    return *this;
-  }
-  Time& operator-=(Duration d) {
-    rep_ -= d;
-    return *this;
-  }
-
-  // Time::Breakdown
-  //
-  // The calendar and wall-clock (aka "civil time") components of an
-  // `absl::Time` in a certain `absl::TimeZone`. This struct is not
-  // intended to represent an instant in time. So, rather than passing
-  // a `Time::Breakdown` to a function, pass an `absl::Time` and an
-  // `absl::TimeZone`.
-  //
-  // Deprecated. Use `absl::TimeZone::CivilInfo`.
-  struct ABSL_DEPRECATED("Use `absl::TimeZone::CivilInfo`.") Breakdown {
-    int64_t year;        // year (e.g., 2013)
-    int month;           // month of year [1:12]
-    int day;             // day of month [1:31]
-    int hour;            // hour of day [0:23]
-    int minute;          // minute of hour [0:59]
-    int second;          // second of minute [0:59]
-    Duration subsecond;  // [Seconds(0):Seconds(1)) if finite
-    int weekday;         // 1==Mon, ..., 7=Sun
-    int yearday;         // day of year [1:366]
-
-    // Note: The following fields exist for backward compatibility
-    // with older APIs.  Accessing these fields directly is a sign of
-    // imprudent logic in the calling code.  Modern time-related code
-    // should only access this data indirectly by way of FormatTime().
-    // These fields are undefined for InfiniteFuture() and InfinitePast().
-    int offset;             // seconds east of UTC
-    bool is_dst;            // is offset non-standard?
-    const char* zone_abbr;  // time-zone abbreviation (e.g., "PST")
-  };
-
-  // Time::In()
-  //
-  // Returns the breakdown of this instant in the given TimeZone.
-  //
-  // Deprecated. Use `absl::TimeZone::At(Time)`.
-  ABSL_INTERNAL_DISABLE_DEPRECATED_DECLARATION_WARNING
-  ABSL_DEPRECATED("Use `absl::TimeZone::At(Time)`.")
-  Breakdown In(TimeZone tz) const;
-  ABSL_INTERNAL_RESTORE_DEPRECATED_DECLARATION_WARNING
-
-  template <typename H>
-  friend H AbslHashValue(H h, Time t) {
-    return H::combine(std::move(h), t.rep_);
-  }
-
- private:
-  friend constexpr Time time_internal::FromUnixDuration(Duration d);
-  friend constexpr Duration time_internal::ToUnixDuration(Time t);
-  friend constexpr bool operator<(Time lhs, Time rhs);
-  friend constexpr bool operator==(Time lhs, Time rhs);
-  friend Duration operator-(Time lhs, Time rhs);
-  friend constexpr Time UniversalEpoch();
-  friend constexpr Time InfiniteFuture();
-  friend constexpr Time InfinitePast();
-  constexpr explicit Time(Duration rep) : rep_(rep) {}
-  Duration rep_;
-};
-
-// Relational Operators
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr bool operator<(Time lhs, Time rhs) {
-  return lhs.rep_ < rhs.rep_;
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr bool operator>(Time lhs, Time rhs) {
-  return rhs < lhs;
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr bool operator>=(Time lhs, Time rhs) {
-  return !(lhs < rhs);
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr bool operator<=(Time lhs, Time rhs) {
-  return !(rhs < lhs);
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr bool operator==(Time lhs, Time rhs) {
-  return lhs.rep_ == rhs.rep_;
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr bool operator!=(Time lhs, Time rhs) {
-  return !(lhs == rhs);
-}
-
-// Additive Operators
-ABSL_ATTRIBUTE_CONST_FUNCTION inline Time operator+(Time lhs, Duration rhs) {
-  return lhs += rhs;
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION inline Time operator+(Duration lhs, Time rhs) {
-  return rhs += lhs;
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION inline Time operator-(Time lhs, Duration rhs) {
-  return lhs -= rhs;
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION inline Duration operator-(Time lhs, Time rhs) {
-  return lhs.rep_ - rhs.rep_;
-}
-
-// UnixEpoch()
-//
-// Returns the `absl::Time` representing "1970-01-01 00:00:00.0 +0000".
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time UnixEpoch() { return Time(); }
-
-// UniversalEpoch()
-//
-// Returns the `absl::Time` representing "0001-01-01 00:00:00.0 +0000", the
-// epoch of the ICU Universal Time Scale.
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time UniversalEpoch() {
-  // 719162 is the number of days from 0001-01-01 to 1970-01-01,
-  // assuming the Gregorian calendar.
-  return Time(
-      time_internal::MakeDuration(-24 * 719162 * int64_t{3600}, uint32_t{0}));
-}
-
-// InfiniteFuture()
-//
-// Returns an `absl::Time` that is infinitely far in the future.
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time InfiniteFuture() {
-  return Time(time_internal::MakeDuration((std::numeric_limits<int64_t>::max)(),
-                                          ~uint32_t{0}));
-}
-
-// InfinitePast()
-//
-// Returns an `absl::Time` that is infinitely far in the past.
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time InfinitePast() {
-  return Time(time_internal::MakeDuration((std::numeric_limits<int64_t>::min)(),
-                                          ~uint32_t{0}));
-}
-
-// FromUnixNanos()
-// FromUnixMicros()
-// FromUnixMillis()
-// FromUnixSeconds()
-// FromTimeT()
-// FromUDate()
-// FromUniversal()
-//
-// Creates an `absl::Time` from a variety of other representations.  See
-// https://unicode-org.github.io/icu/userguide/datetime/universaltimescale.html
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time FromUnixNanos(int64_t ns);
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time FromUnixMicros(int64_t us);
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time FromUnixMillis(int64_t ms);
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time FromUnixSeconds(int64_t s);
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time FromTimeT(time_t t);
-ABSL_ATTRIBUTE_CONST_FUNCTION Time FromUDate(double udate);
-ABSL_ATTRIBUTE_CONST_FUNCTION Time FromUniversal(int64_t universal);
-
-// ToUnixNanos()
-// ToUnixMicros()
-// ToUnixMillis()
-// ToUnixSeconds()
-// ToTimeT()
-// ToUDate()
-// ToUniversal()
-//
-// Converts an `absl::Time` to a variety of other representations.  See
-// https://unicode-org.github.io/icu/userguide/datetime/universaltimescale.html
-//
-// Note that these operations round down toward negative infinity where
-// necessary to adjust to the resolution of the result type.  Beware of
-// possible time_t over/underflow in ToTime{T,val,spec}() on 32-bit platforms.
-ABSL_ATTRIBUTE_CONST_FUNCTION int64_t ToUnixNanos(Time t);
-ABSL_ATTRIBUTE_CONST_FUNCTION int64_t ToUnixMicros(Time t);
-ABSL_ATTRIBUTE_CONST_FUNCTION int64_t ToUnixMillis(Time t);
-ABSL_ATTRIBUTE_CONST_FUNCTION int64_t ToUnixSeconds(Time t);
-ABSL_ATTRIBUTE_CONST_FUNCTION time_t ToTimeT(Time t);
-ABSL_ATTRIBUTE_CONST_FUNCTION double ToUDate(Time t);
-ABSL_ATTRIBUTE_CONST_FUNCTION int64_t ToUniversal(Time t);
-
-// DurationFromTimespec()
-// DurationFromTimeval()
-// ToTimespec()
-// ToTimeval()
-// TimeFromTimespec()
-// TimeFromTimeval()
-// ToTimespec()
-// ToTimeval()
-//
-// Some APIs use a timespec or a timeval as a Duration (e.g., nanosleep(2)
-// and select(2)), while others use them as a Time (e.g. clock_gettime(2)
-// and gettimeofday(2)), so conversion functions are provided for both cases.
-// The "to timespec/val" direction is easily handled via overloading, but
-// for "from timespec/val" the desired type is part of the function name.
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration DurationFromTimespec(timespec ts);
-ABSL_ATTRIBUTE_CONST_FUNCTION Duration DurationFromTimeval(timeval tv);
-ABSL_ATTRIBUTE_CONST_FUNCTION timespec ToTimespec(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION timeval ToTimeval(Duration d);
-ABSL_ATTRIBUTE_CONST_FUNCTION Time TimeFromTimespec(timespec ts);
-ABSL_ATTRIBUTE_CONST_FUNCTION Time TimeFromTimeval(timeval tv);
-ABSL_ATTRIBUTE_CONST_FUNCTION timespec ToTimespec(Time t);
-ABSL_ATTRIBUTE_CONST_FUNCTION timeval ToTimeval(Time t);
-
-// FromChrono()
-//
-// Converts a std::chrono::system_clock::time_point to an absl::Time.
-//
-// Example:
-//
-//   auto tp = std::chrono::system_clock::from_time_t(123);
-//   absl::Time t = absl::FromChrono(tp);
-//   // t == absl::FromTimeT(123)
-ABSL_ATTRIBUTE_PURE_FUNCTION Time
-FromChrono(const std::chrono::system_clock::time_point& tp);
-
-// ToChronoTime()
-//
-// Converts an absl::Time to a std::chrono::system_clock::time_point. If
-// overflow would occur, the returned value will saturate at the min/max time
-// point value instead.
-//
-// Example:
-//
-//   absl::Time t = absl::FromTimeT(123);
-//   auto tp = absl::ToChronoTime(t);
-//   // tp == std::chrono::system_clock::from_time_t(123);
-ABSL_ATTRIBUTE_CONST_FUNCTION std::chrono::system_clock::time_point
-    ToChronoTime(Time);
-
-// AbslParseFlag()
-//
-// Parses the command-line flag string representation `text` into a Time value.
-// Time flags must be specified in a format that matches absl::RFC3339_full.
-//
-// For example:
-//
-//   --start_time=2016-01-02T03:04:05.678+08:00
-//
-// Note: A UTC offset (or 'Z' indicating a zero-offset from UTC) is required.
-//
-// Additionally, if you'd like to specify a time as a count of
-// seconds/milliseconds/etc from the Unix epoch, use an absl::Duration flag
-// and add that duration to absl::UnixEpoch() to get an absl::Time.
-bool AbslParseFlag(absl::string_view text, Time* t, std::string* error);
-
-// AbslUnparseFlag()
-//
-// Unparses a Time value into a command-line string representation using
-// the format specified by `absl::ParseTime()`.
-std::string AbslUnparseFlag(Time t);
-
-ABSL_DEPRECATED("Use AbslParseFlag() instead.")
-bool ParseFlag(const std::string& text, Time* t, std::string* error);
-ABSL_DEPRECATED("Use AbslUnparseFlag() instead.")
-std::string UnparseFlag(Time t);
-
-// TimeZone
-//
-// The `absl::TimeZone` is an opaque, small, value-type class representing a
-// geo-political region within which particular rules are used for converting
-// between absolute and civil times (see https://git.io/v59Ly). `absl::TimeZone`
-// values are named using the TZ identifiers from the IANA Time Zone Database,
-// such as "America/Los_Angeles" or "Australia/Sydney". `absl::TimeZone` values
-// are created from factory functions such as `absl::LoadTimeZone()`. Note:
-// strings like "PST" and "EDT" are not valid TZ identifiers. Prefer to pass by
-// value rather than const reference.
-//
-// For more on the fundamental concepts of time zones, absolute times, and civil
-// times, see https://github.com/google/cctz#fundamental-concepts
-//
-// Examples:
-//
-//   absl::TimeZone utc = absl::UTCTimeZone();
-//   absl::TimeZone pst = absl::FixedTimeZone(-8 * 60 * 60);
-//   absl::TimeZone loc = absl::LocalTimeZone();
-//   absl::TimeZone lax;
-//   if (!absl::LoadTimeZone("America/Los_Angeles", &lax)) {
-//     // handle error case
-//   }
-//
-// See also:
-// - https://github.com/google/cctz
-// - https://www.iana.org/time-zones
-// - https://en.wikipedia.org/wiki/Zoneinfo
-class TimeZone {
- public:
-  explicit TimeZone(time_internal::cctz::time_zone tz) : cz_(tz) {}
-  TimeZone() = default;  // UTC, but prefer UTCTimeZone() to be explicit.
-
-  // Copyable.
-  TimeZone(const TimeZone&) = default;
-  TimeZone& operator=(const TimeZone&) = default;
-
-  explicit operator time_internal::cctz::time_zone() const { return cz_; }
-
-  std::string name() const { return cz_.name(); }
-
-  // TimeZone::CivilInfo
-  //
-  // Information about the civil time corresponding to an absolute time.
-  // This struct is not intended to represent an instant in time. So, rather
-  // than passing a `TimeZone::CivilInfo` to a function, pass an `absl::Time`
-  // and an `absl::TimeZone`.
-  struct CivilInfo {
-    CivilSecond cs;
-    Duration subsecond;
-
-    // Note: The following fields exist for backward compatibility
-    // with older APIs.  Accessing these fields directly is a sign of
-    // imprudent logic in the calling code.  Modern time-related code
-    // should only access this data indirectly by way of FormatTime().
-    // These fields are undefined for InfiniteFuture() and InfinitePast().
-    int offset;             // seconds east of UTC
-    bool is_dst;            // is offset non-standard?
-    const char* zone_abbr;  // time-zone abbreviation (e.g., "PST")
-  };
-
-  // TimeZone::At(Time)
-  //
-  // Returns the civil time for this TimeZone at a certain `absl::Time`.
-  // If the input time is infinite, the output civil second will be set to
-  // CivilSecond::max() or min(), and the subsecond will be infinite.
-  //
-  // Example:
-  //
-  //   const auto epoch = lax.At(absl::UnixEpoch());
-  //   // epoch.cs == 1969-12-31 16:00:00
-  //   // epoch.subsecond == absl::ZeroDuration()
-  //   // epoch.offset == -28800
-  //   // epoch.is_dst == false
-  //   // epoch.abbr == "PST"
-  CivilInfo At(Time t) const;
-
-  // TimeZone::TimeInfo
-  //
-  // Information about the absolute times corresponding to a civil time.
-  // (Subseconds must be handled separately.)
-  //
-  // It is possible for a caller to pass a civil-time value that does
-  // not represent an actual or unique instant in time (due to a shift
-  // in UTC offset in the TimeZone, which results in a discontinuity in
-  // the civil-time components). For example, a daylight-saving-time
-  // transition skips or repeats civil times---in the United States,
-  // March 13, 2011 02:15 never occurred, while November 6, 2011 01:15
-  // occurred twice---so requests for such times are not well-defined.
-  // To account for these possibilities, `absl::TimeZone::TimeInfo` is
-  // richer than just a single `absl::Time`.
-  struct TimeInfo {
-    enum CivilKind {
-      UNIQUE,    // the civil time was singular (pre == trans == post)
-      SKIPPED,   // the civil time did not exist (pre >= trans > post)
-      REPEATED,  // the civil time was ambiguous (pre < trans <= post)
-    } kind;
-    Time pre;    // time calculated using the pre-transition offset
-    Time trans;  // when the civil-time discontinuity occurred
-    Time post;   // time calculated using the post-transition offset
-  };
-
-  // TimeZone::At(CivilSecond)
-  //
-  // Returns an `absl::TimeInfo` containing the absolute time(s) for this
-  // TimeZone at an `absl::CivilSecond`. When the civil time is skipped or
-  // repeated, returns times calculated using the pre-transition and post-
-  // transition UTC offsets, plus the transition time itself.
-  //
-  // Examples:
-  //
-  //   // A unique civil time
-  //   const auto jan01 = lax.At(absl::CivilSecond(2011, 1, 1, 0, 0, 0));
-  //   // jan01.kind == TimeZone::TimeInfo::UNIQUE
-  //   // jan01.pre    is 2011-01-01 00:00:00 -0800
-  //   // jan01.trans  is 2011-01-01 00:00:00 -0800
-  //   // jan01.post   is 2011-01-01 00:00:00 -0800
-  //
-  //   // A Spring DST transition, when there is a gap in civil time
-  //   const auto mar13 = lax.At(absl::CivilSecond(2011, 3, 13, 2, 15, 0));
-  //   // mar13.kind == TimeZone::TimeInfo::SKIPPED
-  //   // mar13.pre   is 2011-03-13 03:15:00 -0700
-  //   // mar13.trans is 2011-03-13 03:00:00 -0700
-  //   // mar13.post  is 2011-03-13 01:15:00 -0800
-  //
-  //   // A Fall DST transition, when civil times are repeated
-  //   const auto nov06 = lax.At(absl::CivilSecond(2011, 11, 6, 1, 15, 0));
-  //   // nov06.kind == TimeZone::TimeInfo::REPEATED
-  //   // nov06.pre   is 2011-11-06 01:15:00 -0700
-  //   // nov06.trans is 2011-11-06 01:00:00 -0800
-  //   // nov06.post  is 2011-11-06 01:15:00 -0800
-  TimeInfo At(CivilSecond ct) const;
-
-  // TimeZone::NextTransition()
-  // TimeZone::PrevTransition()
-  //
-  // Finds the time of the next/previous offset change in this time zone.
-  //
-  // By definition, `NextTransition(t, &trans)` returns false when `t` is
-  // `InfiniteFuture()`, and `PrevTransition(t, &trans)` returns false
-  // when `t` is `InfinitePast()`. If the zone has no transitions, the
-  // result will also be false no matter what the argument.
-  //
-  // Otherwise, when `t` is `InfinitePast()`, `NextTransition(t, &trans)`
-  // returns true and sets `trans` to the first recorded transition. Chains
-  // of calls to `NextTransition()/PrevTransition()` will eventually return
-  // false, but it is unspecified exactly when `NextTransition(t, &trans)`
-  // jumps to false, or what time is set by `PrevTransition(t, &trans)` for
-  // a very distant `t`.
-  //
-  // Note: Enumeration of time-zone transitions is for informational purposes
-  // only. Modern time-related code should not care about when offset changes
-  // occur.
-  //
-  // Example:
-  //   absl::TimeZone nyc;
-  //   if (!absl::LoadTimeZone("America/New_York", &nyc)) { ... }
-  //   const auto now = absl::Now();
-  //   auto t = absl::InfinitePast();
-  //   absl::TimeZone::CivilTransition trans;
-  //   while (t <= now && nyc.NextTransition(t, &trans)) {
-  //     // transition: trans.from -> trans.to
-  //     t = nyc.At(trans.to).trans;
-  //   }
-  struct CivilTransition {
-    CivilSecond from;  // the civil time we jump from
-    CivilSecond to;    // the civil time we jump to
-  };
-  bool NextTransition(Time t, CivilTransition* trans) const;
-  bool PrevTransition(Time t, CivilTransition* trans) const;
-
-  template <typename H>
-  friend H AbslHashValue(H h, TimeZone tz) {
-    return H::combine(std::move(h), tz.cz_);
-  }
-
- private:
-  friend bool operator==(TimeZone a, TimeZone b) { return a.cz_ == b.cz_; }
-  friend bool operator!=(TimeZone a, TimeZone b) { return a.cz_ != b.cz_; }
-  friend std::ostream& operator<<(std::ostream& os, TimeZone tz) {
-    return os << tz.name();
-  }
-
-  time_internal::cctz::time_zone cz_;
-};
-
-// LoadTimeZone()
-//
-// Loads the named zone. May perform I/O on the initial load of the named
-// zone. If the name is invalid, or some other kind of error occurs, returns
-// `false` and `*tz` is set to the UTC time zone.
-inline bool LoadTimeZone(absl::string_view name, TimeZone* tz) {
-  if (name == "localtime") {
-    *tz = TimeZone(time_internal::cctz::local_time_zone());
-    return true;
-  }
-  time_internal::cctz::time_zone cz;
-  const bool b = time_internal::cctz::load_time_zone(std::string(name), &cz);
-  *tz = TimeZone(cz);
-  return b;
-}
-
-// FixedTimeZone()
-//
-// Returns a TimeZone that is a fixed offset (seconds east) from UTC.
-// Note: If the absolute value of the offset is greater than 24 hours
-// you'll get UTC (i.e., no offset) instead.
-inline TimeZone FixedTimeZone(int seconds) {
-  return TimeZone(
-      time_internal::cctz::fixed_time_zone(std::chrono::seconds(seconds)));
-}
-
-// UTCTimeZone()
-//
-// Convenience method returning the UTC time zone.
-inline TimeZone UTCTimeZone() {
-  return TimeZone(time_internal::cctz::utc_time_zone());
-}
-
-// LocalTimeZone()
-//
-// Convenience method returning the local time zone, or UTC if there is
-// no configured local zone.  Warning: Be wary of using LocalTimeZone(),
-// and particularly so in a server process, as the zone configured for the
-// local machine should be irrelevant.  Prefer an explicit zone name.
-inline TimeZone LocalTimeZone() {
-  return TimeZone(time_internal::cctz::local_time_zone());
-}
-
-// ToCivilSecond()
-// ToCivilMinute()
-// ToCivilHour()
-// ToCivilDay()
-// ToCivilMonth()
-// ToCivilYear()
-//
-// Helpers for TimeZone::At(Time) to return particularly aligned civil times.
-//
-// Example:
-//
-//   absl::Time t = ...;
-//   absl::TimeZone tz = ...;
-//   const auto cd = absl::ToCivilDay(t, tz);
-ABSL_ATTRIBUTE_PURE_FUNCTION inline CivilSecond ToCivilSecond(Time t,
-                                                              TimeZone tz) {
-  return tz.At(t).cs;  // already a CivilSecond
-}
-ABSL_ATTRIBUTE_PURE_FUNCTION inline CivilMinute ToCivilMinute(Time t,
-                                                              TimeZone tz) {
-  return CivilMinute(tz.At(t).cs);
-}
-ABSL_ATTRIBUTE_PURE_FUNCTION inline CivilHour ToCivilHour(Time t, TimeZone tz) {
-  return CivilHour(tz.At(t).cs);
-}
-ABSL_ATTRIBUTE_PURE_FUNCTION inline CivilDay ToCivilDay(Time t, TimeZone tz) {
-  return CivilDay(tz.At(t).cs);
-}
-ABSL_ATTRIBUTE_PURE_FUNCTION inline CivilMonth ToCivilMonth(Time t,
-                                                            TimeZone tz) {
-  return CivilMonth(tz.At(t).cs);
-}
-ABSL_ATTRIBUTE_PURE_FUNCTION inline CivilYear ToCivilYear(Time t, TimeZone tz) {
-  return CivilYear(tz.At(t).cs);
-}
-
-// FromCivil()
-//
-// Helper for TimeZone::At(CivilSecond) that provides "order-preserving
-// semantics." If the civil time maps to a unique time, that time is
-// returned. If the civil time is repeated in the given time zone, the
-// time using the pre-transition offset is returned. Otherwise, the
-// civil time is skipped in the given time zone, and the transition time
-// is returned. This means that for any two civil times, ct1 and ct2,
-// (ct1 < ct2) => (FromCivil(ct1) <= FromCivil(ct2)), the equal case
-// being when two non-existent civil times map to the same transition time.
-//
-// Note: Accepts civil times of any alignment.
-ABSL_ATTRIBUTE_PURE_FUNCTION inline Time FromCivil(CivilSecond ct,
-                                                   TimeZone tz) {
-  const auto ti = tz.At(ct);
-  if (ti.kind == TimeZone::TimeInfo::SKIPPED) return ti.trans;
-  return ti.pre;
-}
-
-// TimeConversion
-//
-// An `absl::TimeConversion` represents the conversion of year, month, day,
-// hour, minute, and second values (i.e., a civil time), in a particular
-// `absl::TimeZone`, to a time instant (an absolute time), as returned by
-// `absl::ConvertDateTime()`. Legacy version of `absl::TimeZone::TimeInfo`.
-//
-// Deprecated. Use `absl::TimeZone::TimeInfo`.
-struct ABSL_DEPRECATED("Use `absl::TimeZone::TimeInfo`.") TimeConversion {
-  Time pre;    // time calculated using the pre-transition offset
-  Time trans;  // when the civil-time discontinuity occurred
-  Time post;   // time calculated using the post-transition offset
-
-  enum Kind {
-    UNIQUE,    // the civil time was singular (pre == trans == post)
-    SKIPPED,   // the civil time did not exist
-    REPEATED,  // the civil time was ambiguous
-  };
-  Kind kind;
-
-  bool normalized;  // input values were outside their valid ranges
-};
-
-// ConvertDateTime()
-//
-// Legacy version of `absl::TimeZone::At(absl::CivilSecond)` that takes
-// the civil time as six, separate values (YMDHMS).
-//
-// The input month, day, hour, minute, and second values can be outside
-// of their valid ranges, in which case they will be "normalized" during
-// the conversion.
-//
-// Example:
-//
-//   // "October 32" normalizes to "November 1".
-//   absl::TimeConversion tc =
-//       absl::ConvertDateTime(2013, 10, 32, 8, 30, 0, lax);
-//   // tc.kind == TimeConversion::UNIQUE && tc.normalized == true
-//   // absl::ToCivilDay(tc.pre, tz).month() == 11
-//   // absl::ToCivilDay(tc.pre, tz).day() == 1
-//
-// Deprecated. Use `absl::TimeZone::At(CivilSecond)`.
-ABSL_INTERNAL_DISABLE_DEPRECATED_DECLARATION_WARNING
-ABSL_DEPRECATED("Use `absl::TimeZone::At(CivilSecond)`.")
-TimeConversion ConvertDateTime(int64_t year, int mon, int day, int hour,
-                               int min, int sec, TimeZone tz);
-ABSL_INTERNAL_RESTORE_DEPRECATED_DECLARATION_WARNING
-
-// FromDateTime()
-//
-// A convenience wrapper for `absl::ConvertDateTime()` that simply returns
-// the "pre" `absl::Time`.  That is, the unique result, or the instant that
-// is correct using the pre-transition offset (as if the transition never
-// happened).
-//
-// Example:
-//
-//   absl::Time t = absl::FromDateTime(2017, 9, 26, 9, 30, 0, lax);
-//   // t = 2017-09-26 09:30:00 -0700
-//
-// Deprecated. Use `absl::FromCivil(CivilSecond, TimeZone)`. Note that the
-// behavior of `FromCivil()` differs from `FromDateTime()` for skipped civil
-// times. If you care about that see `absl::TimeZone::At(absl::CivilSecond)`.
-ABSL_DEPRECATED("Use `absl::FromCivil(CivilSecond, TimeZone)`.")
-inline Time FromDateTime(int64_t year, int mon, int day, int hour, int min,
-                         int sec, TimeZone tz) {
-  ABSL_INTERNAL_DISABLE_DEPRECATED_DECLARATION_WARNING
-  return ConvertDateTime(year, mon, day, hour, min, sec, tz).pre;
-  ABSL_INTERNAL_RESTORE_DEPRECATED_DECLARATION_WARNING
-}
-
-// FromTM()
-//
-// Converts the `tm_year`, `tm_mon`, `tm_mday`, `tm_hour`, `tm_min`, and
-// `tm_sec` fields to an `absl::Time` using the given time zone. See ctime(3)
-// for a description of the expected values of the tm fields. If the civil time
-// is unique (see `absl::TimeZone::At(absl::CivilSecond)` above), the matching
-// time instant is returned.  Otherwise, the `tm_isdst` field is consulted to
-// choose between the possible results.  For a repeated civil time, `tm_isdst !=
-// 0` returns the matching DST instant, while `tm_isdst == 0` returns the
-// matching non-DST instant.  For a skipped civil time there is no matching
-// instant, so `tm_isdst != 0` returns the DST instant, and `tm_isdst == 0`
-// returns the non-DST instant, that would have matched if the transition never
-// happened.
-ABSL_ATTRIBUTE_PURE_FUNCTION Time FromTM(const struct tm& tm, TimeZone tz);
-
-// ToTM()
-//
-// Converts the given `absl::Time` to a struct tm using the given time zone.
-// See ctime(3) for a description of the values of the tm fields.
-ABSL_ATTRIBUTE_PURE_FUNCTION struct tm ToTM(Time t, TimeZone tz);
-
-// RFC3339_full
-// RFC3339_sec
-//
-// FormatTime()/ParseTime() format specifiers for RFC3339 date/time strings,
-// with trailing zeros trimmed or with fractional seconds omitted altogether.
-//
-// Note that RFC3339_sec[] matches an ISO 8601 extended format for date and
-// time with UTC offset.  Also note the use of "%Y": RFC3339 mandates that
-// years have exactly four digits, but we allow them to take their natural
-// width.
-ABSL_DLL extern const char RFC3339_full[];  // %Y-%m-%d%ET%H:%M:%E*S%Ez
-ABSL_DLL extern const char RFC3339_sec[];   // %Y-%m-%d%ET%H:%M:%S%Ez
-
-// RFC1123_full
-// RFC1123_no_wday
-//
-// FormatTime()/ParseTime() format specifiers for RFC1123 date/time strings.
-ABSL_DLL extern const char RFC1123_full[];     // %a, %d %b %E4Y %H:%M:%S %z
-ABSL_DLL extern const char RFC1123_no_wday[];  // %d %b %E4Y %H:%M:%S %z
-
-// FormatTime()
-//
-// Formats the given `absl::Time` in the `absl::TimeZone` according to the
-// provided format string. Uses strftime()-like formatting options, with
-// the following extensions:
-//
-//   - %Ez  - RFC3339-compatible numeric UTC offset (+hh:mm or -hh:mm)
-//   - %E*z - Full-resolution numeric UTC offset (+hh:mm:ss or -hh:mm:ss)
-//   - %E#S - Seconds with # digits of fractional precision
-//   - %E*S - Seconds with full fractional precision (a literal '*')
-//   - %E#f - Fractional seconds with # digits of precision
-//   - %E*f - Fractional seconds with full precision (a literal '*')
-//   - %E4Y - Four-character years (-999 ... -001, 0000, 0001 ... 9999)
-//   - %ET  - The RFC3339 "date-time" separator "T"
-//
-// Note that %E0S behaves like %S, and %E0f produces no characters.  In
-// contrast %E*f always produces at least one digit, which may be '0'.
-//
-// Note that %Y produces as many characters as it takes to fully render the
-// year.  A year outside of [-999:9999] when formatted with %E4Y will produce
-// more than four characters, just like %Y.
-//
-// We recommend that format strings include the UTC offset (%z, %Ez, or %E*z)
-// so that the result uniquely identifies a time instant.
-//
-// Example:
-//
-//   absl::CivilSecond cs(2013, 1, 2, 3, 4, 5);
-//   absl::Time t = absl::FromCivil(cs, lax);
-//   std::string f = absl::FormatTime("%H:%M:%S", t, lax);  // "03:04:05"
-//   f = absl::FormatTime("%H:%M:%E3S", t, lax);  // "03:04:05.000"
-//
-// Note: If the given `absl::Time` is `absl::InfiniteFuture()`, the returned
-// string will be exactly "infinite-future". If the given `absl::Time` is
-// `absl::InfinitePast()`, the returned string will be exactly "infinite-past".
-// In both cases the given format string and `absl::TimeZone` are ignored.
-//
-ABSL_ATTRIBUTE_PURE_FUNCTION std::string FormatTime(absl::string_view format,
-                                                    Time t, TimeZone tz);
-
-// Convenience functions that format the given time using the RFC3339_full
-// format.  The first overload uses the provided TimeZone, while the second
-// uses LocalTimeZone().
-ABSL_ATTRIBUTE_PURE_FUNCTION std::string FormatTime(Time t, TimeZone tz);
-ABSL_ATTRIBUTE_PURE_FUNCTION std::string FormatTime(Time t);
-
-// Output stream operator.
-inline std::ostream& operator<<(std::ostream& os, Time t) {
-  return os << FormatTime(t);
-}
-
-// Support for StrFormat(), StrCat() etc.
-template <typename Sink>
-void AbslStringify(Sink& sink, Time t) {
-  sink.Append(FormatTime(t));
-}
-
-// ParseTime()
-//
-// Parses an input string according to the provided format string and
-// returns the corresponding `absl::Time`. Uses strftime()-like formatting
-// options, with the same extensions as FormatTime(), but with the
-// exceptions that %E#S is interpreted as %E*S, and %E#f as %E*f.  %Ez
-// and %E*z also accept the same inputs, which (along with %z) includes
-// 'z' and 'Z' as synonyms for +00:00.  %ET accepts either 'T' or 't'.
-//
-// %Y consumes as many numeric characters as it can, so the matching data
-// should always be terminated with a non-numeric.  %E4Y always consumes
-// exactly four characters, including any sign.
-//
-// Unspecified fields are taken from the default date and time of ...
-//
-//   "1970-01-01 00:00:00.0 +0000"
-//
-// For example, parsing a string of "15:45" (%H:%M) will return an absl::Time
-// that represents "1970-01-01 15:45:00.0 +0000".
-//
-// Note that since ParseTime() returns time instants, it makes the most sense
-// to parse fully-specified date/time strings that include a UTC offset (%z,
-// %Ez, or %E*z).
-//
-// Note also that `absl::ParseTime()` only heeds the fields year, month, day,
-// hour, minute, (fractional) second, and UTC offset.  Other fields, like
-// weekday (%a or %A), while parsed for syntactic validity, are ignored
-// in the conversion.
-//
-// Date and time fields that are out-of-range will be treated as errors
-// rather than normalizing them like `absl::CivilSecond` does.  For example,
-// it is an error to parse the date "Oct 32, 2013" because 32 is out of range.
-//
-// A leap second of ":60" is normalized to ":00" of the following minute
-// with fractional seconds discarded.  The following table shows how the
-// given seconds and subseconds will be parsed:
-//
-//   "59.x" -> 59.x  // exact
-//   "60.x" -> 00.0  // normalized
-//   "00.x" -> 00.x  // exact
-//
-// Errors are indicated by returning false and assigning an error message
-// to the "err" out param if it is non-null.
-//
-// Note: If the input string is exactly "infinite-future", the returned
-// `absl::Time` will be `absl::InfiniteFuture()` and `true` will be returned.
-// If the input string is "infinite-past", the returned `absl::Time` will be
-// `absl::InfinitePast()` and `true` will be returned.
-//
-bool ParseTime(absl::string_view format, absl::string_view input, Time* time,
-               std::string* err);
-
-// Like ParseTime() above, but if the format string does not contain a UTC
-// offset specification (%z/%Ez/%E*z) then the input is interpreted in the
-// given TimeZone.  This means that the input, by itself, does not identify a
-// unique instant.  Being time-zone dependent, it also admits the possibility
-// of ambiguity or non-existence, in which case the "pre" time (as defined
-// by TimeZone::TimeInfo) is returned.  For these reasons we recommend that
-// all date/time strings include a UTC offset so they're context independent.
-bool ParseTime(absl::string_view format, absl::string_view input, TimeZone tz,
-               Time* time, std::string* err);
-
-// ============================================================================
-// Implementation Details Follow
-// ============================================================================
-
-namespace time_internal {
-
-// Creates a Duration with a given representation.
-// REQUIRES: hi,lo is a valid representation of a Duration as specified
-// in time/duration.cc.
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration MakeDuration(int64_t hi,
-                                                              uint32_t lo = 0) {
-  return Duration(hi, lo);
-}
-
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration MakeDuration(int64_t hi,
-                                                              int64_t lo) {
-  return MakeDuration(hi, static_cast<uint32_t>(lo));
-}
-
-// Make a Duration value from a floating-point number, as long as that number
-// is in the range [ 0 .. numeric_limits<int64_t>::max ), that is, as long as
-// it's positive and can be converted to int64_t without risk of UB.
-ABSL_ATTRIBUTE_CONST_FUNCTION inline Duration MakePosDoubleDuration(double n) {
-  const int64_t int_secs = static_cast<int64_t>(n);
-  const uint32_t ticks = static_cast<uint32_t>(
-      std::round((n - static_cast<double>(int_secs)) * kTicksPerSecond));
-  return ticks < kTicksPerSecond
-             ? MakeDuration(int_secs, ticks)
-             : MakeDuration(int_secs + 1, ticks - kTicksPerSecond);
-}
-
-// Creates a normalized Duration from an almost-normalized (sec,ticks)
-// pair. sec may be positive or negative.  ticks must be in the range
-// -kTicksPerSecond < *ticks < kTicksPerSecond.  If ticks is negative it
-// will be normalized to a positive value in the resulting Duration.
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration MakeNormalizedDuration(
-    int64_t sec, int64_t ticks) {
-  return (ticks < 0) ? MakeDuration(sec - 1, ticks + kTicksPerSecond)
-                     : MakeDuration(sec, ticks);
-}
-
-// Provide access to the Duration representation.
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr int64_t GetRepHi(Duration d) {
-  return d.rep_hi_.Get();
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr uint32_t GetRepLo(Duration d) {
-  return d.rep_lo_;
-}
-
-// Returns true iff d is positive or negative infinity.
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr bool IsInfiniteDuration(Duration d) {
-  return GetRepLo(d) == ~uint32_t{0};
-}
-
-// Returns an infinite Duration with the opposite sign.
-// REQUIRES: IsInfiniteDuration(d)
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration OppositeInfinity(Duration d) {
-  return GetRepHi(d) < 0
-             ? MakeDuration((std::numeric_limits<int64_t>::max)(), ~uint32_t{0})
-             : MakeDuration((std::numeric_limits<int64_t>::min)(),
-                            ~uint32_t{0});
-}
-
-// Returns (-n)-1 (equivalently -(n+1)) without avoidable overflow.
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr int64_t NegateAndSubtractOne(
-    int64_t n) {
-  // Note: Good compilers will optimize this expression to ~n when using
-  // a two's-complement representation (which is required for int64_t).
-  return (n < 0) ? -(n + 1) : (-n) - 1;
-}
-
-// Map between a Time and a Duration since the Unix epoch.  Note that these
-// functions depend on the above mentioned choice of the Unix epoch for the
-// Time representation (and both need to be Time friends).  Without this
-// knowledge, we would need to add-in/subtract-out UnixEpoch() respectively.
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time FromUnixDuration(Duration d) {
-  return Time(d);
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration ToUnixDuration(Time t) {
-  return t.rep_;
-}
-
-template <std::intmax_t N>
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration FromInt64(int64_t v,
-                                                           std::ratio<1, N>) {
-  static_assert(0 < N && N <= 1000 * 1000 * 1000, "Unsupported ratio");
-  // Subsecond ratios cannot overflow.
-  return MakeNormalizedDuration(
-      v / N, v % N * kTicksPerNanosecond * 1000 * 1000 * 1000 / N);
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration FromInt64(int64_t v,
-                                                           std::ratio<60>) {
-  return (v <= (std::numeric_limits<int64_t>::max)() / 60 &&
-          v >= (std::numeric_limits<int64_t>::min)() / 60)
-             ? MakeDuration(v * 60)
-             : v > 0 ? InfiniteDuration() : -InfiniteDuration();
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration FromInt64(int64_t v,
-                                                           std::ratio<3600>) {
-  return (v <= (std::numeric_limits<int64_t>::max)() / 3600 &&
-          v >= (std::numeric_limits<int64_t>::min)() / 3600)
-             ? MakeDuration(v * 3600)
-             : v > 0 ? InfiniteDuration() : -InfiniteDuration();
-}
-
-// IsValidRep64<T>(0) is true if the expression `int64_t{std::declval<T>()}` is
-// valid. That is, if a T can be assigned to an int64_t without narrowing.
-template <typename T>
-constexpr auto IsValidRep64(int) -> decltype(int64_t{std::declval<T>()} == 0) {
-  return true;
-}
-template <typename T>
-constexpr auto IsValidRep64(char) -> bool {
-  return false;
-}
-
-// Converts a std::chrono::duration to an absl::Duration.
-template <typename Rep, typename Period>
-ABSL_ATTRIBUTE_PURE_FUNCTION constexpr Duration FromChrono(
-    const std::chrono::duration<Rep, Period>& d) {
-  static_assert(IsValidRep64<Rep>(0), "duration::rep is invalid");
-  return FromInt64(int64_t{d.count()}, Period{});
-}
-
-template <typename Ratio>
-ABSL_ATTRIBUTE_CONST_FUNCTION int64_t ToInt64(Duration d, Ratio) {
-  // Note: This may be used on MSVC, which may have a system_clock period of
-  // std::ratio<1, 10 * 1000 * 1000>
-  return ToInt64Seconds(d * Ratio::den / Ratio::num);
-}
-// Fastpath implementations for the 6 common duration units.
-ABSL_ATTRIBUTE_CONST_FUNCTION inline int64_t ToInt64(Duration d, std::nano) {
-  return ToInt64Nanoseconds(d);
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION inline int64_t ToInt64(Duration d, std::micro) {
-  return ToInt64Microseconds(d);
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION inline int64_t ToInt64(Duration d, std::milli) {
-  return ToInt64Milliseconds(d);
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION inline int64_t ToInt64(Duration d,
-                                                     std::ratio<1>) {
-  return ToInt64Seconds(d);
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION inline int64_t ToInt64(Duration d,
-                                                     std::ratio<60>) {
-  return ToInt64Minutes(d);
-}
-ABSL_ATTRIBUTE_CONST_FUNCTION inline int64_t ToInt64(Duration d,
-                                                     std::ratio<3600>) {
-  return ToInt64Hours(d);
-}
-
-// Converts an absl::Duration to a chrono duration of type T.
-template <typename T>
-ABSL_ATTRIBUTE_CONST_FUNCTION T ToChronoDuration(Duration d) {
-  using Rep = typename T::rep;
-  using Period = typename T::period;
-  static_assert(IsValidRep64<Rep>(0), "duration::rep is invalid");
-  if (time_internal::IsInfiniteDuration(d))
-    return d < ZeroDuration() ? (T::min)() : (T::max)();
-  const auto v = ToInt64(d, Period{});
-  if (v > (std::numeric_limits<Rep>::max)()) return (T::max)();
-  if (v < (std::numeric_limits<Rep>::min)()) return (T::min)();
-  return T{v};
-}
-
-}  // namespace time_internal
-
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr bool operator<(Duration lhs,
-                                                       Duration rhs) {
-  return time_internal::GetRepHi(lhs) != time_internal::GetRepHi(rhs)
-             ? time_internal::GetRepHi(lhs) < time_internal::GetRepHi(rhs)
-         : time_internal::GetRepHi(lhs) == (std::numeric_limits<int64_t>::min)()
-             ? time_internal::GetRepLo(lhs) + 1 <
-                   time_internal::GetRepLo(rhs) + 1
-             : time_internal::GetRepLo(lhs) < time_internal::GetRepLo(rhs);
-}
-
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr bool operator==(Duration lhs,
-                                                        Duration rhs) {
-  return time_internal::GetRepHi(lhs) == time_internal::GetRepHi(rhs) &&
-         time_internal::GetRepLo(lhs) == time_internal::GetRepLo(rhs);
-}
-
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration operator-(Duration d) {
-  // This is a little interesting because of the special cases.
-  //
-  // If rep_lo_ is zero, we have it easy; it's safe to negate rep_hi_, we're
-  // dealing with an integral number of seconds, and the only special case is
-  // the maximum negative finite duration, which can't be negated.
-  //
-  // Infinities stay infinite, and just change direction.
-  //
-  // Finally we're in the case where rep_lo_ is non-zero, and we can borrow
-  // a second's worth of ticks and avoid overflow (as negating int64_t-min + 1
-  // is safe).
-  return time_internal::GetRepLo(d) == 0
-             ? time_internal::GetRepHi(d) ==
-                       (std::numeric_limits<int64_t>::min)()
-                   ? InfiniteDuration()
-                   : time_internal::MakeDuration(-time_internal::GetRepHi(d))
-             : time_internal::IsInfiniteDuration(d)
-                   ? time_internal::OppositeInfinity(d)
-                   : time_internal::MakeDuration(
-                         time_internal::NegateAndSubtractOne(
-                             time_internal::GetRepHi(d)),
-                         time_internal::kTicksPerSecond -
-                             time_internal::GetRepLo(d));
-}
-
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Duration InfiniteDuration() {
-  return time_internal::MakeDuration((std::numeric_limits<int64_t>::max)(),
-                                     ~uint32_t{0});
-}
-
-ABSL_ATTRIBUTE_PURE_FUNCTION constexpr Duration FromChrono(
-    const std::chrono::nanoseconds& d) {
-  return time_internal::FromChrono(d);
-}
-ABSL_ATTRIBUTE_PURE_FUNCTION constexpr Duration FromChrono(
-    const std::chrono::microseconds& d) {
-  return time_internal::FromChrono(d);
-}
-ABSL_ATTRIBUTE_PURE_FUNCTION constexpr Duration FromChrono(
-    const std::chrono::milliseconds& d) {
-  return time_internal::FromChrono(d);
-}
-ABSL_ATTRIBUTE_PURE_FUNCTION constexpr Duration FromChrono(
-    const std::chrono::seconds& d) {
-  return time_internal::FromChrono(d);
-}
-ABSL_ATTRIBUTE_PURE_FUNCTION constexpr Duration FromChrono(
-    const std::chrono::minutes& d) {
-  return time_internal::FromChrono(d);
-}
-ABSL_ATTRIBUTE_PURE_FUNCTION constexpr Duration FromChrono(
-    const std::chrono::hours& d) {
-  return time_internal::FromChrono(d);
-}
-
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time FromUnixNanos(int64_t ns) {
-  return time_internal::FromUnixDuration(Nanoseconds(ns));
-}
-
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time FromUnixMicros(int64_t us) {
-  return time_internal::FromUnixDuration(Microseconds(us));
-}
-
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time FromUnixMillis(int64_t ms) {
-  return time_internal::FromUnixDuration(Milliseconds(ms));
-}
-
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time FromUnixSeconds(int64_t s) {
-  return time_internal::FromUnixDuration(Seconds(s));
-}
-
-ABSL_ATTRIBUTE_CONST_FUNCTION constexpr Time FromTimeT(time_t t) {
-  return time_internal::FromUnixDuration(Seconds(t));
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TIME_TIME_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/types/any.h b/projects/llm_framework/include/abseil-cpp/absl/types/any.h
deleted file mode 100644
index 61f071f1..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/types/any.h
+++ /dev/null
@@ -1,519 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// any.h
-// -----------------------------------------------------------------------------
-//
-// This header file define the `absl::any` type for holding a type-safe value
-// of any type. The 'absl::any` type is useful for providing a way to hold
-// something that is, as yet, unspecified. Such unspecified types
-// traditionally are passed between API boundaries until they are later cast to
-// their "destination" types. To cast to such a destination type, use
-// `absl::any_cast()`. Note that when casting an `absl::any`, you must cast it
-// to an explicit type; implicit conversions will throw.
-//
-// Example:
-//
-//   auto a = absl::any(65);
-//   absl::any_cast<int>(a);         // 65
-//   absl::any_cast<char>(a);        // throws absl::bad_any_cast
-//   absl::any_cast<std::string>(a); // throws absl::bad_any_cast
-//
-// `absl::any` is a C++11 compatible version of the C++17 `std::any` abstraction
-// and is designed to be a drop-in replacement for code compliant with C++17.
-//
-// Traditionally, the behavior of casting to a temporary unspecified type has
-// been accomplished with the `void *` paradigm, where the pointer was to some
-// other unspecified type. `absl::any` provides an "owning" version of `void *`
-// that avoids issues of pointer management.
-//
-// Note: just as in the case of `void *`, use of `absl::any` (and its C++17
-// version `std::any`) is a code smell indicating that your API might not be
-// constructed correctly. We have seen that most uses of `any` are unwarranted,
-// and `absl::any`, like `std::any`, is difficult to use properly. Before using
-// this abstraction, make sure that you should not instead be rewriting your
-// code to be more specific.
-//
-// Abseil has also released an `absl::variant` type (a C++11 compatible version
-// of the C++17 `std::variant`), which is generally preferred for use over
-// `absl::any`.
-#ifndef ABSL_TYPES_ANY_H_
-#define ABSL_TYPES_ANY_H_
-
-#include "absl/base/attributes.h"
-#include "absl/base/config.h"
-#include "absl/utility/utility.h"
-
-#ifdef ABSL_USES_STD_ANY
-
-#include <any>  // IWYU pragma: export
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-using std::any;
-using std::any_cast;
-using std::bad_any_cast;
-using std::make_any;
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#else  // ABSL_USES_STD_ANY
-
-#include <algorithm>
-#include <cstddef>
-#include <initializer_list>
-#include <memory>
-#include <stdexcept>
-#include <type_traits>
-#include <typeinfo>
-#include <utility>
-
-#include "absl/base/internal/fast_type_id.h"
-#include "absl/meta/type_traits.h"
-#include "absl/types/bad_any_cast.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-class any;
-
-// swap()
-//
-// Swaps two `absl::any` values. Equivalent to `x.swap(y) where `x` and `y` are
-// `absl::any` types.
-void swap(any& x, any& y) noexcept;
-
-// make_any()
-//
-// Constructs an `absl::any` of type `T` with the given arguments.
-template <typename T, typename... Args>
-any make_any(Args&&... args);
-
-// Overload of `absl::make_any()` for constructing an `absl::any` type from an
-// initializer list.
-template <typename T, typename U, typename... Args>
-any make_any(std::initializer_list<U> il, Args&&... args);
-
-// any_cast()
-//
-// Statically casts the value of a `const absl::any` type to the given type.
-// This function will throw `absl::bad_any_cast` if the stored value type of the
-// `absl::any` does not match the cast.
-//
-// `any_cast()` can also be used to get a reference to the internal storage iff
-// a reference type is passed as its `ValueType`:
-//
-// Example:
-//
-//   absl::any my_any = std::vector<int>();
-//   absl::any_cast<std::vector<int>&>(my_any).push_back(42);
-template <typename ValueType>
-ValueType any_cast(const any& operand);
-
-// Overload of `any_cast()` to statically cast the value of a non-const
-// `absl::any` type to the given type. This function will throw
-// `absl::bad_any_cast` if the stored value type of the `absl::any` does not
-// match the cast.
-template <typename ValueType>
-ValueType any_cast(any& operand);  // NOLINT(runtime/references)
-
-// Overload of `any_cast()` to statically cast the rvalue of an `absl::any`
-// type. This function will throw `absl::bad_any_cast` if the stored value type
-// of the `absl::any` does not match the cast.
-template <typename ValueType>
-ValueType any_cast(any&& operand);
-
-// Overload of `any_cast()` to statically cast the value of a const pointer
-// `absl::any` type to the given pointer type, or `nullptr` if the stored value
-// type of the `absl::any` does not match the cast.
-template <typename ValueType>
-const ValueType* any_cast(const any* operand) noexcept;
-
-// Overload of `any_cast()` to statically cast the value of a pointer
-// `absl::any` type to the given pointer type, or `nullptr` if the stored value
-// type of the `absl::any` does not match the cast.
-template <typename ValueType>
-ValueType* any_cast(any* operand) noexcept;
-
-// -----------------------------------------------------------------------------
-// absl::any
-// -----------------------------------------------------------------------------
-//
-// An `absl::any` object provides the facility to either store an instance of a
-// type, known as the "contained object", or no value. An `absl::any` is used to
-// store values of types that are unknown at compile time. The `absl::any`
-// object, when containing a value, must contain a value type; storing a
-// reference type is neither desired nor supported.
-//
-// An `absl::any` can only store a type that is copy-constructible; move-only
-// types are not allowed within an `any` object.
-//
-// Example:
-//
-//   auto a = absl::any(65);                 // Literal, copyable
-//   auto b = absl::any(std::vector<int>()); // Default-initialized, copyable
-//   std::unique_ptr<Foo> my_foo;
-//   auto c = absl::any(std::move(my_foo));  // Error, not copy-constructible
-//
-// Note that `absl::any` makes use of decayed types (`absl::decay_t` in this
-// context) to remove const-volatile qualifiers (known as "cv qualifiers"),
-// decay functions to function pointers, etc. We essentially "decay" a given
-// type into its essential type.
-//
-// `absl::any` makes use of decayed types when determining the basic type `T` of
-// the value to store in the any's contained object. In the documentation below,
-// we explicitly denote this by using the phrase "a decayed type of `T`".
-//
-// Example:
-//
-//   const int a = 4;
-//   absl::any foo(a);  // Decay ensures we store an "int", not a "const int&".
-//
-//   void my_function() {}
-//   absl::any bar(my_function);  // Decay ensures we store a function pointer.
-//
-// `absl::any` is a C++11 compatible version of the C++17 `std::any` abstraction
-// and is designed to be a drop-in replacement for code compliant with C++17.
-class any {
- private:
-  template <typename T>
-  struct IsInPlaceType;
-
- public:
-  // Constructors
-
-  // Constructs an empty `absl::any` object (`any::has_value()` will return
-  // `false`).
-  constexpr any() noexcept;
-
-  // Copy constructs an `absl::any` object with a "contained object" of the
-  // passed type of `other` (or an empty `absl::any` if `other.has_value()` is
-  // `false`.
-  any(const any& other)
-      : obj_(other.has_value() ? other.obj_->Clone()
-                               : std::unique_ptr<ObjInterface>()) {}
-
-  // Move constructs an `absl::any` object with a "contained object" of the
-  // passed type of `other` (or an empty `absl::any` if `other.has_value()` is
-  // `false`).
-  any(any&& other) noexcept = default;
-
-  // Constructs an `absl::any` object with a "contained object" of the decayed
-  // type of `T`, which is initialized via `std::forward<T>(value)`.
-  //
-  // This constructor will not participate in overload resolution if the
-  // decayed type of `T` is not copy-constructible.
-  template <
-      typename T, typename VT = absl::decay_t<T>,
-      absl::enable_if_t<!absl::disjunction<
-          std::is_same<any, VT>, IsInPlaceType<VT>,
-          absl::negation<std::is_copy_constructible<VT> > >::value>* = nullptr>
-  any(T&& value) : obj_(new Obj<VT>(in_place, std::forward<T>(value))) {}
-
-  // Constructs an `absl::any` object with a "contained object" of the decayed
-  // type of `T`, which is initialized via `std::forward<T>(value)`.
-  template <typename T, typename... Args, typename VT = absl::decay_t<T>,
-            absl::enable_if_t<absl::conjunction<
-                std::is_copy_constructible<VT>,
-                std::is_constructible<VT, Args...>>::value>* = nullptr>
-  explicit any(in_place_type_t<T> /*tag*/, Args&&... args)
-      : obj_(new Obj<VT>(in_place, std::forward<Args>(args)...)) {}
-
-  // Constructs an `absl::any` object with a "contained object" of the passed
-  // type `VT` as a decayed type of `T`. `VT` is initialized as if
-  // direct-non-list-initializing an object of type `VT` with the arguments
-  // `initializer_list, std::forward<Args>(args)...`.
-  template <
-      typename T, typename U, typename... Args, typename VT = absl::decay_t<T>,
-      absl::enable_if_t<
-          absl::conjunction<std::is_copy_constructible<VT>,
-                            std::is_constructible<VT, std::initializer_list<U>&,
-                                                  Args...>>::value>* = nullptr>
-  explicit any(in_place_type_t<T> /*tag*/, std::initializer_list<U> ilist,
-               Args&&... args)
-      : obj_(new Obj<VT>(in_place, ilist, std::forward<Args>(args)...)) {}
-
-  // Assignment operators
-
-  // Copy assigns an `absl::any` object with a "contained object" of the
-  // passed type.
-  any& operator=(const any& rhs) {
-    any(rhs).swap(*this);
-    return *this;
-  }
-
-  // Move assigns an `absl::any` object with a "contained object" of the
-  // passed type. `rhs` is left in a valid but otherwise unspecified state.
-  any& operator=(any&& rhs) noexcept {
-    any(std::move(rhs)).swap(*this);
-    return *this;
-  }
-
-  // Assigns an `absl::any` object with a "contained object" of the passed type.
-  template <typename T, typename VT = absl::decay_t<T>,
-            absl::enable_if_t<absl::conjunction<
-                absl::negation<std::is_same<VT, any>>,
-                std::is_copy_constructible<VT>>::value>* = nullptr>
-  any& operator=(T&& rhs) {
-    any tmp(in_place_type_t<VT>(), std::forward<T>(rhs));
-    tmp.swap(*this);
-    return *this;
-  }
-
-  // Modifiers
-
-  // any::emplace()
-  //
-  // Emplaces a value within an `absl::any` object by calling `any::reset()`,
-  // initializing the contained value as if direct-non-list-initializing an
-  // object of type `VT` with the arguments `std::forward<Args>(args)...`, and
-  // returning a reference to the new contained value.
-  //
-  // Note: If an exception is thrown during the call to `VT`'s constructor,
-  // `*this` does not contain a value, and any previously contained value has
-  // been destroyed.
-  template <
-      typename T, typename... Args, typename VT = absl::decay_t<T>,
-      absl::enable_if_t<std::is_copy_constructible<VT>::value &&
-                        std::is_constructible<VT, Args...>::value>* = nullptr>
-  VT& emplace(Args&&... args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    reset();  // NOTE: reset() is required here even in the world of exceptions.
-    Obj<VT>* const object_ptr =
-        new Obj<VT>(in_place, std::forward<Args>(args)...);
-    obj_ = std::unique_ptr<ObjInterface>(object_ptr);
-    return object_ptr->value;
-  }
-
-  // Overload of `any::emplace()` to emplace a value within an `absl::any`
-  // object by calling `any::reset()`, initializing the contained value as if
-  // direct-non-list-initializing an object of type `VT` with the arguments
-  // `initializer_list, std::forward<Args>(args)...`, and returning a reference
-  // to the new contained value.
-  //
-  // Note: If an exception is thrown during the call to `VT`'s constructor,
-  // `*this` does not contain a value, and any previously contained value has
-  // been destroyed. The function shall not participate in overload resolution
-  // unless `is_copy_constructible_v<VT>` is `true` and
-  // `is_constructible_v<VT, initializer_list<U>&, Args...>` is `true`.
-  template <
-      typename T, typename U, typename... Args, typename VT = absl::decay_t<T>,
-      absl::enable_if_t<std::is_copy_constructible<VT>::value &&
-                        std::is_constructible<VT, std::initializer_list<U>&,
-                                              Args...>::value>* = nullptr>
-  VT& emplace(std::initializer_list<U> ilist,
-              Args&&... args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    reset();  // NOTE: reset() is required here even in the world of exceptions.
-    Obj<VT>* const object_ptr =
-        new Obj<VT>(in_place, ilist, std::forward<Args>(args)...);
-    obj_ = std::unique_ptr<ObjInterface>(object_ptr);
-    return object_ptr->value;
-  }
-
-  // any::reset()
-  //
-  // Resets the state of the `absl::any` object, destroying the contained object
-  // if present.
-  void reset() noexcept { obj_ = nullptr; }
-
-  // any::swap()
-  //
-  // Swaps the passed value and the value of this `absl::any` object.
-  void swap(any& other) noexcept { obj_.swap(other.obj_); }
-
-  // Observers
-
-  // any::has_value()
-  //
-  // Returns `true` if the `any` object has a contained value, otherwise
-  // returns `false`.
-  bool has_value() const noexcept { return obj_ != nullptr; }
-
-#ifdef ABSL_INTERNAL_HAS_RTTI
-  // Returns: typeid(T) if *this has a contained object of type T, otherwise
-  // typeid(void).
-  const std::type_info& type() const noexcept {
-    if (has_value()) {
-      return obj_->Type();
-    }
-
-    return typeid(void);
-  }
-#endif  // ABSL_INTERNAL_HAS_RTTI
-
- private:
-  // Tagged type-erased abstraction for holding a cloneable object.
-  class ObjInterface {
-   public:
-    virtual ~ObjInterface() = default;
-    virtual std::unique_ptr<ObjInterface> Clone() const = 0;
-    virtual const void* ObjTypeId() const noexcept = 0;
-#ifdef ABSL_INTERNAL_HAS_RTTI
-    virtual const std::type_info& Type() const noexcept = 0;
-#endif  // ABSL_INTERNAL_HAS_RTTI
-  };
-
-  // Hold a value of some queryable type, with an ability to Clone it.
-  template <typename T>
-  class Obj : public ObjInterface {
-   public:
-    template <typename... Args>
-    explicit Obj(in_place_t /*tag*/, Args&&... args)
-        : value(std::forward<Args>(args)...) {}
-
-    std::unique_ptr<ObjInterface> Clone() const final {
-      return std::unique_ptr<ObjInterface>(new Obj(in_place, value));
-    }
-
-    const void* ObjTypeId() const noexcept final { return IdForType<T>(); }
-
-#ifdef ABSL_INTERNAL_HAS_RTTI
-    const std::type_info& Type() const noexcept final { return typeid(T); }
-#endif  // ABSL_INTERNAL_HAS_RTTI
-
-    T value;
-  };
-
-  std::unique_ptr<ObjInterface> CloneObj() const {
-    if (!obj_) return nullptr;
-    return obj_->Clone();
-  }
-
-  template <typename T>
-  constexpr static const void* IdForType() {
-    // Note: This type dance is to make the behavior consistent with typeid.
-    using NormalizedType =
-        typename std::remove_cv<typename std::remove_reference<T>::type>::type;
-
-    return base_internal::FastTypeId<NormalizedType>();
-  }
-
-  const void* GetObjTypeId() const {
-    return obj_ ? obj_->ObjTypeId() : base_internal::FastTypeId<void>();
-  }
-
-  // `absl::any` nonmember functions //
-
-  // Description at the declaration site (top of file).
-  template <typename ValueType>
-  friend ValueType any_cast(const any& operand);
-
-  // Description at the declaration site (top of file).
-  template <typename ValueType>
-  friend ValueType any_cast(any& operand);  // NOLINT(runtime/references)
-
-  // Description at the declaration site (top of file).
-  template <typename T>
-  friend const T* any_cast(const any* operand) noexcept;
-
-  // Description at the declaration site (top of file).
-  template <typename T>
-  friend T* any_cast(any* operand) noexcept;
-
-  std::unique_ptr<ObjInterface> obj_;
-};
-
-// -----------------------------------------------------------------------------
-// Implementation Details
-// -----------------------------------------------------------------------------
-
-constexpr any::any() noexcept = default;
-
-template <typename T>
-struct any::IsInPlaceType : std::false_type {};
-
-template <typename T>
-struct any::IsInPlaceType<in_place_type_t<T>> : std::true_type {};
-
-inline void swap(any& x, any& y) noexcept { x.swap(y); }
-
-// Description at the declaration site (top of file).
-template <typename T, typename... Args>
-any make_any(Args&&... args) {
-  return any(in_place_type_t<T>(), std::forward<Args>(args)...);
-}
-
-// Description at the declaration site (top of file).
-template <typename T, typename U, typename... Args>
-any make_any(std::initializer_list<U> il, Args&&... args) {
-  return any(in_place_type_t<T>(), il, std::forward<Args>(args)...);
-}
-
-// Description at the declaration site (top of file).
-template <typename ValueType>
-ValueType any_cast(const any& operand) {
-  using U = typename std::remove_cv<
-      typename std::remove_reference<ValueType>::type>::type;
-  static_assert(std::is_constructible<ValueType, const U&>::value,
-                "Invalid ValueType");
-  auto* const result = (any_cast<U>)(&operand);
-  if (result == nullptr) {
-    any_internal::ThrowBadAnyCast();
-  }
-  return static_cast<ValueType>(*result);
-}
-
-// Description at the declaration site (top of file).
-template <typename ValueType>
-ValueType any_cast(any& operand) {  // NOLINT(runtime/references)
-  using U = typename std::remove_cv<
-      typename std::remove_reference<ValueType>::type>::type;
-  static_assert(std::is_constructible<ValueType, U&>::value,
-                "Invalid ValueType");
-  auto* result = (any_cast<U>)(&operand);
-  if (result == nullptr) {
-    any_internal::ThrowBadAnyCast();
-  }
-  return static_cast<ValueType>(*result);
-}
-
-// Description at the declaration site (top of file).
-template <typename ValueType>
-ValueType any_cast(any&& operand) {
-  using U = typename std::remove_cv<
-      typename std::remove_reference<ValueType>::type>::type;
-  static_assert(std::is_constructible<ValueType, U>::value,
-                "Invalid ValueType");
-  return static_cast<ValueType>(std::move((any_cast<U&>)(operand)));
-}
-
-// Description at the declaration site (top of file).
-template <typename T>
-const T* any_cast(const any* operand) noexcept {
-  using U =
-      typename std::remove_cv<typename std::remove_reference<T>::type>::type;
-  return operand && operand->GetObjTypeId() == any::IdForType<U>()
-             ? std::addressof(
-                   static_cast<const any::Obj<U>*>(operand->obj_.get())->value)
-             : nullptr;
-}
-
-// Description at the declaration site (top of file).
-template <typename T>
-T* any_cast(any* operand) noexcept {
-  using U =
-      typename std::remove_cv<typename std::remove_reference<T>::type>::type;
-  return operand && operand->GetObjTypeId() == any::IdForType<U>()
-             ? std::addressof(
-                   static_cast<any::Obj<U>*>(operand->obj_.get())->value)
-             : nullptr;
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_USES_STD_ANY
-
-#endif  // ABSL_TYPES_ANY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/types/bad_any_cast.h b/projects/llm_framework/include/abseil-cpp/absl/types/bad_any_cast.h
deleted file mode 100644
index 114cef80..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/types/bad_any_cast.h
+++ /dev/null
@@ -1,75 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// bad_any_cast.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines the `absl::bad_any_cast` type.
-
-#ifndef ABSL_TYPES_BAD_ANY_CAST_H_
-#define ABSL_TYPES_BAD_ANY_CAST_H_
-
-#include <typeinfo>
-
-#include "absl/base/config.h"
-
-#ifdef ABSL_USES_STD_ANY
-
-#include <any>
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-using std::bad_any_cast;
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#else  // ABSL_USES_STD_ANY
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// -----------------------------------------------------------------------------
-// bad_any_cast
-// -----------------------------------------------------------------------------
-//
-// An `absl::bad_any_cast` type is an exception type that is thrown when
-// failing to successfully cast the return value of an `absl::any` object.
-//
-// Example:
-//
-//   auto a = absl::any(65);
-//   absl::any_cast<int>(a);         // 65
-//   try {
-//     absl::any_cast<char>(a);
-//   } catch(const absl::bad_any_cast& e) {
-//     std::cout << "Bad any cast: " << e.what() << '\n';
-//   }
-class bad_any_cast : public std::bad_cast {
- public:
-  ~bad_any_cast() override;
-  const char* what() const noexcept override;
-};
-
-namespace any_internal {
-
-[[noreturn]] void ThrowBadAnyCast();
-
-}  // namespace any_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_USES_STD_ANY
-
-#endif  // ABSL_TYPES_BAD_ANY_CAST_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/types/bad_optional_access.h b/projects/llm_framework/include/abseil-cpp/absl/types/bad_optional_access.h
deleted file mode 100644
index 049e72ad..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/types/bad_optional_access.h
+++ /dev/null
@@ -1,78 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// bad_optional_access.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines the `absl::bad_optional_access` type.
-
-#ifndef ABSL_TYPES_BAD_OPTIONAL_ACCESS_H_
-#define ABSL_TYPES_BAD_OPTIONAL_ACCESS_H_
-
-#include <stdexcept>
-
-#include "absl/base/config.h"
-
-#ifdef ABSL_USES_STD_OPTIONAL
-
-#include <optional>
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-using std::bad_optional_access;
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#else  // ABSL_USES_STD_OPTIONAL
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// -----------------------------------------------------------------------------
-// bad_optional_access
-// -----------------------------------------------------------------------------
-//
-// An `absl::bad_optional_access` type is an exception type that is thrown when
-// attempting to access an `absl::optional` object that does not contain a
-// value.
-//
-// Example:
-//
-//   absl::optional<int> o;
-//
-//   try {
-//     int n = o.value();
-//   } catch(const absl::bad_optional_access& e) {
-//     std::cout << "Bad optional access: " << e.what() << '\n';
-//   }
-class bad_optional_access : public std::exception {
- public:
-  bad_optional_access() = default;
-  ~bad_optional_access() override;
-  const char* what() const noexcept override;
-};
-
-namespace optional_internal {
-
-// throw delegator
-[[noreturn]] ABSL_DLL void throw_bad_optional_access();
-
-}  // namespace optional_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_USES_STD_OPTIONAL
-
-#endif  // ABSL_TYPES_BAD_OPTIONAL_ACCESS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/types/bad_variant_access.h b/projects/llm_framework/include/abseil-cpp/absl/types/bad_variant_access.h
deleted file mode 100644
index 8ab215e9..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/types/bad_variant_access.h
+++ /dev/null
@@ -1,82 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// bad_variant_access.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines the `absl::bad_variant_access` type.
-
-#ifndef ABSL_TYPES_BAD_VARIANT_ACCESS_H_
-#define ABSL_TYPES_BAD_VARIANT_ACCESS_H_
-
-#include <stdexcept>
-
-#include "absl/base/config.h"
-
-#ifdef ABSL_USES_STD_VARIANT
-
-#include <variant>
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-using std::bad_variant_access;
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#else  // ABSL_USES_STD_VARIANT
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// -----------------------------------------------------------------------------
-// bad_variant_access
-// -----------------------------------------------------------------------------
-//
-// An `absl::bad_variant_access` type is an exception type that is thrown in
-// the following cases:
-//
-//   * Calling `absl::get(absl::variant) with an index or type that does not
-//     match the currently selected alternative type
-//   * Calling `absl::visit on an `absl::variant` that is in the
-//     `variant::valueless_by_exception` state.
-//
-// Example:
-//
-//   absl::variant<int, std::string> v;
-//   v = 1;
-//   try {
-//     absl::get<std::string>(v);
-//   } catch(const absl::bad_variant_access& e) {
-//     std::cout << "Bad variant access: " << e.what() << '\n';
-//   }
-class bad_variant_access : public std::exception {
- public:
-  bad_variant_access() noexcept = default;
-  ~bad_variant_access() override;
-  const char* what() const noexcept override;
-};
-
-namespace variant_internal {
-
-[[noreturn]] ABSL_DLL void ThrowBadVariantAccess();
-[[noreturn]] ABSL_DLL void Rethrow();
-
-}  // namespace variant_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_USES_STD_VARIANT
-
-#endif  // ABSL_TYPES_BAD_VARIANT_ACCESS_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/types/compare.h b/projects/llm_framework/include/abseil-cpp/absl/types/compare.h
deleted file mode 100644
index 3cf4a917..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/types/compare.h
+++ /dev/null
@@ -1,505 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// compare.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines the `absl::partial_ordering`, `absl::weak_ordering`,
-// and `absl::strong_ordering` types for storing the results of three way
-// comparisons.
-//
-// Example:
-//   absl::weak_ordering compare(const std::string& a, const std::string& b);
-//
-// These are C++11 compatible versions of the C++20 corresponding types
-// (`std::partial_ordering`, etc.) and are designed to be drop-in replacements
-// for code compliant with C++20.
-
-#ifndef ABSL_TYPES_COMPARE_H_
-#define ABSL_TYPES_COMPARE_H_
-
-#include "absl/base/config.h"
-
-#ifdef ABSL_USES_STD_ORDERING
-
-#include <compare>  // IWYU pragma: export
-#include <type_traits>
-
-#include "absl/meta/type_traits.h"
-
-#else
-
-#include <cstddef>
-#include <cstdint>
-#include <cstdlib>
-#include <type_traits>
-
-#include "absl/base/attributes.h"
-#include "absl/base/macros.h"
-#include "absl/meta/type_traits.h"
-
-#endif
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-#ifdef ABSL_USES_STD_ORDERING
-
-using std::partial_ordering;
-using std::strong_ordering;
-using std::weak_ordering;
-
-#else
-
-namespace compare_internal {
-
-using value_type = int8_t;
-
-class OnlyLiteralZero {
- public:
-#if ABSL_HAVE_ATTRIBUTE(enable_if)
-  // On clang, we can avoid triggering modernize-use-nullptr by only enabling
-  // this overload when the value is a compile time integer constant equal to 0.
-  //
-  // In c++20, this could be a static_assert in a consteval function.
-  constexpr OnlyLiteralZero(int n)  // NOLINT
-      __attribute__((enable_if(n == 0, "Only literal `0` is allowed."))) {}
-#else  // ABSL_HAVE_ATTRIBUTE(enable_if)
-  // Accept only literal zero since it can be implicitly converted to a pointer
-  // to member type. nullptr constants will be caught by the other constructor
-  // which accepts a nullptr_t.
-  //
-  // This constructor is not used for clang since it triggers
-  // modernize-use-nullptr.
-  constexpr OnlyLiteralZero(int OnlyLiteralZero::*) noexcept {}  // NOLINT
-#endif
-
-  // Fails compilation when `nullptr` or integral type arguments other than
-  // `int` are passed. This constructor doesn't accept `int` because literal `0`
-  // has type `int`. Literal `0` arguments will be implicitly converted to
-  // `std::nullptr_t` and accepted by the above constructor, while other `int`
-  // arguments will fail to be converted and cause compilation failure.
-  template <typename T, typename = typename std::enable_if<
-                            std::is_same<T, std::nullptr_t>::value ||
-                            (std::is_integral<T>::value &&
-                             !std::is_same<T, int>::value)>::type>
-  OnlyLiteralZero(T) {  // NOLINT
-    static_assert(sizeof(T) < 0, "Only literal `0` is allowed.");
-  }
-};
-
-enum class eq : value_type {
-  equal = 0,
-  equivalent = equal,
-  nonequal = 1,
-  nonequivalent = nonequal,
-};
-
-enum class ord : value_type { less = -1, greater = 1 };
-
-enum class ncmp : value_type { unordered = -127 };
-
-// Define macros to allow for creation or emulation of C++17 inline variables
-// based on whether the feature is supported. Note: we can't use
-// ABSL_INTERNAL_INLINE_CONSTEXPR here because the variables here are of
-// incomplete types so they need to be defined after the types are complete.
-#ifdef __cpp_inline_variables
-
-// A no-op expansion that can be followed by a semicolon at class level.
-#define ABSL_COMPARE_INLINE_BASECLASS_DECL(name) static_assert(true, "")
-
-#define ABSL_COMPARE_INLINE_SUBCLASS_DECL(type, name) \
-  static const type name
-
-#define ABSL_COMPARE_INLINE_INIT(type, name, init) \
-  inline constexpr type type::name(init)
-
-#else  // __cpp_inline_variables
-
-#define ABSL_COMPARE_INLINE_BASECLASS_DECL(name) \
-  ABSL_CONST_INIT static const T name
-
-// A no-op expansion that can be followed by a semicolon at class level.
-#define ABSL_COMPARE_INLINE_SUBCLASS_DECL(type, name) static_assert(true, "")
-
-#define ABSL_COMPARE_INLINE_INIT(type, name, init) \
-  template <typename T>                            \
-  const T compare_internal::type##_base<T>::name(init)
-
-#endif  // __cpp_inline_variables
-
-// These template base classes allow for defining the values of the constants
-// in the header file (for performance) without using inline variables (which
-// aren't available in C++11).
-template <typename T>
-struct partial_ordering_base {
-  ABSL_COMPARE_INLINE_BASECLASS_DECL(less);
-  ABSL_COMPARE_INLINE_BASECLASS_DECL(equivalent);
-  ABSL_COMPARE_INLINE_BASECLASS_DECL(greater);
-  ABSL_COMPARE_INLINE_BASECLASS_DECL(unordered);
-};
-
-template <typename T>
-struct weak_ordering_base {
-  ABSL_COMPARE_INLINE_BASECLASS_DECL(less);
-  ABSL_COMPARE_INLINE_BASECLASS_DECL(equivalent);
-  ABSL_COMPARE_INLINE_BASECLASS_DECL(greater);
-};
-
-template <typename T>
-struct strong_ordering_base {
-  ABSL_COMPARE_INLINE_BASECLASS_DECL(less);
-  ABSL_COMPARE_INLINE_BASECLASS_DECL(equal);
-  ABSL_COMPARE_INLINE_BASECLASS_DECL(equivalent);
-  ABSL_COMPARE_INLINE_BASECLASS_DECL(greater);
-};
-
-}  // namespace compare_internal
-
-class partial_ordering
-    : public compare_internal::partial_ordering_base<partial_ordering> {
-  explicit constexpr partial_ordering(compare_internal::eq v) noexcept
-      : value_(static_cast<compare_internal::value_type>(v)) {}
-  explicit constexpr partial_ordering(compare_internal::ord v) noexcept
-      : value_(static_cast<compare_internal::value_type>(v)) {}
-  explicit constexpr partial_ordering(compare_internal::ncmp v) noexcept
-      : value_(static_cast<compare_internal::value_type>(v)) {}
-  friend struct compare_internal::partial_ordering_base<partial_ordering>;
-
-  constexpr bool is_ordered() const noexcept {
-    return value_ !=
-           compare_internal::value_type(compare_internal::ncmp::unordered);
-  }
-
- public:
-  ABSL_COMPARE_INLINE_SUBCLASS_DECL(partial_ordering, less);
-  ABSL_COMPARE_INLINE_SUBCLASS_DECL(partial_ordering, equivalent);
-  ABSL_COMPARE_INLINE_SUBCLASS_DECL(partial_ordering, greater);
-  ABSL_COMPARE_INLINE_SUBCLASS_DECL(partial_ordering, unordered);
-
-  // Comparisons
-  friend constexpr bool operator==(
-      partial_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.is_ordered() && v.value_ == 0;
-  }
-  friend constexpr bool operator!=(
-      partial_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return !v.is_ordered() || v.value_ != 0;
-  }
-  friend constexpr bool operator<(
-      partial_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.is_ordered() && v.value_ < 0;
-  }
-  friend constexpr bool operator<=(
-      partial_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.is_ordered() && v.value_ <= 0;
-  }
-  friend constexpr bool operator>(
-      partial_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.is_ordered() && v.value_ > 0;
-  }
-  friend constexpr bool operator>=(
-      partial_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.is_ordered() && v.value_ >= 0;
-  }
-  friend constexpr bool operator==(compare_internal::OnlyLiteralZero,
-                                   partial_ordering v) noexcept {
-    return v.is_ordered() && 0 == v.value_;
-  }
-  friend constexpr bool operator!=(compare_internal::OnlyLiteralZero,
-                                   partial_ordering v) noexcept {
-    return !v.is_ordered() || 0 != v.value_;
-  }
-  friend constexpr bool operator<(compare_internal::OnlyLiteralZero,
-                                  partial_ordering v) noexcept {
-    return v.is_ordered() && 0 < v.value_;
-  }
-  friend constexpr bool operator<=(compare_internal::OnlyLiteralZero,
-                                   partial_ordering v) noexcept {
-    return v.is_ordered() && 0 <= v.value_;
-  }
-  friend constexpr bool operator>(compare_internal::OnlyLiteralZero,
-                                  partial_ordering v) noexcept {
-    return v.is_ordered() && 0 > v.value_;
-  }
-  friend constexpr bool operator>=(compare_internal::OnlyLiteralZero,
-                                   partial_ordering v) noexcept {
-    return v.is_ordered() && 0 >= v.value_;
-  }
-  friend constexpr bool operator==(partial_ordering v1,
-                                   partial_ordering v2) noexcept {
-    return v1.value_ == v2.value_;
-  }
-  friend constexpr bool operator!=(partial_ordering v1,
-                                   partial_ordering v2) noexcept {
-    return v1.value_ != v2.value_;
-  }
-
- private:
-  compare_internal::value_type value_;
-};
-ABSL_COMPARE_INLINE_INIT(partial_ordering, less, compare_internal::ord::less);
-ABSL_COMPARE_INLINE_INIT(partial_ordering, equivalent,
-                         compare_internal::eq::equivalent);
-ABSL_COMPARE_INLINE_INIT(partial_ordering, greater,
-                         compare_internal::ord::greater);
-ABSL_COMPARE_INLINE_INIT(partial_ordering, unordered,
-                         compare_internal::ncmp::unordered);
-
-class weak_ordering
-    : public compare_internal::weak_ordering_base<weak_ordering> {
-  explicit constexpr weak_ordering(compare_internal::eq v) noexcept
-      : value_(static_cast<compare_internal::value_type>(v)) {}
-  explicit constexpr weak_ordering(compare_internal::ord v) noexcept
-      : value_(static_cast<compare_internal::value_type>(v)) {}
-  friend struct compare_internal::weak_ordering_base<weak_ordering>;
-
- public:
-  ABSL_COMPARE_INLINE_SUBCLASS_DECL(weak_ordering, less);
-  ABSL_COMPARE_INLINE_SUBCLASS_DECL(weak_ordering, equivalent);
-  ABSL_COMPARE_INLINE_SUBCLASS_DECL(weak_ordering, greater);
-
-  // Conversions
-  constexpr operator partial_ordering() const noexcept {  // NOLINT
-    return value_ == 0 ? partial_ordering::equivalent
-                       : (value_ < 0 ? partial_ordering::less
-                                     : partial_ordering::greater);
-  }
-  // Comparisons
-  friend constexpr bool operator==(
-      weak_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.value_ == 0;
-  }
-  friend constexpr bool operator!=(
-      weak_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.value_ != 0;
-  }
-  friend constexpr bool operator<(
-      weak_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.value_ < 0;
-  }
-  friend constexpr bool operator<=(
-      weak_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.value_ <= 0;
-  }
-  friend constexpr bool operator>(
-      weak_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.value_ > 0;
-  }
-  friend constexpr bool operator>=(
-      weak_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.value_ >= 0;
-  }
-  friend constexpr bool operator==(compare_internal::OnlyLiteralZero,
-                                   weak_ordering v) noexcept {
-    return 0 == v.value_;
-  }
-  friend constexpr bool operator!=(compare_internal::OnlyLiteralZero,
-                                   weak_ordering v) noexcept {
-    return 0 != v.value_;
-  }
-  friend constexpr bool operator<(compare_internal::OnlyLiteralZero,
-                                  weak_ordering v) noexcept {
-    return 0 < v.value_;
-  }
-  friend constexpr bool operator<=(compare_internal::OnlyLiteralZero,
-                                   weak_ordering v) noexcept {
-    return 0 <= v.value_;
-  }
-  friend constexpr bool operator>(compare_internal::OnlyLiteralZero,
-                                  weak_ordering v) noexcept {
-    return 0 > v.value_;
-  }
-  friend constexpr bool operator>=(compare_internal::OnlyLiteralZero,
-                                   weak_ordering v) noexcept {
-    return 0 >= v.value_;
-  }
-  friend constexpr bool operator==(weak_ordering v1,
-                                   weak_ordering v2) noexcept {
-    return v1.value_ == v2.value_;
-  }
-  friend constexpr bool operator!=(weak_ordering v1,
-                                   weak_ordering v2) noexcept {
-    return v1.value_ != v2.value_;
-  }
-
- private:
-  compare_internal::value_type value_;
-};
-ABSL_COMPARE_INLINE_INIT(weak_ordering, less, compare_internal::ord::less);
-ABSL_COMPARE_INLINE_INIT(weak_ordering, equivalent,
-                         compare_internal::eq::equivalent);
-ABSL_COMPARE_INLINE_INIT(weak_ordering, greater,
-                         compare_internal::ord::greater);
-
-class strong_ordering
-    : public compare_internal::strong_ordering_base<strong_ordering> {
-  explicit constexpr strong_ordering(compare_internal::eq v) noexcept
-      : value_(static_cast<compare_internal::value_type>(v)) {}
-  explicit constexpr strong_ordering(compare_internal::ord v) noexcept
-      : value_(static_cast<compare_internal::value_type>(v)) {}
-  friend struct compare_internal::strong_ordering_base<strong_ordering>;
-
- public:
-  ABSL_COMPARE_INLINE_SUBCLASS_DECL(strong_ordering, less);
-  ABSL_COMPARE_INLINE_SUBCLASS_DECL(strong_ordering, equal);
-  ABSL_COMPARE_INLINE_SUBCLASS_DECL(strong_ordering, equivalent);
-  ABSL_COMPARE_INLINE_SUBCLASS_DECL(strong_ordering, greater);
-
-  // Conversions
-  constexpr operator partial_ordering() const noexcept {  // NOLINT
-    return value_ == 0 ? partial_ordering::equivalent
-                       : (value_ < 0 ? partial_ordering::less
-                                     : partial_ordering::greater);
-  }
-  constexpr operator weak_ordering() const noexcept {  // NOLINT
-    return value_ == 0
-               ? weak_ordering::equivalent
-               : (value_ < 0 ? weak_ordering::less : weak_ordering::greater);
-  }
-  // Comparisons
-  friend constexpr bool operator==(
-      strong_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.value_ == 0;
-  }
-  friend constexpr bool operator!=(
-      strong_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.value_ != 0;
-  }
-  friend constexpr bool operator<(
-      strong_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.value_ < 0;
-  }
-  friend constexpr bool operator<=(
-      strong_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.value_ <= 0;
-  }
-  friend constexpr bool operator>(
-      strong_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.value_ > 0;
-  }
-  friend constexpr bool operator>=(
-      strong_ordering v, compare_internal::OnlyLiteralZero) noexcept {
-    return v.value_ >= 0;
-  }
-  friend constexpr bool operator==(compare_internal::OnlyLiteralZero,
-                                   strong_ordering v) noexcept {
-    return 0 == v.value_;
-  }
-  friend constexpr bool operator!=(compare_internal::OnlyLiteralZero,
-                                   strong_ordering v) noexcept {
-    return 0 != v.value_;
-  }
-  friend constexpr bool operator<(compare_internal::OnlyLiteralZero,
-                                  strong_ordering v) noexcept {
-    return 0 < v.value_;
-  }
-  friend constexpr bool operator<=(compare_internal::OnlyLiteralZero,
-                                   strong_ordering v) noexcept {
-    return 0 <= v.value_;
-  }
-  friend constexpr bool operator>(compare_internal::OnlyLiteralZero,
-                                  strong_ordering v) noexcept {
-    return 0 > v.value_;
-  }
-  friend constexpr bool operator>=(compare_internal::OnlyLiteralZero,
-                                   strong_ordering v) noexcept {
-    return 0 >= v.value_;
-  }
-  friend constexpr bool operator==(strong_ordering v1,
-                                   strong_ordering v2) noexcept {
-    return v1.value_ == v2.value_;
-  }
-  friend constexpr bool operator!=(strong_ordering v1,
-                                   strong_ordering v2) noexcept {
-    return v1.value_ != v2.value_;
-  }
-
- private:
-  compare_internal::value_type value_;
-};
-ABSL_COMPARE_INLINE_INIT(strong_ordering, less, compare_internal::ord::less);
-ABSL_COMPARE_INLINE_INIT(strong_ordering, equal, compare_internal::eq::equal);
-ABSL_COMPARE_INLINE_INIT(strong_ordering, equivalent,
-                         compare_internal::eq::equivalent);
-ABSL_COMPARE_INLINE_INIT(strong_ordering, greater,
-                         compare_internal::ord::greater);
-
-#undef ABSL_COMPARE_INLINE_BASECLASS_DECL
-#undef ABSL_COMPARE_INLINE_SUBCLASS_DECL
-#undef ABSL_COMPARE_INLINE_INIT
-
-#endif  // ABSL_USES_STD_ORDERING
-
-namespace compare_internal {
-// We also provide these comparator adapter functions for internal absl use.
-
-// Helper functions to do a boolean comparison of two keys given a boolean
-// or three-way comparator.
-// SFINAE prevents implicit conversions to bool (such as from int).
-template <typename BoolT,
-          absl::enable_if_t<std::is_same<bool, BoolT>::value, int> = 0>
-constexpr bool compare_result_as_less_than(const BoolT r) { return r; }
-constexpr bool compare_result_as_less_than(const absl::weak_ordering r) {
-  return r < 0;
-}
-
-template <typename Compare, typename K, typename LK>
-constexpr bool do_less_than_comparison(const Compare &compare, const K &x,
-                                       const LK &y) {
-  return compare_result_as_less_than(compare(x, y));
-}
-
-// Helper functions to do a three-way comparison of two keys given a boolean or
-// three-way comparator.
-// SFINAE prevents implicit conversions to int (such as from bool).
-template <typename Int,
-          absl::enable_if_t<std::is_same<int, Int>::value, int> = 0>
-constexpr absl::weak_ordering compare_result_as_ordering(const Int c) {
-  return c < 0 ? absl::weak_ordering::less
-               : c == 0 ? absl::weak_ordering::equivalent
-                        : absl::weak_ordering::greater;
-}
-constexpr absl::weak_ordering compare_result_as_ordering(
-    const absl::weak_ordering c) {
-  return c;
-}
-
-template <
-    typename Compare, typename K, typename LK,
-    absl::enable_if_t<!std::is_same<bool, absl::result_of_t<Compare(
-                                              const K &, const LK &)>>::value,
-                      int> = 0>
-constexpr absl::weak_ordering do_three_way_comparison(const Compare &compare,
-                                                      const K &x, const LK &y) {
-  return compare_result_as_ordering(compare(x, y));
-}
-template <
-    typename Compare, typename K, typename LK,
-    absl::enable_if_t<std::is_same<bool, absl::result_of_t<Compare(
-                                             const K &, const LK &)>>::value,
-                      int> = 0>
-constexpr absl::weak_ordering do_three_way_comparison(const Compare &compare,
-                                                      const K &x, const LK &y) {
-  return compare(x, y) ? absl::weak_ordering::less
-                       : compare(y, x) ? absl::weak_ordering::greater
-                                       : absl::weak_ordering::equivalent;
-}
-
-}  // namespace compare_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TYPES_COMPARE_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/types/internal/optional.h b/projects/llm_framework/include/abseil-cpp/absl/types/internal/optional.h
deleted file mode 100644
index 5731a5bc..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/types/internal/optional.h
+++ /dev/null
@@ -1,352 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-#ifndef ABSL_TYPES_INTERNAL_OPTIONAL_H_
-#define ABSL_TYPES_INTERNAL_OPTIONAL_H_
-
-#include <functional>
-#include <new>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/internal/inline_variable.h"
-#include "absl/memory/memory.h"
-#include "absl/meta/type_traits.h"
-#include "absl/utility/utility.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// Forward declaration
-template <typename T>
-class optional;
-
-namespace optional_internal {
-
-// This tag type is used as a constructor parameter type for `nullopt_t`.
-struct init_t {
-  explicit init_t() = default;
-};
-
-struct empty_struct {};
-
-// This class stores the data in optional<T>.
-// It is specialized based on whether T is trivially destructible.
-// This is the specialization for non trivially destructible type.
-template <typename T, bool unused = std::is_trivially_destructible<T>::value>
-class optional_data_dtor_base {
-  struct dummy_type {
-    static_assert(sizeof(T) % sizeof(empty_struct) == 0, "");
-    // Use an array to avoid GCC 6 placement-new warning.
-    empty_struct data[sizeof(T) / sizeof(empty_struct)];
-  };
-
- protected:
-  // Whether there is data or not.
-  bool engaged_;
-  // Data storage
-  union {
-    T data_;
-    dummy_type dummy_;
-  };
-
-  void destruct() noexcept {
-    if (engaged_) {
-      // `data_` must be initialized if `engaged_` is true.
-#if ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(12, 0)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
-#endif
-      data_.~T();
-#if ABSL_INTERNAL_HAVE_MIN_GNUC_VERSION(12, 0)
-#pragma GCC diagnostic pop
-#endif
-      engaged_ = false;
-    }
-  }
-
-  // dummy_ must be initialized for constexpr constructor.
-  constexpr optional_data_dtor_base() noexcept : engaged_(false), dummy_{{}} {}
-
-  template <typename... Args>
-  constexpr explicit optional_data_dtor_base(in_place_t, Args&&... args)
-      : engaged_(true), data_(std::forward<Args>(args)...) {}
-
-  ~optional_data_dtor_base() { destruct(); }
-};
-
-// Specialization for trivially destructible type.
-template <typename T>
-class optional_data_dtor_base<T, true> {
-  struct dummy_type {
-    static_assert(sizeof(T) % sizeof(empty_struct) == 0, "");
-    // Use array to avoid GCC 6 placement-new warning.
-    empty_struct data[sizeof(T) / sizeof(empty_struct)];
-  };
-
- protected:
-  // Whether there is data or not.
-  bool engaged_;
-  // Data storage
-  union {
-    T data_;
-    dummy_type dummy_;
-  };
-  void destruct() noexcept { engaged_ = false; }
-
-  // dummy_ must be initialized for constexpr constructor.
-  constexpr optional_data_dtor_base() noexcept : engaged_(false), dummy_{{}} {}
-
-  template <typename... Args>
-  constexpr explicit optional_data_dtor_base(in_place_t, Args&&... args)
-      : engaged_(true), data_(std::forward<Args>(args)...) {}
-};
-
-template <typename T>
-class optional_data_base : public optional_data_dtor_base<T> {
- protected:
-  using base = optional_data_dtor_base<T>;
-  using base::base;
-
-  template <typename... Args>
-  void construct(Args&&... args) {
-    // Use dummy_'s address to work around casting cv-qualified T* to void*.
-    ::new (static_cast<void*>(&this->dummy_)) T(std::forward<Args>(args)...);
-    this->engaged_ = true;
-  }
-
-  template <typename U>
-  void assign(U&& u) {
-    if (this->engaged_) {
-      this->data_ = std::forward<U>(u);
-    } else {
-      construct(std::forward<U>(u));
-    }
-  }
-};
-
-// TODO(absl-team): Add another class using
-// std::is_trivially_move_constructible trait when available to match
-// http://cplusplus.github.io/LWG/lwg-defects.html#2900, for types that
-// have trivial move but nontrivial copy.
-// Also, we should be checking is_trivially_copyable here, which is not
-// supported now, so we use is_trivially_* traits instead.
-template <typename T,
-          bool unused = absl::is_trivially_copy_constructible<T>::value&&
-              absl::is_trivially_copy_assignable<typename std::remove_cv<
-                  T>::type>::value&& std::is_trivially_destructible<T>::value>
-class optional_data;
-
-// Trivially copyable types
-template <typename T>
-class optional_data<T, true> : public optional_data_base<T> {
- protected:
-  using optional_data_base<T>::optional_data_base;
-};
-
-template <typename T>
-class optional_data<T, false> : public optional_data_base<T> {
- protected:
-  using optional_data_base<T>::optional_data_base;
-
-  optional_data() = default;
-
-  optional_data(const optional_data& rhs) : optional_data_base<T>() {
-    if (rhs.engaged_) {
-      this->construct(rhs.data_);
-    }
-  }
-
-  optional_data(optional_data&& rhs) noexcept(
-      absl::default_allocator_is_nothrow::value ||
-      std::is_nothrow_move_constructible<T>::value)
-      : optional_data_base<T>() {
-    if (rhs.engaged_) {
-      this->construct(std::move(rhs.data_));
-    }
-  }
-
-  optional_data& operator=(const optional_data& rhs) {
-    if (rhs.engaged_) {
-      this->assign(rhs.data_);
-    } else {
-      this->destruct();
-    }
-    return *this;
-  }
-
-  optional_data& operator=(optional_data&& rhs) noexcept(
-      std::is_nothrow_move_assignable<T>::value&&
-          std::is_nothrow_move_constructible<T>::value) {
-    if (rhs.engaged_) {
-      this->assign(std::move(rhs.data_));
-    } else {
-      this->destruct();
-    }
-    return *this;
-  }
-};
-
-// Ordered by level of restriction, from low to high.
-// Copyable implies movable.
-enum class copy_traits { copyable = 0, movable = 1, non_movable = 2 };
-
-// Base class for enabling/disabling copy/move constructor.
-template <copy_traits>
-class optional_ctor_base;
-
-template <>
-class optional_ctor_base<copy_traits::copyable> {
- public:
-  constexpr optional_ctor_base() = default;
-  optional_ctor_base(const optional_ctor_base&) = default;
-  optional_ctor_base(optional_ctor_base&&) = default;
-  optional_ctor_base& operator=(const optional_ctor_base&) = default;
-  optional_ctor_base& operator=(optional_ctor_base&&) = default;
-};
-
-template <>
-class optional_ctor_base<copy_traits::movable> {
- public:
-  constexpr optional_ctor_base() = default;
-  optional_ctor_base(const optional_ctor_base&) = delete;
-  optional_ctor_base(optional_ctor_base&&) = default;
-  optional_ctor_base& operator=(const optional_ctor_base&) = default;
-  optional_ctor_base& operator=(optional_ctor_base&&) = default;
-};
-
-template <>
-class optional_ctor_base<copy_traits::non_movable> {
- public:
-  constexpr optional_ctor_base() = default;
-  optional_ctor_base(const optional_ctor_base&) = delete;
-  optional_ctor_base(optional_ctor_base&&) = delete;
-  optional_ctor_base& operator=(const optional_ctor_base&) = default;
-  optional_ctor_base& operator=(optional_ctor_base&&) = default;
-};
-
-// Base class for enabling/disabling copy/move assignment.
-template <copy_traits>
-class optional_assign_base;
-
-template <>
-class optional_assign_base<copy_traits::copyable> {
- public:
-  constexpr optional_assign_base() = default;
-  optional_assign_base(const optional_assign_base&) = default;
-  optional_assign_base(optional_assign_base&&) = default;
-  optional_assign_base& operator=(const optional_assign_base&) = default;
-  optional_assign_base& operator=(optional_assign_base&&) = default;
-};
-
-template <>
-class optional_assign_base<copy_traits::movable> {
- public:
-  constexpr optional_assign_base() = default;
-  optional_assign_base(const optional_assign_base&) = default;
-  optional_assign_base(optional_assign_base&&) = default;
-  optional_assign_base& operator=(const optional_assign_base&) = delete;
-  optional_assign_base& operator=(optional_assign_base&&) = default;
-};
-
-template <>
-class optional_assign_base<copy_traits::non_movable> {
- public:
-  constexpr optional_assign_base() = default;
-  optional_assign_base(const optional_assign_base&) = default;
-  optional_assign_base(optional_assign_base&&) = default;
-  optional_assign_base& operator=(const optional_assign_base&) = delete;
-  optional_assign_base& operator=(optional_assign_base&&) = delete;
-};
-
-template <typename T>
-struct ctor_copy_traits {
-  static constexpr copy_traits traits =
-      std::is_copy_constructible<T>::value
-          ? copy_traits::copyable
-          : std::is_move_constructible<T>::value ? copy_traits::movable
-                                                 : copy_traits::non_movable;
-};
-
-template <typename T>
-struct assign_copy_traits {
-  static constexpr copy_traits traits =
-      absl::is_copy_assignable<T>::value && std::is_copy_constructible<T>::value
-          ? copy_traits::copyable
-          : absl::is_move_assignable<T>::value &&
-                    std::is_move_constructible<T>::value
-                ? copy_traits::movable
-                : copy_traits::non_movable;
-};
-
-// Whether T is constructible or convertible from optional<U>.
-template <typename T, typename U>
-struct is_constructible_convertible_from_optional
-    : std::integral_constant<
-          bool, std::is_constructible<T, optional<U>&>::value ||
-                    std::is_constructible<T, optional<U>&&>::value ||
-                    std::is_constructible<T, const optional<U>&>::value ||
-                    std::is_constructible<T, const optional<U>&&>::value ||
-                    std::is_convertible<optional<U>&, T>::value ||
-                    std::is_convertible<optional<U>&&, T>::value ||
-                    std::is_convertible<const optional<U>&, T>::value ||
-                    std::is_convertible<const optional<U>&&, T>::value> {};
-
-// Whether T is constructible or convertible or assignable from optional<U>.
-template <typename T, typename U>
-struct is_constructible_convertible_assignable_from_optional
-    : std::integral_constant<
-          bool, is_constructible_convertible_from_optional<T, U>::value ||
-                    std::is_assignable<T&, optional<U>&>::value ||
-                    std::is_assignable<T&, optional<U>&&>::value ||
-                    std::is_assignable<T&, const optional<U>&>::value ||
-                    std::is_assignable<T&, const optional<U>&&>::value> {};
-
-// Helper function used by [optional.relops], [optional.comp_with_t],
-// for checking whether an expression is convertible to bool.
-bool convertible_to_bool(bool);
-
-// Base class for std::hash<absl::optional<T>>:
-// If std::hash<std::remove_const_t<T>> is enabled, it provides operator() to
-// compute the hash; Otherwise, it is disabled.
-// Reference N4659 23.14.15 [unord.hash].
-template <typename T, typename = size_t>
-struct optional_hash_base {
-  optional_hash_base() = delete;
-  optional_hash_base(const optional_hash_base&) = delete;
-  optional_hash_base(optional_hash_base&&) = delete;
-  optional_hash_base& operator=(const optional_hash_base&) = delete;
-  optional_hash_base& operator=(optional_hash_base&&) = delete;
-};
-
-template <typename T>
-struct optional_hash_base<T, decltype(std::hash<absl::remove_const_t<T> >()(
-                                 std::declval<absl::remove_const_t<T> >()))> {
-  using argument_type = absl::optional<T>;
-  using result_type = size_t;
-  size_t operator()(const absl::optional<T>& opt) const {
-    absl::type_traits_internal::AssertHashEnabled<absl::remove_const_t<T>>();
-    if (opt) {
-      return std::hash<absl::remove_const_t<T> >()(*opt);
-    } else {
-      return static_cast<size_t>(0x297814aaad196e6dULL);
-    }
-  }
-};
-
-}  // namespace optional_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TYPES_INTERNAL_OPTIONAL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/types/internal/span.h b/projects/llm_framework/include/abseil-cpp/absl/types/internal/span.h
deleted file mode 100644
index ab89ba3c..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/types/internal/span.h
+++ /dev/null
@@ -1,139 +0,0 @@
-//
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-#ifndef ABSL_TYPES_INTERNAL_SPAN_H_
-#define ABSL_TYPES_INTERNAL_SPAN_H_
-
-#include <algorithm>
-#include <cstddef>
-#include <string>
-#include <type_traits>
-
-#include "absl/algorithm/algorithm.h"
-#include "absl/base/internal/throw_delegate.h"
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-template <typename T>
-class Span;
-
-namespace span_internal {
-// Wrappers for access to container data pointers.
-template <typename C>
-constexpr auto GetDataImpl(C& c, char) noexcept  // NOLINT(runtime/references)
-    -> decltype(c.data()) {
-  return c.data();
-}
-
-// Before C++17, std::string::data returns a const char* in all cases.
-inline char* GetDataImpl(std::string& s,  // NOLINT(runtime/references)
-                         int) noexcept {
-  return &s[0];
-}
-
-template <typename C>
-constexpr auto GetData(C& c) noexcept  // NOLINT(runtime/references)
-    -> decltype(GetDataImpl(c, 0)) {
-  return GetDataImpl(c, 0);
-}
-
-// Detection idioms for size() and data().
-template <typename C>
-using HasSize =
-    std::is_integral<absl::decay_t<decltype(std::declval<C&>().size())>>;
-
-// We want to enable conversion from vector<T*> to Span<const T* const> but
-// disable conversion from vector<Derived> to Span<Base>. Here we use
-// the fact that U** is convertible to Q* const* if and only if Q is the same
-// type or a more cv-qualified version of U.  We also decay the result type of
-// data() to avoid problems with classes which have a member function data()
-// which returns a reference.
-template <typename T, typename C>
-using HasData =
-    std::is_convertible<absl::decay_t<decltype(GetData(std::declval<C&>()))>*,
-                        T* const*>;
-
-// Extracts value type from a Container
-template <typename C>
-struct ElementType {
-  using type = typename absl::remove_reference_t<C>::value_type;
-};
-
-template <typename T, size_t N>
-struct ElementType<T (&)[N]> {
-  using type = T;
-};
-
-template <typename C>
-using ElementT = typename ElementType<C>::type;
-
-template <typename T>
-using EnableIfMutable =
-    typename std::enable_if<!std::is_const<T>::value, int>::type;
-
-template <template <typename> class SpanT, typename T>
-bool EqualImpl(SpanT<T> a, SpanT<T> b) {
-  static_assert(std::is_const<T>::value, "");
-  return std::equal(a.begin(), a.end(), b.begin(), b.end());
-}
-
-template <template <typename> class SpanT, typename T>
-bool LessThanImpl(SpanT<T> a, SpanT<T> b) {
-  // We can't use value_type since that is remove_cv_t<T>, so we go the long way
-  // around.
-  static_assert(std::is_const<T>::value, "");
-  return std::lexicographical_compare(a.begin(), a.end(), b.begin(), b.end());
-}
-
-template <typename From, typename To>
-using EnableIfConvertibleTo =
-    typename std::enable_if<std::is_convertible<From, To>::value>::type;
-
-// IsView is true for types where the return type of .data() is the same for
-// mutable and const instances. This isn't foolproof, but it's only used to
-// enable a compiler warning.
-template <typename T, typename = void, typename = void>
-struct IsView {
-  static constexpr bool value = false;
-};
-
-template <typename T>
-struct IsView<
-    T, absl::void_t<decltype(span_internal::GetData(std::declval<const T&>()))>,
-    absl::void_t<decltype(span_internal::GetData(std::declval<T&>()))>> {
- private:
-  using Container = std::remove_const_t<T>;
-  using ConstData =
-      decltype(span_internal::GetData(std::declval<const Container&>()));
-  using MutData = decltype(span_internal::GetData(std::declval<Container&>()));
- public:
-  static constexpr bool value = std::is_same<ConstData, MutData>::value;
-};
-
-// These enablers result in 'int' so they can be used as typenames or defaults
-// in template parameters lists.
-template <typename T>
-using EnableIfIsView = std::enable_if_t<IsView<T>::value, int>;
-
-template <typename T>
-using EnableIfNotIsView = std::enable_if_t<!IsView<T>::value, int>;
-
-}  // namespace span_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TYPES_INTERNAL_SPAN_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/types/internal/variant.h b/projects/llm_framework/include/abseil-cpp/absl/types/internal/variant.h
deleted file mode 100644
index 40f57c40..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/types/internal/variant.h
+++ /dev/null
@@ -1,1635 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Implementation details of absl/types/variant.h, pulled into a
-// separate file to avoid cluttering the top of the API header with
-// implementation details.
-
-#ifndef ABSL_TYPES_INTERNAL_VARIANT_H_
-#define ABSL_TYPES_INTERNAL_VARIANT_H_
-
-#include <cassert>
-#include <cstddef>
-#include <cstdlib>
-#include <memory>
-#include <stdexcept>
-#include <tuple>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/config.h"
-#include "absl/base/internal/identity.h"
-#include "absl/base/internal/inline_variable.h"
-#include "absl/base/internal/invoke.h"
-#include "absl/base/macros.h"
-#include "absl/base/optimization.h"
-#include "absl/meta/type_traits.h"
-#include "absl/types/bad_variant_access.h"
-#include "absl/utility/utility.h"
-
-#if !defined(ABSL_USES_STD_VARIANT)
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-template <class... Types>
-class variant;
-
-ABSL_INTERNAL_INLINE_CONSTEXPR(size_t, variant_npos, static_cast<size_t>(-1));
-
-template <class T>
-struct variant_size;
-
-template <std::size_t I, class T>
-struct variant_alternative;
-
-namespace variant_internal {
-
-// NOTE: See specializations below for details.
-template <std::size_t I, class T>
-struct VariantAlternativeSfinae {};
-
-// Requires: I < variant_size_v<T>.
-//
-// Value: The Ith type of Types...
-template <std::size_t I, class T0, class... Tn>
-struct VariantAlternativeSfinae<I, variant<T0, Tn...>>
-    : VariantAlternativeSfinae<I - 1, variant<Tn...>> {};
-
-// Value: T0
-template <class T0, class... Ts>
-struct VariantAlternativeSfinae<0, variant<T0, Ts...>> {
-  using type = T0;
-};
-
-template <std::size_t I, class T>
-using VariantAlternativeSfinaeT = typename VariantAlternativeSfinae<I, T>::type;
-
-// NOTE: Requires T to be a reference type.
-template <class T, class U>
-struct GiveQualsTo;
-
-template <class T, class U>
-struct GiveQualsTo<T&, U> {
-  using type = U&;
-};
-
-template <class T, class U>
-struct GiveQualsTo<T&&, U> {
-  using type = U&&;
-};
-
-template <class T, class U>
-struct GiveQualsTo<const T&, U> {
-  using type = const U&;
-};
-
-template <class T, class U>
-struct GiveQualsTo<const T&&, U> {
-  using type = const U&&;
-};
-
-template <class T, class U>
-struct GiveQualsTo<volatile T&, U> {
-  using type = volatile U&;
-};
-
-template <class T, class U>
-struct GiveQualsTo<volatile T&&, U> {
-  using type = volatile U&&;
-};
-
-template <class T, class U>
-struct GiveQualsTo<volatile const T&, U> {
-  using type = volatile const U&;
-};
-
-template <class T, class U>
-struct GiveQualsTo<volatile const T&&, U> {
-  using type = volatile const U&&;
-};
-
-template <class T, class U>
-using GiveQualsToT = typename GiveQualsTo<T, U>::type;
-
-// Convenience alias, since size_t integral_constant is used a lot in this file.
-template <std::size_t I>
-using SizeT = std::integral_constant<std::size_t, I>;
-
-using NPos = SizeT<variant_npos>;
-
-template <class Variant, class T, class = void>
-struct IndexOfConstructedType {};
-
-template <std::size_t I, class Variant>
-struct VariantAccessResultImpl;
-
-template <std::size_t I, template <class...> class Variantemplate, class... T>
-struct VariantAccessResultImpl<I, Variantemplate<T...>&> {
-  using type = typename absl::variant_alternative<I, variant<T...>>::type&;
-};
-
-template <std::size_t I, template <class...> class Variantemplate, class... T>
-struct VariantAccessResultImpl<I, const Variantemplate<T...>&> {
-  using type =
-      const typename absl::variant_alternative<I, variant<T...>>::type&;
-};
-
-template <std::size_t I, template <class...> class Variantemplate, class... T>
-struct VariantAccessResultImpl<I, Variantemplate<T...>&&> {
-  using type = typename absl::variant_alternative<I, variant<T...>>::type&&;
-};
-
-template <std::size_t I, template <class...> class Variantemplate, class... T>
-struct VariantAccessResultImpl<I, const Variantemplate<T...>&&> {
-  using type =
-      const typename absl::variant_alternative<I, variant<T...>>::type&&;
-};
-
-template <std::size_t I, class Variant>
-using VariantAccessResult =
-    typename VariantAccessResultImpl<I, Variant&&>::type;
-
-// NOTE: This is used instead of std::array to reduce instantiation overhead.
-template <class T, std::size_t Size>
-struct SimpleArray {
-  static_assert(Size != 0, "");
-  T value[Size];
-};
-
-template <class T>
-struct AccessedType {
-  using type = T;
-};
-
-template <class T>
-using AccessedTypeT = typename AccessedType<T>::type;
-
-template <class T, std::size_t Size>
-struct AccessedType<SimpleArray<T, Size>> {
-  using type = AccessedTypeT<T>;
-};
-
-template <class T>
-constexpr T AccessSimpleArray(const T& value) {
-  return value;
-}
-
-template <class T, std::size_t Size, class... SizeT>
-constexpr AccessedTypeT<T> AccessSimpleArray(const SimpleArray<T, Size>& table,
-                                             std::size_t head_index,
-                                             SizeT... tail_indices) {
-  return AccessSimpleArray(table.value[head_index], tail_indices...);
-}
-
-// Note: Intentionally is an alias.
-template <class T>
-using AlwaysZero = SizeT<0>;
-
-template <class Op, class... Vs>
-struct VisitIndicesResultImpl {
-  using type = absl::result_of_t<Op(AlwaysZero<Vs>...)>;
-};
-
-template <class Op, class... Vs>
-using VisitIndicesResultT = typename VisitIndicesResultImpl<Op, Vs...>::type;
-
-template <class ReturnType, class FunctionObject, class EndIndices,
-          class BoundIndices>
-struct MakeVisitationMatrix;
-
-template <class ReturnType, class FunctionObject, std::size_t... Indices>
-constexpr ReturnType call_with_indices(FunctionObject&& function) {
-  static_assert(
-      std::is_same<ReturnType, decltype(std::declval<FunctionObject>()(
-                                   SizeT<Indices>()...))>::value,
-      "Not all visitation overloads have the same return type.");
-  return std::forward<FunctionObject>(function)(SizeT<Indices>()...);
-}
-
-template <class ReturnType, class FunctionObject, std::size_t... BoundIndices>
-struct MakeVisitationMatrix<ReturnType, FunctionObject, index_sequence<>,
-                            index_sequence<BoundIndices...>> {
-  using ResultType = ReturnType (*)(FunctionObject&&);
-  static constexpr ResultType Run() {
-    return &call_with_indices<ReturnType, FunctionObject,
-                              (BoundIndices - 1)...>;
-  }
-};
-
-template <typename Is, std::size_t J>
-struct AppendToIndexSequence;
-
-template <typename Is, std::size_t J>
-using AppendToIndexSequenceT = typename AppendToIndexSequence<Is, J>::type;
-
-template <std::size_t... Is, std::size_t J>
-struct AppendToIndexSequence<index_sequence<Is...>, J> {
-  using type = index_sequence<Is..., J>;
-};
-
-template <class ReturnType, class FunctionObject, class EndIndices,
-          class CurrIndices, class BoundIndices>
-struct MakeVisitationMatrixImpl;
-
-template <class ReturnType, class FunctionObject, class EndIndices,
-          std::size_t... CurrIndices, class BoundIndices>
-struct MakeVisitationMatrixImpl<ReturnType, FunctionObject, EndIndices,
-                                index_sequence<CurrIndices...>, BoundIndices> {
-  using ResultType = SimpleArray<
-      typename MakeVisitationMatrix<ReturnType, FunctionObject, EndIndices,
-                                    index_sequence<>>::ResultType,
-      sizeof...(CurrIndices)>;
-
-  static constexpr ResultType Run() {
-    return {{MakeVisitationMatrix<
-        ReturnType, FunctionObject, EndIndices,
-        AppendToIndexSequenceT<BoundIndices, CurrIndices>>::Run()...}};
-  }
-};
-
-template <class ReturnType, class FunctionObject, std::size_t HeadEndIndex,
-          std::size_t... TailEndIndices, std::size_t... BoundIndices>
-struct MakeVisitationMatrix<ReturnType, FunctionObject,
-                            index_sequence<HeadEndIndex, TailEndIndices...>,
-                            index_sequence<BoundIndices...>>
-    : MakeVisitationMatrixImpl<ReturnType, FunctionObject,
-                               index_sequence<TailEndIndices...>,
-                               absl::make_index_sequence<HeadEndIndex>,
-                               index_sequence<BoundIndices...>> {};
-
-struct UnreachableSwitchCase {
-  template <class Op>
-  [[noreturn]] static VisitIndicesResultT<Op, std::size_t> Run(
-      Op&& /*ignored*/) {
-#if ABSL_HAVE_BUILTIN(__builtin_unreachable) || \
-    (defined(__GNUC__) && !defined(__clang__))
-    __builtin_unreachable();
-#elif defined(_MSC_VER)
-    __assume(false);
-#else
-    // Try to use assert of false being identified as an unreachable intrinsic.
-    // NOTE: We use assert directly to increase chances of exploiting an assume
-    //       intrinsic.
-    assert(false);  // NOLINT
-
-    // Hack to silence potential no return warning -- cause an infinite loop.
-    return Run(std::forward<Op>(op));
-#endif  // Checks for __builtin_unreachable
-  }
-};
-
-template <class Op, std::size_t I>
-struct ReachableSwitchCase {
-  static VisitIndicesResultT<Op, std::size_t> Run(Op&& op) {
-    return absl::base_internal::invoke(std::forward<Op>(op), SizeT<I>());
-  }
-};
-
-// The number 33 is just a guess at a reasonable maximum to our switch. It is
-// not based on any analysis. The reason it is a power of 2 plus 1 instead of a
-// power of 2 is because the number was picked to correspond to a power of 2
-// amount of "normal" alternatives, plus one for the possibility of the user
-// providing "monostate" in addition to the more natural alternatives.
-ABSL_INTERNAL_INLINE_CONSTEXPR(std::size_t, MaxUnrolledVisitCases, 33);
-
-// Note: The default-definition is for unreachable cases.
-template <bool IsReachable>
-struct PickCaseImpl {
-  template <class Op, std::size_t I>
-  using Apply = UnreachableSwitchCase;
-};
-
-template <>
-struct PickCaseImpl</*IsReachable =*/true> {
-  template <class Op, std::size_t I>
-  using Apply = ReachableSwitchCase<Op, I>;
-};
-
-// Note: This form of dance with template aliases is to make sure that we
-//       instantiate a number of templates proportional to the number of variant
-//       alternatives rather than a number of templates proportional to our
-//       maximum unrolled amount of visitation cases (aliases are effectively
-//       "free" whereas other template instantiations are costly).
-template <class Op, std::size_t I, std::size_t EndIndex>
-using PickCase = typename PickCaseImpl<(I < EndIndex)>::template Apply<Op, I>;
-
-template <class ReturnType>
-[[noreturn]] ReturnType TypedThrowBadVariantAccess() {
-  absl::variant_internal::ThrowBadVariantAccess();
-}
-
-// Given N variant sizes, determine the number of cases there would need to be
-// in a single switch-statement that would cover every possibility in the
-// corresponding N-ary visit operation.
-template <std::size_t... NumAlternatives>
-struct NumCasesOfSwitch;
-
-template <std::size_t HeadNumAlternatives, std::size_t... TailNumAlternatives>
-struct NumCasesOfSwitch<HeadNumAlternatives, TailNumAlternatives...> {
-  static constexpr std::size_t value =
-      (HeadNumAlternatives + 1) *
-      NumCasesOfSwitch<TailNumAlternatives...>::value;
-};
-
-template <>
-struct NumCasesOfSwitch<> {
-  static constexpr std::size_t value = 1;
-};
-
-// A switch statement optimizes better than the table of function pointers.
-template <std::size_t EndIndex>
-struct VisitIndicesSwitch {
-  static_assert(EndIndex <= MaxUnrolledVisitCases,
-                "Maximum unrolled switch size exceeded.");
-
-  template <class Op>
-  static VisitIndicesResultT<Op, std::size_t> Run(Op&& op, std::size_t i) {
-    switch (i) {
-      case 0:
-        return PickCase<Op, 0, EndIndex>::Run(std::forward<Op>(op));
-      case 1:
-        return PickCase<Op, 1, EndIndex>::Run(std::forward<Op>(op));
-      case 2:
-        return PickCase<Op, 2, EndIndex>::Run(std::forward<Op>(op));
-      case 3:
-        return PickCase<Op, 3, EndIndex>::Run(std::forward<Op>(op));
-      case 4:
-        return PickCase<Op, 4, EndIndex>::Run(std::forward<Op>(op));
-      case 5:
-        return PickCase<Op, 5, EndIndex>::Run(std::forward<Op>(op));
-      case 6:
-        return PickCase<Op, 6, EndIndex>::Run(std::forward<Op>(op));
-      case 7:
-        return PickCase<Op, 7, EndIndex>::Run(std::forward<Op>(op));
-      case 8:
-        return PickCase<Op, 8, EndIndex>::Run(std::forward<Op>(op));
-      case 9:
-        return PickCase<Op, 9, EndIndex>::Run(std::forward<Op>(op));
-      case 10:
-        return PickCase<Op, 10, EndIndex>::Run(std::forward<Op>(op));
-      case 11:
-        return PickCase<Op, 11, EndIndex>::Run(std::forward<Op>(op));
-      case 12:
-        return PickCase<Op, 12, EndIndex>::Run(std::forward<Op>(op));
-      case 13:
-        return PickCase<Op, 13, EndIndex>::Run(std::forward<Op>(op));
-      case 14:
-        return PickCase<Op, 14, EndIndex>::Run(std::forward<Op>(op));
-      case 15:
-        return PickCase<Op, 15, EndIndex>::Run(std::forward<Op>(op));
-      case 16:
-        return PickCase<Op, 16, EndIndex>::Run(std::forward<Op>(op));
-      case 17:
-        return PickCase<Op, 17, EndIndex>::Run(std::forward<Op>(op));
-      case 18:
-        return PickCase<Op, 18, EndIndex>::Run(std::forward<Op>(op));
-      case 19:
-        return PickCase<Op, 19, EndIndex>::Run(std::forward<Op>(op));
-      case 20:
-        return PickCase<Op, 20, EndIndex>::Run(std::forward<Op>(op));
-      case 21:
-        return PickCase<Op, 21, EndIndex>::Run(std::forward<Op>(op));
-      case 22:
-        return PickCase<Op, 22, EndIndex>::Run(std::forward<Op>(op));
-      case 23:
-        return PickCase<Op, 23, EndIndex>::Run(std::forward<Op>(op));
-      case 24:
-        return PickCase<Op, 24, EndIndex>::Run(std::forward<Op>(op));
-      case 25:
-        return PickCase<Op, 25, EndIndex>::Run(std::forward<Op>(op));
-      case 26:
-        return PickCase<Op, 26, EndIndex>::Run(std::forward<Op>(op));
-      case 27:
-        return PickCase<Op, 27, EndIndex>::Run(std::forward<Op>(op));
-      case 28:
-        return PickCase<Op, 28, EndIndex>::Run(std::forward<Op>(op));
-      case 29:
-        return PickCase<Op, 29, EndIndex>::Run(std::forward<Op>(op));
-      case 30:
-        return PickCase<Op, 30, EndIndex>::Run(std::forward<Op>(op));
-      case 31:
-        return PickCase<Op, 31, EndIndex>::Run(std::forward<Op>(op));
-      case 32:
-        return PickCase<Op, 32, EndIndex>::Run(std::forward<Op>(op));
-      default:
-        ABSL_ASSERT(i == variant_npos);
-        return absl::base_internal::invoke(std::forward<Op>(op), NPos());
-    }
-  }
-};
-
-template <std::size_t... EndIndices>
-struct VisitIndicesFallback {
-  template <class Op, class... SizeT>
-  static VisitIndicesResultT<Op, SizeT...> Run(Op&& op, SizeT... indices) {
-    return AccessSimpleArray(
-        MakeVisitationMatrix<VisitIndicesResultT<Op, SizeT...>, Op,
-                             index_sequence<(EndIndices + 1)...>,
-                             index_sequence<>>::Run(),
-        (indices + 1)...)(std::forward<Op>(op));
-  }
-};
-
-// Take an N-dimensional series of indices and convert them into a single index
-// without loss of information. The purpose of this is to be able to convert an
-// N-ary visit operation into a single switch statement.
-template <std::size_t...>
-struct FlattenIndices;
-
-template <std::size_t HeadSize, std::size_t... TailSize>
-struct FlattenIndices<HeadSize, TailSize...> {
-  template <class... SizeType>
-  static constexpr std::size_t Run(std::size_t head, SizeType... tail) {
-    return head + HeadSize * FlattenIndices<TailSize...>::Run(tail...);
-  }
-};
-
-template <>
-struct FlattenIndices<> {
-  static constexpr std::size_t Run() { return 0; }
-};
-
-// Take a single "flattened" index (flattened by FlattenIndices) and determine
-// the value of the index of one of the logically represented dimensions.
-template <std::size_t I, std::size_t IndexToGet, std::size_t HeadSize,
-          std::size_t... TailSize>
-struct UnflattenIndex {
-  static constexpr std::size_t value =
-      UnflattenIndex<I / HeadSize, IndexToGet - 1, TailSize...>::value;
-};
-
-template <std::size_t I, std::size_t HeadSize, std::size_t... TailSize>
-struct UnflattenIndex<I, 0, HeadSize, TailSize...> {
-  static constexpr std::size_t value = (I % HeadSize);
-};
-
-// The backend for converting an N-ary visit operation into a unary visit.
-template <class IndexSequence, std::size_t... EndIndices>
-struct VisitIndicesVariadicImpl;
-
-template <std::size_t... N, std::size_t... EndIndices>
-struct VisitIndicesVariadicImpl<absl::index_sequence<N...>, EndIndices...> {
-  // A type that can take an N-ary function object and converts it to a unary
-  // function object that takes a single, flattened index, and "unflattens" it
-  // into its individual dimensions when forwarding to the wrapped object.
-  template <class Op>
-  struct FlattenedOp {
-    template <std::size_t I>
-    VisitIndicesResultT<Op, decltype(EndIndices)...> operator()(
-        SizeT<I> /*index*/) && {
-      return base_internal::invoke(
-          std::forward<Op>(op),
-          SizeT<UnflattenIndex<I, N, (EndIndices + 1)...>::value -
-                std::size_t{1}>()...);
-    }
-
-    Op&& op;
-  };
-
-  template <class Op, class... SizeType>
-  static VisitIndicesResultT<Op, decltype(EndIndices)...> Run(Op&& op,
-                                                              SizeType... i) {
-    return VisitIndicesSwitch<NumCasesOfSwitch<EndIndices...>::value>::Run(
-        FlattenedOp<Op>{std::forward<Op>(op)},
-        FlattenIndices<(EndIndices + std::size_t{1})...>::Run(
-            (i + std::size_t{1})...));
-  }
-};
-
-template <std::size_t... EndIndices>
-struct VisitIndicesVariadic
-    : VisitIndicesVariadicImpl<absl::make_index_sequence<sizeof...(EndIndices)>,
-                               EndIndices...> {};
-
-// This implementation will flatten N-ary visit operations into a single switch
-// statement when the number of cases would be less than our maximum specified
-// switch-statement size.
-// TODO(calabrese)
-//   Based on benchmarks, determine whether the function table approach actually
-//   does optimize better than a chain of switch statements and possibly update
-//   the implementation accordingly. Also consider increasing the maximum switch
-//   size.
-template <std::size_t... EndIndices>
-struct VisitIndices
-    : absl::conditional_t<(NumCasesOfSwitch<EndIndices...>::value <=
-                           MaxUnrolledVisitCases),
-                          VisitIndicesVariadic<EndIndices...>,
-                          VisitIndicesFallback<EndIndices...>> {};
-
-template <std::size_t EndIndex>
-struct VisitIndices<EndIndex>
-    : absl::conditional_t<(EndIndex <= MaxUnrolledVisitCases),
-                          VisitIndicesSwitch<EndIndex>,
-                          VisitIndicesFallback<EndIndex>> {};
-
-// Suppress bogus warning on MSVC: MSVC complains that the `reinterpret_cast`
-// below is returning the address of a temporary or local object.
-#ifdef _MSC_VER
-#pragma warning(push)
-#pragma warning(disable : 4172)
-#endif  // _MSC_VER
-
-// TODO(calabrese) std::launder
-// TODO(calabrese) constexpr
-// NOTE: DO NOT REMOVE the `inline` keyword as it is necessary to work around a
-// MSVC bug. See https://github.com/abseil/abseil-cpp/issues/129 for details.
-template <class Self, std::size_t I>
-inline VariantAccessResult<I, Self> AccessUnion(Self&& self, SizeT<I> /*i*/) {
-  return reinterpret_cast<VariantAccessResult<I, Self>>(self);
-}
-
-#ifdef _MSC_VER
-#pragma warning(pop)
-#endif  // _MSC_VER
-
-template <class T>
-void DeducedDestroy(T& self) {  // NOLINT
-  self.~T();
-}
-
-// NOTE: This type exists as a single entity for variant and its bases to
-// befriend. It contains helper functionality that manipulates the state of the
-// variant, such as the implementation of things like assignment and emplace
-// operations.
-struct VariantCoreAccess {
-  template <class VariantType>
-  static typename VariantType::Variant& Derived(VariantType& self) {  // NOLINT
-    return static_cast<typename VariantType::Variant&>(self);
-  }
-
-  template <class VariantType>
-  static const typename VariantType::Variant& Derived(
-      const VariantType& self) {  // NOLINT
-    return static_cast<const typename VariantType::Variant&>(self);
-  }
-
-  template <class VariantType>
-  static void Destroy(VariantType& self) {  // NOLINT
-    Derived(self).destroy();
-    self.index_ = absl::variant_npos;
-  }
-
-  template <class Variant>
-  static void SetIndex(Variant& self, std::size_t i) {  // NOLINT
-    self.index_ = i;
-  }
-
-  template <class Variant>
-  static void InitFrom(Variant& self, Variant&& other) {  // NOLINT
-    VisitIndices<absl::variant_size<Variant>::value>::Run(
-        InitFromVisitor<Variant, Variant&&>{&self,
-                                            std::forward<Variant>(other)},
-        other.index());
-    self.index_ = other.index();
-  }
-
-  // Access a variant alternative, assuming the index is correct.
-  template <std::size_t I, class Variant>
-  static VariantAccessResult<I, Variant> Access(Variant&& self) {
-    // This cast instead of invocation of AccessUnion with an rvalue is a
-    // workaround for msvc. Without this there is a runtime failure when dealing
-    // with rvalues.
-    // TODO(calabrese) Reduce test case and find a simpler workaround.
-    return static_cast<VariantAccessResult<I, Variant>>(
-        variant_internal::AccessUnion(self.state_, SizeT<I>()));
-  }
-
-  // Access a variant alternative, throwing if the index is incorrect.
-  template <std::size_t I, class Variant>
-  static VariantAccessResult<I, Variant> CheckedAccess(Variant&& self) {
-    if (ABSL_PREDICT_FALSE(self.index_ != I)) {
-      TypedThrowBadVariantAccess<VariantAccessResult<I, Variant>>();
-    }
-
-    return Access<I>(std::forward<Variant>(self));
-  }
-
-  // The implementation of the move-assignment operation for a variant.
-  template <class VType>
-  struct MoveAssignVisitor {
-    using DerivedType = typename VType::Variant;
-    template <std::size_t NewIndex>
-    void operator()(SizeT<NewIndex> /*new_i*/) const {
-      if (left->index_ == NewIndex) {
-        Access<NewIndex>(*left) = std::move(Access<NewIndex>(*right));
-      } else {
-        Derived(*left).template emplace<NewIndex>(
-            std::move(Access<NewIndex>(*right)));
-      }
-    }
-
-    void operator()(SizeT<absl::variant_npos> /*new_i*/) const {
-      Destroy(*left);
-    }
-
-    VType* left;
-    VType* right;
-  };
-
-  template <class VType>
-  static MoveAssignVisitor<VType> MakeMoveAssignVisitor(VType* left,
-                                                        VType* other) {
-    return {left, other};
-  }
-
-  // The implementation of the assignment operation for a variant.
-  template <class VType>
-  struct CopyAssignVisitor {
-    using DerivedType = typename VType::Variant;
-    template <std::size_t NewIndex>
-    void operator()(SizeT<NewIndex> /*new_i*/) const {
-      using New =
-          typename absl::variant_alternative<NewIndex, DerivedType>::type;
-
-      if (left->index_ == NewIndex) {
-        Access<NewIndex>(*left) = Access<NewIndex>(*right);
-      } else if (std::is_nothrow_copy_constructible<New>::value ||
-                 !std::is_nothrow_move_constructible<New>::value) {
-        Derived(*left).template emplace<NewIndex>(Access<NewIndex>(*right));
-      } else {
-        Derived(*left) = DerivedType(Derived(*right));
-      }
-    }
-
-    void operator()(SizeT<absl::variant_npos> /*new_i*/) const {
-      Destroy(*left);
-    }
-
-    VType* left;
-    const VType* right;
-  };
-
-  template <class VType>
-  static CopyAssignVisitor<VType> MakeCopyAssignVisitor(VType* left,
-                                                        const VType& other) {
-    return {left, &other};
-  }
-
-  // The implementation of conversion-assignment operations for variant.
-  template <class Left, class QualifiedNew>
-  struct ConversionAssignVisitor {
-    using NewIndex =
-        variant_internal::IndexOfConstructedType<Left, QualifiedNew>;
-
-    void operator()(SizeT<NewIndex::value> /*old_i*/
-    ) const {
-      Access<NewIndex::value>(*left) = std::forward<QualifiedNew>(other);
-    }
-
-    template <std::size_t OldIndex>
-    void operator()(SizeT<OldIndex> /*old_i*/
-    ) const {
-      using New =
-          typename absl::variant_alternative<NewIndex::value, Left>::type;
-      if (std::is_nothrow_constructible<New, QualifiedNew>::value ||
-          !std::is_nothrow_move_constructible<New>::value) {
-        left->template emplace<NewIndex::value>(
-            std::forward<QualifiedNew>(other));
-      } else {
-        // the standard says "equivalent to
-        // operator=(variant(std::forward<T>(t)))", but we use `emplace` here
-        // because the variant's move assignment operator could be deleted.
-        left->template emplace<NewIndex::value>(
-            New(std::forward<QualifiedNew>(other)));
-      }
-    }
-
-    Left* left;
-    QualifiedNew&& other;
-  };
-
-  template <class Left, class QualifiedNew>
-  static ConversionAssignVisitor<Left, QualifiedNew>
-  MakeConversionAssignVisitor(Left* left, QualifiedNew&& qual) {
-    return {left, std::forward<QualifiedNew>(qual)};
-  }
-
-  // Backend for operations for `emplace()` which destructs `*self` then
-  // construct a new alternative with `Args...`.
-  template <std::size_t NewIndex, class Self, class... Args>
-  static typename absl::variant_alternative<NewIndex, Self>::type& Replace(
-      Self* self, Args&&... args) {
-    Destroy(*self);
-    using New = typename absl::variant_alternative<NewIndex, Self>::type;
-    New* const result = ::new (static_cast<void*>(&self->state_))
-        New(std::forward<Args>(args)...);
-    self->index_ = NewIndex;
-    return *result;
-  }
-
-  template <class LeftVariant, class QualifiedRightVariant>
-  struct InitFromVisitor {
-    template <std::size_t NewIndex>
-    void operator()(SizeT<NewIndex> /*new_i*/) const {
-      using Alternative =
-          typename variant_alternative<NewIndex, LeftVariant>::type;
-      ::new (static_cast<void*>(&left->state_)) Alternative(
-          Access<NewIndex>(std::forward<QualifiedRightVariant>(right)));
-    }
-
-    void operator()(SizeT<absl::variant_npos> /*new_i*/) const {
-      // This space intentionally left blank.
-    }
-    LeftVariant* left;
-    QualifiedRightVariant&& right;
-  };
-};
-
-template <class Expected, class... T>
-struct IndexOfImpl;
-
-template <class Expected>
-struct IndexOfImpl<Expected> {
-  using IndexFromEnd = SizeT<0>;
-  using MatchedIndexFromEnd = IndexFromEnd;
-  using MultipleMatches = std::false_type;
-};
-
-template <class Expected, class Head, class... Tail>
-struct IndexOfImpl<Expected, Head, Tail...> : IndexOfImpl<Expected, Tail...> {
-  using IndexFromEnd =
-      SizeT<IndexOfImpl<Expected, Tail...>::IndexFromEnd::value + 1>;
-};
-
-template <class Expected, class... Tail>
-struct IndexOfImpl<Expected, Expected, Tail...>
-    : IndexOfImpl<Expected, Tail...> {
-  using IndexFromEnd =
-      SizeT<IndexOfImpl<Expected, Tail...>::IndexFromEnd::value + 1>;
-  using MatchedIndexFromEnd = IndexFromEnd;
-  using MultipleMatches = std::integral_constant<
-      bool, IndexOfImpl<Expected, Tail...>::MatchedIndexFromEnd::value != 0>;
-};
-
-template <class Expected, class... Types>
-struct IndexOfMeta {
-  using Results = IndexOfImpl<Expected, Types...>;
-  static_assert(!Results::MultipleMatches::value,
-                "Attempted to access a variant by specifying a type that "
-                "matches more than one alternative.");
-  static_assert(Results::MatchedIndexFromEnd::value != 0,
-                "Attempted to access a variant by specifying a type that does "
-                "not match any alternative.");
-  using type = SizeT<sizeof...(Types) - Results::MatchedIndexFromEnd::value>;
-};
-
-template <class Expected, class... Types>
-using IndexOf = typename IndexOfMeta<Expected, Types...>::type;
-
-template <class Variant, class T, std::size_t CurrIndex>
-struct UnambiguousIndexOfImpl;
-
-// Terminating case encountered once we've checked all of the alternatives
-template <class T, std::size_t CurrIndex>
-struct UnambiguousIndexOfImpl<variant<>, T, CurrIndex> : SizeT<CurrIndex> {};
-
-// Case where T is not Head
-template <class Head, class... Tail, class T, std::size_t CurrIndex>
-struct UnambiguousIndexOfImpl<variant<Head, Tail...>, T, CurrIndex>
-    : UnambiguousIndexOfImpl<variant<Tail...>, T, CurrIndex + 1>::type {};
-
-// Case where T is Head
-template <class Head, class... Tail, std::size_t CurrIndex>
-struct UnambiguousIndexOfImpl<variant<Head, Tail...>, Head, CurrIndex>
-    : SizeT<UnambiguousIndexOfImpl<variant<Tail...>, Head, 0>::value ==
-                    sizeof...(Tail)
-                ? CurrIndex
-                : CurrIndex + sizeof...(Tail) + 1> {};
-
-template <class Variant, class T>
-struct UnambiguousIndexOf;
-
-struct NoMatch {
-  struct type {};
-};
-
-template <class... Alts, class T>
-struct UnambiguousIndexOf<variant<Alts...>, T>
-    : std::conditional<UnambiguousIndexOfImpl<variant<Alts...>, T, 0>::value !=
-                           sizeof...(Alts),
-                       UnambiguousIndexOfImpl<variant<Alts...>, T, 0>,
-                       NoMatch>::type::type {};
-
-template <class T, std::size_t /*Dummy*/>
-using UnambiguousTypeOfImpl = T;
-
-template <class Variant, class T>
-using UnambiguousTypeOfT =
-    UnambiguousTypeOfImpl<T, UnambiguousIndexOf<Variant, T>::value>;
-
-template <class H, class... T>
-class VariantStateBase;
-
-// This is an implementation of the "imaginary function" that is described in
-// [variant.ctor]
-// It is used in order to determine which alternative to construct during
-// initialization from some type T.
-template <class Variant, std::size_t I = 0>
-struct ImaginaryFun;
-
-template <std::size_t I>
-struct ImaginaryFun<variant<>, I> {
-  static void Run() = delete;
-};
-
-template <class H, class... T, std::size_t I>
-struct ImaginaryFun<variant<H, T...>, I> : ImaginaryFun<variant<T...>, I + 1> {
-  using ImaginaryFun<variant<T...>, I + 1>::Run;
-
-  // NOTE: const& and && are used instead of by-value due to lack of guaranteed
-  // move elision of C++17. This may have other minor differences, but tests
-  // pass.
-  static SizeT<I> Run(const H&, SizeT<I>);
-  static SizeT<I> Run(H&&, SizeT<I>);
-};
-
-// The following metafunctions are used in constructor and assignment
-// constraints.
-template <class Self, class T>
-struct IsNeitherSelfNorInPlace : std::true_type {};
-
-template <class Self>
-struct IsNeitherSelfNorInPlace<Self, Self> : std::false_type {};
-
-template <class Self, class T>
-struct IsNeitherSelfNorInPlace<Self, in_place_type_t<T>> : std::false_type {};
-
-template <class Self, std::size_t I>
-struct IsNeitherSelfNorInPlace<Self, in_place_index_t<I>> : std::false_type {};
-
-template <class Variant, class T>
-struct IndexOfConstructedType<
-    Variant, T,
-    void_t<decltype(ImaginaryFun<Variant>::Run(std::declval<T>(), {}))>>
-    : decltype(ImaginaryFun<Variant>::Run(std::declval<T>(), {})) {};
-
-template <std::size_t... Is>
-struct ContainsVariantNPos
-    : absl::negation<std::is_same<  // NOLINT
-          std::integer_sequence<bool, 0 <= Is...>,
-          std::integer_sequence<bool, Is != absl::variant_npos...>>> {};
-
-template <class Op, class... QualifiedVariants>
-using RawVisitResult =
-    absl::result_of_t<Op(VariantAccessResult<0, QualifiedVariants>...)>;
-
-// NOTE: The spec requires that all return-paths yield the same type and is not
-// SFINAE-friendly, so we can deduce the return type by examining the first
-// result. If it's not callable, then we get an error, but are compliant and
-// fast to compile.
-// TODO(calabrese) Possibly rewrite in a way that yields better compile errors
-// at the cost of longer compile-times.
-template <class Op, class... QualifiedVariants>
-struct VisitResultImpl {
-  using type =
-      absl::result_of_t<Op(VariantAccessResult<0, QualifiedVariants>...)>;
-};
-
-// Done in two steps intentionally so that we don't cause substitution to fail.
-template <class Op, class... QualifiedVariants>
-using VisitResult = typename VisitResultImpl<Op, QualifiedVariants...>::type;
-
-template <class Op, class... QualifiedVariants>
-struct PerformVisitation {
-  using ReturnType = VisitResult<Op, QualifiedVariants...>;
-
-  template <std::size_t... Is>
-  constexpr ReturnType operator()(SizeT<Is>... indices) const {
-    return Run(typename ContainsVariantNPos<Is...>::type{},
-               absl::index_sequence_for<QualifiedVariants...>(), indices...);
-  }
-
-  template <std::size_t... TupIs, std::size_t... Is>
-  constexpr ReturnType Run(std::false_type /*has_valueless*/,
-                           index_sequence<TupIs...>, SizeT<Is>...) const {
-    static_assert(
-        std::is_same<ReturnType,
-                     absl::result_of_t<Op(VariantAccessResult<
-                                          Is, QualifiedVariants>...)>>::value,
-        "All visitation overloads must have the same return type.");
-    return absl::base_internal::invoke(
-        std::forward<Op>(op),
-        VariantCoreAccess::Access<Is>(
-            std::forward<QualifiedVariants>(std::get<TupIs>(variant_tup)))...);
-  }
-
-  template <std::size_t... TupIs, std::size_t... Is>
-  [[noreturn]] ReturnType Run(std::true_type /*has_valueless*/,
-                              index_sequence<TupIs...>, SizeT<Is>...) const {
-    absl::variant_internal::ThrowBadVariantAccess();
-  }
-
-  // TODO(calabrese) Avoid using a tuple, which causes lots of instantiations
-  // Attempts using lambda variadic captures fail on current GCC.
-  std::tuple<QualifiedVariants&&...> variant_tup;
-  Op&& op;
-};
-
-template <class... T>
-union Union;
-
-// We want to allow for variant<> to be trivial. For that, we need the default
-// constructor to be trivial, which means we can't define it ourselves.
-// Instead, we use a non-default constructor that takes NoopConstructorTag
-// that doesn't affect the triviality of the types.
-struct NoopConstructorTag {};
-
-template <std::size_t I>
-struct EmplaceTag {};
-
-template <>
-union Union<> {
-  constexpr explicit Union(NoopConstructorTag) noexcept {}
-};
-
-// Suppress bogus warning on MSVC: MSVC complains that Union<T...> has a defined
-// deleted destructor from the `std::is_destructible` check below.
-#ifdef _MSC_VER
-#pragma warning(push)
-#pragma warning(disable : 4624)
-#endif  // _MSC_VER
-
-template <class Head, class... Tail>
-union Union<Head, Tail...> {
-  using TailUnion = Union<Tail...>;
-
-  explicit constexpr Union(NoopConstructorTag /*tag*/) noexcept
-      : tail(NoopConstructorTag()) {}
-
-  template <class... P>
-  explicit constexpr Union(EmplaceTag<0>, P&&... args)
-      : head(std::forward<P>(args)...) {}
-
-  template <std::size_t I, class... P>
-  explicit constexpr Union(EmplaceTag<I>, P&&... args)
-      : tail(EmplaceTag<I - 1>{}, std::forward<P>(args)...) {}
-
-  Head head;
-  TailUnion tail;
-};
-
-#ifdef _MSC_VER
-#pragma warning(pop)
-#endif  // _MSC_VER
-
-// TODO(calabrese) Just contain a Union in this union (certain configs fail).
-template <class... T>
-union DestructibleUnionImpl;
-
-template <>
-union DestructibleUnionImpl<> {
-  constexpr explicit DestructibleUnionImpl(NoopConstructorTag) noexcept {}
-};
-
-template <class Head, class... Tail>
-union DestructibleUnionImpl<Head, Tail...> {
-  using TailUnion = DestructibleUnionImpl<Tail...>;
-
-  explicit constexpr DestructibleUnionImpl(NoopConstructorTag /*tag*/) noexcept
-      : tail(NoopConstructorTag()) {}
-
-  template <class... P>
-  explicit constexpr DestructibleUnionImpl(EmplaceTag<0>, P&&... args)
-      : head(std::forward<P>(args)...) {}
-
-  template <std::size_t I, class... P>
-  explicit constexpr DestructibleUnionImpl(EmplaceTag<I>, P&&... args)
-      : tail(EmplaceTag<I - 1>{}, std::forward<P>(args)...) {}
-
-  ~DestructibleUnionImpl() {}
-
-  Head head;
-  TailUnion tail;
-};
-
-// This union type is destructible even if one or more T are not trivially
-// destructible. In the case that all T are trivially destructible, then so is
-// this resultant type.
-template <class... T>
-using DestructibleUnion =
-    absl::conditional_t<std::is_destructible<Union<T...>>::value, Union<T...>,
-                        DestructibleUnionImpl<T...>>;
-
-// Deepest base, containing the actual union and the discriminator
-template <class H, class... T>
-class VariantStateBase {
- protected:
-  using Variant = variant<H, T...>;
-
-  template <class LazyH = H,
-            class ConstructibleH = absl::enable_if_t<
-                std::is_default_constructible<LazyH>::value, LazyH>>
-  constexpr VariantStateBase() noexcept(
-      std::is_nothrow_default_constructible<ConstructibleH>::value)
-      : state_(EmplaceTag<0>()), index_(0) {}
-
-  template <std::size_t I, class... P>
-  explicit constexpr VariantStateBase(EmplaceTag<I> tag, P&&... args)
-      : state_(tag, std::forward<P>(args)...), index_(I) {}
-
-  explicit constexpr VariantStateBase(NoopConstructorTag)
-      : state_(NoopConstructorTag()), index_(variant_npos) {}
-
-  void destroy() {}  // Does nothing (shadowed in child if non-trivial)
-
-  DestructibleUnion<H, T...> state_;
-  std::size_t index_;
-};
-
-using absl::internal::type_identity;
-
-// OverloadSet::Overload() is a unary function which is overloaded to
-// take any of the element types of the variant, by reference-to-const.
-// The return type of the overload on T is type_identity<T>, so that you
-// can statically determine which overload was called.
-//
-// Overload() is not defined, so it can only be called in unevaluated
-// contexts.
-template <typename... Ts>
-struct OverloadSet;
-
-template <typename T, typename... Ts>
-struct OverloadSet<T, Ts...> : OverloadSet<Ts...> {
-  using Base = OverloadSet<Ts...>;
-  static type_identity<T> Overload(const T&);
-  using Base::Overload;
-};
-
-template <>
-struct OverloadSet<> {
-  // For any case not handled above.
-  static void Overload(...);
-};
-
-template <class T>
-using LessThanResult = decltype(std::declval<T>() < std::declval<T>());
-
-template <class T>
-using GreaterThanResult = decltype(std::declval<T>() > std::declval<T>());
-
-template <class T>
-using LessThanOrEqualResult = decltype(std::declval<T>() <= std::declval<T>());
-
-template <class T>
-using GreaterThanOrEqualResult =
-    decltype(std::declval<T>() >= std::declval<T>());
-
-template <class T>
-using EqualResult = decltype(std::declval<T>() == std::declval<T>());
-
-template <class T>
-using NotEqualResult = decltype(std::declval<T>() != std::declval<T>());
-
-using type_traits_internal::is_detected_convertible;
-
-template <class... T>
-using RequireAllHaveEqualT = absl::enable_if_t<
-    absl::conjunction<is_detected_convertible<bool, EqualResult, T>...>::value,
-    bool>;
-
-template <class... T>
-using RequireAllHaveNotEqualT =
-    absl::enable_if_t<absl::conjunction<is_detected_convertible<
-                          bool, NotEqualResult, T>...>::value,
-                      bool>;
-
-template <class... T>
-using RequireAllHaveLessThanT =
-    absl::enable_if_t<absl::conjunction<is_detected_convertible<
-                          bool, LessThanResult, T>...>::value,
-                      bool>;
-
-template <class... T>
-using RequireAllHaveLessThanOrEqualT =
-    absl::enable_if_t<absl::conjunction<is_detected_convertible<
-                          bool, LessThanOrEqualResult, T>...>::value,
-                      bool>;
-
-template <class... T>
-using RequireAllHaveGreaterThanOrEqualT =
-    absl::enable_if_t<absl::conjunction<is_detected_convertible<
-                          bool, GreaterThanOrEqualResult, T>...>::value,
-                      bool>;
-
-template <class... T>
-using RequireAllHaveGreaterThanT =
-    absl::enable_if_t<absl::conjunction<is_detected_convertible<
-                          bool, GreaterThanResult, T>...>::value,
-                      bool>;
-
-// Helper template containing implementations details of variant that can't go
-// in the private section. For convenience, this takes the variant type as a
-// single template parameter.
-template <typename T>
-struct VariantHelper;
-
-template <typename... Ts>
-struct VariantHelper<variant<Ts...>> {
-  // Type metafunction which returns the element type selected if
-  // OverloadSet::Overload() is well-formed when called with argument type U.
-  template <typename U>
-  using BestMatch = decltype(variant_internal::OverloadSet<Ts...>::Overload(
-      std::declval<U>()));
-
-  // Type metafunction which returns true if OverloadSet::Overload() is
-  // well-formed when called with argument type U.
-  // CanAccept can't be just an alias because there is a MSVC bug on parameter
-  // pack expansion involving decltype.
-  template <typename U>
-  struct CanAccept
-      : std::integral_constant<bool, !std::is_void<BestMatch<U>>::value> {};
-
-  // Type metafunction which returns true if Other is an instantiation of
-  // variant, and variants's converting constructor from Other will be
-  // well-formed. We will use this to remove constructors that would be
-  // ill-formed from the overload set.
-  template <typename Other>
-  struct CanConvertFrom;
-
-  template <typename... Us>
-  struct CanConvertFrom<variant<Us...>>
-      : public absl::conjunction<CanAccept<Us>...> {};
-};
-
-// A type with nontrivial copy ctor and trivial move ctor.
-struct TrivialMoveOnly {
-  TrivialMoveOnly(TrivialMoveOnly&&) = default;
-};
-
-// Trait class to detect whether a type is trivially move constructible.
-// A union's defaulted copy/move constructor is deleted if any variant member's
-// copy/move constructor is nontrivial.
-template <typename T>
-struct IsTriviallyMoveConstructible
-    : std::is_move_constructible<Union<T, TrivialMoveOnly>> {};
-
-// To guarantee triviality of all special-member functions that can be trivial,
-// we use a chain of conditional bases for each one.
-// The order of inheritance of bases from child to base are logically:
-//
-// variant
-// VariantCopyAssignBase
-// VariantMoveAssignBase
-// VariantCopyBase
-// VariantMoveBase
-// VariantStateBaseDestructor
-// VariantStateBase
-//
-// Note that there is a separate branch at each base that is dependent on
-// whether or not that corresponding special-member-function can be trivial in
-// the resultant variant type.
-
-template <class... T>
-class VariantStateBaseDestructorNontrivial;
-
-template <class... T>
-class VariantMoveBaseNontrivial;
-
-template <class... T>
-class VariantCopyBaseNontrivial;
-
-template <class... T>
-class VariantMoveAssignBaseNontrivial;
-
-template <class... T>
-class VariantCopyAssignBaseNontrivial;
-
-// Base that is dependent on whether or not the destructor can be trivial.
-template <class... T>
-using VariantStateBaseDestructor =
-    absl::conditional_t<std::is_destructible<Union<T...>>::value,
-                        VariantStateBase<T...>,
-                        VariantStateBaseDestructorNontrivial<T...>>;
-
-// Base that is dependent on whether or not the move-constructor can be
-// implicitly generated by the compiler (trivial or deleted).
-// Previously we were using `std::is_move_constructible<Union<T...>>` to check
-// whether all Ts have trivial move constructor, but it ran into a GCC bug:
-// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84866
-// So we have to use a different approach (i.e. `HasTrivialMoveConstructor`) to
-// work around the bug.
-template <class... T>
-using VariantMoveBase = absl::conditional_t<
-    absl::disjunction<
-        absl::negation<absl::conjunction<std::is_move_constructible<T>...>>,
-        absl::conjunction<IsTriviallyMoveConstructible<T>...>>::value,
-    VariantStateBaseDestructor<T...>, VariantMoveBaseNontrivial<T...>>;
-
-// Base that is dependent on whether or not the copy-constructor can be trivial.
-template <class... T>
-using VariantCopyBase = absl::conditional_t<
-    absl::disjunction<
-        absl::negation<absl::conjunction<std::is_copy_constructible<T>...>>,
-        std::is_copy_constructible<Union<T...>>>::value,
-    VariantMoveBase<T...>, VariantCopyBaseNontrivial<T...>>;
-
-// Base that is dependent on whether or not the move-assign can be trivial.
-template <class... T>
-using VariantMoveAssignBase = absl::conditional_t<
-    absl::disjunction<
-        absl::conjunction<absl::is_move_assignable<Union<T...>>,
-                          std::is_move_constructible<Union<T...>>,
-                          std::is_destructible<Union<T...>>>,
-        absl::negation<absl::conjunction<std::is_move_constructible<T>...,
-                                         // Note: We're not qualifying this with
-                                         // absl:: because it doesn't compile
-                                         // under MSVC.
-                                         is_move_assignable<T>...>>>::value,
-    VariantCopyBase<T...>, VariantMoveAssignBaseNontrivial<T...>>;
-
-// Base that is dependent on whether or not the copy-assign can be trivial.
-template <class... T>
-using VariantCopyAssignBase = absl::conditional_t<
-    absl::disjunction<
-        absl::conjunction<absl::is_copy_assignable<Union<T...>>,
-                          std::is_copy_constructible<Union<T...>>,
-                          std::is_destructible<Union<T...>>>,
-        absl::negation<absl::conjunction<std::is_copy_constructible<T>...,
-                                         // Note: We're not qualifying this with
-                                         // absl:: because it doesn't compile
-                                         // under MSVC.
-                                         is_copy_assignable<T>...>>>::value,
-    VariantMoveAssignBase<T...>, VariantCopyAssignBaseNontrivial<T...>>;
-
-template <class... T>
-using VariantBase = VariantCopyAssignBase<T...>;
-
-template <class... T>
-class VariantStateBaseDestructorNontrivial : protected VariantStateBase<T...> {
- private:
-  using Base = VariantStateBase<T...>;
-
- protected:
-  using Base::Base;
-
-  VariantStateBaseDestructorNontrivial() = default;
-  VariantStateBaseDestructorNontrivial(VariantStateBaseDestructorNontrivial&&) =
-      default;
-  VariantStateBaseDestructorNontrivial(
-      const VariantStateBaseDestructorNontrivial&) = default;
-  VariantStateBaseDestructorNontrivial& operator=(
-      VariantStateBaseDestructorNontrivial&&) = default;
-  VariantStateBaseDestructorNontrivial& operator=(
-      const VariantStateBaseDestructorNontrivial&) = default;
-
-  struct Destroyer {
-    template <std::size_t I>
-    void operator()(SizeT<I> i) const {
-      using Alternative =
-          typename absl::variant_alternative<I, variant<T...>>::type;
-      variant_internal::AccessUnion(self->state_, i).~Alternative();
-    }
-
-    void operator()(SizeT<absl::variant_npos> /*i*/) const {
-      // This space intentionally left blank
-    }
-
-    VariantStateBaseDestructorNontrivial* self;
-  };
-
-  void destroy() { VisitIndices<sizeof...(T)>::Run(Destroyer{this}, index_); }
-
-  ~VariantStateBaseDestructorNontrivial() { destroy(); }
-
- protected:
-  using Base::index_;
-  using Base::state_;
-};
-
-template <class... T>
-class VariantMoveBaseNontrivial : protected VariantStateBaseDestructor<T...> {
- private:
-  using Base = VariantStateBaseDestructor<T...>;
-
- protected:
-  using Base::Base;
-
-  struct Construct {
-    template <std::size_t I>
-    void operator()(SizeT<I> i) const {
-      using Alternative =
-          typename absl::variant_alternative<I, variant<T...>>::type;
-      ::new (static_cast<void*>(&self->state_)) Alternative(
-          variant_internal::AccessUnion(std::move(other->state_), i));
-    }
-
-    void operator()(SizeT<absl::variant_npos> /*i*/) const {}
-
-    VariantMoveBaseNontrivial* self;
-    VariantMoveBaseNontrivial* other;
-  };
-
-  VariantMoveBaseNontrivial() = default;
-  VariantMoveBaseNontrivial(VariantMoveBaseNontrivial&& other) noexcept(
-      absl::conjunction<std::is_nothrow_move_constructible<T>...>::value)
-      : Base(NoopConstructorTag()) {
-    VisitIndices<sizeof...(T)>::Run(Construct{this, &other}, other.index_);
-    index_ = other.index_;
-  }
-
-  VariantMoveBaseNontrivial(VariantMoveBaseNontrivial const&) = default;
-
-  VariantMoveBaseNontrivial& operator=(VariantMoveBaseNontrivial&&) = default;
-  VariantMoveBaseNontrivial& operator=(VariantMoveBaseNontrivial const&) =
-      default;
-
- protected:
-  using Base::index_;
-  using Base::state_;
-};
-
-template <class... T>
-class VariantCopyBaseNontrivial : protected VariantMoveBase<T...> {
- private:
-  using Base = VariantMoveBase<T...>;
-
- protected:
-  using Base::Base;
-
-  VariantCopyBaseNontrivial() = default;
-  VariantCopyBaseNontrivial(VariantCopyBaseNontrivial&&) = default;
-
-  struct Construct {
-    template <std::size_t I>
-    void operator()(SizeT<I> i) const {
-      using Alternative =
-          typename absl::variant_alternative<I, variant<T...>>::type;
-      ::new (static_cast<void*>(&self->state_))
-          Alternative(variant_internal::AccessUnion(other->state_, i));
-    }
-
-    void operator()(SizeT<absl::variant_npos> /*i*/) const {}
-
-    VariantCopyBaseNontrivial* self;
-    const VariantCopyBaseNontrivial* other;
-  };
-
-  VariantCopyBaseNontrivial(VariantCopyBaseNontrivial const& other)
-      : Base(NoopConstructorTag()) {
-    VisitIndices<sizeof...(T)>::Run(Construct{this, &other}, other.index_);
-    index_ = other.index_;
-  }
-
-  VariantCopyBaseNontrivial& operator=(VariantCopyBaseNontrivial&&) = default;
-  VariantCopyBaseNontrivial& operator=(VariantCopyBaseNontrivial const&) =
-      default;
-
- protected:
-  using Base::index_;
-  using Base::state_;
-};
-
-template <class... T>
-class VariantMoveAssignBaseNontrivial : protected VariantCopyBase<T...> {
-  friend struct VariantCoreAccess;
-
- private:
-  using Base = VariantCopyBase<T...>;
-
- protected:
-  using Base::Base;
-
-  VariantMoveAssignBaseNontrivial() = default;
-  VariantMoveAssignBaseNontrivial(VariantMoveAssignBaseNontrivial&&) = default;
-  VariantMoveAssignBaseNontrivial(const VariantMoveAssignBaseNontrivial&) =
-      default;
-  VariantMoveAssignBaseNontrivial& operator=(
-      VariantMoveAssignBaseNontrivial const&) = default;
-
-  VariantMoveAssignBaseNontrivial&
-  operator=(VariantMoveAssignBaseNontrivial&& other) noexcept(
-      absl::conjunction<std::is_nothrow_move_constructible<T>...,
-                        std::is_nothrow_move_assignable<T>...>::value) {
-    VisitIndices<sizeof...(T)>::Run(
-        VariantCoreAccess::MakeMoveAssignVisitor(this, &other), other.index_);
-    return *this;
-  }
-
- protected:
-  using Base::index_;
-  using Base::state_;
-};
-
-template <class... T>
-class VariantCopyAssignBaseNontrivial : protected VariantMoveAssignBase<T...> {
-  friend struct VariantCoreAccess;
-
- private:
-  using Base = VariantMoveAssignBase<T...>;
-
- protected:
-  using Base::Base;
-
-  VariantCopyAssignBaseNontrivial() = default;
-  VariantCopyAssignBaseNontrivial(VariantCopyAssignBaseNontrivial&&) = default;
-  VariantCopyAssignBaseNontrivial(const VariantCopyAssignBaseNontrivial&) =
-      default;
-  VariantCopyAssignBaseNontrivial& operator=(
-      VariantCopyAssignBaseNontrivial&&) = default;
-
-  VariantCopyAssignBaseNontrivial& operator=(
-      const VariantCopyAssignBaseNontrivial& other) {
-    VisitIndices<sizeof...(T)>::Run(
-        VariantCoreAccess::MakeCopyAssignVisitor(this, other), other.index_);
-    return *this;
-  }
-
- protected:
-  using Base::index_;
-  using Base::state_;
-};
-
-////////////////////////////////////////
-// Visitors for Comparison Operations //
-////////////////////////////////////////
-
-template <class... Types>
-struct EqualsOp {
-  const variant<Types...>* v;
-  const variant<Types...>* w;
-
-  constexpr bool operator()(SizeT<absl::variant_npos> /*v_i*/) const {
-    return true;
-  }
-
-  template <std::size_t I>
-  constexpr bool operator()(SizeT<I> /*v_i*/) const {
-    return VariantCoreAccess::Access<I>(*v) == VariantCoreAccess::Access<I>(*w);
-  }
-};
-
-template <class... Types>
-struct NotEqualsOp {
-  const variant<Types...>* v;
-  const variant<Types...>* w;
-
-  constexpr bool operator()(SizeT<absl::variant_npos> /*v_i*/) const {
-    return false;
-  }
-
-  template <std::size_t I>
-  constexpr bool operator()(SizeT<I> /*v_i*/) const {
-    return VariantCoreAccess::Access<I>(*v) != VariantCoreAccess::Access<I>(*w);
-  }
-};
-
-template <class... Types>
-struct LessThanOp {
-  const variant<Types...>* v;
-  const variant<Types...>* w;
-
-  constexpr bool operator()(SizeT<absl::variant_npos> /*v_i*/) const {
-    return false;
-  }
-
-  template <std::size_t I>
-  constexpr bool operator()(SizeT<I> /*v_i*/) const {
-    return VariantCoreAccess::Access<I>(*v) < VariantCoreAccess::Access<I>(*w);
-  }
-};
-
-template <class... Types>
-struct GreaterThanOp {
-  const variant<Types...>* v;
-  const variant<Types...>* w;
-
-  constexpr bool operator()(SizeT<absl::variant_npos> /*v_i*/) const {
-    return false;
-  }
-
-  template <std::size_t I>
-  constexpr bool operator()(SizeT<I> /*v_i*/) const {
-    return VariantCoreAccess::Access<I>(*v) > VariantCoreAccess::Access<I>(*w);
-  }
-};
-
-template <class... Types>
-struct LessThanOrEqualsOp {
-  const variant<Types...>* v;
-  const variant<Types...>* w;
-
-  constexpr bool operator()(SizeT<absl::variant_npos> /*v_i*/) const {
-    return true;
-  }
-
-  template <std::size_t I>
-  constexpr bool operator()(SizeT<I> /*v_i*/) const {
-    return VariantCoreAccess::Access<I>(*v) <= VariantCoreAccess::Access<I>(*w);
-  }
-};
-
-template <class... Types>
-struct GreaterThanOrEqualsOp {
-  const variant<Types...>* v;
-  const variant<Types...>* w;
-
-  constexpr bool operator()(SizeT<absl::variant_npos> /*v_i*/) const {
-    return true;
-  }
-
-  template <std::size_t I>
-  constexpr bool operator()(SizeT<I> /*v_i*/) const {
-    return VariantCoreAccess::Access<I>(*v) >= VariantCoreAccess::Access<I>(*w);
-  }
-};
-
-// Precondition: v.index() == w.index();
-template <class... Types>
-struct SwapSameIndex {
-  variant<Types...>* v;
-  variant<Types...>* w;
-  template <std::size_t I>
-  void operator()(SizeT<I>) const {
-    type_traits_internal::Swap(VariantCoreAccess::Access<I>(*v),
-                               VariantCoreAccess::Access<I>(*w));
-  }
-
-  void operator()(SizeT<variant_npos>) const {}
-};
-
-// TODO(calabrese) do this from a different namespace for proper adl usage
-template <class... Types>
-struct Swap {
-  variant<Types...>* v;
-  variant<Types...>* w;
-
-  void generic_swap() const {
-    variant<Types...> tmp(std::move(*w));
-    VariantCoreAccess::Destroy(*w);
-    VariantCoreAccess::InitFrom(*w, std::move(*v));
-    VariantCoreAccess::Destroy(*v);
-    VariantCoreAccess::InitFrom(*v, std::move(tmp));
-  }
-
-  void operator()(SizeT<absl::variant_npos> /*w_i*/) const {
-    if (!v->valueless_by_exception()) {
-      generic_swap();
-    }
-  }
-
-  template <std::size_t Wi>
-  void operator()(SizeT<Wi> /*w_i*/) {
-    if (v->index() == Wi) {
-      VisitIndices<sizeof...(Types)>::Run(SwapSameIndex<Types...>{v, w}, Wi);
-    } else {
-      generic_swap();
-    }
-  }
-};
-
-template <typename Variant, typename = void, typename... Ts>
-struct VariantHashBase {
-  VariantHashBase() = delete;
-  VariantHashBase(const VariantHashBase&) = delete;
-  VariantHashBase(VariantHashBase&&) = delete;
-  VariantHashBase& operator=(const VariantHashBase&) = delete;
-  VariantHashBase& operator=(VariantHashBase&&) = delete;
-};
-
-struct VariantHashVisitor {
-  template <typename T>
-  size_t operator()(const T& t) {
-    return std::hash<T>{}(t);
-  }
-};
-
-template <typename Variant, typename... Ts>
-struct VariantHashBase<Variant,
-                       absl::enable_if_t<absl::conjunction<
-                           type_traits_internal::IsHashable<Ts>...>::value>,
-                       Ts...> {
-  using argument_type = Variant;
-  using result_type = size_t;
-  size_t operator()(const Variant& var) const {
-    type_traits_internal::AssertHashEnabled<Ts...>();
-    if (var.valueless_by_exception()) {
-      return 239799884;
-    }
-    size_t result = VisitIndices<variant_size<Variant>::value>::Run(
-        PerformVisitation<VariantHashVisitor, const Variant&>{
-            std::forward_as_tuple(var), VariantHashVisitor{}},
-        var.index());
-    // Combine the index and the hash result in order to distinguish
-    // std::variant<int, int> holding the same value as different alternative.
-    return result ^ var.index();
-  }
-};
-
-}  // namespace variant_internal
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // !defined(ABSL_USES_STD_VARIANT)
-#endif  // ABSL_TYPES_INTERNAL_VARIANT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/types/optional.h b/projects/llm_framework/include/abseil-cpp/absl/types/optional.h
deleted file mode 100644
index cf7249cb..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/types/optional.h
+++ /dev/null
@@ -1,779 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// optional.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines the `absl::optional` type for holding a value which
-// may or may not be present. This type is useful for providing value semantics
-// for operations that may either wish to return or hold "something-or-nothing".
-//
-// Example:
-//
-//   // A common way to signal operation failure is to provide an output
-//   // parameter and a bool return type:
-//   bool AcquireResource(const Input&, Resource * out);
-//
-//   // Providing an absl::optional return type provides a cleaner API:
-//   absl::optional<Resource> AcquireResource(const Input&);
-//
-// `absl::optional` is a C++11 compatible version of the C++17 `std::optional`
-// abstraction and is designed to be a drop-in replacement for code compliant
-// with C++17.
-#ifndef ABSL_TYPES_OPTIONAL_H_
-#define ABSL_TYPES_OPTIONAL_H_
-
-#include "absl/base/config.h"   // TODO(calabrese) IWYU removal?
-#include "absl/utility/utility.h"
-
-#ifdef ABSL_USES_STD_OPTIONAL
-
-#include <optional>  // IWYU pragma: export
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-using std::bad_optional_access;
-using std::optional;
-using std::make_optional;
-using std::nullopt_t;
-using std::nullopt;
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#else  // ABSL_USES_STD_OPTIONAL
-
-#include <cassert>
-#include <functional>
-#include <initializer_list>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/attributes.h"
-#include "absl/base/nullability.h"
-#include "absl/base/internal/inline_variable.h"
-#include "absl/meta/type_traits.h"
-#include "absl/types/bad_optional_access.h"
-#include "absl/types/internal/optional.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// nullopt_t
-//
-// Class type for `absl::nullopt` used to indicate an `absl::optional<T>` type
-// that does not contain a value.
-struct nullopt_t {
-  // It must not be default-constructible to avoid ambiguity for opt = {}.
-  explicit constexpr nullopt_t(optional_internal::init_t) noexcept {}
-};
-
-// nullopt
-//
-// A tag constant of type `absl::nullopt_t` used to indicate an empty
-// `absl::optional` in certain functions, such as construction or assignment.
-ABSL_INTERNAL_INLINE_CONSTEXPR(nullopt_t, nullopt,
-                               nullopt_t(optional_internal::init_t()));
-
-// -----------------------------------------------------------------------------
-// absl::optional
-// -----------------------------------------------------------------------------
-//
-// A value of type `absl::optional<T>` holds either a value of `T` or an
-// "empty" value.  When it holds a value of `T`, it stores it as a direct
-// sub-object, so `sizeof(optional<T>)` is approximately
-// `sizeof(T) + sizeof(bool)`.
-//
-// This implementation is based on the specification in the latest draft of the
-// C++17 `std::optional` specification as of May 2017, section 20.6.
-//
-// Differences between `absl::optional<T>` and `std::optional<T>` include:
-//
-//    * `constexpr` is not used for non-const member functions.
-//      (dependency on some differences between C++11 and C++14.)
-//    * `absl::nullopt` and `absl::in_place` are not declared `constexpr`. We
-//      need the inline variable support in C++17 for external linkage.
-//    * Throws `absl::bad_optional_access` instead of
-//      `std::bad_optional_access`.
-//    * `make_optional()` cannot be declared `constexpr` due to the absence of
-//      guaranteed copy elision.
-//    * The move constructor's `noexcept` specification is stronger, i.e. if the
-//      default allocator is non-throwing (via setting
-//      `ABSL_ALLOCATOR_NOTHROW`), it evaluates to `noexcept(true)`, because
-//      we assume
-//       a) move constructors should only throw due to allocation failure and
-//       b) if T's move constructor allocates, it uses the same allocation
-//          function as the default allocator.
-//
-template <typename T>
-class optional : private optional_internal::optional_data<T>,
-                 private optional_internal::optional_ctor_base<
-                     optional_internal::ctor_copy_traits<T>::traits>,
-                 private optional_internal::optional_assign_base<
-                     optional_internal::assign_copy_traits<T>::traits> {
-  using data_base = optional_internal::optional_data<T>;
-
- public:
-  typedef T value_type;
-
-  // Constructors
-
-  // Constructs an `optional` holding an empty value, NOT a default constructed
-  // `T`.
-  constexpr optional() noexcept = default;
-
-  // Constructs an `optional` initialized with `nullopt` to hold an empty value.
-  constexpr optional(nullopt_t) noexcept {}  // NOLINT(runtime/explicit)
-
-  // Copy constructor, standard semantics
-  optional(const optional&) = default;
-
-  // Move constructor, standard semantics
-  optional(optional&&) = default;
-
-  // Constructs a non-empty `optional` direct-initialized value of type `T` from
-  // the arguments `std::forward<Args>(args)...`  within the `optional`.
-  // (The `in_place_t` is a tag used to indicate that the contained object
-  // should be constructed in-place.)
-  template <typename InPlaceT, typename... Args,
-            absl::enable_if_t<absl::conjunction<
-                std::is_same<InPlaceT, in_place_t>,
-                std::is_constructible<T, Args&&...> >::value>* = nullptr>
-  constexpr explicit optional(InPlaceT, Args&&... args)
-      : data_base(in_place_t(), std::forward<Args>(args)...) {}
-
-  // Constructs a non-empty `optional` direct-initialized value of type `T` from
-  // the arguments of an initializer_list and `std::forward<Args>(args)...`.
-  // (The `in_place_t` is a tag used to indicate that the contained object
-  // should be constructed in-place.)
-  template <typename U, typename... Args,
-            typename = typename std::enable_if<std::is_constructible<
-                T, std::initializer_list<U>&, Args&&...>::value>::type>
-  constexpr explicit optional(in_place_t, std::initializer_list<U> il,
-                              Args&&... args)
-      : data_base(in_place_t(), il, std::forward<Args>(args)...) {}
-
-  // Value constructor (implicit)
-  template <
-      typename U = T,
-      typename std::enable_if<
-          absl::conjunction<absl::negation<std::is_same<
-                                in_place_t, typename std::decay<U>::type> >,
-                            absl::negation<std::is_same<
-                                optional<T>, typename std::decay<U>::type> >,
-                            std::is_convertible<U&&, T>,
-                            std::is_constructible<T, U&&> >::value,
-          bool>::type = false>
-  constexpr optional(U&& v) : data_base(in_place_t(), std::forward<U>(v)) {}
-
-  // Value constructor (explicit)
-  template <
-      typename U = T,
-      typename std::enable_if<
-          absl::conjunction<absl::negation<std::is_same<
-                                in_place_t, typename std::decay<U>::type> >,
-                            absl::negation<std::is_same<
-                                optional<T>, typename std::decay<U>::type> >,
-                            absl::negation<std::is_convertible<U&&, T> >,
-                            std::is_constructible<T, U&&> >::value,
-          bool>::type = false>
-  explicit constexpr optional(U&& v)
-      : data_base(in_place_t(), std::forward<U>(v)) {}
-
-  // Converting copy constructor (implicit)
-  template <typename U,
-            typename std::enable_if<
-                absl::conjunction<
-                    absl::negation<std::is_same<T, U> >,
-                    std::is_constructible<T, const U&>,
-                    absl::negation<
-                        optional_internal::
-                            is_constructible_convertible_from_optional<T, U> >,
-                    std::is_convertible<const U&, T> >::value,
-                bool>::type = false>
-  optional(const optional<U>& rhs) {
-    if (rhs) {
-      this->construct(*rhs);
-    }
-  }
-
-  // Converting copy constructor (explicit)
-  template <typename U,
-            typename std::enable_if<
-                absl::conjunction<
-                    absl::negation<std::is_same<T, U>>,
-                    std::is_constructible<T, const U&>,
-                    absl::negation<
-                        optional_internal::
-                            is_constructible_convertible_from_optional<T, U>>,
-                    absl::negation<std::is_convertible<const U&, T>>>::value,
-                bool>::type = false>
-  explicit optional(const optional<U>& rhs) {
-    if (rhs) {
-      this->construct(*rhs);
-    }
-  }
-
-  // Converting move constructor (implicit)
-  template <typename U,
-            typename std::enable_if<
-                absl::conjunction<
-                    absl::negation<std::is_same<T, U> >,
-                    std::is_constructible<T, U&&>,
-                    absl::negation<
-                        optional_internal::
-                            is_constructible_convertible_from_optional<T, U> >,
-                    std::is_convertible<U&&, T> >::value,
-                bool>::type = false>
-  optional(optional<U>&& rhs) {
-    if (rhs) {
-      this->construct(std::move(*rhs));
-    }
-  }
-
-  // Converting move constructor (explicit)
-  template <
-      typename U,
-      typename std::enable_if<
-          absl::conjunction<
-              absl::negation<std::is_same<T, U>>, std::is_constructible<T, U&&>,
-              absl::negation<
-                  optional_internal::is_constructible_convertible_from_optional<
-                      T, U>>,
-              absl::negation<std::is_convertible<U&&, T>>>::value,
-          bool>::type = false>
-  explicit optional(optional<U>&& rhs) {
-    if (rhs) {
-      this->construct(std::move(*rhs));
-    }
-  }
-
-  // Destructor. Trivial if `T` is trivially destructible.
-  ~optional() = default;
-
-  // Assignment Operators
-
-  // Assignment from `nullopt`
-  //
-  // Example:
-  //
-  //   struct S { int value; };
-  //   optional<S> opt = absl::nullopt;  // Could also use opt = { };
-  optional& operator=(nullopt_t) noexcept {
-    this->destruct();
-    return *this;
-  }
-
-  // Copy assignment operator, standard semantics
-  optional& operator=(const optional& src) = default;
-
-  // Move assignment operator, standard semantics
-  optional& operator=(optional&& src) = default;
-
-  // Value assignment operators
-  template <typename U = T,
-            int&...,  // Workaround an internal compiler error in GCC 5 to 10.
-            typename = typename std::enable_if<absl::conjunction<
-                absl::negation<
-                    std::is_same<optional<T>, typename std::decay<U>::type> >,
-                absl::negation<absl::conjunction<
-                    std::is_scalar<T>,
-                    std::is_same<T, typename std::decay<U>::type> > >,
-                std::is_constructible<T, U>,
-                std::is_assignable<T&, U> >::value>::type>
-  optional& operator=(U&& v) {
-    this->assign(std::forward<U>(v));
-    return *this;
-  }
-
-  template <
-      typename U,
-      int&...,  // Workaround an internal compiler error in GCC 5 to 10.
-      typename = typename std::enable_if<absl::conjunction<
-          absl::negation<std::is_same<T, U> >,
-          std::is_constructible<T, const U&>, std::is_assignable<T&, const U&>,
-          absl::negation<
-              optional_internal::
-                  is_constructible_convertible_assignable_from_optional<
-                      T, U> > >::value>::type>
-  optional& operator=(const optional<U>& rhs) {
-    if (rhs) {
-      this->assign(*rhs);
-    } else {
-      this->destruct();
-    }
-    return *this;
-  }
-
-  template <typename U,
-            int&...,  // Workaround an internal compiler error in GCC 5 to 10.
-            typename = typename std::enable_if<absl::conjunction<
-                absl::negation<std::is_same<T, U> >,
-                std::is_constructible<T, U>, std::is_assignable<T&, U>,
-                absl::negation<
-                    optional_internal::
-                        is_constructible_convertible_assignable_from_optional<
-                            T, U> > >::value>::type>
-  optional& operator=(optional<U>&& rhs) {
-    if (rhs) {
-      this->assign(std::move(*rhs));
-    } else {
-      this->destruct();
-    }
-    return *this;
-  }
-
-  // Modifiers
-
-  // optional::reset()
-  //
-  // Destroys the inner `T` value of an `absl::optional` if one is present.
-  ABSL_ATTRIBUTE_REINITIALIZES void reset() noexcept { this->destruct(); }
-
-  // optional::emplace()
-  //
-  // (Re)constructs the underlying `T` in-place with the given forwarded
-  // arguments.
-  //
-  // Example:
-  //
-  //   optional<Foo> opt;
-  //   opt.emplace(arg1,arg2,arg3);  // Constructs Foo(arg1,arg2,arg3)
-  //
-  // If the optional is non-empty, and the `args` refer to subobjects of the
-  // current object, then behaviour is undefined, because the current object
-  // will be destructed before the new object is constructed with `args`.
-  template <typename... Args,
-            typename = typename std::enable_if<
-                std::is_constructible<T, Args&&...>::value>::type>
-  T& emplace(Args&&... args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    this->destruct();
-    this->construct(std::forward<Args>(args)...);
-    return reference();
-  }
-
-  // Emplace reconstruction overload for an initializer list and the given
-  // forwarded arguments.
-  //
-  // Example:
-  //
-  //   struct Foo {
-  //     Foo(std::initializer_list<int>);
-  //   };
-  //
-  //   optional<Foo> opt;
-  //   opt.emplace({1,2,3});  // Constructs Foo({1,2,3})
-  template <typename U, typename... Args,
-            typename = typename std::enable_if<std::is_constructible<
-                T, std::initializer_list<U>&, Args&&...>::value>::type>
-  T& emplace(std::initializer_list<U> il,
-             Args&&... args) ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    this->destruct();
-    this->construct(il, std::forward<Args>(args)...);
-    return reference();
-  }
-
-  // Swaps
-
-  // Swap, standard semantics
-  void swap(optional& rhs) noexcept(
-      std::is_nothrow_move_constructible<T>::value&&
-          type_traits_internal::IsNothrowSwappable<T>::value) {
-    if (*this) {
-      if (rhs) {
-        type_traits_internal::Swap(**this, *rhs);
-      } else {
-        rhs.construct(std::move(**this));
-        this->destruct();
-      }
-    } else {
-      if (rhs) {
-        this->construct(std::move(*rhs));
-        rhs.destruct();
-      } else {
-        // No effect (swap(disengaged, disengaged)).
-      }
-    }
-  }
-
-  // Observers
-
-  // optional::operator->()
-  //
-  // Accesses the underlying `T` value's member `m` of an `optional`. If the
-  // `optional` is empty, behavior is undefined.
-  //
-  // If you need myOpt->foo in constexpr, use (*myOpt).foo instead.
-  absl::Nonnull<const T*> operator->() const ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(this->engaged_);
-    return std::addressof(this->data_);
-  }
-  absl::Nonnull<T*> operator->() ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(this->engaged_);
-    return std::addressof(this->data_);
-  }
-
-  // optional::operator*()
-  //
-  // Accesses the underlying `T` value of an `optional`. If the `optional` is
-  // empty, behavior is undefined.
-  constexpr const T& operator*() const& ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return ABSL_HARDENING_ASSERT(this->engaged_), reference();
-  }
-  T& operator*() & ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(this->engaged_);
-    return reference();
-  }
-  constexpr const T&& operator*() const&& ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return ABSL_HARDENING_ASSERT(this->engaged_), std::move(reference());
-  }
-  T&& operator*() && ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    ABSL_HARDENING_ASSERT(this->engaged_);
-    return std::move(reference());
-  }
-
-  // optional::operator bool()
-  //
-  // Returns false if and only if the `optional` is empty.
-  //
-  //   if (opt) {
-  //     // do something with *opt or opt->;
-  //   } else {
-  //     // opt is empty.
-  //   }
-  //
-  constexpr explicit operator bool() const noexcept { return this->engaged_; }
-
-  // optional::has_value()
-  //
-  // Determines whether the `optional` contains a value. Returns `false` if and
-  // only if `*this` is empty.
-  constexpr bool has_value() const noexcept { return this->engaged_; }
-
-// Suppress bogus warning on MSVC: MSVC complains call to reference() after
-// throw_bad_optional_access() is unreachable.
-#ifdef _MSC_VER
-#pragma warning(push)
-#pragma warning(disable : 4702)
-#endif  // _MSC_VER
-  // optional::value()
-  //
-  // Returns a reference to an `optional`s underlying value. The constness
-  // and lvalue/rvalue-ness of the `optional` is preserved to the view of
-  // the `T` sub-object. Throws `absl::bad_optional_access` when the `optional`
-  // is empty.
-  constexpr const T& value() const& ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return static_cast<bool>(*this)
-               ? reference()
-               : (optional_internal::throw_bad_optional_access(), reference());
-  }
-  T& value() & ABSL_ATTRIBUTE_LIFETIME_BOUND {
-    return static_cast<bool>(*this)
-               ? reference()
-               : (optional_internal::throw_bad_optional_access(), reference());
-  }
-  T&& value() && ABSL_ATTRIBUTE_LIFETIME_BOUND {  // NOLINT(build/c++11)
-    return std::move(
-        static_cast<bool>(*this)
-            ? reference()
-            : (optional_internal::throw_bad_optional_access(), reference()));
-  }
-  constexpr const T&& value()
-      const&& ABSL_ATTRIBUTE_LIFETIME_BOUND {  // NOLINT(build/c++11)
-    return std::move(
-        static_cast<bool>(*this)
-            ? reference()
-            : (optional_internal::throw_bad_optional_access(), reference()));
-  }
-#ifdef _MSC_VER
-#pragma warning(pop)
-#endif  // _MSC_VER
-
-  // optional::value_or()
-  //
-  // Returns either the value of `T` or a passed default `v` if the `optional`
-  // is empty.
-  template <typename U>
-  constexpr T value_or(U&& v) const& {
-    static_assert(std::is_copy_constructible<value_type>::value,
-                  "optional<T>::value_or: T must be copy constructible");
-    static_assert(std::is_convertible<U&&, value_type>::value,
-                  "optional<T>::value_or: U must be convertible to T");
-    return static_cast<bool>(*this) ? **this
-                                    : static_cast<T>(std::forward<U>(v));
-  }
-  template <typename U>
-  T value_or(U&& v) && {  // NOLINT(build/c++11)
-    static_assert(std::is_move_constructible<value_type>::value,
-                  "optional<T>::value_or: T must be move constructible");
-    static_assert(std::is_convertible<U&&, value_type>::value,
-                  "optional<T>::value_or: U must be convertible to T");
-    return static_cast<bool>(*this) ? std::move(**this)
-                                    : static_cast<T>(std::forward<U>(v));
-  }
-
- private:
-  // Private accessors for internal storage viewed as reference to T.
-  constexpr const T& reference() const { return this->data_; }
-  T& reference() { return this->data_; }
-
-  // T constraint checks.  You can't have an optional of nullopt_t, in_place_t
-  // or a reference.
-  static_assert(
-      !std::is_same<nullopt_t, typename std::remove_cv<T>::type>::value,
-      "optional<nullopt_t> is not allowed.");
-  static_assert(
-      !std::is_same<in_place_t, typename std::remove_cv<T>::type>::value,
-      "optional<in_place_t> is not allowed.");
-  static_assert(!std::is_reference<T>::value,
-                "optional<reference> is not allowed.");
-};
-
-// Non-member functions
-
-// swap()
-//
-// Performs a swap between two `absl::optional` objects, using standard
-// semantics.
-template <typename T, typename std::enable_if<
-                          std::is_move_constructible<T>::value &&
-                              type_traits_internal::IsSwappable<T>::value,
-                          bool>::type = false>
-void swap(optional<T>& a, optional<T>& b) noexcept(noexcept(a.swap(b))) {
-  a.swap(b);
-}
-
-// make_optional()
-//
-// Creates a non-empty `optional<T>` where the type of `T` is deduced. An
-// `absl::optional` can also be explicitly instantiated with
-// `make_optional<T>(v)`.
-//
-// Note: `make_optional()` constructions may be declared `constexpr` for
-// trivially copyable types `T`. Non-trivial types require copy elision
-// support in C++17 for `make_optional` to support `constexpr` on such
-// non-trivial types.
-//
-// Example:
-//
-//   constexpr absl::optional<int> opt = absl::make_optional(1);
-//   static_assert(opt.value() == 1, "");
-template <typename T>
-constexpr optional<typename std::decay<T>::type> make_optional(T&& v) {
-  return optional<typename std::decay<T>::type>(std::forward<T>(v));
-}
-
-template <typename T, typename... Args>
-constexpr optional<T> make_optional(Args&&... args) {
-  return optional<T>(in_place_t(), std::forward<Args>(args)...);
-}
-
-template <typename T, typename U, typename... Args>
-constexpr optional<T> make_optional(std::initializer_list<U> il,
-                                    Args&&... args) {
-  return optional<T>(in_place_t(), il, std::forward<Args>(args)...);
-}
-
-// Relational operators [optional.relops]
-
-// Empty optionals are considered equal to each other and less than non-empty
-// optionals. Supports relations between optional<T> and optional<U>, between
-// optional<T> and U, and between optional<T> and nullopt.
-//
-// Note: We're careful to support T having non-bool relationals.
-
-// Requires: The expression, e.g. "*x == *y" shall be well-formed and its result
-// shall be convertible to bool.
-// The C++17 (N4606) "Returns:" statements are translated into
-// code in an obvious way here, and the original text retained as function docs.
-// Returns: If bool(x) != bool(y), false; otherwise if bool(x) == false, true;
-// otherwise *x == *y.
-template <typename T, typename U>
-constexpr auto operator==(const optional<T>& x, const optional<U>& y)
-    -> decltype(optional_internal::convertible_to_bool(*x == *y)) {
-  return static_cast<bool>(x) != static_cast<bool>(y)
-             ? false
-             : static_cast<bool>(x) == false ? true
-                                             : static_cast<bool>(*x == *y);
-}
-
-// Returns: If bool(x) != bool(y), true; otherwise, if bool(x) == false, false;
-// otherwise *x != *y.
-template <typename T, typename U>
-constexpr auto operator!=(const optional<T>& x, const optional<U>& y)
-    -> decltype(optional_internal::convertible_to_bool(*x != *y)) {
-  return static_cast<bool>(x) != static_cast<bool>(y)
-             ? true
-             : static_cast<bool>(x) == false ? false
-                                             : static_cast<bool>(*x != *y);
-}
-// Returns: If !y, false; otherwise, if !x, true; otherwise *x < *y.
-template <typename T, typename U>
-constexpr auto operator<(const optional<T>& x, const optional<U>& y)
-    -> decltype(optional_internal::convertible_to_bool(*x < *y)) {
-  return !y ? false : !x ? true : static_cast<bool>(*x < *y);
-}
-// Returns: If !x, false; otherwise, if !y, true; otherwise *x > *y.
-template <typename T, typename U>
-constexpr auto operator>(const optional<T>& x, const optional<U>& y)
-    -> decltype(optional_internal::convertible_to_bool(*x > *y)) {
-  return !x ? false : !y ? true : static_cast<bool>(*x > *y);
-}
-// Returns: If !x, true; otherwise, if !y, false; otherwise *x <= *y.
-template <typename T, typename U>
-constexpr auto operator<=(const optional<T>& x, const optional<U>& y)
-    -> decltype(optional_internal::convertible_to_bool(*x <= *y)) {
-  return !x ? true : !y ? false : static_cast<bool>(*x <= *y);
-}
-// Returns: If !y, true; otherwise, if !x, false; otherwise *x >= *y.
-template <typename T, typename U>
-constexpr auto operator>=(const optional<T>& x, const optional<U>& y)
-    -> decltype(optional_internal::convertible_to_bool(*x >= *y)) {
-  return !y ? true : !x ? false : static_cast<bool>(*x >= *y);
-}
-
-// Comparison with nullopt [optional.nullops]
-// The C++17 (N4606) "Returns:" statements are used directly here.
-template <typename T>
-constexpr bool operator==(const optional<T>& x, nullopt_t) noexcept {
-  return !x;
-}
-template <typename T>
-constexpr bool operator==(nullopt_t, const optional<T>& x) noexcept {
-  return !x;
-}
-template <typename T>
-constexpr bool operator!=(const optional<T>& x, nullopt_t) noexcept {
-  return static_cast<bool>(x);
-}
-template <typename T>
-constexpr bool operator!=(nullopt_t, const optional<T>& x) noexcept {
-  return static_cast<bool>(x);
-}
-template <typename T>
-constexpr bool operator<(const optional<T>&, nullopt_t) noexcept {
-  return false;
-}
-template <typename T>
-constexpr bool operator<(nullopt_t, const optional<T>& x) noexcept {
-  return static_cast<bool>(x);
-}
-template <typename T>
-constexpr bool operator<=(const optional<T>& x, nullopt_t) noexcept {
-  return !x;
-}
-template <typename T>
-constexpr bool operator<=(nullopt_t, const optional<T>&) noexcept {
-  return true;
-}
-template <typename T>
-constexpr bool operator>(const optional<T>& x, nullopt_t) noexcept {
-  return static_cast<bool>(x);
-}
-template <typename T>
-constexpr bool operator>(nullopt_t, const optional<T>&) noexcept {
-  return false;
-}
-template <typename T>
-constexpr bool operator>=(const optional<T>&, nullopt_t) noexcept {
-  return true;
-}
-template <typename T>
-constexpr bool operator>=(nullopt_t, const optional<T>& x) noexcept {
-  return !x;
-}
-
-// Comparison with T [optional.comp_with_t]
-
-// Requires: The expression, e.g. "*x == v" shall be well-formed and its result
-// shall be convertible to bool.
-// The C++17 (N4606) "Equivalent to:" statements are used directly here.
-template <typename T, typename U>
-constexpr auto operator==(const optional<T>& x, const U& v)
-    -> decltype(optional_internal::convertible_to_bool(*x == v)) {
-  return static_cast<bool>(x) ? static_cast<bool>(*x == v) : false;
-}
-template <typename T, typename U>
-constexpr auto operator==(const U& v, const optional<T>& x)
-    -> decltype(optional_internal::convertible_to_bool(v == *x)) {
-  return static_cast<bool>(x) ? static_cast<bool>(v == *x) : false;
-}
-template <typename T, typename U>
-constexpr auto operator!=(const optional<T>& x, const U& v)
-    -> decltype(optional_internal::convertible_to_bool(*x != v)) {
-  return static_cast<bool>(x) ? static_cast<bool>(*x != v) : true;
-}
-template <typename T, typename U>
-constexpr auto operator!=(const U& v, const optional<T>& x)
-    -> decltype(optional_internal::convertible_to_bool(v != *x)) {
-  return static_cast<bool>(x) ? static_cast<bool>(v != *x) : true;
-}
-template <typename T, typename U>
-constexpr auto operator<(const optional<T>& x, const U& v)
-    -> decltype(optional_internal::convertible_to_bool(*x < v)) {
-  return static_cast<bool>(x) ? static_cast<bool>(*x < v) : true;
-}
-template <typename T, typename U>
-constexpr auto operator<(const U& v, const optional<T>& x)
-    -> decltype(optional_internal::convertible_to_bool(v < *x)) {
-  return static_cast<bool>(x) ? static_cast<bool>(v < *x) : false;
-}
-template <typename T, typename U>
-constexpr auto operator<=(const optional<T>& x, const U& v)
-    -> decltype(optional_internal::convertible_to_bool(*x <= v)) {
-  return static_cast<bool>(x) ? static_cast<bool>(*x <= v) : true;
-}
-template <typename T, typename U>
-constexpr auto operator<=(const U& v, const optional<T>& x)
-    -> decltype(optional_internal::convertible_to_bool(v <= *x)) {
-  return static_cast<bool>(x) ? static_cast<bool>(v <= *x) : false;
-}
-template <typename T, typename U>
-constexpr auto operator>(const optional<T>& x, const U& v)
-    -> decltype(optional_internal::convertible_to_bool(*x > v)) {
-  return static_cast<bool>(x) ? static_cast<bool>(*x > v) : false;
-}
-template <typename T, typename U>
-constexpr auto operator>(const U& v, const optional<T>& x)
-    -> decltype(optional_internal::convertible_to_bool(v > *x)) {
-  return static_cast<bool>(x) ? static_cast<bool>(v > *x) : true;
-}
-template <typename T, typename U>
-constexpr auto operator>=(const optional<T>& x, const U& v)
-    -> decltype(optional_internal::convertible_to_bool(*x >= v)) {
-  return static_cast<bool>(x) ? static_cast<bool>(*x >= v) : false;
-}
-template <typename T, typename U>
-constexpr auto operator>=(const U& v, const optional<T>& x)
-    -> decltype(optional_internal::convertible_to_bool(v >= *x)) {
-  return static_cast<bool>(x) ? static_cast<bool>(v >= *x) : true;
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-namespace std {
-
-// std::hash specialization for absl::optional.
-template <typename T>
-struct hash<absl::optional<T> >
-    : absl::optional_internal::optional_hash_base<T> {};
-
-}  // namespace std
-
-#undef ABSL_MSVC_CONSTEXPR_BUG_IN_UNION_LIKE_CLASS
-
-#endif  // ABSL_USES_STD_OPTIONAL
-
-#endif  // ABSL_TYPES_OPTIONAL_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/types/span.h b/projects/llm_framework/include/abseil-cpp/absl/types/span.h
deleted file mode 100644
index c183aa88..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/types/span.h
+++ /dev/null
@@ -1,753 +0,0 @@
-//
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// span.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines a `Span<T>` type for holding a reference to existing
-// array data. The `Span` object, much like the `absl::string_view` object,
-// does not own such data itself, and the data being referenced by the span must
-// outlive the span itself. Unlike `view` type references, a span can hold a
-// reference to mutable data (and can mutate it for underlying types of
-// non-const T.) A span provides a lightweight way to pass a reference to such
-// data.
-//
-// Additionally, this header file defines `MakeSpan()` and `MakeConstSpan()`
-// factory functions, for clearly creating spans of type `Span<T>` or read-only
-// `Span<const T>` when such types may be difficult to identify due to issues
-// with implicit conversion.
-//
-// The C++20 draft standard includes a `std::span` type. As of June 2020, the
-// differences between `absl::Span` and `std::span` are:
-//    * `absl::Span` has `operator==` (which is likely a design bug,
-//       per https://abseil.io/blog/20180531-regular-types)
-//    * `absl::Span` has the factory functions `MakeSpan()` and
-//      `MakeConstSpan()`
-//    * bounds-checked access to `absl::Span` is accomplished with `at()`
-//    * `absl::Span` has compiler-provided move and copy constructors and
-//      assignment. This is due to them being specified as `constexpr`, but that
-//      implies const in C++11.
-//    * A read-only `absl::Span<const T>` can be implicitly constructed from an
-//      initializer list.
-//    * `absl::Span` has no `bytes()`, `size_bytes()`, `as_bytes()`, or
-//      `as_writable_bytes()` methods
-//    * `absl::Span` has no static extent template parameter, nor constructors
-//      which exist only because of the static extent parameter.
-//    * `absl::Span` has an explicit mutable-reference constructor
-//
-// For more information, see the class comments below.
-#ifndef ABSL_TYPES_SPAN_H_
-#define ABSL_TYPES_SPAN_H_
-
-#include <algorithm>
-#include <cassert>
-#include <cstddef>
-#include <initializer_list>
-#include <iterator>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/attributes.h"
-#include "absl/base/internal/throw_delegate.h"
-#include "absl/base/macros.h"
-#include "absl/base/nullability.h"
-#include "absl/base/optimization.h"
-#include "absl/base/port.h"    // TODO(strel): remove this include
-#include "absl/meta/type_traits.h"
-#include "absl/types/internal/span.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-//------------------------------------------------------------------------------
-// Span
-//------------------------------------------------------------------------------
-//
-// A `Span` is an "array reference" type for holding a reference of contiguous
-// array data; the `Span` object does not and cannot own such data itself. A
-// span provides an easy way to provide overloads for anything operating on
-// contiguous sequences without needing to manage pointers and array lengths
-// manually.
-
-// A span is conceptually a pointer (ptr) and a length (size) into an already
-// existing array of contiguous memory; the array it represents references the
-// elements "ptr[0] .. ptr[size-1]". Passing a properly-constructed `Span`
-// instead of raw pointers avoids many issues related to index out of bounds
-// errors.
-//
-// Spans may also be constructed from containers holding contiguous sequences.
-// Such containers must supply `data()` and `size() const` methods (e.g
-// `std::vector<T>`, `absl::InlinedVector<T, N>`). All implicit conversions to
-// `absl::Span` from such containers will create spans of type `const T`;
-// spans which can mutate their values (of type `T`) must use explicit
-// constructors.
-//
-// A `Span<T>` is somewhat analogous to an `absl::string_view`, but for an array
-// of elements of type `T`, and unlike an `absl::string_view`, a span can hold a
-// reference to mutable data. A user of `Span` must ensure that the data being
-// pointed to outlives the `Span` itself.
-//
-// You can construct a `Span<T>` in several ways:
-//
-//   * Explicitly from a reference to a container type
-//   * Explicitly from a pointer and size
-//   * Implicitly from a container type (but only for spans of type `const T`)
-//   * Using the `MakeSpan()` or `MakeConstSpan()` factory functions.
-//
-// Examples:
-//
-//   // Construct a Span explicitly from a container:
-//   std::vector<int> v = {1, 2, 3, 4, 5};
-//   auto span = absl::Span<const int>(v);
-//
-//   // Construct a Span explicitly from a C-style array:
-//   int a[5] =  {1, 2, 3, 4, 5};
-//   auto span = absl::Span<const int>(a);
-//
-//   // Construct a Span implicitly from a container
-//   void MyRoutine(absl::Span<const int> a) {
-//     ...
-//   }
-//   std::vector v = {1,2,3,4,5};
-//   MyRoutine(v)                     // convert to Span<const T>
-//
-// Note that `Span` objects, in addition to requiring that the memory they
-// point to remains alive, must also ensure that such memory does not get
-// reallocated. Therefore, to avoid undefined behavior, containers with
-// associated spans should not invoke operations that may reallocate memory
-// (such as resizing) or invalidate iterators into the container.
-//
-// One common use for a `Span` is when passing arguments to a routine that can
-// accept a variety of array types (e.g. a `std::vector`, `absl::InlinedVector`,
-// a C-style array, etc.). Instead of creating overloads for each case, you
-// can simply specify a `Span` as the argument to such a routine.
-//
-// Example:
-//
-//   void MyRoutine(absl::Span<const int> a) {
-//     ...
-//   }
-//
-//   std::vector v = {1,2,3,4,5};
-//   MyRoutine(v);
-//
-//   absl::InlinedVector<int, 4> my_inline_vector;
-//   MyRoutine(my_inline_vector);
-//
-//   // Explicit constructor from pointer,size
-//   int* my_array = new int[10];
-//   MyRoutine(absl::Span<const int>(my_array, 10));
-template <typename T>
-class Span {
- private:
-  // Used to determine whether a Span can be constructed from a container of
-  // type C.
-  template <typename C>
-  using EnableIfConvertibleFrom =
-      typename std::enable_if<span_internal::HasData<T, C>::value &&
-                              span_internal::HasSize<C>::value>::type;
-
-  // Used to SFINAE-enable a function when the slice elements are const.
-  template <typename U>
-  using EnableIfValueIsConst =
-      typename std::enable_if<std::is_const<T>::value, U>::type;
-
-  // Used to SFINAE-enable a function when the slice elements are mutable.
-  template <typename U>
-  using EnableIfValueIsMutable =
-      typename std::enable_if<!std::is_const<T>::value, U>::type;
-
- public:
-  using element_type = T;
-  using value_type = absl::remove_cv_t<T>;
-  // TODO(b/316099902) - pointer should be Nullable<T*>, but this makes it hard
-  // to recognize foreach loops as safe.
-  using pointer = T*;
-  using const_pointer = const T*;
-  using reference = T&;
-  using const_reference = const T&;
-  using iterator = pointer;
-  using const_iterator = const_pointer;
-  using reverse_iterator = std::reverse_iterator<iterator>;
-  using const_reverse_iterator = std::reverse_iterator<const_iterator>;
-  using size_type = size_t;
-  using difference_type = ptrdiff_t;
-
-  static const size_type npos = ~(size_type(0));
-
-  constexpr Span() noexcept : Span(nullptr, 0) {}
-  constexpr Span(pointer array, size_type length) noexcept
-      : ptr_(array), len_(length) {}
-
-  // Implicit conversion constructors
-  template <size_t N>
-  constexpr Span(T (&a)[N]) noexcept  // NOLINT(runtime/explicit)
-      : Span(a, N) {}
-
-  // Explicit reference constructor for a mutable `Span<T>` type. Can be
-  // replaced with MakeSpan() to infer the type parameter.
-  template <typename V, typename = EnableIfConvertibleFrom<V>,
-            typename = EnableIfValueIsMutable<V>,
-            typename = span_internal::EnableIfNotIsView<V>>
-  explicit Span(
-      V& v
-          ABSL_ATTRIBUTE_LIFETIME_BOUND) noexcept  // NOLINT(runtime/references)
-      : Span(span_internal::GetData(v), v.size()) {}
-
-  // Implicit reference constructor for a read-only `Span<const T>` type
-  template <typename V, typename = EnableIfConvertibleFrom<V>,
-            typename = EnableIfValueIsConst<V>,
-            typename = span_internal::EnableIfNotIsView<V>>
-  constexpr Span(
-      const V& v
-          ABSL_ATTRIBUTE_LIFETIME_BOUND) noexcept  // NOLINT(runtime/explicit)
-      : Span(span_internal::GetData(v), v.size()) {}
-
-  // Overloads of the above two functions that are only enabled for view types.
-  // This is so we can drop the ABSL_ATTRIBUTE_LIFETIME_BOUND annotation. These
-  // overloads must be made unique by using a different template parameter list
-  // (hence the = 0 for the IsView enabler).
-  template <typename V, typename = EnableIfConvertibleFrom<V>,
-            typename = EnableIfValueIsMutable<V>,
-            span_internal::EnableIfIsView<V> = 0>
-  explicit Span(V& v) noexcept  // NOLINT(runtime/references)
-      : Span(span_internal::GetData(v), v.size()) {}
-  template <typename V, typename = EnableIfConvertibleFrom<V>,
-            typename = EnableIfValueIsConst<V>,
-            span_internal::EnableIfIsView<V> = 0>
-  constexpr Span(const V& v) noexcept  // NOLINT(runtime/explicit)
-      : Span(span_internal::GetData(v), v.size()) {}
-
-  // Implicit constructor from an initializer list, making it possible to pass a
-  // brace-enclosed initializer list to a function expecting a `Span`. Such
-  // spans constructed from an initializer list must be of type `Span<const T>`.
-  //
-  //   void Process(absl::Span<const int> x);
-  //   Process({1, 2, 3});
-  //
-  // Note that as always the array referenced by the span must outlive the span.
-  // Since an initializer list constructor acts as if it is fed a temporary
-  // array (cf. C++ standard [dcl.init.list]/5), it's safe to use this
-  // constructor only when the `std::initializer_list` itself outlives the span.
-  // In order to meet this requirement it's sufficient to ensure that neither
-  // the span nor a copy of it is used outside of the expression in which it's
-  // created:
-  //
-  //   // Assume that this function uses the array directly, not retaining any
-  //   // copy of the span or pointer to any of its elements.
-  //   void Process(absl::Span<const int> ints);
-  //
-  //   // Okay: the std::initializer_list<int> will reference a temporary array
-  //   // that isn't destroyed until after the call to Process returns.
-  //   Process({ 17, 19 });
-  //
-  //   // Not okay: the storage used by the std::initializer_list<int> is not
-  //   // allowed to be referenced after the first line.
-  //   absl::Span<const int> ints = { 17, 19 };
-  //   Process(ints);
-  //
-  //   // Not okay for the same reason as above: even when the elements of the
-  //   // initializer list expression are not temporaries the underlying array
-  //   // is, so the initializer list must still outlive the span.
-  //   const int foo = 17;
-  //   absl::Span<const int> ints = { foo };
-  //   Process(ints);
-  //
-  template <typename LazyT = T,
-            typename = EnableIfValueIsConst<LazyT>>
-  Span(std::initializer_list<value_type> v
-           ABSL_ATTRIBUTE_LIFETIME_BOUND) noexcept  // NOLINT(runtime/explicit)
-      : Span(v.begin(), v.size()) {}
-
-  // Accessors
-
-  // Span::data()
-  //
-  // Returns a pointer to the span's underlying array of data (which is held
-  // outside the span).
-  constexpr pointer data() const noexcept { return ptr_; }
-
-  // Span::size()
-  //
-  // Returns the size of this span.
-  constexpr size_type size() const noexcept { return len_; }
-
-  // Span::length()
-  //
-  // Returns the length (size) of this span.
-  constexpr size_type length() const noexcept { return size(); }
-
-  // Span::empty()
-  //
-  // Returns a boolean indicating whether or not this span is considered empty.
-  constexpr bool empty() const noexcept { return size() == 0; }
-
-  // Span::operator[]
-  //
-  // Returns a reference to the i'th element of this span.
-  constexpr reference operator[](size_type i) const noexcept {
-    return ABSL_HARDENING_ASSERT(i < size()), ptr_[i];
-  }
-
-  // Span::at()
-  //
-  // Returns a reference to the i'th element of this span.
-  constexpr reference at(size_type i) const {
-    return ABSL_PREDICT_TRUE(i < size())  //
-               ? *(data() + i)
-               : (base_internal::ThrowStdOutOfRange(
-                      "Span::at failed bounds check"),
-                  *(data() + i));
-  }
-
-  // Span::front()
-  //
-  // Returns a reference to the first element of this span. The span must not
-  // be empty.
-  constexpr reference front() const noexcept {
-    return ABSL_HARDENING_ASSERT(size() > 0), *data();
-  }
-
-  // Span::back()
-  //
-  // Returns a reference to the last element of this span. The span must not
-  // be empty.
-  constexpr reference back() const noexcept {
-    return ABSL_HARDENING_ASSERT(size() > 0), *(data() + size() - 1);
-  }
-
-  // Span::begin()
-  //
-  // Returns an iterator pointing to the first element of this span, or `end()`
-  // if the span is empty.
-  constexpr iterator begin() const noexcept { return data(); }
-
-  // Span::cbegin()
-  //
-  // Returns a const iterator pointing to the first element of this span, or
-  // `end()` if the span is empty.
-  constexpr const_iterator cbegin() const noexcept { return begin(); }
-
-  // Span::end()
-  //
-  // Returns an iterator pointing just beyond the last element at the
-  // end of this span. This iterator acts as a placeholder; attempting to
-  // access it results in undefined behavior.
-  constexpr iterator end() const noexcept { return data() + size(); }
-
-  // Span::cend()
-  //
-  // Returns a const iterator pointing just beyond the last element at the
-  // end of this span. This iterator acts as a placeholder; attempting to
-  // access it results in undefined behavior.
-  constexpr const_iterator cend() const noexcept { return end(); }
-
-  // Span::rbegin()
-  //
-  // Returns a reverse iterator pointing to the last element at the end of this
-  // span, or `rend()` if the span is empty.
-  constexpr reverse_iterator rbegin() const noexcept {
-    return reverse_iterator(end());
-  }
-
-  // Span::crbegin()
-  //
-  // Returns a const reverse iterator pointing to the last element at the end of
-  // this span, or `crend()` if the span is empty.
-  constexpr const_reverse_iterator crbegin() const noexcept { return rbegin(); }
-
-  // Span::rend()
-  //
-  // Returns a reverse iterator pointing just before the first element
-  // at the beginning of this span. This pointer acts as a placeholder;
-  // attempting to access its element results in undefined behavior.
-  constexpr reverse_iterator rend() const noexcept {
-    return reverse_iterator(begin());
-  }
-
-  // Span::crend()
-  //
-  // Returns a reverse const iterator pointing just before the first element
-  // at the beginning of this span. This pointer acts as a placeholder;
-  // attempting to access its element results in undefined behavior.
-  constexpr const_reverse_iterator crend() const noexcept { return rend(); }
-
-  // Span mutations
-
-  // Span::remove_prefix()
-  //
-  // Removes the first `n` elements from the span.
-  void remove_prefix(size_type n) noexcept {
-    ABSL_HARDENING_ASSERT(size() >= n);
-    ptr_ += n;
-    len_ -= n;
-  }
-
-  // Span::remove_suffix()
-  //
-  // Removes the last `n` elements from the span.
-  void remove_suffix(size_type n) noexcept {
-    ABSL_HARDENING_ASSERT(size() >= n);
-    len_ -= n;
-  }
-
-  // Span::subspan()
-  //
-  // Returns a `Span` starting at element `pos` and of length `len`. Both `pos`
-  // and `len` are of type `size_type` and thus non-negative. Parameter `pos`
-  // must be <= size(). Any `len` value that points past the end of the span
-  // will be trimmed to at most size() - `pos`. A default `len` value of `npos`
-  // ensures the returned subspan continues until the end of the span.
-  //
-  // Examples:
-  //
-  //   std::vector<int> vec = {10, 11, 12, 13};
-  //   absl::MakeSpan(vec).subspan(1, 2);  // {11, 12}
-  //   absl::MakeSpan(vec).subspan(2, 8);  // {12, 13}
-  //   absl::MakeSpan(vec).subspan(1);     // {11, 12, 13}
-  //   absl::MakeSpan(vec).subspan(4);     // {}
-  //   absl::MakeSpan(vec).subspan(5);     // throws std::out_of_range
-  constexpr Span subspan(size_type pos = 0, size_type len = npos) const {
-    return (pos <= size())
-               ? Span(data() + pos, (std::min)(size() - pos, len))
-               : (base_internal::ThrowStdOutOfRange("pos > size()"), Span());
-  }
-
-  // Span::first()
-  //
-  // Returns a `Span` containing first `len` elements. Parameter `len` is of
-  // type `size_type` and thus non-negative. `len` value must be <= size().
-  //
-  // Examples:
-  //
-  //   std::vector<int> vec = {10, 11, 12, 13};
-  //   absl::MakeSpan(vec).first(1);  // {10}
-  //   absl::MakeSpan(vec).first(3);  // {10, 11, 12}
-  //   absl::MakeSpan(vec).first(5);  // throws std::out_of_range
-  constexpr Span first(size_type len) const {
-    return (len <= size())
-               ? Span(data(), len)
-               : (base_internal::ThrowStdOutOfRange("len > size()"), Span());
-  }
-
-  // Span::last()
-  //
-  // Returns a `Span` containing last `len` elements. Parameter `len` is of
-  // type `size_type` and thus non-negative. `len` value must be <= size().
-  //
-  // Examples:
-  //
-  //   std::vector<int> vec = {10, 11, 12, 13};
-  //   absl::MakeSpan(vec).last(1);  // {13}
-  //   absl::MakeSpan(vec).last(3);  // {11, 12, 13}
-  //   absl::MakeSpan(vec).last(5);  // throws std::out_of_range
-  constexpr Span last(size_type len) const {
-    return (len <= size())
-               ? Span(size() - len + data(), len)
-               : (base_internal::ThrowStdOutOfRange("len > size()"), Span());
-  }
-
-  // Support for absl::Hash.
-  template <typename H>
-  friend H AbslHashValue(H h, Span v) {
-    return H::combine(H::combine_contiguous(std::move(h), v.data(), v.size()),
-                      v.size());
-  }
-
- private:
-  pointer ptr_;
-  size_type len_;
-};
-
-template <typename T>
-const typename Span<T>::size_type Span<T>::npos;
-
-// Span relationals
-
-// Equality is compared element-by-element, while ordering is lexicographical.
-// We provide three overloads for each operator to cover any combination on the
-// left or right hand side of mutable Span<T>, read-only Span<const T>, and
-// convertible-to-read-only Span<T>.
-// TODO(zhangxy): Due to MSVC overload resolution bug with partial ordering
-// template functions, 5 overloads per operator is needed as a workaround. We
-// should update them to 3 overloads per operator using non-deduced context like
-// string_view, i.e.
-// - (Span<T>, Span<T>)
-// - (Span<T>, non_deduced<Span<const T>>)
-// - (non_deduced<Span<const T>>, Span<T>)
-
-// operator==
-template <typename T>
-bool operator==(Span<T> a, Span<T> b) {
-  return span_internal::EqualImpl<Span, const T>(a, b);
-}
-template <typename T>
-bool operator==(Span<const T> a, Span<T> b) {
-  return span_internal::EqualImpl<Span, const T>(a, b);
-}
-template <typename T>
-bool operator==(Span<T> a, Span<const T> b) {
-  return span_internal::EqualImpl<Span, const T>(a, b);
-}
-template <
-    typename T, typename U,
-    typename = span_internal::EnableIfConvertibleTo<U, absl::Span<const T>>>
-bool operator==(const U& a, Span<T> b) {
-  return span_internal::EqualImpl<Span, const T>(a, b);
-}
-template <
-    typename T, typename U,
-    typename = span_internal::EnableIfConvertibleTo<U, absl::Span<const T>>>
-bool operator==(Span<T> a, const U& b) {
-  return span_internal::EqualImpl<Span, const T>(a, b);
-}
-
-// operator!=
-template <typename T>
-bool operator!=(Span<T> a, Span<T> b) {
-  return !(a == b);
-}
-template <typename T>
-bool operator!=(Span<const T> a, Span<T> b) {
-  return !(a == b);
-}
-template <typename T>
-bool operator!=(Span<T> a, Span<const T> b) {
-  return !(a == b);
-}
-template <
-    typename T, typename U,
-    typename = span_internal::EnableIfConvertibleTo<U, absl::Span<const T>>>
-bool operator!=(const U& a, Span<T> b) {
-  return !(a == b);
-}
-template <
-    typename T, typename U,
-    typename = span_internal::EnableIfConvertibleTo<U, absl::Span<const T>>>
-bool operator!=(Span<T> a, const U& b) {
-  return !(a == b);
-}
-
-// operator<
-template <typename T>
-bool operator<(Span<T> a, Span<T> b) {
-  return span_internal::LessThanImpl<Span, const T>(a, b);
-}
-template <typename T>
-bool operator<(Span<const T> a, Span<T> b) {
-  return span_internal::LessThanImpl<Span, const T>(a, b);
-}
-template <typename T>
-bool operator<(Span<T> a, Span<const T> b) {
-  return span_internal::LessThanImpl<Span, const T>(a, b);
-}
-template <
-    typename T, typename U,
-    typename = span_internal::EnableIfConvertibleTo<U, absl::Span<const T>>>
-bool operator<(const U& a, Span<T> b) {
-  return span_internal::LessThanImpl<Span, const T>(a, b);
-}
-template <
-    typename T, typename U,
-    typename = span_internal::EnableIfConvertibleTo<U, absl::Span<const T>>>
-bool operator<(Span<T> a, const U& b) {
-  return span_internal::LessThanImpl<Span, const T>(a, b);
-}
-
-// operator>
-template <typename T>
-bool operator>(Span<T> a, Span<T> b) {
-  return b < a;
-}
-template <typename T>
-bool operator>(Span<const T> a, Span<T> b) {
-  return b < a;
-}
-template <typename T>
-bool operator>(Span<T> a, Span<const T> b) {
-  return b < a;
-}
-template <
-    typename T, typename U,
-    typename = span_internal::EnableIfConvertibleTo<U, absl::Span<const T>>>
-bool operator>(const U& a, Span<T> b) {
-  return b < a;
-}
-template <
-    typename T, typename U,
-    typename = span_internal::EnableIfConvertibleTo<U, absl::Span<const T>>>
-bool operator>(Span<T> a, const U& b) {
-  return b < a;
-}
-
-// operator<=
-template <typename T>
-bool operator<=(Span<T> a, Span<T> b) {
-  return !(b < a);
-}
-template <typename T>
-bool operator<=(Span<const T> a, Span<T> b) {
-  return !(b < a);
-}
-template <typename T>
-bool operator<=(Span<T> a, Span<const T> b) {
-  return !(b < a);
-}
-template <
-    typename T, typename U,
-    typename = span_internal::EnableIfConvertibleTo<U, absl::Span<const T>>>
-bool operator<=(const U& a, Span<T> b) {
-  return !(b < a);
-}
-template <
-    typename T, typename U,
-    typename = span_internal::EnableIfConvertibleTo<U, absl::Span<const T>>>
-bool operator<=(Span<T> a, const U& b) {
-  return !(b < a);
-}
-
-// operator>=
-template <typename T>
-bool operator>=(Span<T> a, Span<T> b) {
-  return !(a < b);
-}
-template <typename T>
-bool operator>=(Span<const T> a, Span<T> b) {
-  return !(a < b);
-}
-template <typename T>
-bool operator>=(Span<T> a, Span<const T> b) {
-  return !(a < b);
-}
-template <
-    typename T, typename U,
-    typename = span_internal::EnableIfConvertibleTo<U, absl::Span<const T>>>
-bool operator>=(const U& a, Span<T> b) {
-  return !(a < b);
-}
-template <
-    typename T, typename U,
-    typename = span_internal::EnableIfConvertibleTo<U, absl::Span<const T>>>
-bool operator>=(Span<T> a, const U& b) {
-  return !(a < b);
-}
-
-// MakeSpan()
-//
-// Constructs a mutable `Span<T>`, deducing `T` automatically from either a
-// container or pointer+size.
-//
-// Because a read-only `Span<const T>` is implicitly constructed from container
-// types regardless of whether the container itself is a const container,
-// constructing mutable spans of type `Span<T>` from containers requires
-// explicit constructors. The container-accepting version of `MakeSpan()`
-// deduces the type of `T` by the constness of the pointer received from the
-// container's `data()` member. Similarly, the pointer-accepting version returns
-// a `Span<const T>` if `T` is `const`, and a `Span<T>` otherwise.
-//
-// Examples:
-//
-//   void MyRoutine(absl::Span<MyComplicatedType> a) {
-//     ...
-//   };
-//   // my_vector is a container of non-const types
-//   std::vector<MyComplicatedType> my_vector;
-//
-//   // Constructing a Span implicitly attempts to create a Span of type
-//   // `Span<const T>`
-//   MyRoutine(my_vector);                // error, type mismatch
-//
-//   // Explicitly constructing the Span is verbose
-//   MyRoutine(absl::Span<MyComplicatedType>(my_vector));
-//
-//   // Use MakeSpan() to make an absl::Span<T>
-//   MyRoutine(absl::MakeSpan(my_vector));
-//
-//   // Construct a span from an array ptr+size
-//   absl::Span<T> my_span() {
-//     return absl::MakeSpan(&array[0], num_elements_);
-//   }
-//
-template <int&... ExplicitArgumentBarrier, typename T>
-constexpr Span<T> MakeSpan(absl::Nullable<T*> ptr, size_t size) noexcept {
-  return Span<T>(ptr, size);
-}
-
-template <int&... ExplicitArgumentBarrier, typename T>
-Span<T> MakeSpan(absl::Nullable<T*> begin, absl::Nullable<T*> end) noexcept {
-  return ABSL_HARDENING_ASSERT(begin <= end),
-         Span<T>(begin, static_cast<size_t>(end - begin));
-}
-
-template <int&... ExplicitArgumentBarrier, typename C>
-constexpr auto MakeSpan(C& c) noexcept  // NOLINT(runtime/references)
-    -> decltype(absl::MakeSpan(span_internal::GetData(c), c.size())) {
-  return MakeSpan(span_internal::GetData(c), c.size());
-}
-
-template <int&... ExplicitArgumentBarrier, typename T, size_t N>
-constexpr Span<T> MakeSpan(T (&array)[N]) noexcept {
-  return Span<T>(array, N);
-}
-
-// MakeConstSpan()
-//
-// Constructs a `Span<const T>` as with `MakeSpan`, deducing `T` automatically,
-// but always returning a `Span<const T>`.
-//
-// Examples:
-//
-//   void ProcessInts(absl::Span<const int> some_ints);
-//
-//   // Call with a pointer and size.
-//   int array[3] = { 0, 0, 0 };
-//   ProcessInts(absl::MakeConstSpan(&array[0], 3));
-//
-//   // Call with a [begin, end) pair.
-//   ProcessInts(absl::MakeConstSpan(&array[0], &array[3]));
-//
-//   // Call directly with an array.
-//   ProcessInts(absl::MakeConstSpan(array));
-//
-//   // Call with a contiguous container.
-//   std::vector<int> some_ints = ...;
-//   ProcessInts(absl::MakeConstSpan(some_ints));
-//   ProcessInts(absl::MakeConstSpan(std::vector<int>{ 0, 0, 0 }));
-//
-template <int&... ExplicitArgumentBarrier, typename T>
-constexpr Span<const T> MakeConstSpan(absl::Nullable<T*> ptr,
-                                      size_t size) noexcept {
-  return Span<const T>(ptr, size);
-}
-
-template <int&... ExplicitArgumentBarrier, typename T>
-Span<const T> MakeConstSpan(absl::Nullable<T*> begin,
-                            absl::Nullable<T*> end) noexcept {
-  return ABSL_HARDENING_ASSERT(begin <= end), Span<const T>(begin, end - begin);
-}
-
-template <int&... ExplicitArgumentBarrier, typename C>
-constexpr auto MakeConstSpan(const C& c) noexcept -> decltype(MakeSpan(c)) {
-  return MakeSpan(c);
-}
-
-template <int&... ExplicitArgumentBarrier, typename T, size_t N>
-constexpr Span<const T> MakeConstSpan(const T (&array)[N]) noexcept {
-  return Span<const T>(array, N);
-}
-ABSL_NAMESPACE_END
-}  // namespace absl
-#endif  // ABSL_TYPES_SPAN_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/types/variant.h b/projects/llm_framework/include/abseil-cpp/absl/types/variant.h
deleted file mode 100644
index 56a7e05e..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/types/variant.h
+++ /dev/null
@@ -1,861 +0,0 @@
-// Copyright 2018 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// -----------------------------------------------------------------------------
-// variant.h
-// -----------------------------------------------------------------------------
-//
-// This header file defines an `absl::variant` type for holding a type-safe
-// value of some prescribed set of types (noted as alternative types), and
-// associated functions for managing variants.
-//
-// The `absl::variant` type is a form of type-safe union. An `absl::variant`
-// should always hold a value of one of its alternative types (except in the
-// "valueless by exception state" -- see below). A default-constructed
-// `absl::variant` will hold the value of its first alternative type, provided
-// it is default-constructible.
-//
-// In exceptional cases due to error, an `absl::variant` can hold no
-// value (known as a "valueless by exception" state), though this is not the
-// norm.
-//
-// As with `absl::optional`, an `absl::variant` -- when it holds a value --
-// allocates a value of that type directly within the `variant` itself; it
-// cannot hold a reference, array, or the type `void`; it can, however, hold a
-// pointer to externally managed memory.
-//
-// `absl::variant` is a C++11 compatible version of the C++17 `std::variant`
-// abstraction and is designed to be a drop-in replacement for code compliant
-// with C++17.
-
-#ifndef ABSL_TYPES_VARIANT_H_
-#define ABSL_TYPES_VARIANT_H_
-
-#include "absl/base/config.h"
-#include "absl/utility/utility.h"
-
-#ifdef ABSL_USES_STD_VARIANT
-
-#include <variant>  // IWYU pragma: export
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-using std::bad_variant_access;
-using std::get;
-using std::get_if;
-using std::holds_alternative;
-using std::monostate;
-using std::variant;
-using std::variant_alternative;
-using std::variant_alternative_t;
-using std::variant_npos;
-using std::variant_size;
-using std::variant_size_v;
-using std::visit;
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#else  // ABSL_USES_STD_VARIANT
-
-#include <functional>
-#include <new>
-#include <type_traits>
-#include <utility>
-
-#include "absl/base/macros.h"
-#include "absl/base/port.h"
-#include "absl/meta/type_traits.h"
-#include "absl/types/internal/variant.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// -----------------------------------------------------------------------------
-// absl::variant
-// -----------------------------------------------------------------------------
-//
-// An `absl::variant` type is a form of type-safe union. An `absl::variant` --
-// except in exceptional cases -- always holds a value of one of its alternative
-// types.
-//
-// Example:
-//
-//   // Construct a variant that holds either an integer or a std::string and
-//   // assign it to a std::string.
-//   absl::variant<int, std::string> v = std::string("abc");
-//
-//   // A default-constructed variant will hold a value-initialized value of
-//   // the first alternative type.
-//   auto a = absl::variant<int, std::string>();   // Holds an int of value '0'.
-//
-//   // variants are assignable.
-//
-//   // copy assignment
-//   auto v1 = absl::variant<int, std::string>("abc");
-//   auto v2 = absl::variant<int, std::string>(10);
-//   v2 = v1;  // copy assign
-//
-//   // move assignment
-//   auto v1 = absl::variant<int, std::string>("abc");
-//   v1 = absl::variant<int, std::string>(10);
-//
-//   // assignment through type conversion
-//   a = 128;         // variant contains int
-//   a = "128";       // variant contains std::string
-//
-// An `absl::variant` holding a value of one of its alternative types `T` holds
-// an allocation of `T` directly within the variant itself. An `absl::variant`
-// is not allowed to allocate additional storage, such as dynamic memory, to
-// allocate the contained value. The contained value shall be allocated in a
-// region of the variant storage suitably aligned for all alternative types.
-template <typename... Ts>
-class variant;
-
-// swap()
-//
-// Swaps two `absl::variant` values. This function is equivalent to `v.swap(w)`
-// where `v` and `w` are `absl::variant` types.
-//
-// Note that this function requires all alternative types to be both swappable
-// and move-constructible, because any two variants may refer to either the same
-// type (in which case, they will be swapped) or to two different types (in
-// which case the values will need to be moved).
-//
-template <
-    typename... Ts,
-    absl::enable_if_t<
-        absl::conjunction<std::is_move_constructible<Ts>...,
-                          type_traits_internal::IsSwappable<Ts>...>::value,
-        int> = 0>
-void swap(variant<Ts...>& v, variant<Ts...>& w) noexcept(noexcept(v.swap(w))) {
-  v.swap(w);
-}
-
-// variant_size
-//
-// Returns the number of alternative types available for a given `absl::variant`
-// type as a compile-time constant expression. As this is a class template, it
-// is not generally useful for accessing the number of alternative types of
-// any given `absl::variant` instance.
-//
-// Example:
-//
-//   auto a = absl::variant<int, std::string>;
-//   constexpr int num_types =
-//       absl::variant_size<absl::variant<int, std::string>>();
-//
-//   // You can also use the member constant `value`.
-//   constexpr int num_types =
-//       absl::variant_size<absl::variant<int, std::string>>::value;
-//
-//   // `absl::variant_size` is more valuable for use in generic code:
-//   template <typename Variant>
-//   constexpr bool IsVariantMultivalue() {
-//       return absl::variant_size<Variant>() > 1;
-//   }
-//
-// Note that the set of cv-qualified specializations of `variant_size` are
-// provided to ensure that those specializations compile (especially when passed
-// within template logic).
-template <class T>
-struct variant_size;
-
-template <class... Ts>
-struct variant_size<variant<Ts...>>
-    : std::integral_constant<std::size_t, sizeof...(Ts)> {};
-
-// Specialization of `variant_size` for const qualified variants.
-template <class T>
-struct variant_size<const T> : variant_size<T>::type {};
-
-// Specialization of `variant_size` for volatile qualified variants.
-template <class T>
-struct variant_size<volatile T> : variant_size<T>::type {};
-
-// Specialization of `variant_size` for const volatile qualified variants.
-template <class T>
-struct variant_size<const volatile T> : variant_size<T>::type {};
-
-// variant_alternative
-//
-// Returns the alternative type for a given `absl::variant` at the passed
-// index value as a compile-time constant expression. As this is a class
-// template resulting in a type, it is not useful for access of the run-time
-// value of any given `absl::variant` variable.
-//
-// Example:
-//
-//   // The type of the 0th alternative is "int".
-//   using alternative_type_0
-//     = absl::variant_alternative<0, absl::variant<int, std::string>>::type;
-//
-//   static_assert(std::is_same<alternative_type_0, int>::value, "");
-//
-//   // `absl::variant_alternative` is more valuable for use in generic code:
-//   template <typename Variant>
-//   constexpr bool IsFirstElementTrivial() {
-//       return std::is_trivial_v<variant_alternative<0, Variant>::type>;
-//   }
-//
-// Note that the set of cv-qualified specializations of `variant_alternative`
-// are provided to ensure that those specializations compile (especially when
-// passed within template logic).
-template <std::size_t I, class T>
-struct variant_alternative;
-
-template <std::size_t I, class... Types>
-struct variant_alternative<I, variant<Types...>> {
-  using type =
-      variant_internal::VariantAlternativeSfinaeT<I, variant<Types...>>;
-};
-
-// Specialization of `variant_alternative` for const qualified variants.
-template <std::size_t I, class T>
-struct variant_alternative<I, const T> {
-  using type = const typename variant_alternative<I, T>::type;
-};
-
-// Specialization of `variant_alternative` for volatile qualified variants.
-template <std::size_t I, class T>
-struct variant_alternative<I, volatile T> {
-  using type = volatile typename variant_alternative<I, T>::type;
-};
-
-// Specialization of `variant_alternative` for const volatile qualified
-// variants.
-template <std::size_t I, class T>
-struct variant_alternative<I, const volatile T> {
-  using type = const volatile typename variant_alternative<I, T>::type;
-};
-
-// Template type alias for variant_alternative<I, T>::type.
-//
-// Example:
-//
-//   using alternative_type_0
-//     = absl::variant_alternative_t<0, absl::variant<int, std::string>>;
-//   static_assert(std::is_same<alternative_type_0, int>::value, "");
-template <std::size_t I, class T>
-using variant_alternative_t = typename variant_alternative<I, T>::type;
-
-// holds_alternative()
-//
-// Checks whether the given variant currently holds a given alternative type,
-// returning `true` if so.
-//
-// Example:
-//
-//   absl::variant<int, std::string> foo = 42;
-//   if (absl::holds_alternative<int>(foo)) {
-//       std::cout << "The variant holds an integer";
-//   }
-template <class T, class... Types>
-constexpr bool holds_alternative(const variant<Types...>& v) noexcept {
-  static_assert(
-      variant_internal::UnambiguousIndexOfImpl<variant<Types...>, T,
-                                               0>::value != sizeof...(Types),
-      "The type T must occur exactly once in Types...");
-  return v.index() ==
-         variant_internal::UnambiguousIndexOf<variant<Types...>, T>::value;
-}
-
-// get()
-//
-// Returns a reference to the value currently within a given variant, using
-// either a unique alternative type amongst the variant's set of alternative
-// types, or the variant's index value. Attempting to get a variant's value
-// using a type that is not unique within the variant's set of alternative types
-// is a compile-time error. If the index of the alternative being specified is
-// different from the index of the alternative that is currently stored, throws
-// `absl::bad_variant_access`.
-//
-// Example:
-//
-//   auto a = absl::variant<int, std::string>;
-//
-//   // Get the value by type (if unique).
-//   int i = absl::get<int>(a);
-//
-//   auto b = absl::variant<int, int>;
-//
-//   // Getting the value by a type that is not unique is ill-formed.
-//   int j = absl::get<int>(b);     // Compile Error!
-//
-//   // Getting value by index not ambiguous and allowed.
-//   int k = absl::get<1>(b);
-
-// Overload for getting a variant's lvalue by type.
-template <class T, class... Types>
-constexpr T& get(variant<Types...>& v) {  // NOLINT
-  return variant_internal::VariantCoreAccess::CheckedAccess<
-      variant_internal::IndexOf<T, Types...>::value>(v);
-}
-
-// Overload for getting a variant's rvalue by type.
-template <class T, class... Types>
-constexpr T&& get(variant<Types...>&& v) {
-  return variant_internal::VariantCoreAccess::CheckedAccess<
-      variant_internal::IndexOf<T, Types...>::value>(std::move(v));
-}
-
-// Overload for getting a variant's const lvalue by type.
-template <class T, class... Types>
-constexpr const T& get(const variant<Types...>& v) {
-  return variant_internal::VariantCoreAccess::CheckedAccess<
-      variant_internal::IndexOf<T, Types...>::value>(v);
-}
-
-// Overload for getting a variant's const rvalue by type.
-template <class T, class... Types>
-constexpr const T&& get(const variant<Types...>&& v) {
-  return variant_internal::VariantCoreAccess::CheckedAccess<
-      variant_internal::IndexOf<T, Types...>::value>(std::move(v));
-}
-
-// Overload for getting a variant's lvalue by index.
-template <std::size_t I, class... Types>
-constexpr variant_alternative_t<I, variant<Types...>>& get(
-    variant<Types...>& v) {  // NOLINT
-  return variant_internal::VariantCoreAccess::CheckedAccess<I>(v);
-}
-
-// Overload for getting a variant's rvalue by index.
-template <std::size_t I, class... Types>
-constexpr variant_alternative_t<I, variant<Types...>>&& get(
-    variant<Types...>&& v) {
-  return variant_internal::VariantCoreAccess::CheckedAccess<I>(std::move(v));
-}
-
-// Overload for getting a variant's const lvalue by index.
-template <std::size_t I, class... Types>
-constexpr const variant_alternative_t<I, variant<Types...>>& get(
-    const variant<Types...>& v) {
-  return variant_internal::VariantCoreAccess::CheckedAccess<I>(v);
-}
-
-// Overload for getting a variant's const rvalue by index.
-template <std::size_t I, class... Types>
-constexpr const variant_alternative_t<I, variant<Types...>>&& get(
-    const variant<Types...>&& v) {
-  return variant_internal::VariantCoreAccess::CheckedAccess<I>(std::move(v));
-}
-
-// get_if()
-//
-// Returns a pointer to the value currently stored within a given variant, if
-// present, using either a unique alternative type amongst the variant's set of
-// alternative types, or the variant's index value. If such a value does not
-// exist, returns `nullptr`.
-//
-// As with `get`, attempting to get a variant's value using a type that is not
-// unique within the variant's set of alternative types is a compile-time error.
-
-// Overload for getting a pointer to the value stored in the given variant by
-// index.
-template <std::size_t I, class... Types>
-constexpr absl::add_pointer_t<variant_alternative_t<I, variant<Types...>>>
-get_if(variant<Types...>* v) noexcept {
-  return (v != nullptr && v->index() == I)
-             ? std::addressof(
-                   variant_internal::VariantCoreAccess::Access<I>(*v))
-             : nullptr;
-}
-
-// Overload for getting a pointer to the const value stored in the given
-// variant by index.
-template <std::size_t I, class... Types>
-constexpr absl::add_pointer_t<const variant_alternative_t<I, variant<Types...>>>
-get_if(const variant<Types...>* v) noexcept {
-  return (v != nullptr && v->index() == I)
-             ? std::addressof(
-                   variant_internal::VariantCoreAccess::Access<I>(*v))
-             : nullptr;
-}
-
-// Overload for getting a pointer to the value stored in the given variant by
-// type.
-template <class T, class... Types>
-constexpr absl::add_pointer_t<T> get_if(variant<Types...>* v) noexcept {
-  return absl::get_if<variant_internal::IndexOf<T, Types...>::value>(v);
-}
-
-// Overload for getting a pointer to the const value stored in the given variant
-// by type.
-template <class T, class... Types>
-constexpr absl::add_pointer_t<const T> get_if(
-    const variant<Types...>* v) noexcept {
-  return absl::get_if<variant_internal::IndexOf<T, Types...>::value>(v);
-}
-
-// visit()
-//
-// Calls a provided functor on a given set of variants. `absl::visit()` is
-// commonly used to conditionally inspect the state of a given variant (or set
-// of variants).
-//
-// The functor must return the same type when called with any of the variants'
-// alternatives.
-//
-// Example:
-//
-//   // Define a visitor functor
-//   struct GetVariant {
-//       template<typename T>
-//       void operator()(const T& i) const {
-//         std::cout << "The variant's value is: " << i;
-//       }
-//   };
-//
-//   // Declare our variant, and call `absl::visit()` on it.
-//   // Note that `GetVariant()` returns void in either case.
-//   absl::variant<int, std::string> foo = std::string("foo");
-//   GetVariant visitor;
-//   absl::visit(visitor, foo);  // Prints `The variant's value is: foo'
-template <typename Visitor, typename... Variants>
-variant_internal::VisitResult<Visitor, Variants...> visit(Visitor&& vis,
-                                                          Variants&&... vars) {
-  return variant_internal::
-      VisitIndices<variant_size<absl::decay_t<Variants> >::value...>::Run(
-          variant_internal::PerformVisitation<Visitor, Variants...>{
-              std::forward_as_tuple(std::forward<Variants>(vars)...),
-              std::forward<Visitor>(vis)},
-          vars.index()...);
-}
-
-// monostate
-//
-// The monostate class serves as a first alternative type for a variant for
-// which the first variant type is otherwise not default-constructible.
-struct monostate {};
-
-// `absl::monostate` Relational Operators
-
-constexpr bool operator<(monostate, monostate) noexcept { return false; }
-constexpr bool operator>(monostate, monostate) noexcept { return false; }
-constexpr bool operator<=(monostate, monostate) noexcept { return true; }
-constexpr bool operator>=(monostate, monostate) noexcept { return true; }
-constexpr bool operator==(monostate, monostate) noexcept { return true; }
-constexpr bool operator!=(monostate, monostate) noexcept { return false; }
-
-
-//------------------------------------------------------------------------------
-// `absl::variant` Template Definition
-//------------------------------------------------------------------------------
-template <typename T0, typename... Tn>
-class variant<T0, Tn...> : private variant_internal::VariantBase<T0, Tn...> {
-  static_assert(absl::conjunction<std::is_object<T0>,
-                                  std::is_object<Tn>...>::value,
-                "Attempted to instantiate a variant containing a non-object "
-                "type.");
-  // Intentionally not qualifying `negation` with `absl::` to work around a bug
-  // in MSVC 2015 with inline namespace and variadic template.
-  static_assert(absl::conjunction<negation<std::is_array<T0> >,
-                                  negation<std::is_array<Tn> >...>::value,
-                "Attempted to instantiate a variant containing an array type.");
-  static_assert(absl::conjunction<std::is_nothrow_destructible<T0>,
-                                  std::is_nothrow_destructible<Tn>...>::value,
-                "Attempted to instantiate a variant containing a non-nothrow "
-                "destructible type.");
-
-  friend struct variant_internal::VariantCoreAccess;
-
- private:
-  using Base = variant_internal::VariantBase<T0, Tn...>;
-
- public:
-  // Constructors
-
-  // Constructs a variant holding a default-initialized value of the first
-  // alternative type.
-  constexpr variant() /*noexcept(see 111above)*/ = default;
-
-  // Copy constructor, standard semantics
-  variant(const variant& other) = default;
-
-  // Move constructor, standard semantics
-  variant(variant&& other) /*noexcept(see above)*/ = default;
-
-  // Constructs a variant of an alternative type specified by overload
-  // resolution of the provided forwarding arguments through
-  // direct-initialization.
-  //
-  // Note: If the selected constructor is a constexpr constructor, this
-  // constructor shall be a constexpr constructor.
-  //
-  // NOTE: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0608r1.html
-  // has been voted passed the design phase in the C++ standard meeting in Mar
-  // 2018. It will be implemented and integrated into `absl::variant`.
-  template <
-      class T,
-      std::size_t I = std::enable_if<
-          variant_internal::IsNeitherSelfNorInPlace<variant,
-                                                    absl::decay_t<T> >::value,
-          variant_internal::IndexOfConstructedType<variant, T> >::type::value,
-      class Tj = absl::variant_alternative_t<I, variant>,
-      absl::enable_if_t<std::is_constructible<Tj, T>::value>* = nullptr>
-  constexpr variant(T&& t) noexcept(std::is_nothrow_constructible<Tj, T>::value)
-      : Base(variant_internal::EmplaceTag<I>(), std::forward<T>(t)) {}
-
-  // Constructs a variant of an alternative type from the arguments through
-  // direct-initialization.
-  //
-  // Note: If the selected constructor is a constexpr constructor, this
-  // constructor shall be a constexpr constructor.
-  template <class T, class... Args,
-            typename std::enable_if<std::is_constructible<
-                variant_internal::UnambiguousTypeOfT<variant, T>,
-                Args...>::value>::type* = nullptr>
-  constexpr explicit variant(in_place_type_t<T>, Args&&... args)
-      : Base(variant_internal::EmplaceTag<
-                 variant_internal::UnambiguousIndexOf<variant, T>::value>(),
-             std::forward<Args>(args)...) {}
-
-  // Constructs a variant of an alternative type from an initializer list
-  // and other arguments through direct-initialization.
-  //
-  // Note: If the selected constructor is a constexpr constructor, this
-  // constructor shall be a constexpr constructor.
-  template <class T, class U, class... Args,
-            typename std::enable_if<std::is_constructible<
-                variant_internal::UnambiguousTypeOfT<variant, T>,
-                std::initializer_list<U>&, Args...>::value>::type* = nullptr>
-  constexpr explicit variant(in_place_type_t<T>, std::initializer_list<U> il,
-                             Args&&... args)
-      : Base(variant_internal::EmplaceTag<
-                 variant_internal::UnambiguousIndexOf<variant, T>::value>(),
-             il, std::forward<Args>(args)...) {}
-
-  // Constructs a variant of an alternative type from a provided index,
-  // through value-initialization using the provided forwarded arguments.
-  template <std::size_t I, class... Args,
-            typename std::enable_if<std::is_constructible<
-                variant_internal::VariantAlternativeSfinaeT<I, variant>,
-                Args...>::value>::type* = nullptr>
-  constexpr explicit variant(in_place_index_t<I>, Args&&... args)
-      : Base(variant_internal::EmplaceTag<I>(), std::forward<Args>(args)...) {}
-
-  // Constructs a variant of an alternative type from a provided index,
-  // through value-initialization of an initializer list and the provided
-  // forwarded arguments.
-  template <std::size_t I, class U, class... Args,
-            typename std::enable_if<std::is_constructible<
-                variant_internal::VariantAlternativeSfinaeT<I, variant>,
-                std::initializer_list<U>&, Args...>::value>::type* = nullptr>
-  constexpr explicit variant(in_place_index_t<I>, std::initializer_list<U> il,
-                             Args&&... args)
-      : Base(variant_internal::EmplaceTag<I>(), il,
-             std::forward<Args>(args)...) {}
-
-  // Destructors
-
-  // Destroys the variant's currently contained value, provided that
-  // `absl::valueless_by_exception()` is false.
-  ~variant() = default;
-
-  // Assignment Operators
-
-  // Copy assignment operator
-  variant& operator=(const variant& other) = default;
-
-  // Move assignment operator
-  variant& operator=(variant&& other) /*noexcept(see above)*/ = default;
-
-  // Converting assignment operator
-  //
-  // NOTE: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0608r1.html
-  // has been voted passed the design phase in the C++ standard meeting in Mar
-  // 2018. It will be implemented and integrated into `absl::variant`.
-  template <
-      class T,
-      std::size_t I = std::enable_if<
-          !std::is_same<absl::decay_t<T>, variant>::value,
-          variant_internal::IndexOfConstructedType<variant, T>>::type::value,
-      class Tj = absl::variant_alternative_t<I, variant>,
-      typename std::enable_if<std::is_assignable<Tj&, T>::value &&
-                              std::is_constructible<Tj, T>::value>::type* =
-          nullptr>
-  variant& operator=(T&& t) noexcept(
-      std::is_nothrow_assignable<Tj&, T>::value&&
-          std::is_nothrow_constructible<Tj, T>::value) {
-    variant_internal::VisitIndices<sizeof...(Tn) + 1>::Run(
-        variant_internal::VariantCoreAccess::MakeConversionAssignVisitor(
-            this, std::forward<T>(t)),
-        index());
-
-    return *this;
-  }
-
-
-  // emplace() Functions
-
-  // Constructs a value of the given alternative type T within the variant. The
-  // existing value of the variant is destroyed first (provided that
-  // `absl::valueless_by_exception()` is false). Requires that T is unambiguous
-  // in the variant.
-  //
-  // Example:
-  //
-  //   absl::variant<std::vector<int>, int, std::string> v;
-  //   v.emplace<int>(99);
-  //   v.emplace<std::string>("abc");
-  template <
-      class T, class... Args,
-      typename std::enable_if<std::is_constructible<
-          absl::variant_alternative_t<
-              variant_internal::UnambiguousIndexOf<variant, T>::value, variant>,
-          Args...>::value>::type* = nullptr>
-  T& emplace(Args&&... args) {
-    return variant_internal::VariantCoreAccess::Replace<
-        variant_internal::UnambiguousIndexOf<variant, T>::value>(
-        this, std::forward<Args>(args)...);
-  }
-
-  // Constructs a value of the given alternative type T within the variant using
-  // an initializer list. The existing value of the variant is destroyed first
-  // (provided that `absl::valueless_by_exception()` is false). Requires that T
-  // is unambiguous in the variant.
-  //
-  // Example:
-  //
-  //   absl::variant<std::vector<int>, int, std::string> v;
-  //   v.emplace<std::vector<int>>({0, 1, 2});
-  template <
-      class T, class U, class... Args,
-      typename std::enable_if<std::is_constructible<
-          absl::variant_alternative_t<
-              variant_internal::UnambiguousIndexOf<variant, T>::value, variant>,
-          std::initializer_list<U>&, Args...>::value>::type* = nullptr>
-  T& emplace(std::initializer_list<U> il, Args&&... args) {
-    return variant_internal::VariantCoreAccess::Replace<
-        variant_internal::UnambiguousIndexOf<variant, T>::value>(
-        this, il, std::forward<Args>(args)...);
-  }
-
-  // Destroys the current value of the variant (provided that
-  // `absl::valueless_by_exception()` is false) and constructs a new value at
-  // the given index.
-  //
-  // Example:
-  //
-  //   absl::variant<std::vector<int>, int, int> v;
-  //   v.emplace<1>(99);
-  //   v.emplace<2>(98);
-  //   v.emplace<int>(99);  // Won't compile. 'int' isn't a unique type.
-  template <std::size_t I, class... Args,
-            typename std::enable_if<
-                std::is_constructible<absl::variant_alternative_t<I, variant>,
-                                      Args...>::value>::type* = nullptr>
-  absl::variant_alternative_t<I, variant>& emplace(Args&&... args) {
-    return variant_internal::VariantCoreAccess::Replace<I>(
-        this, std::forward<Args>(args)...);
-  }
-
-  // Destroys the current value of the variant (provided that
-  // `absl::valueless_by_exception()` is false) and constructs a new value at
-  // the given index using an initializer list and the provided arguments.
-  //
-  // Example:
-  //
-  //   absl::variant<std::vector<int>, int, int> v;
-  //   v.emplace<0>({0, 1, 2});
-  template <std::size_t I, class U, class... Args,
-            typename std::enable_if<std::is_constructible<
-                absl::variant_alternative_t<I, variant>,
-                std::initializer_list<U>&, Args...>::value>::type* = nullptr>
-  absl::variant_alternative_t<I, variant>& emplace(std::initializer_list<U> il,
-                                                   Args&&... args) {
-    return variant_internal::VariantCoreAccess::Replace<I>(
-        this, il, std::forward<Args>(args)...);
-  }
-
-  // variant::valueless_by_exception()
-  //
-  // Returns false if and only if the variant currently holds a valid value.
-  constexpr bool valueless_by_exception() const noexcept {
-    return this->index_ == absl::variant_npos;
-  }
-
-  // variant::index()
-  //
-  // Returns the index value of the variant's currently selected alternative
-  // type.
-  constexpr std::size_t index() const noexcept { return this->index_; }
-
-  // variant::swap()
-  //
-  // Swaps the values of two variant objects.
-  //
-  void swap(variant& rhs) noexcept(
-      absl::conjunction<
-          std::is_nothrow_move_constructible<T0>,
-          std::is_nothrow_move_constructible<Tn>...,
-          type_traits_internal::IsNothrowSwappable<T0>,
-          type_traits_internal::IsNothrowSwappable<Tn>...>::value) {
-    return variant_internal::VisitIndices<sizeof...(Tn) + 1>::Run(
-        variant_internal::Swap<T0, Tn...>{this, &rhs}, rhs.index());
-  }
-};
-
-// We need a valid declaration of variant<> for SFINAE and overload resolution
-// to work properly above, but we don't need a full declaration since this type
-// will never be constructed. This declaration, though incomplete, suffices.
-template <>
-class variant<>;
-
-//------------------------------------------------------------------------------
-// Relational Operators
-//------------------------------------------------------------------------------
-//
-// If neither operand is in the `variant::valueless_by_exception` state:
-//
-//   * If the index of both variants is the same, the relational operator
-//     returns the result of the corresponding relational operator for the
-//     corresponding alternative type.
-//   * If the index of both variants is not the same, the relational operator
-//     returns the result of that operation applied to the value of the left
-//     operand's index and the value of the right operand's index.
-//   * If at least one operand is in the valueless_by_exception state:
-//     - A variant in the valueless_by_exception state is only considered equal
-//       to another variant in the valueless_by_exception state.
-//     - If exactly one operand is in the valueless_by_exception state, the
-//       variant in the valueless_by_exception state is less than the variant
-//       that is not in the valueless_by_exception state.
-//
-// Note: The value 1 is added to each index in the relational comparisons such
-// that the index corresponding to the valueless_by_exception state wraps around
-// to 0 (the lowest value for the index type), and the remaining indices stay in
-// the same relative order.
-
-// Equal-to operator
-template <typename... Types>
-constexpr variant_internal::RequireAllHaveEqualT<Types...> operator==(
-    const variant<Types...>& a, const variant<Types...>& b) {
-  return (a.index() == b.index()) &&
-         variant_internal::VisitIndices<sizeof...(Types)>::Run(
-             variant_internal::EqualsOp<Types...>{&a, &b}, a.index());
-}
-
-// Not equal operator
-template <typename... Types>
-constexpr variant_internal::RequireAllHaveNotEqualT<Types...> operator!=(
-    const variant<Types...>& a, const variant<Types...>& b) {
-  return (a.index() != b.index()) ||
-         variant_internal::VisitIndices<sizeof...(Types)>::Run(
-             variant_internal::NotEqualsOp<Types...>{&a, &b}, a.index());
-}
-
-// Less-than operator
-template <typename... Types>
-constexpr variant_internal::RequireAllHaveLessThanT<Types...> operator<(
-    const variant<Types...>& a, const variant<Types...>& b) {
-  return (a.index() != b.index())
-             ? (a.index() + 1) < (b.index() + 1)
-             : variant_internal::VisitIndices<sizeof...(Types)>::Run(
-                   variant_internal::LessThanOp<Types...>{&a, &b}, a.index());
-}
-
-// Greater-than operator
-template <typename... Types>
-constexpr variant_internal::RequireAllHaveGreaterThanT<Types...> operator>(
-    const variant<Types...>& a, const variant<Types...>& b) {
-  return (a.index() != b.index())
-             ? (a.index() + 1) > (b.index() + 1)
-             : variant_internal::VisitIndices<sizeof...(Types)>::Run(
-                   variant_internal::GreaterThanOp<Types...>{&a, &b},
-                   a.index());
-}
-
-// Less-than or equal-to operator
-template <typename... Types>
-constexpr variant_internal::RequireAllHaveLessThanOrEqualT<Types...> operator<=(
-    const variant<Types...>& a, const variant<Types...>& b) {
-  return (a.index() != b.index())
-             ? (a.index() + 1) < (b.index() + 1)
-             : variant_internal::VisitIndices<sizeof...(Types)>::Run(
-                   variant_internal::LessThanOrEqualsOp<Types...>{&a, &b},
-                   a.index());
-}
-
-// Greater-than or equal-to operator
-template <typename... Types>
-constexpr variant_internal::RequireAllHaveGreaterThanOrEqualT<Types...>
-operator>=(const variant<Types...>& a, const variant<Types...>& b) {
-  return (a.index() != b.index())
-             ? (a.index() + 1) > (b.index() + 1)
-             : variant_internal::VisitIndices<sizeof...(Types)>::Run(
-                   variant_internal::GreaterThanOrEqualsOp<Types...>{&a, &b},
-                   a.index());
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-namespace std {
-
-// hash()
-template <>  // NOLINT
-struct hash<absl::monostate> {
-  std::size_t operator()(absl::monostate) const { return 0; }
-};
-
-template <class... T>  // NOLINT
-struct hash<absl::variant<T...>>
-    : absl::variant_internal::VariantHashBase<absl::variant<T...>, void,
-                                              absl::remove_const_t<T>...> {};
-
-}  // namespace std
-
-#endif  // ABSL_USES_STD_VARIANT
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-namespace variant_internal {
-
-// Helper visitor for converting a variant<Ts...>` into another type (mostly
-// variant) that can be constructed from any type.
-template <typename To>
-struct ConversionVisitor {
-  template <typename T>
-  To operator()(T&& v) const {
-    return To(std::forward<T>(v));
-  }
-};
-
-}  // namespace variant_internal
-
-// ConvertVariantTo()
-//
-// Helper functions to convert an `absl::variant` to a variant of another set of
-// types, provided that the alternative type of the new variant type can be
-// converted from any type in the source variant.
-//
-// Example:
-//
-//   absl::variant<name1, name2, float> InternalReq(const Req&);
-//
-//   // name1 and name2 are convertible to name
-//   absl::variant<name, float> ExternalReq(const Req& req) {
-//     return absl::ConvertVariantTo<absl::variant<name, float>>(
-//              InternalReq(req));
-//   }
-template <typename To, typename Variant>
-To ConvertVariantTo(Variant&& variant) {
-  return absl::visit(variant_internal::ConversionVisitor<To>{},
-                     std::forward<Variant>(variant));
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_TYPES_VARIANT_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/utility/internal/if_constexpr.h b/projects/llm_framework/include/abseil-cpp/absl/utility/internal/if_constexpr.h
deleted file mode 100644
index 7a26311d..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/utility/internal/if_constexpr.h
+++ /dev/null
@@ -1,70 +0,0 @@
-// Copyright 2023 The Abseil Authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-// The IfConstexpr and IfConstexprElse utilities in this file are meant to be
-// used to emulate `if constexpr` in pre-C++17 mode in library implementation.
-// The motivation is to allow for avoiding complex SFINAE.
-//
-// The functions passed in must depend on the type(s) of the object(s) that
-// require SFINAE. For example:
-// template<typename T>
-// int MaybeFoo(T& t) {
-//   if constexpr (HasFoo<T>::value) return t.foo();
-//   return 0;
-// }
-//
-// can be written in pre-C++17 as:
-//
-// template<typename T>
-// int MaybeFoo(T& t) {
-//   int i = 0;
-//   absl::utility_internal::IfConstexpr<HasFoo<T>::value>(
-//       [&](const auto& fooer) { i = fooer.foo(); }, t);
-//   return i;
-// }
-
-#ifndef ABSL_UTILITY_INTERNAL_IF_CONSTEXPR_H_
-#define ABSL_UTILITY_INTERNAL_IF_CONSTEXPR_H_
-
-#include <tuple>
-#include <utility>
-
-#include "absl/base/config.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-namespace utility_internal {
-
-template <bool condition, typename TrueFunc, typename FalseFunc,
-          typename... Args>
-auto IfConstexprElse(TrueFunc&& true_func, FalseFunc&& false_func,
-                     Args&&... args) {
-  return std::get<condition>(std::forward_as_tuple(
-      std::forward<FalseFunc>(false_func), std::forward<TrueFunc>(true_func)))(
-      std::forward<Args>(args)...);
-}
-
-template <bool condition, typename Func, typename... Args>
-void IfConstexpr(Func&& func, Args&&... args) {
-  IfConstexprElse<condition>(std::forward<Func>(func), [](auto&&...){},
-                             std::forward<Args>(args)...);
-}
-
-}  // namespace utility_internal
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_UTILITY_INTERNAL_IF_CONSTEXPR_H_
diff --git a/projects/llm_framework/include/abseil-cpp/absl/utility/utility.h b/projects/llm_framework/include/abseil-cpp/absl/utility/utility.h
deleted file mode 100644
index fc0d1f65..00000000
--- a/projects/llm_framework/include/abseil-cpp/absl/utility/utility.h
+++ /dev/null
@@ -1,268 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// This header file contains C++14 versions of standard <utility> header
-// abstractions available within C++17, and are designed to be drop-in
-// replacement for code compliant with C++14 and C++17.
-//
-// The following abstractions are defined:
-//
-//   * apply<Functor, Tuple>         == std::apply<Functor, Tuple>
-//   * exchange<T>                   == std::exchange<T>
-//   * make_from_tuple<T>            == std::make_from_tuple<T>
-//
-// This header file also provides the tag types `in_place_t`, `in_place_type_t`,
-// and `in_place_index_t`, as well as the constant `in_place`, and
-// `constexpr` `std::move()` and `std::forward()` implementations in C++11.
-//
-// References:
-//
-//  https://en.cppreference.com/w/cpp/utility/apply
-//  http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2013/n3658.html
-
-#ifndef ABSL_UTILITY_UTILITY_H_
-#define ABSL_UTILITY_UTILITY_H_
-
-#include <cstddef>
-#include <cstdlib>
-#include <tuple>
-#include <utility>
-
-#include "absl/base/config.h"
-#include "absl/base/internal/inline_variable.h"
-#include "absl/base/internal/invoke.h"
-#include "absl/meta/type_traits.h"
-
-namespace absl {
-ABSL_NAMESPACE_BEGIN
-
-// Historical note: Abseil once provided implementations of these
-// abstractions for platforms that had not yet provided them. Those
-// platforms are no longer supported. New code should simply use the
-// the ones from std directly.
-using std::index_sequence;
-using std::index_sequence_for;
-using std::integer_sequence;
-using std::make_index_sequence;
-using std::make_integer_sequence;
-
-namespace utility_internal {
-
-template <typename T>
-struct InPlaceTypeTag {
-  explicit InPlaceTypeTag() = delete;
-  InPlaceTypeTag(const InPlaceTypeTag&) = delete;
-  InPlaceTypeTag& operator=(const InPlaceTypeTag&) = delete;
-};
-
-template <size_t I>
-struct InPlaceIndexTag {
-  explicit InPlaceIndexTag() = delete;
-  InPlaceIndexTag(const InPlaceIndexTag&) = delete;
-  InPlaceIndexTag& operator=(const InPlaceIndexTag&) = delete;
-};
-
-}  // namespace utility_internal
-
-// Tag types
-
-#ifdef ABSL_USES_STD_OPTIONAL
-
-using std::in_place_t;
-using std::in_place;
-
-#else  // ABSL_USES_STD_OPTIONAL
-
-// in_place_t
-//
-// Tag type used to specify in-place construction, such as with
-// `absl::optional`, designed to be a drop-in replacement for C++17's
-// `std::in_place_t`.
-struct in_place_t {};
-
-ABSL_INTERNAL_INLINE_CONSTEXPR(in_place_t, in_place, {});
-
-#endif  // ABSL_USES_STD_OPTIONAL
-
-#if defined(ABSL_USES_STD_ANY) || defined(ABSL_USES_STD_VARIANT)
-using std::in_place_type;
-using std::in_place_type_t;
-#else
-
-// in_place_type_t
-//
-// Tag type used for in-place construction when the type to construct needs to
-// be specified, such as with `absl::any`, designed to be a drop-in replacement
-// for C++17's `std::in_place_type_t`.
-template <typename T>
-using in_place_type_t = void (*)(utility_internal::InPlaceTypeTag<T>);
-
-template <typename T>
-void in_place_type(utility_internal::InPlaceTypeTag<T>) {}
-#endif  // ABSL_USES_STD_ANY || ABSL_USES_STD_VARIANT
-
-#ifdef ABSL_USES_STD_VARIANT
-using std::in_place_index;
-using std::in_place_index_t;
-#else
-
-// in_place_index_t
-//
-// Tag type used for in-place construction when the type to construct needs to
-// be specified, such as with `absl::any`, designed to be a drop-in replacement
-// for C++17's `std::in_place_index_t`.
-template <size_t I>
-using in_place_index_t = void (*)(utility_internal::InPlaceIndexTag<I>);
-
-template <size_t I>
-void in_place_index(utility_internal::InPlaceIndexTag<I>) {}
-#endif  // ABSL_USES_STD_VARIANT
-
-// Constexpr move and forward
-
-// move()
-//
-// A constexpr version of `std::move()`, designed to be a drop-in replacement
-// for C++14's `std::move()`.
-template <typename T>
-constexpr absl::remove_reference_t<T>&& move(T&& t) noexcept {
-  return static_cast<absl::remove_reference_t<T>&&>(t);
-}
-
-// forward()
-//
-// A constexpr version of `std::forward()`, designed to be a drop-in replacement
-// for C++14's `std::forward()`.
-template <typename T>
-constexpr T&& forward(
-    absl::remove_reference_t<T>& t) noexcept {  // NOLINT(runtime/references)
-  return static_cast<T&&>(t);
-}
-
-namespace utility_internal {
-// Helper method for expanding tuple into a called method.
-template <typename Functor, typename Tuple, std::size_t... Indexes>
-auto apply_helper(Functor&& functor, Tuple&& t, index_sequence<Indexes...>)
-    -> decltype(absl::base_internal::invoke(
-        absl::forward<Functor>(functor),
-        std::get<Indexes>(absl::forward<Tuple>(t))...)) {
-  return absl::base_internal::invoke(
-      absl::forward<Functor>(functor),
-      std::get<Indexes>(absl::forward<Tuple>(t))...);
-}
-
-}  // namespace utility_internal
-
-// apply
-//
-// Invokes a Callable using elements of a tuple as its arguments.
-// Each element of the tuple corresponds to an argument of the call (in order).
-// Both the Callable argument and the tuple argument are perfect-forwarded.
-// For member-function Callables, the first tuple element acts as the `this`
-// pointer. `absl::apply` is designed to be a drop-in replacement for C++17's
-// `std::apply`. Unlike C++17's `std::apply`, this is not currently `constexpr`.
-//
-// Example:
-//
-//   class Foo {
-//    public:
-//     void Bar(int);
-//   };
-//   void user_function1(int, std::string);
-//   void user_function2(std::unique_ptr<Foo>);
-//   auto user_lambda = [](int, int) {};
-//
-//   int main()
-//   {
-//       std::tuple<int, std::string> tuple1(42, "bar");
-//       // Invokes the first user function on int, std::string.
-//       absl::apply(&user_function1, tuple1);
-//
-//       std::tuple<std::unique_ptr<Foo>> tuple2(absl::make_unique<Foo>());
-//       // Invokes the user function that takes ownership of the unique
-//       // pointer.
-//       absl::apply(&user_function2, std::move(tuple2));
-//
-//       auto foo = absl::make_unique<Foo>();
-//       std::tuple<Foo*, int> tuple3(foo.get(), 42);
-//       // Invokes the method Bar on foo with one argument, 42.
-//       absl::apply(&Foo::Bar, tuple3);
-//
-//       std::tuple<int, int> tuple4(8, 9);
-//       // Invokes a lambda.
-//       absl::apply(user_lambda, tuple4);
-//   }
-template <typename Functor, typename Tuple>
-auto apply(Functor&& functor, Tuple&& t)
-    -> decltype(utility_internal::apply_helper(
-        absl::forward<Functor>(functor), absl::forward<Tuple>(t),
-        absl::make_index_sequence<std::tuple_size<
-            typename std::remove_reference<Tuple>::type>::value>{})) {
-  return utility_internal::apply_helper(
-      absl::forward<Functor>(functor), absl::forward<Tuple>(t),
-      absl::make_index_sequence<std::tuple_size<
-          typename std::remove_reference<Tuple>::type>::value>{});
-}
-
-// exchange
-//
-// Replaces the value of `obj` with `new_value` and returns the old value of
-// `obj`.  `absl::exchange` is designed to be a drop-in replacement for C++14's
-// `std::exchange`.
-//
-// Example:
-//
-//   Foo& operator=(Foo&& other) {
-//     ptr1_ = absl::exchange(other.ptr1_, nullptr);
-//     int1_ = absl::exchange(other.int1_, -1);
-//     return *this;
-//   }
-template <typename T, typename U = T>
-T exchange(T& obj, U&& new_value) {
-  T old_value = absl::move(obj);
-  obj = absl::forward<U>(new_value);
-  return old_value;
-}
-
-namespace utility_internal {
-template <typename T, typename Tuple, size_t... I>
-T make_from_tuple_impl(Tuple&& tup, absl::index_sequence<I...>) {
-  return T(std::get<I>(std::forward<Tuple>(tup))...);
-}
-}  // namespace utility_internal
-
-// make_from_tuple
-//
-// Given the template parameter type `T` and a tuple of arguments
-// `std::tuple(arg0, arg1, ..., argN)` constructs an object of type `T` as if by
-// calling `T(arg0, arg1, ..., argN)`.
-//
-// Example:
-//
-//   std::tuple<const char*, size_t> args("hello world", 5);
-//   auto s = absl::make_from_tuple<std::string>(args);
-//   assert(s == "hello");
-//
-template <typename T, typename Tuple>
-constexpr T make_from_tuple(Tuple&& tup) {
-  return utility_internal::make_from_tuple_impl<T>(
-      std::forward<Tuple>(tup),
-      absl::make_index_sequence<
-          std::tuple_size<absl::decay_t<Tuple>>::value>{});
-}
-
-ABSL_NAMESPACE_END
-}  // namespace absl
-
-#endif  // ABSL_UTILITY_UTILITY_H_
diff --git a/projects/llm_framework/include/abseil-cpp/ci/absl_alternate_options.h b/projects/llm_framework/include/abseil-cpp/ci/absl_alternate_options.h
deleted file mode 100644
index a5638591..00000000
--- a/projects/llm_framework/include/abseil-cpp/ci/absl_alternate_options.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// Copyright 2019 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//    https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-// Alternate options.h file, used in continuous integration testing to exercise
-// option settings not used by default.
-
-#ifndef ABSL_CI_ABSL_ALTERNATE_OPTIONS_H_
-#define ABSL_CI_ABSL_ALTERNATE_OPTIONS_H_
-
-#define ABSL_OPTION_USE_STD_ANY 0
-#define ABSL_OPTION_USE_STD_OPTIONAL 0
-#define ABSL_OPTION_USE_STD_STRING_VIEW 0
-#define ABSL_OPTION_USE_STD_VARIANT 0
-#define ABSL_OPTION_USE_STD_ORDERING 0
-#define ABSL_OPTION_USE_INLINE_NAMESPACE 1
-#define ABSL_OPTION_INLINE_NAMESPACE_NAME ns
-#define ABSL_OPTION_HARDENED 1
-
-#endif  // ABSL_CI_ABSL_ALTERNATE_OPTIONS_H_
diff --git a/projects/llm_framework/include/archive.h b/projects/llm_framework/include/archive.h
deleted file mode 100644
index 04601a79..00000000
--- a/projects/llm_framework/include/archive.h
+++ /dev/null
@@ -1,1214 +0,0 @@
-/*-
- * Copyright (c) 2003-2010 Tim Kientzle
- * All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions
- * are met:
- * 1. Redistributions of source code must retain the above copyright
- *    notice, this list of conditions and the following disclaimer.
- * 2. Redistributions in binary form must reproduce the above copyright
- *    notice, this list of conditions and the following disclaimer in the
- *    documentation and/or other materials provided with the distribution.
- *
- * THIS SOFTWARE IS PROVIDED BY THE AUTHOR(S) ``AS IS'' AND ANY EXPRESS OR
- * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
- * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
- * IN NO EVENT SHALL THE AUTHOR(S) BE LIABLE FOR ANY DIRECT, INDIRECT,
- * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
- * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
- * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
- * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
- * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- */
-
-#ifndef ARCHIVE_H_INCLUDED
-#define	ARCHIVE_H_INCLUDED
-
-/*
- * The version number is expressed as a single integer that makes it
- * easy to compare versions at build time: for version a.b.c, the
- * version number is printf("%d%03d%03d",a,b,c).  For example, if you
- * know your application requires version 2.12.108 or later, you can
- * assert that ARCHIVE_VERSION_NUMBER >= 2012108.
- */
-/* Note: Compiler will complain if this does not match archive_entry.h! */
-#define	ARCHIVE_VERSION_NUMBER 3008000
-
-#include <sys/stat.h>
-#include <stddef.h>  /* for wchar_t */
-#include <stdio.h> /* For FILE * */
-#include <time.h> /* For time_t */
-
-/*
- * Note: archive.h is for use outside of libarchive; the configuration
- * headers (config.h, archive_platform.h, etc.) are purely internal.
- * Do NOT use HAVE_XXX configuration macros to control the behavior of
- * this header!  If you must conditionalize, use predefined compiler and/or
- * platform macros.
- */
-#if defined(__BORLANDC__) && __BORLANDC__ >= 0x560
-# include <stdint.h>
-#elif !defined(__WATCOMC__) && !defined(_MSC_VER) && !defined(__INTERIX) && !defined(__BORLANDC__) && !defined(_SCO_DS) && !defined(__osf__) && !defined(__CLANG_INTTYPES_H)
-# include <inttypes.h>
-#endif
-
-/* Get appropriate definitions of 64-bit integer */
-#if !defined(__LA_INT64_T_DEFINED)
-/* Older code relied on the __LA_INT64_T macro; after 4.0 we'll switch to the typedef exclusively. */
-# if ARCHIVE_VERSION_NUMBER < 4000000
-#define __LA_INT64_T la_int64_t
-# endif
-#define __LA_INT64_T_DEFINED
-# if defined(_WIN32) && !defined(__CYGWIN__) && !defined(__WATCOMC__)
-typedef __int64 la_int64_t;
-# else
-# include <unistd.h>  /* ssize_t */
-#  if defined(_SCO_DS) || defined(__osf__)
-typedef long long la_int64_t;
-#  else
-typedef int64_t la_int64_t;
-#  endif
-# endif
-#endif
-
-/* The la_ssize_t should match the type used in 'struct stat' */
-#if !defined(__LA_SSIZE_T_DEFINED)
-/* Older code relied on the __LA_SSIZE_T macro; after 4.0 we'll switch to the typedef exclusively. */
-# if ARCHIVE_VERSION_NUMBER < 4000000
-#define __LA_SSIZE_T la_ssize_t
-# endif
-#define __LA_SSIZE_T_DEFINED
-# if defined(_WIN32) && !defined(__CYGWIN__) && !defined(__WATCOMC__)
-#  if defined(_SSIZE_T_DEFINED) || defined(_SSIZE_T_)
-typedef ssize_t la_ssize_t;
-#  elif defined(_WIN64)
-typedef __int64 la_ssize_t;
-#  else
-typedef long la_ssize_t;
-#  endif
-# else
-# include <unistd.h>  /* ssize_t */
-typedef ssize_t la_ssize_t;
-# endif
-#endif
-
-/* Large file support for Android */
-#if defined(__LIBARCHIVE_BUILD) && defined(__ANDROID__)
-#include "android_lf.h"
-#endif
-
-/*
- * On Windows, define LIBARCHIVE_STATIC if you're building or using a
- * .lib.  The default here assumes you're building a DLL.  Only
- * libarchive source should ever define __LIBARCHIVE_BUILD.
- */
-#if ((defined __WIN32__) || (defined _WIN32) || defined(__CYGWIN__)) && (!defined LIBARCHIVE_STATIC)
-# ifdef __LIBARCHIVE_BUILD
-#  ifdef __GNUC__
-#   define __LA_DECL	__attribute__((dllexport)) extern
-#  else
-#   define __LA_DECL	__declspec(dllexport)
-#  endif
-# else
-#  ifdef __GNUC__
-#   define __LA_DECL
-#  else
-#   define __LA_DECL	__declspec(dllimport)
-#  endif
-# endif
-#elif defined __LIBARCHIVE_ENABLE_VISIBILITY
-#  define __LA_DECL __attribute__((visibility("default")))
-#else
-/* Static libraries or non-Windows needs no special declaration. */
-# define __LA_DECL
-#endif
-
-#if defined(__GNUC__) && __GNUC__ >= 3 && !defined(__MINGW32__)
-#define	__LA_PRINTF(fmtarg, firstvararg) \
-	__attribute__((__format__ (__printf__, fmtarg, firstvararg)))
-#else
-#define	__LA_PRINTF(fmtarg, firstvararg)	/* nothing */
-#endif
-
-#if defined(__GNUC__) && __GNUC__ >= 3 && __GNUC_MINOR__ >= 1
-# define __LA_DEPRECATED __attribute__((deprecated))
-#else
-# define __LA_DEPRECATED
-#endif
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/*
- * The version number is provided as both a macro and a function.
- * The macro identifies the installed header; the function identifies
- * the library version (which may not be the same if you're using a
- * dynamically-linked version of the library).  Of course, if the
- * header and library are very different, you should expect some
- * strangeness.  Don't do that.
- */
-__LA_DECL int		archive_version_number(void);
-
-/*
- * Textual name/version of the library, useful for version displays.
- */
-#define	ARCHIVE_VERSION_ONLY_STRING "3.8.0dev"
-#define	ARCHIVE_VERSION_STRING "libarchive " ARCHIVE_VERSION_ONLY_STRING
-__LA_DECL const char *	archive_version_string(void);
-
-/*
- * Detailed textual name/version of the library and its dependencies.
- * This has the form:
- *    "libarchive x.y.z zlib/a.b.c liblzma/d.e.f ... etc ..."
- * the list of libraries described here will vary depending on how
- * libarchive was compiled.
- */
-__LA_DECL const char *	archive_version_details(void);
-
-/*
- * Returns NULL if libarchive was compiled without the associated library.
- * Otherwise, returns the version number that libarchive was compiled
- * against.
- */
-__LA_DECL const char *  archive_zlib_version(void);
-__LA_DECL const char *  archive_liblzma_version(void);
-__LA_DECL const char *  archive_bzlib_version(void);
-__LA_DECL const char *  archive_liblz4_version(void);
-__LA_DECL const char *  archive_libzstd_version(void);
-
-/* Declare our basic types. */
-struct archive;
-struct archive_entry;
-
-/*
- * Error codes: Use archive_errno() and archive_error_string()
- * to retrieve details.  Unless specified otherwise, all functions
- * that return 'int' use these codes.
- */
-#define	ARCHIVE_EOF	  1	/* Found end of archive. */
-#define	ARCHIVE_OK	  0	/* Operation was successful. */
-#define	ARCHIVE_RETRY	(-10)	/* Retry might succeed. */
-#define	ARCHIVE_WARN	(-20)	/* Partial success. */
-/* For example, if write_header "fails", then you can't push data. */
-#define	ARCHIVE_FAILED	(-25)	/* Current operation cannot complete. */
-/* But if write_header is "fatal," then this archive is dead and useless. */
-#define	ARCHIVE_FATAL	(-30)	/* No more operations are possible. */
-
-/*
- * As far as possible, archive_errno returns standard platform errno codes.
- * Of course, the details vary by platform, so the actual definitions
- * here are stored in "archive_platform.h".  The symbols are listed here
- * for reference; as a rule, clients should not need to know the exact
- * platform-dependent error code.
- */
-/* Unrecognized or invalid file format. */
-/* #define	ARCHIVE_ERRNO_FILE_FORMAT */
-/* Illegal usage of the library. */
-/* #define	ARCHIVE_ERRNO_PROGRAMMER_ERROR */
-/* Unknown or unclassified error. */
-/* #define	ARCHIVE_ERRNO_MISC */
-
-/*
- * Callbacks are invoked to automatically read/skip/write/open/close the
- * archive. You can provide your own for complex tasks (like breaking
- * archives across multiple tapes) or use standard ones built into the
- * library.
- */
-
-/* Returns pointer and size of next block of data from archive. */
-typedef la_ssize_t	archive_read_callback(struct archive *,
-			    void *_client_data, const void **_buffer);
-
-/* Skips at most request bytes from archive and returns the skipped amount.
- * This may skip fewer bytes than requested; it may even skip zero bytes.
- * If you do skip fewer bytes than requested, libarchive will invoke your
- * read callback and discard data as necessary to make up the full skip.
- */
-typedef la_int64_t	archive_skip_callback(struct archive *,
-			    void *_client_data, la_int64_t request);
-
-/* Seeks to specified location in the file and returns the position.
- * Whence values are SEEK_SET, SEEK_CUR, SEEK_END from stdio.h.
- * Return ARCHIVE_FATAL if the seek fails for any reason.
- */
-typedef la_int64_t	archive_seek_callback(struct archive *,
-    void *_client_data, la_int64_t offset, int whence);
-
-/* Returns size actually written, zero on EOF, -1 on error. */
-typedef la_ssize_t	archive_write_callback(struct archive *,
-			    void *_client_data,
-			    const void *_buffer, size_t _length);
-
-typedef int	archive_open_callback(struct archive *, void *_client_data);
-
-typedef int	archive_close_callback(struct archive *, void *_client_data);
-
-typedef int	archive_free_callback(struct archive *, void *_client_data);
-
-/* Switches from one client data object to the next/prev client data object.
- * This is useful for reading from different data blocks such as a set of files
- * that make up one large file.
- */
-typedef int archive_switch_callback(struct archive *, void *_client_data1,
-			    void *_client_data2);
-
-/*
- * Returns a passphrase used for encryption or decryption, NULL on nothing
- * to do and give it up.
- */
-typedef const char *archive_passphrase_callback(struct archive *,
-			    void *_client_data);
-
-/*
- * Codes to identify various stream filters.
- */
-#define	ARCHIVE_FILTER_NONE	0
-#define	ARCHIVE_FILTER_GZIP	1
-#define	ARCHIVE_FILTER_BZIP2	2
-#define	ARCHIVE_FILTER_COMPRESS	3
-#define	ARCHIVE_FILTER_PROGRAM	4
-#define	ARCHIVE_FILTER_LZMA	5
-#define	ARCHIVE_FILTER_XZ	6
-#define	ARCHIVE_FILTER_UU	7
-#define	ARCHIVE_FILTER_RPM	8
-#define	ARCHIVE_FILTER_LZIP	9
-#define	ARCHIVE_FILTER_LRZIP	10
-#define	ARCHIVE_FILTER_LZOP	11
-#define	ARCHIVE_FILTER_GRZIP	12
-#define	ARCHIVE_FILTER_LZ4	13
-#define	ARCHIVE_FILTER_ZSTD	14
-
-#if ARCHIVE_VERSION_NUMBER < 4000000
-#define	ARCHIVE_COMPRESSION_NONE	ARCHIVE_FILTER_NONE
-#define	ARCHIVE_COMPRESSION_GZIP	ARCHIVE_FILTER_GZIP
-#define	ARCHIVE_COMPRESSION_BZIP2	ARCHIVE_FILTER_BZIP2
-#define	ARCHIVE_COMPRESSION_COMPRESS	ARCHIVE_FILTER_COMPRESS
-#define	ARCHIVE_COMPRESSION_PROGRAM	ARCHIVE_FILTER_PROGRAM
-#define	ARCHIVE_COMPRESSION_LZMA	ARCHIVE_FILTER_LZMA
-#define	ARCHIVE_COMPRESSION_XZ		ARCHIVE_FILTER_XZ
-#define	ARCHIVE_COMPRESSION_UU		ARCHIVE_FILTER_UU
-#define	ARCHIVE_COMPRESSION_RPM		ARCHIVE_FILTER_RPM
-#define	ARCHIVE_COMPRESSION_LZIP	ARCHIVE_FILTER_LZIP
-#define	ARCHIVE_COMPRESSION_LRZIP	ARCHIVE_FILTER_LRZIP
-#endif
-
-/*
- * Codes returned by archive_format.
- *
- * Top 16 bits identifies the format family (e.g., "tar"); lower
- * 16 bits indicate the variant.  This is updated by read_next_header.
- * Note that the lower 16 bits will often vary from entry to entry.
- * In some cases, this variation occurs as libarchive learns more about
- * the archive (for example, later entries might utilize extensions that
- * weren't necessary earlier in the archive; in this case, libarchive
- * will change the format code to indicate the extended format that
- * was used).  In other cases, it's because different tools have
- * modified the archive and so different parts of the archive
- * actually have slightly different formats.  (Both tar and cpio store
- * format codes in each entry, so it is quite possible for each
- * entry to be in a different format.)
- */
-#define	ARCHIVE_FORMAT_BASE_MASK		0xff0000
-#define	ARCHIVE_FORMAT_CPIO			0x10000
-#define	ARCHIVE_FORMAT_CPIO_POSIX		(ARCHIVE_FORMAT_CPIO | 1)
-#define	ARCHIVE_FORMAT_CPIO_BIN_LE		(ARCHIVE_FORMAT_CPIO | 2)
-#define	ARCHIVE_FORMAT_CPIO_BIN_BE		(ARCHIVE_FORMAT_CPIO | 3)
-#define	ARCHIVE_FORMAT_CPIO_SVR4_NOCRC		(ARCHIVE_FORMAT_CPIO | 4)
-#define	ARCHIVE_FORMAT_CPIO_SVR4_CRC		(ARCHIVE_FORMAT_CPIO | 5)
-#define	ARCHIVE_FORMAT_CPIO_AFIO_LARGE		(ARCHIVE_FORMAT_CPIO | 6)
-#define	ARCHIVE_FORMAT_CPIO_PWB			(ARCHIVE_FORMAT_CPIO | 7)
-#define	ARCHIVE_FORMAT_SHAR			0x20000
-#define	ARCHIVE_FORMAT_SHAR_BASE		(ARCHIVE_FORMAT_SHAR | 1)
-#define	ARCHIVE_FORMAT_SHAR_DUMP		(ARCHIVE_FORMAT_SHAR | 2)
-#define	ARCHIVE_FORMAT_TAR			0x30000
-#define	ARCHIVE_FORMAT_TAR_USTAR		(ARCHIVE_FORMAT_TAR | 1)
-#define	ARCHIVE_FORMAT_TAR_PAX_INTERCHANGE	(ARCHIVE_FORMAT_TAR | 2)
-#define	ARCHIVE_FORMAT_TAR_PAX_RESTRICTED	(ARCHIVE_FORMAT_TAR | 3)
-#define	ARCHIVE_FORMAT_TAR_GNUTAR		(ARCHIVE_FORMAT_TAR | 4)
-#define	ARCHIVE_FORMAT_ISO9660			0x40000
-#define	ARCHIVE_FORMAT_ISO9660_ROCKRIDGE	(ARCHIVE_FORMAT_ISO9660 | 1)
-#define	ARCHIVE_FORMAT_ZIP			0x50000
-#define	ARCHIVE_FORMAT_EMPTY			0x60000
-#define	ARCHIVE_FORMAT_AR			0x70000
-#define	ARCHIVE_FORMAT_AR_GNU			(ARCHIVE_FORMAT_AR | 1)
-#define	ARCHIVE_FORMAT_AR_BSD			(ARCHIVE_FORMAT_AR | 2)
-#define	ARCHIVE_FORMAT_MTREE			0x80000
-#define	ARCHIVE_FORMAT_RAW			0x90000
-#define	ARCHIVE_FORMAT_XAR			0xA0000
-#define	ARCHIVE_FORMAT_LHA			0xB0000
-#define	ARCHIVE_FORMAT_CAB			0xC0000
-#define	ARCHIVE_FORMAT_RAR			0xD0000
-#define	ARCHIVE_FORMAT_7ZIP			0xE0000
-#define	ARCHIVE_FORMAT_WARC			0xF0000
-#define	ARCHIVE_FORMAT_RAR_V5			0x100000
-
-/*
- * Codes returned by archive_read_format_capabilities().
- *
- * This list can be extended with values between 0 and 0xffff.
- * The original purpose of this list was to let different archive
- * format readers expose their general capabilities in terms of
- * encryption.
- */
-#define ARCHIVE_READ_FORMAT_CAPS_NONE (0) /* no special capabilities */
-#define ARCHIVE_READ_FORMAT_CAPS_ENCRYPT_DATA (1<<0)  /* reader can detect encrypted data */
-#define ARCHIVE_READ_FORMAT_CAPS_ENCRYPT_METADATA (1<<1)  /* reader can detect encryptable metadata (pathname, mtime, etc.) */
-
-/*
- * Codes returned by archive_read_has_encrypted_entries().
- *
- * In case the archive does not support encryption detection at all
- * ARCHIVE_READ_FORMAT_ENCRYPTION_UNSUPPORTED is returned. If the reader
- * for some other reason (e.g. not enough bytes read) cannot say if
- * there are encrypted entries, ARCHIVE_READ_FORMAT_ENCRYPTION_DONT_KNOW
- * is returned.
- */
-#define ARCHIVE_READ_FORMAT_ENCRYPTION_UNSUPPORTED -2
-#define ARCHIVE_READ_FORMAT_ENCRYPTION_DONT_KNOW -1
-
-/*-
- * Basic outline for reading an archive:
- *   1) Ask archive_read_new for an archive reader object.
- *   2) Update any global properties as appropriate.
- *      In particular, you'll certainly want to call appropriate
- *      archive_read_support_XXX functions.
- *   3) Call archive_read_open_XXX to open the archive
- *   4) Repeatedly call archive_read_next_header to get information about
- *      successive archive entries.  Call archive_read_data to extract
- *      data for entries of interest.
- *   5) Call archive_read_free to end processing.
- */
-__LA_DECL struct archive	*archive_read_new(void);
-
-/*
- * The archive_read_support_XXX calls enable auto-detect for this
- * archive handle.  They also link in the necessary support code.
- * For example, if you don't want bzlib linked in, don't invoke
- * support_compression_bzip2().  The "all" functions provide the
- * obvious shorthand.
- */
-
-#if ARCHIVE_VERSION_NUMBER < 4000000
-__LA_DECL int archive_read_support_compression_all(struct archive *)
-		__LA_DEPRECATED;
-__LA_DECL int archive_read_support_compression_bzip2(struct archive *)
-		__LA_DEPRECATED;
-__LA_DECL int archive_read_support_compression_compress(struct archive *)
-		__LA_DEPRECATED;
-__LA_DECL int archive_read_support_compression_gzip(struct archive *)
-		__LA_DEPRECATED;
-__LA_DECL int archive_read_support_compression_lzip(struct archive *)
-		__LA_DEPRECATED;
-__LA_DECL int archive_read_support_compression_lzma(struct archive *)
-		__LA_DEPRECATED;
-__LA_DECL int archive_read_support_compression_none(struct archive *)
-		__LA_DEPRECATED;
-__LA_DECL int archive_read_support_compression_program(struct archive *,
-		     const char *command) __LA_DEPRECATED;
-__LA_DECL int archive_read_support_compression_program_signature
-		(struct archive *, const char *,
-		 const void * /* match */, size_t) __LA_DEPRECATED;
-
-__LA_DECL int archive_read_support_compression_rpm(struct archive *)
-		__LA_DEPRECATED;
-__LA_DECL int archive_read_support_compression_uu(struct archive *)
-		__LA_DEPRECATED;
-__LA_DECL int archive_read_support_compression_xz(struct archive *)
-		__LA_DEPRECATED;
-#endif
-
-__LA_DECL int archive_read_support_filter_all(struct archive *);
-__LA_DECL int archive_read_support_filter_by_code(struct archive *, int);
-__LA_DECL int archive_read_support_filter_bzip2(struct archive *);
-__LA_DECL int archive_read_support_filter_compress(struct archive *);
-__LA_DECL int archive_read_support_filter_gzip(struct archive *);
-__LA_DECL int archive_read_support_filter_grzip(struct archive *);
-__LA_DECL int archive_read_support_filter_lrzip(struct archive *);
-__LA_DECL int archive_read_support_filter_lz4(struct archive *);
-__LA_DECL int archive_read_support_filter_lzip(struct archive *);
-__LA_DECL int archive_read_support_filter_lzma(struct archive *);
-__LA_DECL int archive_read_support_filter_lzop(struct archive *);
-__LA_DECL int archive_read_support_filter_none(struct archive *);
-__LA_DECL int archive_read_support_filter_program(struct archive *,
-		     const char *command);
-__LA_DECL int archive_read_support_filter_program_signature
-		(struct archive *, const char * /* cmd */,
-				    const void * /* match */, size_t);
-__LA_DECL int archive_read_support_filter_rpm(struct archive *);
-__LA_DECL int archive_read_support_filter_uu(struct archive *);
-__LA_DECL int archive_read_support_filter_xz(struct archive *);
-__LA_DECL int archive_read_support_filter_zstd(struct archive *);
-
-__LA_DECL int archive_read_support_format_7zip(struct archive *);
-__LA_DECL int archive_read_support_format_all(struct archive *);
-__LA_DECL int archive_read_support_format_ar(struct archive *);
-__LA_DECL int archive_read_support_format_by_code(struct archive *, int);
-__LA_DECL int archive_read_support_format_cab(struct archive *);
-__LA_DECL int archive_read_support_format_cpio(struct archive *);
-__LA_DECL int archive_read_support_format_empty(struct archive *);
-__LA_DECL int archive_read_support_format_gnutar(struct archive *);
-__LA_DECL int archive_read_support_format_iso9660(struct archive *);
-__LA_DECL int archive_read_support_format_lha(struct archive *);
-__LA_DECL int archive_read_support_format_mtree(struct archive *);
-__LA_DECL int archive_read_support_format_rar(struct archive *);
-__LA_DECL int archive_read_support_format_rar5(struct archive *);
-__LA_DECL int archive_read_support_format_raw(struct archive *);
-__LA_DECL int archive_read_support_format_tar(struct archive *);
-__LA_DECL int archive_read_support_format_warc(struct archive *);
-__LA_DECL int archive_read_support_format_xar(struct archive *);
-/* archive_read_support_format_zip() enables both streamable and seekable
- * zip readers. */
-__LA_DECL int archive_read_support_format_zip(struct archive *);
-/* Reads Zip archives as stream from beginning to end.  Doesn't
- * correctly handle SFX ZIP files or ZIP archives that have been modified
- * in-place. */
-__LA_DECL int archive_read_support_format_zip_streamable(struct archive *);
-/* Reads starting from central directory; requires seekable input. */
-__LA_DECL int archive_read_support_format_zip_seekable(struct archive *);
-
-/* Functions to manually set the format and filters to be used. This is
- * useful to bypass the bidding process when the format and filters to use
- * is known in advance.
- */
-__LA_DECL int archive_read_set_format(struct archive *, int);
-__LA_DECL int archive_read_append_filter(struct archive *, int);
-__LA_DECL int archive_read_append_filter_program(struct archive *,
-    const char *);
-__LA_DECL int archive_read_append_filter_program_signature
-    (struct archive *, const char *, const void * /* match */, size_t);
-
-/* Set various callbacks. */
-__LA_DECL int archive_read_set_open_callback(struct archive *,
-    archive_open_callback *);
-__LA_DECL int archive_read_set_read_callback(struct archive *,
-    archive_read_callback *);
-__LA_DECL int archive_read_set_seek_callback(struct archive *,
-    archive_seek_callback *);
-__LA_DECL int archive_read_set_skip_callback(struct archive *,
-    archive_skip_callback *);
-__LA_DECL int archive_read_set_close_callback(struct archive *,
-    archive_close_callback *);
-/* Callback used to switch between one data object to the next */
-__LA_DECL int archive_read_set_switch_callback(struct archive *,
-    archive_switch_callback *);
-
-/* This sets the first data object. */
-__LA_DECL int archive_read_set_callback_data(struct archive *, void *);
-/* This sets data object at specified index */
-__LA_DECL int archive_read_set_callback_data2(struct archive *, void *,
-    unsigned int);
-/* This adds a data object at the specified index. */
-__LA_DECL int archive_read_add_callback_data(struct archive *, void *,
-    unsigned int);
-/* This appends a data object to the end of list */
-__LA_DECL int archive_read_append_callback_data(struct archive *, void *);
-/* This prepends a data object to the beginning of list */
-__LA_DECL int archive_read_prepend_callback_data(struct archive *, void *);
-
-/* Opening freezes the callbacks. */
-__LA_DECL int archive_read_open1(struct archive *);
-
-/* Convenience wrappers around the above. */
-__LA_DECL int archive_read_open(struct archive *, void *_client_data,
-		     archive_open_callback *, archive_read_callback *,
-		     archive_close_callback *);
-__LA_DECL int archive_read_open2(struct archive *, void *_client_data,
-		     archive_open_callback *, archive_read_callback *,
-		     archive_skip_callback *, archive_close_callback *);
-
-/*
- * A variety of shortcuts that invoke archive_read_open() with
- * canned callbacks suitable for common situations.  The ones that
- * accept a block size handle tape blocking correctly.
- */
-/* Use this if you know the filename.  Note: NULL indicates stdin. */
-__LA_DECL int archive_read_open_filename(struct archive *,
-		     const char *_filename, size_t _block_size);
-/* Use this for reading multivolume files by filenames.
- * NOTE: Must be NULL terminated. Sorting is NOT done. */
-__LA_DECL int archive_read_open_filenames(struct archive *,
-		     const char **_filenames, size_t _block_size);
-__LA_DECL int archive_read_open_filename_w(struct archive *,
-		     const wchar_t *_filename, size_t _block_size);
-#if defined(_WIN32) && !defined(__CYGWIN__)
-__LA_DECL int archive_read_open_filenames_w(struct archive *,
-		     const wchar_t **_filenames, size_t _block_size);
-#endif
-/* archive_read_open_file() is a deprecated synonym for ..._open_filename(). */
-__LA_DECL int archive_read_open_file(struct archive *,
-		     const char *_filename, size_t _block_size) __LA_DEPRECATED;
-/* Read an archive that's stored in memory. */
-__LA_DECL int archive_read_open_memory(struct archive *,
-		     const void * buff, size_t size);
-/* A more involved version that is only used for internal testing. */
-__LA_DECL int archive_read_open_memory2(struct archive *a, const void *buff,
-		     size_t size, size_t read_size);
-/* Read an archive that's already open, using the file descriptor. */
-__LA_DECL int archive_read_open_fd(struct archive *, int _fd,
-		     size_t _block_size);
-/* Read an archive that's already open, using a FILE *. */
-/* Note: DO NOT use this with tape drives. */
-__LA_DECL int archive_read_open_FILE(struct archive *, FILE *_file);
-
-/* Parses and returns next entry header. */
-__LA_DECL int archive_read_next_header(struct archive *,
-		     struct archive_entry **);
-
-/* Parses and returns next entry header using the archive_entry passed in */
-__LA_DECL int archive_read_next_header2(struct archive *,
-		     struct archive_entry *);
-
-/*
- * Retrieve the byte offset in UNCOMPRESSED data where last-read
- * header started.
- */
-__LA_DECL la_int64_t		 archive_read_header_position(struct archive *);
-
-/*
- * Returns 1 if the archive contains at least one encrypted entry.
- * If the archive format not support encryption at all
- * ARCHIVE_READ_FORMAT_ENCRYPTION_UNSUPPORTED is returned.
- * If for any other reason (e.g. not enough data read so far)
- * we cannot say whether there are encrypted entries, then
- * ARCHIVE_READ_FORMAT_ENCRYPTION_DONT_KNOW is returned.
- * In general, this function will return values below zero when the
- * reader is uncertain or totally incapable of encryption support.
- * When this function returns 0 you can be sure that the reader
- * supports encryption detection but no encrypted entries have
- * been found yet.
- *
- * NOTE: If the metadata/header of an archive is also encrypted, you
- * cannot rely on the number of encrypted entries. That is why this
- * function does not return the number of encrypted entries but#
- * just shows that there are some.
- */
-__LA_DECL int	archive_read_has_encrypted_entries(struct archive *);
-
-/*
- * Returns a bitmask of capabilities that are supported by the archive format reader.
- * If the reader has no special capabilities, ARCHIVE_READ_FORMAT_CAPS_NONE is returned.
- */
-__LA_DECL int		 archive_read_format_capabilities(struct archive *);
-
-/* Read data from the body of an entry.  Similar to read(2). */
-__LA_DECL la_ssize_t		 archive_read_data(struct archive *,
-				    void *, size_t);
-
-/* Seek within the body of an entry.  Similar to lseek(2). */
-__LA_DECL la_int64_t archive_seek_data(struct archive *, la_int64_t, int);
-
-/*
- * A zero-copy version of archive_read_data that also exposes the file offset
- * of each returned block.  Note that the client has no way to specify
- * the desired size of the block.  The API does guarantee that offsets will
- * be strictly increasing and that returned blocks will not overlap.
- */
-__LA_DECL int archive_read_data_block(struct archive *a,
-		    const void **buff, size_t *size, la_int64_t *offset);
-
-/*-
- * Some convenience functions that are built on archive_read_data:
- *  'skip': skips entire entry
- *  'into_buffer': writes data into memory buffer that you provide
- *  'into_fd': writes data to specified filedes
- */
-__LA_DECL int archive_read_data_skip(struct archive *);
-__LA_DECL int archive_read_data_into_fd(struct archive *, int fd);
-
-/*
- * Set read options.
- */
-/* Apply option to the format only. */
-__LA_DECL int archive_read_set_format_option(struct archive *_a,
-			    const char *m, const char *o,
-			    const char *v);
-/* Apply option to the filter only. */
-__LA_DECL int archive_read_set_filter_option(struct archive *_a,
-			    const char *m, const char *o,
-			    const char *v);
-/* Apply option to both the format and the filter. */
-__LA_DECL int archive_read_set_option(struct archive *_a,
-			    const char *m, const char *o,
-			    const char *v);
-/* Apply option string to both the format and the filter. */
-__LA_DECL int archive_read_set_options(struct archive *_a,
-			    const char *opts);
-
-/*
- * Add a decryption passphrase.
- */
-__LA_DECL int archive_read_add_passphrase(struct archive *, const char *);
-__LA_DECL int archive_read_set_passphrase_callback(struct archive *,
-			    void *client_data, archive_passphrase_callback *);
-
-
-/*-
- * Convenience function to recreate the current entry (whose header
- * has just been read) on disk.
- *
- * This does quite a bit more than just copy data to disk. It also:
- *  - Creates intermediate directories as required.
- *  - Manages directory permissions:  non-writable directories will
- *    be initially created with write permission enabled; when the
- *    archive is closed, dir permissions are edited to the values specified
- *    in the archive.
- *  - Checks hardlinks:  hardlinks will not be extracted unless the
- *    linked-to file was also extracted within the same session. (TODO)
- */
-
-/* The "flags" argument selects optional behavior, 'OR' the flags you want. */
-
-/* Default: Do not try to set owner/group. */
-#define	ARCHIVE_EXTRACT_OWNER			(0x0001)
-/* Default: Do obey umask, do not restore SUID/SGID/SVTX bits. */
-#define	ARCHIVE_EXTRACT_PERM			(0x0002)
-/* Default: Do not restore mtime/atime. */
-#define	ARCHIVE_EXTRACT_TIME			(0x0004)
-/* Default: Replace existing files. */
-#define	ARCHIVE_EXTRACT_NO_OVERWRITE 		(0x0008)
-/* Default: Try create first, unlink only if create fails with EEXIST. */
-#define	ARCHIVE_EXTRACT_UNLINK			(0x0010)
-/* Default: Do not restore ACLs. */
-#define	ARCHIVE_EXTRACT_ACL			(0x0020)
-/* Default: Do not restore fflags. */
-#define	ARCHIVE_EXTRACT_FFLAGS			(0x0040)
-/* Default: Do not restore xattrs. */
-#define	ARCHIVE_EXTRACT_XATTR 			(0x0080)
-/* Default: Do not try to guard against extracts redirected by symlinks. */
-/* Note: With ARCHIVE_EXTRACT_UNLINK, will remove any intermediate symlink. */
-#define	ARCHIVE_EXTRACT_SECURE_SYMLINKS		(0x0100)
-/* Default: Do not reject entries with '..' as path elements. */
-#define	ARCHIVE_EXTRACT_SECURE_NODOTDOT		(0x0200)
-/* Default: Create parent directories as needed. */
-#define	ARCHIVE_EXTRACT_NO_AUTODIR		(0x0400)
-/* Default: Overwrite files, even if one on disk is newer. */
-#define	ARCHIVE_EXTRACT_NO_OVERWRITE_NEWER	(0x0800)
-/* Detect blocks of 0 and write holes instead. */
-#define	ARCHIVE_EXTRACT_SPARSE			(0x1000)
-/* Default: Do not restore Mac extended metadata. */
-/* This has no effect except on Mac OS. */
-#define	ARCHIVE_EXTRACT_MAC_METADATA		(0x2000)
-/* Default: Use HFS+ compression if it was compressed. */
-/* This has no effect except on Mac OS v10.6 or later. */
-#define	ARCHIVE_EXTRACT_NO_HFS_COMPRESSION	(0x4000)
-/* Default: Do not use HFS+ compression if it was not compressed. */
-/* This has no effect except on Mac OS v10.6 or later. */
-#define	ARCHIVE_EXTRACT_HFS_COMPRESSION_FORCED	(0x8000)
-/* Default: Do not reject entries with absolute paths */
-#define ARCHIVE_EXTRACT_SECURE_NOABSOLUTEPATHS (0x10000)
-/* Default: Do not clear no-change flags when unlinking object */
-#define	ARCHIVE_EXTRACT_CLEAR_NOCHANGE_FFLAGS	(0x20000)
-/* Default: Do not extract atomically (using rename) */
-#define	ARCHIVE_EXTRACT_SAFE_WRITES		(0x40000)
-
-__LA_DECL int archive_read_extract(struct archive *, struct archive_entry *,
-		     int flags);
-__LA_DECL int archive_read_extract2(struct archive *, struct archive_entry *,
-		     struct archive * /* dest */);
-__LA_DECL void	 archive_read_extract_set_progress_callback(struct archive *,
-		     void (*_progress_func)(void *), void *_user_data);
-
-/* Record the dev/ino of a file that will not be written.  This is
- * generally set to the dev/ino of the archive being read. */
-__LA_DECL void		archive_read_extract_set_skip_file(struct archive *,
-		     la_int64_t, la_int64_t);
-
-/* Close the file and release most resources. */
-__LA_DECL int		 archive_read_close(struct archive *);
-/* Release all resources and destroy the object. */
-/* Note that archive_read_free will call archive_read_close for you. */
-__LA_DECL int		 archive_read_free(struct archive *);
-#if ARCHIVE_VERSION_NUMBER < 4000000
-/* Synonym for archive_read_free() for backwards compatibility. */
-__LA_DECL int		 archive_read_finish(struct archive *) __LA_DEPRECATED;
-#endif
-
-/*-
- * To create an archive:
- *   1) Ask archive_write_new for an archive writer object.
- *   2) Set any global properties.  In particular, you should set
- *      the compression and format to use.
- *   3) Call archive_write_open to open the file (most people
- *       will use archive_write_open_file or archive_write_open_fd,
- *       which provide convenient canned I/O callbacks for you).
- *   4) For each entry:
- *      - construct an appropriate struct archive_entry structure
- *      - archive_write_header to write the header
- *      - archive_write_data to write the entry data
- *   5) archive_write_close to close the output
- *   6) archive_write_free to cleanup the writer and release resources
- */
-__LA_DECL struct archive	*archive_write_new(void);
-__LA_DECL int archive_write_set_bytes_per_block(struct archive *,
-		     int bytes_per_block);
-__LA_DECL int archive_write_get_bytes_per_block(struct archive *);
-/* XXX This is badly misnamed; suggestions appreciated. XXX */
-__LA_DECL int archive_write_set_bytes_in_last_block(struct archive *,
-		     int bytes_in_last_block);
-__LA_DECL int archive_write_get_bytes_in_last_block(struct archive *);
-
-/* The dev/ino of a file that won't be archived.  This is used
- * to avoid recursively adding an archive to itself. */
-__LA_DECL int archive_write_set_skip_file(struct archive *,
-    la_int64_t, la_int64_t);
-
-#if ARCHIVE_VERSION_NUMBER < 4000000
-__LA_DECL int archive_write_set_compression_bzip2(struct archive *)
-		__LA_DEPRECATED;
-__LA_DECL int archive_write_set_compression_compress(struct archive *)
-		__LA_DEPRECATED;
-__LA_DECL int archive_write_set_compression_gzip(struct archive *)
-		__LA_DEPRECATED;
-__LA_DECL int archive_write_set_compression_lzip(struct archive *)
-		__LA_DEPRECATED;
-__LA_DECL int archive_write_set_compression_lzma(struct archive *)
-		__LA_DEPRECATED;
-__LA_DECL int archive_write_set_compression_none(struct archive *)
-		__LA_DEPRECATED;
-__LA_DECL int archive_write_set_compression_program(struct archive *,
-		     const char *cmd) __LA_DEPRECATED;
-__LA_DECL int archive_write_set_compression_xz(struct archive *)
-		__LA_DEPRECATED;
-#endif
-
-/* A convenience function to set the filter based on the code. */
-__LA_DECL int archive_write_add_filter(struct archive *, int filter_code);
-__LA_DECL int archive_write_add_filter_by_name(struct archive *,
-		     const char *name);
-__LA_DECL int archive_write_add_filter_b64encode(struct archive *);
-__LA_DECL int archive_write_add_filter_bzip2(struct archive *);
-__LA_DECL int archive_write_add_filter_compress(struct archive *);
-__LA_DECL int archive_write_add_filter_grzip(struct archive *);
-__LA_DECL int archive_write_add_filter_gzip(struct archive *);
-__LA_DECL int archive_write_add_filter_lrzip(struct archive *);
-__LA_DECL int archive_write_add_filter_lz4(struct archive *);
-__LA_DECL int archive_write_add_filter_lzip(struct archive *);
-__LA_DECL int archive_write_add_filter_lzma(struct archive *);
-__LA_DECL int archive_write_add_filter_lzop(struct archive *);
-__LA_DECL int archive_write_add_filter_none(struct archive *);
-__LA_DECL int archive_write_add_filter_program(struct archive *,
-		     const char *cmd);
-__LA_DECL int archive_write_add_filter_uuencode(struct archive *);
-__LA_DECL int archive_write_add_filter_xz(struct archive *);
-__LA_DECL int archive_write_add_filter_zstd(struct archive *);
-
-
-/* A convenience function to set the format based on the code or name. */
-__LA_DECL int archive_write_set_format(struct archive *, int format_code);
-__LA_DECL int archive_write_set_format_by_name(struct archive *,
-		     const char *name);
-/* To minimize link pollution, use one or more of the following. */
-__LA_DECL int archive_write_set_format_7zip(struct archive *);
-__LA_DECL int archive_write_set_format_ar_bsd(struct archive *);
-__LA_DECL int archive_write_set_format_ar_svr4(struct archive *);
-__LA_DECL int archive_write_set_format_cpio(struct archive *);
-__LA_DECL int archive_write_set_format_cpio_bin(struct archive *);
-__LA_DECL int archive_write_set_format_cpio_newc(struct archive *);
-__LA_DECL int archive_write_set_format_cpio_odc(struct archive *);
-__LA_DECL int archive_write_set_format_cpio_pwb(struct archive *);
-__LA_DECL int archive_write_set_format_gnutar(struct archive *);
-__LA_DECL int archive_write_set_format_iso9660(struct archive *);
-__LA_DECL int archive_write_set_format_mtree(struct archive *);
-__LA_DECL int archive_write_set_format_mtree_classic(struct archive *);
-/* TODO: int archive_write_set_format_old_tar(struct archive *); */
-__LA_DECL int archive_write_set_format_pax(struct archive *);
-__LA_DECL int archive_write_set_format_pax_restricted(struct archive *);
-__LA_DECL int archive_write_set_format_raw(struct archive *);
-__LA_DECL int archive_write_set_format_shar(struct archive *);
-__LA_DECL int archive_write_set_format_shar_dump(struct archive *);
-__LA_DECL int archive_write_set_format_ustar(struct archive *);
-__LA_DECL int archive_write_set_format_v7tar(struct archive *);
-__LA_DECL int archive_write_set_format_warc(struct archive *);
-__LA_DECL int archive_write_set_format_xar(struct archive *);
-__LA_DECL int archive_write_set_format_zip(struct archive *);
-__LA_DECL int archive_write_set_format_filter_by_ext(struct archive *a, const char *filename);
-__LA_DECL int archive_write_set_format_filter_by_ext_def(struct archive *a, const char *filename, const char * def_ext);
-__LA_DECL int archive_write_zip_set_compression_deflate(struct archive *);
-__LA_DECL int archive_write_zip_set_compression_store(struct archive *);
-/* Deprecated; use archive_write_open2 instead */
-__LA_DECL int archive_write_open(struct archive *, void *,
-		     archive_open_callback *, archive_write_callback *,
-		     archive_close_callback *);
-__LA_DECL int archive_write_open2(struct archive *, void *,
-		     archive_open_callback *, archive_write_callback *,
-		     archive_close_callback *, archive_free_callback *);
-__LA_DECL int archive_write_open_fd(struct archive *, int _fd);
-__LA_DECL int archive_write_open_filename(struct archive *, const char *_file);
-__LA_DECL int archive_write_open_filename_w(struct archive *,
-		     const wchar_t *_file);
-/* A deprecated synonym for archive_write_open_filename() */
-__LA_DECL int archive_write_open_file(struct archive *, const char *_file)
-		__LA_DEPRECATED;
-__LA_DECL int archive_write_open_FILE(struct archive *, FILE *);
-/* _buffSize is the size of the buffer, _used refers to a variable that
- * will be updated after each write into the buffer. */
-__LA_DECL int archive_write_open_memory(struct archive *,
-			void *_buffer, size_t _buffSize, size_t *_used);
-
-/*
- * Note that the library will truncate writes beyond the size provided
- * to archive_write_header or pad if the provided data is short.
- */
-__LA_DECL int archive_write_header(struct archive *,
-		     struct archive_entry *);
-__LA_DECL la_ssize_t	archive_write_data(struct archive *,
-			    const void *, size_t);
-
-/* This interface is currently only available for archive_write_disk handles.  */
-__LA_DECL la_ssize_t	 archive_write_data_block(struct archive *,
-				    const void *, size_t, la_int64_t);
-
-__LA_DECL int		 archive_write_finish_entry(struct archive *);
-__LA_DECL int		 archive_write_close(struct archive *);
-/* Marks the archive as FATAL so that a subsequent free() operation
- * won't try to close() cleanly.  Provides a fast abort capability
- * when the client discovers that things have gone wrong. */
-__LA_DECL int            archive_write_fail(struct archive *);
-/* This can fail if the archive wasn't already closed, in which case
- * archive_write_free() will implicitly call archive_write_close(). */
-__LA_DECL int		 archive_write_free(struct archive *);
-#if ARCHIVE_VERSION_NUMBER < 4000000
-/* Synonym for archive_write_free() for backwards compatibility. */
-__LA_DECL int		 archive_write_finish(struct archive *) __LA_DEPRECATED;
-#endif
-
-/*
- * Set write options.
- */
-/* Apply option to the format only. */
-__LA_DECL int archive_write_set_format_option(struct archive *_a,
-			    const char *m, const char *o,
-			    const char *v);
-/* Apply option to the filter only. */
-__LA_DECL int archive_write_set_filter_option(struct archive *_a,
-			    const char *m, const char *o,
-			    const char *v);
-/* Apply option to both the format and the filter. */
-__LA_DECL int archive_write_set_option(struct archive *_a,
-			    const char *m, const char *o,
-			    const char *v);
-/* Apply option string to both the format and the filter. */
-__LA_DECL int archive_write_set_options(struct archive *_a,
-			    const char *opts);
-
-/*
- * Set an encryption passphrase.
- */
-__LA_DECL int archive_write_set_passphrase(struct archive *_a, const char *p);
-__LA_DECL int archive_write_set_passphrase_callback(struct archive *,
-			    void *client_data, archive_passphrase_callback *);
-
-/*-
- * ARCHIVE_WRITE_DISK API
- *
- * To create objects on disk:
- *   1) Ask archive_write_disk_new for a new archive_write_disk object.
- *   2) Set any global properties.  In particular, you probably
- *      want to set the options.
- *   3) For each entry:
- *      - construct an appropriate struct archive_entry structure
- *      - archive_write_header to create the file/dir/etc on disk
- *      - archive_write_data to write the entry data
- *   4) archive_write_free to cleanup the writer and release resources
- *
- * In particular, you can use this in conjunction with archive_read()
- * to pull entries out of an archive and create them on disk.
- */
-__LA_DECL struct archive	*archive_write_disk_new(void);
-/* This file will not be overwritten. */
-__LA_DECL int archive_write_disk_set_skip_file(struct archive *,
-    la_int64_t, la_int64_t);
-/* Set flags to control how the next item gets created.
- * This accepts a bitmask of ARCHIVE_EXTRACT_XXX flags defined above. */
-__LA_DECL int		 archive_write_disk_set_options(struct archive *,
-		     int flags);
-/*
- * The lookup functions are given uname/uid (or gname/gid) pairs and
- * return a uid (gid) suitable for this system.  These are used for
- * restoring ownership and for setting ACLs.  The default functions
- * are naive, they just return the uid/gid.  These are small, so reasonable
- * for applications that don't need to preserve ownership; they
- * are probably also appropriate for applications that are doing
- * same-system backup and restore.
- */
-/*
- * The "standard" lookup functions use common system calls to lookup
- * the uname/gname, falling back to the uid/gid if the names can't be
- * found.  They cache lookups and are reasonably fast, but can be very
- * large, so they are not used unless you ask for them.  In
- * particular, these match the specifications of POSIX "pax" and old
- * POSIX "tar".
- */
-__LA_DECL int	 archive_write_disk_set_standard_lookup(struct archive *);
-/*
- * If neither the default (naive) nor the standard (big) functions suit
- * your needs, you can write your own and register them.  Be sure to
- * include a cleanup function if you have allocated private data.
- */
-__LA_DECL int archive_write_disk_set_group_lookup(struct archive *,
-    void * /* private_data */,
-    la_int64_t (*)(void *, const char *, la_int64_t),
-    void (* /* cleanup */)(void *));
-__LA_DECL int archive_write_disk_set_user_lookup(struct archive *,
-    void * /* private_data */,
-    la_int64_t (*)(void *, const char *, la_int64_t),
-    void (* /* cleanup */)(void *));
-__LA_DECL la_int64_t archive_write_disk_gid(struct archive *, const char *, la_int64_t);
-__LA_DECL la_int64_t archive_write_disk_uid(struct archive *, const char *, la_int64_t);
-
-/*
- * ARCHIVE_READ_DISK API
- *
- * This is still evolving and somewhat experimental.
- */
-__LA_DECL struct archive *archive_read_disk_new(void);
-/* The names for symlink modes here correspond to an old BSD
- * command-line argument convention: -L, -P, -H */
-/* Follow all symlinks. */
-__LA_DECL int archive_read_disk_set_symlink_logical(struct archive *);
-/* Follow no symlinks. */
-__LA_DECL int archive_read_disk_set_symlink_physical(struct archive *);
-/* Follow symlink initially, then not. */
-__LA_DECL int archive_read_disk_set_symlink_hybrid(struct archive *);
-/* TODO: Handle Linux stat32/stat64 ugliness. <sigh> */
-__LA_DECL int archive_read_disk_entry_from_file(struct archive *,
-    struct archive_entry *, int /* fd */, const struct stat *);
-/* Look up gname for gid or uname for uid. */
-/* Default implementations are very, very stupid. */
-__LA_DECL const char *archive_read_disk_gname(struct archive *, la_int64_t);
-__LA_DECL const char *archive_read_disk_uname(struct archive *, la_int64_t);
-/* "Standard" implementation uses getpwuid_r, getgrgid_r and caches the
- * results for performance. */
-__LA_DECL int	archive_read_disk_set_standard_lookup(struct archive *);
-/* You can install your own lookups if you like. */
-__LA_DECL int	archive_read_disk_set_gname_lookup(struct archive *,
-    void * /* private_data */,
-    const char *(* /* lookup_fn */)(void *, la_int64_t),
-    void (* /* cleanup_fn */)(void *));
-__LA_DECL int	archive_read_disk_set_uname_lookup(struct archive *,
-    void * /* private_data */,
-    const char *(* /* lookup_fn */)(void *, la_int64_t),
-    void (* /* cleanup_fn */)(void *));
-/* Start traversal. */
-__LA_DECL int	archive_read_disk_open(struct archive *, const char *);
-__LA_DECL int	archive_read_disk_open_w(struct archive *, const wchar_t *);
-/*
- * Request that current entry be visited.  If you invoke it on every
- * directory, you'll get a physical traversal.  This is ignored if the
- * current entry isn't a directory or a link to a directory.  So, if
- * you invoke this on every returned path, you'll get a full logical
- * traversal.
- */
-__LA_DECL int	archive_read_disk_descend(struct archive *);
-__LA_DECL int	archive_read_disk_can_descend(struct archive *);
-__LA_DECL int	archive_read_disk_current_filesystem(struct archive *);
-__LA_DECL int	archive_read_disk_current_filesystem_is_synthetic(struct archive *);
-__LA_DECL int	archive_read_disk_current_filesystem_is_remote(struct archive *);
-/* Request that the access time of the entry visited by traversal be restored. */
-__LA_DECL int  archive_read_disk_set_atime_restored(struct archive *);
-/*
- * Set behavior. The "flags" argument selects optional behavior.
- */
-/* Request that the access time of the entry visited by traversal be restored.
- * This is the same as archive_read_disk_set_atime_restored. */
-#define	ARCHIVE_READDISK_RESTORE_ATIME		(0x0001)
-/* Default: Do not skip an entry which has nodump flags. */
-#define	ARCHIVE_READDISK_HONOR_NODUMP		(0x0002)
-/* Default: Skip a mac resource fork file whose prefix is "._" because of
- * using copyfile. */
-#define	ARCHIVE_READDISK_MAC_COPYFILE		(0x0004)
-/* Default: Traverse mount points. */
-#define	ARCHIVE_READDISK_NO_TRAVERSE_MOUNTS	(0x0008)
-/* Default: Xattrs are read from disk. */
-#define	ARCHIVE_READDISK_NO_XATTR		(0x0010)
-/* Default: ACLs are read from disk. */
-#define	ARCHIVE_READDISK_NO_ACL			(0x0020)
-/* Default: File flags are read from disk. */
-#define	ARCHIVE_READDISK_NO_FFLAGS		(0x0040)
-/* Default: Sparse file information is read from disk. */
-#define	ARCHIVE_READDISK_NO_SPARSE		(0x0080)
-
-__LA_DECL int  archive_read_disk_set_behavior(struct archive *,
-		    int flags);
-
-/*
- * Set archive_match object that will be used in archive_read_disk to
- * know whether an entry should be skipped. The callback function
- * _excluded_func will be invoked when an entry is skipped by the result
- * of archive_match.
- */
-__LA_DECL int	archive_read_disk_set_matching(struct archive *,
-		    struct archive *_matching, void (*_excluded_func)
-		    (struct archive *, void *, struct archive_entry *),
-		    void *_client_data);
-__LA_DECL int	archive_read_disk_set_metadata_filter_callback(struct archive *,
-		    int (*_metadata_filter_func)(struct archive *, void *,
-		    	struct archive_entry *), void *_client_data);
-
-/* Simplified cleanup interface;
- * This calls archive_read_free() or archive_write_free() as needed. */
-__LA_DECL int	archive_free(struct archive *);
-
-/*
- * Accessor functions to read/set various information in
- * the struct archive object:
- */
-
-/* Number of filters in the current filter pipeline. */
-/* Filter #0 is the one closest to the format, -1 is a synonym for the
- * last filter, which is always the pseudo-filter that wraps the
- * client callbacks. */
-__LA_DECL int		 archive_filter_count(struct archive *);
-__LA_DECL la_int64_t	 archive_filter_bytes(struct archive *, int);
-__LA_DECL int		 archive_filter_code(struct archive *, int);
-__LA_DECL const char *	 archive_filter_name(struct archive *, int);
-
-#if ARCHIVE_VERSION_NUMBER < 4000000
-/* These don't properly handle multiple filters, so are deprecated and
- * will eventually be removed. */
-/* As of libarchive 3.0, this is an alias for archive_filter_bytes(a, -1); */
-__LA_DECL la_int64_t	 archive_position_compressed(struct archive *)
-				__LA_DEPRECATED;
-/* As of libarchive 3.0, this is an alias for archive_filter_bytes(a, 0); */
-__LA_DECL la_int64_t	 archive_position_uncompressed(struct archive *)
-				__LA_DEPRECATED;
-/* As of libarchive 3.0, this is an alias for archive_filter_name(a, 0); */
-__LA_DECL const char	*archive_compression_name(struct archive *)
-				__LA_DEPRECATED;
-/* As of libarchive 3.0, this is an alias for archive_filter_code(a, 0); */
-__LA_DECL int		 archive_compression(struct archive *)
-				__LA_DEPRECATED;
-#endif
-
-__LA_DECL int		 archive_errno(struct archive *);
-__LA_DECL const char	*archive_error_string(struct archive *);
-__LA_DECL const char	*archive_format_name(struct archive *);
-__LA_DECL int		 archive_format(struct archive *);
-__LA_DECL void		 archive_clear_error(struct archive *);
-__LA_DECL void		 archive_set_error(struct archive *, int _err,
-			    const char *fmt, ...) __LA_PRINTF(3, 4);
-__LA_DECL void		 archive_copy_error(struct archive *dest,
-			    struct archive *src);
-__LA_DECL int		 archive_file_count(struct archive *);
-
-/*
- * ARCHIVE_MATCH API
- */
-__LA_DECL struct archive *archive_match_new(void);
-__LA_DECL int	archive_match_free(struct archive *);
-
-/*
- * Test if archive_entry is excluded.
- * This is a convenience function. This is the same as calling all
- * archive_match_path_excluded, archive_match_time_excluded
- * and archive_match_owner_excluded.
- */
-__LA_DECL int	archive_match_excluded(struct archive *,
-		    struct archive_entry *);
-
-/*
- * Test if pathname is excluded. The conditions are set by following functions.
- */
-__LA_DECL int	archive_match_path_excluded(struct archive *,
-		    struct archive_entry *);
-/* Control recursive inclusion of directory content when directory is included. Default on. */
-__LA_DECL int	archive_match_set_inclusion_recursion(struct archive *, int);
-/* Add exclusion pathname pattern. */
-__LA_DECL int	archive_match_exclude_pattern(struct archive *, const char *);
-__LA_DECL int	archive_match_exclude_pattern_w(struct archive *,
-		    const wchar_t *);
-/* Add exclusion pathname pattern from file. */
-__LA_DECL int	archive_match_exclude_pattern_from_file(struct archive *,
-		    const char *, int _nullSeparator);
-__LA_DECL int	archive_match_exclude_pattern_from_file_w(struct archive *,
-		    const wchar_t *, int _nullSeparator);
-/* Add inclusion pathname pattern. */
-__LA_DECL int	archive_match_include_pattern(struct archive *, const char *);
-__LA_DECL int	archive_match_include_pattern_w(struct archive *,
-		    const wchar_t *);
-/* Add inclusion pathname pattern from file. */
-__LA_DECL int	archive_match_include_pattern_from_file(struct archive *,
-		    const char *, int _nullSeparator);
-__LA_DECL int	archive_match_include_pattern_from_file_w(struct archive *,
-		    const wchar_t *, int _nullSeparator);
-/*
- * How to get statistic information for inclusion patterns.
- */
-/* Return the amount number of unmatched inclusion patterns. */
-__LA_DECL int	archive_match_path_unmatched_inclusions(struct archive *);
-/* Return the pattern of unmatched inclusion with ARCHIVE_OK.
- * Return ARCHIVE_EOF if there is no inclusion pattern. */
-__LA_DECL int	archive_match_path_unmatched_inclusions_next(
-		    struct archive *, const char **);
-__LA_DECL int	archive_match_path_unmatched_inclusions_next_w(
-		    struct archive *, const wchar_t **);
-
-/*
- * Test if a file is excluded by its time stamp.
- * The conditions are set by following functions.
- */
-__LA_DECL int	archive_match_time_excluded(struct archive *,
-		    struct archive_entry *);
-
-/*
- * Flags to tell a matching type of time stamps. These are used for
- * following functions.
- */
-/* Time flag: mtime to be tested. */
-#define ARCHIVE_MATCH_MTIME	(0x0100)
-/* Time flag: ctime to be tested. */
-#define ARCHIVE_MATCH_CTIME	(0x0200)
-/* Comparison flag: Match the time if it is newer than. */
-#define ARCHIVE_MATCH_NEWER	(0x0001)
-/* Comparison flag: Match the time if it is older than. */
-#define ARCHIVE_MATCH_OLDER	(0x0002)
-/* Comparison flag: Match the time if it is equal to. */
-#define ARCHIVE_MATCH_EQUAL	(0x0010)
-/* Set inclusion time. */
-__LA_DECL int	archive_match_include_time(struct archive *, int _flag,
-		    time_t _sec, long _nsec);
-/* Set inclusion time by a date string. */
-__LA_DECL int	archive_match_include_date(struct archive *, int _flag,
-		    const char *_datestr);
-__LA_DECL int	archive_match_include_date_w(struct archive *, int _flag,
-		    const wchar_t *_datestr);
-/* Set inclusion time by a particular file. */
-__LA_DECL int	archive_match_include_file_time(struct archive *,
-		    int _flag, const char *_pathname);
-__LA_DECL int	archive_match_include_file_time_w(struct archive *,
-		    int _flag, const wchar_t *_pathname);
-/* Add exclusion entry. */
-__LA_DECL int	archive_match_exclude_entry(struct archive *,
-		    int _flag, struct archive_entry *);
-
-/*
- * Test if a file is excluded by its uid ,gid, uname or gname.
- * The conditions are set by following functions.
- */
-__LA_DECL int	archive_match_owner_excluded(struct archive *,
-		    struct archive_entry *);
-/* Add inclusion uid, gid, uname and gname. */
-__LA_DECL int	archive_match_include_uid(struct archive *, la_int64_t);
-__LA_DECL int	archive_match_include_gid(struct archive *, la_int64_t);
-__LA_DECL int	archive_match_include_uname(struct archive *, const char *);
-__LA_DECL int	archive_match_include_uname_w(struct archive *,
-		    const wchar_t *);
-__LA_DECL int	archive_match_include_gname(struct archive *, const char *);
-__LA_DECL int	archive_match_include_gname_w(struct archive *,
-		    const wchar_t *);
-
-/* Utility functions */
-/* Convenience function to sort a NULL terminated list of strings */
-__LA_DECL int archive_utility_string_sort(char **);
-
-#ifdef __cplusplus
-}
-#endif
-
-/* These are meaningless outside of this header. */
-#undef __LA_DECL
-
-#endif /* !ARCHIVE_H_INCLUDED */
diff --git a/projects/llm_framework/include/archive_entry.h b/projects/llm_framework/include/archive_entry.h
deleted file mode 100644
index de2d2ffd..00000000
--- a/projects/llm_framework/include/archive_entry.h
+++ /dev/null
@@ -1,731 +0,0 @@
-/*-
- * Copyright (c) 2003-2008 Tim Kientzle
- * Copyright (c) 2016 Martin Matuska
- * All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions
- * are met:
- * 1. Redistributions of source code must retain the above copyright
- *    notice, this list of conditions and the following disclaimer.
- * 2. Redistributions in binary form must reproduce the above copyright
- *    notice, this list of conditions and the following disclaimer in the
- *    documentation and/or other materials provided with the distribution.
- *
- * THIS SOFTWARE IS PROVIDED BY THE AUTHOR(S) ``AS IS'' AND ANY EXPRESS OR
- * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
- * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
- * IN NO EVENT SHALL THE AUTHOR(S) BE LIABLE FOR ANY DIRECT, INDIRECT,
- * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
- * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
- * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
- * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
- * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- */
-
-#ifndef ARCHIVE_ENTRY_H_INCLUDED
-#define	ARCHIVE_ENTRY_H_INCLUDED
-
-/* Note: Compiler will complain if this does not match archive.h! */
-#define	ARCHIVE_VERSION_NUMBER 3008000
-
-/*
- * Note: archive_entry.h is for use outside of libarchive; the
- * configuration headers (config.h, archive_platform.h, etc.) are
- * purely internal.  Do NOT use HAVE_XXX configuration macros to
- * control the behavior of this header!  If you must conditionalize,
- * use predefined compiler and/or platform macros.
- */
-
-#include <sys/types.h>
-#include <stddef.h>  /* for wchar_t */
-#include <stdint.h>
-#include <time.h>
-
-#if defined(_WIN32) && !defined(__CYGWIN__)
-#include <windows.h>
-#endif
-
-/* Get a suitable 64-bit integer type. */
-#if !defined(__LA_INT64_T_DEFINED)
-# if ARCHIVE_VERSION_NUMBER < 4000000
-#define __LA_INT64_T la_int64_t
-# endif
-#define __LA_INT64_T_DEFINED
-# if defined(_WIN32) && !defined(__CYGWIN__) && !defined(__WATCOMC__)
-typedef __int64 la_int64_t;
-# else
-#include <unistd.h>
-#  if defined(_SCO_DS) || defined(__osf__)
-typedef long long la_int64_t;
-#  else
-typedef int64_t la_int64_t;
-#  endif
-# endif
-#endif
-
-/* The la_ssize_t should match the type used in 'struct stat' */
-#if !defined(__LA_SSIZE_T_DEFINED)
-/* Older code relied on the __LA_SSIZE_T macro; after 4.0 we'll switch to the typedef exclusively. */
-# if ARCHIVE_VERSION_NUMBER < 4000000
-#define __LA_SSIZE_T la_ssize_t
-# endif
-#define __LA_SSIZE_T_DEFINED
-# if defined(_WIN32) && !defined(__CYGWIN__) && !defined(__WATCOMC__)
-#  if defined(_SSIZE_T_DEFINED) || defined(_SSIZE_T_)
-typedef ssize_t la_ssize_t;
-#  elif defined(_WIN64)
-typedef __int64 la_ssize_t;
-#  else
-typedef long la_ssize_t;
-#  endif
-# else
-# include <unistd.h>  /* ssize_t */
-typedef ssize_t la_ssize_t;
-# endif
-#endif
-
-/* Get a suitable definition for mode_t */
-#if ARCHIVE_VERSION_NUMBER >= 3999000
-/* Switch to plain 'int' for libarchive 4.0.  It's less broken than 'mode_t' */
-# define	__LA_MODE_T	int
-#elif defined(_WIN32) && !defined(__CYGWIN__) && !defined(__BORLANDC__) && !defined(__WATCOMC__)
-# define	__LA_MODE_T	unsigned short
-#else
-# define	__LA_MODE_T	mode_t
-#endif
-
-/* Large file support for Android */
-#if defined(__LIBARCHIVE_BUILD) && defined(__ANDROID__)
-#include "android_lf.h"
-#endif
-
-/*
- * On Windows, define LIBARCHIVE_STATIC if you're building or using a
- * .lib.  The default here assumes you're building a DLL.  Only
- * libarchive source should ever define __LIBARCHIVE_BUILD.
- */
-#if ((defined __WIN32__) || (defined _WIN32) || defined(__CYGWIN__)) && (!defined LIBARCHIVE_STATIC)
-# ifdef __LIBARCHIVE_BUILD
-#  ifdef __GNUC__
-#   define __LA_DECL	__attribute__((dllexport)) extern
-#  else
-#   define __LA_DECL	__declspec(dllexport)
-#  endif
-# else
-#  ifdef __GNUC__
-#   define __LA_DECL
-#  else
-#   define __LA_DECL	__declspec(dllimport)
-#  endif
-# endif
-#elif defined __LIBARCHIVE_ENABLE_VISIBILITY
-#  define __LA_DECL __attribute__((visibility("default")))
-#else
-/* Static libraries on all platforms and shared libraries on non-Windows. */
-# define __LA_DECL
-#endif
-
-#if defined(__GNUC__) && __GNUC__ >= 3 && __GNUC_MINOR__ >= 1
-# define __LA_DEPRECATED __attribute__((deprecated))
-#else
-# define __LA_DEPRECATED
-#endif
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/*
- * Description of an archive entry.
- *
- * You can think of this as "struct stat" with some text fields added in.
- *
- * TODO: Add "comment", "charset", and possibly other entries that are
- * supported by "pax interchange" format.  However, GNU, ustar, cpio,
- * and other variants don't support these features, so they're not an
- * excruciatingly high priority right now.
- *
- * TODO: "pax interchange" format allows essentially arbitrary
- * key/value attributes to be attached to any entry.  Supporting
- * such extensions may make this library useful for special
- * applications (e.g., a package manager could attach special
- * package-management attributes to each entry).
- */
-struct archive;
-struct archive_entry;
-
-/*
- * File-type constants.  These are returned from archive_entry_filetype()
- * and passed to archive_entry_set_filetype().
- *
- * These values match S_XXX defines on every platform I've checked,
- * including Windows, AIX, Linux, Solaris, and BSD.  They're
- * (re)defined here because platforms generally don't define the ones
- * they don't support.  For example, Windows doesn't define S_IFLNK or
- * S_IFBLK.  Instead of having a mass of conditional logic and system
- * checks to define any S_XXX values that aren't supported locally,
- * I've just defined a new set of such constants so that
- * libarchive-based applications can manipulate and identify archive
- * entries properly even if the hosting platform can't store them on
- * disk.
- *
- * These values are also used directly within some portable formats,
- * such as cpio.  If you find a platform that varies from these, the
- * correct solution is to leave these alone and translate from these
- * portable values to platform-native values when entries are read from
- * or written to disk.
- */
-/*
- * In libarchive 4.0, we can drop the casts here.
- * They're needed to work around Borland C's broken mode_t.
- */
-#define AE_IFMT		((__LA_MODE_T)0170000)
-#define AE_IFREG	((__LA_MODE_T)0100000)
-#define AE_IFLNK	((__LA_MODE_T)0120000)
-#define AE_IFSOCK	((__LA_MODE_T)0140000)
-#define AE_IFCHR	((__LA_MODE_T)0020000)
-#define AE_IFBLK	((__LA_MODE_T)0060000)
-#define AE_IFDIR	((__LA_MODE_T)0040000)
-#define AE_IFIFO	((__LA_MODE_T)0010000)
-
-/*
- * Symlink types
- */
-#define AE_SYMLINK_TYPE_UNDEFINED	0
-#define AE_SYMLINK_TYPE_FILE		1
-#define AE_SYMLINK_TYPE_DIRECTORY	2
-
-/*
- * Basic object manipulation
- */
-
-__LA_DECL struct archive_entry	*archive_entry_clear(struct archive_entry *);
-/* The 'clone' function does a deep copy; all of the strings are copied too. */
-__LA_DECL struct archive_entry	*archive_entry_clone(struct archive_entry *);
-__LA_DECL void			 archive_entry_free(struct archive_entry *);
-__LA_DECL struct archive_entry	*archive_entry_new(void);
-
-/*
- * This form of archive_entry_new2() will pull character-set
- * conversion information from the specified archive handle.  The
- * older archive_entry_new(void) form is equivalent to calling
- * archive_entry_new2(NULL) and will result in the use of an internal
- * default character-set conversion.
- */
-__LA_DECL struct archive_entry	*archive_entry_new2(struct archive *);
-
-/*
- * Retrieve fields from an archive_entry.
- *
- * There are a number of implicit conversions among these fields.  For
- * example, if a regular string field is set and you read the _w wide
- * character field, the entry will implicitly convert narrow-to-wide
- * using the current locale.  Similarly, dev values are automatically
- * updated when you write devmajor or devminor and vice versa.
- *
- * In addition, fields can be "set" or "unset."  Unset string fields
- * return NULL, non-string fields have _is_set() functions to test
- * whether they've been set.  You can "unset" a string field by
- * assigning NULL; non-string fields have _unset() functions to
- * unset them.
- *
- * Note: There is one ambiguity in the above; string fields will
- * also return NULL when implicit character set conversions fail.
- * This is usually what you want.
- */
-__LA_DECL time_t	 archive_entry_atime(struct archive_entry *);
-__LA_DECL long		 archive_entry_atime_nsec(struct archive_entry *);
-__LA_DECL int		 archive_entry_atime_is_set(struct archive_entry *);
-__LA_DECL time_t	 archive_entry_birthtime(struct archive_entry *);
-__LA_DECL long		 archive_entry_birthtime_nsec(struct archive_entry *);
-__LA_DECL int		 archive_entry_birthtime_is_set(struct archive_entry *);
-__LA_DECL time_t	 archive_entry_ctime(struct archive_entry *);
-__LA_DECL long		 archive_entry_ctime_nsec(struct archive_entry *);
-__LA_DECL int		 archive_entry_ctime_is_set(struct archive_entry *);
-__LA_DECL dev_t		 archive_entry_dev(struct archive_entry *);
-__LA_DECL int		 archive_entry_dev_is_set(struct archive_entry *);
-__LA_DECL dev_t		 archive_entry_devmajor(struct archive_entry *);
-__LA_DECL dev_t		 archive_entry_devminor(struct archive_entry *);
-__LA_DECL __LA_MODE_T	 archive_entry_filetype(struct archive_entry *);
-__LA_DECL int		 archive_entry_filetype_is_set(struct archive_entry *);
-__LA_DECL void		 archive_entry_fflags(struct archive_entry *,
-			    unsigned long * /* set */,
-			    unsigned long * /* clear */);
-__LA_DECL const char	*archive_entry_fflags_text(struct archive_entry *);
-__LA_DECL la_int64_t	 archive_entry_gid(struct archive_entry *);
-__LA_DECL int		 archive_entry_gid_is_set(struct archive_entry *);
-__LA_DECL const char	*archive_entry_gname(struct archive_entry *);
-__LA_DECL const char	*archive_entry_gname_utf8(struct archive_entry *);
-__LA_DECL const wchar_t	*archive_entry_gname_w(struct archive_entry *);
-__LA_DECL void		 archive_entry_set_link_to_hardlink(struct archive_entry *);
-__LA_DECL const char	*archive_entry_hardlink(struct archive_entry *);
-__LA_DECL const char	*archive_entry_hardlink_utf8(struct archive_entry *);
-__LA_DECL const wchar_t	*archive_entry_hardlink_w(struct archive_entry *);
-__LA_DECL int		 archive_entry_hardlink_is_set(struct archive_entry *);
-__LA_DECL la_int64_t	 archive_entry_ino(struct archive_entry *);
-__LA_DECL la_int64_t	 archive_entry_ino64(struct archive_entry *);
-__LA_DECL int		 archive_entry_ino_is_set(struct archive_entry *);
-__LA_DECL __LA_MODE_T	 archive_entry_mode(struct archive_entry *);
-__LA_DECL time_t	 archive_entry_mtime(struct archive_entry *);
-__LA_DECL long		 archive_entry_mtime_nsec(struct archive_entry *);
-__LA_DECL int		 archive_entry_mtime_is_set(struct archive_entry *);
-__LA_DECL unsigned int	 archive_entry_nlink(struct archive_entry *);
-__LA_DECL const char	*archive_entry_pathname(struct archive_entry *);
-__LA_DECL const char	*archive_entry_pathname_utf8(struct archive_entry *);
-__LA_DECL const wchar_t	*archive_entry_pathname_w(struct archive_entry *);
-__LA_DECL __LA_MODE_T	 archive_entry_perm(struct archive_entry *);
-__LA_DECL int		 archive_entry_perm_is_set(struct archive_entry *);
-__LA_DECL int		 archive_entry_rdev_is_set(struct archive_entry *);
-__LA_DECL dev_t		 archive_entry_rdev(struct archive_entry *);
-__LA_DECL dev_t		 archive_entry_rdevmajor(struct archive_entry *);
-__LA_DECL dev_t		 archive_entry_rdevminor(struct archive_entry *);
-__LA_DECL const char	*archive_entry_sourcepath(struct archive_entry *);
-__LA_DECL const wchar_t	*archive_entry_sourcepath_w(struct archive_entry *);
-__LA_DECL la_int64_t	 archive_entry_size(struct archive_entry *);
-__LA_DECL int		 archive_entry_size_is_set(struct archive_entry *);
-__LA_DECL const char	*archive_entry_strmode(struct archive_entry *);
-__LA_DECL void		 archive_entry_set_link_to_symlink(struct archive_entry *);
-__LA_DECL const char	*archive_entry_symlink(struct archive_entry *);
-__LA_DECL const char	*archive_entry_symlink_utf8(struct archive_entry *);
-__LA_DECL int		 archive_entry_symlink_type(struct archive_entry *);
-__LA_DECL const wchar_t	*archive_entry_symlink_w(struct archive_entry *);
-__LA_DECL la_int64_t	 archive_entry_uid(struct archive_entry *);
-__LA_DECL int		 archive_entry_uid_is_set(struct archive_entry *);
-__LA_DECL const char	*archive_entry_uname(struct archive_entry *);
-__LA_DECL const char	*archive_entry_uname_utf8(struct archive_entry *);
-__LA_DECL const wchar_t	*archive_entry_uname_w(struct archive_entry *);
-__LA_DECL int archive_entry_is_data_encrypted(struct archive_entry *);
-__LA_DECL int archive_entry_is_metadata_encrypted(struct archive_entry *);
-__LA_DECL int archive_entry_is_encrypted(struct archive_entry *);
-
-/*
- * Set fields in an archive_entry.
- *
- * Note: Before libarchive 2.4, there were 'set' and 'copy' versions
- * of the string setters.  'copy' copied the actual string, 'set' just
- * stored the pointer.  In libarchive 2.4 and later, strings are
- * always copied.
- */
-
-__LA_DECL void	archive_entry_set_atime(struct archive_entry *, time_t, long);
-__LA_DECL void  archive_entry_unset_atime(struct archive_entry *);
-#if defined(_WIN32) && !defined(__CYGWIN__)
-__LA_DECL void archive_entry_copy_bhfi(struct archive_entry *, BY_HANDLE_FILE_INFORMATION *);
-#endif
-__LA_DECL void	archive_entry_set_birthtime(struct archive_entry *, time_t, long);
-__LA_DECL void  archive_entry_unset_birthtime(struct archive_entry *);
-__LA_DECL void	archive_entry_set_ctime(struct archive_entry *, time_t, long);
-__LA_DECL void  archive_entry_unset_ctime(struct archive_entry *);
-__LA_DECL void	archive_entry_set_dev(struct archive_entry *, dev_t);
-__LA_DECL void	archive_entry_set_devmajor(struct archive_entry *, dev_t);
-__LA_DECL void	archive_entry_set_devminor(struct archive_entry *, dev_t);
-__LA_DECL void	archive_entry_set_filetype(struct archive_entry *, unsigned int);
-__LA_DECL void	archive_entry_set_fflags(struct archive_entry *,
-	    unsigned long /* set */, unsigned long /* clear */);
-/* Returns pointer to start of first invalid token, or NULL if none. */
-/* Note that all recognized tokens are processed, regardless. */
-__LA_DECL const char *archive_entry_copy_fflags_text(struct archive_entry *,
-	    const char *);
-__LA_DECL const char *archive_entry_copy_fflags_text_len(struct archive_entry *,
-	    const char *, size_t);
-__LA_DECL const wchar_t *archive_entry_copy_fflags_text_w(struct archive_entry *,
-	    const wchar_t *);
-__LA_DECL void	archive_entry_set_gid(struct archive_entry *, la_int64_t);
-__LA_DECL void	archive_entry_set_gname(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_set_gname_utf8(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_copy_gname(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_copy_gname_w(struct archive_entry *, const wchar_t *);
-__LA_DECL int	archive_entry_update_gname_utf8(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_set_hardlink(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_set_hardlink_utf8(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_copy_hardlink(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_copy_hardlink_w(struct archive_entry *, const wchar_t *);
-__LA_DECL int	archive_entry_update_hardlink_utf8(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_set_ino(struct archive_entry *, la_int64_t);
-__LA_DECL void	archive_entry_set_ino64(struct archive_entry *, la_int64_t);
-__LA_DECL void	archive_entry_set_link(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_set_link_utf8(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_copy_link(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_copy_link_w(struct archive_entry *, const wchar_t *);
-__LA_DECL int	archive_entry_update_link_utf8(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_set_mode(struct archive_entry *, __LA_MODE_T);
-__LA_DECL void	archive_entry_set_mtime(struct archive_entry *, time_t, long);
-__LA_DECL void  archive_entry_unset_mtime(struct archive_entry *);
-__LA_DECL void	archive_entry_set_nlink(struct archive_entry *, unsigned int);
-__LA_DECL void	archive_entry_set_pathname(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_set_pathname_utf8(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_copy_pathname(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_copy_pathname_w(struct archive_entry *, const wchar_t *);
-__LA_DECL int	archive_entry_update_pathname_utf8(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_set_perm(struct archive_entry *, __LA_MODE_T);
-__LA_DECL void	archive_entry_set_rdev(struct archive_entry *, dev_t);
-__LA_DECL void	archive_entry_set_rdevmajor(struct archive_entry *, dev_t);
-__LA_DECL void	archive_entry_set_rdevminor(struct archive_entry *, dev_t);
-__LA_DECL void	archive_entry_set_size(struct archive_entry *, la_int64_t);
-__LA_DECL void	archive_entry_unset_size(struct archive_entry *);
-__LA_DECL void	archive_entry_copy_sourcepath(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_copy_sourcepath_w(struct archive_entry *, const wchar_t *);
-__LA_DECL void	archive_entry_set_symlink(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_set_symlink_type(struct archive_entry *, int);
-__LA_DECL void	archive_entry_set_symlink_utf8(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_copy_symlink(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_copy_symlink_w(struct archive_entry *, const wchar_t *);
-__LA_DECL int	archive_entry_update_symlink_utf8(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_set_uid(struct archive_entry *, la_int64_t);
-__LA_DECL void	archive_entry_set_uname(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_set_uname_utf8(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_copy_uname(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_copy_uname_w(struct archive_entry *, const wchar_t *);
-__LA_DECL int	archive_entry_update_uname_utf8(struct archive_entry *, const char *);
-__LA_DECL void	archive_entry_set_is_data_encrypted(struct archive_entry *, char is_encrypted);
-__LA_DECL void	archive_entry_set_is_metadata_encrypted(struct archive_entry *, char is_encrypted);
-/*
- * Routines to bulk copy fields to/from a platform-native "struct
- * stat."  Libarchive used to just store a struct stat inside of each
- * archive_entry object, but this created issues when trying to
- * manipulate archives on systems different than the ones they were
- * created on.
- *
- * TODO: On Linux and other LFS systems, provide both stat32 and
- * stat64 versions of these functions and all of the macro glue so
- * that archive_entry_stat is magically defined to
- * archive_entry_stat32 or archive_entry_stat64 as appropriate.
- */
-__LA_DECL const struct stat	*archive_entry_stat(struct archive_entry *);
-__LA_DECL void	archive_entry_copy_stat(struct archive_entry *, const struct stat *);
-
-/*
- * Storage for Mac OS-specific AppleDouble metadata information.
- * Apple-format tar files store a separate binary blob containing
- * encoded metadata with ACL, extended attributes, etc.
- * This provides a place to store that blob.
- */
-
-__LA_DECL const void * archive_entry_mac_metadata(struct archive_entry *, size_t *);
-__LA_DECL void archive_entry_copy_mac_metadata(struct archive_entry *, const void *, size_t);
-
-/*
- * Digest routine. This is used to query the raw hex digest for the
- * given entry. The type of digest is provided as an argument.
- */
-#define ARCHIVE_ENTRY_DIGEST_MD5              0x00000001
-#define ARCHIVE_ENTRY_DIGEST_RMD160           0x00000002
-#define ARCHIVE_ENTRY_DIGEST_SHA1             0x00000003
-#define ARCHIVE_ENTRY_DIGEST_SHA256           0x00000004
-#define ARCHIVE_ENTRY_DIGEST_SHA384           0x00000005
-#define ARCHIVE_ENTRY_DIGEST_SHA512           0x00000006
-
-__LA_DECL const unsigned char * archive_entry_digest(struct archive_entry *, int /* type */);
-
-/*
- * ACL routines.  This used to simply store and return text-format ACL
- * strings, but that proved insufficient for a number of reasons:
- *   = clients need control over uname/uid and gname/gid mappings
- *   = there are many different ACL text formats
- *   = would like to be able to read/convert archives containing ACLs
- *     on platforms that lack ACL libraries
- *
- *  This last point, in particular, forces me to implement a reasonably
- *  complete set of ACL support routines.
- */
-
-/*
- * Permission bits.
- */
-#define	ARCHIVE_ENTRY_ACL_EXECUTE             0x00000001
-#define	ARCHIVE_ENTRY_ACL_WRITE               0x00000002
-#define	ARCHIVE_ENTRY_ACL_READ                0x00000004
-#define	ARCHIVE_ENTRY_ACL_READ_DATA           0x00000008
-#define	ARCHIVE_ENTRY_ACL_LIST_DIRECTORY      0x00000008
-#define	ARCHIVE_ENTRY_ACL_WRITE_DATA          0x00000010
-#define	ARCHIVE_ENTRY_ACL_ADD_FILE            0x00000010
-#define	ARCHIVE_ENTRY_ACL_APPEND_DATA         0x00000020
-#define	ARCHIVE_ENTRY_ACL_ADD_SUBDIRECTORY    0x00000020
-#define	ARCHIVE_ENTRY_ACL_READ_NAMED_ATTRS    0x00000040
-#define	ARCHIVE_ENTRY_ACL_WRITE_NAMED_ATTRS   0x00000080
-#define	ARCHIVE_ENTRY_ACL_DELETE_CHILD        0x00000100
-#define	ARCHIVE_ENTRY_ACL_READ_ATTRIBUTES     0x00000200
-#define	ARCHIVE_ENTRY_ACL_WRITE_ATTRIBUTES    0x00000400
-#define	ARCHIVE_ENTRY_ACL_DELETE              0x00000800
-#define	ARCHIVE_ENTRY_ACL_READ_ACL            0x00001000
-#define	ARCHIVE_ENTRY_ACL_WRITE_ACL           0x00002000
-#define	ARCHIVE_ENTRY_ACL_WRITE_OWNER         0x00004000
-#define	ARCHIVE_ENTRY_ACL_SYNCHRONIZE         0x00008000
-
-#define	ARCHIVE_ENTRY_ACL_PERMS_POSIX1E			\
-	(ARCHIVE_ENTRY_ACL_EXECUTE			\
-	    | ARCHIVE_ENTRY_ACL_WRITE			\
-	    | ARCHIVE_ENTRY_ACL_READ)
-
-#define ARCHIVE_ENTRY_ACL_PERMS_NFS4			\
-	(ARCHIVE_ENTRY_ACL_EXECUTE			\
-	    | ARCHIVE_ENTRY_ACL_READ_DATA		\
-	    | ARCHIVE_ENTRY_ACL_LIST_DIRECTORY 		\
-	    | ARCHIVE_ENTRY_ACL_WRITE_DATA		\
-	    | ARCHIVE_ENTRY_ACL_ADD_FILE		\
-	    | ARCHIVE_ENTRY_ACL_APPEND_DATA		\
-	    | ARCHIVE_ENTRY_ACL_ADD_SUBDIRECTORY	\
-	    | ARCHIVE_ENTRY_ACL_READ_NAMED_ATTRS	\
-	    | ARCHIVE_ENTRY_ACL_WRITE_NAMED_ATTRS	\
-	    | ARCHIVE_ENTRY_ACL_DELETE_CHILD		\
-	    | ARCHIVE_ENTRY_ACL_READ_ATTRIBUTES		\
-	    | ARCHIVE_ENTRY_ACL_WRITE_ATTRIBUTES	\
-	    | ARCHIVE_ENTRY_ACL_DELETE			\
-	    | ARCHIVE_ENTRY_ACL_READ_ACL		\
-	    | ARCHIVE_ENTRY_ACL_WRITE_ACL		\
-	    | ARCHIVE_ENTRY_ACL_WRITE_OWNER		\
-	    | ARCHIVE_ENTRY_ACL_SYNCHRONIZE)
-
-/*
- * Inheritance values (NFS4 ACLs only); included in permset.
- */
-#define	ARCHIVE_ENTRY_ACL_ENTRY_INHERITED                   0x01000000
-#define	ARCHIVE_ENTRY_ACL_ENTRY_FILE_INHERIT                0x02000000
-#define	ARCHIVE_ENTRY_ACL_ENTRY_DIRECTORY_INHERIT           0x04000000
-#define	ARCHIVE_ENTRY_ACL_ENTRY_NO_PROPAGATE_INHERIT        0x08000000
-#define	ARCHIVE_ENTRY_ACL_ENTRY_INHERIT_ONLY                0x10000000
-#define	ARCHIVE_ENTRY_ACL_ENTRY_SUCCESSFUL_ACCESS           0x20000000
-#define	ARCHIVE_ENTRY_ACL_ENTRY_FAILED_ACCESS               0x40000000
-
-#define	ARCHIVE_ENTRY_ACL_INHERITANCE_NFS4			\
-	(ARCHIVE_ENTRY_ACL_ENTRY_FILE_INHERIT			\
-	    | ARCHIVE_ENTRY_ACL_ENTRY_DIRECTORY_INHERIT		\
-	    | ARCHIVE_ENTRY_ACL_ENTRY_NO_PROPAGATE_INHERIT	\
-	    | ARCHIVE_ENTRY_ACL_ENTRY_INHERIT_ONLY		\
-	    | ARCHIVE_ENTRY_ACL_ENTRY_SUCCESSFUL_ACCESS		\
-	    | ARCHIVE_ENTRY_ACL_ENTRY_FAILED_ACCESS		\
-	    | ARCHIVE_ENTRY_ACL_ENTRY_INHERITED)
-
-/* We need to be able to specify combinations of these. */
-#define	ARCHIVE_ENTRY_ACL_TYPE_ACCESS	0x00000100  /* POSIX.1e only */
-#define	ARCHIVE_ENTRY_ACL_TYPE_DEFAULT	0x00000200  /* POSIX.1e only */
-#define	ARCHIVE_ENTRY_ACL_TYPE_ALLOW	0x00000400 /* NFS4 only */
-#define	ARCHIVE_ENTRY_ACL_TYPE_DENY	0x00000800 /* NFS4 only */
-#define	ARCHIVE_ENTRY_ACL_TYPE_AUDIT	0x00001000 /* NFS4 only */
-#define	ARCHIVE_ENTRY_ACL_TYPE_ALARM	0x00002000 /* NFS4 only */
-#define	ARCHIVE_ENTRY_ACL_TYPE_POSIX1E	(ARCHIVE_ENTRY_ACL_TYPE_ACCESS \
-	    | ARCHIVE_ENTRY_ACL_TYPE_DEFAULT)
-#define	ARCHIVE_ENTRY_ACL_TYPE_NFS4	(ARCHIVE_ENTRY_ACL_TYPE_ALLOW \
-	    | ARCHIVE_ENTRY_ACL_TYPE_DENY \
-	    | ARCHIVE_ENTRY_ACL_TYPE_AUDIT \
-	    | ARCHIVE_ENTRY_ACL_TYPE_ALARM)
-
-/* Tag values mimic POSIX.1e */
-#define	ARCHIVE_ENTRY_ACL_USER		10001	/* Specified user. */
-#define	ARCHIVE_ENTRY_ACL_USER_OBJ 	10002	/* User who owns the file. */
-#define	ARCHIVE_ENTRY_ACL_GROUP		10003	/* Specified group. */
-#define	ARCHIVE_ENTRY_ACL_GROUP_OBJ	10004	/* Group who owns the file. */
-#define	ARCHIVE_ENTRY_ACL_MASK		10005	/* Modify group access (POSIX.1e only) */
-#define	ARCHIVE_ENTRY_ACL_OTHER		10006	/* Public (POSIX.1e only) */
-#define	ARCHIVE_ENTRY_ACL_EVERYONE	10107   /* Everyone (NFS4 only) */
-
-/*
- * Set the ACL by clearing it and adding entries one at a time.
- * Unlike the POSIX.1e ACL routines, you must specify the type
- * (access/default) for each entry.  Internally, the ACL data is just
- * a soup of entries.  API calls here allow you to retrieve just the
- * entries of interest.  This design (which goes against the spirit of
- * POSIX.1e) is useful for handling archive formats that combine
- * default and access information in a single ACL list.
- */
-__LA_DECL void	 archive_entry_acl_clear(struct archive_entry *);
-__LA_DECL int	 archive_entry_acl_add_entry(struct archive_entry *,
-	    int /* type */, int /* permset */, int /* tag */,
-	    int /* qual */, const char * /* name */);
-__LA_DECL int	 archive_entry_acl_add_entry_w(struct archive_entry *,
-	    int /* type */, int /* permset */, int /* tag */,
-	    int /* qual */, const wchar_t * /* name */);
-
-/*
- * To retrieve the ACL, first "reset", then repeatedly ask for the
- * "next" entry.  The want_type parameter allows you to request only
- * certain types of entries.
- */
-__LA_DECL int	 archive_entry_acl_reset(struct archive_entry *, int /* want_type */);
-__LA_DECL int	 archive_entry_acl_next(struct archive_entry *, int /* want_type */,
-	    int * /* type */, int * /* permset */, int * /* tag */,
-	    int * /* qual */, const char ** /* name */);
-
-/*
- * Construct a text-format ACL.  The flags argument is a bitmask that
- * can include any of the following:
- *
- * Flags only for archive entries with POSIX.1e ACL:
- * ARCHIVE_ENTRY_ACL_TYPE_ACCESS - Include POSIX.1e "access" entries.
- * ARCHIVE_ENTRY_ACL_TYPE_DEFAULT - Include POSIX.1e "default" entries.
- * ARCHIVE_ENTRY_ACL_STYLE_MARK_DEFAULT - Include "default:" before each
- *    default ACL entry.
- * ARCHIVE_ENTRY_ACL_STYLE_SOLARIS - Output only one colon after "other" and
- *    "mask" entries.
- *
- * Flags only for archive entries with NFSv4 ACL:
- * ARCHIVE_ENTRY_ACL_STYLE_COMPACT - Do not output the minus character for
- *    unset permissions and flags in NFSv4 ACL permission and flag fields
- *
- * Flags for for archive entries with POSIX.1e ACL or NFSv4 ACL:
- * ARCHIVE_ENTRY_ACL_STYLE_EXTRA_ID - Include extra numeric ID field in
- *    each ACL entry.
- * ARCHIVE_ENTRY_ACL_STYLE_SEPARATOR_COMMA - Separate entries with comma
- *    instead of newline.
- */
-#define	ARCHIVE_ENTRY_ACL_STYLE_EXTRA_ID	0x00000001
-#define	ARCHIVE_ENTRY_ACL_STYLE_MARK_DEFAULT	0x00000002
-#define	ARCHIVE_ENTRY_ACL_STYLE_SOLARIS		0x00000004
-#define	ARCHIVE_ENTRY_ACL_STYLE_SEPARATOR_COMMA	0x00000008
-#define	ARCHIVE_ENTRY_ACL_STYLE_COMPACT		0x00000010
-
-__LA_DECL wchar_t *archive_entry_acl_to_text_w(struct archive_entry *,
-	    la_ssize_t * /* len */, int /* flags */);
-__LA_DECL char *archive_entry_acl_to_text(struct archive_entry *,
-	    la_ssize_t * /* len */, int /* flags */);
-__LA_DECL int archive_entry_acl_from_text_w(struct archive_entry *,
-	    const wchar_t * /* wtext */, int /* type */);
-__LA_DECL int archive_entry_acl_from_text(struct archive_entry *,
-	    const char * /* text */, int /* type */);
-
-/* Deprecated constants */
-#define	OLD_ARCHIVE_ENTRY_ACL_STYLE_EXTRA_ID		1024
-#define	OLD_ARCHIVE_ENTRY_ACL_STYLE_MARK_DEFAULT	2048
-
-/* Deprecated functions */
-__LA_DECL const wchar_t	*archive_entry_acl_text_w(struct archive_entry *,
-		    int /* flags */) __LA_DEPRECATED;
-__LA_DECL const char *archive_entry_acl_text(struct archive_entry *,
-		    int /* flags */) __LA_DEPRECATED;
-
-/* Return bitmask of ACL types in an archive entry */
-__LA_DECL int	 archive_entry_acl_types(struct archive_entry *);
-
-/* Return a count of entries matching 'want_type' */
-__LA_DECL int	 archive_entry_acl_count(struct archive_entry *, int /* want_type */);
-
-/* Return an opaque ACL object. */
-/* There's not yet anything clients can actually do with this... */
-struct archive_acl;
-__LA_DECL struct archive_acl *archive_entry_acl(struct archive_entry *);
-
-/*
- * extended attributes
- */
-
-__LA_DECL void	 archive_entry_xattr_clear(struct archive_entry *);
-__LA_DECL void	 archive_entry_xattr_add_entry(struct archive_entry *,
-	    const char * /* name */, const void * /* value */,
-	    size_t /* size */);
-
-/*
- * To retrieve the xattr list, first "reset", then repeatedly ask for the
- * "next" entry.
- */
-
-__LA_DECL int	archive_entry_xattr_count(struct archive_entry *);
-__LA_DECL int	archive_entry_xattr_reset(struct archive_entry *);
-__LA_DECL int	archive_entry_xattr_next(struct archive_entry *,
-	    const char ** /* name */, const void ** /* value */, size_t *);
-
-/*
- * sparse
- */
-
-__LA_DECL void	 archive_entry_sparse_clear(struct archive_entry *);
-__LA_DECL void	 archive_entry_sparse_add_entry(struct archive_entry *,
-	    la_int64_t /* offset */, la_int64_t /* length */);
-
-/*
- * To retrieve the xattr list, first "reset", then repeatedly ask for the
- * "next" entry.
- */
-
-__LA_DECL int	archive_entry_sparse_count(struct archive_entry *);
-__LA_DECL int	archive_entry_sparse_reset(struct archive_entry *);
-__LA_DECL int	archive_entry_sparse_next(struct archive_entry *,
-	    la_int64_t * /* offset */, la_int64_t * /* length */);
-
-/*
- * Utility to match up hardlinks.
- *
- * The 'struct archive_entry_linkresolver' is a cache of archive entries
- * for files with multiple links.  Here's how to use it:
- *   1. Create a lookup object with archive_entry_linkresolver_new()
- *   2. Tell it the archive format you're using.
- *   3. Hand each archive_entry to archive_entry_linkify().
- *      That function will return 0, 1, or 2 entries that should
- *      be written.
- *   4. Call archive_entry_linkify(resolver, NULL) until
- *      no more entries are returned.
- *   5. Call archive_entry_linkresolver_free(resolver) to free resources.
- *
- * The entries returned have their hardlink and size fields updated
- * appropriately.  If an entry is passed in that does not refer to
- * a file with multiple links, it is returned unchanged.  The intention
- * is that you should be able to simply filter all entries through
- * this machine.
- *
- * To make things more efficient, be sure that each entry has a valid
- * nlinks value.  The hardlink cache uses this to track when all links
- * have been found.  If the nlinks value is zero, it will keep every
- * name in the cache indefinitely, which can use a lot of memory.
- *
- * Note that archive_entry_size() is reset to zero if the file
- * body should not be written to the archive.  Pay attention!
- */
-struct archive_entry_linkresolver;
-
-/*
- * There are three different strategies for marking hardlinks.
- * The descriptions below name them after the best-known
- * formats that rely on each strategy:
- *
- * "Old cpio" is the simplest, it always returns any entry unmodified.
- *    As far as I know, only cpio formats use this.  Old cpio archives
- *    store every link with the full body; the onus is on the dearchiver
- *    to detect and properly link the files as they are restored.
- * "tar" is also pretty simple; it caches a copy the first time it sees
- *    any link.  Subsequent appearances are modified to be hardlink
- *    references to the first one without any body.  Used by all tar
- *    formats, although the newest tar formats permit the "old cpio" strategy
- *    as well.  This strategy is very simple for the dearchiver,
- *    and reasonably straightforward for the archiver.
- * "new cpio" is trickier.  It stores the body only with the last
- *    occurrence.  The complication is that we might not
- *    see every link to a particular file in a single session, so
- *    there's no easy way to know when we've seen the last occurrence.
- *    The solution here is to queue one link until we see the next.
- *    At the end of the session, you can enumerate any remaining
- *    entries by calling archive_entry_linkify(NULL) and store those
- *    bodies.  If you have a file with three links l1, l2, and l3,
- *    you'll get the following behavior if you see all three links:
- *           linkify(l1) => NULL   (the resolver stores l1 internally)
- *           linkify(l2) => l1     (resolver stores l2, you write l1)
- *           linkify(l3) => l2, l3 (all links seen, you can write both).
- *    If you only see l1 and l2, you'll get this behavior:
- *           linkify(l1) => NULL
- *           linkify(l2) => l1
- *           linkify(NULL) => l2   (at end, you retrieve remaining links)
- *    As the name suggests, this strategy is used by newer cpio variants.
- *    It's noticeably more complex for the archiver, slightly more complex
- *    for the dearchiver than the tar strategy, but makes it straightforward
- *    to restore a file using any link by simply continuing to scan until
- *    you see a link that is stored with a body.  In contrast, the tar
- *    strategy requires you to rescan the archive from the beginning to
- *    correctly extract an arbitrary link.
- */
-
-__LA_DECL struct archive_entry_linkresolver *archive_entry_linkresolver_new(void);
-__LA_DECL void archive_entry_linkresolver_set_strategy(
-	struct archive_entry_linkresolver *, int /* format_code */);
-__LA_DECL void archive_entry_linkresolver_free(struct archive_entry_linkresolver *);
-__LA_DECL void archive_entry_linkify(struct archive_entry_linkresolver *,
-    struct archive_entry **, struct archive_entry **);
-__LA_DECL struct archive_entry *archive_entry_partial_links(
-    struct archive_entry_linkresolver *res, unsigned int *links);
-#ifdef __cplusplus
-}
-#endif
-
-/* This is meaningless outside of this header. */
-#undef __LA_DECL
-
-#endif /* !ARCHIVE_ENTRY_H_INCLUDED */
diff --git a/projects/llm_framework/include/dwarf.h b/projects/llm_framework/include/dwarf.h
deleted file mode 100644
index 19a4be96..00000000
--- a/projects/llm_framework/include/dwarf.h
+++ /dev/null
@@ -1,1029 +0,0 @@
-/* This file defines standard DWARF types, structures, and macros.
-   Copyright (C) 2000-2011, 2014, 2016, 2017, 2018 Red Hat, Inc.
-   This file is part of elfutils.
-
-   This file is free software; you can redistribute it and/or modify
-   it under the terms of either
-
-     * the GNU Lesser General Public License as published by the Free
-       Software Foundation; either version 3 of the License, or (at
-       your option) any later version
-
-   or
-
-     * the GNU General Public License as published by the Free
-       Software Foundation; either version 2 of the License, or (at
-       your option) any later version
-
-   or both in parallel, as here.
-
-   elfutils is distributed in the hope that it will be useful, but
-   WITHOUT ANY WARRANTY; without even the implied warranty of
-   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
-   General Public License for more details.
-
-   You should have received copies of the GNU General Public License and
-   the GNU Lesser General Public License along with this program.  If
-   not, see <http://www.gnu.org/licenses/>.  */
-
-#ifndef _DWARF_H
-#define	_DWARF_H 1
-
-/* DWARF Unit Header Types.  */
-enum
-  {
-    DW_UT_compile = 0x01,
-    DW_UT_type = 0x02,
-    DW_UT_partial = 0x03,
-    DW_UT_skeleton = 0x04,
-    DW_UT_split_compile = 0x05,
-    DW_UT_split_type = 0x06,
-
-    DW_UT_lo_user = 0x80,
-    DW_UT_hi_user = 0xff
-  };
-
-/* DWARF tags.  */
-enum
-  {
-    DW_TAG_array_type = 0x01,
-    DW_TAG_class_type = 0x02,
-    DW_TAG_entry_point = 0x03,
-    DW_TAG_enumeration_type = 0x04,
-    DW_TAG_formal_parameter = 0x05,
-    /* 0x06 reserved.  */
-    /* 0x07 reserved.  */
-    DW_TAG_imported_declaration = 0x08,
-    /* 0x09 reserved.  */
-    DW_TAG_label = 0x0a,
-    DW_TAG_lexical_block = 0x0b,
-    /* 0x0c reserved.  */
-    DW_TAG_member = 0x0d,
-    /* 0x0e reserved.  */
-    DW_TAG_pointer_type = 0x0f,
-    DW_TAG_reference_type = 0x10,
-    DW_TAG_compile_unit = 0x11,
-    DW_TAG_string_type = 0x12,
-    DW_TAG_structure_type = 0x13,
-    /* 0x14 reserved.  */
-    DW_TAG_subroutine_type = 0x15,
-    DW_TAG_typedef = 0x16,
-    DW_TAG_union_type = 0x17,
-    DW_TAG_unspecified_parameters = 0x18,
-    DW_TAG_variant = 0x19,
-    DW_TAG_common_block = 0x1a,
-    DW_TAG_common_inclusion = 0x1b,
-    DW_TAG_inheritance = 0x1c,
-    DW_TAG_inlined_subroutine = 0x1d,
-    DW_TAG_module = 0x1e,
-    DW_TAG_ptr_to_member_type = 0x1f,
-    DW_TAG_set_type = 0x20,
-    DW_TAG_subrange_type = 0x21,
-    DW_TAG_with_stmt = 0x22,
-    DW_TAG_access_declaration = 0x23,
-    DW_TAG_base_type = 0x24,
-    DW_TAG_catch_block = 0x25,
-    DW_TAG_const_type = 0x26,
-    DW_TAG_constant = 0x27,
-    DW_TAG_enumerator = 0x28,
-    DW_TAG_file_type = 0x29,
-    DW_TAG_friend = 0x2a,
-    DW_TAG_namelist = 0x2b,
-    DW_TAG_namelist_item = 0x2c,
-    DW_TAG_packed_type = 0x2d,
-    DW_TAG_subprogram = 0x2e,
-    DW_TAG_template_type_parameter = 0x2f,
-    DW_TAG_template_value_parameter = 0x30,
-    DW_TAG_thrown_type = 0x31,
-    DW_TAG_try_block = 0x32,
-    DW_TAG_variant_part = 0x33,
-    DW_TAG_variable = 0x34,
-    DW_TAG_volatile_type = 0x35,
-    DW_TAG_dwarf_procedure = 0x36,
-    DW_TAG_restrict_type = 0x37,
-    DW_TAG_interface_type = 0x38,
-    DW_TAG_namespace = 0x39,
-    DW_TAG_imported_module = 0x3a,
-    DW_TAG_unspecified_type = 0x3b,
-    DW_TAG_partial_unit = 0x3c,
-    DW_TAG_imported_unit = 0x3d,
-    /* 0x3e reserved.  Was DW_TAG_mutable_type.  */
-    DW_TAG_condition = 0x3f,
-    DW_TAG_shared_type = 0x40,
-    DW_TAG_type_unit = 0x41,
-    DW_TAG_rvalue_reference_type = 0x42,
-    DW_TAG_template_alias = 0x43,
-    DW_TAG_coarray_type = 0x44,
-    DW_TAG_generic_subrange = 0x45,
-    DW_TAG_dynamic_type = 0x46,
-    DW_TAG_atomic_type = 0x47,
-    DW_TAG_call_site = 0x48,
-    DW_TAG_call_site_parameter = 0x49,
-    DW_TAG_skeleton_unit = 0x4a,
-    DW_TAG_immutable_type = 0x4b,
-
-    DW_TAG_lo_user = 0x4080,
-
-    DW_TAG_MIPS_loop = 0x4081,
-    DW_TAG_format_label = 0x4101,
-    DW_TAG_function_template = 0x4102,
-    DW_TAG_class_template = 0x4103,
-
-    DW_TAG_GNU_BINCL = 0x4104,
-    DW_TAG_GNU_EINCL = 0x4105,
-
-    DW_TAG_GNU_template_template_param = 0x4106,
-    DW_TAG_GNU_template_parameter_pack = 0x4107,
-    DW_TAG_GNU_formal_parameter_pack = 0x4108,
-    DW_TAG_GNU_call_site = 0x4109,
-    DW_TAG_GNU_call_site_parameter = 0x410a,
-
-    DW_TAG_hi_user = 0xffff
-  };
-
-
-/* Children determination encodings.  */
-enum
-  {
-    DW_CHILDREN_no = 0,
-    DW_CHILDREN_yes = 1
-  };
-
-
-/* DWARF attributes encodings.  */
-enum
-  {
-    DW_AT_sibling = 0x01,
-    DW_AT_location = 0x02,
-    DW_AT_name = 0x03,
-    /* 0x04 reserved.  */
-    /* 0x05 reserved.  */
-    /* 0x06 reserved.  */
-    /* 0x07 reserved.  */
-    /* 0x08 reserved.  */
-    DW_AT_ordering = 0x09,
-    /* 0x0a reserved.  */
-    DW_AT_byte_size = 0x0b,
-    DW_AT_bit_offset = 0x0c,  /* Deprecated in DWARF4.  */
-    DW_AT_bit_size = 0x0d,
-    /* 0x0e reserved.  */
-    /* 0x0f reserved.  */
-    DW_AT_stmt_list = 0x10,
-    DW_AT_low_pc = 0x11,
-    DW_AT_high_pc = 0x12,
-    DW_AT_language = 0x13,
-    /* 0x14 reserved.  */
-    DW_AT_discr = 0x15,
-    DW_AT_discr_value = 0x16,
-    DW_AT_visibility = 0x17,
-    DW_AT_import = 0x18,
-    DW_AT_string_length = 0x19,
-    DW_AT_common_reference = 0x1a,
-    DW_AT_comp_dir = 0x1b,
-    DW_AT_const_value = 0x1c,
-    DW_AT_containing_type = 0x1d,
-    DW_AT_default_value = 0x1e,
-    /* 0x1f reserved.  */
-    DW_AT_inline = 0x20,
-    DW_AT_is_optional = 0x21,
-    DW_AT_lower_bound = 0x22,
-    /* 0x23 reserved.  */
-    /* 0x24 reserved.  */
-    DW_AT_producer = 0x25,
-    /* 0x26 reserved.  */
-    DW_AT_prototyped = 0x27,
-    /* 0x28 reserved.  */
-    /* 0x29 reserved.  */
-    DW_AT_return_addr = 0x2a,
-    /* 0x2b reserved.  */
-    DW_AT_start_scope = 0x2c,
-    /* 0x2d reserved.  */
-    DW_AT_bit_stride = 0x2e,
-    DW_AT_upper_bound = 0x2f,
-    /* 0x30 reserved.  */
-    DW_AT_abstract_origin = 0x31,
-    DW_AT_accessibility = 0x32,
-    DW_AT_address_class = 0x33,
-    DW_AT_artificial = 0x34,
-    DW_AT_base_types = 0x35,
-    DW_AT_calling_convention = 0x36,
-    DW_AT_count = 0x37,
-    DW_AT_data_member_location = 0x38,
-    DW_AT_decl_column = 0x39,
-    DW_AT_decl_file = 0x3a,
-    DW_AT_decl_line = 0x3b,
-    DW_AT_declaration = 0x3c,
-    DW_AT_discr_list = 0x3d,
-    DW_AT_encoding = 0x3e,
-    DW_AT_external = 0x3f,
-    DW_AT_frame_base = 0x40,
-    DW_AT_friend = 0x41,
-    DW_AT_identifier_case = 0x42,
-    DW_AT_macro_info = 0x43, /* Deprecated in DWARF5.  */
-    DW_AT_namelist_item = 0x44,
-    DW_AT_priority = 0x45,
-    DW_AT_segment = 0x46,
-    DW_AT_specification = 0x47,
-    DW_AT_static_link = 0x48,
-    DW_AT_type = 0x49,
-    DW_AT_use_location = 0x4a,
-    DW_AT_variable_parameter = 0x4b,
-    DW_AT_virtuality = 0x4c,
-    DW_AT_vtable_elem_location = 0x4d,
-    DW_AT_allocated = 0x4e,
-    DW_AT_associated = 0x4f,
-    DW_AT_data_location = 0x50,
-    DW_AT_byte_stride = 0x51,
-    DW_AT_entry_pc = 0x52,
-    DW_AT_use_UTF8 = 0x53,
-    DW_AT_extension = 0x54,
-    DW_AT_ranges = 0x55,
-    DW_AT_trampoline = 0x56,
-    DW_AT_call_column = 0x57,
-    DW_AT_call_file = 0x58,
-    DW_AT_call_line = 0x59,
-    DW_AT_description = 0x5a,
-    DW_AT_binary_scale = 0x5b,
-    DW_AT_decimal_scale = 0x5c,
-    DW_AT_small = 0x5d,
-    DW_AT_decimal_sign = 0x5e,
-    DW_AT_digit_count = 0x5f,
-    DW_AT_picture_string = 0x60,
-    DW_AT_mutable = 0x61,
-    DW_AT_threads_scaled = 0x62,
-    DW_AT_explicit = 0x63,
-    DW_AT_object_pointer = 0x64,
-    DW_AT_endianity = 0x65,
-    DW_AT_elemental = 0x66,
-    DW_AT_pure = 0x67,
-    DW_AT_recursive = 0x68,
-    DW_AT_signature = 0x69,
-    DW_AT_main_subprogram = 0x6a,
-    DW_AT_data_bit_offset = 0x6b,
-    DW_AT_const_expr = 0x6c,
-    DW_AT_enum_class = 0x6d,
-    DW_AT_linkage_name = 0x6e,
-    DW_AT_string_length_bit_size = 0x6f,
-    DW_AT_string_length_byte_size = 0x70,
-    DW_AT_rank = 0x71,
-    DW_AT_str_offsets_base = 0x72,
-    DW_AT_addr_base = 0x73,
-    DW_AT_rnglists_base = 0x74,
-    /* 0x75 reserved.  */
-    DW_AT_dwo_name = 0x76,
-    DW_AT_reference = 0x77,
-    DW_AT_rvalue_reference = 0x78,
-    DW_AT_macros = 0x79,
-    DW_AT_call_all_calls = 0x7a,
-    DW_AT_call_all_source_calls = 0x7b,
-    DW_AT_call_all_tail_calls = 0x7c,
-    DW_AT_call_return_pc = 0x7d,
-    DW_AT_call_value = 0x7e,
-    DW_AT_call_origin = 0x7f,
-    DW_AT_call_parameter = 0x80,
-    DW_AT_call_pc = 0x81,
-    DW_AT_call_tail_call = 0x82,
-    DW_AT_call_target = 0x83,
-    DW_AT_call_target_clobbered = 0x84,
-    DW_AT_call_data_location = 0x85,
-    DW_AT_call_data_value = 0x86,
-    DW_AT_noreturn = 0x87,
-    DW_AT_alignment = 0x88,
-    DW_AT_export_symbols = 0x89,
-    DW_AT_deleted = 0x8a,
-    DW_AT_defaulted = 0x8b,
-    DW_AT_loclists_base = 0x8c,
-
-    DW_AT_lo_user = 0x2000,
-
-    DW_AT_MIPS_fde = 0x2001,
-    DW_AT_MIPS_loop_begin = 0x2002,
-    DW_AT_MIPS_tail_loop_begin = 0x2003,
-    DW_AT_MIPS_epilog_begin = 0x2004,
-    DW_AT_MIPS_loop_unroll_factor = 0x2005,
-    DW_AT_MIPS_software_pipeline_depth = 0x2006,
-    DW_AT_MIPS_linkage_name = 0x2007,
-    DW_AT_MIPS_stride = 0x2008,
-    DW_AT_MIPS_abstract_name = 0x2009,
-    DW_AT_MIPS_clone_origin = 0x200a,
-    DW_AT_MIPS_has_inlines = 0x200b,
-    DW_AT_MIPS_stride_byte = 0x200c,
-    DW_AT_MIPS_stride_elem = 0x200d,
-    DW_AT_MIPS_ptr_dopetype = 0x200e,
-    DW_AT_MIPS_allocatable_dopetype = 0x200f,
-    DW_AT_MIPS_assumed_shape_dopetype = 0x2010,
-    DW_AT_MIPS_assumed_size = 0x2011,
-
-    /* GNU extensions.  */
-    DW_AT_sf_names = 0x2101,
-    DW_AT_src_info = 0x2102,
-    DW_AT_mac_info = 0x2103,
-    DW_AT_src_coords = 0x2104,
-    DW_AT_body_begin = 0x2105,
-    DW_AT_body_end = 0x2106,
-    DW_AT_GNU_vector = 0x2107,
-    DW_AT_GNU_guarded_by = 0x2108,
-    DW_AT_GNU_pt_guarded_by = 0x2109,
-    DW_AT_GNU_guarded = 0x210a,
-    DW_AT_GNU_pt_guarded = 0x210b,
-    DW_AT_GNU_locks_excluded = 0x210c,
-    DW_AT_GNU_exclusive_locks_required = 0x210d,
-    DW_AT_GNU_shared_locks_required = 0x210e,
-    DW_AT_GNU_odr_signature = 0x210f,
-    DW_AT_GNU_template_name = 0x2110,
-    DW_AT_GNU_call_site_value = 0x2111,
-    DW_AT_GNU_call_site_data_value = 0x2112,
-    DW_AT_GNU_call_site_target = 0x2113,
-    DW_AT_GNU_call_site_target_clobbered = 0x2114,
-    DW_AT_GNU_tail_call = 0x2115,
-    DW_AT_GNU_all_tail_call_sites = 0x2116,
-    DW_AT_GNU_all_call_sites = 0x2117,
-    DW_AT_GNU_all_source_call_sites = 0x2118,
-    DW_AT_GNU_locviews = 0x2137,
-    DW_AT_GNU_entry_view = 0x2138,
-    DW_AT_GNU_macros = 0x2119,
-    DW_AT_GNU_deleted = 0x211a,
-    /* GNU Debug Fission extensions.  */
-    DW_AT_GNU_dwo_name = 0x2130,
-    DW_AT_GNU_dwo_id = 0x2131,
-    DW_AT_GNU_ranges_base = 0x2132,
-    DW_AT_GNU_addr_base = 0x2133,
-    DW_AT_GNU_pubnames = 0x2134,
-    DW_AT_GNU_pubtypes = 0x2135,
-
-    /* https://gcc.gnu.org/wiki/DW_AT_GNU_numerator_denominator  */
-    DW_AT_GNU_numerator = 0x2303,
-    DW_AT_GNU_denominator = 0x2304,
-    /* https://gcc.gnu.org/wiki/DW_AT_GNU_bias  */
-    DW_AT_GNU_bias = 0x2305,
-
-    DW_AT_hi_user = 0x3fff
-  };
-
-/* Old unofficially attribute names.  Should not be used.
-   Will not appear in known-dwarf.h  */
-
-/* DWARF1 array subscripts and element data types.  */
-#define DW_AT_subscr_data	0x0a
-/* DWARF1 enumeration literals.  */
-#define DW_AT_element_list	0x0f
-/* DWARF1 reference for variable to member structure, class or union.  */
-#define DW_AT_member		0x14
-
-/* DWARF form encodings.  */
-enum
-  {
-    DW_FORM_addr = 0x01,
-    DW_FORM_block2 = 0x03,
-    DW_FORM_block4 = 0x04,
-    DW_FORM_data2 = 0x05,
-    DW_FORM_data4 = 0x06,
-    DW_FORM_data8 = 0x07,
-    DW_FORM_string = 0x08,
-    DW_FORM_block = 0x09,
-    DW_FORM_block1 = 0x0a,
-    DW_FORM_data1 = 0x0b,
-    DW_FORM_flag = 0x0c,
-    DW_FORM_sdata = 0x0d,
-    DW_FORM_strp = 0x0e,
-    DW_FORM_udata = 0x0f,
-    DW_FORM_ref_addr = 0x10,
-    DW_FORM_ref1 = 0x11,
-    DW_FORM_ref2 = 0x12,
-    DW_FORM_ref4 = 0x13,
-    DW_FORM_ref8 = 0x14,
-    DW_FORM_ref_udata = 0x15,
-    DW_FORM_indirect = 0x16,
-    DW_FORM_sec_offset = 0x17,
-    DW_FORM_exprloc = 0x18,
-    DW_FORM_flag_present = 0x19,
-    DW_FORM_strx = 0x1a,
-    DW_FORM_addrx = 0x1b,
-    DW_FORM_ref_sup4 = 0x1c,
-    DW_FORM_strp_sup = 0x1d,
-    DW_FORM_data16 = 0x1e,
-    DW_FORM_line_strp = 0x1f,
-    DW_FORM_ref_sig8 = 0x20,
-    DW_FORM_implicit_const = 0x21,
-    DW_FORM_loclistx = 0x22,
-    DW_FORM_rnglistx = 0x23,
-    DW_FORM_ref_sup8 = 0x24,
-    DW_FORM_strx1 = 0x25,
-    DW_FORM_strx2 = 0x26,
-    DW_FORM_strx3 = 0x27,
-    DW_FORM_strx4 = 0x28,
-    DW_FORM_addrx1 = 0x29,
-    DW_FORM_addrx2 = 0x2a,
-    DW_FORM_addrx3 = 0x2b,
-    DW_FORM_addrx4 = 0x2c,
-
-    /* GNU Debug Fission extensions.  */
-    DW_FORM_GNU_addr_index = 0x1f01,
-    DW_FORM_GNU_str_index = 0x1f02,
-
-    DW_FORM_GNU_ref_alt = 0x1f20, /* offset in alternate .debuginfo.  */
-    DW_FORM_GNU_strp_alt = 0x1f21 /* offset in alternate .debug_str. */
-  };
-
-
-/* DWARF location operation encodings.  */
-enum
-  {
-    DW_OP_addr = 0x03,		/* Constant address.  */
-    DW_OP_deref = 0x06,
-    DW_OP_const1u = 0x08,	/* Unsigned 1-byte constant.  */
-    DW_OP_const1s = 0x09,	/* Signed 1-byte constant.  */
-    DW_OP_const2u = 0x0a,	/* Unsigned 2-byte constant.  */
-    DW_OP_const2s = 0x0b,	/* Signed 2-byte constant.  */
-    DW_OP_const4u = 0x0c,	/* Unsigned 4-byte constant.  */
-    DW_OP_const4s = 0x0d,	/* Signed 4-byte constant.  */
-    DW_OP_const8u = 0x0e,	/* Unsigned 8-byte constant.  */
-    DW_OP_const8s = 0x0f,	/* Signed 8-byte constant.  */
-    DW_OP_constu = 0x10,	/* Unsigned LEB128 constant.  */
-    DW_OP_consts = 0x11,	/* Signed LEB128 constant.  */
-    DW_OP_dup = 0x12,
-    DW_OP_drop = 0x13,
-    DW_OP_over = 0x14,
-    DW_OP_pick = 0x15,		/* 1-byte stack index.  */
-    DW_OP_swap = 0x16,
-    DW_OP_rot = 0x17,
-    DW_OP_xderef = 0x18,
-    DW_OP_abs = 0x19,
-    DW_OP_and = 0x1a,
-    DW_OP_div = 0x1b,
-    DW_OP_minus = 0x1c,
-    DW_OP_mod = 0x1d,
-    DW_OP_mul = 0x1e,
-    DW_OP_neg = 0x1f,
-    DW_OP_not = 0x20,
-    DW_OP_or = 0x21,
-    DW_OP_plus = 0x22,
-    DW_OP_plus_uconst = 0x23,	/* Unsigned LEB128 addend.  */
-    DW_OP_shl = 0x24,
-    DW_OP_shr = 0x25,
-    DW_OP_shra = 0x26,
-    DW_OP_xor = 0x27,
-    DW_OP_bra = 0x28,		/* Signed 2-byte constant.  */
-    DW_OP_eq = 0x29,
-    DW_OP_ge = 0x2a,
-    DW_OP_gt = 0x2b,
-    DW_OP_le = 0x2c,
-    DW_OP_lt = 0x2d,
-    DW_OP_ne = 0x2e,
-    DW_OP_skip = 0x2f,		/* Signed 2-byte constant.  */
-    DW_OP_lit0 = 0x30,		/* Literal 0.  */
-    DW_OP_lit1 = 0x31,		/* Literal 1.  */
-    DW_OP_lit2 = 0x32,		/* Literal 2.  */
-    DW_OP_lit3 = 0x33,		/* Literal 3.  */
-    DW_OP_lit4 = 0x34,		/* Literal 4.  */
-    DW_OP_lit5 = 0x35,		/* Literal 5.  */
-    DW_OP_lit6 = 0x36,		/* Literal 6.  */
-    DW_OP_lit7 = 0x37,		/* Literal 7.  */
-    DW_OP_lit8 = 0x38,		/* Literal 8.  */
-    DW_OP_lit9 = 0x39,		/* Literal 9.  */
-    DW_OP_lit10 = 0x3a,		/* Literal 10.  */
-    DW_OP_lit11 = 0x3b,		/* Literal 11.  */
-    DW_OP_lit12 = 0x3c,		/* Literal 12.  */
-    DW_OP_lit13 = 0x3d,		/* Literal 13.  */
-    DW_OP_lit14 = 0x3e,		/* Literal 14.  */
-    DW_OP_lit15 = 0x3f,		/* Literal 15.  */
-    DW_OP_lit16 = 0x40,		/* Literal 16.  */
-    DW_OP_lit17 = 0x41,		/* Literal 17.  */
-    DW_OP_lit18 = 0x42,		/* Literal 18.  */
-    DW_OP_lit19 = 0x43,		/* Literal 19.  */
-    DW_OP_lit20 = 0x44,		/* Literal 20.  */
-    DW_OP_lit21 = 0x45,		/* Literal 21.  */
-    DW_OP_lit22 = 0x46,		/* Literal 22.  */
-    DW_OP_lit23 = 0x47,		/* Literal 23.  */
-    DW_OP_lit24 = 0x48,		/* Literal 24.  */
-    DW_OP_lit25 = 0x49,		/* Literal 25.  */
-    DW_OP_lit26 = 0x4a,		/* Literal 26.  */
-    DW_OP_lit27 = 0x4b,		/* Literal 27.  */
-    DW_OP_lit28 = 0x4c,		/* Literal 28.  */
-    DW_OP_lit29 = 0x4d,		/* Literal 29.  */
-    DW_OP_lit30 = 0x4e,		/* Literal 30.  */
-    DW_OP_lit31 = 0x4f,		/* Literal 31.  */
-    DW_OP_reg0 = 0x50,		/* Register 0.  */
-    DW_OP_reg1 = 0x51,		/* Register 1.  */
-    DW_OP_reg2 = 0x52,		/* Register 2.  */
-    DW_OP_reg3 = 0x53,		/* Register 3.  */
-    DW_OP_reg4 = 0x54,		/* Register 4.  */
-    DW_OP_reg5 = 0x55,		/* Register 5.  */
-    DW_OP_reg6 = 0x56,		/* Register 6.  */
-    DW_OP_reg7 = 0x57,		/* Register 7.  */
-    DW_OP_reg8 = 0x58,		/* Register 8.  */
-    DW_OP_reg9 = 0x59,		/* Register 9.  */
-    DW_OP_reg10 = 0x5a,		/* Register 10.  */
-    DW_OP_reg11 = 0x5b,		/* Register 11.  */
-    DW_OP_reg12 = 0x5c,		/* Register 12.  */
-    DW_OP_reg13 = 0x5d,		/* Register 13.  */
-    DW_OP_reg14 = 0x5e,		/* Register 14.  */
-    DW_OP_reg15 = 0x5f,		/* Register 15.  */
-    DW_OP_reg16 = 0x60,		/* Register 16.  */
-    DW_OP_reg17 = 0x61,		/* Register 17.  */
-    DW_OP_reg18 = 0x62,		/* Register 18.  */
-    DW_OP_reg19 = 0x63,		/* Register 19.  */
-    DW_OP_reg20 = 0x64,		/* Register 20.  */
-    DW_OP_reg21 = 0x65,		/* Register 21.  */
-    DW_OP_reg22 = 0x66,		/* Register 22.  */
-    DW_OP_reg23 = 0x67,		/* Register 24.  */
-    DW_OP_reg24 = 0x68,		/* Register 24.  */
-    DW_OP_reg25 = 0x69,		/* Register 25.  */
-    DW_OP_reg26 = 0x6a,		/* Register 26.  */
-    DW_OP_reg27 = 0x6b,		/* Register 27.  */
-    DW_OP_reg28 = 0x6c,		/* Register 28.  */
-    DW_OP_reg29 = 0x6d,		/* Register 29.  */
-    DW_OP_reg30 = 0x6e,		/* Register 30.  */
-    DW_OP_reg31 = 0x6f,		/* Register 31.  */
-    DW_OP_breg0 = 0x70,		/* Base register 0.  */
-    DW_OP_breg1 = 0x71,		/* Base register 1.  */
-    DW_OP_breg2 = 0x72,		/* Base register 2.  */
-    DW_OP_breg3 = 0x73,		/* Base register 3.  */
-    DW_OP_breg4 = 0x74,		/* Base register 4.  */
-    DW_OP_breg5 = 0x75,		/* Base register 5.  */
-    DW_OP_breg6 = 0x76,		/* Base register 6.  */
-    DW_OP_breg7 = 0x77,		/* Base register 7.  */
-    DW_OP_breg8 = 0x78,		/* Base register 8.  */
-    DW_OP_breg9 = 0x79,		/* Base register 9.  */
-    DW_OP_breg10 = 0x7a,	/* Base register 10.  */
-    DW_OP_breg11 = 0x7b,	/* Base register 11.  */
-    DW_OP_breg12 = 0x7c,	/* Base register 12.  */
-    DW_OP_breg13 = 0x7d,	/* Base register 13.  */
-    DW_OP_breg14 = 0x7e,	/* Base register 14.  */
-    DW_OP_breg15 = 0x7f,	/* Base register 15.  */
-    DW_OP_breg16 = 0x80,	/* Base register 16.  */
-    DW_OP_breg17 = 0x81,	/* Base register 17.  */
-    DW_OP_breg18 = 0x82,	/* Base register 18.  */
-    DW_OP_breg19 = 0x83,	/* Base register 19.  */
-    DW_OP_breg20 = 0x84,	/* Base register 20.  */
-    DW_OP_breg21 = 0x85,	/* Base register 21.  */
-    DW_OP_breg22 = 0x86,	/* Base register 22.  */
-    DW_OP_breg23 = 0x87,	/* Base register 23.  */
-    DW_OP_breg24 = 0x88,	/* Base register 24.  */
-    DW_OP_breg25 = 0x89,	/* Base register 25.  */
-    DW_OP_breg26 = 0x8a,	/* Base register 26.  */
-    DW_OP_breg27 = 0x8b,	/* Base register 27.  */
-    DW_OP_breg28 = 0x8c,	/* Base register 28.  */
-    DW_OP_breg29 = 0x8d,	/* Base register 29.  */
-    DW_OP_breg30 = 0x8e,	/* Base register 30.  */
-    DW_OP_breg31 = 0x8f,	/* Base register 31.  */
-    DW_OP_regx = 0x90,		/* Unsigned LEB128 register.  */
-    DW_OP_fbreg = 0x91,		/* Signed LEB128 offset.  */
-    DW_OP_bregx = 0x92,		/* ULEB128 register followed by SLEB128 off. */
-    DW_OP_piece = 0x93,		/* ULEB128 size of piece addressed. */
-    DW_OP_deref_size = 0x94,	/* 1-byte size of data retrieved.  */
-    DW_OP_xderef_size = 0x95,	/* 1-byte size of data retrieved.  */
-    DW_OP_nop = 0x96,
-    DW_OP_push_object_address = 0x97,
-    DW_OP_call2 = 0x98,
-    DW_OP_call4 = 0x99,
-    DW_OP_call_ref = 0x9a,
-    DW_OP_form_tls_address = 0x9b,/* TLS offset to address in current thread */
-    DW_OP_call_frame_cfa = 0x9c,/* CFA as determined by CFI.  */
-    DW_OP_bit_piece = 0x9d,	/* ULEB128 size and ULEB128 offset in bits.  */
-    DW_OP_implicit_value = 0x9e, /* DW_FORM_block follows opcode.  */
-    DW_OP_stack_value = 0x9f,	 /* No operands, special like DW_OP_piece.  */
-
-    DW_OP_implicit_pointer = 0xa0,
-    DW_OP_addrx = 0xa1,
-    DW_OP_constx = 0xa2,
-    DW_OP_entry_value = 0xa3,
-    DW_OP_const_type = 0xa4,
-    DW_OP_regval_type = 0xa5,
-    DW_OP_deref_type = 0xa6,
-    DW_OP_xderef_type = 0xa7,
-    DW_OP_convert = 0xa8,
-    DW_OP_reinterpret = 0xa9,
-
-    /* GNU extensions.  */
-    DW_OP_GNU_push_tls_address = 0xe0,
-    DW_OP_GNU_uninit = 0xf0,
-    DW_OP_GNU_encoded_addr = 0xf1,
-    DW_OP_GNU_implicit_pointer = 0xf2,
-    DW_OP_GNU_entry_value = 0xf3,
-    DW_OP_GNU_const_type = 0xf4,
-    DW_OP_GNU_regval_type = 0xf5,
-    DW_OP_GNU_deref_type = 0xf6,
-    DW_OP_GNU_convert = 0xf7,
-    DW_OP_GNU_reinterpret = 0xf9,
-    DW_OP_GNU_parameter_ref = 0xfa,
-
-    /* GNU Debug Fission extensions.  */
-    DW_OP_GNU_addr_index = 0xfb,
-    DW_OP_GNU_const_index = 0xfc,
-
-    DW_OP_GNU_variable_value = 0xfd,
-
-    DW_OP_lo_user = 0xe0,	/* Implementation-defined range start.  */
-    DW_OP_hi_user = 0xff	/* Implementation-defined range end.  */
-  };
-
-
-/* DWARF base type encodings.  */
-enum
-  {
-    DW_ATE_void = 0x0,
-    DW_ATE_address = 0x1,
-    DW_ATE_boolean = 0x2,
-    DW_ATE_complex_float = 0x3,
-    DW_ATE_float = 0x4,
-    DW_ATE_signed = 0x5,
-    DW_ATE_signed_char = 0x6,
-    DW_ATE_unsigned = 0x7,
-    DW_ATE_unsigned_char = 0x8,
-    DW_ATE_imaginary_float = 0x9,
-    DW_ATE_packed_decimal = 0xa,
-    DW_ATE_numeric_string = 0xb,
-    DW_ATE_edited = 0xc,
-    DW_ATE_signed_fixed = 0xd,
-    DW_ATE_unsigned_fixed = 0xe,
-    DW_ATE_decimal_float = 0xf,
-    DW_ATE_UTF = 0x10,
-    DW_ATE_UCS = 0x11,
-    DW_ATE_ASCII = 0x12,
-
-    DW_ATE_lo_user = 0x80,
-    DW_ATE_hi_user = 0xff
-  };
-
-
-/* DWARF decimal sign encodings.  */
-enum
-  {
-    DW_DS_unsigned = 1,
-    DW_DS_leading_overpunch = 2,
-    DW_DS_trailing_overpunch = 3,
-    DW_DS_leading_separate = 4,
-    DW_DS_trailing_separate = 5,
-  };
-
-
-/* DWARF endianity encodings.  */
-enum
-  {
-    DW_END_default = 0,
-    DW_END_big = 1,
-    DW_END_little = 2,
-
-    DW_END_lo_user = 0x40,
-    DW_END_hi_user = 0xff
-  };
-
-
-/* DWARF accessibility encodings.  */
-enum
-  {
-    DW_ACCESS_public = 1,
-    DW_ACCESS_protected = 2,
-    DW_ACCESS_private = 3
-  };
-
-
-/* DWARF visibility encodings.  */
-enum
-  {
-    DW_VIS_local = 1,
-    DW_VIS_exported = 2,
-    DW_VIS_qualified = 3
-  };
-
-
-/* DWARF virtuality encodings.  */
-enum
-  {
-    DW_VIRTUALITY_none = 0,
-    DW_VIRTUALITY_virtual = 1,
-    DW_VIRTUALITY_pure_virtual = 2
-  };
-
-
-/* DWARF language encodings.  */
-enum
-  {
-    DW_LANG_C89 = 0x0001,	     /* ISO C:1989 */
-    DW_LANG_C = 0x0002,		     /* C */
-    DW_LANG_Ada83 = 0x0003,	     /* ISO Ada:1983 */
-    DW_LANG_C_plus_plus	= 0x0004,    /* ISO C++:1998 */
-    DW_LANG_Cobol74 = 0x0005,	     /* ISO Cobol:1974 */
-    DW_LANG_Cobol85 = 0x0006,	     /* ISO Cobol:1985 */
-    DW_LANG_Fortran77 = 0x0007,	     /* ISO FORTRAN 77 */
-    DW_LANG_Fortran90 = 0x0008,	     /* ISO Fortran 90 */
-    DW_LANG_Pascal83 = 0x0009,	     /* ISO Pascal:1983 */
-    DW_LANG_Modula2 = 0x000a,	     /* ISO Modula-2:1996 */
-    DW_LANG_Java = 0x000b,	     /* Java */
-    DW_LANG_C99 = 0x000c,	     /* ISO C:1999 */
-    DW_LANG_Ada95 = 0x000d,	     /* ISO Ada:1995 */
-    DW_LANG_Fortran95 = 0x000e,	     /* ISO Fortran 95 */
-    DW_LANG_PLI = 0x000f,	     /* ISO PL/1:1976 */
-    DW_LANG_ObjC = 0x0010,	     /* Objective-C */
-    DW_LANG_ObjC_plus_plus = 0x0011, /* Objective-C++ */
-    DW_LANG_UPC = 0x0012,	     /* Unified Parallel C */
-    DW_LANG_D = 0x0013,		     /* D */
-    DW_LANG_Python = 0x0014,	     /* Python */
-    DW_LANG_OpenCL = 0x0015,	     /* OpenCL */
-    DW_LANG_Go = 0x0016,	     /* Go */
-    DW_LANG_Modula3 = 0x0017,	     /* Modula-3 */
-    DW_LANG_Haskell = 0x0018,	     /* Haskell */
-    DW_LANG_C_plus_plus_03 = 0x0019, /* ISO C++:2003 */
-    DW_LANG_C_plus_plus_11 = 0x001a, /* ISO C++:2011 */
-    DW_LANG_OCaml = 0x001b,	     /* OCaml */
-    DW_LANG_Rust = 0x001c,	     /* Rust */
-    DW_LANG_C11 = 0x001d,	     /* ISO C:2011 */
-    DW_LANG_Swift = 0x001e,	     /* Swift */
-    DW_LANG_Julia = 0x001f,	     /* Julia */
-    DW_LANG_Dylan = 0x0020,	     /* Dylan */
-    DW_LANG_C_plus_plus_14 = 0x0021, /* ISO C++:2014 */
-    DW_LANG_Fortran03 = 0x0022,	     /* ISO/IEC 1539-1:2004 */
-    DW_LANG_Fortran08 = 0x0023,	     /* ISO/IEC 1539-1:2010 */
-    DW_LANG_RenderScript = 0x0024,   /* RenderScript Kernal Language */
-    DW_LANG_BLISS = 0x0025,	     /* BLISS */
-
-    DW_LANG_lo_user = 0x8000,
-    DW_LANG_Mips_Assembler = 0x8001, /* Assembler */
-    DW_LANG_hi_user = 0xffff
-  };
-
-/* Old (typo) '1' != 'I'.  */
-#define DW_LANG_PL1 DW_LANG_PLI
-
-/* DWARF identifier case encodings.  */
-enum
-  {
-    DW_ID_case_sensitive = 0,
-    DW_ID_up_case = 1,
-    DW_ID_down_case = 2,
-    DW_ID_case_insensitive = 3
-  };
-
-
-/* DWARF calling conventions encodings.
-   Used as values of DW_AT_calling_convention for subroutines
-   (normal, program or nocall) or structures, unions and class types
-   (normal, reference or value).  */
-enum
-  {
-    DW_CC_normal = 0x1,
-    DW_CC_program = 0x2,
-    DW_CC_nocall = 0x3,
-    DW_CC_pass_by_reference = 0x4,
-    DW_CC_pass_by_value = 0x5,
-    DW_CC_lo_user = 0x40,
-    DW_CC_hi_user = 0xff
-  };
-
-
-/* DWARF inline encodings.  */
-enum
-  {
-    DW_INL_not_inlined = 0,
-    DW_INL_inlined = 1,
-    DW_INL_declared_not_inlined = 2,
-    DW_INL_declared_inlined = 3
-  };
-
-
-/* DWARF ordering encodings.  */
-enum
-  {
-    DW_ORD_row_major = 0,
-    DW_ORD_col_major = 1
-  };
-
-
-/* DWARF discriminant descriptor encodings.  */
-enum
-  {
-    DW_DSC_label = 0,
-    DW_DSC_range = 1
-  };
-
-/* DWARF defaulted member function encodings.  */
-enum
-  {
-    DW_DEFAULTED_no = 0,
-    DW_DEFAULTED_in_class = 1,
-    DW_DEFAULTED_out_of_class = 2
-  };
-
-/* DWARF line content descriptions.  */
-enum
-  {
-    DW_LNCT_path = 0x1,
-    DW_LNCT_directory_index = 0x2,
-    DW_LNCT_timestamp = 0x3,
-    DW_LNCT_size = 0x4,
-    DW_LNCT_MD5 = 0x5,
-    DW_LNCT_lo_user = 0x2000,
-    DW_LNCT_hi_user = 0x3fff
-  };
-
-/* DWARF standard opcode encodings.  */
-enum
-  {
-    DW_LNS_copy = 1,
-    DW_LNS_advance_pc = 2,
-    DW_LNS_advance_line = 3,
-    DW_LNS_set_file = 4,
-    DW_LNS_set_column = 5,
-    DW_LNS_negate_stmt = 6,
-    DW_LNS_set_basic_block = 7,
-    DW_LNS_const_add_pc = 8,
-    DW_LNS_fixed_advance_pc = 9,
-    DW_LNS_set_prologue_end = 10,
-    DW_LNS_set_epilogue_begin = 11,
-    DW_LNS_set_isa = 12
-  };
-
-
-/* DWARF extended opcode encodings.  */
-enum
-  {
-    DW_LNE_end_sequence = 1,
-    DW_LNE_set_address = 2,
-    DW_LNE_define_file = 3,
-    DW_LNE_set_discriminator = 4,
-
-    DW_LNE_lo_user = 128,
-    DW_LNE_hi_user = 255
-  };
-
-
-/* DWARF macinfo type encodings.  */
-enum
-  {
-    DW_MACINFO_define = 1,
-    DW_MACINFO_undef = 2,
-    DW_MACINFO_start_file = 3,
-    DW_MACINFO_end_file = 4,
-    DW_MACINFO_vendor_ext = 255
-  };
-
-
-/* DWARF debug_macro type encodings.  */
-enum
-  {
-    DW_MACRO_define = 0x01,
-    DW_MACRO_undef = 0x02,
-    DW_MACRO_start_file = 0x03,
-    DW_MACRO_end_file = 0x04,
-    DW_MACRO_define_strp = 0x05,
-    DW_MACRO_undef_strp = 0x06,
-    DW_MACRO_import = 0x07,
-    DW_MACRO_define_sup = 0x08,
-    DW_MACRO_undef_sup = 0x09,
-    DW_MACRO_import_sup = 0x0a,
-    DW_MACRO_define_strx = 0x0b,
-    DW_MACRO_undef_strx = 0x0c,
-    DW_MACRO_lo_user = 0xe0,
-    DW_MACRO_hi_user = 0xff
-  };
-
-/* Old GNU extension names for DWARF5 debug_macro type encodings.
-   There are no equivalents for the supplementary object file (sup)
-   and indirect string references (strx).  */
-#define DW_MACRO_GNU_define		 DW_MACRO_define
-#define DW_MACRO_GNU_undef		 DW_MACRO_undef
-#define DW_MACRO_GNU_start_file		 DW_MACRO_start_file
-#define DW_MACRO_GNU_end_file		 DW_MACRO_end_file
-#define DW_MACRO_GNU_define_indirect	 DW_MACRO_define_strp
-#define DW_MACRO_GNU_undef_indirect	 DW_MACRO_undef_strp
-#define DW_MACRO_GNU_transparent_include DW_MACRO_import
-#define DW_MACRO_GNU_lo_user		 DW_MACRO_lo_user
-#define DW_MACRO_GNU_hi_user		 DW_MACRO_hi_user
-
-
-/* Range list entry encoding.  */
-enum
-  {
-    DW_RLE_end_of_list = 0x0,
-    DW_RLE_base_addressx = 0x1,
-    DW_RLE_startx_endx = 0x2,
-    DW_RLE_startx_length = 0x3,
-    DW_RLE_offset_pair = 0x4,
-    DW_RLE_base_address = 0x5,
-    DW_RLE_start_end = 0x6,
-    DW_RLE_start_length = 0x7
-  };
-
-
-/* Location list entry encoding.  */
-enum
-  {
-    DW_LLE_end_of_list = 0x0,
-    DW_LLE_base_addressx = 0x1,
-    DW_LLE_startx_endx = 0x2,
-    DW_LLE_startx_length = 0x3,
-    DW_LLE_offset_pair = 0x4,
-    DW_LLE_default_location = 0x5,
-    DW_LLE_base_address = 0x6,
-    DW_LLE_start_end = 0x7,
-    DW_LLE_start_length = 0x8
-  };
-
-
-/* GNU DebugFission list entry encodings (.debug_loc.dwo).  */
-enum
-  {
-    DW_LLE_GNU_end_of_list_entry = 0x0,
-    DW_LLE_GNU_base_address_selection_entry = 0x1,
-    DW_LLE_GNU_start_end_entry = 0x2,
-    DW_LLE_GNU_start_length_entry = 0x3
-  };
-
-
-/* DWARF call frame instruction encodings.  */
-enum
-  {
-    DW_CFA_advance_loc = 0x40,
-    DW_CFA_offset = 0x80,
-    DW_CFA_restore = 0xc0,
-    DW_CFA_extended = 0,
-
-    DW_CFA_nop = 0x00,
-    DW_CFA_set_loc = 0x01,
-    DW_CFA_advance_loc1 = 0x02,
-    DW_CFA_advance_loc2 = 0x03,
-    DW_CFA_advance_loc4 = 0x04,
-    DW_CFA_offset_extended = 0x05,
-    DW_CFA_restore_extended = 0x06,
-    DW_CFA_undefined = 0x07,
-    DW_CFA_same_value = 0x08,
-    DW_CFA_register = 0x09,
-    DW_CFA_remember_state = 0x0a,
-    DW_CFA_restore_state = 0x0b,
-    DW_CFA_def_cfa = 0x0c,
-    DW_CFA_def_cfa_register = 0x0d,
-    DW_CFA_def_cfa_offset = 0x0e,
-    DW_CFA_def_cfa_expression = 0x0f,
-    DW_CFA_expression = 0x10,
-    DW_CFA_offset_extended_sf = 0x11,
-    DW_CFA_def_cfa_sf = 0x12,
-    DW_CFA_def_cfa_offset_sf = 0x13,
-    DW_CFA_val_offset = 0x14,
-    DW_CFA_val_offset_sf = 0x15,
-    DW_CFA_val_expression = 0x16,
-
-    DW_CFA_low_user = 0x1c,
-    DW_CFA_MIPS_advance_loc8 = 0x1d,
-    DW_CFA_GNU_window_save = 0x2d,
-    DW_CFA_AARCH64_negate_ra_state = 0x2d,
-    DW_CFA_GNU_args_size = 0x2e,
-    DW_CFA_GNU_negative_offset_extended = 0x2f,
-    DW_CFA_high_user = 0x3f
-  };
-
-/* ID indicating CIE as opposed to FDE in .debug_frame.  */
-enum
-  {
-    DW_CIE_ID_32 = 0xffffffffU,		 /* In 32-bit format CIE header.  */
-    DW_CIE_ID_64 = 0xffffffffffffffffULL /* In 64-bit format CIE header.  */
-  };
-
-
-/* Information for GNU unwind information.  */
-enum
-  {
-    DW_EH_PE_absptr = 0x00,
-    DW_EH_PE_omit = 0xff,
-
-    /* FDE data encoding.  */
-    DW_EH_PE_uleb128 = 0x01,
-    DW_EH_PE_udata2 = 0x02,
-    DW_EH_PE_udata4 = 0x03,
-    DW_EH_PE_udata8 = 0x04,
-    DW_EH_PE_sleb128 = 0x09,
-    DW_EH_PE_sdata2 = 0x0a,
-    DW_EH_PE_sdata4 = 0x0b,
-    DW_EH_PE_sdata8 = 0x0c,
-    DW_EH_PE_signed = 0x08,
-
-    /* FDE flags.  */
-    DW_EH_PE_pcrel = 0x10,
-    DW_EH_PE_textrel = 0x20,
-    DW_EH_PE_datarel = 0x30,
-    DW_EH_PE_funcrel = 0x40,
-    DW_EH_PE_aligned = 0x50,
-
-    DW_EH_PE_indirect = 0x80
-  };
-
-
-/* DWARF XXX.  */
-#define DW_ADDR_none	0
-
-/* Section 7.2.2 of the DWARF3 specification defines a range of escape
-   codes that can appear in the length field of certain DWARF structures.
-
-   These defines enumerate the minimum and maximum values of this range.
-   Currently only the maximum value is used (to indicate that 64-bit
-   values are going to be used in the dwarf data that accompanies the
-   structure).  The other values are reserved.
-
-   Note: There is a typo in DWARF3 spec (published Dec 20, 2005).  In
-   sections 7.4, 7.5.1, 7.19, 7.20 the minimum escape code is referred to
-   as 0xffffff00 whereas in fact it should be 0xfffffff0.  */
-#define DWARF3_LENGTH_MIN_ESCAPE_CODE 0xfffffff0u
-#define DWARF3_LENGTH_MAX_ESCAPE_CODE 0xffffffffu
-#define DWARF3_LENGTH_64_BIT          DWARF3_LENGTH_MAX_ESCAPE_CODE
-
-#endif	/* dwarf.h */
diff --git a/projects/llm_framework/include/elfutils/elf-knowledge.h b/projects/llm_framework/include/elfutils/elf-knowledge.h
deleted file mode 100644
index 6e005fa5..00000000
--- a/projects/llm_framework/include/elfutils/elf-knowledge.h
+++ /dev/null
@@ -1,101 +0,0 @@
-/* Accumulation of various pieces of knowledge about ELF.
-   Copyright (C) 2000-2012, 2014, 2016 Red Hat, Inc.
-   This file is part of elfutils.
-   Written by Ulrich Drepper <drepper@redhat.com>, 2000.
-
-   This file is free software; you can redistribute it and/or modify
-   it under the terms of either
-
-     * the GNU Lesser General Public License as published by the Free
-       Software Foundation; either version 3 of the License, or (at
-       your option) any later version
-
-   or
-
-     * the GNU General Public License as published by the Free
-       Software Foundation; either version 2 of the License, or (at
-       your option) any later version
-
-   or both in parallel, as here.
-
-   elfutils is distributed in the hope that it will be useful, but
-   WITHOUT ANY WARRANTY; without even the implied warranty of
-   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
-   General Public License for more details.
-
-   You should have received copies of the GNU General Public License and
-   the GNU Lesser General Public License along with this program.  If
-   not, see <http://www.gnu.org/licenses/>.  */
-
-#ifndef _ELF_KNOWLEDGE_H
-#define _ELF_KNOWLEDGE_H	1
-
-#include <stdbool.h>
-
-
-/* Test whether a section can be stripped or not.  */
-#define SECTION_STRIP_P(shdr, name, remove_comment) \
-  /* Sections which are allocated are not removed.  */			      \
-  (((shdr)->sh_flags & SHF_ALLOC) == 0					      \
-   /* We never remove .note sections.  */				      \
-   && (shdr)->sh_type != SHT_NOTE					      \
-   && (((shdr)->sh_type) != SHT_PROGBITS				      \
-       /* Never remove .gnu.warning.* sections.  */			      \
-       || (name != NULL							      \
-	   && strncmp (name, ".gnu.warning.", sizeof ".gnu.warning." - 1) != 0\
-	   /* We remove .comment sections only if explicitly told to do so. */\
-	   && (remove_comment						      \
-	       || strcmp (name, ".comment") != 0))))
-
-
-/* Test whether `sh_info' field in section header contains a section
-   index.  There are two kinds of sections doing this:
-
-   - the sections containing relocation information reference in this
-     field the section to which the relocations apply;
-
-   - section with the SHF_INFO_LINK flag set to signal that `sh_info'
-     references a section.  This allows correct handling of unknown
-     sections.  */
-#define SH_INFO_LINK_P(Shdr) \
-  ((Shdr)->sh_type == SHT_REL || (Shdr)->sh_type == SHT_RELA		      \
-   || ((Shdr)->sh_flags & SHF_INFO_LINK) != 0)
-
-
-/* Size of an entry in the hash table.  The ELF specification says all
-   entries are regardless of platform 32-bits in size.  Early 64-bit
-   ports (namely Alpha for Linux) got this wrong.  The wording was not
-   clear.
-
-   Several years later the ABI for the 64-bit S390s was developed.
-   Many things were copied from the IA-64 ABI (which uses the correct
-   32-bit entry size) but it does get the SHT_HASH entry size wrong by
-   using a 64-bit entry size.  So now we need this macro to special
-   case both the alpha and s390x ABIs.  */
-#define SH_ENTSIZE_HASH(Ehdr) \
-  ((Ehdr)->e_machine == EM_ALPHA					      \
-   || ((Ehdr)->e_machine == EM_S390					      \
-       && (Ehdr)->e_ident[EI_CLASS] == ELFCLASS64) ? 8 : 4)
-
-/* GNU Annobin notes are not fully standardized and abuses the owner name.  */
-
-#define ELF_NOTE_GNU_BUILD_ATTRIBUTE_PREFIX "GA"
-
-#define NT_GNU_BUILD_ATTRIBUTE_OPEN 0x100
-#define NT_GNU_BUILD_ATTRIBUTE_FUNC 0x101
-
-#define GNU_BUILD_ATTRIBUTE_TYPE_NUMERIC	'*'
-#define GNU_BUILD_ATTRIBUTE_TYPE_STRING		'$'
-#define GNU_BUILD_ATTRIBUTE_TYPE_BOOL_TRUE	'+'
-#define GNU_BUILD_ATTRIBUTE_TYPE_BOOL_FALSE	'!'
-
-#define GNU_BUILD_ATTRIBUTE_VERSION	1
-#define GNU_BUILD_ATTRIBUTE_STACK_PROT	2
-#define GNU_BUILD_ATTRIBUTE_RELRO	3
-#define GNU_BUILD_ATTRIBUTE_STACK_SIZE	4
-#define GNU_BUILD_ATTRIBUTE_TOOL	5
-#define GNU_BUILD_ATTRIBUTE_ABI		6
-#define GNU_BUILD_ATTRIBUTE_PIC		7
-#define GNU_BUILD_ATTRIBUTE_SHORT_ENUM	8
-
-#endif	/* elf-knowledge.h */
diff --git a/projects/llm_framework/include/elfutils/known-dwarf.h b/projects/llm_framework/include/elfutils/known-dwarf.h
deleted file mode 100644
index 4b8a83c9..00000000
--- a/projects/llm_framework/include/elfutils/known-dwarf.h
+++ /dev/null
@@ -1,782 +0,0 @@
-/* Generated by config/known-dwarf.awk from libdw/dwarf.h contents.  */
-
-#define DWARF_ALL_KNOWN_DW_ACCESS \
-  DWARF_ONE_KNOWN_DW_ACCESS (private, DW_ACCESS_private) \
-  DWARF_ONE_KNOWN_DW_ACCESS (protected, DW_ACCESS_protected) \
-  DWARF_ONE_KNOWN_DW_ACCESS (public, DW_ACCESS_public) \
-  /* End of DW_ACCESS_*.  */
-
-#define DWARF_ALL_KNOWN_DW_AT \
-  DWARF_ONE_KNOWN_DW_AT (GNU_addr_base, DW_AT_GNU_addr_base) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_all_call_sites, DW_AT_GNU_all_call_sites) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_all_source_call_sites, DW_AT_GNU_all_source_call_sites) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_all_tail_call_sites, DW_AT_GNU_all_tail_call_sites) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_bias, DW_AT_GNU_bias) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_call_site_data_value, DW_AT_GNU_call_site_data_value) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_call_site_target, DW_AT_GNU_call_site_target) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_call_site_target_clobbered, DW_AT_GNU_call_site_target_clobbered) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_call_site_value, DW_AT_GNU_call_site_value) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_deleted, DW_AT_GNU_deleted) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_denominator, DW_AT_GNU_denominator) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_dwo_id, DW_AT_GNU_dwo_id) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_dwo_name, DW_AT_GNU_dwo_name) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_entry_view, DW_AT_GNU_entry_view) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_exclusive_locks_required, DW_AT_GNU_exclusive_locks_required) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_guarded, DW_AT_GNU_guarded) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_guarded_by, DW_AT_GNU_guarded_by) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_locks_excluded, DW_AT_GNU_locks_excluded) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_locviews, DW_AT_GNU_locviews) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_macros, DW_AT_GNU_macros) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_numerator, DW_AT_GNU_numerator) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_odr_signature, DW_AT_GNU_odr_signature) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_pt_guarded, DW_AT_GNU_pt_guarded) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_pt_guarded_by, DW_AT_GNU_pt_guarded_by) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_pubnames, DW_AT_GNU_pubnames) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_pubtypes, DW_AT_GNU_pubtypes) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_ranges_base, DW_AT_GNU_ranges_base) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_shared_locks_required, DW_AT_GNU_shared_locks_required) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_tail_call, DW_AT_GNU_tail_call) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_template_name, DW_AT_GNU_template_name) \
-  DWARF_ONE_KNOWN_DW_AT (GNU_vector, DW_AT_GNU_vector) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_abstract_name, DW_AT_MIPS_abstract_name) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_allocatable_dopetype, DW_AT_MIPS_allocatable_dopetype) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_assumed_shape_dopetype, DW_AT_MIPS_assumed_shape_dopetype) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_assumed_size, DW_AT_MIPS_assumed_size) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_clone_origin, DW_AT_MIPS_clone_origin) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_epilog_begin, DW_AT_MIPS_epilog_begin) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_fde, DW_AT_MIPS_fde) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_has_inlines, DW_AT_MIPS_has_inlines) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_linkage_name, DW_AT_MIPS_linkage_name) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_loop_begin, DW_AT_MIPS_loop_begin) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_loop_unroll_factor, DW_AT_MIPS_loop_unroll_factor) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_ptr_dopetype, DW_AT_MIPS_ptr_dopetype) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_software_pipeline_depth, DW_AT_MIPS_software_pipeline_depth) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_stride, DW_AT_MIPS_stride) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_stride_byte, DW_AT_MIPS_stride_byte) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_stride_elem, DW_AT_MIPS_stride_elem) \
-  DWARF_ONE_KNOWN_DW_AT (MIPS_tail_loop_begin, DW_AT_MIPS_tail_loop_begin) \
-  DWARF_ONE_KNOWN_DW_AT (abstract_origin, DW_AT_abstract_origin) \
-  DWARF_ONE_KNOWN_DW_AT (accessibility, DW_AT_accessibility) \
-  DWARF_ONE_KNOWN_DW_AT (addr_base, DW_AT_addr_base) \
-  DWARF_ONE_KNOWN_DW_AT (address_class, DW_AT_address_class) \
-  DWARF_ONE_KNOWN_DW_AT (alignment, DW_AT_alignment) \
-  DWARF_ONE_KNOWN_DW_AT (allocated, DW_AT_allocated) \
-  DWARF_ONE_KNOWN_DW_AT (artificial, DW_AT_artificial) \
-  DWARF_ONE_KNOWN_DW_AT (associated, DW_AT_associated) \
-  DWARF_ONE_KNOWN_DW_AT (base_types, DW_AT_base_types) \
-  DWARF_ONE_KNOWN_DW_AT (binary_scale, DW_AT_binary_scale) \
-  DWARF_ONE_KNOWN_DW_AT (bit_offset, DW_AT_bit_offset) \
-  DWARF_ONE_KNOWN_DW_AT (bit_size, DW_AT_bit_size) \
-  DWARF_ONE_KNOWN_DW_AT (bit_stride, DW_AT_bit_stride) \
-  DWARF_ONE_KNOWN_DW_AT (body_begin, DW_AT_body_begin) \
-  DWARF_ONE_KNOWN_DW_AT (body_end, DW_AT_body_end) \
-  DWARF_ONE_KNOWN_DW_AT (byte_size, DW_AT_byte_size) \
-  DWARF_ONE_KNOWN_DW_AT (byte_stride, DW_AT_byte_stride) \
-  DWARF_ONE_KNOWN_DW_AT (call_all_calls, DW_AT_call_all_calls) \
-  DWARF_ONE_KNOWN_DW_AT (call_all_source_calls, DW_AT_call_all_source_calls) \
-  DWARF_ONE_KNOWN_DW_AT (call_all_tail_calls, DW_AT_call_all_tail_calls) \
-  DWARF_ONE_KNOWN_DW_AT (call_column, DW_AT_call_column) \
-  DWARF_ONE_KNOWN_DW_AT (call_data_location, DW_AT_call_data_location) \
-  DWARF_ONE_KNOWN_DW_AT (call_data_value, DW_AT_call_data_value) \
-  DWARF_ONE_KNOWN_DW_AT (call_file, DW_AT_call_file) \
-  DWARF_ONE_KNOWN_DW_AT (call_line, DW_AT_call_line) \
-  DWARF_ONE_KNOWN_DW_AT (call_origin, DW_AT_call_origin) \
-  DWARF_ONE_KNOWN_DW_AT (call_parameter, DW_AT_call_parameter) \
-  DWARF_ONE_KNOWN_DW_AT (call_pc, DW_AT_call_pc) \
-  DWARF_ONE_KNOWN_DW_AT (call_return_pc, DW_AT_call_return_pc) \
-  DWARF_ONE_KNOWN_DW_AT (call_tail_call, DW_AT_call_tail_call) \
-  DWARF_ONE_KNOWN_DW_AT (call_target, DW_AT_call_target) \
-  DWARF_ONE_KNOWN_DW_AT (call_target_clobbered, DW_AT_call_target_clobbered) \
-  DWARF_ONE_KNOWN_DW_AT (call_value, DW_AT_call_value) \
-  DWARF_ONE_KNOWN_DW_AT (calling_convention, DW_AT_calling_convention) \
-  DWARF_ONE_KNOWN_DW_AT (common_reference, DW_AT_common_reference) \
-  DWARF_ONE_KNOWN_DW_AT (comp_dir, DW_AT_comp_dir) \
-  DWARF_ONE_KNOWN_DW_AT (const_expr, DW_AT_const_expr) \
-  DWARF_ONE_KNOWN_DW_AT (const_value, DW_AT_const_value) \
-  DWARF_ONE_KNOWN_DW_AT (containing_type, DW_AT_containing_type) \
-  DWARF_ONE_KNOWN_DW_AT (count, DW_AT_count) \
-  DWARF_ONE_KNOWN_DW_AT (data_bit_offset, DW_AT_data_bit_offset) \
-  DWARF_ONE_KNOWN_DW_AT (data_location, DW_AT_data_location) \
-  DWARF_ONE_KNOWN_DW_AT (data_member_location, DW_AT_data_member_location) \
-  DWARF_ONE_KNOWN_DW_AT (decimal_scale, DW_AT_decimal_scale) \
-  DWARF_ONE_KNOWN_DW_AT (decimal_sign, DW_AT_decimal_sign) \
-  DWARF_ONE_KNOWN_DW_AT (decl_column, DW_AT_decl_column) \
-  DWARF_ONE_KNOWN_DW_AT (decl_file, DW_AT_decl_file) \
-  DWARF_ONE_KNOWN_DW_AT (decl_line, DW_AT_decl_line) \
-  DWARF_ONE_KNOWN_DW_AT (declaration, DW_AT_declaration) \
-  DWARF_ONE_KNOWN_DW_AT (default_value, DW_AT_default_value) \
-  DWARF_ONE_KNOWN_DW_AT (defaulted, DW_AT_defaulted) \
-  DWARF_ONE_KNOWN_DW_AT (deleted, DW_AT_deleted) \
-  DWARF_ONE_KNOWN_DW_AT (description, DW_AT_description) \
-  DWARF_ONE_KNOWN_DW_AT (digit_count, DW_AT_digit_count) \
-  DWARF_ONE_KNOWN_DW_AT (discr, DW_AT_discr) \
-  DWARF_ONE_KNOWN_DW_AT (discr_list, DW_AT_discr_list) \
-  DWARF_ONE_KNOWN_DW_AT (discr_value, DW_AT_discr_value) \
-  DWARF_ONE_KNOWN_DW_AT (dwo_name, DW_AT_dwo_name) \
-  DWARF_ONE_KNOWN_DW_AT (elemental, DW_AT_elemental) \
-  DWARF_ONE_KNOWN_DW_AT (encoding, DW_AT_encoding) \
-  DWARF_ONE_KNOWN_DW_AT (endianity, DW_AT_endianity) \
-  DWARF_ONE_KNOWN_DW_AT (entry_pc, DW_AT_entry_pc) \
-  DWARF_ONE_KNOWN_DW_AT (enum_class, DW_AT_enum_class) \
-  DWARF_ONE_KNOWN_DW_AT (explicit, DW_AT_explicit) \
-  DWARF_ONE_KNOWN_DW_AT (export_symbols, DW_AT_export_symbols) \
-  DWARF_ONE_KNOWN_DW_AT (extension, DW_AT_extension) \
-  DWARF_ONE_KNOWN_DW_AT (external, DW_AT_external) \
-  DWARF_ONE_KNOWN_DW_AT (frame_base, DW_AT_frame_base) \
-  DWARF_ONE_KNOWN_DW_AT (friend, DW_AT_friend) \
-  DWARF_ONE_KNOWN_DW_AT (high_pc, DW_AT_high_pc) \
-  DWARF_ONE_KNOWN_DW_AT (identifier_case, DW_AT_identifier_case) \
-  DWARF_ONE_KNOWN_DW_AT (import, DW_AT_import) \
-  DWARF_ONE_KNOWN_DW_AT (inline, DW_AT_inline) \
-  DWARF_ONE_KNOWN_DW_AT (is_optional, DW_AT_is_optional) \
-  DWARF_ONE_KNOWN_DW_AT (language, DW_AT_language) \
-  DWARF_ONE_KNOWN_DW_AT (linkage_name, DW_AT_linkage_name) \
-  DWARF_ONE_KNOWN_DW_AT (location, DW_AT_location) \
-  DWARF_ONE_KNOWN_DW_AT (loclists_base, DW_AT_loclists_base) \
-  DWARF_ONE_KNOWN_DW_AT (low_pc, DW_AT_low_pc) \
-  DWARF_ONE_KNOWN_DW_AT (lower_bound, DW_AT_lower_bound) \
-  DWARF_ONE_KNOWN_DW_AT (mac_info, DW_AT_mac_info) \
-  DWARF_ONE_KNOWN_DW_AT (macro_info, DW_AT_macro_info) \
-  DWARF_ONE_KNOWN_DW_AT (macros, DW_AT_macros) \
-  DWARF_ONE_KNOWN_DW_AT (main_subprogram, DW_AT_main_subprogram) \
-  DWARF_ONE_KNOWN_DW_AT (mutable, DW_AT_mutable) \
-  DWARF_ONE_KNOWN_DW_AT (name, DW_AT_name) \
-  DWARF_ONE_KNOWN_DW_AT (namelist_item, DW_AT_namelist_item) \
-  DWARF_ONE_KNOWN_DW_AT (noreturn, DW_AT_noreturn) \
-  DWARF_ONE_KNOWN_DW_AT (object_pointer, DW_AT_object_pointer) \
-  DWARF_ONE_KNOWN_DW_AT (ordering, DW_AT_ordering) \
-  DWARF_ONE_KNOWN_DW_AT (picture_string, DW_AT_picture_string) \
-  DWARF_ONE_KNOWN_DW_AT (priority, DW_AT_priority) \
-  DWARF_ONE_KNOWN_DW_AT (producer, DW_AT_producer) \
-  DWARF_ONE_KNOWN_DW_AT (prototyped, DW_AT_prototyped) \
-  DWARF_ONE_KNOWN_DW_AT (pure, DW_AT_pure) \
-  DWARF_ONE_KNOWN_DW_AT (ranges, DW_AT_ranges) \
-  DWARF_ONE_KNOWN_DW_AT (rank, DW_AT_rank) \
-  DWARF_ONE_KNOWN_DW_AT (recursive, DW_AT_recursive) \
-  DWARF_ONE_KNOWN_DW_AT (reference, DW_AT_reference) \
-  DWARF_ONE_KNOWN_DW_AT (return_addr, DW_AT_return_addr) \
-  DWARF_ONE_KNOWN_DW_AT (rnglists_base, DW_AT_rnglists_base) \
-  DWARF_ONE_KNOWN_DW_AT (rvalue_reference, DW_AT_rvalue_reference) \
-  DWARF_ONE_KNOWN_DW_AT (segment, DW_AT_segment) \
-  DWARF_ONE_KNOWN_DW_AT (sf_names, DW_AT_sf_names) \
-  DWARF_ONE_KNOWN_DW_AT (sibling, DW_AT_sibling) \
-  DWARF_ONE_KNOWN_DW_AT (signature, DW_AT_signature) \
-  DWARF_ONE_KNOWN_DW_AT (small, DW_AT_small) \
-  DWARF_ONE_KNOWN_DW_AT (specification, DW_AT_specification) \
-  DWARF_ONE_KNOWN_DW_AT (src_coords, DW_AT_src_coords) \
-  DWARF_ONE_KNOWN_DW_AT (src_info, DW_AT_src_info) \
-  DWARF_ONE_KNOWN_DW_AT (start_scope, DW_AT_start_scope) \
-  DWARF_ONE_KNOWN_DW_AT (static_link, DW_AT_static_link) \
-  DWARF_ONE_KNOWN_DW_AT (stmt_list, DW_AT_stmt_list) \
-  DWARF_ONE_KNOWN_DW_AT (str_offsets_base, DW_AT_str_offsets_base) \
-  DWARF_ONE_KNOWN_DW_AT (string_length, DW_AT_string_length) \
-  DWARF_ONE_KNOWN_DW_AT (string_length_bit_size, DW_AT_string_length_bit_size) \
-  DWARF_ONE_KNOWN_DW_AT (string_length_byte_size, DW_AT_string_length_byte_size) \
-  DWARF_ONE_KNOWN_DW_AT (threads_scaled, DW_AT_threads_scaled) \
-  DWARF_ONE_KNOWN_DW_AT (trampoline, DW_AT_trampoline) \
-  DWARF_ONE_KNOWN_DW_AT (type, DW_AT_type) \
-  DWARF_ONE_KNOWN_DW_AT (upper_bound, DW_AT_upper_bound) \
-  DWARF_ONE_KNOWN_DW_AT (use_UTF8, DW_AT_use_UTF8) \
-  DWARF_ONE_KNOWN_DW_AT (use_location, DW_AT_use_location) \
-  DWARF_ONE_KNOWN_DW_AT (variable_parameter, DW_AT_variable_parameter) \
-  DWARF_ONE_KNOWN_DW_AT (virtuality, DW_AT_virtuality) \
-  DWARF_ONE_KNOWN_DW_AT (visibility, DW_AT_visibility) \
-  DWARF_ONE_KNOWN_DW_AT (vtable_elem_location, DW_AT_vtable_elem_location) \
-  /* End of DW_AT_*.  */
-
-#define DWARF_ALL_KNOWN_DW_ATE \
-  DWARF_ONE_KNOWN_DW_ATE (ASCII, DW_ATE_ASCII) \
-  DWARF_ONE_KNOWN_DW_ATE (UCS, DW_ATE_UCS) \
-  DWARF_ONE_KNOWN_DW_ATE (UTF, DW_ATE_UTF) \
-  DWARF_ONE_KNOWN_DW_ATE (address, DW_ATE_address) \
-  DWARF_ONE_KNOWN_DW_ATE (boolean, DW_ATE_boolean) \
-  DWARF_ONE_KNOWN_DW_ATE (complex_float, DW_ATE_complex_float) \
-  DWARF_ONE_KNOWN_DW_ATE (decimal_float, DW_ATE_decimal_float) \
-  DWARF_ONE_KNOWN_DW_ATE (edited, DW_ATE_edited) \
-  DWARF_ONE_KNOWN_DW_ATE (float, DW_ATE_float) \
-  DWARF_ONE_KNOWN_DW_ATE (imaginary_float, DW_ATE_imaginary_float) \
-  DWARF_ONE_KNOWN_DW_ATE (numeric_string, DW_ATE_numeric_string) \
-  DWARF_ONE_KNOWN_DW_ATE (packed_decimal, DW_ATE_packed_decimal) \
-  DWARF_ONE_KNOWN_DW_ATE (signed, DW_ATE_signed) \
-  DWARF_ONE_KNOWN_DW_ATE (signed_char, DW_ATE_signed_char) \
-  DWARF_ONE_KNOWN_DW_ATE (signed_fixed, DW_ATE_signed_fixed) \
-  DWARF_ONE_KNOWN_DW_ATE (unsigned, DW_ATE_unsigned) \
-  DWARF_ONE_KNOWN_DW_ATE (unsigned_char, DW_ATE_unsigned_char) \
-  DWARF_ONE_KNOWN_DW_ATE (unsigned_fixed, DW_ATE_unsigned_fixed) \
-  DWARF_ONE_KNOWN_DW_ATE (void, DW_ATE_void) \
-  /* End of DW_ATE_*.  */
-
-#define DWARF_ALL_KNOWN_DW_CC \
-  DWARF_ONE_KNOWN_DW_CC (nocall, DW_CC_nocall) \
-  DWARF_ONE_KNOWN_DW_CC (normal, DW_CC_normal) \
-  DWARF_ONE_KNOWN_DW_CC (pass_by_reference, DW_CC_pass_by_reference) \
-  DWARF_ONE_KNOWN_DW_CC (pass_by_value, DW_CC_pass_by_value) \
-  DWARF_ONE_KNOWN_DW_CC (program, DW_CC_program) \
-  /* End of DW_CC_*.  */
-
-#define DWARF_ALL_KNOWN_DW_CFA \
-  DWARF_ONE_KNOWN_DW_CFA (AARCH64_negate_ra_state, DW_CFA_AARCH64_negate_ra_state) \
-  DWARF_ONE_KNOWN_DW_CFA (GNU_args_size, DW_CFA_GNU_args_size) \
-  DWARF_ONE_KNOWN_DW_CFA (GNU_negative_offset_extended, DW_CFA_GNU_negative_offset_extended) \
-  DWARF_ONE_KNOWN_DW_CFA (GNU_window_save, DW_CFA_GNU_window_save) \
-  DWARF_ONE_KNOWN_DW_CFA (MIPS_advance_loc8, DW_CFA_MIPS_advance_loc8) \
-  DWARF_ONE_KNOWN_DW_CFA (advance_loc, DW_CFA_advance_loc) \
-  DWARF_ONE_KNOWN_DW_CFA (advance_loc1, DW_CFA_advance_loc1) \
-  DWARF_ONE_KNOWN_DW_CFA (advance_loc2, DW_CFA_advance_loc2) \
-  DWARF_ONE_KNOWN_DW_CFA (advance_loc4, DW_CFA_advance_loc4) \
-  DWARF_ONE_KNOWN_DW_CFA (def_cfa, DW_CFA_def_cfa) \
-  DWARF_ONE_KNOWN_DW_CFA (def_cfa_expression, DW_CFA_def_cfa_expression) \
-  DWARF_ONE_KNOWN_DW_CFA (def_cfa_offset, DW_CFA_def_cfa_offset) \
-  DWARF_ONE_KNOWN_DW_CFA (def_cfa_offset_sf, DW_CFA_def_cfa_offset_sf) \
-  DWARF_ONE_KNOWN_DW_CFA (def_cfa_register, DW_CFA_def_cfa_register) \
-  DWARF_ONE_KNOWN_DW_CFA (def_cfa_sf, DW_CFA_def_cfa_sf) \
-  DWARF_ONE_KNOWN_DW_CFA (expression, DW_CFA_expression) \
-  DWARF_ONE_KNOWN_DW_CFA (extended, DW_CFA_extended) \
-  DWARF_ONE_KNOWN_DW_CFA (nop, DW_CFA_nop) \
-  DWARF_ONE_KNOWN_DW_CFA (offset, DW_CFA_offset) \
-  DWARF_ONE_KNOWN_DW_CFA (offset_extended, DW_CFA_offset_extended) \
-  DWARF_ONE_KNOWN_DW_CFA (offset_extended_sf, DW_CFA_offset_extended_sf) \
-  DWARF_ONE_KNOWN_DW_CFA (register, DW_CFA_register) \
-  DWARF_ONE_KNOWN_DW_CFA (remember_state, DW_CFA_remember_state) \
-  DWARF_ONE_KNOWN_DW_CFA (restore, DW_CFA_restore) \
-  DWARF_ONE_KNOWN_DW_CFA (restore_extended, DW_CFA_restore_extended) \
-  DWARF_ONE_KNOWN_DW_CFA (restore_state, DW_CFA_restore_state) \
-  DWARF_ONE_KNOWN_DW_CFA (same_value, DW_CFA_same_value) \
-  DWARF_ONE_KNOWN_DW_CFA (set_loc, DW_CFA_set_loc) \
-  DWARF_ONE_KNOWN_DW_CFA (undefined, DW_CFA_undefined) \
-  DWARF_ONE_KNOWN_DW_CFA (val_expression, DW_CFA_val_expression) \
-  DWARF_ONE_KNOWN_DW_CFA (val_offset, DW_CFA_val_offset) \
-  DWARF_ONE_KNOWN_DW_CFA (val_offset_sf, DW_CFA_val_offset_sf) \
-  /* End of DW_CFA_*.  */
-
-#define DWARF_ALL_KNOWN_DW_CHILDREN \
-  DWARF_ONE_KNOWN_DW_CHILDREN (no, DW_CHILDREN_no) \
-  DWARF_ONE_KNOWN_DW_CHILDREN (yes, DW_CHILDREN_yes) \
-  /* End of DW_CHILDREN_*.  */
-
-#define DWARF_ALL_KNOWN_DW_CIE_ID \
-  DWARF_ONE_KNOWN_DW_CIE_ID (32, DW_CIE_ID_32) \
-  DWARF_ONE_KNOWN_DW_CIE_ID (64, DW_CIE_ID_64) \
-  /* End of DW_CIE_ID_*.  */
-
-#define DWARF_ALL_KNOWN_DW_DEFAULTED \
-  DWARF_ONE_KNOWN_DW_DEFAULTED (in_class, DW_DEFAULTED_in_class) \
-  DWARF_ONE_KNOWN_DW_DEFAULTED (no, DW_DEFAULTED_no) \
-  DWARF_ONE_KNOWN_DW_DEFAULTED (out_of_class, DW_DEFAULTED_out_of_class) \
-  /* End of DW_DEFAULTED_*.  */
-
-#define DWARF_ALL_KNOWN_DW_DS \
-  DWARF_ONE_KNOWN_DW_DS (leading_overpunch, DW_DS_leading_overpunch) \
-  DWARF_ONE_KNOWN_DW_DS (leading_separate, DW_DS_leading_separate) \
-  DWARF_ONE_KNOWN_DW_DS (trailing_overpunch, DW_DS_trailing_overpunch) \
-  DWARF_ONE_KNOWN_DW_DS (trailing_separate, DW_DS_trailing_separate) \
-  DWARF_ONE_KNOWN_DW_DS (unsigned, DW_DS_unsigned) \
-  /* End of DW_DS_*.  */
-
-#define DWARF_ALL_KNOWN_DW_DSC \
-  DWARF_ONE_KNOWN_DW_DSC (label, DW_DSC_label) \
-  DWARF_ONE_KNOWN_DW_DSC (range, DW_DSC_range) \
-  /* End of DW_DSC_*.  */
-
-#define DWARF_ALL_KNOWN_DW_EH_PE \
-  DWARF_ONE_KNOWN_DW_EH_PE (absptr, DW_EH_PE_absptr) \
-  DWARF_ONE_KNOWN_DW_EH_PE (aligned, DW_EH_PE_aligned) \
-  DWARF_ONE_KNOWN_DW_EH_PE (datarel, DW_EH_PE_datarel) \
-  DWARF_ONE_KNOWN_DW_EH_PE (funcrel, DW_EH_PE_funcrel) \
-  DWARF_ONE_KNOWN_DW_EH_PE (indirect, DW_EH_PE_indirect) \
-  DWARF_ONE_KNOWN_DW_EH_PE (omit, DW_EH_PE_omit) \
-  DWARF_ONE_KNOWN_DW_EH_PE (pcrel, DW_EH_PE_pcrel) \
-  DWARF_ONE_KNOWN_DW_EH_PE (sdata2, DW_EH_PE_sdata2) \
-  DWARF_ONE_KNOWN_DW_EH_PE (sdata4, DW_EH_PE_sdata4) \
-  DWARF_ONE_KNOWN_DW_EH_PE (sdata8, DW_EH_PE_sdata8) \
-  DWARF_ONE_KNOWN_DW_EH_PE (signed, DW_EH_PE_signed) \
-  DWARF_ONE_KNOWN_DW_EH_PE (sleb128, DW_EH_PE_sleb128) \
-  DWARF_ONE_KNOWN_DW_EH_PE (textrel, DW_EH_PE_textrel) \
-  DWARF_ONE_KNOWN_DW_EH_PE (udata2, DW_EH_PE_udata2) \
-  DWARF_ONE_KNOWN_DW_EH_PE (udata4, DW_EH_PE_udata4) \
-  DWARF_ONE_KNOWN_DW_EH_PE (udata8, DW_EH_PE_udata8) \
-  DWARF_ONE_KNOWN_DW_EH_PE (uleb128, DW_EH_PE_uleb128) \
-  /* End of DW_EH_PE_*.  */
-
-#define DWARF_ALL_KNOWN_DW_END \
-  DWARF_ONE_KNOWN_DW_END (big, DW_END_big) \
-  DWARF_ONE_KNOWN_DW_END (default, DW_END_default) \
-  DWARF_ONE_KNOWN_DW_END (little, DW_END_little) \
-  /* End of DW_END_*.  */
-
-#define DWARF_ALL_KNOWN_DW_FORM \
-  DWARF_ONE_KNOWN_DW_FORM (GNU_addr_index, DW_FORM_GNU_addr_index) \
-  DWARF_ONE_KNOWN_DW_FORM (GNU_ref_alt, DW_FORM_GNU_ref_alt) \
-  DWARF_ONE_KNOWN_DW_FORM (GNU_str_index, DW_FORM_GNU_str_index) \
-  DWARF_ONE_KNOWN_DW_FORM (GNU_strp_alt, DW_FORM_GNU_strp_alt) \
-  DWARF_ONE_KNOWN_DW_FORM (addr, DW_FORM_addr) \
-  DWARF_ONE_KNOWN_DW_FORM (addrx, DW_FORM_addrx) \
-  DWARF_ONE_KNOWN_DW_FORM (addrx1, DW_FORM_addrx1) \
-  DWARF_ONE_KNOWN_DW_FORM (addrx2, DW_FORM_addrx2) \
-  DWARF_ONE_KNOWN_DW_FORM (addrx3, DW_FORM_addrx3) \
-  DWARF_ONE_KNOWN_DW_FORM (addrx4, DW_FORM_addrx4) \
-  DWARF_ONE_KNOWN_DW_FORM (block, DW_FORM_block) \
-  DWARF_ONE_KNOWN_DW_FORM (block1, DW_FORM_block1) \
-  DWARF_ONE_KNOWN_DW_FORM (block2, DW_FORM_block2) \
-  DWARF_ONE_KNOWN_DW_FORM (block4, DW_FORM_block4) \
-  DWARF_ONE_KNOWN_DW_FORM (data1, DW_FORM_data1) \
-  DWARF_ONE_KNOWN_DW_FORM (data16, DW_FORM_data16) \
-  DWARF_ONE_KNOWN_DW_FORM (data2, DW_FORM_data2) \
-  DWARF_ONE_KNOWN_DW_FORM (data4, DW_FORM_data4) \
-  DWARF_ONE_KNOWN_DW_FORM (data8, DW_FORM_data8) \
-  DWARF_ONE_KNOWN_DW_FORM (exprloc, DW_FORM_exprloc) \
-  DWARF_ONE_KNOWN_DW_FORM (flag, DW_FORM_flag) \
-  DWARF_ONE_KNOWN_DW_FORM (flag_present, DW_FORM_flag_present) \
-  DWARF_ONE_KNOWN_DW_FORM (implicit_const, DW_FORM_implicit_const) \
-  DWARF_ONE_KNOWN_DW_FORM (indirect, DW_FORM_indirect) \
-  DWARF_ONE_KNOWN_DW_FORM (line_strp, DW_FORM_line_strp) \
-  DWARF_ONE_KNOWN_DW_FORM (loclistx, DW_FORM_loclistx) \
-  DWARF_ONE_KNOWN_DW_FORM (ref1, DW_FORM_ref1) \
-  DWARF_ONE_KNOWN_DW_FORM (ref2, DW_FORM_ref2) \
-  DWARF_ONE_KNOWN_DW_FORM (ref4, DW_FORM_ref4) \
-  DWARF_ONE_KNOWN_DW_FORM (ref8, DW_FORM_ref8) \
-  DWARF_ONE_KNOWN_DW_FORM (ref_addr, DW_FORM_ref_addr) \
-  DWARF_ONE_KNOWN_DW_FORM (ref_sig8, DW_FORM_ref_sig8) \
-  DWARF_ONE_KNOWN_DW_FORM (ref_sup4, DW_FORM_ref_sup4) \
-  DWARF_ONE_KNOWN_DW_FORM (ref_sup8, DW_FORM_ref_sup8) \
-  DWARF_ONE_KNOWN_DW_FORM (ref_udata, DW_FORM_ref_udata) \
-  DWARF_ONE_KNOWN_DW_FORM (rnglistx, DW_FORM_rnglistx) \
-  DWARF_ONE_KNOWN_DW_FORM (sdata, DW_FORM_sdata) \
-  DWARF_ONE_KNOWN_DW_FORM (sec_offset, DW_FORM_sec_offset) \
-  DWARF_ONE_KNOWN_DW_FORM (string, DW_FORM_string) \
-  DWARF_ONE_KNOWN_DW_FORM (strp, DW_FORM_strp) \
-  DWARF_ONE_KNOWN_DW_FORM (strp_sup, DW_FORM_strp_sup) \
-  DWARF_ONE_KNOWN_DW_FORM (strx, DW_FORM_strx) \
-  DWARF_ONE_KNOWN_DW_FORM (strx1, DW_FORM_strx1) \
-  DWARF_ONE_KNOWN_DW_FORM (strx2, DW_FORM_strx2) \
-  DWARF_ONE_KNOWN_DW_FORM (strx3, DW_FORM_strx3) \
-  DWARF_ONE_KNOWN_DW_FORM (strx4, DW_FORM_strx4) \
-  DWARF_ONE_KNOWN_DW_FORM (udata, DW_FORM_udata) \
-  /* End of DW_FORM_*.  */
-
-#define DWARF_ALL_KNOWN_DW_ID \
-  DWARF_ONE_KNOWN_DW_ID (case_insensitive, DW_ID_case_insensitive) \
-  DWARF_ONE_KNOWN_DW_ID (case_sensitive, DW_ID_case_sensitive) \
-  DWARF_ONE_KNOWN_DW_ID (down_case, DW_ID_down_case) \
-  DWARF_ONE_KNOWN_DW_ID (up_case, DW_ID_up_case) \
-  /* End of DW_ID_*.  */
-
-#define DWARF_ALL_KNOWN_DW_INL \
-  DWARF_ONE_KNOWN_DW_INL (declared_inlined, DW_INL_declared_inlined) \
-  DWARF_ONE_KNOWN_DW_INL (declared_not_inlined, DW_INL_declared_not_inlined) \
-  DWARF_ONE_KNOWN_DW_INL (inlined, DW_INL_inlined) \
-  DWARF_ONE_KNOWN_DW_INL (not_inlined, DW_INL_not_inlined) \
-  /* End of DW_INL_*.  */
-
-#define DWARF_ALL_KNOWN_DW_LANG \
-  DWARF_ONE_KNOWN_DW_LANG (Ada83, DW_LANG_Ada83) \
-  DWARF_ONE_KNOWN_DW_LANG (Ada95, DW_LANG_Ada95) \
-  DWARF_ONE_KNOWN_DW_LANG (BLISS, DW_LANG_BLISS) \
-  DWARF_ONE_KNOWN_DW_LANG (C, DW_LANG_C) \
-  DWARF_ONE_KNOWN_DW_LANG (C11, DW_LANG_C11) \
-  DWARF_ONE_KNOWN_DW_LANG (C89, DW_LANG_C89) \
-  DWARF_ONE_KNOWN_DW_LANG (C99, DW_LANG_C99) \
-  DWARF_ONE_KNOWN_DW_LANG (C_plus_plus, DW_LANG_C_plus_plus) \
-  DWARF_ONE_KNOWN_DW_LANG (C_plus_plus_03, DW_LANG_C_plus_plus_03) \
-  DWARF_ONE_KNOWN_DW_LANG (C_plus_plus_11, DW_LANG_C_plus_plus_11) \
-  DWARF_ONE_KNOWN_DW_LANG (C_plus_plus_14, DW_LANG_C_plus_plus_14) \
-  DWARF_ONE_KNOWN_DW_LANG (Cobol74, DW_LANG_Cobol74) \
-  DWARF_ONE_KNOWN_DW_LANG (Cobol85, DW_LANG_Cobol85) \
-  DWARF_ONE_KNOWN_DW_LANG (D, DW_LANG_D) \
-  DWARF_ONE_KNOWN_DW_LANG (Dylan, DW_LANG_Dylan) \
-  DWARF_ONE_KNOWN_DW_LANG (Fortran03, DW_LANG_Fortran03) \
-  DWARF_ONE_KNOWN_DW_LANG (Fortran08, DW_LANG_Fortran08) \
-  DWARF_ONE_KNOWN_DW_LANG (Fortran77, DW_LANG_Fortran77) \
-  DWARF_ONE_KNOWN_DW_LANG (Fortran90, DW_LANG_Fortran90) \
-  DWARF_ONE_KNOWN_DW_LANG (Fortran95, DW_LANG_Fortran95) \
-  DWARF_ONE_KNOWN_DW_LANG (Go, DW_LANG_Go) \
-  DWARF_ONE_KNOWN_DW_LANG (Haskell, DW_LANG_Haskell) \
-  DWARF_ONE_KNOWN_DW_LANG (Java, DW_LANG_Java) \
-  DWARF_ONE_KNOWN_DW_LANG (Julia, DW_LANG_Julia) \
-  DWARF_ONE_KNOWN_DW_LANG (Mips_Assembler, DW_LANG_Mips_Assembler) \
-  DWARF_ONE_KNOWN_DW_LANG (Modula2, DW_LANG_Modula2) \
-  DWARF_ONE_KNOWN_DW_LANG (Modula3, DW_LANG_Modula3) \
-  DWARF_ONE_KNOWN_DW_LANG (OCaml, DW_LANG_OCaml) \
-  DWARF_ONE_KNOWN_DW_LANG (ObjC, DW_LANG_ObjC) \
-  DWARF_ONE_KNOWN_DW_LANG (ObjC_plus_plus, DW_LANG_ObjC_plus_plus) \
-  DWARF_ONE_KNOWN_DW_LANG (OpenCL, DW_LANG_OpenCL) \
-  DWARF_ONE_KNOWN_DW_LANG (PLI, DW_LANG_PLI) \
-  DWARF_ONE_KNOWN_DW_LANG (Pascal83, DW_LANG_Pascal83) \
-  DWARF_ONE_KNOWN_DW_LANG (Python, DW_LANG_Python) \
-  DWARF_ONE_KNOWN_DW_LANG (RenderScript, DW_LANG_RenderScript) \
-  DWARF_ONE_KNOWN_DW_LANG (Rust, DW_LANG_Rust) \
-  DWARF_ONE_KNOWN_DW_LANG (Swift, DW_LANG_Swift) \
-  DWARF_ONE_KNOWN_DW_LANG (UPC, DW_LANG_UPC) \
-  /* End of DW_LANG_*.  */
-
-#define DWARF_ALL_KNOWN_DW_LLE \
-  DWARF_ONE_KNOWN_DW_LLE (base_address, DW_LLE_base_address) \
-  DWARF_ONE_KNOWN_DW_LLE (base_addressx, DW_LLE_base_addressx) \
-  DWARF_ONE_KNOWN_DW_LLE (default_location, DW_LLE_default_location) \
-  DWARF_ONE_KNOWN_DW_LLE (end_of_list, DW_LLE_end_of_list) \
-  DWARF_ONE_KNOWN_DW_LLE (offset_pair, DW_LLE_offset_pair) \
-  DWARF_ONE_KNOWN_DW_LLE (start_end, DW_LLE_start_end) \
-  DWARF_ONE_KNOWN_DW_LLE (start_length, DW_LLE_start_length) \
-  DWARF_ONE_KNOWN_DW_LLE (startx_endx, DW_LLE_startx_endx) \
-  DWARF_ONE_KNOWN_DW_LLE (startx_length, DW_LLE_startx_length) \
-  /* End of DW_LLE_*.  */
-
-#define DWARF_ALL_KNOWN_DW_LLE_GNU \
-  DWARF_ONE_KNOWN_DW_LLE_GNU (base_address_selection_entry, DW_LLE_GNU_base_address_selection_entry) \
-  DWARF_ONE_KNOWN_DW_LLE_GNU (end_of_list_entry, DW_LLE_GNU_end_of_list_entry) \
-  DWARF_ONE_KNOWN_DW_LLE_GNU (start_end_entry, DW_LLE_GNU_start_end_entry) \
-  DWARF_ONE_KNOWN_DW_LLE_GNU (start_length_entry, DW_LLE_GNU_start_length_entry) \
-  /* End of DW_LLE_GNU_*.  */
-
-#define DWARF_ALL_KNOWN_DW_LNCT \
-  DWARF_ONE_KNOWN_DW_LNCT (MD5, DW_LNCT_MD5) \
-  DWARF_ONE_KNOWN_DW_LNCT (directory_index, DW_LNCT_directory_index) \
-  DWARF_ONE_KNOWN_DW_LNCT (path, DW_LNCT_path) \
-  DWARF_ONE_KNOWN_DW_LNCT (size, DW_LNCT_size) \
-  DWARF_ONE_KNOWN_DW_LNCT (timestamp, DW_LNCT_timestamp) \
-  /* End of DW_LNCT_*.  */
-
-#define DWARF_ALL_KNOWN_DW_LNE \
-  DWARF_ONE_KNOWN_DW_LNE (define_file, DW_LNE_define_file) \
-  DWARF_ONE_KNOWN_DW_LNE (end_sequence, DW_LNE_end_sequence) \
-  DWARF_ONE_KNOWN_DW_LNE (set_address, DW_LNE_set_address) \
-  DWARF_ONE_KNOWN_DW_LNE (set_discriminator, DW_LNE_set_discriminator) \
-  /* End of DW_LNE_*.  */
-
-#define DWARF_ALL_KNOWN_DW_LNS \
-  DWARF_ONE_KNOWN_DW_LNS (advance_line, DW_LNS_advance_line) \
-  DWARF_ONE_KNOWN_DW_LNS (advance_pc, DW_LNS_advance_pc) \
-  DWARF_ONE_KNOWN_DW_LNS (const_add_pc, DW_LNS_const_add_pc) \
-  DWARF_ONE_KNOWN_DW_LNS (copy, DW_LNS_copy) \
-  DWARF_ONE_KNOWN_DW_LNS (fixed_advance_pc, DW_LNS_fixed_advance_pc) \
-  DWARF_ONE_KNOWN_DW_LNS (negate_stmt, DW_LNS_negate_stmt) \
-  DWARF_ONE_KNOWN_DW_LNS (set_basic_block, DW_LNS_set_basic_block) \
-  DWARF_ONE_KNOWN_DW_LNS (set_column, DW_LNS_set_column) \
-  DWARF_ONE_KNOWN_DW_LNS (set_epilogue_begin, DW_LNS_set_epilogue_begin) \
-  DWARF_ONE_KNOWN_DW_LNS (set_file, DW_LNS_set_file) \
-  DWARF_ONE_KNOWN_DW_LNS (set_isa, DW_LNS_set_isa) \
-  DWARF_ONE_KNOWN_DW_LNS (set_prologue_end, DW_LNS_set_prologue_end) \
-  /* End of DW_LNS_*.  */
-
-#define DWARF_ALL_KNOWN_DW_MACINFO \
-  DWARF_ONE_KNOWN_DW_MACINFO (define, DW_MACINFO_define) \
-  DWARF_ONE_KNOWN_DW_MACINFO (end_file, DW_MACINFO_end_file) \
-  DWARF_ONE_KNOWN_DW_MACINFO (start_file, DW_MACINFO_start_file) \
-  DWARF_ONE_KNOWN_DW_MACINFO (undef, DW_MACINFO_undef) \
-  DWARF_ONE_KNOWN_DW_MACINFO (vendor_ext, DW_MACINFO_vendor_ext) \
-  /* End of DW_MACINFO_*.  */
-
-#define DWARF_ALL_KNOWN_DW_MACRO \
-  DWARF_ONE_KNOWN_DW_MACRO (define, DW_MACRO_define) \
-  DWARF_ONE_KNOWN_DW_MACRO (define_strp, DW_MACRO_define_strp) \
-  DWARF_ONE_KNOWN_DW_MACRO (define_strx, DW_MACRO_define_strx) \
-  DWARF_ONE_KNOWN_DW_MACRO (define_sup, DW_MACRO_define_sup) \
-  DWARF_ONE_KNOWN_DW_MACRO (end_file, DW_MACRO_end_file) \
-  DWARF_ONE_KNOWN_DW_MACRO (import, DW_MACRO_import) \
-  DWARF_ONE_KNOWN_DW_MACRO (import_sup, DW_MACRO_import_sup) \
-  DWARF_ONE_KNOWN_DW_MACRO (start_file, DW_MACRO_start_file) \
-  DWARF_ONE_KNOWN_DW_MACRO (undef, DW_MACRO_undef) \
-  DWARF_ONE_KNOWN_DW_MACRO (undef_strp, DW_MACRO_undef_strp) \
-  DWARF_ONE_KNOWN_DW_MACRO (undef_strx, DW_MACRO_undef_strx) \
-  DWARF_ONE_KNOWN_DW_MACRO (undef_sup, DW_MACRO_undef_sup) \
-  /* End of DW_MACRO_*.  */
-
-#define DWARF_ALL_KNOWN_DW_OP \
-  DWARF_ONE_KNOWN_DW_OP (GNU_addr_index, DW_OP_GNU_addr_index) \
-  DWARF_ONE_KNOWN_DW_OP (GNU_const_index, DW_OP_GNU_const_index) \
-  DWARF_ONE_KNOWN_DW_OP (GNU_const_type, DW_OP_GNU_const_type) \
-  DWARF_ONE_KNOWN_DW_OP (GNU_convert, DW_OP_GNU_convert) \
-  DWARF_ONE_KNOWN_DW_OP (GNU_deref_type, DW_OP_GNU_deref_type) \
-  DWARF_ONE_KNOWN_DW_OP (GNU_encoded_addr, DW_OP_GNU_encoded_addr) \
-  DWARF_ONE_KNOWN_DW_OP (GNU_entry_value, DW_OP_GNU_entry_value) \
-  DWARF_ONE_KNOWN_DW_OP (GNU_implicit_pointer, DW_OP_GNU_implicit_pointer) \
-  DWARF_ONE_KNOWN_DW_OP (GNU_parameter_ref, DW_OP_GNU_parameter_ref) \
-  DWARF_ONE_KNOWN_DW_OP (GNU_push_tls_address, DW_OP_GNU_push_tls_address) \
-  DWARF_ONE_KNOWN_DW_OP (GNU_regval_type, DW_OP_GNU_regval_type) \
-  DWARF_ONE_KNOWN_DW_OP (GNU_reinterpret, DW_OP_GNU_reinterpret) \
-  DWARF_ONE_KNOWN_DW_OP (GNU_uninit, DW_OP_GNU_uninit) \
-  DWARF_ONE_KNOWN_DW_OP (GNU_variable_value, DW_OP_GNU_variable_value) \
-  DWARF_ONE_KNOWN_DW_OP (abs, DW_OP_abs) \
-  DWARF_ONE_KNOWN_DW_OP (addr, DW_OP_addr) \
-  DWARF_ONE_KNOWN_DW_OP (addrx, DW_OP_addrx) \
-  DWARF_ONE_KNOWN_DW_OP (and, DW_OP_and) \
-  DWARF_ONE_KNOWN_DW_OP (bit_piece, DW_OP_bit_piece) \
-  DWARF_ONE_KNOWN_DW_OP (bra, DW_OP_bra) \
-  DWARF_ONE_KNOWN_DW_OP (breg0, DW_OP_breg0) \
-  DWARF_ONE_KNOWN_DW_OP (breg1, DW_OP_breg1) \
-  DWARF_ONE_KNOWN_DW_OP (breg10, DW_OP_breg10) \
-  DWARF_ONE_KNOWN_DW_OP (breg11, DW_OP_breg11) \
-  DWARF_ONE_KNOWN_DW_OP (breg12, DW_OP_breg12) \
-  DWARF_ONE_KNOWN_DW_OP (breg13, DW_OP_breg13) \
-  DWARF_ONE_KNOWN_DW_OP (breg14, DW_OP_breg14) \
-  DWARF_ONE_KNOWN_DW_OP (breg15, DW_OP_breg15) \
-  DWARF_ONE_KNOWN_DW_OP (breg16, DW_OP_breg16) \
-  DWARF_ONE_KNOWN_DW_OP (breg17, DW_OP_breg17) \
-  DWARF_ONE_KNOWN_DW_OP (breg18, DW_OP_breg18) \
-  DWARF_ONE_KNOWN_DW_OP (breg19, DW_OP_breg19) \
-  DWARF_ONE_KNOWN_DW_OP (breg2, DW_OP_breg2) \
-  DWARF_ONE_KNOWN_DW_OP (breg20, DW_OP_breg20) \
-  DWARF_ONE_KNOWN_DW_OP (breg21, DW_OP_breg21) \
-  DWARF_ONE_KNOWN_DW_OP (breg22, DW_OP_breg22) \
-  DWARF_ONE_KNOWN_DW_OP (breg23, DW_OP_breg23) \
-  DWARF_ONE_KNOWN_DW_OP (breg24, DW_OP_breg24) \
-  DWARF_ONE_KNOWN_DW_OP (breg25, DW_OP_breg25) \
-  DWARF_ONE_KNOWN_DW_OP (breg26, DW_OP_breg26) \
-  DWARF_ONE_KNOWN_DW_OP (breg27, DW_OP_breg27) \
-  DWARF_ONE_KNOWN_DW_OP (breg28, DW_OP_breg28) \
-  DWARF_ONE_KNOWN_DW_OP (breg29, DW_OP_breg29) \
-  DWARF_ONE_KNOWN_DW_OP (breg3, DW_OP_breg3) \
-  DWARF_ONE_KNOWN_DW_OP (breg30, DW_OP_breg30) \
-  DWARF_ONE_KNOWN_DW_OP (breg31, DW_OP_breg31) \
-  DWARF_ONE_KNOWN_DW_OP (breg4, DW_OP_breg4) \
-  DWARF_ONE_KNOWN_DW_OP (breg5, DW_OP_breg5) \
-  DWARF_ONE_KNOWN_DW_OP (breg6, DW_OP_breg6) \
-  DWARF_ONE_KNOWN_DW_OP (breg7, DW_OP_breg7) \
-  DWARF_ONE_KNOWN_DW_OP (breg8, DW_OP_breg8) \
-  DWARF_ONE_KNOWN_DW_OP (breg9, DW_OP_breg9) \
-  DWARF_ONE_KNOWN_DW_OP (bregx, DW_OP_bregx) \
-  DWARF_ONE_KNOWN_DW_OP (call2, DW_OP_call2) \
-  DWARF_ONE_KNOWN_DW_OP (call4, DW_OP_call4) \
-  DWARF_ONE_KNOWN_DW_OP (call_frame_cfa, DW_OP_call_frame_cfa) \
-  DWARF_ONE_KNOWN_DW_OP (call_ref, DW_OP_call_ref) \
-  DWARF_ONE_KNOWN_DW_OP (const1s, DW_OP_const1s) \
-  DWARF_ONE_KNOWN_DW_OP (const1u, DW_OP_const1u) \
-  DWARF_ONE_KNOWN_DW_OP (const2s, DW_OP_const2s) \
-  DWARF_ONE_KNOWN_DW_OP (const2u, DW_OP_const2u) \
-  DWARF_ONE_KNOWN_DW_OP (const4s, DW_OP_const4s) \
-  DWARF_ONE_KNOWN_DW_OP (const4u, DW_OP_const4u) \
-  DWARF_ONE_KNOWN_DW_OP (const8s, DW_OP_const8s) \
-  DWARF_ONE_KNOWN_DW_OP (const8u, DW_OP_const8u) \
-  DWARF_ONE_KNOWN_DW_OP (const_type, DW_OP_const_type) \
-  DWARF_ONE_KNOWN_DW_OP (consts, DW_OP_consts) \
-  DWARF_ONE_KNOWN_DW_OP (constu, DW_OP_constu) \
-  DWARF_ONE_KNOWN_DW_OP (constx, DW_OP_constx) \
-  DWARF_ONE_KNOWN_DW_OP (convert, DW_OP_convert) \
-  DWARF_ONE_KNOWN_DW_OP (deref, DW_OP_deref) \
-  DWARF_ONE_KNOWN_DW_OP (deref_size, DW_OP_deref_size) \
-  DWARF_ONE_KNOWN_DW_OP (deref_type, DW_OP_deref_type) \
-  DWARF_ONE_KNOWN_DW_OP (div, DW_OP_div) \
-  DWARF_ONE_KNOWN_DW_OP (drop, DW_OP_drop) \
-  DWARF_ONE_KNOWN_DW_OP (dup, DW_OP_dup) \
-  DWARF_ONE_KNOWN_DW_OP (entry_value, DW_OP_entry_value) \
-  DWARF_ONE_KNOWN_DW_OP (eq, DW_OP_eq) \
-  DWARF_ONE_KNOWN_DW_OP (fbreg, DW_OP_fbreg) \
-  DWARF_ONE_KNOWN_DW_OP (form_tls_address, DW_OP_form_tls_address) \
-  DWARF_ONE_KNOWN_DW_OP (ge, DW_OP_ge) \
-  DWARF_ONE_KNOWN_DW_OP (gt, DW_OP_gt) \
-  DWARF_ONE_KNOWN_DW_OP (implicit_pointer, DW_OP_implicit_pointer) \
-  DWARF_ONE_KNOWN_DW_OP (implicit_value, DW_OP_implicit_value) \
-  DWARF_ONE_KNOWN_DW_OP (le, DW_OP_le) \
-  DWARF_ONE_KNOWN_DW_OP (lit0, DW_OP_lit0) \
-  DWARF_ONE_KNOWN_DW_OP (lit1, DW_OP_lit1) \
-  DWARF_ONE_KNOWN_DW_OP (lit10, DW_OP_lit10) \
-  DWARF_ONE_KNOWN_DW_OP (lit11, DW_OP_lit11) \
-  DWARF_ONE_KNOWN_DW_OP (lit12, DW_OP_lit12) \
-  DWARF_ONE_KNOWN_DW_OP (lit13, DW_OP_lit13) \
-  DWARF_ONE_KNOWN_DW_OP (lit14, DW_OP_lit14) \
-  DWARF_ONE_KNOWN_DW_OP (lit15, DW_OP_lit15) \
-  DWARF_ONE_KNOWN_DW_OP (lit16, DW_OP_lit16) \
-  DWARF_ONE_KNOWN_DW_OP (lit17, DW_OP_lit17) \
-  DWARF_ONE_KNOWN_DW_OP (lit18, DW_OP_lit18) \
-  DWARF_ONE_KNOWN_DW_OP (lit19, DW_OP_lit19) \
-  DWARF_ONE_KNOWN_DW_OP (lit2, DW_OP_lit2) \
-  DWARF_ONE_KNOWN_DW_OP (lit20, DW_OP_lit20) \
-  DWARF_ONE_KNOWN_DW_OP (lit21, DW_OP_lit21) \
-  DWARF_ONE_KNOWN_DW_OP (lit22, DW_OP_lit22) \
-  DWARF_ONE_KNOWN_DW_OP (lit23, DW_OP_lit23) \
-  DWARF_ONE_KNOWN_DW_OP (lit24, DW_OP_lit24) \
-  DWARF_ONE_KNOWN_DW_OP (lit25, DW_OP_lit25) \
-  DWARF_ONE_KNOWN_DW_OP (lit26, DW_OP_lit26) \
-  DWARF_ONE_KNOWN_DW_OP (lit27, DW_OP_lit27) \
-  DWARF_ONE_KNOWN_DW_OP (lit28, DW_OP_lit28) \
-  DWARF_ONE_KNOWN_DW_OP (lit29, DW_OP_lit29) \
-  DWARF_ONE_KNOWN_DW_OP (lit3, DW_OP_lit3) \
-  DWARF_ONE_KNOWN_DW_OP (lit30, DW_OP_lit30) \
-  DWARF_ONE_KNOWN_DW_OP (lit31, DW_OP_lit31) \
-  DWARF_ONE_KNOWN_DW_OP (lit4, DW_OP_lit4) \
-  DWARF_ONE_KNOWN_DW_OP (lit5, DW_OP_lit5) \
-  DWARF_ONE_KNOWN_DW_OP (lit6, DW_OP_lit6) \
-  DWARF_ONE_KNOWN_DW_OP (lit7, DW_OP_lit7) \
-  DWARF_ONE_KNOWN_DW_OP (lit8, DW_OP_lit8) \
-  DWARF_ONE_KNOWN_DW_OP (lit9, DW_OP_lit9) \
-  DWARF_ONE_KNOWN_DW_OP (lt, DW_OP_lt) \
-  DWARF_ONE_KNOWN_DW_OP (minus, DW_OP_minus) \
-  DWARF_ONE_KNOWN_DW_OP (mod, DW_OP_mod) \
-  DWARF_ONE_KNOWN_DW_OP (mul, DW_OP_mul) \
-  DWARF_ONE_KNOWN_DW_OP (ne, DW_OP_ne) \
-  DWARF_ONE_KNOWN_DW_OP (neg, DW_OP_neg) \
-  DWARF_ONE_KNOWN_DW_OP (nop, DW_OP_nop) \
-  DWARF_ONE_KNOWN_DW_OP (not, DW_OP_not) \
-  DWARF_ONE_KNOWN_DW_OP (or, DW_OP_or) \
-  DWARF_ONE_KNOWN_DW_OP (over, DW_OP_over) \
-  DWARF_ONE_KNOWN_DW_OP (pick, DW_OP_pick) \
-  DWARF_ONE_KNOWN_DW_OP (piece, DW_OP_piece) \
-  DWARF_ONE_KNOWN_DW_OP (plus, DW_OP_plus) \
-  DWARF_ONE_KNOWN_DW_OP (plus_uconst, DW_OP_plus_uconst) \
-  DWARF_ONE_KNOWN_DW_OP (push_object_address, DW_OP_push_object_address) \
-  DWARF_ONE_KNOWN_DW_OP (reg0, DW_OP_reg0) \
-  DWARF_ONE_KNOWN_DW_OP (reg1, DW_OP_reg1) \
-  DWARF_ONE_KNOWN_DW_OP (reg10, DW_OP_reg10) \
-  DWARF_ONE_KNOWN_DW_OP (reg11, DW_OP_reg11) \
-  DWARF_ONE_KNOWN_DW_OP (reg12, DW_OP_reg12) \
-  DWARF_ONE_KNOWN_DW_OP (reg13, DW_OP_reg13) \
-  DWARF_ONE_KNOWN_DW_OP (reg14, DW_OP_reg14) \
-  DWARF_ONE_KNOWN_DW_OP (reg15, DW_OP_reg15) \
-  DWARF_ONE_KNOWN_DW_OP (reg16, DW_OP_reg16) \
-  DWARF_ONE_KNOWN_DW_OP (reg17, DW_OP_reg17) \
-  DWARF_ONE_KNOWN_DW_OP (reg18, DW_OP_reg18) \
-  DWARF_ONE_KNOWN_DW_OP (reg19, DW_OP_reg19) \
-  DWARF_ONE_KNOWN_DW_OP (reg2, DW_OP_reg2) \
-  DWARF_ONE_KNOWN_DW_OP (reg20, DW_OP_reg20) \
-  DWARF_ONE_KNOWN_DW_OP (reg21, DW_OP_reg21) \
-  DWARF_ONE_KNOWN_DW_OP (reg22, DW_OP_reg22) \
-  DWARF_ONE_KNOWN_DW_OP (reg23, DW_OP_reg23) \
-  DWARF_ONE_KNOWN_DW_OP (reg24, DW_OP_reg24) \
-  DWARF_ONE_KNOWN_DW_OP (reg25, DW_OP_reg25) \
-  DWARF_ONE_KNOWN_DW_OP (reg26, DW_OP_reg26) \
-  DWARF_ONE_KNOWN_DW_OP (reg27, DW_OP_reg27) \
-  DWARF_ONE_KNOWN_DW_OP (reg28, DW_OP_reg28) \
-  DWARF_ONE_KNOWN_DW_OP (reg29, DW_OP_reg29) \
-  DWARF_ONE_KNOWN_DW_OP (reg3, DW_OP_reg3) \
-  DWARF_ONE_KNOWN_DW_OP (reg30, DW_OP_reg30) \
-  DWARF_ONE_KNOWN_DW_OP (reg31, DW_OP_reg31) \
-  DWARF_ONE_KNOWN_DW_OP (reg4, DW_OP_reg4) \
-  DWARF_ONE_KNOWN_DW_OP (reg5, DW_OP_reg5) \
-  DWARF_ONE_KNOWN_DW_OP (reg6, DW_OP_reg6) \
-  DWARF_ONE_KNOWN_DW_OP (reg7, DW_OP_reg7) \
-  DWARF_ONE_KNOWN_DW_OP (reg8, DW_OP_reg8) \
-  DWARF_ONE_KNOWN_DW_OP (reg9, DW_OP_reg9) \
-  DWARF_ONE_KNOWN_DW_OP (regval_type, DW_OP_regval_type) \
-  DWARF_ONE_KNOWN_DW_OP (regx, DW_OP_regx) \
-  DWARF_ONE_KNOWN_DW_OP (reinterpret, DW_OP_reinterpret) \
-  DWARF_ONE_KNOWN_DW_OP (rot, DW_OP_rot) \
-  DWARF_ONE_KNOWN_DW_OP (shl, DW_OP_shl) \
-  DWARF_ONE_KNOWN_DW_OP (shr, DW_OP_shr) \
-  DWARF_ONE_KNOWN_DW_OP (shra, DW_OP_shra) \
-  DWARF_ONE_KNOWN_DW_OP (skip, DW_OP_skip) \
-  DWARF_ONE_KNOWN_DW_OP (stack_value, DW_OP_stack_value) \
-  DWARF_ONE_KNOWN_DW_OP (swap, DW_OP_swap) \
-  DWARF_ONE_KNOWN_DW_OP (xderef, DW_OP_xderef) \
-  DWARF_ONE_KNOWN_DW_OP (xderef_size, DW_OP_xderef_size) \
-  DWARF_ONE_KNOWN_DW_OP (xderef_type, DW_OP_xderef_type) \
-  DWARF_ONE_KNOWN_DW_OP (xor, DW_OP_xor) \
-  /* End of DW_OP_*.  */
-
-#define DWARF_ALL_KNOWN_DW_ORD \
-  DWARF_ONE_KNOWN_DW_ORD (col_major, DW_ORD_col_major) \
-  DWARF_ONE_KNOWN_DW_ORD (row_major, DW_ORD_row_major) \
-  /* End of DW_ORD_*.  */
-
-#define DWARF_ALL_KNOWN_DW_RLE \
-  DWARF_ONE_KNOWN_DW_RLE (base_address, DW_RLE_base_address) \
-  DWARF_ONE_KNOWN_DW_RLE (base_addressx, DW_RLE_base_addressx) \
-  DWARF_ONE_KNOWN_DW_RLE (end_of_list, DW_RLE_end_of_list) \
-  DWARF_ONE_KNOWN_DW_RLE (offset_pair, DW_RLE_offset_pair) \
-  DWARF_ONE_KNOWN_DW_RLE (start_end, DW_RLE_start_end) \
-  DWARF_ONE_KNOWN_DW_RLE (start_length, DW_RLE_start_length) \
-  DWARF_ONE_KNOWN_DW_RLE (startx_endx, DW_RLE_startx_endx) \
-  DWARF_ONE_KNOWN_DW_RLE (startx_length, DW_RLE_startx_length) \
-  /* End of DW_RLE_*.  */
-
-#define DWARF_ALL_KNOWN_DW_TAG \
-  DWARF_ONE_KNOWN_DW_TAG (GNU_BINCL, DW_TAG_GNU_BINCL) \
-  DWARF_ONE_KNOWN_DW_TAG (GNU_EINCL, DW_TAG_GNU_EINCL) \
-  DWARF_ONE_KNOWN_DW_TAG (GNU_call_site, DW_TAG_GNU_call_site) \
-  DWARF_ONE_KNOWN_DW_TAG (GNU_call_site_parameter, DW_TAG_GNU_call_site_parameter) \
-  DWARF_ONE_KNOWN_DW_TAG (GNU_formal_parameter_pack, DW_TAG_GNU_formal_parameter_pack) \
-  DWARF_ONE_KNOWN_DW_TAG (GNU_template_parameter_pack, DW_TAG_GNU_template_parameter_pack) \
-  DWARF_ONE_KNOWN_DW_TAG (GNU_template_template_param, DW_TAG_GNU_template_template_param) \
-  DWARF_ONE_KNOWN_DW_TAG (MIPS_loop, DW_TAG_MIPS_loop) \
-  DWARF_ONE_KNOWN_DW_TAG (access_declaration, DW_TAG_access_declaration) \
-  DWARF_ONE_KNOWN_DW_TAG (array_type, DW_TAG_array_type) \
-  DWARF_ONE_KNOWN_DW_TAG (atomic_type, DW_TAG_atomic_type) \
-  DWARF_ONE_KNOWN_DW_TAG (base_type, DW_TAG_base_type) \
-  DWARF_ONE_KNOWN_DW_TAG (call_site, DW_TAG_call_site) \
-  DWARF_ONE_KNOWN_DW_TAG (call_site_parameter, DW_TAG_call_site_parameter) \
-  DWARF_ONE_KNOWN_DW_TAG (catch_block, DW_TAG_catch_block) \
-  DWARF_ONE_KNOWN_DW_TAG (class_template, DW_TAG_class_template) \
-  DWARF_ONE_KNOWN_DW_TAG (class_type, DW_TAG_class_type) \
-  DWARF_ONE_KNOWN_DW_TAG (coarray_type, DW_TAG_coarray_type) \
-  DWARF_ONE_KNOWN_DW_TAG (common_block, DW_TAG_common_block) \
-  DWARF_ONE_KNOWN_DW_TAG (common_inclusion, DW_TAG_common_inclusion) \
-  DWARF_ONE_KNOWN_DW_TAG (compile_unit, DW_TAG_compile_unit) \
-  DWARF_ONE_KNOWN_DW_TAG (condition, DW_TAG_condition) \
-  DWARF_ONE_KNOWN_DW_TAG (const_type, DW_TAG_const_type) \
-  DWARF_ONE_KNOWN_DW_TAG (constant, DW_TAG_constant) \
-  DWARF_ONE_KNOWN_DW_TAG (dwarf_procedure, DW_TAG_dwarf_procedure) \
-  DWARF_ONE_KNOWN_DW_TAG (dynamic_type, DW_TAG_dynamic_type) \
-  DWARF_ONE_KNOWN_DW_TAG (entry_point, DW_TAG_entry_point) \
-  DWARF_ONE_KNOWN_DW_TAG (enumeration_type, DW_TAG_enumeration_type) \
-  DWARF_ONE_KNOWN_DW_TAG (enumerator, DW_TAG_enumerator) \
-  DWARF_ONE_KNOWN_DW_TAG (file_type, DW_TAG_file_type) \
-  DWARF_ONE_KNOWN_DW_TAG (formal_parameter, DW_TAG_formal_parameter) \
-  DWARF_ONE_KNOWN_DW_TAG (format_label, DW_TAG_format_label) \
-  DWARF_ONE_KNOWN_DW_TAG (friend, DW_TAG_friend) \
-  DWARF_ONE_KNOWN_DW_TAG (function_template, DW_TAG_function_template) \
-  DWARF_ONE_KNOWN_DW_TAG (generic_subrange, DW_TAG_generic_subrange) \
-  DWARF_ONE_KNOWN_DW_TAG (immutable_type, DW_TAG_immutable_type) \
-  DWARF_ONE_KNOWN_DW_TAG (imported_declaration, DW_TAG_imported_declaration) \
-  DWARF_ONE_KNOWN_DW_TAG (imported_module, DW_TAG_imported_module) \
-  DWARF_ONE_KNOWN_DW_TAG (imported_unit, DW_TAG_imported_unit) \
-  DWARF_ONE_KNOWN_DW_TAG (inheritance, DW_TAG_inheritance) \
-  DWARF_ONE_KNOWN_DW_TAG (inlined_subroutine, DW_TAG_inlined_subroutine) \
-  DWARF_ONE_KNOWN_DW_TAG (interface_type, DW_TAG_interface_type) \
-  DWARF_ONE_KNOWN_DW_TAG (label, DW_TAG_label) \
-  DWARF_ONE_KNOWN_DW_TAG (lexical_block, DW_TAG_lexical_block) \
-  DWARF_ONE_KNOWN_DW_TAG (member, DW_TAG_member) \
-  DWARF_ONE_KNOWN_DW_TAG (module, DW_TAG_module) \
-  DWARF_ONE_KNOWN_DW_TAG (namelist, DW_TAG_namelist) \
-  DWARF_ONE_KNOWN_DW_TAG (namelist_item, DW_TAG_namelist_item) \
-  DWARF_ONE_KNOWN_DW_TAG (namespace, DW_TAG_namespace) \
-  DWARF_ONE_KNOWN_DW_TAG (packed_type, DW_TAG_packed_type) \
-  DWARF_ONE_KNOWN_DW_TAG (partial_unit, DW_TAG_partial_unit) \
-  DWARF_ONE_KNOWN_DW_TAG (pointer_type, DW_TAG_pointer_type) \
-  DWARF_ONE_KNOWN_DW_TAG (ptr_to_member_type, DW_TAG_ptr_to_member_type) \
-  DWARF_ONE_KNOWN_DW_TAG (reference_type, DW_TAG_reference_type) \
-  DWARF_ONE_KNOWN_DW_TAG (restrict_type, DW_TAG_restrict_type) \
-  DWARF_ONE_KNOWN_DW_TAG (rvalue_reference_type, DW_TAG_rvalue_reference_type) \
-  DWARF_ONE_KNOWN_DW_TAG (set_type, DW_TAG_set_type) \
-  DWARF_ONE_KNOWN_DW_TAG (shared_type, DW_TAG_shared_type) \
-  DWARF_ONE_KNOWN_DW_TAG (skeleton_unit, DW_TAG_skeleton_unit) \
-  DWARF_ONE_KNOWN_DW_TAG (string_type, DW_TAG_string_type) \
-  DWARF_ONE_KNOWN_DW_TAG (structure_type, DW_TAG_structure_type) \
-  DWARF_ONE_KNOWN_DW_TAG (subprogram, DW_TAG_subprogram) \
-  DWARF_ONE_KNOWN_DW_TAG (subrange_type, DW_TAG_subrange_type) \
-  DWARF_ONE_KNOWN_DW_TAG (subroutine_type, DW_TAG_subroutine_type) \
-  DWARF_ONE_KNOWN_DW_TAG (template_alias, DW_TAG_template_alias) \
-  DWARF_ONE_KNOWN_DW_TAG (template_type_parameter, DW_TAG_template_type_parameter) \
-  DWARF_ONE_KNOWN_DW_TAG (template_value_parameter, DW_TAG_template_value_parameter) \
-  DWARF_ONE_KNOWN_DW_TAG (thrown_type, DW_TAG_thrown_type) \
-  DWARF_ONE_KNOWN_DW_TAG (try_block, DW_TAG_try_block) \
-  DWARF_ONE_KNOWN_DW_TAG (type_unit, DW_TAG_type_unit) \
-  DWARF_ONE_KNOWN_DW_TAG (typedef, DW_TAG_typedef) \
-  DWARF_ONE_KNOWN_DW_TAG (union_type, DW_TAG_union_type) \
-  DWARF_ONE_KNOWN_DW_TAG (unspecified_parameters, DW_TAG_unspecified_parameters) \
-  DWARF_ONE_KNOWN_DW_TAG (unspecified_type, DW_TAG_unspecified_type) \
-  DWARF_ONE_KNOWN_DW_TAG (variable, DW_TAG_variable) \
-  DWARF_ONE_KNOWN_DW_TAG (variant, DW_TAG_variant) \
-  DWARF_ONE_KNOWN_DW_TAG (variant_part, DW_TAG_variant_part) \
-  DWARF_ONE_KNOWN_DW_TAG (volatile_type, DW_TAG_volatile_type) \
-  DWARF_ONE_KNOWN_DW_TAG (with_stmt, DW_TAG_with_stmt) \
-  /* End of DW_TAG_*.  */
-
-#define DWARF_ALL_KNOWN_DW_UT \
-  DWARF_ONE_KNOWN_DW_UT (compile, DW_UT_compile) \
-  DWARF_ONE_KNOWN_DW_UT (partial, DW_UT_partial) \
-  DWARF_ONE_KNOWN_DW_UT (skeleton, DW_UT_skeleton) \
-  DWARF_ONE_KNOWN_DW_UT (split_compile, DW_UT_split_compile) \
-  DWARF_ONE_KNOWN_DW_UT (split_type, DW_UT_split_type) \
-  DWARF_ONE_KNOWN_DW_UT (type, DW_UT_type) \
-  /* End of DW_UT_*.  */
-
-#define DWARF_ALL_KNOWN_DW_VIRTUALITY \
-  DWARF_ONE_KNOWN_DW_VIRTUALITY (none, DW_VIRTUALITY_none) \
-  DWARF_ONE_KNOWN_DW_VIRTUALITY (pure_virtual, DW_VIRTUALITY_pure_virtual) \
-  DWARF_ONE_KNOWN_DW_VIRTUALITY (virtual, DW_VIRTUALITY_virtual) \
-  /* End of DW_VIRTUALITY_*.  */
-
-#define DWARF_ALL_KNOWN_DW_VIS \
-  DWARF_ONE_KNOWN_DW_VIS (exported, DW_VIS_exported) \
-  DWARF_ONE_KNOWN_DW_VIS (local, DW_VIS_local) \
-  DWARF_ONE_KNOWN_DW_VIS (qualified, DW_VIS_qualified) \
-  /* End of DW_VIS_*.  */
diff --git a/projects/llm_framework/include/elfutils/libdw.h b/projects/llm_framework/include/elfutils/libdw.h
deleted file mode 100644
index 77174d28..00000000
--- a/projects/llm_framework/include/elfutils/libdw.h
+++ /dev/null
@@ -1,1111 +0,0 @@
-/* Interfaces for libdw.
-   Copyright (C) 2002-2010, 2013, 2014, 2016, 2018 Red Hat, Inc.
-   This file is part of elfutils.
-
-   This file is free software; you can redistribute it and/or modify
-   it under the terms of either
-
-     * the GNU Lesser General Public License as published by the Free
-       Software Foundation; either version 3 of the License, or (at
-       your option) any later version
-
-   or
-
-     * the GNU General Public License as published by the Free
-       Software Foundation; either version 2 of the License, or (at
-       your option) any later version
-
-   or both in parallel, as here.
-
-   elfutils is distributed in the hope that it will be useful, but
-   WITHOUT ANY WARRANTY; without even the implied warranty of
-   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
-   General Public License for more details.
-
-   You should have received copies of the GNU General Public License and
-   the GNU Lesser General Public License along with this program.  If
-   not, see <http://www.gnu.org/licenses/>.  */
-
-#ifndef _LIBDW_H
-#define _LIBDW_H	1
-
-#include <gelf.h>
-#include <stdbool.h>
-#include <stddef.h>
-#include <stdint.h>
-
-/* Mode for the session.  */
-typedef enum
-  {
-    DWARF_C_READ,		/* Read .. */
-    DWARF_C_RDWR,		/* Read and write .. */
-    DWARF_C_WRITE,		/* Write .. */
-  }
-Dwarf_Cmd;
-
-
-/* Callback results.  */
-enum
-{
-  DWARF_CB_OK = 0,
-  DWARF_CB_ABORT
-};
-
-
-/* Error values.  */
-enum
-  {
-    DW_TAG_invalid = 0
-#define DW_TAG_invalid	DW_TAG_invalid
-  };
-
-
-/* Type for offset in DWARF file.  */
-typedef GElf_Off Dwarf_Off;
-
-/* Type for address in DWARF file.  */
-typedef GElf_Addr Dwarf_Addr;
-
-/* Integer types.  Big enough to hold any numeric value.  */
-typedef GElf_Xword Dwarf_Word;
-typedef GElf_Sxword Dwarf_Sword;
-/* For the times we know we do not need that much.  */
-typedef GElf_Half Dwarf_Half;
-
-
-/* DWARF abbreviation record.  */
-typedef struct Dwarf_Abbrev Dwarf_Abbrev;
-
-/* Returned to show the last DIE has be returned.  */
-#define DWARF_END_ABBREV ((Dwarf_Abbrev *) -1l)
-
-/* Source code line information for CU.  */
-typedef struct Dwarf_Lines_s Dwarf_Lines;
-
-/* One source code line information.  */
-typedef struct Dwarf_Line_s Dwarf_Line;
-
-/* Source file information.  */
-typedef struct Dwarf_Files_s Dwarf_Files;
-
-/* One address range record.  */
-typedef struct Dwarf_Arange_s Dwarf_Arange;
-
-/* Address ranges of a file.  */
-typedef struct Dwarf_Aranges_s Dwarf_Aranges;
-
-/* CU representation.  */
-struct Dwarf_CU;
-typedef struct Dwarf_CU Dwarf_CU;
-
-/* Macro information.  */
-typedef struct Dwarf_Macro_s Dwarf_Macro;
-
-/* Attribute representation.  */
-typedef struct
-{
-  unsigned int code;
-  unsigned int form;
-  unsigned char *valp;
-  struct Dwarf_CU *cu;
-} Dwarf_Attribute;
-
-
-/* Data block representation.  */
-typedef struct
-{
-  Dwarf_Word length;
-  unsigned char *data;
-} Dwarf_Block;
-
-
-/* DIE information.  */
-typedef struct
-{
-  /* The offset can be computed from the address.  */
-  void *addr;
-  struct Dwarf_CU *cu;
-  Dwarf_Abbrev *abbrev;
-  // XXX We'll see what other information will be needed.
-  long int padding__;
-} Dwarf_Die;
-
-/* Returned to show the last DIE has be returned.  */
-#define DWARF_END_DIE ((Dwarf_Die *) -1l)
-
-
-/* Global symbol information.  */
-typedef struct
-{
-  Dwarf_Off cu_offset;
-  Dwarf_Off die_offset;
-  const char *name;
-} Dwarf_Global;
-
-
-/* One operation in a DWARF location expression.
-   A location expression is an array of these.  */
-typedef struct
-{
-  uint8_t atom;			/* Operation */
-  Dwarf_Word number;		/* Operand */
-  Dwarf_Word number2;		/* Possible second operand */
-  Dwarf_Word offset;		/* Offset in location expression */
-} Dwarf_Op;
-
-
-/* This describes one Common Information Entry read from a CFI section.
-   Pointers here point into the DATA->d_buf block passed to dwarf_next_cfi.  */
-typedef struct
-{
-  Dwarf_Off CIE_id;	 /* Always DW_CIE_ID_64 in Dwarf_CIE structures.  */
-
-  /* Instruction stream describing initial state used by FDEs.  If
-     we did not understand the whole augmentation string and it did
-     not use 'z', then there might be more augmentation data here
-     (and in FDEs) before the actual instructions.  */
-  const uint8_t *initial_instructions;
-  const uint8_t *initial_instructions_end;
-
-  Dwarf_Word code_alignment_factor;
-  Dwarf_Sword data_alignment_factor;
-  Dwarf_Word return_address_register;
-
-  const char *augmentation;	/* Augmentation string.  */
-
-  /* Augmentation data, might be NULL.  The size is correct only if
-     we understood the augmentation string sufficiently.  */
-  const uint8_t *augmentation_data;
-  size_t augmentation_data_size;
-  size_t fde_augmentation_data_size;
-} Dwarf_CIE;
-
-/* This describes one Frame Description Entry read from a CFI section.
-   Pointers here point into the DATA->d_buf block passed to dwarf_next_cfi.  */
-typedef struct
-{
-  /* Section offset of CIE this FDE refers to.  This will never be
-     DW_CIE_ID_64 in an FDE.  If this value is DW_CIE_ID_64, this is
-     actually a Dwarf_CIE structure.  */
-  Dwarf_Off CIE_pointer;
-
-  /* We can't really decode anything further without looking up the CIE
-     and checking its augmentation string.  Here follows the encoded
-     initial_location and address_range, then any augmentation data,
-     then the instruction stream.  This FDE describes PC locations in
-     the byte range [initial_location, initial_location+address_range).
-     When the CIE augmentation string uses 'z', the augmentation data is
-     a DW_FORM_block (self-sized).  Otherwise, when we understand the
-     augmentation string completely, fde_augmentation_data_size gives
-     the number of bytes of augmentation data before the instructions.  */
-  const uint8_t *start;
-  const uint8_t *end;
-} Dwarf_FDE;
-
-/* Each entry in a CFI section is either a CIE described by Dwarf_CIE or
-   an FDE described by Dward_FDE.  Check CIE_id to see which you have.  */
-typedef union
-{
-  Dwarf_Off CIE_id;	 /* Always DW_CIE_ID_64 in Dwarf_CIE structures.  */
-  Dwarf_CIE cie;
-  Dwarf_FDE fde;
-} Dwarf_CFI_Entry;
-
-/* Same as DW_CIE_ID_64 from dwarf.h to keep libdw.h independent.  */
-#define LIBDW_CIE_ID 0xffffffffffffffffULL
-#define dwarf_cfi_cie_p(entry)	((entry)->cie.CIE_id == LIBDW_CIE_ID)
-
-/* Opaque type representing a frame state described by CFI.  */
-typedef struct Dwarf_Frame_s Dwarf_Frame;
-
-/* Opaque type representing a CFI section found in a DWARF or ELF file.  */
-typedef struct Dwarf_CFI_s Dwarf_CFI;
-
-
-/* Handle for debug sessions.  */
-typedef struct Dwarf Dwarf;
-
-
-/* Out-Of-Memory handler.  */
-typedef void (*__noreturn_attribute__ Dwarf_OOM) (void);
-
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* Create a handle for a new debug session.  */
-extern Dwarf *dwarf_begin (int fildes, Dwarf_Cmd cmd);
-
-/* Create a handle for a new debug session for an ELF file.  */
-extern Dwarf *dwarf_begin_elf (Elf *elf, Dwarf_Cmd cmd, Elf_Scn *scngrp);
-
-/* Retrieve ELF descriptor used for DWARF access.  */
-extern Elf *dwarf_getelf (Dwarf *dwarf);
-
-/* Retrieve DWARF descriptor used for a Dwarf_Die or Dwarf_Attribute.
-   A Dwarf_Die or a Dwarf_Attribute is associated with a particular
-   Dwarf_CU handle.  This function returns the DWARF descriptor for
-   that Dwarf_CU.  */
-extern Dwarf *dwarf_cu_getdwarf (Dwarf_CU *cu);
-
-/* Retrieves the DWARF descriptor for debugaltlink data.  Returns NULL
-   if no alternate debug data has been supplied yet.  libdw will try
-   to set the alt file on first use of an alt FORM if not yet explicitly
-   provided by dwarf_setalt.  */
-extern Dwarf *dwarf_getalt (Dwarf *main);
-
-/* Provides the data referenced by the .gnu_debugaltlink section.  The
-   caller should check that MAIN and ALT match (i.e., they have the
-   same build ID).  It is the responsibility of the caller to ensure
-   that the data referenced by ALT stays valid while it is used by
-   MAIN, until dwarf_setalt is called on MAIN with a different
-   descriptor, or dwarf_end.  Must be called before inspecting DIEs
-   that might have alt FORMs.  Otherwise libdw will try to set the
-   alt file itself on first use.  */
-extern void dwarf_setalt (Dwarf *main, Dwarf *alt);
-
-/* Release debugging handling context.  */
-extern int dwarf_end (Dwarf *dwarf);
-
-
-/* Read the header for the DWARF CU.  */
-extern int dwarf_nextcu (Dwarf *dwarf, Dwarf_Off off, Dwarf_Off *next_off,
-			 size_t *header_sizep, Dwarf_Off *abbrev_offsetp,
-			 uint8_t *address_sizep, uint8_t *offset_sizep)
-     __nonnull_attribute__ (3);
-
-/* Read the header of a DWARF CU or type unit.  If TYPE_SIGNATUREP is not
-   null, this reads a type unit from the .debug_types section; otherwise
-   this reads a CU from the .debug_info section.  */
-extern int dwarf_next_unit (Dwarf *dwarf, Dwarf_Off off, Dwarf_Off *next_off,
-			    size_t *header_sizep, Dwarf_Half *versionp,
-			    Dwarf_Off *abbrev_offsetp,
-			    uint8_t *address_sizep, uint8_t *offset_sizep,
-			    uint64_t *type_signaturep, Dwarf_Off *type_offsetp)
-     __nonnull_attribute__ (3);
-
-
-/* Gets the next Dwarf_CU (unit), version, unit type and if available
-   the CU DIE and sub (type) DIE of the unit.  Returns 0 on success,
-   -1 on error or 1 if there are no more units.  To start iterating
-   provide NULL for CU.  If version < 5 the unit type is set from the
-   CU DIE if available (DW_UT_compile for DW_TAG_compile_unit,
-   DW_UT_type for DW_TAG_type_unit or DW_UT_partial for
-   DW_TAG_partial_unit), otherwise it is set to zero.  If unavailable
-   (the version or unit type is unknown) the CU DIE is cleared.
-   Likewise if the sub DIE isn't isn't available (the unit type is not
-   DW_UT_type or DW_UT_split_type) the sub DIE tag is cleared.  */
-extern int dwarf_get_units (Dwarf *dwarf, Dwarf_CU *cu, Dwarf_CU **next_cu,
-			    Dwarf_Half *version, uint8_t *unit_type,
-			    Dwarf_Die *cudie, Dwarf_Die *subdie)
-     __nonnull_attribute__ (3);
-
-/* Provides information and DIEs associated with the given Dwarf_CU
-   unit.  Returns -1 on error, zero on success. Arguments not needed
-   may be NULL.  If they are NULL and aren't known yet, they won't be
-   looked up.  If the subdie doesn't exist for this unit_type it will
-   be cleared.  If there is no unit_id for this unit type it will be
-   set to zero.  */
-extern int dwarf_cu_info (Dwarf_CU *cu,
-			  Dwarf_Half *version, uint8_t *unit_type,
-			  Dwarf_Die *cudie, Dwarf_Die *subdie,
-			  uint64_t *unit_id,
-			  uint8_t *address_size, uint8_t *offset_size);
-
-/* Decode one DWARF CFI entry (CIE or FDE) from the raw section data.
-   The E_IDENT from the originating ELF file indicates the address
-   size and byte order used in the CFI section contained in DATA;
-   EH_FRAME_P should be true for .eh_frame format and false for
-   .debug_frame format.  OFFSET is the byte position in the section
-   to start at; on return *NEXT_OFFSET is filled in with the byte
-   position immediately after this entry.
-
-   On success, returns 0 and fills in *ENTRY; use dwarf_cfi_cie_p to
-   see whether ENTRY->cie or ENTRY->fde is valid.
-
-   On errors, returns -1.  Some format errors will permit safely
-   skipping to the next CFI entry though the current one is unusable.
-   In that case, *NEXT_OFF will be updated before a -1 return.
-
-   If there are no more CFI entries left in the section,
-   returns 1 and sets *NEXT_OFFSET to (Dwarf_Off) -1.  */
-extern int dwarf_next_cfi (const unsigned char e_ident[],
-			   Elf_Data *data, bool eh_frame_p,
-			   Dwarf_Off offset, Dwarf_Off *next_offset,
-			   Dwarf_CFI_Entry *entry)
-  __nonnull_attribute__ (1, 2, 5, 6);
-
-/* Use the CFI in the DWARF .debug_frame section.
-   Returns NULL if there is no such section (not an error).
-   The pointer returned can be used until dwarf_end is called on DWARF,
-   and must not be passed to dwarf_cfi_end.
-   Calling this more than once returns the same pointer.  */
-extern Dwarf_CFI *dwarf_getcfi (Dwarf *dwarf);
-
-/* Use the CFI in the ELF file's exception-handling data.
-   Returns NULL if there is no such data.
-   The pointer returned can be used until elf_end is called on ELF,
-   and must be passed to dwarf_cfi_end before then.
-   Calling this more than once allocates independent data structures.  */
-extern Dwarf_CFI *dwarf_getcfi_elf (Elf *elf);
-
-/* Release resources allocated by dwarf_getcfi_elf.  */
-extern int dwarf_cfi_end (Dwarf_CFI *cache);
-
-
-/* Return DIE at given offset in .debug_info section.  */
-extern Dwarf_Die *dwarf_offdie (Dwarf *dbg, Dwarf_Off offset,
-				Dwarf_Die *result) __nonnull_attribute__ (3);
-
-/* Return DIE at given offset in .debug_types section.  */
-extern Dwarf_Die *dwarf_offdie_types (Dwarf *dbg, Dwarf_Off offset,
-				      Dwarf_Die *result)
-     __nonnull_attribute__ (3);
-
-/* Return offset of DIE.  */
-extern Dwarf_Off dwarf_dieoffset (Dwarf_Die *die);
-
-/* Return offset of DIE in CU.  */
-extern Dwarf_Off dwarf_cuoffset (Dwarf_Die *die);
-
-/* Return CU DIE containing given DIE.  */
-extern Dwarf_Die *dwarf_diecu (Dwarf_Die *die, Dwarf_Die *result,
-			       uint8_t *address_sizep, uint8_t *offset_sizep)
-     __nonnull_attribute__ (2);
-
-/* Given a Dwarf_Die addr returns a (reconstructed) Dwarf_Die, or NULL
-   if the given addr didn't come from a valid Dwarf_Die.  In particular
-   it will make sure that the correct Dwarf_CU pointer is set for the
-   Dwarf_Die, the Dwarf_Abbrev pointer will not be set up yet (it will
-   only be once the Dwarf_Die is used to read attributes, children or
-   siblings).  This functions can be used to keep a reference to a
-   Dwarf_Die which you want to refer to later.  The addr, and the result
-   of this function, is only valid while the associated Dwarf is valid.  */
-extern Dwarf_Die *dwarf_die_addr_die (Dwarf *dbg, void *addr,
-				      Dwarf_Die *result)
-     __nonnull_attribute__ (3);
-
-/* Return the CU DIE and the header info associated with a Dwarf_Die
-   or Dwarf_Attribute.  A Dwarf_Die or a Dwarf_Attribute is associated
-   with a particular Dwarf_CU handle.  This function returns the CU or
-   type unit DIE and header information for that Dwarf_CU.  The
-   returned DIE is either a compile_unit, partial_unit or type_unit.
-   If it is a type_unit, then the type signature and type offset are
-   also provided, otherwise type_offset will be set to zero.  See also
-   dwarf_diecu and dwarf_next_unit.  */
-extern Dwarf_Die *dwarf_cu_die (Dwarf_CU *cu, Dwarf_Die *result,
-				Dwarf_Half *versionp,
-				Dwarf_Off *abbrev_offsetp,
-				uint8_t *address_sizep,
-				uint8_t *offset_sizep,
-				uint64_t *type_signaturep,
-				Dwarf_Off *type_offsetp)
-     __nonnull_attribute__ (2);
-
-/* Return CU DIE containing given address.  */
-extern Dwarf_Die *dwarf_addrdie (Dwarf *dbg, Dwarf_Addr addr,
-				 Dwarf_Die *result) __nonnull_attribute__ (3);
-
-/* Return child of current DIE.  */
-extern int dwarf_child (Dwarf_Die *die, Dwarf_Die *result)
-     __nonnull_attribute__ (2);
-
-/* Locates the first sibling of DIE and places it in RESULT.
-   Returns 0 if a sibling was found, -1 if something went wrong.
-   Returns 1 if no sibling could be found and, if RESULT is not
-   the same as DIE, it sets RESULT->addr to the address of the
-   (non-sibling) DIE that follows this one, or NULL if this DIE
-   was the last one in the compilation unit.  */
-extern int dwarf_siblingof (Dwarf_Die *die, Dwarf_Die *result)
-     __nonnull_attribute__ (2);
-
-/* For type aliases and qualifier type DIEs, which don't modify or
-   change the structural layout of the underlying type, follow the
-   DW_AT_type attribute (recursively) and return the underlying type
-   Dwarf_Die.
-
-   Returns 0 when RESULT contains a Dwarf_Die (possibly equal to the
-   given DIE) that isn't a type alias or qualifier type.  Returns 1
-   when RESULT contains a type alias or qualifier Dwarf_Die that
-   couldn't be peeled further (it doesn't have a DW_TAG_type
-   attribute).  Returns -1 when an error occurred.
-
-   The current DWARF specification defines one type alias tag
-   (DW_TAG_typedef) and seven modifier/qualifier type tags
-   (DW_TAG_const_type, DW_TAG_volatile_type, DW_TAG_restrict_type,
-   DW_TAG_atomic_type, DW_TAG_immutable_type, DW_TAG_packed_type and
-   DW_TAG_shared_type).  This function won't peel modifier type
-   tags that change the way the underlying type is accessed such
-   as the pointer or reference type tags (DW_TAG_pointer_type,
-   DW_TAG_reference_type or DW_TAG_rvalue_reference_type).
-
-   A future version of this function might peel other alias or
-   qualifier type tags if a future DWARF version or GNU extension
-   defines other type aliases or qualifier type tags that don't modify,
-   change the structural layout or the way to access the underlying type.  */
-extern int dwarf_peel_type (Dwarf_Die *die, Dwarf_Die *result)
-    __nonnull_attribute__ (2);
-
-/* Check whether the DIE has children.  */
-extern int dwarf_haschildren (Dwarf_Die *die) __nonnull_attribute__ (1);
-
-/* Walks the attributes of DIE, starting at the one OFFSET bytes in,
-   calling the CALLBACK function for each one.  Stops if the callback
-   function ever returns a value other than DWARF_CB_OK and returns the
-   offset of the offending attribute.  If the end of the attributes
-   is reached 1 is returned.  If something goes wrong -1 is returned and
-   the dwarf error number is set.  */
-extern ptrdiff_t dwarf_getattrs (Dwarf_Die *die,
-				 int (*callback) (Dwarf_Attribute *, void *),
-				 void *arg, ptrdiff_t offset)
-     __nonnull_attribute__ (2);
-
-/* Return tag of given DIE.  */
-extern int dwarf_tag (Dwarf_Die *die) __nonnull_attribute__ (1);
-
-
-/* Return specific attribute of DIE.  */
-extern Dwarf_Attribute *dwarf_attr (Dwarf_Die *die, unsigned int search_name,
-				    Dwarf_Attribute *result)
-     __nonnull_attribute__ (3);
-
-/* Check whether given DIE has specific attribute.  */
-extern int dwarf_hasattr (Dwarf_Die *die, unsigned int search_name);
-
-/* These are the same as dwarf_attr and dwarf_hasattr, respectively,
-   but they resolve an indirect attribute through
-   DW_AT_abstract_origin, DW_AT_specification or, if the DIE is a
-   top-level split CU, the skeleton DIE.  Note that the attribute
-   might come from a DIE in a different CU (possibly from a different
-   Dwarf file).  In that case all attribute information needs to be
-   resolved through the CU associated with the returned
-   Dwarf_Attribute.  The dwarf_form functions already do this
-   automatically.  */
-extern Dwarf_Attribute *dwarf_attr_integrate (Dwarf_Die *die,
-					      unsigned int search_name,
-					      Dwarf_Attribute *result)
-     __nonnull_attribute__ (3);
-extern int dwarf_hasattr_integrate (Dwarf_Die *die, unsigned int search_name);
-
-
-
-
-/* Check whether given attribute has specific form.  */
-extern int dwarf_hasform (Dwarf_Attribute *attr, unsigned int search_form);
-
-/* Return attribute code of given attribute.  */
-extern unsigned int dwarf_whatattr (Dwarf_Attribute *attr);
-
-/* Return form code of given attribute.  */
-extern unsigned int dwarf_whatform (Dwarf_Attribute *attr);
-
-
-/* Return string associated with given attribute.  */
-extern const char *dwarf_formstring (Dwarf_Attribute *attrp);
-
-/* Return unsigned constant represented by attribute.  */
-extern int dwarf_formudata (Dwarf_Attribute *attr, Dwarf_Word *return_uval)
-     __nonnull_attribute__ (2);
-
-/* Return signed constant represented by attribute.  */
-extern int dwarf_formsdata (Dwarf_Attribute *attr, Dwarf_Sword *return_uval)
-     __nonnull_attribute__ (2);
-
-/* Return address represented by attribute.  */
-extern int dwarf_formaddr (Dwarf_Attribute *attr, Dwarf_Addr *return_addr)
-     __nonnull_attribute__ (2);
-
-/* This function is deprecated.  Always use dwarf_formref_die instead.
-   Return reference offset represented by attribute.  */
-extern int dwarf_formref (Dwarf_Attribute *attr, Dwarf_Off *return_offset)
-     __nonnull_attribute__ (2) __deprecated_attribute__;
-
-/* Look up the DIE in a reference-form attribute.  */
-extern Dwarf_Die *dwarf_formref_die (Dwarf_Attribute *attr, Dwarf_Die *die_mem)
-     __nonnull_attribute__ (2);
-
-/* Return block represented by attribute.  */
-extern int dwarf_formblock (Dwarf_Attribute *attr, Dwarf_Block *return_block)
-     __nonnull_attribute__ (2);
-
-/* Return flag represented by attribute.  */
-extern int dwarf_formflag (Dwarf_Attribute *attr, bool *return_bool)
-     __nonnull_attribute__ (2);
-
-
-/* Simplified attribute value access functions.  */
-
-/* Return string in name attribute of DIE.  */
-extern const char *dwarf_diename (Dwarf_Die *die);
-
-/* Return high PC attribute of DIE.  */
-extern int dwarf_highpc (Dwarf_Die *die, Dwarf_Addr *return_addr)
-     __nonnull_attribute__ (2);
-
-/* Return low PC attribute of DIE.  */
-extern int dwarf_lowpc (Dwarf_Die *die, Dwarf_Addr *return_addr)
-     __nonnull_attribute__ (2);
-
-/* Return entry_pc or low_pc attribute of DIE.  */
-extern int dwarf_entrypc (Dwarf_Die *die, Dwarf_Addr *return_addr)
-     __nonnull_attribute__ (2);
-
-/* Return 1 if DIE's lowpc/highpc or ranges attributes match the PC address,
-   0 if not, or -1 for errors.  */
-extern int dwarf_haspc (Dwarf_Die *die, Dwarf_Addr pc);
-
-/* Enumerate the PC address ranges covered by this DIE, covering all
-   addresses where dwarf_haspc returns true.  In the first call OFFSET
-   should be zero and *BASEP need not be initialized.  Returns -1 for
-   errors, zero when there are no more address ranges to report, or a
-   nonzero OFFSET value to pass to the next call.  Each subsequent call
-   must preserve *BASEP from the prior call.  Successful calls fill in
-   *STARTP and *ENDP with a contiguous address range.  */
-extern ptrdiff_t dwarf_ranges (Dwarf_Die *die,
-			       ptrdiff_t offset, Dwarf_Addr *basep,
-			       Dwarf_Addr *startp, Dwarf_Addr *endp);
-
-
-/* Return byte size attribute of DIE.  */
-extern int dwarf_bytesize (Dwarf_Die *die);
-
-/* Return bit size attribute of DIE.  */
-extern int dwarf_bitsize (Dwarf_Die *die);
-
-/* Return bit offset attribute of DIE.  */
-extern int dwarf_bitoffset (Dwarf_Die *die);
-
-/* Return array order attribute of DIE.  */
-extern int dwarf_arrayorder (Dwarf_Die *die);
-
-/* Return source language attribute of DIE.  */
-extern int dwarf_srclang (Dwarf_Die *die);
-
-
-/* Get abbreviation at given offset for given DIE.  */
-extern Dwarf_Abbrev *dwarf_getabbrev (Dwarf_Die *die, Dwarf_Off offset,
-				      size_t *lengthp);
-
-/* Get abbreviation at given offset in .debug_abbrev section.  */
-extern int dwarf_offabbrev (Dwarf *dbg, Dwarf_Off offset, size_t *lengthp,
-			    Dwarf_Abbrev *abbrevp)
-     __nonnull_attribute__ (4);
-
-/* Get abbreviation code.  */
-extern unsigned int dwarf_getabbrevcode (Dwarf_Abbrev *abbrev);
-
-/* Get abbreviation tag.  */
-extern unsigned int dwarf_getabbrevtag (Dwarf_Abbrev *abbrev);
-
-/* Return true if abbreviation is children flag set.  */
-extern int dwarf_abbrevhaschildren (Dwarf_Abbrev *abbrev);
-
-/* Get number of attributes of abbreviation.  */
-extern int dwarf_getattrcnt (Dwarf_Abbrev *abbrev, size_t *attrcntp)
-     __nonnull_attribute__ (2);
-
-/* Get specific attribute of abbreviation.  */
-extern int dwarf_getabbrevattr (Dwarf_Abbrev *abbrev, size_t idx,
-				unsigned int *namep, unsigned int *formp,
-				Dwarf_Off *offset);
-
-/* Get specific attribute of abbreviation and any data encoded with it.
-   Specifically for DW_FORM_implicit_const data will be set to the
-   constant value associated.  */
-extern int dwarf_getabbrevattr_data (Dwarf_Abbrev *abbrev, size_t idx,
-				     unsigned int *namep, unsigned int *formp,
-				     Dwarf_Sword *datap, Dwarf_Off *offset);
-
-/* Get string from-debug_str section.  */
-extern const char *dwarf_getstring (Dwarf *dbg, Dwarf_Off offset,
-				    size_t *lenp);
-
-
-/* Get public symbol information.  */
-extern ptrdiff_t dwarf_getpubnames (Dwarf *dbg,
-				    int (*callback) (Dwarf *, Dwarf_Global *,
-						     void *),
-				    void *arg, ptrdiff_t offset)
-     __nonnull_attribute__ (2);
-
-
-/* Get source file information for CU.  */
-extern int dwarf_getsrclines (Dwarf_Die *cudie, Dwarf_Lines **lines,
-			      size_t *nlines) __nonnull_attribute__ (2, 3);
-
-/* Return one of the source lines of the CU.  */
-extern Dwarf_Line *dwarf_onesrcline (Dwarf_Lines *lines, size_t idx);
-
-/* Get the file source files used in the CU.  */
-extern int dwarf_getsrcfiles (Dwarf_Die *cudie, Dwarf_Files **files,
-			      size_t *nfiles)
-     __nonnull_attribute__ (2);
-
-
-/* Get source for address in CU.  */
-extern Dwarf_Line *dwarf_getsrc_die (Dwarf_Die *cudie, Dwarf_Addr addr);
-
-/* Get source for file and line number.  */
-extern int dwarf_getsrc_file (Dwarf *dbg, const char *fname, int line, int col,
-			      Dwarf_Line ***srcsp, size_t *nsrcs)
-     __nonnull_attribute__ (2, 5, 6);
-
-
-/* Return line address.  */
-extern int dwarf_lineaddr (Dwarf_Line *line, Dwarf_Addr *addrp);
-
-/* Return line VLIW operation index.  */
-extern int dwarf_lineop_index (Dwarf_Line *line, unsigned int *op_indexp);
-
-/* Return line number.  */
-extern int dwarf_lineno (Dwarf_Line *line, int *linep)
-     __nonnull_attribute__ (2);
-
-/* Return column in line.  */
-extern int dwarf_linecol (Dwarf_Line *line, int *colp)
-     __nonnull_attribute__ (2);
-
-/* Return true if record is for beginning of a statement.  */
-extern int dwarf_linebeginstatement (Dwarf_Line *line, bool *flagp)
-     __nonnull_attribute__ (2);
-
-/* Return true if record is for end of sequence.  */
-extern int dwarf_lineendsequence (Dwarf_Line *line, bool *flagp)
-     __nonnull_attribute__ (2);
-
-/* Return true if record is for beginning of a basic block.  */
-extern int dwarf_lineblock (Dwarf_Line *line, bool *flagp)
-     __nonnull_attribute__ (2);
-
-/* Return true if record is for end of prologue.  */
-extern int dwarf_lineprologueend (Dwarf_Line *line, bool *flagp)
-     __nonnull_attribute__ (2);
-
-/* Return true if record is for beginning of epilogue.  */
-extern int dwarf_lineepiloguebegin (Dwarf_Line *line, bool *flagp)
-     __nonnull_attribute__ (2);
-
-/* Return instruction-set architecture in this record.  */
-extern int dwarf_lineisa (Dwarf_Line *line, unsigned int *isap)
-     __nonnull_attribute__ (2);
-
-/* Return code path discriminator in this record.  */
-extern int dwarf_linediscriminator (Dwarf_Line *line, unsigned int *discp)
-     __nonnull_attribute__ (2);
-
-
-/* Find line information for address.  The returned string is NULL when
-   an error occurred, or the file path.  The file path is either absolute
-   or relative to the compilation directory.  See dwarf_decl_file.  */
-extern const char *dwarf_linesrc (Dwarf_Line *line,
-				  Dwarf_Word *mtime, Dwarf_Word *length);
-
-/* Return file information.  The returned string is NULL when
-   an error occurred, or the file path.  The file path is either absolute
-   or relative to the compilation directory.  See dwarf_decl_file.  */
-extern const char *dwarf_filesrc (Dwarf_Files *file, size_t idx,
-				  Dwarf_Word *mtime, Dwarf_Word *length);
-
-/* Return the Dwarf_Files and index associated with the given Dwarf_Line.  */
-extern int dwarf_line_file (Dwarf_Line *line,
-			    Dwarf_Files **files, size_t *idx)
-    __nonnull_attribute__ (2, 3);
-
-/* Return the directory list used in the file information extracted.
-   (*RESULT)[0] is the CU's DW_AT_comp_dir value, and may be null.
-   (*RESULT)[0..*NDIRS-1] are the compile-time include directory path
-   encoded by the compiler.  */
-extern int dwarf_getsrcdirs (Dwarf_Files *files,
-			     const char *const **result, size_t *ndirs)
-  __nonnull_attribute__ (2, 3);
-
-/* Iterates through the debug line units.  Returns 0 on success, -1 on
-   error or 1 if there are no more units.  To start iterating use zero
-   for OFF and set *CU to NULL.  On success NEXT_OFF will be set to
-   the next offset to use.  The *CU will be set if this line table
-   needed a specific CU and needs to be given when calling
-   dwarf_next_lines again (to help dwarf_next_lines quickly find the
-   next CU).  *CU might be set to NULL when it couldn't be found (the
-   compilation directory entry will be the empty string in that case)
-   or for DWARF 5 or later tables, which are self contained.  SRCFILES
-   and SRCLINES may be NULL if the caller is not interested in the
-   actual line or file table.  On success and when not NULL, NFILES
-   and NLINES will be set to the number of files in the file table and
-   number of lines in the line table.  */
-extern int dwarf_next_lines (Dwarf *dwarf, Dwarf_Off off,
-			     Dwarf_Off *next_off, Dwarf_CU **cu,
-			     Dwarf_Files **srcfiles, size_t *nfiles,
-			     Dwarf_Lines **srclines, size_t *nlines)
-  __nonnull_attribute__ (3,4);
-
-/* Return location expression, decoded as a list of operations.  */
-extern int dwarf_getlocation (Dwarf_Attribute *attr, Dwarf_Op **expr,
-			      size_t *exprlen) __nonnull_attribute__ (2, 3);
-
-/* Return location expressions.  If the attribute uses a location list,
-   ADDRESS selects the relevant location expressions from the list.
-   There can be multiple matches, resulting in multiple expressions to
-   return.  EXPRS and EXPRLENS are parallel arrays of NLOCS slots to
-   fill in.  Returns the number of locations filled in, or -1 for
-   errors.  If EXPRS is a null pointer, stores nothing and returns the
-   total number of locations.  A return value of zero means that the
-   location list indicated no value is accessible.  */
-extern int dwarf_getlocation_addr (Dwarf_Attribute *attr, Dwarf_Addr address,
-				   Dwarf_Op **exprs, size_t *exprlens,
-				   size_t nlocs);
-
-/* Enumerate the locations ranges and descriptions covered by the
-   given attribute.  In the first call OFFSET should be zero and
-   *BASEP need not be initialized.  Returns -1 for errors, zero when
-   there are no more locations to report, or a nonzero OFFSET
-   value to pass to the next call.  Each subsequent call must preserve
-   *BASEP from the prior call.  Successful calls fill in *STARTP and
-   *ENDP with a contiguous address range and *EXPR with a pointer to
-   an array of operations with length *EXPRLEN.  If the attribute
-   describes a single location description and not a location list the
-   first call (with OFFSET zero) will return the location description
-   in *EXPR with *STARTP set to zero and *ENDP set to minus one.  */
-extern ptrdiff_t dwarf_getlocations (Dwarf_Attribute *attr,
-				     ptrdiff_t offset, Dwarf_Addr *basep,
-				     Dwarf_Addr *startp, Dwarf_Addr *endp,
-				     Dwarf_Op **expr, size_t *exprlen);
-
-/* Return the block associated with a DW_OP_implicit_value operation.
-   The OP pointer must point into an expression that dwarf_getlocation
-   or dwarf_getlocation_addr has returned given the same ATTR.  */
-extern int dwarf_getlocation_implicit_value (Dwarf_Attribute *attr,
-					     const Dwarf_Op *op,
-					     Dwarf_Block *return_block)
-  __nonnull_attribute__ (2, 3);
-
-/* Return the attribute indicated by a DW_OP_GNU_implicit_pointer operation.
-   The OP pointer must point into an expression that dwarf_getlocation
-   or dwarf_getlocation_addr has returned given the same ATTR.
-   The result is the DW_AT_location or DW_AT_const_value attribute
-   of the OP->number DIE.  */
-extern int dwarf_getlocation_implicit_pointer (Dwarf_Attribute *attr,
-					       const Dwarf_Op *op,
-					       Dwarf_Attribute *result)
-  __nonnull_attribute__ (2, 3);
-
-/* Return the DIE associated with an operation such as
-   DW_OP_GNU_implicit_pointer, DW_OP_GNU_parameter_ref, DW_OP_GNU_convert,
-   DW_OP_GNU_reinterpret, DW_OP_GNU_const_type, DW_OP_GNU_regval_type or
-   DW_OP_GNU_deref_type.  The OP pointer must point into an expression that
-   dwarf_getlocation or dwarf_getlocation_addr has returned given the same
-   ATTR.  The RESULT is a DIE that expresses a type or value needed by the
-   given OP.  */
-extern int dwarf_getlocation_die (Dwarf_Attribute *attr,
-				  const Dwarf_Op *op,
-				  Dwarf_Die *result)
-  __nonnull_attribute__ (2, 3);
-
-/* Return the attribute expressing a value associated with an operation such
-   as DW_OP_implicit_value, DW_OP_GNU_entry_value or DW_OP_GNU_const_type.
-   The OP pointer must point into an expression that dwarf_getlocation
-   or dwarf_getlocation_addr has returned given the same ATTR.
-   The RESULT is a value expressed by an attribute such as DW_AT_location
-   or DW_AT_const_value.  */
-extern int dwarf_getlocation_attr (Dwarf_Attribute *attr,
-				   const Dwarf_Op *op,
-				   Dwarf_Attribute *result)
-  __nonnull_attribute__ (2, 3);
-
-
-/* Compute the byte-size of a type DIE according to DWARF rules.
-   For most types, this is just DW_AT_byte_size.
-   For DW_TAG_array_type it can apply much more complex rules.  */
-extern int dwarf_aggregate_size (Dwarf_Die *die, Dwarf_Word *size);
-
-/* Given a language code, as returned by dwarf_srclan, get the default
-   lower bound for a subrange type without a lower bound attribute.
-   Returns zero on success or -1 on failure when the given language
-   wasn't recognized.  */
-extern int dwarf_default_lower_bound (int lang, Dwarf_Sword *result)
-  __nonnull_attribute__ (2);
-
-/* Return scope DIEs containing PC address.
-   Sets *SCOPES to a malloc'd array of Dwarf_Die structures,
-   and returns the number of elements in the array.
-   (*SCOPES)[0] is the DIE for the innermost scope containing PC,
-   (*SCOPES)[1] is the DIE for the scope containing that scope, and so on.
-   Returns -1 for errors or 0 if no scopes match PC.  */
-extern int dwarf_getscopes (Dwarf_Die *cudie, Dwarf_Addr pc,
-			    Dwarf_Die **scopes);
-
-/* Return scope DIEs containing the given DIE.
-   Sets *SCOPES to a malloc'd array of Dwarf_Die structures,
-   and returns the number of elements in the array.
-   (*SCOPES)[0] is a copy of DIE.
-   (*SCOPES)[1] is the DIE for the scope containing that scope, and so on.
-   Returns -1 for errors or 0 if DIE is not found in any scope entry.  */
-extern int dwarf_getscopes_die (Dwarf_Die *die, Dwarf_Die **scopes);
-
-
-/* Search SCOPES[0..NSCOPES-1] for a variable called NAME.
-   Ignore the first SKIP_SHADOWS scopes that match the name.
-   If MATCH_FILE is not null, accept only declaration in that source file;
-   if MATCH_LINENO or MATCH_LINECOL are also nonzero, accept only declaration
-   at that line and column.
-
-   If successful, fill in *RESULT with the DIE of the variable found,
-   and return N where SCOPES[N] is the scope defining the variable.
-   Return -1 for errors or -2 for no matching variable found.  */
-extern int dwarf_getscopevar (Dwarf_Die *scopes, int nscopes,
-			      const char *name, int skip_shadows,
-			      const char *match_file,
-			      int match_lineno, int match_linecol,
-			      Dwarf_Die *result);
-
-
-
-/* Return list address ranges.  */
-extern int dwarf_getaranges (Dwarf *dbg, Dwarf_Aranges **aranges,
-			     size_t *naranges)
-     __nonnull_attribute__ (2);
-
-/* Return one of the address range entries.  */
-extern Dwarf_Arange *dwarf_onearange (Dwarf_Aranges *aranges, size_t idx);
-
-/* Return information in address range record.  */
-extern int dwarf_getarangeinfo (Dwarf_Arange *arange, Dwarf_Addr *addrp,
-				Dwarf_Word *lengthp, Dwarf_Off *offsetp);
-
-/* Get address range which includes given address.  */
-extern Dwarf_Arange *dwarf_getarange_addr (Dwarf_Aranges *aranges,
-					   Dwarf_Addr addr);
-
-
-
-/* Get functions in CUDIE.  The given callback will be called for all
-   defining DW_TAG_subprograms in the CU DIE tree.  If the callback
-   returns DWARF_CB_ABORT the return value can be used as offset argument
-   to resume the function to find all remaining functions (this is not
-   really recommended, since it needs to rewalk the CU DIE tree first till
-   that offset is found again).  If the callback returns DWARF_CB_OK
-   dwarf_getfuncs will not return but keep calling the callback for each
-   function DIE it finds.  Pass zero for offset on the first call to walk
-   the full CU DIE tree.  If no more functions can be found and the callback
-   returned DWARF_CB_OK then the function returns zero.  */
-extern ptrdiff_t dwarf_getfuncs (Dwarf_Die *cudie,
-				 int (*callback) (Dwarf_Die *, void *),
-				 void *arg, ptrdiff_t offset);
-
-
-/* Return file name containing definition of the given declaration.
-   Of the DECL has an (indirect, see dwarf_attr_integrate) decl_file
-   attribute.  The returned file path is either absolute, or relative
-   to the compilation directory.  Given the decl DIE, the compilation
-   directory can be retrieved through:
-   dwarf_formstring (dwarf_attr (dwarf_diecu (decl, &cudie, NULL, NULL),
-                                 DW_AT_comp_dir, &attr));
-   Returns NULL if no decl_file could be found or an error occurred.  */
-extern const char *dwarf_decl_file (Dwarf_Die *decl);
-
-/* Get line number of beginning of given declaration.  */
-extern int dwarf_decl_line (Dwarf_Die *decl, int *linep)
-     __nonnull_attribute__ (2);
-
-/* Get column number of beginning of given declaration.  */
-extern int dwarf_decl_column (Dwarf_Die *decl, int *colp)
-     __nonnull_attribute__ (2);
-
-
-/* Return nonzero if given function is an abstract inline definition.  */
-extern int dwarf_func_inline (Dwarf_Die *func);
-
-/* Find each concrete inlined instance of the abstract inline definition.  */
-extern int dwarf_func_inline_instances (Dwarf_Die *func,
-					int (*callback) (Dwarf_Die *, void *),
-					void *arg);
-
-
-/* Find the appropriate PC location or locations for function entry
-   breakpoints for the given DW_TAG_subprogram DIE.  Returns -1 for errors.
-   On success, returns the number of breakpoint locations (never zero)
-   and sets *BKPTS to a malloc'd vector of addresses.  */
-extern int dwarf_entry_breakpoints (Dwarf_Die *die, Dwarf_Addr **bkpts);
-
-
-/* Iterate through the macro unit referenced by CUDIE and call
-   CALLBACK for each macro information entry.  To start the iteration,
-   one would pass DWARF_GETMACROS_START for TOKEN.
-
-   The iteration continues while CALLBACK returns DWARF_CB_OK.  If the
-   callback returns DWARF_CB_ABORT, the iteration stops and a
-   continuation token is returned, which can be used to restart the
-   iteration at the point where it ended.  Returns -1 for errors or 0
-   if there are no more macro entries.
-
-   Note that the Dwarf_Macro pointer passed to the callback is only
-   valid for the duration of the callback invocation.
-
-   For backward compatibility, a token of 0 is accepted for starting
-   the iteration as well, but in that case this interface will refuse
-   to serve opcode 0xff from .debug_macro sections.  Such opcode would
-   be considered invalid and would cause dwarf_getmacros to return
-   with error.  */
-#define DWARF_GETMACROS_START PTRDIFF_MIN
-extern ptrdiff_t dwarf_getmacros (Dwarf_Die *cudie,
-				  int (*callback) (Dwarf_Macro *, void *),
-				  void *arg, ptrdiff_t token)
-     __nonnull_attribute__ (2);
-
-/* This is similar in operation to dwarf_getmacros, but selects the
-   unit to iterate through by offset instead of by CU, and always
-   iterates .debug_macro.  This can be used for handling
-   DW_MACRO_GNU_transparent_include's or similar opcodes.
-
-   TOKEN value of DWARF_GETMACROS_START can be used to start the
-   iteration.
-
-   It is not appropriate to obtain macro unit offset by hand from a CU
-   DIE and then request iteration through this interface.  The reason
-   for this is that if a dwarf_macro_getsrcfiles is later called,
-   there would be no way to figure out what DW_AT_comp_dir was present
-   on the CU DIE, and file names referenced in either the macro unit
-   itself, or the .debug_line unit that it references, might be wrong.
-   Use dwarf_getmacros.  */
-extern ptrdiff_t dwarf_getmacros_off (Dwarf *dbg, Dwarf_Off macoff,
-				      int (*callback) (Dwarf_Macro *, void *),
-				      void *arg, ptrdiff_t token)
-  __nonnull_attribute__ (3);
-
-/* Get the source files used by the macro entry.  You shouldn't assume
-   that Dwarf_Files references will remain valid after MACRO becomes
-   invalid.  (Which is to say it's only valid within the
-   dwarf_getmacros* callback.)  Returns 0 for success or a negative
-   value in case of an error.  */
-extern int dwarf_macro_getsrcfiles (Dwarf *dbg, Dwarf_Macro *macro,
-				    Dwarf_Files **files, size_t *nfiles)
-  __nonnull_attribute__ (2, 3, 4);
-
-/* Return macro opcode.  That's a constant that can be either from
-   DW_MACINFO_* domain or DW_MACRO_GNU_* domain.  The two domains have
-   compatible values, so it's OK to use either of them for
-   comparisons.  The only differences is 0xff, which could be either
-   DW_MACINFO_vendor_ext or a vendor-defined DW_MACRO_* constant.  One
-   would need to look if the CU DIE which the iteration was requested
-   for has attribute DW_AT_macro_info, or either of DW_AT_GNU_macros
-   or DW_AT_macros to differentiate the two interpretations.  */
-extern int dwarf_macro_opcode (Dwarf_Macro *macro, unsigned int *opcodep)
-     __nonnull_attribute__ (2);
-
-/* Get number of parameters of MACRO and store it to *PARAMCNTP.  */
-extern int dwarf_macro_getparamcnt (Dwarf_Macro *macro, size_t *paramcntp);
-
-/* Get IDX-th parameter of MACRO (numbered from zero), and stores it
-   to *ATTRIBUTE.  Returns 0 on success or -1 for errors.
-
-   After a successful call, you can query ATTRIBUTE by dwarf_whatform
-   to determine which of the dwarf_formX calls to make to get actual
-   value out of ATTRIBUTE.  Note that calling dwarf_whatattr is not
-   meaningful for pseudo-attributes formed this way.  */
-extern int dwarf_macro_param (Dwarf_Macro *macro, size_t idx,
-			      Dwarf_Attribute *attribute);
-
-/* Return macro parameter with index 0.  This will return -1 if the
-   parameter is not an integral value.  Use dwarf_macro_param for more
-   general access.  */
-extern int dwarf_macro_param1 (Dwarf_Macro *macro, Dwarf_Word *paramp)
-     __nonnull_attribute__ (2);
-
-/* Return macro parameter with index 1.  This will return -1 if the
-   parameter is not an integral or string value.  Use
-   dwarf_macro_param for more general access.  */
-extern int dwarf_macro_param2 (Dwarf_Macro *macro, Dwarf_Word *paramp,
-			       const char **strp);
-
-/* Compute what's known about a call frame when the PC is at ADDRESS.
-   Returns 0 for success or -1 for errors.
-   On success, *FRAME is a malloc'd pointer.  */
-extern int dwarf_cfi_addrframe (Dwarf_CFI *cache,
-				Dwarf_Addr address, Dwarf_Frame **frame)
-  __nonnull_attribute__ (3);
-
-/* Return the DWARF register number used in FRAME to denote
-   the return address in FRAME's caller frame.  The remaining
-   arguments can be non-null to fill in more information.
-
-   Fill [*START, *END) with the PC range to which FRAME's information applies.
-   Fill in *SIGNALP to indicate whether this is a signal-handling frame.
-   If true, this is the implicit call frame that calls a signal handler.
-   This frame's "caller" is actually the interrupted state, not a call;
-   its return address is an exact PC, not a PC after a call instruction.  */
-extern int dwarf_frame_info (Dwarf_Frame *frame,
-			     Dwarf_Addr *start, Dwarf_Addr *end, bool *signalp);
-
-/* Return a DWARF expression that yields the Canonical Frame Address at
-   this frame state.  Returns -1 for errors, or zero for success, with
-   *NOPS set to the number of operations stored at *OPS.  That pointer
-   can be used only as long as FRAME is alive and unchanged.  *NOPS is
-   zero if the CFA cannot be determined here.  Note that if nonempty,
-   *OPS is a DWARF expression, not a location description--append
-   DW_OP_stack_value to a get a location description for the CFA.  */
-extern int dwarf_frame_cfa (Dwarf_Frame *frame, Dwarf_Op **ops, size_t *nops)
-  __nonnull_attribute__ (2);
-
-/* Deliver a DWARF location description that yields the location or
-   value of DWARF register number REGNO in the state described by FRAME.
-
-   Returns -1 for errors or zero for success, setting *NOPS to the
-   number of operations in the array stored at *OPS.  Note the last
-   operation is DW_OP_stack_value if there is no mutable location but
-   only a computable value.
-
-   *NOPS zero with *OPS set to OPS_MEM means CFI says the caller's
-   REGNO is "undefined", i.e. it's call-clobbered and cannot be recovered.
-
-   *NOPS zero with *OPS set to a null pointer means CFI says the
-   caller's REGNO is "same_value", i.e. this frame did not change it;
-   ask the caller frame where to find it.
-
-   For common simple expressions *OPS is OPS_MEM (which is a caller
-   owned array for at least 3 Dwarf_Ops).  For arbitrary DWARF
-   expressions in the CFI, *OPS is an internal pointer that can be
-   used as long as the Dwarf_CFI used to create FRAME remains
-   alive.  */
-extern int dwarf_frame_register (Dwarf_Frame *frame, int regno,
-				 Dwarf_Op ops_mem[3],
-				 Dwarf_Op **ops, size_t *nops)
-  __nonnull_attribute__ (3, 4, 5);
-
-
-/* Return error code of last failing function call.  This value is kept
-   separately for each thread.  */
-extern int dwarf_errno (void);
-
-/* Return error string for ERROR.  If ERROR is zero, return error string
-   for most recent error or NULL is none occurred.  If ERROR is -1 the
-   behaviour is similar to the last case except that not NULL but a legal
-   string is returned.  */
-extern const char *dwarf_errmsg (int err);
-
-
-/* Register new Out-Of-Memory handler.  The old handler is returned.  */
-extern Dwarf_OOM dwarf_new_oom_handler (Dwarf *dbg, Dwarf_OOM handler);
-
-
-/* Inline optimizations.  */
-#ifdef __OPTIMIZE__
-/* Return attribute code of given attribute.  */
-__libdw_extern_inline unsigned int
-dwarf_whatattr (Dwarf_Attribute *attr)
-{
-  return attr == NULL ? 0 : attr->code;
-}
-
-/* Return attribute code of given attribute.  */
-__libdw_extern_inline unsigned int
-dwarf_whatform (Dwarf_Attribute *attr)
-{
-  return attr == NULL ? 0 : attr->form;
-}
-#endif	/* Optimize.  */
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif	/* libdw.h */
diff --git a/projects/llm_framework/include/elfutils/libdwelf.h b/projects/llm_framework/include/elfutils/libdwelf.h
deleted file mode 100644
index 263ca60e..00000000
--- a/projects/llm_framework/include/elfutils/libdwelf.h
+++ /dev/null
@@ -1,147 +0,0 @@
-/* Interfaces for libdwelf. DWARF ELF Low-level Functions.
-   Copyright (C) 2014, 2015, 2016, 2018 Red Hat, Inc.
-   This file is part of elfutils.
-
-   This file is free software; you can redistribute it and/or modify
-   it under the terms of either
-
-     * the GNU Lesser General Public License as published by the Free
-       Software Foundation; either version 3 of the License, or (at
-       your option) any later version
-
-   or
-
-     * the GNU General Public License as published by the Free
-       Software Foundation; either version 2 of the License, or (at
-       your option) any later version
-
-   or both in parallel, as here.
-
-   elfutils is distributed in the hope that it will be useful, but
-   WITHOUT ANY WARRANTY; without even the implied warranty of
-   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
-   General Public License for more details.
-
-   You should have received copies of the GNU General Public License and
-   the GNU Lesser General Public License along with this program.  If
-   not, see <http://www.gnu.org/licenses/>.  */
-
-#ifndef _LIBDWELF_H
-#define _LIBDWELF_H	1
-
-#include "libdw.h"
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* DWARF ELF Low-level Functions (dwelf).
-   Functions starting with dwelf_elf will take a (libelf) Elf object as
-   first argument and might set elf_errno on error.  Functions starting
-   with dwelf_dwarf will take a (libdw) Dwarf object as first argument
-   and might set dwarf_errno on error.  */
-
-/* Returns the name and the CRC32 of the separate debug file from the
-   .gnu_debuglink section if found in the ELF.  Return NULL if the ELF
-   file didn't have a .gnu_debuglink section, had malformed data in the
-   section or some other error occurred.  */
-extern const char *dwelf_elf_gnu_debuglink (Elf *elf, GElf_Word *crc);
-
-/* Returns the name and build ID from the .gnu_debugaltlink section if
-   found in the ELF.  On success, pointers to the name and build ID
-   are written to *NAMEP and *BUILDID_P, and the positive length of
-   the build ID is returned.  Returns 0 if the ELF lacks a
-   .gnu_debugaltlink section.  Returns -1 in case of malformed data or
-   other errors.  */
-extern ssize_t dwelf_dwarf_gnu_debugaltlink (Dwarf *dwarf,
-					     const char **namep,
-					     const void **build_idp);
-
-/* Returns the build ID as found in a NT_GNU_BUILD_ID note from either
-   a SHT_NOTE section or from a PT_NOTE segment if the ELF file
-   doesn't contain any section headers.  On success a pointer to the
-   build ID is written to *BUILDID_P, and the positive length of the
-   build ID is returned.  Returns 0 if the ELF lacks a NT_GNU_BUILD_ID
-   note.  Returns -1 in case of malformed data or other errors.  */
-extern ssize_t dwelf_elf_gnu_build_id (Elf *elf, const void **build_idp);
-
-/* Returns the size of the uncompressed data of a GNU compressed
-   section.  The section name should start with .zdebug (but this
-   isn't checked by this function).  If the section isn't compressed
-   (the section data doesn't start with ZLIB) -1 is returned. If an
-   error occurred -1 is returned and elf_errno is set.  */
-extern ssize_t dwelf_scn_gnu_compressed_size (Elf_Scn *scn);
-
-/* ELF/DWARF string table handling.  */
-typedef struct Dwelf_Strtab Dwelf_Strtab;
-typedef struct Dwelf_Strent Dwelf_Strent;
-
-/* Create a new ELF/DWARF string table object in memory.  ELF string
-   tables have a required zero length null string at offset zero.
-   DWARF string tables don't require such a null entry (unless they
-   are shared with an ELF string table).  If NULLSTR is true then a
-   null entry is always created (even if the string table is empty
-   otherwise).  */
-extern Dwelf_Strtab *dwelf_strtab_init (bool nullstr);
-
-/* Add string STR to string table ST.  Returns NULL if no memory could
-   be allocated.  The given STR is owned by the called and must be
-   valid till dwelf_strtab_free is called.  dwelf_strtab_finalize
-   might copy the string into the final table and dwelf_strent_str
-   might return it, or a reference to an identical copy/substring
-   added to the string table.  */
-extern Dwelf_Strent *dwelf_strtab_add (Dwelf_Strtab *st, const char *str)
-  __nonnull_attribute__ (1, 2);
-
-/* This is an optimized version of dwelf_strtab_add if the length of
-   the string is already known.  LEN is the length of STR including
-   zero terminator.  Calling dwelf_strtab_add (st, str) is similar to
-   calling dwelf_strtab_len (st, str, strlen (str) + 1).  */
-extern Dwelf_Strent *dwelf_strtab_add_len (Dwelf_Strtab *st,
-					   const char *str, size_t len)
-  __nonnull_attribute__ (1, 2);
-
-/* Finalize string table ST and store size and memory location
-   information in DATA d_size and d_buf.  DATA d_type will be set to
-   ELF_T_BYTE, d_off will be zero, d_align will be 1 and d_version
-   will be set to EV_CURRENT.  If no memory could be allocated NULL is
-   returned and DATA->d_buf will be set to NULL.  Otherwise DATA will
-   be returned.  */
-extern Elf_Data *dwelf_strtab_finalize (Dwelf_Strtab *st,
-					Elf_Data *data)
-  __nonnull_attribute__ (1, 2);
-
-/* Get offset in string table for string associated with entry.  Only
-   valid after dwelf_strtab_finalize has been called.  */
-extern size_t dwelf_strent_off (Dwelf_Strent *se)
-  __nonnull_attribute__ (1);
-
-/* Return the string associated with the entry.  */
-extern const char *dwelf_strent_str (Dwelf_Strent *se)
-  __nonnull_attribute__ (1);
-
-/* Free resources allocated for the string table.  This invalidates
-   any Dwelf_Strent references returned earlier. */
-extern void dwelf_strtab_free (Dwelf_Strtab *st)
-  __nonnull_attribute__ (1);
-
-/* Creates a read-only Elf handle from the given file handle.  The
-   file may be compressed and/or contain a linux kernel image header,
-   in which case it is eagerly decompressed in full and the Elf handle
-   is created as if created with elf_memory ().  On decompression or
-   file errors NULL is returned (and elf_errno will be set).  If there
-   was no error, but the file is not an ELF file, then an ELF_K_NONE
-   Elf handle is returned (just like with elf_begin).  The Elf handle
-   should be closed with elf_end ().  The file handle will not be
-   closed.  */
-extern Elf *dwelf_elf_begin (int fd);
-
-/* Returns a human readable string for the given ELF header e_machine
-   value, or NULL if the given number isn't currently known.  */
-extern const char *dwelf_elf_e_machine_string (int machine);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif	/* libdwelf.h */
diff --git a/projects/llm_framework/include/elfutils/libdwfl.h b/projects/llm_framework/include/elfutils/libdwfl.h
deleted file mode 100644
index f98f1d52..00000000
--- a/projects/llm_framework/include/elfutils/libdwfl.h
+++ /dev/null
@@ -1,802 +0,0 @@
-/* Interfaces for libdwfl.
-   Copyright (C) 2005-2010, 2013 Red Hat, Inc.
-   This file is part of elfutils.
-
-   This file is free software; you can redistribute it and/or modify
-   it under the terms of either
-
-     * the GNU Lesser General Public License as published by the Free
-       Software Foundation; either version 3 of the License, or (at
-       your option) any later version
-
-   or
-
-     * the GNU General Public License as published by the Free
-       Software Foundation; either version 2 of the License, or (at
-       your option) any later version
-
-   or both in parallel, as here.
-
-   elfutils is distributed in the hope that it will be useful, but
-   WITHOUT ANY WARRANTY; without even the implied warranty of
-   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
-   General Public License for more details.
-
-   You should have received copies of the GNU General Public License and
-   the GNU Lesser General Public License along with this program.  If
-   not, see <http://www.gnu.org/licenses/>.  */
-
-#ifndef _LIBDWFL_H
-#define _LIBDWFL_H	1
-
-#include "libdw.h"
-#include <stdio.h>
-
-/* Handle for a session using the library.  */
-typedef struct Dwfl Dwfl;
-
-/* Handle for a module.  */
-typedef struct Dwfl_Module Dwfl_Module;
-
-/* Handle describing a line record.  */
-typedef struct Dwfl_Line Dwfl_Line;
-
-/* This holds information common for all the frames of one backtrace for
-   a particular thread/task/TID.  Several threads belong to one Dwfl.  */
-typedef struct Dwfl_Thread Dwfl_Thread;
-
-/* This holds everything we know about the state of the frame at a particular
-   PC location described by an FDE belonging to Dwfl_Thread.  */
-typedef struct Dwfl_Frame Dwfl_Frame;
-
-/* Callbacks.  */
-typedef struct
-{
-  int (*find_elf) (Dwfl_Module *mod, void **userdata,
-		   const char *modname, Dwarf_Addr base,
-		   char **file_name, Elf **elfp);
-
-  int (*find_debuginfo) (Dwfl_Module *mod, void **userdata,
-			 const char *modname, Dwarf_Addr base,
-			 const char *file_name,
-			 const char *debuglink_file, GElf_Word debuglink_crc,
-			 char **debuginfo_file_name);
-
-  /* Fill *ADDR with the loaded address of the section called SECNAME in
-     the given module.  Use (Dwarf_Addr) -1 if this section is omitted from
-     accessible memory.  This is called exactly once for each SHF_ALLOC
-     section that relocations affecting DWARF data refer to, so it can
-     easily be used to collect state about the sections referenced.  */
-  int (*section_address) (Dwfl_Module *mod, void **userdata,
-			  const char *modname, Dwarf_Addr base,
-			  const char *secname,
-			  GElf_Word shndx, const GElf_Shdr *shdr,
-			  Dwarf_Addr *addr);
-
-  char **debuginfo_path;	/* See dwfl_standard_find_debuginfo.  */
-} Dwfl_Callbacks;
-
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* Start a new session with the library.  */
-extern Dwfl *dwfl_begin (const Dwfl_Callbacks *callbacks)
-  __nonnull_attribute__ (1);
-
-
-/* End a session.  */
-extern void dwfl_end (Dwfl *);
-
-/* Return implementation's version string suitable for printing.  */
-extern const char *dwfl_version (Dwfl *);
-
-/* Return error code of last failing function call.  This value is kept
-   separately for each thread.  */
-extern int dwfl_errno (void);
-
-/* Return error string for ERROR.  If ERROR is zero, return error string
-   for most recent error or NULL if none occurred.  If ERROR is -1 the
-   behaviour is similar to the last case except that not NULL but a legal
-   string is returned.  */
-extern const char *dwfl_errmsg (int err);
-
-
-/* Start reporting the current set of segments and modules to the library.
-   All existing segments are wiped.  Existing modules are marked to be
-   deleted, and will not be found via dwfl_addrmodule et al if they are not
-   re-reported before dwfl_report_end is called.  */
-extern void dwfl_report_begin (Dwfl *dwfl);
-
-/* Report that segment NDX begins at PHDR->p_vaddr + BIAS.
-   If NDX is < 0, the value succeeding the last call's NDX
-   is used instead (zero on the first call).  IDENT is ignored.
-
-   If nonzero, the smallest PHDR->p_align value seen sets the
-   effective page size for the address space DWFL describes.
-   This is the granularity at which reported module boundary
-   addresses will be considered to fall in or out of a segment.
-
-   Returns -1 for errors, or NDX (or its assigned replacement) on success.
-
-   Reporting segments at all is optional.  Its only benefit to the caller is to
-   offer this quick lookup via dwfl_addrsegment, or use other segment-based
-   calls.  */
-extern int dwfl_report_segment (Dwfl *dwfl, int ndx,
-				const GElf_Phdr *phdr, GElf_Addr bias,
-				const void *ident);
-
-/* Report that a module called NAME spans addresses [START, END).
-   Returns the module handle, either existing or newly allocated,
-   or returns a null pointer for an allocation error.  */
-extern Dwfl_Module *dwfl_report_module (Dwfl *dwfl, const char *name,
-					Dwarf_Addr start, Dwarf_Addr end);
-
-/* Report a module to address BASE with start and end addresses computed
-   from the ELF program headers in the given file - see the table below.
-   FD may be -1 to open FILE_NAME.  On success, FD is consumed by the
-   library, and the `find_elf' callback will not be used for this module.
-	    ADD_P_VADDR  BASE
-   ET_EXEC  ignored      ignored
-   ET_DYN   false        absolute address where to place the file
-	    true         start address relative to ELF's phdr p_vaddr
-   ET_REL   ignored      absolute address where to place the file
-   ET_CORE  ignored      ignored
-   ET_DYN ELF phdr p_vaddr address can be non-zero if the shared library
-   has been prelinked by tool prelink(8).  */
-extern Dwfl_Module *dwfl_report_elf (Dwfl *dwfl, const char *name,
-				     const char *file_name, int fd,
-				     GElf_Addr base, bool add_p_vaddr);
-
-/* Similar, but report the module for offline use.  All ET_EXEC files
-   being reported must be reported before any relocatable objects.
-   If this is used, dwfl_report_module and dwfl_report_elf may not be
-   used in the same reporting session.  */
-extern Dwfl_Module *dwfl_report_offline (Dwfl *dwfl, const char *name,
-					 const char *file_name, int fd);
-
-
-/* Finish reporting the current set of modules to the library.
-   If REMOVED is not null, it's called for each module that
-   existed before but was not included in the current report.
-   Returns a nonzero return value from the callback.
-   The callback may call dwfl_report_module; doing so with the
-   details of the module being removed prevents its removal.
-   DWFL cannot be used until this function has returned zero.  */
-extern int dwfl_report_end (Dwfl *dwfl,
-			    int (*removed) (Dwfl_Module *, void *,
-					    const char *, Dwarf_Addr,
-					    void *arg),
-			    void *arg);
-
-/* Start reporting additional modules to the library.  No calls but
-   dwfl_report_* can be made on DWFL until dwfl_report_end is called.
-   This is like dwfl_report_begin, but all the old modules are kept on.
-   More dwfl_report_* calls can follow to add more modules.
-   When dwfl_report_end is called, no old modules will be removed.  */
-extern void dwfl_report_begin_add (Dwfl *dwfl);
-
-
-/* Return the name of the module, and for each non-null argument store
-   interesting details: *USERDATA is a location for storing your own
-   pointer, **USERDATA is initially null; *START and *END give the address
-   range covered by the module; *DWBIAS is the address bias for debugging
-   information, and *SYMBIAS for symbol table entries (either is -1 if not
-   yet accessed); *MAINFILE is the name of the ELF file, and *DEBUGFILE the
-   name of the debuginfo file (might be equal to *MAINFILE; either is null
-   if not yet accessed).  */
-extern const char *dwfl_module_info (Dwfl_Module *mod, void ***userdata,
-				     Dwarf_Addr *start, Dwarf_Addr *end,
-				     Dwarf_Addr *dwbias, Dwarf_Addr *symbias,
-				     const char **mainfile,
-				     const char **debugfile);
-
-/* Iterate through the modules, starting the walk with OFFSET == 0.
-   Calls *CALLBACK for each module as long as it returns DWARF_CB_OK.
-   When *CALLBACK returns another value, the walk stops and the
-   return value can be passed as OFFSET to resume it.  Returns 0 when
-   there are no more modules, or -1 for errors.  */
-extern ptrdiff_t dwfl_getmodules (Dwfl *dwfl,
-				  int (*callback) (Dwfl_Module *, void **,
-						   const char *, Dwarf_Addr,
-						   void *arg),
-				  void *arg,
-				  ptrdiff_t offset);
-
-/* Find the module containing the given address.  */
-extern Dwfl_Module *dwfl_addrmodule (Dwfl *dwfl, Dwarf_Addr address);
-
-/* Find the segment, if any, and module, if any, containing ADDRESS.
-   Returns a segment index returned by dwfl_report_segment, or -1
-   if no segment matches the address.  Regardless of the return value,
-   *MOD is always set to the module containing ADDRESS, or to null.  */
-extern int dwfl_addrsegment (Dwfl *dwfl, Dwarf_Addr address, Dwfl_Module **mod);
-
-
-
-/* Report the known build ID bits associated with a module.
-   If VADDR is nonzero, it gives the absolute address where those
-   bits are found within the module.  This can be called at any
-   time, but is usually used immediately after dwfl_report_module.
-   Once the module's main ELF file is opened, the ID note found
-   there takes precedence and cannot be changed.  */
-extern int dwfl_module_report_build_id (Dwfl_Module *mod,
-					const unsigned char *bits, size_t len,
-					GElf_Addr vaddr)
-  __nonnull_attribute__ (2);
-
-/* Extract the build ID bits associated with a module.
-   Returns -1 for errors, 0 if no ID is known, or the number of ID bytes.
-   When an ID is found, *BITS points to it; *VADDR is the absolute address
-   at which the ID bits are found within the module, or 0 if unknown.
-
-   This returns 0 when the module's main ELF file has not yet been loaded
-   and its build ID bits were not reported.  To ensure the ID is always
-   returned when determinable, call dwfl_module_getelf first.  */
-extern int dwfl_module_build_id (Dwfl_Module *mod,
-				 const unsigned char **bits, GElf_Addr *vaddr)
-  __nonnull_attribute__ (2, 3);
-
-
-/*** Standard callbacks ***/
-
-/* These standard find_elf and find_debuginfo callbacks are
-   controlled by a string specifying directories to look in.
-   If `debuginfo_path' is set in the Dwfl_Callbacks structure
-   and the char * it points to is not null, that supplies the
-   string.  Otherwise a default path is used.
-
-   If the first character of the string is + or - that enables or
-   disables CRC32 checksum validation when it's necessary.  The
-   remainder of the string is composed of elements separated by
-   colons.  Each element can start with + or - to override the
-   global checksum behavior.  This flag is never relevant when
-   working with build IDs, but it's always parsed in the path
-   string.  The remainder of the element indicates a directory.
-
-   Searches by build ID consult only the elements naming absolute
-   directory paths.  They look under those directories for a link
-   named ".build-id/xx/yy" or ".build-id/xx/yy.debug", where "xxyy"
-   is the lower-case hexadecimal representation of the ID bytes.
-
-   In searches for debuginfo by name, if the remainder of the
-   element is empty, the directory containing the main file is
-   tried; if it's an absolute path name, the absolute directory path
-   (and any subdirectory of that path) containing the main file is
-   taken as a subdirectory of this path; a relative path name is taken
-   as a subdirectory of the directory containing the main file.
-   Hence for /usr/bin/ls, the default string ":.debug:/usr/lib/debug"
-   says to look in /usr/bin, then /usr/bin/.debug, then the path subdirs
-   under /usr/lib/debug, in the order /usr/lib/debug/usr/bin, then
-   /usr/lib/debug/bin, and finally /usr/lib/debug, for the file name in
-   the .gnu_debuglink section (or "ls.debug" if none was found).  */
-
-/* Standard find_elf callback function working solely on build ID.
-   This can be tried first by any find_elf callback, to use the
-   bits passed to dwfl_module_report_build_id, if any.  */
-extern int dwfl_build_id_find_elf (Dwfl_Module *, void **,
-				   const char *, Dwarf_Addr,
-				   char **, Elf **);
-
-/* Standard find_debuginfo callback function working solely on build ID.
-   This can be tried first by any find_debuginfo callback,
-   to use the build ID bits from the main file when present.  */
-extern int dwfl_build_id_find_debuginfo (Dwfl_Module *, void **,
-					 const char *, Dwarf_Addr,
-					 const char *, const char *,
-					 GElf_Word, char **);
-
-/* Standard find_debuginfo callback function.
-   If a build ID is available, this tries first to use that.
-   If there is no build ID or no valid debuginfo found by ID,
-   it searches the debuginfo path by name, as described above.
-   Any file found in the path is validated by build ID if possible,
-   or else by CRC32 checksum if enabled, and skipped if it does not match.  */
-extern int dwfl_standard_find_debuginfo (Dwfl_Module *, void **,
-					 const char *, Dwarf_Addr,
-					 const char *, const char *,
-					 GElf_Word, char **);
-
-
-/* This callback must be used when using dwfl_offline_* to report modules,
-   if ET_REL is to be supported.  */
-extern int dwfl_offline_section_address (Dwfl_Module *, void **,
-					 const char *, Dwarf_Addr,
-					 const char *, GElf_Word,
-					 const GElf_Shdr *,
-					 Dwarf_Addr *addr);
-
-
-/* Callbacks for working with kernel modules in the running Linux kernel.  */
-extern int dwfl_linux_kernel_find_elf (Dwfl_Module *, void **,
-				       const char *, Dwarf_Addr,
-				       char **, Elf **);
-extern int dwfl_linux_kernel_module_section_address (Dwfl_Module *, void **,
-						     const char *, Dwarf_Addr,
-						     const char *, GElf_Word,
-						     const GElf_Shdr *,
-						     Dwarf_Addr *addr);
-
-/* Call dwfl_report_elf for the running Linux kernel.
-   Returns zero on success, -1 if dwfl_report_module failed,
-   or an errno code if opening the kernel binary failed.  */
-extern int dwfl_linux_kernel_report_kernel (Dwfl *dwfl);
-
-/* Call dwfl_report_module for each kernel module in the running Linux kernel.
-   Returns zero on success, -1 if dwfl_report_module failed,
-   or an errno code if reading the list of modules failed.  */
-extern int dwfl_linux_kernel_report_modules (Dwfl *dwfl);
-
-/* Report a kernel and its modules found on disk, for offline use.
-   If RELEASE starts with '/', it names a directory to look in;
-   if not, it names a directory to find under /lib/modules/;
-   if null, /lib/modules/`uname -r` is used.
-   Returns zero on success, -1 if dwfl_report_module failed,
-   or an errno code if finding the files on disk failed.
-
-   If PREDICATE is not null, it is called with each module to be reported;
-   its arguments are the module name, and the ELF file name or null if unknown,
-   and its return value should be zero to skip the module, one to report it,
-   or -1 to cause the call to fail and return errno.  */
-extern int dwfl_linux_kernel_report_offline (Dwfl *dwfl, const char *release,
-					     int (*predicate) (const char *,
-							       const char *));
-
-/* Examine an ET_CORE file and report modules based on its contents.
-   This can follow a dwfl_report_offline call to bootstrap the
-   DT_DEBUG method of following the dynamic linker link_map chain, in
-   case the core file does not contain enough of the executable's text
-   segment to locate its PT_DYNAMIC in the dump.  In such case you need to
-   supply non-NULL EXECUTABLE, otherwise dynamic libraries will not be loaded
-   into the DWFL map.  This might call dwfl_report_elf on file names found in
-   the dump if reading some link_map files is the only way to ascertain those
-   modules' addresses.  Returns the number of modules reported, or -1 for
-   errors.  */
-extern int dwfl_core_file_report (Dwfl *dwfl, Elf *elf, const char *executable);
-
-/* Call dwfl_report_module for each file mapped into the address space of PID.
-   Returns zero on success, -1 if dwfl_report_module failed,
-   or an errno code if opening the proc files failed.  */
-extern int dwfl_linux_proc_report (Dwfl *dwfl, pid_t pid);
-
-/* Similar, but reads an input stream in the format of Linux /proc/PID/maps
-   files giving module layout, not the file for a live process.  */
-extern int dwfl_linux_proc_maps_report (Dwfl *dwfl, FILE *);
-
-/* Trivial find_elf callback for use with dwfl_linux_proc_report.
-   This uses the module name as a file name directly and tries to open it
-   if it begin with a slash, or handles the magic string "[vdso]".  */
-extern int dwfl_linux_proc_find_elf (Dwfl_Module *mod, void **userdata,
-				     const char *module_name, Dwarf_Addr base,
-				     char **file_name, Elf **);
-
-/* Standard argument parsing for using a standard callback set.  */
-struct argp;
-extern const struct argp *dwfl_standard_argp (void) __const_attribute__;
-
-
-/*** Relocation of addresses from Dwfl ***/
-
-/* Return the number of relocatable bases associated with the module,
-   which is zero for ET_EXEC and one for ET_DYN.  Returns -1 for errors.  */
-extern int dwfl_module_relocations (Dwfl_Module *mod);
-
-/* Return the relocation base index associated with the *ADDRESS location,
-   and adjust *ADDRESS to be an offset relative to that base.
-   Returns -1 for errors.  */
-extern int dwfl_module_relocate_address (Dwfl_Module *mod,
-					 Dwarf_Addr *address);
-
-/* Return the ELF section name for the given relocation base index;
-   if SHNDXP is not null, set *SHNDXP to the ELF section index.
-   For ET_DYN, returns "" and sets *SHNDXP to SHN_ABS; the relocation
-   base is the runtime start address reported for the module.
-   Returns null for errors.  */
-extern const char *dwfl_module_relocation_info (Dwfl_Module *mod,
-						unsigned int idx,
-						GElf_Word *shndxp);
-
-/* Validate that ADDRESS and ADDRESS+OFFSET lie in a known module
-   and both within the same contiguous region for relocation purposes.
-   Returns zero for success and -1 for errors.  */
-extern int dwfl_validate_address (Dwfl *dwfl,
-				  Dwarf_Addr address, Dwarf_Sword offset);
-
-
-/*** ELF access functions ***/
-
-/* Fetch the module main ELF file (where the allocated sections
-   are found) for use with libelf.  If successful, fills in *BIAS
-   with the difference between addresses within the loaded module
-   and those in symbol tables or Dwarf information referring to it.  */
-extern Elf *dwfl_module_getelf (Dwfl_Module *, GElf_Addr *bias)
-  __nonnull_attribute__ (2);
-
-/* Return the number of symbols in the module's symbol table,
-   or -1 for errors.  */
-extern int dwfl_module_getsymtab (Dwfl_Module *mod);
-
-/* Return the index of the first global symbol in the module's symbol
-   table, or -1 for errors.  In each symbol table, all symbols with
-   STB_LOCAL binding precede the weak and global symbols.  This
-   function returns the symbol table index one greater than the last
-   local symbol.  */
-extern int dwfl_module_getsymtab_first_global (Dwfl_Module *mod);
-
-/* Fetch one entry from the module's symbol table.  On errors, returns
-   NULL.  If successful, fills in *SYM and returns the string for st_name.
-   This works like gelf_getsym except that st_value is always adjusted to
-   an absolute value based on the module's location, when the symbol is in
-   an SHF_ALLOC section.  If SHNDXP is non-null, it's set with the section
-   index (whether from st_shndx or extended index table); in case of a
-   symbol in a non-allocated section, *SHNDXP is instead set to -1.
-   Note that since symbols can come from either the main, debug or auxiliary
-   ELF symbol file (either dynsym or symtab) the section index can only
-   be reliably used to compare against special section constants like
-   SHN_UNDEF or SHN_ABS.  It is recommended to use dwfl_module_getsym_info
-   which doesn't have these deficiencies.  */
-extern const char *dwfl_module_getsym (Dwfl_Module *mod, int ndx,
-				       GElf_Sym *sym, GElf_Word *shndxp)
-  __nonnull_attribute__ (3);
-
-/* Fetch one entry from the module's symbol table and the associated
-   address value.  On errors, returns NULL.  If successful, fills in
-   *SYM, *ADDR and returns the string for st_name.  This works like
-   gelf_getsym.  *ADDR is set to the st_value adjusted to an absolute
-   value based on the module's location, when the symbol is in an
-   SHF_ALLOC section.  For non-ET_REL files, if the arch uses function
-   descriptors, and the st_value points to one, *ADDR will be resolved
-   to the actual function entry address.  The SYM->ST_VALUE itself
-   isn't adjusted in any way.  Fills in ELFP, if not NULL, with the
-   ELF file the symbol originally came from.  Note that symbols can
-   come from either the main, debug or auxiliary ELF symbol file
-   (either dynsym or symtab).  If SHNDXP is non-null, it's set with
-   the section index (whether from st_shndx or extended index table);
-   in case of a symbol in a non-allocated section, *SHNDXP is instead
-   set to -1.  Fills in BIAS, if not NULL, with the difference between
-   addresses within the loaded module and those in symbol table of the
-   ELF file.  Note that the address associated with the symbol might
-   be in a different section than the returned symbol.  The section in
-   the main elf file in which returned ADDR falls can be found with
-   dwfl_module_address_section.  */
-extern const char *dwfl_module_getsym_info (Dwfl_Module *mod, int ndx,
-					    GElf_Sym *sym, GElf_Addr *addr,
-					    GElf_Word *shndxp,
-					    Elf **elfp, Dwarf_Addr *bias)
-  __nonnull_attribute__ (3, 4);
-
-/* Find the symbol that ADDRESS lies inside, and return its name.  */
-extern const char *dwfl_module_addrname (Dwfl_Module *mod, GElf_Addr address);
-
-/* Find the symbol associated with ADDRESS.  Return its name or NULL
-   when nothing was found.  If the architecture uses function
-   descriptors, and symbol st_value points to one, ADDRESS will be
-   matched against either the adjusted st_value or the associated
-   function entry value as described in dwfl_module_getsym_info.  If
-   OFFSET is not NULL it will be filled in with the difference from
-   the start of the symbol (or function entry).  If SYM is not NULL it
-   is filled in with the symbol associated with the matched ADDRESS.
-   The SYM->ST_VALUE itself isn't adjusted in any way.  Fills in ELFP,
-   if not NULL, with the ELF file the symbol originally came from.
-   Note that symbols can come from either the main, debug or auxiliary
-   ELF symbol file (either dynsym or symtab).  If SHNDXP is non-null,
-   it's set with the section index (whether from st_shndx or extended
-   index table).  Fills in BIAS, if not NULL, with the difference
-   between addresses within the loaded module and those in symbol
-   table of the ELF file.  Note that the address matched against the
-   symbol might be in a different section than the returned symbol.
-   The section in the main elf file in ADDRESS falls can be found with
-   dwfl_module_address_section.  */
-extern const char *dwfl_module_addrinfo (Dwfl_Module *mod, GElf_Addr address,
-					 GElf_Off *offset, GElf_Sym *sym,
-					 GElf_Word *shndxp, Elf **elfp,
-					 Dwarf_Addr *bias)
-  __nonnull_attribute__ (3);
-
-/* Find the symbol that ADDRESS lies inside, and return detailed
-   information as for dwfl_module_getsym (above).  Note that like
-   dwfl_module_getsym this function also adjusts SYM->ST_VALUE to an
-   absolute value based on the module's location.  ADDRESS is only
-   matched against this adjusted SYM->ST_VALUE.  This means that
-   depending on architecture this might only match symbols that
-   represent function descriptor addresses (and not function entry
-   addresses).  For these reasons it is recommended to use
-   dwfl_module_addrinfo instead.  */
-extern const char *dwfl_module_addrsym (Dwfl_Module *mod, GElf_Addr address,
-					GElf_Sym *sym, GElf_Word *shndxp)
-  __nonnull_attribute__ (3);
-
-/* Find the ELF section that *ADDRESS lies inside and return it.
-   On success, adjusts *ADDRESS to be relative to the section,
-   and sets *BIAS to the difference between addresses used in
-   the returned section's headers and run-time addresses.  */
-extern Elf_Scn *dwfl_module_address_section (Dwfl_Module *mod,
-					     Dwarf_Addr *address,
-					     Dwarf_Addr *bias)
-  __nonnull_attribute__ (2, 3);
-
-
-/*** Dwarf access functions ***/
-
-/* Fetch the module's debug information for use with libdw.
-   If successful, fills in *BIAS with the difference between
-   addresses within the loaded module and those  to use with libdw.  */
-extern Dwarf *dwfl_module_getdwarf (Dwfl_Module *, Dwarf_Addr *bias)
-     __nonnull_attribute__ (2);
-
-/* Get the libdw handle for each module.  */
-extern ptrdiff_t dwfl_getdwarf (Dwfl *,
-				int (*callback) (Dwfl_Module *, void **,
-						 const char *, Dwarf_Addr,
-						 Dwarf *, Dwarf_Addr, void *),
-				void *arg, ptrdiff_t offset);
-
-/* Look up the module containing ADDR and return its debugging information,
-   loading it if necessary.  */
-extern Dwarf *dwfl_addrdwarf (Dwfl *dwfl, Dwarf_Addr addr, Dwarf_Addr *bias)
-     __nonnull_attribute__ (3);
-
-
-/* Find the CU containing ADDR and return its DIE.  */
-extern Dwarf_Die *dwfl_addrdie (Dwfl *dwfl, Dwarf_Addr addr, Dwarf_Addr *bias)
-     __nonnull_attribute__ (3);
-extern Dwarf_Die *dwfl_module_addrdie (Dwfl_Module *mod,
-				       Dwarf_Addr addr, Dwarf_Addr *bias)
-     __nonnull_attribute__ (3);
-
-/* Iterate through the CUs, start with null for LASTCU.  */
-extern Dwarf_Die *dwfl_nextcu (Dwfl *dwfl, Dwarf_Die *lastcu, Dwarf_Addr *bias)
-     __nonnull_attribute__ (3);
-extern Dwarf_Die *dwfl_module_nextcu (Dwfl_Module *mod,
-				      Dwarf_Die *lastcu, Dwarf_Addr *bias)
-     __nonnull_attribute__ (3);
-
-/* Return the module containing the CU DIE.  */
-extern Dwfl_Module *dwfl_cumodule (Dwarf_Die *cudie);
-
-
-/* Cache the source line information for the CU and return the
-   number of Dwfl_Line entries it has.  */
-extern int dwfl_getsrclines (Dwarf_Die *cudie, size_t *nlines);
-
-/* Access one line number entry within the CU.  */
-extern Dwfl_Line *dwfl_onesrcline (Dwarf_Die *cudie, size_t idx);
-
-/* Get source for address.  */
-extern Dwfl_Line *dwfl_module_getsrc (Dwfl_Module *mod, Dwarf_Addr addr);
-extern Dwfl_Line *dwfl_getsrc (Dwfl *dwfl, Dwarf_Addr addr);
-
-/* Get address for source.  */
-extern int dwfl_module_getsrc_file (Dwfl_Module *mod,
-				    const char *fname, int lineno, int column,
-				    Dwfl_Line ***srcsp, size_t *nsrcs);
-
-/* Return the module containing this line record.  */
-extern Dwfl_Module *dwfl_linemodule (Dwfl_Line *line);
-
-/* Return the CU containing this line record.  */
-extern Dwarf_Die *dwfl_linecu (Dwfl_Line *line);
-
-/* Return the source file name and fill in other information.
-   Arguments may be null for unneeded fields.  */
-extern const char *dwfl_lineinfo (Dwfl_Line *line, Dwarf_Addr *addr,
-				  int *linep, int *colp,
-				  Dwarf_Word *mtime, Dwarf_Word *length);
-
-  /* Return the equivalent Dwarf_Line and the bias to apply to its address.  */
-extern Dwarf_Line *dwfl_dwarf_line (Dwfl_Line *line, Dwarf_Addr *bias);
-
-/* Return the compilation directory (AT_comp_dir) from this line's CU.  */
-extern const char *dwfl_line_comp_dir (Dwfl_Line *line);
-
-
-/*** Machine backend access functions ***/
-
-/* Return location expression to find return value given a
-   DW_TAG_subprogram, DW_TAG_subroutine_type, or similar DIE describing
-   function itself (whose DW_AT_type attribute describes its return type).
-   The given DIE must come from the given module.  Returns -1 for errors.
-   Returns zero if the function has no return value (e.g. "void" in C).
-   Otherwise, *LOCOPS gets a location expression to find the return value,
-   and returns the number of operations in the expression.  The pointer is
-   permanently allocated at least as long as the module is live.  */
-extern int dwfl_module_return_value_location (Dwfl_Module *mod,
-					      Dwarf_Die *functypedie,
-					      const Dwarf_Op **locops);
-
-/* Enumerate the DWARF register numbers and their names.
-   For each register, CALLBACK gets its DWARF number, a string describing
-   the register set (such as "integer" or "FPU"), a prefix used in
-   assembler syntax (such as "%" or "$", may be ""), and the name for the
-   register (contains identifier characters only, possibly all digits).
-   The REGNAME string is valid only during the callback. */
-extern int dwfl_module_register_names (Dwfl_Module *mod,
-				       int (*callback) (void *arg,
-							int regno,
-							const char *setname,
-							const char *prefix,
-							const char *regname,
-							int bits, int type),
-				       void *arg);
-
-
-/* Find the CFI for this module.  Returns NULL if there is no CFI.
-   On success, fills in *BIAS with the difference between addresses
-   within the loaded module and those in the CFI referring to it.
-   The pointer returned can be used until the module is cleaned up.
-   Calling these more than once returns the same pointers.
-
-   dwfl_module_dwarf_cfi gets the '.debug_frame' information found with the
-   rest of the DWARF information.  dwfl_module_eh_cfi gets the '.eh_frame'
-   information found linked into the text.  A module might have either or
-   both.  */
-extern Dwarf_CFI *dwfl_module_dwarf_cfi (Dwfl_Module *mod, Dwarf_Addr *bias);
-extern Dwarf_CFI *dwfl_module_eh_cfi (Dwfl_Module *mod, Dwarf_Addr *bias);
-
-
-typedef struct
-{
-  /* Called to iterate through threads.  Returns next TID (thread ID) on
-     success, a negative number on failure and zero if there are no more
-     threads.  dwfl_errno () should be set if negative number has been
-     returned.  *THREAD_ARGP is NULL on first call, and may be optionally
-     set by the implementation. The value set by the implementation will
-     be passed in on the next call to NEXT_THREAD.  THREAD_ARGP is never
-     NULL.  *THREAD_ARGP will be passed to set_initial_registers or
-     thread_detach callbacks together with Dwfl_Thread *thread.  This
-     method must not be NULL.  */
-  pid_t (*next_thread) (Dwfl *dwfl, void *dwfl_arg, void **thread_argp)
-    __nonnull_attribute__ (1);
-
-  /* Called to get a specific thread.  Returns true if there is a
-     thread with the given thread id number, returns false if no such
-     thread exists and will set dwfl_errno in that case.  THREAD_ARGP
-     is never NULL.  *THREAD_ARGP will be passed to
-     set_initial_registers or thread_detach callbacks together with
-     Dwfl_Thread *thread.  This method may be NULL and will then be
-     emulated using the next_thread callback. */
-  bool (*get_thread) (Dwfl *dwfl, pid_t tid, void *dwfl_arg,
-		      void **thread_argp)
-    __nonnull_attribute__ (1);
-
-  /* Called during unwinding to access memory (stack) state.  Returns true for
-     successfully read *RESULT or false and sets dwfl_errno () on failure.
-     This method may be NULL - in such case dwfl_thread_getframes will return
-     only the initial frame.  */
-  bool (*memory_read) (Dwfl *dwfl, Dwarf_Addr addr, Dwarf_Word *result,
-                       void *dwfl_arg)
-    __nonnull_attribute__ (1, 3);
-
-  /* Called on initial unwind to get the initial register state of the first
-     frame.  Should call dwfl_thread_state_registers, possibly multiple times
-     for different ranges and possibly also dwfl_thread_state_register_pc, to
-     fill in initial (DWARF) register values.  After this call, till at least
-     thread_detach is called, the thread is assumed to be frozen, so that it is
-     safe to unwind.  Returns true on success or false and sets dwfl_errno ()
-     on failure.  In the case of a failure thread_detach will not be called.
-     This method must not be NULL.  */
-  bool (*set_initial_registers) (Dwfl_Thread *thread, void *thread_arg)
-    __nonnull_attribute__ (1);
-
-  /* Called by dwfl_end.  All thread_detach method calls have been already
-     done.  This method may be NULL.  */
-  void (*detach) (Dwfl *dwfl, void *dwfl_arg)
-    __nonnull_attribute__ (1);
-
-  /* Called when unwinding is done.  No callback will be called after
-     this method has been called.  Iff set_initial_registers was called for
-     a TID and it returned success thread_detach will be called before the
-     detach method above.  This method may be NULL.  */
-  void (*thread_detach) (Dwfl_Thread *thread, void *thread_arg)
-    __nonnull_attribute__ (1);
-} Dwfl_Thread_Callbacks;
-
-/* PID is the process id associated with the DWFL state.  Architecture of DWFL
-   modules is specified by ELF, ELF must remain valid during DWFL lifetime.
-   Use NULL ELF to detect architecture from DWFL, the function will then detect
-   it from arbitrary Dwfl_Module of DWFL.  DWFL_ARG is the callback backend
-   state.  DWFL_ARG will be provided to the callbacks.  *THREAD_CALLBACKS
-   function pointers must remain valid during lifetime of DWFL.  Function
-   returns true on success, false otherwise.  */
-bool dwfl_attach_state (Dwfl *dwfl, Elf *elf, pid_t pid,
-                        const Dwfl_Thread_Callbacks *thread_callbacks,
-			void *dwfl_arg)
-  __nonnull_attribute__ (1, 4);
-
-/* Calls dwfl_attach_state with Dwfl_Thread_Callbacks setup for extracting
-   thread state from the ELF core file.  Returns the pid number extracted
-   from the core file, or -1 for errors.  */
-extern int dwfl_core_file_attach (Dwfl *dwfl, Elf *elf);
-
-/* Calls dwfl_attach_state with Dwfl_Thread_Callbacks setup for extracting
-   thread state from the proc file system.  Uses ptrace to attach and stop
-   the thread under inspection and detaches when thread_detach is called
-   and unwinding for the thread is done, unless ASSUME_PTRACE_STOPPED is
-   true.  If ASSUME_PTRACE_STOPPED is true the caller should make sure that
-   the thread is ptrace attached and stopped before unwinding by calling
-   either dwfl_thread_getframes or dwfl_getthread_frames.  Returns zero on
-   success, -1 if dwfl_attach_state failed, or an errno code if opening the
-   proc files failed.  */
-extern int dwfl_linux_proc_attach (Dwfl *dwfl, pid_t pid,
-				   bool assume_ptrace_stopped);
-
-/* Return PID for the process associated with DWFL.  Function returns -1 if
-   dwfl_attach_state was not called for DWFL.  */
-pid_t dwfl_pid (Dwfl *dwfl)
-  __nonnull_attribute__ (1);
-
-/* Return DWFL from which THREAD was created using dwfl_getthreads.  */
-Dwfl *dwfl_thread_dwfl (Dwfl_Thread *thread)
-  __nonnull_attribute__ (1);
-
-/* Return positive TID (thread ID) for THREAD.  This function never fails.  */
-pid_t dwfl_thread_tid (Dwfl_Thread *thread)
-  __nonnull_attribute__ (1);
-
-/* Return thread for frame STATE.  This function never fails.  */
-Dwfl_Thread *dwfl_frame_thread (Dwfl_Frame *state)
-  __nonnull_attribute__ (1);
-
-/* Called by Dwfl_Thread_Callbacks.set_initial_registers implementation.
-   For every known continuous block of registers <FIRSTREG..FIRSTREG+NREGS)
-   (inclusive..exclusive) set their content to REGS (array of NREGS items).
-   Function returns false if any of the registers has invalid number.  */
-bool dwfl_thread_state_registers (Dwfl_Thread *thread, int firstreg,
-                                  unsigned nregs, const Dwarf_Word *regs)
-  __nonnull_attribute__ (1, 4);
-
-/* Called by Dwfl_Thread_Callbacks.set_initial_registers implementation.
-   If PC is not contained among DWARF registers passed by
-   dwfl_thread_state_registers on the target architecture pass the PC value
-   here.  */
-void dwfl_thread_state_register_pc (Dwfl_Thread *thread, Dwarf_Word pc)
-  __nonnull_attribute__ (1);
-
-/* Iterate through the threads for a process.  Returns zero if all threads have
-   been processed by the callback, returns -1 on error, or the value of the
-   callback when not DWARF_CB_OK.  -1 returned on error will set dwfl_errno ().
-   Keeps calling the callback with the next thread while the callback returns
-   DWARF_CB_OK, till there are no more threads.  */
-int dwfl_getthreads (Dwfl *dwfl,
-		     int (*callback) (Dwfl_Thread *thread, void *arg),
-		     void *arg)
-  __nonnull_attribute__ (1, 2);
-
-/* Iterate through the frames for a thread.  Returns zero if all frames
-   have been processed by the callback, returns -1 on error, or the value of
-   the callback when not DWARF_CB_OK.  -1 returned on error will
-   set dwfl_errno ().  Some systems return error instead of zero on end of the
-   backtrace, for cross-platform compatibility callers should consider error as
-   a zero.  Keeps calling the callback with the next frame while the callback
-   returns DWARF_CB_OK, till there are no more frames.  On start will call the
-   set_initial_registers callback and on return will call the detach_thread
-   callback of the Dwfl_Thread.  */
-int dwfl_thread_getframes (Dwfl_Thread *thread,
-			   int (*callback) (Dwfl_Frame *state, void *arg),
-			   void *arg)
-  __nonnull_attribute__ (1, 2);
-
-/* Like dwfl_thread_getframes, but specifying the thread by its unique
-   identifier number.  Returns zero if all frames have been processed
-   by the callback, returns -1 on error (and when no thread with
-   the given thread id number exists), or the value of the callback
-   when not DWARF_CB_OK.  -1 returned on error will set dwfl_errno ().  */
-int dwfl_getthread_frames (Dwfl *dwfl, pid_t tid,
-			   int (*callback) (Dwfl_Frame *thread, void *arg),
-			   void *arg)
-  __nonnull_attribute__ (1, 3);
-
-/* Return *PC (program counter) for thread-specific frame STATE.
-   Set *ISACTIVATION according to DWARF frame "activation" definition.
-   Typically you need to subtract 1 from *PC if *ACTIVATION is false to safely
-   find function of the caller.  ACTIVATION may be NULL.  PC must not be NULL.
-   Function returns false if it failed to find *PC.  */
-bool dwfl_frame_pc (Dwfl_Frame *state, Dwarf_Addr *pc, bool *isactivation)
-  __nonnull_attribute__ (1, 2);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif	/* libdwfl.h */
diff --git a/projects/llm_framework/include/elfutils/version.h b/projects/llm_framework/include/elfutils/version.h
deleted file mode 100644
index 8690a24b..00000000
--- a/projects/llm_framework/include/elfutils/version.h
+++ /dev/null
@@ -1,38 +0,0 @@
-/* Version information about elfutils development libraries.
-   Copyright (C) 2008 Red Hat, Inc.
-
-   This file is part of elfutils.
-
-   This file is free software; you can redistribute it and/or modify
-   it under the terms of either
-
-     * the GNU Lesser General Public License as published by the Free
-       Software Foundation; either version 3 of the License, or (at
-       your option) any later version
-
-   or
-
-     * the GNU General Public License as published by the Free
-       Software Foundation; either version 2 of the License, or (at
-       your option) any later version
-
-   or both in parallel, as here.
-
-   elfutils is distributed in the hope that it will be useful, but
-   WITHOUT ANY WARRANTY; without even the implied warranty of
-   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
-   General Public License for more details.
-
-   You should have received copies of the GNU General Public License and
-   the GNU Lesser General Public License along with this program.  If
-   not, see <http://www.gnu.org/licenses/>.  */
-
-#ifndef _ELFUTILS_VERSION_H
-#define _ELFUTILS_VERSION_H	1
-
-#define _ELFUTILS_VERSION	183
-
-#define _ELFUTILS_PREREQ(major, minor) \
-  (_ELFUTILS_VERSION >= ((major) * 1000 + (minor)))
-
-#endif	/* elfutils/version.h */
diff --git a/projects/llm_framework/include/fst/accumulator.h b/projects/llm_framework/include/fst/accumulator.h
deleted file mode 100644
index 5ae19247..00000000
--- a/projects/llm_framework/include/fst/accumulator.h
+++ /dev/null
@@ -1,903 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes to accumulate arc weights. Useful for weight lookahead.
-
-#ifndef FST_ACCUMULATOR_H_
-#define FST_ACCUMULATOR_H_
-
-#include <algorithm>
-#include <functional>
-#include <unordered_map>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arcfilter.h>
-#include <fst/arcsort.h>
-#include <fst/dfs-visit.h>
-#include <fst/expanded-fst.h>
-#include <fst/replace.h>
-
-namespace fst {
-
-// This class accumulates arc weights using the semiring Plus().
-template <class A>
-class DefaultAccumulator {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  DefaultAccumulator() {}
-
-  DefaultAccumulator(const DefaultAccumulator &acc, bool safe = false) {}
-
-  void Init(const Fst<Arc> &fst, bool copy = false) {}
-
-  void SetState(StateId state) {}
-
-  Weight Sum(Weight w, Weight v) { return Plus(w, v); }
-
-  template <class ArcIter>
-  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) {
-    Adder<Weight> adder(w);  // maintains cumulative sum accurately
-    aiter->Seek(begin);
-    for (auto pos = begin; pos < end; aiter->Next(), ++pos)
-      adder.Add(aiter->Value().weight);
-    return adder.Sum();
-  }
-
-  constexpr bool Error() const { return false; }
-
- private:
-  DefaultAccumulator &operator=(const DefaultAccumulator &) = delete;
-};
-
-// This class accumulates arc weights using the log semiring Plus() assuming an
-// arc weight has a WeightConvert specialization to and from log64 weights.
-template <class A>
-class LogAccumulator {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  LogAccumulator() {}
-
-  LogAccumulator(const LogAccumulator &acc, bool safe = false) {}
-
-  void Init(const Fst<Arc> &fst, bool copy = false) {}
-
-  void SetState(StateId s) {}
-
-  Weight Sum(Weight w, Weight v) { return LogPlus(w, v); }
-
-  template <class ArcIter>
-  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) {
-    auto sum = w;
-    aiter->Seek(begin);
-    for (auto pos = begin; pos < end; aiter->Next(), ++pos) {
-      sum = LogPlus(sum, aiter->Value().weight);
-    }
-    return sum;
-  }
-
-  constexpr bool Error() const { return false; }
-
- private:
-  Weight LogPlus(Weight w, Weight v) {
-    if (w == Weight::Zero()) {
-      return v;
-    }
-    const auto f1 = to_log_weight_(w).Value();
-    const auto f2 = to_log_weight_(v).Value();
-    if (f1 > f2) {
-      return to_weight_(Log64Weight(f2 - internal::LogPosExp(f1 - f2)));
-    } else {
-      return to_weight_(Log64Weight(f1 - internal::LogPosExp(f2 - f1)));
-    }
-  }
-
-  const WeightConvert<Weight, Log64Weight> to_log_weight_{};
-  const WeightConvert<Log64Weight, Weight> to_weight_{};
-
-  LogAccumulator &operator=(const LogAccumulator &) = delete;
-};
-
-// Interface for shareable data for fast log accumulator copies. Holds pointers
-// to data only, storage is provided by derived classes.
-class FastLogAccumulatorData {
- public:
-  FastLogAccumulatorData(int arc_limit, int arc_period)
-      : arc_limit_(arc_limit),
-        arc_period_(arc_period),
-        weights_ptr_(nullptr),
-        num_weights_(0),
-        weight_positions_ptr_(nullptr),
-        num_positions_(0) {}
-
-  virtual ~FastLogAccumulatorData() {}
-
-  // Cummulative weight per state for all states s.t. # of arcs > arc_limit_
-  // with arcs in order. The first element per state is Log64Weight::Zero().
-  const double *Weights() const { return weights_ptr_; }
-
-  int NumWeights() const { return num_weights_; }
-
-  // Maps from state to corresponding beginning weight position in weights_.
-  // osition -1 means no pre-computed weights for that state.
-  const int *WeightPositions() const { return weight_positions_ptr_; }
-
-  int NumPositions() const { return num_positions_; }
-
-  int ArcLimit() const { return arc_limit_; }
-
-  int ArcPeriod() const { return arc_period_; }
-
-  // Returns true if the data object is mutable and supports SetData().
-  virtual bool IsMutable() const = 0;
-
-  // Does not take ownership but may invalidate the contents of weights and
-  // weight_positions.
-  virtual void SetData(std::vector<double> *weights,
-                       std::vector<int> *weight_positions) = 0;
-
- protected:
-  void Init(int num_weights, const double *weights, int num_positions,
-            const int *weight_positions) {
-    weights_ptr_ = weights;
-    num_weights_ = num_weights;
-    weight_positions_ptr_ = weight_positions;
-    num_positions_ = num_positions;
-  }
-
- private:
-  const int arc_limit_;
-  const int arc_period_;
-  const double *weights_ptr_;
-  int num_weights_;
-  const int *weight_positions_ptr_;
-  int num_positions_;
-
-  FastLogAccumulatorData(const FastLogAccumulatorData &) = delete;
-  FastLogAccumulatorData &operator=(const FastLogAccumulatorData &) = delete;
-};
-
-// FastLogAccumulatorData with mutable storage; filled by
-// FastLogAccumulator::Init.
-class MutableFastLogAccumulatorData : public FastLogAccumulatorData {
- public:
-  MutableFastLogAccumulatorData(int arc_limit, int arc_period)
-      : FastLogAccumulatorData(arc_limit, arc_period) {}
-
-  bool IsMutable() const override { return true; }
-
-  void SetData(std::vector<double> *weights,
-               std::vector<int> *weight_positions) override {
-    weights_.swap(*weights);
-    weight_positions_.swap(*weight_positions);
-    Init(weights_.size(), weights_.data(), weight_positions_.size(),
-         weight_positions_.data());
-  }
-
- private:
-  std::vector<double> weights_;
-  std::vector<int> weight_positions_;
-
-  MutableFastLogAccumulatorData(const MutableFastLogAccumulatorData &) = delete;
-  MutableFastLogAccumulatorData &operator=(
-      const MutableFastLogAccumulatorData &) = delete;
-};
-
-// This class accumulates arc weights using the log semiring Plus() assuming an
-// arc weight has a WeightConvert specialization to and from log64 weights. The
-// member function Init(fst) has to be called to setup pre-computed weight
-// information.
-template <class A>
-class FastLogAccumulator {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit FastLogAccumulator(ssize_t arc_limit = 20, ssize_t arc_period = 10)
-      : to_log_weight_(),
-        to_weight_(),
-        arc_limit_(arc_limit),
-        arc_period_(arc_period),
-        data_(std::make_shared<MutableFastLogAccumulatorData>(arc_limit,
-                                                              arc_period)),
-        state_weights_(nullptr),
-        error_(false) {}
-
-  explicit FastLogAccumulator(std::shared_ptr<FastLogAccumulatorData> data)
-      : to_log_weight_(),
-        to_weight_(),
-        arc_limit_(data->ArcLimit()),
-        arc_period_(data->ArcPeriod()),
-        data_(data),
-        state_weights_(nullptr),
-        error_(false) {}
-
-  FastLogAccumulator(const FastLogAccumulator<Arc> &acc, bool safe = false)
-      : to_log_weight_(),
-        to_weight_(),
-        arc_limit_(acc.arc_limit_),
-        arc_period_(acc.arc_period_),
-        data_(acc.data_),
-        state_weights_(nullptr),
-        error_(acc.error_) {}
-
-  void SetState(StateId s) {
-    const auto *weights = data_->Weights();
-    const auto *weight_positions = data_->WeightPositions();
-    state_weights_ = nullptr;
-    if (s < data_->NumPositions()) {
-      const auto pos = weight_positions[s];
-      if (pos >= 0) state_weights_ = &(weights[pos]);
-    }
-  }
-
-  Weight Sum(Weight w, Weight v) const { return LogPlus(w, v); }
-
-  template <class ArcIter>
-  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) const {
-    if (error_) return Weight::NoWeight();
-    auto sum = w;
-    // Finds begin and end of pre-stored weights.
-    ssize_t index_begin = -1;
-    ssize_t index_end = -1;
-    ssize_t stored_begin = end;
-    ssize_t stored_end = end;
-    if (state_weights_) {
-      index_begin = begin > 0 ? (begin - 1) / arc_period_ + 1 : 0;
-      index_end = end / arc_period_;
-      stored_begin = index_begin * arc_period_;
-      stored_end = index_end * arc_period_;
-    }
-    // Computes sum before pre-stored weights.
-    if (begin < stored_begin) {
-      const auto pos_end = std::min(stored_begin, end);
-      aiter->Seek(begin);
-      for (auto pos = begin; pos < pos_end; aiter->Next(), ++pos) {
-        sum = LogPlus(sum, aiter->Value().weight);
-      }
-    }
-    // Computes sum between pre-stored weights.
-    if (stored_begin < stored_end) {
-      const auto f1 = state_weights_[index_end];
-      const auto f2 = state_weights_[index_begin];
-      if (f1 < f2) sum = LogPlus(sum, LogMinus(f1, f2));
-      // Commented out for efficiency; adds Zero().
-      /*
-      else {
-        // explicitly computes if cumulative sum lacks precision
-        aiter->Seek(stored_begin);
-        for (auto pos = stored_begin; pos < stored_end; aiter->Next(), ++pos)
-          sum = LogPlus(sum, aiter->Value().weight);
-      }
-      */
-    }
-    // Computes sum after pre-stored weights.
-    if (stored_end < end) {
-      const auto pos_start = std::max(stored_begin, stored_end);
-      aiter->Seek(pos_start);
-      for (auto pos = pos_start; pos < end; aiter->Next(), ++pos) {
-        sum = LogPlus(sum, aiter->Value().weight);
-      }
-    }
-    return sum;
-  }
-
-  template <class FST>
-  void Init(const FST &fst, bool copy = false) {
-    if (copy || !data_->IsMutable()) return;
-    if (data_->NumPositions() != 0 || arc_limit_ < arc_period_) {
-      FSTERROR() << "FastLogAccumulator: Initialization error";
-      error_ = true;
-      return;
-    }
-    std::vector<double> weights;
-    std::vector<int> weight_positions;
-    weight_positions.reserve(CountStates(fst));
-    for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      if (fst.NumArcs(s) >= arc_limit_) {
-        auto sum = FloatLimits<double>::PosInfinity();
-        if (weight_positions.size() <= s) weight_positions.resize(s + 1, -1);
-        weight_positions[s] = weights.size();
-        weights.push_back(sum);
-        size_t narcs = 0;
-        ArcIterator<FST> aiter(fst, s);
-        aiter.SetFlags(kArcWeightValue | kArcNoCache, kArcFlags);
-        for (; !aiter.Done(); aiter.Next()) {
-          const auto &arc = aiter.Value();
-          sum = LogPlus(sum, arc.weight);
-          // Stores cumulative weight distribution per arc_period_.
-          if (++narcs % arc_period_ == 0) weights.push_back(sum);
-        }
-      }
-    }
-    data_->SetData(&weights, &weight_positions);
-  }
-
-  bool Error() const { return error_; }
-
-  std::shared_ptr<FastLogAccumulatorData> GetData() const { return data_; }
-
- private:
-  static double LogPosExp(double x) {
-    return x == FloatLimits<double>::PosInfinity() ? 0.0
-                                                   : log(1.0F + exp(-x));
-  }
-
-  static double LogMinusExp(double x) {
-    return x == FloatLimits<double>::PosInfinity() ? 0.0
-                                                   : log(1.0F - exp(-x));
-  }
-
-  Weight LogPlus(Weight w, Weight v) const {
-    if (w == Weight::Zero()) {
-      return v;
-    }
-    const auto f1 = to_log_weight_(w).Value();
-    const auto f2 = to_log_weight_(v).Value();
-    if (f1 > f2) {
-      return to_weight_(Log64Weight(f2 - LogPosExp(f1 - f2)));
-    } else {
-      return to_weight_(Log64Weight(f1 - LogPosExp(f2 - f1)));
-    }
-  }
-
-  double LogPlus(double f1, Weight v) const {
-    const auto f2 = to_log_weight_(v).Value();
-    if (f1 == FloatLimits<double>::PosInfinity()) {
-      return f2;
-    } else if (f1 > f2) {
-      return f2 - LogPosExp(f1 - f2);
-    } else {
-      return f1 - LogPosExp(f2 - f1);
-    }
-  }
-
-  // Assumes f1 < f2.
-  Weight LogMinus(double f1, double f2) const {
-    if (f2 == FloatLimits<double>::PosInfinity()) {
-      return to_weight_(Log64Weight(f1));
-    } else {
-      return to_weight_(Log64Weight(f1 - LogMinusExp(f2 - f1)));
-    }
-  }
-
-  const WeightConvert<Weight, Log64Weight> to_log_weight_{};
-  const WeightConvert<Log64Weight, Weight> to_weight_{};
-  const ssize_t arc_limit_;   // Minimum number of arcs to pre-compute state.
-  const ssize_t arc_period_;  // Saves cumulative weights per arc_period_.
-  std::shared_ptr<FastLogAccumulatorData> data_;
-  const double *state_weights_;
-  bool error_;
-
-  FastLogAccumulator &operator=(const FastLogAccumulator &) = delete;
-};
-
-// Stores shareable data for cache log accumulator copies. All copies share the
-// same cache.
-template <class Arc>
-class CacheLogAccumulatorData {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  CacheLogAccumulatorData(bool gc, size_t gc_limit)
-      : cache_gc_(gc), cache_limit_(gc_limit), cache_size_(0) {}
-
-  CacheLogAccumulatorData(const CacheLogAccumulatorData<Arc> &data)
-      : cache_gc_(data.cache_gc_),
-        cache_limit_(data.cache_limit_),
-        cache_size_(0) {}
-
-  bool CacheDisabled() const { return cache_gc_ && cache_limit_ == 0; }
-
-  std::vector<double> *GetWeights(StateId s) {
-    auto it = cache_.find(s);
-    if (it != cache_.end()) {
-      it->second.recent = true;
-      return it->second.weights.get();
-    } else {
-      return nullptr;
-    }
-  }
-
-  void AddWeights(StateId s, std::vector<double> *weights) {
-    if (cache_gc_ && cache_size_ >= cache_limit_) GC(false);
-    cache_.insert(std::make_pair(s, CacheState(weights, true)));
-    if (cache_gc_) cache_size_ += weights->capacity() * sizeof(double);
-  }
-
- private:
-  // Cached information for a given state.
-  struct CacheState {
-    std::unique_ptr<std::vector<double>> weights;  // Accumulated weights.
-    bool recent;  // Has this state been accessed since last GC?
-
-    CacheState(std::vector<double> *weights, bool recent)
-        : weights(weights), recent(recent) {}
-  };
-
-  // Garbage collect: Deletes from cache states that have not been accessed
-  // since the last GC ('free_recent = false') until 'cache_size_' is 2/3 of
-  // 'cache_limit_'. If it does not free enough memory, start deleting
-  // recently accessed states.
-  void GC(bool free_recent) {
-    auto cache_target = (2 * cache_limit_) / 3 + 1;
-    auto it = cache_.begin();
-    while (it != cache_.end() && cache_size_ > cache_target) {
-      auto &cs = it->second;
-      if (free_recent || !cs.recent) {
-        cache_size_ -= cs.weights->capacity() * sizeof(double);
-        cache_.erase(it++);
-      } else {
-        cs.recent = false;
-        ++it;
-      }
-    }
-    if (!free_recent && cache_size_ > cache_target) GC(true);
-  }
-
-  std::unordered_map<StateId, CacheState> cache_;  // Cache.
-  bool cache_gc_;       // Enables garbage collection.
-  size_t cache_limit_;  // # of bytes cached.
-  size_t cache_size_;   // # of bytes allowed before GC.
-
-  CacheLogAccumulatorData &operator=(const CacheLogAccumulatorData &) = delete;
-};
-
-// This class accumulates arc weights using the log semiring Plus() has a
-// WeightConvert specialization to and from log64 weights. It is similar to the
-// FastLogAccumator. However here, the accumulated weights are pre-computed and
-// stored only for the states that are visited. The member function Init(fst)
-// has to be called to setup this accumulator.
-template <class Arc>
-class CacheLogAccumulator {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit CacheLogAccumulator(ssize_t arc_limit = 10, bool gc = false,
-                               size_t gc_limit = 10 * 1024 * 1024)
-      : arc_limit_(arc_limit),
-        data_(std::make_shared<CacheLogAccumulatorData<Arc>>(gc, gc_limit)),
-        s_(kNoStateId),
-        error_(false) {}
-
-  CacheLogAccumulator(const CacheLogAccumulator<Arc> &acc, bool safe = false)
-      : arc_limit_(acc.arc_limit_),
-        fst_(acc.fst_ ? acc.fst_->Copy() : nullptr),
-        data_(safe ? std::make_shared<CacheLogAccumulatorData<Arc>>(*acc.data_)
-                   : acc.data_),
-        s_(kNoStateId),
-        error_(acc.error_) {}
-
-  // Argument arc_limit specifies the minimum number of arcs to pre-compute.
-  void Init(const Fst<Arc> &fst, bool copy = false) {
-    if (!copy && fst_) {
-      FSTERROR() << "CacheLogAccumulator: Initialization error";
-      error_ = true;
-      return;
-    }
-    fst_.reset(fst.Copy());
-  }
-
-  void SetState(StateId s, int depth = 0) {
-    if (s == s_) return;
-    s_ = s;
-    if (data_->CacheDisabled() || error_) {
-      weights_ = nullptr;
-      return;
-    }
-    if (!fst_) {
-      FSTERROR() << "CacheLogAccumulator::SetState: Incorrectly initialized";
-      error_ = true;
-      weights_ = nullptr;
-      return;
-    }
-    weights_ = data_->GetWeights(s);
-    if ((weights_ == nullptr) && (fst_->NumArcs(s) >= arc_limit_)) {
-      weights_ = new std::vector<double>;
-      weights_->reserve(fst_->NumArcs(s) + 1);
-      weights_->push_back(FloatLimits<double>::PosInfinity());
-      data_->AddWeights(s, weights_);
-    }
-  }
-
-  Weight Sum(Weight w, Weight v) { return LogPlus(w, v); }
-
-  template <class ArcIter>
-  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) {
-    if (weights_ == nullptr) {
-      auto sum = w;
-      aiter->Seek(begin);
-      for (auto pos = begin; pos < end; aiter->Next(), ++pos) {
-        sum = LogPlus(sum, aiter->Value().weight);
-      }
-      return sum;
-    } else {
-      Extend(end, aiter);
-      const auto &f1 = (*weights_)[end];
-      const auto &f2 = (*weights_)[begin];
-      if (f1 < f2) {
-        return LogPlus(w, LogMinus(f1, f2));
-      } else {
-        // Commented out for efficiency; adds Zero().
-        /*
-        auto sum = w;
-        // Explicitly computes if cumulative sum lacks precision.
-        aiter->Seek(begin);
-        for (auto pos = begin; pos < end; aiter->Next(), ++pos) {
-          sum = LogPlus(sum, aiter->Value().weight);
-        }
-        return sum;
-        */
-        return w;
-      }
-    }
-  }
-
-  // Returns first position from aiter->Position() whose accumulated
-  // value is greater or equal to w (w.r.t. Zero() < One()). The
-  // iterator may be repositioned.
-  template <class ArcIter>
-  size_t LowerBound(Weight w, ArcIter *aiter) {
-    const auto f = to_log_weight_(w).Value();
-    auto pos = aiter->Position();
-    if (weights_) {
-      Extend(fst_->NumArcs(s_), aiter);
-      return std::lower_bound(weights_->begin() + pos + 1, weights_->end(),
-                              f, std::greater<double>()) -
-          weights_->begin() - 1;
-    } else {
-      size_t n = 0;
-      auto x = FloatLimits<double>::PosInfinity();
-      for (aiter->Reset(); !aiter->Done(); aiter->Next(), ++n) {
-        x = LogPlus(x, aiter->Value().weight);
-        if (n >= pos && x <= f) break;
-      }
-      return n;
-    }
-  }
-
-  bool Error() const { return error_; }
-
- private:
-  double LogPosExp(double x) {
-    return x == FloatLimits<double>::PosInfinity() ? 0.0
-                                                   : log(1.0F + exp(-x));
-  }
-
-  double LogMinusExp(double x) {
-    return x == FloatLimits<double>::PosInfinity() ? 0.0
-                                                   : log(1.0F - exp(-x));
-  }
-
-  Weight LogPlus(Weight w, Weight v) {
-    if (w == Weight::Zero()) {
-      return v;
-    }
-    const auto f1 = to_log_weight_(w).Value();
-    const auto f2 = to_log_weight_(v).Value();
-    if (f1 > f2) {
-      return to_weight_(Log64Weight(f2 - LogPosExp(f1 - f2)));
-    } else {
-      return to_weight_(Log64Weight(f1 - LogPosExp(f2 - f1)));
-    }
-  }
-
-  double LogPlus(double f1, Weight v) {
-    const auto f2 = to_log_weight_(v).Value();
-    if (f1 == FloatLimits<double>::PosInfinity()) {
-      return f2;
-    } else if (f1 > f2) {
-      return f2 - LogPosExp(f1 - f2);
-    } else {
-      return f1 - LogPosExp(f2 - f1);
-    }
-  }
-
-  // Assumes f1 < f2.
-  Weight LogMinus(double f1, double f2) {
-    if (f2 == FloatLimits<double>::PosInfinity()) {
-      return to_weight_(Log64Weight(f1));
-    } else {
-      return to_weight_(Log64Weight(f1 - LogMinusExp(f2 - f1)));
-    }
-  }
-
-  // Extends weights up to index 'end'.
-  template <class ArcIter>
-  void Extend(ssize_t end, ArcIter *aiter) {
-    if (weights_->size() <= end) {
-      for (aiter->Seek(weights_->size() - 1); weights_->size() <= end;
-           aiter->Next()) {
-        weights_->push_back(LogPlus(weights_->back(), aiter->Value().weight));
-      }
-    }
-  }
-
-
-  const WeightConvert<Weight, Log64Weight> to_log_weight_{};
-  const WeightConvert<Log64Weight, Weight> to_weight_{};
-  ssize_t arc_limit_;                    // Minimum # of arcs to cache a state.
-  std::vector<double> *weights_;         // Accumulated weights for cur. state.
-  std::unique_ptr<const Fst<Arc>> fst_;  // Input FST.
-  std::shared_ptr<CacheLogAccumulatorData<Arc>> data_;  // Cache data.
-  StateId s_;                                           // Current state.
-  bool error_;
-};
-
-// Stores shareable data for replace accumulator copies.
-template <class Accumulator, class T>
-class ReplaceAccumulatorData {
- public:
-  using Arc = typename Accumulator::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using StateTable = T;
-  using StateTuple = typename StateTable::StateTuple;
-
-  ReplaceAccumulatorData() : state_table_(nullptr) {}
-
-  explicit ReplaceAccumulatorData(
-      const std::vector<Accumulator *> &accumulators)
-      : state_table_(nullptr) {
-    accumulators_.reserve(accumulators.size());
-    for (const auto accumulator : accumulators) {
-      accumulators_.emplace_back(accumulator);
-    }
-  }
-
-  void Init(const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_tuples,
-            const StateTable *state_table) {
-    state_table_ = state_table;
-    accumulators_.resize(fst_tuples.size());
-    for (Label i = 0; i < accumulators_.size(); ++i) {
-      if (!accumulators_[i]) {
-        accumulators_[i].reset(new Accumulator());
-        accumulators_[i]->Init(*(fst_tuples[i].second));
-      }
-      fst_array_.emplace_back(fst_tuples[i].second->Copy());
-    }
-  }
-
-  const StateTuple &GetTuple(StateId s) const { return state_table_->Tuple(s); }
-
-  Accumulator *GetAccumulator(size_t i) { return accumulators_[i].get(); }
-
-  const Fst<Arc> *GetFst(size_t i) const { return fst_array_[i].get(); }
-
- private:
-  const StateTable *state_table_;
-  std::vector<std::unique_ptr<Accumulator>> accumulators_;
-  std::vector<std::unique_ptr<const Fst<Arc>>> fst_array_;
-};
-
-// This class accumulates weights in a ReplaceFst.  The 'Init' method takes as
-// input the argument used to build the ReplaceFst and the ReplaceFst state
-// table. It uses accumulators of type 'Accumulator' in the underlying FSTs.
-template <class Accumulator,
-          class T = DefaultReplaceStateTable<typename Accumulator::Arc>>
-class ReplaceAccumulator {
- public:
-  using Arc = typename Accumulator::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using StateTable = T;
-  using StateTuple = typename StateTable::StateTuple;
-  using Weight = typename Arc::Weight;
-
-  ReplaceAccumulator()
-      : init_(false),
-        data_(std::make_shared<
-              ReplaceAccumulatorData<Accumulator, StateTable>>()),
-        error_(false) {}
-
-  explicit ReplaceAccumulator(const std::vector<Accumulator *> &accumulators)
-      : init_(false),
-        data_(std::make_shared<ReplaceAccumulatorData<Accumulator, StateTable>>(
-            accumulators)),
-        error_(false) {}
-
-  ReplaceAccumulator(const ReplaceAccumulator<Accumulator, StateTable> &acc,
-                     bool safe = false)
-      : init_(acc.init_), data_(acc.data_), error_(acc.error_) {
-    if (!init_) {
-      FSTERROR() << "ReplaceAccumulator: Can't copy unintialized accumulator";
-    }
-    if (safe) FSTERROR() << "ReplaceAccumulator: Safe copy not supported";
-  }
-
-  // Does not take ownership of the state table, the state table is owned by
-  // the ReplaceFst.
-  void Init(const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_tuples,
-            const StateTable *state_table) {
-    init_ = true;
-    data_->Init(fst_tuples, state_table);
-  }
-
-  // Method required by LookAheadMatcher. However, ReplaceAccumulator needs to
-  // be initialized by calling the Init method above before being passed to
-  // LookAheadMatcher.
-  //
-  // TODO(allauzen): Revisit this. Consider creating a method
-  // Init(const ReplaceFst<A, T, C>&, bool) and using friendship to get access
-  // to the innards of ReplaceFst.
-  void Init(const Fst<Arc> &fst, bool copy = false) {
-    if (!init_) {
-      FSTERROR() << "ReplaceAccumulator::Init: Accumulator needs to be"
-                 << " initialized before being passed to LookAheadMatcher";
-      error_ = true;
-    }
-  }
-
-  void SetState(StateId s) {
-    if (!init_) {
-      FSTERROR() << "ReplaceAccumulator::SetState: Incorrectly initialized";
-      error_ = true;
-      return;
-    }
-    auto tuple = data_->GetTuple(s);
-    fst_id_ = tuple.fst_id - 1;  // Replace FST ID is 1-based.
-    data_->GetAccumulator(fst_id_)->SetState(tuple.fst_state);
-    if ((tuple.prefix_id != 0) &&
-        (data_->GetFst(fst_id_)->Final(tuple.fst_state) != Weight::Zero())) {
-      offset_ = 1;
-      offset_weight_ = data_->GetFst(fst_id_)->Final(tuple.fst_state);
-    } else {
-      offset_ = 0;
-      offset_weight_ = Weight::Zero();
-    }
-    aiter_.reset(
-        new ArcIterator<Fst<Arc>>(*data_->GetFst(fst_id_), tuple.fst_state));
-  }
-
-  Weight Sum(Weight w, Weight v) {
-    if (error_) return Weight::NoWeight();
-    return data_->GetAccumulator(fst_id_)->Sum(w, v);
-  }
-
-  template <class ArcIter>
-  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) {
-    if (error_) return Weight::NoWeight();
-    auto sum = begin == end ? Weight::Zero()
-                            : data_->GetAccumulator(fst_id_)->Sum(
-                                  w, aiter_.get(), begin ? begin - offset_ : 0,
-                                  end - offset_);
-    if (begin == 0 && end != 0 && offset_ > 0) sum = Sum(offset_weight_, sum);
-    return sum;
-  }
-
-  bool Error() const { return error_; }
-
- private:
-  bool init_;
-  std::shared_ptr<ReplaceAccumulatorData<Accumulator, StateTable>> data_;
-  Label fst_id_;
-  size_t offset_;
-  Weight offset_weight_;
-  std::unique_ptr<ArcIterator<Fst<Arc>>> aiter_;
-  bool error_;
-};
-
-// SafeReplaceAccumulator accumulates weights in a ReplaceFst and copies of it
-// are always thread-safe copies.
-template <class Accumulator, class T>
-class SafeReplaceAccumulator {
- public:
-  using Arc = typename Accumulator::Arc;
-  using StateId = typename Arc::StateId;
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-  using StateTable = T;
-  using StateTuple = typename StateTable::StateTuple;
-
-  SafeReplaceAccumulator() {}
-
-  SafeReplaceAccumulator(const SafeReplaceAccumulator &copy, bool safe)
-      : SafeReplaceAccumulator(copy) {}
-
-  explicit SafeReplaceAccumulator(
-      const std::vector<Accumulator> &accumulators) {
-    for (const auto &accumulator : accumulators) {
-      accumulators_.emplace_back(accumulator, true);
-    }
-  }
-
-  void Init(const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_tuples,
-            const StateTable *state_table) {
-    state_table_ = state_table;
-    for (Label i = 0; i < fst_tuples.size(); ++i) {
-      if (i == accumulators_.size()) {
-        accumulators_.resize(accumulators_.size() + 1);
-        accumulators_[i].Init(*(fst_tuples[i].second));
-      }
-      fst_array_.emplace_back(fst_tuples[i].second->Copy(true));
-    }
-    init_ = true;
-  }
-
-  void Init(const Fst<Arc> &fst, bool copy = false) {
-    if (!init_) {
-      FSTERROR() << "SafeReplaceAccumulator::Init: Accumulator needs to be"
-                 << " initialized before being passed to LookAheadMatcher";
-      error_ = true;
-    }
-  }
-
-  void SetState(StateId s) {
-    auto tuple = state_table_->Tuple(s);
-    fst_id_ = tuple.fst_id - 1;  // Replace FST ID is 1-based
-    GetAccumulator(fst_id_)->SetState(tuple.fst_state);
-    offset_ = 0;
-    offset_weight_ = Weight::Zero();
-    const auto final_weight = GetFst(fst_id_)->Final(tuple.fst_state);
-    if ((tuple.prefix_id != 0) && (final_weight != Weight::Zero())) {
-      offset_ = 1;
-      offset_weight_ = final_weight;
-    }
-    aiter_.Set(*GetFst(fst_id_), tuple.fst_state);
-  }
-
-  Weight Sum(Weight w, Weight v) {
-    if (error_) return Weight::NoWeight();
-    return GetAccumulator(fst_id_)->Sum(w, v);
-  }
-
-  template <class ArcIter>
-  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) {
-    if (error_) return Weight::NoWeight();
-    if (begin == end) return Weight::Zero();
-    auto sum = GetAccumulator(fst_id_)->Sum(
-        w, aiter_.get(), begin ? begin - offset_ : 0, end - offset_);
-    if (begin == 0 && end != 0 && offset_ > 0) {
-      sum = Sum(offset_weight_, sum);
-    }
-    return sum;
-  }
-
-  bool Error() const { return error_; }
-
- private:
-  class ArcIteratorPtr {
-   public:
-    ArcIteratorPtr() {}
-
-    ArcIteratorPtr(const ArcIteratorPtr &copy) {}
-
-    void Set(const Fst<Arc> &fst, StateId state_id) {
-      ptr_.reset(new ArcIterator<Fst<Arc>>(fst, state_id));
-    }
-
-    ArcIterator<Fst<Arc>> *get() { return ptr_.get(); }
-
-   private:
-    std::unique_ptr<ArcIterator<Fst<Arc>>> ptr_;
-  };
-
-  Accumulator *GetAccumulator(size_t i) { return &accumulators_[i]; }
-
-  const Fst<Arc> *GetFst(size_t i) const { return fst_array_[i].get(); }
-
-  const StateTable *state_table_;
-  std::vector<Accumulator> accumulators_;
-  std::vector<std::shared_ptr<Fst<Arc>>> fst_array_;
-  ArcIteratorPtr aiter_;
-  bool init_ = false;
-  bool error_ = false;
-  Label fst_id_;
-  size_t offset_;
-  Weight offset_weight_;
-};
-
-}  // namespace fst
-
-#endif  // FST_ACCUMULATOR_H_
diff --git a/projects/llm_framework/include/fst/add-on.h b/projects/llm_framework/include/fst/add-on.h
deleted file mode 100644
index 4a95111f..00000000
--- a/projects/llm_framework/include/fst/add-on.h
+++ /dev/null
@@ -1,248 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// FST implementation class to attach an arbitrary object with a read/write
-// method to an FST and its file representation. The FST is given a new type
-// name.
-
-#ifndef FST_ADD_ON_H_
-#define FST_ADD_ON_H_
-
-#include <stddef.h>
-#include <memory>
-#include <string>
-#include <utility>
-
-#include <fst/log.h>
-
-#include <fst/fst.h>
-
-
-namespace fst {
-
-// Identifies stream data as an add-on FST.
-static constexpr int32 kAddOnMagicNumber = 446681434;
-
-// Nothing to save.
-class NullAddOn {
- public:
-  NullAddOn() {}
-
-  static NullAddOn *Read(std::istream &strm, const FstReadOptions &opts) {
-    return new NullAddOn();
-  }
-
-  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
-    return true;
-  }
-};
-
-// Create a new add-on from a pair of add-ons.
-template <class A1, class A2>
-class AddOnPair {
- public:
-  // Argument reference count incremented.
-  AddOnPair(std::shared_ptr<A1> a1, std::shared_ptr<A2> a2)
-      : a1_(std::move(a1)), a2_(std::move(a2)) {}
-
-  const A1 *First() const { return a1_.get(); }
-
-  const A2 *Second() const { return a2_.get(); }
-
-  std::shared_ptr<A1> SharedFirst() const { return a1_; }
-
-  std::shared_ptr<A2> SharedSecond() const { return a2_; }
-
-  static AddOnPair<A1, A2> *Read(std::istream &istrm,
-                                 const FstReadOptions &opts) {
-    A1 *a1 = nullptr;
-    bool have_addon1 = false;
-    ReadType(istrm, &have_addon1);
-    if (have_addon1) a1 = A1::Read(istrm, opts);
-
-    A2 *a2 = nullptr;
-    bool have_addon2 = false;
-    ReadType(istrm, &have_addon2);
-    if (have_addon2) a2 = A2::Read(istrm, opts);
-
-    return new AddOnPair<A1, A2>(std::shared_ptr<A1>(a1),
-                                 std::shared_ptr<A2>(a2));
-  }
-
-  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
-    bool have_addon1 = a1_ != nullptr;
-    WriteType(ostrm, have_addon1);
-    if (have_addon1) a1_->Write(ostrm, opts);
-    bool have_addon2 = a2_ != nullptr;
-    WriteType(ostrm, have_addon2);
-    if (have_addon2) a2_->Write(ostrm, opts);
-    return true;
-  }
-
- private:
-  std::shared_ptr<A1> a1_;
-  std::shared_ptr<A2> a2_;
-};
-
-namespace internal {
-
-// Adds an object of type T to an FST. T must support:
-//
-//     T* Read(std::istream &);
-//     bool Write(std::ostream &);
-//
-// The resulting type is a new FST implementation.
-template <class FST, class T>
-class AddOnImpl : public FstImpl<typename FST::Arc> {
- public:
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::WriteHeader;
-
-  // We make a thread-safe copy of the FST by default since an FST
-  // implementation is expected to not share mutable data between objects.
-  AddOnImpl(const FST &fst, const string &type,
-            std::shared_ptr<T> t = std::shared_ptr<T>())
-      : fst_(fst, true), t_(std::move(t)) {
-    SetType(type);
-    SetProperties(fst_.Properties(kFstProperties, false));
-    SetInputSymbols(fst_.InputSymbols());
-    SetOutputSymbols(fst_.OutputSymbols());
-  }
-
-  // Conversion from const Fst<Arc> & to F always copies the underlying
-  // implementation.
-  AddOnImpl(const Fst<Arc> &fst, const string &type,
-            std::shared_ptr<T> t = std::shared_ptr<T>())
-      : fst_(fst), t_(std::move(t)) {
-    SetType(type);
-    SetProperties(fst_.Properties(kFstProperties, false));
-    SetInputSymbols(fst_.InputSymbols());
-    SetOutputSymbols(fst_.OutputSymbols());
-  }
-
-  // We make a thread-safe copy of the FST by default since an FST
-  // implementation is expected to not share mutable data between objects.
-  AddOnImpl(const AddOnImpl<FST, T> &impl)
-      : fst_(impl.fst_, true), t_(impl.t_) {
-    SetType(impl.Type());
-    SetProperties(fst_.Properties(kCopyProperties, false));
-    SetInputSymbols(fst_.InputSymbols());
-    SetOutputSymbols(fst_.OutputSymbols());
-  }
-
-  StateId Start() const { return fst_.Start(); }
-
-  Weight Final(StateId s) const { return fst_.Final(s); }
-
-  size_t NumArcs(StateId s) const { return fst_.NumArcs(s); }
-
-  size_t NumInputEpsilons(StateId s) const { return fst_.NumInputEpsilons(s); }
-
-  size_t NumOutputEpsilons(StateId s) const {
-    return fst_.NumOutputEpsilons(s);
-  }
-
-  size_t NumStates() const { return fst_.NumStates(); }
-
-  static AddOnImpl<FST, T> *Read(std::istream &strm,
-                                 const FstReadOptions &opts) {
-    FstReadOptions nopts(opts);
-    FstHeader hdr;
-    if (!nopts.header) {
-      hdr.Read(strm, nopts.source);
-      nopts.header = &hdr;
-    }
-    std::unique_ptr<AddOnImpl<FST, T>> impl(
-        new AddOnImpl<FST, T>(nopts.header->FstType()));
-    if (!impl->ReadHeader(strm, nopts, kMinFileVersion, &hdr)) return nullptr;
-    impl.reset();
-    int32 magic_number = 0;
-    ReadType(strm, &magic_number);  // Ensures this is an add-on FST.
-    if (magic_number != kAddOnMagicNumber) {
-      LOG(ERROR) << "AddOnImpl::Read: Bad add-on header: " << nopts.source;
-      return nullptr;
-    }
-    FstReadOptions fopts(opts);
-    fopts.header = nullptr;  // Contained header was written out.
-    std::unique_ptr<FST> fst(FST::Read(strm, fopts));
-    if (!fst) return nullptr;
-    std::shared_ptr<T> t;
-    bool have_addon = false;
-    ReadType(strm, &have_addon);
-    if (have_addon) {  // Reads add-on object if present.
-      t = std::shared_ptr<T>(T::Read(strm, fopts));
-      if (!t) return nullptr;
-    }
-    return new AddOnImpl<FST, T>(*fst, nopts.header->FstType(), t);
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
-    FstHeader hdr;
-    FstWriteOptions nopts(opts);
-    nopts.write_isymbols = false;  // Allows contained FST to hold any symbols.
-    nopts.write_osymbols = false;
-    WriteHeader(strm, nopts, kFileVersion, &hdr);
-    WriteType(strm, kAddOnMagicNumber);  // Ensures this is an add-on FST.
-    FstWriteOptions fopts(opts);
-    fopts.write_header = true;  // Forces writing contained header.
-    if (!fst_.Write(strm, fopts)) return false;
-    bool have_addon = !!t_;
-    WriteType(strm, have_addon);
-    // Writes add-on object if present.
-    if (have_addon) t_->Write(strm, opts);
-    return true;
-  }
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const {
-    fst_.InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const {
-    fst_.InitArcIterator(s, data);
-  }
-
-  FST &GetFst() { return fst_; }
-
-  const FST &GetFst() const { return fst_; }
-
-  const T *GetAddOn() const { return t_.get(); }
-
-  std::shared_ptr<T> GetSharedAddOn() const { return t_; }
-
-  void SetAddOn(std::shared_ptr<T> t) { t_ = t; }
-
- private:
-  explicit AddOnImpl(const string &type) : t_() {
-    SetType(type);
-    SetProperties(kExpanded);
-  }
-
-  // Current file format version.
-  static constexpr int kFileVersion = 1;
-  // Minimum file format version supported.
-  static constexpr int kMinFileVersion = 1;
-
-  FST fst_;
-  std::shared_ptr<T> t_;
-
-  AddOnImpl &operator=(const AddOnImpl &) = delete;
-};
-
-template <class FST, class T>
-constexpr int AddOnImpl<FST, T>::kFileVersion;
-
-template <class FST, class T>
-constexpr int AddOnImpl<FST, T>::kMinFileVersion;
-
-}  // namespace internal
-}  // namespace fst
-
-#endif  // FST_ADD_ON_H_
diff --git a/projects/llm_framework/include/fst/arc-arena.h b/projects/llm_framework/include/fst/arc-arena.h
deleted file mode 100644
index 13fe918a..00000000
--- a/projects/llm_framework/include/fst/arc-arena.h
+++ /dev/null
@@ -1,232 +0,0 @@
-#ifndef FST_ARC_ARENA_H_
-#define FST_ARC_ARENA_H_
-
-#include <deque>
-#include <memory>
-#include <utility>
-#include <fst/fst.h>
-#include <fst/memory.h>
-#include <unordered_map>
-
-namespace fst {
-
-// ArcArena is used for fast allocation of contiguous arrays of arcs.
-//
-// To create an arc array:
-//   for each state:
-//     for each arc:
-//       arena.PushArc();
-//     // Commits these arcs and returns pointer to them.
-//     Arc *arcs = arena.GetArcs();
-//
-//     OR
-//
-//     arena.DropArcs();  // Throws away current arcs, reuse the space.
-//
-// The arcs returned are guaranteed to be contiguous and the pointer returned
-// will never be invalidated until the arena is cleared for reuse.
-//
-// The contents of the arena can be released with a call to arena.Clear() after
-// which the arena will restart with an initial allocation capable of holding at
-// least all of the arcs requested in the last usage before Clear() making
-// subsequent uses of the Arena more efficient.
-//
-// The max_retained_size option can limit the amount of arc space requested on
-// Clear() to avoid excess growth from intermittent high usage.
-template <typename Arc>
-class ArcArena {
- public:
-  explicit ArcArena(size_t block_size = 256,
-                    size_t max_retained_size = 1e6)
-      : block_size_(block_size),
-        max_retained_size_(max_retained_size) {
-    blocks_.emplace_back(MakeSharedBlock(block_size_));
-    first_block_size_ = block_size_;
-    total_size_ = block_size_;
-    arcs_ = blocks_.back().get();
-    end_ = arcs_ + block_size_;
-    next_ = arcs_;
-  }
-
-  ArcArena(const ArcArena& copy)
-      : arcs_(copy.arcs_), next_(copy.next_), end_(copy.end_),
-        block_size_(copy.block_size_),
-        first_block_size_(copy.first_block_size_),
-        total_size_(copy.total_size_),
-        max_retained_size_(copy.max_retained_size_),
-        blocks_(copy.blocks_) {
-    NewBlock(block_size_);
-  }
-
-  void ReserveArcs(size_t n) {
-    if (next_ + n < end_) return;
-    NewBlock(n);
-  }
-
-  void PushArc(const Arc& arc) {
-    if (next_ == end_) {
-      size_t length = next_ - arcs_;
-      NewBlock(length * 2);
-    }
-    *next_ = arc;
-    ++next_;
-  }
-
-  const Arc* GetArcs() {
-    const auto *arcs = arcs_;
-    arcs_ = next_;
-    return arcs;
-  }
-
-  void DropArcs() { next_ = arcs_; }
-
-  size_t Size() { return total_size_; }
-
-  void Clear() {
-    blocks_.resize(1);
-    if (total_size_ > first_block_size_) {
-      first_block_size_ = std::min(max_retained_size_, total_size_);
-      blocks_.back() = MakeSharedBlock(first_block_size_);
-    }
-    total_size_ = first_block_size_;
-    arcs_ = blocks_.back().get();
-    end_ = arcs_ + first_block_size_;
-    next_ = arcs_;
-  }
-
- private:
-  // Allocates a new block with capacity of at least n or block_size,
-  // copying incomplete arc sequence from old block to new block.
-  void NewBlock(size_t n) {
-    const auto length = next_ - arcs_;
-    const auto new_block_size = std::max(n, block_size_);
-    total_size_ += new_block_size;
-    blocks_.emplace_back(MakeSharedBlock(new_block_size));
-    std::copy(arcs_, next_, blocks_.back().get());
-    arcs_ = blocks_.back().get();
-    next_ = arcs_ + length;
-    end_ = arcs_ + new_block_size;
-  }
-
-  std::shared_ptr<Arc> MakeSharedBlock(size_t size) {
-    return std::shared_ptr<Arc>(new Arc[size], std::default_delete<Arc[]>());
-  }
-
-  Arc *arcs_;
-  Arc *next_;
-  const Arc *end_;
-  size_t block_size_;
-  size_t first_block_size_;
-  size_t total_size_;
-  size_t max_retained_size_;
-  std::list<std::shared_ptr<Arc>> blocks_;
-};
-
-// ArcArenaStateStore uses a resusable ArcArena to store arc arrays and does not
-// require that the Expander call ReserveArcs first.
-//
-// TODO(tombagby): Make cache type configurable.
-// TODO(tombagby): Provide ThreadLocal/Concurrent configuration.
-template <class A>
-class ArcArenaStateStore {
- public:
-  using Arc = A;
-  using Weight = typename Arc::Weight;
-  using StateId = typename Arc::StateId;
-
-  ArcArenaStateStore() : arena_(64 * 1024) {
-  }
-
-  class State {
-   public:
-    Weight Final() const { return final_; }
-
-    size_t NumInputEpsilons() const { return niepsilons_; }
-
-    size_t NumOutputEpsilons() const { return noepsilons_; }
-
-    size_t NumArcs() const { return narcs_; }
-
-    const Arc &GetArc(size_t n) const { return arcs_[n]; }
-
-    const Arc *Arcs() const { return arcs_; }
-
-    int* MutableRefCount() const { return nullptr; }
-
-   private:
-    State(Weight weight, int32 niepsilons, int32 noepsilons, int32 narcs,
-          const Arc *arcs)
-        : final_(std::move(weight)),
-          niepsilons_(niepsilons),
-          noepsilons_(noepsilons),
-          narcs_(narcs),
-          arcs_(arcs) {}
-
-    Weight final_;
-    size_t niepsilons_;
-    size_t noepsilons_;
-    size_t narcs_;
-    const Arc *arcs_;
-
-    friend class ArcArenaStateStore<Arc>;
-  };
-
-  template <class Expander>
-  State *FindOrExpand(Expander &expander, StateId state_id) {  // NOLINT
-    auto it = cache_.insert(std::pair<StateId, State*>(state_id, nullptr));
-    if (!it.second) return it.first->second;
-    // Needs a new state.
-    StateBuilder builder(&arena_);
-    expander.Expand(state_id, &builder);
-    const auto arcs = arena_.GetArcs();
-    size_t narcs = builder.narcs_;
-    size_t niepsilons = 0;
-    size_t noepsilons = 0;
-    for (size_t i = 0; i < narcs; ++i) {
-      if (arcs[i].ilabel == 0) ++niepsilons;
-      if (arcs[i].olabel == 0) ++noepsilons;
-    }
-    states_.emplace_back(
-        State(builder.final_, niepsilons, noepsilons, narcs, arcs));
-    // Places it in the cache.
-    auto state = &states_.back();
-    it.first->second = state;
-    return state;
-  }
-
-  State *Find(StateId state_id) const {
-    auto it = cache_.find(state_id);
-    return (it == cache_.end()) ? nullptr : it->second;
-  }
-
- private:
-  class StateBuilder {
-   public:
-    explicit StateBuilder(ArcArena<Arc>* arena)
-       : arena_(arena), final_(Weight::Zero()), narcs_(0) {}
-
-    void SetFinal(Weight weight) { final_ = std::move(weight); }
-
-    void ReserveArcs(size_t n) { arena_->ReserveArcs(n); }
-
-    void AddArc(const Arc &arc) {
-      ++narcs_;
-      arena_->PushArc(arc);
-    }
-
-   private:
-    friend class ArcArenaStateStore<Arc>;
-
-    ArcArena<Arc> *arena_;
-    Weight final_;
-    size_t narcs_;
-  };
-
-  std::unordered_map<StateId, State *> cache_;
-  std::deque<State> states_;
-  ArcArena<Arc> arena_;
-};
-
-}  // namespace fst
-
-#endif  // FST_ARC_ARENA_H_
diff --git a/projects/llm_framework/include/fst/arc-map.h b/projects/llm_framework/include/fst/arc-map.h
deleted file mode 100644
index 24db4911..00000000
--- a/projects/llm_framework/include/fst/arc-map.h
+++ /dev/null
@@ -1,1285 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to map over/transform arcs e.g., change semirings or
-// implement project/invert. Consider using when operation does
-// not change the number of arcs (except possibly superfinal arcs).
-
-#ifndef FST_ARC_MAP_H_
-#define FST_ARC_MAP_H_
-
-#include <string>
-#include <unordered_map>
-#include <utility>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-// Determines how final weights are mapped.
-enum MapFinalAction {
-  // A final weight is mapped into a final weight. An error is raised if this
-  // is not possible.
-  MAP_NO_SUPERFINAL,
-  // A final weight is mapped to an arc to the superfinal state when the result
-  // cannot be represented as a final weight. The superfinal state will be
-  // added only if it is needed.
-  MAP_ALLOW_SUPERFINAL,
-  // A final weight is mapped to an arc to the superfinal state unless the
-  // result can be represented as a final weight of weight Zero(). The
-  // superfinal state is always added (if the input is not the empty FST).
-  MAP_REQUIRE_SUPERFINAL
-};
-
-// Determines how symbol tables are mapped.
-enum MapSymbolsAction {
-  // Symbols should be cleared in the result by the map.
-  MAP_CLEAR_SYMBOLS,
-  // Symbols should be copied from the input FST by the map.
-  MAP_COPY_SYMBOLS,
-  // Symbols should not be modified in the result by the map itself.
-  // (They may set by the mapper).
-  MAP_NOOP_SYMBOLS
-};
-
-// The ArcMapper interfaces defines how arcs and final weights are mapped.
-// This is useful for implementing operations that do not change the number of
-// arcs (except possibly superfinal arcs).
-//
-// template <class A, class B>
-// class ArcMapper {
-//  public:
-//   using FromArc = A;
-//   using ToArc = B;
-//
-//   // Maps an arc type FromArc to arc type ToArc.
-//   ToArc operator()(const FromArc &arc);
-//
-//   // Specifies final action the mapper requires (see above).
-//   // The mapper will be passed final weights as arcs of the form
-//   // Arc(0, 0, weight, kNoStateId).
-//   MapFinalAction FinalAction() const;
-//
-//   // Specifies input symbol table action the mapper requires (see above).
-//   MapSymbolsAction InputSymbolsAction() const;
-//
-//   // Specifies output symbol table action the mapper requires (see above).
-//   MapSymbolsAction OutputSymbolsAction() const;
-//
-//   // This specifies the known properties of an FST mapped by this mapper. It
-//   takes as argument the input FSTs's known properties.
-//   uint64 Properties(uint64 props) const;
-// };
-//
-// The ArcMap functions and classes below will use the FinalAction()
-// method of the mapper to determine how to treat final weights, e.g., whether
-// to add a superfinal state. They will use the Properties() method to set the
-// result FST properties.
-//
-// We include a various map versions below. One dimension of variation is
-// whether the mapping mutates its input, writes to a new result FST, or is an
-// on-the-fly FST. Another dimension is how we pass the mapper. We allow passing
-// the mapper by pointer for cases that we need to change the state of the
-// user's mapper.  This is the case with the EncodeMapper, which is reused
-// during decoding. We also include map versions that pass the mapper by value
-// or const reference when this suffices.
-
-// Maps an arc type A using a mapper function object C, passed
-// by pointer.  This version modifies its Fst input.
-template <class A, class C>
-void ArcMap(MutableFst<A> *fst, C *mapper) {
-  using FromArc = A;
-  using ToArc = A;
-  using Weight = typename FromArc::Weight;
-  if (mapper->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    fst->SetInputSymbols(nullptr);
-  }
-  if (mapper->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    fst->SetOutputSymbols(nullptr);
-  }
-  if (fst->Start() == kNoStateId) return;
-  const auto props = fst->Properties(kFstProperties, false);
-  const auto final_action = mapper->FinalAction();
-  auto superfinal = kNoStateId;
-  if (final_action == MAP_REQUIRE_SUPERFINAL) {
-    superfinal = fst->AddState();
-    fst->SetFinal(superfinal, Weight::One());
-  }
-  for (StateIterator<MutableFst<FromArc>> siter(*fst); !siter.Done();
-       siter.Next()) {
-    const auto state = siter.Value();
-    for (MutableArcIterator<MutableFst<FromArc>> aiter(fst, state);
-         !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      aiter.SetValue((*mapper)(arc));
-    }
-    switch (final_action) {
-      case MAP_NO_SUPERFINAL:
-      default: {
-        const FromArc arc(0, 0, fst->Final(state), kNoStateId);
-        const auto final_arc = (*mapper)(arc);
-        if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
-          FSTERROR() << "ArcMap: Non-zero arc labels for superfinal arc";
-          fst->SetProperties(kError, kError);
-        }
-        fst->SetFinal(state, final_arc.weight);
-        break;
-      }
-      case MAP_ALLOW_SUPERFINAL: {
-        if (state != superfinal) {
-          const FromArc arc(0, 0, fst->Final(state), kNoStateId);
-          auto final_arc = (*mapper)(arc);
-          if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
-            // Add a superfinal state if not already done.
-            if (superfinal == kNoStateId) {
-              superfinal = fst->AddState();
-              fst->SetFinal(superfinal, Weight::One());
-            }
-            final_arc.nextstate = superfinal;
-            fst->AddArc(state, std::move(final_arc));
-            fst->SetFinal(state, Weight::Zero());
-          } else {
-            fst->SetFinal(state, final_arc.weight);
-          }
-        }
-        break;
-      }
-      case MAP_REQUIRE_SUPERFINAL: {
-        if (state != superfinal) {
-          const FromArc arc(0, 0, fst->Final(state), kNoStateId);
-          const auto final_arc = (*mapper)(arc);
-          if (final_arc.ilabel != 0 || final_arc.olabel != 0 ||
-              final_arc.weight != Weight::Zero()) {
-            fst->AddArc(state, ToArc(final_arc.ilabel, final_arc.olabel,
-                                     final_arc.weight, superfinal));
-          }
-          fst->SetFinal(state, Weight::Zero());
-        }
-        break;
-      }
-    }
-  }
-  fst->SetProperties(mapper->Properties(props), kFstProperties);
-}
-
-// Maps an arc type A using a mapper function object C, passed by value. This
-// version modifies its FST input.
-template <class A, class C>
-void ArcMap(MutableFst<A> *fst, C mapper) {
-  ArcMap(fst, &mapper);
-}
-
-// Maps an arc type A to an arc type B using mapper function object C,
-// passed by pointer. This version writes the mapped input FST to an
-// output MutableFst.
-template <class A, class B, class C>
-void ArcMap(const Fst<A> &ifst, MutableFst<B> *ofst, C *mapper) {
-  using FromArc = A;
-  using StateId = typename FromArc::StateId;
-  ofst->DeleteStates();
-  if (mapper->InputSymbolsAction() == MAP_COPY_SYMBOLS) {
-    ofst->SetInputSymbols(ifst.InputSymbols());
-  } else if (mapper->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    ofst->SetInputSymbols(nullptr);
-  }
-  if (mapper->OutputSymbolsAction() == MAP_COPY_SYMBOLS) {
-    ofst->SetOutputSymbols(ifst.OutputSymbols());
-  } else if (mapper->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    ofst->SetOutputSymbols(nullptr);
-  }
-  const auto iprops = ifst.Properties(kCopyProperties, false);
-  if (ifst.Start() == kNoStateId) {
-    if (iprops & kError) ofst->SetProperties(kError, kError);
-    return;
-  }
-  const auto final_action = mapper->FinalAction();
-  if (ifst.Properties(kExpanded, false)) {
-    ofst->ReserveStates(
-        CountStates(ifst) + (final_action == MAP_NO_SUPERFINAL ? 0 : 1));
-  }
-  // Adds all states.
-  for (StateIterator<Fst<A>> siter(ifst); !siter.Done(); siter.Next()) {
-    ofst->AddState();
-  }
-  StateId superfinal = kNoStateId;
-  if (final_action == MAP_REQUIRE_SUPERFINAL) {
-    superfinal = ofst->AddState();
-    ofst->SetFinal(superfinal, B::Weight::One());
-  }
-  for (StateIterator<Fst<A>> siter(ifst); !siter.Done(); siter.Next()) {
-    StateId s = siter.Value();
-    if (s == ifst.Start()) ofst->SetStart(s);
-    ofst->ReserveArcs(
-        s, ifst.NumArcs(s) + (final_action != MAP_NO_SUPERFINAL ? 1 : 0));
-    for (ArcIterator<Fst<A>> aiter(ifst, s); !aiter.Done(); aiter.Next()) {
-      ofst->AddArc(s, (*mapper)(aiter.Value()));
-    }
-    switch (final_action) {
-      case MAP_NO_SUPERFINAL:
-      default: {
-        B final_arc = (*mapper)(A(0, 0, ifst.Final(s), kNoStateId));
-        if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
-          FSTERROR() << "ArcMap: Non-zero arc labels for superfinal arc";
-          ofst->SetProperties(kError, kError);
-        }
-        ofst->SetFinal(s, final_arc.weight);
-        break;
-      }
-      case MAP_ALLOW_SUPERFINAL: {
-        B final_arc = (*mapper)(A(0, 0, ifst.Final(s), kNoStateId));
-        if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
-          // Add a superfinal state if not already done.
-          if (superfinal == kNoStateId) {
-            superfinal = ofst->AddState();
-            ofst->SetFinal(superfinal, B::Weight::One());
-          }
-          final_arc.nextstate = superfinal;
-          ofst->AddArc(s, std::move(final_arc));
-          ofst->SetFinal(s, B::Weight::Zero());
-        } else {
-          ofst->SetFinal(s, final_arc.weight);
-        }
-        break;
-      }
-      case MAP_REQUIRE_SUPERFINAL: {
-        B final_arc = (*mapper)(A(0, 0, ifst.Final(s), kNoStateId));
-        if (final_arc.ilabel != 0 || final_arc.olabel != 0 ||
-            final_arc.weight != B::Weight::Zero()) {
-          ofst->AddArc(s, B(final_arc.ilabel, final_arc.olabel,
-                            final_arc.weight, superfinal));
-        }
-        ofst->SetFinal(s, B::Weight::Zero());
-        break;
-      }
-    }
-  }
-  const auto oprops = ofst->Properties(kFstProperties, false);
-  ofst->SetProperties(mapper->Properties(iprops) | oprops, kFstProperties);
-}
-
-// Maps an arc type A to an arc type B using mapper function
-// object C, passed by value. This version writes the mapped input
-// Fst to an output MutableFst.
-template <class A, class B, class C>
-void ArcMap(const Fst<A> &ifst, MutableFst<B> *ofst, C mapper) {
-  ArcMap(ifst, ofst, &mapper);
-}
-
-struct ArcMapFstOptions : public CacheOptions {
-  // ArcMapFst default caching behaviour is to do no caching. Most mappers are
-  // cheap and therefore we save memory by not doing caching.
-  ArcMapFstOptions() : CacheOptions(true, 0) {}
-
-  explicit ArcMapFstOptions(const CacheOptions &opts) : CacheOptions(opts) {}
-};
-
-template <class A, class B, class C>
-class ArcMapFst;
-
-namespace internal {
-
-// Implementation of delayed ArcMapFst.
-template <class A, class B, class C>
-class ArcMapFstImpl : public CacheImpl<B> {
- public:
-  using Arc = B;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<B>::SetType;
-  using FstImpl<B>::SetProperties;
-  using FstImpl<B>::SetInputSymbols;
-  using FstImpl<B>::SetOutputSymbols;
-
-  using CacheImpl<B>::EmplaceArc;
-  using CacheImpl<B>::HasArcs;
-  using CacheImpl<B>::HasFinal;
-  using CacheImpl<B>::HasStart;
-  using CacheImpl<B>::PushArc;
-  using CacheImpl<B>::SetArcs;
-  using CacheImpl<B>::SetFinal;
-  using CacheImpl<B>::SetStart;
-
-  friend class StateIterator<ArcMapFst<A, B, C>>;
-
-  ArcMapFstImpl(const Fst<A> &fst, const C &mapper,
-                const ArcMapFstOptions &opts)
-      : CacheImpl<B>(opts),
-        fst_(fst.Copy()),
-        mapper_(new C(mapper)),
-        own_mapper_(true),
-        superfinal_(kNoStateId),
-        nstates_(0) {
-    Init();
-  }
-
-  ArcMapFstImpl(const Fst<A> &fst, C *mapper, const ArcMapFstOptions &opts)
-      : CacheImpl<B>(opts),
-        fst_(fst.Copy()),
-        mapper_(mapper),
-        own_mapper_(false),
-        superfinal_(kNoStateId),
-        nstates_(0) {
-    Init();
-  }
-
-  ArcMapFstImpl(const ArcMapFstImpl<A, B, C> &impl)
-      : CacheImpl<B>(impl),
-        fst_(impl.fst_->Copy(true)),
-        mapper_(new C(*impl.mapper_)),
-        own_mapper_(true),
-        superfinal_(kNoStateId),
-        nstates_(0) {
-    Init();
-  }
-
-  ~ArcMapFstImpl() override {
-    if (own_mapper_) delete mapper_;
-  }
-
-  StateId Start() {
-    if (!HasStart()) SetStart(FindOState(fst_->Start()));
-    return CacheImpl<B>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) {
-      switch (final_action_) {
-        case MAP_NO_SUPERFINAL:
-        default: {
-          const auto final_arc =
-              (*mapper_)(A(0, 0, fst_->Final(FindIState(s)), kNoStateId));
-          if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
-            FSTERROR() << "ArcMapFst: Non-zero arc labels for superfinal arc";
-            SetProperties(kError, kError);
-          }
-          SetFinal(s, final_arc.weight);
-          break;
-        }
-        case MAP_ALLOW_SUPERFINAL: {
-          if (s == superfinal_) {
-            SetFinal(s, Weight::One());
-          } else {
-            const auto final_arc =
-                (*mapper_)(A(0, 0, fst_->Final(FindIState(s)), kNoStateId));
-            if (final_arc.ilabel == 0 && final_arc.olabel == 0) {
-              SetFinal(s, final_arc.weight);
-            } else {
-              SetFinal(s, Weight::Zero());
-            }
-          }
-          break;
-        }
-        case MAP_REQUIRE_SUPERFINAL: {
-          SetFinal(s, s == superfinal_ ? Weight::One() : Weight::Zero());
-          break;
-        }
-      }
-    }
-    return CacheImpl<B>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<B>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<B>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<B>::NumOutputEpsilons(s);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && (fst_->Properties(kError, false) ||
-                            (mapper_->Properties(0) & kError))) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<B> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<B>::InitArcIterator(s, data);
-  }
-
-  void Expand(StateId s) {
-    // Add exiting arcs.
-    if (s == superfinal_) {
-      SetArcs(s);
-      return;
-    }
-    for (ArcIterator<Fst<A>> aiter(*fst_, FindIState(s)); !aiter.Done();
-         aiter.Next()) {
-      auto aarc = aiter.Value();
-      aarc.nextstate = FindOState(aarc.nextstate);
-      PushArc(s, (*mapper_)(aarc));
-    }
-
-    // Check for superfinal arcs.
-    if (!HasFinal(s) || Final(s) == Weight::Zero()) {
-      switch (final_action_) {
-        case MAP_NO_SUPERFINAL:
-        default:
-          break;
-        case MAP_ALLOW_SUPERFINAL: {
-          auto final_arc =
-              (*mapper_)(A(0, 0, fst_->Final(FindIState(s)), kNoStateId));
-          if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
-            if (superfinal_ == kNoStateId) superfinal_ = nstates_++;
-            final_arc.nextstate = superfinal_;
-            PushArc(s, std::move(final_arc));
-          }
-          break;
-        }
-        case MAP_REQUIRE_SUPERFINAL: {
-          const auto final_arc =
-              (*mapper_)(A(0, 0, fst_->Final(FindIState(s)), kNoStateId));
-          if (final_arc.ilabel != 0 || final_arc.olabel != 0 ||
-              final_arc.weight != B::Weight::Zero()) {
-            EmplaceArc(s, final_arc.ilabel, final_arc.olabel, final_arc.weight,
-                       superfinal_);
-          }
-          break;
-        }
-      }
-    }
-    SetArcs(s);
-  }
-
- private:
-  void Init() {
-    SetType("map");
-    if (mapper_->InputSymbolsAction() == MAP_COPY_SYMBOLS) {
-      SetInputSymbols(fst_->InputSymbols());
-    } else if (mapper_->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-      SetInputSymbols(nullptr);
-    }
-    if (mapper_->OutputSymbolsAction() == MAP_COPY_SYMBOLS) {
-      SetOutputSymbols(fst_->OutputSymbols());
-    } else if (mapper_->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-      SetOutputSymbols(nullptr);
-    }
-    if (fst_->Start() == kNoStateId) {
-      final_action_ = MAP_NO_SUPERFINAL;
-      SetProperties(kNullProperties);
-    } else {
-      final_action_ = mapper_->FinalAction();
-      uint64 props = fst_->Properties(kCopyProperties, false);
-      SetProperties(mapper_->Properties(props));
-      if (final_action_ == MAP_REQUIRE_SUPERFINAL) superfinal_ = 0;
-    }
-  }
-
-  // Maps from output state to input state.
-  StateId FindIState(StateId s) {
-    if (superfinal_ == kNoStateId || s < superfinal_) {
-      return s;
-    } else {
-      return s - 1;
-    }
-  }
-
-  // Maps from input state to output state.
-  StateId FindOState(StateId is) {
-    auto os = is;
-    if (!(superfinal_ == kNoStateId || is < superfinal_)) ++os;
-    if (os >= nstates_) nstates_ = os + 1;
-    return os;
-  }
-
-  std::unique_ptr<const Fst<A>> fst_;
-  C *mapper_;
-  const bool own_mapper_;
-  MapFinalAction final_action_;
-  StateId superfinal_;
-  StateId nstates_;
-};
-
-}  // namespace internal
-
-// Maps an arc type A to an arc type B using Mapper function object
-// C. This version is a delayed FST.
-template <class A, class B, class C>
-class ArcMapFst : public ImplToFst<internal::ArcMapFstImpl<A, B, C>> {
- public:
-  using Arc = B;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<B>;
-  using State = typename Store::State;
-  using Impl = internal::ArcMapFstImpl<A, B, C>;
-
-  friend class ArcIterator<ArcMapFst<A, B, C>>;
-  friend class StateIterator<ArcMapFst<A, B, C>>;
-
-  ArcMapFst(const Fst<A> &fst, const C &mapper, const ArcMapFstOptions &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, mapper, opts)) {}
-
-  ArcMapFst(const Fst<A> &fst, C *mapper, const ArcMapFstOptions &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, mapper, opts)) {}
-
-  ArcMapFst(const Fst<A> &fst, const C &mapper)
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, mapper, ArcMapFstOptions())) {}
-
-  ArcMapFst(const Fst<A> &fst, C *mapper)
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, mapper, ArcMapFstOptions())) {}
-
-  // See Fst<>::Copy() for doc.
-  ArcMapFst(const ArcMapFst<A, B, C> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this ArcMapFst. See Fst<>::Copy() for further doc.
-  ArcMapFst<A, B, C> *Copy(bool safe = false) const override {
-    return new ArcMapFst<A, B, C>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<B> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<B> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
- protected:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
- private:
-  ArcMapFst &operator=(const ArcMapFst &) = delete;
-};
-
-// Specialization for ArcMapFst.
-//
-// This may be derived from.
-template <class A, class B, class C>
-class StateIterator<ArcMapFst<A, B, C>> : public StateIteratorBase<B> {
- public:
-  using StateId = typename B::StateId;
-
-  explicit StateIterator(const ArcMapFst<A, B, C> &fst)
-      : impl_(fst.GetImpl()),
-        siter_(*impl_->fst_),
-        s_(0),
-        superfinal_(impl_->final_action_ == MAP_REQUIRE_SUPERFINAL) {
-    CheckSuperfinal();
-  }
-
-  bool Done() const final { return siter_.Done() && !superfinal_; }
-
-  StateId Value() const final { return s_; }
-
-  void Next() final {
-    ++s_;
-    if (!siter_.Done()) {
-      siter_.Next();
-      CheckSuperfinal();
-    } else if (superfinal_) {
-      superfinal_ = false;
-    }
-  }
-
-  void Reset() final {
-    s_ = 0;
-    siter_.Reset();
-    superfinal_ = impl_->final_action_ == MAP_REQUIRE_SUPERFINAL;
-    CheckSuperfinal();
-  }
-
- private:
-  void CheckSuperfinal() {
-    if (impl_->final_action_ != MAP_ALLOW_SUPERFINAL || superfinal_) return;
-    if (!siter_.Done()) {
-      const auto final_arc =
-          (*impl_->mapper_)(A(0, 0, impl_->fst_->Final(s_), kNoStateId));
-      if (final_arc.ilabel != 0 || final_arc.olabel != 0) superfinal_ = true;
-    }
-  }
-
-  const internal::ArcMapFstImpl<A, B, C> *impl_;
-  StateIterator<Fst<A>> siter_;
-  StateId s_;
-  bool superfinal_;  // True if there is a superfinal state and not done.
-};
-
-// Specialization for ArcMapFst.
-template <class A, class B, class C>
-class ArcIterator<ArcMapFst<A, B, C>>
-    : public CacheArcIterator<ArcMapFst<A, B, C>> {
- public:
-  using StateId = typename A::StateId;
-
-  ArcIterator(const ArcMapFst<A, B, C> &fst, StateId s)
-      : CacheArcIterator<ArcMapFst<A, B, C>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class A, class B, class C>
-inline void ArcMapFst<A, B, C>::InitStateIterator(
-    StateIteratorData<B> *data) const {
-  data->base = new StateIterator<ArcMapFst<A, B, C>>(*this);
-}
-
-// Utility Mappers.
-
-// Mapper that returns its input.
-template <class A>
-class IdentityArcMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-
-  constexpr ToArc operator()(const FromArc &arc) const { return arc; }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr uint64 Properties(uint64 props) const { return props; }
-};
-
-// Mapper that converts all input symbols to epsilon.
-template <class A>
-class InputEpsilonMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-
-  constexpr ToArc operator()(const FromArc &arc) const {
-    return ToArc(0, arc.olabel, arc.weight, arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr uint64 Properties(uint64 props) const {
-    return (props & kSetArcProperties) | kIEpsilons | kILabelSorted;
-  }
-};
-
-// Mapper that converts all output symbols to epsilon.
-template <class A>
-class OutputEpsilonMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-
-  constexpr ToArc operator()(const FromArc &arc) const {
-    return ToArc(arc.ilabel, 0, arc.weight, arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  constexpr uint64 Properties(uint64 props) const {
-    return (props & kSetArcProperties) | kOEpsilons | kOLabelSorted;
-  }
-};
-
-// Mapper that returns its input with final states redirected to a single
-// super-final state.
-template <class A>
-class SuperFinalMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-  using Label = typename FromArc::Label;
-  using Weight = typename FromArc::Weight;;
-
-  // Arg allows setting super-final label.
-  explicit SuperFinalMapper(Label final_label = 0)
-      : final_label_(final_label) {}
-
-  ToArc operator()(const FromArc &arc) const {
-    // Super-final arc.
-    if (arc.nextstate == kNoStateId && arc.weight != Weight::Zero()) {
-      return ToArc(final_label_, final_label_, arc.weight, kNoStateId);
-    } else {
-      return arc;
-    }
-  }
-
-  constexpr MapFinalAction FinalAction() const {
-    return MAP_REQUIRE_SUPERFINAL;
-  }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    if (final_label_ == 0) {
-      return props & kAddSuperFinalProperties;
-    } else {
-      return props & kAddSuperFinalProperties &
-          kILabelInvariantProperties & kOLabelInvariantProperties;
-    }
-  }
-
- private:
-  Label final_label_;
-};
-
-// Mapper that leaves labels and nextstate unchanged and constructs a new weight
-// from the underlying value of the arc weight. If no weight converter is
-// explictly specified, requires that there is a WeightConvert class
-// specialization that converts the weights.
-template <class A, class B,
-          class C = WeightConvert<typename A::Weight, typename B::Weight>>
-class WeightConvertMapper {
- public:
-  using FromArc = A;
-  using ToArc = B;
-  using Converter = C;
-  using FromWeight = typename FromArc::Weight;
-  using ToWeight = typename ToArc::Weight;
-
-  constexpr explicit WeightConvertMapper(const Converter &c = Converter())
-      : convert_weight_(c) {}
-
-  constexpr ToArc operator()(const FromArc &arc) const {
-    return ToArc(arc.ilabel, arc.olabel, convert_weight_(arc.weight),
-                 arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr uint64 Properties(uint64 props) const { return props; }
-
- private:
-  const Converter convert_weight_;
-};
-
-// Non-precision-changing weight conversions; consider using more efficient
-// Cast method instead.
-
-using StdToLogMapper = WeightConvertMapper<StdArc, LogArc>;
-
-using LogToStdMapper = WeightConvertMapper<LogArc, StdArc>;
-
-// Precision-changing weight conversions.
-
-using StdToLog64Mapper = WeightConvertMapper<StdArc, Log64Arc>;
-
-using LogToLog64Mapper = WeightConvertMapper<LogArc, Log64Arc>;
-
-using Log64ToStdMapper = WeightConvertMapper<Log64Arc, StdArc>;
-
-using Log64ToLogMapper = WeightConvertMapper<Log64Arc, LogArc>;
-
-// Mapper from A to GallicArc<A>.
-template <class A, GallicType G = GALLIC_LEFT>
-class ToGallicMapper {
- public:
-  using FromArc = A;
-  using ToArc = GallicArc<A, G>;
-
-  using SW = StringWeight<typename A::Label, GallicStringType(G)>;
-  using AW = typename FromArc::Weight;
-  using GW = typename ToArc::Weight;
-
-  ToArc operator()(const FromArc &arc) const {
-    // Super-final arc.
-    if (arc.nextstate == kNoStateId && arc.weight != AW::Zero()) {
-      return ToArc(0, 0, GW(SW::One(), arc.weight), kNoStateId);
-    // Super-non-final arc.
-    } else if (arc.nextstate == kNoStateId) {
-      return ToArc(0, 0, GW::Zero(), kNoStateId);
-    // Epsilon label.
-    } else if (arc.olabel == 0) {
-      return ToArc(arc.ilabel, arc.ilabel, GW(SW::One(), arc.weight),
-                   arc.nextstate);
-    // Regular label.
-    } else {
-      return ToArc(arc.ilabel, arc.ilabel, GW(SW(arc.olabel), arc.weight),
-                   arc.nextstate);
-    }
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return ProjectProperties(props, true) & kWeightInvariantProperties;
-  }
-};
-
-// Mapper from GallicArc<A> to A.
-template <class A, GallicType G = GALLIC_LEFT>
-class FromGallicMapper {
- public:
-  using FromArc = GallicArc<A, G>;
-  using ToArc = A;
-
-  using Label = typename ToArc::Label;
-  using AW = typename ToArc::Weight;
-  using GW = typename FromArc::Weight;
-
-  explicit FromGallicMapper(Label superfinal_label = 0)
-      : superfinal_label_(superfinal_label), error_(false) {}
-
-  ToArc operator()(const FromArc &arc) const {
-    // 'Super-non-final' arc.
-    if (arc.nextstate == kNoStateId && arc.weight == GW::Zero()) {
-      return A(arc.ilabel, 0, AW::Zero(), kNoStateId);
-    }
-    Label l = kNoLabel;
-    AW weight;
-    if (!Extract(arc.weight, &weight, &l) || arc.ilabel != arc.olabel) {
-      FSTERROR() << "FromGallicMapper: Unrepresentable weight: " << arc.weight
-                 << " for arc with ilabel = " << arc.ilabel
-                 << ", olabel = " << arc.olabel
-                 << ", nextstate = " << arc.nextstate;
-      error_ = true;
-    }
-    if (arc.ilabel == 0 && l != 0 && arc.nextstate == kNoStateId) {
-      return ToArc(superfinal_label_, l, weight, arc.nextstate);
-    } else {
-      return ToArc(arc.ilabel, l, weight, arc.nextstate);
-    }
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_ALLOW_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 inprops) const {
-    uint64 outprops = inprops & kOLabelInvariantProperties &
-                      kWeightInvariantProperties & kAddSuperFinalProperties;
-    if (error_) outprops |= kError;
-    return outprops;
-  }
-
- private:
-  template <GallicType GT>
-  static bool Extract(const GallicWeight<Label, AW, GT> &gallic_weight,
-                      typename A::Weight *weight, typename A::Label *label) {
-    using GWT = StringWeight<Label, GallicStringType(GT)>;
-    const GWT &w1 = gallic_weight.Value1();
-    const AW &w2 = gallic_weight.Value2();
-    typename GWT::Iterator iter1(w1);
-    const Label l = w1.Size() == 1 ? iter1.Value() : 0;
-    if (l == kStringInfinity || l == kStringBad || w1.Size() > 1) return false;
-    *label = l;
-    *weight = w2;
-    return true;
-  }
-
-  static bool Extract(const GallicWeight<Label, AW, GALLIC> &gallic_weight,
-                      typename A::Weight *weight, typename A::Label *label) {
-    if (gallic_weight.Size() > 1) return false;
-    if (gallic_weight.Size() == 0) {
-      *label = 0;
-      *weight = A::Weight::Zero();
-      return true;
-    }
-    return Extract<GALLIC_RESTRICT>(gallic_weight.Back(), weight, label);
-  }
-
-  const Label superfinal_label_;
-  mutable bool error_;
-};
-
-// Mapper from GallicArc<A> to A.
-template <class A, GallicType G = GALLIC_LEFT>
-class GallicToNewSymbolsMapper {
- public:
-  using FromArc = GallicArc<A, G>;
-  using ToArc = A;
-
-  using Label = typename ToArc::Label;
-  using StateId = typename ToArc::StateId;
-  using AW = typename ToArc::Weight;
-  using GW = typename FromArc::Weight;
-  using SW = StringWeight<Label, GallicStringType(G)>;
-
-  explicit GallicToNewSymbolsMapper(MutableFst<ToArc> *fst)
-      : fst_(fst),
-        lmax_(0),
-        osymbols_(fst->OutputSymbols()),
-        isymbols_(nullptr),
-        error_(false) {
-    fst_->DeleteStates();
-    state_ = fst_->AddState();
-    fst_->SetStart(state_);
-    fst_->SetFinal(state_, AW::One());
-    if (osymbols_) {
-      string name = osymbols_->Name() + "_from_gallic";
-      fst_->SetInputSymbols(new SymbolTable(name));
-      isymbols_ = fst_->MutableInputSymbols();
-      const int64 zero = 0;
-      isymbols_->AddSymbol(osymbols_->Find(zero), 0);
-    } else {
-      fst_->SetInputSymbols(nullptr);
-    }
-  }
-
-  ToArc operator()(const FromArc &arc) {
-    // Super-non-final arc.
-    if (arc.nextstate == kNoStateId && arc.weight == GW::Zero()) {
-      return ToArc(arc.ilabel, 0, AW::Zero(), kNoStateId);
-    }
-    SW w1 = arc.weight.Value1();
-    AW w2 = arc.weight.Value2();
-    Label l;
-    if (w1.Size() == 0) {
-      l = 0;
-    } else {
-      auto insert_result = map_.insert(std::make_pair(w1, kNoLabel));
-      if (!insert_result.second) {
-        l = insert_result.first->second;
-      } else {
-        l = ++lmax_;
-        insert_result.first->second = l;
-        StringWeightIterator<SW> iter1(w1);
-        StateId n;
-        string s;
-        for (size_t i = 0, p = state_; i < w1.Size();
-             ++i, iter1.Next(), p = n) {
-          n = i == w1.Size() - 1 ? state_ : fst_->AddState();
-          fst_->AddArc(p, ToArc(i ? 0 : l, iter1.Value(), AW::One(), n));
-          if (isymbols_) {
-            if (i) s = s + "_";
-            s = s + osymbols_->Find(iter1.Value());
-          }
-        }
-        if (isymbols_) isymbols_->AddSymbol(s, l);
-      }
-    }
-    if (l == kStringInfinity || l == kStringBad || arc.ilabel != arc.olabel) {
-      FSTERROR() << "GallicToNewSymbolMapper: Unrepresentable weight: " << l;
-      error_ = true;
-    }
-    return ToArc(arc.ilabel, l, w2, arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_ALLOW_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 inprops) const {
-    uint64 outprops = inprops & kOLabelInvariantProperties &
-                      kWeightInvariantProperties & kAddSuperFinalProperties;
-    if (error_) outprops |= kError;
-    return outprops;
-  }
-
- private:
-  class StringKey {
-   public:
-    size_t operator()(const SW &x) const { return x.Hash(); }
-  };
-
-  using Map = std::unordered_map<SW, Label, StringKey>;
-
-  MutableFst<ToArc> *fst_;
-  Map map_;
-  Label lmax_;
-  StateId state_;
-  const SymbolTable *osymbols_;
-  SymbolTable *isymbols_;
-  mutable bool error_;
-};
-
-// TODO(kbg): Add common base class for those mappers which do nothing except
-// mutate their weights.
-
-// Mapper to add a constant to all weights.
-template <class A>
-class PlusMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-  using Weight = typename FromArc::Weight;
-
-  constexpr explicit PlusMapper(Weight weight) : weight_(std::move(weight)) {}
-
-  ToArc operator()(const FromArc &arc) const {
-    if (arc.weight == Weight::Zero()) return arc;
-    return ToArc(arc.ilabel, arc.olabel, Plus(arc.weight, weight_),
-                 arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr uint64 Properties(uint64 props) const {
-    return props & kWeightInvariantProperties;
-  }
-
- private:
-  const Weight weight_;
-};
-
-// Mapper to (right) multiply a constant to all weights.
-template <class A>
-class TimesMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-  using Weight = typename FromArc::Weight;
-
-  constexpr explicit TimesMapper(Weight weight) : weight_(std::move(weight)) {}
-
-  ToArc operator()(const FromArc &arc) const {
-    if (arc.weight == Weight::Zero()) return arc;
-    return ToArc(arc.ilabel, arc.olabel, Times(arc.weight, weight_),
-                 arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr uint64 Properties(uint64 props) const {
-    return props & kWeightInvariantProperties;
-  }
-
- private:
-  const Weight weight_;
-};
-
-// Mapper to take all weights to a constant power. The power argument is stored
-// as a double, so if there is a floating-point power implementation for this
-// weight type, it will take precedence. Otherwise, the power argument's 53 bits
-// of integer precision will be implicitly converted to a size_t and the default
-// power implementation (iterated multiplication) will be used instead.
-template <class A>
-class PowerMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-  using Weight = typename FromArc::Weight;
-
-  explicit PowerMapper(double power) : power_(power) {}
-
-  ToArc operator()(const FromArc &arc) const {
-    return ToArc(arc.ilabel, arc.olabel, Power(arc.weight, power_),
-                 arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr uint64 Properties(uint64 props) const {
-    return props & kWeightInvariantProperties;
-  }
-
- private:
-  const double power_;
-};
-
-// Mapper to reciprocate all non-Zero() weights.
-template <class A>
-class InvertWeightMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-  using Weight = typename FromArc::Weight;
-
-  ToArc operator()(const FromArc &arc) const {
-    if (arc.weight == Weight::Zero()) return arc;
-    return ToArc(arc.ilabel, arc.olabel, Divide(Weight::One(), arc.weight),
-                 arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr uint64 Properties(uint64 props) const {
-    return props & kWeightInvariantProperties;
-  }
-};
-
-// Mapper to map all non-Zero() weights to One().
-template <class A, class B = A>
-class RmWeightMapper {
- public:
-  using FromArc = A;
-  using ToArc = B;
-  using FromWeight = typename FromArc::Weight;
-  using ToWeight = typename ToArc::Weight;
-
-  ToArc operator()(const FromArc &arc) const {
-    return ToArc(arc.ilabel, arc.olabel,
-                 arc.weight != FromWeight::Zero() ?
-                 ToWeight::One() : ToWeight::Zero(),
-                 arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr uint64 Properties(uint64 props) const {
-    return (props & kWeightInvariantProperties) | kUnweighted;
-  }
-};
-
-// Mapper to quantize all weights.
-template <class A, class B = A>
-class QuantizeMapper {
- public:
-  using FromArc = A;
-  using ToArc = B;
-  using FromWeight = typename FromArc::Weight;
-  using ToWeight = typename ToArc::Weight;
-
-  QuantizeMapper() : delta_(kDelta) {}
-
-  explicit QuantizeMapper(float d) : delta_(d) {}
-
-  ToArc operator()(const FromArc &arc) const {
-    return ToArc(arc.ilabel, arc.olabel, arc.weight.Quantize(delta_),
-                 arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr uint64 Properties(uint64 props) const {
-    return props & kWeightInvariantProperties;
-  }
-
- private:
-  const float delta_;
-};
-
-// Mapper from A to B under the assumption:
-//
-//    B::Weight = A::Weight::ReverseWeight
-//    B::Label == A::Label
-//    B::StateId == A::StateId
-//
-// The weight is reversed, while the label and nextstate are preserved.
-template <class A, class B>
-class ReverseWeightMapper {
- public:
-  using FromArc = A;
-  using ToArc = B;
-
-  constexpr ToArc operator()(const FromArc &arc) const {
-    return ToArc(arc.ilabel, arc.olabel, arc.weight.Reverse(), arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr uint64 Properties(uint64 props) const { return props; }
-};
-
-}  // namespace fst
-
-#endif  // FST_ARC_MAP_H_
diff --git a/projects/llm_framework/include/fst/arc.h b/projects/llm_framework/include/fst/arc.h
deleted file mode 100644
index 651b11df..00000000
--- a/projects/llm_framework/include/fst/arc.h
+++ /dev/null
@@ -1,317 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Commonly used FST arc types.
-
-#ifndef FST_ARC_H_
-#define FST_ARC_H_
-
-#include <climits>
-#include <string>
-#include <type_traits>
-#include <utility>
-
-
-#include <fst/expectation-weight.h>
-#include <fst/float-weight.h>
-#include <fst/lexicographic-weight.h>
-#include <fst/power-weight.h>
-#include <fst/product-weight.h>
-#include <fst/signed-log-weight.h>
-#include <fst/sparse-power-weight.h>
-#include <fst/string-weight.h>
-
-
-namespace fst {
-
-template <class W>
-struct ArcTpl {
- public:
-  using Weight = W;
-  using Label = int;
-  using StateId = int;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  ArcTpl() noexcept(std::is_nothrow_default_constructible<Weight>::value) {}
-
-  template <class T>
-  ArcTpl(Label ilabel, Label olabel, T &&weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::forward<T>(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const auto *const type =
-        new string(Weight::Type() == "tropical" ? "standard" : Weight::Type());
-    return *type;
-  }
-};
-
-using StdArc = ArcTpl<TropicalWeight>;
-using LogArc = ArcTpl<LogWeight>;
-using Log64Arc = ArcTpl<Log64Weight>;
-using SignedLogArc = ArcTpl<SignedLogWeight>;
-using SignedLog64Arc = ArcTpl<SignedLog64Weight>;
-using MinMaxArc = ArcTpl<MinMaxWeight>;
-
-// Arc with integer labels and state IDs and string weights.
-template <StringType S = STRING_LEFT>
-struct StringArc {
- public:
-  using Label = int;
-  using Weight = StringWeight<int, S>;
-  using StateId = int;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  StringArc() = default;
-
-  template <class W>
-  StringArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::forward<W>(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const auto *const type = new string(
-        S == STRING_LEFT ? "left_standard_string"
-                         : (S == STRING_RIGHT ? "right_standard_string"
-                                              : "restricted_standard_string"));
-    return *type;
-  }
-};
-
-// Arc with label and state Id type the same as template arg and with
-// weights over the Gallic semiring w.r.t the output labels and weights of A.
-template <class A, GallicType G = GALLIC_LEFT>
-struct GallicArc {
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = GallicWeight<Label, typename Arc::Weight, G>;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  GallicArc() = default;
-
-  template <class W>
-  GallicArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::forward<W>(weight)),
-        nextstate(nextstate) {}
-
-  explicit GallicArc(const Arc &arc)
-      : ilabel(arc.ilabel), olabel(arc.ilabel), weight(arc.olabel, arc.weight),
-        nextstate(arc.nextstate) {}
-
-  static const string &Type() {
-    static const auto *const type = new string(
-        (G == GALLIC_LEFT
-             ? "left_gallic_"
-             : (G == GALLIC_RIGHT
-                    ? "right_gallic_"
-                    : (G == GALLIC_RESTRICT
-                           ? "restricted_gallic_"
-                           : (G == GALLIC_MIN ? "min_gallic_" : "gallic_")))) +
-        Arc::Type());
-    return *type;
-  }
-};
-
-// Arc with the reverse of the weight found in its template arg.
-template <class A>
-struct ReverseArc {
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using AWeight = typename Arc::Weight;
-  using Weight = typename AWeight::ReverseWeight;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  ReverseArc() = default;
-
-  template <class W>
-  ReverseArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::forward<W>(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const auto *const type = new string("reverse_" + Arc::Type());
-    return *type;
-  }
-};
-
-// Arc with integer labels and state IDs and lexicographic weights.
-template <class Weight1, class Weight2>
-struct LexicographicArc {
-  using Label = int;
-  using StateId = int;
-  using Weight = LexicographicWeight<Weight1, Weight2>;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  LexicographicArc() = default;
-
-  template <class W>
-  LexicographicArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::forward<W>(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const string *const type = new string(Weight::Type());
-    return *type;
-  }
-};
-
-// Arc with integer labels and state IDs and product weights.
-template <class Weight1, class Weight2>
-struct ProductArc {
-  using Label = int;
-  using StateId = int;
-  using Weight = ProductWeight<Weight1, Weight2>;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  ProductArc() = default;
-
-  template <class W>
-  ProductArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::forward<W>(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const auto *const type = new string(Weight::Type());
-    return *type;
-  }
-};
-
-// Arc with label and state ID type the same as first template argument and with
-// weights over the n-th Cartesian power of the weight type of the template
-// argument.
-template <class A, size_t n>
-struct PowerArc {
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = PowerWeight<typename Arc::Weight, n>;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  PowerArc() = default;
-
-  template <class W>
-  PowerArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::forward<W>(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const auto *const type =
-        new string(Arc::Type() + "_^" + std::to_string(n));
-    return *type;
-  }
-};
-
-// Arc with label and state ID type the same as first template argument and with
-// weights over the arbitrary Cartesian power of the weight type.
-template <class A, class K = int>
-struct SparsePowerArc {
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::Label;
-  using Weight = SparsePowerWeight<typename Arc::Weight, K>;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  SparsePowerArc() = default;
-
-  template <class W>
-  SparsePowerArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::forward<W>(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const string *const type = [] {
-      string type = Arc::Type() + "_^n";
-      if (sizeof(K) != sizeof(uint32)) {
-        type += "_" + std::to_string(CHAR_BIT * sizeof(K));
-      }
-      return new string(type);
-    }();
-    return *type;
-  }
-};
-
-// Arc with label and state ID type the same as first template argument and with
-// expectation weight over the first template argument's weight type and the
-// second template argument.
-template <class A, class X2>
-struct ExpectationArc {
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using X1 = typename Arc::Weight;
-  using Weight = ExpectationWeight<X1, X2>;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  ExpectationArc() = default;
-
-  template <class W>
-  ExpectationArc(Label ilabel, Label olabel, W &&weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::forward<W>(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const auto *const type =
-        new string("expectation_" + Arc::Type() + "_" + X2::Type());
-    return *type;
-  }
-};
-
-}  // namespace fst
-
-#endif  // FST_ARC_H_
diff --git a/projects/llm_framework/include/fst/arcfilter.h b/projects/llm_framework/include/fst/arcfilter.h
deleted file mode 100644
index 598e543f..00000000
--- a/projects/llm_framework/include/fst/arcfilter.h
+++ /dev/null
@@ -1,93 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function objects to restrict which arcs are traversed in an FST.
-
-#ifndef FST_ARCFILTER_H_
-#define FST_ARCFILTER_H_
-
-
-#include <fst/fst.h>
-#include <fst/util.h>
-
-
-namespace fst {
-
-// True for all arcs.
-template <class Arc>
-class AnyArcFilter {
- public:
-  bool operator()(const Arc &arc) const { return true; }
-};
-
-// True for (input/output) epsilon arcs.
-template <class Arc>
-class EpsilonArcFilter {
- public:
-  bool operator()(const Arc &arc) const {
-    return arc.ilabel == 0 && arc.olabel == 0;
-  }
-};
-
-// True for input epsilon arcs.
-template <class Arc>
-class InputEpsilonArcFilter {
- public:
-  bool operator()(const Arc &arc) const { return arc.ilabel == 0; }
-};
-
-// True for output epsilon arcs.
-template <class Arc>
-class OutputEpsilonArcFilter {
- public:
-  bool operator()(const Arc &arc) const { return arc.olabel == 0; }
-};
-
-// True if specified label matches (doesn't match) when keep_match is
-// true (false).
-template <class Arc>
-class LabelArcFilter {
- public:
-  using Label = typename Arc::Label;
-
-  explicit LabelArcFilter(Label label, bool match_input = true,
-                          bool keep_match = true)
-      : label_(label), match_input_(match_input), keep_match_(keep_match) {}
-
-  bool operator()(const Arc &arc) const {
-    const bool match = (match_input_ ? arc.ilabel : arc.olabel) == label_;
-    return keep_match_ ? match : !match;
-  }
-
- private:
-  const Label label_;
-  const bool match_input_;
-  const bool keep_match_;
-};
-
-// True if specified labels match (don't match) when keep_match is true (false).
-template <class Arc>
-class MultiLabelArcFilter {
- public:
-  using Label = typename Arc::Label;
-
-  explicit MultiLabelArcFilter(bool match_input = true, bool keep_match = true)
-      : match_input_(match_input), keep_match_(keep_match) {}
-
-  bool operator()(const Arc &arc) const {
-    const Label label = match_input_ ? arc.ilabel : arc.olabel;
-    const bool match = labels_.Find(label) != labels_.End();
-    return keep_match_ ? match : !match;
-  }
-
-  void AddLabel(Label label) { labels_.Insert(label); }
-
- private:
-  CompactSet<Label, kNoLabel> labels_;
-  const bool match_input_;
-  const bool keep_match_;
-};
-
-}  // namespace fst
-
-#endif  // FST_ARCFILTER_H_
diff --git a/projects/llm_framework/include/fst/arcsort.h b/projects/llm_framework/include/fst/arcsort.h
deleted file mode 100644
index b5ab50e0..00000000
--- a/projects/llm_framework/include/fst/arcsort.h
+++ /dev/null
@@ -1,211 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to sort arcs in an FST.
-
-#ifndef FST_ARCSORT_H_
-#define FST_ARCSORT_H_
-
-#include <algorithm>
-#include <string>
-#include <vector>
-
-#include <fst/cache.h>
-#include <fst/state-map.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-template <class Arc, class Compare>
-class ArcSortMapper {
- public:
-  using FromArc = Arc;
-  using ToArc = Arc;
-
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  constexpr ArcSortMapper(const Fst<Arc> &fst, const Compare &comp)
-      : fst_(fst), comp_(comp), i_(0) {}
-
-  // Allows updating Fst argument; pass only if changed.
-  ArcSortMapper(const ArcSortMapper<Arc, Compare> &mapper,
-                const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? *fst : mapper.fst_), comp_(mapper.comp_), i_(0) {}
-
-  StateId Start() { return fst_.Start(); }
-
-  Weight Final(StateId s) const { return fst_.Final(s); }
-
-  void SetState(StateId s) {
-    i_ = 0;
-    arcs_.clear();
-    arcs_.reserve(fst_.NumArcs(s));
-    for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
-      arcs_.push_back(aiter.Value());
-    }
-    std::sort(arcs_.begin(), arcs_.end(), comp_);
-  }
-
-  bool Done() const { return i_ >= arcs_.size(); }
-
-  const Arc &Value() const { return arcs_[i_]; }
-
-  void Next() { ++i_; }
-
-  MapSymbolsAction InputSymbolsAction() const { return MAP_COPY_SYMBOLS; }
-
-  MapSymbolsAction OutputSymbolsAction() const { return MAP_COPY_SYMBOLS; }
-
-  uint64 Properties(uint64 props) const { return comp_.Properties(props); }
-
- private:
-  const Fst<Arc> &fst_;
-  const Compare &comp_;
-  std::vector<Arc> arcs_;
-  ssize_t i_;  // current arc position
-
-  ArcSortMapper &operator=(const ArcSortMapper &) = delete;
-};
-
-// Sorts the arcs in an FST according to function object 'comp' of type Compare.
-// This version modifies its input. Comparison function objects ILabelCompare
-// and OLabelCompare are provided by the library. In general, Compare must meet
-// the requirements for a  comparison function object (e.g., similar to those
-// used by std::sort). It must also have a member Properties(uint64) that
-// specifies the known properties of the sorted FST; it takes as argument the
-// input FST's known properties before the sort.
-//
-// Complexity:
-//
-// - Time: O(v d log d)
-// - Space: O(d)
-//
-// where v = # of states and d = maximum out-degree.
-template <class Arc, class Compare>
-void ArcSort(MutableFst<Arc> *fst, Compare comp) {
-  ArcSortMapper<Arc, Compare> mapper(*fst, comp);
-  StateMap(fst, mapper);
-}
-
-using ArcSortFstOptions = CacheOptions;
-
-// Sorts the arcs in an FST according to function object 'comp' of type Compare.
-// This version is a delayed FST. Comparsion function objects ILabelCompare and
-// OLabelCompare are provided by the library. In general, Compare must meet the
-// requirements for a comparision function object (e.g., similar to those
-// used by std::sort). It must also have a member Properties(uint64) that
-// specifies the known properties of the sorted FST; it takes as argument the
-// input FST's known properties.
-//
-// Complexity:
-//
-// - Time: O(v d log d)
-// - Space: O(d)
-//
-// where v = # of states visited, d = maximum out-degree of states visited.
-// Constant time and space to visit an input state is assumed and exclusive of
-// caching.
-template <class Arc, class Compare>
-class ArcSortFst : public StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>> {
-  using StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>>::GetImpl;
-
- public:
-  using StateId = typename Arc::StateId;
-  using Mapper = ArcSortMapper<Arc, Compare>;
-
-  ArcSortFst(const Fst<Arc> &fst, const Compare &comp)
-      : StateMapFst<Arc, Arc, Mapper>(fst,
-                                      ArcSortMapper<Arc, Compare>(fst, comp)) {}
-
-  ArcSortFst(const Fst<Arc> &fst, const Compare &comp,
-             const ArcSortFstOptions &opts)
-      : StateMapFst<Arc, Arc, Mapper>(fst, Mapper(fst, comp), opts) {}
-
-  // See Fst<>::Copy() for doc.
-  ArcSortFst(const ArcSortFst<Arc, Compare> &fst, bool safe = false)
-      : StateMapFst<Arc, Arc, Mapper>(fst, safe) {}
-
-  // Gets a copy of this ArcSortFst. See Fst<>::Copy() for further doc.
-  ArcSortFst<Arc, Compare> *Copy(bool safe = false) const override {
-    return new ArcSortFst(*this, safe);
-  }
-
-  size_t NumArcs(StateId s) const override {
-    return GetImpl()->GetFst()->NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) const override {
-    return GetImpl()->GetFst()->NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) const override {
-    return GetImpl()->GetFst()->NumOutputEpsilons(s);
-  }
-};
-
-// Specialization for ArcSortFst.
-template <class Arc, class Compare>
-class StateIterator<ArcSortFst<Arc, Compare>>
-    : public StateIterator<StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>>> {
- public:
-  explicit StateIterator(const ArcSortFst<Arc, Compare> &fst)
-      : StateIterator<StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>>>(fst) {
-  }
-};
-
-// Specialization for ArcSortFst.
-template <class Arc, class Compare>
-class ArcIterator<ArcSortFst<Arc, Compare>>
-    : public ArcIterator<StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>>> {
- public:
-  ArcIterator(const ArcSortFst<Arc, Compare> &fst, typename Arc::StateId s)
-      : ArcIterator<StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>>>(fst,
-                                                                        s) {}
-};
-
-// Compare class for comparing input labels of arcs.
-template <class Arc>
-class ILabelCompare {
- public:
-  constexpr ILabelCompare() {}
-
-  constexpr bool operator()(const Arc &arc1, const Arc &arc2) const {
-    return arc1.ilabel < arc2.ilabel;
-  }
-
-  constexpr uint64 Properties(uint64 props) const {
-    return (props & kArcSortProperties) | kILabelSorted |
-           (props & kAcceptor ? kOLabelSorted : 0);
-  }
-};
-
-// Compare class for comparing output labels of arcs.
-template <class Arc>
-class OLabelCompare {
- public:
-  constexpr OLabelCompare() {}
-
-  constexpr bool operator()(const Arc &arc1, const Arc &arc2) const {
-    return arc1.olabel < arc2.olabel;
-  }
-
-  constexpr uint64 Properties(uint64 props) const {
-    return (props & kArcSortProperties) | kOLabelSorted |
-           (props & kAcceptor ? kILabelSorted : 0);
-  }
-};
-
-// Useful aliases when using StdArc.
-
-template <class Compare>
-using StdArcSortFst = ArcSortFst<StdArc, Compare>;
-
-using StdILabelCompare = ILabelCompare<StdArc>;
-
-using StdOLabelCompare = OLabelCompare<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_ARCSORT_H_
diff --git a/projects/llm_framework/include/fst/bi-table.h b/projects/llm_framework/include/fst/bi-table.h
deleted file mode 100644
index 9651cfe8..00000000
--- a/projects/llm_framework/include/fst/bi-table.h
+++ /dev/null
@@ -1,480 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes for representing a bijective mapping between an arbitrary entry
-// of type T and a signed integral ID.
-
-#ifndef FST_BI_TABLE_H_
-#define FST_BI_TABLE_H_
-
-#include <deque>
-#include <functional>
-#include <memory>
-#include <type_traits>
-#include <unordered_map>
-#include <unordered_set>
-#include <vector>
-
-#include <fst/log.h>
-#include <fst/memory.h>
-#include <unordered_set>
-
-namespace fst {
-
-// Bitables model bijective mappings between entries of an arbitrary type T and
-// an signed integral ID of type I. The IDs are allocated starting from 0 in
-// order.
-//
-// template <class I, class T>
-// class BiTable {
-//  public:
-//
-//   // Required constructors.
-//   BiTable();
-//
-//   // Looks up integer ID from entry. If it doesn't exist and insert
-//   / is true, adds it; otherwise, returns -1.
-//   I FindId(const T &entry, bool insert = true);
-//
-//   // Looks up entry from integer ID.
-//   const T &FindEntry(I) const;
-//
-//   // Returns number of stored entries.
-//   I Size() const;
-// };
-
-// An implementation using a hash map for the entry to ID mapping. H is the
-// hash function and E is the equality function. If passed to the constructor,
-// ownership is given to this class.
-template <class I, class T, class H, class E = std::equal_to<T>>
-class HashBiTable {
- public:
-  // Reserves space for table_size elements. If passing H and E to the
-  // constructor, this class owns them.
-  explicit HashBiTable(size_t table_size = 0, H *h = nullptr, E *e = nullptr) :
-      hash_func_(h ? h : new H()), hash_equal_(e ? e : new E()),
-      entry2id_(table_size, *hash_func_, *hash_equal_) {
-    if (table_size) id2entry_.reserve(table_size);
-  }
-
-  HashBiTable(const HashBiTable<I, T, H, E> &table)
-      : hash_func_(new H(*table.hash_func_)),
-        hash_equal_(new E(*table.hash_equal_)),
-        entry2id_(table.entry2id_.begin(), table.entry2id_.end(),
-                  table.entry2id_.size(), *hash_func_, *hash_equal_),
-        id2entry_(table.id2entry_) {}
-
-  I FindId(const T &entry, bool insert = true) {
-    if (!insert) {
-      const auto it = entry2id_.find(entry);
-      return it == entry2id_.end() ? -1 : it->second - 1;
-    }
-    I &id_ref = entry2id_[entry];
-    if (id_ref == 0) {  // T not found; stores and assigns a new ID.
-      id2entry_.push_back(entry);
-      id_ref = id2entry_.size();
-    }
-    return id_ref - 1;  // NB: id_ref = ID + 1.
-  }
-
-  const T &FindEntry(I s) const { return id2entry_[s]; }
-
-  I Size() const { return id2entry_.size(); }
-
-  // TODO(riley): Add fancy clear-to-size, as in CompactHashBiTable.
-  void Clear() {
-    entry2id_.clear();
-    id2entry_.clear();
-  }
-
- private:
-  std::unique_ptr<H> hash_func_;
-  std::unique_ptr<E> hash_equal_;
-  std::unordered_map<T, I, H, E> entry2id_;
-  std::vector<T> id2entry_;
-};
-
-// Enables alternative hash set representations below.
-enum HSType { HS_STL = 0, HS_DENSE = 1, HS_SPARSE = 2, HS_FLAT = 3 };
-
-// Default hash set is STL hash_set.
-template <class K, class H, class E, HSType HS>
-struct HashSet : public std::unordered_set<K, H, E, PoolAllocator<K>> {
-  explicit HashSet(size_t n = 0, const H &h = H(), const E &e = E())
-      : std::unordered_set<K, H, E, PoolAllocator<K>>(n, h, e) {}
-
-  void rehash(size_t n) {}
-};
-
-// An implementation using a hash set for the entry to ID mapping. The hash set
-// holds keys which are either the ID or kCurrentKey. These keys can be mapped
-// to entries either by looking up in the entry vector or, if kCurrentKey, in
-// current_entry_. The hash and key equality functions map to entries first. H
-// is the hash function and E is the equality function. If passed to the
-// constructor, ownership is given to this class.
-// TODO(rybach): remove support for (deprecated and unused) HS_DENSE, HS_SPARSE.
-template <class I, class T, class H, class E = std::equal_to<T>,
-          HSType HS = HS_FLAT>
-class CompactHashBiTable {
- public:
-  friend class HashFunc;
-  friend class HashEqual;
-
-  // Reserves space for table_size elements. If passing H and E to the
-  // constructor, this class owns them.
-  explicit CompactHashBiTable(size_t table_size = 0, H *h = nullptr,
-                              E *e = nullptr) :
-        hash_func_(h ? h : new H()), hash_equal_(e ? e : new E()),
-        compact_hash_func_(*this), compact_hash_equal_(*this),
-        keys_(table_size, compact_hash_func_, compact_hash_equal_) {
-    if (table_size) id2entry_.reserve(table_size);
-  }
-
-  CompactHashBiTable(const CompactHashBiTable<I, T, H, E, HS> &table)
-      : hash_func_(new H(*table.hash_func_)),
-        hash_equal_(new E(*table.hash_equal_)),
-        compact_hash_func_(*this), compact_hash_equal_(*this),
-        keys_(table.keys_.size(), compact_hash_func_, compact_hash_equal_),
-        id2entry_(table.id2entry_) {
-    keys_.insert(table.keys_.begin(), table.keys_.end());
-  }
-
-  I FindId(const T &entry, bool insert = true) {
-    current_entry_ = &entry;
-    if (insert) {
-      auto result = keys_.insert(kCurrentKey);
-      if (!result.second) return *result.first;  // Already exists.
-      // Overwrites kCurrentKey with a new key value; this is safe because it
-      // doesn't affect hashing or equality testing.
-      I key = id2entry_.size();
-      const_cast<I &>(*result.first) = key;
-      id2entry_.push_back(entry);
-      return key;
-    }
-    const auto it = keys_.find(kCurrentKey);
-    return it == keys_.end() ? -1 : *it;
-  }
-
-  const T &FindEntry(I s) const { return id2entry_[s]; }
-
-  I Size() const { return id2entry_.size(); }
-
-  // Clears content; with argument, erases last n IDs.
-  void Clear(ssize_t n = -1) {
-    if (n < 0 || n >= id2entry_.size()) {  // Clears completely.
-      keys_.clear();
-      id2entry_.clear();
-    } else if (n == id2entry_.size() - 1) {  // Leaves only key 0.
-      const T entry = FindEntry(0);
-      keys_.clear();
-      id2entry_.clear();
-      FindId(entry, true);
-    } else {
-      while (n-- > 0) {
-        I key = id2entry_.size() - 1;
-        keys_.erase(key);
-        id2entry_.pop_back();
-      }
-      keys_.rehash(0);
-    }
-  }
-
- private:
-  static_assert(std::is_signed<I>::value, "I must be a signed type");
-  // ... otherwise >= kCurrentKey comparisons as used below don't work.
-  // TODO(rybach): (1) remove kEmptyKey, kDeletedKey, (2) don't use >= for key
-  // comparison, (3) allow unsigned key types.
-  static constexpr I kCurrentKey = -1;
-  static constexpr I kEmptyKey = -2;
-  static constexpr I kDeletedKey = -3;
-
-  class HashFunc {
-   public:
-    explicit HashFunc(const CompactHashBiTable &ht) : ht_(&ht) {}
-
-    size_t operator()(I k) const {
-      if (k >= kCurrentKey) {
-        return (*ht_->hash_func_)(ht_->Key2Entry(k));
-      } else {
-        return 0;
-      }
-    }
-
-   private:
-    const CompactHashBiTable *ht_;
-  };
-
-  class HashEqual {
-   public:
-    explicit HashEqual(const CompactHashBiTable &ht) : ht_(&ht) {}
-
-    bool operator()(I k1, I k2) const {
-      if (k1 == k2) {
-        return true;
-      } else if (k1 >= kCurrentKey && k2 >= kCurrentKey) {
-        return (*ht_->hash_equal_)(ht_->Key2Entry(k1), ht_->Key2Entry(k2));
-      } else {
-        return false;
-      }
-    }
-
-   private:
-    const CompactHashBiTable *ht_;
-  };
-
-  using KeyHashSet = HashSet<I, HashFunc, HashEqual, HS>;
-
-  const T &Key2Entry(I k) const {
-    if (k == kCurrentKey) {
-      return *current_entry_;
-    } else {
-      return id2entry_[k];
-    }
-  }
-
-  std::unique_ptr<H> hash_func_;
-  std::unique_ptr<E> hash_equal_;
-  HashFunc compact_hash_func_;
-  HashEqual compact_hash_equal_;
-  KeyHashSet keys_;
-  std::vector<T> id2entry_;
-  const T *current_entry_;
-};
-
-template <class I, class T, class H, class E, HSType HS>
-constexpr I CompactHashBiTable<I, T, H, E, HS>::kCurrentKey;
-
-template <class I, class T, class H, class E, HSType HS>
-constexpr I CompactHashBiTable<I, T, H, E, HS>::kEmptyKey;
-
-template <class I, class T, class H, class E, HSType HS>
-constexpr I CompactHashBiTable<I, T, H, E, HS>::kDeletedKey;
-
-// An implementation using a vector for the entry to ID mapping. It is passed a
-// function object FP that should fingerprint entries uniquely to an integer
-// that can used as a vector index. Normally, VectorBiTable constructs the FP
-// object. The user can instead pass in this object; in that case, VectorBiTable
-// takes its ownership.
-template <class I, class T, class FP>
-class VectorBiTable {
- public:
-  // Reserves table_size cells of space. If passing FP argument to the
-  // constructor, this class owns it.
-  explicit VectorBiTable(FP *fp = nullptr, size_t table_size = 0) :
-      fp_(fp ? fp : new FP()) {
-    if (table_size) id2entry_.reserve(table_size);
-  }
-
-  VectorBiTable(const VectorBiTable<I, T, FP> &table)
-      : fp_(new FP(*table.fp_)), fp2id_(table.fp2id_),
-        id2entry_(table.id2entry_) {}
-
-  I FindId(const T &entry, bool insert = true) {
-    ssize_t fp = (*fp_)(entry);
-    if (fp >= fp2id_.size()) fp2id_.resize(fp + 1);
-    I &id_ref = fp2id_[fp];
-    if (id_ref == 0) {  // T not found.
-      if (insert) {     // Stores and assigns a new ID.
-        id2entry_.push_back(entry);
-        id_ref = id2entry_.size();
-      } else {
-        return -1;
-      }
-    }
-    return id_ref - 1;  // NB: id_ref = ID + 1.
-  }
-
-  const T &FindEntry(I s) const { return id2entry_[s]; }
-
-  I Size() const { return id2entry_.size(); }
-
-  const FP &Fingerprint() const { return *fp_; }
-
- private:
-  std::unique_ptr<FP> fp_;
-  std::vector<I> fp2id_;
-  std::vector<T> id2entry_;
-};
-
-// An implementation using a vector and a compact hash table. The selecting
-// functor S returns true for entries to be hashed in the vector. The
-// fingerprinting functor FP returns a unique fingerprint for each entry to be
-// hashed in the vector (these need to be suitable for indexing in a vector).
-// The hash functor H is used when hashing entry into the compact hash table.
-// If passed to the constructor, ownership is given to this class.
-template <class I, class T, class S, class FP, class H, HSType HS = HS_DENSE>
-class VectorHashBiTable {
- public:
-  friend class HashFunc;
-  friend class HashEqual;
-
-  explicit VectorHashBiTable(S *s, FP *fp, H *h, size_t vector_size = 0,
-                             size_t entry_size = 0)
-      : selector_(s), fp_(fp), h_(h), hash_func_(*this), hash_equal_(*this),
-        keys_(0, hash_func_, hash_equal_) {
-    if (vector_size) fp2id_.reserve(vector_size);
-    if (entry_size) id2entry_.reserve(entry_size);
-  }
-
-  VectorHashBiTable(const VectorHashBiTable<I, T, S, FP, H, HS> &table)
-      : selector_(new S(table.s_)), fp_(new FP(*table.fp_)),
-        h_(new H(*table.h_)), id2entry_(table.id2entry_),
-        fp2id_(table.fp2id_), hash_func_(*this), hash_equal_(*this),
-        keys_(table.keys_.size(), hash_func_, hash_equal_) {
-    keys_.insert(table.keys_.begin(), table.keys_.end());
-  }
-
-  I FindId(const T &entry, bool insert = true) {
-    if ((*selector_)(entry)) {  // Uses the vector if selector_(entry) == true.
-      uint64 fp = (*fp_)(entry);
-      if (fp2id_.size() <= fp) fp2id_.resize(fp + 1, 0);
-      if (fp2id_[fp] == 0) {  // T not found.
-        if (insert) {         // Stores and assigns a new ID.
-          id2entry_.push_back(entry);
-          fp2id_[fp] = id2entry_.size();
-        } else {
-          return -1;
-        }
-      }
-      return fp2id_[fp] - 1;  // NB: assoc_value = ID + 1.
-    } else {                  // Uses the hash table otherwise.
-      current_entry_ = &entry;
-      const auto it = keys_.find(kCurrentKey);
-      if (it == keys_.end()) {
-        if (insert) {
-          I key = id2entry_.size();
-          id2entry_.push_back(entry);
-          keys_.insert(key);
-          return key;
-        } else {
-          return -1;
-        }
-      } else {
-        return *it;
-      }
-    }
-  }
-
-  const T &FindEntry(I s) const { return id2entry_[s]; }
-
-  I Size() const { return id2entry_.size(); }
-
-  const S &Selector() const { return *selector_; }
-
-  const FP &Fingerprint() const { return *fp_; }
-
-  const H &Hash() const { return *h_; }
-
- private:
-  static constexpr I kCurrentKey = -1;
-  static constexpr I kEmptyKey = -2;
-
-  class HashFunc {
-   public:
-    explicit HashFunc(const VectorHashBiTable &ht) : ht_(&ht) {}
-
-    size_t operator()(I k) const {
-      if (k >= kCurrentKey) {
-        return (*(ht_->h_))(ht_->Key2Entry(k));
-      } else {
-        return 0;
-      }
-    }
-
-   private:
-    const VectorHashBiTable *ht_;
-  };
-
-  class HashEqual {
-   public:
-    explicit HashEqual(const VectorHashBiTable &ht) : ht_(&ht) {}
-
-    bool operator()(I k1, I k2) const {
-      if (k1 >= kCurrentKey && k2 >= kCurrentKey) {
-        return ht_->Key2Entry(k1) == ht_->Key2Entry(k2);
-      } else {
-        return k1 == k2;
-      }
-    }
-
-   private:
-    const VectorHashBiTable *ht_;
-  };
-
-  using KeyHashSet = HashSet<I, HashFunc, HashEqual, HS>;
-
-  const T &Key2Entry(I k) const {
-    if (k == kCurrentKey) {
-      return *current_entry_;
-    } else {
-      return id2entry_[k];
-    }
-  }
-
-  std::unique_ptr<S> selector_;  // True if entry hashed into vector.
-  std::unique_ptr<FP> fp_;       // Fingerprint used for hashing into vector.
-  std::unique_ptr<H> h_;         // Hash funcion used for hashing into hash_set.
-
-  std::vector<T> id2entry_;  // Maps state IDs to entry.
-  std::vector<I> fp2id_;     // Maps entry fingerprints to IDs.
-
-  // Compact implementation of the hash table mapping entries to state IDs
-  // using the hash function h_.
-  HashFunc hash_func_;
-  HashEqual hash_equal_;
-  KeyHashSet keys_;
-  const T *current_entry_;
-};
-
-template <class I, class T, class S, class FP, class H, HSType HS>
-constexpr I VectorHashBiTable<I, T, S, FP, H, HS>::kCurrentKey;
-
-template <class I, class T, class S, class FP, class H, HSType HS>
-constexpr I VectorHashBiTable<I, T, S, FP, H, HS>::kEmptyKey;
-
-// An implementation using a hash map for the entry to ID mapping. This version
-// permits erasing of arbitrary states. The entry T must have == defined and
-// its default constructor must produce a entry that will never be seen. F is
-// the hash function.
-template <class I, class T, class F>
-class ErasableBiTable {
- public:
-  ErasableBiTable() : first_(0) {}
-
-  I FindId(const T &entry, bool insert = true) {
-    I &id_ref = entry2id_[entry];
-    if (id_ref == 0) {  // T not found.
-      if (insert) {     // Stores and assigns a new ID.
-        id2entry_.push_back(entry);
-        id_ref = id2entry_.size() + first_;
-      } else {
-        return -1;
-      }
-    }
-    return id_ref - 1;  // NB: id_ref = ID + 1.
-  }
-
-  const T &FindEntry(I s) const { return id2entry_[s - first_]; }
-
-  I Size() const { return id2entry_.size(); }
-
-  void Erase(I s) {
-    auto &ref = id2entry_[s - first_];
-    entry2id_.erase(ref);
-    ref = empty_entry_;
-    while (!id2entry_.empty() && id2entry_.front() == empty_entry_) {
-      id2entry_.pop_front();
-      ++first_;
-    }
-  }
-
- private:
-  std::unordered_map<T, I, F> entry2id_;
-  std::deque<T> id2entry_;
-  const T empty_entry_;
-  I first_;  // I of first element in the deque.
-};
-
-}  // namespace fst
-
-#endif  // FST_BI_TABLE_H_
diff --git a/projects/llm_framework/include/fst/cache.h b/projects/llm_framework/include/fst/cache.h
deleted file mode 100644
index 13b7cf81..00000000
--- a/projects/llm_framework/include/fst/cache.h
+++ /dev/null
@@ -1,1327 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// An FST implementation that caches FST elements of a delayed computation.
-
-#ifndef FST_CACHE_H_
-#define FST_CACHE_H_
-
-#include <algorithm>
-#include <functional>
-#include <list>
-#include <vector>
-
-#include <fst/flags.h>
-#include <fst/log.h>
-
-#include <fst/vector-fst.h>
-
-#include <unordered_map>
-
-DECLARE_bool(fst_default_cache_gc);
-DECLARE_int64(fst_default_cache_gc_limit);
-
-namespace fst {
-
-// Options for controlling caching behavior; higher level than CacheImplOptions.
-struct CacheOptions {
-  bool gc;          // Enables GC.
-  size_t gc_limit;  // Number of bytes allowed before GC.
-
-  explicit CacheOptions(bool gc = FLAGS_fst_default_cache_gc,
-                        size_t gc_limit = FLAGS_fst_default_cache_gc_limit)
-      : gc(gc), gc_limit(gc_limit) {}
-};
-
-// Options for controlling caching behavior, at a lower level than
-// CacheOptions; templated on the cache store and allows passing the store.
-template <class CacheStore>
-struct CacheImplOptions {
-  bool gc;            // Enables GC.
-  size_t gc_limit;    // Number of bytes allowed before GC.
-  CacheStore *store;  // Cache store.
-  bool own_store;     // Should CacheImpl takes ownership of the store?
-
-  explicit CacheImplOptions(bool gc = FLAGS_fst_default_cache_gc,
-                            size_t gc_limit = FLAGS_fst_default_cache_gc_limit,
-                            CacheStore *store = nullptr)
-      : gc(gc), gc_limit(gc_limit), store(store), own_store(true) {}
-
-  explicit CacheImplOptions(const CacheOptions &opts)
-      : gc(opts.gc), gc_limit(opts.gc_limit), store(nullptr), own_store(true) {}
-};
-
-// Cache flags.
-constexpr uint32 kCacheFinal = 0x0001;   // Final weight has been cached.
-constexpr uint32 kCacheArcs = 0x0002;    // Arcs have been cached.
-constexpr uint32 kCacheInit = 0x0004;    // Initialized by GC.
-constexpr uint32 kCacheRecent = 0x0008;  // Visited since GC.
-constexpr uint32 kCacheFlags =
-    kCacheFinal | kCacheArcs | kCacheInit | kCacheRecent;
-
-// Cache state, with arcs stored in a per-state std::vector.
-template <class A, class M = PoolAllocator<A>>
-class CacheState {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using ArcAllocator = M;
-  using StateAllocator =
-      typename ArcAllocator::template rebind<CacheState<A, M>>::other;
-
-  // Provides STL allocator for arcs.
-  explicit CacheState(const ArcAllocator &alloc)
-      : final_(Weight::Zero()),
-        niepsilons_(0),
-        noepsilons_(0),
-        arcs_(alloc),
-        flags_(0),
-        ref_count_(0) {}
-
-  CacheState(const CacheState<A> &state, const ArcAllocator &alloc)
-      : final_(state.Final()),
-        niepsilons_(state.NumInputEpsilons()),
-        noepsilons_(state.NumOutputEpsilons()),
-        arcs_(state.arcs_.begin(), state.arcs_.end(), alloc),
-        flags_(state.Flags()),
-        ref_count_(0) {}
-
-  void Reset() {
-    final_ = Weight::Zero();
-    niepsilons_ = 0;
-    noepsilons_ = 0;
-    ref_count_ = 0;
-    flags_ = 0;
-    arcs_.clear();
-  }
-
-  Weight Final() const { return final_; }
-
-  size_t NumInputEpsilons() const { return niepsilons_; }
-
-  size_t NumOutputEpsilons() const { return noepsilons_; }
-
-  size_t NumArcs() const { return arcs_.size(); }
-
-  const Arc &GetArc(size_t n) const { return arcs_[n]; }
-
-  // Used by the ArcIterator<Fst<Arc>> efficient implementation.
-  const Arc *Arcs() const { return !arcs_.empty() ? &arcs_[0] : nullptr; }
-
-  // Accesses flags; used by the caller.
-  uint32 Flags() const { return flags_; }
-
-  // Accesses ref count; used by the caller.
-  int RefCount() const { return ref_count_; }
-
-  void SetFinal(Weight weight) { final_ = std::move(weight); }
-
-  void ReserveArcs(size_t n) { arcs_.reserve(n); }
-
-  // Adds one arc at a time with all needed book-keeping; use PushArc and
-  // SetArcs for a more efficient alternative.
-  void AddArc(const Arc &arc) {
-    IncrementNumEpsilons(arc);
-    arcs_.push_back(arc);
-  }
-
-  void AddArc(Arc &&arc) {
-    IncrementNumEpsilons(arc);
-    arcs_.push_back(std::move(arc));
-  }
-
-  // Adds one arc at a time with delayed book-keeping; finalize with SetArcs().
-  void PushArc(const Arc &arc) { arcs_.push_back(arc); }
-
-  void PushArc(Arc &&arc) { arcs_.push_back(std::move(arc)); }
-
-  // Adds one arc at a time with delayed book-keeping; finalize with SetArcs().
-  template <class... T>
-  void EmplaceArc(T &&... ctor_args) {
-    arcs_.emplace_back(std::forward<T>(ctor_args)...);
-  }
-
-  // Finalizes arcs book-keeping; call only once.
-  void SetArcs() {
-    for (const auto &arc : arcs_) {
-      IncrementNumEpsilons(arc);
-    }
-  }
-
-  // Modifies nth arc.
-  void SetArc(const Arc &arc, size_t n) {
-    if (arcs_[n].ilabel == 0) --niepsilons_;
-    if (arcs_[n].olabel == 0) --noepsilons_;
-    IncrementNumEpsilons(arc);
-    arcs_[n] = arc;
-  }
-
-  // Deletes all arcs.
-  void DeleteArcs() {
-    niepsilons_ = 0;
-    noepsilons_ = 0;
-    arcs_.clear();
-  }
-
-  void DeleteArcs(size_t n) {
-    for (size_t i = 0; i < n; ++i) {
-      if (arcs_.back().ilabel == 0) --niepsilons_;
-      if (arcs_.back().olabel == 0) --noepsilons_;
-      arcs_.pop_back();
-    }
-  }
-
-  // Sets status flags; used by the caller.
-  void SetFlags(uint32 flags, uint32 mask) const {
-    flags_ &= ~mask;
-    flags_ |= flags;
-  }
-
-  // Mutates reference counts; used by the caller.
-
-  int IncrRefCount() const { return ++ref_count_; }
-
-  int DecrRefCount() const { return --ref_count_; }
-
-  // Used by the ArcIterator<Fst<Arc>> efficient implementation.
-  int *MutableRefCount() const { return &ref_count_; }
-
-  // Used for state class allocation.
-  void *operator new(size_t size, StateAllocator *alloc) {
-    return alloc->allocate(1);
-  }
-
-  // For state destruction and memory freeing.
-  static void Destroy(CacheState<Arc> *state, StateAllocator *alloc) {
-    if (state) {
-      state->~CacheState<Arc>();
-      alloc->deallocate(state, 1);
-    }
-  }
-
- private:
-  // Update the number of epsilons as a result of having added an arc.
-  void IncrementNumEpsilons(const Arc &arc) {
-    if (arc.ilabel == 0) ++niepsilons_;
-    if (arc.olabel == 0) ++noepsilons_;
-  }
-
-  Weight final_;                         // Final weight.
-  size_t niepsilons_;                    // # of input epsilons.
-  size_t noepsilons_;                    // # of output epsilons.
-  std::vector<Arc, ArcAllocator> arcs_;  // Arcs representation.
-  mutable uint32 flags_;
-  mutable int ref_count_;  // If 0, available for GC.
-};
-
-// Cache store, allocating and storing states, providing a mapping from state
-// IDs to cached states, and an iterator over these states. The state template
-// argument must implement the CacheState interface. The state for a StateId s
-// is constructed when requested by GetMutableState(s) if it is not yet stored.
-// Initially, a state has a reference count of zero, but the user may increment
-// or decrement this to control the time of destruction. In particular, a state
-// is destroyed when:
-//
-// 1. This instance is destroyed, or
-// 2. Clear() or Delete() is called, or
-// 3. Possibly (implementation-dependently) when:
-//    - Garbage collection is enabled (as defined by opts.gc),
-//    - The cache store size exceeds the limits (as defined by opts.gc_limits),
-//    - The state's reference count is zero, and
-//    - The state is not the most recently requested state.
-//
-// template <class S>
-// class CacheStore {
-//  public:
-//   using State = S;
-//   using Arc = typename State::Arc;
-//   using StateId = typename Arc::StateId;
-//
-//   // Required constructors/assignment operators.
-//   explicit CacheStore(const CacheOptions &opts);
-//
-//   // Returns nullptr if state is not stored.
-//   const State *GetState(StateId s);
-//
-//   // Creates state if state is not stored.
-//   State *GetMutableState(StateId s);
-//
-//   // Similar to State::AddArc() but updates cache store book-keeping.
-//   void AddArc(State *state, const Arc &arc);
-//
-//   // Similar to State::SetArcs() but updates cache store book-keeping; call
-//   // only once.
-//   void SetArcs(State *state);
-//
-//   // Similar to State::DeleteArcs() but updates cache store book-keeping.
-//
-//   void DeleteArcs(State *state);
-//
-//   void DeleteArcs(State *state, size_t n);
-//
-//   // Deletes all cached states.
-//   void Clear();
-//
-//   // Number of cached states.
-//   StateId CountStates();
-//
-//   // Iterates over cached states (in an arbitrary order); only needed if
-//   // opts.gc is true.
-//   bool Done() const;      // End of iteration.
-//   StateId Value() const;  // Current state.
-//   void Next();            // Advances to next state (when !Done).
-//   void Reset();           // Returns to initial condition.
-//   void Delete();          // Deletes current state and advances to next.
-// };
-
-// Container cache stores.
-
-// This class uses a vector of pointers to states to store cached states.
-template <class S>
-class VectorCacheStore {
- public:
-  using State = S;
-  using Arc = typename State::Arc;
-  using StateId = typename Arc::StateId;
-  using StateList = std::list<StateId, PoolAllocator<StateId>>;
-
-  // Required constructors/assignment operators.
-  explicit VectorCacheStore(const CacheOptions &opts) : cache_gc_(opts.gc) {
-    Clear();
-    Reset();
-  }
-
-  VectorCacheStore(const VectorCacheStore<S> &store)
-      : cache_gc_(store.cache_gc_) {
-    CopyStates(store);
-    Reset();
-  }
-
-  ~VectorCacheStore() { Clear(); }
-
-  VectorCacheStore<State> &operator=(const VectorCacheStore<State> &store) {
-    if (this != &store) {
-      CopyStates(store);
-      Reset();
-    }
-    return *this;
-  }
-
-  bool InBounds(StateId s) const {
-    return s < static_cast<StateId>(state_vec_.size());
-  }
-
-  // Returns nullptr if state is not stored.
-  const State *GetState(StateId s) const {
-    return InBounds(s) ? state_vec_[s] : nullptr;
-  }
-
-  // Creates state if state is not stored.
-  State *GetMutableState(StateId s) {
-    State *state = nullptr;
-    if (InBounds(s)) {
-      state = state_vec_[s];
-    } else {
-      state_vec_.resize(s + 1, nullptr);
-    }
-    if (!state) {
-      state = new (&state_alloc_) State(arc_alloc_);
-      state_vec_[s] = state;
-      if (cache_gc_) state_list_.push_back(s);
-    }
-    return state;
-  }
-
-  // Similar to State::AddArc() but updates cache store book-keeping
-  void AddArc(State *state, const Arc &arc) { state->AddArc(arc); }
-
-  // Similar to State::SetArcs() but updates cache store book-keeping; call
-  // only once.
-  void SetArcs(State *state) { state->SetArcs(); }
-
-  // Deletes all arcs.
-  void DeleteArcs(State *state) { state->DeleteArcs(); }
-
-  // Deletes some arcs.
-  void DeleteArcs(State *state, size_t n) { state->DeleteArcs(n); }
-
-  // Deletes all cached states.
-  void Clear() {
-    for (State *s : state_vec_) {
-      State::Destroy(s, &state_alloc_);
-    }
-    state_vec_.clear();
-    state_list_.clear();
-  }
-
-  StateId CountStates() const {
-    return std::count_if(state_vec_.begin(), state_vec_.end(),
-                         [](const State *s) { return s != nullptr; });
-  }
-
-  // Iterates over cached states (in an arbitrary order); only works if GC is
-  // enabled (o.w. avoiding state_list_ overhead).
-  bool Done() const { return iter_ == state_list_.end(); }
-
-  StateId Value() const { return *iter_; }
-
-  void Next() { ++iter_; }
-
-  void Reset() { iter_ = state_list_.begin(); }
-
-  // Deletes current state and advances to next.
-  void Delete() {
-    State::Destroy(state_vec_[*iter_], &state_alloc_);
-    state_vec_[*iter_] = nullptr;
-    state_list_.erase(iter_++);
-  }
-
- private:
-  void CopyStates(const VectorCacheStore<State> &store) {
-    Clear();
-    state_vec_.reserve(store.state_vec_.size());
-    for (size_t s = 0; s < store.state_vec_.size(); ++s) {
-      State *state = nullptr;
-      const auto *store_state = store.state_vec_[s];
-      if (store_state) {
-        state = new (&state_alloc_) State(*store_state, arc_alloc_);
-        if (cache_gc_) state_list_.push_back(s);
-      }
-      state_vec_.push_back(state);
-    }
-  }
-
-  bool cache_gc_;                               // Supports iteration when true.
-  std::vector<State *> state_vec_;              // Vector of states (or null).
-  StateList state_list_;                        // List of states.
-  typename StateList::iterator iter_;           // State list iterator.
-  typename State::StateAllocator state_alloc_;  // For state allocation.
-  typename State::ArcAllocator arc_alloc_;      // For arc allocation.
-};
-
-// This class uses a hash map from state IDs to pointers to cached states.
-template <class S>
-class HashCacheStore {
- public:
-  using State = S;
-  using Arc = typename State::Arc;
-  using StateId = typename Arc::StateId;
-
-  using StateMap =
-      std::unordered_map<StateId, State *, std::hash<StateId>,
-                          std::equal_to<StateId>,
-                          PoolAllocator<std::pair<const StateId, State *>>>;
-
-  // Required constructors/assignment operators.
-  explicit HashCacheStore(const CacheOptions &opts) {
-    Clear();
-    Reset();
-  }
-
-  HashCacheStore(const HashCacheStore<S> &store) {
-    CopyStates(store);
-    Reset();
-  }
-
-  ~HashCacheStore() { Clear(); }
-
-  HashCacheStore<State> &operator=(const HashCacheStore<State> &store) {
-    if (this != &store) {
-      CopyStates(store);
-      Reset();
-    }
-    return *this;
-  }
-
-  // Returns nullptr if state is not stored.
-  const State *GetState(StateId s) const {
-    const auto it = state_map_.find(s);
-    return it != state_map_.end() ? it->second : nullptr;
-  }
-
-  // Creates state if state is not stored.
-  State *GetMutableState(StateId s) {
-    auto *&state = state_map_[s];
-    if (!state) state = new (&state_alloc_) State(arc_alloc_);
-    return state;
-  }
-
-  // Similar to State::AddArc() but updates cache store book-keeping.
-  void AddArc(State *state, const Arc &arc) { state->AddArc(arc); }
-
-  // Similar to State::SetArcs() but updates internal cache size; call only
-  // once.
-  void SetArcs(State *state) { state->SetArcs(); }
-
-  // Deletes all arcs.
-  void DeleteArcs(State *state) { state->DeleteArcs(); }
-
-  // Deletes some arcs.
-  void DeleteArcs(State *state, size_t n) { state->DeleteArcs(n); }
-
-  // Deletes all cached states.
-  void Clear() {
-    for (auto it = state_map_.begin(); it != state_map_.end(); ++it) {
-      State::Destroy(it->second, &state_alloc_);
-    }
-    state_map_.clear();
-  }
-
-  StateId CountStates() const { return state_map_.size(); }
-
-  // Iterates over cached states (in an arbitrary order).
-  bool Done() const { return iter_ == state_map_.end(); }
-
-  StateId Value() const { return iter_->first; }
-
-  void Next() { ++iter_; }
-
-  void Reset() { iter_ = state_map_.begin(); }
-
-  // Deletes current state and advances to next.
-  void Delete() {
-    State::Destroy(iter_->second, &state_alloc_);
-    state_map_.erase(iter_++);
-  }
-
- private:
-  void CopyStates(const HashCacheStore<State> &store) {
-    Clear();
-    for (auto it = store.state_map_.begin(); it != store.state_map_.end();
-         ++it) {
-      state_map_[it->first] =
-          new (&state_alloc_) State(*it->second, arc_alloc_);
-    }
-  }
-
-  StateMap state_map_;                          // Map from state ID to state.
-  typename StateMap::iterator iter_;            // State map iterator.
-  typename State::StateAllocator state_alloc_;  // For state allocation.
-  typename State::ArcAllocator arc_alloc_;      // For arc allocation.
-};
-
-// Garbage-colllection cache stores.
-
-// This class implements a simple garbage collection scheme when
-// 'opts.gc_limit = 0'. In particular, the first cached state is reused for each
-// new state so long as the reference count is zero on the to-be-reused state.
-// Otherwise, the full underlying store is used. The caller can increment the
-// reference count to inhibit the GC of in-use states (e.g., in an ArcIterator).
-//
-// The typical use case for this optimization is when a single pass over a
-// cached
-// FST is performed with only one-state expanded at a time.
-template <class CacheStore>
-class FirstCacheStore {
- public:
-  using State = typename CacheStore::State;
-  using Arc = typename State::Arc;
-  using StateId = typename Arc::StateId;
-
-  // Required constructors/assignment operators.
-  explicit FirstCacheStore(const CacheOptions &opts)
-      : store_(opts),
-        cache_gc_(opts.gc_limit == 0),  // opts.gc ignored historically.
-        cache_first_state_id_(kNoStateId),
-        cache_first_state_(nullptr) {}
-
-  FirstCacheStore(const FirstCacheStore<CacheStore> &store)
-      : store_(store.store_),
-        cache_gc_(store.cache_gc_),
-        cache_first_state_id_(store.cache_first_state_id_),
-        cache_first_state_(store.cache_first_state_id_ != kNoStateId
-                               ? store_.GetMutableState(0)
-                               : nullptr) {}
-
-  FirstCacheStore<CacheStore> &operator=(
-      const FirstCacheStore<CacheStore> &store) {
-    if (this != &store) {
-      store_ = store.store_;
-      cache_gc_ = store.cache_gc_;
-      cache_first_state_id_ = store.cache_first_state_id_;
-      cache_first_state_ = store.cache_first_state_id_ != kNoStateId
-                               ? store_.GetMutableState(0)
-                               : nullptr;
-    }
-    return *this;
-  }
-
-  // Returns nullptr if state is not stored.
-  const State *GetState(StateId s) const {
-    // store_ state 0 may hold first cached state; the rest are shifted by 1.
-    return s == cache_first_state_id_ ? cache_first_state_
-                                      : store_.GetState(s + 1);
-  }
-
-  // Creates state if state is not stored.
-  State *GetMutableState(StateId s) {
-    // store_ state 0 used to hold first cached state; the rest are shifted by
-    // 1.
-    if (cache_first_state_id_ == s) {
-      return cache_first_state_;  // Request for first cached state.
-    }
-    if (cache_gc_) {
-      if (cache_first_state_id_ == kNoStateId) {
-        cache_first_state_id_ = s;  // Sets first cached state.
-        cache_first_state_ = store_.GetMutableState(0);
-        cache_first_state_->SetFlags(kCacheInit, kCacheInit);
-        cache_first_state_->ReserveArcs(2 * kAllocSize);
-        return cache_first_state_;
-      } else if (cache_first_state_->RefCount() == 0) {
-        cache_first_state_id_ = s;  // Updates first cached state.
-        cache_first_state_->Reset();
-        cache_first_state_->SetFlags(kCacheInit, kCacheInit);
-        return cache_first_state_;
-      } else {  // Keeps first cached state.
-        cache_first_state_->SetFlags(0, kCacheInit);  // Clears initialized bit.
-        cache_gc_ = false;                            // Disables GC.
-      }
-    }
-    auto *state = store_.GetMutableState(s + 1);
-    return state;
-  }
-
-  // Similar to State::AddArc() but updates cache store book-keeping.
-  void AddArc(State *state, const Arc &arc) { store_.AddArc(state, arc); }
-
-  // Similar to State::SetArcs() but updates internal cache size; call only
-  // once.
-  void SetArcs(State *state) { store_.SetArcs(state); }
-
-  // Deletes all arcs
-  void DeleteArcs(State *state) { store_.DeleteArcs(state); }
-
-  // Deletes some arcs
-  void DeleteArcs(State *state, size_t n) { store_.DeleteArcs(state, n); }
-
-  // Deletes all cached states
-  void Clear() {
-    store_.Clear();
-    cache_first_state_id_ = kNoStateId;
-    cache_first_state_ = nullptr;
-  }
-
-  StateId CountStates() const { return store_.CountStates(); }
-
-  // Iterates over cached states (in an arbitrary order). Only needed if GC is
-  // enabled.
-  bool Done() const { return store_.Done(); }
-
-  StateId Value() const {
-    // store_ state 0 may hold first cached state; rest shifted + 1.
-    const auto s = store_.Value();
-    return s ? s - 1 : cache_first_state_id_;
-  }
-
-  void Next() { store_.Next(); }
-
-  void Reset() { store_.Reset(); }
-
-  // Deletes current state and advances to next.
-  void Delete() {
-    if (Value() == cache_first_state_id_) {
-      cache_first_state_id_ = kNoStateId;
-      cache_first_state_ = nullptr;
-    }
-    store_.Delete();
-  }
-
- private:
-  CacheStore store_;              // Underlying store.
-  bool cache_gc_;                 // GC enabled.
-  StateId cache_first_state_id_;  // First cached state ID.
-  State *cache_first_state_;      // First cached state.
-};
-
-// This class implements mark-sweep garbage collection on an underlying cache
-// store. If GC is enabled, garbage collection of states is performed in a
-// rough approximation of LRU order once when 'gc_limit' bytes is reached. The
-// caller can increment the reference count to inhibit the GC of in-use state
-// (e.g., in an ArcIterator). With GC enabled, the 'gc_limit' parameter allows
-// the caller to trade-off time vs. space.
-template <class CacheStore>
-class GCCacheStore {
- public:
-  using State = typename CacheStore::State;
-  using Arc = typename State::Arc;
-  using StateId = typename Arc::StateId;
-
-  // Required constructors/assignment operators.
-  explicit GCCacheStore(const CacheOptions &opts)
-      : store_(opts),
-        cache_gc_request_(opts.gc),
-        cache_limit_(opts.gc_limit > kMinCacheLimit ? opts.gc_limit
-                                                    : kMinCacheLimit),
-        cache_gc_(false),
-        cache_size_(0) {}
-
-  // Returns 0 if state is not stored.
-  const State *GetState(StateId s) const { return store_.GetState(s); }
-
-  // Creates state if state is not stored
-  State *GetMutableState(StateId s) {
-    auto *state = store_.GetMutableState(s);
-    if (cache_gc_request_ && !(state->Flags() & kCacheInit)) {
-      state->SetFlags(kCacheInit, kCacheInit);
-      cache_size_ += sizeof(State) + state->NumArcs() * sizeof(Arc);
-      // GC is enabled once an uninited state (from underlying store) is seen.
-      cache_gc_ = true;
-      if (cache_size_ > cache_limit_) GC(state, false);
-    }
-    return state;
-  }
-
-  // Similar to State::AddArc() but updates cache store book-keeping.
-  void AddArc(State *state, const Arc &arc) {
-    store_.AddArc(state, arc);
-    if (cache_gc_ && (state->Flags() & kCacheInit)) {
-      cache_size_ += sizeof(Arc);
-      if (cache_size_ > cache_limit_) GC(state, false);
-    }
-  }
-
-  // Similar to State::SetArcs() but updates internal cache size; call only
-  // once.
-  void SetArcs(State *state) {
-    store_.SetArcs(state);
-    if (cache_gc_ && (state->Flags() & kCacheInit)) {
-      cache_size_ += state->NumArcs() * sizeof(Arc);
-      if (cache_size_ > cache_limit_) GC(state, false);
-    }
-  }
-
-  // Deletes all arcs.
-  void DeleteArcs(State *state) {
-    if (cache_gc_ && (state->Flags() & kCacheInit)) {
-      cache_size_ -= state->NumArcs() * sizeof(Arc);
-    }
-    store_.DeleteArcs(state);
-  }
-
-  // Deletes some arcs.
-  void DeleteArcs(State *state, size_t n) {
-    if (cache_gc_ && (state->Flags() & kCacheInit)) {
-      cache_size_ -= n * sizeof(Arc);
-    }
-    store_.DeleteArcs(state, n);
-  }
-
-  // Deletes all cached states.
-  void Clear() {
-    store_.Clear();
-    cache_size_ = 0;
-  }
-
-  StateId CountStates() const { return store_.CountStates(); }
-
-  // Iterates over cached states (in an arbitrary order); only needed if GC is
-  // enabled.
-  bool Done() const { return store_.Done(); }
-
-  StateId Value() const { return store_.Value(); }
-
-  void Next() { store_.Next(); }
-
-  void Reset() { store_.Reset(); }
-
-  // Deletes current state and advances to next.
-  void Delete() {
-    if (cache_gc_) {
-      const auto *state = store_.GetState(Value());
-      if (state->Flags() & kCacheInit) {
-        cache_size_ -= sizeof(State) + state->NumArcs() * sizeof(Arc);
-      }
-    }
-    store_.Delete();
-  }
-
-  // Removes from the cache store (not referenced-counted and not the current)
-  // states that have not been accessed since the last GC until at most
-  // cache_fraction * cache_limit_ bytes are cached. If that fails to free
-  // enough, attempts to uncaching recently visited states as well. If still
-  // unable to free enough memory, then widens cache_limit_.
-  void GC(const State *current, bool free_recent, float cache_fraction = 0.666);
-
-  // Returns the current cache size in bytes or 0 if GC is disabled.
-  size_t CacheSize() const { return cache_size_; }
-
-  // Returns the cache limit in bytes.
-  size_t CacheLimit() const { return cache_limit_; }
-
- private:
-  static constexpr size_t kMinCacheLimit = 8096;  // Minimum cache limit.
-
-  CacheStore store_;       // Underlying store.
-  bool cache_gc_request_;  // GC requested but possibly not yet enabled.
-  size_t cache_limit_;     // Number of bytes allowed before GC.
-  bool cache_gc_;          // GC enabled
-  size_t cache_size_;      // Number of bytes cached.
-};
-
-template <class CacheStore>
-void GCCacheStore<CacheStore>::GC(const State *current, bool free_recent,
-                                  float cache_fraction) {
-  if (!cache_gc_) return;
-  VLOG(2) << "GCCacheStore: Enter GC: object = "
-          << "(" << this << "), free recently cached = " << free_recent
-          << ", cache size = " << cache_size_
-          << ", cache frac = " << cache_fraction
-          << ", cache limit = " << cache_limit_ << "\n";
-  size_t cache_target = cache_fraction * cache_limit_;
-  store_.Reset();
-  while (!store_.Done()) {
-    auto *state = store_.GetMutableState(store_.Value());
-    if (cache_size_ > cache_target && state->RefCount() == 0 &&
-        (free_recent || !(state->Flags() & kCacheRecent)) && state != current) {
-      if (state->Flags() & kCacheInit) {
-        size_t size = sizeof(State) + state->NumArcs() * sizeof(Arc);
-        if (size < cache_size_) {
-          cache_size_ -= size;
-        }
-      }
-      store_.Delete();
-    } else {
-      state->SetFlags(0, kCacheRecent);
-      store_.Next();
-    }
-  }
-  if (!free_recent && cache_size_ > cache_target) {  // Recurses on recent.
-    GC(current, true, cache_fraction);
-  } else if (cache_target > 0) {  // Widens cache limit.
-    while (cache_size_ > cache_target) {
-      cache_limit_ *= 2;
-      cache_target *= 2;
-    }
-  } else if (cache_size_ > 0) {
-    FSTERROR() << "GCCacheStore:GC: Unable to free all cached states";
-  }
-  VLOG(2) << "GCCacheStore: Exit GC: object = "
-          << "(" << this << "), free recently cached = " << free_recent
-          << ", cache size = " << cache_size_
-          << ", cache frac = " << cache_fraction
-          << ", cache limit = " << cache_limit_ << "\n";
-}
-
-template <class CacheStore>
-constexpr size_t GCCacheStore<CacheStore>::kMinCacheLimit;
-
-// This class is the default cache state and store used by CacheBaseImpl.
-// It uses VectorCacheStore for storage decorated by FirstCacheStore
-// and GCCacheStore to do (optional) garbage collection.
-template <class Arc>
-class DefaultCacheStore
-    : public GCCacheStore<FirstCacheStore<VectorCacheStore<CacheState<Arc>>>> {
- public:
-  explicit DefaultCacheStore(const CacheOptions &opts)
-      : GCCacheStore<FirstCacheStore<VectorCacheStore<CacheState<Arc>>>>(opts) {
-  }
-};
-
-namespace internal {
-
-// This class is used to cache FST elements stored in states of type State
-// (see CacheState) with the flags used to indicate what has been cached. Use
-// HasStart(), HasFinal(), and HasArcs() to determine if cached and SetStart(),
-// SetFinal(), AddArc(), (or PushArc() and SetArcs()) to cache. Note that you
-// must set the final weight even if the state is non-final to mark it as
-// cached. The state storage method and any garbage collection policy are
-// determined by the cache store. If the store is passed in with the options,
-// CacheBaseImpl takes ownership.
-template <class State,
-          class CacheStore = DefaultCacheStore<typename State::Arc>>
-class CacheBaseImpl : public FstImpl<typename State::Arc> {
- public:
-  using Arc = typename State::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = CacheStore;
-
-  using FstImpl<Arc>::Type;
-  using FstImpl<Arc>::Properties;
-
-  explicit CacheBaseImpl(const CacheOptions &opts = CacheOptions())
-      : has_start_(false),
-        cache_start_(kNoStateId),
-        nknown_states_(0),
-        min_unexpanded_state_id_(0),
-        max_expanded_state_id_(-1),
-        cache_gc_(opts.gc),
-        cache_limit_(opts.gc_limit),
-        cache_store_(new CacheStore(opts)),
-        new_cache_store_(true),
-        own_cache_store_(true) {}
-
-  explicit CacheBaseImpl(const CacheImplOptions<CacheStore> &opts)
-      : has_start_(false),
-        cache_start_(kNoStateId),
-        nknown_states_(0),
-        min_unexpanded_state_id_(0),
-        max_expanded_state_id_(-1),
-        cache_gc_(opts.gc),
-        cache_limit_(opts.gc_limit),
-        cache_store_(opts.store ? opts.store : new CacheStore(CacheOptions(
-                                                   opts.gc, opts.gc_limit))),
-        new_cache_store_(!opts.store),
-        own_cache_store_(opts.store ? opts.own_store : true) {}
-
-  // Preserve gc parameters. If preserve_cache is true, also preserves
-  // cache data.
-  CacheBaseImpl(const CacheBaseImpl<State, CacheStore> &impl,
-                bool preserve_cache = false)
-      : FstImpl<Arc>(),
-        has_start_(false),
-        cache_start_(kNoStateId),
-        nknown_states_(0),
-        min_unexpanded_state_id_(0),
-        max_expanded_state_id_(-1),
-        cache_gc_(impl.cache_gc_),
-        cache_limit_(impl.cache_limit_),
-        cache_store_(new CacheStore(CacheOptions(cache_gc_, cache_limit_))),
-        new_cache_store_(impl.new_cache_store_ || !preserve_cache),
-        own_cache_store_(true) {
-    if (preserve_cache) {
-      *cache_store_ = *impl.cache_store_;
-      has_start_ = impl.has_start_;
-      cache_start_ = impl.cache_start_;
-      nknown_states_ = impl.nknown_states_;
-      expanded_states_ = impl.expanded_states_;
-      min_unexpanded_state_id_ = impl.min_unexpanded_state_id_;
-      max_expanded_state_id_ = impl.max_expanded_state_id_;
-    }
-  }
-
-  ~CacheBaseImpl() override { if (own_cache_store_) delete cache_store_; }
-
-  void SetStart(StateId s) {
-    cache_start_ = s;
-    has_start_ = true;
-    if (s >= nknown_states_) nknown_states_ = s + 1;
-  }
-
-  void SetFinal(StateId s, Weight weight) {
-    auto *state = cache_store_->GetMutableState(s);
-    state->SetFinal(std::move(weight));
-    static constexpr auto flags = kCacheFinal | kCacheRecent;
-    state->SetFlags(flags, flags);
-  }
-
-// Disabled to ensure PushArc not AddArc is used in existing code
-// TODO(sorenj): re-enable for backing store
-#if 0
-  // AddArc adds a single arc to a state and does incremental cache
-  // book-keeping. For efficiency, prefer PushArc and SetArcs below
-  // when possible.
-  void AddArc(StateId s, const Arc &arc) {
-    auto *state = cache_store_->GetMutableState(s);
-    cache_store_->AddArc(state, arc);
-    if (arc.nextstate >= nknown_states_)
-      nknown_states_ = arc.nextstate + 1;
-    SetExpandedState(s);
-    static constexpr auto flags = kCacheArcs | kCacheRecent;
-    state->SetFlags(flags, flags);
-  }
-#endif
-
-  // Adds a single arc to a state but delays cache book-keeping. SetArcs must
-  // be called when all PushArc and EmplaceArc calls at a state are complete.
-  // Do not mix with calls to AddArc.
-  void PushArc(StateId s, const Arc &arc) {
-    auto *state = cache_store_->GetMutableState(s);
-    state->PushArc(arc);
-  }
-
-  void PushArc(StateId s, Arc &&arc) {
-    auto *state = cache_store_->GetMutableState(s);
-    state->PushArc(std::move(arc));
-  }
-
-  // Adds a single arc to a state but delays cache book-keeping. SetArcs must
-  // be called when all PushArc and EmplaceArc calls at a state are complete.
-  // Do not mix with calls to AddArc.
-  template <class... T>
-  void EmplaceArc(StateId s, T &&... ctor_args) {
-    auto *state = cache_store_->GetMutableState(s);
-    state->EmplaceArc(std::forward<T>(ctor_args)...);
-  }
-
-  // Marks arcs of a state as cached and does cache book-keeping after all
-  // calls to PushArc have been completed. Do not mix with calls to AddArc.
-  void SetArcs(StateId s) {
-    auto *state = cache_store_->GetMutableState(s);
-    cache_store_->SetArcs(state);
-    const auto narcs = state->NumArcs();
-    for (size_t a = 0; a < narcs; ++a) {
-      const auto &arc = state->GetArc(a);
-      if (arc.nextstate >= nknown_states_) nknown_states_ = arc.nextstate + 1;
-    }
-    SetExpandedState(s);
-    static constexpr auto flags = kCacheArcs | kCacheRecent;
-    state->SetFlags(flags, flags);
-  }
-
-  void ReserveArcs(StateId s, size_t n) {
-    auto *state = cache_store_->GetMutableState(s);
-    state->ReserveArcs(n);
-  }
-
-  void DeleteArcs(StateId s) {
-    auto *state = cache_store_->GetMutableState(s);
-    cache_store_->DeleteArcs(state);
-  }
-
-  void DeleteArcs(StateId s, size_t n) {
-    auto *state = cache_store_->GetMutableState(s);
-    cache_store_->DeleteArcs(state, n);
-  }
-
-  void Clear() {
-    nknown_states_ = 0;
-    min_unexpanded_state_id_ = 0;
-    max_expanded_state_id_ = -1;
-    has_start_ = false;
-    cache_start_ = kNoStateId;
-    cache_store_->Clear();
-  }
-
-  // Is the start state cached?
-  bool HasStart() const {
-    if (!has_start_ && Properties(kError)) has_start_ = true;
-    return has_start_;
-  }
-
-  // Is the final weight of the state cached?
-  bool HasFinal(StateId s) const {
-    const auto *state = cache_store_->GetState(s);
-    if (state && state->Flags() & kCacheFinal) {
-      state->SetFlags(kCacheRecent, kCacheRecent);
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  // Are arcs of the state cached?
-  bool HasArcs(StateId s) const {
-    const auto *state = cache_store_->GetState(s);
-    if (state && state->Flags() & kCacheArcs) {
-      state->SetFlags(kCacheRecent, kCacheRecent);
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  StateId Start() const { return cache_start_; }
-
-  Weight Final(StateId s) const {
-    const auto *state = cache_store_->GetState(s);
-    return state->Final();
-  }
-
-  size_t NumArcs(StateId s) const {
-    const auto *state = cache_store_->GetState(s);
-    return state->NumArcs();
-  }
-
-  size_t NumInputEpsilons(StateId s) const {
-    const auto *state = cache_store_->GetState(s);
-    return state->NumInputEpsilons();
-  }
-
-  size_t NumOutputEpsilons(StateId s) const {
-    const auto *state = cache_store_->GetState(s);
-    return state->NumOutputEpsilons();
-  }
-
-  // Provides information needed for generic arc iterator.
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const {
-    const auto *state = cache_store_->GetState(s);
-    data->base = nullptr;
-    data->narcs = state->NumArcs();
-    data->arcs = state->Arcs();
-    data->ref_count = state->MutableRefCount();
-    state->IncrRefCount();
-  }
-
-  // Number of known states.
-  StateId NumKnownStates() const { return nknown_states_; }
-
-  // Updates number of known states, taking into account the passed state ID.
-  void UpdateNumKnownStates(StateId s) {
-    if (s >= nknown_states_) nknown_states_ = s + 1;
-  }
-
-  // Finds the mininum never-expanded state ID.
-  StateId MinUnexpandedState() const {
-    while (min_unexpanded_state_id_ <= max_expanded_state_id_ &&
-           ExpandedState(min_unexpanded_state_id_)) {
-      ++min_unexpanded_state_id_;
-    }
-    return min_unexpanded_state_id_;
-  }
-
-  // Returns maximum ever-expanded state ID.
-  StateId MaxExpandedState() const { return max_expanded_state_id_; }
-
-  void SetExpandedState(StateId s) {
-    if (s > max_expanded_state_id_) max_expanded_state_id_ = s;
-    if (s < min_unexpanded_state_id_) return;
-    if (s == min_unexpanded_state_id_) ++min_unexpanded_state_id_;
-    if (cache_gc_ || cache_limit_ == 0) {
-      if (expanded_states_.size() <= static_cast<size_t>(s))
-        expanded_states_.resize(s + 1, false);
-      expanded_states_[s] = true;
-    }
-  }
-
-  bool ExpandedState(StateId s) const {
-    if (cache_gc_ || cache_limit_ == 0) {
-      return expanded_states_[s];
-    } else if (new_cache_store_) {
-      return cache_store_->GetState(s) != nullptr;
-    } else {
-      // If the cache was not created by this class, then the cached state needs
-      // to be inspected to update nknown_states_.
-      return false;
-    }
-  }
-
-  const CacheStore *GetCacheStore() const { return cache_store_; }
-
-  CacheStore *GetCacheStore() { return cache_store_; }
-
-  // Caching on/off switch, limit and size accessors.
-
-  bool GetCacheGc() const { return cache_gc_; }
-
-  size_t GetCacheLimit() const { return cache_limit_; }
-
- private:
-  mutable bool has_start_;                   // Is the start state cached?
-  StateId cache_start_;                      // ID of start state.
-  StateId nknown_states_;                    // Number of known states.
-  std::vector<bool> expanded_states_;        // States that have been expanded.
-  mutable StateId min_unexpanded_state_id_;  // Minimum never-expanded state ID
-  mutable StateId max_expanded_state_id_;    // Maximum ever-expanded state ID
-  bool cache_gc_;                            // GC enabled.
-  size_t cache_limit_;       // Number of bytes allowed before GC.
-  CacheStore *cache_store_;  // The store of cached states.
-  bool new_cache_store_;     // Was the store was created by class?
-  bool own_cache_store_;     // Is the store owned by class?
-
-  CacheBaseImpl &operator=(const CacheBaseImpl &impl) = delete;
-};
-
-// A CacheBaseImpl with the default cache state type.
-template <class Arc>
-class CacheImpl : public CacheBaseImpl<CacheState<Arc>> {
- public:
-  using State = CacheState<Arc>;
-
-  CacheImpl() {}
-
-  explicit CacheImpl(const CacheOptions &opts)
-      : CacheBaseImpl<CacheState<Arc>>(opts) {}
-
-  CacheImpl(const CacheImpl<Arc> &impl, bool preserve_cache = false)
-      : CacheBaseImpl<State>(impl, preserve_cache) {}
-
- private:
-  CacheImpl &operator=(const CacheImpl &impl) = delete;
-};
-
-}  // namespace internal
-
-// Use this to make a state iterator for a CacheBaseImpl-derived FST, which must
-// have Arc and Store types defined. Note this iterator only returns those
-// states reachable from the initial state, so consider implementing a
-// class-specific one.
-//
-// This class may be derived from.
-template <class FST>
-class CacheStateIterator : public StateIteratorBase<typename FST::Arc> {
- public:
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = typename FST::Store;
-  using State = typename Store::State;
-  using Impl = internal::CacheBaseImpl<State, Store>;
-
-  CacheStateIterator(const FST &fst, Impl *impl)
-      : fst_(fst), impl_(impl), s_(0) {
-    fst_.Start();  // Forces start state.
-  }
-
-  bool Done() const final {
-    if (s_ < impl_->NumKnownStates()) return false;
-    for (StateId u = impl_->MinUnexpandedState(); u < impl_->NumKnownStates();
-         u = impl_->MinUnexpandedState()) {
-      // Forces state expansion.
-      ArcIterator<FST> aiter(fst_, u);
-      aiter.SetFlags(kArcValueFlags, kArcValueFlags | kArcNoCache);
-      for (; !aiter.Done(); aiter.Next()) {
-        impl_->UpdateNumKnownStates(aiter.Value().nextstate);
-      }
-      impl_->SetExpandedState(u);
-      if (s_ < impl_->NumKnownStates()) return false;
-    }
-    return true;
-  }
-
-  StateId Value() const final { return s_; }
-
-  void Next() final { ++s_; }
-
-  void Reset() final { s_ = 0; }
-
- private:
-  const FST &fst_;
-  Impl *impl_;
-  StateId s_;
-};
-
-// Used to make an arc iterator for a CacheBaseImpl-derived FST, which must
-// have Arc and State types defined.
-template <class FST>
-class CacheArcIterator {
- public:
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = typename FST::Store;
-  using State = typename Store::State;
-  using Impl = internal::CacheBaseImpl<State, Store>;
-
-  CacheArcIterator(Impl *impl, StateId s) : i_(0) {
-    state_ = impl->GetCacheStore()->GetMutableState(s);
-    state_->IncrRefCount();
-  }
-
-  ~CacheArcIterator() { state_->DecrRefCount(); }
-
-  bool Done() const { return i_ >= state_->NumArcs(); }
-
-  const Arc &Value() const { return state_->GetArc(i_); }
-
-  void Next() { ++i_; }
-
-  size_t Position() const { return i_; }
-
-  void Reset() { i_ = 0; }
-
-  void Seek(size_t a) { i_ = a; }
-
-  constexpr uint32 Flags() const { return kArcValueFlags; }
-
-  void SetFlags(uint32 flags, uint32 mask) {}
-
- private:
-  const State *state_;
-  size_t i_;
-
-  CacheArcIterator(const CacheArcIterator &) = delete;
-  CacheArcIterator &operator=(const CacheArcIterator &) = delete;
-};
-
-// Use this to make a mutable arc iterator for a CacheBaseImpl-derived FST,
-// which must have types Arc and Store defined.
-template <class FST>
-class CacheMutableArcIterator
-    : public MutableArcIteratorBase<typename FST::Arc> {
- public:
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = typename FST::Store;
-  using State = typename Store::State;
-  using Impl = internal::CacheBaseImpl<State, Store>;
-
-  // User must call MutateCheck() in the constructor.
-  CacheMutableArcIterator(Impl *impl, StateId s) : i_(0), s_(s), impl_(impl) {
-    state_ = impl_->GetCacheStore()->GetMutableState(s_);
-    state_->IncrRefCount();
-  }
-
-  ~CacheMutableArcIterator() override { state_->DecrRefCount(); }
-
-  bool Done() const final { return i_ >= state_->NumArcs(); }
-
-  const Arc &Value() const final { return state_->GetArc(i_); }
-
-  void Next() final { ++i_; }
-
-  size_t Position() const final { return i_; }
-
-  void Reset() final { i_ = 0; }
-
-  void Seek(size_t a) final { i_ = a; }
-
-  void SetValue(const Arc &arc) final { state_->SetArc(arc, i_); }
-
-  uint32 Flags() const final { return kArcValueFlags; }
-
-  void SetFlags(uint32, uint32) final {}
-
- private:
-  size_t i_;
-  StateId s_;
-  Impl *impl_;
-  State *state_;
-
-  CacheMutableArcIterator(const CacheMutableArcIterator &) = delete;
-  CacheMutableArcIterator &operator=(const CacheMutableArcIterator &) = delete;
-};
-
-// Wrap existing CacheStore implementation to use with ExpanderFst.
-template <class CacheStore>
-class ExpanderCacheStore {
- public:
-  using State = typename CacheStore::State;
-  using Arc = typename CacheStore::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit ExpanderCacheStore(const CacheOptions &opts = CacheOptions())
-      : store_(opts) {}
-
-  template <class Expander>
-  State *FindOrExpand(Expander &expander, StateId s) {  // NOLINT
-    auto *state = store_.GetMutableState(s);
-    if (state->Flags()) {
-      state->SetFlags(kCacheRecent, kCacheRecent);
-    } else {
-      StateBuilder builder(state);
-      expander.Expand(s, &builder);
-      state->SetFlags(kCacheFlags, kCacheFlags);
-      store_.SetArcs(state);
-    }
-    return state;
-  }
-
- private:
-  CacheStore store_;
-
-  struct StateBuilder {
-    State *state;
-
-    explicit StateBuilder(State *state_) : state(state_) {}
-
-    void AddArc(const Arc &arc) { state->PushArc(arc); }
-
-    void AddArc(Arc &&arc) { state->PushArc(std::move(arc)); }
-
-    void SetFinal(Weight weight) { state->SetFinal(std::move(weight)); }
-  };
-};
-
-}  // namespace fst
-
-#endif  // FST_CACHE_H_
diff --git a/projects/llm_framework/include/fst/closure.h b/projects/llm_framework/include/fst/closure.h
deleted file mode 100644
index 13beea9c..00000000
--- a/projects/llm_framework/include/fst/closure.h
+++ /dev/null
@@ -1,134 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to compute the concatenative closure of an FST.
-
-#ifndef FST_CLOSURE_H_
-#define FST_CLOSURE_H_
-
-#include <algorithm>
-#include <vector>
-
-#include <fst/mutable-fst.h>
-#include <fst/rational.h>
-
-
-namespace fst {
-
-// Computes the concatenative closure. This version modifies its
-// MutableFst input. If an FST transduces string x to y with weight a,
-// then its closure transduces x to y with weight a, xx to yy with
-// weight Times(a, a), xxx to yyy with with Times(Times(a, a), a),
-// etc. If closure_type == CLOSURE_STAR, then the empty string is
-// transduced to itself with weight Weight::One() as well.
-//
-// Complexity:
-//
-//   Time: O(V)
-//   Space: O(V)
-//
-// where V is the number of states.
-template <class Arc>
-void Closure(MutableFst<Arc> *fst, ClosureType closure_type) {
-  using Weight = typename Arc::Weight;
-  const auto props = fst->Properties(kFstProperties, false);
-  const auto start = fst->Start();
-  for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
-       siter.Next()) {
-    const auto s = siter.Value();
-    const auto weight = fst->Final(s);
-    if (weight != Weight::Zero()) fst->AddArc(s, Arc(0, 0, weight, start));
-  }
-  if (closure_type == CLOSURE_STAR) {
-    fst->ReserveStates(fst->NumStates() + 1);
-    const auto nstart = fst->AddState();
-    fst->SetStart(nstart);
-    fst->SetFinal(nstart, Weight::One());
-    if (start != kNoLabel) fst->AddArc(nstart, Arc(0, 0, Weight::One(), start));
-  }
-  fst->SetProperties(ClosureProperties(props, closure_type == CLOSURE_STAR),
-                     kFstProperties);
-}
-
-// Computes the concatenative closure. This version modifies its
-// RationalFst input.
-template <class Arc>
-void Closure(RationalFst<Arc> *fst, ClosureType closure_type) {
-  fst->GetMutableImpl()->AddClosure(closure_type);
-}
-
-struct ClosureFstOptions : RationalFstOptions {
-  ClosureType type;
-
-  ClosureFstOptions(const RationalFstOptions &opts,
-                    ClosureType type = CLOSURE_STAR)
-      : RationalFstOptions(opts), type(type) {}
-
-  explicit ClosureFstOptions(ClosureType type = CLOSURE_STAR) : type(type) {}
-};
-
-// Computes the concatenative closure. This version is a delayed FST. If an FST
-// transduces string x to y with weight a, then its closure transduces x to y
-// with weight a, xx to yy with weight Times(a, a), xxx to yyy with weight
-// Times(Times(a, a), a), etc. If closure_type == CLOSURE_STAR, then the empty
-// string is transduced to itself with weight Weight::One() as well.
-//
-// Complexity:
-//
-//   Time: O(v)
-//   Space: O(v)
-//
-// where v is the number of states visited. Constant time and space to visit an
-// input state or arc is assumed and exclusive of caching.
-template <class A>
-class ClosureFst : public RationalFst<A> {
- public:
-  using Arc = A;
-
-  ClosureFst(const Fst<Arc> &fst, ClosureType closure_type) {
-    GetMutableImpl()->InitClosure(fst, closure_type);
-  }
-
-  ClosureFst(const Fst<Arc> &fst, const ClosureFstOptions &opts)
-      : RationalFst<A>(opts) {
-    GetMutableImpl()->InitClosure(fst, opts.type);
-  }
-
-  // See Fst<>::Copy() for doc.
-  ClosureFst(const ClosureFst<Arc> &fst, bool safe = false)
-      : RationalFst<A>(fst, safe) {}
-
-  // Gets a copy of this ClosureFst. See Fst<>::Copy() for further doc.
-  ClosureFst<A> *Copy(bool safe = false) const override {
-    return new ClosureFst<A>(*this, safe);
-  }
-
- private:
-  using ImplToFst<internal::RationalFstImpl<Arc>>::GetImpl;
-  using ImplToFst<internal::RationalFstImpl<Arc>>::GetMutableImpl;
-};
-
-// Specialization for ClosureFst.
-template <class Arc>
-class StateIterator<ClosureFst<Arc>> : public StateIterator<RationalFst<Arc>> {
- public:
-  explicit StateIterator(const ClosureFst<Arc> &fst)
-      : StateIterator<RationalFst<Arc>>(fst) {}
-};
-
-// Specialization for ClosureFst.
-template <class Arc>
-class ArcIterator<ClosureFst<Arc>> : public ArcIterator<RationalFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const ClosureFst<Arc> &fst, StateId s)
-      : ArcIterator<RationalFst<Arc>>(fst, s) {}
-};
-
-// Useful alias when using StdArc.
-using StdClosureFst = ClosureFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_CLOSURE_H_
diff --git a/projects/llm_framework/include/fst/compact-fst.h b/projects/llm_framework/include/fst/compact-fst.h
deleted file mode 100644
index 402c87b7..00000000
--- a/projects/llm_framework/include/fst/compact-fst.h
+++ /dev/null
@@ -1,1564 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// FST Class for memory-efficient representation of common types of
-// FSTs: linear automata, acceptors, unweighted FSTs, ...
-
-#ifndef FST_COMPACT_FST_H_
-#define FST_COMPACT_FST_H_
-
-#include <climits>
-#include <iterator>
-#include <memory>
-#include <tuple>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/expanded-fst.h>
-#include <fst/fst-decl.h>  // For optional argument declarations
-#include <fst/mapped-file.h>
-#include <fst/matcher.h>
-#include <fst/test-properties.h>
-#include <fst/util.h>
-
-
-namespace fst {
-
-struct CompactFstOptions : public CacheOptions {
-  // The default caching behaviour is to do no caching. Most compactors are
-  // cheap and therefore we save memory by not doing caching.
-  CompactFstOptions() : CacheOptions(true, 0) {}
-
-  explicit CompactFstOptions(const CacheOptions &opts) : CacheOptions(opts) {}
-};
-
-// New upcoming (Fst) Compactor interface - currently used internally
-// by CompactFstImpl.
-//
-// class Compactor {
-//  public:
-//   // Constructor from the Fst to be compacted.
-//   Compactor(const Fst<Arc> &fst, ...);
-//   // Copy constructor
-//   Compactor(const Compactor &compactor, bool safe = false)
-//   // Default constructor (optional, see comment below).
-//   Compactor();
-//
-//   // Returns the start state, number of states, and total number of arcs
-//   // of the compacted Fst
-//   StateId Start() const;
-//   StateId NumStates() const;
-//   size_t NumArcs() const;
-//
-//   // Accessor class for state attributes.
-//   class State {
-//    public:
-//     State();  // Required, corresponds to kNoStateId.
-//     State(const Compactor *c, StateId);  // Accessor for StateId 's'.
-//     StateId GetStateId() const;
-//     Weight Final() const;
-//     size_t NumArcs() const;
-//     Arc GetArc(size_t i, uint32 f) const;
-//   };
-//
-//   // Modifies 'state' accessor to provide access to state id 's'.
-//   void SetState(StateId s, State *state);
-//   // Tests whether 'fst' can be compacted by this compactor.
-//   bool IsCompatible(const Fst<A> &fst) const;
-//   // Return the properties that are always true for an fst
-//   // compacted using this compactor
-//   uint64 Properties() const;
-//   // Return a string identifying the type of compactor.
-//   static const string &Type();
-//   // Return true if an error has occured.
-//   bool Error() const;
-//   // Writes a compactor to a file.
-//   bool Write(std::ostream &strm, const FstWriteOptions &opts) const;
-//   // Reads a compactor from a file.
-//   static Compactor*Read(std::istream &strm, const FstReadOptions &opts,
-//                         const FstHeader &hdr);
-// };
-//
-
-// Old (Arc) Compactor Interface:
-//
-// The ArcCompactor class determines how arcs and final weights are compacted
-// and expanded.
-//
-// Final weights are treated as transitions to the superfinal state, i.e.,
-// ilabel = olabel = kNoLabel and nextstate = kNoStateId.
-//
-// There are two types of compactors:
-//
-// * Fixed out-degree compactors: 'compactor.Size()' returns a positive integer
-//   's'. An FST can be compacted by this compactor only if each state has
-//   exactly 's' outgoing transitions (counting a non-Zero() final weight as a
-//   transition). A typical example is a compactor for string FSTs, i.e.,
-//   's == 1'.
-//
-// * Variable out-degree compactors: 'compactor.Size() == -1'. There are no
-//   out-degree restrictions for these compactors.
-//
-// Interface:
-//
-// class ArcCompactor {
-//  public:
-//   // Element is the type of the compacted transitions.
-//   using Element = ...
-//
-//   // Returns the compacted representation of a transition 'arc'
-//   // at a state 's'.
-//   Element Compact(StateId s, const Arc &arc);
-//
-//   // Returns the transition at state 's' represented by the compacted
-//   // transition 'e'.
-//   Arc Expand(StateId s, const Element &e) const;
-//
-//   // Returns -1 for variable out-degree compactors, and the mandatory
-//   // out-degree otherwise.
-//   ssize_t Size() const;
-//
-//   // Tests whether an FST can be compacted by this compactor.
-//   bool Compatible(const Fst<A> &fst) const;
-//
-//   // Returns the properties that are always true for an FST compacted using
-//   // this compactor
-//   uint64 Properties() const;
-//
-//   // Returns a string identifying the type of compactor.
-//   static const string &Type();
-//
-//   // Writes a compactor to a file.
-//   bool Write(std::ostream &strm) const;
-//
-//   // Reads a compactor from a file.
-//   static ArcCompactor *Read(std::istream &strm);
-//
-//   // Default constructor (optional, see comment below).
-//   ArcCompactor();
-// };
-//
-// The default constructor is only required for FST_REGISTER to work (i.e.,
-// enabling Convert() and the command-line utilities to work with this new
-// compactor). However, a default constructor always needs to be specified for
-// this code to compile, but one can have it simply raise an error when called,
-// like so:
-//
-// Compactor::Compactor() {
-//   FSTERROR() << "Compactor: No default constructor";
-// }
-
-// Default implementation data for CompactFst, which can shared between
-// otherwise independent copies.
-//
-// The implementation contains two arrays: 'states_' and 'compacts_'.
-//
-// For fixed out-degree compactors, the 'states_' array is unallocated. The
-// 'compacts_' contains the compacted transitions. Its size is 'ncompacts_'.
-// The outgoing transitions at a given state are stored consecutively. For a
-// given state 's', its 'compactor.Size()' outgoing transitions (including
-// superfinal transition when 's' is final), are stored in position
-// ['s*compactor.Size()', '(s+1)*compactor.Size()').
-//
-// For variable out-degree compactors, the states_ array has size
-// 'nstates_ + 1' and contains pointers to positions into 'compacts_'. For a
-// given state 's', the compacted transitions of 's' are stored in positions
-// ['states_[s]', 'states_[s + 1]') in 'compacts_'. By convention,
-// 'states_[nstates_] == ncompacts_'.
-//
-// In both cases, the superfinal transitions (when 's' is final, i.e.,
-// 'Final(s) != Weight::Zero()') are stored first.
-//
-// The unsigned type U is used to represent indices into the compacts_ array.
-template <class Element, class Unsigned>
-class DefaultCompactStore {
- public:
-  DefaultCompactStore()
-      : states_(nullptr),
-        compacts_(nullptr),
-        nstates_(0),
-        ncompacts_(0),
-        narcs_(0),
-        start_(kNoStateId),
-        error_(false) {}
-
-  template <class Arc, class Compactor>
-  DefaultCompactStore(const Fst<Arc> &fst, const Compactor &compactor);
-
-  template <class Iterator, class Compactor>
-  DefaultCompactStore(const Iterator &begin, const Iterator &end,
-                      const Compactor &compactor);
-
-  ~DefaultCompactStore() {
-    if (!states_region_) delete[] states_;
-    if (!compacts_region_) delete[] compacts_;
-  }
-
-  template <class Compactor>
-  static DefaultCompactStore<Element, Unsigned> *Read(
-      std::istream &strm, const FstReadOptions &opts, const FstHeader &hdr,
-      const Compactor &compactor);
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const;
-
-  Unsigned States(ssize_t i) const { return states_[i]; }
-
-  const Element &Compacts(size_t i) const { return compacts_[i]; }
-
-  size_t NumStates() const { return nstates_; }
-
-  size_t NumCompacts() const { return ncompacts_; }
-
-  size_t NumArcs() const { return narcs_; }
-
-  ssize_t Start() const { return start_; }
-
-  bool Error() const { return error_; }
-
-  // Returns a string identifying the type of data storage container.
-  static const string &Type();
-
- private:
-  std::unique_ptr<MappedFile> states_region_;
-  std::unique_ptr<MappedFile> compacts_region_;
-  Unsigned *states_;
-  Element *compacts_;
-  size_t nstates_;
-  size_t ncompacts_;
-  size_t narcs_;
-  ssize_t start_;
-  bool error_;
-};
-
-template <class Element, class Unsigned>
-template <class Arc, class Compactor>
-DefaultCompactStore<Element, Unsigned>::DefaultCompactStore(
-    const Fst<Arc> &fst, const Compactor &compactor)
-    : states_(nullptr),
-      compacts_(nullptr),
-      nstates_(0),
-      ncompacts_(0),
-      narcs_(0),
-      start_(kNoStateId),
-      error_(false) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  start_ = fst.Start();
-  // Counts # of states and arcs.
-  StateId nfinals = 0;
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    ++nstates_;
-    const auto s = siter.Value();
-    narcs_ += fst.NumArcs(s);
-    if (fst.Final(s) != Weight::Zero()) ++nfinals;
-  }
-  if (compactor.Size() == -1) {
-    states_ = new Unsigned[nstates_ + 1];
-    ncompacts_ = narcs_ + nfinals;
-    compacts_ = new Element[ncompacts_];
-    states_[nstates_] = ncompacts_;
-  } else {
-    states_ = nullptr;
-    ncompacts_ = nstates_ * compactor.Size();
-    if ((narcs_ + nfinals) != ncompacts_) {
-      FSTERROR() << "DefaultCompactStore: Compactor incompatible with FST";
-      error_ = true;
-      return;
-    }
-    compacts_ = new Element[ncompacts_];
-  }
-  size_t pos = 0;
-  size_t fpos = 0;
-  for (size_t s = 0; s < nstates_; ++s) {
-    fpos = pos;
-    if (compactor.Size() == -1) states_[s] = pos;
-    if (fst.Final(s) != Weight::Zero()) {
-      compacts_[pos++] = compactor.Compact(
-          s, Arc(kNoLabel, kNoLabel, fst.Final(s), kNoStateId));
-    }
-    for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-      compacts_[pos++] = compactor.Compact(s, aiter.Value());
-    }
-    if ((compactor.Size() != -1) && (pos != fpos + compactor.Size())) {
-      FSTERROR() << "DefaultCompactStore: Compactor incompatible with FST";
-      error_ = true;
-      return;
-    }
-  }
-  if (pos != ncompacts_) {
-    FSTERROR() << "DefaultCompactStore: Compactor incompatible with FST";
-    error_ = true;
-    return;
-  }
-}
-
-template <class Element, class Unsigned>
-template <class Iterator, class Compactor>
-DefaultCompactStore<Element, Unsigned>::DefaultCompactStore(
-    const Iterator &begin, const Iterator &end, const Compactor &compactor)
-    : states_(nullptr),
-      compacts_(nullptr),
-      nstates_(0),
-      ncompacts_(0),
-      narcs_(0),
-      start_(kNoStateId),
-      error_(false) {
-  using Arc = typename Compactor::Arc;
-  using Weight = typename Arc::Weight;
-  if (compactor.Size() != -1) {
-    ncompacts_ = std::distance(begin, end);
-    if (compactor.Size() == 1) {
-      // For strings, allows implicit final weight. Empty input is the empty
-      // string.
-      if (ncompacts_ == 0) {
-        ++ncompacts_;
-      } else {
-        const auto arc =
-            compactor.Expand(ncompacts_ - 1, *(begin + (ncompacts_ - 1)));
-        if (arc.ilabel != kNoLabel) ++ncompacts_;
-      }
-    }
-    if (ncompacts_ % compactor.Size()) {
-      FSTERROR() << "DefaultCompactStore: Size of input container incompatible"
-                 << " with compactor";
-      error_ = true;
-      return;
-    }
-    if (ncompacts_ == 0) return;
-    start_ = 0;
-    nstates_ = ncompacts_ / compactor.Size();
-    compacts_ = new Element[ncompacts_];
-    size_t i = 0;
-    Iterator it = begin;
-    for (; it != end; ++it, ++i) {
-      compacts_[i] = *it;
-      if (compactor.Expand(i, *it).ilabel != kNoLabel) ++narcs_;
-    }
-    if (i < ncompacts_) {
-      compacts_[i] = compactor.Compact(
-          i, Arc(kNoLabel, kNoLabel, Weight::One(), kNoStateId));
-    }
-  } else {
-    if (std::distance(begin, end) == 0) return;
-    // Count # of states, arcs and compacts.
-    auto it = begin;
-    for (size_t i = 0; it != end; ++it, ++i) {
-      const auto arc = compactor.Expand(i, *it);
-      if (arc.ilabel != kNoLabel) {
-        ++narcs_;
-        ++ncompacts_;
-      } else {
-        ++nstates_;
-        if (arc.weight != Weight::Zero()) ++ncompacts_;
-      }
-    }
-    start_ = 0;
-    compacts_ = new Element[ncompacts_];
-    states_ = new Unsigned[nstates_ + 1];
-    states_[nstates_] = ncompacts_;
-    size_t i = 0;
-    size_t s = 0;
-    for (it = begin; it != end; ++it) {
-      const auto arc = compactor.Expand(i, *it);
-      if (arc.ilabel != kNoLabel) {
-        compacts_[i++] = *it;
-      } else {
-        states_[s++] = i;
-        if (arc.weight != Weight::Zero()) compacts_[i++] = *it;
-      }
-    }
-    if ((s != nstates_) || (i != ncompacts_)) {
-      FSTERROR() << "DefaultCompactStore: Ill-formed input container";
-      error_ = true;
-      return;
-    }
-  }
-}
-
-template <class Element, class Unsigned>
-template <class Compactor>
-DefaultCompactStore<Element, Unsigned>
-    *DefaultCompactStore<Element, Unsigned>::Read(std::istream &strm,
-                                                  const FstReadOptions &opts,
-                                                  const FstHeader &hdr,
-                                                  const Compactor &compactor) {
-  std::unique_ptr<DefaultCompactStore<Element, Unsigned>> data(
-      new DefaultCompactStore<Element, Unsigned>());
-  data->start_ = hdr.Start();
-  data->nstates_ = hdr.NumStates();
-  data->narcs_ = hdr.NumArcs();
-  if (compactor.Size() == -1) {
-    if ((hdr.GetFlags() & FstHeader::IS_ALIGNED) && !AlignInput(strm)) {
-      LOG(ERROR) << "DefaultCompactStore::Read: Alignment failed: "
-                 << opts.source;
-      return nullptr;
-    }
-    auto b = (data->nstates_ + 1) * sizeof(Unsigned);
-    data->states_region_.reset(MappedFile::Map(
-        &strm, opts.mode == FstReadOptions::MAP, opts.source, b));
-    if (!strm || !data->states_region_) {
-      LOG(ERROR) << "DefaultCompactStore::Read: Read failed: " << opts.source;
-      return nullptr;
-    }
-    data->states_ =
-        static_cast<Unsigned *>(data->states_region_->mutable_data());
-  } else {
-    data->states_ = nullptr;
-  }
-  data->ncompacts_ = compactor.Size() == -1 ? data->states_[data->nstates_]
-                                            : data->nstates_ * compactor.Size();
-  if ((hdr.GetFlags() & FstHeader::IS_ALIGNED) && !AlignInput(strm)) {
-    LOG(ERROR) << "DefaultCompactStore::Read: Alignment failed: "
-               << opts.source;
-    return nullptr;
-  }
-  size_t b = data->ncompacts_ * sizeof(Element);
-  data->compacts_region_.reset(
-      MappedFile::Map(&strm, opts.mode == FstReadOptions::MAP, opts.source, b));
-  if (!strm || !data->compacts_region_) {
-    LOG(ERROR) << "DefaultCompactStore::Read: Read failed: " << opts.source;
-    return nullptr;
-  }
-  data->compacts_ =
-      static_cast<Element *>(data->compacts_region_->mutable_data());
-  return data.release();
-}
-
-template <class Element, class Unsigned>
-bool DefaultCompactStore<Element, Unsigned>::Write(
-    std::ostream &strm, const FstWriteOptions &opts) const {
-  if (states_) {
-    if (opts.align && !AlignOutput(strm)) {
-      LOG(ERROR) << "DefaultCompactStore::Write: Alignment failed: "
-                 << opts.source;
-      return false;
-    }
-    strm.write(reinterpret_cast<char *>(states_),
-               (nstates_ + 1) * sizeof(Unsigned));
-  }
-  if (opts.align && !AlignOutput(strm)) {
-    LOG(ERROR) << "DefaultCompactStore::Write: Alignment failed: "
-               << opts.source;
-    return false;
-  }
-  strm.write(reinterpret_cast<char *>(compacts_), ncompacts_ * sizeof(Element));
-  strm.flush();
-  if (!strm) {
-    LOG(ERROR) << "DefaultCompactStore::Write: Write failed: " << opts.source;
-    return false;
-  }
-  return true;
-}
-
-template <class Element, class Unsigned>
-const string &DefaultCompactStore<Element, Unsigned>::Type() {
-  static const string *const type = new string("compact");
-  return *type;
-}
-
-template <class C, class U, class S> class DefaultCompactState;
-
-// Wraps an arc compactor and a compact store as a new Fst compactor.
-template <class C, class U,
-          class S = DefaultCompactStore<typename C::Element, U>>
-class DefaultCompactor {
- public:
-  using ArcCompactor = C;
-  using Unsigned = U;
-  using CompactStore = S;
-  using Element = typename C::Element;
-  using Arc = typename C::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using State = DefaultCompactState<C, U, S>;
-  friend State;
-
-  DefaultCompactor()
-      : arc_compactor_(nullptr), compact_store_(nullptr) {}
-
-  // Constructs from Fst.
-  DefaultCompactor(const Fst<Arc> &fst,
-                   std::shared_ptr<ArcCompactor> arc_compactor)
-      : arc_compactor_(std::move(arc_compactor)),
-        compact_store_(std::make_shared<S>(fst, *arc_compactor_)) {}
-
-  DefaultCompactor(const Fst<Arc> &fst,
-                   std::shared_ptr<DefaultCompactor<C, U, S>> compactor)
-      : arc_compactor_(compactor->arc_compactor_),
-        compact_store_(compactor->compact_store_ == nullptr ?
-                       std::make_shared<S>(fst, *arc_compactor_) :
-                       compactor->compact_store_) {}
-
-  // Constructs from CompactStore.
-  DefaultCompactor(std::shared_ptr<ArcCompactor> arc_compactor,
-                   std::shared_ptr<CompactStore> compact_store)
-      : arc_compactor_(std::move(arc_compactor)),
-        compact_store_(std::move(compact_store)) {}
-
-  // Constructs from set of compact elements (when arc_compactor.Size() != -1).
-  template <class Iterator>
-  DefaultCompactor(const Iterator &b, const Iterator &e,
-                   std::shared_ptr<C> arc_compactor)
-      : arc_compactor_(std::move(arc_compactor)),
-        compact_store_(std::make_shared<S>(b, e, *arc_compactor_)) {}
-
-  // Copy constructor.
-  DefaultCompactor(const DefaultCompactor<C, U, S> &compactor)
-      : arc_compactor_(std::make_shared<C>(*compactor.GetArcCompactor())),
-        compact_store_(compactor.SharedCompactStore()) {}
-
-  template <class OtherC>
-  explicit DefaultCompactor(const DefaultCompactor<OtherC, U, S> &compactor)
-      : arc_compactor_(std::make_shared<C>(*compactor.GetArcCompactor())),
-        compact_store_(compactor.SharedCompactStore()) {}
-
-  StateId Start() const { return compact_store_->Start(); }
-  StateId NumStates() const { return compact_store_->NumStates(); }
-  size_t NumArcs() const { return compact_store_->NumArcs(); }
-
-  void SetState(StateId s, State *state) const {
-    if (state->GetStateId() != s) state->Set(this, s);
-  }
-
-  static DefaultCompactor<C, U, S> *Read(std::istream &strm,
-                                         const FstReadOptions &opts,
-                                         const FstHeader &hdr) {
-    std::shared_ptr<C> arc_compactor(C::Read(strm));
-    if (arc_compactor == nullptr) return nullptr;
-    std::shared_ptr<S> compact_store(S::Read(strm, opts, hdr, *arc_compactor));
-    if (compact_store == nullptr) return nullptr;
-    return new DefaultCompactor<C, U, S>(arc_compactor, compact_store);
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
-    return arc_compactor_->Write(strm) && compact_store_->Write(strm, opts);
-  }
-
-  uint64 Properties() const { return arc_compactor_->Properties(); }
-
-  bool IsCompatible(const Fst<Arc> &fst) const {
-    return arc_compactor_->Compatible(fst);
-  }
-
-  bool Error() const { return compact_store_->Error(); }
-
-  bool HasFixedOutdegree() const { return arc_compactor_->Size() != -1; }
-
-  static const string &Type() {
-    static const string *const type = [] {
-      string type = "compact";
-      if (sizeof(U) != sizeof(uint32)) type += std::to_string(8 * sizeof(U));
-      type += "_";
-      type += C::Type();
-      if (CompactStore::Type() != "compact") {
-        type += "_";
-        type += CompactStore::Type();
-      }
-      return new string(type);
-    }();
-    return *type;
-  }
-
-  const ArcCompactor *GetArcCompactor() const { return arc_compactor_.get(); }
-  CompactStore *GetCompactStore() const { return compact_store_.get(); }
-
-  std::shared_ptr<ArcCompactor> SharedArcCompactor() const {
-    return arc_compactor_;
-  }
-
-  std::shared_ptr<CompactStore> SharedCompactStore() const {
-    return compact_store_;
-  }
-
-  // TODO(allauzen): remove dependencies on this method and make private.
-  Arc ComputeArc(StateId s, Unsigned i, uint32 f) const {
-    return arc_compactor_->Expand(s, compact_store_->Compacts(i), f);
-  }
-
- private:
-  std::pair<Unsigned, Unsigned> CompactsRange(StateId s) const {
-    std::pair<size_t, size_t> range;
-    if (HasFixedOutdegree()) {
-      range.first = s * arc_compactor_->Size();
-      range.second = arc_compactor_->Size();
-    } else {
-      range.first = compact_store_->States(s);
-      range.second = compact_store_->States(s + 1) - range.first;
-    }
-    return range;
-  }
-
- private:
-  std::shared_ptr<ArcCompactor> arc_compactor_;
-  std::shared_ptr<CompactStore> compact_store_;
-};
-
-// Default implementation of state attributes accessor class for
-// DefaultCompactor. Use of efficient specialization strongly encouraged.
-template <class C, class U, class S>
-class DefaultCompactState {
- public:
-  using Arc = typename C::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  DefaultCompactState() = default;
-
-  DefaultCompactState(const DefaultCompactor<C, U, S> *compactor, StateId s)
-      : compactor_(compactor),
-        s_(s),
-        range_(compactor->CompactsRange(s)),
-        has_final_(
-            range_.second != 0 &&
-            compactor->ComputeArc(s, range_.first,
-                                 kArcILabelValue).ilabel == kNoLabel) {
-    if (has_final_) {
-      ++range_.first;
-      --range_.second;
-    }
-  }
-
-  void Set(const DefaultCompactor<C, U, S> *compactor, StateId s) {
-    compactor_ = compactor;
-    s_ = s;
-    range_ = compactor->CompactsRange(s);
-    if (range_.second != 0 &&
-        compactor->ComputeArc(s, range_.first, kArcILabelValue).ilabel
-        == kNoLabel) {
-      has_final_ = true;
-      ++range_.first;
-      --range_.second;
-    } else {
-      has_final_ = false;
-    }
-  }
-
-  StateId GetStateId() const { return s_; }
-
-  Weight Final() const {
-    if (!has_final_) return Weight::Zero();
-    return compactor_->ComputeArc(s_, range_.first - 1, kArcWeightValue).weight;
-  }
-
-  size_t NumArcs() const { return range_.second; }
-
-  Arc GetArc(size_t i, uint32 f) const {
-    return compactor_->ComputeArc(s_, range_.first + i, f);
-  }
-
- private:
-  const DefaultCompactor<C, U, S> *compactor_ = nullptr;  // borrowed ref.
-  StateId s_ = kNoStateId;
-  std::pair<U, U> range_ = {0, 0};
-  bool has_final_ = false;
-};
-
-// Specialization for DefaultCompactStore.
-template <class C, class U>
-class DefaultCompactState<C, U, DefaultCompactStore<typename C::Element, U>> {
- public:
-  using Arc = typename C::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using CompactStore = DefaultCompactStore<typename C::Element, U>;
-
-  DefaultCompactState() = default;
-
-  DefaultCompactState(
-      const DefaultCompactor<C, U, CompactStore> *compactor, StateId s)
-      : arc_compactor_(compactor->GetArcCompactor()), s_(s) {
-    Init(compactor);
-  }
-
-  void Set(const DefaultCompactor<C, U, CompactStore> *compactor, StateId s) {
-    arc_compactor_ = compactor->GetArcCompactor();
-    s_ = s;
-    has_final_ = false;
-    Init(compactor);
-  }
-
-  StateId GetStateId() const { return s_; }
-
-  Weight Final() const {
-    if (!has_final_) return Weight::Zero();
-    return arc_compactor_->Expand(s_, *(compacts_ - 1), kArcWeightValue).weight;
-  }
-
-  size_t NumArcs() const { return num_arcs_; }
-
-  Arc GetArc(size_t i, uint32 f) const {
-    return arc_compactor_->Expand(s_, compacts_[i], f);
-  }
-
- private:
-  void Init(const DefaultCompactor<C, U, CompactStore> *compactor) {
-    const auto *store = compactor->GetCompactStore();
-    U offset;
-    if (!compactor->HasFixedOutdegree()) {  // Variable out-degree compactor.
-      offset = store->States(s_);
-      num_arcs_ = store->States(s_ + 1) - offset;
-    } else {  // Fixed out-degree compactor.
-      offset = s_ * arc_compactor_->Size();
-      num_arcs_ = arc_compactor_->Size();
-    }
-    if (num_arcs_ > 0) {
-      compacts_ = &(store->Compacts(offset));
-      if (arc_compactor_->Expand(s_, *compacts_, kArcILabelValue).ilabel
-          == kNoStateId) {
-        ++compacts_;
-        --num_arcs_;
-        has_final_ = true;
-      }
-    }
-  }
-
- private:
-  const C *arc_compactor_ = nullptr;               // Borrowed reference.
-  const typename C::Element *compacts_ = nullptr;  // Borrowed reference.
-  StateId s_ = kNoStateId;
-  U num_arcs_ = 0;
-  bool has_final_ = false;
-};
-
-template <class Arc, class ArcCompactor, class Unsigned, class CompactStore,
-          class CacheStore>
-class CompactFst;
-
-template <class F, class G>
-void Cast(const F &, G *);
-
-namespace internal {
-
-// Implementation class for CompactFst, which contains parametrizeable
-// Fst data storage (DefaultCompactStore by default) and Fst cache.
-template <class Arc, class C, class CacheStore = DefaultCacheStore<Arc>>
-class CompactFstImpl
-    : public CacheBaseImpl<typename CacheStore::State, CacheStore> {
- public:
-  using Weight = typename Arc::Weight;
-  using StateId = typename Arc::StateId;
-  using Compactor = C;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::Properties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-  using FstImpl<Arc>::WriteHeader;
-
-  using ImplBase = CacheBaseImpl<typename CacheStore::State, CacheStore>;
-  using ImplBase::PushArc;
-  using ImplBase::HasArcs;
-  using ImplBase::HasFinal;
-  using ImplBase::HasStart;
-  using ImplBase::SetArcs;
-  using ImplBase::SetFinal;
-  using ImplBase::SetStart;
-
-  CompactFstImpl()
-      : ImplBase(CompactFstOptions()),
-        compactor_() {
-    SetType(Compactor::Type());
-    SetProperties(kNullProperties | kStaticProperties);
-  }
-
-  CompactFstImpl(const Fst<Arc> &fst, std::shared_ptr<Compactor> compactor,
-                 const CompactFstOptions &opts)
-      : ImplBase(opts),
-        compactor_(std::make_shared<Compactor>(fst, compactor)) {
-    SetType(Compactor::Type());
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-    if (compactor_->Error()) SetProperties(kError, kError);
-    uint64 copy_properties = fst.Properties(kMutable, false) ?
-        fst.Properties(kCopyProperties, true):
-        CheckProperties(fst,
-                        kCopyProperties & ~kWeightedCycles & ~kUnweightedCycles,
-                        kCopyProperties);
-    if ((copy_properties & kError) || !compactor_->IsCompatible(fst)) {
-      FSTERROR() << "CompactFstImpl: Input Fst incompatible with compactor";
-      SetProperties(kError, kError);
-      return;
-    }
-    SetProperties(copy_properties | kStaticProperties);
-  }
-
-  CompactFstImpl(std::shared_ptr<Compactor> compactor,
-                 const CompactFstOptions &opts)
-      : ImplBase(opts),
-        compactor_(compactor) {
-    SetType(Compactor::Type());
-    SetProperties(kStaticProperties | compactor_->Properties());
-    if (compactor_->Error()) SetProperties(kError, kError);
-  }
-
-  CompactFstImpl(const CompactFstImpl<Arc, Compactor, CacheStore> &impl)
-      : ImplBase(impl),
-        compactor_(impl.compactor_ == nullptr ?
-                   std::make_shared<Compactor>() :
-                   std::make_shared<Compactor>(*impl.compactor_)) {
-    SetType(impl.Type());
-    SetProperties(impl.Properties());
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  // Allows to change the cache store from OtherI to I.
-  template <class OtherCacheStore>
-  CompactFstImpl(const CompactFstImpl<Arc, Compactor, OtherCacheStore> &impl)
-      : ImplBase(CacheOptions(impl.GetCacheGc(), impl.GetCacheLimit())),
-        compactor_(impl.compactor_ == nullptr ?
-                   std::make_shared<Compactor>() :
-                   std::make_shared<Compactor>(*impl.compactor_)) {
-    SetType(impl.Type());
-    SetProperties(impl.Properties());
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  StateId Start() {
-    if (!HasStart()) SetStart(compactor_->Start());
-    return ImplBase::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (HasFinal(s)) return ImplBase::Final(s);
-    compactor_->SetState(s, &state_);
-    return state_.Final();
-  }
-
-  StateId NumStates() const {
-    if (Properties(kError)) return 0;
-    return compactor_->NumStates();
-  }
-
-  size_t NumArcs(StateId s) {
-    if (HasArcs(s)) return ImplBase::NumArcs(s);
-    compactor_->SetState(s, &state_);
-    return state_.NumArcs();
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s) && !Properties(kILabelSorted)) Expand(s);
-    if (HasArcs(s)) return ImplBase::NumInputEpsilons(s);
-    return CountEpsilons(s, false);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s) && !Properties(kOLabelSorted)) Expand(s);
-    if (HasArcs(s)) return ImplBase::NumOutputEpsilons(s);
-    return CountEpsilons(s, true);
-  }
-
-  size_t CountEpsilons(StateId s, bool output_epsilons) {
-    compactor_->SetState(s, &state_);
-    const uint32 f = output_epsilons ? kArcOLabelValue : kArcILabelValue;
-    size_t num_eps = 0;
-    for (size_t i = 0; i < state_.NumArcs(); ++i) {
-      const auto& arc = state_.GetArc(i, f);
-      const auto label = output_epsilons ? arc.olabel : arc.ilabel;
-      if (label == 0)
-        ++num_eps;
-      else if (label > 0)
-        break;
-    }
-    return num_eps;
-  }
-
-  static CompactFstImpl<Arc, Compactor, CacheStore> *Read(
-      std::istream &strm, const FstReadOptions &opts) {
-    std::unique_ptr<CompactFstImpl<Arc, Compactor, CacheStore>> impl(
-      new CompactFstImpl<Arc, Compactor, CacheStore>());
-    FstHeader hdr;
-    if (!impl->ReadHeader(strm, opts, kMinFileVersion, &hdr)) {
-      return nullptr;
-    }
-    // Ensures compatibility.
-    if (hdr.Version() == kAlignedFileVersion) {
-      hdr.SetFlags(hdr.GetFlags() | FstHeader::IS_ALIGNED);
-    }
-    impl->compactor_ = std::shared_ptr<Compactor>(
-        Compactor::Read(strm, opts, hdr));
-    if (!impl->compactor_) {
-      return nullptr;
-    }
-    return impl.release();
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
-    FstHeader hdr;
-    hdr.SetStart(compactor_->Start());
-    hdr.SetNumStates(compactor_->NumStates());
-    hdr.SetNumArcs(compactor_->NumArcs());
-    // Ensures compatibility.
-    const auto file_version = opts.align ? kAlignedFileVersion : kFileVersion;
-    WriteHeader(strm, opts, file_version, &hdr);
-    return compactor_->Write(strm, opts);
-  }
-
-  // Provides information needed for generic state iterator.
-  void InitStateIterator(StateIteratorData<Arc> *data) const {
-    data->base = nullptr;
-    data->nstates = compactor_->NumStates();
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    ImplBase::InitArcIterator(s, data);
-  }
-
-  void Expand(StateId s) {
-    compactor_->SetState(s, &state_);
-    for (size_t i = 0; i < state_.NumArcs(); ++i)
-      PushArc(s, state_.GetArc(i, kArcValueFlags));
-    SetArcs(s);
-    if (!HasFinal(s)) SetFinal(s, state_.Final());
-  }
-
-  const Compactor *GetCompactor() const { return compactor_.get(); }
-  std::shared_ptr<Compactor> SharedCompactor() const { return compactor_; }
-  void SetCompactor(std::shared_ptr<Compactor> compactor) {
-    // TODO(allauzen): is this correct? is this needed?
-    // TODO(allauzen): consider removing and forcing this through direct calls
-    // to compactor.
-    compactor_ = compactor;
-  }
-
-  // Properties always true of this FST class.
-  static constexpr uint64 kStaticProperties = kExpanded;
-
- protected:
-  template <class OtherArc, class OtherCompactor, class OtherCacheStore>
-  explicit CompactFstImpl(
-    const CompactFstImpl<OtherArc, OtherCompactor, OtherCacheStore> &impl)
-    : compactor_(std::make_shared<Compactor>(*impl.GetCompactor())) {
-    SetType(impl.Type());
-    SetProperties(impl.Properties());
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
- private:
-  // Allows access during write.
-  template <class AnyArc, class ArcCompactor, class Unsigned,
-            class CompactStore, class AnyCacheStore>
-  friend class ::fst::CompactFst;  // allow access during write.
-
-  // Current unaligned file format version.
-  static constexpr int kFileVersion = 2;
-  // Current aligned file format version.
-  static constexpr int kAlignedFileVersion = 1;
-  // Minimum file format version supported.
-  static constexpr int kMinFileVersion = 1;
-
-  std::shared_ptr<Compactor> compactor_;
-  typename Compactor::State state_;
-};
-
-template <class Arc, class Compactor, class CacheStore>
-constexpr uint64 CompactFstImpl<Arc, Compactor, CacheStore>::kStaticProperties;
-
-template <class Arc, class Compactor, class CacheStore>
-constexpr int CompactFstImpl<Arc, Compactor, CacheStore>::kFileVersion;
-
-template <class Arc, class Compactor, class CacheStore>
-constexpr int CompactFstImpl<Arc, Compactor, CacheStore>::kAlignedFileVersion;
-
-template <class Arc, class Compactor, class CacheStore>
-constexpr int CompactFstImpl<Arc, Compactor, CacheStore>::kMinFileVersion;
-
-}  // namespace internal
-
-// This class attaches interface to implementation and handles reference
-// counting, delegating most methods to ImplToExpandedFst. The Unsigned type
-// is used to represent indices into the compact arc array. (Template
-// argument defaults are declared in fst-decl.h.)
-template <class A, class ArcCompactor, class Unsigned, class CompactStore,
-          class CacheStore>
-class CompactFst
-    : public ImplToExpandedFst<internal::CompactFstImpl<
-          A,
-          DefaultCompactor<ArcCompactor, Unsigned, CompactStore>,
-          CacheStore>> {
- public:
-  template <class F, class G>
-  void friend Cast(const F &, G *);
-
-  using Arc = A;
-  using StateId = typename A::StateId;
-  using Compactor = DefaultCompactor<ArcCompactor, Unsigned, CompactStore>;
-  using Impl = internal::CompactFstImpl<A, Compactor, CacheStore>;
-  using Store = CacheStore;  // for CacheArcIterator
-
-  friend class StateIterator<
-      CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>>;
-  friend class ArcIterator<
-      CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>>;
-
-  CompactFst() : ImplToExpandedFst<Impl>(std::make_shared<Impl>()) {}
-
-  // If data is not nullptr, it is assumed to be already initialized.
-  explicit CompactFst(
-      const Fst<A> &fst,
-      const ArcCompactor &compactor = ArcCompactor(),
-      const CompactFstOptions &opts = CompactFstOptions(),
-      std::shared_ptr<CompactStore> data = std::shared_ptr<CompactStore>())
-      : ImplToExpandedFst<Impl>(
-            std::make_shared<Impl>(
-                fst,
-                std::make_shared<Compactor>(
-                    std::make_shared<ArcCompactor>(compactor), data),
-                opts)) {}
-
-  // If data is not nullptr, it is assumed to be already initialized.
-  CompactFst(
-      const Fst<Arc> &fst,
-      std::shared_ptr<ArcCompactor> compactor,
-      const CompactFstOptions &opts = CompactFstOptions(),
-      std::shared_ptr<CompactStore> data = std::shared_ptr<CompactStore>())
-      : ImplToExpandedFst<Impl>(
-            std::make_shared<Impl>(fst,
-                                   std::make_shared<Compactor>(compactor, data),
-                                   opts)) {}
-
-  // The following 2 constructors take as input two iterators delimiting a set
-  // of (already) compacted transitions, starting with the transitions out of
-  // the initial state. The format of the input differs for fixed out-degree
-  // and variable out-degree compactors.
-  //
-  // - For fixed out-degree compactors, the final weight (encoded as a
-  // compacted transition) needs to be given only for final states. All strings
-  // (compactor of size 1) will be assume to be terminated by a final state
-  // even when the final state is not implicitely given.
-  //
-  // - For variable out-degree compactors, the final weight (encoded as a
-  // compacted transition) needs to be given for all states and must appeared
-  // first in the list (for state s, final weight of s, followed by outgoing
-  // transitons in s).
-  //
-  // These 2 constructors allows the direct construction of a CompactFst
-  // without first creating a more memory-hungry regular FST. This is useful
-  // when memory usage is severely constrained.
-  template <class Iterator>
-  explicit CompactFst(const Iterator &begin, const Iterator &end,
-                      const ArcCompactor &compactor = ArcCompactor(),
-                      const CompactFstOptions &opts = CompactFstOptions())
-      : ImplToExpandedFst<Impl>(
-            std::make_shared<Impl>(
-                std::make_shared<Compactor>(
-                    begin, end, std::make_shared<ArcCompactor>(compactor)),
-                opts)) {}
-
-  template <class Iterator>
-  CompactFst(const Iterator &begin, const Iterator &end,
-             std::shared_ptr<ArcCompactor> compactor,
-             const CompactFstOptions &opts = CompactFstOptions())
-      : ImplToExpandedFst<Impl>(
-            std::make_shared<Impl>(
-                std::make_shared<Compactor>(begin, end, compactor), opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  CompactFst(
-      const CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>
-      &fst,
-      bool safe = false)
-      : ImplToExpandedFst<Impl>(fst, safe) {}
-
-  // Get a copy of this CompactFst. See Fst<>::Copy() for further doc.
-  CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore> *Copy(
-      bool safe = false) const override {
-    return new CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>(
-        *this, safe);
-  }
-
-  // Read a CompactFst from an input stream; return nullptr on error
-  static CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore> *Read(
-      std::istream &strm, const FstReadOptions &opts) {
-    auto *impl = Impl::Read(strm, opts);
-    return impl ? new CompactFst<A, ArcCompactor, Unsigned, CompactStore,
-                                 CacheStore>(std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  // Read a CompactFst from a file; return nullptr on error
-  // Empty filename reads from standard input
-  static CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore> *Read(
-      const string &filename) {
-    auto *impl = ImplToExpandedFst<Impl>::Read(filename);
-    return impl ? new CompactFst<A, ArcCompactor, Unsigned, CompactStore,
-                                 CacheStore>(std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return GetImpl()->Write(strm, opts);
-  }
-
-  bool Write(const string &filename) const override {
-    return Fst<Arc>::WriteFile(filename);
-  }
-
-  template <class FST>
-  static bool WriteFst(const FST &fst, const ArcCompactor &compactor,
-                       std::ostream &strm, const FstWriteOptions &opts);
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const override {
-    GetImpl()->InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
-  MatcherBase<Arc> *InitMatcher(MatchType match_type) const override {
-    return new SortedMatcher<
-        CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>>(
-        *this, match_type);
-  }
-
-  template <class Iterator>
-  void SetCompactElements(const Iterator &b, const Iterator &e) {
-    GetMutableImpl()->SetCompactor(std::make_shared<Compactor>(
-        b, e, std::make_shared<ArcCompactor>()));
-  }
-
- private:
-  using ImplToFst<Impl, ExpandedFst<Arc>>::GetImpl;
-  using ImplToFst<Impl, ExpandedFst<Arc>>::GetMutableImpl;
-
-  explicit CompactFst(std::shared_ptr<Impl> impl)
-      : ImplToExpandedFst<Impl>(impl) {}
-
-  // Use overloading to extract the type of the argument.
-  static Impl *GetImplIfCompactFst(
-      const CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>
-          &compact_fst) {
-    return compact_fst.GetImpl();
-  }
-
-  // This does not give privileged treatment to subclasses of CompactFst.
-  template <typename NonCompactFst>
-  static Impl *GetImplIfCompactFst(const NonCompactFst &fst) {
-    return nullptr;
-  }
-
-  CompactFst &operator=(const CompactFst &fst) = delete;
-};
-
-// Writes FST in Compact format, with a possible pass over the machine before
-// writing to compute the number of states and arcs.
-template <class A, class ArcCompactor, class Unsigned, class CompactStore,
-          class CacheStore>
-template <class FST>
-bool CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>::WriteFst(
-    const FST &fst, const ArcCompactor &compactor, std::ostream &strm,
-    const FstWriteOptions &opts) {
-  using Arc = A;
-  using Weight = typename A::Weight;
-  using Element = typename ArcCompactor::Element;
-  const auto file_version =
-      opts.align ? Impl::kAlignedFileVersion : Impl::kFileVersion;
-  size_t num_arcs = -1;
-  size_t num_states = -1;
-  auto first_pass_compactor = compactor;
-  if (auto *impl = GetImplIfCompactFst(fst)) {
-    num_arcs = impl->GetCompactor()->GetCompactStore()->NumArcs();
-    num_states = impl->GetCompactor()->GetCompactStore()->NumStates();
-    first_pass_compactor = *impl->GetCompactor()->GetArcCompactor();
-  } else {
-    // A first pass is needed to compute the state of the compactor, which
-    // is saved ahead of the rest of the data structures. This unfortunately
-    // means forcing a complete double compaction when writing in this format.
-    // TODO(allauzen): eliminate mutable state from compactors.
-    num_arcs = 0;
-    num_states = 0;
-    for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      ++num_states;
-      if (fst.Final(s) != Weight::Zero()) {
-        first_pass_compactor.Compact(
-            s, Arc(kNoLabel, kNoLabel, fst.Final(s), kNoStateId));
-      }
-      for (ArcIterator<FST> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-        ++num_arcs;
-        first_pass_compactor.Compact(s, aiter.Value());
-      }
-    }
-  }
-  FstHeader hdr;
-  hdr.SetStart(fst.Start());
-  hdr.SetNumStates(num_states);
-  hdr.SetNumArcs(num_arcs);
-  string type = "compact";
-  if (sizeof(Unsigned) != sizeof(uint32)) {
-    type += std::to_string(CHAR_BIT * sizeof(Unsigned));
-  }
-  type += "_";
-  type += ArcCompactor::Type();
-  if (CompactStore::Type() != "compact") {
-    type += "_";
-    type += CompactStore::Type();
-  }
-  const auto copy_properties = fst.Properties(kCopyProperties, true);
-  if ((copy_properties & kError) || !compactor.Compatible(fst)) {
-    FSTERROR() << "Fst incompatible with compactor";
-    return false;
-  }
-  uint64 properties = copy_properties | Impl::kStaticProperties;
-  internal::FstImpl<Arc>::WriteFstHeader(fst, strm, opts, file_version, type,
-                                         properties, &hdr);
-  first_pass_compactor.Write(strm);
-  if (first_pass_compactor.Size() == -1) {
-    if (opts.align && !AlignOutput(strm)) {
-      LOG(ERROR) << "CompactFst::Write: Alignment failed: " << opts.source;
-      return false;
-    }
-    Unsigned compacts = 0;
-    for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      strm.write(reinterpret_cast<const char *>(&compacts), sizeof(compacts));
-      if (fst.Final(s) != Weight::Zero()) {
-        ++compacts;
-      }
-      compacts += fst.NumArcs(s);
-    }
-    strm.write(reinterpret_cast<const char *>(&compacts), sizeof(compacts));
-  }
-  if (opts.align && !AlignOutput(strm)) {
-    LOG(ERROR) << "Could not align file during write after writing states";
-  }
-  const auto &second_pass_compactor = compactor;
-  Element element;
-  for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    if (fst.Final(s) != Weight::Zero()) {
-      element = second_pass_compactor.Compact(
-          s, A(kNoLabel, kNoLabel, fst.Final(s), kNoStateId));
-      strm.write(reinterpret_cast<const char *>(&element), sizeof(element));
-    }
-    for (ArcIterator<FST> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-      element = second_pass_compactor.Compact(s, aiter.Value());
-      strm.write(reinterpret_cast<const char *>(&element), sizeof(element));
-    }
-  }
-  strm.flush();
-  if (!strm) {
-    LOG(ERROR) << "CompactFst write failed: " << opts.source;
-    return false;
-  }
-  return true;
-}
-
-// Specialization for CompactFst; see generic version in fst.h for sample
-// usage (but use the CompactFst type!). This version should inline.
-template <class Arc, class ArcCompactor, class Unsigned, class CompactStore,
-          class CacheStore>
-class StateIterator<
-    CompactFst<Arc, ArcCompactor, Unsigned, CompactStore, CacheStore>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  explicit StateIterator(
-      const CompactFst<Arc, ArcCompactor, Unsigned, CompactStore,
-                       CacheStore> &fst)
-      : nstates_(fst.GetImpl()->NumStates()), s_(0) {}
-
-  bool Done() const { return s_ >= nstates_; }
-
-  StateId Value() const { return s_; }
-
-  void Next() { ++s_; }
-
-  void Reset() { s_ = 0; }
-
- private:
-  StateId nstates_;
-  StateId s_;
-};
-
-// Specialization for CompactFst. Never caches,
-// always iterates over the underlying compact elements.
-template <class Arc, class ArcCompactor, class Unsigned,
-          class CompactStore, class CacheStore>
-class ArcIterator<CompactFst<
-    Arc, ArcCompactor, Unsigned, CompactStore, CacheStore>> {
- public:
-  using StateId = typename Arc::StateId;
-  using Element = typename ArcCompactor::Element;
-  using Compactor = DefaultCompactor<ArcCompactor, Unsigned, CompactStore>;
-  using State = typename Compactor::State;
-
-  ArcIterator(const CompactFst<Arc, ArcCompactor, Unsigned, CompactStore,
-                               CacheStore> &fst,
-              StateId s)
-      : state_(fst.GetImpl()->GetCompactor(), s),
-        pos_(0),
-        flags_(kArcValueFlags) {}
-
-  bool Done() const { return pos_ >= state_.NumArcs(); }
-
-  const Arc &Value() const {
-    arc_ = state_.GetArc(pos_, flags_);
-    return arc_;
-  }
-
-  void Next() { ++pos_; }
-
-  size_t Position() const { return pos_; }
-
-  void Reset() { pos_ = 0; }
-
-  void Seek(size_t pos) { pos_ = pos; }
-
-  uint32 Flags() const { return flags_; }
-
-  void SetFlags(uint32 f, uint32 m) {
-    flags_ &= ~m;
-    flags_ |= (f & kArcValueFlags);
-  }
-
- private:
-  State state_;
-  size_t pos_;
-  mutable Arc arc_;
-  uint32 flags_;
-};
-
-// ArcCompactor for unweighted string FSTs.
-template <class A>
-class StringCompactor {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Element = Label;
-
-  Element Compact(StateId s, const Arc &arc) const { return arc.ilabel; }
-
-  Arc Expand(StateId s, const Element &p, uint32 f = kArcValueFlags) const {
-    return Arc(p, p, Weight::One(), p != kNoLabel ? s + 1 : kNoStateId);
-  }
-
-  constexpr ssize_t Size() const { return 1; }
-
-  constexpr uint64 Properties() const {
-    return kString | kAcceptor | kUnweighted;
-  }
-
-  bool Compatible(const Fst<Arc> &fst) const {
-    const auto props = Properties();
-    return fst.Properties(props, true) == props;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string("string");
-    return *type;
-  }
-
-  bool Write(std::ostream &strm) const { return true; }
-
-  static StringCompactor *Read(std::istream &strm) {
-    return new StringCompactor;
-  }
-};
-
-// ArcCompactor for weighted string FSTs.
-template <class A>
-class WeightedStringCompactor {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Element = std::pair<Label, Weight>;
-
-  Element Compact(StateId s, const Arc &arc) const {
-    return std::make_pair(arc.ilabel, arc.weight);
-  }
-
-  Arc Expand(StateId s, const Element &p, uint32 f = kArcValueFlags) const {
-    return Arc(p.first, p.first, p.second,
-               p.first != kNoLabel ? s + 1 : kNoStateId);
-  }
-
-  constexpr ssize_t Size() const { return 1; }
-
-  constexpr uint64 Properties() const { return kString | kAcceptor; }
-
-  bool Compatible(const Fst<Arc> &fst) const {
-    const auto props = Properties();
-    return fst.Properties(props, true) == props;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string("weighted_string");
-    return *type;
-  }
-
-  bool Write(std::ostream &strm) const { return true; }
-
-  static WeightedStringCompactor *Read(std::istream &strm) {
-    return new WeightedStringCompactor;
-  }
-};
-
-// ArcCompactor for unweighted acceptor FSTs.
-template <class A>
-class UnweightedAcceptorCompactor {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Element = std::pair<Label, StateId>;
-
-  Element Compact(StateId s, const Arc &arc) const {
-    return std::make_pair(arc.ilabel, arc.nextstate);
-  }
-
-  Arc Expand(StateId s, const Element &p, uint32 f = kArcValueFlags) const {
-    return Arc(p.first, p.first, Weight::One(), p.second);
-  }
-
-  constexpr ssize_t Size() const { return -1; }
-
-  constexpr uint64 Properties() const { return kAcceptor | kUnweighted; }
-
-  bool Compatible(const Fst<Arc> &fst) const {
-    const auto props = Properties();
-    return fst.Properties(props, true) == props;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string("unweighted_acceptor");
-    return *type;
-  }
-
-  bool Write(std::ostream &strm) const { return true; }
-
-  static UnweightedAcceptorCompactor *Read(std::istream &istrm) {
-    return new UnweightedAcceptorCompactor;
-  }
-};
-
-// ArcCompactor for weighted acceptor FSTs.
-template <class A>
-class AcceptorCompactor {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Element = std::pair<std::pair<Label, Weight>, StateId>;
-
-  Element Compact(StateId s, const Arc &arc) const {
-    return std::make_pair(std::make_pair(arc.ilabel, arc.weight),
-                          arc.nextstate);
-  }
-
-  Arc Expand(StateId s, const Element &p, uint32 f = kArcValueFlags) const {
-    return Arc(p.first.first, p.first.first, p.first.second, p.second);
-  }
-
-  constexpr ssize_t Size() const { return -1; }
-
-  constexpr uint64 Properties() const { return kAcceptor; }
-
-  bool Compatible(const Fst<Arc> &fst) const {
-    const auto props = Properties();
-    return fst.Properties(props, true) == props;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string("acceptor");
-    return *type;
-  }
-
-  bool Write(std::ostream &strm) const { return true; }
-
-  static AcceptorCompactor *Read(std::istream &strm) {
-    return new AcceptorCompactor;
-  }
-};
-
-// ArcCompactor for unweighted FSTs.
-template <class A>
-class UnweightedCompactor {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Element = std::pair<std::pair<Label, Label>, StateId>;
-
-  Element Compact(StateId s, const Arc &arc) const {
-    return std::make_pair(std::make_pair(arc.ilabel, arc.olabel),
-                          arc.nextstate);
-  }
-
-  Arc Expand(StateId s, const Element &p, uint32 f = kArcValueFlags) const {
-    return Arc(p.first.first, p.first.second, Weight::One(), p.second);
-  }
-
-  constexpr ssize_t Size() const { return -1; }
-
-  constexpr uint64 Properties() const { return kUnweighted; }
-
-  bool Compatible(const Fst<Arc> &fst) const {
-    const auto props = Properties();
-    return fst.Properties(props, true) == props;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string("unweighted");
-    return *type;
-  }
-
-  bool Write(std::ostream &strm) const { return true; }
-
-  static UnweightedCompactor *Read(std::istream &strm) {
-    return new UnweightedCompactor;
-  }
-};
-
-template <class Arc, class Unsigned /* = uint32 */>
-using CompactStringFst = CompactFst<Arc, StringCompactor<Arc>, Unsigned>;
-
-template <class Arc, class Unsigned /* = uint32 */>
-using CompactWeightedStringFst =
-    CompactFst<Arc, WeightedStringCompactor<Arc>, Unsigned>;
-
-template <class Arc, class Unsigned /* = uint32 */>
-using CompactAcceptorFst = CompactFst<Arc, AcceptorCompactor<Arc>, Unsigned>;
-
-template <class Arc, class Unsigned /* = uint32 */>
-using CompactUnweightedFst =
-    CompactFst<Arc, UnweightedCompactor<Arc>, Unsigned>;
-
-template <class Arc, class Unsigned /* = uint32 */>
-using CompactUnweightedAcceptorFst =
-    CompactFst<Arc, UnweightedAcceptorCompactor<Arc>, Unsigned>;
-
-using StdCompactStringFst = CompactStringFst<StdArc, uint32>;
-
-using StdCompactWeightedStringFst = CompactWeightedStringFst<StdArc, uint32>;
-
-using StdCompactAcceptorFst = CompactAcceptorFst<StdArc, uint32>;
-
-using StdCompactUnweightedFst = CompactUnweightedFst<StdArc, uint32>;
-
-using StdCompactUnweightedAcceptorFst =
-    CompactUnweightedAcceptorFst<StdArc, uint32>;
-
-}  // namespace fst
-
-#endif  // FST_COMPACT_FST_H_
diff --git a/projects/llm_framework/include/fst/compat.h b/projects/llm_framework/include/fst/compat.h
deleted file mode 100644
index 73ed5737..00000000
--- a/projects/llm_framework/include/fst/compat.h
+++ /dev/null
@@ -1,130 +0,0 @@
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_LIB_COMPAT_H_
-#define FST_LIB_COMPAT_H_
-
-#include <climits>
-#include <cstdlib>
-#include <cstring>
-#include <iostream>
-#include <string>
-#include <vector>
-
-// Makes copy constructor and operator= private
-// Deprecated: now just use =delete.
-#define DISALLOW_COPY_AND_ASSIGN(type)    \
-  type(const type&);                      \
-  void operator=(const type&)
-
-#if defined(__GNUC__) || defined(__clang__)
-#define OPENFST_DEPRECATED(message) __attribute__((deprecated(message)))
-#elif defined(_MSC_VER)
-#define OPENFST_DEPRECATED(message) __declspec(deprecated(message))
-#else
-#define OPENFST_DEPRECATED(message)
-#endif
-
-#include <fst/config.h>
-#include <fst/types.h>
-#include <fst/lock.h>
-#include <fst/flags.h>
-#include <fst/log.h>
-#include <fst/icu.h>
-
-using std::string;
-
-void FailedNewHandler();
-
-#ifdef _MSC_VER
-#include <intrin.h>
-const char* basename(const char* path);
-#define __builtin_popcount __popcnt
-
-#ifdef _M_X64
-// Using 64-bit MSVC intrinsics.
-#define __builtin_popcountll __popcnt64
-inline unsigned int __builtin_ctzll(std::uint64_t w) {
-  unsigned long v;
-  return _BitScanForward64(&v, std::uint32_t(w)) ? v : 0;
-}
-#else
-// Using 32-bit MSVC intrinsics.
-inline unsigned int __builtin_popcountll(std::uint64_t w) {
-  return __popcnt(std::uint32_t(w)) + __popcnt(std::uint32_t(w >> 32));
-}
-inline unsigned int __builtin_ctzll(std::uint64_t w) {
-  unsigned long v;
-  return (_BitScanForward(&v, std::uint32_t(w)) ? v :
-          _BitScanForward(&v, std::uint32_t(w >> 32)) ? v + 32 : 0);
-}
-#endif  // _M_X64
-#endif  // _MSC_VER
-
-namespace fst {
-
-// Downcasting.
-template<typename To, typename From>
-inline To down_cast(From* f) { return static_cast<To>(f); }
-
-// Bitcasting.
-template <class Dest, class Source>
-inline Dest bit_cast(const Source &source) {
-  static_assert(sizeof(Dest) == sizeof(Source),
-                "Bitcasting unsafe for specified types");
-  Dest dest;
-  memcpy(&dest, &source, sizeof(dest));
-  return dest;
-}
-
-// Check sums
-class CheckSummer {
- public:
-  CheckSummer() : count_(0) {
-    check_sum_.resize(kCheckSumLength, '\0');
-  }
-
-  void Reset() {
-    count_ = 0;
-    for (int i = 0; i < kCheckSumLength; ++i) check_sum_[i] = '\0';
-  }
-
-  void Update(void const *data, int size) {
-    const char *p = reinterpret_cast<const char *>(data);
-    for (int i = 0; i < size; ++i) {
-      check_sum_[(count_++) % kCheckSumLength] ^= p[i];
-    }
-  }
-
-  void Update(string const &data) {
-    for (int i = 0; i < data.size(); ++i) {
-      check_sum_[(count_++) % kCheckSumLength] ^= data[i];
-    }
-  }
-
-  string Digest() { return check_sum_; }
-
- private:
-  static const int kCheckSumLength = 32;
-  int count_;
-  string check_sum_;
-
-  CheckSummer(const CheckSummer &) = delete;
-  CheckSummer &operator=(const CheckSummer &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_LIB_COMPAT_H_
diff --git a/projects/llm_framework/include/fst/complement.h b/projects/llm_framework/include/fst/complement.h
deleted file mode 100644
index 64eebc03..00000000
--- a/projects/llm_framework/include/fst/complement.h
+++ /dev/null
@@ -1,277 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to complement an FST.
-
-#ifndef FST_COMPLEMENT_H_
-#define FST_COMPLEMENT_H_
-
-#include <algorithm>
-#include <string>
-#include <vector>
-#include <fst/log.h>
-
-#include <fst/fst.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-template <class Arc>
-class ComplementFst;
-
-namespace internal {
-
-// Implementation of delayed ComplementFst. The algorithm used completes the
-// (deterministic) FSA and then exchanges final and non-final states.
-// Completion, i.e. ensuring that all labels can be read from every state, is
-// accomplished by using ρ-labels, which match all labels that are otherwise
-// not found leaving a state. The first state in the output is reserved to be a
-// new state that is the destination of all ρ-labels. Each remaining output
-// state s corresponds to input state s - 1. The first arc in the output at
-// these states is the ρ-label, the remaining arcs correspond to the input
-// arcs.
-template <class A>
-class ComplementFstImpl : public FstImpl<A> {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<A>::SetType;
-  using FstImpl<A>::SetProperties;
-  using FstImpl<A>::SetInputSymbols;
-  using FstImpl<A>::SetOutputSymbols;
-
-  friend class StateIterator<ComplementFst<Arc>>;
-  friend class ArcIterator<ComplementFst<Arc>>;
-
-  explicit ComplementFstImpl(const Fst<Arc> &fst) : fst_(fst.Copy()) {
-    SetType("complement");
-    uint64 props = fst.Properties(kILabelSorted, false);
-    SetProperties(ComplementProperties(props), kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  ComplementFstImpl(const ComplementFstImpl<Arc> &impl)
-      : fst_(impl.fst_->Copy()) {
-    SetType("complement");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  StateId Start() const {
-    if (Properties(kError)) return kNoStateId;
-    auto start = fst_->Start();
-    return start != kNoStateId ? start + 1 : 0;
-  }
-
-  // Exchange final and non-final states; makes ρ-destination state final.
-  Weight Final(StateId s) const {
-    if (s == 0 || fst_->Final(s - 1) == Weight::Zero()) {
-      return Weight::One();
-    } else {
-      return Weight::Zero();
-    }
-  }
-
-  size_t NumArcs(StateId s) const {
-    return s == 0 ? 1 : fst_->NumArcs(s - 1) + 1;
-  }
-
-  size_t NumInputEpsilons(StateId s) const {
-    return s == 0 ? 0 : fst_->NumInputEpsilons(s - 1);
-  }
-
-  size_t NumOutputEpsilons(StateId s) const {
-    return s == 0 ? 0 : fst_->NumOutputEpsilons(s - 1);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && fst_->Properties(kError, false)) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
- private:
-  std::unique_ptr<const Fst<Arc>> fst_;
-};
-
-}  // namespace internal
-
-// Complements an automaton. This is a library-internal operation that
-// introduces a (negative) ρ-label; use Difference/DifferenceFst in user code,
-// which will not see this label. This version is a delayed FST.
-//
-// This class attaches interface to implementation and handles
-// reference counting, delegating most methods to ImplToFst.
-template <class A>
-class ComplementFst : public ImplToFst<internal::ComplementFstImpl<A>> {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Impl = internal::ComplementFstImpl<Arc>;
-
-  friend class StateIterator<ComplementFst<Arc>>;
-  friend class ArcIterator<ComplementFst<Arc>>;
-
-  explicit ComplementFst(const Fst<Arc> &fst)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst)) {
-    static constexpr auto props =
-        kUnweighted | kNoEpsilons | kIDeterministic | kAcceptor;
-    if (fst.Properties(props, true) != props) {
-      FSTERROR() << "ComplementFst: Argument not an unweighted "
-                 << "epsilon-free deterministic acceptor";
-      GetImpl()->SetProperties(kError, kError);
-    }
-  }
-
-  // See Fst<>::Copy() for doc.
-  ComplementFst(const ComplementFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Gets a copy of this FST. See Fst<>::Copy() for further doc.
-  ComplementFst<Arc> *Copy(bool safe = false) const override {
-    return new ComplementFst<Arc>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  inline void InitArcIterator(StateId s,
-                              ArcIteratorData<Arc> *data) const override;
-
-  // Label that represents the ρ-transition; we use a negative value private to
-  // the library and which will preserve FST label sort order.
-  static const Label kRhoLabel = -2;
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-
-  ComplementFst &operator=(const ComplementFst &) = delete;
-};
-
-template <class Arc>
-const typename Arc::Label ComplementFst<Arc>::kRhoLabel;
-
-// Specialization for ComplementFst.
-template <class Arc>
-class StateIterator<ComplementFst<Arc>> : public StateIteratorBase<Arc> {
- public:
-  using StateId = typename Arc::StateId;
-
-  explicit StateIterator(const ComplementFst<Arc> &fst)
-      : siter_(*fst.GetImpl()->fst_), s_(0) {}
-
-  bool Done() const final { return s_ > 0 && siter_.Done(); }
-
-  StateId Value() const final { return s_; }
-
-  void Next() final {
-    if (s_ != 0) siter_.Next();
-    ++s_;
-  }
-
-  void Reset() final {
-    siter_.Reset();
-    s_ = 0;
-  }
-
- private:
-  StateIterator<Fst<Arc>> siter_;
-  StateId s_;
-};
-
-// Specialization for ComplementFst.
-template <class Arc>
-class ArcIterator<ComplementFst<Arc>> : public ArcIteratorBase<Arc> {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  ArcIterator(const ComplementFst<Arc> &fst, StateId s) : s_(s), pos_(0) {
-    if (s_ != 0) {
-      aiter_.reset(new ArcIterator<Fst<Arc>>(*fst.GetImpl()->fst_, s - 1));
-    }
-  }
-
-  bool Done() const final {
-    if (s_ != 0) {
-      return pos_ > 0 && aiter_->Done();
-    } else {
-      return pos_ > 0;
-    }
-  }
-
-  // Adds the ρ-label to the ρ destination state.
-  const Arc &Value() const final {
-    if (pos_ == 0) {
-      arc_.ilabel = arc_.olabel = ComplementFst<Arc>::kRhoLabel;
-      arc_.weight = Weight::One();
-      arc_.nextstate = 0;
-    } else {
-      arc_ = aiter_->Value();
-      ++arc_.nextstate;
-    }
-    return arc_;
-  }
-
-  void Next() final {
-    if (s_ != 0 && pos_ > 0) aiter_->Next();
-    ++pos_;
-  }
-
-  size_t Position() const final { return pos_; }
-
-  void Reset() final {
-    if (s_ != 0) aiter_->Reset();
-    pos_ = 0;
-  }
-
-  void Seek(size_t a) final {
-    if (s_ != 0) {
-      if (a == 0) {
-        aiter_->Reset();
-      } else {
-        aiter_->Seek(a - 1);
-      }
-    }
-    pos_ = a;
-  }
-
-  uint32 Flags() const final { return kArcValueFlags; }
-
-  void SetFlags(uint32, uint32) final {}
-
- private:
-  std::unique_ptr<ArcIterator<Fst<Arc>>> aiter_;
-  StateId s_;
-  size_t pos_;
-  mutable Arc arc_;
-};
-
-template <class Arc>
-inline void ComplementFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<ComplementFst<Arc>>(*this);
-}
-
-template <class Arc>
-inline void ComplementFst<Arc>::InitArcIterator(StateId s,
-    ArcIteratorData<Arc> *data) const {
-  data->base = new ArcIterator<ComplementFst<Arc>>(*this, s);
-}
-
-// Useful alias when using StdArc.
-using StdComplementFst = ComplementFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_COMPLEMENT_H_
diff --git a/projects/llm_framework/include/fst/compose-filter.h b/projects/llm_framework/include/fst/compose-filter.h
deleted file mode 100644
index 7251e273..00000000
--- a/projects/llm_framework/include/fst/compose-filter.h
+++ /dev/null
@@ -1,571 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes for filtering the composition matches, e.g. for correct epsilon
-// handling.
-
-#ifndef FST_COMPOSE_FILTER_H_
-#define FST_COMPOSE_FILTER_H_
-
-#include <fst/filter-state.h>
-#include <fst/fst-decl.h>  // For optional argument declarations
-#include <fst/fst.h>
-#include <fst/matcher.h>
-
-
-namespace fst {
-
-// Composition filters determine which matches are allowed to proceed. The
-// filter's state is represeted by the type ComposeFilter::FilterState.
-// The basic filters handle correct epsilon matching. Their interface is:
-//
-// template <class M1, class M2>
-// class ComposeFilter {
-//  public:
-//   using Matcher1 = ...;
-//   using Matcher2 = ...;
-//   using FST1 = typename M1::FST;
-//   using FST2 = typename M2::FST;
-//   using FilterState = ...;
-//
-//   using Arc = typename FST1::Arc;
-//   using StateId = typename Arc::StateId;
-//   using Weight = typename Arc::Weight;
-//
-//   // Required constructor.
-//   ComposeFilter(const FST1 &fst1, const FST2 &fst2,
-//                 M1 *matcher1 = nullptr, M2 *matcher2 = nullptr);
-//
-//   // If safe=true, the copy is thread-safe. See Fst<>::Copy()
-//   // for further doc.
-//   ComposeFilter(const ComposeFilter<M1, M2> &filter,
-//                 bool safe = false);
-//
-//   // Return start state of filter.
-//   FilterState Start() const;
-//
-//   // Specifies current composition state.
-//   void SetState(StateId s1, StateId s2, const FilterState &fs);
-//
-//   // Apply filter at current composition state to these transitions. If an
-//   // arc label to be matched is kNolabel, then that side does not consume a
-//   // symbol. Returns the new filter state or, if disallowed,
-//   // FilterState::NoState(). The filter is permitted to modify its inputs
-//   // (e.g. for optimization reasons).
-//   FilterState FilterArc(Arc *arc1, Arc *arc2) const;
-
-//   // Apply filter at current composition state to these final weights
-//   // (cf. superfinal transitions). The filter may modify its inputs
-//   // (e.g. for optimization reasons).
-//   void FilterFinal(Weight *w1, Weight *w2) const;
-//
-//   // Return the respective matchers. Ownership stays with filter. These
-//   // methods allow the filter to access and possibly modify the compositio
-//   // matchers (useful, e.g., with lookahead).
-//
-//   Matcher1 *GetMatcher1();
-//
-//   Matcher2 *GetMatcher2();
-//
-//   // This specifies how the filter affects the composition result properties.
-//   It takes as argument the properties that would apply with a trivial
-//   // composition filter.
-//   uint64 Properties(uint64 props) const;
-// };
-//
-// This filter allows only exact matching of symbols from FST1 with on FST2;
-// e.g., no special interpretation of epsilons.
-template <class M1, class M2 /* = M1 */>
-class NullComposeFilter {
- public:
-  using Matcher1 = M1;
-  using Matcher2 = M2;
-  using FST1 = typename M1::FST;
-  using FST2 = typename M2::FST;
-  using FilterState = TrivialFilterState;
-
-  using Arc = typename FST1::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  NullComposeFilter(const FST1 &fst1, const FST2 &fst2,
-                    Matcher1 *matcher1 = nullptr, Matcher2 *matcher2 = nullptr)
-      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
-        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()) {}
-
-  NullComposeFilter(const NullComposeFilter<M1, M2> &filter, bool safe = false)
-      : matcher1_(filter.matcher1_->Copy(safe)),
-        matcher2_(filter.matcher2_->Copy(safe)),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()) {}
-
-  FilterState Start() const { return FilterState(true); }
-
-  void SetState(StateId, StateId, const FilterState &) {}
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    return (arc1->olabel == kNoLabel || arc2->ilabel == kNoLabel)
-               ? FilterState::NoState()
-               : FilterState(true);
-  }
-
-  void FilterFinal(Weight *, Weight *) const {}
-
-  Matcher1 *GetMatcher1() { return matcher1_.get(); }
-
-  Matcher2 *GetMatcher2() { return matcher2_.get(); }
-
-  uint64 Properties(uint64 props) const { return props; }
-
- private:
-  std::unique_ptr<Matcher1> matcher1_;
-  std::unique_ptr<Matcher2> matcher2_;
-  const FST1 &fst1_;
-  const FST2 &fst2_;
-};
-
-// This filter allows all epsilon matches, potentially resulting in redundant
-// epsilon paths. The use of this filter gives correct results iff one of the
-// following conditions hold:
-//
-//  (1) The semiring is idempotent,
-//  (2) the first FST is output-epsilon free, or
-//  (3) the second FST is input-epsilon free.
-//
-// For (1), redundant epsilon paths may be created but won't hurt correctness.
-// For (2) and (3), no redundant paths are created.
-template <class M1, class M2 /* = M1 */>
-class TrivialComposeFilter {
- public:
-  using Matcher1 = M1;
-  using Matcher2 = M2;
-  using FST1 = typename M1::FST;
-  using FST2 = typename M2::FST;
-  using FilterState = TrivialFilterState;
-
-  using Arc = typename FST1::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  TrivialComposeFilter(const FST1 &fst1, const FST2 &fst2,
-                       Matcher1 *matcher1 = nullptr,
-                       Matcher2 *matcher2 = nullptr)
-      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
-        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()) {}
-
-  TrivialComposeFilter(const TrivialComposeFilter<Matcher1, Matcher2> &filter,
-                       bool safe = false)
-      : matcher1_(filter.matcher1_->Copy(safe)),
-        matcher2_(filter.matcher2_->Copy(safe)),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()) {}
-
-  FilterState Start() const { return FilterState(true); }
-
-  void SetState(StateId, StateId, const FilterState &) {}
-
-  FilterState FilterArc(Arc *, Arc *) const { return FilterState(true); }
-
-  void FilterFinal(Weight *, Weight *) const {}
-
-  Matcher1 *GetMatcher1() { return matcher1_.get(); }
-
-  Matcher2 *GetMatcher2() { return matcher2_.get(); }
-
-  uint64 Properties(uint64 props) const { return props; }
-
- private:
-  std::unique_ptr<Matcher1> matcher1_;
-  std::unique_ptr<Matcher2> matcher2_;
-  const FST1 &fst1_;
-  const FST2 &fst2_;
-};
-
-// This filter requires epsilons on FST1 to be read before epsilons on FST2.
-template <class M1, class M2 /* = M1 */>
-class SequenceComposeFilter {
- public:
-  using Matcher1 = M1;
-  using Matcher2 = M2;
-  using FST1 = typename M1::FST;
-  using FST2 = typename M2::FST;
-  using FilterState = CharFilterState;
-
-  using Arc = typename FST1::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  SequenceComposeFilter(const FST1 &fst1, const FST2 &fst2,
-                        Matcher1 *matcher1 = nullptr,
-                        Matcher2 *matcher2 = nullptr)
-      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
-        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
-        fst1_(matcher1_->GetFst()),
-        s1_(kNoStateId),
-        s2_(kNoStateId),
-        fs_(kNoStateId) {}
-
-  SequenceComposeFilter(const SequenceComposeFilter<Matcher1, Matcher2> &filter,
-                        bool safe = false)
-      : matcher1_(filter.matcher1_->Copy(safe)),
-        matcher2_(filter.matcher2_->Copy(safe)),
-        fst1_(matcher1_->GetFst()),
-        s1_(kNoStateId),
-        s2_(kNoStateId),
-        fs_(kNoStateId) {}
-
-  FilterState Start() const { return FilterState(0); }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    if (s1_ == s1 && s2_ == s2 && fs == fs_) return;
-    s1_ = s1;
-    s2_ = s2;
-    fs_ = fs;
-    const auto na1 = internal::NumArcs(fst1_, s1);
-    const auto ne1 = internal::NumOutputEpsilons(fst1_, s1);
-    const bool fin1 = internal::Final(fst1_, s1) != Weight::Zero();
-    alleps1_ = na1 == ne1 && !fin1;
-    noeps1_ = ne1 == 0;
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    if (arc1->olabel == kNoLabel) {
-      return alleps1_ ? FilterState::NoState() : noeps1_ ? FilterState(0)
-                                                         : FilterState(1);
-    } else if (arc2->ilabel == kNoLabel) {
-      return fs_ != FilterState(0) ? FilterState::NoState() : FilterState(0);
-    } else {
-      return arc1->olabel == 0 ? FilterState::NoState() : FilterState(0);
-    }
-  }
-
-  void FilterFinal(Weight *, Weight *) const {}
-
-  Matcher1 *GetMatcher1() { return matcher1_.get(); }
-
-  Matcher2 *GetMatcher2() { return matcher2_.get(); }
-
-  uint64 Properties(uint64 props) const { return props; }
-
- private:
-  std::unique_ptr<Matcher1> matcher1_;
-  std::unique_ptr<Matcher2> matcher2_;
-  const FST1 &fst1_;
-  StateId s1_;      // Current fst1_ state.
-  StateId s2_;      // Current fst2_ state.
-  FilterState fs_;  // Current filter state.
-  bool alleps1_;   // Only epsilons (and non-final) leaving s1_?
-  bool noeps1_;    // No epsilons leaving s1_?
-};
-
-// This filter requires epsilons on FST2 to be read before epsilons on FST1.
-template <class M1, class M2 /* = M1 */>
-class AltSequenceComposeFilter {
- public:
-  using Matcher1 = M1;
-  using Matcher2 = M2;
-  using FST1 = typename M1::FST;
-  using FST2 = typename M2::FST;
-  using FilterState = CharFilterState;
-
-  using Arc = typename FST1::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  AltSequenceComposeFilter(const FST1 &fst1, const FST2 &fst2,
-                           Matcher1 *matcher1 = nullptr,
-                           Matcher2 *matcher2 = nullptr)
-      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
-        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
-        fst2_(matcher2_->GetFst()),
-        s1_(kNoStateId),
-        s2_(kNoStateId),
-        fs_(kNoStateId) {}
-
-  AltSequenceComposeFilter(
-      const AltSequenceComposeFilter<Matcher1, Matcher2> &filter,
-      bool safe = false)
-      : matcher1_(filter.matcher1_->Copy(safe)),
-        matcher2_(filter.matcher2_->Copy(safe)),
-        fst2_(matcher2_->GetFst()),
-        s1_(kNoStateId),
-        s2_(kNoStateId),
-        fs_(kNoStateId) {}
-
-  FilterState Start() const { return FilterState(0); }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    if (s1_ == s1 && s2_ == s2 && fs == fs_) return;
-    s1_ = s1;
-    s2_ = s2;
-    fs_ = fs;
-    const auto na2 = internal::NumArcs(fst2_, s2);
-    const auto ne2 = internal::NumInputEpsilons(fst2_, s2);
-    const bool fin2 = internal::Final(fst2_, s2) != Weight::Zero();
-    alleps2_ = na2 == ne2 && !fin2;
-    noeps2_ = ne2 == 0;
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    if (arc2->ilabel == kNoLabel) {
-      return alleps2_ ? FilterState::NoState() : noeps2_ ? FilterState(0)
-                                                         : FilterState(1);
-    } else if (arc1->olabel == kNoLabel) {
-      return fs_ == FilterState(1) ? FilterState::NoState() : FilterState(0);
-    } else {
-      return arc1->olabel == 0 ? FilterState::NoState() : FilterState(0);
-    }
-  }
-
-  void FilterFinal(Weight *, Weight *) const {}
-
-  Matcher1 *GetMatcher1() { return matcher1_.get(); }
-
-  Matcher2 *GetMatcher2() { return matcher2_.get(); }
-
-  uint64 Properties(uint64 props) const { return props; }
-
- private:
-  std::unique_ptr<Matcher1> matcher1_;
-  std::unique_ptr<Matcher2> matcher2_;
-  const FST2 &fst2_;
-  StateId s1_;      // Current fst1_ state.
-  StateId s2_;      // Current fst2_ state.
-  FilterState fs_;  // Current filter state.
-  bool alleps2_;    // Only epsilons (and non-final) leaving s2_?
-  bool noeps2_;     // No epsilons leaving s2_?
-};
-
-// This filter requires epsilons on FST1 to be matched with epsilons on FST2
-// whenever possible. (Template arg default declared in fst-decl.h.)
-template <class M1, class M2 /* = M1 */>
-class MatchComposeFilter {
- public:
-  using Matcher1 = M1;
-  using Matcher2 = M2;
-  using FST1 = typename M1::FST;
-  using FST2 = typename M2::FST;
-  using FilterState = CharFilterState;
-
-  using Arc = typename FST1::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  MatchComposeFilter(const FST1 &fst1, const FST2 &fst2,
-                     Matcher1 *matcher1 = nullptr, Matcher2 *matcher2 = nullptr)
-      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
-        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()),
-        s1_(kNoStateId),
-        s2_(kNoStateId),
-        fs_(kNoStateId) {}
-
-  MatchComposeFilter(const MatchComposeFilter<Matcher1, Matcher2> &filter,
-                     bool safe = false)
-      : matcher1_(filter.matcher1_->Copy(safe)),
-        matcher2_(filter.matcher2_->Copy(safe)),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()),
-        s1_(kNoStateId),
-        s2_(kNoStateId),
-        fs_(kNoStateId) {}
-
-  FilterState Start() const { return FilterState(0); }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    if (s1_ == s1 && s2_ == s2 && fs == fs_) return;
-    s1_ = s1;
-    s2_ = s2;
-    fs_ = fs;
-    size_t na1 = internal::NumArcs(fst1_, s1);
-    size_t ne1 = internal::NumOutputEpsilons(fst1_, s1);
-    bool f1 = internal::Final(fst1_, s1) != Weight::Zero();
-    alleps1_ = na1 == ne1 && !f1;
-    noeps1_ = ne1 == 0;
-    size_t na2 = internal::NumArcs(fst2_, s2);
-    size_t ne2 = internal::NumInputEpsilons(fst2_, s2);
-    bool f2 = internal::Final(fst2_, s2) != Weight::Zero();
-    alleps2_ = na2 == ne2 && !f2;
-    noeps2_ = ne2 == 0;
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    if (arc2->ilabel == kNoLabel) {  // Epsilon in FST1.
-      return fs_ == FilterState(0)
-                 ? (noeps2_
-                        ? FilterState(0)
-                        : (alleps2_ ? FilterState::NoState() : FilterState(1)))
-                 : (fs_ == FilterState(1) ? FilterState(1)
-                                          : FilterState::NoState());
-    } else if (arc1->olabel == kNoLabel) {  // Epsilon in FST2.
-      return fs_ == FilterState(0)
-                 ? (noeps1_
-                        ? FilterState(0)
-                        : (alleps1_ ? FilterState::NoState() : FilterState(2)))
-                 : (fs_ == FilterState(2) ? FilterState(2)
-                                          : FilterState::NoState());
-    } else if (arc1->olabel == 0) {  // Epsilon in both.
-      return fs_ == FilterState(0) ? FilterState(0) : FilterState::NoState();
-    } else {  // Both are non-epsilons.
-      return FilterState(0);
-    }
-  }
-
-  void FilterFinal(Weight *, Weight *) const {}
-
-  Matcher1 *GetMatcher1() { return matcher1_.get(); }
-
-  Matcher2 *GetMatcher2() { return matcher2_.get(); }
-
-  uint64 Properties(uint64 props) const { return props; }
-
- private:
-  std::unique_ptr<Matcher1> matcher1_;
-  std::unique_ptr<Matcher2> matcher2_;
-  const FST1 &fst1_;
-  const FST2 &fst2_;
-  StateId s1_;      // Current fst1_ state.
-  StateId s2_;      // Current fst2_ state.
-  FilterState fs_;  // Current filter state ID.
-  bool alleps1_;    // Only epsilson (and non-final) leaving s1?
-  bool alleps2_;    // Only epsilons (and non-final) leaving s2?
-  bool noeps1_;     // No epsilons leaving s1?
-  bool noeps2_;     // No epsilons leaving s2?
-};
-
-// This filter disallows matching epsilons on FST1 with epsilons on FST2,
-// but allows all other matches, potentially resulting in redundant
-// epsilon paths. The use of this filter gives correct results iff one of the
-// following conditions hold:
-//
-//  (1) The semiring is idempotent,
-//  (2) the first FST is output-epsilon free, or
-//  (3) the second FST is input-epsilon free.
-//
-// For (1), redundant epsilon paths may be created but won't hurt correctness.
-// For (2) and (3), no redundant paths are created.
-template <class M1, class M2 /* = M1 */>
-class NoMatchComposeFilter {
- public:
-  using Matcher1 = M1;
-  using Matcher2 = M2;
-  using FST1 = typename M1::FST;
-  using FST2 = typename M2::FST;
-  using FilterState = TrivialFilterState;
-
-  using Arc = typename FST1::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  NoMatchComposeFilter(const FST1 &fst1, const FST2 &fst2,
-                       Matcher1 *matcher1 = nullptr,
-                       Matcher2 *matcher2 = nullptr)
-      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
-        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()) {}
-
-  NoMatchComposeFilter(const NoMatchComposeFilter<Matcher1, Matcher2> &filter,
-                       bool safe = false)
-      : matcher1_(filter.matcher1_->Copy(safe)),
-        matcher2_(filter.matcher2_->Copy(safe)),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()) {}
-
-  FilterState Start() const { return FilterState(true); }
-
-  void SetState(StateId, StateId, const FilterState &) {}
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    return FilterState(arc1->olabel != 0 || arc2->ilabel != 0);
-  }
-
-  void FilterFinal(Weight *, Weight *) const {}
-
-  Matcher1 *GetMatcher1() { return matcher1_.get(); }
-
-  Matcher2 *GetMatcher2() { return matcher2_.get(); }
-
-  uint64 Properties(uint64 props) const { return props; }
-
- private:
-  std::unique_ptr<Matcher1> matcher1_;
-  std::unique_ptr<Matcher2> matcher2_;
-  const FST1 &fst1_;
-  const FST2 &fst2_;
-};
-
-// This filter works with the MultiEpsMatcher to determine if multi-epsilons are
-// preserved in the composition output (rather than rewritten as 0) and
-// ensures correct properties.
-template <class Filter>
-class MultiEpsFilter {
- public:
-  using Matcher1 = typename Filter::Matcher1;
-  using Matcher2 = typename Filter::Matcher2;
-  using FST1 = typename Filter::FST1;
-  using FST2 = typename Filter::FST2;
-  using FilterState = typename Filter::FilterState;
-
-  using Arc = typename Filter::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  MultiEpsFilter(const FST1 &fst1, const FST2 &fst2,
-                 Matcher1 *matcher1 = nullptr, Matcher2 *matcher2 = nullptr,
-                 bool keep_multi_eps = false)
-      : filter_(fst1, fst2, matcher1, matcher2),
-        keep_multi_eps_(keep_multi_eps) {}
-
-  MultiEpsFilter(const MultiEpsFilter<Filter> &filter, bool safe = false)
-      : filter_(filter.filter_, safe),
-        keep_multi_eps_(filter.keep_multi_eps_) {}
-
-  FilterState Start() const { return filter_.Start(); }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    return filter_.SetState(s1, s2, fs);
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    const auto fs = filter_.FilterArc(arc1, arc2);
-    if (keep_multi_eps_) {
-      if (arc1->olabel == kNoLabel) arc1->ilabel = arc2->ilabel;
-      if (arc2->ilabel == kNoLabel) arc2->olabel = arc1->olabel;
-    }
-    return fs;
-  }
-
-  void FilterFinal(Weight *w1, Weight *w2) const {
-    return filter_.FilterFinal(w1, w2);
-  }
-
-  Matcher1 *GetMatcher1() { return filter_.GetMatcher1(); }
-
-  Matcher2 *GetMatcher2() { return filter_.GetMatcher2(); }
-
-  uint64 Properties(uint64 iprops) const {
-    const auto oprops = filter_.Properties(iprops);
-    return oprops & kILabelInvariantProperties & kOLabelInvariantProperties;
-  }
-
- private:
-  Filter filter_;
-  bool keep_multi_eps_;
-};
-
-}  // namespace fst
-
-#endif  // FST_COMPOSE_FILTER_H_
diff --git a/projects/llm_framework/include/fst/compose.h b/projects/llm_framework/include/fst/compose.h
deleted file mode 100644
index 1066d097..00000000
--- a/projects/llm_framework/include/fst/compose.h
+++ /dev/null
@@ -1,1035 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to compute the composition of two FSTs.
-
-#ifndef FST_COMPOSE_H_
-#define FST_COMPOSE_H_
-
-#include <algorithm>
-#include <memory>
-#include <utility>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/compose-filter.h>
-#include <fst/fst-decl.h>  // For optional argument declarations
-#include <fst/lookahead-filter.h>
-#include <fst/matcher.h>
-#include <fst/state-table.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-// Delayed composition options templated on the arc type, the matcher,
-// the composition filter, and the composition state table.  By
-// default, the matchers, filter, and state table are constructed by
-// composition. If set below, the user can instead pass in these
-// objects; in that case, ComposeFst takes their ownership. This
-// version controls composition implemented between generic Fst<Arc>
-// types and a shared matcher type M for Fst<Arc>. This should be
-// adequate for most applications, giving a reasonable tradeoff
-// between efficiency and code sharing (but see ComposeFstImplOptions).
-template <class Arc, class M = Matcher<Fst<Arc>>,
-          class Filter = SequenceComposeFilter<M>,
-          class StateTable =
-              GenericComposeStateTable<Arc, typename Filter::FilterState>>
-struct ComposeFstOptions : public CacheOptions {
-  M *matcher1;              // FST1 matcher.
-  M *matcher2;              // FST2 matcher.
-  Filter *filter;           // Composition filter.
-  StateTable *state_table;  // Composition state table.
-
-  explicit ComposeFstOptions(const CacheOptions &opts = CacheOptions(),
-                             M *matcher1 = nullptr, M *matcher2 = nullptr,
-                             Filter *filter = nullptr,
-                             StateTable *state_table = nullptr)
-      : CacheOptions(opts),
-        matcher1(matcher1),
-        matcher2(matcher2),
-        filter(filter),
-        state_table(state_table) {}
-};
-
-// Forward declaration of ComposeFstMatcher.
-template <class C, class F, class T>
-class ComposeFstMatcher;
-
-// Delayed composition options templated on the two matcher types, the
-// composition filter, the composition state table and the cache store. By
-// default, the matchers, filter, state table and cache store are constructed
-// by composition. If set below, the user can instead pass in these objects; in
-// that case, ComposeFst takes their ownership. This version controls
-// composition implemented using arbitrary matchers (of the same arc type but
-// otherwise arbitrary FST type). The user must ensure the matchers are
-// compatible. These options permit the most efficient use, but shares the
-// least code. This is for advanced use only in the most demanding or
-// specialized applications that can benefit from it; otherwise, prefer
-// ComposeFstOptions).
-template <class M1, class M2, class Filter = SequenceComposeFilter<M1, M2>,
-          class StateTable = GenericComposeStateTable<
-              typename M1::Arc, typename Filter::FilterState>,
-          class CacheStore = DefaultCacheStore<typename M1::Arc>>
-struct ComposeFstImplOptions : public CacheImplOptions<CacheStore> {
-  M1 *matcher1;    // FST1 matcher (see matcher.h)....
-  M2 *matcher2;    // FST2 matcher.
-  Filter *filter;  // Composition filter (see compose-filter.h).
-  StateTable
-    *state_table;        // Composition state table (see compose-state-table.h).
-  bool own_state_table;   // ComposeFstImpl takes ownership of 'state_table'?
-  bool allow_noncommute;  // Allow non-commutative weights
-
-  explicit ComposeFstImplOptions(const CacheOptions &opts,
-                                 M1 *matcher1 = nullptr, M2 *matcher2 = nullptr,
-                                 Filter *filter = nullptr,
-                                 StateTable *state_table = nullptr)
-      : CacheImplOptions<CacheStore>(opts),
-        matcher1(matcher1),
-        matcher2(matcher2),
-        filter(filter),
-        state_table(state_table),
-        own_state_table(true),
-        allow_noncommute(false) {}
-
-  explicit ComposeFstImplOptions(const CacheImplOptions<CacheStore> &opts,
-                                 M1 *matcher1 = nullptr, M2 *matcher2 = nullptr,
-                                 Filter *filter = nullptr,
-                                 StateTable *state_table = nullptr)
-      : CacheImplOptions<CacheStore>(opts),
-        matcher1(matcher1),
-        matcher2(matcher2),
-        filter(filter),
-        state_table(state_table),
-        own_state_table(true),
-        allow_noncommute(false) {}
-
-  ComposeFstImplOptions()
-      : matcher1(nullptr),
-        matcher2(nullptr),
-        filter(nullptr),
-        state_table(nullptr),
-        own_state_table(true),
-        allow_noncommute(false) {}
-};
-
-namespace internal {
-
-// Implementation of delayed composition. This base class is common to the
-// variants with different matchers, composition filters and state tables.
-template <class Arc, class CacheStore = DefaultCacheStore<Arc>,
-          class F = ComposeFst<Arc, CacheStore>>
-class ComposeFstImplBase
-    : public CacheBaseImpl<typename CacheStore::State, CacheStore> {
- public:
-  using FST = F;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using State = typename CacheStore::State;
-  using CacheImpl = CacheBaseImpl<State, CacheStore>;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::Properties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheImpl::HasStart;
-  using CacheImpl::HasFinal;
-  using CacheImpl::HasArcs;
-  using CacheImpl::SetFinal;
-  using CacheImpl::SetStart;
-
-  ComposeFstImplBase(const CacheImplOptions<CacheStore> &opts)
-      : CacheImpl(opts) {}
-
-  ComposeFstImplBase(const CacheOptions &opts) : CacheImpl(opts) {}
-
-  ComposeFstImplBase(const ComposeFstImplBase &impl) : CacheImpl(impl, true) {
-    SetType(impl.Type());
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  virtual ComposeFstImplBase *Copy() const = 0;
-
-  ~ComposeFstImplBase() override {}
-
-  StateId Start() {
-    if (!HasStart()) {
-      const auto start = ComputeStart();
-      if (start != kNoStateId) SetStart(start);
-    }
-    return CacheImpl::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) SetFinal(s, ComputeFinal(s));
-    return CacheImpl::Final(s);
-  }
-
-  virtual void Expand(StateId s) = 0;
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl::NumOutputEpsilons(s);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl::InitArcIterator(s, data);
-  }
-
-  virtual MatcherBase<Arc> *InitMatcher(const F &fst,
-                                        MatchType match_type) const {
-    // Use the default matcher if no override is provided.
-    return nullptr;
-  }
-
- protected:
-  virtual StateId ComputeStart() = 0;
-  virtual Weight ComputeFinal(StateId s) = 0;
-};
-
-// Implementation of delayed composition templated on the matchers (see
-// matcher.h), composition filter (see compose-filter.h) and the composition
-// state table (see compose-state-table.h).
-template <class CacheStore, class Filter, class StateTable>
-class ComposeFstImpl
-    : public ComposeFstImplBase<typename CacheStore::Arc, CacheStore> {
- public:
-  using Matcher1 = typename Filter::Matcher1;
-  using Matcher2 = typename Filter::Matcher2;
-
-  using FST1 = typename Matcher1::FST;
-  using FST2 = typename Matcher2::FST;
-
-  using Arc = typename CacheStore::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FilterState = typename Filter::FilterState;
-  using State = typename CacheStore::State;
-
-  using CacheImpl = CacheBaseImpl<State, CacheStore>;
-
-  using StateTuple = typename StateTable::StateTuple;
-
-  friend class ComposeFstMatcher<CacheStore, Filter, StateTable>;
-
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-
-  template <class M1, class M2>
-  ComposeFstImpl(const FST1 &fst1, const FST2 &fst2,
-                 const ComposeFstImplOptions<M1, M2, Filter, StateTable,
-                                             CacheStore> &opts);
-
-  ComposeFstImpl(const ComposeFstImpl &impl)
-      : ComposeFstImplBase<Arc, CacheStore>(impl),
-        filter_(new Filter(*impl.filter_, true)),
-        matcher1_(filter_->GetMatcher1()),
-        matcher2_(filter_->GetMatcher2()),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()),
-        state_table_(new StateTable(*impl.state_table_)),
-        own_state_table_(true),
-        match_type_(impl.match_type_) {}
-
-  ~ComposeFstImpl() override {
-    if (own_state_table_) delete state_table_;
-  }
-
-  ComposeFstImpl *Copy() const override { return new ComposeFstImpl(*this); }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) &&
-        (fst1_.Properties(kError, false) || fst2_.Properties(kError, false) ||
-         (matcher1_->Properties(0) & kError) ||
-         (matcher2_->Properties(0) & kError) |
-             (filter_->Properties(0) & kError) ||
-         state_table_->Error())) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  // Arranges it so that the first arg to OrderedExpand is the Fst
-  // that will be matched on.
-  void Expand(StateId s) override {
-    const auto &tuple = state_table_->Tuple(s);
-    const auto s1 = tuple.StateId1();
-    const auto s2 = tuple.StateId2();
-    filter_->SetState(s1, s2, tuple.GetFilterState());
-    if (MatchInput(s1, s2)) {
-      OrderedExpand(s, fst2_, s2, fst1_, s1, matcher2_, true);
-    } else {
-      OrderedExpand(s, fst1_, s1, fst2_, s2, matcher1_, false);
-    }
-  }
-
-  const FST1 &GetFst1() const { return fst1_; }
-
-  const FST2 &GetFst2() const { return fst2_; }
-
-  const Matcher1 *GetMatcher1() const { return matcher1_; }
-
-  Matcher1 *GetMatcher1() { return matcher1_; }
-
-  const Matcher2 *GetMatcher2() const { return matcher2_; }
-
-  Matcher2 *GetMatcher2() { return matcher2_; }
-
-  const Filter *GetFilter() const { return filter_.get(); }
-
-  Filter *GetFilter() { return filter_.get(); }
-
-  const StateTable *GetStateTable() const { return state_table_; }
-
-  StateTable *GetStateTable() { return state_table_; }
-
-  MatcherBase<Arc> *InitMatcher(const ComposeFst<Arc, CacheStore> &fst,
-                                MatchType match_type) const override {
-    const auto test_props = match_type == MATCH_INPUT
-                                ? kFstProperties & ~kILabelInvariantProperties
-                                : kFstProperties & ~kOLabelInvariantProperties;
-    // If both matchers support 'match_type' and we have a guarantee that a
-    // call to 'filter_->FilterArc(arc1, arc2)' will not modify the ilabel of
-    // arc1 when MATCH_INPUT or the olabel or arc2 when MATCH_OUTPUT, then
-    // ComposeFstMatcher can be used.
-    if ((matcher1_->Type(false) == match_type) &&
-        (matcher2_->Type(false) == match_type) &&
-        (filter_->Properties(test_props) == test_props)) {
-      return new ComposeFstMatcher<
-        CacheStore, Filter, StateTable>(&fst, match_type);
-    }
-    return nullptr;
-  }
-
- private:
-  // This does that actual matching of labels in the composition. The
-  // arguments are ordered so matching is called on state 'sa' of
-  // 'fsta' for each arc leaving state 'sb' of 'fstb'. The 'match_input' arg
-  // determines whether the input or output label of arcs at 'sb' is
-  // the one to match on.
-  template <class FST, class Matcher>
-  void OrderedExpand(StateId s, const Fst<Arc> &, StateId sa, const FST &fstb,
-                     StateId sb, Matcher *matchera, bool match_input) {
-    matchera->SetState(sa);
-    // First processes non-consuming symbols (e.g., epsilons) on FSTA.
-    const Arc loop(match_input ? 0 : kNoLabel, match_input ? kNoLabel : 0,
-                   Weight::One(), sb);
-    MatchArc(s, matchera, loop, match_input);
-    // Then processes matches on FSTB.
-    for (ArcIterator<FST> iterb(fstb, sb); !iterb.Done(); iterb.Next()) {
-      MatchArc(s, matchera, iterb.Value(), match_input);
-    }
-    CacheImpl::SetArcs(s);
-  }
-
-  // Matches a single transition from 'fstb' against 'fata' at 's'.
-  template <class Matcher>
-  void MatchArc(StateId s, Matcher *matchera, const Arc &arc,
-                bool match_input) {
-    if (matchera->Find(match_input ? arc.olabel : arc.ilabel)) {
-      for (; !matchera->Done(); matchera->Next()) {
-        auto arca = matchera->Value();
-        auto arcb = arc;
-        if (match_input) {
-          const auto &fs = filter_->FilterArc(&arcb, &arca);
-          if (fs != FilterState::NoState()) AddArc(s, arcb, arca, fs);
-        } else {
-          const auto &fs = filter_->FilterArc(&arca, &arcb);
-          if (fs != FilterState::NoState()) AddArc(s, arca, arcb, fs);
-        }
-      }
-    }
-  }
-
-  // Add a matching transition at 's'.
-  void AddArc(StateId s, const Arc &arc1, const Arc &arc2,
-              const FilterState &f) {
-    const StateTuple tuple(arc1.nextstate, arc2.nextstate, f);
-    CacheImpl::EmplaceArc(
-        s, arc1.ilabel, arc2.olabel, Times(arc1.weight, arc2.weight),
-        state_table_->FindState(tuple));
-  }
-
-  StateId ComputeStart() override {
-    const auto s1 = fst1_.Start();
-    if (s1 == kNoStateId) return kNoStateId;
-    const auto s2 = fst2_.Start();
-    if (s2 == kNoStateId) return kNoStateId;
-    const auto &fs = filter_->Start();
-    const StateTuple tuple(s1, s2, fs);
-    return state_table_->FindState(tuple);
-  }
-
-  Weight ComputeFinal(StateId s) override {
-    const auto &tuple = state_table_->Tuple(s);
-    const auto s1 = tuple.StateId1();
-    auto final1 = matcher1_->Final(s1);
-    if (final1 == Weight::Zero()) return final1;
-    const auto s2 = tuple.StateId2();
-    auto final2 = matcher2_->Final(s2);
-    if (final2 == Weight::Zero()) return final2;
-    filter_->SetState(s1, s2, tuple.GetFilterState());
-    filter_->FilterFinal(&final1, &final2);
-    return Times(final1, final2);
-  }
-
-  // Determines which side to match on per composition state.
-  bool MatchInput(StateId s1, StateId s2) {
-    switch (match_type_) {
-      case MATCH_INPUT:
-        return true;
-      case MATCH_OUTPUT:
-        return false;
-      default:  // MATCH_BOTH
-        const auto priority1 = matcher1_->Priority(s1);
-        const auto priority2 = matcher2_->Priority(s2);
-        if (priority1 == kRequirePriority && priority2 == kRequirePriority) {
-          FSTERROR() << "ComposeFst: Both sides can't require match";
-          SetProperties(kError, kError);
-          return true;
-        }
-        if (priority1 == kRequirePriority) return false;
-        if (priority2 == kRequirePriority) {
-          return true;
-        }
-        return priority1 <= priority2;
-    }
-  }
-
-  // Identifies and verifies the capabilities of the matcher to be used for
-  // composition.
-  void SetMatchType();
-
-  std::unique_ptr<Filter> filter_;
-  Matcher1 *matcher1_;  // Borrowed reference.
-  Matcher2 *matcher2_;  // Borrowed reference.
-  const FST1 &fst1_;
-  const FST2 &fst2_;
-  StateTable *state_table_;
-  bool own_state_table_;
-
-  MatchType match_type_;
-};
-
-template <class CacheStore, class Filter, class StateTable>
-template <class M1, class M2>
-ComposeFstImpl<CacheStore, Filter, StateTable>::ComposeFstImpl(
-    const FST1 &fst1, const FST2 &fst2,
-    const ComposeFstImplOptions<M1, M2, Filter, StateTable, CacheStore> &opts)
-    : ComposeFstImplBase<Arc, CacheStore>(opts),
-      filter_(opts.filter
-                  ? opts.filter
-                  : new Filter(fst1, fst2, opts.matcher1, opts.matcher2)),
-      matcher1_(filter_->GetMatcher1()),
-      matcher2_(filter_->GetMatcher2()),
-      fst1_(matcher1_->GetFst()),
-      fst2_(matcher2_->GetFst()),
-      state_table_(opts.state_table ? opts.state_table
-                                    : new StateTable(fst1_, fst2_)),
-      own_state_table_(opts.state_table ? opts.own_state_table : true) {
-  SetType("compose");
-  if (!CompatSymbols(fst2.InputSymbols(), fst1.OutputSymbols())) {
-    FSTERROR() << "ComposeFst: Output symbol table of 1st argument "
-               << "does not match input symbol table of 2nd argument";
-    SetProperties(kError, kError);
-  }
-  SetInputSymbols(fst1_.InputSymbols());
-  SetOutputSymbols(fst2_.OutputSymbols());
-  SetMatchType();
-  VLOG(2) << "ComposeFstImpl: Match type: " << match_type_;
-  if (match_type_ == MATCH_NONE) SetProperties(kError, kError);
-  const auto fprops1 = fst1.Properties(kFstProperties, false);
-  const auto fprops2 = fst2.Properties(kFstProperties, false);
-  const auto mprops1 = matcher1_->Properties(fprops1);
-  const auto mprops2 = matcher2_->Properties(fprops2);
-  const auto cprops = ComposeProperties(mprops1, mprops2);
-  SetProperties(filter_->Properties(cprops), kCopyProperties);
-  if (state_table_->Error()) SetProperties(kError, kError);
-}
-
-template <class CacheStore, class Filter, class StateTable>
-void ComposeFstImpl<CacheStore, Filter, StateTable>::SetMatchType() {
-  // Ensures any required matching is possible and known.
-  if ((matcher1_->Flags() & kRequireMatch) &&
-      matcher1_->Type(true) != MATCH_OUTPUT) {
-    FSTERROR() << "ComposeFst: 1st argument cannot perform required matching "
-               << "(sort?).";
-    match_type_ = MATCH_NONE;
-    return;
-  }
-  if ((matcher2_->Flags() & kRequireMatch) &&
-      matcher2_->Type(true) != MATCH_INPUT) {
-    FSTERROR() << "ComposeFst: 2nd argument cannot perform required matching "
-               << "(sort?).";
-    match_type_ = MATCH_NONE;
-    return;
-  }
-  // Finds which sides to match on (favoring minimal testing of capabilities).
-  const auto type1 = matcher1_->Type(false);
-  const auto type2 = matcher2_->Type(false);
-  if (type1 == MATCH_OUTPUT && type2 == MATCH_INPUT) {
-    match_type_ = MATCH_BOTH;
-  } else if (type1 == MATCH_OUTPUT) {
-    match_type_ = MATCH_OUTPUT;
-  } else if (type2 == MATCH_INPUT) {
-    match_type_ = MATCH_INPUT;
-  } else if (matcher1_->Type(true) == MATCH_OUTPUT) {
-    match_type_ = MATCH_OUTPUT;
-  } else if (matcher2_->Type(true) == MATCH_INPUT) {
-    match_type_ = MATCH_INPUT;
-  } else {
-    FSTERROR() << "ComposeFst: 1st argument cannot match on output labels "
-               << "and 2nd argument cannot match on input labels (sort?).";
-    match_type_ = MATCH_NONE;
-  }
-}
-
-}  // namespace internal
-
-// Computes the composition of two transducers. This version is a delayed FST.
-// If FST1 transduces string x to y with weight a and FST2 transduces y to z
-// with weight b, then their composition transduces string x to z with weight
-// Times(x, z).
-//
-// The output labels of the first transducer or the input labels of the second
-// transducer must be sorted (with the default matcher). The weights need to
-// form a commutative semiring (valid for TropicalWeight and LogWeight).
-//
-// Complexity:
-//
-// Assuming the first FST is unsorted and the second is sorted,
-//
-//   Time: O(v1 v2 d1 (log d2 + m2)),
-//   Space: O(v1 v2)
-//
-// where vi = # of states visited, di = maximum out-degree, and mi the
-// maximum multiplicity of the states visited, for the ith FST. Constant time
-// and space to visit an input state or arc is assumed and exclusive of caching.
-//
-// Caveats:
-// - ComposeFst does not trim its output (since it is a delayed operation).
-// - The efficiency of composition can be strongly affected by several factors:
-//   - the choice of which transducer is sorted - prefer sorting the FST
-//     that has the greater average out-degree.
-//   - the amount of non-determinism
-//   - the presence and location of epsilon transitions - avoid epsilon
-//     transitions on the output side of the first transducer or
-//     the input side of the second transducer or prefer placing
-//     them later in a path since they delay matching and can
-//     introduce non-coaccessible states and transitions.
-//
-// This class attaches interface to implementation and handles reference
-// counting, delegating most methods to ImplToFst. The CacheStore specifies the
-// cache store (default declared in fst-decl.h).
-template <class A, class CacheStore /* = DefaultCacheStore<A> */>
-class ComposeFst
-    : public ImplToFst<internal::ComposeFstImplBase<A, CacheStore>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = CacheStore;
-  using State = typename CacheStore::State;
-
-  using Impl = internal::ComposeFstImplBase<A, CacheStore>;
-
-  friend class ArcIterator<ComposeFst<Arc, CacheStore>>;
-  friend class StateIterator<ComposeFst<Arc, CacheStore>>;
-  template <class, class, class> friend class ComposeFstMatcher;
-
-  // Compose specifying only caching options.
-  ComposeFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-             const CacheOptions &opts = CacheOptions())
-      : ImplToFst<Impl>(CreateBase(fst1, fst2, opts)) {}
-
-  // Compose specifying one shared matcher type M. Requires that the input FSTs
-  // and matcher FST types be Fst<Arc>. Recommended for best code-sharing and
-  // matcher compatiblity.
-  template <class Matcher, class Filter, class StateTuple>
-  ComposeFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-             const ComposeFstOptions<Arc, Matcher, Filter, StateTuple> &opts)
-      : ImplToFst<Impl>(CreateBase1(fst1, fst2, opts)) {}
-
-  // Compose specifying two matcher types Matcher1 and Matcher2. Requires input
-  // FST (of the same Arc type, but o.w. arbitrary) match the corresponding
-  // matcher FST types). Recommended only for advanced use in demanding or
-  // specialized applications due to potential code bloat and matcher
-  // incompatibilities.
-  template <class Matcher1, class Matcher2, class Filter, class StateTuple>
-  ComposeFst(const typename Matcher1::FST &fst1,
-             const typename Matcher2::FST &fst2,
-             const ComposeFstImplOptions<Matcher1, Matcher2, Filter, StateTuple,
-                                         CacheStore> &opts)
-      : ImplToFst<Impl>(CreateBase2(fst1, fst2, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  ComposeFst(const ComposeFst<A, CacheStore> &fst, bool safe = false)
-      : ImplToFst<Impl>(safe ? std::shared_ptr<Impl>(fst.GetImpl()->Copy())
-                             : fst.GetSharedImpl()) {}
-
-  // Get a copy of this ComposeFst. See Fst<>::Copy() for further doc.
-  ComposeFst<A, CacheStore> *Copy(bool safe = false) const override {
-    return new ComposeFst<A, CacheStore>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
-  MatcherBase<Arc> *InitMatcher(MatchType match_type) const override {
-    return GetImpl()->InitMatcher(*this, match_type);
-  }
-
- protected:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  explicit ComposeFst(std::shared_ptr<Impl> impl) : ImplToFst<Impl>(impl) {}
-
-  // Create compose implementation specifying two matcher types.
-  template <class Matcher1, class Matcher2, class Filter, class StateTuple>
-  static std::shared_ptr<Impl> CreateBase2(
-      const typename Matcher1::FST &fst1, const typename Matcher2::FST &fst2,
-      const ComposeFstImplOptions<Matcher1, Matcher2, Filter, StateTuple,
-                                  CacheStore> &opts) {
-    auto impl = std::make_shared<
-        internal::ComposeFstImpl<CacheStore, Filter, StateTuple>>(fst1, fst2,
-                                                                  opts);
-    if (!(Weight::Properties() & kCommutative) && !opts.allow_noncommute) {
-      const auto props1 = fst1.Properties(kUnweighted, true);
-      const auto props2 = fst2.Properties(kUnweighted, true);
-      if (!(props1 & kUnweighted) && !(props2 & kUnweighted)) {
-        FSTERROR() << "ComposeFst: Weights must be a commutative semiring: "
-                   << Weight::Type();
-        impl->SetProperties(kError, kError);
-      }
-    }
-    return impl;
-  }
-
-  // Create compose implementation specifying one matcher type; requires that
-  // input and matcher FST types be Fst<Arc>.
-  template <class Matcher, class Filter, class StateTuple>
-  static std::shared_ptr<Impl> CreateBase1(
-      const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-      const ComposeFstOptions<Arc, Matcher, Filter, StateTuple> &opts) {
-    ComposeFstImplOptions<Matcher, Matcher, Filter, StateTuple, CacheStore>
-        nopts(opts, opts.matcher1, opts.matcher2, opts.filter,
-              opts.state_table);
-    return CreateBase2(fst1, fst2, nopts);
-  }
-
-  // Create compose implementation specifying no matcher type.
-  static std::shared_ptr<Impl> CreateBase(const Fst<Arc> &fst1,
-                                          const Fst<Arc> &fst2,
-                                          const CacheOptions &opts) {
-    switch (LookAheadMatchType(fst1, fst2)) {  // Check for lookahead matchers
-      default:
-      case MATCH_NONE: {  // Default composition (no look-ahead).
-        ComposeFstOptions<Arc> nopts(opts);
-        return CreateBase1(fst1, fst2, nopts);
-      }
-      case MATCH_OUTPUT: {  // Lookahead on fst1.
-        using M = typename DefaultLookAhead<Arc, MATCH_OUTPUT>::FstMatcher;
-        using F = typename DefaultLookAhead<Arc, MATCH_OUTPUT>::ComposeFilter;
-        ComposeFstOptions<Arc, M, F> nopts(opts);
-        return CreateBase1(fst1, fst2, nopts);
-      }
-      case MATCH_INPUT: {  // Lookahead on fst2
-        using M = typename DefaultLookAhead<Arc, MATCH_INPUT>::FstMatcher;
-        using F = typename DefaultLookAhead<Arc, MATCH_INPUT>::ComposeFilter;
-        ComposeFstOptions<Arc, M, F> nopts(opts);
-        return CreateBase1(fst1, fst2, nopts);
-      }
-    }
-  }
-
- private:
-  ComposeFst &operator=(const ComposeFst &fst) = delete;
-};
-
-// Specialization for ComposeFst.
-template <class Arc, class CacheStore>
-class StateIterator<ComposeFst<Arc, CacheStore>>
-    : public CacheStateIterator<ComposeFst<Arc, CacheStore>> {
- public:
-  explicit StateIterator(const ComposeFst<Arc, CacheStore> &fst)
-      : CacheStateIterator<ComposeFst<Arc, CacheStore>>(fst,
-                                                        fst.GetMutableImpl()) {}
-};
-
-// Specialization for ComposeFst.
-template <class Arc, class CacheStore>
-class ArcIterator<ComposeFst<Arc, CacheStore>>
-    : public CacheArcIterator<ComposeFst<Arc, CacheStore>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const ComposeFst<Arc, CacheStore> &fst, StateId s)
-      : CacheArcIterator<ComposeFst<Arc, CacheStore>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc, class CacheStore>
-inline void ComposeFst<Arc, CacheStore>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<ComposeFst<Arc, CacheStore>>(*this);
-}
-
-// Specialized matcher for ComposeFst. Supports MATCH_INPUT or MATCH_OUTPUT,
-// iff the underlying matchers for the two FSTS being composed support
-// MATCH_INPUT or MATCH_OUTPUT, respectively.
-template <class CacheStore, class Filter, class StateTable>
-class ComposeFstMatcher : public MatcherBase<typename CacheStore::Arc> {
- public:
-  using Arc = typename CacheStore::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Matcher1 = typename Filter::Matcher1;
-  using Matcher2 = typename Filter::Matcher2;
-  using FilterState = typename Filter::FilterState;
-
-  using StateTuple = typename StateTable::StateTuple;
-  using Impl = internal::ComposeFstImpl<CacheStore, Filter, StateTable>;
-
-  // The compose FST arg must match the filter and state table types.
-  // This makes a copy of the FST.
-  ComposeFstMatcher(const ComposeFst<Arc, CacheStore> &fst,
-                    MatchType match_type)
-      : owned_fst_(fst.Copy()),
-        fst_(*owned_fst_),
-        impl_(static_cast<const Impl *>(fst_.GetImpl())),
-        s_(kNoStateId),
-        match_type_(match_type),
-        matcher1_(impl_->matcher1_->Copy()),
-        matcher2_(impl_->matcher2_->Copy()),
-        current_loop_(false),
-        loop_(kNoLabel, 0, Weight::One(), kNoStateId) {
-    if (match_type_ == MATCH_OUTPUT) std::swap(loop_.ilabel, loop_.olabel);
-  }
-
-  // The compose FST arg must match the filter and state table types.
-  // This doesn't copy the FST (although it may copy components).
-  ComposeFstMatcher(const ComposeFst<Arc, CacheStore> *fst,
-                    MatchType match_type)
-      : fst_(*fst),
-        impl_(static_cast<const Impl *>(fst_.GetImpl())),
-        s_(kNoStateId),
-        match_type_(match_type),
-        matcher1_(impl_->matcher1_->Copy()),
-        matcher2_(impl_->matcher2_->Copy()),
-        current_loop_(false),
-        loop_(kNoLabel, 0, Weight::One(), kNoStateId) {
-    if (match_type_ == MATCH_OUTPUT) std::swap(loop_.ilabel, loop_.olabel);
-  }
-
-  // This makes a copy of the FST.
-  ComposeFstMatcher(
-      const ComposeFstMatcher<CacheStore, Filter, StateTable> &matcher,
-      bool safe = false)
-      : owned_fst_(matcher.fst_.Copy(safe)),
-        fst_(*owned_fst_),
-        impl_(static_cast<const Impl *>(fst_.GetImpl())),
-        s_(kNoStateId),
-        match_type_(matcher.match_type_),
-        matcher1_(matcher.matcher1_->Copy(safe)),
-        matcher2_(matcher.matcher2_->Copy(safe)),
-        current_loop_(false),
-        loop_(kNoLabel, 0, Weight::One(), kNoStateId) {
-    if (match_type_ == MATCH_OUTPUT) std::swap(loop_.ilabel, loop_.olabel);
-  }
-
-  ComposeFstMatcher<CacheStore, Filter, StateTable> *Copy(
-      bool safe = false) const override {
-    return new ComposeFstMatcher<CacheStore, Filter, StateTable>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override {
-    if ((matcher1_->Type(test) == MATCH_NONE) ||
-        (matcher2_->Type(test) == MATCH_NONE)) {
-      return MATCH_NONE;
-    }
-    if (((matcher1_->Type(test) == MATCH_UNKNOWN) &&
-         (matcher2_->Type(test) == MATCH_UNKNOWN)) ||
-        ((matcher1_->Type(test) == MATCH_UNKNOWN) &&
-         (matcher2_->Type(test) == match_type_)) ||
-        ((matcher1_->Type(test) == match_type_) &&
-         (matcher2_->Type(test) == MATCH_UNKNOWN))) {
-      return MATCH_UNKNOWN;
-    }
-    if ((matcher1_->Type(test) == match_type_) &&
-        (matcher2_->Type(test) == match_type_)) {
-      return match_type_;
-    }
-    return MATCH_NONE;
-  }
-
-  const Fst<Arc> &GetFst() const override { return fst_; }
-
-  uint64 Properties(uint64 inprops) const override {
-    return inprops;
-  }
-
-  void SetState(StateId s) final {
-    if (s_ == s) return;
-    s_ = s;
-    const auto &tuple = impl_->state_table_->Tuple(s);
-    matcher1_->SetState(tuple.StateId1());
-    matcher2_->SetState(tuple.StateId2());
-    loop_.nextstate = s_;
-  }
-
-  bool Find(Label label) final {
-    bool found = false;
-    current_loop_ = false;
-    if (label == 0) {
-      current_loop_ = true;
-      found = true;
-    }
-    if (match_type_ == MATCH_INPUT) {
-      found = found || FindLabel(label, matcher1_.get(), matcher2_.get());
-    } else {  // match_type_ == MATCH_OUTPUT
-      found = found || FindLabel(label, matcher2_.get(), matcher1_.get());
-    }
-    return found;
-  }
-
-  bool Done() const final {
-    return !current_loop_ && matcher1_->Done() && matcher2_->Done();
-  }
-
-  const Arc &Value() const final { return current_loop_ ? loop_ : arc_; }
-
-  void Next() final {
-    if (current_loop_) {
-      current_loop_ = false;
-    } else if (match_type_ == MATCH_INPUT) {
-      FindNext(matcher1_.get(), matcher2_.get());
-    } else {  // match_type_ == MATCH_OUTPUT
-      FindNext(matcher2_.get(), matcher1_.get());
-    }
-  }
-
-  ssize_t Priority(StateId s) final { return fst_.NumArcs(s); }
-
- private:
-  // Processes a match with the filter and creates resulting arc.
-  bool MatchArc(StateId s, Arc arc1,
-                Arc arc2) {  // FIXME(kbg): copy but not assignment.
-    const auto &fs = impl_->filter_->FilterArc(&arc1, &arc2);
-    if (fs == FilterState::NoState()) return false;
-    const StateTuple tuple(arc1.nextstate, arc2.nextstate, fs);
-    arc_.ilabel = arc1.ilabel;
-    arc_.olabel = arc2.olabel;
-    arc_.weight = Times(arc1.weight, arc2.weight);
-    arc_.nextstate = impl_->state_table_->FindState(tuple);
-    return true;
-  }
-
-  // Finds the first match allowed by the filter.
-  template <class MatcherA, class MatcherB>
-  bool FindLabel(Label label, MatcherA *matchera, MatcherB *matcherb) {
-    if (matchera->Find(label)) {
-      matcherb->Find(match_type_ == MATCH_INPUT ? matchera->Value().olabel
-                                                : matchera->Value().ilabel);
-      return FindNext(matchera, matcherb);
-    }
-    return false;
-  }
-
-  // Finds the next match allowed by the filter, returning true iff such a
-  // match is found.
-  template <class MatcherA, class MatcherB>
-  bool FindNext(MatcherA *matchera, MatcherB *matcherb) {
-    // State when entering this function:
-    // 'matchera' is pointed to a match x, y for label x, and a match for y was
-    // requested on 'matcherb'.
-    while (!matchera->Done() || !matcherb->Done()) {
-      if (matcherb->Done()) {
-        // If no more matches for y on 'matcherb', moves forward on 'matchera'
-        // until a match x, y' is found such that there is a match for y' on
-        // 'matcherb'.
-        matchera->Next();
-        while (!matchera->Done() &&
-               !matcherb->Find(match_type_ == MATCH_INPUT
-                                   ? matchera->Value().olabel
-                                   : matchera->Value().ilabel)) {
-          matchera->Next();
-        }
-      }
-      while (!matcherb->Done()) {
-        // 'matchera' is pointing to a match x, y' ('arca') and 'matcherb' is
-        // pointing to a match y', z' ('arcb'). If combining these two arcs is
-        // allowed by the filter (hence resulting in an arc x, z') return true.
-        // Position 'matcherb' on the next potential match for y' before
-        // returning.
-        const auto &arca = matchera->Value();
-        const auto &arcb = matcherb->Value();
-        // Position 'matcherb' on the next potential match for y'.
-        matcherb->Next();
-        // Returns true If combining these two arcs is allowed by the filter
-        // (hence resulting in an arc x, z'); otherwise consider next match
-        // for y' on 'matcherb'.
-        if (MatchArc(s_, match_type_ == MATCH_INPUT ? arca : arcb,
-                     match_type_ == MATCH_INPUT ? arcb : arca)) {
-          return true;
-        }
-      }
-    }
-    // Both 'matchera' and 'matcherb' are done, no more match to analyse.
-    return false;
-  }
-
-  std::unique_ptr<const ComposeFst<Arc, CacheStore>> owned_fst_;
-  const ComposeFst<Arc, CacheStore> &fst_;
-  const Impl *impl_;
-  StateId s_;
-  MatchType match_type_;
-  std::unique_ptr<Matcher1> matcher1_;
-  std::unique_ptr<Matcher2> matcher2_;
-  bool current_loop_;
-  Arc loop_;
-  Arc arc_;
-};
-
-// Useful alias when using StdArc.
-using StdComposeFst = ComposeFst<StdArc>;
-
-enum ComposeFilter {
-  AUTO_FILTER,
-  NULL_FILTER,
-  TRIVIAL_FILTER,
-  SEQUENCE_FILTER,
-  ALT_SEQUENCE_FILTER,
-  MATCH_FILTER,
-  NO_MATCH_FILTER
-};
-
-struct ComposeOptions {
-  bool connect;               // Connect output?
-  ComposeFilter filter_type;  // Pre-defined filter to use.
-
-  explicit ComposeOptions(bool connect = true,
-                          ComposeFilter filter_type = AUTO_FILTER)
-      : connect(connect), filter_type(filter_type) {}
-};
-
-// Computes the composition of two transducers. This version writes
-// the composed FST into a MutableFst. If FST1 transduces string x to
-// y with weight a and FST2 transduces y to z with weight b, then
-// their composition transduces string x to z with weight
-// Times(x, z).
-//
-// The output labels of the first transducer or the input labels of
-// the second transducer must be sorted.  The weights need to form a
-// commutative semiring (valid for TropicalWeight and LogWeight).
-//
-// Complexity:
-//
-// Assuming the first FST is unsorted and the second is sorted:
-//
-//   Time: O(V1 V2 D1 (log D2 + M2)),
-//   Space: O(V1 V2 D1 M2)
-//
-// where Vi = # of states, Di = maximum out-degree, and Mi is the maximum
-// multiplicity, for the ith FST.
-//
-// Caveats:
-//
-// - Compose trims its output.
-// - The efficiency of composition can be strongly affected by several factors:
-//   - the choice of which transducer is sorted - prefer sorting the FST
-//     that has the greater average out-degree.
-//   - the amount of non-determinism
-//   - the presence and location of epsilon transitions - avoid epsilon
-//     transitions on the output side of the first transducer or
-//     the input side of the second transducer or prefer placing
-//     them later in a path since they delay matching and can
-//     introduce non-coaccessible states and transitions.
-template <class Arc>
-void Compose(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-             MutableFst<Arc> *ofst,
-             const ComposeOptions &opts = ComposeOptions()) {
-  using M = Matcher<Fst<Arc>>;
-  // In each case, we cache only the last state for fastest copy.
-  switch (opts.filter_type) {
-    case AUTO_FILTER: {
-      CacheOptions nopts;
-      nopts.gc_limit = 0;
-      *ofst = ComposeFst<Arc>(ifst1, ifst2, nopts);
-      break;
-    }
-    case NULL_FILTER: {
-      ComposeFstOptions<Arc, M, NullComposeFilter<M>> copts;
-      copts.gc_limit = 0;
-      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-      break;
-    }
-    case SEQUENCE_FILTER: {
-      ComposeFstOptions<Arc, M, SequenceComposeFilter<M>> copts;
-      copts.gc_limit = 0;
-      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-      break;
-    }
-    case ALT_SEQUENCE_FILTER: {
-      ComposeFstOptions<Arc, M, AltSequenceComposeFilter<M>> copts;
-      copts.gc_limit = 0;
-      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-      break;
-    }
-    case MATCH_FILTER: {
-      ComposeFstOptions<Arc, M, MatchComposeFilter<M>> copts;
-      copts.gc_limit = 0;
-      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-      break;
-    }
-    case NO_MATCH_FILTER: {
-      ComposeFstOptions<Arc, M, NoMatchComposeFilter<M>> copts;
-      copts.gc_limit = 0;
-      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-      break;
-    }
-    case TRIVIAL_FILTER: {
-      ComposeFstOptions<Arc, M, TrivialComposeFilter<M>> copts;
-      copts.gc_limit = 0;
-      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-      break;
-    }
-  }
-  if (opts.connect) Connect(ofst);
-}
-
-}  // namespace fst
-
-#endif  // FST_COMPOSE_H_
diff --git a/projects/llm_framework/include/fst/concat.h b/projects/llm_framework/include/fst/concat.h
deleted file mode 100644
index 74d22c22..00000000
--- a/projects/llm_framework/include/fst/concat.h
+++ /dev/null
@@ -1,220 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to compute the concatenation of two FSTs.
-
-#ifndef FST_CONCAT_H_
-#define FST_CONCAT_H_
-
-#include <algorithm>
-#include <vector>
-
-#include <fst/mutable-fst.h>
-#include <fst/rational.h>
-
-
-namespace fst {
-
-// Computes the concatenation (product) of two FSTs. If FST1 transduces string
-// x to y with weight a and FST2 transduces string w to v with weight b, then
-// their concatenation transduces string xw to yv with weight Times(a, b).
-//
-// This version modifies its MutableFst argument (in first position).
-//
-// Complexity:
-//
-//   Time: O(V1 + V2 + E2)
-//   Space: O(V1 + V2 + E2)
-//
-// where Vi is the number of states, and Ei is the number of arcs, of the ith
-// FST.
-template <class Arc>
-void Concat(MutableFst<Arc> *fst1, const Fst<Arc> &fst2) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  // Checks that the symbol table are compatible.
-  if (!CompatSymbols(fst1->InputSymbols(), fst2.InputSymbols()) ||
-      !CompatSymbols(fst1->OutputSymbols(), fst2.OutputSymbols())) {
-    FSTERROR() << "Concat: Input/output symbol tables of 1st argument "
-               << "does not match input/output symbol tables of 2nd argument";
-    fst1->SetProperties(kError, kError);
-    return;
-  }
-  const auto props1 = fst1->Properties(kFstProperties, false);
-  const auto props2 = fst2.Properties(kFstProperties, false);
-  const auto start1 = fst1->Start();
-  if (start1 == kNoStateId) {
-    if (props2 & kError) fst1->SetProperties(kError, kError);
-    return;
-  }
-  const auto numstates1 = fst1->NumStates();
-  if (fst2.Properties(kExpanded, false)) {
-    fst1->ReserveStates(numstates1 + CountStates(fst2));
-  }
-  for (StateIterator<Fst<Arc>> siter2(fst2); !siter2.Done(); siter2.Next()) {
-    const auto s1 = fst1->AddState();
-    const auto s2 = siter2.Value();
-    fst1->SetFinal(s1, fst2.Final(s2));
-    fst1->ReserveArcs(s1, fst2.NumArcs(s2));
-    for (ArcIterator<Fst<Arc>> aiter(fst2, s2); !aiter.Done(); aiter.Next()) {
-      auto arc = aiter.Value();
-      arc.nextstate += numstates1;
-      fst1->AddArc(s1, arc);
-    }
-  }
-  const auto start2 = fst2.Start();
-  for (StateId s1 = 0; s1 < numstates1; ++s1) {
-    const auto weight = fst1->Final(s1);
-    if (weight != Weight::Zero()) {
-      fst1->SetFinal(s1, Weight::Zero());
-      if (start2 != kNoStateId) {
-        fst1->AddArc(s1, Arc(0, 0, weight, start2 + numstates1));
-      }
-    }
-  }
-  if (start2 != kNoStateId) {
-    fst1->SetProperties(ConcatProperties(props1, props2), kFstProperties);
-  }
-}
-
-// Computes the concatentation of two FSTs.  This version modifies its
-// MutableFst argument (in second position).
-//
-// Complexity:
-//
-//   Time: O(V1 + E1)
-//   Space: O(V1 + E1)
-//
-// where Vi is the number of states, and Ei is the number of arcs, of the ith
-// FST.
-template <class Arc>
-void Concat(const Fst<Arc> &fst1, MutableFst<Arc> *fst2) {
-  using Weight = typename Arc::Weight;
-  // Checks that the symbol table are compatible.
-  if (!CompatSymbols(fst1.InputSymbols(), fst2->InputSymbols()) ||
-      !CompatSymbols(fst1.OutputSymbols(), fst2->OutputSymbols())) {
-    FSTERROR() << "Concat: Input/output symbol tables of 1st argument "
-               << "does not match input/output symbol tables of 2nd argument";
-    fst2->SetProperties(kError, kError);
-    return;
-  }
-  const auto props1 = fst1.Properties(kFstProperties, false);
-  const auto props2 = fst2->Properties(kFstProperties, false);
-  const auto start2 = fst2->Start();
-  if (start2 == kNoStateId) {
-    if (props1 & kError) fst2->SetProperties(kError, kError);
-    return;
-  }
-  const auto numstates2 = fst2->NumStates();
-  if (fst1.Properties(kExpanded, false)) {
-    fst2->ReserveStates(numstates2 + CountStates(fst1));
-  }
-  for (StateIterator<Fst<Arc>> siter(fst1); !siter.Done(); siter.Next()) {
-    const auto s1 = siter.Value();
-    const auto s2 = fst2->AddState();
-    const auto weight = fst1.Final(s1);
-    if (weight != Weight::Zero()) {
-      fst2->ReserveArcs(s2, fst1.NumArcs(s1) + 1);
-      fst2->AddArc(s2, Arc(0, 0, weight, start2));
-    } else {
-      fst2->ReserveArcs(s2, fst1.NumArcs(s1));
-    }
-    for (ArcIterator<Fst<Arc>> aiter(fst1, s1); !aiter.Done(); aiter.Next()) {
-      auto arc = aiter.Value();
-      arc.nextstate += numstates2;
-      fst2->AddArc(s2, arc);
-    }
-  }
-  const auto start1 = fst1.Start();
-  if (start1 != kNoStateId) {
-    fst2->SetStart(start1 + numstates2);
-    fst2->SetProperties(ConcatProperties(props1, props2), kFstProperties);
-  } else {
-    fst2->SetStart(fst2->AddState());
-  }
-}
-
-// Computes the concatentation of two FSTs. This version modifies its
-// RationalFst input (in first position).
-template <class Arc>
-void Concat(RationalFst<Arc> *fst1, const Fst<Arc> &fst2) {
-  fst1->GetMutableImpl()->AddConcat(fst2, true);
-}
-
-// Computes the concatentation of two FSTs. This version modifies its
-// RationalFst input (in second position).
-template <class Arc>
-void Concat(const Fst<Arc> &fst1, RationalFst<Arc> *fst2) {
-  fst2->GetMutableImpl()->AddConcat(fst1, false);
-}
-
-using ConcatFstOptions = RationalFstOptions;
-
-// Computes the concatenation (product) of two FSTs; this version is a delayed
-// FST. If FST1 transduces string x to y with weight a and FST2 transduces
-// string w to v with weight b, then their concatenation transduces string xw
-// to yv with Times(a, b).
-//
-// Complexity:
-//
-//   Time: O(v1 + e1 + v2 + e2),
-//   Space: O(v1 + v2)
-//
-// where vi is the number of states visited, and ei is the number of arcs
-// visited, of the ith FST. Constant time and space to visit an input state or
-// arc is assumed and exclusive of caching.
-template <class A>
-class ConcatFst : public RationalFst<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  ConcatFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {
-    GetMutableImpl()->InitConcat(fst1, fst2);
-  }
-
-  ConcatFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-            const ConcatFstOptions &opts)
-      : RationalFst<Arc>(opts) {
-    GetMutableImpl()->InitConcat(fst1, fst2);
-  }
-
-  // See Fst<>::Copy() for doc.
-  ConcatFst(const ConcatFst<Arc> &fst, bool safe = false)
-      : RationalFst<Arc>(fst, safe) {}
-
-  // Get a copy of this ConcatFst. See Fst<>::Copy() for further doc.
-  ConcatFst<Arc> *Copy(bool safe = false) const override {
-    return new ConcatFst<Arc>(*this, safe);
-  }
-
- private:
-  using ImplToFst<internal::RationalFstImpl<Arc>>::GetImpl;
-  using ImplToFst<internal::RationalFstImpl<Arc>>::GetMutableImpl;
-};
-
-// Specialization for ConcatFst.
-template <class Arc>
-class StateIterator<ConcatFst<Arc>> : public StateIterator<RationalFst<Arc>> {
- public:
-  explicit StateIterator(const ConcatFst<Arc> &fst)
-      : StateIterator<RationalFst<Arc>>(fst) {}
-};
-
-// Specialization for ConcatFst.
-template <class Arc>
-class ArcIterator<ConcatFst<Arc>> : public ArcIterator<RationalFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const ConcatFst<Arc> &fst, StateId s)
-      : ArcIterator<RationalFst<Arc>>(fst, s) {}
-};
-
-// Useful alias when using StdArc.
-using StdConcatFst = ConcatFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_CONCAT_H_
diff --git a/projects/llm_framework/include/fst/config.h b/projects/llm_framework/include/fst/config.h
deleted file mode 100644
index 32f2a653..00000000
--- a/projects/llm_framework/include/fst/config.h
+++ /dev/null
@@ -1,3 +0,0 @@
-// Windows-specific OpenFst config file
-// No dynamic registration.
-#define FST_NO_DYNAMIC_LINKING 1
diff --git a/projects/llm_framework/include/fst/config.h.in b/projects/llm_framework/include/fst/config.h.in
deleted file mode 100644
index 7815dfcd..00000000
--- a/projects/llm_framework/include/fst/config.h.in
+++ /dev/null
@@ -1,11 +0,0 @@
-// OpenFst config file 
-
-/* Define to 1 if you have the ICU library. */
-#undef HAVE_ICU
-
-/* Define to 1 if the system has the type `std::tr1::hash<long long
-   unsigned>'. */
-#define HAVE_STD__TR1__HASH_LONG_LONG_UNSIGNED_ 1
-
-/* Define to 1 if the system has the type `__gnu_cxx::slist<int>'. */
-#define HAVE___GNU_CXX__SLIST_INT_ 1
diff --git a/projects/llm_framework/include/fst/connect.h b/projects/llm_framework/include/fst/connect.h
deleted file mode 100644
index 4c989292..00000000
--- a/projects/llm_framework/include/fst/connect.h
+++ /dev/null
@@ -1,323 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes and functions to remove unsuccessful paths from an FST.
-
-#ifndef FST_CONNECT_H_
-#define FST_CONNECT_H_
-
-#include <algorithm>
-#include <memory>
-#include <vector>
-
-#include <fst/dfs-visit.h>
-#include <fst/mutable-fst.h>
-#include <fst/union-find.h>
-
-
-namespace fst {
-
-// Finds and returns connected components. Use with Visit().
-template <class Arc>
-class CcVisitor {
- public:
-  using Weight = typename Arc::Weight;
-  using StateId = typename Arc::StateId;
-
-  // cc[i]: connected component number for state i.
-  explicit CcVisitor(std::vector<StateId> *cc)
-      : comps_(new UnionFind<StateId>(0, kNoStateId)), cc_(cc), nstates_(0) {}
-
-  // comps: connected components equiv classes.
-  explicit CcVisitor(UnionFind<StateId> *comps)
-      : comps_(comps), cc_(nullptr), nstates_(0) {}
-
-  ~CcVisitor() {
-    if (cc_) delete comps_;
-  }
-
-  void InitVisit(const Fst<Arc> &fst) {}
-
-  bool InitState(StateId s, StateId root) {
-    ++nstates_;
-    if (comps_->FindSet(s) == kNoStateId) comps_->MakeSet(s);
-    return true;
-  }
-
-  bool WhiteArc(StateId s, const Arc &arc) {
-    comps_->MakeSet(arc.nextstate);
-    comps_->Union(s, arc.nextstate);
-    return true;
-  }
-
-  bool GreyArc(StateId s, const Arc &arc) {
-    comps_->Union(s, arc.nextstate);
-    return true;
-  }
-
-  bool BlackArc(StateId s, const Arc &arc) {
-    comps_->Union(s, arc.nextstate);
-    return true;
-  }
-
-  void FinishState(StateId s) {}
-
-  void FinishVisit() {
-    if (cc_) GetCcVector(cc_);
-  }
-
-  // Returns number of components.
-  // cc[i]: connected component number for state i.
-  int GetCcVector(std::vector<StateId> *cc) {
-    cc->clear();
-    cc->resize(nstates_, kNoStateId);
-    StateId ncomp = 0;
-    for (StateId s = 0; s < nstates_; ++s) {
-      const auto rep = comps_->FindSet(s);
-      auto &comp = (*cc)[rep];
-      if (comp == kNoStateId) {
-        comp = ncomp;
-        ++ncomp;
-      }
-      (*cc)[s] = comp;
-    }
-    return ncomp;
-  }
-
- private:
-  UnionFind<StateId> *comps_;  // Components.
-  std::vector<StateId> *cc_;   // State's cc number.
-  StateId nstates_;            // State count.
-};
-
-// Finds and returns strongly-connected components, accessible and
-// coaccessible states and related properties. Uses Tarjan's single
-// DFS SCC algorithm (see Aho, et al, "Design and Analysis of Computer
-// Algorithms", 189pp). Use with DfsVisit();
-template <class Arc>
-class SccVisitor {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // scc[i]: strongly-connected component number for state i.
-  //   SCC numbers will be in topological order for acyclic input.
-  // access[i]: accessibility of state i.
-  // coaccess[i]: coaccessibility of state i.
-  // Any of above can be NULL.
-  // props: related property bits (cyclicity, initial cyclicity,
-  //   accessibility, coaccessibility) set/cleared (o.w. unchanged).
-  SccVisitor(std::vector<StateId> *scc, std::vector<bool> *access,
-             std::vector<bool> *coaccess, uint64 *props)
-      : scc_(scc), access_(access), coaccess_(coaccess), props_(props) {}
-  explicit SccVisitor(uint64 *props)
-      : scc_(nullptr), access_(nullptr), coaccess_(nullptr), props_(props) {}
-
-  void InitVisit(const Fst<Arc> &fst);
-
-  bool InitState(StateId s, StateId root);
-
-  bool TreeArc(StateId s, const Arc &arc) { return true; }
-
-  bool BackArc(StateId s, const Arc &arc) {
-    const auto t = arc.nextstate;
-    if ((*dfnumber_)[t] < (*lowlink_)[s]) (*lowlink_)[s] = (*dfnumber_)[t];
-    if ((*coaccess_)[t]) (*coaccess_)[s] = true;
-    *props_ |= kCyclic;
-    *props_ &= ~kAcyclic;
-    if (t == start_) {
-      *props_ |= kInitialCyclic;
-      *props_ &= ~kInitialAcyclic;
-    }
-    return true;
-  }
-
-  bool ForwardOrCrossArc(StateId s, const Arc &arc) {
-    const auto t = arc.nextstate;
-    if ((*dfnumber_)[t] < (*dfnumber_)[s] /* cross edge */ && (*onstack_)[t] &&
-        (*dfnumber_)[t] < (*lowlink_)[s]) {
-      (*lowlink_)[s] = (*dfnumber_)[t];
-    }
-    if ((*coaccess_)[t]) (*coaccess_)[s] = true;
-    return true;
-  }
-
-  // Last argument always ignored, but required by the interface.
-  void FinishState(StateId state, StateId p, const Arc *);
-
-  void FinishVisit() {
-    // Numbers SCCs in topological order when acyclic.
-    if (scc_) {
-      for (size_t s = 0; s < scc_->size(); ++s) {
-        (*scc_)[s] = nscc_ - 1 - (*scc_)[s];
-      }
-    }
-    if (coaccess_internal_) delete coaccess_;
-    dfnumber_.reset();
-    lowlink_.reset();
-    onstack_.reset();
-    scc_stack_.reset();
-  }
-
- private:
-  std::vector<StateId> *scc_;    // State's scc number.
-  std::vector<bool> *access_;    // State's accessibility.
-  std::vector<bool> *coaccess_;  // State's coaccessibility.
-  uint64 *props_;
-  const Fst<Arc> *fst_;
-  StateId start_;
-  StateId nstates_;  // State count.
-  StateId nscc_;     // SCC count.
-  bool coaccess_internal_;
-  std::unique_ptr<std::vector<StateId>> dfnumber_;  // State discovery times.
-  std::unique_ptr<std::vector<StateId>>
-      lowlink_;  // lowlink[state] == dfnumber[state] => SCC root
-  std::unique_ptr<std::vector<bool>> onstack_;  // Is a state on the SCC stack?
-  std::unique_ptr<std::vector<StateId>>
-      scc_stack_;  // SCC stack, with random access.
-};
-
-template <class Arc>
-inline void SccVisitor<Arc>::InitVisit(const Fst<Arc> &fst) {
-  if (scc_) scc_->clear();
-  if (access_) access_->clear();
-  if (coaccess_) {
-    coaccess_->clear();
-    coaccess_internal_ = false;
-  } else {
-    coaccess_ = new std::vector<bool>;
-    coaccess_internal_ = true;
-  }
-  *props_ |= kAcyclic | kInitialAcyclic | kAccessible | kCoAccessible;
-  *props_ &= ~(kCyclic | kInitialCyclic | kNotAccessible | kNotCoAccessible);
-  fst_ = &fst;
-  start_ = fst.Start();
-  nstates_ = 0;
-  nscc_ = 0;
-  dfnumber_.reset(new std::vector<StateId>());
-  lowlink_.reset(new std::vector<StateId>());
-  onstack_.reset(new std::vector<bool>());
-  scc_stack_.reset(new std::vector<StateId>());
-}
-
-template <class Arc>
-inline bool SccVisitor<Arc>::InitState(StateId s, StateId root) {
-  scc_stack_->push_back(s);
-  if (static_cast<StateId>(dfnumber_->size()) <= s) {
-    if (scc_) scc_->resize(s + 1, -1);
-    if (access_) access_->resize(s + 1, false);
-    coaccess_->resize(s + 1, false);
-    dfnumber_->resize(s + 1, -1);
-    lowlink_->resize(s + 1, -1);
-    onstack_->resize(s + 1, false);
-  }
-  (*dfnumber_)[s] = nstates_;
-  (*lowlink_)[s] = nstates_;
-  (*onstack_)[s] = true;
-  if (root == start_) {
-    if (access_) (*access_)[s] = true;
-  } else {
-    if (access_) (*access_)[s] = false;
-    *props_ |= kNotAccessible;
-    *props_ &= ~kAccessible;
-  }
-  ++nstates_;
-  return true;
-}
-
-template <class Arc>
-inline void SccVisitor<Arc>::FinishState(StateId s, StateId p, const Arc *) {
-  if (fst_->Final(s) != Weight::Zero()) (*coaccess_)[s] = true;
-  if ((*dfnumber_)[s] == (*lowlink_)[s]) {  // Root of new SCC.
-    bool scc_coaccess = false;
-    auto i = scc_stack_->size();
-    StateId t;
-    do {
-      t = (*scc_stack_)[--i];
-      if ((*coaccess_)[t]) scc_coaccess = true;
-    } while (s != t);
-    do {
-      t = scc_stack_->back();
-      if (scc_) (*scc_)[t] = nscc_;
-      if (scc_coaccess) (*coaccess_)[t] = true;
-      (*onstack_)[t] = false;
-      scc_stack_->pop_back();
-    } while (s != t);
-    if (!scc_coaccess) {
-      *props_ |= kNotCoAccessible;
-      *props_ &= ~kCoAccessible;
-    }
-    ++nscc_;
-  }
-  if (p != kNoStateId) {
-    if ((*coaccess_)[s]) (*coaccess_)[p] = true;
-    if ((*lowlink_)[s] < (*lowlink_)[p]) (*lowlink_)[p] = (*lowlink_)[s];
-  }
-}
-
-// Trims an FST, removing states and arcs that are not on successful paths.
-// This version modifies its input.
-//
-// Complexity:
-//
-//   Time:  O(V + E)
-//   Space: O(V + E)
-//
-// where V = # of states and E = # of arcs.
-template <class Arc>
-void Connect(MutableFst<Arc> *fst) {
-  using StateId = typename Arc::StateId;
-  std::vector<bool> access;
-  std::vector<bool> coaccess;
-  uint64 props = 0;
-  SccVisitor<Arc> scc_visitor(nullptr, &access, &coaccess, &props);
-  DfsVisit(*fst, &scc_visitor);
-  std::vector<StateId> dstates;
-  dstates.reserve(access.size());
-  for (StateId s = 0; s < access.size(); ++s) {
-    if (!access[s] || !coaccess[s]) dstates.push_back(s);
-  }
-  fst->DeleteStates(dstates);
-  fst->SetProperties(kAccessible | kCoAccessible, kAccessible | kCoAccessible);
-}
-
-// Returns an acyclic FST where each SCC in the input FST has been condensed to
-// a single state with transitions between SCCs retained and within SCCs
-// dropped. Also populates 'scc' with a mapping from input to output states.
-template <class Arc>
-void Condense(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-              std::vector<typename Arc::StateId> *scc) {
-  using StateId = typename Arc::StateId;
-  ofst->DeleteStates();
-  uint64 props = 0;
-  SccVisitor<Arc> scc_visitor(scc, nullptr, nullptr, &props);
-  DfsVisit(ifst, &scc_visitor);
-  const auto iter = std::max_element(scc->cbegin(), scc->cend());
-  if (iter == scc->cend()) return;
-  const StateId num_condensed_states = 1 + *iter;
-  ofst->ReserveStates(num_condensed_states);
-  for (StateId c = 0; c < num_condensed_states; ++c) {
-    ofst->AddState();
-  }
-  for (StateId s = 0; s < scc->size(); ++s) {
-    const auto c = (*scc)[s];
-    if (s == ifst.Start()) ofst->SetStart(c);
-    const auto weight = ifst.Final(s);
-    if (weight != Arc::Weight::Zero())
-      ofst->SetFinal(c, Plus(ofst->Final(c), weight));
-    for (ArcIterator<Fst<Arc>> aiter(ifst, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto nextc = (*scc)[arc.nextstate];
-      if (nextc != c) {
-        Arc condensed_arc = arc;
-        condensed_arc.nextstate = nextc;
-        ofst->AddArc(c, std::move(condensed_arc));
-      }
-    }
-  }
-  ofst->SetProperties(kAcyclic | kInitialAcyclic, kAcyclic | kInitialAcyclic);
-}
-
-}  // namespace fst
-
-#endif  // FST_CONNECT_H_
diff --git a/projects/llm_framework/include/fst/const-fst.h b/projects/llm_framework/include/fst/const-fst.h
deleted file mode 100644
index 09c81c7b..00000000
--- a/projects/llm_framework/include/fst/const-fst.h
+++ /dev/null
@@ -1,485 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Simple concrete immutable FST whose states and arcs are each stored in
-// single arrays.
-
-#ifndef FST_CONST_FST_H_
-#define FST_CONST_FST_H_
-
-#include <climits>
-#include <string>
-#include <vector>
-
-// Google-only...
-// ...Google-only
-#include <fst/log.h>
-
-#include <fst/expanded-fst.h>
-#include <fst/fst-decl.h>
-#include <fst/mapped-file.h>
-#include <fst/test-properties.h>
-#include <fst/util.h>
-
-
-namespace fst {
-
-template <class A, class Unsigned>
-class ConstFst;
-
-template <class F, class G>
-void Cast(const F &, G *);
-
-namespace internal {
-
-// States and arcs each implemented by single arrays, templated on the
-// Arc definition. Unsigned is used to represent indices into the arc array.
-template <class A, class Unsigned>
-class ConstFstImpl : public FstImpl<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<A>::SetInputSymbols;
-  using FstImpl<A>::SetOutputSymbols;
-  using FstImpl<A>::SetType;
-  using FstImpl<A>::SetProperties;
-  using FstImpl<A>::Properties;
-
-  ConstFstImpl()
-      : states_(nullptr),
-        arcs_(nullptr),
-        narcs_(0),
-        nstates_(0),
-        start_(kNoStateId) {
-    string type = "const";
-    if (sizeof(Unsigned) != sizeof(uint32)) {
-      type += std::to_string(CHAR_BIT * sizeof(Unsigned));
-    }
-    SetType(type);
-    SetProperties(kNullProperties | kStaticProperties);
-  }
-
-  explicit ConstFstImpl(const Fst<Arc> &fst);
-
-  StateId Start() const { return start_; }
-
-  Weight Final(StateId s) const { return states_[s].weight; }
-
-  StateId NumStates() const { return nstates_; }
-
-  size_t NumArcs(StateId s) const { return states_[s].narcs; }
-
-  size_t NumInputEpsilons(StateId s) const { return states_[s].niepsilons; }
-
-  size_t NumOutputEpsilons(StateId s) const { return states_[s].noepsilons; }
-
-  static ConstFstImpl<Arc, Unsigned> *Read(std::istream &strm,
-                                           const FstReadOptions &opts);
-
-  const Arc *Arcs(StateId s) const { return arcs_ + states_[s].pos; }
-
-  // Provide information needed for generic state iterator.
-  void InitStateIterator(StateIteratorData<Arc> *data) const {
-    data->base = nullptr;
-    data->nstates = nstates_;
-  }
-
-  // Provide information needed for the generic arc iterator.
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const {
-    data->base = nullptr;
-    data->arcs = arcs_ + states_[s].pos;
-    data->narcs = states_[s].narcs;
-    data->ref_count = nullptr;
-  }
-
- private:
-  // Used to find narcs_ and nstates_ in Write.
-  friend class ConstFst<Arc, Unsigned>;
-
-  // States implemented by array *states_ below, arcs by (single) *arcs_.
-  struct ConstState {
-    Weight weight;        // Final weight.
-    Unsigned pos;         // Start of state's arcs in *arcs_.
-    Unsigned narcs;       // Number of arcs (per state).
-    Unsigned niepsilons;  // Number of input epsilons.
-    Unsigned noepsilons;  // Number of output epsilons.
-
-    ConstState() : weight(Weight::Zero()) {}
-  };
-
-  // Properties always true of this FST class.
-  static constexpr uint64 kStaticProperties = kExpanded;
-  // Current unaligned file format version. The unaligned version was added and
-  // made the default since the aligned version does not work on pipes.
-  static constexpr int kFileVersion = 2;
-  // Current aligned file format version.
-  static constexpr int kAlignedFileVersion = 1;
-  // Minimum file format version supported.
-  static constexpr int kMinFileVersion = 1;
-
-  std::unique_ptr<MappedFile> states_region_;  // Mapped file for states.
-  std::unique_ptr<MappedFile> arcs_region_;    // Mapped file for arcs.
-  ConstState *states_;                         // States representation.
-  Arc *arcs_;                                  // Arcs representation.
-  size_t narcs_;                               // Number of arcs.
-  StateId nstates_;                            // Number of states.
-  StateId start_;                              // Initial state.
-
-  ConstFstImpl(const ConstFstImpl &) = delete;
-  ConstFstImpl &operator=(const ConstFstImpl &) = delete;
-};
-
-template <class Arc, class Unsigned>
-constexpr uint64 ConstFstImpl<Arc, Unsigned>::kStaticProperties;
-
-template <class Arc, class Unsigned>
-constexpr int ConstFstImpl<Arc, Unsigned>::kFileVersion;
-
-template <class Arc, class Unsigned>
-constexpr int ConstFstImpl<Arc, Unsigned>::kAlignedFileVersion;
-
-template <class Arc, class Unsigned>
-constexpr int ConstFstImpl<Arc, Unsigned>::kMinFileVersion;
-
-template <class Arc, class Unsigned>
-ConstFstImpl<Arc, Unsigned>::ConstFstImpl(const Fst<Arc> &fst)
-    : narcs_(0), nstates_(0) {
-  string type = "const";
-  if (sizeof(Unsigned) != sizeof(uint32)) {
-    type += std::to_string(CHAR_BIT * sizeof(Unsigned));
-  }
-  SetType(type);
-  SetInputSymbols(fst.InputSymbols());
-  SetOutputSymbols(fst.OutputSymbols());
-  start_ = fst.Start();
-  // Counts states and arcs.
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    ++nstates_;
-    narcs_ += fst.NumArcs(siter.Value());
-  }
-  states_region_.reset(MappedFile::Allocate(nstates_ * sizeof(*states_)));
-  arcs_region_.reset(MappedFile::Allocate(narcs_ * sizeof(*arcs_)));
-  states_ = reinterpret_cast<ConstState *>(states_region_->mutable_data());
-  arcs_ = reinterpret_cast<Arc *>(arcs_region_->mutable_data());
-  size_t pos = 0;
-  for (StateId s = 0; s < nstates_; ++s) {
-    states_[s].weight = fst.Final(s);
-    states_[s].pos = pos;
-    states_[s].narcs = 0;
-    states_[s].niepsilons = 0;
-    states_[s].noepsilons = 0;
-    for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      ++states_[s].narcs;
-      if (arc.ilabel == 0) ++states_[s].niepsilons;
-      if (arc.olabel == 0) ++states_[s].noepsilons;
-      arcs_[pos] = arc;
-      ++pos;
-    }
-  }
-  const auto props =
-      fst.Properties(kMutable, false)
-          ? fst.Properties(kCopyProperties, true)
-          : CheckProperties(
-                fst, kCopyProperties & ~kWeightedCycles & ~kUnweightedCycles,
-                kCopyProperties);
-  SetProperties(props | kStaticProperties);
-}
-
-template <class Arc, class Unsigned>
-ConstFstImpl<Arc, Unsigned> *ConstFstImpl<Arc, Unsigned>::Read(
-    std::istream &strm, const FstReadOptions &opts) {
-  using ConstState = typename ConstFstImpl<Arc, Unsigned>::ConstState;
-  std::unique_ptr<ConstFstImpl<Arc, Unsigned>> impl(
-      new ConstFstImpl<Arc, Unsigned>());
-  FstHeader hdr;
-  if (!impl->ReadHeader(strm, opts, kMinFileVersion, &hdr)) return nullptr;
-  impl->start_ = hdr.Start();
-  impl->nstates_ = hdr.NumStates();
-  impl->narcs_ = hdr.NumArcs();
-  // Ensures compatibility.
-  if (hdr.Version() == kAlignedFileVersion) {
-    hdr.SetFlags(hdr.GetFlags() | FstHeader::IS_ALIGNED);
-  }
-  if ((hdr.GetFlags() & FstHeader::IS_ALIGNED) && !AlignInput(strm)) {
-    LOG(ERROR) << "ConstFst::Read: Alignment failed: " << opts.source;
-    return nullptr;
-  }
-  size_t b = impl->nstates_ * sizeof(ConstState);
-  impl->states_region_.reset(
-      MappedFile::Map(&strm, opts.mode == FstReadOptions::MAP, opts.source, b));
-  if (!strm || !impl->states_region_) {
-    LOG(ERROR) << "ConstFst::Read: Read failed: " << opts.source;
-    return nullptr;
-  }
-  impl->states_ =
-      reinterpret_cast<ConstState *>(impl->states_region_->mutable_data());
-  if ((hdr.GetFlags() & FstHeader::IS_ALIGNED) && !AlignInput(strm)) {
-    LOG(ERROR) << "ConstFst::Read: Alignment failed: " << opts.source;
-    return nullptr;
-  }
-  b = impl->narcs_ * sizeof(Arc);
-  impl->arcs_region_.reset(
-      MappedFile::Map(&strm, opts.mode == FstReadOptions::MAP, opts.source, b));
-  if (!strm || !impl->arcs_region_) {
-    LOG(ERROR) << "ConstFst::Read: Read failed: " << opts.source;
-    return nullptr;
-  }
-  impl->arcs_ = reinterpret_cast<Arc *>(impl->arcs_region_->mutable_data());
-  return impl.release();
-}
-
-}  // namespace internal
-
-// Simple concrete immutable FST. This class attaches interface to
-// implementation and handles reference counting, delegating most methods to
-// ImplToExpandedFst. The unsigned type U is used to represent indices into the
-// arc array (default declared in fst-decl.h).
-template <class A, class Unsigned>
-class ConstFst : public ImplToExpandedFst<internal::ConstFstImpl<A, Unsigned>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  using Impl = internal::ConstFstImpl<A, Unsigned>;
-  using ConstState = typename Impl::ConstState;
-
-  friend class StateIterator<ConstFst<Arc, Unsigned>>;
-  friend class ArcIterator<ConstFst<Arc, Unsigned>>;
-
-  template <class F, class G>
-  void friend Cast(const F &, G *);
-
-  ConstFst() : ImplToExpandedFst<Impl>(std::make_shared<Impl>()) {}
-
-  explicit ConstFst(const Fst<Arc> &fst)
-      : ImplToExpandedFst<Impl>(std::make_shared<Impl>(fst)) {}
-
-  ConstFst(const ConstFst<A, Unsigned> &fst, bool safe = false)
-      : ImplToExpandedFst<Impl>(fst) {}
-
-  // Gets a copy of this ConstFst. See Fst<>::Copy() for further doc.
-  ConstFst<A, Unsigned> *Copy(bool safe = false) const override {
-    return new ConstFst<A, Unsigned>(*this, safe);
-  }
-
-  // Reads a ConstFst from an input stream, returning nullptr on error.
-  static ConstFst<A, Unsigned> *Read(std::istream &strm,
-                                     const FstReadOptions &opts) {
-    auto *impl = Impl::Read(strm, opts);
-    return impl ? new ConstFst<A, Unsigned>(std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  // Read a ConstFst from a file; return nullptr on error; empty filename reads
-  // from standard input.
-  static ConstFst<A, Unsigned> *Read(const string &filename) {
-    auto *impl = ImplToExpandedFst<Impl>::Read(filename);
-    return impl ? new ConstFst<A, Unsigned>(std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return WriteFst(*this, strm, opts);
-  }
-
-  bool Write(const string &filename) const override {
-    return Fst<Arc>::WriteFile(filename);
-  }
-
-  template <class FST>
-  static bool WriteFst(const FST &fst, std::ostream &strm,
-                       const FstWriteOptions &opts);
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const override {
-    GetImpl()->InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetImpl()->InitArcIterator(s, data);
-  }
-
- private:
-  explicit ConstFst(std::shared_ptr<Impl> impl)
-      : ImplToExpandedFst<Impl>(impl) {}
-
-  using ImplToFst<Impl, ExpandedFst<Arc>>::GetImpl;
-
-  // Uses overloading to extract the type of the argument.
-  static const Impl *GetImplIfConstFst(const ConstFst &const_fst) {
-    return const_fst.GetImpl();
-  }
-
-  // NB: this does not give privileged treatment to subtypes of ConstFst.
-  template <typename FST>
-  static Impl *GetImplIfConstFst(const FST &fst) {
-    return nullptr;
-  }
-
-  ConstFst &operator=(const ConstFst &) = delete;
-};
-
-// Writes FST in Const format, potentially with a pass over the machine before
-// writing to compute number of states and arcs.
-template <class Arc, class Unsigned>
-template <class FST>
-bool ConstFst<Arc, Unsigned>::WriteFst(const FST &fst, std::ostream &strm,
-                                       const FstWriteOptions &opts) {
-  const auto file_version =
-      opts.align ? internal::ConstFstImpl<Arc, Unsigned>::kAlignedFileVersion
-                 : internal::ConstFstImpl<Arc, Unsigned>::kFileVersion;
-  size_t num_arcs = 0;    // To silence -Wsometimes-uninitialized warnings.
-  size_t num_states = 0;  // Ditto.
-  std::streamoff start_offset = 0;
-  bool update_header = true;
-  if (const auto *impl = GetImplIfConstFst(fst)) {
-    num_arcs = impl->narcs_;
-    num_states = impl->nstates_;
-    update_header = false;
-  } else if (opts.stream_write || (start_offset = strm.tellp()) == -1) {
-    // precompute values needed for header when we cannot seek to rewrite it.
-    num_arcs = 0;
-    num_states = 0;
-    for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-      num_arcs += fst.NumArcs(siter.Value());
-      ++num_states;
-    }
-    update_header = false;
-  }
-  FstHeader hdr;
-  hdr.SetStart(fst.Start());
-  hdr.SetNumStates(num_states);
-  hdr.SetNumArcs(num_arcs);
-  string type = "const";
-  if (sizeof(Unsigned) != sizeof(uint32)) {
-    type += std::to_string(CHAR_BIT * sizeof(Unsigned));
-  }
-  const auto properties =
-      fst.Properties(kCopyProperties, true) |
-      internal::ConstFstImpl<Arc, Unsigned>::kStaticProperties;
-  internal::FstImpl<Arc>::WriteFstHeader(fst, strm, opts, file_version, type,
-                                         properties, &hdr);
-  if (opts.align && !AlignOutput(strm)) {
-    LOG(ERROR) << "Could not align file during write after header";
-    return false;
-  }
-  size_t pos = 0;
-  size_t states = 0;
-  typename ConstFst<Arc, Unsigned>::ConstState state;
-  for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    state.weight = fst.Final(s);
-    state.pos = pos;
-    state.narcs = fst.NumArcs(s);
-    state.niepsilons = fst.NumInputEpsilons(s);
-    state.noepsilons = fst.NumOutputEpsilons(s);
-    strm.write(reinterpret_cast<const char *>(&state), sizeof(state));
-    pos += state.narcs;
-    ++states;
-  }
-  hdr.SetNumStates(states);
-  hdr.SetNumArcs(pos);
-  if (opts.align && !AlignOutput(strm)) {
-    LOG(ERROR) << "Could not align file during write after writing states";
-  }
-  for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-    for (ArcIterator<FST> aiter(fst, siter.Value()); !aiter.Done();
-         aiter.Next()) {
-      const auto &arc = aiter.Value();
-// Google-only...
-#ifdef MEMORY_SANITIZER
-      // arc may contain padding which has unspecified contents. Tell MSAN to
-      // not complain about it when writing it to a file.
-      ANNOTATE_MEMORY_IS_INITIALIZED(reinterpret_cast<const char *>(&arc),
-                                     sizeof(arc));
-#endif
-      // ...Google-only
-      strm.write(reinterpret_cast<const char *>(&arc), sizeof(arc));
-    }
-  }
-  strm.flush();
-  if (!strm) {
-    LOG(ERROR) << "ConstFst::WriteFst: write failed: " << opts.source;
-    return false;
-  }
-  if (update_header) {
-    return internal::FstImpl<Arc>::UpdateFstHeader(
-        fst, strm, opts, file_version, type, properties, &hdr, start_offset);
-  } else {
-    if (hdr.NumStates() != num_states) {
-      LOG(ERROR) << "Inconsistent number of states observed during write";
-      return false;
-    }
-    if (hdr.NumArcs() != num_arcs) {
-      LOG(ERROR) << "Inconsistent number of arcs observed during write";
-      return false;
-    }
-  }
-  return true;
-}
-
-// Specialization for ConstFst; see generic version in fst.h for sample usage
-// (but use the ConstFst type instead). This version should inline.
-template <class Arc, class Unsigned>
-class StateIterator<ConstFst<Arc, Unsigned>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  explicit StateIterator(const ConstFst<Arc, Unsigned> &fst)
-      : nstates_(fst.GetImpl()->NumStates()), s_(0) {}
-
-  bool Done() const { return s_ >= nstates_; }
-
-  StateId Value() const { return s_; }
-
-  void Next() { ++s_; }
-
-  void Reset() { s_ = 0; }
-
- private:
-  const StateId nstates_;
-  StateId s_;
-};
-
-// Specialization for ConstFst; see generic version in fst.h for sample usage
-// (but use the ConstFst type instead). This version should inline.
-template <class Arc, class Unsigned>
-class ArcIterator<ConstFst<Arc, Unsigned>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const ConstFst<Arc, Unsigned> &fst, StateId s)
-      : arcs_(fst.GetImpl()->Arcs(s)),
-        narcs_(fst.GetImpl()->NumArcs(s)),
-        i_(0) {}
-
-  bool Done() const { return i_ >= narcs_; }
-
-  const Arc &Value() const { return arcs_[i_]; }
-
-  void Next() { ++i_; }
-
-  size_t Position() const { return i_; }
-
-  void Reset() { i_ = 0; }
-
-  void Seek(size_t a) { i_ = a; }
-
-  constexpr uint32 Flags() const { return kArcValueFlags; }
-
-  void SetFlags(uint32, uint32) {}
-
- private:
-  const Arc *arcs_;
-  size_t narcs_;
-  size_t i_;
-};
-
-// A useful alias when using StdArc.
-using StdConstFst = ConstFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_CONST_FST_H_
diff --git a/projects/llm_framework/include/fst/determinize.h b/projects/llm_framework/include/fst/determinize.h
deleted file mode 100644
index 736a140e..00000000
--- a/projects/llm_framework/include/fst/determinize.h
+++ /dev/null
@@ -1,1093 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to determinize an FST.
-
-#ifndef FST_DETERMINIZE_H_
-#define FST_DETERMINIZE_H_
-
-#include <algorithm>
-#include <climits>
-#include <forward_list>
-#include <map>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arc-map.h>
-#include <fst/bi-table.h>
-#include <fst/cache.h>
-#include <fst/factor-weight.h>
-#include <fst/filter-state.h>
-#include <fst/prune.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-// Common divisors are used in determinization to compute transition weights.
-// In the simplest case, it is the same as semiring Plus, but other choices
-// permit more efficient determinization when the output contains strings.
-
-// The default common divisor uses the semiring Plus.
-template <class W>
-struct DefaultCommonDivisor {
- public:
-  using Weight = W;
-
-  Weight operator()(const Weight &w1, const Weight &w2) const {
-    return Plus(w1, w2);
-  }
-};
-
-// The label common divisor for a (left) string semiring selects a single
-// letter common prefix or the empty string. This is used in the
-// determinization of output strings so that at most a single letter will
-// appear in the output of a transtion.
-template <typename Label, StringType S>
-struct LabelCommonDivisor {
- public:
-  using Weight = StringWeight<Label, S>;
-
-  Weight operator()(const Weight &w1, const Weight &w2) const {
-    typename Weight::Iterator iter1(w1);
-    typename Weight::Iterator iter2(w2);
-    if (!(StringWeight<Label, S>::Properties() & kLeftSemiring)) {
-      FSTERROR() << "LabelCommonDivisor: Weight needs to be left semiring";
-      return Weight::NoWeight();
-    } else if (w1.Size() == 0 || w2.Size() == 0) {
-      return Weight::One();
-    } else if (w1 == Weight::Zero()) {
-      return Weight(iter2.Value());
-    } else if (w2 == Weight::Zero()) {
-      return Weight(iter1.Value());
-    } else if (iter1.Value() == iter2.Value()) {
-      return Weight(iter1.Value());
-    } else {
-      return Weight::One();
-    }
-  }
-};
-
-// The gallic common divisor uses the label common divisor on the string
-// component and the common divisor on the weight component, which defaults to
-// the default common divisor.
-template <class Label, class W, GallicType G,
-          class CommonDivisor = DefaultCommonDivisor<W>>
-class GallicCommonDivisor {
- public:
-  using Weight = GallicWeight<Label, W, G>;
-
-  Weight operator()(const Weight &w1, const Weight &w2) const {
-    return Weight(label_common_divisor_(w1.Value1(), w2.Value1()),
-                  weight_common_divisor_(w1.Value2(), w2.Value2()));
-  }
-
- private:
-  LabelCommonDivisor<Label, GallicStringType(G)> label_common_divisor_;
-  CommonDivisor weight_common_divisor_;
-};
-
-// Specialization for general GALLIC weight.
-template <class Label, class W, class CommonDivisor>
-class GallicCommonDivisor<Label, W, GALLIC, CommonDivisor> {
- public:
-  using Weight = GallicWeight<Label, W, GALLIC>;
-  using GRWeight = GallicWeight<Label, W, GALLIC_RESTRICT>;
-  using Iterator =
-      UnionWeightIterator<GRWeight, GallicUnionWeightOptions<Label, W>>;
-
-  Weight operator()(const Weight &w1, const Weight &w2) const {
-    auto weight = GRWeight::Zero();
-    for (Iterator iter(w1); !iter.Done(); iter.Next()) {
-      weight = common_divisor_(weight, iter.Value());
-    }
-    for (Iterator iter(w2); !iter.Done(); iter.Next()) {
-      weight = common_divisor_(weight, iter.Value());
-    }
-    return weight == GRWeight::Zero() ? Weight::Zero() : Weight(weight);
-  }
-
- private:
-  GallicCommonDivisor<Label, W, GALLIC_RESTRICT, CommonDivisor> common_divisor_;
-};
-
-namespace internal {
-
-// Represents an element in a subset
-template <class Arc>
-struct DeterminizeElement {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  DeterminizeElement(StateId s, Weight weight)
-      : state_id(s), weight(std::move(weight)) {}
-
-  inline bool operator==(const DeterminizeElement<Arc> &element) const {
-    return state_id == element.state_id && weight == element.weight;
-  }
-
-  inline bool operator!=(const DeterminizeElement<Arc> &element) const {
-    return !(*this == element);
-  }
-
-  inline bool operator<(const DeterminizeElement<Arc> &element) const {
-    return state_id < element.state_id;
-  }
-
-  StateId state_id;  // Input state ID.
-  Weight weight;     // Residual weight.
-};
-
-// Represents a weighted subset and determinization filter state
-template <typename A, typename FilterState>
-struct DeterminizeStateTuple {
-  using Arc = A;
-  using Element = DeterminizeElement<Arc>;
-  using Subset = std::forward_list<Element>;
-
-  DeterminizeStateTuple() : filter_state(FilterState::NoState()) {}
-
-  inline bool operator==(
-      const DeterminizeStateTuple<Arc, FilterState> &tuple) const {
-    return (tuple.filter_state == filter_state) && (tuple.subset == subset);
-  }
-
-  inline bool operator!=(
-      const DeterminizeStateTuple<Arc, FilterState> &tuple) const {
-    return (tuple.filter_state != filter_state) || (tuple.subset != subset);
-  }
-
-  Subset subset;
-  FilterState filter_state;
-};
-
-// Proto-transition for determinization.
-template <class StateTuple>
-struct DeterminizeArc {
-  using Arc = typename StateTuple::Arc;
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-
-  DeterminizeArc()
-      : label(kNoLabel), weight(Weight::Zero()), dest_tuple(nullptr) {}
-
-  explicit DeterminizeArc(const Arc &arc)
-      : label(arc.ilabel), weight(Weight::Zero()), dest_tuple(new StateTuple) {}
-
-  Label label;             // Arc label.
-  Weight weight;           // Arc weight.
-  StateTuple *dest_tuple;  // Destination subset and filter state.
-};
-
-}  // namespace internal
-
-// Determinization filters are used to compute destination state tuples based
-// on the source tuple, transition, and destination element or on similar
-// super-final transition information. The filter operates on a map between a
-// label and the corresponding destination state tuples. It must define the map
-// type LabelMap. The default filter is used for weighted determinization.
-// A determinize filter for implementing weighted determinization.
-template <class Arc>
-class DefaultDeterminizeFilter {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FilterState = CharFilterState;
-  using Element = internal::DeterminizeElement<Arc>;
-  using StateTuple = internal::DeterminizeStateTuple<Arc, FilterState>;
-  using LabelMap = std::map<Label, internal::DeterminizeArc<StateTuple>>;
-
-  // This is needed e.g. to go into the gallic domain for transducers.
-  template <class A>
-  struct rebind {
-    using Other = DefaultDeterminizeFilter<A>;
-  };
-
-  explicit DefaultDeterminizeFilter(const Fst<Arc> &fst) : fst_(fst.Copy()) {}
-
-  // This is needed (e.g.) to go into the gallic domain for transducers.
-  // Ownership of the templated filter argument is given to this class.
-  template <class Filter>
-  DefaultDeterminizeFilter(const Fst<Arc> &fst, Filter *filter)
-      : fst_(fst.Copy()) {
-    delete filter;
-  }
-
-  // Copy constructor; the FST can be passed if it has been deep-copied.
-  DefaultDeterminizeFilter(const DefaultDeterminizeFilter<Arc> &filter,
-                           const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? fst->Copy() : filter.fst_->Copy()) {}
-
-  FilterState Start() const { return FilterState(0); }
-
-  // Does no work.
-  void SetState(StateId s, const StateTuple &tuple) {}
-
-  // Filters transition, possibly modifying label map. Returns true if arc is
-  // added to the label map.
-  bool FilterArc(const Arc &arc, const Element &src_element,
-                 Element &&dest_element, LabelMap *label_map) const {
-    // Adds element to unique state tuple for arc label.
-    auto &det_arc = (*label_map)[arc.ilabel];
-    if (det_arc.label == kNoLabel) {
-      det_arc = internal::DeterminizeArc<StateTuple>(arc);
-      det_arc.dest_tuple->filter_state = FilterState(0);
-    }
-    det_arc.dest_tuple->subset.push_front(std::move(dest_element));
-    return true;
-  }
-
-  // Filters super-final transition, returning new final weight.
-  Weight FilterFinal(Weight weight, const Element &element) { return weight; }
-
-  static uint64 Properties(uint64 props) { return props; }
-
- private:
-  std::unique_ptr<Fst<Arc>> fst_;
-};
-
-// Determinization state table interface:
-//
-// template <class Arc, class FilterState>
-// class DeterminizeStateTable {
-//  public:
-//   using StateId = typename Arc::StateId;
-//   using StateTuple = internal::DeterminizeStateTuple<Arc, FilterState>;
-//
-//   // Required sub-class. This is needed (e.g.) to go into the gallic domain.
-//   template <class B, class G>
-//   struct rebind {
-//     using Other = DeterminizeStateTable<B, G>;
-//   }
-//
-//   // Required constuctor.
-//   DeterminizeStateTable();
-//
-//   // Required copy constructor that does not copy state.
-//   DeterminizeStateTable(const DeterminizeStateTable<Arc, FilterState>
-//   &table);
-//
-//   // Looks up state ID by state tuple; if it doesn't exist, then adds it.
-//   // FindState takes ownership of the state tuple argument so that it
-//   // doesn't have to copy it if it creates a new state.
-//   StateId FindState(StateTuple *tuple);
-//
-//   // Looks up state tuple by ID.
-//   const StateTuple *Tuple(StateId id) const;
-// };
-
-// The default determinization state table based on the compact hash bi-table.
-template <class Arc, class FilterState>
-class DefaultDeterminizeStateTable {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using StateTuple = internal::DeterminizeStateTuple<Arc, FilterState>;
-  using Element = typename StateTuple::Element;
-  using Subset = typename StateTuple::Subset;
-
-  template <class B, class G>
-  struct rebind {
-    using Other = DefaultDeterminizeStateTable<B, G>;
-  };
-
-  explicit DefaultDeterminizeStateTable(size_t table_size = 0)
-      : table_size_(table_size), tuples_(table_size_) {}
-
-  DefaultDeterminizeStateTable(
-      const DefaultDeterminizeStateTable<Arc, FilterState> &table)
-      : table_size_(table.table_size_), tuples_(table_size_) {}
-
-  ~DefaultDeterminizeStateTable() {
-    for (StateId s = 0; s < tuples_.Size(); ++s) delete tuples_.FindEntry(s);
-  }
-
-  // Finds the state corresponding to a state tuple. Only creates a new state if
-  // the tuple is not found. FindState takes ownership of the tuple argument so
-  // that it doesn't have to copy it if it creates a new state.
-  StateId FindState(StateTuple *tuple) {
-    const StateId ns = tuples_.Size();
-    const auto s = tuples_.FindId(tuple);
-    if (s != ns) delete tuple;  // Tuple found.
-    return s;
-  }
-
-  const StateTuple *Tuple(StateId s) { return tuples_.FindEntry(s); }
-
- private:
-  // Comparison object for StateTuples.
-  class StateTupleEqual {
-   public:
-    bool operator()(const StateTuple *tuple1, const StateTuple *tuple2) const {
-      return *tuple1 == *tuple2;
-    }
-  };
-
-  // Hash function for StateTuples.
-  class StateTupleKey {
-   public:
-    size_t operator()(const StateTuple *tuple) const {
-      size_t h = tuple->filter_state.Hash();
-      for (auto it = tuple->subset.begin(); it != tuple->subset.end(); ++it) {
-        const size_t h1 = it->state_id;
-        static constexpr auto lshift = 5;
-        static constexpr auto rshift = CHAR_BIT * sizeof(size_t) - 5;
-        h ^= h << 1 ^ h1 << lshift ^ h1 >> rshift ^ it->weight.Hash();
-      }
-      return h;
-    }
-  };
-
-  size_t table_size_;
-  CompactHashBiTable<StateId, StateTuple *, StateTupleKey, StateTupleEqual,
-                     HS_STL>
-      tuples_;
-
-  DefaultDeterminizeStateTable &operator=(
-      const DefaultDeterminizeStateTable &) = delete;
-};
-
-// Determinization type.
-enum DeterminizeType {
-  // Input transducer is known to be functional (or error).
-  DETERMINIZE_FUNCTIONAL,  // Input transducer is functional (error if not).
-  // Input transducer is not known to be functional.
-  DETERMINIZE_NONFUNCTIONAL,
-  // Input transducer is not known to be functional but only keep the min of
-  // of ambiguous outputs.
-  DETERMINIZE_DISAMBIGUATE
-};
-
-// Options for finite-state transducer determinization templated on the arc
-// type, common divisor, the determinization filter and the state table.
-// DeterminizeFst takes ownership of the determinization filter and state table,
-// if provided.
-template <class Arc,
-          class CommonDivisor = DefaultCommonDivisor<typename Arc::Weight>,
-          class Filter = DefaultDeterminizeFilter<Arc>,
-          class StateTable =
-              DefaultDeterminizeStateTable<Arc, typename Filter::FilterState>>
-struct DeterminizeFstOptions : public CacheOptions {
-  using Label = typename Arc::Label;
-
-  float delta;                // Quantization delta for subset weights.
-  Label subsequential_label;  // Label used for residual final output
-                              // when producing subsequential transducers.
-  DeterminizeType type;       // Determinization type.
-  bool increment_subsequential_label;  // When creating several subsequential
-                                       // arcs at a given state, make their
-                                       // label distinct by incrementing.
-  Filter *filter;                      // Determinization filter;
-                                       // DeterminizeFst takes ownership.
-  StateTable *state_table;             // Determinization state table;
-                                       // DeterminizeFst takes ownership.
-
-  explicit DeterminizeFstOptions(const CacheOptions &opts, float delta = kDelta,
-                                 Label subsequential_label = 0,
-                                 DeterminizeType type = DETERMINIZE_FUNCTIONAL,
-                                 bool increment_subsequential_label = false,
-                                 Filter *filter = nullptr,
-                                 StateTable *state_table = nullptr)
-      : CacheOptions(opts),
-        delta(delta),
-        subsequential_label(subsequential_label),
-        type(type),
-        increment_subsequential_label(increment_subsequential_label),
-        filter(filter),
-        state_table(state_table) {}
-
-  explicit DeterminizeFstOptions(float delta = kDelta,
-                                 Label subsequential_label = 0,
-                                 DeterminizeType type = DETERMINIZE_FUNCTIONAL,
-                                 bool increment_subsequential_label = false,
-                                 Filter *filter = nullptr,
-                                 StateTable *state_table = nullptr)
-      : delta(delta),
-        subsequential_label(subsequential_label),
-        type(type),
-        increment_subsequential_label(increment_subsequential_label),
-        filter(filter),
-        state_table(state_table) {}
-};
-
-namespace internal {
-
-// Implementation of delayed DeterminizeFst. This base class is
-// common to the variants that implement acceptor and transducer
-// determinization.
-template <class Arc>
-class DeterminizeFstImplBase : public CacheImpl<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::Properties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheBaseImpl<CacheState<Arc>>::HasStart;
-  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
-  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
-  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
-  using CacheBaseImpl<CacheState<Arc>>::SetStart;
-
-  template <class CommonDivisor, class Filter, class StateTable>
-  DeterminizeFstImplBase(
-      const Fst<Arc> &fst,
-      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable> &opts)
-      : CacheImpl<Arc>(opts), fst_(fst.Copy()) {
-    SetType("determinize");
-    const auto iprops = fst.Properties(kFstProperties, false);
-    const auto dprops =
-        DeterminizeProperties(iprops, opts.subsequential_label != 0,
-                              opts.type == DETERMINIZE_NONFUNCTIONAL
-                                  ? opts.increment_subsequential_label
-                                  : true);
-    SetProperties(Filter::Properties(dprops), kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  DeterminizeFstImplBase(const DeterminizeFstImplBase<Arc> &impl)
-      : CacheImpl<Arc>(impl), fst_(impl.fst_->Copy(true)) {
-    SetType("determinize");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  virtual DeterminizeFstImplBase<Arc> *Copy() const = 0;
-
-  StateId Start() {
-    if (!HasStart()) {
-      const auto start = ComputeStart();
-      if (start != kNoStateId) SetStart(start);
-    }
-    return CacheImpl<Arc>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) SetFinal(s, ComputeFinal(s));
-    return CacheImpl<Arc>::Final(s);
-  }
-
-  virtual void Expand(StateId s) = 0;
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumOutputEpsilons(s);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<Arc>::InitArcIterator(s, data);
-  }
-
-  virtual StateId ComputeStart() = 0;
-
-  virtual Weight ComputeFinal(StateId s) = 0;
-
-  const Fst<Arc> &GetFst() const { return *fst_; }
-
- private:
-  std::unique_ptr<const Fst<Arc>> fst_;  // Input FST.
-};
-
-// Implementation of delayed determinization for weighted acceptors.
-template <class Arc, class CommonDivisor, class Filter, class StateTable>
-class DeterminizeFsaImpl : public DeterminizeFstImplBase<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FilterState = typename Filter::FilterState;
-  using StateTuple = internal::DeterminizeStateTuple<Arc, FilterState>;
-  using Element = typename StateTuple::Element;
-  using Subset = typename StateTuple::Subset;
-  using LabelMap = typename Filter::LabelMap;
-
-  using FstImpl<Arc>::SetProperties;
-  using DeterminizeFstImplBase<Arc>::GetFst;
-  using DeterminizeFstImplBase<Arc>::SetArcs;
-
-  DeterminizeFsaImpl(
-      const Fst<Arc> &fst, const std::vector<Weight> *in_dist,
-      std::vector<Weight> *out_dist,
-      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable> &opts)
-      : DeterminizeFstImplBase<Arc>(fst, opts),
-        delta_(opts.delta),
-        in_dist_(in_dist),
-        out_dist_(out_dist),
-        filter_(opts.filter ? opts.filter : new Filter(fst)),
-        state_table_(opts.state_table ? opts.state_table : new StateTable()) {
-    if (!fst.Properties(kAcceptor, true)) {
-      FSTERROR() << "DeterminizeFst: Argument not an acceptor";
-      SetProperties(kError, kError);
-    }
-    if (!(Weight::Properties() & kLeftSemiring)) {
-      FSTERROR() << "DeterminizeFst: Weight must be left distributive: "
-                 << Weight::Type();
-      SetProperties(kError, kError);
-    }
-    if (out_dist_) out_dist_->clear();
-  }
-
-  DeterminizeFsaImpl(
-      const DeterminizeFsaImpl<Arc, CommonDivisor, Filter, StateTable> &impl)
-      : DeterminizeFstImplBase<Arc>(impl),
-        delta_(impl.delta_),
-        in_dist_(nullptr),
-        out_dist_(nullptr),
-        filter_(new Filter(*impl.filter_, &GetFst())),
-        state_table_(new StateTable(*impl.state_table_)) {
-    if (impl.out_dist_) {
-      FSTERROR() << "DeterminizeFsaImpl: Cannot copy with out_dist vector";
-      SetProperties(kError, kError);
-    }
-  }
-
-  DeterminizeFsaImpl<Arc, CommonDivisor, Filter, StateTable> *Copy()
-      const override {
-    return new DeterminizeFsaImpl<Arc, CommonDivisor, Filter, StateTable>(
-        *this);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && (GetFst().Properties(kError, false))) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  StateId ComputeStart() override {
-    const auto s = GetFst().Start();
-    if (s == kNoStateId) return kNoStateId;
-    auto *tuple = new StateTuple;
-    tuple->subset.emplace_front(s, Weight::One());
-    tuple->filter_state = filter_->Start();
-    return FindState(tuple);
-  }
-
-  Weight ComputeFinal(StateId s) override {
-    const auto *tuple = state_table_->Tuple(s);
-    filter_->SetState(s, *tuple);
-    auto final_weight = Weight::Zero();
-    for (auto it = tuple->subset.begin(); it != tuple->subset.end(); ++it) {
-      const auto &element = *it;
-      final_weight =
-          Plus(final_weight,
-               Times(element.weight, GetFst().Final(element.state_id)));
-      final_weight = filter_->FilterFinal(final_weight, element);
-      if (!final_weight.Member()) SetProperties(kError, kError);
-    }
-    return final_weight;
-  }
-
-  StateId FindState(StateTuple *tuple) {
-    const auto s = state_table_->FindState(tuple);
-    if (in_dist_ && out_dist_->size() <= s) {
-      out_dist_->push_back(ComputeDistance(tuple->subset));
-    }
-    return s;
-  }
-
-  // Computes distance from a state to the final states in the DFA given the
-  // distances in the NFA.
-  Weight ComputeDistance(const Subset &subset) {
-    auto outd = Weight::Zero();
-    for (auto it = subset.begin(); it != subset.end(); ++it) {
-      const auto &element = *it;
-      const auto ind =
-          (element.state_id < in_dist_->size() ? (*in_dist_)[element.state_id]
-                                               : Weight::Zero());
-      outd = Plus(outd, Times(element.weight, ind));
-    }
-    return outd;
-  }
-
-  // Computes the outgoing transitions from a state, creating new destination
-  // states as needed.
-  void Expand(StateId s) override {
-    LabelMap label_map;
-    GetLabelMap(s, &label_map);
-    for (auto it = label_map.begin(); it != label_map.end(); ++it) {
-      AddArc(s, std::move(it->second));
-    }
-    SetArcs(s);
-  }
-
- private:
-  using DetArc = internal::DeterminizeArc<StateTuple>;
-
-  // Constructs proto-determinization transition, including destination subset,
-  // per label.
-  void GetLabelMap(StateId s, LabelMap *label_map) {
-    const auto *src_tuple = state_table_->Tuple(s);
-    filter_->SetState(s, *src_tuple);
-    for (auto it = src_tuple->subset.begin(); it != src_tuple->subset.end();
-         ++it) {
-      const auto &src_element = *it;
-      for (ArcIterator<Fst<Arc>> aiter(GetFst(), src_element.state_id);
-           !aiter.Done(); aiter.Next()) {
-        const auto &arc = aiter.Value();
-        Element dest_element(arc.nextstate,
-                             Times(src_element.weight, arc.weight));
-        filter_->FilterArc(arc, src_element, std::move(dest_element),
-                           label_map);
-      }
-    }
-    for (auto it = label_map->begin(); it != label_map->end(); ++it) {
-      NormArc(&it->second);
-    }
-  }
-
-  // Sorts subsets and removes duplicate elements, normalizing transition and
-  // subset weights.
-  void NormArc(DetArc *det_arc) {
-    auto *dest_tuple = det_arc->dest_tuple;
-    dest_tuple->subset.sort();
-    auto piter = dest_tuple->subset.begin();
-    for (auto diter = dest_tuple->subset.begin();
-         diter != dest_tuple->subset.end();) {
-      auto &dest_element = *diter;
-      auto &prev_element = *piter;
-      // Computes arc weight.
-      det_arc->weight = common_divisor_(det_arc->weight, dest_element.weight);
-      if (piter != diter && dest_element.state_id == prev_element.state_id) {
-        // Found duplicate state: sums state weight and deletes duplicate.
-        prev_element.weight = Plus(prev_element.weight, dest_element.weight);
-        if (!prev_element.weight.Member()) SetProperties(kError, kError);
-        ++diter;
-        dest_tuple->subset.erase_after(piter);
-      } else {
-        piter = diter;
-        ++diter;
-      }
-    }
-    // Divides out label weight from destination subset elements, quantizing to
-    // ensure comparisons are effective.
-    for (auto diter = dest_tuple->subset.begin();
-         diter != dest_tuple->subset.end(); ++diter) {
-      auto &dest_element = *diter;
-      dest_element.weight =
-          Divide(dest_element.weight, det_arc->weight, DIVIDE_LEFT);
-      dest_element.weight = dest_element.weight.Quantize(delta_);
-    }
-  }
-
-  // Adds an arc from state S to the destination state associated with state
-  // tuple in det_arc as created by GetLabelMap.
-  void AddArc(StateId s, DetArc &&det_arc) {
-    CacheImpl<Arc>::EmplaceArc(
-        s, det_arc.label, det_arc.label, std::move(det_arc.weight),
-        FindState(det_arc.dest_tuple));
-  }
-
-  float delta_;                         // Quantization delta for weights.
-  const std::vector<Weight> *in_dist_;  // Distance to final NFA states.
-  std::vector<Weight> *out_dist_;       // Distance to final DFA states.
-
-  // FIXME(kbg): Ought to be static const?
-  CommonDivisor common_divisor_;
-  std::unique_ptr<Filter> filter_;
-  std::unique_ptr<StateTable> state_table_;
-};
-
-// Implementation of delayed determinization for transducers. Transducer
-// determinization is implemented by mapping the input to the Gallic semiring as
-// an acceptor whose weights contain the output strings and using acceptor
-// determinization above to determinize that acceptor.
-template <class Arc, GallicType G, class CommonDivisor, class Filter,
-          class StateTable>
-class DeterminizeFstImpl : public DeterminizeFstImplBase<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using ToMapper = ToGallicMapper<Arc, G>;
-  using ToArc = typename ToMapper::ToArc;
-  using ToFst = ArcMapFst<Arc, ToArc, ToMapper>;
-  using FromMapper = FromGallicMapper<Arc, G>;
-  using FromFst = ArcMapFst<ToArc, Arc, FromMapper>;
-
-  using ToCommonDivisor = GallicCommonDivisor<Label, Weight, G, CommonDivisor>;
-  using ToFilter = typename Filter::template rebind<ToArc>::Other;
-  using ToFilterState = typename ToFilter::FilterState;
-  using ToStateTable =
-      typename StateTable::template rebind<ToArc, ToFilterState>::Other;
-  using FactorIterator = GallicFactor<Label, Weight, G>;
-
-  using FstImpl<Arc>::SetProperties;
-  using DeterminizeFstImplBase<Arc>::GetFst;
-  using CacheBaseImpl<CacheState<Arc>>::GetCacheGc;
-  using CacheBaseImpl<CacheState<Arc>>::GetCacheLimit;
-
-  DeterminizeFstImpl(
-      const Fst<Arc> &fst,
-      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable> &opts)
-      : DeterminizeFstImplBase<Arc>(fst, opts),
-        delta_(opts.delta),
-        subsequential_label_(opts.subsequential_label),
-        increment_subsequential_label_(opts.increment_subsequential_label) {
-    if (opts.state_table) {
-      FSTERROR() << "DeterminizeFst: "
-                 << "A state table can not be passed with transducer input";
-      SetProperties(kError, kError);
-      return;
-    }
-    Init(GetFst(), opts.filter);
-  }
-
-  DeterminizeFstImpl(
-      const DeterminizeFstImpl<Arc, G, CommonDivisor, Filter, StateTable> &impl)
-      : DeterminizeFstImplBase<Arc>(impl),
-        delta_(impl.delta_),
-        subsequential_label_(impl.subsequential_label_),
-        increment_subsequential_label_(impl.increment_subsequential_label_) {
-    Init(GetFst(), nullptr);
-  }
-
-  DeterminizeFstImpl<Arc, G, CommonDivisor, Filter, StateTable> *Copy()
-      const override {
-    return new DeterminizeFstImpl<Arc, G, CommonDivisor, Filter, StateTable>(
-        *this);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && (GetFst().Properties(kError, false) ||
-                            from_fst_->Properties(kError, false))) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  StateId ComputeStart() override { return from_fst_->Start(); }
-
-  Weight ComputeFinal(StateId s) override { return from_fst_->Final(s); }
-
-  void Expand(StateId s) override {
-    for (ArcIterator<FromFst> aiter(*from_fst_, s); !aiter.Done();
-         aiter.Next()) {
-      CacheImpl<Arc>::PushArc(s, aiter.Value());
-    }
-    CacheImpl<Arc>::SetArcs(s);
-  }
-
- private:
-  // Initialization of transducer determinization implementation, which is
-  // defined after DeterminizeFst since it calls it.
-  void Init(const Fst<Arc> &fst, Filter *filter);
-
-  float delta_;
-  Label subsequential_label_;
-  bool increment_subsequential_label_;
-  std::unique_ptr<FromFst> from_fst_;
-};
-
-}  // namespace internal
-
-// Determinizes a weighted transducer. This version is a delayed
-// FST. The result will be an equivalent FST that has the property
-// that no state has two transitions with the same input label.
-// For this algorithm, epsilon transitions are treated as regular
-// symbols (cf. RmEpsilon).
-//
-// The transducer must be functional. The weights must be (weakly) left
-// divisible (valid for TropicalWeight and LogWeight for instance) and be
-// zero-sum-free if for all a, b: (Plus(a, b) == 0) => a = b = 0.
-//
-// Complexity:
-//
-//   Determinizable: exponential (polynomial in the size of the output).
-//   Non-determinizable: does not terminate.
-//
-// The determinizable automata include all unweighted and all acyclic input.
-//
-// For more information, see:
-//
-// Mohri, M. 1997. Finite-state transducers in language and speech processing.
-// Computational Linguistics 23(2): 269-311.
-//
-// This class attaches interface to implementation and handles reference
-// counting, delegating most methods to ImplToFst.
-template <class A>
-class DeterminizeFst : public ImplToFst<internal::DeterminizeFstImplBase<A>> {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::DeterminizeFstImplBase<Arc>;
-
-  friend class ArcIterator<DeterminizeFst<Arc>>;
-  friend class StateIterator<DeterminizeFst<Arc>>;
-
-  template <class B, GallicType G, class CommonDivisor, class Filter,
-            class StateTable>
-  friend class DeterminizeFstImpl;
-
-  explicit DeterminizeFst(const Fst<A> &fst)
-      : ImplToFst<Impl>(CreateImpl(fst)) {}
-
-  template <class CommonDivisor, class Filter, class StateTable>
-  DeterminizeFst(
-      const Fst<Arc> &fst,
-      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable>
-          &opts =
-              DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable>())
-      : ImplToFst<Impl>(CreateImpl(fst, opts)) {}
-
-  // This acceptor-only version additionally computes the distance to final
-  // states in the output if provided with those distances for the input; this
-  // is useful for e.g., computing the k-shortest unique paths.
-  template <class CommonDivisor, class Filter, class StateTable>
-  DeterminizeFst(
-      const Fst<Arc> &fst, const std::vector<Weight> *in_dist,
-      std::vector<Weight> *out_dist,
-      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable>
-          &opts =
-              DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable>())
-      : ImplToFst<Impl>(
-            std::make_shared<internal::DeterminizeFsaImpl<Arc, CommonDivisor,
-                                                          Filter, StateTable>>(
-                fst, in_dist, out_dist, opts)) {
-    if (!fst.Properties(kAcceptor, true)) {
-      FSTERROR() << "DeterminizeFst: "
-                 << "Distance to final states computed for acceptors only";
-      GetMutableImpl()->SetProperties(kError, kError);
-    }
-  }
-
-  // See Fst<>::Copy() for doc.
-  DeterminizeFst(const DeterminizeFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(safe ? std::shared_ptr<Impl>(fst.GetImpl()->Copy())
-                             : fst.GetSharedImpl()) {}
-
-  // Get a copy of this DeterminizeFst. See Fst<>::Copy() for further doc.
-  DeterminizeFst<Arc> *Copy(bool safe = false) const override {
-    return new DeterminizeFst<Arc>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  static std::shared_ptr<Impl> CreateImpl(const Fst<Arc> &fst) {
-    using D = DefaultCommonDivisor<Weight>;
-    using F = DefaultDeterminizeFilter<Arc>;
-    using T = DefaultDeterminizeStateTable<Arc, typename F::FilterState>;
-    const DeterminizeFstOptions<Arc, D, F, T> opts;
-    return CreateImpl(fst, opts);
-  }
-
-  template <class CommonDivisor, class Filter, class StateTable>
-  static std::shared_ptr<Impl> CreateImpl(
-      const Fst<Arc> &fst,
-      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable>
-          &opts) {
-    if (fst.Properties(kAcceptor, true)) {
-      // Calls implementation for acceptors.
-      return std::make_shared<
-          internal::DeterminizeFsaImpl<Arc, CommonDivisor, Filter, StateTable>>(
-          fst, nullptr, nullptr, opts);
-    } else if (opts.type == DETERMINIZE_DISAMBIGUATE) {
-      auto rv = std::make_shared<internal::DeterminizeFstImpl<
-          Arc, GALLIC_MIN, CommonDivisor, Filter, StateTable>>(fst, opts);
-      if (!(Weight::Properties() & kPath)) {
-        FSTERROR() << "DeterminizeFst: Weight needs to have the "
-                   << "path property to disambiguate output: "
-                   << Weight::Type();
-        rv->SetProperties(kError, kError);
-      }
-      // Calls disambiguating implementation for non-functional transducers.
-      return rv;
-    } else if (opts.type == DETERMINIZE_FUNCTIONAL) {
-      // Calls implementation for functional transducers.
-      return std::make_shared<internal::DeterminizeFstImpl<
-          Arc, GALLIC_RESTRICT, CommonDivisor, Filter, StateTable>>(fst, opts);
-    } else {  // opts.type == DETERMINIZE_NONFUNCTIONAL
-      // Calls implementation for non functional transducers;
-      return std::make_shared<internal::DeterminizeFstImpl<
-          Arc, GALLIC, CommonDivisor, Filter, StateTable>>(fst, opts);
-    }
-  }
-
-  DeterminizeFst &operator=(const DeterminizeFst &) = delete;
-};
-
-namespace internal {
-
-// Initialization of transducer determinization implementation, which is defined
-// after DeterminizeFst since it calls it.
-template <class Arc, GallicType G, class D, class Filter, class T>
-void DeterminizeFstImpl<Arc, G, D, Filter, T>::Init(const Fst<Arc> &fst, Filter *filter) {
-  // Mapper to an acceptor.
-  const ToFst to_fst(fst, ToMapper());
-  auto *to_filter = filter ? new ToFilter(to_fst, filter) : nullptr;
-  // This recursive call terminates since it is to a (non-recursive)
-  // different constructor.
-  const CacheOptions copts(GetCacheGc(), GetCacheLimit());
-  const DeterminizeFstOptions<ToArc, ToCommonDivisor, ToFilter, ToStateTable>
-      dopts(copts, delta_, 0, DETERMINIZE_FUNCTIONAL, false, to_filter);
-  // Uses acceptor-only constructor to avoid template recursion.
-  const DeterminizeFst<ToArc> det_fsa(to_fst, nullptr, nullptr, dopts);
-  // Mapper back to transducer.
-  const FactorWeightOptions<ToArc> fopts(
-      CacheOptions(true, 0), delta_, kFactorFinalWeights, subsequential_label_,
-      subsequential_label_, increment_subsequential_label_,
-      increment_subsequential_label_);
-  const FactorWeightFst<ToArc, FactorIterator> factored_fst(det_fsa, fopts);
-  from_fst_.reset(new FromFst(factored_fst, FromMapper(subsequential_label_)));
-}
-
-}  // namespace internal
-
-// Specialization for DeterminizeFst.
-template <class Arc>
-class StateIterator<DeterminizeFst<Arc>>
-    : public CacheStateIterator<DeterminizeFst<Arc>> {
- public:
-  explicit StateIterator(const DeterminizeFst<Arc> &fst)
-      : CacheStateIterator<DeterminizeFst<Arc>>(fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for DeterminizeFst.
-template <class Arc>
-class ArcIterator<DeterminizeFst<Arc>>
-    : public CacheArcIterator<DeterminizeFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const DeterminizeFst<Arc> &fst, StateId s)
-      : CacheArcIterator<DeterminizeFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc>
-inline void DeterminizeFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<DeterminizeFst<Arc>>(*this);
-}
-
-// Useful aliases when using StdArc.
-using StdDeterminizeFst = DeterminizeFst<StdArc>;
-
-template <class Arc>
-struct DeterminizeOptions {
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  float delta;                // Quantization delta for subset weights.
-  Weight weight_threshold;    // Pruning weight threshold.
-  StateId state_threshold;    // Pruning state threshold.
-  Label subsequential_label;  // Label used for residual final output.
-  DeterminizeType type;
-  bool increment_subsequential_label;  // When creating several subsequential
-                                       // arcs at a given state, make their
-                                       // label distinct by incrementation?
-
-  explicit DeterminizeOptions(float delta = kDelta,
-                              Weight weight_threshold = Weight::Zero(),
-                              StateId state_threshold = kNoStateId,
-                              Label subsequential_label = 0,
-                              DeterminizeType type = DETERMINIZE_FUNCTIONAL,
-                              bool increment_subsequential_label = false)
-      : delta(delta),
-        weight_threshold(std::move(weight_threshold)),
-        state_threshold(state_threshold),
-        subsequential_label(subsequential_label),
-        type(type),
-        increment_subsequential_label(increment_subsequential_label) {}
-};
-
-// Determinizes a weighted transducer. This version writes the
-// determinized Fst to an output MutableFst. The result will be an
-// equivalent FST that has the property that no state has two
-// transitions with the same input label. For this algorithm, epsilon
-// transitions are treated as regular symbols (cf. RmEpsilon).
-//
-// The transducer must be functional. The weights must be (weakly)
-// left divisible (valid for TropicalWeight and LogWeight).
-//
-// Complexity:
-//
-//   Determinizable: exponential (polynomial in the size of the output)
-//   Non-determinizable: does not terminate
-//
-// The determinizable automata include all unweighted and all acyclic input.
-template <class Arc>
-void Determinize(
-    const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-    const DeterminizeOptions<Arc> &opts = DeterminizeOptions<Arc>()) {
-  using Weight = typename Arc::Weight;
-  DeterminizeFstOptions<Arc> nopts;
-  nopts.delta = opts.delta;
-  nopts.subsequential_label = opts.subsequential_label;
-  nopts.type = opts.type;
-  nopts.increment_subsequential_label = opts.increment_subsequential_label;
-  nopts.gc_limit = 0;  // Caches only the last state for fastest copy.
-  if (opts.weight_threshold != Weight::Zero() ||
-      opts.state_threshold != kNoStateId) {
-    if (ifst.Properties(kAcceptor, false)) {
-      std::vector<Weight> idistance;
-      std::vector<Weight> odistance;
-      ShortestDistance(ifst, &idistance, true);
-      DeterminizeFst<Arc> dfst(ifst, &idistance, &odistance, nopts);
-      PruneOptions<Arc, AnyArcFilter<Arc>> popts(
-          opts.weight_threshold, opts.state_threshold, AnyArcFilter<Arc>(),
-          &odistance);
-      Prune(dfst, ofst, popts);
-    } else {
-      *ofst = DeterminizeFst<Arc>(ifst, nopts);
-      Prune(ofst, opts.weight_threshold, opts.state_threshold);
-    }
-  } else {
-    *ofst = DeterminizeFst<Arc>(ifst, nopts);
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_DETERMINIZE_H_
diff --git a/projects/llm_framework/include/fst/dfs-visit.h b/projects/llm_framework/include/fst/dfs-visit.h
deleted file mode 100644
index a7b18a6c..00000000
--- a/projects/llm_framework/include/fst/dfs-visit.h
+++ /dev/null
@@ -1,202 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Depth-first search visitation. See visit.h for more general search queue
-// disciplines.
-
-#ifndef FST_DFS_VISIT_H_
-#define FST_DFS_VISIT_H_
-
-#include <stack>
-#include <vector>
-
-#include <fst/arcfilter.h>
-#include <fst/fst.h>
-
-
-namespace fst {
-
-// Visitor Interface: class determining actions taken during a depth-first
-// search-style visit. If any of the boolean member functions return false, the
-// DFS is aborted by first calling FinishState() on all currently grey states
-// and then calling FinishVisit().
-//
-// This is similar to the more general visitor interface in visit.h, except
-// that FinishState returns additional information appropriate only for a DFS
-// and some methods names here are better suited to a DFS.
-//
-// template <class Arc>
-// class Visitor {
-//  public:
-//   using StateId = typename Arc::StateId;
-//
-//   Visitor(T *return_data);
-//
-//   // Invoked before DFS visit.
-//   void InitVisit(const Fst<Arc> &fst);
-//
-//   // Invoked when state discovered (2nd arg is DFS tree root).
-//   bool InitState(StateId s, StateId root);
-//
-//   // Invoked when tree arc to white/undiscovered state examined.
-//   bool TreeArc(StateId s, const Arc &arc);
-//
-//   // Invoked when back arc to grey/unfinished state examined.
-//   bool BackArc(StateId s, const Arc &arc);
-//
-//   // Invoked when forward or cross arc to black/finished state examined.
-//   bool ForwardOrCrossArc(StateId s, const Arc &arc);
-//
-//   // Invoked when state finished ('s' is tree root, 'parent' is kNoStateId,
-//   // and 'arc' is nullptr).
-//   void FinishState(StateId s, StateId parent, const Arc *arc);
-//
-//   // Invoked after DFS visit.
-//   void FinishVisit();
-// };
-
-namespace internal {
-
-// An FST state's DFS stack state.
-template <class FST>
-struct DfsState {
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-
-  DfsState(const FST &fst, StateId s) : state_id(s), arc_iter(fst, s) {}
-
-  void *operator new(size_t size, MemoryPool<DfsState<FST>> *pool) {
-    return pool->Allocate();
-  }
-
-  static void Destroy(DfsState<FST> *dfs_state,
-                      MemoryPool<DfsState<FST>> *pool) {
-    if (dfs_state) {
-      dfs_state->~DfsState<FST>();
-      pool->Free(dfs_state);
-    }
-  }
-
-  StateId state_id;           // FST state.
-  ArcIterator<FST> arc_iter;  // The corresponding arcs.
-};
-
-}  // namespace internal
-
-// Performs depth-first visitation. Visitor class argument determines actions
-// and contains any return data. ArcFilter determines arcs that are considered.
-// If 'access_only' is true, performs visitation only to states accessible from
-// the initial state.
-//
-// Note this is similar to Visit() in visit.h called with a LIFO queue, except
-// this version has a Visitor class specialized and augmented for a DFS.
-template <class FST, class Visitor, class ArcFilter>
-void DfsVisit(const FST &fst, Visitor *visitor, ArcFilter filter,
-              bool access_only = false) {
-  visitor->InitVisit(fst);
-  const auto start = fst.Start();
-  if (start == kNoStateId) {
-    visitor->FinishVisit();
-    return;
-  }
-  // An FST state's DFS status
-  static constexpr uint8 kDfsWhite = 0;  // Undiscovered.
-  static constexpr uint8 kDfsGrey = 1;   // Discovered but unfinished.
-  static constexpr uint8 kDfsBlack = 2;  // Finished.
-  std::vector<uint8> state_color;
-  std::stack<internal::DfsState<FST> *> state_stack;  // DFS execution stack.
-  MemoryPool<internal::DfsState<FST>> state_pool;     // Pool for DFSStates.
-  auto nstates = start + 1;  // Number of known states in general case.
-  bool expanded = false;
-  if (fst.Properties(kExpanded, false)) {  // Tests if expanded case, then
-    nstates = CountStates(fst);            // uses ExpandedFst::NumStates().
-    expanded = true;
-  }
-  state_color.resize(nstates, kDfsWhite);
-  StateIterator<FST> siter(fst);
-  // Continue DFS while true.
-  bool dfs = true;
-  // Iterate over trees in DFS forest.
-  for (auto root = start; dfs && root < nstates;) {
-    state_color[root] = kDfsGrey;
-    state_stack.push(new (&state_pool) internal::DfsState<FST>(fst, root));
-    dfs = visitor->InitState(root, root);
-    while (!state_stack.empty()) {
-      auto *dfs_state = state_stack.top();
-      const auto s = dfs_state->state_id;
-      if (s >= static_cast<typename FST::Arc::StateId>(state_color.size())) {
-        nstates = s + 1;
-        state_color.resize(nstates, kDfsWhite);
-      }
-      ArcIterator<FST> &aiter = dfs_state->arc_iter;
-      if (!dfs || aiter.Done()) {
-        state_color[s] = kDfsBlack;
-        internal::DfsState<FST>::Destroy(dfs_state, &state_pool);
-        state_stack.pop();
-        if (!state_stack.empty()) {
-          auto *parent_state = state_stack.top();
-          auto &piter = parent_state->arc_iter;
-          visitor->FinishState(s, parent_state->state_id, &piter.Value());
-          piter.Next();
-        } else {
-          visitor->FinishState(s, kNoStateId, nullptr);
-        }
-        continue;
-      }
-      const auto &arc = aiter.Value();
-      if (arc.nextstate >= state_color.size()) {
-        nstates = arc.nextstate + 1;
-        state_color.resize(nstates, kDfsWhite);
-      }
-      if (!filter(arc)) {
-        aiter.Next();
-        continue;
-      }
-      const auto next_color = state_color[arc.nextstate];
-      switch (next_color) {
-        default:
-        case kDfsWhite:
-          dfs = visitor->TreeArc(s, arc);
-          if (!dfs) break;
-          state_color[arc.nextstate] = kDfsGrey;
-          state_stack.push(new (&state_pool)
-                               internal::DfsState<FST>(fst, arc.nextstate));
-          dfs = visitor->InitState(arc.nextstate, root);
-          break;
-        case kDfsGrey:
-          dfs = visitor->BackArc(s, arc);
-          aiter.Next();
-          break;
-        case kDfsBlack:
-          dfs = visitor->ForwardOrCrossArc(s, arc);
-          aiter.Next();
-          break;
-      }
-    }
-    if (access_only) break;
-    // Finds next tree root.
-    for (root = root == start ? 0 : root + 1;
-         root < nstates && state_color[root] != kDfsWhite; ++root) {
-    }
-    // Checks for a state beyond the largest known state.
-    if (!expanded && root == nstates) {
-      for (; !siter.Done(); siter.Next()) {
-        if (siter.Value() == nstates) {
-          ++nstates;
-          state_color.push_back(kDfsWhite);
-          break;
-        }
-      }
-    }
-  }
-  visitor->FinishVisit();
-}
-
-template <class Arc, class Visitor>
-void DfsVisit(const Fst<Arc> &fst, Visitor *visitor) {
-  DfsVisit(fst, visitor, AnyArcFilter<Arc>());
-}
-
-}  // namespace fst
-
-#endif  // FST_DFS_VISIT_H_
diff --git a/projects/llm_framework/include/fst/difference.h b/projects/llm_framework/include/fst/difference.h
deleted file mode 100644
index f073b0e1..00000000
--- a/projects/llm_framework/include/fst/difference.h
+++ /dev/null
@@ -1,205 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to compute the difference between two FSAs.
-
-#ifndef FST_DIFFERENCE_H_
-#define FST_DIFFERENCE_H_
-
-#include <memory>
-
-
-#include <fst/cache.h>
-#include <fst/complement.h>
-#include <fst/compose.h>
-
-
-namespace fst {
-
-template <class Arc, class M = Matcher<Fst<Arc>>,
-          class Filter = SequenceComposeFilter<M>,
-          class StateTable =
-              GenericComposeStateTable<Arc, typename Filter::FilterState>>
-struct DifferenceFstOptions
-    : public ComposeFstOptions<Arc, M, Filter, StateTable> {
-  explicit DifferenceFstOptions(const CacheOptions &opts = CacheOptions(),
-                                M *matcher1 = nullptr, M *matcher2 = nullptr,
-                                Filter *filter = nullptr,
-                                StateTable *state_table = nullptr)
-      : ComposeFstOptions<Arc, M, Filter, StateTable>(opts, matcher1, matcher2,
-                                                      filter, state_table) {}
-};
-
-// Computes the difference between two FSAs. This version is a delayed FST.
-// Only strings that are in the first automaton but not in second are retained
-// in the result.
-//
-// The first argument must be an acceptor; the second argument must be an
-// unweighted, epsilon-free, deterministic acceptor. One of the arguments must
-// be label-sorted.
-//
-// Complexity: same as ComposeFst.
-//
-// Caveats: same as ComposeFst.
-template <class A>
-class DifferenceFst : public ComposeFst<A> {
- public:
-  using Arc = A;
-  using Weight = typename Arc::Weight;
-  using StateId = typename Arc::StateId;
-
-  using ComposeFst<Arc>::CreateBase1;
-
-  // A - B = A ^ B'.
-  DifferenceFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-                const CacheOptions &opts = CacheOptions())
-      : ComposeFst<Arc>(CreateDifferenceImplWithCacheOpts(fst1, fst2, opts)) {
-    if (!fst1.Properties(kAcceptor, true)) {
-      FSTERROR() << "DifferenceFst: 1st argument not an acceptor";
-      GetImpl()->SetProperties(kError, kError);
-    }
-  }
-
-  template <class Matcher, class Filter, class StateTable>
-  DifferenceFst(
-      const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-      const DifferenceFstOptions<Arc, Matcher, Filter, StateTable> &opts)
-      : ComposeFst<Arc>(
-            CreateDifferenceImplWithDifferenceOpts(fst1, fst2, opts)) {
-    if (!fst1.Properties(kAcceptor, true)) {
-      FSTERROR() << "DifferenceFst: 1st argument not an acceptor";
-      GetImpl()->SetProperties(kError, kError);
-    }
-  }
-
-  // See Fst<>::Copy() for doc.
-  DifferenceFst(const DifferenceFst<Arc> &fst, bool safe = false)
-      : ComposeFst<Arc>(fst, safe) {}
-
-  // Get a copy of this DifferenceFst. See Fst<>::Copy() for further doc.
-  DifferenceFst<Arc> *Copy(bool safe = false) const override {
-    return new DifferenceFst<Arc>(*this, safe);
-  }
-
- private:
-  using Impl = internal::ComposeFstImplBase<Arc>;
-  using ImplToFst<Impl>::GetImpl;
-
-  static std::shared_ptr<Impl> CreateDifferenceImplWithCacheOpts(
-      const Fst<Arc> &fst1, const Fst<Arc> &fst2, const CacheOptions &opts) {
-    using RM = RhoMatcher<Matcher<Fst<A>>>;
-    ComplementFst<Arc> cfst(fst2);
-    ComposeFstOptions<A, RM> copts(
-        CacheOptions(), new RM(fst1, MATCH_NONE),
-        new RM(cfst, MATCH_INPUT, ComplementFst<Arc>::kRhoLabel));
-    return CreateBase1(fst1, cfst, copts);
-  }
-
-  template <class Matcher, class Filter, class StateTable>
-  static std::shared_ptr<Impl> CreateDifferenceImplWithDifferenceOpts(
-      const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-      const DifferenceFstOptions<Arc, Matcher, Filter, StateTable> &opts) {
-    using RM = RhoMatcher<Matcher>;
-    ComplementFst<Arc> cfst(fst2);
-    ComposeFstOptions<Arc, RM> copts(opts);
-    copts.matcher1 = new RM(fst1, MATCH_NONE, kNoLabel, MATCHER_REWRITE_ALWAYS,
-                            opts.matcher1);
-    copts.matcher2 = new RM(cfst, MATCH_INPUT, ComplementFst<Arc>::kRhoLabel,
-                            MATCHER_REWRITE_ALWAYS, opts.matcher2);
-    return CreateBase1(fst1, cfst, copts);
-  }
-};
-
-// Specialization for DifferenceFst.
-template <class Arc>
-class StateIterator<DifferenceFst<Arc>>
-    : public StateIterator<ComposeFst<Arc>> {
- public:
-  explicit StateIterator(const DifferenceFst<Arc> &fst)
-      : StateIterator<ComposeFst<Arc>>(fst) {}
-};
-
-// Specialization for DifferenceFst.
-template <class Arc>
-class ArcIterator<DifferenceFst<Arc>> : public ArcIterator<ComposeFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const DifferenceFst<Arc> &fst, StateId s)
-      : ArcIterator<ComposeFst<Arc>>(fst, s) {}
-};
-
-using DifferenceOptions = ComposeOptions;
-
-// Useful alias when using StdArc.
-using StdDifferenceFst = DifferenceFst<StdArc>;
-
-using DifferenceOptions = ComposeOptions;
-
-// Computes the difference between two FSAs. This version writes the difference
-// to an output MutableFst. Only strings that are in the first automaton but not
-// in the second are retained in the result.
-//
-// The first argument must be an acceptor; the second argument must be an
-// unweighted, epsilon-free, deterministic acceptor. One of the arguments must
-// be label-sorted.
-//
-// Complexity: same as Compose.
-//
-// Caveats: same as Compose.
-template <class Arc>
-void Difference(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-                MutableFst<Arc> *ofst,
-                const DifferenceOptions &opts = DifferenceOptions()) {
-  using M = Matcher<Fst<Arc>>;
-  // In each case, we cache only the last state for fastest copy.
-  switch (opts.filter_type) {
-    case AUTO_FILTER: {
-      CacheOptions nopts;
-      nopts.gc_limit = 0;
-      *ofst = DifferenceFst<Arc>(ifst1, ifst2, nopts);
-      break;
-    }
-    case SEQUENCE_FILTER: {
-      DifferenceFstOptions<Arc> dopts;
-      dopts.gc_limit = 0;
-      *ofst = DifferenceFst<Arc>(ifst1, ifst2, dopts);
-      break;
-    }
-    case ALT_SEQUENCE_FILTER: {
-      DifferenceFstOptions<Arc, M, AltSequenceComposeFilter<M>> dopts;
-      dopts.gc_limit = 0;
-      *ofst = DifferenceFst<Arc>(ifst1, ifst2, dopts);
-      break;
-    }
-    case MATCH_FILTER: {
-      DifferenceFstOptions<Arc, M, MatchComposeFilter<M>> dopts;
-      dopts.gc_limit = 0;
-      *ofst = DifferenceFst<Arc>(ifst1, ifst2, dopts);
-      break;
-    }
-    case NO_MATCH_FILTER: {
-      DifferenceFstOptions<Arc, M, NoMatchComposeFilter<M>> dopts;
-      dopts.gc_limit = 0;
-      *ofst = DifferenceFst<Arc>(ifst1, ifst2, dopts);
-      break;
-    }
-    case NULL_FILTER: {
-      DifferenceFstOptions<Arc, M, NullComposeFilter<M>> dopts;
-      dopts.gc_limit = 0;
-      *ofst = DifferenceFst<Arc>(ifst1, ifst2, dopts);
-      break;
-    }
-    case TRIVIAL_FILTER: {
-      DifferenceFstOptions<Arc, M, TrivialComposeFilter<M>> dopts;
-      dopts.gc_limit = 0;
-      *ofst = DifferenceFst<Arc>(ifst1, ifst2, dopts);
-      break;
-    }
-  }
-  if (opts.connect) Connect(ofst);
-}
-
-}  // namespace fst
-
-#endif  // FST_DIFFERENCE_H_
diff --git a/projects/llm_framework/include/fst/disambiguate.h b/projects/llm_framework/include/fst/disambiguate.h
deleted file mode 100644
index 7e3ba37c..00000000
--- a/projects/llm_framework/include/fst/disambiguate.h
+++ /dev/null
@@ -1,564 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to disambiguate an FST.
-
-#ifndef FST_DISAMBIGUATE_H_
-#define FST_DISAMBIGUATE_H_
-
-#include <list>
-#include <map>
-#include <set>
-#include <utility>
-#include <vector>
-
-#include <fst/arcsort.h>
-#include <fst/compose.h>
-#include <fst/connect.h>
-#include <fst/determinize.h>
-#include <fst/dfs-visit.h>
-#include <fst/project.h>
-#include <fst/prune.h>
-#include <fst/state-map.h>
-#include <fst/state-table.h>
-#include <fst/union-find.h>
-#include <fst/verify.h>
-
-
-namespace fst {
-
-template <class Arc>
-struct DisambiguateOptions : public DeterminizeOptions<Arc> {
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit DisambiguateOptions(float delta = kDelta,
-                               Weight weight = Weight::Zero(),
-                               StateId n = kNoStateId, Label label = 0)
-      : DeterminizeOptions<Arc>(delta, std::move(weight), n, label,
-                                DETERMINIZE_FUNCTIONAL) {}
-};
-
-namespace internal {
-
-// A determinization filter based on a subset element relation. The relation is
-// assumed to be reflexive and symmetric.
-template <class Arc, class Relation>
-class RelationDeterminizeFilter {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FilterState = IntegerFilterState<StateId>;
-  using StateTuple = DeterminizeStateTuple<Arc, FilterState>;
-  using Subset = typename StateTuple::Subset;
-  using Element = typename StateTuple::Element;
-  using LabelMap = std::multimap<Label, DeterminizeArc<StateTuple>>;
-
-  // This is needed (e.g.) to go into the gallic domain for transducers; there
-  // is no need to rebind the relation since its use here only depends on the
-  // state IDs.
-  template <class A>
-  struct rebind {
-    using Other = RelationDeterminizeFilter<A, Relation>;
-  };
-
-  explicit RelationDeterminizeFilter(const Fst<Arc> &fst)
-      : fst_(fst.Copy()), r_(new Relation()), s_(kNoStateId), head_(nullptr) {}
-
-  // Ownership of the relation is given to this class.
-  RelationDeterminizeFilter(const Fst<Arc> &fst, Relation *r)
-      : fst_(fst.Copy()), r_(r), s_(kNoStateId), head_(0) {}
-
-  // Ownership of the relation is given to this class.
-  RelationDeterminizeFilter(const Fst<Arc> &fst, Relation *r,
-                            std::vector<StateId> *head)
-      : fst_(fst.Copy()), r_(r), s_(kNoStateId), head_(head) {}
-
-  // This is needed, e.g., to go into the gallic domain for transducers.
-  // Ownership of the templated filter argument is given to this class.
-  template <class Filter>
-  RelationDeterminizeFilter(const Fst<Arc> &fst, Filter *filter)
-      : fst_(fst.Copy()),
-        r_(new Relation(filter->GetRelation())),
-        s_(kNoStateId),
-        head_(filter->GetHeadStates()) {
-    delete filter;
-  }
-
-  // Copy constructor; the FST can be passed if it has been deep-copied.
-  RelationDeterminizeFilter(const RelationDeterminizeFilter &filter,
-                            const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? fst->Copy() : filter.fst_->Copy()),
-        r_(new Relation(*filter.r_)),
-        s_(kNoStateId),
-        head_() {}
-
-  FilterState Start() const { return FilterState(fst_->Start()); }
-
-  void SetState(StateId s, const StateTuple &tuple) {
-    if (s_ != s) {
-      s_ = s;
-      tuple_ = &tuple;
-      const auto head = tuple.filter_state.GetState();
-      is_final_ = fst_->Final(head) != Weight::Zero();
-      if (head_) {
-        if (head_->size() <= s) head_->resize(s + 1, kNoStateId);
-        (*head_)[s] = head;
-      }
-    }
-  }
-
-  // Filters transition, possibly modifying label map. Returns true if arc is
-  // added to label map.
-  bool FilterArc(const Arc &arc, const Element &src_element,
-                 const Element &dest_element, LabelMap *label_map) const;
-
-  // Filters super-final transition, returning new final weight.
-  Weight FilterFinal(const Weight final_weight, const Element &element) const {
-    return is_final_ ? final_weight : Weight::Zero();
-  }
-
-  static uint64 Properties(uint64 props) {
-    return props & ~(kIDeterministic | kODeterministic);
-  }
-
-  const Relation &GetRelation() { return *r_; }
-
-  std::vector<StateId> *GetHeadStates() { return head_; }
-
- private:
-  // Pairs arc labels with state tuples with possible heads and empty subsets.
-  void InitLabelMap(LabelMap *label_map) const;
-
-  std::unique_ptr<Fst<Arc>> fst_;  // Input FST.
-  std::unique_ptr<Relation> r_;    // Relation compatible with inv. trans. fnc.
-  StateId s_;                      // Current state.
-  const StateTuple *tuple_;        // Current tuple.
-  bool is_final_;                  // Is the current head state final?
-  std::vector<StateId> *head_;     // Head state for a given state,
-                                   // owned by the Disambiguator.
-};
-
-template <class Arc, class Relation>
-bool RelationDeterminizeFilter<Arc, Relation>::FilterArc(
-    const Arc &arc, const Element &src_element, const Element &dest_element,
-    LabelMap *label_map) const {
-  bool added = false;
-  if (label_map->empty()) InitLabelMap(label_map);
-  // Adds element to state tuple if element state is related to tuple head.
-  for (auto liter = label_map->lower_bound(arc.ilabel);
-       liter != label_map->end() && liter->first == arc.ilabel; ++liter) {
-    auto *dest_tuple = liter->second.dest_tuple;
-    const auto dest_head = dest_tuple->filter_state.GetState();
-    if ((*r_)(dest_element.state_id, dest_head)) {
-      dest_tuple->subset.push_front(dest_element);
-      added = true;
-    }
-  }
-  return added;
-}
-
-template <class Arc, class Relation>
-void RelationDeterminizeFilter<Arc, Relation>::InitLabelMap(
-    LabelMap *label_map) const {
-  const auto src_head = tuple_->filter_state.GetState();
-  Label label = kNoLabel;
-  StateId nextstate = kNoStateId;
-  for (ArcIterator<Fst<Arc>> aiter(*fst_, src_head); !aiter.Done();
-       aiter.Next()) {
-    const auto &arc = aiter.Value();
-    // Continues if multiarc.
-    if (arc.ilabel == label && arc.nextstate == nextstate) continue;
-    DeterminizeArc<StateTuple> det_arc(arc);
-    det_arc.dest_tuple->filter_state = FilterState(arc.nextstate);
-    label_map->insert(std::make_pair(arc.ilabel, det_arc));
-    label = arc.ilabel;
-    nextstate = arc.nextstate;
-  }
-}
-
-// Helper class to disambiguate an FST via Disambiguate().
-template <class Arc>
-class Disambiguator {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // IDs arcs with state ID and arc position. Arc position -1 indicates final
-  // (super-final transition).
-  using ArcId = std::pair<StateId, ssize_t>;
-
-  Disambiguator() : error_(false) {}
-
-  void Disambiguate(
-      const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-      const DisambiguateOptions<Arc> &opts = DisambiguateOptions<Arc>()) {
-    VectorFst<Arc> sfst(ifst);
-    Connect(&sfst);
-    ArcSort(&sfst, ArcCompare());
-    PreDisambiguate(sfst, ofst, opts);
-    ArcSort(ofst, ArcCompare());
-    FindAmbiguities(*ofst);
-    RemoveSplits(ofst);
-    MarkAmbiguities();
-    RemoveAmbiguities(ofst);
-    if (error_) ofst->SetProperties(kError, kError);
-  }
-
- private:
-  // Comparison functor for comparing input labels and next states of arcs. This
-  // sort order facilitates the predisambiguation.
-  class ArcCompare {
-   public:
-    bool operator()(const Arc &arc1, const Arc &arc2) const {
-      return arc1.ilabel < arc2.ilabel ||
-             (arc1.ilabel == arc2.ilabel && arc1.nextstate < arc2.nextstate);
-    }
-
-    uint64 Properties(uint64 props) const {
-      return (props & kArcSortProperties) | kILabelSorted |
-             (props & kAcceptor ? kOLabelSorted : 0);
-    }
-  };
-
-  // Comparison functor for comparing transitions represented by their arc ID.
-  // This sort order facilitates ambiguity detection.
-  class ArcIdCompare {
-   public:
-    explicit ArcIdCompare(const std::vector<StateId> &head) : head_(head) {}
-
-    bool operator()(const ArcId &a1, const ArcId &a2) const {
-      // Sort first by source head state...
-      const auto src1 = a1.first;
-      const auto src2 = a2.first;
-      const auto head1 = head_[src1];
-      const auto head2 = head_[src2];
-      if (head1 < head2) return true;
-      if (head2 < head1) return false;
-      // ...then by source state...
-      if (src1 < src2) return true;
-      if (src2 < src1) return false;
-      // ...then by position.
-      return a1.second < a2.second;
-    }
-
-   private:
-    const std::vector<StateId> &head_;
-  };
-
-  // A relation that determines if two states share a common future.
-  class CommonFuture {
-   public:
-    using StateTable = GenericComposeStateTable<Arc, TrivialFilterState>;
-    using StateTuple = typename StateTable::StateTuple;
-
-    // Needed for compilation with DeterminizeRelationFilter.
-    CommonFuture() {
-      FSTERROR() << "Disambiguate::CommonFuture: FST not provided";
-    }
-
-    explicit CommonFuture(const Fst<Arc> &ifst) {
-      using M = Matcher<Fst<Arc>>;
-      ComposeFstOptions<Arc, M, NullComposeFilter<M>> opts;
-      // Ensures composition is between acceptors.
-      const bool trans = ifst.Properties(kNotAcceptor, true);
-      const auto *fsa =
-          trans ? new ProjectFst<Arc>(ifst, PROJECT_INPUT) : &ifst;
-      opts.state_table = new StateTable(*fsa, *fsa);
-      const ComposeFst<Arc> cfst(*fsa, *fsa, opts);
-      std::vector<bool> coaccess;
-      uint64 props = 0;
-      SccVisitor<Arc> scc_visitor(nullptr, nullptr, &coaccess, &props);
-      DfsVisit(cfst, &scc_visitor);
-      for (StateId s = 0; s < coaccess.size(); ++s) {
-        if (coaccess[s]) {
-          related_.insert(opts.state_table->Tuple(s).StatePair());
-        }
-      }
-      if (trans) delete fsa;
-    }
-
-    bool operator()(const StateId s1, StateId s2) const {
-      return related_.count(std::make_pair(s1, s2)) > 0;
-    }
-
-   private:
-    // States s1 and s2 resp. are in this relation iff they there is a
-    // path from s1 to a final state that has the same label as some
-    // path from s2 to a final state.
-    std::set<std::pair<StateId, StateId>> related_;
-  };
-
-  using ArcIdMap = std::multimap<ArcId, ArcId, ArcIdCompare>;
-
-  // Inserts candidate into the arc ID map.
-  inline void InsertCandidate(StateId s1, StateId s2, const ArcId &a1,
-                              const ArcId &a2) {
-    candidates_->insert(head_[s1] > head_[s2] ? std::make_pair(a1, a2)
-                                              : std::make_pair(a2, a1));
-  }
-
-  // Returns the arc corresponding to ArcId a.
-  static Arc GetArc(const Fst<Arc> &fst, ArcId aid) {
-    if (aid.second == -1) {  // Returns super-final transition.
-      return Arc(kNoLabel, kNoLabel, fst.Final(aid.first), kNoStateId);
-    } else {
-      ArcIterator<Fst<Arc>> aiter(fst, aid.first);
-      aiter.Seek(aid.second);
-      return aiter.Value();
-    }
-  }
-
-  // Outputs an equivalent FST whose states are subsets of states that have a
-  // future path in common.
-  void PreDisambiguate(const ExpandedFst<Arc> &ifst, MutableFst<Arc> *ofst,
-                       const DisambiguateOptions<Arc> &opts);
-
-  // Finds transitions that are ambiguous candidates in the result of
-  // PreDisambiguate.
-  void FindAmbiguities(const ExpandedFst<Arc> &fst);
-
-  // Finds transition pairs that are ambiguous candidates from two specified
-  // source states.
-  void FindAmbiguousPairs(const ExpandedFst<Arc> &fst, StateId s1, StateId s2);
-
-  // Marks ambiguous transitions to be removed.
-  void MarkAmbiguities();
-
-  // Deletes spurious ambiguous transitions (due to quantization).
-  void RemoveSplits(MutableFst<Arc> *ofst);
-
-  // Deletes actual ambiguous transitions.
-  void RemoveAmbiguities(MutableFst<Arc> *ofst);
-
-  // States s1 and s2 are in this relation iff there is a path from the initial
-  // state to s1 that has the same label as some path from the initial state to
-  // s2. We store only state pairs s1, s2 such that s1 <= s2.
-  std::set<std::pair<StateId, StateId>> coreachable_;
-
-  // Queue of disambiguation-related states to be processed. We store only
-  // state pairs s1, s2 such that s1 <= s2.
-  std::list<std::pair<StateId, StateId>> queue_;
-
-  // Head state in the pre-disambiguation for a given state.
-  std::vector<StateId> head_;
-
-  // Maps from a candidate ambiguous arc A to each ambiguous candidate arc B
-  // with the same label and destination state as A, whose source state s' is
-  // coreachable with the source state s of A, and for which head(s') < head(s).
-  std::unique_ptr<ArcIdMap> candidates_;
-
-  // Set of ambiguous transitions to be removed.
-  std::set<ArcId> ambiguous_;
-
-  // States to merge due to quantization issues.
-  std::unique_ptr<UnionFind<StateId>> merge_;
-  // Marks error condition.
-  bool error_;
-
-  Disambiguator(const Disambiguator &) = delete;
-  Disambiguator &operator=(const Disambiguator &) = delete;
-};
-
-template <class Arc>
-void Disambiguator<Arc>::PreDisambiguate(const ExpandedFst<Arc> &ifst,
-                                         MutableFst<Arc> *ofst,
-                                         const DisambiguateOptions<Arc> &opts) {
-  using CommonDivisor = DefaultCommonDivisor<Weight>;
-  using Filter = RelationDeterminizeFilter<Arc, CommonFuture>;
-  // Subset elements with states s1 and s2 (resp.) are in this relation iff they
-  // there is a path from s1 to a final state that has the same label as some
-  // path from s2 to a final state.
-  auto *common_future = new CommonFuture(ifst);
-  DeterminizeFstOptions<Arc, CommonDivisor, Filter> nopts;
-  nopts.delta = opts.delta;
-  nopts.subsequential_label = opts.subsequential_label;
-  nopts.filter = new Filter(ifst, common_future, &head_);
-  // The filter takes ownership of 'common_future', and determinization takes
-  // ownership of the filter itself.
-  nopts.gc_limit = 0;  // Cache only the last state for fastest copy.
-  if (opts.weight_threshold != Weight::Zero() ||
-      opts.state_threshold != kNoStateId) {
-    /* TODO(riley): fails regression test; understand why
-    if (ifst.Properties(kAcceptor, true)) {
-      std::vector<Weight> idistance, odistance;
-      ShortestDistance(ifst, &idistance, true);
-      DeterminizeFst<Arc> dfst(ifst, &idistance, &odistance, nopts);
-      PruneOptions< Arc, AnyArcFilter<Arc>> popts(opts.weight_threshold,
-                                                   opts.state_threshold,
-                                                   AnyArcFilter<Arc>(),
-                                                   &odistance);
-      Prune(dfst, ofst, popts);
-      } else */ {
-      *ofst = DeterminizeFst<Arc>(ifst, nopts);
-      Prune(ofst, opts.weight_threshold, opts.state_threshold);
-    }
-  } else {
-    *ofst = DeterminizeFst<Arc>(ifst, nopts);
-  }
-  head_.resize(ofst->NumStates(), kNoStateId);
-}
-
-template <class Arc>
-void Disambiguator<Arc>::FindAmbiguities(const ExpandedFst<Arc> &fst) {
-  if (fst.Start() == kNoStateId) return;
-  candidates_.reset(new ArcIdMap(ArcIdCompare(head_)));
-  const auto start_pr = std::make_pair(fst.Start(), fst.Start());
-  coreachable_.insert(start_pr);
-  queue_.push_back(start_pr);
-  while (!queue_.empty()) {
-    const auto &pr = queue_.front();
-    const auto s1 = pr.first;
-    const auto s2 = pr.second;
-    queue_.pop_front();
-    FindAmbiguousPairs(fst, s1, s2);
-  }
-}
-
-template <class Arc>
-void Disambiguator<Arc>::FindAmbiguousPairs(const ExpandedFst<Arc> &fst,
-                                            StateId s1, StateId s2) {
-  if (fst.NumArcs(s2) > fst.NumArcs(s1)) FindAmbiguousPairs(fst, s2, s1);
-  SortedMatcher<Fst<Arc>> matcher(fst, MATCH_INPUT);
-  matcher.SetState(s2);
-  for (ArcIterator<Fst<Arc>> aiter(fst, s1); !aiter.Done(); aiter.Next()) {
-    const auto &arc1 = aiter.Value();
-    const ArcId a1(s1, aiter.Position());
-    if (matcher.Find(arc1.ilabel)) {
-      for (; !matcher.Done(); matcher.Next()) {
-        const auto &arc2 = matcher.Value();
-        // Continues on implicit epsilon match.
-        if (arc2.ilabel == kNoLabel) continue;
-        const ArcId a2(s2, matcher.Position());
-        // Actual transition is ambiguous.
-        if (s1 != s2 && arc1.nextstate == arc2.nextstate) {
-          InsertCandidate(s1, s2, a1, a2);
-        }
-        const auto spr = arc1.nextstate <= arc2.nextstate
-                             ? std::make_pair(arc1.nextstate, arc2.nextstate)
-                             : std::make_pair(arc2.nextstate, arc1.nextstate);
-        // Not already marked as coreachable?
-        if (coreachable_.insert(spr).second) {
-          // Only possible if state split by quantization issues.
-          if (spr.first != spr.second &&
-              head_[spr.first] == head_[spr.second]) {
-            if (!merge_) {
-              merge_.reset(new UnionFind<StateId>(fst.NumStates(), kNoStateId));
-              merge_->MakeAllSet(fst.NumStates());
-            }
-            merge_->Union(spr.first, spr.second);
-          } else {
-            queue_.push_back(spr);
-          }
-        }
-      }
-    }
-  }
-  // Super-final transition is ambiguous.
-  if (s1 != s2 && fst.Final(s1) != Weight::Zero() &&
-      fst.Final(s2) != Weight::Zero()) {
-    const ArcId a1(s1, -1);
-    const ArcId a2(s2, -1);
-    InsertCandidate(s1, s2, a1, a2);
-  }
-}
-
-template <class Arc>
-void Disambiguator<Arc>::MarkAmbiguities() {
-  if (!candidates_) return;
-  for (auto it = candidates_->begin(); it != candidates_->end(); ++it) {
-    const auto a = it->first;
-    const auto b = it->second;
-    // If b is not to be removed, then a is.
-    if (ambiguous_.count(b) == 0) ambiguous_.insert(a);
-  }
-  coreachable_.clear();
-  candidates_.reset();
-}
-
-template <class Arc>
-void Disambiguator<Arc>::RemoveSplits(MutableFst<Arc> *ofst) {
-  if (!merge_) return;
-  // Merges split states to remove spurious ambiguities.
-  for (StateIterator<MutableFst<Arc>> siter(*ofst); !siter.Done();
-       siter.Next()) {
-    for (MutableArcIterator<MutableFst<Arc>> aiter(ofst, siter.Value());
-         !aiter.Done(); aiter.Next()) {
-      auto arc = aiter.Value();
-      const auto nextstate = merge_->FindSet(arc.nextstate);
-      if (nextstate != arc.nextstate) {
-        arc.nextstate = nextstate;
-        aiter.SetValue(arc);
-      }
-    }
-  }
-  // Repeats search for actual ambiguities on modified FST.
-  coreachable_.clear();
-  merge_.reset();
-  candidates_.reset();
-  FindAmbiguities(*ofst);
-  if (merge_) {  // Shouldn't get here; sanity test.
-    FSTERROR() << "Disambiguate: Unable to remove spurious ambiguities";
-    error_ = true;
-    return;
-  }
-}
-
-template <class Arc>
-void Disambiguator<Arc>::RemoveAmbiguities(MutableFst<Arc> *ofst) {
-  if (ambiguous_.empty()) return;
-  // Adds dead state to redirect ambiguous transitions to be removed.
-  const auto dead = ofst->AddState();
-  for (auto it = ambiguous_.begin(); it != ambiguous_.end(); ++it) {
-    const auto pos = it->second;
-    if (pos >= 0) {  // Actual transition.
-      MutableArcIterator<MutableFst<Arc>> aiter(ofst, it->first);
-      aiter.Seek(pos);
-      auto arc = aiter.Value();
-      arc.nextstate = dead;
-      aiter.SetValue(arc);
-    } else {  // Super-final transition.
-      ofst->SetFinal(it->first, Weight::Zero());
-    }
-  }
-  Connect(ofst);
-  ambiguous_.clear();
-}
-
-}  // namespace internal
-
-// Disambiguates a weighted FST. This version writes the disambiguated FST to an
-// output MutableFst. The result will be an equivalent FST that has the
-// property that there are not two distinct paths from the initial state to a
-// final state with the same input labeling.
-//
-// The weights must be (weakly) left divisible (valid for Tropical and
-// LogWeight).
-//
-// Complexity:
-//
-//   Disambiguable: exponential (polynomial in the size of the output).
-//   Non-disambiguable: does not terminate.
-//
-// The disambiguable transducers include all automata and functional transducers
-// that are unweighted or that are acyclic or that are unambiguous.
-//
-// For more information, see:
-//
-// Mohri, M. and Riley, M. 2015. On the disambiguation of weighted automata.
-// In CIAA, pages 263-278.
-template <class Arc>
-void Disambiguate(
-    const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-    const DisambiguateOptions<Arc> &opts = DisambiguateOptions<Arc>()) {
-  internal::Disambiguator<Arc> disambiguator;
-  disambiguator.Disambiguate(ifst, ofst, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_DISAMBIGUATE_H_
diff --git a/projects/llm_framework/include/fst/edit-fst.h b/projects/llm_framework/include/fst/edit-fst.h
deleted file mode 100644
index e7bdfd61..00000000
--- a/projects/llm_framework/include/fst/edit-fst.h
+++ /dev/null
@@ -1,702 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// An FST implementation that allows non-destructive edit operations on an
-// existing FST.
-//
-// The EditFst class enables non-destructive edit operations on a wrapped
-// ExpandedFst. The implementation uses copy-on-write semantics at the node
-// level: if a user has an underlying fst on which he or she wants to perform a
-// relatively small number of edits (read: mutations), then this implementation
-// will copy the edited node to an internal MutableFst and perform any edits in
-// situ on that copied node. This class supports all the methods of MutableFst
-// except for DeleteStates(const std::vector<StateId> &); thus, new nodes may
-// also be
-// added, and one may add transitions from existing nodes of the wrapped fst to
-// new nodes.
-//
-// N.B.: The documentation for Fst::Copy(true) says that its behavior is
-// undefined if invoked on an fst that has already been accessed.  This class
-// requires that the Fst implementation it wraps provides consistent, reliable
-// behavior when its Copy(true) method is invoked, where consistent means
-// the graph structure, graph properties and state numbering and do not change.
-// VectorFst and CompactFst, for example, are both well-behaved in this regard.
-
-#ifndef FST_EDIT_FST_H_
-#define FST_EDIT_FST_H_
-
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-
-
-namespace fst {
-namespace internal {
-
-// The EditFstData class is a container for all mutable data for EditFstImpl;
-// also, this class provides most of the actual implementation of what EditFst
-// does (that is, most of EditFstImpl's methods delegate to methods in this, the
-// EditFstData class).  Instances of this class are reference-counted and can be
-// shared between otherwise independent EditFstImpl instances. This scheme
-// allows EditFstImpl to implement the thread-safe, copy-on-write semantics
-// required by Fst::Copy(true).
-//
-// template parameters:
-//   A the type of arc to use
-//   WrappedFstT the type of fst wrapped by the EditFst instance that
-//     this EditFstData instance is backing
-//   MutableFstT the type of mutable fst to use internally for edited states;
-//     crucially, MutableFstT::Copy(false) *must* yield an fst that is
-//     thread-safe for reading (VectorFst, for example, has this property)
-template <typename Arc, typename WrappedFstT = ExpandedFst<Arc>,
-          typename MutableFstT = VectorFst<Arc>>
-class EditFstData {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  EditFstData() : num_new_states_(0) {}
-
-  EditFstData(const EditFstData &other)
-      : edits_(other.edits_),
-        external_to_internal_ids_(other.external_to_internal_ids_),
-        edited_final_weights_(other.edited_final_weights_),
-        num_new_states_(other.num_new_states_) {}
-
-  ~EditFstData() {}
-
-  static EditFstData<Arc, WrappedFstT, MutableFstT> *Read(
-      std::istream &strm, const FstReadOptions &opts);
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
-    // Serialize all private data members of this class.
-    FstWriteOptions edits_opts(opts);
-    edits_opts.write_header = true;  // Force writing contained header.
-    edits_.Write(strm, edits_opts);
-    WriteType(strm, external_to_internal_ids_);
-    WriteType(strm, edited_final_weights_);
-    WriteType(strm, num_new_states_);
-    if (!strm) {
-      LOG(ERROR) << "EditFstData::Write: Write failed: " << opts.source;
-      return false;
-    }
-    return true;
-  }
-
-  StateId NumNewStates() const { return num_new_states_; }
-
-  // accessor methods for the fst holding edited states
-  StateId EditedStart() const { return edits_.Start(); }
-
-  Weight Final(StateId s, const WrappedFstT *wrapped) const {
-    auto final_weight_it = GetFinalWeightIterator(s);
-    if (final_weight_it == NotInFinalWeightMap()) {
-      auto it = GetEditedIdMapIterator(s);
-      return it == NotInEditedMap() ? wrapped->Final(s)
-                                    : edits_.Final(it->second);
-    } else {
-      return final_weight_it->second;
-    }
-  }
-
-  size_t NumArcs(StateId s, const WrappedFstT *wrapped) const {
-    auto it = GetEditedIdMapIterator(s);
-    return it == NotInEditedMap() ? wrapped->NumArcs(s)
-                                  : edits_.NumArcs(it->second);
-  }
-
-  size_t NumInputEpsilons(StateId s, const WrappedFstT *wrapped) const {
-    auto it = GetEditedIdMapIterator(s);
-    return it == NotInEditedMap() ? wrapped->NumInputEpsilons(s)
-                                  : edits_.NumInputEpsilons(it->second);
-  }
-
-  size_t NumOutputEpsilons(StateId s, const WrappedFstT *wrapped) const {
-    auto it = GetEditedIdMapIterator(s);
-    return it == NotInEditedMap() ? wrapped->NumOutputEpsilons(s)
-                                  : edits_.NumOutputEpsilons(it->second);
-  }
-
-  void SetEditedProperties(uint64 props, uint64 mask) {
-    edits_.SetProperties(props, mask);
-  }
-
-  // Non-const MutableFst operations.
-
-  // Sets the start state for this FST.
-  void SetStart(StateId s) { edits_.SetStart(s); }
-
-  // Sets the final state for this FST.
-  Weight SetFinal(StateId s, Weight w, const WrappedFstT *wrapped) {
-    Weight old_weight = Final(s, wrapped);
-    auto it = GetEditedIdMapIterator(s);
-    // If we haven't already edited state s, don't add it to edited_ (which can
-    // be expensive if s has many transitions); just use the
-    // edited_final_weights_ map.
-    if (it == NotInEditedMap()) {
-      edited_final_weights_[s] = w;
-    } else {
-      edits_.SetFinal(GetEditableInternalId(s, wrapped), w);
-    }
-    return old_weight;
-  }
-
-  // Adds a new state to this FST, initially with no arcs.
-  StateId AddState(StateId curr_num_states) {
-    StateId internal_state_id = edits_.AddState();
-    StateId external_state_id = curr_num_states;
-    external_to_internal_ids_[external_state_id] = internal_state_id;
-    num_new_states_++;
-    return external_state_id;
-  }
-
-  // Adds the specified arc to the specified state of this FST.
-  const Arc *AddArc(StateId s, const Arc &arc, const WrappedFstT *wrapped) {
-    const auto internal_id = GetEditableInternalId(s, wrapped);
-    const auto num_arcs = edits_.NumArcs(internal_id);
-    ArcIterator<MutableFstT> arc_it(edits_, internal_id);
-    const Arc *prev_arc = nullptr;
-    if (num_arcs > 0) {
-      // grab the final arc associated with this state in edits_
-      arc_it.Seek(num_arcs - 1);
-      prev_arc = &(arc_it.Value());
-    }
-    edits_.AddArc(internal_id, arc);
-    return prev_arc;
-  }
-
-  void DeleteStates() {
-    edits_.DeleteStates();
-    num_new_states_ = 0;
-    external_to_internal_ids_.clear();
-    edited_final_weights_.clear();
-  }
-
-  // Removes all but the first n outgoing arcs of the specified state.
-  void DeleteArcs(StateId s, size_t n, const WrappedFstT *wrapped) {
-    edits_.DeleteArcs(GetEditableInternalId(s, wrapped), n);
-  }
-
-  // Removes all outgoing arcs from the specified state.
-  void DeleteArcs(StateId s, const WrappedFstT *wrapped) {
-    edits_.DeleteArcs(GetEditableInternalId(s, wrapped));
-  }
-
-  // End methods for non-const MutableFst operations.
-
-  // Provides information for the generic arc iterator.
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data,
-                       const WrappedFstT *wrapped) const {
-    auto id_map_it = GetEditedIdMapIterator(s);
-    if (id_map_it == NotInEditedMap()) {
-      VLOG(3) << "EditFstData::InitArcIterator: iterating on state " << s
-              << " of original fst";
-      wrapped->InitArcIterator(s, data);
-    } else {
-      VLOG(2) << "EditFstData::InitArcIterator: iterating on edited state " << s
-              << " (internal state id: " << id_map_it->second << ")";
-      edits_.InitArcIterator(id_map_it->second, data);
-    }
-  }
-
-  // Provides information for the generic mutable arc iterator.
-  void InitMutableArcIterator(StateId s, MutableArcIteratorData<Arc> *data,
-                              const WrappedFstT *wrapped) {
-    data->base = new MutableArcIterator<MutableFstT>(
-        &edits_, GetEditableInternalId(s, wrapped));
-  }
-
-  // Prints out the map from external to internal state id's (for debugging
-  // purposes).
-  void PrintMap() {
-    for (auto map_it = external_to_internal_ids_.begin();
-         map_it != NotInEditedMap(); ++map_it) {
-      LOG(INFO) << "(external,internal)=(" << map_it->first << ","
-                << map_it->second << ")";
-    }
-  }
-
- private:
-  // Returns the iterator of the map from external to internal state id's
-  // of edits_ for the specified external state id.
-  typename std::unordered_map<StateId, StateId>::const_iterator
-      GetEditedIdMapIterator(StateId s) const {
-    return external_to_internal_ids_.find(s);
-  }
-
-  typename std::unordered_map<StateId, StateId>::const_iterator
-      NotInEditedMap() const {
-    return external_to_internal_ids_.end();
-  }
-
-  typename std::unordered_map<StateId, Weight>::const_iterator
-      GetFinalWeightIterator(StateId s) const {
-    return edited_final_weights_.find(s);
-  }
-
-  typename std::unordered_map<StateId, Weight>::const_iterator
-      NotInFinalWeightMap() const {
-    return edited_final_weights_.end();
-  }
-
-  // Returns the internal state ID of the specified external ID if the state has
-  // already been made editable, or else copies the state from wrapped_ to
-  // edits_ and returns the state id of the newly editable state in edits_.
-  StateId GetEditableInternalId(StateId s, const WrappedFstT *wrapped) {
-    auto id_map_it = GetEditedIdMapIterator(s);
-    if (id_map_it == NotInEditedMap()) {
-      StateId new_internal_id = edits_.AddState();
-      VLOG(2) << "EditFstData::GetEditableInternalId: editing state " << s
-              << " of original fst; new internal state id:" << new_internal_id;
-      external_to_internal_ids_[s] = new_internal_id;
-      for (ArcIterator<Fst<Arc>> arc_iterator(*wrapped, s);
-           !arc_iterator.Done(); arc_iterator.Next()) {
-        edits_.AddArc(new_internal_id, arc_iterator.Value());
-      }
-      // Copies the final weight.
-      auto final_weight_it = GetFinalWeightIterator(s);
-      if (final_weight_it == NotInFinalWeightMap()) {
-        edits_.SetFinal(new_internal_id, wrapped->Final(s));
-      } else {
-        edits_.SetFinal(new_internal_id, final_weight_it->second);
-        edited_final_weights_.erase(s);
-      }
-      return new_internal_id;
-    } else {
-      return id_map_it->second;
-    }
-  }
-
-  // A mutable FST (by default, a VectorFst) to contain new states, and/or
-  // copies of states from a wrapped ExpandedFst that have been modified in
-  // some way.
-  MutableFstT edits_;
-  // A mapping from external state IDs to the internal IDs of states that
-  // appear in edits_.
-  std::unordered_map<StateId, StateId> external_to_internal_ids_;
-  // A mapping from external state IDs to final state weights assigned to
-  // those states.  The states in this map are *only* those whose final weight
-  // has been modified; if any other part of the state has been modified,
-  // the entire state is copied to edits_, and all modifications reside there.
-  std::unordered_map<StateId, Weight> edited_final_weights_;
-  // The number of new states added to this mutable fst impl, which is <= the
-  // number of states in edits_ (since edits_ contains both edited *and* new
-  // states).
-  StateId num_new_states_;
-};
-
-// EditFstData method implementations: just the Read method.
-template <typename A, typename WrappedFstT, typename MutableFstT>
-EditFstData<A, WrappedFstT, MutableFstT> *
-EditFstData<A, WrappedFstT, MutableFstT>::Read(std::istream &strm,
-                                               const FstReadOptions &opts) {
-  auto *data = new EditFstData<A, WrappedFstT, MutableFstT>();
-  // next read in MutabelFstT machine that stores edits
-  FstReadOptions edits_opts(opts);
-  // Contained header was written out, so read it in.
-  edits_opts.header = nullptr;
-
-  // Because our internal representation of edited states is a solid object
-  // of type MutableFstT (defaults to VectorFst<A>) and not a pointer,
-  // and because the static Read method allocates a new object on the heap,
-  // we need to call Read, check if there was a failure, use
-  // MutableFstT::operator= to assign the object (not the pointer) to the
-  // edits_ data member (which will increase the ref count by 1 on the impl)
-  // and, finally, delete the heap-allocated object.
-  std::unique_ptr<MutableFstT> edits(MutableFstT::Read(strm, edits_opts));
-  if (!edits) return nullptr;
-  data->edits_ = *edits;
-  edits.reset();
-  // Finally, reads in rest of private data members.
-  ReadType(strm, &data->external_to_internal_ids_);
-  ReadType(strm, &data->edited_final_weights_);
-  ReadType(strm, &data->num_new_states_);
-  if (!strm) {
-    LOG(ERROR) << "EditFst::Read: read failed: " << opts.source;
-    return nullptr;
-  }
-  return data;
-}
-
-// This class enables non-destructive edit operations on a wrapped ExpandedFst.
-// The implementation uses copy-on-write semantics at the node level: if a user
-// has an underlying fst on which he or she wants to perform a relatively small
-// number of edits (read: mutations), then this implementation will copy the
-// edited node to an internal MutableFst and perform any edits in situ on that
-// copied node. This class supports all the methods of MutableFst except for
-// DeleteStates(const std::vector<StateId> &); thus, new nodes may also be
-// added, and
-// one may add transitions from existing nodes of the wrapped fst to new nodes.
-//
-// template parameters:
-//   A the type of arc to use
-//   WrappedFstT the type of fst wrapped by the EditFst instance that
-//     this EditFstImpl instance is backing
-//   MutableFstT the type of mutable fst to use internally for edited states;
-//     crucially, MutableFstT::Copy(false) *must* yield an fst that is
-//     thread-safe for reading (VectorFst, for example, has this property)
-template <typename A, typename WrappedFstT = ExpandedFst<A>,
-          typename MutableFstT = VectorFst<A>>
-class EditFstImpl : public FstImpl<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-  using FstImpl<Arc>::WriteHeader;
-
-  // Constructs an editable FST implementation with no states. Effectively, this
-  // initially-empty fst will in every way mimic the behavior of a
-  // VectorFst---more precisely, a VectorFstImpl instance---but with slightly
-  // slower performance (by a constant factor), due to the fact that
-  // this class maintains a mapping between external state id's and
-  // their internal equivalents.
-  EditFstImpl() : wrapped_(new MutableFstT()) {
-    FstImpl<Arc>::SetType("edit");
-    InheritPropertiesFromWrapped();
-    data_ = std::make_shared<EditFstData<Arc, WrappedFstT, MutableFstT>>();
-  }
-
-  // Wraps the specified ExpandedFst. This constructor requires that the
-  // specified Fst is an ExpandedFst instance. This requirement is only enforced
-  // at runtime. (See below for the reason.)
-  //
-  // This library uses the pointer-to-implementation or "PIMPL" design pattern.
-  // In particular, to make it convenient to bind an implementation class to its
-  // interface, there are a pair of template "binder" classes, one for immutable
-  // and one for mutable fst's (ImplToFst and ImplToMutableFst, respectively).
-  // As it happens, the API for the ImplToMutableFst<I,F> class requires that
-  // the implementation class--the template parameter "I"--have a constructor
-  // taking a const Fst<A> reference.  Accordingly, the constructor here must
-  // perform a static_cast to the WrappedFstT type required by EditFst and
-  // therefore EditFstImpl.
-  explicit EditFstImpl(const Fst<Arc> &wrapped)
-      : wrapped_(static_cast<WrappedFstT *>(wrapped.Copy())) {
-    FstImpl<Arc>::SetType("edit");
-    data_ = std::make_shared<EditFstData<Arc, WrappedFstT, MutableFstT>>();
-    // have edits_ inherit all properties from wrapped_
-    data_->SetEditedProperties(wrapped_->Properties(kFstProperties, false),
-                               kFstProperties);
-    InheritPropertiesFromWrapped();
-  }
-
-  // A copy constructor for this implementation class, used to implement
-  // the Copy() method of the Fst interface.
-  EditFstImpl(const EditFstImpl &impl)
-      : FstImpl<Arc>(),
-        wrapped_(static_cast<WrappedFstT *>(impl.wrapped_->Copy(true))),
-        data_(impl.data_) {
-    SetProperties(impl.Properties());
-  }
-
-  // const Fst/ExpandedFst operations, declared in the Fst and ExpandedFst
-  // interfaces
-  StateId Start() const {
-    const auto edited_start = data_->EditedStart();
-    return edited_start == kNoStateId ? wrapped_->Start() : edited_start;
-  }
-
-  Weight Final(StateId s) const { return data_->Final(s, wrapped_.get()); }
-
-  size_t NumArcs(StateId s) const { return data_->NumArcs(s, wrapped_.get()); }
-
-  size_t NumInputEpsilons(StateId s) const {
-    return data_->NumInputEpsilons(s, wrapped_.get());
-  }
-
-  size_t NumOutputEpsilons(StateId s) const {
-    return data_->NumOutputEpsilons(s, wrapped_.get());
-  }
-
-  StateId NumStates() const {
-    return wrapped_->NumStates() + data_->NumNewStates();
-  }
-
-  static EditFstImpl<Arc, WrappedFstT, MutableFstT> *Read(
-      std::istream &strm, const FstReadOptions &opts);
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
-    FstHeader hdr;
-    hdr.SetStart(Start());
-    hdr.SetNumStates(NumStates());
-    FstWriteOptions header_opts(opts);
-    // Allows the contained FST to hold any symbols.
-    header_opts.write_isymbols = false;
-    header_opts.write_osymbols = false;
-    WriteHeader(strm, header_opts, kFileVersion, &hdr);
-    // First, serializes the wrapped FST to stream.
-    FstWriteOptions wrapped_opts(opts);
-    // Forcse writing the contained header.
-    wrapped_opts.write_header = true;
-    wrapped_->Write(strm, wrapped_opts);
-    data_->Write(strm, opts);
-    strm.flush();
-    if (!strm) {
-      LOG(ERROR) << "EditFst::Write: Write failed: " << opts.source;
-      return false;
-    }
-    return true;
-  }
-
-  // Sets the start state for this FST.
-  void SetStart(StateId s) {
-    MutateCheck();
-    data_->SetStart(s);
-    SetProperties(SetStartProperties(FstImpl<Arc>::Properties()));
-  }
-
-  // Sets the final state for this fst.
-  void SetFinal(StateId s, Weight weight) {
-    MutateCheck();
-    Weight old_weight = data_->SetFinal(s, weight, wrapped_.get());
-    SetProperties(
-        SetFinalProperties(FstImpl<Arc>::Properties(), old_weight, weight));
-  }
-
-  // Adds a new state to this fst, initially with no arcs.
-  StateId AddState() {
-    MutateCheck();
-    SetProperties(AddStateProperties(FstImpl<Arc>::Properties()));
-    return data_->AddState(NumStates());
-  }
-
-  // Adds the specified arc to the specified state of this fst.
-  void AddArc(StateId s, const Arc &arc) {
-    MutateCheck();
-    const auto *prev_arc = data_->AddArc(s, arc, wrapped_.get());
-    SetProperties(
-        AddArcProperties(FstImpl<Arc>::Properties(), s, arc, prev_arc));
-  }
-
-  void DeleteStates(const std::vector<StateId> &dstates) {
-    FSTERROR() << ": EditFstImpl::DeleteStates(const std::vector<StateId>&): "
-               << " not implemented";
-    SetProperties(kError, kError);
-  }
-
-  // Deletes all states in this fst.
-  void DeleteStates();
-
-  // Removes all but the first n outgoing arcs of the specified state.
-  void DeleteArcs(StateId s, size_t n) {
-    MutateCheck();
-    data_->DeleteArcs(s, n, wrapped_.get());
-    SetProperties(DeleteArcsProperties(FstImpl<Arc>::Properties()));
-  }
-
-  // Removes all outgoing arcs from the specified state.
-  void DeleteArcs(StateId s) {
-    MutateCheck();
-    data_->DeleteArcs(s, wrapped_.get());
-    SetProperties(DeleteArcsProperties(FstImpl<Arc>::Properties()));
-  }
-
-  void ReserveStates(StateId s) {}
-
-  void ReserveArcs(StateId s, size_t n) {}
-
-  // Ends non-const MutableFst operations.
-
-  // Provides information for the generic state iterator.
-  void InitStateIterator(StateIteratorData<Arc> *data) const {
-    data->base = nullptr;
-    data->nstates = NumStates();
-  }
-
-  // Provides information for the generic arc iterator.
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const {
-    data_->InitArcIterator(s, data, wrapped_.get());
-  }
-
-  // Provides information for the generic mutable arc iterator.
-  void InitMutableArcIterator(StateId s, MutableArcIteratorData<Arc> *data) {
-    MutateCheck();
-    data_->InitMutableArcIterator(s, data, wrapped_.get());
-  }
-
- private:
-  // Properties always true of this FST class.
-  static constexpr uint64 kStaticProperties = kExpanded | kMutable;
-  // Current file format version.
-  static constexpr int kFileVersion = 2;
-  // Minimum file format version supported
-  static constexpr int kMinFileVersion = 2;
-
-  // Causes this FST to inherit all the properties from its wrapped FST, except
-  // for the two properties that always apply to EditFst instances: kExpanded
-  // and kMutable.
-  void InheritPropertiesFromWrapped() {
-    SetProperties(wrapped_->Properties(kCopyProperties, false) |
-                  kStaticProperties);
-    SetInputSymbols(wrapped_->InputSymbols());
-    SetOutputSymbols(wrapped_->OutputSymbols());
-  }
-
-  // This method ensures that any operations that alter the mutable data
-  // portion of this EditFstImpl cause the data_ member to be copied when its
-  // reference count is greater than 1.  Note that this method is distinct from
-  // MutableFst::Mutate, which gets invoked whenever one of the basic mutation
-  // methods defined in MutableFst is invoked, such as SetInputSymbols.
-  // The MutateCheck here in EditFstImpl is invoked whenever one of the
-  // mutating methods specifically related to the types of edits provided
-  // by EditFst is performed, such as changing an arc of an existing state
-  // of the wrapped fst via a MutableArcIterator, or adding a new state via
-  // AddState().
-  void MutateCheck() {
-    if (!data_.unique()) {
-      data_ =
-          std::make_shared<EditFstData<Arc, WrappedFstT, MutableFstT>>(*data_);
-    }
-  }
-
-  // The FST that this FST wraps. The purpose of this class is to enable
-  // non-destructive edits on this wrapped FST.
-  std::unique_ptr<const WrappedFstT> wrapped_;
-  // The mutable data for this EditFst instance, with delegates for all the
-  // methods that can mutate data.
-  std::shared_ptr<EditFstData<Arc, WrappedFstT, MutableFstT>> data_;
-};
-
-template <typename Arc, typename WrappedFstT, typename MutableFstT>
-constexpr uint64 EditFstImpl<Arc, WrappedFstT, MutableFstT>::kStaticProperties;
-
-template <typename Arc, typename WrappedFstT, typename MutableFstT>
-constexpr int EditFstImpl<Arc, WrappedFstT, MutableFstT>::kFileVersion;
-
-template <typename Arc, typename WrappedFstT, typename MutableFstT>
-constexpr int EditFstImpl<Arc, WrappedFstT, MutableFstT>::kMinFileVersion;
-
-template <typename Arc, typename WrappedFstT, typename MutableFstT>
-inline void EditFstImpl<Arc, WrappedFstT, MutableFstT>::DeleteStates() {
-  data_->DeleteStates();
-  // we are deleting all states, so just forget about pointer to wrapped_
-  // and do what default constructor does: set wrapped_ to a new VectorFst
-  wrapped_.reset(new MutableFstT());
-  const auto new_props =
-      DeleteAllStatesProperties(FstImpl<Arc>::Properties(), kStaticProperties);
-  FstImpl<Arc>::SetProperties(new_props);
-}
-
-template <typename Arc, typename WrappedFstT, typename MutableFstT>
-EditFstImpl<Arc, WrappedFstT, MutableFstT> *
-EditFstImpl<Arc, WrappedFstT, MutableFstT>::Read(std::istream &strm,
-                                                 const FstReadOptions &opts) {
-  auto *impl = new EditFstImpl();
-  FstHeader hdr;
-  if (!impl->ReadHeader(strm, opts, kMinFileVersion, &hdr)) return nullptr;
-  impl->SetStart(hdr.Start());
-  // Reads in wrapped FST.
-  FstReadOptions wrapped_opts(opts);
-  // Contained header was written out, so reads it in too.
-  wrapped_opts.header = nullptr;
-  std::unique_ptr<Fst<Arc>> wrapped_fst(Fst<Arc>::Read(strm, wrapped_opts));
-  if (!wrapped_fst) return nullptr;
-  impl->wrapped_.reset(static_cast<WrappedFstT *>(wrapped_fst.release()));
-  impl->data_ = std::shared_ptr<EditFstData<Arc, WrappedFstT, MutableFstT>>(
-      EditFstData<Arc, WrappedFstT, MutableFstT>::Read(strm, opts));
-  if (!impl->data_) return nullptr;
-  return impl;
-}
-
-}  // namespace internal
-
-// Concrete, editable FST.  This class attaches interface to implementation.
-template <typename A, typename WrappedFstT = ExpandedFst<A>,
-          typename MutableFstT = VectorFst<A>>
-class EditFst : public ImplToMutableFst<
-                    internal::EditFstImpl<A, WrappedFstT, MutableFstT>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  using Impl = internal::EditFstImpl<Arc, WrappedFstT, MutableFstT>;
-
-  friend class MutableArcIterator<EditFst<Arc, WrappedFstT, MutableFstT>>;
-
-  EditFst() : ImplToMutableFst<Impl>(std::make_shared<Impl>()) {}
-
-  explicit EditFst(const Fst<Arc> &fst)
-      : ImplToMutableFst<Impl>(std::make_shared<Impl>(fst)) {}
-
-  explicit EditFst(const WrappedFstT &fst)
-      : ImplToMutableFst<Impl>(std::make_shared<Impl>(fst)) {}
-
-  // See Fst<>::Copy() for doc.
-  EditFst(const EditFst<Arc, WrappedFstT, MutableFstT> &fst, bool safe = false)
-      : ImplToMutableFst<Impl>(fst, safe) {}
-
-  ~EditFst() override {}
-
-  // Gets a copy of this EditFst. See Fst<>::Copy() for further doc.
-  EditFst<Arc, WrappedFstT, MutableFstT> *Copy(
-      bool safe = false) const override {
-    return new EditFst<Arc, WrappedFstT, MutableFstT>(*this, safe);
-  }
-
-  EditFst<Arc, WrappedFstT, MutableFstT> &operator=(
-      const EditFst<Arc, WrappedFstT, MutableFstT> &fst) {
-    SetImpl(fst.GetSharedImpl());
-    return *this;
-  }
-
-  EditFst<Arc, WrappedFstT, MutableFstT> &operator=(
-      const Fst<Arc> &fst) override {
-    SetImpl(std::make_shared<Impl>(fst));
-    return *this;
-  }
-
-  // Reads an EditFst from an input stream, returning nullptr on error.
-  static EditFst<Arc, WrappedFstT, MutableFstT> *Read(
-      std::istream &strm, const FstReadOptions &opts) {
-    auto *impl = Impl::Read(strm, opts);
-    return impl ? new EditFst<Arc>(std::shared_ptr<Impl>(impl)) : nullptr;
-  }
-
-  // Reads an EditFst from a file, returning nullptr on error. If the filename
-  // argument is an empty string, it reads from standard input.
-  static EditFst<Arc, WrappedFstT, MutableFstT> *Read(const string &filename) {
-    auto *impl = ImplToExpandedFst<Impl, MutableFst<Arc>>::Read(filename);
-    return impl ? new EditFst<Arc, WrappedFstT, MutableFstT>(
-                      std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return GetImpl()->Write(strm, opts);
-  }
-
-  bool Write(const string &filename) const override {
-    return Fst<Arc>::WriteFile(filename);
-  }
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const override {
-    GetImpl()->InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetImpl()->InitArcIterator(s, data);
-  }
-
-  void InitMutableArcIterator(StateId s,
-                              MutableArcIteratorData<A> *data) override {
-    GetMutableImpl()->InitMutableArcIterator(s, data);
-  }
-
- private:
-  explicit EditFst(std::shared_ptr<Impl> impl) : ImplToMutableFst<Impl>(impl) {}
-
-  using ImplToFst<Impl, MutableFst<Arc>>::GetImpl;
-  using ImplToFst<Impl, MutableFst<Arc>>::GetMutableImpl;
-  using ImplToFst<Impl, MutableFst<Arc>>::SetImpl;
-};
-
-}  // namespace fst
-
-#endif  // FST_EDIT_FST_H_
diff --git a/projects/llm_framework/include/fst/encode.h b/projects/llm_framework/include/fst/encode.h
deleted file mode 100644
index f251bbfc..00000000
--- a/projects/llm_framework/include/fst/encode.h
+++ /dev/null
@@ -1,556 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to encode and decode an FST.
-
-#ifndef FST_ENCODE_H_
-#define FST_ENCODE_H_
-
-#include <iostream>
-#include <memory>
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-#include <fstream>
-
-#include <fst/arc-map.h>
-#include <fst/rmfinalepsilon.h>
-
-
-namespace fst {
-
-enum EncodeType { ENCODE = 1, DECODE = 2 };
-
-static constexpr uint32 kEncodeLabels = 0x0001;
-static constexpr uint32 kEncodeWeights = 0x0002;
-static constexpr uint32 kEncodeFlags = 0x0003;
-
-namespace internal {
-
-static constexpr uint32 kEncodeHasISymbols = 0x0004;
-static constexpr uint32 kEncodeHasOSymbols = 0x0008;
-
-// Identifies stream data as an encode table (and its endianity)
-static const int32 kEncodeMagicNumber = 2129983209;
-
-// The following class encapsulates implementation details for the encoding and
-// decoding of label/weight tuples used for encoding and decoding of FSTs. The
-// EncodeTable is bidirectional. I.e, it stores both the Tuple of encode labels
-// and weights to a unique label, and the reverse.
-template <class Arc>
-class EncodeTable {
- public:
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-
-  // Encoded data consists of arc input/output labels and arc weight.
-  struct Tuple {
-    Tuple() {}
-
-    Tuple(Label ilabel_, Label olabel_, Weight weight_)
-        : ilabel(ilabel_), olabel(olabel_), weight(std::move(weight_)) {}
-
-    Tuple(const Tuple &tuple)
-        : ilabel(tuple.ilabel),
-          olabel(tuple.olabel),
-          weight(std::move(tuple.weight)) {}
-
-    Label ilabel;
-    Label olabel;
-    Weight weight;
-  };
-
-  // Comparison object for hashing EncodeTable Tuple(s).
-  class TupleEqual {
-   public:
-    bool operator()(const Tuple *x, const Tuple *y) const {
-      return (x->ilabel == y->ilabel && x->olabel == y->olabel &&
-              x->weight == y->weight);
-    }
-  };
-
-  // Hash function for EncodeTabe Tuples. Based on the encode flags
-  // we either hash the labels, weights or combination of them.
-  class TupleKey {
-   public:
-    TupleKey() : encode_flags_(kEncodeLabels | kEncodeWeights) {}
-
-    TupleKey(const TupleKey &key) : encode_flags_(key.encode_flags_) {}
-
-    explicit TupleKey(uint32 encode_flags) : encode_flags_(encode_flags) {}
-
-    size_t operator()(const Tuple *x) const {
-      size_t hash = x->ilabel;
-      static constexpr int lshift = 5;
-      static constexpr int rshift = CHAR_BIT * sizeof(size_t) - 5;
-      if (encode_flags_ & kEncodeLabels) {
-        hash = hash << lshift ^ hash >> rshift ^ x->olabel;
-      }
-      if (encode_flags_ & kEncodeWeights) {
-        hash = hash << lshift ^ hash >> rshift ^ x->weight.Hash();
-      }
-      return hash;
-    }
-
-   private:
-    int32 encode_flags_;
-  };
-
-  explicit EncodeTable(uint32 encode_flags)
-      : flags_(encode_flags), encode_hash_(1024, TupleKey(encode_flags)) {}
-
-  using EncodeHash = std::unordered_map<const Tuple *, Label, TupleKey,
-                                        TupleEqual>;
-
-  // Given an arc, encodes either input/output labels or input/costs or both.
-  Label Encode(const Arc &arc) {
-    std::unique_ptr<Tuple> tuple(
-        new Tuple(arc.ilabel, flags_ & kEncodeLabels ? arc.olabel : 0,
-                  flags_ & kEncodeWeights ? arc.weight : Weight::One()));
-    auto insert_result = encode_hash_.insert(
-        std::make_pair(tuple.get(), encode_tuples_.size() + 1));
-    if (insert_result.second) encode_tuples_.push_back(std::move(tuple));
-    return insert_result.first->second;
-  }
-
-  // Given an arc, looks up its encoded label or returns kNoLabel if not found.
-  Label GetLabel(const Arc &arc) const {
-    const Tuple tuple(arc.ilabel, flags_ & kEncodeLabels ? arc.olabel : 0,
-                      flags_ & kEncodeWeights ? arc.weight : Weight::One());
-    auto it = encode_hash_.find(&tuple);
-    return (it == encode_hash_.end()) ?  kNoLabel : it->second;
-  }
-
-  // Given an encoded arc label, decodes back to input/output labels and costs.
-  const Tuple *Decode(Label key) const {
-    if (key < 1 || key > encode_tuples_.size()) {
-      LOG(ERROR) << "EncodeTable::Decode: Unknown decode key: " << key;
-      return nullptr;
-    }
-    return encode_tuples_[key - 1].get();
-  }
-
-  size_t Size() const { return encode_tuples_.size(); }
-
-  bool Write(std::ostream &strm, const string &source) const;
-
-  static EncodeTable<Arc> *Read(std::istream &strm, const string &source);
-
-  uint32 Flags() const { return flags_ & kEncodeFlags; }
-
-  const SymbolTable *InputSymbols() const { return isymbols_.get(); }
-
-  const SymbolTable *OutputSymbols() const { return osymbols_.get(); }
-
-  void SetInputSymbols(const SymbolTable *syms) {
-    if (syms) {
-      isymbols_.reset(syms->Copy());
-      flags_ |= kEncodeHasISymbols;
-    } else {
-      isymbols_.reset();
-      flags_ &= ~kEncodeHasISymbols;
-    }
-  }
-
-  void SetOutputSymbols(const SymbolTable *syms) {
-    if (syms) {
-      osymbols_.reset(syms->Copy());
-      flags_ |= kEncodeHasOSymbols;
-    } else {
-      osymbols_.reset();
-      flags_ &= ~kEncodeHasOSymbols;
-    }
-  }
-
- private:
-  uint32 flags_;
-  std::vector<std::unique_ptr<Tuple>> encode_tuples_;
-  EncodeHash encode_hash_;
-  std::unique_ptr<SymbolTable> isymbols_;  // Pre-encoded input symbol table.
-  std::unique_ptr<SymbolTable> osymbols_;  // Pre-encoded output symbol table.
-
-  EncodeTable(const EncodeTable &) = delete;
-  EncodeTable &operator=(const EncodeTable &) = delete;
-};
-
-template <class Arc>
-bool EncodeTable<Arc>::Write(std::ostream &strm,
-                                  const string &source) const {
-  WriteType(strm, kEncodeMagicNumber);
-  WriteType(strm, flags_);
-  const int64 size = encode_tuples_.size();
-  WriteType(strm, size);
-  for (const auto &tuple : encode_tuples_) {
-    WriteType(strm, tuple->ilabel);
-    WriteType(strm, tuple->olabel);
-    tuple->weight.Write(strm);
-  }
-  if (flags_ & kEncodeHasISymbols) isymbols_->Write(strm);
-  if (flags_ & kEncodeHasOSymbols) osymbols_->Write(strm);
-  strm.flush();
-  if (!strm) {
-    LOG(ERROR) << "EncodeTable::Write: Write failed: " << source;
-    return false;
-  }
-  return true;
-}
-
-template <class Arc>
-EncodeTable<Arc> *EncodeTable<Arc>::Read(std::istream &strm,
-                                         const string &source) {
-  int32 magic_number = 0;
-  ReadType(strm, &magic_number);
-  if (magic_number != kEncodeMagicNumber) {
-    LOG(ERROR) << "EncodeTable::Read: Bad encode table header: " << source;
-    return nullptr;
-  }
-  uint32 flags;
-  ReadType(strm, &flags);
-  int64 size;
-  ReadType(strm, &size);
-  if (!strm) {
-    LOG(ERROR) << "EncodeTable::Read: Read failed: " << source;
-    return nullptr;
-  }
-  std::unique_ptr<EncodeTable<Arc>> table(new EncodeTable<Arc>(flags));
-  for (int64 i = 0; i < size; ++i) {
-    std::unique_ptr<Tuple> tuple(new Tuple());
-    ReadType(strm, &tuple->ilabel);
-    ReadType(strm, &tuple->olabel);
-    tuple->weight.Read(strm);
-    if (!strm) {
-      LOG(ERROR) << "EncodeTable::Read: Read failed: " << source;
-      return nullptr;
-    }
-    table->encode_tuples_.push_back(std::move(tuple));
-    table->encode_hash_[table->encode_tuples_.back().get()] =
-        table->encode_tuples_.size();
-  }
-  if (flags & kEncodeHasISymbols) {
-    table->isymbols_.reset(SymbolTable::Read(strm, source));
-  }
-  if (flags & kEncodeHasOSymbols) {
-    table->osymbols_.reset(SymbolTable::Read(strm, source));
-  }
-  return table.release();
-}
-
-}  // namespace internal
-
-// A mapper to encode/decode weighted transducers. Encoding of an FST is used
-// for performing classical determinization or minimization on a weighted
-// transducer viewing it as an unweighted acceptor over encoded labels.
-//
-// The mapper stores the encoding in a local hash table (EncodeTable). This
-// table is shared (and reference-counted) between the encoder and decoder.
-// A decoder has read-only access to the EncodeTable.
-//
-// The EncodeMapper allows on the fly encoding of the machine. As the
-// EncodeTable is generated the same table may by used to decode the machine
-// on the fly. For example in the following sequence of operations
-//
-//  Encode -> Determinize -> Decode
-//
-// we will use the encoding table generated during the encode step in the
-// decode, even though the encoding is not complete.
-template <class Arc>
-class EncodeMapper {
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-
- public:
-  EncodeMapper(uint32 flags, EncodeType type)
-      : flags_(flags),
-        type_(type),
-        table_(std::make_shared<internal::EncodeTable<Arc>>(flags)),
-        error_(false) {}
-
-  EncodeMapper(const EncodeMapper &mapper)
-      : flags_(mapper.flags_),
-        type_(mapper.type_),
-        table_(mapper.table_),
-        error_(false) {}
-
-  // Copy constructor but setting the type, typically to DECODE.
-  EncodeMapper(const EncodeMapper &mapper, EncodeType type)
-      : flags_(mapper.flags_),
-        type_(type),
-        table_(mapper.table_),
-        error_(mapper.error_) {}
-
-  Arc operator()(const Arc &arc);
-
-  MapFinalAction FinalAction() const {
-    return (type_ == ENCODE && (flags_ & kEncodeWeights))
-               ? MAP_REQUIRE_SUPERFINAL
-               : MAP_NO_SUPERFINAL;
-  }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 inprops) {
-    uint64 outprops = inprops;
-    if (error_) outprops |= kError;
-    uint64 mask = kFstProperties;
-    if (flags_ & kEncodeLabels) {
-      mask &= kILabelInvariantProperties & kOLabelInvariantProperties;
-    }
-    if (flags_ & kEncodeWeights) {
-      mask &= kILabelInvariantProperties & kWeightInvariantProperties &
-              (type_ == ENCODE ? kAddSuperFinalProperties
-                               : kRmSuperFinalProperties);
-    }
-    return outprops & mask;
-  }
-
-  uint32 Flags() const { return flags_; }
-
-  EncodeType Type() const { return type_; }
-
-  bool Write(std::ostream &strm, const string &source) const {
-    return table_->Write(strm, source);
-  }
-
-  bool Write(const string &filename) const {
-    std::ofstream strm(filename,
-                             std::ios_base::out | std::ios_base::binary);
-    if (!strm) {
-      LOG(ERROR) << "EncodeMap: Can't open file: " << filename;
-      return false;
-    }
-    return Write(strm, filename);
-  }
-
-  static EncodeMapper<Arc> *Read(std::istream &strm, const string &source,
-                               EncodeType type = ENCODE) {
-    auto *table = internal::EncodeTable<Arc>::Read(strm, source);
-    return table ? new EncodeMapper(table->Flags(), type, table) : nullptr;
-  }
-
-  static EncodeMapper<Arc> *Read(const string &filename,
-                                 EncodeType type = ENCODE) {
-    std::ifstream strm(filename,
-                            std::ios_base::in | std::ios_base::binary);
-    if (!strm) {
-      LOG(ERROR) << "EncodeMap: Can't open file: " << filename;
-      return nullptr;
-    }
-    return Read(strm, filename, type);
-  }
-
-  const SymbolTable *InputSymbols() const { return table_->InputSymbols(); }
-
-  const SymbolTable *OutputSymbols() const { return table_->OutputSymbols(); }
-
-  void SetInputSymbols(const SymbolTable *syms) {
-    table_->SetInputSymbols(syms);
-  }
-
-  void SetOutputSymbols(const SymbolTable *syms) {
-    table_->SetOutputSymbols(syms);
-  }
-
- private:
-  uint32 flags_;
-  EncodeType type_;
-  std::shared_ptr<internal::EncodeTable<Arc>> table_;
-  bool error_;
-
-  explicit EncodeMapper(uint32 flags, EncodeType type,
-                        internal::EncodeTable<Arc> *table)
-      : flags_(flags), type_(type), table_(table), error_(false) {}
-
-  EncodeMapper &operator=(const EncodeMapper &) = delete;
-};
-
-template <class Arc>
-Arc EncodeMapper<Arc>::operator()(const Arc &arc) {
-  if (type_ == ENCODE) {
-    if ((arc.nextstate == kNoStateId && !(flags_ & kEncodeWeights)) ||
-        (arc.nextstate == kNoStateId && (flags_ & kEncodeWeights) &&
-         arc.weight == Weight::Zero())) {
-      return arc;
-    } else {
-      const auto label = table_->Encode(arc);
-      return Arc(label, flags_ & kEncodeLabels ? label : arc.olabel,
-                 flags_ & kEncodeWeights ? Weight::One() : arc.weight,
-                 arc.nextstate);
-    }
-  } else {  // type_ == DECODE
-    if (arc.nextstate == kNoStateId) {
-      return arc;
-    } else {
-      if (arc.ilabel == 0) return arc;
-      if (flags_ & kEncodeLabels && arc.ilabel != arc.olabel) {
-        FSTERROR() << "EncodeMapper: Label-encoded arc has different "
-                      "input and output labels";
-        error_ = true;
-      }
-      if (flags_ & kEncodeWeights && arc.weight != Weight::One()) {
-        FSTERROR() << "EncodeMapper: Weight-encoded arc has non-trivial weight";
-        error_ = true;
-      }
-      const auto tuple = table_->Decode(arc.ilabel);
-      if (!tuple) {
-        FSTERROR() << "EncodeMapper: Decode failed";
-        error_ = true;
-        return Arc(kNoLabel, kNoLabel, Weight::NoWeight(), arc.nextstate);
-      } else {
-        return Arc(tuple->ilabel,
-                   flags_ & kEncodeLabels ? tuple->olabel : arc.olabel,
-                   flags_ & kEncodeWeights ? tuple->weight : arc.weight,
-                   arc.nextstate);
-      }
-    }
-  }
-}
-
-// Complexity: O(E + V).
-template <class Arc>
-inline void Encode(MutableFst<Arc> *fst, EncodeMapper<Arc> *mapper) {
-  mapper->SetInputSymbols(fst->InputSymbols());
-  mapper->SetOutputSymbols(fst->OutputSymbols());
-  ArcMap(fst, mapper);
-}
-
-template <class Arc>
-inline void Decode(MutableFst<Arc> *fst, const EncodeMapper<Arc> &mapper) {
-  ArcMap(fst, EncodeMapper<Arc>(mapper, DECODE));
-  RmFinalEpsilon(fst);
-  fst->SetInputSymbols(mapper.InputSymbols());
-  fst->SetOutputSymbols(mapper.OutputSymbols());
-}
-
-// On-the-fly encoding of an input FST.
-//
-// Complexity:
-//
-//   Construction: O(1)
-//   Traversal: O(e + v)
-//
-// where e is the number of arcs visited and v is the number of states visited.
-// Constant time and space to visit an input state or arc is assumed and
-// exclusive of caching.
-template <class Arc>
-class EncodeFst : public ArcMapFst<Arc, Arc, EncodeMapper<Arc>> {
- public:
-  using Mapper = EncodeMapper<Arc>;
-  using Impl = internal::ArcMapFstImpl<Arc, Arc, Mapper>;
-
-  EncodeFst(const Fst<Arc> &fst, Mapper *encoder)
-      : ArcMapFst<Arc, Arc, Mapper>(fst, encoder, ArcMapFstOptions()) {
-    encoder->SetInputSymbols(fst.InputSymbols());
-    encoder->SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  EncodeFst(const Fst<Arc> &fst, const Mapper &encoder)
-      : ArcMapFst<Arc, Arc, Mapper>(fst, encoder, ArcMapFstOptions()) {}
-
-  // See Fst<>::Copy() for doc.
-  EncodeFst(const EncodeFst<Arc> &fst, bool copy = false)
-      : ArcMapFst<Arc, Arc, Mapper>(fst, copy) {}
-
-  // Makes a copy of this EncodeFst. See Fst<>::Copy() for further doc.
-  EncodeFst<Arc> *Copy(bool safe = false) const override {
-    if (safe) {
-      FSTERROR() << "EncodeFst::Copy(true): Not allowed";
-      GetImpl()->SetProperties(kError, kError);
-    }
-    return new EncodeFst(*this);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-};
-
-// On-the-fly decoding of an input FST.
-//
-// Complexity:
-//
-//   Construction: O(1).
-//   Traversal: O(e + v)
-//
-// Constant time and space to visit an input state or arc is assumed and
-// exclusive of caching.
-template <class Arc>
-class DecodeFst : public ArcMapFst<Arc, Arc, EncodeMapper<Arc>> {
- public:
-  using Mapper = EncodeMapper<Arc>;
-  using Impl = internal::ArcMapFstImpl<Arc, Arc, Mapper>;
-  using ImplToFst<Impl>::GetImpl;
-
-  DecodeFst(const Fst<Arc> &fst, const Mapper &encoder)
-      : ArcMapFst<Arc, Arc, Mapper>(fst, Mapper(encoder, DECODE),
-                                    ArcMapFstOptions()) {
-    GetMutableImpl()->SetInputSymbols(encoder.InputSymbols());
-    GetMutableImpl()->SetOutputSymbols(encoder.OutputSymbols());
-  }
-
-  // See Fst<>::Copy() for doc.
-  DecodeFst(const DecodeFst<Arc> &fst, bool safe = false)
-      : ArcMapFst<Arc, Arc, Mapper>(fst, safe) {}
-
-  // Makes a copy of this DecodeFst. See Fst<>::Copy() for further doc.
-  DecodeFst<Arc> *Copy(bool safe = false) const override {
-    return new DecodeFst(*this, safe);
-  }
-
- private:
-  using ImplToFst<Impl>::GetMutableImpl;
-};
-
-// Specialization for EncodeFst.
-template <class Arc>
-class StateIterator<EncodeFst<Arc>>
-    : public StateIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>> {
- public:
-  explicit StateIterator(const EncodeFst<Arc> &fst)
-      : StateIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>>(fst) {}
-};
-
-// Specialization for EncodeFst.
-template <class Arc>
-class ArcIterator<EncodeFst<Arc>>
-    : public ArcIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>> {
- public:
-  ArcIterator(const EncodeFst<Arc> &fst, typename Arc::StateId s)
-      : ArcIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>>(fst, s) {}
-};
-
-// Specialization for DecodeFst.
-template <class Arc>
-class StateIterator<DecodeFst<Arc>>
-    : public StateIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>> {
- public:
-  explicit StateIterator(const DecodeFst<Arc> &fst)
-      : StateIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>>(fst) {}
-};
-
-// Specialization for DecodeFst.
-template <class Arc>
-class ArcIterator<DecodeFst<Arc>>
-    : public ArcIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>> {
- public:
-  ArcIterator(const DecodeFst<Arc> &fst, typename Arc::StateId s)
-      : ArcIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>>(fst, s) {}
-};
-
-// Useful aliases when using StdArc.
-
-using StdEncodeFst = EncodeFst<StdArc>;
-
-using StdDecodeFst = DecodeFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_ENCODE_H_
diff --git a/projects/llm_framework/include/fst/epsnormalize.h b/projects/llm_framework/include/fst/epsnormalize.h
deleted file mode 100644
index 18105fb1..00000000
--- a/projects/llm_framework/include/fst/epsnormalize.h
+++ /dev/null
@@ -1,61 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function that implements epsilon-normalization.
-
-#ifndef FST_EPSNORMALIZE_H_
-#define FST_EPSNORMALIZE_H_
-
-
-#include <fst/arc-map.h>
-#include <fst/factor-weight.h>
-#include <fst/invert.h>
-#include <fst/rmepsilon.h>
-
-
-namespace fst {
-
-enum EpsNormalizeType { EPS_NORM_INPUT, EPS_NORM_OUTPUT };
-
-// Returns an equivalent FST that is epsilon-normalized. An acceptor is
-// epsilon-normalized if it is epsilon-removed. A transducer is input
-// epsilon-normalized if additionally if on each path any epsilon input
-// label follows all non-epsilon input labels. Output epsilon-normalized
-// is defined similarly.
-//
-// For more information, see:
-//
-// Mohri, M. 2002. Generic epsilon-removal and input epsilon-normalization
-// algorithms for weighted transducers. International Journal of Computer
-// Science, 13(1): 129-143, 2002.
-template <class Arc>
-void EpsNormalize(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-                  EpsNormalizeType type = EPS_NORM_INPUT) {
-  EpsNormalize<Arc, GALLIC>(ifst, ofst, type);
-}
-
-// Same as above, except allows specifying explicitly the gallic weight type.
-template <class Arc, GallicType G>
-void EpsNormalize(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-                  EpsNormalizeType type) {
-  VectorFst<GallicArc<Arc, G>> gfst;
-  std::unique_ptr<SymbolTable> symbols;
-  if (type == EPS_NORM_INPUT) {
-    ArcMap(ifst, &gfst, ToGallicMapper<Arc, G>());
-    if (ifst.OutputSymbols()) symbols.reset(ifst.OutputSymbols()->Copy());
-  } else {  // type == EPS_NORM_OUTPUT
-    ArcMap(InvertFst<Arc>(ifst), &gfst, ToGallicMapper<Arc, G>());
-    if (ifst.InputSymbols()) symbols.reset(ifst.InputSymbols()->Copy());
-  }
-  RmEpsilon(&gfst);
-  FactorWeightFst<GallicArc<Arc, G>,
-                  GallicFactor<typename Arc::Label, typename Arc::Weight, G>>
-      fwfst(gfst);
-  ArcMap(fwfst, ofst, FromGallicMapper<Arc, G>());
-  ofst->SetOutputSymbols(symbols.get());
-  if (type == EPS_NORM_OUTPUT) Invert(ofst);
-}
-
-}  // namespace fst
-
-#endif  // FST_EPSNORMALIZE_H_
diff --git a/projects/llm_framework/include/fst/equal.h b/projects/llm_framework/include/fst/equal.h
deleted file mode 100644
index ed89c6ce..00000000
--- a/projects/llm_framework/include/fst/equal.h
+++ /dev/null
@@ -1,169 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function to test equality of two FSTs.
-
-#ifndef FST_EQUAL_H_
-#define FST_EQUAL_H_
-
-#include <fst/log.h>
-
-#include <fst/fst.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-constexpr uint32 kEqualFsts = 0x0001;
-constexpr uint32 kEqualFstTypes = 0x0002;
-constexpr uint32 kEqualCompatProperties = 0x0004;
-constexpr uint32 kEqualCompatSymbols = 0x0008;
-constexpr uint32 kEqualAll =
-    kEqualFsts | kEqualFstTypes | kEqualCompatProperties | kEqualCompatSymbols;
-
-class WeightApproxEqual {
- public:
-  explicit WeightApproxEqual(float delta) : delta_(delta) {}
-
-  template <class Weight>
-  bool operator()(const Weight &w1, const Weight &w2) const {
-    return ApproxEqual(w1, w2, delta_);
-  }
-
- private:
-  float delta_;
-};
-
-// Tests if two Fsts have the same states and arcs in the same order (when
-// etype & kEqualFst).
-// Also optional checks equality of Fst types (etype & kEqualFstTypes) and
-// compatibility of stored properties (etype & kEqualCompatProperties) and
-// of symbol tables (etype & kEqualCompatSymbols).
-template <class Arc, class WeightEqual>
-bool Equal(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-           WeightEqual weight_equal, uint32 etype = kEqualFsts) {
-  if ((etype & kEqualFstTypes) && (fst1.Type() != fst2.Type())) {
-    VLOG(1) << "Equal: Mismatched FST types (" << fst1.Type() << " != "
-            << fst2.Type() << ")";
-    return false;
-  }
-  if ((etype & kEqualCompatProperties) &&
-      !CompatProperties(fst1.Properties(kCopyProperties, false),
-                        fst2.Properties(kCopyProperties, false))) {
-    VLOG(1) << "Equal: Properties not compatible";
-    return false;
-  }
-  if (etype & kEqualCompatSymbols) {
-    if (!CompatSymbols(fst1.InputSymbols(), fst2.InputSymbols(), false)) {
-      VLOG(1) << "Equal: Input symbols not compatible";
-      return false;
-    }
-    if (!CompatSymbols(fst1.OutputSymbols(), fst2.OutputSymbols(), false)) {
-      VLOG(1) << "Equal: Output symbols not compatible";
-      return false;
-    }
-  }
-  if (!(etype & kEqualFsts)) return true;
-  if (fst1.Start() != fst2.Start()) {
-    VLOG(1) << "Equal: Mismatched start states (" << fst1.Start() << " != "
-            << fst2.Start() << ")";
-    return false;
-  }
-  StateIterator<Fst<Arc>> siter1(fst1);
-  StateIterator<Fst<Arc>> siter2(fst2);
-  while (!siter1.Done() || !siter2.Done()) {
-    if (siter1.Done() || siter2.Done()) {
-      VLOG(1) << "Equal: Mismatched number of states";
-      return false;
-    }
-    const auto s1 = siter1.Value();
-    const auto s2 = siter2.Value();
-    if (s1 != s2) {
-      VLOG(1) << "Equal: Mismatched states (" << s1 << "!= "
-              << s2 << ")";
-      return false;
-    }
-    const auto &final1 = fst1.Final(s1);
-    const auto &final2 = fst2.Final(s2);
-    if (!weight_equal(final1, final2)) {
-      VLOG(1) << "Equal: Mismatched final weights at state " << s1
-              << " (" << final1 << " != " << final2 << ")";
-      return false;
-    }
-    ArcIterator<Fst<Arc>> aiter1(fst1, s1);
-    ArcIterator<Fst<Arc>> aiter2(fst2, s2);
-    for (auto a = 0; !aiter1.Done() || !aiter2.Done(); ++a) {
-      if (aiter1.Done() || aiter2.Done()) {
-        VLOG(1) << "Equal: Mismatched number of arcs at state " << s1;
-        return false;
-      }
-      const auto &arc1 = aiter1.Value();
-      const auto &arc2 = aiter2.Value();
-      if (arc1.ilabel != arc2.ilabel) {
-        VLOG(1) << "Equal: Mismatched arc input labels at state " << s1
-                << ", arc " << a << " (" << arc1.ilabel << " != "
-                << arc2.ilabel << ")";
-        return false;
-      } else if (arc1.olabel != arc2.olabel) {
-        VLOG(1) << "Equal: Mismatched arc output labels at state " << s1
-                << ", arc " << a << " (" << arc1.olabel << " != "
-                << arc2.olabel << ")";
-        return false;
-      } else if (!weight_equal(arc1.weight, arc2.weight)) {
-        VLOG(1) << "Equal: Mismatched arc weights at state " << s1
-                << ", arc " << a << " (" << arc1.weight << " != "
-                << arc2.weight << ")";
-        return false;
-      } else if (arc1.nextstate != arc2.nextstate) {
-        VLOG(1) << "Equal: Mismatched next state at state " << s1
-                << ", arc " << a << " (" << arc1.nextstate << " != "
-                << arc2.nextstate << ")";
-        return false;
-      }
-      aiter1.Next();
-      aiter2.Next();
-    }
-    // Sanity checks: should never fail.
-    if (fst1.NumArcs(s1) != fst2.NumArcs(s2)) {
-      FSTERROR() << "Equal: Inconsistent arc counts at state " << s1
-                 << " (" << fst1.NumArcs(s1) << " != "
-                 << fst2.NumArcs(s2) << ")";
-      return false;
-    }
-    if (fst1.NumInputEpsilons(s1) != fst2.NumInputEpsilons(s2)) {
-      FSTERROR() << "Equal: Inconsistent input epsilon counts at state " << s1
-                 << " (" << fst1.NumInputEpsilons(s1) << " != "
-                 << fst2.NumInputEpsilons(s2) << ")";
-      return false;
-    }
-    if (fst1.NumOutputEpsilons(s1) != fst2.NumOutputEpsilons(s2)) {
-      FSTERROR() << "Equal: Inconsistent output epsilon counts at state " << s1
-                 << " (" << fst1.NumOutputEpsilons(s1) << " != "
-                 << fst2.NumOutputEpsilons(s2) << ")";
-    }
-    siter1.Next();
-    siter2.Next();
-  }
-  return true;
-}
-
-template <class Arc>
-bool Equal(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-           float delta = kDelta, uint32 etype = kEqualFsts) {
-  return Equal(fst1, fst2, WeightApproxEqual(delta), etype);
-}
-
-// Support double deltas without forcing all clients to cast to float.
-// Without this overload, Equal<Arc, WeightEqual=double> will be chosen,
-// since it is a better match than double -> float narrowing, but
-// the instantiation will fail.
-template <class Arc>
-bool Equal(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-           double delta, uint32 etype = kEqualFsts) {
-  return Equal(fst1, fst2, WeightApproxEqual(static_cast<float>(delta)), etype);
-}
-
-
-}  // namespace fst
-
-#endif  // FST_EQUAL_H_
diff --git a/projects/llm_framework/include/fst/equivalent.h b/projects/llm_framework/include/fst/equivalent.h
deleted file mode 100644
index cf3fdb61..00000000
--- a/projects/llm_framework/include/fst/equivalent.h
+++ /dev/null
@@ -1,230 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to determine the equivalence of two FSTs.
-
-#ifndef FST_EQUIVALENT_H_
-#define FST_EQUIVALENT_H_
-
-#include <algorithm>
-#include <deque>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-#include <fst/log.h>
-
-#include <fst/encode.h>
-#include <fst/push.h>
-#include <fst/union-find.h>
-#include <fst/vector-fst.h>
-
-
-namespace fst {
-namespace internal {
-
-// Traits-like struct holding utility functions/typedefs/constants for
-// the equivalence algorithm.
-//
-// Encoding device: in order to make the statesets of the two acceptors
-// disjoint, we map Arc::StateId on the type MappedId. The states of
-// the first acceptor are mapped on odd numbers (s -> 2s + 1), and
-// those of the second one on even numbers (s -> 2s + 2). The number 0
-// is reserved for an implicit (non-final) dead state (required for
-// the correct treatment of non-coaccessible states; kNoStateId is mapped to
-// kDeadState for both acceptors). The union-find algorithm operates on the
-// mapped IDs.
-template <class Arc>
-struct EquivalenceUtil {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using MappedId = StateId;  // ID for an equivalence class.
-
-  // MappedId for an implicit dead state.
-  static constexpr MappedId kDeadState = 0;
-
-  // MappedId for lookup failure.
-  static constexpr MappedId kInvalidId = -1;
-
-  // Maps state ID to the representative of the corresponding
-  // equivalence class. The parameter 'which_fst' takes the values 1
-  // and 2, identifying the input FST.
-  static MappedId MapState(StateId s, int32 which_fst) {
-    return (kNoStateId == s) ? kDeadState
-                             : (static_cast<MappedId>(s) << 1) + which_fst;
-  }
-
-  // Maps set ID to State ID.
-  static StateId UnMapState(MappedId id) {
-    return static_cast<StateId>((--id) >> 1);
-  }
-
-  // Convenience function: checks if state with MappedId s is final in
-  // acceptor fa.
-  static bool IsFinal(const Fst<Arc> &fa, MappedId s) {
-    return (kDeadState == s) ? false
-                             : (fa.Final(UnMapState(s)) != Weight::Zero());
-  }
-  // Convenience function: returns the representative of ID in sets,
-  // creating a new set if needed.
-  static MappedId FindSet(UnionFind<MappedId> *sets, MappedId id) {
-    const auto repr = sets->FindSet(id);
-    if (repr != kInvalidId) {
-      return repr;
-    } else {
-      sets->MakeSet(id);
-      return id;
-    }
-  }
-};
-
-template <class Arc>
-constexpr
-    typename EquivalenceUtil<Arc>::MappedId EquivalenceUtil<Arc>::kDeadState;
-
-template <class Arc>
-constexpr
-    typename EquivalenceUtil<Arc>::MappedId EquivalenceUtil<Arc>::kInvalidId;
-
-}  // namespace internal
-
-// Equivalence checking algorithm: determines if the two FSTs fst1 and fst2
-// are equivalent. The input FSTs must be deterministic input-side epsilon-free
-// acceptors, unweighted or with weights over a left semiring. Two acceptors are
-// considered equivalent if they accept exactly the same set of strings (with
-// the same weights).
-//
-// The algorithm (cf. Aho, Hopcroft and Ullman, "The Design and Analysis of
-// Computer Programs") successively constructs sets of states that can be
-// reached by the same prefixes, starting with a set containing the start states
-// of both acceptors. A disjoint tree forest (the union-find algorithm) is used
-// to represent the sets of states. The algorithm returns false if one of the
-// constructed sets contains both final and non-final states. Returns an
-// optional error value (useful when FLAGS_error_fatal = false).
-//
-// Complexity:
-//
-// Quasi-linear, i.e., O(n G(n)), where
-//
-//   n = |S1| + |S2| is the number of states in both acceptors
-//
-//   G(n) is a very slowly growing function that can be approximated
-//        by 4 by all practical purposes.
-template <class Arc>
-bool Equivalent(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-                float delta = kDelta, bool *error = nullptr) {
-  using Weight = typename Arc::Weight;
-  if (error) *error = false;
-  // Check that the symbol table are compatible.
-  if (!CompatSymbols(fst1.InputSymbols(), fst2.InputSymbols()) ||
-      !CompatSymbols(fst1.OutputSymbols(), fst2.OutputSymbols())) {
-    FSTERROR() << "Equivalent: Input/output symbol tables of 1st argument "
-               << "do not match input/output symbol tables of 2nd argument";
-    if (error) *error = true;
-    return false;
-  }
-  // Check properties first.
-  static constexpr auto props = kNoEpsilons | kIDeterministic | kAcceptor;
-  if (fst1.Properties(props, true) != props) {
-    FSTERROR() << "Equivalent: 1st argument not an"
-               << " epsilon-free deterministic acceptor";
-    if (error) *error = true;
-    return false;
-  }
-  if (fst2.Properties(props, true) != props) {
-    FSTERROR() << "Equivalent: 2nd argument not an"
-               << " epsilon-free deterministic acceptor";
-    if (error) *error = true;
-    return false;
-  }
-  if ((fst1.Properties(kUnweighted, true) != kUnweighted) ||
-      (fst2.Properties(kUnweighted, true) != kUnweighted)) {
-    VectorFst<Arc> efst1(fst1);
-    VectorFst<Arc> efst2(fst2);
-    Push(&efst1, REWEIGHT_TO_INITIAL, delta);
-    Push(&efst2, REWEIGHT_TO_INITIAL, delta);
-    ArcMap(&efst1, QuantizeMapper<Arc>(delta));
-    ArcMap(&efst2, QuantizeMapper<Arc>(delta));
-    EncodeMapper<Arc> mapper(kEncodeWeights | kEncodeLabels, ENCODE);
-    ArcMap(&efst1, &mapper);
-    ArcMap(&efst2, &mapper);
-    return Equivalent(efst1, efst2);
-  }
-  using Util = internal::EquivalenceUtil<Arc>;
-  using MappedId = typename Util::MappedId;
-  enum { FST1 = 1, FST2 = 2 };  // Required by Util::MapState(...)
-  auto s1 = Util::MapState(fst1.Start(), FST1);
-  auto s2 = Util::MapState(fst2.Start(), FST2);
-  // The union-find structure.
-  UnionFind<MappedId> eq_classes(1000, Util::kInvalidId);
-  // Initializes the union-find structure.
-  eq_classes.MakeSet(s1);
-  eq_classes.MakeSet(s2);
-  // Data structure for the (partial) acceptor transition function of fst1 and
-  // fst2: input labels mapped to pairs of MappedIds representing destination
-  // states of the corresponding arcs in fst1 and fst2, respectively.
-  using Label2StatePairMap =
-      std::unordered_map<typename Arc::Label, std::pair<MappedId, MappedId>>;
-  Label2StatePairMap arc_pairs;
-  // Pairs of MappedId's to be processed, organized in a queue.
-  std::deque<std::pair<MappedId, MappedId>> q;
-  bool ret = true;
-  // Returns early if the start states differ w.r.t. finality.
-  if (Util::IsFinal(fst1, s1) != Util::IsFinal(fst2, s2)) ret = false;
-  // Main loop: explores the two acceptors in a breadth-first manner, updating
-  // the equivalence relation on the statesets. Loop invariant: each block of
-  // the states contains either final states only or non-final states only.
-  for (q.push_back(std::make_pair(s1, s2)); ret && !q.empty(); q.pop_front()) {
-    s1 = q.front().first;
-    s2 = q.front().second;
-    // Representatives of the equivalence classes of s1/s2.
-    const auto rep1 = Util::FindSet(&eq_classes, s1);
-    const auto rep2 = Util::FindSet(&eq_classes, s2);
-    if (rep1 != rep2) {
-      eq_classes.Union(rep1, rep2);
-      arc_pairs.clear();
-      // Copies outgoing arcs starting at s1 into the hash-table.
-      if (Util::kDeadState != s1) {
-        ArcIterator<Fst<Arc>> arc_iter(fst1, Util::UnMapState(s1));
-        for (; !arc_iter.Done(); arc_iter.Next()) {
-          const auto &arc = arc_iter.Value();
-          // Zero-weight arcs are treated as if they did not exist.
-          if (arc.weight != Weight::Zero()) {
-            arc_pairs[arc.ilabel].first = Util::MapState(arc.nextstate, FST1);
-          }
-        }
-      }
-      // Copies outgoing arcs starting at s2 into the hashtable.
-      if (Util::kDeadState != s2) {
-        ArcIterator<Fst<Arc>> arc_iter(fst2, Util::UnMapState(s2));
-        for (; !arc_iter.Done(); arc_iter.Next()) {
-          const auto &arc = arc_iter.Value();
-          // Zero-weight arcs are treated as if they did not exist.
-          if (arc.weight != Weight::Zero()) {
-            arc_pairs[arc.ilabel].second = Util::MapState(arc.nextstate, FST2);
-          }
-        }
-      }
-      // Iterates through the hashtable and process pairs of target states.
-      for (const auto &arc_iter : arc_pairs) {
-        const auto &pair = arc_iter.second;
-        if (Util::IsFinal(fst1, pair.first) !=
-            Util::IsFinal(fst2, pair.second)) {
-          // Detected inconsistency: return false.
-          ret = false;
-          break;
-        }
-        q.push_back(pair);
-      }
-    }
-  }
-  if (fst1.Properties(kError, false) || fst2.Properties(kError, false)) {
-    if (error) *error = true;
-    return false;
-  }
-  return ret;
-}
-
-}  // namespace fst
-
-#endif  // FST_EQUIVALENT_H_
diff --git a/projects/llm_framework/include/fst/expanded-fst.h b/projects/llm_framework/include/fst/expanded-fst.h
deleted file mode 100644
index 2c7d514c..00000000
--- a/projects/llm_framework/include/fst/expanded-fst.h
+++ /dev/null
@@ -1,179 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Generic FST augmented with state count-interface class definition.
-
-#ifndef FST_EXPANDED_FST_H_
-#define FST_EXPANDED_FST_H_
-
-#include <sys/types.h>
-#include <istream>
-#include <memory>
-#include <string>
-
-#include <fst/log.h>
-#include <fstream>
-
-#include <fst/fst.h>
-
-
-namespace fst {
-
-// A generic FST plus state count.
-template <class A>
-class ExpandedFst : public Fst<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  virtual StateId NumStates() const = 0;  // State count
-
-  // Get a copy of this ExpandedFst. See Fst<>::Copy() for further doc.
-  ExpandedFst<Arc> *Copy(bool safe = false) const override = 0;
-
-  // Read an ExpandedFst from an input stream; return NULL on error.
-  static ExpandedFst<Arc> *Read(std::istream &strm,
-                                const FstReadOptions &opts) {
-    FstReadOptions ropts(opts);
-    FstHeader hdr;
-    if (ropts.header) {
-      hdr = *opts.header;
-    } else {
-      if (!hdr.Read(strm, opts.source)) return nullptr;
-      ropts.header = &hdr;
-    }
-    if (!(hdr.Properties() & kExpanded)) {
-      LOG(ERROR) << "ExpandedFst::Read: Not an ExpandedFst: " << ropts.source;
-      return nullptr;
-    }
-    const auto reader =
-        FstRegister<Arc>::GetRegister()->GetReader(hdr.FstType());
-    if (!reader) {
-      LOG(ERROR) << "ExpandedFst::Read: Unknown FST type \"" << hdr.FstType()
-                 << "\" (arc type = \"" << A::Type() << "\"): " << ropts.source;
-      return nullptr;
-    }
-    auto *fst = reader(strm, ropts);
-    if (!fst) return nullptr;
-    return static_cast<ExpandedFst<Arc> *>(fst);
-  }
-
-  // Read an ExpandedFst from a file; return NULL on error.
-  // Empty filename reads from standard input.
-  static ExpandedFst<Arc> *Read(const string &filename) {
-    if (!filename.empty()) {
-      std::ifstream strm(filename,
-                              std::ios_base::in | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "ExpandedFst::Read: Can't open file: " << filename;
-        return nullptr;
-      }
-      return Read(strm, FstReadOptions(filename));
-    } else {
-      return Read(std::cin, FstReadOptions("standard input"));
-    }
-  }
-};
-
-namespace internal {
-
-//  ExpandedFst<A> case - abstract methods.
-template <class Arc>
-inline typename Arc::Weight Final(const ExpandedFst<Arc> &fst,
-                                  typename Arc::StateId s) {
-  return fst.Final(s);
-}
-
-template <class Arc>
-inline ssize_t NumArcs(const ExpandedFst<Arc> &fst, typename Arc::StateId s) {
-  return fst.NumArcs(s);
-}
-
-template <class Arc>
-inline ssize_t NumInputEpsilons(const ExpandedFst<Arc> &fst,
-                                typename Arc::StateId s) {
-  return fst.NumInputEpsilons(s);
-}
-
-template <class Arc>
-inline ssize_t NumOutputEpsilons(const ExpandedFst<Arc> &fst,
-                                 typename Arc::StateId s) {
-  return fst.NumOutputEpsilons(s);
-}
-
-}  // namespace internal
-
-// A useful alias when using StdArc.
-using StdExpandedFst = ExpandedFst<StdArc>;
-
-// This is a helper class template useful for attaching an ExpandedFst
-// interface to its implementation, handling reference counting. It
-// delegates to ImplToFst the handling of the Fst interface methods.
-template <class Impl, class FST = ExpandedFst<typename Impl::Arc>>
-class ImplToExpandedFst : public ImplToFst<Impl, FST> {
- public:
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  StateId NumStates() const override { return GetImpl()->NumStates(); }
-
- protected:
-  using ImplToFst<Impl, FST>::GetImpl;
-
-  explicit ImplToExpandedFst(std::shared_ptr<Impl> impl)
-      : ImplToFst<Impl, FST>(impl) {}
-
-  ImplToExpandedFst(const ImplToExpandedFst<Impl, FST> &fst, bool safe)
-      : ImplToFst<Impl, FST>(fst, safe) {}
-
-  static Impl *Read(std::istream &strm, const FstReadOptions &opts) {
-    return Impl::Read(strm, opts);
-  }
-
-  // Read FST implementation from a file; return NULL on error.
-  // Empty filename reads from standard input.
-  static Impl *Read(const string &filename) {
-    if (!filename.empty()) {
-      std::ifstream strm(filename,
-                              std::ios_base::in | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "ExpandedFst::Read: Can't open file: " << filename;
-        return nullptr;
-      }
-      return Impl::Read(strm, FstReadOptions(filename));
-    } else {
-      return Impl::Read(std::cin, FstReadOptions("standard input"));
-    }
-  }
-};
-
-// Function to return the number of states in an FST, counting them
-// if necessary.
-template <class Arc>
-typename Arc::StateId CountStates(const Fst<Arc> &fst) {
-  if (fst.Properties(kExpanded, false)) {
-    const auto *efst = static_cast<const ExpandedFst<Arc> *>(&fst);
-    return efst->NumStates();
-  } else {
-    typename Arc::StateId nstates = 0;
-    for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-      ++nstates;
-    }
-    return nstates;
-  }
-}
-
-// Function to return the number of arcs in an FST.
-template <class Arc>
-typename Arc::StateId CountArcs(const Fst<Arc> &fst) {
-  size_t narcs = 0;
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    narcs += fst.NumArcs(siter.Value());
-  }
-  return narcs;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXPANDED_FST_H_
diff --git a/projects/llm_framework/include/fst/expectation-weight.h b/projects/llm_framework/include/fst/expectation-weight.h
deleted file mode 100644
index f996cbc6..00000000
--- a/projects/llm_framework/include/fst/expectation-weight.h
+++ /dev/null
@@ -1,134 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Expectation semiring as described by Jason Eisner:
-// See: doi=10.1.1.22.9398
-// Multiplex semiring operations and identities:
-//    One: <One, Zero>
-//    Zero: <Zero, Zero>
-//    Plus: <a1, b1> + <a2, b2> = < (a1 + a2) , (b1 + b2) >
-//    Times: <a1, b1> * <a2, b2> = < (a1 * a2) , [(a1 * b2) + (a2 * b1)] >
-//    Division: Undefined (currently)
-//
-// Usually used to store the pair <probability, random_variable> so that
-// ShortestDistance[Fst<ArcTpl<ExpectationWeight<P, V>>>]
-//    == < PosteriorProbability, Expected_Value[V] >
-
-#ifndef FST_EXPECTATION_WEIGHT_H_
-#define FST_EXPECTATION_WEIGHT_H_
-
-#include <string>
-
-#include <fst/log.h>
-
-#include <fst/pair-weight.h>
-#include <fst/product-weight.h>
-
-
-namespace fst {
-
-// X1 is usually a probability weight like LogWeight.
-// X2 is usually a random variable or vector (see SignedLogWeight or
-// SparsePowerWeight).
-//
-// If X1 is distinct from X2, it is required that there is an external product
-// between X1 and X2 and if both semriring are commutative, or left or right
-// semirings, then result must have those properties.
-template <class X1, class X2>
-class ExpectationWeight : public PairWeight<X1, X2> {
- public:
-  using PairWeight<X1, X2>::Value1;
-  using PairWeight<X1, X2>::Value2;
-
-  using PairWeight<X1, X2>::Reverse;
-  using PairWeight<X1, X2>::Quantize;
-  using PairWeight<X1, X2>::Member;
-
-  using ReverseWeight =
-      ExpectationWeight<typename X1::ReverseWeight, typename X2::ReverseWeight>;
-
-  ExpectationWeight() : PairWeight<X1, X2>(Zero()) {}
-
-  explicit ExpectationWeight(const PairWeight<X1, X2> &weight)
-      : PairWeight<X1, X2>(weight) {}
-
-  ExpectationWeight(const X1 &x1, const X2 &x2) : PairWeight<X1, X2>(x1, x2) {}
-
-  static const ExpectationWeight &Zero() {
-    static const ExpectationWeight zero(X1::Zero(), X2::Zero());
-    return zero;
-  }
-
-  static const ExpectationWeight &One() {
-    static const ExpectationWeight one(X1::One(), X2::Zero());
-    return one;
-  }
-
-  static const ExpectationWeight &NoWeight() {
-    static const ExpectationWeight no_weight(X1::NoWeight(), X2::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type =
-        new string("expectation_" + X1::Type() + "_" + X2::Type());
-    return *type;
-  }
-
-  PairWeight<X1, X2> Quantize(float delta = kDelta) const {
-    return ExpectationWeight(PairWeight<X1, X2>::Quantize());
-  }
-
-  ReverseWeight Reverse() const {
-    return ReverseWeight(PairWeight<X1, X2>::Reverse());
-  }
-
-  bool Member() const { return PairWeight<X1, X2>::Member(); }
-
-  static constexpr uint64 Properties() {
-    return X1::Properties() & X2::Properties() &
-           (kLeftSemiring | kRightSemiring | kCommutative | kIdempotent);
-  }
-};
-
-template <class X1, class X2>
-inline ExpectationWeight<X1, X2> Plus(const ExpectationWeight<X1, X2> &w1,
-                                      const ExpectationWeight<X1, X2> &w2) {
-  return ExpectationWeight<X1, X2>(Plus(w1.Value1(), w2.Value1()),
-                                   Plus(w1.Value2(), w2.Value2()));
-}
-
-template <class X1, class X2>
-inline ExpectationWeight<X1, X2> Times(const ExpectationWeight<X1, X2> &w1,
-                                       const ExpectationWeight<X1, X2> &w2) {
-  return ExpectationWeight<X1, X2>(
-      Times(w1.Value1(), w2.Value1()),
-      Plus(Times(w1.Value1(), w2.Value2()), Times(w1.Value2(), w2.Value1())));
-}
-
-template <class X1, class X2>
-inline ExpectationWeight<X1, X2> Divide(const ExpectationWeight<X1, X2> &w1,
-                                        const ExpectationWeight<X1, X2> &w2,
-                                        DivideType typ = DIVIDE_ANY) {
-  FSTERROR() << "ExpectationWeight::Divide: Not implemented";
-  return ExpectationWeight<X1, X2>::NoWeight();
-}
-
-// This function object generates weights by calling the underlying generators
-// for the template weight types, like all other pair weight types. This is
-// intended primarily for testing.
-template <class X1, class X2>
-class WeightGenerate<ExpectationWeight<X1, X2>>
-    : public WeightGenerate<PairWeight<X1, X2>> {
- public:
-  using Weight = ExpectationWeight<X1, X2>;
-  using Generate = WeightGenerate<PairWeight<X1, X2>>;
-
-  explicit WeightGenerate(bool allow_zero = true) : Generate(allow_zero) {}
-
-  Weight operator()() const { return Weight(Generate::operator()()); }
-};
-
-}  // namespace fst
-
-#endif  // FST_EXPECTATION_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/extensions/compress/compress-script.h b/projects/llm_framework/include/fst/extensions/compress/compress-script.h
deleted file mode 100644
index bad238aa..00000000
--- a/projects/llm_framework/include/fst/extensions/compress/compress-script.h
+++ /dev/null
@@ -1,53 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Declarations of 'scriptable' versions of compression operations, that is,
-// those that can be called with FstClass-type arguments.
-
-#ifndef FST_EXTENSIONS_COMPRESS_COMPRESS_SCRIPT_H_
-#define FST_EXTENSIONS_COMPRESS_COMPRESS_SCRIPT_H_
-
-#include <string>
-#include <tuple>
-
-#include <fst/log.h>
-#include <fst/extensions/compress/compress.h>
-#include <fst/mutable-fst.h>
-#include <fst/util.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-typedef std::tuple<const FstClass &, const string &, const bool> CompressArgs;
-
-template <class Arc>
-void Compress(CompressArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  const string &filename = std::get<1>(*args);
-  const bool gzip = std::get<2>(*args);
-
-  if (!fst::Compress(fst, filename, gzip)) FSTERROR() << "Compress: failed";
-}
-
-void Compress(const FstClass &fst, const string &filename, const bool gzip);
-
-typedef std::tuple<const string &, MutableFstClass *, const bool>
-    DecompressArgs;
-
-template <class Arc>
-void Decompress(DecompressArgs *args) {
-  const string &filename = std::get<0>(*args);
-  MutableFst<Arc> *fst = std::get<1>(*args)->GetMutableFst<Arc>();
-  const bool gzip = std::get<2>(*args);
-
-  if (!fst::Decompress(filename, fst, gzip))
-    FSTERROR() << "Decompress: failed";
-}
-
-void Decompress(const string &filename, MutableFstClass *fst, const bool gzip);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_COMPRESS_COMPRESS_SCRIPT_H_
diff --git a/projects/llm_framework/include/fst/extensions/compress/compress.h b/projects/llm_framework/include/fst/extensions/compress/compress.h
deleted file mode 100644
index aa94848f..00000000
--- a/projects/llm_framework/include/fst/extensions/compress/compress.h
+++ /dev/null
@@ -1,906 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Compresses and decompresses unweighted FSTs.
-
-#ifndef FST_EXTENSIONS_COMPRESS_COMPRESS_H_
-#define FST_EXTENSIONS_COMPRESS_COMPRESS_H_
-
-#include <algorithm>
-#include <cstdio>
-#include <iostream>
-#include <map>
-#include <memory>
-#include <queue>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/log.h>
-#include <fst/extensions/compress/elias.h>
-#include <fst/extensions/compress/gzfile.h>
-#include <fst/encode.h>
-#include <fst/fst.h>
-#include <fst/mutable-fst.h>
-#include <fst/statesort.h>
-
-namespace fst {
-
-// Identifies stream data as a vanilla compressed FST.
-static const int32 kCompressMagicNumber = 1858869554;
-// Identifies stream data as (probably) a Gzip file accidentally read from
-// a vanilla stream, without gzip support.
-static const int32 kGzipMagicNumber = 0x8b1f;
-// Selects the two most significant bytes.
-constexpr uint32 kGzipMask = 0xffffffff >> 16;
-
-namespace internal {
-
-// Expands a Lempel Ziv code and returns the set of code words. expanded_code[i]
-// is the i^th Lempel Ziv codeword.
-template <class Var, class Edge>
-bool ExpandLZCode(const std::vector<std::pair<Var, Edge>> &code,
-                  std::vector<std::vector<Edge>> *expanded_code) {
-  expanded_code->resize(code.size());
-  for (int i = 0; i < code.size(); ++i) {
-    if (code[i].first > i) {
-      LOG(ERROR) << "ExpandLZCode: Not a valid code";
-      return false;
-    }
-    if (code[i].first == 0) {
-      (*expanded_code)[i].resize(1, code[i].second);
-    } else {
-      (*expanded_code)[i].resize((*expanded_code)[code[i].first - 1].size() +
-                                 1);
-      std::copy((*expanded_code)[code[i].first - 1].begin(),
-                (*expanded_code)[code[i].first - 1].end(),
-                (*expanded_code)[i].begin());
-      (*expanded_code)[i][(*expanded_code)[code[i].first - 1].size()] =
-          code[i].second;
-    }
-  }
-  return true;
-}
-
-}  // namespace internal
-
-// Lempel Ziv on data structure Edge, with a less than operator
-// EdgeLessThan and an equals operator  EdgeEquals.
-// Edge has a value defaultedge which it never takes and
-// Edge is defined, it is initialized to defaultedge
-template <class Var, class Edge, class EdgeLessThan, class EdgeEquals>
-class LempelZiv {
- public:
-  LempelZiv() : dict_number_(0), default_edge_() {
-    root_.current_number = dict_number_++;
-    root_.current_edge = default_edge_;
-    decode_vector_.push_back(std::make_pair(0, default_edge_));
-  }
-  // Encodes a vector input into output
-  void BatchEncode(const std::vector<Edge> &input,
-                   std::vector<std::pair<Var, Edge>> *output);
-
-  // Decodes codedvector to output. Returns false if
-  // the index exceeds the size.
-  bool BatchDecode(const std::vector<std::pair<Var, Edge>> &input,
-                   std::vector<Edge> *output);
-
-  // Decodes a single dictionary element. Returns false
-  // if the index exceeds the size.
-  bool SingleDecode(const Var &index, Edge *output) {
-    if (index >= decode_vector_.size()) {
-      LOG(ERROR) << "LempelZiv::SingleDecode: "
-                 << "Index exceeded the dictionary size";
-      return false;
-    } else {
-      *output = decode_vector_[index].second;
-      return true;
-    }
-  }
-
-  ~LempelZiv() {
-    for (auto it = (root_.next_number).begin(); it != (root_.next_number).end();
-         ++it) {
-      CleanUp(it->second);
-    }
-  }
-  // Adds a single dictionary element while decoding
-  //  void AddDictElement(const std::pair<Var, Edge> &newdict) {
-  //    EdgeEquals InstEdgeEquals;
-  //  if (InstEdgeEquals(newdict.second, default_edge_) != 1)
-  //     decode_vector_.push_back(newdict);
-  //  }
-
- private:
-  // Node datastructure is used for encoding
-
-  struct Node {
-    Var current_number;
-    Edge current_edge;
-    std::map<Edge, Node *, EdgeLessThan> next_number;
-  };
-
-  void CleanUp(Node *temp) {
-    for (auto it = (temp->next_number).begin(); it != (temp->next_number).end();
-         ++it) {
-      CleanUp(it->second);
-    }
-    delete temp;
-  }
-  Node root_;
-  Var dict_number_;
-  // decode_vector_ is used for decoding
-  std::vector<std::pair<Var, Edge>> decode_vector_;
-  Edge default_edge_;
-};
-
-template <class Var, class Edge, class EdgeLessThan, class EdgeEquals>
-void LempelZiv<Var, Edge, EdgeLessThan, EdgeEquals>::BatchEncode(
-    const std::vector<Edge> &input, std::vector<std::pair<Var, Edge>> *output) {
-  for (typename std::vector<Edge>::const_iterator it = input.begin();
-       it != input.end(); ++it) {
-    Node *temp_node = &root_;
-    while (it != input.end()) {
-      auto next = (temp_node->next_number).find(*it);
-      if (next != (temp_node->next_number).end()) {
-        temp_node = next->second;
-        ++it;
-      } else {
-        break;
-      }
-    }
-    if (it == input.end() && temp_node->current_number != 0) {
-      output->push_back(
-          std::make_pair(temp_node->current_number, default_edge_));
-    } else if (it != input.end()) {
-      output->push_back(std::make_pair(temp_node->current_number, *it));
-      Node *new_node = new (Node);
-      new_node->current_number = dict_number_++;
-      new_node->current_edge = *it;
-      (temp_node->next_number)[*it] = new_node;
-    }
-    if (it == input.end()) break;
-  }
-}
-
-template <class Var, class Edge, class EdgeLessThan, class EdgeEquals>
-bool LempelZiv<Var, Edge, EdgeLessThan, EdgeEquals>::BatchDecode(
-    const std::vector<std::pair<Var, Edge>> &input, std::vector<Edge> *output) {
-  for (typename std::vector<std::pair<Var, Edge>>::const_iterator it =
-           input.begin();
-       it != input.end(); ++it) {
-    std::vector<Edge> temp_output;
-    EdgeEquals InstEdgeEquals;
-    if (InstEdgeEquals(it->second, default_edge_) != 1) {
-      decode_vector_.push_back(*it);
-      temp_output.push_back(it->second);
-    }
-    Var temp_integer = it->first;
-    if (temp_integer >= decode_vector_.size()) {
-      LOG(ERROR) << "LempelZiv::BatchDecode: "
-                 << "Index exceeded the dictionary size";
-      return false;
-    } else {
-      while (temp_integer != 0) {
-        temp_output.push_back(decode_vector_[temp_integer].second);
-        temp_integer = decode_vector_[temp_integer].first;
-      }
-      std::reverse(temp_output.begin(), temp_output.end());
-      output->insert(output->end(), temp_output.begin(), temp_output.end());
-    }
-  }
-  return true;
-}
-
-// The main Compressor class
-template <class Arc>
-class Compressor {
- public:
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Label Label;
-  typedef typename Arc::Weight Weight;
-
-  Compressor() {}
-
-  // Compresses fst into a boolean vector code. Returns true on sucesss.
-  bool Compress(const Fst<Arc> &fst, std::ostream &strm);
-
-  // Decompresses the boolean vector into Fst. Returns true on sucesss.
-  bool Decompress(std::istream &strm, const string &source,
-                  MutableFst<Arc> *fst);
-
-  // Finds the BFS order of a fst
-  void BfsOrder(const ExpandedFst<Arc> &fst, std::vector<StateId> *order);
-
-  // Preprocessing step to convert fst to a isomorphic fst
-  // Returns a preproccess fst and a dictionary
-  void Preprocess(const Fst<Arc> &fst, MutableFst<Arc> *preprocessedfst,
-                  EncodeMapper<Arc> *encoder);
-
-  // Performs Lempel Ziv and outputs a stream of integers
-  // and sends it to a stream
-  void EncodeProcessedFst(const ExpandedFst<Arc> &fst, std::ostream &strm);
-
-  // Decodes fst from the stream
-  void DecodeProcessedFst(const std::vector<StateId> &input,
-                          MutableFst<Arc> *fst, bool unweighted);
-
-  // Converts buffer_code_ to uint8 and writes to a stream.
-
-  // Writes the boolean file to the stream
-  void WriteToStream(std::ostream &strm);
-
-  // Writes the weights to the stream
-  void WriteWeight(const std::vector<Weight> &input, std::ostream &strm);
-
-  void ReadWeight(std::istream &strm, std::vector<Weight> *output);
-
-  // Same as fst::Decode without the line RmFinalEpsilon(fst)
-  void DecodeForCompress(MutableFst<Arc> *fst, const EncodeMapper<Arc> &mapper);
-
-  // Updates the buffer_code_
-  template <class CVar>
-  void WriteToBuffer(CVar input) {
-    std::vector<bool> current_code;
-    Elias<CVar>::DeltaEncode(input, &current_code);
-    if (!buffer_code_.empty()) {
-      buffer_code_.insert(buffer_code_.end(), current_code.begin(),
-                          current_code.end());
-    } else {
-      buffer_code_.assign(current_code.begin(), current_code.end());
-    }
-  }
-
- private:
-  struct LZLabel {
-    LZLabel() : label(0) {}
-    Label label;
-  };
-
-  struct LabelLessThan {
-    bool operator()(const LZLabel &labelone, const LZLabel &labeltwo) const {
-      return labelone.label < labeltwo.label;
-    }
-  };
-
-  struct LabelEquals {
-    bool operator()(const LZLabel &labelone, const LZLabel &labeltwo) const {
-      return labelone.label == labeltwo.label;
-    }
-  };
-
-  struct Transition {
-    Transition() : nextstate(0), label(0), weight(Weight::Zero()) {}
-
-    StateId nextstate;
-    Label label;
-    Weight weight;
-  };
-
-  struct TransitionLessThan {
-    bool operator()(const Transition &transition_one,
-                    const Transition &transition_two) const {
-      if (transition_one.nextstate == transition_two.nextstate)
-        return transition_one.label < transition_two.label;
-      else
-        return transition_one.nextstate < transition_two.nextstate;
-    }
-  } transition_less_than;
-
-  struct TransitionEquals {
-    bool operator()(const Transition &transition_one,
-                    const Transition &transition_two) const {
-      return transition_one.nextstate == transition_two.nextstate &&
-             transition_one.label == transition_two.label;
-    }
-  } transition_equals;
-
-  struct OldDictCompare {
-    bool operator()(const std::pair<StateId, Transition> &pair_one,
-                    const std::pair<StateId, Transition> &pair_two) const {
-      if ((pair_one.second).nextstate == (pair_two.second).nextstate)
-        return (pair_one.second).label < (pair_two.second).label;
-      else
-        return (pair_one.second).nextstate < (pair_two.second).nextstate;
-    }
-  } old_dict_compare;
-
-  std::vector<bool> buffer_code_;
-  std::vector<Weight> arc_weight_;
-  std::vector<Weight> final_weight_;
-};
-
-template <class Arc>
-inline void Compressor<Arc>::DecodeForCompress(
-    MutableFst<Arc> *fst, const EncodeMapper<Arc> &mapper) {
-  ArcMap(fst, EncodeMapper<Arc>(mapper, DECODE));
-  fst->SetInputSymbols(mapper.InputSymbols());
-  fst->SetOutputSymbols(mapper.OutputSymbols());
-}
-
-// Compressor::BfsOrder
-template <class Arc>
-void Compressor<Arc>::BfsOrder(const ExpandedFst<Arc> &fst,
-                               std::vector<StateId> *order) {
-  Arc arc;
-  StateId bfs_visit_number = 0;
-  std::queue<StateId> states_queue;
-  order->assign(fst.NumStates(), kNoStateId);
-  states_queue.push(fst.Start());
-  (*order)[fst.Start()] = bfs_visit_number++;
-  while (!states_queue.empty()) {
-    for (ArcIterator<Fst<Arc>> aiter(fst, states_queue.front()); !aiter.Done();
-         aiter.Next()) {
-      arc = aiter.Value();
-      StateId nextstate = arc.nextstate;
-      if ((*order)[nextstate] == kNoStateId) {
-        (*order)[nextstate] = bfs_visit_number++;
-        states_queue.push(nextstate);
-      }
-    }
-    states_queue.pop();
-  }
-
-  // If the FST is unconnected, then the following
-  // code finds them
-  while (bfs_visit_number < fst.NumStates()) {
-    int unseen_state = 0;
-    for (unseen_state = 0; unseen_state < fst.NumStates(); ++unseen_state) {
-      if ((*order)[unseen_state] == kNoStateId) break;
-    }
-    states_queue.push(unseen_state);
-    (*order)[unseen_state] = bfs_visit_number++;
-    while (!states_queue.empty()) {
-      for (ArcIterator<Fst<Arc>> aiter(fst, states_queue.front());
-           !aiter.Done(); aiter.Next()) {
-        arc = aiter.Value();
-        StateId nextstate = arc.nextstate;
-        if ((*order)[nextstate] == kNoStateId) {
-          (*order)[nextstate] = bfs_visit_number++;
-          states_queue.push(nextstate);
-        }
-      }
-      states_queue.pop();
-    }
-  }
-}
-
-template <class Arc>
-void Compressor<Arc>::Preprocess(const Fst<Arc> &fst,
-                                 MutableFst<Arc> *preprocessedfst,
-                                 EncodeMapper<Arc> *encoder) {
-  *preprocessedfst = fst;
-  if (!preprocessedfst->NumStates()) {
-    return;
-  }
-  // Relabels the edges and develops a dictionary
-  Encode(preprocessedfst, encoder);
-  std::vector<StateId> order;
-  // Finds the BFS sorting order of the fst
-  BfsOrder(*preprocessedfst, &order);
-  // Reorders the states according to the BFS order
-  StateSort(preprocessedfst, order);
-}
-
-template <class Arc>
-void Compressor<Arc>::EncodeProcessedFst(const ExpandedFst<Arc> &fst,
-                                         std::ostream &strm) {
-  std::vector<StateId> output;
-  LempelZiv<StateId, LZLabel, LabelLessThan, LabelEquals> dict_new;
-  LempelZiv<StateId, Transition, TransitionLessThan, TransitionEquals> dict_old;
-  std::vector<LZLabel> current_new_input;
-  std::vector<Transition> current_old_input;
-  std::vector<std::pair<StateId, LZLabel>> current_new_output;
-  std::vector<std::pair<StateId, Transition>> current_old_output;
-  std::vector<StateId> final_states;
-
-  StateId number_of_states = fst.NumStates();
-
-  StateId seen_states = 0;
-  // Adding the number of states
-  WriteToBuffer<StateId>(number_of_states);
-
-  for (StateId state = 0; state < number_of_states; ++state) {
-    current_new_input.clear();
-    current_old_input.clear();
-    current_new_output.clear();
-    current_old_output.clear();
-    if (state > seen_states) ++seen_states;
-
-    // Collecting the final states
-    if (fst.Final(state) != Weight::Zero()) {
-      final_states.push_back(state);
-      final_weight_.push_back(fst.Final(state));
-    }
-
-    // Reading the states
-    for (ArcIterator<Fst<Arc>> aiter(fst, state); !aiter.Done(); aiter.Next()) {
-      Arc arc = aiter.Value();
-      if (arc.nextstate > seen_states) {  // RILEY: > or >= ?
-        ++seen_states;
-        LZLabel temp_label;
-        temp_label.label = arc.ilabel;
-        arc_weight_.push_back(arc.weight);
-        current_new_input.push_back(temp_label);
-      } else {
-        Transition temp_transition;
-        temp_transition.nextstate = arc.nextstate;
-        temp_transition.label = arc.ilabel;
-        temp_transition.weight = arc.weight;
-        current_old_input.push_back(temp_transition);
-      }
-    }
-    // Adding new states
-    dict_new.BatchEncode(current_new_input, &current_new_output);
-    WriteToBuffer<StateId>(current_new_output.size());
-
-    for (auto it = current_new_output.begin(); it != current_new_output.end();
-         ++it) {
-      WriteToBuffer<StateId>(it->first);
-      WriteToBuffer<Label>((it->second).label);
-    }
-    // Adding old states by sorting and using difference coding
-    std::sort(current_old_input.begin(), current_old_input.end(),
-              transition_less_than);
-    for (auto it = current_old_input.begin(); it != current_old_input.end();
-         ++it) {
-      arc_weight_.push_back(it->weight);
-    }
-    dict_old.BatchEncode(current_old_input, &current_old_output);
-    std::vector<StateId> dict_old_temp;
-    std::vector<Transition> transition_old_temp;
-    for (auto it = current_old_output.begin(); it != current_old_output.end();
-         ++it) {
-      dict_old_temp.push_back(it->first);
-      transition_old_temp.push_back(it->second);
-    }
-    if (!transition_old_temp.empty()) {
-      if ((transition_old_temp.back()).nextstate == 0 &&
-          (transition_old_temp.back()).label == 0) {
-        transition_old_temp.pop_back();
-      }
-    }
-    std::sort(dict_old_temp.begin(), dict_old_temp.end());
-    std::sort(transition_old_temp.begin(), transition_old_temp.end(),
-              transition_less_than);
-
-    WriteToBuffer<StateId>(dict_old_temp.size());
-    if (dict_old_temp.size() != transition_old_temp.size())
-      WriteToBuffer<int>(1);
-    else
-      WriteToBuffer<int>(0);
-
-    StateId previous;
-    if (!dict_old_temp.empty()) {
-      WriteToBuffer<StateId>(dict_old_temp.front());
-      previous = dict_old_temp.front();
-    }
-    if (dict_old_temp.size() > 1) {
-      for (auto it = dict_old_temp.begin() + 1; it != dict_old_temp.end();
-           ++it) {
-        WriteToBuffer<StateId>(*it - previous);
-        previous = *it;
-      }
-    }
-    if (!transition_old_temp.empty()) {
-      WriteToBuffer<StateId>((transition_old_temp.front()).nextstate);
-      previous = (transition_old_temp.front()).nextstate;
-      WriteToBuffer<Label>((transition_old_temp.front()).label);
-    }
-    if (transition_old_temp.size() > 1) {
-      for (auto it = transition_old_temp.begin() + 1;
-           it != transition_old_temp.end(); ++it) {
-        WriteToBuffer<StateId>(it->nextstate - previous);
-        previous = it->nextstate;
-        WriteToBuffer<StateId>(it->label);
-      }
-    }
-  }
-  // Adding final states
-  WriteToBuffer<StateId>(final_states.size());
-  if (!final_states.empty()) {
-    for (auto it = final_states.begin(); it != final_states.end(); ++it) {
-      WriteToBuffer<StateId>(*it);
-    }
-  }
-  WriteToStream(strm);
-  uint8 unweighted = (fst.Properties(kUnweighted, true) == kUnweighted);
-  WriteType(strm, unweighted);
-  if (unweighted == 0) {
-    WriteWeight(arc_weight_, strm);
-    WriteWeight(final_weight_, strm);
-  }
-}
-
-template <class Arc>
-void Compressor<Arc>::DecodeProcessedFst(const std::vector<StateId> &input,
-                                         MutableFst<Arc> *fst,
-                                         bool unweighted) {
-  LempelZiv<StateId, LZLabel, LabelLessThan, LabelEquals> dict_new;
-  LempelZiv<StateId, Transition, TransitionLessThan, TransitionEquals> dict_old;
-  std::vector<std::pair<StateId, LZLabel>> current_new_input;
-  std::vector<std::pair<StateId, Transition>> current_old_input;
-  std::vector<LZLabel> current_new_output;
-  std::vector<Transition> current_old_output;
-  std::vector<std::pair<StateId, Transition>> actual_old_dict_numbers;
-  std::vector<Transition> actual_old_dict_transitions;
-  auto arc_weight_it = arc_weight_.begin();
-  Transition default_transition;
-  StateId seen_states = 1;
-
-  // Adding states.
-  const StateId num_states = input.front();
-  if (num_states > 0) {
-    const StateId start_state = fst->AddState();
-    fst->SetStart(start_state);
-    for (StateId state = 1; state < num_states; ++state) {
-      fst->AddState();
-    }
-  }
-
-  typename std::vector<StateId>::const_iterator main_it = input.begin();
-  ++main_it;
-
-  for (StateId current_state = 0; current_state < num_states; ++current_state) {
-    if (current_state >= seen_states) ++seen_states;
-    current_new_input.clear();
-    current_new_output.clear();
-    current_old_input.clear();
-    current_old_output.clear();
-    // New states
-    StateId current_number_new_elements = *main_it;
-    ++main_it;
-    for (StateId new_integer = 0; new_integer < current_number_new_elements;
-         ++new_integer) {
-      std::pair<StateId, LZLabel> temp_new_dict_element;
-      temp_new_dict_element.first = *main_it;
-      ++main_it;
-      LZLabel temp_label;
-      temp_label.label = *main_it;
-      ++main_it;
-      temp_new_dict_element.second = temp_label;
-      current_new_input.push_back(temp_new_dict_element);
-    }
-    dict_new.BatchDecode(current_new_input, &current_new_output);
-    for (auto it = current_new_output.begin(); it != current_new_output.end();
-         ++it) {
-      if (!unweighted) {
-        fst->AddArc(current_state,
-                    Arc(it->label, it->label, *arc_weight_it, seen_states++));
-        ++arc_weight_it;
-      } else {
-        fst->AddArc(current_state,
-                    Arc(it->label, it->label, Weight::One(), seen_states++));
-      }
-    }
-
-    // Old states dictionary
-    StateId current_number_old_elements = *main_it;
-    ++main_it;
-    StateId is_zero_removed = *main_it;
-    ++main_it;
-    StateId previous = 0;
-    actual_old_dict_numbers.clear();
-    for (StateId new_integer = 0; new_integer < current_number_old_elements;
-         ++new_integer) {
-      std::pair<StateId, Transition> pair_temp_transition;
-      if (new_integer == 0) {
-        pair_temp_transition.first = *main_it;
-        previous = *main_it;
-      } else {
-        pair_temp_transition.first = *main_it + previous;
-        previous = pair_temp_transition.first;
-      }
-      ++main_it;
-      Transition temp_test;
-      if (!dict_old.SingleDecode(pair_temp_transition.first, &temp_test)) {
-        FSTERROR() << "Compressor::Decode: failed";
-        fst->DeleteStates();
-        fst->SetProperties(kError, kError);
-        return;
-      }
-      pair_temp_transition.second = temp_test;
-      actual_old_dict_numbers.push_back(pair_temp_transition);
-    }
-
-    // Reordering the dictionary elements
-    std::sort(actual_old_dict_numbers.begin(), actual_old_dict_numbers.end(),
-              old_dict_compare);
-
-    // Transitions
-    previous = 0;
-    actual_old_dict_transitions.clear();
-
-    for (StateId new_integer = 0;
-         new_integer < current_number_old_elements - is_zero_removed;
-         ++new_integer) {
-      Transition temp_transition;
-      if (new_integer == 0) {
-        temp_transition.nextstate = *main_it;
-        previous = *main_it;
-      } else {
-        temp_transition.nextstate = *main_it + previous;
-        previous = temp_transition.nextstate;
-      }
-      ++main_it;
-      temp_transition.label = *main_it;
-      ++main_it;
-      actual_old_dict_transitions.push_back(temp_transition);
-    }
-
-    if (is_zero_removed == 1) {
-      actual_old_dict_transitions.push_back(default_transition);
-    }
-
-    auto trans_it = actual_old_dict_transitions.begin();
-    auto dict_it = actual_old_dict_numbers.begin();
-
-    while (trans_it != actual_old_dict_transitions.end() &&
-           dict_it != actual_old_dict_numbers.end()) {
-      if (dict_it->first == 0) {
-        ++dict_it;
-      } else {
-        std::pair<StateId, Transition> temp_pair;
-        if (transition_equals(*trans_it, default_transition) == 1) {
-          temp_pair.first = dict_it->first;
-          temp_pair.second = default_transition;
-          ++dict_it;
-        } else if (transition_less_than(dict_it->second, *trans_it) == 1) {
-          temp_pair.first = dict_it->first;
-          temp_pair.second = *trans_it;
-          ++dict_it;
-        } else {
-          temp_pair.first = 0;
-          temp_pair.second = *trans_it;
-        }
-        ++trans_it;
-        current_old_input.push_back(temp_pair);
-      }
-    }
-    while (trans_it != actual_old_dict_transitions.end()) {
-      std::pair<StateId, Transition> temp_pair;
-      temp_pair.first = 0;
-      temp_pair.second = *trans_it;
-      ++trans_it;
-      current_old_input.push_back(temp_pair);
-    }
-
-    // Adding old elements in the dictionary
-    if (!dict_old.BatchDecode(current_old_input, &current_old_output)) {
-      FSTERROR() << "Compressor::Decode: Failed";
-      fst->DeleteStates();
-      fst->SetProperties(kError, kError);
-      return;
-    }
-
-    for (auto it = current_old_output.begin(); it != current_old_output.end();
-         ++it) {
-      if (!unweighted) {
-        fst->AddArc(current_state,
-                    Arc(it->label, it->label, *arc_weight_it, it->nextstate));
-        ++arc_weight_it;
-      } else {
-        fst->AddArc(current_state,
-                    Arc(it->label, it->label, Weight::One(), it->nextstate));
-      }
-    }
-  }
-  // Adding the final states
-  StateId number_of_final_states = *main_it;
-  if (number_of_final_states > 0) {
-    ++main_it;
-    for (StateId temp_int = 0; temp_int < number_of_final_states; ++temp_int) {
-      if (!unweighted) {
-        fst->SetFinal(*main_it, final_weight_[temp_int]);
-      } else {
-        fst->SetFinal(*main_it, Weight(0));
-      }
-      ++main_it;
-    }
-  }
-}
-
-template <class Arc>
-void Compressor<Arc>::ReadWeight(std::istream &strm,
-                                 std::vector<Weight> *output) {
-  int64 size;
-  Weight weight;
-  ReadType(strm, &size);
-  for (int64 i = 0; i < size; ++i) {
-    weight.Read(strm);
-    output->push_back(weight);
-  }
-}
-
-template <class Arc>
-bool Compressor<Arc>::Decompress(std::istream &strm, const string &source,
-                                 MutableFst<Arc> *fst) {
-  fst->DeleteStates();
-  int32 magic_number = 0;
-  ReadType(strm, &magic_number);
-  if (magic_number != kCompressMagicNumber) {
-    LOG(ERROR) << "Decompress: Bad compressed Fst: " << source;
-    // If the most significant two bytes of the magic number match the
-    // gzip magic number, then we are probably reading a gzip file as an
-    // ordinary stream.
-    if ((magic_number & kGzipMask) == kGzipMagicNumber) {
-      LOG(ERROR) << "Decompress: Fst appears to be compressed with Gzip, but "
-                    "gzip decompression was not requested. Try with "
-                    "the --gzip flag"
-                    ".";
-    }
-    return false;
-  }
-  std::unique_ptr<EncodeMapper<Arc>> encoder(
-      EncodeMapper<Arc>::Read(strm, "Decoding", DECODE));
-  std::vector<bool> bool_code;
-  uint8 block;
-  uint8 msb = 128;
-  int64 data_size;
-  ReadType(strm, &data_size);
-  for (int64 i = 0; i < data_size; ++i) {
-    ReadType(strm, &block);
-    for (int j = 0; j < 8; ++j) {
-      uint8 temp = msb & block;
-      if (temp == 128)
-        bool_code.push_back(1);
-      else
-        bool_code.push_back(0);
-      block = block << 1;
-    }
-  }
-  std::vector<StateId> int_code;
-  Elias<StateId>::BatchDecode(bool_code, &int_code);
-  bool_code.clear();
-  uint8 unweighted;
-  ReadType(strm, &unweighted);
-  if (unweighted == 0) {
-    ReadWeight(strm, &arc_weight_);
-    ReadWeight(strm, &final_weight_);
-  }
-  DecodeProcessedFst(int_code, fst, unweighted);
-  DecodeForCompress(fst, *encoder);
-  return !fst->Properties(kError, false);
-}
-
-template <class Arc>
-void Compressor<Arc>::WriteWeight(const std::vector<Weight> &input,
-                                  std::ostream &strm) {
-  int64 size = input.size();
-  WriteType(strm, size);
-  for (typename std::vector<Weight>::const_iterator it = input.begin();
-       it != input.end(); ++it) {
-    it->Write(strm);
-  }
-}
-
-template <class Arc>
-void Compressor<Arc>::WriteToStream(std::ostream &strm) {
-  while (buffer_code_.size() % 8 != 0) buffer_code_.push_back(1);
-  int64 data_size = buffer_code_.size() / 8;
-  WriteType(strm, data_size);
-  std::vector<bool>::const_iterator it;
-  int64 i;
-  uint8 block;
-  for (it = buffer_code_.begin(), i = 0; it != buffer_code_.end(); ++it, ++i) {
-    if (i % 8 == 0) {
-      if (i > 0) WriteType(strm, block);
-      block = 0;
-    } else {
-      block = block << 1;
-    }
-    block |= *it;
-  }
-  WriteType(strm, block);
-}
-
-template <class Arc>
-bool Compressor<Arc>::Compress(const Fst<Arc> &fst, std::ostream &strm) {
-  VectorFst<Arc> processedfst;
-  EncodeMapper<Arc> encoder(kEncodeLabels, ENCODE);
-  Preprocess(fst, &processedfst, &encoder);
-  WriteType(strm, kCompressMagicNumber);
-  encoder.Write(strm, "encoder stream");
-  EncodeProcessedFst(processedfst, strm);
-  return true;
-}
-
-// Convenience functions that call the compressor and decompressor.
-
-template <class Arc>
-void Compress(const Fst<Arc> &fst, std::ostream &strm) {
-  Compressor<Arc> comp;
-  comp.Compress(fst, strm);
-}
-
-// Returns true on success.
-template <class Arc>
-bool Compress(const Fst<Arc> &fst, const string &file_name,
-              const bool gzip = false) {
-  if (gzip) {
-    if (file_name.empty()) {
-      std::stringstream strm;
-      Compress(fst, strm);
-      OGzFile gzfile(fileno(stdout));
-      gzfile.write(strm);
-      if (!gzfile) {
-        LOG(ERROR) << "Compress: Can't write to file: stdout";
-        return false;
-      }
-    } else {
-      std::stringstream strm;
-      Compress(fst, strm);
-      OGzFile gzfile(file_name);
-      if (!gzfile) {
-        LOG(ERROR) << "Compress: Can't open file: " << file_name;
-        return false;
-      }
-      gzfile.write(strm);
-      if (!gzfile) {
-        LOG(ERROR) << "Compress: Can't write to file: " << file_name;
-        return false;
-      }
-    }
-  } else if (file_name.empty()) {
-    Compress(fst, std::cout);
-  } else {
-    std::ofstream strm(file_name,
-                             std::ios_base::out | std::ios_base::binary);
-    if (!strm) {
-      LOG(ERROR) << "Compress: Can't open file: " << file_name;
-      return false;
-    }
-    Compress(fst, strm);
-  }
-  return true;
-}
-
-template <class Arc>
-void Decompress(std::istream &strm, const string &source,
-                MutableFst<Arc> *fst) {
-  Compressor<Arc> comp;
-  comp.Decompress(strm, source, fst);
-}
-
-// Returns true on success.
-template <class Arc>
-bool Decompress(const string &file_name, MutableFst<Arc> *fst,
-                const bool gzip = false) {
-  if (gzip) {
-    if (file_name.empty()) {
-      IGzFile gzfile(fileno(stdin));
-      Decompress(*gzfile.read(), "stdin", fst);
-      if (!gzfile) {
-        LOG(ERROR) << "Decompress: Can't read from file: stdin";
-        return false;
-      }
-    } else {
-      IGzFile gzfile(file_name);
-      if (!gzfile) {
-        LOG(ERROR) << "Decompress: Can't open file: " << file_name;
-        return false;
-      }
-      Decompress(*gzfile.read(), file_name, fst);
-      if (!gzfile) {
-        LOG(ERROR) << "Decompress: Can't read from file: " << file_name;
-        return false;
-      }
-    }
-  } else if (file_name.empty()) {
-    Decompress(std::cin, "stdin", fst);
-  } else {
-    std::ifstream strm(file_name,
-                            std::ios_base::in | std::ios_base::binary);
-    if (!strm) {
-      LOG(ERROR) << "Decompress: Can't open file: " << file_name;
-      return false;
-    }
-    Decompress(strm, file_name, fst);
-  }
-  return true;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_COMPRESS_COMPRESS_H_
diff --git a/projects/llm_framework/include/fst/extensions/compress/elias.h b/projects/llm_framework/include/fst/extensions/compress/elias.h
deleted file mode 100644
index 6c576d33..00000000
--- a/projects/llm_framework/include/fst/extensions/compress/elias.h
+++ /dev/null
@@ -1,97 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Compresses and decompresses unweighted FSTs.
-
-#ifndef FST_EXTENSIONS_COMPRESS_ELIAS_H_
-#define FST_EXTENSIONS_COMPRESS_ELIAS_H_
-
-#include <stack>
-#include <vector>
-
-#include <fst/compat.h>
-namespace fst {
-
-template <class Var>
-class Elias {
- public:
-  // Gamma encoding is a subroutine for Delta encoding
-  static void GammaEncode(const Var &input, std::vector<bool> *code);
-
-  // Elias Delta encoding for a single integer
-  static void DeltaEncode(const Var &input, std::vector<bool> *code);
-
-  // Batch decoding of a set of integers
-  static void BatchDecode(const std::vector<bool> &input,
-                          std::vector<Var> *output);
-};
-
-template <class Var>
-void Elias<Var>::GammaEncode(const Var &input, std::vector<bool> *code) {
-  Var input_copy = input;
-  std::stack<bool> reverse_code;
-  while (input_copy > 0) {
-    reverse_code.push(input_copy % 2);
-    input_copy = input_copy / 2;
-  }
-  for (Var auxvar = 0; auxvar < reverse_code.size() - 1; auxvar++)
-    code->push_back(0);
-  while (reverse_code.empty() != 1) {
-    code->push_back(reverse_code.top());
-    reverse_code.pop();
-  }
-}
-template <class Var>
-void Elias<Var>::DeltaEncode(const Var &input, std::vector<bool> *code) {
-  Var input_copy = input + 1;
-  std::stack<bool> reverse_remainder;
-  Var auxvar = 0;
-  while (input_copy != 0) {
-    reverse_remainder.push(input_copy % 2);
-    input_copy = input_copy / 2;
-    auxvar = auxvar + 1;
-  }
-  GammaEncode(auxvar, code);
-  reverse_remainder.pop();
-  while (reverse_remainder.empty() != 1) {
-    code->push_back(reverse_remainder.top());
-    reverse_remainder.pop();
-  }
-}
-
-template <class Var>
-void Elias<Var>::BatchDecode(const std::vector<bool> &input,
-                             std::vector<Var> *output) {
-  Var lead_zeros = 0;
-  Var remainder_bits = 0;
-  Var current_word = 1;
-  Var value = 1;
-  std::vector<bool>::const_iterator it = input.begin();
-  while (it != input.end()) {
-    lead_zeros = 0;
-    remainder_bits = 0;
-    current_word = 1;
-    value = 1;
-    while (*it != 1) {
-      it++;
-      lead_zeros++;
-    }
-    it++;
-    while (lead_zeros > 0) {
-      lead_zeros--;
-      current_word = 2 * current_word + *it;
-      it++;
-    }
-    current_word--;
-    while (current_word > 0) {
-      value = 2 * value + *it;
-      current_word--;
-      it++;
-    }
-    output->push_back(value - 1);
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_COMPRESS_ELIAS_H_
diff --git a/projects/llm_framework/include/fst/extensions/compress/gzfile.h b/projects/llm_framework/include/fst/extensions/compress/gzfile.h
deleted file mode 100644
index 2d2e6243..00000000
--- a/projects/llm_framework/include/fst/extensions/compress/gzfile.h
+++ /dev/null
@@ -1,127 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Resource handles for gzip files written to or read from stringstreams. These
-// are necessary to provide the compression routines with streams reading from
-// or writing to compressed files (or the UNIX standard streams), and are not
-// intended for general use.
-
-#ifndef FST_EXTENSIONS_COMPRESS_GZFILE_H_
-#define FST_EXTENSIONS_COMPRESS_GZFILE_H_
-
-#include <cstddef>
-#include <memory>
-#include <sstream>
-#include <string>
-
-#include <fst/compat.h>
-#include <fst/log.h>
-#include <fst/fst.h>
-#include <zlib.h>
-
-using std::stringstream;
-using std::unique_ptr;
-
-namespace fst {
-
-// Gives the zlib gzFile type an OO-like interface. String inputs are all
-// C-style strings. The caller is responsible to get the file modes appropriate
-// for the IO methods being called, and for opening the file descriptors
-// if that constructor is used. The ! operator can be used to check for errors
-// after construction or read/writing.
-class GzFile {
- public:
-  GzFile(const char *filename, const char *mode)
-      : gzfile_(gzopen(filename, mode)), error_(check_handle()) {
-  }
-
-  // The caller is responsible to ensure the corresponding FD is open and has
-  // the needed modes ("r" for reading, "w" or "a" for writing).
-  explicit GzFile(const int fd, const char *mode)
-      : gzfile_(gzdopen(fd, mode)), error_(check_handle()), close_me_(false) {
-  }
-
-  // If the instance was constructed from an FD, flush the buffer; otherwise,
-  // close the file, which flushes the buffer as a side-effect.
-  ~GzFile() { close_me_ ? gzclose(gzfile_) : gzflush(gzfile_, Z_FINISH); }
-
-  inline bool operator!() const { return error_; }
-
-  // Returns false on EOF and sets error if short read does not reach an EOF.
-  int read(void *buf, unsigned int size) {
-    int bytes_read = gzread(gzfile_, buf, size);
-    if ((bytes_read < size) && !gzeof(gzfile_)) error_ = true;
-    return bytes_read;
-  }
-
-  // Sets error on short writes.
-  void write(const char *buf, unsigned int size) {
-    if (gzwrite(gzfile_, buf, size) != size) error_ = true;
-  }
-
- private:
-  // gzopen and gzdopen signal failure by returning null.
-  bool check_handle() { return gzfile_ == nullptr; }
-
-  gzFile gzfile_ = nullptr;
-  bool error_ = false;
-  bool close_me_ = false;
-};
-
-// Resource handle for writing stringstream to GzFile.
-class OGzFile {
- public:
-  explicit OGzFile(const string &filename) : OGzFile(filename.c_str()) {}
-
-  explicit OGzFile(const char *filename) : gz_(GzFile(filename, mode_)) {}
-
-  explicit OGzFile(const int fd) : gz_(GzFile(fd, mode_)) {}
-
-  inline bool operator!() const { return !gz_; }
-
-  void write(const std::stringstream &ssbuf) {
-    string sbuf = ssbuf.str();
-    gz_.write(sbuf.data(), sbuf.size());
-  }
-
- private:
-  GzFile gz_;
-  static constexpr auto &mode_ = "wb";
-};
-
-// Resource handle for reading stringstream from GzFile.
-class IGzFile {
- public:
-  explicit IGzFile(const string &filename) : IGzFile(filename.c_str()) {}
-
-  explicit IGzFile(const char *filename) : gz_(GzFile(filename, mode_)) {}
-
-  explicit IGzFile(const int fd) : gz_(GzFile(fd, mode_)) {}
-
-  inline bool operator!() const { return !gz_; }
-
-  // This is a great case for "move", but GCC 4 is missing the C+11 standard
-  // move constructor for stringstream, so a unique_ptr is the next best thing.
-  std::unique_ptr<std::stringstream> read() {
-    char buf[bufsize_];
-    std::unique_ptr<std::stringstream> sstrm(new std::stringstream);
-    // We always read at least once, and the result of the last read is always
-    // pushed onto the stringstream. We use the "write" member because << onto
-    // a stringstream stops at the null byte, which might be data!
-    int bytes_read;
-    while ((bytes_read = gz_.read(buf, bufsize_)) == bufsize_)
-      sstrm->write(buf, bufsize_);
-    sstrm->write(buf, bytes_read);
-    return sstrm;
-  }
-
- private:
-  GzFile gz_;
-  static constexpr auto &mode_ = "rb";
-  // This is the same size as the default internal buffer for zlib.
-  static const size_t bufsize_ = 8192;
-};
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_COMPRESS_GZFILE_H_
diff --git a/projects/llm_framework/include/fst/extensions/compress/randmod.h b/projects/llm_framework/include/fst/extensions/compress/randmod.h
deleted file mode 100644
index bfe2de5b..00000000
--- a/projects/llm_framework/include/fst/extensions/compress/randmod.h
+++ /dev/null
@@ -1,102 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Generates a random FST according to a class-specific transition model.
-
-#ifndef FST_EXTENSIONS_COMPRESS_RANDMOD_H_
-#define FST_EXTENSIONS_COMPRESS_RANDMOD_H_
-
-#include <cstdlib>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/mutable-fst.h>
-
-namespace fst {
-
-template <class Arc, class G>
-class RandMod {
- public:
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Label Label;
-  typedef typename Arc::Weight Weight;
-
-  // Generates random FST with 'nstates' with 'nclasses' in the probability
-  // generation model, and 'nlabels' in the alphabet. If 'trans' = true, then
-  // a transducer is generated; iff 'generate_' is non-null, the output is
-  // randomly weighted.
-  RandMod(StateId nstates, StateId nclasses, Label nlabels, bool trans,
-          const G *generate)
-      : nstates_(nstates),
-        nclasses_(nclasses),
-        nlabels_(nlabels),
-        trans_(trans),
-        generate_(generate) {
-    for (StateId s = 0; s < nstates; ++s) {
-      classes_.push_back(rand() % nclasses);  // NOLINT
-    }
-  }
-
-  // Generates a random FST according to a class-specific transition model
-  void Generate(StdMutableFst *fst) {
-    StateId start = rand() % nstates_;  // NOLINT
-    fst->DeleteStates();
-    for (StateId s = 0; s < nstates_; ++s) {
-      fst->AddState();
-      if (s == start) fst->SetStart(start);
-      for (StateId n = 0; n <= nstates_; ++n) {
-        Arc arc;
-        StateId d = n == nstates_ ? kNoStateId : n;
-        if (!RandArc(s, d, &arc)) continue;
-        if (d == kNoStateId) {  // A super-final transition?
-          fst->SetFinal(s, arc.weight);
-        } else {
-          fst->AddArc(s, arc);
-        }
-      }
-    }
-  }
-
- private:
-  // Generates a transition from s to d. If d == kNoStateId, a superfinal
-  // transition is generated. Returns false if no transition generated.
-  bool RandArc(StateId s, StateId d, Arc *arc) {
-    StateId sclass = classes_[s];
-    StateId dclass = d != kNoStateId ? classes_[d] : 0;
-
-    int r = sclass + dclass + 2;
-    if ((rand() % r) != 0)  // NOLINT
-      return false;
-
-    arc->nextstate = d;
-
-    Label ilabel = kNoLabel;
-    Label olabel = kNoLabel;
-    if (d != kNoStateId) {
-      ilabel = (dclass % nlabels_) + 1;
-      if (trans_)
-        olabel = (sclass % nlabels_) + 1;
-      else
-        olabel = ilabel;
-    }
-
-    Weight weight = Weight::One();
-    if (generate_) weight = (*generate_)();
-
-    arc->ilabel = ilabel;
-    arc->olabel = olabel;
-    arc->weight = weight;
-    return true;
-  }
-
-  StateId nstates_;
-  StateId nclasses_;
-  Label nlabels_;
-  bool trans_;
-  const G *generate_;
-  std::vector<StateId> classes_;
-};
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_COMPRESS_RANDMOD_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/compile-strings.h b/projects/llm_framework/include/fst/extensions/far/compile-strings.h
deleted file mode 100644
index 96271c17..00000000
--- a/projects/llm_framework/include/fst/extensions/far/compile-strings.h
+++ /dev/null
@@ -1,260 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_FAR_COMPILE_STRINGS_H_
-#define FST_EXTENSIONS_FAR_COMPILE_STRINGS_H_
-
-#ifndef _MSC_VER
-#include <libgen.h>
-#else
-#include <fst/compat.h>
-#endif
-
-#include <fstream>
-#include <istream>
-#include <string>
-#include <vector>
-
-#include <fst/extensions/far/far.h>
-#include <fstream>
-#include <fst/string.h>
-
-namespace fst {
-
-// Constructs a reader that provides FSTs from a file (stream) either on a
-// line-by-line basis or on a per-stream basis. Note that the freshly
-// constructed reader is already set to the first input.
-//
-// Sample usage:
-//
-//   for (StringReader<Arc> reader(...); !reader.Done(); reader.Next()) {
-//     auto *fst = reader.GetVectorFst();
-//   }
-template <class Arc>
-class StringReader {
- public:
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-
-  enum EntryType { LINE = 1, FILE = 2 };
-
-  StringReader(std::istream &istrm, const string &source, EntryType entry_type,
-               StringTokenType token_type, bool allow_negative_labels,
-               const SymbolTable *syms = nullptr,
-               Label unknown_label = kNoStateId)
-      : nline_(0),
-        istrm_(istrm),
-        source_(source),
-        entry_type_(entry_type),
-        token_type_(token_type),
-        symbols_(syms),
-        done_(false),
-        compiler_(token_type, syms, unknown_label, allow_negative_labels) {
-    Next();  // Initialize the reader to the first input.
-  }
-
-  bool Done() { return done_; }
-
-  void Next() {
-    VLOG(1) << "Processing source " << source_ << " at line " << nline_;
-    if (!istrm_) {  // We're done if we have no more input.
-      done_ = true;
-      return;
-    }
-    if (entry_type_ == LINE) {
-      getline(istrm_, content_);
-      ++nline_;
-    } else {
-      content_.clear();
-      string line;
-      while (getline(istrm_, line)) {
-        ++nline_;
-        content_.append(line);
-        content_.append("\n");
-      }
-    }
-    if (!istrm_ && content_.empty())  // We're also done if we read off all the
-      done_ = true;                   // whitespace at the end of a file.
-  }
-
-  VectorFst<Arc> *GetVectorFst(bool keep_symbols = false) {
-    std::unique_ptr<VectorFst<Arc>> fst(new VectorFst<Arc>());
-    if (keep_symbols) {
-      fst->SetInputSymbols(symbols_);
-      fst->SetOutputSymbols(symbols_);
-    }
-    if (compiler_(content_, fst.get())) {
-      return fst.release();
-    } else {
-      return nullptr;
-    }
-  }
-
-  CompactStringFst<Arc> *GetCompactFst(bool keep_symbols = false) {
-    std::unique_ptr<CompactStringFst<Arc>> fst;
-    if (keep_symbols) {
-      VectorFst<Arc> tmp;
-      tmp.SetInputSymbols(symbols_);
-      tmp.SetOutputSymbols(symbols_);
-      fst.reset(new CompactStringFst<Arc>(tmp));
-    } else {
-      fst.reset(new CompactStringFst<Arc>());
-    }
-    if (compiler_(content_, fst.get())) {
-      return fst.release();
-    } else {
-      return nullptr;
-    }
-  }
-
- private:
-  size_t nline_;
-  std::istream &istrm_;
-  string source_;
-  EntryType entry_type_;
-  StringTokenType token_type_;
-  const SymbolTable *symbols_;
-  bool done_;
-  StringCompiler<Arc> compiler_;
-  string content_;  // The actual content of the input stream's next FST.
-
-  StringReader(const StringReader &) = delete;
-  StringReader &operator=(const StringReader &) = delete;
-};
-
-// Computes the minimal length required to encode each line number as a decimal
-// number.
-int KeySize(const char *filename);
-
-template <class Arc>
-void FarCompileStrings(const std::vector<string> &in_fnames,
-                       const string &out_fname, const string &fst_type,
-                       const FarType &far_type, int32 generate_keys,
-                       FarEntryType fet, FarTokenType tt,
-                       const string &symbols_fname,
-                       const string &unknown_symbol, bool keep_symbols,
-                       bool initial_symbols, bool allow_negative_labels,
-                       const string &key_prefix, const string &key_suffix) {
-  typename StringReader<Arc>::EntryType entry_type;
-  if (fet == FET_LINE) {
-    entry_type = StringReader<Arc>::LINE;
-  } else if (fet == FET_FILE) {
-    entry_type = StringReader<Arc>::FILE;
-  } else {
-    FSTERROR() << "FarCompileStrings: Unknown entry type";
-    return;
-  }
-  StringTokenType token_type;
-  if (tt == FTT_SYMBOL) {
-    token_type = StringTokenType::SYMBOL;
-  } else if (tt == FTT_BYTE) {
-    token_type = StringTokenType::BYTE;
-  } else if (tt == FTT_UTF8) {
-    token_type = StringTokenType::UTF8;
-  } else {
-    FSTERROR() << "FarCompileStrings: Unknown token type";
-    return;
-  }
-  bool compact;
-  if (fst_type.empty() || (fst_type == "vector")) {
-    compact = false;
-  } else if (fst_type == "compact") {
-    compact = true;
-  } else {
-    FSTERROR() << "FarCompileStrings: Unknown FST type: " << fst_type;
-    return;
-  }
-  std::unique_ptr<const SymbolTable> syms;
-  typename Arc::Label unknown_label = kNoLabel;
-  if (!symbols_fname.empty()) {
-    const SymbolTableTextOptions opts(allow_negative_labels);
-    syms.reset(SymbolTable::ReadText(symbols_fname, opts));
-    if (!syms) {
-      LOG(ERROR) << "FarCompileStrings: Error reading symbol table: "
-                 << symbols_fname;
-      return;
-    }
-    if (!unknown_symbol.empty()) {
-      unknown_label = syms->Find(unknown_symbol);
-      if (unknown_label == kNoLabel) {
-        FSTERROR() << "FarCompileStrings: Label \"" << unknown_label
-                   << "\" missing from symbol table: " << symbols_fname;
-        return;
-      }
-    }
-  }
-  std::unique_ptr<FarWriter<Arc>> far_writer(
-      FarWriter<Arc>::Create(out_fname, far_type));
-  if (!far_writer) return;
-  int n = 0;
-  for (const auto &in_fname : in_fnames) {
-    if (generate_keys == 0 && in_fname.empty()) {
-      FSTERROR() << "FarCompileStrings: Read from a file instead of stdin or"
-                 << " set the --generate_keys flags.";
-      return;
-    }
-    int key_size =
-        generate_keys ? generate_keys : (entry_type == StringReader<Arc>::FILE
-                                             ? 1 : KeySize(in_fname.c_str()));
-    std::ifstream fstrm;
-    if (!in_fname.empty()) {
-      fstrm.open(in_fname);
-      if (!fstrm) {
-        FSTERROR() << "FarCompileStrings: Can't open file: " << in_fname;
-        return;
-      }
-    }
-    std::istream &istrm = fstrm.is_open() ? fstrm : std::cin;
-    bool keep_syms = keep_symbols;
-    for (StringReader<Arc> reader(
-             istrm, in_fname.empty() ? "stdin" : in_fname, entry_type,
-             token_type, allow_negative_labels, syms.get(), unknown_label);
-         !reader.Done(); reader.Next()) {
-      ++n;
-      std::unique_ptr<const Fst<Arc>> fst;
-      if (compact) {
-        fst.reset(reader.GetCompactFst(keep_syms));
-      } else {
-        fst.reset(reader.GetVectorFst(keep_syms));
-      }
-      if (initial_symbols) keep_syms = false;
-      if (!fst) {
-        FSTERROR() << "FarCompileStrings: Compiling string number " << n
-                   << " in file " << in_fname << " failed with token_type = "
-                   << (tt == FTT_BYTE
-                           ? "byte"
-                           : (tt == FTT_UTF8
-                                  ? "utf8"
-                                  : (tt == FTT_SYMBOL ? "symbol" : "unknown")))
-                   << " and entry_type = "
-                   << (fet == FET_LINE
-                           ? "line"
-                           : (fet == FET_FILE ? "file" : "unknown"));
-        return;
-      }
-      std::ostringstream keybuf;
-      keybuf.width(key_size);
-      keybuf.fill('0');
-      keybuf << n;
-      string key;
-      if (generate_keys > 0) {
-        key = keybuf.str();
-      } else {
-        auto *filename = new char[in_fname.size() + 1];
-        strcpy(filename, in_fname.c_str());
-        key = basename(filename);
-        if (entry_type != StringReader<Arc>::FILE) {
-          key += "-";
-          key += keybuf.str();
-        }
-        delete[] filename;
-      }
-      far_writer->Add(key_prefix + key + key_suffix, *fst);
-    }
-    if (generate_keys == 0) n = 0;
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_COMPILE_STRINGS_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/create.h b/projects/llm_framework/include/fst/extensions/far/create.h
deleted file mode 100644
index 31c95896..00000000
--- a/projects/llm_framework/include/fst/extensions/far/create.h
+++ /dev/null
@@ -1,46 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Creates a finite-state archive from component FSTs.
-
-#ifndef FST_EXTENSIONS_FAR_CREATE_H_
-#define FST_EXTENSIONS_FAR_CREATE_H_
-
-#include <sstream>
-#include <string>
-#include <vector>
-
-#include <fst/extensions/far/far.h>
-
-namespace fst {
-
-template <class Arc>
-void FarCreate(const std::vector<string> &in_fnames, const string &out_fname,
-               const int32 generate_keys, const FarType &far_type,
-               const string &key_prefix, const string &key_suffix) {
-  std::unique_ptr<FarWriter<Arc>> far_writer(
-      FarWriter<Arc>::Create(out_fname, far_type));
-  if (!far_writer) return;
-  for (size_t i = 0; i < in_fnames.size(); ++i) {
-    std::unique_ptr<Fst<Arc>> ifst(Fst<Arc>::Read(in_fnames[i]));
-    if (!ifst) return;
-    string key;
-    if (generate_keys > 0) {
-      std::ostringstream keybuf;
-      keybuf.width(generate_keys);
-      keybuf.fill('0');
-      keybuf << i + 1;
-      key = keybuf.str();
-    } else {
-      auto *filename = new char[in_fnames[i].size() + 1];
-      strcpy(filename, in_fnames[i].c_str());
-      key = basename(filename);
-      delete[] filename;
-    }
-    far_writer->Add(key_prefix + key + key_suffix, *ifst);
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_CREATE_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/equal.h b/projects/llm_framework/include/fst/extensions/far/equal.h
deleted file mode 100644
index 04511d21..00000000
--- a/projects/llm_framework/include/fst/extensions/far/equal.h
+++ /dev/null
@@ -1,69 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_FAR_EQUAL_H_
-#define FST_EXTENSIONS_FAR_EQUAL_H_
-
-#include <memory>
-#include <string>
-
-#include <fst/extensions/far/far.h>
-#include <fst/equal.h>
-
-namespace fst {
-
-template <class Arc>
-bool FarEqual(const string &filename1, const string &filename2,
-              float delta = kDelta, const string &begin_key = string(),
-              const string &end_key = string()) {
-  std::unique_ptr<FarReader<Arc>> reader1(FarReader<Arc>::Open(filename1));
-  if (!reader1) {
-    LOG(ERROR) << "FarEqual: Could not open FAR file " << filename1;
-    return false;
-  }
-  std::unique_ptr<FarReader<Arc>> reader2(FarReader<Arc>::Open(filename2));
-  if (!reader2) {
-    LOG(ERROR) << "FarEqual: Could not open FAR file " << filename2;
-    return false;
-  }
-  if (!begin_key.empty()) {
-    bool find_begin1 = reader1->Find(begin_key);
-    bool find_begin2 = reader2->Find(begin_key);
-    if (!find_begin1 || !find_begin2) {
-      bool ret = !find_begin1 && !find_begin2;
-      if (!ret) {
-        LOG(ERROR) << "FarEqual: Key " << begin_key << " missing from "
-                   << (find_begin1 ? "second" : "first") << " archive";
-      }
-      return ret;
-    }
-  }
-  for (; !reader1->Done() && !reader2->Done();
-       reader1->Next(), reader2->Next()) {
-    const auto &key1 = reader1->GetKey();
-    const auto &key2 = reader2->GetKey();
-    if (!end_key.empty() && end_key < key1 && end_key < key2) {
-      return true;
-    }
-    if (key1 != key2) {
-      LOG(ERROR) << "FarEqual: Mismatched keys " << key1 << " and " << key2;
-      return false;
-    }
-    if (!Equal(*(reader1->GetFst()), *(reader2->GetFst()), delta)) {
-      LOG(ERROR) << "FarEqual: FSTs for key " << key1 << " are not equal";
-      return false;
-    }
-  }
-  if (!reader1->Done() || !reader2->Done()) {
-    LOG(ERROR) << "FarEqual: Key "
-               << (reader1->Done() ? reader2->GetKey() : reader1->GetKey())
-               << " missing from " << (reader2->Done() ? "first" : "second")
-               << " archive";
-    return false;
-  }
-  return true;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_EQUAL_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/extract.h b/projects/llm_framework/include/fst/extensions/far/extract.h
deleted file mode 100644
index e707835f..00000000
--- a/projects/llm_framework/include/fst/extensions/far/extract.h
+++ /dev/null
@@ -1,118 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Extracts component FSTs from an finite-state archive.
-
-#ifndef FST_EXTENSIONS_FAR_EXTRACT_H_
-#define FST_EXTENSIONS_FAR_EXTRACT_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include <fst/extensions/far/far.h>
-#include <fst/util.h>
-
-namespace fst {
-
-template <class Arc>
-inline void FarWriteFst(const Fst<Arc> *fst, string key, string *okey,
-                        int *nrep, int32 generate_filenames, int i,
-                        const string &filename_prefix,
-                        const string &filename_suffix) {
-  if (key == *okey) {
-    ++*nrep;
-  } else {
-    *nrep = 0;
-  }
-  *okey = key;
-  string ofilename;
-  if (generate_filenames) {
-    std::ostringstream tmp;
-    tmp.width(generate_filenames);
-    tmp.fill('0');
-    tmp << i;
-    ofilename = tmp.str();
-  } else {
-    if (*nrep > 0) {
-      std::ostringstream tmp;
-      tmp << '.' << nrep;
-      key.append(tmp.str().data(), tmp.str().size());
-    }
-    ofilename = key;
-  }
-  fst->Write(filename_prefix + ofilename + filename_suffix);
-}
-
-template <class Arc>
-void FarExtract(const std::vector<string> &ifilenames, int32 generate_filenames,
-                const string &keys, const string &key_separator,
-                const string &range_delimiter, const string &filename_prefix,
-                const string &filename_suffix) {
-  std::unique_ptr<FarReader<Arc>> far_reader(
-      FarReader<Arc>::Open(ifilenames));
-  if (!far_reader) return;
-  string okey;
-  int nrep = 0;
-  std::vector<char *> key_vector;
-  // User has specified a set of FSTs to extract, where some of these may in
-  // fact be ranges.
-  if (!keys.empty()) {
-    auto *keys_cstr = new char[keys.size() + 1];
-    strcpy(keys_cstr, keys.c_str());
-    SplitString(keys_cstr, key_separator.c_str(), &key_vector, true);
-    int i = 0;
-    for (size_t k = 0; k < key_vector.size(); ++k, ++i) {
-      string key = key_vector[k];
-      auto *key_cstr = new char[key.size() + 1];
-      strcpy(key_cstr, key.c_str());
-      std::vector<char *> range_vector;
-      SplitString(key_cstr, range_delimiter.c_str(), &range_vector, false);
-      if (range_vector.size() == 1) {  // Not a range
-        if (!far_reader->Find(key)) {
-          LOG(ERROR) << "FarExtract: Cannot find key " << key;
-          return;
-        }
-        const auto *fst = far_reader->GetFst();
-        FarWriteFst(fst, key, &okey, &nrep, generate_filenames, i,
-                    filename_prefix, filename_suffix);
-      } else if (range_vector.size() == 2) {  // A legal range
-        string begin_key = range_vector[0];
-        string end_key = range_vector[1];
-        if (begin_key.empty() || end_key.empty()) {
-          LOG(ERROR) << "FarExtract: Illegal range specification " << key;
-          return;
-        }
-        if (!far_reader->Find(begin_key)) {
-          LOG(ERROR) << "FarExtract: Cannot find key " << begin_key;
-          return;
-        }
-        for (; !far_reader->Done(); far_reader->Next(), ++i) {
-          const auto &ikey = far_reader->GetKey();
-          if (end_key < ikey) break;
-          const auto *fst = far_reader->GetFst();
-          FarWriteFst(fst, ikey, &okey, &nrep, generate_filenames, i,
-                      filename_prefix, filename_suffix);
-        }
-      } else {
-        LOG(ERROR) << "FarExtract: Illegal range specification " << key;
-        return;
-      }
-      delete[] key_cstr;
-    }
-    delete[] keys_cstr;
-    return;
-  }
-  // Nothing specified, so just extracts everything.
-  for (size_t i = 1; !far_reader->Done(); far_reader->Next(), ++i) {
-    const auto &key = far_reader->GetKey();
-    const auto *fst = far_reader->GetFst();
-    FarWriteFst(fst, key, &okey, &nrep, generate_filenames, i, filename_prefix,
-                filename_suffix);
-  }
-  return;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_EXTRACT_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/far-class.h b/projects/llm_framework/include/fst/extensions/far/far-class.h
deleted file mode 100644
index 50d7be9b..00000000
--- a/projects/llm_framework/include/fst/extensions/far/far-class.h
+++ /dev/null
@@ -1,258 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Scripting API support for FarReader and FarWriter.
-
-#ifndef FST_EXTENSIONS_FAR_FAR_CLASS_H_
-#define FST_EXTENSIONS_FAR_FAR_CLASS_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include <fst/extensions/far/far.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fstscript.h>
-
-namespace fst {
-namespace script {
-
-
-// FarReader API.
-
-// Virtual interface implemented by each concrete FarReaderImpl<A>.
-// See the FarReader interface in far.h for the exact semantics.
-class FarReaderImplBase {
- public:
-  virtual const string &ArcType() const = 0;
-  virtual bool Done() const = 0;
-  virtual bool Error() const = 0;
-  virtual const string &GetKey() const = 0;
-  virtual const FstClass *GetFstClass() const = 0;
-  virtual bool Find(const string &key) = 0;
-  virtual void Next() = 0;
-  virtual void Reset() = 0;
-  virtual FarType Type() const = 0;
-  virtual ~FarReaderImplBase() {}
-};
-
-// Templated implementation.
-template <class Arc>
-class FarReaderClassImpl : public FarReaderImplBase {
- public:
-  explicit FarReaderClassImpl(const string &filename)
-      : impl_(FarReader<Arc>::Open(filename)) {}
-
-  explicit FarReaderClassImpl(const std::vector<string> &filenames)
-      : impl_(FarReader<Arc>::Open(filenames)) {}
-
-  const string &ArcType() const final { return Arc::Type(); }
-
-  bool Done() const final { return impl_->Done(); }
-
-  bool Error() const final { return impl_->Error(); }
-
-  bool Find(const string &key) final { return impl_->Find(key); }
-
-  const FstClass *GetFstClass() const final {
-    fstc_.reset(new FstClass(*impl_->GetFst()));
-    return fstc_.get();
-  }
-
-  const string &GetKey() const final { return impl_->GetKey(); }
-
-  void Next() final { return impl_->Next(); }
-
-  void Reset() final { impl_->Reset(); }
-
-  FarType Type() const final { return impl_->Type(); }
-
-  const FarReader<Arc> *GetImpl() const { return impl_.get(); }
-
-  FarReader<Arc> *GetImpl() { return impl_.get(); }
-
- private:
-  std::unique_ptr<FarReader<Arc>> impl_;
-  mutable std::unique_ptr<FstClass> fstc_;
-};
-
-
-class FarReaderClass;
-
-using OpenFarReaderClassArgs =
-    WithReturnValue<FarReaderClass *, const std::vector<string> &>;
-
-// Untemplated user-facing class holding a templated pimpl.
-class FarReaderClass {
- public:
-  const string &ArcType() const { return impl_->ArcType(); }
-
-  bool Done() const { return impl_->Done(); }
-
-  // Returns True if the impl is null (i.e., due to read failure).
-  // Attempting to call any other function will result in null dereference.
-  bool Error() const { return (impl_) ? impl_->Error() : true; }
-
-  bool Find(const string &key) { return impl_->Find(key); }
-
-  const FstClass *GetFstClass() const { return impl_->GetFstClass(); }
-
-  const string &GetKey() const { return impl_->GetKey(); }
-
-  void Next() { impl_->Next(); }
-
-  void Reset() { impl_->Reset(); }
-
-  FarType Type() const { return impl_->Type(); }
-
-  template <class Arc>
-  const FarReader<Arc> *GetFarReader() const {
-    if (Arc::Type() != ArcType()) return nullptr;
-    const FarReaderClassImpl<Arc> *typed_impl =
-        static_cast<FarReaderClassImpl<Arc> *>(impl_.get());
-    return typed_impl->GetImpl();
-  }
-
-  template <class Arc>
-  FarReader<Arc> *GetFarReader() {
-    if (Arc::Type() != ArcType()) return nullptr;
-    FarReaderClassImpl<Arc> *typed_impl =
-        static_cast<FarReaderClassImpl<Arc> *>(impl_.get());
-    return typed_impl->GetImpl();
-  }
-
-  template <class Arc>
-  friend void OpenFarReaderClass(OpenFarReaderClassArgs *args);
-
-  // Defined in the CC.
-
-  static FarReaderClass *Open(const string &filename);
-
-  static FarReaderClass *Open(const std::vector<string> &filenames);
-
- private:
-  template <class Arc>
-  explicit FarReaderClass(FarReaderClassImpl<Arc> *impl) : impl_(impl) {}
-
-  std::unique_ptr<FarReaderImplBase> impl_;
-};
-
-// These exist solely for registration purposes; users should call the
-// static method FarReaderClass::Open instead.
-
-template <class Arc>
-void OpenFarReaderClass(OpenFarReaderClassArgs *args) {
-  args->retval = new FarReaderClass(new FarReaderClassImpl<Arc>(args->args));
-}
-
-// FarWriter API.
-
-// Virtual interface implemented by each concrete FarWriterImpl<A>.
-class FarWriterImplBase {
- public:
-  // Unlike the lower-level library, this returns a boolean to signal failure
-  // due to non-conformant arc types.
-  virtual bool Add(const string &key, const FstClass &fst) = 0;
-  virtual const string &ArcType() const = 0;
-  virtual bool Error() const = 0;
-  virtual FarType Type() const = 0;
-  virtual ~FarWriterImplBase() {}
-};
-
-
-// Templated implementation.
-template <class Arc>
-class FarWriterClassImpl : public FarWriterImplBase {
- public:
-  explicit FarWriterClassImpl(const string &filename,
-                              FarType type = FAR_DEFAULT)
-      : impl_(FarWriter<Arc>::Create(filename, type)) {}
-
-  bool Add(const string &key, const FstClass &fst) final {
-    if (ArcType() != fst.ArcType()) {
-      FSTERROR() << "Cannot write FST with " << fst.ArcType() << " arcs to "
-                 << "FAR with " << ArcType() << " arcs";
-      return false;
-    }
-    impl_->Add(key, *(fst.GetFst<Arc>()));
-    return true;
-  }
-
-  const string &ArcType() const final { return Arc::Type(); }
-
-  bool Error() const final { return impl_->Error(); }
-
-  FarType Type() const final { return impl_->Type(); }
-
-  const FarWriter<Arc> *GetImpl() const { return impl_.get(); }
-
-  FarWriter<Arc> *GetImpl() { return impl_.get(); }
-
- private:
-  std::unique_ptr<FarWriter<Arc>> impl_;
-};
-
-
-class FarWriterClass;
-
-using CreateFarWriterClassInnerArgs = std::pair<const string &, FarType>;
-
-using CreateFarWriterClassArgs =
-    WithReturnValue<FarWriterClass *, CreateFarWriterClassInnerArgs>;
-
-// Untemplated user-facing class holding a templated pimpl.
-class FarWriterClass {
- public:
-  static FarWriterClass *Create(const string &filename, const string &arc_type,
-                                FarType type = FAR_DEFAULT);
-
-  bool Add(const string &key, const FstClass &fst) {
-    return impl_->Add(key, fst);
-  }
-
-  // Returns True if the impl is null (i.e., due to construction failure).
-  // Attempting to call any other function will result in null dereference.
-  bool Error() const { return (impl_) ? impl_->Error() : true; }
-
-  const string &ArcType() const { return impl_->ArcType(); }
-
-  FarType Type() const { return impl_->Type(); }
-
-  template <class Arc>
-  const FarWriter<Arc> *GetFarWriter() const {
-    if (Arc::Type() != ArcType()) return nullptr;
-    const FarWriterClassImpl<Arc> *typed_impl =
-        static_cast<FarWriterClassImpl<Arc> *>(impl_.get());
-    return typed_impl->GetImpl();
-  }
-
-  template <class Arc>
-  FarWriter<Arc> *GetFarWriter() {
-    if (Arc::Type() != ArcType()) return nullptr;
-    FarWriterClassImpl<Arc> *typed_impl =
-        static_cast<FarWriterClassImpl<Arc> *>(impl_.get());
-    return typed_impl->GetImpl();
-  }
-
-  template <class Arc>
-  friend void CreateFarWriterClass(CreateFarWriterClassArgs *args);
-
- private:
-  template <class Arc>
-  explicit FarWriterClass(FarWriterClassImpl<Arc> *impl) : impl_(impl) {}
-
-  std::unique_ptr<FarWriterImplBase> impl_;
-};
-
-// This exists solely for registration purposes; users should call the
-// static method FarWriterClass::Create instead.
-template <class Arc>
-void CreateFarWriterClass(CreateFarWriterClassArgs *args) {
-  args->retval = new FarWriterClass(new FarWriterClassImpl<Arc>(
-      std::get<0>(args->args), std::get<1>(args->args)));
-}
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_FAR_CLASS_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/far.h b/projects/llm_framework/include/fst/extensions/far/far.h
deleted file mode 100644
index c24c7dab..00000000
--- a/projects/llm_framework/include/fst/extensions/far/far.h
+++ /dev/null
@@ -1,481 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Finite-State Transducer (FST) archive classes.
-
-#ifndef FST_EXTENSIONS_FAR_FAR_H_
-#define FST_EXTENSIONS_FAR_FAR_H_
-
-#include <iostream>
-#include <sstream>
-
-#include <fst/log.h>
-#include <fst/extensions/far/stlist.h>
-#include <fst/extensions/far/sttable.h>
-#include <fst/fst.h>
-#include <fst/vector-fst.h>
-#include <fstream>
-
-namespace fst {
-
-enum FarEntryType { FET_LINE, FET_FILE };
-
-enum FarTokenType { FTT_SYMBOL, FTT_BYTE, FTT_UTF8 };
-
-inline bool IsFst(const string &filename) {
-  std::ifstream strm(filename, std::ios_base::in | std::ios_base::binary);
-  if (!strm) return false;
-  return IsFstHeader(strm, filename);
-}
-
-// FST archive header class
-class FarHeader {
- public:
-  const string &ArcType() const { return arctype_; }
-
-  const string &FarType() const { return fartype_; }
-
-  bool Read(const string &filename) {
-    FstHeader fsthdr;
-    if (filename.empty()) {
-      // Header reading unsupported on stdin. Assumes STList and StdArc.
-      fartype_ = "stlist";
-      arctype_ = "standard";
-      return true;
-    } else if (IsSTTable(filename)) {  // Checks if STTable.
-      ReadSTTableHeader(filename, &fsthdr);
-      fartype_ = "sttable";
-      arctype_ = fsthdr.ArcType().empty() ? "unknown" : fsthdr.ArcType();
-      return true;
-    } else if (IsSTList(filename)) {  // Checks if STList.
-      ReadSTListHeader(filename, &fsthdr);
-      fartype_ = "stlist";
-      arctype_ = fsthdr.ArcType().empty() ? "unknown" : fsthdr.ArcType();
-      return true;
-    } else if (IsFst(filename)) {  // Checks if FST.
-      std::ifstream istrm(filename,
-                               std::ios_base::in | std::ios_base::binary);
-      fsthdr.Read(istrm, filename);
-      fartype_ = "fst";
-      arctype_ = fsthdr.ArcType().empty() ? "unknown" : fsthdr.ArcType();
-      return true;
-    }
-    return false;
-  }
-
- private:
-  string fartype_;
-  string arctype_;
-};
-
-enum FarType {
-  FAR_DEFAULT = 0,
-  FAR_STTABLE = 1,
-  FAR_STLIST = 2,
-  FAR_FST = 3,
-};
-
-// This class creates an archive of FSTs.
-template <class A>
-class FarWriter {
- public:
-  using Arc = A;
-
-  // Creates a new (empty) FST archive; returns null on error.
-  static FarWriter *Create(const string &filename, FarType type = FAR_DEFAULT);
-
-  // Adds an FST to the end of an archive. Keys must be non-empty and
-  // in lexicographic order. FSTs must have a suitable write method.
-  virtual void Add(const string &key, const Fst<Arc> &fst) = 0;
-
-  virtual FarType Type() const = 0;
-
-  virtual bool Error() const = 0;
-
-  virtual ~FarWriter() {}
-
- protected:
-  FarWriter() {}
-};
-
-// This class iterates through an existing archive of FSTs.
-template <class A>
-class FarReader {
- public:
-  using Arc = A;
-
-  // Opens an existing FST archive in a single file; returns null on error.
-  // Sets current position to the beginning of the achive.
-  static FarReader *Open(const string &filename);
-
-  // Opens an existing FST archive in multiple files; returns null on error.
-  // Sets current position to the beginning of the achive.
-  static FarReader *Open(const std::vector<string> &filenames);
-
-  // Resets current position to beginning of archive.
-  virtual void Reset() = 0;
-
-  // Sets current position to first entry >= key.  Returns true if a match.
-  virtual bool Find(const string &key) = 0;
-
-  // Current position at end of archive?
-  virtual bool Done() const = 0;
-
-  // Move current position to next FST.
-  virtual void Next() = 0;
-
-  // Returns key at the current position. This reference is invalidated if
-  // the current position in the archive is changed.
-  virtual const string &GetKey() const = 0;
-
-  // Returns pointer to FST at the current position. This is invalidated if
-  // the current position in the archive is changed.
-  virtual const Fst<Arc> *GetFst() const = 0;
-
-  virtual FarType Type() const = 0;
-
-  virtual bool Error() const = 0;
-
-  virtual ~FarReader() {}
-
- protected:
-  FarReader() {}
-};
-
-template <class Arc>
-class FstWriter {
- public:
-  void operator()(std::ostream &strm, const Fst<Arc> &fst) const {
-    fst.Write(strm, FstWriteOptions());
-  }
-};
-
-template <class A>
-class STTableFarWriter : public FarWriter<A> {
- public:
-  using Arc = A;
-
-  static STTableFarWriter *Create(const string &filename) {
-    auto *writer = STTableWriter<Fst<Arc>, FstWriter<Arc>>::Create(filename);
-    return new STTableFarWriter(writer);
-  }
-
-  void Add(const string &key, const Fst<Arc> &fst) final {
-    writer_->Add(key, fst);
-  }
-
-  FarType Type() const final { return FAR_STTABLE; }
-
-  bool Error() const final { return writer_->Error(); }
-
- private:
-  explicit STTableFarWriter(STTableWriter<Fst<Arc>, FstWriter<Arc>> *writer)
-      : writer_(writer) {}
-
-  std::unique_ptr<STTableWriter<Fst<Arc>, FstWriter<Arc>>> writer_;
-};
-
-template <class A>
-class STListFarWriter : public FarWriter<A> {
- public:
-  using Arc = A;
-
-  static STListFarWriter *Create(const string &filename) {
-    auto *writer = STListWriter<Fst<Arc>, FstWriter<Arc>>::Create(filename);
-    return new STListFarWriter(writer);
-  }
-
-  void Add(const string &key, const Fst<Arc> &fst) final {
-    writer_->Add(key, fst);
-  }
-
-  constexpr FarType Type() const final { return FAR_STLIST; }
-
-  bool Error() const final { return writer_->Error(); }
-
- private:
-  explicit STListFarWriter(STListWriter<Fst<Arc>, FstWriter<Arc>> *writer)
-      : writer_(writer) {}
-
-  std::unique_ptr<STListWriter<Fst<Arc>, FstWriter<Arc>>> writer_;
-};
-
-template <class A>
-class FstFarWriter : public FarWriter<A> {
- public:
-  using Arc = A;
-
-  explicit FstFarWriter(const string &filename)
-      : filename_(filename), error_(false), written_(false) {}
-
-  static FstFarWriter *Create(const string &filename) {
-    return new FstFarWriter(filename);
-  }
-
-  void Add(const string &key, const Fst<A> &fst) final {
-    if (written_) {
-      LOG(WARNING) << "FstFarWriter::Add: only one FST supported,"
-                   << " subsequent entries discarded.";
-    } else {
-      error_ = !fst.Write(filename_);
-      written_ = true;
-    }
-  }
-
-  constexpr FarType Type() const final { return FAR_FST; }
-
-  bool Error() const final { return error_; }
-
-  ~FstFarWriter() final {}
-
- private:
-  string filename_;
-  bool error_;
-  bool written_;
-};
-
-template <class Arc>
-FarWriter<Arc> *FarWriter<Arc>::Create(const string &filename, FarType type) {
-  switch (type) {
-    case FAR_DEFAULT:
-      if (filename.empty()) return STListFarWriter<Arc>::Create(filename);
-    case FAR_STTABLE:
-      return STTableFarWriter<Arc>::Create(filename);
-    case FAR_STLIST:
-      return STListFarWriter<Arc>::Create(filename);
-    case FAR_FST:
-      return FstFarWriter<Arc>::Create(filename);
-    default:
-      LOG(ERROR) << "FarWriter::Create: Unknown FAR type";
-      return nullptr;
-  }
-}
-
-template <class Arc>
-class FstReader {
- public:
-  Fst<Arc> *operator()(std::istream &strm) const {
-    return Fst<Arc>::Read(strm, FstReadOptions());
-  }
-};
-
-template <class A>
-class STTableFarReader : public FarReader<A> {
- public:
-  using Arc = A;
-
-  static STTableFarReader *Open(const string &filename) {
-    auto *reader = STTableReader<Fst<Arc>, FstReader<Arc>>::Open(filename);
-    if (!reader || reader->Error()) return nullptr;
-    return new STTableFarReader(reader);
-  }
-
-  static STTableFarReader *Open(const std::vector<string> &filenames) {
-    auto *reader = STTableReader<Fst<Arc>, FstReader<Arc>>::Open(filenames);
-    if (!reader || reader->Error()) return nullptr;
-    return new STTableFarReader(reader);
-  }
-
-  void Reset() final { reader_->Reset(); }
-
-  bool Find(const string &key) final { return reader_->Find(key); }
-
-  bool Done() const final { return reader_->Done(); }
-
-  void Next() final { return reader_->Next(); }
-
-  const string &GetKey() const final { return reader_->GetKey(); }
-
-  const Fst<Arc> *GetFst() const final { return reader_->GetEntry(); }
-
-  constexpr FarType Type() const final { return FAR_STTABLE; }
-
-  bool Error() const final { return reader_->Error(); }
-
- private:
-  explicit STTableFarReader(STTableReader<Fst<Arc>, FstReader<Arc>> *reader)
-      : reader_(reader) {}
-
-  std::unique_ptr<STTableReader<Fst<Arc>, FstReader<Arc>>> reader_;
-};
-
-template <class A>
-class STListFarReader : public FarReader<A> {
- public:
-  using Arc = A;
-
-  static STListFarReader *Open(const string &filename) {
-    auto *reader = STListReader<Fst<Arc>, FstReader<Arc>>::Open(filename);
-    if (!reader || reader->Error()) return nullptr;
-    return new STListFarReader(reader);
-  }
-
-  static STListFarReader *Open(const std::vector<string> &filenames) {
-    auto *reader = STListReader<Fst<Arc>, FstReader<Arc>>::Open(filenames);
-    if (!reader || reader->Error()) return nullptr;
-    return new STListFarReader(reader);
-  }
-
-  void Reset() final { reader_->Reset(); }
-
-  bool Find(const string &key) final { return reader_->Find(key); }
-
-  bool Done() const final { return reader_->Done(); }
-
-  void Next() final { return reader_->Next(); }
-
-  const string &GetKey() const final { return reader_->GetKey(); }
-
-  const Fst<Arc> *GetFst() const final { return reader_->GetEntry(); }
-
-  constexpr FarType Type() const final { return FAR_STLIST; }
-
-  bool Error() const final { return reader_->Error(); }
-
- private:
-  explicit STListFarReader(STListReader<Fst<Arc>, FstReader<Arc>> *reader)
-      : reader_(reader) {}
-
-  std::unique_ptr<STListReader<Fst<Arc>, FstReader<Arc>>> reader_;
-};
-
-template <class A>
-class FstFarReader : public FarReader<A> {
- public:
-  using Arc = A;
-
-  static FstFarReader *Open(const string &filename) {
-    std::vector<string> filenames;
-    filenames.push_back(filename);
-    return new FstFarReader<Arc>(filenames);
-  }
-
-  static FstFarReader *Open(const std::vector<string> &filenames) {
-    return new FstFarReader<Arc>(filenames);
-  }
-
-  explicit FstFarReader(const std::vector<string> &filenames)
-      : keys_(filenames), has_stdin_(false), pos_(0), error_(false) {
-    std::sort(keys_.begin(), keys_.end());
-    streams_.resize(keys_.size(), 0);
-    for (size_t i = 0; i < keys_.size(); ++i) {
-      if (keys_[i].empty()) {
-        if (!has_stdin_) {
-          streams_[i] = &std::cin;
-          has_stdin_ = true;
-        } else {
-          FSTERROR() << "FstFarReader::FstFarReader: standard input should "
-                        "only appear once in the input file list";
-          error_ = true;
-          return;
-        }
-      } else {
-        streams_[i] = new std::ifstream(
-            keys_[i], std::ios_base::in | std::ios_base::binary);
-        if (streams_[i]->fail()) {
-          FSTERROR() << "FstFarReader::FstFarReader: Error reading file: "
-                     << filenames[i];
-          error_ = true;
-          return;
-        }
-      }
-    }
-    if (pos_ >= keys_.size()) return;
-    ReadFst();
-  }
-
-  void Reset() final {
-    if (has_stdin_) {
-      FSTERROR()
-          << "FstFarReader::Reset: Operation not supported on standard input";
-      error_ = true;
-      return;
-    }
-    pos_ = 0;
-    ReadFst();
-  }
-
-  bool Find(const string &key) final {
-    if (has_stdin_) {
-      FSTERROR()
-          << "FstFarReader::Find: Operation not supported on standard input";
-      error_ = true;
-      return false;
-    }
-    pos_ = 0;  // TODO
-    ReadFst();
-    return true;
-  }
-
-  bool Done() const final { return error_ || pos_ >= keys_.size(); }
-
-  void Next() final {
-    ++pos_;
-    ReadFst();
-  }
-
-  const string &GetKey() const final { return keys_[pos_]; }
-
-  const Fst<Arc> *GetFst() const final { return fst_.get(); }
-
-  constexpr FarType Type() const final { return FAR_FST; }
-
-  bool Error() const final { return error_; }
-
-  ~FstFarReader() final {
-    for (size_t i = 0; i < keys_.size(); ++i) {
-      if (streams_[i] != &std::cin) {
-        delete streams_[i];
-      }
-    }
-  }
-
- private:
-  void ReadFst() {
-    fst_.reset();
-    if (pos_ >= keys_.size()) return;
-    streams_[pos_]->seekg(0);
-    fst_.reset(Fst<Arc>::Read(*streams_[pos_], FstReadOptions()));
-    if (!fst_) {
-      FSTERROR() << "FstFarReader: Error reading Fst from: " << keys_[pos_];
-      error_ = true;
-    }
-  }
-
-  std::vector<string> keys_;
-  std::vector<std::istream *> streams_;
-  bool has_stdin_;
-  size_t pos_;
-  mutable std::unique_ptr<Fst<Arc>> fst_;
-  mutable bool error_;
-};
-
-template <class Arc>
-FarReader<Arc> *FarReader<Arc>::Open(const string &filename) {
-  if (filename.empty())
-    return STListFarReader<Arc>::Open(filename);
-  else if (IsSTTable(filename))
-    return STTableFarReader<Arc>::Open(filename);
-  else if (IsSTList(filename))
-    return STListFarReader<Arc>::Open(filename);
-  else if (IsFst(filename))
-    return FstFarReader<Arc>::Open(filename);
-  return nullptr;
-}
-
-template <class Arc>
-FarReader<Arc> *FarReader<Arc>::Open(const std::vector<string> &filenames) {
-  if (!filenames.empty() && filenames[0].empty())
-    return STListFarReader<Arc>::Open(filenames);
-  else if (!filenames.empty() && IsSTTable(filenames[0]))
-    return STTableFarReader<Arc>::Open(filenames);
-  else if (!filenames.empty() && IsSTList(filenames[0]))
-    return STListFarReader<Arc>::Open(filenames);
-  else if (!filenames.empty() && IsFst(filenames[0]))
-    return FstFarReader<Arc>::Open(filenames);
-  return nullptr;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_FAR_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/farlib.h b/projects/llm_framework/include/fst/extensions/far/farlib.h
deleted file mode 100644
index c9bb1710..00000000
--- a/projects/llm_framework/include/fst/extensions/far/farlib.h
+++ /dev/null
@@ -1,19 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// A finite-state archive (FAR) is used to store an indexable collection of
-// FSTs in a single file. Utilities are provided to create FARs from FSTs,
-// to iterate over FARs, and to extract specific FSTs from FARs.
-
-#ifndef FST_EXTENSIONS_FAR_FARLIB_H_
-#define FST_EXTENSIONS_FAR_FARLIB_H_
-
-#include <fst/extensions/far/compile-strings.h>
-#include <fst/extensions/far/create.h>
-#include <fst/extensions/far/extract.h>
-#include <fst/extensions/far/far.h>
-#include <fst/extensions/far/getters.h>
-#include <fst/extensions/far/info.h>
-#include <fst/extensions/far/print-strings.h>
-
-#endif  // FST_EXTENSIONS_FAR_FARLIB_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/farscript.h b/projects/llm_framework/include/fst/extensions/far/farscript.h
deleted file mode 100644
index 4bd11a94..00000000
--- a/projects/llm_framework/include/fst/extensions/far/farscript.h
+++ /dev/null
@@ -1,269 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Convenience file for including all of the FAR operations, or registering
-// them for new arc types.
-
-#ifndef FST_EXTENSIONS_FAR_FARSCRIPT_H_
-#define FST_EXTENSIONS_FAR_FARSCRIPT_H_
-
-#include <string>
-#include <vector>
-
-#include <fst/types.h>
-#include <fst/extensions/far/compile-strings.h>
-#include <fst/extensions/far/create.h>
-#include <fst/extensions/far/equal.h>
-#include <fst/extensions/far/extract.h>
-#include <fst/extensions/far/far.h>
-#include <fst/extensions/far/far-class.h>
-#include <fst/extensions/far/info.h>
-#include <fst/extensions/far/isomorphic.h>
-#include <fst/extensions/far/print-strings.h>
-#include <fst/extensions/far/script-impl.h>
-#include <fst/script/arg-packs.h>
-
-namespace fst {
-namespace script {
-
-// Note: it is safe to pass these strings as references because this struct is
-// only used to pass them deeper in the call graph. Be sure you understand why
-// this is so before using this struct for anything else!
-struct FarCompileStringsArgs {
-  const std::vector<string> &in_fnames;
-  const string &out_fname;
-  const string &fst_type;
-  const FarType &far_type;
-  const int32 generate_keys;
-  const FarEntryType fet;
-  const FarTokenType tt;
-  const string &symbols_fname;
-  const string &unknown_symbol;
-  const bool keep_symbols;
-  const bool initial_symbols;
-  const bool allow_negative_labels;
-  const string &key_prefix;
-  const string &key_suffix;
-
-  FarCompileStringsArgs(const std::vector<string> &in_fnames,
-                        const string &out_fname, const string &fst_type,
-                        const FarType &far_type, int32 generate_keys,
-                        FarEntryType fet, FarTokenType tt,
-                        const string &symbols_fname,
-                        const string &unknown_symbol, bool keep_symbols,
-                        bool initial_symbols, bool allow_negative_labels,
-                        const string &key_prefix, const string &key_suffix)
-      : in_fnames(in_fnames),
-        out_fname(out_fname),
-        fst_type(fst_type),
-        far_type(far_type),
-        generate_keys(generate_keys),
-        fet(fet),
-        tt(tt),
-        symbols_fname(symbols_fname),
-        unknown_symbol(unknown_symbol),
-        keep_symbols(keep_symbols),
-        initial_symbols(initial_symbols),
-        allow_negative_labels(allow_negative_labels),
-        key_prefix(key_prefix),
-        key_suffix(key_suffix) {}
-};
-
-template <class Arc>
-void FarCompileStrings(FarCompileStringsArgs *args) {
-  FarCompileStrings<Arc>(
-      args->in_fnames, args->out_fname, args->fst_type, args->far_type,
-      args->generate_keys, args->fet, args->tt, args->symbols_fname,
-      args->unknown_symbol, args->keep_symbols, args->initial_symbols,
-      args->allow_negative_labels, args->key_prefix, args->key_suffix);
-}
-
-void FarCompileStrings(const std::vector<string> &in_fnames,
-                       const string &out_fname, const string &arc_type,
-                       const string &fst_type, const FarType &far_type,
-                       int32 generate_keys, FarEntryType fet, FarTokenType tt,
-                       const string &symbols_fname,
-                       const string &unknown_symbol, bool keep_symbols,
-                       bool initial_symbols, bool allow_negative_labels,
-                       const string &key_prefix, const string &key_suffix);
-
-// Note: it is safe to pass these strings as references because this struct is
-// only used to pass them deeper in the call graph. Be sure you understand why
-// this is so before using this struct for anything else!
-struct FarCreateArgs {
-  const std::vector<string> &in_fnames;
-  const string &out_fname;
-  const int32 generate_keys;
-  const FarType &far_type;
-  const string &key_prefix;
-  const string &key_suffix;
-
-  FarCreateArgs(const std::vector<string> &in_fnames, const string &out_fname,
-                const int32 generate_keys, const FarType &far_type,
-                const string &key_prefix, const string &key_suffix)
-      : in_fnames(in_fnames),
-        out_fname(out_fname),
-        generate_keys(generate_keys),
-        far_type(far_type),
-        key_prefix(key_prefix),
-        key_suffix(key_suffix) {}
-};
-
-template <class Arc>
-void FarCreate(FarCreateArgs *args) {
-  FarCreate<Arc>(args->in_fnames, args->out_fname, args->generate_keys,
-                 args->far_type, args->key_prefix, args->key_suffix);
-}
-
-void FarCreate(const std::vector<string> &in_fnames, const string &out_fname,
-               const string &arc_type, const int32 generate_keys,
-               const FarType &far_type, const string &key_prefix,
-               const string &key_suffix);
-
-using FarEqualInnerArgs = std::tuple<const string &, const string &, float,
-                                     const string &, const string &>;
-
-using FarEqualArgs = WithReturnValue<bool, FarEqualInnerArgs>;
-
-template <class Arc>
-void FarEqual(FarEqualArgs *args) {
-  args->retval = fst::FarEqual<Arc>(
-      std::get<0>(args->args), std::get<1>(args->args), std::get<2>(args->args),
-      std::get<3>(args->args), std::get<4>(args->args));
-}
-
-bool FarEqual(const string &filename1, const string &filename2,
-              const string &arc_type, float delta = kDelta,
-              const string &begin_key = string(),
-              const string &end_key = string());
-
-using FarExtractArgs =
-    std::tuple<const std::vector<string> &, int32, const string &,
-               const string &, const string &, const string &, const string &>;
-
-template <class Arc>
-void FarExtract(FarExtractArgs *args) {
-  fst::FarExtract<Arc>(std::get<0>(*args), std::get<1>(*args),
-                           std::get<2>(*args), std::get<3>(*args),
-                           std::get<4>(*args), std::get<5>(*args),
-                           std::get<6>(*args));
-}
-
-void FarExtract(const std::vector<string> &ifilenames, const string &arc_type,
-                int32 generate_filenames, const string &keys,
-                const string &key_separator, const string &range_delimiter,
-                const string &filename_prefix, const string &filename_suffix);
-
-using FarInfoArgs = std::tuple<const std::vector<string> &, const string &,
-                               const string &, const bool>;
-
-template <class Arc>
-void FarInfo(FarInfoArgs *args) {
-  fst::FarInfo<Arc>(std::get<0>(*args), std::get<1>(*args),
-                        std::get<2>(*args), std::get<3>(*args));
-}
-
-void FarInfo(const std::vector<string> &filenames, const string &arc_type,
-             const string &begin_key, const string &end_key,
-             const bool list_fsts);
-
-using GetFarInfoArgs = std::tuple<const std::vector<string> &, const string &,
-                                  const string &, const bool, FarInfoData *>;
-
-template <class Arc>
-void GetFarInfo(GetFarInfoArgs *args) {
-  fst::GetFarInfo<Arc>(std::get<0>(*args), std::get<1>(*args),
-                           std::get<2>(*args), std::get<3>(*args),
-                           std::get<4>(*args));
-}
-
-void GetFarInfo(const std::vector<string> &filenames, const string &arc_type,
-                const string &begin_key, const string &end_key,
-                const bool list_fsts, FarInfoData *);
-
-using FarIsomorphicInnerArgs = std::tuple<const string &, const string &, float,
-                                          const string &, const string &>;
-
-using FarIsomorphicArgs = WithReturnValue<bool, FarIsomorphicInnerArgs>;
-
-template <class Arc>
-void FarIsomorphic(FarIsomorphicArgs *args) {
-  args->retval = fst::FarIsomorphic<Arc>(
-      std::get<0>(args->args), std::get<1>(args->args), std::get<2>(args->args),
-      std::get<3>(args->args), std::get<4>(args->args));
-}
-
-bool FarIsomorphic(const string &filename1, const string &filename2,
-                   const string &arc_type, float delta = kDelta,
-                   const string &begin_key = string(),
-                   const string &end_key = string());
-
-struct FarPrintStringsArgs {
-  const std::vector<string> &ifilenames;
-  const FarEntryType entry_type;
-  const FarTokenType token_type;
-  const string &begin_key;
-  const string &end_key;
-  const bool print_key;
-  const bool print_weight;
-  const string &symbols_fname;
-  const bool initial_symbols;
-  const int32 generate_filenames;
-  const string &filename_prefix;
-  const string &filename_suffix;
-
-  FarPrintStringsArgs(const std::vector<string> &ifilenames,
-                      const FarEntryType entry_type,
-                      const FarTokenType token_type, const string &begin_key,
-                      const string &end_key, const bool print_key,
-                      const bool print_weight, const string &symbols_fname,
-                      const bool initial_symbols,
-                      const int32 generate_filenames,
-                      const string &filename_prefix,
-                      const string &filename_suffix)
-      : ifilenames(ifilenames),
-        entry_type(entry_type),
-        token_type(token_type),
-        begin_key(begin_key),
-        end_key(end_key),
-        print_key(print_key),
-        print_weight(print_weight),
-        symbols_fname(symbols_fname),
-        initial_symbols(initial_symbols),
-        generate_filenames(generate_filenames),
-        filename_prefix(filename_prefix),
-        filename_suffix(filename_suffix) {}
-};
-
-template <class Arc>
-void FarPrintStrings(FarPrintStringsArgs *args) {
-  fst::FarPrintStrings<Arc>(
-      args->ifilenames, args->entry_type, args->token_type, args->begin_key,
-      args->end_key, args->print_key, args->print_weight, args->symbols_fname,
-      args->initial_symbols, args->generate_filenames, args->filename_prefix,
-      args->filename_suffix);
-}
-
-void FarPrintStrings(const std::vector<string> &ifilenames,
-                     const string &arc_type, const FarEntryType entry_type,
-                     const FarTokenType token_type, const string &begin_key,
-                     const string &end_key, const bool print_key,
-                     const bool print_weight, const string &symbols_fname,
-                     const bool initial_symbols, const int32 generate_filenames,
-                     const string &filename_prefix,
-                     const string &filename_suffix);
-
-}  // namespace script
-}  // namespace fst
-
-#define REGISTER_FST_FAR_OPERATIONS(ArcType)                                 \
-  REGISTER_FST_OPERATION(FarCompileStrings, ArcType, FarCompileStringsArgs); \
-  REGISTER_FST_OPERATION(FarCreate, ArcType, FarCreateArgs);                 \
-  REGISTER_FST_OPERATION(FarEqual, ArcType, FarEqualArgs);                   \
-  REGISTER_FST_OPERATION(FarExtract, ArcType, FarExtractArgs);               \
-  REGISTER_FST_OPERATION(FarInfo, ArcType, FarInfoArgs);                     \
-  REGISTER_FST_OPERATION(FarIsomorphic, ArcType, FarIsomorphicArgs);         \
-  REGISTER_FST_OPERATION(FarPrintStrings, ArcType, FarPrintStringsArgs);     \
-  REGISTER_FST_OPERATION(GetFarInfo, ArcType, GetFarInfoArgs)
-
-#endif  // FST_EXTENSIONS_FAR_FARSCRIPT_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/getters.h b/projects/llm_framework/include/fst/extensions/far/getters.h
deleted file mode 100644
index 3dde4194..00000000
--- a/projects/llm_framework/include/fst/extensions/far/getters.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes and functions for registering and invoking FAR main
-// functions that support multiple and extensible arc types.
-
-#ifndef FST_EXTENSIONS_FAR_GETTERS_H_
-#define FST_EXTENSIONS_FAR_GETTERS_H_
-
-#include <fst/flags.h>
-#include <fst/extensions/far/far.h>
-
-namespace fst {
-namespace script {
-
-FarType GetFarType(const string &str);
-
-bool GetFarEntryType(const string &str, FarEntryType *entry_type);
-
-bool GetFarTokenType(const string &str, FarTokenType *token_type);
-
-void ExpandArgs(int argc, char **argv, int *argcp, char ***argvp);
-
-}  // namespace script
-
-string GetFarTypeString(FarType type);
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_GETTERS_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/info.h b/projects/llm_framework/include/fst/extensions/far/info.h
deleted file mode 100644
index 0391c1f4..00000000
--- a/projects/llm_framework/include/fst/extensions/far/info.h
+++ /dev/null
@@ -1,147 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_FAR_INFO_H_
-#define FST_EXTENSIONS_FAR_INFO_H_
-
-#include <iomanip>
-#include <memory>
-#include <set>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fst/extensions/far/far.h>
-#include <fst/extensions/far/getters.h>
-
-namespace fst {
-
-template <class Arc>
-void AccumulateStatesAndArcs(const Fst<Arc> &fst, size_t *nstate, size_t *narc,
-                             size_t *nfinal) {
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done();
-       siter.Next(), ++(*nstate)) {
-    ArcIterator<Fst<Arc>> aiter(fst, siter.Value());
-    for (; !aiter.Done(); aiter.Next(), ++(*narc)) {
-    }
-    if (fst.Final(siter.Value()) != Arc::Weight::Zero()) ++(*nfinal);
-  }
-}
-
-struct KeyInfo {
-  string key;
-  string type;
-  size_t nstate = 0;
-  size_t narc = 0;
-  size_t nfinal = 0;
-};
-
-struct FarInfoData {
-  std::vector<KeyInfo> key_infos;
-  string far_type;
-  string arc_type;
-  size_t nfst = 0;
-  size_t nstate = 0;
-  size_t narc = 0;
-  size_t nfinal = 0;
-  std::set<string> fst_types;
-};
-
-template <class Arc>
-void GetFarInfo(const std::vector<string> &filenames, const string &begin_key,
-                const string &end_key, const bool list_fsts,
-                FarInfoData *far_info) {
-  *far_info = FarInfoData();
-  std::unique_ptr<FarReader<Arc>> reader(FarReader<Arc>::Open(filenames));
-  if (!reader) {
-    LOG(ERROR) << "GetFarInfo: failed to create far reader.";
-    return;
-  }
-  if (!begin_key.empty()) reader->Find(begin_key);
-
-  for (; !reader->Done(); reader->Next()) {
-    const auto &key = reader->GetKey();
-    if (!end_key.empty() && end_key < key) break;
-    ++far_info->nfst;
-    const auto *fst = reader->GetFst();
-    far_info->fst_types.insert(fst->Type());
-    if (list_fsts) {
-      KeyInfo info;
-      info.key = key;
-      info.type = fst->Type();
-      AccumulateStatesAndArcs(*fst, &info.nstate, &info.narc, &info.nfinal);
-      far_info->nstate += info.nstate;
-      far_info->narc += info.narc;
-      far_info->nfinal += info.nfinal;
-      far_info->key_infos.push_back(info);
-    } else {
-      AccumulateStatesAndArcs(*fst, &far_info->nstate, &far_info->narc,
-                              &far_info->nfinal);
-    }
-  }
-  far_info->far_type = GetFarTypeString(reader->Type());
-  far_info->arc_type = Arc::Type();
-}
-
-template <class Arc>
-void FarInfo(const std::vector<string> &filenames, const string &begin_key,
-             const string &end_key, const bool list_fsts) {
-  FarInfoData info;
-  GetFarInfo<Arc>(filenames, begin_key, end_key, list_fsts, &info);
-  if (!list_fsts) {
-    std::cout << std::left << std::setw(50) << "far type" << info.far_type
-              << std::endl;
-    std::cout << std::left << std::setw(50) << "arc type" << Arc::Type()
-              << std::endl;
-    std::cout << std::left << std::setw(50) << "fst type";
-    for (auto iter = info.fst_types.begin(); iter != info.fst_types.end();
-         ++iter) {
-      if (iter != info.fst_types.begin()) std::cout << ",";
-      std::cout << *iter;
-    }
-    std::cout << std::endl;
-    std::cout << std::left << std::setw(50) << "# of FSTs" << info.nfst
-              << std::endl;
-    std::cout << std::left << std::setw(50) << "total # of states"
-              << info.nstate << std::endl;
-    std::cout << std::left << std::setw(50) << "total # of arcs" << info.narc
-              << std::endl;
-    std::cout << std::left << std::setw(50) << "total # of final states"
-              << info.nfinal << std::endl;
-  } else {
-    // FIXME(kbg): Grok, then document this.
-    int wkey = 10;
-    int wtype = 10;
-    int wnstate = 14;
-    int wnarc = 12;
-    int wnfinal = 20;
-    for (const auto &key_info : info.key_infos) {
-      if (key_info.key.size() + 2 > wkey) wkey = key_info.key.size() + 2;
-      if (key_info.type.size() + 2 > wtype) wtype = key_info.type.size() + 2;
-      if (ceil(log10(key_info.nstate)) + 2 > wnstate) {
-        wnstate = ceil(log10(key_info.nstate)) + 2;
-      }
-      if (ceil(log10(key_info.narc)) + 2 > wnarc) {
-        wnarc = ceil(log10(key_info.narc)) + 2;
-      }
-      if (ceil(log10(key_info.nfinal)) + 2 > wnfinal) {
-        wnfinal = ceil(log10(key_info.nfinal)) + 2;
-      }
-    }
-    std::cout << std::left << std::setw(wkey) << "key" << std::setw(wtype)
-              << "type" << std::right << std::setw(wnstate) << "# of states"
-              << std::setw(wnarc) << "# of arcs" << std::setw(wnfinal)
-              << "# of final states" << std::endl;
-    for (const auto &key_info : info.key_infos) {
-      std::cout << std::left << std::setw(wkey) << key_info.key
-                << std::setw(wtype) << key_info.type << std::right
-                << std::setw(wnstate) << key_info.nstate << std::setw(wnarc)
-                << key_info.narc << std::setw(wnfinal) << key_info.nfinal
-                << std::endl;
-    }
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_INFO_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/isomorphic.h b/projects/llm_framework/include/fst/extensions/far/isomorphic.h
deleted file mode 100644
index 1e6e9cb3..00000000
--- a/projects/llm_framework/include/fst/extensions/far/isomorphic.h
+++ /dev/null
@@ -1,69 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_FAR_ISOMORPHIC_H_
-#define FST_EXTENSIONS_FAR_ISOMORPHIC_H_
-
-#include <memory>
-#include <string>
-
-#include <fst/extensions/far/far.h>
-#include <fst/isomorphic.h>
-
-namespace fst {
-
-template <class Arc>
-bool FarIsomorphic(const string &filename1, const string &filename2,
-                   float delta = kDelta, const string &begin_key = string(),
-                   const string &end_key = string()) {
-  std::unique_ptr<FarReader<Arc>> reader1(FarReader<Arc>::Open(filename1));
-  if (!reader1) {
-    LOG(ERROR) << "FarIsomorphic: Cannot open FAR file " << filename1;
-    return false;
-  }
-  std::unique_ptr<FarReader<Arc>> reader2(FarReader<Arc>::Open(filename2));
-  if (!reader2) {
-    LOG(ERROR) << "FarIsomorphic: Cannot open FAR file " << filename2;
-    return false;
-  }
-  if (!begin_key.empty()) {
-    bool find_begin1 = reader1->Find(begin_key);
-    bool find_begin2 = reader2->Find(begin_key);
-    if (!find_begin1 || !find_begin2) {
-      bool ret = !find_begin1 && !find_begin2;
-      if (!ret) {
-        VLOG(1) << "FarIsomorphic: Key " << begin_key << " missing from "
-                << (find_begin1 ? "second" : "first") << " archive.";
-      }
-      return ret;
-    }
-  }
-  for (; !reader1->Done() && !reader2->Done();
-       reader1->Next(), reader2->Next()) {
-    const auto &key1 = reader1->GetKey();
-    const auto &key2 = reader2->GetKey();
-    if (!end_key.empty() && end_key < key1 && end_key < key2) return true;
-    if (key1 != key2) {
-      LOG(ERROR) << "FarIsomorphic: Mismatched keys " << key1 << " and "
-                 << key2;
-      return false;
-    }
-    if (!Isomorphic(*(reader1->GetFst()), *(reader2->GetFst()), delta)) {
-      LOG(ERROR) << "FarIsomorphic: FSTs for key " << key1
-                 << " are not isomorphic";
-      return false;
-    }
-  }
-  if (!reader1->Done() || !reader2->Done()) {
-    LOG(ERROR) << "FarIsomorphic: Key "
-               << (reader1->Done() ? reader2->GetKey() : reader1->GetKey())
-               << " missing form " << (reader2->Done() ? "first" : "second")
-               << " archive";
-    return false;
-  }
-  return true;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_ISOMORPHIC_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/print-strings.h b/projects/llm_framework/include/fst/extensions/far/print-strings.h
deleted file mode 100644
index dc428401..00000000
--- a/projects/llm_framework/include/fst/extensions/far/print-strings.h
+++ /dev/null
@@ -1,105 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Outputs as strings the string FSTs in a finite-state archive.
-
-#ifndef FST_EXTENSIONS_FAR_PRINT_STRINGS_H_
-#define FST_EXTENSIONS_FAR_PRINT_STRINGS_H_
-
-#include <iomanip>
-#include <string>
-#include <vector>
-
-#include <fst/flags.h>
-#include <fst/extensions/far/far.h>
-#include <fstream>
-#include <fst/shortest-distance.h>
-#include <fst/string.h>
-
-DECLARE_string(far_field_separator);
-
-namespace fst {
-
-template <class Arc>
-void FarPrintStrings(const std::vector<string> &ifilenames,
-                     FarEntryType entry_type, FarTokenType far_token_type,
-                     const string &begin_key, const string &end_key,
-                     bool print_key, bool print_weight,
-                     const string &symbols_fname, bool initial_symbols,
-                     int32 generate_filenames, const string &filename_prefix,
-                     const string &filename_suffix) {
-  StringTokenType token_type;
-  if (far_token_type == FTT_SYMBOL) {
-    token_type = StringTokenType::SYMBOL;
-  } else if (far_token_type == FTT_BYTE) {
-    token_type = StringTokenType::BYTE;
-  } else if (far_token_type == FTT_UTF8) {
-    token_type = StringTokenType::UTF8;
-  } else {
-    FSTERROR() << "FarPrintStrings: Unknown token type";
-    return;
-  }
-  std::unique_ptr<const SymbolTable> syms;
-  if (!symbols_fname.empty()) {
-    // TODO(kbg): Allow negative flag?
-    const SymbolTableTextOptions opts(true);
-    syms.reset(SymbolTable::ReadText(symbols_fname, opts));
-    if (!syms) {
-      LOG(ERROR) << "FarPrintStrings: Error reading symbol table "
-                 << symbols_fname;
-      return;
-    }
-  }
-  std::unique_ptr<FarReader<Arc>> far_reader(FarReader<Arc>::Open(ifilenames));
-  if (!far_reader) return;
-  if (!begin_key.empty()) far_reader->Find(begin_key);
-  string okey;
-  int nrep = 0;
-  for (int i = 1; !far_reader->Done(); far_reader->Next(), ++i) {
-    const auto &key = far_reader->GetKey();
-    if (!end_key.empty() && end_key < key) break;
-    if (okey == key) {
-      ++nrep;
-    } else {
-      nrep = 0;
-    }
-    okey = key;
-    const auto *fst = far_reader->GetFst();
-    if (i == 1 && initial_symbols && !syms && fst->InputSymbols())
-      syms.reset(fst->InputSymbols()->Copy());
-    string str;
-    VLOG(2) << "Handling key: " << key;
-    StringPrinter<Arc> string_printer(token_type,
-                                      syms ? syms.get() : fst->InputSymbols());
-    string_printer(*fst, &str);
-    if (entry_type == FET_LINE) {
-      if (print_key) std::cout << key << FLAGS_far_field_separator[0];
-      std::cout << str;
-      if (print_weight)
-        std::cout << FLAGS_far_field_separator[0] << ShortestDistance(*fst);
-      std::cout << std::endl;
-    } else if (entry_type == FET_FILE) {
-      std::stringstream sstrm;
-      if (generate_filenames) {
-        sstrm.fill('0');
-        sstrm << std::right << std::setw(generate_filenames) << i;
-      } else {
-        sstrm << key;
-        if (nrep > 0) sstrm << "." << nrep;
-      }
-      string filename;
-      filename = filename_prefix + sstrm.str() + filename_suffix;
-      std::ofstream ostrm(filename);
-      if (!ostrm) {
-        LOG(ERROR) << "FarPrintStrings: Can't open file: " << filename;
-        return;
-      }
-      ostrm << str;
-      if (token_type == StringTokenType::SYMBOL) ostrm << "\n";
-    }
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_PRINT_STRINGS_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/script-impl.h b/projects/llm_framework/include/fst/extensions/far/script-impl.h
deleted file mode 100644
index a0586cc3..00000000
--- a/projects/llm_framework/include/fst/extensions/far/script-impl.h
+++ /dev/null
@@ -1,23 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes and functions for registering and invoking Far main
-// functions that support multiple and extensible arc types.
-
-#ifndef FST_EXTENSIONS_FAR_SCRIPT_IMPL_H_
-#define FST_EXTENSIONS_FAR_SCRIPT_IMPL_H_
-
-#include <string>
-
-#include <fst/compat.h>
-namespace fst {
-namespace script {
-
-string LoadArcTypeFromFar(const string &far_fname);
-
-string LoadArcTypeFromFst(const string &fst_fname);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_SCRIPT_IMPL_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/stlist.h b/projects/llm_framework/include/fst/extensions/far/stlist.h
deleted file mode 100644
index b155e17e..00000000
--- a/projects/llm_framework/include/fst/extensions/far/stlist.h
+++ /dev/null
@@ -1,273 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// A generic (string,type) list file format.
-//
-// This is a stripped-down version of STTable that does not support the Find()
-// operation but that does support reading/writting from standard in/out.
-
-#ifndef FST_EXTENSIONS_FAR_STLIST_H_
-#define FST_EXTENSIONS_FAR_STLIST_H_
-
-#include <algorithm>
-#include <functional>
-#include <iostream>
-#include <memory>
-#include <queue>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fstream>
-#include <fst/util.h>
-
-namespace fst {
-
-static constexpr int32 kSTListMagicNumber = 5656924;
-static constexpr int32 kSTListFileVersion = 1;
-
-// String-type list writing class for object of type T using a functor Writer.
-// The Writer functor must provide at least the following interface:
-//
-//   struct Writer {
-//     void operator()(std::ostream &, const T &) const;
-//   };
-template <class T, class Writer>
-class STListWriter {
- public:
-  explicit STListWriter(const string &filename)
-      : stream_(filename.empty() ? &std::cout : new std::ofstream(
-                                                    filename,
-                                                    std::ios_base::out |
-                                                        std::ios_base::binary)),
-        error_(false) {
-    WriteType(*stream_, kSTListMagicNumber);
-    WriteType(*stream_, kSTListFileVersion);
-    if (!stream_) {
-      FSTERROR() << "STListWriter::STListWriter: Error writing to file: "
-                 << filename;
-      error_ = true;
-    }
-  }
-
-  static STListWriter<T, Writer> *Create(const string &filename) {
-    return new STListWriter<T, Writer>(filename);
-  }
-
-  void Add(const string &key, const T &t) {
-    if (key == "") {
-      FSTERROR() << "STListWriter::Add: Key empty: " << key;
-      error_ = true;
-    } else if (key < last_key_) {
-      FSTERROR() << "STListWriter::Add: Key out of order: " << key;
-      error_ = true;
-    }
-    if (error_) return;
-    last_key_ = key;
-    WriteType(*stream_, key);
-    entry_writer_(*stream_, t);
-  }
-
-  bool Error() const { return error_; }
-
-  ~STListWriter() {
-    WriteType(*stream_, string());
-    if (stream_ != &std::cout) delete stream_;
-  }
-
- private:
-  Writer entry_writer_;
-  std::ostream *stream_;  // Output stream.
-  string last_key_;       // Last key.
-  bool error_;
-
-  STListWriter(const STListWriter &) = delete;
-  STListWriter &operator=(const STListWriter &) = delete;
-};
-
-// String-type list reading class for object of type T using a functor Reader.
-// Reader must provide at least the following interface:
-//
-//   struct Reader {
-//     T *operator()(std::istream &) const;
-//   };
-template <class T, class Reader>
-class STListReader {
- public:
-  explicit STListReader(const std::vector<string> &filenames)
-      : sources_(filenames), error_(false) {
-    streams_.resize(filenames.size(), 0);
-    bool has_stdin = false;
-    for (size_t i = 0; i < filenames.size(); ++i) {
-      if (filenames[i].empty()) {
-        if (!has_stdin) {
-          streams_[i] = &std::cin;
-          sources_[i] = "stdin";
-          has_stdin = true;
-        } else {
-          FSTERROR() << "STListReader::STListReader: Cannot read multiple "
-                     << "inputs from standard input";
-          error_ = true;
-          return;
-        }
-      } else {
-        streams_[i] = new std::ifstream(
-            filenames[i], std::ios_base::in | std::ios_base::binary);
-        if (streams_[i]->fail()) {
-          FSTERROR() << "STListReader::STListReader: Error reading file: "
-                     << filenames[i];
-          error_ = true;
-          return;
-        }
-      }
-      int32 magic_number = 0;
-      ReadType(*streams_[i], &magic_number);
-      int32 file_version = 0;
-      ReadType(*streams_[i], &file_version);
-      if (magic_number != kSTListMagicNumber) {
-        FSTERROR() << "STListReader::STListReader: Wrong file type: "
-                   << filenames[i];
-        error_ = true;
-        return;
-      }
-      if (file_version != kSTListFileVersion) {
-        FSTERROR() << "STListReader::STListReader: Wrong file version: "
-                   << filenames[i];
-        error_ = true;
-        return;
-      }
-      string key;
-      ReadType(*streams_[i], &key);
-      if (!key.empty()) heap_.push(std::make_pair(key, i));
-      if (!*streams_[i]) {
-        FSTERROR() << "STListReader: Error reading file: " << sources_[i];
-        error_ = true;
-        return;
-      }
-    }
-    if (heap_.empty()) return;
-    const auto current = heap_.top().second;
-    entry_.reset(entry_reader_(*streams_[current]));
-    if (!entry_ || !*streams_[current]) {
-      FSTERROR() << "STListReader: Error reading entry for key "
-                 << heap_.top().first << ", file " << sources_[current];
-      error_ = true;
-    }
-  }
-
-  ~STListReader() {
-    for (auto &stream : streams_) {
-      if (stream != &std::cin) delete stream;
-    }
-  }
-
-  static STListReader<T, Reader> *Open(const string &filename) {
-    std::vector<string> filenames;
-    filenames.push_back(filename);
-    return new STListReader<T, Reader>(filenames);
-  }
-
-  static STListReader<T, Reader> *Open(const std::vector<string> &filenames) {
-    return new STListReader<T, Reader>(filenames);
-  }
-
-  void Reset() {
-    FSTERROR() << "STListReader::Reset: Operation not supported";
-    error_ = true;
-  }
-
-  bool Find(const string &key) {
-    FSTERROR() << "STListReader::Find: Operation not supported";
-    error_ = true;
-    return false;
-  }
-
-  bool Done() const { return error_ || heap_.empty(); }
-
-  void Next() {
-    if (error_) return;
-    auto current = heap_.top().second;
-    string key;
-    heap_.pop();
-    ReadType(*(streams_[current]), &key);
-    if (!*streams_[current]) {
-      FSTERROR() << "STListReader: Error reading file: " << sources_[current];
-      error_ = true;
-      return;
-    }
-    if (!key.empty()) heap_.push(std::make_pair(key, current));
-    if (!heap_.empty()) {
-      current = heap_.top().second;
-      entry_.reset(entry_reader_(*streams_[current]));
-      if (!entry_ || !*streams_[current]) {
-        FSTERROR() << "STListReader: Error reading entry for key: "
-                   << heap_.top().first << ", file: " << sources_[current];
-        error_ = true;
-      }
-    }
-  }
-
-  const string &GetKey() const { return heap_.top().first; }
-
-  const T *GetEntry() const { return entry_.get(); }
-
-  bool Error() const { return error_; }
-
- private:
-  Reader entry_reader_;                  // Read functor.
-  std::vector<std::istream *> streams_;  // Input streams.
-  std::vector<string> sources_;          // Corresponding filenames.
-  std::priority_queue<
-      std::pair<string, size_t>, std::vector<std::pair<string, size_t>>,
-      std::greater<std::pair<string, size_t>>> heap_;  // (Key, stream id) heap
-  mutable std::unique_ptr<T> entry_;  // The currently read entry.
-  bool error_;
-
-  STListReader(const STListReader &) = delete;
-  STListReader &operator=(const STListReader &) = delete;
-};
-
-// String-type list header reading function, templated on the entry header type.
-// The Header type must provide at least the following interface:
-//
-//  struct Header {
-//    void Read(std::istream &strm, const string &filename);
-//  };
-template <class Header>
-bool ReadSTListHeader(const string &filename, Header *header) {
-  if (filename.empty()) {
-    LOG(ERROR) << "ReadSTListHeader: Can't read header from standard input";
-    return false;
-  }
-  std::ifstream strm(filename, std::ios_base::in | std::ios_base::binary);
-  if (!strm) {
-    LOG(ERROR) << "ReadSTListHeader: Could not open file: " << filename;
-    return false;
-  }
-  int32 magic_number = 0;
-  ReadType(strm, &magic_number);
-  int32 file_version = 0;
-  ReadType(strm, &file_version);
-  if (magic_number != kSTListMagicNumber) {
-    LOG(ERROR) << "ReadSTListHeader: Wrong file type: " << filename;
-    return false;
-  }
-  if (file_version != kSTListFileVersion) {
-    LOG(ERROR) << "ReadSTListHeader: Wrong file version: " << filename;
-    return false;
-  }
-  string key;
-  ReadType(strm, &key);
-  header->Read(strm, filename + ":" + key);
-  if (!strm) {
-    LOG(ERROR) << "ReadSTListHeader: Error reading file: " << filename;
-    return false;
-  }
-  return true;
-}
-
-bool IsSTList(const string &filename);
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_STLIST_H_
diff --git a/projects/llm_framework/include/fst/extensions/far/sttable.h b/projects/llm_framework/include/fst/extensions/far/sttable.h
deleted file mode 100644
index 2a01bb16..00000000
--- a/projects/llm_framework/include/fst/extensions/far/sttable.h
+++ /dev/null
@@ -1,353 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// A generic string-to-type table file format.
-//
-// This is not meant as a generalization of SSTable. This is more of a simple
-// replacement for SSTable in order to provide an open-source implementation
-// of the FAR format for the external version of the FST library.
-
-#ifndef FST_EXTENSIONS_FAR_STTABLE_H_
-#define FST_EXTENSIONS_FAR_STTABLE_H_
-
-#include <algorithm>
-#include <istream>
-#include <memory>
-
-#include <fstream>
-#include <fst/util.h>
-
-namespace fst {
-
-static constexpr int32 kSTTableMagicNumber = 2125656924;
-static constexpr int32 kSTTableFileVersion = 1;
-
-// String-type table writing class for an object of type T using a functor
-// Writer. The Writer functor must provide at least the following interface:
-//
-//   struct Writer {
-//     void operator()(std::ostream &, const T &) const;
-//   };
-template <class T, class Writer>
-class STTableWriter {
- public:
-  explicit STTableWriter(const string &filename)
-      : stream_(filename, std::ios_base::out | std::ios_base::binary),
-        error_(false) {
-    WriteType(stream_, kSTTableMagicNumber);
-    WriteType(stream_, kSTTableFileVersion);
-    if (stream_.fail()) {
-      FSTERROR() << "STTableWriter::STTableWriter: Error writing to file: "
-                 << filename;
-      error_ = true;
-    }
-  }
-
-  static STTableWriter<T, Writer> *Create(const string &filename) {
-    if (filename.empty()) {
-      LOG(ERROR) << "STTableWriter: Writing to standard out unsupported.";
-      return nullptr;
-    }
-    return new STTableWriter<T, Writer>(filename);
-  }
-
-  void Add(const string &key, const T &t) {
-    if (key == "") {
-      FSTERROR() << "STTableWriter::Add: Key empty: " << key;
-      error_ = true;
-    } else if (key < last_key_) {
-      FSTERROR() << "STTableWriter::Add: Key out of order: " << key;
-      error_ = true;
-    }
-    if (error_) return;
-    last_key_ = key;
-    positions_.push_back(stream_.tellp());
-    WriteType(stream_, key);
-    entry_writer_(stream_, t);
-  }
-
-  bool Error() const { return error_; }
-
-  ~STTableWriter() {
-    WriteType(stream_, positions_);
-    WriteType(stream_, static_cast<int64>(positions_.size()));
-  }
-
- private:
-  Writer entry_writer_;
-  std::ofstream stream_;
-  std::vector<int64> positions_;  // Position in file of each key-entry pair.
-  string last_key_;               // Last key.
-  bool error_;
-
-  STTableWriter(const STTableWriter &) = delete;
-  STTableWriter &operator=(const STTableWriter &) = delete;
-};
-
-// String-type table reading class for object of type T using a functor Reader.
-// Reader must provide at least the following interface:
-//
-//   struct Reader {
-//     T *operator()(std::istream &) const;
-//   };
-//
-template <class T, class Reader>
-class STTableReader {
- public:
-  explicit STTableReader(const std::vector<string> &filenames)
-      : sources_(filenames), error_(false) {
-    compare_.reset(new Compare(&keys_));
-    keys_.resize(filenames.size());
-    streams_.resize(filenames.size(), 0);
-    positions_.resize(filenames.size());
-    for (size_t i = 0; i < filenames.size(); ++i) {
-      streams_[i] = new std::ifstream(
-          filenames[i], std::ios_base::in | std::ios_base::binary);
-      if (streams_[i]->fail()) {
-        FSTERROR() << "STTableReader::STTableReader: Error reading file: "
-                 << filenames[i];
-        error_ = true;
-        return;
-      }
-      int32 magic_number = 0;
-      ReadType(*streams_[i], &magic_number);
-      int32 file_version = 0;
-      ReadType(*streams_[i], &file_version);
-      if (magic_number != kSTTableMagicNumber) {
-        FSTERROR() << "STTableReader::STTableReader: Wrong file type: "
-                   << filenames[i];
-        error_ = true;
-        return;
-      }
-      if (file_version != kSTTableFileVersion) {
-        FSTERROR() << "STTableReader::STTableReader: Wrong file version: "
-                   << filenames[i];
-        error_ = true;
-        return;
-      }
-      int64 num_entries;
-      streams_[i]->seekg(-static_cast<int>(sizeof(int64)), std::ios_base::end);
-      ReadType(*streams_[i], &num_entries);
-      if (num_entries > 0) {
-        streams_[i]->seekg(-static_cast<int>(sizeof(int64)) * (num_entries + 1),
-                           std::ios_base::end);
-        positions_[i].resize(num_entries);
-        for (size_t j = 0; (j < num_entries) && (!streams_[i]->fail()); ++j) {
-          ReadType(*streams_[i], &(positions_[i][j]));
-        }
-        streams_[i]->seekg(positions_[i][0]);
-        if (streams_[i]->fail()) {
-          FSTERROR() << "STTableReader::STTableReader: Error reading file: "
-                     << filenames[i];
-          error_ = true;
-          return;
-        }
-      }
-    }
-    MakeHeap();
-  }
-
-  ~STTableReader() {
-    for (auto &stream : streams_) delete stream;
-  }
-
-  static STTableReader<T, Reader> *Open(const string &filename) {
-    if (filename.empty()) {
-      LOG(ERROR) << "STTableReader: Operation not supported on standard input";
-      return nullptr;
-    }
-    std::vector<string> filenames;
-    filenames.push_back(filename);
-    return new STTableReader<T, Reader>(filenames);
-  }
-
-  static STTableReader<T, Reader> *Open(const std::vector<string> &filenames) {
-    return new STTableReader<T, Reader>(filenames);
-  }
-
-  void Reset() {
-    if (error_) return;
-    for (size_t i = 0; i < streams_.size(); ++i)
-      streams_[i]->seekg(positions_[i].front());
-    MakeHeap();
-  }
-
-  bool Find(const string &key) {
-    if (error_) return false;
-    for (size_t i = 0; i < streams_.size(); ++i) LowerBound(i, key);
-    MakeHeap();
-    if (heap_.empty()) return false;
-    return keys_[current_] == key;
-  }
-
-  bool Done() const { return error_ || heap_.empty(); }
-
-  void Next() {
-    if (error_) return;
-    if (streams_[current_]->tellg() <= positions_[current_].back()) {
-      ReadType(*(streams_[current_]), &(keys_[current_]));
-      if (streams_[current_]->fail()) {
-        FSTERROR() << "STTableReader: Error reading file: "
-                   << sources_[current_];
-        error_ = true;
-        return;
-      }
-      std::push_heap(heap_.begin(), heap_.end(), *compare_);
-    } else {
-      heap_.pop_back();
-    }
-    if (!heap_.empty()) PopHeap();
-  }
-
-  const string &GetKey() const { return keys_[current_]; }
-
-  const T *GetEntry() const { return entry_.get(); }
-
-  bool Error() const { return error_; }
-
- private:
-  // Comparison functor used to compare stream IDs in the heap.
-  struct Compare {
-    explicit Compare(const std::vector<string> *keys) : keys(keys) {}
-
-    bool operator()(size_t i, size_t j) const {
-      return (*keys)[i] > (*keys)[j];
-    };
-
-   private:
-    const std::vector<string> *keys;
-  };
-
-  // Positions the stream at the position corresponding to the lower bound for
-  // the specified key.
-  void LowerBound(size_t id, const string &find_key) {
-    auto *strm = streams_[id];
-    const auto &positions = positions_[id];
-    if (positions.empty()) return;
-    size_t low = 0;
-    size_t high = positions.size() - 1;
-    while (low < high) {
-      size_t mid = (low + high) / 2;
-      strm->seekg(positions[mid]);
-      string key;
-      ReadType(*strm, &key);
-      if (key > find_key) {
-        high = mid;
-      } else if (key < find_key) {
-        low = mid + 1;
-      } else {
-        for (size_t i = mid; i > low; --i) {
-          strm->seekg(positions[i - 1]);
-          ReadType(*strm, &key);
-          if (key != find_key) {
-            strm->seekg(positions[i]);
-            return;
-          }
-        }
-        strm->seekg(positions[low]);
-        return;
-      }
-    }
-    strm->seekg(positions[low]);
-  }
-
-  // Adds all streams to the heap.
-  void MakeHeap() {
-    heap_.clear();
-    for (size_t i = 0; i < streams_.size(); ++i) {
-      if (positions_[i].empty()) continue;
-      ReadType(*streams_[i], &(keys_[i]));
-      if (streams_[i]->fail()) {
-        FSTERROR() << "STTableReader: Error reading file: " << sources_[i];
-        error_ = true;
-        return;
-      }
-      heap_.push_back(i);
-    }
-    if (heap_.empty()) return;
-    std::make_heap(heap_.begin(), heap_.end(), *compare_);
-    PopHeap();
-  }
-
-  // Positions the stream with the lowest key at the top of the heap, sets
-  // current_ to the ID of that stream, and reads the current entry from that
-  // stream.
-  void PopHeap() {
-    std::pop_heap(heap_.begin(), heap_.end(), *compare_);
-    current_ = heap_.back();
-    entry_.reset(entry_reader_(*streams_[current_]));
-    if (!entry_) error_ = true;
-    if (streams_[current_]->fail()) {
-      FSTERROR() << "STTableReader: Error reading entry for key: "
-                 << keys_[current_] << ", file: " << sources_[current_];
-      error_ = true;
-    }
-  }
-
-  Reader entry_reader_;
-  std::vector<std::istream *> streams_;        // Input streams.
-  std::vector<string> sources_;                // Corresponding file names.
-  std::vector<std::vector<int64>> positions_;  // Index of positions.
-  std::vector<string> keys_;  // Lowest unread key for each stream.
-  std::vector<int64> heap_;   // Heap containing ID of streams with unread keys.
-  int64 current_;             // ID of current stream to be read.
-  std::unique_ptr<Compare> compare_;  // Functor comparing stream IDs.
-  mutable std::unique_ptr<T> entry_;  // The currently read entry.
-  bool error_;
-};
-
-// String-type table header reading function template on the entry header type.
-// The Header type must provide at least the following interface:
-//
-//   struct Header {
-//     void Read(std::istream &istrm, const string &filename);
-//   };
-template <class Header>
-bool ReadSTTableHeader(const string &filename, Header *header) {
-  if (filename.empty()) {
-    LOG(ERROR) << "ReadSTTable: Can't read header from standard input";
-    return false;
-  }
-  std::ifstream strm(filename, std::ios_base::in | std::ios_base::binary);
-  if (!strm) {
-    LOG(ERROR) << "ReadSTTableHeader: Could not open file: " << filename;
-    return false;
-  }
-  int32 magic_number = 0;
-  ReadType(strm, &magic_number);
-  int32 file_version = 0;
-  ReadType(strm, &file_version);
-  if (magic_number != kSTTableMagicNumber) {
-    LOG(ERROR) << "ReadSTTableHeader: Wrong file type: " << filename;
-    return false;
-  }
-  if (file_version != kSTTableFileVersion) {
-    LOG(ERROR) << "ReadSTTableHeader: Wrong file version: " << filename;
-    return false;
-  }
-  int64 i = -1;
-  strm.seekg(-static_cast<int>(sizeof(int64)), std::ios_base::end);
-  ReadType(strm, &i);  // Reads number of entries
-  if (strm.fail()) {
-    LOG(ERROR) << "ReadSTTableHeader: Error reading file: " << filename;
-    return false;
-  }
-  if (i == 0) return true;  // No entry header to read.
-  strm.seekg(-2 * static_cast<int>(sizeof(int64)), std::ios_base::end);
-  ReadType(strm, &i);  // Reads position for last entry in file.
-  strm.seekg(i);
-  string key;
-  ReadType(strm, &key);
-  header->Read(strm, filename + ":" + key);
-  if (strm.fail()) {
-    LOG(ERROR) << "ReadSTTableHeader: Error reading file: " << filename;
-    return false;
-  }
-  return true;
-}
-
-bool IsSTTable(const string &filename);
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_STTABLE_H_
diff --git a/projects/llm_framework/include/fst/extensions/linear/linear-fst-data-builder.h b/projects/llm_framework/include/fst/extensions/linear/linear-fst-data-builder.h
deleted file mode 100644
index a6ac7279..00000000
--- a/projects/llm_framework/include/fst/extensions/linear/linear-fst-data-builder.h
+++ /dev/null
@@ -1,1074 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_BUILDER_H_
-#define FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_BUILDER_H_
-
-#include <map>
-#include <queue>
-#include <set>
-#include <sstream>
-#include <stack>
-#include <string>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/log.h>
-#include <fst/fst.h>
-#include <fst/symbol-table.h>
-#include <fst/util.h>
-
-#include <fst/extensions/linear/linear-fst-data.h>
-
-namespace fst {
-
-// Forward declaration
-template <class A>
-class FeatureGroupBuilder;
-
-// For logging purposes
-inline string TranslateLabel(int64 label, const SymbolTable *syms);
-template <class Iterator>
-string JoinLabels(Iterator begin, Iterator end, const SymbolTable *syms);
-template <class Label>
-string JoinLabels(const std::vector<Label> &labels, const SymbolTable *syms);
-
-// Guesses the appropriate boundary label (start- or end-of-sentence)
-// for all labels equal to `boundary` and modifies the `sequence`
-// in-place. Returns the number of positions that are still uncertain.
-template <class A>
-typename A::Label GuessStartOrEnd(std::vector<typename A::Label> *sequence,
-                                  typename A::Label boundary);
-
-// Builds a `LinearFstData` object by adding words and feature
-// weights. A few conventions:
-//
-// - Input labels forms a dense non-empty range from 1 to `MaxInputLabel()`.
-// - Feature labels, output labels are > 0.
-// - Being a discriminative linear model, it only makes sense to use tropical
-// semirings.
-template <class A>
-class LinearFstDataBuilder {
- public:
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-
-  // Constructs a builder with associated symbol tables for diagonstic
-  // output. Each of these symbol tables may also be nullptr.
-  explicit LinearFstDataBuilder(const SymbolTable *isyms = nullptr,
-                                const SymbolTable *fsyms = nullptr,
-                                const SymbolTable *osyms = nullptr)
-      : error_(false),
-        max_future_size_(0),
-        max_input_label_(1),
-        isyms_(isyms),
-        fsyms_(fsyms),
-        osyms_(osyms) {}
-
-  // Tests whether the builder has encountered any error. No operation
-  // is valid if the builder is already at error state. All other
-  // public methods should check this before any actual operations.
-  bool Error() const { return error_; }
-
-  // Adds a word and its feature labels to the vocabulary; this
-  // version allows the word to have any output label. Returns true
-  // iff the word is added.
-  //
-  // This may fail if the word is added twice or if the feature labels
-  // are non-positive.
-  bool AddWord(Label word, const std::vector<Label> &features);
-
-  // Adds a word and its feature labels to the vocabulary; this
-  // version puts constraint on possible output labels the word can
-  // have. `possible_output` must not be empty. Returns true iff the
-  // word is added.
-  //
-  // In addition to the reasons above in the two-parameter version,
-  // this may also fail if `possible_output` is empty or any output
-  // label in it is non-positive.
-  bool AddWord(Label word, const std::vector<Label> &word_features,
-               const std::vector<Label> &possible_output);
-
-  // Creates a new feature group with specified future size (size of
-  // the look-ahead window), returns the group id to be used for
-  // adding actual feature weights or a negative number when called at
-  // error state.
-  //
-  // This does not fail unless called at error state.
-  int AddGroup(size_t future_size);
-
-  // Adds an instance of feature weight to the specified feature
-  // group. If some weight has already been added with the same
-  // feature, the product of the old and new weights are
-  // stored. Returns true iff the weight is added. A weight is not
-  // added when the context has ill-formed context involving start-,
-  // end-of-sentence marks.
-  //
-  // For two features to be within the same group, it must satisfy
-  // that (1) they have the same future size; (2) the two either have
-  // disjoint context or one is the back-off context of the
-  // other. Furthermore, for all features in a single group, there
-  // must be one and only one other context (not necessarily an active
-  // feature) that the feature immediately backs off to (i.e. there is
-  // no other context that is the back-off of the first and backs off
-  // to the second).
-  //
-  // Consider for example features with zero look-ahead of the form
-  // (input, OUTPUT).
-  //
-  // - The following two features can be put in the same group because
-  // their context is disjoint: (a a a, A A), (b, B B);
-  //
-  // - The following two features can be put in the same group because
-  // one is the back-off context of the other: (a a a, A A), (a a, A
-  // A);
-  //
-  // - The following two features can NOT be put in the same group
-  // because there is overlap but neither is the other's back-off: (a
-  // a a, A), (a a, A A);
-  //
-  // - Finally, the following three features cannot be in a same group
-  // because the first one can immediately back off to either of the
-  // rest: (a a a, A A), (a a, A A), (a a a, A).
-  //
-  // The easiest way to satisfy the constraints is to create a feature
-  // group for each feature template. However, better feature grouping
-  // may help improve speed.
-  //
-  // This may fail if any of input or output labels are non-positive,
-  // or if any call to `FeatureGroupBuilder<>::AddWeight()` fails.
-  bool AddWeight(size_t group, const std::vector<Label> &input,
-                 const std::vector<Label> &output, Weight weight);
-
-  // Returns a newly created `LinearFstData` object or nullptr in case
-  // of failure. The caller takes the ownership of the memory. No
-  // other methods shall be called after this --- this is enforced by
-  // putting the builder at error state, even when a
-  // `LinearFstData<>` object is successfully built.
-  //
-  // This may fail if the call to any `FeatureGroupBuilder<>::Dump()`
-  // fails.
-  LinearFstData<A> *Dump();
-
- private:
-  bool error_;
-  CompactSet<Label, kNoLabel> all_output_labels_;
-  std::map<Label, std::set<Label>> word_output_map_, word_feat_map_;
-  std::map<Label, std::set<size_t>> feat_groups_;
-  std::vector<std::unique_ptr<FeatureGroupBuilder<A>>> groups_;
-  size_t max_future_size_;
-  Label max_input_label_;
-  const SymbolTable *isyms_, *fsyms_, *osyms_;
-
-  LinearFstDataBuilder(const LinearFstDataBuilder &) = delete;
-  LinearFstDataBuilder &operator=(const LinearFstDataBuilder &) = delete;
-};
-
-// Builds a LinearFstData tailored for a LinearClassifierFst. The
-// major difference between an ordinary LinearFstData that works on
-// taggers and a LinearFstData that works on classifiers is that
-// feature groups are divided into sections by the prediction class
-// label. For a prediction label `pred` and a logical group id
-// `group`, the actual group id is `group * num_classes + pred -
-// 1`.
-//
-// This layout saves us from recording output labels in each single
-// FeatureGroup. Because there is no need for any delaying, stripping
-// the output allows features with different shapes but using the same
-// set of feature label mapping to reside in a single FeatureGroup.
-template <class A>
-class LinearClassifierFstDataBuilder {
- public:
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-
-  // Constructs a builder for a `num_classes`-class classifier,
-  // optinally with associated symbol tables for diagnostic
-  // output. The output labels (i.e. prediction) must be in the range
-  // of [1, num_classes].
-  explicit LinearClassifierFstDataBuilder(size_t num_classes,
-                                          const SymbolTable *isyms = nullptr,
-                                          const SymbolTable *fsyms = nullptr,
-                                          const SymbolTable *osyms = nullptr)
-      : error_(false),
-        num_classes_(num_classes),
-        num_groups_(0),
-        builder_(isyms, fsyms, osyms) {}
-
-  // Tests whether the builder has encountered any error. Similar to
-  // LinearFstDataBuilder<>::Error().
-  bool Error() const { return error_; }
-
-  // Same as LinearFstDataBuilder<>::AddWord().
-  bool AddWord(Label word, const std::vector<Label> &features);
-
-  // Adds a logical feature group. Similar to
-  // LinearFstDataBuilder<>::AddGroup(), with the exception that the
-  // returned group id is the logical group id. Also there is no need
-  // for "future" in a classifier.
-  int AddGroup();
-
-  // Adds an instance of feature weight to the specified logical
-  // feature group. Instead of a vector of output, only a single
-  // prediction is needed as the output.
-  //
-  // This may fail if `pred` is not in the range of [1, num_classes_].
-  bool AddWeight(size_t group, const std::vector<Label> &input, Label pred,
-                 Weight weight);
-
-  // Returns a newly created `LinearFstData` object or nullptr in case of
-  // failure.
-  LinearFstData<A> *Dump();
-
- private:
-  std::vector<Label> empty_;
-  bool error_;
-  size_t num_classes_, num_groups_;
-  LinearFstDataBuilder<A> builder_;
-};
-
-// Builds a single feature group. Usually used in
-// `LinearFstDataBuilder::AddWeight()`. See that method for the
-// constraints on grouping features.
-template <class A>
-class FeatureGroupBuilder {
- public:
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-
-  // Constructs a builder with the given future size. All features
-  // added to the group will have look-ahead windows of this size.
-  FeatureGroupBuilder(size_t future_size, const SymbolTable *fsyms,
-                      const SymbolTable *osyms)
-      : error_(false), future_size_(future_size), fsyms_(fsyms), osyms_(osyms) {
-    // This edge is special; see doc of class `FeatureGroup` on the
-    // details.
-    start_ = trie_.Insert(trie_.Root(), InputOutputLabel(kNoLabel, kNoLabel));
-  }
-
-  // Tests whether the builder has encountered any error. No operation
-  // is valid if the builder is already at error state. All other
-  // public methods should check this before any actual operations.
-  bool Error() const { return error_; }
-
-  // Adds a feature weight with the given context. Returns true iff
-  // the weight is added. A weight is not added if it has ill-formed
-  // context involving start-, end-of-sentence marks.
-  //
-  // Note: `input` is the sequence of input
-  // features, instead of input labels themselves. `input` must be at
-  // least as long as `future_size`; `output` may be empty, but
-  // usually should be non-empty because an empty output context is
-  // useless in discriminative modelling. All labels in both `input`
-  // and `output` must be > 0 (this is checked in
-  // `LinearFstDataBuilder::AddWeight()`). See
-  // LinearFstDataBuilder<>::AddWeight for more details.
-  //
-  // This may fail if the input is smaller than the look-ahead window.
-  bool AddWeight(const std::vector<Label> &input,
-                 const std::vector<Label> &output, Weight weight);
-
-  // Creates an actual FeatureGroup<> object. Connects back-off links;
-  // pre-accumulates weights from back-off features. Returns nullptr if
-  // there is any violation in unique immediate back-off
-  // constraints.
-  //
-  // Regardless of whether the call succeeds or not, the error flag is
-  // always set before this returns, to prevent repeated dumping.
-  //
-  // TODO(wuke): check overlapping top-level contexts (see
-  // `DumpOverlappingContext()` in tests).
-  FeatureGroup<A> *Dump(size_t max_future_size);
-
- private:
-  typedef typename FeatureGroup<A>::InputOutputLabel InputOutputLabel;
-  typedef typename FeatureGroup<A>::InputOutputLabelHash InputOutputLabelHash;
-  typedef typename FeatureGroup<A>::WeightBackLink WeightBackLink;
-  // Nested trie topology uses more memory but we can traverse a
-  // node's children easily, which is required in `BuildBackLinks()`.
-  typedef NestedTrieTopology<InputOutputLabel, InputOutputLabelHash> Topology;
-  typedef MutableTrie<InputOutputLabel, WeightBackLink, Topology> Trie;
-
-  // Finds the first node with an arc with `label` following the
-  // back-off chain of `parent`. Returns the node index or
-  // `kNoTrieNodeId` when not found. The number of hops is stored in
-  // `hop` when it is not `nullptr`.
-  //
-  // This does not fail.
-  int FindFirstMatch(InputOutputLabel label, int parent, int *hop) const;
-
-  // Links each node to its immediate back-off. root is linked to -1.
-  //
-  // This may fail when the unique immediate back-off constraint is
-  // violated.
-  void BuildBackLinks();
-
-  // Traces back on the back-chain for each node to multiply the
-  // weights from back-offs to the node itself.
-  //
-  // This does not fail.
-  void PreAccumulateWeights();
-
-  // Reconstruct the path from trie root to given node for logging.
-  bool TrieDfs(const Topology &topology, int cur, int target,
-               std::vector<InputOutputLabel> *path) const;
-  string TriePath(int node, const Topology &topology) const;
-
-  bool error_;
-  size_t future_size_;
-  Trie trie_;
-  int start_;
-  const SymbolTable *fsyms_, *osyms_;
-
-  FeatureGroupBuilder(const FeatureGroupBuilder &) = delete;
-  FeatureGroupBuilder &operator=(const FeatureGroupBuilder &) = delete;
-};
-
-//
-// Implementation of methods in `LinearFstDataBuilder`
-//
-template <class A>
-bool LinearFstDataBuilder<A>::AddWord(Label word,
-                                      const std::vector<Label> &features) {
-  if (error_) {
-    FSTERROR() << "Calling LinearFstDataBuilder<>::AddWord() at error state";
-    return false;
-  }
-  if (word == LinearFstData<A>::kStartOfSentence ||
-      word == LinearFstData<A>::kEndOfSentence) {
-    LOG(WARNING) << "Ignored: adding boundary label: "
-                 << TranslateLabel(word, isyms_)
-                 << "(start-of-sentence=" << LinearFstData<A>::kStartOfSentence
-                 << ", end-of-sentence=" << LinearFstData<A>::kEndOfSentence
-                 << ")";
-    return false;
-  }
-  if (word <= 0) {
-    error_ = true;
-    FSTERROR() << "Word label must be > 0; got " << word;
-    return false;
-  }
-  if (word > max_input_label_) max_input_label_ = word;
-  // Make sure the word hasn't been added before
-  if (word_feat_map_.find(word) != word_feat_map_.end()) {
-    error_ = true;
-    FSTERROR() << "Input word " << TranslateLabel(word, isyms_)
-               << " is added twice";
-    return false;
-  }
-  // Store features
-  std::set<Label> *feats = &word_feat_map_[word];
-  for (size_t i = 0; i < features.size(); ++i) {
-    Label feat = features[i];
-    if (feat <= 0) {
-      error_ = true;
-      FSTERROR() << "Feature label must be > 0; got " << feat;
-      return false;
-    }
-    feats->insert(feat);
-  }
-  return true;
-}
-
-template <class A>
-bool LinearFstDataBuilder<A>::AddWord(
-    Label word, const std::vector<Label> &word_features,
-    const std::vector<Label> &possible_output) {
-  if (error_) {
-    FSTERROR() << "Calling LinearFstDataBuilder<>::AddWord() at error state";
-    return false;
-  }
-  if (!AddWord(word, word_features)) return false;
-  // Store possible output constraint
-  if (possible_output.empty()) {
-    error_ = true;
-    FSTERROR() << "Empty possible output constraint; "
-               << "use the two-parameter version if no constraint is need.";
-    return false;
-  }
-  std::set<Label> *outputs = &word_output_map_[word];
-  for (size_t i = 0; i < possible_output.size(); ++i) {
-    Label output = possible_output[i];
-    if (output == LinearFstData<A>::kStartOfSentence ||
-        output == LinearFstData<A>::kEndOfSentence) {
-      LOG(WARNING) << "Ignored: word = " << TranslateLabel(word, isyms_)
-                   << ": adding boundary label as possible output: " << output
-                   << "(start-of-sentence="
-                   << LinearFstData<A>::kStartOfSentence
-                   << ", end-of-sentence=" << LinearFstData<A>::kEndOfSentence
-                   << ")";
-      continue;
-    }
-    if (output <= 0) {
-      error_ = true;
-      FSTERROR() << "Output label must be > 0; got " << output;
-      return false;
-    }
-    outputs->insert(output);
-    all_output_labels_.Insert(output);
-  }
-  return true;
-}
-
-template <class A>
-inline int LinearFstDataBuilder<A>::AddGroup(size_t future_size) {
-  if (error_) {
-    FSTERROR() << "Calling LinearFstDataBuilder<>::AddGroup() at error state";
-    return -1;
-  }
-  size_t ret = groups_.size();
-  groups_.emplace_back(new FeatureGroupBuilder<A>(future_size, fsyms_, osyms_));
-  if (future_size > max_future_size_) max_future_size_ = future_size;
-  return ret;
-}
-
-template <class A>
-bool LinearFstDataBuilder<A>::AddWeight(size_t group,
-                                        const std::vector<Label> &input,
-                                        const std::vector<Label> &output,
-                                        Weight weight) {
-  if (error_) {
-    FSTERROR() << "Calling LinearFstDataBuilder<>::AddWeight() at error state";
-    return false;
-  }
-  // Check well-formedness of boundary marks on the input.
-  {
-    bool start_in_middle = false, end_in_middle = false;
-    for (int i = 1; i < input.size(); ++i) {
-      if (input[i] == LinearFstData<A>::kStartOfSentence &&
-          input[i - 1] != LinearFstData<A>::kStartOfSentence)
-        start_in_middle = true;
-      if (input[i - 1] == LinearFstData<A>::kEndOfSentence &&
-          input[i] != LinearFstData<A>::kEndOfSentence)
-        end_in_middle = true;
-    }
-    if (start_in_middle) {
-      LOG(WARNING) << "Ignored: start-of-sentence in the middle of the input!";
-      LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-      LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-      return false;
-    }
-    if (end_in_middle) {
-      LOG(WARNING) << "Ignored: end-of-sentence in the middle of the input!";
-      LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-      LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-      return false;
-    }
-  }
-  // Check well-formedness of boundary marks on the output.
-  {
-    bool non_first_start = false, non_last_end = false;
-    for (int i = 1; i < output.size(); ++i) {
-      if (output[i] == LinearFstData<A>::kStartOfSentence)
-        non_first_start = true;
-      if (output[i - 1] == LinearFstData<A>::kEndOfSentence)
-        non_last_end = true;
-    }
-    if (non_first_start) {
-      LOG(WARNING) << "Ignored: start-of-sentence not appearing "
-                   << "as the first label in the output!";
-      LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-      LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-      return false;
-    }
-    if (non_last_end) {
-      LOG(WARNING) << "Ignored: end-of-sentence not appearing "
-                   << "as the last label in the output!";
-      LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-      LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-      return false;
-    }
-  }
-
-  for (size_t i = 0; i < input.size(); ++i) {
-    Label feat = input[i];
-    if (feat != LinearFstData<A>::kStartOfSentence &&
-        feat != LinearFstData<A>::kEndOfSentence && feat <= 0) {
-      error_ = true;
-      FSTERROR() << "Feature label must be > 0; got " << feat;
-      return false;
-    }
-    feat_groups_[feat].insert(group);
-  }
-  for (size_t i = 0; i < output.size(); ++i) {
-    Label label = output[i];
-    if (label != LinearFstData<A>::kStartOfSentence &&
-        label != LinearFstData<A>::kEndOfSentence && label <= 0) {
-      error_ = true;
-      FSTERROR() << "Output label must be > 0; got " << label;
-      return false;
-    }
-    if (label != LinearFstData<A>::kStartOfSentence &&
-        label != LinearFstData<A>::kEndOfSentence)
-      all_output_labels_.Insert(label);
-  }
-
-  // Everything looks good at this point (more checks on the way in
-  // the feature group). Add this feature weight.
-  bool added = groups_[group]->AddWeight(input, output, weight);
-  if (groups_[group]->Error()) {
-    error_ = true;
-    FSTERROR() << "FeatureGroupBuilder<>::AddWeight() failed";
-    return false;
-  }
-  return added;
-}
-
-template <class A>
-LinearFstData<A> *LinearFstDataBuilder<A>::Dump() {
-  if (error_) {
-    FSTERROR() << "Calling LinearFstDataBuilder<>::Dump() at error state";
-    return nullptr;
-  }
-
-  std::unique_ptr<LinearFstData<A>> data(new LinearFstData<A>());
-  data->max_future_size_ = max_future_size_;
-  data->max_input_label_ = max_input_label_;
-
-  // Feature groups; free builders after it's dumped.
-  data->groups_.resize(groups_.size());
-  for (int group = 0; group != groups_.size(); ++group) {
-    FeatureGroup<A> *new_group = groups_[group]->Dump(max_future_size_);
-    if (new_group == nullptr) {
-      error_ = true;
-      FSTERROR() << "Error in dumping group " << group;
-      return nullptr;
-    }
-    data->groups_[group].reset(new_group);
-    groups_[group].reset();
-    VLOG(1) << "Group " << group << ": " << new_group->Stats();
-  }
-
-  // Per-group feature mapping
-  data->group_feat_map_.Init(data->NumGroups(), max_input_label_ + 1);
-  for (Label word = 1; word <= max_input_label_; ++word) {
-    typename std::map<Label, std::set<Label>>::const_iterator it =
-        word_feat_map_.find(word);
-    if (it == word_feat_map_.end()) continue;
-    for (typename std::set<Label>::const_iterator oit = it->second.begin();
-         oit != it->second.end(); ++oit) {
-      Label feat = *oit;
-      typename std::map<Label, std::set<size_t>>::const_iterator jt =
-          feat_groups_.find(feat);
-      if (jt == feat_groups_.end()) continue;
-      for (std::set<size_t>::const_iterator git = jt->second.begin();
-           git != jt->second.end(); ++git) {
-        size_t group_id = *git;
-        if (!data->group_feat_map_.Set(group_id, word, feat)) {
-          error_ = true;
-          return nullptr;
-        }
-      }
-    }
-  }
-
-  // Possible output labels
-  {
-    std::vector<typename LinearFstData<A>::InputAttribute> *input_attribs =
-        &data->input_attribs_;
-    std::vector<Label> *output_pool = &data->output_pool_;
-    input_attribs->resize(max_input_label_ + 1);
-    for (Label word = 0; word <= max_input_label_; ++word) {
-      typename std::map<Label, std::set<Label>>::const_iterator it =
-          word_output_map_.find(word);
-      if (it == word_output_map_.end()) {
-        (*input_attribs)[word].output_begin = 0;
-        (*input_attribs)[word].output_length = 0;
-      } else {
-        (*input_attribs)[word].output_begin = output_pool->size();
-        (*input_attribs)[word].output_length = it->second.size();
-        for (typename std::set<Label>::const_iterator oit = it->second.begin();
-             oit != it->second.end(); ++oit) {
-          Label olabel = *oit;
-          output_pool->push_back(olabel);
-        }
-      }
-    }
-  }
-
-  for (typename CompactSet<Label, kNoLabel>::const_iterator it =
-           all_output_labels_.Begin();
-       it != all_output_labels_.End(); ++it)
-    data->output_set_.push_back(*it);
-
-  error_ = true;  // prevent future calls on this object
-  return data.release();
-}
-
-//
-// Implementation of methods in `LinearClassifierFstDataBuilder`
-//
-template <class A>
-inline bool LinearClassifierFstDataBuilder<A>::AddWord(
-    Label word, const std::vector<Label> &features) {
-  if (error_) {
-    FSTERROR() << "Calling LinearClassifierFstDataBuilder<>::AddWord() at "
-                  "error state";
-    return false;
-  }
-  bool added = builder_.AddWord(word, features);
-  if (builder_.Error()) error_ = true;
-  return added;
-}
-
-template <class A>
-inline int LinearClassifierFstDataBuilder<A>::AddGroup() {
-  if (error_) {
-    FSTERROR() << "Calling LinearClassifierFstDataBuilder<>::AddGroup() at "
-                  "error state";
-    return -1;
-  }
-  for (int i = 0; i < num_classes_; ++i) builder_.AddGroup(0);
-  if (builder_.Error()) {
-    error_ = true;
-    return -1;
-  }
-  return num_groups_++;
-}
-
-template <class A>
-inline bool LinearClassifierFstDataBuilder<A>::AddWeight(
-    size_t group, const std::vector<Label> &input, Label pred, Weight weight) {
-  if (error_) {
-    FSTERROR() << "Calling LinearClassifierFstDataBuilder<>::AddWeight() at "
-                  "error state";
-    return false;
-  }
-  if (pred <= 0 || pred > num_classes_) {
-    FSTERROR() << "Out-of-range prediction label: " << pred
-               << " (num classes = " << num_classes_ << ")";
-    error_ = true;
-    return false;
-  }
-  size_t real_group = group * num_classes_ + pred - 1;
-  bool added = builder_.AddWeight(real_group, input, empty_, weight);
-  if (builder_.Error()) error_ = true;
-  return added;
-}
-
-template <class A>
-inline LinearFstData<A> *LinearClassifierFstDataBuilder<A>::Dump() {
-  if (error_) {
-    FSTERROR()
-        << "Calling LinearClassifierFstDataBuilder<>::Dump() at error state";
-    return nullptr;
-  }
-  LinearFstData<A> *data = builder_.Dump();
-  error_ = true;
-  return data;
-}
-
-//
-// Implementation of methods in `FeatureGroupBuilder`
-//
-template <class A>
-bool FeatureGroupBuilder<A>::AddWeight(const std::vector<Label> &input,
-                                       const std::vector<Label> &output,
-                                       Weight weight) {
-  if (error_) {
-    FSTERROR() << "Calling FeatureGroupBuilder<>::AddWeight() at error state";
-    return false;
-  }
-
-  // `LinearFstDataBuilder<>::AddWeight()` ensures prefix/suffix
-  // properties for us. We can directly count.
-  int num_input_start = 0;
-  while (num_input_start < input.size() &&
-         input[num_input_start] == LinearFstData<A>::kStartOfSentence)
-    ++num_input_start;
-  int num_output_start = 0;
-  while (num_output_start < output.size() &&
-         output[num_output_start] == LinearFstData<A>::kStartOfSentence)
-    ++num_output_start;
-  int num_input_end = 0;
-  for (int i = input.size() - 1;
-       i >= 0 && input[i] == LinearFstData<A>::kEndOfSentence; --i)
-    ++num_input_end;
-  int num_output_end = 0;
-  for (int i = output.size() - 1;
-       i >= 0 && output[i] == LinearFstData<A>::kEndOfSentence; --i)
-    ++num_output_end;
-
-  DCHECK_LE(num_output_end, 1);
-
-  if (input.size() - num_input_start < future_size_) {
-    LOG(WARNING) << "Ignored: start-of-sentence in the future!";
-    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-    LOG(WARNING) << "\tOutput: " << JoinLabels(output, fsyms_);
-    return false;
-  }
-  if (num_input_start > 0 &&
-      input.size() - future_size_ - num_input_start <
-          output.size() - num_output_start) {
-    LOG(WARNING) << "Ignored: matching start-of-sentence with actual output!";
-    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-    LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-    return false;
-  }
-  if (num_output_start > 0 &&
-      input.size() - future_size_ - num_input_start >
-          output.size() - num_output_start) {
-    LOG(WARNING) << "Ignored: matching start-of-sentence with actual input!";
-    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-    LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-    return false;
-  }
-  // The following two require `num_output_end` <= 1.
-  if (num_input_end > future_size_ && num_input_end - future_size_ != 1) {
-    LOG(WARNING) << "Ignored: matching end-of-sentence with actual output!";
-    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-    LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-    return false;
-  }
-  if (num_output_end > 0 &&
-      ((input.size() == future_size_ && future_size_ != num_input_end) ||
-       (input.size() > future_size_ &&
-        num_input_end != future_size_ + num_output_end))) {
-    LOG(WARNING) << "Ignored: matching end-of-sentence with actual input!";
-    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-    LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-    return false;
-  }
-  // Check if the context has no other labels than boundary marks
-  // (such features are useless).
-  if (num_input_start + num_input_end == input.size() &&
-      num_output_start + num_output_end == output.size()) {
-    LOG(WARNING)
-        << "Ignored: feature context consisting of only boundary marks!";
-    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-    LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-    return false;
-  }
-
-  // Start point for insertion in the trie. Insert at `start_` iff the
-  // beginning of the context is non-consumed start-of-sentence.
-  int cur = (num_input_start == 0 && num_output_start <= future_size_)
-                ? trie_.Root()
-                : start_;
-  // Skip all input start-of-sentence marks
-  size_t ipos = num_input_start;
-  // Skip to keep at most `future_size_` start-of-sentence marks
-  size_t opos =
-      num_output_start <= future_size_ ? 0 : num_output_start - future_size_;
-  // Skip `num_output_end` end-of-sentence marks on both input and output
-  size_t iend = !input.empty() ? input.size() - num_output_end : 0,
-         oend = output.size() - num_output_end;
-  // Further, when output is empty, keep at most `future_size_`
-  // end-of-sentence marks on input.
-  if (output.empty() && num_input_end > future_size_)
-    iend = input.size() - num_input_end + future_size_;
-
-  // Actual feature context is (input[ipos:iend], output[opos:oend]).
-
-  // Pad `kNoLabel` as don't cares on the shorter of actual `input`
-  // and `output`.
-  const size_t effective_input_size = iend - ipos,
-               effective_output_size = oend - opos;
-  if (effective_input_size > effective_output_size) {
-    for (size_t pad = effective_input_size - effective_output_size; pad != 0;
-         --pad, ++ipos)
-      cur = trie_.Insert(cur, InputOutputLabel(input[ipos], kNoLabel));
-  } else if (effective_input_size < effective_output_size) {
-    for (size_t pad = effective_output_size - effective_input_size; pad != 0;
-         --pad, ++opos)
-      cur = trie_.Insert(cur, InputOutputLabel(kNoLabel, output[opos]));
-  }
-  CHECK_EQ(iend - ipos, oend - opos);
-  for (; ipos != iend; ++ipos, ++opos)
-    cur = trie_.Insert(cur, InputOutputLabel(input[ipos], output[opos]));
-  // We only need to attach final weight when there is an output
-  // end-of-sentence. When there is only end-of-sentence on the input,
-  // they are all consumed as the end-of-sentence paddings from
-  // `LinearFstImpl<>::ShiftBuffer()`. `LinearFstImpl<>::Expand()`
-  // and `LinearFstImpl<>::MatchInput()` ensures no other
-  // transition takes place after consuming the padding.
-  if (num_output_end > 0 || (output.empty() && num_input_end > future_size_))
-    trie_[cur].final_weight = Times(trie_[cur].final_weight, weight);
-  else
-    trie_[cur].weight = Times(trie_[cur].weight, weight);
-
-  return true;
-}
-
-template <class A>
-FeatureGroup<A> *FeatureGroupBuilder<A>::Dump(size_t max_future_size) {
-  if (error_) {
-    FSTERROR() << "Calling FeatureGroupBuilder<>::PreAccumulateWeights() "
-               << "at error state";
-    return nullptr;
-  }
-
-  if (max_future_size < future_size_) {
-    error_ = true;
-    FSTERROR() << "max_future_size (= " << max_future_size
-               << ") is smaller the builder's future_size (= " << future_size_
-               << ")";
-    return nullptr;
-  }
-
-  BuildBackLinks();
-  if (error_) return nullptr;
-  PreAccumulateWeights();  // does not fail
-
-  FeatureGroup<A> *ret =
-      new FeatureGroup<A>(max_future_size - future_size_, start_);
-
-  // Walk around the trie to compute next states
-  ret->next_state_.resize(trie_.NumNodes());
-  const Topology &topology = trie_.TrieTopology();
-  for (int i = 0; i < topology.NumNodes(); ++i) {
-    int next = i;
-    while (next != topology.Root() && topology.ChildrenOf(next).empty() &&
-           trie_[next].final_weight ==
-               trie_[trie_[next].back_link].final_weight)
-      next = trie_[next].back_link;
-    ret->next_state_[i] = next;
-  }
-
-  // Copy the trie
-  typename FeatureGroup<A>::Trie store_trie(trie_);
-  ret->trie_.swap(store_trie);
-
-  // Put the builder at error state to prevent repeated call of `Dump()`.
-  error_ = true;
-  return ret;
-}
-
-template <class A>
-int FeatureGroupBuilder<A>::FindFirstMatch(InputOutputLabel label, int parent,
-                                           int *hop) const {
-  int hop_count = 0;
-  int ret = kNoTrieNodeId;
-  for (; parent >= 0; parent = trie_[parent].back_link, ++hop_count) {
-    int next = trie_.Find(parent, label);
-    if (next != kNoTrieNodeId) {
-      ret = next;
-      break;
-    }
-  }
-  if (hop != nullptr) *hop = hop_count;
-  return ret;
-}
-
-template <class A>
-void FeatureGroupBuilder<A>::BuildBackLinks() {
-  // Breadth first search from the root. In the case where we only
-  // have the input label, the immedate back-off is simply the longest
-  // suffix of the current node that is also in the trie. For a node
-  // reached from its parent with label L, we can simply walk through
-  // the parent's back-off chain to find the first state with an arc
-  // of the same label L. The uniqueness is always
-  // guanranteed. However, in the case with both input and output
-  // labels, it is possible to back off by removing first labels from
-  // either side, which in general causes non-uniqueness.
-
-  const Topology &topology = trie_.TrieTopology();
-  std::queue<int> q;  // all enqueued or visited nodes have known links
-
-  // Note: nodes have back link initialized to -1 in their
-  // constructor.
-  q.push(trie_.Root());
-  while (!error_ && !q.empty()) {
-    int parent = q.front();
-    q.pop();
-    // Find links for every child
-    const typename Topology::NextMap &children = topology.ChildrenOf(parent);
-    for (typename Topology::NextMap::const_iterator eit = children.begin();
-         eit != children.end(); ++eit) {
-      const std::pair<InputOutputLabel, int> &edge = *eit;
-      InputOutputLabel label = edge.first;
-      int child = edge.second;
-      if (label.input == kNoLabel || label.output == kNoLabel) {
-        // Label pairs from root to here all have one and only one
-        // `kNoLabel` on the same side; equivalent to the
-        // "longest-suffix" case.
-        trie_[child].back_link =
-            FindFirstMatch(label, trie_[parent].back_link, nullptr);
-      } else {
-        // Neither side is `kNoLabel` at this point, there are
-        // three possible ways to back-off: if the parent backs
-        // off to some context with only one side non-empty, the
-        // empty side may remain empty; or else an exact match of
-        // both sides is needed. Try to find all three possible
-        // backs and look for the closest one (in terms of hops
-        // along the parent's back-off chain).
-        int only_input_hop, only_output_hop, full_hop;
-        int only_input_link =
-                FindFirstMatch(InputOutputLabel(label.input, kNoLabel), parent,
-                               &only_input_hop),
-            only_output_link =
-                FindFirstMatch(InputOutputLabel(kNoLabel, label.output), parent,
-                               &only_output_hop),
-            full_link =
-                FindFirstMatch(label, trie_[parent].back_link, &full_hop);
-        if (only_input_link != -1 && only_output_link != -1) {
-          error_ = true;
-          FSTERROR() << "Branching back-off chain:\n"
-                     << "\tnode " << child << ": " << TriePath(child, topology)
-                     << "\n"
-                     << "\tcan back-off to node " << only_input_link << ": "
-                     << TriePath(only_input_link, topology) << "\n"
-                     << "\tcan back-off to node " << only_output_link << ": "
-                     << TriePath(only_output_link, topology);
-          return;
-        } else if (full_link != -1) {
-          ++full_hop;
-          if (full_hop <= only_input_hop && full_hop <= only_output_hop) {
-            trie_[child].back_link = full_link;
-          } else {
-            error_ = true;
-            int problem_link = only_input_link != kNoTrieNodeId
-                                   ? only_input_link
-                                   : only_output_link;
-            CHECK_NE(problem_link, kNoTrieNodeId);
-            FSTERROR() << "Branching back-off chain:\n"
-                       << "\tnode " << child << ": "
-                       << TriePath(child, topology) << "\n"
-                       << "\tcan back-off to node " << full_link << ": "
-                       << TriePath(full_link, topology) << "\n"
-                       << "tcan back-off to node " << problem_link << ": "
-                       << TriePath(problem_link, topology);
-            return;
-          }
-        } else {
-          trie_[child].back_link =
-              only_input_link != -1 ? only_input_link : only_output_link;
-        }
-      }
-      if (error_) break;
-      // Point to empty context (root) when no back-off can be found
-      if (trie_[child].back_link == -1) trie_[child].back_link = 0;
-      q.push(child);
-    }
-  }
-}
-
-template <class A>
-void FeatureGroupBuilder<A>::PreAccumulateWeights() {
-  std::vector<bool> visited(trie_.NumNodes(), false);
-  visited[trie_.Root()] = true;
-
-  for (size_t i = 0; i != trie_.NumNodes(); ++i) {
-    std::stack<int> back_offs;
-    for (int j = i; !visited[j]; j = trie_[j].back_link) back_offs.push(j);
-    while (!back_offs.empty()) {
-      int j = back_offs.top();
-      back_offs.pop();
-      WeightBackLink &node = trie_[j];
-      node.weight = Times(node.weight, trie_[node.back_link].weight);
-      node.final_weight =
-          Times(node.final_weight, trie_[node.back_link].final_weight);
-      visited[j] = true;
-    }
-  }
-}
-
-template <class A>
-bool FeatureGroupBuilder<A>::TrieDfs(
-    const Topology &topology, int cur, int target,
-    std::vector<InputOutputLabel> *path) const {
-  if (cur == target) return true;
-  const typename Topology::NextMap &children = topology.ChildrenOf(cur);
-  for (typename Topology::NextMap::const_iterator eit = children.begin();
-       eit != children.end(); ++eit) {
-    const std::pair<InputOutputLabel, int> &edge = *eit;
-    path->push_back(edge.first);
-    if (TrieDfs(topology, edge.second, target, path)) return true;
-    path->pop_back();
-  }
-  return false;
-}
-
-template <class A>
-string FeatureGroupBuilder<A>::TriePath(int node,
-                                        const Topology &topology) const {
-  std::vector<InputOutputLabel> labels;
-  TrieDfs(topology, topology.Root(), node, &labels);
-  bool first = true;
-  std::ostringstream strm;
-  for (typename std::vector<InputOutputLabel>::const_iterator it =
-           labels.begin();
-       it != labels.end(); ++it) {
-    InputOutputLabel i = *it;
-    if (first)
-      first = false;
-    else
-      strm << ", ";
-    strm << "(" << TranslateLabel(i.input, fsyms_) << ", "
-         << TranslateLabel(i.output, osyms_) << ")";
-  }
-  return strm.str();
-}
-
-inline string TranslateLabel(int64 label, const SymbolTable *syms) {
-  string ret;
-  if (syms != nullptr) ret += syms->Find(label);
-  if (ret.empty()) {
-    std::ostringstream strm;
-    strm << '<' << label << '>';
-    ret = strm.str();
-  }
-  return ret;
-}
-
-template <class Iterator>
-string JoinLabels(Iterator begin, Iterator end, const SymbolTable *syms) {
-  if (begin == end) return "<empty>";
-  std::ostringstream strm;
-  bool first = true;
-  for (Iterator it = begin; it != end; ++it) {
-    if (first)
-      first = false;
-    else
-      strm << '|';
-    strm << TranslateLabel(*it, syms);
-  }
-  return strm.str();
-}
-
-template <class Label>
-string JoinLabels(const std::vector<Label> &labels, const SymbolTable *syms) {
-  return JoinLabels(labels.begin(), labels.end(), syms);
-}
-
-template <class A>
-typename A::Label GuessStartOrEnd(std::vector<typename A::Label> *sequence,
-                                  typename A::Label boundary) {
-  const size_t length = sequence->size();
-  std::vector<bool> non_boundary_on_left(length, false),
-      non_boundary_on_right(length, false);
-  for (size_t i = 1; i < length; ++i) {
-    non_boundary_on_left[i] =
-        non_boundary_on_left[i - 1] || (*sequence)[i - 1] != boundary;
-    non_boundary_on_right[length - 1 - i] = non_boundary_on_right[length - i] ||
-                                            (*sequence)[length - i] != boundary;
-  }
-  int unresolved = 0;
-  for (size_t i = 0; i < length; ++i) {
-    if ((*sequence)[i] != boundary) continue;
-    const bool left = non_boundary_on_left[i], right = non_boundary_on_right[i];
-    if (left && right) {
-      // Boundary in the middle
-      LOG(WARNING) << "Boundary label in the middle of the sequence! position: "
-                   << i << "; boundary: " << boundary
-                   << "; sequence: " << JoinLabels(*sequence, nullptr);
-      LOG(WARNING)
-          << "This is an invalid sequence anyway so I will set it to start.";
-      (*sequence)[i] = LinearFstData<A>::kStartOfSentence;
-    } else if (left && !right) {
-      // Can only be end
-      (*sequence)[i] = LinearFstData<A>::kEndOfSentence;
-    } else if (!left && right) {
-      // Can only be start
-      (*sequence)[i] = LinearFstData<A>::kStartOfSentence;
-    } else {
-      // !left && !right; can't really tell
-      ++unresolved;
-    }
-  }
-  return unresolved;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_BUILDER_H_
diff --git a/projects/llm_framework/include/fst/extensions/linear/linear-fst-data.h b/projects/llm_framework/include/fst/extensions/linear/linear-fst-data.h
deleted file mode 100644
index 3b39c29d..00000000
--- a/projects/llm_framework/include/fst/extensions/linear/linear-fst-data.h
+++ /dev/null
@@ -1,526 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Data structures for storing and looking up the actual feature weights.
-
-#ifndef FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_H_
-#define FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_H_
-
-#include <memory>
-#include <numeric>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/fst.h>
-
-#include <fst/extensions/linear/trie.h>
-
-namespace fst {
-
-// Forward declarations
-template <class A>
-class LinearFstDataBuilder;
-template <class A>
-class FeatureGroup;
-
-// Immutable data storage of the feature weights in a linear
-// model. Produces state tuples that represent internal states of a
-// LinearTaggerFst. Object of this class can only be constructed via
-// either `LinearFstDataBuilder::Dump()` or `LinearFstData::Read()`
-// and usually used as refcount'd object shared across mutiple
-// `LinearTaggerFst` copies.
-//
-// TODO(wuke): more efficient trie implementation
-template <class A>
-class LinearFstData {
- public:
-  friend class LinearFstDataBuilder<A>;  // For builder access
-
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-
-  // Sentence boundary labels. Both of them are negative labels other
-  // than `kNoLabel`.
-  static const Label kStartOfSentence;
-  static const Label kEndOfSentence;
-
-  // Constructs empty data; for non-trivial ways of construction see
-  // `Read()` and `LinearFstDataBuilder`.
-  LinearFstData()
-      : max_future_size_(0), max_input_label_(1), input_attribs_(1) {}
-
-  // Appends the state tuple of the start state to `output`, where
-  // each tuple holds the node ids of a trie for each feature group.
-  void EncodeStartState(std::vector<Label> *output) const {
-    for (int i = 0; i < NumGroups(); ++i) output->push_back(GroupStartState(i));
-  }
-
-  // Takes a transition from the trie states stored in
-  // `(trie_state_begin, trie_state_end)` with input label `ilabel`
-  // and output label `olabel`; appends the destination state tuple to
-  // `next` and multiplies the weight of the transition onto
-  // `weight`. `next` should be the shifted input buffer of the caller
-  // in `LinearTaggerFstImpl` (i.e. of size `LinearTaggerFstImpl::delay_`;
-  // the last element is `ilabel`).
-  template <class Iterator>
-  void TakeTransition(Iterator buffer_end, Iterator trie_state_begin,
-                      Iterator trie_state_end, Label ilabel, Label olabel,
-                      std::vector<Label> *next, Weight *weight) const;
-
-  // Returns the final weight of the given trie state sequence.
-  template <class Iterator>
-  Weight FinalWeight(Iterator trie_state_begin, Iterator trie_state_end) const;
-
-  // Returns the start trie state of the given group.
-  Label GroupStartState(int group_id) const {
-    return groups_[group_id]->Start();
-  }
-
-  // Takes a transition only within the given group. Returns the
-  // destination trie state and multiplies the weight onto `weight`.
-  Label GroupTransition(int group_id, int trie_state, Label ilabel,
-                        Label olabel, Weight *weight) const;
-
-  // Returns the final weight of the given trie state in the given group.
-  Weight GroupFinalWeight(int group_id, int trie_state) const {
-    return groups_[group_id]->FinalWeight(trie_state);
-  }
-
-  Label MinInputLabel() const { return 1; }
-
-  Label MaxInputLabel() const { return max_input_label_; }
-
-  // Returns the maximum future size of all feature groups. Future is
-  // the look-ahead window of a feature, e.g. if a feature looks at
-  // the next 2 words after the current input, then the future size is
-  // 2. There is no look-ahead for output. Features inside a single
-  // `FeatureGroup` must have equal future size.
-  size_t MaxFutureSize() const { return max_future_size_; }
-
-  // Returns the number of feature groups
-  size_t NumGroups() const { return groups_.size(); }
-
-  // Returns the range of possible output labels for an input label.
-  std::pair<typename std::vector<Label>::const_iterator,
-            typename std::vector<Label>::const_iterator>
-  PossibleOutputLabels(Label word) const;
-
-  static LinearFstData<A> *Read(std::istream &strm);  // NOLINT
-  std::ostream &Write(std::ostream &strm) const;      // NOLINT
-
- private:
-  // Offsets in `output_pool_`
-  struct InputAttribute {
-    size_t output_begin, output_length;
-
-    std::istream &Read(std::istream &strm);         // NOLINT
-    std::ostream &Write(std::ostream &strm) const;  // NOLINT
-  };
-
-  // Mapping from input label to per-group feature label
-  class GroupFeatureMap;
-
-  // Translates the input label into input feature label of group
-  // `group`; returns `kNoLabel` when there is no feature for that
-  // group.
-  Label FindFeature(size_t group, Label word) const;
-
-  size_t max_future_size_;
-  Label max_input_label_;
-  std::vector<std::unique_ptr<const FeatureGroup<A>>> groups_;
-  std::vector<InputAttribute> input_attribs_;
-  std::vector<Label> output_pool_, output_set_;
-  GroupFeatureMap group_feat_map_;
-
-  LinearFstData(const LinearFstData &) = delete;
-  LinearFstData &operator=(const LinearFstData &) = delete;
-};
-
-template <class A>
-const typename A::Label LinearFstData<A>::kStartOfSentence = -3;
-template <class A>
-const typename A::Label LinearFstData<A>::kEndOfSentence = -2;
-
-template <class A>
-template <class Iterator>
-void LinearFstData<A>::TakeTransition(Iterator buffer_end,
-                                      Iterator trie_state_begin,
-                                      Iterator trie_state_end, Label ilabel,
-                                      Label olabel, std::vector<Label> *next,
-                                      Weight *weight) const {
-  DCHECK_EQ(trie_state_end - trie_state_begin, groups_.size());
-  DCHECK(ilabel > 0 || ilabel == kEndOfSentence);
-  DCHECK(olabel > 0 || olabel == kStartOfSentence);
-  size_t group_id = 0;
-  for (Iterator it = trie_state_begin; it != trie_state_end; ++it, ++group_id) {
-    size_t delay = groups_[group_id]->Delay();
-    // On the buffer, there may also be `kStartOfSentence` from the
-    // initial empty buffer.
-    Label real_ilabel = delay == 0 ? ilabel : *(buffer_end - delay);
-    next->push_back(
-        GroupTransition(group_id, *it, real_ilabel, olabel, weight));
-  }
-}
-
-template <class A>
-typename A::Label LinearFstData<A>::GroupTransition(int group_id,
-                                                    int trie_state,
-                                                    Label ilabel, Label olabel,
-                                                    Weight *weight) const {
-  Label group_ilabel = FindFeature(group_id, ilabel);
-  return groups_[group_id]->Walk(trie_state, group_ilabel, olabel, weight);
-}
-
-template <class A>
-template <class Iterator>
-inline typename A::Weight LinearFstData<A>::FinalWeight(
-    Iterator trie_state_begin, Iterator trie_state_end) const {
-  DCHECK_EQ(trie_state_end - trie_state_begin, groups_.size());
-  size_t group_id = 0;
-  Weight accum = Weight::One();
-  for (Iterator it = trie_state_begin; it != trie_state_end; ++it, ++group_id)
-    accum = Times(accum, GroupFinalWeight(group_id, *it));
-  return accum;
-}
-
-template <class A>
-inline std::pair<typename std::vector<typename A::Label>::const_iterator,
-                 typename std::vector<typename A::Label>::const_iterator>
-LinearFstData<A>::PossibleOutputLabels(Label word) const {
-  const InputAttribute &attrib = input_attribs_[word];
-  if (attrib.output_length == 0)
-    return std::make_pair(output_set_.begin(), output_set_.end());
-  else
-    return std::make_pair(
-        output_pool_.begin() + attrib.output_begin,
-        output_pool_.begin() + attrib.output_begin + attrib.output_length);
-}
-
-template <class A>
-inline LinearFstData<A> *LinearFstData<A>::Read(std::istream &strm) {  // NOLINT
-  std::unique_ptr<LinearFstData<A>> data(new LinearFstData<A>());
-  ReadType(strm, &(data->max_future_size_));
-  ReadType(strm, &(data->max_input_label_));
-  // Feature groups
-  size_t num_groups = 0;
-  ReadType(strm, &num_groups);
-  data->groups_.resize(num_groups);
-  for (size_t i = 0; i < num_groups; ++i)
-    data->groups_[i].reset(FeatureGroup<A>::Read(strm));
-  // Other data
-  ReadType(strm, &(data->input_attribs_));
-  ReadType(strm, &(data->output_pool_));
-  ReadType(strm, &(data->output_set_));
-  ReadType(strm, &(data->group_feat_map_));
-  if (strm) {
-    return data.release();
-  } else {
-    return nullptr;
-  }
-}
-
-template <class A>
-inline std::ostream &LinearFstData<A>::Write(
-    std::ostream &strm) const {  // NOLINT
-  WriteType(strm, max_future_size_);
-  WriteType(strm, max_input_label_);
-  // Feature groups
-  WriteType(strm, groups_.size());
-  for (size_t i = 0; i < groups_.size(); ++i) {
-    groups_[i]->Write(strm);
-  }
-  // Other data
-  WriteType(strm, input_attribs_);
-  WriteType(strm, output_pool_);
-  WriteType(strm, output_set_);
-  WriteType(strm, group_feat_map_);
-  return strm;
-}
-
-template <class A>
-typename A::Label LinearFstData<A>::FindFeature(size_t group,
-                                                Label word) const {
-  DCHECK(word > 0 || word == kStartOfSentence || word == kEndOfSentence);
-  if (word == kStartOfSentence || word == kEndOfSentence)
-    return word;
-  else
-    return group_feat_map_.Find(group, word);
-}
-
-template <class A>
-inline std::istream &LinearFstData<A>::InputAttribute::Read(
-    std::istream &strm) {  // NOLINT
-  ReadType(strm, &output_begin);
-  ReadType(strm, &output_length);
-  return strm;
-}
-
-template <class A>
-inline std::ostream &LinearFstData<A>::InputAttribute::Write(
-    std::ostream &strm) const {  // NOLINT
-  WriteType(strm, output_begin);
-  WriteType(strm, output_length);
-  return strm;
-}
-
-// Forward declaration
-template <class A>
-class FeatureGroupBuilder;
-
-// An immutable grouping of features with similar context shape. Like
-// `LinearFstData`, this can only be constructed via `Read()` or
-// via its builder.
-//
-// Internally it uses a trie to store all feature n-grams and their
-// weights. The label of a trie edge is a pair (feat, olabel) of
-// labels. They can be either positive (ordinary label), `kNoLabel`,
-// `kStartOfSentence`, or `kEndOfSentence`. `kNoLabel` usually means
-// matching anything, with one exception: from the root of the trie,
-// there is a special (kNoLabel, kNoLabel) that leads to the implicit
-// start-of-sentence state. This edge is never actually matched
-// (`FindFirstMatch()` ensures this).
-template <class A>
-class FeatureGroup {
- public:
-  friend class FeatureGroupBuilder<A>;  // for builder access
-
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-
-  int Start() const { return start_; }
-
-  // Finds destination node from `cur` by consuming `ilabel` and
-  // `olabel`. The transition weight is multiplied onto `weight`.
-  int Walk(int cur, Label ilabel, Label olabel, Weight *weight) const;
-
-  // Returns the final weight of the current trie state. Only valid if
-  // the state is already known to be part of a final state (see
-  // `LinearFstData<>::CanBeFinal()`).
-  Weight FinalWeight(int trie_state) const {
-    return trie_[trie_state].final_weight;
-  }
-
-  static FeatureGroup<A> *Read(std::istream &strm) {  // NOLINT
-    size_t delay;
-    ReadType(strm, &delay);
-    int start;
-    ReadType(strm, &start);
-    Trie trie;
-    ReadType(strm, &trie);
-    std::unique_ptr<FeatureGroup<A>> ret(new FeatureGroup<A>(delay, start));
-    ret->trie_.swap(trie);
-    ReadType(strm, &ret->next_state_);
-    if (strm) {
-      return ret.release();
-    } else {
-      return nullptr;
-    }
-  }
-
-  std::ostream &Write(std::ostream &strm) const {  // NOLINT
-    WriteType(strm, delay_);
-    WriteType(strm, start_);
-    WriteType(strm, trie_);
-    WriteType(strm, next_state_);
-    return strm;
-  }
-
-  size_t Delay() const { return delay_; }
-
-  string Stats() const;
-
- private:
-  // Label along the arcs on the trie. `kNoLabel` means anything
-  // (non-negative label) can match; both sides holding `kNoLabel`
-  // is not allow; otherwise the label is > 0 (enforced by
-  // `LinearFstDataBuilder::AddWeight()`).
-  struct InputOutputLabel;
-  struct InputOutputLabelHash;
-
-  // Data to be stored on the trie
-  struct WeightBackLink {
-    int back_link;
-    Weight weight, final_weight;
-
-    WeightBackLink()
-        : back_link(kNoTrieNodeId),
-          weight(Weight::One()),
-          final_weight(Weight::One()) {}
-
-    std::istream &Read(std::istream &strm) {  // NOLINT
-      ReadType(strm, &back_link);
-      ReadType(strm, &weight);
-      ReadType(strm, &final_weight);
-      return strm;
-    }
-
-    std::ostream &Write(std::ostream &strm) const {  // NOLINT
-      WriteType(strm, back_link);
-      WriteType(strm, weight);
-      WriteType(strm, final_weight);
-      return strm;
-    }
-  };
-
-  typedef FlatTrieTopology<InputOutputLabel, InputOutputLabelHash> Topology;
-  typedef MutableTrie<InputOutputLabel, WeightBackLink, Topology> Trie;
-
-  explicit FeatureGroup(size_t delay, int start)
-      : delay_(delay), start_(start) {}
-
-  // Finds the first node with an arc with `label` following the
-  // back-off chain of `parent`. Returns the node index or
-  // `kNoTrieNodeId` when not found.
-  int FindFirstMatch(InputOutputLabel label, int parent) const;
-
-  size_t delay_;
-  int start_;
-  Trie trie_;
-  // Where to go after hitting this state. When we reach a state with
-  // no child and with no additional final weight (i.e. its final
-  // weight is the same as its back-off), we can immediately go to its
-  // back-off state.
-  std::vector<int> next_state_;
-
-  FeatureGroup(const FeatureGroup &) = delete;
-  FeatureGroup &operator=(const FeatureGroup &) = delete;
-};
-
-template <class A>
-struct FeatureGroup<A>::InputOutputLabel {
-  Label input, output;
-
-  InputOutputLabel(Label i = kNoLabel, Label o = kNoLabel)
-      : input(i), output(o) {}
-
-  bool operator==(InputOutputLabel that) const {
-    return input == that.input && output == that.output;
-  }
-
-  std::istream &Read(std::istream &strm) {  // NOLINT
-    ReadType(strm, &input);
-    ReadType(strm, &output);
-    return strm;
-  }
-
-  std::ostream &Write(std::ostream &strm) const {  // NOLINT
-    WriteType(strm, input);
-    WriteType(strm, output);
-    return strm;
-  }
-};
-
-template <class A>
-struct FeatureGroup<A>::InputOutputLabelHash {
-  size_t operator()(InputOutputLabel label) const {
-    return static_cast<size_t>(label.input * 7853 + label.output);
-  }
-};
-
-template <class A>
-int FeatureGroup<A>::Walk(int cur, Label ilabel, Label olabel,
-                          Weight *weight) const {
-  // Note: user of this method need to ensure `ilabel` and `olabel`
-  // are valid (e.g. see DCHECKs in
-  // `LinearFstData<>::TakeTransition()` and
-  // `LinearFstData<>::FindFeature()`).
-  int next;
-  if (ilabel == LinearFstData<A>::kStartOfSentence) {
-    // An observed start-of-sentence only occurs in the beginning of
-    // the input, when this feature group is delayed (i.e. there is
-    // another feature group with a larger future size). The actual
-    // input hasn't arrived so stay at the start state.
-    DCHECK_EQ(cur, start_);
-    next = start_;
-  } else {
-    // First, try exact match
-    next = FindFirstMatch(InputOutputLabel(ilabel, olabel), cur);
-    // Then try with don't cares
-    if (next == kNoTrieNodeId)
-      next = FindFirstMatch(InputOutputLabel(ilabel, kNoLabel), cur);
-    if (next == kNoTrieNodeId)
-      next = FindFirstMatch(InputOutputLabel(kNoLabel, olabel), cur);
-    // All failed, go to empty context
-    if (next == kNoTrieNodeId) next = trie_.Root();
-    *weight = Times(*weight, trie_[next].weight);
-    next = next_state_[next];
-  }
-  return next;
-}
-
-template <class A>
-inline int FeatureGroup<A>::FindFirstMatch(InputOutputLabel label,
-                                           int parent) const {
-  if (label.input == kNoLabel && label.output == kNoLabel)
-    return kNoTrieNodeId;  // very important; see class doc.
-  for (; parent != kNoTrieNodeId; parent = trie_[parent].back_link) {
-    int next = trie_.Find(parent, label);
-    if (next != kNoTrieNodeId) return next;
-  }
-  return kNoTrieNodeId;
-}
-
-template <class A>
-inline string FeatureGroup<A>::Stats() const {
-  std::ostringstream strm;
-  int num_states = 2;
-  for (int i = 2; i < next_state_.size(); ++i)
-    num_states += i == next_state_[i];
-  strm << trie_.NumNodes() << " node(s); " << num_states << " state(s)";
-  return strm.str();
-}
-
-template <class A>
-class LinearFstData<A>::GroupFeatureMap {
- public:
-  GroupFeatureMap() {}
-
-  void Init(size_t num_groups, size_t num_words) {
-    num_groups_ = num_groups;
-    pool_.clear();
-    pool_.resize(num_groups * num_words, kNoLabel);
-  }
-
-  Label Find(size_t group_id, Label ilabel) const {
-    return pool_[IndexOf(group_id, ilabel)];
-  }
-
-  bool Set(size_t group_id, Label ilabel, Label feat) {
-    size_t i = IndexOf(group_id, ilabel);
-    if (pool_[i] != kNoLabel && pool_[i] != feat) {
-      FSTERROR() << "Feature group " << group_id
-                 << " already has feature for word " << ilabel;
-      return false;
-    }
-    pool_[i] = feat;
-    return true;
-  }
-
-  std::istream &Read(std::istream &strm) {  // NOLINT
-    ReadType(strm, &num_groups_);
-    ReadType(strm, &pool_);
-    return strm;
-  }
-
-  std::ostream &Write(std::ostream &strm) const {  // NOLINT
-    WriteType(strm, num_groups_);
-    WriteType(strm, pool_);
-    return strm;
-  }
-
- private:
-  size_t IndexOf(size_t group_id, Label ilabel) const {
-    return ilabel * num_groups_ + group_id;
-  }
-
-  size_t num_groups_;
-  // `pool_[ilabel * num_groups_ + group_id]` is the feature active
-  // for group `group_id` with input `ilabel`
-  std::vector<Label> pool_;
-};
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_H_
diff --git a/projects/llm_framework/include/fst/extensions/linear/linear-fst.h b/projects/llm_framework/include/fst/extensions/linear/linear-fst.h
deleted file mode 100644
index e4239a87..00000000
--- a/projects/llm_framework/include/fst/extensions/linear/linear-fst.h
+++ /dev/null
@@ -1,1173 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes for building, storing and representing log-linear models as FSTs.
-
-#ifndef FST_EXTENSIONS_LINEAR_LINEAR_FST_H_
-#define FST_EXTENSIONS_LINEAR_LINEAR_FST_H_
-
-#include <algorithm>
-#include <iostream>
-#include <memory>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/log.h>
-#include <fst/extensions/pdt/collection.h>
-#include <fst/bi-table.h>
-#include <fst/cache.h>
-#include <fstream>
-#include <fst/fst.h>
-#include <fst/matcher.h>
-#include <fst/symbol-table.h>
-
-#include <fst/extensions/linear/linear-fst-data.h>
-
-namespace fst {
-
-// Forward declaration of the specialized matcher for both
-// LinearTaggerFst and LinearClassifierFst.
-template <class F>
-class LinearFstMatcherTpl;
-
-namespace internal {
-
-// Implementation class for on-the-fly generated LinearTaggerFst with
-// special optimization in matching.
-template <class A>
-class LinearTaggerFstImpl : public CacheImpl<A> {
- public:
-  using FstImpl<A>::SetType;
-  using FstImpl<A>::SetProperties;
-  using FstImpl<A>::SetInputSymbols;
-  using FstImpl<A>::SetOutputSymbols;
-  using FstImpl<A>::WriteHeader;
-
-  using CacheBaseImpl<CacheState<A>>::PushArc;
-  using CacheBaseImpl<CacheState<A>>::HasArcs;
-  using CacheBaseImpl<CacheState<A>>::HasFinal;
-  using CacheBaseImpl<CacheState<A>>::HasStart;
-  using CacheBaseImpl<CacheState<A>>::SetArcs;
-  using CacheBaseImpl<CacheState<A>>::SetFinal;
-  using CacheBaseImpl<CacheState<A>>::SetStart;
-
-  typedef A Arc;
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-  typedef typename A::StateId StateId;
-  typedef typename Collection<StateId, Label>::SetIterator NGramIterator;
-
-  // Constructs an empty FST by default.
-  LinearTaggerFstImpl()
-      : CacheImpl<A>(CacheOptions()),
-        data_(std::make_shared<LinearFstData<A>>()),
-        delay_(0) {
-    SetType("linear-tagger");
-  }
-
-  // Constructs the FST with given data storage and symbol
-  // tables.
-  //
-  // TODO(wuke): when there is no constraint on output we can delay
-  // less than `data->MaxFutureSize` positions.
-  LinearTaggerFstImpl(const LinearFstData<Arc> *data, const SymbolTable *isyms,
-                      const SymbolTable *osyms, CacheOptions opts)
-      : CacheImpl<A>(opts), data_(data), delay_(data->MaxFutureSize()) {
-    SetType("linear-tagger");
-    SetProperties(kILabelSorted, kFstProperties);
-    SetInputSymbols(isyms);
-    SetOutputSymbols(osyms);
-    ReserveStubSpace();
-  }
-
-  // Copy by sharing the underlying data storage.
-  LinearTaggerFstImpl(const LinearTaggerFstImpl &impl)
-      : CacheImpl<A>(impl), data_(impl.data_), delay_(impl.delay_) {
-    SetType("linear-tagger");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-    ReserveStubSpace();
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      StateId start = FindStartState();
-      SetStart(start);
-    }
-    return CacheImpl<A>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) {
-      state_stub_.clear();
-      FillState(s, &state_stub_);
-      if (CanBeFinal(state_stub_))
-        SetFinal(s, data_->FinalWeight(InternalBegin(state_stub_),
-                                       InternalEnd(state_stub_)));
-      else
-        SetFinal(s, Weight::Zero());
-    }
-    return CacheImpl<A>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<A>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<A>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<A>::NumOutputEpsilons(s);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<A> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<A>::InitArcIterator(s, data);
-  }
-
-  // Computes the outgoing transitions from a state, creating new
-  // destination states as needed.
-  void Expand(StateId s);
-
-  // Appends to `arcs` all out-going arcs from state `s` that matches `label` as
-  // the input label.
-  void MatchInput(StateId s, Label ilabel, std::vector<Arc> *arcs);
-
-  static LinearTaggerFstImpl *Read(std::istream &strm,
-                                   const FstReadOptions &opts);
-
-  bool Write(std::ostream &strm,  // NOLINT
-             const FstWriteOptions &opts) const {
-    FstHeader header;
-    header.SetStart(kNoStateId);
-    WriteHeader(strm, opts, kFileVersion, &header);
-    data_->Write(strm);
-    if (!strm) {
-      LOG(ERROR) << "LinearTaggerFst::Write: Write failed: " << opts.source;
-      return false;
-    }
-    return true;
-  }
-
- private:
-  static const int kMinFileVersion;
-  static const int kFileVersion;
-
-  // A collection of functions to access parts of the state tuple. A
-  // state tuple is a vector of `Label`s with two parts:
-  //   [buffer] [internal].
-  //
-  // - [buffer] is a buffer of observed input labels with length
-  // `delay_`. `LinearFstData<A>::kStartOfSentence`
-  // (resp. `LinearFstData<A>::kEndOfSentence`) are used as
-  // paddings when the buffer has fewer than `delay_` elements, which
-  // can only appear as the prefix (resp. suffix) of the buffer.
-  //
-  // - [internal] is the internal state tuple for `LinearFstData`
-  typename std::vector<Label>::const_iterator BufferBegin(
-      const std::vector<Label> &state) const {
-    return state.begin();
-  }
-
-  typename std::vector<Label>::const_iterator BufferEnd(
-      const std::vector<Label> &state) const {
-    return state.begin() + delay_;
-  }
-
-  typename std::vector<Label>::const_iterator InternalBegin(
-      const std::vector<Label> &state) const {
-    return state.begin() + delay_;
-  }
-
-  typename std::vector<Label>::const_iterator InternalEnd(
-      const std::vector<Label> &state) const {
-    return state.end();
-  }
-
-  // The size of state tuples are fixed, reserve them in stubs
-  void ReserveStubSpace() {
-    state_stub_.reserve(delay_ + data_->NumGroups());
-    next_stub_.reserve(delay_ + data_->NumGroups());
-  }
-
-  // Computes the start state tuple and maps it to the start state id.
-  StateId FindStartState() {
-    // Empty buffer with start-of-sentence paddings
-    state_stub_.clear();
-    state_stub_.resize(delay_, LinearFstData<A>::kStartOfSentence);
-    // Append internal states
-    data_->EncodeStartState(&state_stub_);
-    return FindState(state_stub_);
-  }
-
-  // Tests whether the buffer in `(begin, end)` is empty.
-  bool IsEmptyBuffer(typename std::vector<Label>::const_iterator begin,
-                     typename std::vector<Label>::const_iterator end) const {
-    // The following is guanranteed by `ShiftBuffer()`:
-    // - buffer[i] == LinearFstData<A>::kEndOfSentence =>
-    //       buffer[i+x] == LinearFstData<A>::kEndOfSentence
-    // - buffer[i] == LinearFstData<A>::kStartOfSentence =>
-    //       buffer[i-x] == LinearFstData<A>::kStartOfSentence
-    return delay_ == 0 || *(end - 1) == LinearFstData<A>::kStartOfSentence ||
-           *begin == LinearFstData<A>::kEndOfSentence;
-  }
-
-  // Tests whether the given state tuple can be a final state. A state
-  // is final iff there is no observed input in the buffer.
-  bool CanBeFinal(const std::vector<Label> &state) {
-    return IsEmptyBuffer(BufferBegin(state), BufferEnd(state));
-  }
-
-  // Finds state corresponding to an n-gram. Creates new state if n-gram not
-  // found.
-  StateId FindState(const std::vector<Label> &ngram) {
-    StateId sparse = ngrams_.FindId(ngram, true);
-    StateId dense = condensed_.FindId(sparse, true);
-    return dense;
-  }
-
-  // Appends after `output` the state tuple corresponding to the state id. The
-  // state id must exist.
-  void FillState(StateId s, std::vector<Label> *output) {
-    s = condensed_.FindEntry(s);
-    for (NGramIterator it = ngrams_.FindSet(s); !it.Done(); it.Next()) {
-      Label label = it.Element();
-      output->push_back(label);
-    }
-  }
-
-  // Shifts the buffer in `state` by appending `ilabel` and popping
-  // the one in the front as the return value. `next_stub_` is a
-  // shifted buffer of size `delay_` where the first `delay_ - 1`
-  // elements are the last `delay_ - 1` elements in the buffer of
-  // `state`. The last (if any) element in `next_stub_` will be
-  // `ilabel` after the call returns.
-  Label ShiftBuffer(const std::vector<Label> &state, Label ilabel,
-                    std::vector<Label> *next_stub_);
-
-  // Builds an arc from state tuple `state` consuming `ilabel` and
-  // `olabel`. `next_stub_` is the buffer filled in `ShiftBuffer`.
-  Arc MakeArc(const std::vector<Label> &state, Label ilabel, Label olabel,
-              std::vector<Label> *next_stub_);
-
-  // Expands arcs from state `s`, equivalent to state tuple `state`,
-  // with input `ilabel`. `next_stub_` is the buffer filled in
-  // `ShiftBuffer`.
-  void ExpandArcs(StateId s, const std::vector<Label> &state, Label ilabel,
-                  std::vector<Label> *next_stub_);
-
-  // Appends arcs from state `s`, equivalent to state tuple `state`,
-  // with input `ilabel` to `arcs`. `next_stub_` is the buffer filled
-  // in `ShiftBuffer`.
-  void AppendArcs(StateId s, const std::vector<Label> &state, Label ilabel,
-                  std::vector<Label> *next_stub_, std::vector<Arc> *arcs);
-
-  std::shared_ptr<const LinearFstData<A>> data_;
-  size_t delay_;
-  // Mapping from internal state tuple to *non-consecutive* ids
-  Collection<StateId, Label> ngrams_;
-  // Mapping from non-consecutive id to actual state id
-  CompactHashBiTable<StateId, StateId, std::hash<StateId>> condensed_;
-  // Two frequently used vectors, reuse to avoid repeated heap
-  // allocation
-  std::vector<Label> state_stub_, next_stub_;
-
-  LinearTaggerFstImpl &operator=(const LinearTaggerFstImpl &) = delete;
-};
-
-template <class A>
-const int LinearTaggerFstImpl<A>::kMinFileVersion = 1;
-
-template <class A>
-const int LinearTaggerFstImpl<A>::kFileVersion = 1;
-
-template <class A>
-inline typename A::Label LinearTaggerFstImpl<A>::ShiftBuffer(
-    const std::vector<Label> &state, Label ilabel,
-    std::vector<Label> *next_stub_) {
-  DCHECK(ilabel > 0 || ilabel == LinearFstData<A>::kEndOfSentence);
-  if (delay_ == 0) {
-    DCHECK_GT(ilabel, 0);
-    return ilabel;
-  } else {
-    (*next_stub_)[BufferEnd(*next_stub_) - next_stub_->begin() - 1] = ilabel;
-    return *BufferBegin(state);
-  }
-}
-
-template <class A>
-inline A LinearTaggerFstImpl<A>::MakeArc(const std::vector<Label> &state,
-                                         Label ilabel, Label olabel,
-                                         std::vector<Label> *next_stub_) {
-  DCHECK(ilabel > 0 || ilabel == LinearFstData<A>::kEndOfSentence);
-  DCHECK(olabel > 0 || olabel == LinearFstData<A>::kStartOfSentence);
-  Weight weight(Weight::One());
-  data_->TakeTransition(BufferEnd(state), InternalBegin(state),
-                        InternalEnd(state), ilabel, olabel, next_stub_,
-                        &weight);
-  StateId nextstate = FindState(*next_stub_);
-  // Restore `next_stub_` to its size before the call
-  next_stub_->resize(delay_);
-  // In the actual arc, we use epsilons instead of boundaries.
-  return A(ilabel == LinearFstData<A>::kEndOfSentence ? 0 : ilabel,
-           olabel == LinearFstData<A>::kStartOfSentence ? 0 : olabel, weight,
-           nextstate);
-}
-
-template <class A>
-inline void LinearTaggerFstImpl<A>::ExpandArcs(StateId s,
-                                               const std::vector<Label> &state,
-                                               Label ilabel,
-                                               std::vector<Label> *next_stub_) {
-  // Input label to constrain the output with, observed `delay_` steps
-  // back. `ilabel` is the input label to be put on the arc, which
-  // fires features.
-  Label obs_ilabel = ShiftBuffer(state, ilabel, next_stub_);
-  if (obs_ilabel == LinearFstData<A>::kStartOfSentence) {
-    // This happens when input is shorter than `delay_`.
-    PushArc(s, MakeArc(state, ilabel, LinearFstData<A>::kStartOfSentence,
-                       next_stub_));
-  } else {
-    std::pair<typename std::vector<typename A::Label>::const_iterator,
-              typename std::vector<typename A::Label>::const_iterator> range =
-        data_->PossibleOutputLabels(obs_ilabel);
-    for (typename std::vector<typename A::Label>::const_iterator it =
-             range.first;
-         it != range.second; ++it)
-      PushArc(s, MakeArc(state, ilabel, *it, next_stub_));
-  }
-}
-
-// TODO(wuke): this has much in duplicate with `ExpandArcs()`
-template <class A>
-inline void LinearTaggerFstImpl<A>::AppendArcs(StateId /*s*/,
-                                               const std::vector<Label> &state,
-                                               Label ilabel,
-                                               std::vector<Label> *next_stub_,
-                                               std::vector<Arc> *arcs) {
-  // Input label to constrain the output with, observed `delay_` steps
-  // back. `ilabel` is the input label to be put on the arc, which
-  // fires features.
-  Label obs_ilabel = ShiftBuffer(state, ilabel, next_stub_);
-  if (obs_ilabel == LinearFstData<A>::kStartOfSentence) {
-    // This happens when input is shorter than `delay_`.
-    arcs->push_back(
-        MakeArc(state, ilabel, LinearFstData<A>::kStartOfSentence, next_stub_));
-  } else {
-    std::pair<typename std::vector<typename A::Label>::const_iterator,
-              typename std::vector<typename A::Label>::const_iterator> range =
-        data_->PossibleOutputLabels(obs_ilabel);
-    for (typename std::vector<typename A::Label>::const_iterator it =
-             range.first;
-         it != range.second; ++it)
-      arcs->push_back(MakeArc(state, ilabel, *it, next_stub_));
-  }
-}
-
-template <class A>
-void LinearTaggerFstImpl<A>::Expand(StateId s) {
-  VLOG(3) << "Expand " << s;
-  state_stub_.clear();
-  FillState(s, &state_stub_);
-
-  // Precompute the first `delay_ - 1` elements in the buffer of
-  // next states, which are identical for different input/output.
-  next_stub_.clear();
-  next_stub_.resize(delay_);
-  if (delay_ > 0)
-    std::copy(BufferBegin(state_stub_) + 1, BufferEnd(state_stub_),
-              next_stub_.begin());
-
-  // Epsilon transition for flushing out the next observed input
-  if (!IsEmptyBuffer(BufferBegin(state_stub_), BufferEnd(state_stub_)))
-    ExpandArcs(s, state_stub_, LinearFstData<A>::kEndOfSentence, &next_stub_);
-
-  // Non-epsilon input when we haven't flushed
-  if (delay_ == 0 ||
-      *(BufferEnd(state_stub_) - 1) != LinearFstData<A>::kEndOfSentence)
-    for (Label ilabel = data_->MinInputLabel();
-         ilabel <= data_->MaxInputLabel(); ++ilabel)
-      ExpandArcs(s, state_stub_, ilabel, &next_stub_);
-
-  SetArcs(s);
-}
-
-template <class A>
-void LinearTaggerFstImpl<A>::MatchInput(StateId s, Label ilabel,
-                                        std::vector<Arc> *arcs) {
-  state_stub_.clear();
-  FillState(s, &state_stub_);
-
-  // Precompute the first `delay_ - 1` elements in the buffer of
-  // next states, which are identical for different input/output.
-  next_stub_.clear();
-  next_stub_.resize(delay_);
-  if (delay_ > 0)
-    std::copy(BufferBegin(state_stub_) + 1, BufferEnd(state_stub_),
-              next_stub_.begin());
-
-  if (ilabel == 0) {
-    // Epsilon transition for flushing out the next observed input
-    if (!IsEmptyBuffer(BufferBegin(state_stub_), BufferEnd(state_stub_)))
-      AppendArcs(s, state_stub_, LinearFstData<A>::kEndOfSentence, &next_stub_,
-                 arcs);
-  } else {
-    // Non-epsilon input when we haven't flushed
-    if (delay_ == 0 ||
-        *(BufferEnd(state_stub_) - 1) != LinearFstData<A>::kEndOfSentence)
-      AppendArcs(s, state_stub_, ilabel, &next_stub_, arcs);
-  }
-}
-
-template <class A>
-inline LinearTaggerFstImpl<A> *LinearTaggerFstImpl<A>::Read(
-    std::istream &strm, const FstReadOptions &opts) {  // NOLINT
-  std::unique_ptr<LinearTaggerFstImpl<A>> impl(new LinearTaggerFstImpl<A>());
-  FstHeader header;
-  if (!impl->ReadHeader(strm, opts, kMinFileVersion, &header)) {
-    return nullptr;
-  }
-  impl->data_ = std::shared_ptr<LinearFstData<A>>(LinearFstData<A>::Read(strm));
-  if (!impl->data_) {
-    return nullptr;
-  }
-  impl->delay_ = impl->data_->MaxFutureSize();
-  impl->ReserveStubSpace();
-  return impl.release();
-}
-
-}  // namespace internal
-
-// This class attaches interface to implementation and handles
-// reference counting, delegating most methods to ImplToFst.
-template <class A>
-class LinearTaggerFst : public ImplToFst<internal::LinearTaggerFstImpl<A>> {
- public:
-  friend class ArcIterator<LinearTaggerFst<A>>;
-  friend class StateIterator<LinearTaggerFst<A>>;
-  friend class LinearFstMatcherTpl<LinearTaggerFst<A>>;
-
-  typedef A Arc;
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-  typedef typename A::StateId StateId;
-  typedef DefaultCacheStore<A> Store;
-  typedef typename Store::State State;
-  using Impl = internal::LinearTaggerFstImpl<A>;
-
-  LinearTaggerFst() : ImplToFst<Impl>(std::make_shared<Impl>()) {}
-
-  explicit LinearTaggerFst(LinearFstData<A> *data,
-                           const SymbolTable *isyms = nullptr,
-                           const SymbolTable *osyms = nullptr,
-                           CacheOptions opts = CacheOptions())
-      : ImplToFst<Impl>(std::make_shared<Impl>(data, isyms, osyms, opts)) {}
-
-  explicit LinearTaggerFst(const Fst<A> &fst)
-      : ImplToFst<Impl>(std::make_shared<Impl>()) {
-    LOG(FATAL) << "LinearTaggerFst: no constructor from arbitrary FST.";
-  }
-
-  // See Fst<>::Copy() for doc.
-  LinearTaggerFst(const LinearTaggerFst<A> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this LinearTaggerFst. See Fst<>::Copy() for further doc.
-  LinearTaggerFst<A> *Copy(bool safe = false) const override {
-    return new LinearTaggerFst<A>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<A> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<A> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
-  MatcherBase<A> *InitMatcher(MatchType match_type) const override {
-    return new LinearFstMatcherTpl<LinearTaggerFst<A>>(this, match_type);
-  }
-
-  static LinearTaggerFst<A> *Read(const string &filename) {
-    if (!filename.empty()) {
-      std::ifstream strm(filename,
-                              std::ios_base::in | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "LinearTaggerFst::Read: Can't open file: " << filename;
-        return nullptr;
-      }
-      return Read(strm, FstReadOptions(filename));
-    } else {
-      return Read(std::cin, FstReadOptions("standard input"));
-    }
-  }
-
-  static LinearTaggerFst<A> *Read(std::istream &in,  // NOLINT
-                                  const FstReadOptions &opts) {
-    auto *impl = Impl::Read(in, opts);
-    return impl ? new LinearTaggerFst<A>(std::shared_ptr<Impl>(impl)) : nullptr;
-  }
-
-  bool Write(const string &filename) const override {
-    if (!filename.empty()) {
-      std::ofstream strm(filename,
-                               std::ios_base::out | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "LinearTaggerFst::Write: Can't open file: " << filename;
-        return false;
-      }
-      return Write(strm, FstWriteOptions(filename));
-    } else {
-      return Write(std::cout, FstWriteOptions("standard output"));
-    }
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return GetImpl()->Write(strm, opts);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  explicit LinearTaggerFst(std::shared_ptr<Impl> impl)
-      : ImplToFst<Impl>(impl) {}
-
-  void operator=(const LinearTaggerFst<A> &fst) = delete;
-};
-
-// Specialization for LinearTaggerFst.
-template <class Arc>
-class StateIterator<LinearTaggerFst<Arc>>
-    : public CacheStateIterator<LinearTaggerFst<Arc>> {
- public:
-  explicit StateIterator(const LinearTaggerFst<Arc> &fst)
-      : CacheStateIterator<LinearTaggerFst<Arc>>(fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for LinearTaggerFst.
-template <class Arc>
-class ArcIterator<LinearTaggerFst<Arc>>
-    : public CacheArcIterator<LinearTaggerFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const LinearTaggerFst<Arc> &fst, StateId s)
-      : CacheArcIterator<LinearTaggerFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc>
-inline void LinearTaggerFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<LinearTaggerFst<Arc>>(*this);
-}
-
-namespace internal {
-
-// Implementation class for on-the-fly generated LinearClassifierFst with
-// special optimization in matching.
-template <class A>
-class LinearClassifierFstImpl : public CacheImpl<A> {
- public:
-  using FstImpl<A>::SetType;
-  using FstImpl<A>::SetProperties;
-  using FstImpl<A>::SetInputSymbols;
-  using FstImpl<A>::SetOutputSymbols;
-  using FstImpl<A>::WriteHeader;
-
-  using CacheBaseImpl<CacheState<A>>::PushArc;
-  using CacheBaseImpl<CacheState<A>>::HasArcs;
-  using CacheBaseImpl<CacheState<A>>::HasFinal;
-  using CacheBaseImpl<CacheState<A>>::HasStart;
-  using CacheBaseImpl<CacheState<A>>::SetArcs;
-  using CacheBaseImpl<CacheState<A>>::SetFinal;
-  using CacheBaseImpl<CacheState<A>>::SetStart;
-
-  typedef A Arc;
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-  typedef typename A::StateId StateId;
-  typedef typename Collection<StateId, Label>::SetIterator NGramIterator;
-
-  // Constructs an empty FST by default.
-  LinearClassifierFstImpl()
-      : CacheImpl<A>(CacheOptions()),
-        data_(std::make_shared<LinearFstData<A>>()) {
-    SetType("linear-classifier");
-    num_classes_ = 0;
-    num_groups_ = 0;
-  }
-
-  // Constructs the FST with given data storage, number of classes and
-  // symbol tables.
-  LinearClassifierFstImpl(const LinearFstData<Arc> *data, size_t num_classes,
-                          const SymbolTable *isyms, const SymbolTable *osyms,
-                          CacheOptions opts)
-      : CacheImpl<A>(opts),
-        data_(data),
-        num_classes_(num_classes),
-        num_groups_(data_->NumGroups() / num_classes_) {
-    SetType("linear-classifier");
-    SetProperties(kILabelSorted, kFstProperties);
-    SetInputSymbols(isyms);
-    SetOutputSymbols(osyms);
-    ReserveStubSpace();
-  }
-
-  // Copy by sharing the underlying data storage.
-  LinearClassifierFstImpl(const LinearClassifierFstImpl &impl)
-      : CacheImpl<A>(impl),
-        data_(impl.data_),
-        num_classes_(impl.num_classes_),
-        num_groups_(impl.num_groups_) {
-    SetType("linear-classifier");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-    ReserveStubSpace();
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      StateId start = FindStartState();
-      SetStart(start);
-    }
-    return CacheImpl<A>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) {
-      state_stub_.clear();
-      FillState(s, &state_stub_);
-      SetFinal(s, FinalWeight(state_stub_));
-    }
-    return CacheImpl<A>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<A>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<A>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<A>::NumOutputEpsilons(s);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<A> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<A>::InitArcIterator(s, data);
-  }
-
-  // Computes the outgoing transitions from a state, creating new
-  // destination states as needed.
-  void Expand(StateId s);
-
-  // Appends to `arcs` all out-going arcs from state `s` that matches
-  // `label` as the input label.
-  void MatchInput(StateId s, Label ilabel, std::vector<Arc> *arcs);
-
-  static LinearClassifierFstImpl<A> *Read(std::istream &strm,
-                                          const FstReadOptions &opts);
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
-    FstHeader header;
-    header.SetStart(kNoStateId);
-    WriteHeader(strm, opts, kFileVersion, &header);
-    data_->Write(strm);
-    WriteType(strm, num_classes_);
-    if (!strm) {
-      LOG(ERROR) << "LinearClassifierFst::Write: Write failed: " << opts.source;
-      return false;
-    }
-    return true;
-  }
-
- private:
-  static const int kMinFileVersion;
-  static const int kFileVersion;
-
-  // A collection of functions to access parts of the state tuple. A
-  // state tuple is a vector of `Label`s with two parts:
-  //   [prediction] [internal].
-  //
-  // - [prediction] is a single label of the predicted class. A state
-  //   must have a positive class label, unless it is the start state.
-  //
-  // - [internal] is the internal state tuple for `LinearFstData` of
-  //   the given class; or kNoTrieNodeId's if in start state.
-  Label &Prediction(std::vector<Label> &state) { return state[0]; }  // NOLINT
-  Label Prediction(const std::vector<Label> &state) const { return state[0]; }
-
-  Label &InternalAt(std::vector<Label> &state, int index) {  // NOLINT
-    return state[index + 1];
-  }
-  Label InternalAt(const std::vector<Label> &state, int index) const {
-    return state[index + 1];
-  }
-
-  // The size of state tuples are fixed, reserve them in stubs
-  void ReserveStubSpace() {
-    size_t size = 1 + num_groups_;
-    state_stub_.reserve(size);
-    next_stub_.reserve(size);
-  }
-
-  // Computes the start state tuple and maps it to the start state id.
-  StateId FindStartState() {
-    // A start state tuple has no prediction
-    state_stub_.clear();
-    state_stub_.push_back(kNoLabel);
-    // For a start state, we don't yet know where we are in the tries.
-    for (size_t i = 0; i < num_groups_; ++i)
-      state_stub_.push_back(kNoTrieNodeId);
-    return FindState(state_stub_);
-  }
-
-  // Tests if the state tuple represents the start state.
-  bool IsStartState(const std::vector<Label> &state) const {
-    return state[0] == kNoLabel;
-  }
-
-  // Computes the actual group id in the data storage.
-  int GroupId(Label pred, int group) const {
-    return group * num_classes_ + pred - 1;
-  }
-
-  // Finds out the final weight of the given state. A state is final
-  // iff it is not the start.
-  Weight FinalWeight(const std::vector<Label> &state) const {
-    if (IsStartState(state)) {
-      return Weight::Zero();
-    }
-    Label pred = Prediction(state);
-    DCHECK_GT(pred, 0);
-    DCHECK_LE(pred, num_classes_);
-    Weight final_weight = Weight::One();
-    for (size_t group = 0; group < num_groups_; ++group) {
-      int group_id = GroupId(pred, group);
-      int trie_state = InternalAt(state, group);
-      final_weight =
-          Times(final_weight, data_->GroupFinalWeight(group_id, trie_state));
-    }
-    return final_weight;
-  }
-
-  // Finds state corresponding to an n-gram. Creates new state if n-gram not
-  // found.
-  StateId FindState(const std::vector<Label> &ngram) {
-    StateId sparse = ngrams_.FindId(ngram, true);
-    StateId dense = condensed_.FindId(sparse, true);
-    return dense;
-  }
-
-  // Appends after `output` the state tuple corresponding to the state id. The
-  // state id must exist.
-  void FillState(StateId s, std::vector<Label> *output) {
-    s = condensed_.FindEntry(s);
-    for (NGramIterator it = ngrams_.FindSet(s); !it.Done(); it.Next()) {
-      Label label = it.Element();
-      output->push_back(label);
-    }
-  }
-
-  std::shared_ptr<const LinearFstData<A>> data_;
-  // Division of groups in `data_`; num_classes_ * num_groups_ ==
-  // data_->NumGroups().
-  size_t num_classes_, num_groups_;
-  // Mapping from internal state tuple to *non-consecutive* ids
-  Collection<StateId, Label> ngrams_;
-  // Mapping from non-consecutive id to actual state id
-  CompactHashBiTable<StateId, StateId, std::hash<StateId>> condensed_;
-  // Two frequently used vectors, reuse to avoid repeated heap
-  // allocation
-  std::vector<Label> state_stub_, next_stub_;
-
-  void operator=(const LinearClassifierFstImpl<A> &) = delete;
-};
-
-template <class A>
-const int LinearClassifierFstImpl<A>::kMinFileVersion = 0;
-
-template <class A>
-const int LinearClassifierFstImpl<A>::kFileVersion = 0;
-
-template <class A>
-void LinearClassifierFstImpl<A>::Expand(StateId s) {
-  VLOG(3) << "Expand " << s;
-  state_stub_.clear();
-  FillState(s, &state_stub_);
-  next_stub_.clear();
-  next_stub_.resize(1 + num_groups_);
-
-  if (IsStartState(state_stub_)) {
-    // Make prediction
-    for (Label pred = 1; pred <= num_classes_; ++pred) {
-      Prediction(next_stub_) = pred;
-      for (int i = 0; i < num_groups_; ++i)
-        InternalAt(next_stub_, i) = data_->GroupStartState(GroupId(pred, i));
-      PushArc(s, A(0, pred, Weight::One(), FindState(next_stub_)));
-    }
-  } else {
-    Label pred = Prediction(state_stub_);
-    DCHECK_GT(pred, 0);
-    DCHECK_LE(pred, num_classes_);
-    for (Label ilabel = data_->MinInputLabel();
-         ilabel <= data_->MaxInputLabel(); ++ilabel) {
-      Prediction(next_stub_) = pred;
-      Weight weight = Weight::One();
-      for (int i = 0; i < num_groups_; ++i)
-        InternalAt(next_stub_, i) =
-            data_->GroupTransition(GroupId(pred, i), InternalAt(state_stub_, i),
-                                   ilabel, pred, &weight);
-      PushArc(s, A(ilabel, 0, weight, FindState(next_stub_)));
-    }
-  }
-
-  SetArcs(s);
-}
-
-template <class A>
-void LinearClassifierFstImpl<A>::MatchInput(StateId s, Label ilabel,
-                                            std::vector<Arc> *arcs) {
-  state_stub_.clear();
-  FillState(s, &state_stub_);
-  next_stub_.clear();
-  next_stub_.resize(1 + num_groups_);
-
-  if (IsStartState(state_stub_)) {
-    // Make prediction if `ilabel` is epsilon.
-    if (ilabel == 0) {
-      for (Label pred = 1; pred <= num_classes_; ++pred) {
-        Prediction(next_stub_) = pred;
-        for (int i = 0; i < num_groups_; ++i)
-          InternalAt(next_stub_, i) = data_->GroupStartState(GroupId(pred, i));
-        arcs->push_back(A(0, pred, Weight::One(), FindState(next_stub_)));
-      }
-    }
-  } else if (ilabel != 0) {
-    Label pred = Prediction(state_stub_);
-    Weight weight = Weight::One();
-    Prediction(next_stub_) = pred;
-    for (int i = 0; i < num_groups_; ++i)
-      InternalAt(next_stub_, i) = data_->GroupTransition(
-          GroupId(pred, i), InternalAt(state_stub_, i), ilabel, pred, &weight);
-    arcs->push_back(A(ilabel, 0, weight, FindState(next_stub_)));
-  }
-}
-
-template <class A>
-inline LinearClassifierFstImpl<A> *LinearClassifierFstImpl<A>::Read(
-    std::istream &strm, const FstReadOptions &opts) {
-  std::unique_ptr<LinearClassifierFstImpl<A>> impl(
-      new LinearClassifierFstImpl<A>());
-  FstHeader header;
-  if (!impl->ReadHeader(strm, opts, kMinFileVersion, &header)) {
-    return nullptr;
-  }
-  impl->data_ = std::shared_ptr<LinearFstData<A>>(LinearFstData<A>::Read(strm));
-  if (!impl->data_) {
-    return nullptr;
-  }
-  ReadType(strm, &impl->num_classes_);
-  if (!strm) {
-    return nullptr;
-  }
-  impl->num_groups_ = impl->data_->NumGroups() / impl->num_classes_;
-  if (impl->num_groups_ * impl->num_classes_ != impl->data_->NumGroups()) {
-    FSTERROR() << "Total number of feature groups is not a multiple of the "
-                  "number of classes: num groups = "
-               << impl->data_->NumGroups()
-               << ", num classes = " << impl->num_classes_;
-    return nullptr;
-  }
-  impl->ReserveStubSpace();
-  return impl.release();
-}
-
-}  // namespace internal
-
-// This class attaches interface to implementation and handles
-// reference counting, delegating most methods to ImplToFst.
-template <class A>
-class LinearClassifierFst
-    : public ImplToFst<internal::LinearClassifierFstImpl<A>> {
- public:
-  friend class ArcIterator<LinearClassifierFst<A>>;
-  friend class StateIterator<LinearClassifierFst<A>>;
-  friend class LinearFstMatcherTpl<LinearClassifierFst<A>>;
-
-  typedef A Arc;
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-  typedef typename A::StateId StateId;
-  typedef DefaultCacheStore<A> Store;
-  typedef typename Store::State State;
-  using Impl = internal::LinearClassifierFstImpl<A>;
-
-  LinearClassifierFst() : ImplToFst<Impl>(std::make_shared<Impl>()) {}
-
-  explicit LinearClassifierFst(LinearFstData<A> *data, size_t num_classes,
-                               const SymbolTable *isyms = nullptr,
-                               const SymbolTable *osyms = nullptr,
-                               CacheOptions opts = CacheOptions())
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(data, num_classes, isyms, osyms, opts)) {}
-
-  explicit LinearClassifierFst(const Fst<A> &fst)
-      : ImplToFst<Impl>(std::make_shared<Impl>()) {
-    LOG(FATAL) << "LinearClassifierFst: no constructor from arbitrary FST.";
-  }
-
-  // See Fst<>::Copy() for doc.
-  LinearClassifierFst(const LinearClassifierFst<A> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this LinearClassifierFst. See Fst<>::Copy() for further doc.
-  LinearClassifierFst<A> *Copy(bool safe = false) const override {
-    return new LinearClassifierFst<A>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<A> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<A> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
-  MatcherBase<A> *InitMatcher(MatchType match_type) const override {
-    return new LinearFstMatcherTpl<LinearClassifierFst<A>>(this, match_type);
-  }
-
-  static LinearClassifierFst<A> *Read(const string &filename) {
-    if (!filename.empty()) {
-      std::ifstream strm(filename,
-                              std::ios_base::in | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "LinearClassifierFst::Read: Can't open file: "
-                   << filename;
-        return nullptr;
-      }
-      return Read(strm, FstReadOptions(filename));
-    } else {
-      return Read(std::cin, FstReadOptions("standard input"));
-    }
-  }
-
-  static LinearClassifierFst<A> *Read(std::istream &in,
-                                      const FstReadOptions &opts) {
-    auto *impl = Impl::Read(in, opts);
-    return impl ? new LinearClassifierFst<A>(std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  bool Write(const string &filename) const override {
-    if (!filename.empty()) {
-      std::ofstream strm(filename,
-                               std::ios_base::out | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "ProdLmFst::Write: Can't open file: " << filename;
-        return false;
-      }
-      return Write(strm, FstWriteOptions(filename));
-    } else {
-      return Write(std::cout, FstWriteOptions("standard output"));
-    }
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return GetImpl()->Write(strm, opts);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  explicit LinearClassifierFst(std::shared_ptr<Impl> impl)
-      : ImplToFst<Impl>(impl) {}
-
-  void operator=(const LinearClassifierFst<A> &fst) = delete;
-};
-
-// Specialization for LinearClassifierFst.
-template <class Arc>
-class StateIterator<LinearClassifierFst<Arc>>
-    : public CacheStateIterator<LinearClassifierFst<Arc>> {
- public:
-  explicit StateIterator(const LinearClassifierFst<Arc> &fst)
-      : CacheStateIterator<LinearClassifierFst<Arc>>(fst,
-                                                     fst.GetMutableImpl()) {}
-};
-
-// Specialization for LinearClassifierFst.
-template <class Arc>
-class ArcIterator<LinearClassifierFst<Arc>>
-    : public CacheArcIterator<LinearClassifierFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const LinearClassifierFst<Arc> &fst, StateId s)
-      : CacheArcIterator<LinearClassifierFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc>
-inline void LinearClassifierFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<LinearClassifierFst<Arc>>(*this);
-}
-
-// Specialized Matcher for LinearFsts. This matcher only supports
-// matching from the input side. This is intentional because comparing
-// the scores of different input sequences with the same output
-// sequence is meaningless in a discriminative model.
-template <class F>
-class LinearFstMatcherTpl : public MatcherBase<typename F::Arc> {
- public:
-  typedef typename F::Arc Arc;
-  typedef typename Arc::Label Label;
-  typedef typename Arc::Weight Weight;
-  typedef typename Arc::StateId StateId;
-  typedef F FST;
-
-  // This makes a copy of the FST.
-  LinearFstMatcherTpl(const FST &fst, MatchType match_type)
-      : owned_fst_(fst.Copy()),
-        fst_(*owned_fst_),
-        match_type_(match_type),
-        s_(kNoStateId),
-        current_loop_(false),
-        loop_(kNoLabel, 0, Weight::One(), kNoStateId),
-        cur_arc_(0),
-        error_(false) {
-    switch (match_type_) {
-      case MATCH_INPUT:
-      case MATCH_OUTPUT:
-      case MATCH_NONE:
-        break;
-      default:
-        FSTERROR() << "LinearFstMatcherTpl: Bad match type";
-        match_type_ = MATCH_NONE;
-        error_ = true;
-    }
-  }
-
-  // This doesn't copy the FST.
-  LinearFstMatcherTpl(const FST *fst, MatchType match_type)
-      : fst_(*fst),
-        match_type_(match_type),
-        s_(kNoStateId),
-        current_loop_(false),
-        loop_(kNoLabel, 0, Weight::One(), kNoStateId),
-        cur_arc_(0),
-        error_(false) {
-    switch (match_type_) {
-      case MATCH_INPUT:
-      case MATCH_OUTPUT:
-      case MATCH_NONE:
-        break;
-      default:
-        FSTERROR() << "LinearFstMatcherTpl: Bad match type";
-        match_type_ = MATCH_NONE;
-        error_ = true;
-    }
-  }
-
-  // This makes a copy of the FST.
-  LinearFstMatcherTpl(const LinearFstMatcherTpl<F> &matcher, bool safe = false)
-      : owned_fst_(matcher.fst_.Copy(safe)),
-        fst_(*owned_fst_),
-        match_type_(matcher.match_type_),
-        s_(kNoStateId),
-        current_loop_(false),
-        loop_(matcher.loop_),
-        cur_arc_(0),
-        error_(matcher.error_) {}
-
-  LinearFstMatcherTpl<F> *Copy(bool safe = false) const override {
-    return new LinearFstMatcherTpl<F>(*this, safe);
-  }
-
-  MatchType Type(bool /*test*/) const override {
-    // `MATCH_INPUT` is the only valid type
-    return match_type_ == MATCH_INPUT ? match_type_ : MATCH_NONE;
-  }
-
-  void SetState(StateId s) final {
-    if (s_ == s) return;
-    s_ = s;
-    // `MATCH_INPUT` is the only valid type
-    if (match_type_ != MATCH_INPUT) {
-      FSTERROR() << "LinearFstMatcherTpl: Bad match type";
-      error_ = true;
-    }
-    loop_.nextstate = s;
-  }
-
-  bool Find(Label label) final {
-    if (error_) {
-      current_loop_ = false;
-      return false;
-    }
-    current_loop_ = label == 0;
-    if (label == kNoLabel) label = 0;
-    arcs_.clear();
-    cur_arc_ = 0;
-    fst_.GetMutableImpl()->MatchInput(s_, label, &arcs_);
-    return current_loop_ || !arcs_.empty();
-  }
-
-  bool Done() const final {
-    return !(current_loop_ || cur_arc_ < arcs_.size());
-  }
-
-  const Arc &Value() const final {
-    return current_loop_ ? loop_ : arcs_[cur_arc_];
-  }
-
-  void Next() final {
-    if (current_loop_)
-      current_loop_ = false;
-    else
-      ++cur_arc_;
-  }
-
-  ssize_t Priority(StateId s) final { return kRequirePriority; }
-
-  const FST &GetFst() const override { return fst_; }
-
-  uint64 Properties(uint64 props) const override {
-    if (error_) props |= kError;
-    return props;
-  }
-
-  uint32 Flags() const override { return kRequireMatch; }
-
- private:
-  std::unique_ptr<const FST> owned_fst_;
-  const FST &fst_;
-  MatchType match_type_;  // Type of match to perform.
-  StateId s_;             // Current state.
-  bool current_loop_;     // Current arc is the implicit loop.
-  Arc loop_;              // For non-consuming symbols.
-  // All out-going arcs matching the label in last Find() call.
-  std::vector<Arc> arcs_;
-  size_t cur_arc_;  // Index to the arc that `Value()` should return.
-  bool error_;      // Error encountered.
-};
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_LINEAR_LINEAR_FST_H_
diff --git a/projects/llm_framework/include/fst/extensions/linear/linearscript.h b/projects/llm_framework/include/fst/extensions/linear/linearscript.h
deleted file mode 100644
index 54106d20..00000000
--- a/projects/llm_framework/include/fst/extensions/linear/linearscript.h
+++ /dev/null
@@ -1,391 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_LINEAR_LINEARSCRIPT_H_
-#define FST_EXTENSIONS_LINEAR_LINEARSCRIPT_H_
-
-#include <istream>
-#include <sstream>
-#include <string>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/extensions/linear/linear-fst-data-builder.h>
-#include <fst/extensions/linear/linear-fst.h>
-#include <fstream>
-#include <fst/symbol-table.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/script-impl.h>
-
-DECLARE_string(delimiter);
-DECLARE_string(empty_symbol);
-DECLARE_string(start_symbol);
-DECLARE_string(end_symbol);
-DECLARE_bool(classifier);
-
-namespace fst {
-namespace script {
-typedef std::tuple<const string &, const string &, const string &, char **, int,
-                   const string &, const string &, const string &,
-                   const string &>
-    LinearCompileArgs;
-
-bool ValidateDelimiter();
-bool ValidateEmptySymbol();
-
-// Returns the proper label given the symbol. For symbols other than
-// `FLAGS_start_symbol` or `FLAGS_end_symbol`, looks up the symbol
-// table to decide the label. Depending on whether
-// `FLAGS_start_symbol` and `FLAGS_end_symbol` are identical, it
-// either returns `kNoLabel` for later processing or decides the label
-// right away.
-template <class Arc>
-inline typename Arc::Label LookUp(const string &str, SymbolTable *syms) {
-  if (str == FLAGS_start_symbol)
-    return str == FLAGS_end_symbol ? kNoLabel
-                                   : LinearFstData<Arc>::kStartOfSentence;
-  else if (str == FLAGS_end_symbol)
-    return LinearFstData<Arc>::kEndOfSentence;
-  else
-    return syms->AddSymbol(str);
-}
-
-// Splits `str` with `delim` as the delimiter and stores the labels in
-// `output`.
-template <class Arc>
-void SplitAndPush(const string &str, const char delim, SymbolTable *syms,
-                  std::vector<typename Arc::Label> *output) {
-  if (str == FLAGS_empty_symbol) return;
-  std::istringstream strm(str);
-  string buf;
-  while (std::getline(strm, buf, delim))
-    output->push_back(LookUp<Arc>(buf, syms));
-}
-
-// Like `std::replace_copy` but returns the number of modifications
-template <class InputIterator, class OutputIterator, class T>
-size_t ReplaceCopy(InputIterator first, InputIterator last,
-                   OutputIterator result, const T &old_value,
-                   const T &new_value) {
-  size_t changes = 0;
-  while (first != last) {
-    if (*first == old_value) {
-      *result = new_value;
-      ++changes;
-    } else {
-      *result = *first;
-    }
-    ++first;
-    ++result;
-  }
-  return changes;
-}
-
-template <class Arc>
-bool GetVocabRecord(const string &vocab, std::istream &strm,  // NOLINT
-                    SymbolTable *isyms, SymbolTable *fsyms, SymbolTable *osyms,
-                    typename Arc::Label *word,
-                    std::vector<typename Arc::Label> *feature_labels,
-                    std::vector<typename Arc::Label> *possible_labels,
-                    size_t *num_line);
-
-template <class Arc>
-bool GetModelRecord(const string &model, std::istream &strm,  // NOLINT
-                    SymbolTable *fsyms, SymbolTable *osyms,
-                    std::vector<typename Arc::Label> *input_labels,
-                    std::vector<typename Arc::Label> *output_labels,
-                    typename Arc::Weight *weight, size_t *num_line);
-
-// Reads in vocabulary file. Each line is in the following format
-//
-//   word <whitespace> features [ <whitespace> possible output ]
-//
-// where features and possible output are `FLAGS_delimiter`-delimited lists of
-// tokens
-template <class Arc>
-void AddVocab(const string &vocab, SymbolTable *isyms, SymbolTable *fsyms,
-              SymbolTable *osyms, LinearFstDataBuilder<Arc> *builder) {
-  std::ifstream in(vocab);
-  if (!in) LOG(FATAL) << "Can't open file: " << vocab;
-  size_t num_line = 0, num_added = 0;
-  std::vector<string> fields;
-  std::vector<typename Arc::Label> feature_labels, possible_labels;
-  typename Arc::Label word;
-  while (GetVocabRecord<Arc>(vocab, in, isyms, fsyms, osyms, &word,
-                             &feature_labels, &possible_labels, &num_line)) {
-    if (word == kNoLabel) {
-      LOG(WARNING) << "Ignored: boundary word: " << fields[0];
-      continue;
-    }
-    if (possible_labels.empty())
-      num_added += builder->AddWord(word, feature_labels);
-    else
-      num_added += builder->AddWord(word, feature_labels, possible_labels);
-  }
-  VLOG(1) << "Read " << num_added << " words in " << num_line << " lines from "
-          << vocab;
-}
-
-template <class Arc>
-void AddVocab(const string &vocab, SymbolTable *isyms, SymbolTable *fsyms,
-              SymbolTable *osyms,
-              LinearClassifierFstDataBuilder<Arc> *builder) {
-  std::ifstream in(vocab);
-  if (!in) LOG(FATAL) << "Can't open file: " << vocab;
-  size_t num_line = 0, num_added = 0;
-  std::vector<string> fields;
-  std::vector<typename Arc::Label> feature_labels, possible_labels;
-  typename Arc::Label word;
-  while (GetVocabRecord<Arc>(vocab, in, isyms, fsyms, osyms, &word,
-                             &feature_labels, &possible_labels, &num_line)) {
-    if (!possible_labels.empty())
-      LOG(FATAL)
-          << "Classifier vocabulary should not have possible output constraint";
-    if (word == kNoLabel) {
-      LOG(WARNING) << "Ignored: boundary word: " << fields[0];
-      continue;
-    }
-    num_added += builder->AddWord(word, feature_labels);
-  }
-  VLOG(1) << "Read " << num_added << " words in " << num_line << " lines from "
-          << vocab;
-}
-
-// Reads in model file. The first line is an integer designating the
-// size of future window in the input sequences. After this, each line
-// is in the following format
-//
-//   input sequence <whitespace> output sequence <whitespace> weight
-//
-// input sequence is a `FLAGS_delimiter`-delimited sequence of feature
-// labels (see `AddVocab()`) . output sequence is a
-// `FLAGS_delimiter`-delimited sequence of output labels where the
-// last label is the output of the feature position before the history
-// boundary.
-template <class Arc>
-void AddModel(const string &model, SymbolTable *fsyms, SymbolTable *osyms,
-              LinearFstDataBuilder<Arc> *builder) {
-  std::ifstream in(model);
-  if (!in) LOG(FATAL) << "Can't open file: " << model;
-  string line;
-  std::getline(in, line);
-  if (!in) LOG(FATAL) << "Empty file: " << model;
-  size_t future_size;
-  {
-    std::istringstream strm(line);
-    strm >> future_size;
-    if (!strm) LOG(FATAL) << "Can't read future size: " << model;
-  }
-  size_t num_line = 1, num_added = 0;
-  const int group = builder->AddGroup(future_size);
-  VLOG(1) << "Group " << group << ": from " << model << "; future size is "
-          << future_size << ".";
-  // Add the rest of lines as a single feature group
-  std::vector<string> fields;
-  std::vector<typename Arc::Label> input_labels, output_labels;
-  typename Arc::Weight weight;
-  while (GetModelRecord<Arc>(model, in, fsyms, osyms, &input_labels,
-                             &output_labels, &weight, &num_line)) {
-    if (output_labels.empty())
-      LOG(FATAL) << "Empty output sequence in source " << model << ", line "
-                 << num_line;
-
-    const typename Arc::Label marks[] = {LinearFstData<Arc>::kStartOfSentence,
-                                         LinearFstData<Arc>::kEndOfSentence};
-
-    std::vector<typename Arc::Label> copy_input(input_labels.size()),
-        copy_output(output_labels.size());
-    for (int i = 0; i < 2; ++i) {
-      for (int j = 0; j < 2; ++j) {
-        size_t num_input_changes =
-            ReplaceCopy(input_labels.begin(), input_labels.end(),
-                        copy_input.begin(), kNoLabel, marks[i]);
-        size_t num_output_changes =
-            ReplaceCopy(output_labels.begin(), output_labels.end(),
-                        copy_output.begin(), kNoLabel, marks[j]);
-        if ((num_input_changes > 0 || i == 0) &&
-            (num_output_changes > 0 || j == 0))
-          num_added +=
-              builder->AddWeight(group, copy_input, copy_output, weight);
-      }
-    }
-  }
-  VLOG(1) << "Group " << group << ": read " << num_added << " weight(s) in "
-          << num_line << " lines.";
-}
-
-template <class Arc>
-void AddModel(const string &model, SymbolTable *fsyms, SymbolTable *osyms,
-              LinearClassifierFstDataBuilder<Arc> *builder) {
-  std::ifstream in(model);
-  if (!in) LOG(FATAL) << "Can't open file: " << model;
-  string line;
-  std::getline(in, line);
-  if (!in) LOG(FATAL) << "Empty file: " << model;
-  size_t future_size;
-  {
-    std::istringstream strm(line);
-    strm >> future_size;
-    if (!strm) LOG(FATAL) << "Can't read future size: " << model;
-  }
-  if (future_size != 0)
-    LOG(FATAL) << "Classifier model must have future size = 0; got "
-               << future_size << " from " << model;
-  size_t num_line = 1, num_added = 0;
-  const int group = builder->AddGroup();
-  VLOG(1) << "Group " << group << ": from " << model << "; future size is "
-          << future_size << ".";
-  // Add the rest of lines as a single feature group
-  std::vector<string> fields;
-  std::vector<typename Arc::Label> input_labels, output_labels;
-  typename Arc::Weight weight;
-  while (GetModelRecord<Arc>(model, in, fsyms, osyms, &input_labels,
-                             &output_labels, &weight, &num_line)) {
-    if (output_labels.size() != 1)
-      LOG(FATAL) << "Output not a single label in source " << model << ", line "
-                 << num_line;
-
-    const typename Arc::Label marks[] = {LinearFstData<Arc>::kStartOfSentence,
-                                         LinearFstData<Arc>::kEndOfSentence};
-
-    typename Arc::Label pred = output_labels[0];
-
-    std::vector<typename Arc::Label> copy_input(input_labels.size());
-    for (int i = 0; i < 2; ++i) {
-      size_t num_input_changes =
-          ReplaceCopy(input_labels.begin(), input_labels.end(),
-                      copy_input.begin(), kNoLabel, marks[i]);
-      if (num_input_changes > 0 || i == 0)
-        num_added += builder->AddWeight(group, copy_input, pred, weight);
-    }
-  }
-  VLOG(1) << "Group " << group << ": read " << num_added << " weight(s) in "
-          << num_line << " lines.";
-}
-
-void SplitByWhitespace(const string &str, std::vector<string> *out);
-int ScanNumClasses(char **models, int models_length);
-
-template <class Arc>
-void LinearCompileTpl(LinearCompileArgs *args) {
-  const string &epsilon_symbol = std::get<0>(*args);
-  const string &unknown_symbol = std::get<1>(*args);
-  const string &vocab = std::get<2>(*args);
-  char **models = std::get<3>(*args);
-  const int models_length = std::get<4>(*args);
-  const string &out = std::get<5>(*args);
-  const string &save_isymbols = std::get<6>(*args);
-  const string &save_fsymbols = std::get<7>(*args);
-  const string &save_osymbols = std::get<8>(*args);
-
-  SymbolTable isyms,  // input (e.g. word tokens)
-      osyms,          // output (e.g. tags)
-      fsyms;          // feature (e.g. word identity, suffix, etc.)
-  isyms.AddSymbol(epsilon_symbol);
-  osyms.AddSymbol(epsilon_symbol);
-  fsyms.AddSymbol(epsilon_symbol);
-  isyms.AddSymbol(unknown_symbol);
-
-  VLOG(1) << "start-of-sentence label is "
-          << LinearFstData<Arc>::kStartOfSentence;
-  VLOG(1) << "end-of-sentence label is " << LinearFstData<Arc>::kEndOfSentence;
-
-  if (FLAGS_classifier) {
-    int num_classes = ScanNumClasses(models, models_length);
-    LinearClassifierFstDataBuilder<Arc> builder(num_classes, &isyms, &fsyms,
-                                                &osyms);
-
-    AddVocab(vocab, &isyms, &fsyms, &osyms, &builder);
-    for (int i = 0; i < models_length; ++i)
-      AddModel(models[i], &fsyms, &osyms, &builder);
-
-    LinearClassifierFst<Arc> fst(builder.Dump(), num_classes, &isyms, &osyms);
-    fst.Write(out);
-  } else {
-    LinearFstDataBuilder<Arc> builder(&isyms, &fsyms, &osyms);
-
-    AddVocab(vocab, &isyms, &fsyms, &osyms, &builder);
-    for (int i = 0; i < models_length; ++i)
-      AddModel(models[i], &fsyms, &osyms, &builder);
-
-    LinearTaggerFst<Arc> fst(builder.Dump(), &isyms, &osyms);
-    fst.Write(out);
-  }
-
-  if (!save_isymbols.empty()) isyms.WriteText(save_isymbols);
-  if (!save_fsymbols.empty()) fsyms.WriteText(save_fsymbols);
-  if (!save_osymbols.empty()) osyms.WriteText(save_osymbols);
-}
-
-void LinearCompile(const string &arc_type, const string &epsilon_symbol,
-                   const string &unknown_symbol, const string &vocab,
-                   char **models, int models_len, const string &out,
-                   const string &save_isymbols, const string &save_fsymbols,
-                   const string &save_osymbols);
-
-template <class Arc>
-bool GetVocabRecord(const string &vocab, std::istream &strm,  // NOLINT
-                    SymbolTable *isyms, SymbolTable *fsyms, SymbolTable *osyms,
-                    typename Arc::Label *word,
-                    std::vector<typename Arc::Label> *feature_labels,
-                    std::vector<typename Arc::Label> *possible_labels,
-                    size_t *num_line) {
-  string line;
-  if (!std::getline(strm, line)) return false;
-  ++(*num_line);
-
-  std::vector<string> fields;
-  SplitByWhitespace(line, &fields);
-  if (fields.size() != 3)
-    LOG(FATAL) << "Wrong number of fields in source " << vocab << ", line "
-               << num_line;
-
-  feature_labels->clear();
-  possible_labels->clear();
-
-  *word = LookUp<Arc>(fields[0], isyms);
-
-  const char delim = FLAGS_delimiter[0];
-  SplitAndPush<Arc>(fields[1], delim, fsyms, feature_labels);
-  SplitAndPush<Arc>(fields[2], delim, osyms, possible_labels);
-
-  return true;
-}
-
-template <class Arc>
-bool GetModelRecord(const string &model, std::istream &strm,  // NOLINT
-                    SymbolTable *fsyms, SymbolTable *osyms,
-                    std::vector<typename Arc::Label> *input_labels,
-                    std::vector<typename Arc::Label> *output_labels,
-                    typename Arc::Weight *weight, size_t *num_line) {
-  string line;
-  if (!std::getline(strm, line)) return false;
-  ++(*num_line);
-
-  std::vector<string> fields;
-  SplitByWhitespace(line, &fields);
-  if (fields.size() != 3)
-    LOG(FATAL) << "Wrong number of fields in source " << model << ", line "
-               << num_line;
-
-  input_labels->clear();
-  output_labels->clear();
-
-  const char delim = FLAGS_delimiter[0];
-  SplitAndPush<Arc>(fields[0], delim, fsyms, input_labels);
-  SplitAndPush<Arc>(fields[1], delim, osyms, output_labels);
-
-  *weight = StrToWeight<typename Arc::Weight>(fields[2], model, *num_line);
-
-  GuessStartOrEnd<Arc>(input_labels, kNoLabel);
-  GuessStartOrEnd<Arc>(output_labels, kNoLabel);
-
-  return true;
-}
-}  // namespace script
-}  // namespace fst
-
-#define REGISTER_FST_LINEAR_OPERATIONS(Arc) \
-  REGISTER_FST_OPERATION(LinearCompileTpl, Arc, LinearCompileArgs);
-
-#endif  // FST_EXTENSIONS_LINEAR_LINEARSCRIPT_H_
diff --git a/projects/llm_framework/include/fst/extensions/linear/loglinear-apply.h b/projects/llm_framework/include/fst/extensions/linear/loglinear-apply.h
deleted file mode 100644
index 1b5d2eaf..00000000
--- a/projects/llm_framework/include/fst/extensions/linear/loglinear-apply.h
+++ /dev/null
@@ -1,77 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_LINEAR_LOGLINEAR_APPLY_H_
-#define FST_EXTENSIONS_LINEAR_LOGLINEAR_APPLY_H_
-
-#include <fst/compat.h>
-#include <fst/arc.h>
-#include <fst/arc-map.h>
-#include <fst/compose.h>
-#include <fst/determinize.h>
-#include <fst/float-weight.h>
-#include <fst/fst.h>
-#include <fst/minimize.h>
-#include <fst/mutable-fst.h>
-#include <fst/project.h>
-#include <fst/rmepsilon.h>
-#include <fst/vector-fst.h>
-
-namespace fst {
-
-// Applies a FST model as a discriminative model to weighted input
-// `ifst`. `A` is an arc type with tropical weight of all the
-// input/output FSTs.
-//
-// In general, consider `ifst` an unnormalized probability
-// distribution between its input X and output Y, P(X, Y); and `lfst`
-// a group of unnormalized probability distributions of all its output
-// Z for every input Y, Q(Z|Y). `normalize` controls whether Q is
-// normalized for every Y before chaining with P(X, Y). I.e., for a
-// path (X, Y, Z) in `ofst` (where Y is hidden),
-//
-// - When `normalize` is true, its weight is P(X, Y) Q(Z|Y) / sum_z Q(z|Y);
-// - When `normalize` is false, its weight is P(X, Y) Q(Z|Y).
-template <class A>
-void LogLinearApply(const Fst<A> &ifst, const Fst<A> &lfst, MutableFst<A> *ofst,
-                    bool normalize = true) {
-  LogLinearApply<A, LogArc>(ifst, lfst, ofst, normalize);
-}
-
-// This version gives finer control over the arc type (`B`) to be used
-// in normalization. `B` is an arc type with log weight (e.g. `LogArc`
-// or `Log64Arc`).
-template <class A, class B>
-void LogLinearApply(const Fst<A> &ifst, const Fst<A> &lfst, MutableFst<A> *ofst,
-                    bool normalize = true) {
-  if (normalize) {
-    VectorFst<A> unnormalized_ofst, rescored_ifsa;
-    Compose(ifst, lfst, &unnormalized_ofst);
-    {
-      VectorFst<A> tropical_ifsa(unnormalized_ofst);
-      Project(&tropical_ifsa, PROJECT_INPUT);
-      {
-        VectorFst<B> minimal_log_ifsa;
-        {
-          VectorFst<B> log_ifsa;
-          ArcMap(tropical_ifsa, &log_ifsa, WeightConvertMapper<A, B>());
-          RmEpsilon(&log_ifsa);
-          Determinize(log_ifsa, &minimal_log_ifsa);
-        }
-        Minimize(&minimal_log_ifsa);
-        ArcMap(&minimal_log_ifsa, InvertWeightMapper<B>());
-        ArcMap(minimal_log_ifsa, &tropical_ifsa, WeightConvertMapper<B, A>());
-      }
-      ArcSort(&tropical_ifsa, OLabelCompare<A>());
-      Compose(tropical_ifsa, ifst, &rescored_ifsa);
-    }
-    ArcSort(&rescored_ifsa, OLabelCompare<A>());
-    Compose(rescored_ifsa, unnormalized_ofst, ofst);
-  } else {
-    Compose(ifst, lfst, ofst);
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_LINEAR_LOGLINEAR_APPLY_H_
diff --git a/projects/llm_framework/include/fst/extensions/linear/trie.h b/projects/llm_framework/include/fst/extensions/linear/trie.h
deleted file mode 100644
index b5ddb3ad..00000000
--- a/projects/llm_framework/include/fst/extensions/linear/trie.h
+++ /dev/null
@@ -1,444 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_LINEAR_TRIE_H_
-#define FST_EXTENSIONS_LINEAR_TRIE_H_
-
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/util.h>
-
-namespace fst {
-
-const int kNoTrieNodeId = -1;
-
-// Forward declarations of all available trie topologies.
-template <class L, class H>
-class NestedTrieTopology;
-template <class L, class H>
-class FlatTrieTopology;
-
-// A pair of parent node id and label, part of a trie edge
-template <class L>
-struct ParentLabel {
-  int parent;
-  L label;
-
-  ParentLabel() {}
-  ParentLabel(int p, L l) : parent(p), label(l) {}
-
-  bool operator==(const ParentLabel &that) const {
-    return parent == that.parent && label == that.label;
-  }
-
-  std::istream &Read(std::istream &strm) {  // NOLINT
-    ReadType(strm, &parent);
-    ReadType(strm, &label);
-    return strm;
-  }
-
-  std::ostream &Write(std::ostream &strm) const {  // NOLINT
-    WriteType(strm, parent);
-    WriteType(strm, label);
-    return strm;
-  }
-};
-
-template <class L, class H>
-struct ParentLabelHash {
-  size_t operator()(const ParentLabel<L> &pl) const {
-    return static_cast<size_t>(pl.parent * 7853 + H()(pl.label));
-  }
-};
-
-// The trie topology in a nested tree of hash maps; allows efficient
-// iteration over children of a specific node.
-template <class L, class H>
-class NestedTrieTopology {
- public:
-  typedef L Label;
-  typedef H Hash;
-  typedef std::unordered_map<L, int, H> NextMap;
-
-  class const_iterator {
-   public:
-    typedef std::forward_iterator_tag iterator_category;
-    typedef std::pair<ParentLabel<L>, int> value_type;
-    typedef std::ptrdiff_t difference_type;
-    typedef const value_type *pointer;
-    typedef const value_type &reference;
-
-    friend class NestedTrieTopology<L, H>;
-
-    const_iterator() : ptr_(nullptr), cur_node_(kNoTrieNodeId), cur_edge_() {}
-
-    reference operator*() {
-      UpdateStub();
-      return stub_;
-    }
-    pointer operator->() {
-      UpdateStub();
-      return &stub_;
-    }
-
-    const_iterator &operator++();
-    const_iterator &operator++(int);  // NOLINT
-
-    bool operator==(const const_iterator &that) const {
-      return ptr_ == that.ptr_ && cur_node_ == that.cur_node_ &&
-             cur_edge_ == that.cur_edge_;
-    }
-    bool operator!=(const const_iterator &that) const {
-      return !(*this == that);
-    }
-
-   private:
-    const_iterator(const NestedTrieTopology *ptr, int cur_node)
-        : ptr_(ptr), cur_node_(cur_node) {
-      SetProperCurEdge();
-    }
-
-    void SetProperCurEdge() {
-      if (cur_node_ < ptr_->NumNodes())
-        cur_edge_ = ptr_->nodes_[cur_node_]->begin();
-      else
-        cur_edge_ = ptr_->nodes_[0]->begin();
-    }
-
-    void UpdateStub() {
-      stub_.first = ParentLabel<L>(cur_node_, cur_edge_->first);
-      stub_.second = cur_edge_->second;
-    }
-
-    const NestedTrieTopology *ptr_;
-    int cur_node_;
-    typename NextMap::const_iterator cur_edge_;
-    value_type stub_;
-  };
-
-  NestedTrieTopology();
-  NestedTrieTopology(const NestedTrieTopology &that);
-  ~NestedTrieTopology();
-  void swap(NestedTrieTopology &that);
-  NestedTrieTopology &operator=(const NestedTrieTopology &that);
-  bool operator==(const NestedTrieTopology &that) const;
-  bool operator!=(const NestedTrieTopology &that) const;
-
-  int Root() const { return 0; }
-  size_t NumNodes() const { return nodes_.size(); }
-  int Insert(int parent, const L &label);
-  int Find(int parent, const L &label) const;
-  const NextMap &ChildrenOf(int parent) const { return *nodes_[parent]; }
-
-  std::istream &Read(std::istream &strm);         // NOLINT
-  std::ostream &Write(std::ostream &strm) const;  // NOLINT
-
-  const_iterator begin() const { return const_iterator(this, 0); }
-  const_iterator end() const { return const_iterator(this, NumNodes()); }
-
- private:
-  std::vector<NextMap *>
-      nodes_;  // Use pointers to avoid copying the maps when the
-               // vector grows
-};
-
-template <class L, class H>
-NestedTrieTopology<L, H>::NestedTrieTopology() {
-  nodes_.push_back(new NextMap);
-}
-
-template <class L, class H>
-NestedTrieTopology<L, H>::NestedTrieTopology(const NestedTrieTopology &that) {
-  nodes_.reserve(that.nodes_.size());
-  for (size_t i = 0; i < that.nodes_.size(); ++i) {
-    NextMap *node = that.nodes_[i];
-    nodes_.push_back(new NextMap(*node));
-  }
-}
-
-template <class L, class H>
-NestedTrieTopology<L, H>::~NestedTrieTopology() {
-  for (size_t i = 0; i < nodes_.size(); ++i) {
-    NextMap *node = nodes_[i];
-    delete node;
-  }
-}
-
-// TODO(wuke): std::swap compatibility
-template <class L, class H>
-inline void NestedTrieTopology<L, H>::swap(NestedTrieTopology &that) {
-  nodes_.swap(that.nodes_);
-}
-
-template <class L, class H>
-inline NestedTrieTopology<L, H> &NestedTrieTopology<L, H>::operator=(
-    const NestedTrieTopology &that) {
-  NestedTrieTopology copy(that);
-  swap(copy);
-  return *this;
-}
-
-template <class L, class H>
-inline bool NestedTrieTopology<L, H>::operator==(
-    const NestedTrieTopology &that) const {
-  if (NumNodes() != that.NumNodes()) return false;
-  for (int i = 0; i < NumNodes(); ++i)
-    if (ChildrenOf(i) != that.ChildrenOf(i)) return false;
-  return true;
-}
-
-template <class L, class H>
-inline bool NestedTrieTopology<L, H>::operator!=(
-    const NestedTrieTopology &that) const {
-  return !(*this == that);
-}
-
-template <class L, class H>
-inline int NestedTrieTopology<L, H>::Insert(int parent, const L &label) {
-  int ret = Find(parent, label);
-  if (ret == kNoTrieNodeId) {
-    ret = NumNodes();
-    (*nodes_[parent])[label] = ret;
-    nodes_.push_back(new NextMap);
-  }
-  return ret;
-}
-
-template <class L, class H>
-inline int NestedTrieTopology<L, H>::Find(int parent, const L &label) const {
-  typename NextMap::const_iterator it = nodes_[parent]->find(label);
-  return it == nodes_[parent]->end() ? kNoTrieNodeId : it->second;
-}
-
-template <class L, class H>
-inline std::istream &NestedTrieTopology<L, H>::Read(
-    std::istream &strm) {  // NOLINT
-  NestedTrieTopology new_trie;
-  size_t num_nodes;
-  if (!ReadType(strm, &num_nodes)) return strm;
-  for (size_t i = 1; i < num_nodes; ++i) new_trie.nodes_.push_back(new NextMap);
-  for (size_t i = 0; i < num_nodes; ++i) ReadType(strm, new_trie.nodes_[i]);
-  if (strm) swap(new_trie);
-  return strm;
-}
-
-template <class L, class H>
-inline std::ostream &NestedTrieTopology<L, H>::Write(
-    std::ostream &strm) const {  // NOLINT
-  WriteType(strm, NumNodes());
-  for (size_t i = 0; i < NumNodes(); ++i) WriteType(strm, *nodes_[i]);
-  return strm;
-}
-
-template <class L, class H>
-inline typename NestedTrieTopology<L, H>::const_iterator
-    &NestedTrieTopology<L, H>::const_iterator::operator++() {
-  ++cur_edge_;
-  if (cur_edge_ == ptr_->nodes_[cur_node_]->end()) {
-    ++cur_node_;
-    while (cur_node_ < ptr_->NumNodes() && ptr_->nodes_[cur_node_]->empty())
-      ++cur_node_;
-    SetProperCurEdge();
-  }
-  return *this;
-}
-
-template <class L, class H>
-inline typename NestedTrieTopology<L, H>::const_iterator
-    &NestedTrieTopology<L, H>::const_iterator::operator++(int) {  // NOLINT
-  const_iterator save(*this);
-  ++(*this);
-  return save;
-}
-
-// The trie topology in a single hash map; only allows iteration over
-// all the edges in arbitrary order.
-template <class L, class H>
-class FlatTrieTopology {
- private:
-  typedef std::unordered_map<ParentLabel<L>, int, ParentLabelHash<L, H>>
-      NextMap;
-
- public:
-  // Iterator over edges as std::pair<ParentLabel<L>, int>
-  typedef typename NextMap::const_iterator const_iterator;
-  typedef L Label;
-  typedef H Hash;
-
-  FlatTrieTopology() {}
-  FlatTrieTopology(const FlatTrieTopology &that) : next_(that.next_) {}
-  template <class T>
-  explicit FlatTrieTopology(const T &that);
-
-  // TODO(wuke): std::swap compatibility
-  void swap(FlatTrieTopology &that) { next_.swap(that.next_); }
-
-  bool operator==(const FlatTrieTopology &that) const {
-    return next_ == that.next_;
-  }
-  bool operator!=(const FlatTrieTopology &that) const {
-    return !(*this == that);
-  }
-
-  int Root() const { return 0; }
-  size_t NumNodes() const { return next_.size() + 1; }
-  int Insert(int parent, const L &label);
-  int Find(int parent, const L &label) const;
-
-  std::istream &Read(std::istream &strm) {  // NOLINT
-    return ReadType(strm, &next_);
-  }
-  std::ostream &Write(std::ostream &strm) const {  // NOLINT
-    return WriteType(strm, next_);
-  }
-
-  const_iterator begin() const { return next_.begin(); }
-  const_iterator end() const { return next_.end(); }
-
- private:
-  NextMap next_;
-};
-
-template <class L, class H>
-template <class T>
-FlatTrieTopology<L, H>::FlatTrieTopology(const T &that)
-    : next_(that.begin(), that.end()) {}
-
-template <class L, class H>
-inline int FlatTrieTopology<L, H>::Insert(int parent, const L &label) {
-  int ret = Find(parent, label);
-  if (ret == kNoTrieNodeId) {
-    ret = NumNodes();
-    next_[ParentLabel<L>(parent, label)] = ret;
-  }
-  return ret;
-}
-
-template <class L, class H>
-inline int FlatTrieTopology<L, H>::Find(int parent, const L &label) const {
-  typename NextMap::const_iterator it =
-      next_.find(ParentLabel<L>(parent, label));
-  return it == next_.end() ? kNoTrieNodeId : it->second;
-}
-
-// A collection of implementations of the trie data structure. The key
-// is a sequence of type `L` which must be hashable. The value is of
-// `V` which must be default constructible and copyable. In addition,
-// a value object is stored for each node in the trie therefore
-// copying `V` should be cheap.
-//
-// One can access the store values with an integer node id, using the
-// [] operator. A valid node id can be obtained by the following ways:
-//
-// 1. Using the `Root()` method to get the node id of the root.
-//
-// 2. Iterating through 0 to `NumNodes() - 1`. The node ids are dense
-// so every integer in this range is a valid node id.
-//
-// 3. Using the node id returned from a successful `Insert()` or
-// `Find()` call.
-//
-// 4. Iterating over the trie edges with an `EdgeIterator` and using
-// the node ids returned from its `Parent()` and `Child()` methods.
-//
-// Below is an example of inserting keys into the trie:
-//
-//   const string words[] = {"hello", "health", "jello"};
-//   Trie<char, bool> dict;
-//   for (auto word : words) {
-//     int cur = dict.Root();
-//     for (char c : word) {
-//       cur = dict.Insert(cur, c);
-//     }
-//     dict[cur] = true;
-//   }
-//
-// And the following is an example of looking up the longest prefix of
-// a string using the trie constructed above:
-//
-//   string query = "healed";
-//   size_t prefix_length = 0;
-//   int cur = dict.Find(dict.Root(), query[prefix_length]);
-//   while (prefix_length < query.size() &&
-//     cur != Trie<char, bool>::kNoNodeId) {
-//     ++prefix_length;
-//     cur = dict.Find(cur, query[prefix_length]);
-//   }
-template <class L, class V, class T>
-class MutableTrie {
- public:
-  template <class LL, class VV, class TT>
-  friend class MutableTrie;
-
-  typedef L Label;
-  typedef V Value;
-  typedef T Topology;
-
-  // Constructs a trie with only the root node.
-  MutableTrie() {}
-
-  // Conversion from another trie of a possiblly different
-  // topology. The underlying topology must supported conversion.
-  template <class S>
-  explicit MutableTrie(const MutableTrie<L, V, S> &that)
-      : topology_(that.topology_), values_(that.values_) {}
-
-  // TODO(wuke): std::swap compatibility
-  void swap(MutableTrie &that) {
-    topology_.swap(that.topology_);
-    values_.swap(that.values_);
-  }
-
-  int Root() const { return topology_.Root(); }
-  size_t NumNodes() const { return topology_.NumNodes(); }
-
-  // Inserts an edge with given `label` at node `parent`. Returns the
-  // child node id. If the node already exists, returns the node id
-  // right away.
-  int Insert(int parent, const L &label) {
-    int ret = topology_.Insert(parent, label);
-    values_.resize(NumNodes());
-    return ret;
-  }
-
-  // Finds the node id of the node from `parent` via `label`. Returns
-  // `kNoTrieNodeId` when such a node does not exist.
-  int Find(int parent, const L &label) const {
-    return topology_.Find(parent, label);
-  }
-
-  const T &TrieTopology() const { return topology_; }
-
-  // Accesses the value stored for the given node.
-  V &operator[](int node_id) { return values_[node_id]; }
-  const V &operator[](int node_id) const { return values_[node_id]; }
-
-  // Comparison by content
-  bool operator==(const MutableTrie &that) const {
-    return topology_ == that.topology_ && values_ == that.values_;
-  }
-
-  bool operator!=(const MutableTrie &that) const { return !(*this == that); }
-
-  std::istream &Read(std::istream &strm) {  // NOLINT
-    ReadType(strm, &topology_);
-    ReadType(strm, &values_);
-    return strm;
-  }
-  std::ostream &Write(std::ostream &strm) const {  // NOLINT
-    WriteType(strm, topology_);
-    WriteType(strm, values_);
-    return strm;
-  }
-
- private:
-  T topology_;
-  std::vector<V> values_;
-};
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_LINEAR_TRIE_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/compose.h b/projects/llm_framework/include/fst/extensions/mpdt/compose.h
deleted file mode 100644
index 47714e37..00000000
--- a/projects/llm_framework/include/fst/extensions/mpdt/compose.h
+++ /dev/null
@@ -1,267 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Compose an MPDT and an FST.
-
-#ifndef FST_EXTENSIONS_MPDT_COMPOSE_H_
-#define FST_EXTENSIONS_MPDT_COMPOSE_H_
-
-#include <list>
-
-#include <fst/extensions/mpdt/mpdt.h>
-#include <fst/extensions/pdt/compose.h>
-#include <fst/compose.h>
-
-namespace fst {
-
-template <class Filter>
-class MPdtParenFilter {
- public:
-  using FST1 = typename Filter::FST1;
-  using FST2 = typename Filter::FST2;
-  using Arc = typename Filter::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Matcher1 = typename Filter::Matcher1;
-  using Matcher2 = typename Filter::Matcher2;
-
-  using StackId = StateId;
-  using ParenStack = internal::MPdtStack<StackId, Label>;
-  using FilterState1 = typename Filter::FilterState;
-  using FilterState2 = IntegerFilterState<StackId>;
-  using FilterState = PairFilterState<FilterState1, FilterState2>;
-
-  MPdtParenFilter(const FST1 &fst1, const FST2 &fst2,
-                  Matcher1 *matcher1 = nullptr, Matcher2 *matcher2 = nullptr,
-                  const std::vector<std::pair<Label, Label>> *parens = nullptr,
-                  const std::vector<Label> *assignments = nullptr,
-                  bool expand = false, bool keep_parens = true)
-      : filter_(fst1, fst2, matcher1, matcher2),
-        parens_(parens ? *parens : std::vector<std::pair<Label, Label>>()),
-        assignments_(assignments ? *assignments : std::vector<Label>()),
-        expand_(expand),
-        keep_parens_(keep_parens),
-        fs_(FilterState::NoState()),
-        stack_(parens_, assignments_),
-        paren_id_(-1) {
-    if (parens) {
-      for (const auto &pair : *parens) {
-        parens_.push_back(pair);
-        GetMatcher1()->AddOpenParen(pair.first);
-        GetMatcher2()->AddOpenParen(pair.first);
-        if (!expand_) {
-          GetMatcher1()->AddCloseParen(pair.second);
-          GetMatcher2()->AddCloseParen(pair.second);
-        }
-      }
-    }
-  }
-
-  MPdtParenFilter(const MPdtParenFilter &filter, bool safe = false)
-      : filter_(filter.filter_, safe),
-        parens_(filter.parens_),
-        expand_(filter.expand_),
-        keep_parens_(filter.keep_parens_),
-        fs_(FilterState::NoState()),
-        stack_(filter.parens_, filter.assignments_),
-        paren_id_(-1) {}
-
-  FilterState Start() const {
-    return FilterState(filter_.Start(), FilterState2(0));
-  }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    fs_ = fs;
-    filter_.SetState(s1, s2, fs_.GetState1());
-    if (!expand_) return;
-    const auto paren_id = stack_.Top(fs.GetState2().GetState());
-    if (paren_id != paren_id_) {
-      if (paren_id_ != -1) {
-        GetMatcher1()->RemoveCloseParen(parens_[paren_id_].second);
-        GetMatcher2()->RemoveCloseParen(parens_[paren_id_].second);
-      }
-      paren_id_ = paren_id;
-      if (paren_id_ != -1) {
-        GetMatcher1()->AddCloseParen(parens_[paren_id_].second);
-        GetMatcher2()->AddCloseParen(parens_[paren_id_].second);
-      }
-    }
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    const auto fs1 = filter_.FilterArc(arc1, arc2);
-    const auto &fs2 = fs_.GetState2();
-    if (fs1 == FilterState1::NoState()) return FilterState::NoState();
-    if (arc1->olabel == kNoLabel && arc2->ilabel) {  // arc2 parentheses.
-      if (keep_parens_) {
-        arc1->ilabel = arc2->ilabel;
-      } else if (arc2->ilabel) {
-        arc2->olabel = arc1->ilabel;
-      }
-      return FilterParen(arc2->ilabel, fs1, fs2);
-    } else if (arc2->ilabel == kNoLabel && arc1->olabel) {  // arc1 parentheses
-      if (keep_parens_) {
-        arc2->olabel = arc1->olabel;
-      } else {
-        arc1->ilabel = arc2->olabel;
-      }
-      return FilterParen(arc1->olabel, fs1, fs2);
-    } else {
-      return FilterState(fs1, fs2);
-    }
-  }
-
-  void FilterFinal(Weight *w1, Weight *w2) const {
-    if (fs_.GetState2().GetState() != 0) *w1 = Weight::Zero();
-    filter_.FilterFinal(w1, w2);
-  }
-
-  // Returns respective matchers; ownership stays with filter.
-
-  Matcher1 *GetMatcher1() { return filter_.GetMatcher1(); }
-
-  Matcher2 *GetMatcher2() { return filter_.GetMatcher2(); }
-
-  uint64 Properties(uint64 iprops) const {
-    const auto oprops = filter_.Properties(iprops);
-    return oprops & kILabelInvariantProperties & kOLabelInvariantProperties;
-  }
-
- private:
-  const FilterState FilterParen(Label label, const FilterState1 &fs1,
-                                const FilterState2 &fs2) const {
-    if (!expand_) return FilterState(fs1, fs2);
-    const auto stack_id = stack_.Find(fs2.GetState(), label);
-    if (stack_id < 0) {
-      return FilterState::NoState();
-    } else {
-      return FilterState(fs1, FilterState2(stack_id));
-    }
-  }
-
-  Filter filter_;
-  std::vector<std::pair<Label, Label>> parens_;
-  std::vector<Label> assignments_;
-  bool expand_;       // Expands to FST?
-  bool keep_parens_;  // Retains parentheses in output?
-  FilterState fs_;    // Current filter state.
-  mutable ParenStack stack_;
-  ssize_t paren_id_;
-};
-
-// Class to setup composition options for MPDT composition. Default is to take
-// the MPDT as the first composition argument.
-template <class Arc, bool left_pdt = true>
-class MPdtComposeFstOptions
-    : public ComposeFstOptions<Arc, ParenMatcher<Fst<Arc>>,
-                               MPdtParenFilter<AltSequenceComposeFilter<
-                                   ParenMatcher<Fst<Arc>> >> > {
- public:
-  using Label = typename Arc::Label;
-  using MPdtMatcher = ParenMatcher<Fst<Arc>>;
-  using MPdtFilter = MPdtParenFilter<AltSequenceComposeFilter<MPdtMatcher>>;
-
-  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::matcher1;
-  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::matcher2;
-  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::filter;
-
-  MPdtComposeFstOptions(const Fst<Arc> &ifst1,
-                        const std::vector<std::pair<Label, Label>> &parens,
-                        const std::vector<typename Arc::Label> &assignments,
-                        const Fst<Arc> &ifst2, bool expand = false,
-                        bool keep_parens = true) {
-    matcher1 = new MPdtMatcher(ifst1, MATCH_OUTPUT, kParenList);
-    matcher2 = new MPdtMatcher(ifst2, MATCH_INPUT, kParenLoop);
-    filter = new MPdtFilter(ifst1, ifst2, matcher1, matcher2, &parens,
-                            &assignments, expand, keep_parens);
-  }
-};
-
-// Class to setup composition options for PDT with FST composition.
-// Specialization is for the FST as the first composition argument.
-template <class Arc>
-class MPdtComposeFstOptions<Arc, false>
-    : public ComposeFstOptions<
-          Arc, ParenMatcher<Fst<Arc>>,
-          MPdtParenFilter<SequenceComposeFilter<ParenMatcher<Fst<Arc>> >> > {
- public:
-  using Label = typename Arc::Label;
-  using MPdtMatcher = ParenMatcher<Fst<Arc>>;
-  using MPdtFilter = MPdtParenFilter<SequenceComposeFilter<MPdtMatcher>>;
-
-  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::matcher1;
-  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::matcher2;
-  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::filter;
-
-  MPdtComposeFstOptions(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-                        const std::vector<std::pair<Label, Label>> &parens,
-                        const std::vector<typename Arc::Label> &assignments,
-                        bool expand = false, bool keep_parens = true) {
-    matcher1 = new MPdtMatcher(ifst1, MATCH_OUTPUT, kParenLoop);
-    matcher2 = new MPdtMatcher(ifst2, MATCH_INPUT, kParenList);
-    filter = new MPdtFilter(ifst1, ifst2, matcher1, matcher2, &parens,
-                            &assignments, expand, keep_parens);
-  }
-};
-
-struct MPdtComposeOptions {
-  bool connect;                  // Connect output?
-  PdtComposeFilter filter_type;  // Which pre-defined filter to use.
-
-  explicit MPdtComposeOptions(bool connect = true,
-                              PdtComposeFilter filter_type = PAREN_FILTER)
-      : connect(connect), filter_type(filter_type) {}
-};
-
-// Composes multi-pushdown transducer (MPDT) encoded as an FST (1st arg) and an
-// FST (2nd arg) with the result also an MPDT encoded as an FST (3rd arg). In
-// theMPDTs, some transitions are labeled with open or close parentheses (and
-// associated with a stack). To be interpreted as an MPDT, the parents on each
-// stack must balance on a path (see MPdtExpand()). The open-close parenthesis
-// label pairs are passed using the parens arguments, and the stack assignments
-// are passed using the assignments argument.
-template <class Arc>
-void Compose(
-    const Fst<Arc> &ifst1,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens,
-    const std::vector<typename Arc::Label> &assignments, const Fst<Arc> &ifst2,
-    MutableFst<Arc> *ofst,
-    const MPdtComposeOptions &opts = MPdtComposeOptions()) {
-  bool expand = opts.filter_type != PAREN_FILTER;
-  bool keep_parens = opts.filter_type != EXPAND_FILTER;
-  MPdtComposeFstOptions<Arc, true> copts(ifst1, parens, assignments, ifst2,
-                                         expand, keep_parens);
-  copts.gc_limit = 0;
-  *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-  if (opts.connect) Connect(ofst);
-}
-
-// Composes an FST (1st arg) and a multi-pushdown transducer (MPDT) encoded as
-// an FST (2nd arg) with the result also an MPDT encoded as an FST (3rd arg).
-// In the MPDTs, some transitions are labeled with open or close parentheses
-// (and associated with a stack). To be interpreted as an MPDT, the parents on
-// each stack must balance on a path (see MPdtExpand()). The open-close
-// parenthesis label pairs are passed using the parens arguments, and the stack
-// assignments are passed using the assignments argument.
-template <class Arc>
-void Compose(
-    const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens,
-    const std::vector<typename Arc::Label> &assignments, MutableFst<Arc> *ofst,
-    const MPdtComposeOptions &opts = MPdtComposeOptions()) {
-  bool expand = opts.filter_type != PAREN_FILTER;
-  bool keep_parens = opts.filter_type != EXPAND_FILTER;
-  MPdtComposeFstOptions<Arc, false> copts(ifst1, ifst2, parens, assignments,
-                                          expand, keep_parens);
-  copts.gc_limit = 0;
-  *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-  if (opts.connect) Connect(ofst);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_MPDT_COMPOSE_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/expand.h b/projects/llm_framework/include/fst/extensions/mpdt/expand.h
deleted file mode 100644
index 0e06cdce..00000000
--- a/projects/llm_framework/include/fst/extensions/mpdt/expand.h
+++ /dev/null
@@ -1,335 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Expands an MPDT to an FST.
-
-#ifndef FST_EXTENSIONS_MPDT_EXPAND_H_
-#define FST_EXTENSIONS_MPDT_EXPAND_H_
-
-#include <vector>
-
-#include <fst/extensions/mpdt/mpdt.h>
-#include <fst/extensions/pdt/paren.h>
-#include <fst/cache.h>
-#include <fst/mutable-fst.h>
-#include <fst/queue.h>
-#include <fst/state-table.h>
-#include <fst/test-properties.h>
-
-namespace fst {
-
-template <class Arc>
-struct MPdtExpandFstOptions : public CacheOptions {
-  bool keep_parentheses;
-  internal::MPdtStack<typename Arc::StateId, typename Arc::Label> *stack;
-  PdtStateTable<typename Arc::StateId, typename Arc::StateId> *state_table;
-
-  MPdtExpandFstOptions(
-      const CacheOptions &opts = CacheOptions(), bool kp = false,
-      internal::MPdtStack<typename Arc::StateId, typename Arc::Label> *s =
-          nullptr,
-      PdtStateTable<typename Arc::StateId, typename Arc::StateId> *st = nullptr)
-      : CacheOptions(opts), keep_parentheses(kp), stack(s), state_table(st) {}
-};
-
-// Properties for an expanded PDT.
-inline uint64 MPdtExpandProperties(uint64 inprops) {
-  return inprops & (kAcceptor | kAcyclic | kInitialAcyclic | kUnweighted);
-}
-
-namespace internal {
-
-// Implementation class for ExpandFst
-template <class Arc>
-class MPdtExpandFstImpl : public CacheImpl<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using StackId = StateId;
-  using StateTuple = PdtStateTuple<StateId, StackId>;
-  using ParenStack = internal::MPdtStack<StateId, Label>;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::Properties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheBaseImpl<CacheState<Arc>>::PushArc;
-  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
-  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
-  using CacheBaseImpl<CacheState<Arc>>::HasStart;
-  using CacheBaseImpl<CacheState<Arc>>::SetArcs;
-  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
-  using CacheBaseImpl<CacheState<Arc>>::SetStart;
-
-  MPdtExpandFstImpl(const Fst<Arc> &fst,
-                    const std::vector<std::pair<Label, Label>> &parens,
-                    const std::vector<Label> &assignments,
-                    const MPdtExpandFstOptions<Arc> &opts)
-      : CacheImpl<Arc>(opts),
-        fst_(fst.Copy()),
-        stack_(opts.stack ? opts.stack : new ParenStack(parens, assignments)),
-        state_table_(opts.state_table ? opts.state_table
-                                      : new PdtStateTable<StateId, StackId>()),
-        own_stack_(!opts.stack),
-        own_state_table_(!opts.state_table),
-        keep_parentheses_(opts.keep_parentheses) {
-    SetType("expand");
-    const auto props = fst.Properties(kFstProperties, false);
-    SetProperties(MPdtExpandProperties(props), kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  MPdtExpandFstImpl(const MPdtExpandFstImpl &impl)
-      : CacheImpl<Arc>(impl),
-        fst_(impl.fst_->Copy(true)),
-        stack_(new ParenStack(*impl.stack_)),
-        state_table_(new PdtStateTable<StateId, StackId>()),
-        own_stack_(true),
-        own_state_table_(true),
-        keep_parentheses_(impl.keep_parentheses_) {
-    SetType("expand");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  ~MPdtExpandFstImpl() override {
-    if (own_stack_) delete stack_;
-    if (own_state_table_) delete state_table_;
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      const auto s = fst_->Start();
-      if (s == kNoStateId) return kNoStateId;
-      const StateTuple tuple(s, 0);
-      const auto start = state_table_->FindState(tuple);
-      SetStart(start);
-    }
-    return CacheImpl<Arc>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) {
-      const auto &tuple = state_table_->Tuple(s);
-      const auto weight = fst_->Final(tuple.state_id);
-      SetFinal(s,
-               (weight != Weight::Zero() && tuple.stack_id == 0)
-                   ? weight
-                   : Weight::Zero());
-    }
-    return CacheImpl<Arc>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) ExpandState(s);
-    return CacheImpl<Arc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) ExpandState(s);
-    return CacheImpl<Arc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) ExpandState(s);
-    return CacheImpl<Arc>::NumOutputEpsilons(s);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) ExpandState(s);
-    CacheImpl<Arc>::InitArcIterator(s, data);
-  }
-
-  // Computes the outgoing transitions from a state, creating new destination
-  // states as needed.
-  void ExpandState(StateId s) {
-    const auto tuple = state_table_->Tuple(s);
-    for (ArcIterator<Fst<Arc>> aiter(*fst_, tuple.state_id); !aiter.Done();
-         aiter.Next()) {
-      auto arc = aiter.Value();
-      const auto stack_id = stack_->Find(tuple.stack_id, arc.ilabel);
-      if (stack_id == -1) {
-        continue;  // Non-matching close parenthesis.
-      } else if ((stack_id != tuple.stack_id) && !keep_parentheses_) {
-        arc.ilabel = arc.olabel = 0;  // Stack push/pop.
-      }
-      const StateTuple ntuple(arc.nextstate, stack_id);
-      arc.nextstate = state_table_->FindState(ntuple);
-      PushArc(s, arc);
-    }
-    SetArcs(s);
-  }
-
-  const ParenStack &GetStack() const { return *stack_; }
-
-  const PdtStateTable<StateId, StackId> &GetStateTable() const {
-    return *state_table_;
-  }
-
- private:
-  std::unique_ptr<const Fst<Arc>> fst_;
-  ParenStack *stack_;
-  PdtStateTable<StateId, StackId> *state_table_;
-  const bool own_stack_;
-  const bool own_state_table_;
-  const bool keep_parentheses_;
-
-  MPdtExpandFstImpl &operator=(const MPdtExpandFstImpl &) = delete;
-};
-
-}  // namespace internal
-
-// Expands a multi-pushdown transducer (MPDT) encoded as an FST into an FST.
-// This version is a delayed FST. In the MPDT, some transitions are labeled with
-// open or close parentheses. To be interpreted as an MPDT, the parens for each
-// stack must balance on a path. The open-close parenthesis label
-// pairs are passed using the parens argument, and the assignment of those pairs
-// to stacks is passed using the assignments argument. Expansion enforces the
-// parenthesis constraints. The MPDT must be
-// expandable as an FST.
-//
-// This class attaches interface to implementation and handles
-// reference counting, delegating most methods to ImplToFst.
-template <class A>
-class MPdtExpandFst : public ImplToFst<internal::MPdtExpandFstImpl<A>> {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using StackId = StateId;
-  using ParenStack = internal::MPdtStack<StackId, Label>;
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::MPdtExpandFstImpl<Arc>;
-
-  friend class ArcIterator<MPdtExpandFst<Arc>>;
-  friend class StateIterator<MPdtExpandFst<Arc>>;
-
-  MPdtExpandFst(const Fst<Arc> &fst,
-                const std::vector<std::pair<Label, Label>> &parens,
-                const std::vector<Label> &assignments)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, parens, assignments,
-                                               MPdtExpandFstOptions<Arc>())) {}
-
-  MPdtExpandFst(const Fst<Arc> &fst,
-                const std::vector<std::pair<Label, Label>> &parens,
-                const std::vector<Label> &assignments,
-                const MPdtExpandFstOptions<Arc> &opts)
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, parens, assignments, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  MPdtExpandFst(const MPdtExpandFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this ExpandFst. See Fst<>::Copy() for further doc.
-  MPdtExpandFst<Arc> *Copy(bool safe = false) const override {
-    return new MPdtExpandFst<A>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
-  const ParenStack &GetStack() const { return GetImpl()->GetStack(); }
-
-  const PdtStateTable<StateId, StackId> &GetStateTable() const {
-    return GetImpl()->GetStateTable();
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  void operator=(const MPdtExpandFst &) = delete;
-};
-
-// Specialization for MPdtExpandFst.
-template <class Arc>
-class StateIterator<MPdtExpandFst<Arc>>
-    : public CacheStateIterator<MPdtExpandFst<Arc>> {
- public:
-  explicit StateIterator(const MPdtExpandFst<Arc> &fst)
-      : CacheStateIterator<MPdtExpandFst<Arc>>(fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for MPdtExpandFst.
-template <class Arc>
-class ArcIterator<MPdtExpandFst<Arc>>
-    : public CacheArcIterator<MPdtExpandFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const MPdtExpandFst<Arc> &fst, StateId s)
-      : CacheArcIterator<MPdtExpandFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->ExpandState(s);
-  }
-};
-
-template <class Arc>
-inline void MPdtExpandFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<MPdtExpandFst<Arc>>(*this);
-}
-
-struct MPdtExpandOptions {
-  bool connect;
-  bool keep_parentheses;
-
-  explicit MPdtExpandOptions(bool connect = true, bool keep_parentheses = false)
-      : connect(connect), keep_parentheses(keep_parentheses) {}
-};
-
-// Expands a multi-pushdown transducer (MPDT) encoded as an FST into an FST.
-// This version writes the expanded PDT to a mutable FST. In the MPDT, some
-// transitions are labeled with open or close parentheses. To be interpreted as
-// an MPDT, the parens for each stack must balance on a path. The open-close
-// parenthesis label pair sets are passed using the parens argument, and the
-// assignment of those pairs to stacks is passed using the assignments argument.
-// The expansion enforces the parenthesis constraints. The MPDT must be
-// expandable as an FST.
-template <class Arc>
-void Expand(const Fst<Arc> &ifst,
-            const std::vector<
-            std::pair<typename Arc::Label, typename Arc::Label>> &parens,
-            const std::vector<typename Arc::Label> &assignments,
-            MutableFst<Arc> *ofst, const MPdtExpandOptions &opts) {
-  MPdtExpandFstOptions<Arc> eopts;
-  eopts.gc_limit = 0;
-  eopts.keep_parentheses = opts.keep_parentheses;
-  *ofst = MPdtExpandFst<Arc>(ifst, parens, assignments, eopts);
-  if (opts.connect) Connect(ofst);
-}
-
-// Expands a multi-pushdown transducer (MPDT) encoded as an FST into an FST.
-// This version writes the expanded PDT to a mutable FST. In the MPDT, some
-// transitions are labeled with open or close parentheses. To be interpreted as
-// an MPDT, the parens for each stack must balance on a path. The open-close
-// parenthesis label pair sets are passed using the parens argument, and the
-// assignment of those pairs to stacks is passed using the assignments argument.
-// The expansion enforces the parenthesis constraints. The MPDT must be
-// expandable as an FST.
-template <class Arc>
-void Expand(const Fst<Arc> &ifst,
-            const std::vector<std::pair<typename Arc::Label,
-            typename Arc::Label>> &parens,
-            const std::vector<typename Arc::Label> &assignments,
-            MutableFst<Arc> *ofst, bool connect = true,
-            bool keep_parentheses = false) {
-  const MPdtExpandOptions opts(connect, keep_parentheses);
-  Expand(ifst, parens, assignments, ofst, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_MPDT_EXPAND_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/info.h b/projects/llm_framework/include/fst/extensions/mpdt/info.h
deleted file mode 100644
index 512fcfa8..00000000
--- a/projects/llm_framework/include/fst/extensions/mpdt/info.h
+++ /dev/null
@@ -1,190 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Prints information about an MPDT.
-
-#ifndef FST_EXTENSIONS_MPDT_INFO_H_
-#define FST_EXTENSIONS_MPDT_INFO_H_
-
-#include <unordered_map>
-#include <vector>
-
-#include <fst/extensions/mpdt/mpdt.h>
-#include <fst/fst.h>
-
-namespace fst {
-
-// Compute various information about MPDTs, helper class for mpdtinfo.cc.
-template <class Arc, typename Arc::Label nlevels = 2>
-class MPdtInfo {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  MPdtInfo(const Fst<Arc> &fst,
-           const std::vector<std::pair<Label, Label>> &parens,
-           const std::vector<Label> &assignments);
-
-  const string &FstType() const { return fst_type_; }
-
-  const string &ArcType() const { return Arc::Type(); }
-
-  int64 NumStates() const { return nstates_; }
-
-  int64 NumArcs() const { return narcs_; }
-
-  int64 NumLevels() const { return nlevels; }
-
-  int64 NumOpenParens(Label level) const { return nopen_parens_[level]; }
-
-  int64 NumCloseParens(Label level) const { return nclose_parens_[level]; }
-
-  int64 NumUniqueOpenParens(Label level) const {
-    return nuniq_open_parens_[level];
-  }
-
-  int64 NumUniqueCloseParens(Label level) const {
-    return nuniq_close_parens_[level];
-  }
-  int64 NumOpenParenStates(Label level) const {
-    return nopen_paren_states_[level];
-  }
-
-  int64 NumCloseParenStates(Label level) const {
-    return nclose_paren_states_[level];
-  }
-
-  void Print();
-
- private:
-  string fst_type_;
-  int64 nstates_;
-  int64 narcs_;
-  int64 nopen_parens_[nlevels];
-  int64 nclose_parens_[nlevels];
-  int64 nuniq_open_parens_[nlevels];
-  int64 nuniq_close_parens_[nlevels];
-  int64 nopen_paren_states_[nlevels];
-  int64 nclose_paren_states_[nlevels];
-  bool error_;
-};
-
-template <class Arc, typename Arc::Label nlevels>
-MPdtInfo<Arc, nlevels>::MPdtInfo(
-    const Fst<Arc> &fst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens,
-    const std::vector<typename Arc::Label> &assignments)
-    : fst_type_(fst.Type()), nstates_(0), narcs_(0), error_(false) {
-  std::unordered_map<Label, size_t> paren_map;
-  std::unordered_set<Label> paren_set;
-  std::unordered_map<Label, int> paren_levels;
-  std::unordered_set<StateId> open_paren_state_set;
-  std::unordered_set<StateId> close_paren_state_set;
-  if (parens.size() != assignments.size()) {
-    FSTERROR() << "MPdtInfo: Parens of different size from assignments";
-    error_ = true;
-    return;
-  }
-  for (Label i = 0; i < assignments.size(); ++i) {
-    // Assignments here start at 0, so assuming the human-readable version has
-    // them starting at 1, we should subtract 1 here.
-    Label level = assignments[i] - 1;
-    if (level < 0 || level >= nlevels) {
-      FSTERROR() << "MPdtInfo: Specified level " << level << " out of bounds";
-      error_ = true;
-      return;
-    }
-    const auto &pair = parens[i];
-    paren_levels[pair.first] = level;
-    paren_levels[pair.second] = level;
-    paren_map[pair.first] = i;
-    paren_map[pair.second] = i;
-  }
-  for (Label i = 0; i < nlevels; ++i) {
-    nopen_parens_[i] = 0;
-    nclose_parens_[i] = 0;
-    nuniq_open_parens_[i] = 0;
-    nuniq_close_parens_[i] = 0;
-    nopen_paren_states_[i] = 0;
-    nclose_paren_states_[i] = 0;
-  }
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    ++nstates_;
-    const auto s = siter.Value();
-    for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      ++narcs_;
-      const auto it = paren_map.find(arc.ilabel);
-      if (it != paren_map.end()) {
-        const auto open_paren = parens[it->second].first;
-        const auto close_paren = parens[it->second].second;
-        const auto level = paren_levels[arc.ilabel];
-        if (arc.ilabel == open_paren) {
-          ++nopen_parens_[level];
-          if (!paren_set.count(open_paren)) {
-            ++nuniq_open_parens_[level];
-            paren_set.insert(open_paren);
-          }
-          if (!open_paren_state_set.count(arc.nextstate)) {
-            ++nopen_paren_states_[level];
-            open_paren_state_set.insert(arc.nextstate);
-          }
-        } else {
-          ++nclose_parens_[level];
-          if (!paren_set.count(close_paren)) {
-            ++nuniq_close_parens_[level];
-            paren_set.insert(close_paren);
-          }
-          if (!close_paren_state_set.count(s)) {
-            ++nclose_paren_states_[level];
-            close_paren_state_set.insert(s);
-          }
-        }
-      }
-    }
-  }
-}
-
-template <class Arc, typename Arc::Label nlevels>
-void MPdtInfo<Arc, nlevels>::Print() {
-  const auto old = std::cout.setf(std::ios::left);
-  std::cout.width(50);
-  std::cout << "fst type" << FstType() << std::endl;
-  std::cout.width(50);
-  std::cout << "arc type" << ArcType() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of states" << NumStates() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of arcs" << NumArcs() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of levels" << NumLevels() << std::endl;
-  std::cout.width(50);
-  for (typename Arc::Label i = 0; i < nlevels; ++i) {
-    int level = i + 1;
-    std::cout << "# of open parentheses at levelel " << level << "\t"
-              << NumOpenParens(i) << std::endl;
-    std::cout.width(50);
-    std::cout << "# of close parentheses at levelel " << level << "\t"
-              << NumCloseParens(i) << std::endl;
-    std::cout.width(50);
-    std::cout << "# of unique open parentheses at levelel " << level << "\t"
-              << NumUniqueOpenParens(i) << std::endl;
-    std::cout.width(50);
-    std::cout << "# of unique close parentheses at levelel " << level << "\t"
-              << NumUniqueCloseParens(i) << std::endl;
-    std::cout.width(50);
-    std::cout << "# of open parenthesis dest. states at levelel " << level
-              << "\t" << NumOpenParenStates(i) << std::endl;
-    std::cout.width(50);
-    std::cout << "# of close parenthesis source states at levelel " << level
-              << "\t" << NumCloseParenStates(i) << std::endl;
-    std::cout.width(50);
-  }
-  std::cout.setf(old);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_MPDT_INFO_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/mpdt.h b/projects/llm_framework/include/fst/extensions/mpdt/mpdt.h
deleted file mode 100644
index 77ff7574..00000000
--- a/projects/llm_framework/include/fst/extensions/mpdt/mpdt.h
+++ /dev/null
@@ -1,357 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Common classes for Multi Pushdown Transducer (MPDT) expansion/traversal.
-
-#ifndef FST_EXTENSIONS_MPDT_MPDT_H_
-#define FST_EXTENSIONS_MPDT_MPDT_H_
-
-#include <array>
-#include <functional>
-#include <map>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/extensions/pdt/pdt.h>
-
-namespace fst {
-
-enum MPdtType {
-  MPDT_READ_RESTRICT,   // Can only read from first empty stack
-  MPDT_WRITE_RESTRICT,  // Can only write to first empty stack
-  MPDT_NO_RESTRICT,     // No read-write restrictions
-};
-
-namespace internal {
-
-// PLEASE READ THIS CAREFULLY:
-//
-// When USEVECTOR is set, the stack configurations --- the statewise
-// representation of the StackId's for each substack --- is stored in a vector.
-// I would like to do this using an array for efficiency reasons, thus the
-// definition of StackConfig below. However, while this *works* in that tests
-// pass, etc. It causes a memory leak in the compose and expand tests, evidently
-// in the map[] that is being used to store the mapping between these
-// StackConfigs and the external StackId that the caller sees. There are no
-// memory leaks when I use a vector, only with this StackId. Why there should be
-// memory leaks given that I am not mallocing anything is a mystery. In case you
-// were wondering, clearing the map at the end does not help.
-
-template <typename StackId, typename Level, Level nlevels>
-struct StackConfig {
-  StackConfig() : array_() {}
-
-  StackConfig(const StackConfig<StackId, Level, nlevels> &config) {
-    array_ = config.array_;
-  }
-
-  StackId &operator[](const int index) { return array_[index]; }
-
-  const StackId &operator[](const int index) const { return array_[index]; }
-
-  StackConfig &operator=(const StackConfig<StackId, Level, nlevels> &config) {
-    if (this == &config) return *this;
-    array_ = config.array_;
-    return *this;
-  }
-
-  std::array<StackId, nlevels> array_;
-};
-
-template <typename StackId, typename Level, Level nlevels>
-class CompConfig {
- public:
-  using Config = StackConfig<StackId, Level, nlevels>;
-
-  bool operator()(const Config &x, const Config &y) const {
-    for (Level level = 0; level < nlevels; ++level) {
-      if (x.array_[level] < y.array_[level]) {
-        return true;
-      } else if (x.array_[level] > y.array_[level]) {
-        return false;
-      }
-    }
-    return false;
-  }
-};
-
-// Defines the KeyPair type used as the key to MPdtStack.paren_id_map_. The hash
-// function is provided as a separate struct to match templating syntax.
-template <typename Level>
-struct KeyPair {
-  Level level;
-  size_t underlying_id;
-
-  KeyPair(Level level, size_t id) : level(level), underlying_id(id) {}
-
-  inline bool operator==(const KeyPair<Level> &rhs) const {
-    return level == rhs.level && underlying_id == rhs.underlying_id;
-  }
-};
-
-template <typename Level>
-struct KeyPairHasher {
-  inline size_t operator()(const KeyPair<Level> &keypair) const {
-    return std::hash<Level>()(keypair.level) ^
-           (std::hash<size_t>()(keypair.underlying_id) << 1);
-  }
-};
-
-template <typename StackId, typename Level, Level nlevels = 2,
-          MPdtType restrict = MPDT_READ_RESTRICT>
-class MPdtStack {
- public:
-  using Label = Level;
-  using Config = StackConfig<StackId, Level, nlevels>;
-  using ConfigToStackId =
-      std::map<Config, StackId, CompConfig<StackId, Level, nlevels>>;
-
-  MPdtStack(const std::vector<std::pair<Label, Label>> &parens,
-            const std::vector<Level> &assignments);
-
-  MPdtStack(const MPdtStack &mstack);
-
-  ~MPdtStack() {
-    for (Level level = 0; level < nlevels; ++level) delete stacks_[level];
-  }
-
-  StackId Find(StackId stack_id, Label label);
-
-  // For now we do not implement Pop since this is needed only for
-  // ShortestPath().
-
-  // For Top we find the first non-empty config, and find the paren ID of that
-  // (or -1) if there is none, then map that to the external stack_id to return.
-  ssize_t Top(StackId stack_id) const {
-    if (stack_id == -1) return -1;
-    const auto config = InternalStackIds(stack_id);
-    Level level = 0;
-    StackId underlying_id = -1;
-    for (; level < nlevels; ++level) {
-      if (!Empty(config, level)) {
-        underlying_id = stacks_[level]->Top(config[level]);
-        break;
-      }
-    }
-    if (underlying_id == -1) return -1;
-    const auto it = paren_id_map_.find(KeyPair<Level>(level, underlying_id));
-    if (it == paren_id_map_.end()) return -1;  // NB: shouldn't happen.
-    return it->second;
-  }
-
-  ssize_t ParenId(Label label) const {
-    const auto it = paren_map_.find(label);
-    return it != paren_map_.end() ? it->second : -1;
-  }
-
-  // TODO(rws): For debugging purposes only: remove later.
-  string PrintConfig(const Config &config) const {
-    string result = "[";
-    for (Level i = 0; i < nlevels; ++i) {
-      char s[128];
-      snprintf(s, sizeof(s), "%d", config[i]);
-      result += string(s);
-      if (i < nlevels - 1) result += ", ";
-    }
-    result += "]";
-    return result;
-  }
-
-  bool Error() { return error_; }
-
-  // Each component stack has an internal stack ID for a given configuration and
-  // label.
-  // This function maps a configuration of those to the stack ID the caller
-  // sees.
-  inline StackId ExternalStackId(const Config &config) {
-    const auto it = config_to_stack_id_map_.find(config);
-    StackId result;
-    if (it == config_to_stack_id_map_.end()) {
-      result = next_stack_id_++;
-      config_to_stack_id_map_.insert(
-          std::pair<Config, StackId>(config, result));
-      stack_id_to_config_map_[result] = config;
-    } else {
-      result = it->second;
-    }
-    return result;
-  }
-
-  // Retrieves the internal stack ID for a corresponding external stack ID.
-  inline const Config InternalStackIds(StackId stack_id) const {
-    auto it = stack_id_to_config_map_.find(stack_id);
-    if (it == stack_id_to_config_map_.end()) {
-      it = stack_id_to_config_map_.find(-1);
-    }
-    return it->second;
-  }
-
-  inline bool Empty(const Config &config, Level level) const {
-    return config[level] <= 0;
-  }
-
-  inline bool AllEmpty(const Config &config) {
-    for (Level level = 0; level < nlevels; ++level) {
-      if (!Empty(config, level)) return false;
-    }
-    return true;
-  }
-
-  bool error_;
-  Label min_paren_;
-  Label max_paren_;
-  // Stores level of each paren.
-  std::unordered_map<Label, Label> paren_levels_;
-  std::vector<std::pair<Label, Label>> parens_;  // As in pdt.h.
-  std::unordered_map<Label, size_t> paren_map_;  // As in pdt.h.
-  // Maps between internal paren_id and external paren_id.
-  std::unordered_map<KeyPair<Level>, size_t, KeyPairHasher<Level>>
-      paren_id_map_;
-  // Maps between internal stack ids and external stack id.
-  ConfigToStackId config_to_stack_id_map_;
-  std::unordered_map<StackId, Config> stack_id_to_config_map_;
-  StackId next_stack_id_;
-  // Array of stacks.
-  PdtStack<StackId, Label> *stacks_[nlevels];
-};
-
-template <typename StackId, typename Level, Level nlevels, MPdtType restrict>
-MPdtStack<StackId, Level, nlevels, restrict>::MPdtStack(
-    const std::vector<std::pair<Level, Level>> &parens,  // NB: Label = Level.
-    const std::vector<Level> &assignments)
-    : error_(false),
-      min_paren_(kNoLabel),
-      max_paren_(kNoLabel),
-      parens_(parens),
-      next_stack_id_(1) {
-  using Label = Level;
-  if (parens.size() != assignments.size()) {
-    FSTERROR() << "MPdtStack: Parens of different size from assignments";
-    error_ = true;
-    return;
-  }
-  std::vector<std::pair<Label, Label>> vectors[nlevels];
-  for (Level i = 0; i < assignments.size(); ++i) {
-    // Assignments here start at 0, so assuming the human-readable version has
-    // them starting at 1, we should subtract 1 here
-    const auto level = assignments[i] - 1;
-    if (level < 0 || level >= nlevels) {
-      FSTERROR() << "MPdtStack: Specified level " << level << " out of bounds";
-      error_ = true;
-      return;
-    }
-    const auto &pair = parens[i];
-    vectors[level].push_back(pair);
-    paren_levels_[pair.first] = level;
-    paren_levels_[pair.second] = level;
-    paren_map_[pair.first] = i;
-    paren_map_[pair.second] = i;
-    const KeyPair<Level> key(level, vectors[level].size() - 1);
-    paren_id_map_[key] = i;
-    if (min_paren_ == kNoLabel || pair.first < min_paren_) {
-      min_paren_ = pair.first;
-    }
-    if (pair.second < min_paren_) min_paren_ = pair.second;
-    if (max_paren_ == kNoLabel || pair.first > max_paren_) {
-      max_paren_ = pair.first;
-    }
-    if (pair.second > max_paren_) max_paren_ = pair.second;
-  }
-  using Config = StackConfig<StackId, Level, nlevels>;
-  Config neg_one;
-  Config zero;
-  for (Level level = 0; level < nlevels; ++level) {
-    stacks_[level] = new PdtStack<StackId, Label>(vectors[level]);
-    neg_one[level] = -1;
-    zero[level] = 0;
-  }
-  config_to_stack_id_map_[neg_one] = -1;
-  config_to_stack_id_map_[zero] = 0;
-  stack_id_to_config_map_[-1] = neg_one;
-  stack_id_to_config_map_[0] = zero;
-}
-
-template <typename StackId, typename Level, Level nlevels, MPdtType restrict>
-MPdtStack<StackId, Level, nlevels, restrict>::MPdtStack(
-    const MPdtStack<StackId, Level, nlevels, restrict> &mstack)
-    : error_(mstack.error_),
-      min_paren_(mstack.min_paren_),
-      max_paren_(mstack.max_paren_),
-      parens_(mstack.parens_),
-      next_stack_id_(mstack.next_stack_id_) {
-  for (const auto &kv : mstack.paren_levels_) {
-    paren_levels_[kv.first] = kv.second;
-  }
-  for (const auto &paren : mstack.parens_) parens_.push_back(paren);
-  for (const auto &kv : mstack.paren_map_) {
-    paren_map_[kv.first] = kv.second;
-  }
-  for (const auto &kv : mstack.paren_id_map_) {
-    paren_id_map_[kv.first] = kv.second;
-  }
-  for (auto it = mstack.config_to_stack_id_map_.begin();
-       it != mstack.config_to_stack_id_map_.end(); ++it) {
-    config_to_stack_id_map_[it->first] = it->second;
-  }
-  for (const auto &kv : mstack.stack_id_to_config_map_) {
-    using Config = StackConfig<StackId, Level, nlevels>;
-    const Config config(kv.second);
-    stack_id_to_config_map_[kv.first] = config;
-  }
-  for (Level level = 0; level < nlevels; ++level)
-    stacks_[level] = mstack.stacks_[level];
-}
-
-template <typename StackId, typename Level, Level nlevels, MPdtType restrict>
-StackId MPdtStack<StackId, Level, nlevels, restrict>::Find(StackId stack_id,
-                                                           Level label) {
-  // Non-paren.
-  if (min_paren_ == kNoLabel || label < min_paren_ || label > max_paren_) {
-    return stack_id;
-  }
-  const auto it = paren_map_.find(label);
-  // Non-paren.
-  if (it == paren_map_.end()) return stack_id;
-  ssize_t paren_id = it->second;
-  // Gets the configuration associated with this stack_id.
-  const auto config = InternalStackIds(stack_id);
-  // Gets the level.
-  const auto level = paren_levels_.find(label)->second;
-  // If the label is an open paren we push:
-  //
-  // 1) if the restrict type is not MPDT_WRITE_RESTRICT, or
-  // 2) the restrict type is MPDT_WRITE_RESTRICT, and all the stacks above the
-  // level are empty.
-  if (label == parens_[paren_id].first) {  // Open paren.
-    if (restrict == MPDT_WRITE_RESTRICT) {
-      for (Level upper_level = 0; upper_level < level; ++upper_level) {
-        if (!Empty(config, upper_level)) return -1;  // Non-empty stack blocks.
-      }
-    }
-    // If the label is an close paren we pop:
-    //
-    // 1) if the restrict type is not MPDT_READ_RESTRICT, or
-    // 2) the restrict type is MPDT_READ_RESTRICT, and all the stacks above the
-    // level are empty.
-  } else if (restrict == MPDT_READ_RESTRICT) {
-    for (Level lower_level = 0; lower_level < level; ++lower_level) {
-      if (!Empty(config, lower_level)) return -1;  // Non-empty stack blocks.
-    }
-  }
-  const auto nid = stacks_[level]->Find(config[level], label);
-  // If the new ID is -1, that means that there is no valid transition at the
-  // level we want.
-  if (nid == -1) {
-    return -1;
-  } else {
-    using Config = StackConfig<StackId, Level, nlevels>;
-    Config nconfig(config);
-    nconfig[level] = nid;
-    return ExternalStackId(nconfig);
-  }
-}
-
-}  // namespace internal
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_MPDT_MPDT_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/mpdtlib.h b/projects/llm_framework/include/fst/extensions/mpdt/mpdtlib.h
deleted file mode 100644
index bbf206f1..00000000
--- a/projects/llm_framework/include/fst/extensions/mpdt/mpdtlib.h
+++ /dev/null
@@ -1,18 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// This is an experimental multipush-down transducer (MPDT) library. An MPDT is
-// encoded as an FST, where some transitions are labeled with open or close
-// parentheses, each mated pair of which is associated to one stack. To be
-// interpreted as an MPDT, the parentheses within a stack must balance on a
-// path.
-
-#ifndef FST_EXTENSIONS_MPDT_MPDTLIB_H_
-#define FST_EXTENSIONS_MPDT_MPDTLIB_H_
-
-#include <fst/extensions/mpdt/compose.h>
-#include <fst/extensions/mpdt/expand.h>
-#include <fst/extensions/mpdt/mpdt.h>
-#include <fst/extensions/mpdt/reverse.h>
-
-#endif  // FST_EXTENSIONS_MPDT_MPDTLIB_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/mpdtscript.h b/projects/llm_framework/include/fst/extensions/mpdt/mpdtscript.h
deleted file mode 100644
index 66340992..00000000
--- a/projects/llm_framework/include/fst/extensions/mpdt/mpdtscript.h
+++ /dev/null
@@ -1,156 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Convenience file for including all MPDT operations at once, and/or
-// registering them for new arc types.
-
-#ifndef FST_EXTENSIONS_MPDT_MPDTSCRIPT_H_
-#define FST_EXTENSIONS_MPDT_MPDTSCRIPT_H_
-
-#include <algorithm>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-#include <fst/compose.h>  // for ComposeOptions
-#include <fst/util.h>
-
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/shortest-path.h>
-
-#include <fst/extensions/mpdt/compose.h>
-#include <fst/extensions/mpdt/expand.h>
-#include <fst/extensions/mpdt/info.h>
-#include <fst/extensions/mpdt/reverse.h>
-
-#include <fst/extensions/pdt/pdtscript.h>  // For LabelClassPair,
-                                               // FstClassPair, and to detect
-                                               // any collisions.
-
-namespace fst {
-namespace script {
-
-using MPdtComposeArgs =
-    std::tuple<const FstClass &, const FstClass &,
-               const std::vector<LabelPair> &, const std::vector<int64> &,
-               MutableFstClass *, const MPdtComposeOptions &, bool>;
-
-template <class Arc>
-void MPdtCompose(MPdtComposeArgs *args) {
-  const Fst<Arc> &ifst1 = *(std::get<0>(*args).GetFst<Arc>());
-  const Fst<Arc> &ifst2 = *(std::get<1>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<4>(*args)->GetMutableFst<Arc>();
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<2>(*args).size());
-  std::copy(std::get<2>(*args).begin(), std::get<2>(*args).end(),
-            typed_parens.begin());
-  using Level = typename Arc::Label;
-  std::vector<Level> typed_assignments(std::get<3>(*args).size());
-  std::copy(std::get<3>(*args).begin(), std::get<3>(*args).end(),
-            typed_assignments.begin());
-  if (std::get<6>(*args)) {
-    Compose(ifst1, typed_parens, typed_assignments, ifst2, ofst,
-            std::get<5>(*args));
-  } else {
-    Compose(ifst1, ifst2, typed_parens, typed_assignments, ofst,
-            std::get<5>(*args));
-  }
-}
-
-void MPdtCompose(const FstClass &ifst1, const FstClass &ifst2,
-                 const std::vector<LabelPair> &parens,
-                 const std::vector<int64> &assignments, MutableFstClass *ofst,
-                 const MPdtComposeOptions &copts, bool left_pdt);
-
-using MPdtExpandArgs =
-    std::tuple<const FstClass &, const std::vector<LabelPair> &,
-               const std::vector<int64> &, MutableFstClass *,
-               const MPdtExpandOptions &>;
-
-template <class Arc>
-void MPdtExpand(MPdtExpandArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<3>(*args)->GetMutableFst<Arc>();
-  // In case Arc::Label is not the same as FstClass::Label, we make copies.
-  // Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_parens.begin());
-  using Level = typename Arc::Label;
-  std::vector<Level> typed_assignments(std::get<2>(*args).size());
-  std::copy(std::get<2>(*args).begin(), std::get<2>(*args).end(),
-            typed_assignments.begin());
-  Expand(fst, typed_parens, typed_assignments, ofst,
-         MPdtExpandOptions(std::get<4>(*args).connect,
-                           std::get<4>(*args).keep_parentheses));
-}
-
-void MPdtExpand(const FstClass &ifst, const std::vector<LabelPair> &parens,
-                const std::vector<int64> &assignments, MutableFstClass *ofst,
-                const MPdtExpandOptions &opts);
-
-using MPdtReverseArgs =
-    std::tuple<const FstClass &, const std::vector<LabelPair> &,
-               std::vector<int64> *, MutableFstClass *>;
-
-template <class Arc>
-void MPdtReverse(MPdtReverseArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<3>(*args)->GetMutableFst<Arc>();
-  // In case Arc::Label is not the same as FstClass::Label, we make copies.
-  // Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_parens.begin());
-  using Level = typename Arc::Label;
-  std::vector<Level> typed_assignments(std::get<2>(*args)->size());
-  std::copy(std::get<2>(*args)->begin(), std::get<2>(*args)->end(),
-            typed_assignments.begin());
-  Reverse(fst, typed_parens, &typed_assignments, ofst);
-  // Reassign stack assignments to input assignment vector.
-  std::copy(typed_assignments.begin(), typed_assignments.end(),
-            std::get<2>(*args)->begin());
-}
-
-void MPdtReverse(const FstClass &ifst, const std::vector<LabelPair> &parens,
-                 std::vector<int64> *assignments, MutableFstClass *ofst);
-
-using PrintMPdtInfoArgs =
-    std::tuple<const FstClass &, const std::vector<LabelPair> &,
-               const std::vector<int64> &>;
-
-template <class Arc>
-void PrintMPdtInfo(PrintMPdtInfoArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  // In case Arc::Label is not the same as FstClass::Label, we make copies.
-  // Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_parens.begin());
-  using Level = typename Arc::Label;
-  std::vector<Level> typed_assignments(std::get<2>(*args).size());
-  std::copy(std::get<2>(*args).begin(), std::get<2>(*args).end(),
-            typed_assignments.begin());
-  MPdtInfo<Arc> mpdtinfo(fst, typed_parens, typed_assignments);
-  mpdtinfo.Print();
-}
-
-void PrintMPdtInfo(const FstClass &ifst, const std::vector<LabelPair> &parens,
-                   const std::vector<int64> &assignments);
-
-}  // namespace script
-}  // namespace fst
-
-#define REGISTER_FST_MPDT_OPERATIONS(ArcType)                    \
-  REGISTER_FST_OPERATION(MPdtCompose, ArcType, MPdtComposeArgs); \
-  REGISTER_FST_OPERATION(MPdtExpand, ArcType, MPdtExpandArgs);   \
-  REGISTER_FST_OPERATION(MPdtReverse, ArcType, MPdtReverseArgs); \
-  REGISTER_FST_OPERATION(PrintMPdtInfo, ArcType, PrintMPdtInfoArgs)
-#endif  // FST_EXTENSIONS_MPDT_MPDTSCRIPT_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/read_write_utils.h b/projects/llm_framework/include/fst/extensions/mpdt/read_write_utils.h
deleted file mode 100644
index 38f9f963..00000000
--- a/projects/llm_framework/include/fst/extensions/mpdt/read_write_utils.h
+++ /dev/null
@@ -1,86 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Definition of ReadLabelTriples based on ReadLabelPairs, like that in
-// nlp/fst/lib/util.h for pairs, and similarly for WriteLabelTriples.
-
-#ifndef FST_EXTENSIONS_MPDT_READ_WRITE_UTILS_H_
-#define FST_EXTENSIONS_MPDT_READ_WRITE_UTILS_H_
-
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fstream>
-#include <fst/test-properties.h>
-
-namespace fst {
-
-// Returns true on success.
-template <typename Label>
-bool ReadLabelTriples(const string &filename,
-                      std::vector<std::pair<Label, Label>> *pairs,
-                      std::vector<Label> *assignments,
-                      bool allow_negative = false) {
-  std::ifstream fstrm(filename);
-  if (!fstrm) {
-    LOG(ERROR) << "ReadIntTriples: Can't open file: " << filename;
-    return false;
-  }
-  static constexpr auto kLineLen = 8096;
-  char line[kLineLen];
-  size_t nline = 0;
-  pairs->clear();
-  while (fstrm.getline(line, kLineLen)) {
-    ++nline;
-    std::vector<char *> col;
-    SplitString(line, "\n\t ", &col, true);
-    // Empty line or comment?
-    if (col.empty() || col[0][0] == '\0' || col[0][0] == '#') continue;
-    if (col.size() != 3) {
-      LOG(ERROR) << "ReadLabelTriples: Bad number of columns, "
-                 << "file = " << filename << ", line = " << nline;
-      return false;
-    }
-    bool err;
-    const Label i1 = StrToInt64(col[0], filename, nline, allow_negative, &err);
-    if (err) return false;
-    const Label i2 = StrToInt64(col[1], filename, nline, allow_negative, &err);
-    if (err) return false;
-    using Level = Label;
-    const Level i3 = StrToInt64(col[2], filename, nline, allow_negative, &err);
-    if (err) return false;
-    pairs->push_back(std::make_pair(i1, i2));
-    assignments->push_back(i3);
-  }
-  return true;
-}
-
-// Returns true on success.
-template <typename Label>
-bool WriteLabelTriples(const string &filename,
-                       const std::vector<std::pair<Label, Label>> &pairs,
-                       const std::vector<Label> &assignments) {
-  if (pairs.size() != assignments.size()) {
-    LOG(ERROR) << "WriteLabelTriples: Pairs and assignments of different sizes";
-    return false;
-  }
-  std::ofstream fstrm(filename);
-  if (!fstrm) {
-    LOG(ERROR) << "WriteLabelTriples: Can't open file: " << filename;
-    return false;
-  }
-  for (size_t n = 0; n < pairs.size(); ++n)
-    fstrm << pairs[n].first << "\t" << pairs[n].second << "\t" << assignments[n]
-          << "\n";
-  if (!fstrm) {
-    LOG(ERROR) << "WriteLabelTriples: Write failed: "
-               << (filename.empty() ? "standard output" : filename);
-    return false;
-  }
-  return true;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_MPDT_READ_WRITE_UTILS_H_
diff --git a/projects/llm_framework/include/fst/extensions/mpdt/reverse.h b/projects/llm_framework/include/fst/extensions/mpdt/reverse.h
deleted file mode 100644
index 413dd3f0..00000000
--- a/projects/llm_framework/include/fst/extensions/mpdt/reverse.h
+++ /dev/null
@@ -1,54 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Reverses an MPDT.
-
-#ifndef FST_EXTENSIONS_MPDT_REVERSE_H_
-#define FST_EXTENSIONS_MPDT_REVERSE_H_
-
-#include <limits>
-#include <vector>
-
-#include <fst/mutable-fst.h>
-#include <fst/relabel.h>
-#include <fst/reverse.h>
-
-namespace fst {
-
-// Reverses a multi-stack pushdown transducer (MPDT) encoded as an FST.
-template <class Arc, class RevArc>
-void Reverse(
-    const Fst<Arc> &ifst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens,
-    std::vector<typename Arc::Label> *assignments, MutableFst<RevArc> *ofst) {
-  using Label = typename Arc::Label;
-  // Reverses FST component.
-  Reverse(ifst, ofst);
-  // Exchanges open and close parenthesis pairs.
-  std::vector<std::pair<Label, Label>> relabel_pairs;
-  relabel_pairs.reserve(2 * parens.size());
-  for (const auto &pair : parens) {
-    relabel_pairs.emplace_back(pair.first, pair.second);
-    relabel_pairs.emplace_back(pair.second, pair.first);
-  }
-  Relabel(ofst, relabel_pairs, relabel_pairs);
-  // Computes new bounds for the stack assignments.
-  Label max_level = -1;
-  Label min_level = std::numeric_limits<Label>::max();
-  for (const auto assignment : *assignments) {
-    if (assignment < min_level) {
-      min_level = assignment;
-    } else if (assignment > max_level) {
-      max_level = assignment;
-    }
-  }
-  // Actually reverses stack assignments.
-  for (auto &assignment : *assignments) {
-    assignment = (max_level - assignment) + min_level;
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_MPDT_REVERSE_H_
diff --git a/projects/llm_framework/include/fst/extensions/ngram/bitmap-index.h b/projects/llm_framework/include/fst/extensions/ngram/bitmap-index.h
deleted file mode 100644
index 09ad6d7d..00000000
--- a/projects/llm_framework/include/fst/extensions/ngram/bitmap-index.h
+++ /dev/null
@@ -1,168 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_NGRAM_BITMAP_INDEX_H_
-#define FST_EXTENSIONS_NGRAM_BITMAP_INDEX_H_
-
-#include <utility>
-#include <vector>
-
-#include <fst/compat.h>
-
-// This class is a bitstring storage class with an index that allows
-// seeking to the Nth set or clear bit in time O(Log(N)) where N is
-// the length of the bit vector.  In addition, it allows counting set or
-// clear bits over ranges in constant time.
-//
-// This is accomplished by maintaining an "secondary" index of limited
-// size in bits that maintains a running count of the number of bits set
-// in each block of bitmap data.  A block is defined as the number of
-// uint64 values that can fit in the secondary index before an overflow
-// occurs.
-//
-// To handle overflows, a "primary" index containing a running count of
-// bits set in each block is created using the type uint64.
-
-namespace fst {
-
-class BitmapIndex {
- public:
-  static size_t StorageSize(size_t size) {
-    return ((size + kStorageBlockMask) >> kStorageLogBitSize);
-  }
-
-  BitmapIndex() : bits_(nullptr), size_(0) {}
-
-  bool Get(size_t index) const {
-    return (bits_[index >> kStorageLogBitSize] &
-            (kOne << (index & kStorageBlockMask))) != 0;
-  }
-
-  static void Set(uint64* bits, size_t index) {
-    bits[index >> kStorageLogBitSize] |= (kOne << (index & kStorageBlockMask));
-  }
-
-  static void Clear(uint64* bits, size_t index) {
-    bits[index >> kStorageLogBitSize] &= ~(kOne << (index & kStorageBlockMask));
-  }
-
-  size_t Bits() const { return size_; }
-
-  size_t ArraySize() const { return StorageSize(size_); }
-
-  // Returns the number of one bits in the bitmap
-  size_t GetOnesCount() const {
-    return primary_index_[primary_index_size() - 1];
-  }
-
-  // Returns the number of one bits in positions 0 to limit - 1.
-  // REQUIRES: limit <= Bits()
-  size_t Rank1(size_t end) const;
-
-  // Returns the number of one bits in the range start to end - 1.
-  // REQUIRES: limit <= Bits()
-  size_t GetOnesCountInRange(size_t start, size_t end) const {
-    return Rank1(end) - Rank1(start);
-  }
-
-  // Returns the number of zero bits in positions 0 to limit - 1.
-  // REQUIRES: limit <= Bits()
-  size_t Rank0(size_t end) const { return end - Rank1(end); }
-
-  // Returns the number of zero bits in the range start to end - 1.
-  // REQUIRES: limit <= Bits()
-  size_t GetZeroesCountInRange(size_t start, size_t end) const {
-    return end - start - GetOnesCountInRange(start, end);
-  }
-
-  // Return true if any bit between begin inclusive and end exclusive
-  // is set.  0 <= begin <= end <= Bits() is required.
-  //
-  bool TestRange(size_t start, size_t end) const {
-    return Rank1(end) > Rank1(start);
-  }
-
-  // Returns the offset to the nth set bit (zero based)
-  // or Bits() if index >= number of ones
-  size_t Select1(size_t bit_index) const;
-
-  // Returns the offset to the nth clear bit (zero based)
-  // or Bits() if index > number of
-  size_t Select0(size_t bit_index) const;
-
-  // Returns the offset of the nth and nth+1 clear bit (zero based),
-  // equivalent to two calls to Select0, but more efficient.
-  std::pair<size_t, size_t> Select0s(size_t bit_index) const;
-
-  // Rebuilds from index for the associated Bitmap, should be called
-  // whenever changes have been made to the Bitmap or else behavior
-  // of the indexed bitmap methods will be undefined.
-  void BuildIndex(const uint64* bits, size_t size);
-
-  // the secondary index accumulates counts until it can possibly overflow
-  // this constant computes the number of uint64 units that can fit into
-  // units the size of uint16.
-  static const uint64 kOne = 1;
-  static const uint32 kStorageBitSize = 64;
-  static const uint32 kStorageLogBitSize = 6;
-  static const uint32 kSecondaryBlockSize =
-      ((1 << 16) - 1) >> kStorageLogBitSize;
-
- private:
-  static const uint32 kStorageBlockMask = kStorageBitSize - 1;
-
-  // returns, from the index, the count of ones up to array_index
-  size_t get_index_ones_count(size_t array_index) const;
-
-  // because the indexes, both primary and secondary, contain a running
-  // count of the population of one bits contained in [0,i), there is
-  // no reason to have an element in the zeroth position as this value would
-  // necessarily be zero.  (The bits are indexed in a zero based way.)  Thus
-  // we don't store the 0th element in either index.  Both of the following
-  // functions, if greater than 0, must be decremented by one before retreiving
-  // the value from the corresponding array.
-  // returns the 1 + the block that contains the bitindex in question
-  // the inverted version works the same but looks for zeros using an inverted
-  // view of the index
-  size_t find_primary_block(size_t bit_index) const;
-
-  size_t find_inverted_primary_block(size_t bit_index) const;
-
-  // similarly, the secondary index (which resets its count to zero at
-  // the end of every kSecondaryBlockSize entries) does not store the element
-  // at 0.  Note that the rem_bit_index parameter is the number of bits
-  // within the secondary block, after the bits accounted for by the primary
-  // block have been removed (i.e. the remaining bits)  And, because we
-  // reset to zero with each new block, there is no need to store those
-  // actual zeros.
-  // returns 1 + the secondary block that contains the bitindex in question
-  size_t find_secondary_block(size_t block, size_t rem_bit_index) const;
-
-  size_t find_inverted_secondary_block(size_t block,
-                                       size_t rem_bit_index) const;
-
-  // We create a primary index based upon the number of secondary index
-  // blocks.  The primary index uses fields wide enough to accomodate any
-  // index of the bitarray so cannot overflow
-  // The primary index is the actual running
-  // count of one bits set for all blocks (and, thus, all uint64s).
-  size_t primary_index_size() const {
-    return (ArraySize() + kSecondaryBlockSize - 1) / kSecondaryBlockSize;
-  }
-
-  const uint64* bits_;
-  size_t size_;
-
-  // The primary index contains the running popcount of all blocks
-  // which means the nth value contains the popcounts of
-  // [0,n*kSecondaryBlockSize], however, the 0th element is omitted.
-  std::vector<uint32> primary_index_;
-  // The secondary index contains the running popcount of the associated
-  // bitmap.  It is the same length (in units of uint16) as the
-  // bitmap's map is in units of uint64s.
-  std::vector<uint16> secondary_index_;
-};
-
-}  // end namespace fst
-
-#endif  // FST_EXTENSIONS_NGRAM_BITMAP_INDEX_H_
diff --git a/projects/llm_framework/include/fst/extensions/ngram/ngram-fst.h b/projects/llm_framework/include/fst/extensions/ngram/ngram-fst.h
deleted file mode 100644
index 8183a8c0..00000000
--- a/projects/llm_framework/include/fst/extensions/ngram/ngram-fst.h
+++ /dev/null
@@ -1,1027 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// NgramFst implements a n-gram language model based upon the LOUDS data
-// structure.  Please refer to "Unary Data Structures for Language Models"
-// http://research.google.com/pubs/archive/37218.pdf
-
-#ifndef FST_EXTENSIONS_NGRAM_NGRAM_FST_H_
-#define FST_EXTENSIONS_NGRAM_NGRAM_FST_H_
-
-#include <stddef.h>
-#include <string.h>
-#include <algorithm>
-#include <iostream>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/log.h>
-#include <fstream>
-#include <fst/extensions/ngram/bitmap-index.h>
-#include <fst/fstlib.h>
-#include <fst/mapped-file.h>
-
-namespace fst {
-template <class A>
-class NGramFst;
-template <class A>
-class NGramFstMatcher;
-
-// Instance data containing mutable state for bookkeeping repeated access to
-// the same state.
-template <class A>
-struct NGramFstInst {
-  typedef typename A::Label Label;
-  typedef typename A::StateId StateId;
-  typedef typename A::Weight Weight;
-  StateId state_;
-  size_t num_futures_;
-  size_t offset_;
-  size_t node_;
-  StateId node_state_;
-  std::vector<Label> context_;
-  StateId context_state_;
-  NGramFstInst()
-      : state_(kNoStateId),
-        node_state_(kNoStateId),
-        context_state_(kNoStateId) {}
-};
-
-namespace internal {
-
-// Implementation class for LOUDS based NgramFst interface.
-template <class A>
-class NGramFstImpl : public FstImpl<A> {
-  using FstImpl<A>::SetInputSymbols;
-  using FstImpl<A>::SetOutputSymbols;
-  using FstImpl<A>::SetType;
-  using FstImpl<A>::WriteHeader;
-
-  friend class ArcIterator<NGramFst<A>>;
-  friend class NGramFstMatcher<A>;
-
- public:
-  using FstImpl<A>::InputSymbols;
-  using FstImpl<A>::SetProperties;
-  using FstImpl<A>::Properties;
-
-  typedef A Arc;
-  typedef typename A::Label Label;
-  typedef typename A::StateId StateId;
-  typedef typename A::Weight Weight;
-
-  NGramFstImpl() {
-    SetType("ngram");
-    SetInputSymbols(nullptr);
-    SetOutputSymbols(nullptr);
-    SetProperties(kStaticProperties);
-  }
-
-  NGramFstImpl(const Fst<A> &fst, std::vector<StateId> *order_out);
-
-  explicit NGramFstImpl(const Fst<A> &fst) : NGramFstImpl(fst, nullptr) {}
-
-  NGramFstImpl(const NGramFstImpl &other) {
-    FSTERROR() << "Copying NGramFst Impls is not supported, use safe = false.";
-    SetProperties(kError, kError);
-  }
-
-  ~NGramFstImpl() override {
-    if (owned_) {
-      delete[] data_;
-    }
-  }
-
-  static NGramFstImpl<A> *Read(std::istream &strm,  // NOLINT
-                               const FstReadOptions &opts) {
-    NGramFstImpl<A> *impl = new NGramFstImpl();
-    FstHeader hdr;
-    if (!impl->ReadHeader(strm, opts, kMinFileVersion, &hdr)) return 0;
-    uint64 num_states, num_futures, num_final;
-    const size_t offset =
-        sizeof(num_states) + sizeof(num_futures) + sizeof(num_final);
-    // Peek at num_states and num_futures to see how much more needs to be read.
-    strm.read(reinterpret_cast<char *>(&num_states), sizeof(num_states));
-    strm.read(reinterpret_cast<char *>(&num_futures), sizeof(num_futures));
-    strm.read(reinterpret_cast<char *>(&num_final), sizeof(num_final));
-    size_t size = Storage(num_states, num_futures, num_final);
-    MappedFile *data_region = MappedFile::Allocate(size);
-    char *data = reinterpret_cast<char *>(data_region->mutable_data());
-    // Copy num_states, num_futures and num_final back into data.
-    memcpy(data, reinterpret_cast<char *>(&num_states), sizeof(num_states));
-    memcpy(data + sizeof(num_states), reinterpret_cast<char *>(&num_futures),
-           sizeof(num_futures));
-    memcpy(data + sizeof(num_states) + sizeof(num_futures),
-           reinterpret_cast<char *>(&num_final), sizeof(num_final));
-    strm.read(data + offset, size - offset);
-    if (strm.fail()) {
-      delete impl;
-      return nullptr;
-    }
-    impl->Init(data, false, data_region);
-    return impl;
-  }
-
-  bool Write(std::ostream &strm,  // NOLINT
-             const FstWriteOptions &opts) const {
-    FstHeader hdr;
-    hdr.SetStart(Start());
-    hdr.SetNumStates(num_states_);
-    WriteHeader(strm, opts, kFileVersion, &hdr);
-    strm.write(data_, StorageSize());
-    return !strm.fail();
-  }
-
-  StateId Start() const { return start_; }
-
-  Weight Final(StateId state) const {
-    if (final_index_.Get(state)) {
-      return final_probs_[final_index_.Rank1(state)];
-    } else {
-      return Weight::Zero();
-    }
-  }
-
-  size_t NumArcs(StateId state, NGramFstInst<A> *inst = nullptr) const {
-    if (inst == nullptr) {
-      const std::pair<size_t, size_t> zeros =
-          (state == 0) ? select_root_ : future_index_.Select0s(state);
-      return zeros.second - zeros.first - 1;
-    }
-    SetInstFuture(state, inst);
-    return inst->num_futures_ + ((state == 0) ? 0 : 1);
-  }
-
-  size_t NumInputEpsilons(StateId state) const {
-    // State 0 has no parent, thus no backoff.
-    if (state == 0) return 0;
-    return 1;
-  }
-
-  size_t NumOutputEpsilons(StateId state) const {
-    return NumInputEpsilons(state);
-  }
-
-  StateId NumStates() const { return num_states_; }
-
-  void InitStateIterator(StateIteratorData<A> *data) const {
-    data->base = 0;
-    data->nstates = num_states_;
-  }
-
-  static size_t Storage(uint64 num_states, uint64 num_futures,
-                        uint64 num_final) {
-    uint64 b64;
-    Weight weight;
-    Label label;
-    size_t offset =
-        sizeof(num_states) + sizeof(num_futures) + sizeof(num_final);
-    offset +=
-        sizeof(b64) * (BitmapIndex::StorageSize(num_states * 2 + 1) +
-                       BitmapIndex::StorageSize(num_futures + num_states + 1) +
-                       BitmapIndex::StorageSize(num_states));
-    offset += (num_states + 1) * sizeof(label) + num_futures * sizeof(label);
-    // Pad for alignemnt, see
-    // http://en.wikipedia.org/wiki/Data_structure_alignment#Computing_padding
-    offset = (offset + sizeof(weight) - 1) & ~(sizeof(weight) - 1);
-    offset += (num_states + 1) * sizeof(weight) + num_final * sizeof(weight) +
-              (num_futures + 1) * sizeof(weight);
-    return offset;
-  }
-
-  void SetInstFuture(StateId state, NGramFstInst<A> *inst) const {
-    if (inst->state_ != state) {
-      inst->state_ = state;
-      const std::pair<size_t, size_t> zeros = future_index_.Select0s(state);
-      inst->num_futures_ = zeros.second - zeros.first - 1;
-      inst->offset_ = future_index_.Rank1(zeros.first + 1);
-    }
-  }
-
-  void SetInstNode(NGramFstInst<A> *inst) const {
-    if (inst->node_state_ != inst->state_) {
-      inst->node_state_ = inst->state_;
-      inst->node_ = context_index_.Select1(inst->state_);
-    }
-  }
-
-  void SetInstContext(NGramFstInst<A> *inst) const {
-    SetInstNode(inst);
-    if (inst->context_state_ != inst->state_) {
-      inst->context_state_ = inst->state_;
-      inst->context_.clear();
-      size_t node = inst->node_;
-      while (node != 0) {
-        inst->context_.push_back(context_words_[context_index_.Rank1(node)]);
-        node = context_index_.Select1(context_index_.Rank0(node) - 1);
-      }
-    }
-  }
-
-  // Access to the underlying representation
-  const char *GetData(size_t *data_size) const {
-    *data_size = StorageSize();
-    return data_;
-  }
-
-  void Init(const char *data, bool owned, MappedFile *file = nullptr);
-
-  const std::vector<Label> &GetContext(StateId s, NGramFstInst<A> *inst) const {
-    SetInstFuture(s, inst);
-    SetInstContext(inst);
-    return inst->context_;
-  }
-
-  size_t StorageSize() const {
-    return Storage(num_states_, num_futures_, num_final_);
-  }
-
-  void GetStates(const std::vector<Label> &context,
-                 std::vector<StateId> *states) const;
-
- private:
-  StateId Transition(const std::vector<Label> &context, Label future) const;
-
-  // Properties always true for this Fst class.
-  static const uint64 kStaticProperties =
-      kAcceptor | kIDeterministic | kODeterministic | kEpsilons | kIEpsilons |
-      kOEpsilons | kILabelSorted | kOLabelSorted | kWeighted | kCyclic |
-      kInitialAcyclic | kNotTopSorted | kAccessible | kCoAccessible |
-      kNotString | kExpanded;
-  // Current file format version.
-  static const int kFileVersion = 4;
-  // Minimum file format version supported.
-  static const int kMinFileVersion = 4;
-
-  std::unique_ptr<MappedFile> data_region_;
-  const char *data_ = nullptr;
-  bool owned_ = false;  // True if we own data_
-  StateId start_ = fst::kNoStateId;
-  uint64 num_states_ = 0;
-  uint64 num_futures_ = 0;
-  uint64 num_final_ = 0;
-  std::pair<size_t, size_t> select_root_;
-  const Label *root_children_ = nullptr;
-  // borrowed references
-  const uint64 *context_ = nullptr;
-  const uint64 *future_ = nullptr;
-  const uint64 *final_ = nullptr;
-  const Label *context_words_ = nullptr;
-  const Label *future_words_ = nullptr;
-  const Weight *backoff_ = nullptr;
-  const Weight *final_probs_ = nullptr;
-  const Weight *future_probs_ = nullptr;
-  BitmapIndex context_index_;
-  BitmapIndex future_index_;
-  BitmapIndex final_index_;
-};
-
-template <typename A>
-inline void NGramFstImpl<A>::GetStates(
-    const std::vector<Label> &context,
-    std::vector<typename A::StateId> *states) const {
-  states->clear();
-  states->push_back(0);
-  typename std::vector<Label>::const_reverse_iterator cit = context.rbegin();
-  const Label *children = root_children_;
-  size_t num_children = select_root_.second - 2;
-  const Label *loc = std::lower_bound(children, children + num_children, *cit);
-  if (loc == children + num_children || *loc != *cit) return;
-  size_t node = 2 + loc - children;
-  states->push_back(context_index_.Rank1(node));
-  if (context.size() == 1) return;
-  size_t node_rank = context_index_.Rank1(node);
-  std::pair<size_t, size_t> zeros =
-      node_rank == 0 ? select_root_ : context_index_.Select0s(node_rank);
-  size_t first_child = zeros.first + 1;
-  ++cit;
-  if (context_index_.Get(first_child) != false) {
-    size_t last_child = zeros.second - 1;
-    while (cit != context.rend()) {
-      children = context_words_ + context_index_.Rank1(first_child);
-      loc = std::lower_bound(children, children + last_child - first_child + 1,
-                             *cit);
-      if (loc == children + last_child - first_child + 1 || *loc != *cit) {
-        break;
-      }
-      ++cit;
-      node = first_child + loc - children;
-      states->push_back(context_index_.Rank1(node));
-      node_rank = context_index_.Rank1(node);
-      zeros =
-          node_rank == 0 ? select_root_ : context_index_.Select0s(node_rank);
-      first_child = zeros.first + 1;
-      if (context_index_.Get(first_child) == false) break;
-      last_child = zeros.second - 1;
-    }
-  }
-}
-
-}  // namespace internal
-
-/*****************************************************************************/
-template <class A>
-class NGramFst : public ImplToExpandedFst<internal::NGramFstImpl<A>> {
-  friend class ArcIterator<NGramFst<A>>;
-  friend class NGramFstMatcher<A>;
-
- public:
-  typedef A Arc;
-  typedef typename A::StateId StateId;
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-  typedef internal::NGramFstImpl<A> Impl;
-
-  explicit NGramFst(const Fst<A> &dst)
-      : ImplToExpandedFst<Impl>(std::make_shared<Impl>(dst, nullptr)) {}
-
-  NGramFst(const Fst<A> &fst, std::vector<StateId> *order_out)
-      : ImplToExpandedFst<Impl>(std::make_shared<Impl>(fst, order_out)) {}
-
-  // Because the NGramFstImpl is a const stateless data structure, there
-  // is never a need to do anything beside copy the reference.
-  NGramFst(const NGramFst<A> &fst, bool safe = false)
-      : ImplToExpandedFst<Impl>(fst, false) {}
-
-  NGramFst() : ImplToExpandedFst<Impl>(std::make_shared<Impl>()) {}
-
-  // Non-standard constructor to initialize NGramFst directly from data.
-  NGramFst(const char *data, bool owned)
-      : ImplToExpandedFst<Impl>(std::make_shared<Impl>()) {
-    GetMutableImpl()->Init(data, owned, nullptr);
-  }
-
-  // Get method that gets the data associated with Init().
-  const char *GetData(size_t *data_size) const {
-    return GetImpl()->GetData(data_size);
-  }
-
-  const std::vector<Label> GetContext(StateId s) const {
-    return GetImpl()->GetContext(s, &inst_);
-  }
-
-  // Consumes as much as possible of context from right to left, returns the
-  // the states corresponding to the increasingly conditioned input sequence.
-  void GetStates(const std::vector<Label> &context,
-                 std::vector<StateId> *state) const {
-    return GetImpl()->GetStates(context, state);
-  }
-
-  size_t NumArcs(StateId s) const override {
-    return GetImpl()->NumArcs(s, &inst_);
-  }
-
-  NGramFst<A> *Copy(bool safe = false) const override {
-    return new NGramFst(*this, safe);
-  }
-
-  static NGramFst<A> *Read(std::istream &strm, const FstReadOptions &opts) {
-    Impl *impl = Impl::Read(strm, opts);
-    return impl ? new NGramFst<A>(std::shared_ptr<Impl>(impl)) : nullptr;
-  }
-
-  static NGramFst<A> *Read(const string &filename) {
-    if (!filename.empty()) {
-      std::ifstream strm(filename,
-                              std::ios_base::in | std::ios_base::binary);
-      if (!strm.good()) {
-        LOG(ERROR) << "NGramFst::Read: Can't open file: " << filename;
-        return nullptr;
-      }
-      return Read(strm, FstReadOptions(filename));
-    } else {
-      return Read(std::cin, FstReadOptions("standard input"));
-    }
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return GetImpl()->Write(strm, opts);
-  }
-
-  bool Write(const string &filename) const override {
-    return Fst<A>::WriteFile(filename);
-  }
-
-  inline void InitStateIterator(StateIteratorData<A> *data) const override {
-    GetImpl()->InitStateIterator(data);
-  }
-
-  inline void InitArcIterator(StateId s,
-                              ArcIteratorData<A> *data) const override;
-
-  MatcherBase<A> *InitMatcher(MatchType match_type) const override {
-    return new NGramFstMatcher<A>(this, match_type);
-  }
-
-  size_t StorageSize() const { return GetImpl()->StorageSize(); }
-
-  static bool HasRequiredProps(const Fst<A> &fst) {
-    static const auto props =
-        kAcceptor | kIDeterministic | kILabelSorted | kIEpsilons | kAccessible;
-    return fst.Properties(props, true) == props;
-  }
-
-  static bool HasRequiredStructure(const Fst<A> &fst) {
-    if (!HasRequiredProps(fst)) {
-      return false;
-    }
-    typename A::StateId unigram = fst.Start();
-    while (true) {  // Follows epsilon arc chain to find unigram state.
-      if (unigram == fst::kNoStateId) return false;  // No unigram state.
-      typename fst::ArcIterator<Fst<A>> aiter(fst, unigram);
-      if (aiter.Done() || aiter.Value().ilabel != 0) break;
-      unigram = aiter.Value().nextstate;
-      aiter.Next();
-    }
-    // Other requirement: all states other than unigram an epsilon arc.
-    for (fst::StateIterator<Fst<A>> siter(fst); !siter.Done();
-         siter.Next()) {
-      const typename A::StateId &state = siter.Value();
-      fst::ArcIterator<Fst<A>> aiter(fst, state);
-      if (state != unigram) {
-        if (aiter.Done()) return false;
-        if (aiter.Value().ilabel != 0) return false;
-        aiter.Next();
-        if (!aiter.Done() && aiter.Value().ilabel == 0) return false;
-      }
-    }
-    return true;
-  }
-
- private:
-  using ImplToExpandedFst<Impl, ExpandedFst<A>>::GetImpl;
-  using ImplToExpandedFst<Impl, ExpandedFst<A>>::GetMutableImpl;
-
-  explicit NGramFst(std::shared_ptr<Impl> impl)
-      : ImplToExpandedFst<Impl>(impl) {}
-
-  mutable NGramFstInst<A> inst_;
-};
-
-template <class A>
-inline void NGramFst<A>::InitArcIterator(StateId s,
-                                         ArcIteratorData<A> *data) const {
-  GetImpl()->SetInstFuture(s, &inst_);
-  GetImpl()->SetInstNode(&inst_);
-  data->base = new ArcIterator<NGramFst<A>>(*this, s);
-}
-
-namespace internal {
-
-template <typename A>
-NGramFstImpl<A>::NGramFstImpl(const Fst<A> &fst,
-                              std::vector<StateId> *order_out) {
-  typedef A Arc;
-  typedef typename Arc::Label Label;
-  typedef typename Arc::Weight Weight;
-  typedef typename Arc::StateId StateId;
-  SetType("ngram");
-  SetInputSymbols(fst.InputSymbols());
-  SetOutputSymbols(fst.OutputSymbols());
-  SetProperties(kStaticProperties);
-
-  // Check basic requirements for an OpenGrm language model Fst.
-  if (!NGramFst<A>::HasRequiredProps(fst)) {
-    FSTERROR() << "NGramFst only accepts OpenGrm language models as input";
-    SetProperties(kError, kError);
-    return;
-  }
-
-  int64 num_states = CountStates(fst);
-  Label *context = new Label[num_states];
-
-  // Find the unigram state by starting from the start state, following
-  // epsilons.
-  StateId unigram = fst.Start();
-  while (1) {
-    if (unigram == kNoStateId) {
-      FSTERROR() << "Could not identify unigram state";
-      SetProperties(kError, kError);
-      return;
-    }
-    ArcIterator<Fst<A>> aiter(fst, unigram);
-    if (aiter.Done()) {
-      LOG(WARNING) << "Unigram state " << unigram << " has no arcs.";
-      break;
-    }
-    if (aiter.Value().ilabel != 0) break;
-    unigram = aiter.Value().nextstate;
-  }
-
-  // Each state's context is determined by the subtree it is under from the
-  // unigram state.
-  std::queue<std::pair<StateId, Label>> label_queue;
-  std::vector<bool> visited(num_states);
-  // Force an epsilon link to the start state.
-  label_queue.push(std::make_pair(fst.Start(), 0));
-  for (ArcIterator<Fst<A>> aiter(fst, unigram); !aiter.Done(); aiter.Next()) {
-    label_queue.push(
-        std::make_pair(aiter.Value().nextstate, aiter.Value().ilabel));
-  }
-  // investigate states in breadth first fashion to assign context words.
-  while (!label_queue.empty()) {
-    std::pair<StateId, Label> &now = label_queue.front();
-    if (!visited[now.first]) {
-      context[now.first] = now.second;
-      visited[now.first] = true;
-      for (ArcIterator<Fst<A>> aiter(fst, now.first); !aiter.Done();
-           aiter.Next()) {
-        const Arc &arc = aiter.Value();
-        if (arc.ilabel != 0) {
-          label_queue.push(std::make_pair(arc.nextstate, now.second));
-        }
-      }
-    }
-    label_queue.pop();
-  }
-  visited.clear();
-
-  // The arc from the start state should be assigned an epsilon to put it
-  // in front of the all other labels (which makes Start state 1 after
-  // unigram which is state 0).
-  context[fst.Start()] = 0;
-
-  // Build the tree of contexts fst by reversing the epsilon arcs from fst.
-  VectorFst<Arc> context_fst;
-  uint64 num_final = 0;
-  for (int i = 0; i < num_states; ++i) {
-    if (fst.Final(i) != Weight::Zero()) {
-      ++num_final;
-    }
-    context_fst.SetFinal(context_fst.AddState(), fst.Final(i));
-  }
-  context_fst.SetStart(unigram);
-  context_fst.SetInputSymbols(fst.InputSymbols());
-  context_fst.SetOutputSymbols(fst.OutputSymbols());
-  int64 num_context_arcs = 0;
-  int64 num_futures = 0;
-  for (StateIterator<Fst<A>> siter(fst); !siter.Done(); siter.Next()) {
-    const StateId &state = siter.Value();
-    num_futures += fst.NumArcs(state) - fst.NumInputEpsilons(state);
-    ArcIterator<Fst<A>> aiter(fst, state);
-    if (!aiter.Done()) {
-      const Arc &arc = aiter.Value();
-      // this arc goes from state to arc.nextstate, so create an arc from
-      // arc.nextstate to state to reverse it.
-      if (arc.ilabel == 0) {
-        context_fst.AddArc(arc.nextstate, Arc(context[state], context[state],
-                                              arc.weight, state));
-        num_context_arcs++;
-      }
-    }
-  }
-  if (num_context_arcs != context_fst.NumStates() - 1) {
-    FSTERROR() << "Number of contexts arcs != number of states - 1";
-    SetProperties(kError, kError);
-    return;
-  }
-  if (context_fst.NumStates() != num_states) {
-    FSTERROR() << "Number of contexts != number of states";
-    SetProperties(kError, kError);
-    return;
-  }
-  int64 context_props =
-      context_fst.Properties(kIDeterministic | kILabelSorted, true);
-  if (!(context_props & kIDeterministic)) {
-    FSTERROR() << "Input Fst is not structured properly";
-    SetProperties(kError, kError);
-    return;
-  }
-  if (!(context_props & kILabelSorted)) {
-    ArcSort(&context_fst, ILabelCompare<Arc>());
-  }
-
-  delete[] context;
-
-  uint64 b64;
-  Weight weight;
-  Label label = kNoLabel;
-  const size_t storage = Storage(num_states, num_futures, num_final);
-  MappedFile *data_region = MappedFile::Allocate(storage);
-  char *data = reinterpret_cast<char *>(data_region->mutable_data());
-  memset(data, 0, storage);
-  size_t offset = 0;
-  memcpy(data + offset, reinterpret_cast<char *>(&num_states),
-         sizeof(num_states));
-  offset += sizeof(num_states);
-  memcpy(data + offset, reinterpret_cast<char *>(&num_futures),
-         sizeof(num_futures));
-  offset += sizeof(num_futures);
-  memcpy(data + offset, reinterpret_cast<char *>(&num_final),
-         sizeof(num_final));
-  offset += sizeof(num_final);
-  uint64 *context_bits = reinterpret_cast<uint64 *>(data + offset);
-  offset += BitmapIndex::StorageSize(num_states * 2 + 1) * sizeof(b64);
-  uint64 *future_bits = reinterpret_cast<uint64 *>(data + offset);
-  offset +=
-      BitmapIndex::StorageSize(num_futures + num_states + 1) * sizeof(b64);
-  uint64 *final_bits = reinterpret_cast<uint64 *>(data + offset);
-  offset += BitmapIndex::StorageSize(num_states) * sizeof(b64);
-  Label *context_words = reinterpret_cast<Label *>(data + offset);
-  offset += (num_states + 1) * sizeof(label);
-  Label *future_words = reinterpret_cast<Label *>(data + offset);
-  offset += num_futures * sizeof(label);
-  offset = (offset + sizeof(weight) - 1) & ~(sizeof(weight) - 1);
-  Weight *backoff = reinterpret_cast<Weight *>(data + offset);
-  offset += (num_states + 1) * sizeof(weight);
-  Weight *final_probs = reinterpret_cast<Weight *>(data + offset);
-  offset += num_final * sizeof(weight);
-  Weight *future_probs = reinterpret_cast<Weight *>(data + offset);
-  int64 context_arc = 0, future_arc = 0, context_bit = 0, future_bit = 0,
-        final_bit = 0;
-
-  // pseudo-root bits
-  BitmapIndex::Set(context_bits, context_bit++);
-  ++context_bit;
-  context_words[context_arc] = label;
-  backoff[context_arc] = Weight::Zero();
-  context_arc++;
-
-  ++future_bit;
-  if (order_out) {
-    order_out->clear();
-    order_out->resize(num_states);
-  }
-
-  std::queue<StateId> context_q;
-  context_q.push(context_fst.Start());
-  StateId state_number = 0;
-  while (!context_q.empty()) {
-    const StateId &state = context_q.front();
-    if (order_out) {
-      (*order_out)[state] = state_number;
-    }
-
-    const Weight final_weight = context_fst.Final(state);
-    if (final_weight != Weight::Zero()) {
-      BitmapIndex::Set(final_bits, state_number);
-      final_probs[final_bit] = final_weight;
-      ++final_bit;
-    }
-
-    for (ArcIterator<VectorFst<A>> aiter(context_fst, state); !aiter.Done();
-         aiter.Next()) {
-      const Arc &arc = aiter.Value();
-      context_words[context_arc] = arc.ilabel;
-      backoff[context_arc] = arc.weight;
-      ++context_arc;
-      BitmapIndex::Set(context_bits, context_bit++);
-      context_q.push(arc.nextstate);
-    }
-    ++context_bit;
-
-    for (ArcIterator<Fst<A>> aiter(fst, state); !aiter.Done(); aiter.Next()) {
-      const Arc &arc = aiter.Value();
-      if (arc.ilabel != 0) {
-        future_words[future_arc] = arc.ilabel;
-        future_probs[future_arc] = arc.weight;
-        ++future_arc;
-        BitmapIndex::Set(future_bits, future_bit++);
-      }
-    }
-    ++future_bit;
-    ++state_number;
-    context_q.pop();
-  }
-
-  if ((state_number != num_states) || (context_bit != num_states * 2 + 1) ||
-      (context_arc != num_states) || (future_arc != num_futures) ||
-      (future_bit != num_futures + num_states + 1) ||
-      (final_bit != num_final)) {
-    FSTERROR() << "Structure problems detected during construction";
-    SetProperties(kError, kError);
-    return;
-  }
-
-  Init(data, false, data_region);
-}
-
-template <typename A>
-inline void NGramFstImpl<A>::Init(const char *data, bool owned,
-                                  MappedFile *data_region) {
-  if (owned_) {
-    delete[] data_;
-  }
-  data_region_.reset(data_region);
-  owned_ = owned;
-  data_ = data;
-  size_t offset = 0;
-  num_states_ = *(reinterpret_cast<const uint64 *>(data_ + offset));
-  offset += sizeof(num_states_);
-  num_futures_ = *(reinterpret_cast<const uint64 *>(data_ + offset));
-  offset += sizeof(num_futures_);
-  num_final_ = *(reinterpret_cast<const uint64 *>(data_ + offset));
-  offset += sizeof(num_final_);
-  uint64 bits;
-  size_t context_bits = num_states_ * 2 + 1;
-  size_t future_bits = num_futures_ + num_states_ + 1;
-  context_ = reinterpret_cast<const uint64 *>(data_ + offset);
-  offset += BitmapIndex::StorageSize(context_bits) * sizeof(bits);
-  future_ = reinterpret_cast<const uint64 *>(data_ + offset);
-  offset += BitmapIndex::StorageSize(future_bits) * sizeof(bits);
-  final_ = reinterpret_cast<const uint64 *>(data_ + offset);
-  offset += BitmapIndex::StorageSize(num_states_) * sizeof(bits);
-  context_words_ = reinterpret_cast<const Label *>(data_ + offset);
-  offset += (num_states_ + 1) * sizeof(*context_words_);
-  future_words_ = reinterpret_cast<const Label *>(data_ + offset);
-  offset += num_futures_ * sizeof(*future_words_);
-  offset = (offset + sizeof(*backoff_) - 1) & ~(sizeof(*backoff_) - 1);
-  backoff_ = reinterpret_cast<const Weight *>(data_ + offset);
-  offset += (num_states_ + 1) * sizeof(*backoff_);
-  final_probs_ = reinterpret_cast<const Weight *>(data_ + offset);
-  offset += num_final_ * sizeof(*final_probs_);
-  future_probs_ = reinterpret_cast<const Weight *>(data_ + offset);
-
-  context_index_.BuildIndex(context_, context_bits);
-  future_index_.BuildIndex(future_, future_bits);
-  final_index_.BuildIndex(final_, num_states_);
-
-  select_root_ = context_index_.Select0s(0);
-  if (context_index_.Rank1(0) != 0 || select_root_.first != 1 ||
-      context_index_.Get(2) == false) {
-    FSTERROR() << "Malformed file";
-    SetProperties(kError, kError);
-    return;
-  }
-  root_children_ = context_words_ + context_index_.Rank1(2);
-  start_ = 1;
-}
-
-template <typename A>
-inline typename A::StateId NGramFstImpl<A>::Transition(
-    const std::vector<Label> &context, Label future) const {
-  const Label *children = root_children_;
-  size_t num_children = select_root_.second - 2;
-  const Label *loc =
-      std::lower_bound(children, children + num_children, future);
-  if (loc == children + num_children || *loc != future) {
-    return context_index_.Rank1(0);
-  }
-  size_t node = 2 + loc - children;
-  size_t node_rank = context_index_.Rank1(node);
-  std::pair<size_t, size_t> zeros =
-      (node_rank == 0) ? select_root_ : context_index_.Select0s(node_rank);
-  size_t first_child = zeros.first + 1;
-  if (context_index_.Get(first_child) == false) {
-    return context_index_.Rank1(node);
-  }
-  size_t last_child = zeros.second - 1;
-  for (int word = context.size() - 1; word >= 0; --word) {
-    children = context_words_ + context_index_.Rank1(first_child);
-    loc = std::lower_bound(children, children + last_child - first_child + 1,
-                           context[word]);
-    if (loc == children + last_child - first_child + 1 ||
-        *loc != context[word]) {
-      break;
-    }
-    node = first_child + loc - children;
-    node_rank = context_index_.Rank1(node);
-    zeros =
-        (node_rank == 0) ? select_root_ : context_index_.Select0s(node_rank);
-    first_child = zeros.first + 1;
-    if (context_index_.Get(first_child) == false) break;
-    last_child = zeros.second - 1;
-  }
-  return context_index_.Rank1(node);
-}
-
-}  // namespace internal
-
-/*****************************************************************************/
-template <class A>
-class NGramFstMatcher : public MatcherBase<A> {
- public:
-  typedef A Arc;
-  typedef typename A::Label Label;
-  typedef typename A::StateId StateId;
-  typedef typename A::Weight Weight;
-
-  // This makes a copy of the FST.
-  NGramFstMatcher(const NGramFst<A> &fst, MatchType match_type)
-      : owned_fst_(fst.Copy()),
-        fst_(*owned_fst_),
-        inst_(fst_.inst_),
-        match_type_(match_type),
-        current_loop_(false),
-        loop_(kNoLabel, 0, A::Weight::One(), kNoStateId) {
-    if (match_type_ == MATCH_OUTPUT) {
-      std::swap(loop_.ilabel, loop_.olabel);
-    }
-  }
-
-  // This doesn't copy the FST.
-  NGramFstMatcher(const NGramFst<A> *fst, MatchType match_type)
-      : fst_(*fst),
-        inst_(fst_.inst_),
-        match_type_(match_type),
-        current_loop_(false),
-        loop_(kNoLabel, 0, A::Weight::One(), kNoStateId) {
-    if (match_type_ == MATCH_OUTPUT) {
-      std::swap(loop_.ilabel, loop_.olabel);
-    }
-  }
-
-  // This makes a copy of the FST.
-  NGramFstMatcher(const NGramFstMatcher<A> &matcher, bool safe = false)
-      : owned_fst_(matcher.fst_.Copy(safe)),
-        fst_(*owned_fst_),
-        inst_(matcher.inst_),
-        match_type_(matcher.match_type_),
-        current_loop_(false),
-        loop_(kNoLabel, 0, A::Weight::One(), kNoStateId) {
-    if (match_type_ == MATCH_OUTPUT) {
-      std::swap(loop_.ilabel, loop_.olabel);
-    }
-  }
-
-  NGramFstMatcher<A> *Copy(bool safe = false) const override {
-    return new NGramFstMatcher<A>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override { return match_type_; }
-
-  const Fst<A> &GetFst() const override { return fst_; }
-
-  uint64 Properties(uint64 props) const override { return props; }
-
-  void SetState(StateId s) final {
-    fst_.GetImpl()->SetInstFuture(s, &inst_);
-    current_loop_ = false;
-  }
-
-  bool Find(Label label) final {
-    const Label nolabel = kNoLabel;
-    done_ = true;
-    if (label == 0 || label == nolabel) {
-      if (label == 0) {
-        current_loop_ = true;
-        loop_.nextstate = inst_.state_;
-      }
-      // The unigram state has no epsilon arc.
-      if (inst_.state_ != 0) {
-        arc_.ilabel = arc_.olabel = 0;
-        fst_.GetImpl()->SetInstNode(&inst_);
-        arc_.nextstate = fst_.GetImpl()->context_index_.Rank1(
-            fst_.GetImpl()->context_index_.Select1(
-                fst_.GetImpl()->context_index_.Rank0(inst_.node_) - 1));
-        arc_.weight = fst_.GetImpl()->backoff_[inst_.state_];
-        done_ = false;
-      }
-    } else {
-      current_loop_ = false;
-      const Label *start = fst_.GetImpl()->future_words_ + inst_.offset_;
-      const Label *end = start + inst_.num_futures_;
-      const Label *search = std::lower_bound(start, end, label);
-      if (search != end && *search == label) {
-        size_t state = search - start;
-        arc_.ilabel = arc_.olabel = label;
-        arc_.weight = fst_.GetImpl()->future_probs_[inst_.offset_ + state];
-        fst_.GetImpl()->SetInstContext(&inst_);
-        arc_.nextstate = fst_.GetImpl()->Transition(inst_.context_, label);
-        done_ = false;
-      }
-    }
-    return !Done();
-  }
-
-  bool Done() const final { return !current_loop_ && done_; }
-
-  const Arc &Value() const final { return (current_loop_) ? loop_ : arc_; }
-
-  void Next() final {
-    if (current_loop_) {
-      current_loop_ = false;
-    } else {
-      done_ = true;
-    }
-  }
-
-  ssize_t Priority(StateId s) final { return fst_.NumArcs(s); }
-
- private:
-  std::unique_ptr<NGramFst<A>> owned_fst_;
-  const NGramFst<A> &fst_;
-  NGramFstInst<A> inst_;
-  MatchType match_type_;  // Supplied by caller
-  bool done_;
-  Arc arc_;
-  bool current_loop_;  // Current arc is the implicit loop
-  Arc loop_;
-};
-
-/*****************************************************************************/
-// Specialization for NGramFst; see generic version in fst.h
-// for sample usage (but use the ProdLmFst type!). This version
-// should inline.
-template <class A>
-class StateIterator<NGramFst<A>> : public StateIteratorBase<A> {
- public:
-  typedef typename A::StateId StateId;
-
-  explicit StateIterator(const NGramFst<A> &fst)
-      : s_(0), num_states_(fst.NumStates()) {}
-
-  bool Done() const final { return s_ >= num_states_; }
-
-  StateId Value() const final { return s_; }
-
-  void Next() final { ++s_; }
-
-  void Reset() final { s_ = 0; }
-
- private:
-  StateId s_;
-  StateId num_states_;
-};
-
-/*****************************************************************************/
-template <class A>
-class ArcIterator<NGramFst<A>> : public ArcIteratorBase<A> {
- public:
-  typedef A Arc;
-  typedef typename A::Label Label;
-  typedef typename A::StateId StateId;
-  typedef typename A::Weight Weight;
-
-  ArcIterator(const NGramFst<A> &fst, StateId state)
-      : lazy_(~0), impl_(fst.GetImpl()), i_(0), flags_(kArcValueFlags) {
-    inst_ = fst.inst_;
-    impl_->SetInstFuture(state, &inst_);
-    impl_->SetInstNode(&inst_);
-  }
-
-  bool Done() const final {
-    return i_ >=
-           ((inst_.node_ == 0) ? inst_.num_futures_ : inst_.num_futures_ + 1);
-  }
-
-  const Arc &Value() const final {
-    bool eps = (inst_.node_ != 0 && i_ == 0);
-    StateId state = (inst_.node_ == 0) ? i_ : i_ - 1;
-    if (flags_ & lazy_ & (kArcILabelValue | kArcOLabelValue)) {
-      arc_.ilabel = arc_.olabel =
-          eps ? 0 : impl_->future_words_[inst_.offset_ + state];
-      lazy_ &= ~(kArcILabelValue | kArcOLabelValue);
-    }
-    if (flags_ & lazy_ & kArcNextStateValue) {
-      if (eps) {
-        arc_.nextstate =
-            impl_->context_index_.Rank1(impl_->context_index_.Select1(
-                impl_->context_index_.Rank0(inst_.node_) - 1));
-      } else {
-        if (lazy_ & kArcNextStateValue) {
-          impl_->SetInstContext(&inst_);  // first time only.
-        }
-        arc_.nextstate = impl_->Transition(
-            inst_.context_, impl_->future_words_[inst_.offset_ + state]);
-      }
-      lazy_ &= ~kArcNextStateValue;
-    }
-    if (flags_ & lazy_ & kArcWeightValue) {
-      arc_.weight = eps ? impl_->backoff_[inst_.state_]
-                        : impl_->future_probs_[inst_.offset_ + state];
-      lazy_ &= ~kArcWeightValue;
-    }
-    return arc_;
-  }
-
-  void Next() final {
-    ++i_;
-    lazy_ = ~0;
-  }
-
-  size_t Position() const final { return i_; }
-
-  void Reset() final {
-    i_ = 0;
-    lazy_ = ~0;
-  }
-
-  void Seek(size_t a) final {
-    if (i_ != a) {
-      i_ = a;
-      lazy_ = ~0;
-    }
-  }
-
-  uint32 Flags() const final { return flags_; }
-
-  void SetFlags(uint32 flags, uint32 mask) final {
-    flags_ &= ~mask;
-    flags_ |= (flags & kArcValueFlags);
-  }
-
- private:
-  mutable Arc arc_;
-  mutable uint32 lazy_;
-  const internal::NGramFstImpl<A> *impl_;  // Borrowed reference.
-  mutable NGramFstInst<A> inst_;
-
-  size_t i_;
-  uint32 flags_;
-};
-
-}  // namespace fst
-#endif  // FST_EXTENSIONS_NGRAM_NGRAM_FST_H_
diff --git a/projects/llm_framework/include/fst/extensions/ngram/nthbit.h b/projects/llm_framework/include/fst/extensions/ngram/nthbit.h
deleted file mode 100644
index 1e6ec635..00000000
--- a/projects/llm_framework/include/fst/extensions/ngram/nthbit.h
+++ /dev/null
@@ -1,49 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_NGRAM_NTHBIT_H_
-#define FST_EXTENSIONS_NGRAM_NTHBIT_H_
-
-#include <fst/types.h>
-#include <fst/compat.h>
-
-#ifdef __BMI2__
-// PDEP requires BMI2.
-
-// Returns the position (0-63) of the r-th 1 bit in v.
-// 1 <= r <= CountOnes(v) <= 64.  Therefore, v must not be 0.
-inline uint32 nth_bit(uint64 v, uint32 r) {
-  // PDEP example from https://stackoverflow.com/a/27453505
-  return __builtin_ctzll(_pdep_u64(uint64{1} << (r - 1), v));
-}
-
-#else  // !defined(__BMI2__)
-
-extern const uint32 nth_bit_bit_offset[];
-
-// Returns the position (0-63) of the r-th 1 bit in v.
-// 1 <= r <= CountOnes(v) <= 64.  Therefore, v must not be 0.
-inline uint32 nth_bit(uint64 v, uint32 r) {
-  uint32 shift = 0;
-  uint32 c = __builtin_popcount(v & 0xffffffff);
-  uint32 mask = -(r > c);
-  r -= c & mask;
-  shift += (32 & mask);
-
-  c = __builtin_popcount((v >> shift) & 0xffff);
-  mask = -(r > c);
-  r -= c & mask;
-  shift += (16 & mask);
-
-  c = __builtin_popcount((v >> shift) & 0xff);
-  mask = -(r > c);
-  r -= c & mask;
-  shift += (8 & mask);
-
-  return shift +
-         ((nth_bit_bit_offset[(v >> shift) & 0xff] >> ((r - 1) << 2)) & 0xf);
-}
-
-#endif  // !defined(__BMI2__)
-
-#endif  // FST_EXTENSIONS_NGRAM_NTHBIT_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/collection.h b/projects/llm_framework/include/fst/extensions/pdt/collection.h
deleted file mode 100644
index ae34aba7..00000000
--- a/projects/llm_framework/include/fst/extensions/pdt/collection.h
+++ /dev/null
@@ -1,107 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to store a collection of ordered (multi-)sets with elements of type T.
-
-#ifndef FST_EXTENSIONS_PDT_COLLECTION_H_
-#define FST_EXTENSIONS_PDT_COLLECTION_H_
-
-#include <functional>
-#include <vector>
-
-#include <fst/log.h>
-#include <fst/bi-table.h>
-
-namespace fst {
-
-// Stores a collection of non-empty, ordered (multi-)sets with elements of type
-// T. A default constructor, operator==, and an STL-style hash functor must be
-// defined on the elements. Provides signed integer ID (of type I) for each
-// unique set. The IDs are allocated starting from 0 in order.
-template <class I, class T>
-class Collection {
- public:
-  struct Node {  // Trie node.
-    I node_id;   // Root is kNoNodeId;
-    T element;
-
-    Node() : node_id(kNoNodeId), element(T()) {}
-
-    Node(I i, const T &t) : node_id(i), element(t) {}
-
-    bool operator==(const Node &n) const {
-      return n.node_id == node_id && n.element == element;
-    }
-  };
-
-  struct NodeHash {
-    size_t operator()(const Node &n) const {
-      static constexpr auto kPrime = 7853;
-      return n.node_id + hash_(n.element) * kPrime;
-    }
-  };
-
-  using NodeTable = CompactHashBiTable<I, Node, NodeHash>;
-
-  class SetIterator {
-   public:
-    SetIterator(I id, Node node, NodeTable *node_table)
-        : id_(id), node_(node), node_table_(node_table) {}
-
-    bool Done() const { return id_ == kNoNodeId; }
-
-    const T &Element() const { return node_.element; }
-
-    void Next() {
-      id_ = node_.node_id;
-      if (id_ != kNoNodeId) node_ = node_table_->FindEntry(id_);
-    }
-
-   private:
-    I id_;       // Iterator set node ID.
-    Node node_;  // Iterator set node.
-    NodeTable *node_table_;
-  };
-
-  Collection() {}
-
-  // Looks up integer ID from ordered multi-se, and if it doesn't exist and
-  // insert is true, then adds it. Otherwise returns -1.
-  I FindId(const std::vector<T> &set, bool insert = true) {
-    I node_id = kNoNodeId;
-    for (ssize_t i = set.size() - 1; i >= 0; --i) {
-      Node node(node_id, set[i]);
-      node_id = node_table_.FindId(node, insert);
-      if (node_id == -1) break;
-    }
-    return node_id;
-  }
-
-  // Finds ordered (multi-)set given integer ID. Returns set iterator to
-  // traverse result.
-  SetIterator FindSet(I id) {
-    if (id < 0 || id >= node_table_.Size()) {
-      return SetIterator(kNoNodeId, Node(kNoNodeId, T()), &node_table_);
-    } else {
-      return SetIterator(id, node_table_.FindEntry(id), &node_table_);
-    }
-  }
-
-  I Size() const { return node_table_.Size(); }
-
- private:
-  static constexpr I kNoNodeId = -1;
-  static const std::hash<T> hash_;
-
-  NodeTable node_table_;
-};
-
-template <class I, class T>
-constexpr I Collection<I, T>::kNoNodeId;
-
-template <class I, class T>
-const std::hash<T> Collection<I, T>::hash_ = {};
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_COLLECTION_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/compose.h b/projects/llm_framework/include/fst/extensions/pdt/compose.h
deleted file mode 100644
index 525d613a..00000000
--- a/projects/llm_framework/include/fst/extensions/pdt/compose.h
+++ /dev/null
@@ -1,493 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Composes a PDT and an FST.
-
-#ifndef FST_EXTENSIONS_PDT_COMPOSE_H_
-#define FST_EXTENSIONS_PDT_COMPOSE_H_
-
-#include <list>
-
-#include <fst/extensions/pdt/pdt.h>
-#include <fst/compose.h>
-
-namespace fst {
-
-// Returns paren arcs for Find(kNoLabel).
-constexpr uint32 kParenList = 0x00000001;
-
-// Returns a kNolabel loop for Find(paren).
-constexpr uint32 kParenLoop = 0x00000002;
-
-// This class is a matcher that treats parens as multi-epsilon labels.
-// It is most efficient if the parens are in a range non-overlapping with
-// the non-paren labels.
-template <class F>
-class ParenMatcher {
- public:
-  using FST = F;
-  using M = SortedMatcher<FST>;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // This makes a copy of the FST.
-  ParenMatcher(const FST &fst, MatchType match_type,
-               uint32 flags = (kParenLoop | kParenList))
-      : matcher_(fst, match_type), match_type_(match_type), flags_(flags) {
-    if (match_type == MATCH_INPUT) {
-      loop_.ilabel = kNoLabel;
-      loop_.olabel = 0;
-    } else {
-      loop_.ilabel = 0;
-      loop_.olabel = kNoLabel;
-    }
-    loop_.weight = Weight::One();
-    loop_.nextstate = kNoStateId;
-  }
-
-  // This doesn't copy the FST.
-  ParenMatcher(const FST *fst, MatchType match_type,
-               uint32 flags = (kParenLoop | kParenList))
-      : matcher_(fst, match_type), match_type_(match_type), flags_(flags) {
-    if (match_type == MATCH_INPUT) {
-      loop_.ilabel = kNoLabel;
-      loop_.olabel = 0;
-    } else {
-      loop_.ilabel = 0;
-      loop_.olabel = kNoLabel;
-    }
-    loop_.weight = Weight::One();
-    loop_.nextstate = kNoStateId;
-  }
-
-  // This makes a copy of the FST.
-  ParenMatcher(const ParenMatcher<FST> &matcher, bool safe = false)
-      : matcher_(matcher.matcher_, safe),
-        match_type_(matcher.match_type_),
-        flags_(matcher.flags_),
-        open_parens_(matcher.open_parens_),
-        close_parens_(matcher.close_parens_),
-        loop_(matcher.loop_) {
-    loop_.nextstate = kNoStateId;
-  }
-
-  ParenMatcher<FST> *Copy(bool safe = false) const {
-    return new ParenMatcher<FST>(*this, safe);
-  }
-
-  MatchType Type(bool test) const { return matcher_.Type(test); }
-
-  void SetState(StateId s) {
-    matcher_.SetState(s);
-    loop_.nextstate = s;
-  }
-
-  bool Find(Label match_label);
-
-  bool Done() const { return done_; }
-
-  const Arc &Value() const { return paren_loop_ ? loop_ : matcher_.Value(); }
-
-  void Next();
-
-  Weight Final(StateId s) { return matcher_.Final(s); }
-
-  ssize_t Priority(StateId s) { return matcher_.Priority(s); }
-
-  const FST &GetFst() const { return matcher_.GetFst(); }
-
-  uint64 Properties(uint64 props) const { return matcher_.Properties(props); }
-
-  uint32 Flags() const { return matcher_.Flags(); }
-
-  void AddOpenParen(Label label) {
-    if (label == 0) {
-      FSTERROR() << "ParenMatcher: Bad open paren label: 0";
-    } else {
-      open_parens_.Insert(label);
-    }
-  }
-
-  void AddCloseParen(Label label) {
-    if (label == 0) {
-      FSTERROR() << "ParenMatcher: Bad close paren label: 0";
-    } else {
-      close_parens_.Insert(label);
-    }
-  }
-
-  void RemoveOpenParen(Label label) {
-    if (label == 0) {
-      FSTERROR() << "ParenMatcher: Bad open paren label: 0";
-    } else {
-      open_parens_.Erase(label);
-    }
-  }
-
-  void RemoveCloseParen(Label label) {
-    if (label == 0) {
-      FSTERROR() << "ParenMatcher: Bad close paren label: 0";
-    } else {
-      close_parens_.Erase(label);
-    }
-  }
-
-  void ClearOpenParens() { open_parens_.Clear(); }
-
-  void ClearCloseParens() { close_parens_.Clear(); }
-
-  bool IsOpenParen(Label label) const { return open_parens_.Member(label); }
-
-  bool IsCloseParen(Label label) const { return close_parens_.Member(label); }
-
- private:
-  // Advances matcher to next open paren, returning true if it exists.
-  bool NextOpenParen();
-
-  // Advances matcher to next close paren, returning true if it exists.
-  bool NextCloseParen();
-
-  M matcher_;
-  MatchType match_type_;  // Type of match to perform.
-  uint32 flags_;
-  // Open paren label set.
-  CompactSet<Label, kNoLabel> open_parens_;
-  // Close paren label set.
-  CompactSet<Label, kNoLabel> close_parens_;
-  bool open_paren_list_;   // Matching open paren list?
-  bool close_paren_list_;  // Matching close paren list?
-  bool paren_loop_;        // Current arc is the implicit paren loop?
-  mutable Arc loop_;       // For non-consuming symbols.
-  bool done_;              // Matching done?
-
-  ParenMatcher &operator=(const ParenMatcher &) = delete;
-};
-
-template <class FST>
-inline bool ParenMatcher<FST>::Find(Label match_label) {
-  open_paren_list_ = false;
-  close_paren_list_ = false;
-  paren_loop_ = false;
-  done_ = false;
-  // Returns all parenthesis arcs.
-  if (match_label == kNoLabel && (flags_ & kParenList)) {
-    if (open_parens_.LowerBound() != kNoLabel) {
-      matcher_.LowerBound(open_parens_.LowerBound());
-      open_paren_list_ = NextOpenParen();
-      if (open_paren_list_) return true;
-    }
-    if (close_parens_.LowerBound() != kNoLabel) {
-      matcher_.LowerBound(close_parens_.LowerBound());
-      close_paren_list_ = NextCloseParen();
-      if (close_paren_list_) return true;
-    }
-  }
-  // Returns the implicit paren loop.
-  if (match_label > 0 && (flags_ & kParenLoop) &&
-      (IsOpenParen(match_label) || IsCloseParen(match_label))) {
-    paren_loop_ = true;
-    return true;
-  }
-  // Returns all other labels.
-  if (matcher_.Find(match_label)) return true;
-  done_ = true;
-  return false;
-}
-
-template <class FST>
-inline void ParenMatcher<FST>::Next() {
-  if (paren_loop_) {
-    paren_loop_ = false;
-    done_ = true;
-  } else if (open_paren_list_) {
-    matcher_.Next();
-    open_paren_list_ = NextOpenParen();
-    if (open_paren_list_) return;
-    if (close_parens_.LowerBound() != kNoLabel) {
-      matcher_.LowerBound(close_parens_.LowerBound());
-      close_paren_list_ = NextCloseParen();
-      if (close_paren_list_) return;
-    }
-    done_ = !matcher_.Find(kNoLabel);
-  } else if (close_paren_list_) {
-    matcher_.Next();
-    close_paren_list_ = NextCloseParen();
-    if (close_paren_list_) return;
-    done_ = !matcher_.Find(kNoLabel);
-  } else {
-    matcher_.Next();
-    done_ = matcher_.Done();
-  }
-}
-
-// Advances matcher to next open paren, returning true if it exists.
-template <class FST>
-inline bool ParenMatcher<FST>::NextOpenParen() {
-  for (; !matcher_.Done(); matcher_.Next()) {
-    Label label = match_type_ == MATCH_INPUT ? matcher_.Value().ilabel
-                                             : matcher_.Value().olabel;
-    if (label > open_parens_.UpperBound()) return false;
-    if (IsOpenParen(label)) return true;
-  }
-  return false;
-}
-
-// Advances matcher to next close paren, returning true if it exists.
-template <class FST>
-inline bool ParenMatcher<FST>::NextCloseParen() {
-  for (; !matcher_.Done(); matcher_.Next()) {
-    Label label = match_type_ == MATCH_INPUT ? matcher_.Value().ilabel
-                                             : matcher_.Value().olabel;
-    if (label > close_parens_.UpperBound()) return false;
-    if (IsCloseParen(label)) return true;
-  }
-  return false;
-}
-
-template <class Filter>
-class ParenFilter {
- public:
-  using FST1 = typename Filter::FST1;
-  using FST2 = typename Filter::FST2;
-  using Arc = typename Filter::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Matcher1 = typename Filter::Matcher1;
-  using Matcher2 = typename Filter::Matcher2;
-
-  using StackId = StateId;
-  using ParenStack = PdtStack<StackId, Label>;
-  using FilterState1 = typename Filter::FilterState;
-  using FilterState2 = IntegerFilterState<StackId>;
-  using FilterState = PairFilterState<FilterState1, FilterState2>;
-
-  ParenFilter(const FST1 &fst1, const FST2 &fst2, Matcher1 *matcher1 = nullptr,
-              Matcher2 *matcher2 = nullptr,
-              const std::vector<std::pair<Label, Label>> *parens = nullptr,
-              bool expand = false, bool keep_parens = true)
-      : filter_(fst1, fst2, matcher1, matcher2),
-        parens_(parens ? *parens : std::vector<std::pair<Label, Label>>()),
-        expand_(expand),
-        keep_parens_(keep_parens),
-        fs_(FilterState::NoState()),
-        stack_(parens_),
-        paren_id_(-1) {
-    if (parens) {
-      for (const auto &pair : *parens) {
-        parens_.push_back(pair);
-        GetMatcher1()->AddOpenParen(pair.first);
-        GetMatcher2()->AddOpenParen(pair.first);
-        if (!expand_) {
-          GetMatcher1()->AddCloseParen(pair.second);
-          GetMatcher2()->AddCloseParen(pair.second);
-        }
-      }
-    }
-  }
-
-  ParenFilter(const ParenFilter &filter, bool safe = false)
-      : filter_(filter.filter_, safe),
-        parens_(filter.parens_),
-        expand_(filter.expand_),
-        keep_parens_(filter.keep_parens_),
-        fs_(FilterState::NoState()),
-        stack_(filter.parens_),
-        paren_id_(-1) {}
-
-  FilterState Start() const {
-    return FilterState(filter_.Start(), FilterState2(0));
-  }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    fs_ = fs;
-    filter_.SetState(s1, s2, fs_.GetState1());
-    if (!expand_) return;
-    ssize_t paren_id = stack_.Top(fs.GetState2().GetState());
-    if (paren_id != paren_id_) {
-      if (paren_id_ != -1) {
-        GetMatcher1()->RemoveCloseParen(parens_[paren_id_].second);
-        GetMatcher2()->RemoveCloseParen(parens_[paren_id_].second);
-      }
-      paren_id_ = paren_id;
-      if (paren_id_ != -1) {
-        GetMatcher1()->AddCloseParen(parens_[paren_id_].second);
-        GetMatcher2()->AddCloseParen(parens_[paren_id_].second);
-      }
-    }
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    const auto fs1 = filter_.FilterArc(arc1, arc2);
-    const auto &fs2 = fs_.GetState2();
-    if (fs1 == FilterState1::NoState()) return FilterState::NoState();
-    if (arc1->olabel == kNoLabel && arc2->ilabel) {  // arc2 parentheses.
-      if (keep_parens_) {
-        arc1->ilabel = arc2->ilabel;
-      } else if (arc2->ilabel) {
-        arc2->olabel = arc1->ilabel;
-      }
-      return FilterParen(arc2->ilabel, fs1, fs2);
-    } else if (arc2->ilabel == kNoLabel && arc1->olabel) {  // arc1 parentheses.
-      if (keep_parens_) {
-        arc2->olabel = arc1->olabel;
-      } else {
-        arc1->ilabel = arc2->olabel;
-      }
-      return FilterParen(arc1->olabel, fs1, fs2);
-    } else {
-      return FilterState(fs1, fs2);
-    }
-  }
-
-  void FilterFinal(Weight *w1, Weight *w2) const {
-    if (fs_.GetState2().GetState() != 0) *w1 = Weight::Zero();
-    filter_.FilterFinal(w1, w2);
-  }
-
-  // Returns respective matchers; ownership stays with filter.
-
-  Matcher1 *GetMatcher1() { return filter_.GetMatcher1(); }
-
-  Matcher2 *GetMatcher2() { return filter_.GetMatcher2(); }
-
-  uint64 Properties(uint64 iprops) const {
-    return filter_.Properties(iprops) & kILabelInvariantProperties &
-           kOLabelInvariantProperties;
-  }
-
- private:
-  const FilterState FilterParen(Label label, const FilterState1 &fs1,
-                                const FilterState2 &fs2) const {
-    if (!expand_) return FilterState(fs1, fs2);
-    const auto stack_id = stack_.Find(fs2.GetState(), label);
-    if (stack_id < 0) {
-      return FilterState::NoState();
-    } else {
-      return FilterState(fs1, FilterState2(stack_id));
-    }
-  }
-
-  Filter filter_;
-  std::vector<std::pair<Label, Label>> parens_;
-  bool expand_;       // Expands to FST?
-  bool keep_parens_;  // Retains parentheses in output?
-  FilterState fs_;    // Current filter state.
-  mutable ParenStack stack_;
-  ssize_t paren_id_;
-};
-
-// Class to setup composition options for PDT composition. Default is to take
-// the PDT as the first composition argument.
-template <class Arc, bool left_pdt = true>
-class PdtComposeFstOptions
-    : public ComposeFstOptions<
-          Arc, ParenMatcher<Fst<Arc>>,
-          ParenFilter<AltSequenceComposeFilter<ParenMatcher<Fst<Arc>>>>> {
- public:
-  using Label = typename Arc::Label;
-  using PdtMatcher = ParenMatcher<Fst<Arc>>;
-  using PdtFilter = ParenFilter<AltSequenceComposeFilter<PdtMatcher>>;
-
-  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::matcher1;
-  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::matcher2;
-  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::filter;
-
-  PdtComposeFstOptions(const Fst<Arc> &ifst1,
-                       const std::vector<std::pair<Label, Label>> &parens,
-                       const Fst<Arc> &ifst2, bool expand = false,
-                       bool keep_parens = true) {
-    matcher1 = new PdtMatcher(ifst1, MATCH_OUTPUT, kParenList);
-    matcher2 = new PdtMatcher(ifst2, MATCH_INPUT, kParenLoop);
-    filter = new PdtFilter(ifst1, ifst2, matcher1, matcher2, &parens, expand,
-                           keep_parens);
-  }
-};
-
-// Class to setup composition options for PDT with FST composition.
-// Specialization is for the FST as the first composition argument.
-template <class Arc>
-class PdtComposeFstOptions<Arc, false>
-    : public ComposeFstOptions<
-          Arc, ParenMatcher<Fst<Arc>>,
-          ParenFilter<SequenceComposeFilter<ParenMatcher<Fst<Arc>>>>> {
- public:
-  using Label = typename Arc::Label;
-  using PdtMatcher = ParenMatcher<Fst<Arc>>;
-  using PdtFilter = ParenFilter<SequenceComposeFilter<PdtMatcher>>;
-
-  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::matcher1;
-  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::matcher2;
-  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::filter;
-
-  PdtComposeFstOptions(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-                       const std::vector<std::pair<Label, Label>> &parens,
-                       bool expand = false, bool keep_parens = true) {
-    matcher1 = new PdtMatcher(ifst1, MATCH_OUTPUT, kParenLoop);
-    matcher2 = new PdtMatcher(ifst2, MATCH_INPUT, kParenList);
-    filter = new PdtFilter(ifst1, ifst2, matcher1, matcher2, &parens, expand,
-                           keep_parens);
-  }
-};
-
-enum PdtComposeFilter {
-  PAREN_FILTER,         // Bar-Hillel construction; keeps parentheses.
-  EXPAND_FILTER,        // Bar-Hillel + expansion; removes parentheses.
-  EXPAND_PAREN_FILTER,  // Bar-Hillel + expansion; keeps parentheses.
-};
-
-struct PdtComposeOptions {
-  bool connect;                  // Connect output?
-  PdtComposeFilter filter_type;  // Pre-defined filter to use.
-
-  explicit PdtComposeOptions(bool connect = true,
-                             PdtComposeFilter filter_type = PAREN_FILTER)
-      : connect(connect), filter_type(filter_type) {}
-};
-
-// Composes pushdown transducer (PDT) encoded as an FST (1st arg) and an FST
-// (2nd arg) with the result also a PDT encoded as an FST (3rd arg). In the
-// PDTs, some transitions are labeled with open or close parentheses. To be
-// interpreted as a PDT, the parens must balance on a path (see PdtExpand()).
-// The open-close parenthesis label pairs are passed using the parens argument.
-template <class Arc>
-void Compose(const Fst<Arc> &ifst1,
-             const std::vector<
-                 std::pair<typename Arc::Label, typename Arc::Label>> &parens,
-             const Fst<Arc> &ifst2, MutableFst<Arc> *ofst,
-             const PdtComposeOptions &opts = PdtComposeOptions()) {
-  bool expand = opts.filter_type != PAREN_FILTER;
-  bool keep_parens = opts.filter_type != EXPAND_FILTER;
-  PdtComposeFstOptions<Arc, true> copts(ifst1, parens, ifst2, expand,
-                                        keep_parens);
-  copts.gc_limit = 0;
-  *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-  if (opts.connect) Connect(ofst);
-}
-
-// Composes an FST (1st arg) and pushdown transducer (PDT) encoded as an FST
-// (2nd arg) with the result also a PDT encoded as an FST (3rd arg). In the
-// PDTs, some transitions are labeled with open or close parentheses. To be
-// interpreted as a PDT, the parens must balance on a path (see ExpandFst()).
-// The open-close parenthesis label pairs are passed using the parens argument.
-template <class Arc>
-void Compose(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-             const std::vector<
-                 std::pair<typename Arc::Label, typename Arc::Label>> &parens,
-             MutableFst<Arc> *ofst,
-             const PdtComposeOptions &opts = PdtComposeOptions()) {
-  bool expand = opts.filter_type != PAREN_FILTER;
-  bool keep_parens = opts.filter_type != EXPAND_FILTER;
-  PdtComposeFstOptions<Arc, false> copts(ifst1, ifst2, parens, expand,
-                                         keep_parens);
-  copts.gc_limit = 0;
-  *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-  if (opts.connect) Connect(ofst);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_COMPOSE_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/expand.h b/projects/llm_framework/include/fst/extensions/pdt/expand.h
deleted file mode 100644
index eeee781e..00000000
--- a/projects/llm_framework/include/fst/extensions/pdt/expand.h
+++ /dev/null
@@ -1,933 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Expands a PDT to an FST.
-
-#ifndef FST_EXTENSIONS_PDT_EXPAND_H_
-#define FST_EXTENSIONS_PDT_EXPAND_H_
-
-#include <forward_list>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/extensions/pdt/paren.h>
-#include <fst/extensions/pdt/pdt.h>
-#include <fst/extensions/pdt/reverse.h>
-#include <fst/extensions/pdt/shortest-path.h>
-#include <fst/cache.h>
-#include <fst/mutable-fst.h>
-#include <fst/queue.h>
-#include <fst/state-table.h>
-#include <fst/test-properties.h>
-
-namespace fst {
-
-template <class Arc>
-struct PdtExpandFstOptions : public CacheOptions {
-  bool keep_parentheses;
-  PdtStack<typename Arc::StateId, typename Arc::Label> *stack;
-  PdtStateTable<typename Arc::StateId, typename Arc::StateId> *state_table;
-
-  explicit PdtExpandFstOptions(
-      const CacheOptions &opts = CacheOptions(), bool keep_parentheses = false,
-      PdtStack<typename Arc::StateId, typename Arc::Label> *stack = nullptr,
-      PdtStateTable<typename Arc::StateId, typename Arc::StateId> *state_table =
-          nullptr)
-      : CacheOptions(opts),
-        keep_parentheses(keep_parentheses),
-        stack(stack),
-        state_table(state_table) {}
-};
-
-namespace internal {
-
-// Implementation class for PdtExpandFst.
-template <class Arc>
-class PdtExpandFstImpl : public CacheImpl<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using StackId = StateId;
-  using StateTuple = PdtStateTuple<StateId, StackId>;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::Properties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheBaseImpl<CacheState<Arc>>::PushArc;
-  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
-  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
-  using CacheBaseImpl<CacheState<Arc>>::HasStart;
-  using CacheBaseImpl<CacheState<Arc>>::SetArcs;
-  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
-  using CacheBaseImpl<CacheState<Arc>>::SetStart;
-
-  PdtExpandFstImpl(const Fst<Arc> &fst,
-                   const std::vector<std::pair<Label, Label>> &parens,
-                   const PdtExpandFstOptions<Arc> &opts)
-      : CacheImpl<Arc>(opts),
-        fst_(fst.Copy()),
-        stack_(opts.stack ? opts.stack : new PdtStack<StateId, Label>(parens)),
-        state_table_(opts.state_table ? opts.state_table
-                                      : new PdtStateTable<StateId, StackId>()),
-        own_stack_(opts.stack == 0),
-        own_state_table_(opts.state_table == 0),
-        keep_parentheses_(opts.keep_parentheses) {
-    SetType("expand");
-    const auto props = fst.Properties(kFstProperties, false);
-    SetProperties(PdtExpandProperties(props), kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  PdtExpandFstImpl(const PdtExpandFstImpl &impl)
-      : CacheImpl<Arc>(impl),
-        fst_(impl.fst_->Copy(true)),
-        stack_(new PdtStack<StateId, Label>(*impl.stack_)),
-        state_table_(new PdtStateTable<StateId, StackId>()),
-        own_stack_(true),
-        own_state_table_(true),
-        keep_parentheses_(impl.keep_parentheses_) {
-    SetType("expand");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  ~PdtExpandFstImpl() override {
-    if (own_stack_) delete stack_;
-    if (own_state_table_) delete state_table_;
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      const auto s = fst_->Start();
-      if (s == kNoStateId) return kNoStateId;
-      StateTuple tuple(s, 0);
-      const auto start = state_table_->FindState(tuple);
-      SetStart(start);
-    }
-    return CacheImpl<Arc>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) {
-      const auto &tuple = state_table_->Tuple(s);
-      const auto weight = fst_->Final(tuple.state_id);
-      if (weight != Weight::Zero() && tuple.stack_id == 0)
-        SetFinal(s, weight);
-      else
-        SetFinal(s, Weight::Zero());
-    }
-    return CacheImpl<Arc>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) ExpandState(s);
-    return CacheImpl<Arc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) ExpandState(s);
-    return CacheImpl<Arc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) ExpandState(s);
-    return CacheImpl<Arc>::NumOutputEpsilons(s);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) ExpandState(s);
-    CacheImpl<Arc>::InitArcIterator(s, data);
-  }
-
-  // Computes the outgoing transitions from a state, creating new destination
-  // states as needed.
-  void ExpandState(StateId s) {
-    StateTuple tuple = state_table_->Tuple(s);
-    for (ArcIterator<Fst<Arc>> aiter(*fst_, tuple.state_id); !aiter.Done();
-         aiter.Next()) {
-      auto arc = aiter.Value();
-      const auto stack_id = stack_->Find(tuple.stack_id, arc.ilabel);
-      if (stack_id == -1) {  // Non-matching close parenthesis.
-        continue;
-      } else if ((stack_id != tuple.stack_id) && !keep_parentheses_) {
-        // Stack push/pop.
-        arc.ilabel = 0;
-        arc.olabel = 0;
-      }
-      StateTuple ntuple(arc.nextstate, stack_id);
-      arc.nextstate = state_table_->FindState(ntuple);
-      PushArc(s, arc);
-    }
-    SetArcs(s);
-  }
-
-  const PdtStack<StackId, Label> &GetStack() const { return *stack_; }
-
-  const PdtStateTable<StateId, StackId> &GetStateTable() const {
-    return *state_table_;
-  }
-
- private:
-  // Properties for an expanded PDT.
-  inline uint64 PdtExpandProperties(uint64 inprops) {
-    return inprops & (kAcceptor | kAcyclic | kInitialAcyclic | kUnweighted);
-  }
-
-  std::unique_ptr<const Fst<Arc>> fst_;
-  PdtStack<StackId, Label> *stack_;
-  PdtStateTable<StateId, StackId> *state_table_;
-  bool own_stack_;
-  bool own_state_table_;
-  bool keep_parentheses_;
-};
-
-}  // namespace internal
-
-// Expands a pushdown transducer (PDT) encoded as an FST into an FST. This
-// version is a delayed FST. In the PDT, some transitions are labeled with open
-// or close parentheses. To be interpreted as a PDT, the parens must balance on
-// a path. The open-close parenthesis label pairs are passed using the parens
-// argument. The expansion enforces the parenthesis constraints. The PDT must be
-// expandable as an FST.
-//
-// This class attaches interface to implementation and handles reference
-// counting, delegating most methods to ImplToFst.
-template <class A>
-class PdtExpandFst : public ImplToFst<internal::PdtExpandFstImpl<A>> {
- public:
-  using Arc = A;
-
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using StackId = StateId;
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::PdtExpandFstImpl<Arc>;
-
-  friend class ArcIterator<PdtExpandFst<Arc>>;
-  friend class StateIterator<PdtExpandFst<Arc>>;
-
-  PdtExpandFst(const Fst<Arc> &fst,
-               const std::vector<std::pair<Label, Label>> &parens)
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, parens, PdtExpandFstOptions<A>())) {}
-
-  PdtExpandFst(const Fst<Arc> &fst,
-               const std::vector<std::pair<Label, Label>> &parens,
-               const PdtExpandFstOptions<Arc> &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, parens, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  PdtExpandFst(const PdtExpandFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Gets a copy of this ExpandFst. See Fst<>::Copy() for further doc.
-  PdtExpandFst<Arc> *Copy(bool safe = false) const override {
-    return new PdtExpandFst<Arc>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
-  const PdtStack<StackId, Label> &GetStack() const {
-    return GetImpl()->GetStack();
-  }
-
-  const PdtStateTable<StateId, StackId> &GetStateTable() const {
-    return GetImpl()->GetStateTable();
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  void operator=(const PdtExpandFst &) = delete;
-};
-
-// Specialization for PdtExpandFst.
-template <class Arc>
-class StateIterator<PdtExpandFst<Arc>>
-    : public CacheStateIterator<PdtExpandFst<Arc>> {
- public:
-  explicit StateIterator(const PdtExpandFst<Arc> &fst)
-      : CacheStateIterator<PdtExpandFst<Arc>>(fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for PdtExpandFst.
-template <class Arc>
-class ArcIterator<PdtExpandFst<Arc>>
-    : public CacheArcIterator<PdtExpandFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const PdtExpandFst<Arc> &fst, StateId s)
-      : CacheArcIterator<PdtExpandFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->ExpandState(s);
-  }
-};
-
-template <class Arc>
-inline void PdtExpandFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<PdtExpandFst<Arc>>(*this);
-}
-
-// PrunedExpand prunes the delayed expansion of a pushdown transducer (PDT)
-// encoded as an FST into an FST. In the PDT, some transitions are labeled with
-// open or close parentheses. To be interpreted as a PDT, the parens must
-// balance on a path. The open-close parenthesis label pairs are passed
-// using the parens argument. The expansion enforces the parenthesis
-// constraints.
-//
-// The algorithm works by visiting the delayed ExpandFst using a shortest-stack
-// first queue discipline and relies on the shortest-distance information
-// computed using a reverse shortest-path call to perform the pruning.
-//
-// The algorithm maintains the same state ordering between the ExpandFst being
-// visited (efst_) and the result of pruning written into the MutableFst (ofst_)
-// to improve readability.
-template <class Arc>
-class PdtPrunedExpand {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using StackId = StateId;
-  using Stack = PdtStack<StackId, Label>;
-  using StateTable = PdtStateTable<StateId, StackId>;
-  using SetIterator = typename internal::PdtBalanceData<Arc>::SetIterator;
-
-  // Constructor taking as input a PDT specified by by an input FST and a vector
-  // of parentheses. The keep_parentheses argument specifies whether parentheses
-  // are replaced by epsilons or not during the expansion. The cache options are
-  // passed to the underlying ExpandFst.
-  PdtPrunedExpand(const Fst<Arc> &ifst,
-                  const std::vector<std::pair<Label, Label>> &parens,
-                  bool keep_parentheses = false,
-                  const CacheOptions &opts = CacheOptions())
-      : ifst_(ifst.Copy()),
-        keep_parentheses_(keep_parentheses),
-        stack_(parens),
-        efst_(ifst, parens,
-              PdtExpandFstOptions<Arc>(opts, true, &stack_, &state_table_)),
-        queue_(state_table_, stack_, stack_length_, distance_, fdistance_),
-        error_(false) {
-    Reverse(*ifst_, parens, &rfst_);
-    VectorFst<Arc> path;
-    reverse_shortest_path_.reset(new PdtShortestPath<Arc, FifoQueue<StateId>>(
-        rfst_, parens,
-        PdtShortestPathOptions<Arc, FifoQueue<StateId>>(true, false)));
-    reverse_shortest_path_->ShortestPath(&path);
-    error_ = (path.Properties(kError, true) == kError);
-    balance_data_.reset(reverse_shortest_path_->GetBalanceData()->Reverse(
-        rfst_.NumStates(), 10, -1));
-    InitCloseParenMultimap(parens);
-  }
-
-  bool Error() const { return error_; }
-
-  // Expands and prunes the input PDT according to the provided weight
-  // threshold, wirting the result into an output mutable FST.
-  void Expand(MutableFst<Arc> *ofst, const Weight &threshold);
-
- private:
-  static constexpr uint8 kEnqueued = 0x01;
-  static constexpr uint8 kExpanded = 0x02;
-  static constexpr uint8 kSourceState = 0x04;
-
-  // Comparison functor used by the queue:
-  //
-  // 1. States corresponding to shortest stack first, and
-  // 2. for stacks of matching length, reverse lexicographic order is used, and
-  // 3. for states with the same stack, shortest-first order is used.
-  class StackCompare {
-   public:
-    StackCompare(const StateTable &state_table, const Stack &stack,
-                 const std::vector<StackId> &stack_length,
-                 const std::vector<Weight> &distance,
-                 const std::vector<Weight> &fdistance)
-        : state_table_(state_table),
-          stack_(stack),
-          stack_length_(stack_length),
-          distance_(distance),
-          fdistance_(fdistance) {}
-
-    bool operator()(StateId s1, StateId s2) const {
-      auto si1 = state_table_.Tuple(s1).stack_id;
-      auto si2 = state_table_.Tuple(s2).stack_id;
-      if (stack_length_[si1] < stack_length_[si2]) return true;
-      if (stack_length_[si1] > stack_length_[si2]) return false;
-      // If stack IDs are equal, use A*.
-      if (si1 == si2) {
-        return less_(Distance(s1), Distance(s2));
-      }
-      // If lengths are equal, uses reverse lexicographic order.
-      for (; si1 != si2; si1 = stack_.Pop(si1), si2 = stack_.Pop(si2)) {
-        if (stack_.Top(si1) < stack_.Top(si2)) return true;
-        if (stack_.Top(si1) > stack_.Top(si2)) return false;
-      }
-      return false;
-    }
-
-   private:
-    Weight Distance(StateId s) const {
-      return (s < distance_.size()) && (s < fdistance_.size())
-                 ? Times(distance_[s], fdistance_[s])
-                 : Weight::Zero();
-    }
-
-    const StateTable &state_table_;
-    const Stack &stack_;
-    const std::vector<StackId> &stack_length_;
-    const std::vector<Weight> &distance_;
-    const std::vector<Weight> &fdistance_;
-    const NaturalLess<Weight> less_;
-  };
-
-  class ShortestStackFirstQueue
-      : public ShortestFirstQueue<StateId, StackCompare> {
-   public:
-    ShortestStackFirstQueue(const PdtStateTable<StateId, StackId> &state_table,
-                            const Stack &stack,
-                            const std::vector<StackId> &stack_length,
-                            const std::vector<Weight> &distance,
-                            const std::vector<Weight> &fdistance)
-        : ShortestFirstQueue<StateId, StackCompare>(StackCompare(
-              state_table, stack, stack_length, distance, fdistance)) {}
-  };
-
-  void InitCloseParenMultimap(
-      const std::vector<std::pair<Label, Label>> &parens);
-
-  Weight DistanceToDest(StateId source, StateId dest) const;
-
-  uint8 Flags(StateId s) const;
-
-  void SetFlags(StateId s, uint8 flags, uint8 mask);
-
-  Weight Distance(StateId s) const;
-
-  void SetDistance(StateId s, Weight weight);
-
-  Weight FinalDistance(StateId s) const;
-
-  void SetFinalDistance(StateId s, Weight weight);
-
-  StateId SourceState(StateId s) const;
-
-  void SetSourceState(StateId s, StateId p);
-
-  void AddStateAndEnqueue(StateId s);
-
-  void Relax(StateId s, const Arc &arc, Weight weight);
-
-  bool PruneArc(StateId s, const Arc &arc);
-
-  void ProcStart();
-
-  void ProcFinal(StateId s);
-
-  bool ProcNonParen(StateId s, const Arc &arc, bool add_arc);
-
-  bool ProcOpenParen(StateId s, const Arc &arc, StackId si, StackId nsi);
-
-  bool ProcCloseParen(StateId s, const Arc &arc);
-
-  void ProcDestStates(StateId s, StackId si);
-
-  // Input PDT.
-  std::unique_ptr<Fst<Arc>> ifst_;
-  // Reversed PDT.
-  VectorFst<Arc> rfst_;
-  // Keep parentheses in ofst?
-  const bool keep_parentheses_;
-  // State table for efst_.
-  StateTable state_table_;
-  // Stack trie.
-  Stack stack_;
-  // Expanded PDT.
-  PdtExpandFst<Arc> efst_;
-  // Length of stack for given stack ID.
-  std::vector<StackId> stack_length_;
-  // Distance from initial state in efst_/ofst.
-  std::vector<Weight> distance_;
-  // Distance to final states in efst_/ofst.
-  std::vector<Weight> fdistance_;
-  // Queue used to visit efst_.
-  ShortestStackFirstQueue queue_;
-  // Construction time failure?
-  bool error_;
-  // Status flags for states in efst_/ofst.
-  std::vector<uint8> flags_;
-  // PDT source state for each expanded state.
-  std::vector<StateId> sources_;
-  // Shortest path for rfst_.
-  std::unique_ptr<PdtShortestPath<Arc, FifoQueue<StateId>>>
-      reverse_shortest_path_;
-  std::unique_ptr<internal::PdtBalanceData<Arc>> balance_data_;
-  // Maps open paren arcs to balancing close paren arcs.
-  typename PdtShortestPath<Arc, FifoQueue<StateId>>::CloseParenMultimap
-      close_paren_multimap_;
-  MutableFst<Arc> *ofst_;  // Output FST.
-  Weight limit_;           // Weight limit.
-
-  // Maps a state s in ifst (i.e., the source of a closed paranthesis matching
-  // the top of current_stack_id_ to final states in efst_.
-  std::unordered_map<StateId, Weight> dest_map_;
-  // Stack ID of the states currently at the top of the queue, i.e., the states
-  // currently being popped and processed.
-  StackId current_stack_id_;
-  ssize_t current_paren_id_;  // Paren ID at top of current stack.
-  ssize_t cached_stack_id_;
-  StateId cached_source_;
-  // The set of pairs of destination states and weights to final states for the
-  // source state cached_source_ and the paren ID cached_paren_id_; i.e., the
-  // set of source states of a closed parenthesis with paren ID cached_paren_id
-  // balancing an incoming open parenthesis with paren ID cached_paren_id_ in
-  // state cached_source_.
-  std::forward_list<std::pair<StateId, Weight>> cached_dest_list_;
-  NaturalLess<Weight> less_;
-};
-
-// Initializes close paren multimap, mapping pairs (s, paren_id) to all the arcs
-// out of s labeled with close parenthese for paren_id.
-template <class Arc>
-void PdtPrunedExpand<Arc>::InitCloseParenMultimap(
-    const std::vector<std::pair<Label, Label>> &parens) {
-  std::unordered_map<Label, Label> paren_map;
-  for (size_t i = 0; i < parens.size(); ++i) {
-    const auto &pair = parens[i];
-    paren_map[pair.first] = i;
-    paren_map[pair.second] = i;
-  }
-  for (StateIterator<Fst<Arc>> siter(*ifst_); !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    for (ArcIterator<Fst<Arc>> aiter(*ifst_, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto it = paren_map.find(arc.ilabel);
-      if (it == paren_map.end()) continue;
-      if (arc.ilabel == parens[it->second].second) {  // Close paren.
-        const internal::ParenState<Arc> key(it->second, s);
-        close_paren_multimap_.emplace(key, arc);
-      }
-    }
-  }
-}
-
-// Returns the weight of the shortest balanced path from source to dest
-// in ifst_; dest must be the source state of a close paren arc.
-template <class Arc>
-typename Arc::Weight PdtPrunedExpand<Arc>::DistanceToDest(StateId source,
-                                                          StateId dest) const {
-  using SearchState =
-      typename PdtShortestPath<Arc, FifoQueue<StateId>>::SearchState;
-  const SearchState ss(source + 1, dest + 1);
-  const auto distance =
-      reverse_shortest_path_->GetShortestPathData().Distance(ss);
-  VLOG(2) << "D(" << source << ", " << dest << ") =" << distance;
-  return distance;
-}
-
-// Returns the flags for state s in ofst_.
-template <class Arc>
-uint8 PdtPrunedExpand<Arc>::Flags(StateId s) const {
-  return s < flags_.size() ? flags_[s] : 0;
-}
-
-// Modifies the flags for state s in ofst_.
-template <class Arc>
-void PdtPrunedExpand<Arc>::SetFlags(StateId s, uint8 flags, uint8 mask) {
-  while (flags_.size() <= s) flags_.push_back(0);
-  flags_[s] &= ~mask;
-  flags_[s] |= flags & mask;
-}
-
-// Returns the shortest distance from the initial state to s in ofst_.
-template <class Arc>
-typename Arc::Weight PdtPrunedExpand<Arc>::Distance(StateId s) const {
-  return s < distance_.size() ? distance_[s] : Weight::Zero();
-}
-
-// Sets the shortest distance from the initial state to s in ofst_.
-template <class Arc>
-void PdtPrunedExpand<Arc>::SetDistance(StateId s, Weight weight) {
-  while (distance_.size() <= s) distance_.push_back(Weight::Zero());
-  distance_[s] = std::move(weight);
-}
-
-// Returns the shortest distance from s to the final states in ofst_.
-template <class Arc>
-typename Arc::Weight PdtPrunedExpand<Arc>::FinalDistance(StateId s) const {
-  return s < fdistance_.size() ? fdistance_[s] : Weight::Zero();
-}
-
-// Sets the shortest distance from s to the final states in ofst_.
-template <class Arc>
-void PdtPrunedExpand<Arc>::SetFinalDistance(StateId s, Weight weight) {
-  while (fdistance_.size() <= s) fdistance_.push_back(Weight::Zero());
-  fdistance_[s] = std::move(weight);
-}
-
-// Returns the PDT source state of state s in ofst_.
-template <class Arc>
-typename Arc::StateId PdtPrunedExpand<Arc>::SourceState(StateId s) const {
-  return s < sources_.size() ? sources_[s] : kNoStateId;
-}
-
-// Sets the PDT source state of state s in ofst_ to state p'in ifst_.
-template <class Arc>
-void PdtPrunedExpand<Arc>::SetSourceState(StateId s, StateId p) {
-  while (sources_.size() <= s) sources_.push_back(kNoStateId);
-  sources_[s] = p;
-}
-
-// Adds state s of efst_ to ofst_ and inserts it in the queue, modifying the
-// flags for s accordingly.
-template <class Arc>
-void PdtPrunedExpand<Arc>::AddStateAndEnqueue(StateId s) {
-  if (!(Flags(s) & (kEnqueued | kExpanded))) {
-    while (ofst_->NumStates() <= s) ofst_->AddState();
-    queue_.Enqueue(s);
-    SetFlags(s, kEnqueued, kEnqueued);
-  } else if (Flags(s) & kEnqueued) {
-    queue_.Update(s);
-  }
-  // TODO(allauzen): Check everything is fine when kExpanded?
-}
-
-// Relaxes arc out of state s in ofst_ as follows:
-//
-// 1. If the distance to s times the weight of arc is smaller than
-//   the currently stored distance for arc.nextstate, updates
-//   Distance(arc.nextstate) with a new estimate
-// 2. If fd is less than the currently stored distance from arc.nextstate to the
-// final state, updates with new estimate.
-template <class Arc>
-void PdtPrunedExpand<Arc>::Relax(StateId s, const Arc &arc, Weight fd) {
-  const auto nd = Times(Distance(s), arc.weight);
-  if (less_(nd, Distance(arc.nextstate))) {
-    SetDistance(arc.nextstate, nd);
-    SetSourceState(arc.nextstate, SourceState(s));
-  }
-  if (less_(fd, FinalDistance(arc.nextstate))) {
-    SetFinalDistance(arc.nextstate, fd);
-  }
-  VLOG(2) << "Relax: " << s << ", d[s] = " << Distance(s) << ", to "
-          << arc.nextstate << ", d[ns] = " << Distance(arc.nextstate)
-          << ", nd = " << nd;
-}
-
-// Returns whether the arc out of state s in efst needs pruned.
-template <class Arc>
-bool PdtPrunedExpand<Arc>::PruneArc(StateId s, const Arc &arc) {
-  VLOG(2) << "Prune ?";
-  auto fd = Weight::Zero();
-  if ((cached_source_ != SourceState(s)) ||
-      (cached_stack_id_ != current_stack_id_)) {
-    cached_source_ = SourceState(s);
-    cached_stack_id_ = current_stack_id_;
-    cached_dest_list_.clear();
-    if (cached_source_ != ifst_->Start()) {
-      for (auto set_iter =
-               balance_data_->Find(current_paren_id_, cached_source_);
-           !set_iter.Done(); set_iter.Next()) {
-        auto dest = set_iter.Element();
-        const auto it = dest_map_.find(dest);
-        cached_dest_list_.push_front(*it);
-      }
-    } else {
-      // TODO(allauzen): queue discipline should prevent this from ever
-      // happening.
-      // Replace by a check.
-      cached_dest_list_.push_front(
-          std::make_pair(rfst_.Start() - 1, Weight::One()));
-    }
-  }
-  for (auto it = cached_dest_list_.begin(); it != cached_dest_list_.end();
-       ++it) {
-    const auto d =
-        DistanceToDest(state_table_.Tuple(arc.nextstate).state_id, it->first);
-    fd = Plus(fd, Times(d, it->second));
-  }
-  Relax(s, arc, fd);
-  return less_(limit_, Times(Distance(s), Times(arc.weight, fd)));
-}
-
-// Adds start state of efst_ to ofst_, enqueues it, and initializes the distance
-// data structures.
-template <class Arc>
-void PdtPrunedExpand<Arc>::ProcStart() {
-  const auto s = efst_.Start();
-  AddStateAndEnqueue(s);
-  ofst_->SetStart(s);
-  SetSourceState(s, ifst_->Start());
-  current_stack_id_ = 0;
-  current_paren_id_ = -1;
-  stack_length_.push_back(0);
-  const auto r = rfst_.Start() - 1;
-  cached_source_ = ifst_->Start();
-  cached_stack_id_ = 0;
-  cached_dest_list_.push_front(std::make_pair(r, Weight::One()));
-  const PdtStateTuple<StateId, StackId> tuple(r, 0);
-  SetFinalDistance(state_table_.FindState(tuple), Weight::One());
-  SetDistance(s, Weight::One());
-  const auto d = DistanceToDest(ifst_->Start(), r);
-  SetFinalDistance(s, d);
-  VLOG(2) << d;
-}
-
-// Makes s final in ofst_ if shortest accepting path ending in s is below
-// threshold.
-template <class Arc>
-void PdtPrunedExpand<Arc>::ProcFinal(StateId s) {
-  const auto weight = efst_.Final(s);
-  if (weight == Weight::Zero()) return;
-  if (less_(limit_, Times(Distance(s), weight))) return;
-  ofst_->SetFinal(s, weight);
-}
-
-// Returns true when an arc (or meta-arc) leaving state s in efst_ is below the
-// threshold. When add_arc is true, arc is added to ofst_.
-template <class Arc>
-bool PdtPrunedExpand<Arc>::ProcNonParen(StateId s, const Arc &arc,
-                                        bool add_arc) {
-  VLOG(2) << "ProcNonParen: " << s << " to " << arc.nextstate << ", "
-          << arc.ilabel << ":" << arc.olabel << " / " << arc.weight
-          << ", add_arc = " << (add_arc ? "true" : "false");
-  if (PruneArc(s, arc)) return false;
-  if (add_arc) ofst_->AddArc(s, arc);
-  AddStateAndEnqueue(arc.nextstate);
-  return true;
-}
-
-// Processes an open paren arc leaving state s in ofst_. When the arc is labeled
-// with an open paren,
-//
-// 1. Considers each (shortest) balanced path starting in s by taking the arc
-// and ending by a close paren balancing the open paren of as a meta-arc,
-// processing and pruning each meta-arc as a non-paren arc, inserting its
-// destination to the queue;
-// 2. if at least one of these meta-arcs has not been pruned, adds the
-// destination of arc to ofst_ as a new source state for the stack ID nsi, and
-// inserts it in the queue.
-template <class Arc>
-bool PdtPrunedExpand<Arc>::ProcOpenParen(StateId s, const Arc &arc, StackId si,
-                                         StackId nsi) {
-  // Updates the stack length when needed.
-  while (stack_length_.size() <= nsi) stack_length_.push_back(-1);
-  if (stack_length_[nsi] == -1) stack_length_[nsi] = stack_length_[si] + 1;
-  const auto ns = arc.nextstate;
-  VLOG(2) << "Open paren: " << s << "(" << state_table_.Tuple(s).state_id
-          << ") to " << ns << "(" << state_table_.Tuple(ns).state_id << ")";
-  bool proc_arc = false;
-  auto fd = Weight::Zero();
-  const auto paren_id = stack_.ParenId(arc.ilabel);
-  std::forward_list<StateId> sources;
-  for (auto set_iter =
-           balance_data_->Find(paren_id, state_table_.Tuple(ns).state_id);
-       !set_iter.Done(); set_iter.Next()) {
-    sources.push_front(set_iter.Element());
-  }
-  for (const auto source : sources) {
-    VLOG(2) << "Close paren source: " << source;
-    const internal::ParenState<Arc> paren_state(paren_id, source);
-    for (auto it = close_paren_multimap_.find(paren_state);
-         it != close_paren_multimap_.end() && paren_state == it->first; ++it) {
-      auto meta_arc = it->second;
-      const PdtStateTuple<StateId, StackId> tuple(meta_arc.nextstate, si);
-      meta_arc.nextstate = state_table_.FindState(tuple);
-      const auto state_id = state_table_.Tuple(ns).state_id;
-      const auto d = DistanceToDest(state_id, source);
-      VLOG(2) << state_id << ", " << source;
-      VLOG(2) << "Meta arc weight = " << arc.weight << " Times " << d
-              << " Times " << meta_arc.weight;
-      meta_arc.weight = Times(arc.weight, Times(d, meta_arc.weight));
-      proc_arc |= ProcNonParen(s, meta_arc, false);
-      fd = Plus(
-          fd,
-          Times(Times(DistanceToDest(state_table_.Tuple(ns).state_id, source),
-                      it->second.weight),
-                FinalDistance(meta_arc.nextstate)));
-    }
-  }
-  if (proc_arc) {
-    VLOG(2) << "Proc open paren " << s << " to " << arc.nextstate;
-    ofst_->AddArc(
-        s, keep_parentheses_ ? arc : Arc(0, 0, arc.weight, arc.nextstate));
-    AddStateAndEnqueue(arc.nextstate);
-    const auto nd = Times(Distance(s), arc.weight);
-    if (less_(nd, Distance(arc.nextstate))) SetDistance(arc.nextstate, nd);
-    // FinalDistance not necessary for source state since pruning decided using
-    // meta-arcs above.  But this is a problem with A*, hence the following.
-    if (less_(fd, FinalDistance(arc.nextstate)))
-      SetFinalDistance(arc.nextstate, fd);
-    SetFlags(arc.nextstate, kSourceState, kSourceState);
-  }
-  return proc_arc;
-}
-
-// Checks that shortest path through close paren arc in efst_ is below
-// threshold, and if so, adds it to ofst_.
-template <class Arc>
-bool PdtPrunedExpand<Arc>::ProcCloseParen(StateId s, const Arc &arc) {
-  const auto weight =
-      Times(Distance(s), Times(arc.weight, FinalDistance(arc.nextstate)));
-  if (less_(limit_, weight)) return false;
-  ofst_->AddArc(s,
-                keep_parentheses_ ? arc : Arc(0, 0, arc.weight, arc.nextstate));
-  return true;
-}
-
-// When state s in ofst_ is a source state for stack ID si, identifies all the
-// corresponding possible destination states, that is, all the states in ifst_
-// that have an outgoing close paren arc balancing the incoming open paren taken
-// to get to s. For each such state t, computes the shortest distance from (t,
-// si) to the final states in ofst_. Stores this information in dest_map_.
-template <class Arc>
-void PdtPrunedExpand<Arc>::ProcDestStates(StateId s, StackId si) {
-  if (!(Flags(s) & kSourceState)) return;
-  if (si != current_stack_id_) {
-    dest_map_.clear();
-    current_stack_id_ = si;
-    current_paren_id_ = stack_.Top(current_stack_id_);
-    VLOG(2) << "StackID " << si << " dequeued for first time";
-  }
-  // TODO(allauzen): clean up source state business; rename current function to
-  // ProcSourceState.
-  SetSourceState(s, state_table_.Tuple(s).state_id);
-  const auto paren_id = stack_.Top(si);
-  for (auto set_iter =
-           balance_data_->Find(paren_id, state_table_.Tuple(s).state_id);
-       !set_iter.Done(); set_iter.Next()) {
-    const auto dest_state = set_iter.Element();
-    if (dest_map_.find(dest_state) != dest_map_.end()) continue;
-    auto dest_weight = Weight::Zero();
-    internal::ParenState<Arc> paren_state(paren_id, dest_state);
-    for (auto it = close_paren_multimap_.find(paren_state);
-         it != close_paren_multimap_.end() && paren_state == it->first; ++it) {
-      const auto &arc = it->second;
-      const PdtStateTuple<StateId, StackId> tuple(arc.nextstate,
-                                                  stack_.Pop(si));
-      dest_weight =
-          Plus(dest_weight,
-               Times(arc.weight, FinalDistance(state_table_.FindState(tuple))));
-    }
-    dest_map_[dest_state] = dest_weight;
-    VLOG(2) << "State " << dest_state << " is a dest state for stack ID " << si
-            << " with weight " << dest_weight;
-  }
-}
-
-// Expands and prunes the input PDT, writing the result in ofst.
-template <class Arc>
-void PdtPrunedExpand<Arc>::Expand(MutableFst<Arc> *ofst,
-                                  const typename Arc::Weight &threshold) {
-  ofst_ = ofst;
-  if (error_) {
-    ofst_->SetProperties(kError, kError);
-    return;
-  }
-  ofst_->DeleteStates();
-  ofst_->SetInputSymbols(ifst_->InputSymbols());
-  ofst_->SetOutputSymbols(ifst_->OutputSymbols());
-  limit_ = Times(DistanceToDest(ifst_->Start(), rfst_.Start() - 1), threshold);
-  flags_.clear();
-  ProcStart();
-  while (!queue_.Empty()) {
-    const auto s = queue_.Head();
-    queue_.Dequeue();
-    SetFlags(s, kExpanded, kExpanded | kEnqueued);
-    VLOG(2) << s << " dequeued!";
-    ProcFinal(s);
-    StackId stack_id = state_table_.Tuple(s).stack_id;
-    ProcDestStates(s, stack_id);
-    for (ArcIterator<PdtExpandFst<Arc>> aiter(efst_, s); !aiter.Done();
-         aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto nextstack_id = state_table_.Tuple(arc.nextstate).stack_id;
-      if (stack_id == nextstack_id) {
-        ProcNonParen(s, arc, true);
-      } else if (stack_id == stack_.Pop(nextstack_id)) {
-        ProcOpenParen(s, arc, stack_id, nextstack_id);
-      } else {
-        ProcCloseParen(s, arc);
-      }
-    }
-    VLOG(2) << "d[" << s << "] = " << Distance(s) << ", fd[" << s
-            << "] = " << FinalDistance(s);
-  }
-}
-
-// Expand functions.
-
-template <class Arc>
-struct PdtExpandOptions {
-  using Weight = typename Arc::Weight;
-
-  bool connect;
-  bool keep_parentheses;
-  Weight weight_threshold;
-
-  PdtExpandOptions(bool connect = true, bool keep_parentheses = false,
-                   Weight weight_threshold = Weight::Zero())
-      : connect(connect),
-        keep_parentheses(keep_parentheses),
-        weight_threshold(std::move(weight_threshold)) {}
-};
-
-// Expands a pushdown transducer (PDT) encoded as an FST into an FST. This
-// version writes the expanded PDT to a mutable FST. In the PDT, some
-// transitions are labeled with open or close parentheses. To be interpreted as
-// a PDT, the parens must balance on a path. The open-close parenthesis label
-// pairs are passed using the parens argument. Expansion enforces the
-// parenthesis constraints. The PDT must be expandable as an FST.
-template <class Arc>
-void Expand(
-    const Fst<Arc> &ifst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens,
-    MutableFst<Arc> *ofst, const PdtExpandOptions<Arc> &opts) {
-  PdtExpandFstOptions<Arc> eopts;
-  eopts.gc_limit = 0;
-  if (opts.weight_threshold == Arc::Weight::Zero()) {
-    eopts.keep_parentheses = opts.keep_parentheses;
-    *ofst = PdtExpandFst<Arc>(ifst, parens, eopts);
-  } else {
-    PdtPrunedExpand<Arc> pruned_expand(ifst, parens, opts.keep_parentheses);
-    pruned_expand.Expand(ofst, opts.weight_threshold);
-  }
-  if (opts.connect) Connect(ofst);
-}
-
-// Expands a pushdown transducer (PDT) encoded as an FST into an FST. This
-// version writes the expanded PDT result to a mutable FST. In the PDT, some
-// transitions are labeled with open or close parentheses. To be interpreted as
-// a PDT, the parens must balance on a path. The open-close parenthesis label
-// pairs are passed using the parents argument. Expansion enforces the
-// parenthesis constraints. The PDT must be expandable as an FST.
-template <class Arc>
-void Expand(const Fst<Arc> &ifst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-    &parens, MutableFst<Arc> *ofst, bool connect = true,
-    bool keep_parentheses = false) {
-  const PdtExpandOptions<Arc> opts(connect, keep_parentheses);
-  Expand(ifst, parens, ofst, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_EXPAND_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/getters.h b/projects/llm_framework/include/fst/extensions/pdt/getters.h
deleted file mode 100644
index 69dd150d..00000000
--- a/projects/llm_framework/include/fst/extensions/pdt/getters.h
+++ /dev/null
@@ -1,22 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_PDT_GETTERS_H_
-#define FST_EXTENSIONS_PDT_GETTERS_H_
-
-#include <string>
-
-#include <fst/extensions/pdt/compose.h>
-#include <fst/extensions/pdt/replace.h>
-
-namespace fst {
-namespace script {
-
-bool GetPdtComposeFilter(const string &str, PdtComposeFilter *cf);
-
-bool GetPdtParserType(const string &str, PdtParserType *pt);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_GETTERS_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/info.h b/projects/llm_framework/include/fst/extensions/pdt/info.h
deleted file mode 100644
index 3de54772..00000000
--- a/projects/llm_framework/include/fst/extensions/pdt/info.h
+++ /dev/null
@@ -1,152 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Prints information about a PDT.
-
-#ifndef FST_EXTENSIONS_PDT_INFO_H_
-#define FST_EXTENSIONS_PDT_INFO_H_
-
-#include <unordered_map>
-#include <unordered_set>
-#include <vector>
-
-#include <fst/extensions/pdt/pdt.h>
-#include <fst/fst.h>
-
-namespace fst {
-
-// Compute various information about PDTs.
-template <class Arc>
-class PdtInfo {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  PdtInfo(const Fst<Arc> &fst,
-          const std::vector<std::pair<Label, Label>> &parents);
-
-  const string &FstType() const { return fst_type_; }
-
-  const string &ArcType() const { return Arc::Type(); }
-
-  int64 NumStates() const { return nstates_; }
-
-  int64 NumArcs() const { return narcs_; }
-
-  int64 NumOpenParens() const { return nopen_parens_; }
-
-  int64 NumCloseParens() const { return nclose_parens_; }
-
-  int64 NumUniqueOpenParens() const { return nuniq_open_parens_; }
-
-  int64 NumUniqueCloseParens() const { return nuniq_close_parens_; }
-
-  int64 NumOpenParenStates() const { return nopen_paren_states_; }
-
-  int64 NumCloseParenStates() const { return nclose_paren_states_; }
-
- private:
-  string fst_type_;
-  int64 nstates_;
-  int64 narcs_;
-  int64 nopen_parens_;
-  int64 nclose_parens_;
-  int64 nuniq_open_parens_;
-  int64 nuniq_close_parens_;
-  int64 nopen_paren_states_;
-  int64 nclose_paren_states_;
-};
-
-template <class Arc>
-PdtInfo<Arc>::PdtInfo(
-    const Fst<Arc> &fst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens)
-    : fst_type_(fst.Type()),
-      nstates_(0),
-      narcs_(0),
-      nopen_parens_(0),
-      nclose_parens_(0),
-      nuniq_open_parens_(0),
-      nuniq_close_parens_(0),
-      nopen_paren_states_(0),
-      nclose_paren_states_(0) {
-  std::unordered_map<Label, size_t> paren_map;
-  std::unordered_set<Label> paren_set;
-  std::unordered_set<StateId> open_paren_state_set;
-  std::unordered_set<StateId> close_paren_state_set;
-  for (size_t i = 0; i < parens.size(); ++i) {
-    const auto &pair = parens[i];
-    paren_map[pair.first] = i;
-    paren_map[pair.second] = i;
-  }
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    ++nstates_;
-    const auto s = siter.Value();
-    for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      ++narcs_;
-      const auto it = paren_map.find(arc.ilabel);
-      if (it != paren_map.end()) {
-        const auto open_paren = parens[it->second].first;
-        const auto close_paren = parens[it->second].second;
-        if (arc.ilabel == open_paren) {
-          ++nopen_parens_;
-          if (!paren_set.count(open_paren)) {
-            ++nuniq_open_parens_;
-            paren_set.insert(open_paren);
-          }
-          if (!open_paren_state_set.count(arc.nextstate)) {
-            ++nopen_paren_states_;
-            open_paren_state_set.insert(arc.nextstate);
-          }
-        } else {
-          ++nclose_parens_;
-          if (!paren_set.count(close_paren)) {
-            ++nuniq_close_parens_;
-            paren_set.insert(close_paren);
-          }
-          if (!close_paren_state_set.count(s)) {
-            ++nclose_paren_states_;
-            close_paren_state_set.insert(s);
-          }
-        }
-      }
-    }
-  }
-}
-
-template <class Arc>
-void PrintPdtInfo(const PdtInfo<Arc> &info) {
-  const auto old = std::cout.setf(std::ios::left);
-  std::cout.width(50);
-  std::cout << "fst type" << info.FstType() << std::endl;
-  std::cout.width(50);
-  std::cout << "arc type" << info.ArcType() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of states" << info.NumStates() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of arcs" << info.NumArcs() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of open parentheses" << info.NumOpenParens() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of close parentheses" << info.NumCloseParens() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of unique open parentheses" << info.NumUniqueOpenParens()
-            << std::endl;
-  std::cout.width(50);
-  std::cout << "# of unique close parentheses" << info.NumUniqueCloseParens()
-            << std::endl;
-  std::cout.width(50);
-  std::cout << "# of open parenthesis dest. states" << info.NumOpenParenStates()
-            << std::endl;
-  std::cout.width(50);
-  std::cout << "# of close parenthesis source states"
-            << info.NumCloseParenStates() << std::endl;
-  std::cout.setf(old);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_INFO_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/paren.h b/projects/llm_framework/include/fst/extensions/pdt/paren.h
deleted file mode 100644
index 0347848d..00000000
--- a/projects/llm_framework/include/fst/extensions/pdt/paren.h
+++ /dev/null
@@ -1,440 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Common classes for PDT parentheses.
-
-#ifndef FST_EXTENSIONS_PDT_PAREN_H_
-#define FST_EXTENSIONS_PDT_PAREN_H_
-
-#include <algorithm>
-#include <set>
-#include <unordered_map>
-#include <unordered_set>
-
-#include <fst/log.h>
-
-#include <fst/extensions/pdt/collection.h>
-#include <fst/extensions/pdt/pdt.h>
-#include <fst/dfs-visit.h>
-#include <fst/fst.h>
-
-
-namespace fst {
-namespace internal {
-
-// ParenState: Pair of an open (close) parenthesis and its destination (source)
-// state.
-
-template <class Arc>
-struct ParenState {
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-
-  Label paren_id;    // ID of open (close) paren.
-  StateId state_id;  // Destination (source) state of open (close) paren.
-
-  explicit ParenState(Label paren_id = kNoLabel, StateId state_id = kNoStateId)
-      : paren_id(paren_id), state_id(state_id) {}
-
-  bool operator==(const ParenState<Arc> &other) const {
-    if (&other == this) return true;
-    return other.paren_id == paren_id && other.state_id == state_id;
-  }
-
-  bool operator!=(const ParenState<Arc> &other) const {
-    return !(other == *this);
-  }
-
-  struct Hash {
-    size_t operator()(const ParenState<Arc> &pstate) const {
-      static constexpr auto prime = 7853;
-      return pstate.paren_id + pstate.state_id * prime;
-    }
-  };
-};
-
-// Creates an FST-style const iterator from an STL-style map.
-template <class Map>
-class MapIterator {
- public:
-  using StlIterator = typename Map::const_iterator;
-  using ValueType = typename Map::mapped_type;
-
-  MapIterator(const Map &map, StlIterator it)
-      : begin_(it), end_(map.end()), it_(it) {}
-
-  bool Done() const { return it_ == end_ || it_->first != begin_->first; }
-
-  ValueType Value() const { return it_->second; }
-
-  void Next() { ++it_; }
-
-  void Reset() { it_ = begin_; }
-
- private:
-  const StlIterator begin_;
-  const StlIterator end_;
-  StlIterator it_;
-};
-
-// PdtParenReachable: Provides various parenthesis reachability information.
-
-template <class Arc>
-class PdtParenReachable {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-
-  using State = ParenState<Arc>;
-  using StateHash = typename State::Hash;
-
-  // Maps from state ID to reachable paren IDs from (to) that state.
-  using ParenMultimap = std::unordered_multimap<StateId, Label>;
-
-  // Maps from paren ID and state ID to reachable state set ID.
-  using StateSetMap = std::unordered_map<State, ssize_t, StateHash>;
-
-  // Maps from paren ID and state ID to arcs exiting that state with that
-  // Label.
-  using ParenArcMultimap = std::unordered_map<State, Arc, StateHash>;
-
-  using ParenIterator = MapIterator<ParenMultimap>;
-
-  using ParenArcIterator = MapIterator<ParenArcMultimap>;
-
-  using SetIterator = typename Collection<ssize_t, StateId>::SetIterator;
-
-  // Computes close (open) parenthesis reachability information for a PDT with
-  // bounded stack.
-  PdtParenReachable(const Fst<Arc> &fst,
-                    const std::vector<std::pair<Label, Label>> &parens,
-                    bool close)
-      : fst_(fst), parens_(parens), close_(close), error_(false) {
-    paren_map_.reserve(2 * parens.size());
-    for (size_t i = 0; i < parens.size(); ++i) {
-      const auto &pair = parens[i];
-      paren_map_[pair.first] = i;
-      paren_map_[pair.second] = i;
-    }
-    if (close_) {
-      const auto start = fst.Start();
-      if (start == kNoStateId) return;
-      if (!DFSearch(start)) {
-        FSTERROR() << "PdtReachable: Underlying cyclicity not supported";
-        error_ = true;
-      }
-    } else {
-      FSTERROR() << "PdtParenReachable: Open paren info not implemented";
-      error_ = true;
-    }
-  }
-
-  bool Error() const { return error_; }
-
-  // Given a state ID, returns an iterator over paren IDs for close (open)
-  // parens reachable from that state along balanced paths.
-  ParenIterator FindParens(StateId s) const {
-    return ParenIterator(paren_multimap_, paren_multimap_.find(s));
-  }
-
-  // Given a paren ID and a state ID s, returns an iterator over states that can
-  // be reached along balanced paths from (to) s that have have close (open)
-  // parentheses matching the paren ID exiting (entering) those states.
-  SetIterator FindStates(Label paren_id, StateId s) const {
-    const State paren_state(paren_id, s);
-    const auto it = set_map_.find(paren_state);
-    if (it == set_map_.end()) {
-      return state_sets_.FindSet(-1);
-    } else {
-      return state_sets_.FindSet(it->second);
-    }
-  }
-
-  // Given a paren ID and a state ID s, return an iterator over arcs that exit
-  // (enter) s and are labeled with a close (open) parenthesis matching the
-  // paren ID.
-  ParenArcIterator FindParenArcs(Label paren_id, StateId s) const {
-    const State paren_state(paren_id, s);
-    return ParenArcIterator(paren_arc_multimap_,
-                            paren_arc_multimap_.find(paren_state));
-  }
-
- private:
-  // Returns false when cycle detected during DFS gathering paren and state set
-  // information.
-  bool DFSearch(StateId s);
-
-  // Unions state sets together gathered by the DFS.
-  void ComputeStateSet(StateId s);
-
-  // Gathers state set(s) from state.
-  void UpdateStateSet(StateId nextstate, std::set<Label> *paren_set,
-                      std::vector<std::set<StateId>> *state_sets) const;
-
-  const Fst<Arc> &fst_;
-  // Paren IDs to labels.
-  const std::vector<std::pair<Label, Label>> &parens_;
-  // Close/open paren info?
-  const bool close_;
-  // Labels to paren IDs.
-  std::unordered_map<Label, Label> paren_map_;
-  // Paren reachability.
-  ParenMultimap paren_multimap_;
-  // Paren arcs.
-  ParenArcMultimap paren_arc_multimap_;
-  // DFS states.
-  std::vector<uint8> state_color_;
-  // Reachable states to IDs.
-  mutable Collection<ssize_t, StateId> state_sets_;
-  // IDs to reachable states.
-  StateSetMap set_map_;
-  bool error_;
-
-  PdtParenReachable(const PdtParenReachable &) = delete;
-  PdtParenReachable &operator=(const PdtParenReachable &) = delete;
-};
-
-// Gathers paren and state set information.
-template <class Arc>
-bool PdtParenReachable<Arc>::DFSearch(StateId s) {
-  static constexpr uint8 kWhiteState = 0x01;  // Undiscovered.
-  static constexpr uint8 kGreyState = 0x02;   // Discovered & unfinished.
-  static constexpr uint8 kBlackState = 0x04;  // Finished.
-  if (s >= state_color_.size()) state_color_.resize(s + 1, kWhiteState);
-  if (state_color_[s] == kBlackState) return true;
-  if (state_color_[s] == kGreyState) return false;
-  state_color_[s] = kGreyState;
-  for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
-    const auto &arc = aiter.Value();
-    const auto it = paren_map_.find(arc.ilabel);
-    if (it != paren_map_.end()) {  // Paren?
-      const auto paren_id = it->second;
-      if (arc.ilabel == parens_[paren_id].first) {  // Open paren?
-        if (!DFSearch(arc.nextstate)) return false;
-        for (auto set_iter = FindStates(paren_id, arc.nextstate);
-             !set_iter.Done(); set_iter.Next()) {
-          for (auto paren_arc_iter =
-                   FindParenArcs(paren_id, set_iter.Element());
-               !paren_arc_iter.Done(); paren_arc_iter.Next()) {
-            const auto &cparc = paren_arc_iter.Value();
-            if (!DFSearch(cparc.nextstate)) return false;
-          }
-        }
-      }
-    } else if (!DFSearch(arc.nextstate)) {  // Non-paren.
-      return false;
-    }
-  }
-  ComputeStateSet(s);
-  state_color_[s] = kBlackState;
-  return true;
-}
-
-// Unions state sets.
-template <class Arc>
-void PdtParenReachable<Arc>::ComputeStateSet(StateId s) {
-  std::set<Label> paren_set;
-  std::vector<std::set<StateId>> state_sets(parens_.size());
-  for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
-    const auto &arc = aiter.Value();
-    const auto it = paren_map_.find(arc.ilabel);
-    if (it != paren_map_.end()) {  // Paren?
-      const auto paren_id = it->second;
-      if (arc.ilabel == parens_[paren_id].first) {  // Open paren?
-        for (auto set_iter = FindStates(paren_id, arc.nextstate);
-             !set_iter.Done(); set_iter.Next()) {
-          for (auto paren_arc_iter =
-                   FindParenArcs(paren_id, set_iter.Element());
-               !paren_arc_iter.Done(); paren_arc_iter.Next()) {
-            const auto &cparc = paren_arc_iter.Value();
-            UpdateStateSet(cparc.nextstate, &paren_set, &state_sets);
-          }
-        }
-      } else {  // Close paren.
-        paren_set.insert(paren_id);
-        state_sets[paren_id].insert(s);
-        const State paren_state(paren_id, s);
-        paren_arc_multimap_.insert(std::make_pair(paren_state, arc));
-      }
-    } else {  // Non-paren.
-      UpdateStateSet(arc.nextstate, &paren_set, &state_sets);
-    }
-  }
-  std::vector<StateId> state_set;
-  for (auto paren_iter = paren_set.begin(); paren_iter != paren_set.end();
-       ++paren_iter) {
-    state_set.clear();
-    const auto paren_id = *paren_iter;
-    paren_multimap_.insert(std::make_pair(s, paren_id));
-    for (auto state_iter = state_sets[paren_id].begin();
-         state_iter != state_sets[paren_id].end(); ++state_iter) {
-      state_set.push_back(*state_iter);
-    }
-    const State paren_state(paren_id, s);
-    set_map_[paren_state] = state_sets_.FindId(state_set);
-  }
-}
-
-// Gathers state sets.
-template <class Arc>
-void PdtParenReachable<Arc>::UpdateStateSet(
-    StateId nextstate, std::set<Label> *paren_set,
-    std::vector<std::set<StateId>> *state_sets) const {
-  for (auto paren_iter = FindParens(nextstate); !paren_iter.Done();
-       paren_iter.Next()) {
-    const auto paren_id = paren_iter.Value();
-    paren_set->insert(paren_id);
-    for (auto set_iter = FindStates(paren_id, nextstate); !set_iter.Done();
-         set_iter.Next()) {
-      (*state_sets)[paren_id].insert(set_iter.Element());
-    }
-  }
-}
-
-// Stores balancing parenthesis data for a PDT. Unlike PdtParenReachable above
-// this allows on-the-fly construction (e.g., in PdtShortestPath).
-template <class Arc>
-class PdtBalanceData {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-
-  using State = ParenState<Arc>;
-  using StateHash = typename State::Hash;
-
-  // Set for open parens.
-  using OpenParenSet = std::unordered_set<State, StateHash>;
-
-  // Maps from open paren destination state to parenthesis ID.
-  using OpenParenMap = std::unordered_multimap<StateId, Label>;
-
-  // Maps from open paren state to source states of matching close parens
-  using CloseParenMap = std::unordered_multimap<State, StateId, StateHash>;
-
-  // Maps from open paren state to close source set ID.
-  using CloseSourceMap = std::unordered_map<State, ssize_t, StateHash>;
-
-  using SetIterator = typename Collection<ssize_t, StateId>::SetIterator;
-
-  PdtBalanceData() {}
-
-  void Clear() {
-    open_paren_map_.clear();
-    close_paren_map_.clear();
-  }
-
-  // Adds an open parenthesis with destination state open_dest.
-  void OpenInsert(Label paren_id, StateId open_dest) {
-    const State key(paren_id, open_dest);
-    if (!open_paren_set_.count(key)) {
-      open_paren_set_.insert(key);
-      open_paren_map_.emplace(open_dest, paren_id);
-    }
-  }
-
-  // Adds a matching closing parenthesis with source state close_source
-  // balancing an open_parenthesis with destination state open_dest if
-  // OpenInsert() previously called.
-  void CloseInsert(Label paren_id, StateId open_dest, StateId close_source) {
-    const State key(paren_id, open_dest);
-    if (open_paren_set_.count(key)) {
-      close_paren_map_.emplace(key, close_source);
-    }
-  }
-
-  // Finds close paren source states matching an open parenthesis. The following
-  // methods are then used to iterate through those matching states. Should be
-  // called only after FinishInsert(open_dest).
-  SetIterator Find(Label paren_id, StateId open_dest) {
-    const State key(paren_id, open_dest);
-    const auto it = close_source_map_.find(key);
-    if (it == close_source_map_.end()) {
-      return close_source_sets_.FindSet(-1);
-    } else {
-      return close_source_sets_.FindSet(it->second);
-    }
-  }
-
-  // Called when all open and close parenthesis insertions (w.r.t. open
-  // parentheses entering state open_dest) are finished. Must be called before
-  // Find(open_dest).
-  void FinishInsert(StateId open_dest) {
-    std::vector<StateId> close_sources;
-    for (auto oit = open_paren_map_.find(open_dest);
-         oit != open_paren_map_.end() && oit->first == open_dest;) {
-      const auto paren_id = oit->second;
-      close_sources.clear();
-      const State key(paren_id, open_dest);
-      open_paren_set_.erase(open_paren_set_.find(key));
-      for (auto cit = close_paren_map_.find(key);
-           cit != close_paren_map_.end() && cit->first == key;) {
-        close_sources.push_back(cit->second);
-        close_paren_map_.erase(cit++);
-      }
-      std::sort(close_sources.begin(), close_sources.end());
-      auto unique_end = std::unique(close_sources.begin(), close_sources.end());
-      close_sources.resize(unique_end - close_sources.begin());
-      if (!close_sources.empty()) {
-        close_source_map_[key] = close_source_sets_.FindId(close_sources);
-      }
-      open_paren_map_.erase(oit++);
-    }
-  }
-
-  // Returns a new balance data object representing the reversed balance
-  // information.
-  PdtBalanceData<Arc> *Reverse(StateId num_states, StateId num_split,
-                               StateId state_id_shift) const;
-
- private:
-  // Open paren at destintation state?
-  OpenParenSet open_paren_set_;
-  // Open parens per state.
-  OpenParenMap open_paren_map_;
-  // Current open destination state.
-  State open_dest_;
-  // Current open paren/state.
-  typename OpenParenMap::const_iterator open_iter_;
-  // Close states to (open paren, state).
-  CloseParenMap close_paren_map_;
-  // (Paren, state) to set ID.
-  CloseSourceMap close_source_map_;
-  mutable Collection<ssize_t, StateId> close_source_sets_;
-};
-
-// Return a new balance data object representing the reversed balance
-// information.
-template <class Arc>
-PdtBalanceData<Arc> *PdtBalanceData<Arc>::Reverse(
-    StateId num_states, StateId num_split, StateId state_id_shift) const {
-  auto *bd = new PdtBalanceData<Arc>;
-  std::unordered_set<StateId> close_sources;
-  const auto split_size = num_states / num_split;
-  for (StateId i = 0; i < num_states; i += split_size) {
-    close_sources.clear();
-    for (auto it = close_source_map_.begin(); it != close_source_map_.end();
-         ++it) {
-      const auto &okey = it->first;
-      const auto open_dest = okey.state_id;
-      const auto paren_id = okey.paren_id;
-      for (auto set_iter = close_source_sets_.FindSet(it->second);
-           !set_iter.Done(); set_iter.Next()) {
-        const auto close_source = set_iter.Element();
-        if ((close_source < i) || (close_source >= i + split_size)) continue;
-        close_sources.insert(close_source + state_id_shift);
-        bd->OpenInsert(paren_id, close_source + state_id_shift);
-        bd->CloseInsert(paren_id, close_source + state_id_shift,
-                        open_dest + state_id_shift);
-      }
-    }
-    for (auto it = close_sources.begin(); it != close_sources.end(); ++it) {
-      bd->FinishInsert(*it);
-    }
-  }
-  return bd;
-}
-
-}  // namespace internal
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_PAREN_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/pdt.h b/projects/llm_framework/include/fst/extensions/pdt/pdt.h
deleted file mode 100644
index d6c5946f..00000000
--- a/projects/llm_framework/include/fst/extensions/pdt/pdt.h
+++ /dev/null
@@ -1,165 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Common classes for PDT expansion/traversal.
-
-#ifndef FST_EXTENSIONS_PDT_PDT_H_
-#define FST_EXTENSIONS_PDT_PDT_H_
-
-#include <map>
-#include <set>
-#include <unordered_map>
-
-#include <fst/compat.h>
-#include <fst/log.h>
-#include <fst/fst.h>
-#include <fst/state-table.h>
-
-namespace fst {
-
-// Provides bijection between parenthesis stacks and signed integral stack IDs.
-// Each stack ID is unique to each distinct stack. The open-close parenthesis
-// label pairs are passed using the parens argument.
-template <typename StackId, typename Label>
-class PdtStack {
- public:
-  // The stacks are stored in a tree. The nodes are stored in a vector. Each
-  // node represents the top of some stack and is identified by its position in
-  // the vector. Its' parent node represents the stack with the top popped and
-  // its children are stored in child_map_ and accessed by stack_id and label.
-  // The paren_id is
-  // the position in parens of the parenthesis for that node.
-  struct StackNode {
-    StackId parent_id;
-    size_t paren_id;
-
-    StackNode(StackId p, size_t i) : parent_id(p), paren_id(i) {}
-  };
-
-  explicit PdtStack(const std::vector<std::pair<Label, Label>> &parens)
-      : parens_(parens), min_paren_(kNoLabel), max_paren_(kNoLabel) {
-    for (size_t i = 0; i < parens.size(); ++i) {
-      const auto &pair = parens[i];
-      paren_map_[pair.first] = i;
-      paren_map_[pair.second] = i;
-      if (min_paren_ == kNoLabel || pair.first < min_paren_) {
-        min_paren_ = pair.first;
-      }
-      if (pair.second < min_paren_) min_paren_ = pair.second;
-      if (max_paren_ == kNoLabel || pair.first > max_paren_) {
-        max_paren_ = pair.first;
-      }
-      if (pair.second > max_paren_) max_paren_ = pair.second;
-    }
-    nodes_.push_back(StackNode(-1, -1));  // Tree root.
-  }
-
-  // Returns stack ID given the current stack ID (0 if empty) and label read.
-  // Pushes onto the stack if the label is an open parenthesis, returning the
-  // new stack ID. Pops the stack if the label is a close parenthesis that
-  // matches the top of the stack, returning the parent stack ID. Returns -1 if
-  // label is an unmatched close parenthesis. Otherwise, returns the current
-  // stack ID.
-  StackId Find(StackId stack_id, Label label) {
-    if (min_paren_ == kNoLabel || label < min_paren_ || label > max_paren_) {
-      return stack_id;  // Non-paren.
-    }
-    const auto it = paren_map_.find(label);
-    // Non-paren.
-    if (it == paren_map_.end()) return stack_id;
-    const auto paren_id = it->second;
-    // Open paren.
-    if (label == parens_[paren_id].first) {
-      auto &child_id = child_map_[std::make_pair(stack_id, label)];
-      if (child_id == 0) {  // Child not found; pushes label.
-        child_id = nodes_.size();
-        nodes_.push_back(StackNode(stack_id, paren_id));
-      }
-      return child_id;
-    }
-    const auto &node = nodes_[stack_id];
-    // Matching close paren.
-    if (paren_id == node.paren_id) return node.parent_id;
-    // Non-matching close paren.
-    return -1;
-  }
-
-  // Returns the stack ID obtained by popping the label at the top of the
-  // current stack ID.
-  StackId Pop(StackId stack_id) const { return nodes_[stack_id].parent_id; }
-
-  // Returns the paren ID at the top of the stack.
-  ssize_t Top(StackId stack_id) const { return nodes_[stack_id].paren_id; }
-
-  ssize_t ParenId(Label label) const {
-    const auto it = paren_map_.find(label);
-    if (it == paren_map_.end()) return -1;  // Non-paren.
-    return it->second;
-  }
-
- private:
-  struct ChildHash {
-    size_t operator()(const std::pair<StackId, Label> &pair) const {
-      static constexpr size_t prime = 7853;
-      return static_cast<size_t>(pair.first) +
-             static_cast<size_t>(pair.second) * prime;
-    }
-  };
-
-  std::vector<std::pair<Label, Label>> parens_;
-  std::vector<StackNode> nodes_;
-  std::unordered_map<Label, size_t> paren_map_;
-  // Child of stack node w.r.t label
-  std::unordered_map<std::pair<StackId, Label>, StackId, ChildHash> child_map_;
-  Label min_paren_;
-  Label max_paren_;
-};
-
-// State tuple for PDT expansion.
-template <typename S, typename K>
-struct PdtStateTuple {
-  using StateId = S;
-  using StackId = K;
-
-  StateId state_id;
-  StackId stack_id;
-
-  PdtStateTuple(StateId state_id = kNoStateId, StackId stack_id = -1)
-      : state_id(state_id), stack_id(stack_id) {}
-};
-
-// Equality of PDT state tuples.
-template <typename S, typename K>
-inline bool operator==(const PdtStateTuple<S, K> &x,
-                       const PdtStateTuple<S, K> &y) {
-  if (&x == &y) return true;
-  return x.state_id == y.state_id && x.stack_id == y.stack_id;
-}
-
-// Hash function object for PDT state tuples
-template <class T>
-class PdtStateHash {
- public:
-  size_t operator()(const T &tuple) const {
-    static constexpr auto prime = 7853;
-    return tuple.state_id + tuple.stack_id * prime;
-  }
-};
-
-// Tuple to PDT state bijection.
-template <class StateId, class StackId>
-class PdtStateTable : public CompactHashStateTable<
-                          PdtStateTuple<StateId, StackId>,
-                          PdtStateHash<PdtStateTuple<StateId, StackId>>> {
- public:
-  PdtStateTable() {}
-
-  PdtStateTable(const PdtStateTable &other) {}
-
- private:
-  PdtStateTable &operator=(const PdtStateTable &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_PDT_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/pdtlib.h b/projects/llm_framework/include/fst/extensions/pdt/pdtlib.h
deleted file mode 100644
index 41d98986..00000000
--- a/projects/llm_framework/include/fst/extensions/pdt/pdtlib.h
+++ /dev/null
@@ -1,19 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// This is an experimental push-down transducer (PDT) library. A PDT is
-// encoded as an FST, where some transitions are labeled with open or close
-// parentheses. To be interpreted as a PDT, the parentheses must balance on a
-// path.
-
-#ifndef FST_EXTENSIONS_PDT_PDTLIB_H_
-#define FST_EXTENSIONS_PDT_PDTLIB_H_
-
-#include <fst/extensions/pdt/compose.h>
-#include <fst/extensions/pdt/expand.h>
-#include <fst/extensions/pdt/pdt.h>
-#include <fst/extensions/pdt/replace.h>
-#include <fst/extensions/pdt/reverse.h>
-#include <fst/extensions/pdt/shortest-path.h>
-
-#endif  // FST_EXTENSIONS_PDT_PDTLIB_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/pdtscript.h b/projects/llm_framework/include/fst/extensions/pdt/pdtscript.h
deleted file mode 100644
index bdf5b4ea..00000000
--- a/projects/llm_framework/include/fst/extensions/pdt/pdtscript.h
+++ /dev/null
@@ -1,244 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Convenience file for including all PDT operations at once, and/or
-// registering them for new arc types.
-
-#ifndef FST_EXTENSIONS_PDT_PDTSCRIPT_H_
-#define FST_EXTENSIONS_PDT_PDTSCRIPT_H_
-
-#include <algorithm>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-#include <fst/compose.h>  // for ComposeOptions
-#include <fst/util.h>
-
-#include <fst/script/arg-packs.h>
-#include <fst/script/fstscript.h>
-#include <fst/script/shortest-path.h>
-
-#include <fst/extensions/pdt/compose.h>
-#include <fst/extensions/pdt/expand.h>
-#include <fst/extensions/pdt/info.h>
-#include <fst/extensions/pdt/replace.h>
-#include <fst/extensions/pdt/reverse.h>
-#include <fst/extensions/pdt/shortest-path.h>
-
-namespace fst {
-namespace script {
-
-using PdtComposeArgs =
-    std::tuple<const FstClass &, const FstClass &,
-               const std::vector<LabelPair> &, MutableFstClass *,
-               const PdtComposeOptions &, bool>;
-
-template <class Arc>
-void PdtCompose(PdtComposeArgs *args) {
-  const Fst<Arc> &ifst1 = *(std::get<0>(*args).GetFst<Arc>());
-  const Fst<Arc> &ifst2 = *(std::get<1>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<3>(*args)->GetMutableFst<Arc>();
-  // In case Arc::Label is not the same as FstClass::Label, we make a
-  // copy. Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<2>(*args).size());
-  std::copy(std::get<2>(*args).begin(), std::get<2>(*args).end(),
-            typed_parens.begin());
-  if (std::get<5>(*args)) {
-    Compose(ifst1, typed_parens, ifst2, ofst, std::get<4>(*args));
-  } else {
-    Compose(ifst1, ifst2, typed_parens, ofst, std::get<4>(*args));
-  }
-}
-
-void PdtCompose(const FstClass &ifst1, const FstClass &ifst2,
-                const std::vector<LabelPair> &parens,
-                MutableFstClass *ofst, const PdtComposeOptions &opts,
-                bool left_pdt);
-
-struct PdtExpandOptions {
-  bool connect;
-  bool keep_parentheses;
-  const WeightClass &weight_threshold;
-
-  PdtExpandOptions(bool c, bool k, const WeightClass &w)
-      : connect(c), keep_parentheses(k), weight_threshold(w) {}
-};
-
-using PdtExpandArgs =
-    std::tuple<const FstClass &, const std::vector<LabelPair> &,
-               MutableFstClass *, const PdtExpandOptions &>;
-
-template <class Arc>
-void PdtExpand(PdtExpandArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
-  // In case Arc::Label is not the same as FstClass::Label, we make a
-  // copy. Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_parens.begin());
-  Expand(fst, typed_parens, ofst,
-         fst::PdtExpandOptions<Arc>(
-             std::get<3>(*args).connect, std::get<3>(*args).keep_parentheses,
-             *(std::get<3>(*args)
-                   .weight_threshold.GetWeight<typename Arc::Weight>())));
-}
-
-void PdtExpand(const FstClass &ifst, const std::vector<LabelPair> &parens,
-               MutableFstClass *ofst, const PdtExpandOptions &opts);
-
-void PdtExpand(const FstClass &ifst, const std::vector<LabelPair> &parens,
-               MutableFstClass *ofst, bool connect, bool keep_parentheses,
-               const WeightClass &weight_threshold);
-
-using PdtReplaceArgs =
-    std::tuple<const std::vector<LabelFstClassPair> &, MutableFstClass *,
-               std::vector<LabelPair> *, int64, PdtParserType, int64,
-               const string &, const string &>;
-
-template <class Arc>
-void PdtReplace(PdtReplaceArgs *args) {
-  const auto &untyped_pairs = std::get<0>(*args);
-  auto size = untyped_pairs.size();
-  std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>> typed_pairs(
-      size);
-  for (size_t i = 0; i < size; ++i) {
-    typed_pairs[i].first = untyped_pairs[i].first;
-    typed_pairs[i].second = untyped_pairs[i].second->GetFst<Arc>();
-  }
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens;
-  const PdtReplaceOptions<Arc> opts(std::get<3>(*args), std::get<4>(*args),
-                                    std::get<5>(*args), std::get<6>(*args),
-                                    std::get<7>(*args));
-  Replace(typed_pairs, ofst, &typed_parens, opts);
-  // Copies typed parens into arg3.
-  std::get<2>(*args)->resize(typed_parens.size());
-  std::copy(typed_parens.begin(), typed_parens.end(),
-            std::get<2>(*args)->begin());
-}
-
-void PdtReplace(const std::vector<LabelFstClassPair> &pairs,
-                MutableFstClass *ofst, std::vector<LabelPair> *parens,
-                int64 root, PdtParserType parser_type = PDT_LEFT_PARSER,
-                int64 start_paren_labels = kNoLabel,
-                const string &left_paren_prefix = "(_",
-                const string &right_paren_prefix = "_)");
-
-using PdtReverseArgs =
-    std::tuple<const FstClass &, const std::vector<LabelPair> &,
-               MutableFstClass *>;
-
-template <class Arc>
-void PdtReverse(PdtReverseArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
-  // In case Arc::Label is not the same as FstClass::Label, we make a
-  // copy. Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_parens.begin());
-  Reverse(fst, typed_parens, ofst);
-}
-
-void PdtReverse(const FstClass &ifst, const std::vector<LabelPair> &,
-                MutableFstClass *ofst);
-
-// PDT SHORTESTPATH
-
-struct PdtShortestPathOptions {
-  QueueType queue_type;
-  bool keep_parentheses;
-  bool path_gc;
-
-  PdtShortestPathOptions(QueueType qt = FIFO_QUEUE, bool kp = false,
-                         bool gc = true)
-      : queue_type(qt), keep_parentheses(kp), path_gc(gc) {}
-};
-
-using PdtShortestPathArgs =
-    std::tuple<const FstClass &, const std::vector<LabelPair> &,
-               MutableFstClass *, const PdtShortestPathOptions &>;
-
-template <class Arc>
-void PdtShortestPath(PdtShortestPathArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
-  const PdtShortestPathOptions &opts = std::get<3>(*args);
-  // In case Arc::Label is not the same as FstClass::Label, we make a
-  // copy. Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_parens.begin());
-  switch (opts.queue_type) {
-    default:
-      FSTERROR() << "Unknown queue type: " << opts.queue_type;
-    case FIFO_QUEUE: {
-      using Queue = FifoQueue<typename Arc::StateId>;
-      fst::PdtShortestPathOptions<Arc, Queue> spopts(opts.keep_parentheses,
-                                                         opts.path_gc);
-      ShortestPath(fst, typed_parens, ofst, spopts);
-      return;
-    }
-    case LIFO_QUEUE: {
-      using Queue = LifoQueue<typename Arc::StateId>;
-      fst::PdtShortestPathOptions<Arc, Queue> spopts(opts.keep_parentheses,
-                                                         opts.path_gc);
-      ShortestPath(fst, typed_parens, ofst, spopts);
-      return;
-    }
-    case STATE_ORDER_QUEUE: {
-      using Queue = StateOrderQueue<typename Arc::StateId>;
-      fst::PdtShortestPathOptions<Arc, Queue> spopts(opts.keep_parentheses,
-                                                         opts.path_gc);
-      ShortestPath(fst, typed_parens, ofst, spopts);
-      return;
-    }
-  }
-}
-
-void PdtShortestPath(const FstClass &ifst,
-    const std::vector<LabelPair> &parens, MutableFstClass *ofst,
-    const PdtShortestPathOptions &opts = PdtShortestPathOptions());
-
-// PRINT INFO
-
-using PrintPdtInfoArgs =
-    std::pair<const FstClass &, const std::vector<LabelPair> &>;
-
-template <class Arc>
-void PrintPdtInfo(PrintPdtInfoArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  // In case Arc::Label is not the same as FstClass::Label, we make a
-  // copy. Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_parens.begin());
-  PdtInfo<Arc> pdtinfo(fst, typed_parens);
-  PrintPdtInfo(pdtinfo);
-}
-
-void PrintPdtInfo(const FstClass &ifst, const std::vector<LabelPair> &parens);
-
-}  // namespace script
-}  // namespace fst
-
-#define REGISTER_FST_PDT_OPERATIONS(ArcType)                             \
-  REGISTER_FST_OPERATION(PdtCompose, ArcType, PdtComposeArgs);           \
-  REGISTER_FST_OPERATION(PdtExpand, ArcType, PdtExpandArgs);             \
-  REGISTER_FST_OPERATION(PdtReplace, ArcType, PdtReplaceArgs);           \
-  REGISTER_FST_OPERATION(PdtReverse, ArcType, PdtReverseArgs);           \
-  REGISTER_FST_OPERATION(PdtShortestPath, ArcType, PdtShortestPathArgs); \
-  REGISTER_FST_OPERATION(PrintPdtInfo, ArcType, PrintPdtInfoArgs)
-#endif  // FST_EXTENSIONS_PDT_PDTSCRIPT_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/replace.h b/projects/llm_framework/include/fst/extensions/pdt/replace.h
deleted file mode 100644
index 76341e8d..00000000
--- a/projects/llm_framework/include/fst/extensions/pdt/replace.h
+++ /dev/null
@@ -1,827 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Recursively replaces FST arcs with other FSTs, returning a PDT.
-
-#ifndef FST_EXTENSIONS_PDT_REPLACE_H_
-#define FST_EXTENSIONS_PDT_REPLACE_H_
-
-#include <map>
-#include <memory>
-#include <set>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/replace.h>
-#include <fst/replace-util.h>
-#include <fst/symbol-table-ops.h>
-
-namespace fst {
-namespace internal {
-
-// Hash to paren IDs
-template <typename S>
-struct ReplaceParenHash {
-  size_t operator()(const std::pair<size_t, S> &paren) const {
-    static constexpr auto prime = 7853;
-    return paren.first + paren.second * prime;
-  }
-};
-
-}  // namespace internal
-
-// Parser types characterize the PDT construction method. When applied to a CFG,
-// each non-terminal is encoded as a DFA that accepts precisely the RHS's of
-// productions of that non-terminal. For parsing (rather than just recognition),
-// production numbers can used as outputs (placed as early as possible) in the
-// DFAs promoted to DFTs. For more information on the strongly regular
-// construction, see:
-//
-// Mohri, M., and Pereira, F. 1998. Dynamic compilation of weighted context-free
-// grammars. In Proc. ACL, pages 891-897.
-enum PdtParserType {
-  // Top-down construction. Applied to a simple LL(1) grammar (among others),
-  // gives a DPDA. If promoted to a DPDT, with outputs being production
-  // numbers, gives a leftmost derivation. Left recursive grammars are
-  // problematic in use.
-  PDT_LEFT_PARSER,
-
-  // Top-down construction. Similar to PDT_LEFT_PARSE except bounded-stack
-  // (expandable as an FST) result with regular or, more generally, strongly
-  // regular grammars. Epsilons may replace some parentheses, which may
-  // introduce some non-determinism.
-  PDT_LEFT_SR_PARSER,
-
-  /* TODO(riley):
-  // Bottom-up construction. Applied to a LR(0) grammar, gives a DPDA.
-  // If promoted to a DPDT, with outputs being the production nubmers,
-  // gives the reverse of a rightmost derivation.
-  PDT_RIGHT_PARSER,
-  */
-};
-
-template <class Arc>
-struct PdtReplaceOptions {
-  using Label = typename Arc::Label;
-
-  explicit PdtReplaceOptions(Label root,
-                             PdtParserType type = PDT_LEFT_PARSER,
-                             Label start_paren_labels = kNoLabel,
-                             string left_paren_prefix = "(_",
-                             string right_paren_prefix = ")_") :
-      root(root), type(type), start_paren_labels(start_paren_labels),
-      left_paren_prefix(std::move(left_paren_prefix)),
-      right_paren_prefix(std::move(right_paren_prefix)) {}
-
-  Label root;
-  PdtParserType type;
-  Label start_paren_labels;
-  const string left_paren_prefix;
-  const string right_paren_prefix;
-};
-
-// PdtParser: Base PDT parser class common to specific parsers.
-
-template <class Arc>
-class PdtParser {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using LabelFstPair = std::pair<Label, const Fst<Arc> *>;
-  using LabelPair = std::pair<Label, Label>;
-  using LabelStatePair = std::pair<Label, StateId>;
-  using StateWeightPair = std::pair<StateId, Weight>;
-  using ParenKey = std::pair<size_t, StateId>;
-  using ParenMap =
-      std::unordered_map<ParenKey, size_t, internal::ReplaceParenHash<StateId>>;
-
-  PdtParser(const std::vector<LabelFstPair> &fst_array,
-            const PdtReplaceOptions<Arc> &opts) :
-      root_(opts.root), start_paren_labels_(opts.start_paren_labels),
-      left_paren_prefix_(std::move(opts.left_paren_prefix)),
-      right_paren_prefix_(std::move(opts.right_paren_prefix)),
-      error_(false) {
-    for (size_t i = 0; i < fst_array.size(); ++i) {
-      if (!CompatSymbols(fst_array[0].second->InputSymbols(),
-                         fst_array[i].second->InputSymbols())) {
-        FSTERROR() << "PdtParser: Input symbol table of input FST " << i
-                   << " does not match input symbol table of 0th input FST";
-        error_ = true;
-      }
-      if (!CompatSymbols(fst_array[0].second->OutputSymbols(),
-                         fst_array[i].second->OutputSymbols())) {
-        FSTERROR() << "PdtParser: Output symbol table of input FST " << i
-                   << " does not match input symbol table of 0th input FST";
-        error_ = true;
-      }
-      fst_array_.emplace_back(fst_array[i].first, fst_array[i].second->Copy());
-      // Builds map from non-terminal label to FST ID.
-      label2id_[fst_array[i].first] = i;
-    }
-  }
-
-  virtual ~PdtParser() {
-    for (auto &pair : fst_array_) delete pair.second;
-  }
-
-  // Constructs the output PDT, dependent on the derived parser type.
-  virtual void GetParser(MutableFst<Arc> *ofst,
-                         std::vector<LabelPair> *parens) = 0;
-
- protected:
-  const std::vector<LabelFstPair> &FstArray() const { return fst_array_; }
-
-  Label Root() const { return root_; }
-
-  // Maps from non-terminal label to corresponding FST ID, or returns
-  // kNoStateId to signal lookup failure.
-  StateId Label2Id(Label l) const {
-    auto it = label2id_.find(l);
-    return it == label2id_.end() ? kNoStateId : it->second;
-  }
-
-  // Maps from output state to input FST label, state pair, or returns a
-  // (kNoLabel, kNoStateId) pair to signal lookup failure.
-  LabelStatePair GetLabelStatePair(StateId os) const {
-    if (os >= label_state_pairs_.size()) {
-      static const LabelStatePair no_pair(kNoLabel, kNoLabel);
-      return no_pair;
-    } else {
-      return label_state_pairs_[os];
-    }
-  }
-
-  // Maps to output state from input FST (label, state) pair, or returns
-  // kNoStateId to signal lookup failure.
-  StateId GetState(const LabelStatePair &lsp) const {
-    auto it = state_map_.find(lsp);
-    if (it == state_map_.end()) {
-      return kNoStateId;
-    } else {
-      return it->second;
-    }
-  }
-
-  // Builds single FST combining all referenced input FSTs, leaving in the
-  // non-termnals for now; also tabulates the PDT states that correspond to the
-  // start and final states of the input FSTs.
-  void CreateFst(MutableFst<Arc> *ofst, std::vector<StateId> *open_dest,
-                 std::vector<std::vector<StateWeightPair>> *close_src);
-
-  // Assigns parenthesis labels from total allocated paren IDs.
-  void AssignParenLabels(size_t total_nparens, std::vector<LabelPair> *parens) {
-    parens->clear();
-    for (size_t paren_id = 0; paren_id < total_nparens; ++paren_id) {
-      const auto open_paren = start_paren_labels_ + paren_id;
-      const auto close_paren = open_paren + total_nparens;
-      parens->emplace_back(open_paren, close_paren);
-    }
-  }
-
-  // Determines how non-terminal instances are assigned parentheses IDs.
-  virtual size_t AssignParenIds(const Fst<Arc> &ofst,
-                                ParenMap *paren_map) const = 0;
-
-  // Changes a non-terminal transition to an open parenthesis transition
-  // redirected to the PDT state specified in the open_dest argument, when
-  // indexed by the input FST ID for the non-terminal. Adds close parenthesis
-  // transitions (with specified weights) from the PDT states specified in the
-  // close_src argument, when indexed by the input FST ID for the non-terminal,
-  // to the former destination state of the non-terminal transition. The
-  // paren_map argument gives the parenthesis ID for a given non-terminal FST ID
-  // and destination state pair. The close_non_term_weight vector specifies
-  // non-terminals for which the non-terminal arc weight should be applied on
-  // the close parenthesis (multiplying the close_src weight above) rather than
-  // on the open parenthesis. If no paren ID is found, then an epsilon replaces
-  // the parenthesis that would carry the non-terminal arc weight and the other
-  // parenthesis is omitted (appropriate for the strongly-regular case).
-  void AddParensToFst(
-      const std::vector<LabelPair> &parens,
-      const ParenMap &paren_map,
-      const std::vector<StateId> &open_dest,
-      const std::vector<std::vector<StateWeightPair>> &close_src,
-      const std::vector<bool> &close_non_term_weight,
-      MutableFst<Arc> *ofst);
-
-  // Ensures that parentheses arcs are added to the symbol table.
-  void AddParensToSymbolTables(const std::vector<LabelPair> &parens,
-                               MutableFst<Arc> *ofst);
-
- private:
-  std::vector<LabelFstPair> fst_array_;
-  Label root_;
-  // Index to use for the first parenthesis.
-  Label start_paren_labels_;
-  const string left_paren_prefix_;
-  const string right_paren_prefix_;
-  // Maps from non-terminal label to FST ID.
-  std::unordered_map<Label, StateId> label2id_;
-  // Given an output state, specifies the input FST (label, state) pair.
-  std::vector<LabelStatePair> label_state_pairs_;
-  // Given an FST (label, state) pair, specifies the output FST state ID.
-  std::map<LabelStatePair, StateId> state_map_;
-  bool error_;
-};
-
-template <class Arc>
-void PdtParser<Arc>::CreateFst(
-    MutableFst<Arc> *ofst, std::vector<StateId> *open_dest,
-    std::vector<std::vector<StateWeightPair>> *close_src) {
-  ofst->DeleteStates();
-  if (error_) {
-    ofst->SetProperties(kError, kError);
-    return;
-  }
-  open_dest->resize(fst_array_.size(), kNoStateId);
-  close_src->resize(fst_array_.size());
-  // Queue of non-terminals to replace.
-  std::deque<Label> non_term_queue;
-  non_term_queue.push_back(root_);
-  // Has a non-terminal been enqueued?
-  std::vector<bool> enqueued(fst_array_.size(), false);
-  enqueued[label2id_[root_]] = true;
-  Label max_label = kNoLabel;
-  for (StateId soff = 0; !non_term_queue.empty(); soff = ofst->NumStates()) {
-    const auto label = non_term_queue.front();
-    non_term_queue.pop_front();
-    StateId fst_id = Label2Id(label);
-    const auto *ifst = fst_array_[fst_id].second;
-    for (StateIterator<Fst<Arc>> siter(*ifst); !siter.Done(); siter.Next()) {
-      const auto is = siter.Value();
-      const auto os = ofst->AddState();
-      const LabelStatePair lsp(label, is);
-      label_state_pairs_.push_back(lsp);
-      state_map_[lsp] = os;
-      if (is == ifst->Start()) {
-        (*open_dest)[fst_id] = os;
-        if (label == root_) ofst->SetStart(os);
-      }
-      if (ifst->Final(is) != Weight::Zero()) {
-        if (label == root_) ofst->SetFinal(os, ifst->Final(is));
-        (*close_src)[fst_id].emplace_back(os, ifst->Final(is));
-      }
-      for (ArcIterator<Fst<Arc>> aiter(*ifst, is); !aiter.Done();
-           aiter.Next()) {
-        auto arc = aiter.Value();
-        arc.nextstate += soff;
-        if (max_label == kNoLabel || arc.olabel > max_label)
-          max_label = arc.olabel;
-        const auto nfst_id = Label2Id(arc.olabel);
-        if (nfst_id != kNoStateId) {
-          if (fst_array_[nfst_id].second->Start() == kNoStateId) continue;
-          if (!enqueued[nfst_id]) {
-            non_term_queue.push_back(arc.olabel);
-            enqueued[nfst_id] = true;
-          }
-        }
-        ofst->AddArc(os, arc);
-      }
-    }
-  }
-  if (start_paren_labels_ == kNoLabel)
-    start_paren_labels_ = max_label + 1;
-}
-
-template <class Arc>
-void PdtParser<Arc>::AddParensToFst(
-    const std::vector<LabelPair> &parens,
-    const ParenMap &paren_map,
-    const std::vector<StateId> &open_dest,
-    const std::vector<std::vector<StateWeightPair>> &close_src,
-    const std::vector<bool> &close_non_term_weight,
-    MutableFst<Arc> *ofst) {
-  StateId dead_state = kNoStateId;
-  using MIter = MutableArcIterator<MutableFst<Arc>>;
-  for (StateIterator<Fst<Arc>> siter(*ofst); !siter.Done(); siter.Next()) {
-    StateId os = siter.Value();
-    std::unique_ptr<MIter> aiter(new MIter(ofst, os));
-    for (auto n = 0; !aiter->Done(); aiter->Next(), ++n) {
-      const auto arc = aiter->Value();  // A reference here may go stale.
-      StateId nfst_id = Label2Id(arc.olabel);
-      if (nfst_id != kNoStateId) {
-        // Gets parentheses.
-        const ParenKey paren_key(nfst_id, arc.nextstate);
-        auto it = paren_map.find(paren_key);
-        Label open_paren = 0;
-        Label close_paren = 0;
-        if (it != paren_map.end()) {
-          const auto paren_id = it->second;
-          open_paren = parens[paren_id].first;
-          close_paren = parens[paren_id].second;
-        }
-        // Sets open parenthesis.
-        if (open_paren != 0 || !close_non_term_weight[nfst_id]) {
-          const auto open_weight =
-              close_non_term_weight[nfst_id] ? Weight::One() : arc.weight;
-          const Arc sarc(open_paren, open_paren, open_weight,
-                         open_dest[nfst_id]);
-          aiter->SetValue(sarc);
-        } else {
-          if (dead_state == kNoStateId) {
-            dead_state = ofst->AddState();
-          }
-          const Arc sarc(0, 0, Weight::One(), dead_state);
-          aiter->SetValue(sarc);
-        }
-        // Adds close parentheses.
-        if (close_paren != 0 || close_non_term_weight[nfst_id]) {
-          for (size_t i = 0; i < close_src[nfst_id].size(); ++i) {
-            const auto &pair = close_src[nfst_id][i];
-            const auto close_weight = close_non_term_weight[nfst_id]
-                                          ? Times(arc.weight, pair.second)
-                                          : pair.second;
-            const Arc farc(close_paren, close_paren, close_weight,
-                           arc.nextstate);
-
-            ofst->AddArc(pair.first, farc);
-            if (os == pair.first) {  // Invalidated iterator.
-              aiter.reset(new MIter(ofst, os));
-              aiter->Seek(n);
-            }
-          }
-        }
-      }
-    }
-  }
-}
-
-template <class Arc>
-void PdtParser<Arc>::AddParensToSymbolTables(
-    const std::vector<LabelPair> &parens, MutableFst<Arc> *ofst) {
-  auto size = parens.size();
-  if (ofst->InputSymbols()) {
-    if (!AddAuxiliarySymbols(left_paren_prefix_, start_paren_labels_, size,
-                             ofst->MutableInputSymbols())) {
-      ofst->SetProperties(kError, kError);
-      return;
-    }
-    if (!AddAuxiliarySymbols(right_paren_prefix_, start_paren_labels_ + size,
-                             size, ofst->MutableInputSymbols())) {
-      ofst->SetProperties(kError, kError);
-      return;
-    }
-  }
-  if (ofst->OutputSymbols()) {
-    if (!AddAuxiliarySymbols(left_paren_prefix_, start_paren_labels_, size,
-                             ofst->MutableOutputSymbols())) {
-      ofst->SetProperties(kError, kError);
-      return;
-    }
-    if (!AddAuxiliarySymbols(right_paren_prefix_, start_paren_labels_ + size,
-                             size, ofst->MutableOutputSymbols())) {
-      ofst->SetProperties(kError, kError);
-      return;
-    }
-  }
-}
-
-// Builds a PDT by recursive replacement top-down, where the call and return are
-// encoded in the parentheses.
-template <class Arc>
-class PdtLeftParser final : public PdtParser<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using LabelFstPair = typename PdtParser<Arc>::LabelFstPair;
-  using LabelPair = typename PdtParser<Arc>::LabelPair;
-  using LabelStatePair = typename PdtParser<Arc>::LabelStatePair;
-  using StateWeightPair = typename PdtParser<Arc>::StateWeightPair;
-  using ParenKey = typename PdtParser<Arc>::ParenKey;
-  using ParenMap = typename PdtParser<Arc>::ParenMap;
-
-  using PdtParser<Arc>::AddParensToFst;
-  using PdtParser<Arc>::AddParensToSymbolTables;
-  using PdtParser<Arc>::AssignParenLabels;
-  using PdtParser<Arc>::CreateFst;
-  using PdtParser<Arc>::FstArray;
-  using PdtParser<Arc>::GetLabelStatePair;
-  using PdtParser<Arc>::GetState;
-  using PdtParser<Arc>::Label2Id;
-  using PdtParser<Arc>::Root;
-
-  PdtLeftParser(const std::vector<LabelFstPair> &fst_array,
-                const PdtReplaceOptions<Arc> &opts) :
-      PdtParser<Arc>(fst_array, opts) { }
-
-  void GetParser(MutableFst<Arc> *ofst,
-                 std::vector<LabelPair> *parens) override;
-
- protected:
-  // Assigns a unique parenthesis ID for each non-terminal, destination
-  // state pair.
-  size_t AssignParenIds(const Fst<Arc> &ofst,
-                        ParenMap *paren_map) const override;
-};
-
-template <class Arc>
-void PdtLeftParser<Arc>::GetParser(
-    MutableFst<Arc> *ofst,
-    std::vector<LabelPair> *parens) {
-  ofst->DeleteStates();
-  parens->clear();
-  const auto &fst_array = FstArray();
-  // Map that gives the paren ID for a (non-terminal, dest. state) pair
-  // (which can be unique).
-  ParenMap paren_map;
-  // Specifies the open parenthesis destination state for a given non-terminal.
-  // The source is the non-terminal instance source state.
-  std::vector<StateId> open_dest(fst_array.size(), kNoStateId);
-  // Specifies close parenthesis source states and weights for a given
-  // non-terminal. The destination is the non-terminal instance destination
-  // state.
-  std::vector<std::vector<StateWeightPair>> close_src(fst_array.size());
-  // Specifies non-terminals for which the non-terminal arc weight
-  // should be applied on the close parenthesis (multiplying the
-  // 'close_src' weight above) rather than on the open parenthesis.
-  std::vector<bool> close_non_term_weight(fst_array.size(), false);
-  CreateFst(ofst, &open_dest, &close_src);
-  auto total_nparens = AssignParenIds(*ofst, &paren_map);
-  AssignParenLabels(total_nparens, parens);
-  AddParensToFst(*parens, paren_map, open_dest, close_src,
-                 close_non_term_weight, ofst);
-  if (!fst_array.empty()) {
-    ofst->SetInputSymbols(fst_array[0].second->InputSymbols());
-    ofst->SetOutputSymbols(fst_array[0].second->OutputSymbols());
-  }
-  AddParensToSymbolTables(*parens, ofst);
-}
-
-template <class Arc>
-size_t PdtLeftParser<Arc>::AssignParenIds(
-    const Fst<Arc> &ofst,
-    ParenMap *paren_map) const {
-  // Number of distinct parenthesis pairs per FST.
-  std::vector<size_t> nparens(FstArray().size(), 0);
-  // Number of distinct parenthesis pairs overall.
-  size_t total_nparens = 0;
-  for (StateIterator<Fst<Arc>> siter(ofst); !siter.Done(); siter.Next()) {
-    const auto os = siter.Value();
-    for (ArcIterator<Fst<Arc>> aiter(ofst, os); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto nfst_id = Label2Id(arc.olabel);
-      if (nfst_id != kNoStateId) {
-        const ParenKey paren_key(nfst_id, arc.nextstate);
-        auto it = paren_map->find(paren_key);
-        if (it == paren_map->end()) {
-          // Assigns new paren ID for this (FST, dest state) pair.
-          (*paren_map)[paren_key] = nparens[nfst_id]++;
-          if (nparens[nfst_id] > total_nparens)
-            total_nparens = nparens[nfst_id];
-        }
-      }
-    }
-  }
-  return total_nparens;
-}
-
-// Similar to PdtLeftParser but:
-//
-// 1. Uses epsilons rather than parentheses labels for any non-terminal
-//    instances within a left- (right-) linear dependency SCC,
-// 2. Allocates a paren ID uniquely for each such dependency SCC (rather than
-//    non-terminal = dependency state) and destination state.
-template <class Arc>
-class PdtLeftSRParser final : public PdtParser<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using LabelFstPair = typename PdtParser<Arc>::LabelFstPair;
-  using LabelPair = typename PdtParser<Arc>::LabelPair;
-  using LabelStatePair = typename PdtParser<Arc>::LabelStatePair;
-  using StateWeightPair = typename PdtParser<Arc>::StateWeightPair;
-  using ParenKey = typename PdtParser<Arc>::ParenKey;
-  using ParenMap = typename PdtParser<Arc>::ParenMap;
-
-  using PdtParser<Arc>::AddParensToFst;
-  using PdtParser<Arc>::AddParensToSymbolTables;
-  using PdtParser<Arc>::AssignParenLabels;
-  using PdtParser<Arc>::CreateFst;
-  using PdtParser<Arc>::FstArray;
-  using PdtParser<Arc>::GetLabelStatePair;
-  using PdtParser<Arc>::GetState;
-  using PdtParser<Arc>::Label2Id;
-  using PdtParser<Arc>::Root;
-
-  PdtLeftSRParser(const std::vector<LabelFstPair> &fst_array,
-                  const PdtReplaceOptions<Arc> &opts) :
-      PdtParser<Arc>(fst_array, opts),
-      replace_util_(fst_array, ReplaceUtilOptions(opts.root)) { }
-
-  void GetParser(MutableFst<Arc> *ofst,
-                 std::vector<LabelPair> *parens) override;
-
- protected:
-  // Assigns a unique parenthesis ID for each non-terminal, destination state
-  // pair when the non-terminal refers to a non-linear FST. Otherwise, assigns
-  // a unique parenthesis ID for each dependency SCC, destination state pair if
-  // the non-terminal instance is between
-  // SCCs. Otherwise does nothing.
-  size_t AssignParenIds(const Fst<Arc> &ofst,
-                        ParenMap *paren_map) const override;
-
-  // Returns dependency SCC for given label.
-  size_t SCC(Label label) const { return replace_util_.SCC(label); }
-
-  // Is a given dependency SCC left-linear?
-  bool SCCLeftLinear(size_t scc_id) const {
-    const auto ll_props = kReplaceSCCLeftLinear | kReplaceSCCNonTrivial;
-    const auto scc_props = replace_util_.SCCProperties(scc_id);
-    return (scc_props & ll_props) == ll_props;
-  }
-
-  // Is a given dependency SCC right-linear?
-  bool SCCRightLinear(size_t scc_id) const {
-    const auto lr_props = kReplaceSCCRightLinear | kReplaceSCCNonTrivial;
-    const auto scc_props = replace_util_.SCCProperties(scc_id);
-    return (scc_props & lr_props) == lr_props;
-  }
-
-  // Components of left- (right-) linear dependency SCC; empty o.w.
-  const std::vector<size_t> &SCCComps(size_t scc_id) const {
-    if (scc_comps_.empty()) GetSCCComps();
-    return scc_comps_[scc_id];
-  }
-
-  // Returns the representative state of an SCC. For left-linear grammars, it
-  // is one of the initial states. For right-linear grammars, it is one of the
-  // non-terminal destination states; otherwise, it is kNoStateId.
-  StateId RepState(size_t scc_id) const {
-    if (SCCComps(scc_id).empty()) return kNoStateId;
-    const auto fst_id = SCCComps(scc_id).front();
-    const auto &fst_array = FstArray();
-    const auto label = fst_array[fst_id].first;
-    const auto *ifst = fst_array[fst_id].second;
-    if (SCCLeftLinear(scc_id)) {
-      const LabelStatePair lsp(label, ifst->Start());
-      return GetState(lsp);
-    } else {  // Right-linear.
-      const LabelStatePair lsp(label, *NonTermDests(fst_id).begin());
-      return GetState(lsp);
-    }
-    return kNoStateId;
-  }
-
- private:
-  // Merges initial (final) states of in a left- (right-) linear dependency SCC
-  // after dealing with the non-terminal arc and final weights.
-  void ProcSCCs(MutableFst<Arc> *ofst,
-                std::vector<StateId> *open_dest,
-                std::vector<std::vector<StateWeightPair>> *close_src,
-                std::vector<bool> *close_non_term_weight) const;
-
-  // Computes components of left- (right-) linear dependency SCC.
-  void GetSCCComps() const {
-    const std::vector<LabelFstPair> &fst_array = FstArray();
-    for (size_t i = 0; i < fst_array.size(); ++i) {
-      const auto label = fst_array[i].first;
-      const auto scc_id = SCC(label);
-      if (scc_comps_.size() <= scc_id) scc_comps_.resize(scc_id + 1);
-      if (SCCLeftLinear(scc_id) || SCCRightLinear(scc_id)) {
-        scc_comps_[scc_id].push_back(i);
-      }
-    }
-  }
-
-  const std::set<StateId> &NonTermDests(StateId fst_id) const {
-    if (non_term_dests_.empty()) GetNonTermDests();
-    return non_term_dests_[fst_id];
-  }
-
-  // Finds non-terminal destination states for right-linear FSTS, or does
-  // nothing if not found.
-  void GetNonTermDests() const;
-
-  // Dependency SCC info.
-  mutable ReplaceUtil<Arc> replace_util_;
-  // Components of left- (right-) linear dependency SCCs, or empty otherwise.
-  mutable std::vector<std::vector<size_t>> scc_comps_;
-  // States that have non-terminals entering them for each (right-linear) FST.
-  mutable std::vector<std::set<StateId>> non_term_dests_;
-};
-
-template <class Arc>
-void PdtLeftSRParser<Arc>::GetParser(
-    MutableFst<Arc> *ofst,
-    std::vector<LabelPair> *parens) {
-  ofst->DeleteStates();
-  parens->clear();
-  const auto &fst_array = FstArray();
-  // Map that gives the paren ID for a (non-terminal, dest. state) pair.
-  ParenMap paren_map;
-  // Specifies the open parenthesis destination state for a given non-terminal.
-  // The source is the non-terminal instance source state.
-  std::vector<StateId> open_dest(fst_array.size(), kNoStateId);
-  // Specifies close parenthesis source states and weights for a given
-  // non-terminal. The destination is the non-terminal instance destination
-  // state.
-  std::vector<std::vector<StateWeightPair>> close_src(fst_array.size());
-  // Specifies non-terminals for which the non-terminal arc weight should be
-  // applied on the close parenthesis (multiplying the close_src weight above)
-  // rather than on the open parenthesis.
-  std::vector<bool> close_non_term_weight(fst_array.size(), false);
-  CreateFst(ofst, &open_dest, &close_src);
-  ProcSCCs(ofst, &open_dest, &close_src, &close_non_term_weight);
-  const auto total_nparens = AssignParenIds(*ofst, &paren_map);
-  AssignParenLabels(total_nparens, parens);
-  AddParensToFst(*parens, paren_map, open_dest, close_src,
-                 close_non_term_weight, ofst);
-  if (!fst_array.empty()) {
-    ofst->SetInputSymbols(fst_array[0].second->InputSymbols());
-    ofst->SetOutputSymbols(fst_array[0].second->OutputSymbols());
-  }
-  AddParensToSymbolTables(*parens, ofst);
-  Connect(ofst);
-}
-
-template <class Arc>
-void PdtLeftSRParser<Arc>::ProcSCCs(
-    MutableFst<Arc> *ofst,
-    std::vector<StateId> *open_dest,
-    std::vector<std::vector<StateWeightPair>> *close_src,
-    std::vector<bool> *close_non_term_weight) const {
-  const auto &fst_array = FstArray();
-  for (StateIterator<Fst<Arc>> siter(*ofst); !siter.Done(); siter.Next()) {
-    const auto os = siter.Value();
-    const auto label = GetLabelStatePair(os).first;
-    const auto is = GetLabelStatePair(os).second;
-    const auto fst_id = Label2Id(label);
-    const auto scc_id = SCC(label);
-    const auto rs = RepState(scc_id);
-    const auto *ifst = fst_array[fst_id].second;
-    // SCC LEFT-LINEAR: puts non-terminal weights on close parentheses. Merges
-    // initial states into SCC representative state and updates open_dest.
-    if (SCCLeftLinear(scc_id)) {
-      (*close_non_term_weight)[fst_id] = true;
-      if (is == ifst->Start() && os != rs) {
-        for (ArcIterator<Fst<Arc>> aiter(*ofst, os); !aiter.Done();
-             aiter.Next()) {
-          const auto &arc = aiter.Value();
-          ofst->AddArc(rs, arc);
-        }
-        ofst->DeleteArcs(os);
-        if (os == ofst->Start())
-          ofst->SetStart(rs);
-        (*open_dest)[fst_id] = rs;
-      }
-    }
-    // SCC RIGHT-LINEAR: pushes back final weights onto non-terminals, if
-    // possible, or adds weighted epsilons to the SCC representative state.
-    // Merges final states into SCC representative state and updates close_src.
-    if (SCCRightLinear(scc_id)) {
-      for (MutableArcIterator<MutableFst<Arc>> aiter(ofst, os); !aiter.Done();
-           aiter.Next()) {
-        auto arc = aiter.Value();
-        const auto idest = GetLabelStatePair(arc.nextstate).second;
-        if (NonTermDests(fst_id).count(idest) > 0) {
-          if (ofst->Final(arc.nextstate) != Weight::Zero()) {
-            ofst->SetFinal(arc.nextstate, Weight::Zero());
-            ofst->SetFinal(rs, Weight::One());
-          }
-          arc.weight = Times(arc.weight, ifst->Final(idest));
-          arc.nextstate = rs;
-          aiter.SetValue(arc);
-        }
-      }
-      const auto final_weight = ifst->Final(is);
-      if (final_weight != Weight::Zero() &&
-          NonTermDests(fst_id).count(is) == 0) {
-        ofst->AddArc(os, Arc(0, 0, final_weight, rs));
-        if (ofst->Final(os) != Weight::Zero()) {
-          ofst->SetFinal(os, Weight::Zero());
-          ofst->SetFinal(rs, Weight::One());
-        }
-      }
-      if (is == ifst->Start()) {
-        (*close_src)[fst_id].clear();
-        (*close_src)[fst_id].emplace_back(rs, Weight::One());
-      }
-    }
-  }
-}
-
-template <class Arc>
-void PdtLeftSRParser<Arc>::GetNonTermDests() const {
-  const auto &fst_array = FstArray();
-  non_term_dests_.resize(fst_array.size());
-  for (size_t fst_id = 0; fst_id < fst_array.size(); ++fst_id) {
-    const auto label = fst_array[fst_id].first;
-    const auto scc_id = SCC(label);
-    if (SCCRightLinear(scc_id)) {
-      const auto *ifst = fst_array[fst_id].second;
-      for (StateIterator<Fst<Arc>> siter(*ifst); !siter.Done(); siter.Next()) {
-        const auto is = siter.Value();
-        for (ArcIterator<Fst<Arc>> aiter(*ifst, is); !aiter.Done();
-             aiter.Next()) {
-          const auto &arc = aiter.Value();
-          if (Label2Id(arc.olabel) != kNoStateId) {
-            non_term_dests_[fst_id].insert(arc.nextstate);
-          }
-        }
-      }
-    }
-  }
-}
-
-template <class Arc>
-size_t PdtLeftSRParser<Arc>::AssignParenIds(
-    const Fst<Arc> &ofst,
-    ParenMap *paren_map) const {
-  const auto &fst_array = FstArray();
-  // Number of distinct parenthesis pairs per FST.
-  std::vector<size_t> nparens(fst_array.size(), 0);
-  // Number of distinct parenthesis pairs overall.
-  size_t total_nparens = 0;
-  for (StateIterator<Fst<Arc>> siter(ofst); !siter.Done(); siter.Next()) {
-    const auto os = siter.Value();
-    const auto label = GetLabelStatePair(os).first;
-    const auto scc_id = SCC(label);
-    for (ArcIterator<Fst<Arc>> aiter(ofst, os); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto nfst_id = Label2Id(arc.olabel);
-      if (nfst_id != kNoStateId) {
-        size_t nscc_id = SCC(arc.olabel);
-        bool nscc_linear = !SCCComps(nscc_id).empty();
-        // Assigns a parenthesis ID for the non-terminal transition
-        // if the non-terminal belongs to a (left-/right-) linear dependency
-        // SCC or if the transition is in an FST from a different SCC
-        if (!nscc_linear || scc_id != nscc_id) {
-          // For (left-/right-) linear SCCs instead of using nfst_id, we
-          // will use its SCC prototype pfst_id for assigning distinct
-          // parenthesis IDs.
-          const auto pfst_id =
-              nscc_linear ? SCCComps(nscc_id).front() : nfst_id;
-          ParenKey paren_key(pfst_id, arc.nextstate);
-          const auto it = paren_map->find(paren_key);
-          if (it == paren_map->end()) {
-            // Assigns new paren ID for this (FST/SCC, dest. state) pair.
-            if (nscc_linear) {
-              // This is mapping we'll need, but we also store (harmlessly)
-              // for the prototype below so we can easily keep count per SCC.
-              const ParenKey nparen_key(nfst_id, arc.nextstate);
-              (*paren_map)[nparen_key] = nparens[pfst_id];
-            }
-            (*paren_map)[paren_key] = nparens[pfst_id]++;
-            if (nparens[pfst_id] > total_nparens) {
-              total_nparens = nparens[pfst_id];
-            }
-          }
-        }
-      }
-    }
-  }
-  return total_nparens;
-}
-
-// Builds a pushdown transducer (PDT) from an RTN specification. The result is
-// a PDT written to a mutable FST where some transitions are labeled with
-// open or close parentheses. To be interpreted as a PDT, the parens must
-// balance on a path (see PdtExpand()). The open/close parenthesis label pairs
-// are returned in the parens argument.
-template <class Arc>
-void Replace(
-    const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
-        &ifst_array,
-    MutableFst<Arc> *ofst,
-    std::vector<std::pair<typename Arc::Label, typename Arc::Label>> *parens,
-    const PdtReplaceOptions<Arc> &opts) {
-  switch (opts.type) {
-    case PDT_LEFT_PARSER:
-      {
-        PdtLeftParser<Arc> pr(ifst_array, opts);
-        pr.GetParser(ofst, parens);
-        return;
-      }
-    case PDT_LEFT_SR_PARSER:
-      {
-        PdtLeftSRParser<Arc> pr(ifst_array, opts);
-        pr.GetParser(ofst, parens);
-        return;
-      }
-    default:
-      FSTERROR() << "Replace: Unknown PDT parser type: " << opts.type;
-      ofst->DeleteStates();
-      ofst->SetProperties(kError, kError);
-      parens->clear();
-      return;
-  }
-}
-
-// Variant where the only user-controlled arguments is the root ID.
-template <class Arc>
-void Replace(
-    const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
-        &ifst_array,
-    MutableFst<Arc> *ofst,
-    std::vector<std::pair<typename Arc::Label, typename Arc::Label>> *parens,
-    typename Arc::Label root) {
-  PdtReplaceOptions<Arc> opts(root);
-  Replace(ifst_array, ofst, parens, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_REPLACE_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/reverse.h b/projects/llm_framework/include/fst/extensions/pdt/reverse.h
deleted file mode 100644
index 3bd44e21..00000000
--- a/projects/llm_framework/include/fst/extensions/pdt/reverse.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Expands a PDT to an FST.
-
-#ifndef FST_EXTENSIONS_PDT_REVERSE_H_
-#define FST_EXTENSIONS_PDT_REVERSE_H_
-
-#include <vector>
-
-#include <fst/mutable-fst.h>
-#include <fst/relabel.h>
-#include <fst/reverse.h>
-
-namespace fst {
-
-// Reverses a pushdown transducer (PDT) encoded as an FST.
-template <class Arc, class RevArc>
-void Reverse(const Fst<Arc> &ifst,
-             const std::vector<
-                 std::pair<typename Arc::Label, typename Arc::Label>> &parens,
-             MutableFst<RevArc> *ofst) {
-  using Label = typename Arc::Label;
-  // Reverses FST component.
-  Reverse(ifst, ofst);
-  // Exchanges open and close parenthesis pairs.
-  std::vector<std::pair<Label, Label>> relabel_pairs;
-  relabel_pairs.reserve(2 * parens.size());
-  for (const auto &pair : parens) {
-    relabel_pairs.emplace_back(pair.first, pair.second);
-    relabel_pairs.emplace_back(pair.second, pair.first);
-  }
-  Relabel(ofst, relabel_pairs, relabel_pairs);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_REVERSE_H_
diff --git a/projects/llm_framework/include/fst/extensions/pdt/shortest-path.h b/projects/llm_framework/include/fst/extensions/pdt/shortest-path.h
deleted file mode 100644
index 591c1ea4..00000000
--- a/projects/llm_framework/include/fst/extensions/pdt/shortest-path.h
+++ /dev/null
@@ -1,715 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions to find shortest paths in a PDT.
-
-#ifndef FST_EXTENSIONS_PDT_SHORTEST_PATH_H_
-#define FST_EXTENSIONS_PDT_SHORTEST_PATH_H_
-
-#include <stack>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/extensions/pdt/paren.h>
-#include <fst/extensions/pdt/pdt.h>
-#include <fst/shortest-path.h>
-
-namespace fst {
-
-template <class Arc, class Queue>
-struct PdtShortestPathOptions {
-  bool keep_parentheses;
-  bool path_gc;
-
-  PdtShortestPathOptions(bool keep_parentheses = false, bool path_gc = true)
-      : keep_parentheses(keep_parentheses), path_gc(path_gc) {}
-};
-
-namespace internal {
-
-// Flags for shortest path data.
-
-constexpr uint8 kPdtInited = 0x01;
-constexpr uint8 kPdtFinal = 0x02;
-constexpr uint8 kPdtMarked = 0x04;
-
-// Stores shortest path tree info Distance(), Parent(), and ArcParent()
-// information keyed on two types:
-//
-// 1. SearchState: This is a usual node in a shortest path tree but:
-//    a. is w.r.t a PDT search state (a pair of a PDT state and a "start" state,
-//    either the PDT start state or the destination state of an open
-//    parenthesis).
-//    b. the Distance() is from this "start" state to the search state.
-//    c. Parent().state is kNoLabel for the "start" state.
-//
-// 2. ParenSpec: This connects shortest path trees depending on the the
-// parenthesis taken. Given the parenthesis spec:
-//    a. the Distance() is from the Parent() "start" state to the parenthesis
-//    destination state.
-//    b. The ArcParent() is the parenthesis arc.
-template <class Arc>
-class PdtShortestPathData {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  struct SearchState {
-    StateId state;  // PDT state.
-    StateId start;  // PDT paren "start" state.
-
-    SearchState(StateId s = kNoStateId, StateId t = kNoStateId)
-        : state(s), start(t) {}
-
-    bool operator==(const SearchState &other) const {
-      if (&other == this) return true;
-      return other.state == state && other.start == start;
-    }
-  };
-
-  // Specifies paren ID, source and dest "start" states of a paren. These are
-  // the "start" states of the respective sub-graphs.
-  struct ParenSpec {
-    ParenSpec(Label paren_id = kNoLabel, StateId src_start = kNoStateId,
-              StateId dest_start = kNoStateId)
-        : paren_id(paren_id), src_start(src_start), dest_start(dest_start) {}
-
-    Label paren_id;
-    StateId src_start;   // Sub-graph "start" state for paren source.
-    StateId dest_start;  // Sub-graph "start" state for paren dest.
-
-    bool operator==(const ParenSpec &other) const {
-      if (&other == this) return true;
-      return (other.paren_id == paren_id &&
-              other.src_start == other.src_start &&
-              other.dest_start == dest_start);
-    }
-  };
-
-  struct SearchData {
-    SearchData()
-        : distance(Weight::Zero()),
-          parent(kNoStateId, kNoStateId),
-          paren_id(kNoLabel),
-          flags(0) {}
-
-    Weight distance;     // Distance to this state from PDT "start" state.
-    SearchState parent;  // Parent state in shortest path tree.
-    int16 paren_id;      // If parent arc has paren, paren ID (or kNoLabel).
-    uint8 flags;         // First byte reserved for PdtShortestPathData use.
-  };
-
-  PdtShortestPathData(bool gc)
-      : gc_(gc), nstates_(0), ngc_(0), finished_(false) {}
-
-  ~PdtShortestPathData() {
-    VLOG(1) << "opm size: " << paren_map_.size();
-    VLOG(1) << "# of search states: " << nstates_;
-    if (gc_) VLOG(1) << "# of GC'd search states: " << ngc_;
-  }
-
-  void Clear() {
-    search_map_.clear();
-    search_multimap_.clear();
-    paren_map_.clear();
-    state_ = SearchState(kNoStateId, kNoStateId);
-    nstates_ = 0;
-    ngc_ = 0;
-  }
-
-  // TODO(kbg): Currently copying SearchState and passing a const reference to
-  // ParenSpec. Benchmark to confirm this is the right thing to do.
-
-  Weight Distance(SearchState s) const { return GetSearchData(s)->distance; }
-
-  Weight Distance(const ParenSpec &paren) const {
-    return GetSearchData(paren)->distance;
-  }
-
-  SearchState Parent(SearchState s) const { return GetSearchData(s)->parent; }
-
-  SearchState Parent(const ParenSpec &paren) const {
-    return GetSearchData(paren)->parent;
-  }
-
-  Label ParenId(SearchState s) const { return GetSearchData(s)->paren_id; }
-
-  uint8 Flags(SearchState s) const { return GetSearchData(s)->flags; }
-
-  void SetDistance(SearchState s, Weight weight) {
-    GetSearchData(s)->distance = std::move(weight);
-  }
-
-  void SetDistance(const ParenSpec &paren, Weight weight) {
-    GetSearchData(paren)->distance = std::move(weight);
-  }
-
-  void SetParent(SearchState s, SearchState p) { GetSearchData(s)->parent = p; }
-
-  void SetParent(const ParenSpec &paren, SearchState p) {
-    GetSearchData(paren)->parent = p;
-  }
-
-  void SetParenId(SearchState s, Label p) {
-    if (p >= 32768) {
-      FSTERROR() << "PdtShortestPathData: Paren ID does not fit in an int16";
-    }
-    GetSearchData(s)->paren_id = p;
-  }
-
-  void SetFlags(SearchState s, uint8 f, uint8 mask) {
-    auto *data = GetSearchData(s);
-    data->flags &= ~mask;
-    data->flags |= f & mask;
-  }
-
-  void GC(StateId s);
-
-  void Finish() { finished_ = true; }
-
- private:
-  // Hash for search state.
-  struct SearchStateHash {
-    size_t operator()(const SearchState &s) const {
-      static constexpr auto prime = 7853;
-      return s.state + s.start * prime;
-    }
-  };
-
-  // Hash for paren map.
-  struct ParenHash {
-    size_t operator()(const ParenSpec &paren) const {
-      static constexpr auto prime0 = 7853;
-      static constexpr auto prime1 = 7867;
-      return paren.paren_id + paren.src_start * prime0 +
-             paren.dest_start * prime1;
-    }
-  };
-
-  using SearchMap =
-      std::unordered_map<SearchState, SearchData, SearchStateHash>;
-
-  using SearchMultimap = std::unordered_multimap<StateId, StateId>;
-
-  // Hash map from paren spec to open paren data.
-  using ParenMap = std::unordered_map<ParenSpec, SearchData, ParenHash>;
-
-  SearchData *GetSearchData(SearchState s) const {
-    if (s == state_) return state_data_;
-    if (finished_) {
-      auto it = search_map_.find(s);
-      if (it == search_map_.end()) return &null_search_data_;
-      state_ = s;
-      return state_data_ = &(it->second);
-    } else {
-      state_ = s;
-      state_data_ = &search_map_[s];
-      if (!(state_data_->flags & kPdtInited)) {
-        ++nstates_;
-        if (gc_) search_multimap_.insert(std::make_pair(s.start, s.state));
-        state_data_->flags = kPdtInited;
-      }
-      return state_data_;
-    }
-  }
-
-  SearchData *GetSearchData(ParenSpec paren) const {
-    if (paren == paren_) return paren_data_;
-    if (finished_) {
-      auto it = paren_map_.find(paren);
-      if (it == paren_map_.end()) return &null_search_data_;
-      paren_ = paren;
-      return state_data_ = &(it->second);
-    } else {
-      paren_ = paren;
-      return paren_data_ = &paren_map_[paren];
-    }
-  }
-
-  mutable SearchMap search_map_;            // Maps from search state to data.
-  mutable SearchMultimap search_multimap_;  // Maps from "start" to subgraph.
-  mutable ParenMap paren_map_;              // Maps paren spec to search data.
-  mutable SearchState state_;               // Last state accessed.
-  mutable SearchData *state_data_;          // Last state data accessed.
-  mutable ParenSpec paren_;                 // Last paren spec accessed.
-  mutable SearchData *paren_data_;          // Last paren data accessed.
-  bool gc_;                                 // Allow GC?
-  mutable size_t nstates_;                  // Total number of search states.
-  size_t ngc_;                              // Number of GC'd search states.
-  mutable SearchData null_search_data_;     // Null search data.
-  bool finished_;                           // Read-only access when true.
-
-  PdtShortestPathData(const PdtShortestPathData &) = delete;
-  PdtShortestPathData &operator=(const PdtShortestPathData &) = delete;
-};
-
-// Deletes inaccessible search data from a given "start" (open paren dest)
-// state. Assumes "final" (close paren source or PDT final) states have
-// been flagged kPdtFinal.
-template <class Arc>
-void PdtShortestPathData<Arc>::GC(StateId start) {
-  if (!gc_) return;
-  std::vector<StateId> finals;
-  for (auto it = search_multimap_.find(start);
-       it != search_multimap_.end() && it->first == start; ++it) {
-    const SearchState s(it->second, start);
-    if (search_map_[s].flags & kPdtFinal) finals.push_back(s.state);
-  }
-  // Mark phase.
-  for (const auto state : finals) {
-    SearchState ss(state, start);
-    while (ss.state != kNoLabel) {
-      auto &sdata = search_map_[ss];
-      if (sdata.flags & kPdtMarked) break;
-      sdata.flags |= kPdtMarked;
-      const auto p = sdata.parent;
-      if (p.start != start && p.start != kNoLabel) {  // Entering sub-subgraph.
-        const ParenSpec paren(sdata.paren_id, ss.start, p.start);
-        ss = paren_map_[paren].parent;
-      } else {
-        ss = p;
-      }
-    }
-  }
-  // Sweep phase.
-  auto it = search_multimap_.find(start);
-  while (it != search_multimap_.end() && it->first == start) {
-    const SearchState s(it->second, start);
-    auto mit = search_map_.find(s);
-    const SearchData &data = mit->second;
-    if (!(data.flags & kPdtMarked)) {
-      search_map_.erase(mit);
-      ++ngc_;
-    }
-    search_multimap_.erase(it++);
-  }
-}
-
-}  // namespace internal
-
-// This computes the single source shortest (balanced) path (SSSP) through a
-// weighted PDT that has a bounded stack (i.e., is expandable as an FST). It is
-// a generalization of the classic SSSP graph algorithm that removes a state s
-// from a queue (defined by a user-provided queue type) and relaxes the
-// destination states of transitions leaving s. In this PDT version, states that
-// have entering open parentheses are treated as source states for a sub-graph
-// SSSP problem with the shortest path up to the open parenthesis being first
-// saved. When a close parenthesis is then encountered any balancing open
-// parenthesis is examined for this saved information and multiplied back. In
-// this way, each sub-graph is entered only once rather than repeatedly. If
-// every state in the input PDT has the property that there is a unique "start"
-// state for it with entering open parentheses, then this algorithm is quite
-// straightforward. In general, this will not be the case, so the algorithm
-// (implicitly) creates a new graph where each state is a pair of an original
-// state and a possible parenthesis "start" state for that state.
-template <class Arc, class Queue>
-class PdtShortestPath {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using SpData = internal::PdtShortestPathData<Arc>;
-  using SearchState = typename SpData::SearchState;
-  using ParenSpec = typename SpData::ParenSpec;
-  using CloseSourceIterator =
-      typename internal::PdtBalanceData<Arc>::SetIterator;
-
-  PdtShortestPath(const Fst<Arc> &ifst,
-                  const std::vector<std::pair<Label, Label>> &parens,
-                  const PdtShortestPathOptions<Arc, Queue> &opts)
-      : ifst_(ifst.Copy()),
-        parens_(parens),
-        keep_parens_(opts.keep_parentheses),
-        start_(ifst.Start()),
-        sp_data_(opts.path_gc),
-        error_(false) {
-    // TODO(kbg): Make this a compile-time static_assert once:
-    // 1) All weight properties are made constexpr for all weight types.
-    // 2) We have a pleasant way to "deregister" this oepration for non-path
-    //    semirings so an informative error message is produced. The best
-    //    solution will probably involve some kind of SFINAE magic.
-    if ((Weight::Properties() & (kPath | kRightSemiring)) !=
-        (kPath | kRightSemiring)) {
-      FSTERROR() << "PdtShortestPath: Weight needs to have the path"
-                 << " property and be right distributive: " << Weight::Type();
-      error_ = true;
-    }
-    for (Label i = 0; i < parens.size(); ++i) {
-      const auto &pair = parens[i];
-      paren_map_[pair.first] = i;
-      paren_map_[pair.second] = i;
-    }
-  }
-
-  ~PdtShortestPath() {
-    VLOG(1) << "# of input states: " << CountStates(*ifst_);
-    VLOG(1) << "# of enqueued: " << nenqueued_;
-    VLOG(1) << "cpmm size: " << close_paren_multimap_.size();
-  }
-
-  void ShortestPath(MutableFst<Arc> *ofst) {
-    Init(ofst);
-    GetDistance(start_);
-    GetPath();
-    sp_data_.Finish();
-    if (error_) ofst->SetProperties(kError, kError);
-  }
-
-  const internal::PdtShortestPathData<Arc> &GetShortestPathData() const {
-    return sp_data_;
-  }
-
-  internal::PdtBalanceData<Arc> *GetBalanceData() { return &balance_data_; }
-
- public:
-  // Hash multimap from close paren label to an paren arc.
-  using CloseParenMultimap =
-      std::unordered_multimap<internal::ParenState<Arc>, Arc,
-                              typename internal::ParenState<Arc>::Hash>;
-
-  const CloseParenMultimap &GetCloseParenMultimap() const {
-    return close_paren_multimap_;
-  }
-
- private:
-  void Init(MutableFst<Arc> *ofst);
-
-  void GetDistance(StateId start);
-
-  void ProcFinal(SearchState s);
-
-  void ProcArcs(SearchState s);
-
-  void ProcOpenParen(Label paren_id, SearchState s, StateId nexstate,
-                     const Weight &weight);
-
-  void ProcCloseParen(Label paren_id, SearchState s, const Weight &weight);
-
-  void ProcNonParen(SearchState s, StateId nextstate, const Weight &weight);
-
-  void Relax(SearchState s, SearchState t, StateId nextstate,
-             const Weight &weight, Label paren_id);
-
-  void Enqueue(SearchState d);
-
-  void GetPath();
-
-  Arc GetPathArc(SearchState s, SearchState p, Label paren_id, bool open);
-
-  std::unique_ptr<Fst<Arc>> ifst_;
-  MutableFst<Arc> *ofst_;
-  const std::vector<std::pair<Label, Label>> &parens_;
-  bool keep_parens_;
-  Queue *state_queue_;
-  StateId start_;
-  Weight fdistance_;
-  SearchState f_parent_;
-  SpData sp_data_;
-  std::unordered_map<Label, Label> paren_map_;
-  CloseParenMultimap close_paren_multimap_;
-  internal::PdtBalanceData<Arc> balance_data_;
-  ssize_t nenqueued_;
-  bool error_;
-
-  static constexpr uint8 kEnqueued = 0x10;
-  static constexpr uint8 kExpanded = 0x20;
-  static constexpr uint8 kFinished = 0x40;
-
-  static const Arc kNoArc;
-};
-
-template <class Arc, class Queue>
-void PdtShortestPath<Arc, Queue>::Init(MutableFst<Arc> *ofst) {
-  ofst_ = ofst;
-  ofst->DeleteStates();
-  ofst->SetInputSymbols(ifst_->InputSymbols());
-  ofst->SetOutputSymbols(ifst_->OutputSymbols());
-  if (ifst_->Start() == kNoStateId) return;
-  fdistance_ = Weight::Zero();
-  f_parent_ = SearchState(kNoStateId, kNoStateId);
-  sp_data_.Clear();
-  close_paren_multimap_.clear();
-  balance_data_.Clear();
-  nenqueued_ = 0;
-  // Finds open parens per destination state and close parens per source state.
-  for (StateIterator<Fst<Arc>> siter(*ifst_); !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    for (ArcIterator<Fst<Arc>> aiter(*ifst_, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto it = paren_map_.find(arc.ilabel);
-      if (it != paren_map_.end()) {  // Is a paren?
-        const auto paren_id = it->second;
-        if (arc.ilabel == parens_[paren_id].first) {  // Open paren.
-          balance_data_.OpenInsert(paren_id, arc.nextstate);
-        } else {  // Close paren.
-          const internal::ParenState<Arc> paren_state(paren_id, s);
-          close_paren_multimap_.emplace(paren_state, arc);
-        }
-      }
-    }
-  }
-}
-
-// Computes the shortest distance stored in a recursive way. Each sub-graph
-// (i.e., different paren "start" state) begins with weight One().
-template <class Arc, class Queue>
-void PdtShortestPath<Arc, Queue>::GetDistance(StateId start) {
-  if (start == kNoStateId) return;
-  Queue state_queue;
-  state_queue_ = &state_queue;
-  const SearchState q(start, start);
-  Enqueue(q);
-  sp_data_.SetDistance(q, Weight::One());
-  while (!state_queue_->Empty()) {
-    const auto state = state_queue_->Head();
-    state_queue_->Dequeue();
-    const SearchState s(state, start);
-    sp_data_.SetFlags(s, 0, kEnqueued);
-    ProcFinal(s);
-    ProcArcs(s);
-    sp_data_.SetFlags(s, kExpanded, kExpanded);
-  }
-  sp_data_.SetFlags(q, kFinished, kFinished);
-  balance_data_.FinishInsert(start);
-  sp_data_.GC(start);
-}
-
-// Updates best complete path.
-template <class Arc, class Queue>
-void PdtShortestPath<Arc, Queue>::ProcFinal(SearchState s) {
-  if (ifst_->Final(s.state) != Weight::Zero() && s.start == start_) {
-    const auto weight = Times(sp_data_.Distance(s), ifst_->Final(s.state));
-    if (fdistance_ != Plus(fdistance_, weight)) {
-      if (f_parent_.state != kNoStateId) {
-        sp_data_.SetFlags(f_parent_, 0, internal::kPdtFinal);
-      }
-      sp_data_.SetFlags(s, internal::kPdtFinal, internal::kPdtFinal);
-      fdistance_ = Plus(fdistance_, weight);
-      f_parent_ = s;
-    }
-  }
-}
-
-// Processes all arcs leaving the state s.
-template <class Arc, class Queue>
-void PdtShortestPath<Arc, Queue>::ProcArcs(SearchState s) {
-  for (ArcIterator<Fst<Arc>> aiter(*ifst_, s.state); !aiter.Done();
-       aiter.Next()) {
-    const auto &arc = aiter.Value();
-    const auto weight = Times(sp_data_.Distance(s), arc.weight);
-    const auto it = paren_map_.find(arc.ilabel);
-    if (it != paren_map_.end()) {  // Is a paren?
-      const auto paren_id = it->second;
-      if (arc.ilabel == parens_[paren_id].first) {
-        ProcOpenParen(paren_id, s, arc.nextstate, weight);
-      } else {
-        ProcCloseParen(paren_id, s, weight);
-      }
-    } else {
-      ProcNonParen(s, arc.nextstate, weight);
-    }
-  }
-}
-
-// Saves the shortest path info for reaching this parenthesis and starts a new
-// SSSP in the sub-graph pointed to by the parenthesis if previously unvisited.
-// Otherwise it finds any previously encountered closing parentheses and relaxes
-// them using the recursively stored shortest distance to them.
-template <class Arc, class Queue>
-inline void PdtShortestPath<Arc, Queue>::ProcOpenParen(Label paren_id,
-                                                       SearchState s,
-                                                       StateId nextstate,
-                                                       const Weight &weight) {
-  const SearchState d(nextstate, nextstate);
-  const ParenSpec paren(paren_id, s.start, d.start);
-  const auto pdist = sp_data_.Distance(paren);
-  if (pdist != Plus(pdist, weight)) {
-    sp_data_.SetDistance(paren, weight);
-    sp_data_.SetParent(paren, s);
-    const auto dist = sp_data_.Distance(d);
-    if (dist == Weight::Zero()) {
-      auto *state_queue = state_queue_;
-      GetDistance(d.start);
-      state_queue_ = state_queue;
-    } else if (!(sp_data_.Flags(d) & kFinished)) {
-      FSTERROR()
-          << "PdtShortestPath: open parenthesis recursion: not bounded stack";
-      error_ = true;
-    }
-    for (auto set_iter = balance_data_.Find(paren_id, nextstate);
-         !set_iter.Done(); set_iter.Next()) {
-      const SearchState cpstate(set_iter.Element(), d.start);
-      const internal::ParenState<Arc> paren_state(paren_id, cpstate.state);
-      for (auto cpit = close_paren_multimap_.find(paren_state);
-           cpit != close_paren_multimap_.end() && paren_state == cpit->first;
-           ++cpit) {
-        const auto &cparc = cpit->second;
-        const auto cpw =
-            Times(weight, Times(sp_data_.Distance(cpstate), cparc.weight));
-        Relax(cpstate, s, cparc.nextstate, cpw, paren_id);
-      }
-    }
-  }
-}
-
-// Saves the correspondence between each closing parenthesis and its balancing
-// open parenthesis info. Relaxes any close parenthesis destination state that
-// has a balancing previously encountered open parenthesis.
-template <class Arc, class Queue>
-inline void PdtShortestPath<Arc, Queue>::ProcCloseParen(Label paren_id,
-                                                        SearchState s,
-                                                        const Weight &weight) {
-  const internal::ParenState<Arc> paren_state(paren_id, s.start);
-  if (!(sp_data_.Flags(s) & kExpanded)) {
-    balance_data_.CloseInsert(paren_id, s.start, s.state);
-    sp_data_.SetFlags(s, internal::kPdtFinal, internal::kPdtFinal);
-  }
-}
-
-// Classical relaxation for non-parentheses.
-template <class Arc, class Queue>
-inline void PdtShortestPath<Arc, Queue>::ProcNonParen(SearchState s,
-                                                      StateId nextstate,
-                                                      const Weight &weight) {
-  Relax(s, s, nextstate, weight, kNoLabel);
-}
-
-// Classical relaxation on the search graph for an arc with destination state
-// nexstate from state s. State t is in the same sub-graph as nextstate (i.e.,
-// has the same paren "start").
-template <class Arc, class Queue>
-inline void PdtShortestPath<Arc, Queue>::Relax(SearchState s, SearchState t,
-                                               StateId nextstate,
-                                               const Weight &weight,
-                                               Label paren_id) {
-  const SearchState d(nextstate, t.start);
-  Weight dist = sp_data_.Distance(d);
-  if (dist != Plus(dist, weight)) {
-    sp_data_.SetParent(d, s);
-    sp_data_.SetParenId(d, paren_id);
-    sp_data_.SetDistance(d, Plus(dist, weight));
-    Enqueue(d);
-  }
-}
-
-template <class Arc, class Queue>
-inline void PdtShortestPath<Arc, Queue>::Enqueue(SearchState s) {
-  if (!(sp_data_.Flags(s) & kEnqueued)) {
-    state_queue_->Enqueue(s.state);
-    sp_data_.SetFlags(s, kEnqueued, kEnqueued);
-    ++nenqueued_;
-  } else {
-    state_queue_->Update(s.state);
-  }
-}
-
-// Follows parent pointers to find the shortest path. A stack is used since the
-// shortest distance is stored recursively.
-template <class Arc, class Queue>
-void PdtShortestPath<Arc, Queue>::GetPath() {
-  SearchState s = f_parent_;
-  SearchState d = SearchState(kNoStateId, kNoStateId);
-  StateId s_p = kNoStateId;
-  StateId d_p = kNoStateId;
-  auto arc = kNoArc;
-  Label paren_id = kNoLabel;
-  std::stack<ParenSpec> paren_stack;
-  while (s.state != kNoStateId) {
-    d_p = s_p;
-    s_p = ofst_->AddState();
-    if (d.state == kNoStateId) {
-      ofst_->SetFinal(s_p, ifst_->Final(f_parent_.state));
-    } else {
-      if (paren_id != kNoLabel) {                     // Paren?
-        if (arc.ilabel == parens_[paren_id].first) {  // Open paren?
-          paren_stack.pop();
-        } else {  // Close paren?
-          const ParenSpec paren(paren_id, d.start, s.start);
-          paren_stack.push(paren);
-        }
-        if (!keep_parens_) arc.ilabel = arc.olabel = 0;
-      }
-      arc.nextstate = d_p;
-      ofst_->AddArc(s_p, arc);
-    }
-    d = s;
-    s = sp_data_.Parent(d);
-    paren_id = sp_data_.ParenId(d);
-    if (s.state != kNoStateId) {
-      arc = GetPathArc(s, d, paren_id, false);
-    } else if (!paren_stack.empty()) {
-      const ParenSpec paren = paren_stack.top();
-      s = sp_data_.Parent(paren);
-      paren_id = paren.paren_id;
-      arc = GetPathArc(s, d, paren_id, true);
-    }
-  }
-  ofst_->SetStart(s_p);
-  ofst_->SetProperties(
-      ShortestPathProperties(ofst_->Properties(kFstProperties, false)),
-      kFstProperties);
-}
-
-// Finds transition with least weight between two states with label matching
-// paren_id and open/close paren type or a non-paren if kNoLabel.
-template <class Arc, class Queue>
-Arc PdtShortestPath<Arc, Queue>::GetPathArc(SearchState s, SearchState d,
-                                            Label paren_id, bool open_paren) {
-  auto path_arc = kNoArc;
-  for (ArcIterator<Fst<Arc>> aiter(*ifst_, s.state); !aiter.Done();
-       aiter.Next()) {
-    const auto &arc = aiter.Value();
-    if (arc.nextstate != d.state) continue;
-    Label arc_paren_id = kNoLabel;
-    const auto it = paren_map_.find(arc.ilabel);
-    if (it != paren_map_.end()) {
-      arc_paren_id = it->second;
-      bool arc_open_paren = (arc.ilabel == parens_[arc_paren_id].first);
-      if (arc_open_paren != open_paren) continue;
-    }
-    if (arc_paren_id != paren_id) continue;
-    if (arc.weight == Plus(arc.weight, path_arc.weight)) path_arc = arc;
-  }
-  if (path_arc.nextstate == kNoStateId) {
-    FSTERROR() << "PdtShortestPath::GetPathArc: Failed to find arc";
-    error_ = true;
-  }
-  return path_arc;
-}
-
-template <class Arc, class Queue>
-const Arc PdtShortestPath<Arc, Queue>::kNoArc = Arc(kNoLabel, kNoLabel,
-                                                    Weight::Zero(), kNoStateId);
-
-// Functional variants.
-
-template <class Arc, class Queue>
-void ShortestPath(
-    const Fst<Arc> &ifst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens,
-    MutableFst<Arc> *ofst, const PdtShortestPathOptions<Arc, Queue> &opts) {
-  PdtShortestPath<Arc, Queue> psp(ifst, parens, opts);
-  psp.ShortestPath(ofst);
-}
-
-template <class Arc>
-void ShortestPath(
-    const Fst<Arc> &ifst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens,
-    MutableFst<Arc> *ofst) {
-  using Q = FifoQueue<typename Arc::StateId>;
-  const PdtShortestPathOptions<Arc, Q> opts;
-  PdtShortestPath<Arc, Q> psp(ifst, parens, opts);
-  psp.ShortestPath(ofst);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_SHORTEST_PATH_H_
diff --git a/projects/llm_framework/include/fst/extensions/special/phi-fst.h b/projects/llm_framework/include/fst/extensions/special/phi-fst.h
deleted file mode 100644
index db40d611..00000000
--- a/projects/llm_framework/include/fst/extensions/special/phi-fst.h
+++ /dev/null
@@ -1,183 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_SPECIAL_PHI_FST_H_
-#define FST_EXTENSIONS_SPECIAL_PHI_FST_H_
-
-#include <memory>
-#include <string>
-
-#include <fst/const-fst.h>
-#include <fst/matcher-fst.h>
-#include <fst/matcher.h>
-
-DECLARE_int64(phi_fst_phi_label);
-DECLARE_bool(phi_fst_phi_loop);
-DECLARE_string(phi_fst_rewrite_mode);
-
-namespace fst {
-namespace internal {
-
-template <class Label>
-class PhiFstMatcherData {
- public:
-  PhiFstMatcherData(
-      Label phi_label = FLAGS_phi_fst_phi_label,
-      bool phi_loop = FLAGS_phi_fst_phi_loop,
-      MatcherRewriteMode rewrite_mode = RewriteMode(FLAGS_phi_fst_rewrite_mode))
-      : phi_label_(phi_label),
-        phi_loop_(phi_loop),
-        rewrite_mode_(rewrite_mode) {}
-
-  PhiFstMatcherData(const PhiFstMatcherData &data)
-      : phi_label_(data.phi_label_),
-        phi_loop_(data.phi_loop_),
-        rewrite_mode_(data.rewrite_mode_) {}
-
-  static PhiFstMatcherData<Label> *Read(std::istream &istrm,
-                                        const FstReadOptions &read) {
-    auto *data = new PhiFstMatcherData<Label>();
-    ReadType(istrm, &data->phi_label_);
-    ReadType(istrm, &data->phi_loop_);
-    int32 rewrite_mode;
-    ReadType(istrm, &rewrite_mode);
-    data->rewrite_mode_ = static_cast<MatcherRewriteMode>(rewrite_mode);
-    return data;
-  }
-
-  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
-    WriteType(ostrm, phi_label_);
-    WriteType(ostrm, phi_loop_);
-    WriteType(ostrm, static_cast<int32>(rewrite_mode_));
-    return !ostrm ? false : true;
-  }
-
-  Label PhiLabel() const { return phi_label_; }
-
-  bool PhiLoop() const { return phi_loop_; }
-
-  MatcherRewriteMode RewriteMode() const { return rewrite_mode_; }
-
- private:
-  static MatcherRewriteMode RewriteMode(const string &mode) {
-    if (mode == "auto") return MATCHER_REWRITE_AUTO;
-    if (mode == "always") return MATCHER_REWRITE_ALWAYS;
-    if (mode == "never") return MATCHER_REWRITE_NEVER;
-    LOG(WARNING) << "PhiFst: Unknown rewrite mode: " << mode << ". "
-                 << "Defaulting to auto.";
-    return MATCHER_REWRITE_AUTO;
-  }
-
-  Label phi_label_;
-  bool phi_loop_;
-  MatcherRewriteMode rewrite_mode_;
-};
-
-}  // namespace internal
-
-constexpr uint8 kPhiFstMatchInput = 0x01;   // Input matcher is PhiMatcher.
-constexpr uint8 kPhiFstMatchOutput = 0x02;  // Output matcher is PhiMatcher.
-
-template <class M, uint8 flags = kPhiFstMatchInput | kPhiFstMatchOutput>
-class PhiFstMatcher : public PhiMatcher<M> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename M::Arc;
-  using StateId = typename Arc::StateId;
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-  using MatcherData = internal::PhiFstMatcherData<Label>;
-
-  enum : uint8 { kFlags = flags };
-
-  // This makes a copy of the FST.
-  PhiFstMatcher(const FST &fst, MatchType match_type,
-      std::shared_ptr<MatcherData> data = std::make_shared<MatcherData>())
-      : PhiMatcher<M>(fst, match_type,
-                      PhiLabel(match_type, data ? data->PhiLabel()
-                                                : MatcherData().PhiLabel()),
-                      data ? data->PhiLoop() : MatcherData().PhiLoop(),
-                      data ? data->RewriteMode() : MatcherData().RewriteMode()),
-        data_(data) {}
-
-  // This doesn't copy the FST.
-  PhiFstMatcher(const FST *fst, MatchType match_type,
-      std::shared_ptr<MatcherData> data = std::make_shared<MatcherData>())
-      : PhiMatcher<M>(fst, match_type,
-                      PhiLabel(match_type, data ? data->PhiLabel()
-                                                : MatcherData().PhiLabel()),
-                      data ? data->PhiLoop() : MatcherData().PhiLoop(),
-                      data ? data->RewriteMode() : MatcherData().RewriteMode()),
-        data_(data) {}
-
-  // This makes a copy of the FST.
-  PhiFstMatcher(const PhiFstMatcher<M, flags> &matcher, bool safe = false)
-      : PhiMatcher<M>(matcher, safe), data_(matcher.data_) {}
-
-  PhiFstMatcher<M, flags> *Copy(bool safe = false) const override {
-    return new PhiFstMatcher<M, flags>(*this, safe);
-  }
-
-  const MatcherData *GetData() const { return data_.get(); }
-
-  std::shared_ptr<MatcherData> GetSharedData() const { return data_; }
-
- private:
-  static Label PhiLabel(MatchType match_type, Label label) {
-    if (match_type == MATCH_INPUT && flags & kPhiFstMatchInput) return label;
-    if (match_type == MATCH_OUTPUT && flags & kPhiFstMatchOutput) return label;
-    return kNoLabel;
-  }
-
-  std::shared_ptr<MatcherData> data_;
-};
-
-extern const char phi_fst_type[];
-extern const char input_phi_fst_type[];
-extern const char output_phi_fst_type[];
-
-using StdPhiFst =
-    MatcherFst<ConstFst<StdArc>, PhiFstMatcher<SortedMatcher<ConstFst<StdArc>>>,
-               phi_fst_type>;
-
-using LogPhiFst =
-    MatcherFst<ConstFst<LogArc>, PhiFstMatcher<SortedMatcher<ConstFst<LogArc>>>,
-               phi_fst_type>;
-
-using Log64PhiFst = MatcherFst<ConstFst<Log64Arc>,
-                               PhiFstMatcher<SortedMatcher<ConstFst<Log64Arc>>>,
-                               input_phi_fst_type>;
-
-using StdInputPhiFst =
-    MatcherFst<ConstFst<StdArc>, PhiFstMatcher<SortedMatcher<ConstFst<StdArc>>,
-                                               kPhiFstMatchInput>,
-               input_phi_fst_type>;
-
-using LogInputPhiFst =
-    MatcherFst<ConstFst<LogArc>, PhiFstMatcher<SortedMatcher<ConstFst<LogArc>>,
-                                               kPhiFstMatchInput>,
-               input_phi_fst_type>;
-
-using Log64InputPhiFst = MatcherFst<
-    ConstFst<Log64Arc>,
-    PhiFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kPhiFstMatchInput>,
-    input_phi_fst_type>;
-
-using StdOutputPhiFst =
-    MatcherFst<ConstFst<StdArc>, PhiFstMatcher<SortedMatcher<ConstFst<StdArc>>,
-                                               kPhiFstMatchOutput>,
-               output_phi_fst_type>;
-
-using LogOutputPhiFst =
-    MatcherFst<ConstFst<LogArc>, PhiFstMatcher<SortedMatcher<ConstFst<LogArc>>,
-                                               kPhiFstMatchOutput>,
-               output_phi_fst_type>;
-
-using Log64OutputPhiFst = MatcherFst<
-    ConstFst<Log64Arc>,
-    PhiFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kPhiFstMatchOutput>,
-    output_phi_fst_type>;
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_SPECIAL_PHI_FST_H_
diff --git a/projects/llm_framework/include/fst/extensions/special/rho-fst.h b/projects/llm_framework/include/fst/extensions/special/rho-fst.h
deleted file mode 100644
index ababecc0..00000000
--- a/projects/llm_framework/include/fst/extensions/special/rho-fst.h
+++ /dev/null
@@ -1,172 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_SPECIAL_RHO_FST_H_
-#define FST_EXTENSIONS_SPECIAL_RHO_FST_H_
-
-#include <memory>
-#include <string>
-
-#include <fst/const-fst.h>
-#include <fst/matcher-fst.h>
-#include <fst/matcher.h>
-
-DECLARE_int64(rho_fst_rho_label);
-DECLARE_string(rho_fst_rewrite_mode);
-
-namespace fst {
-namespace internal {
-
-template <class Label>
-class RhoFstMatcherData {
- public:
-  explicit RhoFstMatcherData(
-      Label rho_label = FLAGS_rho_fst_rho_label,
-      MatcherRewriteMode rewrite_mode = RewriteMode(FLAGS_rho_fst_rewrite_mode))
-      : rho_label_(rho_label), rewrite_mode_(rewrite_mode) {}
-
-  RhoFstMatcherData(const RhoFstMatcherData &data)
-      : rho_label_(data.rho_label_), rewrite_mode_(data.rewrite_mode_) {}
-
-  static RhoFstMatcherData<Label> *Read(std::istream &istrm,
-                                    const FstReadOptions &read) {
-    auto *data = new RhoFstMatcherData<Label>();
-    ReadType(istrm, &data->rho_label_);
-    int32 rewrite_mode;
-    ReadType(istrm, &rewrite_mode);
-    data->rewrite_mode_ = static_cast<MatcherRewriteMode>(rewrite_mode);
-    return data;
-  }
-
-  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
-    WriteType(ostrm, rho_label_);
-    WriteType(ostrm, static_cast<int32>(rewrite_mode_));
-    return !ostrm ? false : true;
-  }
-
-  Label RhoLabel() const { return rho_label_; }
-
-  MatcherRewriteMode RewriteMode() const { return rewrite_mode_; }
-
- private:
-  static MatcherRewriteMode RewriteMode(const string &mode) {
-    if (mode == "auto") return MATCHER_REWRITE_AUTO;
-    if (mode == "always") return MATCHER_REWRITE_ALWAYS;
-    if (mode == "never") return MATCHER_REWRITE_NEVER;
-    LOG(WARNING) << "RhoFst: Unknown rewrite mode: " << mode << ". "
-                 << "Defaulting to auto.";
-    return MATCHER_REWRITE_AUTO;
-  }
-
-  Label rho_label_;
-  MatcherRewriteMode rewrite_mode_;
-};
-
-}  // namespace internal
-
-constexpr uint8 kRhoFstMatchInput = 0x01;   // Input matcher is RhoMatcher.
-constexpr uint8 kRhoFstMatchOutput = 0x02;  // Output matcher is RhoMatcher.
-
-template <class M, uint8 flags = kRhoFstMatchInput | kRhoFstMatchOutput>
-class RhoFstMatcher : public RhoMatcher<M> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename M::Arc;
-  using StateId = typename Arc::StateId;
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-  using MatcherData = internal::RhoFstMatcherData<Label>;
-
-  enum : uint8 { kFlags = flags };
-
-  // This makes a copy of the FST.
-  RhoFstMatcher(
-      const FST &fst, MatchType match_type,
-      std::shared_ptr<MatcherData> data = std::make_shared<MatcherData>())
-      : RhoMatcher<M>(fst, match_type,
-                      RhoLabel(match_type, data ? data->RhoLabel()
-                                                : MatcherData().RhoLabel()),
-                      data ? data->RewriteMode() : MatcherData().RewriteMode()),
-        data_(data) {}
-
-  // This doesn't copy the FST.
-  RhoFstMatcher(
-      const FST *fst, MatchType match_type,
-      std::shared_ptr<MatcherData> data = std::make_shared<MatcherData>())
-      : RhoMatcher<M>(fst, match_type,
-                      RhoLabel(match_type, data ? data->RhoLabel()
-                                                : MatcherData().RhoLabel()),
-                      data ? data->RewriteMode() : MatcherData().RewriteMode()),
-        data_(data) {}
-
-  // This makes a copy of the FST.
-  RhoFstMatcher(const RhoFstMatcher<M, flags> &matcher, bool safe = false)
-      : RhoMatcher<M>(matcher, safe), data_(matcher.data_) {}
-
-  RhoFstMatcher<M, flags> *Copy(bool safe = false) const override {
-    return new RhoFstMatcher<M, flags>(*this, safe);
-  }
-
-  const MatcherData *GetData() const { return data_.get(); }
-
-  std::shared_ptr<MatcherData> GetSharedData() const { return data_; }
-
- private:
-  static Label RhoLabel(MatchType match_type, Label label) {
-    if (match_type == MATCH_INPUT && flags & kRhoFstMatchInput) return label;
-    if (match_type == MATCH_OUTPUT && flags & kRhoFstMatchOutput) return label;
-    return kNoLabel;
-  }
-
-  std::shared_ptr<MatcherData> data_;
-};
-
-extern const char rho_fst_type[];
-extern const char input_rho_fst_type[];
-extern const char output_rho_fst_type[];
-
-using StdRhoFst =
-    MatcherFst<ConstFst<StdArc>, RhoFstMatcher<SortedMatcher<ConstFst<StdArc>>>,
-               rho_fst_type>;
-
-using LogRhoFst =
-    MatcherFst<ConstFst<LogArc>, RhoFstMatcher<SortedMatcher<ConstFst<LogArc>>>,
-               rho_fst_type>;
-
-using Log64RhoFst = MatcherFst<ConstFst<Log64Arc>,
-                               RhoFstMatcher<SortedMatcher<ConstFst<Log64Arc>>>,
-                               input_rho_fst_type>;
-
-using StdInputRhoFst =
-    MatcherFst<ConstFst<StdArc>, RhoFstMatcher<SortedMatcher<ConstFst<StdArc>>,
-                                               kRhoFstMatchInput>,
-               input_rho_fst_type>;
-
-using LogInputRhoFst =
-    MatcherFst<ConstFst<LogArc>, RhoFstMatcher<SortedMatcher<ConstFst<LogArc>>,
-                                               kRhoFstMatchInput>,
-               input_rho_fst_type>;
-
-using Log64InputRhoFst = MatcherFst<
-    ConstFst<Log64Arc>,
-    RhoFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kRhoFstMatchInput>,
-    input_rho_fst_type>;
-
-using StdOutputRhoFst =
-    MatcherFst<ConstFst<StdArc>, RhoFstMatcher<SortedMatcher<ConstFst<StdArc>>,
-                                               kRhoFstMatchOutput>,
-               output_rho_fst_type>;
-
-using LogOutputRhoFst =
-    MatcherFst<ConstFst<LogArc>, RhoFstMatcher<SortedMatcher<ConstFst<LogArc>>,
-                                               kRhoFstMatchOutput>,
-               output_rho_fst_type>;
-
-using Log64OutputRhoFst = MatcherFst<
-    ConstFst<Log64Arc>,
-    RhoFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kRhoFstMatchOutput>,
-    output_rho_fst_type>;
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_SPECIAL_RHO_FST_H_
diff --git a/projects/llm_framework/include/fst/extensions/special/sigma-fst.h b/projects/llm_framework/include/fst/extensions/special/sigma-fst.h
deleted file mode 100644
index 0d6cfd03..00000000
--- a/projects/llm_framework/include/fst/extensions/special/sigma-fst.h
+++ /dev/null
@@ -1,176 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_SPECIAL_SIGMA_FST_H_
-#define FST_EXTENSIONS_SPECIAL_SIGMA_FST_H_
-
-#include <memory>
-#include <string>
-
-#include <fst/const-fst.h>
-#include <fst/matcher-fst.h>
-#include <fst/matcher.h>
-
-DECLARE_int64(sigma_fst_sigma_label);
-DECLARE_string(sigma_fst_rewrite_mode);
-
-namespace fst {
-namespace internal {
-
-template <class Label>
-class SigmaFstMatcherData {
- public:
-  explicit SigmaFstMatcherData(Label sigma_label = FLAGS_sigma_fst_sigma_label,
-      MatcherRewriteMode rewrite_mode =
-                          RewriteMode(FLAGS_sigma_fst_rewrite_mode))
-      : sigma_label_(sigma_label), rewrite_mode_(rewrite_mode) {}
-
-  SigmaFstMatcherData(const SigmaFstMatcherData &data)
-      : sigma_label_(data.sigma_label_), rewrite_mode_(data.rewrite_mode_) {}
-
-  static SigmaFstMatcherData<Label> *Read(std::istream &istrm,
-                                      const FstReadOptions &read) {
-    auto *data = new SigmaFstMatcherData<Label>();
-    ReadType(istrm, &data->sigma_label_);
-    int32 rewrite_mode;
-    ReadType(istrm, &rewrite_mode);
-    data->rewrite_mode_ = static_cast<MatcherRewriteMode>(rewrite_mode);
-    return data;
-  }
-
-  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
-    WriteType(ostrm, sigma_label_);
-    WriteType(ostrm, static_cast<int32>(rewrite_mode_));
-    return !ostrm ? false : true;
-  }
-
-  Label SigmaLabel() const { return sigma_label_; }
-
-  MatcherRewriteMode RewriteMode() const { return rewrite_mode_; }
-
- private:
-  static MatcherRewriteMode RewriteMode(const string &mode) {
-    if (mode == "auto") return MATCHER_REWRITE_AUTO;
-    if (mode == "always") return MATCHER_REWRITE_ALWAYS;
-    if (mode == "never") return MATCHER_REWRITE_NEVER;
-    LOG(WARNING) << "SigmaFst: Unknown rewrite mode: " << mode << ". "
-                 << "Defaulting to auto.";
-    return MATCHER_REWRITE_AUTO;
-  }
-
-  Label sigma_label_;
-  MatcherRewriteMode rewrite_mode_;
-};
-
-}  // namespace internal
-
-constexpr uint8 kSigmaFstMatchInput = 0x01;   // Input matcher is SigmaMatcher.
-constexpr uint8 kSigmaFstMatchOutput = 0x02;  // Output matcher is SigmaMatcher.
-
-template <class M, uint8 flags = kSigmaFstMatchInput | kSigmaFstMatchOutput>
-class SigmaFstMatcher : public SigmaMatcher<M> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename M::Arc;
-  using StateId = typename Arc::StateId;
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-  using MatcherData = internal::SigmaFstMatcherData<Label>;
-
-  enum : uint8 { kFlags = flags };
-
-  // This makes a copy of the FST.
-  SigmaFstMatcher(
-      const FST &fst, MatchType match_type,
-      std::shared_ptr<MatcherData> data = std::make_shared<MatcherData>())
-      : SigmaMatcher<M>(
-            fst, match_type,
-            SigmaLabel(match_type,
-                       data ? data->SigmaLabel() : MatcherData().SigmaLabel()),
-            data ? data->RewriteMode() : MatcherData().RewriteMode()),
-        data_(data) {}
-
-  // This doesn't copy the FST.
-  SigmaFstMatcher(
-      const FST *fst, MatchType match_type,
-      std::shared_ptr<MatcherData> data = std::make_shared<MatcherData>())
-      : SigmaMatcher<M>(
-            fst, match_type,
-            SigmaLabel(match_type,
-                       data ? data->SigmaLabel() : MatcherData().SigmaLabel()),
-            data ? data->RewriteMode() : MatcherData().RewriteMode()),
-        data_(data) {}
-
-  // This makes a copy of the FST.
-  SigmaFstMatcher(const SigmaFstMatcher<M, flags> &matcher, bool safe = false)
-      : SigmaMatcher<M>(matcher, safe), data_(matcher.data_) {}
-
-  SigmaFstMatcher<M, flags> *Copy(bool safe = false) const override {
-    return new SigmaFstMatcher<M, flags>(*this, safe);
-  }
-
-  const MatcherData *GetData() const { return data_.get(); }
-
-  std::shared_ptr<MatcherData> GetSharedData() const { return data_; }
-
- private:
-  static Label SigmaLabel(MatchType match_type, Label label) {
-    if (match_type == MATCH_INPUT && flags & kSigmaFstMatchInput) return label;
-    if (match_type == MATCH_OUTPUT && flags & kSigmaFstMatchOutput)
-      return label;
-    return kNoLabel;
-  }
-
-  std::shared_ptr<MatcherData> data_;
-};
-
-extern const char sigma_fst_type[];
-extern const char input_sigma_fst_type[];
-extern const char output_sigma_fst_type[];
-
-using StdSigmaFst = MatcherFst<ConstFst<StdArc>,
-                               SigmaFstMatcher<SortedMatcher<ConstFst<StdArc>>>,
-                               sigma_fst_type>;
-
-using LogSigmaFst = MatcherFst<ConstFst<LogArc>,
-                               SigmaFstMatcher<SortedMatcher<ConstFst<LogArc>>>,
-                               sigma_fst_type>;
-
-using Log64SigmaFst =
-    MatcherFst<ConstFst<Log64Arc>,
-               SigmaFstMatcher<SortedMatcher<ConstFst<Log64Arc>>>,
-               input_sigma_fst_type>;
-
-using StdInputSigmaFst = MatcherFst<
-    ConstFst<StdArc>,
-    SigmaFstMatcher<SortedMatcher<ConstFst<StdArc>>, kSigmaFstMatchInput>,
-    input_sigma_fst_type>;
-
-using LogInputSigmaFst = MatcherFst<
-    ConstFst<LogArc>,
-    SigmaFstMatcher<SortedMatcher<ConstFst<LogArc>>, kSigmaFstMatchInput>,
-    input_sigma_fst_type>;
-
-using Log64InputSigmaFst = MatcherFst<
-    ConstFst<Log64Arc>,
-    SigmaFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kSigmaFstMatchInput>,
-    input_sigma_fst_type>;
-
-using StdOutputSigmaFst = MatcherFst<
-    ConstFst<StdArc>,
-    SigmaFstMatcher<SortedMatcher<ConstFst<StdArc>>, kSigmaFstMatchOutput>,
-    output_sigma_fst_type>;
-
-using LogOutputSigmaFst = MatcherFst<
-    ConstFst<LogArc>,
-    SigmaFstMatcher<SortedMatcher<ConstFst<LogArc>>, kSigmaFstMatchOutput>,
-    output_sigma_fst_type>;
-
-using Log64OutputSigmaFst = MatcherFst<
-    ConstFst<Log64Arc>,
-    SigmaFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kSigmaFstMatchOutput>,
-    output_sigma_fst_type>;
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_SPECIAL_SIGMA_FST_H_
diff --git a/projects/llm_framework/include/fst/factor-weight.h b/projects/llm_framework/include/fst/factor-weight.h
deleted file mode 100644
index 30e02320..00000000
--- a/projects/llm_framework/include/fst/factor-weight.h
+++ /dev/null
@@ -1,496 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes to factor weights in an FST.
-
-#ifndef FST_FACTOR_WEIGHT_H_
-#define FST_FACTOR_WEIGHT_H_
-
-#include <algorithm>
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-constexpr uint32 kFactorFinalWeights = 0x00000001;
-constexpr uint32 kFactorArcWeights = 0x00000002;
-
-template <class Arc>
-struct FactorWeightOptions : CacheOptions {
-  using Label = typename Arc::Label;
-
-  float delta;
-  uint32 mode;         // Factor arc weights and/or final weights.
-  Label final_ilabel;  // Input label of arc when factoring final weights.
-  Label final_olabel;  // Output label of arc when factoring final weights.
-  bool increment_final_ilabel;  // When factoring final w' results in > 1 arcs
-  bool increment_final_olabel;  // at state, increment labels to make distinct?
-
-  explicit FactorWeightOptions(const CacheOptions &opts, float delta = kDelta,
-                               uint32 mode = kFactorArcWeights |
-                                             kFactorFinalWeights,
-                               Label final_ilabel = 0, Label final_olabel = 0,
-                               bool increment_final_ilabel = false,
-                               bool increment_final_olabel = false)
-      : CacheOptions(opts),
-        delta(delta),
-        mode(mode),
-        final_ilabel(final_ilabel),
-        final_olabel(final_olabel),
-        increment_final_ilabel(increment_final_ilabel),
-        increment_final_olabel(increment_final_olabel) {}
-
-  explicit FactorWeightOptions(float delta = kDelta,
-                               uint32 mode = kFactorArcWeights |
-                                             kFactorFinalWeights,
-                               Label final_ilabel = 0, Label final_olabel = 0,
-                               bool increment_final_ilabel = false,
-                               bool increment_final_olabel = false)
-      : delta(delta),
-        mode(mode),
-        final_ilabel(final_ilabel),
-        final_olabel(final_olabel),
-        increment_final_ilabel(increment_final_ilabel),
-        increment_final_olabel(increment_final_olabel) {}
-};
-
-// A factor iterator takes as argument a weight w and returns a sequence of
-// pairs of weights (xi, yi) such that the sum of the products xi times yi is
-// equal to w. If w is fully factored, the iterator should return nothing.
-//
-// template <class W>
-// class FactorIterator {
-//  public:
-//   explicit FactorIterator(W w);
-//
-//   bool Done() const;
-//
-//   void Next();
-//
-//   std::pair<W, W> Value() const;
-//
-//   void Reset();
-// }
-
-// Factors trivially.
-template <class W>
-class IdentityFactor {
- public:
-  explicit IdentityFactor(const W &weight) {}
-
-  bool Done() const { return true; }
-
-  void Next() {}
-
-  std::pair<W, W> Value() const { return std::make_pair(W::One(), W::One()); }
-
-  void Reset() {}
-};
-
-// Factors a StringWeight w as 'ab' where 'a' is a label.
-template <typename Label, StringType S = STRING_LEFT>
-class StringFactor {
- public:
-  explicit StringFactor(const StringWeight<Label, S> &weight)
-      : weight_(weight), done_(weight.Size() <= 1) {}
-
-  bool Done() const { return done_; }
-
-  void Next() { done_ = true; }
-
-  std::pair<StringWeight<Label, S>, StringWeight<Label, S>> Value() const {
-    using Weight = StringWeight<Label, S>;
-    typename Weight::Iterator siter(weight_);
-    Weight w1(siter.Value());
-    Weight w2;
-    for (siter.Next(); !siter.Done(); siter.Next()) w2.PushBack(siter.Value());
-    return std::make_pair(w1, w2);
-  }
-
-  void Reset() { done_ = weight_.Size() <= 1; }
-
- private:
-  const StringWeight<Label, S> weight_;
-  bool done_;
-};
-
-// Factor a GallicWeight using StringFactor.
-template <class Label, class W, GallicType G = GALLIC_LEFT>
-class GallicFactor {
- public:
-  using GW = GallicWeight<Label, W, G>;
-
-  explicit GallicFactor(const GW &weight)
-      : weight_(weight), done_(weight.Value1().Size() <= 1) {}
-
-  bool Done() const { return done_; }
-
-  void Next() { done_ = true; }
-
-  std::pair<GW, GW> Value() const {
-    StringFactor<Label, GallicStringType(G)> siter(weight_.Value1());
-    GW w1(siter.Value().first, weight_.Value2());
-    GW w2(siter.Value().second, W::One());
-    return std::make_pair(w1, w2);
-  }
-
-  void Reset() { done_ = weight_.Value1().Size() <= 1; }
-
- private:
-  const GW weight_;
-  bool done_;
-};
-
-// Specialization for the (general) GALLIC type GallicWeight.
-template <class Label, class W>
-class GallicFactor<Label, W, GALLIC> {
- public:
-  using GW = GallicWeight<Label, W, GALLIC>;
-  using GRW = GallicWeight<Label, W, GALLIC_RESTRICT>;
-
-  explicit GallicFactor(const GW &weight)
-      : iter_(weight),
-        done_(weight.Size() == 0 ||
-              (weight.Size() == 1 && weight.Back().Value1().Size() <= 1)) {}
-
-  bool Done() const { return done_ || iter_.Done(); }
-
-  void Next() { iter_.Next(); }
-
-  void Reset() { iter_.Reset(); }
-
-  std::pair<GW, GW> Value() const {
-    const auto weight = iter_.Value();
-    StringFactor<Label, GallicStringType(GALLIC_RESTRICT)> siter(
-        weight.Value1());
-    GRW w1(siter.Value().first, weight.Value2());
-    GRW w2(siter.Value().second, W::One());
-    return std::make_pair(GW(w1), GW(w2));
-  }
-
- private:
-  UnionWeightIterator<GRW, GallicUnionWeightOptions<Label, W>> iter_;
-  bool done_;
-};
-
-namespace internal {
-
-// Implementation class for FactorWeight
-template <class Arc, class FactorIterator>
-class FactorWeightFstImpl : public CacheImpl<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheBaseImpl<CacheState<Arc>>::EmplaceArc;
-  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
-  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
-  using CacheBaseImpl<CacheState<Arc>>::HasStart;
-  using CacheBaseImpl<CacheState<Arc>>::SetArcs;
-  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
-  using CacheBaseImpl<CacheState<Arc>>::SetStart;
-
-  struct Element {
-    Element() {}
-
-    Element(StateId s, Weight weight_) : state(s), weight(std::move(weight_)) {}
-
-    StateId state;  // Input state ID.
-    Weight weight;  // Residual weight.
-  };
-
-  FactorWeightFstImpl(const Fst<Arc> &fst, const FactorWeightOptions<Arc> &opts)
-      : CacheImpl<Arc>(opts),
-        fst_(fst.Copy()),
-        delta_(opts.delta),
-        mode_(opts.mode),
-        final_ilabel_(opts.final_ilabel),
-        final_olabel_(opts.final_olabel),
-        increment_final_ilabel_(opts.increment_final_ilabel),
-        increment_final_olabel_(opts.increment_final_olabel) {
-    SetType("factor_weight");
-    const auto props = fst.Properties(kFstProperties, false);
-    SetProperties(FactorWeightProperties(props), kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-    if (mode_ == 0) {
-      LOG(WARNING) << "FactorWeightFst: Factor mode is set to 0; "
-                   << "factoring neither arc weights nor final weights";
-    }
-  }
-
-  FactorWeightFstImpl(const FactorWeightFstImpl<Arc, FactorIterator> &impl)
-      : CacheImpl<Arc>(impl),
-        fst_(impl.fst_->Copy(true)),
-        delta_(impl.delta_),
-        mode_(impl.mode_),
-        final_ilabel_(impl.final_ilabel_),
-        final_olabel_(impl.final_olabel_),
-        increment_final_ilabel_(impl.increment_final_ilabel_),
-        increment_final_olabel_(impl.increment_final_olabel_) {
-    SetType("factor_weight");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      const auto s = fst_->Start();
-      if (s == kNoStateId) return kNoStateId;
-      SetStart(FindState(Element(fst_->Start(), Weight::One())));
-    }
-    return CacheImpl<Arc>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) {
-      const auto &element = elements_[s];
-      // TODO(sorenj): fix so cast is unnecessary
-      const auto weight =
-          element.state == kNoStateId
-              ? element.weight
-              : (Weight)Times(element.weight, fst_->Final(element.state));
-      FactorIterator siter(weight);
-      if (!(mode_ & kFactorFinalWeights) || siter.Done()) {
-        SetFinal(s, weight);
-      } else {
-        SetFinal(s, Weight::Zero());
-      }
-    }
-    return CacheImpl<Arc>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumOutputEpsilons(s);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && fst_->Properties(kError, false)) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<Arc>::InitArcIterator(s, data);
-  }
-
-  // Finds state corresponding to an element, creating new state if element not
-  // found.
-  StateId FindState(const Element &element) {
-    if (!(mode_ & kFactorArcWeights) && element.weight == Weight::One() &&
-        element.state != kNoStateId) {
-      while (unfactored_.size() <= element.state)
-        unfactored_.push_back(kNoStateId);
-      if (unfactored_[element.state] == kNoStateId) {
-        unfactored_[element.state] = elements_.size();
-        elements_.push_back(element);
-      }
-      return unfactored_[element.state];
-    } else {
-      const auto insert_result =
-          element_map_.insert(std::make_pair(element, elements_.size()));
-      if (insert_result.second) {
-        elements_.push_back(element);
-      }
-      return insert_result.first->second;
-    }
-  }
-
-  // Computes the outgoing transitions from a state, creating new destination
-  // states as needed.
-  void Expand(StateId s) {
-    const auto element = elements_[s];
-    if (element.state != kNoStateId) {
-      for (ArcIterator<Fst<Arc>> ait(*fst_, element.state); !ait.Done();
-           ait.Next()) {
-        const auto &arc = ait.Value();
-        auto weight = Times(element.weight, arc.weight);
-        FactorIterator fiter(weight);
-        if (!(mode_ & kFactorArcWeights) || fiter.Done()) {
-          const auto dest = FindState(Element(arc.nextstate, Weight::One()));
-          EmplaceArc(s, arc.ilabel, arc.olabel, std::move(weight), dest);
-        } else {
-          for (; !fiter.Done(); fiter.Next()) {
-            auto pair = fiter.Value();
-            const auto dest =
-                FindState(Element(arc.nextstate, pair.second.Quantize(delta_)));
-            EmplaceArc(s, arc.ilabel, arc.olabel, std::move(pair.first), dest);
-          }
-        }
-      }
-    }
-    if ((mode_ & kFactorFinalWeights) &&
-        ((element.state == kNoStateId) ||
-         (fst_->Final(element.state) != Weight::Zero()))) {
-      const auto weight =
-          element.state == kNoStateId
-              ? element.weight
-              : Times(element.weight, fst_->Final(element.state));
-      auto ilabel = final_ilabel_;
-      auto olabel = final_olabel_;
-      for (FactorIterator fiter(weight); !fiter.Done(); fiter.Next()) {
-        auto pair = fiter.Value();
-        const auto dest =
-            FindState(Element(kNoStateId, pair.second.Quantize(delta_)));
-        EmplaceArc(s, ilabel, olabel, std::move(pair.first), dest);
-        if (increment_final_ilabel_) ++ilabel;
-        if (increment_final_olabel_) ++olabel;
-      }
-    }
-    SetArcs(s);
-  }
-
- private:
-  // Equality function for Elements, assume weights have been quantized.
-  class ElementEqual {
-   public:
-    bool operator()(const Element &x, const Element &y) const {
-      return x.state == y.state && x.weight == y.weight;
-    }
-  };
-
-  // Hash function for Elements to Fst states.
-  class ElementKey {
-   public:
-    size_t operator()(const Element &x) const {
-      static constexpr auto prime = 7853;
-      return static_cast<size_t>(x.state * prime + x.weight.Hash());
-    }
-  };
-
-  using ElementMap =
-      std::unordered_map<Element, StateId, ElementKey, ElementEqual>;
-
-  std::unique_ptr<const Fst<Arc>> fst_;
-  float delta_;
-  uint32 mode_;         // Factoring arc and/or final weights.
-  Label final_ilabel_;  // ilabel of arc created when factoring final weights.
-  Label final_olabel_;  // olabel of arc created when factoring final weights.
-  bool increment_final_ilabel_;    // When factoring final weights results in
-  bool increment_final_olabel_;    // mutiple arcs, increment labels?
-  std::vector<Element> elements_;  // mapping from FST state to Element.
-  ElementMap element_map_;         // mapping from Element to FST state.
-  // Mapping between old/new StateId for states that do not need to be factored
-  // when mode_ is 0 or kFactorFinalWeights.
-  std::vector<StateId> unfactored_;
-};
-
-}  // namespace internal
-
-// FactorWeightFst takes as template parameter a FactorIterator as defined
-// above. The result of weight factoring is a transducer equivalent to the
-// input whose path weights have been factored according to the FactorIterator.
-// States and transitions will be added as necessary. The algorithm is a
-// generalization to arbitrary weights of the second step of the input
-// epsilon-normalization algorithm.
-//
-// This class attaches interface to implementation and handles reference
-// counting, delegating most methods to ImplToFst.
-template <class A, class FactorIterator>
-class FactorWeightFst
-    : public ImplToFst<internal::FactorWeightFstImpl<A, FactorIterator>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::FactorWeightFstImpl<Arc, FactorIterator>;
-
-  friend class ArcIterator<FactorWeightFst<Arc, FactorIterator>>;
-  friend class StateIterator<FactorWeightFst<Arc, FactorIterator>>;
-
-  explicit FactorWeightFst(const Fst<Arc> &fst)
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, FactorWeightOptions<Arc>())) {}
-
-  FactorWeightFst(const Fst<Arc> &fst, const FactorWeightOptions<Arc> &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  FactorWeightFst(const FactorWeightFst<Arc, FactorIterator> &fst, bool copy)
-      : ImplToFst<Impl>(fst, copy) {}
-
-  // Get a copy of this FactorWeightFst. See Fst<>::Copy() for further doc.
-  FactorWeightFst<Arc, FactorIterator> *Copy(bool copy = false) const override {
-    return new FactorWeightFst<Arc, FactorIterator>(*this, copy);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  FactorWeightFst &operator=(const FactorWeightFst &) = delete;
-};
-
-// Specialization for FactorWeightFst.
-template <class Arc, class FactorIterator>
-class StateIterator<FactorWeightFst<Arc, FactorIterator>>
-    : public CacheStateIterator<FactorWeightFst<Arc, FactorIterator>> {
- public:
-  explicit StateIterator(const FactorWeightFst<Arc, FactorIterator> &fst)
-      : CacheStateIterator<FactorWeightFst<Arc, FactorIterator>>(
-            fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for FactorWeightFst.
-template <class Arc, class FactorIterator>
-class ArcIterator<FactorWeightFst<Arc, FactorIterator>>
-    : public CacheArcIterator<FactorWeightFst<Arc, FactorIterator>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const FactorWeightFst<Arc, FactorIterator> &fst, StateId s)
-      : CacheArcIterator<FactorWeightFst<Arc, FactorIterator>>(
-            fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc, class FactorIterator>
-inline void FactorWeightFst<Arc, FactorIterator>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<FactorWeightFst<Arc, FactorIterator>>(*this);
-}
-
-}  // namespace fst
-
-#endif  // FST_FACTOR_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/filter-state.h b/projects/llm_framework/include/fst/filter-state.h
deleted file mode 100644
index 289b628e..00000000
--- a/projects/llm_framework/include/fst/filter-state.h
+++ /dev/null
@@ -1,199 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes for storing filter state in various algorithms like composition.
-
-#ifndef FST_FILTER_STATE_H_
-#define FST_FILTER_STATE_H_
-
-#include <forward_list>
-#include <utility>
-
-#include <fst/fst-decl.h>  // For optional argument declarations
-#include <fst/fst.h>
-#include <fst/matcher.h>
-
-
-namespace fst {
-
-// The filter state interface represents the state of a (e.g., composition)
-// filter.
-//
-// class FilterState {
-//  public:
-//   // Required constructors.
-//
-//   FilterState();
-//
-//   FilterState(const FilterState &fs);
-//
-//   // An invalid filter state.
-//   static const FilterState NoState();
-//
-//   // Maps state to integer for hashing.
-//   size_t Hash() const;
-//
-//   // Equality of filter states.
-//   bool operator==(const FilterState &fs) const;
-//
-//   // Inequality of filter states.
-//   bool operator!=(const FilterState &fs) const;
-//
-//   // Assignment to filter states.
-//   FilterState &operator=(const FilterState& fs);
-// };
-
-// Filter state that is a signed integral type.
-template <typename T>
-class IntegerFilterState {
- public:
-  IntegerFilterState() : state_(kNoStateId) {}
-
-  explicit IntegerFilterState(T s) : state_(s) {}
-
-  static const IntegerFilterState NoState() { return IntegerFilterState(); }
-
-  size_t Hash() const { return static_cast<size_t>(state_); }
-
-  bool operator==(const IntegerFilterState &fs) const {
-    return state_ == fs.state_;
-  }
-
-  bool operator!=(const IntegerFilterState &fs) const {
-    return state_ != fs.state_;
-  }
-
-  T GetState() const { return state_; }
-
-  void SetState(T state) { state_ = state; }
-
- private:
-  T state_;
-};
-
-using CharFilterState = IntegerFilterState<signed char>;
-using ShortFilterState = IntegerFilterState<short>;  // NOLINT
-using IntFilterState = IntegerFilterState<int>;
-
-// Filter state that is a weight (class).
-template <class W>
-class WeightFilterState {
- public:
-  WeightFilterState() : weight_(W::Zero()) {}
-
-  explicit WeightFilterState(W weight) : weight_(std::move(weight)) {}
-
-  static const WeightFilterState NoState() { return WeightFilterState(); }
-
-  size_t Hash() const { return weight_.Hash(); }
-
-  bool operator==(const WeightFilterState &fs) const {
-    return weight_ == fs.weight_;
-  }
-
-  bool operator!=(const WeightFilterState &fs) const {
-    return weight_ != fs.weight_;
-  }
-
-  W GetWeight() const { return weight_; }
-
-  void SetWeight(W weight) { weight_ = std::move(weight); }
-
- private:
-  W weight_;
-};
-
-// Filter state is a list of signed integer types T. Order matters
-// for equality.
-template <typename T>
-class ListFilterState {
- public:
-  ListFilterState() {}
-
-  explicit ListFilterState(T s) { list_.push_front(s); }
-
-  static const ListFilterState NoState() { return ListFilterState(kNoStateId); }
-
-  size_t Hash() const {
-    size_t h = 0;
-    for (const auto &elem : list_) h ^= h << 1 ^ elem;
-    return h;
-  }
-
-  bool operator==(const ListFilterState &fs) const { return list_ == fs.list_; }
-
-  bool operator!=(const ListFilterState &fs) const { return list_ != fs.list_; }
-
-  const std::forward_list<T> &GetState() const { return list_; }
-
-  std::forward_list<T> *GetMutableState() { return &list_; }
-
-  void SetState(const std::forward_list<T> &state) { list_ = state; }
-
- private:
-  std::forward_list<T> list_;
-};
-
-// Filter state that is the combination of two filter states.
-template <class FS1, class FS2>
-class PairFilterState {
- public:
-  PairFilterState() : fs1_(FS1::NoState()), fs2_(FS2::NoState()) {}
-
-  PairFilterState(const FS1 &fs1, const FS2 &fs2) : fs1_(fs1), fs2_(fs2) {}
-
-  static const PairFilterState NoState() { return PairFilterState(); }
-
-  size_t Hash() const {
-    const auto h1 = fs1_.Hash();
-    static constexpr auto lshift = 5;
-    static constexpr auto rshift = CHAR_BIT * sizeof(size_t) - 5;
-    return h1 << lshift ^ h1 >> rshift ^ fs2_.Hash();
-  }
-
-  bool operator==(const PairFilterState &fs) const {
-    return fs1_ == fs.fs1_ && fs2_ == fs.fs2_;
-  }
-
-  bool operator!=(const PairFilterState &fs) const {
-    return fs1_ != fs.fs1_ || fs2_ != fs.fs2_;
-  }
-
-  const FS1 &GetState1() const { return fs1_; }
-
-  const FS2 &GetState2() const { return fs2_; }
-
-  void SetState(const FS1 &fs1, const FS2 &fs2) {
-    fs1_ = fs1;
-    fs2_ = fs2;
-  }
-
- private:
-  FS1 fs1_;
-  FS2 fs2_;
-};
-
-// Single non-blocking filter state.
-class TrivialFilterState {
- public:
-  explicit TrivialFilterState(bool state = false) : state_(state) {}
-
-  static const TrivialFilterState NoState() { return TrivialFilterState(); }
-
-  size_t Hash() const { return 0; }
-
-  bool operator==(const TrivialFilterState &fs) const {
-    return state_ == fs.state_;
-  }
-
-  bool operator!=(const TrivialFilterState &fs) const {
-    return state_ != fs.state_;
-  }
-
- private:
-  bool state_;
-};
-
-}  // namespace fst
-
-#endif  // FST_FILTER_STATE_H_
diff --git a/projects/llm_framework/include/fst/flags.h b/projects/llm_framework/include/fst/flags.h
deleted file mode 100644
index 6ef1257f..00000000
--- a/projects/llm_framework/include/fst/flags.h
+++ /dev/null
@@ -1,219 +0,0 @@
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Google-style flag handling declarations and inline definitions.
-
-#ifndef FST_LIB_FLAGS_H_
-#define FST_LIB_FLAGS_H_
-
-#include <cstdlib>
-
-#include <iostream>
-#include <map>
-#include <set>
-#include <sstream>
-#include <string>
-
-#include <fst/lock.h>
-#include <fst/types.h>
-
-#include "gflags/gflags.h"
-#include "glog/logging.h"
-
-using std::string;
-
-// FLAGS USAGE:
-//
-// Definition example:
-//
-//    DEFINE_int32(length, 0, "length");
-//
-// This defines variable FLAGS_length, initialized to 0.
-//
-// Declaration example:
-//
-//    DECLARE_int32(length);
-//
-// SET_FLAGS() can be used to set flags from the command line
-// using, for example, '--length=2'.
-//
-// ShowUsage() can be used to print out command and flag usage.
-
-// #define DECLARE_bool(name) extern bool FLAGS_ ## name
-// #define DECLARE_string(name) extern string FLAGS_ ## name
-// #define DECLARE_int32(name) extern int32 FLAGS_ ## name
-// #define DECLARE_int64(name) extern int64 FLAGS_ ## name
-// #define DECLARE_double(name) extern double FLAGS_ ## name
-
-template <typename T>
-struct FlagDescription {
-  FlagDescription(T* addr, const char* doc, const char* type, const char* file,
-                  const T val)
-      : address(addr),
-        doc_string(doc),
-        type_name(type),
-        file_name(file),
-        default_value(val) {}
-
-  T* address;
-  const char* doc_string;
-  const char* type_name;
-  const char* file_name;
-  const T default_value;
-};
-
-template <typename T>
-class FlagRegister {
- public:
-  static FlagRegister<T>* GetRegister() {
-    static auto reg = new FlagRegister<T>;
-    return reg;
-  }
-
-  const FlagDescription<T>& GetFlagDescription(const string& name) const {
-    fst::MutexLock l(&flag_lock_);
-    auto it = flag_table_.find(name);
-    return it != flag_table_.end() ? it->second : 0;
-  }
-
-  void SetDescription(const string& name, const FlagDescription<T>& desc) {
-    fst::MutexLock l(&flag_lock_);
-    flag_table_.insert(make_pair(name, desc));
-  }
-
-  bool SetFlag(const string& val, bool* address) const {
-    if (val == "true" || val == "1" || val.empty()) {
-      *address = true;
-      return true;
-    } else if (val == "false" || val == "0") {
-      *address = false;
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  bool SetFlag(const string& val, string* address) const {
-    *address = val;
-    return true;
-  }
-
-  bool SetFlag(const string& val, int32* address) const {
-    char* p = 0;
-    *address = strtol(val.c_str(), &p, 0);
-    return !val.empty() && *p == '\0';
-  }
-
-  bool SetFlag(const string& val, int64* address) const {
-    char* p = 0;
-    *address = strtoll(val.c_str(), &p, 0);
-    return !val.empty() && *p == '\0';
-  }
-
-  bool SetFlag(const string& val, double* address) const {
-    char* p = 0;
-    *address = strtod(val.c_str(), &p);
-    return !val.empty() && *p == '\0';
-  }
-
-  bool SetFlag(const string& arg, const string& val) const {
-    for (typename std::map<string, FlagDescription<T>>::const_iterator it =
-             flag_table_.begin();
-         it != flag_table_.end(); ++it) {
-      const string& name = it->first;
-      const FlagDescription<T>& desc = it->second;
-      if (arg == name) return SetFlag(val, desc.address);
-    }
-    return false;
-  }
-
-  void GetUsage(std::set<std::pair<string, string>>* usage_set) const {
-    for (auto it = flag_table_.begin(); it != flag_table_.end(); ++it) {
-      const string& name = it->first;
-      const FlagDescription<T>& desc = it->second;
-      string usage = "  --" + name;
-      usage += ": type = ";
-      usage += desc.type_name;
-      usage += ", default = ";
-      usage += GetDefault(desc.default_value) + "\n  ";
-      usage += desc.doc_string;
-      usage_set->insert(make_pair(desc.file_name, usage));
-    }
-  }
-
- private:
-  string GetDefault(bool default_value) const {
-    return default_value ? "true" : "false";
-  }
-
-  string GetDefault(const string& default_value) const {
-    return "\"" + default_value + "\"";
-  }
-
-  template <class V>
-  string GetDefault(const V& default_value) const {
-    std::ostringstream strm;
-    strm << default_value;
-    return strm.str();
-  }
-
-  mutable fst::Mutex flag_lock_;  // Multithreading lock.
-  std::map<string, FlagDescription<T>> flag_table_;
-};
-
-template <typename T>
-class FlagRegisterer {
- public:
-  FlagRegisterer(const string& name, const FlagDescription<T>& desc) {
-    auto registr = FlagRegister<T>::GetRegister();
-    registr->SetDescription(name, desc);
-  }
-
- private:
-  FlagRegisterer(const FlagRegisterer&) = delete;
-  FlagRegisterer& operator=(const FlagRegisterer&) = delete;
-};
-
-#define DEFINE_VAR(type, name, value, doc)             \
-  type FLAGS_##name = value;                           \
-  static FlagRegisterer<type> name##_flags_registerer( \
-      #name,                                           \
-      FlagDescription<type>(&FLAGS_##name, doc, #type, __FILE__, value))
-
-// #define DEFINE_bool(name, value, doc) DEFINE_VAR(bool, name, value, doc)
-// #define DEFINE_string(name, value, doc) \
-//   DEFINE_VAR(string, name, value, doc)
-// #define DEFINE_int32(name, value, doc) DEFINE_VAR(int32, name, value, doc)
-// #define DEFINE_int64(name, value, doc) DEFINE_VAR(int64, name, value, doc)
-// #define DEFINE_double(name, value, doc) DEFINE_VAR(double, name, value, doc)
-
-// Temporary directory.
-DECLARE_string(tmpdir);
-
-void SetFlags(const char* usage, int* argc, char*** argv, bool remove_flags,
-              const char* src = "");
-
-#define SET_FLAGS(usage, argc, argv, rmflags) \
-  gflags::ParseCommandLineFlags(argc, argv, true)
-// SetFlags(usage, argc, argv, rmflags, __FILE__)
-
-// Deprecated; for backward compatibility.
-inline void InitFst(const char* usage, int* argc, char*** argv, bool rmflags) {
-  return SetFlags(usage, argc, argv, rmflags);
-}
-
-void ShowUsage(bool long_usage = true);
-
-#endif  // FST_LIB_FLAGS_H_
diff --git a/projects/llm_framework/include/fst/float-weight.h b/projects/llm_framework/include/fst/float-weight.h
deleted file mode 100644
index 8fade684..00000000
--- a/projects/llm_framework/include/fst/float-weight.h
+++ /dev/null
@@ -1,858 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Float weight set and associated semiring operation definitions.
-
-#ifndef FST_FLOAT_WEIGHT_H_
-#define FST_FLOAT_WEIGHT_H_
-
-#include <algorithm>
-#include <climits>
-#include <cmath>
-#include <cstdlib>
-#include <cstring>
-#include <limits>
-#include <sstream>
-#include <string>
-#include <type_traits>
-
-#include <fst/util.h>
-#include <fst/weight.h>
-
-
-namespace fst {
-
-// Numeric limits class.
-template <class T>
-class FloatLimits {
- public:
-  static constexpr T PosInfinity() {
-    return std::numeric_limits<T>::infinity();
-  }
-
-  static constexpr T NegInfinity() { return -PosInfinity(); }
-
-  static constexpr T NumberBad() { return std::numeric_limits<T>::quiet_NaN(); }
-};
-
-// Weight class to be templated on floating-points types.
-template <class T = float>
-class FloatWeightTpl {
- public:
-  using ValueType = T;
-
-  FloatWeightTpl() noexcept {}
-
-  constexpr FloatWeightTpl(T f) : value_(f) {}  // NOLINT
-
-  // TODO(mjansche): Leave implicit once Android NDK r18 is the default.
-  FloatWeightTpl(const FloatWeightTpl<T> &) = default;
-  FloatWeightTpl(FloatWeightTpl<T> &&) noexcept = default;
-
-  FloatWeightTpl<T> &operator=(const FloatWeightTpl<T> &) = default;
-  FloatWeightTpl<T> &operator=(FloatWeightTpl<T> &&) noexcept = default;
-
-  std::istream &Read(std::istream &strm) { return ReadType(strm, &value_); }
-
-  std::ostream &Write(std::ostream &strm) const {
-    return WriteType(strm, value_);
-  }
-
-  size_t Hash() const {
-    size_t hash = 0;
-    // Avoid using union, which would be undefined behavior.
-    // Use memcpy, similar to bit_cast, but sizes may be different.
-    // This should be optimized into a single move instruction by
-    // any reasonable compiler.
-    std::memcpy(&hash, &value_, std::min(sizeof(hash), sizeof(value_)));
-    return hash;
-  }
-
-  constexpr const T &Value() const { return value_; }
-
- protected:
-  void SetValue(const T &f) { value_ = f; }
-
-  static constexpr const char *GetPrecisionString() {
-    return sizeof(T) == 4 ? ""
-        : sizeof(T) == 1 ? "8"
-        : sizeof(T) == 2 ? "16"
-        : sizeof(T) == 8 ? "64"
-        : "unknown";
-  }
-
- private:
-  T value_;
-};
-
-// Single-precision float weight.
-using FloatWeight = FloatWeightTpl<float>;
-
-template <class T>
-constexpr bool operator==(const FloatWeightTpl<T> &w1,
-                          const FloatWeightTpl<T> &w2) {
-#if (defined(__i386__) || defined(__x86_64__)) && !defined(__SSE2_MATH__)
-// With i387 instructions, excess precision on a weight in an 80-bit
-// register may cause it to compare unequal to that same weight when
-// stored to memory.  This breaks =='s reflexivity, in turn breaking
-// NaturalLess.
-#error "Please compile with -msse -mfpmath=sse, or equivalent."
-#endif
-  return w1.Value() == w2.Value();
-}
-
-// These seemingly unnecessary overloads are actually needed to make
-// comparisons like FloatWeightTpl<float> == float compile.  If only the
-// templated version exists, the FloatWeightTpl<float>(float) conversion
-// won't be found.
-constexpr bool operator==(const FloatWeightTpl<float> &w1,
-                          const FloatWeightTpl<float> &w2) {
-  return operator==<float>(w1, w2);
-}
-
-constexpr bool operator==(const FloatWeightTpl<double> &w1,
-                          const FloatWeightTpl<double> &w2) {
-  return operator==<double>(w1, w2);
-}
-
-template <class T>
-constexpr bool operator!=(const FloatWeightTpl<T> &w1,
-                          const FloatWeightTpl<T> &w2) {
-  return !(w1 == w2);
-}
-
-constexpr bool operator!=(const FloatWeightTpl<float> &w1,
-                          const FloatWeightTpl<float> &w2) {
-  return operator!=<float>(w1, w2);
-}
-
-constexpr bool operator!=(const FloatWeightTpl<double> &w1,
-                          const FloatWeightTpl<double> &w2) {
-  return operator!=<double>(w1, w2);
-}
-
-template <class T>
-constexpr bool FloatApproxEqual(T w1, T w2, float delta = kDelta) {
-  return w1 <= w2 + delta && w2 <= w1 + delta;
-}
-
-template <class T>
-constexpr bool ApproxEqual(const FloatWeightTpl<T> &w1,
-                           const FloatWeightTpl<T> &w2, float delta = kDelta) {
-  return FloatApproxEqual(w1.Value(), w2.Value(), delta);
-}
-
-template <class T>
-inline std::ostream &operator<<(std::ostream &strm,
-                                const FloatWeightTpl<T> &w) {
-  if (w.Value() == FloatLimits<T>::PosInfinity()) {
-    return strm << "Infinity";
-  } else if (w.Value() == FloatLimits<T>::NegInfinity()) {
-    return strm << "-Infinity";
-  } else if (w.Value() != w.Value()) {  // Fails for IEEE NaN.
-    return strm << "BadNumber";
-  } else {
-    return strm << w.Value();
-  }
-}
-
-template <class T>
-inline std::istream &operator>>(std::istream &strm, FloatWeightTpl<T> &w) {
-  string s;
-  strm >> s;
-  if (s == "Infinity") {
-    w = FloatWeightTpl<T>(FloatLimits<T>::PosInfinity());
-  } else if (s == "-Infinity") {
-    w = FloatWeightTpl<T>(FloatLimits<T>::NegInfinity());
-  } else {
-    char *p;
-    T f = strtod(s.c_str(), &p);
-    if (p < s.c_str() + s.size()) {
-      strm.clear(std::ios::badbit);
-    } else {
-      w = FloatWeightTpl<T>(f);
-    }
-  }
-  return strm;
-}
-
-// Tropical semiring: (min, +, inf, 0).
-template <class T>
-class TropicalWeightTpl : public FloatWeightTpl<T> {
- public:
-  using typename FloatWeightTpl<T>::ValueType;
-  using FloatWeightTpl<T>::Value;
-  using ReverseWeight = TropicalWeightTpl<T>;
-  using Limits = FloatLimits<T>;
-
-  TropicalWeightTpl() noexcept : FloatWeightTpl<T>() {}
-
-  constexpr TropicalWeightTpl(T f) : FloatWeightTpl<T>(f) {}
-
-  static constexpr TropicalWeightTpl<T> Zero() { return Limits::PosInfinity(); }
-
-  static constexpr TropicalWeightTpl<T> One() { return 0; }
-
-  static constexpr TropicalWeightTpl<T> NoWeight() {
-    return Limits::NumberBad();
-  }
-
-  static const string &Type() {
-    static const string *const type =
-        new string(string("tropical") +
-                   FloatWeightTpl<T>::GetPrecisionString());
-    return *type;
-  }
-
-  constexpr bool Member() const {
-    // All floating point values except for NaNs and negative infinity are valid
-    // tropical weights.
-    //
-    // Testing membership of a given value can be done by simply checking that
-    // it is strictly greater than negative infinity, which fails for negative
-    // infinity itself but also for NaNs. This can usually be accomplished in a
-    // single instruction (such as *UCOMI* on x86) without branching logic.
-    //
-    // An additional wrinkle involves constexpr correctness of floating point
-    // comparisons against NaN. GCC is uneven when it comes to which expressions
-    // it considers compile-time constants. In particular, current versions of
-    // GCC do not always consider (nan < inf) to be a constant expression, but
-    // do consider (inf < nan) to be a constant expression. (See
-    // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88173 and
-    // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88683 for details.) In order
-    // to allow Member() to be a constexpr function accepted by GCC, we write
-    // the comparison here as (-inf < v).
-    return Limits::NegInfinity() < Value();
-  }
-
-  TropicalWeightTpl<T> Quantize(float delta = kDelta) const {
-    if (!Member() || Value() == Limits::PosInfinity()) {
-      return *this;
-    } else {
-      return TropicalWeightTpl<T>(std::floor(Value() / delta + 0.5F) * delta);
-    }
-  }
-
-  constexpr TropicalWeightTpl<T> Reverse() const { return *this; }
-
-  static constexpr uint64 Properties() {
-    return kLeftSemiring | kRightSemiring | kCommutative | kPath | kIdempotent;
-  }
-};
-
-// Single precision tropical weight.
-using TropicalWeight = TropicalWeightTpl<float>;
-
-template <class T>
-constexpr TropicalWeightTpl<T> Plus(const TropicalWeightTpl<T> &w1,
-                                    const TropicalWeightTpl<T> &w2) {
-  return (!w1.Member() || !w2.Member()) ? TropicalWeightTpl<T>::NoWeight()
-      : w1.Value() < w2.Value() ? w1 : w2;
-}
-
-// See comment at operator==(FloatWeightTpl<float>, FloatWeightTpl<float>)
-// for why these overloads are present.
-constexpr TropicalWeightTpl<float> Plus(const TropicalWeightTpl<float> &w1,
-                                        const TropicalWeightTpl<float> &w2) {
-  return Plus<float>(w1, w2);
-}
-
-constexpr TropicalWeightTpl<double> Plus(const TropicalWeightTpl<double> &w1,
-                                         const TropicalWeightTpl<double> &w2) {
-  return Plus<double>(w1, w2);
-}
-
-template <class T>
-constexpr TropicalWeightTpl<T> Times(const TropicalWeightTpl<T> &w1,
-                                     const TropicalWeightTpl<T> &w2) {
-  // The following is safe in the context of the Tropical (and Log) semiring
-  // for all IEEE floating point values, including infinities and NaNs,
-  // because:
-  //
-  // * If one or both of the floating point Values is NaN and hence not a
-  //   Member, the result of addition below is NaN, so the result is not a
-  //   Member. This supersedes all other cases, so we only consider non-NaN
-  //   values next.
-  //
-  // * If both Values are finite, there is no issue.
-  //
-  // * If one of the Values is infinite, or if both are infinities with the
-  //   same sign, the result of floating point addition is the same infinity,
-  //   so there is no issue.
-  //
-  // * If both of the Values are infinities with opposite signs, the result of
-  //   adding IEEE floating point -inf + inf is NaN and hence not a Member. But
-  //   since -inf was not a Member to begin with, returning a non-Member result
-  //   is fine as well.
-  return TropicalWeightTpl<T>(w1.Value() + w2.Value());
-}
-
-constexpr TropicalWeightTpl<float> Times(const TropicalWeightTpl<float> &w1,
-                                         const TropicalWeightTpl<float> &w2) {
-  return Times<float>(w1, w2);
-}
-
-constexpr TropicalWeightTpl<double> Times(const TropicalWeightTpl<double> &w1,
-                                          const TropicalWeightTpl<double> &w2) {
-  return Times<double>(w1, w2);
-}
-
-template <class T>
-constexpr TropicalWeightTpl<T> Divide(const TropicalWeightTpl<T> &w1,
-                                      const TropicalWeightTpl<T> &w2,
-                                      DivideType typ = DIVIDE_ANY) {
-  // The following is safe in the context of the Tropical (and Log) semiring
-  // for all IEEE floating point values, including infinities and NaNs,
-  // because:
-  //
-  // * If one or both of the floating point Values is NaN and hence not a
-  //   Member, the result of subtraction below is NaN, so the result is not a
-  //   Member. This supersedes all other cases, so we only consider non-NaN
-  //   values next.
-  //
-  // * If both Values are finite, there is no issue.
-  //
-  // * If w2.Value() is -inf (and hence w2 is not a Member), the result of ?:
-  //   below is NoWeight, which is not a Member.
-  //
-  //   Whereas in IEEE floating point semantics 0/inf == 0, this does not carry
-  //   over to this semiring (since TropicalWeight(-inf) would be the analogue
-  //   of floating point inf) and instead Divide(Zero(), TropicalWeight(-inf))
-  //   is NoWeight().
-  //
-  // * If w2.Value() is inf (and hence w2 is Zero), the resulting floating
-  //   point value is either NaN (if w1 is Zero or if w1.Value() is NaN) and
-  //   hence not a Member, or it is -inf and hence not a Member; either way,
-  //   division by Zero results in a non-Member result.
-  using Weight = TropicalWeightTpl<T>;
-  return w2.Member() ? Weight(w1.Value() - w2.Value()) : Weight::NoWeight();
-}
-
-constexpr TropicalWeightTpl<float> Divide(const TropicalWeightTpl<float> &w1,
-                                          const TropicalWeightTpl<float> &w2,
-                                          DivideType typ = DIVIDE_ANY) {
-  return Divide<float>(w1, w2, typ);
-}
-
-constexpr TropicalWeightTpl<double> Divide(const TropicalWeightTpl<double> &w1,
-                                           const TropicalWeightTpl<double> &w2,
-                                           DivideType typ = DIVIDE_ANY) {
-  return Divide<double>(w1, w2, typ);
-}
-
-// Power(w, n) calculates the n-th power of w with respect to semiring Times.
-//
-// In the case of the Tropical (and Log) semiring, the exponent n is not
-// restricted to be an integer. It can be a floating point value, for example.
-//
-// In weight.h, a narrower and hence more broadly applicable version of
-// Power(w, n) is defined for arbitrary weight types and non-negative integer
-// exponents n (of type size_t) and implemented in terms of repeated
-// multiplication using Times.
-//
-// Without further provisions this means that, when an expression such as
-//
-//   Power(TropicalWeightTpl<float>::One(), static_cast<size_t>(2))
-//
-// is specified, the overload of Power() is ambiguous. The template function
-// below could be instantiated as
-//
-//   Power<float, size_t>(const TropicalWeightTpl<float> &, size_t)
-//
-// and the template function defined in weight.h (further specialized below)
-// could be instantiated as
-//
-//   Power<TropicalWeightTpl<float>>(const TropicalWeightTpl<float> &, size_t)
-//
-// That would lead to two definitions with identical signatures, which results
-// in a compilation error. To avoid that, we hide the definition of Power<T, V>
-// when V is size_t, so only Power<W> is visible. Power<W> is further
-// specialized to Power<TropicalWeightTpl<...>>, and the overloaded definition
-// of Power<T, V> is made conditionally available only to that template
-// specialization.
-
-template <class T, class V,
-          bool Enable = !std::is_same<V, size_t>::value,
-          typename std::enable_if<Enable>::type * = nullptr>
-constexpr TropicalWeightTpl<T> Power(const TropicalWeightTpl<T> &w, V n) {
-  using Weight = TropicalWeightTpl<T>;
-  return (!w.Member() || n != n) ? Weight::NoWeight()
-      : (n == 0 || w == Weight::One()) ? Weight::One()
-      : Weight(w.Value() * n);
-}
-
-// Specializes the library-wide template to use the above implementation; rules
-// of function template instantiation require this be a full instantiation.
-
-template <>
-constexpr TropicalWeightTpl<float> Power<TropicalWeightTpl<float>>(
-    const TropicalWeightTpl<float> &weight, size_t n) {
-  return Power<float, size_t, true>(weight, n);
-}
-
-template <>
-constexpr TropicalWeightTpl<double> Power<TropicalWeightTpl<double>>(
-    const TropicalWeightTpl<double> &weight, size_t n) {
-  return Power<double, size_t, true>(weight, n);
-}
-
-
-// Log semiring: (log(e^-x + e^-y), +, inf, 0).
-template <class T>
-class LogWeightTpl : public FloatWeightTpl<T> {
- public:
-  using typename FloatWeightTpl<T>::ValueType;
-  using FloatWeightTpl<T>::Value;
-  using ReverseWeight = LogWeightTpl;
-  using Limits = FloatLimits<T>;
-
-  LogWeightTpl() noexcept : FloatWeightTpl<T>() {}
-
-  constexpr LogWeightTpl(T f) : FloatWeightTpl<T>(f) {}
-
-  static constexpr LogWeightTpl Zero() { return Limits::PosInfinity(); }
-
-  static constexpr LogWeightTpl One() { return 0; }
-
-  static constexpr LogWeightTpl NoWeight() { return Limits::NumberBad(); }
-
-  static const string &Type() {
-    static const string *const type =
-        new string(string("log") + FloatWeightTpl<T>::GetPrecisionString());
-    return *type;
-  }
-
-  constexpr bool Member() const {
-    // The comments for TropicalWeightTpl<>::Member() apply here unchanged.
-    return Limits::NegInfinity() < Value();
-  }
-
-  LogWeightTpl<T> Quantize(float delta = kDelta) const {
-    if (!Member() || Value() == Limits::PosInfinity()) {
-      return *this;
-    } else {
-      return LogWeightTpl<T>(std::floor(Value() / delta + 0.5F) * delta);
-    }
-  }
-
-  constexpr LogWeightTpl<T> Reverse() const { return *this; }
-
-  static constexpr uint64 Properties() {
-    return kLeftSemiring | kRightSemiring | kCommutative;
-  }
-};
-
-// Single-precision log weight.
-using LogWeight = LogWeightTpl<float>;
-
-// Double-precision log weight.
-using Log64Weight = LogWeightTpl<double>;
-
-namespace internal {
-
-// -log(e^-x + e^-y) = x - LogPosExp(y - x), assuming x >= 0.0.
-inline double LogPosExp(double x) {
-  DCHECK(!(x < 0));  // NB: NaN values are allowed.
-  return log1p(exp(-x));
-}
-
-// -log(e^-x - e^-y) = x - LogNegExp(y - x), assuming x > 0.0.
-inline double LogNegExp(double x) {
-  DCHECK_GT(x, 0);
-  return log1p(-exp(-x));
-}
-
-// a +_log b = -log(e^-a + e^-b) = KahanLogSum(a, b, ...).
-// Kahan compensated summation provides an error bound that is
-// independent of the number of addends. Assumes b >= a;
-// c is the compensation.
-inline double KahanLogSum(double a, double b, double *c) {
-  DCHECK_GE(b, a);
-  double y = -LogPosExp(b - a) - *c;
-  double t = a + y;
-  *c = (t - a) - y;
-  return t;
-}
-
-// a -_log b = -log(e^-a - e^-b) = KahanLogDiff(a, b, ...).
-// Kahan compensated summation provides an error bound that is
-// independent of the number of addends. Assumes b > a;
-// c is the compensation.
-inline double KahanLogDiff(double a, double b, double *c) {
-  DCHECK_GT(b, a);
-  double y = -LogNegExp(b - a) - *c;
-  double t = a + y;
-  *c = (t - a) - y;
-  return t;
-}
-
-}  // namespace internal
-
-template <class T>
-inline LogWeightTpl<T> Plus(const LogWeightTpl<T> &w1,
-                            const LogWeightTpl<T> &w2) {
-  using Limits = FloatLimits<T>;
-  const T f1 = w1.Value();
-  const T f2 = w2.Value();
-  if (f1 == Limits::PosInfinity()) {
-    return w2;
-  } else if (f2 == Limits::PosInfinity()) {
-    return w1;
-  } else if (f1 > f2) {
-    return LogWeightTpl<T>(f2 - internal::LogPosExp(f1 - f2));
-  } else {
-    return LogWeightTpl<T>(f1 - internal::LogPosExp(f2 - f1));
-  }
-}
-
-inline LogWeightTpl<float> Plus(const LogWeightTpl<float> &w1,
-                                const LogWeightTpl<float> &w2) {
-  return Plus<float>(w1, w2);
-}
-
-inline LogWeightTpl<double> Plus(const LogWeightTpl<double> &w1,
-                                 const LogWeightTpl<double> &w2) {
-  return Plus<double>(w1, w2);
-}
-
-template <class T>
-constexpr LogWeightTpl<T> Times(const LogWeightTpl<T> &w1,
-                                const LogWeightTpl<T> &w2) {
-  // The comments for Times(Tropical...) above apply here unchanged.
-  return LogWeightTpl<T>(w1.Value() + w2.Value());
-}
-
-constexpr LogWeightTpl<float> Times(const LogWeightTpl<float> &w1,
-                                    const LogWeightTpl<float> &w2) {
-  return Times<float>(w1, w2);
-}
-
-constexpr LogWeightTpl<double> Times(const LogWeightTpl<double> &w1,
-                                     const LogWeightTpl<double> &w2) {
-  return Times<double>(w1, w2);
-}
-
-template <class T>
-constexpr LogWeightTpl<T> Divide(const LogWeightTpl<T> &w1,
-                                 const LogWeightTpl<T> &w2,
-                                 DivideType typ = DIVIDE_ANY) {
-  // The comments for Divide(Tropical...) above apply here unchanged.
-  using Weight = LogWeightTpl<T>;
-  return w2.Member() ? Weight(w1.Value() - w2.Value()) : Weight::NoWeight();
-}
-
-constexpr LogWeightTpl<float> Divide(const LogWeightTpl<float> &w1,
-                                     const LogWeightTpl<float> &w2,
-                                     DivideType typ = DIVIDE_ANY) {
-  return Divide<float>(w1, w2, typ);
-}
-
-constexpr LogWeightTpl<double> Divide(const LogWeightTpl<double> &w1,
-                                      const LogWeightTpl<double> &w2,
-                                      DivideType typ = DIVIDE_ANY) {
-  return Divide<double>(w1, w2, typ);
-}
-
-// The comments for Power<>(Tropical...) above apply here unchanged.
-
-template <class T, class V,
-          bool Enable = !std::is_same<V, size_t>::value,
-          typename std::enable_if<Enable>::type * = nullptr>
-constexpr LogWeightTpl<T> Power(const LogWeightTpl<T> &w, V n) {
-  using Weight = LogWeightTpl<T>;
-  return (!w.Member() || n != n) ? Weight::NoWeight()
-      : (n == 0 || w == Weight::One()) ? Weight::One()
-      : Weight(w.Value() * n);
-}
-
-// Specializes the library-wide template to use the above implementation; rules
-// of function template instantiation require this be a full instantiation.
-
-template <>
-constexpr LogWeightTpl<float> Power<LogWeightTpl<float>>(
-    const LogWeightTpl<float> &weight, size_t n) {
-  return Power<float, size_t, true>(weight, n);
-}
-
-template <>
-constexpr LogWeightTpl<double> Power<LogWeightTpl<double>>(
-    const LogWeightTpl<double> &weight, size_t n) {
-  return Power<double, size_t, true>(weight, n);
-}
-
-// Specialization using the Kahan compensated summation.
-template <class T>
-class Adder<LogWeightTpl<T>> {
- public:
-  using Weight = LogWeightTpl<T>;
-
-  explicit Adder(Weight w = Weight::Zero())
-      : sum_(w.Value()),
-        c_(0.0) { }
-
-  Weight Add(const Weight &w) {
-    using Limits = FloatLimits<T>;
-    const T f = w.Value();
-    if (f == Limits::PosInfinity()) {
-      return Sum();
-    } else if (sum_ == Limits::PosInfinity()) {
-      sum_ = f;
-      c_ = 0.0;
-    } else if (f > sum_) {
-      sum_ = internal::KahanLogSum(sum_, f, &c_);
-    } else {
-      sum_ = internal::KahanLogSum(f, sum_, &c_);
-    }
-    return Sum();
-  }
-
-  Weight Sum() { return Weight(sum_); }
-
-  void Reset(Weight w = Weight::Zero()) {
-    sum_ = w.Value();
-    c_ = 0.0;
-  }
-
- private:
-  double sum_;
-  double c_;   // Kahan compensation.
-};
-
-// MinMax semiring: (min, max, inf, -inf).
-template <class T>
-class MinMaxWeightTpl : public FloatWeightTpl<T> {
- public:
-  using typename FloatWeightTpl<T>::ValueType;
-  using FloatWeightTpl<T>::Value;
-  using ReverseWeight = MinMaxWeightTpl<T>;
-  using Limits = FloatLimits<T>;
-
-  MinMaxWeightTpl() noexcept : FloatWeightTpl<T>() {}
-
-  constexpr MinMaxWeightTpl(T f) : FloatWeightTpl<T>(f) {}  // NOLINT
-
-  static constexpr MinMaxWeightTpl Zero() { return Limits::PosInfinity(); }
-
-  static constexpr MinMaxWeightTpl One() { return Limits::NegInfinity(); }
-
-  static constexpr MinMaxWeightTpl NoWeight() { return Limits::NumberBad(); }
-
-  static const string &Type() {
-    static const string *const type =
-        new string(string("minmax") + FloatWeightTpl<T>::GetPrecisionString());
-    return *type;
-  }
-
-  // Fails for IEEE NaN.
-  constexpr bool Member() const { return Value() == Value(); }
-
-  MinMaxWeightTpl<T> Quantize(float delta = kDelta) const {
-    // If one of infinities, or a NaN.
-    if (!Member() ||
-        Value() == Limits::NegInfinity() || Value() == Limits::PosInfinity()) {
-      return *this;
-    } else {
-      return MinMaxWeightTpl<T>(std::floor(Value() / delta + 0.5F) * delta);
-    }
-  }
-
-  constexpr MinMaxWeightTpl<T> Reverse() const { return *this; }
-
-  static constexpr uint64 Properties() {
-    return kLeftSemiring | kRightSemiring | kCommutative | kIdempotent | kPath;
-  }
-};
-
-// Single-precision min-max weight.
-using MinMaxWeight = MinMaxWeightTpl<float>;
-
-// Min.
-template <class T>
-constexpr MinMaxWeightTpl<T> Plus(const MinMaxWeightTpl<T> &w1,
-                               const MinMaxWeightTpl<T> &w2) {
-  return (!w1.Member() || !w2.Member()) ? MinMaxWeightTpl<T>::NoWeight()
-      : w1.Value() < w2.Value() ? w1 : w2;
-}
-
-constexpr MinMaxWeightTpl<float> Plus(const MinMaxWeightTpl<float> &w1,
-                                      const MinMaxWeightTpl<float> &w2) {
-  return Plus<float>(w1, w2);
-}
-
-constexpr MinMaxWeightTpl<double> Plus(const MinMaxWeightTpl<double> &w1,
-                                       const MinMaxWeightTpl<double> &w2) {
-  return Plus<double>(w1, w2);
-}
-
-// Max.
-template <class T>
-constexpr MinMaxWeightTpl<T> Times(const MinMaxWeightTpl<T> &w1,
-                                   const MinMaxWeightTpl<T> &w2) {
-  return (!w1.Member() || !w2.Member()) ? MinMaxWeightTpl<T>::NoWeight()
-      : w1.Value() >= w2.Value() ? w1 : w2;
-}
-
-constexpr MinMaxWeightTpl<float> Times(const MinMaxWeightTpl<float> &w1,
-                                       const MinMaxWeightTpl<float> &w2) {
-  return Times<float>(w1, w2);
-}
-
-constexpr MinMaxWeightTpl<double> Times(const MinMaxWeightTpl<double> &w1,
-                                        const MinMaxWeightTpl<double> &w2) {
-  return Times<double>(w1, w2);
-}
-
-// Defined only for special cases.
-template <class T>
-constexpr MinMaxWeightTpl<T> Divide(const MinMaxWeightTpl<T> &w1,
-                                    const MinMaxWeightTpl<T> &w2,
-                                    DivideType typ = DIVIDE_ANY) {
-  return w1.Value() >= w2.Value() ? w1 : MinMaxWeightTpl<T>::NoWeight();
-}
-
-constexpr MinMaxWeightTpl<float> Divide(const MinMaxWeightTpl<float> &w1,
-                                        const MinMaxWeightTpl<float> &w2,
-                                        DivideType typ = DIVIDE_ANY) {
-  return Divide<float>(w1, w2, typ);
-}
-
-constexpr MinMaxWeightTpl<double> Divide(const MinMaxWeightTpl<double> &w1,
-                                         const MinMaxWeightTpl<double> &w2,
-                                         DivideType typ = DIVIDE_ANY) {
-  return Divide<double>(w1, w2, typ);
-}
-
-// Converts to tropical.
-template <>
-struct WeightConvert<LogWeight, TropicalWeight> {
-  constexpr TropicalWeight operator()(const LogWeight &w) const {
-    return w.Value();
-  }
-};
-
-template <>
-struct WeightConvert<Log64Weight, TropicalWeight> {
-  constexpr TropicalWeight operator()(const Log64Weight &w) const {
-    return w.Value();
-  }
-};
-
-// Converts to log.
-template <>
-struct WeightConvert<TropicalWeight, LogWeight> {
-  constexpr LogWeight operator()(const TropicalWeight &w) const {
-    return w.Value();
-  }
-};
-
-template <>
-struct WeightConvert<Log64Weight, LogWeight> {
-  constexpr LogWeight operator()(const Log64Weight &w) const {
-    return w.Value();
-  }
-};
-
-// Converts to log64.
-template <>
-struct WeightConvert<TropicalWeight, Log64Weight> {
-  constexpr Log64Weight operator()(const TropicalWeight &w) const {
-    return w.Value();
-  }
-};
-
-template <>
-struct WeightConvert<LogWeight, Log64Weight> {
-  constexpr Log64Weight operator()(const LogWeight &w) const {
-    return w.Value();
-  }
-};
-
-// This function object returns random integers chosen from [0,
-// num_random_weights). The boolean 'allow_zero' determines whether Zero() and
-// zero divisors should be returned in the random weight generation. This is
-// intended primary for testing.
-template <class Weight>
-class FloatWeightGenerate {
- public:
-  explicit FloatWeightGenerate(
-      bool allow_zero = true,
-      const size_t num_random_weights = kNumRandomWeights)
-      : allow_zero_(allow_zero), num_random_weights_(num_random_weights) {}
-
-  Weight operator()() const {
-    const int n = rand() % (num_random_weights_ + allow_zero_);  // NOLINT
-    if (allow_zero_ && n == num_random_weights_) return Weight::Zero();
-    return Weight(n);
-  }
-
- private:
-  // Permits Zero() and zero divisors.
-  const bool allow_zero_;
-  // Number of alternative random weights.
-  const size_t num_random_weights_;
-};
-
-template <class T>
-class WeightGenerate<TropicalWeightTpl<T>>
-    : public FloatWeightGenerate<TropicalWeightTpl<T>> {
- public:
-  using Weight = TropicalWeightTpl<T>;
-  using Generate = FloatWeightGenerate<Weight>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t num_random_weights = kNumRandomWeights)
-      : Generate(allow_zero, num_random_weights) {}
-
-  Weight operator()() const { return Weight(Generate::operator()()); }
-};
-
-template <class T>
-class WeightGenerate<LogWeightTpl<T>>
-    : public FloatWeightGenerate<LogWeightTpl<T>> {
- public:
-  using Weight = LogWeightTpl<T>;
-  using Generate = FloatWeightGenerate<Weight>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t num_random_weights = kNumRandomWeights)
-      : Generate(allow_zero, num_random_weights) {}
-
-  Weight operator()() const { return Weight(Generate::operator()()); }
-};
-
-// This function object returns random integers chosen from [0,
-// num_random_weights). The boolean 'allow_zero' determines whether Zero() and
-// zero divisors should be returned in the random weight generation. This is
-// intended primary for testing.
-template <class T>
-class WeightGenerate<MinMaxWeightTpl<T>> {
- public:
-  using Weight = MinMaxWeightTpl<T>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t num_random_weights = kNumRandomWeights)
-      : allow_zero_(allow_zero), num_random_weights_(num_random_weights) {}
-
-  Weight operator()() const {
-    const int n = (rand() %  // NOLINT
-                   (2 * num_random_weights_ + allow_zero_)) -
-                  num_random_weights_;
-    if (allow_zero_ && n == num_random_weights_) {
-      return Weight::Zero();
-    } else if (n == -num_random_weights_) {
-      return Weight::One();
-    } else {
-      return Weight(n);
-    }
-  }
-
- private:
-  // Permits Zero() and zero divisors.
-  const bool allow_zero_;
-  // Number of alternative random weights.
-  const size_t num_random_weights_;
-};
-
-}  // namespace fst
-
-#endif  // FST_FLOAT_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/fst-decl.h b/projects/llm_framework/include/fst/fst-decl.h
deleted file mode 100644
index 9319dfc9..00000000
--- a/projects/llm_framework/include/fst/fst-decl.h
+++ /dev/null
@@ -1,254 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// This file contains declarations of classes in the Fst template library.
-
-#ifndef FST_FST_DECL_H_
-#define FST_FST_DECL_H_
-
-#include <sys/types.h>
-#include <memory>  // for allocator<>
-
-#include <fst/types.h>
-
-namespace fst {
-
-// Symbol table and iterator.
-
-class SymbolTable;
-
-class SymbolTableIterator;
-
-// Weight templates and weights.
-
-template <class T>
-class FloatWeightTpl;
-
-template <class T>
-class TropicalWeightTpl;
-
-template <class T>
-class LogWeightTpl;
-
-template <class T>
-class MinMaxWeightTpl;
-
-using FloatWeight = FloatWeightTpl<float>;
-
-using TropicalWeight = TropicalWeightTpl<float>;
-
-using LogWeight = LogWeightTpl<float>;
-
-using MinMaxWeight = MinMaxWeightTpl<float>;
-
-// Arc templates and arcs.
-
-template <class Weight>
-struct ArcTpl;
-
-using StdArc = ArcTpl<TropicalWeight>;
-
-using LogArc = ArcTpl<LogWeight>;
-
-// Stores.
-
-template <class Element, class U>
-class DefaultCompactStore;
-
-template <class Arc>
-class DefaultCacheStore;
-
-// FST templates.
-
-template <class A, class ArcCompactor, class Unsigned = uint32,
-    class CompactStore = DefaultCompactStore<typename ArcCompactor::Element, Unsigned>,
-    class CacheStore = DefaultCacheStore<A>>
-class CompactFst;
-
-template <class Arc, class U = uint32>
-class ConstFst;
-
-template <class Arc, class Weight, class Matcher>
-class EditFst;
-
-template <class Arc>
-class ExpandedFst;
-
-template <class Arc>
-class Fst;
-
-template <class Arc>
-class MutableFst;
-
-template <class A, class Allocator = std::allocator<A>>
-class VectorState;
-
-template <class A, class State = VectorState<A>>
-class VectorFst;
-
-template <class Arc, class U = ssize_t>
-class DefaultReplaceStateTable;
-
-// On-the-fly operations.
-
-template <class Arc, class Compare>
-class ArcSortFst;
-
-template <class Arc>
-class ClosureFst;
-
-template <class A, class Store = DefaultCacheStore<A>>
-class ComposeFst;
-
-template <class Arc>
-class ConcatFst;
-
-template <class Arc>
-class DeterminizeFst;
-
-template <class Arc>
-class DifferenceFst;
-
-template <class Arc>
-class IntersectFst;
-
-template <class Arc>
-class InvertFst;
-
-template <class AArc, class BArc, class Mapper>
-class ArcMapFst;
-
-template <class Arc>
-class ProjectFst;
-
-template <class AArc, class BArc, class Selector>
-class RandGenFst;
-
-template <class Arc>
-class RelabelFst;
-
-template <class A, class StateTable = DefaultReplaceStateTable<A>,
-          class Store = DefaultCacheStore<A>>
-class ReplaceFst;
-
-template <class Arc>
-class RmEpsilonFst;
-
-template <class Arc>
-class UnionFst;
-
-// Heap.
-
-template <class T, class Compare>
-class Heap;
-
-// Compactors.
-
-template <class Arc>
-class AcceptorCompactor;
-
-template <class Arc>
-class StringCompactor;
-
-template <class Arc>
-class UnweightedAcceptorCompactor;
-
-template <class Arc>
-class UnweightedCompactor;
-
-template <class Arc>
-class WeightedStringCompactor;
-
-// Compact FSTs.
-
-template <class Arc, class U = uint32>
-using CompactStringFst = CompactFst<Arc, StringCompactor<Arc>, U>;
-
-template <class Arc, class U = uint32>
-using CompactWeightedStringFst =
-    CompactFst<Arc, WeightedStringCompactor<Arc>, U>;
-
-template <class Arc, class U = uint32>
-using CompactAcceptorFst = CompactFst<Arc, AcceptorCompactor<Arc>, U>;
-
-template <class Arc, class U = uint32>
-using CompactUnweightedFst = CompactFst<Arc, UnweightedCompactor<Arc>, U>;
-
-template <class Arc, class U = uint32>
-using CompactUnweightedAcceptorFst =
-    CompactFst<Arc, UnweightedAcceptorCompactor<Arc>, U>;
-
-// StdArc aliases for FSTs.
-
-using StdConstFst = ConstFst<StdArc>;
-using StdExpandedFst = ExpandedFst<StdArc>;
-using StdFst = Fst<StdArc>;
-using StdMutableFst = MutableFst<StdArc>;
-using StdVectorFst = VectorFst<StdArc>;
-
-// StdArc aliases for on-the-fly operations.
-
-template <class Compare>
-using StdArcSortFst = ArcSortFst<StdArc, Compare>;
-
-using StdClosureFst = ClosureFst<StdArc>;
-
-using StdComposeFst = ComposeFst<StdArc>;
-
-using StdConcatFst = ConcatFst<StdArc>;
-
-using StdDeterminizeFst = DeterminizeFst<StdArc>;
-
-using StdDifferenceFst = DifferenceFst<StdArc>;
-
-using StdIntersectFst = IntersectFst<StdArc>;
-
-using StdInvertFst = InvertFst<StdArc>;
-
-using StdProjectFst = ProjectFst<StdArc>;
-
-using StdRelabelFst = RelabelFst<StdArc>;
-
-using StdReplaceFst = ReplaceFst<StdArc>;
-
-using StdRmEpsilonFst = RmEpsilonFst<StdArc>;
-
-using StdUnionFst = UnionFst<StdArc>;
-
-// Filter states.
-
-template <class T>
-class IntegerFilterState;
-
-using CharFilterState = IntegerFilterState<signed char>;
-
-using ShortFilterState = IntegerFilterState<short>;  // NOLINT
-
-using IntFilterState = IntegerFilterState<int>;
-
-// Matchers and filters.
-
-template <class FST>
-class Matcher;
-
-template <class M1, class M2 = M1>
-class NullComposeFilter;
-
-template <class M1, class M2 = M1>
-class TrivialComposeFilter;
-
-template <class M1, class M2 = M1>
-class SequenceComposeFilter;
-
-template <class M1, class M2 = M1>
-class AltSequenceComposeFilter;
-
-template <class M1, class M2 = M1>
-class MatchComposeFilter;
-
-template <class M1, class M2 = M1>
-class NoMatchComposeFilter;
-
-}  // namespace fst
-
-#endif  // FST_FST_DECL_H_
diff --git a/projects/llm_framework/include/fst/fst.h b/projects/llm_framework/include/fst/fst.h
deleted file mode 100644
index 20e6bb3c..00000000
--- a/projects/llm_framework/include/fst/fst.h
+++ /dev/null
@@ -1,1007 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// FST abstract base class definition, state and arc iterator interface, and
-// suggested base implementation.
-
-#ifndef FST_FST_H_
-#define FST_FST_H_
-
-#include <sys/types.h>
-
-#include <cmath>
-#include <cstddef>
-
-#include <iostream>
-#include <memory>
-#include <sstream>
-#include <string>
-#include <utility>
-
-#include <fst/compat.h>
-#include <fst/flags.h>
-#include <fst/log.h>
-#include <fstream>
-
-#include <fst/arc.h>
-#include <fst/memory.h>
-#include <fst/properties.h>
-#include <fst/register.h>
-#include <fst/symbol-table.h>
-#include <fst/util.h>
-
-
-DECLARE_bool(fst_align);
-
-namespace fst {
-
-bool IsFstHeader(std::istream &, const string &);
-
-class FstHeader;
-
-template <class Arc>
-struct StateIteratorData;
-
-template <class Arc>
-struct ArcIteratorData;
-
-template <class Arc>
-class MatcherBase;
-
-struct FstReadOptions {
-  // FileReadMode(s) are advisory, there are many conditions than prevent a
-  // file from being mapped, READ mode will be selected in these cases with
-  // a warning indicating why it was chosen.
-  enum FileReadMode { READ, MAP };
-
-  string source;                // Where you're reading from.
-  const FstHeader *header;      // Pointer to FST header; if non-zero, use
-                                // this info (don't read a stream header).
-  const SymbolTable *isymbols;  // Pointer to input symbols; if non-zero, use
-                                // this info (read and skip stream isymbols)
-  const SymbolTable *osymbols;  // Pointer to output symbols; if non-zero, use
-                                // this info (read and skip stream osymbols)
-  FileReadMode mode;            // Read or map files (advisory, if possible)
-  bool read_isymbols;           // Read isymbols, if any (default: true).
-  bool read_osymbols;           // Read osymbols, if any (default: true).
-
-  explicit FstReadOptions(const string &source = "<unspecified>",
-                          const FstHeader *header = nullptr,
-                          const SymbolTable *isymbols = nullptr,
-                          const SymbolTable *osymbols = nullptr);
-
-  explicit FstReadOptions(const string &source, const SymbolTable *isymbols,
-                          const SymbolTable *osymbols = nullptr);
-
-  // Helper function to convert strings FileReadModes into their enum value.
-  static FileReadMode ReadMode(const string &mode);
-
-  // Outputs a debug string for the FstReadOptions object.
-  string DebugString() const;
-};
-
-struct FstWriteOptions {
-  string source;        // Where you're writing to.
-  bool write_header;    // Write the header?
-  bool write_isymbols;  // Write input symbols?
-  bool write_osymbols;  // Write output symbols?
-  bool align;           // Write data aligned (may fail on pipes)?
-  bool stream_write;    // Avoid seek operations in writing.
-
-  explicit FstWriteOptions(const string &source = "<unspecifed>",
-                           bool write_header = true, bool write_isymbols = true,
-                           bool write_osymbols = true,
-                           bool align = FLAGS_fst_align,
-                           bool stream_write = false)
-      : source(source),
-        write_header(write_header),
-        write_isymbols(write_isymbols),
-        write_osymbols(write_osymbols),
-        align(align),
-        stream_write(stream_write) {}
-};
-
-// Header class.
-//
-// This is the recommended file header representation.
-
-class FstHeader {
- public:
-  enum {
-    HAS_ISYMBOLS = 0x1,  // Has input symbol table.
-    HAS_OSYMBOLS = 0x2,  // Has output symbol table.
-    IS_ALIGNED = 0x4,    // Memory-aligned (where appropriate).
-  } Flags;
-
-  FstHeader() : version_(0), flags_(0), properties_(0), start_(-1),
-      numstates_(0), numarcs_(0) {}
-
-  const string &FstType() const { return fsttype_; }
-
-  const string &ArcType() const { return arctype_; }
-
-  int32 Version() const { return version_; }
-
-  int32 GetFlags() const { return flags_; }
-
-  uint64 Properties() const { return properties_; }
-
-  int64 Start() const { return start_; }
-
-  int64 NumStates() const { return numstates_; }
-
-  int64 NumArcs() const { return numarcs_; }
-
-  void SetFstType(const string &type) { fsttype_ = type; }
-
-  void SetArcType(const string &type) { arctype_ = type; }
-
-  void SetVersion(int32 version) { version_ = version; }
-
-  void SetFlags(int32 flags) { flags_ = flags; }
-
-  void SetProperties(uint64 properties) { properties_ = properties; }
-
-  void SetStart(int64 start) { start_ = start; }
-
-  void SetNumStates(int64 numstates) { numstates_ = numstates; }
-
-  void SetNumArcs(int64 numarcs) { numarcs_ = numarcs; }
-
-  bool Read(std::istream &strm, const string &source,
-            bool rewind = false);
-
-  bool Write(std::ostream &strm, const string &source) const;
-
-  // Outputs a debug string for the FstHeader object.
-  string DebugString() const;
-
- private:
-  string fsttype_;     // E.g. "vector".
-  string arctype_;     // E.g. "standard".
-  int32 version_;      // Type version number.
-  int32 flags_;        // File format bits.
-  uint64 properties_;  // FST property bits.
-  int64 start_;        // Start state.
-  int64 numstates_;    // # of states.
-  int64 numarcs_;      // # of arcs.
-};
-
-// Specifies matcher action.
-enum MatchType {
-  MATCH_INPUT = 1,   // Match input label.
-  MATCH_OUTPUT = 2,  // Match output label.
-  MATCH_BOTH = 3,    // Match input or output label.
-  MATCH_NONE = 4,    // Match nothing.
-  MATCH_UNKNOWN = 5
-};  // Otherwise, match type unknown.
-
-constexpr int kNoLabel = -1;    // Not a valid label.
-constexpr int kNoStateId = -1;  // Not a valid state ID.
-
-// A generic FST, templated on the arc definition, with common-demoninator
-// methods (use StateIterator and ArcIterator to iterate over its states and
-// arcs).
-template <class A>
-class Fst {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  virtual ~Fst() {}
-
-  // Initial state.
-  virtual StateId Start() const = 0;
-
-  // State's final weight.
-  virtual Weight Final(StateId) const = 0;
-
-  // State's arc count.
-  virtual size_t NumArcs(StateId) const = 0;
-
-  // State's input epsilon count.
-  virtual size_t NumInputEpsilons(StateId) const = 0;
-
-  // State's output epsilon count.
-  virtual size_t NumOutputEpsilons(StateId) const = 0;
-
-  // Property bits. If test = false, return stored properties bits for mask
-  // (some possibly unknown); if test = true, return property bits for mask
-  // (computing o.w. unknown).
-  virtual uint64 Properties(uint64 mask, bool test) const = 0;
-
-  // FST type name.
-  virtual const string &Type() const = 0;
-
-  // Gets a copy of this Fst. The copying behaves as follows:
-  //
-  // (1) The copying is constant time if safe = false or if safe = true
-  // and is on an otherwise unaccessed FST.
-  //
-  // (2) If safe = true, the copy is thread-safe in that the original
-  // and copy can be safely accessed (but not necessarily mutated) by
-  // separate threads. For some FST types, 'Copy(true)' should only be
-  // called on an FST that has not otherwise been accessed. Behavior is
-  // otherwise undefined.
-  //
-  // (3) If a MutableFst is copied and then mutated, then the original is
-  // unmodified and vice versa (often by a copy-on-write on the initial
-  // mutation, which may not be constant time).
-  virtual Fst<Arc> *Copy(bool safe = false) const = 0;
-
-  // Reads an FST from an input stream; returns nullptr on error.
-  static Fst<Arc> *Read(std::istream &strm, const FstReadOptions &opts) {
-    FstReadOptions ropts(opts);
-    FstHeader hdr;
-    if (ropts.header) {
-      hdr = *opts.header;
-    } else {
-      if (!hdr.Read(strm, opts.source)) return nullptr;
-      ropts.header = &hdr;
-    }
-    const auto &fst_type = hdr.FstType();
-    const auto reader = FstRegister<Arc>::GetRegister()->GetReader(fst_type);
-    if (!reader) {
-      LOG(ERROR) << "Fst::Read: Unknown FST type " << fst_type
-                 << " (arc type = " << Arc::Type() << "): " << ropts.source;
-      return nullptr;
-    }
-    return reader(strm, ropts);
-  }
-
-  // Reads an FST from a file; returns nullptr on error. An empty filename
-  // results in reading from standard input.
-  static Fst<Arc> *Read(const string &filename) {
-    if (!filename.empty()) {
-      std::ifstream strm(filename,
-                              std::ios_base::in | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "Fst::Read: Can't open file: " << filename;
-        return nullptr;
-      }
-      return Read(strm, FstReadOptions(filename));
-    } else {
-      return Read(std::cin, FstReadOptions("standard input"));
-    }
-  }
-
-  // Writes an FST to an output stream; returns false on error.
-  virtual bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
-    LOG(ERROR) << "Fst::Write: No write stream method for " << Type()
-               << " FST type";
-    return false;
-  }
-
-  // Writes an FST to a file; returns false on error; an empty filename
-  // results in writing to standard output.
-  virtual bool Write(const string &filename) const {
-    LOG(ERROR) << "Fst::Write: No write filename method for " << Type()
-               << " FST type";
-    return false;
-  }
-
-  // Returns input label symbol table; return nullptr if not specified.
-  virtual const SymbolTable *InputSymbols() const = 0;
-
-  // Return output label symbol table; return nullptr if not specified.
-  virtual const SymbolTable *OutputSymbols() const = 0;
-
-  // For generic state iterator construction (not normally called directly by
-  // users). Does not copy the FST.
-  virtual void InitStateIterator(StateIteratorData<Arc> *data) const = 0;
-
-  // For generic arc iterator construction (not normally called directly by
-  // users). Does not copy the FST.
-  virtual void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const = 0;
-
-  // For generic matcher construction (not normally called directly by users).
-  // Does not copy the FST.
-  virtual MatcherBase<Arc> *InitMatcher(MatchType match_type) const;
-
- protected:
-  bool WriteFile(const string &filename) const {
-    if (!filename.empty()) {
-      std::ofstream strm(filename,
-                               std::ios_base::out | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "Fst::Write: Can't open file: " << filename;
-        return false;
-      }
-      bool val = Write(strm, FstWriteOptions(filename));
-      if (!val) LOG(ERROR) << "Fst::Write failed: " << filename;
-      return val;
-    } else {
-      return Write(std::cout, FstWriteOptions("standard output"));
-    }
-  }
-};
-
-// A useful alias when using StdArc.
-using StdFst = Fst<StdArc>;
-
-// State and arc iterator definitions.
-//
-// State iterator interface templated on the Arc definition; used for
-// StateIterator specializations returned by the InitStateIterator FST method.
-template <class Arc>
-class StateIteratorBase {
- public:
-  using StateId = typename Arc::StateId;
-
-  virtual ~StateIteratorBase() {}
-
-  // End of iterator?
-  virtual bool Done() const = 0;
-  // Returns current state (when !Done()).
-  virtual StateId Value() const = 0;
-  // Advances to next state (when !Done()).
-  virtual void Next() = 0;
-  // Resets to initial condition.
-  virtual void Reset() = 0;
-};
-
-// StateIterator initialization data.
-
-template <class Arc>
-struct StateIteratorData {
-  using StateId = typename Arc::StateId;
-
-  // Specialized iterator if non-zero.
-  StateIteratorBase<Arc> *base;
-  // Otherwise, the total number of states.
-  StateId nstates;
-
-  StateIteratorData() : base(nullptr), nstates(0) {}
-
-  StateIteratorData(const StateIteratorData &) = delete;
-  StateIteratorData &operator=(const StateIteratorData &) = delete;
-};
-
-// Generic state iterator, templated on the FST definition (a wrapper
-// around a pointer to a specific one). Here is a typical use:
-//
-//   for (StateIterator<StdFst> siter(fst);
-//        !siter.Done();
-//        siter.Next()) {
-//     StateId s = siter.Value();
-//     ...
-//   }
-// There is no copying of the FST.
-template <class FST>
-class StateIterator {
- public:
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-
-  explicit StateIterator(const FST &fst) : s_(0) {
-    fst.InitStateIterator(&data_);
-  }
-
-  ~StateIterator() { delete data_.base; }
-
-  bool Done() const {
-    return data_.base ? data_.base->Done() : s_ >= data_.nstates;
-  }
-
-  StateId Value() const { return data_.base ? data_.base->Value() : s_; }
-
-  void Next() {
-    if (data_.base) {
-      data_.base->Next();
-    } else {
-      ++s_;
-    }
-  }
-
-  void Reset() {
-    if (data_.base) {
-      data_.base->Reset();
-    } else {
-      s_ = 0;
-    }
-  }
-
- private:
-  StateIteratorData<Arc> data_;
-  StateId s_;
-};
-
-// Flags to control the behavior on an arc iterator.
-static constexpr uint32 kArcILabelValue =
-    0x0001;  // Value() gives valid ilabel.
-static constexpr uint32 kArcOLabelValue = 0x0002;  //  "       "     " olabel.
-static constexpr uint32 kArcWeightValue = 0x0004;  //  "       "     " weight.
-static constexpr uint32 kArcNextStateValue =
-    0x0008;                                    //  "       "     " nextstate.
-static constexpr uint32 kArcNoCache = 0x0010;  // No need to cache arcs.
-
-static constexpr uint32 kArcValueFlags =
-    kArcILabelValue | kArcOLabelValue | kArcWeightValue | kArcNextStateValue;
-
-static constexpr uint32 kArcFlags = kArcValueFlags | kArcNoCache;
-
-// Arc iterator interface, templated on the arc definition; used for arc
-// iterator specializations that are returned by the InitArcIterator FST method.
-template <class Arc>
-class ArcIteratorBase {
- public:
-  using StateId = typename Arc::StateId;
-
-  virtual ~ArcIteratorBase() {}
-
-  // End of iterator?
-  virtual bool Done() const = 0;
-  // Returns current arc (when !Done()).
-  virtual const Arc &Value() const = 0;
-  // Advances to next arc (when !Done()).
-  virtual void Next() = 0;
-  // Returns current position.
-  virtual size_t Position() const = 0;
-  // Returns to initial condition.
-  virtual void Reset() = 0;
-  // Advances to arbitrary arc by position.
-  virtual void Seek(size_t) = 0;
-  // Returns current behavorial flags
-  virtual uint32 Flags() const = 0;
-  // Sets behavorial flags.
-  virtual void SetFlags(uint32, uint32) = 0;
-};
-
-// ArcIterator initialization data.
-template <class Arc>
-struct ArcIteratorData {
-  ArcIteratorData()
-      : base(nullptr), arcs(nullptr), narcs(0), ref_count(nullptr) {}
-
-  ArcIteratorData(const ArcIteratorData &) = delete;
-
-  ArcIteratorData &operator=(const ArcIteratorData &) = delete;
-
-  ArcIteratorBase<Arc> *base;  // Specialized iterator if non-zero.
-  const Arc *arcs;             // O.w. arcs pointer
-  size_t narcs;                // ... and arc count.
-  int *ref_count;              // ... and reference count if non-zero.
-};
-
-// Generic arc iterator, templated on the FST definition (a wrapper around a
-// pointer to a specific one). Here is a typical use:
-//
-//   for (ArcIterator<StdFst> aiter(fst, s);
-//        !aiter.Done();
-//         aiter.Next()) {
-//     StdArc &arc = aiter.Value();
-//     ...
-//   }
-// There is no copying of the FST.
-template <class FST>
-class ArcIterator {
- public:
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const FST &fst, StateId s) : i_(0) {
-    fst.InitArcIterator(s, &data_);
-  }
-
-  explicit ArcIterator(const ArcIteratorData<Arc> &data) : data_(data), i_(0) {
-    if (data_.ref_count) ++(*data_.ref_count);
-  }
-
-  ~ArcIterator() {
-    if (data_.base) {
-      delete data_.base;
-    } else if (data_.ref_count) {
-      --(*data_.ref_count);
-    }
-  }
-
-  bool Done() const {
-    return data_.base ? data_.base->Done() : i_ >= data_.narcs;
-  }
-
-  const Arc &Value() const {
-    return data_.base ? data_.base->Value() : data_.arcs[i_];
-  }
-
-  void Next() {
-    if (data_.base) {
-      data_.base->Next();
-    } else {
-      ++i_;
-    }
-  }
-
-  void Reset() {
-    if (data_.base) {
-      data_.base->Reset();
-    } else {
-      i_ = 0;
-    }
-  }
-
-  void Seek(size_t a) {
-    if (data_.base) {
-      data_.base->Seek(a);
-    } else {
-      i_ = a;
-    }
-  }
-
-  size_t Position() const { return data_.base ? data_.base->Position() : i_; }
-
-  uint32 Flags() const {
-    if (data_.base) {
-      return data_.base->Flags();
-    } else {
-      return kArcValueFlags;
-    }
-  }
-
-  void SetFlags(uint32 flags, uint32 mask) {
-    if (data_.base) data_.base->SetFlags(flags, mask);
-  }
-
- private:
-  ArcIteratorData<Arc> data_;
-  size_t i_;
-};
-
-}  // namespace fst
-
-// ArcIterator placement operator new and destroy function; new needs to be in
-// the global namespace.
-
-template <class FST>
-void *operator new(size_t size,
-                   fst::MemoryPool<fst::ArcIterator<FST>> *pool) {
-  return pool->Allocate();
-}
-
-namespace fst {
-
-template <class FST>
-void Destroy(ArcIterator<FST> *aiter, MemoryPool<ArcIterator<FST>> *pool) {
-  if (aiter) {
-    aiter->~ArcIterator<FST>();
-    pool->Free(aiter);
-  }
-}
-
-// Matcher definitions.
-
-template <class Arc>
-MatcherBase<Arc> *Fst<Arc>::InitMatcher(MatchType match_type) const {
-  return nullptr;  // One should just use the default matcher.
-}
-
-// FST accessors, useful in high-performance applications.
-
-namespace internal {
-
-// General case, requires non-abstract, 'final' methods. Use for inlining.
-
-template <class F>
-inline typename F::Arc::Weight Final(const F &fst, typename F::Arc::StateId s) {
-  return fst.F::Final(s);
-}
-
-template <class F>
-inline ssize_t NumArcs(const F &fst, typename F::Arc::StateId s) {
-  return fst.F::NumArcs(s);
-}
-
-template <class F>
-inline ssize_t NumInputEpsilons(const F &fst, typename F::Arc::StateId s) {
-  return fst.F::NumInputEpsilons(s);
-}
-
-template <class F>
-inline ssize_t NumOutputEpsilons(const F &fst, typename F::Arc::StateId s) {
-  return fst.F::NumOutputEpsilons(s);
-}
-
-// Fst<Arc> case, abstract methods.
-
-template <class Arc>
-inline typename Arc::Weight Final(const Fst<Arc> &fst,
-                                  typename Arc::StateId s) {
-  return fst.Final(s);
-}
-
-template <class Arc>
-inline size_t NumArcs(const Fst<Arc> &fst, typename Arc::StateId s) {
-  return fst.NumArcs(s);
-}
-
-template <class Arc>
-inline size_t NumInputEpsilons(const Fst<Arc> &fst, typename Arc::StateId s) {
-  return fst.NumInputEpsilons(s);
-}
-
-template <class Arc>
-inline size_t NumOutputEpsilons(const Fst<Arc> &fst, typename Arc::StateId s) {
-  return fst.NumOutputEpsilons(s);
-}
-
-// FST implementation base.
-//
-// This is the recommended FST implementation base class. It will handle
-// reference counts, property bits, type information and symbols.
-//
-// Users are discouraged, but not prohibited, from subclassing this outside the
-// FST library.
-template <class Arc>
-class FstImpl {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  FstImpl() : properties_(0), type_("null") {}
-
-  FstImpl(const FstImpl<Arc> &impl)
-      : properties_(impl.properties_),
-        type_(impl.type_),
-        isymbols_(impl.isymbols_ ? impl.isymbols_->Copy() : nullptr),
-        osymbols_(impl.osymbols_ ? impl.osymbols_->Copy() : nullptr) {}
-
-  FstImpl(FstImpl<Arc> &&impl) noexcept;
-
-  virtual ~FstImpl() {}
-
-  FstImpl &operator=(const FstImpl<Arc> &impl) {
-    properties_ = impl.properties_;
-    type_ = impl.type_;
-    isymbols_ = impl.isymbols_ ? impl.isymbols_->Copy() : nullptr;
-    osymbols_ = impl.osymbols_ ? impl.osymbols_->Copy() : nullptr;
-    return *this;
-  }
-
-  FstImpl &operator=(FstImpl<Arc> &&impl) noexcept;
-
-  const string &Type() const { return type_; }
-
-  void SetType(const string &type) { type_ = type; }
-
-  virtual uint64 Properties() const { return properties_; }
-
-  virtual uint64 Properties(uint64 mask) const { return properties_ & mask; }
-
-  void SetProperties(uint64 props) {
-    properties_ &= kError;  // kError can't be cleared.
-    properties_ |= props;
-  }
-
-  void SetProperties(uint64 props, uint64 mask) {
-    properties_ &= ~mask | kError;  // kError can't be cleared.
-    properties_ |= props & mask;
-  }
-
-  // Allows (only) setting error bit on const FST implementations.
-  void SetProperties(uint64 props, uint64 mask) const {
-    if (mask != kError) {
-      FSTERROR() << "FstImpl::SetProperties() const: Can only set kError";
-    }
-    properties_ |= kError;
-  }
-
-  const SymbolTable *InputSymbols() const { return isymbols_.get(); }
-
-  const SymbolTable *OutputSymbols() const { return osymbols_.get(); }
-
-  SymbolTable *InputSymbols() { return isymbols_.get(); }
-
-  SymbolTable *OutputSymbols() { return osymbols_.get(); }
-
-  void SetInputSymbols(const SymbolTable *isyms) {
-    isymbols_.reset(isyms ? isyms->Copy() : nullptr);
-  }
-
-  void SetOutputSymbols(const SymbolTable *osyms) {
-    osymbols_.reset(osyms ? osyms->Copy() : nullptr);
-  }
-
-  // Reads header and symbols from input stream, initializes FST, and returns
-  // the header. If opts.header is non-null, skips reading and uses the option
-  // value instead. If opts.[io]symbols is non-null, reads in (if present), but
-  // uses the option value.
-  bool ReadHeader(std::istream &strm, const FstReadOptions &opts,
-                  int min_version, FstHeader *hdr);
-
-  // Writes header and symbols to output stream. If opts.header is false, skips
-  // writing header. If opts.[io]symbols is false, skips writing those symbols.
-  // This method is needed for implementations that implement Write methods.
-  void WriteHeader(std::ostream &strm, const FstWriteOptions &opts,
-                   int version, FstHeader *hdr) const {
-    if (opts.write_header) {
-      hdr->SetFstType(type_);
-      hdr->SetArcType(Arc::Type());
-      hdr->SetVersion(version);
-      hdr->SetProperties(properties_);
-      int32 file_flags = 0;
-      if (isymbols_ && opts.write_isymbols) {
-        file_flags |= FstHeader::HAS_ISYMBOLS;
-      }
-      if (osymbols_ && opts.write_osymbols) {
-        file_flags |= FstHeader::HAS_OSYMBOLS;
-      }
-      if (opts.align) file_flags |= FstHeader::IS_ALIGNED;
-      hdr->SetFlags(file_flags);
-      hdr->Write(strm, opts.source);
-    }
-    if (isymbols_ && opts.write_isymbols) isymbols_->Write(strm);
-    if (osymbols_ && opts.write_osymbols) osymbols_->Write(strm);
-  }
-
-  // Writes out header and symbols to output stream. If opts.header is false,
-  // skips writing header. If opts.[io]symbols is false, skips writing those
-  // symbols. `type` is the FST type being written. This method is used in the
-  // cross-type serialization methods Fst::WriteFst.
-  static void WriteFstHeader(const Fst<Arc> &fst, std::ostream &strm,
-                             const FstWriteOptions &opts, int version,
-                             const string &type, uint64 properties,
-                             FstHeader *hdr) {
-    if (opts.write_header) {
-      hdr->SetFstType(type);
-      hdr->SetArcType(Arc::Type());
-      hdr->SetVersion(version);
-      hdr->SetProperties(properties);
-      int32 file_flags = 0;
-      if (fst.InputSymbols() && opts.write_isymbols) {
-        file_flags |= FstHeader::HAS_ISYMBOLS;
-      }
-      if (fst.OutputSymbols() && opts.write_osymbols) {
-        file_flags |= FstHeader::HAS_OSYMBOLS;
-      }
-      if (opts.align) file_flags |= FstHeader::IS_ALIGNED;
-      hdr->SetFlags(file_flags);
-      hdr->Write(strm, opts.source);
-    }
-    if (fst.InputSymbols() && opts.write_isymbols) {
-      fst.InputSymbols()->Write(strm);
-    }
-    if (fst.OutputSymbols() && opts.write_osymbols) {
-      fst.OutputSymbols()->Write(strm);
-    }
-  }
-
-  // In serialization routines where the header cannot be written until after
-  // the machine has been serialized, this routine can be called to seek to the
-  // beginning of the file an rewrite the header with updated fields. It
-  // repositions the file pointer back at the end of the file. Returns true on
-  // success, false on failure.
-  static bool UpdateFstHeader(const Fst<Arc> &fst, std::ostream &strm,
-                              const FstWriteOptions &opts, int version,
-                              const string &type, uint64 properties,
-                              FstHeader *hdr, size_t header_offset) {
-    strm.seekp(header_offset);
-    if (!strm) {
-      LOG(ERROR) << "Fst::UpdateFstHeader: Write failed: " << opts.source;
-      return false;
-    }
-    WriteFstHeader(fst, strm, opts, version, type, properties, hdr);
-    if (!strm) {
-      LOG(ERROR) << "Fst::UpdateFstHeader: Write failed: " << opts.source;
-      return false;
-    }
-    strm.seekp(0, std::ios_base::end);
-    if (!strm) {
-      LOG(ERROR) << "Fst::UpdateFstHeader: Write failed: " << opts.source;
-      return false;
-    }
-    return true;
-  }
-
- protected:
-  mutable uint64 properties_;  // Property bits.
-
- private:
-  string type_;  // Unique name of FST class.
-  std::unique_ptr<SymbolTable> isymbols_;
-  std::unique_ptr<SymbolTable> osymbols_;
-};
-
-template <class Arc>
-inline FstImpl<Arc>::FstImpl(FstImpl<Arc> &&) noexcept = default;
-
-template <class Arc>
-inline FstImpl<Arc> &FstImpl<Arc>::operator=(
-    FstImpl<Arc> &&) noexcept = default;
-
-template <class Arc>
-bool FstImpl<Arc>::ReadHeader(std::istream &strm, const FstReadOptions &opts,
-                              int min_version, FstHeader *hdr) {
-  if (opts.header) {
-    *hdr = *opts.header;
-  } else if (!hdr->Read(strm, opts.source)) {
-    return false;
-  }
-  if (FLAGS_v >= 2) {
-    LOG(INFO) << "FstImpl::ReadHeader: source: " << opts.source
-              << ", fst_type: " << hdr->FstType()
-              << ", arc_type: " << Arc::Type()
-              << ", version: " << hdr->Version()
-              << ", flags: " << hdr->GetFlags();
-  }
-  if (hdr->FstType() != type_) {
-    LOG(ERROR) << "FstImpl::ReadHeader: FST not of type " << type_
-               << ": " << opts.source;
-    return false;
-  }
-  if (hdr->ArcType() != Arc::Type()) {
-    LOG(ERROR) << "FstImpl::ReadHeader: Arc not of type " << Arc::Type()
-               << ": " << opts.source;
-    return false;
-  }
-  if (hdr->Version() < min_version) {
-    LOG(ERROR) << "FstImpl::ReadHeader: Obsolete " << type_
-               << " FST version: " << opts.source;
-    return false;
-  }
-  properties_ = hdr->Properties();
-  if (hdr->GetFlags() & FstHeader::HAS_ISYMBOLS) {
-    isymbols_.reset(SymbolTable::Read(strm, opts.source));
-  }
-  // Deletes input symbol table.
-  if (!opts.read_isymbols) SetInputSymbols(nullptr);
-  if (hdr->GetFlags() & FstHeader::HAS_OSYMBOLS) {
-    osymbols_.reset(SymbolTable::Read(strm, opts.source));
-  }
-  // Deletes output symbol table.
-  if (!opts.read_osymbols) SetOutputSymbols(nullptr);
-  if (opts.isymbols) {
-    isymbols_.reset(opts.isymbols->Copy());
-  }
-  if (opts.osymbols) {
-    osymbols_.reset(opts.osymbols->Copy());
-  }
-  return true;
-}
-
-}  // namespace internal
-
-template <class Arc>
-uint64 TestProperties(const Fst<Arc> &fst, uint64 mask, uint64 *known);
-
-// This is a helper class template useful for attaching an FST interface to
-// its implementation, handling reference counting.
-template <class Impl, class FST = Fst<typename Impl::Arc>>
-class ImplToFst : public FST {
- public:
-  using Arc = typename Impl::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  StateId Start() const override { return impl_->Start(); }
-
-  Weight Final(StateId s) const override { return impl_->Final(s); }
-
-  size_t NumArcs(StateId s) const override { return impl_->NumArcs(s); }
-
-  size_t NumInputEpsilons(StateId s) const override {
-    return impl_->NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) const override {
-    return impl_->NumOutputEpsilons(s);
-  }
-
-  uint64 Properties(uint64 mask, bool test) const override {
-    if (test) {
-      uint64 knownprops, testprops = TestProperties(*this, mask, &knownprops);
-      impl_->SetProperties(testprops, knownprops);
-      return testprops & mask;
-    } else {
-      return impl_->Properties(mask);
-    }
-  }
-
-  const string &Type() const override { return impl_->Type(); }
-
-  const SymbolTable *InputSymbols() const override {
-    return impl_->InputSymbols();
-  }
-
-  const SymbolTable *OutputSymbols() const override {
-    return impl_->OutputSymbols();
-  }
-
- protected:
-  explicit ImplToFst(std::shared_ptr<Impl> impl) : impl_(std::move(impl)) {}
-
-  // This constructor presumes there is a copy constructor for the
-  // implementation.
-  ImplToFst(const ImplToFst<Impl, FST> &fst, bool safe) {
-    if (safe) {
-      impl_ = std::make_shared<Impl>(*(fst.impl_));
-    } else {
-      impl_ = fst.impl_;
-    }
-  }
-
-  ImplToFst() = delete;
-
-  ImplToFst(const ImplToFst<Impl, FST> &fst) : impl_(fst.impl_) {}
-
-  ImplToFst(ImplToFst<Impl, FST> &&fst) noexcept
-      : impl_(std::move(fst.impl_)) {
-    fst.impl_ = std::make_shared<Impl>();
-  }
-
-  ImplToFst<Impl, FST> &operator=(const ImplToFst<Impl, FST> &fst) {
-    impl_ = fst.impl_;
-    return *this;
-  }
-
-  ImplToFst<Impl, FST> &operator=(ImplToFst<Impl, FST> &&fst) noexcept {
-    if (this != &fst) {
-      impl_ = std::move(fst.impl_);
-      fst.impl_ = std::make_shared<Impl>();
-    }
-    return *this;
-  }
-
-  // Returns raw pointers to the shared object.
-  const Impl *GetImpl() const { return impl_.get(); }
-
-  Impl *GetMutableImpl() const { return impl_.get(); }
-
-  // Returns a ref-counted smart poiner to the implementation.
-  std::shared_ptr<Impl> GetSharedImpl() const { return impl_; }
-
-  bool Unique() const { return impl_.unique(); }
-
-  void SetImpl(std::shared_ptr<Impl> impl) { impl_ = std::move(impl); }
-
- private:
-  template <class IFST, class OFST>
-  friend void Cast(const IFST &ifst, OFST *ofst);
-
-  std::shared_ptr<Impl> impl_;
-};
-
-// Converts FSTs by casting their implementations, where this makes sense
-// (which excludes implementations with weight-dependent virtual methods).
-// Must be a friend of the FST classes involved (currently the concrete FSTs:
-// ConstFst, CompactFst, and VectorFst). This can only be safely used for arc
-// types that have identical storage characteristics. As with an FST
-// copy constructor and Copy() method, this is a constant time operation
-// (but subject to copy-on-write if it is a MutableFst and modified).
-template <class IFST, class OFST>
-void Cast(const IFST &ifst, OFST *ofst) {
-  using OImpl = typename OFST::Impl;
-  ofst->impl_ = std::shared_ptr<OImpl>(ifst.impl_,
-      reinterpret_cast<OImpl *>(ifst.impl_.get()));
-}
-
-// FST serialization.
-
-template <class Arc>
-string FstToString(const Fst<Arc> &fst,
-                   const FstWriteOptions &options =
-                       FstWriteOptions("FstToString")) {
-  std::ostringstream ostrm;
-  fst.Write(ostrm, options);
-  return ostrm.str();
-}
-
-template <class Arc>
-void FstToString(const Fst<Arc> &fst, string *result) {
-  *result = FstToString(fst);
-}
-
-template <class Arc>
-void FstToString(const Fst<Arc> &fst, string *result,
-                 const FstWriteOptions &options) {
-  *result = FstToString(fst, options);
-}
-
-template <class Arc>
-Fst<Arc> *StringToFst(const string &s) {
-  std::istringstream istrm(s);
-  return Fst<Arc>::Read(istrm, FstReadOptions("StringToFst"));
-}
-
-}  // namespace fst
-
-#endif  // FST_FST_H_
diff --git a/projects/llm_framework/include/fst/fstlib.h b/projects/llm_framework/include/fst/fstlib.h
deleted file mode 100644
index e8b1c3a1..00000000
--- a/projects/llm_framework/include/fst/fstlib.h
+++ /dev/null
@@ -1,130 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// This is a library for constructing, combining, optimizing, and searching
-// "weighted finite-state transducers" (FSTs). Weighted finite-state transducers
-// are automata where each transition has an input label, an output label, and a
-// weight. The more familiar finite-state acceptor is represented as a
-// transducer with each transition's input and output the same. Finite-state
-// acceptors are used to represent sets of strings (specifically, "regular" or
-// "rational sets"); finite-state transducers are used to represent binary
-// relations between pairs of strings (specifically, "rational transductions").
-// The weights can be used to represent the cost of taking a particular
-// transition.
-//
-// In this library, transducers are templated on the Arc (transition)
-// definition, which allows changing the label, weight, and state ID sets.
-// Labels and state IDs are restricted to signed integral types but the weight
-// can be an arbitrary type whose members satisfy certain algebraic ("semiring")
-// properties.
-//
-// This convenience file includes all other FST header files.
-
-#ifndef FST_FSTLIB_H_
-#define FST_FSTLIB_H_
-
-
-// Abstract FST classes.
-#include <fst/expanded-fst.h>
-#include <fst/fst.h>
-#include <fst/mutable-fst.h>
-
-// Concrete FST classes.
-#include <fst/compact-fst.h>
-#include <fst/const-fst.h>
-#include <fst/edit-fst.h>
-#include <fst/vector-fst.h>
-
-// FST algorithms and delayed FST classes.
-#include <fst/arc-map.h>
-#include <fst/arcsort.h>
-#include <fst/closure.h>
-#include <fst/compose.h>
-#include <fst/concat.h>
-#include <fst/connect.h>
-#include <fst/determinize.h>
-#include <fst/difference.h>
-#include <fst/disambiguate.h>
-#include <fst/encode.h>
-#include <fst/epsnormalize.h>
-#include <fst/equal.h>
-#include <fst/equivalent.h>
-#include <fst/factor-weight.h>
-#include <fst/intersect.h>
-#include <fst/invert.h>
-#include <fst/isomorphic.h>
-#include <fst/map.h>
-#include <fst/minimize.h>
-#include <fst/project.h>
-#include <fst/prune.h>
-#include <fst/push.h>
-#include <fst/randequivalent.h>
-#include <fst/randgen.h>
-#include <fst/rational.h>
-#include <fst/relabel.h>
-#include <fst/replace-util.h>
-#include <fst/replace.h>
-#include <fst/reverse.h>
-#include <fst/reweight.h>
-#include <fst/rmepsilon.h>
-#include <fst/rmfinalepsilon.h>
-#include <fst/shortest-distance.h>
-#include <fst/shortest-path.h>
-#include <fst/state-map.h>
-#include <fst/statesort.h>
-#include <fst/synchronize.h>
-#include <fst/topsort.h>
-#include <fst/union.h>
-#include <fst/verify.h>
-#include <fst/visit.h>
-
-// Weights.
-#include <fst/expectation-weight.h>
-#include <fst/float-weight.h>
-#include <fst/lexicographic-weight.h>
-#include <fst/pair-weight.h>
-#include <fst/power-weight.h>
-#include <fst/product-weight.h>
-#include <fst/signed-log-weight.h>
-#include <fst/sparse-power-weight.h>
-#include <fst/sparse-tuple-weight.h>
-#include <fst/string-weight.h>
-#include <fst/tuple-weight.h>
-#include <fst/weight.h>
-
-// Auxiliary classes for composition.
-#include <fst/compose-filter.h>
-#include <fst/lookahead-filter.h>
-#include <fst/lookahead-matcher.h>
-#include <fst/matcher-fst.h>
-#include <fst/matcher.h>
-#include <fst/state-table.h>
-
-// Data structures.
-#include <fst/heap.h>
-#include <fst/interval-set.h>
-#include <fst/queue.h>
-#include <fst/union-find.h>
-
-// Miscellaneous.
-#include <fst/accumulator.h>
-#include <fst/add-on.h>
-#include <fst/arc.h>
-#include <fst/arcfilter.h>
-#include <fst/cache.h>
-#include <fst/complement.h>
-#include <fst/dfs-visit.h>
-#include <fst/generic-register.h>
-#include <fst/label-reachable.h>
-#include <fst/partition.h>
-#include <fst/properties.h>
-#include <fst/register.h>
-#include <fst/state-reachable.h>
-#include <fst/string.h>
-#include <fst/symbol-table-ops.h>
-#include <fst/symbol-table.h>
-#include <fst/test-properties.h>
-#include <fst/util.h>
-
-
-#endif  // FST_FSTLIB_H_
diff --git a/projects/llm_framework/include/fst/generic-register.h b/projects/llm_framework/include/fst/generic-register.h
deleted file mode 100644
index ea6b8fe1..00000000
--- a/projects/llm_framework/include/fst/generic-register.h
+++ /dev/null
@@ -1,126 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_GENERIC_REGISTER_H_
-#define FST_GENERIC_REGISTER_H_
-
-#include <fst/compat.h>
-#ifndef FST_NO_DYNAMIC_LINKING
-#include <dlfcn.h>
-#endif
-#include <map>
-#include <string>
-
-#include <fst/types.h>
-#include <fst/log.h>
-
-// Generic class representing a globally-stored correspondence between
-// objects of KeyType and EntryType.
-//
-// KeyType must:
-//
-// * be such as can be stored as a key in a std::map<>.
-// * be concatenable with a const char* with the + operator
-//   (or you must subclass and redefine LoadEntryFromSharedObject)
-//
-// EntryType must be default constructible.
-//
-// The third template parameter should be the type of a subclass of this class
-// (think CRTP). This is to allow GetRegister() to instantiate and return an
-// object of the appropriate type.
-
-namespace fst {
-
-template <class KeyType, class EntryType, class RegisterType>
-class GenericRegister {
- public:
-  using Key = KeyType;
-  using Entry = EntryType;
-
-  static RegisterType *GetRegister() {
-    static auto reg = new RegisterType;
-    return reg;
-  }
-
-  void SetEntry(const KeyType &key, const EntryType &entry) {
-    MutexLock l(&register_lock_);
-    register_table_.insert(std::make_pair(key, entry));
-  }
-
-  EntryType GetEntry(const KeyType &key) const {
-    const auto *entry = LookupEntry(key);
-    if (entry) {
-      return *entry;
-    } else {
-      return LoadEntryFromSharedObject(key);
-    }
-  }
-
-  virtual ~GenericRegister() {}
-
- protected:
-  // Override this if you want to be able to load missing definitions from
-  // shared object files.
-  virtual EntryType LoadEntryFromSharedObject(const KeyType &key) const {
-#ifdef FST_NO_DYNAMIC_LINKING
-    return EntryType();
-#else
-    const auto so_filename = ConvertKeyToSoFilename(key);
-    void *handle = dlopen(so_filename.c_str(), RTLD_LAZY);
-    if (handle == nullptr) {
-      LOG(ERROR) << "GenericRegister::GetEntry: " << dlerror();
-      return EntryType();
-    }
-#ifdef RUN_MODULE_INITIALIZERS
-    RUN_MODULE_INITIALIZERS();
-#endif
-    // We assume that the DSO constructs a static object in its global scope
-    // that does the registration. Thus we need only load it, not call any
-    // methods.
-    const auto *entry = this->LookupEntry(key);
-    if (entry == nullptr) {
-      LOG(ERROR) << "GenericRegister::GetEntry: "
-                 << "lookup failed in shared object: " << so_filename;
-      return EntryType();
-    }
-    return *entry;
-#endif  // FST_NO_DYNAMIC_LINKING
-  }
-
-  // Override this to define how to turn a key into an SO filename.
-  virtual string ConvertKeyToSoFilename(const KeyType &key) const = 0;
-
-  virtual const EntryType *LookupEntry(const KeyType &key) const {
-    MutexLock l(&register_lock_);
-    const auto it = register_table_.find(key);
-    if (it != register_table_.end()) {
-      return &it->second;
-    } else {
-      return nullptr;
-    }
-  }
-
- private:
-  mutable Mutex register_lock_;
-  std::map<KeyType, EntryType> register_table_;
-};
-
-// Generic register-er class capable of creating new register entries in the
-// given RegisterType template parameter. This type must define types Key and
-// Entry, and have appropriate static GetRegister() and instance SetEntry()
-// functions. An easy way to accomplish this is to have RegisterType be the
-// type of a subclass of GenericRegister.
-template <class RegisterType>
-class GenericRegisterer {
- public:
-  using Key = typename RegisterType::Key;
-  using Entry = typename RegisterType::Entry;
-
-  GenericRegisterer(Key key, Entry entry) {
-    RegisterType::GetRegister()->SetEntry(key, entry);
-  }
-};
-
-}  // namespace fst
-
-#endif  // FST_GENERIC_REGISTER_H_
diff --git a/projects/llm_framework/include/fst/heap.h b/projects/llm_framework/include/fst/heap.h
deleted file mode 100644
index 041a4bb9..00000000
--- a/projects/llm_framework/include/fst/heap.h
+++ /dev/null
@@ -1,168 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Implementation of a heap as in STL, but allows tracking positions in heap
-// using a key. The key can be used to do an in-place update of values in the
-// heap.
-
-#ifndef FST_HEAP_H_
-#define FST_HEAP_H_
-
-#include <utility>
-#include <vector>
-
-#include <fst/compat.h>
-namespace fst {
-
-// A templated heap implementation that supports in-place update of values.
-//
-// The templated heap implementation is a little different from the STL
-// priority_queue and the *_heap operations in STL. This heap supports
-// indexing of values in the heap via an associated key.
-//
-// Each value is internally associated with a key which is returned to the
-// calling functions on heap insert. This key can be used to later update
-// the specific value in the heap.
-//
-// T: the element type of the hash. It can be POD, Data or a pointer to Data.
-// Compare: comparison functor for determining min-heapness.
-template <class T, class Compare>
-class Heap {
- public:
-  using Value = T;
-
-  static constexpr int kNoKey = -1;
-
-  // Initializes with a specific comparator.
-  explicit Heap(Compare comp = Compare()) : comp_(comp), size_(0) {}
-
-  // Inserts a value into the heap.
-  int Insert(const Value &value) {
-    if (size_ < values_.size()) {
-      values_[size_] = value;
-      pos_[key_[size_]] = size_;
-    } else {
-      values_.push_back(value);
-      pos_.push_back(size_);
-      key_.push_back(size_);
-    }
-    ++size_;
-    return Insert(value, size_ - 1);
-  }
-
-  // Updates a value at position given by the key. The pos_ array is first
-  // indexed by the key. The position gives the position in the heap array.
-  // Once we have the position we can then use the standard heap operations
-  // to calculate the parent and child positions.
-  void Update(int key, const Value &value) {
-    const auto i = pos_[key];
-    const bool is_better = comp_(value, values_[Parent(i)]);
-    values_[i] = value;
-    if (is_better) {
-      Insert(value, i);
-    } else {
-      Heapify(i);
-    }
-  }
-
-  // Returns the least value.
-  Value Pop() {
-    Value top = values_.front();
-    Swap(0, size_-1);
-    size_--;
-    Heapify(0);
-    return top;
-  }
-
-  // Returns the least value w.r.t.  the comparison function from the
-  // heap.
-  const Value &Top() const { return values_.front(); }
-
-  // Returns the element for the given key.
-  const Value &Get(int key) const { return values_[pos_[key]]; }
-
-  // Checks if the heap is empty.
-  bool Empty() const { return size_ == 0; }
-
-  void Clear() { size_ = 0; }
-
-  int Size() const { return size_; }
-
-  void Reserve(int size) {
-    values_.reserve(size);
-    pos_.reserve(size);
-    key_.reserve(size);
-  }
-
-  const Compare &GetCompare() const { return comp_; }
-
- private:
-  // The following private routines are used in a supportive role
-  // for managing the heap and keeping the heap properties.
-
-  // Computes left child of parent.
-  static int Left(int i) {
-    return 2 * (i + 1) - 1;  // 0 -> 1, 1 -> 3
-  }
-
-  // Computes right child of parent.
-  static int Right(int i) {
-    return 2 * (i + 1);  // 0 -> 2, 1 -> 4
-  }
-
-  // Given a child computes parent.
-  static int Parent(int i) {
-    return (i - 1) / 2;  // 0 -> 0, 1 -> 0, 2 -> 0,  3 -> 1,  4 -> 1, ...
-  }
-
-  // Swaps a child and parent. Use to move element up/down tree. Note the use of
-  // a little trick here. When we swap we need to swap:
-  //
-  // - the value
-  // - the associated keys
-  // - the position of the value in the heap
-  void Swap(int j, int k) {
-    const auto tkey = key_[j];
-    pos_[key_[j] = key_[k]] = j;
-    pos_[key_[k] = tkey] = k;
-    using std::swap;
-    swap(values_[j], values_[k]);
-  }
-
-  // Heapifies the subtree rooted at index i.
-  void Heapify(int i) {
-    const auto l = Left(i);
-    const auto r = Right(i);
-    auto largest = (l < size_ && comp_(values_[l], values_[i])) ? l : i;
-    if (r < size_ && comp_(values_[r], values_[largest])) largest = r;
-    if (largest != i) {
-      Swap(i, largest);
-      Heapify(largest);
-    }
-  }
-
-  // Inserts (updates) element at subtree rooted at index i.
-  int Insert(const Value &value, int i) {
-    int p;
-    while (i > 0 && !comp_(values_[p = Parent(i)], value)) {
-      Swap(i, p);
-      i = p;
-    }
-    return key_[i];
-  }
-
- private:
-  const Compare comp_;
-
-  std::vector<int> pos_;
-  std::vector<int> key_;
-  std::vector<Value> values_;
-  int size_;
-};
-
-template <class T, class Compare>
-constexpr int Heap<T, Compare>::kNoKey;
-
-}  // namespace fst
-
-#endif  // FST_HEAP_H_
diff --git a/projects/llm_framework/include/fst/icu.h b/projects/llm_framework/include/fst/icu.h
deleted file mode 100644
index 5459da24..00000000
--- a/projects/llm_framework/include/fst/icu.h
+++ /dev/null
@@ -1,129 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// This library implements an unrestricted Thompson/Pike UTF-8 parser and
-// serializer. UTF-8 is a restricted subset of this byte stream encoding. For
-// a description of the encoding details, see:
-//
-//     http://en.wikipedia.org/wiki/UTF-8
-
-#ifndef FST_ICU_H_
-#define FST_ICU_H_
-
-#include <sstream>
-#include <vector>
-
-#include <fst/log.h>
-
-namespace fst {
-
-// Trivial function to copy bytestrings into vectors of labels, truncating
-// if necessary. It is possible to use this sensibly with as little as 8 bits
-// of Label precision. This returns `true` deterministically for compatibility.
-template <class Label>
-bool ByteStringToLabels(const string &str, std::vector<Label> *labels) {
-  for (const unsigned char ch : str) labels->push_back(ch);
-  return true;
-}
-
-// This function writes UTF-8 strings into a vector of Labels, truncating if
-// necessary. It is possible to use this sensibly with as little as 16 bits of
-// Label precision (i.e., when all characters are within the Basic Multilingual
-// Plane). With 21 bits, one can label all UTF-8 labelpoints, including those
-// from the various Astral Planes. Naturally, it is safe to use this with larger
-// Labels (e.g., 64 bits).
-template <class Label>
-bool UTF8StringToLabels(const string &str, std::vector<Label> *labels) {
-  for (auto it = str.begin(); it != str.end();) {
-    int c = *it & 0xff;
-    ++it;
-    if ((c & 0x80) == 0) {
-      labels->push_back(c);
-    } else {
-      if ((c & 0xc0) == 0x80) {
-        LOG(ERROR) << "UTF8StringToLabels: Continuation byte as lead byte";
-        return false;
-      }
-      int count =
-          (c >= 0xc0) + (c >= 0xe0) + (c >= 0xf0) + (c >= 0xf8) + (c >= 0xfc);
-      int32 label = c & ((1 << (6 - count)) - 1);
-      while (count != 0) {
-        if (it == str.end()) {
-          LOG(ERROR) << "UTF8StringToLabels: Truncated UTF-8 byte sequence";
-          return false;
-        }
-        char cb = *it;
-        ++it;
-        if ((cb & 0xc0) != 0x80) {
-          LOG(ERROR) << "UTF8StringToLabels: Missing/invalid continuation byte";
-          return false;
-        }
-        label = (label << 6) | (cb & 0x3f);
-        --count;
-      }
-      if (label < 0) {
-        // Should be unreachable.
-        LOG(ERROR) << "UTF8StringToLabels: Invalid character found: " << c;
-        return false;
-      }
-      labels->push_back(label);
-    }
-  }
-  return true;
-}
-
-template <class Label>
-bool LabelsToByteString(const std::vector<Label> &labels, string *str) {
-  std::ostringstream ostrm;
-  for (const char label : labels) {
-    if (label != 0) ostrm << label;
-  }
-  *str = ostrm.str();
-  return !!ostrm;
-}
-
-template <class Label>
-bool LabelsToUTF8String(const std::vector<Label> &labels, string *str) {
-  std::ostringstream ostrm;
-  for (const int32 label : labels) {
-    if (label < 0) {
-      LOG(ERROR) << "LabelsToUTF8String: Invalid character found: " << label;
-      return false;
-    } else if (label == 0) {
-      continue;
-    } else if (label < 0x80) {
-      ostrm << static_cast<char>(label);
-    } else if (label < 0x800) {
-      ostrm << static_cast<char>((label >> 6) | 0xc0);
-      ostrm << static_cast<char>((label & 0x3f) | 0x80);
-    } else if (label < 0x10000) {
-      ostrm << static_cast<char>((label >> 12) | 0xe0);
-      ostrm << static_cast<char>(((label >> 6) & 0x3f) | 0x80);
-      ostrm << static_cast<char>((label & 0x3f) | 0x80);
-    } else if (label < 0x200000) {
-      ostrm << static_cast<char>((label >> 18) | 0xf0);
-      ostrm << static_cast<char>(((label >> 12) & 0x3f) | 0x80);
-      ostrm << static_cast<char>(((label >> 6) & 0x3f) | 0x80);
-      ostrm << static_cast<char>((label & 0x3f) | 0x80);
-    } else if (label < 0x4000000) {
-      ostrm << static_cast<char>((label >> 24) | 0xf8);
-      ostrm << static_cast<char>(((label >> 18) & 0x3f) | 0x80);
-      ostrm << static_cast<char>(((label >> 12) & 0x3f) | 0x80);
-      ostrm << static_cast<char>(((label >> 6) & 0x3f) | 0x80);
-      ostrm << static_cast<char>((label & 0x3f) | 0x80);
-    } else {
-      ostrm << static_cast<char>((label >> 30) | 0xfc);
-      ostrm << static_cast<char>(((label >> 24) & 0x3f) | 0x80);
-      ostrm << static_cast<char>(((label >> 18) & 0x3f) | 0x80);
-      ostrm << static_cast<char>(((label >> 12) & 0x3f) | 0x80);
-      ostrm << static_cast<char>(((label >> 6) & 0x3f) | 0x80);
-      ostrm << static_cast<char>((label & 0x3f) | 0x80);
-    }
-  }
-  *str = ostrm.str();
-  return !!ostrm;
-}
-
-}  // namespace fst
-
-#endif  // FST_ICU_H_
diff --git a/projects/llm_framework/include/fst/intersect.h b/projects/llm_framework/include/fst/intersect.h
deleted file mode 100644
index fa66206d..00000000
--- a/projects/llm_framework/include/fst/intersect.h
+++ /dev/null
@@ -1,181 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to compute the intersection of two FSAs.
-
-#ifndef FST_INTERSECT_H_
-#define FST_INTERSECT_H_
-
-#include <algorithm>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/compose.h>
-
-
-namespace fst {
-
-using IntersectOptions = ComposeOptions;
-
-template <class Arc, class M = Matcher<Fst<Arc>>,
-          class Filter = SequenceComposeFilter<M>,
-          class StateTable =
-              GenericComposeStateTable<Arc, typename Filter::FilterState>>
-struct IntersectFstOptions
-    : public ComposeFstOptions<Arc, M, Filter, StateTable> {
-  IntersectFstOptions() {}
-
-  explicit IntersectFstOptions(const CacheOptions &opts, M *matcher1 = nullptr,
-                               M *matcher2 = nullptr, Filter *filter = nullptr,
-                               StateTable *state_table = nullptr)
-      : ComposeFstOptions<Arc, M, Filter, StateTable>(opts, matcher1, matcher2,
-                                                      filter, state_table) {}
-};
-
-// Computes the intersection (Hadamard product) of two FSAs. This version is a
-// delayed FST. Only strings that are in both automata are retained in the
-// result.
-//
-// The two arguments must be acceptors. One of the arguments must be
-// label-sorted.
-//
-// Complexity: same as ComposeFst.
-//
-// Caveats: same as ComposeFst.
-template <class A>
-class IntersectFst : public ComposeFst<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using ComposeFst<A>::CreateBase;
-  using ComposeFst<A>::CreateBase1;
-  using ComposeFst<A>::Properties;
-
-  IntersectFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-               const CacheOptions &opts = CacheOptions())
-      : ComposeFst<Arc>(CreateBase(fst1, fst2, opts)) {
-    const bool acceptors =
-        fst1.Properties(kAcceptor, true) && fst2.Properties(kAcceptor, true);
-    if (!acceptors) {
-      FSTERROR() << "IntersectFst: Input FSTs are not acceptors";
-      GetMutableImpl()->SetProperties(kError);
-    }
-  }
-
-  template <class M, class Filter, class StateTable>
-  IntersectFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-               const IntersectFstOptions<Arc, M, Filter, StateTable> &opts)
-      : ComposeFst<Arc>(CreateBase1(fst1, fst2, opts)) {
-    const bool acceptors =
-        fst1.Properties(kAcceptor, true) && fst2.Properties(kAcceptor, true);
-    if (!acceptors) {
-      FSTERROR() << "IntersectFst: input FSTs are not acceptors";
-      GetMutableImpl()->SetProperties(kError);
-    }
-  }
-
-  // See Fst<>::Copy() for doc.
-  IntersectFst(const IntersectFst<Arc> &fst, bool safe = false)
-      : ComposeFst<Arc>(fst, safe) {}
-
-  // Get a copy of this IntersectFst. See Fst<>::Copy() for further doc.
-  IntersectFst<Arc> *Copy(bool safe = false) const override {
-    return new IntersectFst<Arc>(*this, safe);
-  }
-
- private:
-  using ImplToFst<internal::ComposeFstImplBase<A>>::GetImpl;
-  using ImplToFst<internal::ComposeFstImplBase<A>>::GetMutableImpl;
-};
-
-// Specialization for IntersectFst.
-template <class Arc>
-class StateIterator<IntersectFst<Arc>> : public StateIterator<ComposeFst<Arc>> {
- public:
-  explicit StateIterator(const IntersectFst<Arc> &fst)
-      : StateIterator<ComposeFst<Arc>>(fst) {}
-};
-
-// Specialization for IntersectFst.
-template <class Arc>
-class ArcIterator<IntersectFst<Arc>> : public ArcIterator<ComposeFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const IntersectFst<Arc> &fst, StateId s)
-      : ArcIterator<ComposeFst<Arc>>(fst, s) {}
-};
-
-// Useful alias when using StdArc.
-using StdIntersectFst = IntersectFst<StdArc>;
-
-// Computes the intersection (Hadamard product) of two FSAs. This version
-// writes the intersection to an output MurableFst. Only strings that are in
-// both automata are retained in the result.
-//
-// The two arguments must be acceptors. One of the arguments must be
-// label-sorted.
-//
-// Complexity: same as Compose.
-//
-// Caveats: same as Compose.
-template <class Arc>
-void Intersect(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-               MutableFst<Arc> *ofst,
-               const IntersectOptions &opts = IntersectOptions()) {
-  using M = Matcher<Fst<Arc>>;
-  // In each case, we cache only the last state for fastest copy.
-  switch (opts.filter_type) {
-    case AUTO_FILTER: {
-      CacheOptions nopts;
-      nopts.gc_limit = 0;
-      *ofst = IntersectFst<Arc>(ifst1, ifst2, nopts);
-      break;
-    }
-    case SEQUENCE_FILTER: {
-      IntersectFstOptions<Arc> iopts;
-      iopts.gc_limit = 0;
-      *ofst = IntersectFst<Arc>(ifst1, ifst2, iopts);
-      break;
-    }
-    case ALT_SEQUENCE_FILTER: {
-      IntersectFstOptions<Arc, M, AltSequenceComposeFilter<M>> iopts;
-      iopts.gc_limit = 0;
-      *ofst = IntersectFst<Arc>(ifst1, ifst2, iopts);
-      break;
-    }
-    case MATCH_FILTER: {
-      IntersectFstOptions<Arc, M, MatchComposeFilter<M>> iopts;
-      iopts.gc_limit = 0;
-      *ofst = IntersectFst<Arc>(ifst1, ifst2, iopts);
-      break;
-    }
-    case NO_MATCH_FILTER: {
-      IntersectFstOptions<Arc, M, NoMatchComposeFilter<M>> iopts;
-      iopts.gc_limit = 0;
-      *ofst = IntersectFst<Arc>(ifst1, ifst2, iopts);
-      break;
-    }
-    case NULL_FILTER: {
-      IntersectFstOptions<Arc, M, NullComposeFilter<M>> iopts;
-      iopts.gc_limit = 0;
-      *ofst = IntersectFst<Arc>(ifst1, ifst2, iopts);
-      break;
-    }
-    case TRIVIAL_FILTER: {
-      IntersectFstOptions<Arc, M, TrivialComposeFilter<M>> iopts;
-      iopts.gc_limit = 0;
-      *ofst = IntersectFst<Arc>(ifst1, ifst2, iopts);
-      break;
-    }
-  }
-  if (opts.connect) Connect(ofst);
-}
-
-}  // namespace fst
-
-#endif  // FST_INTERSECT_H_
diff --git a/projects/llm_framework/include/fst/interval-set.h b/projects/llm_framework/include/fst/interval-set.h
deleted file mode 100644
index 0942ea00..00000000
--- a/projects/llm_framework/include/fst/interval-set.h
+++ /dev/null
@@ -1,398 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to represent and operate on sets of intervals.
-
-#ifndef FST_INTERVAL_SET_H_
-#define FST_INTERVAL_SET_H_
-
-#include <algorithm>
-#include <iostream>
-#include <vector>
-
-
-#include <fst/util.h>
-
-
-namespace fst {
-
-// Half-open integral interval [a, b) of signed integers of type T.
-template <class T>
-struct IntInterval {
-  T begin;
-  T end;
-
-  IntInterval() : begin(-1), end(-1) {}
-
-  IntInterval(T begin, T end) : begin(begin), end(end) {}
-
-  bool operator<(const IntInterval<T> &i) const {
-    return begin < i.begin || (begin == i.begin && end > i.end);
-  }
-
-  bool operator==(const IntInterval<T> &i) const {
-    return begin == i.begin && end == i.end;
-  }
-
-  bool operator!=(const IntInterval<T> &i) const {
-    return begin != i.begin || end != i.end;
-  }
-
-  std::istream &Read(std::istream &strm) {
-    T n;
-    ReadType(strm, &n);
-    begin = n;
-    ReadType(strm, &n);
-    end = n;
-    return strm;
-  }
-
-  std::ostream &Write(std::ostream &strm) const {
-    T n = begin;
-    WriteType(strm, n);
-    n = end;
-    WriteType(strm, n);
-    return strm;
-  }
-};
-
-// Stores IntIntervals<T> in a vector. In addition, keeps the count of points in
-// all intervals.
-template <class T>
-class VectorIntervalStore {
- public:
-  using Interval = IntInterval<T>;
-  using Iterator = typename std::vector<Interval>::const_iterator;
-
-  VectorIntervalStore() : count_(-1) {}
-
-  std::vector<Interval> *MutableIntervals() { return &intervals_; }
-
-  const Interval *Intervals() const { return intervals_.data(); }
-
-  T Size() const { return intervals_.size(); }
-
-  T Count() const { return count_; }
-
-  void SetCount(T count) { count_ = count; }
-
-  void Clear() {
-    intervals_.clear();
-    count_ = 0;
-  }
-
-  Iterator begin() const { return intervals_.begin(); }
-
-  Iterator end() const { return intervals_.end(); }
-
-  std::istream &Read(std::istream &strm) {
-    ReadType(strm, &intervals_);
-    return ReadType(strm, &count_);
-  }
-
-  std::ostream &Write(std::ostream &strm) const {
-    WriteType(strm, intervals_);
-    return WriteType(strm, count_);
-  }
-
- private:
-  std::vector<Interval> intervals_;
-  T count_;
-};
-
-// Stores and operates on a set of half-open integral intervals [a, b)
-// of signed integers of type T.
-template <class T, class Store = VectorIntervalStore<T>>
-class IntervalSet {
- public:
-  using Interval = IntInterval<T>;
-
-  template <class... A>
-  explicit IntervalSet(A... args) : intervals_(args...) {}
-
-  // Returns the interval set as a vector.
-  std::vector<Interval> *MutableIntervals() {
-    return intervals_.MutableIntervals();
-  }
-
-  // Returns a pointer to an array of Size() elements.
-  const Interval *Intervals() const { return intervals_.Intervals(); }
-
-  bool Empty() const { return Size() == 0; }
-
-  T Size() const { return intervals_.Size(); }
-
-  // Number of points in the intervals (undefined if not normalized).
-  T Count() const { return intervals_.Count(); }
-
-  void Clear() { intervals_.Clear(); }
-
-  // Adds an interval set to the set. The result may not be normalized.
-  void Union(const IntervalSet<T, Store> &iset) {
-    intervals_.MutableIntervals()->insert(intervals_.MutableIntervals()->end(),
-                                          iset.intervals_.begin(),
-                                          iset.intervals_.end());
-  }
-
-  // Requires intervals be normalized.
-  bool Member(T value) const {
-    const Interval interval(value, value);
-    auto lb = std::lower_bound(intervals_.begin(), intervals_.end(), interval);
-    if (lb == intervals_.begin()) return false;
-    return (--lb)->end > value;
-  }
-
-  // Requires intervals be normalized.
-  bool operator==(const IntervalSet<T, Store> &iset) const {
-    return Size() == iset.Size() &&
-           std::equal(intervals_.begin(), intervals_.end(),
-                      iset.intervals_.begin());
-  }
-
-  // Requires intervals be normalized.
-  bool operator!=(const IntervalSet<T, Store> &iset) const {
-    return Size() != iset.Size() ||
-           !std::equal(intervals_.begin(), intervals_.end(),
-                       iset.intervals_.begin());
-  }
-
-  bool Singleton() const {
-    return Size() == 1 &&
-           intervals_.begin()->begin + 1 == intervals_.begin()->end;
-  }
-
-  // Sorts, collapses overlapping and adjacent interals, and sets count.
-  void Normalize();
-
-  // Intersects an interval set with the set. Requires intervals be normalized.
-  // The result is normalized.
-  void Intersect(const IntervalSet<T, Store> &iset,
-                 IntervalSet<T, Store> *oset) const;
-
-  // Complements the set w.r.t [0, maxval). Requires intervals be normalized.
-  // The result is normalized.
-  void Complement(T maxval, IntervalSet<T, Store> *oset) const;
-
-  // Subtract an interval set from the set. Requires intervals be normalized.
-  // The result is normalized.
-  void Difference(const IntervalSet<T, Store> &iset,
-                  IntervalSet<T, Store> *oset) const;
-
-  // Determines if an interval set overlaps with the set. Requires intervals be
-  // normalized.
-  bool Overlaps(const IntervalSet<T, Store> &iset) const;
-
-  // Determines if an interval set overlaps with the set but neither is
-  // contained in the other. Requires intervals be normalized.
-  bool StrictlyOverlaps(const IntervalSet<T, Store> &iset) const;
-
-  // Determines if an interval set is contained within the set. Requires
-  // intervals be normalized.
-  bool Contains(const IntervalSet<T, Store> &iset) const;
-
-  std::istream &Read(std::istream &strm) { return intervals_.Read(strm); }
-
-  std::ostream &Write(std::ostream &strm) const {
-    return intervals_.Write(strm);
-  }
-
-  typename Store::Iterator begin() const { return intervals_.begin(); }
-
-  typename Store::Iterator end() const { return intervals_.end(); }
-
- private:
-  Store intervals_;
-};
-
-// Sorts, collapses overlapping and adjacent intervals, and sets count.
-template <typename T, class Store>
-void IntervalSet<T, Store>::Normalize() {
-  auto &intervals = *intervals_.MutableIntervals();
-  std::sort(intervals.begin(), intervals.end());
-  T count = 0;
-  T size = 0;
-  for (T i = 0; i < intervals.size(); ++i) {
-    auto &inti = intervals[i];
-    if (inti.begin == inti.end) continue;
-    for (T j = i + 1; j < intervals.size(); ++j) {
-      auto &intj = intervals[j];
-      if (intj.begin > inti.end) break;
-      if (intj.end > inti.end) inti.end = intj.end;
-      ++i;
-    }
-    count += inti.end - inti.begin;
-    intervals[size++] = inti;
-  }
-  intervals.resize(size);
-  intervals_.SetCount(count);
-}
-
-// Intersects an interval set with the set. Requires intervals be normalized.
-// The result is normalized.
-template <typename T, class Store>
-void IntervalSet<T, Store>::Intersect(const IntervalSet<T, Store> &iset,
-                                      IntervalSet<T, Store> *oset) const {
-  auto *ointervals = oset->MutableIntervals();
-  auto it1 = intervals_.begin();
-  auto it2 = iset.intervals_.begin();
-  ointervals->clear();
-  T count = 0;
-  while (it1 != intervals_.end() && it2 != iset.intervals_.end()) {
-    if (it1->end <= it2->begin) {
-      ++it1;
-    } else if (it2->end <= it1->begin) {
-      ++it2;
-    } else {
-      ointervals->emplace_back(std::max(it1->begin, it2->begin),
-                               std::min(it1->end, it2->end));
-      count += ointervals->back().end - ointervals->back().begin;
-      if (it1->end < it2->end) {
-        ++it1;
-      } else {
-        ++it2;
-      }
-    }
-  }
-  oset->intervals_.SetCount(count);
-}
-
-// Complements the set w.r.t [0, maxval). Requires intervals be normalized.
-// The result is normalized.
-template <typename T, class Store>
-void IntervalSet<T, Store>::Complement(T maxval,
-                                       IntervalSet<T, Store> *oset) const {
-  auto *ointervals = oset->MutableIntervals();
-  ointervals->clear();
-  T count = 0;
-  Interval interval;
-  interval.begin = 0;
-  for (auto it = intervals_.begin(); it != intervals_.end(); ++it) {
-    interval.end = std::min(it->begin, maxval);
-    if ((interval.begin) < (interval.end)) {
-      ointervals->push_back(interval);
-      count += interval.end - interval.begin;
-    }
-    interval.begin = it->end;
-  }
-  interval.end = maxval;
-  if ((interval.begin) < (interval.end)) {
-    ointervals->push_back(interval);
-    count += interval.end - interval.begin;
-  }
-  oset->intervals_.SetCount(count);
-}
-
-// Subtract an interval set from the set. Requires intervals be normalized.
-// The result is normalized.
-template <typename T, class Store>
-void IntervalSet<T, Store>::Difference(const IntervalSet<T, Store> &iset,
-                                       IntervalSet<T, Store> *oset) const {
-  if (Empty()) {
-    oset->MutableIntervals()->clear();
-    oset->intervals_.SetCount(0);
-  } else {
-    IntervalSet<T, Store> cset;
-    iset.Complement(intervals_.Intervals()[intervals_.Size() - 1].end, &cset);
-    Intersect(cset, oset);
-  }
-}
-
-// Determines if an interval set overlaps with the set. Requires intervals be
-// normalized.
-template <typename T, class Store>
-bool IntervalSet<T, Store>::Overlaps(const IntervalSet<T, Store> &iset) const {
-  auto it1 = intervals_.begin();
-  auto it2 = iset.intervals_.begin();
-  while (it1 != intervals_.end() && it2 != iset.intervals_.end()) {
-    if (it1->end <= it2->begin) {
-      ++it1;
-    } else if (it2->end <= it1->begin) {
-      ++it2;
-    } else {
-      return true;
-    }
-  }
-  return false;
-}
-
-// Determines if an interval set overlaps with the set but neither is contained
-// in the other. Requires intervals be normalized.
-template <typename T, class Store>
-bool IntervalSet<T, Store>::StrictlyOverlaps(
-    const IntervalSet<T, Store> &iset) const {
-  auto it1 = intervals_.begin();
-  auto it2 = iset.intervals_.begin();
-  bool only1 = false;    // Point in intervals_ but not intervals.
-  bool only2 = false;    // Point in intervals but not intervals_.
-  bool overlap = false;  // Point in both intervals_ and intervals.
-  while (it1 != intervals_.end() && it2 != iset.intervals_.end()) {
-    if (it1->end <= it2->begin) {  // no overlap - it1 first
-      only1 = true;
-      ++it1;
-    } else if (it2->end <= it1->begin) {  // no overlap - it2 first
-      only2 = true;
-      ++it2;
-    } else if (it2->begin == it1->begin && it2->end == it1->end) {  // equals
-      overlap = true;
-      ++it1;
-      ++it2;
-    } else if (it2->begin <= it1->begin && it2->end >= it1->end) {  // 1 c 2
-      only2 = true;
-      overlap = true;
-      ++it1;
-    } else if (it1->begin <= it2->begin && it1->end >= it2->end) {  // 2 c 1
-      only1 = true;
-      overlap = true;
-      ++it2;
-    } else {  // Strict overlap.
-      only1 = true;
-      only2 = true;
-      overlap = true;
-    }
-    if (only1 == true && only2 == true && overlap == true) return true;
-  }
-  if (it1 != intervals_.end()) only1 = true;
-  if (it2 != iset.intervals_.end()) only2 = true;
-  return only1 == true && only2 == true && overlap == true;
-}
-
-// Determines if an interval set is contained within the set. Requires intervals
-// be normalized.
-template <typename T, class Store>
-bool IntervalSet<T, Store>::Contains(const IntervalSet<T, Store> &iset) const {
-  if (iset.Count() > Count()) return false;
-  auto it1 = intervals_.begin();
-  auto it2 = iset.intervals_.begin();
-  while (it1 != intervals_.end() && it2 != iset.intervals_.end()) {
-    if ((it1->end) <= (it2->begin)) {  // No overlap; it1 first.
-      ++it1;
-    } else if ((it2->begin) < (it1->begin) ||
-               (it2->end) > (it1->end)) {  // No C.
-      return false;
-    } else if (it2->end == it1->end) {
-      ++it1;
-      ++it2;
-    } else {
-      ++it2;
-    }
-  }
-  return it2 == iset.intervals_.end();
-}
-
-template <typename T, class Store>
-std::ostream &operator<<(std::ostream &strm, const IntervalSet<T, Store> &s) {
-  strm << "{";
-  for (T i = 0; i < s.Size(); ++i) {
-    if (i > 0) {
-      strm << ",";
-    }
-    const auto &interval = s.Intervals()[i];
-    strm << "[" << interval.begin << "," << interval.end << ")";
-  }
-  strm << "}";
-  return strm;
-}
-
-}  // namespace fst
-
-#endif  // FST_INTERVAL_SET_H_
diff --git a/projects/llm_framework/include/fst/invert.h b/projects/llm_framework/include/fst/invert.h
deleted file mode 100644
index bd243c62..00000000
--- a/projects/llm_framework/include/fst/invert.h
+++ /dev/null
@@ -1,139 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to invert an FST.
-
-#ifndef FST_INVERT_H_
-#define FST_INVERT_H_
-
-#include <fst/arc-map.h>
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-// Mapper to implement inversion of an arc.
-template <class A>
-struct InvertMapper {
-  using FromArc = A;
-  using ToArc = A;
-
-  InvertMapper() {}
-
-  ToArc operator()(const FromArc &arc) const {
-    return ToArc(arc.olabel, arc.ilabel, arc.weight, arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const {
-     return MAP_NO_SUPERFINAL;
-  }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return InvertProperties(props);
-  }
-};
-
-// Inverts the transduction corresponding to an FST by exchanging the
-// FST's input and output labels.
-//
-// Complexity:
-//
-//   Time: O(V + E)
-//   Space: O(1)
-//
-// where V is the number of states and E is the number of arcs.
-template <class Arc>
-inline void Invert(const Fst<Arc> &ifst, MutableFst<Arc> *ofst) {
-  std::unique_ptr<SymbolTable> input(
-      ifst.InputSymbols() ? ifst.InputSymbols()->Copy() : nullptr);
-  std::unique_ptr<SymbolTable> output(
-      ifst.OutputSymbols() ? ifst.OutputSymbols()->Copy() : nullptr);
-  ArcMap(ifst, ofst, InvertMapper<Arc>());
-  ofst->SetInputSymbols(output.get());
-  ofst->SetOutputSymbols(input.get());
-}
-
-// Destructive variant of the above.
-template <class Arc>
-inline void Invert(MutableFst<Arc> *fst) {
-  std::unique_ptr<SymbolTable> input(
-      fst->InputSymbols() ? fst->InputSymbols()->Copy() : nullptr);
-  std::unique_ptr<SymbolTable> output(
-      fst->OutputSymbols() ? fst->OutputSymbols()->Copy() : nullptr);
-  ArcMap(fst, InvertMapper<Arc>());
-  fst->SetInputSymbols(output.get());
-  fst->SetOutputSymbols(input.get());
-}
-
-// Inverts the transduction corresponding to an FST by exchanging the
-// FST's input and output labels. This version is a delayed FST.
-//
-// Complexity:
-//
-//   Time: O(v + e)
-//   Space: O(1)
-//
-// where v is the number of states visited and e is the number of arcs visited.
-// Constant time and to visit an input state or arc is assumed and exclusive of
-// caching.
-template <class A>
-class InvertFst : public ArcMapFst<A, A, InvertMapper<A>> {
- public:
-  using Arc = A;
-
-  using Mapper = InvertMapper<Arc>;
-  using Impl = internal::ArcMapFstImpl<A, A, InvertMapper<A>>;
-
-  explicit InvertFst(const Fst<Arc> &fst)
-      : ArcMapFst<Arc, Arc, Mapper>(fst, Mapper()) {
-    GetMutableImpl()->SetOutputSymbols(fst.InputSymbols());
-    GetMutableImpl()->SetInputSymbols(fst.OutputSymbols());
-  }
-
-  // See Fst<>::Copy() for doc.
-  InvertFst(const InvertFst<Arc> &fst, bool safe = false)
-      : ArcMapFst<Arc, Arc, Mapper>(fst, safe) {}
-
-  // Get a copy of this InvertFst. See Fst<>::Copy() for further doc.
-  InvertFst<Arc> *Copy(bool safe = false) const override {
-    return new InvertFst(*this, safe);
-  }
-
- private:
-  using ImplToFst<Impl>::GetMutableImpl;
-};
-
-// Specialization for InvertFst.
-template <class Arc>
-class StateIterator<InvertFst<Arc>>
-    : public StateIterator<ArcMapFst<Arc, Arc, InvertMapper<Arc>>> {
- public:
-  explicit StateIterator(const InvertFst<Arc> &fst)
-      : StateIterator<ArcMapFst<Arc, Arc, InvertMapper<Arc>>>(fst) {}
-};
-
-// Specialization for InvertFst.
-template <class Arc>
-class ArcIterator<InvertFst<Arc>>
-    : public ArcIterator<ArcMapFst<Arc, Arc, InvertMapper<Arc>>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const InvertFst<Arc> &fst, StateId s)
-      : ArcIterator<ArcMapFst<Arc, Arc, InvertMapper<Arc>>>(fst, s) {}
-};
-
-// Useful alias when using StdArc.
-using StdInvertFst = InvertFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_INVERT_H_
diff --git a/projects/llm_framework/include/fst/isomorphic.h b/projects/llm_framework/include/fst/isomorphic.h
deleted file mode 100644
index b100b0a6..00000000
--- a/projects/llm_framework/include/fst/isomorphic.h
+++ /dev/null
@@ -1,183 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function to test two FSTs are isomorphic, i.e., they are equal up to a state
-// and arc re-ordering. FSTs should be deterministic when viewed as
-// unweighted automata.
-
-#ifndef FST_ISOMORPHIC_H_
-#define FST_ISOMORPHIC_H_
-
-#include <algorithm>
-#include <list>
-#include <type_traits>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/fst.h>
-
-
-namespace fst {
-namespace internal {
-
-// Orders weights for equality checking.
-template <class Weight, typename std::enable_if<
-                            IsIdempotent<Weight>::value>::type * = nullptr>
-bool WeightCompare(const Weight &w1, const Weight &w2, float delta,
-                   bool *error) {
-  return NaturalLess<Weight>()(w1, w2);
-}
-
-template <class Weight, typename std::enable_if<
-                            !IsIdempotent<Weight>::value>::type * = nullptr>
-bool WeightCompare(const Weight &w1, const Weight &w2, float delta,
-                   bool *error) {
-  // No natural order; use hash.
-  const auto q1 = w1.Quantize(delta);
-  const auto q2 = w2.Quantize(delta);
-  auto n1 = q1.Hash();
-  auto n2 = q2.Hash();
-  // Hash not unique; very unlikely to happen.
-  if (n1 == n2 && q1 != q2) {
-    VLOG(1) << "Isomorphic: Weight hash collision";
-    *error = true;
-  }
-  return n1 < n2;
-}
-
-template <class Arc>
-class Isomorphism {
-  using StateId = typename Arc::StateId;
-
- public:
-  Isomorphism(const Fst<Arc> &fst1, const Fst<Arc> &fst2, float delta)
-      : fst1_(fst1.Copy()),
-        fst2_(fst2.Copy()),
-        delta_(delta),
-        error_(false),
-        comp_(delta, &error_) {}
-
-  // Checks if input FSTs are isomorphic.
-  bool IsIsomorphic() {
-    if (fst1_->Start() == kNoStateId && fst2_->Start() == kNoStateId) {
-      return true;
-    }
-    if (fst1_->Start() == kNoStateId || fst2_->Start() == kNoStateId) {
-      return false;
-    }
-    PairState(fst1_->Start(), fst2_->Start());
-    while (!queue_.empty()) {
-      const auto &pr = queue_.front();
-      if (!IsIsomorphicState(pr.first, pr.second)) return false;
-      queue_.pop_front();
-    }
-    return true;
-  }
-
-  bool Error() const { return error_; }
-
- private:
-  // Orders arcs for equality checking.
-  class ArcCompare {
-   public:
-    ArcCompare(float delta, bool *error) : delta_(delta), error_(error) {}
-
-    bool operator()(const Arc &arc1, const Arc &arc2) const {
-      if (arc1.ilabel < arc2.ilabel) return true;
-      if (arc1.ilabel > arc2.ilabel) return false;
-      if (arc1.olabel < arc2.olabel) return true;
-      if (arc1.olabel > arc2.olabel) return false;
-      return WeightCompare(arc1.weight, arc2.weight, delta_, error_);
-    }
-
-   private:
-    float delta_;
-    bool *error_;
-  };
-
-  // Maintains state correspondences and queue.
-  bool PairState(StateId s1, StateId s2) {
-    if (state_pairs_.size() <= s1) state_pairs_.resize(s1 + 1, kNoStateId);
-    if (state_pairs_[s1] == s2) {
-      return true;  // already seen this pair
-    } else if (state_pairs_[s1] != kNoStateId) {
-      return false;  // s1 already paired with another s2
-    }
-    state_pairs_[s1] = s2;
-    queue_.push_back(std::make_pair(s1, s2));
-    return true;
-  }
-
-  // Checks if state pair is isomorphic
-  bool IsIsomorphicState(StateId s1, StateId s2);
-
-  std::unique_ptr<Fst<Arc>> fst1_;
-  std::unique_ptr<Fst<Arc>> fst2_;
-  float delta_;                          // Weight equality delta.
-  std::vector<Arc> arcs1_;               // For sorting arcs on FST1.
-  std::vector<Arc> arcs2_;               // For sorting arcs on FST2.
-  std::vector<StateId> state_pairs_;     // Maintains state correspondences.
-  std::list<std::pair<StateId, StateId>> queue_;  // Queue of state pairs.
-  bool error_;                           // Error flag.
-  ArcCompare comp_;
-};
-
-template <class Arc>
-bool Isomorphism<Arc>::IsIsomorphicState(StateId s1, StateId s2) {
-  if (!ApproxEqual(fst1_->Final(s1), fst2_->Final(s2), delta_)) return false;
-  auto narcs1 = fst1_->NumArcs(s1);
-  auto narcs2 = fst2_->NumArcs(s2);
-  if (narcs1 != narcs2) return false;
-  ArcIterator<Fst<Arc>> aiter1(*fst1_, s1);
-  ArcIterator<Fst<Arc>> aiter2(*fst2_, s2);
-  arcs1_.clear();
-  arcs1_.reserve(narcs1);
-  arcs2_.clear();
-  arcs2_.reserve(narcs2);
-  for (; !aiter1.Done(); aiter1.Next(), aiter2.Next()) {
-    arcs1_.push_back(aiter1.Value());
-    arcs2_.push_back(aiter2.Value());
-  }
-  std::sort(arcs1_.begin(), arcs1_.end(), comp_);
-  std::sort(arcs2_.begin(), arcs2_.end(), comp_);
-  for (size_t i = 0; i < arcs1_.size(); ++i) {
-    const auto &arc1 = arcs1_[i];
-    const auto &arc2 = arcs2_[i];
-    if (arc1.ilabel != arc2.ilabel) return false;
-    if (arc1.olabel != arc2.olabel) return false;
-    if (!ApproxEqual(arc1.weight, arc2.weight, delta_)) return false;
-    if (!PairState(arc1.nextstate, arc2.nextstate)) return false;
-    if (i > 0) {  // Checks for non-determinism.
-      const auto &arc0 = arcs1_[i - 1];
-      if (arc1.ilabel == arc0.ilabel && arc1.olabel == arc0.olabel &&
-          ApproxEqual(arc1.weight, arc0.weight, delta_)) {
-        VLOG(1) << "Isomorphic: Non-determinism as an unweighted automaton";
-        error_ = true;
-        return false;
-      }
-    }
-  }
-  return true;
-}
-
-}  // namespace internal
-
-// Tests if two FSTs have the same states and arcs up to a reordering.
-// Inputs should be non-deterministic when viewed as unweighted automata.
-template <class Arc>
-bool Isomorphic(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-                float delta = kDelta) {
-  internal::Isomorphism<Arc> iso(fst1, fst2, delta);
-  bool result = iso.IsIsomorphic();
-  if (iso.Error()) {
-    FSTERROR() << "Isomorphic: Cannot determine if inputs are isomorphic";
-    return false;
-  } else {
-    return result;
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_ISOMORPHIC_H_
diff --git a/projects/llm_framework/include/fst/label-reachable.h b/projects/llm_framework/include/fst/label-reachable.h
deleted file mode 100644
index f3d7f2bc..00000000
--- a/projects/llm_framework/include/fst/label-reachable.h
+++ /dev/null
@@ -1,511 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to determine if a non-epsilon label can be read as the first
-// non-epsilon symbol along some path from a given state.
-
-#ifndef FST_LABEL_REACHABLE_H_
-#define FST_LABEL_REACHABLE_H_
-
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/accumulator.h>
-#include <fst/arcsort.h>
-#include <fst/interval-set.h>
-#include <fst/state-reachable.h>
-#include <fst/util.h>
-#include <fst/vector-fst.h>
-
-
-namespace fst {
-
-// Stores shareable data for label reachable class copies.
-template <typename Label>
-class LabelReachableData {
- public:
-  using LabelIntervalSet = IntervalSet<Label>;
-  using Interval = typename LabelIntervalSet::Interval;
-
-  explicit LabelReachableData(bool reach_input, bool keep_relabel_data = true)
-      : reach_input_(reach_input),
-        keep_relabel_data_(keep_relabel_data),
-        have_relabel_data_(true),
-        final_label_(kNoLabel) {}
-
-  ~LabelReachableData() {}
-
-  bool ReachInput() const { return reach_input_; }
-
-  std::vector<LabelIntervalSet> *MutableIntervalSets() {
-    return &interval_sets_;
-  }
-
-  const LabelIntervalSet &GetIntervalSet(int s) const {
-    return interval_sets_[s];
-  }
-
-  int NumIntervalSets() const { return interval_sets_.size(); }
-
-  std::unordered_map<Label, Label> *Label2Index() {
-    if (!have_relabel_data_) {
-      FSTERROR() << "LabelReachableData: No relabeling data";
-    }
-    return &label2index_;
-  }
-
-  void SetFinalLabel(Label final_label) { final_label_ = final_label; }
-
-  Label FinalLabel() const { return final_label_; }
-
-  static LabelReachableData<Label> *Read(std::istream &istrm,
-                                         const FstReadOptions &opts) {
-    auto *data = new LabelReachableData<Label>();
-    ReadType(istrm, &data->reach_input_);
-    ReadType(istrm, &data->keep_relabel_data_);
-    data->have_relabel_data_ = data->keep_relabel_data_;
-    if (data->keep_relabel_data_) ReadType(istrm, &data->label2index_);
-    ReadType(istrm, &data->final_label_);
-    ReadType(istrm, &data->interval_sets_);
-    return data;
-  }
-
-  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
-    WriteType(ostrm, reach_input_);
-    WriteType(ostrm, keep_relabel_data_);
-    if (keep_relabel_data_) WriteType(ostrm, label2index_);
-    WriteType(ostrm, FinalLabel());
-    WriteType(ostrm, interval_sets_);
-    return true;
-  }
-
- private:
-  LabelReachableData() {}
-
-  bool reach_input_;                              // Input labels considered?
-  bool keep_relabel_data_;                        // Save label2index_ to file?
-  bool have_relabel_data_;                        // Using label2index_?
-  Label final_label_;                             // Final label.
-  std::unordered_map<Label, Label> label2index_;  // Finds index for a label.
-  std::vector<LabelIntervalSet> interval_sets_;   // Interval sets per state.
-};
-
-// Tests reachability of labels from a given state. If reach_input is true, then
-// input labels are considered, o.w. output labels are considered. To test for
-// reachability from a state s, first do SetState(s), then a label l can be
-// reached from state s of FST f iff Reach(r) is true where r = Relabel(l). The
-// relabeling is required to ensure a compact representation of the reachable
-// labels.
-
-// The whole FST can be relabeled instead with Relabel(&f, reach_input) so that
-// the test Reach(r) applies directly to the labels of the transformed FST f.
-// The relabeled FST will also be sorted appropriately for composition.
-//
-// Reachablity of a final state from state s (via an epsilon path) can be
-// tested with ReachFinal().
-//
-// Reachability can also be tested on the set of labels specified by an arc
-// iterator, useful for FST composition. In particular, Reach(aiter, ...) is
-// true if labels on the input (output) side of the transitions of the arc
-// iterator, when iter_input is true (false), can be reached from the state s.
-// The iterator labels must have already been relabeled.
-//
-// With the arc iterator test of reachability, the begin position, end position
-// and accumulated arc weight of the matches can be returned. The optional
-// template argument controls how reachable arc weights are accumulated. The
-// default uses semiring Plus(). Alternative ones can be used to distribute the
-// weights in composition in various ways.
-template <class Arc, class Accumulator = DefaultAccumulator<Arc>,
-          class D = LabelReachableData<typename Arc::Label>>
-class LabelReachable {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using Data = D;
-
-  using LabelIntervalSet = typename Data::LabelIntervalSet;
-
-  using Interval = typename LabelIntervalSet::Interval;
-
-  LabelReachable(const Fst<Arc> &fst, bool reach_input,
-                 Accumulator *accumulator = nullptr,
-                 bool keep_relabel_data = true)
-      : fst_(new VectorFst<Arc>(fst)),
-        s_(kNoStateId),
-        data_(std::make_shared<Data>(reach_input, keep_relabel_data)),
-        accumulator_(accumulator ? accumulator : new Accumulator()),
-        ncalls_(0),
-        nintervals_(0),
-        reach_fst_input_(false),
-        error_(false) {
-    const auto ins = fst_->NumStates();
-    TransformFst();
-    FindIntervals(ins);
-    fst_.reset();
-  }
-
-  explicit LabelReachable(std::shared_ptr<Data> data,
-                          Accumulator *accumulator = nullptr)
-      : s_(kNoStateId),
-        data_(std::move(data)),
-        accumulator_(accumulator ? accumulator : new Accumulator()),
-        ncalls_(0),
-        nintervals_(0),
-        reach_fst_input_(false),
-        error_(false) {}
-
-  LabelReachable(const LabelReachable<Arc, Accumulator, Data> &reachable,
-                 bool safe = false)
-      : s_(kNoStateId),
-        data_(reachable.data_),
-        accumulator_(new Accumulator(*reachable.accumulator_, safe)),
-        ncalls_(0),
-        nintervals_(0),
-        reach_fst_input_(reachable.reach_fst_input_),
-        error_(reachable.error_) {}
-
-  ~LabelReachable() {
-    if (ncalls_ > 0) {
-      VLOG(2) << "# of calls: " << ncalls_;
-      VLOG(2) << "# of intervals/call: " << (nintervals_ / ncalls_);
-    }
-  }
-
-  // Relabels w.r.t labels that give compact label sets.
-  Label Relabel(Label label) {
-    if (label == 0 || error_) return label;
-    auto &label2index = *data_->Label2Index();
-    auto &relabel = label2index[label];
-    if (!relabel) relabel = label2index.size() + 1;  // Adds new label.
-    return relabel;
-  }
-
-  // Relabels FST w.r.t to labels that give compact label sets.
-  void Relabel(MutableFst<Arc> *fst, bool relabel_input) {
-    for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
-         siter.Next()) {
-      for (MutableArcIterator<MutableFst<Arc>> aiter(fst, siter.Value());
-           !aiter.Done(); aiter.Next()) {
-        auto arc = aiter.Value();
-        if (relabel_input) {
-          arc.ilabel = Relabel(arc.ilabel);
-        } else {
-          arc.olabel = Relabel(arc.olabel);
-        }
-        aiter.SetValue(arc);
-      }
-    }
-    if (relabel_input) {
-      ArcSort(fst, ILabelCompare<Arc>());
-      fst->SetInputSymbols(nullptr);
-    } else {
-      ArcSort(fst, OLabelCompare<Arc>());
-      fst->SetOutputSymbols(nullptr);
-    }
-  }
-
-  // Returns relabeling pairs (cf. relabel.h::Relabel()). If avoid_collisions is
-  // true, extra pairs are added to ensure no collisions when relabeling
-  // automata that have labels unseen here.
-  void RelabelPairs(std::vector<std::pair<Label, Label>> *pairs,
-                    bool avoid_collisions = false) {
-    pairs->clear();
-    const auto &label2index = *data_->Label2Index();
-    // Maps labels to their new values in [1, label2index().size()].
-    for (auto it = label2index.begin(); it != label2index.end(); ++it) {
-      if (it->second != data_->FinalLabel()) {
-        pairs->push_back(std::make_pair(it->first, it->second));
-      }
-    }
-    if (avoid_collisions) {
-      // Ensures any label in [1, label2index().size()] is mapped either
-      // by the above step or to label2index() + 1 (to avoid collisions).
-      for (size_t i = 1; i <= label2index.size(); ++i) {
-        const auto it = label2index.find(i);
-        if (it == label2index.end() || it->second == data_->FinalLabel()) {
-          pairs->push_back(std::make_pair(i, label2index.size() + 1));
-        }
-      }
-    }
-  }
-
-  // Set current state. Optionally set state associated
-  // with arc iterator to be passed to Reach.
-  void SetState(StateId s, StateId aiter_s = kNoStateId) {
-    s_ = s;
-    if (aiter_s != kNoStateId) {
-      accumulator_->SetState(aiter_s);
-      if (accumulator_->Error()) error_ = true;
-    }
-  }
-
-  // Can reach this label from current state?
-  // Original labels must be transformed by the Relabel methods above.
-  bool Reach(Label label) const {
-    if (label == 0 || error_) return false;
-    return data_->GetIntervalSet(s_).Member(label);
-  }
-
-  // Can reach final state (via epsilon transitions) from this state?
-  bool ReachFinal() const {
-    if (error_) return false;
-    return data_->GetIntervalSet(s_).Member(data_->FinalLabel());
-  }
-
-  // Initialize with secondary FST to be used with Reach(Iterator,...).
-  // If reach_input = true, then arc input labels are considered in
-  // Reach(aiter, ...), o.w. output labels are considered. If copy is true, then
-  // the FST is a copy of the FST used in the previous call to this method
-  // (useful to avoid unnecessary updates).
-  template <class FST>
-  void ReachInit(const FST &fst, bool reach_input, bool copy = false) {
-    reach_fst_input_ = reach_input;
-    if (!fst.Properties(reach_fst_input_ ? kILabelSorted : kOLabelSorted,
-                        true)) {
-      FSTERROR() << "LabelReachable::ReachInit: Fst is not sorted";
-      error_ = true;
-    }
-    accumulator_->Init(fst, copy);
-    if (accumulator_->Error()) error_ = true;
-  }
-
-  // Can reach any arc iterator label between iterator positions
-  // aiter_begin and aiter_end?
-  // Arc iterator labels must be transformed by the Relabel methods
-  // above. If compute_weight is true, user may call ReachWeight().
-  template <class Iterator>
-  bool Reach(Iterator *aiter, ssize_t aiter_begin, ssize_t aiter_end,
-             bool compute_weight) {
-    if (error_) return false;
-    const auto &interval_set = data_->GetIntervalSet(s_);
-    ++ncalls_;
-    nintervals_ += interval_set.Size();
-    reach_begin_ = -1;
-    reach_end_ = -1;
-    reach_weight_ = Weight::Zero();
-    const auto flags = aiter->Flags();  // Save flags to restore them on exit.
-    aiter->SetFlags(kArcNoCache, kArcNoCache);  // Makes caching optional.
-    aiter->Seek(aiter_begin);
-    if (2 * (aiter_end - aiter_begin) < interval_set.Size()) {
-      // Checks each arc against intervals, setting arc iterator flags to only
-      // compute the ilabel or olabel values, since they are the only values
-      // required for most of the arcs processed.
-      aiter->SetFlags(reach_fst_input_ ? kArcILabelValue : kArcOLabelValue,
-                      kArcValueFlags);
-      Label reach_label = kNoLabel;
-      for (auto aiter_pos = aiter_begin; aiter_pos < aiter_end;
-           aiter->Next(), ++aiter_pos) {
-        const auto &arc = aiter->Value();
-        const auto label = reach_fst_input_ ? arc.ilabel : arc.olabel;
-        if (label == reach_label || Reach(label)) {
-          reach_label = label;
-          if (reach_begin_ < 0) reach_begin_ = aiter_pos;
-          reach_end_ = aiter_pos + 1;
-          if (compute_weight) {
-            if (!(aiter->Flags() & kArcWeightValue)) {
-              // If arc.weight wasn't computed by the call to aiter->Value()
-              // above, we need to call aiter->Value() again after having set
-              // the arc iterator flags to compute the arc weight value.
-              aiter->SetFlags(kArcWeightValue, kArcValueFlags);
-              const auto &arcb = aiter->Value();
-              // Call the accumulator.
-              reach_weight_ = accumulator_->Sum(reach_weight_, arcb.weight);
-              // Only ilabel or olabel required to process the following arcs.
-              aiter->SetFlags(
-                  reach_fst_input_ ? kArcILabelValue : kArcOLabelValue,
-                  kArcValueFlags);
-            } else {
-              // Calls the accumulator.
-              reach_weight_ = accumulator_->Sum(reach_weight_, arc.weight);
-            }
-          }
-        }
-      }
-    } else {
-      // Checks each interval against arcs.
-      auto begin_low = aiter_begin;
-      auto end_low = aiter_begin;
-      for (const auto &interval : interval_set) {
-        begin_low = LowerBound(aiter, end_low, aiter_end, interval.begin);
-        end_low = LowerBound(aiter, begin_low, aiter_end, interval.end);
-        if (end_low - begin_low > 0) {
-          if (reach_begin_ < 0) reach_begin_ = begin_low;
-          reach_end_ = end_low;
-          if (compute_weight) {
-            aiter->SetFlags(kArcWeightValue, kArcValueFlags);
-            reach_weight_ =
-                accumulator_->Sum(reach_weight_, aiter, begin_low, end_low);
-          }
-        }
-      }
-    }
-    aiter->SetFlags(flags, kArcFlags);  // Restores original flag values.
-    return reach_begin_ >= 0;
-  }
-
-  // Returns iterator position of first matching arc.
-  ssize_t ReachBegin() const { return reach_begin_; }
-
-  // Returns iterator position one past last matching arc.
-  ssize_t ReachEnd() const { return reach_end_; }
-
-  // Return the sum of the weights for matching arcs. Valid only if
-  // compute_weight was true in Reach() call.
-  Weight ReachWeight() const { return reach_weight_; }
-
-  // Access to the relabeling map. Excludes epsilon (0) label but
-  // includes kNoLabel that is used internally for super-final
-  // transitons.
-  const std::unordered_map<Label, Label> &Label2Index() const {
-    return *data_->Label2Index();
-  }
-
-  const Data *GetData() const { return data_.get(); }
-
-  std::shared_ptr<Data> GetSharedData() const { return data_; }
-
-  bool Error() const { return error_ || accumulator_->Error(); }
-
- private:
-  // Redirects labeled arcs (input or output labels determined by ReachInput())
-  // to new label-specific final states. Each original final state is
-  // redirected via a transition labeled with kNoLabel to a new
-  // kNoLabel-specific final state. Creates super-initial state for all states
-  // with zero in-degree.
-  void TransformFst() {
-    auto ins = fst_->NumStates();
-    auto ons = ins;
-    std::vector<ssize_t> indeg(ins, 0);
-    // Redirects labeled arcs to new final states.
-    for (StateId s = 0; s < ins; ++s) {
-      for (MutableArcIterator<VectorFst<Arc>> aiter(fst_.get(), s);
-           !aiter.Done(); aiter.Next()) {
-        auto arc = aiter.Value();
-        const auto label = data_->ReachInput() ? arc.ilabel : arc.olabel;
-        if (label) {
-          auto insert_result = label2state_.insert(std::make_pair(label, ons));
-          if (insert_result.second) {
-            indeg.push_back(0);
-            ++ons;
-          }
-          arc.nextstate = label2state_[label];
-          aiter.SetValue(arc);
-        }
-        ++indeg[arc.nextstate];  // Finds in-degrees for next step.
-      }
-      // Redirects final weights to new final state.
-      auto final_weight = fst_->Final(s);
-      if (final_weight != Weight::Zero()) {
-        auto insert_result = label2state_.insert(std::make_pair(kNoLabel, ons));
-        if (insert_result.second) {
-          indeg.push_back(0);
-          ++ons;
-        }
-        const auto nextstate = label2state_[kNoLabel];
-        fst_->EmplaceArc(s, kNoLabel, kNoLabel, std::move(final_weight),
-                         nextstate);
-        ++indeg[nextstate];  // Finds in-degrees for next step.
-        fst_->SetFinal(s, Weight::Zero());
-      }
-    }
-    // Adds new final states to the FST.
-    while (fst_->NumStates() < ons) {
-      StateId s = fst_->AddState();
-      fst_->SetFinal(s, Weight::One());
-    }
-    // Creates a super-initial state for all states with zero in-degree.
-    const auto start = fst_->AddState();
-    fst_->SetStart(start);
-    for (StateId s = 0; s < start; ++s) {
-      if (indeg[s] == 0) {
-        fst_->EmplaceArc(start, 0, 0, Weight::One(), s);
-      }
-    }
-  }
-
-  void FindIntervals(StateId ins) {
-    StateReachable<Arc, Label, LabelIntervalSet> state_reachable(*fst_);
-    if (state_reachable.Error()) {
-      error_ = true;
-      return;
-    }
-    auto &state2index = state_reachable.State2Index();
-    auto &interval_sets = *data_->MutableIntervalSets();
-    interval_sets = state_reachable.IntervalSets();
-    interval_sets.resize(ins);
-    auto &label2index = *data_->Label2Index();
-    for (const auto &kv : label2state_) {
-      Label i = state2index[kv.second];
-      label2index[kv.first] = i;
-      if (kv.first == kNoLabel) data_->SetFinalLabel(i);
-    }
-    label2state_.clear();
-    double nintervals = 0;
-    ssize_t non_intervals = 0;
-    for (StateId s = 0; s < ins; ++s) {
-      nintervals += interval_sets[s].Size();
-      if (interval_sets[s].Size() > 1) {
-        ++non_intervals;
-        VLOG(3) << "state: " << s
-                << " # of intervals: " << interval_sets[s].Size();
-      }
-    }
-    VLOG(2) << "# of states: " << ins;
-    VLOG(2) << "# of intervals: " << nintervals;
-    VLOG(2) << "# of intervals/state: " << nintervals / ins;
-    VLOG(2) << "# of non-interval states: " << non_intervals;
-  }
-
-  template <class Iterator>
-  ssize_t LowerBound(Iterator *aiter, ssize_t aiter_begin, ssize_t aiter_end,
-                     Label match_label) const {
-    // Only needs to compute the ilabel or olabel of arcs when performing the
-    // binary search.
-    aiter->SetFlags(reach_fst_input_ ? kArcILabelValue : kArcOLabelValue,
-                    kArcValueFlags);
-    ssize_t low = aiter_begin;
-    ssize_t high = aiter_end;
-    while (low < high) {
-      const ssize_t mid = low + (high - low) / 2;
-      aiter->Seek(mid);
-      auto label =
-          reach_fst_input_ ? aiter->Value().ilabel : aiter->Value().olabel;
-      if (label < match_label) {
-        low = mid + 1;
-      } else {
-        high = mid;
-      }
-    }
-    aiter->Seek(low);
-    aiter->SetFlags(kArcValueFlags, kArcValueFlags);
-    return low;
-  }
-
-  std::unique_ptr<VectorFst<Arc>> fst_;
-  // Current state
-  StateId s_;
-  // Finds final state for a label
-  std::unordered_map<Label, StateId> label2state_;
-  // Iterator position of first match.
-  ssize_t reach_begin_;
-  // Iterator position after last match.
-  ssize_t reach_end_;
-  // Gives weight sum of arc iterator arcs with reachable labels.
-  Weight reach_weight_;
-  // Shareable data between copies.
-  std::shared_ptr<Data> data_;
-  // Sums arc weights.
-  std::unique_ptr<Accumulator> accumulator_;
-  double ncalls_;
-  double nintervals_;
-  bool reach_fst_input_;
-  bool error_;
-};
-
-}  // namespace fst
-
-#endif  // FST_LABEL_REACHABLE_H_
diff --git a/projects/llm_framework/include/fst/lexicographic-weight.h b/projects/llm_framework/include/fst/lexicographic-weight.h
deleted file mode 100644
index af3c6db6..00000000
--- a/projects/llm_framework/include/fst/lexicographic-weight.h
+++ /dev/null
@@ -1,173 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Lexicographic weight set and associated semiring operation definitions.
-//
-// A lexicographic weight is a sequence of weights, each of which must have the
-// path property and Times() must be (strongly) cancellative
-// (for all a,b,c != Zero(): Times(c, a) = Times(c, b) => a = b,
-// Times(a, c) = Times(b, c) => a = b).
-// The + operation on two weights a and b is the lexicographically
-// prior of a and b.
-
-#ifndef FST_LEXICOGRAPHIC_WEIGHT_H_
-#define FST_LEXICOGRAPHIC_WEIGHT_H_
-
-#include <cstdlib>
-
-#include <string>
-
-#include <fst/log.h>
-
-#include <fst/pair-weight.h>
-#include <fst/weight.h>
-
-
-namespace fst {
-
-template <class W1, class W2>
-class LexicographicWeight : public PairWeight<W1, W2> {
- public:
-  using ReverseWeight = LexicographicWeight<typename W1::ReverseWeight,
-                                            typename W2::ReverseWeight>;
-
-  using PairWeight<W1, W2>::Value1;
-  using PairWeight<W1, W2>::Value2;
-  using PairWeight<W1, W2>::SetValue1;
-  using PairWeight<W1, W2>::SetValue2;
-  using PairWeight<W1, W2>::Zero;
-  using PairWeight<W1, W2>::One;
-  using PairWeight<W1, W2>::NoWeight;
-  using PairWeight<W1, W2>::Quantize;
-  using PairWeight<W1, W2>::Reverse;
-
-  LexicographicWeight() {}
-
-  explicit LexicographicWeight(const PairWeight<W1, W2> &w)
-      : PairWeight<W1, W2>(w) {}
-
-  LexicographicWeight(W1 w1, W2 w2) : PairWeight<W1, W2>(w1, w2) {
-    if ((W1::Properties() & kPath) != kPath) {
-      FSTERROR() << "LexicographicWeight must "
-                 << "have the path property: " << W1::Type();
-      SetValue1(W1::NoWeight());
-    }
-    if ((W2::Properties() & kPath) != kPath) {
-      FSTERROR() << "LexicographicWeight must "
-                 << "have the path property: " << W2::Type();
-      SetValue2(W2::NoWeight());
-    }
-  }
-
-  static const LexicographicWeight &Zero() {
-    static const LexicographicWeight zero(PairWeight<W1, W2>::Zero());
-    return zero;
-  }
-
-  static const LexicographicWeight &One() {
-    static const LexicographicWeight one(PairWeight<W1, W2>::One());
-    return one;
-  }
-
-  static const LexicographicWeight &NoWeight() {
-    static const LexicographicWeight no_weight(PairWeight<W1, W2>::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type =
-        new string(W1::Type() + "_LT_" + W2::Type());
-    return *type;
-  }
-
-  bool Member() const {
-    if (!Value1().Member() || !Value2().Member()) return false;
-    // Lexicographic weights cannot mix zeroes and non-zeroes.
-    if (Value1() == W1::Zero() && Value2() == W2::Zero()) return true;
-    if (Value1() != W1::Zero() && Value2() != W2::Zero()) return true;
-    return false;
-  }
-
-  LexicographicWeight Quantize(float delta = kDelta) const {
-    return LexicographicWeight(PairWeight<W1, W2>::Quantize());
-  }
-
-  ReverseWeight Reverse() const {
-    return ReverseWeight(PairWeight<W1, W2>::Reverse());
-  }
-
-  static constexpr uint64 Properties() {
-    return W1::Properties() & W2::Properties() &
-           (kLeftSemiring | kRightSemiring | kPath | kIdempotent |
-            kCommutative);
-  }
-};
-
-template <class W1, class W2>
-inline LexicographicWeight<W1, W2> Plus(const LexicographicWeight<W1, W2> &w,
-                                        const LexicographicWeight<W1, W2> &v) {
-  if (!w.Member() || !v.Member()) {
-    return LexicographicWeight<W1, W2>::NoWeight();
-  }
-  NaturalLess<W1> less1;
-  NaturalLess<W2> less2;
-  if (less1(w.Value1(), v.Value1())) return w;
-  if (less1(v.Value1(), w.Value1())) return v;
-  if (less2(w.Value2(), v.Value2())) return w;
-  if (less2(v.Value2(), w.Value2())) return v;
-  return w;
-}
-
-template <class W1, class W2>
-inline LexicographicWeight<W1, W2> Times(const LexicographicWeight<W1, W2> &w,
-                                         const LexicographicWeight<W1, W2> &v) {
-  return LexicographicWeight<W1, W2>(Times(w.Value1(), v.Value1()),
-                                     Times(w.Value2(), v.Value2()));
-}
-
-template <class W1, class W2>
-inline LexicographicWeight<W1, W2> Divide(const LexicographicWeight<W1, W2> &w,
-                                          const LexicographicWeight<W1, W2> &v,
-                                          DivideType typ = DIVIDE_ANY) {
-  return LexicographicWeight<W1, W2>(Divide(w.Value1(), v.Value1(), typ),
-                                     Divide(w.Value2(), v.Value2(), typ));
-}
-
-// This function object generates weights by calling the underlying generators
-// for the templated weight types, like all other pair weight types. However,
-// for lexicographic weights, we cannot generate zeroes for the two subweights
-// separately: weights are members iff both members are zero or both members
-// are non-zero. This is intended primarily for testing.
-template <class W1, class W2>
-class WeightGenerate<LexicographicWeight<W1, W2>> {
- public:
-  using Weight = LexicographicWeight<W1, W1>;
-  using Generate1 = WeightGenerate<W1>;
-  using Generate2 = WeightGenerate<W2>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t num_random_weights = kNumRandomWeights)
-      : generator1_(false, num_random_weights),
-        generator2_(false, num_random_weights), allow_zero_(allow_zero),
-        num_random_weights_(num_random_weights) {}
-
-  Weight operator()() const {
-    if (allow_zero_) {
-      const int n = rand() % (num_random_weights_ + 1);  // NOLINT
-      if (n == num_random_weights_) return Weight(W1::Zero(), W2::Zero());
-    }
-    return Weight(generator1_(), generator2_());
-  }
-
- private:
-  const Generate1 generator1_;
-  const Generate2 generator2_;
-  // Permits Zero() and zero divisors.
-  const bool allow_zero_;
-  // The number of alternative random weights.
-  const size_t num_random_weights_;
-};
-
-}  // namespace fst
-
-#endif  // FST_LEXICOGRAPHIC_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/lock.h b/projects/llm_framework/include/fst/lock.h
deleted file mode 100644
index 388d42f5..00000000
--- a/projects/llm_framework/include/fst/lock.h
+++ /dev/null
@@ -1,62 +0,0 @@
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Google-compatibility locking declarations and inline definitions.
-
-#ifndef FST_LIB_LOCK_H_
-#define FST_LIB_LOCK_H_
-
-#include <mutex>
-
-namespace fst {
-
-using namespace std;
-
-class Mutex {
- public:
-  Mutex() {}
-
-  inline void Lock() { mu_.lock(); }
-
-  inline void Unlock() { mu_.unlock(); }
-
- private:
-  std::mutex mu_;
-
-  Mutex(const Mutex &) = delete;
-  Mutex &operator=(const Mutex &) = delete;
-};
-
-class MutexLock {
- public:
-  explicit MutexLock(Mutex *mu) : mu_(mu) { mu_->Lock(); }
-
-  ~MutexLock() { mu_->Unlock(); }
-
- private:
-  Mutex *mu_;
-
-  MutexLock(const MutexLock &) = delete;
-  MutexLock &operator=(const MutexLock &) = delete;
-};
-
-// Currently, we don't use a separate reader lock.
-// TODO(kbg): Implement this with std::shared_mutex once C++17 becomes widely
-// available.
-using ReaderMutexLock = MutexLock;
-
-}  // namespace fst
-
-#endif  // FST_LIB_LOCK_H_
diff --git a/projects/llm_framework/include/fst/log.h b/projects/llm_framework/include/fst/log.h
deleted file mode 100644
index 6a8f3017..00000000
--- a/projects/llm_framework/include/fst/log.h
+++ /dev/null
@@ -1,78 +0,0 @@
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Google-style logging declarations and inline definitions.
-
-#ifndef FST_LIB_LOG_H_
-#define FST_LIB_LOG_H_
-
-#include <cassert>
-#include <iostream>
-#include <string>
-
-#include <fst/flags.h>
-#include <fst/types.h>
-
-using std::string;
-
-DECLARE_int32(v);
-
-class LogMessage {
- public:
-  LogMessage(const string& type) : fatal_(type == "FATAL") {
-    std::cerr << type << ": ";
-  }
-  ~LogMessage() {
-    std::cerr << std::endl;
-    if (fatal_) exit(1);
-  }
-  std::ostream& stream() { return std::cerr; }
-
- private:
-  bool fatal_;
-};
-
-// #define LOG(type) LogMessage(#type).stream()
-// #define VLOG(level) if ((level) <= FLAGS_v) LOG(INFO)
-
-// Checks
-inline void FstCheck(bool x, const char* expr, const char* file, int line) {
-  if (!x) {
-    LOG(FATAL) << "Check failed: \"" << expr << "\" file: " << file
-               << " line: " << line;
-  }
-}
-
-// #define CHECK(x) FstCheck(static_cast<bool>(x), #x, __FILE__, __LINE__)
-// #define CHECK_EQ(x, y) CHECK((x) == (y))
-// #define CHECK_LT(x, y) CHECK((x) < (y))
-// #define CHECK_GT(x, y) CHECK((x) > (y))
-// #define CHECK_LE(x, y) CHECK((x) <= (y))
-// #define CHECK_GE(x, y) CHECK((x) >= (y))
-// #define CHECK_NE(x, y) CHECK((x) != (y))
-
-// Debug checks
-// #define DCHECK(x) assert(x)
-// #define DCHECK_EQ(x, y) DCHECK((x) == (y))
-// #define DCHECK_LT(x, y) DCHECK((x) < (y))
-// #define DCHECK_GT(x, y) DCHECK((x) > (y))
-// #define DCHECK_LE(x, y) DCHECK((x) <= (y))
-// #define DCHECK_GE(x, y) DCHECK((x) >= (y))
-// #define DCHECK_NE(x, y) DCHECK((x) != (y))
-
-// Ports
-#define ATTRIBUTE_DEPRECATED __attribute__((deprecated))
-
-#endif  // FST_LIB_LOG_H_
diff --git a/projects/llm_framework/include/fst/lookahead-filter.h b/projects/llm_framework/include/fst/lookahead-filter.h
deleted file mode 100644
index df3fc6c9..00000000
--- a/projects/llm_framework/include/fst/lookahead-filter.h
+++ /dev/null
@@ -1,623 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Composition filters to support lookahead matchers, useful for improving
-// composition efficiency with certain inputs.
-
-#ifndef FST_LOOKAHEAD_FILTER_H_
-#define FST_LOOKAHEAD_FILTER_H_
-
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/filter-state.h>
-#include <fst/fst.h>
-#include <fst/lookahead-matcher.h>
-
-
-namespace fst {
-
-// Identifies and verifies the capabilities of the matcher to be used for
-// lookahead with the composition filters below. This version is passed two
-// matchers.
-template <class Matcher1, class Matcher2>
-MatchType LookAheadMatchType(const Matcher1 &m1, const Matcher2 &m2) {
-  const auto type1 = m1.Type(false);
-  const auto type2 = m2.Type(false);
-  if (type1 == MATCH_OUTPUT && m1.Flags() & kOutputLookAheadMatcher) {
-    return MATCH_OUTPUT;
-  } else if (type2 == MATCH_INPUT && m2.Flags() & kInputLookAheadMatcher) {
-    return MATCH_INPUT;
-  } else if (m1.Flags() & kOutputLookAheadMatcher &&
-             m1.Type(true) == MATCH_OUTPUT) {
-    return MATCH_OUTPUT;
-  } else if (m2.Flags() & kInputLookAheadMatcher &&
-             m2.Type(true) == MATCH_INPUT) {
-    return MATCH_INPUT;
-  } else {
-    return MATCH_NONE;
-  }
-}
-
-// Identifies and verifies the capabilities of the matcher to be used for
-// lookahead with the composition filters below. This version uses the FST's
-// default matchers.
-template <class Arc>
-MatchType LookAheadMatchType(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {
-  LookAheadMatcher<Fst<Arc>> matcher1(fst1, MATCH_OUTPUT);
-  LookAheadMatcher<Fst<Arc>> matcher2(fst2, MATCH_INPUT);
-  return LookAheadMatchType(matcher1, matcher2);
-}
-
-// LookAheadSelector is a helper class for selecting among possibly distinct
-// FST and matcher types without using a common base class. This lets us avoid
-// virtual function calls. It stores and returns the appropriate FSTs and
-// matcher for lookahead. It is templated on the matcher types. General case
-// has no methods.
-template <class Matcher1, class Matcher2, MatchType MT>
-class LookAheadSelector {};
-
-// Stores and returns the appropriate FST and matcher for lookahead. Specialized
-// for two matchers of same type with the (match) type argument determining
-// which is used for lookahead.
-template <class Matcher, MatchType MT>
-class LookAheadSelector<Matcher, Matcher, MT> {
- public:
-  using FST = typename Matcher::FST;
-
-  LookAheadSelector(Matcher *lmatcher1, Matcher *lmatcher2, MatchType type)
-      : lmatcher1_(lmatcher1->Copy()),
-        lmatcher2_(lmatcher2->Copy()),
-        type_(type) {}
-
-  LookAheadSelector(const LookAheadSelector<Matcher, Matcher, MT> &selector)
-      : lmatcher1_(selector.lmatcher1_->Copy()),
-        lmatcher2_(selector.lmatcher2_->Copy()),
-        type_(selector.type_) {}
-
-  const FST &GetFst() const {
-    return type_ == MATCH_OUTPUT ? lmatcher2_->GetFst() : lmatcher1_->GetFst();
-  }
-
-  Matcher *GetMatcher() const {
-    return type_ == MATCH_OUTPUT ? lmatcher1_.get() : lmatcher2_.get();
-  }
-
- private:
-  std::unique_ptr<Matcher> lmatcher1_;
-  std::unique_ptr<Matcher> lmatcher2_;
-  MatchType type_;
-};
-
-// Stores and returns the appropriate FST and matcher for lookahead.
-// Specialized for lookahead on input labels.
-template <class Matcher1, class Matcher2>
-class LookAheadSelector<Matcher1, Matcher2, MATCH_INPUT> {
- public:
-  using FST1 = typename Matcher1::FST;
-
-  LookAheadSelector(Matcher1 *lmatcher1, Matcher2 *lmatcher2, MatchType)
-      : fst_(lmatcher1->GetFst().Copy()), lmatcher_(lmatcher2->Copy()) {}
-
-  LookAheadSelector(
-      const LookAheadSelector<Matcher1, Matcher2, MATCH_INPUT> &selector)
-      : fst_(selector.fst_->Copy()), lmatcher_(selector.lmatcher_->Copy()) {}
-
-  const FST1 &GetFst() const { return *fst_; }
-
-  Matcher2 *GetMatcher() const { return lmatcher_.get(); }
-
- private:
-  std::unique_ptr<const FST1> fst_;
-  std::unique_ptr<Matcher2> lmatcher_;
-};
-
-// Stores and returns the appropriate FST and matcher for lookahead.
-// Specialized for lookahead on output labels.
-template <class Matcher1, class Matcher2>
-class LookAheadSelector<Matcher1, Matcher2, MATCH_OUTPUT> {
- public:
-  using FST2 = typename Matcher2::FST;
-
-  LookAheadSelector(Matcher1 *lmatcher1, Matcher2 *lmatcher2, MatchType)
-      : fst_(lmatcher2->GetFst().Copy()), lmatcher_(lmatcher1->Copy()) {}
-
-  LookAheadSelector(
-      const LookAheadSelector<Matcher1, Matcher2, MATCH_OUTPUT> &selector)
-      : fst_(selector.fst_->Copy()), lmatcher_(selector.lmatcher_->Copy()) {}
-
-  const FST2 &GetFst() const { return *fst_; }
-
-  Matcher1 *GetMatcher() const { return lmatcher_.get(); }
-
- private:
-  std::unique_ptr<const FST2> fst_;
-  std::unique_ptr<Matcher1> lmatcher_;
-};
-
-// This filter uses a lookahead matcher in FilterArc(arc1, arc2) to examine the
-// future of the composition state (arc1.nextstate, arc2.nextstate), blocking
-// moving forward when its determined to be
-// non-coaccessible. It is templated on an underlying filter, typically the
-// epsilon filter. Which matcher is the lookahead matcher is determined by the
-// template argument MT unless it is MATCH_BOTH. In that case, both matcher
-// arguments must be lookahead matchers of the same type and one will be
-// selected by LookAheadMatchType() based on their capability.
-template <class Filter, class M1 = LookAheadMatcher<typename Filter::FST1>,
-          class M2 = M1, MatchType MT = MATCH_BOTH>
-class LookAheadComposeFilter {
- public:
-  using Arc = typename Filter::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FST1 = typename Filter::FST1;
-  using FST2 = typename Filter::FST2;
-  using Matcher1 = typename Filter::Matcher1;
-  using Matcher2 = typename Filter::Matcher2;
-  using FilterState = typename Filter::FilterState;
-
-  LookAheadComposeFilter(const FST1 &fst1, const FST2 &fst2, M1 *matcher1,
-                         M2 *matcher2)
-      : filter_(fst1, fst2, matcher1, matcher2),
-        lookahead_type_(MT == MATCH_BOTH
-                            ? LookAheadMatchType(*filter_.GetMatcher1(),
-                                                 *filter_.GetMatcher2())
-                            : MT),
-        selector_(filter_.GetMatcher1(), filter_.GetMatcher2(),
-                  lookahead_type_),
-        flags_(lookahead_type_ == MATCH_OUTPUT
-                   ? filter_.GetMatcher1()->Flags()
-                   : filter_.GetMatcher2()->Flags()) {
-    if (lookahead_type_ == MATCH_NONE) {
-      FSTERROR() << "LookAheadComposeFilter: 1st argument cannot "
-                 << "match/look-ahead on output labels and 2nd argument "
-                 << "cannot match/look-ahead on input labels";
-    }
-    selector_.GetMatcher()->InitLookAheadFst(selector_.GetFst());
-  }
-
-  LookAheadComposeFilter(
-      const LookAheadComposeFilter<Filter, M1, M2, MT> &filter,
-      bool safe = false)
-      : filter_(filter.filter_, safe),
-        lookahead_type_(filter.lookahead_type_),
-        selector_(filter_.GetMatcher1(), filter_.GetMatcher2(),
-                  lookahead_type_),
-        flags_(filter.flags_) {
-    selector_.GetMatcher()->InitLookAheadFst(selector_.GetFst(), true);
-  }
-
-  FilterState Start() const { return filter_.Start(); }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    filter_.SetState(s1, s2, fs);
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    lookahead_arc_ = false;
-    const FilterState &fs = filter_.FilterArc(arc1, arc2);
-    if (fs == FilterState::NoState()) return FilterState::NoState();
-    return LookAheadOutput() ? LookAheadFilterArc(arc1, arc2, fs)
-                             : LookAheadFilterArc(arc2, arc1, fs);
-  }
-
-  void FilterFinal(Weight *weight1, Weight *weight2) const {
-    filter_.FilterFinal(weight1, weight2);
-  }
-
-  // Returns matchers; ownership stays with filter.
-
-  Matcher1 *GetMatcher1() { return filter_.GetMatcher1(); }
-
-  Matcher2 *GetMatcher2() { return filter_.GetMatcher2(); }
-
-  const LookAheadSelector<Matcher1, Matcher2, MT> &Selector() const {
-    return selector_;
-  }
-
-  uint64 Properties(uint64 inprops) const {
-    auto outprops = filter_.Properties(inprops);
-    if (lookahead_type_ == MATCH_NONE) outprops |= kError;
-    return outprops;
-  }
-
-  uint32 LookAheadFlags() const { return flags_; }
-
-  bool LookAheadArc() const { return lookahead_arc_; }
-
-  bool LookAheadOutput() const {
-    if (MT == MATCH_OUTPUT) {
-      return true;
-    } else if (MT == MATCH_INPUT) {
-      return false;
-    } else if (lookahead_type_ == MATCH_OUTPUT) {
-      return true;
-    } else {
-      return false;
-    }
-  }
-
- private:
-  FilterState LookAheadFilterArc(Arc *arca, Arc *arcb,
-                                 const FilterState &fs) const {
-    auto &labela = LookAheadOutput() ? arca->olabel : arca->ilabel;
-    if (labela != 0 && !(flags_ & kLookAheadNonEpsilons)) return fs;
-    if (labela == 0 && !(flags_ & kLookAheadEpsilons)) return fs;
-    lookahead_arc_ = true;
-    selector_.GetMatcher()->SetState(arca->nextstate);
-    return selector_.GetMatcher()->LookAheadFst(selector_.GetFst(),
-                                                arcb->nextstate)
-               ? fs
-               : FilterState::NoState();
-  }
-
-  Filter filter_;             // Underlying filter.
-  MatchType lookahead_type_;  // Lookahead match type.
-  LookAheadSelector<Matcher1, Matcher2, MT> selector_;
-  uint32 flags_;                // Lookahead flags.
-  mutable bool lookahead_arc_;  // Look-ahead performed at last FilterArc()?
-
-  LookAheadComposeFilter &operator=(const LookAheadComposeFilter &) = delete;
-};
-
-// This filter adds weight-pushing to a lookahead composition filter using the
-// LookAheadWeight() method of matcher argument. It is templated on an
-// underlying lookahead filter, typically the basic lookahead filter.
-// Weight-pushing in composition brings weights forward as much as possible
-// based on the lookahead information.
-template <class Filter, class M1 = LookAheadMatcher<typename Filter::FST1>,
-          class M2 = M1, MatchType MT = MATCH_BOTH>
-class PushWeightsComposeFilter {
- public:
-  using Arc = typename Filter::Arc;
-  using StateId = typename Filter::StateId;
-  using Weight = typename Filter::Weight;
-
-  using FST1 = typename Filter::FST1;
-  using FST2 = typename Filter::FST2;
-  using Matcher1 = typename Filter::Matcher1;
-  using Matcher2 = typename Filter::Matcher2;
-
-  using FilterState1 = typename Filter::FilterState;
-  using FilterState2 = WeightFilterState<Weight>;
-  using FilterState = PairFilterState<FilterState1, FilterState2>;
-
-  PushWeightsComposeFilter(const FST1 &fst1, const FST2 &fst2, M1 *matcher1,
-                           M2 *matcher2)
-      : filter_(fst1, fst2, matcher1, matcher2), fs_(FilterState::NoState()) {}
-
-  PushWeightsComposeFilter(
-      const PushWeightsComposeFilter<Filter, M1, M2, MT> &filter,
-      bool safe = false)
-      : filter_(filter.filter_, safe), fs_(FilterState::NoState()) {}
-
-  FilterState Start() const {
-    return FilterState(filter_.Start(), FilterState2(Weight::One()));
-  }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    fs_ = fs;
-    filter_.SetState(s1, s2, fs.GetState1());
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    const auto &fs1 = filter_.FilterArc(arc1, arc2);
-    if (fs1 == FilterState1::NoState()) return FilterState::NoState();
-    if (!(LookAheadFlags() & kLookAheadWeight)) {
-      return FilterState(fs1, FilterState2(Weight::One()));
-    }
-    const auto &lweight = filter_.LookAheadArc()
-                              ? Selector().GetMatcher()->LookAheadWeight()
-                              : Weight::One();
-    const auto &fs2 = fs_.GetState2();
-    const auto &fweight = fs2.GetWeight();
-    // Disallows Zero() weight futures.
-    if (lweight == Weight::Zero()) return FilterState::NoState();
-    arc2->weight = Divide(Times(arc2->weight, lweight), fweight);
-    return FilterState(fs1, FilterState2(lweight.Quantize()));
-  }
-
-  void FilterFinal(Weight *weight1, Weight *weight2) const {
-    filter_.FilterFinal(weight1, weight2);
-    if (!(LookAheadFlags() & kLookAheadWeight) || *weight1 == Weight::Zero()) {
-      return;
-    }
-    const auto &fs2 = fs_.GetState2();
-    const auto &fweight = fs2.GetWeight();
-    *weight1 = Divide(*weight1, fweight);
-  }
-
-  // Returns matchers; ownership states with filter.
-
-  Matcher1 *GetMatcher1() { return filter_.GetMatcher1(); }
-
-  Matcher2 *GetMatcher2() { return filter_.GetMatcher2(); }
-
-  const LookAheadSelector<Matcher1, Matcher2, MT> &Selector() const {
-    return filter_.Selector();
-  }
-
-  uint32 LookAheadFlags() const { return filter_.LookAheadFlags(); }
-
-  bool LookAheadArc() const { return filter_.LookAheadArc(); }
-
-  bool LookAheadOutput() const { return filter_.LookAheadOutput(); }
-
-  uint64 Properties(uint64 props) const {
-    return filter_.Properties(props) & kWeightInvariantProperties;
-  }
-
- private:
-  Filter filter_;   // Underlying filter.
-  FilterState fs_;  // Current filter state.
-
-  PushWeightsComposeFilter &operator=(const PushWeightsComposeFilter &) =
-      delete;
-};
-
-// This filter adds label-pushing to a lookahead composition filter using the
-// LookAheadPrefix() method of the matcher argument. It is templated on an
-// underlying filter, typically the basic lookahead or weight-pushing lookahead
-// filter. Label-pushing in composition matches labels as early as possible
-// based on the lookahead information.
-template <class Filter, class M1 = LookAheadMatcher<typename Filter::FST1>,
-          class M2 = M1, MatchType MT = MATCH_BOTH>
-class PushLabelsComposeFilter {
- public:
-  using Arc = typename Filter::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FST1 = typename Filter::FST1;
-  using FST2 = typename Filter::FST2;
-  using Matcher1 = MultiEpsMatcher<typename Filter::Matcher1>;
-  using Matcher2 = MultiEpsMatcher<typename Filter::Matcher2>;
-  using FilterState1 = typename Filter::FilterState;
-  using FilterState2 = IntegerFilterState<Label>;
-  using FilterState = PairFilterState<FilterState1, FilterState2>;
-
-  PushLabelsComposeFilter(const FST1 &fst1, const FST2 &fst2, M1 *matcher1,
-                          M2 *matcher2)
-      : filter_(fst1, fst2, matcher1, matcher2),
-        fs_(FilterState::NoState()),
-        fst1_(filter_.GetMatcher1()->GetFst()),
-        fst2_(filter_.GetMatcher2()->GetFst()),
-        matcher1_(fst1_, MATCH_OUTPUT,
-                  filter_.LookAheadOutput() ? kMultiEpsList : kMultiEpsLoop,
-                  filter_.GetMatcher1(), false),
-        matcher2_(fst2_, MATCH_INPUT,
-                  filter_.LookAheadOutput() ? kMultiEpsLoop : kMultiEpsList,
-                  filter_.GetMatcher2(), false) {}
-
-  PushLabelsComposeFilter(
-      const PushLabelsComposeFilter<Filter, M1, M2, MT> &filter,
-      bool safe = false)
-      : filter_(filter.filter_, safe),
-        fs_(FilterState::NoState()),
-        fst1_(filter_.GetMatcher1()->GetFst()),
-        fst2_(filter_.GetMatcher2()->GetFst()),
-        matcher1_(fst1_, MATCH_OUTPUT,
-                  filter_.LookAheadOutput() ? kMultiEpsList : kMultiEpsLoop,
-                  filter_.GetMatcher1(), false),
-        matcher2_(fst2_, MATCH_INPUT,
-                  filter_.LookAheadOutput() ? kMultiEpsLoop : kMultiEpsList,
-                  filter_.GetMatcher2(), false) {}
-
-  FilterState Start() const {
-    return FilterState(filter_.Start(), FilterState2(kNoLabel));
-  }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    fs_ = fs;
-    filter_.SetState(s1, s2, fs.GetState1());
-    if (!(LookAheadFlags() & kLookAheadPrefix)) return;
-    narcsa_ = LookAheadOutput() ? internal::NumArcs(fst1_, s1)
-                                : internal::NumArcs(fst2_, s2);
-    const auto &fs2 = fs_.GetState2();
-    const auto &flabel = fs2.GetState();
-    GetMatcher1()->ClearMultiEpsLabels();
-    GetMatcher2()->ClearMultiEpsLabels();
-    if (flabel != kNoLabel) {                   // Have a lookahead label?
-      GetMatcher1()->AddMultiEpsLabel(flabel);  // Yes, make it a multi-epsilon
-      GetMatcher2()->AddMultiEpsLabel(flabel);  // label so that it matches the
-    }                                           // implicit epsilon arc to be
-  }                                             // modified below when pushing.
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    if (!(LookAheadFlags() & kLookAheadPrefix)) {
-      return FilterState(filter_.FilterArc(arc1, arc2), FilterState2(kNoLabel));
-    }
-    const auto &fs2 = fs_.GetState2();
-    const auto &flabel = fs2.GetState();
-    if (flabel != kNoLabel) {  // Have a lookahead label?
-      return LookAheadOutput() ? PushedLabelFilterArc(arc1, arc2, flabel)
-                               : PushedLabelFilterArc(arc2, arc1, flabel);
-    }
-    const auto &fs1 = filter_.FilterArc(arc1, arc2);
-    if (fs1 == FilterState1::NoState()) return FilterState::NoState();
-    if (!filter_.LookAheadArc())
-      return FilterState(fs1, FilterState2(kNoLabel));
-    return LookAheadOutput() ? PushLabelFilterArc(arc1, arc2, fs1)
-                             : PushLabelFilterArc(arc2, arc1, fs1);
-  }
-
-  void FilterFinal(Weight *weight1, Weight *weight2) const {
-    filter_.FilterFinal(weight1, weight2);
-    if (!(LookAheadFlags() & kLookAheadPrefix) || *weight1 == Weight::Zero()) {
-      return;
-    }
-    const auto &fs2 = fs_.GetState2();
-    const auto &flabel = fs2.GetState();
-    if (flabel != kNoLabel) *weight1 = Weight::Zero();
-  }
-
-  // Returns matchers; ownership states with filter.
-
-  Matcher1 *GetMatcher1() { return &matcher1_; }
-
-  Matcher2 *GetMatcher2() { return &matcher2_; }
-
-  uint64 Properties(uint64 iprops) const {
-    const auto oprops = filter_.Properties(iprops);
-    if (LookAheadOutput()) {
-      return oprops & kOLabelInvariantProperties;
-    } else {
-      return oprops & kILabelInvariantProperties;
-    }
-  }
-
- private:
-  const LookAheadSelector<typename Filter::Matcher1, typename Filter::Matcher2,
-                          MT>
-      &Selector() const {
-    return filter_.Selector();
-  }
-
-  // Consumes an already pushed label.
-  FilterState PushedLabelFilterArc(Arc *arca, Arc *arcb, Label flabel) const {
-    auto &labela = LookAheadOutput() ? arca->olabel : arca->ilabel;
-    const auto &labelb = LookAheadOutput() ? arcb->ilabel : arcb->olabel;
-    if (labelb != kNoLabel) {
-      return FilterState::NoState();  // Blocks non-(multi-)epsilon label
-    } else if (labela == flabel) {
-      labela = 0;  // Converts match to multi-epsilon to epsilon.
-      return Start();
-    } else if (labela == 0) {
-      if (narcsa_ == 1) return fs_;  // Takes epsilon, keeping state with label.
-      Selector().GetMatcher()->SetState(arca->nextstate);
-      if (Selector().GetMatcher()->LookAheadLabel(flabel)) {
-        return fs_;  // Takes epsilon, keeping state with label.
-      } else {
-        return FilterState::NoState();  // Blocks non-coaccessible path.
-      }
-    } else {
-      return FilterState::NoState();  // Blocks mismatch to multi-epsilon label.
-    }
-  }
-
-  // Pushes a label forward when possible.
-  FilterState PushLabelFilterArc(Arc *arca, Arc *arcb,
-                                 const FilterState1 &fs1) const {
-    auto &labela = LookAheadOutput() ? arca->olabel : arca->ilabel;
-    const auto &labelb = LookAheadOutput() ? arcb->olabel : arcb->ilabel;
-    if (labelb != 0) {  // No place to push.
-      return FilterState(fs1, FilterState2(kNoLabel));
-    }
-    if (labela != 0 &&  // Wrong lookahead prefix type?
-        LookAheadFlags() & kLookAheadNonEpsilonPrefix) {
-      return FilterState(fs1, FilterState2(kNoLabel));
-    }
-    Arc larc(kNoLabel, kNoLabel, Weight::Zero(), kNoStateId);
-    if (Selector().GetMatcher()->LookAheadPrefix(&larc)) {  // Have prefix arc?
-      labela = LookAheadOutput() ? larc.ilabel : larc.olabel;
-      arcb->ilabel = larc.ilabel;  // Goes forward on that arc,
-      arcb->olabel = larc.olabel;  // thus pushing the label.
-      arcb->weight = Times(arcb->weight, larc.weight);
-      arcb->nextstate = larc.nextstate;
-      return FilterState(fs1, FilterState2(labela));
-    } else {
-      return FilterState(fs1, FilterState2(kNoLabel));
-    }
-  }
-
-  uint32 LookAheadFlags() const { return filter_.LookAheadFlags(); }
-
-  bool LookAheadArc() const { return filter_.LookAheadArc(); }
-
-  bool LookAheadOutput() const { return filter_.LookAheadOutput(); }
-
-  Filter filter_;   // Underlying filter.
-  FilterState fs_;  // Current filter state.
-  const FST1 &fst1_;
-  const FST2 &fst2_;
-  Matcher1 matcher1_;  // Multi-epsilon matcher for fst1_.
-  Matcher2 matcher2_;  // Multi-epsilon matcher for fst2_.
-  ssize_t narcsa_;     // Number of arcs leaving look-ahead match FST.
-
-  PushLabelsComposeFilter &operator=(const PushLabelsComposeFilter &) = delete;
-};
-
-// Convenience class for setting up composition with a default lookahead matcher
-// and filter.
-template <class Arc, MatchType type>
-class DefaultLookAhead {
- public:
-  using M = Matcher<Fst<Arc>>;
-  using ComposeFilter = SequenceComposeFilter<M>;
-  using FstMatcher = M;
-};
-
-// Specializes for MATCH_INPUT to allow lookahead.
-template <class Arc>
-class DefaultLookAhead<Arc, MATCH_INPUT> {
- public:
-  using M = LookAheadMatcher<Fst<Arc>>;
-  using SF = SequenceComposeFilter<M>;
-  using ComposeFilter = LookAheadComposeFilter<SF, M>;
-  using FstMatcher = M;
-};
-
-// Specializes for MATCH_OUTPUT to allow lookahead.
-template <class Arc>
-class DefaultLookAhead<Arc, MATCH_OUTPUT> {
- public:
-  using M = LookAheadMatcher<Fst<Arc>>;
-  using SF = AltSequenceComposeFilter<M>;
-  using ComposeFilter = LookAheadComposeFilter<SF, M>;
-  using FstMatcher = M;
-};
-
-// Specializes for StdArc to allow weight and label pushing.
-template <>
-class DefaultLookAhead<StdArc, MATCH_INPUT> {
- public:
-  using M = LookAheadMatcher<Fst<StdArc>>;
-  using SF = SequenceComposeFilter<M>;
-  using LF = LookAheadComposeFilter<SF, M>;
-  using WF = PushWeightsComposeFilter<LF, M>;
-  using ComposeFilter = PushLabelsComposeFilter<WF, M>;
-  using FstMatcher = M;
-};
-
-// Specializes for StdArc to allow weight and label pushing.
-template <>
-class DefaultLookAhead<StdArc, MATCH_OUTPUT> {
- public:
-  using M = LookAheadMatcher<Fst<StdArc>>;
-  using SF = AltSequenceComposeFilter<M>;
-  using LF = LookAheadComposeFilter<SF, M>;
-  using WF = PushWeightsComposeFilter<LF, M>;
-  using ComposeFilter = PushLabelsComposeFilter<WF, M>;
-  using FstMatcher = M;
-};
-
-// Specializes for LogArc to allow weight and label pushing.
-template <>
-class DefaultLookAhead<LogArc, MATCH_INPUT> {
- public:
-  using M = LookAheadMatcher<Fst<LogArc>>;
-  using SF = SequenceComposeFilter<M>;
-  using LF = LookAheadComposeFilter<SF, M>;
-  using WF = PushWeightsComposeFilter<LF, M>;
-  using ComposeFilter = PushLabelsComposeFilter<WF, M>;
-  using FstMatcher = M;
-};
-
-// Specializes for LogArc to allow weight and label pushing.
-template <>
-class DefaultLookAhead<LogArc, MATCH_OUTPUT> {
- public:
-  using M = LookAheadMatcher<Fst<LogArc>>;
-  using SF = AltSequenceComposeFilter<M>;
-  using LF = LookAheadComposeFilter<SF, M>;
-  using WF = PushWeightsComposeFilter<LF, M>;
-  using ComposeFilter = PushLabelsComposeFilter<WF, M>;
-  using FstMatcher = M;
-};
-
-}  // namespace fst
-
-#endif  // FST_LOOKAHEAD_FILTER_H_
diff --git a/projects/llm_framework/include/fst/lookahead-matcher.h b/projects/llm_framework/include/fst/lookahead-matcher.h
deleted file mode 100644
index d542cfea..00000000
--- a/projects/llm_framework/include/fst/lookahead-matcher.h
+++ /dev/null
@@ -1,841 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes to add lookahead to FST matchers, useful for improving composition
-// efficiency with certain inputs.
-
-#ifndef FST_LOOKAHEAD_MATCHER_H_
-#define FST_LOOKAHEAD_MATCHER_H_
-
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include <fst/flags.h>
-#include <fst/log.h>
-
-#include <fst/add-on.h>
-#include <fst/const-fst.h>
-#include <fst/fst.h>
-#include <fst/label-reachable.h>
-#include <fst/matcher.h>
-
-
-DECLARE_string(save_relabel_ipairs);
-DECLARE_string(save_relabel_opairs);
-
-namespace fst {
-
-// Lookahead matches extend the matcher interface with following additional
-// methods:
-//
-// template <class FST>
-// class LookAheadMatcher {
-//  public:
-//   using Arc = typename FST::Arc;
-//   using Label = typename Arc::Label;
-//   using StateId = typename Arc::StateId;
-//   using Weight = typename Arc::Weight;
-//
-//  // Required constructors.
-//  // This makes a copy of the FST.
-//  LookAheadMatcher(const FST &fst, MatchType match_type);
-//  // This doesn't copy the FST.
-//  LookAheadMatcher(const FST *fst, MatchType match_type);
-//  // This makes a copy of the FST.
-//  // See Copy() below.
-//  LookAheadMatcher(const LookAheadMatcher &matcher, bool safe = false);
-//
-//   // If safe = true, the copy is thread-safe (except the lookahead FST is
-//   // preserved). See Fst<>::Copy() for further doc.
-//   LookaheadMatcher<FST> *Copy(bool safe = false) const override;
-
-//  // Below are methods for looking ahead for a match to a label and more
-//  // generally, to a rational set. Each returns false if there is definitely
-//  // not a match and returns true if there possibly is a match.
-//
-//  // Optionally pre-specifies the lookahead FST that will be passed to
-//  // LookAheadFst() for possible precomputation. If copy is true, then the FST
-//  // argument is a copy of the FST used in the previous call to this method
-//  // (to avoid unnecessary updates).
-//  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) override;
-//
-//  // Are there paths from a state in the lookahead FST that can be read from
-//  // the curent matcher state?
-//  bool LookAheadFst(const Fst<Arc> &fst, StateId s) override;
-//
-//  // Can the label be read from the current matcher state after possibly
-//  // following epsilon transitions?
-//  bool LookAheadLabel(Label label) const override;
-//
-//  // The following methods allow looking ahead for an arbitrary rational set
-//  // of strings, specified by an FST and a state from which to begin the
-//  // matching. If the lookahead FST is a transducer, this looks on the side
-//  // different from the matcher's match_type (cf. composition).
-//  // Is there is a single non-epsilon arc found in the lookahead FST that
-//  // begins the path (after possibly following any epsilons) in the last call
-//  // to LookAheadFst? If so, return true and copy it to the arc argument;
-//  // otherwise, return false. Non-trivial implementations are useful for
-//  // label-pushing in composition.
-//  bool LookAheadPrefix(Arc *arc) override;
-//
-//  // Gives an estimate of the combined weight of the paths in the lookahead
-//  // and matcher FSTs for the last call to LookAheadFst. Non-trivial
-//  // implementations are useful for weight-pushing in composition.
-//  Weight LookAheadWeight() const override;
-// };
-
-// Look-ahead flags.
-// Matcher is a lookahead matcher when match_type is MATCH_INPUT.
-constexpr uint32 kInputLookAheadMatcher = 0x00000010;
-
-// Matcher is a lookahead matcher when match_type is MATCH_OUTPUT.
-constexpr uint32 kOutputLookAheadMatcher = 0x00000020;
-
-// Is a non-trivial implementation of LookAheadWeight() method defined and
-// if so, should it be used?
-constexpr uint32 kLookAheadWeight = 0x00000040;
-
-// Is a non-trivial implementation of LookAheadPrefix() method defined and
-// if so, should it be used?
-constexpr uint32 kLookAheadPrefix = 0x00000080;
-
-// Look-ahead of matcher FST non-epsilon arcs?
-constexpr uint32 kLookAheadNonEpsilons = 0x00000100;
-
-// Look-ahead of matcher FST epsilon arcs?
-constexpr uint32 kLookAheadEpsilons = 0x00000200;
-
-// Ignore epsilon paths for the lookahead prefix? This gives correct results in
-// composition only with an appropriate composition filter since it depends on
-// the filter blocking the ignored paths.
-constexpr uint32 kLookAheadNonEpsilonPrefix = 0x00000400;
-
-// For LabelLookAheadMatcher, save relabeling data to file?
-constexpr uint32 kLookAheadKeepRelabelData = 0x00000800;
-
-// Flags used for lookahead matchers.
-constexpr uint32 kLookAheadFlags = 0x00000ff0;
-
-// LookAhead Matcher interface, templated on the Arc definition; used
-// for lookahead matcher specializations that are returned by the
-// InitMatcher() Fst method.
-template <class Arc>
-class LookAheadMatcherBase : public MatcherBase<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  virtual void InitLookAheadFst(const Fst<Arc> &, bool copy = false) = 0;
-  virtual bool LookAheadFst(const Fst<Arc> &, StateId) = 0;
-  virtual bool LookAheadLabel(Label) const = 0;
-
-  // Suggested concrete implementation of lookahead methods.
-
-  bool LookAheadPrefix(Arc *arc) const {
-    if (prefix_arc_.nextstate != kNoStateId) {
-      *arc = prefix_arc_;
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  Weight LookAheadWeight() const { return weight_; }
-
- protected:
-  // Concrete implementations for lookahead helper methods.
-
-  void ClearLookAheadWeight() { weight_ = Weight::One(); }
-
-  void SetLookAheadWeight(Weight weight) { weight_ = std::move(weight); }
-
-  void ClearLookAheadPrefix() { prefix_arc_.nextstate = kNoStateId; }
-
-  void SetLookAheadPrefix(Arc arc) { prefix_arc_ = std::move(arc); }
-
- private:
-  Arc prefix_arc_;
-  Weight weight_;
-};
-
-// Doesn't actually lookahead, just declares that the future looks good.
-template <class M>
-class TrivialLookAheadMatcher
-    : public LookAheadMatcherBase<typename M::FST::Arc> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // This makes a copy of the FST.
-  TrivialLookAheadMatcher(const FST &fst, MatchType match_type)
-      : matcher_(fst, match_type) {}
-
-  // This doesn't copy the FST.
-  TrivialLookAheadMatcher(const FST *fst, MatchType match_type)
-      : matcher_(fst, match_type) {}
-
-  // This makes a copy of the FST.
-  TrivialLookAheadMatcher(const TrivialLookAheadMatcher<M> &lmatcher,
-                          bool safe = false)
-      : matcher_(lmatcher.matcher_, safe) {}
-
-  TrivialLookAheadMatcher<M> *Copy(bool safe = false) const override {
-    return new TrivialLookAheadMatcher<M>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override { return matcher_.Type(test); }
-
-  void SetState(StateId s) final { return matcher_.SetState(s); }
-
-  bool Find(Label label) final { return matcher_.Find(label); }
-
-  bool Done() const final { return matcher_.Done(); }
-
-  const Arc &Value() const final { return matcher_.Value(); }
-
-  void Next() final { matcher_.Next(); }
-
-  Weight Final(StateId s) const final { return matcher_.Final(s); }
-
-  ssize_t Priority(StateId s) final { return matcher_.Priority(s); }
-
-  const FST &GetFst() const override { return matcher_.GetFst(); }
-
-  uint64 Properties(uint64 props) const override {
-    return matcher_.Properties(props);
-  }
-
-  uint32 Flags() const override {
-    return matcher_.Flags() | kInputLookAheadMatcher | kOutputLookAheadMatcher;
-  }
-
-  // Lookahead methods (all trivial).
-
-  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) override {}
-
-  bool LookAheadFst(const Fst<Arc> &, StateId) final { return true; }
-
-  bool LookAheadLabel(Label) const final { return true; }
-
-  bool LookAheadPrefix(Arc *) const { return false; }
-
-  Weight LookAheadWeight() const { return Weight::One(); }
-
- private:
-  M matcher_;
-};
-
-// Look-ahead of one transition. Template argument flags accepts flags to
-// control behavior.
-template <class M,
-          uint32 flags = kLookAheadNonEpsilons | kLookAheadEpsilons |
-                         kLookAheadWeight | kLookAheadPrefix>
-class ArcLookAheadMatcher : public LookAheadMatcherBase<typename M::FST::Arc> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using MatcherData = NullAddOn;
-
-  using LookAheadMatcherBase<Arc>::ClearLookAheadWeight;
-  using LookAheadMatcherBase<Arc>::LookAheadWeight;
-  using LookAheadMatcherBase<Arc>::SetLookAheadWeight;
-  using LookAheadMatcherBase<Arc>::ClearLookAheadPrefix;
-  using LookAheadMatcherBase<Arc>::LookAheadPrefix;
-  using LookAheadMatcherBase<Arc>::SetLookAheadPrefix;
-
-  enum : uint32 { kFlags = flags };
-
-  // This makes a copy of the FST.
-  ArcLookAheadMatcher(
-      const FST &fst, MatchType match_type,
-      std::shared_ptr<MatcherData> data = std::shared_ptr<MatcherData>())
-      : matcher_(fst, match_type),
-        fst_(matcher_.GetFst()),
-        lfst_(nullptr),
-        state_(kNoStateId) {}
-
-  // This doesn't copy the FST.
-  ArcLookAheadMatcher(
-      const FST *fst, MatchType match_type,
-      std::shared_ptr<MatcherData> data = std::shared_ptr<MatcherData>())
-      : matcher_(fst, match_type),
-        fst_(matcher_.GetFst()),
-        lfst_(nullptr),
-        state_(kNoStateId) {}
-
-  // This makes a copy of the FST.
-  ArcLookAheadMatcher(const ArcLookAheadMatcher<M, flags> &lmatcher,
-                      bool safe = false)
-      : matcher_(lmatcher.matcher_, safe),
-        fst_(matcher_.GetFst()),
-        lfst_(lmatcher.lfst_),
-        state_(kNoStateId) {}
-
-  // General matcher methods.
-  ArcLookAheadMatcher<M, flags> *Copy(bool safe = false) const override {
-    return new ArcLookAheadMatcher<M, flags>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override { return matcher_.Type(test); }
-
-  void SetState(StateId s) final {
-    state_ = s;
-    matcher_.SetState(s);
-  }
-
-  bool Find(Label label) final { return matcher_.Find(label); }
-
-  bool Done() const final { return matcher_.Done(); }
-
-  const Arc &Value() const final { return matcher_.Value(); }
-
-  void Next() final { matcher_.Next(); }
-
-  Weight Final(StateId s) const final { return matcher_.Final(s); }
-
-  ssize_t Priority(StateId s) final { return matcher_.Priority(s); }
-
-  const FST &GetFst() const override { return fst_; }
-
-  uint64 Properties(uint64 props) const override {
-    return matcher_.Properties(props);
-  }
-
-  uint32 Flags() const override {
-    return matcher_.Flags() | kInputLookAheadMatcher | kOutputLookAheadMatcher |
-           kFlags;
-  }
-
-  const MatcherData *GetData() const { return nullptr; }
-
-  std::shared_ptr<MatcherData> GetSharedData() const {
-    return std::shared_ptr<MatcherData>();
-  }
-
-  // Look-ahead methods.
-
-  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) override {
-    lfst_ = &fst;
-  }
-
-  // Checks if there is a matching (possibly super-final) transition
-  // at (state_, s).
-  bool LookAheadFst(const Fst<Arc> &, StateId) final;
-
-  bool LookAheadLabel(Label label) const final { return matcher_.Find(label); }
-
- private:
-  mutable M matcher_;
-  const FST &fst_;        // Matcher FST.
-  const Fst<Arc> *lfst_;  // Look-ahead FST.
-  StateId state_;         // Matcher state.
-};
-
-template <class M, uint32 flags>
-bool ArcLookAheadMatcher<M, flags>::LookAheadFst(const Fst<Arc> &fst,
-                                                 StateId s) {
-  if (&fst != lfst_) InitLookAheadFst(fst);
-  bool result = false;
-  ssize_t nprefix = 0;
-  if (kFlags & kLookAheadWeight) ClearLookAheadWeight();
-  if (kFlags & kLookAheadPrefix) ClearLookAheadPrefix();
-  if (fst_.Final(state_) != Weight::Zero() &&
-      lfst_->Final(s) != Weight::Zero()) {
-    if (!(kFlags & (kLookAheadWeight | kLookAheadPrefix))) return true;
-    ++nprefix;
-    if (kFlags & kLookAheadWeight) {
-      SetLookAheadWeight(
-          Plus(LookAheadWeight(), Times(fst_.Final(state_), lfst_->Final(s))));
-    }
-    result = true;
-  }
-  if (matcher_.Find(kNoLabel)) {
-    if (!(kFlags & (kLookAheadWeight | kLookAheadPrefix))) return true;
-    ++nprefix;
-    if (kFlags & kLookAheadWeight) {
-      for (; !matcher_.Done(); matcher_.Next()) {
-        SetLookAheadWeight(Plus(LookAheadWeight(), matcher_.Value().weight));
-      }
-    }
-    result = true;
-  }
-  for (ArcIterator<Fst<Arc>> aiter(*lfst_, s); !aiter.Done(); aiter.Next()) {
-    const auto &arc = aiter.Value();
-    Label label = kNoLabel;
-    switch (matcher_.Type(false)) {
-      case MATCH_INPUT:
-        label = arc.olabel;
-        break;
-      case MATCH_OUTPUT:
-        label = arc.ilabel;
-        break;
-      default:
-        FSTERROR() << "ArcLookAheadMatcher::LookAheadFst: Bad match type";
-        return true;
-    }
-    if (label == 0) {
-      if (!(kFlags & (kLookAheadWeight | kLookAheadPrefix))) return true;
-      if (!(kFlags & kLookAheadNonEpsilonPrefix)) ++nprefix;
-      if (kFlags & kLookAheadWeight) {
-        SetLookAheadWeight(Plus(LookAheadWeight(), arc.weight));
-      }
-      result = true;
-    } else if (matcher_.Find(label)) {
-      if (!(kFlags & (kLookAheadWeight | kLookAheadPrefix))) return true;
-      for (; !matcher_.Done(); matcher_.Next()) {
-        ++nprefix;
-        if (kFlags & kLookAheadWeight) {
-          SetLookAheadWeight(Plus(LookAheadWeight(),
-                                  Times(arc.weight, matcher_.Value().weight)));
-        }
-        if ((kFlags & kLookAheadPrefix) && nprefix == 1)
-          SetLookAheadPrefix(arc);
-      }
-      result = true;
-    }
-  }
-  if (kFlags & kLookAheadPrefix) {
-    if (nprefix == 1) {
-      ClearLookAheadWeight();  // Avoids double counting.
-    } else {
-      ClearLookAheadPrefix();
-    }
-  }
-  return result;
-}
-
-// Template argument flags accepts flags to control behavior. It must include
-// precisely one of kInputLookAheadMatcher or kOutputLookAheadMatcher.
-template <class M,
-          uint32 flags = kLookAheadEpsilons | kLookAheadWeight |
-                         kLookAheadPrefix | kLookAheadNonEpsilonPrefix |
-                         kLookAheadKeepRelabelData,
-          class Accumulator = DefaultAccumulator<typename M::Arc>,
-          class Reachable = LabelReachable<typename M::Arc, Accumulator>>
-class LabelLookAheadMatcher
-    : public LookAheadMatcherBase<typename M::FST::Arc> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using MatcherData = typename Reachable::Data;
-
-  using LookAheadMatcherBase<Arc>::ClearLookAheadWeight;
-  using LookAheadMatcherBase<Arc>::LookAheadWeight;
-  using LookAheadMatcherBase<Arc>::SetLookAheadWeight;
-  using LookAheadMatcherBase<Arc>::ClearLookAheadPrefix;
-  using LookAheadMatcherBase<Arc>::LookAheadPrefix;
-  using LookAheadMatcherBase<Arc>::SetLookAheadPrefix;
-
-  enum : uint32 { kFlags = flags };
-
-  // This makes a copy of the FST.
-  LabelLookAheadMatcher(
-      const FST &fst, MatchType match_type,
-      std::shared_ptr<MatcherData> data = std::shared_ptr<MatcherData>(),
-      Accumulator *accumulator = nullptr)
-      : matcher_(fst, match_type),
-        lfst_(nullptr),
-        state_(kNoStateId),
-        error_(false) {
-    Init(fst, match_type, data, accumulator);
-  }
-
-  // This doesn't copy the FST.
-  LabelLookAheadMatcher(
-      const FST *fst, MatchType match_type,
-      std::shared_ptr<MatcherData> data = std::shared_ptr<MatcherData>(),
-      Accumulator *accumulator = nullptr)
-      : matcher_(fst, match_type),
-        lfst_(nullptr),
-        state_(kNoStateId),
-        error_(false) {
-    Init(*fst, match_type, data, accumulator);
-  }
-
-  // This makes a copy of the FST.
-  LabelLookAheadMatcher(
-      const LabelLookAheadMatcher<M, flags, Accumulator, Reachable> &lmatcher,
-      bool safe = false)
-      : matcher_(lmatcher.matcher_, safe),
-        lfst_(lmatcher.lfst_),
-        label_reachable_(lmatcher.label_reachable_
-                             ? new Reachable(*lmatcher.label_reachable_, safe)
-                             : nullptr),
-        state_(kNoStateId),
-        error_(lmatcher.error_) {}
-
-  LabelLookAheadMatcher<M, flags, Accumulator, Reachable> *Copy(
-      bool safe = false) const override {
-    return new LabelLookAheadMatcher<M, flags, Accumulator, Reachable>(*this,
-                                                                       safe);
-  }
-
-  MatchType Type(bool test) const override { return matcher_.Type(test); }
-
-  void SetState(StateId s) final {
-    if (state_ == s) return;
-    state_ = s;
-    match_set_state_ = false;
-    reach_set_state_ = false;
-  }
-
-  bool Find(Label label) final {
-    if (!match_set_state_) {
-      matcher_.SetState(state_);
-      match_set_state_ = true;
-    }
-    return matcher_.Find(label);
-  }
-
-  bool Done() const final { return matcher_.Done(); }
-
-  const Arc &Value() const final { return matcher_.Value(); }
-
-  void Next() final { matcher_.Next(); }
-
-  Weight Final(StateId s) const final { return matcher_.Final(s); }
-
-  ssize_t Priority(StateId s) final { return matcher_.Priority(s); }
-
-  const FST &GetFst() const override { return matcher_.GetFst(); }
-
-  uint64 Properties(uint64 inprops) const override {
-    auto outprops = matcher_.Properties(inprops);
-    if (error_ || (label_reachable_ && label_reachable_->Error())) {
-      outprops |= kError;
-    }
-    return outprops;
-  }
-
-  uint32 Flags() const override {
-    if (label_reachable_ && label_reachable_->GetData()->ReachInput()) {
-      return matcher_.Flags() | kFlags | kInputLookAheadMatcher;
-    } else if (label_reachable_ && !label_reachable_->GetData()->ReachInput()) {
-      return matcher_.Flags() | kFlags | kOutputLookAheadMatcher;
-    } else {
-      return matcher_.Flags();
-    }
-  }
-
-  const MatcherData *GetData() const {
-    return label_reachable_ ? label_reachable_->GetData() : nullptr;
-  };
-
-  std::shared_ptr<MatcherData> GetSharedData() const {
-    return label_reachable_ ? label_reachable_->GetSharedData()
-                            : std::shared_ptr<MatcherData>();
-  }
-  // Checks if there is a matching (possibly super-final) transition at
-  // (state_, s).
-  template <class LFST>
-  bool LookAheadFst(const LFST &fst, StateId s);
-
-  // Required to make class concrete.
-  bool LookAheadFst(const Fst<Arc> &fst, StateId s) final {
-    return LookAheadFst<Fst<Arc>>(fst, s);
-  }
-
-  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) override {
-    lfst_ = &fst;
-    if (label_reachable_) {
-      const bool reach_input = Type(false) == MATCH_OUTPUT;
-      label_reachable_->ReachInit(fst, reach_input, copy);
-    }
-  }
-
-  template <class LFST>
-  void InitLookAheadFst(const LFST &fst, bool copy = false) {
-    lfst_ = static_cast<const Fst<Arc> *>(&fst);
-    if (label_reachable_) {
-      const bool reach_input = Type(false) == MATCH_OUTPUT;
-      label_reachable_->ReachInit(fst, reach_input, copy);
-    }
-  }
-
-  bool LookAheadLabel(Label label) const final {
-    if (label == 0) return true;
-    if (label_reachable_) {
-      if (!reach_set_state_) {
-        label_reachable_->SetState(state_);
-        reach_set_state_ = true;
-      }
-      return label_reachable_->Reach(label);
-    } else {
-      return true;
-    }
-  }
-
- private:
-  void Init(const FST &fst, MatchType match_type,
-            std::shared_ptr<MatcherData> data,
-            Accumulator *accumulator) {
-    if (!(kFlags & (kInputLookAheadMatcher | kOutputLookAheadMatcher))) {
-      FSTERROR() << "LabelLookaheadMatcher: Bad matcher flags: " << kFlags;
-      error_ = true;
-    }
-    const bool reach_input = match_type == MATCH_INPUT;
-    if (data) {
-      if (reach_input == data->ReachInput()) {
-        label_reachable_.reset(new Reachable(data, accumulator));
-      }
-    } else if ((reach_input && (kFlags & kInputLookAheadMatcher)) ||
-               (!reach_input && (kFlags & kOutputLookAheadMatcher))) {
-      label_reachable_.reset(new Reachable(fst, reach_input, accumulator,
-                                           kFlags & kLookAheadKeepRelabelData));
-    }
-  }
-
-  mutable M matcher_;
-  const Fst<Arc> *lfst_;                        // Look-ahead FST.
-  std::unique_ptr<Reachable> label_reachable_;  // Label reachability info.
-  StateId state_;                               // Matcher state.
-  bool match_set_state_;                        // matcher_.SetState called?
-  mutable bool reach_set_state_;                // reachable_.SetState called?
-  bool error_;                                  // Error encountered?
-};
-
-template <class M, uint32 flags, class Accumulator, class Reachable>
-template <class LFST>
-inline bool LabelLookAheadMatcher<M, flags, Accumulator,
-                                  Reachable>::LookAheadFst(const LFST &fst,
-                                                           StateId s) {
-  if (static_cast<const Fst<Arc> *>(&fst) != lfst_) InitLookAheadFst(fst);
-  ClearLookAheadWeight();
-  ClearLookAheadPrefix();
-  if (!label_reachable_) return true;
-  label_reachable_->SetState(state_, s);
-  reach_set_state_ = true;
-  bool compute_weight = kFlags & kLookAheadWeight;
-  bool compute_prefix = kFlags & kLookAheadPrefix;
-  ArcIterator<LFST> aiter(fst, s);
-  aiter.SetFlags(kArcNoCache, kArcNoCache);  // Makes caching optional.
-  const bool reach_arc = label_reachable_->Reach(
-      &aiter, 0, internal::NumArcs(*lfst_, s), compute_weight);
-  const auto lfinal = internal::Final(*lfst_, s);
-  const bool reach_final =
-      lfinal != Weight::Zero() && label_reachable_->ReachFinal();
-  if (reach_arc) {
-    const auto begin = label_reachable_->ReachBegin();
-    const auto end = label_reachable_->ReachEnd();
-    if (compute_prefix && end - begin == 1 && !reach_final) {
-      aiter.Seek(begin);
-      SetLookAheadPrefix(aiter.Value());
-      compute_weight = false;
-    } else if (compute_weight) {
-      SetLookAheadWeight(label_reachable_->ReachWeight());
-    }
-  }
-  if (reach_final && compute_weight) {
-    SetLookAheadWeight(reach_arc ? Plus(LookAheadWeight(), lfinal) : lfinal);
-  }
-  return reach_arc || reach_final;
-}
-
-// Label-lookahead relabeling class.
-template <class Arc, class Data = LabelReachableData<typename Arc::Label>>
-class LabelLookAheadRelabeler {
- public:
-  using Label = typename Arc::Label;
-  using Reachable = LabelReachable<Arc, DefaultAccumulator<Arc>, Data>;
-
-  // Relabels matcher FST (initialization function object).
-  template <typename Impl>
-  explicit LabelLookAheadRelabeler(std::shared_ptr<Impl> *impl);
-
-  // Relabels arbitrary FST. Class LFST should be a label-lookahead FST.
-  template <class LFST>
-  static void Relabel(MutableFst<Arc> *fst, const LFST &mfst,
-                      bool relabel_input) {
-    const auto *data = mfst.GetAddOn();
-    Reachable reachable(data->First() ? data->SharedFirst()
-                                      : data->SharedSecond());
-    reachable.Relabel(fst, relabel_input);
-  }
-
-  // Returns relabeling pairs (cf. relabel.h::Relabel()). Class LFST should be a
-  // label-lookahead FST. If avoid_collisions is true, extra pairs are added to
-  // ensure no collisions when relabeling automata that have labels unseen here.
-  template <class LFST>
-  static void RelabelPairs(const LFST &mfst,
-                           std::vector<std::pair<Label, Label>> *pairs,
-                           bool avoid_collisions = false) {
-    const auto *data = mfst.GetAddOn();
-    Reachable reachable(data->First() ? data->SharedFirst()
-                                      : data->SharedSecond());
-    reachable.RelabelPairs(pairs, avoid_collisions);
-  }
-};
-
-template <class Arc, class Data>
-template <typename Impl>
-inline LabelLookAheadRelabeler<Arc, Data>::LabelLookAheadRelabeler(
-    std::shared_ptr<Impl> *impl) {
-  Fst<Arc> &fst = (*impl)->GetFst();
-  auto data = (*impl)->GetSharedAddOn();
-  const auto name = (*impl)->Type();
-  const bool is_mutable = fst.Properties(kMutable, false);
-  std::unique_ptr<MutableFst<Arc>> mfst;
-  if (is_mutable) {
-    mfst.reset(static_cast<MutableFst<Arc> *>(&fst));
-  } else {
-    mfst.reset(new VectorFst<Arc>(fst));
-  }
-  if (data->First()) {  // reach_input.
-    Reachable reachable(data->SharedFirst());
-    reachable.Relabel(mfst.get(), true);
-    if (!FLAGS_save_relabel_ipairs.empty()) {
-      std::vector<std::pair<Label, Label>> pairs;
-      reachable.RelabelPairs(&pairs, true);
-      WriteLabelPairs(FLAGS_save_relabel_ipairs, pairs);
-    }
-  } else {
-    Reachable reachable(data->SharedSecond());
-    reachable.Relabel(mfst.get(), false);
-    if (!FLAGS_save_relabel_opairs.empty()) {
-      std::vector<std::pair<Label, Label>> pairs;
-      reachable.RelabelPairs(&pairs, true);
-      WriteLabelPairs(FLAGS_save_relabel_opairs, pairs);
-    }
-  }
-  if (!is_mutable) {
-    *impl = std::make_shared<Impl>(*mfst, name);
-    (*impl)->SetAddOn(data);
-  }
-}
-
-// Generic lookahead matcher, templated on the FST definition (a wrapper around
-// a pointer to specific one).
-template <class F>
-class LookAheadMatcher {
- public:
-  using FST = F;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using LBase = LookAheadMatcherBase<Arc>;
-
-  // This makes a copy of the FST.
-  LookAheadMatcher(const FST &fst, MatchType match_type)
-      : owned_fst_(fst.Copy()),
-        base_(owned_fst_->InitMatcher(match_type)),
-        lookahead_(false) {
-    if (!base_) base_.reset(new SortedMatcher<FST>(owned_fst_.get(),
-                                                   match_type));
-  }
-
-  // This doesn't copy the FST.
-  LookAheadMatcher(const FST *fst, MatchType match_type)
-      : base_(fst->InitMatcher(match_type)),
-        lookahead_(false) {
-    if (!base_) base_.reset(new SortedMatcher<FST>(fst, match_type));
-  }
-
-  // This makes a copy of the FST.
-  LookAheadMatcher(const LookAheadMatcher<FST> &matcher, bool safe = false)
-      : base_(matcher.base_->Copy(safe)),
-        lookahead_(matcher.lookahead_) { }
-
-  // Takes ownership of base.
-  explicit LookAheadMatcher(MatcherBase<Arc> *base)
-      : base_(base), lookahead_(false) {}
-
-  LookAheadMatcher<FST> *Copy(bool safe = false) const {
-    return new LookAheadMatcher<FST>(*this, safe);
-  }
-
-  MatchType Type(bool test) const { return base_->Type(test); }
-
-  void SetState(StateId s) { base_->SetState(s); }
-
-  bool Find(Label label) { return base_->Find(label); }
-
-  bool Done() const { return base_->Done(); }
-
-  const Arc &Value() const { return base_->Value(); }
-
-  void Next() { base_->Next(); }
-
-  Weight Final(StateId s) const { return base_->Final(s); }
-
-  ssize_t Priority(StateId s) { return base_->Priority(s); }
-
-  const FST &GetFst() const {
-    return static_cast<const FST &>(base_->GetFst());
-  }
-
-  uint64 Properties(uint64 props) const { return base_->Properties(props); }
-
-  uint32 Flags() const { return base_->Flags(); }
-
-  bool LookAheadLabel(Label label) const {
-    if (LookAheadCheck()) {
-      return static_cast<LBase *>(base_.get())->LookAheadLabel(label);
-    } else {
-      return true;
-    }
-  }
-
-  bool LookAheadFst(const Fst<Arc> &fst, StateId s) {
-    if (LookAheadCheck()) {
-      return static_cast<LBase *>(base_.get())->LookAheadFst(fst, s);
-    } else {
-      return true;
-    }
-  }
-
-  Weight LookAheadWeight() const {
-    if (LookAheadCheck()) {
-      return static_cast<LBase *>(base_.get())->LookAheadWeight();
-    } else {
-      return Weight::One();
-    }
-  }
-
-  bool LookAheadPrefix(Arc *arc) const {
-    if (LookAheadCheck()) {
-      return static_cast<LBase *>(base_.get())->LookAheadPrefix(arc);
-    } else {
-      return false;
-    }
-  }
-
-  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) {
-    if (LookAheadCheck()) {
-      static_cast<LBase *>(base_.get())->InitLookAheadFst(fst, copy);
-    }
-  }
-
- private:
-  bool LookAheadCheck() const {
-    if (!lookahead_) {
-      lookahead_ =
-          base_->Flags() & (kInputLookAheadMatcher | kOutputLookAheadMatcher);
-      if (!lookahead_) {
-        FSTERROR() << "LookAheadMatcher: No look-ahead matcher defined";
-      }
-    }
-    return lookahead_;
-  }
-
-  std::unique_ptr<const FST> owned_fst_;
-  std::unique_ptr<MatcherBase<Arc>> base_;
-  mutable bool lookahead_;
-
-  LookAheadMatcher &operator=(const LookAheadMatcher &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_LOOKAHEAD_MATCHER_H_
diff --git a/projects/llm_framework/include/fst/map.h b/projects/llm_framework/include/fst/map.h
deleted file mode 100644
index 5d663ae8..00000000
--- a/projects/llm_framework/include/fst/map.h
+++ /dev/null
@@ -1,110 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Compatibility file for old-style Map() functions and MapFst class that have
-// been renamed to ArcMap (cf. StateMap).
-
-#ifndef FST_MAP_H_
-#define FST_MAP_H_
-
-
-#include <fst/arc-map.h>
-
-
-namespace fst {
-
-template <class A, class C>
-void Map(MutableFst<A> *fst, C *mapper) {
-  ArcMap(fst, mapper);
-}
-
-template <class A, class C>
-void Map(MutableFst<A> *fst, C mapper) {
-  ArcMap(fst, mapper);
-}
-
-template <class A, class B, class C>
-void Map(const Fst<A> &ifst, MutableFst<B> *ofst, C *mapper) {
-  ArcMap(ifst, ofst, mapper);
-}
-
-template <class A, class B, class C>
-void Map(const Fst<A> &ifst, MutableFst<B> *ofst, C mapper) {
-  ArcMap(ifst, ofst, mapper);
-}
-
-using MapFstOptions = ArcMapFstOptions;
-
-template <class A, class B, class C>
-class MapFst : public ArcMapFst<A, B, C> {
- public:
-  using FromArc = A;
-  using ToArc = B;
-
-  using StateId = typename ToArc::StateId;
-  using Weight = typename ToArc::Weight;
-
-  using State = CacheState<B>;
-
-  MapFst(const Fst<A> &fst, const C &mapper, const MapFstOptions &opts)
-      : ArcMapFst<A, B, C>(fst, mapper, opts) {}
-
-  MapFst(const Fst<A> &fst, C *mapper, const MapFstOptions &opts)
-      : ArcMapFst<A, B, C>(fst, mapper, opts) {}
-
-  MapFst(const Fst<A> &fst, const C &mapper)
-      : ArcMapFst<A, B, C>(fst, mapper) {}
-
-  MapFst(const Fst<A> &fst, C *mapper) : ArcMapFst<A, B, C>(fst, mapper) {}
-
-  // See Fst<>::Copy() for doc.
-  MapFst(const MapFst<A, B, C> &fst, bool safe = false)
-      : ArcMapFst<A, B, C>(fst, safe) {}
-
-  // Get a copy of this MapFst. See Fst<>::Copy() for further doc.
-  MapFst<A, B, C> *Copy(bool safe = false) const override {
-    return new MapFst(*this, safe);
-  }
-};
-
-// Specialization for MapFst.
-template <class A, class B, class C>
-class StateIterator<MapFst<A, B, C>>
-    : public StateIterator<ArcMapFst<A, B, C>> {
- public:
-  explicit StateIterator(const ArcMapFst<A, B, C> &fst)
-      : StateIterator<ArcMapFst<A, B, C>>(fst) {}
-};
-
-// Specialization for MapFst.
-template <class A, class B, class C>
-class ArcIterator<MapFst<A, B, C>> : public ArcIterator<ArcMapFst<A, B, C>> {
- public:
-  ArcIterator(const ArcMapFst<A, B, C> &fst, typename A::StateId s)
-      : ArcIterator<ArcMapFst<A, B, C>>(fst, s) {}
-};
-
-// For backwards compatibility only; use IdentityArcMapper otherwise.
-template <class A>
-struct IdentityMapper {
-  using FromArc = A;
-  using ToArc = A;
-
-  ToArc operator()(const FromArc &arc) const { return arc; }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const { return props; }
-};
-
-}  // namespace fst
-
-#endif  // FST_MAP_H_
diff --git a/projects/llm_framework/include/fst/mapped-file.h b/projects/llm_framework/include/fst/mapped-file.h
deleted file mode 100644
index adb33c28..00000000
--- a/projects/llm_framework/include/fst/mapped-file.h
+++ /dev/null
@@ -1,81 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_MAPPED_FILE_H_
-#define FST_MAPPED_FILE_H_
-
-#include <cstddef>
-#include <istream>
-#include <string>
-
-#include <fst/compat.h>
-#include <fst/flags.h>
-
-namespace fst {
-
-// A memory region is a simple abstraction for allocated memory or data from
-// memory-mapped files. If mmap is null, then data represents an owned region
-// of size bytes. Otherwise, mmap and size refer to the mapping and data is a
-// casted pointer to a region contained within [mmap, mmap + size). If size is
-// 0, then mmap and data refer to a block of memory managed externally by some
-// other allocator. The offset is used when allocating memory to providing
-// padding for alignment.
-struct MemoryRegion {
-  void *data;
-  void *mmap;
-  size_t size;
-  int offset;
-};
-
-class MappedFile {
- public:
-  ~MappedFile();
-
-  void *mutable_data() const { return region_.data; }
-
-  const void *data() const { return region_.data; }
-
-  // Returns a MappedFile object that contains the contents of the input stream
-  // strm starting from the current file position with size bytes. The memorymap
-  // bool is advisory, and Map will default to allocating and reading. The
-  // source argument needs to contain the filename that was used to open the
-  // input stream.
-  static MappedFile *Map(std::istream *istrm, bool memorymap,
-                         const string &source, size_t size);
-
-  // Returns a MappedFile object that contains the contents of the file referred
-  // to by the file descriptor starting from pos with size bytes. If the
-  // memory mapping fails, nullptr is returned. In contrast to Map(), this
-  // factory function does not backoff to allocating and reading.
-  static MappedFile *MapFromFileDescriptor(int fd, int pos, size_t size);
-
-  // Creates a MappedFile object with a new[]'ed block of memory of size. The
-  // align argument can be used to specify a desired block alignment.
-  // This is RECOMMENDED FOR INTERNAL USE ONLY as it may change in future
-  // releases.
-  static MappedFile *Allocate(size_t size, int align = kArchAlignment);
-
-  // Creates a MappedFile object pointing to a borrowed reference to data. This
-  // block of memory is not owned by the MappedFile object and will not be
-  // freed. This is RECOMMENDED FOR INTERNAL USE ONLY, may change in future
-  // releases.
-  static MappedFile *Borrow(void *data);
-
-  // Alignment required for mapping structures in bytes. Regions of memory that
-  // are not aligned upon a 128-bit boundary are read from the file instead.
-  // This is consistent with the alignment boundary set in ConstFst and
-  // CompactFst.
-  static constexpr int kArchAlignment = 16;
-
-  static constexpr size_t kMaxReadChunk = 256 * 1024 * 1024;  // 256 MB.
-
- private:
-  explicit MappedFile(const MemoryRegion &region);
-
-  MemoryRegion region_;
-  MappedFile(const MappedFile &) = delete;
-  MappedFile &operator=(const MappedFile &) = delete;
-};
-}  // namespace fst
-
-#endif  // FST_MAPPED_FILE_H_
diff --git a/projects/llm_framework/include/fst/matcher-fst.h b/projects/llm_framework/include/fst/matcher-fst.h
deleted file mode 100644
index 61e95820..00000000
--- a/projects/llm_framework/include/fst/matcher-fst.h
+++ /dev/null
@@ -1,347 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to add a matcher to an FST.
-
-#ifndef FST_MATCHER_FST_H_
-#define FST_MATCHER_FST_H_
-
-#include <memory>
-#include <string>
-
-#include <fst/add-on.h>
-#include <fst/const-fst.h>
-#include <fst/lookahead-matcher.h>
-
-
-namespace fst {
-
-// Writeable matchers have the same interface as Matchers (as defined in
-// matcher.h) along with the following additional methods:
-//
-// template <class F>
-// class Matcher {
-//  public:
-//   using FST = F;
-//   ...
-//   using MatcherData = ...;   // Initialization data.
-//
-//   // Constructor with additional argument for external initialization data;
-//   // matcher increments its reference count on construction and decrements
-//   // the reference count, and deletes once the reference count has reached
-//   // zero.
-//   Matcher(const FST &fst, MatchType type, MatcherData *data);
-//
-//   // Returns pointer to initialization data that can be passed to a Matcher
-//   // constructor.
-//   MatcherData *GetData() const;
-// };
-
-// The matcher initialization data class must also provide the following
-// interface:
-//
-// class MatcherData {
-// public:
-//   // Required copy constructor.
-//   MatcherData(const MatcherData &);
-//
-//   // Required I/O methods.
-//   static MatcherData *Read(std::istream &istrm, const FstReadOptions &opts);
-//   bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const;
-// };
-
-// Trivial (no-op) MatcherFst initializer functor.
-template <class M>
-class NullMatcherFstInit {
- public:
-  using MatcherData = typename M::MatcherData;
-  using Data = AddOnPair<MatcherData, MatcherData>;
-  using Impl = internal::AddOnImpl<typename M::FST, Data>;
-
-  explicit NullMatcherFstInit(std::shared_ptr<Impl> *) {}
-};
-
-// Class adding a matcher to an FST type. Creates a new FST whose name is given
-// by N. An optional functor Init can be used to initialize the FST. The Data
-// template parameter allows the user to select the type of the add-on.
-template <
-    class F, class M, const char *Name, class Init = NullMatcherFstInit<M>,
-    class Data = AddOnPair<typename M::MatcherData, typename M::MatcherData>>
-class MatcherFst : public ImplToExpandedFst<internal::AddOnImpl<F, Data>> {
- public:
-  using FST = F;
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-
-  using FstMatcher = M;
-  using MatcherData = typename FstMatcher::MatcherData;
-
-  using Impl = internal::AddOnImpl<FST, Data>;
-  using D = Data;
-
-  friend class StateIterator<MatcherFst<FST, FstMatcher, Name, Init, Data>>;
-  friend class ArcIterator<MatcherFst<FST, FstMatcher, Name, Init, Data>>;
-
-  MatcherFst() : ImplToExpandedFst<Impl>(std::make_shared<Impl>(FST(), Name)) {}
-
-  explicit MatcherFst(const FST &fst, std::shared_ptr<Data> data = nullptr)
-      : ImplToExpandedFst<Impl>(data ? CreateImpl(fst, Name, data)
-                                     : CreateDataAndImpl(fst, Name)) {}
-
-  explicit MatcherFst(const Fst<Arc> &fst)
-      : ImplToExpandedFst<Impl>(CreateDataAndImpl(fst, Name)) {}
-
-  // See Fst<>::Copy() for doc.
-  MatcherFst(const MatcherFst<FST, FstMatcher, Name, Init, Data> &fst,
-             bool safe = false)
-      : ImplToExpandedFst<Impl>(fst, safe) {}
-
-  // Get a copy of this MatcherFst. See Fst<>::Copy() for further doc.
-  MatcherFst<FST, FstMatcher, Name, Init, Data> *Copy(
-      bool safe = false) const override {
-    return new MatcherFst<FST, FstMatcher, Name, Init, Data>(*this, safe);
-  }
-
-  // Read a MatcherFst from an input stream; return nullptr on error
-  static MatcherFst<FST, M, Name, Init, Data> *Read(
-      std::istream &strm, const FstReadOptions &opts) {
-    auto *impl = Impl::Read(strm, opts);
-    return impl ? new MatcherFst<FST, FstMatcher, Name, Init, Data>(
-                      std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  // Read a MatcherFst from a file; return nullptr on error
-  // Empty filename reads from standard input
-  static MatcherFst<FST, FstMatcher, Name, Init, Data> *Read(
-      const string &filename) {
-    auto *impl = ImplToExpandedFst<Impl>::Read(filename);
-    return impl ? new MatcherFst<FST, FstMatcher, Name, Init, Data>(
-                      std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return GetImpl()->Write(strm, opts);
-  }
-
-  bool Write(const string &filename) const override {
-    return Fst<Arc>::WriteFile(filename);
-  }
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const override {
-    return GetImpl()->InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    return GetImpl()->InitArcIterator(s, data);
-  }
-
-  FstMatcher *InitMatcher(MatchType match_type) const override {
-    return new FstMatcher(&GetFst(), match_type, GetSharedData(match_type));
-  }
-
-  const FST &GetFst() const { return GetImpl()->GetFst(); }
-
-  const Data *GetAddOn() const { return GetImpl()->GetAddOn(); }
-
-  std::shared_ptr<Data> GetSharedAddOn() const {
-    return GetImpl()->GetSharedAddOn();
-  }
-
-  const MatcherData *GetData(MatchType match_type) const {
-    const auto *data = GetAddOn();
-    return match_type == MATCH_INPUT ? data->First() : data->Second();
-  }
-
-  std::shared_ptr<MatcherData> GetSharedData(MatchType match_type) const {
-    const auto *data = GetAddOn();
-    return match_type == MATCH_INPUT ? data->SharedFirst()
-                                     : data->SharedSecond();
-  }
-
- protected:
-  using ImplToFst<Impl, ExpandedFst<Arc>>::GetImpl;
-
-  static std::shared_ptr<Impl> CreateDataAndImpl(const FST &fst,
-                                                 const string &name) {
-    FstMatcher imatcher(fst, MATCH_INPUT);
-    FstMatcher omatcher(fst, MATCH_OUTPUT);
-    return CreateImpl(fst, name,
-                      std::make_shared<Data>(imatcher.GetSharedData(),
-                                             omatcher.GetSharedData()));
-  }
-
-  static std::shared_ptr<Impl> CreateDataAndImpl(const Fst<Arc> &fst,
-                                                 const string &name) {
-    FST result(fst);
-    return CreateDataAndImpl(result, name);
-  }
-
-  static std::shared_ptr<Impl> CreateImpl(const FST &fst, const string &name,
-                                          std::shared_ptr<Data> data) {
-    auto impl = std::make_shared<Impl>(fst, name);
-    impl->SetAddOn(data);
-    Init init(&impl);
-    return impl;
-  }
-
-  explicit MatcherFst(std::shared_ptr<Impl> impl)
-      : ImplToExpandedFst<Impl>(impl) {}
-
- private:
-  MatcherFst &operator=(const MatcherFst &) = delete;
-};
-
-// Specialization for MatcherFst.
-template <class FST, class M, const char *Name, class Init>
-class StateIterator<MatcherFst<FST, M, Name, Init>>
-    : public StateIterator<FST> {
- public:
-  explicit StateIterator(const MatcherFst<FST, M, Name, Init> &fst)
-      : StateIterator<FST>(fst.GetImpl()->GetFst()) {}
-};
-
-// Specialization for MatcherFst.
-template <class FST, class M, const char *Name, class Init>
-class ArcIterator<MatcherFst<FST, M, Name, Init>> : public ArcIterator<FST> {
- public:
-  using StateId = typename FST::Arc::StateId;
-
-  ArcIterator(const MatcherFst<FST, M, Name, Init> &fst,
-              typename FST::Arc::StateId s)
-      : ArcIterator<FST>(fst.GetImpl()->GetFst(), s) {}
-};
-
-// Specialization for MatcherFst.
-template <class F, class M, const char *Name, class Init>
-class Matcher<MatcherFst<F, M, Name, Init>> {
- public:
-  using FST = MatcherFst<F, M, Name, Init>;
-  using Arc = typename F::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-
-  Matcher(const FST &fst, MatchType match_type)
-      : matcher_(fst.InitMatcher(match_type)) {}
-
-  Matcher(const Matcher<FST> &matcher) : matcher_(matcher.matcher_->Copy()) {}
-
-  Matcher<FST> *Copy() const { return new Matcher<FST>(*this); }
-
-  MatchType Type(bool test) const { return matcher_->Type(test); }
-
-  void SetState(StateId s) { matcher_->SetState(s); }
-
-  bool Find(Label label) { return matcher_->Find(label); }
-
-  bool Done() const { return matcher_->Done(); }
-
-  const Arc &Value() const { return matcher_->Value(); }
-
-  void Next() { matcher_->Next(); }
-
-  uint64 Properties(uint64 props) const { return matcher_->Properties(props); }
-
-  uint32 Flags() const { return matcher_->Flags(); }
-
- private:
-  std::unique_ptr<M> matcher_;
-};
-
-// Specialization for MatcherFst.
-template <class F, class M, const char *Name, class Init>
-class LookAheadMatcher<MatcherFst<F, M, Name, Init>> {
- public:
-  using FST = MatcherFst<F, M, Name, Init>;
-  using Arc = typename F::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  LookAheadMatcher(const FST &fst, MatchType match_type)
-      : matcher_(fst.InitMatcher(match_type)) {}
-
-  LookAheadMatcher(const LookAheadMatcher<FST> &matcher, bool safe = false)
-      : matcher_(matcher.matcher_->Copy(safe)) {}
-
-  // General matcher methods.
-  LookAheadMatcher<FST> *Copy(bool safe = false) const {
-    return new LookAheadMatcher<FST>(*this, safe);
-  }
-
-  MatchType Type(bool test) const { return matcher_->Type(test); }
-
-  void SetState(StateId s) { matcher_->SetState(s); }
-
-  bool Find(Label label) { return matcher_->Find(label); }
-
-  bool Done() const { return matcher_->Done(); }
-
-  const Arc &Value() const { return matcher_->Value(); }
-
-  void Next() { matcher_->Next(); }
-
-  const FST &GetFst() const { return matcher_->GetFst(); }
-
-  uint64 Properties(uint64 props) const { return matcher_->Properties(props); }
-
-  uint32 Flags() const { return matcher_->Flags(); }
-
-  bool LookAheadLabel(Label label) const {
-    return matcher_->LookAheadLabel(label);
-  }
-
-  bool LookAheadFst(const Fst<Arc> &fst, StateId s) {
-    return matcher_->LookAheadFst(fst, s);
-  }
-
-  Weight LookAheadWeight() const { return matcher_->LookAheadWeight(); }
-
-  bool LookAheadPrefix(Arc *arc) const {
-    return matcher_->LookAheadPrefix(arc);
-  }
-
-  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) {
-    matcher_->InitLookAheadFst(fst, copy);
-  }
-
- private:
-  std::unique_ptr<M> matcher_;
-};
-
-// Useful aliases when using StdArc.
-
-extern const char arc_lookahead_fst_type[];
-
-using StdArcLookAheadFst =
-    MatcherFst<ConstFst<StdArc>,
-               ArcLookAheadMatcher<SortedMatcher<ConstFst<StdArc>>>,
-               arc_lookahead_fst_type>;
-
-extern const char ilabel_lookahead_fst_type[];
-extern const char olabel_lookahead_fst_type[];
-
-constexpr auto ilabel_lookahead_flags =
-    kInputLookAheadMatcher | kLookAheadWeight | kLookAheadPrefix |
-    kLookAheadEpsilons | kLookAheadNonEpsilonPrefix;
-
-constexpr auto olabel_lookahead_flags =
-    kOutputLookAheadMatcher | kLookAheadWeight | kLookAheadPrefix |
-    kLookAheadEpsilons | kLookAheadNonEpsilonPrefix;
-
-using StdILabelLookAheadFst = MatcherFst<
-    ConstFst<StdArc>,
-    LabelLookAheadMatcher<SortedMatcher<ConstFst<StdArc>>,
-                          ilabel_lookahead_flags, FastLogAccumulator<StdArc>>,
-    ilabel_lookahead_fst_type, LabelLookAheadRelabeler<StdArc>>;
-
-using StdOLabelLookAheadFst = MatcherFst<
-    ConstFst<StdArc>,
-    LabelLookAheadMatcher<SortedMatcher<ConstFst<StdArc>>,
-                          olabel_lookahead_flags, FastLogAccumulator<StdArc>>,
-    olabel_lookahead_fst_type, LabelLookAheadRelabeler<StdArc>>;
-
-}  // namespace fst
-
-#endif  // FST_MATCHER_FST_H_
diff --git a/projects/llm_framework/include/fst/matcher.h b/projects/llm_framework/include/fst/matcher.h
deleted file mode 100644
index d9528d68..00000000
--- a/projects/llm_framework/include/fst/matcher.h
+++ /dev/null
@@ -1,1575 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes to allow matching labels leaving FST states.
-
-#ifndef FST_MATCHER_H_
-#define FST_MATCHER_H_
-
-#include <algorithm>
-#include <memory>
-#include <unordered_map>
-#include <utility>
-
-#include <fst/log.h>
-
-#include <fst/mutable-fst.h>  // for all internal FST accessors.
-
-
-namespace fst {
-
-// Matchers find and iterate through requested labels at FST states. In the
-// simplest form, these are just some associative map or search keyed on labels.
-// More generally, they may implement matching special labels that represent
-// sets of labels such as sigma (all), rho (rest), or phi (fail). The Matcher
-// interface is:
-//
-// template <class F>
-// class Matcher {
-//  public:
-//   using FST = F;
-//   using Arc = typename FST::Arc;
-//   using Label = typename Arc::Label;
-//   using StateId = typename Arc::StateId;
-//   using Weight = typename Arc::Weight;
-//
-//   // Required constructors. Note:
-//   // -- the constructors that copy the FST arg are useful for
-//   // letting the matcher manage the FST through copies
-//   // (esp with 'safe' copies); e.g. ComposeFst depends on this.
-//   // -- the constructor that does not copy is useful when the
-//   // the FST is mutated during the lifetime of the matcher
-//   // (o.w. the matcher would have its own unmutated deep copy).
-//
-//   // This makes a copy of the FST.
-//   Matcher(const FST &fst, MatchType type);
-//   // This doesn't copy the FST.
-//   Matcher(const FST *fst, MatchType type);
-//   // This makes a copy of the FST.
-//   // See Copy() below.
-//   Matcher(const Matcher &matcher, bool safe = false);
-//
-//   // If safe = true, the copy is thread-safe. See Fst<>::Copy() for
-//   // further doc.
-//   Matcher<FST> *Copy(bool safe = false) const override;
-//
-//   // Returns the match type that can be provided (depending on compatibility
-//   of the input FST). It is either the requested match type, MATCH_NONE, or
-//   MATCH_UNKNOWN. If test is false, a costly testing is avoided, but
-//   MATCH_UNKNOWN may be returned. If test is true, a definite answer is
-//   returned, but may involve more costly computation (e.g., visiting the FST).
-//   MatchType Type(bool test) const override;
-//
-//   // Specifies the current state.
-//   void SetState(StateId s) final;
-//
-//   // Finds matches to a label at the current state, returning true if a match
-//   // found. kNoLabel matches any non-consuming transitions, e.g., epsilon
-//   // transitions, which do not require a matching symbol.
-//   bool Find(Label label) final;
-//
-//   // Iterator methods. Note that initially and after SetState() these have
-//   undefined behavior until Find() is called.
-//
-//   bool Done() const final;
-//
-//   const Arc &Value() const final;
-//
-//   void Next() final;
-//
-//   // Returns final weight of a state.
-//   Weight Final(StateId) const final;
-//
-//   // Indicates preference for being the side used for matching in
-//   // composition. If the value is kRequirePriority, then it is
-//   // mandatory that it be used. Calling this method without passing the
-//   // current state of the matcher invalidates the state of the matcher.
-//   ssize_t Priority(StateId s) final;
-//
-//   // This specifies the known FST properties as viewed from this matcher. It
-//   // takes as argument the input FST's known properties.
-//   uint64 Properties(uint64 props) const override;
-//
-//   // Returns matcher flags.
-//   uint32 Flags() const override;
-//
-//   // Returns matcher FST.
-//   const FST &GetFst() const override;
-// };
-
-// Basic matcher flags.
-
-// Matcher needs to be used as the matching side in composition for
-// at least one state (has kRequirePriority).
-constexpr uint32 kRequireMatch = 0x00000001;
-
-// Flags used for basic matchers (see also lookahead.h).
-constexpr uint32 kMatcherFlags = kRequireMatch;
-
-// Matcher priority that is mandatory.
-constexpr ssize_t kRequirePriority = -1;
-
-// Matcher interface, templated on the Arc definition; used for matcher
-// specializations that are returned by the InitMatcher FST method.
-template <class A>
-class MatcherBase {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  virtual ~MatcherBase() {}
-
-  // Virtual interface.
-
-  virtual MatcherBase<Arc> *Copy(bool safe = false) const = 0;
-  virtual MatchType Type(bool) const = 0;
-  virtual void SetState(StateId) = 0;
-  virtual bool Find(Label) = 0;
-  virtual bool Done() const = 0;
-  virtual const Arc &Value() const = 0;
-  virtual void Next() = 0;
-  virtual const Fst<Arc> &GetFst() const = 0;
-  virtual uint64 Properties(uint64) const = 0;
-
-  // Trivial implementations that can be used by derived classes. Full
-  // devirtualization is expected for any derived class marked final.
-  virtual uint32 Flags() const { return 0; }
-
-  virtual Weight Final(StateId s) const { return internal::Final(GetFst(), s); }
-
-  virtual ssize_t Priority(StateId s) { return internal::NumArcs(GetFst(), s); }
-};
-
-// A matcher that expects sorted labels on the side to be matched.
-// If match_type == MATCH_INPUT, epsilons match the implicit self-loop
-// Arc(kNoLabel, 0, Weight::One(), current_state) as well as any
-// actual epsilon transitions. If match_type == MATCH_OUTPUT, then
-// Arc(0, kNoLabel, Weight::One(), current_state) is instead matched.
-template <class F>
-class SortedMatcher : public MatcherBase<typename F::Arc> {
- public:
-  using FST = F;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using MatcherBase<Arc>::Flags;
-  using MatcherBase<Arc>::Properties;
-
-  // Labels >= binary_label will be searched for by binary search;
-  // o.w. linear search is used.
-  // This makes a copy of the FST.
-  SortedMatcher(const FST &fst, MatchType match_type, Label binary_label = 1)
-      : SortedMatcher(fst.Copy(), match_type, binary_label) {
-    owned_fst_.reset(&fst_);
-  }
-
-  // Labels >= binary_label will be searched for by binary search;
-  // o.w. linear search is used.
-  // This doesn't copy the FST.
-  SortedMatcher(const FST *fst, MatchType match_type, Label binary_label = 1)
-      : fst_(*fst),
-        state_(kNoStateId),
-        aiter_(nullptr),
-        match_type_(match_type),
-        binary_label_(binary_label),
-        match_label_(kNoLabel),
-        narcs_(0),
-        loop_(kNoLabel, 0, Weight::One(), kNoStateId),
-        error_(false),
-        aiter_pool_(1) {
-    switch (match_type_) {
-      case MATCH_INPUT:
-      case MATCH_NONE:
-        break;
-      case MATCH_OUTPUT:
-        std::swap(loop_.ilabel, loop_.olabel);
-        break;
-      default:
-        FSTERROR() << "SortedMatcher: Bad match type";
-        match_type_ = MATCH_NONE;
-        error_ = true;
-    }
-  }
-
-  // This makes a copy of the FST.
-  SortedMatcher(const SortedMatcher<FST> &matcher, bool safe = false)
-      : owned_fst_(matcher.fst_.Copy(safe)),
-        fst_(*owned_fst_),
-        state_(kNoStateId),
-        aiter_(nullptr),
-        match_type_(matcher.match_type_),
-        binary_label_(matcher.binary_label_),
-        match_label_(kNoLabel),
-        narcs_(0),
-        loop_(matcher.loop_),
-        error_(matcher.error_),
-        aiter_pool_(1) {}
-
-  ~SortedMatcher() override { Destroy(aiter_, &aiter_pool_); }
-
-  SortedMatcher<FST> *Copy(bool safe = false) const override {
-    return new SortedMatcher<FST>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override {
-    if (match_type_ == MATCH_NONE) return match_type_;
-    const auto true_prop =
-        match_type_ == MATCH_INPUT ? kILabelSorted : kOLabelSorted;
-    const auto false_prop =
-        match_type_ == MATCH_INPUT ? kNotILabelSorted : kNotOLabelSorted;
-    const auto props = fst_.Properties(true_prop | false_prop, test);
-    if (props & true_prop) {
-      return match_type_;
-    } else if (props & false_prop) {
-      return MATCH_NONE;
-    } else {
-      return MATCH_UNKNOWN;
-    }
-  }
-
-  void SetState(StateId s) final {
-    if (state_ == s) return;
-    state_ = s;
-    if (match_type_ == MATCH_NONE) {
-      FSTERROR() << "SortedMatcher: Bad match type";
-      error_ = true;
-    }
-    Destroy(aiter_, &aiter_pool_);
-    aiter_ = new (&aiter_pool_) ArcIterator<FST>(fst_, s);
-    aiter_->SetFlags(kArcNoCache, kArcNoCache);
-    narcs_ = internal::NumArcs(fst_, s);
-    loop_.nextstate = s;
-  }
-
-  bool Find(Label match_label) final {
-    exact_match_ = true;
-    if (error_) {
-      current_loop_ = false;
-      match_label_ = kNoLabel;
-      return false;
-    }
-    current_loop_ = match_label == 0;
-    match_label_ = match_label == kNoLabel ? 0 : match_label;
-    if (Search()) {
-      return true;
-    } else {
-      return current_loop_;
-    }
-  }
-
-  // Positions matcher to the first position where inserting match_label would
-  // maintain the sort order.
-  void LowerBound(Label label) {
-    exact_match_ = false;
-    current_loop_ = false;
-    if (error_) {
-      match_label_ = kNoLabel;
-      return;
-    }
-    match_label_ = label;
-    Search();
-  }
-
-  // After Find(), returns false if no more exact matches.
-  // After LowerBound(), returns false if no more arcs.
-  bool Done() const final {
-    if (current_loop_) return false;
-    if (aiter_->Done()) return true;
-    if (!exact_match_) return false;
-    aiter_->SetFlags(match_type_ == MATCH_INPUT ?
-        kArcILabelValue : kArcOLabelValue,
-        kArcValueFlags);
-    return GetLabel() != match_label_;
-  }
-
-  const Arc &Value() const final {
-    if (current_loop_) return loop_;
-    aiter_->SetFlags(kArcValueFlags, kArcValueFlags);
-    return aiter_->Value();
-  }
-
-  void Next() final {
-    if (current_loop_) {
-      current_loop_ = false;
-    } else {
-      aiter_->Next();
-    }
-  }
-
-  Weight Final(StateId s) const final {
-    return MatcherBase<Arc>::Final(s);
-  }
-
-  ssize_t Priority(StateId s) final {
-    return MatcherBase<Arc>::Priority(s);
-  }
-
-  const FST &GetFst() const override { return fst_; }
-
-  uint64 Properties(uint64 inprops) const override {
-    return inprops | (error_ ? kError : 0);
-  }
-
-  size_t Position() const { return aiter_ ? aiter_->Position() : 0; }
-
- private:
-  Label GetLabel() const {
-    const auto &arc = aiter_->Value();
-    return match_type_ == MATCH_INPUT ? arc.ilabel : arc.olabel;
-  }
-
-  bool BinarySearch();
-  bool LinearSearch();
-  bool Search();
-
-  std::unique_ptr<const FST> owned_fst_;   // FST ptr if owned.
-  const FST &fst_;           // FST for matching.
-  StateId state_;            // Matcher state.
-  ArcIterator<FST> *aiter_;  // Iterator for current state.
-  MatchType match_type_;     // Type of match to perform.
-  Label binary_label_;       // Least label for binary search.
-  Label match_label_;        // Current label to be matched.
-  size_t narcs_;             // Current state arc count.
-  Arc loop_;                 // For non-consuming symbols.
-  bool current_loop_;        // Current arc is the implicit loop.
-  bool exact_match_;         // Exact match or lower bound?
-  bool error_;               // Error encountered?
-  MemoryPool<ArcIterator<FST>> aiter_pool_;  // Pool of arc iterators.
-};
-
-// Returns true iff match to match_label_. The arc iterator is positioned at the
-// lower bound, that is, the first element greater than or equal to
-// match_label_, or the end if all elements are less than match_label_.
-// If multiple elements are equal to the `match_label_`, returns the rightmost
-// one.
-template <class FST>
-inline bool SortedMatcher<FST>::BinarySearch() {
-  size_t size = narcs_;
-  if (size == 0) {
-    return false;
-  }
-  size_t high = size - 1;
-  while (size > 1) {
-    const size_t half = size / 2;
-    const size_t mid = high - half;
-    aiter_->Seek(mid);
-    if (GetLabel() >= match_label_) {
-      high = mid;
-    }
-    size -= half;
-  }
-  aiter_->Seek(high);
-  const auto label = GetLabel();
-  if (label == match_label_) {
-    return true;
-  }
-  if (label < match_label_) {
-    aiter_->Next();
-  }
-  return false;
-}
-
-// Returns true iff match to match_label_, positioning arc iterator at lower
-// bound.
-template <class FST>
-inline bool SortedMatcher<FST>::LinearSearch() {
-  for (aiter_->Reset(); !aiter_->Done(); aiter_->Next()) {
-    const auto label = GetLabel();
-    if (label == match_label_) return true;
-    if (label > match_label_) break;
-  }
-  return false;
-}
-
-// Returns true iff match to match_label_, positioning arc iterator at lower
-// bound.
-template <class FST>
-inline bool SortedMatcher<FST>::Search() {
-  aiter_->SetFlags(match_type_ == MATCH_INPUT ?
-                   kArcILabelValue : kArcOLabelValue,
-                   kArcValueFlags);
-  if (match_label_ >= binary_label_) {
-    return BinarySearch();
-  } else {
-    return LinearSearch();
-  }
-}
-
-// A matcher that stores labels in a per-state hash table populated upon the
-// first visit to that state. Sorting is not required. Treatment of
-// epsilons are the same as with SortedMatcher.
-template <class F>
-class HashMatcher : public MatcherBase<typename F::Arc> {
- public:
-  using FST = F;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using MatcherBase<Arc>::Flags;
-  using MatcherBase<Arc>::Final;
-  using MatcherBase<Arc>::Priority;
-
-  // This makes a copy of the FST.
-  HashMatcher(const FST &fst, MatchType match_type)
-      : HashMatcher(fst.Copy(), match_type) {
-    owned_fst_.reset(&fst_);
-  }
-
-  // This doesn't copy the FST.
-  HashMatcher(const FST *fst, MatchType match_type)
-      : fst_(*fst),
-        state_(kNoStateId),
-        match_type_(match_type),
-        loop_(kNoLabel, 0, Weight::One(), kNoStateId),
-        error_(false),
-        state_table_(std::make_shared<StateTable>()) {
-    switch (match_type_) {
-      case MATCH_INPUT:
-      case MATCH_NONE:
-        break;
-      case MATCH_OUTPUT:
-        std::swap(loop_.ilabel, loop_.olabel);
-        break;
-      default:
-        FSTERROR() << "HashMatcher: Bad match type";
-        match_type_ = MATCH_NONE;
-        error_ = true;
-    }
-  }
-
-  // This makes a copy of the FST.
-  HashMatcher(const HashMatcher<FST> &matcher, bool safe = false)
-      : owned_fst_(matcher.fst_.Copy(safe)),
-        fst_(*owned_fst_),
-        state_(kNoStateId),
-        match_type_(matcher.match_type_),
-        loop_(matcher.loop_),
-        error_(matcher.error_),
-        state_table_(
-            safe ? std::make_shared<StateTable>() : matcher.state_table_) {}
-
-  HashMatcher<FST> *Copy(bool safe = false) const override {
-    return new HashMatcher<FST>(*this, safe);
-  }
-
-  // The argument is ignored as there are no relevant properties to test.
-  MatchType Type(bool test) const override { return match_type_; }
-
-  void SetState(StateId s) final;
-
-  bool Find(Label label) final {
-    current_loop_ = label == 0;
-    if (label == 0) {
-      Search(label);
-      return true;
-    }
-    if (label == kNoLabel) label = 0;
-    return Search(label);
-  }
-
-  bool Done() const final {
-    if (current_loop_) return false;
-    return label_it_ == label_end_;
-  }
-
-  const Arc &Value() const final {
-    if (current_loop_) return loop_;
-    aiter_->Seek(label_it_->second);
-    return aiter_->Value();
-  }
-
-  void Next() final {
-    if (current_loop_) {
-      current_loop_ = false;
-    } else {
-      ++label_it_;
-    }
-  }
-
-  const FST &GetFst() const override { return fst_; }
-
-  uint64 Properties(uint64 inprops) const override {
-    return inprops | (error_ ? kError : 0);
-  }
-
- private:
-  Label GetLabel() const {
-    const auto &arc = aiter_->Value();
-    return match_type_ == MATCH_INPUT ? arc.ilabel : arc.olabel;
-  }
-
-  bool Search(Label match_label);
-
-  using LabelTable = std::unordered_multimap<Label, size_t>;
-  using StateTable = std::unordered_map<StateId, std::unique_ptr<LabelTable>>;
-
-  std::unique_ptr<const FST> owned_fst_;  // ptr to FST if owned.
-  const FST &fst_;     // FST for matching.
-  StateId state_;      // Matcher state.
-  MatchType match_type_;
-  Arc loop_;            // The implicit loop itself.
-  bool current_loop_;   // Is the current arc the implicit loop?
-  bool error_;          // Error encountered?
-  std::unique_ptr<ArcIterator<FST>> aiter_;
-  std::shared_ptr<StateTable> state_table_;  // Table from state to label table.
-  LabelTable *label_table_;  // Pointer to current state's label table.
-  typename LabelTable::iterator label_it_;   // Position for label.
-  typename LabelTable::iterator label_end_;  // Position for last label + 1.
-};
-
-template <class FST>
-void HashMatcher<FST>::SetState(typename FST::Arc::StateId s) {
-  if (state_ == s) return;
-  // Resets everything for the state.
-  state_ = s;
-  loop_.nextstate = state_;
-  aiter_.reset(new ArcIterator<FST>(fst_, state_));
-  if (match_type_ == MATCH_NONE) {
-    FSTERROR() << "HashMatcher: Bad match type";
-    error_ = true;
-  }
-  // Attempts to insert a new label table.
-  auto it_and_success = state_table_->emplace(
-    state_, std::unique_ptr<LabelTable>(new LabelTable()));
-  // Sets instance's pointer to the label table for this state.
-  label_table_ = it_and_success.first->second.get();
-  // If it already exists, no additional work is done and we simply return.
-  if (!it_and_success.second) return;
-  // Otherwise, populate this new table.
-  // Populates the label table.
-  label_table_->reserve(internal::NumArcs(fst_, state_));
-  const auto aiter_flags =
-      (match_type_ == MATCH_INPUT ? kArcILabelValue : kArcOLabelValue) |
-      kArcNoCache;
-  aiter_->SetFlags(aiter_flags, kArcFlags);
-  for (; !aiter_->Done(); aiter_->Next()) {
-    label_table_->emplace(GetLabel(), aiter_->Position());
-  }
-  aiter_->SetFlags(kArcValueFlags, kArcValueFlags);
-}
-
-template <class FST>
-inline bool HashMatcher<FST>::Search(typename FST::Arc::Label match_label) {
-  auto range = label_table_->equal_range(match_label);
-  label_it_ = range.first;
-  label_end_ = range.second;
-  if (label_it_ == label_end_) return false;
-  aiter_->Seek(label_it_->second);
-  return true;
-}
-
-// Specifies whether we rewrite both the input and output sides during matching.
-enum MatcherRewriteMode {
-  MATCHER_REWRITE_AUTO = 0,  // Rewrites both sides iff acceptor.
-  MATCHER_REWRITE_ALWAYS,
-  MATCHER_REWRITE_NEVER
-};
-
-// For any requested label that doesn't match at a state, this matcher
-// considers the *unique* transition that matches the label 'phi_label'
-// (phi = 'fail'), and recursively looks for a match at its
-// destination.  When 'phi_loop' is true, if no match is found but a
-// phi self-loop is found, then the phi transition found is returned
-// with the phi_label rewritten as the requested label (both sides if
-// an acceptor, or if 'rewrite_both' is true and both input and output
-// labels of the found transition are 'phi_label').  If 'phi_label' is
-// kNoLabel, this special matching is not done.  PhiMatcher is
-// templated itself on a matcher, which is used to perform the
-// underlying matching.  By default, the underlying matcher is
-// constructed by PhiMatcher. The user can instead pass in this
-// object; in that case, PhiMatcher takes its ownership.
-// Phi non-determinism not supported. No non-consuming symbols other
-// than epsilon supported with the underlying template argument matcher.
-template <class M>
-class PhiMatcher : public MatcherBase<typename M::Arc> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // This makes a copy of the FST (w/o 'matcher' arg).
-  PhiMatcher(const FST &fst, MatchType match_type, Label phi_label = kNoLabel,
-             bool phi_loop = true,
-             MatcherRewriteMode rewrite_mode = MATCHER_REWRITE_AUTO,
-             M *matcher = nullptr)
-      : matcher_(matcher ? matcher : new M(fst, match_type)),
-        match_type_(match_type),
-        phi_label_(phi_label),
-        state_(kNoStateId),
-        phi_loop_(phi_loop),
-        error_(false) {
-    if (match_type == MATCH_BOTH) {
-      FSTERROR() << "PhiMatcher: Bad match type";
-      match_type_ = MATCH_NONE;
-      error_ = true;
-    }
-    if (rewrite_mode == MATCHER_REWRITE_AUTO) {
-      rewrite_both_ = fst.Properties(kAcceptor, true);
-    } else if (rewrite_mode == MATCHER_REWRITE_ALWAYS) {
-      rewrite_both_ = true;
-    } else {
-      rewrite_both_ = false;
-    }
-  }
-
-  // This doesn't copy the FST.
-  PhiMatcher(const FST *fst, MatchType match_type, Label phi_label = kNoLabel,
-             bool phi_loop = true,
-             MatcherRewriteMode rewrite_mode = MATCHER_REWRITE_AUTO,
-             M *matcher = nullptr)
-      : PhiMatcher(*fst, match_type, phi_label, phi_loop, rewrite_mode,
-                   matcher ? matcher : new M(fst, match_type)) { }
-
-
-  // This makes a copy of the FST.
-  PhiMatcher(const PhiMatcher<M> &matcher, bool safe = false)
-      : matcher_(new M(*matcher.matcher_, safe)),
-        match_type_(matcher.match_type_),
-        phi_label_(matcher.phi_label_),
-        rewrite_both_(matcher.rewrite_both_),
-        state_(kNoStateId),
-        phi_loop_(matcher.phi_loop_),
-        error_(matcher.error_) {}
-
-  PhiMatcher<M> *Copy(bool safe = false) const override {
-    return new PhiMatcher<M>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override { return matcher_->Type(test); }
-
-  void SetState(StateId s) final {
-    if (state_ == s) return;
-    matcher_->SetState(s);
-    state_ = s;
-    has_phi_ = phi_label_ != kNoLabel;
-  }
-
-  bool Find(Label match_label) final;
-
-  bool Done() const final { return matcher_->Done(); }
-
-  const Arc &Value() const final {
-    if ((phi_match_ == kNoLabel) && (phi_weight_ == Weight::One())) {
-      return matcher_->Value();
-    } else if (phi_match_ == 0) {  // Virtual epsilon loop.
-      phi_arc_ = Arc(kNoLabel, 0, Weight::One(), state_);
-      if (match_type_ == MATCH_OUTPUT) {
-        std::swap(phi_arc_.ilabel, phi_arc_.olabel);
-      }
-      return phi_arc_;
-    } else {
-      phi_arc_ = matcher_->Value();
-      phi_arc_.weight = Times(phi_weight_, phi_arc_.weight);
-      if (phi_match_ != kNoLabel) {  // Phi loop match.
-        if (rewrite_both_) {
-          if (phi_arc_.ilabel == phi_label_) phi_arc_.ilabel = phi_match_;
-          if (phi_arc_.olabel == phi_label_) phi_arc_.olabel = phi_match_;
-        } else if (match_type_ == MATCH_INPUT) {
-          phi_arc_.ilabel = phi_match_;
-        } else {
-          phi_arc_.olabel = phi_match_;
-        }
-      }
-      return phi_arc_;
-    }
-  }
-
-  void Next() final { matcher_->Next(); }
-
-  Weight Final(StateId s) const final {
-    auto weight = matcher_->Final(s);
-    if (phi_label_ == kNoLabel || weight != Weight::Zero()) {
-      return weight;
-    }
-    weight = Weight::One();
-    matcher_->SetState(s);
-    while (matcher_->Final(s) == Weight::Zero()) {
-      if (!matcher_->Find(phi_label_ == 0 ? -1 : phi_label_)) break;
-      weight = Times(weight, matcher_->Value().weight);
-      if (s == matcher_->Value().nextstate) {
-        return Weight::Zero();  // Does not follow phi self-loops.
-      }
-      s = matcher_->Value().nextstate;
-      matcher_->SetState(s);
-    }
-    weight = Times(weight, matcher_->Final(s));
-    return weight;
-  }
-
-  ssize_t Priority(StateId s) final {
-    if (phi_label_ != kNoLabel) {
-      matcher_->SetState(s);
-      const bool has_phi = matcher_->Find(phi_label_ == 0 ? -1 : phi_label_);
-      return has_phi ? kRequirePriority : matcher_->Priority(s);
-    } else {
-      return matcher_->Priority(s);
-    }
-  }
-
-  const FST &GetFst() const override { return matcher_->GetFst(); }
-
-  uint64 Properties(uint64 props) const override;
-
-  uint32 Flags() const override {
-    if (phi_label_ == kNoLabel || match_type_ == MATCH_NONE) {
-      return matcher_->Flags();
-    }
-    return matcher_->Flags() | kRequireMatch;
-  }
-
-  Label PhiLabel() const { return phi_label_; }
-
- private:
-  mutable std::unique_ptr<M> matcher_;
-  MatchType match_type_;  // Type of match requested.
-  Label phi_label_;       // Label that represents the phi transition.
-  bool rewrite_both_;     // Rewrite both sides when both are phi_label_?
-  bool has_phi_;          // Are there possibly phis at the current state?
-  Label phi_match_;       // Current label that matches phi loop.
-  mutable Arc phi_arc_;   // Arc to return.
-  StateId state_;         // Matcher state.
-  Weight phi_weight_;     // Product of the weights of phi transitions taken.
-  bool phi_loop_;         // When true, phi self-loop are allowed and treated
-                          // as rho (required for Aho-Corasick).
-  bool error_;            // Error encountered?
-
-  PhiMatcher &operator=(const PhiMatcher &) = delete;
-};
-
-template <class M>
-inline bool PhiMatcher<M>::Find(Label label) {
-  if (label == phi_label_ && phi_label_ != kNoLabel && phi_label_ != 0) {
-    FSTERROR() << "PhiMatcher::Find: bad label (phi): " << phi_label_;
-    error_ = true;
-    return false;
-  }
-  matcher_->SetState(state_);
-  phi_match_ = kNoLabel;
-  phi_weight_ = Weight::One();
-  // If phi_label_ == 0, there are no more true epsilon arcs.
-  if (phi_label_ == 0) {
-    if (label == kNoLabel) {
-      return false;
-    }
-    if (label == 0) {  // but a virtual epsilon loop needs to be returned.
-      if (!matcher_->Find(kNoLabel)) {
-        return matcher_->Find(0);
-      } else {
-        phi_match_ = 0;
-        return true;
-      }
-    }
-  }
-  if (!has_phi_ || label == 0 || label == kNoLabel) {
-    return matcher_->Find(label);
-  }
-  auto s = state_;
-  while (!matcher_->Find(label)) {
-    // Look for phi transition (if phi_label_ == 0, we need to look
-    // for -1 to avoid getting the virtual self-loop)
-    if (!matcher_->Find(phi_label_ == 0 ? -1 : phi_label_)) return false;
-    if (phi_loop_ && matcher_->Value().nextstate == s) {
-      phi_match_ = label;
-      return true;
-    }
-    phi_weight_ = Times(phi_weight_, matcher_->Value().weight);
-    s = matcher_->Value().nextstate;
-    matcher_->Next();
-    if (!matcher_->Done()) {
-      FSTERROR() << "PhiMatcher: Phi non-determinism not supported";
-      error_ = true;
-    }
-    matcher_->SetState(s);
-  }
-  return true;
-}
-
-template <class M>
-inline uint64 PhiMatcher<M>::Properties(uint64 inprops) const {
-  auto outprops = matcher_->Properties(inprops);
-  if (error_) outprops |= kError;
-  if (match_type_ == MATCH_NONE) {
-    return outprops;
-  } else if (match_type_ == MATCH_INPUT) {
-    if (phi_label_ == 0) {
-      outprops &= ~kEpsilons | ~kIEpsilons | ~kOEpsilons;
-      outprops |= kNoEpsilons | kNoIEpsilons;
-    }
-    if (rewrite_both_) {
-      return outprops &
-             ~(kODeterministic | kNonODeterministic | kString | kILabelSorted |
-               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
-    } else {
-      return outprops &
-             ~(kODeterministic | kAcceptor | kString | kILabelSorted |
-               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
-    }
-  } else if (match_type_ == MATCH_OUTPUT) {
-    if (phi_label_ == 0) {
-      outprops &= ~kEpsilons | ~kIEpsilons | ~kOEpsilons;
-      outprops |= kNoEpsilons | kNoOEpsilons;
-    }
-    if (rewrite_both_) {
-      return outprops &
-             ~(kIDeterministic | kNonIDeterministic | kString | kILabelSorted |
-               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
-    } else {
-      return outprops &
-             ~(kIDeterministic | kAcceptor | kString | kILabelSorted |
-               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
-    }
-  } else {
-    // Shouldn't ever get here.
-    FSTERROR() << "PhiMatcher: Bad match type: " << match_type_;
-    return 0;
-  }
-}
-
-// For any requested label that doesn't match at a state, this matcher
-// considers all transitions that match the label 'rho_label' (rho =
-// 'rest').  Each such rho transition found is returned with the
-// rho_label rewritten as the requested label (both sides if an
-// acceptor, or if 'rewrite_both' is true and both input and output
-// labels of the found transition are 'rho_label').  If 'rho_label' is
-// kNoLabel, this special matching is not done.  RhoMatcher is
-// templated itself on a matcher, which is used to perform the
-// underlying matching.  By default, the underlying matcher is
-// constructed by RhoMatcher.  The user can instead pass in this
-// object; in that case, RhoMatcher takes its ownership.
-// No non-consuming symbols other than epsilon supported with
-// the underlying template argument matcher.
-template <class M>
-class RhoMatcher : public MatcherBase<typename M::Arc> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // This makes a copy of the FST (w/o 'matcher' arg).
-  RhoMatcher(const FST &fst, MatchType match_type, Label rho_label = kNoLabel,
-             MatcherRewriteMode rewrite_mode = MATCHER_REWRITE_AUTO,
-             M *matcher = nullptr)
-      : matcher_(matcher ? matcher : new M(fst, match_type)),
-        match_type_(match_type),
-        rho_label_(rho_label),
-        error_(false),
-        state_(kNoStateId),
-        has_rho_(false) {
-    if (match_type == MATCH_BOTH) {
-      FSTERROR() << "RhoMatcher: Bad match type";
-      match_type_ = MATCH_NONE;
-      error_ = true;
-    }
-    if (rho_label == 0) {
-      FSTERROR() << "RhoMatcher: 0 cannot be used as rho_label";
-      rho_label_ = kNoLabel;
-      error_ = true;
-    }
-    if (rewrite_mode == MATCHER_REWRITE_AUTO) {
-      rewrite_both_ = fst.Properties(kAcceptor, true);
-    } else if (rewrite_mode == MATCHER_REWRITE_ALWAYS) {
-      rewrite_both_ = true;
-    } else {
-      rewrite_both_ = false;
-    }
-  }
-
-  // This doesn't copy the FST.
-  RhoMatcher(const FST *fst, MatchType match_type, Label rho_label = kNoLabel,
-             MatcherRewriteMode rewrite_mode = MATCHER_REWRITE_AUTO,
-             M *matcher = nullptr)
-      : RhoMatcher(*fst, match_type, rho_label, rewrite_mode,
-                   matcher ? matcher : new M(fst, match_type)) { }
-
-  // This makes a copy of the FST.
-  RhoMatcher(const RhoMatcher<M> &matcher, bool safe = false)
-      : matcher_(new M(*matcher.matcher_, safe)),
-        match_type_(matcher.match_type_),
-        rho_label_(matcher.rho_label_),
-        rewrite_both_(matcher.rewrite_both_),
-        error_(matcher.error_),
-        state_(kNoStateId),
-        has_rho_(false) {}
-
-  RhoMatcher<M> *Copy(bool safe = false) const override {
-    return new RhoMatcher<M>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override { return matcher_->Type(test); }
-
-  void SetState(StateId s) final {
-    if (state_ == s) return;
-    state_ = s;
-    matcher_->SetState(s);
-    has_rho_ = rho_label_ != kNoLabel;
-  }
-
-  bool Find(Label label) final {
-    if (label == rho_label_ && rho_label_ != kNoLabel) {
-      FSTERROR() << "RhoMatcher::Find: bad label (rho)";
-      error_ = true;
-      return false;
-    }
-    if (matcher_->Find(label)) {
-      rho_match_ = kNoLabel;
-      return true;
-    } else if (has_rho_ && label != 0 && label != kNoLabel &&
-               (has_rho_ = matcher_->Find(rho_label_))) {
-      rho_match_ = label;
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  bool Done() const final { return matcher_->Done(); }
-
-  const Arc &Value() const final {
-    if (rho_match_ == kNoLabel) {
-      return matcher_->Value();
-    } else {
-      rho_arc_ = matcher_->Value();
-      if (rewrite_both_) {
-        if (rho_arc_.ilabel == rho_label_) rho_arc_.ilabel = rho_match_;
-        if (rho_arc_.olabel == rho_label_) rho_arc_.olabel = rho_match_;
-      } else if (match_type_ == MATCH_INPUT) {
-        rho_arc_.ilabel = rho_match_;
-      } else {
-        rho_arc_.olabel = rho_match_;
-      }
-      return rho_arc_;
-    }
-  }
-
-  void Next() final { matcher_->Next(); }
-
-  Weight Final(StateId s) const final { return matcher_->Final(s); }
-
-  ssize_t Priority(StateId s) final {
-    state_ = s;
-    matcher_->SetState(s);
-    has_rho_ = matcher_->Find(rho_label_);
-    if (has_rho_) {
-      return kRequirePriority;
-    } else {
-      return matcher_->Priority(s);
-    }
-  }
-
-  const FST &GetFst() const override { return matcher_->GetFst(); }
-
-  uint64 Properties(uint64 props) const override;
-
-  uint32 Flags() const override {
-    if (rho_label_ == kNoLabel || match_type_ == MATCH_NONE) {
-      return matcher_->Flags();
-    }
-    return matcher_->Flags() | kRequireMatch;
-  }
-
-  Label RhoLabel() const { return rho_label_; }
-
- private:
-  std::unique_ptr<M> matcher_;
-  MatchType match_type_;  // Type of match requested.
-  Label rho_label_;       // Label that represents the rho transition
-  bool rewrite_both_;     // Rewrite both sides when both are rho_label_?
-  Label rho_match_;       // Current label that matches rho transition.
-  mutable Arc rho_arc_;   // Arc to return when rho match.
-  bool error_;            // Error encountered?
-  StateId state_;         // Matcher state.
-  bool has_rho_;          // Are there possibly rhos at the current state?
-};
-
-template <class M>
-inline uint64 RhoMatcher<M>::Properties(uint64 inprops) const {
-  auto outprops = matcher_->Properties(inprops);
-  if (error_) outprops |= kError;
-  if (match_type_ == MATCH_NONE) {
-    return outprops;
-  } else if (match_type_ == MATCH_INPUT) {
-    if (rewrite_both_) {
-      return outprops &
-             ~(kODeterministic | kNonODeterministic | kString | kILabelSorted |
-               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
-    } else {
-      return outprops &
-             ~(kODeterministic | kAcceptor | kString | kILabelSorted |
-               kNotILabelSorted);
-    }
-  } else if (match_type_ == MATCH_OUTPUT) {
-    if (rewrite_both_) {
-      return outprops &
-             ~(kIDeterministic | kNonIDeterministic | kString | kILabelSorted |
-               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
-    } else {
-      return outprops &
-             ~(kIDeterministic | kAcceptor | kString | kOLabelSorted |
-               kNotOLabelSorted);
-    }
-  } else {
-    // Shouldn't ever get here.
-    FSTERROR() << "RhoMatcher: Bad match type: " << match_type_;
-    return 0;
-  }
-}
-
-// For any requested label, this matcher considers all transitions
-// that match the label 'sigma_label' (sigma = "any"), and this in
-// additions to transitions with the requested label.  Each such sigma
-// transition found is returned with the sigma_label rewritten as the
-// requested label (both sides if an acceptor, or if 'rewrite_both' is
-// true and both input and output labels of the found transition are
-// 'sigma_label').  If 'sigma_label' is kNoLabel, this special
-// matching is not done.  SigmaMatcher is templated itself on a
-// matcher, which is used to perform the underlying matching.  By
-// default, the underlying matcher is constructed by SigmaMatcher.
-// The user can instead pass in this object; in that case,
-// SigmaMatcher takes its ownership.  No non-consuming symbols other
-// than epsilon supported with the underlying template argument matcher.
-template <class M>
-class SigmaMatcher : public MatcherBase<typename M::Arc> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // This makes a copy of the FST (w/o 'matcher' arg).
-  SigmaMatcher(const FST &fst, MatchType match_type,
-               Label sigma_label = kNoLabel,
-               MatcherRewriteMode rewrite_mode = MATCHER_REWRITE_AUTO,
-               M *matcher = nullptr)
-      : matcher_(matcher ? matcher : new M(fst, match_type)),
-        match_type_(match_type),
-        sigma_label_(sigma_label),
-        error_(false),
-        state_(kNoStateId) {
-    if (match_type == MATCH_BOTH) {
-      FSTERROR() << "SigmaMatcher: Bad match type";
-      match_type_ = MATCH_NONE;
-      error_ = true;
-    }
-    if (sigma_label == 0) {
-      FSTERROR() << "SigmaMatcher: 0 cannot be used as sigma_label";
-      sigma_label_ = kNoLabel;
-      error_ = true;
-    }
-    if (rewrite_mode == MATCHER_REWRITE_AUTO) {
-      rewrite_both_ = fst.Properties(kAcceptor, true);
-    } else if (rewrite_mode == MATCHER_REWRITE_ALWAYS) {
-      rewrite_both_ = true;
-    } else {
-      rewrite_both_ = false;
-    }
-  }
-
-  // This doesn't copy the FST.
-  SigmaMatcher(const FST *fst, MatchType match_type,
-               Label sigma_label = kNoLabel,
-               MatcherRewriteMode rewrite_mode = MATCHER_REWRITE_AUTO,
-             M *matcher = nullptr)
-      : SigmaMatcher(*fst, match_type, sigma_label, rewrite_mode,
-                     matcher ? matcher : new M(fst, match_type)) { }
-
-  // This makes a copy of the FST.
-  SigmaMatcher(const SigmaMatcher<M> &matcher, bool safe = false)
-      : matcher_(new M(*matcher.matcher_, safe)),
-        match_type_(matcher.match_type_),
-        sigma_label_(matcher.sigma_label_),
-        rewrite_both_(matcher.rewrite_both_),
-        error_(matcher.error_),
-        state_(kNoStateId) {}
-
-  SigmaMatcher<M> *Copy(bool safe = false) const override {
-    return new SigmaMatcher<M>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override { return matcher_->Type(test); }
-
-  void SetState(StateId s) final {
-    if (state_ == s) return;
-    state_ = s;
-    matcher_->SetState(s);
-    has_sigma_ =
-        (sigma_label_ != kNoLabel) ? matcher_->Find(sigma_label_) : false;
-  }
-
-  bool Find(Label match_label) final {
-    match_label_ = match_label;
-    if (match_label == sigma_label_ && sigma_label_ != kNoLabel) {
-      FSTERROR() << "SigmaMatcher::Find: bad label (sigma)";
-      error_ = true;
-      return false;
-    }
-    if (matcher_->Find(match_label)) {
-      sigma_match_ = kNoLabel;
-      return true;
-    } else if (has_sigma_ && match_label != 0 && match_label != kNoLabel &&
-               matcher_->Find(sigma_label_)) {
-      sigma_match_ = match_label;
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  bool Done() const final { return matcher_->Done(); }
-
-  const Arc &Value() const final {
-    if (sigma_match_ == kNoLabel) {
-      return matcher_->Value();
-    } else {
-      sigma_arc_ = matcher_->Value();
-      if (rewrite_both_) {
-        if (sigma_arc_.ilabel == sigma_label_) sigma_arc_.ilabel = sigma_match_;
-        if (sigma_arc_.olabel == sigma_label_) sigma_arc_.olabel = sigma_match_;
-      } else if (match_type_ == MATCH_INPUT) {
-        sigma_arc_.ilabel = sigma_match_;
-      } else {
-        sigma_arc_.olabel = sigma_match_;
-      }
-      return sigma_arc_;
-    }
-  }
-
-  void Next() final {
-    matcher_->Next();
-    if (matcher_->Done() && has_sigma_ && (sigma_match_ == kNoLabel) &&
-        (match_label_ > 0)) {
-      matcher_->Find(sigma_label_);
-      sigma_match_ = match_label_;
-    }
-  }
-
-  Weight Final(StateId s) const final { return matcher_->Final(s); }
-
-  ssize_t Priority(StateId s) final {
-    if (sigma_label_ != kNoLabel) {
-      SetState(s);
-      return has_sigma_ ? kRequirePriority : matcher_->Priority(s);
-    } else {
-      return matcher_->Priority(s);
-    }
-  }
-
-  const FST &GetFst() const override { return matcher_->GetFst(); }
-
-  uint64 Properties(uint64 props) const override;
-
-  uint32 Flags() const override {
-    if (sigma_label_ == kNoLabel || match_type_ == MATCH_NONE) {
-      return matcher_->Flags();
-    }
-    return matcher_->Flags() | kRequireMatch;
-  }
-
-  Label SigmaLabel() const { return sigma_label_; }
-
- private:
-  std::unique_ptr<M> matcher_;
-  MatchType match_type_;   // Type of match requested.
-  Label sigma_label_;      // Label that represents the sigma transition.
-  bool rewrite_both_;      // Rewrite both sides when both are sigma_label_?
-  bool has_sigma_;         // Are there sigmas at the current state?
-  Label sigma_match_;      // Current label that matches sigma transition.
-  mutable Arc sigma_arc_;  // Arc to return when sigma match.
-  Label match_label_;      // Label being matched.
-  bool error_;             // Error encountered?
-  StateId state_;          // Matcher state.
-};
-
-template <class M>
-inline uint64 SigmaMatcher<M>::Properties(uint64 inprops) const {
-  auto outprops = matcher_->Properties(inprops);
-  if (error_) outprops |= kError;
-  if (match_type_ == MATCH_NONE) {
-    return outprops;
-  } else if (rewrite_both_) {
-    return outprops &
-           ~(kIDeterministic | kNonIDeterministic | kODeterministic |
-             kNonODeterministic | kILabelSorted | kNotILabelSorted |
-             kOLabelSorted | kNotOLabelSorted | kString);
-  } else if (match_type_ == MATCH_INPUT) {
-    return outprops &
-           ~(kIDeterministic | kNonIDeterministic | kODeterministic |
-             kNonODeterministic | kILabelSorted | kNotILabelSorted | kString |
-             kAcceptor);
-  } else if (match_type_ == MATCH_OUTPUT) {
-    return outprops &
-           ~(kIDeterministic | kNonIDeterministic | kODeterministic |
-             kNonODeterministic | kOLabelSorted | kNotOLabelSorted | kString |
-             kAcceptor);
-  } else {
-    // Shouldn't ever get here.
-    FSTERROR() << "SigmaMatcher: Bad match type: " << match_type_;
-    return 0;
-  }
-}
-
-// Flags for MultiEpsMatcher.
-
-// Return multi-epsilon arcs for Find(kNoLabel).
-const uint32 kMultiEpsList = 0x00000001;
-
-// Return a kNolabel loop for Find(multi_eps).
-const uint32 kMultiEpsLoop = 0x00000002;
-
-// MultiEpsMatcher: allows treating multiple non-0 labels as
-// non-consuming labels in addition to 0 that is always
-// non-consuming. Precise behavior controlled by 'flags' argument. By
-// default, the underlying matcher is constructed by
-// MultiEpsMatcher. The user can instead pass in this object; in that
-// case, MultiEpsMatcher takes its ownership iff 'own_matcher' is
-// true.
-template <class M>
-class MultiEpsMatcher {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // This makes a copy of the FST (w/o 'matcher' arg).
-  MultiEpsMatcher(const FST &fst, MatchType match_type,
-                  uint32 flags = (kMultiEpsLoop | kMultiEpsList),
-                  M *matcher = nullptr, bool own_matcher = true)
-      : matcher_(matcher ? matcher : new M(fst, match_type)),
-        flags_(flags),
-        own_matcher_(matcher ? own_matcher : true) {
-    Init(match_type);
-  }
-
-  // This doesn't copy the FST.
-  MultiEpsMatcher(const FST *fst, MatchType match_type,
-                  uint32 flags = (kMultiEpsLoop | kMultiEpsList),
-                  M *matcher = nullptr, bool own_matcher = true)
-      : matcher_(matcher ? matcher : new M(fst, match_type)),
-        flags_(flags),
-        own_matcher_(matcher ? own_matcher : true) {
-    Init(match_type);
-  }
-
-  // This makes a copy of the FST.
-  MultiEpsMatcher(const MultiEpsMatcher<M> &matcher, bool safe = false)
-      : matcher_(new M(*matcher.matcher_, safe)),
-        flags_(matcher.flags_),
-        own_matcher_(true),
-        multi_eps_labels_(matcher.multi_eps_labels_),
-        loop_(matcher.loop_) {
-    loop_.nextstate = kNoStateId;
-  }
-
-  ~MultiEpsMatcher() {
-    if (own_matcher_) delete matcher_;
-  }
-
-  MultiEpsMatcher<M> *Copy(bool safe = false) const {
-    return new MultiEpsMatcher<M>(*this, safe);
-  }
-
-  MatchType Type(bool test) const { return matcher_->Type(test); }
-
-  void SetState(StateId state) {
-    matcher_->SetState(state);
-    loop_.nextstate = state;
-  }
-
-  bool Find(Label label);
-
-  bool Done() const { return done_; }
-
-  const Arc &Value() const { return current_loop_ ? loop_ : matcher_->Value(); }
-
-  void Next() {
-    if (!current_loop_) {
-      matcher_->Next();
-      done_ = matcher_->Done();
-      if (done_ && multi_eps_iter_ != multi_eps_labels_.End()) {
-        ++multi_eps_iter_;
-        while ((multi_eps_iter_ != multi_eps_labels_.End()) &&
-               !matcher_->Find(*multi_eps_iter_)) {
-          ++multi_eps_iter_;
-        }
-        if (multi_eps_iter_ != multi_eps_labels_.End()) {
-          done_ = false;
-        } else {
-          done_ = !matcher_->Find(kNoLabel);
-        }
-      }
-    } else {
-      done_ = true;
-    }
-  }
-
-  const FST &GetFst() const { return matcher_->GetFst(); }
-
-  uint64 Properties(uint64 props) const { return matcher_->Properties(props); }
-
-  const M *GetMatcher() const { return matcher_; }
-
-  Weight Final(StateId s) const { return matcher_->Final(s); }
-
-  uint32 Flags() const { return matcher_->Flags(); }
-
-  ssize_t Priority(StateId s) { return matcher_->Priority(s); }
-
-  void AddMultiEpsLabel(Label label) {
-    if (label == 0) {
-      FSTERROR() << "MultiEpsMatcher: Bad multi-eps label: 0";
-    } else {
-      multi_eps_labels_.Insert(label);
-    }
-  }
-
-  void RemoveMultiEpsLabel(Label label) {
-    if (label == 0) {
-      FSTERROR() << "MultiEpsMatcher: Bad multi-eps label: 0";
-    } else {
-      multi_eps_labels_.Erase(label);
-    }
-  }
-
-  void ClearMultiEpsLabels() { multi_eps_labels_.Clear(); }
-
- private:
-  void Init(MatchType match_type) {
-    if (match_type == MATCH_INPUT) {
-      loop_.ilabel = kNoLabel;
-      loop_.olabel = 0;
-    } else {
-      loop_.ilabel = 0;
-      loop_.olabel = kNoLabel;
-    }
-    loop_.weight = Weight::One();
-    loop_.nextstate = kNoStateId;
-  }
-
-  M *matcher_;
-  uint32 flags_;
-  bool own_matcher_;  // Does this class delete the matcher?
-
-  // Multi-eps label set.
-  CompactSet<Label, kNoLabel> multi_eps_labels_;
-  typename CompactSet<Label, kNoLabel>::const_iterator multi_eps_iter_;
-
-  bool current_loop_;  // Current arc is the implicit loop?
-  mutable Arc loop_;   // For non-consuming symbols.
-  bool done_;          // Matching done?
-
-  MultiEpsMatcher &operator=(const MultiEpsMatcher &) = delete;
-};
-
-template <class M>
-inline bool MultiEpsMatcher<M>::Find(Label label) {
-  multi_eps_iter_ = multi_eps_labels_.End();
-  current_loop_ = false;
-  bool ret;
-  if (label == 0) {
-    ret = matcher_->Find(0);
-  } else if (label == kNoLabel) {
-    if (flags_ & kMultiEpsList) {
-      // Returns all non-consuming arcs (including epsilon).
-      multi_eps_iter_ = multi_eps_labels_.Begin();
-      while ((multi_eps_iter_ != multi_eps_labels_.End()) &&
-             !matcher_->Find(*multi_eps_iter_)) {
-        ++multi_eps_iter_;
-      }
-      if (multi_eps_iter_ != multi_eps_labels_.End()) {
-        ret = true;
-      } else {
-        ret = matcher_->Find(kNoLabel);
-      }
-    } else {
-      // Returns all epsilon arcs.
-      ret = matcher_->Find(kNoLabel);
-    }
-  } else if ((flags_ & kMultiEpsLoop) &&
-             multi_eps_labels_.Find(label) != multi_eps_labels_.End()) {
-    // Returns implicit loop.
-    current_loop_ = true;
-    ret = true;
-  } else {
-    ret = matcher_->Find(label);
-  }
-  done_ = !ret;
-  return ret;
-}
-
-// This class discards any implicit matches (e.g., the implicit epsilon
-// self-loops in the SortedMatcher). Matchers are most often used in
-// composition/intersection where the implicit matches are needed
-// e.g. for epsilon processing. However, if a matcher is simply being
-// used to look-up explicit label matches, this class saves the user
-// from having to check for and discard the unwanted implicit matches
-// themselves.
-template <class M>
-class ExplicitMatcher : public MatcherBase<typename M::Arc> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // This makes a copy of the FST.
-  ExplicitMatcher(const FST &fst, MatchType match_type, M *matcher = nullptr)
-      : matcher_(matcher ? matcher : new M(fst, match_type)),
-        match_type_(match_type),
-        error_(false) {}
-
-  // This doesn't copy the FST.
-  ExplicitMatcher(const FST *fst, MatchType match_type, M *matcher = nullptr)
-      : matcher_(matcher ? matcher : new M(fst, match_type)),
-        match_type_(match_type),
-        error_(false) {}
-
-  // This makes a copy of the FST.
-  ExplicitMatcher(const ExplicitMatcher<M> &matcher, bool safe = false)
-      : matcher_(new M(*matcher.matcher_, safe)),
-        match_type_(matcher.match_type_),
-        error_(matcher.error_) {}
-
-  ExplicitMatcher<M> *Copy(bool safe = false) const override {
-    return new ExplicitMatcher<M>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override { return matcher_->Type(test); }
-
-  void SetState(StateId s) final { matcher_->SetState(s); }
-
-  bool Find(Label label) final {
-    matcher_->Find(label);
-    CheckArc();
-    return !Done();
-  }
-
-  bool Done() const final { return matcher_->Done(); }
-
-  const Arc &Value() const final { return matcher_->Value(); }
-
-  void Next() final {
-    matcher_->Next();
-    CheckArc();
-  }
-
-  Weight Final(StateId s) const final { return matcher_->Final(s); }
-
-  ssize_t Priority(StateId s) final { return matcher_->Priority(s); }
-
-  const FST &GetFst() const final { return matcher_->GetFst(); }
-
-  uint64 Properties(uint64 inprops) const override {
-    return matcher_->Properties(inprops);
-  }
-
-  const M *GetMatcher() const { return matcher_.get(); }
-
-  uint32 Flags() const override { return matcher_->Flags(); }
-
- private:
-  // Checks current arc if available and explicit. If not available, stops. If
-  // not explicit, checks next ones.
-  void CheckArc() {
-    for (; !matcher_->Done(); matcher_->Next()) {
-      const auto label = match_type_ == MATCH_INPUT ? matcher_->Value().ilabel
-                                                    : matcher_->Value().olabel;
-      if (label != kNoLabel) return;
-    }
-  }
-
-  std::unique_ptr<M> matcher_;
-  MatchType match_type_;  // Type of match requested.
-  bool error_;            // Error encountered?
-};
-
-// Generic matcher, templated on the FST definition.
-//
-// Here is a typical use:
-//
-//   Matcher<StdFst> matcher(fst, MATCH_INPUT);
-//   matcher.SetState(state);
-//   if (matcher.Find(label))
-//     for (; !matcher.Done(); matcher.Next()) {
-//       auto &arc = matcher.Value();
-//       ...
-//     }
-template <class F>
-class Matcher {
- public:
-  using FST = F;
-  using Arc = typename F::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // This makes a copy of the FST.
-  Matcher(const FST &fst, MatchType match_type)
-      : owned_fst_(fst.Copy()),
-        base_(owned_fst_->InitMatcher(match_type)) {
-    if (!base_) base_.reset(new SortedMatcher<FST>(owned_fst_.get(),
-                                                   match_type));
-  }
-
-  // This doesn't copy the FST.
-  Matcher(const FST *fst, MatchType match_type)
-      : base_(fst->InitMatcher(match_type)) {
-    if (!base_) base_.reset(new SortedMatcher<FST>(fst, match_type));
-  }
-
-  // This makes a copy of the FST.
-  Matcher(const Matcher<FST> &matcher, bool safe = false)
-      : base_(matcher.base_->Copy(safe)) { }
-
-  // Takes ownership of the provided matcher.
-  explicit Matcher(MatcherBase<Arc> *base_matcher)
-      : base_(base_matcher) { }
-
-  Matcher<FST> *Copy(bool safe = false) const {
-    return new Matcher<FST>(*this, safe);
-  }
-
-  MatchType Type(bool test) const { return base_->Type(test); }
-
-  void SetState(StateId s) { base_->SetState(s); }
-
-  bool Find(Label label) { return base_->Find(label); }
-
-  bool Done() const { return base_->Done(); }
-
-  const Arc &Value() const { return base_->Value(); }
-
-  void Next() { base_->Next(); }
-
-  const FST &GetFst() const {
-    return static_cast<const FST &>(base_->GetFst());
-  }
-
-  uint64 Properties(uint64 props) const { return base_->Properties(props); }
-
-  Weight Final(StateId s) const { return base_->Final(s); }
-
-  uint32 Flags() const { return base_->Flags() & kMatcherFlags; }
-
-  ssize_t Priority(StateId s) { return base_->Priority(s); }
-
- private:
-  std::unique_ptr<const FST> owned_fst_;
-  std::unique_ptr<MatcherBase<Arc>> base_;
-};
-
-}  // namespace fst
-
-#endif  // FST_MATCHER_H_
diff --git a/projects/llm_framework/include/fst/memory.h b/projects/llm_framework/include/fst/memory.h
deleted file mode 100644
index c1f0bddc..00000000
--- a/projects/llm_framework/include/fst/memory.h
+++ /dev/null
@@ -1,443 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// FST memory utilities.
-
-#ifndef FST_MEMORY_H_
-#define FST_MEMORY_H_
-
-#include <list>
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include <fst/types.h>
-#include <fst/log.h>
-#include <fstream>
-
-namespace fst {
-
-// Default block allocation size.
-constexpr int kAllocSize = 64;
-
-// Minimum number of allocations per block.
-constexpr int kAllocFit = 4;
-
-// Base class for MemoryArena that allows (e.g.) MemoryArenaCollection to
-// easily manipulate collections of variously sized arenas.
-class MemoryArenaBase {
- public:
-  virtual ~MemoryArenaBase() {}
-  virtual size_t Size() const = 0;
-};
-
-namespace internal {
-
-// Allocates 'size' unintialized memory chunks of size object_size from
-// underlying blocks of (at least) size 'block_size * object_size'.
-// All blocks are freed when this class is deleted. Result of allocate() will
-// be aligned to object_size.
-template <size_t object_size>
-class MemoryArenaImpl : public MemoryArenaBase {
- public:
-  enum { kObjectSize = object_size };
-
-  explicit MemoryArenaImpl(size_t block_size = kAllocSize)
-      : block_size_(block_size * kObjectSize), block_pos_(0) {
-    blocks_.emplace_front(new char[block_size_]);
-  }
-
-  void *Allocate(size_t size) {
-    const auto byte_size = size * kObjectSize;
-    if (byte_size * kAllocFit > block_size_) {
-      // Large block; adds new large block.
-      auto *ptr = new char[byte_size];
-      blocks_.emplace_back(ptr);
-      return ptr;
-    }
-    if (block_pos_ + byte_size > block_size_) {
-      // Doesn't fit; adds new standard block.
-      auto *ptr = new char[block_size_];
-      block_pos_ = 0;
-      blocks_.emplace_front(ptr);
-    }
-    // Fits; uses current block.
-    auto *ptr = blocks_.front().get() + block_pos_;
-    block_pos_ += byte_size;
-    return ptr;
-  }
-
-  size_t Size() const override { return kObjectSize; }
-
- private:
-  const size_t block_size_;  // Default block size in bytes.
-  size_t block_pos_;   // Current position in block in bytes.
-  std::list<std::unique_ptr<char[]>> blocks_;  // List of allocated blocks.
-};
-
-}  // namespace internal
-
-template <typename T>
-using MemoryArena = internal::MemoryArenaImpl<sizeof(T)>;
-
-// Base class for MemoryPool that allows (e.g.) MemoryPoolCollection to easily
-// manipulate collections of variously sized pools.
-class MemoryPoolBase {
- public:
-  virtual ~MemoryPoolBase() {}
-  virtual size_t Size() const = 0;
-};
-
-namespace internal {
-
-// Allocates and frees initially uninitialized memory chunks of size
-// object_size. Keeps an internal list of freed chunks that are reused (as is)
-// on the next allocation if available. Chunks are constructed in blocks of size
-// 'pool_size'.
-template <size_t object_size>
-class MemoryPoolImpl : public MemoryPoolBase {
- public:
-  enum { kObjectSize = object_size };
-
-  struct Link {
-    char buf[kObjectSize];
-    Link *next;
-  };
-
-  explicit MemoryPoolImpl(size_t pool_size)
-      : mem_arena_(pool_size), free_list_(nullptr) {}
-
-  void *Allocate() {
-    if (free_list_ == nullptr) {
-      auto *link = static_cast<Link *>(mem_arena_.Allocate(1));
-      link->next = nullptr;
-      return link;
-    } else {
-      auto *link = free_list_;
-      free_list_ = link->next;
-      return link;
-    }
-  }
-
-  void Free(void *ptr) {
-    if (ptr) {
-      auto *link = static_cast<Link *>(ptr);
-      link->next = free_list_;
-      free_list_ = link;
-    }
-  }
-
-  size_t Size() const override { return kObjectSize; }
-
- private:
-  MemoryArena<Link> mem_arena_;
-  Link *free_list_;
-
-  MemoryPoolImpl(const MemoryPoolImpl &) = delete;
-  MemoryPoolImpl &operator=(const MemoryPoolImpl &) = delete;
-};
-
-}  // namespace internal
-
-// Allocates and frees initially uninitialized memory chunks of size sizeof(T).
-// All memory is freed when the class is deleted. The result of Allocate() will
-// be suitably memory-aligned. Combined with placement operator new and destroy
-// functions for the T class, this can be used to improve allocation efficiency.
-// See nlp/fst/lib/visit.h (global new) and nlp/fst/lib/dfs-visit.h (class new)
-// for examples.
-template <typename T>
-class MemoryPool : public internal::MemoryPoolImpl<sizeof(T)> {
- public:
-  // 'pool_size' specifies the size of the initial pool and how it is extended.
-  MemoryPool(size_t pool_size = kAllocSize)
-      : internal::MemoryPoolImpl<sizeof(T)>(pool_size) {}
-};
-
-// Stores a collection of memory arenas.
-class MemoryArenaCollection {
- public:
-  // 'block_size' specifies the block size of the arenas.
-  explicit MemoryArenaCollection(size_t block_size = kAllocSize)
-      : block_size_(block_size), ref_count_(1) {}
-
-  template <typename T>
-  MemoryArena<T> *Arena() {
-    if (sizeof(T) >= arenas_.size()) arenas_.resize(sizeof(T) + 1);
-    MemoryArenaBase *arena = arenas_[sizeof(T)].get();
-    if (arena == nullptr) {
-      arena = new MemoryArena<T>(block_size_);
-      arenas_[sizeof(T)].reset(arena);
-    }
-    return static_cast<MemoryArena<T> *>(arena);
-  }
-
-  size_t BlockSize() const { return block_size_; }
-
-  size_t RefCount() const { return ref_count_; }
-
-  size_t IncrRefCount() { return ++ref_count_; }
-
-  size_t DecrRefCount() { return --ref_count_; }
-
- private:
-  size_t block_size_;
-  size_t ref_count_;
-  std::vector<std::unique_ptr<MemoryArenaBase>> arenas_;
-};
-
-// Stores a collection of memory pools
-class MemoryPoolCollection {
- public:
-  // 'pool_size' specifies the size of initial pool and how it is extended.
-  explicit MemoryPoolCollection(size_t pool_size = kAllocSize)
-      : pool_size_(pool_size), ref_count_(1) {}
-
-  template <typename T>
-  MemoryPool<T> *Pool() {
-    if (sizeof(T) >= pools_.size()) pools_.resize(sizeof(T) + 1);
-    MemoryPoolBase *pool = pools_[sizeof(T)].get();
-    if (pool == nullptr) {
-      pool = new MemoryPool<T>(pool_size_);
-      pools_[sizeof(T)].reset(pool);
-    }
-    return static_cast<MemoryPool<T> *>(pool);
-  }
-
-  size_t PoolSize() const { return pool_size_; }
-
-  size_t RefCount() const { return ref_count_; }
-
-  size_t IncrRefCount() { return ++ref_count_; }
-
-  size_t DecrRefCount() { return --ref_count_; }
-
- private:
-  size_t pool_size_;
-  size_t ref_count_;
-  std::vector<std::unique_ptr<MemoryPoolBase>> pools_;
-};
-
-// STL allocator using memory arenas. Memory is allocated from underlying
-// blocks of size 'block_size * sizeof(T)'. Memory is freed only when all
-// objects using this allocator are destroyed and there is otherwise no reuse
-// (unlike PoolAllocator).
-//
-// This allocator has object-local state so it should not be used with splicing
-// or swapping operations between objects created with different allocators nor
-// should it be used if copies must be thread-safe. The result of allocate()
-// will be suitably memory-aligned.
-template <typename T>
-class BlockAllocator {
- public:
-  using Allocator = std::allocator<T>;
-  using size_type = typename Allocator::size_type;
-  using difference_type = typename Allocator::difference_type;
-  using pointer = typename Allocator::pointer;
-  using const_pointer = typename Allocator::const_pointer;
-  using reference = typename Allocator::reference;
-  using const_reference = typename Allocator::const_reference;
-  using value_type = typename Allocator::value_type;
-
-  template <typename U>
-  struct rebind {
-    using other = BlockAllocator<U>;
-  };
-
-  explicit BlockAllocator(size_t block_size = kAllocSize)
-      : arenas_(new MemoryArenaCollection(block_size)) {}
-
-  BlockAllocator(const BlockAllocator<T> &arena_alloc)
-      : arenas_(arena_alloc.Arenas()) {
-    Arenas()->IncrRefCount();
-  }
-
-  template <typename U>
-  explicit BlockAllocator(const BlockAllocator<U> &arena_alloc)
-      : arenas_(arena_alloc.Arenas()) {
-    Arenas()->IncrRefCount();
-  }
-
-  ~BlockAllocator() {
-    if (Arenas()->DecrRefCount() == 0) delete Arenas();
-  }
-
-  pointer address(reference ref) const { return Allocator().address(ref); }
-
-  const_pointer address(const_reference ref) const {
-    return Allocator().address(ref);
-  }
-
-  size_type max_size() const { return Allocator().max_size(); }
-
-  template <class U, class... Args>
-  void construct(U *p, Args &&... args) {
-    Allocator().construct(p, std::forward<Args>(args)...);
-  }
-
-  void destroy(pointer p) { Allocator().destroy(p); }
-
-  pointer allocate(size_type n, const void *hint = nullptr) {
-    if (n * kAllocFit <= kAllocSize) {
-      return static_cast<pointer>(Arena()->Allocate(n));
-    } else {
-      return Allocator().allocate(n, hint);
-    }
-  }
-
-  void deallocate(pointer p, size_type n) {
-    if (n * kAllocFit > kAllocSize) Allocator().deallocate(p, n);
-  }
-
-  MemoryArenaCollection *Arenas() const { return arenas_; }
-
- private:
-  MemoryArena<T> *Arena() { return arenas_->Arena<T>(); }
-
-  MemoryArenaCollection *arenas_;
-
-  BlockAllocator<T> operator=(const BlockAllocator<T> &);
-};
-
-template <typename T, typename U>
-bool operator==(const BlockAllocator<T> &alloc1,
-                const BlockAllocator<U> &alloc2) {
-  return false;
-}
-
-template <typename T, typename U>
-bool operator!=(const BlockAllocator<T> &alloc1,
-                const BlockAllocator<U> &alloc2) {
-  return true;
-}
-
-// STL allocator using memory pools. Memory is allocated from underlying
-// blocks of size 'block_size * sizeof(T)'. Keeps an internal list of freed
-// chunks thare are reused on the next allocation.
-//
-// This allocator has object-local state so it should not be used with splicing
-// or swapping operations between objects created with different allocators nor
-// should it be used if copies must be thread-safe. The result of allocate()
-// will be suitably memory-aligned.
-template <typename T>
-class PoolAllocator {
- public:
-  using Allocator = std::allocator<T>;
-  using size_type = typename Allocator::size_type;
-  using difference_type = typename Allocator::difference_type;
-  using pointer = typename Allocator::pointer;
-  using const_pointer = typename Allocator::const_pointer;
-  using reference = typename Allocator::reference;
-  using const_reference = typename Allocator::const_reference;
-  using value_type = typename Allocator::value_type;
-
-  template <typename U>
-  struct rebind {
-    using other = PoolAllocator<U>;
-  };
-
-  explicit PoolAllocator(size_t pool_size = kAllocSize)
-      : pools_(new MemoryPoolCollection(pool_size)) {}
-
-  PoolAllocator(const PoolAllocator<T> &pool_alloc)
-      : pools_(pool_alloc.Pools()) {
-    Pools()->IncrRefCount();
-  }
-
-  template <typename U>
-  explicit PoolAllocator(const PoolAllocator<U> &pool_alloc)
-      : pools_(pool_alloc.Pools()) {
-    Pools()->IncrRefCount();
-  }
-
-  ~PoolAllocator() {
-    if (Pools()->DecrRefCount() == 0) delete Pools();
-  }
-
-  pointer address(reference ref) const { return Allocator().address(ref); }
-
-  const_pointer address(const_reference ref) const {
-    return Allocator().address(ref);
-  }
-
-  size_type max_size() const { return Allocator().max_size(); }
-
-  template <class U, class... Args>
-  void construct(U *p, Args &&... args) {
-    Allocator().construct(p, std::forward<Args>(args)...);
-  }
-
-  void destroy(pointer p) { Allocator().destroy(p); }
-
-  pointer allocate(size_type n, const void *hint = nullptr) {
-    if (n == 1) {
-      return static_cast<pointer>(Pool<1>()->Allocate());
-    } else if (n == 2) {
-      return static_cast<pointer>(Pool<2>()->Allocate());
-    } else if (n <= 4) {
-      return static_cast<pointer>(Pool<4>()->Allocate());
-    } else if (n <= 8) {
-      return static_cast<pointer>(Pool<8>()->Allocate());
-    } else if (n <= 16) {
-      return static_cast<pointer>(Pool<16>()->Allocate());
-    } else if (n <= 32) {
-      return static_cast<pointer>(Pool<32>()->Allocate());
-    } else if (n <= 64) {
-      return static_cast<pointer>(Pool<64>()->Allocate());
-    } else {
-      return Allocator().allocate(n, hint);
-    }
-  }
-
-  void deallocate(pointer p, size_type n) {
-    if (n == 1) {
-      Pool<1>()->Free(p);
-    } else if (n == 2) {
-      Pool<2>()->Free(p);
-    } else if (n <= 4) {
-      Pool<4>()->Free(p);
-    } else if (n <= 8) {
-      Pool<8>()->Free(p);
-    } else if (n <= 16) {
-      Pool<16>()->Free(p);
-    } else if (n <= 32) {
-      Pool<32>()->Free(p);
-    } else if (n <= 64) {
-      Pool<64>()->Free(p);
-    } else {
-      Allocator().deallocate(p, n);
-    }
-  }
-
-  MemoryPoolCollection *Pools() const { return pools_; }
-
- private:
-  template <int n>
-  struct TN {
-    T buf[n];
-  };
-
-  template <int n>
-  MemoryPool<TN<n>> *Pool() {
-    return pools_->Pool<TN<n>>();
-  }
-
-  MemoryPoolCollection *pools_;
-
-  PoolAllocator<T> operator=(const PoolAllocator<T> &);
-};
-
-template <typename T, typename U>
-bool operator==(const PoolAllocator<T> &alloc1,
-                const PoolAllocator<U> &alloc2) {
-  return false;
-}
-
-template <typename T, typename U>
-bool operator!=(const PoolAllocator<T> &alloc1,
-                const PoolAllocator<U> &alloc2) {
-  return true;
-}
-
-}  // namespace fst
-
-#endif  // FST_MEMORY_H_
diff --git a/projects/llm_framework/include/fst/minimize.h b/projects/llm_framework/include/fst/minimize.h
deleted file mode 100644
index 9f17a22a..00000000
--- a/projects/llm_framework/include/fst/minimize.h
+++ /dev/null
@@ -1,568 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to minimize an FST.
-
-#ifndef FST_MINIMIZE_H_
-#define FST_MINIMIZE_H_
-
-#include <cmath>
-
-#include <algorithm>
-#include <map>
-#include <queue>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arcsort.h>
-#include <fst/connect.h>
-#include <fst/dfs-visit.h>
-#include <fst/encode.h>
-#include <fst/factor-weight.h>
-#include <fst/fst.h>
-#include <fst/mutable-fst.h>
-#include <fst/partition.h>
-#include <fst/push.h>
-#include <fst/queue.h>
-#include <fst/reverse.h>
-#include <fst/shortest-distance.h>
-#include <fst/state-map.h>
-
-
-namespace fst {
-namespace internal {
-
-// Comparator for creating partition.
-template <class Arc>
-class StateComparator {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  StateComparator(const Fst<Arc> &fst, const Partition<StateId> &partition)
-      : fst_(fst), partition_(partition) {}
-
-  // Compares state x with state y based on sort criteria.
-  bool operator()(const StateId x, const StateId y) const {
-    // Checks for final state equivalence.
-    const auto xfinal = fst_.Final(x).Hash();
-    const auto yfinal = fst_.Final(y).Hash();
-    if (xfinal < yfinal) {
-      return true;
-    } else if (xfinal > yfinal) {
-      return false;
-    }
-    // Checks for number of arcs.
-    if (fst_.NumArcs(x) < fst_.NumArcs(y)) return true;
-    if (fst_.NumArcs(x) > fst_.NumArcs(y)) return false;
-    // If the number of arcs are equal, checks for arc match.
-    for (ArcIterator<Fst<Arc>> aiter1(fst_, x), aiter2(fst_, y);
-         !aiter1.Done() && !aiter2.Done(); aiter1.Next(), aiter2.Next()) {
-      const auto &arc1 = aiter1.Value();
-      const auto &arc2 = aiter2.Value();
-      if (arc1.ilabel < arc2.ilabel) return true;
-      if (arc1.ilabel > arc2.ilabel) return false;
-      if (partition_.ClassId(arc1.nextstate) <
-          partition_.ClassId(arc2.nextstate))
-        return true;
-      if (partition_.ClassId(arc1.nextstate) >
-          partition_.ClassId(arc2.nextstate))
-        return false;
-    }
-    return false;
-  }
-
- private:
-  const Fst<Arc> &fst_;
-  const Partition<StateId> &partition_;
-};
-
-// Computes equivalence classes for cyclic unweighted acceptors. For cyclic
-// minimization we use the classic Hopcroft minimization algorithm, which has
-// complexity O(E log V) where E is the number of arcs and V is the number of
-// states.
-//
-// For more information, see:
-//
-//  Hopcroft, J. 1971. An n Log n algorithm for minimizing states in a finite
-//  automaton. Ms, Stanford University.
-//
-// Note: the original presentation of the paper was for a finite automaton (==
-// deterministic, unweighted acceptor), but we also apply it to the
-// nondeterministic case, where it is also applicable as long as the semiring is
-// idempotent (if the semiring is not idempotent, there are some complexities
-// in keeping track of the weight when there are multiple arcs to states that
-// will be merged, and we don't deal with this).
-template <class Arc, class Queue>
-class CyclicMinimizer {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using ClassId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using RevArc = ReverseArc<Arc>;
-
-  explicit CyclicMinimizer(const ExpandedFst<Arc> &fst) {
-    Initialize(fst);
-    Compute(fst);
-  }
-
-  const Partition<StateId> &GetPartition() const { return P_; }
-
- private:
-  // StateILabelHasher is a hashing object that computes a hash-function
-  // of an FST state that depends only on the set of ilabels on arcs leaving
-  // the state [note: it assumes that the arcs are ilabel-sorted].
-  // In order to work correctly for non-deterministic automata, multiple
-  // instances of the same ilabel count the same as a single instance.
-  class StateILabelHasher {
-   public:
-    explicit StateILabelHasher(const Fst<Arc> &fst) : fst_(fst) {}
-
-    using Label = typename Arc::Label;
-    using StateId = typename Arc::StateId;
-
-    size_t operator()(const StateId s) {
-      const size_t p1 = 7603;
-      const size_t p2 = 433024223;
-      size_t result = p2;
-      size_t current_ilabel = kNoLabel;
-      for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
-        Label this_ilabel = aiter.Value().ilabel;
-        if (this_ilabel != current_ilabel) {  // Ignores repeats.
-          result = p1 * result + this_ilabel;
-          current_ilabel = this_ilabel;
-        }
-      }
-      return result;
-    }
-
-   private:
-    const Fst<Arc> &fst_;
-  };
-
-  class ArcIterCompare {
-   public:
-    explicit ArcIterCompare(const Partition<StateId> &partition)
-        : partition_(partition) {}
-
-    ArcIterCompare(const ArcIterCompare &comp) : partition_(comp.partition_) {}
-
-    // Compares two iterators based on their input labels.
-    bool operator()(const ArcIterator<Fst<RevArc>> *x,
-                    const ArcIterator<Fst<RevArc>> *y) const {
-      const auto &xarc = x->Value();
-      const auto &yarc = y->Value();
-      return xarc.ilabel > yarc.ilabel;
-    }
-
-   private:
-    const Partition<StateId> &partition_;
-  };
-
-  using ArcIterQueue =
-      std::priority_queue<ArcIterator<Fst<RevArc>> *,
-                          std::vector<ArcIterator<Fst<RevArc>> *>,
-                          ArcIterCompare>;
-
- private:
-  // Prepartitions the space into equivalence classes. We ensure that final and
-  // non-final states always go into different equivalence classes, and we use
-  // class StateILabelHasher to make sure that most of the time, states with
-  // different sets of ilabels on arcs leaving them, go to different partitions.
-  // Note: for the O(n) guarantees we don't rely on the goodness of this
-  // hashing function---it just provides a bonus speedup.
-  void PrePartition(const ExpandedFst<Arc> &fst) {
-    VLOG(5) << "PrePartition";
-    StateId next_class = 0;
-    auto num_states = fst.NumStates();
-    // Allocates a temporary vector to store the initial class mappings, so that
-    // we can allocate the classes all at once.
-    std::vector<StateId> state_to_initial_class(num_states);
-    {
-      // We maintain two maps from hash-value to class---one for final states
-      // (final-prob == One()) and one for non-final states
-      // (final-prob == Zero()). We are processing unweighted acceptors, so the
-      // are the only two possible values.
-      using HashToClassMap = std::unordered_map<size_t, StateId>;
-      HashToClassMap hash_to_class_nonfinal;
-      HashToClassMap hash_to_class_final;
-      StateILabelHasher hasher(fst);
-      for (StateId s = 0; s < num_states; ++s) {
-        size_t hash = hasher(s);
-        HashToClassMap &this_map =
-            (fst.Final(s) != Weight::Zero() ? hash_to_class_final
-                                            : hash_to_class_nonfinal);
-        // Avoids two map lookups by using 'insert' instead of 'find'.
-        auto p = this_map.insert(std::make_pair(hash, next_class));
-        state_to_initial_class[s] = p.second ? next_class++ : p.first->second;
-      }
-      // Lets the unordered_maps go out of scope before we allocate the classes,
-      // to reduce the maximum amount of memory used.
-    }
-    P_.AllocateClasses(next_class);
-    for (StateId s = 0; s < num_states; ++s) {
-      P_.Add(s, state_to_initial_class[s]);
-    }
-    for (StateId c = 0; c < next_class; ++c) L_.Enqueue(c);
-    VLOG(5) << "Initial Partition: " << P_.NumClasses();
-  }
-
-  // Creates inverse transition Tr_ = rev(fst), loops over states in FST and
-  // splits on final, creating two blocks in the partition corresponding to
-  // final, non-final.
-  void Initialize(const ExpandedFst<Arc> &fst) {
-    // Constructs Tr.
-    Reverse(fst, &Tr_);
-    ILabelCompare<RevArc> ilabel_comp;
-    ArcSort(&Tr_, ilabel_comp);
-    // Tells the partition how many elements to allocate. The first state in
-    // Tr_ is super-final state.
-    P_.Initialize(Tr_.NumStates() - 1);
-    // Prepares initial partition.
-    PrePartition(fst);
-    // Allocates arc iterator queue.
-    ArcIterCompare comp(P_);
-    aiter_queue_.reset(new ArcIterQueue(comp));
-  }
-  // Partitions all classes with destination C.
-  void Split(ClassId C) {
-    // Prepares priority queue: opens arc iterator for each state in C, and
-    // inserts into priority queue.
-    for (PartitionIterator<StateId> siter(P_, C); !siter.Done(); siter.Next()) {
-      StateId s = siter.Value();
-      if (Tr_.NumArcs(s + 1)) {
-        aiter_queue_->push(new ArcIterator<Fst<RevArc>>(Tr_, s + 1));
-      }
-    }
-    // Now pops arc iterator from queue, splits entering equivalence class, and
-    // re-inserts updated iterator into queue.
-    Label prev_label = -1;
-    while (!aiter_queue_->empty()) {
-      std::unique_ptr<ArcIterator<Fst<RevArc>>> aiter(aiter_queue_->top());
-      aiter_queue_->pop();
-      if (aiter->Done()) continue;
-      const auto &arc = aiter->Value();
-      auto from_state = aiter->Value().nextstate - 1;
-      auto from_label = arc.ilabel;
-      if (prev_label != from_label) P_.FinalizeSplit(&L_);
-      auto from_class = P_.ClassId(from_state);
-      if (P_.ClassSize(from_class) > 1) P_.SplitOn(from_state);
-      prev_label = from_label;
-      aiter->Next();
-      if (!aiter->Done()) aiter_queue_->push(aiter.release());
-    }
-    P_.FinalizeSplit(&L_);
-  }
-
-  // Main loop for Hopcroft minimization.
-  void Compute(const Fst<Arc> &fst) {
-    // Processes active classes (FIFO, or FILO).
-    while (!L_.Empty()) {
-      const auto C = L_.Head();
-      L_.Dequeue();
-      Split(C);  // Splits on C, all labels in C.
-    }
-  }
-
- private:
-  // Partioning of states into equivalence classes.
-  Partition<StateId> P_;
-  // Set of active classes to be processed in partition P.
-  Queue L_;
-  // Reverses transition function.
-  VectorFst<RevArc> Tr_;
-  // Priority queue of open arc iterators for all states in the splitter
-  // equivalence class.
-  std::unique_ptr<ArcIterQueue> aiter_queue_;
-};
-
-// Computes equivalence classes for acyclic FST.
-//
-// Complexity:
-//
-//   O(E)
-//
-// where E is the number of arcs.
-//
-// For more information, see:
-//
-// Revuz, D. 1992. Minimization of acyclic deterministic automata in linear
-// time. Theoretical Computer Science 92(1): 181-189.
-template <class Arc>
-class AcyclicMinimizer {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using ClassId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit AcyclicMinimizer(const ExpandedFst<Arc> &fst) {
-    Initialize(fst);
-    Refine(fst);
-  }
-
-  const Partition<StateId> &GetPartition() { return partition_; }
-
- private:
-  // DFS visitor to compute the height (distance) to final state.
-  class HeightVisitor {
-   public:
-    HeightVisitor() : max_height_(0), num_states_(0) {}
-
-    // Invoked before DFS visit.
-    void InitVisit(const Fst<Arc> &fst) {}
-
-    // Invoked when state is discovered (2nd arg is DFS tree root).
-    bool InitState(StateId s, StateId root) {
-      // Extends height array and initialize height (distance) to 0.
-      for (StateId i = height_.size(); i <= s; ++i) height_.push_back(-1);
-      if (s >= num_states_) num_states_ = s + 1;
-      return true;
-    }
-
-    // Invoked when tree arc examined (to undiscovered state).
-    bool TreeArc(StateId s, const Arc &arc) { return true; }
-
-    // Invoked when back arc examined (to unfinished state).
-    bool BackArc(StateId s, const Arc &arc) { return true; }
-
-    // Invoked when forward or cross arc examined (to finished state).
-    bool ForwardOrCrossArc(StateId s, const Arc &arc) {
-      if (height_[arc.nextstate] + 1 > height_[s]) {
-        height_[s] = height_[arc.nextstate] + 1;
-      }
-      return true;
-    }
-
-    // Invoked when state finished (parent is kNoStateId for tree root).
-    void FinishState(StateId s, StateId parent, const Arc *parent_arc) {
-      if (height_[s] == -1) height_[s] = 0;
-      const auto h = height_[s] + 1;
-      if (parent >= 0) {
-        if (h > height_[parent]) height_[parent] = h;
-        if (h > max_height_) max_height_ = h;
-      }
-    }
-
-    // Invoked after DFS visit.
-    void FinishVisit() {}
-
-    size_t max_height() const { return max_height_; }
-
-    const std::vector<StateId> &height() const { return height_; }
-
-    size_t num_states() const { return num_states_; }
-
-   private:
-    std::vector<StateId> height_;
-    size_t max_height_;
-    size_t num_states_;
-  };
-
- private:
-  // Cluster states according to height (distance to final state)
-  void Initialize(const Fst<Arc> &fst) {
-    // Computes height (distance to final state).
-    HeightVisitor hvisitor;
-    DfsVisit(fst, &hvisitor);
-    // Creates initial partition based on height.
-    partition_.Initialize(hvisitor.num_states());
-    partition_.AllocateClasses(hvisitor.max_height() + 1);
-    const auto &hstates = hvisitor.height();
-    for (StateId s = 0; s < hstates.size(); ++s) partition_.Add(s, hstates[s]);
-  }
-
-  // Refines states based on arc sort (out degree, arc equivalence).
-  void Refine(const Fst<Arc> &fst) {
-    using EquivalenceMap = std::map<StateId, StateId, StateComparator<Arc>>;
-    StateComparator<Arc> comp(fst, partition_);
-    // Starts with tail (height = 0).
-    auto height = partition_.NumClasses();
-    for (StateId h = 0; h < height; ++h) {
-      EquivalenceMap equiv_classes(comp);
-      // Sorts states within equivalence class.
-      PartitionIterator<StateId> siter(partition_, h);
-      equiv_classes[siter.Value()] = h;
-      for (siter.Next(); !siter.Done(); siter.Next()) {
-        auto insert_result =
-            equiv_classes.insert(std::make_pair(siter.Value(), kNoStateId));
-        if (insert_result.second) {
-          insert_result.first->second = partition_.AddClass();
-        }
-      }
-      // Creates refined partition.
-      for (siter.Reset(); !siter.Done();) {
-        const auto s = siter.Value();
-        const auto old_class = partition_.ClassId(s);
-        const auto new_class = equiv_classes[s];
-        // A move operation can invalidate the iterator, so we first update
-        // the iterator to the next element before we move the current element
-        // out of the list.
-        siter.Next();
-        if (old_class != new_class) partition_.Move(s, new_class);
-      }
-    }
-  }
-
- private:
-  Partition<StateId> partition_;
-};
-
-// Given a partition and a Mutable FST, merges states of Fst in place (i.e.,
-// destructively). Merging works by taking the first state in a class of the
-// partition to be the representative state for the class. Each arc is then
-// reconnected to this state. All states in the class are merged by adding
-// their arcs to the representative state.
-template <class Arc>
-void MergeStates(const Partition<typename Arc::StateId> &partition,
-                 MutableFst<Arc> *fst) {
-  using StateId = typename Arc::StateId;
-  std::vector<StateId> state_map(partition.NumClasses());
-  for (StateId i = 0; i < partition.NumClasses(); ++i) {
-    PartitionIterator<StateId> siter(partition, i);
-    state_map[i] = siter.Value();  // First state in partition.
-  }
-  // Relabels destination states.
-  for (StateId c = 0; c < partition.NumClasses(); ++c) {
-    for (PartitionIterator<StateId> siter(partition, c); !siter.Done();
-         siter.Next()) {
-      const auto s = siter.Value();
-      for (MutableArcIterator<MutableFst<Arc>> aiter(fst, s); !aiter.Done();
-           aiter.Next()) {
-        auto arc = aiter.Value();
-        arc.nextstate = state_map[partition.ClassId(arc.nextstate)];
-        if (s == state_map[c]) {  // For the first state, just sets destination.
-          aiter.SetValue(arc);
-        } else {
-          fst->AddArc(state_map[c], std::move(arc));
-        }
-      }
-    }
-  }
-  fst->SetStart(state_map[partition.ClassId(fst->Start())]);
-  Connect(fst);
-}
-
-template <class Arc>
-void AcceptorMinimize(MutableFst<Arc> *fst,
-                      bool allow_acyclic_minimization = true) {
-  if (!(fst->Properties(kAcceptor | kUnweighted, true) ==
-        (kAcceptor | kUnweighted))) {
-    FSTERROR() << "FST is not an unweighted acceptor";
-    fst->SetProperties(kError, kError);
-    return;
-  }
-  // Connects FST before minimization, handles disconnected states.
-  Connect(fst);
-  if (fst->NumStates() == 0) return;
-  if (allow_acyclic_minimization && fst->Properties(kAcyclic, true)) {
-    // Acyclic minimization (Revuz).
-    VLOG(2) << "Acyclic minimization";
-    ArcSort(fst, ILabelCompare<Arc>());
-    AcyclicMinimizer<Arc> minimizer(*fst);
-    MergeStates(minimizer.GetPartition(), fst);
-  } else {
-    // Either the FST has cycles, or it's generated from non-deterministic input
-    // (which the Revuz algorithm can't handle), so use the cyclic minimization
-    // algorithm of Hopcroft.
-    VLOG(2) << "Cyclic minimization";
-    CyclicMinimizer<Arc, LifoQueue<typename Arc::StateId>> minimizer(*fst);
-    MergeStates(minimizer.GetPartition(), fst);
-  }
-  // Merges in appropriate semiring
-  ArcUniqueMapper<Arc> mapper(*fst);
-  StateMap(fst, mapper);
-}
-
-}  // namespace internal
-
-// In place minimization of deterministic weighted automata and transducers,
-// and also non-deterministic ones if they use an idempotent semiring.
-// For transducers, if the 'sfst' argument is not null, the algorithm
-// produces a compact factorization of the minimal transducer.
-//
-// In the acyclic deterministic case, we use an algorithm from Revuz that is
-// linear in the number of arcs (edges) in the machine.
-//
-// In the cyclic or non-deterministic case, we use the classical Hopcroft
-// minimization (which was presented for the deterministic case but which
-// also works for non-deterministic FSTs); this has complexity O(e log v).
-//
-template <class Arc>
-void Minimize(MutableFst<Arc> *fst, MutableFst<Arc> *sfst = nullptr,
-              float delta = kShortestDelta, bool allow_nondet = false) {
-  using Weight = typename Arc::Weight;
-  const auto props = fst->Properties(
-      kAcceptor | kIDeterministic | kWeighted | kUnweighted, true);
-  bool allow_acyclic_minimization;
-  if (props & kIDeterministic) {
-    allow_acyclic_minimization = true;
-  } else {
-    // Our approach to minimization of non-deterministic FSTs will only work in
-    // idempotent semirings---for non-deterministic inputs, a state could have
-    // multiple transitions to states that will get merged, and we'd have to
-    // sum their weights. The algorithm doesn't handle that.
-    if (!(Weight::Properties() & kIdempotent)) {
-      fst->SetProperties(kError, kError);
-      FSTERROR() << "Cannot minimize a non-deterministic FST over a "
-                    "non-idempotent semiring";
-      return;
-    } else if (!allow_nondet) {
-      fst->SetProperties(kError, kError);
-      FSTERROR() << "Refusing to minimize a non-deterministic FST with "
-                 << "allow_nondet = false";
-      return;
-    }
-    // The Revuz algorithm won't work for nondeterministic inputs, so if the
-    // input is nondeterministic, we'll have to pass a bool saying not to use
-    // that algorithm. We check at this level rather than in AcceptorMinimize(),
-    // because it's possible that the FST at this level could be deterministic,
-    // but a harmless type of non-determinism could be introduced by Encode()
-    // (thanks to kEncodeWeights, if the FST has epsilons and has a final
-    // weight with weights equal to some epsilon arc.)
-    allow_acyclic_minimization = false;
-  }
-  if (!(props & kAcceptor)) {  // Weighted transducer.
-    VectorFst<GallicArc<Arc, GALLIC_LEFT>> gfst;
-    ArcMap(*fst, &gfst, ToGallicMapper<Arc, GALLIC_LEFT>());
-    fst->DeleteStates();
-    gfst.SetProperties(kAcceptor, kAcceptor);
-    Push(&gfst, REWEIGHT_TO_INITIAL, delta);
-    ArcMap(&gfst, QuantizeMapper<GallicArc<Arc, GALLIC_LEFT>>(delta));
-    EncodeMapper<GallicArc<Arc, GALLIC_LEFT>> encoder(
-        kEncodeLabels | kEncodeWeights, ENCODE);
-    Encode(&gfst, &encoder);
-    internal::AcceptorMinimize(&gfst, allow_acyclic_minimization);
-    Decode(&gfst, encoder);
-    if (!sfst) {
-      FactorWeightFst<GallicArc<Arc, GALLIC_LEFT>,
-                      GallicFactor<typename Arc::Label, Weight, GALLIC_LEFT>>
-          fwfst(gfst);
-      std::unique_ptr<SymbolTable> osyms(
-          fst->OutputSymbols() ? fst->OutputSymbols()->Copy() : nullptr);
-      ArcMap(fwfst, fst, FromGallicMapper<Arc, GALLIC_LEFT>());
-      fst->SetOutputSymbols(osyms.get());
-    } else {
-      sfst->SetOutputSymbols(fst->OutputSymbols());
-      GallicToNewSymbolsMapper<Arc, GALLIC_LEFT> mapper(sfst);
-      ArcMap(gfst, fst, &mapper);
-      fst->SetOutputSymbols(sfst->InputSymbols());
-    }
-  } else if (props & kWeighted) {  // Weighted acceptor.
-    Push(fst, REWEIGHT_TO_INITIAL, delta);
-    ArcMap(fst, QuantizeMapper<Arc>(delta));
-    EncodeMapper<Arc> encoder(kEncodeLabels | kEncodeWeights, ENCODE);
-    Encode(fst, &encoder);
-    internal::AcceptorMinimize(fst, allow_acyclic_minimization);
-    Decode(fst, encoder);
-  } else {  // Unweighted acceptor.
-    internal::AcceptorMinimize(fst, allow_acyclic_minimization);
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_MINIMIZE_H_
diff --git a/projects/llm_framework/include/fst/mutable-fst.h b/projects/llm_framework/include/fst/mutable-fst.h
deleted file mode 100644
index 9031770d..00000000
--- a/projects/llm_framework/include/fst/mutable-fst.h
+++ /dev/null
@@ -1,398 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Expanded FST augmented with mutators; interface class definition and
-// mutable arc iterator interface.
-
-#ifndef FST_MUTABLE_FST_H_
-#define FST_MUTABLE_FST_H_
-
-#include <stddef.h>
-#include <sys/types.h>
-
-#include <istream>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-#include <fstream>
-
-#include <fst/expanded-fst.h>
-
-
-namespace fst {
-
-template <class Arc>
-struct MutableArcIteratorData;
-
-// Abstract interface for an expanded FST which also supports mutation
-// operations. To modify arcs, use MutableArcIterator.
-template <class A>
-class MutableFst : public ExpandedFst<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  virtual MutableFst<Arc> &operator=(const Fst<Arc> &fst) = 0;
-
-  MutableFst<Arc> &operator=(const MutableFst<Arc> &fst) {
-    return operator=(static_cast<const Fst<Arc> &>(fst));
-  }
-
-  // Sets the initial state.
-  virtual void SetStart(StateId) = 0;
-
-  // Sets a state's final weight.
-  virtual void SetFinal(StateId, Weight) = 0;
-
-  // Sets property bits w.r.t. mask.
-  virtual void SetProperties(uint64 props, uint64 mask) = 0;
-
-  // Adds a state and returns its ID.
-  virtual StateId AddState() = 0;
-
-  // Adds an arc to state.
-  virtual void AddArc(StateId, const Arc &arc) = 0;
-
-  // Adds an arc (passed by rvalue reference) to state. Allows subclasses
-  // to optionally implement move semantics. Defaults to lvalue overload.
-  virtual void AddArc(StateId state, Arc &&arc) { AddArc(state, arc); }
-
-  // Deletes some states, preserving original StateId ordering.
-  virtual void DeleteStates(const std::vector<StateId> &) = 0;
-
-  // Delete all states.
-  virtual void DeleteStates() = 0;
-
-  // Delete some arcs at a given state.
-  virtual void DeleteArcs(StateId, size_t n) = 0;
-
-  // Delete all arcs at a given state.
-  virtual void DeleteArcs(StateId) = 0;
-
-  // Optional, best effort only.
-  virtual void ReserveStates(StateId n) {}
-
-  // Optional, best effort only.
-  virtual void ReserveArcs(StateId s, size_t n) {}
-
-  // Returns input label symbol table or nullptr if not specified.
-  const SymbolTable *InputSymbols() const override = 0;
-
-  // Returns output label symbol table or nullptr if not specified.
-  const SymbolTable *OutputSymbols() const override = 0;
-
-  // Returns input label symbol table or nullptr if not specified.
-  virtual SymbolTable *MutableInputSymbols() = 0;
-
-  // Returns output label symbol table or nullptr if not specified.
-  virtual SymbolTable *MutableOutputSymbols() = 0;
-
-  // Sets input label symbol table; pass nullptr to delete table.
-  virtual void SetInputSymbols(const SymbolTable *isyms) = 0;
-
-  // Sets output label symbol table; pass nullptr to delete table.
-  virtual void SetOutputSymbols(const SymbolTable *osyms) = 0;
-
-  // Gets a copy of this MutableFst. See Fst<>::Copy() for further doc.
-  MutableFst<A> *Copy(bool safe = false) const override = 0;
-
-  // Reads a MutableFst from an input stream, returning nullptr on error.
-  static MutableFst<Arc> *Read(std::istream &strm, const FstReadOptions &opts) {
-    FstReadOptions ropts(opts);
-    FstHeader hdr;
-    if (ropts.header) {
-      hdr = *opts.header;
-    } else {
-      if (!hdr.Read(strm, opts.source)) return nullptr;
-      ropts.header = &hdr;
-    }
-    if (!(hdr.Properties() & kMutable)) {
-      LOG(ERROR) << "MutableFst::Read: Not a MutableFst: " << ropts.source;
-      return nullptr;
-    }
-    const auto &fst_type = hdr.FstType();
-    const auto reader = FstRegister<Arc>::GetRegister()->GetReader(fst_type);
-    if (!reader) {
-      LOG(ERROR) << "MutableFst::Read: Unknown FST type \"" << fst_type
-                 << "\" (arc type = \"" << A::Type() << "\"): " << ropts.source;
-      return nullptr;
-    }
-    auto *fst = reader(strm, ropts);
-    if (!fst) return nullptr;
-    return static_cast<MutableFst<Arc> *>(fst);
-  }
-
-  // Reads a MutableFst from a file; returns nullptr on error. An empty
-  // filename results in reading from standard input. If convert is true,
-  // convert to a mutable FST subclass (given by convert_type) in the case
-  // that the input FST is non-mutable.
-  static MutableFst<Arc> *Read(const string &filename, bool convert = false,
-                               const string &convert_type = "vector") {
-    if (convert == false) {
-      if (!filename.empty()) {
-        std::ifstream strm(filename,
-                                std::ios_base::in | std::ios_base::binary);
-        if (!strm) {
-          LOG(ERROR) << "MutableFst::Read: Can't open file: " << filename;
-          return nullptr;
-        }
-        return Read(strm, FstReadOptions(filename));
-      } else {
-        return Read(std::cin, FstReadOptions("standard input"));
-      }
-    } else {  // Converts to 'convert_type' if not mutable.
-      std::unique_ptr<Fst<Arc>> ifst(Fst<Arc>::Read(filename));
-      if (!ifst) return nullptr;
-      if (ifst->Properties(kMutable, false)) {
-        return static_cast<MutableFst<Arc> *>(ifst.release());
-      } else {
-        std::unique_ptr<Fst<Arc>> ofst(Convert(*ifst, convert_type));
-        ifst.reset();
-        if (!ofst) return nullptr;
-        if (!ofst->Properties(kMutable, false)) {
-          LOG(ERROR) << "MutableFst: Bad convert type: " << convert_type;
-        }
-        return static_cast<MutableFst<Arc> *>(ofst.release());
-      }
-    }
-  }
-
-  // For generic mutuble arc iterator construction; not normally called
-  // directly by users.
-  virtual void InitMutableArcIterator(StateId s,
-                                      MutableArcIteratorData<Arc> *data) = 0;
-};
-
-// Mutable arc iterator interface, templated on the Arc definition. This is
-// used by mutable arc iterator specializations that are returned by the
-// InitMutableArcIterator MutableFst method.
-template <class Arc>
-class MutableArcIteratorBase : public ArcIteratorBase<Arc> {
- public:
-  // Sets current arc.
-  virtual void SetValue(const Arc &) = 0;
-};
-
-template <class Arc>
-struct MutableArcIteratorData {
-  MutableArcIteratorBase<Arc> *base;  // Specific iterator.
-};
-
-// Generic mutable arc iterator, templated on the FST definition; a wrapper
-// around a pointer to a more specific one.
-//
-// Here is a typical use:
-//
-//   for (MutableArcIterator<StdFst> aiter(&fst, s);
-//        !aiter.Done();
-//         aiter.Next()) {
-//     StdArc arc = aiter.Value();
-//     arc.ilabel = 7;
-//     aiter.SetValue(arc);
-//     ...
-//   }
-//
-// This version requires function calls.
-template <class FST>
-class MutableArcIterator {
- public:
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-
-  MutableArcIterator(FST *fst, StateId s) {
-    fst->InitMutableArcIterator(s, &data_);
-  }
-
-  ~MutableArcIterator() { delete data_.base; }
-
-  bool Done() const { return data_.base->Done(); }
-
-  const Arc &Value() const { return data_.base->Value(); }
-
-  void Next() { data_.base->Next(); }
-
-  size_t Position() const { return data_.base->Position(); }
-
-  void Reset() { data_.base->Reset(); }
-
-  void Seek(size_t a) { data_.base->Seek(a); }
-
-  void SetValue(const Arc &arc) { data_.base->SetValue(arc); }
-
-  uint32 Flags() const { return data_.base->Flags(); }
-
-  void SetFlags(uint32 flags, uint32 mask) {
-    return data_.base->SetFlags(flags, mask);
-  }
-
- private:
-  MutableArcIteratorData<Arc> data_;
-
-  MutableArcIterator(const MutableArcIterator &) = delete;
-  MutableArcIterator &operator=(const MutableArcIterator &) = delete;
-};
-
-namespace internal {
-
-// MutableFst<A> case: abstract methods.
-template <class Arc>
-inline typename Arc::Weight Final(const MutableFst<Arc> &fst,
-                                  typename Arc::StateId s) {
-  return fst.Final(s);
-}
-
-template <class Arc>
-inline ssize_t NumArcs(const MutableFst<Arc> &fst, typename Arc::StateId s) {
-  return fst.NumArcs(s);
-}
-
-template <class Arc>
-inline ssize_t NumInputEpsilons(const MutableFst<Arc> &fst,
-                                typename Arc::StateId s) {
-  return fst.NumInputEpsilons(s);
-}
-
-template <class Arc>
-inline ssize_t NumOutputEpsilons(const MutableFst<Arc> &fst,
-                                 typename Arc::StateId s) {
-  return fst.NumOutputEpsilons(s);
-}
-
-}  // namespace internal
-
-// A useful alias when using StdArc.
-using StdMutableFst = MutableFst<StdArc>;
-
-// This is a helper class template useful for attaching a MutableFst interface
-// to its implementation, handling reference counting and COW semantics.
-template <class Impl, class FST = MutableFst<typename Impl::Arc>>
-class ImplToMutableFst : public ImplToExpandedFst<Impl, FST> {
- public:
-  using Arc = typename Impl::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using ImplToExpandedFst<Impl, FST>::operator=;
-
-  void SetStart(StateId s) override {
-    MutateCheck();
-    GetMutableImpl()->SetStart(s);
-  }
-
-  void SetFinal(StateId s, Weight weight) override {
-    MutateCheck();
-    GetMutableImpl()->SetFinal(s, std::move(weight));
-  }
-
-  void SetProperties(uint64 props, uint64 mask) override {
-    // Can skip mutate check if extrinsic properties don't change,
-    // since it is then safe to update all (shallow) copies
-    const auto exprops = kExtrinsicProperties & mask;
-    if (GetImpl()->Properties(exprops) != (props & exprops)) MutateCheck();
-    GetMutableImpl()->SetProperties(props, mask);
-  }
-
-  StateId AddState() override {
-    MutateCheck();
-    return GetMutableImpl()->AddState();
-  }
-
-  void AddArc(StateId s, const Arc &arc) override {
-    MutateCheck();
-    GetMutableImpl()->AddArc(s, arc);
-  }
-
-  void AddArc(StateId s, Arc &&arc) override {
-    MutateCheck();
-    GetMutableImpl()->AddArc(s, std::move(arc));
-  }
-
-  void DeleteStates(const std::vector<StateId> &dstates) override {
-    MutateCheck();
-    GetMutableImpl()->DeleteStates(dstates);
-  }
-
-  void DeleteStates() override {
-    if (!Unique()) {
-      const auto *isymbols = GetImpl()->InputSymbols();
-      const auto *osymbols = GetImpl()->OutputSymbols();
-      SetImpl(std::make_shared<Impl>());
-      GetMutableImpl()->SetInputSymbols(isymbols);
-      GetMutableImpl()->SetOutputSymbols(osymbols);
-    } else {
-      GetMutableImpl()->DeleteStates();
-    }
-  }
-
-  void DeleteArcs(StateId s, size_t n) override {
-    MutateCheck();
-    GetMutableImpl()->DeleteArcs(s, n);
-  }
-
-  void DeleteArcs(StateId s) override {
-    MutateCheck();
-    GetMutableImpl()->DeleteArcs(s);
-  }
-
-  void ReserveStates(StateId s) override {
-    MutateCheck();
-    GetMutableImpl()->ReserveStates(s);
-  }
-
-  void ReserveArcs(StateId s, size_t n) override {
-    MutateCheck();
-    GetMutableImpl()->ReserveArcs(s, n);
-  }
-
-  const SymbolTable *InputSymbols() const override {
-    return GetImpl()->InputSymbols();
-  }
-
-  const SymbolTable *OutputSymbols() const override {
-    return GetImpl()->OutputSymbols();
-  }
-
-  SymbolTable *MutableInputSymbols() override {
-    MutateCheck();
-    return GetMutableImpl()->InputSymbols();
-  }
-
-  SymbolTable *MutableOutputSymbols() override {
-    MutateCheck();
-    return GetMutableImpl()->OutputSymbols();
-  }
-
-  void SetInputSymbols(const SymbolTable *isyms) override {
-    MutateCheck();
-    GetMutableImpl()->SetInputSymbols(isyms);
-  }
-
-  void SetOutputSymbols(const SymbolTable *osyms) override {
-    MutateCheck();
-    GetMutableImpl()->SetOutputSymbols(osyms);
-  }
-
- protected:
-  using ImplToExpandedFst<Impl, FST>::GetImpl;
-  using ImplToExpandedFst<Impl, FST>::GetMutableImpl;
-  using ImplToExpandedFst<Impl, FST>::Unique;
-  using ImplToExpandedFst<Impl, FST>::SetImpl;
-  using ImplToExpandedFst<Impl, FST>::InputSymbols;
-
-  explicit ImplToMutableFst(std::shared_ptr<Impl> impl)
-      : ImplToExpandedFst<Impl, FST>(impl) {}
-
-  ImplToMutableFst(const ImplToMutableFst<Impl, FST> &fst, bool safe)
-      : ImplToExpandedFst<Impl, FST>(fst, safe) {}
-
-  void MutateCheck() {
-    if (!Unique()) SetImpl(std::make_shared<Impl>(*this));
-  }
-};
-
-}  // namespace fst
-
-#endif  // FST_MUTABLE_FST_H_
diff --git a/projects/llm_framework/include/fst/pair-weight.h b/projects/llm_framework/include/fst/pair-weight.h
deleted file mode 100644
index 1f2e963c..00000000
--- a/projects/llm_framework/include/fst/pair-weight.h
+++ /dev/null
@@ -1,155 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Pair weight templated base class for weight classes that contain two weights
-// (e.g. Product, Lexicographic).
-
-#ifndef FST_PAIR_WEIGHT_H_
-#define FST_PAIR_WEIGHT_H_
-
-#include <climits>
-#include <stack>
-#include <string>
-#include <utility>
-
-#include <fst/flags.h>
-#include <fst/log.h>
-
-#include <fst/weight.h>
-
-
-namespace fst {
-
-template <class W1, class W2>
-class PairWeight {
- public:
-  using ReverseWeight =
-      PairWeight<typename W1::ReverseWeight, typename W2::ReverseWeight>;
-
-  PairWeight() {}
-
-  PairWeight(W1 w1, W2 w2) : value1_(std::move(w1)), value2_(std::move(w2)) {}
-
-  static const PairWeight<W1, W2> &Zero() {
-    static const PairWeight zero(W1::Zero(), W2::Zero());
-    return zero;
-  }
-
-  static const PairWeight<W1, W2> &One() {
-    static const PairWeight one(W1::One(), W2::One());
-    return one;
-  }
-
-  static const PairWeight<W1, W2> &NoWeight() {
-    static const PairWeight no_weight(W1::NoWeight(), W2::NoWeight());
-    return no_weight;
-  }
-
-  std::istream &Read(std::istream &strm) {
-    value1_.Read(strm);
-    return value2_.Read(strm);
-  }
-
-  std::ostream &Write(std::ostream &strm) const {
-    value1_.Write(strm);
-    return value2_.Write(strm);
-  }
-
-  bool Member() const { return value1_.Member() && value2_.Member(); }
-
-  size_t Hash() const {
-    const auto h1 = value1_.Hash();
-    const auto h2 = value2_.Hash();
-    static constexpr int lshift = 5;
-    static constexpr int rshift = CHAR_BIT * sizeof(size_t) - 5;
-    return h1 << lshift ^ h1 >> rshift ^ h2;
-  }
-
-  PairWeight<W1, W2> Quantize(float delta = kDelta) const {
-    return PairWeight<W1, W2>(value1_.Quantize(delta), value2_.Quantize(delta));
-  }
-
-  ReverseWeight Reverse() const {
-    return ReverseWeight(value1_.Reverse(), value2_.Reverse());
-  }
-
-  const W1 &Value1() const { return value1_; }
-
-  const W2 &Value2() const { return value2_; }
-
-  void SetValue1(const W1 &weight) { value1_ = weight; }
-
-  void SetValue2(const W2 &weight) { value2_ = weight; }
-
- private:
-  W1 value1_;
-  W2 value2_;
-};
-
-template <class W1, class W2>
-inline bool operator==(const PairWeight<W1, W2> &w1,
-                       const PairWeight<W1, W2> &w2) {
-  return w1.Value1() == w2.Value1() && w1.Value2() == w2.Value2();
-}
-
-template <class W1, class W2>
-inline bool operator!=(const PairWeight<W1, W2> &w1,
-                       const PairWeight<W1, W2> &w2) {
-  return w1.Value1() != w2.Value1() || w1.Value2() != w2.Value2();
-}
-
-template <class W1, class W2>
-inline bool ApproxEqual(const PairWeight<W1, W2> &w1,
-                        const PairWeight<W1, W2> &w2, float delta = kDelta) {
-  return ApproxEqual(w1.Value1(), w2.Value1(), delta) &&
-         ApproxEqual(w1.Value2(), w2.Value2(), delta);
-}
-
-template <class W1, class W2>
-inline std::ostream &operator<<(std::ostream &strm,
-                                const PairWeight<W1, W2> &weight) {
-  CompositeWeightWriter writer(strm);
-  writer.WriteBegin();
-  writer.WriteElement(weight.Value1());
-  writer.WriteElement(weight.Value2());
-  writer.WriteEnd();
-  return strm;
-}
-
-template <class W1, class W2>
-inline std::istream &operator>>(std::istream &strm,
-                                PairWeight<W1, W2> &weight) {
-  CompositeWeightReader reader(strm);
-  reader.ReadBegin();
-  W1 w1;
-  reader.ReadElement(&w1);
-  weight.SetValue1(w1);
-  W2 w2;
-  reader.ReadElement(&w2, true);
-  weight.SetValue2(w2);
-  reader.ReadEnd();
-  return strm;
-}
-
-// This function object returns weights by calling the underlying generators
-// and forming a pair. This is intended primarily for testing.
-template <class W1, class W2>
-class WeightGenerate<PairWeight<W1, W2>> {
- public:
-  using Weight = PairWeight<W1, W2>;
-  using Generate1 = WeightGenerate<W1>;
-  using Generate2 = WeightGenerate<W2>;
-
-  explicit WeightGenerate(bool allow_zero = true)
-      : generate1_(allow_zero), generate2_(allow_zero) {}
-
-  Weight operator()() const { return Weight(generate1_(), generate2_()); }
-
- private:
-  Generate1 generate1_;
-  Generate2 generate2_;
-};
-
-}  // namespace fst
-
-#endif  // FST_PAIR_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/partition.h b/projects/llm_framework/include/fst/partition.h
deleted file mode 100644
index 5dbbe46a..00000000
--- a/projects/llm_framework/include/fst/partition.h
+++ /dev/null
@@ -1,305 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to create a partition of states.
-
-#ifndef FST_PARTITION_H_
-#define FST_PARTITION_H_
-
-#include <algorithm>
-#include <vector>
-
-
-#include <fst/queue.h>
-
-
-namespace fst {
-namespace internal {
-
-template <typename T>
-class PartitionIterator;
-
-// Defines a partitioning of elements, used to represent equivalence classes
-// for FST operations like minimization. T must be a signed integer type.
-//
-// The elements are numbered from 0 to num_elements - 1.
-// Initialize(num_elements) sets up the class for a given number of elements.
-// We maintain a partition of these elements into classes. The classes are also
-// numbered from zero; you can add a class with AddClass(), or add them in bulk
-// with AllocateClasses(num_classes). Initially the elements are not assigned
-// to any class; you set up the initial mapping from elements to classes by
-// calling Add(element_id, class_id). You can also move an element to a
-// different class by calling Move(element_id, class_id).
-//
-// We also support a rather specialized interface that allows you to efficiently
-// split classes in the Hopcroft minimization algorithm. This maintains a
-// binary partition of each class.  Let's call these, rather arbitrarily, the
-// 'yes' subset and the 'no' subset of each class, and assume that by default,
-// each element of a class is in its 'no' subset. When one calls
-// SplitOn(element_id), element_id is moved to the 'yes' subset of its class.
-// (If it was already in the 'yes' set, it just stays there). The aim is to
-// enable (later) splitting the class in two in time no greater than the time
-// already spent calling SplitOn() for that class. We keep a list of the classes
-// which have nonempty 'yes' sets, as visited_classes_. When one calls
-// FinalizeSplit(Queue *l), for each class in visited_classes_ whose 'yes'
-// and 'no' sets are both nonempty, it will create a new class consisting of
-// the smaller of the two subsets (and this class will be added to the queue),
-// and the old class will now be the larger of the two subsets. This call also
-// resets all the yes/no partitions so that everything is in the 'no' subsets.
-//
-// One cannot use the Move() function if SplitOn() has been called without
-// a subsequent call to FinalizeSplit()
-template <typename T>
-class Partition {
- public:
-  Partition() {}
-
-  explicit Partition(T num_elements) { Initialize(num_elements); }
-
-  // Creates an empty partition for num_elements. This means that the elements
-  // are not assigned to a class (i.e class_index = -1); you should set up the
-  // number of classes using AllocateClasses() or AddClass(), and allocate each
-  // element to a class by calling Add(element, class_id).
-  void Initialize(size_t num_elements) {
-    elements_.resize(num_elements);
-    classes_.reserve(num_elements);
-    classes_.clear();
-    yes_counter_ = 1;
-  }
-
-  // Adds a class; returns new number of classes.
-  T AddClass() {
-    auto num_classes = classes_.size();
-    classes_.resize(num_classes + 1);
-    return num_classes;
-  }
-
-  // Adds 'num_classes' new (empty) classes.
-  void AllocateClasses(T num_classes) {
-    classes_.resize(classes_.size() + num_classes);
-  }
-
-  // Adds element_id to class_id. element_id should already have been allocated
-  // by calling Initialize(num_elements)---or the constructor taking
-  // num_elements---with num_elements > element_id. element_id must not
-  // currently be a member of any class; once elements have been added to a
-  // class, use the Move() method to move them from one class to another.
-  void Add(T element_id, T class_id) {
-    auto &this_element = elements_[element_id];
-    auto &this_class = classes_[class_id];
-    ++this_class.size;
-    // Adds the element to the 'no' subset of the class.
-    auto no_head = this_class.no_head;
-    if (no_head >= 0) elements_[no_head].prev_element = element_id;
-    this_class.no_head = element_id;
-    this_element.class_id = class_id;
-    // Adds to the 'no' subset of the class.
-    this_element.yes = 0;
-    this_element.next_element = no_head;
-    this_element.prev_element = -1;
-  }
-
-  // Moves element_id from 'no' subset of its current class to 'no' subset of
-  // class class_id. This may not work correctly if you have called SplitOn()
-  // [for any element] and haven't subsequently called FinalizeSplit().
-  void Move(T element_id, T class_id) {
-    auto elements = &(elements_[0]);
-    auto &element = elements[element_id];
-    auto &old_class = classes_[element.class_id];
-    --old_class.size;
-    // Excises the element from the 'no' list of its old class, where it is
-    // assumed to be.
-    if (element.prev_element >= 0) {
-      elements[element.prev_element].next_element = element.next_element;
-    } else {
-      old_class.no_head = element.next_element;
-    }
-    if (element.next_element >= 0) {
-      elements[element.next_element].prev_element = element.prev_element;
-    }
-    // Adds to new class.
-    Add(element_id, class_id);
-  }
-
-  // Moves element_id to the 'yes' subset of its class if it was in the 'no'
-  // subset, and marks the class as having been visited.
-  void SplitOn(T element_id) {
-    auto elements = &(elements_[0]);
-    auto &element = elements[element_id];
-    if (element.yes == yes_counter_) {
-      return;  // Already in the 'yes' set; nothing to do.
-    }
-    auto class_id = element.class_id;
-    auto &this_class = classes_[class_id];
-    // Excises the element from the 'no' list of its class.
-    if (element.prev_element >= 0) {
-      elements[element.prev_element].next_element = element.next_element;
-    } else {
-      this_class.no_head = element.next_element;
-    }
-    if (element.next_element >= 0) {
-      elements[element.next_element].prev_element = element.prev_element;
-    }
-    // Adds the element to the 'yes' list.
-    if (this_class.yes_head >= 0) {
-      elements[this_class.yes_head].prev_element = element_id;
-    } else {
-      visited_classes_.push_back(class_id);
-    }
-    element.yes = yes_counter_;
-    element.next_element = this_class.yes_head;
-    element.prev_element = -1;
-    this_class.yes_head = element_id;
-    this_class.yes_size++;
-  }
-
-  // This should be called after one has possibly called SplitOn for one or more
-  // elements, thus moving those elements to the 'yes' subset for their class.
-  // For each class that has a nontrivial split (i.e., it's not the case that
-  // all members are in the 'yes' or 'no' subset), this function creates a new
-  // class containing the smaller of the two subsets of elements, leaving the
-  // larger group of elements in the old class. The identifier of the new class
-  // will be added to the queue provided as the pointer L. This method then
-  // moves all elements to the 'no' subset of their class.
-  template <class Queue>
-  void FinalizeSplit(Queue *queue) {
-    for (const auto &visited_class : visited_classes_) {
-      const auto new_class = SplitRefine(visited_class);
-      if (new_class != -1 && queue) queue->Enqueue(new_class);
-    }
-    visited_classes_.clear();
-    // Incrementation sets all the 'yes' members of the elements to false.
-    ++yes_counter_;
-  }
-
-  const T ClassId(T element_id) const { return elements_[element_id].class_id; }
-
-  const size_t ClassSize(T class_id) const { return classes_[class_id].size; }
-
-  const T NumClasses() const { return classes_.size(); }
-
- private:
-  friend class PartitionIterator<T>;
-
-  // Information about a given element.
-  struct Element {
-    T class_id;      // Class ID of this element.
-    T yes;           // This is to be interpreted as a bool, true if it's in the
-                     // 'yes' set of this class. The interpretation as bool is
-                     // (yes == yes_counter_ ? true : false).
-    T next_element;  // Next element in the 'no' list or 'yes' list of this
-                     // class, whichever of the two we belong to (think of
-                     // this as the 'next' in a doubly-linked list, although
-                     // it is an index into the elements array). Negative
-                     // values corresponds to null.
-    T prev_element;  // Previous element in the 'no' or 'yes' doubly linked
-                     // list. Negative values corresponds to null.
-  };
-
-  // Information about a given class.
-  struct Class {
-    Class() : size(0), yes_size(0), no_head(-1), yes_head(-1) {}
-    T size;      // Total number of elements in this class ('no' plus 'yes'
-                 // subsets).
-    T yes_size;  // Total number of elements of 'yes' subset of this class.
-    T no_head;   // Index of head element of doubly-linked list in 'no' subset.
-                 // Everything is in the 'no' subset until you call SplitOn().
-                 // -1 means no element.
-    T yes_head;  // Index of head element of doubly-linked list in 'yes' subset.
-                 // -1 means no element.
-  };
-
-  // This method, called from FinalizeSplit(), checks whether a class has to
-  // be split (a class will be split only if its 'yes' and 'no' subsets are
-  // both nonempty, but one can assume that since this function was called, the
-  // 'yes' subset is nonempty). It splits by taking the smaller subset and
-  // making it a new class, and leaving the larger subset of elements in the
-  // 'no' subset of the old class. It returns the new class if created, or -1
-  // if none was created.
-  T SplitRefine(T class_id) {
-    auto yes_size = classes_[class_id].yes_size;
-    auto size = classes_[class_id].size;
-    auto no_size = size - yes_size;
-    if (no_size == 0) {
-      // All members are in the 'yes' subset, so we don't have to create a new
-      // class, just move them all to the 'no' subset.
-      classes_[class_id].no_head = classes_[class_id].yes_head;
-      classes_[class_id].yes_head = -1;
-      classes_[class_id].yes_size = 0;
-      return -1;
-    } else {
-      auto new_class_id = classes_.size();
-      classes_.resize(classes_.size() + 1);
-      auto &old_class = classes_[class_id];
-      auto &new_class = classes_[new_class_id];
-      // The new_class will have the values from the constructor.
-      if (no_size < yes_size) {
-        // Moves the 'no' subset to new class ('no' subset).
-        new_class.no_head = old_class.no_head;
-        new_class.size = no_size;
-        // And makes the 'yes' subset of the old class ('no' subset).
-        old_class.no_head = old_class.yes_head;
-        old_class.yes_head = -1;
-        old_class.size = yes_size;
-        old_class.yes_size = 0;
-      } else {
-        // Moves the 'yes' subset to the new class (to the 'no' subset)
-        new_class.size = yes_size;
-        new_class.no_head = old_class.yes_head;
-        // Retains only the 'no' subset in the old class.
-        old_class.size = no_size;
-        old_class.yes_size = 0;
-        old_class.yes_head = -1;
-      }
-      auto elements = &(elements_[0]);
-      // Updates the 'class_id' of all the elements we moved.
-      for (auto e = new_class.no_head; e >= 0; e = elements[e].next_element) {
-        elements[e].class_id = new_class_id;
-      }
-      return new_class_id;
-    }
-  }
-
-  // elements_[i] contains all info about the i'th element.
-  std::vector<Element> elements_;
-  // classes_[i] contains all info about the i'th class.
-  std::vector<Class> classes_;
-  // Set of visited classes to be used in split refine.
-  std::vector<T> visited_classes_;
-  // yes_counter_ is used in interpreting the 'yes' members of class Element.
-  // If element.yes == yes_counter_, we interpret that element as being in the
-  // 'yes' subset of its class. This allows us to, in effect, set all those
-  // bools to false at a stroke by incrementing yes_counter_.
-  T yes_counter_;
-};
-
-// Iterates over members of the 'no' subset of a class in a partition. (When
-// this is used, everything is in the 'no' subset).
-template <typename T>
-class PartitionIterator {
- public:
-  using Element = typename Partition<T>::Element;
-
-  PartitionIterator(const Partition<T> &partition, T class_id)
-      : partition_(partition),
-        element_id_(partition_.classes_[class_id].no_head),
-        class_id_(class_id) {}
-
-  bool Done() { return element_id_ < 0; }
-
-  const T Value() { return element_id_; }
-
-  void Next() { element_id_ = partition_.elements_[element_id_].next_element; }
-
-  void Reset() { element_id_ = partition_.classes_[class_id_].no_head; }
-
- private:
-  const Partition<T> &partition_;
-  T element_id_;
-  T class_id_;
-};
-
-}  // namespace internal
-}  // namespace fst
-
-#endif  // FST_PARTITION_H_
diff --git a/projects/llm_framework/include/fst/power-weight.h b/projects/llm_framework/include/fst/power-weight.h
deleted file mode 100644
index f2f3cbdb..00000000
--- a/projects/llm_framework/include/fst/power-weight.h
+++ /dev/null
@@ -1,168 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Cartesian power weight semiring operation definitions.
-
-#ifndef FST_POWER_WEIGHT_H_
-#define FST_POWER_WEIGHT_H_
-
-#include <string>
-
-#include <fst/tuple-weight.h>
-#include <fst/weight.h>
-
-
-namespace fst {
-
-// Cartesian power semiring: W ^ n
-//
-// Forms:
-//  - a left semimodule when W is a left semiring,
-//  - a right semimodule when W is a right semiring,
-//  - a bisemimodule when W is a semiring,
-//    the free semimodule of rank n over W
-// The Times operation is overloaded to provide the left and right scalar
-// products.
-template <class W, size_t n>
-class PowerWeight : public TupleWeight<W, n> {
- public:
-  using ReverseWeight = PowerWeight<typename W::ReverseWeight, n>;
-
-  PowerWeight() {}
-
-  explicit PowerWeight(const TupleWeight<W, n> &weight)
-      : TupleWeight<W, n>(weight) {}
-
-  template <class Iterator>
-  PowerWeight(Iterator begin, Iterator end) : TupleWeight<W, n>(begin, end) {}
-
-  // Initialize component `index` to `weight`; initialize all other components
-  // to `default_weight`
-  PowerWeight(size_t index, const W &weight,
-              const W &default_weight = W::Zero())
-      : TupleWeight<W, n>(index, weight, default_weight) {}
-
-  static const PowerWeight &Zero() {
-    static const PowerWeight zero(TupleWeight<W, n>::Zero());
-    return zero;
-  }
-
-  static const PowerWeight &One() {
-    static const PowerWeight one(TupleWeight<W, n>::One());
-    return one;
-  }
-
-  static const PowerWeight &NoWeight() {
-    static const PowerWeight no_weight(TupleWeight<W, n>::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type =
-        new string(W::Type() + "_^" + std::to_string(n));
-    return *type;
-  }
-
-  static constexpr uint64 Properties() {
-    return W::Properties() &
-           (kLeftSemiring | kRightSemiring | kCommutative | kIdempotent);
-  }
-
-  PowerWeight Quantize(float delta = kDelta) const {
-    return PowerWeight(TupleWeight<W, n>::Quantize(delta));
-  }
-
-  ReverseWeight Reverse() const {
-    return ReverseWeight(TupleWeight<W, n>::Reverse());
-  }
-};
-
-// Semiring plus operation.
-template <class W, size_t n>
-inline PowerWeight<W, n> Plus(const PowerWeight<W, n> &w1,
-                              const PowerWeight<W, n> &w2) {
-  PowerWeight<W, n> result;
-  for (size_t i = 0; i < n; ++i) {
-    result.SetValue(i, Plus(w1.Value(i), w2.Value(i)));
-  }
-  return result;
-}
-
-// Semiring times operation.
-template <class W, size_t n>
-inline PowerWeight<W, n> Times(const PowerWeight<W, n> &w1,
-                               const PowerWeight<W, n> &w2) {
-  PowerWeight<W, n> result;
-  for (size_t i = 0; i < n; ++i) {
-    result.SetValue(i, Times(w1.Value(i), w2.Value(i)));
-  }
-  return result;
-}
-
-// Semiring divide operation.
-template <class W, size_t n>
-inline PowerWeight<W, n> Divide(const PowerWeight<W, n> &w1,
-                                const PowerWeight<W, n> &w2,
-                                DivideType type = DIVIDE_ANY) {
-  PowerWeight<W, n> result;
-  for (size_t i = 0; i < n; ++i) {
-    result.SetValue(i, Divide(w1.Value(i), w2.Value(i), type));
-  }
-  return result;
-}
-
-// Semimodule left scalar product.
-template <class W, size_t n>
-inline PowerWeight<W, n> Times(const W &scalar,
-                               const PowerWeight<W, n> &weight) {
-  PowerWeight<W, n> result;
-  for (size_t i = 0; i < n; ++i) {
-    result.SetValue(i, Times(scalar, weight.Value(i)));
-  }
-  return result;
-}
-
-// Semimodule right scalar product.
-template <class W, size_t n>
-inline PowerWeight<W, n> Times(const PowerWeight<W, n> &weight,
-                               const W &scalar) {
-  PowerWeight<W, n> result;
-  for (size_t i = 0; i < n; ++i) {
-    result.SetValue(i, Times(weight.Value(i), scalar));
-  }
-  return result;
-}
-
-// Semimodule dot product.
-template <class W, size_t n>
-inline W DotProduct(const PowerWeight<W, n> &w1, const PowerWeight<W, n> &w2) {
-  W result(W::Zero());
-  for (size_t i = 0; i < n; ++i) {
-    result = Plus(result, Times(w1.Value(i), w2.Value(i)));
-  }
-  return result;
-}
-
-// This function object generates weights over the Cartesian power of rank
-// n over the underlying weight. This is intended primarily for testing.
-template <class W, size_t n>
-class WeightGenerate<PowerWeight<W, n>> {
- public:
-  using Weight = PowerWeight<W, n>;
-  using Generate = WeightGenerate<W>;
-
-  explicit WeightGenerate(bool allow_zero = true) : generate_(allow_zero) {}
-
-  Weight operator()() const {
-    Weight result;
-    for (size_t i = 0; i < n; ++i) result.SetValue(i, generate_());
-    return result;
-  }
-
- private:
-  Generate generate_;
-};
-
-}  // namespace fst
-
-#endif  // FST_POWER_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/product-weight.h b/projects/llm_framework/include/fst/product-weight.h
deleted file mode 100644
index 56a18be1..00000000
--- a/projects/llm_framework/include/fst/product-weight.h
+++ /dev/null
@@ -1,107 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Product weight set and associated semiring operation definitions.
-
-#ifndef FST_PRODUCT_WEIGHT_H_
-#define FST_PRODUCT_WEIGHT_H_
-
-#include <string>
-#include <utility>
-
-#include <fst/pair-weight.h>
-#include <fst/weight.h>
-
-
-namespace fst {
-
-// Product semiring: W1 * W2.
-template <class W1, class W2>
-class ProductWeight : public PairWeight<W1, W2> {
- public:
-  using ReverseWeight =
-      ProductWeight<typename W1::ReverseWeight, typename W2::ReverseWeight>;
-
-  ProductWeight() {}
-
-  explicit ProductWeight(const PairWeight<W1, W2> &weight)
-      : PairWeight<W1, W2>(weight) {}
-
-  ProductWeight(W1 w1, W2 w2)
-      : PairWeight<W1, W2>(std::move(w1), std::move(w2)) {}
-
-  static const ProductWeight &Zero() {
-    static const ProductWeight zero(PairWeight<W1, W2>::Zero());
-    return zero;
-  }
-
-  static const ProductWeight &One() {
-    static const ProductWeight one(PairWeight<W1, W2>::One());
-    return one;
-  }
-
-  static const ProductWeight &NoWeight() {
-    static const ProductWeight no_weight(PairWeight<W1, W2>::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type =
-        new string(W1::Type() + "_X_" + W2::Type());
-    return *type;
-  }
-
-  static constexpr uint64 Properties() {
-    return W1::Properties() & W2::Properties() &
-           (kLeftSemiring | kRightSemiring | kCommutative | kIdempotent);
-  }
-
-  ProductWeight Quantize(float delta = kDelta) const {
-    return ProductWeight(PairWeight<W1, W2>::Quantize(delta));
-  }
-
-  ReverseWeight Reverse() const {
-    return ReverseWeight(PairWeight<W1, W2>::Reverse());
-  }
-};
-
-template <class W1, class W2>
-inline ProductWeight<W1, W2> Plus(const ProductWeight<W1, W2> &w1,
-                                  const ProductWeight<W1, W2> &w2) {
-  return ProductWeight<W1, W2>(Plus(w1.Value1(), w2.Value1()),
-                               Plus(w1.Value2(), w2.Value2()));
-}
-
-template <class W1, class W2>
-inline ProductWeight<W1, W2> Times(const ProductWeight<W1, W2> &w1,
-                                   const ProductWeight<W1, W2> &w2) {
-  return ProductWeight<W1, W2>(Times(w1.Value1(), w2.Value1()),
-                               Times(w1.Value2(), w2.Value2()));
-}
-
-template <class W1, class W2>
-inline ProductWeight<W1, W2> Divide(const ProductWeight<W1, W2> &w1,
-                                    const ProductWeight<W1, W2> &w2,
-                                    DivideType typ = DIVIDE_ANY) {
-  return ProductWeight<W1, W2>(Divide(w1.Value1(), w2.Value1(), typ),
-                               Divide(w1.Value2(), w2.Value2(), typ));
-}
-
-// This function object generates weights by calling the underlying generators
-// for the template weight types, like all other pair weight types. This is
-// intended primarily for testing.
-template <class W1, class W2>
-class WeightGenerate<ProductWeight<W1, W2>> :
-    public WeightGenerate<PairWeight<W1, W2>> {
- public:
-  using Weight = ProductWeight<W1, W2>;
-  using Generate = WeightGenerate<PairWeight<W1, W2>>;
-
-  explicit WeightGenerate(bool allow_zero = true) : Generate(allow_zero) {}
-
-  Weight operator()() const { return Weight(Generate::operator()()); }
-};
-
-}  // namespace fst
-
-#endif  // FST_PRODUCT_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/project.h b/projects/llm_framework/include/fst/project.h
deleted file mode 100644
index 5a82cf14..00000000
--- a/projects/llm_framework/include/fst/project.h
+++ /dev/null
@@ -1,159 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to project an FST on to its domain or range.
-
-#ifndef FST_PROJECT_H_
-#define FST_PROJECT_H_
-
-#include <fst/arc-map.h>
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-// This specifies whether to project on input or output.
-enum ProjectType { PROJECT_INPUT = 1, PROJECT_OUTPUT = 2 };
-
-// Mapper to implement projection per arc.
-template <class A>
-class ProjectMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-
-  constexpr explicit ProjectMapper(ProjectType project_type)
-      : project_type_(project_type) {}
-
-  ToArc operator()(const FromArc &arc) const {
-    const auto label = project_type_ == PROJECT_INPUT ? arc.ilabel : arc.olabel;
-    return ToArc(label, label, arc.weight, arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const {
-    return MAP_NO_SUPERFINAL;
-  }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return project_type_ == PROJECT_INPUT ? MAP_COPY_SYMBOLS
-                                          : MAP_CLEAR_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return project_type_ == PROJECT_OUTPUT ? MAP_COPY_SYMBOLS
-                                           : MAP_CLEAR_SYMBOLS;
-  }
-
-  constexpr uint64 Properties(uint64 props) const {
-    return ProjectProperties(props, project_type_ == PROJECT_INPUT);
-  }
-
- private:
-  const ProjectType project_type_;
-};
-
-// Projects an FST onto its domain or range by either copying each arcs' input
-// label to the output label or vice versa.
-//
-// Complexity:
-//
-//   Time: O(V + E)
-//   Space: O(1)
-//
-// where V is the number of states and E is the number of arcs.
-template <class Arc>
-inline void Project(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-                    ProjectType project_type) {
-  ArcMap(ifst, ofst, ProjectMapper<Arc>(project_type));
-  switch (project_type) {
-    case PROJECT_INPUT:
-      ofst->SetOutputSymbols(ifst.InputSymbols());
-      return;
-    case PROJECT_OUTPUT:
-      ofst->SetInputSymbols(ifst.OutputSymbols());
-      return;
-  }
-}
-
-// Destructive variant of the above.
-template <class Arc>
-inline void Project(MutableFst<Arc> *fst, ProjectType project_type) {
-  ArcMap(fst, ProjectMapper<Arc>(project_type));
-  switch (project_type) {
-    case PROJECT_INPUT:
-      fst->SetOutputSymbols(fst->InputSymbols());
-      return;
-    case PROJECT_OUTPUT:
-      fst->SetInputSymbols(fst->OutputSymbols());
-      return;
-  }
-}
-
-// Projects an FST onto its domain or range by either copying each arc's input
-// label to the output label or vice versa. This version is a delayed FST.
-//
-// Complexity:
-//
-//   Time: O(v + e)
-//   Space: O(1)
-//
-// where v is the number of states visited and e is the number of arcs visited.
-// Constant time and to visit an input state or arc is assumed and exclusive of
-// caching.
-template <class A>
-class ProjectFst : public ArcMapFst<A, A, ProjectMapper<A>> {
- public:
-  using FromArc = A;
-  using ToArc = A;
-
-  using Impl = internal::ArcMapFstImpl<A, A, ProjectMapper<A>>;
-
-  ProjectFst(const Fst<A> &fst, ProjectType project_type)
-      : ArcMapFst<A, A, ProjectMapper<A>>(fst, ProjectMapper<A>(project_type)) {
-    if (project_type == PROJECT_INPUT) {
-      GetMutableImpl()->SetOutputSymbols(fst.InputSymbols());
-    }
-    if (project_type == PROJECT_OUTPUT) {
-      GetMutableImpl()->SetInputSymbols(fst.OutputSymbols());
-    }
-  }
-
-  // See Fst<>::Copy() for doc.
-  ProjectFst(const ProjectFst<A> &fst, bool safe = false)
-      : ArcMapFst<A, A, ProjectMapper<A>>(fst, safe) {}
-
-  // Gets a copy of this ProjectFst. See Fst<>::Copy() for further doc.
-  ProjectFst<A> *Copy(bool safe = false) const override {
-    return new ProjectFst(*this, safe);
-  }
-
- private:
-  using ImplToFst<Impl>::GetMutableImpl;
-};
-
-// Specialization for ProjectFst.
-template <class A>
-class StateIterator<ProjectFst<A>>
-    : public StateIterator<ArcMapFst<A, A, ProjectMapper<A>>> {
- public:
-  explicit StateIterator(const ProjectFst<A> &fst)
-      : StateIterator<ArcMapFst<A, A, ProjectMapper<A>>>(fst) {}
-};
-
-// Specialization for ProjectFst.
-template <class A>
-class ArcIterator<ProjectFst<A>>
-    : public ArcIterator<ArcMapFst<A, A, ProjectMapper<A>>> {
- public:
-  using StateId = typename A::StateId;
-
-  ArcIterator(const ProjectFst<A> &fst, StateId s)
-      : ArcIterator<ArcMapFst<A, A, ProjectMapper<A>>>(fst, s) {}
-};
-
-// Useful alias when using StdArc.
-using StdProjectFst = ProjectFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_PROJECT_H_
diff --git a/projects/llm_framework/include/fst/properties.h b/projects/llm_framework/include/fst/properties.h
deleted file mode 100644
index 157247a6..00000000
--- a/projects/llm_framework/include/fst/properties.h
+++ /dev/null
@@ -1,468 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// FST property bits.
-
-#ifndef FST_PROPERTIES_H_
-#define FST_PROPERTIES_H_
-
-#include <sys/types.h>
-#include <vector>
-
-#include <fst/compat.h>
-
-namespace fst {
-
-// The property bits here assert facts about an FST. If individual bits are
-// added, then the composite properties below, the property functions and
-// property names in properties.cc, and TestProperties() in test-properties.h
-// should be updated.
-
-// BINARY PROPERTIES
-//
-// For each property below, there is a single bit. If it is set, the property is
-// true. If it is not set, the property is false.
-
-// The Fst is an ExpandedFst.
-constexpr uint64 kExpanded = 0x0000000000000001ULL;
-
-// The Fst is a MutableFst.
-constexpr uint64 kMutable = 0x0000000000000002ULL;
-
-// An error was detected while constructing/using the FST.
-constexpr uint64 kError = 0x0000000000000004ULL;
-
-// TRINARY PROPERTIES
-//
-// For each of these properties below there is a pair of property bits, one
-// positive and one negative. If the positive bit is set, the property is true.
-// If the negative bit is set, the property is false. If neither is set, the
-// property has unknown value. Both should never be simultaneously set. The
-// individual positive and negative bit pairs should be adjacent with the
-// positive bit at an odd and lower position.
-
-// ilabel == olabel for each arc.
-constexpr uint64 kAcceptor = 0x0000000000010000ULL;
-// ilabel != olabel for some arc.
-constexpr uint64 kNotAcceptor = 0x0000000000020000ULL;
-
-// ilabels unique leaving each state.
-constexpr uint64 kIDeterministic = 0x0000000000040000ULL;
-// ilabels not unique leaving some state.
-constexpr uint64 kNonIDeterministic = 0x0000000000080000ULL;
-
-// olabels unique leaving each state.
-constexpr uint64 kODeterministic = 0x0000000000100000ULL;
-// olabels not unique leaving some state.
-constexpr uint64 kNonODeterministic = 0x0000000000200000ULL;
-
-// FST has input/output epsilons.
-constexpr uint64 kEpsilons = 0x0000000000400000ULL;
-// FST has no input/output epsilons.
-constexpr uint64 kNoEpsilons = 0x0000000000800000ULL;
-
-// FST has input epsilons.
-constexpr uint64 kIEpsilons = 0x0000000001000000ULL;
-// FST has no input epsilons.
-constexpr uint64 kNoIEpsilons = 0x0000000002000000ULL;
-
-// FST has output epsilons.
-constexpr uint64 kOEpsilons = 0x0000000004000000ULL;
-// FST has no output epsilons.
-constexpr uint64 kNoOEpsilons = 0x0000000008000000ULL;
-
-// ilabels sorted wrt < for each state.
-constexpr uint64 kILabelSorted = 0x0000000010000000ULL;
-// ilabels not sorted wrt < for some state.
-constexpr uint64 kNotILabelSorted = 0x0000000020000000ULL;
-
-// olabels sorted wrt < for each state.
-constexpr uint64 kOLabelSorted = 0x0000000040000000ULL;
-// olabels not sorted wrt < for some state.
-constexpr uint64 kNotOLabelSorted = 0x0000000080000000ULL;
-
-// Non-trivial arc or final weights.
-constexpr uint64 kWeighted = 0x0000000100000000ULL;
-// Only trivial arc and final weights.
-constexpr uint64 kUnweighted = 0x0000000200000000ULL;
-
-// FST has cycles.
-constexpr uint64 kCyclic = 0x0000000400000000ULL;
-// FST has no cycles.
-constexpr uint64 kAcyclic = 0x0000000800000000ULL;
-
-// FST has cycles containing the initial state.
-constexpr uint64 kInitialCyclic = 0x0000001000000000ULL;
-// FST has no cycles containing the initial state.
-constexpr uint64 kInitialAcyclic = 0x0000002000000000ULL;
-
-// FST is topologically sorted.
-constexpr uint64 kTopSorted = 0x0000004000000000ULL;
-// FST is not topologically sorted.
-constexpr uint64 kNotTopSorted = 0x0000008000000000ULL;
-
-// All states reachable from the initial state.
-constexpr uint64 kAccessible = 0x0000010000000000ULL;
-// Not all states reachable from the initial state.
-constexpr uint64 kNotAccessible = 0x0000020000000000ULL;
-
-// All states can reach a final state.
-constexpr uint64 kCoAccessible = 0x0000040000000000ULL;
-// Not all states can reach a final state.
-constexpr uint64 kNotCoAccessible = 0x0000080000000000ULL;
-
-// If NumStates() > 0, then state 0 is initial, state NumStates() - 1 is final,
-// there is a transition from each non-final state i to state i + 1, and there
-// are no other transitions.
-constexpr uint64 kString = 0x0000100000000000ULL;
-
-// Not a string FST.
-constexpr uint64 kNotString = 0x0000200000000000ULL;
-
-// FST has least one weighted cycle.
-constexpr uint64 kWeightedCycles = 0x0000400000000000ULL;
-
-// Only unweighted cycles.
-constexpr uint64 kUnweightedCycles = 0x0000800000000000ULL;
-
-// COMPOSITE PROPERTIES
-
-// Properties of an empty machine.
-constexpr uint64 kNullProperties =
-    kAcceptor | kIDeterministic | kODeterministic | kNoEpsilons | kNoIEpsilons |
-    kNoOEpsilons | kILabelSorted | kOLabelSorted | kUnweighted | kAcyclic |
-    kInitialAcyclic | kTopSorted | kAccessible | kCoAccessible | kString |
-    kUnweightedCycles;
-
-// Properties that are preserved when an FST is copied.
-constexpr uint64 kCopyProperties =
-    kError | kAcceptor | kNotAcceptor | kIDeterministic | kNonIDeterministic |
-    kODeterministic | kNonODeterministic | kEpsilons | kNoEpsilons |
-    kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons | kILabelSorted |
-    kNotILabelSorted | kOLabelSorted | kNotOLabelSorted | kWeighted |
-    kUnweighted | kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic |
-    kTopSorted | kNotTopSorted | kAccessible | kNotAccessible | kCoAccessible |
-    kNotCoAccessible | kString | kNotString | kWeightedCycles |
-    kUnweightedCycles;
-
-// Properties that are intrinsic to the FST.
-constexpr uint64 kIntrinsicProperties =
-    kExpanded | kMutable | kAcceptor | kNotAcceptor | kIDeterministic |
-    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
-    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
-    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
-    kWeighted | kUnweighted | kCyclic | kAcyclic | kInitialCyclic |
-    kInitialAcyclic | kTopSorted | kNotTopSorted | kAccessible |
-    kNotAccessible | kCoAccessible | kNotCoAccessible | kString | kNotString |
-    kWeightedCycles | kUnweightedCycles;
-
-// Properties that are (potentially) extrinsic to the FST.
-constexpr uint64 kExtrinsicProperties = kError;
-
-// Properties that are preserved when an FST start state is set.
-constexpr uint64 kSetStartProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
-    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
-    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
-    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
-    kWeighted | kUnweighted | kCyclic | kAcyclic | kTopSorted | kNotTopSorted |
-    kCoAccessible | kNotCoAccessible | kWeightedCycles | kUnweightedCycles;
-
-// Properties that are preserved when an FST final weight is set.
-constexpr uint64 kSetFinalProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
-    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
-    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
-    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
-    kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic | kTopSorted |
-    kNotTopSorted | kAccessible | kNotAccessible | kWeightedCycles |
-    kUnweightedCycles;
-
-// Properties that are preserved when an FST state is added.
-constexpr uint64 kAddStateProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
-    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
-    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
-    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
-    kWeighted | kUnweighted | kCyclic | kAcyclic | kInitialCyclic |
-    kInitialAcyclic | kTopSorted | kNotTopSorted | kNotAccessible |
-    kNotCoAccessible | kNotString | kWeightedCycles | kUnweightedCycles;
-
-// Properties that are preserved when an FST arc is added.
-constexpr uint64 kAddArcProperties =
-    kExpanded | kMutable | kError | kNotAcceptor | kNonIDeterministic |
-    kNonODeterministic | kEpsilons | kIEpsilons | kOEpsilons |
-    kNotILabelSorted | kNotOLabelSorted | kWeighted | kCyclic | kInitialCyclic |
-    kNotTopSorted | kAccessible | kCoAccessible | kWeightedCycles;
-
-// Properties that are preserved when an FST arc is set.
-constexpr uint64 kSetArcProperties = kExpanded | kMutable | kError;
-
-// Properties that are preserved when FST states are deleted.
-constexpr uint64 kDeleteStatesProperties =
-    kExpanded | kMutable | kError | kAcceptor | kIDeterministic |
-    kODeterministic | kNoEpsilons | kNoIEpsilons | kNoOEpsilons |
-    kILabelSorted | kOLabelSorted | kUnweighted | kAcyclic | kInitialAcyclic |
-    kTopSorted | kUnweightedCycles;
-
-// Properties that are preserved when FST arcs are deleted.
-constexpr uint64 kDeleteArcsProperties =
-    kExpanded | kMutable | kError | kAcceptor | kIDeterministic |
-    kODeterministic | kNoEpsilons | kNoIEpsilons | kNoOEpsilons |
-    kILabelSorted | kOLabelSorted | kUnweighted | kAcyclic | kInitialAcyclic |
-    kTopSorted | kNotAccessible | kNotCoAccessible | kUnweightedCycles;
-
-// Properties that are preserved when an FST's states are reordered.
-constexpr uint64 kStateSortProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
-    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
-    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
-    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
-    kWeighted | kUnweighted | kCyclic | kAcyclic | kInitialCyclic |
-    kInitialAcyclic | kAccessible | kNotAccessible | kCoAccessible |
-    kNotCoAccessible | kWeightedCycles | kUnweightedCycles;
-
-// Properties that are preserved when an FST's arcs are reordered.
-constexpr uint64 kArcSortProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
-    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
-    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
-    kWeighted | kUnweighted | kCyclic | kAcyclic | kInitialCyclic |
-    kInitialAcyclic | kTopSorted | kNotTopSorted | kAccessible |
-    kNotAccessible | kCoAccessible | kNotCoAccessible | kString | kNotString |
-    kWeightedCycles | kUnweightedCycles;
-
-// Properties that are preserved when an FST's input labels are changed.
-constexpr uint64 kILabelInvariantProperties =
-    kExpanded | kMutable | kError | kODeterministic | kNonODeterministic |
-    kOEpsilons | kNoOEpsilons | kOLabelSorted | kNotOLabelSorted | kWeighted |
-    kUnweighted | kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic |
-    kTopSorted | kNotTopSorted | kAccessible | kNotAccessible | kCoAccessible |
-    kNotCoAccessible | kString | kNotString | kWeightedCycles |
-    kUnweightedCycles;
-
-// Properties that are preserved when an FST's output labels are changed.
-constexpr uint64 kOLabelInvariantProperties =
-    kExpanded | kMutable | kError | kIDeterministic | kNonIDeterministic |
-    kIEpsilons | kNoIEpsilons | kILabelSorted | kNotILabelSorted | kWeighted |
-    kUnweighted | kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic |
-    kTopSorted | kNotTopSorted | kAccessible | kNotAccessible | kCoAccessible |
-    kNotCoAccessible | kString | kNotString | kWeightedCycles |
-    kUnweightedCycles;
-
-// Properties that are preserved when an FST's weights are changed. This
-// assumes that the set of states that are non-final is not changed.
-constexpr uint64 kWeightInvariantProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
-    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
-    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
-    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
-    kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic | kTopSorted |
-    kNotTopSorted | kAccessible | kNotAccessible | kCoAccessible |
-    kNotCoAccessible | kString | kNotString;
-
-// Properties that are preserved when a superfinal state is added and an FST's
-// final weights are directed to it via new transitions.
-constexpr uint64 kAddSuperFinalProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor |
-    kNonIDeterministic | kNonODeterministic | kEpsilons | kIEpsilons |
-    kOEpsilons | kNotILabelSorted | kNotOLabelSorted | kWeighted | kUnweighted |
-    kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic | kNotTopSorted |
-    kNotAccessible | kCoAccessible | kNotCoAccessible | kNotString |
-    kWeightedCycles | kUnweightedCycles;
-
-// Properties that are preserved when a superfinal state is removed and the
-// epsilon transitions directed to it are made final weights.
-constexpr uint64 kRmSuperFinalProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
-    kODeterministic | kNoEpsilons | kNoIEpsilons | kNoOEpsilons |
-    kILabelSorted | kOLabelSorted | kWeighted | kUnweighted | kCyclic |
-    kAcyclic | kInitialCyclic | kInitialAcyclic | kTopSorted | kAccessible |
-    kCoAccessible | kNotCoAccessible | kString | kWeightedCycles |
-    kUnweightedCycles;
-
-// All binary properties.
-constexpr uint64 kBinaryProperties = 0x0000000000000007ULL;
-
-// All trinary properties.
-constexpr uint64 kTrinaryProperties = 0x0000ffffffff0000ULL;
-
-// COMPUTED PROPERTIES
-
-// 1st bit of trinary properties.
-constexpr uint64 kPosTrinaryProperties = kTrinaryProperties &
-    0x5555555555555555ULL;
-
-// 2nd bit of trinary properties.
-constexpr uint64 kNegTrinaryProperties = kTrinaryProperties &
-    0xaaaaaaaaaaaaaaaaULL;
-
-// All properties.
-constexpr uint64 kFstProperties = kBinaryProperties | kTrinaryProperties;
-
-// PROPERTY FUNCTIONS and STRING NAMES (defined in properties.cc).
-
-// Below are functions for getting property bit vectors when executing
-// mutation operations.
-inline uint64 SetStartProperties(uint64 inprops);
-
-template <typename Weight>
-uint64 SetFinalProperties(uint64 inprops, const Weight &old_weight,
-                          const Weight &new_weight);
-
-inline uint64 AddStateProperties(uint64 inprops);
-
-template <typename A>
-uint64 AddArcProperties(uint64 inprops, typename A::StateId s, const A &arc,
-                        const A *prev_arc);
-
-inline uint64 DeleteStatesProperties(uint64 inprops);
-
-inline uint64 DeleteAllStatesProperties(uint64 inprops, uint64 staticProps);
-
-inline uint64 DeleteArcsProperties(uint64 inprops);
-
-uint64 ClosureProperties(uint64 inprops, bool star, bool delayed = false);
-
-uint64 ComplementProperties(uint64 inprops);
-
-uint64 ComposeProperties(uint64 inprops1, uint64 inprops2);
-
-uint64 ConcatProperties(uint64 inprops1, uint64 inprops2, bool delayed = false);
-
-uint64 DeterminizeProperties(uint64 inprops, bool has_subsequential_label,
-                             bool distinct_psubsequential_labels);
-
-uint64 FactorWeightProperties(uint64 inprops);
-
-uint64 InvertProperties(uint64 inprops);
-
-uint64 ProjectProperties(uint64 inprops, bool project_input);
-
-uint64 RandGenProperties(uint64 inprops, bool weighted);
-
-uint64 RelabelProperties(uint64 inprops);
-
-uint64 ReplaceProperties(const std::vector<uint64> &inprops, size_t root,
-                         bool epsilon_on_call, bool epsilon_on_return,
-                         bool out_epsilon_on_call, bool out_epsilon_on_return,
-                         bool replace_transducer, bool no_empty_fst,
-                         bool all_ilabel_sorted, bool all_olabel_sorted,
-                         bool all_negative_or_dense);
-
-uint64 ReverseProperties(uint64 inprops, bool has_superinitial);
-
-uint64 ReweightProperties(uint64 inprops);
-
-uint64 RmEpsilonProperties(uint64 inprops, bool delayed = false);
-
-uint64 ShortestPathProperties(uint64 props, bool tree = false);
-
-uint64 SynchronizeProperties(uint64 inprops);
-
-uint64 UnionProperties(uint64 inprops1, uint64 inprops2, bool delayed = false);
-
-// Definitions of inlined functions.
-
-uint64 SetStartProperties(uint64 inprops) {
-  auto outprops = inprops & kSetStartProperties;
-  if (inprops & kAcyclic) {
-    outprops |= kInitialAcyclic;
-  }
-  return outprops;
-}
-
-uint64 AddStateProperties(uint64 inprops) {
-  return inprops & kAddStateProperties;
-}
-
-uint64 DeleteStatesProperties(uint64 inprops) {
-  return inprops & kDeleteStatesProperties;
-}
-
-uint64 DeleteAllStatesProperties(uint64 inprops, uint64 staticprops) {
-  const auto outprops = inprops & kError;
-  return outprops | kNullProperties | staticprops;
-}
-
-uint64 DeleteArcsProperties(uint64 inprops) {
-  return inprops & kDeleteArcsProperties;
-}
-
-// Definitions of template functions.
-
-template <typename Weight>
-uint64 SetFinalProperties(uint64 inprops, const Weight &old_weight,
-                          const Weight &new_weight) {
-  auto outprops = inprops;
-  if (old_weight != Weight::Zero() && old_weight != Weight::One()) {
-    outprops &= ~kWeighted;
-  }
-  if (new_weight != Weight::Zero() && new_weight != Weight::One()) {
-    outprops |= kWeighted;
-    outprops &= ~kUnweighted;
-  }
-  outprops &= kSetFinalProperties | kWeighted | kUnweighted;
-  return outprops;
-}
-
-/// Gets the properties for the MutableFst::AddArc method.
-///
-/// \param inprops  the current properties of the FST
-/// \param s        the ID of the state to which an arc is being added.
-/// \param arc      the arc being added to the state with the specified ID
-/// \param prev_arc the previously-added (or "last") arc of state s, or nullptr
-//                  if s currently has no arcs.
-template <typename Arc>
-uint64 AddArcProperties(uint64 inprops, typename Arc::StateId s,
-                        const Arc &arc, const Arc *prev_arc) {
-  using Weight = typename Arc::Weight;
-  auto outprops = inprops;
-  if (arc.ilabel != arc.olabel) {
-    outprops |= kNotAcceptor;
-    outprops &= ~kAcceptor;
-  }
-  if (arc.ilabel == 0) {
-    outprops |= kIEpsilons;
-    outprops &= ~kNoIEpsilons;
-    if (arc.olabel == 0) {
-      outprops |= kEpsilons;
-      outprops &= ~kNoEpsilons;
-    }
-  }
-  if (arc.olabel == 0) {
-    outprops |= kOEpsilons;
-    outprops &= ~kNoOEpsilons;
-  }
-  if (prev_arc) {
-    if (prev_arc->ilabel > arc.ilabel) {
-      outprops |= kNotILabelSorted;
-      outprops &= ~kILabelSorted;
-    }
-    if (prev_arc->olabel > arc.olabel) {
-      outprops |= kNotOLabelSorted;
-      outprops &= ~kOLabelSorted;
-    }
-  }
-  if (arc.weight != Weight::Zero() && arc.weight != Weight::One()) {
-    outprops |= kWeighted;
-    outprops &= ~kUnweighted;
-  }
-  if (arc.nextstate <= s) {
-    outprops |= kNotTopSorted;
-    outprops &= ~kTopSorted;
-  }
-  outprops &= kAddArcProperties | kAcceptor | kNoEpsilons | kNoIEpsilons |
-              kNoOEpsilons | kILabelSorted | kOLabelSorted | kUnweighted |
-              kTopSorted;
-  if (outprops & kTopSorted) {
-    outprops |= kAcyclic | kInitialAcyclic;
-  }
-  return outprops;
-}
-
-extern const char *PropertyNames[];
-
-}  // namespace fst
-
-#endif  // FST_PROPERTIES_H_
diff --git a/projects/llm_framework/include/fst/prune.h b/projects/llm_framework/include/fst/prune.h
deleted file mode 100644
index 9e7c04bd..00000000
--- a/projects/llm_framework/include/fst/prune.h
+++ /dev/null
@@ -1,341 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions implementing pruning.
-
-#ifndef FST_PRUNE_H_
-#define FST_PRUNE_H_
-
-#include <type_traits>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arcfilter.h>
-#include <fst/heap.h>
-#include <fst/shortest-distance.h>
-
-
-namespace fst {
-namespace internal {
-
-template <class StateId, class Weight>
-class PruneCompare {
- public:
-  PruneCompare(const std::vector<Weight> &idistance,
-               const std::vector<Weight> &fdistance)
-      : idistance_(idistance), fdistance_(fdistance) {}
-
-  bool operator()(const StateId x, const StateId y) const {
-    const auto wx = Times(IDistance(x), FDistance(x));
-    const auto wy = Times(IDistance(y), FDistance(y));
-    return less_(wx, wy);
-  }
-
- private:
-  Weight IDistance(const StateId s) const {
-    return s < idistance_.size() ? idistance_[s] : Weight::Zero();
-  }
-
-  Weight FDistance(const StateId s) const {
-    return s < fdistance_.size() ? fdistance_[s] : Weight::Zero();
-  }
-
-  const std::vector<Weight> &idistance_;
-  const std::vector<Weight> &fdistance_;
-  NaturalLess<Weight> less_;
-};
-
-}  // namespace internal
-
-template <class Arc, class ArcFilter>
-struct PruneOptions {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit PruneOptions(const Weight &weight_threshold = Weight::Zero(),
-                        StateId state_threshold = kNoStateId,
-                        ArcFilter filter = ArcFilter(),
-                        std::vector<Weight> *distance = nullptr,
-                        float delta = kDelta, bool threshold_initial = false)
-      : weight_threshold(std::move(weight_threshold)),
-        state_threshold(state_threshold),
-        filter(std::move(filter)),
-        distance(distance),
-        delta(delta),
-        threshold_initial(threshold_initial) {}
-
-  // Pruning weight threshold.
-  Weight weight_threshold;
-  // Pruning state threshold.
-  StateId state_threshold;
-  // Arc filter.
-  ArcFilter filter;
-  // If non-zero, passes in pre-computed shortest distance to final states.
-  const std::vector<Weight> *distance;
-  // Determines the degree of convergence required when computing shortest
-  // distances.
-  float delta;
-  // Determines if the shortest path weight is left (true) or right
-  // (false) multiplied by the threshold to get the limit for
-  // keeping a state or arc (matters if the semiring is not
-  // commutative).
-  bool threshold_initial;
-};
-
-// Pruning algorithm: this version modifies its input and it takes an options
-// class as an argument. After pruning the FST contains states and arcs that
-// belong to a successful path in the FST whose weight is no more than the
-// weight of the shortest path Times() the provided weight threshold. When the
-// state threshold is not kNoStateId, the output FST is further restricted to
-// have no more than the number of states in opts.state_threshold. Weights must
-// have the path property. The weight of any cycle needs to be bounded; i.e.,
-//
-//   Plus(weight, Weight::One()) == Weight::One()
-template <class Arc, class ArcFilter,
-          typename std::enable_if<IsPath<typename Arc::Weight>::value>::type * =
-              nullptr>
-void Prune(MutableFst<Arc> *fst, const PruneOptions<Arc, ArcFilter> &opts =
-                                     PruneOptions<Arc, ArcFilter>()) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using StateHeap = Heap<StateId, internal::PruneCompare<StateId, Weight>>;
-  auto ns = fst->NumStates();
-  if (ns < 1) return;
-  std::vector<Weight> idistance(ns, Weight::Zero());
-  std::vector<Weight> tmp;
-  if (!opts.distance) {
-    tmp.reserve(ns);
-    ShortestDistance(*fst, &tmp, true, opts.delta);
-  }
-  const auto *fdistance = opts.distance ? opts.distance : &tmp;
-  if ((opts.state_threshold == 0) || (fdistance->size() <= fst->Start()) ||
-      ((*fdistance)[fst->Start()] == Weight::Zero())) {
-    fst->DeleteStates();
-    return;
-  }
-  internal::PruneCompare<StateId, Weight> compare(idistance, *fdistance);
-  StateHeap heap(compare);
-  std::vector<bool> visited(ns, false);
-  std::vector<size_t> enqueued(ns, StateHeap::kNoKey);
-  std::vector<StateId> dead;
-  dead.push_back(fst->AddState());
-  NaturalLess<Weight> less;
-  auto s = fst->Start();
-  const auto limit = opts.threshold_initial ?
-      Times(opts.weight_threshold, (*fdistance)[s]) :
-      Times((*fdistance)[s], opts.weight_threshold);
-  StateId num_visited = 0;
-
-  if (!less(limit, (*fdistance)[s])) {
-    idistance[s] = Weight::One();
-    enqueued[s] = heap.Insert(s);
-    ++num_visited;
-  }
-  while (!heap.Empty()) {
-    s = heap.Top();
-    heap.Pop();
-    enqueued[s] = StateHeap::kNoKey;
-    visited[s] = true;
-    if (less(limit, Times(idistance[s], fst->Final(s)))) {
-      fst->SetFinal(s, Weight::Zero());
-    }
-    for (MutableArcIterator<MutableFst<Arc>> aiter(fst, s); !aiter.Done();
-         aiter.Next()) {
-      auto arc = aiter.Value();  // Copy intended.
-      if (!opts.filter(arc)) continue;
-      const auto weight = Times(Times(idistance[s], arc.weight),
-                                arc.nextstate < fdistance->size() ?
-                                (*fdistance)[arc.nextstate] : Weight::Zero());
-      if (less(limit, weight)) {
-        arc.nextstate = dead[0];
-        aiter.SetValue(arc);
-        continue;
-      }
-      if (less(Times(idistance[s], arc.weight), idistance[arc.nextstate])) {
-        idistance[arc.nextstate] = Times(idistance[s], arc.weight);
-      }
-      if (visited[arc.nextstate]) continue;
-      if ((opts.state_threshold != kNoStateId) &&
-          (num_visited >= opts.state_threshold)) {
-        continue;
-      }
-      if (enqueued[arc.nextstate] == StateHeap::kNoKey) {
-        enqueued[arc.nextstate] = heap.Insert(arc.nextstate);
-        ++num_visited;
-      } else {
-        heap.Update(enqueued[arc.nextstate], arc.nextstate);
-      }
-    }
-  }
-  for (StateId i = 0; i < visited.size(); ++i) {
-    if (!visited[i]) dead.push_back(i);
-  }
-  fst->DeleteStates(dead);
-}
-
-template <class Arc, class ArcFilter,
-          typename std::enable_if<!IsPath<typename Arc::Weight>::value>::type
-              * = nullptr>
-void Prune(MutableFst<Arc> *fst, const PruneOptions<Arc, ArcFilter> &opts =
-                                     PruneOptions<Arc, ArcFilter>()) {
-  FSTERROR() << "Prune: Weight needs to have the path property: "
-             << Arc::Weight::Type();
-  fst->SetProperties(kError, kError);
-}
-
-// Pruning algorithm: this version modifies its input and takes the
-// pruning threshold as an argument. It deletes states and arcs in the
-// FST that do not belong to a successful path whose weight is more
-// than the weight of the shortest path Times() the provided weight
-// threshold. When the state threshold is not kNoStateId, the output
-// FST is further restricted to have no more than the number of states
-// in opts.state_threshold. Weights must have the path property. The
-// weight of any cycle needs to be bounded; i.e.,
-//
-//   Plus(weight, Weight::One()) == Weight::One()
-template <class Arc>
-void Prune(MutableFst<Arc> *fst, typename Arc::Weight weight_threshold,
-           typename Arc::StateId state_threshold = kNoStateId,
-           float delta = kDelta) {
-  const PruneOptions<Arc, AnyArcFilter<Arc>> opts(
-      weight_threshold, state_threshold, AnyArcFilter<Arc>(), nullptr, delta);
-  Prune(fst, opts);
-}
-
-// Pruning algorithm: this version writes the pruned input FST to an
-// output MutableFst and it takes an options class as an argument. The
-// output FST contains states and arcs that belong to a successful
-// path in the input FST whose weight is more than the weight of the
-// shortest path Times() the provided weight threshold. When the state
-// threshold is not kNoStateId, the output FST is further restricted
-// to have no more than the number of states in
-// opts.state_threshold. Weights have the path property.  The weight
-// of any cycle needs to be bounded; i.e.,
-//
-//   Plus(weight, Weight::One()) == Weight::One()
-template <class Arc, class ArcFilter,
-          typename std::enable_if<IsPath<typename Arc::Weight>::value>::type * =
-              nullptr>
-void Prune(
-    const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-    const PruneOptions<Arc, ArcFilter> &opts = PruneOptions<Arc, ArcFilter>()) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using StateHeap = Heap<StateId, internal::PruneCompare<StateId, Weight>>;
-  ofst->DeleteStates();
-  ofst->SetInputSymbols(ifst.InputSymbols());
-  ofst->SetOutputSymbols(ifst.OutputSymbols());
-  if (ifst.Start() == kNoStateId) return;
-  NaturalLess<Weight> less;
-  if (less(opts.weight_threshold, Weight::One()) ||
-      (opts.state_threshold == 0)) {
-    return;
-  }
-  std::vector<Weight> idistance;
-  std::vector<Weight> tmp;
-  if (!opts.distance) ShortestDistance(ifst, &tmp, true, opts.delta);
-  const auto *fdistance = opts.distance ? opts.distance : &tmp;
-  if ((fdistance->size() <= ifst.Start()) ||
-      ((*fdistance)[ifst.Start()] == Weight::Zero())) {
-    return;
-  }
-  internal::PruneCompare<StateId, Weight> compare(idistance, *fdistance);
-  StateHeap heap(compare);
-  std::vector<StateId> copy;
-  std::vector<size_t> enqueued;
-  std::vector<bool> visited;
-  auto s = ifst.Start();
-  const auto limit = opts.threshold_initial ?
-      Times(opts.weight_threshold, (*fdistance)[s]) :
-      Times((*fdistance)[s], opts.weight_threshold);
-  while (copy.size() <= s) copy.push_back(kNoStateId);
-  copy[s] = ofst->AddState();
-  ofst->SetStart(copy[s]);
-  while (idistance.size() <= s) idistance.push_back(Weight::Zero());
-  idistance[s] = Weight::One();
-  while (enqueued.size() <= s) {
-    enqueued.push_back(StateHeap::kNoKey);
-    visited.push_back(false);
-  }
-  enqueued[s] = heap.Insert(s);
-  while (!heap.Empty()) {
-    s = heap.Top();
-    heap.Pop();
-    enqueued[s] = StateHeap::kNoKey;
-    visited[s] = true;
-    if (!less(limit, Times(idistance[s], ifst.Final(s)))) {
-      ofst->SetFinal(copy[s], ifst.Final(s));
-    }
-    for (ArcIterator<Fst<Arc>> aiter(ifst, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      if (!opts.filter(arc)) continue;
-      const auto weight = Times(Times(idistance[s], arc.weight),
-                                arc.nextstate < fdistance->size() ?
-                                (*fdistance)[arc.nextstate] : Weight::Zero());
-      if (less(limit, weight)) continue;
-      if ((opts.state_threshold != kNoStateId) &&
-          (ofst->NumStates() >= opts.state_threshold)) {
-        continue;
-      }
-      while (idistance.size() <= arc.nextstate) {
-        idistance.push_back(Weight::Zero());
-      }
-      if (less(Times(idistance[s], arc.weight), idistance[arc.nextstate])) {
-        idistance[arc.nextstate] = Times(idistance[s], arc.weight);
-      }
-      while (copy.size() <= arc.nextstate) copy.push_back(kNoStateId);
-      if (copy[arc.nextstate] == kNoStateId) {
-        copy[arc.nextstate] = ofst->AddState();
-      }
-      ofst->AddArc(copy[s], Arc(arc.ilabel, arc.olabel, arc.weight,
-                                copy[arc.nextstate]));
-      while (enqueued.size() <= arc.nextstate) {
-        enqueued.push_back(StateHeap::kNoKey);
-        visited.push_back(false);
-      }
-      if (visited[arc.nextstate]) continue;
-      if (enqueued[arc.nextstate] == StateHeap::kNoKey) {
-        enqueued[arc.nextstate] = heap.Insert(arc.nextstate);
-      } else {
-        heap.Update(enqueued[arc.nextstate], arc.nextstate);
-      }
-    }
-  }
-}
-
-template <class Arc, class ArcFilter,
-          typename std::enable_if<!IsPath<typename Arc::Weight>::value>::type
-              * = nullptr>
-void Prune(const Fst<Arc> &, MutableFst<Arc> *ofst,
-           const PruneOptions<Arc, ArcFilter> &) {
-  FSTERROR() << "Prune: Weight needs to have the path property: "
-             << Arc::Weight::Type();
-  ofst->SetProperties(kError, kError);
-}
-
-// Pruning algorithm: this version writes the pruned input FST to an
-// output MutableFst and simply takes the pruning threshold as an
-// argument. The output FST contains states and arcs that belong to a
-// successful path in the input FST whose weight is no more than the
-// weight of the shortest path Times() the provided weight
-// threshold. When the state threshold is not kNoStateId, the output
-// FST is further restricted to have no more than the number of states
-// in opts.state_threshold. Weights must have the path property. The
-// weight of any cycle needs to be bounded; i.e.,
-//
-// Plus(weight, Weight::One()) = Weight::One();
-template <class Arc>
-void Prune(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-           typename Arc::Weight weight_threshold,
-           typename Arc::StateId state_threshold = kNoStateId,
-           float delta = kDelta) {
-  const PruneOptions<Arc, AnyArcFilter<Arc>> opts(
-      weight_threshold, state_threshold, AnyArcFilter<Arc>(), nullptr, delta);
-  Prune(ifst, ofst, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_PRUNE_H_
diff --git a/projects/llm_framework/include/fst/push.h b/projects/llm_framework/include/fst/push.h
deleted file mode 100644
index 1f772739..00000000
--- a/projects/llm_framework/include/fst/push.h
+++ /dev/null
@@ -1,155 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to reweight/push an FST, and utility functions to weigh and reweight
-// an FST.
-
-#ifndef FST_PUSH_H_
-#define FST_PUSH_H_
-
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arc-map.h>
-#include <fst/factor-weight.h>
-#include <fst/fst.h>
-#include <fst/reweight.h>
-#include <fst/shortest-distance.h>
-
-
-namespace fst {
-
-// Computes the total weight (sum of the weights of all accepting paths) from
-// the output of ShortestDistance, using the shortest distance from the final
-// state when reverse is true and from the initial state otherwise.
-template <class Arc>
-typename Arc::Weight ComputeTotalWeight(
-    const Fst<Arc> &fst, const std::vector<typename Arc::Weight> &distance,
-    bool reverse) {
-  if (reverse) {
-    return fst.Start() < distance.size() ? distance[fst.Start()]
-                                         : Arc::Weight::Zero();
-  }
-  auto sum = Arc::Weight::Zero();
-  for (typename Arc::StateId s = 0; s < distance.size(); ++s) {
-    sum = Plus(sum, Times(distance[s], fst.Final(s)));
-  }
-  return sum;
-}
-
-// Divides the weight of every accepting path by a fixed weight. This weight
-// is also divided at the final state if at_final is true and at the initial
-// state otherwise.
-template <class Arc>
-void RemoveWeight(MutableFst<Arc> *fst, const typename Arc::Weight &weight,
-                  bool at_final) {
-  using Weight = typename Arc::Weight;
-  if ((weight == Weight::One()) || (weight == Weight::Zero())) return;
-  if (at_final) {
-    for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
-         siter.Next()) {
-      fst->SetFinal(siter.Value(),
-                    Divide(fst->Final(siter.Value()), weight, DIVIDE_RIGHT));
-    }
-  } else {
-    const auto start = fst->Start();
-    for (MutableArcIterator<MutableFst<Arc>> aiter(fst, start); !aiter.Done();
-         aiter.Next()) {
-      auto arc = aiter.Value();
-      arc.weight = Divide(arc.weight, weight, DIVIDE_LEFT);
-      aiter.SetValue(arc);
-    }
-    fst->SetFinal(start, Divide(fst->Final(start), weight, DIVIDE_LEFT));
-  }
-}
-
-// Pushes the weights in FST in the direction defined by TYPE. If
-// pushing towards the initial state, the sum of the weight of the
-// outgoing transitions and final weight at a non-initial state is
-// equal to One() in the resulting machine. If pushing towards the
-// final state, the same property holds on the reverse machine.
-//
-// Weight needs to be left distributive when pushing towards the
-// initial state and right distributive when pushing towards the final
-// states.
-template <class Arc>
-void Push(MutableFst<Arc> *fst, ReweightType type, float delta = kDelta,
-          bool remove_total_weight = false) {
-  using Weight = typename Arc::Weight;
-  std::vector<Weight> distance;
-  ShortestDistance(*fst, &distance, type == REWEIGHT_TO_INITIAL, delta);
-  auto total_weight = Weight::One();
-  if (remove_total_weight) {
-    total_weight =
-        ComputeTotalWeight(*fst, distance, type == REWEIGHT_TO_INITIAL);
-  }
-  Reweight(fst, distance, type);
-  if (remove_total_weight) {
-    RemoveWeight(fst, total_weight, type == REWEIGHT_TO_FINAL);
-  }
-}
-
-constexpr uint32 kPushWeights = 0x0001;
-constexpr uint32 kPushLabels = 0x0002;
-constexpr uint32 kPushRemoveTotalWeight = 0x0004;
-constexpr uint32 kPushRemoveCommonAffix = 0x0008;
-
-// Pushes the weights and/or labels of the input FST into the output
-// mutable FST by pushing weights and/or labels (as determined by the
-// ptype argument) towards the initial state or final states (as
-// determined by the rtype template parameter). The weight type must
-// be left distributive when pushing weights towards the initial state, and
-// right distribution when pushing weights towards the final states.
-template <class Arc, ReweightType rtype>
-void Push(const Fst<Arc> &ifst, MutableFst<Arc> *ofst, uint32 ptype,
-          float delta = kDelta) {
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-  if ((ptype & (kPushWeights | kPushLabels)) == kPushWeights) {
-    *ofst = ifst;
-    Push(ofst, rtype, delta, ptype & kPushRemoveTotalWeight);
-  } else if (ptype & kPushLabels) {
-    const auto gtype =
-        rtype == REWEIGHT_TO_INITIAL ? GALLIC_LEFT : GALLIC_RIGHT;
-    using GallicWeight = typename GallicArc<Arc, gtype>::Weight;
-    std::vector<GallicWeight> gdistance;
-    VectorFst<GallicArc<Arc, gtype>> gfst;
-    ArcMap(ifst, &gfst, ToGallicMapper<Arc, gtype>());
-    if (ptype & kPushWeights) {
-      ShortestDistance(gfst, &gdistance, rtype == REWEIGHT_TO_INITIAL, delta);
-    } else {
-      ArcMapFst<Arc, Arc, RmWeightMapper<Arc>> uwfst(ifst,
-                                                      RmWeightMapper<Arc>());
-      ArcMapFst<Arc, GallicArc<Arc, gtype>, ToGallicMapper<Arc, gtype>> guwfst(
-          uwfst, ToGallicMapper<Arc, gtype>());
-      ShortestDistance(guwfst, &gdistance, rtype == REWEIGHT_TO_INITIAL, delta);
-    }
-    auto total_weight = GallicWeight::One();
-    if (ptype & (kPushRemoveTotalWeight | kPushRemoveCommonAffix)) {
-      total_weight =
-          ComputeTotalWeight(gfst, gdistance, rtype == REWEIGHT_TO_INITIAL);
-      total_weight = GallicWeight(
-          ptype & kPushRemoveCommonAffix
-              ? total_weight.Value1()
-              : StringWeight<Label, GallicStringType(gtype)>::One(),
-          ptype & kPushRemoveTotalWeight ? total_weight.Value2()
-                                         : Weight::One());
-    }
-    Reweight(&gfst, gdistance, rtype);
-    if (ptype & (kPushRemoveTotalWeight | kPushRemoveCommonAffix)) {
-      RemoveWeight(&gfst, total_weight, rtype == REWEIGHT_TO_FINAL);
-    }
-    FactorWeightFst<GallicArc<Arc, gtype>, GallicFactor<Label, Weight, gtype>>
-        fwfst(gfst);
-    ArcMap(fwfst, ofst, FromGallicMapper<Arc, gtype>());
-    ofst->SetOutputSymbols(ifst.OutputSymbols());
-  } else {
-    LOG(WARNING) << "Push: pushing type is set to 0, so not pushing";
-    *ofst = ifst;
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_PUSH_H_
diff --git a/projects/llm_framework/include/fst/queue.h b/projects/llm_framework/include/fst/queue.h
deleted file mode 100644
index f57d176e..00000000
--- a/projects/llm_framework/include/fst/queue.h
+++ /dev/null
@@ -1,948 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes for various FST state queues with a unified interface.
-
-#ifndef FST_QUEUE_H_
-#define FST_QUEUE_H_
-
-#include <deque>
-#include <memory>
-#include <type_traits>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arcfilter.h>
-#include <fst/connect.h>
-#include <fst/heap.h>
-#include <fst/topsort.h>
-#include <fst/weight.h>
-
-
-namespace fst {
-
-// The Queue interface is:
-//
-// template <class S>
-// class Queue {
-//  public:
-//   using StateId = S;
-//
-//   // Constructor: may need args (e.g., FST, comparator) for some queues.
-//   Queue(...) override;
-//
-//   // Returns the head of the queue.
-//   StateId Head() const override;
-//
-//   // Inserts a state.
-//   void Enqueue(StateId s) override;
-//
-//   // Removes the head of the queue.
-//   void Dequeue() override;
-//
-//   // Updates ordering of state s when weight changes, if necessary.
-//   void Update(StateId s) override;
-//
-//   // Is the queue empty?
-//   bool Empty() const override;
-//
-//   // Removes all states from the queue.
-//   void Clear() override;
-// };
-
-// State queue types.
-enum QueueType {
-  TRIVIAL_QUEUE = 0,         // Single state queue.
-  FIFO_QUEUE = 1,            // First-in, first-out queue.
-  LIFO_QUEUE = 2,            // Last-in, first-out queue.
-  SHORTEST_FIRST_QUEUE = 3,  // Shortest-first queue.
-  TOP_ORDER_QUEUE = 4,       // Topologically-ordered queue.
-  STATE_ORDER_QUEUE = 5,     // State ID-ordered queue.
-  SCC_QUEUE = 6,             // Component graph top-ordered meta-queue.
-  AUTO_QUEUE = 7,            // Auto-selected queue.
-  OTHER_QUEUE = 8
-};
-
-// QueueBase, templated on the StateId, is a virtual base class shared by all
-// queues considered by AutoQueue.
-template <class S>
-class QueueBase {
- public:
-  using StateId = S;
-
-  virtual ~QueueBase() {}
-
-  // Concrete implementation.
-
-  explicit QueueBase(QueueType type) : queue_type_(type), error_(false) {}
-
-  void SetError(bool error) { error_ = error; }
-
-  bool Error() const { return error_; }
-
-  QueueType Type() const { return queue_type_; }
-
-  // Virtual interface.
-
-  virtual StateId Head() const = 0;
-  virtual void Enqueue(StateId) = 0;
-  virtual void Dequeue() = 0;
-  virtual void Update(StateId) = 0;
-  virtual bool Empty() const = 0;
-  virtual void Clear() = 0;
-
- private:
-  QueueType queue_type_;
-  bool error_;
-};
-
-// Trivial queue discipline; one may enqueue at most one state at a time. It
-// can be used for strongly connected components with only one state and no
-// self-loops.
-template <class S>
-class TrivialQueue : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  TrivialQueue() : QueueBase<StateId>(TRIVIAL_QUEUE), front_(kNoStateId) {}
-
-  virtual ~TrivialQueue() = default;
-
-  StateId Head() const final { return front_; }
-
-  void Enqueue(StateId s) final { front_ = s; }
-
-  void Dequeue() final { front_ = kNoStateId; }
-
-  void Update(StateId) final {}
-
-  bool Empty() const final { return front_ == kNoStateId; }
-
-  void Clear() final { front_ = kNoStateId; }
-
- private:
-  StateId front_;
-};
-
-// First-in, first-out queue discipline.
-//
-// This is not a final class.
-template <class S>
-class FifoQueue : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  FifoQueue() : QueueBase<StateId>(FIFO_QUEUE) {}
-
-  virtual ~FifoQueue() = default;
-
-  StateId Head() const override { return queue_.back(); }
-
-  void Enqueue(StateId s) override { queue_.push_front(s); }
-
-  void Dequeue() override { queue_.pop_back(); }
-
-  void Update(StateId) override {}
-
-  bool Empty() const override { return queue_.empty(); }
-
-  void Clear() override { queue_.clear(); }
-
- private:
-  std::deque<StateId> queue_;
-};
-
-// Last-in, first-out queue discipline.
-template <class S>
-class LifoQueue : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  LifoQueue() : QueueBase<StateId>(LIFO_QUEUE) {}
-
-  virtual ~LifoQueue() = default;
-
-  StateId Head() const final { return queue_.front(); }
-
-  void Enqueue(StateId s) final { queue_.push_front(s); }
-
-  void Dequeue() final { queue_.pop_front(); }
-
-  void Update(StateId) final {}
-
-  bool Empty() const final { return queue_.empty(); }
-
-  void Clear() final { queue_.clear(); }
-
- private:
-  std::deque<StateId> queue_;
-};
-
-// Shortest-first queue discipline, templated on the StateId and as well as a
-// comparison functor used to compare two StateIds. If a (single) state's order
-// changes, it can be reordered in the queue with a call to Update(). If update
-// is false, call to Update() does not reorder the queue.
-//
-// This is not a final class.
-template <typename S, typename Compare, bool update = true>
-class ShortestFirstQueue : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  explicit ShortestFirstQueue(Compare comp)
-      : QueueBase<StateId>(SHORTEST_FIRST_QUEUE), heap_(comp) {}
-
-  virtual ~ShortestFirstQueue() = default;
-
-  StateId Head() const override { return heap_.Top(); }
-
-  void Enqueue(StateId s) override {
-    if (update) {
-      for (StateId i = key_.size(); i <= s; ++i) key_.push_back(kNoStateId);
-      key_[s] = heap_.Insert(s);
-    } else {
-      heap_.Insert(s);
-    }
-  }
-
-  void Dequeue() override {
-    if (update) {
-      key_[heap_.Pop()] = kNoStateId;
-    } else {
-      heap_.Pop();
-    }
-  }
-
-  void Update(StateId s) override {
-    if (!update) return;
-    if (s >= key_.size() || key_[s] == kNoStateId) {
-      Enqueue(s);
-    } else {
-      heap_.Update(key_[s], s);
-    }
-  }
-
-  bool Empty() const override { return heap_.Empty(); }
-
-  void Clear() override {
-    heap_.Clear();
-    if (update) key_.clear();
-  }
-
-  const Compare &GetCompare() const { return heap_.GetCompare(); }
-
- private:
-  Heap<StateId, Compare> heap_;
-  std::vector<ssize_t> key_;
-};
-
-namespace internal {
-
-// Given a vector that maps from states to weights, and a comparison functor
-// for weights, this class defines a comparison function object between states.
-template <typename StateId, typename Less>
-class StateWeightCompare {
- public:
-  using Weight = typename Less::Weight;
-
-  StateWeightCompare(const std::vector<Weight> &weights, const Less &less)
-      : weights_(weights), less_(less) {}
-
-  bool operator()(const StateId s1, const StateId s2) const {
-    return less_(weights_[s1], weights_[s2]);
-  }
-
- private:
-  // Borrowed references.
-  const std::vector<Weight> &weights_;
-  const Less &less_;
-};
-
-}  // namespace internal
-
-// Shortest-first queue discipline, templated on the StateId and Weight, is
-// specialized to use the weight's natural order for the comparison function.
-template <typename S, typename Weight>
-class NaturalShortestFirstQueue
-    : public ShortestFirstQueue<
-          S, internal::StateWeightCompare<S, NaturalLess<Weight>>> {
- public:
-  using StateId = S;
-  using Compare = internal::StateWeightCompare<StateId, NaturalLess<Weight>>;
-
-  explicit NaturalShortestFirstQueue(const std::vector<Weight> &distance)
-      : ShortestFirstQueue<StateId, Compare>(Compare(distance, less_)) {}
-
-  virtual ~NaturalShortestFirstQueue() = default;
-
- private:
-  // This is non-static because the constructor for non-idempotent weights will
-  // result in an error.
-  const NaturalLess<Weight> less_{};
-};
-
-// In a shortest path computation on a lattice-like FST, we may keep many old
-// nonviable paths as a part of the search. Since the search process always
-// expands the lowest cost path next, that lowest cost path may be a very old
-// nonviable path instead of one we expect to lead to a shortest path.
-//
-// For instance, suppose that the current best path in an alignment has
-// traversed 500 arcs with a cost of 10. We may also have a bad path in
-// the queue that has traversed only 40 arcs but also has a cost of 10.
-// This path is very unlikely to lead to a reasonable alignment, so this queue
-// can prune it from the search space.
-//
-// This queue relies on the caller using a shortest-first exploration order
-// like this:
-//   while (true) {
-//     StateId head = queue.Head();
-//     queue.Dequeue();
-//     for (const auto& arc : GetArcs(fst, head)) {
-//       queue.Enqueue(arc.nextstate);
-//     }
-//   }
-// We use this assumption to guess that there is an arc between Head and the
-// Enqueued state; this is how the number of path steps is measured.
-template <typename S, typename Weight>
-class PruneNaturalShortestFirstQueue
-    : public NaturalShortestFirstQueue<S, Weight> {
- public:
-  using StateId = S;
-  using Base = NaturalShortestFirstQueue<StateId, Weight>;
-
-  explicit PruneNaturalShortestFirstQueue(const std::vector<Weight> &distance,
-                                          int threshold)
-      : Base(distance),
-        threshold_(threshold),
-        head_steps_(0),
-        max_head_steps_(0) {}
-
-  ~PruneNaturalShortestFirstQueue() override = default;
-
-  StateId Head() const override {
-    const auto head = Base::Head();
-    // Stores the number of steps from the start of the graph to this state
-    // along the shortest-weight path.
-    if (head < steps_.size()) {
-      max_head_steps_ = std::max(steps_[head], max_head_steps_);
-      head_steps_ = steps_[head];
-    }
-    return head;
-  }
-
-  void Enqueue(StateId s) override {
-    // We assume that there is an arc between the Head() state and this
-    // Enqueued state.
-    const ssize_t state_steps = head_steps_ + 1;
-    if (s >= steps_.size()) {
-      steps_.resize(s + 1, state_steps);
-    }
-    // This is the number of arcs in the minimum cost path from Start to s.
-    steps_[s] = state_steps;
-    if (state_steps > (max_head_steps_ - threshold_) || threshold_ < 0) {
-      Base::Enqueue(s);
-    }
-  }
-
- private:
-  // A dense map from StateId to the number of arcs in the minimum weight
-  // path from Start to this state.
-  std::vector<ssize_t> steps_;
-  // We only keep paths that are within this number of arcs (not weight!)
-  // of the longest path.
-  const ssize_t threshold_;
-
-  // The following are mutable because Head() is const.
-  // The number of arcs traversed in the minimum cost path from the start
-  // state to the current Head() state.
-  mutable ssize_t head_steps_;
-  // The maximum number of arcs traversed by any low-cost path so far.
-  mutable ssize_t max_head_steps_;
-};
-
-// Topological-order queue discipline, templated on the StateId. States are
-// ordered in the queue topologically. The FST must be acyclic.
-template <class S>
-class TopOrderQueue : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  // This constructor computes the topological order. It accepts an arc filter
-  // to limit the transitions considered in that computation (e.g., only the
-  // epsilon graph).
-  template <class Arc, class ArcFilter>
-  TopOrderQueue(const Fst<Arc> &fst, ArcFilter filter)
-      : QueueBase<StateId>(TOP_ORDER_QUEUE),
-        front_(0),
-        back_(kNoStateId),
-        order_(0),
-        state_(0) {
-    bool acyclic;
-    TopOrderVisitor<Arc> top_order_visitor(&order_, &acyclic);
-    DfsVisit(fst, &top_order_visitor, filter);
-    if (!acyclic) {
-      FSTERROR() << "TopOrderQueue: FST is not acyclic";
-      QueueBase<S>::SetError(true);
-    }
-    state_.resize(order_.size(), kNoStateId);
-  }
-
-  // This constructor is passed the pre-computed topological order.
-  explicit TopOrderQueue(const std::vector<StateId> &order)
-      : QueueBase<StateId>(TOP_ORDER_QUEUE),
-        front_(0),
-        back_(kNoStateId),
-        order_(order),
-        state_(order.size(), kNoStateId) {}
-
-  virtual ~TopOrderQueue() = default;
-
-  StateId Head() const final { return state_[front_]; }
-
-  void Enqueue(StateId s) final {
-    if (front_ > back_) {
-      front_ = back_ = order_[s];
-    } else if (order_[s] > back_) {
-      back_ = order_[s];
-    } else if (order_[s] < front_) {
-      front_ = order_[s];
-    }
-    state_[order_[s]] = s;
-  }
-
-  void Dequeue() final {
-    state_[front_] = kNoStateId;
-    while ((front_ <= back_) && (state_[front_] == kNoStateId)) ++front_;
-  }
-
-  void Update(StateId) final {}
-
-  bool Empty() const final { return front_ > back_; }
-
-  void Clear() final {
-    for (StateId s = front_; s <= back_; ++s) state_[s] = kNoStateId;
-    back_ = kNoStateId;
-    front_ = 0;
-  }
-
- private:
-  StateId front_;
-  StateId back_;
-  std::vector<StateId> order_;
-  std::vector<StateId> state_;
-};
-
-// State order queue discipline, templated on the StateId. States are ordered in
-// the queue by state ID.
-template <class S>
-class StateOrderQueue : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  StateOrderQueue()
-      : QueueBase<StateId>(STATE_ORDER_QUEUE), front_(0), back_(kNoStateId) {}
-
-  virtual ~StateOrderQueue() = default;
-
-  StateId Head() const final { return front_; }
-
-  void Enqueue(StateId s) final {
-    if (front_ > back_) {
-      front_ = back_ = s;
-    } else if (s > back_) {
-      back_ = s;
-    } else if (s < front_) {
-      front_ = s;
-    }
-    while (enqueued_.size() <= s) enqueued_.push_back(false);
-    enqueued_[s] = true;
-  }
-
-  void Dequeue() final {
-    enqueued_[front_] = false;
-    while ((front_ <= back_) && (enqueued_[front_] == false)) ++front_;
-  }
-
-  void Update(StateId) final {}
-
-  bool Empty() const final { return front_ > back_; }
-
-  void Clear() final {
-    for (StateId i = front_; i <= back_; ++i) enqueued_[i] = false;
-    front_ = 0;
-    back_ = kNoStateId;
-  }
-
- private:
-  StateId front_;
-  StateId back_;
-  std::vector<bool> enqueued_;
-};
-
-// SCC topological-order meta-queue discipline, templated on the StateId and a
-// queue used inside each SCC. It visits the SCCs of an FST in topological
-// order. Its constructor is passed the queues to to use within an SCC.
-template <class S, class Queue>
-class SccQueue : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  // Constructor takes a vector specifying the SCC number per state and a
-  // vector giving the queue to use per SCC number.
-  SccQueue(const std::vector<StateId> &scc,
-           std::vector<std::unique_ptr<Queue>> *queue)
-      : QueueBase<StateId>(SCC_QUEUE),
-        queue_(queue),
-        scc_(scc),
-        front_(0),
-        back_(kNoStateId) {}
-
-  virtual ~SccQueue() = default;
-
-  StateId Head() const final {
-    while ((front_ <= back_) &&
-           (((*queue_)[front_] && (*queue_)[front_]->Empty()) ||
-            (((*queue_)[front_] == nullptr) &&
-             ((front_ >= trivial_queue_.size()) ||
-              (trivial_queue_[front_] == kNoStateId))))) {
-      ++front_;
-    }
-    if ((*queue_)[front_]) {
-      return (*queue_)[front_]->Head();
-    } else {
-      return trivial_queue_[front_];
-    }
-  }
-
-  void Enqueue(StateId s) final {
-    if (front_ > back_) {
-      front_ = back_ = scc_[s];
-    } else if (scc_[s] > back_) {
-      back_ = scc_[s];
-    } else if (scc_[s] < front_) {
-      front_ = scc_[s];
-    }
-    if ((*queue_)[scc_[s]]) {
-      (*queue_)[scc_[s]]->Enqueue(s);
-    } else {
-      while (trivial_queue_.size() <= scc_[s]) {
-        trivial_queue_.push_back(kNoStateId);
-      }
-      trivial_queue_[scc_[s]] = s;
-    }
-  }
-
-  void Dequeue() final {
-    if ((*queue_)[front_]) {
-      (*queue_)[front_]->Dequeue();
-    } else if (front_ < trivial_queue_.size()) {
-      trivial_queue_[front_] = kNoStateId;
-    }
-  }
-
-  void Update(StateId s) final {
-    if ((*queue_)[scc_[s]]) (*queue_)[scc_[s]]->Update(s);
-  }
-
-  bool Empty() const final {
-    // Queues SCC number back_ is not empty unless back_ == front_.
-    if (front_ < back_) {
-      return false;
-    } else if (front_ > back_) {
-      return true;
-    } else if ((*queue_)[front_]) {
-      return (*queue_)[front_]->Empty();
-    } else {
-      return (front_ >= trivial_queue_.size()) ||
-             (trivial_queue_[front_] == kNoStateId);
-    }
-  }
-
-  void Clear() final {
-    for (StateId i = front_; i <= back_; ++i) {
-      if ((*queue_)[i]) {
-        (*queue_)[i]->Clear();
-      } else if (i < trivial_queue_.size()) {
-        trivial_queue_[i] = kNoStateId;
-      }
-    }
-    front_ = 0;
-    back_ = kNoStateId;
-  }
-
- private:
-  std::vector<std::unique_ptr<Queue>> *queue_;
-  const std::vector<StateId> &scc_;
-  mutable StateId front_;
-  StateId back_;
-  std::vector<StateId> trivial_queue_;
-};
-
-// Automatic queue discipline. It selects a queue discipline for a given FST
-// based on its properties.
-template <class S>
-class AutoQueue : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  // This constructor takes a state distance vector that, if non-null and if
-  // the Weight type has the path property, will entertain the shortest-first
-  // queue using the natural order w.r.t to the distance.
-  template <class Arc, class ArcFilter>
-  AutoQueue(const Fst<Arc> &fst,
-            const std::vector<typename Arc::Weight> *distance, ArcFilter filter)
-      : QueueBase<StateId>(AUTO_QUEUE) {
-    using Weight = typename Arc::Weight;
-    using Less = NaturalLess<Weight>;
-    using Compare = internal::StateWeightCompare<StateId, Less>;
-    // First checks if the FST is known to have these properties.
-    const auto props =
-        fst.Properties(kAcyclic | kCyclic | kTopSorted | kUnweighted, false);
-    if ((props & kTopSorted) || fst.Start() == kNoStateId) {
-      queue_.reset(new StateOrderQueue<StateId>());
-      VLOG(2) << "AutoQueue: using state-order discipline";
-    } else if (props & kAcyclic) {
-      queue_.reset(new TopOrderQueue<StateId>(fst, filter));
-      VLOG(2) << "AutoQueue: using top-order discipline";
-    } else if ((props & kUnweighted) && (Weight::Properties() & kIdempotent)) {
-      queue_.reset(new LifoQueue<StateId>());
-      VLOG(2) << "AutoQueue: using LIFO discipline";
-    } else {
-      uint64 properties;
-      // Decomposes into strongly-connected components.
-      SccVisitor<Arc> scc_visitor(&scc_, nullptr, nullptr, &properties);
-      DfsVisit(fst, &scc_visitor, filter);
-      auto nscc = *std::max_element(scc_.begin(), scc_.end()) + 1;
-      std::vector<QueueType> queue_types(nscc);
-      std::unique_ptr<Less> less;
-      std::unique_ptr<Compare> comp;
-      if (distance && (Weight::Properties() & kPath) == kPath) {
-        less.reset(new Less);
-        comp.reset(new Compare(*distance, *less));
-      }
-      // Finds the queue type to use per SCC.
-      bool unweighted;
-      bool all_trivial;
-      SccQueueType(fst, scc_, &queue_types, filter, less.get(), &all_trivial,
-                   &unweighted);
-      // If unweighted and semiring is idempotent, uses LIFO queue.
-      if (unweighted) {
-        queue_.reset(new LifoQueue<StateId>());
-        VLOG(2) << "AutoQueue: using LIFO discipline";
-        return;
-      }
-      // If all the SCC are trivial, the FST is acyclic and the scc number gives
-      // the topological order.
-      if (all_trivial) {
-        queue_.reset(new TopOrderQueue<StateId>(scc_));
-        VLOG(2) << "AutoQueue: using top-order discipline";
-        return;
-      }
-      VLOG(2) << "AutoQueue: using SCC meta-discipline";
-      queues_.resize(nscc);
-      for (StateId i = 0; i < nscc; ++i) {
-        switch (queue_types[i]) {
-          case TRIVIAL_QUEUE:
-            queues_[i].reset();
-            VLOG(3) << "AutoQueue: SCC #" << i << ": using trivial discipline";
-            break;
-          case SHORTEST_FIRST_QUEUE:
-            queues_[i].reset(
-                new ShortestFirstQueue<StateId, Compare, false>(*comp));
-            VLOG(3) << "AutoQueue: SCC #" << i
-                    << ": using shortest-first discipline";
-            break;
-          case LIFO_QUEUE:
-            queues_[i].reset(new LifoQueue<StateId>());
-            VLOG(3) << "AutoQueue: SCC #" << i << ": using LIFO discipline";
-            break;
-          case FIFO_QUEUE:
-          default:
-            queues_[i].reset(new FifoQueue<StateId>());
-            VLOG(3) << "AutoQueue: SCC #" << i << ": using FIFO discipine";
-            break;
-        }
-      }
-      queue_.reset(new SccQueue<StateId, QueueBase<StateId>>(scc_, &queues_));
-    }
-  }
-
-  virtual ~AutoQueue() = default;
-
-  StateId Head() const final { return queue_->Head(); }
-
-  void Enqueue(StateId s) final { queue_->Enqueue(s); }
-
-  void Dequeue() final { queue_->Dequeue(); }
-
-  void Update(StateId s) final { queue_->Update(s); }
-
-  bool Empty() const final { return queue_->Empty(); }
-
-  void Clear() final { queue_->Clear(); }
-
- private:
-  template <class Arc, class ArcFilter, class Less>
-  static void SccQueueType(const Fst<Arc> &fst, const std::vector<StateId> &scc,
-                           std::vector<QueueType> *queue_types,
-                           ArcFilter filter, Less *less, bool *all_trivial,
-                           bool *unweighted);
-
-  std::unique_ptr<QueueBase<StateId>> queue_;
-  std::vector<std::unique_ptr<QueueBase<StateId>>> queues_;
-  std::vector<StateId> scc_;
-};
-
-// Examines the states in an FST's strongly connected components and determines
-// which type of queue to use per SCC. Stores result as a vector of QueueTypes
-// which is assumed to have length equal to the number of SCCs. An arc filter
-// is used to limit the transitions considered (e.g., only the epsilon graph).
-// The argument all_trivial is set to true if every queue is the trivial queue.
-// The argument unweighted is set to true if the semiring is idempotent and all
-// the arc weights are equal to Zero() or One().
-template <class StateId>
-template <class Arc, class ArcFilter, class Less>
-void AutoQueue<StateId>::SccQueueType(const Fst<Arc> &fst,
-                                      const std::vector<StateId> &scc,
-                                      std::vector<QueueType> *queue_type,
-                                      ArcFilter filter, Less *less,
-                                      bool *all_trivial, bool *unweighted) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  *all_trivial = true;
-  *unweighted = true;
-  for (StateId i = 0; i < queue_type->size(); ++i) {
-    (*queue_type)[i] = TRIVIAL_QUEUE;
-  }
-  for (StateIterator<Fst<Arc>> sit(fst); !sit.Done(); sit.Next()) {
-    const auto state = sit.Value();
-    for (ArcIterator<Fst<Arc>> ait(fst, state); !ait.Done(); ait.Next()) {
-      const auto &arc = ait.Value();
-      if (!filter(arc)) continue;
-      if (scc[state] == scc[arc.nextstate]) {
-        auto &type = (*queue_type)[scc[state]];
-        if (!less || ((*less)(arc.weight, Weight::One()))) {
-          type = FIFO_QUEUE;
-        } else if ((type == TRIVIAL_QUEUE) || (type == LIFO_QUEUE)) {
-          if (!(Weight::Properties() & kIdempotent) ||
-              (arc.weight != Weight::Zero() && arc.weight != Weight::One())) {
-            type = SHORTEST_FIRST_QUEUE;
-          } else {
-            type = LIFO_QUEUE;
-          }
-        }
-        if (type != TRIVIAL_QUEUE) *all_trivial = false;
-      }
-      if (!(Weight::Properties() & kIdempotent) ||
-          (arc.weight != Weight::Zero() && arc.weight != Weight::One())) {
-        *unweighted = false;
-      }
-    }
-  }
-}
-
-// An A* estimate is a function object that maps from a state ID to an
-// estimate of the shortest distance to the final states.
-
-// A trivial A* estimate, yielding a queue which behaves the same in Dijkstra's
-// algorithm.
-template <typename StateId, typename Weight>
-struct TrivialAStarEstimate {
-  constexpr Weight operator()(StateId) const { return Weight::One(); }
-};
-
-// A non-trivial A* estimate using a vector of the estimated future costs.
-template <typename StateId, typename Weight>
-class NaturalAStarEstimate {
- public:
-  NaturalAStarEstimate(const std::vector<Weight> &beta) : beta_(beta) {}
-
-  const Weight &operator()(StateId s) const {
-     return (s < beta_.size()) ? beta_[s] : kZero;
-  }
-
- private:
-  static constexpr Weight kZero = Weight::Zero();
-
-  const std::vector<Weight> &beta_;
-};
-
-template <typename Arc, typename Weight>
-constexpr Weight NaturalAStarEstimate<Arc, Weight>::kZero;
-
-// Given a vector that maps from states to weights representing the shortest
-// distance from the initial state, a comparison function object between
-// weights, and an estimate of the shortest distance to the final states, this
-// class defines a comparison function object between states.
-template <typename S, typename Less, typename Estimate>
-class AStarWeightCompare {
- public:
-  using StateId = S;
-  using Weight = typename Less::Weight;
-
-  AStarWeightCompare(const std::vector<Weight> &weights, const Less &less,
-                     const Estimate &estimate)
-      : weights_(weights), less_(less), estimate_(estimate) {}
-
-  bool operator()(StateId s1, StateId s2) const {
-    const auto w1 = Times(weights_[s1], estimate_(s1));
-    const auto w2 = Times(weights_[s2], estimate_(s2));
-    return less_(w1, w2);
-  }
-
-  const Estimate &GetEstimate() const { return estimate_; }
-
- private:
-  const std::vector<Weight> &weights_;
-  const Less &less_;
-  const Estimate &estimate_;
-};
-
-// A* queue discipline templated on StateId, Weight, and Estimate.
-template <typename S, typename Weight, typename Estimate>
-class NaturalAStarQueue : public ShortestFirstQueue<
-          S, AStarWeightCompare<S, NaturalLess<Weight>, Estimate>> {
- public:
-  using StateId = S;
-  using Compare = AStarWeightCompare<StateId, NaturalLess<Weight>, Estimate>;
-
-  NaturalAStarQueue(const std::vector<Weight> &distance,
-                    const Estimate &estimate)
-      : ShortestFirstQueue<StateId, Compare>(
-            Compare(distance, less_, estimate)) {}
-
-  ~NaturalAStarQueue() = default;
-
- private:
-  // This is non-static because the constructor for non-idempotent weights will
-  // result in an error.
-  const NaturalLess<Weight> less_{};
-};
-
-// A state equivalence class is a function object that maps from a state ID to
-// an equivalence class (state) ID. The trivial equivalence class maps a state
-// ID to itself.
-template <typename StateId>
-struct TrivialStateEquivClass {
-  StateId operator()(StateId s) const { return s; }
-};
-
-// Distance-based pruning queue discipline: Enqueues a state only when its
-// shortest distance (so far), as specified by distance, is less than (as
-// specified by comp) the shortest distance Times() the threshold to any state
-// in the same equivalence class, as specified by the functor class_func. The
-// underlying queue discipline is specified by queue. The ownership of queue is
-// given to this class.
-//
-// This is not a final class.
-template <typename Queue, typename Less, typename ClassFnc>
-class PruneQueue : public QueueBase<typename Queue::StateId> {
- public:
-  using StateId = typename Queue::StateId;
-  using Weight = typename Less::Weight;
-
-  PruneQueue(const std::vector<Weight> &distance, Queue *queue,
-             const Less &less, const ClassFnc &class_fnc, Weight threshold)
-      : QueueBase<StateId>(OTHER_QUEUE),
-        distance_(distance),
-        queue_(queue),
-        less_(less),
-        class_fnc_(class_fnc),
-        threshold_(std::move(threshold)) {}
-
-  virtual ~PruneQueue() = default;
-
-  StateId Head() const override { return queue_->Head(); }
-
-  void Enqueue(StateId s) override {
-    const auto c = class_fnc_(s);
-    if (c >= class_distance_.size()) {
-      class_distance_.resize(c + 1, Weight::Zero());
-    }
-    if (less_(distance_[s], class_distance_[c])) {
-      class_distance_[c] = distance_[s];
-    }
-    // Enqueues only if below threshold limit.
-    const auto limit = Times(class_distance_[c], threshold_);
-    if (less_(distance_[s], limit)) queue_->Enqueue(s);
-  }
-
-  void Dequeue() override { queue_->Dequeue(); }
-
-  void Update(StateId s) override {
-    const auto c = class_fnc_(s);
-    if (less_(distance_[s], class_distance_[c])) {
-      class_distance_[c] = distance_[s];
-    }
-    queue_->Update(s);
-  }
-
-  bool Empty() const override { return queue_->Empty(); }
-
-  void Clear() override { queue_->Clear(); }
-
- private:
-  const std::vector<Weight> &distance_;  // Shortest distance to state.
-  std::unique_ptr<Queue> queue_;
-  const Less &less_;                    // Borrowed reference.
-  const ClassFnc &class_fnc_;           // Equivalence class functor.
-  Weight threshold_;                    // Pruning weight threshold.
-  std::vector<Weight> class_distance_;  // Shortest distance to class.
-};
-
-// Pruning queue discipline (see above) using the weight's natural order for the
-// comparison function. The ownership of the queue argument is given to this
-// class.
-template <typename Queue, typename Weight, typename ClassFnc>
-class NaturalPruneQueue final
-    : public PruneQueue<Queue, NaturalLess<Weight>, ClassFnc> {
- public:
-  using StateId = typename Queue::StateId;
-
-  NaturalPruneQueue(const std::vector<Weight> &distance, Queue *queue,
-                    const ClassFnc &class_fnc, Weight threshold)
-      : PruneQueue<Queue, NaturalLess<Weight>, ClassFnc>(
-            distance, queue, NaturalLess<Weight>(), class_fnc, threshold) {}
-
-  virtual ~NaturalPruneQueue() = default;
-};
-
-// Filter-based pruning queue discipline: enqueues a state only if allowed by
-// the filter, specified by the state filter functor argument. The underlying
-// queue discipline is specified by the queue argument. The ownership of the
-// queue is given to this class.
-template <typename Queue, typename Filter>
-class FilterQueue : public QueueBase<typename Queue::StateId> {
- public:
-  using StateId = typename Queue::StateId;
-
-  FilterQueue(Queue *queue, const Filter &filter)
-      : QueueBase<StateId>(OTHER_QUEUE), queue_(queue), filter_(filter) {}
-
-  virtual ~FilterQueue() = default;
-
-  StateId Head() const final { return queue_->Head(); }
-
-  // Enqueues only if allowed by state filter.
-  void Enqueue(StateId s) final {
-    if (filter_(s)) queue_->Enqueue(s);
-  }
-
-  void Dequeue() final { queue_->Dequeue(); }
-
-  void Update(StateId s) final {}
-
-  bool Empty() const final { return queue_->Empty(); }
-
-  void Clear() final { queue_->Clear(); }
-
- private:
-  std::unique_ptr<Queue> queue_;
-  const Filter &filter_;
-};
-
-}  // namespace fst
-
-#endif  // FST_QUEUE_H_
diff --git a/projects/llm_framework/include/fst/randequivalent.h b/projects/llm_framework/include/fst/randequivalent.h
deleted file mode 100644
index 73108b46..00000000
--- a/projects/llm_framework/include/fst/randequivalent.h
+++ /dev/null
@@ -1,114 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Tests if two FSTS are equivalent by checking if random strings from one FST
-// are transduced the same by both FSTs.
-
-#ifndef FST_RANDEQUIVALENT_H_
-#define FST_RANDEQUIVALENT_H_
-
-#include <fst/log.h>
-
-#include <fst/arcsort.h>
-#include <fst/compose.h>
-#include <fst/project.h>
-#include <fst/randgen.h>
-#include <fst/shortest-distance.h>
-#include <fst/vector-fst.h>
-
-
-namespace fst {
-
-// Test if two FSTs are stochastically equivalent by randomly generating
-// random paths through the FSTs.
-//
-// For each randomly generated path, the algorithm computes for each
-// of the two FSTs the sum of the weights of all the successful paths
-// sharing the same input and output labels as the considered randomly
-// generated path and checks that these two values are within a user-specified
-// delta. Returns optional error value (when FLAGS_error_fatal = false).
-template <class Arc, class ArcSelector>
-bool RandEquivalent(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-                    int32 num_paths, float delta,
-                    const RandGenOptions<ArcSelector> &opts,
-                    bool *error = nullptr) {
-  using Weight = typename Arc::Weight;
-  if (error) *error = false;
-  // Checks that the symbol table are compatible.
-  if (!CompatSymbols(fst1.InputSymbols(), fst2.InputSymbols()) ||
-      !CompatSymbols(fst1.OutputSymbols(), fst2.OutputSymbols())) {
-    FSTERROR() << "RandEquivalent: Input/output symbol tables of 1st "
-               << "argument do not match input/output symbol tables of 2nd "
-               << "argument";
-    if (error) *error = true;
-    return false;
-  }
-  static const ILabelCompare<Arc> icomp;
-  static const OLabelCompare<Arc> ocomp;
-  VectorFst<Arc> sfst1(fst1);
-  VectorFst<Arc> sfst2(fst2);
-  Connect(&sfst1);
-  Connect(&sfst2);
-  ArcSort(&sfst1, icomp);
-  ArcSort(&sfst2, icomp);
-  bool result = true;
-  for (int32 n = 0; n < num_paths; ++n) {
-    VectorFst<Arc> path;
-    const auto &fst = rand() % 2 ? sfst1 : sfst2;  // NOLINT
-    RandGen(fst, &path, opts);
-    VectorFst<Arc> ipath(path);
-    VectorFst<Arc> opath(path);
-    Project(&ipath, PROJECT_INPUT);
-    Project(&opath, PROJECT_OUTPUT);
-    VectorFst<Arc> cfst1, pfst1;
-    Compose(ipath, sfst1, &cfst1);
-    ArcSort(&cfst1, ocomp);
-    Compose(cfst1, opath, &pfst1);
-    // Gives up if there are epsilon cycles in a non-idempotent semiring.
-    if (!(Weight::Properties() & kIdempotent) &&
-        pfst1.Properties(kCyclic, true)) {
-      continue;
-    }
-    const auto sum1 = ShortestDistance(pfst1);
-    VectorFst<Arc> cfst2;
-    Compose(ipath, sfst2, &cfst2);
-    ArcSort(&cfst2, ocomp);
-    VectorFst<Arc> pfst2;
-    Compose(cfst2, opath, &pfst2);
-    // Gives up if there are epsilon cycles in a non-idempotent semiring.
-    if (!(Weight::Properties() & kIdempotent) &&
-        pfst2.Properties(kCyclic, true)) {
-      continue;
-    }
-    const auto sum2 = ShortestDistance(pfst2);
-    if (!ApproxEqual(sum1, sum2, delta)) {
-      VLOG(1) << "Sum1 = " << sum1;
-      VLOG(1) << "Sum2 = " << sum2;
-      result = false;
-      break;
-    }
-  }
-  if (fst1.Properties(kError, false) || fst2.Properties(kError, false)) {
-    if (error) *error = true;
-    return false;
-  }
-  return result;
-}
-
-// Tests if two FSTs are equivalent by randomly generating a nnum_paths paths
-// (no longer than the path_length) using a user-specified seed, optionally
-// indicating an error setting an optional error argument to true.
-template <class Arc>
-bool RandEquivalent(const Fst<Arc> &fst1, const Fst<Arc> &fst2, int32 num_paths,
-                    float delta = kDelta, time_t seed = time(nullptr),
-                    int32 max_length = std::numeric_limits<int32>::max(),
-                    bool *error = nullptr) {
-  const UniformArcSelector<Arc> uniform_selector(seed);
-  const RandGenOptions<UniformArcSelector<Arc>> opts(uniform_selector,
-                                                     max_length);
-  return RandEquivalent(fst1, fst2, num_paths, delta, opts, error);
-}
-
-}  // namespace fst
-
-#endif  // FST_RANDEQUIVALENT_H_
diff --git a/projects/llm_framework/include/fst/randgen.h b/projects/llm_framework/include/fst/randgen.h
deleted file mode 100644
index 5bcd9fd0..00000000
--- a/projects/llm_framework/include/fst/randgen.h
+++ /dev/null
@@ -1,756 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes and functions to generate random paths through an FST.
-
-#ifndef FST_RANDGEN_H_
-#define FST_RANDGEN_H_
-
-#include <math.h>
-#include <stddef.h>
-#include <limits>
-#include <map>
-#include <memory>
-#include <random>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/accumulator.h>
-#include <fst/cache.h>
-#include <fst/dfs-visit.h>
-#include <fst/float-weight.h>
-#include <fst/fst-decl.h>
-#include <fst/fst.h>
-#include <fst/mutable-fst.h>
-#include <fst/properties.h>
-#include <fst/util.h>
-#include <fst/weight.h>
-
-namespace fst {
-
-// The RandGenFst class is roughly similar to ArcMapFst in that it takes two
-// template parameters denoting the input and output arc types. However, it also
-// takes an additional template parameter which specifies a sampler object which
-// samples (with replacement) arcs from an FST state. The sampler in turn takes
-// a template parameter for a selector object which actually chooses the arc.
-//
-// Arc selector functors are used to select a random transition given an FST
-// state s, returning a number N such that 0 <= N <= NumArcs(s). If N is
-// NumArcs(s), then the final weight is selected; otherwise the N-th arc is
-// selected. It is assumed these are not applied to any state which is neither
-// final nor has any arcs leaving it.
-
-// Randomly selects a transition using the uniform distribution. This class is
-// not thread-safe.
-template <class Arc>
-class UniformArcSelector {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // Constructs a selector with a non-deterministic seed.
-  UniformArcSelector() : rand_(std::random_device()()) {}
-  // Constructs a selector with a given seed.
-  explicit UniformArcSelector(uint64 seed) : rand_(seed) {}
-
-  size_t operator()(const Fst<Arc> &fst, StateId s) const {
-    const auto n = fst.NumArcs(s) + (fst.Final(s) != Weight::Zero());
-    return static_cast<size_t>(
-        std::uniform_int_distribution<>(0, n - 1)(rand_));
-  }
-
- private:
-  mutable std::mt19937_64 rand_;
-};
-
-// Randomly selects a transition w.r.t. the weights treated as negative log
-// probabilities after normalizing for the total weight leaving the state. Zero
-// transitions are disregarded. It assumed that Arc::Weight::Value() accesses
-// the floating point representation of the weight. This class is not
-// thread-safe.
-template <class Arc>
-class LogProbArcSelector {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // Constructs a selector with a non-deterministic seed.
-  LogProbArcSelector() : seed_(std::random_device()()), rand_(seed_) {}
-  // Constructs a selector with a given seed.
-  explicit LogProbArcSelector(uint64 seed) : seed_(seed), rand_(seed) {}
-
-  size_t operator()(const Fst<Arc> &fst, StateId s) const {
-    // Finds total weight leaving state.
-    auto sum = Log64Weight::Zero();
-    ArcIterator<Fst<Arc>> aiter(fst, s);
-    for (; !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      sum = Plus(sum, to_log_weight_(arc.weight));
-    }
-    sum = Plus(sum, to_log_weight_(fst.Final(s)));
-    const double threshold =
-        std::uniform_real_distribution<>(0, exp(-sum.Value()))(rand_);
-    auto p = Log64Weight::Zero();
-    size_t n = 0;
-    for (aiter.Reset(); !aiter.Done(); aiter.Next(), ++n) {
-      p = Plus(p, to_log_weight_(aiter.Value().weight));
-      if (exp(-p.Value()) > threshold) return n;
-    }
-    return n;
-  }
-
-  uint64 Seed() const { return seed_; }
-
- protected:
-  Log64Weight ToLogWeight(const Weight &weight) const {
-    return to_log_weight_(weight);
-  }
-
-  std::mt19937_64 &MutableRand() const { return rand_; }
-
- private:
-  const uint64 seed_;
-  mutable std::mt19937_64 rand_;
-  const WeightConvert<Weight, Log64Weight> to_log_weight_{};
-};
-
-// Useful alias when using StdArc.
-using StdArcSelector = LogProbArcSelector<StdArc>;
-
-// Same as LogProbArcSelector but use CacheLogAccumulator to cache the weight
-// accumulation computations. This class is not thread-safe.
-template <class Arc>
-class FastLogProbArcSelector : public LogProbArcSelector<Arc> {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using LogProbArcSelector<Arc>::MutableRand;
-  using LogProbArcSelector<Arc>::ToLogWeight;
-  using LogProbArcSelector<Arc>::operator();
-
-  // Constructs a selector with a non-deterministic seed.
-  FastLogProbArcSelector() : LogProbArcSelector<Arc>() {}
-  // Constructs a selector with a given seed.
-  explicit FastLogProbArcSelector(uint64 seed) : LogProbArcSelector<Arc>(
-      seed) {}
-
-  size_t operator()(const Fst<Arc> &fst, StateId s,
-                    CacheLogAccumulator<Arc> *accumulator) const {
-    accumulator->SetState(s);
-    ArcIterator<Fst<Arc>> aiter(fst, s);
-    // Finds total weight leaving state.
-    const double sum =
-        ToLogWeight(accumulator->Sum(fst.Final(s), &aiter, 0, fst.NumArcs(s)))
-            .Value();
-    const double r = -log(std::uniform_real_distribution<>(0, 1)(
-        MutableRand()));
-    Weight w = from_log_weight_(r + sum);
-    aiter.Reset();
-    return accumulator->LowerBound(w, &aiter);
-  }
-
- private:
-  const WeightConvert<Log64Weight, Weight> from_log_weight_{};
-};
-
-// Random path state info maintained by RandGenFst and passed to samplers.
-template <typename Arc>
-struct RandState {
-  using StateId = typename Arc::StateId;
-
-  StateId state_id;  // Current input FST state.
-  size_t nsamples;   // Number of samples to be sampled at this state.
-  size_t length;     // Length of path to this random state.
-  size_t select;     // Previous sample arc selection.
-  const RandState<Arc> *parent;  // Previous random state on this path.
-
-  explicit RandState(StateId state_id, size_t nsamples = 0, size_t length = 0,
-                     size_t select = 0, const RandState<Arc> *parent = nullptr)
-      : state_id(state_id),
-        nsamples(nsamples),
-        length(length),
-        select(select),
-        parent(parent) {}
-
-  RandState() : RandState(kNoStateId) {}
-};
-
-// This class, given an arc selector, samples, with replacement, multiple random
-// transitions from an FST's state. This is a generic version with a
-// straightforward use of the arc selector. Specializations may be defined for
-// arc selectors for greater efficiency or special behavior.
-template <class Arc, class Selector>
-class ArcSampler {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // The max_length argument may be interpreted (or ignored) by a selector as
-  // it chooses. This generic version interprets this literally.
-  ArcSampler(const Fst<Arc> &fst, const Selector &selector,
-             int32 max_length = std::numeric_limits<int32>::max())
-      : fst_(fst), selector_(selector), max_length_(max_length) {}
-
-  // Allow updating FST argument; pass only if changed.
-  ArcSampler(const ArcSampler<Arc, Selector> &sampler,
-             const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? *fst : sampler.fst_),
-        selector_(sampler.selector_),
-        max_length_(sampler.max_length_) {
-    Reset();
-  }
-
-  // Samples a fixed number of samples from the given state. The length argument
-  // specifies the length of the path to the state. Returns true if the samples
-  // were collected. No samples may be collected if either there are no
-  // transitions leaving the state and the state is non-final, or if the path
-  // length has been exceeded. Iterator members are provided to read the samples
-  // in the order in which they were collected.
-  bool Sample(const RandState<Arc> &rstate) {
-    sample_map_.clear();
-    if ((fst_.NumArcs(rstate.state_id) == 0 &&
-         fst_.Final(rstate.state_id) == Weight::Zero()) ||
-        rstate.length == max_length_) {
-      Reset();
-      return false;
-    }
-    for (size_t i = 0; i < rstate.nsamples; ++i) {
-      ++sample_map_[selector_(fst_, rstate.state_id)];
-    }
-    Reset();
-    return true;
-  }
-
-  // More samples?
-  bool Done() const { return sample_iter_ == sample_map_.end(); }
-
-  // Gets the next sample.
-  void Next() { ++sample_iter_; }
-
-  std::pair<size_t, size_t> Value() const { return *sample_iter_; }
-
-  void Reset() { sample_iter_ = sample_map_.begin(); }
-
-  bool Error() const { return false; }
-
- private:
-  const Fst<Arc> &fst_;
-  const Selector &selector_;
-  const int32 max_length_;
-
-  // Stores (N, K) as described for Value().
-  std::map<size_t, size_t> sample_map_;
-  std::map<size_t, size_t>::const_iterator sample_iter_;
-
-  ArcSampler<Arc, Selector> &operator=(const ArcSampler &) = delete;
-};
-
-// Samples one sample of num_to_sample dimensions from a multinomial
-// distribution parameterized by a vector of probabilities. The result
-// container should be pre-initialized (e.g., an empty map or a zeroed vector
-// sized the same as the vector of probabilities.
-// probs.size()).
-template <class Result, class RNG>
-void OneMultinomialSample(const std::vector<double> &probs,
-                          size_t num_to_sample, Result *result, RNG *rng) {
-  // Left-over probability mass.
-  double norm = 0;
-  for (double p : probs) norm += p;
-  // Left-over number of samples needed.
-  for (size_t i = 0; i < probs.size(); ++i) {
-    size_t num_sampled = 0;
-    if (probs[i] > 0) {
-      std::binomial_distribution<> d(num_to_sample, probs[i] / norm);
-      num_sampled = d(*rng);
-    }
-    if (num_sampled != 0) (*result)[i] = num_sampled;
-    norm -= probs[i];
-    num_to_sample -= num_sampled;
-  }
-}
-
-// Specialization for FastLogProbArcSelector.
-template <class Arc>
-class ArcSampler<Arc, FastLogProbArcSelector<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Accumulator = CacheLogAccumulator<Arc>;
-  using Selector = FastLogProbArcSelector<Arc>;
-
-  ArcSampler(const Fst<Arc> &fst, const Selector &selector,
-             int32 max_length = std::numeric_limits<int32>::max())
-      : fst_(fst),
-        selector_(selector),
-        max_length_(max_length),
-        accumulator_(new Accumulator()) {
-    accumulator_->Init(fst);
-    rng_.seed(selector_.Seed());
-  }
-
-  ArcSampler(const ArcSampler<Arc, Selector> &sampler,
-             const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? *fst : sampler.fst_),
-        selector_(sampler.selector_),
-        max_length_(sampler.max_length_) {
-    if (fst) {
-      accumulator_.reset(new Accumulator());
-      accumulator_->Init(*fst);
-    } else {  // Shallow copy.
-      accumulator_.reset(new Accumulator(*sampler.accumulator_));
-    }
-  }
-
-  bool Sample(const RandState<Arc> &rstate) {
-    sample_map_.clear();
-    if ((fst_.NumArcs(rstate.state_id) == 0 &&
-         fst_.Final(rstate.state_id) == Weight::Zero()) ||
-        rstate.length == max_length_) {
-      Reset();
-      return false;
-    }
-    if (fst_.NumArcs(rstate.state_id) + 1 < rstate.nsamples) {
-      MultinomialSample(rstate);
-      Reset();
-      return true;
-    }
-    for (size_t i = 0; i < rstate.nsamples; ++i) {
-      ++sample_map_[selector_(fst_, rstate.state_id, accumulator_.get())];
-    }
-    Reset();
-    return true;
-  }
-
-  bool Done() const { return sample_iter_ == sample_map_.end(); }
-
-  void Next() { ++sample_iter_; }
-
-  std::pair<size_t, size_t> Value() const { return *sample_iter_; }
-
-  void Reset() { sample_iter_ = sample_map_.begin(); }
-
-  bool Error() const { return accumulator_->Error(); }
-
- private:
-  using RNG = std::mt19937;
-
-  // Sample according to the multinomial distribution of rstate.nsamples draws
-  // from p_.
-  void MultinomialSample(const RandState<Arc> &rstate) {
-    p_.clear();
-    for (ArcIterator<Fst<Arc>> aiter(fst_, rstate.state_id); !aiter.Done();
-         aiter.Next()) {
-      p_.push_back(exp(-to_log_weight_(aiter.Value().weight).Value()));
-    }
-    if (fst_.Final(rstate.state_id) != Weight::Zero()) {
-      p_.push_back(exp(-to_log_weight_(fst_.Final(rstate.state_id)).Value()));
-    }
-    if (rstate.nsamples < std::numeric_limits<RNG::result_type>::max()) {
-      OneMultinomialSample(p_, rstate.nsamples, &sample_map_, &rng_);
-    } else {
-      for (size_t i = 0; i < p_.size(); ++i) {
-        sample_map_[i] = ceil(p_[i] * rstate.nsamples);
-      }
-    }
-  }
-
-  const Fst<Arc> &fst_;
-  const Selector &selector_;
-  const int32 max_length_;
-
-  // Stores (N, K) for Value().
-  std::map<size_t, size_t> sample_map_;
-  std::map<size_t, size_t>::const_iterator sample_iter_;
-
-  std::unique_ptr<Accumulator> accumulator_;
-  RNG rng_;                // Random number generator.
-  std::vector<double> p_;  // Multinomial parameters.
-  const WeightConvert<Weight, Log64Weight> to_log_weight_{};
-};
-
-// Options for random path generation with RandGenFst. The template argument is
-// a sampler, typically the class ArcSampler. Ownership of the sampler is taken
-// by RandGenFst.
-template <class Sampler>
-struct RandGenFstOptions : public CacheOptions {
-  Sampler *sampler;          // How to sample transitions at a state.
-  int32 npath;               // Number of paths to generate.
-  bool weighted;             // Is the output tree weighted by path count, or
-                             // is it just an unweighted DAG?
-  bool remove_total_weight;  // Remove total weight when output is weighted.
-
-  RandGenFstOptions(const CacheOptions &opts, Sampler *sampler, int32 npath = 1,
-                    bool weighted = true, bool remove_total_weight = false)
-      : CacheOptions(opts),
-        sampler(sampler),
-        npath(npath),
-        weighted(weighted),
-        remove_total_weight(remove_total_weight) {}
-};
-
-namespace internal {
-
-// Implementation of RandGenFst.
-template <class FromArc, class ToArc, class Sampler>
-class RandGenFstImpl : public CacheImpl<ToArc> {
- public:
-  using FstImpl<ToArc>::SetType;
-  using FstImpl<ToArc>::SetProperties;
-  using FstImpl<ToArc>::SetInputSymbols;
-  using FstImpl<ToArc>::SetOutputSymbols;
-
-  using CacheBaseImpl<CacheState<ToArc>>::EmplaceArc;
-  using CacheBaseImpl<CacheState<ToArc>>::HasArcs;
-  using CacheBaseImpl<CacheState<ToArc>>::HasFinal;
-  using CacheBaseImpl<CacheState<ToArc>>::HasStart;
-  using CacheBaseImpl<CacheState<ToArc>>::SetArcs;
-  using CacheBaseImpl<CacheState<ToArc>>::SetFinal;
-  using CacheBaseImpl<CacheState<ToArc>>::SetStart;
-
-  using Label = typename FromArc::Label;
-  using StateId = typename FromArc::StateId;
-  using FromWeight = typename FromArc::Weight;
-
-  using ToWeight = typename ToArc::Weight;
-
-  RandGenFstImpl(const Fst<FromArc> &fst,
-                 const RandGenFstOptions<Sampler> &opts)
-      : CacheImpl<ToArc>(opts),
-        fst_(fst.Copy()),
-        sampler_(opts.sampler),
-        npath_(opts.npath),
-        weighted_(opts.weighted),
-        remove_total_weight_(opts.remove_total_weight),
-        superfinal_(kNoLabel) {
-    SetType("randgen");
-    SetProperties(
-        RandGenProperties(fst.Properties(kFstProperties, false), weighted_),
-        kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  RandGenFstImpl(const RandGenFstImpl &impl)
-      : CacheImpl<ToArc>(impl),
-        fst_(impl.fst_->Copy(true)),
-        sampler_(new Sampler(*impl.sampler_, fst_.get())),
-        npath_(impl.npath_),
-        weighted_(impl.weighted_),
-        superfinal_(kNoLabel) {
-    SetType("randgen");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      const auto s = fst_->Start();
-      if (s == kNoStateId) return kNoStateId;
-      SetStart(state_table_.size());
-      state_table_.emplace_back(
-          new RandState<FromArc>(s, npath_, 0, 0, nullptr));
-    }
-    return CacheImpl<ToArc>::Start();
-  }
-
-  ToWeight Final(StateId s) {
-    if (!HasFinal(s)) Expand(s);
-    return CacheImpl<ToArc>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<ToArc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<ToArc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<ToArc>::NumOutputEpsilons(s);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) &&
-        (fst_->Properties(kError, false) || sampler_->Error())) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<ToArc>::Properties(mask);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<ToArc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<ToArc>::InitArcIterator(s, data);
-  }
-
-  // Computes the outgoing transitions from a state, creating new destination
-  // states as needed.
-  void Expand(StateId s) {
-    if (s == superfinal_) {
-      SetFinal(s, ToWeight::One());
-      SetArcs(s);
-      return;
-    }
-    SetFinal(s, ToWeight::Zero());
-    const auto &rstate = *state_table_[s];
-    sampler_->Sample(rstate);
-    ArcIterator<Fst<FromArc>> aiter(*fst_, rstate.state_id);
-    const auto narcs = fst_->NumArcs(rstate.state_id);
-    for (; !sampler_->Done(); sampler_->Next()) {
-      const auto &sample_pair = sampler_->Value();
-      const auto pos = sample_pair.first;
-      const auto count = sample_pair.second;
-      double prob = static_cast<double>(count) / rstate.nsamples;
-      if (pos < narcs) {  // Regular transition.
-        aiter.Seek(sample_pair.first);
-        const auto &aarc = aiter.Value();
-        auto weight =
-            weighted_ ? to_weight_(Log64Weight(-log(prob))) : ToWeight::One();
-        EmplaceArc(s, aarc.ilabel, aarc.olabel, std::move(weight),
-                   state_table_.size());
-        auto *nrstate = new RandState<FromArc>(aarc.nextstate, count,
-                                               rstate.length + 1, pos, &rstate);
-        state_table_.emplace_back(nrstate);
-      } else {  // Super-final transition.
-        if (weighted_) {
-          const auto weight =
-              remove_total_weight_
-                  ? to_weight_(Log64Weight(-log(prob)))
-                  : to_weight_(Log64Weight(-log(prob * npath_)));
-          SetFinal(s, weight);
-        } else {
-          if (superfinal_ == kNoLabel) {
-            superfinal_ = state_table_.size();
-            state_table_.emplace_back(
-                new RandState<FromArc>(kNoStateId, 0, 0, 0, nullptr));
-          }
-          for (size_t n = 0; n < count; ++n) {
-            EmplaceArc(s, 0, 0, ToWeight::One(), superfinal_);
-          }
-        }
-      }
-    }
-    SetArcs(s);
-  }
-
- private:
-  const std::unique_ptr<Fst<FromArc>> fst_;
-  std::unique_ptr<Sampler> sampler_;
-  const int32 npath_;
-  std::vector<std::unique_ptr<RandState<FromArc>>> state_table_;
-  const bool weighted_;
-  bool remove_total_weight_;
-  StateId superfinal_;
-  const WeightConvert<Log64Weight, ToWeight> to_weight_{};
-};
-
-}  // namespace internal
-
-// FST class to randomly generate paths through an FST, with details controlled
-// by RandGenOptionsFst. Output format is a tree weighted by the path count.
-template <class FromArc, class ToArc, class Sampler>
-class RandGenFst
-    : public ImplToFst<internal::RandGenFstImpl<FromArc, ToArc, Sampler>> {
- public:
-  using Label = typename FromArc::Label;
-  using StateId = typename FromArc::StateId;
-  using Weight = typename FromArc::Weight;
-
-  using Store = DefaultCacheStore<FromArc>;
-  using State = typename Store::State;
-
-  using Impl = internal::RandGenFstImpl<FromArc, ToArc, Sampler>;
-
-  friend class ArcIterator<RandGenFst<FromArc, ToArc, Sampler>>;
-  friend class StateIterator<RandGenFst<FromArc, ToArc, Sampler>>;
-
-  RandGenFst(const Fst<FromArc> &fst, const RandGenFstOptions<Sampler> &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  RandGenFst(const RandGenFst<FromArc, ToArc, Sampler> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this RandGenFst. See Fst<>::Copy() for further doc.
-  RandGenFst<FromArc, ToArc, Sampler> *Copy(bool safe = false) const override {
-    return new RandGenFst<FromArc, ToArc, Sampler>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<ToArc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<ToArc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  RandGenFst &operator=(const RandGenFst &) = delete;
-};
-
-// Specialization for RandGenFst.
-template <class FromArc, class ToArc, class Sampler>
-class StateIterator<RandGenFst<FromArc, ToArc, Sampler>>
-    : public CacheStateIterator<RandGenFst<FromArc, ToArc, Sampler>> {
- public:
-  explicit StateIterator(const RandGenFst<FromArc, ToArc, Sampler> &fst)
-      : CacheStateIterator<RandGenFst<FromArc, ToArc, Sampler>>(
-            fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for RandGenFst.
-template <class FromArc, class ToArc, class Sampler>
-class ArcIterator<RandGenFst<FromArc, ToArc, Sampler>>
-    : public CacheArcIterator<RandGenFst<FromArc, ToArc, Sampler>> {
- public:
-  using StateId = typename FromArc::StateId;
-
-  ArcIterator(const RandGenFst<FromArc, ToArc, Sampler> &fst, StateId s)
-      : CacheArcIterator<RandGenFst<FromArc, ToArc, Sampler>>(
-            fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class FromArc, class ToArc, class Sampler>
-inline void RandGenFst<FromArc, ToArc, Sampler>::InitStateIterator(
-    StateIteratorData<ToArc> *data) const {
-  data->base = new StateIterator<RandGenFst<FromArc, ToArc, Sampler>>(*this);
-}
-
-// Options for random path generation.
-template <class Selector>
-struct RandGenOptions {
-  const Selector &selector;  // How an arc is selected at a state.
-  int32 max_length;          // Maximum path length.
-  int32 npath;               // Number of paths to generate.
-  bool weighted;             // Is the output tree weighted by path count, or
-                             // is it just an unweighted DAG?
-  bool remove_total_weight;  // Remove total weight when output is weighted?
-
-  explicit RandGenOptions(const Selector &selector,
-                          int32 max_length = std::numeric_limits<int32>::max(),
-                          int32 npath = 1, bool weighted = false,
-                          bool remove_total_weight = false)
-      : selector(selector),
-        max_length(max_length),
-        npath(npath),
-        weighted(weighted),
-        remove_total_weight(remove_total_weight) {}
-};
-
-namespace internal {
-
-template <class FromArc, class ToArc>
-class RandGenVisitor {
- public:
-  using StateId = typename FromArc::StateId;
-  using Weight = typename FromArc::Weight;
-
-  explicit RandGenVisitor(MutableFst<ToArc> *ofst) : ofst_(ofst) {}
-
-  void InitVisit(const Fst<FromArc> &ifst) {
-    ifst_ = &ifst;
-    ofst_->DeleteStates();
-    ofst_->SetInputSymbols(ifst.InputSymbols());
-    ofst_->SetOutputSymbols(ifst.OutputSymbols());
-    if (ifst.Properties(kError, false)) ofst_->SetProperties(kError, kError);
-    path_.clear();
-  }
-
-  constexpr bool InitState(StateId, StateId) const { return true; }
-
-  bool TreeArc(StateId, const ToArc &arc) {
-    if (ifst_->Final(arc.nextstate) == Weight::Zero()) {
-      path_.push_back(arc);
-    } else {
-      OutputPath();
-    }
-    return true;
-  }
-
-  bool BackArc(StateId, const FromArc &) {
-    FSTERROR() << "RandGenVisitor: cyclic input";
-    ofst_->SetProperties(kError, kError);
-    return false;
-  }
-
-  bool ForwardOrCrossArc(StateId, const FromArc &) {
-    OutputPath();
-    return true;
-  }
-
-  void FinishState(StateId s, StateId p, const FromArc *) {
-    if (p != kNoStateId && ifst_->Final(s) == Weight::Zero()) path_.pop_back();
-  }
-
-  void FinishVisit() {}
-
- private:
-  void OutputPath() {
-    if (ofst_->Start() == kNoStateId) {
-      const auto start = ofst_->AddState();
-      ofst_->SetStart(start);
-    }
-    auto src = ofst_->Start();
-    for (size_t i = 0; i < path_.size(); ++i) {
-      const auto dest = ofst_->AddState();
-      const ToArc arc(path_[i].ilabel, path_[i].olabel, Weight::One(), dest);
-      ofst_->AddArc(src, arc);
-      src = dest;
-    }
-    ofst_->SetFinal(src, Weight::One());
-  }
-
-  const Fst<FromArc> *ifst_;
-  MutableFst<ToArc> *ofst_;
-  std::vector<ToArc> path_;
-
-  RandGenVisitor(const RandGenVisitor &) = delete;
-  RandGenVisitor &operator=(const RandGenVisitor &) = delete;
-};
-
-}  // namespace internal
-
-// Randomly generate paths through an FST; details controlled by
-// RandGenOptions.
-template <class FromArc, class ToArc, class Selector>
-void RandGen(const Fst<FromArc> &ifst, MutableFst<ToArc> *ofst,
-             const RandGenOptions<Selector> &opts) {
-  using Sampler = ArcSampler<FromArc, Selector>;
-  auto *sampler = new Sampler(ifst, opts.selector, opts.max_length);
-  RandGenFstOptions<Sampler> fopts(CacheOptions(true, 0), sampler, opts.npath,
-                                   opts.weighted, opts.remove_total_weight);
-  RandGenFst<FromArc, ToArc, Sampler> rfst(ifst, fopts);
-  if (opts.weighted) {
-    *ofst = rfst;
-  } else {
-    internal::RandGenVisitor<FromArc, ToArc> rand_visitor(ofst);
-    DfsVisit(rfst, &rand_visitor);
-  }
-}
-
-// Randomly generate a path through an FST with the uniform distribution
-// over the transitions.
-template <class FromArc, class ToArc>
-void RandGen(const Fst<FromArc> &ifst, MutableFst<ToArc> *ofst) {
-  const UniformArcSelector<FromArc> uniform_selector;
-  RandGenOptions<UniformArcSelector<ToArc>> opts(uniform_selector);
-  RandGen(ifst, ofst, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_RANDGEN_H_
diff --git a/projects/llm_framework/include/fst/rational.h b/projects/llm_framework/include/fst/rational.h
deleted file mode 100644
index 184ebf3f..00000000
--- a/projects/llm_framework/include/fst/rational.h
+++ /dev/null
@@ -1,307 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// An FST implementation and base interface for delayed unions, concatenations,
-// and closures.
-
-#ifndef FST_RATIONAL_H_
-#define FST_RATIONAL_H_
-
-#include <algorithm>
-#include <string>
-#include <vector>
-
-#include <fst/mutable-fst.h>
-#include <fst/replace.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-using RationalFstOptions = CacheOptions;
-
-// This specifies whether to add the empty string.
-enum ClosureType {
-  CLOSURE_STAR = 0,  // Add the empty string.
-  CLOSURE_PLUS = 1   // Don't add the empty string.
-};
-
-template <class Arc>
-class RationalFst;
-
-template <class Arc>
-void Union(RationalFst<Arc> *fst1, const Fst<Arc> &fst2);
-
-template <class Arc>
-void Concat(RationalFst<Arc> *fst1, const Fst<Arc> &fst2);
-
-template <class Arc>
-void Concat(const Fst<Arc> &fst1, RationalFst<Arc> *fst2);
-
-template <class Arc>
-void Closure(RationalFst<Arc> *fst, ClosureType closure_type);
-
-namespace internal {
-
-// Implementation class for delayed unions, concatenations and closures.
-template <class A>
-class RationalFstImpl : public FstImpl<A> {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::WriteHeader;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  explicit RationalFstImpl(const RationalFstOptions &opts)
-      : nonterminals_(0), replace_options_(opts, 0) {
-    SetType("rational");
-    fst_tuples_.emplace_back(0, nullptr);
-  }
-
-  RationalFstImpl(const RationalFstImpl<Arc> &impl)
-      : rfst_(impl.rfst_),
-        nonterminals_(impl.nonterminals_),
-        replace_(impl.replace_ ? impl.replace_->Copy(true) : nullptr),
-        replace_options_(impl.replace_options_) {
-    SetType("rational");
-    fst_tuples_.reserve(impl.fst_tuples_.size());
-    for (const auto &pair : impl.fst_tuples_) {
-      fst_tuples_.emplace_back(pair.first,
-                               pair.second ? pair.second->Copy(true) : nullptr);
-    }
-  }
-
-  ~RationalFstImpl() override {
-    for (auto &tuple : fst_tuples_) delete tuple.second;
-  }
-
-  StateId Start() { return Replace()->Start(); }
-
-  Weight Final(StateId s) { return Replace()->Final(s); }
-
-  size_t NumArcs(StateId s) { return Replace()->NumArcs(s); }
-
-  size_t NumInputEpsilons(StateId s) { return Replace()->NumInputEpsilons(s); }
-
-  size_t NumOutputEpsilons(StateId s) {
-    return Replace()->NumOutputEpsilons(s);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && Replace()->Properties(kError, false)) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  // Implementation of UnionFst(fst1, fst2).
-  void InitUnion(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {
-    replace_.reset();
-    const auto props1 = fst1.Properties(kFstProperties, false);
-    const auto props2 = fst2.Properties(kFstProperties, false);
-    SetInputSymbols(fst1.InputSymbols());
-    SetOutputSymbols(fst1.OutputSymbols());
-    rfst_.AddState();
-    rfst_.AddState();
-    rfst_.SetStart(0);
-    rfst_.SetFinal(1, Weight::One());
-    rfst_.SetInputSymbols(fst1.InputSymbols());
-    rfst_.SetOutputSymbols(fst1.OutputSymbols());
-    nonterminals_ = 2;
-    rfst_.EmplaceArc(0, 0, -1, Weight::One(), 1);
-    rfst_.EmplaceArc(0, 0, -2, Weight::One(), 1);
-    fst_tuples_.emplace_back(-1, fst1.Copy());
-    fst_tuples_.emplace_back(-2, fst2.Copy());
-    SetProperties(UnionProperties(props1, props2, true), kCopyProperties);
-  }
-
-  // Implementation of ConcatFst(fst1, fst2).
-  void InitConcat(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {
-    replace_.reset();
-    const auto props1 = fst1.Properties(kFstProperties, false);
-    const auto props2 = fst2.Properties(kFstProperties, false);
-    SetInputSymbols(fst1.InputSymbols());
-    SetOutputSymbols(fst1.OutputSymbols());
-    rfst_.AddState();
-    rfst_.AddState();
-    rfst_.AddState();
-    rfst_.SetStart(0);
-    rfst_.SetFinal(2, Weight::One());
-    rfst_.SetInputSymbols(fst1.InputSymbols());
-    rfst_.SetOutputSymbols(fst1.OutputSymbols());
-    nonterminals_ = 2;
-    rfst_.EmplaceArc(0, 0, -1, Weight::One(), 1);
-    rfst_.EmplaceArc(1, 0, -2, Weight::One(), 2);
-    fst_tuples_.emplace_back(-1, fst1.Copy());
-    fst_tuples_.emplace_back(-2, fst2.Copy());
-    SetProperties(ConcatProperties(props1, props2, true), kCopyProperties);
-  }
-
-  // Implementation of ClosureFst(fst, closure_type).
-  void InitClosure(const Fst<Arc> &fst, ClosureType closure_type) {
-    replace_.reset();
-    const auto props = fst.Properties(kFstProperties, false);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-    if (closure_type == CLOSURE_STAR) {
-      rfst_.AddState();
-      rfst_.SetStart(0);
-      rfst_.SetFinal(0, Weight::One());
-      rfst_.EmplaceArc(0, 0, -1, Weight::One(), 0);
-    } else {
-      rfst_.AddState();
-      rfst_.AddState();
-      rfst_.SetStart(0);
-      rfst_.SetFinal(1, Weight::One());
-      rfst_.EmplaceArc(0, 0, -1, Weight::One(), 1);
-      rfst_.EmplaceArc(1, 0, 0, Weight::One(), 0);
-    }
-    rfst_.SetInputSymbols(fst.InputSymbols());
-    rfst_.SetOutputSymbols(fst.OutputSymbols());
-    fst_tuples_.emplace_back(-1, fst.Copy());
-    nonterminals_ = 1;
-    SetProperties(ClosureProperties(props, closure_type == CLOSURE_STAR, true),
-                  kCopyProperties);
-  }
-
-  // Implementation of Union(Fst &, RationalFst *).
-  void AddUnion(const Fst<Arc> &fst) {
-    replace_.reset();
-    const auto props1 = FstImpl<A>::Properties();
-    const auto props2 = fst.Properties(kFstProperties, false);
-    VectorFst<Arc> afst;
-    afst.AddState();
-    afst.AddState();
-    afst.SetStart(0);
-    afst.SetFinal(1, Weight::One());
-    ++nonterminals_;
-    afst.EmplaceArc(0, 0, -nonterminals_, Weight::One(), 1);
-    Union(&rfst_, afst);
-    fst_tuples_.emplace_back(-nonterminals_, fst.Copy());
-    SetProperties(UnionProperties(props1, props2, true), kCopyProperties);
-  }
-
-  // Implementation of Concat(Fst &, RationalFst *).
-  void AddConcat(const Fst<Arc> &fst, bool append) {
-    replace_.reset();
-    const auto props1 = FstImpl<A>::Properties();
-    const auto props2 = fst.Properties(kFstProperties, false);
-    VectorFst<Arc> afst;
-    afst.AddState();
-    afst.AddState();
-    afst.SetStart(0);
-    afst.SetFinal(1, Weight::One());
-    ++nonterminals_;
-    afst.EmplaceArc(0, 0, -nonterminals_, Weight::One(), 1);
-    if (append) {
-      Concat(&rfst_, afst);
-    } else {
-      Concat(afst, &rfst_);
-    }
-    fst_tuples_.emplace_back(-nonterminals_, fst.Copy());
-    SetProperties(ConcatProperties(props1, props2, true), kCopyProperties);
-  }
-
-  // Implementation of Closure(RationalFst *, closure_type).
-  void AddClosure(ClosureType closure_type) {
-    replace_.reset();
-    const auto props = FstImpl<A>::Properties();
-    Closure(&rfst_, closure_type);
-    SetProperties(ClosureProperties(props, closure_type == CLOSURE_STAR, true),
-                  kCopyProperties);
-  }
-
-  // Returns the underlying ReplaceFst, preserving ownership of the underlying
-  // object.
-  ReplaceFst<Arc> *Replace() const {
-    if (!replace_) {
-      fst_tuples_[0].second = rfst_.Copy();
-      replace_.reset(new ReplaceFst<Arc>(fst_tuples_, replace_options_));
-    }
-    return replace_.get();
-  }
-
- private:
-  // Rational topology machine, using negative non-terminals.
-  VectorFst<Arc> rfst_;
-  // Number of nonterminals used.
-  Label nonterminals_;
-  // Contains the nonterminals and their corresponding FSTs.
-  mutable std::vector<std::pair<Label, const Fst<Arc> *>> fst_tuples_;
-  // Underlying ReplaceFst.
-  mutable std::unique_ptr<ReplaceFst<Arc>> replace_;
-  const ReplaceFstOptions<Arc> replace_options_;
-};
-
-}  // namespace internal
-
-// Parent class for the delayed rational operations (union, concatenation, and
-// closure). This class attaches interface to implementation and handles
-// reference counting, delegating most methods to ImplToFst.
-template <class A>
-class RationalFst : public ImplToFst<internal::RationalFstImpl<A>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  using Impl = internal::RationalFstImpl<Arc>;
-
-  friend class StateIterator<RationalFst<Arc>>;
-  friend class ArcIterator<RationalFst<Arc>>;
-  friend void Union<>(RationalFst<Arc> *fst1, const Fst<Arc> &fst2);
-  friend void Concat<>(RationalFst<Arc> *fst1, const Fst<Arc> &fst2);
-  friend void Concat<>(const Fst<Arc> &fst1, RationalFst<Arc> *fst2);
-  friend void Closure<>(RationalFst<Arc> *fst, ClosureType closure_type);
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const override {
-    GetImpl()->Replace()->InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetImpl()->Replace()->InitArcIterator(s, data);
-  }
-
- protected:
-  using ImplToFst<Impl>::GetImpl;
-
-  explicit RationalFst(const RationalFstOptions &opts = RationalFstOptions())
-      : ImplToFst<Impl>(std::make_shared<Impl>(opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  RationalFst(const RationalFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
- private:
-  RationalFst &operator=(const RationalFst &) = delete;
-};
-
-// Specialization for RationalFst.
-template <class Arc>
-class StateIterator<RationalFst<Arc>> : public StateIterator<ReplaceFst<Arc>> {
- public:
-  explicit StateIterator(const RationalFst<Arc> &fst)
-      : StateIterator<ReplaceFst<Arc>>(*(fst.GetImpl()->Replace())) {}
-};
-
-// Specialization for RationalFst.
-template <class Arc>
-class ArcIterator<RationalFst<Arc>> : public CacheArcIterator<ReplaceFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const RationalFst<Arc> &fst, StateId s)
-      : ArcIterator<ReplaceFst<Arc>>(*(fst.GetImpl()->Replace()), s) {}
-};
-
-}  // namespace fst
-
-#endif  // FST_RATIONAL_H_
diff --git a/projects/llm_framework/include/fst/register.h b/projects/llm_framework/include/fst/register.h
deleted file mode 100644
index 2d1a6ea7..00000000
--- a/projects/llm_framework/include/fst/register.h
+++ /dev/null
@@ -1,115 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes for registering derived FST for generic reading.
-
-#ifndef FST_REGISTER_H_
-#define FST_REGISTER_H_
-
-#include <string>
-#include <type_traits>
-
-
-#include <fst/compat.h>
-#include <fst/generic-register.h>
-#include <fst/util.h>
-
-
-#include <fst/types.h>
-#include <fst/log.h>
-
-namespace fst {
-
-template <class Arc>
-class Fst;
-
-struct FstReadOptions;
-
-// This class represents a single entry in a FstRegister
-template <class Arc>
-struct FstRegisterEntry {
-  using Reader = Fst<Arc> *(*)(std::istream &istrm, const FstReadOptions &opts);
-  using Converter = Fst<Arc> *(*)(const Fst<Arc> &fst);
-
-  Reader reader;
-  Converter converter;
-
-  explicit FstRegisterEntry(Reader reader = nullptr,
-                            Converter converter = nullptr)
-      : reader(reader), converter(converter) {}
-};
-
-// This class maintains the correspondence between a string describing
-// an FST type, and its reader and converter.
-template <class Arc>
-class FstRegister
-    : public GenericRegister<string, FstRegisterEntry<Arc>, FstRegister<Arc>> {
- public:
-  using Reader = typename FstRegisterEntry<Arc>::Reader;
-  using Converter = typename FstRegisterEntry<Arc>::Converter;
-
-  const Reader GetReader(const string &type) const {
-    return this->GetEntry(type).reader;
-  }
-
-  const Converter GetConverter(const string &type) const {
-    return this->GetEntry(type).converter;
-  }
-
- protected:
-  string ConvertKeyToSoFilename(const string &key) const override {
-    string legal_type(key);
-    ConvertToLegalCSymbol(&legal_type);
-    return legal_type + "-fst.so";
-  }
-};
-
-// This class registers an FST type for generic reading and creating.
-// The type must have a default constructor and a copy constructor from
-// Fst<Arc>.
-template <class FST>
-class FstRegisterer : public GenericRegisterer<FstRegister<typename FST::Arc>> {
- public:
-  using Arc = typename FST::Arc;
-  using Entry = typename FstRegister<Arc>::Entry;
-  using Reader = typename FstRegister<Arc>::Reader;
-
-  FstRegisterer()
-      : GenericRegisterer<FstRegister<typename FST::Arc>>(FST().Type(),
-                                                          BuildEntry()) {}
-
- private:
-  static Fst<Arc> *ReadGeneric(
-      std::istream &strm, const FstReadOptions &opts) {
-    static_assert(std::is_base_of<Fst<Arc>, FST>::value,
-                  "FST class does not inherit from Fst<Arc>");
-    return FST::Read(strm, opts);
-  }
-
-  static Entry BuildEntry() {
-    return Entry(&ReadGeneric, &FstRegisterer<FST>::Convert);
-  }
-
-  static Fst<Arc> *Convert(const Fst<Arc> &fst) { return new FST(fst); }
-};
-
-// Convenience macro to generate static FstRegisterer instance.
-#define REGISTER_FST(FST, Arc) \
-  static fst::FstRegisterer<FST<Arc>> FST##_##Arc##_registerer
-
-// Converts an FST to the specified type.
-template <class Arc>
-Fst<Arc> *Convert(const Fst<Arc> &fst, const string &fst_type) {
-  auto *reg = FstRegister<Arc>::GetRegister();
-  const auto converter = reg->GetConverter(fst_type);
-  if (!converter) {
-    FSTERROR() << "Fst::Convert: Unknown FST type " << fst_type << " (arc type "
-               << Arc::Type() << ")";
-    return nullptr;
-  }
-  return converter(fst);
-}
-
-}  // namespace fst
-
-#endif  // FST_REGISTER_H_
diff --git a/projects/llm_framework/include/fst/relabel.h b/projects/llm_framework/include/fst/relabel.h
deleted file mode 100644
index 0979b077..00000000
--- a/projects/llm_framework/include/fst/relabel.h
+++ /dev/null
@@ -1,472 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to relabel an FST (either on input or output).
-
-#ifndef FST_RELABEL_H_
-#define FST_RELABEL_H_
-
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/test-properties.h>
-
-
-#include <unordered_map>
-
-namespace fst {
-
-// Relabels either the input labels or output labels. The old to
-// new labels are specified using a vector of std::pair<Label, Label>.
-// Any label associations not specified are assumed to be identity
-// mapping. The destination labels must be valid labels (e.g., not kNoLabel).
-template <class Arc>
-void Relabel(
-    MutableFst<Arc> *fst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &ipairs,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &opairs) {
-  using Label = typename Arc::Label;
-  const auto props = fst->Properties(kFstProperties, false);
-  // Constructs label-to-label maps.
-  const std::unordered_map<Label, Label> input_map(
-      ipairs.begin(), ipairs.end());
-  const std::unordered_map<Label, Label> output_map(
-      opairs.begin(), opairs.end());
-  for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
-       siter.Next()) {
-    for (MutableArcIterator<MutableFst<Arc>> aiter(fst, siter.Value());
-         !aiter.Done(); aiter.Next()) {
-      auto arc = aiter.Value();
-      // Relabels input.
-      auto it = input_map.find(arc.ilabel);
-      if (it != input_map.end()) {
-        if (it->second == kNoLabel) {
-          FSTERROR() << "Input symbol ID " << arc.ilabel
-                     << " missing from target vocabulary";
-          fst->SetProperties(kError, kError);
-          return;
-        }
-        arc.ilabel = it->second;
-      }
-      // Relabels output.
-      it = output_map.find(arc.olabel);
-      if (it != output_map.end()) {
-        if (it->second == kNoLabel) {
-          FSTERROR() << "Output symbol id " << arc.olabel
-                     << " missing from target vocabulary";
-          fst->SetProperties(kError, kError);
-          return;
-        }
-        arc.olabel = it->second;
-      }
-      aiter.SetValue(arc);
-    }
-  }
-  fst->SetProperties(RelabelProperties(props), kFstProperties);
-}
-
-// Relabels either the input labels or output labels. The old to
-// new labels are specified using pairs of old and new symbol tables.
-// The tables must contain (at least) all labels on the appropriate side of the
-// FST. If the 'unknown_i(o)symbol' is non-empty, it is used to label any
-// missing symbol in new_i(o)symbols table.
-template <class Arc>
-void Relabel(MutableFst<Arc> *fst,
-             const SymbolTable *old_isymbols, const SymbolTable *new_isymbols,
-             const string &unknown_isymbol, bool attach_new_isymbols,
-             const SymbolTable *old_osymbols, const SymbolTable *new_osymbols,
-             const string &unknown_osymbol, bool attach_new_osymbols) {
-  using Label = typename Arc::Label;
-  // Constructs vectors of input-side label pairs.
-  std::vector<std::pair<Label, Label>> ipairs;
-  if (old_isymbols && new_isymbols) {
-    size_t num_missing_syms = 0;
-    Label unknown_ilabel = kNoLabel;
-    if (!unknown_isymbol.empty()) {
-      unknown_ilabel = new_isymbols->Find(unknown_isymbol);
-      if (unknown_ilabel == kNoLabel) {
-        VLOG(1) << "Input symbol '" << unknown_isymbol
-                << "' missing from target symbol table";
-        ++num_missing_syms;
-      }
-    }
-
-    for (SymbolTableIterator siter(*old_isymbols); !siter.Done();
-         siter.Next()) {
-      const auto old_index = siter.Value();
-      const auto symbol = siter.Symbol();
-      auto new_index = new_isymbols->Find(siter.Symbol());
-      if (new_index == kNoLabel) {
-        if (unknown_ilabel != kNoLabel) {
-          new_index = unknown_ilabel;
-        } else {
-          VLOG(1) << "Input symbol ID " << old_index << " symbol '" << symbol
-                  << "' missing from target symbol table";
-          ++num_missing_syms;
-        }
-      }
-      ipairs.push_back(std::make_pair(old_index, new_index));
-    }
-    if (num_missing_syms > 0) {
-      LOG(WARNING) << "Target symbol table missing: " << num_missing_syms
-                   << " input symbols";
-    }
-    if (attach_new_isymbols) fst->SetInputSymbols(new_isymbols);
-  }
-  // Constructs vectors of output-side label pairs.
-  std::vector<std::pair<Label, Label>> opairs;
-  if (old_osymbols && new_osymbols) {
-    size_t num_missing_syms = 0;
-    Label unknown_olabel = kNoLabel;
-    if (!unknown_osymbol.empty()) {
-      unknown_olabel = new_osymbols->Find(unknown_osymbol);
-      if (unknown_olabel == kNoLabel) {
-        VLOG(1) << "Output symbol '" << unknown_osymbol
-                << "' missing from target symbol table";
-        ++num_missing_syms;
-      }
-    }
-
-    for (SymbolTableIterator siter(*old_osymbols); !siter.Done();
-         siter.Next()) {
-      const auto old_index = siter.Value();
-      const auto symbol = siter.Symbol();
-      auto new_index = new_osymbols->Find(siter.Symbol());
-      if (new_index == kNoLabel) {
-        if (unknown_olabel != kNoLabel) {
-          new_index = unknown_olabel;
-        } else {
-          VLOG(1) << "Output symbol ID " << old_index << " symbol '" << symbol
-                  << "' missing from target symbol table";
-          ++num_missing_syms;
-        }
-      }
-      opairs.push_back(std::make_pair(old_index, new_index));
-    }
-    if (num_missing_syms > 0) {
-      LOG(WARNING) << "Target symbol table missing: " << num_missing_syms
-                   << " output symbols";
-    }
-    if (attach_new_osymbols) fst->SetOutputSymbols(new_osymbols);
-  }
-  // Calls relabel using vector of relabel pairs.
-  Relabel(fst, ipairs, opairs);
-}
-
-// Same as previous but no special allowance for unknown symbols. Kept
-// for backward compat.
-template <class Arc>
-void Relabel(MutableFst<Arc> *fst, const SymbolTable *old_isymbols,
-             const SymbolTable *new_isymbols, bool attach_new_isymbols,
-             const SymbolTable *old_osymbols, const SymbolTable *new_osymbols,
-             bool attach_new_osymbols) {
-  Relabel(fst,
-          old_isymbols, new_isymbols, "" /* no unknown isymbol */,
-          attach_new_isymbols,
-          old_osymbols, new_osymbols, "" /* no unknown ioymbol */,
-          attach_new_osymbols);
-}
-
-
-// Relabels either the input labels or output labels. The old to
-// new labels are specified using symbol tables. Any label associations not
-// specified are assumed to be identity mapping.
-template <class Arc>
-void Relabel(MutableFst<Arc> *fst, const SymbolTable *new_isymbols,
-             const SymbolTable *new_osymbols) {
-  Relabel(fst, fst->InputSymbols(), new_isymbols, true, fst->OutputSymbols(),
-          new_osymbols, true);
-}
-
-using RelabelFstOptions = CacheOptions;
-
-template <class Arc>
-class RelabelFst;
-
-namespace internal {
-
-// Relabels an FST from one symbol set to another. Relabeling can either be on
-// input or output space. RelabelFst implements a delayed version of the
-// relabel. Arcs are relabeled on the fly and not cached; i.e., each request is
-// recomputed.
-template <class Arc>
-class RelabelFstImpl : public CacheImpl<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::WriteHeader;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheImpl<Arc>::PushArc;
-  using CacheImpl<Arc>::HasArcs;
-  using CacheImpl<Arc>::HasFinal;
-  using CacheImpl<Arc>::HasStart;
-  using CacheImpl<Arc>::SetArcs;
-  using CacheImpl<Arc>::SetFinal;
-  using CacheImpl<Arc>::SetStart;
-
-  friend class StateIterator<RelabelFst<Arc>>;
-
-  RelabelFstImpl(const Fst<Arc> &fst,
-                 const std::vector<std::pair<Label, Label>> &ipairs,
-                 const std::vector<std::pair<Label, Label>> &opairs,
-                 const RelabelFstOptions &opts)
-      : CacheImpl<Arc>(opts),
-        fst_(fst.Copy()),
-        input_map_(ipairs.begin(), ipairs.end()),
-        output_map_(opairs.begin(), opairs.end()),
-        relabel_input_(!ipairs.empty()),
-        relabel_output_(!opairs.empty()) {
-    SetProperties(RelabelProperties(fst.Properties(kCopyProperties, false)));
-    SetType("relabel");
-  }
-
-  RelabelFstImpl(const Fst<Arc> &fst,
-                 const SymbolTable *old_isymbols,
-                 const SymbolTable *new_isymbols,
-                 const SymbolTable *old_osymbols,
-                 const SymbolTable *new_osymbols,
-                 const RelabelFstOptions &opts)
-      : CacheImpl<Arc>(opts),
-        fst_(fst.Copy()),
-        relabel_input_(false),
-        relabel_output_(false) {
-    SetType("relabel");
-    SetProperties(RelabelProperties(fst.Properties(kCopyProperties, false)));
-    SetInputSymbols(old_isymbols);
-    SetOutputSymbols(old_osymbols);
-    if (old_isymbols && new_isymbols &&
-        old_isymbols->LabeledCheckSum() != new_isymbols->LabeledCheckSum()) {
-      for (SymbolTableIterator siter(*old_isymbols); !siter.Done();
-           siter.Next()) {
-        input_map_[siter.Value()] = new_isymbols->Find(siter.Symbol());
-      }
-      SetInputSymbols(new_isymbols);
-      relabel_input_ = true;
-    }
-    if (old_osymbols && new_osymbols &&
-        old_osymbols->LabeledCheckSum() != new_osymbols->LabeledCheckSum()) {
-      for (SymbolTableIterator siter(*old_osymbols); !siter.Done();
-           siter.Next()) {
-        output_map_[siter.Value()] = new_osymbols->Find(siter.Symbol());
-      }
-      SetOutputSymbols(new_osymbols);
-      relabel_output_ = true;
-    }
-  }
-
-  RelabelFstImpl(const RelabelFstImpl<Arc> &impl)
-      : CacheImpl<Arc>(impl),
-        fst_(impl.fst_->Copy(true)),
-        input_map_(impl.input_map_),
-        output_map_(impl.output_map_),
-        relabel_input_(impl.relabel_input_),
-        relabel_output_(impl.relabel_output_) {
-    SetType("relabel");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  StateId Start() {
-    if (!HasStart()) SetStart(fst_->Start());
-    return CacheImpl<Arc>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) SetFinal(s, fst_->Final(s));
-    return CacheImpl<Arc>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumOutputEpsilons(s);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && fst_->Properties(kError, false)) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<Arc>::InitArcIterator(s, data);
-  }
-
-  void Expand(StateId s) {
-    for (ArcIterator<Fst<Arc>> aiter(*fst_, s); !aiter.Done(); aiter.Next()) {
-      auto arc = aiter.Value();
-      if (relabel_input_) {
-        auto it = input_map_.find(arc.ilabel);
-        if (it != input_map_.end()) arc.ilabel = it->second;
-      }
-      if (relabel_output_) {
-        auto it = output_map_.find(arc.olabel);
-        if (it != output_map_.end()) {
-          arc.olabel = it->second;
-        }
-      }
-      PushArc(s, std::move(arc));
-    }
-    SetArcs(s);
-  }
-
- private:
-  std::unique_ptr<const Fst<Arc>> fst_;
-
-  std::unordered_map<Label, Label> input_map_;
-  std::unordered_map<Label, Label> output_map_;
-  bool relabel_input_;
-  bool relabel_output_;
-};
-
-}  // namespace internal
-
-// This class attaches interface to implementation and handles
-// reference counting, delegating most methods to ImplToFst.
-template <class A>
-class RelabelFst : public ImplToFst<internal::RelabelFstImpl<A>> {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::RelabelFstImpl<Arc>;
-
-  friend class ArcIterator<RelabelFst<A>>;
-  friend class StateIterator<RelabelFst<A>>;
-
-  RelabelFst(const Fst<Arc> &fst,
-             const std::vector<std::pair<Label, Label>> &ipairs,
-             const std::vector<std::pair<Label, Label>> &opairs,
-             const RelabelFstOptions &opts = RelabelFstOptions())
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, ipairs, opairs, opts)) {}
-
-  RelabelFst(const Fst<Arc> &fst, const SymbolTable *new_isymbols,
-             const SymbolTable *new_osymbols,
-             const RelabelFstOptions &opts = RelabelFstOptions())
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, fst.InputSymbols(), new_isymbols,
-                                   fst.OutputSymbols(), new_osymbols, opts)) {}
-
-  RelabelFst(const Fst<Arc> &fst, const SymbolTable *old_isymbols,
-             const SymbolTable *new_isymbols, const SymbolTable *old_osymbols,
-             const SymbolTable *new_osymbols,
-             const RelabelFstOptions &opts = RelabelFstOptions())
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, old_isymbols, new_isymbols,
-                                               old_osymbols, new_osymbols,
-                                               opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  RelabelFst(const RelabelFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Gets a copy of this RelabelFst. See Fst<>::Copy() for further doc.
-  RelabelFst<Arc> *Copy(bool safe = false) const override {
-    return new RelabelFst<Arc>(*this, safe);
-  }
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    return GetMutableImpl()->InitArcIterator(s, data);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  RelabelFst &operator=(const RelabelFst &) = delete;
-};
-
-// Specialization for RelabelFst.
-template <class Arc>
-class StateIterator<RelabelFst<Arc>> : public StateIteratorBase<Arc> {
- public:
-  using StateId = typename Arc::StateId;
-
-  explicit StateIterator(const RelabelFst<Arc> &fst)
-      : impl_(fst.GetImpl()), siter_(*impl_->fst_), s_(0) {}
-
-  bool Done() const final { return siter_.Done(); }
-
-  StateId Value() const final { return s_; }
-
-  void Next() final {
-    if (!siter_.Done()) {
-      ++s_;
-      siter_.Next();
-    }
-  }
-
-  void Reset() final {
-    s_ = 0;
-    siter_.Reset();
-  }
-
- private:
-  const internal::RelabelFstImpl<Arc>* impl_;
-  StateIterator<Fst<Arc>> siter_;
-  StateId s_;
-
-  StateIterator(const StateIterator &) = delete;
-  StateIterator &operator=(const StateIterator &) = delete;
-};
-
-// Specialization for RelabelFst.
-template <class Arc>
-class ArcIterator<RelabelFst<Arc>> : public CacheArcIterator<RelabelFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const RelabelFst<Arc> &fst, StateId s)
-      : CacheArcIterator<RelabelFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc>
-inline void RelabelFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<RelabelFst<Arc>>(*this);
-}
-
-// Useful alias when using StdArc.
-using StdRelabelFst = RelabelFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_RELABEL_H_
diff --git a/projects/llm_framework/include/fst/replace-util.h b/projects/llm_framework/include/fst/replace-util.h
deleted file mode 100644
index 42c69824..00000000
--- a/projects/llm_framework/include/fst/replace-util.h
+++ /dev/null
@@ -1,629 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Utility classes for the recursive replacement of FSTs (RTNs).
-
-#ifndef FST_REPLACE_UTIL_H_
-#define FST_REPLACE_UTIL_H_
-
-#include <map>
-#include <unordered_map>
-#include <unordered_set>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/connect.h>
-#include <fst/mutable-fst.h>
-#include <fst/topsort.h>
-#include <fst/vector-fst.h>
-
-
-namespace fst {
-
-// This specifies what labels to output on the call or return arc. Note that
-// REPLACE_LABEL_INPUT and REPLACE_LABEL_OUTPUT will produce transducers when
-// applied to acceptors.
-enum ReplaceLabelType {
-  // Epsilon labels on both input and output.
-  REPLACE_LABEL_NEITHER = 1,
-  // Non-epsilon labels on input and epsilon on output.
-  REPLACE_LABEL_INPUT = 2,
-  // Epsilon on input and non-epsilon on output.
-  REPLACE_LABEL_OUTPUT = 3,
-  // Non-epsilon labels on both input and output.
-  REPLACE_LABEL_BOTH = 4
-};
-
-// By default ReplaceUtil will copy the input label of the replace arc.
-// The call_label_type and return_label_type options specify how to manage
-// the labels of the call arc and the return arc of the replace FST
-struct ReplaceUtilOptions {
-  int64 root;                          // Root rule for expansion.
-  ReplaceLabelType call_label_type;    // How to label call arc.
-  ReplaceLabelType return_label_type;  // How to label return arc.
-  int64 return_label;                  // Label to put on return arc.
-
-  explicit ReplaceUtilOptions(
-      int64 root = kNoLabel,
-      ReplaceLabelType call_label_type = REPLACE_LABEL_INPUT,
-      ReplaceLabelType return_label_type = REPLACE_LABEL_NEITHER,
-      int64 return_label = 0)
-      : root(root),
-        call_label_type(call_label_type),
-        return_label_type(return_label_type),
-        return_label(return_label) {}
-
-  // For backwards compatibility.
-  ReplaceUtilOptions(int64 root, bool epsilon_replace_arc)
-      : ReplaceUtilOptions(root,
-                           epsilon_replace_arc ? REPLACE_LABEL_NEITHER
-                                               : REPLACE_LABEL_INPUT) {}
-};
-
-// Every non-terminal on a path appears as the first label on that path in every
-// FST associated with a given SCC of the replace dependency graph. This would
-// be true if the SCC were formed from left-linear grammar rules.
-constexpr uint8 kReplaceSCCLeftLinear = 0x01;
-// Every non-terminal on a path appears as the final label on that path in every
-// FST associated with a given SCC of the replace dependency graph. This would
-// be true if the SCC were formed from right-linear grammar rules.
-constexpr uint8 kReplaceSCCRightLinear = 0x02;
-// The SCC in the replace dependency graph has more than one state or a
-// self-loop.
-constexpr uint8 kReplaceSCCNonTrivial = 0x04;
-
-// Defined in replace.h.
-template <class Arc>
-void Replace(
-    const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>> &,
-    MutableFst<Arc> *, const ReplaceUtilOptions &);
-
-// Utility class for the recursive replacement of FSTs (RTNs). The user provides
-// a set of label/FST pairs at construction. These are used by methods for
-// testing cyclic dependencies and connectedness and doing RTN connection and
-// specific FST replacement by label or for various optimization properties. The
-// modified results can be obtained with the GetFstPairs() or
-// GetMutableFstPairs() methods.
-template <class Arc>
-class ReplaceUtil {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstPair = std::pair<Label, const Fst<Arc> *>;
-  using MutableFstPair = std::pair<Label, MutableFst<Arc> *>;
-  using NonTerminalHash = std::unordered_map<Label, Label>;
-
-  // Constructs from mutable FSTs; FST ownership is given to ReplaceUtil.
-  ReplaceUtil(const std::vector<MutableFstPair> &fst_pairs,
-              const ReplaceUtilOptions &opts);
-
-  // Constructs from FSTs; FST ownership is retained by caller.
-  ReplaceUtil(const std::vector<FstPair> &fst_pairs,
-              const ReplaceUtilOptions &opts);
-
-  // Constructs from ReplaceFst internals; FST ownership is retained by caller.
-  ReplaceUtil(const std::vector<std::unique_ptr<const Fst<Arc>>> &fst_array,
-              const NonTerminalHash &nonterminal_hash,
-              const ReplaceUtilOptions &opts);
-
-  ~ReplaceUtil() {
-    for (Label i = 0; i < fst_array_.size(); ++i) delete fst_array_[i];
-  }
-
-  // True if the non-terminal dependencies are cyclic. Cyclic dependencies will
-  // result in an unexpandable FST.
-  bool CyclicDependencies() const {
-    GetDependencies(false);
-    return depprops_ & kCyclic;
-  }
-
-  // Returns the strongly-connected component ID in the dependency graph of the
-  // replace FSTS.
-  StateId SCC(Label label) const {
-    GetDependencies(false);
-    const auto it = nonterminal_hash_.find(label);
-    if (it == nonterminal_hash_.end()) return kNoStateId;
-    return depscc_[it->second];
-  }
-
-  // Returns properties for the strongly-connected component in the dependency
-  // graph of the replace FSTs. If the SCC is kReplaceSCCLeftLinear or
-  // kReplaceSCCRightLinear, that SCC can be represented as finite-state despite
-  // any cyclic dependencies, but not by the usual replacement operation (see
-  // fst/extensions/pdt/replace.h).
-  uint8 SCCProperties(StateId scc_id) {
-    GetSCCProperties();
-    return depsccprops_[scc_id];
-  }
-
-  // Returns true if no useless FSTs, states or transitions are present in the
-  // RTN.
-  bool Connected() const {
-    GetDependencies(false);
-    uint64 props = kAccessible | kCoAccessible;
-    for (Label i = 0; i < fst_array_.size(); ++i) {
-      if (!fst_array_[i]) continue;
-      if (fst_array_[i]->Properties(props, true) != props || !depaccess_[i]) {
-        return false;
-      }
-    }
-    return true;
-  }
-
-  // Removes useless FSTs, states and transitions from the RTN.
-  void Connect();
-
-  // Replaces FSTs specified by labels, unless there are cyclic dependencies.
-  void ReplaceLabels(const std::vector<Label> &labels);
-
-  // Replaces FSTs that have at most nstates states, narcs arcs and nnonterm
-  // non-terminals (updating in reverse dependency order), unless there are
-  // cyclic dependencies.
-  void ReplaceBySize(size_t nstates, size_t narcs, size_t nnonterms);
-
-  // Replaces singleton FSTS, unless there are cyclic dependencies.
-  void ReplaceTrivial() { ReplaceBySize(2, 1, 1); }
-
-  // Replaces non-terminals that have at most ninstances instances (updating in
-  // dependency order), unless there are cyclic dependencies.
-  void ReplaceByInstances(size_t ninstances);
-
-  // Replaces non-terminals that have only one instance, unless there are cyclic
-  // dependencies.
-  void ReplaceUnique() { ReplaceByInstances(1); }
-
-  // Returns label/FST pairs, retaining FST ownership.
-  void GetFstPairs(std::vector<FstPair> *fst_pairs);
-
-  // Returns label/mutable FST pairs, giving FST ownership over to the caller.
-  void GetMutableFstPairs(std::vector<MutableFstPair> *mutable_fst_pairs);
-
- private:
-  // FST statistics.
-  struct ReplaceStats {
-    StateId nstates;  // Number of states.
-    StateId nfinal;   // Number of final states.
-    size_t narcs;     // Number of arcs.
-    Label nnonterms;  // Number of non-terminals in FST.
-    size_t nref;      // Number of non-terminal instances referring to this FST.
-    // Number of times that ith FST references this FST
-    std::map<Label, size_t> inref;
-    // Number of times that this FST references the ith FST
-    std::map<Label, size_t> outref;
-
-    ReplaceStats() : nstates(0), nfinal(0), narcs(0), nnonterms(0), nref(0) {}
-  };
-
-  // Checks that Mutable FSTs exists, creating them if necessary.
-  void CheckMutableFsts();
-
-  // Computes the dependency graph for the RTN, computing dependency statistics
-  // if stats is true.
-  void GetDependencies(bool stats) const;
-
-  void ClearDependencies() const {
-    depfst_.DeleteStates();
-    stats_.clear();
-    depprops_ = 0;
-    depsccprops_.clear();
-    have_stats_ = false;
-  }
-
-  // Gets topological order of dependencies, returning false with cyclic input.
-  bool GetTopOrder(const Fst<Arc> &fst, std::vector<Label> *toporder) const;
-
-  // Updates statistics to reflect the replacement of the jth FST.
-  void UpdateStats(Label j);
-
-  // Computes the properties for the strongly-connected component in the
-  // dependency graph of the replace FSTs.
-  void GetSCCProperties() const;
-
-  Label root_label_;                                  // Root non-terminal.
-  Label root_fst_;                                    // Root FST ID.
-  ReplaceLabelType call_label_type_;                  // See Replace().
-  ReplaceLabelType return_label_type_;                // See Replace().
-  int64 return_label_;                                // See Replace().
-  std::vector<const Fst<Arc> *> fst_array_;           // FST per ID.
-  std::vector<MutableFst<Arc> *> mutable_fst_array_;  // Mutable FST per ID.
-  std::vector<Label> nonterminal_array_;              // FST ID to non-terminal.
-  NonTerminalHash nonterminal_hash_;                  // Non-terminal to FST ID.
-  mutable VectorFst<Arc> depfst_;                     // FST ID dependencies.
-  mutable std::vector<StateId> depscc_;               // FST SCC ID.
-  mutable std::vector<bool> depaccess_;               // FST ID accessibility.
-  mutable uint64 depprops_;                           // Dependency FST props.
-  mutable bool have_stats_;                  // Have dependency statistics?
-  mutable std::vector<ReplaceStats> stats_;  // Per-FST statistics.
-  mutable std::vector<uint8> depsccprops_;   // SCC properties.
-  ReplaceUtil(const ReplaceUtil &) = delete;
-  ReplaceUtil &operator=(const ReplaceUtil &) = delete;
-};
-
-template <class Arc>
-ReplaceUtil<Arc>::ReplaceUtil(const std::vector<MutableFstPair> &fst_pairs,
-                              const ReplaceUtilOptions &opts)
-    : root_label_(opts.root),
-      call_label_type_(opts.call_label_type),
-      return_label_type_(opts.return_label_type),
-      return_label_(opts.return_label),
-      depprops_(0),
-      have_stats_(false) {
-  fst_array_.push_back(nullptr);
-  mutable_fst_array_.push_back(nullptr);
-  nonterminal_array_.push_back(kNoLabel);
-  for (const auto &fst_pair : fst_pairs) {
-    const auto label = fst_pair.first;
-    auto *fst = fst_pair.second;
-    nonterminal_hash_[label] = fst_array_.size();
-    nonterminal_array_.push_back(label);
-    fst_array_.push_back(fst);
-    mutable_fst_array_.push_back(fst);
-  }
-  root_fst_ = nonterminal_hash_[root_label_];
-  if (!root_fst_) {
-    FSTERROR() << "ReplaceUtil: No root FST for label: " << root_label_;
-  }
-}
-
-template <class Arc>
-ReplaceUtil<Arc>::ReplaceUtil(const std::vector<FstPair> &fst_pairs,
-                              const ReplaceUtilOptions &opts)
-    : root_label_(opts.root),
-      call_label_type_(opts.call_label_type),
-      return_label_type_(opts.return_label_type),
-      return_label_(opts.return_label),
-      depprops_(0),
-      have_stats_(false) {
-  fst_array_.push_back(nullptr);
-  nonterminal_array_.push_back(kNoLabel);
-  for (const auto &fst_pair : fst_pairs) {
-    const auto label = fst_pair.first;
-    const auto *fst = fst_pair.second;
-    nonterminal_hash_[label] = fst_array_.size();
-    nonterminal_array_.push_back(label);
-    fst_array_.push_back(fst->Copy());
-  }
-  root_fst_ = nonterminal_hash_[root_label_];
-  if (!root_fst_) {
-    FSTERROR() << "ReplaceUtil: No root FST for label: " << root_label_;
-  }
-}
-
-template <class Arc>
-ReplaceUtil<Arc>::ReplaceUtil(
-    const std::vector<std::unique_ptr<const Fst<Arc>>> &fst_array,
-    const NonTerminalHash &nonterminal_hash, const ReplaceUtilOptions &opts)
-    : root_fst_(opts.root),
-      call_label_type_(opts.call_label_type),
-      return_label_type_(opts.return_label_type),
-      return_label_(opts.return_label),
-      nonterminal_array_(fst_array.size()),
-      nonterminal_hash_(nonterminal_hash),
-      depprops_(0),
-      have_stats_(false) {
-  fst_array_.push_back(nullptr);
-  for (size_t i = 1; i < fst_array.size(); ++i) {
-    fst_array_.push_back(fst_array[i]->Copy());
-  }
-  for (auto it = nonterminal_hash.begin(); it != nonterminal_hash.end(); ++it) {
-    nonterminal_array_[it->second] = it->first;
-  }
-  root_label_ = nonterminal_array_[root_fst_];
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::GetDependencies(bool stats) const {
-  if (depfst_.NumStates() > 0) {
-    if (stats && !have_stats_) {
-      ClearDependencies();
-    } else {
-      return;
-    }
-  }
-  have_stats_ = stats;
-  if (have_stats_) stats_.reserve(fst_array_.size());
-  for (Label i = 0; i < fst_array_.size(); ++i) {
-    depfst_.AddState();
-    depfst_.SetFinal(i, Weight::One());
-    if (have_stats_) stats_.push_back(ReplaceStats());
-  }
-  depfst_.SetStart(root_fst_);
-  // An arc from each state (representing the FST) to the state representing the
-  // FST being replaced
-  for (Label i = 0; i < fst_array_.size(); ++i) {
-    const auto *ifst = fst_array_[i];
-    if (!ifst) continue;
-    for (StateIterator<Fst<Arc>> siter(*ifst); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      if (have_stats_) {
-        ++stats_[i].nstates;
-        if (ifst->Final(s) != Weight::Zero()) ++stats_[i].nfinal;
-      }
-      for (ArcIterator<Fst<Arc>> aiter(*ifst, s); !aiter.Done();
-           aiter.Next()) {
-        if (have_stats_) ++stats_[i].narcs;
-        const auto &arc = aiter.Value();
-        auto it = nonterminal_hash_.find(arc.olabel);
-        if (it != nonterminal_hash_.end()) {
-          const auto j = it->second;
-          depfst_.EmplaceArc(i, arc.olabel, arc.olabel, Weight::One(), j);
-          if (have_stats_) {
-            ++stats_[i].nnonterms;
-            ++stats_[j].nref;
-            ++stats_[j].inref[i];
-            ++stats_[i].outref[j];
-          }
-        }
-      }
-    }
-  }
-  // Computes accessibility info.
-  SccVisitor<Arc> scc_visitor(&depscc_, &depaccess_, nullptr, &depprops_);
-  DfsVisit(depfst_, &scc_visitor);
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::UpdateStats(Label j) {
-  if (!have_stats_) {
-    FSTERROR() << "ReplaceUtil::UpdateStats: Stats not available";
-    return;
-  }
-  if (j == root_fst_) return;  // Can't replace root.
-  for (auto in = stats_[j].inref.begin(); in != stats_[j].inref.end(); ++in) {
-    const auto i = in->first;
-    const auto ni = in->second;
-    stats_[i].nstates += stats_[j].nstates * ni;
-    stats_[i].narcs += (stats_[j].narcs + 1) * ni;
-    stats_[i].nnonterms += (stats_[j].nnonterms - 1) * ni;
-    stats_[i].outref.erase(j);
-    for (auto out = stats_[j].outref.begin(); out != stats_[j].outref.end();
-         ++out) {
-      const auto k = out->first;
-      const auto nk = out->second;
-      stats_[i].outref[k] += ni * nk;
-    }
-  }
-  for (auto out = stats_[j].outref.begin(); out != stats_[j].outref.end();
-       ++out) {
-    const auto k = out->first;
-    const auto nk = out->second;
-    stats_[k].nref -= nk;
-    stats_[k].inref.erase(j);
-    for (auto in = stats_[j].inref.begin(); in != stats_[j].inref.end(); ++in) {
-      const auto i = in->first;
-      const auto ni = in->second;
-      stats_[k].inref[i] += ni * nk;
-      stats_[k].nref += ni * nk;
-    }
-  }
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::CheckMutableFsts() {
-  if (mutable_fst_array_.empty()) {
-    for (Label i = 0; i < fst_array_.size(); ++i) {
-      if (!fst_array_[i]) {
-        mutable_fst_array_.push_back(nullptr);
-      } else {
-        mutable_fst_array_.push_back(new VectorFst<Arc>(*fst_array_[i]));
-        delete fst_array_[i];
-        fst_array_[i] = mutable_fst_array_[i];
-      }
-    }
-  }
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::Connect() {
-  CheckMutableFsts();
-  static constexpr auto props = kAccessible | kCoAccessible;
-  for (auto *mutable_fst : mutable_fst_array_) {
-    if (!mutable_fst) continue;
-    if (mutable_fst->Properties(props, false) != props) {
-      fst::Connect(mutable_fst);
-    }
-  }
-  GetDependencies(false);
-  for (Label i = 0; i < mutable_fst_array_.size(); ++i) {
-    auto *fst = mutable_fst_array_[i];
-    if (fst && !depaccess_[i]) {
-      delete fst;
-      fst_array_[i] = nullptr;
-      mutable_fst_array_[i] = nullptr;
-    }
-  }
-  ClearDependencies();
-}
-
-template <class Arc>
-bool ReplaceUtil<Arc>::GetTopOrder(const Fst<Arc> &fst,
-                                   std::vector<Label> *toporder) const {
-  // Finds topological order of dependencies.
-  std::vector<StateId> order;
-  bool acyclic = false;
-  TopOrderVisitor<Arc> top_order_visitor(&order, &acyclic);
-  DfsVisit(fst, &top_order_visitor);
-  if (!acyclic) {
-    LOG(WARNING) << "ReplaceUtil::GetTopOrder: Cyclical label dependencies";
-    return false;
-  }
-  toporder->resize(order.size());
-  for (Label i = 0; i < order.size(); ++i) (*toporder)[order[i]] = i;
-  return true;
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::ReplaceLabels(const std::vector<Label> &labels) {
-  CheckMutableFsts();
-  std::unordered_set<Label> label_set;
-  for (const auto label : labels) {
-    // Can't replace root.
-    if (label != root_label_) label_set.insert(label);
-  }
-  // Finds FST dependencies restricted to the labels requested.
-  GetDependencies(false);
-  VectorFst<Arc> pfst(depfst_);
-  for (StateId i = 0; i < pfst.NumStates(); ++i) {
-    std::vector<Arc> arcs;
-    for (ArcIterator<VectorFst<Arc>> aiter(pfst, i); !aiter.Done();
-         aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto label = nonterminal_array_[arc.nextstate];
-      if (label_set.count(label) > 0) arcs.push_back(arc);
-    }
-    pfst.DeleteArcs(i);
-    for (auto &arc : arcs) pfst.AddArc(i, std::move(arc));
-  }
-  std::vector<Label> toporder;
-  if (!GetTopOrder(pfst, &toporder)) {
-    ClearDependencies();
-    return;
-  }
-  // Visits FSTs in reverse topological order of dependencies and performs
-  // replacements.
-  for (Label o = toporder.size() - 1; o >= 0; --o) {
-    std::vector<FstPair> fst_pairs;
-    auto s = toporder[o];
-    for (ArcIterator<VectorFst<Arc>> aiter(pfst, s); !aiter.Done();
-         aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto label = nonterminal_array_[arc.nextstate];
-      const auto *fst = fst_array_[arc.nextstate];
-      fst_pairs.push_back(std::make_pair(label, fst));
-    }
-    if (fst_pairs.empty()) continue;
-    const auto label = nonterminal_array_[s];
-    const auto *fst = fst_array_[s];
-    fst_pairs.push_back(std::make_pair(label, fst));
-    const ReplaceUtilOptions opts(label, call_label_type_, return_label_type_,
-                                  return_label_);
-    Replace(fst_pairs, mutable_fst_array_[s], opts);
-  }
-  ClearDependencies();
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::ReplaceBySize(size_t nstates, size_t narcs,
-                                     size_t nnonterms) {
-  std::vector<Label> labels;
-  GetDependencies(true);
-  std::vector<Label> toporder;
-  if (!GetTopOrder(depfst_, &toporder)) {
-    ClearDependencies();
-    return;
-  }
-  for (Label o = toporder.size() - 1; o >= 0; --o) {
-    const auto j = toporder[o];
-    if (stats_[j].nstates <= nstates && stats_[j].narcs <= narcs &&
-        stats_[j].nnonterms <= nnonterms) {
-      labels.push_back(nonterminal_array_[j]);
-      UpdateStats(j);
-    }
-  }
-  ReplaceLabels(labels);
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::ReplaceByInstances(size_t ninstances) {
-  std::vector<Label> labels;
-  GetDependencies(true);
-  std::vector<Label> toporder;
-  if (!GetTopOrder(depfst_, &toporder)) {
-    ClearDependencies();
-    return;
-  }
-  for (Label o = 0; o < toporder.size(); ++o) {
-    const auto j = toporder[o];
-    if (stats_[j].nref <= ninstances) {
-      labels.push_back(nonterminal_array_[j]);
-      UpdateStats(j);
-    }
-  }
-  ReplaceLabels(labels);
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::GetFstPairs(std::vector<FstPair> *fst_pairs) {
-  CheckMutableFsts();
-  fst_pairs->clear();
-  for (Label i = 0; i < fst_array_.size(); ++i) {
-    const auto label = nonterminal_array_[i];
-    const auto *fst = fst_array_[i];
-    if (!fst) continue;
-    fst_pairs->push_back(std::make_pair(label, fst));
-  }
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::GetMutableFstPairs(
-    std::vector<MutableFstPair> *mutable_fst_pairs) {
-  CheckMutableFsts();
-  mutable_fst_pairs->clear();
-  for (Label i = 0; i < mutable_fst_array_.size(); ++i) {
-    const auto label = nonterminal_array_[i];
-    const auto *fst = mutable_fst_array_[i];
-    if (!fst) continue;
-    mutable_fst_pairs->push_back(std::make_pair(label, fst->Copy()));
-  }
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::GetSCCProperties() const {
-  if (!depsccprops_.empty()) return;
-  GetDependencies(false);
-  if (depscc_.empty()) return;
-  for (StateId scc = 0; scc < depscc_.size(); ++scc) {
-    depsccprops_.push_back(kReplaceSCCLeftLinear | kReplaceSCCRightLinear);
-  }
-  if (!(depprops_ & kCyclic)) return;  // No cyclic dependencies.
-  // Checks for self-loops in the dependency graph.
-  for (StateId scc = 0; scc < depscc_.size(); ++scc) {
-    for (ArcIterator<Fst<Arc> > aiter(depfst_, scc);
-         !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      if (arc.nextstate == scc) {  // SCC has a self loop.
-        depsccprops_[scc] |= kReplaceSCCNonTrivial;
-      }
-    }
-  }
-  std::vector<bool> depscc_visited(depscc_.size(), false);
-  for (Label i = 0; i < fst_array_.size(); ++i) {
-    const auto *fst = fst_array_[i];
-    if (!fst) continue;
-    const auto depscc = depscc_[i];
-    if (depscc_visited[depscc]) {  // SCC has more than one state.
-      depsccprops_[depscc] |= kReplaceSCCNonTrivial;
-    }
-    depscc_visited[depscc] = true;
-    std::vector<StateId> fstscc;  // SCCs of the current FST.
-    uint64 fstprops;
-    SccVisitor<Arc> scc_visitor(&fstscc, nullptr, nullptr, &fstprops);
-    DfsVisit(*fst, &scc_visitor);
-    for (StateIterator<Fst<Arc>> siter(*fst); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      for (ArcIterator<Fst<Arc>> aiter(*fst, s); !aiter.Done(); aiter.Next()) {
-        const auto &arc = aiter.Value();
-        auto it = nonterminal_hash_.find(arc.olabel);
-        if (it == nonterminal_hash_.end() || depscc_[it->second] != depscc) {
-          continue;  // Skips if a terminal or a non-terminal not in SCC.
-        }
-        const bool arc_in_cycle = fstscc[s] == fstscc[arc.nextstate];
-        // Left linear iff all non-terminals are initial.
-        if (s != fst->Start() || arc_in_cycle) {
-          depsccprops_[depscc] &= ~kReplaceSCCLeftLinear;
-        }
-        // Right linear iff all non-terminals are final.
-        if (fst->Final(arc.nextstate) == Weight::Zero() || arc_in_cycle) {
-          depsccprops_[depscc] &= ~kReplaceSCCRightLinear;
-        }
-      }
-    }
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_REPLACE_UTIL_H_
diff --git a/projects/llm_framework/include/fst/replace.h b/projects/llm_framework/include/fst/replace.h
deleted file mode 100644
index 8b6b18ca..00000000
--- a/projects/llm_framework/include/fst/replace.h
+++ /dev/null
@@ -1,1492 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes for the recursive replacement of FSTs.
-
-#ifndef FST_REPLACE_H_
-#define FST_REPLACE_H_
-
-#include <set>
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/expanded-fst.h>
-#include <fst/fst-decl.h>  // For optional argument declarations.
-#include <fst/fst.h>
-#include <fst/matcher.h>
-#include <fst/replace-util.h>
-#include <fst/state-table.h>
-#include <fst/test-properties.h>
-
-namespace fst {
-
-// Replace state tables have the form:
-//
-// template <class Arc, class P>
-// class ReplaceStateTable {
-//  public:
-//   using Label = typename Arc::Label Label;
-//   using StateId = typename Arc::StateId;
-//
-//   using PrefixId = P;
-//   using StateTuple = ReplaceStateTuple<StateId, PrefixId>;
-//   using StackPrefix = ReplaceStackPrefix<Label, StateId>;
-//
-//   // Required constructor.
-//   ReplaceStateTable(
-//       const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_list,
-//       Label root);
-//
-//   // Required copy constructor that does not copy state.
-//   ReplaceStateTable(const ReplaceStateTable<Arc, PrefixId> &table);
-//
-//   // Looks up state ID by tuple, adding it if it doesn't exist.
-//   StateId FindState(const StateTuple &tuple);
-//
-//   // Looks up state tuple by ID.
-//   const StateTuple &Tuple(StateId id) const;
-//
-//   // Lookus up prefix ID by stack prefix, adding it if it doesn't exist.
-//   PrefixId FindPrefixId(const StackPrefix &stack_prefix);
-//
-//  // Looks up stack prefix by ID.
-//  const StackPrefix &GetStackPrefix(PrefixId id) const;
-// };
-
-// Tuple that uniquely defines a state in replace.
-template <class S, class P>
-struct ReplaceStateTuple {
-  using StateId = S;
-  using PrefixId = P;
-
-  ReplaceStateTuple(PrefixId prefix_id = -1, StateId fst_id = kNoStateId,
-                    StateId fst_state = kNoStateId)
-      : prefix_id(prefix_id), fst_id(fst_id), fst_state(fst_state) {}
-
-  PrefixId prefix_id;  // Index in prefix table.
-  StateId fst_id;      // Current FST being walked.
-  StateId fst_state;   // Current state in FST being walked (not to be
-                       // confused with the thse StateId of the combined FST).
-};
-
-// Equality of replace state tuples.
-template <class StateId, class PrefixId>
-inline bool operator==(const ReplaceStateTuple<StateId, PrefixId> &x,
-                       const ReplaceStateTuple<StateId, PrefixId> &y) {
-  return x.prefix_id == y.prefix_id && x.fst_id == y.fst_id &&
-         x.fst_state == y.fst_state;
-}
-
-// Functor returning true for tuples corresponding to states in the root FST.
-template <class StateId, class PrefixId>
-class ReplaceRootSelector {
- public:
-  bool operator()(const ReplaceStateTuple<StateId, PrefixId> &tuple) const {
-    return tuple.prefix_id == 0;
-  }
-};
-
-// Functor for fingerprinting replace state tuples.
-template <class StateId, class PrefixId>
-class ReplaceFingerprint {
- public:
-  explicit ReplaceFingerprint(const std::vector<uint64> *size_array)
-      : size_array_(size_array) {}
-
-  uint64 operator()(const ReplaceStateTuple<StateId, PrefixId> &tuple) const {
-    return tuple.prefix_id * size_array_->back() +
-           size_array_->at(tuple.fst_id - 1) + tuple.fst_state;
-  }
-
- private:
-  const std::vector<uint64> *size_array_;
-};
-
-// Useful when the fst_state uniquely define the tuple.
-template <class StateId, class PrefixId>
-class ReplaceFstStateFingerprint {
- public:
-  uint64 operator()(const ReplaceStateTuple<StateId, PrefixId> &tuple) const {
-    return tuple.fst_state;
-  }
-};
-
-// A generic hash function for replace state tuples.
-template <typename S, typename P>
-class ReplaceHash {
- public:
-  size_t operator()(const ReplaceStateTuple<S, P>& t) const {
-    static constexpr size_t prime0 = 7853;
-    static constexpr size_t prime1 = 7867;
-    return t.prefix_id + t.fst_id * prime0 + t.fst_state * prime1;
-  }
-};
-
-// Container for stack prefix.
-template <class Label, class StateId>
-class ReplaceStackPrefix {
- public:
-  struct PrefixTuple {
-    PrefixTuple(Label fst_id = kNoLabel, StateId nextstate = kNoStateId)
-        : fst_id(fst_id), nextstate(nextstate) {}
-
-    Label fst_id;
-    StateId nextstate;
-  };
-
-  ReplaceStackPrefix() {}
-
-  ReplaceStackPrefix(const ReplaceStackPrefix &other)
-      : prefix_(other.prefix_) {}
-
-  void Push(StateId fst_id, StateId nextstate) {
-    prefix_.push_back(PrefixTuple(fst_id, nextstate));
-  }
-
-  void Pop() { prefix_.pop_back(); }
-
-  const PrefixTuple &Top() const { return prefix_[prefix_.size() - 1]; }
-
-  size_t Depth() const { return prefix_.size(); }
-
- public:
-  std::vector<PrefixTuple> prefix_;
-};
-
-// Equality stack prefix classes.
-template <class Label, class StateId>
-inline bool operator==(const ReplaceStackPrefix<Label, StateId> &x,
-                       const ReplaceStackPrefix<Label, StateId> &y) {
-  if (x.prefix_.size() != y.prefix_.size()) return false;
-  for (size_t i = 0; i < x.prefix_.size(); ++i) {
-    if (x.prefix_[i].fst_id != y.prefix_[i].fst_id ||
-        x.prefix_[i].nextstate != y.prefix_[i].nextstate) {
-      return false;
-    }
-  }
-  return true;
-}
-
-// Hash function for stack prefix to prefix id.
-template <class Label, class StateId>
-class ReplaceStackPrefixHash {
- public:
-  size_t operator()(const ReplaceStackPrefix<Label, StateId> &prefix) const {
-    size_t sum = 0;
-    for (const auto &pair : prefix.prefix_) {
-      static constexpr size_t prime = 7863;
-      sum += pair.fst_id + pair.nextstate * prime;
-    }
-    return sum;
-  }
-};
-
-// Replace state tables.
-
-// A two-level state table for replace. Warning: calls CountStates to compute
-// the number of states of each component FST.
-template <class Arc, class P = ssize_t>
-class VectorHashReplaceStateTable {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-
-  using PrefixId = P;
-
-  using StateTuple = ReplaceStateTuple<StateId, PrefixId>;
-  using StateTable =
-      VectorHashStateTable<ReplaceStateTuple<StateId, PrefixId>,
-                           ReplaceRootSelector<StateId, PrefixId>,
-                           ReplaceFstStateFingerprint<StateId, PrefixId>,
-                           ReplaceFingerprint<StateId, PrefixId>>;
-  using StackPrefix = ReplaceStackPrefix<Label, StateId>;
-  using StackPrefixTable =
-      CompactHashBiTable<PrefixId, StackPrefix,
-                         ReplaceStackPrefixHash<Label, StateId>>;
-
-  VectorHashReplaceStateTable(
-      const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_list,
-      Label root)
-      : root_size_(0) {
-    size_array_.push_back(0);
-    for (const auto &fst_pair : fst_list) {
-      if (fst_pair.first == root) {
-        root_size_ = CountStates(*(fst_pair.second));
-        size_array_.push_back(size_array_.back());
-      } else {
-        size_array_.push_back(size_array_.back() +
-                              CountStates(*(fst_pair.second)));
-      }
-    }
-    state_table_.reset(
-        new StateTable(new ReplaceRootSelector<StateId, PrefixId>,
-                       new ReplaceFstStateFingerprint<StateId, PrefixId>,
-                       new ReplaceFingerprint<StateId, PrefixId>(&size_array_),
-                       root_size_, root_size_ + size_array_.back()));
-  }
-
-  VectorHashReplaceStateTable(
-      const VectorHashReplaceStateTable<Arc, PrefixId> &table)
-      : root_size_(table.root_size_),
-        size_array_(table.size_array_),
-        prefix_table_(table.prefix_table_) {
-    state_table_.reset(
-        new StateTable(new ReplaceRootSelector<StateId, PrefixId>,
-                       new ReplaceFstStateFingerprint<StateId, PrefixId>,
-                       new ReplaceFingerprint<StateId, PrefixId>(&size_array_),
-                       root_size_, root_size_ + size_array_.back()));
-  }
-
-  StateId FindState(const StateTuple &tuple) {
-    return state_table_->FindState(tuple);
-  }
-
-  const StateTuple &Tuple(StateId id) const { return state_table_->Tuple(id); }
-
-  PrefixId FindPrefixId(const StackPrefix &prefix) {
-    return prefix_table_.FindId(prefix);
-  }
-
-  const StackPrefix& GetStackPrefix(PrefixId id) const {
-    return prefix_table_.FindEntry(id);
-  }
-
- private:
-  StateId root_size_;
-  std::vector<uint64> size_array_;
-  std::unique_ptr<StateTable> state_table_;
-  StackPrefixTable prefix_table_;
-};
-
-// Default replace state table.
-template <class Arc, class P /* = size_t */>
-class DefaultReplaceStateTable
-    : public CompactHashStateTable<ReplaceStateTuple<typename Arc::StateId, P>,
-                                   ReplaceHash<typename Arc::StateId, P>> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-
-  using PrefixId = P;
-  using StateTuple = ReplaceStateTuple<StateId, PrefixId>;
-  using StateTable =
-      CompactHashStateTable<StateTuple, ReplaceHash<StateId, PrefixId>>;
-  using StackPrefix = ReplaceStackPrefix<Label, StateId>;
-  using StackPrefixTable =
-      CompactHashBiTable<PrefixId, StackPrefix,
-                         ReplaceStackPrefixHash<Label, StateId>>;
-
-  using StateTable::FindState;
-  using StateTable::Tuple;
-
-  DefaultReplaceStateTable(
-      const std::vector<std::pair<Label, const Fst<Arc> *>> &, Label) {}
-
-  DefaultReplaceStateTable(const DefaultReplaceStateTable<Arc, PrefixId> &table)
-      : StateTable(), prefix_table_(table.prefix_table_) {}
-
-  PrefixId FindPrefixId(const StackPrefix &prefix) {
-    return prefix_table_.FindId(prefix);
-  }
-
-  const StackPrefix &GetStackPrefix(PrefixId id) const {
-    return prefix_table_.FindEntry(id);
-  }
-
- private:
-  StackPrefixTable prefix_table_;
-};
-
-// By default ReplaceFst will copy the input label of the replace arc.
-// The call_label_type and return_label_type options specify how to manage
-// the labels of the call arc and the return arc of the replace FST
-template <class Arc, class StateTable = DefaultReplaceStateTable<Arc>,
-          class CacheStore = DefaultCacheStore<Arc>>
-struct ReplaceFstOptions : CacheImplOptions<CacheStore> {
-  using Label = typename Arc::Label;
-
-  // Index of root rule for expansion.
-  Label root;
-  // How to label call arc.
-  ReplaceLabelType call_label_type = REPLACE_LABEL_INPUT;
-  // How to label return arc.
-  ReplaceLabelType return_label_type = REPLACE_LABEL_NEITHER;
-  // Specifies output label to put on call arc; if kNoLabel, use existing label
-  // on call arc. Otherwise, use this field as the output label.
-  Label call_output_label = kNoLabel;
-  // Specifies label to put on return arc.
-  Label return_label = 0;
-  // Take ownership of input FSTs?
-  bool take_ownership = false;
-  // Pointer to optional pre-constructed state table.
-  StateTable *state_table = nullptr;
-
-  explicit ReplaceFstOptions(const CacheImplOptions<CacheStore> &opts,
-                             Label root = kNoLabel)
-      : CacheImplOptions<CacheStore>(opts), root(root) {}
-
-  explicit ReplaceFstOptions(const CacheOptions &opts, Label root = kNoLabel)
-      : CacheImplOptions<CacheStore>(opts), root(root) {}
-
-  // FIXME(kbg): There are too many constructors here. Come up with a consistent
-  // position for call_output_label (probably the very end) so that it is
-  // possible to express all the remaining constructors with a single
-  // default-argument constructor. Also move clients off of the "backwards
-  // compatibility" constructor, for good.
-
-  explicit ReplaceFstOptions(Label root) : root(root) {}
-
-  explicit ReplaceFstOptions(Label root, ReplaceLabelType call_label_type,
-                             ReplaceLabelType return_label_type,
-                             Label return_label)
-      : root(root),
-        call_label_type(call_label_type),
-        return_label_type(return_label_type),
-        return_label(return_label) {}
-
-  explicit ReplaceFstOptions(Label root, ReplaceLabelType call_label_type,
-                             ReplaceLabelType return_label_type,
-                             Label call_output_label, Label return_label)
-      : root(root),
-        call_label_type(call_label_type),
-        return_label_type(return_label_type),
-        call_output_label(call_output_label),
-        return_label(return_label) {}
-
-  explicit ReplaceFstOptions(const ReplaceUtilOptions &opts)
-      : ReplaceFstOptions(opts.root, opts.call_label_type,
-                          opts.return_label_type, opts.return_label) {}
-
-  ReplaceFstOptions() : root(kNoLabel) {}
-
-  // For backwards compatibility.
-  ReplaceFstOptions(int64 root, bool epsilon_replace_arc)
-      : root(root),
-        call_label_type(epsilon_replace_arc ? REPLACE_LABEL_NEITHER
-                                            : REPLACE_LABEL_INPUT),
-        call_output_label(epsilon_replace_arc ? 0 : kNoLabel) {}
-};
-
-
-// Forward declaration.
-template <class Arc, class StateTable, class CacheStore>
-class ReplaceFstMatcher;
-
-template <class Arc>
-using FstList = std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>;
-
-// Returns true if label type on arc results in epsilon input label.
-inline bool EpsilonOnInput(ReplaceLabelType label_type) {
-  return label_type == REPLACE_LABEL_NEITHER ||
-         label_type == REPLACE_LABEL_OUTPUT;
-}
-
-// Returns true if label type on arc results in epsilon input label.
-inline bool EpsilonOnOutput(ReplaceLabelType label_type) {
-  return label_type == REPLACE_LABEL_NEITHER ||
-         label_type == REPLACE_LABEL_INPUT;
-}
-
-// Returns true if for either the call or return arc ilabel != olabel.
-template <class Label>
-bool ReplaceTransducer(ReplaceLabelType call_label_type,
-                       ReplaceLabelType return_label_type,
-                       Label call_output_label) {
-  return call_label_type == REPLACE_LABEL_INPUT ||
-         call_label_type == REPLACE_LABEL_OUTPUT ||
-         (call_label_type == REPLACE_LABEL_BOTH &&
-          call_output_label != kNoLabel) ||
-         return_label_type == REPLACE_LABEL_INPUT ||
-         return_label_type == REPLACE_LABEL_OUTPUT;
-}
-
-template <class Arc>
-uint64 ReplaceFstProperties(typename Arc::Label root_label,
-                            const FstList<Arc> &fst_list,
-                            ReplaceLabelType call_label_type,
-                            ReplaceLabelType return_label_type,
-                            typename Arc::Label call_output_label,
-                            bool *sorted_and_non_empty) {
-  using Label = typename Arc::Label;
-  std::vector<uint64> inprops;
-  bool all_ilabel_sorted = true;
-  bool all_olabel_sorted = true;
-  bool all_non_empty = true;
-  // All nonterminals are negative?
-  bool all_negative = true;
-  // All nonterminals are positive and form a dense range containing 1?
-  bool dense_range = true;
-  Label root_fst_idx = 0;
-  for (Label i = 0; i < fst_list.size(); ++i) {
-    const auto label = fst_list[i].first;
-    if (label >= 0) all_negative = false;
-    if (label > fst_list.size() || label <= 0) dense_range = false;
-    if (label == root_label) root_fst_idx = i;
-    const auto *fst = fst_list[i].second;
-    if (fst->Start() == kNoStateId) all_non_empty = false;
-    if (!fst->Properties(kILabelSorted, false)) all_ilabel_sorted = false;
-    if (!fst->Properties(kOLabelSorted, false)) all_olabel_sorted = false;
-    inprops.push_back(fst->Properties(kCopyProperties, false));
-  }
-  const auto props = ReplaceProperties(
-      inprops, root_fst_idx, EpsilonOnInput(call_label_type),
-      EpsilonOnInput(return_label_type), EpsilonOnOutput(call_label_type),
-      EpsilonOnOutput(return_label_type),
-      ReplaceTransducer(call_label_type, return_label_type, call_output_label),
-      all_non_empty, all_ilabel_sorted, all_olabel_sorted,
-      all_negative || dense_range);
-  const bool sorted = props & (kILabelSorted | kOLabelSorted);
-  *sorted_and_non_empty = all_non_empty && sorted;
-  return props;
-}
-
-namespace internal {
-
-// The replace implementation class supports a dynamic expansion of a recursive
-// transition network represented as label/FST pairs with dynamic replacable
-// arcs.
-template <class Arc, class StateTable, class CacheStore>
-class ReplaceFstImpl
-    : public CacheBaseImpl<typename CacheStore::State, CacheStore> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using State = typename CacheStore::State;
-  using CacheImpl = CacheBaseImpl<State, CacheStore>;
-  using PrefixId = typename StateTable::PrefixId;
-  using StateTuple = ReplaceStateTuple<StateId, PrefixId>;
-  using StackPrefix = ReplaceStackPrefix<Label, StateId>;
-  using NonTerminalHash = std::unordered_map<Label, Label>;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::WriteHeader;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-  using FstImpl<Arc>::InputSymbols;
-  using FstImpl<Arc>::OutputSymbols;
-
-  using CacheImpl::PushArc;
-  using CacheImpl::HasArcs;
-  using CacheImpl::HasFinal;
-  using CacheImpl::HasStart;
-  using CacheImpl::SetArcs;
-  using CacheImpl::SetFinal;
-  using CacheImpl::SetStart;
-
-  friend class ReplaceFstMatcher<Arc, StateTable, CacheStore>;
-
-  ReplaceFstImpl(const FstList<Arc> &fst_list,
-                 const ReplaceFstOptions<Arc, StateTable, CacheStore> &opts)
-      : CacheImpl(opts),
-        call_label_type_(opts.call_label_type),
-        return_label_type_(opts.return_label_type),
-        call_output_label_(opts.call_output_label),
-        return_label_(opts.return_label),
-        state_table_(opts.state_table ? opts.state_table
-                                      : new StateTable(fst_list, opts.root)) {
-    SetType("replace");
-    // If the label is epsilon, then all replace label options are equivalent,
-    // so we set the label types to NEITHER for simplicity.
-    if (call_output_label_ == 0) call_label_type_ = REPLACE_LABEL_NEITHER;
-    if (return_label_ == 0) return_label_type_ = REPLACE_LABEL_NEITHER;
-    if (!fst_list.empty()) {
-      SetInputSymbols(fst_list[0].second->InputSymbols());
-      SetOutputSymbols(fst_list[0].second->OutputSymbols());
-    }
-    fst_array_.push_back(nullptr);
-    for (Label i = 0; i < fst_list.size(); ++i) {
-      const auto label = fst_list[i].first;
-      const auto *fst = fst_list[i].second;
-      nonterminal_hash_[label] = fst_array_.size();
-      nonterminal_set_.insert(label);
-      fst_array_.emplace_back(opts.take_ownership ? fst : fst->Copy());
-      if (i) {
-        if (!CompatSymbols(InputSymbols(), fst->InputSymbols())) {
-          FSTERROR() << "ReplaceFstImpl: Input symbols of FST " << i
-                     << " do not match input symbols of base FST (0th FST)";
-          SetProperties(kError, kError);
-        }
-        if (!CompatSymbols(OutputSymbols(), fst->OutputSymbols())) {
-          FSTERROR() << "ReplaceFstImpl: Output symbols of FST " << i
-                     << " do not match output symbols of base FST (0th FST)";
-          SetProperties(kError, kError);
-        }
-      }
-    }
-    const auto nonterminal = nonterminal_hash_[opts.root];
-    if ((nonterminal == 0) && (fst_array_.size() > 1)) {
-      FSTERROR() << "ReplaceFstImpl: No FST corresponding to root label "
-                 << opts.root << " in the input tuple vector";
-      SetProperties(kError, kError);
-    }
-    root_ = (nonterminal > 0) ? nonterminal : 1;
-    bool all_non_empty_and_sorted = false;
-    SetProperties(ReplaceFstProperties(opts.root, fst_list, call_label_type_,
-                                       return_label_type_, call_output_label_,
-                                       &all_non_empty_and_sorted));
-    // Enables optional caching as long as sorted and all non-empty.
-    always_cache_ = !all_non_empty_and_sorted;
-    VLOG(2) << "ReplaceFstImpl::ReplaceFstImpl: always_cache = "
-            << (always_cache_ ? "true" : "false");
-  }
-
-  ReplaceFstImpl(const ReplaceFstImpl &impl)
-      : CacheImpl(impl),
-        call_label_type_(impl.call_label_type_),
-        return_label_type_(impl.return_label_type_),
-        call_output_label_(impl.call_output_label_),
-        return_label_(impl.return_label_),
-        always_cache_(impl.always_cache_),
-        state_table_(new StateTable(*(impl.state_table_))),
-        nonterminal_set_(impl.nonterminal_set_),
-        nonterminal_hash_(impl.nonterminal_hash_),
-        root_(impl.root_) {
-    SetType("replace");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-    fst_array_.reserve(impl.fst_array_.size());
-    fst_array_.emplace_back(nullptr);
-    for (Label i = 1; i < impl.fst_array_.size(); ++i) {
-      fst_array_.emplace_back(impl.fst_array_[i]->Copy(true));
-    }
-  }
-
-  // Computes the dependency graph of the replace class and returns
-  // true if the dependencies are cyclic. Cyclic dependencies will result
-  // in an un-expandable FST.
-  bool CyclicDependencies() const {
-    const ReplaceUtilOptions opts(root_);
-    ReplaceUtil<Arc> replace_util(fst_array_, nonterminal_hash_, opts);
-    return replace_util.CyclicDependencies();
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      if (fst_array_.size() == 1) {
-        SetStart(kNoStateId);
-        return kNoStateId;
-      } else {
-        const auto fst_start = fst_array_[root_]->Start();
-        if (fst_start == kNoStateId) return kNoStateId;
-        const auto prefix = GetPrefixId(StackPrefix());
-        const auto start =
-            state_table_->FindState(StateTuple(prefix, root_, fst_start));
-        SetStart(start);
-        return start;
-      }
-    } else {
-      return CacheImpl::Start();
-    }
-  }
-
-  Weight Final(StateId s) {
-    if (HasFinal(s)) return CacheImpl::Final(s);
-    const auto &tuple = state_table_->Tuple(s);
-    auto weight = Weight::Zero();
-    if (tuple.prefix_id == 0) {
-      const auto fst_state = tuple.fst_state;
-      weight = fst_array_[tuple.fst_id]->Final(fst_state);
-    }
-    if (always_cache_ || HasArcs(s)) SetFinal(s, weight);
-    return weight;
-  }
-
-  size_t NumArcs(StateId s) {
-    if (HasArcs(s)) {
-      return CacheImpl::NumArcs(s);
-    } else if (always_cache_) {  // If always caching, expands and caches state.
-      Expand(s);
-      return CacheImpl::NumArcs(s);
-    } else {  // Otherwise computes the number of arcs without expanding.
-      const auto tuple = state_table_->Tuple(s);
-      if (tuple.fst_state == kNoStateId) return 0;
-      auto num_arcs = fst_array_[tuple.fst_id]->NumArcs(tuple.fst_state);
-      if (ComputeFinalArc(tuple, nullptr)) ++num_arcs;
-      return num_arcs;
-    }
-  }
-
-  // Returns whether a given label is a non-terminal.
-  bool IsNonTerminal(Label label) const {
-    if (label < *nonterminal_set_.begin() ||
-        label > *nonterminal_set_.rbegin()) {
-      return false;
-    } else {
-      return nonterminal_hash_.count(label);
-    }
-    // TODO(allauzen): be smarter and take advantage of all_dense or
-    // all_negative. Also use this in ComputeArc. This would require changes to
-    // Replace so that recursing into an empty FST lead to a non co-accessible
-    // state instead of deleting the arc as done currently. The current use
-    // correct, since labels are sorted if all_non_empty is true.
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (HasArcs(s)) {
-      return CacheImpl::NumInputEpsilons(s);
-    } else if (always_cache_ || !Properties(kILabelSorted)) {
-      // If always caching or if the number of input epsilons is too expensive
-      // to compute without caching (i.e., not ilabel-sorted), then expands and
-      // caches state.
-      Expand(s);
-      return CacheImpl::NumInputEpsilons(s);
-    } else {
-      // Otherwise, computes the number of input epsilons without caching.
-      const auto tuple = state_table_->Tuple(s);
-      if (tuple.fst_state == kNoStateId) return 0;
-      size_t num = 0;
-      if (!EpsilonOnInput(call_label_type_)) {
-        // If EpsilonOnInput(c) is false, all input epsilon arcs
-        // are also input epsilons arcs in the underlying machine.
-        num = fst_array_[tuple.fst_id]->NumInputEpsilons(tuple.fst_state);
-      } else {
-        // Otherwise, one need to consider that all non-terminal arcs
-        // in the underlying machine also become input epsilon arc.
-        ArcIterator<Fst<Arc>> aiter(*fst_array_[tuple.fst_id], tuple.fst_state);
-        for (; !aiter.Done() && ((aiter.Value().ilabel == 0) ||
-                                 IsNonTerminal(aiter.Value().olabel));
-             aiter.Next()) {
-          ++num;
-        }
-      }
-      if (EpsilonOnInput(return_label_type_) &&
-          ComputeFinalArc(tuple, nullptr)) {
-        ++num;
-      }
-      return num;
-    }
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (HasArcs(s)) {
-      return CacheImpl::NumOutputEpsilons(s);
-    } else if (always_cache_ || !Properties(kOLabelSorted)) {
-      // If always caching or if the number of output epsilons is too expensive
-      // to compute without caching (i.e., not olabel-sorted), then expands and
-      // caches state.
-      Expand(s);
-      return CacheImpl::NumOutputEpsilons(s);
-    } else {
-      // Otherwise, computes the number of output epsilons without caching.
-      const auto tuple = state_table_->Tuple(s);
-      if (tuple.fst_state == kNoStateId) return 0;
-      size_t num = 0;
-      if (!EpsilonOnOutput(call_label_type_)) {
-        // If EpsilonOnOutput(c) is false, all output epsilon arcs are also
-        // output epsilons arcs in the underlying machine.
-        num = fst_array_[tuple.fst_id]->NumOutputEpsilons(tuple.fst_state);
-      } else {
-        // Otherwise, one need to consider that all non-terminal arcs in the
-        // underlying machine also become output epsilon arc.
-        ArcIterator<Fst<Arc>> aiter(*fst_array_[tuple.fst_id], tuple.fst_state);
-        for (; !aiter.Done() && ((aiter.Value().olabel == 0) ||
-                                 IsNonTerminal(aiter.Value().olabel));
-             aiter.Next()) {
-          ++num;
-        }
-      }
-      if (EpsilonOnOutput(return_label_type_) &&
-          ComputeFinalArc(tuple, nullptr)) {
-        ++num;
-      }
-      return num;
-    }
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if (mask & kError) {
-      for (Label i = 1; i < fst_array_.size(); ++i) {
-        if (fst_array_[i]->Properties(kError, false)) {
-          SetProperties(kError, kError);
-        }
-      }
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  // Returns the base arc iterator, and if arcs have not been computed yet,
-  // extends and recurses for new arcs.
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl::InitArcIterator(s, data);
-    // TODO(allauzen): Set behaviour of generic iterator.
-    // Warning: ArcIterator<ReplaceFst<A>>::InitCache() relies on current
-    // behaviour.
-  }
-
-  // Extends current state (walk arcs one level deep).
-  void Expand(StateId s) {
-    const auto tuple = state_table_->Tuple(s);
-    if (tuple.fst_state == kNoStateId) {  // Local FST is empty.
-      SetArcs(s);
-      return;
-    }
-    ArcIterator<Fst<Arc>> aiter(*fst_array_[tuple.fst_id], tuple.fst_state);
-    Arc arc;
-    // Creates a final arc when needed.
-    if (ComputeFinalArc(tuple, &arc)) PushArc(s, std::move(arc));
-    // Expands all arcs leaving the state.
-    for (; !aiter.Done(); aiter.Next()) {
-      if (ComputeArc(tuple, aiter.Value(), &arc)) PushArc(s, std::move(arc));
-    }
-    SetArcs(s);
-  }
-
-  void Expand(StateId s, const StateTuple &tuple,
-              const ArcIteratorData<Arc> &data) {
-    if (tuple.fst_state == kNoStateId) {  // Local FST is empty.
-      SetArcs(s);
-      return;
-    }
-    ArcIterator<Fst<Arc>> aiter(data);
-    Arc arc;
-    // Creates a final arc when needed.
-    if (ComputeFinalArc(tuple, &arc)) AddArc(s, arc);
-    // Expands all arcs leaving the state.
-    for (; !aiter.Done(); aiter.Next()) {
-      if (ComputeArc(tuple, aiter.Value(), &arc)) AddArc(s, arc);
-    }
-    SetArcs(s);
-  }
-
-  // If acpp is null, only returns true if a final arcp is required, but does
-  // not actually compute it.
-  bool ComputeFinalArc(const StateTuple &tuple, Arc *arcp,
-                       uint32 flags = kArcValueFlags) {
-    const auto fst_state = tuple.fst_state;
-    if (fst_state == kNoStateId) return false;
-    // If state is final, pops the stack.
-    if (fst_array_[tuple.fst_id]->Final(fst_state) != Weight::Zero() &&
-        tuple.prefix_id) {
-      if (arcp) {
-        arcp->ilabel = (EpsilonOnInput(return_label_type_)) ? 0 : return_label_;
-        arcp->olabel =
-            (EpsilonOnOutput(return_label_type_)) ? 0 : return_label_;
-        if (flags & kArcNextStateValue) {
-          const auto &stack = state_table_->GetStackPrefix(tuple.prefix_id);
-          const auto prefix_id = PopPrefix(stack);
-          const auto &top = stack.Top();
-          arcp->nextstate = state_table_->FindState(
-              StateTuple(prefix_id, top.fst_id, top.nextstate));
-        }
-        if (flags & kArcWeightValue) {
-          arcp->weight = fst_array_[tuple.fst_id]->Final(fst_state);
-        }
-      }
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  // Computes an arc in the FST corresponding to one in the underlying machine.
-  // Returns false if the underlying arc corresponds to no arc in the resulting
-  // FST.
-  bool ComputeArc(const StateTuple &tuple, const Arc &arc, Arc *arcp,
-                  uint32 flags = kArcValueFlags) {
-    if (!EpsilonOnInput(call_label_type_) &&
-        (flags == (flags & (kArcILabelValue | kArcWeightValue)))) {
-      *arcp = arc;
-      return true;
-    }
-    if (arc.olabel == 0 || arc.olabel < *nonterminal_set_.begin() ||
-        arc.olabel > *nonterminal_set_.rbegin()) {  // Expands local FST.
-      const auto nextstate =
-          flags & kArcNextStateValue
-              ? state_table_->FindState(
-                    StateTuple(tuple.prefix_id, tuple.fst_id, arc.nextstate))
-              : kNoStateId;
-      *arcp = Arc(arc.ilabel, arc.olabel, arc.weight, nextstate);
-    } else {
-      // Checks for non-terminal.
-      const auto it = nonterminal_hash_.find(arc.olabel);
-      if (it != nonterminal_hash_.end()) {  // Recurses into non-terminal.
-        const auto nonterminal = it->second;
-        const auto nt_prefix =
-            PushPrefix(state_table_->GetStackPrefix(tuple.prefix_id),
-                       tuple.fst_id, arc.nextstate);
-        // If the start state is valid, replace; othewise, the arc is implicitly
-        // deleted.
-        const auto nt_start = fst_array_[nonterminal]->Start();
-        if (nt_start != kNoStateId) {
-          const auto nt_nextstate = flags & kArcNextStateValue
-                                        ? state_table_->FindState(StateTuple(
-                                              nt_prefix, nonterminal, nt_start))
-                                        : kNoStateId;
-          const auto ilabel =
-              (EpsilonOnInput(call_label_type_)) ? 0 : arc.ilabel;
-          const auto olabel =
-              (EpsilonOnOutput(call_label_type_))
-                  ? 0
-                  : ((call_output_label_ == kNoLabel) ? arc.olabel
-                                                      : call_output_label_);
-          *arcp = Arc(ilabel, olabel, arc.weight, nt_nextstate);
-        } else {
-          return false;
-        }
-      } else {
-        const auto nextstate =
-            flags & kArcNextStateValue
-                ? state_table_->FindState(
-                      StateTuple(tuple.prefix_id, tuple.fst_id, arc.nextstate))
-                : kNoStateId;
-        *arcp = Arc(arc.ilabel, arc.olabel, arc.weight, nextstate);
-      }
-    }
-    return true;
-  }
-
-  // Returns the arc iterator flags supported by this FST.
-  uint32 ArcIteratorFlags() const {
-    uint32 flags = kArcValueFlags;
-    if (!always_cache_) flags |= kArcNoCache;
-    return flags;
-  }
-
-  StateTable *GetStateTable() const { return state_table_.get(); }
-
-  const Fst<Arc> *GetFst(Label fst_id) const {
-    return fst_array_[fst_id].get();
-  }
-
-  Label GetFstId(Label nonterminal) const {
-    const auto it = nonterminal_hash_.find(nonterminal);
-    if (it == nonterminal_hash_.end()) {
-      FSTERROR() << "ReplaceFstImpl::GetFstId: Nonterminal not found: "
-                 << nonterminal;
-    }
-    return it->second;
-  }
-
-  // Returns true if label type on call arc results in epsilon input label.
-  bool EpsilonOnCallInput() { return EpsilonOnInput(call_label_type_); }
-
- private:
-  // The unique index into stack prefix table.
-  PrefixId GetPrefixId(const StackPrefix &prefix) {
-    return state_table_->FindPrefixId(prefix);
-  }
-
-  // The prefix ID after a stack pop.
-  PrefixId PopPrefix(StackPrefix prefix) {
-    prefix.Pop();
-    return GetPrefixId(prefix);
-  }
-
-  // The prefix ID after a stack push.
-  PrefixId PushPrefix(StackPrefix prefix, Label fst_id, StateId nextstate) {
-    prefix.Push(fst_id, nextstate);
-    return GetPrefixId(prefix);
-  }
-
-  // Runtime options
-  ReplaceLabelType call_label_type_;    // How to label call arc.
-  ReplaceLabelType return_label_type_;  // How to label return arc.
-  int64 call_output_label_;  // Specifies output label to put on call arc
-  int64 return_label_;       // Specifies label to put on return arc.
-  bool always_cache_;        // Disable optional caching of arc iterator?
-
-  // State table.
-  std::unique_ptr<StateTable> state_table_;
-
-  // Replace components.
-  std::set<Label> nonterminal_set_;
-  NonTerminalHash nonterminal_hash_;
-  std::vector<std::unique_ptr<const Fst<Arc>>> fst_array_;
-  Label root_;
-};
-
-}  // namespace internal
-
-//
-// ReplaceFst supports dynamic replacement of arcs in one FST with another FST.
-// This replacement is recursive. ReplaceFst can be used to support a variety of
-// delayed constructions such as recursive
-// transition networks, union, or closure. It is constructed with an array of
-// FST(s). One FST represents the root (or topology) machine. The root FST
-// refers to other FSTs by recursively replacing arcs labeled as non-terminals
-// with the matching non-terminal FST. Currently the ReplaceFst uses the output
-// symbols of the arcs to determine whether the arc is a non-terminal arc or
-// not. A non-terminal can be any label that is not a non-zero terminal label in
-// the output alphabet.
-//
-// Note that the constructor uses a vector of pairs. These correspond to the
-// tuple of non-terminal Label and corresponding FST. For example to implement
-// the closure operation we need 2 FSTs. The first root FST is a single
-// self-loop arc on the start state.
-//
-// The ReplaceFst class supports an optionally caching arc iterator.
-//
-// The ReplaceFst needs to be built such that it is known to be ilabel- or
-// olabel-sorted (see usage below).
-//
-// Observe that Matcher<Fst<A>> will use the optionally caching arc iterator
-// when available (the FST is ilabel-sorted and matching on the input, or the
-// FST is olabel -orted and matching on the output).  In order to obtain the
-// most efficient behaviour, it is recommended to set call_label_type to
-// REPLACE_LABEL_INPUT or REPLACE_LABEL_BOTH and return_label_type to
-// REPLACE_LABEL_OUTPUT or REPLACE_LABEL_NEITHER. This means that the call arc
-// does not have epsilon on the input side and the return arc has epsilon on the
-// input side) and matching on the input side.
-//
-// This class attaches interface to implementation and handles reference
-// counting, delegating most methods to ImplToFst.
-template <class A, class T /* = DefaultReplaceStateTable<A> */,
-          class CacheStore /* = DefaultCacheStore<A> */>
-class ReplaceFst
-    : public ImplToFst<internal::ReplaceFstImpl<A, T, CacheStore>> {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using StateTable = T;
-  using Store = CacheStore;
-  using State = typename CacheStore::State;
-  using Impl = internal::ReplaceFstImpl<Arc, StateTable, CacheStore>;
-  using CacheImpl = internal::CacheBaseImpl<State, CacheStore>;
-
-  using ImplToFst<Impl>::Properties;
-
-  friend class ArcIterator<ReplaceFst<Arc, StateTable, CacheStore>>;
-  friend class StateIterator<ReplaceFst<Arc, StateTable, CacheStore>>;
-  friend class ReplaceFstMatcher<Arc, StateTable, CacheStore>;
-
-  ReplaceFst(const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_array,
-             Label root)
-      : ImplToFst<Impl>(std::make_shared<Impl>(
-            fst_array, ReplaceFstOptions<Arc, StateTable, CacheStore>(root))) {}
-
-  ReplaceFst(const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_array,
-             const ReplaceFstOptions<Arc, StateTable, CacheStore> &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst_array, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  ReplaceFst(const ReplaceFst<Arc, StateTable, CacheStore> &fst,
-             bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this ReplaceFst. See Fst<>::Copy() for further doc.
-  ReplaceFst<Arc, StateTable, CacheStore> *Copy(
-      bool safe = false) const override {
-    return new ReplaceFst<Arc, StateTable, CacheStore>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
-  MatcherBase<Arc> *InitMatcher(MatchType match_type) const override {
-    if ((GetImpl()->ArcIteratorFlags() & kArcNoCache) &&
-        ((match_type == MATCH_INPUT && Properties(kILabelSorted, false)) ||
-         (match_type == MATCH_OUTPUT && Properties(kOLabelSorted, false)))) {
-      return new ReplaceFstMatcher<Arc, StateTable, CacheStore>
-          (this, match_type);
-    } else {
-      VLOG(2) << "Not using replace matcher";
-      return nullptr;
-    }
-  }
-
-  bool CyclicDependencies() const { return GetImpl()->CyclicDependencies(); }
-
-  const StateTable &GetStateTable() const {
-    return *GetImpl()->GetStateTable();
-  }
-
-  const Fst<Arc> &GetFst(Label nonterminal) const {
-    return *GetImpl()->GetFst(GetImpl()->GetFstId(nonterminal));
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  ReplaceFst &operator=(const ReplaceFst &) = delete;
-};
-
-// Specialization for ReplaceFst.
-template <class Arc, class StateTable, class CacheStore>
-class StateIterator<ReplaceFst<Arc, StateTable, CacheStore>>
-    : public CacheStateIterator<ReplaceFst<Arc, StateTable, CacheStore>> {
- public:
-  explicit StateIterator(const ReplaceFst<Arc, StateTable, CacheStore> &fst)
-      : CacheStateIterator<ReplaceFst<Arc, StateTable, CacheStore>>(
-            fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for ReplaceFst, implementing optional caching. It is be used
-// as follows:
-//
-//   ReplaceFst<A> replace;
-//   ArcIterator<ReplaceFst<A>> aiter(replace, s);
-//   // Note: ArcIterator< Fst<A>> is always a caching arc iterator.
-//   aiter.SetFlags(kArcNoCache, kArcNoCache);
-//   // Uses the arc iterator, no arc will be cached, no state will be expanded.
-//   // Arc flags can be used to decide which component of the arc need to be
-//   computed.
-//   aiter.SetFlags(kArcILabelValue, kArcValueFlags);
-//   // Wants the ilabel for this arc.
-//   aiter.Value();  // Does not compute the destination state.
-//   aiter.Next();
-//   aiter.SetFlags(kArcNextStateValue, kArcNextStateValue);
-//   // Wants the ilabel and next state for this arc.
-//   aiter.Value();  // Does compute the destination state and inserts it
-//                   // in the replace state table.
-//   // No additional arcs have been cached at this point.
-template <class Arc, class StateTable, class CacheStore>
-class ArcIterator<ReplaceFst<Arc, StateTable, CacheStore>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  using StateTuple = typename StateTable::StateTuple;
-
-  ArcIterator(const ReplaceFst<Arc, StateTable, CacheStore> &fst, StateId s)
-      : fst_(fst),
-        s_(s),
-        pos_(0),
-        offset_(0),
-        flags_(kArcValueFlags),
-        arcs_(nullptr),
-        data_flags_(0),
-        final_flags_(0) {
-    cache_data_.ref_count = nullptr;
-    local_data_.ref_count = nullptr;
-    // If FST does not support optional caching, forces caching.
-    if (!(fst_.GetImpl()->ArcIteratorFlags() & kArcNoCache) &&
-        !(fst_.GetImpl()->HasArcs(s_))) {
-      fst_.GetMutableImpl()->Expand(s_);
-    }
-    // If state is already cached, use cached arcs array.
-    if (fst_.GetImpl()->HasArcs(s_)) {
-      (fst_.GetImpl())
-          ->internal::template CacheBaseImpl<
-              typename CacheStore::State,
-              CacheStore>::InitArcIterator(s_, &cache_data_);
-      num_arcs_ = cache_data_.narcs;
-      arcs_ = cache_data_.arcs;      // arcs_ is a pointer to the cached arcs.
-      data_flags_ = kArcValueFlags;  // All the arc member values are valid.
-    } else {  // Otherwise delay decision until Value() is called.
-      tuple_ = fst_.GetImpl()->GetStateTable()->Tuple(s_);
-      if (tuple_.fst_state == kNoStateId) {
-        num_arcs_ = 0;
-      } else {
-        // The decision to cache or not to cache has been defered until Value()
-        // or
-        // SetFlags() is called. However, the arc iterator is set up now to be
-        // ready for non-caching in order to keep the Value() method simple and
-        // efficient.
-        const auto *rfst = fst_.GetImpl()->GetFst(tuple_.fst_id);
-        rfst->InitArcIterator(tuple_.fst_state, &local_data_);
-        // arcs_ is a pointer to the arcs in the underlying machine.
-        arcs_ = local_data_.arcs;
-        // Computes the final arc (but not its destination state) if a final arc
-        // is required.
-        bool has_final_arc = fst_.GetMutableImpl()->ComputeFinalArc(
-            tuple_, &final_arc_, kArcValueFlags & ~kArcNextStateValue);
-        // Sets the arc value flags that hold for final_arc_.
-        final_flags_ = kArcValueFlags & ~kArcNextStateValue;
-        // Computes the number of arcs.
-        num_arcs_ = local_data_.narcs;
-        if (has_final_arc) ++num_arcs_;
-        // Sets the offset between the underlying arc positions and the
-        // positions
-        // in the arc iterator.
-        offset_ = num_arcs_ - local_data_.narcs;
-        // Defers the decision to cache or not until Value() or SetFlags() is
-        // called.
-        data_flags_ = 0;
-      }
-    }
-  }
-
-  ~ArcIterator() {
-    if (cache_data_.ref_count) --(*cache_data_.ref_count);
-    if (local_data_.ref_count) --(*local_data_.ref_count);
-  }
-
-  void ExpandAndCache() const  {
-    // TODO(allauzen): revisit this.
-    // fst_.GetImpl()->Expand(s_, tuple_, local_data_);
-    // (fst_.GetImpl())->CacheImpl<A>*>::InitArcIterator(s_,
-    //                                               &cache_data_);
-    //
-    fst_.InitArcIterator(s_, &cache_data_);  // Expand and cache state.
-    arcs_ = cache_data_.arcs;      // arcs_ is a pointer to the cached arcs.
-    data_flags_ = kArcValueFlags;  // All the arc member values are valid.
-    offset_ = 0;                   // No offset.
-  }
-
-  void Init() {
-    if (flags_ & kArcNoCache) {  // If caching is disabled
-      // arcs_ is a pointer to the arcs in the underlying machine.
-      arcs_ = local_data_.arcs;
-      // Sets the arcs value flags that hold for arcs_.
-      data_flags_ = kArcWeightValue;
-      if (!fst_.GetMutableImpl()->EpsilonOnCallInput()) {
-        data_flags_ |= kArcILabelValue;
-      }
-      // Sets the offset between the underlying arc positions and the positions
-      // in the arc iterator.
-      offset_ = num_arcs_ - local_data_.narcs;
-    } else {
-      ExpandAndCache();
-    }
-  }
-
-  bool Done() const { return pos_ >= num_arcs_; }
-
-  const Arc &Value() const {
-    // If data_flags_ is 0, non-caching was not requested.
-    if (!data_flags_) {
-      // TODO(allauzen): Revisit this.
-      if (flags_ & kArcNoCache) {
-        // Should never happen.
-        FSTERROR() << "ReplaceFst: Inconsistent arc iterator flags";
-      }
-      ExpandAndCache();
-    }
-    if (pos_ - offset_ >= 0) {  // The requested arc is not the final arc.
-      const auto &arc = arcs_[pos_ - offset_];
-      if ((data_flags_ & flags_) == (flags_ & kArcValueFlags)) {
-        // If the value flags match the recquired value flags then returns the
-        // arc.
-        return arc;
-      } else {
-        // Otherwise, compute the corresponding arc on-the-fly.
-        fst_.GetMutableImpl()->ComputeArc(tuple_, arc, &arc_,
-                                          flags_ & kArcValueFlags);
-        return arc_;
-      }
-    } else {  // The requested arc is the final arc.
-      if ((final_flags_ & flags_) != (flags_ & kArcValueFlags)) {
-        // If the arc value flags that hold for the final arc do not match the
-        // requested value flags, then
-        // final_arc_ needs to be updated.
-        fst_.GetMutableImpl()->ComputeFinalArc(tuple_, &final_arc_,
-                                               flags_ & kArcValueFlags);
-        final_flags_ = flags_ & kArcValueFlags;
-      }
-      return final_arc_;
-    }
-  }
-
-  void Next() { ++pos_; }
-
-  size_t Position() const { return pos_; }
-
-  void Reset() { pos_ = 0; }
-
-  void Seek(size_t pos) { pos_ = pos; }
-
-  uint32 Flags() const { return flags_; }
-
-  void SetFlags(uint32 flags, uint32 mask) {
-    // Updates the flags taking into account what flags are supported
-    // by the FST.
-    flags_ &= ~mask;
-    flags_ |= (flags & fst_.GetImpl()->ArcIteratorFlags());
-    // If non-caching is not requested (and caching has not already been
-    // performed), then flush data_flags_ to request caching during the next
-    // call to Value().
-    if (!(flags_ & kArcNoCache) && data_flags_ != kArcValueFlags) {
-      if (!fst_.GetImpl()->HasArcs(s_)) data_flags_ = 0;
-    }
-    // If data_flags_ has been flushed but non-caching is requested before
-    // calling Value(), then set up the iterator for non-caching.
-    if ((flags & kArcNoCache) && (!data_flags_)) Init();
-  }
-
- private:
-  const ReplaceFst<Arc, StateTable, CacheStore> &fst_;  // Reference to the FST.
-  StateId s_;                                           // State in the FST.
-  mutable StateTuple tuple_;  // Tuple corresponding to state_.
-
-  ssize_t pos_;             // Current position.
-  mutable ssize_t offset_;  // Offset between position in iterator and in arcs_.
-  ssize_t num_arcs_;        // Number of arcs at state_.
-  uint32 flags_;            // Behavorial flags for the arc iterator
-  mutable Arc arc_;         // Memory to temporarily store computed arcs.
-
-  mutable ArcIteratorData<Arc> cache_data_;  // Arc iterator data in cache.
-  mutable ArcIteratorData<Arc> local_data_;  // Arc iterator data in local FST.
-
-  mutable const Arc *arcs_;     // Array of arcs.
-  mutable uint32 data_flags_;   // Arc value flags valid for data in arcs_.
-  mutable Arc final_arc_;       // Final arc (when required).
-  mutable uint32 final_flags_;  // Arc value flags valid for final_arc_.
-
-  ArcIterator(const ArcIterator &) = delete;
-  ArcIterator &operator=(const ArcIterator &) = delete;
-};
-
-template <class Arc, class StateTable, class CacheStore>
-class ReplaceFstMatcher : public MatcherBase<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FST = ReplaceFst<Arc, StateTable, CacheStore>;
-  using LocalMatcher = MultiEpsMatcher<Matcher<Fst<Arc>>>;
-
-  using StateTuple = typename StateTable::StateTuple;
-
-  // This makes a copy of the FST.
-  ReplaceFstMatcher(const ReplaceFst<Arc, StateTable, CacheStore> &fst,
-                    MatchType match_type)
-      : owned_fst_(fst.Copy()),
-        fst_(*owned_fst_),
-        impl_(fst_.GetMutableImpl()),
-        s_(fst::kNoStateId),
-        match_type_(match_type),
-        current_loop_(false),
-        final_arc_(false),
-        loop_(kNoLabel, 0, Weight::One(), kNoStateId) {
-    if (match_type_ == fst::MATCH_OUTPUT) {
-      std::swap(loop_.ilabel, loop_.olabel);
-    }
-    InitMatchers();
-  }
-
-  // This doesn't copy the FST.
-  ReplaceFstMatcher(const ReplaceFst<Arc, StateTable, CacheStore> *fst,
-                    MatchType match_type)
-      : fst_(*fst),
-        impl_(fst_.GetMutableImpl()),
-        s_(fst::kNoStateId),
-        match_type_(match_type),
-        current_loop_(false),
-        final_arc_(false),
-        loop_(kNoLabel, 0, Weight::One(), kNoStateId) {
-    if (match_type_ == fst::MATCH_OUTPUT) {
-      std::swap(loop_.ilabel, loop_.olabel);
-    }
-    InitMatchers();
-  }
-
-  // This makes a copy of the FST.
-  ReplaceFstMatcher(
-      const ReplaceFstMatcher<Arc, StateTable, CacheStore> &matcher,
-      bool safe = false)
-      : owned_fst_(matcher.fst_.Copy(safe)),
-        fst_(*owned_fst_),
-        impl_(fst_.GetMutableImpl()),
-        s_(fst::kNoStateId),
-        match_type_(matcher.match_type_),
-        current_loop_(false),
-        final_arc_(false),
-        loop_(fst::kNoLabel, 0, Weight::One(), fst::kNoStateId) {
-    if (match_type_ == fst::MATCH_OUTPUT) {
-      std::swap(loop_.ilabel, loop_.olabel);
-    }
-    InitMatchers();
-  }
-
-  // Creates a local matcher for each component FST in the RTN. LocalMatcher is
-  // a multi-epsilon wrapper matcher. MultiEpsilonMatcher is used to match each
-  // non-terminal arc, since these non-terminal
-  // turn into epsilons on recursion.
-  void InitMatchers() {
-    const auto &fst_array = impl_->fst_array_;
-    matcher_.resize(fst_array.size());
-    for (Label i = 0; i < fst_array.size(); ++i) {
-      if (fst_array[i]) {
-        matcher_[i].reset(
-            new LocalMatcher(*fst_array[i], match_type_, kMultiEpsList));
-        auto it = impl_->nonterminal_set_.begin();
-        for (; it != impl_->nonterminal_set_.end(); ++it) {
-          matcher_[i]->AddMultiEpsLabel(*it);
-        }
-      }
-    }
-  }
-
-  ReplaceFstMatcher<Arc, StateTable, CacheStore> *Copy(
-      bool safe = false) const override {
-    return new ReplaceFstMatcher<Arc, StateTable, CacheStore>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override {
-    if (match_type_ == MATCH_NONE) return match_type_;
-    const auto true_prop =
-        match_type_ == MATCH_INPUT ? kILabelSorted : kOLabelSorted;
-    const auto false_prop =
-        match_type_ == MATCH_INPUT ? kNotILabelSorted : kNotOLabelSorted;
-    const auto props = fst_.Properties(true_prop | false_prop, test);
-    if (props & true_prop) {
-      return match_type_;
-    } else if (props & false_prop) {
-      return MATCH_NONE;
-    } else {
-      return MATCH_UNKNOWN;
-    }
-  }
-
-  const Fst<Arc> &GetFst() const override { return fst_; }
-
-  uint64 Properties(uint64 props) const override { return props; }
-
-  // Sets the state from which our matching happens.
-  void SetState(StateId s) final {
-    if (s_ == s) return;
-    s_ = s;
-    tuple_ = impl_->GetStateTable()->Tuple(s_);
-    if (tuple_.fst_state == kNoStateId) {
-      done_ = true;
-      return;
-    }
-    // Gets current matcher, used for non-epsilon matching.
-    current_matcher_ = matcher_[tuple_.fst_id].get();
-    current_matcher_->SetState(tuple_.fst_state);
-    loop_.nextstate = s_;
-    final_arc_ = false;
-  }
-
-  // Searches for label from previous set state. If label == 0, first
-  // hallucinate an epsilon loop; otherwise use the underlying matcher to
-  // search for the label or epsilons. Note since the ReplaceFst recursion
-  // on non-terminal arcs causes epsilon transitions to be created we use
-  // MultiEpsilonMatcher to search for possible matches of non-terminals. If the
-  // component FST
-  // reaches a final state we also need to add the exiting final arc.
-  bool Find(Label label) final {
-    bool found = false;
-    label_ = label;
-    if (label_ == 0 || label_ == kNoLabel) {
-      // Computes loop directly, avoiding Replace::ComputeArc.
-      if (label_ == 0) {
-        current_loop_ = true;
-        found = true;
-      }
-      // Searches for matching multi-epsilons.
-      final_arc_ = impl_->ComputeFinalArc(tuple_, nullptr);
-      found = current_matcher_->Find(kNoLabel) || final_arc_ || found;
-    } else {
-      // Searches on a sub machine directly using sub machine matcher.
-      found = current_matcher_->Find(label_);
-    }
-    return found;
-  }
-
-  bool Done() const final {
-    return !current_loop_ && !final_arc_ && current_matcher_->Done();
-  }
-
-  const Arc &Value() const final {
-    if (current_loop_) return loop_;
-    if (final_arc_) {
-      impl_->ComputeFinalArc(tuple_, &arc_);
-      return arc_;
-    }
-    const auto &component_arc = current_matcher_->Value();
-    impl_->ComputeArc(tuple_, component_arc, &arc_);
-    return arc_;
-  }
-
-  void Next() final {
-    if (current_loop_) {
-      current_loop_ = false;
-      return;
-    }
-    if (final_arc_) {
-      final_arc_ = false;
-      return;
-    }
-    current_matcher_->Next();
-  }
-
-  ssize_t Priority(StateId s) final { return fst_.NumArcs(s); }
-
- private:
-  std::unique_ptr<const ReplaceFst<Arc, StateTable, CacheStore>> owned_fst_;
-  const ReplaceFst<Arc, StateTable, CacheStore> &fst_;
-  internal::ReplaceFstImpl<Arc, StateTable, CacheStore> *impl_;
-  LocalMatcher *current_matcher_;
-  std::vector<std::unique_ptr<LocalMatcher>> matcher_;
-  StateId s_;             // Current state.
-  Label label_;           // Current label.
-  MatchType match_type_;  // Supplied by caller.
-  mutable bool done_;
-  mutable bool current_loop_;  // Current arc is the implicit loop.
-  mutable bool final_arc_;     // Current arc for exiting recursion.
-  mutable StateTuple tuple_;   // Tuple corresponding to state_.
-  mutable Arc arc_;
-  Arc loop_;
-
-  ReplaceFstMatcher &operator=(const ReplaceFstMatcher &) = delete;
-};
-
-template <class Arc, class StateTable, class CacheStore>
-inline void ReplaceFst<Arc, StateTable, CacheStore>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base =
-      new StateIterator<ReplaceFst<Arc, StateTable, CacheStore>>(*this);
-}
-
-using StdReplaceFst = ReplaceFst<StdArc>;
-
-// Recursively replaces arcs in the root FSTs with other FSTs.
-// This version writes the result of replacement to an output MutableFst.
-//
-// Replace supports replacement of arcs in one Fst with another FST. This
-// replacement is recursive. Replace takes an array of FST(s). One FST
-// represents the root (or topology) machine. The root FST refers to other FSTs
-// by recursively replacing arcs labeled as non-terminals with the matching
-// non-terminal FST. Currently Replace uses the output symbols of the arcs to
-// determine whether the arc is a non-terminal arc or not. A non-terminal can be
-// any label that is not a non-zero terminal label in the output alphabet.
-//
-// Note that input argument is a vector of pairs. These correspond to the tuple
-// of non-terminal Label and corresponding FST.
-template <class Arc>
-void Replace(const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
-                 &ifst_array,
-             MutableFst<Arc> *ofst,
-             ReplaceFstOptions<Arc> opts = ReplaceFstOptions<Arc>()) {
-  opts.gc = true;
-  opts.gc_limit = 0;  // Caches only the last state for fastest copy.
-  *ofst = ReplaceFst<Arc>(ifst_array, opts);
-}
-
-template <class Arc>
-void Replace(const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
-                 &ifst_array,
-             MutableFst<Arc> *ofst, const ReplaceUtilOptions &opts) {
-  Replace(ifst_array, ofst, ReplaceFstOptions<Arc>(opts));
-}
-
-// For backwards compatibility.
-template <class Arc>
-void Replace(const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
-                 &ifst_array,
-             MutableFst<Arc> *ofst, typename Arc::Label root,
-             bool epsilon_on_replace) {
-  Replace(ifst_array, ofst, ReplaceFstOptions<Arc>(root, epsilon_on_replace));
-}
-
-template <class Arc>
-void Replace(const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
-                 &ifst_array,
-             MutableFst<Arc> *ofst, typename Arc::Label root) {
-  Replace(ifst_array, ofst, ReplaceFstOptions<Arc>(root));
-}
-
-}  // namespace fst
-
-#endif  // FST_REPLACE_H_
diff --git a/projects/llm_framework/include/fst/reverse.h b/projects/llm_framework/include/fst/reverse.h
deleted file mode 100644
index 7c7c89db..00000000
--- a/projects/llm_framework/include/fst/reverse.h
+++ /dev/null
@@ -1,116 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to sort arcs in an FST.
-
-#ifndef FST_REVERSE_H_
-#define FST_REVERSE_H_
-
-#include <algorithm>
-#include <vector>
-
-#include <fst/cache.h>
-
-
-namespace fst {
-
-// Reverses an FST. The reversed result is written to an output mutable FST.
-// If A transduces string x to y with weight a, then the reverse of A
-// transduces the reverse of x to the reverse of y with weight a.Reverse().
-//
-// Typically, a = a.Reverse() and an arc is its own reverse (e.g., for
-// TropicalWeight or LogWeight). In general, e.g., when the weights only form a
-// left or right semiring, the output arc type must match the input arc type
-// except having the reversed Weight type.
-//
-// When require_superinitial is false, a superinitial state is not created in
-// the reversed FST iff the input FST has exactly one final state (which becomes
-// the initial state of the reversed FST) with a final weight of semiring One,
-// or if it does not belong to any cycle. When require_superinitial is true, a
-// superinitial state is always created.
-template <class FromArc, class ToArc>
-void Reverse(const Fst<FromArc> &ifst, MutableFst<ToArc> *ofst,
-             bool require_superinitial = true) {
-  using StateId = typename FromArc::StateId;
-  using FromWeight = typename FromArc::Weight;
-  using ToWeight = typename ToArc::Weight;
-  ofst->DeleteStates();
-  ofst->SetInputSymbols(ifst.InputSymbols());
-  ofst->SetOutputSymbols(ifst.OutputSymbols());
-  if (ifst.Properties(kExpanded, false)) {
-    ofst->ReserveStates(CountStates(ifst) + 1);
-  }
-  StateId istart = ifst.Start();
-  StateId ostart = kNoStateId;
-  StateId offset = 0;
-  uint64 dfs_iprops = 0;
-  uint64 dfs_oprops = 0;
-  if (!require_superinitial) {
-    for (StateIterator<Fst<FromArc>> siter(ifst); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      if (ifst.Final(s) == FromWeight::Zero()) continue;
-      if (ostart != kNoStateId) {
-        ostart = kNoStateId;
-        break;
-      } else {
-        ostart = s;
-      }
-    }
-    if (ostart != kNoStateId && ifst.Final(ostart) != FromWeight::One()) {
-      std::vector<StateId> scc;
-      SccVisitor<FromArc> scc_visitor(&scc, nullptr, nullptr, &dfs_iprops);
-      DfsVisit(ifst, &scc_visitor);
-      if (count(scc.begin(), scc.end(), scc[ostart]) > 1) {
-        ostart = kNoStateId;
-      } else {
-        for (ArcIterator<Fst<FromArc>> aiter(ifst, ostart); !aiter.Done();
-             aiter.Next()) {
-          if (aiter.Value().nextstate == ostart) {
-            ostart = kNoStateId;
-            break;
-          }
-        }
-      }
-      if (ostart != kNoStateId) dfs_oprops = kInitialAcyclic;
-    }
-  }
-  if (ostart == kNoStateId) {  // Super-initial requested or needed.
-    ostart = ofst->AddState();
-    offset = 1;
-  }
-  for (StateIterator<Fst<FromArc>> siter(ifst); !siter.Done(); siter.Next()) {
-    const auto is = siter.Value();
-    const auto os = is + offset;
-    while (ofst->NumStates() <= os) ofst->AddState();
-    if (is == istart) ofst->SetFinal(os, ToWeight::One());
-    const auto weight = ifst.Final(is);
-    if ((weight != FromWeight::Zero()) && (offset == 1)) {
-      const ToArc oarc(0, 0, weight.Reverse(), os);
-      ofst->AddArc(0, oarc);
-    }
-    for (ArcIterator<Fst<FromArc>> aiter(ifst, is); !aiter.Done();
-         aiter.Next()) {
-      const auto &iarc = aiter.Value();
-      const auto nos = iarc.nextstate + offset;
-      auto weight = iarc.weight.Reverse();
-      if (!offset && (nos == ostart)) {
-        weight = Times(ifst.Final(ostart).Reverse(), weight);
-      }
-      const ToArc oarc(iarc.ilabel, iarc.olabel, weight, os);
-      while (ofst->NumStates() <= nos) ofst->AddState();
-      ofst->AddArc(nos, oarc);
-    }
-  }
-  ofst->SetStart(ostart);
-  if (offset == 0 && ostart == istart) {
-    ofst->SetFinal(ostart, ifst.Final(ostart).Reverse());
-  }
-  const auto iprops = ifst.Properties(kCopyProperties, false) | dfs_iprops;
-  const auto oprops = ofst->Properties(kFstProperties, false) | dfs_oprops;
-  ofst->SetProperties(ReverseProperties(iprops, offset == 1) | oprops,
-                      kFstProperties);
-}
-
-}  // namespace fst
-
-#endif  // FST_REVERSE_H_
diff --git a/projects/llm_framework/include/fst/reweight.h b/projects/llm_framework/include/fst/reweight.h
deleted file mode 100644
index 64e68cb7..00000000
--- a/projects/llm_framework/include/fst/reweight.h
+++ /dev/null
@@ -1,127 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function to reweight an FST.
-
-#ifndef FST_REWEIGHT_H_
-#define FST_REWEIGHT_H_
-
-#include <vector>
-#include <fst/log.h>
-
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-enum ReweightType { REWEIGHT_TO_INITIAL, REWEIGHT_TO_FINAL };
-
-// Reweights an FST according to a vector of potentials in a given direction.
-// The weight must be left distributive when reweighting towards the initial
-// state and right distributive when reweighting towards the final states.
-//
-// An arc of weight w, with an origin state of potential p and destination state
-// of potential q, is reweighted by p^-1 \otimes (w \otimes q) when reweighting
-// torwards the initial state, and by (p \otimes w) \otimes q^-1 when
-// reweighting towards the final states.
-template <class Arc>
-void Reweight(MutableFst<Arc> *fst,
-              const std::vector<typename Arc::Weight> &potential,
-              ReweightType type) {
-  using Weight = typename Arc::Weight;
-  if (fst->NumStates() == 0) return;
-  // TODO(kbg): Make this a compile-time static_assert once we have a pleasant
-  // way to "deregister" this operation for non-distributive semirings so an
-  // informative error message is produced.
-  if (type == REWEIGHT_TO_FINAL && !(Weight::Properties() & kRightSemiring)) {
-    FSTERROR() << "Reweight: Reweighting to the final states requires "
-               << "Weight to be right distributive: " << Weight::Type();
-    fst->SetProperties(kError, kError);
-    return;
-  }
-  // TODO(kbg): Make this a compile-time static_assert once we have a pleasant
-  // way to "deregister" this operation for non-distributive semirings so an
-  // informative error message is produced.
-  if (type == REWEIGHT_TO_INITIAL && !(Weight::Properties() & kLeftSemiring)) {
-    FSTERROR() << "Reweight: Reweighting to the initial state requires "
-               << "Weight to be left distributive: " << Weight::Type();
-    fst->SetProperties(kError, kError);
-    return;
-  }
-  StateIterator<MutableFst<Arc>> siter(*fst);
-  for (; !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    if (s == potential.size()) break;
-    const auto &weight = potential[s];
-    if (weight != Weight::Zero()) {
-      for (MutableArcIterator<MutableFst<Arc>> aiter(fst, s); !aiter.Done();
-           aiter.Next()) {
-        auto arc = aiter.Value();
-        if (arc.nextstate >= potential.size()) continue;
-        const auto &nextweight = potential[arc.nextstate];
-        if (nextweight == Weight::Zero()) continue;
-        if (type == REWEIGHT_TO_INITIAL) {
-          arc.weight =
-              Divide(Times(arc.weight, nextweight), weight, DIVIDE_LEFT);
-        }
-        if (type == REWEIGHT_TO_FINAL) {
-          arc.weight =
-              Divide(Times(weight, arc.weight), nextweight, DIVIDE_RIGHT);
-        }
-        aiter.SetValue(arc);
-      }
-      if (type == REWEIGHT_TO_INITIAL) {
-        fst->SetFinal(s, Divide(fst->Final(s), weight, DIVIDE_LEFT));
-      }
-    }
-    if (type == REWEIGHT_TO_FINAL) {
-      fst->SetFinal(s, Times(weight, fst->Final(s)));
-    }
-  }
-  // This handles elements past the end of the potentials array.
-  for (; !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    if (type == REWEIGHT_TO_FINAL) {
-      fst->SetFinal(s, Times(Weight::Zero(), fst->Final(s)));
-    }
-  }
-  const auto startweight = fst->Start() < potential.size()
-                               ? potential[fst->Start()]
-                               : Weight::Zero();
-  if ((startweight != Weight::One()) && (startweight != Weight::Zero())) {
-    if (fst->Properties(kInitialAcyclic, true) & kInitialAcyclic) {
-      const auto s = fst->Start();
-      for (MutableArcIterator<MutableFst<Arc>> aiter(fst, s); !aiter.Done();
-           aiter.Next()) {
-        auto arc = aiter.Value();
-        if (type == REWEIGHT_TO_INITIAL) {
-          arc.weight = Times(startweight, arc.weight);
-        } else {
-          arc.weight = Times(Divide(Weight::One(), startweight, DIVIDE_RIGHT),
-                             arc.weight);
-        }
-        aiter.SetValue(arc);
-      }
-      if (type == REWEIGHT_TO_INITIAL) {
-        fst->SetFinal(s, Times(startweight, fst->Final(s)));
-      } else {
-        fst->SetFinal(s, Times(Divide(Weight::One(), startweight, DIVIDE_RIGHT),
-                               fst->Final(s)));
-      }
-    } else {
-      const auto s = fst->AddState();
-      const auto weight =
-          (type == REWEIGHT_TO_INITIAL)
-              ? startweight
-              : Divide(Weight::One(), startweight, DIVIDE_RIGHT);
-      fst->AddArc(s, Arc(0, 0, weight, fst->Start()));
-      fst->SetStart(s);
-    }
-  }
-  fst->SetProperties(ReweightProperties(fst->Properties(kFstProperties, false)),
-                     kFstProperties);
-}
-
-}  // namespace fst
-
-#endif  // FST_REWEIGHT_H_
diff --git a/projects/llm_framework/include/fst/rmepsilon.h b/projects/llm_framework/include/fst/rmepsilon.h
deleted file mode 100644
index 5135bf2d..00000000
--- a/projects/llm_framework/include/fst/rmepsilon.h
+++ /dev/null
@@ -1,548 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes that implemement epsilon-removal.
-
-#ifndef FST_RMEPSILON_H_
-#define FST_RMEPSILON_H_
-
-#include <forward_list>
-#include <stack>
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arcfilter.h>
-#include <fst/cache.h>
-#include <fst/connect.h>
-#include <fst/factor-weight.h>
-#include <fst/invert.h>
-#include <fst/prune.h>
-#include <fst/queue.h>
-#include <fst/shortest-distance.h>
-#include <fst/topsort.h>
-
-
-namespace fst {
-
-template <class Arc, class Queue>
-class RmEpsilonOptions
-    : public ShortestDistanceOptions<Arc, Queue, EpsilonArcFilter<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  bool connect;             // Connect output
-  Weight weight_threshold;  // Pruning weight threshold.
-  StateId state_threshold;  // Pruning state threshold.
-
-  explicit RmEpsilonOptions(Queue *queue, float delta = kShortestDelta,
-                            bool connect = true,
-                            Weight weight_threshold = Weight::Zero(),
-                            StateId state_threshold = kNoStateId)
-      : ShortestDistanceOptions<Arc, Queue, EpsilonArcFilter<Arc>>(
-            queue, EpsilonArcFilter<Arc>(), kNoStateId, delta),
-        connect(connect),
-        weight_threshold(std::move(weight_threshold)),
-        state_threshold(state_threshold) {}
-};
-
-namespace internal {
-
-// Computation state of the epsilon-removal algorithm.
-template <class Arc, class Queue>
-class RmEpsilonState {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  RmEpsilonState(const Fst<Arc> &fst, std::vector<Weight> *distance,
-                 const RmEpsilonOptions<Arc, Queue> &opts)
-      : fst_(fst),
-        distance_(distance),
-        sd_state_(fst_, distance, opts, true),
-        expand_id_(0) {}
-
-  void Expand(StateId s);
-
-  std::vector<Arc> &Arcs() { return arcs_; }
-
-  const Weight &Final() const { return final_; }
-
-  bool Error() const { return sd_state_.Error(); }
-
- private:
-  struct Element {
-    Label ilabel;
-    Label olabel;
-    StateId nextstate;
-
-    Element() {}
-
-    Element(Label ilabel, Label olabel, StateId nexstate)
-        : ilabel(ilabel), olabel(olabel), nextstate(nexstate) {}
-  };
-
-  struct ElementHash {
-   public:
-    size_t operator()(const Element &element) const {
-      static constexpr size_t prime0 = 7853;
-      static constexpr size_t prime1 = 7867;
-      return static_cast<size_t>(element.nextstate) +
-             static_cast<size_t>(element.ilabel) * prime0 +
-             static_cast<size_t>(element.olabel) * prime1;
-    }
-  };
-
-  class ElementEqual {
-   public:
-    bool operator()(const Element &e1, const Element &e2) const {
-      return (e1.ilabel == e2.ilabel) && (e1.olabel == e2.olabel) &&
-             (e1.nextstate == e2.nextstate);
-    }
-  };
-
-  using ElementMap = std::unordered_map<Element, std::pair<StateId, size_t>,
-                                        ElementHash, ElementEqual>;
-
-  const Fst<Arc> &fst_;
-  // Distance from state being expanded in epsilon-closure.
-  std::vector<Weight> *distance_;
-  // Shortest distance algorithm computation state.
-  internal::ShortestDistanceState<Arc, Queue, EpsilonArcFilter<Arc>> sd_state_;
-  // Maps an element to a pair corresponding to a position in the arcs vector
-  // of the state being expanded. The element corresopnds to the position in
-  // the arcs_ vector if p.first is equal to the state being expanded.
-  ElementMap element_map_;
-  EpsilonArcFilter<Arc> eps_filter_;
-  std::stack<StateId> eps_queue_;  // Queue used to visit the epsilon-closure.
-  std::vector<bool> visited_;      // True if the state has been visited.
-  std::forward_list<StateId> visited_states_;  // List of visited states.
-  std::vector<Arc> arcs_;                      // Arcs of state being expanded.
-  Weight final_;       // Final weight of state being expanded.
-  StateId expand_id_;  // Unique ID for each call to Expand
-
-  RmEpsilonState(const RmEpsilonState &) = delete;
-  RmEpsilonState &operator=(const RmEpsilonState &) = delete;
-};
-
-template <class Arc, class Queue>
-void RmEpsilonState<Arc, Queue>::Expand(typename Arc::StateId source) {
-  final_ = Weight::Zero();
-  arcs_.clear();
-  sd_state_.ShortestDistance(source);
-  if (sd_state_.Error()) return;
-  eps_queue_.push(source);
-  while (!eps_queue_.empty()) {
-    const auto state = eps_queue_.top();
-    eps_queue_.pop();
-    while (visited_.size() <= state) visited_.push_back(false);
-    if (visited_[state]) continue;
-    visited_[state] = true;
-    visited_states_.push_front(state);
-    for (ArcIterator<Fst<Arc>> aiter(fst_, state); !aiter.Done();
-         aiter.Next()) {
-      auto arc = aiter.Value();
-      arc.weight = Times((*distance_)[state], arc.weight);
-      if (eps_filter_(arc)) {
-        while (visited_.size() <= arc.nextstate) visited_.push_back(false);
-        if (!visited_[arc.nextstate]) eps_queue_.push(arc.nextstate);
-      } else {
-        const Element element(arc.ilabel, arc.olabel, arc.nextstate);
-        auto insert_result = element_map_.insert(
-            std::make_pair(element, std::make_pair(expand_id_, arcs_.size())));
-        if (insert_result.second) {
-          arcs_.push_back(std::move(arc));
-        } else {
-          if (insert_result.first->second.first == expand_id_) {
-            auto &weight = arcs_[insert_result.first->second.second].weight;
-            weight = Plus(weight, arc.weight);
-          } else {
-            insert_result.first->second.first = expand_id_;
-            insert_result.first->second.second = arcs_.size();
-            arcs_.push_back(std::move(arc));
-          }
-        }
-      }
-    }
-    final_ = Plus(final_, Times((*distance_)[state], fst_.Final(state)));
-  }
-  while (!visited_states_.empty()) {
-    visited_[visited_states_.front()] = false;
-    visited_states_.pop_front();
-  }
-  ++expand_id_;
-}
-
-}  // namespace internal
-
-// Removes epsilon-transitions (when both the input and output label are an
-// epsilon) from a transducer. The result will be an equivalent FST that has no
-// such epsilon transitions. This version modifies its input. It allows fine
-// control via the options argument; see below for a simpler interface.
-//
-// The distance vector will be used to hold the shortest distances during the
-// epsilon-closure computation. The state queue discipline and convergence delta
-// are taken in the options argument.
-template <class Arc, class Queue>
-void RmEpsilon(MutableFst<Arc> *fst,
-               std::vector<typename Arc::Weight> *distance,
-               const RmEpsilonOptions<Arc, Queue> &opts) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  if (fst->Start() == kNoStateId) return;
-  // noneps_in[s] will be set to true iff s admits a non-epsilon incoming
-  // transition or is the start state.
-  std::vector<bool> noneps_in(fst->NumStates(), false);
-  noneps_in[fst->Start()] = true;
-  for (size_t i = 0; i < fst->NumStates(); ++i) {
-    for (ArcIterator<Fst<Arc>> aiter(*fst, i); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      if (arc.ilabel != 0 || arc.olabel != 0) {
-        noneps_in[arc.nextstate] = true;
-      }
-    }
-  }
-  // States sorted in topological order when (acyclic) or generic topological
-  // order (cyclic).
-  std::vector<StateId> states;
-  states.reserve(fst->NumStates());
-  if (fst->Properties(kTopSorted, false) & kTopSorted) {
-    for (size_t i = 0; i < fst->NumStates(); i++) states.push_back(i);
-  } else if (fst->Properties(kAcyclic, false) & kAcyclic) {
-    std::vector<StateId> order;
-    bool acyclic;
-    TopOrderVisitor<Arc> top_order_visitor(&order, &acyclic);
-    DfsVisit(*fst, &top_order_visitor, EpsilonArcFilter<Arc>());
-    // Sanity check: should be acyclic if property bit is set.
-    if (!acyclic) {
-      FSTERROR() << "RmEpsilon: Inconsistent acyclic property bit";
-      fst->SetProperties(kError, kError);
-      return;
-    }
-    states.resize(order.size());
-    for (StateId i = 0; i < order.size(); i++) states[order[i]] = i;
-  } else {
-    uint64 props;
-    std::vector<StateId> scc;
-    SccVisitor<Arc> scc_visitor(&scc, nullptr, nullptr, &props);
-    DfsVisit(*fst, &scc_visitor, EpsilonArcFilter<Arc>());
-    std::vector<StateId> first(scc.size(), kNoStateId);
-    std::vector<StateId> next(scc.size(), kNoStateId);
-    for (StateId i = 0; i < scc.size(); i++) {
-      if (first[scc[i]] != kNoStateId) next[i] = first[scc[i]];
-      first[scc[i]] = i;
-    }
-    for (StateId i = 0; i < first.size(); i++) {
-      for (auto j = first[i]; j != kNoStateId; j = next[j]) {
-        states.push_back(j);
-      }
-    }
-  }
-  internal::RmEpsilonState<Arc, Queue> rmeps_state(*fst, distance, opts);
-  while (!states.empty()) {
-    const auto state = states.back();
-    states.pop_back();
-    if (!noneps_in[state] &&
-        (opts.connect || opts.weight_threshold != Weight::Zero() ||
-         opts.state_threshold != kNoStateId)) {
-      continue;
-    }
-    rmeps_state.Expand(state);
-    fst->SetFinal(state, rmeps_state.Final());
-    fst->DeleteArcs(state);
-    auto &arcs = rmeps_state.Arcs();
-    fst->ReserveArcs(state, arcs.size());
-    while (!arcs.empty()) {
-      fst->AddArc(state, arcs.back());
-      arcs.pop_back();
-    }
-  }
-  if (opts.connect || opts.weight_threshold != Weight::Zero() ||
-      opts.state_threshold != kNoStateId) {
-    for (size_t s = 0; s < fst->NumStates(); ++s) {
-      if (!noneps_in[s]) fst->DeleteArcs(s);
-    }
-  }
-  if (rmeps_state.Error()) fst->SetProperties(kError, kError);
-  fst->SetProperties(
-      RmEpsilonProperties(fst->Properties(kFstProperties, false)),
-      kFstProperties);
-  if (opts.weight_threshold != Weight::Zero() ||
-      opts.state_threshold != kNoStateId) {
-    Prune(fst, opts.weight_threshold, opts.state_threshold);
-  }
-  if (opts.connect && opts.weight_threshold == Weight::Zero() &&
-      opts.state_threshold == kNoStateId) {
-    Connect(fst);
-  }
-}
-
-// Removes epsilon-transitions (when both the input and output label
-// are an epsilon) from a transducer. The result will be an equivalent
-// FST that has no such epsilon transitions. This version modifies its
-// input. It has a simplified interface; see above for a version that
-// allows finer control.
-//
-// Complexity:
-//
-// - Time:
-//
-//   Unweighted: O(v^2 + ve).
-//   Acyclic: O(v^2 + V e).
-//   Tropical semiring: O(v^2 log V + ve).
-//   General: exponential.
-//
-// - Space: O(vE)
-//
-// where v is the number of states visited and e is the number of arcs visited.
-//
-// For more information, see:
-//
-// Mohri, M. 2002. Generic epsilon-removal and input epsilon-normalization
-// algorithms for weighted transducers. International Journal of Computer
-// Science 13(1): 129-143.
-template <class Arc>
-void RmEpsilon(MutableFst<Arc> *fst, bool connect = true,
-               typename Arc::Weight weight_threshold = Arc::Weight::Zero(),
-               typename Arc::StateId state_threshold = kNoStateId,
-               float delta = kShortestDelta) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  std::vector<Weight> distance;
-  AutoQueue<StateId> state_queue(*fst, &distance, EpsilonArcFilter<Arc>());
-  RmEpsilonOptions<Arc, AutoQueue<StateId>> opts(
-      &state_queue, delta, connect, weight_threshold, state_threshold);
-  RmEpsilon(fst, &distance, opts);
-}
-
-struct RmEpsilonFstOptions : CacheOptions {
-  float delta;
-
-  explicit RmEpsilonFstOptions(const CacheOptions &opts,
-                               float delta = kShortestDelta)
-      : CacheOptions(opts), delta(delta) {}
-
-  explicit RmEpsilonFstOptions(float delta = kShortestDelta) : delta(delta) {}
-};
-
-namespace internal {
-
-// Implementation of delayed RmEpsilonFst.
-template <class Arc>
-class RmEpsilonFstImpl : public CacheImpl<Arc> {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-
-  using FstImpl<Arc>::Properties;
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
-  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
-  using CacheBaseImpl<CacheState<Arc>>::HasStart;
-  using CacheBaseImpl<CacheState<Arc>>::PushArc;
-  using CacheBaseImpl<CacheState<Arc>>::SetArcs;
-  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
-  using CacheBaseImpl<CacheState<Arc>>::SetStart;
-
-  RmEpsilonFstImpl(const Fst<Arc> &fst, const RmEpsilonFstOptions &opts)
-      : CacheImpl<Arc>(opts),
-        fst_(fst.Copy()),
-        delta_(opts.delta),
-        rmeps_state_(
-            *fst_, &distance_,
-            RmEpsilonOptions<Arc, FifoQueue<StateId>>(&queue_, delta_, false)) {
-    SetType("rmepsilon");
-    SetProperties(
-        RmEpsilonProperties(fst.Properties(kFstProperties, false), true),
-        kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  RmEpsilonFstImpl(const RmEpsilonFstImpl &impl)
-      : CacheImpl<Arc>(impl),
-        fst_(impl.fst_->Copy(true)),
-        delta_(impl.delta_),
-        rmeps_state_(
-            *fst_, &distance_,
-            RmEpsilonOptions<Arc, FifoQueue<StateId>>(&queue_, delta_, false)) {
-    SetType("rmepsilon");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  StateId Start() {
-    if (!HasStart()) SetStart(fst_->Start());
-    return CacheImpl<Arc>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) Expand(s);
-    return CacheImpl<Arc>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumOutputEpsilons(s);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) &&
-        (fst_->Properties(kError, false) || rmeps_state_.Error())) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<Arc>::InitArcIterator(s, data);
-  }
-
-  void Expand(StateId s) {
-    rmeps_state_.Expand(s);
-    SetFinal(s, rmeps_state_.Final());
-    auto &arcs = rmeps_state_.Arcs();
-    while (!arcs.empty()) {
-      PushArc(s, std::move(arcs.back()));
-      arcs.pop_back();
-    }
-    SetArcs(s);
-  }
-
- private:
-  std::unique_ptr<const Fst<Arc>> fst_;
-  float delta_;
-  std::vector<Weight> distance_;
-  FifoQueue<StateId> queue_;
-  internal::RmEpsilonState<Arc, FifoQueue<StateId>> rmeps_state_;
-};
-
-}  // namespace internal
-
-// Removes epsilon-transitions (when both the input and output label are an
-// epsilon) from a transducer. The result will be an equivalent FST that has no
-// such epsilon transitions. This version is a
-// delayed FST.
-//
-// Complexity:
-//
-// - Time:
-//   Unweighted: O(v^2 + ve).
-//   General: exponential.
-//
-// - Space: O(vE)
-//
-// where v is the number of states visited and e is the number of arcs visited.
-// Constant time to visit an input state or arc is assumed and exclusive of
-// caching.
-//
-// For more information, see:
-//
-// Mohri, M. 2002. Generic epsilon-removal and input epsilon-normalization
-// algorithms for weighted transducers. International Journal of Computer
-// Science 13(1): 129-143.
-//
-// This class attaches interface to implementation and handles
-// reference counting, delegating most methods to ImplToFst.
-template <class A>
-class RmEpsilonFst : public ImplToFst<internal::RmEpsilonFstImpl<A>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::RmEpsilonFstImpl<Arc>;
-
-  friend class ArcIterator<RmEpsilonFst<Arc>>;
-  friend class StateIterator<RmEpsilonFst<Arc>>;
-
-  explicit RmEpsilonFst(const Fst<Arc> &fst)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, RmEpsilonFstOptions())) {}
-
-  RmEpsilonFst(const Fst<A> &fst, const RmEpsilonFstOptions &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  RmEpsilonFst(const RmEpsilonFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this RmEpsilonFst. See Fst<>::Copy() for further doc.
-  RmEpsilonFst<Arc> *Copy(bool safe = false) const override {
-    return new RmEpsilonFst<Arc>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  RmEpsilonFst &operator=(const RmEpsilonFst &) = delete;
-};
-
-// Specialization for RmEpsilonFst.
-template <class Arc>
-class StateIterator<RmEpsilonFst<Arc>>
-    : public CacheStateIterator<RmEpsilonFst<Arc>> {
- public:
-  explicit StateIterator(const RmEpsilonFst<Arc> &fst)
-      : CacheStateIterator<RmEpsilonFst<Arc>>(fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for RmEpsilonFst.
-template <class Arc>
-class ArcIterator<RmEpsilonFst<Arc>>
-    : public CacheArcIterator<RmEpsilonFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const RmEpsilonFst<Arc> &fst, StateId s)
-      : CacheArcIterator<RmEpsilonFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc>
-inline void RmEpsilonFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<RmEpsilonFst<Arc>>(*this);
-}
-
-// Useful alias when using StdArc.
-using StdRmEpsilonFst = RmEpsilonFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_RMEPSILON_H_
diff --git a/projects/llm_framework/include/fst/rmfinalepsilon.h b/projects/llm_framework/include/fst/rmfinalepsilon.h
deleted file mode 100644
index 87e3a714..00000000
--- a/projects/llm_framework/include/fst/rmfinalepsilon.h
+++ /dev/null
@@ -1,80 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function to remove of final states that have epsilon-only input arcs.
-
-#ifndef FST_RMFINALEPSILON_H_
-#define FST_RMFINALEPSILON_H_
-
-#include <unordered_set>
-#include <vector>
-
-#include <fst/connect.h>
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-// Removes final states that have epsilon-only input arcs.
-template <class Arc>
-void RmFinalEpsilon(MutableFst<Arc> *fst) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  // Determines the coaccesibility of states.
-  std::vector<bool> access;
-  std::vector<bool> coaccess;
-  uint64 props = 0;
-  SccVisitor<Arc> scc_visitor(nullptr, &access, &coaccess, &props);
-  DfsVisit(*fst, &scc_visitor);
-  // Finds potential list of removable final states. These are final states that
-  // have no outgoing transitions or final states that have a non-coaccessible
-  // future.
-  std::unordered_set<StateId> finals;
-  for (StateIterator<Fst<Arc>> siter(*fst); !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    if (fst->Final(s) != Weight::Zero()) {
-      bool future_coaccess = false;
-      for (ArcIterator<Fst<Arc>> aiter(*fst, s); !aiter.Done(); aiter.Next()) {
-        const auto &arc = aiter.Value();
-        if (coaccess[arc.nextstate]) {
-          future_coaccess = true;
-          break;
-        }
-      }
-      if (!future_coaccess) finals.insert(s);
-    }
-  }
-  // Moves the final weight.
-  std::vector<Arc> arcs;
-  for (StateIterator<Fst<Arc>> siter(*fst); !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    auto weight = fst->Final(s);
-    arcs.clear();
-    for (ArcIterator<Fst<Arc>> aiter(*fst, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      // Next state is in the list of finals.
-      if (finals.find(arc.nextstate) != finals.end()) {
-        // Sums up all epsilon arcs.
-        if (arc.ilabel == 0 && arc.olabel == 0) {
-          weight = Plus(Times(fst->Final(arc.nextstate), arc.weight), weight);
-        } else {
-          arcs.push_back(arc);
-        }
-      } else {
-        arcs.push_back(arc);
-      }
-    }
-    // If some arcs (epsilon arcs) were deleted, delete all arcs and add back
-    // only the non-epsilon arcs.
-    if (arcs.size() < fst->NumArcs(s)) {
-      fst->DeleteArcs(s);
-      fst->SetFinal(s, weight);
-      for (const auto &arc : arcs) fst->AddArc(s, arc);
-    }
-  }
-  Connect(fst);
-}
-
-}  // namespace fst
-
-#endif  // FST_RMFINALEPSILON_H_
diff --git a/projects/llm_framework/include/fst/script/arc-class.h b/projects/llm_framework/include/fst/script/arc-class.h
deleted file mode 100644
index 551266d7..00000000
--- a/projects/llm_framework/include/fst/script/arc-class.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_ARC_CLASS_H_
-#define FST_SCRIPT_ARC_CLASS_H_
-
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-// A struct representing an arc while ignoring arc type. It is passed as an
-// argument to AddArc.
-
-struct ArcClass {
-  template <class Arc>
-  explicit ArcClass(const Arc &arc)
-      : ilabel(arc.ilabel), olabel(arc.olabel), weight(arc.weight),
-        nextstate(arc.nextstate) {}
-
-  ArcClass(int64 ilabel, int64 olabel, const WeightClass &weight,
-           int64 nextstate)
-      : ilabel(ilabel), olabel(olabel), weight(weight), nextstate(nextstate) {}
-
-  template <class Arc>
-  Arc GetArc() const {
-    return Arc(ilabel, olabel, *(weight.GetWeight<typename Arc::Weight>()),
-               nextstate);
-  }
-
-  int64 ilabel;
-  int64 olabel;
-  WeightClass weight;
-  int64 nextstate;
-};
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_ARC_CLASS_H_
diff --git a/projects/llm_framework/include/fst/script/arciterator-class.h b/projects/llm_framework/include/fst/script/arciterator-class.h
deleted file mode 100644
index 8e4ca4f8..00000000
--- a/projects/llm_framework/include/fst/script/arciterator-class.h
+++ /dev/null
@@ -1,212 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_ARCITERATOR_CLASS_H_
-#define FST_SCRIPT_ARCITERATOR_CLASS_H_
-
-#include <memory>
-#include <utility>
-
-#include <fst/fstlib.h>
-#include <fst/script/fst-class.h>
-
-// Scripting API support for ArcIterator.
-//
-// A call to Value() causes the underlying arc to be used to construct the
-// associated ArcClass.
-
-namespace fst {
-namespace script {
-
-// Non-mutable arc iterators.
-
-// Virtual interface implemented by each concrete ArcIteratorImpl<F>.
-class ArcIteratorImplBase {
- public:
-  virtual bool Done() const = 0;
-  virtual uint32 Flags() const = 0;
-  virtual void Next() = 0;
-  virtual size_t Position() const = 0;
-  virtual void Reset() = 0;
-  virtual void Seek(size_t a) = 0;
-  virtual void SetFlags(uint32 flags, uint32 mask) = 0;
-  virtual ArcClass Value() const = 0;
-  virtual ~ArcIteratorImplBase() {}
-};
-
-// Templated implementation.
-template <class Arc>
-class ArcIteratorClassImpl : public ArcIteratorImplBase {
- public:
-  explicit ArcIteratorClassImpl(const Fst<Arc> &fst, int64 s)
-      : aiter_(fst, s) {}
-
-  bool Done() const final { return aiter_.Done(); }
-
-  uint32 Flags() const final { return aiter_.Flags(); }
-
-  void Next() final { aiter_.Next(); }
-
-  size_t Position() const final { return aiter_.Position(); }
-
-  void Reset() final { aiter_.Reset(); }
-
-  void Seek(size_t a) final { aiter_.Seek(a); }
-
-  void SetFlags(uint32 flags, uint32 mask) final {
-    aiter_.SetFlags(flags, mask);
-  }
-
-  // This is returned by value because it has not yet been constructed, and
-  // is likely to participate in return-value optimization.
-  ArcClass Value() const final { return ArcClass(aiter_.Value()); }
-
-  ~ArcIteratorClassImpl() final {}
-
- private:
-  ArcIterator<Fst<Arc>> aiter_;
-};
-
-class ArcIteratorClass;
-
-using InitArcIteratorClassArgs =
-    std::tuple<const FstClass &, int64, ArcIteratorClass *>;
-
-// Untemplated user-facing class holding a templated pimpl.
-class ArcIteratorClass {
- public:
-  ArcIteratorClass(const FstClass &fst, int64 s);
-
-  template <class Arc>
-  ArcIteratorClass(const Fst<Arc> &fst, int64 s)
-      : impl_(new ArcIteratorClassImpl<Arc>(fst, s)) {}
-
-  bool Done() const { return impl_->Done(); }
-
-  uint32 Flags() const { return impl_->Flags(); }
-
-  void Next() { impl_->Next(); }
-
-  size_t Position() const { return impl_->Position(); }
-
-  void Reset() { impl_->Reset(); }
-
-  void Seek(size_t a) { impl_->Seek(a); }
-
-  void SetFlags(uint32 flags, uint32 mask) { impl_->SetFlags(flags, mask); }
-
-  ArcClass Value() const { return impl_->Value(); }
-
-  template <class Arc>
-  friend void InitArcIteratorClass(InitArcIteratorClassArgs *args);
-
- private:
-  std::unique_ptr<ArcIteratorImplBase> impl_;
-};
-
-template <class Arc>
-void InitArcIteratorClass(InitArcIteratorClassArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  std::get<2>(*args)->impl_.reset(
-      new ArcIteratorClassImpl<Arc>(fst, std::get<1>(*args)));
-}
-
-// Mutable arc iterators.
-
-// Virtual interface implemented by each concrete MutableArcIteratorImpl<F>.
-class MutableArcIteratorImplBase : public ArcIteratorImplBase {
- public:
-  virtual void SetValue(const ArcClass &) = 0;
-
-  ~MutableArcIteratorImplBase() override {}
-};
-
-// Templated implementation.
-template <class Arc>
-class MutableArcIteratorClassImpl
-    : public MutableArcIteratorImplBase {
- public:
-  explicit MutableArcIteratorClassImpl(MutableFst<Arc> *fst, int64 s)
-      : aiter_(fst, s) {}
-
-  bool Done() const final { return aiter_.Done(); }
-
-  uint32 Flags() const final { return aiter_.Flags(); }
-
-  void Next() final { aiter_.Next(); }
-
-  size_t Position() const final { return aiter_.Position(); }
-
-  void Reset() final { aiter_.Reset(); }
-
-  void Seek(size_t a) final { aiter_.Seek(a); }
-
-  void SetFlags(uint32 flags, uint32 mask) final {
-    aiter_.SetFlags(flags, mask);
-  }
-
-  void SetValue(const Arc &arc) { aiter_.SetValue(arc); }
-
-  void SetValue(const ArcClass &ac) final { aiter_.SetValue(ac.GetArc<Arc>()); }
-
-  // This is returned by value because it has not yet been constructed, and
-  // is likely to participate in return-value optimization.
-  ArcClass Value() const final { return ArcClass(aiter_.Value()); }
-
-  ~MutableArcIteratorClassImpl() override {}
-
- private:
-  MutableArcIterator<MutableFst<Arc>> aiter_;
-};
-
-class MutableArcIteratorClass;
-
-using InitMutableArcIteratorClassArgs =
-    std::tuple<MutableFstClass *, int64, MutableArcIteratorClass *>;
-
-// Untemplated user-facing class holding a templated pimpl.
-class MutableArcIteratorClass {
- public:
-  MutableArcIteratorClass(MutableFstClass *fst, int64 s);
-
-  template <class Arc>
-  MutableArcIteratorClass(MutableFst<Arc> *fst, int64 s)
-      : impl_(new MutableArcIteratorClassImpl<Arc>(fst, s)) {}
-
-  bool Done() const { return impl_->Done(); }
-
-  uint32 Flags() const { return impl_->Flags(); }
-
-  void Next() { impl_->Next(); }
-
-  size_t Position() const { return impl_->Position(); }
-
-  void Reset() { impl_->Reset(); }
-
-  void Seek(size_t a) { impl_->Seek(a); }
-
-  void SetFlags(uint32 flags, uint32 mask) { impl_->SetFlags(flags, mask); }
-
-  void SetValue(const ArcClass &ac) { impl_->SetValue(ac); }
-
-  ArcClass Value() const { return impl_->Value(); }
-
-  template <class Arc>
-  friend void InitMutableArcIteratorClass(
-      InitMutableArcIteratorClassArgs *args);
-
- private:
-  std::unique_ptr<MutableArcIteratorImplBase> impl_;
-};
-
-template <class Arc>
-void InitMutableArcIteratorClass(InitMutableArcIteratorClassArgs *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  std::get<2>(*args)->impl_.reset(
-      new MutableArcIteratorClassImpl<Arc>(fst, std::get<1>(*args)));
-}
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_ARCITERATOR_CLASS_H_
diff --git a/projects/llm_framework/include/fst/script/arcsort.h b/projects/llm_framework/include/fst/script/arcsort.h
deleted file mode 100644
index 3e56fe5c..00000000
--- a/projects/llm_framework/include/fst/script/arcsort.h
+++ /dev/null
@@ -1,44 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_ARCSORT_H_
-#define FST_SCRIPT_ARCSORT_H_
-
-#include <utility>
-
-#include <fst/arcsort.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-enum ArcSortType {
-  ILABEL_SORT,
-  OLABEL_SORT
-};
-
-using ArcSortArgs = std::pair<MutableFstClass *, ArcSortType>;
-
-template <class Arc>
-void ArcSort(ArcSortArgs *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  switch (std::get<1>(*args)) {
-    case ILABEL_SORT: {
-      const ILabelCompare<Arc> icomp;
-      ArcSort(fst, icomp);
-      return;
-    }
-    case OLABEL_SORT: {
-      const OLabelCompare<Arc> ocomp;
-      ArcSort(fst, ocomp);
-      return;
-    }
-  }
-}
-
-void ArcSort(MutableFstClass *ofst, ArcSortType);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_ARCSORT_H_
diff --git a/projects/llm_framework/include/fst/script/arg-packs.h b/projects/llm_framework/include/fst/script/arg-packs.h
deleted file mode 100644
index 93cd4a05..00000000
--- a/projects/llm_framework/include/fst/script/arg-packs.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// std::pair and std::tuple are used for the arguments of FstClass operations.
-//
-// If a function with a return value is required, use the WithReturnValue
-// template as follows:
-//
-// WithReturnValue<bool, std::tuple<...>>
-
-#ifndef FST_SCRIPT_ARG_PACKS_H_
-#define FST_SCRIPT_ARG_PACKS_H_
-
-#include <type_traits>
-
-namespace fst {
-namespace script {
-
-// Tack this on to an existing type to add a return value. The syntax for
-// accessing the args is then slightly more stilted, as you must do an extra
-// member access (since the args are stored as a member of this class).
-
-template <class Retval, class ArgTuple>
-struct WithReturnValue {
-  // Avoid reference-to-reference if ArgTuple is a reference.
-  using Args = typename std::remove_reference<ArgTuple>::type;
-
-  Retval retval;
-  const Args &args;
-
-  explicit WithReturnValue(const Args &args) : args(args) {}
-};
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_ARG_PACKS_H_
diff --git a/projects/llm_framework/include/fst/script/closure.h b/projects/llm_framework/include/fst/script/closure.h
deleted file mode 100644
index 7c68604a..00000000
--- a/projects/llm_framework/include/fst/script/closure.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_CLOSURE_H_
-#define FST_SCRIPT_CLOSURE_H_
-
-#include <utility>
-
-#include <fst/closure.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using ClosureArgs = std::pair<MutableFstClass *, const ClosureType>;
-
-template <class Arc>
-void Closure(ClosureArgs *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  Closure(fst, std::get<1>(*args));
-}
-
-void Closure(MutableFstClass *ofst, ClosureType closure_type);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_CLOSURE_H_
diff --git a/projects/llm_framework/include/fst/script/compile-impl.h b/projects/llm_framework/include/fst/script/compile-impl.h
deleted file mode 100644
index 943a0b72..00000000
--- a/projects/llm_framework/include/fst/script/compile-impl.h
+++ /dev/null
@@ -1,217 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to to compile a binary FST from textual input.
-
-#ifndef FST_SCRIPT_COMPILE_IMPL_H_
-#define FST_SCRIPT_COMPILE_IMPL_H_
-
-#include <iostream>
-#include <memory>
-#include <sstream>
-#include <string>
-#include <vector>
-
-#include <fst/fst.h>
-#include <fst/util.h>
-#include <fst/vector-fst.h>
-#include <unordered_map>
-
-DECLARE_string(fst_field_separator);
-
-namespace fst {
-
-// Compile a binary Fst from textual input, helper class for fstcompile.cc
-// WARNING: Stand-alone use of this class not recommended, most code should
-// read/write using the binary format which is much more efficient.
-template <class Arc>
-class FstCompiler {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // WARNING: use of negative labels not recommended as it may cause conflicts.
-  // If add_symbols_ is true, then the symbols will be dynamically added to the
-  // symbol tables. This is only useful if you set the (i/o)keep flag to attach
-  // the final symbol table, or use the accessors. (The input symbol tables are
-  // const and therefore not changed.)
-  FstCompiler(std::istream &istrm, const string &source,  // NOLINT
-              const SymbolTable *isyms, const SymbolTable *osyms,
-              const SymbolTable *ssyms, bool accep, bool ikeep,
-              bool okeep, bool nkeep, bool allow_negative_labels = false) {
-    std::unique_ptr<SymbolTable> misyms(isyms ? isyms->Copy() : nullptr);
-    std::unique_ptr<SymbolTable> mosyms(osyms ? osyms->Copy() : nullptr);
-    std::unique_ptr<SymbolTable> mssyms(ssyms ? ssyms->Copy() : nullptr);
-    Init(istrm, source, misyms.get(), mosyms.get(), mssyms.get(), accep,
-         ikeep, okeep, nkeep, allow_negative_labels, false);
-  }
-
-  FstCompiler(std::istream &istrm, const string &source,  // NOLINT
-              SymbolTable *isyms, SymbolTable *osyms, SymbolTable *ssyms,
-              bool accep, bool ikeep, bool okeep, bool nkeep,
-              bool allow_negative_labels, bool add_symbols) {
-    Init(istrm, source, isyms, osyms, ssyms, accep, ikeep, okeep, nkeep,
-         allow_negative_labels, add_symbols);
-  }
-
-  void Init(std::istream &istrm, const string &source,  // NOLINT
-            SymbolTable *isyms, SymbolTable *osyms, SymbolTable *ssyms,
-            bool accep, bool ikeep, bool okeep, bool nkeep,
-            bool allow_negative_labels, bool add_symbols) {
-    nline_ = 0;
-    source_ = source;
-    isyms_ = isyms;
-    osyms_ = osyms;
-    ssyms_ = ssyms;
-    nstates_ = 0;
-    keep_state_numbering_ = nkeep;
-    allow_negative_labels_ = allow_negative_labels;
-    add_symbols_ = add_symbols;
-    bool start_state_populated = false;
-    char line[kLineLen];
-    const string separator = FLAGS_fst_field_separator + "\n";
-    while (istrm.getline(line, kLineLen)) {
-      ++nline_;
-      std::vector<char *> col;
-      SplitString(line, separator.c_str(), &col, true);
-      if (col.empty() || col[0][0] == '\0')
-        continue;
-      if (col.size() > 5 || (col.size() > 4 && accep) ||
-          (col.size() == 3 && !accep)) {
-        FSTERROR() << "FstCompiler: Bad number of columns, source = " << source_
-                   << ", line = " << nline_;
-        fst_.SetProperties(kError, kError);
-        return;
-      }
-      StateId s = StrToStateId(col[0]);
-      while (s >= fst_.NumStates()) fst_.AddState();
-      if (!start_state_populated) {
-        fst_.SetStart(s);
-        start_state_populated = true;
-      }
-
-      Arc arc;
-      StateId d = s;
-      switch (col.size()) {
-        case 1:
-          fst_.SetFinal(s, Weight::One());
-          break;
-        case 2:
-          fst_.SetFinal(s, StrToWeight(col[1], true));
-          break;
-        case 3:
-          arc.nextstate = d = StrToStateId(col[1]);
-          arc.ilabel = StrToILabel(col[2]);
-          arc.olabel = arc.ilabel;
-          arc.weight = Weight::One();
-          fst_.AddArc(s, arc);
-          break;
-        case 4:
-          arc.nextstate = d = StrToStateId(col[1]);
-          arc.ilabel = StrToILabel(col[2]);
-          if (accep) {
-            arc.olabel = arc.ilabel;
-            arc.weight = StrToWeight(col[3], true);
-          } else {
-            arc.olabel = StrToOLabel(col[3]);
-            arc.weight = Weight::One();
-          }
-          fst_.AddArc(s, arc);
-          break;
-        case 5:
-          arc.nextstate = d = StrToStateId(col[1]);
-          arc.ilabel = StrToILabel(col[2]);
-          arc.olabel = StrToOLabel(col[3]);
-          arc.weight = StrToWeight(col[4], true);
-          fst_.AddArc(s, arc);
-      }
-      while (d >= fst_.NumStates()) fst_.AddState();
-    }
-    if (ikeep) fst_.SetInputSymbols(isyms);
-    if (okeep) fst_.SetOutputSymbols(osyms);
-  }
-
-  const VectorFst<Arc> &Fst() const { return fst_; }
-
- private:
-  // Maximum line length in text file.
-  static constexpr int kLineLen = 8096;
-
-  StateId StrToId(const char *s, SymbolTable *syms, const char *name,
-                  bool allow_negative = false) const {
-    StateId n = 0;
-    if (syms) {
-      n = (add_symbols_) ? syms->AddSymbol(s) : syms->Find(s);
-      if (n == -1 || (!allow_negative && n < 0)) {
-        FSTERROR() << "FstCompiler: Symbol \"" << s
-                   << "\" is not mapped to any integer " << name
-                   << ", symbol table = " << syms->Name()
-                   << ", source = " << source_ << ", line = " << nline_;
-        fst_.SetProperties(kError, kError);
-      }
-    } else {
-      char *p;
-      n = strtoll(s, &p, 10);
-      if (p < s + strlen(s) || (!allow_negative && n < 0)) {
-        FSTERROR() << "FstCompiler: Bad " << name << " integer = \"" << s
-                   << "\", source = " << source_ << ", line = " << nline_;
-        fst_.SetProperties(kError, kError);
-      }
-    }
-    return n;
-  }
-
-  StateId StrToStateId(const char *s) {
-    StateId n = StrToId(s, ssyms_, "state ID");
-    if (keep_state_numbering_) return n;
-    // Remaps state IDs to make dense set.
-    const auto it = states_.find(n);
-    if (it == states_.end()) {
-      states_[n] = nstates_;
-      return nstates_++;
-    } else {
-      return it->second;
-    }
-  }
-
-  StateId StrToILabel(const char *s) const {
-    return StrToId(s, isyms_, "arc ilabel", allow_negative_labels_);
-  }
-
-  StateId StrToOLabel(const char *s) const {
-    return StrToId(s, osyms_, "arc olabel", allow_negative_labels_);
-  }
-
-  Weight StrToWeight(const char *s, bool allow_zero) const {
-    Weight w;
-    std::istringstream strm(s);
-    strm >> w;
-    if (!strm || (!allow_zero && w == Weight::Zero())) {
-      FSTERROR() << "FstCompiler: Bad weight = \"" << s
-                 << "\", source = " << source_ << ", line = " << nline_;
-      fst_.SetProperties(kError, kError);
-      w = Weight::NoWeight();
-    }
-    return w;
-  }
-
-  mutable VectorFst<Arc> fst_;
-  size_t nline_;
-  string source_;       // Text FST source name.
-  SymbolTable *isyms_;  // ilabel symbol table (not owned).
-  SymbolTable *osyms_;  // olabel symbol table (not owned).
-  SymbolTable *ssyms_;  // slabel symbol table (not owned).
-  std::unordered_map<StateId, StateId> states_;  // State ID map.
-  StateId nstates_;                              // Number of seen states.
-  bool keep_state_numbering_;
-  bool allow_negative_labels_;  // Not recommended; may cause conflicts.
-  bool add_symbols_;            // Add to symbol tables on-the fly.
-
-  FstCompiler(const FstCompiler &) = delete;
-  FstCompiler &operator=(const FstCompiler &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_SCRIPT_COMPILE_IMPL_H_
diff --git a/projects/llm_framework/include/fst/script/compile.h b/projects/llm_framework/include/fst/script/compile.h
deleted file mode 100644
index c82ed477..00000000
--- a/projects/llm_framework/include/fst/script/compile.h
+++ /dev/null
@@ -1,98 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_COMPILE_H_
-#define FST_SCRIPT_COMPILE_H_
-
-#include <istream>
-#include <memory>
-
-#include <fst/script/arg-packs.h>
-#include <fst/script/compile-impl.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-// This operation exists in two forms. 1 is a void operation which writes the
-// compiled machine to disk; 2 returns an FstClass. I/O should normally be done
-// using the binary format for efficiency, so users are STRONGLY ENCOURAGED to
-// use 1 or to construct FSTs using the C++ FST mutation operations.
-
-// Note: it is safe to pass these strings as references because
-// this struct is only used to pass them deeper in the call graph.
-// Be sure you understand why this is so before using this struct
-// for anything else!
-struct CompileFstInnerArgs {
-  std::istream &istrm;
-  const string &source;
-  const string &fst_type;
-  const fst::SymbolTable *isyms;
-  const fst::SymbolTable *osyms;
-  const fst::SymbolTable *ssyms;
-  const bool accep;
-  const bool ikeep;
-  const bool okeep;
-  const bool nkeep;
-  const bool allow_negative_labels;
-
-  CompileFstInnerArgs(std::istream &istrm, const string &source,
-                      const string &fst_type, const fst::SymbolTable *isyms,
-                      const fst::SymbolTable *osyms,
-                      const fst::SymbolTable *ssyms, bool accep, bool ikeep,
-                      bool okeep, bool nkeep,
-                      bool allow_negative_labels = false)
-      : istrm(istrm),
-        source(source),
-        fst_type(fst_type),
-        isyms(isyms),
-        osyms(osyms),
-        ssyms(ssyms),
-        accep(accep),
-        ikeep(ikeep),
-        okeep(okeep),
-        nkeep(nkeep),
-        allow_negative_labels(allow_negative_labels) {}
-};
-
-using CompileFstArgs = WithReturnValue<FstClass *, CompileFstInnerArgs>;
-
-template <class Arc>
-void CompileFstInternal(CompileFstArgs *args) {
-  using fst::Convert;
-  using fst::Fst;
-  using fst::FstCompiler;
-  FstCompiler<Arc> fstcompiler(
-      args->args.istrm, args->args.source, args->args.isyms, args->args.osyms,
-      args->args.ssyms, args->args.accep, args->args.ikeep, args->args.okeep,
-      args->args.nkeep, args->args.allow_negative_labels);
-  const Fst<Arc> *fst = &fstcompiler.Fst();
-  std::unique_ptr<const Fst<Arc>> owned_fst;
-  if (args->args.fst_type != "vector") {
-    owned_fst.reset(Convert<Arc>(*fst, args->args.fst_type));
-    if (!owned_fst) {
-      FSTERROR() << "Failed to convert FST to desired type: "
-                 << args->args.fst_type;
-    }
-    fst = owned_fst.get();
-  }
-  args->retval = fst ? new FstClass(*fst) : nullptr;
-}
-
-void CompileFst(std::istream &istrm, const string &source, const string &dest,
-                const string &fst_type, const string &arc_type,
-                const SymbolTable *isyms, const SymbolTable *osyms,
-                const SymbolTable *ssyms, bool accep, bool ikeep, bool okeep,
-                bool nkeep, bool allow_negative_labels);
-
-FstClass *CompileFstInternal(std::istream &istrm, const string &source,
-                             const string &fst_type, const string &arc_type,
-                             const SymbolTable *isyms, const SymbolTable *osyms,
-                             const SymbolTable *ssyms, bool accep, bool ikeep,
-                             bool okeep, bool nkeep,
-                             bool allow_negative_labels);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_COMPILE_H_
diff --git a/projects/llm_framework/include/fst/script/compose.h b/projects/llm_framework/include/fst/script/compose.h
deleted file mode 100644
index a1735803..00000000
--- a/projects/llm_framework/include/fst/script/compose.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_COMPOSE_H_
-#define FST_SCRIPT_COMPOSE_H_
-
-#include <tuple>
-
-#include <fst/compose.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using ComposeArgs = std::tuple<const FstClass &, const FstClass &,
-                               MutableFstClass *, const ComposeOptions &>;
-
-template <class Arc>
-void Compose(ComposeArgs *args) {
-  const Fst<Arc> &ifst1 = *(std::get<0>(*args).GetFst<Arc>());
-  const Fst<Arc> &ifst2 = *(std::get<1>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
-  const auto &opts = std::get<3>(*args);
-  Compose(ifst1, ifst2, ofst, opts);
-}
-
-void Compose(const FstClass &ifst1, const FstClass &ifst2,
-             MutableFstClass *ofst,
-             const ComposeOptions &opts = ComposeOptions());
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_COMPOSE_H_
diff --git a/projects/llm_framework/include/fst/script/concat.h b/projects/llm_framework/include/fst/script/concat.h
deleted file mode 100644
index 4bf8dc61..00000000
--- a/projects/llm_framework/include/fst/script/concat.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_CONCAT_H_
-#define FST_SCRIPT_CONCAT_H_
-
-#include <utility>
-
-#include <fst/concat.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using ConcatArgs1 = std::pair<MutableFstClass *, const FstClass &>;
-
-template <class Arc>
-void Concat(ConcatArgs1 *args) {
-  MutableFst<Arc> *ofst = std::get<0>(*args)->GetMutableFst<Arc>();
-  const Fst<Arc> &ifst = *(std::get<1>(*args).GetFst<Arc>());
-  Concat(ofst, ifst);
-}
-
-using ConcatArgs2 = std::pair<const FstClass &, MutableFstClass *>;
-
-template <class Arc>
-void Concat(ConcatArgs2 *args) {
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  Concat(ifst, ofst);
-}
-
-void Concat(MutableFstClass *ofst, const FstClass &ifst);
-
-void Concat(const FstClass &ifst, MutableFstClass *ofst);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_CONCAT_H_
diff --git a/projects/llm_framework/include/fst/script/connect.h b/projects/llm_framework/include/fst/script/connect.h
deleted file mode 100644
index 030102ac..00000000
--- a/projects/llm_framework/include/fst/script/connect.h
+++ /dev/null
@@ -1,23 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_CONNECT_H_
-#define FST_SCRIPT_CONNECT_H_
-
-#include <fst/connect.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-template <class Arc>
-void Connect(MutableFstClass *fst) {
-  Connect(fst->GetMutableFst<Arc>());
-}
-
-void Connect(MutableFstClass *fst);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_CONNECT_H_
diff --git a/projects/llm_framework/include/fst/script/convert.h b/projects/llm_framework/include/fst/script/convert.h
deleted file mode 100644
index 1a6eeaa3..00000000
--- a/projects/llm_framework/include/fst/script/convert.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_CONVERT_H_
-#define FST_SCRIPT_CONVERT_H_
-
-#include <memory>
-#include <string>
-#include <utility>
-
-#include <fst/register.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using ConvertInnerArgs = std::pair<const FstClass &, const string &>;
-
-using ConvertArgs = WithReturnValue<FstClass *, ConvertInnerArgs>;
-
-template <class Arc>
-void Convert(ConvertArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(args->args).GetFst<Arc>());
-  const string &new_type = std::get<1>(args->args);
-  std::unique_ptr<Fst<Arc>> result(Convert(fst, new_type));
-  args->retval = result ? new FstClass(*result) : nullptr;
-}
-
-FstClass *Convert(const FstClass &fst, const string &new_type);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_CONVERT_H_
diff --git a/projects/llm_framework/include/fst/script/decode.h b/projects/llm_framework/include/fst/script/decode.h
deleted file mode 100644
index 09f25391..00000000
--- a/projects/llm_framework/include/fst/script/decode.h
+++ /dev/null
@@ -1,49 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_DECODE_H_
-#define FST_SCRIPT_DECODE_H_
-
-#include <memory>
-#include <string>
-#include <utility>
-
-#include <fst/encode.h>
-#include <fst/script/encodemapper-class.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using DecodeArgs1 = std::pair<MutableFstClass *, const string &>;
-
-template <class Arc>
-void Decode(DecodeArgs1 *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  std::unique_ptr<EncodeMapper<Arc>> decoder(
-      EncodeMapper<Arc>::Read(std::get<1>(*args), DECODE));
-  if (!decoder) {
-    fst->SetProperties(kError, kError);
-    return;
-  }
-  Decode(fst, *decoder);
-}
-
-using DecodeArgs2 = std::pair<MutableFstClass *, const EncodeMapperClass &>;
-
-template <class Arc>
-void Decode(DecodeArgs2 *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  const EncodeMapper<Arc> &encoder =
-      *(std::get<1>(*args).GetEncodeMapper<Arc>());
-  Decode(fst, encoder);
-}
-
-void Decode(MutableFstClass *fst, const string &coder_fname);
-
-void Decode(MutableFstClass *fst, const EncodeMapperClass &encoder);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_DECODE_H_
diff --git a/projects/llm_framework/include/fst/script/determinize.h b/projects/llm_framework/include/fst/script/determinize.h
deleted file mode 100644
index 383a8fe4..00000000
--- a/projects/llm_framework/include/fst/script/determinize.h
+++ /dev/null
@@ -1,59 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_DETERMINIZE_H_
-#define FST_SCRIPT_DETERMINIZE_H_
-
-#include <tuple>
-
-#include <fst/determinize.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-struct DeterminizeOptions {
-  const float delta;
-  const WeightClass &weight_threshold;
-  const int64 state_threshold;
-  const int64 subsequential_label;
-  const DeterminizeType det_type;
-  const bool increment_subsequential_label;
-
-  DeterminizeOptions(float delta, const WeightClass &weight_threshold,
-                     int64 state_threshold = kNoStateId,
-                     int64 subsequential_label = 0,
-                     DeterminizeType det_type = DETERMINIZE_FUNCTIONAL,
-                     bool increment_subsequential_label = false)
-      : delta(delta),
-        weight_threshold(weight_threshold),
-        state_threshold(state_threshold),
-        subsequential_label(subsequential_label),
-        det_type(det_type),
-        increment_subsequential_label(increment_subsequential_label) {}
-};
-
-using DeterminizeArgs = std::tuple<const FstClass &, MutableFstClass *,
-                                   const DeterminizeOptions &>;
-
-template <class Arc>
-void Determinize(DeterminizeArgs *args) {
-  using Weight = typename Arc::Weight;
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  const auto &opts = std::get<2>(*args);
-  const auto weight_threshold = *(opts.weight_threshold.GetWeight<Weight>());
-  const fst::DeterminizeOptions<Arc> detargs(opts.delta, weight_threshold,
-      opts.state_threshold, opts.subsequential_label, opts.det_type,
-      opts.increment_subsequential_label);
-  Determinize(ifst, ofst, detargs);
-}
-
-void Determinize(const FstClass &ifst, MutableFstClass *ofst,
-                 const DeterminizeOptions &opts);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_DETERMINIZE_H_
diff --git a/projects/llm_framework/include/fst/script/difference.h b/projects/llm_framework/include/fst/script/difference.h
deleted file mode 100644
index 7af6200a..00000000
--- a/projects/llm_framework/include/fst/script/difference.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_DIFFERENCE_H_
-#define FST_SCRIPT_DIFFERENCE_H_
-
-#include <tuple>
-
-#include <fst/difference.h>
-#include <fst/script/compose.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using DifferenceArgs = std::tuple<const FstClass &, const FstClass &,
-                                  MutableFstClass *, const ComposeOptions &>;
-
-template <class Arc>
-void Difference(DifferenceArgs *args) {
-  const Fst<Arc> &ifst1 = *(std::get<0>(*args).GetFst<Arc>());
-  const Fst<Arc> &ifst2 = *(std::get<1>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
-  const auto &opts = std::get<3>(*args);
-  Difference(ifst1, ifst2, ofst, opts);
-}
-
-void Difference(const FstClass &ifst1, const FstClass &ifst2,
-                MutableFstClass *ofst,
-                const ComposeOptions &opts = ComposeOptions());
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_DIFFERENCE_H_
diff --git a/projects/llm_framework/include/fst/script/disambiguate.h b/projects/llm_framework/include/fst/script/disambiguate.h
deleted file mode 100644
index acc1fba2..00000000
--- a/projects/llm_framework/include/fst/script/disambiguate.h
+++ /dev/null
@@ -1,54 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_DISAMBIGUATE_H_
-#define FST_SCRIPT_DISAMBIGUATE_H_
-
-#include <tuple>
-#include <utility>
-
-#include <fst/disambiguate.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-struct DisambiguateOptions {
-  const float delta;
-  const WeightClass &weight_threshold;
-  const int64 state_threshold;
-  const int64 subsequential_label;
-
-  DisambiguateOptions(float delta, const WeightClass &weight_threshold,
-                      int64 state_threshold = kNoStateId,
-                      int64 subsequential_label = 0)
-      : delta(delta),
-        weight_threshold(weight_threshold),
-        state_threshold(state_threshold),
-        subsequential_label(subsequential_label) {}
-};
-
-using DisambiguateArgs = std::tuple<const FstClass &, MutableFstClass *,
-                                     const DisambiguateOptions &>;
-
-template <class Arc>
-void Disambiguate(DisambiguateArgs *args) {
-  using Weight = typename Arc::Weight;
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  const auto &opts = std::get<2>(*args);
-  const auto weight_threshold = *(opts.weight_threshold.GetWeight<Weight>());
-  const fst::DisambiguateOptions<Arc> disargs(opts.delta, weight_threshold,
-                                                  opts.state_threshold,
-                                                  opts.subsequential_label);
-  Disambiguate(ifst, ofst, disargs);
-}
-
-void Disambiguate(const FstClass &ifst, MutableFstClass *ofst,
-                  const DisambiguateOptions &opts);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_DISAMBIGUATE_H_
diff --git a/projects/llm_framework/include/fst/script/draw-impl.h b/projects/llm_framework/include/fst/script/draw-impl.h
deleted file mode 100644
index f204b2e6..00000000
--- a/projects/llm_framework/include/fst/script/draw-impl.h
+++ /dev/null
@@ -1,227 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to draw a binary FST by producing a text file in dot format, a helper
-// class to fstdraw.cc.
-
-#ifndef FST_SCRIPT_DRAW_IMPL_H_
-#define FST_SCRIPT_DRAW_IMPL_H_
-
-#include <ostream>
-#include <sstream>
-#include <string>
-
-#include <fst/fst.h>
-#include <fst/util.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-
-// Print a binary FST in GraphViz textual format (helper class for fstdraw.cc).
-// WARNING: Stand-alone use not recommend.
-template <class Arc>
-class FstDrawer {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  FstDrawer(const Fst<Arc> &fst, const SymbolTable *isyms,
-            const SymbolTable *osyms, const SymbolTable *ssyms, bool accep,
-            const string &title, float width, float height, bool portrait,
-            bool vertical, float ranksep, float nodesep, int fontsize,
-            int precision, const string &float_format, bool show_weight_one)
-      : fst_(fst),
-        isyms_(isyms),
-        osyms_(osyms),
-        ssyms_(ssyms),
-        accep_(accep && fst.Properties(kAcceptor, true)),
-        ostrm_(nullptr),
-        title_(title),
-        width_(width),
-        height_(height),
-        portrait_(portrait),
-        vertical_(vertical),
-        ranksep_(ranksep),
-        nodesep_(nodesep),
-        fontsize_(fontsize),
-        precision_(precision),
-        float_format_(float_format),
-        show_weight_one_(show_weight_one) {}
-
-  // Draws FST to an output buffer.
-  void Draw(std::ostream *strm, const string &dest) {
-    ostrm_ = strm;
-    SetStreamState(ostrm_);
-    dest_ = dest;
-    StateId start = fst_.Start();
-    if (start == kNoStateId) return;
-    PrintString("digraph FST {\n");
-    if (vertical_) {
-      PrintString("rankdir = BT;\n");
-    } else {
-      PrintString("rankdir = LR;\n");
-    }
-    PrintString("size = \"");
-    Print(width_);
-    PrintString(",");
-    Print(height_);
-    PrintString("\";\n");
-    if (!dest_.empty()) PrintString("label = \"" + title_ + "\";\n");
-    PrintString("center = 1;\n");
-    if (portrait_) {
-      PrintString("orientation = Portrait;\n");
-    } else {
-      PrintString("orientation = Landscape;\n");
-    }
-    PrintString("ranksep = \"");
-    Print(ranksep_);
-    PrintString("\";\n");
-    PrintString("nodesep = \"");
-    Print(nodesep_);
-    PrintString("\";\n");
-    // Initial state first.
-    DrawState(start);
-    for (StateIterator<Fst<Arc>> siter(fst_); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      if (s != start) DrawState(s);
-    }
-    PrintString("}\n");
-  }
-
- private:
-  void SetStreamState(std::ostream* strm) const {
-    strm->precision(precision_);
-    if (float_format_ == "e")
-        strm->setf(std::ios_base::scientific, std::ios_base::floatfield);
-    if (float_format_ == "f")
-        strm->setf(std::ios_base::fixed, std::ios_base::floatfield);
-    // O.w. defaults to "g" per standard lib.
-  }
-
-  void PrintString(const string &str) const { *ostrm_ << str; }
-
-  // Escapes backslash and double quote if these occur in the string. Dot will
-  // not deal gracefully with these if they are not escaped.
-  static string Escape(const string &str) {
-    string ns;
-    for (char c : str) {
-      if (c == '\\' || c == '"') ns.push_back('\\');
-      ns.push_back(c);
-    }
-    return ns;
-  }
-
-  void PrintId(StateId id, const SymbolTable *syms, const char *name) const {
-    if (syms) {
-      auto symbol = syms->Find(id);
-      if (symbol.empty()) {
-        FSTERROR() << "FstDrawer: Integer " << id
-                   << " is not mapped to any textual symbol"
-                   << ", symbol table = " << syms->Name()
-                   << ", destination = " << dest_;
-        symbol = "?";
-      }
-      PrintString(Escape(symbol));
-    } else {
-      PrintString(std::to_string(id));
-    }
-  }
-
-  void PrintStateId(StateId s) const { PrintId(s, ssyms_, "state ID"); }
-
-  void PrintILabel(Label label) const {
-    PrintId(label, isyms_, "arc input label");
-  }
-
-  void PrintOLabel(Label label) const {
-    PrintId(label, osyms_, "arc output label");
-  }
-
-  void PrintWeight(Weight w) const {
-    // Weight may have double quote characters in it, so escape it.
-    PrintString(Escape(ToString(w)));
-  }
-
-  template <class T>
-  void Print(T t) const { *ostrm_ << t; }
-
-  template <class T>
-  string ToString(T t) const {
-    std::stringstream ss;
-    SetStreamState(&ss);
-    ss << t;
-    return ss.str();
-  }
-
-  void DrawState(StateId s) const {
-    Print(s);
-    PrintString(" [label = \"");
-    PrintStateId(s);
-    const auto weight = fst_.Final(s);
-    if (weight != Weight::Zero()) {
-      if (show_weight_one_ || (weight != Weight::One())) {
-        PrintString("/");
-        PrintWeight(weight);
-      }
-      PrintString("\", shape = doublecircle,");
-    } else {
-      PrintString("\", shape = circle,");
-    }
-    if (s == fst_.Start()) {
-      PrintString(" style = bold,");
-    } else {
-      PrintString(" style = solid,");
-    }
-    PrintString(" fontsize = ");
-    Print(fontsize_);
-    PrintString("]\n");
-    for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      PrintString("\t");
-      Print(s);
-      PrintString(" -> ");
-      Print(arc.nextstate);
-      PrintString(" [label = \"");
-      PrintILabel(arc.ilabel);
-      if (!accep_) {
-        PrintString(":");
-        PrintOLabel(arc.olabel);
-      }
-      if (show_weight_one_ || (arc.weight != Weight::One())) {
-        PrintString("/");
-        PrintWeight(arc.weight);
-      }
-      PrintString("\", fontsize = ");
-      Print(fontsize_);
-      PrintString("];\n");
-    }
-  }
-
-  const Fst<Arc> &fst_;
-  const SymbolTable *isyms_;  // ilabel symbol table.
-  const SymbolTable *osyms_;  // olabel symbol table.
-  const SymbolTable *ssyms_;  // slabel symbol table.
-  bool accep_;                // Print as acceptor when possible.
-  std::ostream *ostrm_;       // Drawn FST destination.
-  string dest_;               // Drawn FST destination name.
-
-  string title_;
-  float width_;
-  float height_;
-  bool portrait_;
-  bool vertical_;
-  float ranksep_;
-  float nodesep_;
-  int fontsize_;
-  int precision_;
-  string float_format_;
-  bool show_weight_one_;
-
-  FstDrawer(const FstDrawer &) = delete;
-  FstDrawer &operator=(const FstDrawer &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_SCRIPT_DRAW_IMPL_H_
diff --git a/projects/llm_framework/include/fst/script/draw.h b/projects/llm_framework/include/fst/script/draw.h
deleted file mode 100644
index cb37df1e..00000000
--- a/projects/llm_framework/include/fst/script/draw.h
+++ /dev/null
@@ -1,85 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_DRAW_H_
-#define FST_SCRIPT_DRAW_H_
-
-#include <ostream>
-
-#include <fst/script/draw-impl.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-// Note: it is safe to pass these strings as references because
-// this struct is only used to pass them deeper in the call graph.
-// Be sure you understand why this is so before using this struct
-// for anything else!
-struct FstDrawerArgs {
-  const FstClass &fst;
-  const SymbolTable *isyms;
-  const SymbolTable *osyms;
-  const SymbolTable *ssyms;
-  const bool accep;
-  const string &title;
-  const float width;
-  const float height;
-  const bool portrait;
-  const bool vertical;
-  const float ranksep;
-  const float nodesep;
-  const int fontsize;
-  const int precision;
-  const string &float_format;  // NOLINT
-  const bool show_weight_one;
-  std::ostream *ostrm;
-  const string &dest;
-
-  FstDrawerArgs(const FstClass &fst, const SymbolTable *isyms,
-                const SymbolTable *osyms, const SymbolTable *ssyms, bool accep,
-                const string &title, float width, float height, bool portrait,
-                bool vertical, float ranksep, float nodesep, int fontsize,
-                int precision, const string &float_format,
-                bool show_weight_one, std::ostream *ostrm,  const string &dest)
-      : fst(fst),
-        isyms(isyms),
-        osyms(osyms),
-        ssyms(ssyms),
-        accep(accep),
-        title(title),
-        width(width),
-        height(height),
-        portrait(portrait),
-        vertical(vertical),
-        ranksep(ranksep),
-        nodesep(nodesep),
-        fontsize(fontsize),
-        precision(precision),
-        float_format(float_format),
-        show_weight_one(show_weight_one),
-        ostrm(ostrm),
-        dest(dest) {}
-};
-
-template <class Arc>
-void DrawFst(FstDrawerArgs *args) {
-  const Fst<Arc> &fst = *(args->fst.GetFst<Arc>());
-  FstDrawer<Arc> fstdrawer(fst, args->isyms, args->osyms, args->ssyms,
-      args->accep, args->title, args->width, args->height, args->portrait,
-      args->vertical, args->ranksep, args->nodesep, args->fontsize,
-      args->precision, args->float_format, args->show_weight_one);
-  fstdrawer.Draw(args->ostrm, args->dest);
-}
-
-void DrawFst(const FstClass &fst, const SymbolTable *isyms,
-             const SymbolTable *osyms, const SymbolTable *ssyms, bool accep,
-             const string &title, float width, float height, bool portrait,
-             bool vertical, float ranksep, float nodesep, int fontsize,
-             int precision, const string &float_format, bool show_weight_one,
-             std::ostream *ostrm, const string &dest);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_DRAW_H_
diff --git a/projects/llm_framework/include/fst/script/encode.h b/projects/llm_framework/include/fst/script/encode.h
deleted file mode 100644
index 6a869680..00000000
--- a/projects/llm_framework/include/fst/script/encode.h
+++ /dev/null
@@ -1,51 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_ENCODE_H_
-#define FST_SCRIPT_ENCODE_H_
-
-#include <memory>
-#include <string>
-#include <tuple>
-#include <utility>
-
-#include <fst/encode.h>
-#include <fst/script/encodemapper-class.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using EncodeArgs1 = std::tuple<MutableFstClass *, uint32, bool, const string &>;
-
-template <class Arc>
-void Encode(EncodeArgs1 *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  const string &coder_fname = std::get<3>(*args);
-  // If true, reuse encode from disk. If false, make a new encoder and just use
-  // the filename argument as the destination state.
-  std::unique_ptr<EncodeMapper<Arc>> encoder(
-      std::get<2>(*args) ? EncodeMapper<Arc>::Read(coder_fname, ENCODE)
-                         : new EncodeMapper<Arc>(std::get<1>(*args), ENCODE));
-  Encode(fst, encoder.get());
-  if (!std::get<2>(*args)) encoder->Write(coder_fname);
-}
-
-using EncodeArgs2 = std::pair<MutableFstClass *, EncodeMapperClass *>;
-
-template <class Arc>
-void Encode(EncodeArgs2 *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  EncodeMapper<Arc> *encoder = std::get<1>(*args)->GetEncodeMapper<Arc>();
-  Encode(fst, encoder);
-}
-
-void Encode(MutableFstClass *fst, uint32 flags, bool reuse_encoder,
-            const string &coder_fname);
-
-void Encode(MutableFstClass *fst, EncodeMapperClass *encoder);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_ENCODE_H_
diff --git a/projects/llm_framework/include/fst/script/encodemapper-class.h b/projects/llm_framework/include/fst/script/encodemapper-class.h
deleted file mode 100644
index b62824f1..00000000
--- a/projects/llm_framework/include/fst/script/encodemapper-class.h
+++ /dev/null
@@ -1,169 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_ENCODEMAPPER_CLASS_H_
-#define FST_SCRIPT_ENCODEMAPPER_CLASS_H_
-
-#include <memory>
-#include <string>
-#include <iostream>
-
-#include <fst/fstlib.h>
-#include <fst/script/arc-class.h>
-#include <fst/script/fst-class.h>
-
-// Scripting API support for EncodeMapper.
-
-namespace fst {
-namespace script {
-
-// Virtual interface implemented by each concrete EncodeMapperClassImpl<A>.
-class EncodeMapperImplBase {
- public:
-  // Returns an encoded ArcClass.
-  virtual ArcClass operator()(const ArcClass &a) = 0;
-  virtual const string &ArcType() const = 0;
-  virtual uint32 Flags() const = 0;
-  virtual uint64 Properties(uint64 inprops) = 0;
-  virtual EncodeType Type() const = 0;
-  virtual const SymbolTable *InputSymbols() const = 0;
-  virtual const SymbolTable *OutputSymbols() const = 0;
-  virtual void SetInputSymbols(const SymbolTable *syms) = 0;
-  virtual void SetOutputSymbols(const SymbolTable *syms) = 0;
-  virtual const string &WeightType() const = 0;
-  virtual ~EncodeMapperImplBase() {}
-};
-
-// Templated implementation.
-template <class Arc>
-class EncodeMapperClassImpl : public EncodeMapperImplBase {
- public:
-  EncodeMapperClassImpl(uint32 flags, EncodeType type)
-      : encoder_(flags, type) {}
-
-  ArcClass operator()(const ArcClass &a) final;
-
-  const string &ArcType() const final { return Arc::Type(); }
-
-  uint32 Flags() const final { return encoder_.Flags(); }
-
-  uint64 Properties(uint64 inprops) final {
-    return encoder_.Properties(inprops);
-  }
-
-  EncodeType Type() const final { return encoder_.Type(); }
-
-  const SymbolTable *InputSymbols() const final {
-    return encoder_.InputSymbols();
-  }
-
-  const SymbolTable *OutputSymbols() const final {
-    return encoder_.OutputSymbols();
-  }
-
-  void SetInputSymbols(const SymbolTable *syms) final {
-    encoder_.SetInputSymbols(syms);
-  }
-
-  void SetOutputSymbols(const SymbolTable *syms) final {
-    encoder_.SetOutputSymbols(syms);
-  }
-
-  const string &WeightType() const final { return Arc::Weight::Type(); }
-
-  ~EncodeMapperClassImpl() override {}
-
-  EncodeMapper<Arc> *GetImpl() const { return &encoder_; }
-
-  EncodeMapper<Arc> *GetImpl() { return &encoder_; }
-
- private:
-  EncodeMapper<Arc> encoder_;
-};
-
-// This is returned by value because it is very likely to undergo return-value
-// optimization.
-template <class Arc>
-inline ArcClass EncodeMapperClassImpl<Arc>::operator()(const ArcClass &a) {
-  Arc arc(a.ilabel, a.olabel, *(a.weight.GetWeight<typename Arc::Weight>()),
-          a.nextstate);
-  return ArcClass(encoder_(arc));
-}
-
-class EncodeMapperClass;
-
-using InitEncodeMapperClassArgs =
-    std::tuple<uint32, EncodeType, EncodeMapperClass *>;
-
-class EncodeMapperClass {
- public:
-  EncodeMapperClass(const string &arc_type, uint32 flags, EncodeType type);
-
-  template <class Arc>
-  EncodeMapperClass(uint32 flags, EncodeType type)
-      : impl_(new EncodeMapperClassImpl<Arc>(flags, type)) {}
-
-  ArcClass operator()(const ArcClass &arc) { return (*impl_)(arc); }
-
-  const string &ArcType() const { return impl_->ArcType(); }
-
-  uint32 Flags() const { return impl_->Flags(); }
-
-  uint64 Properties(uint64 inprops) { return impl_->Properties(inprops); }
-
-  EncodeType Type() const { return impl_->Type(); }
-
-  const SymbolTable *InputSymbols() const { return impl_->InputSymbols(); }
-
-  const SymbolTable *OutputSymbols() const { return impl_->OutputSymbols(); }
-
-  void SetInputSymbols(const SymbolTable *syms) {
-    impl_->SetInputSymbols(syms);
-  }
-
-  void SetOutputSymbols(const SymbolTable *syms) {
-    impl_->SetOutputSymbols(syms);
-  }
-
-  const string &WeightType() const { return impl_->WeightType(); }
-
-  template <class Arc>
-  friend void InitEncodeMapperClass(InitEncodeMapperClassArgs *args);
-
-  // Naturally, this exists in non-const and const forms. Encoding arcs or FSTs
-  // mutates the underlying encoder; decoding them does not.
-
-  template <class Arc>
-  EncodeMapper<Arc> *GetEncodeMapper() {
-    if (Arc::Type() != ArcType()) {
-      return nullptr;
-    } else {
-      auto *typed_impl = static_cast<EncodeMapperClassImpl<Arc> *>(impl_.get());
-      return typed_impl->GetImpl();
-    }
-  }
-
-  template <class Arc>
-  const EncodeMapper<Arc> *GetEncodeMapper() const {
-    if (Arc::Type() != ArcType()) {
-      return nullptr;
-    } else {
-      auto *typed_impl = static_cast<EncodeMapperClassImpl<Arc> *>(impl_.get());
-      return typed_impl->GetImpl();
-    }
-  }
-
- private:
-  std::unique_ptr<EncodeMapperImplBase> impl_;
-};
-
-template <class Arc>
-void InitEncodeMapperClass(InitEncodeMapperClassArgs *args) {
-  std::get<2>(*args)->impl_.reset(
-      new EncodeMapperClassImpl<Arc>(std::get<0>(*args), std::get<1>(*args)));
-}
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_ENCODEMAPPER_CLASS_H_
diff --git a/projects/llm_framework/include/fst/script/epsnormalize.h b/projects/llm_framework/include/fst/script/epsnormalize.h
deleted file mode 100644
index b55fefae..00000000
--- a/projects/llm_framework/include/fst/script/epsnormalize.h
+++ /dev/null
@@ -1,31 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_EPSNORMALIZE_H_
-#define FST_SCRIPT_EPSNORMALIZE_H_
-
-#include <tuple>
-
-#include <fst/epsnormalize.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using EpsNormalizeArgs = std::tuple<const FstClass &, MutableFstClass *,
-                                    EpsNormalizeType>;
-
-template <class Arc>
-void EpsNormalize(EpsNormalizeArgs *args) {
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  EpsNormalize(ifst, ofst, std::get<2>(*args));
-}
-
-void EpsNormalize(const FstClass &ifst, MutableFstClass *ofst,
-                  EpsNormalizeType norm_type = EPS_NORM_INPUT);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_EPSNORMALIZE_H_
diff --git a/projects/llm_framework/include/fst/script/equal.h b/projects/llm_framework/include/fst/script/equal.h
deleted file mode 100644
index 79ea9aa4..00000000
--- a/projects/llm_framework/include/fst/script/equal.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_EQUAL_H_
-#define FST_SCRIPT_EQUAL_H_
-
-#include <tuple>
-
-#include <fst/equal.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using EqualInnerArgs = std::tuple<const FstClass &, const FstClass &, float>;
-
-using EqualArgs = WithReturnValue<bool, EqualInnerArgs>;
-
-template <class Arc>
-void Equal(EqualArgs *args) {
-  const Fst<Arc> &fst1 = *(std::get<0>(args->args).GetFst<Arc>());
-  const Fst<Arc> &fst2 = *(std::get<1>(args->args).GetFst<Arc>());
-  args->retval = Equal(fst1, fst2, std::get<2>(args->args));
-}
-
-bool Equal(const FstClass &fst1, const FstClass &fst2, float delta = kDelta);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_EQUAL_H_
diff --git a/projects/llm_framework/include/fst/script/equivalent.h b/projects/llm_framework/include/fst/script/equivalent.h
deleted file mode 100644
index 7cdff45e..00000000
--- a/projects/llm_framework/include/fst/script/equivalent.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_EQUIVALENT_H_
-#define FST_SCRIPT_EQUIVALENT_H_
-
-#include <tuple>
-
-#include <fst/equivalent.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using EquivalentInnerArgs = std::tuple<const FstClass &, const FstClass &,
-                                       float>;
-
-using EquivalentArgs = WithReturnValue<bool, EquivalentInnerArgs>;
-
-template <class Arc>
-void Equivalent(EquivalentArgs *args) {
-  const Fst<Arc> &fst1 = *(std::get<0>(args->args).GetFst<Arc>());
-  const Fst<Arc> &fst2 = *(std::get<1>(args->args).GetFst<Arc>());
-  args->retval = Equivalent(fst1, fst2, std::get<2>(args->args));
-}
-
-bool Equivalent(const FstClass &fst1, const FstClass &fst2,
-                float delta = kDelta);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_EQUIVALENT_H_
diff --git a/projects/llm_framework/include/fst/script/fst-class.h b/projects/llm_framework/include/fst/script/fst-class.h
deleted file mode 100644
index 07319fc7..00000000
--- a/projects/llm_framework/include/fst/script/fst-class.h
+++ /dev/null
@@ -1,530 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_FST_CLASS_H_
-#define FST_SCRIPT_FST_CLASS_H_
-
-#include <algorithm>
-#include <limits>
-#include <string>
-#include <type_traits>
-
-#include <fst/expanded-fst.h>
-#include <fst/fst.h>
-#include <fst/mutable-fst.h>
-#include <fst/vector-fst.h>
-#include <fst/script/arc-class.h>
-#include <fst/script/weight-class.h>
-
-// Classes to support "boxing" all existing types of FST arcs in a single
-// FstClass which hides the arc types. This allows clients to load
-// and work with FSTs without knowing the arc type. These classes are only
-// recommended for use in high-level scripting applications. Most users should
-// use the lower-level templated versions corresponding to these classes.
-
-namespace fst {
-namespace script {
-
-// Abstract base class defining the set of functionalities implemented in all
-// impls and passed through by all bases. Below FstClassBase the class
-// hierarchy bifurcates; FstClassImplBase serves as the base class for all
-// implementations (of which FstClassImpl is currently the only one) and
-// FstClass serves as the base class for all interfaces.
-
-class FstClassBase {
- public:
-  virtual const string &ArcType() const = 0;
-  virtual WeightClass Final(int64) const = 0;
-  virtual const string &FstType() const = 0;
-  virtual const SymbolTable *InputSymbols() const = 0;
-  virtual size_t NumArcs(int64) const = 0;
-  virtual size_t NumInputEpsilons(int64) const = 0;
-  virtual size_t NumOutputEpsilons(int64) const = 0;
-  virtual const SymbolTable *OutputSymbols() const = 0;
-  virtual uint64 Properties(uint64, bool) const = 0;
-  virtual int64 Start() const = 0;
-  virtual const string &WeightType() const = 0;
-  virtual bool ValidStateId(int64) const = 0;
-  virtual bool Write(const string &) const = 0;
-  virtual bool Write(std::ostream &, const string &) const = 0;
-  virtual ~FstClassBase() {}
-};
-
-// Adds all the MutableFst methods.
-class FstClassImplBase : public FstClassBase {
- public:
-  virtual bool AddArc(int64, const ArcClass &) = 0;
-  virtual int64 AddState() = 0;
-  virtual FstClassImplBase *Copy() = 0;
-  virtual bool DeleteArcs(int64, size_t) = 0;
-  virtual bool DeleteArcs(int64) = 0;
-  virtual bool DeleteStates(const std::vector<int64> &) = 0;
-  virtual void DeleteStates() = 0;
-  virtual SymbolTable *MutableInputSymbols() = 0;
-  virtual SymbolTable *MutableOutputSymbols() = 0;
-  virtual int64 NumStates() const = 0;
-  virtual bool ReserveArcs(int64, size_t) = 0;
-  virtual void ReserveStates(int64) = 0;
-  virtual void SetInputSymbols(SymbolTable *) = 0;
-  virtual bool SetFinal(int64, const WeightClass &) = 0;
-  virtual void SetOutputSymbols(SymbolTable *) = 0;
-  virtual void SetProperties(uint64, uint64) = 0;
-  virtual bool SetStart(int64) = 0;
-  ~FstClassImplBase() override {}
-};
-
-// Containiner class wrapping an Fst<Arc>, hiding its arc type. Whether this
-// Fst<Arc> pointer refers to a special kind of FST (e.g. a MutableFst) is
-// known by the type of interface class that owns the pointer to this
-// container.
-
-template <class Arc>
-class FstClassImpl : public FstClassImplBase {
- public:
-  explicit FstClassImpl(Fst<Arc> *impl, bool should_own = false)
-      : impl_(should_own ? impl : impl->Copy()) {}
-
-  explicit FstClassImpl(const Fst<Arc> &impl) : impl_(impl.Copy()) {}
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  bool AddArc(int64 s, const ArcClass &ac) final {
-    if (!ValidStateId(s)) return false;
-    // Note that we do not check that the destination state is valid, so users
-    // can add arcs before they add the corresponding states. Verify can be
-    // used to determine whether any arc has a nonexisting destination.
-    Arc arc(ac.ilabel, ac.olabel, *ac.weight.GetWeight<typename Arc::Weight>(),
-            ac.nextstate);
-    static_cast<MutableFst<Arc> *>(impl_.get())->AddArc(s, arc);
-    return true;
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  int64 AddState() final {
-    return static_cast<MutableFst<Arc> *>(impl_.get())->AddState();
-  }
-
-  const string &ArcType() const final { return Arc::Type(); }
-
-  FstClassImpl *Copy() final { return new FstClassImpl<Arc>(impl_.get()); }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  bool DeleteArcs(int64 s, size_t n) final {
-    if (!ValidStateId(s)) return false;
-    static_cast<MutableFst<Arc> *>(impl_.get())->DeleteArcs(s, n);
-    return true;
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  bool DeleteArcs(int64 s) final {
-    if (!ValidStateId(s)) return false;
-    static_cast<MutableFst<Arc> *>(impl_.get())->DeleteArcs(s);
-    return true;
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  bool DeleteStates(const std::vector<int64> &dstates) final {
-    for (const auto &state : dstates)
-      if (!ValidStateId(state)) return false;
-    // Warning: calling this method with any integers beyond the precision of
-    // the underlying FST will result in truncation.
-    std::vector<typename Arc::StateId> typed_dstates(dstates.size());
-    std::copy(dstates.begin(), dstates.end(), typed_dstates.begin());
-    static_cast<MutableFst<Arc> *>(impl_.get())->DeleteStates(typed_dstates);
-    return true;
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  void DeleteStates() final {
-    static_cast<MutableFst<Arc> *>(impl_.get())->DeleteStates();
-  }
-
-  WeightClass Final(int64 s) const final {
-    if (!ValidStateId(s)) return WeightClass::NoWeight(WeightType());
-    WeightClass w(impl_->Final(s));
-    return w;
-  }
-
-  const string &FstType() const final { return impl_->Type(); }
-
-  const SymbolTable *InputSymbols() const final {
-    return impl_->InputSymbols();
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  SymbolTable *MutableInputSymbols() final {
-    return static_cast<MutableFst<Arc> *>(impl_.get())->MutableInputSymbols();
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  SymbolTable *MutableOutputSymbols() final {
-    return static_cast<MutableFst<Arc> *>(impl_.get())->MutableOutputSymbols();
-  }
-
-  // Signals failure by returning size_t max.
-  size_t NumArcs(int64 s) const final {
-    return ValidStateId(s) ? impl_->NumArcs(s)
-                           : std::numeric_limits<size_t>::max();
-  }
-
-  // Signals failure by returning size_t max.
-  size_t NumInputEpsilons(int64 s) const final {
-    return ValidStateId(s) ? impl_->NumInputEpsilons(s)
-                           : std::numeric_limits<size_t>::max();
-  }
-
-  // Signals failure by returning size_t max.
-  size_t NumOutputEpsilons(int64 s) const final {
-    return ValidStateId(s) ? impl_->NumOutputEpsilons(s)
-                           : std::numeric_limits<size_t>::max();
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  int64 NumStates() const final {
-    return static_cast<MutableFst<Arc> *>(impl_.get())->NumStates();
-  }
-
-  uint64 Properties(uint64 mask, bool test) const final {
-    return impl_->Properties(mask, test);
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  bool ReserveArcs(int64 s, size_t n) final {
-    if (!ValidStateId(s)) return false;
-    static_cast<MutableFst<Arc> *>(impl_.get())->ReserveArcs(s, n);
-    return true;
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  void ReserveStates(int64 s) final {
-    static_cast<MutableFst<Arc> *>(impl_.get())->ReserveStates(s);
-  }
-
-  const SymbolTable *OutputSymbols() const final {
-    return impl_->OutputSymbols();
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  void SetInputSymbols(SymbolTable *isyms) final {
-    static_cast<MutableFst<Arc> *>(impl_.get())->SetInputSymbols(isyms);
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  bool SetFinal(int64 s, const WeightClass &weight) final {
-    if (!ValidStateId(s)) return false;
-    static_cast<MutableFst<Arc> *>(impl_.get())
-        ->SetFinal(s, *weight.GetWeight<typename Arc::Weight>());
-    return true;
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  void SetOutputSymbols(SymbolTable *osyms) final {
-    static_cast<MutableFst<Arc> *>(impl_.get())->SetOutputSymbols(osyms);
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  void SetProperties(uint64 props, uint64 mask) final {
-    static_cast<MutableFst<Arc> *>(impl_.get())->SetProperties(props, mask);
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  bool SetStart(int64 s) final {
-    if (!ValidStateId(s)) return false;
-    static_cast<MutableFst<Arc> *>(impl_.get())->SetStart(s);
-    return true;
-  }
-
-  int64 Start() const final { return impl_->Start(); }
-
-  bool ValidStateId(int64 s) const final {
-    // This cowardly refuses to count states if the FST is not yet expanded.
-    if (!Properties(kExpanded, true)) {
-      FSTERROR() << "Cannot get number of states for unexpanded FST";
-      return false;
-    }
-    // If the FST is already expanded, CountStates calls NumStates.
-    if (s < 0 || s >= CountStates(*impl_)) {
-      FSTERROR() << "State ID " << s << " not valid";
-      return false;
-    }
-    return true;
-  }
-
-  const string &WeightType() const final { return Arc::Weight::Type(); }
-
-  bool Write(const string &fname) const final { return impl_->Write(fname); }
-
-  bool Write(std::ostream &ostr, const string &fname) const final {
-    const FstWriteOptions opts(fname);
-    return impl_->Write(ostr, opts);
-  }
-
-  ~FstClassImpl() override {}
-
-  Fst<Arc> *GetImpl() const { return impl_.get(); }
-
- private:
-  std::unique_ptr<Fst<Arc>> impl_;
-};
-
-// BASE CLASS DEFINITIONS
-
-class MutableFstClass;
-
-class FstClass : public FstClassBase {
- public:
-  FstClass() : impl_(nullptr) {}
-
-  template <class Arc>
-  explicit FstClass(const Fst<Arc> &fst) : impl_(new FstClassImpl<Arc>(fst)) {}
-
-  FstClass(const FstClass &other)
-      : impl_(other.impl_ == nullptr ? nullptr : other.impl_->Copy()) {}
-
-  FstClass &operator=(const FstClass &other) {
-    impl_.reset(other.impl_ == nullptr ? nullptr : other.impl_->Copy());
-    return *this;
-  }
-
-  WeightClass Final(int64 s) const final { return impl_->Final(s); }
-
-  const string &ArcType() const final { return impl_->ArcType(); }
-
-  const string &FstType() const final { return impl_->FstType(); }
-
-  const SymbolTable *InputSymbols() const final {
-    return impl_->InputSymbols();
-  }
-
-  size_t NumArcs(int64 s) const final { return impl_->NumArcs(s); }
-
-  size_t NumInputEpsilons(int64 s) const final {
-    return impl_->NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(int64 s) const final {
-    return impl_->NumOutputEpsilons(s);
-  }
-
-  const SymbolTable *OutputSymbols() const final {
-    return impl_->OutputSymbols();
-  }
-
-  uint64 Properties(uint64 mask, bool test) const final {
-    // Special handling for FSTs with a null impl.
-    if (!impl_) return kError & mask;
-    return impl_->Properties(mask, test);
-  }
-
-  static FstClass *Read(const string &fname);
-
-  static FstClass *Read(std::istream &istrm, const string &source);
-
-  int64 Start() const final { return impl_->Start(); }
-
-  bool ValidStateId(int64 s) const final { return impl_->ValidStateId(s); }
-
-  const string &WeightType() const final { return impl_->WeightType(); }
-
-  // Helper that logs an ERROR if the weight type of an FST and a WeightClass
-  // don't match.
-
-  bool WeightTypesMatch(const WeightClass &weight, const string &op_name) const;
-
-  bool Write(const string &fname) const final { return impl_->Write(fname); }
-
-  bool Write(std::ostream &ostr, const string &fname) const final {
-    return impl_->Write(ostr, fname);
-  }
-
-  ~FstClass() override {}
-
-  // These methods are required by IO registration.
-
-  template <class Arc>
-  static FstClassImplBase *Convert(const FstClass &other) {
-    FSTERROR() << "Doesn't make sense to convert any class to type FstClass";
-    return nullptr;
-  }
-
-  template <class Arc>
-  static FstClassImplBase *Create() {
-    FSTERROR() << "Doesn't make sense to create an FstClass with a "
-               << "particular arc type";
-    return nullptr;
-  }
-
-  template <class Arc>
-  const Fst<Arc> *GetFst() const {
-    if (Arc::Type() != ArcType()) {
-      return nullptr;
-    } else {
-      FstClassImpl<Arc> *typed_impl =
-          static_cast<FstClassImpl<Arc> *>(impl_.get());
-      return typed_impl->GetImpl();
-    }
-  }
-
-  template <class Arc>
-  static FstClass *Read(std::istream &stream, const FstReadOptions &opts) {
-    if (!opts.header) {
-      LOG(ERROR) << "FstClass::Read: Options header not specified";
-      return nullptr;
-    }
-    const FstHeader &hdr = *opts.header;
-    if (hdr.Properties() & kMutable) {
-      return ReadTypedFst<MutableFstClass, MutableFst<Arc>>(stream, opts);
-    } else {
-      return ReadTypedFst<FstClass, Fst<Arc>>(stream, opts);
-    }
-  }
-
- protected:
-  explicit FstClass(FstClassImplBase *impl) : impl_(impl) {}
-
-  const FstClassImplBase *GetImpl() const { return impl_.get(); }
-
-  FstClassImplBase *GetImpl() { return impl_.get(); }
-
-  // Generic template method for reading an arc-templated FST of type
-  // UnderlyingT, and returning it wrapped as FstClassT, with appropriat
-  // error checking. Called from arc-templated Read() static methods.
-  template <class FstClassT, class UnderlyingT>
-  static FstClassT *ReadTypedFst(std::istream &stream,
-                                 const FstReadOptions &opts) {
-    std::unique_ptr<UnderlyingT> u(UnderlyingT::Read(stream, opts));
-    return u ? new FstClassT(*u) : nullptr;
-  }
-
- private:
-  std::unique_ptr<FstClassImplBase> impl_;
-};
-
-// Specific types of FstClass with special properties
-
-class MutableFstClass : public FstClass {
- public:
-  bool AddArc(int64 s, const ArcClass &ac) {
-    if (!WeightTypesMatch(ac.weight, "AddArc")) return false;
-    return GetImpl()->AddArc(s, ac);
-  }
-
-  int64 AddState() { return GetImpl()->AddState(); }
-
-  bool DeleteArcs(int64 s, size_t n) { return GetImpl()->DeleteArcs(s, n); }
-
-  bool DeleteArcs(int64 s) { return GetImpl()->DeleteArcs(s); }
-
-  bool DeleteStates(const std::vector<int64> &dstates) {
-    return GetImpl()->DeleteStates(dstates);
-  }
-
-  void DeleteStates() { GetImpl()->DeleteStates(); }
-
-  SymbolTable *MutableInputSymbols() {
-    return GetImpl()->MutableInputSymbols();
-  }
-
-  SymbolTable *MutableOutputSymbols() {
-    return GetImpl()->MutableOutputSymbols();
-  }
-
-  int64 NumStates() const { return GetImpl()->NumStates(); }
-
-  bool ReserveArcs(int64 s, size_t n) { return GetImpl()->ReserveArcs(s, n); }
-
-  void ReserveStates(int64 s) { GetImpl()->ReserveStates(s); }
-
-  static MutableFstClass *Read(const string &fname, bool convert = false);
-
-  void SetInputSymbols(SymbolTable *isyms) {
-    GetImpl()->SetInputSymbols(isyms);
-  }
-
-  bool SetFinal(int64 s, const WeightClass &weight) {
-    if (!WeightTypesMatch(weight, "SetFinal")) return false;
-    return GetImpl()->SetFinal(s, weight);
-  }
-
-  void SetOutputSymbols(SymbolTable *osyms) {
-    GetImpl()->SetOutputSymbols(osyms);
-  }
-
-  void SetProperties(uint64 props, uint64 mask) {
-    GetImpl()->SetProperties(props, mask);
-  }
-
-  bool SetStart(int64 s) { return GetImpl()->SetStart(s); }
-
-  template <class Arc>
-  explicit MutableFstClass(const MutableFst<Arc> &fst) : FstClass(fst) {}
-
-  // These methods are required by IO registration.
-
-  template <class Arc>
-  static FstClassImplBase *Convert(const FstClass &other) {
-    FSTERROR() << "Doesn't make sense to convert any class to type "
-               << "MutableFstClass";
-    return nullptr;
-  }
-
-  template <class Arc>
-  static FstClassImplBase *Create() {
-    FSTERROR() << "Doesn't make sense to create a MutableFstClass with a "
-               << "particular arc type";
-    return nullptr;
-  }
-
-  template <class Arc>
-  MutableFst<Arc> *GetMutableFst() {
-    Fst<Arc> *fst = const_cast<Fst<Arc> *>(this->GetFst<Arc>());
-    MutableFst<Arc> *mfst = static_cast<MutableFst<Arc> *>(fst);
-    return mfst;
-  }
-
-  template <class Arc>
-  static MutableFstClass *Read(std::istream &stream,
-                               const FstReadOptions &opts) {
-    std::unique_ptr<MutableFst<Arc>> mfst(MutableFst<Arc>::Read(stream, opts));
-    return mfst ? new MutableFstClass(*mfst) : nullptr;
-  }
-
- protected:
-  explicit MutableFstClass(FstClassImplBase *impl) : FstClass(impl) {}
-};
-
-class VectorFstClass : public MutableFstClass {
- public:
-  explicit VectorFstClass(FstClassImplBase *impl) : MutableFstClass(impl) {}
-
-  explicit VectorFstClass(const FstClass &other);
-
-  explicit VectorFstClass(const string &arc_type);
-
-  static VectorFstClass *Read(const string &fname);
-
-  template <class Arc>
-  static VectorFstClass *Read(std::istream &stream,
-                              const FstReadOptions &opts) {
-    std::unique_ptr<VectorFst<Arc>> mfst(VectorFst<Arc>::Read(stream, opts));
-    return mfst ? new VectorFstClass(*mfst) : nullptr;
-  }
-
-  template <class Arc>
-  explicit VectorFstClass(const VectorFst<Arc> &fst) : MutableFstClass(fst) {}
-
-  template <class Arc>
-  static FstClassImplBase *Convert(const FstClass &other) {
-    return new FstClassImpl<Arc>(new VectorFst<Arc>(*other.GetFst<Arc>()),
-                                 true);
-  }
-
-  template <class Arc>
-  static FstClassImplBase *Create() {
-    return new FstClassImpl<Arc>(new VectorFst<Arc>(), true);
-  }
-};
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_FST_CLASS_H_
diff --git a/projects/llm_framework/include/fst/script/fstscript-decl.h b/projects/llm_framework/include/fst/script/fstscript-decl.h
deleted file mode 100644
index 294d0159..00000000
--- a/projects/llm_framework/include/fst/script/fstscript-decl.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Forward declarations for the FST and FST script classes.
-
-#ifndef FST_SCRIPT_FSTSCRIPT_DECL_H_
-#define FST_SCRIPT_FSTSCRIPT_DECL_H_
-
-#include <fst/fst-decl.h>
-
-namespace fst {
-namespace script {
-
-class ArcClass;
-
-class ArcIteratorClass;
-class MutableArcIteratorClass;
-
-class EncodeMapperClass;
-
-class FstClass;
-class MutableFstClass;
-class VectorFstClass;
-
-class StateIteratorClass;
-
-class WeightClass;
-
-}  // namespace script
-}  // namespace fst;
-
-#endif  // FST_SCRIPT_FSTSCRIPT_DECL_H_
diff --git a/projects/llm_framework/include/fst/script/fstscript.h b/projects/llm_framework/include/fst/script/fstscript.h
deleted file mode 100644
index 45f16175..00000000
--- a/projects/llm_framework/include/fst/script/fstscript.h
+++ /dev/null
@@ -1,155 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// The FST script interface permits users to interact with FSTs without knowing
-// their arc type. It does this by mapping compile-time polymorphism (in the
-// form of a arc-templated FST types) onto a shared virtual interface. It also
-// supports arc extension via a DSO interface. Due to the overhead of virtual
-// dispatch and registered function lookups, the script API is somewhat slower
-// then library API provided by types like StdVectorFst, but has the advantage
-// that it is designed not to crash (and to provide useful debugging
-// information) upon common user errors like passing invalid indices or
-// attempting comparison of incompatible FSTs. It is used both by the FST
-// binaries and the Python extension.
-//
-// This header includes all of the FST script functionality.
-
-#ifndef FST_SCRIPT_FSTSCRIPT_H_
-#define FST_SCRIPT_FSTSCRIPT_H_
-
-// Major classes
-#include <fst/script/arciterator-class.h>
-#include <fst/script/encodemapper-class.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/stateiterator-class.h>
-#include <fst/script/text-io.h>
-#include <fst/script/weight-class.h>
-
-// Flag-to-enum parsers.
-#include <fst/script/getters.h>
-// Templates like Operation<> and Apply<>.
-#include <fst/script/script-impl.h>
-
-// Operations.
-#include <fst/script/arcsort.h>
-#include <fst/script/closure.h>
-#include <fst/script/compile.h>
-#include <fst/script/compose.h>
-#include <fst/script/concat.h>
-#include <fst/script/connect.h>
-#include <fst/script/convert.h>
-#include <fst/script/decode.h>
-#include <fst/script/determinize.h>
-#include <fst/script/difference.h>
-#include <fst/script/disambiguate.h>
-#include <fst/script/draw.h>
-#include <fst/script/encode.h>
-#include <fst/script/epsnormalize.h>
-#include <fst/script/equal.h>
-#include <fst/script/equivalent.h>
-#include <fst/script/info.h>
-#include <fst/script/intersect.h>
-#include <fst/script/invert.h>
-#include <fst/script/isomorphic.h>
-#include <fst/script/map.h>
-#include <fst/script/minimize.h>
-#include <fst/script/print.h>
-#include <fst/script/project.h>
-#include <fst/script/prune.h>
-#include <fst/script/push.h>
-#include <fst/script/randequivalent.h>
-#include <fst/script/randgen.h>
-#include <fst/script/relabel.h>
-#include <fst/script/replace.h>
-#include <fst/script/reverse.h>
-#include <fst/script/reweight.h>
-#include <fst/script/rmepsilon.h>
-#include <fst/script/shortest-distance.h>
-#include <fst/script/shortest-path.h>
-#include <fst/script/synchronize.h>
-#include <fst/script/topsort.h>
-#include <fst/script/union.h>
-#include <fst/script/verify.h>
-
-// This class is necessary because registering each of the operations
-// separately overfills the stack, as there's so many of them.
-namespace fst {
-namespace script {
-
-template <class Arc>
-class AllFstOperationsRegisterer {
- public:
-  AllFstOperationsRegisterer() {
-    RegisterBatch1();
-    RegisterBatch2();
-  }
-
- private:
-  void RegisterBatch1() {
-    REGISTER_FST_OPERATION(ArcSort, Arc, ArcSortArgs);
-    REGISTER_FST_OPERATION(Closure, Arc, ClosureArgs);
-    REGISTER_FST_OPERATION(CompileFstInternal, Arc, CompileFstArgs);
-    REGISTER_FST_OPERATION(Compose, Arc, ComposeArgs);
-    REGISTER_FST_OPERATION(Concat, Arc, ConcatArgs1);
-    REGISTER_FST_OPERATION(Concat, Arc, ConcatArgs2);
-    REGISTER_FST_OPERATION(Connect, Arc, MutableFstClass);
-    REGISTER_FST_OPERATION(Convert, Arc, ConvertArgs);
-    REGISTER_FST_OPERATION(Decode, Arc, DecodeArgs1);
-    REGISTER_FST_OPERATION(Decode, Arc, DecodeArgs2);
-    REGISTER_FST_OPERATION(Determinize, Arc, DeterminizeArgs);
-    REGISTER_FST_OPERATION(Difference, Arc, DifferenceArgs);
-    REGISTER_FST_OPERATION(Disambiguate, Arc, DisambiguateArgs);
-    REGISTER_FST_OPERATION(DrawFst, Arc, FstDrawerArgs);
-    REGISTER_FST_OPERATION(Encode, Arc, EncodeArgs1);
-    REGISTER_FST_OPERATION(Encode, Arc, EncodeArgs2);
-    REGISTER_FST_OPERATION(EpsNormalize, Arc, EpsNormalizeArgs);
-    REGISTER_FST_OPERATION(Equal, Arc, EqualArgs);
-    REGISTER_FST_OPERATION(Equivalent, Arc, EquivalentArgs);
-    REGISTER_FST_OPERATION(PrintFstInfo, Arc, InfoArgs);
-    REGISTER_FST_OPERATION(GetFstInfo, Arc, GetInfoArgs);
-    REGISTER_FST_OPERATION(InitArcIteratorClass, Arc,
-                           InitArcIteratorClassArgs);
-    REGISTER_FST_OPERATION(InitEncodeMapperClass, Arc,
-                           InitEncodeMapperClassArgs);
-    REGISTER_FST_OPERATION(InitMutableArcIteratorClass, Arc,
-                           InitMutableArcIteratorClassArgs);
-    REGISTER_FST_OPERATION(InitStateIteratorClass, Arc,
-                           InitStateIteratorClassArgs);
-  }
-
-  void RegisterBatch2() {
-    REGISTER_FST_OPERATION(Intersect, Arc, IntersectArgs);
-    REGISTER_FST_OPERATION(Invert, Arc, MutableFstClass);
-    REGISTER_FST_OPERATION(Map, Arc, MapArgs);
-    REGISTER_FST_OPERATION(Minimize, Arc, MinimizeArgs);
-    REGISTER_FST_OPERATION(PrintFst, Arc, FstPrinterArgs);
-    REGISTER_FST_OPERATION(Project, Arc, ProjectArgs);
-    REGISTER_FST_OPERATION(Prune, Arc, PruneArgs1);
-    REGISTER_FST_OPERATION(Prune, Arc, PruneArgs2);
-    REGISTER_FST_OPERATION(Push, Arc, PushArgs1);
-    REGISTER_FST_OPERATION(Push, Arc, PushArgs2);
-    REGISTER_FST_OPERATION(RandEquivalent, Arc, RandEquivalentArgs);
-    REGISTER_FST_OPERATION(RandGen, Arc, RandGenArgs);
-    REGISTER_FST_OPERATION(Relabel, Arc, RelabelArgs1);
-    REGISTER_FST_OPERATION(Relabel, Arc, RelabelArgs2);
-    REGISTER_FST_OPERATION(Replace, Arc, ReplaceArgs);
-    REGISTER_FST_OPERATION(Reverse, Arc, ReverseArgs);
-    REGISTER_FST_OPERATION(Reweight, Arc, ReweightArgs);
-    REGISTER_FST_OPERATION(RmEpsilon, Arc, RmEpsilonArgs);
-    REGISTER_FST_OPERATION(ShortestDistance, Arc, ShortestDistanceArgs1);
-    REGISTER_FST_OPERATION(ShortestDistance, Arc, ShortestDistanceArgs2);
-    REGISTER_FST_OPERATION(ShortestPath, Arc, ShortestPathArgs);
-    REGISTER_FST_OPERATION(Synchronize, Arc, SynchronizeArgs);
-    REGISTER_FST_OPERATION(TopSort, Arc, TopSortArgs);
-    REGISTER_FST_OPERATION(Union, Arc, UnionArgs);
-    REGISTER_FST_OPERATION(Verify, Arc, VerifyArgs);
-  }
-};
-
-}  // namespace script
-}  // namespace fst
-
-#define REGISTER_FST_OPERATIONS(Arc) \
-  AllFstOperationsRegisterer<Arc> register_all_fst_operations##Arc;
-
-#endif  // FST_SCRIPT_FSTSCRIPT_H_
diff --git a/projects/llm_framework/include/fst/script/getters.h b/projects/llm_framework/include/fst/script/getters.h
deleted file mode 100644
index 5cd727e8..00000000
--- a/projects/llm_framework/include/fst/script/getters.h
+++ /dev/null
@@ -1,76 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Getters for converting command-line arguments into the appropriate enums
-// or bitmasks, with the simplest ones defined as inline.
-
-#ifndef FST_SCRIPT_GETTERS_H_
-#define FST_SCRIPT_GETTERS_H_
-
-#include <string>
-
-#include <fst/compose.h>          // For ComposeFilter.
-#include <fst/determinize.h>      // For DeterminizeType.
-#include <fst/encode.h>           // For kEncodeLabels (etc.).
-#include <fst/epsnormalize.h>     // For EpsNormalizeType.
-#include <fst/project.h>          // For ProjectType.
-#include <fst/push.h>             // For kPushWeights (etc.).
-#include <fst/queue.h>            // For QueueType.
-#include <fst/rational.h>         // For ClosureType.
-#include <fst/script/arcsort.h>       // For ArcSortType.
-#include <fst/script/map.h>           // For MapType.
-#include <fst/script/script-impl.h>   // For RandArcSelection.
-
-#include <fst/log.h>
-
-namespace fst {
-namespace script {
-
-bool GetArcSortType(const string &str, ArcSortType *sort_type);
-
-inline ClosureType GetClosureType(bool closure_plus) {
-  return closure_plus ? CLOSURE_PLUS : CLOSURE_STAR;
-}
-
-bool GetComposeFilter(const string &str, ComposeFilter *compose_filter);
-
-bool GetDeterminizeType(const string &str, DeterminizeType *det_type);
-
-inline uint32 GetEncodeFlags(bool encode_labels, bool encode_weights) {
-  return (encode_labels ? kEncodeLabels : 0) |
-         (encode_weights ? kEncodeWeights : 0);
-}
-
-inline EpsNormalizeType GetEpsNormalizeType(bool eps_norm_output) {
-  return eps_norm_output ? EPS_NORM_OUTPUT : EPS_NORM_INPUT;
-}
-
-bool GetMapType(const string &str, MapType *map_type);
-
-inline ProjectType GetProjectType(bool project_output) {
-  return project_output ? PROJECT_OUTPUT : PROJECT_INPUT;
-}
-
-inline uint32 GetPushFlags(bool push_weights, bool push_labels,
-                           bool remove_total_weight, bool remove_common_affix) {
-  return ((push_weights ? kPushWeights : 0) |
-          (push_labels ? kPushLabels : 0) |
-          (remove_total_weight ? kPushRemoveTotalWeight : 0) |
-          (remove_common_affix ? kPushRemoveCommonAffix : 0));
-}
-
-bool GetQueueType(const string &str, QueueType *queue_type);
-
-bool GetRandArcSelection(const string &str, RandArcSelection *ras);
-
-bool GetReplaceLabelType(const string &str, bool epsilon_on_replace,
-                         ReplaceLabelType *rlt);
-
-inline ReweightType GetReweightType(bool to_final) {
-  return to_final ? REWEIGHT_TO_FINAL : REWEIGHT_TO_INITIAL;
-}
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_GETTERS_H_
diff --git a/projects/llm_framework/include/fst/script/info-impl.h b/projects/llm_framework/include/fst/script/info-impl.h
deleted file mode 100644
index e8956498..00000000
--- a/projects/llm_framework/include/fst/script/info-impl.h
+++ /dev/null
@@ -1,314 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to compute various information about FSTs, a helper class for
-// fstinfo.cc.
-
-#ifndef FST_SCRIPT_INFO_IMPL_H_
-#define FST_SCRIPT_INFO_IMPL_H_
-
-#include <map>
-#include <string>
-#include <vector>
-
-#include <fst/connect.h>
-#include <fst/dfs-visit.h>
-#include <fst/fst.h>
-#include <fst/lookahead-matcher.h>
-#include <fst/matcher.h>
-#include <fst/queue.h>
-#include <fst/test-properties.h>
-#include <fst/verify.h>
-#include <fst/visit.h>
-
-namespace fst {
-
-// Compute various information about FSTs, helper class for fstinfo.cc.
-// WARNING: Stand-alone use of this class is not recommended, most code
-// should call directly the relevant library functions: Fst<Arc>::NumStates,
-// Fst<Arc>::NumArcs, TestProperties, etc.
-class FstInfo {
- public:
-  FstInfo() {}
-
-  // When info_type is "short" (or "auto" and not an ExpandedFst) then only
-  // minimal info is computed and can be requested.
-  template <typename Arc>
-  FstInfo(const Fst<Arc> &fst, bool test_properties,
-          const string &arc_filter_type = "any",
-          const string &info_type = "auto", bool verify = true)
-      : fst_type_(fst.Type()),
-        input_symbols_(fst.InputSymbols() ? fst.InputSymbols()->Name()
-                                          : "none"),
-        output_symbols_(fst.OutputSymbols() ? fst.OutputSymbols()->Name()
-                                            : "none"),
-        nstates_(0),
-        narcs_(0),
-        start_(kNoStateId),
-        nfinal_(0),
-        nepsilons_(0),
-        niepsilons_(0),
-        noepsilons_(0),
-        ilabel_mult_(0.0),
-        olabel_mult_(0.0),
-        naccess_(0),
-        ncoaccess_(0),
-        nconnect_(0),
-        ncc_(0),
-        nscc_(0),
-        input_match_type_(MATCH_NONE),
-        output_match_type_(MATCH_NONE),
-        input_lookahead_(false),
-        output_lookahead_(false),
-        properties_(0),
-        arc_filter_type_(arc_filter_type),
-        long_info_(true),
-        arc_type_(Arc::Type()) {
-    using Label = typename Arc::Label;
-    using StateId = typename Arc::StateId;
-    using Weight = typename Arc::Weight;
-    if (info_type == "long") {
-      long_info_ = true;
-    } else if (info_type == "short") {
-      long_info_ = false;
-    } else if (info_type == "auto") {
-      long_info_ = fst.Properties(kExpanded, false);
-    } else {
-      FSTERROR() << "Bad info type: " << info_type;
-      return;
-    }
-    if (!long_info_) return;
-    // If the FST is not sane, we return.
-    if (verify && !Verify(fst)) {
-      FSTERROR() << "FstInfo: Verify: FST not well-formed";
-      return;
-    }
-    start_ = fst.Start();
-    properties_ = fst.Properties(kFstProperties, test_properties);
-    for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-      ++nstates_;
-      const auto s = siter.Value();
-      if (fst.Final(s) != Weight::Zero()) ++nfinal_;
-      std::map<Label, size_t> ilabel_count;
-      std::map<Label, size_t> olabel_count;
-      for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-        const auto &arc = aiter.Value();
-        ++narcs_;
-        if (arc.ilabel == 0 && arc.olabel == 0) ++nepsilons_;
-        if (arc.ilabel == 0) ++niepsilons_;
-        if (arc.olabel == 0) ++noepsilons_;
-        ++ilabel_count[arc.ilabel];
-        ++olabel_count[arc.olabel];
-      }
-      for (auto it = ilabel_count.begin(); it != ilabel_count.end(); ++it) {
-        ilabel_mult_ += it->second * it->second;
-      }
-      for (auto it = olabel_count.begin(); it != olabel_count.end(); ++it) {
-        olabel_mult_ += it->second * it->second;
-      }
-    }
-    if (narcs_ > 0) {
-      ilabel_mult_ /= narcs_;
-      olabel_mult_ /= narcs_;
-    }
-    {
-      std::vector<StateId> cc;
-      CcVisitor<Arc> cc_visitor(&cc);
-      FifoQueue<StateId> fifo_queue;
-      if (arc_filter_type == "any") {
-        Visit(fst, &cc_visitor, &fifo_queue);
-      } else if (arc_filter_type == "epsilon") {
-        Visit(fst, &cc_visitor, &fifo_queue, EpsilonArcFilter<Arc>());
-      } else if (arc_filter_type == "iepsilon") {
-        Visit(fst, &cc_visitor, &fifo_queue, InputEpsilonArcFilter<Arc>());
-      } else if (arc_filter_type == "oepsilon") {
-        Visit(fst, &cc_visitor, &fifo_queue, OutputEpsilonArcFilter<Arc>());
-      } else {
-        FSTERROR() << "Bad arc filter type: " << arc_filter_type;
-        return;
-      }
-      for (StateId s = 0; s < cc.size(); ++s) {
-        if (cc[s] >= ncc_) ncc_ = cc[s] + 1;
-      }
-    }
-    {
-      std::vector<StateId> scc;
-      std::vector<bool> access, coaccess;
-      uint64 props = 0;
-      SccVisitor<Arc> scc_visitor(&scc, &access, &coaccess, &props);
-      if (arc_filter_type == "any") {
-        DfsVisit(fst, &scc_visitor);
-      } else if (arc_filter_type == "epsilon") {
-        DfsVisit(fst, &scc_visitor, EpsilonArcFilter<Arc>());
-      } else if (arc_filter_type == "iepsilon") {
-        DfsVisit(fst, &scc_visitor, InputEpsilonArcFilter<Arc>());
-      } else if (arc_filter_type == "oepsilon") {
-        DfsVisit(fst, &scc_visitor, OutputEpsilonArcFilter<Arc>());
-      } else {
-        FSTERROR() << "Bad arc filter type: " << arc_filter_type;
-        return;
-      }
-      for (StateId s = 0; s < scc.size(); ++s) {
-        if (access[s]) ++naccess_;
-        if (coaccess[s]) ++ncoaccess_;
-        if (access[s] && coaccess[s]) ++nconnect_;
-        if (scc[s] >= nscc_) nscc_ = scc[s] + 1;
-      }
-    }
-    LookAheadMatcher<Fst<Arc>> imatcher(fst, MATCH_INPUT);
-    input_match_type_ = imatcher.Type(test_properties);
-    input_lookahead_ = imatcher.Flags() & kInputLookAheadMatcher;
-    LookAheadMatcher<Fst<Arc>> omatcher(fst, MATCH_OUTPUT);
-    output_match_type_ = omatcher.Type(test_properties);
-    output_lookahead_ = omatcher.Flags() & kOutputLookAheadMatcher;
-  }
-
-  // Short info.
-
-  const string &FstType() const { return fst_type_; }
-
-  const string &ArcType() const { return arc_type_; }
-
-  const string &InputSymbols() const { return input_symbols_; }
-
-  const string &OutputSymbols() const { return output_symbols_; }
-
-  bool LongInfo() const { return long_info_; }
-
-  const string &ArcFilterType() const { return arc_filter_type_; }
-
-  // Long info.
-
-  MatchType InputMatchType() const {
-    CheckLong();
-    return input_match_type_;
-  }
-
-  MatchType OutputMatchType() const {
-    CheckLong();
-    return output_match_type_;
-  }
-
-  bool InputLookAhead() const {
-    CheckLong();
-    return input_lookahead_;
-  }
-
-  bool OutputLookAhead() const {
-    CheckLong();
-    return output_lookahead_;
-  }
-
-  int64 NumStates() const {
-    CheckLong();
-    return nstates_;
-  }
-
-  size_t NumArcs() const {
-    CheckLong();
-    return narcs_;
-  }
-
-  int64 Start() const {
-    CheckLong();
-    return start_;
-  }
-
-  size_t NumFinal() const {
-    CheckLong();
-    return nfinal_;
-  }
-
-  size_t NumEpsilons() const {
-    CheckLong();
-    return nepsilons_;
-  }
-
-  size_t NumInputEpsilons() const {
-    CheckLong();
-    return niepsilons_;
-  }
-
-  size_t NumOutputEpsilons() const {
-    CheckLong();
-    return noepsilons_;
-  }
-
-  double InputLabelMultiplicity() const {
-    CheckLong();
-    return ilabel_mult_;
-  }
-
-  double OutputLabelMultiplicity() const {
-    CheckLong();
-    return olabel_mult_;
-  }
-
-  size_t NumAccessible() const {
-    CheckLong();
-    return naccess_;
-  }
-
-  size_t NumCoAccessible() const {
-    CheckLong();
-    return ncoaccess_;
-  }
-
-  size_t NumConnected() const {
-    CheckLong();
-    return nconnect_;
-  }
-
-  size_t NumCc() const {
-    CheckLong();
-    return ncc_;
-  }
-
-  size_t NumScc() const {
-    CheckLong();
-    return nscc_;
-  }
-
-  uint64 Properties() const {
-    CheckLong();
-    return properties_;
-  }
-
- private:
-  void CheckLong() const {
-    if (!long_info_)
-      FSTERROR() << "FstInfo: Method only available with long info signature";
-  }
-
-  string fst_type_;
-  string input_symbols_;
-  string output_symbols_;
-  int64 nstates_;
-  size_t narcs_;
-  int64 start_;
-  size_t nfinal_;
-  size_t nepsilons_;
-  size_t niepsilons_;
-  size_t noepsilons_;
-  double ilabel_mult_;
-  double olabel_mult_;
-  size_t naccess_;
-  size_t ncoaccess_;
-  size_t nconnect_;
-  size_t ncc_;
-  size_t nscc_;
-  MatchType input_match_type_;
-  MatchType output_match_type_;
-  bool input_lookahead_;
-  bool output_lookahead_;
-  uint64 properties_;
-  string arc_filter_type_;
-  bool long_info_;
-  string arc_type_;
-};
-
-void PrintFstInfoImpl(const FstInfo &fstinfo, bool pipe = false);
-
-}  // namespace fst
-
-#endif  // FST_SCRIPT_INFO_IMPL_H_
diff --git a/projects/llm_framework/include/fst/script/info.h b/projects/llm_framework/include/fst/script/info.h
deleted file mode 100644
index 039d06d8..00000000
--- a/projects/llm_framework/include/fst/script/info.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_INFO_H_
-#define FST_SCRIPT_INFO_H_
-
-#include <string>
-#include <tuple>
-
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/info-impl.h>
-
-namespace fst {
-namespace script {
-
-using InfoArgs = std::tuple<const FstClass &, bool, const string &,
-                            const string &, bool, bool>;
-
-template <class Arc>
-void PrintFstInfo(InfoArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  const FstInfo fstinfo(fst, std::get<1>(*args), std::get<2>(*args),
-                        std::get<3>(*args), std::get<4>(*args));
-  PrintFstInfoImpl(fstinfo, std::get<5>(*args));
-  if (std::get<5>(*args)) fst.Write("");
-}
-
-void PrintFstInfo(const FstClass &f, bool test_properties,
-                  const string &arc_filter, const string &info_type, bool pipe,
-                  bool verify);
-
-using GetInfoArgs = std::tuple<const FstClass &, bool, const string &,
-                               const string &, bool, FstInfo *>;
-
-template <class Arc>
-void GetFstInfo(GetInfoArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  *(std::get<5>(*args)) = FstInfo(fst, std::get<1>(*args), std::get<2>(*args),
-                                  std::get<3>(*args), std::get<4>(*args));
-}
-
-void GetFstInfo(const FstClass &fst, bool test_properties,
-                const string &arc_filter, const string &info_type, bool verify,
-                FstInfo *info);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_INFO_H_
diff --git a/projects/llm_framework/include/fst/script/intersect.h b/projects/llm_framework/include/fst/script/intersect.h
deleted file mode 100644
index 229bd56f..00000000
--- a/projects/llm_framework/include/fst/script/intersect.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_INTERSECT_H_
-#define FST_SCRIPT_INTERSECT_H_
-
-#include <tuple>
-
-#include <fst/intersect.h>
-#include <fst/script/compose.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using IntersectArgs = std::tuple<const FstClass &, const FstClass &,
-                                 MutableFstClass *, const ComposeOptions &>;
-
-template <class Arc>
-void Intersect(IntersectArgs *args) {
-  const Fst<Arc> &ifst1 = *(std::get<0>(*args).GetFst<Arc>());
-  const Fst<Arc> &ifst2 = *(std::get<1>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
-  const auto &opts = std::get<3>(*args);
-  Intersect(ifst1, ifst2, ofst, opts);
-}
-
-void Intersect(const FstClass &ifst, const FstClass &ifst2,
-               MutableFstClass *ofst,
-               const ComposeOptions &opts = ComposeOptions());
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_INTERSECT_H_
diff --git a/projects/llm_framework/include/fst/script/invert.h b/projects/llm_framework/include/fst/script/invert.h
deleted file mode 100644
index 5bc31317..00000000
--- a/projects/llm_framework/include/fst/script/invert.h
+++ /dev/null
@@ -1,23 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_INVERT_H_
-#define FST_SCRIPT_INVERT_H_
-
-#include <fst/invert.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-template <class Arc>
-void Invert(MutableFstClass *fst) {
-  Invert(fst->GetMutableFst<Arc>());
-}
-
-void Invert(MutableFstClass *fst);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_INVERT_H_
diff --git a/projects/llm_framework/include/fst/script/isomorphic.h b/projects/llm_framework/include/fst/script/isomorphic.h
deleted file mode 100644
index 94ea77f9..00000000
--- a/projects/llm_framework/include/fst/script/isomorphic.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_ISOMORPHIC_H_
-#define FST_SCRIPT_ISOMORPHIC_H_
-
-#include <tuple>
-
-#include <fst/isomorphic.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using IsomorphicInnerArgs = std::tuple<const FstClass &, const FstClass &,
-                                       float>;
-
-using IsomorphicArgs = WithReturnValue<bool, IsomorphicInnerArgs>;
-
-template <class Arc>
-void Isomorphic(IsomorphicArgs *args) {
-  const Fst<Arc> &fst1 = *(std::get<0>(args->args).GetFst<Arc>());
-  const Fst<Arc> &fst2 = *(std::get<1>(args->args).GetFst<Arc>());
-  args->retval = Isomorphic(fst1, fst2, std::get<2>(args->args));
-}
-
-bool Isomorphic(const FstClass &fst1, const FstClass &fst2,
-                float delta = kDelta);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_ISOMORPHIC_H_
diff --git a/projects/llm_framework/include/fst/script/map.h b/projects/llm_framework/include/fst/script/map.h
deleted file mode 100644
index 158d98aa..00000000
--- a/projects/llm_framework/include/fst/script/map.h
+++ /dev/null
@@ -1,158 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_MAP_H_
-#define FST_SCRIPT_MAP_H_
-
-#include <memory>
-#include <tuple>
-
-#include <fst/arc-map.h>
-#include <fst/state-map.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-template <class M>
-Fst<typename M::ToArc> *ArcMap(const Fst<typename M::FromArc> &fst,
-                               const M &mapper) {
-  using ToArc = typename M::ToArc;
-  auto *ofst = new VectorFst<ToArc>;
-  ArcMap(fst, ofst, mapper);
-  return ofst;
-}
-
-template <class M>
-Fst<typename M::ToArc> *StateMap(const Fst<typename M::FromArc> &fst,
-                                 const M &mapper) {
-  using ToArc = typename M::ToArc;
-  auto *ofst = new VectorFst<ToArc>;
-  StateMap(fst, ofst, mapper);
-  return ofst;
-}
-
-enum MapType {
-  ARC_SUM_MAPPER,
-  ARC_UNIQUE_MAPPER,
-  IDENTITY_MAPPER,
-  INPUT_EPSILON_MAPPER,
-  INVERT_MAPPER,
-  OUTPUT_EPSILON_MAPPER,
-  PLUS_MAPPER,
-  POWER_MAPPER,
-  QUANTIZE_MAPPER,
-  RMWEIGHT_MAPPER,
-  SUPERFINAL_MAPPER,
-  TIMES_MAPPER,
-  TO_LOG_MAPPER,
-  TO_LOG64_MAPPER,
-  TO_STD_MAPPER
-};
-
-using MapInnerArgs =
-    std::tuple<const FstClass &, MapType, float, double, const WeightClass &>;
-
-using MapArgs = WithReturnValue<FstClass *, MapInnerArgs>;
-
-template <class Arc>
-void Map(MapArgs *args) {
-  using Weight = typename Arc::Weight;
-  const Fst<Arc> &ifst = *(std::get<0>(args->args).GetFst<Arc>());
-  const auto map_type = std::get<1>(args->args);
-  switch (map_type) {
-    case ARC_SUM_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(StateMap(ifst, ArcSumMapper<Arc>(ifst)));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case ARC_UNIQUE_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(
-          StateMap(ifst, ArcUniqueMapper<Arc>(ifst)));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case IDENTITY_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, IdentityArcMapper<Arc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case INPUT_EPSILON_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, InputEpsilonMapper<Arc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case INVERT_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, InvertWeightMapper<Arc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case OUTPUT_EPSILON_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, OutputEpsilonMapper<Arc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case PLUS_MAPPER: {
-      const auto weight = *(std::get<4>(args->args).GetWeight<Weight>());
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, PlusMapper<Arc>(weight)));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case POWER_MAPPER: {
-      const auto power = std::get<3>(args->args);
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, PowerMapper<Arc>(power)));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case QUANTIZE_MAPPER: {
-      const auto delta = std::get<2>(args->args);
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, QuantizeMapper<Arc>(delta)));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case RMWEIGHT_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, RmWeightMapper<Arc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case SUPERFINAL_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, SuperFinalMapper<Arc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case TIMES_MAPPER: {
-      const auto weight = *(std::get<4>(args->args).GetWeight<Weight>());
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, TimesMapper<Arc>(weight)));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case TO_LOG_MAPPER: {
-      std::unique_ptr<Fst<LogArc>> ofst(
-          ArcMap(ifst, WeightConvertMapper<Arc, LogArc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case TO_LOG64_MAPPER: {
-      std::unique_ptr<Fst<Log64Arc>> ofst(
-          ArcMap(ifst, WeightConvertMapper<Arc, Log64Arc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case TO_STD_MAPPER: {
-      std::unique_ptr<Fst<StdArc>> ofst(
-          ArcMap(ifst, WeightConvertMapper<Arc, StdArc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-  }
-}
-
-FstClass *Map(const FstClass &ifst, MapType map_type, float delta, double power,
-              const WeightClass &weight);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_MAP_H_
diff --git a/projects/llm_framework/include/fst/script/minimize.h b/projects/llm_framework/include/fst/script/minimize.h
deleted file mode 100644
index 773e8ef1..00000000
--- a/projects/llm_framework/include/fst/script/minimize.h
+++ /dev/null
@@ -1,33 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_MINIMIZE_H_
-#define FST_SCRIPT_MINIMIZE_H_
-
-#include <tuple>
-
-#include <fst/minimize.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using MinimizeArgs = std::tuple<MutableFstClass *, MutableFstClass *, float,
-                                bool>;
-
-template <class Arc>
-void Minimize(MinimizeArgs *args) {
-  MutableFst<Arc> *ofst1 = std::get<0>(*args)->GetMutableFst<Arc>();
-  MutableFst<Arc> *ofst2 = (std::get<1>(*args) ?
-                            std::get<1>(*args)->GetMutableFst<Arc>() :
-                            nullptr);
-  Minimize(ofst1, ofst2, std::get<2>(*args), std::get<3>(*args));
-}
-
-void Minimize(MutableFstClass *ofst1, MutableFstClass *ofst2 = nullptr,
-              float delta = kShortestDelta, bool allow_nondet = false);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_MINIMIZE_H_
diff --git a/projects/llm_framework/include/fst/script/print-impl.h b/projects/llm_framework/include/fst/script/print-impl.h
deleted file mode 100644
index 539c6d8f..00000000
--- a/projects/llm_framework/include/fst/script/print-impl.h
+++ /dev/null
@@ -1,132 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Stand-alone class to print out binary FSTs in the AT&T format, a helper
-// class for fstprint.cc.
-
-#ifndef FST_SCRIPT_PRINT_IMPL_H_
-#define FST_SCRIPT_PRINT_IMPL_H_
-
-#include <ostream>
-#include <sstream>
-#include <string>
-
-#include <fst/fstlib.h>
-#include <fst/util.h>
-
-DECLARE_string(fst_field_separator);
-
-namespace fst {
-
-// Print a binary FST in textual format (helper class for fstprint.cc).
-// WARNING: Stand-alone use of this class not recommended, most code should
-// read/write using the binary format which is much more efficient.
-template <class Arc>
-class FstPrinter {
- public:
-  using StateId = typename Arc::StateId;
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-
-  FstPrinter(const Fst<Arc> &fst, const SymbolTable *isyms,
-             const SymbolTable *osyms, const SymbolTable *ssyms, bool accep,
-             bool show_weight_one, const string &field_separator,
-             const string &missing_symbol = "")
-      : fst_(fst),
-        isyms_(isyms),
-        osyms_(osyms),
-        ssyms_(ssyms),
-        accep_(accep && fst.Properties(kAcceptor, true)),
-        ostrm_(nullptr),
-        show_weight_one_(show_weight_one),
-        sep_(field_separator),
-        missing_symbol_(missing_symbol) {}
-
-  // Prints FST to an output stream.
-  void Print(std::ostream *ostrm, const string &dest) {
-    ostrm_ = ostrm;
-    dest_ = dest;
-    const auto start = fst_.Start();
-    if (start == kNoStateId) return;
-    // Initial state first.
-    PrintState(start);
-    for (StateIterator<Fst<Arc>> siter(fst_); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      if (s != start) PrintState(s);
-    }
-  }
-
- private:
-  void PrintId(StateId id, const SymbolTable *syms, const char *name) const {
-    if (syms) {
-      string symbol = syms->Find(id);
-      if (symbol.empty()) {
-        if (missing_symbol_.empty()) {
-          FSTERROR() << "FstPrinter: Integer " << id
-                     << " is not mapped to any textual symbol"
-                     << ", symbol table = " << syms->Name()
-                     << ", destination = " << dest_;
-          symbol = "?";
-        } else {
-          symbol = missing_symbol_;
-        }
-      }
-      *ostrm_ << symbol;
-    } else {
-      *ostrm_ << id;
-    }
-  }
-
-  void PrintStateId(StateId s) const { PrintId(s, ssyms_, "state ID"); }
-
-  void PrintILabel(Label l) const { PrintId(l, isyms_, "arc input label"); }
-
-  void PrintOLabel(Label l) const { PrintId(l, osyms_, "arc output label"); }
-
-  void PrintState(StateId s) const {
-    bool output = false;
-    for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      PrintStateId(s);
-      *ostrm_ << sep_;
-      PrintStateId(arc.nextstate);
-      *ostrm_ << sep_;
-      PrintILabel(arc.ilabel);
-      if (!accep_) {
-        *ostrm_ << sep_;
-        PrintOLabel(arc.olabel);
-      }
-      if (show_weight_one_ || arc.weight != Weight::One())
-        *ostrm_ << sep_ << arc.weight;
-      *ostrm_ << "\n";
-      output = true;
-    }
-    const auto weight = fst_.Final(s);
-    if (weight != Weight::Zero() || !output) {
-      PrintStateId(s);
-      if (show_weight_one_ || weight != Weight::One()) {
-        *ostrm_ << sep_ << weight;
-      }
-      *ostrm_ << "\n";
-    }
-  }
-
-  const Fst<Arc> &fst_;
-  const SymbolTable *isyms_;  // ilabel symbol table.
-  const SymbolTable *osyms_;  // olabel symbol table.
-  const SymbolTable *ssyms_;  // slabel symbol table.
-  bool accep_;                // Print as acceptor when possible?
-  std::ostream *ostrm_;       // Text FST destination.
-  string dest_;               // Text FST destination name.
-  bool show_weight_one_;      // Print weights equal to Weight::One()?
-  string sep_;                // Separator character between fields.
-  string missing_symbol_;     // Symbol to print when lookup fails (default
-                              // "" means raise error).
-                              //
-  FstPrinter(const FstPrinter &) = delete;
-  FstPrinter &operator=(const FstPrinter &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_SCRIPT_PRINT_IMPL_H_
diff --git a/projects/llm_framework/include/fst/script/print.h b/projects/llm_framework/include/fst/script/print.h
deleted file mode 100644
index 687606b3..00000000
--- a/projects/llm_framework/include/fst/script/print.h
+++ /dev/null
@@ -1,79 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_PRINT_H_
-#define FST_SCRIPT_PRINT_H_
-
-#include <ostream>
-
-#include <fst/flags.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/print-impl.h>
-
-DECLARE_string(fst_field_separator);
-
-namespace fst {
-namespace script {
-
-// Note: it is safe to pass these strings as references because
-// this struct is only used to pass them deeper in the call graph.
-// Be sure you understand why this is so before using this struct
-// for anything else!
-struct FstPrinterArgs {
-  const FstClass &fst;
-  const SymbolTable *isyms;
-  const SymbolTable *osyms;
-  const SymbolTable *ssyms;
-  const bool accept;
-  const bool show_weight_one;
-  std::ostream *ostrm;
-  const string &dest;
-  const string &sep;  // NOLINT
-  const string &missing_symbol;
-
-  FstPrinterArgs(const FstClass &fst, const SymbolTable *isyms,
-                 const SymbolTable *osyms, const SymbolTable *ssyms,
-                 bool accept, bool show_weight_one, std::ostream *ostrm,
-                 const string &dest, const string &sep,
-                 const string &missing_sym = "")
-      : fst(fst),
-        isyms(isyms),
-        osyms(osyms),
-        ssyms(ssyms),
-        accept(accept),
-        show_weight_one(show_weight_one),
-        ostrm(ostrm),
-        dest(dest),
-        sep(sep),
-        missing_symbol(missing_sym) {}
-};
-
-template <class Arc>
-void PrintFst(FstPrinterArgs *args) {
-  const Fst<Arc> &fst = *(args->fst.GetFst<Arc>());
-  FstPrinter<Arc> fstprinter(fst, args->isyms, args->osyms, args->ssyms,
-                             args->accept, args->show_weight_one, args->sep,
-                             args->missing_symbol);
-  fstprinter.Print(args->ostrm, args->dest);
-}
-
-void PrintFst(const FstClass &fst, std::ostream &ostrm, const string &dest,
-              const SymbolTable *isyms, const SymbolTable *osyms,
-              const SymbolTable *ssyms, bool accept, bool show_weight_one,
-              const string &missing_sym = "");
-
-// The same, but with more sensible defaults.
-template <class Arc>
-void PrintFst(const Fst<Arc> &fst, std::ostream &ostrm, const string &dest = "",
-              const SymbolTable *isyms = nullptr,
-              const SymbolTable *osyms = nullptr,
-              const SymbolTable *ssyms = nullptr) {
-  const string sep = FLAGS_fst_field_separator.substr(0, 1);
-  FstPrinter<Arc> fstprinter(fst, isyms, osyms, ssyms, true, true, sep);
-  fstprinter.Print(&ostrm, dest);
-}
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_PRINT_H_
diff --git a/projects/llm_framework/include/fst/script/project.h b/projects/llm_framework/include/fst/script/project.h
deleted file mode 100644
index 13edeb1d..00000000
--- a/projects/llm_framework/include/fst/script/project.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_PROJECT_H_
-#define FST_SCRIPT_PROJECT_H_
-
-#include <utility>
-
-#include <fst/project.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using ProjectArgs = std::pair<MutableFstClass *, ProjectType>;
-
-template <class Arc>
-void Project(ProjectArgs *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  Project(fst, std::get<1>(*args));
-}
-
-void Project(MutableFstClass *fst, ProjectType project_type);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_PROJECT_H_
diff --git a/projects/llm_framework/include/fst/script/prune.h b/projects/llm_framework/include/fst/script/prune.h
deleted file mode 100644
index ed10b540..00000000
--- a/projects/llm_framework/include/fst/script/prune.h
+++ /dev/null
@@ -1,51 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_PRUNE_H_
-#define FST_SCRIPT_PRUNE_H_
-
-#include <tuple>
-#include <utility>
-
-#include <fst/prune.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-using PruneArgs1 = std::tuple<const FstClass &, MutableFstClass *,
-                              const WeightClass &, int64, float>;
-
-template <class Arc>
-void Prune(PruneArgs1 *args) {
-  using Weight = typename Arc::Weight;
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  const auto weight_threshold = *(std::get<2>(*args).GetWeight<Weight>());
-  Prune(ifst, ofst, weight_threshold, std::get<3>(*args), std::get<4>(*args));
-}
-
-using PruneArgs2 = std::tuple<MutableFstClass *, const WeightClass &, int64,
-                               float>;
-
-template <class Arc>
-void Prune(PruneArgs2 *args) {
-  using Weight = typename Arc::Weight;
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  const auto weight_threshold = *(std::get<1>(*args).GetWeight<Weight>());
-  Prune(fst, weight_threshold, std::get<2>(*args), std::get<3>(*args));
-}
-
-void Prune(const FstClass &ifst, MutableFstClass *ofst,
-           const WeightClass &weight_threshold,
-           int64 state_threshold = kNoStateId,
-           float delta = kDelta);
-
-void Prune(MutableFstClass *fst, const WeightClass &weight_threshold,
-           int64 state_threshold = kNoStateId, float delta = kDelta);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_PRUNE_H_
diff --git a/projects/llm_framework/include/fst/script/push.h b/projects/llm_framework/include/fst/script/push.h
deleted file mode 100644
index 018cd8f8..00000000
--- a/projects/llm_framework/include/fst/script/push.h
+++ /dev/null
@@ -1,53 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_PUSH_H_
-#define FST_SCRIPT_PUSH_H_
-
-#include <tuple>
-
-#include <fst/push.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using PushArgs1 = std::tuple<MutableFstClass *, ReweightType, float, bool>;
-
-template <class Arc>
-void Push(PushArgs1 *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  Push(fst, std::get<1>(*args), std::get<2>(*args), std::get<3>(*args));
-}
-
-using PushArgs2 = std::tuple<const FstClass &, MutableFstClass *, uint32,
-                             ReweightType, float>;
-
-template <class Arc>
-void Push(PushArgs2 *args) {
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  switch (std::get<3>(*args)) {
-    case REWEIGHT_TO_FINAL: {
-      Push<Arc, REWEIGHT_TO_FINAL>(ifst, ofst, std::get<2>(*args),
-                                   std::get<4>(*args));
-      return;
-    }
-    case REWEIGHT_TO_INITIAL: {
-      Push<Arc, REWEIGHT_TO_INITIAL>(ifst, ofst, std::get<2>(*args),
-                                     std::get<4>(*args));
-      return;
-    }
-  }
-}
-
-void Push(MutableFstClass *fst, ReweightType rew_type, float delta = kDelta,
-          bool remove_total_weight = false);
-
-void Push(const FstClass &ifst, MutableFstClass *ofst, uint32 flags,
-          ReweightType rew_type, float delta = kDelta);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_PUSH_H_
diff --git a/projects/llm_framework/include/fst/script/randequivalent.h b/projects/llm_framework/include/fst/script/randequivalent.h
deleted file mode 100644
index 945f8a06..00000000
--- a/projects/llm_framework/include/fst/script/randequivalent.h
+++ /dev/null
@@ -1,67 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_RANDEQUIVALENT_H_
-#define FST_SCRIPT_RANDEQUIVALENT_H_
-
-#include <climits>
-#include <ctime>
-
-#include <tuple>
-
-#include <fst/randequivalent.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/script-impl.h>
-
-namespace fst {
-namespace script {
-
-using RandEquivalentInnerArgs = std::tuple<const FstClass &, const FstClass &,
-    int32, float, time_t, const RandGenOptions<RandArcSelection> &>;
-
-using RandEquivalentArgs = WithReturnValue<bool, RandEquivalentInnerArgs>;
-
-template <class Arc>
-void RandEquivalent(RandEquivalentArgs *args) {
-  const Fst<Arc> &fst1 = *(std::get<0>(args->args).GetFst<Arc>());
-  const Fst<Arc> &fst2 = *(std::get<1>(args->args).GetFst<Arc>());
-  const auto seed = std::get<4>(args->args);
-  const auto &opts = std::get<5>(args->args);
-  switch (opts.selector) {
-    case UNIFORM_ARC_SELECTOR: {
-      const UniformArcSelector<Arc> selector(seed);
-      const RandGenOptions<UniformArcSelector<Arc>> ropts(selector,
-                                                          opts.max_length);
-      args->retval = RandEquivalent(fst1, fst2, std::get<2>(args->args),
-                                    std::get<3>(args->args), ropts);
-      return;
-    }
-    case FAST_LOG_PROB_ARC_SELECTOR: {
-      const FastLogProbArcSelector<Arc> selector(seed);
-      const RandGenOptions<FastLogProbArcSelector<Arc>> ropts(selector,
-                                                              opts.max_length);
-      args->retval = RandEquivalent(fst1, fst2, std::get<2>(args->args),
-                                    std::get<3>(args->args), ropts);
-      return;
-    }
-    case LOG_PROB_ARC_SELECTOR: {
-      const LogProbArcSelector<Arc> selector(seed);
-      const RandGenOptions<LogProbArcSelector<Arc>> ropts(selector,
-                                                          opts.max_length);
-      args->retval = RandEquivalent(fst1, fst2, std::get<2>(args->args),
-                                    std::get<3>(args->args), ropts);
-      return;
-    }
-  }
-}
-
-bool RandEquivalent(const FstClass &fst1, const FstClass &fst2, int32 npath = 1,
-    float delta = kDelta, time_t seed = time(nullptr),
-    const RandGenOptions<RandArcSelection> &opts =
-        RandGenOptions<RandArcSelection>(UNIFORM_ARC_SELECTOR));
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_RANDEQUIVALENT_H_
diff --git a/projects/llm_framework/include/fst/script/randgen.h b/projects/llm_framework/include/fst/script/randgen.h
deleted file mode 100644
index 5ce79d01..00000000
--- a/projects/llm_framework/include/fst/script/randgen.h
+++ /dev/null
@@ -1,63 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_RANDGEN_H_
-#define FST_SCRIPT_RANDGEN_H_
-
-#include <ctime>
-
-#include <tuple>
-
-#include <fst/randgen.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/script-impl.h>
-
-namespace fst {
-namespace script {
-
-using RandGenArgs = std::tuple<const FstClass &, MutableFstClass *, time_t,
-                               const RandGenOptions<RandArcSelection> &>;
-
-template <class Arc>
-void RandGen(RandGenArgs *args) {
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  const time_t seed = std::get<2>(*args);
-  const auto &opts = std::get<3>(*args);
-  switch (opts.selector) {
-    case UNIFORM_ARC_SELECTOR: {
-      const UniformArcSelector<Arc> selector(seed);
-      const RandGenOptions<UniformArcSelector<Arc>> ropts(
-          selector, opts.max_length, opts.npath, opts.weighted,
-          opts.remove_total_weight);
-      RandGen(ifst, ofst, ropts);
-      return;
-    }
-    case FAST_LOG_PROB_ARC_SELECTOR: {
-      const FastLogProbArcSelector<Arc> selector(seed);
-      const RandGenOptions<FastLogProbArcSelector<Arc>> ropts(
-          selector, opts.max_length, opts.npath, opts.weighted,
-          opts.remove_total_weight);
-      RandGen(ifst, ofst, ropts);
-      return;
-    }
-    case LOG_PROB_ARC_SELECTOR: {
-      const LogProbArcSelector<Arc> selector(seed);
-      const RandGenOptions<LogProbArcSelector<Arc>> ropts(
-          selector, opts.max_length, opts.npath, opts.weighted,
-          opts.remove_total_weight);
-      RandGen(ifst, ofst, ropts);
-      return;
-    }
-  }
-}
-
-void RandGen(const FstClass &ifst, MutableFstClass *ofst,
-             time_t seed = time(nullptr),
-             const RandGenOptions<RandArcSelection> &opts =
-                 RandGenOptions<RandArcSelection>(UNIFORM_ARC_SELECTOR));
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_RANDGEN_H_
diff --git a/projects/llm_framework/include/fst/script/register.h b/projects/llm_framework/include/fst/script/register.h
deleted file mode 100644
index d66e7ade..00000000
--- a/projects/llm_framework/include/fst/script/register.h
+++ /dev/null
@@ -1,99 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_REGISTER_H_
-#define FST_SCRIPT_REGISTER_H_
-
-#include <istream>
-#include <string>
-
-#include <fst/generic-register.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/weight-class.h>
-
-// Holds methods and classes responsible for maintaining
-// the register for FstClass arc types.
-
-namespace fst {
-namespace script {
-
-// Registers for reading and converting various kinds of FST classes.
-
-// This class definition is to avoid a nested class definition inside the
-// IORegistration struct.
-
-template <class Reader, class Creator, class Converter>
-struct FstClassRegEntry {
-  Reader reader;
-  Creator creator;
-  Converter converter;
-
-  FstClassRegEntry(Reader r, Creator cr, Converter co)
-      : reader(r), creator(cr), converter(co) {}
-
-  FstClassRegEntry()
-      : reader(nullptr), creator(nullptr), converter(nullptr) {}
-};
-
-template <class Reader, class Creator, class Converter>
-class FstClassIORegister
-    : public GenericRegister<string,
-                             FstClassRegEntry<Reader, Creator, Converter>,
-                             FstClassIORegister<Reader, Creator, Converter>> {
- public:
-  Reader GetReader(const string &arc_type) const {
-    return this->GetEntry(arc_type).reader;
-  }
-
-  Creator GetCreator(const string &arc_type) const {
-    return this->GetEntry(arc_type).creator;
-  }
-
-  Converter GetConverter(const string &arc_type) const {
-    return this->GetEntry(arc_type).converter;
-  }
-
- protected:
-  string ConvertKeyToSoFilename(const string &key) const final {
-    string legal_type(key);
-    ConvertToLegalCSymbol(&legal_type);
-    return legal_type + "-arc.so";
-  }
-};
-
-// Struct containing everything needed to register a particular type
-// of FST class (e.g., a plain FstClass, or a MutableFstClass, etc.).
-template <class FstClassType>
-struct IORegistration {
-  using Reader = FstClassType *(*)(std::istream &stream,
-                                   const FstReadOptions &opts);
-
-  using Creator = FstClassImplBase *(*)();
-
-  using Converter = FstClassImplBase *(*)(const FstClass &other);
-
-  using Entry = FstClassRegEntry<Reader, Creator, Converter>;
-
-  // FST class Register.
-  using Register = FstClassIORegister<Reader, Creator, Converter>;
-
-  // FST class Register-er.
-  using Registerer =
-      GenericRegisterer<FstClassIORegister<Reader, Creator, Converter>>;
-};
-
-#define REGISTER_FST_CLASS(Class, Arc)                                   \
-  static IORegistration<Class>::Registerer Class##_##Arc##_registerer(   \
-      Arc::Type(),                                                       \
-      IORegistration<Class>::Entry(Class::Read<Arc>, Class::Create<Arc>, \
-                                   Class::Convert<Arc>))
-
-#define REGISTER_FST_CLASSES(Arc)           \
-  REGISTER_FST_CLASS(FstClass, Arc);        \
-  REGISTER_FST_CLASS(MutableFstClass, Arc); \
-  REGISTER_FST_CLASS(VectorFstClass, Arc);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_REGISTER_H_
diff --git a/projects/llm_framework/include/fst/script/relabel.h b/projects/llm_framework/include/fst/script/relabel.h
deleted file mode 100644
index 74443490..00000000
--- a/projects/llm_framework/include/fst/script/relabel.h
+++ /dev/null
@@ -1,64 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_RELABEL_H_
-#define FST_SCRIPT_RELABEL_H_
-
-#include <algorithm>
-#include <tuple>
-#include <utility>
-#include <vector>
-
-#include <fst/relabel.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using RelabelArgs1 = std::tuple<MutableFstClass *, const SymbolTable *,
-                                const SymbolTable *, const string &, bool,
-                                const SymbolTable *, const SymbolTable *,
-                                const string &, bool>;
-
-template <class Arc>
-void Relabel(RelabelArgs1 *args) {
-  MutableFst<Arc> *ofst = std::get<0>(*args)->GetMutableFst<Arc>();
-  Relabel(ofst, std::get<1>(*args), std::get<2>(*args), std::get<3>(*args),
-          std::get<4>(*args), std::get<5>(*args), std::get<6>(*args),
-          std::get<7>(*args), std::get<8>(*args));
-}
-
-using LabelPair = std::pair<int64, int64>;
-
-using RelabelArgs2 = std::tuple<MutableFstClass *,
-                                const std::vector<LabelPair> &,
-                                const std::vector<LabelPair> &>;
-
-template <class Arc>
-void Relabel(RelabelArgs2 *args) {
-  MutableFst<Arc> *ofst = std::get<0>(*args)->GetMutableFst<Arc>();
-  using LabelPair = std::pair<typename Arc::Label, typename Arc::Label>;
-  // In case the MutableFstClass::Label is not the same as Arc::Label,
-  // make a copy.
-  std::vector<LabelPair> typed_ipairs(std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_ipairs.begin());
-  std::vector<LabelPair> typed_opairs(std::get<2>(*args).size());
-  std::copy(std::get<2>(*args).begin(), std::get<2>(*args).end(),
-            typed_opairs.begin());
-  Relabel(ofst, typed_ipairs, typed_opairs);
-}
-
-void Relabel(MutableFstClass *ofst,
-             const SymbolTable *old_isymbols, const SymbolTable *new_isymbols,
-             const string &unknown_isymbol,  bool attach_new_isymbols,
-             const SymbolTable *old_osymbols, const SymbolTable *new_osymbols,
-             const string &unknown_osymbol, bool attach_new_osymbols);
-
-void Relabel(MutableFstClass *ofst, const std::vector<LabelPair> &ipairs,
-             const std::vector<LabelPair> &opairs);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_RELABEL_H_
diff --git a/projects/llm_framework/include/fst/script/replace.h b/projects/llm_framework/include/fst/script/replace.h
deleted file mode 100644
index 926ece24..00000000
--- a/projects/llm_framework/include/fst/script/replace.h
+++ /dev/null
@@ -1,72 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_REPLACE_H_
-#define FST_SCRIPT_REPLACE_H_
-
-#include <tuple>
-#include <utility>
-#include <vector>
-
-#include <fst/replace.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-struct ReplaceOptions {
-  const int64 root;                          // Root rule for expansion.
-  const ReplaceLabelType call_label_type;    // How to label call arc.
-  const ReplaceLabelType return_label_type;  // How to label return arc.
-  const int64 return_label;                  // Specifies return arc label.
-
-  explicit ReplaceOptions(int64 root,
-      ReplaceLabelType call_label_type = REPLACE_LABEL_INPUT,
-      ReplaceLabelType return_label_type = REPLACE_LABEL_NEITHER,
-      int64 return_label = 0)
-      : root(root),
-        call_label_type(call_label_type),
-        return_label_type(return_label_type),
-        return_label(return_label) {}
-};
-
-using LabelFstClassPair = std::pair<int64, const FstClass *>;
-
-using ReplaceArgs = std::tuple<const std::vector<LabelFstClassPair> &,
-                               MutableFstClass *, const ReplaceOptions &>;
-
-template <class Arc>
-void Replace(ReplaceArgs *args) {
-  using LabelFstPair = std::pair<typename Arc::Label, const Fst<Arc> *>;
-  // Now that we know the arc type, we construct a vector of
-  // std::pair<real label, real fst> that the real Replace will use.
-  const auto &untyped_pairs = std::get<0>(*args);
-  std::vector<LabelFstPair> typed_pairs;
-  typed_pairs.reserve(untyped_pairs.size());
-  for (const auto &untyped_pair : untyped_pairs) {
-    typed_pairs.emplace_back(untyped_pair.first,  // Converts label.
-                             untyped_pair.second->GetFst<Arc>());
-  }
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  const auto &opts = std::get<2>(*args);
-  ReplaceFstOptions<Arc> typed_opts(opts.root, opts.call_label_type,
-                                    opts.return_label_type, opts.return_label);
-  ReplaceFst<Arc> rfst(typed_pairs, typed_opts);
-  // Checks for cyclic dependencies before attempting expansion.
-  if (rfst.CyclicDependencies()) {
-    FSTERROR() << "Replace: Cyclic dependencies detected; cannot expand";
-    ofst->SetProperties(kError, kError);
-    return;
-  }
-  typed_opts.gc = true;     // Caching options to speed up batch copy.
-  typed_opts.gc_limit = 0;
-  *ofst = rfst;
-}
-
-void Replace(const std::vector<LabelFstClassPair> &pairs,
-             MutableFstClass *ofst, const ReplaceOptions &opts);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_REPLACE_H_
diff --git a/projects/llm_framework/include/fst/script/reverse.h b/projects/llm_framework/include/fst/script/reverse.h
deleted file mode 100644
index badd96b5..00000000
--- a/projects/llm_framework/include/fst/script/reverse.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_REVERSE_H_
-#define FST_SCRIPT_REVERSE_H_
-
-#include <tuple>
-
-#include <fst/reverse.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using ReverseArgs = std::tuple<const FstClass &, MutableFstClass *, bool>;
-
-template <class Arc>
-void Reverse(ReverseArgs *args) {
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  Reverse(ifst, ofst, std::get<2>(*args));
-}
-
-void Reverse(const FstClass &ifst, MutableFstClass *ofst,
-             bool require_superinitial = true);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_REVERSE_H_
diff --git a/projects/llm_framework/include/fst/script/reweight.h b/projects/llm_framework/include/fst/script/reweight.h
deleted file mode 100644
index 3893ad85..00000000
--- a/projects/llm_framework/include/fst/script/reweight.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_REWEIGHT_H_
-#define FST_SCRIPT_REWEIGHT_H_
-
-#include <tuple>
-#include <vector>
-
-#include <fst/reweight.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/script-impl.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-using ReweightArgs = std::tuple<MutableFstClass *,
-                                const std::vector<WeightClass> &, ReweightType>;
-
-template <class Arc>
-void Reweight(ReweightArgs *args) {
-  using Weight = typename Arc::Weight;
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  const std::vector<WeightClass> &potentials = std::get<1>(*args);
-  std::vector<Weight> typed_potentials;
-  internal::CopyWeights(potentials, &typed_potentials);
-  Reweight(fst, typed_potentials, std::get<2>(*args));
-}
-
-void Reweight(MutableFstClass *fst, const std::vector<WeightClass> &potentials,
-              ReweightType reweight_type);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_REWEIGHT_H_
diff --git a/projects/llm_framework/include/fst/script/rmepsilon.h b/projects/llm_framework/include/fst/script/rmepsilon.h
deleted file mode 100644
index 42986c85..00000000
--- a/projects/llm_framework/include/fst/script/rmepsilon.h
+++ /dev/null
@@ -1,109 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_RMEPSILON_H_
-#define FST_SCRIPT_RMEPSILON_H_
-
-#include <utility>
-#include <vector>
-
-#include <fst/queue.h>
-#include <fst/rmepsilon.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/shortest-distance.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-struct RmEpsilonOptions : public ShortestDistanceOptions {
-  const bool connect;
-  const WeightClass &weight_threshold;
-  const int64 state_threshold;
-
-  RmEpsilonOptions(QueueType queue_type, bool connect,
-                   const WeightClass &weight_threshold,
-                   int64 state_threshold = kNoStateId, float delta = kDelta)
-      : ShortestDistanceOptions(queue_type, EPSILON_ARC_FILTER, kNoStateId,
-                                delta),
-        connect(connect),
-        weight_threshold(weight_threshold),
-        state_threshold(state_threshold) {}
-};
-
-namespace internal {
-
-// Code to implement switching on queue types.
-
-template <class Arc, class Queue>
-void RmEpsilon(MutableFst<Arc> *fst,
-               std::vector<typename Arc::Weight> *distance,
-               const RmEpsilonOptions &opts, Queue *queue) {
-  using Weight = typename Arc::Weight;
-  const fst::RmEpsilonOptions<Arc, Queue> ropts(
-      queue, opts.delta, opts.connect,
-      *opts.weight_threshold.GetWeight<Weight>(), opts.state_threshold);
-  RmEpsilon(fst, distance, ropts);
-}
-
-template <class Arc>
-void RmEpsilon(MutableFst<Arc> *fst, const RmEpsilonOptions &opts) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  std::vector<Weight> distance;
-  switch (opts.queue_type) {
-    case AUTO_QUEUE: {
-      AutoQueue<StateId> queue(*fst, &distance, EpsilonArcFilter<Arc>());
-      RmEpsilon(fst, &distance, opts, &queue);
-      return;
-    }
-    case FIFO_QUEUE: {
-      FifoQueue<StateId> queue;
-      RmEpsilon(fst, &distance, opts, &queue);
-      return;
-    }
-    case LIFO_QUEUE: {
-      LifoQueue<StateId> queue;
-      RmEpsilon(fst, &distance, opts, &queue);
-      return;
-    }
-    case SHORTEST_FIRST_QUEUE: {
-      NaturalShortestFirstQueue<StateId, Weight> queue(distance);
-      RmEpsilon(fst, &distance, opts, &queue);
-      return;
-    }
-    case STATE_ORDER_QUEUE: {
-      StateOrderQueue<StateId> queue;
-      RmEpsilon(fst, &distance, opts, &queue);
-      return;
-    }
-    case TOP_ORDER_QUEUE: {
-      TopOrderQueue<StateId> queue(*fst, EpsilonArcFilter<Arc>());
-      internal::RmEpsilon(fst, &distance, opts, &queue);
-      return;
-    }
-    default: {
-      FSTERROR() << "RmEpsilon: Unknown queue type: " << opts.queue_type;
-      fst->SetProperties(kError, kError);
-      return;
-    }
-  }
-}
-
-}  // namespace internal
-
-using RmEpsilonArgs = std::pair<MutableFstClass *, const RmEpsilonOptions &>;
-
-template <class Arc>
-void RmEpsilon(RmEpsilonArgs *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  const auto &opts = std::get<1>(*args);
-  internal::RmEpsilon(fst, opts);
-}
-
-void RmEpsilon(MutableFstClass *fst, const RmEpsilonOptions &opts);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_RMEPSILON_H_
diff --git a/projects/llm_framework/include/fst/script/script-impl.h b/projects/llm_framework/include/fst/script/script-impl.h
deleted file mode 100644
index 33c2853a..00000000
--- a/projects/llm_framework/include/fst/script/script-impl.h
+++ /dev/null
@@ -1,211 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// This file defines the registration mechanism for new operations.
-// These operations are designed to enable scripts to work with FST classes
-// at a high level.
-//
-// If you have a new arc type and want these operations to work with FSTs
-// with that arc type, see below for the registration steps
-// you must take.
-//
-// These methods are only recommended for use in high-level scripting
-// applications. Most users should use the lower-level templated versions
-// corresponding to these.
-//
-// If you have a new arc type you'd like these operations to work with,
-// use the REGISTER_FST_OPERATIONS macro defined in fstscript.h.
-//
-// If you have a custom operation you'd like to define, you need four
-// components. In the following, assume you want to create a new operation
-// with the signature
-//
-//    void Foo(const FstClass &ifst, MutableFstClass *ofst);
-//
-//  You need:
-//
-//  1) A way to bundle the args that your new Foo operation will take, as
-//     a single struct. The template structs in arg-packs.h provide a handy
-//     way to do this. In Foo's case, that might look like this:
-//
-//       using FooArgs = std::pair<const FstClass &, MutableFstClass *>;
-//
-//     Note: this package of args is going to be passed by non-const pointer.
-//
-//  2) A function template that is able to perform Foo, given the args and
-//     arc type. Yours might look like this:
-//
-//       template<class Arc>
-//       void Foo(FooArgs *args) {
-//          // Pulls out the actual, arc-templated FSTs.
-//          const Fst<Arc> &ifst = std::get<0>(*args).GetFst<Arc>();
-//          MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-//          // Actually perform Foo on ifst and ofst.
-//       }
-//
-//  3) a client-facing function for your operation. This would look like
-//     the following:
-//
-//     void Foo(const FstClass &ifst, MutableFstClass *ofst) {
-//       // Check that the arc types of the FSTs match
-//       if (!ArcTypesMatch(ifst, *ofst, "Foo")) return;
-//       // package the args
-//       FooArgs args(ifst, ofst);
-//       // Finally, call the operation
-//       Apply<Operation<FooArgs>>("Foo", ifst->ArcType(), &args);
-//     }
-//
-//  The Apply<> function template takes care of the link between 2 and 3,
-//  provided you also have:
-//
-//  4) A registration for your new operation, on the arc types you care about.
-//     This can be provided easily by the REGISTER_FST_OPERATION macro in
-//     operations.h:
-//
-//       REGISTER_FST_OPERATION(Foo, StdArc, FooArgs);
-//       REGISTER_FST_OPERATION(Foo, MyArc, FooArgs);
-//       // .. etc
-//
-//
-//  That's it! Now when you call Foo(const FstClass &, MutableFstClass *),
-//  it dispatches (in #3) via the Apply<> function to the correct
-//  instantiation of the template function in #2.
-//
-
-#ifndef FST_SCRIPT_SCRIPT_IMPL_H_
-#define FST_SCRIPT_SCRIPT_IMPL_H_
-
-// This file contains general-purpose templates which are used in the
-// implementation of the operations.
-
-#include <string>
-#include <utility>
-
-#include <fst/generic-register.h>
-#include <fst/script/fst-class.h>
-
-#include <fst/log.h>
-
-namespace fst {
-namespace script {
-
-enum RandArcSelection {
-  UNIFORM_ARC_SELECTOR,
-  LOG_PROB_ARC_SELECTOR,
-  FAST_LOG_PROB_ARC_SELECTOR
-};
-
-// A generic register for operations with various kinds of signatures.
-// Needed since every function signature requires a new registration class.
-// The std::pair<string, string> is understood to be the operation name and arc
-// type; subclasses (or typedefs) need only provide the operation signature.
-
-template <class OperationSignature>
-class GenericOperationRegister
-    : public GenericRegister<std::pair<string, string>, OperationSignature,
-                             GenericOperationRegister<OperationSignature>> {
- public:
-  void RegisterOperation(const string &operation_name, const string &arc_type,
-                         OperationSignature op) {
-    this->SetEntry(std::make_pair(operation_name, arc_type), op);
-  }
-
-  OperationSignature GetOperation(const string &operation_name,
-                                  const string &arc_type) {
-    return this->GetEntry(std::make_pair(operation_name, arc_type));
-  }
-
- protected:
-  string ConvertKeyToSoFilename(
-      const std::pair<string, string> &key) const final {
-    // Uses the old-style FST for now.
-    string legal_type(key.second);  // The arc type.
-    ConvertToLegalCSymbol(&legal_type);
-    return legal_type + "-arc.so";
-  }
-};
-
-// Operation package: everything you need to register a new type of operation.
-// The ArgPack should be the type that's passed into each wrapped function;
-// for instance, it might be a struct containing all the args. It's always
-// passed by pointer, so const members should be used to enforce constness where
-// it's needed. Return values should be implemented as a member of ArgPack as
-// well.
-
-template <class Args>
-struct Operation {
-  using ArgPack = Args;
-
-  using OpType = void (*)(ArgPack *args);
-
-  // The register (hash) type.
-  using Register = GenericOperationRegister<OpType>;
-
-  // The register-er type
-  using Registerer = GenericRegisterer<Register>;
-};
-
-// Macro for registering new types of operations.
-
-#define REGISTER_FST_OPERATION(Op, Arc, ArgPack)               \
-  static fst::script::Operation<ArgPack>::Registerer       \
-      arc_dispatched_operation_##ArgPack##Op##Arc##_registerer \
-      (std::make_pair(#Op, Arc::Type()), Op<Arc>)
-
-// Template function to apply an operation by name.
-
-template <class OpReg>
-void Apply(const string &op_name, const string &arc_type,
-           typename OpReg::ArgPack *args) {
-  const auto op = OpReg::Register::GetRegister()->GetOperation(op_name,
-                                                               arc_type);
-  if (!op) {
-    FSTERROR() << "No operation found for " << op_name << " on "
-               << "arc type " << arc_type;
-    return;
-  }
-  op(args);
-}
-
-namespace internal {
-
-// Helper that logs to ERROR if the arc types of m and n don't match,
-// assuming that both m and n implement .ArcType(). The op_name argument is
-// used to construct the error message.
-template <class M, class N>
-bool ArcTypesMatch(const M &m, const N &n, const string &op_name) {
-  if (m.ArcType() != n.ArcType()) {
-    FSTERROR() << "Arguments with non-matching arc types passed to "
-               << op_name << ":\t" << m.ArcType() << " and " << n.ArcType();
-    return false;
-  }
-  return true;
-}
-
-// From untyped to typed weights.
-template <class Weight>
-void CopyWeights(const std::vector<WeightClass> &weights,
-                 std::vector<Weight> *typed_weights) {
-  typed_weights->clear();
-  typed_weights->reserve(weights.size());
-  for (const auto &weight : weights) {
-    typed_weights->push_back(*weight.GetWeight<Weight>());
-  }
-}
-
-// From typed to untyped weights.
-template <class Weight>
-void CopyWeights(const std::vector<Weight> &typed_weights,
-                 std::vector<WeightClass> *weights) {
-  weights->clear();
-  weights->reserve(typed_weights.size());
-  for (const auto &typed_weight : typed_weights) {
-    weights->emplace_back(typed_weight);
-  }
-}
-
-}  // namespace internal
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_SCRIPT_IMPL_H_
diff --git a/projects/llm_framework/include/fst/script/shortest-distance.h b/projects/llm_framework/include/fst/script/shortest-distance.h
deleted file mode 100644
index a44a6c9b..00000000
--- a/projects/llm_framework/include/fst/script/shortest-distance.h
+++ /dev/null
@@ -1,214 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_SHORTEST_DISTANCE_H_
-#define FST_SCRIPT_SHORTEST_DISTANCE_H_
-
-#include <tuple>
-#include <vector>
-
-#include <fst/queue.h>
-#include <fst/shortest-distance.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/prune.h>
-#include <fst/script/script-impl.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-enum ArcFilterType {
-  ANY_ARC_FILTER,
-  EPSILON_ARC_FILTER,
-  INPUT_EPSILON_ARC_FILTER,
-  OUTPUT_EPSILON_ARC_FILTER
-};
-
-struct ShortestDistanceOptions {
-  const QueueType queue_type;
-  const ArcFilterType arc_filter_type;
-  const int64 source;
-  const float delta;
-
-  ShortestDistanceOptions(QueueType queue_type, ArcFilterType arc_filter_type,
-                          int64 source, float delta)
-      : queue_type(queue_type),
-        arc_filter_type(arc_filter_type),
-        source(source),
-        delta(delta) {}
-};
-
-namespace internal {
-
-// Code to implement switching on queue and arc filter types.
-
-template <class Arc, class Queue, class ArcFilter>
-struct QueueConstructor {
-  using Weight = typename Arc::Weight;
-
-  static Queue *Construct(const Fst<Arc> &, const std::vector<Weight> *) {
-    return new Queue();
-  }
-};
-
-// Specializations to support queues with different constructors.
-
-template <class Arc, class ArcFilter>
-struct QueueConstructor<Arc, AutoQueue<typename Arc::StateId>, ArcFilter> {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  //  template<class Arc, class ArcFilter>
-  static AutoQueue<StateId> *Construct(const Fst<Arc> &fst,
-                                       const std::vector<Weight> *distance) {
-    return new AutoQueue<StateId>(fst, distance, ArcFilter());
-  }
-};
-
-template <class Arc, class ArcFilter>
-struct QueueConstructor<
-    Arc, NaturalShortestFirstQueue<typename Arc::StateId, typename Arc::Weight>,
-    ArcFilter> {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  static NaturalShortestFirstQueue<StateId, Weight> *Construct(
-      const Fst<Arc> &, const std::vector<Weight> *distance) {
-    return new NaturalShortestFirstQueue<StateId, Weight>(*distance);
-  }
-};
-
-template <class Arc, class ArcFilter>
-struct QueueConstructor<Arc, TopOrderQueue<typename Arc::StateId>, ArcFilter> {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  static TopOrderQueue<StateId> *Construct(const Fst<Arc> &fst,
-                                           const std::vector<Weight> *) {
-    return new TopOrderQueue<StateId>(fst, ArcFilter());
-  }
-};
-
-template <class Arc, class Queue, class ArcFilter>
-void ShortestDistance(const Fst<Arc> &fst,
-                      std::vector<typename Arc::Weight> *distance,
-                      const ShortestDistanceOptions &opts) {
-  std::unique_ptr<Queue> queue(
-      QueueConstructor<Arc, Queue, ArcFilter>::Construct(fst, distance));
-  const fst::ShortestDistanceOptions<Arc, Queue, ArcFilter> sopts(
-      queue.get(), ArcFilter(), opts.source, opts.delta);
-  ShortestDistance(fst, distance, sopts);
-}
-
-template <class Arc, class Queue>
-void ShortestDistance(const Fst<Arc> &fst,
-                      std::vector<typename Arc::Weight> *distance,
-                      const ShortestDistanceOptions &opts) {
-  switch (opts.arc_filter_type) {
-    case ANY_ARC_FILTER: {
-      ShortestDistance<Arc, Queue, AnyArcFilter<Arc>>(fst, distance, opts);
-      return;
-    }
-    case EPSILON_ARC_FILTER: {
-      ShortestDistance<Arc, Queue, EpsilonArcFilter<Arc>>(fst, distance, opts);
-      return;
-    }
-    case INPUT_EPSILON_ARC_FILTER: {
-      ShortestDistance<Arc, Queue, InputEpsilonArcFilter<Arc>>(fst, distance,
-                                                               opts);
-      return;
-    }
-    case OUTPUT_EPSILON_ARC_FILTER: {
-      ShortestDistance<Arc, Queue, OutputEpsilonArcFilter<Arc>>(fst, distance,
-                                                                opts);
-      return;
-    }
-    default: {
-      FSTERROR() << "ShortestDistance: Unknown arc filter type: "
-                 << opts.arc_filter_type;
-      distance->clear();
-      distance->resize(1, Arc::Weight::NoWeight());
-      return;
-    }
-  }
-}
-
-}  // namespace internal
-
-using ShortestDistanceArgs1 =
-    std::tuple<const FstClass &, std::vector<WeightClass> *,
-               const ShortestDistanceOptions &>;
-
-template <class Arc>
-void ShortestDistance(ShortestDistanceArgs1 *args) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  const auto &opts = std::get<2>(*args);
-  std::vector<Weight> typed_distance;
-  switch (opts.queue_type) {
-    case AUTO_QUEUE: {
-      internal::ShortestDistance<Arc, AutoQueue<StateId>>(fst, &typed_distance,
-                                                          opts);
-      break;
-    }
-    case FIFO_QUEUE: {
-      internal::ShortestDistance<Arc, FifoQueue<StateId>>(fst, &typed_distance,
-                                                          opts);
-      break;
-    }
-    case LIFO_QUEUE: {
-      internal::ShortestDistance<Arc, LifoQueue<StateId>>(fst, &typed_distance,
-                                                          opts);
-      break;
-    }
-    case SHORTEST_FIRST_QUEUE: {
-      internal::ShortestDistance<Arc,
-                                 NaturalShortestFirstQueue<StateId, Weight>>(
-          fst, &typed_distance, opts);
-      break;
-    }
-    case STATE_ORDER_QUEUE: {
-      internal::ShortestDistance<Arc, StateOrderQueue<StateId>>(
-          fst, &typed_distance, opts);
-      break;
-    }
-    case TOP_ORDER_QUEUE: {
-      internal::ShortestDistance<Arc, TopOrderQueue<StateId>>(
-          fst, &typed_distance, opts);
-      break;
-    }
-    default: {
-      FSTERROR() << "ShortestDistance: Unknown queue type: " << opts.queue_type;
-      typed_distance.clear();
-      typed_distance.resize(1, Arc::Weight::NoWeight());
-      break;
-    }
-  }
-  internal::CopyWeights(typed_distance, std::get<1>(*args));
-}
-
-using ShortestDistanceArgs2 =
-    std::tuple<const FstClass &, std::vector<WeightClass> *, bool, double>;
-
-template <class Arc>
-void ShortestDistance(ShortestDistanceArgs2 *args) {
-  using Weight = typename Arc::Weight;
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  std::vector<Weight> typed_distance;
-  ShortestDistance(fst, &typed_distance, std::get<2>(*args),
-                   std::get<3>(*args));
-  internal::CopyWeights(typed_distance, std::get<1>(*args));
-}
-
-void ShortestDistance(const FstClass &fst, std::vector<WeightClass> *distance,
-                      const ShortestDistanceOptions &opts);
-
-void ShortestDistance(const FstClass &ifst, std::vector<WeightClass> *distance,
-                      bool reverse = false,
-                      double delta = fst::kShortestDelta);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_SHORTEST_DISTANCE_H_
diff --git a/projects/llm_framework/include/fst/script/shortest-path.h b/projects/llm_framework/include/fst/script/shortest-path.h
deleted file mode 100644
index 86bc88da..00000000
--- a/projects/llm_framework/include/fst/script/shortest-path.h
+++ /dev/null
@@ -1,116 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_SHORTEST_PATH_H_
-#define FST_SCRIPT_SHORTEST_PATH_H_
-
-#include <memory>
-#include <vector>
-
-#include <fst/shortest-path.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/shortest-distance.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-// Slightly simplified interface: `has_distance` and `first_path` are disabled.
-
-struct ShortestPathOptions : public ShortestDistanceOptions {
-  const int32 nshortest;
-  const bool unique;
-  const WeightClass &weight_threshold;
-  const int64 state_threshold;
-
-  ShortestPathOptions(QueueType queue_type, int32 nshortest, bool unique,
-                      float delta, const WeightClass &weight_threshold,
-                      int64 state_threshold = kNoStateId)
-      : ShortestDistanceOptions(queue_type, ANY_ARC_FILTER, kNoStateId, delta),
-        nshortest(nshortest),
-        unique(unique),
-        weight_threshold(weight_threshold),
-        state_threshold(state_threshold) {}
-};
-
-namespace internal {
-
-// Code to implement switching on queue types.
-
-template <class Arc, class Queue>
-void ShortestPath(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-                  std::vector<typename Arc::Weight> *distance,
-                  const ShortestPathOptions &opts) {
-  using ArcFilter = AnyArcFilter<Arc>;
-  using Weight = typename Arc::Weight;
-  const std::unique_ptr<Queue> queue(
-      QueueConstructor<Arc, Queue, ArcFilter>::Construct(ifst, distance));
-  const fst::ShortestPathOptions<Arc, Queue, ArcFilter> sopts(
-      queue.get(), ArcFilter(), opts.nshortest, opts.unique,
-      /* has_distance=*/false, opts.delta, /* first_path=*/false,
-      *opts.weight_threshold.GetWeight<Weight>(), opts.state_threshold);
-  ShortestPath(ifst, ofst, distance, sopts);
-}
-
-template <class Arc>
-void ShortestPath(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-                  const ShortestPathOptions &opts) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  std::vector<Weight> distance;
-  switch (opts.queue_type) {
-    case AUTO_QUEUE: {
-      ShortestPath<Arc, AutoQueue<StateId>>(ifst, ofst, &distance, opts);
-      return;
-    }
-    case FIFO_QUEUE: {
-      ShortestPath<Arc, FifoQueue<StateId>>(ifst, ofst, &distance, opts);
-      return;
-    }
-    case LIFO_QUEUE: {
-      ShortestPath<Arc, LifoQueue<StateId>>(ifst, ofst, &distance, opts);
-      return;
-    }
-    case SHORTEST_FIRST_QUEUE: {
-      ShortestPath<Arc, NaturalShortestFirstQueue<StateId, Weight>>(ifst, ofst,
-                                                                    &distance,
-                                                                    opts);
-      return;
-    }
-    case STATE_ORDER_QUEUE: {
-      ShortestPath<Arc, StateOrderQueue<StateId>>(ifst, ofst, &distance, opts);
-      return;
-    }
-    case TOP_ORDER_QUEUE: {
-      ShortestPath<Arc, TopOrderQueue<StateId>>(ifst, ofst, &distance, opts);
-      return;
-    }
-    default: {
-      FSTERROR() << "ShortestPath: Unknown queue type: "
-                 << opts.queue_type;
-      ofst->SetProperties(kError, kError);
-      return;
-    }
-  }
-}
-
-}  // namespace internal
-
-using ShortestPathArgs = std::tuple<const FstClass &, MutableFstClass *,
-                                    const ShortestPathOptions &>;
-
-template <class Arc>
-void ShortestPath(ShortestPathArgs *args) {
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  const ShortestPathOptions &opts = std::get<2>(*args);
-  internal::ShortestPath(ifst, ofst, opts);
-}
-
-void ShortestPath(const FstClass &ifst, MutableFstClass *ofst,
-                  const ShortestPathOptions &opts);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_SHORTEST_PATH_H_
diff --git a/projects/llm_framework/include/fst/script/stateiterator-class.h b/projects/llm_framework/include/fst/script/stateiterator-class.h
deleted file mode 100644
index f6fddfe6..00000000
--- a/projects/llm_framework/include/fst/script/stateiterator-class.h
+++ /dev/null
@@ -1,85 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_STATEITERATOR_CLASS_H_
-#define FST_SCRIPT_STATEITERATOR_CLASS_H_
-
-#include <memory>
-
-#include <fst/fstlib.h>
-#include <fst/script/fst-class.h>
-
-// Scripting API support for StateIterator.
-
-namespace fst {
-namespace script {
-
-// Virtual interface implemented by each concrete StateIteratorImpl<F>.
-class StateIteratorImplBase {
- public:
-  virtual bool Done() const = 0;
-  virtual int64 Value() const = 0;
-  virtual void Next() = 0;
-  virtual void Reset() = 0;
-  virtual ~StateIteratorImplBase() {}
-};
-
-// Templated implementation.
-template <class Arc>
-class StateIteratorClassImpl : public StateIteratorImplBase {
- public:
-  explicit StateIteratorClassImpl(const Fst<Arc> &fst) : siter_(fst) {}
-
-  bool Done() const final { return siter_.Done(); }
-
-  int64 Value() const final { return siter_.Value(); }
-
-  void Next() final { siter_.Next(); }
-
-  void Reset() final { siter_.Reset(); }
-
-  ~StateIteratorClassImpl() override {}
-
- private:
-  StateIterator<Fst<Arc>> siter_;
-};
-
-class StateIteratorClass;
-
-using InitStateIteratorClassArgs =
-    std::pair<const FstClass &, StateIteratorClass *>;
-
-// Untemplated user-facing class holding a templated pimpl.
-class StateIteratorClass {
- public:
-  explicit StateIteratorClass(const FstClass &fst);
-
-  template <class Arc>
-  explicit StateIteratorClass(const Fst<Arc> &fst)
-      : impl_(new StateIteratorClassImpl<Arc>(fst)) {}
-
-  bool Done() const { return impl_->Done(); }
-
-  int64 Value() const { return impl_->Value(); }
-
-  void Next() { impl_->Next(); }
-
-  void Reset() { impl_->Reset(); }
-
-  template <class Arc>
-  friend void InitStateIteratorClass(InitStateIteratorClassArgs *args);
-
- private:
-  std::unique_ptr<StateIteratorImplBase> impl_;
-};
-
-template <class Arc>
-void InitStateIteratorClass(InitStateIteratorClassArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  std::get<1>(*args)->impl_.reset(new StateIteratorClassImpl<Arc>(fst));
-}
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_STATEITERATOR_CLASS_H_
diff --git a/projects/llm_framework/include/fst/script/synchronize.h b/projects/llm_framework/include/fst/script/synchronize.h
deleted file mode 100644
index 01df151a..00000000
--- a/projects/llm_framework/include/fst/script/synchronize.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_SYNCHRONIZE_H_
-#define FST_SCRIPT_SYNCHRONIZE_H_
-
-#include <utility>
-
-#include <fst/synchronize.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using SynchronizeArgs = std::pair<const FstClass &, MutableFstClass *>;
-
-template <class Arc>
-void Synchronize(SynchronizeArgs *args) {
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  Synchronize(ifst, ofst);
-}
-
-void Synchronize(const FstClass &ifst, MutableFstClass *ofst);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_SYNCHRONIZE_H_
diff --git a/projects/llm_framework/include/fst/script/text-io.h b/projects/llm_framework/include/fst/script/text-io.h
deleted file mode 100644
index 464bf885..00000000
--- a/projects/llm_framework/include/fst/script/text-io.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Utilities for reading and writing textual strings representing states,
-// labels, and weights and files specifying label-label pairs and potentials
-// (state-weight pairs).
-
-#ifndef FST_SCRIPT_TEXT_IO_H__
-#define FST_SCRIPT_TEXT_IO_H__
-
-#include <string>
-#include <vector>
-
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-bool ReadPotentials(const string &weight_type, const string &filename,
-                    std::vector<WeightClass> *potentials);
-
-bool WritePotentials(const string &filename,
-                     const std::vector<WeightClass> &potentials);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_TEXT_IO_H__
diff --git a/projects/llm_framework/include/fst/script/topsort.h b/projects/llm_framework/include/fst/script/topsort.h
deleted file mode 100644
index fb6738d7..00000000
--- a/projects/llm_framework/include/fst/script/topsort.h
+++ /dev/null
@@ -1,26 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_TOPSORT_H_
-#define FST_SCRIPT_TOPSORT_H_
-
-#include <fst/topsort.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using TopSortArgs = WithReturnValue<bool, MutableFstClass *>;
-
-template <class Arc>
-void TopSort(TopSortArgs *args) {
-  args->retval = TopSort(args->args->GetMutableFst<Arc>());
-}
-
-bool TopSort(MutableFstClass *fst);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_TOPSORT_H_
diff --git a/projects/llm_framework/include/fst/script/union.h b/projects/llm_framework/include/fst/script/union.h
deleted file mode 100644
index 9493e2b1..00000000
--- a/projects/llm_framework/include/fst/script/union.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_UNION_H_
-#define FST_SCRIPT_UNION_H_
-
-#include <utility>
-
-#include <fst/union.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using UnionArgs = std::pair<MutableFstClass *, const FstClass &>;
-
-template <class Arc>
-void Union(UnionArgs *args) {
-  MutableFst<Arc> *fst1 = std::get<0>(*args)->GetMutableFst<Arc>();
-  const Fst<Arc> &fst2 = *(std::get<1>(*args).GetFst<Arc>());
-  Union(fst1, fst2);
-}
-
-void Union(MutableFstClass *fst1, const FstClass &fst2);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_UNION_H_
diff --git a/projects/llm_framework/include/fst/script/verify.h b/projects/llm_framework/include/fst/script/verify.h
deleted file mode 100644
index 52f58641..00000000
--- a/projects/llm_framework/include/fst/script/verify.h
+++ /dev/null
@@ -1,27 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_VERIFY_H_
-#define FST_SCRIPT_VERIFY_H_
-
-#include <fst/verify.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using VerifyArgs = WithReturnValue<bool, const FstClass &>;
-
-template <class Arc>
-void Verify(VerifyArgs *args) {
-  const Fst<Arc> &fst = *(args->args.GetFst<Arc>());
-  args->retval = Verify(fst);
-}
-
-bool Verify(const FstClass &fst);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_VERIFY_H_
diff --git a/projects/llm_framework/include/fst/script/weight-class.h b/projects/llm_framework/include/fst/script/weight-class.h
deleted file mode 100644
index 6dadf92c..00000000
--- a/projects/llm_framework/include/fst/script/weight-class.h
+++ /dev/null
@@ -1,235 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Represents a generic weight in an FST; that is, represents a specific type
-// of weight underneath while hiding that type from a client.
-
-#ifndef FST_SCRIPT_WEIGHT_CLASS_H_
-#define FST_SCRIPT_WEIGHT_CLASS_H_
-
-#include <memory>
-#include <ostream>
-#include <string>
-
-#include <fst/arc.h>
-#include <fst/generic-register.h>
-#include <fst/util.h>
-#include <fst/weight.h>
-
-namespace fst {
-namespace script {
-
-class WeightImplBase {
- public:
-  virtual WeightImplBase *Copy() const = 0;
-  virtual void Print(std::ostream *o) const = 0;
-  virtual const string &Type() const = 0;
-  virtual string ToString() const = 0;
-  virtual bool Member() const = 0;
-  virtual bool operator==(const WeightImplBase &other) const = 0;
-  virtual bool operator!=(const WeightImplBase &other) const = 0;
-  virtual WeightImplBase &PlusEq(const WeightImplBase &other) = 0;
-  virtual WeightImplBase &TimesEq(const WeightImplBase &other) = 0;
-  virtual WeightImplBase &DivideEq(const WeightImplBase &other) = 0;
-  virtual WeightImplBase &PowerEq(size_t n) = 0;
-  virtual ~WeightImplBase() {}
-};
-
-template <class W>
-class WeightClassImpl : public WeightImplBase {
- public:
-  explicit WeightClassImpl(const W &weight) : weight_(weight) {}
-
-  WeightClassImpl<W> *Copy() const final {
-    return new WeightClassImpl<W>(weight_);
-  }
-
-  const string &Type() const final { return W::Type(); }
-
-  void Print(std::ostream *ostrm) const final { *ostrm << weight_; }
-
-  string ToString() const final {
-    string str;
-    WeightToStr(weight_, &str);
-    return str;
-  }
-
-  bool Member() const final { return weight_.Member(); }
-
-  bool operator==(const WeightImplBase &other) const final {
-    const auto *typed_other = static_cast<const WeightClassImpl<W> *>(&other);
-    return weight_ == typed_other->weight_;
-  }
-
-  bool operator!=(const WeightImplBase &other) const final {
-    return !(*this == other);
-  }
-
-  WeightClassImpl<W> &PlusEq(const WeightImplBase &other) final {
-    const auto *typed_other = static_cast<const WeightClassImpl<W> *>(&other);
-    weight_ = Plus(weight_, typed_other->weight_);
-    return *this;
-  }
-
-  WeightClassImpl<W> &TimesEq(const WeightImplBase &other) final {
-    const auto *typed_other = static_cast<const WeightClassImpl<W> *>(&other);
-    weight_ = Times(weight_, typed_other->weight_);
-    return *this;
-  }
-
-  WeightClassImpl<W> &DivideEq(const WeightImplBase &other) final {
-    const auto *typed_other = static_cast<const WeightClassImpl<W> *>(&other);
-    weight_ = Divide(weight_, typed_other->weight_);
-    return *this;
-  }
-
-  WeightClassImpl<W> &PowerEq(size_t n) final {
-    weight_ = Power(weight_, n);
-    return *this;
-  }
-
-  W *GetImpl() { return &weight_; }
-
- private:
-  W weight_;
-};
-
-
-class WeightClass {
- public:
-  WeightClass() = default;
-
-  template <class W>
-  explicit WeightClass(const W &weight)
-      : impl_(new WeightClassImpl<W>(weight)) {}
-
-  template <class W>
-  explicit WeightClass(const WeightClassImpl<W> &impl)
-      : impl_(new WeightClassImpl<W>(impl)) {}
-
-  WeightClass(const string &weight_type, const string &weight_str);
-
-  WeightClass(const WeightClass &other)
-      : impl_(other.impl_ ? other.impl_->Copy() : nullptr) {}
-
-  WeightClass &operator=(const WeightClass &other) {
-    impl_.reset(other.impl_ ? other.impl_->Copy() : nullptr);
-    return *this;
-  }
-
-  static constexpr const char *__ZERO__ = "__ZERO__";  // NOLINT
-
-  static WeightClass Zero(const string &weight_type);
-
-  static constexpr const char *__ONE__ = "__ONE__";  // NOLINT
-
-  static WeightClass One(const string &weight_type);
-
-  static constexpr const char *__NOWEIGHT__ = "__NOWEIGHT__";  // NOLINT
-
-  static WeightClass NoWeight(const string &weight_type);
-
-  template <class W>
-  const W *GetWeight() const {
-    if (W::Type() != impl_->Type()) {
-       return nullptr;
-    } else {
-      auto *typed_impl = static_cast<WeightClassImpl<W> *>(impl_.get());
-      return typed_impl->GetImpl();
-    }
-  }
-
-  string ToString() const { return (impl_) ? impl_->ToString() : "none"; }
-
-  const string &Type() const {
-    if (impl_) return impl_->Type();
-    static const string *const no_type = new string("none");
-    return *no_type;
-  }
-
-  bool Member() const { return impl_ && impl_->Member(); }
-
-  bool WeightTypesMatch(const WeightClass &other, const string &op_name) const;
-
-  friend bool operator==(const WeightClass &lhs, const WeightClass &rhs);
-
-  friend WeightClass Plus(const WeightClass &lhs, const WeightClass &rhs);
-
-  friend WeightClass Times(const WeightClass &lhs, const WeightClass &rhs);
-
-  friend WeightClass Divide(const WeightClass &lhs, const WeightClass &rhs);
-
-  friend WeightClass Power(const WeightClass &w, size_t n);
-
- private:
-  const WeightImplBase *GetImpl() const { return impl_.get(); }
-
-  WeightImplBase *GetImpl() { return impl_.get(); }
-
-  std::unique_ptr<WeightImplBase> impl_;
-
-  friend std::ostream &operator<<(std::ostream &o, const WeightClass &c);
-};
-
-bool operator==(const WeightClass &lhs, const WeightClass &rhs);
-
-bool operator!=(const WeightClass &lhs, const WeightClass &rhs);
-
-WeightClass Plus(const WeightClass &lhs, const WeightClass &rhs);
-
-WeightClass Times(const WeightClass &lhs, const WeightClass &rhs);
-
-WeightClass Divide(const WeightClass &lhs, const WeightClass &rhs);
-
-WeightClass Power(const WeightClass &w, size_t n);
-
-std::ostream &operator<<(std::ostream &o, const WeightClass &c);
-
-// Registration for generic weight types.
-
-using StrToWeightImplBaseT = WeightImplBase *(*)(const string &str,
-                                                 const string &src,
-                                                 size_t nline);
-
-template <class W>
-WeightImplBase *StrToWeightImplBase(const string &str, const string &src,
-                                    size_t nline) {
-  if (str == WeightClass::__ZERO__)
-    return new WeightClassImpl<W>(W::Zero());
-  else if (str == WeightClass::__ONE__)
-    return new WeightClassImpl<W>(W::One());
-  else if (str == WeightClass::__NOWEIGHT__)
-    return new WeightClassImpl<W>(W::NoWeight());
-  return new WeightClassImpl<W>(StrToWeight<W>(str, src, nline));
-}
-
-class WeightClassRegister : public GenericRegister<string, StrToWeightImplBaseT,
-                                                   WeightClassRegister> {
- protected:
-  string ConvertKeyToSoFilename(const string &key) const final {
-    string legal_type(key);
-    ConvertToLegalCSymbol(&legal_type);
-    return legal_type + ".so";
-  }
-};
-
-using WeightClassRegisterer = GenericRegisterer<WeightClassRegister>;
-
-// Internal version; needs to be called by wrapper in order for macro args to
-// expand.
-#define REGISTER_FST_WEIGHT__(Weight, line)                \
-  static WeightClassRegisterer weight_registerer##_##line( \
-      Weight::Type(), StrToWeightImplBase<Weight>)
-
-// This layer is where __FILE__ and __LINE__ are expanded.
-#define REGISTER_FST_WEIGHT_EXPANDER(Weight, line) \
-  REGISTER_FST_WEIGHT__(Weight, line)
-
-// Macro for registering new weight types; clients call this.
-#define REGISTER_FST_WEIGHT(Weight) \
-  REGISTER_FST_WEIGHT_EXPANDER(Weight, __LINE__)
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_WEIGHT_CLASS_H_
diff --git a/projects/llm_framework/include/fst/set-weight.h b/projects/llm_framework/include/fst/set-weight.h
deleted file mode 100644
index dd665f0d..00000000
--- a/projects/llm_framework/include/fst/set-weight.h
+++ /dev/null
@@ -1,618 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Weights consisting of sets (of integral Labels) and
-// associated semiring operation definitions using intersect
-// and union.
-
-#ifndef FST_SET_WEIGHT_H_
-#define FST_SET_WEIGHT_H_
-
-#include <cstdlib>
-
-#include <algorithm>
-#include <list>
-#include <string>
-#include <vector>
-
-#include <fst/union-weight.h>
-#include <fst/weight.h>
-
-
-namespace fst {
-
-constexpr int kSetEmpty = 0;         // Label for the empty set.
-constexpr int kSetUniv = -1;         // Label for the universal set.
-constexpr int kSetBad = -2;          // Label for a non-set.
-constexpr char kSetSeparator = '_';  // Label separator in sets.
-
-// Determines whether to use (intersect, union) or (union, intersect)
-// as (+, *) for the semiring. SET_INTERSECT_UNION_RESTRICTED is a
-// restricted version of (intersect, union) that requires summed
-// arguments to be equal (or an error is signalled), useful for
-// algorithms that require a unique labelled path weight.  SET_BOOLEAN
-// treats all non-Zero() elements as equivalent (with Zero() ==
-// UnivSet()), useful for algorithms that don't really depend on the
-// detailed sets.
-enum SetType { SET_INTERSECT_UNION = 0,
-               SET_UNION_INTERSECT = 1,
-               SET_INTERSECT_UNION_RESTRICT = 2,
-               SET_BOOLEAN = 3 };
-
-template <class>
-class SetWeightIterator;
-
-// Set semiring of integral labels.
-template <typename Label_, SetType S = SET_INTERSECT_UNION>
-class SetWeight {
- public:
-  using Label = Label_;
-  using ReverseWeight = SetWeight<Label, S>;
-  using Iterator = SetWeightIterator<SetWeight>;
-  friend class SetWeightIterator<SetWeight>;
-  // Allow type-converting copy and move constructors private access.
-  template <typename L2, SetType S2>
-  friend class SetWeight;
-
-  SetWeight() {}
-
-  // Input should be positive, sorted and unique.
-  template <typename Iterator>
-  SetWeight(const Iterator &begin, const Iterator &end) {
-    for (auto iter = begin; iter != end; ++iter) PushBack(*iter);
-  }
-
-  // Input should be positive. (Non-positive value has
-  // special internal meaning w.r.t. integral constants above.)
-  explicit SetWeight(Label label) { PushBack(label); }
-
-  template <SetType S2>
-  explicit SetWeight(const SetWeight<Label, S2> &w)
-    : first_(w.first_), rest_(w.rest_) {}
-
-  template <SetType S2>
-  explicit SetWeight(SetWeight<Label, S2> &&w)
-    : first_(w.first_), rest_(std::move(w.rest_)) { w.Clear(); }
-
-  template <SetType S2>
-  SetWeight &operator=(const SetWeight<Label, S2> &w) {
-    first_ = w.first_;
-    rest_ = w.rest_;
-    return *this;
-  }
-
-  template <SetType S2>
-  SetWeight &operator=(SetWeight<Label, S2> &&w) {
-    first_ = w.first_;
-    rest_ = std::move(w.rest_);
-    w.Clear();
-    return *this;
-  }
-
-  static const SetWeight &Zero() {
-    return S == SET_UNION_INTERSECT ? EmptySet() : UnivSet();
-  }
-
-  static const SetWeight &One() {
-    return S == SET_UNION_INTERSECT ? UnivSet() : EmptySet();
-  }
-
-  static const SetWeight &NoWeight() {
-    static const auto *const no_weight = new SetWeight(Label(kSetBad));
-    return *no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string(
-        S == SET_UNION_INTERSECT
-        ? "union_intersect_set"
-        : (S == SET_INTERSECT_UNION
-           ? "intersect_union_set"
-           : (S == SET_INTERSECT_UNION_RESTRICT
-              ? "restricted_set_intersect_union"
-              : "boolean_set")));
-    return *type;
-  }
-
-  bool Member() const;
-
-  std::istream &Read(std::istream &strm);
-
-  std::ostream &Write(std::ostream &strm) const;
-
-  size_t Hash() const;
-
-  SetWeight Quantize(float delta = kDelta) const { return *this; }
-
-  ReverseWeight Reverse() const;
-
-  static constexpr uint64 Properties() {
-    return kIdempotent | kLeftSemiring | kRightSemiring | kCommutative;
-  }
-
-  // These operations combined with the SetWeightIterator
-  // provide the access and mutation of the set internal elements.
-
-  // The empty set.
-  static const SetWeight &EmptySet() {
-    static const auto *const empty = new SetWeight(Label(kSetEmpty));
-    return *empty;
-  }
-
-  // The univeral set.
-  static const SetWeight &UnivSet() {
-    static const auto *const univ = new SetWeight(Label(kSetUniv));
-    return *univ;
-  }
-
-  // Clear existing SetWeight.
-  void Clear() {
-    first_ = kSetEmpty;
-    rest_.clear();
-  }
-
-  size_t Size() const { return first_ == kSetEmpty ? 0 : rest_.size() + 1; }
-
-  Label Back() {
-    if (rest_.empty()) {
-      return first_;
-    } else {
-      return rest_.back();
-    }
-  }
-
-  // Caller must add in sort order and be unique (or error signalled).
-  // Input should also be positive. Non-positive value (for the first
-  // push) has special internal meaning w.r.t. integral constants above.
-  void PushBack(Label label) {
-    if (first_ == kSetEmpty) {
-      first_ = label;
-    } else {
-      if (label <= Back() || label <= 0) {
-        FSTERROR() << "SetWeight: labels must be positive, added"
-                   << " in sort order and be unique.";
-        rest_.push_back(Label(kSetBad));
-      }
-      rest_.push_back(label);
-    }
-  }
-
- private:
-  Label first_ = kSetEmpty;  // First label in set (kSetEmpty if empty).
-  std::list<Label> rest_;    // Remaining labels in set.
-};
-
-// Traverses set in forward direction.
-template <class SetWeight_>
-class SetWeightIterator {
- public:
-  using Weight = SetWeight_;
-  using Label = typename Weight::Label;
-
-  explicit SetWeightIterator(const Weight &w)
-      : first_(w.first_), rest_(w.rest_), init_(true), iter_(rest_.begin()) {}
-
-  bool Done() const {
-    if (init_) {
-      return first_ == kSetEmpty;
-    } else {
-      return iter_ == rest_.end();
-    }
-  }
-
-  const Label &Value() const { return init_ ? first_ : *iter_; }
-
-  void Next() {
-    if (init_) {
-      init_ = false;
-    } else {
-      ++iter_;
-    }
-  }
-
-  void Reset() {
-    init_ = true;
-    iter_ = rest_.begin();
-  }
-
- private:
-  const Label &first_;
-  const decltype(Weight::rest_) &rest_;
-  bool init_;  // In the initialized state?
-  typename std::remove_reference<decltype(Weight::rest_)>::type::const_iterator iter_;
-};
-
-
-// SetWeight member functions follow that require SetWeightIterator
-
-template <typename Label, SetType S>
-inline std::istream &SetWeight<Label, S>::Read(std::istream &strm) {
-  Clear();
-  int32 size;
-  ReadType(strm, &size);
-  for (int32 i = 0; i < size; ++i) {
-    Label label;
-    ReadType(strm, &label);
-    PushBack(label);
-  }
-  return strm;
-}
-
-template <typename Label, SetType S>
-inline std::ostream &SetWeight<Label, S>::Write(std::ostream &strm) const {
-  const int32 size = Size();
-  WriteType(strm, size);
-  for (Iterator iter(*this); !iter.Done(); iter.Next()) {
-    WriteType(strm, iter.Value());
-  }
-  return strm;
-}
-
-template <typename Label, SetType S>
-inline bool SetWeight<Label, S>::Member() const {
-  Iterator iter(*this);
-  return iter.Value() != Label(kSetBad);
-}
-
-template <typename Label, SetType S>
-inline typename SetWeight<Label, S>::ReverseWeight
-SetWeight<Label, S>::Reverse() const {
-  return *this;
-}
-
-template <typename Label, SetType S>
-inline size_t SetWeight<Label, S>::Hash() const {
-  using Weight = SetWeight<Label, S>;
-  if (S == SET_BOOLEAN) {
-    return *this == Weight::Zero() ? 0 : 1;
-  } else {
-    size_t h = 0;
-    for (Iterator iter(*this); !iter.Done(); iter.Next()) {
-      h ^= h << 1 ^ iter.Value();
-    }
-    return h;
-  }
-}
-
-// Default ==
-template <typename Label, SetType S>
-inline bool operator==(const SetWeight<Label, S> &w1,
-                       const SetWeight<Label, S> &w2) {
-  if (w1.Size() != w2.Size()) return false;
-  using Iterator = typename SetWeight<Label, S>::Iterator;
-  Iterator iter1(w1);
-  Iterator iter2(w2);
-  for (; !iter1.Done(); iter1.Next(), iter2.Next()) {
-    if (iter1.Value() != iter2.Value()) return false;
-  }
-  return true;
-}
-
-// Boolean ==
-template <typename Label>
-inline bool operator==(const SetWeight<Label, SET_BOOLEAN> &w1,
-                       const SetWeight<Label, SET_BOOLEAN> &w2) {
-  // x == kSetEmpty if x \nin {kUnivSet, kSetBad}
-  if (!w1.Member() || !w2.Member()) return false;
-  using Iterator = typename SetWeight<Label, SET_BOOLEAN>::Iterator;
-  Iterator iter1(w1);
-  Iterator iter2(w2);
-  Label label1 = iter1.Done() ? kSetEmpty : iter1.Value();
-  Label label2 = iter2.Done() ? kSetEmpty : iter2.Value();
-  if (label1 == kSetUniv) return label2 == kSetUniv;
-  if (label2 == kSetUniv) return label1 == kSetUniv;
-  return true;
-}
-
-template <typename Label, SetType S>
-inline bool operator!=(const SetWeight<Label, S> &w1,
-                       const SetWeight<Label, S> &w2) {
-  return !(w1 == w2);
-}
-
-template <typename Label, SetType S>
-inline bool ApproxEqual(const SetWeight<Label, S> &w1,
-                        const SetWeight<Label, S> &w2,
-                        float delta = kDelta) {
-  return w1 == w2;
-}
-
-template <typename Label, SetType S>
-inline std::ostream &operator<<(std::ostream &strm,
-                                const SetWeight<Label, S> &weight) {
-  typename SetWeight<Label, S>::Iterator iter(weight);
-  if (iter.Done()) {
-    return strm << "EmptySet";
-  } else if (iter.Value() == Label(kSetUniv)) {
-    return strm << "UnivSet";
-  } else if (iter.Value() == Label(kSetBad)) {
-    return strm << "BadSet";
-  } else {
-    for (size_t i = 0; !iter.Done(); ++i, iter.Next()) {
-      if (i > 0) strm << kSetSeparator;
-      strm << iter.Value();
-    }
-  }
-  return strm;
-}
-
-template <typename Label, SetType S>
-inline std::istream &operator>>(std::istream &strm,
-                                SetWeight<Label, S> &weight) {
-  string str;
-  strm >> str;
-  using Weight = SetWeight<Label, S>;
-  if (str == "EmptySet") {
-    weight = Weight(Label(kSetEmpty));
-  } else if (str == "UnivSet") {
-    weight = Weight(Label(kSetUniv));
-  } else {
-    weight.Clear();
-    char *p = nullptr;
-    for (const char *cs = str.c_str(); !p || *p != '\0'; cs = p + 1) {
-      const Label label = strtoll(cs, &p, 10);
-      if (p == cs || (*p != 0 && *p != kSetSeparator)) {
-        strm.clear(std::ios::badbit);
-        break;
-      }
-      weight.PushBack(label);
-    }
-  }
-  return strm;
-}
-
-template <typename Label, SetType S>
-inline SetWeight<Label, S> Union(
-    const SetWeight<Label, S> &w1,
-    const SetWeight<Label, S> &w2) {
-  using Weight = SetWeight<Label, S>;
-  using Iterator = typename SetWeight<Label, S>::Iterator;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::EmptySet()) return w2;
-  if (w2 == Weight::EmptySet()) return w1;
-  if (w1 == Weight::UnivSet()) return w1;
-  if (w2 == Weight::UnivSet()) return w2;
-  Iterator it1(w1);
-  Iterator it2(w2);
-  Weight result;
-  while (!it1.Done() && !it2.Done()) {
-    const auto v1 = it1.Value();
-    const auto v2 = it2.Value();
-    if (v1 < v2) {
-      result.PushBack(v1);
-      it1.Next();
-    } else if (v1 > v2) {
-      result.PushBack(v2);
-      it2.Next();
-    } else {
-      result.PushBack(v1);
-      it1.Next();
-      it2.Next();
-    }
-  }
-  for (; !it1.Done(); it1.Next()) result.PushBack(it1.Value());
-  for (; !it2.Done(); it2.Next()) result.PushBack(it2.Value());
-  return result;
-}
-
-template <typename Label, SetType S>
-inline SetWeight<Label, S> Intersect(
-    const SetWeight<Label, S> &w1,
-    const SetWeight<Label, S> &w2) {
-  using Weight = SetWeight<Label, S>;
-  using Iterator = typename SetWeight<Label, S>::Iterator;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::EmptySet()) return w1;
-  if (w2 == Weight::EmptySet()) return w2;
-  if (w1 == Weight::UnivSet()) return w2;
-  if (w2 == Weight::UnivSet()) return w1;
-  Iterator it1(w1);
-  Iterator it2(w2);
-  Weight result;
-  while (!it1.Done() && !it2.Done()) {
-    const auto v1 = it1.Value();
-    const auto v2 = it2.Value();
-    if (v1 < v2) {
-      it1.Next();
-    } else if (v1 > v2) {
-      it2.Next();
-    } else {
-      result.PushBack(v1);
-      it1.Next();
-      it2.Next();
-    }
-  }
-  return result;
-}
-
-template <typename Label, SetType S>
-inline SetWeight<Label, S> Difference(
-    const SetWeight<Label, S> &w1,
-    const SetWeight<Label, S> &w2) {
-  using Weight = SetWeight<Label, S>;
-  using Iterator = typename SetWeight<Label, S>::Iterator;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::EmptySet()) return w1;
-  if (w2 == Weight::EmptySet()) return w1;
-  if (w2 == Weight::UnivSet()) return Weight::EmptySet();
-  Iterator it1(w1);
-  Iterator it2(w2);
-  Weight result;
-  while (!it1.Done() && !it2.Done()) {
-    const auto v1 = it1.Value();
-    const auto v2 = it2.Value();
-    if (v1 < v2) {
-      result.PushBack(v1);
-      it1.Next();
-    } else if (v1 > v2) {
-      it2.Next();
-    } else {
-      it1.Next();
-      it2.Next();
-    }
-  }
-  for (; !it1.Done(); it1.Next()) result.PushBack(it1.Value());
-  return result;
-}
-
-// Default: Plus = Intersect.
-template <typename Label, SetType S>
-inline SetWeight<Label, S> Plus(
-    const SetWeight<Label, S> &w1,
-    const SetWeight<Label, S> &w2) {
-  return Intersect(w1, w2);
-}
-
-// Plus = Union.
-template <typename Label>
-inline SetWeight<Label, SET_UNION_INTERSECT> Plus(
-    const SetWeight<Label, SET_UNION_INTERSECT> &w1,
-    const SetWeight<Label, SET_UNION_INTERSECT> &w2) {
-  return Union(w1, w2);
-}
-
-// Plus = Set equality is required (for non-Zero() input). The
-// restriction is useful (e.g., in determinization) to ensure the input
-// has a unique labelled path weight.
-template <typename Label>
-inline SetWeight<Label, SET_INTERSECT_UNION_RESTRICT> Plus(
-    const SetWeight<Label, SET_INTERSECT_UNION_RESTRICT> &w1,
-    const SetWeight<Label, SET_INTERSECT_UNION_RESTRICT> &w2) {
-  using Weight = SetWeight<Label, SET_INTERSECT_UNION_RESTRICT>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::Zero()) return w2;
-  if (w2 == Weight::Zero()) return w1;
-  if (w1 != w2) {
-    FSTERROR() << "SetWeight::Plus: Unequal arguments "
-               << "(non-unique labelled path weights?)"
-               << " w1 = " << w1 << " w2 = " << w2;
-    return Weight::NoWeight();
-  }
-  return w1;
-}
-
-// Plus = Or.
-template <typename Label>
-inline SetWeight<Label, SET_BOOLEAN> Plus(
-    const SetWeight<Label, SET_BOOLEAN> &w1,
-    const SetWeight<Label, SET_BOOLEAN> &w2) {
-  using Weight = SetWeight<Label, SET_BOOLEAN>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::One()) return w1;
-  if (w2 == Weight::One()) return w2;
-  return Weight::Zero();
-}
-
-// Default: Times = Union.
-template <typename Label, SetType S>
-inline SetWeight<Label, S> Times(
-    const SetWeight<Label, S> &w1,
-    const SetWeight<Label, S> &w2) {
-  return Union(w1, w2);
-}
-
-// Times = Intersect.
-template <typename Label>
-inline SetWeight<Label, SET_UNION_INTERSECT> Times(
-    const SetWeight<Label, SET_UNION_INTERSECT> &w1,
-    const SetWeight<Label, SET_UNION_INTERSECT> &w2) {
-  return Intersect(w1, w2);
-}
-
-// Times = And.
-template <typename Label>
-inline SetWeight<Label, SET_BOOLEAN> Times(
-    const SetWeight<Label, SET_BOOLEAN> &w1,
-    const SetWeight<Label, SET_BOOLEAN> &w2) {
-  using Weight = SetWeight<Label, SET_BOOLEAN>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::One()) return w2;
-  return w1;
-}
-
-// Divide = Difference.
-template <typename Label, SetType S>
-inline SetWeight<Label, S> Divide(const SetWeight<Label, S> &w1,
-                                  const SetWeight<Label, S> &w2,
-                                  DivideType divide_type = DIVIDE_ANY) {
-  return Difference(w1, w2);
-}
-
-// Divide = dividend (or the universal set if the
-// dividend == divisor).
-template <typename Label>
-inline SetWeight<Label, SET_UNION_INTERSECT> Divide(
-    const SetWeight<Label, SET_UNION_INTERSECT> &w1,
-    const SetWeight<Label, SET_UNION_INTERSECT> &w2,
-    DivideType divide_type = DIVIDE_ANY) {
-  using Weight = SetWeight<Label, SET_UNION_INTERSECT>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == w2) return Weight::UnivSet();
-  return w1;
-}
-
-// Divide = Or Not.
-template <typename Label>
-inline SetWeight<Label, SET_BOOLEAN> Divide(
-    const SetWeight<Label, SET_BOOLEAN> &w1,
-    const SetWeight<Label, SET_BOOLEAN> &w2,
-    DivideType divide_type = DIVIDE_ANY) {
-  using Weight = SetWeight<Label, SET_BOOLEAN>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::One()) return w1;
-  if (w2 == Weight::Zero()) return Weight::One();
-  return Weight::Zero();
-}
-
-// Converts between different set types.
-template <typename Label, SetType S1, SetType S2>
-struct WeightConvert<SetWeight<Label, S1>, SetWeight<Label, S2>> {
-  SetWeight<Label, S2> operator()(const SetWeight<Label, S1> &w1) const {
-    using Iterator = SetWeightIterator<SetWeight<Label, S1>>;
-    SetWeight<Label, S2> w2;
-    for (Iterator iter(w1); !iter.Done(); iter.Next())
-      w2.PushBack(iter.Value());
-    return w2;
-  }
-};
-
-// This function object generates SetWeights that are random integer sets
-// from {1, ... , alphabet_size}^{0, max_set_length} U { Zero }. This is
-// intended primarily for testing.
-template <class Label, SetType S>
-class WeightGenerate<SetWeight<Label, S>> {
- public:
-  using Weight = SetWeight<Label, S>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t alphabet_size = kNumRandomWeights,
-                          size_t max_set_length = kNumRandomWeights)
-      : allow_zero_(allow_zero),
-        alphabet_size_(alphabet_size),
-        max_set_length_(max_set_length) {}
-
-  Weight operator()() const {
-    const size_t n = rand() % (max_set_length_ + allow_zero_);  // NOLINT
-    if (allow_zero_ && n == max_set_length_) return Weight::Zero();
-    std::vector<Label> labels;
-    for (size_t i = 0; i < n; ++i) {
-      labels.push_back(rand() % alphabet_size_ + 1);  // NOLINT
-    }
-    std::sort(labels.begin(), labels.end());
-    const auto labels_end = std::unique(labels.begin(), labels.end());
-    labels.resize(labels_end - labels.begin());
-    return Weight(labels.begin(), labels.end());
-  }
-
- private:
-  // Permits Zero() and zero divisors.
-  const bool allow_zero_;
-  // Alphabet size for random weights.
-  const size_t alphabet_size_;
-  // Number of alternative random weights.
-  const size_t max_set_length_;
-};
-
-}  // namespace fst
-
-#endif  // FST_SET_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/shortest-distance.h b/projects/llm_framework/include/fst/shortest-distance.h
deleted file mode 100644
index b1c5b26c..00000000
--- a/projects/llm_framework/include/fst/shortest-distance.h
+++ /dev/null
@@ -1,351 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to find shortest distance in an FST.
-
-#ifndef FST_SHORTEST_DISTANCE_H_
-#define FST_SHORTEST_DISTANCE_H_
-
-#include <cstddef>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arcfilter.h>
-#include <fst/cache.h>
-#include <fst/queue.h>
-#include <fst/reverse.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-// A representable float for shortest distance and shortest path algorithms.
-constexpr float kShortestDelta = 1e-6;
-
-template <class Arc, class Queue, class ArcFilter>
-struct ShortestDistanceOptions {
-  using StateId = typename Arc::StateId;
-
-  Queue *state_queue;    // Queue discipline used; owned by caller.
-  ArcFilter arc_filter;  // Arc filter (e.g., limit to only epsilon graph).
-  StateId source;        // If kNoStateId, use the FST's initial state.
-  float delta;           // Determines the degree of convergence required
-  bool first_path;       // For a semiring with the path property (o.w.
-                         // undefined), compute the shortest-distances along
-                         // along the first path to a final state found
-                         // by the algorithm. That path is the shortest-path
-                         // only if the FST has a unique final state (or all
-                         // the final states have the same final weight), the
-                         // queue discipline is shortest-first and all the
-                         // weights in the FST are between One() and Zero()
-                         // according to NaturalLess.
-
-  ShortestDistanceOptions(Queue *state_queue, ArcFilter arc_filter,
-                          StateId source = kNoStateId,
-                          float delta = kShortestDelta,
-                          bool first_path = false)
-      : state_queue(state_queue),
-        arc_filter(arc_filter),
-        source(source),
-        delta(delta),
-        first_path(first_path) {}
-};
-
-namespace internal {
-
-// Computation state of the shortest-distance algorithm. Reusable information
-// is maintained across calls to member function ShortestDistance(source) when
-// retain is true for improved efficiency when calling multiple times from
-// different source states (e.g., in epsilon removal). Contrary to the usual
-// conventions, fst may not be freed before this class. Vector distance
-// should not be modified by the user between these calls. The Error() method
-// returns true iff an error was encountered.
-template <class Arc, class Queue, class ArcFilter>
-class ShortestDistanceState {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  ShortestDistanceState(
-      const Fst<Arc> &fst, std::vector<Weight> *distance,
-      const ShortestDistanceOptions<Arc, Queue, ArcFilter> &opts, bool retain)
-      : fst_(fst),
-        distance_(distance),
-        state_queue_(opts.state_queue),
-        arc_filter_(opts.arc_filter),
-        delta_(opts.delta),
-        first_path_(opts.first_path),
-        retain_(retain),
-        source_id_(0),
-        error_(false) {
-    distance_->clear();
-    if (fst.Properties(kExpanded, false) == kExpanded) {
-      const auto num_states = CountStates(fst);
-      distance_->reserve(num_states);
-      adder_.reserve(num_states);
-      radder_.reserve(num_states);
-      enqueued_.reserve(num_states);
-    }
-  }
-
-  void ShortestDistance(StateId source);
-
-  bool Error() const { return error_; }
-
- private:
-  void EnsureDistanceIndexIsValid(std::size_t index) {
-    while (distance_->size() <= index) {
-      distance_->push_back(Weight::Zero());
-      adder_.push_back(Adder<Weight>());
-      radder_.push_back(Adder<Weight>());
-      enqueued_.push_back(false);
-    }
-    DCHECK_LT(index, distance_->size());
-  }
-
-  void EnsureSourcesIndexIsValid(std::size_t index) {
-    while (sources_.size() <= index) {
-      sources_.push_back(kNoStateId);
-    }
-    DCHECK_LT(index, sources_.size());
-  }
-
-  const Fst<Arc> &fst_;
-  std::vector<Weight> *distance_;
-  Queue *state_queue_;
-  ArcFilter arc_filter_;
-  const float delta_;
-  const bool first_path_;
-  const bool retain_;  // Retain and reuse information across calls.
-
-  std::vector<Adder<Weight>> adder_;   // Sums distance_ accurately.
-  std::vector<Adder<Weight>> radder_;  // Relaxation distance.
-  std::vector<bool> enqueued_;         // Is state enqueued?
-  std::vector<StateId> sources_;       // Source ID for ith state in distance_,
-                                       // (r)adder_, and enqueued_ if retained.
-  StateId source_id_;                  // Unique ID characterizing each call.
-  bool error_;
-};
-
-// Compute the shortest distance; if source is kNoStateId, uses the initial
-// state of the FST.
-template <class Arc, class Queue, class ArcFilter>
-void ShortestDistanceState<Arc, Queue, ArcFilter>::ShortestDistance(
-    StateId source) {
-  if (fst_.Start() == kNoStateId) {
-    if (fst_.Properties(kError, false)) error_ = true;
-    return;
-  }
-  if (!(Weight::Properties() & kRightSemiring)) {
-    FSTERROR() << "ShortestDistance: Weight needs to be right distributive: "
-               << Weight::Type();
-    error_ = true;
-    return;
-  }
-  if (first_path_ && !(Weight::Properties() & kPath)) {
-    FSTERROR() << "ShortestDistance: The first_path option is disallowed when "
-               << "Weight does not have the path property: " << Weight::Type();
-    error_ = true;
-    return;
-  }
-  state_queue_->Clear();
-  if (!retain_) {
-    distance_->clear();
-    adder_.clear();
-    radder_.clear();
-    enqueued_.clear();
-  }
-  if (source == kNoStateId) source = fst_.Start();
-  EnsureDistanceIndexIsValid(source);
-  if (retain_) {
-    EnsureSourcesIndexIsValid(source);
-    sources_[source] = source_id_;
-  }
-  (*distance_)[source] = Weight::One();
-  adder_[source].Reset(Weight::One());
-  radder_[source].Reset(Weight::One());
-  enqueued_[source] = true;
-  state_queue_->Enqueue(source);
-  while (!state_queue_->Empty()) {
-    const auto state = state_queue_->Head();
-    state_queue_->Dequeue();
-    EnsureDistanceIndexIsValid(state);
-    if (first_path_ && (fst_.Final(state) != Weight::Zero())) break;
-    enqueued_[state] = false;
-    const auto r = radder_[state].Sum();
-    radder_[state].Reset();
-    for (ArcIterator<Fst<Arc>> aiter(fst_, state); !aiter.Done();
-         aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto nextstate = arc.nextstate;
-      if (!arc_filter_(arc)) continue;
-      EnsureDistanceIndexIsValid(nextstate);
-      if (retain_) {
-        EnsureSourcesIndexIsValid(nextstate);
-        if (sources_[nextstate] != source_id_) {
-          (*distance_)[nextstate] = Weight::Zero();
-          adder_[nextstate].Reset();
-          radder_[nextstate].Reset();
-          enqueued_[nextstate] = false;
-          sources_[nextstate] = source_id_;
-        }
-      }
-      auto &nd = (*distance_)[nextstate];
-      auto &na = adder_[nextstate];
-      auto &nr = radder_[nextstate];
-      auto weight = Times(r, arc.weight);
-      if (!ApproxEqual(nd, Plus(nd, weight), delta_)) {
-        nd = na.Add(weight);
-        nr.Add(weight);
-        if (!nd.Member() || !nr.Sum().Member()) {
-          error_ = true;
-          return;
-        }
-        if (!enqueued_[nextstate]) {
-          state_queue_->Enqueue(nextstate);
-          enqueued_[nextstate] = true;
-        } else {
-          state_queue_->Update(nextstate);
-        }
-      }
-    }
-  }
-  ++source_id_;
-  if (fst_.Properties(kError, false)) error_ = true;
-}
-
-}  // namespace internal
-
-// Shortest-distance algorithm: this version allows fine control
-// via the options argument. See below for a simpler interface.
-//
-// This computes the shortest distance from the opts.source state to each
-// visited state S and stores the value in the distance vector. An
-// unvisited state S has distance Zero(), which will be stored in the
-// distance vector if S is less than the maximum visited state. The state
-// queue discipline, arc filter, and convergence delta are taken in the
-// options argument. The distance vector will contain a unique element for
-// which Member() is false if an error was encountered.
-//
-// The weights must must be right distributive and k-closed (i.e., 1 +
-// x + x^2 + ... + x^(k +1) = 1 + x + x^2 + ... + x^k).
-//
-// Complexity:
-//
-// Depends on properties of the semiring and the queue discipline.
-//
-// For more information, see:
-//
-// Mohri, M. 2002. Semiring framework and algorithms for shortest-distance
-// problems, Journal of Automata, Languages and
-// Combinatorics 7(3): 321-350, 2002.
-template <class Arc, class Queue, class ArcFilter>
-void ShortestDistance(
-    const Fst<Arc> &fst, std::vector<typename Arc::Weight> *distance,
-    const ShortestDistanceOptions<Arc, Queue, ArcFilter> &opts) {
-  internal::ShortestDistanceState<Arc, Queue, ArcFilter> sd_state(fst, distance,
-                                                                  opts, false);
-  sd_state.ShortestDistance(opts.source);
-  if (sd_state.Error()) {
-    distance->assign(1, Arc::Weight::NoWeight());
-  }
-}
-
-// Shortest-distance algorithm: simplified interface. See above for a version
-// that permits finer control.
-//
-// If reverse is false, this computes the shortest distance from the initial
-// state to each state S and stores the value in the distance vector. If
-// reverse is true, this computes the shortest distance from each state to the
-// final states. An unvisited state S has distance Zero(), which will be stored
-// in the distance vector if S is less than the maximum visited state. The
-// state queue discipline is automatically-selected. The distance vector will
-// contain a unique element for which Member() is false if an error was
-// encountered.
-//
-// The weights must must be right (left) distributive if reverse is false (true)
-// and k-closed (i.e., 1 + x + x^2 + ... + x^(k +1) = 1 + x + x^2 + ... + x^k).
-//
-// Arc weights must satisfy the property that the sum of the weights of one or
-// more paths from some state S to T is never Zero(). In particular, arc weights
-// are never Zero().
-//
-// Complexity:
-//
-// Depends on properties of the semiring and the queue discipline.
-//
-// For more information, see:
-//
-// Mohri, M. 2002. Semiring framework and algorithms for
-// shortest-distance problems, Journal of Automata, Languages and
-// Combinatorics 7(3): 321-350, 2002.
-template <class Arc>
-void ShortestDistance(const Fst<Arc> &fst,
-                      std::vector<typename Arc::Weight> *distance,
-                      bool reverse = false, float delta = kShortestDelta) {
-  using StateId = typename Arc::StateId;
-  if (!reverse) {
-    AnyArcFilter<Arc> arc_filter;
-    AutoQueue<StateId> state_queue(fst, distance, arc_filter);
-    const ShortestDistanceOptions<Arc, AutoQueue<StateId>, AnyArcFilter<Arc>>
-        opts(&state_queue, arc_filter, kNoStateId, delta);
-    ShortestDistance(fst, distance, opts);
-  } else {
-    using ReverseArc = ReverseArc<Arc>;
-    using ReverseWeight = typename ReverseArc::Weight;
-    AnyArcFilter<ReverseArc> rarc_filter;
-    VectorFst<ReverseArc> rfst;
-    Reverse(fst, &rfst);
-    std::vector<ReverseWeight> rdistance;
-    AutoQueue<StateId> state_queue(rfst, &rdistance, rarc_filter);
-    const ShortestDistanceOptions<ReverseArc, AutoQueue<StateId>,
-                                  AnyArcFilter<ReverseArc>>
-        ropts(&state_queue, rarc_filter, kNoStateId, delta);
-    ShortestDistance(rfst, &rdistance, ropts);
-    distance->clear();
-    if (rdistance.size() == 1 && !rdistance[0].Member()) {
-      distance->assign(1, Arc::Weight::NoWeight());
-      return;
-    }
-    DCHECK_GE(rdistance.size(), 1);  // reversing added one state
-    distance->reserve(rdistance.size() - 1);
-    while (distance->size() < rdistance.size() - 1) {
-      distance->push_back(rdistance[distance->size() + 1].Reverse());
-    }
-  }
-}
-
-// Return the sum of the weight of all successful paths in an FST, i.e., the
-// shortest-distance from the initial state to the final states. Returns a
-// weight such that Member() is false if an error was encountered.
-template <class Arc>
-typename Arc::Weight ShortestDistance(const Fst<Arc> &fst,
-                                      float delta = kShortestDelta) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  std::vector<Weight> distance;
-  if (Weight::Properties() & kRightSemiring) {
-    ShortestDistance(fst, &distance, false, delta);
-    if (distance.size() == 1 && !distance[0].Member()) {
-      return Arc::Weight::NoWeight();
-    }
-    Adder<Weight> adder;  // maintains cumulative sum accurately
-    for (StateId state = 0; state < distance.size(); ++state) {
-      adder.Add(Times(distance[state], fst.Final(state)));
-    }
-    return adder.Sum();
-  } else {
-    ShortestDistance(fst, &distance, true, delta);
-    const auto state = fst.Start();
-    if (distance.size() == 1 && !distance[0].Member()) {
-      return Arc::Weight::NoWeight();
-    }
-    return state != kNoStateId && state < distance.size() ? distance[state]
-                                                          : Weight::Zero();
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_SHORTEST_DISTANCE_H_
diff --git a/projects/llm_framework/include/fst/shortest-path.h b/projects/llm_framework/include/fst/shortest-path.h
deleted file mode 100644
index 7bd04711..00000000
--- a/projects/llm_framework/include/fst/shortest-path.h
+++ /dev/null
@@ -1,549 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions to find shortest paths in an FST.
-
-#ifndef FST_SHORTEST_PATH_H_
-#define FST_SHORTEST_PATH_H_
-
-#include <functional>
-#include <type_traits>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/determinize.h>
-#include <fst/queue.h>
-#include <fst/shortest-distance.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-template <class Arc, class Queue, class ArcFilter>
-struct ShortestPathOptions
-    : public ShortestDistanceOptions<Arc, Queue, ArcFilter> {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  int32 nshortest;    // Returns n-shortest paths.
-  bool unique;        // Only returns paths with distinct input strings.
-  bool has_distance;  // Distance vector already contains the
-                      // shortest distance from the initial state.
-  bool first_path;    // Single shortest path stops after finding the first
-                      // path to a final state; that path is the shortest path
-                      // only when:
-                      // (1) using the ShortestFirstQueue with all the weights
-                      // in the FST being between One() and Zero() according to
-                      // NaturalLess or when
-                      // (2) using the NaturalAStarQueue with an admissible
-                      // and consistent estimate.
-  Weight weight_threshold;  // Pruning weight threshold.
-  StateId state_threshold;  // Pruning state threshold.
-
-  ShortestPathOptions(Queue *queue, ArcFilter filter, int32 nshortest = 1,
-                      bool unique = false, bool has_distance = false,
-                      float delta = kShortestDelta, bool first_path = false,
-                      Weight weight_threshold = Weight::Zero(),
-                      StateId state_threshold = kNoStateId)
-      : ShortestDistanceOptions<Arc, Queue, ArcFilter>(queue, filter,
-                                                       kNoStateId, delta),
-        nshortest(nshortest),
-        unique(unique),
-        has_distance(has_distance),
-        first_path(first_path),
-        weight_threshold(std::move(weight_threshold)),
-        state_threshold(state_threshold) {}
-};
-
-namespace internal {
-
-constexpr size_t kNoArc = -1;
-
-// Helper function for SingleShortestPath building the shortest path as a left-
-// to-right machine backwards from the best final state. It takes the input
-// FST passed to SingleShortestPath and the parent vector and f_parent returned
-// by that function, and builds the result into the provided output mutable FS
-// This is not normally called by users; see ShortestPath instead.
-template <class Arc>
-void SingleShortestPathBacktrace(
-    const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-    const std::vector<std::pair<typename Arc::StateId, size_t>> &parent,
-    typename Arc::StateId f_parent) {
-  using StateId = typename Arc::StateId;
-  ofst->DeleteStates();
-  ofst->SetInputSymbols(ifst.InputSymbols());
-  ofst->SetOutputSymbols(ifst.OutputSymbols());
-  StateId s_p = kNoStateId;
-  StateId d_p = kNoStateId;
-  for (StateId state = f_parent, d = kNoStateId; state != kNoStateId;
-       d = state, state = parent[state].first) {
-    d_p = s_p;
-    s_p = ofst->AddState();
-    if (d == kNoStateId) {
-      ofst->SetFinal(s_p, ifst.Final(f_parent));
-    } else {
-      ArcIterator<Fst<Arc>> aiter(ifst, state);
-      aiter.Seek(parent[d].second);
-      auto arc = aiter.Value();
-      arc.nextstate = d_p;
-      ofst->AddArc(s_p, std::move(arc));
-    }
-  }
-  ofst->SetStart(s_p);
-  if (ifst.Properties(kError, false)) ofst->SetProperties(kError, kError);
-  ofst->SetProperties(
-      ShortestPathProperties(ofst->Properties(kFstProperties, false), true),
-      kFstProperties);
-}
-
-// Helper function for SingleShortestPath building a tree of shortest paths to
-// every final state in the input FST. It takes the input FST and parent values
-// computed by SingleShortestPath and builds into the output mutable FST the
-// subtree of ifst that consists only of the best paths to all final states.
-// This is not normally called by users; see ShortestPath instead.
-template <class Arc>
-void SingleShortestTree(
-    const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-    const std::vector<std::pair<typename Arc::StateId, size_t>> &parent) {
-  ofst->DeleteStates();
-  ofst->SetInputSymbols(ifst.InputSymbols());
-  ofst->SetOutputSymbols(ifst.OutputSymbols());
-  ofst->SetStart(ifst.Start());
-  for (StateIterator<Fst<Arc>> siter(ifst); !siter.Done(); siter.Next()) {
-    ofst->AddState();
-    ofst->SetFinal(siter.Value(), ifst.Final(siter.Value()));
-  }
-  for (const auto &pair : parent) {
-    if (pair.first != kNoStateId && pair.second != kNoArc) {
-      ArcIterator<Fst<Arc>> aiter(ifst, pair.first);
-      aiter.Seek(pair.second);
-      ofst->AddArc(pair.first, aiter.Value());
-    }
-  }
-  if (ifst.Properties(kError, false)) ofst->SetProperties(kError, kError);
-  ofst->SetProperties(
-      ShortestPathProperties(ofst->Properties(kFstProperties, false), true),
-      kFstProperties);
-}
-
-// Implements the stopping criterion when ShortestPathOptions::first_path
-// is set to true:
-//   operator()(s, d, f) == true
-//   iff every successful path through state 's' has a cost greater or equal
-//   to 'f' under the assumption that 'd' is the shortest distance to state 's'.
-// Correct when using the ShortestFirstQueue with all the weights in the FST
-// being between One() and Zero() according to NaturalLess
-template <typename S, typename W, typename Queue>
-struct FirstPathSelect {
-  FirstPathSelect(const Queue &) {}
-  bool operator()(S s, W d, W f) const { return f == Plus(d, f); }
-};
-
-// Specialisation for A*.
-// Correct when the estimate is admissible and consistent.
-template <typename S, typename W, typename Estimate>
-class FirstPathSelect<S, W, NaturalAStarQueue<S, W, Estimate>> {
- public:
-  using Queue = NaturalAStarQueue<S, W, Estimate>;
-
-  FirstPathSelect(const Queue &state_queue)
-    : estimate_(state_queue.GetCompare().GetEstimate()) {}
-
-  bool operator()(S s, W d, W f) const {
-    return f == Plus(Times(d, estimate_(s)), f);
-  }
-
- private:
-  const Estimate &estimate_;
-};
-
-// Shortest-path algorithm. It builds the output mutable FST so that it contains
-// the shortest path in the input FST; distance returns the shortest distances
-// from the source state to each state in the input FST, and the options struct
-// is
-// used to specify options such as the queue discipline, the arc filter and
-// delta. The super_final option is an output parameter indicating the final
-// state, and the parent argument is used for the storage of the backtrace path
-// for each state 1 to n, (i.e., the best previous state and the arc that
-// transition to state n.) The shortest path is the lowest weight path w.r.t.
-// the natural semiring order. The weights need to be right distributive and
-// have the path (kPath) property. False is returned if an error is encountered.
-//
-// This is not normally called by users; see ShortestPath instead (with n = 1).
-template <class Arc, class Queue, class ArcFilter>
-bool SingleShortestPath(
-    const Fst<Arc> &ifst, std::vector<typename Arc::Weight> *distance,
-    const ShortestPathOptions<Arc, Queue, ArcFilter> &opts,
-    typename Arc::StateId *f_parent,
-    std::vector<std::pair<typename Arc::StateId, size_t>> *parent) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  static_assert(IsPath<Weight>::value, "Weight must have path property.");
-  static_assert((Weight::Properties() & kRightSemiring) == kRightSemiring,
-                "Weight must be right distributive.");
-  parent->clear();
-  *f_parent = kNoStateId;
-  if (ifst.Start() == kNoStateId) return true;
-  std::vector<bool> enqueued;
-  auto state_queue = opts.state_queue;
-  const auto source = (opts.source == kNoStateId) ? ifst.Start() : opts.source;
-  bool final_seen = false;
-  auto f_distance = Weight::Zero();
-  distance->clear();
-  state_queue->Clear();
-  while (distance->size() < source) {
-    distance->push_back(Weight::Zero());
-    enqueued.push_back(false);
-    parent->emplace_back(kNoStateId, kNoArc);
-  }
-  distance->push_back(Weight::One());
-  parent->emplace_back(kNoStateId, kNoArc);
-  state_queue->Enqueue(source);
-  enqueued.push_back(true);
-  while (!state_queue->Empty()) {
-    const auto s = state_queue->Head();
-    state_queue->Dequeue();
-    enqueued[s] = false;
-    const auto sd = (*distance)[s];
-    // If we are using a shortest queue, no other path is going to be shorter
-    // than f_distance at this point.
-    using FirstPath = FirstPathSelect<StateId, Weight, Queue>;
-    if (opts.first_path && final_seen &&
-        FirstPath(*state_queue)(s, sd, f_distance)) {
-      break;
-    }
-    if (ifst.Final(s) != Weight::Zero()) {
-      const auto plus = Plus(f_distance, Times(sd, ifst.Final(s)));
-      if (f_distance != plus) {
-        f_distance = plus;
-        *f_parent = s;
-      }
-      if (!f_distance.Member()) return false;
-      final_seen = true;
-    }
-    for (ArcIterator<Fst<Arc>> aiter(ifst, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      while (distance->size() <= arc.nextstate) {
-        distance->push_back(Weight::Zero());
-        enqueued.push_back(false);
-        parent->emplace_back(kNoStateId, kNoArc);
-      }
-      auto &nd = (*distance)[arc.nextstate];
-      const auto weight = Times(sd, arc.weight);
-      if (nd != Plus(nd, weight)) {
-        nd = Plus(nd, weight);
-        if (!nd.Member()) return false;
-        (*parent)[arc.nextstate] = std::make_pair(s, aiter.Position());
-        if (!enqueued[arc.nextstate]) {
-          state_queue->Enqueue(arc.nextstate);
-          enqueued[arc.nextstate] = true;
-        } else {
-          state_queue->Update(arc.nextstate);
-        }
-      }
-    }
-  }
-  return true;
-}
-
-template <class StateId, class Weight>
-class ShortestPathCompare {
- public:
-  ShortestPathCompare(const std::vector<std::pair<StateId, Weight>> &pairs,
-                      const std::vector<Weight> &distance, StateId superfinal,
-                      float delta)
-      : pairs_(pairs),
-        distance_(distance),
-        superfinal_(superfinal),
-        delta_(delta) {}
-
-  bool operator()(const StateId x, const StateId y) const {
-    const auto &px = pairs_[x];
-    const auto &py = pairs_[y];
-    const auto wx = Times(PWeight(px.first), px.second);
-    const auto wy = Times(PWeight(py.first), py.second);
-    // Penalize complete paths to ensure correct results with inexact weights.
-    // This forms a strict weak order so long as ApproxEqual(a, b) =>
-    // ApproxEqual(a, c) for all c s.t. less_(a, c) && less_(c, b).
-    if (px.first == superfinal_ && py.first != superfinal_) {
-      return less_(wy, wx) || ApproxEqual(wx, wy, delta_);
-    } else if (py.first == superfinal_ && px.first != superfinal_) {
-      return less_(wy, wx) && !ApproxEqual(wx, wy, delta_);
-    } else {
-      return less_(wy, wx);
-    }
-  }
-
- private:
-  Weight PWeight(StateId state) const {
-    return (state == superfinal_)
-               ? Weight::One()
-               : (state < distance_.size()) ? distance_[state] : Weight::Zero();
-  }
-
-  const std::vector<std::pair<StateId, Weight>> &pairs_;
-  const std::vector<Weight> &distance_;
-  const StateId superfinal_;
-  const float delta_;
-  NaturalLess<Weight> less_;
-};
-
-// N-Shortest-path algorithm: implements the core n-shortest path algorithm.
-// The output is built reversed. See below for versions with more options and
-// *not reversed*.
-//
-// The output mutable FST contains the REVERSE of n'shortest paths in the input
-// FST; distance must contain the shortest distance from each state to a final
-// state in the input FST; delta is the convergence delta.
-//
-// The n-shortest paths are the n-lowest weight paths w.r.t. the natural
-// semiring order. The single path that can be read from the ith of at most n
-// transitions leaving the initial state of the input FST is the ith shortest
-// path. Disregarding the initial state and initial transitions, the
-// n-shortest paths, in fact, form a tree rooted at the single final state.
-//
-// The weights need to be left and right distributive (kSemiring) and have the
-// path (kPath) property.
-//
-// Arc weights must satisfy the property that the sum of the weights of one or
-// more paths from some state S to T is never Zero(). In particular, arc weights
-// are never Zero().
-//
-// For more information, see:
-//
-// Mohri, M, and Riley, M. 2002. An efficient algorithm for the n-best-strings
-// problem. In Proc. ICSLP.
-//
-// The algorithm relies on the shortest-distance algorithm. There are some
-// issues with the pseudo-code as written in the paper (viz., line 11).
-//
-// IMPLEMENTATION NOTE: The input FST can be a delayed FST and at any state in
-// its expansion the values of distance vector need only be defined at that time
-// for the states that are known to exist.
-template <class Arc, class RevArc>
-void NShortestPath(const Fst<RevArc> &ifst, MutableFst<Arc> *ofst,
-                   const std::vector<typename Arc::Weight> &distance,
-                   int32 nshortest, float delta = kShortestDelta,
-                   typename Arc::Weight weight_threshold = Arc::Weight::Zero(),
-                   typename Arc::StateId state_threshold = kNoStateId) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using Pair = std::pair<StateId, Weight>;
-  static_assert((Weight::Properties() & kPath) == kPath,
-                "Weight must have path property.");
-  static_assert((Weight::Properties() & kSemiring) == kSemiring,
-                "Weight must be distributive.");
-  if (nshortest <= 0) return;
-  ofst->DeleteStates();
-  ofst->SetInputSymbols(ifst.InputSymbols());
-  ofst->SetOutputSymbols(ifst.OutputSymbols());
-  // Each state in ofst corresponds to a path with weight w from the initial
-  // state of ifst to a state s in ifst, that can be characterized by a pair
-  // (s, w). The vector pairs maps each state in ofst to the corresponding
-  // pair maps states in ofst to the corresponding pair (s, w).
-  std::vector<Pair> pairs;
-  // The supefinal state is denoted by kNoStateId. The distance from the
-  // superfinal state to the final state is semiring One, so
-  // `distance[kNoStateId]` is not needed.
-  const ShortestPathCompare<StateId, Weight> compare(pairs, distance,
-                                                     kNoStateId, delta);
-  const NaturalLess<Weight> less;
-  if (ifst.Start() == kNoStateId || distance.size() <= ifst.Start() ||
-      distance[ifst.Start()] == Weight::Zero() ||
-      less(weight_threshold, Weight::One()) || state_threshold == 0) {
-    if (ifst.Properties(kError, false)) ofst->SetProperties(kError, kError);
-    return;
-  }
-  ofst->SetStart(ofst->AddState());
-  const auto final_state = ofst->AddState();
-  ofst->SetFinal(final_state, Weight::One());
-  while (pairs.size() <= final_state) {
-    pairs.push_back(std::make_pair(kNoStateId, Weight::Zero()));
-  }
-  pairs[final_state] = std::make_pair(ifst.Start(), Weight::One());
-  std::vector<StateId> heap;
-  heap.push_back(final_state);
-  const auto limit = Times(distance[ifst.Start()], weight_threshold);
-  // r[s + 1], s state in fst, is the number of states in ofst which
-  // corresponding pair contains s, i.e., it is number of paths computed so far
-  // to s. Valid for s == kNoStateId (the superfinal state).
-  std::vector<int> r;
-  while (!heap.empty()) {
-    std::pop_heap(heap.begin(), heap.end(), compare);
-    const auto state = heap.back();
-    const auto p = pairs[state];
-    heap.pop_back();
-    const auto d =
-        (p.first == kNoStateId)
-            ? Weight::One()
-            : (p.first < distance.size()) ? distance[p.first] : Weight::Zero();
-    if (less(limit, Times(d, p.second)) ||
-        (state_threshold != kNoStateId &&
-         ofst->NumStates() >= state_threshold)) {
-      continue;
-    }
-    while (r.size() <= p.first + 1) r.push_back(0);
-    ++r[p.first + 1];
-    if (p.first == kNoStateId) {
-      ofst->AddArc(ofst->Start(), Arc(0, 0, Weight::One(), state));
-    }
-    if ((p.first == kNoStateId) && (r[p.first + 1] == nshortest)) break;
-    if (r[p.first + 1] > nshortest) continue;
-    if (p.first == kNoStateId) continue;
-    for (ArcIterator<Fst<RevArc>> aiter(ifst, p.first); !aiter.Done();
-         aiter.Next()) {
-      const auto &rarc = aiter.Value();
-      Arc arc(rarc.ilabel, rarc.olabel, rarc.weight.Reverse(), rarc.nextstate);
-      const auto weight = Times(p.second, arc.weight);
-      const auto next = ofst->AddState();
-      pairs.push_back(std::make_pair(arc.nextstate, weight));
-      arc.nextstate = state;
-      ofst->AddArc(next, std::move(arc));
-      heap.push_back(next);
-      std::push_heap(heap.begin(), heap.end(), compare);
-    }
-    const auto final_weight = ifst.Final(p.first).Reverse();
-    if (final_weight != Weight::Zero()) {
-      const auto weight = Times(p.second, final_weight);
-      const auto next = ofst->AddState();
-      pairs.push_back(std::make_pair(kNoStateId, weight));
-      ofst->AddArc(next, Arc(0, 0, final_weight, state));
-      heap.push_back(next);
-      std::push_heap(heap.begin(), heap.end(), compare);
-    }
-  }
-  Connect(ofst);
-  if (ifst.Properties(kError, false)) ofst->SetProperties(kError, kError);
-  ofst->SetProperties(
-      ShortestPathProperties(ofst->Properties(kFstProperties, false)),
-      kFstProperties);
-}
-
-}  // namespace internal
-
-// N-Shortest-path algorithm: this version allows finer control via the options
-// argument. See below for a simpler interface. The output mutable FST contains
-// the n-shortest paths in the input FST; the distance argument is used to
-// return the shortest distances from the source state to each state in the
-// input FST, and the options struct is used to specify the number of paths to
-// return, whether they need to have distinct input strings, the queue
-// discipline, the arc filter and the convergence delta.
-//
-// The n-shortest paths are the n-lowest weight paths w.r.t. the natural
-// semiring order. The single path that can be read from the ith of at most n
-// transitions leaving the initial state of the output FST is the ith shortest
-// path.
-// Disregarding the initial state and initial transitions, The n-shortest paths,
-// in fact, form a tree rooted at the single final state.
-//
-// The weights need to be right distributive and have the path (kPath) property.
-// They need to be left distributive as well for nshortest > 1.
-//
-// For more information, see:
-//
-// Mohri, M, and Riley, M. 2002. An efficient algorithm for the n-best-strings
-// problem. In Proc. ICSLP.
-//
-// The algorithm relies on the shortest-distance algorithm. There are some
-// issues with the pseudo-code as written in the paper (viz., line 11).
-template <class Arc, class Queue, class ArcFilter,
-          typename std::enable_if<IsPath<typename Arc::Weight>::value>::type * =
-              nullptr>
-void ShortestPath(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-                  std::vector<typename Arc::Weight> *distance,
-                  const ShortestPathOptions<Arc, Queue, ArcFilter> &opts) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using RevArc = ReverseArc<Arc>;
-  if (opts.nshortest == 1) {
-    std::vector<std::pair<StateId, size_t>> parent;
-    StateId f_parent;
-    if (internal::SingleShortestPath(ifst, distance, opts, &f_parent,
-                                     &parent)) {
-      internal::SingleShortestPathBacktrace(ifst, ofst, parent, f_parent);
-    } else {
-      ofst->SetProperties(kError, kError);
-    }
-    return;
-  }
-  if (opts.nshortest <= 0) return;
-  if (!opts.has_distance) {
-    ShortestDistance(ifst, distance, opts);
-    if (distance->size() == 1 && !(*distance)[0].Member()) {
-      ofst->SetProperties(kError, kError);
-      return;
-    }
-  }
-  // Algorithm works on the reverse of 'fst'; 'distance' is the distance to the
-  // final state in 'rfst', 'ofst' is built as the reverse of the tree of
-  // n-shortest path in 'rfst'.
-  VectorFst<RevArc> rfst;
-  Reverse(ifst, &rfst);
-  auto d = Weight::Zero();
-  for (ArcIterator<VectorFst<RevArc>> aiter(rfst, 0); !aiter.Done();
-       aiter.Next()) {
-    const auto &arc = aiter.Value();
-    const auto state = arc.nextstate - 1;
-    if (state < distance->size()) {
-      d = Plus(d, Times(arc.weight.Reverse(), (*distance)[state]));
-    }
-  }
-  // TODO(kbg): Avoid this expensive vector operation.
-  distance->insert(distance->begin(), d);
-  if (!opts.unique) {
-    internal::NShortestPath(rfst, ofst, *distance, opts.nshortest, opts.delta,
-                            opts.weight_threshold, opts.state_threshold);
-  } else {
-    std::vector<Weight> ddistance;
-    DeterminizeFstOptions<RevArc> dopts(opts.delta);
-    DeterminizeFst<RevArc> dfst(rfst, distance, &ddistance, dopts);
-    internal::NShortestPath(dfst, ofst, ddistance, opts.nshortest, opts.delta,
-                            opts.weight_threshold, opts.state_threshold);
-  }
-  // TODO(kbg): Avoid this expensive vector operation.
-  distance->erase(distance->begin());
-}
-
-template <class Arc, class Queue, class ArcFilter,
-          typename std::enable_if<!IsPath<typename Arc::Weight>::value>::type
-              * = nullptr>
-void ShortestPath(const Fst<Arc> &, MutableFst<Arc> *ofst,
-                  std::vector<typename Arc::Weight> *,
-                  const ShortestPathOptions<Arc, Queue, ArcFilter> &) {
-  FSTERROR() << "ShortestPath: Weight needs to have the "
-             << "path property and be distributive: " << Arc::Weight::Type();
-  ofst->SetProperties(kError, kError);
-}
-
-// Shortest-path algorithm: simplified interface. See above for a version that
-// allows finer control. The output mutable FST contains the n-shortest paths
-// in the input FST. The queue discipline is automatically selected. When unique
-// is true, only paths with distinct input label sequences are returned.
-//
-// The n-shortest paths are the n-lowest weight paths w.r.t. the natural
-// semiring order. The single path that can be read from the ith of at most n
-// transitions leaving the initial state of the output FST is the ith best path.
-// The weights need to be right distributive and have the path (kPath) property.
-template <class Arc>
-void ShortestPath(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-                  int32 nshortest = 1, bool unique = false,
-                  bool first_path = false,
-                  typename Arc::Weight weight_threshold = Arc::Weight::Zero(),
-                  typename Arc::StateId state_threshold = kNoStateId,
-                  float delta = kShortestDelta) {
-  using StateId = typename Arc::StateId;
-  std::vector<typename Arc::Weight> distance;
-  AnyArcFilter<Arc> arc_filter;
-  AutoQueue<StateId> state_queue(ifst, &distance, arc_filter);
-  const ShortestPathOptions<Arc, AutoQueue<StateId>, AnyArcFilter<Arc>> opts(
-      &state_queue, arc_filter, nshortest, unique, false, delta, first_path,
-      weight_threshold, state_threshold);
-  ShortestPath(ifst, ofst, &distance, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_SHORTEST_PATH_H_
diff --git a/projects/llm_framework/include/fst/signed-log-weight.h b/projects/llm_framework/include/fst/signed-log-weight.h
deleted file mode 100644
index 41339bc8..00000000
--- a/projects/llm_framework/include/fst/signed-log-weight.h
+++ /dev/null
@@ -1,440 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// LogWeight along with sign information that represents the value X in the
-// linear domain as <sign(X), -ln(|X|)>
-//
-// The sign is a TropicalWeight:
-//  positive, TropicalWeight.Value() > 0.0, recommended value 1.0
-//  negative, TropicalWeight.Value() <= 0.0, recommended value -1.0
-
-#ifndef FST_SIGNED_LOG_WEIGHT_H_
-#define FST_SIGNED_LOG_WEIGHT_H_
-
-#include <cstdlib>
-
-#include <fst/float-weight.h>
-#include <fst/pair-weight.h>
-#include <fst/product-weight.h>
-
-
-namespace fst {
-template <class T>
-class SignedLogWeightTpl : public PairWeight<TropicalWeight, LogWeightTpl<T>> {
- public:
-  using X1 = TropicalWeight;
-  using X2 = LogWeightTpl<T>;
-  using ReverseWeight = SignedLogWeightTpl;
-
-  using PairWeight<X1, X2>::Value1;
-  using PairWeight<X1, X2>::Value2;
-
-  SignedLogWeightTpl() noexcept : PairWeight<X1, X2>() {}
-
-  explicit SignedLogWeightTpl(const PairWeight<X1, X2> &w)
-      : PairWeight<X1, X2>(w) {}
-
-  SignedLogWeightTpl(const X1 &x1, const X2 &x2) : PairWeight<X1, X2>(x1, x2) {}
-
-  static const SignedLogWeightTpl &Zero() {
-    static const SignedLogWeightTpl zero(X1(1.0), X2::Zero());
-    return zero;
-  }
-
-  static const SignedLogWeightTpl &One() {
-    static const SignedLogWeightTpl one(X1(1.0), X2::One());
-    return one;
-  }
-
-  static const SignedLogWeightTpl &NoWeight() {
-    static const SignedLogWeightTpl no_weight(X1(1.0), X2::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type =
-        new string("signed_log_" + X1::Type() + "_" + X2::Type());
-    return *type;
-  }
-
-  bool IsPositive() const { return Value1().Value() > 0; }
-
-  SignedLogWeightTpl Quantize(float delta = kDelta) const {
-    return SignedLogWeightTpl(PairWeight<X1, X2>::Quantize(delta));
-  }
-
-  ReverseWeight Reverse() const {
-    return SignedLogWeightTpl(PairWeight<X1, X2>::Reverse());
-  }
-
-  bool Member() const { return PairWeight<X1, X2>::Member(); }
-
-  // Neither idempotent nor path.
-  static constexpr uint64 Properties() {
-    return kLeftSemiring | kRightSemiring | kCommutative;
-  }
-
-  size_t Hash() const {
-    size_t h1;
-    if (Value2() == X2::Zero() || IsPositive()) {
-      h1 = TropicalWeight(1.0).Hash();
-    } else {
-      h1 = TropicalWeight(-1.0).Hash();
-    }
-    size_t h2 = Value2().Hash();
-    static constexpr int lshift = 5;
-    static constexpr int rshift = CHAR_BIT * sizeof(size_t) - 5;
-    return h1 << lshift ^ h1 >> rshift ^ h2;
-  }
-};
-
-template <class T>
-inline SignedLogWeightTpl<T> Plus(const SignedLogWeightTpl<T> &w1,
-                                  const SignedLogWeightTpl<T> &w2) {
-  using X1 = TropicalWeight;
-  using X2 = LogWeightTpl<T>;
-  if (!w1.Member() || !w2.Member()) return SignedLogWeightTpl<T>::NoWeight();
-  const auto s1 = w1.IsPositive();
-  const auto s2 = w2.IsPositive();
-  const bool equal = (s1 == s2);
-  const auto f1 = w1.Value2().Value();
-  const auto f2 = w2.Value2().Value();
-  if (f1 == FloatLimits<T>::PosInfinity()) {
-    return w2;
-  } else if (f2 == FloatLimits<T>::PosInfinity()) {
-    return w1;
-  } else if (f1 == f2) {
-    if (equal) {
-      return SignedLogWeightTpl<T>(X1(w1.Value1()), X2(f2 - log(2.0F)));
-    } else {
-      return SignedLogWeightTpl<T>::Zero();
-    }
-  } else if (f1 > f2) {
-    if (equal) {
-      return SignedLogWeightTpl<T>(X1(w1.Value1()),
-                                   X2(f2 - internal::LogPosExp(f1 - f2)));
-    } else {
-      return SignedLogWeightTpl<T>(X1(w2.Value1()),
-                                   X2((f2 - internal::LogNegExp(f1 - f2))));
-    }
-  } else {
-    if (equal) {
-      return SignedLogWeightTpl<T>(X1(w2.Value1()),
-                                   X2((f1 - internal::LogPosExp(f2 - f1))));
-    } else {
-      return SignedLogWeightTpl<T>(X1(w1.Value1()),
-                                   X2((f1 - internal::LogNegExp(f2 - f1))));
-    }
-  }
-}
-
-template <class T>
-inline SignedLogWeightTpl<T> Minus(const SignedLogWeightTpl<T> &w1,
-                                   const SignedLogWeightTpl<T> &w2) {
-  SignedLogWeightTpl<T> minus_w2(-w2.Value1().Value(), w2.Value2());
-  return Plus(w1, minus_w2);
-}
-
-template <class T>
-inline SignedLogWeightTpl<T> Times(const SignedLogWeightTpl<T> &w1,
-                                   const SignedLogWeightTpl<T> &w2) {
-  using X2 = LogWeightTpl<T>;
-  if (!w1.Member() || !w2.Member()) return SignedLogWeightTpl<T>::NoWeight();
-  const auto s1 = w1.IsPositive();
-  const auto s2 = w2.IsPositive();
-  const auto f1 = w1.Value2().Value();
-  const auto f2 = w2.Value2().Value();
-  if (s1 == s2) {
-    return SignedLogWeightTpl<T>(TropicalWeight(1.0), X2(f1 + f2));
-  } else {
-    return SignedLogWeightTpl<T>(TropicalWeight(-1.0), X2(f1 + f2));
-  }
-}
-
-template <class T>
-inline SignedLogWeightTpl<T> Divide(const SignedLogWeightTpl<T> &w1,
-                                    const SignedLogWeightTpl<T> &w2,
-                                    DivideType typ = DIVIDE_ANY) {
-  using X2 = LogWeightTpl<T>;
-  if (!w1.Member() || !w2.Member()) return SignedLogWeightTpl<T>::NoWeight();
-  const auto s1 = w1.IsPositive();
-  const auto s2 = w2.IsPositive();
-  const auto f1 = w1.Value2().Value();
-  const auto f2 = w2.Value2().Value();
-  if (f2 == FloatLimits<T>::PosInfinity()) {
-    return SignedLogWeightTpl<T>(TropicalWeight(1.0),
-                                 X2(FloatLimits<T>::NumberBad()));
-  } else if (f1 == FloatLimits<T>::PosInfinity()) {
-    return SignedLogWeightTpl<T>(TropicalWeight(1.0),
-                                 X2(FloatLimits<T>::PosInfinity()));
-  } else if (s1 == s2) {
-    return SignedLogWeightTpl<T>(TropicalWeight(1.0), X2(f1 - f2));
-  } else {
-    return SignedLogWeightTpl<T>(TropicalWeight(-1.0), X2(f1 - f2));
-  }
-}
-
-template <class T>
-inline bool ApproxEqual(const SignedLogWeightTpl<T> &w1,
-                        const SignedLogWeightTpl<T> &w2, float delta = kDelta) {
-  using X2 = LogWeightTpl<T>;
-  if (w1.IsPositive() == w2.IsPositive()) {
-    return ApproxEqual(w1.Value2(), w2.Value2(), delta);
-  } else {
-    return ApproxEqual(w1.Value2(), X2::Zero(), delta)
-        && ApproxEqual(w2.Value2(), X2::Zero(), delta);
-  }
-}
-
-template <class T>
-inline bool operator==(const SignedLogWeightTpl<T> &w1,
-                       const SignedLogWeightTpl<T> &w2) {
-  using X2 = LogWeightTpl<T>;
-  if (w1.IsPositive() == w2.IsPositive()) {
-    return w1.Value2() == w2.Value2();
-  } else {
-    return w1.Value2() == X2::Zero()
-        && w2.Value2() == X2::Zero();
-  }
-}
-
-template <class T>
-inline bool operator!=(const SignedLogWeightTpl<T> &w1,
-                       const SignedLogWeightTpl<T> &w2) {
-  return !(w1 == w2);
-}
-
-// Single-precision signed-log weight.
-using SignedLogWeight = SignedLogWeightTpl<float>;
-
-// Double-precision signed-log weight.
-using SignedLog64Weight = SignedLogWeightTpl<double>;
-
-template <class W1, class W2>
-bool SignedLogConvertCheck(W1 weight) {
-  if (weight.Value1().Value() < 0.0) {
-    FSTERROR() << "WeightConvert: Can't convert weight " << weight
-               << " from " << W1::Type() << " to " << W2::Type();
-    return false;
-  }
-  return true;
-}
-
-// Specialization using the Kahan compensated summation
-template <class T>
-class Adder<SignedLogWeightTpl<T>> {
- public:
-  using Weight = SignedLogWeightTpl<T>;
-  using X1 = TropicalWeight;
-  using X2 = LogWeightTpl<T>;
-
-  explicit Adder(Weight w = Weight::Zero())
-      : ssum_(w.IsPositive()),
-        sum_(w.Value2().Value()),
-        c_(0.0) { }
-
-  Weight Add(const Weight &w) {
-    const auto sw = w.IsPositive();
-    const auto f = w.Value2().Value();
-    const bool equal = (ssum_ == sw);
-
-    if (!Sum().Member() || f == FloatLimits<T>::PosInfinity()) {
-      return Sum();
-    } else if (!w.Member() || sum_ == FloatLimits<T>::PosInfinity()) {
-      sum_ = f;
-      ssum_ = sw;
-      c_ = 0.0;
-    } else if (f == sum_) {
-      if (equal) {
-        sum_ = internal::KahanLogSum(sum_, f, &c_);
-      } else {
-        sum_ = FloatLimits<T>::PosInfinity();
-        ssum_ = true;
-        c_ = 0.0;
-      }
-    } else if (f > sum_) {
-      if (equal) {
-        sum_ = internal::KahanLogSum(sum_, f, &c_);
-      } else {
-        sum_ = internal::KahanLogDiff(sum_, f, &c_);
-      }
-    } else {
-      if (equal) {
-        sum_ = internal::KahanLogSum(f, sum_, &c_);
-      } else {
-        sum_ = internal::KahanLogDiff(f, sum_, &c_);
-        ssum_ = sw;
-      }
-    }
-    return Sum();
-  }
-
-  Weight Sum() { return Weight(X1(ssum_ ? 1.0 : -1.0), X2(sum_)); }
-
-  void Reset(Weight w = Weight::Zero()) {
-    ssum_ = w.IsPositive();
-    sum_ = w.Value2().Value();
-    c_ = 0.0;
-  }
-
- private:
-  bool ssum_;   // true iff sign of sum is positive
-  double sum_;  // unsigned sum
-  double c_;    // Kahan compensation
-};
-
-// Converts to tropical.
-template <>
-struct WeightConvert<SignedLogWeight, TropicalWeight> {
-  TropicalWeight operator()(const SignedLogWeight &weight) const {
-    if (!SignedLogConvertCheck<SignedLogWeight, TropicalWeight>(weight)) {
-      return TropicalWeight::NoWeight();
-    }
-    return TropicalWeight(weight.Value2().Value());
-  }
-};
-
-template <>
-struct WeightConvert<SignedLog64Weight, TropicalWeight> {
-  TropicalWeight operator()(const SignedLog64Weight &weight) const {
-    if (!SignedLogConvertCheck<SignedLog64Weight, TropicalWeight>(weight)) {
-      return TropicalWeight::NoWeight();
-    }
-    return TropicalWeight(weight.Value2().Value());
-  }
-};
-
-// Converts to log.
-template <>
-struct WeightConvert<SignedLogWeight, LogWeight> {
-  LogWeight operator()(const SignedLogWeight &weight) const {
-    if (!SignedLogConvertCheck<SignedLogWeight, LogWeight>(weight)) {
-      return LogWeight::NoWeight();
-    }
-    return LogWeight(weight.Value2().Value());
-  }
-};
-
-template <>
-struct WeightConvert<SignedLog64Weight, LogWeight> {
-  LogWeight operator()(const SignedLog64Weight &weight) const {
-    if (!SignedLogConvertCheck<SignedLog64Weight, LogWeight>(weight)) {
-      return LogWeight::NoWeight();
-    }
-    return LogWeight(weight.Value2().Value());
-  }
-};
-
-// Converts to log64.
-template <>
-struct WeightConvert<SignedLogWeight, Log64Weight> {
-  Log64Weight operator()(const SignedLogWeight &weight) const {
-    if (!SignedLogConvertCheck<SignedLogWeight, Log64Weight>(weight)) {
-      return Log64Weight::NoWeight();
-    }
-    return Log64Weight(weight.Value2().Value());
-  }
-};
-
-template <>
-struct WeightConvert<SignedLog64Weight, Log64Weight> {
-  Log64Weight operator()(const SignedLog64Weight &weight) const {
-    if (!SignedLogConvertCheck<SignedLog64Weight, Log64Weight>(weight)) {
-      return Log64Weight::NoWeight();
-    }
-    return Log64Weight(weight.Value2().Value());
-  }
-};
-
-// Converts to signed log.
-template <>
-struct WeightConvert<TropicalWeight, SignedLogWeight> {
-  SignedLogWeight operator()(const TropicalWeight &weight) const {
-    return SignedLogWeight(1.0, weight.Value());
-  }
-};
-
-template <>
-struct WeightConvert<LogWeight, SignedLogWeight> {
-  SignedLogWeight operator()(const LogWeight &weight) const {
-    return SignedLogWeight(1.0, weight.Value());
-  }
-};
-
-template <>
-struct WeightConvert<Log64Weight, SignedLogWeight> {
-  SignedLogWeight operator()(const Log64Weight &weight) const {
-    return SignedLogWeight(1.0, weight.Value());
-  }
-};
-
-template <>
-struct WeightConvert<SignedLog64Weight, SignedLogWeight> {
-  SignedLogWeight operator()(const SignedLog64Weight &weight) const {
-    return SignedLogWeight(weight.Value1(), weight.Value2().Value());
-  }
-};
-
-// Converts to signed log64.
-template <>
-struct WeightConvert<TropicalWeight, SignedLog64Weight> {
-  SignedLog64Weight operator()(const TropicalWeight &weight) const {
-    return SignedLog64Weight(1.0, weight.Value());
-  }
-};
-
-template <>
-struct WeightConvert<LogWeight, SignedLog64Weight> {
-  SignedLog64Weight operator()(const LogWeight &weight) const {
-    return SignedLog64Weight(1.0, weight.Value());
-  }
-};
-
-template <>
-struct WeightConvert<Log64Weight, SignedLog64Weight> {
-  SignedLog64Weight operator()(const Log64Weight &weight) const {
-    return SignedLog64Weight(1.0, weight.Value());
-  }
-};
-
-template <>
-struct WeightConvert<SignedLogWeight, SignedLog64Weight> {
-  SignedLog64Weight operator()(const SignedLogWeight &weight) const {
-    return SignedLog64Weight(weight.Value1(), weight.Value2().Value());
-  }
-};
-
-// This function object returns SignedLogWeightTpl<T>'s that are random integers
-// chosen from [0, num_random_weights) times a random sign. This is intended
-// primarily for testing.
-template <class T>
-class WeightGenerate<SignedLogWeightTpl<T>> {
- public:
-  using Weight = SignedLogWeightTpl<T>;
-  using X1 = typename Weight::X1;
-  using X2 = typename Weight::X2;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t num_random_weights = kNumRandomWeights)
-    : allow_zero_(allow_zero), num_random_weights_(num_random_weights) {}
-
-  Weight operator()() const {
-    static const X1 negative_one(-1.0);
-    static const X1 positive_one(+1.0);
-    const int m = rand() % 2;                                    // NOLINT
-    const int n = rand() % (num_random_weights_ + allow_zero_);  // NOLINT
-    return Weight((m == 0) ? negative_one : positive_one,
-                  (allow_zero_ && n == num_random_weights_) ?
-                   X2::Zero() : X2(n));
-  }
-
- private:
-  // Permits Zero() and zero divisors.
-  const bool allow_zero_;
-  // Number of alternative random weights.
-  const size_t num_random_weights_;
-};
-
-}  // namespace fst
-
-#endif  // FST_SIGNED_LOG_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/sparse-power-weight.h b/projects/llm_framework/include/fst/sparse-power-weight.h
deleted file mode 100644
index b7f20981..00000000
--- a/projects/llm_framework/include/fst/sparse-power-weight.h
+++ /dev/null
@@ -1,209 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Cartesian power weight semiring operation definitions, using
-// SparseTupleWeight as underlying representation.
-
-#ifndef FST_SPARSE_POWER_WEIGHT_H_
-#define FST_SPARSE_POWER_WEIGHT_H_
-
-#include <climits>
-#include <string>
-
-#include <fst/sparse-tuple-weight.h>
-#include <fst/weight.h>
-
-
-namespace fst {
-
-// Sparse cartesian power semiring: W ^ n
-//
-// Forms:
-//
-//  - a left semimodule when W is a left semiring,
-//  - a right semimodule when W is a right semiring,
-//  - a bisemimodule when W is a semiring,
-//    the free semimodule of rank n over W
-//
-// The Times operation is overloaded to provide the left and right scalar
-// products.
-//
-// K is the key value type. kNoKey (-1) is reserved for internal use
-template <class W, class K = int>
-class SparsePowerWeight : public SparseTupleWeight<W, K> {
- public:
-  using ReverseWeight = SparsePowerWeight<typename W::ReverseWeight, K>;
-
-  SparsePowerWeight() {}
-
-  explicit SparsePowerWeight(const SparseTupleWeight<W, K> &weight)
-      : SparseTupleWeight<W, K>(weight) {}
-
-  template <class Iterator>
-  SparsePowerWeight(Iterator begin, Iterator end)
-      : SparseTupleWeight<W, K>(begin, end) {}
-
-  // Initialize component `key` to `weight`, with `default_weight` for all
-  // other components.
-  SparsePowerWeight(const K &key, const W &weight,
-                    const W &default_weight = W::Zero())
-      : SparseTupleWeight<W, K>(key, weight, default_weight) {}
-
-  static const SparsePowerWeight &Zero() {
-    static const SparsePowerWeight zero(SparseTupleWeight<W, K>::Zero());
-    return zero;
-  }
-
-  static const SparsePowerWeight &One() {
-    static const SparsePowerWeight one(SparseTupleWeight<W, K>::One());
-    return one;
-  }
-
-  static const SparsePowerWeight &NoWeight() {
-    static const SparsePowerWeight no_weight(
-        SparseTupleWeight<W, K>::NoWeight());
-    return no_weight;
-  }
-
-  // Overide this: Overwrite the Type method to reflect the key type if using
-  // a non-default key type.
-  static const string &Type() {
-    static const string *const type = [] {
-      string type = W::Type() + "_^n";
-      if (sizeof(K) != sizeof(uint32)) {
-        type += "_" + std::to_string(CHAR_BIT * sizeof(K));
-      }
-      return new string(type);
-    }();
-    return *type;
-  }
-
-  static constexpr uint64 Properties() {
-    return W::Properties() &
-           (kLeftSemiring | kRightSemiring | kCommutative | kIdempotent);
-  }
-
-  SparsePowerWeight Quantize(float delta = kDelta) const {
-    return SparsePowerWeight(SparseTupleWeight<W, K>::Quantize(delta));
-  }
-
-  ReverseWeight Reverse() const {
-    return ReverseWeight(SparseTupleWeight<W, K>::Reverse());
-  }
-};
-
-template <class W, class K, class M>
-inline SparsePowerWeight<W, K> SparsePowerWeightMap(
-    const SparsePowerWeight<W, K> &w1,
-    const SparsePowerWeight<W, K> &w2,
-    const M &operator_mapper) {
-  SparsePowerWeight<W, K> result;
-  SparseTupleWeightMap(&result, w1, w2, operator_mapper);
-  return result;
-}
-
-// Semimodule plus operation.
-template <class W, class K>
-inline SparsePowerWeight<W, K> Plus(const SparsePowerWeight<W, K> &w1,
-                                    const SparsePowerWeight<W, K> &w2) {
-  return SparsePowerWeightMap(w1, w2, [](const K &k, const W &v1, const W &v2) {
-    return Plus(v1, v2);
-  });
-}
-
-// Semimodule times operation.
-template <class W, class K>
-inline SparsePowerWeight<W, K> Times(const SparsePowerWeight<W, K> &w1,
-                                     const SparsePowerWeight<W, K> &w2) {
-  return SparsePowerWeightMap(w1, w2, [](const K &k, const W &v1, const W &v2) {
-    return Times(v1, v2);
-  });
-}
-
-// Semimodule divide operation.
-template <class W, class K>
-inline SparsePowerWeight<W, K> Divide(const SparsePowerWeight<W, K> &w1,
-                                      const SparsePowerWeight<W, K> &w2,
-                                      DivideType type = DIVIDE_ANY) {
-  return SparsePowerWeightMap(w1, w2,
-                              [type](const K &k, const W &v1, const W &v2) {
-                                return Divide(v1, v2, type);
-                              });
-}
-
-// Semimodule dot product operation.
-template <class W, class K>
-inline const W &DotProduct(const SparsePowerWeight<W, K> &w1,
-                           const SparsePowerWeight<W, K> &w2) {
-  const SparsePowerWeight<W, K> product = Times(w1, w2);
-  W result(W::Zero());
-  for (SparseTupleWeightIterator<W, K> it(product); !it.Done(); it.Next()) {
-    result = Plus(result, it.Value().second);
-  }
-  return result;
-}
-
-template <class W, class K>
-inline bool ApproxEqual(const SparsePowerWeight<W, K> &w1,
-                        const SparsePowerWeight<W, K> &w2,
-                        float delta = kDelta) {
-  auto result = SparsePowerWeightMap(
-      w1, w2, [delta](const K &k, const W &v1, const W &v2) {
-        return ApproxEqual(v1, v2, delta) ? W::One() : W::Zero();
-      });
-  return result == SparsePowerWeight<W, K>::One();
-}
-
-template <class W, class K>
-inline SparsePowerWeight<W, K> Times(const W &k,
-                                     const SparsePowerWeight<W, K> &w2) {
-  const SparseTupleWeight<W, K> t2(k);
-  const SparsePowerWeight<W, K> w1(t2);
-  return Times(w1, w2);
-}
-
-template <class W, class K>
-inline SparsePowerWeight<W, K> Times(const SparsePowerWeight<W, K> &w1,
-                                     const W &k) {
-  const SparseTupleWeight<W, K> t2(k);
-  const SparsePowerWeight<W, K> w2(t2);
-  return Times(w1, w2);
-}
-
-template <class W, class K>
-inline SparsePowerWeight<W, K> Divide(const SparsePowerWeight<W, K> &w1,
-                                      const W &k,
-                                      DivideType divide_type = DIVIDE_ANY) {
-  const SparseTupleWeight<W, K> t2(k);
-  const SparsePowerWeight<W, K> w2(t2);
-  return Divide(w1, w2, divide_type);
-}
-
-// This function object generates weights over the Cartesian power of rank
-// n over the underlying weight. This is intended primarily for testing.
-template <class W, class K>
-class WeightGenerate<SparsePowerWeight<W, K>> {
- public:
-  using Weight = SparsePowerWeight<W, K>;
-  using Generate = WeightGenerate<W>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t sparse_power_rank = 3)
-      : generate_(allow_zero), sparse_power_rank_(sparse_power_rank) {}
-
-  Weight operator()() const {
-    Weight weight;
-    for (size_t i = 1; i <= sparse_power_rank_; ++i) {
-      weight.PushBack(i, generate_(), true);
-    }
-    return weight;
-  }
-
- private:
-  const Generate generate_;
-  const size_t sparse_power_rank_;
-};
-
-}  // namespace fst
-
-#endif  // FST_SPARSE_POWER_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/sparse-tuple-weight.h b/projects/llm_framework/include/fst/sparse-tuple-weight.h
deleted file mode 100644
index 68c9c663..00000000
--- a/projects/llm_framework/include/fst/sparse-tuple-weight.h
+++ /dev/null
@@ -1,422 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Sparse version of tuple-weight, based on tuple-weight.h.
-// Internally stores sparse key, value pairs in linked list. The default value
-// element is the assumed value of unset keys. Internal singleton
-// implementation that stores first key, value pair as a initialized member
-// variable to avoid unnecessary allocation on heap. Use
-// SparseTupleWeightIterator to iterate through the key,value pairs. Note:
-// this does NOT iterate through the default value.
-//
-// Sparse tuple weight set operation definitions.
-
-#ifndef FST_SPARSE_TUPLE_WEIGHT_H_
-#define FST_SPARSE_TUPLE_WEIGHT_H_
-
-#include <algorithm>
-#include <list>
-#include <stack>
-#include <string>
-#include <unordered_map>
-#include <utility>
-
-
-#include <fst/weight.h>
-
-
-namespace fst {
-
-template <class W, class K>
-class SparseTupleWeightIterator;
-
-// Arbitrary dimension tuple weight, stored as a sorted linked-list.
-// W is any weight class, and K is the key value type. kNoKey (-1) is reserved
-// for internal use.
-template <class W, class K = int>
-class SparseTupleWeight {
- public:
-  using ReverseWeight = SparseTupleWeight<typename W::ReverseWeight, K>;
-
-  using Iterator = SparseTupleWeightIterator<W, K>;
-  using Pair = std::pair<K, W>;
-  using Weight = W;
-  using Index = K;
-
-  constexpr static K kNoKey = -1;
-
-  SparseTupleWeight() { Init(); }
-
-  ~SparseTupleWeight() noexcept = default;
-
-  template <class Iterator>
-  SparseTupleWeight(Iterator begin, Iterator end) {
-    Init();
-    // Assumes input iterator is sorted.
-    for (auto it = begin; it != end; ++it) PushBack(*it);
-  }
-
-  // Initialize component `key` to `weight`, with `default_weight` for all
-  // other components.
-  SparseTupleWeight(const K &key, const W &weight, const W &default_weight)
-      : default_(default_weight),
-        first_(weight == default_weight ? kNoKey : key, weight) {}
-
-  explicit SparseTupleWeight(const W &weight) { Init(weight); }
-
-  SparseTupleWeight(const SparseTupleWeight &weight) {
-    Init(weight.DefaultValue());
-    SetDefaultValue(weight.DefaultValue());
-    for (Iterator it(weight); !it.Done(); it.Next()) {
-      PushBack(it.Value());
-    }
-  }
-
-  SparseTupleWeight(SparseTupleWeight &&weight) noexcept
-      // Don't move the default, so weight.default_ is still valid.
-      : default_(weight.default_),  // NOLINT
-        first_(std::move(weight.first_)),
-        rest_(std::move(weight.rest_)) {
-    // move leaves the source in a valid but unspecified state.
-    // Make sure the source weight is empty.
-    weight.first_ = Pair(kNoKey, W::NoWeight());
-    weight.rest_.clear();
-  }
-
-  static const SparseTupleWeight &Zero() {
-    static const SparseTupleWeight zero(W::Zero());
-    return zero;
-  }
-
-  static const SparseTupleWeight &One() {
-    static const SparseTupleWeight one(W::One());
-    return one;
-  }
-
-  static const SparseTupleWeight &NoWeight() {
-    static const SparseTupleWeight no_weight(W::NoWeight());
-    return no_weight;
-  }
-
-  std::istream &Read(std::istream &strm) {
-    ReadType(strm, &default_);
-    ReadType(strm, &first_);
-    return ReadType(strm, &rest_);
-  }
-
-  std::ostream &Write(std::ostream &strm) const {
-    WriteType(strm, default_);
-    WriteType(strm, first_);
-    return WriteType(strm, rest_);
-  }
-
-  SparseTupleWeight &operator=(const SparseTupleWeight &weight) {
-    if (this == &weight) return *this;  // Checks for identity.
-    Init(weight.DefaultValue());
-    for (Iterator it(weight); !it.Done(); it.Next()) {
-      PushBack(it.Value());
-    }
-    return *this;
-  }
-
-  SparseTupleWeight &operator=(SparseTupleWeight &&weight) noexcept {
-    if (this == &weight) return *this;  // Checks for identity.
-    // Don't move the default, so weight.default_ is still valid.
-    default_ = weight.default_;
-    first_ = std::move(weight.first_);
-    rest_ = std::move(weight.rest_);
-    // move leaves the source in a valid but unspecified state.
-    // Make sure the source weight is empty.
-    weight.first_ = Pair(kNoKey, W::NoWeight());
-    weight.rest_.clear();
-    return *this;
-  }
-
-  bool Member() const {
-    if (!DefaultValue().Member()) return false;
-    for (Iterator it(*this); !it.Done(); it.Next()) {
-      if (!it.Value().second.Member()) return false;
-    }
-    return true;
-  }
-
-  // Assumes H() function exists for the hash of the key value.
-  size_t Hash() const {
-    size_t h = 0;
-    static const std::hash<K> H;
-    for (Iterator it(*this); !it.Done(); it.Next()) {
-      h = 5 * h + H(it.Value().first);
-      h = 13 * h + it.Value().second.Hash();
-    }
-    return h;
-  }
-
-  SparseTupleWeight Quantize(float delta = kDelta) const {
-    SparseTupleWeight weight;
-    for (Iterator it(*this); !it.Done(); it.Next()) {
-      weight.PushBack(it.Value().first, it.Value().second.Quantize(delta));
-    }
-    return weight;
-  }
-
-  ReverseWeight Reverse() const {
-    ReverseWeight weight(DefaultValue().Reverse());
-    for (Iterator it(*this); !it.Done(); it.Next()) {
-      weight.PushBack(it.Value().first, it.Value().second.Reverse());
-    }
-    return weight;
-  }
-
-  void Init(const W &default_value = W::Zero()) {
-    first_ = Pair(kNoKey, W::NoWeight());
-    // Initialized to the reserved key value.
-    default_ = default_value;
-    rest_.clear();
-  }
-
-  size_t Size() const {
-    if (first_.first == kNoKey) {
-      return 0;
-    } else {
-      return rest_.size() + 1;
-    }
-  }
-
-  inline void PushBack(const K &key, const W &weight,
-                       bool default_value_check = true) {
-    PushBack(std::make_pair(key, weight), default_value_check);
-  }
-
-  inline void PushBack(const Pair &pair, bool default_value_check = true) {
-    if (default_value_check && pair.second == default_) return;
-    if (first_.first == kNoKey) {
-      first_ = pair;
-    } else {
-      rest_.push_back(pair);
-    }
-  }
-
-  // Returns the `key`-th component, or the default value if not set.
-  const W &Value(const K &key) const {
-    // TODO(rybach): Consider binary search.
-    Iterator iter(*this);
-    for (; !iter.Done() && iter.Value().first < key; iter.Next()) continue;
-    return !iter.Done() && iter.Value().first == key ? iter.Value().second
-                                                     : DefaultValue();
-  }
-
-  void SetValue(const K &key, const W &w) {
-    if (w == DefaultValue()) {
-      ClearValue(key);
-    } else {
-      SetValueToNonDefault(key, w);
-    }
-  }
-
-  void SetDefaultValue(const W &value) { default_ = value; }
-
-  const W &DefaultValue() const { return default_; }
-
- private:
-  void SetValueToNonDefault(const K &key, const W &w) {
-    // Don't use SparseTupleWeightIterator, since that's const.
-    if (first_.first == kNoKey) {
-      first_ = Pair(key, w);
-    } else if (key < first_.first) {
-      rest_.push_front(first_);
-      first_ = Pair(key, w);
-    } else if (key == first_.first) {
-      first_.second = w;
-    } else {
-      const auto i =
-          std::find_if(rest_.begin(), rest_.end(),
-                       [key](const Pair &p) { return p.first >= key; });
-      if (i != rest_.end() && i->first == key) {
-        i->second = w;
-      } else {
-        rest_.insert(i, Pair(key, w));
-      }
-    }
-  }
-
-  // Removes the weight value for `key`, having the effect of setting
-  // it to `DefaultValue()`.
-  void ClearValue(const K &key) {
-    if (key == first_.first) {
-      if (!rest_.empty()) {
-        first_ = rest_.front();
-        rest_.pop_front();
-      } else {
-        first_.first = kNoKey;
-      }
-    } else if (key > first_.first) {
-      const auto i =
-          std::find_if(rest_.begin(), rest_.end(),
-                       [key](const Pair &p) { return p.first >= key; });
-      if (i != rest_.end() && i->first == key) {
-        rest_.erase(i);
-      }
-    }
-  }
-
-  // Assumed default value of uninitialized keys, by default W::Zero().
-  W default_;
-
-  // Key values pairs are first stored in first_, then fill rest_ this way we
-  // can avoid dynamic allocation in the common case where the weight is a
-  // single key/value pair.
-  Pair first_;
-  std::list<Pair> rest_;
-
-  friend class SparseTupleWeightIterator<W, K>;
-};
-
-// Declare storage for kNoKey since it is passed by reference.
-template <class W, class K>
-constexpr K SparseTupleWeight<W, K>::kNoKey;
-
-template <class W, class K>
-class SparseTupleWeightIterator {
- public:
-  using Pair = typename SparseTupleWeight<W, K>::Pair;
-  using const_iterator = typename std::list<Pair>::const_iterator;
-  using iterator = typename std::list<Pair>::iterator;
-
-  explicit SparseTupleWeightIterator(const SparseTupleWeight<W, K> &weight)
-      : first_(weight.first_),
-        rest_(weight.rest_),
-        init_(true),
-        iter_(rest_.begin()) {}
-
-  bool Done() const {
-    if (init_) {
-      return first_.first == SparseTupleWeight<W, K>::kNoKey;
-    } else {
-      return iter_ == rest_.end();
-    }
-  }
-
-  const Pair &Value() const { return init_ ? first_ : *iter_; }
-
-  void Next() {
-    if (init_) {
-      init_ = false;
-    } else {
-      ++iter_;
-    }
-  }
-
-  void Reset() {
-    init_ = true;
-    iter_ = rest_.begin();
-  }
-
- private:
-  const Pair &first_;
-  const std::list<Pair> &rest_;
-  bool init_;  // In the initialized state?
-  const_iterator iter_;
-};
-
-// M must be callable as a function W(K, W, W).
-// K will be kNoKey when mapping the default value.
-template <class W, class K, class M>
-inline void SparseTupleWeightMap(SparseTupleWeight<W, K> *result,
-                                 const SparseTupleWeight<W, K> &w1,
-                                 const SparseTupleWeight<W, K> &w2,
-                                 const M &operator_mapper) {
-  SparseTupleWeightIterator<W, K> w1_it(w1);
-  SparseTupleWeightIterator<W, K> w2_it(w2);
-  const auto &v1_def = w1.DefaultValue();
-  const auto &v2_def = w2.DefaultValue();
-  result->SetDefaultValue(
-      operator_mapper(SparseTupleWeight<W, K>::kNoKey, v1_def, v2_def));
-  while (!w1_it.Done() || !w2_it.Done()) {
-    const auto &k1 = (w1_it.Done()) ? w2_it.Value().first : w1_it.Value().first;
-    const auto &k2 = (w2_it.Done()) ? w1_it.Value().first : w2_it.Value().first;
-    const auto &v1 = (w1_it.Done()) ? v1_def : w1_it.Value().second;
-    const auto &v2 = (w2_it.Done()) ? v2_def : w2_it.Value().second;
-    if (k1 == k2) {
-      result->PushBack(k1, operator_mapper(k1, v1, v2));
-      if (!w1_it.Done()) w1_it.Next();
-      if (!w2_it.Done()) w2_it.Next();
-    } else if (k1 < k2) {
-      result->PushBack(k1, operator_mapper(k1, v1, v2_def));
-      w1_it.Next();
-    } else {
-      result->PushBack(k2, operator_mapper(k2, v1_def, v2));
-      w2_it.Next();
-    }
-  }
-}
-
-template <class W, class K>
-inline bool operator==(const SparseTupleWeight<W, K> &w1,
-                       const SparseTupleWeight<W, K> &w2) {
-  const auto &v1_def = w1.DefaultValue();
-  const auto &v2_def = w2.DefaultValue();
-  if (v1_def != v2_def) return false;
-  SparseTupleWeightIterator<W, K> w1_it(w1);
-  SparseTupleWeightIterator<W, K> w2_it(w2);
-  while (!w1_it.Done() || !w2_it.Done()) {
-    const auto &k1 = (w1_it.Done()) ? w2_it.Value().first : w1_it.Value().first;
-    const auto &k2 = (w2_it.Done()) ? w1_it.Value().first : w2_it.Value().first;
-    const auto &v1 = (w1_it.Done()) ? v1_def : w1_it.Value().second;
-    const auto &v2 = (w2_it.Done()) ? v2_def : w2_it.Value().second;
-    if (k1 == k2) {
-      if (v1 != v2) return false;
-      if (!w1_it.Done()) w1_it.Next();
-      if (!w2_it.Done()) w2_it.Next();
-    } else if (k1 < k2) {
-      if (v1 != v2_def) return false;
-      w1_it.Next();
-    } else {
-      if (v1_def != v2) return false;
-      w2_it.Next();
-    }
-  }
-  return true;
-}
-
-template <class W, class K>
-inline bool operator!=(const SparseTupleWeight<W, K> &w1,
-                       const SparseTupleWeight<W, K> &w2) {
-  return !(w1 == w2);
-}
-
-template <class W, class K>
-inline std::ostream &operator<<(std::ostream &strm,
-                                const SparseTupleWeight<W, K> &weight) {
-  CompositeWeightWriter writer(strm);
-  writer.WriteBegin();
-  writer.WriteElement(weight.DefaultValue());
-  for (SparseTupleWeightIterator<W, K> it(weight); !it.Done(); it.Next()) {
-    writer.WriteElement(it.Value().first);
-    writer.WriteElement(it.Value().second);
-  }
-  writer.WriteEnd();
-  return strm;
-}
-
-template <class W, class K>
-inline std::istream &operator>>(std::istream &strm,
-                                SparseTupleWeight<W, K> &weight) {
-  CompositeWeightReader reader(strm);
-  reader.ReadBegin();
-  W def;
-  bool more = reader.ReadElement(&def);
-  weight.Init(def);
-  while (more) {
-    K key;
-    reader.ReadElement(&key);
-    W v;
-    more = reader.ReadElement(&v);
-    weight.PushBack(key, v);
-  }
-  reader.ReadEnd();
-  return strm;
-}
-
-}  // namespace fst
-
-#endif  // FST_SPARSE_TUPLE_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/state-map.h b/projects/llm_framework/include/fst/state-map.h
deleted file mode 100644
index 065f1c0b..00000000
--- a/projects/llm_framework/include/fst/state-map.h
+++ /dev/null
@@ -1,613 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to map over/transform states e.g., sort transitions.
-//
-// Consider using when operation does not change the number of states.
-
-#ifndef FST_STATE_MAP_H_
-#define FST_STATE_MAP_H_
-
-#include <algorithm>
-#include <string>
-#include <utility>
-
-#include <fst/log.h>
-
-#include <fst/arc-map.h>
-#include <fst/cache.h>
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-// StateMapper Interface. The class determines how states are mapped; useful for
-// implementing operations that do not change the number of states.
-//
-// class StateMapper {
-//  public:
-//   using FromArc = A;
-//   using ToArc = B;
-//
-//   // Typical constructor.
-//   StateMapper(const Fst<A> &fst);
-//
-//   // Required copy constructor that allows updating FST argument;
-//   // pass only if relevant and changed.
-//   StateMapper(const StateMapper &mapper, const Fst<A> *fst = 0);
-//
-//   // Specifies initial state of result.
-//   B::StateId Start() const;
-//   // Specifies state's final weight in result.
-//   B::Weight Final(B::StateId state) const;
-//
-//   // These methods iterate through a state's arcs in result.
-//
-//   // Specifies state to iterate over.
-//   void SetState(B::StateId state);
-//
-//   // End of arcs?
-//   bool Done() const;
-//
-//   // Current arc.
-//   const B &Value() const;
-//
-//   // Advances to next arc (when !Done)
-//   void Next();
-//
-//   // Specifies input symbol table action the mapper requires (see above).
-//   MapSymbolsAction InputSymbolsAction() const;
-//
-//   // Specifies output symbol table action the mapper requires (see above).
-//   MapSymbolsAction OutputSymbolsAction() const;
-//
-//   // This specifies the known properties of an FST mapped by this
-//   // mapper. It takes as argument the input FST's known properties.
-//   uint64 Properties(uint64 props) const;
-// };
-//
-// We include a various state map versions below. One dimension of variation is
-// whether the mapping mutates its input, writes to a new result FST, or is an
-// on-the-fly Fst. Another dimension is how we pass the mapper. We allow passing
-// the mapper by pointer for cases that we need to change the state of the
-// user's mapper. We also include map versions that pass the mapper by value or
-// const reference when this suffices.
-
-// Maps an arc type A using a mapper function object C, passed by pointer. This
-// version modifies the input FST.
-template <class A, class C>
-void StateMap(MutableFst<A> *fst, C *mapper) {
-  if (mapper->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    fst->SetInputSymbols(nullptr);
-  }
-  if (mapper->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    fst->SetOutputSymbols(nullptr);
-  }
-  if (fst->Start() == kNoStateId) return;
-  const auto props = fst->Properties(kFstProperties, false);
-  fst->SetStart(mapper->Start());
-  for (StateIterator<Fst<A>> siter(*fst); !siter.Done(); siter.Next()) {
-    const auto state = siter.Value();
-    mapper->SetState(state);
-    fst->DeleteArcs(state);
-    for (; !mapper->Done(); mapper->Next()) {
-      fst->AddArc(state, mapper->Value());
-    }
-    fst->SetFinal(state, mapper->Final(state));
-  }
-  fst->SetProperties(mapper->Properties(props), kFstProperties);
-}
-
-// Maps an arc type A using a mapper function object C, passed by value.
-// This version modifies the input FST.
-template <class A, class C>
-void StateMap(MutableFst<A> *fst, C mapper) {
-  StateMap(fst, &mapper);
-}
-
-// Maps an arc type A to an arc type B using mapper functor C, passed by
-// pointer. This version writes to an output FST.
-template <class A, class B, class C>
-void StateMap(const Fst<A> &ifst, MutableFst<B> *ofst, C *mapper) {
-  ofst->DeleteStates();
-  if (mapper->InputSymbolsAction() == MAP_COPY_SYMBOLS) {
-    ofst->SetInputSymbols(ifst.InputSymbols());
-  } else if (mapper->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    ofst->SetInputSymbols(nullptr);
-  }
-  if (mapper->OutputSymbolsAction() == MAP_COPY_SYMBOLS) {
-    ofst->SetOutputSymbols(ifst.OutputSymbols());
-  } else if (mapper->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    ofst->SetOutputSymbols(nullptr);
-  }
-  const auto iprops = ifst.Properties(kCopyProperties, false);
-  if (ifst.Start() == kNoStateId) {
-    if (iprops & kError) ofst->SetProperties(kError, kError);
-    return;
-  }
-  // Adds all states.
-  if (ifst.Properties(kExpanded, false)) ofst->ReserveStates(CountStates(ifst));
-  for (StateIterator<Fst<A>> siter(ifst); !siter.Done(); siter.Next()) {
-    ofst->AddState();
-  }
-  ofst->SetStart(mapper->Start());
-  for (StateIterator<Fst<A>> siter(ifst); !siter.Done(); siter.Next()) {
-    const auto state = siter.Value();
-    mapper->SetState(state);
-    for (; !mapper->Done(); mapper->Next()) {
-      ofst->AddArc(state, mapper->Value());
-    }
-    ofst->SetFinal(state, mapper->Final(state));
-  }
-  const auto oprops = ofst->Properties(kFstProperties, false);
-  ofst->SetProperties(mapper->Properties(iprops) | oprops, kFstProperties);
-}
-
-// Maps an arc type A to an arc type B using mapper functor object C, passed by
-// value. This version writes to an output FST.
-template <class A, class B, class C>
-void StateMap(const Fst<A> &ifst, MutableFst<B> *ofst, C mapper) {
-  StateMap(ifst, ofst, &mapper);
-}
-
-using StateMapFstOptions = CacheOptions;
-
-template <class A, class B, class C>
-class StateMapFst;
-
-// Facade around StateIteratorBase<A> inheriting from StateIteratorBase<B>.
-template <class A, class B>
-class StateMapStateIteratorBase : public StateIteratorBase<B> {
- public:
-  using Arc = B;
-  using StateId = typename Arc::StateId;
-
-  explicit StateMapStateIteratorBase(StateIteratorBase<A> *base)
-      : base_(base) {}
-
-  bool Done() const final { return base_->Done(); }
-
-  StateId Value() const final { return base_->Value(); }
-
-  void Next() final { base_->Next(); }
-
-  void Reset() final { base_->Reset(); }
-
- private:
-  std::unique_ptr<StateIteratorBase<A>> base_;
-
-  StateMapStateIteratorBase() = delete;
-};
-
-namespace internal {
-
-// Implementation of delayed StateMapFst.
-template <class A, class B, class C>
-class StateMapFstImpl : public CacheImpl<B> {
- public:
-  using Arc = B;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<B>::SetType;
-  using FstImpl<B>::SetProperties;
-  using FstImpl<B>::SetInputSymbols;
-  using FstImpl<B>::SetOutputSymbols;
-
-  using CacheImpl<B>::PushArc;
-  using CacheImpl<B>::HasArcs;
-  using CacheImpl<B>::HasFinal;
-  using CacheImpl<B>::HasStart;
-  using CacheImpl<B>::SetArcs;
-  using CacheImpl<B>::SetFinal;
-  using CacheImpl<B>::SetStart;
-
-  friend class StateIterator<StateMapFst<A, B, C>>;
-
-  StateMapFstImpl(const Fst<A> &fst, const C &mapper,
-                  const StateMapFstOptions &opts)
-      : CacheImpl<B>(opts),
-        fst_(fst.Copy()),
-        mapper_(new C(mapper, fst_.get())),
-        own_mapper_(true) {
-    Init();
-  }
-
-  StateMapFstImpl(const Fst<A> &fst, C *mapper, const StateMapFstOptions &opts)
-      : CacheImpl<B>(opts),
-        fst_(fst.Copy()),
-        mapper_(mapper),
-        own_mapper_(false) {
-    Init();
-  }
-
-  StateMapFstImpl(const StateMapFstImpl<A, B, C> &impl)
-      : CacheImpl<B>(impl),
-        fst_(impl.fst_->Copy(true)),
-        mapper_(new C(*impl.mapper_, fst_.get())),
-        own_mapper_(true) {
-    Init();
-  }
-
-  ~StateMapFstImpl() override {
-    if (own_mapper_) delete mapper_;
-  }
-
-  StateId Start() {
-    if (!HasStart()) SetStart(mapper_->Start());
-    return CacheImpl<B>::Start();
-  }
-
-  Weight Final(StateId state) {
-    if (!HasFinal(state)) SetFinal(state, mapper_->Final(state));
-    return CacheImpl<B>::Final(state);
-  }
-
-  size_t NumArcs(StateId state) {
-    if (!HasArcs(state)) Expand(state);
-    return CacheImpl<B>::NumArcs(state);
-  }
-
-  size_t NumInputEpsilons(StateId state) {
-    if (!HasArcs(state)) Expand(state);
-    return CacheImpl<B>::NumInputEpsilons(state);
-  }
-
-  size_t NumOutputEpsilons(StateId state) {
-    if (!HasArcs(state)) Expand(state);
-    return CacheImpl<B>::NumOutputEpsilons(state);
-  }
-
-  void InitStateIterator(StateIteratorData<B> *datb) const {
-    StateIteratorData<A> data;
-    fst_->InitStateIterator(&data);
-    datb->base = data.base ? new StateMapStateIteratorBase<A, B>(data.base)
-        : nullptr;
-    datb->nstates = data.nstates;
-  }
-
-  void InitArcIterator(StateId state, ArcIteratorData<B> *data) {
-    if (!HasArcs(state)) Expand(state);
-    CacheImpl<B>::InitArcIterator(state, data);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && (fst_->Properties(kError, false) ||
-                            (mapper_->Properties(0) & kError))) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  void Expand(StateId state) {
-    // Adds exiting arcs.
-    for (mapper_->SetState(state); !mapper_->Done(); mapper_->Next()) {
-      PushArc(state, mapper_->Value());
-    }
-    SetArcs(state);
-  }
-
-  const Fst<A> *GetFst() const { return fst_.get(); }
-
- private:
-  void Init() {
-    SetType("statemap");
-    if (mapper_->InputSymbolsAction() == MAP_COPY_SYMBOLS) {
-      SetInputSymbols(fst_->InputSymbols());
-    } else if (mapper_->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-      SetInputSymbols(nullptr);
-    }
-    if (mapper_->OutputSymbolsAction() == MAP_COPY_SYMBOLS) {
-      SetOutputSymbols(fst_->OutputSymbols());
-    } else if (mapper_->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-      SetOutputSymbols(nullptr);
-    }
-    const auto props = fst_->Properties(kCopyProperties, false);
-    SetProperties(mapper_->Properties(props));
-  }
-
-  std::unique_ptr<const Fst<A>> fst_;
-  C *mapper_;
-  bool own_mapper_;
-};
-
-}  // namespace internal
-
-// Maps an arc type A to an arc type B using Mapper function object
-// C. This version is a delayed FST.
-template <class A, class B, class C>
-class StateMapFst : public ImplToFst<internal::StateMapFstImpl<A, B, C>> {
- public:
-  friend class ArcIterator<StateMapFst<A, B, C>>;
-
-  using Arc = B;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::StateMapFstImpl<A, B, C>;
-
-  StateMapFst(const Fst<A> &fst, const C &mapper,
-              const StateMapFstOptions &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, mapper, opts)) {}
-
-  StateMapFst(const Fst<A> &fst, C *mapper, const StateMapFstOptions &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, mapper, opts)) {}
-
-  StateMapFst(const Fst<A> &fst, const C &mapper)
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, mapper, StateMapFstOptions())) {}
-
-  StateMapFst(const Fst<A> &fst, C *mapper)
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, mapper, StateMapFstOptions())) {}
-
-  // See Fst<>::Copy() for doc.
-  StateMapFst(const StateMapFst<A, B, C> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this StateMapFst. See Fst<>::Copy() for further doc.
-  StateMapFst<A, B, C> *Copy(bool safe = false) const override {
-    return new StateMapFst<A, B, C>(*this, safe);
-  }
-
-  void InitStateIterator(StateIteratorData<B> *data) const override {
-    GetImpl()->InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId state, ArcIteratorData<B> *data) const override {
-    GetMutableImpl()->InitArcIterator(state, data);
-  }
-
- protected:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
- private:
-  StateMapFst &operator=(const StateMapFst &) = delete;
-};
-
-// Specialization for StateMapFst.
-template <class A, class B, class C>
-class ArcIterator<StateMapFst<A, B, C>>
-    : public CacheArcIterator<StateMapFst<A, B, C>> {
- public:
-  using StateId = typename A::StateId;
-
-  ArcIterator(const StateMapFst<A, B, C> &fst, StateId state)
-      : CacheArcIterator<StateMapFst<A, B, C>>(fst.GetMutableImpl(), state) {
-    if (!fst.GetImpl()->HasArcs(state)) fst.GetMutableImpl()->Expand(state);
-  }
-};
-
-// Utility mappers.
-
-// Mapper that returns its input.
-template <class Arc>
-class IdentityStateMapper {
- public:
-  using FromArc = Arc;
-  using ToArc = Arc;
-
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit IdentityStateMapper(const Fst<Arc> &fst) : fst_(fst) {}
-
-  // Allows updating FST argument; pass only if changed.
-  IdentityStateMapper(const IdentityStateMapper<Arc> &mapper,
-                      const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? *fst : mapper.fst_) {}
-
-  StateId Start() const { return fst_.Start(); }
-
-  Weight Final(StateId state) const { return fst_.Final(state); }
-
-  void SetState(StateId state) {
-    aiter_.reset(new ArcIterator<Fst<Arc>>(fst_, state));
-  }
-
-  bool Done() const { return aiter_->Done(); }
-
-  const Arc &Value() const { return aiter_->Value(); }
-
-  void Next() { aiter_->Next(); }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const { return props; }
-
- private:
-  const Fst<Arc> &fst_;
-  std::unique_ptr<ArcIterator<Fst<Arc>>> aiter_;
-};
-
-template <class Arc>
-class ArcSumMapper {
- public:
-  using FromArc = Arc;
-  using ToArc = Arc;
-
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit ArcSumMapper(const Fst<Arc> &fst) : fst_(fst), i_(0) {}
-
-  // Allows updating FST argument; pass only if changed.
-  ArcSumMapper(const ArcSumMapper<Arc> &mapper, const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? *fst : mapper.fst_), i_(0) {}
-
-  StateId Start() const { return fst_.Start(); }
-
-  Weight Final(StateId state) const { return fst_.Final(state); }
-
-  void SetState(StateId state) {
-    i_ = 0;
-    arcs_.clear();
-    arcs_.reserve(fst_.NumArcs(state));
-    for (ArcIterator<Fst<Arc>> aiter(fst_, state); !aiter.Done();
-         aiter.Next()) {
-      arcs_.push_back(aiter.Value());
-    }
-    // First sorts the exiting arcs by input label, output label and destination
-    // state and then sums weights of arcs with the same input label, output
-    // label, and destination state.
-    std::sort(arcs_.begin(), arcs_.end(), comp_);
-    size_t narcs = 0;
-    for (const auto &arc : arcs_) {
-      if (narcs > 0 && equal_(arc, arcs_[narcs - 1])) {
-        arcs_[narcs - 1].weight = Plus(arcs_[narcs - 1].weight, arc.weight);
-      } else {
-        arcs_[narcs] = arc;
-        ++narcs;
-      }
-    }
-    arcs_.resize(narcs);
-  }
-
-  bool Done() const { return i_ >= arcs_.size(); }
-
-  const Arc &Value() const { return arcs_[i_]; }
-
-  void Next() { ++i_; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return props & kArcSortProperties & kDeleteArcsProperties &
-           kWeightInvariantProperties;
-  }
-
- private:
-  struct Compare {
-    bool operator()(const Arc &x, const Arc &y) const {
-      if (x.ilabel < y.ilabel) return true;
-      if (x.ilabel > y.ilabel) return false;
-      if (x.olabel < y.olabel) return true;
-      if (x.olabel > y.olabel) return false;
-      if (x.nextstate < y.nextstate) return true;
-      if (x.nextstate > y.nextstate) return false;
-      return false;
-    }
-  };
-
-  struct Equal {
-    bool operator()(const Arc &x, const Arc &y) const {
-      return (x.ilabel == y.ilabel && x.olabel == y.olabel &&
-              x.nextstate == y.nextstate);
-    }
-  };
-
-  const Fst<Arc> &fst_;
-  Compare comp_;
-  Equal equal_;
-  std::vector<Arc> arcs_;
-  ssize_t i_;  // Current arc position.
-
-  ArcSumMapper &operator=(const ArcSumMapper &) = delete;
-};
-
-template <class Arc>
-class ArcUniqueMapper {
- public:
-  using FromArc = Arc;
-  using ToArc = Arc;
-
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit ArcUniqueMapper(const Fst<Arc> &fst) : fst_(fst), i_(0) {}
-
-  // Allows updating FST argument; pass only if changed.
-  ArcUniqueMapper(const ArcUniqueMapper<Arc> &mapper,
-                  const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? *fst : mapper.fst_), i_(0) {}
-
-  StateId Start() const { return fst_.Start(); }
-
-  Weight Final(StateId state) const { return fst_.Final(state); }
-
-  void SetState(StateId state) {
-    i_ = 0;
-    arcs_.clear();
-    arcs_.reserve(fst_.NumArcs(state));
-    for (ArcIterator<Fst<Arc>> aiter(fst_, state); !aiter.Done();
-         aiter.Next()) {
-      arcs_.push_back(aiter.Value());
-    }
-    // First sorts the exiting arcs by input label, output label and destination
-    // state and then uniques identical arcs.
-    std::sort(arcs_.begin(), arcs_.end(), comp_);
-    arcs_.erase(std::unique(arcs_.begin(), arcs_.end(), equal_), arcs_.end());
-  }
-
-  bool Done() const { return i_ >= arcs_.size(); }
-
-  const Arc &Value() const { return arcs_[i_]; }
-
-  void Next() { ++i_; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return props & kArcSortProperties & kDeleteArcsProperties;
-  }
-
- private:
-  struct Compare {
-    bool operator()(const Arc &x, const Arc &y) const {
-      if (x.ilabel < y.ilabel) return true;
-      if (x.ilabel > y.ilabel) return false;
-      if (x.olabel < y.olabel) return true;
-      if (x.olabel > y.olabel) return false;
-      if (x.nextstate < y.nextstate) return true;
-      if (x.nextstate > y.nextstate) return false;
-      return false;
-    }
-  };
-
-  struct Equal {
-    bool operator()(const Arc &x, const Arc &y) const {
-      return (x.ilabel == y.ilabel && x.olabel == y.olabel &&
-              x.nextstate == y.nextstate && x.weight == y.weight);
-    }
-  };
-
-  const Fst<Arc> &fst_;
-  Compare comp_;
-  Equal equal_;
-  std::vector<Arc> arcs_;
-  size_t i_;  // Current arc position.
-
-  ArcUniqueMapper &operator=(const ArcUniqueMapper &) = delete;
-};
-
-// Useful aliases when using StdArc.
-
-using StdArcSumMapper = ArcSumMapper<StdArc>;
-
-using StdArcUniqueMapper = ArcUniqueMapper<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_STATE_MAP_H_
diff --git a/projects/llm_framework/include/fst/state-reachable.h b/projects/llm_framework/include/fst/state-reachable.h
deleted file mode 100644
index 36b5559a..00000000
--- a/projects/llm_framework/include/fst/state-reachable.h
+++ /dev/null
@@ -1,224 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to determine whether a given (final) state can be reached from some
-// other given state.
-
-#ifndef FST_STATE_REACHABLE_H_
-#define FST_STATE_REACHABLE_H_
-
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/connect.h>
-#include <fst/dfs-visit.h>
-#include <fst/fst.h>
-#include <fst/interval-set.h>
-#include <fst/vector-fst.h>
-
-
-namespace fst {
-
-// Computes the (final) states reachable from a given state in an FST. After
-// this visitor has been called, a final state f can be reached from a state
-// s iff (*isets)[s].Member(state2index[f]) is true, where (*isets[s]) is a
-// set of half-open inteval of final state indices and state2index[f] maps from
-// a final state to its index. If state2index is empty, it is filled-in with
-// suitable indices. If it is non-empty, those indices are used; in this case,
-// the final states must have out-degree 0.
-template <class Arc, class I = typename Arc::StateId, class S = IntervalSet<I>>
-class IntervalReachVisitor {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Index = I;
-  using ISet = S;
-  using Interval = typename ISet::Interval;
-
-  IntervalReachVisitor(const Fst<Arc> &fst, std::vector<S> *isets,
-                       std::vector<Index> *state2index)
-      : fst_(fst),
-        isets_(isets),
-        state2index_(state2index),
-        index_(state2index->empty() ? 1 : -1),
-        error_(false) {
-    isets_->clear();
-  }
-
-  void InitVisit(const Fst<Arc> &) { error_ = false; }
-
-  bool InitState(StateId s, StateId r) {
-    while (isets_->size() <= s) isets_->push_back(S());
-    while (state2index_->size() <= s) state2index_->push_back(-1);
-    if (fst_.Final(s) != Weight::Zero()) {
-      // Create tree interval.
-      auto *intervals = (*isets_)[s].MutableIntervals();
-      if (index_ < 0) {  // Uses state2index_ map to set index.
-        if (fst_.NumArcs(s) > 0) {
-          FSTERROR() << "IntervalReachVisitor: state2index map must be empty "
-                     << "for this FST";
-          error_ = true;
-          return false;
-        }
-        const auto index = (*state2index_)[s];
-        if (index < 0) {
-          FSTERROR() << "IntervalReachVisitor: state2index map incomplete";
-          error_ = true;
-          return false;
-        }
-        intervals->push_back(Interval(index, index + 1));
-      } else {  // Use pre-order index.
-        intervals->push_back(Interval(index_, index_ + 1));
-        (*state2index_)[s] = index_++;
-      }
-    }
-    return true;
-  }
-
-  constexpr bool TreeArc(StateId, const Arc &) const { return true; }
-
-  bool BackArc(StateId s, const Arc &arc) {
-    FSTERROR() << "IntervalReachVisitor: Cyclic input";
-    error_ = true;
-    return false;
-  }
-
-  bool ForwardOrCrossArc(StateId s, const Arc &arc) {
-    // Non-tree interval.
-    (*isets_)[s].Union((*isets_)[arc.nextstate]);
-    return true;
-  }
-
-  void FinishState(StateId s, StateId p, const Arc *) {
-    if (index_ >= 0 && fst_.Final(s) != Weight::Zero()) {
-      auto *intervals = (*isets_)[s].MutableIntervals();
-      (*intervals)[0].end = index_;  // Updates tree interval end.
-    }
-    (*isets_)[s].Normalize();
-    if (p != kNoStateId) {
-      (*isets_)[p].Union((*isets_)[s]);  // Propagates intervals to parent.
-    }
-  }
-
-  void FinishVisit() {}
-
-  bool Error() const { return error_; }
-
- private:
-  const Fst<Arc> &fst_;
-  std::vector<ISet> *isets_;
-  std::vector<Index> *state2index_;
-  Index index_;
-  bool error_;
-};
-
-// Tests reachability of final states from a given state. To test for
-// reachability from a state s, first do SetState(s). Then a final state f can
-// be reached from state s of FST iff Reach(f) is true. The input can be cyclic,
-// but no cycle may contain a final state.
-template <class Arc, class I = typename Arc::StateId, class S = IntervalSet<I>>
-class StateReachable {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Index = I;
-  using ISet = S;
-  using Interval = typename ISet::Interval;
-
-  explicit StateReachable(const Fst<Arc> &fst) : error_(false) {
-    if (fst.Properties(kAcyclic, true)) {
-      AcyclicStateReachable(fst);
-    } else {
-      CyclicStateReachable(fst);
-    }
-  }
-
-  explicit StateReachable(const StateReachable<Arc> &reachable) {
-    FSTERROR() << "Copy constructor for state reachable class "
-               << "not implemented.";
-    error_ = true;
-  }
-
-  // Sets current state.
-  void SetState(StateId s) { s_ = s; }
-
-  // Can reach this final state from current state?
-  bool Reach(StateId s) {
-    if (s >= state2index_.size()) return false;
-    const auto i = state2index_[s];
-    if (i < 0) {
-      FSTERROR() << "StateReachable: State non-final: " << s;
-      error_ = true;
-      return false;
-    }
-    return isets_[s_].Member(i);
-  }
-
-  // Access to the state-to-index mapping. Unassigned states have index -1.
-  std::vector<Index> &State2Index() { return state2index_; }
-
-  // Access to the interval sets. These specify the reachability to the final
-  // states as intervals of the final state indices.
-  const std::vector<ISet> &IntervalSets() { return isets_; }
-
-  bool Error() const { return error_; }
-
- private:
-  void AcyclicStateReachable(const Fst<Arc> &fst) {
-    IntervalReachVisitor<Arc, StateId, ISet> reach_visitor(fst, &isets_,
-                                                           &state2index_);
-    DfsVisit(fst, &reach_visitor);
-    if (reach_visitor.Error()) error_ = true;
-  }
-
-  void CyclicStateReachable(const Fst<Arc> &fst) {
-    // Finds state reachability on the acyclic condensation FST.
-    VectorFst<Arc> cfst;
-    std::vector<StateId> scc;
-    Condense(fst, &cfst, &scc);
-    StateReachable reachable(cfst);
-    if (reachable.Error()) {
-      error_ = true;
-      return;
-    }
-    // Gets the number of states per SCC.
-    std::vector<size_t> nscc;
-    for (StateId s = 0; s < scc.size(); ++s) {
-      const auto c = scc[s];
-      while (c >= nscc.size()) nscc.push_back(0);
-      ++nscc[c];
-    }
-    // Constructs the interval sets and state index mapping for the original
-    // FST from the condensation FST.
-    state2index_.resize(scc.size(), -1);
-    isets_.resize(scc.size());
-    for (StateId s = 0; s < scc.size(); ++s) {
-      const auto c = scc[s];
-      isets_[s] = reachable.IntervalSets()[c];
-      state2index_[s] = reachable.State2Index()[c];
-      // Checks that each final state in an input FST is not contained in a
-      // cycle (i.e., not in a non-trivial SCC).
-      if (cfst.Final(c) != Weight::Zero() && nscc[c] > 1) {
-        FSTERROR() << "StateReachable: Final state contained in a cycle";
-        error_ = true;
-        return;
-      }
-    }
-  }
-
-  StateId s_;                       // Current state.
-  std::vector<ISet> isets_;         // Interval sets per state.
-  std::vector<Index> state2index_;  // Finds index for a final state.
-  bool error_;
-
-  StateReachable &operator=(const StateReachable &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_STATE_REACHABLE_H_
diff --git a/projects/llm_framework/include/fst/state-table.h b/projects/llm_framework/include/fst/state-table.h
deleted file mode 100644
index a5067592..00000000
--- a/projects/llm_framework/include/fst/state-table.h
+++ /dev/null
@@ -1,494 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes for representing the mapping between state tuples and state IDs.
-
-#ifndef FST_STATE_TABLE_H_
-#define FST_STATE_TABLE_H_
-
-#include <deque>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/bi-table.h>
-#include <fst/expanded-fst.h>
-#include <fst/filter-state.h>
-
-
-namespace fst {
-
-// State tables determine the bijective mapping between state tuples (e.g., in
-// composition, triples of two FST states and a composition filter state) and
-// their corresponding state IDs. They are classes, templated on state tuples,
-// with the following interface:
-//
-// template <class T>
-// class StateTable {
-//  public:
-//   using StateTuple = T;
-//
-//   // Required constructors.
-//   StateTable();
-//
-//   StateTable(const StateTable &);
-//
-//   // Looks up state ID by tuple. If it doesn't exist, then add it.
-//   StateId FindState(const StateTuple &tuple);
-//
-//   // Looks up state tuple by state ID.
-//   const StateTuple<StateId> &Tuple(StateId s) const;
-//
-//   // # of stored tuples.
-//   StateId Size() const;
-// };
-//
-// A state tuple has the form:
-//
-// template <class S>
-// struct StateTuple {
-//   using StateId = S;
-//
-//   // Required constructors.
-//
-//   StateTuple();
-//
-//   StateTuple(const StateTuple &tuple);
-// };
-
-// An implementation using a hash map for the tuple to state ID mapping. The
-// state tuple T must support operator==.
-template <class T, class H>
-class HashStateTable : public HashBiTable<typename T::StateId, T, H> {
- public:
-  using StateTuple = T;
-  using StateId = typename StateTuple::StateId;
-
-  using HashBiTable<StateId, StateTuple, H>::FindId;
-  using HashBiTable<StateId, StateTuple, H>::FindEntry;
-  using HashBiTable<StateId, StateTuple, H>::Size;
-
-  HashStateTable() : HashBiTable<StateId, StateTuple, H>() {}
-
-  explicit HashStateTable(size_t table_size)
-      : HashBiTable<StateId, StateTuple, H>(table_size) {}
-
-  StateId FindState(const StateTuple &tuple) { return FindId(tuple); }
-
-  const StateTuple &Tuple(StateId s) const { return FindEntry(s); }
-};
-
-// An implementation using a hash map for the tuple to state ID mapping. The
-// state tuple T must support operator==.
-template <class T, class H>
-class CompactHashStateTable
-    : public CompactHashBiTable<typename T::StateId, T, H> {
- public:
-  using StateTuple = T;
-  using StateId = typename StateTuple::StateId;
-
-  using CompactHashBiTable<StateId, StateTuple, H>::FindId;
-  using CompactHashBiTable<StateId, StateTuple, H>::FindEntry;
-  using CompactHashBiTable<StateId, StateTuple, H>::Size;
-
-  CompactHashStateTable() : CompactHashBiTable<StateId, StateTuple, H>() {}
-
-  explicit CompactHashStateTable(size_t table_size)
-      : CompactHashBiTable<StateId, StateTuple, H>(table_size) {}
-
-  StateId FindState(const StateTuple &tuple) { return FindId(tuple); }
-
-  const StateTuple &Tuple(StateId s) const { return FindEntry(s); }
-};
-
-// An implementation using a vector for the tuple to state mapping. It is
-// passed a fingerprint functor that should fingerprint tuples uniquely to an
-// integer that can used as a vector index. Normally, VectorStateTable
-// constructs the fingerprint functor. Alternately, the user can pass this
-// object, in which case the table takes ownership.
-template <class T, class FP>
-class VectorStateTable : public VectorBiTable<typename T::StateId, T, FP> {
- public:
-  using StateTuple = T;
-  using StateId = typename StateTuple::StateId;
-
-  using VectorBiTable<StateId, StateTuple, FP>::FindId;
-  using VectorBiTable<StateId, StateTuple, FP>::FindEntry;
-  using VectorBiTable<StateId, StateTuple, FP>::Size;
-  using VectorBiTable<StateId, StateTuple, FP>::Fingerprint;
-
-  explicit VectorStateTable(FP *fingerprint = nullptr, size_t table_size = 0)
-      : VectorBiTable<StateId, StateTuple, FP>(fingerprint, table_size) {}
-
-  StateId FindState(const StateTuple &tuple) { return FindId(tuple); }
-
-  const StateTuple &Tuple(StateId s) const { return FindEntry(s); }
-};
-
-// An implementation using a vector and a compact hash table. The selection
-// functor returns true for tuples to be hashed in the vector. The fingerprint
-// functor should fingerprint tuples uniquely to an integer that can be used as
-// a vector index. A hash functor is used when hashing tuples into the compact
-// hash table.
-template <class T, class Select, class FP, class H>
-class VectorHashStateTable
-    : public VectorHashBiTable<typename T::StateId, T, Select, FP, H> {
- public:
-  using StateTuple = T;
-  using StateId = typename StateTuple::StateId;
-
-  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::FindId;
-  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::FindEntry;
-  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::Size;
-  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::Selector;
-  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::Fingerprint;
-  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::Hash;
-
-  VectorHashStateTable(Select *select, FP *fingerprint, H *hash,
-                       size_t vector_size = 0, size_t tuple_size = 0)
-      : VectorHashBiTable<StateId, StateTuple, Select, FP, H>(
-            select, fingerprint, hash, vector_size, tuple_size) {}
-
-  StateId FindState(const StateTuple &tuple) { return FindId(tuple); }
-
-  const StateTuple &Tuple(StateId s) const { return FindEntry(s); }
-};
-
-// An implementation using a hash map to map from tuples to state IDs. This
-// version permits erasing of states. The state tuple's default constructor
-// must produce a tuple that will never be seen and the table must suppor
-// operator==.
-template <class T, class H>
-class ErasableStateTable : public ErasableBiTable<typename T::StateId, T, H> {
- public:
-  using StateTuple = T;
-  using StateId = typename StateTuple::StateId;
-
-  using ErasableBiTable<StateId, StateTuple, H>::FindId;
-  using ErasableBiTable<StateId, StateTuple, H>::FindEntry;
-  using ErasableBiTable<StateId, StateTuple, H>::Size;
-  using ErasableBiTable<StateId, StateTuple, H>::Erase;
-
-  ErasableStateTable() : ErasableBiTable<StateId, StateTuple, H>() {}
-
-  StateId FindState(const StateTuple &tuple) { return FindId(tuple); }
-
-  const StateTuple &Tuple(StateId s) const { return FindEntry(s); }
-};
-
-// The composition state table has the form:
-//
-// template <class Arc, class FilterState>
-// class ComposeStateTable {
-//  public:
-//   using StateId = typename Arc::StateId;
-//
-//   // Required constructors.
-//
-//   ComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2);
-//   ComposeStateTable(const ComposeStateTable<Arc, FilterState> &table);
-//
-//   // Looks up a state ID by tuple, adding it if doesn't exist.
-//   StateId FindState(const StateTuple &tuple);
-//
-//   // Looks up a tuple by state ID.
-//   const ComposeStateTuple<StateId> &Tuple(StateId s) const;
-//
-//   // The number of of stored tuples.
-//   StateId Size() const;
-//
-//   // Return true if error was encountered.
-//   bool Error() const;
-// };
-//
-// The following interface is used to represent the composition state.
-//
-// template <class S, class FS>
-// class CompositionStateTuple {
-//  public:
-//   using StateId = typename StateId;
-//   using FS = FilterState;
-//
-//   // Required constructors.
-//   StateTuple();
-//   StateTuple(StateId s1, StateId s2, const FilterState &fs);
-//
-//   StateId StateId1() const;
-//   StateId StateId2() const;
-//
-//   FilterState GetFilterState() const;
-//
-//   std::pair<StateId, StateId> StatePair() const;
-//
-//   size_t Hash() const;
-//
-//   friend bool operator==(const StateTuple& x, const StateTuple &y);
-// }
-//
-template <typename S, typename FS>
-class DefaultComposeStateTuple {
- public:
-  using StateId = S;
-  using FilterState = FS;
-
-  DefaultComposeStateTuple()
-      : state_pair_(kNoStateId, kNoStateId), fs_(FilterState::NoState()) {}
-
-  DefaultComposeStateTuple(StateId s1, StateId s2, const FilterState &fs)
-      : state_pair_(s1, s2), fs_(fs) {}
-
-  StateId StateId1() const { return state_pair_.first; }
-
-  StateId StateId2() const { return state_pair_.second; }
-
-  FilterState GetFilterState() const { return fs_; }
-
-  const std::pair<StateId, StateId> &StatePair() const { return state_pair_; }
-
-  friend bool operator==(const DefaultComposeStateTuple &x,
-                         const DefaultComposeStateTuple &y) {
-    return (&x == &y) || (x.state_pair_ == y.state_pair_ && x.fs_ == y.fs_);
-  }
-
-  size_t Hash() const {
-    return static_cast<size_t>(StateId1()) +
-           static_cast<size_t>(StateId2()) * 7853u +
-           GetFilterState().Hash() * 7867u;
-  }
-
- private:
-  std::pair<StateId, StateId> state_pair_;
-  FilterState fs_;  // State of composition filter.
-};
-
-// Specialization for TrivialFilterState that does not explicitely store the
-// filter state since it is always the unique non-blocking state.
-template <typename S>
-class DefaultComposeStateTuple<S, TrivialFilterState> {
- public:
-  using StateId = S;
-  using FilterState = TrivialFilterState;
-
-  DefaultComposeStateTuple()
-      : state_pair_(kNoStateId, kNoStateId) {}
-
-  DefaultComposeStateTuple(StateId s1, StateId s2, const FilterState &)
-      : state_pair_(s1, s2) {}
-
-  StateId StateId1() const { return state_pair_.first; }
-
-  StateId StateId2() const { return state_pair_.second; }
-
-  FilterState GetFilterState() const { return FilterState(true); }
-
-  const std::pair<StateId, StateId> &StatePair() const { return state_pair_; }
-
-  friend bool operator==(const DefaultComposeStateTuple &x,
-                         const DefaultComposeStateTuple &y) {
-    return (&x == &y) || (x.state_pair_ == y.state_pair_);
-  }
-
-  size_t Hash() const { return StateId1() + StateId2() * 7853; }
-
- private:
-  std::pair<StateId, StateId> state_pair_;
-};
-
-// Hashing of composition state tuples.
-template <typename T>
-class ComposeHash {
- public:
-  size_t operator()(const T &t) const { return t.Hash(); }
-};
-
-// A HashStateTable over composition tuples.
-template <typename Arc, typename FilterState,
-          typename StateTuple =
-              DefaultComposeStateTuple<typename Arc::StateId, FilterState>,
-          typename StateTable =
-              CompactHashStateTable<StateTuple, ComposeHash<StateTuple>>>
-class GenericComposeStateTable : public StateTable {
- public:
-  using StateId = typename Arc::StateId;
-
-  GenericComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {}
-
-  GenericComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-                           size_t table_size)
-      : StateTable(table_size) {}
-
-  constexpr bool Error() const { return false; }
-
- private:
-  GenericComposeStateTable &operator=(const GenericComposeStateTable &table) =
-      delete;
-};
-
-//  Fingerprint for general composition tuples.
-template <typename StateTuple>
-class ComposeFingerprint {
- public:
-  using StateId = typename StateTuple::StateId;
-
-  // Required but suboptimal constructor.
-  ComposeFingerprint() : mult1_(8192), mult2_(8192) {
-    LOG(WARNING) << "TupleFingerprint: # of FST states should be provided.";
-  }
-
-  // Constructor is provided the sizes of the input FSTs.
-  ComposeFingerprint(StateId nstates1, StateId nstates2)
-      : mult1_(nstates1), mult2_(nstates1 * nstates2) {}
-
-  size_t operator()(const StateTuple &tuple) {
-    return tuple.StateId1() + tuple.StateId2() * mult1_ +
-           tuple.GetFilterState().Hash() * mult2_;
-  }
-
- private:
-  const ssize_t mult1_;
-  const ssize_t mult2_;
-};
-
-// Useful when the first composition state determines the tuple.
-template <typename StateTuple>
-class ComposeState1Fingerprint {
- public:
-  size_t operator()(const StateTuple &tuple) { return tuple.StateId1(); }
-};
-
-// Useful when the second composition state determines the tuple.
-template <typename StateTuple>
-class ComposeState2Fingerprint {
- public:
-  size_t operator()(const StateTuple &tuple) { return tuple.StateId2(); }
-};
-
-// A VectorStateTable over composition tuples. This can be used when the
-// product of number of states in FST1 and FST2 (and the composition filter
-// state hash) is manageable. If the FSTs are not expanded FSTs, they will
-// first have their states counted.
-template <typename Arc, typename StateTuple>
-class ProductComposeStateTable
-    : public VectorStateTable<StateTuple, ComposeFingerprint<StateTuple>> {
- public:
-  using StateId = typename Arc::StateId;
-  using StateTable =
-      VectorStateTable<StateTuple, ComposeFingerprint<StateTuple>>;
-
-  ProductComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-                           size_t table_size = 0)
-      : StateTable(new ComposeFingerprint<StateTuple>(CountStates(fst1),
-                                                      CountStates(fst2)),
-                   table_size) {}
-
-  ProductComposeStateTable(
-      const ProductComposeStateTable<Arc, StateTuple> &table)
-      : StateTable(new ComposeFingerprint<StateTuple>(table.Fingerprint())) {}
-
-  constexpr bool Error() const { return false; }
-
- private:
-  ProductComposeStateTable &operator=(const ProductComposeStateTable &table) =
-      delete;
-};
-
-// A vector-backed table over composition tuples which can be used when the
-// first FST is a string (i.e., satisfies kString property) and the second is
-// deterministic and epsilon-free. It should be used with a composition filter
-// that creates at most one filter state per tuple under these conditions (e.g.,
-// SequenceComposeFilter or MatchComposeFilter).
-template <typename Arc, typename StateTuple>
-class StringDetComposeStateTable
-    : public VectorStateTable<StateTuple,
-                              ComposeState1Fingerprint<StateTuple>> {
- public:
-  using StateId = typename Arc::StateId;
-  using StateTable =
-      VectorStateTable<StateTuple, ComposeState1Fingerprint<StateTuple>>;
-
-  StringDetComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2)
-      : error_(false) {
-    static constexpr auto props2 = kIDeterministic | kNoIEpsilons;
-    if (fst1.Properties(kString, true) != kString) {
-      FSTERROR() << "StringDetComposeStateTable: 1st FST is not a string";
-      error_ = true;
-    } else if (fst2.Properties(props2, true) != props2) {
-      FSTERROR() << "StringDetComposeStateTable: 2nd FST is not deterministic "
-                    "and epsilon-free";
-      error_ = true;
-    }
-  }
-
-  StringDetComposeStateTable(
-      const StringDetComposeStateTable<Arc, StateTuple> &table)
-      : StateTable(table), error_(table.error_) {}
-
-  bool Error() const { return error_; }
-
- private:
-  bool error_;
-
-  StringDetComposeStateTable &operator=(const StringDetComposeStateTable &) =
-      delete;
-};
-
-// A vector-backed table over composition tuples which can be used when the
-// first FST is deterministic and epsilon-free and the second is a string (i.e.,
-// satisfies kString). It should be used with a composition filter that creates
-// at most one filter state per tuple under these conditions (e.g.,
-// SequenceComposeFilter or MatchComposeFilter).
-template <typename Arc, typename StateTuple>
-class DetStringComposeStateTable
-    : public VectorStateTable<StateTuple,
-                              ComposeState2Fingerprint<StateTuple>> {
- public:
-  using StateId = typename Arc::StateId;
-  using StateTable =
-      VectorStateTable<StateTuple, ComposeState2Fingerprint<StateTuple>>;
-
-  DetStringComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2)
-      : error_(false) {
-    static constexpr auto props = kODeterministic | kNoOEpsilons;
-    if (fst1.Properties(props, true) != props) {
-      FSTERROR() << "StringDetComposeStateTable: 1st FST is not "
-                 << "input-deterministic and epsilon-free";
-      error_ = true;
-    } else if (fst2.Properties(kString, true) != kString) {
-      FSTERROR() << "DetStringComposeStateTable: 2nd FST is not a string";
-      error_ = true;
-    }
-  }
-
-  DetStringComposeStateTable(
-      const DetStringComposeStateTable<Arc, StateTuple> &table)
-      : StateTable(table), error_(table.error_) {}
-
-  bool Error() const { return error_; }
-
- private:
-  bool error_;
-
-  DetStringComposeStateTable &operator=(const DetStringComposeStateTable &) =
-      delete;
-};
-
-// An erasable table over composition tuples. The Erase(StateId) method can be
-// called if the user either is sure that composition will never return to that
-// tuple or doesn't care that if it does, it is assigned a new state ID.
-template <typename Arc, typename StateTuple>
-class ErasableComposeStateTable
-    : public ErasableStateTable<StateTuple, ComposeHash<StateTuple>> {
- public:
-  ErasableComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {}
-
-  constexpr bool Error() const { return false; }
-
- private:
-  ErasableComposeStateTable &operator=(const ErasableComposeStateTable &table) =
-      delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_STATE_TABLE_H_
diff --git a/projects/llm_framework/include/fst/statesort.h b/projects/llm_framework/include/fst/statesort.h
deleted file mode 100644
index 346c7d32..00000000
--- a/projects/llm_framework/include/fst/statesort.h
+++ /dev/null
@@ -1,74 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function to sort states of an FST.
-
-#ifndef FST_STATESORT_H_
-#define FST_STATESORT_H_
-
-#include <algorithm>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-// Sorts the input states of an FST. order[i] gives the the state ID after
-// sorting that corresponds to the state ID i before sorting; it must
-// therefore be a permutation of the input FST's states ID sequence.
-template <class Arc>
-void StateSort(MutableFst<Arc> *fst,
-               const std::vector<typename Arc::StateId> &order) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  if (order.size() != fst->NumStates()) {
-    FSTERROR() << "StateSort: Bad order vector size: " << order.size();
-    fst->SetProperties(kError, kError);
-    return;
-  }
-  if (fst->Start() == kNoStateId) return;
-  const auto props = fst->Properties(kStateSortProperties, false);
-  std::vector<bool> done(order.size(), false);
-  std::vector<Arc> arcsa;
-  std::vector<Arc> arcsb;
-  fst->SetStart(order[fst->Start()]);
-  for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
-       siter.Next()) {
-    auto s1 = siter.Value();
-    StateId s2;
-    if (done[s1]) continue;
-    auto final1 = fst->Final(s1);
-    auto final2 = Weight::Zero();
-    arcsa.clear();
-    for (ArcIterator<MutableFst<Arc>> aiter(*fst, s1); !aiter.Done();
-         aiter.Next()) {
-      arcsa.push_back(aiter.Value());
-    }
-    for (; !done[s1]; s1 = s2, final1 = final2, std::swap(arcsa, arcsb)) {
-      s2 = order[s1];
-      if (!done[s2]) {
-        final2 = fst->Final(s2);
-        arcsb.clear();
-        for (ArcIterator<MutableFst<Arc>> aiter(*fst, s2); !aiter.Done();
-             aiter.Next()) {
-          arcsb.push_back(aiter.Value());
-        }
-      }
-      fst->SetFinal(s2, final1);
-      fst->DeleteArcs(s2);
-      for (auto arc : arcsa) {  // Copy intended.
-        arc.nextstate = order[arc.nextstate];
-        fst->AddArc(s2, arc);
-      }
-      done[s1] = true;
-    }
-  }
-  fst->SetProperties(props, kFstProperties);
-}
-
-}  // namespace fst
-
-#endif  // FST_STATESORT_H_
diff --git a/projects/llm_framework/include/fst/string-weight.h b/projects/llm_framework/include/fst/string-weight.h
deleted file mode 100644
index fb3e70f1..00000000
--- a/projects/llm_framework/include/fst/string-weight.h
+++ /dev/null
@@ -1,807 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// String weight set and associated semiring operation definitions.
-
-#ifndef FST_STRING_WEIGHT_H_
-#define FST_STRING_WEIGHT_H_
-
-#include <cstdlib>
-
-#include <list>
-#include <string>
-#include <vector>
-
-#include <fst/product-weight.h>
-#include <fst/union-weight.h>
-#include <fst/weight.h>
-
-
-namespace fst {
-
-constexpr int kStringInfinity = -1;     // Label for the infinite string.
-constexpr int kStringBad = -2;          // Label for a non-string.
-constexpr char kStringSeparator = '_';  // Label separator in strings.
-
-// Determines whether to use left or right string semiring.  Includes a
-// 'restricted' version that signals an error if proper prefixes/suffixes
-// would otherwise be returned by Plus, useful with various
-// algorithms that require functional transducer input with the
-// string semirings.
-enum StringType { STRING_LEFT = 0, STRING_RIGHT = 1, STRING_RESTRICT = 2 };
-
-constexpr StringType ReverseStringType(StringType s) {
-  return s == STRING_LEFT ? STRING_RIGHT
-                          : (s == STRING_RIGHT ? STRING_LEFT : STRING_RESTRICT);
-}
-
-template <class>
-class StringWeightIterator;
-template <class>
-class StringWeightReverseIterator;
-
-// String semiring: (longest_common_prefix/suffix, ., Infinity, Epsilon)
-template <typename Label_, StringType S = STRING_LEFT>
-class StringWeight {
- public:
-  using Label = Label_;
-  using ReverseWeight = StringWeight<Label, ReverseStringType(S)>;
-  using Iterator = StringWeightIterator<StringWeight>;
-  using ReverseIterator = StringWeightReverseIterator<StringWeight>;
-
-  friend class StringWeightIterator<StringWeight>;
-  friend class StringWeightReverseIterator<StringWeight>;
-
-  StringWeight() {}
-
-  template <typename Iterator>
-  StringWeight(const Iterator &begin, const Iterator &end) {
-    for (auto iter = begin; iter != end; ++iter) PushBack(*iter);
-  }
-
-  explicit StringWeight(Label label) { PushBack(label); }
-
-  static const StringWeight &Zero() {
-    static const auto *const zero = new StringWeight(Label(kStringInfinity));
-    return *zero;
-  }
-
-  static const StringWeight &One() {
-    static const auto *const one = new StringWeight();
-    return *one;
-  }
-
-  static const StringWeight &NoWeight() {
-    static const auto *const no_weight = new StringWeight(Label(kStringBad));
-    return *no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string(
-        S == STRING_LEFT
-            ? "left_string"
-            : (S == STRING_RIGHT ? "right_string" : "restricted_string"));
-    return *type;
-  }
-
-  bool Member() const;
-
-  std::istream &Read(std::istream &strm);
-
-  std::ostream &Write(std::ostream &strm) const;
-
-  size_t Hash() const;
-
-  StringWeight Quantize(float delta = kDelta) const { return *this; }
-
-  ReverseWeight Reverse() const;
-
-  static constexpr uint64 Properties() {
-    return kIdempotent |
-           (S == STRING_LEFT ? kLeftSemiring
-                             : (S == STRING_RIGHT
-                                    ? kRightSemiring
-                                    : /* S == STRING_RESTRICT */ kLeftSemiring |
-                                          kRightSemiring));
-  }
-
-  // These operations combined with the StringWeightIterator and
-  // StringWeightReverseIterator provide the access and mutation of the string
-  // internal elements.
-
-  // Clear existing StringWeight.
-  void Clear() {
-    first_ = 0;
-    rest_.clear();
-  }
-
-  size_t Size() const { return first_ ? rest_.size() + 1 : 0; }
-
-  void PushFront(Label label) {
-    if (first_) rest_.push_front(first_);
-    first_ = label;
-  }
-
-  void PushBack(Label label) {
-    if (!first_) {
-      first_ = label;
-    } else {
-      rest_.push_back(label);
-    }
-  }
-
- private:
-  Label first_ = 0;        // First label in string (0 if empty).
-  std::list<Label> rest_;  // Remaining labels in string.
-};
-
-// Traverses string in forward direction.
-template <class StringWeight_>
-class StringWeightIterator {
- public:
-  using Weight = StringWeight_;
-  using Label = typename Weight::Label;
-
-  explicit StringWeightIterator(const Weight &w)
-      : first_(w.first_), rest_(w.rest_), init_(true), iter_(rest_.begin()) {}
-
-  bool Done() const {
-    if (init_) {
-      return first_ == 0;
-    } else {
-      return iter_ == rest_.end();
-    }
-  }
-
-  const Label &Value() const { return init_ ? first_ : *iter_; }
-
-  void Next() {
-    if (init_) {
-      init_ = false;
-    } else {
-      ++iter_;
-    }
-  }
-
-  void Reset() {
-    init_ = true;
-    iter_ = rest_.begin();
-  }
-
- private:
-  const Label &first_;
-  const decltype(Weight::rest_) &rest_;
-  bool init_;  // In the initialized state?
-  typename std::remove_reference<decltype (Weight::rest_)>::type::const_iterator iter_;
-};
-
-// Traverses string in backward direction.
-template <class StringWeight_>
-class StringWeightReverseIterator {
- public:
-  using Weight = StringWeight_;
-  using Label = typename Weight::Label;
-
-  explicit StringWeightReverseIterator(const Weight &w)
-      : first_(w.first_),
-        rest_(w.rest_),
-        fin_(first_ == Label()),
-        iter_(rest_.rbegin()) {}
-
-  bool Done() const { return fin_; }
-
-  const Label &Value() const { return iter_ == rest_.rend() ? first_ : *iter_; }
-
-  void Next() {
-    if (iter_ == rest_.rend()) {
-      fin_ = true;
-    } else {
-      ++iter_;
-    }
-  }
-
-  void Reset() {
-    fin_ = false;
-    iter_ = rest_.rbegin();
-  }
-
- private:
-  const Label &first_;
-  const decltype(Weight::rest_) &rest_;
-  bool fin_;  // In the final state?
-  typename std::remove_reference<decltype (Weight::rest_)>::type::const_reverse_iterator iter_;
-};
-
-// StringWeight member functions follow that require
-// StringWeightIterator or StringWeightReverseIterator.
-
-template <typename Label, StringType S>
-inline std::istream &StringWeight<Label, S>::Read(std::istream &strm) {
-  Clear();
-  int32 size;
-  ReadType(strm, &size);
-  for (int32 i = 0; i < size; ++i) {
-    Label label;
-    ReadType(strm, &label);
-    PushBack(label);
-  }
-  return strm;
-}
-
-template <typename Label, StringType S>
-inline std::ostream &StringWeight<Label, S>::Write(std::ostream &strm) const {
-  const int32 size = Size();
-  WriteType(strm, size);
-  for (Iterator iter(*this); !iter.Done(); iter.Next()) {
-    WriteType(strm, iter.Value());
-  }
-  return strm;
-}
-
-template <typename Label, StringType S>
-inline bool StringWeight<Label, S>::Member() const {
-  Iterator iter(*this);
-  return iter.Value() != Label(kStringBad);
-}
-
-template <typename Label, StringType S>
-inline typename StringWeight<Label, S>::ReverseWeight
-StringWeight<Label, S>::Reverse() const {
-  ReverseWeight rweight;
-  for (Iterator iter(*this); !iter.Done(); iter.Next()) {
-    rweight.PushFront(iter.Value());
-  }
-  return rweight;
-}
-
-template <typename Label, StringType S>
-inline size_t StringWeight<Label, S>::Hash() const {
-  size_t h = 0;
-  for (Iterator iter(*this); !iter.Done(); iter.Next()) {
-    h ^= h << 1 ^ iter.Value();
-  }
-  return h;
-}
-
-template <typename Label, StringType S>
-inline bool operator==(const StringWeight<Label, S> &w1,
-                       const StringWeight<Label, S> &w2) {
-  if (w1.Size() != w2.Size()) return false;
-  using Iterator = typename StringWeight<Label, S>::Iterator;
-  Iterator iter1(w1);
-  Iterator iter2(w2);
-  for (; !iter1.Done(); iter1.Next(), iter2.Next()) {
-    if (iter1.Value() != iter2.Value()) return false;
-  }
-  return true;
-}
-
-template <typename Label, StringType S>
-inline bool operator!=(const StringWeight<Label, S> &w1,
-                       const StringWeight<Label, S> &w2) {
-  return !(w1 == w2);
-}
-
-template <typename Label, StringType S>
-inline bool ApproxEqual(const StringWeight<Label, S> &w1,
-                        const StringWeight<Label, S> &w2,
-                        float delta = kDelta) {
-  return w1 == w2;
-}
-
-template <typename Label, StringType S>
-inline std::ostream &operator<<(std::ostream &strm,
-                                const StringWeight<Label, S> &weight) {
-  typename StringWeight<Label, S>::Iterator iter(weight);
-  if (iter.Done()) {
-    return strm << "Epsilon";
-  } else if (iter.Value() == Label(kStringInfinity)) {
-    return strm << "Infinity";
-  } else if (iter.Value() == Label(kStringBad)) {
-    return strm << "BadString";
-  } else {
-    for (size_t i = 0; !iter.Done(); ++i, iter.Next()) {
-      if (i > 0) strm << kStringSeparator;
-      strm << iter.Value();
-    }
-  }
-  return strm;
-}
-
-template <typename Label, StringType S>
-inline std::istream &operator>>(std::istream &strm,
-                                StringWeight<Label, S> &weight) {
-  string str;
-  strm >> str;
-  using Weight = StringWeight<Label, S>;
-  if (str == "Infinity") {
-    weight = Weight::Zero();
-  } else if (str == "Epsilon") {
-    weight = Weight::One();
-  } else {
-    weight.Clear();
-    char *p = nullptr;
-    for (const char *cs = str.c_str(); !p || *p != '\0'; cs = p + 1) {
-      const Label label = strtoll(cs, &p, 10);
-      if (p == cs || (*p != 0 && *p != kStringSeparator)) {
-        strm.clear(std::ios::badbit);
-        break;
-      }
-      weight.PushBack(label);
-    }
-  }
-  return strm;
-}
-
-// Default is for the restricted string semiring. String equality is required
-// (for non-Zero() input). The restriction is used (e.g., in determinization)
-// to ensure the input is functional.
-template <typename Label, StringType S>
-inline StringWeight<Label, S> Plus(const StringWeight<Label, S> &w1,
-                                   const StringWeight<Label, S> &w2) {
-  using Weight = StringWeight<Label, S>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::Zero()) return w2;
-  if (w2 == Weight::Zero()) return w1;
-  if (w1 != w2) {
-    FSTERROR() << "StringWeight::Plus: Unequal arguments "
-               << "(non-functional FST?)"
-               << " w1 = " << w1 << " w2 = " << w2;
-    return Weight::NoWeight();
-  }
-  return w1;
-}
-
-// Longest common prefix for left string semiring.
-template <typename Label>
-inline StringWeight<Label, STRING_LEFT> Plus(
-    const StringWeight<Label, STRING_LEFT> &w1,
-    const StringWeight<Label, STRING_LEFT> &w2) {
-  using Weight = StringWeight<Label, STRING_LEFT>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::Zero()) return w2;
-  if (w2 == Weight::Zero()) return w1;
-  Weight sum;
-  typename Weight::Iterator iter1(w1);
-  typename Weight::Iterator iter2(w2);
-  for (; !iter1.Done() && !iter2.Done() && iter1.Value() == iter2.Value();
-       iter1.Next(), iter2.Next()) {
-    sum.PushBack(iter1.Value());
-  }
-  return sum;
-}
-
-// Longest common suffix for right string semiring.
-template <typename Label>
-inline StringWeight<Label, STRING_RIGHT> Plus(
-    const StringWeight<Label, STRING_RIGHT> &w1,
-    const StringWeight<Label, STRING_RIGHT> &w2) {
-  using Weight = StringWeight<Label, STRING_RIGHT>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::Zero()) return w2;
-  if (w2 == Weight::Zero()) return w1;
-  Weight sum;
-  typename Weight::ReverseIterator iter1(w1);
-  typename Weight::ReverseIterator iter2(w2);
-  for (; !iter1.Done() && !iter2.Done() && iter1.Value() == iter2.Value();
-       iter1.Next(), iter2.Next()) {
-    sum.PushFront(iter1.Value());
-  }
-  return sum;
-}
-
-template <typename Label, StringType S>
-inline StringWeight<Label, S> Times(const StringWeight<Label, S> &w1,
-                                    const StringWeight<Label, S> &w2) {
-  using Weight = StringWeight<Label, S>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::Zero() || w2 == Weight::Zero()) return Weight::Zero();
-  Weight product(w1);
-  for (typename Weight::Iterator iter(w2); !iter.Done(); iter.Next()) {
-    product.PushBack(iter.Value());
-  }
-  return product;
-}
-
-// Left division in a left string semiring.
-template <typename Label, StringType S>
-inline StringWeight<Label, S> DivideLeft(const StringWeight<Label, S> &w1,
-                                         const StringWeight<Label, S> &w2) {
-  using Weight = StringWeight<Label, S>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w2 == Weight::Zero()) {
-    return Weight(Label(kStringBad));
-  } else if (w1 == Weight::Zero()) {
-    return Weight::Zero();
-  }
-  Weight result;
-  typename Weight::Iterator iter(w1);
-  size_t i = 0;
-  for (; !iter.Done() && i < w2.Size(); iter.Next(), ++i) {
-  }
-  for (; !iter.Done(); iter.Next()) result.PushBack(iter.Value());
-  return result;
-}
-
-// Right division in a right string semiring.
-template <typename Label, StringType S>
-inline StringWeight<Label, S> DivideRight(const StringWeight<Label, S> &w1,
-                                          const StringWeight<Label, S> &w2) {
-  using Weight = StringWeight<Label, S>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w2 == Weight::Zero()) {
-    return Weight(Label(kStringBad));
-  } else if (w1 == Weight::Zero()) {
-    return Weight::Zero();
-  }
-  Weight result;
-  typename Weight::ReverseIterator iter(w1);
-  size_t i = 0;
-  for (; !iter.Done() && i < w2.Size(); iter.Next(), ++i) {
-  }
-  for (; !iter.Done(); iter.Next()) result.PushFront(iter.Value());
-  return result;
-}
-
-// Default is the restricted string semiring.
-template <typename Label, StringType S>
-inline StringWeight<Label, S> Divide(const StringWeight<Label, S> &w1,
-                                     const StringWeight<Label, S> &w2,
-                                     DivideType divide_type) {
-  using Weight = StringWeight<Label, S>;
-  if (divide_type == DIVIDE_LEFT) {
-    return DivideLeft(w1, w2);
-  } else if (divide_type == DIVIDE_RIGHT) {
-    return DivideRight(w1, w2);
-  } else {
-    FSTERROR() << "StringWeight::Divide: "
-               << "Only explicit left or right division is defined "
-               << "for the " << Weight::Type() << " semiring";
-    return Weight::NoWeight();
-  }
-}
-
-// Left division in the left string semiring.
-template <typename Label>
-inline StringWeight<Label, STRING_LEFT> Divide(
-    const StringWeight<Label, STRING_LEFT> &w1,
-    const StringWeight<Label, STRING_LEFT> &w2, DivideType divide_type) {
-  if (divide_type != DIVIDE_LEFT) {
-    FSTERROR() << "StringWeight::Divide: Only left division is defined "
-               << "for the left string semiring";
-    return StringWeight<Label, STRING_LEFT>::NoWeight();
-  }
-  return DivideLeft(w1, w2);
-}
-
-// Right division in the right string semiring.
-template <typename Label>
-inline StringWeight<Label, STRING_RIGHT> Divide(
-    const StringWeight<Label, STRING_RIGHT> &w1,
-    const StringWeight<Label, STRING_RIGHT> &w2, DivideType divide_type) {
-  if (divide_type != DIVIDE_RIGHT) {
-    FSTERROR() << "StringWeight::Divide: Only right division is defined "
-               << "for the right string semiring";
-    return StringWeight<Label, STRING_RIGHT>::NoWeight();
-  }
-  return DivideRight(w1, w2);
-}
-
-// This function object generates StringWeights that are random integer strings
-// from {1, ... , alphabet_size)^{0, max_string_length} U { Zero }. This is
-// intended primarily for testing.
-template <class Label, StringType S>
-class WeightGenerate<StringWeight<Label, S>> {
- public:
-  using Weight = StringWeight<Label, S>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t alphabet_size = kNumRandomWeights,
-                          size_t max_string_length = kNumRandomWeights)
-      : allow_zero_(allow_zero),
-        alphabet_size_(alphabet_size),
-        max_string_length_(max_string_length) {}
-
-  Weight operator()() const {
-    size_t n = rand() % (max_string_length_ + allow_zero_);  // NOLINT
-    if (allow_zero_ && n == max_string_length_) return Weight::Zero();
-    std::vector<Label> labels;
-    labels.reserve(n);
-    for (size_t i = 0; i < n; ++i) {
-      labels.push_back(rand() % alphabet_size_ + 1);  // NOLINT
-    }
-    return Weight(labels.begin(), labels.end());
-  }
-
- private:
-  // Permits Zero() and zero divisors.
-  const bool allow_zero_;
-  // Alphabet size for random weights.
-  const size_t alphabet_size_;
-  // Number of alternative random weights.
-  const size_t max_string_length_;
-};
-
-// Determines whether to use left, right, or (general) gallic semiring. Includes
-// a restricted version that signals an error if proper string prefixes or
-// suffixes would otherwise be returned by string Plus. This is useful with
-// algorithms that require functional transducer input. Also includes min
-// version that changes the Plus to keep only the lowest W weight string.
-enum GallicType {
-  GALLIC_LEFT = 0,
-  GALLIC_RIGHT = 1,
-  GALLIC_RESTRICT = 2,
-  GALLIC_MIN = 3,
-  GALLIC = 4
-};
-
-constexpr StringType GallicStringType(GallicType g) {
-  return g == GALLIC_LEFT
-             ? STRING_LEFT
-             : (g == GALLIC_RIGHT ? STRING_RIGHT : STRING_RESTRICT);
-}
-
-constexpr GallicType ReverseGallicType(GallicType g) {
-  return g == GALLIC_LEFT
-             ? GALLIC_RIGHT
-             : (g == GALLIC_RIGHT
-                    ? GALLIC_LEFT
-                    : (g == GALLIC_RESTRICT
-                           ? GALLIC_RESTRICT
-                           : (g == GALLIC_MIN ? GALLIC_MIN : GALLIC)));
-}
-
-// Product of string weight and an arbitraryy weight.
-template <class Label, class W, GallicType G = GALLIC_LEFT>
-struct GallicWeight
-    : public ProductWeight<StringWeight<Label, GallicStringType(G)>, W> {
-  using ReverseWeight =
-      GallicWeight<Label, typename W::ReverseWeight, ReverseGallicType(G)>;
-  using SW = StringWeight<Label, GallicStringType(G)>;
-
-  using ProductWeight<SW, W>::Properties;
-
-  GallicWeight() {}
-
-  GallicWeight(SW w1, W w2) : ProductWeight<SW, W>(w1, w2) {}
-
-  explicit GallicWeight(const string &s, int *nread = nullptr)
-      : ProductWeight<SW, W>(s, nread) {}
-
-  explicit GallicWeight(const ProductWeight<SW, W> &w)
-      : ProductWeight<SW, W>(w) {}
-
-  static const GallicWeight &Zero() {
-    static const GallicWeight zero(ProductWeight<SW, W>::Zero());
-    return zero;
-  }
-
-  static const GallicWeight &One() {
-    static const GallicWeight one(ProductWeight<SW, W>::One());
-    return one;
-  }
-
-  static const GallicWeight &NoWeight() {
-    static const GallicWeight no_weight(ProductWeight<SW, W>::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string(
-        G == GALLIC_LEFT
-            ? "left_gallic"
-            : (G == GALLIC_RIGHT
-                   ? "right_gallic"
-                   : (G == GALLIC_RESTRICT
-                          ? "restricted_gallic"
-                          : (G == GALLIC_MIN ? "min_gallic" : "gallic"))));
-    return *type;
-  }
-
-  GallicWeight Quantize(float delta = kDelta) const {
-    return GallicWeight(ProductWeight<SW, W>::Quantize(delta));
-  }
-
-  ReverseWeight Reverse() const {
-    return ReverseWeight(ProductWeight<SW, W>::Reverse());
-  }
-};
-
-// Default plus.
-template <class Label, class W, GallicType G>
-inline GallicWeight<Label, W, G> Plus(const GallicWeight<Label, W, G> &w,
-                                      const GallicWeight<Label, W, G> &v) {
-  return GallicWeight<Label, W, G>(Plus(w.Value1(), v.Value1()),
-                                   Plus(w.Value2(), v.Value2()));
-}
-
-// Min gallic plus.
-template <class Label, class W>
-inline GallicWeight<Label, W, GALLIC_MIN> Plus(
-    const GallicWeight<Label, W, GALLIC_MIN> &w1,
-    const GallicWeight<Label, W, GALLIC_MIN> &w2) {
-  static const NaturalLess<W> less;
-  return less(w1.Value2(), w2.Value2()) ? w1 : w2;
-}
-
-template <class Label, class W, GallicType G>
-inline GallicWeight<Label, W, G> Times(const GallicWeight<Label, W, G> &w,
-                                       const GallicWeight<Label, W, G> &v) {
-  return GallicWeight<Label, W, G>(Times(w.Value1(), v.Value1()),
-                                   Times(w.Value2(), v.Value2()));
-}
-
-template <class Label, class W, GallicType G>
-inline GallicWeight<Label, W, G> Divide(const GallicWeight<Label, W, G> &w,
-                                        const GallicWeight<Label, W, G> &v,
-                                        DivideType divide_type = DIVIDE_ANY) {
-  return GallicWeight<Label, W, G>(Divide(w.Value1(), v.Value1(), divide_type),
-                                   Divide(w.Value2(), v.Value2(), divide_type));
-}
-
-// This function object generates gallic weights by calling an underlying
-// product weight generator. This is intended primarily for testing.
-template <class Label, class W, GallicType G>
-class WeightGenerate<GallicWeight<Label, W, G>>
-    : public WeightGenerate<
-          ProductWeight<StringWeight<Label, GallicStringType(G)>, W>> {
- public:
-  using Weight = GallicWeight<Label, W, G>;
-  using Generate = WeightGenerate<
-      ProductWeight<StringWeight<Label, GallicStringType(G)>, W>>;
-
-  explicit WeightGenerate(bool allow_zero = true) : generate_(allow_zero) {}
-
-  Weight operator()() const { return Weight(generate_()); }
-
- private:
-  const Generate generate_;
-};
-
-// Union weight options for (general) GALLIC type.
-template <class Label, class W>
-struct GallicUnionWeightOptions {
-  using ReverseOptions = GallicUnionWeightOptions<Label, W>;
-  using GW = GallicWeight<Label, W, GALLIC_RESTRICT>;
-  using SW = StringWeight<Label, GallicStringType(GALLIC_RESTRICT)>;
-  using SI = StringWeightIterator<SW>;
-
-  // Military order.
-  struct Compare {
-    bool operator()(const GW &w1, const GW &w2) const {
-      const SW &s1 = w1.Value1();
-      const SW &s2 = w2.Value1();
-      if (s1.Size() < s2.Size()) return true;
-      if (s1.Size() > s2.Size()) return false;
-      SI iter1(s1);
-      SI iter2(s2);
-      while (!iter1.Done()) {
-        const auto l1 = iter1.Value();
-        const auto l2 = iter2.Value();
-        if (l1 < l2) return true;
-        if (l1 > l2) return false;
-        iter1.Next();
-        iter2.Next();
-      }
-      return false;
-    }
-  };
-
-  // Adds W weights when string part equal.
-  struct Merge {
-    GW operator()(const GW &w1, const GW &w2) const {
-      return GW(w1.Value1(), Plus(w1.Value2(), w2.Value2()));
-    }
-  };
-};
-
-// Specialization for the (general) GALLIC type.
-template <class Label, class W>
-struct GallicWeight<Label, W, GALLIC>
-    : public UnionWeight<GallicWeight<Label, W, GALLIC_RESTRICT>,
-                         GallicUnionWeightOptions<Label, W>> {
-  using GW = GallicWeight<Label, W, GALLIC_RESTRICT>;
-  using SW = StringWeight<Label, GallicStringType(GALLIC_RESTRICT)>;
-  using SI = StringWeightIterator<SW>;
-  using UW = UnionWeight<GW, GallicUnionWeightOptions<Label, W>>;
-  using UI = UnionWeightIterator<GW, GallicUnionWeightOptions<Label, W>>;
-  using ReverseWeight = GallicWeight<Label, W, GALLIC>;
-
-  using UW::Properties;
-
-  GallicWeight() {}
-
-  // Copy constructor.
-  GallicWeight(const UW &weight) : UW(weight) {}  // NOLINT
-
-  // Singleton constructors: create a GALLIC weight containing a single
-  // GALLIC_RESTRICT weight. Takes as argument (1) a GALLIC_RESTRICT weight or
-  // (2) the two components of a GALLIC_RESTRICT weight.
-  explicit GallicWeight(const GW &weight) : UW(weight) {}
-
-  GallicWeight(SW w1, W w2) : UW(GW(w1, w2)) {}
-
-  explicit GallicWeight(const string &str, int *nread = nullptr)
-      : UW(str, nread) {}
-
-  static const GallicWeight<Label, W, GALLIC> &Zero() {
-    static const GallicWeight<Label, W, GALLIC> zero(UW::Zero());
-    return zero;
-  }
-
-  static const GallicWeight<Label, W, GALLIC> &One() {
-    static const GallicWeight<Label, W, GALLIC> one(UW::One());
-    return one;
-  }
-
-  static const GallicWeight<Label, W, GALLIC> &NoWeight() {
-    static const GallicWeight<Label, W, GALLIC> no_weight(UW::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string("gallic");
-    return *type;
-  }
-
-  GallicWeight<Label, W, GALLIC> Quantize(float delta = kDelta) const {
-    return UW::Quantize(delta);
-  }
-
-  ReverseWeight Reverse() const { return UW::Reverse(); }
-};
-
-// (General) gallic plus.
-template <class Label, class W>
-inline GallicWeight<Label, W, GALLIC> Plus(
-    const GallicWeight<Label, W, GALLIC> &w1,
-    const GallicWeight<Label, W, GALLIC> &w2) {
-  using GW = GallicWeight<Label, W, GALLIC_RESTRICT>;
-  using UW = UnionWeight<GW, GallicUnionWeightOptions<Label, W>>;
-  return Plus(static_cast<UW>(w1), static_cast<UW>(w2));
-}
-
-// (General) gallic times.
-template <class Label, class W>
-inline GallicWeight<Label, W, GALLIC> Times(
-    const GallicWeight<Label, W, GALLIC> &w1,
-    const GallicWeight<Label, W, GALLIC> &w2) {
-  using GW = GallicWeight<Label, W, GALLIC_RESTRICT>;
-  using UW = UnionWeight<GW, GallicUnionWeightOptions<Label, W>>;
-  return Times(static_cast<UW>(w1), static_cast<UW>(w2));
-}
-
-// (General) gallic divide.
-template <class Label, class W>
-inline GallicWeight<Label, W, GALLIC> Divide(
-    const GallicWeight<Label, W, GALLIC> &w1,
-    const GallicWeight<Label, W, GALLIC> &w2,
-    DivideType divide_type = DIVIDE_ANY) {
-  using GW = GallicWeight<Label, W, GALLIC_RESTRICT>;
-  using UW = UnionWeight<GW, GallicUnionWeightOptions<Label, W>>;
-  return Divide(static_cast<UW>(w1), static_cast<UW>(w2), divide_type);
-}
-
-// This function object generates gallic weights by calling an underlying
-// union weight generator. This is intended primarily for testing.
-template <class Label, class W>
-class WeightGenerate<GallicWeight<Label, W, GALLIC>>
-    : public WeightGenerate<UnionWeight<GallicWeight<Label, W, GALLIC_RESTRICT>,
-                                        GallicUnionWeightOptions<Label, W>>> {
- public:
-  using Weight = GallicWeight<Label, W, GALLIC>;
-  using Generate =
-      WeightGenerate<UnionWeight<GallicWeight<Label, W, GALLIC_RESTRICT>,
-                                 GallicUnionWeightOptions<Label, W>>>;
-
-  explicit WeightGenerate(bool allow_zero = true) : generate_(allow_zero) {}
-
-  Weight operator()() const { return Weight(generate_()); }
-
- private:
-  const Generate generate_;
-};
-
-}  // namespace fst
-
-#endif  // FST_STRING_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/string.h b/projects/llm_framework/include/fst/string.h
deleted file mode 100644
index 07ebfddd..00000000
--- a/projects/llm_framework/include/fst/string.h
+++ /dev/null
@@ -1,286 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Utilities to convert strings into FSTs.
-
-#ifndef FST_STRING_H_
-#define FST_STRING_H_
-
-#include <memory>
-#include <sstream>
-#include <string>
-#include <vector>
-
-#include <fst/flags.h>
-#include <fst/log.h>
-
-#include <fst/compact-fst.h>
-#include <fst/icu.h>
-#include <fst/mutable-fst.h>
-#include <fst/util.h>
-
-
-DECLARE_string(fst_field_separator);
-
-namespace fst {
-
-enum StringTokenType { SYMBOL = 1, BYTE = 2, UTF8 = 3 };
-
-namespace internal {
-
-template <class Label>
-bool ConvertSymbolToLabel(const char *str, const SymbolTable *syms,
-                          Label unknown_label, bool allow_negative,
-                          Label *output) {
-  int64 n;
-  if (syms) {
-    n = syms->Find(str);
-    if ((n == -1) && (unknown_label != kNoLabel)) n = unknown_label;
-    if (n == -1 || (!allow_negative && n < 0)) {
-      LOG(ERROR) << "ConvertSymbolToLabel: Symbol \"" << str
-                 << "\" is not mapped to any integer label, symbol table = "
-                 << syms->Name();
-      return false;
-    }
-  } else {
-    char *p;
-    n = strtoll(str, &p, 10);
-    if (p < str + strlen(str) || (!allow_negative && n < 0)) {
-      LOG(ERROR) << "ConvertSymbolToLabel: Bad label integer "
-                 << "= \"" << str << "\"";
-      return false;
-    }
-  }
-  *output = n;
-  return true;
-}
-
-template <class Label>
-bool ConvertStringToLabels(const string &str, StringTokenType token_type,
-                           const SymbolTable *syms, Label unknown_label,
-                           bool allow_negative, std::vector<Label> *labels) {
-  labels->clear();
-  if (token_type == StringTokenType::BYTE) {
-    labels->reserve(str.size());
-    return ByteStringToLabels(str, labels);
-  } else if (token_type == StringTokenType::UTF8) {
-    return UTF8StringToLabels(str, labels);
-  } else {
-    std::unique_ptr<char[]> c_str(new char[str.size() + 1]);
-    str.copy(c_str.get(), str.size());
-    c_str[str.size()] = 0;
-    std::vector<char *> vec;
-    const string separator = "\n" + FLAGS_fst_field_separator;
-    SplitString(c_str.get(), separator.c_str(), &vec, true);
-    for (const char *c : vec) {
-      Label label;
-      if (!ConvertSymbolToLabel(c, syms, unknown_label, allow_negative,
-                                &label)) {
-        return false;
-      }
-      labels->push_back(label);
-    }
-  }
-  return true;
-}
-
-// The sep string is used as a separator between symbols, unless it is nullptr,
-// in which case the last character of FLAGS_fst_field_separator is used.
-template <class Label>
-bool LabelsToSymbolString(const std::vector<Label> &labels, string *str,
-                          const SymbolTable &syms,
-                          const string *sep = nullptr) {
-  std::stringstream ostrm;
-  string delim = "";
-  for (auto label : labels) {
-    ostrm << delim;
-    const string &symbol = syms.Find(label);
-    if (symbol.empty()) {
-      LOG(ERROR) << "LabelToSymbolString: Label " << label
-                 << " is not mapped onto any textual symbol in symbol table "
-                 << syms.Name();
-      return false;
-    }
-    ostrm << symbol;
-    delim = sep != nullptr ? *sep : string(1, FLAGS_fst_field_separator.back());
-  }
-  *str = ostrm.str();
-  return !!ostrm;
-}
-
-}  // namespace internal
-
-// TODO(kbg): Move this just before StringPrinter.
-// Converts an FST to a vector of output labels. To get input labels, use
-// Project or Invert. Returns true on success. Use only with string FSTs; may
-// loop for non-string FSTs.
-template <class Arc>
-bool StringFstToOutputLabels(const Fst<Arc> &fst,
-                             std::vector<typename Arc::Label> *labels) {
-  labels->clear();
-  auto s = fst.Start();
-  if (s == kNoStateId) {
-    LOG(ERROR) << "StringFstToOutputLabels: Invalid start state";
-    return false;
-  }
-  while (fst.Final(s) == Arc::Weight::Zero()) {
-    ArcIterator<Fst<Arc>> aiter(fst, s);
-    if (aiter.Done()) {
-      LOG(ERROR) << "StringFstToOutputLabels: Does not reach final state";
-      return false;
-    }
-    const auto &arc = aiter.Value();
-    labels->push_back(arc.olabel);
-    s = arc.nextstate;
-    aiter.Next();
-    if (!aiter.Done()) {
-      LOG(ERROR) << "StringFstToOutputLabels: State has multiple outgoing arcs";
-      return false;
-    }
-  }
-  if (fst.NumArcs(s) != 0) {
-    LOG(ERROR) << "FstToOutputLabels: Final state has outgoing arc(s)";
-    return false;
-  }
-  return true;
-}
-
-// TODO(kbg): Move this just before StringPrinter.
-// Converts a list of symbols to a string. Returns true on success. If the token
-// type is SYMBOL and sep is provided, it is used to separate textual symbols.
-// If the token type is SYMBOL and it is not provided, the last character of
-// FLAGS_fst_field_separator is used.
-template <class Label>
-bool LabelsToString(const std::vector<Label> &labels, string *str,
-                    StringTokenType ttype = BYTE,
-                    const SymbolTable *syms = nullptr,
-                    const string *sep = nullptr) {
-  switch (ttype) {
-    case StringTokenType::BYTE: {
-      return LabelsToByteString(labels, str);
-    }
-    case StringTokenType::UTF8: {
-      return LabelsToUTF8String(labels, str);
-    }
-    case StringTokenType::SYMBOL: {
-      return internal::LabelsToSymbolString(labels, str, *syms, sep);
-    }
-  }
-  return false;
-}
-
-// Functor for compiling a string in an FST.
-template <class Arc>
-class StringCompiler {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit StringCompiler(StringTokenType token_type,
-                          const SymbolTable *syms = nullptr,
-                          Label unknown_label = kNoLabel,
-                          bool allow_negative = false)
-      : token_type_(token_type),
-        syms_(syms),
-        unknown_label_(unknown_label),
-        allow_negative_(allow_negative) {}
-
-  // Compiles string into an FST.
-  template <class FST>
-  bool operator()(const string &str, FST *fst) const {
-    std::vector<Label> labels;
-    if (!internal::ConvertStringToLabels(str, token_type_, syms_,
-                                         unknown_label_, allow_negative_,
-                                         &labels)) {
-      return false;
-    }
-    Compile(labels, fst);
-    return true;
-  }
-
-  template <class FST>
-  bool operator()(const string &str, FST *fst, Weight weight) const {
-    std::vector<Label> labels;
-    if (!internal::ConvertStringToLabels(str, token_type_, syms_,
-                                         unknown_label_, allow_negative_,
-                                         &labels)) {
-      return false;
-    }
-    Compile(labels, fst, std::move(weight));
-    return true;
-  }
-
- private:
-  void Compile(const std::vector<Label> &labels, MutableFst<Arc> *fst,
-               Weight weight = Weight::One()) const {
-    fst->DeleteStates();
-    while (fst->NumStates() <= labels.size()) fst->AddState();
-    for (StateId i = 0; i < labels.size(); ++i) {
-      fst->AddArc(i, Arc(labels[i], labels[i], Weight::One(), i + 1));
-    }
-    fst->SetStart(0);
-    fst->SetFinal(labels.size(), std::move(weight));
-  }
-
-  template <class Unsigned>
-  void Compile(const std::vector<Label> &labels,
-               CompactStringFst<Arc, Unsigned> *fst) const {
-    fst->SetCompactElements(labels.begin(), labels.end());
-  }
-
-  template <class Unsigned>
-  void Compile(const std::vector<Label> &labels,
-               CompactWeightedStringFst<Arc, Unsigned> *fst,
-               const Weight &weight = Weight::One()) const {
-    std::vector<std::pair<Label, Weight>> compacts;
-    compacts.reserve(labels.size() + 1);
-    for (StateId i = 0; i < static_cast<StateId>(labels.size()) - 1; ++i) {
-      compacts.emplace_back(labels[i], Weight::One());
-    }
-    compacts.emplace_back(!labels.empty() ? labels.back() : kNoLabel, weight);
-    fst->SetCompactElements(compacts.begin(), compacts.end());
-  }
-
-  const StringTokenType token_type_;
-  const SymbolTable *syms_;    // Symbol table (used when token type is symbol).
-  const Label unknown_label_;  // Label for token missing from symbol table.
-  const bool allow_negative_;  // Negative labels allowed?
-
-  StringCompiler(const StringCompiler &) = delete;
-  StringCompiler &operator=(const StringCompiler &) = delete;
-};
-
-// Helpers for StringPrinter.
-
-// Functor for printing a string FST as a string.
-template <class Arc>
-class StringPrinter {
- public:
-  using Label = typename Arc::Label;
-
-  explicit StringPrinter(StringTokenType token_type,
-                         const SymbolTable *syms = nullptr)
-      : token_type_(token_type), syms_(syms) {}
-
-  // Converts the FST into a string. With SYMBOL token type, sep is used as a
-  // separator between symbols, unless it is nullptr, in which case the last
-  // character of FLAGS_fst_field_separator is used. Returns true on success.
-  bool operator()(const Fst<Arc> &fst, string *str,
-                  const string *sep = nullptr) const {
-    std::vector<Label> labels;
-    return (StringFstToOutputLabels(fst, &labels) &&
-            LabelsToString(labels, str, token_type_, syms_, sep));
-  }
-
- private:
-  const StringTokenType token_type_;
-  const SymbolTable *syms_;
-
-  StringPrinter(const StringPrinter &) = delete;
-  StringPrinter &operator=(const StringPrinter &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_STRING_H_
diff --git a/projects/llm_framework/include/fst/symbol-table-ops.h b/projects/llm_framework/include/fst/symbol-table-ops.h
deleted file mode 100644
index 38f9d819..00000000
--- a/projects/llm_framework/include/fst/symbol-table-ops.h
+++ /dev/null
@@ -1,76 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SYMBOL_TABLE_OPS_H_
-#define FST_SYMBOL_TABLE_OPS_H_
-
-#include <string>
-#include <unordered_set>
-#include <vector>
-
-
-#include <fst/fst.h>
-#include <fst/symbol-table.h>
-
-
-namespace fst {
-
-// Returns a minimal symbol table containing only symbols referenced by the
-// passed fst.  Symbols preserve their original numbering, so fst does not
-// require relabeling.
-template <class Arc>
-SymbolTable *PruneSymbolTable(const Fst<Arc> &fst, const SymbolTable &syms,
-                              bool input) {
-  std::unordered_set<typename Arc::Label> seen;
-  seen.insert(0);  // Always keep epsilon.
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    for (ArcIterator<Fst<Arc>> aiter(fst, siter.Value()); !aiter.Done();
-         aiter.Next()) {
-      const auto sym = (input) ? aiter.Value().ilabel : aiter.Value().olabel;
-      seen.insert(sym);
-    }
-  }
-  auto *pruned = new SymbolTable(syms.Name() + "_pruned");
-  for (SymbolTableIterator stiter(syms); !stiter.Done(); stiter.Next()) {
-    const auto label = stiter.Value();
-    if (seen.count(label)) pruned->AddSymbol(stiter.Symbol(), label);
-  }
-  return pruned;
-}
-
-// Relabels a symbol table to make it a contiguous mapping.
-SymbolTable *CompactSymbolTable(const SymbolTable &syms);
-
-// Merges two SymbolTables, all symbols from left will be merged into right
-// with the same ids.  Symbols in right that have conflicting ids with those
-// in left will be assigned to value assigned from the left SymbolTable.
-// The returned symbol table will never modify symbol assignments from the left
-// side, but may do so on the right.  If right_relabel_output is non-NULL, it
-// will be assigned true if the symbols from the right table needed to be
-// reassigned.
-// A potential use case is to Compose two Fst's that have different symbol
-// tables.  You can reconcile them in the following way:
-//   Fst<Arc> a, b;
-//   bool relabel;
-//   std::unique_ptr<SymbolTable> bnew(MergeSymbolTable(a.OutputSymbols(),
-//                                     b.InputSymbols(), &relabel);
-//   if (relabel) {
-//     Relabel(b, bnew.get(), nullptr);
-//   }
-//   b.SetInputSymbols(bnew);
-SymbolTable *MergeSymbolTable(const SymbolTable &left, const SymbolTable &right,
-                              bool *right_relabel_output = nullptr);
-
-// Read the symbol table from any Fst::Read()able file, without loading the
-// corresponding Fst.  Returns nullptr if the Fst does not contain a symbol
-// table or the symbol table cannot be read.
-SymbolTable *FstReadSymbols(const string &filename, bool input);
-
-// Adds a contiguous range of symbols to a symbol table using a simple prefix
-// for the string, returning false if the inserted symbol string clashes with
-// any currently present.
-bool AddAuxiliarySymbols(const string &prefix, int64 start_label,
-                         int64 nlabels, SymbolTable *syms);
-
-}  // namespace fst
-#endif  // FST_SYMBOL_TABLE_OPS_H_
diff --git a/projects/llm_framework/include/fst/symbol-table.h b/projects/llm_framework/include/fst/symbol-table.h
deleted file mode 100644
index 2884bc3b..00000000
--- a/projects/llm_framework/include/fst/symbol-table.h
+++ /dev/null
@@ -1,445 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes to provide symbol-to-integer and integer-to-symbol mappings.
-
-#ifndef FST_SYMBOL_TABLE_H_
-#define FST_SYMBOL_TABLE_H_
-
-#include <functional>
-#include <ios>
-#include <iostream>
-#include <memory>
-#include <sstream>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/flags.h>
-#include <fst/log.h>
-#include <fstream>
-#include <map>
-
-DECLARE_bool(fst_compat_symbols);
-
-namespace fst {
-
-constexpr int64 kNoSymbol = -1;
-
-// WARNING: Reading via symbol table read options should
-//          not be used. This is a temporary work around for
-//          reading symbol ranges of previously stored symbol sets.
-struct SymbolTableReadOptions {
-  SymbolTableReadOptions() {}
-
-  SymbolTableReadOptions(
-      std::vector<std::pair<int64, int64>> string_hash_ranges,
-      const string &source)
-      : string_hash_ranges(std::move(string_hash_ranges)), source(source) {}
-
-  std::vector<std::pair<int64, int64>> string_hash_ranges;
-  string source;
-};
-
-struct SymbolTableTextOptions {
-  explicit SymbolTableTextOptions(bool allow_negative_labels = false);
-
-  bool allow_negative_labels;
-  string fst_field_separator;
-};
-
-namespace internal {
-
-extern const int kLineLen;
-
-// List of symbols with a dense hash for looking up symbol index, rehashing at
-// 75% occupancy.
-class DenseSymbolMap {
- public:
-  DenseSymbolMap();
-
-  DenseSymbolMap(const DenseSymbolMap &x);
-
-  std::pair<int64, bool> InsertOrFind(const string &key);
-
-  int64 Find(const string &key) const;
-
-  size_t Size() const { return symbols_.size(); }
-
-  const string &GetSymbol(size_t idx) const { return symbols_[idx]; }
-
-  void RemoveSymbol(size_t idx);
-
- private:
-  // num_buckets must be power of 2.
-  void Rehash(size_t num_buckets);
-
-  int64 empty_;
-  std::vector<string> symbols_;
-  std::hash<string> str_hash_;
-  std::vector<int64> buckets_;
-  uint64 hash_mask_;
-};
-
-class SymbolTableImpl {
- public:
-  explicit SymbolTableImpl(const string &name)
-      : name_(name),
-        available_key_(0),
-        dense_key_limit_(0),
-        check_sum_finalized_(false) {}
-
-  SymbolTableImpl(const SymbolTableImpl &impl)
-      : name_(impl.name_),
-        available_key_(impl.available_key_),
-        dense_key_limit_(impl.dense_key_limit_),
-        symbols_(impl.symbols_),
-        idx_key_(impl.idx_key_),
-        key_map_(impl.key_map_),
-        check_sum_finalized_(false) {}
-
-  int64 AddSymbol(const string &symbol, int64 key);
-
-  int64 AddSymbol(const string &symbol) {
-    return AddSymbol(symbol, available_key_);
-  }
-
-  // Removes the symbol with the given key. The removal is costly
-  // (O(NumSymbols)) and may reduce the efficiency of Find() because of a
-  // potentially reduced size of the dense key interval.
-  void RemoveSymbol(int64 key);
-
-  static SymbolTableImpl *ReadText(
-      std::istream &strm, const string &name,
-      const SymbolTableTextOptions &opts = SymbolTableTextOptions());
-
-  static SymbolTableImpl* Read(std::istream &strm,
-                               const SymbolTableReadOptions &opts);
-
-  bool Write(std::ostream &strm) const;
-
-  // Return the string associated with the key. If the key is out of
-  // range (<0, >max), return an empty string.
-  string Find(int64 key) const {
-    int64 idx = key;
-    if (key < 0 || key >= dense_key_limit_) {
-      const auto it = key_map_.find(key);
-      if (it == key_map_.end()) return "";
-      idx = it->second;
-    }
-    if (idx < 0 || idx >= symbols_.Size()) return "";
-    return symbols_.GetSymbol(idx);
-  }
-
-  // Returns the key associated with the symbol; if the symbol
-  // does not exists, returns kNoSymbol.
-  int64 Find(const string &symbol) const {
-    int64 idx = symbols_.Find(symbol);
-    if (idx == kNoSymbol || idx < dense_key_limit_) return idx;
-    return idx_key_[idx - dense_key_limit_];
-  }
-
-  bool Member(int64 key) const { return !Find(key).empty(); }
-
-  bool Member(const string &symbol) const { return Find(symbol) != kNoSymbol; }
-
-  int64 GetNthKey(ssize_t pos) const {
-    if (pos < 0 || pos >= symbols_.Size()) return kNoSymbol;
-    if (pos < dense_key_limit_) return pos;
-    return Find(symbols_.GetSymbol(pos));
-  }
-
-  const string &Name() const { return name_; }
-
-  void SetName(const string &new_name) { name_ = new_name; }
-
-  const string &CheckSum() const {
-    MaybeRecomputeCheckSum();
-    return check_sum_string_;
-  }
-
-  const string &LabeledCheckSum() const {
-    MaybeRecomputeCheckSum();
-    return labeled_check_sum_string_;
-  }
-
-  int64 AvailableKey() const { return available_key_; }
-
-  size_t NumSymbols() const { return symbols_.Size(); }
-
- private:
-  // Recomputes the checksums (both of them) if we've had changes since the last
-  // computation (i.e., if check_sum_finalized_ is false).
-  // Takes ~2.5 microseconds (dbg) or ~230 nanoseconds (opt) on a 2.67GHz Xeon
-  // if the checksum is up-to-date (requiring no recomputation).
-  void MaybeRecomputeCheckSum() const;
-
-  string name_;
-  int64 available_key_;
-  int64 dense_key_limit_;
-
-  DenseSymbolMap symbols_;
-  // Maps index to key for index >= dense_key_limit:
-  //   key = idx_key_[index - dense_key_limit]
-  std::vector<int64> idx_key_;
-  // Maps key to index for key >= dense_key_limit_.
-  //  index = key_map_[key]
-  std::map<int64, int64> key_map_;
-
-  mutable bool check_sum_finalized_;
-  mutable string check_sum_string_;
-  mutable string labeled_check_sum_string_;
-  mutable Mutex check_sum_mutex_;
-};
-
-}  // namespace internal
-
-// Symbol (string) to integer (and reverse) mapping.
-//
-// The SymbolTable implements the mappings of labels to strings and reverse.
-// SymbolTables are used to describe the alphabet of the input and output
-// labels for arcs in a Finite State Transducer.
-//
-// SymbolTables are reference-counted and can therefore be shared across
-// multiple machines. For example a language model grammar G, with a
-// SymbolTable for the words in the language model can share this symbol
-// table with the lexical representation L o G.
-class SymbolTable {
- public:
-  // Constructs symbol table with an optional name.
-  explicit SymbolTable(const string &name = "<unspecified>")
-      : impl_(std::make_shared<internal::SymbolTableImpl>(name)) {}
-
-  virtual ~SymbolTable() {}
-
-  // Reads a text representation of the symbol table from an istream. Pass a
-  // name to give the resulting SymbolTable.
-  static SymbolTable *ReadText(
-      std::istream &strm, const string &name,
-      const SymbolTableTextOptions &opts = SymbolTableTextOptions()) {
-    auto *impl = internal::SymbolTableImpl::ReadText(strm, name, opts);
-    return impl ? new SymbolTable(impl) : nullptr;
-  }
-
-  // Reads a text representation of the symbol table.
-  static SymbolTable *ReadText(const string &filename,
-      const SymbolTableTextOptions &opts = SymbolTableTextOptions()) {
-    std::ifstream strm(filename, std::ios_base::in);
-    if (!strm.good()) {
-      LOG(ERROR) << "SymbolTable::ReadText: Can't open file " << filename;
-      return nullptr;
-    }
-    return ReadText(strm, filename, opts);
-  }
-
-  // WARNING: Reading via symbol table read options should not be used. This is
-  // a temporary work-around.
-  static SymbolTable* Read(std::istream &strm,
-                           const SymbolTableReadOptions &opts) {
-    auto *impl = internal::SymbolTableImpl::Read(strm, opts);
-    return (impl) ? new SymbolTable(impl) : nullptr;
-  }
-
-  // Reads a binary dump of the symbol table from a stream.
-  static SymbolTable *Read(std::istream &strm,
-                           const string &source) {
-    SymbolTableReadOptions opts;
-    opts.source = source;
-    return Read(strm, opts);
-  }
-
-  // Reads a binary dump of the symbol table.
-  static SymbolTable *Read(const string& filename) {
-    std::ifstream strm(filename,
-                            std::ios_base::in | std::ios_base::binary);
-    if (!strm.good()) {
-      LOG(ERROR) << "SymbolTable::Read: Can't open file " << filename;
-      return nullptr;
-    }
-    return Read(strm, filename);
-  }
-
-  // DERIVABLE INTERFACE
-
-  // Creates a reference counted copy.
-  virtual SymbolTable *Copy() const { return new SymbolTable(*this); }
-
-  // Adds a symbol with given key to table. A symbol table also keeps track of
-  // the last available key (highest key value in the symbol table).
-  virtual int64 AddSymbol(const string &symbol, int64 key) {
-    MutateCheck();
-    return impl_->AddSymbol(symbol, key);
-  }
-
-  // Adds a symbol to the table. The associated value key is automatically
-  // assigned by the symbol table.
-  virtual int64 AddSymbol(const string &symbol) {
-    MutateCheck();
-    return impl_->AddSymbol(symbol);
-  }
-
-  // Adds another symbol table to this table. All key values will be offset
-  // by the current available key (highest key value in the symbol table).
-  // Note string symbols with the same key value will still have the same
-  // key value after the symbol table has been merged, but a different
-  // value. Adding symbol tables do not result in changes in the base table.
-  virtual void AddTable(const SymbolTable &table);
-
-  // Returns the current available key (i.e., highest key + 1) in the symbol
-  // table.
-  virtual int64 AvailableKey() const { return impl_->AvailableKey(); }
-
-  // Return the label-agnostic MD5 check-sum for this table. All new symbols
-  // added to the table will result in an updated checksum. Deprecated.
-  virtual const string &CheckSum() const { return impl_->CheckSum(); }
-
-  virtual int64 GetNthKey(ssize_t pos) const { return impl_->GetNthKey(pos); }
-
-  // Returns the string associated with the key; if the key is out of
-  // range (<0, >max), returns an empty string.
-  virtual string Find(int64 key) const { return impl_->Find(key); }
-
-  // Returns the key associated with the symbol; if the symbol does not exist,
-  // kNoSymbol is returned.
-  virtual int64 Find(const string &symbol) const { return impl_->Find(symbol); }
-
-  // Same as CheckSum(), but returns an label-dependent version.
-  virtual const string &LabeledCheckSum() const {
-    return impl_->LabeledCheckSum();
-  }
-
-  virtual bool Member(int64 key) const { return impl_->Member(key); }
-
-  virtual bool Member(const string &symbol) const {
-    return impl_->Member(symbol);
-  }
-
-  // Returns the name of the symbol table.
-  virtual const string &Name() const { return impl_->Name(); }
-
-  // Returns the current number of symbols in table (not necessarily equal to
-  // AvailableKey()).
-  virtual size_t NumSymbols() const { return impl_->NumSymbols(); }
-
-  virtual void RemoveSymbol(int64 key) {
-    MutateCheck();
-    return impl_->RemoveSymbol(key);
-  }
-
-  // Sets the name of the symbol table.
-  virtual void SetName(const string &new_name) {
-    MutateCheck();
-    impl_->SetName(new_name);
-  }
-
-  virtual bool Write(std::ostream &strm) const { return impl_->Write(strm); }
-
-  virtual bool Write(const string &filename) const {
-    std::ofstream strm(filename,
-                             std::ios_base::out | std::ios_base::binary);
-    if (!strm.good()) {
-      LOG(ERROR) << "SymbolTable::Write: Can't open file " << filename;
-      return false;
-    }
-    return Write(strm);
-  }
-
-  // Dump a text representation of the symbol table via a stream.
-  virtual bool WriteText(std::ostream &strm,
-      const SymbolTableTextOptions &opts = SymbolTableTextOptions()) const;
-
-  // Dump an text representation of the symbol table.
-  virtual bool WriteText(const string &filename) const {
-    std::ofstream strm(filename);
-    if (!strm.good()) {
-      LOG(ERROR) << "SymbolTable::WriteText: Can't open file " << filename;
-      return false;
-    }
-    return WriteText(strm);
-  }
-
- private:
-  explicit SymbolTable(internal::SymbolTableImpl *impl) : impl_(impl) {}
-
-  void MutateCheck() {
-    if (!impl_.unique()) impl_.reset(new internal::SymbolTableImpl(*impl_));
-  }
-
-  const internal::SymbolTableImpl *Impl() const { return impl_.get(); }
-
- private:
-  std::shared_ptr<internal::SymbolTableImpl> impl_;
-};
-
-// Iterator class for symbols in a symbol table.
-class SymbolTableIterator {
- public:
-  explicit SymbolTableIterator(const SymbolTable &table)
-      : table_(table),
-        pos_(0),
-        nsymbols_(table.NumSymbols()),
-        key_(table.GetNthKey(0)) {}
-
-  ~SymbolTableIterator() {}
-
-  // Returns whether iterator is done.
-  bool Done() const { return (pos_ == nsymbols_); }
-
-  // Return the key of the current symbol.
-  int64 Value() const { return key_; }
-
-  // Return the string of the current symbol.
-  string Symbol() const { return table_.Find(key_); }
-
-  // Advances iterator.
-  void Next() {
-    ++pos_;
-    if (pos_ < nsymbols_) key_ = table_.GetNthKey(pos_);
-  }
-
-  // Resets iterator.
-  void Reset() {
-    pos_ = 0;
-    key_ = table_.GetNthKey(0);
-  }
-
- private:
-  const SymbolTable &table_;
-  ssize_t pos_;
-  size_t nsymbols_;
-  int64 key_;
-};
-
-// Relabels a symbol table as specified by the input vector of pairs
-// (old label, new label). The new symbol table only retains symbols
-// for which a relabeling is explicitly specified.
-//
-// TODO(allauzen): consider adding options to allow for some form of implicit
-// identity relabeling.
-template <class Label>
-SymbolTable *RelabelSymbolTable(const SymbolTable *table,
-    const std::vector<std::pair<Label, Label>> &pairs) {
-  auto *new_table = new SymbolTable(
-      table->Name().empty() ? string()
-                            : (string("relabeled_") + table->Name()));
-  for (const auto &pair : pairs) {
-    new_table->AddSymbol(table->Find(pair.first), pair.second);
-  }
-  return new_table;
-}
-
-// Returns true if the two symbol tables have equal checksums. Passing in
-// nullptr for either table always returns true.
-bool CompatSymbols(const SymbolTable *syms1, const SymbolTable *syms2,
-                   bool warning = true);
-
-// Symbol Table serialization.
-
-void SymbolTableToString(const SymbolTable *table, string *result);
-
-SymbolTable *StringToSymbolTable(const string &str);
-
-}  // namespace fst
-
-#endif  // FST_SYMBOL_TABLE_H_
diff --git a/projects/llm_framework/include/fst/synchronize.h b/projects/llm_framework/include/fst/synchronize.h
deleted file mode 100644
index 0c1c6862..00000000
--- a/projects/llm_framework/include/fst/synchronize.h
+++ /dev/null
@@ -1,414 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Synchronize an FST with bounded delay.
-
-#ifndef FST_SYNCHRONIZE_H_
-#define FST_SYNCHRONIZE_H_
-
-#include <algorithm>
-#include <string>
-#include <unordered_map>
-#include <unordered_set>
-#include <utility>
-#include <vector>
-
-#include <fst/cache.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-using SynchronizeFstOptions = CacheOptions;
-
-namespace internal {
-
-// Implementation class for SynchronizeFst.
-// TODO(kbg,sorenj): Refactor to guarantee thread-safety.
-
-template <class Arc>
-class SynchronizeFstImpl : public CacheImpl<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheBaseImpl<CacheState<Arc>>::EmplaceArc;
-  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
-  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
-  using CacheBaseImpl<CacheState<Arc>>::HasStart;
-  using CacheBaseImpl<CacheState<Arc>>::SetArcs;
-  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
-  using CacheBaseImpl<CacheState<Arc>>::SetStart;
-
-  using String = basic_string<Label>;
-
-  struct Element {
-    Element() {}
-
-    Element(StateId state_, const String *i, const String *o)
-        : state(state_), istring(i), ostring(o) {}
-
-    StateId state;          // Input state ID.
-    const String *istring;  // Residual input labels.
-    const String *ostring;  // Residual output labels.
-    // Residual strings are represented by const pointers to
-    // basic_string<Label> and are stored in a hash_set. The pointed
-    // memory is owned by the hash_set string_set_.
-  };
-
-  SynchronizeFstImpl(const Fst<Arc> &fst, const SynchronizeFstOptions &opts)
-      : CacheImpl<Arc>(opts), fst_(fst.Copy()) {
-    SetType("synchronize");
-    const auto props = fst.Properties(kFstProperties, false);
-    SetProperties(SynchronizeProperties(props), kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  SynchronizeFstImpl(const SynchronizeFstImpl &impl)
-      : CacheImpl<Arc>(impl), fst_(impl.fst_->Copy(true)) {
-    SetType("synchronize");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  ~SynchronizeFstImpl() override {
-    for (const auto *ptr : string_set_) delete ptr;
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      auto start = fst_->Start();
-      if (start == kNoStateId) return kNoStateId;
-      const auto *empty = FindString(new String());
-      start = FindState(Element(fst_->Start(), empty, empty));
-      SetStart(start);
-    }
-    return CacheImpl<Arc>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) {
-      const auto &element = elements_[s];
-      const auto weight = element.state == kNoStateId
-                              ? Weight::One()
-                              : fst_->Final(element.state);
-      if ((weight != Weight::Zero()) && (element.istring)->empty() &&
-          (element.ostring)->empty()) {
-        SetFinal(s, weight);
-      } else {
-        SetFinal(s, Weight::Zero());
-      }
-    }
-    return CacheImpl<Arc>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumOutputEpsilons(s);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, returning other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && fst_->Properties(kError, false)) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<Arc>::InitArcIterator(s, data);
-  }
-
-  // Returns the first character of the string obtained by concatenating the
-  // string and the label.
-  Label Car(const String *str, Label label = 0) const {
-    if (!str->empty()) {
-      return (*str)[0];
-    } else {
-      return label;
-    }
-  }
-
-  // Computes the residual string obtained by removing the first
-  // character in the concatenation of the string and the label.
-  const String *Cdr(const String *str, Label label = 0) {
-    auto *r = new String();
-    for (size_t i = 1; i < str->size(); ++i) r->push_back((*str)[i]);
-    if (label && !(str->empty())) r->push_back(label);
-    return FindString(r);
-  }
-
-  // Computes the concatenation of the string and the label.
-  const String *Concat(const String *str, Label label = 0) {
-    auto *r = new String();
-    for (size_t i = 0; i < str->size(); ++i) r->push_back((*str)[i]);
-    if (label) r->push_back(label);
-    return FindString(r);
-  }
-
-  // Tests if the concatenation of the string and label is empty.
-  bool Empty(const String *str, Label label = 0) const {
-    if (str->empty()) {
-      return label == 0;
-    } else {
-      return false;
-    }
-  }
-
-  // Finds the string pointed by s in the hash set. Transfers the pointer
-  // ownership to the hash set.
-  const String *FindString(const String *str) {
-    const auto insert_result = string_set_.insert(str);
-    if (!insert_result.second) {
-      delete str;
-    }
-    return *insert_result.first;
-  }
-
-  // Finds state corresponding to an element. Creates new state if element
-  // is not found.
-  StateId FindState(const Element &element) {
-    const auto insert_result =
-        element_map_.insert(std::make_pair(element, elements_.size()));
-    if (insert_result.second) {
-      elements_.push_back(element);
-    }
-    return insert_result.first->second;
-  }
-
-  // Computes the outgoing transitions from a state, creating new destination
-  // states as needed.
-  void Expand(StateId s) {
-    const auto element = elements_[s];
-    if (element.state != kNoStateId) {
-      for (ArcIterator<Fst<Arc>> aiter(*fst_, element.state); !aiter.Done();
-           aiter.Next()) {
-        const auto &arc = aiter.Value();
-        if (!Empty(element.istring, arc.ilabel) &&
-            !Empty(element.ostring, arc.olabel)) {
-          const auto *istring = Cdr(element.istring, arc.ilabel);
-          const auto *ostring = Cdr(element.ostring, arc.olabel);
-          EmplaceArc(s, Car(element.istring, arc.ilabel),
-                     Car(element.ostring, arc.olabel), arc.weight,
-                     FindState(Element(arc.nextstate, istring, ostring)));
-        } else {
-          const auto *istring = Concat(element.istring, arc.ilabel);
-          const auto *ostring = Concat(element.ostring, arc.olabel);
-          EmplaceArc(s, 0, 0, arc.weight,
-                     FindState(Element(arc.nextstate, istring, ostring)));
-        }
-      }
-    }
-    const auto weight = element.state == kNoStateId
-                            ? Weight::One()
-                            : fst_->Final(element.state);
-    if ((weight != Weight::Zero()) &&
-        ((element.istring)->size() + (element.ostring)->size() > 0)) {
-      const auto *istring = Cdr(element.istring);
-      const auto *ostring = Cdr(element.ostring);
-      EmplaceArc(s, Car(element.istring), Car(element.ostring), weight,
-                 FindState(Element(kNoStateId, istring, ostring)));
-    }
-    SetArcs(s);
-  }
-
- private:
-  // Equality function for Elements; assumes strings have been hashed.
-  class ElementEqual {
-   public:
-    bool operator()(const Element &x, const Element &y) const {
-      return x.state == y.state && x.istring == y.istring &&
-             x.ostring == y.ostring;
-    }
-  };
-
-  // Hash function for Elements to FST states.
-  class ElementKey {
-   public:
-    size_t operator()(const Element &x) const {
-      size_t key = x.state;
-      key = (key << 1) ^ (x.istring)->size();
-      for (size_t i = 0; i < (x.istring)->size(); ++i) {
-        key = (key << 1) ^ (*x.istring)[i];
-      }
-      key = (key << 1) ^ (x.ostring)->size();
-      for (size_t i = 0; i < (x.ostring)->size(); ++i) {
-        key = (key << 1) ^ (*x.ostring)[i];
-      }
-      return key;
-    }
-  };
-
-  // Equality function for strings.
-  class StringEqual {
-   public:
-    bool operator()(const String *const &x, const String *const &y) const {
-      if (x->size() != y->size()) return false;
-      for (size_t i = 0; i < x->size(); ++i) {
-        if ((*x)[i] != (*y)[i]) return false;
-      }
-      return true;
-    }
-  };
-
-  // Hash function for set of strings
-  class StringKey {
-   public:
-    size_t operator()(const String *const &x) const {
-      size_t key = x->size();
-      for (size_t i = 0; i < x->size(); ++i) key = (key << 1) ^ (*x)[i];
-      return key;
-    }
-  };
-
-  using ElementMap =
-      std::unordered_map<Element, StateId, ElementKey, ElementEqual>;
-  using StringSet = std::unordered_set<const String *, StringKey, StringEqual>;
-
-  std::unique_ptr<const Fst<Arc>> fst_;
-  std::vector<Element> elements_;  // Maps FST state to Elements.
-  ElementMap element_map_;         // Maps Elements to FST state.
-  StringSet string_set_;
-};
-
-}  // namespace internal
-
-// Synchronizes a transducer. This version is a delayed FST. The result is an
-// equivalent FST that has the property that during the traversal of a path,
-// the delay is either zero or strictly increasing, where the delay is the
-// difference between the number of non-epsilon output labels and input labels
-// along the path.
-//
-// For the algorithm to terminate, the input transducer must have bounded
-// delay, i.e., the delay of every cycle must be zero.
-//
-// Complexity:
-//
-// - A has bounded delay: exponential.
-// - A does not have bounded delay: does not terminate.
-//
-// For more information, see:
-//
-// Mohri, M. 2003. Edit-distance of weighted automata: General definitions and
-// algorithms. International Journal of Computer Science 14(6): 957-982.
-//
-// This class attaches interface to implementation and handles reference
-// counting, delegating most methods to ImplToFst.
-template <class A>
-class SynchronizeFst : public ImplToFst<internal::SynchronizeFstImpl<A>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::SynchronizeFstImpl<A>;
-
-  friend class ArcIterator<SynchronizeFst<A>>;
-  friend class StateIterator<SynchronizeFst<A>>;
-
-  explicit SynchronizeFst(
-      const Fst<A> &fst,
-      const SynchronizeFstOptions &opts = SynchronizeFstOptions())
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  SynchronizeFst(const SynchronizeFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Gets a copy of this SynchronizeFst. See Fst<>::Copy() for further doc.
-  SynchronizeFst<Arc> *Copy(bool safe = false) const override {
-    return new SynchronizeFst<Arc>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  SynchronizeFst &operator=(const SynchronizeFst &) = delete;
-};
-
-// Specialization for SynchronizeFst.
-template <class Arc>
-class StateIterator<SynchronizeFst<Arc>>
-    : public CacheStateIterator<SynchronizeFst<Arc>> {
- public:
-  explicit StateIterator(const SynchronizeFst<Arc> &fst)
-      : CacheStateIterator<SynchronizeFst<Arc>>(fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for SynchronizeFst.
-template <class Arc>
-class ArcIterator<SynchronizeFst<Arc>>
-    : public CacheArcIterator<SynchronizeFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const SynchronizeFst<Arc> &fst, StateId s)
-      : CacheArcIterator<SynchronizeFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc>
-inline void SynchronizeFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<SynchronizeFst<Arc>>(*this);
-}
-
-// Synchronizes a transducer. This version writes the synchronized result to a
-// MutableFst. The result will be an equivalent FST that has the property that
-// during the traversal of a path, the delay is either zero or strictly
-// increasing, where the delay is the difference between the number of
-// non-epsilon output labels and input labels along the path.
-//
-// For the algorithm to terminate, the input transducer must have bounded
-// delay, i.e., the delay of every cycle must be zero.
-//
-// Complexity:
-//
-// - A has bounded delay: exponential.
-// - A does not have bounded delay: does not terminate.
-//
-// For more information, see:
-//
-// Mohri, M. 2003. Edit-distance of weighted automata: General definitions and
-// algorithms. International Journal of Computer Science 14(6): 957-982.
-template <class Arc>
-void Synchronize(const Fst<Arc> &ifst, MutableFst<Arc> *ofst) {
-  // Caches only the last state for fastest copy.
-  const SynchronizeFstOptions opts(FLAGS_fst_default_cache_gc, 0);
-  *ofst = SynchronizeFst<Arc>(ifst, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_SYNCHRONIZE_H_
diff --git a/projects/llm_framework/include/fst/test-properties.h b/projects/llm_framework/include/fst/test-properties.h
deleted file mode 100644
index 677ed01b..00000000
--- a/projects/llm_framework/include/fst/test-properties.h
+++ /dev/null
@@ -1,246 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions to manipulate and test property bits.
-
-#ifndef FST_TEST_PROPERTIES_H_
-#define FST_TEST_PROPERTIES_H_
-
-#include <unordered_set>
-
-#include <fst/flags.h>
-#include <fst/log.h>
-
-#include <fst/connect.h>
-#include <fst/dfs-visit.h>
-
-
-DECLARE_bool(fst_verify_properties);
-
-namespace fst {
-// namespace internal {
-
-// For a binary property, the bit is always returned set. For a trinary (i.e.,
-// two-bit) property, both bits are returned set iff either corresponding input
-// bit is set.
-inline uint64 KnownProperties(uint64 props) {
-  return kBinaryProperties | (props & kTrinaryProperties) |
-         ((props & kPosTrinaryProperties) << 1) |
-         ((props & kNegTrinaryProperties) >> 1);
-}
-
-// Tests compatibility between two sets of properties.
-inline bool CompatProperties(uint64 props1, uint64 props2) {
-  const auto known_props1 = KnownProperties(props1);
-  const auto known_props2 = KnownProperties(props2);
-  const auto known_props = known_props1 & known_props2;
-  const auto incompat_props = (props1 & known_props) ^ (props2 & known_props);
-  if (incompat_props) {
-    uint64 prop = 1;
-    for (int i = 0; i < 64; ++i, prop <<= 1) {
-      if (prop & incompat_props) {
-        LOG(ERROR) << "CompatProperties: Mismatch: " << PropertyNames[i]
-                   << ": props1 = " << (props1 & prop ? "true" : "false")
-                   << ", props2 = " << (props2 & prop ? "true" : "false");
-      }
-    }
-    return false;
-  } else {
-    return true;
-  }
-}
-
-// Computes FST property values defined in properties.h. The value of each
-// property indicated in the mask will be determined and returned (these will
-// never be unknown here). In the course of determining the properties
-// specifically requested in the mask, certain other properties may be
-// determined (those with little additional expense) and their values will be
-// returned as well. The complete set of known properties (whether true or
-// false) determined by this operation will be assigned to the the value pointed
-// to by KNOWN. If 'use_stored' is true, pre-computed FST properties may be used
-// when possible. 'mask & required_mask' is used to determine whether the stored
-// propertoes can be used. This routine is seldom called directly; instead it is
-// used to implement fst.Properties(mask, true).
-template <class Arc>
-uint64 ComputeProperties(const Fst<Arc> &fst, uint64 mask, uint64 *known,
-                         bool use_stored) {
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  const auto fst_props = fst.Properties(kFstProperties, false);  // FST-stored.
-  // Check stored FST properties first if allowed.
-  if (use_stored) {
-    const auto known_props = KnownProperties(fst_props);
-    // If FST contains required info, return it.
-    if ((known_props & mask) == mask) {
-      if (known) *known = known_props;
-      return fst_props;
-    }
-  }
-  // Computes (trinary) properties explicitly.
-  // Initialize with binary properties (already known).
-  uint64 comp_props = fst_props & kBinaryProperties;
-  // Computes these trinary properties with a DFS. We compute only those that
-  // need a DFS here, since we otherwise would like to avoid a DFS since its
-  // stack could grow large.
-  uint64 dfs_props = kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic |
-                     kAccessible | kNotAccessible | kCoAccessible |
-                     kNotCoAccessible;
-  std::vector<StateId> scc;
-  if (mask & (dfs_props | kWeightedCycles | kUnweightedCycles)) {
-    SccVisitor<Arc> scc_visitor(&scc, nullptr, nullptr, &comp_props);
-    DfsVisit(fst, &scc_visitor);
-  }
-  // Computes any remaining trinary properties via a state and arcs iterations
-  if (mask & ~(kBinaryProperties | dfs_props)) {
-    comp_props |= kAcceptor | kNoEpsilons | kNoIEpsilons | kNoOEpsilons |
-                  kILabelSorted | kOLabelSorted | kUnweighted | kTopSorted |
-                  kString;
-    if (mask & (kIDeterministic | kNonIDeterministic)) {
-      comp_props |= kIDeterministic;
-    }
-    if (mask & (kODeterministic | kNonODeterministic)) {
-      comp_props |= kODeterministic;
-    }
-    if (mask & (dfs_props | kWeightedCycles | kUnweightedCycles)) {
-      comp_props |= kUnweightedCycles;
-    }
-    std::unique_ptr<std::unordered_set<Label>> ilabels;
-    std::unique_ptr<std::unordered_set<Label>> olabels;
-    StateId nfinal = 0;
-    for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-      StateId s = siter.Value();
-      Arc prev_arc;
-      // Creates these only if we need to.
-      if (mask & (kIDeterministic | kNonIDeterministic)) {
-        ilabels.reset(new std::unordered_set<Label>());
-      }
-      if (mask & (kODeterministic | kNonODeterministic)) {
-        olabels.reset(new std::unordered_set<Label>());
-      }
-      bool first_arc = true;
-      for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-        const auto &arc = aiter.Value();
-        if (ilabels && ilabels->find(arc.ilabel) != ilabels->end()) {
-          comp_props |= kNonIDeterministic;
-          comp_props &= ~kIDeterministic;
-        }
-        if (olabels && olabels->find(arc.olabel) != olabels->end()) {
-          comp_props |= kNonODeterministic;
-          comp_props &= ~kODeterministic;
-        }
-        if (arc.ilabel != arc.olabel) {
-          comp_props |= kNotAcceptor;
-          comp_props &= ~kAcceptor;
-        }
-        if (arc.ilabel == 0 && arc.olabel == 0) {
-          comp_props |= kEpsilons;
-          comp_props &= ~kNoEpsilons;
-        }
-        if (arc.ilabel == 0) {
-          comp_props |= kIEpsilons;
-          comp_props &= ~kNoIEpsilons;
-        }
-        if (arc.olabel == 0) {
-          comp_props |= kOEpsilons;
-          comp_props &= ~kNoOEpsilons;
-        }
-        if (!first_arc) {
-          if (arc.ilabel < prev_arc.ilabel) {
-            comp_props |= kNotILabelSorted;
-            comp_props &= ~kILabelSorted;
-          }
-          if (arc.olabel < prev_arc.olabel) {
-            comp_props |= kNotOLabelSorted;
-            comp_props &= ~kOLabelSorted;
-          }
-        }
-        if (arc.weight != Weight::One() && arc.weight != Weight::Zero()) {
-          comp_props |= kWeighted;
-          comp_props &= ~kUnweighted;
-          if ((comp_props & kUnweightedCycles) &&
-              scc[s] == scc[arc.nextstate]) {
-            comp_props |= kWeightedCycles;
-            comp_props &= ~kUnweightedCycles;
-          }
-        }
-        if (arc.nextstate <= s) {
-          comp_props |= kNotTopSorted;
-          comp_props &= ~kTopSorted;
-        }
-        if (arc.nextstate != s + 1) {
-          comp_props |= kNotString;
-          comp_props &= ~kString;
-        }
-        prev_arc = arc;
-        first_arc = false;
-        if (ilabels) ilabels->insert(arc.ilabel);
-        if (olabels) olabels->insert(arc.olabel);
-      }
-
-      if (nfinal > 0) {  // Final state not last.
-        comp_props |= kNotString;
-        comp_props &= ~kString;
-      }
-      const auto final_weight = fst.Final(s);
-      if (final_weight != Weight::Zero()) {  // Final state.
-        if (final_weight != Weight::One()) {
-          comp_props |= kWeighted;
-          comp_props &= ~kUnweighted;
-        }
-        ++nfinal;
-      } else {  // Non-final state.
-        if (fst.NumArcs(s) != 1) {
-          comp_props |= kNotString;
-          comp_props &= ~kString;
-        }
-      }
-    }
-    if (fst.Start() != kNoStateId && fst.Start() != 0) {
-      comp_props |= kNotString;
-      comp_props &= ~kString;
-    }
-  }
-  if (known) *known = KnownProperties(comp_props);
-  return comp_props;
-}
-
-// This is a wrapper around ComputeProperties that will cause a fatal error if
-// the stored properties and the computed properties are incompatible when
-// FLAGS_fst_verify_properties is true. This routine is seldom called directly;
-// instead it is used to implement fst.Properties(mask, true).
-template <class Arc>
-uint64 TestProperties(const Fst<Arc> &fst, uint64 mask, uint64 *known) {
-  if (FLAGS_fst_verify_properties) {
-    const auto stored_props = fst.Properties(kFstProperties, false);
-    const auto computed_props = ComputeProperties(fst, mask, known, false);
-    if (!CompatProperties(stored_props, computed_props)) {
-      FSTERROR() << "TestProperties: stored FST properties incorrect"
-                 << " (stored: props1, computed: props2)";
-    }
-    return computed_props;
-  } else {
-    return ComputeProperties(fst, mask, known, true);
-  }
-}
-
-// If all the properties of 'fst' corresponding to 'check_mask' are known,
-// returns the stored properties. Otherwise, the properties corresponding to
-// both 'check_mask' and 'test_mask' are computed. This is used to check for
-// newly-added properties that might not be set in old binary files.
-template <class Arc>
-uint64 CheckProperties(const Fst<Arc> &fst, uint64 check_mask,
-                       uint64 test_mask) {
-  auto props = fst.Properties(kFstProperties, false);
-  if (FLAGS_fst_verify_properties) {
-    props = TestProperties(fst, check_mask | test_mask, nullptr);
-  } else if ((KnownProperties(props) & check_mask) != check_mask) {
-    props = ComputeProperties(fst, check_mask | test_mask, nullptr, false);
-  }
-  return props & (check_mask | test_mask);
-}
-
-//}  // namespace internal
-}  // namespace fst
-
-#endif  // FST_TEST_PROPERTIES_H_
diff --git a/projects/llm_framework/include/fst/test/algo_test.h b/projects/llm_framework/include/fst/test/algo_test.h
deleted file mode 100644
index 4c5854a0..00000000
--- a/projects/llm_framework/include/fst/test/algo_test.h
+++ /dev/null
@@ -1,1414 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Regression test for various FST algorithms.
-
-#ifndef FST_TEST_ALGO_TEST_H_
-#define FST_TEST_ALGO_TEST_H_
-
-#include <fst/log.h>
-
-#include <fst/fstlib.h>
-#include <fst/test/rand-fst.h>
-
-DECLARE_int32(repeat);  // defined in ./algo_test.cc
-
-namespace fst {
-
-// Mapper to change input and output label of every transition into
-// epsilons.
-template <class A>
-class EpsMapper {
- public:
-  EpsMapper() {}
-
-  A operator()(const A &arc) const {
-    return A(0, 0, arc.weight, arc.nextstate);
-  }
-
-  uint64 Properties(uint64 props) const {
-    props &= ~kNotAcceptor;
-    props |= kAcceptor;
-    props &= ~kNoIEpsilons & ~kNoOEpsilons & ~kNoEpsilons;
-    props |= kIEpsilons | kOEpsilons | kEpsilons;
-    props &= ~kNotILabelSorted & ~kNotOLabelSorted;
-    props |= kILabelSorted | kOLabelSorted;
-    return props;
-  }
-
-  MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  MapSymbolsAction InputSymbolsAction() const { return MAP_COPY_SYMBOLS; }
-
-  MapSymbolsAction OutputSymbolsAction() const { return MAP_COPY_SYMBOLS; }
-};
-
-// Generic - no lookahead.
-template <class Arc>
-void LookAheadCompose(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-                      MutableFst<Arc> *ofst) {
-  Compose(ifst1, ifst2, ofst);
-}
-
-// Specialized and epsilon olabel acyclic - lookahead.
-void LookAheadCompose(const Fst<StdArc> &ifst1, const Fst<StdArc> &ifst2,
-                      MutableFst<StdArc> *ofst) {
-  std::vector<StdArc::StateId> order;
-  bool acyclic;
-  TopOrderVisitor<StdArc> visitor(&order, &acyclic);
-  DfsVisit(ifst1, &visitor, OutputEpsilonArcFilter<StdArc>());
-  if (acyclic) {  // no ifst1 output epsilon cycles?
-    StdOLabelLookAheadFst lfst1(ifst1);
-    StdVectorFst lfst2(ifst2);
-    LabelLookAheadRelabeler<StdArc>::Relabel(&lfst2, lfst1, true);
-    Compose(lfst1, lfst2, ofst);
-  } else {
-    Compose(ifst1, ifst2, ofst);
-  }
-}
-
-// This class tests a variety of identities and properties that must
-// hold for various algorithms on weighted FSTs.
-template <class Arc, class WeightGenerator>
-class WeightedTester {
- public:
-  typedef typename Arc::Label Label;
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Weight Weight;
-
-  WeightedTester(time_t seed, const Fst<Arc> &zero_fst, const Fst<Arc> &one_fst,
-                 const Fst<Arc> &univ_fst, WeightGenerator *weight_generator)
-      : seed_(seed),
-        zero_fst_(zero_fst),
-        one_fst_(one_fst),
-        univ_fst_(univ_fst),
-        weight_generator_(weight_generator) {}
-
-  void Test(const Fst<Arc> &T1, const Fst<Arc> &T2, const Fst<Arc> &T3) {
-    TestRational(T1, T2, T3);
-    TestMap(T1);
-    TestCompose(T1, T2, T3);
-    TestSort(T1);
-    TestOptimize(T1);
-    TestSearch(T1);
-  }
-
- private:
-  // Tests rational operations with identities
-  void TestRational(const Fst<Arc> &T1, const Fst<Arc> &T2,
-                    const Fst<Arc> &T3) {
-    {
-      VLOG(1) << "Check destructive and delayed union are equivalent.";
-      VectorFst<Arc> U1(T1);
-      Union(&U1, T2);
-      UnionFst<Arc> U2(T1, T2);
-      CHECK(Equiv(U1, U2));
-    }
-
-    {
-      VLOG(1) << "Check destructive and delayed concatenation are equivalent.";
-      VectorFst<Arc> C1(T1);
-      Concat(&C1, T2);
-      ConcatFst<Arc> C2(T1, T2);
-      CHECK(Equiv(C1, C2));
-      VectorFst<Arc> C3(T2);
-      Concat(T1, &C3);
-      CHECK(Equiv(C3, C2));
-    }
-
-    {
-      VLOG(1) << "Check destructive and delayed closure* are equivalent.";
-      VectorFst<Arc> C1(T1);
-      Closure(&C1, CLOSURE_STAR);
-      ClosureFst<Arc> C2(T1, CLOSURE_STAR);
-      CHECK(Equiv(C1, C2));
-    }
-
-    {
-      VLOG(1) << "Check destructive and delayed closure+ are equivalent.";
-      VectorFst<Arc> C1(T1);
-      Closure(&C1, CLOSURE_PLUS);
-      ClosureFst<Arc> C2(T1, CLOSURE_PLUS);
-      CHECK(Equiv(C1, C2));
-    }
-
-    {
-      VLOG(1) << "Check union is associative (destructive).";
-      VectorFst<Arc> U1(T1);
-      Union(&U1, T2);
-      Union(&U1, T3);
-
-      VectorFst<Arc> U3(T2);
-      Union(&U3, T3);
-      VectorFst<Arc> U4(T1);
-      Union(&U4, U3);
-
-      CHECK(Equiv(U1, U4));
-    }
-
-    {
-      VLOG(1) << "Check union is associative (delayed).";
-      UnionFst<Arc> U1(T1, T2);
-      UnionFst<Arc> U2(U1, T3);
-
-      UnionFst<Arc> U3(T2, T3);
-      UnionFst<Arc> U4(T1, U3);
-
-      CHECK(Equiv(U2, U4));
-    }
-
-    {
-      VLOG(1) << "Check union is associative (destructive delayed).";
-      UnionFst<Arc> U1(T1, T2);
-      Union(&U1, T3);
-
-      UnionFst<Arc> U3(T2, T3);
-      UnionFst<Arc> U4(T1, U3);
-
-      CHECK(Equiv(U1, U4));
-    }
-
-    {
-      VLOG(1) << "Check concatenation is associative (destructive).";
-      VectorFst<Arc> C1(T1);
-      Concat(&C1, T2);
-      Concat(&C1, T3);
-
-      VectorFst<Arc> C3(T2);
-      Concat(&C3, T3);
-      VectorFst<Arc> C4(T1);
-      Concat(&C4, C3);
-
-      CHECK(Equiv(C1, C4));
-    }
-
-    {
-      VLOG(1) << "Check concatenation is associative (delayed).";
-      ConcatFst<Arc> C1(T1, T2);
-      ConcatFst<Arc> C2(C1, T3);
-
-      ConcatFst<Arc> C3(T2, T3);
-      ConcatFst<Arc> C4(T1, C3);
-
-      CHECK(Equiv(C2, C4));
-    }
-
-    {
-      VLOG(1) << "Check concatenation is associative (destructive delayed).";
-      ConcatFst<Arc> C1(T1, T2);
-      Concat(&C1, T3);
-
-      ConcatFst<Arc> C3(T2, T3);
-      ConcatFst<Arc> C4(T1, C3);
-
-      CHECK(Equiv(C1, C4));
-    }
-
-    if (Weight::Properties() & kLeftSemiring) {
-      VLOG(1) << "Check concatenation left distributes"
-              << " over union (destructive).";
-
-      VectorFst<Arc> U1(T1);
-      Union(&U1, T2);
-      VectorFst<Arc> C1(T3);
-      Concat(&C1, U1);
-
-      VectorFst<Arc> C2(T3);
-      Concat(&C2, T1);
-      VectorFst<Arc> C3(T3);
-      Concat(&C3, T2);
-      VectorFst<Arc> U2(C2);
-      Union(&U2, C3);
-
-      CHECK(Equiv(C1, U2));
-    }
-
-    if (Weight::Properties() & kRightSemiring) {
-      VLOG(1) << "Check concatenation right distributes"
-              << " over union (destructive).";
-      VectorFst<Arc> U1(T1);
-      Union(&U1, T2);
-      VectorFst<Arc> C1(U1);
-      Concat(&C1, T3);
-
-      VectorFst<Arc> C2(T1);
-      Concat(&C2, T3);
-      VectorFst<Arc> C3(T2);
-      Concat(&C3, T3);
-      VectorFst<Arc> U2(C2);
-      Union(&U2, C3);
-
-      CHECK(Equiv(C1, U2));
-    }
-
-    if (Weight::Properties() & kLeftSemiring) {
-      VLOG(1) << "Check concatenation left distributes over union (delayed).";
-      UnionFst<Arc> U1(T1, T2);
-      ConcatFst<Arc> C1(T3, U1);
-
-      ConcatFst<Arc> C2(T3, T1);
-      ConcatFst<Arc> C3(T3, T2);
-      UnionFst<Arc> U2(C2, C3);
-
-      CHECK(Equiv(C1, U2));
-    }
-
-    if (Weight::Properties() & kRightSemiring) {
-      VLOG(1) << "Check concatenation right distributes over union (delayed).";
-      UnionFst<Arc> U1(T1, T2);
-      ConcatFst<Arc> C1(U1, T3);
-
-      ConcatFst<Arc> C2(T1, T3);
-      ConcatFst<Arc> C3(T2, T3);
-      UnionFst<Arc> U2(C2, C3);
-
-      CHECK(Equiv(C1, U2));
-    }
-
-    if (Weight::Properties() & kLeftSemiring) {
-      VLOG(1) << "Check T T* == T+ (destructive).";
-      VectorFst<Arc> S(T1);
-      Closure(&S, CLOSURE_STAR);
-      VectorFst<Arc> C(T1);
-      Concat(&C, S);
-
-      VectorFst<Arc> P(T1);
-      Closure(&P, CLOSURE_PLUS);
-
-      CHECK(Equiv(C, P));
-    }
-
-    if (Weight::Properties() & kRightSemiring) {
-      VLOG(1) << "Check T* T == T+ (destructive).";
-      VectorFst<Arc> S(T1);
-      Closure(&S, CLOSURE_STAR);
-      VectorFst<Arc> C(S);
-      Concat(&C, T1);
-
-      VectorFst<Arc> P(T1);
-      Closure(&P, CLOSURE_PLUS);
-
-      CHECK(Equiv(C, P));
-    }
-
-    if (Weight::Properties() & kLeftSemiring) {
-      VLOG(1) << "Check T T* == T+ (delayed).";
-      ClosureFst<Arc> S(T1, CLOSURE_STAR);
-      ConcatFst<Arc> C(T1, S);
-
-      ClosureFst<Arc> P(T1, CLOSURE_PLUS);
-
-      CHECK(Equiv(C, P));
-    }
-
-    if (Weight::Properties() & kRightSemiring) {
-      VLOG(1) << "Check T* T == T+ (delayed).";
-      ClosureFst<Arc> S(T1, CLOSURE_STAR);
-      ConcatFst<Arc> C(S, T1);
-
-      ClosureFst<Arc> P(T1, CLOSURE_PLUS);
-
-      CHECK(Equiv(C, P));
-    }
-  }
-
-  // Tests map-based operations.
-  void TestMap(const Fst<Arc> &T) {
-    {
-      VLOG(1) << "Check destructive and delayed projection are equivalent.";
-      VectorFst<Arc> P1(T);
-      Project(&P1, PROJECT_INPUT);
-      ProjectFst<Arc> P2(T, PROJECT_INPUT);
-      CHECK(Equiv(P1, P2));
-    }
-
-    {
-      VLOG(1) << "Check destructive and delayed inversion are equivalent.";
-      VectorFst<Arc> I1(T);
-      Invert(&I1);
-      InvertFst<Arc> I2(T);
-      CHECK(Equiv(I1, I2));
-    }
-
-    {
-      VLOG(1) << "Check Pi_1(T) = Pi_2(T^-1) (destructive).";
-      VectorFst<Arc> P1(T);
-      VectorFst<Arc> I1(T);
-      Project(&P1, PROJECT_INPUT);
-      Invert(&I1);
-      Project(&I1, PROJECT_OUTPUT);
-      CHECK(Equiv(P1, I1));
-    }
-
-    {
-      VLOG(1) << "Check Pi_2(T) = Pi_1(T^-1) (destructive).";
-      VectorFst<Arc> P1(T);
-      VectorFst<Arc> I1(T);
-      Project(&P1, PROJECT_OUTPUT);
-      Invert(&I1);
-      Project(&I1, PROJECT_INPUT);
-      CHECK(Equiv(P1, I1));
-    }
-
-    {
-      VLOG(1) << "Check Pi_1(T) = Pi_2(T^-1) (delayed).";
-      ProjectFst<Arc> P1(T, PROJECT_INPUT);
-      InvertFst<Arc> I1(T);
-      ProjectFst<Arc> P2(I1, PROJECT_OUTPUT);
-      CHECK(Equiv(P1, P2));
-    }
-
-    {
-      VLOG(1) << "Check Pi_2(T) = Pi_1(T^-1) (delayed).";
-      ProjectFst<Arc> P1(T, PROJECT_OUTPUT);
-      InvertFst<Arc> I1(T);
-      ProjectFst<Arc> P2(I1, PROJECT_INPUT);
-      CHECK(Equiv(P1, P2));
-    }
-
-    {
-      VLOG(1) << "Check destructive relabeling";
-      static const int kNumLabels = 10;
-      // set up relabeling pairs
-      std::vector<Label> labelset(kNumLabels);
-      for (size_t i = 0; i < kNumLabels; ++i) labelset[i] = i;
-      for (size_t i = 0; i < kNumLabels; ++i) {
-        using std::swap;
-        swap(labelset[i], labelset[rand() % kNumLabels]);
-      }
-
-      std::vector<std::pair<Label, Label>> ipairs1(kNumLabels);
-      std::vector<std::pair<Label, Label>> opairs1(kNumLabels);
-      for (size_t i = 0; i < kNumLabels; ++i) {
-        ipairs1[i] = std::make_pair(i, labelset[i]);
-        opairs1[i] = std::make_pair(labelset[i], i);
-      }
-      VectorFst<Arc> R(T);
-      Relabel(&R, ipairs1, opairs1);
-
-      std::vector<std::pair<Label, Label>> ipairs2(kNumLabels);
-      std::vector<std::pair<Label, Label>> opairs2(kNumLabels);
-      for (size_t i = 0; i < kNumLabels; ++i) {
-        ipairs2[i] = std::make_pair(labelset[i], i);
-        opairs2[i] = std::make_pair(i, labelset[i]);
-      }
-      Relabel(&R, ipairs2, opairs2);
-      CHECK(Equiv(R, T));
-
-      VLOG(1) << "Check on-the-fly relabeling";
-      RelabelFst<Arc> Rdelay(T, ipairs1, opairs1);
-
-      RelabelFst<Arc> RRdelay(Rdelay, ipairs2, opairs2);
-      CHECK(Equiv(RRdelay, T));
-    }
-
-    {
-      VLOG(1) << "Check encoding/decoding (destructive).";
-      VectorFst<Arc> D(T);
-      uint32 encode_props = 0;
-      if (rand() % 2) encode_props |= kEncodeLabels;
-      if (rand() % 2) encode_props |= kEncodeWeights;
-      EncodeMapper<Arc> encoder(encode_props, ENCODE);
-      Encode(&D, &encoder);
-      Decode(&D, encoder);
-      CHECK(Equiv(D, T));
-    }
-
-    {
-      VLOG(1) << "Check encoding/decoding (delayed).";
-      uint32 encode_props = 0;
-      if (rand() % 2) encode_props |= kEncodeLabels;
-      if (rand() % 2) encode_props |= kEncodeWeights;
-      EncodeMapper<Arc> encoder(encode_props, ENCODE);
-      EncodeFst<Arc> E(T, &encoder);
-      VectorFst<Arc> Encoded(E);
-      DecodeFst<Arc> D(Encoded, encoder);
-      CHECK(Equiv(D, T));
-    }
-
-    {
-      VLOG(1) << "Check gallic mappers (constructive).";
-      ToGallicMapper<Arc> to_mapper;
-      FromGallicMapper<Arc> from_mapper;
-      VectorFst<GallicArc<Arc>> G;
-      VectorFst<Arc> F;
-      ArcMap(T, &G, to_mapper);
-      ArcMap(G, &F, from_mapper);
-      CHECK(Equiv(T, F));
-    }
-
-    {
-      VLOG(1) << "Check gallic mappers (delayed).";
-      ToGallicMapper<Arc> to_mapper;
-      FromGallicMapper<Arc> from_mapper;
-      ArcMapFst<Arc, GallicArc<Arc>, ToGallicMapper<Arc>> G(T, to_mapper);
-      ArcMapFst<GallicArc<Arc>, Arc, FromGallicMapper<Arc>> F(G, from_mapper);
-      CHECK(Equiv(T, F));
-    }
-  }
-
-  // Tests compose-based operations.
-  void TestCompose(const Fst<Arc> &T1, const Fst<Arc> &T2, const Fst<Arc> &T3) {
-    if (!(Weight::Properties() & kCommutative)) return;
-
-    VectorFst<Arc> S1(T1);
-    VectorFst<Arc> S2(T2);
-    VectorFst<Arc> S3(T3);
-
-    ILabelCompare<Arc> icomp;
-    OLabelCompare<Arc> ocomp;
-
-    ArcSort(&S1, ocomp);
-    ArcSort(&S2, ocomp);
-    ArcSort(&S3, icomp);
-
-    {
-      VLOG(1) << "Check composition is associative.";
-      ComposeFst<Arc> C1(S1, S2);
-      ComposeFst<Arc> C2(C1, S3);
-      ComposeFst<Arc> C3(S2, S3);
-      ComposeFst<Arc> C4(S1, C3);
-
-      CHECK(Equiv(C2, C4));
-    }
-
-    {
-      VLOG(1) << "Check composition left distributes over union.";
-      UnionFst<Arc> U1(S2, S3);
-      ComposeFst<Arc> C1(S1, U1);
-
-      ComposeFst<Arc> C2(S1, S2);
-      ComposeFst<Arc> C3(S1, S3);
-      UnionFst<Arc> U2(C2, C3);
-
-      CHECK(Equiv(C1, U2));
-    }
-
-    {
-      VLOG(1) << "Check composition right distributes over union.";
-      UnionFst<Arc> U1(S1, S2);
-      ComposeFst<Arc> C1(U1, S3);
-
-      ComposeFst<Arc> C2(S1, S3);
-      ComposeFst<Arc> C3(S2, S3);
-      UnionFst<Arc> U2(C2, C3);
-
-      CHECK(Equiv(C1, U2));
-    }
-
-    VectorFst<Arc> A1(S1);
-    VectorFst<Arc> A2(S2);
-    VectorFst<Arc> A3(S3);
-    Project(&A1, PROJECT_OUTPUT);
-    Project(&A2, PROJECT_INPUT);
-    Project(&A3, PROJECT_INPUT);
-
-    {
-      VLOG(1) << "Check intersection is commutative.";
-      IntersectFst<Arc> I1(A1, A2);
-      IntersectFst<Arc> I2(A2, A1);
-      CHECK(Equiv(I1, I2));
-    }
-
-    {
-      VLOG(1) << "Check all epsilon filters leads to equivalent results.";
-      typedef Matcher<Fst<Arc>> M;
-      ComposeFst<Arc> C1(S1, S2);
-      ComposeFst<Arc> C2(
-          S1, S2, ComposeFstOptions<Arc, M, AltSequenceComposeFilter<M>>());
-      ComposeFst<Arc> C3(S1, S2,
-                         ComposeFstOptions<Arc, M, MatchComposeFilter<M>>());
-
-      CHECK(Equiv(C1, C2));
-      CHECK(Equiv(C1, C3));
-
-      if ((Weight::Properties() & kIdempotent) ||
-          S1.Properties(kNoOEpsilons, false) ||
-          S2.Properties(kNoIEpsilons, false)) {
-        ComposeFst<Arc> C4(
-            S1, S2, ComposeFstOptions<Arc, M, TrivialComposeFilter<M>>());
-        CHECK(Equiv(C1, C4));
-        ComposeFst<Arc> C5(
-            S1, S2, ComposeFstOptions<Arc, M, NoMatchComposeFilter<M>>());
-        CHECK(Equiv(C1, C5));
-      }
-
-      if (S1.Properties(kNoOEpsilons, false) &&
-          S2.Properties(kNoIEpsilons, false)) {
-        ComposeFst<Arc> C6(S1, S2,
-                           ComposeFstOptions<Arc, M, NullComposeFilter<M>>());
-        CHECK(Equiv(C1, C6));
-      }
-    }
-
-    {
-      VLOG(1) << "Check look-ahead filters lead to equivalent results.";
-      VectorFst<Arc> C1, C2;
-      Compose(S1, S2, &C1);
-      LookAheadCompose(S1, S2, &C2);
-      CHECK(Equiv(C1, C2));
-    }
-  }
-
-  // Tests sorting operations
-  void TestSort(const Fst<Arc> &T) {
-    ILabelCompare<Arc> icomp;
-    OLabelCompare<Arc> ocomp;
-
-    {
-      VLOG(1) << "Check arc sorted Fst is equivalent to its input.";
-      VectorFst<Arc> S1(T);
-      ArcSort(&S1, icomp);
-      CHECK(Equiv(T, S1));
-    }
-
-    {
-      VLOG(1) << "Check destructive and delayed arcsort are equivalent.";
-      VectorFst<Arc> S1(T);
-      ArcSort(&S1, icomp);
-      ArcSortFst<Arc, ILabelCompare<Arc>> S2(T, icomp);
-      CHECK(Equiv(S1, S2));
-    }
-
-    {
-      VLOG(1) << "Check ilabel sorting vs. olabel sorting with inversions.";
-      VectorFst<Arc> S1(T);
-      VectorFst<Arc> S2(T);
-      ArcSort(&S1, icomp);
-      Invert(&S2);
-      ArcSort(&S2, ocomp);
-      Invert(&S2);
-      CHECK(Equiv(S1, S2));
-    }
-
-    {
-      VLOG(1) << "Check topologically sorted Fst is equivalent to its input.";
-      VectorFst<Arc> S1(T);
-      TopSort(&S1);
-      CHECK(Equiv(T, S1));
-    }
-
-    {
-      VLOG(1) << "Check reverse(reverse(T)) = T";
-      for (int i = 0; i < 2; ++i) {
-        VectorFst<ReverseArc<Arc>> R1;
-        VectorFst<Arc> R2;
-        bool require_superinitial = i == 1;
-        Reverse(T, &R1, require_superinitial);
-        Reverse(R1, &R2, require_superinitial);
-        CHECK(Equiv(T, R2));
-      }
-    }
-  }
-
-  // Tests optimization operations
-  void TestOptimize(const Fst<Arc> &T) {
-    uint64 tprops = T.Properties(kFstProperties, true);
-    uint64 wprops = Weight::Properties();
-
-    VectorFst<Arc> A(T);
-    Project(&A, PROJECT_INPUT);
-    {
-      VLOG(1) << "Check connected FST is equivalent to its input.";
-      VectorFst<Arc> C1(T);
-      Connect(&C1);
-      CHECK(Equiv(T, C1));
-    }
-
-    if ((wprops & kSemiring) == kSemiring &&
-        (tprops & kAcyclic || wprops & kIdempotent)) {
-      VLOG(1) << "Check epsilon-removed FST is equivalent to its input.";
-      VectorFst<Arc> R1(T);
-      RmEpsilon(&R1);
-      CHECK(Equiv(T, R1));
-
-      VLOG(1) << "Check destructive and delayed epsilon removal"
-              << "are equivalent.";
-      RmEpsilonFst<Arc> R2(T);
-      CHECK(Equiv(R1, R2));
-
-      VLOG(1) << "Check an FST with a large proportion"
-              << " of epsilon transitions:";
-      // Maps all transitions of T to epsilon-transitions and append
-      // a non-epsilon transition.
-      VectorFst<Arc> U;
-      ArcMap(T, &U, EpsMapper<Arc>());
-      VectorFst<Arc> V;
-      V.SetStart(V.AddState());
-      Arc arc(1, 1, Weight::One(), V.AddState());
-      V.AddArc(V.Start(), arc);
-      V.SetFinal(arc.nextstate, Weight::One());
-      Concat(&U, V);
-      // Check that epsilon-removal preserves the shortest-distance
-      // from the initial state to the final states.
-      std::vector<Weight> d;
-      ShortestDistance(U, &d, true);
-      Weight w = U.Start() < d.size() ? d[U.Start()] : Weight::Zero();
-      VectorFst<Arc> U1(U);
-      RmEpsilon(&U1);
-      ShortestDistance(U1, &d, true);
-      Weight w1 = U1.Start() < d.size() ? d[U1.Start()] : Weight::Zero();
-      CHECK(ApproxEqual(w, w1, kTestDelta));
-      RmEpsilonFst<Arc> U2(U);
-      ShortestDistance(U2, &d, true);
-      Weight w2 = U2.Start() < d.size() ? d[U2.Start()] : Weight::Zero();
-      CHECK(ApproxEqual(w, w2, kTestDelta));
-    }
-
-    if ((wprops & kSemiring) == kSemiring && tprops & kAcyclic) {
-      VLOG(1) << "Check determinized FSA is equivalent to its input.";
-      DeterminizeFst<Arc> D(A);
-      CHECK(Equiv(A, D));
-
-      {
-        VLOG(1) << "Check determinized FST is equivalent to its input.";
-        DeterminizeFstOptions<Arc> opts;
-        opts.type = DETERMINIZE_NONFUNCTIONAL;
-        DeterminizeFst<Arc> DT(T, opts);
-        CHECK(Equiv(T, DT));
-      }
-
-      if ((wprops & (kPath | kCommutative)) == (kPath | kCommutative)) {
-        VLOG(1) << "Check pruning in determinization";
-        VectorFst<Arc> P;
-        Weight threshold = (*weight_generator_)();
-        DeterminizeOptions<Arc> opts;
-        opts.weight_threshold = threshold;
-        Determinize(A, &P, opts);
-        CHECK(P.Properties(kIDeterministic, true));
-        CHECK(PruneEquiv(A, P, threshold));
-      }
-
-      if ((wprops & kPath) == kPath) {
-        VLOG(1) << "Check min-determinization";
-
-        // Ensures no input epsilons
-        VectorFst<Arc> R(T);
-        std::vector<std::pair<Label, Label>> ipairs, opairs;
-        ipairs.push_back(std::pair<Label, Label>(0, 1));
-        Relabel(&R, ipairs, opairs);
-
-        VectorFst<Arc> M;
-        DeterminizeOptions<Arc> opts;
-        opts.type = DETERMINIZE_DISAMBIGUATE;
-        Determinize(R, &M, opts);
-        CHECK(M.Properties(kIDeterministic, true));
-        CHECK(MinRelated(M, R));
-      }
-
-      int n;
-      {
-        VLOG(1) << "Check size(min(det(A))) <= size(det(A))"
-                << " and  min(det(A)) equiv det(A)";
-        VectorFst<Arc> M(D);
-        n = M.NumStates();
-        Minimize(&M, static_cast<MutableFst<Arc> *>(nullptr), kDelta);
-        CHECK(Equiv(D, M));
-        CHECK(M.NumStates() <= n);
-        n = M.NumStates();
-      }
-
-      if (n && (wprops & kIdempotent) == kIdempotent &&
-          A.Properties(kNoEpsilons, true)) {
-        VLOG(1) << "Check that Revuz's algorithm leads to the"
-                << " same number of states as Brozozowski's algorithm";
-
-        // Skip test if A is the empty machine or contains epsilons or
-        // if the semiring is not idempotent (to avoid floating point
-        // errors)
-        VectorFst<Arc> R;
-        Reverse(A, &R);
-        RmEpsilon(&R);
-        DeterminizeFst<Arc> DR(R);
-        VectorFst<Arc> RD;
-        Reverse(DR, &RD);
-        DeterminizeFst<Arc> DRD(RD);
-        VectorFst<Arc> M(DRD);
-        CHECK_EQ(n + 1, M.NumStates());  // Accounts for the epsilon transition
-                                         // to the initial state
-      }
-    }
-
-    if ((wprops & kSemiring) == kSemiring && tprops & kAcyclic) {
-      VLOG(1) << "Check disambiguated FSA is equivalent to its input.";
-      VectorFst<Arc> R(A), D;
-      RmEpsilon(&R);
-      Disambiguate(R, &D);
-      CHECK(Equiv(R, D));
-      VLOG(1) << "Check disambiguated FSA is unambiguous";
-      CHECK(Unambiguous(D));
-
-      /* TODO(riley): find out why this fails
-      if ((wprops & (kPath | kCommutative)) == (kPath | kCommutative)) {
-        VLOG(1)  << "Check pruning in disambiguation";
-        VectorFst<Arc> P;
-        Weight threshold = (*weight_generator_)();
-        DisambiguateOptions<Arc> opts;
-        opts.weight_threshold = threshold;
-        Disambiguate(R, &P, opts);
-        CHECK(Unambiguous(P));
-        CHECK(PruneEquiv(A, P, threshold));
-      }
-      */
-    }
-
-    if (Arc::Type() == LogArc::Type() || Arc::Type() == StdArc::Type()) {
-      VLOG(1) << "Check reweight(T) equiv T";
-      std::vector<Weight> potential;
-      VectorFst<Arc> RI(T);
-      VectorFst<Arc> RF(T);
-      while (potential.size() < RI.NumStates())
-        potential.push_back((*weight_generator_)());
-
-      Reweight(&RI, potential, REWEIGHT_TO_INITIAL);
-      CHECK(Equiv(T, RI));
-
-      Reweight(&RF, potential, REWEIGHT_TO_FINAL);
-      CHECK(Equiv(T, RF));
-    }
-
-    if ((wprops & kIdempotent) || (tprops & kAcyclic)) {
-      VLOG(1) << "Check pushed FST is equivalent to input FST.";
-      // Pushing towards the final state.
-      if (wprops & kRightSemiring) {
-        VectorFst<Arc> P1;
-        Push<Arc, REWEIGHT_TO_FINAL>(T, &P1, kPushLabels);
-        CHECK(Equiv(T, P1));
-
-        VectorFst<Arc> P2;
-        Push<Arc, REWEIGHT_TO_FINAL>(T, &P2, kPushWeights);
-        CHECK(Equiv(T, P2));
-
-        VectorFst<Arc> P3;
-        Push<Arc, REWEIGHT_TO_FINAL>(T, &P3, kPushLabels | kPushWeights);
-        CHECK(Equiv(T, P3));
-      }
-
-      // Pushing towards the initial state.
-      if (wprops & kLeftSemiring) {
-        VectorFst<Arc> P1;
-        Push<Arc, REWEIGHT_TO_INITIAL>(T, &P1, kPushLabels);
-        CHECK(Equiv(T, P1));
-
-        VectorFst<Arc> P2;
-        Push<Arc, REWEIGHT_TO_INITIAL>(T, &P2, kPushWeights);
-        CHECK(Equiv(T, P2));
-        VectorFst<Arc> P3;
-        Push<Arc, REWEIGHT_TO_INITIAL>(T, &P3, kPushLabels | kPushWeights);
-        CHECK(Equiv(T, P3));
-      }
-    }
-
-    if ((wprops & (kPath | kCommutative)) == (kPath | kCommutative)) {
-      VLOG(1) << "Check pruning algorithm";
-      {
-        VLOG(1) << "Check equiv. of constructive and destructive algorithms";
-        Weight thresold = (*weight_generator_)();
-        VectorFst<Arc> P1(T);
-        Prune(&P1, thresold);
-        VectorFst<Arc> P2;
-        Prune(T, &P2, thresold);
-        CHECK(Equiv(P1, P2));
-      }
-
-      {
-        VLOG(1) << "Check prune(reverse) equiv reverse(prune)";
-        Weight thresold = (*weight_generator_)();
-        VectorFst<ReverseArc<Arc>> R;
-        VectorFst<Arc> P1(T);
-        VectorFst<Arc> P2;
-        Prune(&P1, thresold);
-        Reverse(T, &R);
-        Prune(&R, thresold.Reverse());
-        Reverse(R, &P2);
-        CHECK(Equiv(P1, P2));
-      }
-      {
-        VLOG(1) << "Check: ShortestDistance(A - prune(A))"
-                << " > ShortestDistance(A) times Threshold";
-        Weight threshold = (*weight_generator_)();
-        VectorFst<Arc> P;
-        Prune(A, &P, threshold);
-        CHECK(PruneEquiv(A, P, threshold));
-      }
-    }
-    if (tprops & kAcyclic) {
-      VLOG(1) << "Check synchronize(T) equiv T";
-      SynchronizeFst<Arc> S(T);
-      CHECK(Equiv(T, S));
-    }
-  }
-
-  // Tests search operations
-  void TestSearch(const Fst<Arc> &T) {
-    uint64 wprops = Weight::Properties();
-
-    VectorFst<Arc> A(T);
-    Project(&A, PROJECT_INPUT);
-
-    if ((wprops & (kPath | kRightSemiring)) == (kPath | kRightSemiring)) {
-      VLOG(1) << "Check 1-best weight.";
-      VectorFst<Arc> path;
-      ShortestPath(T, &path);
-      Weight tsum = ShortestDistance(T);
-      Weight psum = ShortestDistance(path);
-      CHECK(ApproxEqual(tsum, psum, kTestDelta));
-    }
-
-    if ((wprops & (kPath | kSemiring)) == (kPath | kSemiring)) {
-      VLOG(1) << "Check n-best weights";
-      VectorFst<Arc> R(A);
-      RmEpsilon(&R, /*connect=*/ true, Arc::Weight::Zero(), kNoStateId,
-                kDelta);
-      int nshortest = rand() % kNumRandomShortestPaths + 2;
-      VectorFst<Arc> paths;
-      ShortestPath(R, &paths, nshortest, /*unique=*/ true,
-                   /*first_path=*/ false, Weight::Zero(), kNumShortestStates,
-                   kDelta);
-      std::vector<Weight> distance;
-      ShortestDistance(paths, &distance, true, kDelta);
-      StateId pstart = paths.Start();
-      if (pstart != kNoStateId) {
-        ArcIterator<Fst<Arc>> piter(paths, pstart);
-        for (; !piter.Done(); piter.Next()) {
-          StateId s = piter.Value().nextstate;
-          Weight nsum = s < distance.size()
-                            ? Times(piter.Value().weight, distance[s])
-                            : Weight::Zero();
-          VectorFst<Arc> path;
-          ShortestPath(R, &path, 1, false, false, Weight::Zero(), kNoStateId,
-                       kDelta);
-          Weight dsum = ShortestDistance(path, kDelta);
-          CHECK(ApproxEqual(nsum, dsum, kTestDelta));
-          ArcMap(&path, RmWeightMapper<Arc>());
-          VectorFst<Arc> S;
-          Difference(R, path, &S);
-          R = S;
-        }
-      }
-    }
-  }
-
-  // Tests if two FSTS are equivalent by checking if random
-  // strings from one FST are transduced the same by both FSTs.
-  template <class A>
-  bool Equiv(const Fst<A> &fst1, const Fst<A> &fst2) {
-    VLOG(1) << "Check FSTs for sanity (including property bits).";
-    CHECK(Verify(fst1));
-    CHECK(Verify(fst2));
-
-    // Ensures seed used once per instantiation.
-    static UniformArcSelector<A> uniform_selector(seed_);
-    RandGenOptions<UniformArcSelector<A>> opts(uniform_selector,
-                                               kRandomPathLength);
-    return RandEquivalent(fst1, fst2, kNumRandomPaths, kTestDelta, opts);
-  }
-
-  // Tests FSA is unambiguous
-  bool Unambiguous(const Fst<Arc> &fst) {
-    VectorFst<StdArc> sfst, dfst;
-    VectorFst<LogArc> lfst1, lfst2;
-    Map(fst, &sfst, RmWeightMapper<Arc, StdArc>());
-    Determinize(sfst, &dfst);
-    Map(fst, &lfst1, RmWeightMapper<Arc, LogArc>());
-    Map(dfst, &lfst2, RmWeightMapper<StdArc, LogArc>());
-    return Equiv(lfst1, lfst2);
-  }
-
-  // Ensures input-epsilon free transducers fst1 and fst2 have the
-  // same domain and that for each string pair '(is, os)' in fst1,
-  // '(is, os)' is the minimum weight match to 'is' in fst2.
-  template <class A>
-  bool MinRelated(const Fst<A> &fst1, const Fst<A> &fst2) {
-    // Same domain
-    VectorFst<Arc> P1(fst1), P2(fst2);
-    Project(&P1, PROJECT_INPUT);
-    Project(&P2, PROJECT_INPUT);
-    if (!Equiv(P1, P2)) {
-      LOG(ERROR) << "Inputs not equivalent";
-      return false;
-    }
-
-    // Ensures seed used once per instantiation.
-    static UniformArcSelector<A> uniform_selector(seed_);
-    RandGenOptions<UniformArcSelector<A>> opts(uniform_selector,
-                                               kRandomPathLength);
-
-    VectorFst<Arc> path, paths1, paths2;
-    for (ssize_t n = 0; n < kNumRandomPaths; ++n) {
-      RandGen(fst1, &path, opts);
-      Invert(&path);
-      Map(&path, RmWeightMapper<Arc>());
-      Compose(path, fst2, &paths1);
-      Weight sum1 = ShortestDistance(paths1);
-      Compose(paths1, path, &paths2);
-      Weight sum2 = ShortestDistance(paths2);
-      if (!ApproxEqual(Plus(sum1, sum2), sum2, kTestDelta)) {
-        LOG(ERROR) << "Sums not equivalent: " << sum1 << " " << sum2;
-        return false;
-      }
-    }
-    return true;
-  }
-
-  // Tests ShortestDistance(A - P) >=
-  // ShortestDistance(A) times Threshold.
-  template <class A>
-  bool PruneEquiv(const Fst<A> &fst, const Fst<A> &pfst, Weight threshold) {
-    VLOG(1) << "Check FSTs for sanity (including property bits).";
-    CHECK(Verify(fst));
-    CHECK(Verify(pfst));
-
-    DifferenceFst<Arc> D(fst, DeterminizeFst<Arc>(RmEpsilonFst<Arc>(
-                                  ArcMapFst<Arc, Arc, RmWeightMapper<Arc>>(
-                                      pfst, RmWeightMapper<Arc>()))));
-    Weight sum1 = Times(ShortestDistance(fst), threshold);
-    Weight sum2 = ShortestDistance(D);
-    return ApproxEqual(Plus(sum1, sum2), sum1, kTestDelta);
-  }
-
-  // Random seed.
-  int seed_;
-  // FST with no states
-  VectorFst<Arc> zero_fst_;
-  // FST with one state that accepts epsilon.
-  VectorFst<Arc> one_fst_;
-  // FST with one state that accepts all strings.
-  VectorFst<Arc> univ_fst_;
-  // Generates weights used in testing.
-  WeightGenerator *weight_generator_;
-  // Maximum random path length.
-  static const int kRandomPathLength;
-  // Number of random paths to explore.
-  static const int kNumRandomPaths;
-  // Maximum number of nshortest paths.
-  static const int kNumRandomShortestPaths;
-  // Maximum number of nshortest states.
-  static const int kNumShortestStates;
-  // Delta for equivalence tests.
-  static const float kTestDelta;
-
-  WeightedTester(const WeightedTester &) = delete;
-  WeightedTester &operator=(const WeightedTester &) = delete;
-};
-
-template <class A, class WG>
-const int WeightedTester<A, WG>::kRandomPathLength = 25;
-
-template <class A, class WG>
-const int WeightedTester<A, WG>::kNumRandomPaths = 100;
-
-template <class A, class WG>
-const int WeightedTester<A, WG>::kNumRandomShortestPaths = 100;
-
-template <class A, class WG>
-const int WeightedTester<A, WG>::kNumShortestStates = 10000;
-
-template <class A, class WG>
-const float WeightedTester<A, WG>::kTestDelta = .05;
-
-// This class tests a variety of identities and properties that must
-// hold for various algorithms on unweighted FSAs and that are not tested
-// by WeightedTester. Only the specialization does anything interesting.
-template <class Arc>
-class UnweightedTester {
- public:
-  UnweightedTester(const Fst<Arc> &zero_fsa, const Fst<Arc> &one_fsa,
-                   const Fst<Arc> &univ_fsa) {}
-
-  void Test(const Fst<Arc> &A1, const Fst<Arc> &A2, const Fst<Arc> &A3) {}
-};
-
-// Specialization for StdArc. This should work for any commutative,
-// idempotent semiring when restricted to the unweighted case
-// (being isomorphic to the boolean semiring).
-template <>
-class UnweightedTester<StdArc> {
- public:
-  typedef StdArc Arc;
-  typedef Arc::Label Label;
-  typedef Arc::StateId StateId;
-  typedef Arc::Weight Weight;
-
-  UnweightedTester(const Fst<Arc> &zero_fsa, const Fst<Arc> &one_fsa,
-                   const Fst<Arc> &univ_fsa)
-      : zero_fsa_(zero_fsa), one_fsa_(one_fsa), univ_fsa_(univ_fsa) {}
-
-  void Test(const Fst<Arc> &A1, const Fst<Arc> &A2, const Fst<Arc> &A3) {
-    TestRational(A1, A2, A3);
-    TestIntersect(A1, A2, A3);
-    TestOptimize(A1);
-  }
-
- private:
-  // Tests rational operations with identities
-  void TestRational(const Fst<Arc> &A1, const Fst<Arc> &A2,
-                    const Fst<Arc> &A3) {
-    {
-      VLOG(1) << "Check the union contains its arguments (destructive).";
-      VectorFst<Arc> U(A1);
-      Union(&U, A2);
-
-      CHECK(Subset(A1, U));
-      CHECK(Subset(A2, U));
-    }
-
-    {
-      VLOG(1) << "Check the union contains its arguments (delayed).";
-      UnionFst<Arc> U(A1, A2);
-
-      CHECK(Subset(A1, U));
-      CHECK(Subset(A2, U));
-    }
-
-    {
-      VLOG(1) << "Check if A^n c A* (destructive).";
-      VectorFst<Arc> C(one_fsa_);
-      int n = rand() % 5;
-      for (int i = 0; i < n; ++i) Concat(&C, A1);
-
-      VectorFst<Arc> S(A1);
-      Closure(&S, CLOSURE_STAR);
-      CHECK(Subset(C, S));
-    }
-
-    {
-      VLOG(1) << "Check if A^n c A* (delayed).";
-      int n = rand() % 5;
-      Fst<Arc> *C = new VectorFst<Arc>(one_fsa_);
-      for (int i = 0; i < n; ++i) {
-        ConcatFst<Arc> *F = new ConcatFst<Arc>(*C, A1);
-        delete C;
-        C = F;
-      }
-      ClosureFst<Arc> S(A1, CLOSURE_STAR);
-      CHECK(Subset(*C, S));
-      delete C;
-    }
-  }
-
-  // Tests intersect-based operations.
-  void TestIntersect(const Fst<Arc> &A1, const Fst<Arc> &A2,
-                     const Fst<Arc> &A3) {
-    VectorFst<Arc> S1(A1);
-    VectorFst<Arc> S2(A2);
-    VectorFst<Arc> S3(A3);
-
-    ILabelCompare<Arc> comp;
-
-    ArcSort(&S1, comp);
-    ArcSort(&S2, comp);
-    ArcSort(&S3, comp);
-
-    {
-      VLOG(1) << "Check the intersection is contained in its arguments.";
-      IntersectFst<Arc> I1(S1, S2);
-      CHECK(Subset(I1, S1));
-      CHECK(Subset(I1, S2));
-    }
-
-    {
-      VLOG(1) << "Check union distributes over intersection.";
-      IntersectFst<Arc> I1(S1, S2);
-      UnionFst<Arc> U1(I1, S3);
-
-      UnionFst<Arc> U2(S1, S3);
-      UnionFst<Arc> U3(S2, S3);
-      ArcSortFst<Arc, ILabelCompare<Arc>> S4(U3, comp);
-      IntersectFst<Arc> I2(U2, S4);
-
-      CHECK(Equiv(U1, I2));
-    }
-
-    VectorFst<Arc> C1;
-    VectorFst<Arc> C2;
-    Complement(S1, &C1);
-    Complement(S2, &C2);
-    ArcSort(&C1, comp);
-    ArcSort(&C2, comp);
-
-    {
-      VLOG(1) << "Check S U S' = Sigma*";
-      UnionFst<Arc> U(S1, C1);
-      CHECK(Equiv(U, univ_fsa_));
-    }
-
-    {
-      VLOG(1) << "Check S n S' = {}";
-      IntersectFst<Arc> I(S1, C1);
-      CHECK(Equiv(I, zero_fsa_));
-    }
-
-    {
-      VLOG(1) << "Check (S1' U S2') == (S1 n S2)'";
-      UnionFst<Arc> U(C1, C2);
-
-      IntersectFst<Arc> I(S1, S2);
-      VectorFst<Arc> C3;
-      Complement(I, &C3);
-      CHECK(Equiv(U, C3));
-    }
-
-    {
-      VLOG(1) << "Check (S1' n S2') == (S1 U S2)'";
-      IntersectFst<Arc> I(C1, C2);
-
-      UnionFst<Arc> U(S1, S2);
-      VectorFst<Arc> C3;
-      Complement(U, &C3);
-      CHECK(Equiv(I, C3));
-    }
-  }
-
-  // Tests optimization operations
-  void TestOptimize(const Fst<Arc> &A) {
-    {
-      VLOG(1) << "Check determinized FSA is equivalent to its input.";
-      DeterminizeFst<Arc> D(A);
-      CHECK(Equiv(A, D));
-    }
-
-    {
-      VLOG(1) << "Check disambiguated FSA is equivalent to its input.";
-      VectorFst<Arc> R(A), D;
-      RmEpsilon(&R);
-
-      Disambiguate(R, &D);
-      CHECK(Equiv(R, D));
-    }
-
-    {
-      VLOG(1) << "Check minimized FSA is equivalent to its input.";
-      int n;
-      {
-        RmEpsilonFst<Arc> R(A);
-        DeterminizeFst<Arc> D(R);
-        VectorFst<Arc> M(D);
-        Minimize(&M, static_cast<MutableFst<Arc> *>(nullptr), kDelta);
-        CHECK(Equiv(A, M));
-        n = M.NumStates();
-      }
-
-      if (n) {  // Skip test if A is the empty machine
-        VLOG(1) << "Check that Hopcroft's and Revuz's algorithms lead to the"
-                << " same number of states as Brozozowski's algorithm";
-        VectorFst<Arc> R;
-        Reverse(A, &R);
-        RmEpsilon(&R);
-        DeterminizeFst<Arc> DR(R);
-        VectorFst<Arc> RD;
-        Reverse(DR, &RD);
-        DeterminizeFst<Arc> DRD(RD);
-        VectorFst<Arc> M(DRD);
-        CHECK_EQ(n + 1, M.NumStates());  // Accounts for the epsilon transition
-                                         // to the initial state
-      }
-    }
-  }
-
-  // Tests if two FSAS are equivalent.
-  bool Equiv(const Fst<Arc> &fsa1, const Fst<Arc> &fsa2) {
-    VLOG(1) << "Check FSAs for sanity (including property bits).";
-    CHECK(Verify(fsa1));
-    CHECK(Verify(fsa2));
-
-    VectorFst<Arc> vfsa1(fsa1);
-    VectorFst<Arc> vfsa2(fsa2);
-    RmEpsilon(&vfsa1);
-    RmEpsilon(&vfsa2);
-    DeterminizeFst<Arc> dfa1(vfsa1);
-    DeterminizeFst<Arc> dfa2(vfsa2);
-
-    // Test equivalence using union-find algorithm
-    bool equiv1 = Equivalent(dfa1, dfa2);
-
-    // Test equivalence by checking if (S1 - S2) U (S2 - S1) is empty
-    ILabelCompare<Arc> comp;
-    VectorFst<Arc> sdfa1(dfa1);
-    ArcSort(&sdfa1, comp);
-    VectorFst<Arc> sdfa2(dfa2);
-    ArcSort(&sdfa2, comp);
-
-    DifferenceFst<Arc> dfsa1(sdfa1, sdfa2);
-    DifferenceFst<Arc> dfsa2(sdfa2, sdfa1);
-
-    VectorFst<Arc> ufsa(dfsa1);
-    Union(&ufsa, dfsa2);
-    Connect(&ufsa);
-    bool equiv2 = ufsa.NumStates() == 0;
-
-    // Check two equivalence tests match
-    CHECK((equiv1 && equiv2) || (!equiv1 && !equiv2));
-
-    return equiv1;
-  }
-
-  // Tests if FSA1 is a subset of FSA2 (disregarding weights).
-  bool Subset(const Fst<Arc> &fsa1, const Fst<Arc> &fsa2) {
-    VLOG(1) << "Check FSAs (incl. property bits) for sanity";
-    CHECK(Verify(fsa1));
-    CHECK(Verify(fsa2));
-
-    VectorFst<StdArc> vfsa1;
-    VectorFst<StdArc> vfsa2;
-    RmEpsilon(&vfsa1);
-    RmEpsilon(&vfsa2);
-    ILabelCompare<StdArc> comp;
-    ArcSort(&vfsa1, comp);
-    ArcSort(&vfsa2, comp);
-    IntersectFst<StdArc> ifsa(vfsa1, vfsa2);
-    DeterminizeFst<StdArc> dfa1(vfsa1);
-    DeterminizeFst<StdArc> dfa2(ifsa);
-    return Equivalent(dfa1, dfa2);
-  }
-
-  // Returns complement Fsa
-  void Complement(const Fst<Arc> &ifsa, MutableFst<Arc> *ofsa) {
-    RmEpsilonFst<Arc> rfsa(ifsa);
-    DeterminizeFst<Arc> dfa(rfsa);
-    DifferenceFst<Arc> cfsa(univ_fsa_, dfa);
-    *ofsa = cfsa;
-  }
-
-  // FSA with no states
-  VectorFst<Arc> zero_fsa_;
-
-  // FSA with one state that accepts epsilon.
-  VectorFst<Arc> one_fsa_;
-
-  // FSA with one state that accepts all strings.
-  VectorFst<Arc> univ_fsa_;
-};
-
-// This class tests a variety of identities and properties that must
-// hold for various FST algorithms. It randomly generates FSTs, using
-// function object 'weight_generator' to select weights. 'WeightTester'
-// and 'UnweightedTester' are then called.
-template <class Arc, class WeightGenerator>
-class AlgoTester {
- public:
-  typedef typename Arc::Label Label;
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Weight Weight;
-
-  AlgoTester(WeightGenerator generator, int seed)
-      : weight_generator_(generator) {
-    one_fst_.AddState();
-    one_fst_.SetStart(0);
-    one_fst_.SetFinal(0, Weight::One());
-
-    univ_fst_.AddState();
-    univ_fst_.SetStart(0);
-    univ_fst_.SetFinal(0, Weight::One());
-    for (int i = 0; i < kNumRandomLabels; ++i)
-      univ_fst_.AddArc(0, Arc(i, i, Weight::One(), 0));
-
-    weighted_tester_ = new WeightedTester<Arc, WeightGenerator>(
-        seed, zero_fst_, one_fst_, univ_fst_, &weight_generator_);
-
-    unweighted_tester_ =
-        new UnweightedTester<Arc>(zero_fst_, one_fst_, univ_fst_);
-  }
-
-  ~AlgoTester() {
-    delete weighted_tester_;
-    delete unweighted_tester_;
-  }
-
-  void MakeRandFst(MutableFst<Arc> *fst) {
-    RandFst<Arc, WeightGenerator>(kNumRandomStates, kNumRandomArcs,
-                                  kNumRandomLabels, kAcyclicProb,
-                                  &weight_generator_, fst);
-  }
-
-  void Test() {
-    VLOG(1) << "weight type = " << Weight::Type();
-
-    for (int i = 0; i < FLAGS_repeat; ++i) {
-      // Random transducers
-      VectorFst<Arc> T1;
-      VectorFst<Arc> T2;
-      VectorFst<Arc> T3;
-      MakeRandFst(&T1);
-      MakeRandFst(&T2);
-      MakeRandFst(&T3);
-      weighted_tester_->Test(T1, T2, T3);
-
-      VectorFst<Arc> A1(T1);
-      VectorFst<Arc> A2(T2);
-      VectorFst<Arc> A3(T3);
-      Project(&A1, PROJECT_OUTPUT);
-      Project(&A2, PROJECT_INPUT);
-      Project(&A3, PROJECT_INPUT);
-      ArcMap(&A1, rm_weight_mapper_);
-      ArcMap(&A2, rm_weight_mapper_);
-      ArcMap(&A3, rm_weight_mapper_);
-      unweighted_tester_->Test(A1, A2, A3);
-    }
-  }
-
- private:
-  // Generates weights used in testing.
-  WeightGenerator weight_generator_;
-
-  // FST with no states
-  VectorFst<Arc> zero_fst_;
-
-  // FST with one state that accepts epsilon.
-  VectorFst<Arc> one_fst_;
-
-  // FST with one state that accepts all strings.
-  VectorFst<Arc> univ_fst_;
-
-  // Tests weighted FSTs
-  WeightedTester<Arc, WeightGenerator> *weighted_tester_;
-
-  // Tests unweighted FSTs
-  UnweightedTester<Arc> *unweighted_tester_;
-
-  // Mapper to remove weights from an Fst
-  RmWeightMapper<Arc> rm_weight_mapper_;
-
-  // Maximum number of states in random test Fst.
-  static const int kNumRandomStates;
-
-  // Maximum number of arcs in random test Fst.
-  static const int kNumRandomArcs;
-
-  // Number of alternative random labels.
-  static const int kNumRandomLabels;
-
-  // Probability to force an acyclic Fst
-  static const float kAcyclicProb;
-
-  // Maximum random path length.
-  static const int kRandomPathLength;
-
-  // Number of random paths to explore.
-  static const int kNumRandomPaths;
-
-  AlgoTester(const AlgoTester &) = delete;
-  AlgoTester &operator=(const AlgoTester &) = delete;
-};
-
-template <class A, class G>
-const int AlgoTester<A, G>::kNumRandomStates = 10;
-
-template <class A, class G>
-const int AlgoTester<A, G>::kNumRandomArcs = 25;
-
-template <class A, class G>
-const int AlgoTester<A, G>::kNumRandomLabels = 5;
-
-template <class A, class G>
-const float AlgoTester<A, G>::kAcyclicProb = .25;
-
-template <class A, class G>
-const int AlgoTester<A, G>::kRandomPathLength = 25;
-
-template <class A, class G>
-const int AlgoTester<A, G>::kNumRandomPaths = 100;
-
-}  // namespace fst
-
-#endif  // FST_TEST_ALGO_TEST_H_
diff --git a/projects/llm_framework/include/fst/test/fst_test.h b/projects/llm_framework/include/fst/test/fst_test.h
deleted file mode 100644
index 7d536d90..00000000
--- a/projects/llm_framework/include/fst/test/fst_test.h
+++ /dev/null
@@ -1,318 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Regression test for FST classes.
-
-#ifndef FST_TEST_FST_TEST_H_
-#define FST_TEST_FST_TEST_H_
-
-#include <fst/equal.h>
-#include <fstream>
-#include <fst/matcher.h>
-#include <fst/vector-fst.h>
-#include <fst/verify.h>
-
-DECLARE_string(tmpdir);
-
-namespace fst {
-
-// This tests an Fst F that is assumed to have a copy method from an
-// arbitrary Fst. Some test functions make further assumptions mostly
-// obvious from their name. These tests are written as member temple
-// functions that take a test fst as its argument so that different
-// Fsts in the interface hierarchy can be tested separately and so
-// that we can instantiate only those tests that make sense for a
-// particular Fst.
-template <class F>
-class FstTester {
- public:
-  typedef typename F::Arc Arc;
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Weight Weight;
-  typedef typename Arc::Label Label;
-
-  FstTester() {
-    VectorFst<Arc> vfst;
-    InitFst(&vfst, 128);
-    testfst_ = new F(vfst);
-  }
-
-  explicit FstTester(F *testfst) : testfst_(testfst) {}
-
-  ~FstTester() { delete testfst_; }
-
-  // This verifies the contents described in InitFst() using
-  // methods defined in a generic Fst.
-  template <class G>
-  void TestBase(const G &fst) const {
-    CHECK(Verify(fst));
-    CHECK_EQ(fst.Start(), 0);
-    StateId ns = 0;
-    StateIterator<G> siter(fst);
-    Matcher<G> matcher(fst, MATCH_INPUT);
-    MatchType match_type = matcher.Type(true);
-    for (; !siter.Done(); siter.Next()) {
-    }
-    for (siter.Reset(); !siter.Done(); siter.Next()) {
-      StateId s = siter.Value();
-      matcher.SetState(s);
-      CHECK_EQ(fst.Final(s), NthWeight(s));
-      size_t na = 0;
-      ArcIterator<G> aiter(fst, s);
-      for (; !aiter.Done(); aiter.Next()) {
-      }
-      for (aiter.Reset(); !aiter.Done(); aiter.Next()) {
-        ++na;
-        const Arc &arc = aiter.Value();
-        CHECK_EQ(arc.ilabel, na);
-        CHECK_EQ(arc.olabel, 0);
-        CHECK_EQ(arc.weight, NthWeight(na));
-        CHECK_EQ(arc.nextstate, s);
-        if (match_type == MATCH_INPUT) {
-          CHECK(matcher.Find(arc.ilabel));
-          CHECK_EQ(matcher.Value().ilabel, arc.ilabel);
-        }
-      }
-      CHECK_EQ(na, s);
-      CHECK_EQ(na, aiter.Position());
-      CHECK_EQ(fst.NumArcs(s), s);
-      CHECK_EQ(fst.NumInputEpsilons(s), 0);
-      CHECK_EQ(fst.NumOutputEpsilons(s), s);
-      CHECK(!matcher.Find(s + 1));     // out-of-range
-      CHECK(!matcher.Find(kNoLabel));  // no explicit epsilons
-      CHECK(matcher.Find(0));
-      CHECK_EQ(matcher.Value().ilabel, kNoLabel);  // implicit epsilon loop
-      ++ns;
-    }
-    CHECK(fst.Properties(kNotAcceptor, true));
-    CHECK(fst.Properties(kOEpsilons, true));
-  }
-
-  void TestBase() const { TestBase(*testfst_); }
-
-  // This verifies methods specfic to an ExpandedFst.
-  template <class G>
-  void TestExpanded(const G &fst) const {
-    StateId ns = 0;
-    for (StateIterator<G> siter(fst); !siter.Done(); siter.Next()) {
-      ++ns;
-    }
-    CHECK_EQ(fst.NumStates(), ns);
-    CHECK(fst.Properties(kExpanded, false));
-  }
-
-  void TestExpanded() const { TestExpanded(*testfst_); }
-
-  // This verifies methods specific to a MutableFst.
-  template <class G>
-  void TestMutable(G *fst) const {
-    for (StateIterator<G> siter(*fst); !siter.Done(); siter.Next()) {
-      StateId s = siter.Value();
-      size_t na = 0;
-      size_t ni = fst->NumInputEpsilons(s);
-      MutableArcIterator<G> aiter(fst, s);
-      for (; !aiter.Done(); aiter.Next()) {
-      }
-      for (aiter.Reset(); !aiter.Done(); aiter.Next()) {
-        ++na;
-        Arc arc = aiter.Value();
-        arc.ilabel = 0;
-        aiter.SetValue(arc);
-        arc = aiter.Value();
-        CHECK_EQ(arc.ilabel, 0);
-        CHECK_EQ(fst->NumInputEpsilons(s), ni + 1);
-        arc.ilabel = na;
-        aiter.SetValue(arc);
-        CHECK_EQ(fst->NumInputEpsilons(s), ni);
-      }
-    }
-
-    G *cfst1 = fst->Copy();
-    cfst1->DeleteStates();
-    CHECK_EQ(cfst1->NumStates(), 0);
-    delete cfst1;
-
-    G *cfst2 = fst->Copy();
-    for (StateIterator<G> siter(*cfst2); !siter.Done(); siter.Next()) {
-      StateId s = siter.Value();
-      cfst2->DeleteArcs(s);
-      CHECK_EQ(cfst2->NumArcs(s), 0);
-      CHECK_EQ(cfst2->NumInputEpsilons(s), 0);
-      CHECK_EQ(cfst2->NumOutputEpsilons(s), 0);
-    }
-    delete cfst2;
-  }
-
-  void TestMutable() { TestMutable(testfst_); }
-
-  // This verifies the copy methods.
-  template <class G>
-  void TestAssign(G *fst) const {
-    // Assignment from G
-    G afst1;
-    afst1 = *fst;
-    CHECK(Equal(*fst, afst1));
-
-    // Assignment from Fst
-    G afst2;
-    afst2 = *static_cast<const Fst<Arc> *>(fst);
-    CHECK(Equal(*fst, afst2));
-
-    // Assignment from self
-    afst2.operator=(afst2);
-    CHECK(Equal(*fst, afst2));
-  }
-
-  void TestAssign() { TestAssign(testfst_); }
-
-  // This verifies the copy methods.
-  template <class G>
-  void TestCopy(const G &fst) const {
-    // Copy from G
-    G c1fst(fst);
-    TestBase(c1fst);
-
-    // Copy from Fst
-    const G c2fst(static_cast<const Fst<Arc> &>(fst));
-    TestBase(c2fst);
-
-    // Copy from self
-    const G *c3fst = fst.Copy();
-    TestBase(*c3fst);
-    delete c3fst;
-  }
-
-  void TestCopy() const { TestCopy(*testfst_); }
-
-  // This verifies the read/write methods.
-  template <class G>
-  void TestIO(const G &fst) const {
-    const string filename = FLAGS_tmpdir + "/test.fst";
-    const string aligned = FLAGS_tmpdir + "/aligned.fst";
-    {
-      // write/read
-      CHECK(fst.Write(filename));
-      G *ffst = G::Read(filename);
-      CHECK(ffst);
-      TestBase(*ffst);
-      delete ffst;
-    }
-
-    {
-      // generic read/cast/test
-      Fst<Arc> *gfst = Fst<Arc>::Read(filename);
-      CHECK(gfst);
-      G *dfst = static_cast<G *>(gfst);
-      TestBase(*dfst);
-
-      // generic write/read/test
-      CHECK(gfst->Write(filename));
-      Fst<Arc> *hfst = Fst<Arc>::Read(filename);
-      CHECK(hfst);
-      TestBase(*hfst);
-      delete gfst;
-      delete hfst;
-    }
-
-    {
-      // check mmaping by first writing the file with the aligned attribute set
-      {
-        std::ofstream ostr(aligned);
-        FstWriteOptions opts;
-        opts.source = aligned;
-        opts.align = true;
-        CHECK(fst.Write(ostr, opts));
-      }
-      std::ifstream istr(aligned);
-      FstReadOptions opts;
-      opts.mode = FstReadOptions::ReadMode("map");
-      opts.source = aligned;
-      G *gfst = G::Read(istr, opts);
-      CHECK(gfst);
-      TestBase(*gfst);
-      delete gfst;
-    }
-
-    // check mmaping of unaligned files to make sure it does not fail.
-    {
-      {
-        std::ofstream ostr(aligned);
-        FstWriteOptions opts;
-        opts.source = aligned;
-        opts.align = false;
-        CHECK(fst.Write(ostr, opts));
-      }
-      std::ifstream istr(aligned);
-      FstReadOptions opts;
-      opts.mode = FstReadOptions::ReadMode("map");
-      opts.source = aligned;
-      G *gfst = G::Read(istr, opts);
-      CHECK(gfst);
-      TestBase(*gfst);
-      delete gfst;
-    }
-
-    // expanded write/read/test
-    if (fst.Properties(kExpanded, false)) {
-      ExpandedFst<Arc> *efst = ExpandedFst<Arc>::Read(filename);
-      CHECK(efst);
-      TestBase(*efst);
-      TestExpanded(*efst);
-      delete efst;
-    }
-
-    // mutable write/read/test
-    if (fst.Properties(kMutable, false)) {
-      MutableFst<Arc> *mfst = MutableFst<Arc>::Read(filename);
-      CHECK(mfst);
-      TestBase(*mfst);
-      TestExpanded(*mfst);
-      TestMutable(mfst);
-      delete mfst;
-    }
-  }
-
-  void TestIO() const { TestIO(*testfst_); }
-
- private:
-  // This constructs test FSTs. Given a mutable FST, will leave
-  // the FST as follows:
-  // (I) NumStates() = nstates
-  // (II) Start() = 0
-  // (III) Final(s) =  NthWeight(s)
-  // (IV) For state s:
-  //     (a) NumArcs(s) == s
-  //     (b) For ith arc of s:
-  //         (1) ilabel = i
-  //         (2) olabel = 0
-  //         (3) weight = NthWeight(i)
-  //         (4) nextstate = s
-  void InitFst(MutableFst<Arc> *fst, size_t nstates) const {
-    fst->DeleteStates();
-    CHECK_GT(nstates, 0);
-
-    for (StateId s = 0; s < nstates; ++s) {
-      fst->AddState();
-      fst->SetFinal(s, NthWeight(s));
-      for (size_t i = 1; i <= s; ++i) {
-        Arc arc(i, 0, NthWeight(i), s);
-        fst->AddArc(s, arc);
-      }
-    }
-
-    fst->SetStart(0);
-  }
-
-  // Generates One() + ... + One() (n times)
-  Weight NthWeight(int n) const {
-    Weight w = Weight::Zero();
-    for (int i = 0; i < n; ++i) w = Plus(w, Weight::One());
-    return w;
-  }
-
-  F *testfst_;  // what we're testing
-};
-
-}  // namespace fst
-
-#endif  // FST_TEST_FST_TEST_H_
diff --git a/projects/llm_framework/include/fst/test/rand-fst.h b/projects/llm_framework/include/fst/test/rand-fst.h
deleted file mode 100644
index f2f34c67..00000000
--- a/projects/llm_framework/include/fst/test/rand-fst.h
+++ /dev/null
@@ -1,90 +0,0 @@
-#ifndef FST_TEST_RAND_FST_H_
-#define FST_TEST_RAND_FST_H_
-
-#include <fst/log.h>
-#include <fst/mutable-fst.h>
-#include <fst/verify.h>
-
-namespace fst {
-
-// Generates a random FST.
-template <class Arc, class WeightGenerator>
-void RandFst(const int num_random_states, const int num_random_arcs,
-             const int num_random_labels, const float acyclic_prob,
-             WeightGenerator *weight_generator, MutableFst<Arc> *fst) {
-  typedef typename Arc::Label Label;
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Weight Weight;
-
-  // Determines direction of the arcs wrt state numbering. This way we
-  // can force acyclicity when desired.
-  enum ArcDirection {
-    ANY_DIRECTION = 0,
-    FORWARD_DIRECTION = 1,
-    REVERSE_DIRECTION = 2,
-    NUM_DIRECTIONS = 3
-  };
-
-  ArcDirection arc_direction = ANY_DIRECTION;
-  if (rand() / (RAND_MAX + 1.0) < acyclic_prob)
-    arc_direction = rand() % 2 ? FORWARD_DIRECTION : REVERSE_DIRECTION;
-
-  fst->DeleteStates();
-  StateId ns = rand() % num_random_states;
-
-  if (ns == 0) return;
-  for (StateId s = 0; s < ns; ++s) fst->AddState();
-
-  StateId start = rand() % ns;
-  fst->SetStart(start);
-
-  size_t na = rand() % num_random_arcs;
-  for (size_t n = 0; n < na; ++n) {
-    StateId s = rand() % ns;
-    Arc arc;
-    arc.ilabel = rand() % num_random_labels;
-    arc.olabel = rand() % num_random_labels;
-    arc.weight = (*weight_generator)();
-    arc.nextstate = rand() % ns;
-
-    if ((arc_direction == FORWARD_DIRECTION ||
-         arc_direction == REVERSE_DIRECTION) &&
-        s == arc.nextstate) {
-      continue;  // skips self-loops
-    }
-
-    if ((arc_direction == FORWARD_DIRECTION && s > arc.nextstate) ||
-        (arc_direction == REVERSE_DIRECTION && s < arc.nextstate)) {
-      StateId t = s;  // reverses arcs
-      s = arc.nextstate;
-      arc.nextstate = t;
-    }
-
-    fst->AddArc(s, arc);
-  }
-
-  StateId nf = rand() % (ns + 1);
-  for (StateId n = 0; n < nf; ++n) {
-    StateId s = rand() % ns;
-    Weight final = (*weight_generator)();
-    fst->SetFinal(s, final);
-  }
-  VLOG(1) << "Check FST for sanity (including property bits).";
-  CHECK(Verify(*fst));
-
-  // Get/compute all properties.
-  uint64 props = fst->Properties(kFstProperties, true);
-
-  // Select random set of properties to be unknown.
-  uint64 mask = 0;
-  for (int n = 0; n < 8; ++n) {
-    mask |= rand() & 0xff;
-    mask <<= 8;
-  }
-  mask &= ~kTrinaryProperties;
-  fst->SetProperties(props & ~mask, mask);
-}
-
-}  // namespace fst
-
-#endif  // FST_TEST_RAND_FST_H_
diff --git a/projects/llm_framework/include/fst/test/weight-tester.h b/projects/llm_framework/include/fst/test/weight-tester.h
deleted file mode 100644
index b7de665e..00000000
--- a/projects/llm_framework/include/fst/test/weight-tester.h
+++ /dev/null
@@ -1,207 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Utility class for regression testing of FST weights.
-
-#ifndef FST_TEST_WEIGHT_TESTER_H_
-#define FST_TEST_WEIGHT_TESTER_H_
-
-#include <iostream>
-#include <sstream>
-
-#include <utility>
-
-#include <fst/log.h>
-#include <fst/weight.h>
-
-namespace fst {
-
-// This class tests a variety of identities and properties that must
-// hold for the Weight class to be well-defined. It calls function object
-// WEIGHT_GENERATOR to select weights that are used in the tests.
-template <class Weight, class WeightGenerator>
-class WeightTester {
- public:
-  WeightTester(WeightGenerator generator)
-      : weight_generator_(std::move(generator)) {}
-
-  void Test(int iterations, bool test_division = true) {
-    for (int i = 0; i < iterations; ++i) {
-      // Selects the test weights.
-      const Weight w1(weight_generator_());
-      const Weight w2(weight_generator_());
-      const Weight w3(weight_generator_());
-
-      VLOG(1) << "weight type = " << Weight::Type();
-      VLOG(1) << "w1 = " << w1;
-      VLOG(1) << "w2 = " << w2;
-      VLOG(1) << "w3 = " << w3;
-
-      TestSemiring(w1, w2, w3);
-      if (test_division) TestDivision(w1, w2);
-      TestReverse(w1, w2);
-      TestEquality(w1, w2, w3);
-      TestIO(w1);
-      TestCopy(w1);
-    }
-  }
-
- private:
-  // Note in the tests below we use ApproxEqual rather than == and add
-  // kDelta to inequalities where the weights might be inexact.
-
-  // Tests (Plus, Times, Zero, One) defines a commutative semiring.
-  void TestSemiring(Weight w1, Weight w2, Weight w3) {
-    // Checks that the operations are closed.
-    CHECK(Plus(w1, w2).Member());
-    CHECK(Times(w1, w2).Member());
-
-    // Checks that the operations are associative.
-    CHECK(ApproxEqual(Plus(w1, Plus(w2, w3)), Plus(Plus(w1, w2), w3)));
-    CHECK(ApproxEqual(Times(w1, Times(w2, w3)), Times(Times(w1, w2), w3)));
-
-    // Checks the identity elements.
-    CHECK(Plus(w1, Weight::Zero()) == w1);
-    CHECK(Plus(Weight::Zero(), w1) == w1);
-    CHECK(Times(w1, Weight::One()) == w1);
-    CHECK(Times(Weight::One(), w1) == w1);
-
-    // Check the no weight element.
-    CHECK(!Weight::NoWeight().Member());
-    CHECK(!Plus(w1, Weight::NoWeight()).Member());
-    CHECK(!Plus(Weight::NoWeight(), w1).Member());
-    CHECK(!Times(w1, Weight::NoWeight()).Member());
-    CHECK(!Times(Weight::NoWeight(), w1).Member());
-
-    // Checks that the operations commute.
-    CHECK(ApproxEqual(Plus(w1, w2), Plus(w2, w1)));
-
-    if (Weight::Properties() & kCommutative)
-      CHECK(ApproxEqual(Times(w1, w2), Times(w2, w1)));
-
-    // Checks Zero() is the annihilator.
-    CHECK(Times(w1, Weight::Zero()) == Weight::Zero());
-    CHECK(Times(Weight::Zero(), w1) == Weight::Zero());
-
-    // Check Power(w, 0) is Weight::One()
-    CHECK(Power(w1, 0) == Weight::One());
-
-    // Check Power(w, 1) is w
-    CHECK(Power(w1, 1) == w1);
-
-    // Check Power(w, 3) is Times(w, Times(w, w))
-    CHECK(Power(w1, 3) == Times(w1, Times(w1, w1)));
-
-    // Checks distributivity.
-    if (Weight::Properties() & kLeftSemiring) {
-      CHECK(ApproxEqual(Times(w1, Plus(w2, w3)),
-                        Plus(Times(w1, w2), Times(w1, w3))));
-    }
-    if (Weight::Properties() & kRightSemiring)
-      CHECK(ApproxEqual(Times(Plus(w1, w2), w3),
-                        Plus(Times(w1, w3), Times(w2, w3))));
-
-    if (Weight::Properties() & kIdempotent) CHECK(Plus(w1, w1) == w1);
-
-    if (Weight::Properties() & kPath)
-      CHECK(Plus(w1, w2) == w1 || Plus(w1, w2) == w2);
-
-    // Ensure weights form a left or right semiring.
-    CHECK(Weight::Properties() & (kLeftSemiring | kRightSemiring));
-
-    // Check when Times() is commutative that it is marked as a semiring.
-    if (Weight::Properties() & kCommutative)
-      CHECK(Weight::Properties() & kSemiring);
-  }
-
-  // Tests division operation.
-  void TestDivision(Weight w1, Weight w2) {
-    Weight p = Times(w1, w2);
-
-    if (Weight::Properties() & kLeftSemiring) {
-      Weight d = Divide(p, w1, DIVIDE_LEFT);
-      if (d.Member()) CHECK(ApproxEqual(p, Times(w1, d)));
-      CHECK(!Divide(w1, Weight::NoWeight(), DIVIDE_LEFT).Member());
-      CHECK(!Divide(Weight::NoWeight(), w1, DIVIDE_LEFT).Member());
-    }
-
-    if (Weight::Properties() & kRightSemiring) {
-      Weight d = Divide(p, w2, DIVIDE_RIGHT);
-      if (d.Member()) CHECK(ApproxEqual(p, Times(d, w2)));
-      CHECK(!Divide(w1, Weight::NoWeight(), DIVIDE_RIGHT).Member());
-      CHECK(!Divide(Weight::NoWeight(), w1, DIVIDE_RIGHT).Member());
-    }
-
-    if (Weight::Properties() & kCommutative) {
-      Weight d = Divide(p, w1, DIVIDE_RIGHT);
-      if (d.Member()) CHECK(ApproxEqual(p, Times(d, w1)));
-    }
-  }
-
-  // Tests reverse operation.
-  void TestReverse(Weight w1, Weight w2) {
-    typedef typename Weight::ReverseWeight ReverseWeight;
-
-    ReverseWeight rw1 = w1.Reverse();
-    ReverseWeight rw2 = w2.Reverse();
-
-    CHECK(rw1.Reverse() == w1);
-    CHECK(Plus(w1, w2).Reverse() == Plus(rw1, rw2));
-    CHECK(Times(w1, w2).Reverse() == Times(rw2, rw1));
-  }
-
-  // Tests == is an equivalence relation.
-  void TestEquality(Weight w1, Weight w2, Weight w3) {
-    // Checks reflexivity.
-    CHECK(w1 == w1);
-
-    // Checks symmetry.
-    CHECK((w1 == w2) == (w2 == w1));
-
-    // Checks transitivity.
-    if (w1 == w2 && w2 == w3) CHECK(w1 == w3);
-  }
-
-  // Tests binary serialization and textual I/O.
-  void TestIO(Weight w) {
-    // Tests binary I/O
-    {
-      std::ostringstream os;
-      w.Write(os);
-      os.flush();
-      std::istringstream is(os.str());
-      Weight v;
-      v.Read(is);
-      CHECK_EQ(w, v);
-    }
-
-    // Tests textual I/O.
-    {
-      std::ostringstream os;
-      os << w;
-      std::istringstream is(os.str());
-      Weight v(Weight::One());
-      is >> v;
-      CHECK(ApproxEqual(w, v));
-    }
-  }
-
-  // Tests copy constructor and assignment operator
-  void TestCopy(Weight w) {
-    Weight x = w;
-    CHECK(w == x);
-
-    x = Weight(w);
-    CHECK(w == x);
-
-    x.operator=(x);
-    CHECK(w == x);
-  }
-
-  // Generates weights used in testing.
-  WeightGenerator weight_generator_;
-};
-
-}  // namespace fst
-
-#endif  // FST_TEST_WEIGHT_TESTER_H_
diff --git a/projects/llm_framework/include/fst/topsort.h b/projects/llm_framework/include/fst/topsort.h
deleted file mode 100644
index cae3e154..00000000
--- a/projects/llm_framework/include/fst/topsort.h
+++ /dev/null
@@ -1,95 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Topological sort of FSTs.
-
-#ifndef FST_TOPSORT_H_
-#define FST_TOPSORT_H_
-
-#include <memory>
-#include <vector>
-
-
-#include <fst/dfs-visit.h>
-#include <fst/fst.h>
-#include <fst/statesort.h>
-
-
-namespace fst {
-
-// DFS visitor class to return topological ordering.
-template <class Arc>
-class TopOrderVisitor {
- public:
-  using StateId = typename Arc::StateId;
-
-  // If acyclic, order[i] gives the topological position of StateId i;
-  // otherwise it is unchanged. acyclic_ will be true iff the FST has no
-  // cycles. The caller retains ownership of the state order vector.
-  TopOrderVisitor(std::vector<StateId> *order, bool *acyclic)
-      : order_(order), acyclic_(acyclic) {}
-
-  void InitVisit(const Fst<Arc> &fst) {
-    finish_.reset(new std::vector<StateId>());
-    *acyclic_ = true;
-  }
-
-  constexpr bool InitState(StateId, StateId) const { return true; }
-
-  constexpr bool TreeArc(StateId, const Arc &) const { return true; }
-
-  bool BackArc(StateId, const Arc &) { return (*acyclic_ = false); }
-
-  constexpr bool ForwardOrCrossArc(StateId, const Arc &) const { return true; }
-
-  void FinishState(StateId s, StateId, const Arc *) { finish_->push_back(s); }
-
-  void FinishVisit() {
-    if (*acyclic_) {
-      order_->clear();
-      for (StateId s = 0; s < finish_->size(); ++s) {
-        order_->push_back(kNoStateId);
-      }
-      for (StateId s = 0; s < finish_->size(); ++s) {
-        (*order_)[(*finish_)[finish_->size() - s - 1]] = s;
-      }
-    }
-    finish_.reset();
-  }
-
- private:
-  std::vector<StateId> *order_;
-  bool *acyclic_;
-  // States in finish-time order.
-  std::unique_ptr<std::vector<StateId>> finish_;
-};
-
-// Topologically sorts its input if acyclic, modifying it. Otherwise, the input
-// is unchanged. When sorted, all transitions are from lower to higher state
-// IDs.
-//
-// Complexity:
-//
-//   Time:  O(V + E)
-//   Space: O(V + E)
-//
-// where V is the number of states and E is the number of arcs.
-template <class Arc>
-bool TopSort(MutableFst<Arc> *fst) {
-  std::vector<typename Arc::StateId> order;
-  bool acyclic;
-  TopOrderVisitor<Arc> top_order_visitor(&order, &acyclic);
-  DfsVisit(*fst, &top_order_visitor);
-  if (acyclic) {
-    StateSort(fst, order);
-    fst->SetProperties(kAcyclic | kInitialAcyclic | kTopSorted,
-                       kAcyclic | kInitialAcyclic | kTopSorted);
-  } else {
-    fst->SetProperties(kCyclic | kNotTopSorted, kCyclic | kNotTopSorted);
-  }
-  return acyclic;
-}
-
-}  // namespace fst
-
-#endif  // FST_TOPSORT_H_
diff --git a/projects/llm_framework/include/fst/tuple-weight.h b/projects/llm_framework/include/fst/tuple-weight.h
deleted file mode 100644
index c8ecab76..00000000
--- a/projects/llm_framework/include/fst/tuple-weight.h
+++ /dev/null
@@ -1,163 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Tuple weight set operation definitions.
-
-#ifndef FST_TUPLE_WEIGHT_H_
-#define FST_TUPLE_WEIGHT_H_
-
-#include <algorithm>
-#include <array>
-#include <functional>
-#include <string>
-#include <vector>
-
-#include <fst/flags.h>
-#include <fst/log.h>
-
-#include <fst/weight.h>
-
-
-namespace fst {
-
-// n-tuple weight, element of the n-th Cartesian power of W.
-template <class W, size_t n>
-class TupleWeight {
- public:
-  using ReverseWeight = TupleWeight<typename W::ReverseWeight, n>;
-
-  using Weight = W;
-  using Index = size_t;
-
-  template <class Iterator>
-  TupleWeight(Iterator begin, Iterator end) {
-    std::copy(begin, end, values_.begin());
-  }
-
-  explicit TupleWeight(const W &weight = W::Zero()) { values_.fill(weight); }
-
-  // Initialize component `index` to `weight`; initialize all other components
-  // to `default_weight`
-  TupleWeight(Index index, const W &weight, const W &default_weight)
-      : TupleWeight(default_weight) {
-    values_[index] = weight;
-  }
-
-  static const TupleWeight<W, n> &Zero() {
-    static const TupleWeight<W, n> zero(W::Zero());
-    return zero;
-  }
-
-  static const TupleWeight<W, n> &One() {
-    static const TupleWeight<W, n> one(W::One());
-    return one;
-  }
-
-  static const TupleWeight<W, n> &NoWeight() {
-    static const TupleWeight<W, n> no_weight(W::NoWeight());
-    return no_weight;
-  }
-
-  constexpr static size_t Length() { return n; }
-
-  std::istream &Read(std::istream &istrm) {
-    for (size_t i = 0; i < n; ++i) values_[i].Read(istrm);
-    return istrm;
-  }
-
-  std::ostream &Write(std::ostream &ostrm) const {
-    for (size_t i = 0; i < n; ++i) values_[i].Write(ostrm);
-    return ostrm;
-  }
-
-  bool Member() const {
-    return std::all_of(values_.begin(), values_.end(),
-                       std::mem_fn(&W::Member));
-  }
-
-  size_t Hash() const {
-    uint64 hash = 0;
-    for (size_t i = 0; i < n; ++i) hash = 5 * hash + values_[i].Hash();
-    return size_t(hash);
-  }
-
-  TupleWeight<W, n> Quantize(float delta = kDelta) const {
-    TupleWeight<W, n> weight;
-    for (size_t i = 0; i < n; ++i) {
-      weight.values_[i] = values_[i].Quantize(delta);
-    }
-    return weight;
-  }
-
-  ReverseWeight Reverse() const {
-    TupleWeight<W, n> w;
-    for (size_t i = 0; i < n; ++i) w.values_[i] = values_[i].Reverse();
-    return w;
-  }
-
-  const W &Value(size_t i) const { return values_[i]; }
-
-  void SetValue(size_t i, const W &w) { values_[i] = w; }
-
- private:
-  std::array<W, n> values_;
-};
-
-template <class W, size_t n>
-inline bool operator==(const TupleWeight<W, n> &w1,
-                       const TupleWeight<W, n> &w2) {
-  for (size_t i = 0; i < n; ++i) {
-    if (w1.Value(i) != w2.Value(i)) return false;
-  }
-  return true;
-}
-
-template <class W, size_t n>
-inline bool operator!=(const TupleWeight<W, n> &w1,
-                       const TupleWeight<W, n> &w2) {
-  for (size_t i = 0; i < n; ++i) {
-    if (w1.Value(i) != w2.Value(i)) return true;
-  }
-  return false;
-}
-
-template <class W, size_t n>
-inline bool ApproxEqual(const TupleWeight<W, n> &w1,
-                        const TupleWeight<W, n> &w2, float delta = kDelta) {
-  for (size_t i = 0; i < n; ++i) {
-    if (!ApproxEqual(w1.Value(i), w2.Value(i), delta)) return false;
-  }
-  return true;
-}
-
-template <class W, size_t n>
-inline std::ostream &operator<<(std::ostream &strm,
-                                const TupleWeight<W, n> &w) {
-  CompositeWeightWriter writer(strm);
-  writer.WriteBegin();
-  for (size_t i = 0; i < n; ++i) writer.WriteElement(w.Value(i));
-  writer.WriteEnd();
-  return strm;
-}
-
-template <class W, size_t n>
-inline std::istream &operator>>(std::istream &strm, TupleWeight<W, n> &w) {
-  CompositeWeightReader reader(strm);
-  reader.ReadBegin();
-  W v;
-  // Reads first n-1 elements.
-  static_assert(n > 0, "Size must be positive.");
-  for (size_t i = 0; i < n - 1; ++i) {
-    reader.ReadElement(&v);
-    w.SetValue(i, v);
-  }
-  // Reads n-th element.
-  reader.ReadElement(&v, true);
-  w.SetValue(n - 1, v);
-  reader.ReadEnd();
-  return strm;
-}
-
-}  // namespace fst
-
-#endif  // FST_TUPLE_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/types.h b/projects/llm_framework/include/fst/types.h
deleted file mode 100644
index 9c0b7998..00000000
--- a/projects/llm_framework/include/fst/types.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Various type definitions (mostly for Google compatibility).
-
-#include <cstddef>       // For std::ptrdiff_t.
-#include <cstdlib>       // for ssize_t.
-#include <cstdint>       // for ?int*_t.
-
-#ifndef FST_LIB_TYPES_H_
-#define FST_LIB_TYPES_H_
-
-using int8 = int8_t;
-using int16 = int16_t;
-using int32 = int32_t;
-using int64 = int64_t;
-
-using uint8 = uint8_t;
-using uint16 = uint16_t;
-using uint32 = uint32_t;
-using uint64 = uint64_t;
-
-#ifdef _MSC_VER
-// Not really Windows-specific: they should have used ptrdiff_t in the first
-// place. But on Windows there has never been ssize_t.
-using ssize_t = std::ptrdiff_t;
-#endif  // _MSC_VER
-
-#endif  // FST_LIB_TYPES_H_
diff --git a/projects/llm_framework/include/fst/union-find.h b/projects/llm_framework/include/fst/union-find.h
deleted file mode 100644
index b5c7f10b..00000000
--- a/projects/llm_framework/include/fst/union-find.h
+++ /dev/null
@@ -1,84 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Union-find algorithm for dense sets of non-negative integers, implemented
-// using disjoint tree forests with rank heuristics and path compression.
-
-#ifndef FST_UNION_FIND_H_
-#define FST_UNION_FIND_H_
-
-#include <stack>
-#include <vector>
-
-namespace fst {
-
-// Union-Find algorithm for dense sets of non-negative integers.
-template <class T>
-class UnionFind {
- public:
-  // Creates a disjoint set forest for the range [0; max); 'fail' is a value
-  // indicating that an element hasn't been initialized using MakeSet(...).
-  // The upper bound of the range can be reset (increased) using MakeSet(...).
-  UnionFind(T max, T fail) : parent_(max, fail), rank_(max), fail_(fail) {}
-
-  // Finds the representative of the set 'item' belongs to, performing path
-  // compression if necessary.
-  T FindSet(T item) {
-    if (item >= parent_.size() || item == fail_ || parent_[item] == fail_) {
-      return fail_;
-    }
-    auto *p = &parent_[item];
-    for (; *p != item; item = *p, p = &parent_[item]) exec_stack_.push(p);
-    for (; !exec_stack_.empty(); exec_stack_.pop()) *exec_stack_.top() = *p;
-    return *p;
-  }
-
-  // Creates the (destructive) union of the sets x and y belong to.
-  void Union(T x, T y) { Link(FindSet(x), FindSet(y)); }
-
-  // Initialization of an element: creates a singleton set containing 'item'.
-  // The range [0; max) is reset if item >= max.
-  T MakeSet(T item) {
-    if (item >= parent_.size()) {
-      // New value in parent_ should be initialized to fail_.
-      const auto nitem = item > 0 ? 2 * item : 2;
-      parent_.resize(nitem, fail_);
-      rank_.resize(nitem);
-    }
-    parent_[item] = item;
-    return item;
-  }
-
-  // Initialization of all elements starting from 0 to max - 1 to distinct sets.
-  void MakeAllSet(T max) {
-    parent_.resize(max);
-    for (T item = 0; item < max; ++item) parent_[item] = item;
-  }
-
- private:
-  // Links trees rooted in 'x' and 'y'.
-  void Link(T x, T y) {
-    if (x == y) return;
-    if (rank_[x] > rank_[y]) {
-      parent_[y] = x;
-    } else {
-      parent_[x] = y;
-      if (rank_[x] == rank_[y]) {
-        ++rank_[y];
-      }
-    }
-  }
-
-  UnionFind(const UnionFind &) = delete;
-
-  UnionFind &operator=(const UnionFind &) = delete;
-
-  std::vector<T> parent_;       // Parent nodes.
-  std::vector<int> rank_;       // Rank of an element = min. depth in tree.
-  T fail_;                      // Value indicating lookup failure.
-  std::stack<T *> exec_stack_;  // Used for path compression.
-};
-
-}  // namespace fst
-
-#endif  // FST_UNION_FIND_H_
diff --git a/projects/llm_framework/include/fst/union-weight.h b/projects/llm_framework/include/fst/union-weight.h
deleted file mode 100644
index bb2dea96..00000000
--- a/projects/llm_framework/include/fst/union-weight.h
+++ /dev/null
@@ -1,505 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Union weight set and associated semiring operation definitions.
-//
-// TODO(riley): add in normalizer functor
-
-#ifndef FST_UNION_WEIGHT_H_
-#define FST_UNION_WEIGHT_H_
-
-#include <cstdlib>
-
-#include <iostream>
-#include <list>
-#include <sstream>
-#include <string>
-#include <utility>
-
-#include <fst/weight.h>
-
-
-namespace fst {
-
-// Example UnionWeightOptions for UnionWeight template below. The Merge
-// operation is used to collapse elements of the set and the Compare function
-// to efficiently implement the merge. In the simplest case, merge would just
-// apply with equality of set elements so the result is a set (and not a
-// multiset). More generally, this can be used to maintain the multiplicity or
-// other such weight associated with the set elements (cf. Gallic weights).
-
-// template <class W>
-// struct UnionWeightOptions {
-//   // Comparison function C is a total order on W that is monotonic w.r.t. to
-//   // Times: for all a, b,c != Zero(): C(a, b) => C(ca, cb) and is
-//   // anti-monotonic w.r.rt to Divide: C(a, b) => C(c/b, c/a).
-//   //
-//   // For all a, b: only one of C(a, b), C(b, a) or a ~ b must true where
-//   // ~ is an equivalence relation on W. Also we require a ~ b iff
-//   // a.Reverse() ~ b.Reverse().
-//   using Compare = NaturalLess<W>;
-//
-//   // How to combine two weights if a ~ b as above. For all a, b: a ~ b =>
-//   // merge(a, b) ~ a, Merge must define a semiring endomorphism from the
-//   // unmerged weight sets to the merged weight sets.
-//   struct Merge {
-//     W operator()(const W &w1, const W &w2) const { return w1; }
-//   };
-//
-//   // For ReverseWeight.
-//   using ReverseOptions = UnionWeightOptions<ReverseWeight>;
-// };
-
-template <class W, class O>
-class UnionWeight;
-
-template <class W, class O>
-class UnionWeightIterator;
-
-template <class W, class O>
-class UnionWeightReverseIterator;
-
-template <class W, class O>
-bool operator==(const UnionWeight<W, O> &, const UnionWeight<W, O> &);
-
-// Semiring that uses Times() and One() from W and union and the empty set
-// for Plus() and Zero(), respectively. Template argument O specifies the union
-// weight options as above.
-template <class W, class O>
-class UnionWeight {
- public:
-  using Weight = W;
-  using Compare = typename O::Compare;
-  using Merge = typename O::Merge;
-
-  using ReverseWeight =
-      UnionWeight<typename W::ReverseWeight, typename O::ReverseOptions>;
-
-  friend class UnionWeightIterator<W, O>;
-  friend class UnionWeightReverseIterator<W, O>;
-  friend bool operator==
-      <>(const UnionWeight<W, O> &, const UnionWeight<W, O> &);
-
-  // Sets represented as first_ weight + rest_ weights. Uses first_ as
-  // NoWeight() to indicate the union weight Zero() ask the empty set. Uses
-  // rest_ containing NoWeight() to indicate the union weight NoWeight().
-  UnionWeight() : first_(W::NoWeight()) {}
-
-  explicit UnionWeight(W weight) : first_(weight) {
-    if (weight == W::NoWeight()) rest_.push_back(weight);
-  }
-
-  static const UnionWeight<W, O> &Zero() {
-    static const UnionWeight<W, O> zero(W::NoWeight());
-    return zero;
-  }
-
-  static const UnionWeight<W, O> &One() {
-    static const UnionWeight<W, O> one(W::One());
-    return one;
-  }
-
-  static const UnionWeight<W, O> &NoWeight() {
-    static const UnionWeight<W, O> no_weight(W::Zero(), W::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string(W::Type() + "_union");
-    return *type;
-  }
-
-  static constexpr uint64 Properties() {
-    return W::Properties() &
-           (kLeftSemiring | kRightSemiring | kCommutative | kIdempotent);
-  }
-
-  bool Member() const;
-
-  std::istream &Read(std::istream &strm);
-
-  std::ostream &Write(std::ostream &strm) const;
-
-  size_t Hash() const;
-
-  UnionWeight<W, O> Quantize(float delta = kDelta) const;
-
-  ReverseWeight Reverse() const;
-
-  // These operations combined with the UnionWeightIterator and
-  // UnionWeightReverseIterator provide the access and mutation of the union
-  // weight internal elements.
-
-  // Common initializer among constructors; clears existing UnionWeight.
-  void Clear() {
-    first_ = W::NoWeight();
-    rest_.clear();
-  }
-
-  size_t Size() const { return first_.Member() ? rest_.size() + 1 : 0; }
-
-  const W &Back() const { return rest_.empty() ? first_ : rest_.back(); }
-
-  // When srt is true, assumes elements added sorted w.r.t Compare and merging
-  // of weights performed as needed. Otherwise, just ensures first_ is the
-  // least element wrt Compare.
-  void PushBack(W weight, bool srt);
-
-  // Sorts the elements of the set. Assumes that first_, if present, is the
-  // least element.
-  void Sort() { rest_.sort(comp_); }
-
- private:
-  W &Back() {
-    if (rest_.empty()) {
-      return first_;
-    } else {
-      return rest_.back();
-    }
-  }
-
-  UnionWeight(W w1, W w2) : first_(std::move(w1)), rest_(1, std::move(w2)) {}
-
-  W first_;            // First weight in set.
-  std::list<W> rest_;  // Remaining weights in set.
-  Compare comp_;
-  Merge merge_;
-};
-
-template <class W, class O>
-void UnionWeight<W, O>::PushBack(W weight, bool srt) {
-  if (!weight.Member()) {
-    rest_.push_back(std::move(weight));
-  } else if (!first_.Member()) {
-    first_ = std::move(weight);
-  } else if (srt) {
-    auto &back = Back();
-    if (comp_(back, weight)) {
-      rest_.push_back(std::move(weight));
-    } else {
-      back = merge_(back, std::move(weight));
-    }
-  } else {
-    if (comp_(first_, weight)) {
-      rest_.push_back(std::move(weight));
-    } else {
-      rest_.push_back(first_);
-      first_ = std::move(weight);
-    }
-  }
-}
-
-// Traverses union weight in the forward direction.
-template <class W, class O>
-class UnionWeightIterator {
- public:
-  explicit UnionWeightIterator(const UnionWeight<W, O> &weight)
-      : first_(weight.first_),
-        rest_(weight.rest_),
-        init_(true),
-        it_(rest_.begin()) {}
-
-  bool Done() const { return init_ ? !first_.Member() : it_ == rest_.end(); }
-
-  const W &Value() const { return init_ ? first_ : *it_; }
-
-  void Next() {
-    if (init_) {
-      init_ = false;
-    } else {
-      ++it_;
-    }
-  }
-
-  void Reset() {
-    init_ = true;
-    it_ = rest_.begin();
-  }
-
- private:
-  const W &first_;
-  const std::list<W> &rest_;
-  bool init_;  // in the initialized state?
-  typename std::list<W>::const_iterator it_;
-};
-
-// Traverses union weight in backward direction.
-template <typename L, class O>
-class UnionWeightReverseIterator {
- public:
-  explicit UnionWeightReverseIterator(const UnionWeight<L, O> &weight)
-      : first_(weight.first_),
-        rest_(weight.rest_),
-        fin_(!first_.Member()),
-        it_(rest_.rbegin()) {}
-
-  bool Done() const { return fin_; }
-
-  const L &Value() const { return it_ == rest_.rend() ? first_ : *it_; }
-
-  void Next() {
-    if (it_ == rest_.rend()) {
-      fin_ = true;
-    } else {
-      ++it_;
-    }
-  }
-
-  void Reset() {
-    fin_ = !first_.Member();
-    it_ = rest_.rbegin();
-  }
-
- private:
-  const L &first_;
-  const std::list<L> &rest_;
-  bool fin_;  // in the final state?
-  typename std::list<L>::const_reverse_iterator it_;
-};
-
-// UnionWeight member functions follow that require UnionWeightIterator.
-template <class W, class O>
-inline std::istream &UnionWeight<W, O>::Read(std::istream &istrm) {
-  Clear();
-  int32 size;
-  ReadType(istrm, &size);
-  for (int i = 0; i < size; ++i) {
-    W weight;
-    ReadType(istrm, &weight);
-    PushBack(weight, true);
-  }
-  return istrm;
-}
-
-template <class W, class O>
-inline std::ostream &UnionWeight<W, O>::Write(std::ostream &ostrm) const {
-  const int32 size = Size();
-  WriteType(ostrm, size);
-  for (UnionWeightIterator<W, O> it(*this); !it.Done(); it.Next()) {
-    WriteType(ostrm, it.Value());
-  }
-  return ostrm;
-}
-
-template <class W, class O>
-inline bool UnionWeight<W, O>::Member() const {
-  if (Size() <= 1) return true;
-  for (UnionWeightIterator<W, O> it(*this); !it.Done(); it.Next()) {
-    if (!it.Value().Member()) return false;
-  }
-  return true;
-}
-
-template <class W, class O>
-inline UnionWeight<W, O> UnionWeight<W, O>::Quantize(float delta) const {
-  UnionWeight<W, O> weight;
-  for (UnionWeightIterator<W, O> it(*this); !it.Done(); it.Next()) {
-    weight.PushBack(it.Value().Quantize(delta), true);
-  }
-  return weight;
-}
-
-template <class W, class O>
-inline typename UnionWeight<W, O>::ReverseWeight UnionWeight<W, O>::Reverse()
-    const {
-  ReverseWeight weight;
-  for (UnionWeightIterator<W, O> it(*this); !it.Done(); it.Next()) {
-    weight.PushBack(it.Value().Reverse(), false);
-  }
-  weight.Sort();
-  return weight;
-}
-
-template <class W, class O>
-inline size_t UnionWeight<W, O>::Hash() const {
-  size_t h = 0;
-  static constexpr int lshift = 5;
-  static constexpr int rshift = CHAR_BIT * sizeof(size_t) - lshift;
-  for (UnionWeightIterator<W, O> it(*this); !it.Done(); it.Next()) {
-    h = h << lshift ^ h >> rshift ^ it.Value().Hash();
-  }
-  return h;
-}
-
-// Requires union weight has been canonicalized.
-template <class W, class O>
-inline bool operator==(const UnionWeight<W, O> &w1,
-                       const UnionWeight<W, O> &w2) {
-  if (w1.Size() != w2.Size()) return false;
-  UnionWeightIterator<W, O> it1(w1);
-  UnionWeightIterator<W, O> it2(w2);
-  for (; !it1.Done(); it1.Next(), it2.Next()) {
-    if (it1.Value() != it2.Value()) return false;
-  }
-  return true;
-}
-
-// Requires union weight has been canonicalized.
-template <class W, class O>
-inline bool operator!=(const UnionWeight<W, O> &w1,
-                       const UnionWeight<W, O> &w2) {
-  return !(w1 == w2);
-}
-
-// Requires union weight has been canonicalized.
-template <class W, class O>
-inline bool ApproxEqual(const UnionWeight<W, O> &w1,
-                        const UnionWeight<W, O> &w2, float delta = kDelta) {
-  if (w1.Size() != w2.Size()) return false;
-  UnionWeightIterator<W, O> it1(w1);
-  UnionWeightIterator<W, O> it2(w2);
-  for (; !it1.Done(); it1.Next(), it2.Next()) {
-    if (!ApproxEqual(it1.Value(), it2.Value(), delta)) return false;
-  }
-  return true;
-}
-
-template <class W, class O>
-inline std::ostream &operator<<(std::ostream &ostrm,
-                                const UnionWeight<W, O> &weight) {
-  UnionWeightIterator<W, O> it(weight);
-  if (it.Done()) {
-    return ostrm << "EmptySet";
-  } else if (!weight.Member()) {
-    return ostrm << "BadSet";
-  } else {
-    CompositeWeightWriter writer(ostrm);
-    writer.WriteBegin();
-    for (; !it.Done(); it.Next()) writer.WriteElement(it.Value());
-    writer.WriteEnd();
-  }
-  return ostrm;
-}
-
-template <class W, class O>
-inline std::istream &operator>>(std::istream &istrm,
-                                UnionWeight<W, O> &weight) {
-  string s;
-  istrm >> s;
-  if (s == "EmptySet") {
-    weight = UnionWeight<W, O>::Zero();
-  } else if (s == "BadSet") {
-    weight = UnionWeight<W, O>::NoWeight();
-  } else {
-    weight = UnionWeight<W, O>::Zero();
-    std::istringstream sstrm(s);
-    CompositeWeightReader reader(sstrm);
-    reader.ReadBegin();
-    bool more = true;
-    while (more) {
-      W v;
-      more = reader.ReadElement(&v);
-      weight.PushBack(v, true);
-    }
-    reader.ReadEnd();
-  }
-  return istrm;
-}
-
-template <class W, class O>
-inline UnionWeight<W, O> Plus(const UnionWeight<W, O> &w1,
-                              const UnionWeight<W, O> &w2) {
-  if (!w1.Member() || !w2.Member()) return UnionWeight<W, O>::NoWeight();
-  if (w1 == UnionWeight<W, O>::Zero()) return w2;
-  if (w2 == UnionWeight<W, O>::Zero()) return w1;
-  UnionWeightIterator<W, O> it1(w1);
-  UnionWeightIterator<W, O> it2(w2);
-  UnionWeight<W, O> sum;
-  typename O::Compare comp;
-  while (!it1.Done() && !it2.Done()) {
-    const auto v1 = it1.Value();
-    const auto v2 = it2.Value();
-    if (comp(v1, v2)) {
-      sum.PushBack(v1, true);
-      it1.Next();
-    } else {
-      sum.PushBack(v2, true);
-      it2.Next();
-    }
-  }
-  for (; !it1.Done(); it1.Next()) sum.PushBack(it1.Value(), true);
-  for (; !it2.Done(); it2.Next()) sum.PushBack(it2.Value(), true);
-  return sum;
-}
-
-template <class W, class O>
-inline UnionWeight<W, O> Times(const UnionWeight<W, O> &w1,
-                               const UnionWeight<W, O> &w2) {
-  if (!w1.Member() || !w2.Member()) return UnionWeight<W, O>::NoWeight();
-  if (w1 == UnionWeight<W, O>::Zero() || w2 == UnionWeight<W, O>::Zero()) {
-    return UnionWeight<W, O>::Zero();
-  }
-  UnionWeightIterator<W, O> it1(w1);
-  UnionWeightIterator<W, O> it2(w2);
-  UnionWeight<W, O> prod1;
-  for (; !it1.Done(); it1.Next()) {
-    UnionWeight<W, O> prod2;
-    for (; !it2.Done(); it2.Next()) {
-      prod2.PushBack(Times(it1.Value(), it2.Value()), true);
-    }
-    prod1 = Plus(prod1, prod2);
-    it2.Reset();
-  }
-  return prod1;
-}
-
-template <class W, class O>
-inline UnionWeight<W, O> Divide(const UnionWeight<W, O> &w1,
-                                const UnionWeight<W, O> &w2, DivideType typ) {
-  if (!w1.Member() || !w2.Member()) return UnionWeight<W, O>::NoWeight();
-  if (w1 == UnionWeight<W, O>::Zero() || w2 == UnionWeight<W, O>::Zero()) {
-    return UnionWeight<W, O>::Zero();
-  }
-  UnionWeightIterator<W, O> it1(w1);
-  UnionWeightReverseIterator<W, O> it2(w2);
-  UnionWeight<W, O> quot;
-  if (w1.Size() == 1) {
-    for (; !it2.Done(); it2.Next()) {
-      quot.PushBack(Divide(it1.Value(), it2.Value(), typ), true);
-    }
-  } else if (w2.Size() == 1) {
-    for (; !it1.Done(); it1.Next()) {
-      quot.PushBack(Divide(it1.Value(), it2.Value(), typ), true);
-    }
-  } else {
-    quot = UnionWeight<W, O>::NoWeight();
-  }
-  return quot;
-}
-
-// This function object generates weights over the union of weights for the
-// underlying generators for the template weight types. This is intended
-// primarily for testing.
-template <class W, class O>
-class WeightGenerate<UnionWeight<W, O>> {
- public:
-  using Weight = UnionWeight<W, O>;
-  using Generate = WeightGenerate<W>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t num_random_weights = kNumRandomWeights)
-      : generate_(false), allow_zero_(allow_zero),
-        num_random_weights_(num_random_weights) {}
-
-  Weight operator()() const {
-    const int n = rand() % (num_random_weights_ + 1);  // NOLINT
-    if (allow_zero_ && n == num_random_weights_) {
-      return Weight::Zero();
-    } else if (n % 2 == 0) {
-      return Weight(generate_());
-    } else {
-      return Plus(Weight(generate_()), Weight(generate_()));
-    }
-  }
-
- private:
-  Generate generate_;
-  // Permits Zero() and zero divisors.
-  bool allow_zero_;
-  // The number of alternative random weights.
-  const size_t num_random_weights_;
-};
-
-}  // namespace fst
-
-#endif  // FST_UNION_WEIGHT_H_
diff --git a/projects/llm_framework/include/fst/union.h b/projects/llm_framework/include/fst/union.h
deleted file mode 100644
index 257099b8..00000000
--- a/projects/llm_framework/include/fst/union.h
+++ /dev/null
@@ -1,157 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to compute the union of two FSTs.
-
-#ifndef FST_UNION_H_
-#define FST_UNION_H_
-
-#include <algorithm>
-#include <utility>
-#include <vector>
-
-#include <fst/mutable-fst.h>
-#include <fst/rational.h>
-
-
-namespace fst {
-
-// Computes the union (sum) of two FSTs. This version writes the union to an
-// output MutableFst. If A transduces string x to y with weight a and B
-// transduces string w to v with weight b, then their union transduces x to y
-// with weight a and w to v with weight b.
-//
-// Complexity:
-//
-//   Time: (V_2 + E_2)
-//   Space: O(V_2 + E_2)
-//
-// where Vi is the number of states, and Ei is the number of arcs, in the ith
-// FST.
-template <class Arc>
-void Union(MutableFst<Arc> *fst1, const Fst<Arc> &fst2) {
-  using Weight = typename Arc::Weight;
-  // Checks for symbol table compatibility.
-  if (!CompatSymbols(fst1->InputSymbols(), fst2.InputSymbols()) ||
-      !CompatSymbols(fst1->OutputSymbols(), fst2.OutputSymbols())) {
-    FSTERROR() << "Union: Input/output symbol tables of 1st argument "
-               << "do not match input/output symbol tables of 2nd argument";
-    fst1->SetProperties(kError, kError);
-    return;
-  }
-  const auto numstates1 = fst1->NumStates();
-  const bool initial_acyclic1 = fst1->Properties(kInitialAcyclic, true);
-  const auto props1 = fst1->Properties(kFstProperties, false);
-  const auto props2 = fst2.Properties(kFstProperties, false);
-  const auto start2 = fst2.Start();
-  if (start2 == kNoStateId) {
-    if (props2 & kError) fst1->SetProperties(kError, kError);
-    return;
-  }
-  if (fst2.Properties(kExpanded, false)) {
-    fst1->ReserveStates(numstates1 + CountStates(fst2) +
-                        (initial_acyclic1 ? 0 : 1));
-  }
-  for (StateIterator<Fst<Arc>> siter(fst2); !siter.Done(); siter.Next()) {
-    const auto s1 = fst1->AddState();
-    const auto s2 = siter.Value();
-    fst1->SetFinal(s1, fst2.Final(s2));
-    fst1->ReserveArcs(s1, fst2.NumArcs(s2));
-    for (ArcIterator<Fst<Arc>> aiter(fst2, s2); !aiter.Done(); aiter.Next()) {
-      auto arc = aiter.Value();  // Copy intended.
-      arc.nextstate += numstates1;
-      fst1->AddArc(s1, std::move(arc));
-    }
-  }
-  const auto start1 = fst1->Start();
-  if (start1 == kNoStateId) {
-    fst1->SetStart(start2);
-    fst1->SetProperties(props2, kCopyProperties);
-    return;
-  }
-  if (initial_acyclic1) {
-    fst1->AddArc(start1, Arc(0, 0, Weight::One(), start2 + numstates1));
-  } else {
-    const auto nstart1 = fst1->AddState();
-    fst1->SetStart(nstart1);
-    fst1->AddArc(nstart1, Arc(0, 0, Weight::One(), start1));
-    fst1->AddArc(nstart1, Arc(0, 0, Weight::One(), start2 + numstates1));
-  }
-  fst1->SetProperties(UnionProperties(props1, props2), kFstProperties);
-}
-
-// Computes the union of two FSTs, modifying the RationalFst argument.
-template <class Arc>
-void Union(RationalFst<Arc> *fst1, const Fst<Arc> &fst2) {
-  fst1->GetMutableImpl()->AddUnion(fst2);
-}
-
-using UnionFstOptions = RationalFstOptions;
-
-// Computes the union (sum) of two FSTs. This version is a delayed FST. If A
-// transduces string x to y with weight a and B transduces string w to v with
-// weight b, then their union transduces x to y with weight a and w to v with
-// weight b.
-//
-// Complexity:
-//
-//   Time: O(v_1 + e_1 + v_2 + e_2)
-//   Space: O(v_1 + v_2)
-//
-// where vi is the number of states visited, and ei is the number of arcs
-// visited, in the ith FST. Constant time and space to visit an input state or
-// arc is assumed and exclusive of caching.
-template <class A>
-class UnionFst : public RationalFst<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  UnionFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {
-    GetMutableImpl()->InitUnion(fst1, fst2);
-  }
-
-  UnionFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-           const UnionFstOptions &opts)
-      : RationalFst<Arc>(opts) {
-    GetMutableImpl()->InitUnion(fst1, fst2);
-  }
-
-  // See Fst<>::Copy() for doc.
-  UnionFst(const UnionFst<Arc> &fst, bool safe = false)
-      : RationalFst<Arc>(fst, safe) {}
-
-  // Gets a copy of this UnionFst. See Fst<>::Copy() for further doc.
-  UnionFst<Arc> *Copy(bool safe = false) const override {
-    return new UnionFst<Arc>(*this, safe);
-  }
-
- private:
-  using ImplToFst<internal::RationalFstImpl<Arc>>::GetImpl;
-  using ImplToFst<internal::RationalFstImpl<Arc>>::GetMutableImpl;
-};
-
-// Specialization for UnionFst.
-template <class Arc>
-class StateIterator<UnionFst<Arc>> : public StateIterator<RationalFst<Arc>> {
- public:
-  explicit StateIterator(const UnionFst<Arc> &fst)
-      : StateIterator<RationalFst<Arc>>(fst) {}
-};
-
-// Specialization for UnionFst.
-template <class Arc>
-class ArcIterator<UnionFst<Arc>> : public ArcIterator<RationalFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const UnionFst<Arc> &fst, StateId s)
-      : ArcIterator<RationalFst<Arc>>(fst, s) {}
-};
-
-using StdUnionFst = UnionFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_UNION_H_
diff --git a/projects/llm_framework/include/fst/util.h b/projects/llm_framework/include/fst/util.h
deleted file mode 100644
index c7520213..00000000
--- a/projects/llm_framework/include/fst/util.h
+++ /dev/null
@@ -1,400 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// FST utility inline definitions.
-
-#ifndef FST_UTIL_H_
-#define FST_UTIL_H_
-
-#include <iostream>
-#include <iterator>
-#include <list>
-#include <map>
-#include <set>
-#include <sstream>
-#include <string>
-#include <type_traits>
-#include <unordered_map>
-#include <unordered_set>
-#include <utility>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/types.h>
-#include <fst/log.h>
-#include <fstream>
-
-#include <fst/flags.h>
-#include <unordered_map>
-
-
-// Utility for error handling.
-
-DECLARE_bool(fst_error_fatal);
-
-#define FSTERROR() \
-  (FLAGS_fst_error_fatal ? LOG(FATAL) : LOG(ERROR))
-
-namespace fst {
-
-// Utility for type I/O.
-
-// Reads types from an input stream.
-
-// Generic case.
-template <class T,
-    typename std::enable_if<std::is_class<T>::value, T>::type* = nullptr>
-inline std::istream &ReadType(std::istream &strm, T *t) {
-  return t->Read(strm);
-}
-
-// Numeric (boolean, integral, floating-point) case.
-template <class T,
-    typename std::enable_if<std::is_arithmetic<T>::value, T>::type* = nullptr>
-inline std::istream &ReadType(std::istream &strm, T *t) {
-  return strm.read(reinterpret_cast<char *>(t), sizeof(T)); \
-}
-
-// String case.
-inline std::istream &ReadType(std::istream &strm, string *s) {  // NOLINT
-  s->clear();
-  int32 ns = 0;
-  strm.read(reinterpret_cast<char *>(&ns), sizeof(ns));
-  for (int32 i = 0; i < ns; ++i) {
-    char c;
-    strm.read(&c, 1);
-    *s += c;
-  }
-  return strm;
-}
-
-// Pair case.
-template <typename S, typename T>
-inline std::istream &ReadType(std::istream &strm, std::pair<S, T> *p) {
-  ReadType(strm, &p->first);
-  ReadType(strm, &p->second);
-  return strm;
-}
-
-template <typename S, typename T>
-inline std::istream &ReadType(std::istream &strm, std::pair<const S, T> *p) {
-  ReadType(strm, const_cast<S *>(&p->first));
-  ReadType(strm, &p->second);
-  return strm;
-}
-
-namespace internal {
-template <class C, class ReserveFn>
-std::istream &ReadContainerType(std::istream &strm, C *c, ReserveFn reserve) {
-  c->clear();
-  int64 n = 0;
-  ReadType(strm, &n);
-  reserve(c, n);
-  auto insert = std::inserter(*c, c->begin());
-  for (int64 i = 0; i < n; ++i) {
-    typename C::value_type value;
-    ReadType(strm, &value);
-    *insert = value;
-  }
-  return strm;
-}
-}  // namespace internal
-
-template <typename T, typename A>
-std::istream &ReadType(std::istream &strm, std::vector<T,A> *c) {
-  return internal::ReadContainerType(
-      strm, c, [](decltype(c) v, int n) { v->reserve(n); });
-}
-
-template <typename T, typename A>
-std::istream &ReadType(std::istream &strm, std::list<T,A> *c) {
-  return internal::ReadContainerType(strm, c, [](decltype(c) v, int n) {});
-}
-
-template <typename T, typename L, typename A>
-std::istream &ReadType(std::istream &strm, std::set<T,L,A> *c) {
-  return internal::ReadContainerType(strm, c, [](decltype(c) v, int n) {});
-}
-
-template <typename K, typename V, typename L, typename A>
-std::istream &ReadType(std::istream &strm, std::map<K,V,L,A> *c) {
-  return internal::ReadContainerType(strm, c, [](decltype(c) v, int n) {});
-}
-
-template <typename T, typename H, typename E, typename A>
-std::istream &ReadType(std::istream &strm, std::unordered_set<T,H,E,A> *c) {
-  return internal::ReadContainerType(
-      strm, c, [](decltype(c) v, int n) { v->reserve(n); });
-}
-
-template <typename K, typename V, typename H, typename E, typename A>
-std::istream &ReadType(std::istream &strm, std::unordered_map<K,V,H,E,A> *c) {
-  return internal::ReadContainerType(
-      strm, c, [](decltype(c) v, int n) { v->reserve(n); });
-}
-
-// Writes types to an output stream.
-
-// Generic case.
-template <class T,
-    typename std::enable_if<std::is_class<T>::value, T>::type* = nullptr>
-inline std::ostream &WriteType(std::ostream &strm, const T t) {
-  t.Write(strm);
-  return strm;
-}
-
-// Numeric (boolean, integral, floating-point) case.
-template <class T,
-    typename std::enable_if<std::is_arithmetic<T>::value, T>::type* = nullptr>
-inline std::ostream &WriteType(std::ostream &strm, const T t) {
-  return strm.write(reinterpret_cast<const char *>(&t), sizeof(T));
-}
-
-// String case.
-inline std::ostream &WriteType(std::ostream &strm, const string &s) {  // NOLINT
-  int32 ns = s.size();
-  strm.write(reinterpret_cast<const char *>(&ns), sizeof(ns));
-  return strm.write(s.data(), ns);
-}
-
-// Pair case.
-template <typename S, typename T>
-inline std::ostream &WriteType(std::ostream &strm,
-                               const std::pair<S, T> &p) {  // NOLINT
-  WriteType(strm, p.first);
-  WriteType(strm, p.second);
-  return strm;
-}
-
-namespace internal {
-template <class C>
-std::ostream &WriteContainer(std::ostream &strm, const C &c) {
-  const int64 n = c.size();
-  WriteType(strm, n);
-  for (const auto &e : c) {
-    WriteType(strm, e);
-  }
-  return strm;
-}
-}  // namespace internal
-
-template <typename T, typename A>
-std::ostream &WriteType(std::ostream &strm, const std::vector<T,A> &c) {
-  return internal::WriteContainer(strm, c);
-}
-
-template <typename T, typename A>
-std::ostream &WriteType(std::ostream &strm, const std::list<T,A> &c) {
-  return internal::WriteContainer(strm, c);
-}
-
-template <typename T, typename L, typename A>
-std::ostream &WriteType(std::ostream &strm, const std::set<T,L,A> &c) {
-  return internal::WriteContainer(strm, c);
-}
-
-template <typename K, typename V, typename L, typename A>
-std::ostream &WriteType(std::ostream &strm, const std::map<K,V,L,A> &c) {
-  return internal::WriteContainer(strm, c);
-}
-
-template <typename T, typename H, typename E, typename A>
-std::ostream &WriteType(std::ostream &strm, const std::unordered_set<T,H,E,A> &c) {
-  return internal::WriteContainer(strm, c);
-}
-
-template <typename K, typename V, typename H, typename E, typename A>
-std::ostream &WriteType(std::ostream &strm, const std::unordered_map<K,V,H,E,A> &c) {
-  return internal::WriteContainer(strm, c);
-}
-
-// Utilities for converting between int64 or Weight and string.
-
-int64 StrToInt64(const string &s, const string &src, size_t nline,
-                 bool allow_negative, bool *error = nullptr);
-
-template <typename Weight>
-Weight StrToWeight(const string &s, const string &src, size_t nline) {
-  Weight w;
-  std::istringstream strm(s);
-  strm >> w;
-  if (!strm) {
-    FSTERROR() << "StrToWeight: Bad weight = \"" << s << "\", source = " << src
-               << ", line = " << nline;
-    return Weight::NoWeight();
-  }
-  return w;
-}
-
-template <typename Weight>
-void WeightToStr(Weight w, string *s) {
-  std::ostringstream strm;
-  strm.precision(9);
-  strm << w;
-  s->append(strm.str().data(), strm.str().size());
-}
-
-// Utilities for reading/writing integer pairs (typically labels)
-
-// Modifies line using a vector of pointers to a buffer beginning with line.
-void SplitString(char *line, const char *delim, std::vector<char *> *vec,
-                 bool omit_empty_strings);
-
-template <typename I>
-bool ReadIntPairs(const string &filename, std::vector<std::pair<I, I>> *pairs,
-                  bool allow_negative = false) {
-  std::ifstream strm(filename, std::ios_base::in);
-  if (!strm) {
-    LOG(ERROR) << "ReadIntPairs: Can't open file: " << filename;
-    return false;
-  }
-  const int kLineLen = 8096;
-  char line[kLineLen];
-  size_t nline = 0;
-  pairs->clear();
-  while (strm.getline(line, kLineLen)) {
-    ++nline;
-    std::vector<char *> col;
-    SplitString(line, "\n\t ", &col, true);
-    // empty line or comment?
-    if (col.empty() || col[0][0] == '\0' || col[0][0] == '#') continue;
-    if (col.size() != 2) {
-      LOG(ERROR) << "ReadIntPairs: Bad number of columns, "
-                 << "file = " << filename << ", line = " << nline;
-      return false;
-    }
-    bool err;
-    I i1 = StrToInt64(col[0], filename, nline, allow_negative, &err);
-    if (err) return false;
-    I i2 = StrToInt64(col[1], filename, nline, allow_negative, &err);
-    if (err) return false;
-    pairs->push_back(std::make_pair(i1, i2));
-  }
-  return true;
-}
-
-template <typename I>
-bool WriteIntPairs(const string &filename,
-                   const std::vector<std::pair<I, I>> &pairs) {
-  std::ostream *strm = &std::cout;
-  if (!filename.empty()) {
-    strm = new std::ofstream(filename);
-    if (!*strm) {
-      LOG(ERROR) << "WriteIntPairs: Can't open file: " << filename;
-      return false;
-    }
-  }
-  for (ssize_t n = 0; n < pairs.size(); ++n) {
-    *strm << pairs[n].first << "\t" << pairs[n].second << "\n";
-  }
-  if (!*strm) {
-    LOG(ERROR) << "WriteIntPairs: Write failed: "
-               << (filename.empty() ? "standard output" : filename);
-    return false;
-  }
-  if (strm != &std::cout) delete strm;
-  return true;
-}
-
-// Utilities for reading/writing label pairs.
-
-template <typename Label>
-bool ReadLabelPairs(const string &filename,
-                    std::vector<std::pair<Label, Label>> *pairs,
-                    bool allow_negative = false) {
-  return ReadIntPairs(filename, pairs, allow_negative);
-}
-
-template <typename Label>
-bool WriteLabelPairs(const string &filename,
-                     const std::vector<std::pair<Label, Label>> &pairs) {
-  return WriteIntPairs(filename, pairs);
-}
-
-// Utilities for converting a type name to a legal C symbol.
-
-void ConvertToLegalCSymbol(string *s);
-
-// Utilities for stream I/O.
-
-bool AlignInput(std::istream &strm);
-bool AlignOutput(std::ostream &strm);
-
-// An associative container for which testing membership is faster than an STL
-// set if members are restricted to an interval that excludes most non-members.
-// A Key must have ==, !=, and < operators defined. Element NoKey should be a
-// key that marks an uninitialized key and is otherwise unused. Find() returns
-// an STL const_iterator to the match found, otherwise it equals End().
-template <class Key, Key NoKey>
-class CompactSet {
- public:
-  using const_iterator = typename std::set<Key>::const_iterator;
-
-  CompactSet() : min_key_(NoKey), max_key_(NoKey) {}
-
-  CompactSet(const CompactSet<Key, NoKey> &compact_set)
-      : set_(compact_set.set_),
-        min_key_(compact_set.min_key_),
-        max_key_(compact_set.max_key_) {}
-
-  void Insert(Key key) {
-    set_.insert(key);
-    if (min_key_ == NoKey || key < min_key_) min_key_ = key;
-    if (max_key_ == NoKey || max_key_ < key) max_key_ = key;
-  }
-
-  void Erase(Key key) {
-    set_.erase(key);
-    if (set_.empty()) {
-      min_key_ = max_key_ = NoKey;
-    } else if (key == min_key_) {
-      ++min_key_;
-    } else if (key == max_key_) {
-      --max_key_;
-    }
-  }
-
-  void Clear() {
-    set_.clear();
-    min_key_ = max_key_ = NoKey;
-  }
-
-  const_iterator Find(Key key) const {
-    if (min_key_ == NoKey || key < min_key_ || max_key_ < key) {
-      return set_.end();
-    } else {
-      return set_.find(key);
-    }
-  }
-
-  bool Member(Key key) const {
-    if (min_key_ == NoKey || key < min_key_ || max_key_ < key) {
-      return false;  // out of range
-    } else if (min_key_ != NoKey && max_key_ + 1 == min_key_ + set_.size()) {
-      return true;  // dense range
-    } else {
-      return set_.count(key);
-    }
-  }
-
-  const_iterator Begin() const { return set_.begin(); }
-
-  const_iterator End() const { return set_.end(); }
-
-  // All stored keys are greater than or equal to this value.
-  Key LowerBound() const { return min_key_; }
-
-  // All stored keys are less than or equal to this value.
-  Key UpperBound() const { return max_key_; }
-
- private:
-  std::set<Key> set_;
-  Key min_key_;
-  Key max_key_;
-
-  void operator=(const CompactSet &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_UTIL_H_
diff --git a/projects/llm_framework/include/fst/vector-fst.h b/projects/llm_framework/include/fst/vector-fst.h
deleted file mode 100644
index 7514bc55..00000000
--- a/projects/llm_framework/include/fst/vector-fst.h
+++ /dev/null
@@ -1,796 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Simple concrete, mutable FST whose states and arcs are stored in STL vectors.
-
-#ifndef FST_VECTOR_FST_H_
-#define FST_VECTOR_FST_H_
-
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/fst-decl.h>  // For optional argument declarations
-#include <fst/mutable-fst.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-template <class A, class S>
-class VectorFst;
-
-template <class F, class G>
-void Cast(const F &, G *);
-
-// Arcs (of type A) implemented by an STL vector per state. M specifies Arc
-// allocator (default declared in fst-decl.h).
-template <class A, class M /* = std::allocator<A> */>
-class VectorState {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using ArcAllocator = M;
-  using StateAllocator =
-      typename ArcAllocator::template rebind<VectorState<Arc, M>>::other;
-
-  // Provide STL allocator for arcs.
-  explicit VectorState(const ArcAllocator &alloc)
-      : final_(Weight::Zero()), niepsilons_(0), noepsilons_(0), arcs_(alloc) {}
-
-  VectorState(const VectorState<A, M> &state, const ArcAllocator &alloc)
-      : final_(state.Final()),
-        niepsilons_(state.NumInputEpsilons()),
-        noepsilons_(state.NumOutputEpsilons()),
-        arcs_(state.arcs_.begin(), state.arcs_.end(), alloc) {}
-
-  void Reset() {
-    final_ = Weight::Zero();
-    niepsilons_ = 0;
-    noepsilons_ = 0;
-    arcs_.clear();
-  }
-
-  Weight Final() const { return final_; }
-
-  size_t NumInputEpsilons() const { return niepsilons_; }
-
-  size_t NumOutputEpsilons() const { return noepsilons_; }
-
-  size_t NumArcs() const { return arcs_.size(); }
-
-  const Arc &GetArc(size_t n) const { return arcs_[n]; }
-
-  const Arc *Arcs() const { return !arcs_.empty() ? &arcs_[0] : nullptr; }
-
-  Arc *MutableArcs() { return !arcs_.empty() ? &arcs_[0] : nullptr; }
-
-  void ReserveArcs(size_t n) { arcs_.reserve(n); }
-
-  void SetFinal(Weight weight) { final_ = std::move(weight); }
-
-  void SetNumInputEpsilons(size_t n) { niepsilons_ = n; }
-
-  void SetNumOutputEpsilons(size_t n) { noepsilons_ = n; }
-
-  void AddArc(const Arc &arc) {
-    IncrementNumEpsilons(arc);
-    arcs_.push_back(arc);
-  }
-
-  void AddArc(Arc &&arc) {
-    IncrementNumEpsilons(arc);
-    arcs_.push_back(std::move(arc));
-  }
-
-  template <class... T>
-  void EmplaceArc(T&&... ctor_args) {
-    arcs_.emplace_back(std::forward<T>(ctor_args)...);
-    IncrementNumEpsilons(arcs_.back());
-  }
-
-  void SetArc(const Arc &arc, size_t n) {
-    if (arcs_[n].ilabel == 0) --niepsilons_;
-    if (arcs_[n].olabel == 0) --noepsilons_;
-    IncrementNumEpsilons(arc);
-    arcs_[n] = arc;
-  }
-
-  void DeleteArcs() {
-    niepsilons_ = 0;
-    noepsilons_ = 0;
-    arcs_.clear();
-  }
-
-  void DeleteArcs(size_t n) {
-    for (size_t i = 0; i < n; ++i) {
-      if (arcs_.back().ilabel == 0) --niepsilons_;
-      if (arcs_.back().olabel == 0) --noepsilons_;
-      arcs_.pop_back();
-    }
-  }
-
-  // For state class allocation.
-  void *operator new(size_t size, StateAllocator *alloc) {
-    return alloc->allocate(1);
-  }
-
-  // For state destruction and memory freeing.
-  static void Destroy(VectorState<A, M> *state, StateAllocator *alloc) {
-    if (state) {
-      state->~VectorState<A, M>();
-      alloc->deallocate(state, 1);
-    }
-  }
-
- private:
-  // Update the number of epsilons as a result of having added an arc.
-  void IncrementNumEpsilons(const Arc &arc) {
-    if (arc.ilabel == 0) ++niepsilons_;
-    if (arc.olabel == 0) ++noepsilons_;
-  }
-
-  Weight final_;                       // Final weight.
-  size_t niepsilons_;                  // # of input epsilons
-  size_t noepsilons_;                  // # of output epsilons
-  std::vector<A, ArcAllocator> arcs_;  // Arc container.
-};
-
-namespace internal {
-
-// States are implemented by STL vectors, templated on the
-// State definition. This does not manage the Fst properties.
-template <class S>
-class VectorFstBaseImpl : public FstImpl<typename S::Arc> {
- public:
-  using State = S;
-  using Arc = typename State::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  VectorFstBaseImpl() : start_(kNoStateId) {}
-
-  ~VectorFstBaseImpl() override {
-    for (size_t s = 0; s < states_.size(); ++s) {
-      State::Destroy(states_[s], &state_alloc_);
-    }
-  }
-
-  // Copying is not permitted.
-  VectorFstBaseImpl(const VectorFstBaseImpl<S> &) = delete;
-  VectorFstBaseImpl<S> &operator=(const VectorFstBaseImpl<S> &) = delete;
-
-  // Moving is permitted.
-  VectorFstBaseImpl(VectorFstBaseImpl<S> &&impl) noexcept
-      : FstImpl<typename S::Arc>(),
-        states_(std::move(impl.states_)),
-        start_(impl.start_) {
-    impl.states_.clear();
-    impl.start_ = kNoStateId;
-  }
-
-  VectorFstBaseImpl<S> &operator=(VectorFstBaseImpl<S> &&impl) noexcept {
-    states_ = std::move(impl.states_);
-    start_ = impl.start_;
-    impl.states_.clear();
-    impl.start_ = kNoStateId;
-    return *this;
-  }
-
-  StateId Start() const { return start_; }
-
-  Weight Final(StateId state) const { return states_[state]->Final(); }
-
-  StateId NumStates() const { return states_.size(); }
-
-  size_t NumArcs(StateId state) const { return states_[state]->NumArcs(); }
-
-  size_t NumInputEpsilons(StateId state) const {
-    return GetState(state)->NumInputEpsilons();
-  }
-
-  size_t NumOutputEpsilons(StateId state) const {
-    return GetState(state)->NumOutputEpsilons();
-  }
-
-  void SetStart(StateId state) { start_ = state; }
-
-  void SetFinal(StateId state, Weight weight) {
-    states_[state]->SetFinal(std::move(weight));
-  }
-
-  StateId AddState() {
-    states_.push_back(new (&state_alloc_) State(arc_alloc_));
-    return states_.size() - 1;
-  }
-
-  StateId AddState(State *state) {
-    states_.push_back(state);
-    return states_.size() - 1;
-  }
-
-  void AddArc(StateId state, const Arc &arc) { states_[state]->AddArc(arc); }
-
-  void AddArc(StateId state, Arc &&arc) {
-    states_[state]->AddArc(std::move(arc));
-  }
-
-  template <class... T>
-  void EmplaceArc(StateId state, T&&... ctor_args) {
-    states_[state]->EmplaceArc(std::forward<T>(ctor_args)...);
-  }
-
-  void DeleteStates(const std::vector<StateId> &dstates) {
-    std::vector<StateId> newid(states_.size(), 0);
-    for (size_t i = 0; i < dstates.size(); ++i) newid[dstates[i]] = kNoStateId;
-    StateId nstates = 0;
-    for (StateId state = 0; state < states_.size(); ++state) {
-      if (newid[state] != kNoStateId) {
-        newid[state] = nstates;
-        if (state != nstates) states_[nstates] = states_[state];
-        ++nstates;
-      } else {
-        State::Destroy(states_[state], &state_alloc_);
-      }
-    }
-    states_.resize(nstates);
-    for (StateId state = 0; state < states_.size(); ++state) {
-      auto *arcs = states_[state]->MutableArcs();
-      size_t narcs = 0;
-      auto nieps = states_[state]->NumInputEpsilons();
-      auto noeps = states_[state]->NumOutputEpsilons();
-      for (size_t i = 0; i < states_[state]->NumArcs(); ++i) {
-        const auto t = newid[arcs[i].nextstate];
-        if (t != kNoStateId) {
-          arcs[i].nextstate = t;
-          if (i != narcs) arcs[narcs] = arcs[i];
-          ++narcs;
-        } else {
-          if (arcs[i].ilabel == 0) --nieps;
-          if (arcs[i].olabel == 0) --noeps;
-        }
-      }
-      states_[state]->DeleteArcs(states_[state]->NumArcs() - narcs);
-      states_[state]->SetNumInputEpsilons(nieps);
-      states_[state]->SetNumOutputEpsilons(noeps);
-    }
-    if (Start() != kNoStateId) SetStart(newid[Start()]);
-  }
-
-  void DeleteStates() {
-    for (size_t state = 0; state < states_.size(); ++state) {
-      State::Destroy(states_[state], &state_alloc_);
-    }
-    states_.clear();
-    SetStart(kNoStateId);
-  }
-
-  void DeleteArcs(StateId state, size_t n) { states_[state]->DeleteArcs(n); }
-
-  void DeleteArcs(StateId state) { states_[state]->DeleteArcs(); }
-
-  State *GetState(StateId state) { return states_[state]; }
-
-  const State *GetState(StateId state) const { return states_[state]; }
-
-  void SetState(StateId state, State *vstate) { states_[state] = vstate; }
-
-  void ReserveStates(StateId n) { states_.reserve(n); }
-
-  void ReserveArcs(StateId state, size_t n) { states_[state]->ReserveArcs(n); }
-
-  // Provide information needed for generic state iterator.
-  void InitStateIterator(StateIteratorData<Arc> *data) const {
-    data->base = nullptr;
-    data->nstates = states_.size();
-  }
-
-  // Provide information needed for generic arc iterator.
-  void InitArcIterator(StateId state, ArcIteratorData<Arc> *data) const {
-    data->base = nullptr;
-    data->narcs = states_[state]->NumArcs();
-    data->arcs = states_[state]->Arcs();
-    data->ref_count = nullptr;
-  }
-
- private:
-  std::vector<State *> states_;                 // States represenation.
-  StateId start_;                               // Initial state.
-  typename State::StateAllocator state_alloc_;  // For state allocation.
-  typename State::ArcAllocator arc_alloc_;      // For arc allocation.
-};
-
-// This is a VectorFstBaseImpl container that holds VectorStates and manages FST
-// properties.
-template <class S>
-class VectorFstImpl : public VectorFstBaseImpl<S> {
- public:
-  using State = S;
-  using Arc = typename State::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::Properties;
-
-  using VectorFstBaseImpl<S>::Start;
-  using VectorFstBaseImpl<S>::NumStates;
-  using VectorFstBaseImpl<S>::GetState;
-  using VectorFstBaseImpl<S>::ReserveArcs;
-
-  friend class MutableArcIterator<VectorFst<Arc, S>>;
-
-  using BaseImpl = VectorFstBaseImpl<S>;
-
-  VectorFstImpl() {
-    SetType("vector");
-    SetProperties(kNullProperties | kStaticProperties);
-  }
-
-  explicit VectorFstImpl(const Fst<Arc> &fst);
-
-  static VectorFstImpl<S> *Read(std::istream &strm, const FstReadOptions &opts);
-
-  void SetStart(StateId state) {
-    BaseImpl::SetStart(state);
-    SetProperties(SetStartProperties(Properties()));
-  }
-
-  void SetFinal(StateId state, Weight weight) {
-    const auto old_weight = BaseImpl::Final(state);
-    const auto properties =
-        SetFinalProperties(Properties(), old_weight, weight);
-    BaseImpl::SetFinal(state, std::move(weight));
-    SetProperties(properties);
-  }
-
-  StateId AddState() {
-    const auto state = BaseImpl::AddState();
-    SetProperties(AddStateProperties(Properties()));
-    return state;
-  }
-
-  void AddArc(StateId state, const Arc &arc) {
-    BaseImpl::AddArc(state, arc);
-    UpdatePropertiesAfterAddArc(state);
-  }
-
-  void AddArc(StateId state, Arc &&arc) {
-    BaseImpl::AddArc(state, std::move(arc));
-    UpdatePropertiesAfterAddArc(state);
-  }
-
-  template <class... T>
-  void EmplaceArc(StateId state, T&&... ctor_args) {
-    BaseImpl::EmplaceArc(state, std::forward<T>(ctor_args)...);
-    UpdatePropertiesAfterAddArc(state);
-  }
-
-  void DeleteStates(const std::vector<StateId> &dstates) {
-    BaseImpl::DeleteStates(dstates);
-    SetProperties(DeleteStatesProperties(Properties()));
-  }
-
-  void DeleteStates() {
-    BaseImpl::DeleteStates();
-    SetProperties(DeleteAllStatesProperties(Properties(), kStaticProperties));
-  }
-
-  void DeleteArcs(StateId state, size_t n) {
-    BaseImpl::DeleteArcs(state, n);
-    SetProperties(DeleteArcsProperties(Properties()));
-  }
-
-  void DeleteArcs(StateId state) {
-    BaseImpl::DeleteArcs(state);
-    SetProperties(DeleteArcsProperties(Properties()));
-  }
-
-  // Properties always true of this FST class
-  static constexpr uint64 kStaticProperties = kExpanded | kMutable;
-
- private:
-  void UpdatePropertiesAfterAddArc(StateId state) {
-    auto *vstate = GetState(state);
-    const size_t num_arcs{vstate->NumArcs()};
-    if (num_arcs) {
-      const auto &arc = vstate->GetArc(num_arcs - 1);
-      const auto *parc = (num_arcs < 2)
-                         ? nullptr
-                         : &(vstate->GetArc(num_arcs - 2));
-      SetProperties(AddArcProperties(Properties(), state, arc, parc));
-    }
-  }
-
-  // Minimum file format version supported.
-  static constexpr int kMinFileVersion = 2;
-};
-
-template <class S>
-constexpr uint64 VectorFstImpl<S>::kStaticProperties;
-
-template <class S>
-constexpr int VectorFstImpl<S>::kMinFileVersion;
-
-template <class S>
-VectorFstImpl<S>::VectorFstImpl(const Fst<Arc> &fst) {
-  SetType("vector");
-  SetInputSymbols(fst.InputSymbols());
-  SetOutputSymbols(fst.OutputSymbols());
-  BaseImpl::SetStart(fst.Start());
-  if (fst.Properties(kExpanded, false)) {
-    BaseImpl::ReserveStates(CountStates(fst));
-  }
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    const auto state = siter.Value();
-    BaseImpl::AddState();
-    BaseImpl::SetFinal(state, fst.Final(state));
-    ReserveArcs(state, fst.NumArcs(state));
-    for (ArcIterator<Fst<Arc>> aiter(fst, state); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      BaseImpl::AddArc(state, arc);
-    }
-  }
-  SetProperties(fst.Properties(kCopyProperties, false) | kStaticProperties);
-}
-
-template <class S>
-VectorFstImpl<S> *VectorFstImpl<S>::Read(std::istream &strm,
-                                         const FstReadOptions &opts) {
-  std::unique_ptr<VectorFstImpl<S>> impl(new VectorFstImpl());
-  FstHeader hdr;
-  if (!impl->ReadHeader(strm, opts, kMinFileVersion, &hdr)) return nullptr;
-  impl->BaseImpl::SetStart(hdr.Start());
-  if (hdr.NumStates() != kNoStateId) impl->ReserveStates(hdr.NumStates());
-  StateId state = 0;
-  for (; hdr.NumStates() == kNoStateId || state < hdr.NumStates(); ++state) {
-    Weight weight;
-    if (!weight.Read(strm)) break;
-    impl->BaseImpl::AddState();
-    auto *vstate = impl->GetState(state);
-    vstate->SetFinal(weight);
-    int64 narcs;
-    ReadType(strm, &narcs);
-    if (!strm) {
-      LOG(ERROR) << "VectorFst::Read: Read failed: " << opts.source;
-      return nullptr;
-    }
-    impl->ReserveArcs(state, narcs);
-    for (int64 i = 0; i < narcs; ++i) {
-      Arc arc;
-      ReadType(strm, &arc.ilabel);
-      ReadType(strm, &arc.olabel);
-      arc.weight.Read(strm);
-      ReadType(strm, &arc.nextstate);
-      if (!strm) {
-        LOG(ERROR) << "VectorFst::Read: Read failed: " << opts.source;
-        return nullptr;
-      }
-      impl->BaseImpl::AddArc(state, std::move(arc));
-    }
-  }
-  if (hdr.NumStates() != kNoStateId && state != hdr.NumStates()) {
-    LOG(ERROR) << "VectorFst::Read: Unexpected end of file: " << opts.source;
-    return nullptr;
-  }
-  return impl.release();
-}
-
-}  // namespace internal
-
-// Simple concrete, mutable FST. This class attaches interface to implementation
-// and handles reference counting, delegating most methods to ImplToMutableFst.
-// Also supports ReserveStates and ReserveArcs methods (cf. STL vector methods).
-// The second optional template argument gives the State definition.
-template <class A, class S /* = VectorState<A> */>
-class VectorFst : public ImplToMutableFst<internal::VectorFstImpl<S>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  using State = S;
-  using Impl = internal::VectorFstImpl<State>;
-
-  friend class StateIterator<VectorFst<Arc, State>>;
-  friend class ArcIterator<VectorFst<Arc, State>>;
-  friend class MutableArcIterator<VectorFst<A, S>>;
-
-  template <class F, class G>
-  friend void Cast(const F &, G *);
-
-  VectorFst() : ImplToMutableFst<Impl>(std::make_shared<Impl>()) {}
-
-  explicit VectorFst(const Fst<Arc> &fst)
-      : ImplToMutableFst<Impl>(std::make_shared<Impl>(fst)) {}
-
-  VectorFst(const VectorFst<Arc, State> &fst, bool safe = false)
-      : ImplToMutableFst<Impl>(fst) {}
-
-  VectorFst(VectorFst<Arc, State> &&) noexcept;
-
-  // Get a copy of this VectorFst. See Fst<>::Copy() for further doc.
-  VectorFst<Arc, State> *Copy(bool safe = false) const override {
-    return new VectorFst<Arc, State>(*this, safe);
-  }
-
-  VectorFst<Arc, State> &operator=(const VectorFst<Arc, State> &) = default;
-
-  VectorFst<Arc, State> &operator=(VectorFst<Arc, State> &&) noexcept;
-
-  VectorFst<Arc, State> &operator=(const Fst<Arc> &fst) override {
-    if (this != &fst) SetImpl(std::make_shared<Impl>(fst));
-    return *this;
-  }
-
-  template <class... T>
-  void EmplaceArc(StateId state, T&&... ctor_args) {
-    MutateCheck();
-    GetMutableImpl()->EmplaceArc(state, std::forward<T>(ctor_args)...);
-  }
-
-  // Reads a VectorFst from an input stream, returning nullptr on error.
-  static VectorFst<Arc, State> *Read(std::istream &strm,
-                                     const FstReadOptions &opts) {
-    auto *impl = Impl::Read(strm, opts);
-    return impl ? new VectorFst<Arc, State>(std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  // Read a VectorFst from a file, returning nullptr on error; empty filename
-  // reads from standard input.
-  static VectorFst<Arc, State> *Read(const string &filename) {
-    auto *impl = ImplToExpandedFst<Impl, MutableFst<Arc>>::Read(filename);
-    return impl ? new VectorFst<Arc, State>(std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return WriteFst(*this, strm, opts);
-  }
-
-  bool Write(const string &filename) const override {
-    return Fst<Arc>::WriteFile(filename);
-  }
-
-  template <class FST>
-  static bool WriteFst(const FST &fst, std::ostream &strm,
-                       const FstWriteOptions &opts);
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const override {
-    GetImpl()->InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetImpl()->InitArcIterator(s, data);
-  }
-
-  inline void InitMutableArcIterator(StateId s,
-                                     MutableArcIteratorData<Arc> *) override;
-
-  using ImplToMutableFst<Impl, MutableFst<Arc>>::ReserveArcs;
-  using ImplToMutableFst<Impl, MutableFst<Arc>>::ReserveStates;
-
- private:
-  using ImplToMutableFst<Impl, MutableFst<Arc>>::GetImpl;
-  using ImplToMutableFst<Impl, MutableFst<Arc>>::GetMutableImpl;
-  using ImplToMutableFst<Impl, MutableFst<Arc>>::MutateCheck;
-  using ImplToMutableFst<Impl, MutableFst<Arc>>::SetImpl;
-
-  explicit VectorFst(std::shared_ptr<Impl> impl)
-      : ImplToMutableFst<Impl>(impl) {}
-};
-
-template <class Arc, class State>
-inline VectorFst<Arc, State>::VectorFst(
-    VectorFst<Arc, State> &&fst) noexcept = default;
-
-template <class Arc, class State>
-inline VectorFst<Arc, State> &VectorFst<Arc, State>::operator=(
-    VectorFst<Arc, State> &&fst) noexcept = default;
-
-// Writes FST to file in Vector format, potentially with a pass over the machine
-// before writing to compute number of states.
-template <class Arc, class State>
-template <class FST>
-bool VectorFst<Arc, State>::WriteFst(const FST &fst, std::ostream &strm,
-                                     const FstWriteOptions &opts) {
-  static constexpr int file_version = 2;
-  bool update_header = true;
-  FstHeader hdr;
-  hdr.SetStart(fst.Start());
-  hdr.SetNumStates(kNoStateId);
-  std::streampos start_offset = 0;
-  if (fst.Properties(kExpanded, false) || opts.stream_write ||
-      (start_offset = strm.tellp()) != -1) {
-    hdr.SetNumStates(CountStates(fst));
-    update_header = false;
-  }
-  const auto properties =
-      fst.Properties(kCopyProperties, false) | Impl::kStaticProperties;
-  internal::FstImpl<Arc>::WriteFstHeader(fst, strm, opts, file_version,
-                                         "vector", properties, &hdr);
-  StateId num_states = 0;
-  for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    fst.Final(s).Write(strm);
-    const int64 narcs = fst.NumArcs(s);
-    WriteType(strm, narcs);
-    for (ArcIterator<FST> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      WriteType(strm, arc.ilabel);
-      WriteType(strm, arc.olabel);
-      arc.weight.Write(strm);
-      WriteType(strm, arc.nextstate);
-    }
-    ++num_states;
-  }
-  strm.flush();
-  if (!strm) {
-    LOG(ERROR) << "VectorFst::Write: Write failed: " << opts.source;
-    return false;
-  }
-  if (update_header) {
-    hdr.SetNumStates(num_states);
-    return internal::FstImpl<Arc>::UpdateFstHeader(
-        fst, strm, opts, file_version, "vector", properties, &hdr,
-        start_offset);
-  } else {
-    if (num_states != hdr.NumStates()) {
-      LOG(ERROR) << "Inconsistent number of states observed during write";
-      return false;
-    }
-  }
-  return true;
-}
-
-// Specialization for VectorFst; see generic version in fst.h for sample usage
-// (but use the VectorFst type instead). This version should inline.
-template <class Arc, class State>
-class StateIterator<VectorFst<Arc, State>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  explicit StateIterator(const VectorFst<Arc, State> &fst)
-      : nstates_(fst.GetImpl()->NumStates()), s_(0) {}
-
-  bool Done() const { return s_ >= nstates_; }
-
-  StateId Value() const { return s_; }
-
-  void Next() { ++s_; }
-
-  void Reset() { s_ = 0; }
-
- private:
-  const StateId nstates_;
-  StateId s_;
-};
-
-// Specialization for VectorFst; see generic version in fst.h for sample usage
-// (but use the VectorFst type instead). This version should inline.
-template <class Arc, class State>
-class ArcIterator<VectorFst<Arc, State>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const VectorFst<Arc, State> &fst, StateId s)
-      : arcs_(fst.GetImpl()->GetState(s)->Arcs()),
-        narcs_(fst.GetImpl()->GetState(s)->NumArcs()),
-        i_(0) {}
-
-  bool Done() const { return i_ >= narcs_; }
-
-  const Arc &Value() const { return arcs_[i_]; }
-
-  void Next() { ++i_; }
-
-  void Reset() { i_ = 0; }
-
-  void Seek(size_t a) { i_ = a; }
-
-  size_t Position() const { return i_; }
-
-  constexpr uint32 Flags() const { return kArcValueFlags; }
-
-  void SetFlags(uint32, uint32) {}
-
- private:
-  const Arc *arcs_;
-  size_t narcs_;
-  size_t i_;
-};
-
-// Specialization for VectorFst; see generic version in mutable-fst.h for sample
-// usage (but use the VectorFst type instead). This version should inline.
-template <class Arc, class State>
-class MutableArcIterator<VectorFst<Arc, State>>
-    : public MutableArcIteratorBase<Arc> {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  MutableArcIterator(VectorFst<Arc, State> *fst, StateId s) : i_(0) {
-    fst->MutateCheck();
-    state_ = fst->GetMutableImpl()->GetState(s);
-    properties_ = &fst->GetImpl()->properties_;
-  }
-
-  bool Done() const final { return i_ >= state_->NumArcs(); }
-
-  const Arc &Value() const final { return state_->GetArc(i_); }
-
-  void Next() final { ++i_; }
-
-  size_t Position() const final { return i_; }
-
-  void Reset() final { i_ = 0; }
-
-  void Seek(size_t a) final { i_ = a; }
-
-  void SetValue(const Arc &arc) final {
-    const auto &oarc = state_->GetArc(i_);
-    if (oarc.ilabel != oarc.olabel) *properties_ &= ~kNotAcceptor;
-    if (oarc.ilabel == 0) {
-      *properties_ &= ~kIEpsilons;
-      if (oarc.olabel == 0) *properties_ &= ~kEpsilons;
-    }
-    if (oarc.olabel == 0) *properties_ &= ~kOEpsilons;
-    if (oarc.weight != Weight::Zero() && oarc.weight != Weight::One()) {
-      *properties_ &= ~kWeighted;
-    }
-    state_->SetArc(arc, i_);
-    if (arc.ilabel != arc.olabel) {
-      *properties_ |= kNotAcceptor;
-      *properties_ &= ~kAcceptor;
-    }
-    if (arc.ilabel == 0) {
-      *properties_ |= kIEpsilons;
-      *properties_ &= ~kNoIEpsilons;
-      if (arc.olabel == 0) {
-        *properties_ |= kEpsilons;
-        *properties_ &= ~kNoEpsilons;
-      }
-    }
-    if (arc.olabel == 0) {
-      *properties_ |= kOEpsilons;
-      *properties_ &= ~kNoOEpsilons;
-    }
-    if (arc.weight != Weight::Zero() && arc.weight != Weight::One()) {
-      *properties_ |= kWeighted;
-      *properties_ &= ~kUnweighted;
-    }
-    *properties_ &= kSetArcProperties | kAcceptor | kNotAcceptor | kEpsilons |
-                    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons |
-                    kNoOEpsilons | kWeighted | kUnweighted;
-  }
-
-  uint32 Flags() const final { return kArcValueFlags; }
-
-  void SetFlags(uint32, uint32) final {}
-
- private:
-  State *state_;
-  uint64 *properties_;
-  size_t i_;
-};
-
-// Provides information needed for the generic mutable arc iterator.
-template <class Arc, class State>
-inline void VectorFst<Arc, State>::InitMutableArcIterator(
-    StateId s, MutableArcIteratorData<Arc> *data) {
-  data->base = new MutableArcIterator<VectorFst<Arc, State>>(this, s);
-}
-
-// A useful alias when using StdArc.
-using StdVectorFst = VectorFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_VECTOR_FST_H_
diff --git a/projects/llm_framework/include/fst/verify.h b/projects/llm_framework/include/fst/verify.h
deleted file mode 100644
index 2ea8a64a..00000000
--- a/projects/llm_framework/include/fst/verify.h
+++ /dev/null
@@ -1,100 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function to verify an FST's contents.
-
-#ifndef FST_VERIFY_H_
-#define FST_VERIFY_H_
-
-#include <fst/log.h>
-
-#include <fst/fst.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-// Verifies that an Fst's contents are sane.
-template <class Arc>
-bool Verify(const Fst<Arc> &fst, bool allow_negative_labels = false) {
-  using StateId = typename Arc::StateId;
-  const auto start = fst.Start();
-  const auto *isyms = fst.InputSymbols();
-  const auto *osyms = fst.OutputSymbols();
-  // Count states
-  StateId ns = 0;
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) ++ns;
-  if (start == kNoStateId && ns > 0) {
-    LOG(ERROR) << "Verify: FST start state ID not set";
-    return false;
-  } else if (start >= ns) {
-    LOG(ERROR) << "Verify: FST start state ID exceeds number of states";
-    return false;
-  }
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    auto state = siter.Value();
-    size_t na = 0;
-    for (ArcIterator<Fst<Arc>> aiter(fst, state); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      if (!allow_negative_labels && arc.ilabel < 0) {
-        LOG(ERROR) << "Verify: FST input label ID of arc at position " << na
-                   << " of state " << state << " is negative";
-        return false;
-      } else if (isyms && isyms->Find(arc.ilabel).empty()) {
-        LOG(ERROR) << "Verify: FST input label ID " << arc.ilabel
-                   << " of arc at position " << na << " of state " << state
-                   << " is missing from input symbol table \"" << isyms->Name()
-                   << "\"";
-        return false;
-      } else if (!allow_negative_labels && arc.olabel < 0) {
-        LOG(ERROR) << "Verify: FST output label ID of arc at position " << na
-                   << " of state " << state << " is negative";
-        return false;
-      } else if (osyms && osyms->Find(arc.olabel).empty()) {
-        LOG(ERROR) << "Verify: FST output label ID " << arc.olabel
-                   << " of arc at position " << na << " of state " << state
-                   << " is missing from output symbol table \"" << osyms->Name()
-                   << "\"";
-        return false;
-      } else if (!arc.weight.Member()) {
-        LOG(ERROR) << "Verify: FST weight of arc at position " << na
-                   << " of state " << state << " is invalid";
-        return false;
-      } else if (arc.nextstate < 0) {
-        LOG(ERROR) << "Verify: FST destination state ID of arc at position "
-                   << na << " of state " << state << " is negative";
-        return false;
-      } else if (arc.nextstate >= ns) {
-        LOG(ERROR) << "Verify: FST destination state ID of arc at position "
-                   << na << " of state " << state
-                   << " exceeds number of states";
-        return false;
-      }
-      ++na;
-    }
-    if (!fst.Final(state).Member()) {
-      LOG(ERROR) << "Verify: FST final weight of state " << state
-                 << " is invalid";
-      return false;
-    }
-  }
-  const auto fst_props = fst.Properties(kFstProperties, false);
-  if (fst_props & kError) {
-    LOG(ERROR) << "Verify: FST error property is set";
-    return false;
-  }
-  uint64 known_props;
-  uint64 test_props =
-      ComputeProperties(fst, kFstProperties, &known_props, false);
-  if (!CompatProperties(fst_props, test_props)) {
-    LOG(ERROR) << "Verify: Stored FST properties incorrect "
-               << "(props1 = stored props, props2 = tested)";
-    return false;
-  } else {
-    return true;
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_VERIFY_H_
diff --git a/projects/llm_framework/include/fst/visit.h b/projects/llm_framework/include/fst/visit.h
deleted file mode 100644
index c6658047..00000000
--- a/projects/llm_framework/include/fst/visit.h
+++ /dev/null
@@ -1,321 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Queue-dependent visitation of finite-state transducers. See also dfs-visit.h.
-
-#ifndef FST_VISIT_H_
-#define FST_VISIT_H_
-
-
-#include <fst/arcfilter.h>
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-// Visitor Interface: class determining actions taken during a visit. If any of
-// the boolean member functions return false, the visit is aborted by first
-// calling FinishState() on all unfinished (grey) states and then calling
-// FinishVisit().
-//
-// Note this is more general than the visitor interface in dfs-visit.h but lacks
-// some DFS-specific behavior.
-//
-// template <class Arc>
-// class Visitor {
-//  public:
-//   using StateId = typename Arc::StateId;
-//
-//   Visitor(T *return_data);
-//
-//   // Invoked before visit.
-//   void InitVisit(const Fst<Arc> &fst);
-//
-//   // Invoked when state discovered (2nd arg is visitation root).
-//   bool InitState(StateId s, StateId root);
-//
-//   // Invoked when arc to white/undiscovered state examined.
-//   bool WhiteArc(StateId s, const Arc &arc);
-//
-//   // Invoked when arc to grey/unfinished state examined.
-//   bool GreyArc(StateId s, const Arc &arc);
-//
-//   // Invoked when arc to black/finished state examined.
-//   bool BlackArc(StateId s, const Arc &arc);
-//
-//   // Invoked when state finished.
-//   void FinishState(StateId s);
-//
-//   // Invoked after visit.
-//   void FinishVisit();
-// };
-
-// Performs queue-dependent visitation. Visitor class argument determines
-// actions and contains any return data. ArcFilter determines arcs that are
-// considered. If 'access_only' is true, performs visitation only to states
-// accessible from the initial state.
-template <class FST, class Visitor, class Queue, class ArcFilter>
-void Visit(const FST &fst, Visitor *visitor, Queue *queue, ArcFilter filter,
-           bool access_only = false) {
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-  visitor->InitVisit(fst);
-  const auto start = fst.Start();
-  if (start == kNoStateId) {
-    visitor->FinishVisit();
-    return;
-  }
-  // An FST's state's visit color.
-  static constexpr uint8 kWhiteState = 0x01;  // Undiscovered.
-  static constexpr uint8 kGreyState = 0x02;   // Discovered & unfinished.
-  static constexpr uint8 kBlackState = 0x04;  // Finished.
-  // We destroy an iterator as soon as possible and mark it so.
-  static constexpr uint8 kArcIterDone = 0x08;
-  std::vector<uint8> state_status;
-  std::vector<ArcIterator<FST> *> arc_iterator;
-  MemoryPool<ArcIterator<FST>> aiter_pool;
-  StateId nstates = start + 1;  // Number of known states in general case.
-  bool expanded = false;
-  if (fst.Properties(kExpanded, false)) {  // Tests if expanded, then uses
-    nstates = CountStates(fst);            // ExpandedFst::NumStates().
-    expanded = true;
-  }
-  state_status.resize(nstates, kWhiteState);
-  arc_iterator.resize(nstates);
-  StateIterator<Fst<Arc>> siter(fst);
-  // Continues visit while true.
-  bool visit = true;
-  // Iterates over trees in visit forest.
-  for (auto root = start; visit && root < nstates;) {
-    visit = visitor->InitState(root, root);
-    state_status[root] = kGreyState;
-    queue->Enqueue(root);
-    while (!queue->Empty()) {
-      auto state = queue->Head();
-      if (state >= state_status.size()) {
-        nstates = state + 1;
-        state_status.resize(nstates, kWhiteState);
-        arc_iterator.resize(nstates);
-      }
-      // Creates arc iterator if needed.
-      if (!arc_iterator[state] && !(state_status[state] & kArcIterDone) &&
-          visit) {
-        arc_iterator[state] = new (&aiter_pool) ArcIterator<FST>(fst, state);
-      }
-      // Deletes arc iterator if done.
-      auto *aiter = arc_iterator[state];
-      if ((aiter && aiter->Done()) || !visit) {
-        Destroy(aiter, &aiter_pool);
-        arc_iterator[state] = nullptr;
-        state_status[state] |= kArcIterDone;
-      }
-      // Dequeues state and marks black if done.
-      if (state_status[state] & kArcIterDone) {
-        queue->Dequeue();
-        visitor->FinishState(state);
-        state_status[state] = kBlackState;
-        continue;
-      }
-      const auto &arc = aiter->Value();
-      if (arc.nextstate >= state_status.size()) {
-        nstates = arc.nextstate + 1;
-        state_status.resize(nstates, kWhiteState);
-        arc_iterator.resize(nstates);
-      }
-      // Visits respective arc types.
-      if (filter(arc)) {
-        // Enqueues destination state and marks grey if white.
-        if (state_status[arc.nextstate] == kWhiteState) {
-          visit = visitor->WhiteArc(state, arc);
-          if (!visit) continue;
-          visit = visitor->InitState(arc.nextstate, root);
-          state_status[arc.nextstate] = kGreyState;
-          queue->Enqueue(arc.nextstate);
-        } else if (state_status[arc.nextstate] == kBlackState) {
-          visit = visitor->BlackArc(state, arc);
-        } else {
-          visit = visitor->GreyArc(state, arc);
-        }
-      }
-      aiter->Next();
-      // Destroys an iterator ASAP for efficiency.
-      if (aiter->Done()) {
-        Destroy(aiter, &aiter_pool);
-        arc_iterator[state] = nullptr;
-        state_status[state] |= kArcIterDone;
-      }
-    }
-    if (access_only) break;
-    // Finds next tree root.
-    for (root = (root == start) ? 0 : root + 1;
-         root < nstates && state_status[root] != kWhiteState; ++root) {
-    }
-    // Check for a state beyond the largest known state.
-    if (!expanded && root == nstates) {
-      for (; !siter.Done(); siter.Next()) {
-        if (siter.Value() == nstates) {
-          ++nstates;
-          state_status.push_back(kWhiteState);
-          arc_iterator.push_back(nullptr);
-          break;
-        }
-      }
-    }
-  }
-  visitor->FinishVisit();
-}
-
-template <class Arc, class Visitor, class Queue>
-inline void Visit(const Fst<Arc> &fst, Visitor *visitor, Queue *queue) {
-  Visit(fst, visitor, queue, AnyArcFilter<Arc>());
-}
-
-// Copies input FST to mutable FST following queue order.
-template <class A>
-class CopyVisitor {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  explicit CopyVisitor(MutableFst<Arc> *ofst) : ifst_(nullptr), ofst_(ofst) {}
-
-  void InitVisit(const Fst<A> &ifst) {
-    ifst_ = &ifst;
-    ofst_->DeleteStates();
-    ofst_->SetStart(ifst_->Start());
-  }
-
-  bool InitState(StateId state, StateId) {
-    while (ofst_->NumStates() <= state) ofst_->AddState();
-    return true;
-  }
-
-  bool WhiteArc(StateId state, const Arc &arc) {
-    ofst_->AddArc(state, arc);
-    return true;
-  }
-
-  bool GreyArc(StateId state, const Arc &arc) {
-    ofst_->AddArc(state, arc);
-    return true;
-  }
-
-  bool BlackArc(StateId state, const Arc &arc) {
-    ofst_->AddArc(state, arc);
-    return true;
-  }
-
-  void FinishState(StateId state) {
-    ofst_->SetFinal(state, ifst_->Final(state));
-  }
-
-  void FinishVisit() {}
-
- private:
-  const Fst<Arc> *ifst_;
-  MutableFst<Arc> *ofst_;
-};
-
-// Visits input FST up to a state limit following queue order.
-template <class A>
-class PartialVisitor {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  explicit PartialVisitor(StateId maxvisit)
-      : fst_(nullptr), maxvisit_(maxvisit) {}
-
-  void InitVisit(const Fst<A> &ifst) {
-    fst_ = &ifst;
-    ninit_ = 0;
-    nfinish_ = 0;
-  }
-
-  bool InitState(StateId state, StateId root) {
-    ++ninit_;
-    return ninit_ <= maxvisit_;
-  }
-
-  bool WhiteArc(StateId state, const Arc &arc) { return true; }
-
-  bool GreyArc(StateId state, const Arc &arc) { return true; }
-
-  bool BlackArc(StateId state, const Arc &arc) { return true; }
-
-  void FinishState(StateId state) {
-    fst_->Final(state);  // Visits super-final arc.
-    ++nfinish_;
-  }
-
-  void FinishVisit() {}
-
-  StateId NumInitialized() { return ninit_; }
-
-  StateId NumFinished() { return nfinish_; }
-
- private:
-  const Fst<Arc> *fst_;
-  StateId maxvisit_;
-  StateId ninit_;
-  StateId nfinish_;
-};
-
-// Copies input FST to mutable FST up to a state limit following queue order.
-template <class A>
-class PartialCopyVisitor : public CopyVisitor<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  using CopyVisitor<A>::WhiteArc;
-
-  PartialCopyVisitor(MutableFst<Arc> *ofst, StateId maxvisit,
-                     bool copy_grey = true, bool copy_black = true)
-      : CopyVisitor<A>(ofst), maxvisit_(maxvisit),
-        copy_grey_(copy_grey), copy_black_(copy_black) {}
-
-  void InitVisit(const Fst<A> &ifst) {
-    CopyVisitor<A>::InitVisit(ifst);
-    ninit_ = 0;
-    nfinish_ = 0;
-  }
-
-  bool InitState(StateId state, StateId root) {
-    CopyVisitor<A>::InitState(state, root);
-    ++ninit_;
-    return ninit_ <= maxvisit_;
-  }
-
-  bool GreyArc(StateId state, const Arc &arc) {
-    if (copy_grey_) return CopyVisitor<A>::GreyArc(state, arc);
-    return true;
-  }
-
-  bool BlackArc(StateId state, const Arc &arc) {
-    if (copy_black_) return CopyVisitor<A>::BlackArc(state, arc);
-    return true;
-  }
-
-  void FinishState(StateId state) {
-    CopyVisitor<A>::FinishState(state);
-    ++nfinish_;
-  }
-
-  void FinishVisit() {}
-
-  StateId NumInitialized() { return ninit_; }
-
-  StateId NumFinished() { return nfinish_; }
-
- private:
-  StateId maxvisit_;
-  StateId ninit_;
-  StateId nfinish_;
-  const bool copy_grey_;
-  const bool copy_black_;
-};
-
-}  // namespace fst
-
-#endif  // FST_VISIT_H_
diff --git a/projects/llm_framework/include/fst/weight.h b/projects/llm_framework/include/fst/weight.h
deleted file mode 100644
index ea012ec6..00000000
--- a/projects/llm_framework/include/fst/weight.h
+++ /dev/null
@@ -1,389 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// General weight set and associated semiring operation definitions.
-
-#ifndef FST_WEIGHT_H_
-#define FST_WEIGHT_H_
-
-#include <cctype>
-#include <cmath>
-#include <iostream>
-#include <sstream>
-#include <type_traits>
-#include <utility>
-
-#include <fst/compat.h>
-#include <fst/log.h>
-
-#include <fst/util.h>
-
-
-DECLARE_string(fst_weight_parentheses);
-DECLARE_string(fst_weight_separator);
-
-namespace fst {
-
-// A semiring is specified by two binary operations Plus and Times and two
-// designated elements Zero and One with the following properties:
-//
-//   Plus: associative, commutative, and has Zero as its identity.
-//
-//   Times: associative and has identity One, distributes w.r.t. Plus, and
-//     has Zero as an annihilator:
-//          Times(Zero(), a) == Times(a, Zero()) = Zero().
-//
-// A left semiring distributes on the left; a right semiring is similarly
-// defined.
-//
-// A Weight class must have binary functions Plus and Times and static member
-// functions Zero() and One() and these must form (at least) a left or right
-// semiring.
-//
-// In addition, the following should be defined for a Weight:
-//
-//   Member: predicate on set membership.
-//
-//   NoWeight: static member function that returns an element that is
-//      not a set member; used to signal an error.
-//
-//   >>: reads textual representation of a weight.
-//
-//   <<: prints textual representation of a weight.
-//
-//   Read(istream &istrm): reads binary representation of a weight.
-//
-//   Write(ostream &ostrm): writes binary representation of a weight.
-//
-//   Hash: maps weight to size_t.
-//
-//   ApproxEqual: approximate equality (for inexact weights)
-//
-//   Quantize: quantizes w.r.t delta (for inexact weights)
-//
-//   Divide:
-//     - In a left semiring, for all a, b, b', c:
-//       if Times(a, b) = c, Divide(c, a, DIVIDE_LEFT) = b' and b'.Member(),
-//       then Times(a, b') = c.
-//     - In a right semiring, for all a, a', b, c:
-//       if Times(a, b) = c, Divide(c, b, DIVIDE_RIGHT) = a' and a'.Member(),
-//       then Times(a', b) = c.
-//     - In a commutative semiring,
-//        * for all a, c:
-//          Divide(c, a, DIVIDE_ANY) = Divide(c, a, DIVIDE_LEFT)
-//           = Divide(c, a, DIVIDE_RIGHT)
-//        * for all a, b, b', c:
-//          if Times(a, b), Divide(c, a, DIVIDE_ANY) = b' and b'.Member(),
-//          then Times(a, b') = c
-//     - In the case where there exist no b such that c = Times(a, b), the
-//       return value of Divide(c, a, DIVIDE_LEFT) is unspecified. Returning
-//       Weight::NoWeight() is recommemded but not required in order to
-//       allow the most efficient implementation.
-//     - All algorithms in this library only call Divide(c, a) when it is
-//       guaranteed that there exists a b such that c = Times(a, b).
-//
-//   ReverseWeight: the type of the corresponding reverse weight.
-//
-//     Typically the same type as Weight for a (both left and right) semiring.
-//     For the left string semiring, it is the right string semiring.
-//
-//   Reverse: a mapping from Weight to ReverseWeight s.t.
-//
-//     --> Reverse(Reverse(a)) = a
-//     --> Reverse(Plus(a, b)) = Plus(Reverse(a), Reverse(b))
-//     --> Reverse(Times(a, b)) = Times(Reverse(b), Reverse(a))
-//     Typically the identity mapping in a (both left and right) semiring.
-//     In the left string semiring, it maps to the reverse string in the right
-//     string semiring.
-//
-//   Properties: specifies additional properties that hold:
-//      LeftSemiring: indicates weights form a left semiring.
-//      RightSemiring: indicates weights form a right semiring.
-//      Commutative: for all a,b: Times(a,b) == Times(b,a)
-//      Idempotent: for all a: Plus(a, a) == a.
-//      Path: for all a, b: Plus(a, b) == a or Plus(a, b) == b.
-
-// CONSTANT DEFINITIONS
-
-// A representable float near .001.
-constexpr float kDelta = 1.0F / 1024.0F;
-
-// For all a, b, c: Times(c, Plus(a, b)) = Plus(Times(c, a), Times(c, b)).
-constexpr uint64 kLeftSemiring = 0x0000000000000001ULL;
-
-// For all a, b, c: Times(Plus(a, b), c) = Plus(Times(a, c), Times(b, c)).
-constexpr uint64 kRightSemiring = 0x0000000000000002ULL;
-
-constexpr uint64 kSemiring = kLeftSemiring | kRightSemiring;
-
-// For all a, b: Times(a, b) = Times(b, a).
-constexpr uint64 kCommutative = 0x0000000000000004ULL;
-
-// For all a: Plus(a, a) = a.
-constexpr uint64 kIdempotent = 0x0000000000000008ULL;
-
-// For all a, b: Plus(a, b) = a or Plus(a, b) = b.
-constexpr uint64 kPath = 0x0000000000000010ULL;
-
-// For random weight generation: default number of distinct weights.
-// This is also used for a few other weight generation defaults.
-constexpr size_t kNumRandomWeights = 5;
-
-// Weight property boolean constants needed for SFINAE.
-
-// MSVC compiler bug workaround: an expression containing W::Properties() cannot
-// be directly used as a value argument to std::enable_if or integral_constant.
-// WeightPropertiesThunk<W>::Properties works instead, however.
-namespace bug {
-template <class W>
-struct WeightPropertiesThunk {
-  WeightPropertiesThunk() = delete;
-  constexpr static const uint64 Properties = W::Properties();
-};
-
-template <class W, uint64 props>
-using TestWeightProperties = std::integral_constant<bool,
-        (WeightPropertiesThunk<W>::Properties & props) == props>;
-}  // namespace bug
-
-template <class W>
-using IsIdempotent = bug::TestWeightProperties<W, kIdempotent>;
-
-template <class W>
-using IsPath = bug::TestWeightProperties<W, kPath>;
-
-
-// Determines direction of division.
-enum DivideType {
-  DIVIDE_LEFT,   // left division
-  DIVIDE_RIGHT,  // right division
-  DIVIDE_ANY
-};  // division in a commutative semiring
-
-// NATURAL ORDER
-//
-// By definition:
-//
-//                 a <= b iff a + b = a
-//
-// The natural order is a negative partial order iff the semiring is
-// idempotent. It is trivially monotonic for plus. It is left
-// (resp. right) monotonic for times iff the semiring is left
-// (resp. right) distributive. It is a total order iff the semiring
-// has the path property.
-//
-// For more information, see:
-//
-// Mohri, M. 2002. Semiring framework and algorithms for shortest-distance
-// problems, Journal of Automata, Languages and
-// Combinatorics 7(3): 321-350, 2002.
-//
-// We define the strict version of this order below.
-
-template <class W>
-class NaturalLess {
-public:
-  using Weight = W;
-
-  NaturalLess() {
-    if (!(W::Properties() & kIdempotent)) {
-      FSTERROR() << "NaturalLess: Weight type is not idempotent: " << W::Type();
-    }
-  }
-
-  bool operator()(const W &w1, const W &w2) const {
-    return (Plus(w1, w2) == w1) && w1 != w2;
-  }
-};
-
-// Power is the iterated product for arbitrary semirings such that Power(w, 0)
-// is One() for the semiring, and Power(w, n) = Times(Power(w, n - 1), w).
-template <class Weight>
-Weight Power(const Weight &weight, size_t n) {
-  auto result = Weight::One();
-  for (size_t i = 0; i < n; ++i) result = Times(result, weight);
-  return result;
-}
-
-// Simple default adder class. Specializations might be more complex.
-template <class Weight>
-class Adder {
- public:
-  explicit Adder(Weight w = Weight::Zero()) : sum_(w) { }
-
-  Weight Add(const Weight &w) {
-    sum_ = Plus(sum_, w);
-    return sum_;
-  }
-
-  Weight Sum() { return sum_; }
-
-  void Reset(Weight w = Weight::Zero()) { sum_ = w; }
-
- private:
-  Weight sum_;
-};
-
-// General weight converter: raises error.
-template <class W1, class W2>
-struct WeightConvert {
-  W2 operator()(W1 w1) const {
-    FSTERROR() << "WeightConvert: Can't convert weight from \"" << W1::Type()
-               << "\" to \"" << W2::Type();
-    return W2::NoWeight();
-  }
-};
-
-// Specialized weight converter to self.
-template <class W>
-struct WeightConvert<W, W> {
-  W operator()(W weight) const { return weight; }
-};
-
-// General random weight generator: raises error.
-template <class W>
-struct WeightGenerate {
-  W operator()() const {
-    FSTERROR() << "WeightGenerate: No random generator for " << W::Type();
-    return W::NoWeight();
-  }
-};
-
-namespace internal {
-
-class CompositeWeightIO {
- public:
-  CompositeWeightIO();
-  CompositeWeightIO(char separator, std::pair<char, char> parentheses);
-
-  std::pair<char, char> parentheses() const {
-    return {open_paren_, close_paren_};
-  }
-  char separator() const { return separator_; }
-
-  bool error() const { return error_; }
-
- protected:
-  const char separator_;
-  const char open_paren_;
-  const char close_paren_;
-
- private:
-  bool error_;
-};
-
-}  // namespace internal
-
-// Helper class for writing textual composite weights.
-class CompositeWeightWriter : public internal::CompositeWeightIO {
- public:
-  // Uses configuration from flags (FLAGS_fst_weight_separator,
-  // FLAGS_fst_weight_parentheses).
-  explicit CompositeWeightWriter(std::ostream &ostrm);
-
-  // parentheses defines the opening and closing parenthesis characters.
-  // Set parentheses = {0, 0} to disable writing parenthesis.
-  CompositeWeightWriter(std::ostream &ostrm, char separator,
-                        std::pair<char, char> parentheses);
-
-  CompositeWeightWriter(const CompositeWeightWriter &) = delete;
-  CompositeWeightWriter &operator=(const CompositeWeightWriter &) = delete;
-
-  // Writes open parenthesis to a stream if option selected.
-  void WriteBegin();
-
-  // Writes element to a stream.
-  template <class T>
-  void WriteElement(const T &comp) {
-    if (i_++ > 0) ostrm_ << separator_;
-    ostrm_ << comp;
-  }
-
-  // Writes close parenthesis to a stream if option selected.
-  void WriteEnd();
-
- private:
-  std::ostream &ostrm_;
-  int i_ = 0;  // Element position.
-};
-
-// Helper class for reading textual composite weights. Elements are separated by
-// a separator character. There must be at least one element per textual
-// representation.  Parentheses characters should be set if the composite
-// weights themselves contain composite weights to ensure proper parsing.
-class CompositeWeightReader : public internal::CompositeWeightIO {
- public:
-  // Uses configuration from flags (FLAGS_fst_weight_separator,
-  // FLAGS_fst_weight_parentheses).
-  explicit CompositeWeightReader(std::istream &istrm);
-
-  // parentheses defines the opening and closing parenthesis characters.
-  // Set parentheses = {0, 0} to disable reading parenthesis.
-  CompositeWeightReader(std::istream &istrm, char separator,
-                        std::pair<char, char> parentheses);
-
-  CompositeWeightReader(const CompositeWeightReader &) = delete;
-  CompositeWeightReader &operator=(const CompositeWeightReader &) = delete;
-
-  // Reads open parenthesis from a stream if option selected.
-  void ReadBegin();
-
-  // Reads element from a stream. The second argument, when true, indicates that
-  // this will be the last element (allowing more forgiving formatting of the
-  // last element). Returns false when last element is read.
-  template <class T>
-  bool ReadElement(T *comp, bool last = false);
-
-  // Finalizes reading.
-  void ReadEnd();
-
- private:
-  std::istream &istrm_;  // Input stream.
-  int c_ = 0;            // Last character read, or EOF.
-  int depth_ = 0;        // Weight parentheses depth.
-};
-
-template <class T>
-inline bool CompositeWeightReader::ReadElement(T *comp, bool last) {
-  string s;
-  const bool has_parens = open_paren_ != 0;
-  while ((c_ != std::istream::traits_type::eof()) && !std::isspace(c_) &&
-         (c_ != separator_ || depth_ > 1 || last) &&
-         (c_ != close_paren_ || depth_ != 1)) {
-    s += c_;
-    // If parentheses encountered before separator, they must be matched.
-    if (has_parens && c_ == open_paren_) {
-      ++depth_;
-    } else if (has_parens && c_ == close_paren_) {
-      // Failure on unmatched parentheses.
-      if (depth_ == 0) {
-        FSTERROR() << "CompositeWeightReader: Unmatched close paren: "
-                   << "Is the fst_weight_parentheses flag set correctly?";
-        istrm_.clear(std::ios::badbit);
-        return false;
-      }
-      --depth_;
-    }
-    c_ = istrm_.get();
-  }
-  if (s.empty()) {
-    FSTERROR() << "CompositeWeightReader: Empty element: "
-               << "Is the fst_weight_parentheses flag set correctly?";
-    istrm_.clear(std::ios::badbit);
-    return false;
-  }
-  std::istringstream istrm(s);
-  istrm >> *comp;
-  // Skips separator/close parenthesis.
-  if (c_ != std::istream::traits_type::eof() && !std::isspace(c_)) {
-    c_ = istrm_.get();
-  }
-  const bool is_eof = c_ == std::istream::traits_type::eof();
-  // Clears fail bit if just EOF.
-  if (is_eof && !istrm_.bad()) istrm_.clear(std::ios::eofbit);
-  return !is_eof && !std::isspace(c_);
-}
-
-}  // namespace fst
-
-#endif  // FST_WEIGHT_H_
diff --git a/projects/llm_framework/include/gelf.h b/projects/llm_framework/include/gelf.h
deleted file mode 100644
index 7a3c87aa..00000000
--- a/projects/llm_framework/include/gelf.h
+++ /dev/null
@@ -1,342 +0,0 @@
-/* This file defines generic ELF types, structures, and macros.
-   Copyright (C) 1999, 2000, 2001, 2002, 2004, 2005, 2007, 2015 Red Hat, Inc.
-   This file is part of elfutils.
-
-   This file is free software; you can redistribute it and/or modify
-   it under the terms of either
-
-     * the GNU Lesser General Public License as published by the Free
-       Software Foundation; either version 3 of the License, or (at
-       your option) any later version
-
-   or
-
-     * the GNU General Public License as published by the Free
-       Software Foundation; either version 2 of the License, or (at
-       your option) any later version
-
-   or both in parallel, as here.
-
-   elfutils is distributed in the hope that it will be useful, but
-   WITHOUT ANY WARRANTY; without even the implied warranty of
-   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
-   General Public License for more details.
-
-   You should have received copies of the GNU General Public License and
-   the GNU Lesser General Public License along with this program.  If
-   not, see <http://www.gnu.org/licenses/>.  */
-
-#ifndef _GELF_H
-#define	_GELF_H 1
-
-#include <libelf.h>
-
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* Class independent type definitions.  Correctly speaking this is not
-   true.  We assume that 64-bit binaries are the largest class and
-   therefore all other classes can be represented without loss.  */
-
-/* Type for a 16-bit quantity.  */
-typedef Elf64_Half GElf_Half;
-
-/* Types for signed and unsigned 32-bit quantities.  */
-typedef Elf64_Word GElf_Word;
-typedef	Elf64_Sword GElf_Sword;
-
-/* Types for signed and unsigned 64-bit quantities.  */
-typedef Elf64_Xword GElf_Xword;
-typedef	Elf64_Sxword GElf_Sxword;
-
-/* Type of addresses.  */
-typedef Elf64_Addr GElf_Addr;
-
-/* Type of file offsets.  */
-typedef Elf64_Off GElf_Off;
-
-
-/* The ELF file header.  This appears at the start of every ELF file.  */
-typedef Elf64_Ehdr GElf_Ehdr;
-
-/* Section header.  */
-typedef Elf64_Shdr GElf_Shdr;
-
-/* Section index.  */
-/* XXX This should probably be a larger type in preparation of times when
-   regular section indices can be larger.  */
-typedef Elf64_Section GElf_Section;
-
-/* Symbol table entry.  */
-typedef Elf64_Sym GElf_Sym;
-
-/* The syminfo section if available contains additional information about
-   every dynamic symbol.  */
-typedef Elf64_Syminfo GElf_Syminfo;
-
-/* Relocation table entry without addend (in section of type SHT_REL).  */
-typedef Elf64_Rel GElf_Rel;
-
-/* Relocation table entry with addend (in section of type SHT_RELA).  */
-typedef Elf64_Rela GElf_Rela;
-
-/* Program segment header.  */
-typedef Elf64_Phdr GElf_Phdr;
-
-/* Header of a compressed section.  */
-typedef Elf64_Chdr GElf_Chdr;
-
-/* Dynamic section entry.  */
-typedef Elf64_Dyn GElf_Dyn;
-
-
-/* Version definition sections.  */
-typedef Elf64_Verdef GElf_Verdef;
-
-/* Auxiliary version information.  */
-typedef Elf64_Verdaux GElf_Verdaux;
-
-/* Version dependency section.  */
-typedef Elf64_Verneed GElf_Verneed;
-
-/* Auxiliary needed version information.  */
-typedef Elf64_Vernaux GElf_Vernaux;
-
-
-/* Type for version symbol information.  */
-typedef Elf64_Versym GElf_Versym;
-
-
-/* Auxiliary vector.  */
-typedef Elf64_auxv_t GElf_auxv_t;
-
-
-/* Note section contents.  */
-typedef Elf64_Nhdr GElf_Nhdr;
-
-
-/* Move structure.  */
-typedef Elf64_Move GElf_Move;
-
-
-/* Library list structure.  */
-typedef Elf64_Lib GElf_Lib;
-
-
-/* How to extract and insert information held in the st_info field.  */
-
-#define GELF_ST_BIND(val)		ELF64_ST_BIND (val)
-#define GELF_ST_TYPE(val)		ELF64_ST_TYPE (val)
-#define GELF_ST_INFO(bind, type)	ELF64_ST_INFO (bind, type)
-
-/* How to extract information held in the st_other field.  */
-
-#define GELF_ST_VISIBILITY(val)		ELF64_ST_VISIBILITY (val)
-
-
-/* How to extract and insert information held in the r_info field.  */
-
-#define GELF_R_SYM(info)		ELF64_R_SYM (info)
-#define GELF_R_TYPE(info)		ELF64_R_TYPE (info)
-#define GELF_R_INFO(sym, type)		ELF64_R_INFO (sym, type)
-
-
-/* How to extract and insert information held in the m_info field.  */
-#define GELF_M_SYM(info)		ELF64_M_SYM (info)
-#define GELF_M_SIZE(info)		ELF64_M_SIZE (info)
-#define GELF_M_INFO(sym, size)		ELF64_M_INFO (sym, size)
-
-
-/* Get class of the file associated with ELF.  */
-extern int gelf_getclass (Elf *__elf);
-
-
-/* Return size of array of COUNT elements of the type denoted by TYPE
-   in the external representation.  The binary class is taken from ELF.
-   The result is based on version VERSION of the ELF standard.  */
-extern size_t gelf_fsize (Elf *__elf, Elf_Type __type, size_t __count,
-			  unsigned int __version);
-
-/* Retrieve object file header.  */
-extern GElf_Ehdr *gelf_getehdr (Elf *__elf, GElf_Ehdr *__dest);
-
-/* Update the ELF header.  */
-extern int gelf_update_ehdr (Elf *__elf, GElf_Ehdr *__src);
-
-/* Create new ELF header if none exists.  Creates an Elf32_Ehdr if CLASS
-   is ELFCLASS32 or an Elf64_Ehdr if CLASS is ELFCLASS64.  Returns NULL
-   on error.  */
-extern void *gelf_newehdr (Elf *__elf, int __class);
-
-/* Get section at OFFSET.  */
-extern Elf_Scn *gelf_offscn (Elf *__elf, GElf_Off __offset);
-
-/* Retrieve section header.  */
-extern GElf_Shdr *gelf_getshdr (Elf_Scn *__scn, GElf_Shdr *__dst);
-
-/* Update section header.  */
-extern int gelf_update_shdr (Elf_Scn *__scn, GElf_Shdr *__src);
-
-/* Retrieve program header table entry.  */
-extern GElf_Phdr *gelf_getphdr (Elf *__elf, int __ndx, GElf_Phdr *__dst);
-
-/* Update the program header.  */
-extern int gelf_update_phdr (Elf *__elf, int __ndx, GElf_Phdr *__src);
-
-/* Create new program header with PHNUM entries.  Creates either an
-   Elf32_Phdr or an Elf64_Phdr depending on whether the given ELF is
-   ELFCLASS32 or ELFCLASS64.  Returns NULL on error.  */
-extern void *gelf_newphdr (Elf *__elf, size_t __phnum);
-
-/* Get compression header of section if any.  Returns NULL and sets
-   elf_errno if the section isn't compressed or an error occurred.  */
-extern GElf_Chdr *gelf_getchdr (Elf_Scn *__scn, GElf_Chdr *__dst);
-
-/* Convert data structure from the representation in the file represented
-   by ELF to their memory representation.  */
-extern Elf_Data *gelf_xlatetom (Elf *__elf, Elf_Data *__dest,
-				const Elf_Data *__src, unsigned int __encode);
-
-/* Convert data structure from to the representation in memory
-   represented by ELF file representation.  */
-extern Elf_Data *gelf_xlatetof (Elf *__elf, Elf_Data *__dest,
-				const Elf_Data *__src, unsigned int __encode);
-
-
-/* Retrieve REL relocation info at the given index.  */
-extern GElf_Rel *gelf_getrel (Elf_Data *__data, int __ndx, GElf_Rel *__dst);
-
-/* Retrieve RELA relocation info at the given index.  */
-extern GElf_Rela *gelf_getrela (Elf_Data *__data, int __ndx, GElf_Rela *__dst);
-
-/* Update REL relocation information at given index.  */
-extern int gelf_update_rel (Elf_Data *__dst, int __ndx, GElf_Rel *__src);
-
-/* Update RELA relocation information at given index.  */
-extern int gelf_update_rela (Elf_Data *__dst, int __ndx, GElf_Rela *__src);
-
-
-/* Retrieve symbol information from the symbol table at the given index.  */
-extern GElf_Sym *gelf_getsym (Elf_Data *__data, int __ndx, GElf_Sym *__dst);
-
-/* Update symbol information in the symbol table at the given index.  */
-extern int gelf_update_sym (Elf_Data *__data, int __ndx, GElf_Sym *__src);
-
-
-/* Retrieve symbol information and separate section index from the
-   symbol table at the given index.  */
-extern GElf_Sym *gelf_getsymshndx (Elf_Data *__symdata, Elf_Data *__shndxdata,
-				   int __ndx, GElf_Sym *__sym,
-				   Elf32_Word *__xshndx);
-
-/* Update symbol information and separate section index in the symbol
-   table at the given index.  */
-extern int gelf_update_symshndx (Elf_Data *__symdata, Elf_Data *__shndxdata,
-				 int __ndx, GElf_Sym *__sym,
-				 Elf32_Word __xshndx);
-
-
-/* Retrieve additional symbol information from the symbol table at the
-   given index.  */
-extern GElf_Syminfo *gelf_getsyminfo (Elf_Data *__data, int __ndx,
-				      GElf_Syminfo *__dst);
-
-/* Update additional symbol information in the symbol table at the
-   given index.  */
-extern int gelf_update_syminfo (Elf_Data *__data, int __ndx,
-				GElf_Syminfo *__src);
-
-
-/* Get information from dynamic table at the given index.  */
-extern GElf_Dyn *gelf_getdyn (Elf_Data *__data, int __ndx, GElf_Dyn *__dst);
-
-/* Update information in dynamic table at the given index.  */
-extern int gelf_update_dyn (Elf_Data *__dst, int __ndx, GElf_Dyn *__src);
-
-
-/* Get move structure at the given index.  */
-extern GElf_Move *gelf_getmove (Elf_Data *__data, int __ndx, GElf_Move *__dst);
-
-/* Update move structure at the given index.  */
-extern int gelf_update_move (Elf_Data *__data, int __ndx,
-			     GElf_Move *__src);
-
-
-/* Get library from table at the given index.  */
-extern GElf_Lib *gelf_getlib (Elf_Data *__data, int __ndx, GElf_Lib *__dst);
-
-/* Update library in table at the given index.  */
-extern int gelf_update_lib (Elf_Data *__data, int __ndx, GElf_Lib *__src);
-
-
-
-/* Retrieve symbol version information at given index.  */
-extern GElf_Versym *gelf_getversym (Elf_Data *__data, int __ndx,
-				    GElf_Versym *__dst);
-
-/* Update symbol version information.  */
-extern int gelf_update_versym (Elf_Data *__data, int __ndx,
-			       GElf_Versym *__src);
-
-
-/* Retrieve required symbol version information at given offset.  */
-extern GElf_Verneed *gelf_getverneed (Elf_Data *__data, int __offset,
-				      GElf_Verneed *__dst);
-
-/* Update required symbol version information.  */
-extern int gelf_update_verneed (Elf_Data *__data, int __offset,
-				GElf_Verneed *__src);
-
-/* Retrieve additional required symbol version information at given offset.  */
-extern GElf_Vernaux *gelf_getvernaux (Elf_Data *__data, int __offset,
-				      GElf_Vernaux *__dst);
-
-/* Update additional required symbol version information.  */
-extern int gelf_update_vernaux (Elf_Data *__data, int __offset,
-				GElf_Vernaux *__src);
-
-
-/* Retrieve symbol version definition information at given offset.  */
-extern GElf_Verdef *gelf_getverdef (Elf_Data *__data, int __offset,
-				    GElf_Verdef *__dst);
-
-/* Update symbol version definition information.  */
-extern int gelf_update_verdef (Elf_Data *__data, int __offset,
-			       GElf_Verdef *__src);
-
-/* Retrieve additional symbol version definition information at given
-   offset.  */
-extern GElf_Verdaux *gelf_getverdaux (Elf_Data *__data, int __offset,
-				      GElf_Verdaux *__dst);
-
-/* Update additional symbol version definition information.  */
-extern int gelf_update_verdaux (Elf_Data *__data, int __offset,
-				GElf_Verdaux *__src);
-
-
-/* Get auxv entry at the given index.  */
-extern GElf_auxv_t *gelf_getauxv (Elf_Data *__data, int __ndx,
-				  GElf_auxv_t *__dst);
-
-/* Update auxv entry at the given index.  */
-extern int gelf_update_auxv (Elf_Data *__data, int __ndx, GElf_auxv_t *__src);
-
-
-/* Get note header at the given offset into the data, and the offsets of
-   the note's name and descriptor data.  Returns the offset of the next
-   note header, or 0 for an invalid offset or corrupt note header.  */
-extern size_t gelf_getnote (Elf_Data *__data, size_t __offset,
-			    GElf_Nhdr *__result,
-			    size_t *__name_offset, size_t *__desc_offset);
-
-
-/* Compute simple checksum from permanent parts of the ELF file.  */
-extern long int gelf_checksum (Elf *__elf);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif	/* gelf.h */
diff --git a/projects/llm_framework/include/gflags/defines.h b/projects/llm_framework/include/gflags/defines.h
deleted file mode 100644
index 944ed7db..00000000
--- a/projects/llm_framework/include/gflags/defines.h
+++ /dev/null
@@ -1,48 +0,0 @@
-/* Generated from defines.h.in during build configuration using CMake. */
-
-// Note: This header file is only used internally. It is not part of public interface!
-//       Any cmakedefine is defined using the -D flag instead when Bazel is used.
-//       For Bazel, this file is thus not used to avoid a private file in $(GENDIR).
-
-#ifndef GFLAGS_DEFINES_H_
-#define GFLAGS_DEFINES_H_
-
-
-// Define if you build this library for a MS Windows OS.
-/* #undef OS_WINDOWS */
-
-// Define if you have the <stdint.h> header file.
-#define HAVE_STDINT_H
-
-// Define if you have the <sys/types.h> header file.
-#define HAVE_SYS_TYPES_H
-
-// Define if you have the <inttypes.h> header file.
-#define HAVE_INTTYPES_H
-
-// Define if you have the <sys/stat.h> header file.
-#define HAVE_SYS_STAT_H
-
-// Define if you have the <unistd.h> header file.
-#define HAVE_UNISTD_H
-
-// Define if you have the <fnmatch.h> header file.
-#define HAVE_FNMATCH_H
-
-// Define if you have the <shlwapi.h> header file (Windows 2000/XP).
-/* #undef HAVE_SHLWAPI_H */
-
-// Define if you have the strtoll function.
-#define HAVE_STRTOLL
-
-// Define if you have the strtoq function.
-/* #undef HAVE_STRTOQ */
-
-// Define if you have the <pthread.h> header file.
-/* #undef HAVE_PTHREAD */
-
-// Define if your pthread library defines the type pthread_rwlock_t
-/* #undef HAVE_RWLOCK */
-
-
-#endif // GFLAGS_DEFINES_H_
diff --git a/projects/llm_framework/include/gflags/gflags.h b/projects/llm_framework/include/gflags/gflags.h
deleted file mode 100644
index 9273da8d..00000000
--- a/projects/llm_framework/include/gflags/gflags.h
+++ /dev/null
@@ -1,626 +0,0 @@
-// Copyright (c) 2006, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// ---
-// Revamped and reorganized by Craig Silverstein
-//
-// This is the file that should be included by any file which declares
-// or defines a command line flag or wants to parse command line flags
-// or print a program usage message (which will include information about
-// flags).  Executive summary, in the form of an example foo.cc file:
-//
-//    #include "foo.h"         // foo.h has a line "DECLARE_int32(start);"
-//    #include "validators.h"  // hypothetical file defining ValidateIsFile()
-//
-//    DEFINE_int32(end, 1000, "The last record to read");
-//
-//    DEFINE_string(filename, "my_file.txt", "The file to read");
-//    // Crash if the specified file does not exist.
-//    static bool dummy = RegisterFlagValidator(&FLAGS_filename,
-//                                              &ValidateIsFile);
-//
-//    DECLARE_bool(verbose); // some other file has a DEFINE_bool(verbose, ...)
-//
-//    void MyFunc() {
-//      if (FLAGS_verbose) printf("Records %d-%d\n", FLAGS_start, FLAGS_end);
-//    }
-//
-//    Then, at the command-line:
-//       ./foo --noverbose --start=5 --end=100
-//
-// For more details, see
-//    doc/gflags.html
-//
-// --- A note about thread-safety:
-//
-// We describe many functions in this routine as being thread-hostile,
-// thread-compatible, or thread-safe.  Here are the meanings we use:
-//
-// thread-safe: it is safe for multiple threads to call this routine
-//   (or, when referring to a class, methods of this class)
-//   concurrently.
-// thread-hostile: it is not safe for multiple threads to call this
-//   routine (or methods of this class) concurrently.  In gflags,
-//   most thread-hostile routines are intended to be called early in,
-//   or even before, main() -- that is, before threads are spawned.
-// thread-compatible: it is safe for multiple threads to read from
-//   this variable (when applied to variables), or to call const
-//   methods of this class (when applied to classes), as long as no
-//   other thread is writing to the variable or calling non-const
-//   methods of this class.
-
-#ifndef GFLAGS_GFLAGS_H_
-#define GFLAGS_GFLAGS_H_
-
-#include <string>
-#include <vector>
-
-#include "gflags/gflags_declare.h" // IWYU pragma: export
-
-
-// We always want to export variables defined in user code
-#ifndef GFLAGS_DLL_DEFINE_FLAG
-#  if GFLAGS_IS_A_DLL && defined(_MSC_VER)
-#    define GFLAGS_DLL_DEFINE_FLAG __declspec(dllexport)
-#  else
-#    define GFLAGS_DLL_DEFINE_FLAG
-#  endif
-#endif
-
-
-namespace GFLAGS_NAMESPACE {
-
-
-// --------------------------------------------------------------------
-// To actually define a flag in a file, use DEFINE_bool,
-// DEFINE_string, etc. at the bottom of this file.  You may also find
-// it useful to register a validator with the flag.  This ensures that
-// when the flag is parsed from the commandline, or is later set via
-// SetCommandLineOption, we call the validation function. It is _not_
-// called when you assign the value to the flag directly using the = operator.
-//
-// The validation function should return true if the flag value is valid, and
-// false otherwise. If the function returns false for the new setting of the
-// flag, the flag will retain its current value. If it returns false for the
-// default value, ParseCommandLineFlags() will die.
-//
-// This function is safe to call at global construct time (as in the
-// example below).
-//
-// Example use:
-//    static bool ValidatePort(const char* flagname, int32 value) {
-//       if (value > 0 && value < 32768)   // value is ok
-//         return true;
-//       printf("Invalid value for --%s: %d\n", flagname, (int)value);
-//       return false;
-//    }
-//    DEFINE_int32(port, 0, "What port to listen on");
-//    static bool dummy = RegisterFlagValidator(&FLAGS_port, &ValidatePort);
-
-// Returns true if successfully registered, false if not (because the
-// first argument doesn't point to a command-line flag, or because a
-// validator is already registered for this flag).
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const bool*        flag, bool (*validate_fn)(const char*, bool));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const int32*       flag, bool (*validate_fn)(const char*, int32));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const uint32*      flag, bool (*validate_fn)(const char*, uint32));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const int64*       flag, bool (*validate_fn)(const char*, int64));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const uint64*      flag, bool (*validate_fn)(const char*, uint64));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const double*      flag, bool (*validate_fn)(const char*, double));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const std::string* flag, bool (*validate_fn)(const char*, const std::string&));
-
-// Convenience macro for the registration of a flag validator
-#define DEFINE_validator(name, validator) \
-    static const bool name##_validator_registered = \
-            GFLAGS_NAMESPACE::RegisterFlagValidator(&FLAGS_##name, validator)
-
-
-// --------------------------------------------------------------------
-// These methods are the best way to get access to info about the
-// list of commandline flags.  Note that these routines are pretty slow.
-//   GetAllFlags: mostly-complete info about the list, sorted by file.
-//   ShowUsageWithFlags: pretty-prints the list to stdout (what --help does)
-//   ShowUsageWithFlagsRestrict: limit to filenames with restrict as a substr
-//
-// In addition to accessing flags, you can also access argv[0] (the program
-// name) and argv (the entire commandline), which we sock away a copy of.
-// These variables are static, so you should only set them once.
-//
-// No need to export this data only structure from DLL, avoiding VS warning 4251.
-struct CommandLineFlagInfo {
-  std::string name;            // the name of the flag
-  std::string type;            // the type of the flag: int32, etc
-  std::string description;     // the "help text" associated with the flag
-  std::string current_value;   // the current value, as a string
-  std::string default_value;   // the default value, as a string
-  std::string filename;        // 'cleaned' version of filename holding the flag
-  bool has_validator_fn;       // true if RegisterFlagValidator called on this flag
-  bool is_default;             // true if the flag has the default value and
-                               // has not been set explicitly from the cmdline
-                               // or via SetCommandLineOption
-  const void* flag_ptr;        // pointer to the flag's current value (i.e. FLAGS_foo)
-};
-
-// Using this inside of a validator is a recipe for a deadlock.
-// TODO(user) Fix locking when validators are running, to make it safe to
-// call validators during ParseAllFlags.
-// Also make sure then to uncomment the corresponding unit test in
-// gflags_unittest.sh
-extern GFLAGS_DLL_DECL void GetAllFlags(std::vector<CommandLineFlagInfo>* OUTPUT);
-// These two are actually defined in gflags_reporting.cc.
-extern GFLAGS_DLL_DECL void ShowUsageWithFlags(const char *argv0);  // what --help does
-extern GFLAGS_DLL_DECL void ShowUsageWithFlagsRestrict(const char *argv0, const char *restrict);
-
-// Create a descriptive string for a flag.
-// Goes to some trouble to make pretty line breaks.
-extern GFLAGS_DLL_DECL std::string DescribeOneFlag(const CommandLineFlagInfo& flag);
-
-// Thread-hostile; meant to be called before any threads are spawned.
-extern GFLAGS_DLL_DECL void SetArgv(int argc, const char** argv);
-
-// The following functions are thread-safe as long as SetArgv() is
-// only called before any threads start.
-extern GFLAGS_DLL_DECL const std::vector<std::string>& GetArgvs();
-extern GFLAGS_DLL_DECL const char* GetArgv();                      // all of argv as a string
-extern GFLAGS_DLL_DECL const char* GetArgv0();                     // only argv0
-extern GFLAGS_DLL_DECL uint32 GetArgvSum();                        // simple checksum of argv
-extern GFLAGS_DLL_DECL const char* ProgramInvocationName();        // argv0, or "UNKNOWN" if not set
-extern GFLAGS_DLL_DECL const char* ProgramInvocationShortName();   // basename(argv0)
-
-// ProgramUsage() is thread-safe as long as SetUsageMessage() is only
-// called before any threads start.
-extern GFLAGS_DLL_DECL const char* ProgramUsage();                 // string set by SetUsageMessage()
-
-// VersionString() is thread-safe as long as SetVersionString() is only
-// called before any threads start.
-extern GFLAGS_DLL_DECL const char* VersionString();                // string set by SetVersionString()
-
-
-
-// --------------------------------------------------------------------
-// Normally you access commandline flags by just saying "if (FLAGS_foo)"
-// or whatever, and set them by calling "FLAGS_foo = bar" (or, more
-// commonly, via the DEFINE_foo macro).  But if you need a bit more
-// control, we have programmatic ways to get/set the flags as well.
-// These programmatic ways to access flags are thread-safe, but direct
-// access is only thread-compatible.
-
-// Return true iff the flagname was found.
-// OUTPUT is set to the flag's value, or unchanged if we return false.
-extern GFLAGS_DLL_DECL bool GetCommandLineOption(const char* name, std::string* OUTPUT);
-
-// Return true iff the flagname was found. OUTPUT is set to the flag's
-// CommandLineFlagInfo or unchanged if we return false.
-extern GFLAGS_DLL_DECL bool GetCommandLineFlagInfo(const char* name, CommandLineFlagInfo* OUTPUT);
-
-// Return the CommandLineFlagInfo of the flagname.  exit() if name not found.
-// Example usage, to check if a flag's value is currently the default value:
-//   if (GetCommandLineFlagInfoOrDie("foo").is_default) ...
-extern GFLAGS_DLL_DECL CommandLineFlagInfo GetCommandLineFlagInfoOrDie(const char* name);
-
-enum GFLAGS_DLL_DECL FlagSettingMode {
-  // update the flag's value (can call this multiple times).
-  SET_FLAGS_VALUE,
-  // update the flag's value, but *only if* it has not yet been updated
-  // with SET_FLAGS_VALUE, SET_FLAG_IF_DEFAULT, or "FLAGS_xxx = nondef".
-  SET_FLAG_IF_DEFAULT,
-  // set the flag's default value to this.  If the flag has not yet updated
-  // yet (via SET_FLAGS_VALUE, SET_FLAG_IF_DEFAULT, or "FLAGS_xxx = nondef")
-  // change the flag's current value to the new default value as well.
-  SET_FLAGS_DEFAULT
-};
-
-// Set a particular flag ("command line option").  Returns a string
-// describing the new value that the option has been set to.  The
-// return value API is not well-specified, so basically just depend on
-// it to be empty if the setting failed for some reason -- the name is
-// not a valid flag name, or the value is not a valid value -- and
-// non-empty else.
-
-// SetCommandLineOption uses set_mode == SET_FLAGS_VALUE (the common case)
-extern GFLAGS_DLL_DECL std::string SetCommandLineOption        (const char* name, const char* value);
-extern GFLAGS_DLL_DECL std::string SetCommandLineOptionWithMode(const char* name, const char* value, FlagSettingMode set_mode);
-
-
-// --------------------------------------------------------------------
-// Saves the states (value, default value, whether the user has set
-// the flag, registered validators, etc) of all flags, and restores
-// them when the FlagSaver is destroyed.  This is very useful in
-// tests, say, when you want to let your tests change the flags, but
-// make sure that they get reverted to the original states when your
-// test is complete.
-//
-// Example usage:
-//   void TestFoo() {
-//     FlagSaver s1;
-//     FLAG_foo = false;
-//     FLAG_bar = "some value";
-//
-//     // test happens here.  You can return at any time
-//     // without worrying about restoring the FLAG values.
-//   }
-//
-// Note: This class is marked with GFLAGS_ATTRIBUTE_UNUSED because all
-// the work is done in the constructor and destructor, so in the standard
-// usage example above, the compiler would complain that it's an
-// unused variable.
-//
-// This class is thread-safe.  However, its destructor writes to
-// exactly the set of flags that have changed value during its
-// lifetime, so concurrent _direct_ access to those flags
-// (i.e. FLAGS_foo instead of {Get,Set}CommandLineOption()) is unsafe.
-
-class GFLAGS_DLL_DECL FlagSaver {
- public:
-  FlagSaver();
-  ~FlagSaver();
-
- private:
-  class FlagSaverImpl* impl_;   // we use pimpl here to keep API steady
-
-  FlagSaver(const FlagSaver&);  // no copying!
-  void operator=(const FlagSaver&);
-}__attribute((unused));
-
-// --------------------------------------------------------------------
-// Some deprecated or hopefully-soon-to-be-deprecated functions.
-
-// This is often used for logging.  TODO(csilvers): figure out a better way
-extern GFLAGS_DLL_DECL std::string CommandlineFlagsIntoString();
-// Usually where this is used, a FlagSaver should be used instead.
-extern GFLAGS_DLL_DECL
-bool ReadFlagsFromString(const std::string& flagfilecontents,
-                         const char* prog_name,
-                         bool errors_are_fatal);  // uses SET_FLAGS_VALUE
-
-// These let you manually implement --flagfile functionality.
-// DEPRECATED.
-extern GFLAGS_DLL_DECL bool AppendFlagsIntoFile(const std::string& filename, const char* prog_name);
-extern GFLAGS_DLL_DECL bool ReadFromFlagsFile(const std::string& filename, const char* prog_name, bool errors_are_fatal);   // uses SET_FLAGS_VALUE
-
-
-// --------------------------------------------------------------------
-// Useful routines for initializing flags from the environment.
-// In each case, if 'varname' does not exist in the environment
-// return defval.  If 'varname' does exist but is not valid
-// (e.g., not a number for an int32 flag), abort with an error.
-// Otherwise, return the value.  NOTE: for booleans, for true use
-// 't' or 'T' or 'true' or '1', for false 'f' or 'F' or 'false' or '0'.
-
-extern GFLAGS_DLL_DECL bool BoolFromEnv(const char *varname, bool defval);
-extern GFLAGS_DLL_DECL int32 Int32FromEnv(const char *varname, int32 defval);
-extern GFLAGS_DLL_DECL uint32 Uint32FromEnv(const char *varname, uint32 defval);
-extern GFLAGS_DLL_DECL int64 Int64FromEnv(const char *varname, int64 defval);
-extern GFLAGS_DLL_DECL uint64 Uint64FromEnv(const char *varname, uint64 defval);
-extern GFLAGS_DLL_DECL double DoubleFromEnv(const char *varname, double defval);
-extern GFLAGS_DLL_DECL const char *StringFromEnv(const char *varname, const char *defval);
-
-
-// --------------------------------------------------------------------
-// The next two functions parse gflags from main():
-
-// Set the "usage" message for this program.  For example:
-//   string usage("This program does nothing.  Sample usage:\n");
-//   usage += argv[0] + " <uselessarg1> <uselessarg2>";
-//   SetUsageMessage(usage);
-// Do not include commandline flags in the usage: we do that for you!
-// Thread-hostile; meant to be called before any threads are spawned.
-extern GFLAGS_DLL_DECL void SetUsageMessage(const std::string& usage);
-
-// Sets the version string, which is emitted with --version.
-// For instance: SetVersionString("1.3");
-// Thread-hostile; meant to be called before any threads are spawned.
-extern GFLAGS_DLL_DECL void SetVersionString(const std::string& version);
-
-
-// Looks for flags in argv and parses them.  Rearranges argv to put
-// flags first, or removes them entirely if remove_flags is true.
-// If a flag is defined more than once in the command line or flag
-// file, the last definition is used.  Returns the index (into argv)
-// of the first non-flag argument.
-// See top-of-file for more details on this function.
-#ifndef SWIG   // In swig, use ParseCommandLineFlagsScript() instead.
-extern GFLAGS_DLL_DECL uint32 ParseCommandLineFlags(int *argc, char*** argv, bool remove_flags);
-#endif
-
-
-// Calls to ParseCommandLineNonHelpFlags and then to
-// HandleCommandLineHelpFlags can be used instead of a call to
-// ParseCommandLineFlags during initialization, in order to allow for
-// changing default values for some FLAGS (via
-// e.g. SetCommandLineOptionWithMode calls) between the time of
-// command line parsing and the time of dumping help information for
-// the flags as a result of command line parsing.  If a flag is
-// defined more than once in the command line or flag file, the last
-// definition is used.  Returns the index (into argv) of the first
-// non-flag argument.  (If remove_flags is true, will always return 1.)
-extern GFLAGS_DLL_DECL uint32 ParseCommandLineNonHelpFlags(int *argc, char*** argv, bool remove_flags);
-
-// This is actually defined in gflags_reporting.cc.
-// This function is misnamed (it also handles --version, etc.), but
-// it's too late to change that now. :-(
-extern GFLAGS_DLL_DECL void HandleCommandLineHelpFlags();   // in gflags_reporting.cc
-
-// Allow command line reparsing.  Disables the error normally
-// generated when an unknown flag is found, since it may be found in a
-// later parse.  Thread-hostile; meant to be called before any threads
-// are spawned.
-extern GFLAGS_DLL_DECL void AllowCommandLineReparsing();
-
-// Reparse the flags that have not yet been recognized.  Only flags
-// registered since the last parse will be recognized.  Any flag value
-// must be provided as part of the argument using "=", not as a
-// separate command line argument that follows the flag argument.
-// Intended for handling flags from dynamically loaded libraries,
-// since their flags are not registered until they are loaded.
-extern GFLAGS_DLL_DECL void ReparseCommandLineNonHelpFlags();
-
-// Clean up memory allocated by flags.  This is only needed to reduce
-// the quantity of "potentially leaked" reports emitted by memory
-// debugging tools such as valgrind.  It is not required for normal
-// operation, or for the google perftools heap-checker.  It must only
-// be called when the process is about to exit, and all threads that
-// might access flags are quiescent.  Referencing flags after this is
-// called will have unexpected consequences.  This is not safe to run
-// when multiple threads might be running: the function is
-// thread-hostile.
-extern GFLAGS_DLL_DECL void ShutDownCommandLineFlags();
-
-
-// --------------------------------------------------------------------
-// Now come the command line flag declaration/definition macros that
-// will actually be used.  They're kind of hairy.  A major reason
-// for this is initialization: we want people to be able to access
-// variables in global constructors and have that not crash, even if
-// their global constructor runs before the global constructor here.
-// (Obviously, we can't guarantee the flags will have the correct
-// default value in that case, but at least accessing them is safe.)
-// The only way to do that is have flags point to a static buffer.
-// So we make one, using a union to ensure proper alignment, and
-// then use placement-new to actually set up the flag with the
-// correct default value.  In the same vein, we have to worry about
-// flag access in global destructors, so FlagRegisterer has to be
-// careful never to destroy the flag-values it constructs.
-//
-// Note that when we define a flag variable FLAGS_<name>, we also
-// preemptively define a junk variable, FLAGS_no<name>.  This is to
-// cause a link-time error if someone tries to define 2 flags with
-// names like "logging" and "nologging".  We do this because a bool
-// flag FLAG can be set from the command line to true with a "-FLAG"
-// argument, and to false with a "-noFLAG" argument, and so this can
-// potentially avert confusion.
-//
-// We also put flags into their own namespace.  It is purposefully
-// named in an opaque way that people should have trouble typing
-// directly.  The idea is that DEFINE puts the flag in the weird
-// namespace, and DECLARE imports the flag from there into the current
-// namespace.  The net result is to force people to use DECLARE to get
-// access to a flag, rather than saying "extern GFLAGS_DLL_DECL bool FLAGS_whatever;"
-// or some such instead.  We want this so we can put extra
-// functionality (like sanity-checking) in DECLARE if we want, and
-// make sure it is picked up everywhere.
-//
-// We also put the type of the variable in the namespace, so that
-// people can't DECLARE_int32 something that they DEFINE_bool'd
-// elsewhere.
-
-class GFLAGS_DLL_DECL FlagRegisterer {
- public:
-  // We instantiate this template ctor for all supported types,
-  // so it is possible to place implementation of the FlagRegisterer ctor in
-  // .cc file.
-  // Calling this constructor with unsupported type will produce linker error.
-  template <typename FlagType>
-  FlagRegisterer(const char* name,
-                 const char* help, const char* filename,
-                 FlagType* current_storage, FlagType* defvalue_storage);
-};
-
-// Force compiler to not generate code for the given template specialization.
-#if defined(_MSC_VER) && _MSC_VER < 1800 // Visual Studio 2013 version 12.0
-  #define GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(type)
-#else
-  #define GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(type)                  \
-    extern template GFLAGS_DLL_DECL FlagRegisterer::FlagRegisterer(  \
-        const char* name, const char* help, const char* filename,    \
-        type* current_storage, type* defvalue_storage)
-#endif
-
-// Do this for all supported flag types.
-GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(bool);
-GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(int32);
-GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(uint32);
-GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(int64);
-GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(uint64);
-GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(double);
-GFLAGS_DECLARE_FLAG_REGISTERER_CTOR(std::string);
-
-#undef GFLAGS_DECLARE_FLAG_REGISTERER_CTOR
-
-// If your application #defines STRIP_FLAG_HELP to a non-zero value
-// before #including this file, we remove the help message from the
-// binary file. This can reduce the size of the resulting binary
-// somewhat, and may also be useful for security reasons.
-
-extern GFLAGS_DLL_DECL const char kStrippedFlagHelp[];
-
-
-} // namespace GFLAGS_NAMESPACE
-
-
-#ifndef SWIG  // In swig, ignore the main flag declarations
-
-#if defined(STRIP_FLAG_HELP) && STRIP_FLAG_HELP > 0
-// Need this construct to avoid the 'defined but not used' warning.
-#define MAYBE_STRIPPED_HELP(txt) \
-   (false ? (txt) : GFLAGS_NAMESPACE::kStrippedFlagHelp)
-#else
-#define MAYBE_STRIPPED_HELP(txt) txt
-#endif
-
-// Each command-line flag has two variables associated with it: one
-// with the current value, and one with the default value.  However,
-// we have a third variable, which is where value is assigned; it's a
-// constant.  This guarantees that FLAG_##value is initialized at
-// static initialization time (e.g. before program-start) rather than
-// than global construction time (which is after program-start but
-// before main), at least when 'value' is a compile-time constant.  We
-// use a small trick for the "default value" variable, and call it
-// FLAGS_no<name>.  This serves the second purpose of assuring a
-// compile error if someone tries to define a flag named no<name>
-// which is illegal (--foo and --nofoo both affect the "foo" flag).
-#define DEFINE_VARIABLE(type, shorttype, name, value, help)             \
-  namespace fL##shorttype {                                             \
-    static const type FLAGS_nono##name = value;                         \
-    /* We always want to export defined variables, dll or no */         \
-    GFLAGS_DLL_DEFINE_FLAG type FLAGS_##name = FLAGS_nono##name;        \
-    static type FLAGS_no##name = FLAGS_nono##name;                      \
-    static GFLAGS_NAMESPACE::FlagRegisterer o_##name(                   \
-      #name, MAYBE_STRIPPED_HELP(help), __FILE__,                       \
-      &FLAGS_##name, &FLAGS_no##name);                                  \
-  }                                                                     \
-  using fL##shorttype::FLAGS_##name
-
-// For DEFINE_bool, we want to do the extra check that the passed-in
-// value is actually a bool, and not a string or something that can be
-// coerced to a bool.  These declarations (no definition needed!) will
-// help us do that, and never evaluate From, which is important.
-// We'll use 'sizeof(IsBool(val))' to distinguish. This code requires
-// that the compiler have different sizes for bool & double. Since
-// this is not guaranteed by the standard, we check it with a
-// COMPILE_ASSERT.
-namespace fLB {
-struct CompileAssert {};
-typedef CompileAssert expected_sizeof_double_neq_sizeof_bool[
-                      (sizeof(double) != sizeof(bool)) ? 1 : -1];
-template<typename From> double GFLAGS_DLL_DECL IsBoolFlag(const From& from);
-GFLAGS_DLL_DECL bool IsBoolFlag(bool from);
-}  // namespace fLB
-
-// Here are the actual DEFINE_*-macros. The respective DECLARE_*-macros
-// are in a separate include, gflags_declare.h, for reducing
-// the physical transitive size for DECLARE use.
-#define DEFINE_bool(name, val, txt)                                     \
-  namespace fLB {                                                       \
-    typedef ::fLB::CompileAssert FLAG_##name##_value_is_not_a_bool[     \
-            (sizeof(::fLB::IsBoolFlag(val)) != sizeof(double))? 1: -1]; \
-  }                                                                     \
-  DEFINE_VARIABLE(bool, B, name, val, txt)
-
-#define DEFINE_int32(name, val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::int32, I, \
-                   name, val, txt)
-
-#define DEFINE_uint32(name,val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::uint32, U, \
-                   name, val, txt)
-
-#define DEFINE_int64(name, val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::int64, I64, \
-                   name, val, txt)
-
-#define DEFINE_uint64(name,val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::uint64, U64, \
-                   name, val, txt)
-
-#define DEFINE_double(name, val, txt) \
-   DEFINE_VARIABLE(double, D, name, val, txt)
-
-// Strings are trickier, because they're not a POD, so we can't
-// construct them at static-initialization time (instead they get
-// constructed at global-constructor time, which is much later).  To
-// try to avoid crashes in that case, we use a char buffer to store
-// the string, which we can static-initialize, and then placement-new
-// into it later.  It's not perfect, but the best we can do.
-
-namespace fLS {
-
-inline clstring* dont_pass0toDEFINE_string(char *stringspot,
-                                           const char *value) {
-  return new(stringspot) clstring(value);
-}
-inline clstring* dont_pass0toDEFINE_string(char *stringspot,
-                                           const clstring &value) {
-  return new(stringspot) clstring(value);
-}
-inline clstring* dont_pass0toDEFINE_string(char *stringspot,
-                                           int value);
-
-// Auxiliary class used to explicitly call destructor of string objects
-// allocated using placement new during static program deinitialization.
-// The destructor MUST be an inline function such that the explicit
-// destruction occurs in the same compilation unit as the placement new.
-class StringFlagDestructor {
-  void *current_storage_;
-  void *defvalue_storage_;
-
-public: 
-
-  StringFlagDestructor(void *current, void *defvalue)
-  : current_storage_(current), defvalue_storage_(defvalue) {}
-
-  ~StringFlagDestructor() {
-    reinterpret_cast<clstring*>(current_storage_ )->~clstring();
-    reinterpret_cast<clstring*>(defvalue_storage_)->~clstring();
-  }
-};
-
-}  // namespace fLS
-
-// We need to define a var named FLAGS_no##name so people don't define
-// --string and --nostring.  And we need a temporary place to put val
-// so we don't have to evaluate it twice.  Two great needs that go
-// great together!
-// The weird 'using' + 'extern' inside the fLS namespace is to work around
-// an unknown compiler bug/issue with the gcc 4.2.1 on SUSE 10.  See
-//    http://code.google.com/p/google-gflags/issues/detail?id=20
-#define DEFINE_string(name, val, txt)                                       \
-  namespace fLS {                                                           \
-    using ::fLS::clstring;                                                  \
-    using ::fLS::StringFlagDestructor;                                      \
-    static union { void* align; char s[sizeof(clstring)]; } s_##name[2];    \
-    clstring* const FLAGS_no##name = ::fLS::                                \
-                                   dont_pass0toDEFINE_string(s_##name[0].s, \
-                                                             val);          \
-    static GFLAGS_NAMESPACE::FlagRegisterer o_##name(                       \
-        #name, MAYBE_STRIPPED_HELP(txt), __FILE__,                          \
-        FLAGS_no##name, new (s_##name[1].s) clstring(*FLAGS_no##name));     \
-    static StringFlagDestructor d_##name(s_##name[0].s, s_##name[1].s);     \
-    extern GFLAGS_DLL_DEFINE_FLAG clstring& FLAGS_##name;                   \
-    using fLS::FLAGS_##name;                                                \
-    clstring& FLAGS_##name = *FLAGS_no##name;                               \
-  }                                                                         \
-  using fLS::FLAGS_##name
-
-#endif  // SWIG
-
-
-
-
-
-#endif  // GFLAGS_GFLAGS_H_
diff --git a/projects/llm_framework/include/gflags/gflags_completions.h b/projects/llm_framework/include/gflags/gflags_completions.h
deleted file mode 100644
index 2fa0db6d..00000000
--- a/projects/llm_framework/include/gflags/gflags_completions.h
+++ /dev/null
@@ -1,121 +0,0 @@
-// Copyright (c) 2008, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-//
-// ---
-
-//
-// Implement helpful bash-style command line flag completions
-//
-// ** Functional API:
-// HandleCommandLineCompletions() should be called early during
-// program startup, but after command line flag code has been
-// initialized, such as the beginning of HandleCommandLineHelpFlags().
-// It checks the value of the flag --tab_completion_word.  If this
-// flag is empty, nothing happens here.  If it contains a string,
-// however, then HandleCommandLineCompletions() will hijack the
-// process, attempting to identify the intention behind this
-// completion.  Regardless of the outcome of this deduction, the
-// process will be terminated, similar to --helpshort flag
-// handling.
-//
-// ** Overview of Bash completions:
-// Bash can be told to programatically determine completions for the
-// current 'cursor word'.  It does this by (in this case) invoking a
-// command with some additional arguments identifying the command
-// being executed, the word being completed, and the previous word
-// (if any).  Bash then expects a sequence of output lines to be
-// printed to stdout.  If these lines all contain a common prefix
-// longer than the cursor word, bash will replace the cursor word
-// with that common prefix, and display nothing.  If there isn't such
-// a common prefix, bash will display the lines in pages using 'more'.
-//
-// ** Strategy taken for command line completions:
-// If we can deduce either the exact flag intended, or a common flag
-// prefix, we'll output exactly that.  Otherwise, if information
-// must be displayed to the user, we'll take the opportunity to add
-// some helpful information beyond just the flag name (specifically,
-// we'll include the default flag value and as much of the flag's
-// description as can fit on a single terminal line width, as specified
-// by the flag --tab_completion_columns).  Furthermore, we'll try to
-// make bash order the output such that the most useful or relevent
-// flags are the most likely to be shown at the top.
-//
-// ** Additional features:
-// To assist in finding that one really useful flag, substring matching
-// was implemented.  Before pressing a <TAB> to get completion for the
-// current word, you can append one or more '?' to the flag to do
-// substring matching.  Here's the semantics:
-//   --foo<TAB>     Show me all flags with names prefixed by 'foo'
-//   --foo?<TAB>    Show me all flags with 'foo' somewhere in the name
-//   --foo??<TAB>   Same as prior case, but also search in module
-//                  definition path for 'foo'
-//   --foo???<TAB>  Same as prior case, but also search in flag
-//                  descriptions for 'foo'
-// Finally, we'll trim the output to a relatively small number of
-// flags to keep bash quiet about the verbosity of output.  If one
-// really wanted to see all possible matches, appending a '+' to the
-// search word will force the exhaustive list of matches to be printed.
-//
-// ** How to have bash accept completions from a binary:
-// Bash requires that it be informed about each command that programmatic
-// completion should be enabled for.  Example addition to a .bashrc
-// file would be (your path to gflags_completions.sh file may differ):
-
-/*
-$ complete -o bashdefault -o default -o nospace -C                            \
- '/home/build/eng/bash/bash_completions.sh --tab_completion_columns $COLUMNS' \
-  time  env  binary_name  another_binary  [...]
-*/
-
-// This would allow the following to work:
-//   $ /path/to/binary_name --vmodule<TAB>
-// Or:
-//   $ ./bin/path/another_binary --gfs_u<TAB>
-// (etc)
-//
-// Sadly, it appears that bash gives no easy way to force this behavior for
-// all commands.  That's where the "time" in the above example comes in.
-// If you haven't specifically added a command to the list of completion
-// supported commands, you can still get completions by prefixing the
-// entire command with "env".
-//   $ env /some/brand/new/binary --vmod<TAB>
-// Assuming that "binary" is a newly compiled binary, this should still
-// produce the expected completion output.
-
-
-#ifndef GFLAGS_COMPLETIONS_H_
-#define GFLAGS_COMPLETIONS_H_
-
-namespace gflags {
-
-extern void HandleCommandLineCompletions(void);
-
-}
-
-#endif  // GFLAGS_COMPLETIONS_H_
diff --git a/projects/llm_framework/include/gflags/gflags_declare.h b/projects/llm_framework/include/gflags/gflags_declare.h
deleted file mode 100644
index 69cf1129..00000000
--- a/projects/llm_framework/include/gflags/gflags_declare.h
+++ /dev/null
@@ -1,156 +0,0 @@
-// Copyright (c) 1999, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// ---
-//
-// Revamped and reorganized by Craig Silverstein
-//
-// This is the file that should be included by any file which declares
-// command line flag.
-
-#ifndef GFLAGS_DECLARE_H_
-#define GFLAGS_DECLARE_H_
-
-
-// ---------------------------------------------------------------------------
-// Namespace of gflags library symbols.
-#define GFLAGS_NAMESPACE gflags
-
-// ---------------------------------------------------------------------------
-// Windows DLL import/export.
-
-// Whether gflags library is a DLL.
-//
-// Set to 1 by default when the shared gflags library was built on Windows.
-// Must be overwritten when this header file is used with the optionally also
-// built static library instead; set by CMake's INTERFACE_COMPILE_DEFINITIONS.
-#ifndef GFLAGS_IS_A_DLL
-#  define GFLAGS_IS_A_DLL 1
-#endif
-
-// We always want to import the symbols of the gflags library.
-#ifndef GFLAGS_DLL_DECL
-#  if GFLAGS_IS_A_DLL && defined(_MSC_VER)
-#    define GFLAGS_DLL_DECL __declspec(dllimport)
-#  elif defined(__GNUC__) && __GNUC__ >= 4
-#    define GFLAGS_DLL_DECL __attribute__((visibility("default")))
-#  else
-#    define GFLAGS_DLL_DECL
-#  endif
-#endif
-
-// We always want to import variables declared in user code.
-#ifndef GFLAGS_DLL_DECLARE_FLAG
-#  if GFLAGS_IS_A_DLL && defined(_MSC_VER)
-#    define GFLAGS_DLL_DECLARE_FLAG __declspec(dllimport)
-#  elif defined(__GNUC__) && __GNUC__ >= 4
-#    define GFLAGS_DLL_DECLARE_FLAG __attribute__((visibility("default")))
-#  else
-#    define GFLAGS_DLL_DECLARE_FLAG
-#  endif
-#endif
-
-// ---------------------------------------------------------------------------
-// Flag types
-#include <string>
-#if 1
-#  include <stdint.h>                   // the normal place uint32_t is defined
-#elif 1
-#  include <sys/types.h>                // the normal place u_int32_t is defined
-#elif 1
-#  include <inttypes.h>                 // a third place for uint32_t or u_int32_t
-#endif
-
-namespace GFLAGS_NAMESPACE {
-
-#if 1 // C99
-typedef int32_t          int32;
-typedef uint32_t         uint32;
-typedef int64_t          int64;
-typedef uint64_t         uint64;
-#elif 0 // BSD
-typedef int32_t          int32;
-typedef u_int32_t        uint32;
-typedef int64_t          int64;
-typedef u_int64_t        uint64;
-#elif 0 // Windows
-typedef __int32          int32;
-typedef unsigned __int32 uint32;
-typedef __int64          int64;
-typedef unsigned __int64 uint64;
-#else
-#  error Do not know how to define a 32-bit integer quantity on your system
-#endif
-
-} // namespace GFLAGS_NAMESPACE
-
-
-namespace fLS {
-
-// The meaning of "string" might be different between now and when the
-// macros below get invoked (e.g., if someone is experimenting with
-// other string implementations that get defined after this file is
-// included).  Save the current meaning now and use it in the macros.
-typedef std::string clstring;
-
-} // namespace fLS
-
-
-#define DECLARE_VARIABLE(type, shorttype, name) \
-  /* We always want to import declared variables, dll or no */ \
-  namespace fL##shorttype { extern GFLAGS_DLL_DECLARE_FLAG type FLAGS_##name; } \
-  using fL##shorttype::FLAGS_##name
-
-#define DECLARE_bool(name) \
-  DECLARE_VARIABLE(bool, B, name)
-
-#define DECLARE_int32(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::int32, I, name)
-
-#define DECLARE_uint32(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::uint32, U, name)
-
-#define DECLARE_int64(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::int64, I64, name)
-
-#define DECLARE_uint64(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::uint64, U64, name)
-
-#define DECLARE_double(name) \
-  DECLARE_VARIABLE(double, D, name)
-
-#define DECLARE_string(name) \
-  /* We always want to import declared variables, dll or no */ \
-  namespace fLS { \
-  extern GFLAGS_DLL_DECLARE_FLAG ::fLS::clstring& FLAGS_##name; \
-  } \
-  using fLS::FLAGS_##name
-
-
-#endif  // GFLAGS_DECLARE_H_
diff --git a/projects/llm_framework/include/glog/log_severity.h b/projects/llm_framework/include/glog/log_severity.h
deleted file mode 100644
index 99945a42..00000000
--- a/projects/llm_framework/include/glog/log_severity.h
+++ /dev/null
@@ -1,92 +0,0 @@
-// Copyright (c) 2007, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef BASE_LOG_SEVERITY_H__
-#define BASE_LOG_SEVERITY_H__
-
-// Annoying stuff for windows -- makes sure clients can import these functions
-#ifndef GOOGLE_GLOG_DLL_DECL
-# if defined(_WIN32) && !defined(__CYGWIN__)
-#   define GOOGLE_GLOG_DLL_DECL  __declspec(dllimport)
-# else
-#   define GOOGLE_GLOG_DLL_DECL
-# endif
-#endif
-
-// Variables of type LogSeverity are widely taken to lie in the range
-// [0, NUM_SEVERITIES-1].  Be careful to preserve this assumption if
-// you ever need to change their values or add a new severity.
-typedef int LogSeverity;
-
-const int GLOG_INFO = 0, GLOG_WARNING = 1, GLOG_ERROR = 2, GLOG_FATAL = 3,
-  NUM_SEVERITIES = 4;
-#ifndef GLOG_NO_ABBREVIATED_SEVERITIES
-# ifdef ERROR
-#  error ERROR macro is defined. Define GLOG_NO_ABBREVIATED_SEVERITIES before including logging.h. See the document for detail.
-# endif
-const int INFO = GLOG_INFO, WARNING = GLOG_WARNING,
-  ERROR = GLOG_ERROR, FATAL = GLOG_FATAL;
-#endif
-
-// DFATAL is FATAL in debug mode, ERROR in normal mode
-#ifdef NDEBUG
-#define DFATAL_LEVEL ERROR
-#else
-#define DFATAL_LEVEL FATAL
-#endif
-
-extern GOOGLE_GLOG_DLL_DECL const char* const LogSeverityNames[NUM_SEVERITIES];
-
-// NDEBUG usage helpers related to (RAW_)DCHECK:
-//
-// DEBUG_MODE is for small !NDEBUG uses like
-//   if (DEBUG_MODE) foo.CheckThatFoo();
-// instead of substantially more verbose
-//   #ifndef NDEBUG
-//     foo.CheckThatFoo();
-//   #endif
-//
-// IF_DEBUG_MODE is for small !NDEBUG uses like
-//   IF_DEBUG_MODE( string error; )
-//   DCHECK(Foo(&error)) << error;
-// instead of substantially more verbose
-//   #ifndef NDEBUG
-//     string error;
-//     DCHECK(Foo(&error)) << error;
-//   #endif
-//
-#ifdef NDEBUG
-enum { DEBUG_MODE = 0 };
-#define IF_DEBUG_MODE(x)
-#else
-enum { DEBUG_MODE = 1 };
-#define IF_DEBUG_MODE(x) x
-#endif
-
-#endif  // BASE_LOG_SEVERITY_H__
diff --git a/projects/llm_framework/include/glog/logging.h b/projects/llm_framework/include/glog/logging.h
deleted file mode 100644
index 4cd247e1..00000000
--- a/projects/llm_framework/include/glog/logging.h
+++ /dev/null
@@ -1,1662 +0,0 @@
-// Copyright (c) 1999, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-//
-// Author: Ray Sidney
-//
-// This file contains #include information about logging-related stuff.
-// Pretty much everybody needs to #include this file so that they can
-// log various happenings.
-//
-#ifndef _LOGGING_H_
-#define _LOGGING_H_
-
-#include <errno.h>
-#include <string.h>
-#include <time.h>
-#include <iosfwd>
-#include <ostream>
-#include <sstream>
-#include <string>
-#if 1
-# include <unistd.h>
-#endif
-#include <vector>
-
-#if defined(_MSC_VER)
-#define GLOG_MSVC_PUSH_DISABLE_WARNING(n) __pragma(warning(push)) \
-                                     __pragma(warning(disable:n))
-#define GLOG_MSVC_POP_WARNING() __pragma(warning(pop))
-#else
-#define GLOG_MSVC_PUSH_DISABLE_WARNING(n)
-#define GLOG_MSVC_POP_WARNING()
-#endif
-
-// Annoying stuff for windows -- makes sure clients can import these functions
-#ifndef GOOGLE_GLOG_DLL_DECL
-# if defined(_WIN32) && !defined(__CYGWIN__)
-#   define GOOGLE_GLOG_DLL_DECL  __declspec(dllimport)
-# else
-#   define GOOGLE_GLOG_DLL_DECL
-# endif
-#endif
-
-// We care a lot about number of bits things take up.  Unfortunately,
-// systems define their bit-specific ints in a lot of different ways.
-// We use our own way, and have a typedef to get there.
-// Note: these commands below may look like "#if 1" or "#if 0", but
-// that's because they were constructed that way at ./configure time.
-// Look at logging.h.in to see how they're calculated (based on your config).
-#if 1
-#include <stdint.h>             // the normal place uint16_t is defined
-#endif
-#if 1
-#include <sys/types.h>          // the normal place u_int16_t is defined
-#endif
-#if 1
-#include <inttypes.h>           // a third place for uint16_t or u_int16_t
-#endif
-
-#if 0
-#include <gflags/gflags.h>
-#endif
-
-namespace google {
-
-#if 1      // the C99 format
-typedef int32_t int32;
-typedef uint32_t uint32;
-typedef int64_t int64;
-typedef uint64_t uint64;
-#elif 1   // the BSD format
-typedef int32_t int32;
-typedef u_int32_t uint32;
-typedef int64_t int64;
-typedef u_int64_t uint64;
-#elif 0    // the windows (vc7) format
-typedef __int32 int32;
-typedef unsigned __int32 uint32;
-typedef __int64 int64;
-typedef unsigned __int64 uint64;
-#else
-#error Do not know how to define a 32-bit integer quantity on your system
-#endif
-
-}
-
-// The global value of GOOGLE_STRIP_LOG. All the messages logged to
-// LOG(XXX) with severity less than GOOGLE_STRIP_LOG will not be displayed.
-// If it can be determined at compile time that the message will not be
-// printed, the statement will be compiled out.
-//
-// Example: to strip out all INFO and WARNING messages, use the value
-// of 2 below. To make an exception for WARNING messages from a single
-// file, add "#define GOOGLE_STRIP_LOG 1" to that file _before_ including
-// base/logging.h
-#ifndef GOOGLE_STRIP_LOG
-#define GOOGLE_STRIP_LOG 0
-#endif
-
-// GCC can be told that a certain branch is not likely to be taken (for
-// instance, a CHECK failure), and use that information in static analysis.
-// Giving it this information can help it optimize for the common case in
-// the absence of better information (ie. -fprofile-arcs).
-//
-#ifndef GOOGLE_PREDICT_BRANCH_NOT_TAKEN
-#if 1
-#define GOOGLE_PREDICT_BRANCH_NOT_TAKEN(x) (__builtin_expect(x, 0))
-#else
-#define GOOGLE_PREDICT_BRANCH_NOT_TAKEN(x) x
-#endif
-#endif
-
-#ifndef GOOGLE_PREDICT_FALSE
-#if 1
-#define GOOGLE_PREDICT_FALSE(x) (__builtin_expect(x, 0))
-#else
-#define GOOGLE_PREDICT_FALSE(x) x
-#endif
-#endif
-
-#ifndef GOOGLE_PREDICT_TRUE
-#if 1
-#define GOOGLE_PREDICT_TRUE(x) (__builtin_expect(!!(x), 1))
-#else
-#define GOOGLE_PREDICT_TRUE(x) x
-#endif
-#endif
-
-
-// Make a bunch of macros for logging.  The way to log things is to stream
-// things to LOG(<a particular severity level>).  E.g.,
-//
-//   LOG(INFO) << "Found " << num_cookies << " cookies";
-//
-// You can capture log messages in a string, rather than reporting them
-// immediately:
-//
-//   vector<string> errors;
-//   LOG_STRING(ERROR, &errors) << "Couldn't parse cookie #" << cookie_num;
-//
-// This pushes back the new error onto 'errors'; if given a NULL pointer,
-// it reports the error via LOG(ERROR).
-//
-// You can also do conditional logging:
-//
-//   LOG_IF(INFO, num_cookies > 10) << "Got lots of cookies";
-//
-// You can also do occasional logging (log every n'th occurrence of an
-// event):
-//
-//   LOG_EVERY_N(INFO, 10) << "Got the " << google::COUNTER << "th cookie";
-//
-// The above will cause log messages to be output on the 1st, 11th, 21st, ...
-// times it is executed.  Note that the special google::COUNTER value is used
-// to identify which repetition is happening.
-//
-// You can also do occasional conditional logging (log every n'th
-// occurrence of an event, when condition is satisfied):
-//
-//   LOG_IF_EVERY_N(INFO, (size > 1024), 10) << "Got the " << google::COUNTER
-//                                           << "th big cookie";
-//
-// You can log messages the first N times your code executes a line. E.g.
-//
-//   LOG_FIRST_N(INFO, 20) << "Got the " << google::COUNTER << "th cookie";
-//
-// Outputs log messages for the first 20 times it is executed.
-//
-// Analogous SYSLOG, SYSLOG_IF, and SYSLOG_EVERY_N macros are available.
-// These log to syslog as well as to the normal logs.  If you use these at
-// all, you need to be aware that syslog can drastically reduce performance,
-// especially if it is configured for remote logging!  Don't use these
-// unless you fully understand this and have a concrete need to use them.
-// Even then, try to minimize your use of them.
-//
-// There are also "debug mode" logging macros like the ones above:
-//
-//   DLOG(INFO) << "Found cookies";
-//
-//   DLOG_IF(INFO, num_cookies > 10) << "Got lots of cookies";
-//
-//   DLOG_EVERY_N(INFO, 10) << "Got the " << google::COUNTER << "th cookie";
-//
-// All "debug mode" logging is compiled away to nothing for non-debug mode
-// compiles.
-//
-// We also have
-//
-//   LOG_ASSERT(assertion);
-//   DLOG_ASSERT(assertion);
-//
-// which is syntactic sugar for {,D}LOG_IF(FATAL, assert fails) << assertion;
-//
-// There are "verbose level" logging macros.  They look like
-//
-//   VLOG(1) << "I'm printed when you run the program with --v=1 or more";
-//   VLOG(2) << "I'm printed when you run the program with --v=2 or more";
-//
-// These always log at the INFO log level (when they log at all).
-// The verbose logging can also be turned on module-by-module.  For instance,
-//    --vmodule=mapreduce=2,file=1,gfs*=3 --v=0
-// will cause:
-//   a. VLOG(2) and lower messages to be printed from mapreduce.{h,cc}
-//   b. VLOG(1) and lower messages to be printed from file.{h,cc}
-//   c. VLOG(3) and lower messages to be printed from files prefixed with "gfs"
-//   d. VLOG(0) and lower messages to be printed from elsewhere
-//
-// The wildcarding functionality shown by (c) supports both '*' (match
-// 0 or more characters) and '?' (match any single character) wildcards.
-//
-// There's also VLOG_IS_ON(n) "verbose level" condition macro. To be used as
-//
-//   if (VLOG_IS_ON(2)) {
-//     // do some logging preparation and logging
-//     // that can't be accomplished with just VLOG(2) << ...;
-//   }
-//
-// There are also VLOG_IF, VLOG_EVERY_N and VLOG_IF_EVERY_N "verbose level"
-// condition macros for sample cases, when some extra computation and
-// preparation for logs is not needed.
-//   VLOG_IF(1, (size > 1024))
-//      << "I'm printed when size is more than 1024 and when you run the "
-//         "program with --v=1 or more";
-//   VLOG_EVERY_N(1, 10)
-//      << "I'm printed every 10th occurrence, and when you run the program "
-//         "with --v=1 or more. Present occurence is " << google::COUNTER;
-//   VLOG_IF_EVERY_N(1, (size > 1024), 10)
-//      << "I'm printed on every 10th occurence of case when size is more "
-//         " than 1024, when you run the program with --v=1 or more. ";
-//         "Present occurence is " << google::COUNTER;
-//
-// The supported severity levels for macros that allow you to specify one
-// are (in increasing order of severity) INFO, WARNING, ERROR, and FATAL.
-// Note that messages of a given severity are logged not only in the
-// logfile for that severity, but also in all logfiles of lower severity.
-// E.g., a message of severity FATAL will be logged to the logfiles of
-// severity FATAL, ERROR, WARNING, and INFO.
-//
-// There is also the special severity of DFATAL, which logs FATAL in
-// debug mode, ERROR in normal mode.
-//
-// Very important: logging a message at the FATAL severity level causes
-// the program to terminate (after the message is logged).
-//
-// Unless otherwise specified, logs will be written to the filename
-// "<program name>.<hostname>.<user name>.log.<severity level>.", followed
-// by the date, time, and pid (you can't prevent the date, time, and pid
-// from being in the filename).
-//
-// The logging code takes two flags:
-//     --v=#           set the verbose level
-//     --logtostderr   log all the messages to stderr instead of to logfiles
-
-// LOG LINE PREFIX FORMAT
-//
-// Log lines have this form:
-//
-//     Lmmdd hh:mm:ss.uuuuuu threadid file:line] msg...
-//
-// where the fields are defined as follows:
-//
-//   L                A single character, representing the log level
-//                    (eg 'I' for INFO)
-//   mm               The month (zero padded; ie May is '05')
-//   dd               The day (zero padded)
-//   hh:mm:ss.uuuuuu  Time in hours, minutes and fractional seconds
-//   threadid         The space-padded thread ID as returned by GetTID()
-//                    (this matches the PID on Linux)
-//   file             The file name
-//   line             The line number
-//   msg              The user-supplied message
-//
-// Example:
-//
-//   I1103 11:57:31.739339 24395 google.cc:2341] Command line: ./some_prog
-//   I1103 11:57:31.739403 24395 google.cc:2342] Process id 24395
-//
-// NOTE: although the microseconds are useful for comparing events on
-// a single machine, clocks on different machines may not be well
-// synchronized.  Hence, use caution when comparing the low bits of
-// timestamps from different machines.
-
-#ifndef DECLARE_VARIABLE
-#define MUST_UNDEF_GFLAGS_DECLARE_MACROS
-#define DECLARE_VARIABLE(type, shorttype, name, tn)                     \
-  namespace fL##shorttype {                                             \
-    extern GOOGLE_GLOG_DLL_DECL type FLAGS_##name;                      \
-  }                                                                     \
-  using fL##shorttype::FLAGS_##name
-
-// bool specialization
-#define DECLARE_bool(name) \
-  DECLARE_VARIABLE(bool, B, name, bool)
-
-// int32 specialization
-#define DECLARE_int32(name) \
-  DECLARE_VARIABLE(google::int32, I, name, int32)
-
-// Special case for string, because we have to specify the namespace
-// std::string, which doesn't play nicely with our FLAG__namespace hackery.
-#define DECLARE_string(name)                                            \
-  namespace fLS {                                                       \
-    extern GOOGLE_GLOG_DLL_DECL std::string& FLAGS_##name;              \
-  }                                                                     \
-  using fLS::FLAGS_##name
-#endif
-
-// Set whether log messages go to stderr instead of logfiles
-DECLARE_bool(logtostderr);
-
-// Set whether log messages go to stderr in addition to logfiles.
-DECLARE_bool(alsologtostderr);
-
-// Set color messages logged to stderr (if supported by terminal).
-DECLARE_bool(colorlogtostderr);
-
-// Log messages at a level >= this flag are automatically sent to
-// stderr in addition to log files.
-DECLARE_int32(stderrthreshold);
-
-// Set whether the log prefix should be prepended to each line of output.
-DECLARE_bool(log_prefix);
-
-// Log messages at a level <= this flag are buffered.
-// Log messages at a higher level are flushed immediately.
-DECLARE_int32(logbuflevel);
-
-// Sets the maximum number of seconds which logs may be buffered for.
-DECLARE_int32(logbufsecs);
-
-// Log suppression level: messages logged at a lower level than this
-// are suppressed.
-DECLARE_int32(minloglevel);
-
-// If specified, logfiles are written into this directory instead of the
-// default logging directory.
-DECLARE_string(log_dir);
-
-// Set the log file mode.
-DECLARE_int32(logfile_mode);
-
-// Sets the path of the directory into which to put additional links
-// to the log files.
-DECLARE_string(log_link);
-
-DECLARE_int32(v);  // in vlog_is_on.cc
-
-// Sets the maximum log file size (in MB).
-DECLARE_int32(max_log_size);
-
-// Sets whether to avoid logging to the disk if the disk is full.
-DECLARE_bool(stop_logging_if_full_disk);
-
-#ifdef MUST_UNDEF_GFLAGS_DECLARE_MACROS
-#undef MUST_UNDEF_GFLAGS_DECLARE_MACROS
-#undef DECLARE_VARIABLE
-#undef DECLARE_bool
-#undef DECLARE_int32
-#undef DECLARE_string
-#endif
-
-// Log messages below the GOOGLE_STRIP_LOG level will be compiled away for
-// security reasons. See LOG(severtiy) below.
-
-// A few definitions of macros that don't generate much code.  Since
-// LOG(INFO) and its ilk are used all over our code, it's
-// better to have compact code for these operations.
-
-#if GOOGLE_STRIP_LOG == 0
-#define COMPACT_GOOGLE_LOG_INFO google::LogMessage( \
-      __FILE__, __LINE__)
-#define LOG_TO_STRING_INFO(message) google::LogMessage( \
-      __FILE__, __LINE__, google::GLOG_INFO, message)
-#else
-#define COMPACT_GOOGLE_LOG_INFO google::NullStream()
-#define LOG_TO_STRING_INFO(message) google::NullStream()
-#endif
-
-#if GOOGLE_STRIP_LOG <= 1
-#define COMPACT_GOOGLE_LOG_WARNING google::LogMessage( \
-      __FILE__, __LINE__, google::GLOG_WARNING)
-#define LOG_TO_STRING_WARNING(message) google::LogMessage( \
-      __FILE__, __LINE__, google::GLOG_WARNING, message)
-#else
-#define COMPACT_GOOGLE_LOG_WARNING google::NullStream()
-#define LOG_TO_STRING_WARNING(message) google::NullStream()
-#endif
-
-#if GOOGLE_STRIP_LOG <= 2
-#define COMPACT_GOOGLE_LOG_ERROR google::LogMessage( \
-      __FILE__, __LINE__, google::GLOG_ERROR)
-#define LOG_TO_STRING_ERROR(message) google::LogMessage( \
-      __FILE__, __LINE__, google::GLOG_ERROR, message)
-#else
-#define COMPACT_GOOGLE_LOG_ERROR google::NullStream()
-#define LOG_TO_STRING_ERROR(message) google::NullStream()
-#endif
-
-#if GOOGLE_STRIP_LOG <= 3
-#define COMPACT_GOOGLE_LOG_FATAL google::LogMessageFatal( \
-      __FILE__, __LINE__)
-#define LOG_TO_STRING_FATAL(message) google::LogMessage( \
-      __FILE__, __LINE__, google::GLOG_FATAL, message)
-#else
-#define COMPACT_GOOGLE_LOG_FATAL google::NullStreamFatal()
-#define LOG_TO_STRING_FATAL(message) google::NullStreamFatal()
-#endif
-
-#if defined(NDEBUG) && !defined(DCHECK_ALWAYS_ON)
-#define DCHECK_IS_ON() 0
-#else
-#define DCHECK_IS_ON() 1
-#endif
-
-// For DFATAL, we want to use LogMessage (as opposed to
-// LogMessageFatal), to be consistent with the original behavior.
-#if !DCHECK_IS_ON()
-#define COMPACT_GOOGLE_LOG_DFATAL COMPACT_GOOGLE_LOG_ERROR
-#elif GOOGLE_STRIP_LOG <= 3
-#define COMPACT_GOOGLE_LOG_DFATAL google::LogMessage( \
-      __FILE__, __LINE__, google::GLOG_FATAL)
-#else
-#define COMPACT_GOOGLE_LOG_DFATAL google::NullStreamFatal()
-#endif
-
-#define GOOGLE_LOG_INFO(counter) google::LogMessage(__FILE__, __LINE__, google::GLOG_INFO, counter, &google::LogMessage::SendToLog)
-#define SYSLOG_INFO(counter) \
-  google::LogMessage(__FILE__, __LINE__, google::GLOG_INFO, counter, \
-  &google::LogMessage::SendToSyslogAndLog)
-#define GOOGLE_LOG_WARNING(counter)  \
-  google::LogMessage(__FILE__, __LINE__, google::GLOG_WARNING, counter, \
-  &google::LogMessage::SendToLog)
-#define SYSLOG_WARNING(counter)  \
-  google::LogMessage(__FILE__, __LINE__, google::GLOG_WARNING, counter, \
-  &google::LogMessage::SendToSyslogAndLog)
-#define GOOGLE_LOG_ERROR(counter)  \
-  google::LogMessage(__FILE__, __LINE__, google::GLOG_ERROR, counter, \
-  &google::LogMessage::SendToLog)
-#define SYSLOG_ERROR(counter)  \
-  google::LogMessage(__FILE__, __LINE__, google::GLOG_ERROR, counter, \
-  &google::LogMessage::SendToSyslogAndLog)
-#define GOOGLE_LOG_FATAL(counter) \
-  google::LogMessage(__FILE__, __LINE__, google::GLOG_FATAL, counter, \
-  &google::LogMessage::SendToLog)
-#define SYSLOG_FATAL(counter) \
-  google::LogMessage(__FILE__, __LINE__, google::GLOG_FATAL, counter, \
-  &google::LogMessage::SendToSyslogAndLog)
-#define GOOGLE_LOG_DFATAL(counter) \
-  google::LogMessage(__FILE__, __LINE__, google::DFATAL_LEVEL, counter, \
-  &google::LogMessage::SendToLog)
-#define SYSLOG_DFATAL(counter) \
-  google::LogMessage(__FILE__, __LINE__, google::DFATAL_LEVEL, counter, \
-  &google::LogMessage::SendToSyslogAndLog)
-
-#if defined(WIN32) || defined(_WIN32) || defined(__WIN32__) || defined(__CYGWIN__) || defined(__CYGWIN32__)
-// A very useful logging macro to log windows errors:
-#define LOG_SYSRESULT(result) \
-  if (FAILED(HRESULT_FROM_WIN32(result))) { \
-    LPSTR message = NULL; \
-    LPSTR msg = reinterpret_cast<LPSTR>(&message); \
-    DWORD message_length = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | \
-                         FORMAT_MESSAGE_FROM_SYSTEM, \
-                         0, result, 0, msg, 100, NULL); \
-    if (message_length > 0) { \
-      google::LogMessage(__FILE__, __LINE__, google::GLOG_ERROR, 0, \
-          &google::LogMessage::SendToLog).stream() \
-          << reinterpret_cast<const char*>(message); \
-      LocalFree(message); \
-    } \
-  }
-#endif
-
-// We use the preprocessor's merging operator, "##", so that, e.g.,
-// LOG(INFO) becomes the token GOOGLE_LOG_INFO.  There's some funny
-// subtle difference between ostream member streaming functions (e.g.,
-// ostream::operator<<(int) and ostream non-member streaming functions
-// (e.g., ::operator<<(ostream&, string&): it turns out that it's
-// impossible to stream something like a string directly to an unnamed
-// ostream. We employ a neat hack by calling the stream() member
-// function of LogMessage which seems to avoid the problem.
-#define LOG(severity) COMPACT_GOOGLE_LOG_ ## severity.stream()
-#define SYSLOG(severity) SYSLOG_ ## severity(0).stream()
-
-namespace google {
-
-// They need the definitions of integer types.
-#include "glog/log_severity.h"
-#include "glog/vlog_is_on.h"
-
-// Initialize google's logging library. You will see the program name
-// specified by argv0 in log outputs.
-GOOGLE_GLOG_DLL_DECL void InitGoogleLogging(const char* argv0);
-
-// Shutdown google's logging library.
-GOOGLE_GLOG_DLL_DECL void ShutdownGoogleLogging();
-
-// Install a function which will be called after LOG(FATAL).
-GOOGLE_GLOG_DLL_DECL void InstallFailureFunction(void (*fail_func)());
-
-class LogSink;  // defined below
-
-// If a non-NULL sink pointer is given, we push this message to that sink.
-// For LOG_TO_SINK we then do normal LOG(severity) logging as well.
-// This is useful for capturing messages and passing/storing them
-// somewhere more specific than the global log of the process.
-// Argument types:
-//   LogSink* sink;
-//   LogSeverity severity;
-// The cast is to disambiguate NULL arguments.
-#define LOG_TO_SINK(sink, severity) \
-  google::LogMessage(                                    \
-      __FILE__, __LINE__,                                               \
-      google::GLOG_ ## severity,                         \
-      static_cast<google::LogSink*>(sink), true).stream()
-#define LOG_TO_SINK_BUT_NOT_TO_LOGFILE(sink, severity)                  \
-  google::LogMessage(                                    \
-      __FILE__, __LINE__,                                               \
-      google::GLOG_ ## severity,                         \
-      static_cast<google::LogSink*>(sink), false).stream()
-
-// If a non-NULL string pointer is given, we write this message to that string.
-// We then do normal LOG(severity) logging as well.
-// This is useful for capturing messages and storing them somewhere more
-// specific than the global log of the process.
-// Argument types:
-//   string* message;
-//   LogSeverity severity;
-// The cast is to disambiguate NULL arguments.
-// NOTE: LOG(severity) expands to LogMessage().stream() for the specified
-// severity.
-#define LOG_TO_STRING(severity, message) \
-  LOG_TO_STRING_##severity(static_cast<string*>(message)).stream()
-
-// If a non-NULL pointer is given, we push the message onto the end
-// of a vector of strings; otherwise, we report it with LOG(severity).
-// This is handy for capturing messages and perhaps passing them back
-// to the caller, rather than reporting them immediately.
-// Argument types:
-//   LogSeverity severity;
-//   vector<string> *outvec;
-// The cast is to disambiguate NULL arguments.
-#define LOG_STRING(severity, outvec) \
-  LOG_TO_STRING_##severity(static_cast<std::vector<std::string>*>(outvec)).stream()
-
-#define LOG_IF(severity, condition) \
-  static_cast<void>(0),             \
-  !(condition) ? (void) 0 : google::LogMessageVoidify() & LOG(severity)
-#define SYSLOG_IF(severity, condition) \
-  static_cast<void>(0),                \
-  !(condition) ? (void) 0 : google::LogMessageVoidify() & SYSLOG(severity)
-
-#define LOG_ASSERT(condition)  \
-  LOG_IF(FATAL, !(condition)) << "Assert failed: " #condition
-#define SYSLOG_ASSERT(condition) \
-  SYSLOG_IF(FATAL, !(condition)) << "Assert failed: " #condition
-
-// CHECK dies with a fatal error if condition is not true.  It is *not*
-// controlled by DCHECK_IS_ON(), so the check will be executed regardless of
-// compilation mode.  Therefore, it is safe to do things like:
-//    CHECK(fp->Write(x) == 4)
-#define CHECK(condition)  \
-      LOG_IF(FATAL, GOOGLE_PREDICT_BRANCH_NOT_TAKEN(!(condition))) \
-             << "Check failed: " #condition " "
-
-// A container for a string pointer which can be evaluated to a bool -
-// true iff the pointer is NULL.
-struct CheckOpString {
-  CheckOpString(std::string* str) : str_(str) { }
-  // No destructor: if str_ is non-NULL, we're about to LOG(FATAL),
-  // so there's no point in cleaning up str_.
-  operator bool() const {
-    return GOOGLE_PREDICT_BRANCH_NOT_TAKEN(str_ != NULL);
-  }
-  std::string* str_;
-};
-
-// Function is overloaded for integral types to allow static const
-// integrals declared in classes and not defined to be used as arguments to
-// CHECK* macros. It's not encouraged though.
-template <class T>
-inline const T&       GetReferenceableValue(const T&           t) { return t; }
-inline char           GetReferenceableValue(char               t) { return t; }
-inline unsigned char  GetReferenceableValue(unsigned char      t) { return t; }
-inline signed char    GetReferenceableValue(signed char        t) { return t; }
-inline short          GetReferenceableValue(short              t) { return t; }
-inline unsigned short GetReferenceableValue(unsigned short     t) { return t; }
-inline int            GetReferenceableValue(int                t) { return t; }
-inline unsigned int   GetReferenceableValue(unsigned int       t) { return t; }
-inline long           GetReferenceableValue(long               t) { return t; }
-inline unsigned long  GetReferenceableValue(unsigned long      t) { return t; }
-inline long long      GetReferenceableValue(long long          t) { return t; }
-inline unsigned long long GetReferenceableValue(unsigned long long t) {
-  return t;
-}
-
-// This is a dummy class to define the following operator.
-struct DummyClassToDefineOperator {};
-
-}
-
-// Define global operator<< to declare using ::operator<<.
-// This declaration will allow use to use CHECK macros for user
-// defined classes which have operator<< (e.g., stl_logging.h).
-inline std::ostream& operator<<(
-    std::ostream& out, const google::DummyClassToDefineOperator&) {
-  return out;
-}
-
-namespace google {
-
-// This formats a value for a failing CHECK_XX statement.  Ordinarily,
-// it uses the definition for operator<<, with a few special cases below.
-template <typename T>
-inline void MakeCheckOpValueString(std::ostream* os, const T& v) {
-  (*os) << v;
-}
-
-// Overrides for char types provide readable values for unprintable
-// characters.
-template <> GOOGLE_GLOG_DLL_DECL
-void MakeCheckOpValueString(std::ostream* os, const char& v);
-template <> GOOGLE_GLOG_DLL_DECL
-void MakeCheckOpValueString(std::ostream* os, const signed char& v);
-template <> GOOGLE_GLOG_DLL_DECL
-void MakeCheckOpValueString(std::ostream* os, const unsigned char& v);
-
-// Build the error message string. Specify no inlining for code size.
-template <typename T1, typename T2>
-std::string* MakeCheckOpString(const T1& v1, const T2& v2, const char* exprtext)
-    __attribute__((noinline));
-
-namespace base {
-namespace internal {
-
-// If "s" is less than base_logging::INFO, returns base_logging::INFO.
-// If "s" is greater than base_logging::FATAL, returns
-// base_logging::ERROR.  Otherwise, returns "s".
-LogSeverity NormalizeSeverity(LogSeverity s);
-
-}  // namespace internal
-
-// A helper class for formatting "expr (V1 vs. V2)" in a CHECK_XX
-// statement.  See MakeCheckOpString for sample usage.  Other
-// approaches were considered: use of a template method (e.g.,
-// base::BuildCheckOpString(exprtext, base::Print<T1>, &v1,
-// base::Print<T2>, &v2), however this approach has complications
-// related to volatile arguments and function-pointer arguments).
-class GOOGLE_GLOG_DLL_DECL CheckOpMessageBuilder {
- public:
-  // Inserts "exprtext" and " (" to the stream.
-  explicit CheckOpMessageBuilder(const char *exprtext);
-  // Deletes "stream_".
-  ~CheckOpMessageBuilder();
-  // For inserting the first variable.
-  std::ostream* ForVar1() { return stream_; }
-  // For inserting the second variable (adds an intermediate " vs. ").
-  std::ostream* ForVar2();
-  // Get the result (inserts the closing ")").
-  std::string* NewString();
-
- private:
-  std::ostringstream *stream_;
-};
-
-}  // namespace base
-
-template <typename T1, typename T2>
-std::string* MakeCheckOpString(const T1& v1, const T2& v2, const char* exprtext) {
-  base::CheckOpMessageBuilder comb(exprtext);
-  MakeCheckOpValueString(comb.ForVar1(), v1);
-  MakeCheckOpValueString(comb.ForVar2(), v2);
-  return comb.NewString();
-}
-
-// Helper functions for CHECK_OP macro.
-// The (int, int) specialization works around the issue that the compiler
-// will not instantiate the template version of the function on values of
-// unnamed enum type - see comment below.
-#define DEFINE_CHECK_OP_IMPL(name, op) \
-  template <typename T1, typename T2> \
-  inline std::string* name##Impl(const T1& v1, const T2& v2,    \
-                            const char* exprtext) { \
-    if (GOOGLE_PREDICT_TRUE(v1 op v2)) return NULL; \
-    else return MakeCheckOpString(v1, v2, exprtext); \
-  } \
-  inline std::string* name##Impl(int v1, int v2, const char* exprtext) { \
-    return name##Impl<int, int>(v1, v2, exprtext); \
-  }
-
-// We use the full name Check_EQ, Check_NE, etc. in case the file including
-// base/logging.h provides its own #defines for the simpler names EQ, NE, etc.
-// This happens if, for example, those are used as token names in a
-// yacc grammar.
-DEFINE_CHECK_OP_IMPL(Check_EQ, ==)  // Compilation error with CHECK_EQ(NULL, x)?
-DEFINE_CHECK_OP_IMPL(Check_NE, !=)  // Use CHECK(x == NULL) instead.
-DEFINE_CHECK_OP_IMPL(Check_LE, <=)
-DEFINE_CHECK_OP_IMPL(Check_LT, < )
-DEFINE_CHECK_OP_IMPL(Check_GE, >=)
-DEFINE_CHECK_OP_IMPL(Check_GT, > )
-#undef DEFINE_CHECK_OP_IMPL
-
-// Helper macro for binary operators.
-// Don't use this macro directly in your code, use CHECK_EQ et al below.
-
-#if defined(STATIC_ANALYSIS)
-// Only for static analysis tool to know that it is equivalent to assert
-#define CHECK_OP_LOG(name, op, val1, val2, log) CHECK((val1) op (val2))
-#elif DCHECK_IS_ON()
-// In debug mode, avoid constructing CheckOpStrings if possible,
-// to reduce the overhead of CHECK statments by 2x.
-// Real DCHECK-heavy tests have seen 1.5x speedups.
-
-// The meaning of "string" might be different between now and
-// when this macro gets invoked (e.g., if someone is experimenting
-// with other string implementations that get defined after this
-// file is included).  Save the current meaning now and use it
-// in the macro.
-typedef std::string _Check_string;
-#define CHECK_OP_LOG(name, op, val1, val2, log)                         \
-  while (google::_Check_string* _result =                \
-         google::Check##name##Impl(                      \
-             google::GetReferenceableValue(val1),        \
-             google::GetReferenceableValue(val2),        \
-             #val1 " " #op " " #val2))                                  \
-    log(__FILE__, __LINE__,                                             \
-        google::CheckOpString(_result)).stream()
-#else
-// In optimized mode, use CheckOpString to hint to compiler that
-// the while condition is unlikely.
-#define CHECK_OP_LOG(name, op, val1, val2, log)                         \
-  while (google::CheckOpString _result =                 \
-         google::Check##name##Impl(                      \
-             google::GetReferenceableValue(val1),        \
-             google::GetReferenceableValue(val2),        \
-             #val1 " " #op " " #val2))                                  \
-    log(__FILE__, __LINE__, _result).stream()
-#endif  // STATIC_ANALYSIS, DCHECK_IS_ON()
-
-#if GOOGLE_STRIP_LOG <= 3
-#define CHECK_OP(name, op, val1, val2) \
-  CHECK_OP_LOG(name, op, val1, val2, google::LogMessageFatal)
-#else
-#define CHECK_OP(name, op, val1, val2) \
-  CHECK_OP_LOG(name, op, val1, val2, google::NullStreamFatal)
-#endif // STRIP_LOG <= 3
-
-// Equality/Inequality checks - compare two values, and log a FATAL message
-// including the two values when the result is not as expected.  The values
-// must have operator<<(ostream, ...) defined.
-//
-// You may append to the error message like so:
-//   CHECK_NE(1, 2) << ": The world must be ending!";
-//
-// We are very careful to ensure that each argument is evaluated exactly
-// once, and that anything which is legal to pass as a function argument is
-// legal here.  In particular, the arguments may be temporary expressions
-// which will end up being destroyed at the end of the apparent statement,
-// for example:
-//   CHECK_EQ(string("abc")[1], 'b');
-//
-// WARNING: These don't compile correctly if one of the arguments is a pointer
-// and the other is NULL. To work around this, simply static_cast NULL to the
-// type of the desired pointer.
-
-#define CHECK_EQ(val1, val2) CHECK_OP(_EQ, ==, val1, val2)
-#define CHECK_NE(val1, val2) CHECK_OP(_NE, !=, val1, val2)
-#define CHECK_LE(val1, val2) CHECK_OP(_LE, <=, val1, val2)
-#define CHECK_LT(val1, val2) CHECK_OP(_LT, < , val1, val2)
-#define CHECK_GE(val1, val2) CHECK_OP(_GE, >=, val1, val2)
-#define CHECK_GT(val1, val2) CHECK_OP(_GT, > , val1, val2)
-
-// Check that the input is non NULL.  This very useful in constructor
-// initializer lists.
-
-#define CHECK_NOTNULL(val) \
-  google::CheckNotNull(__FILE__, __LINE__, "'" #val "' Must be non NULL", (val))
-
-// Helper functions for string comparisons.
-// To avoid bloat, the definitions are in logging.cc.
-#define DECLARE_CHECK_STROP_IMPL(func, expected) \
-  GOOGLE_GLOG_DLL_DECL std::string* Check##func##expected##Impl( \
-      const char* s1, const char* s2, const char* names);
-DECLARE_CHECK_STROP_IMPL(strcmp, true)
-DECLARE_CHECK_STROP_IMPL(strcmp, false)
-DECLARE_CHECK_STROP_IMPL(strcasecmp, true)
-DECLARE_CHECK_STROP_IMPL(strcasecmp, false)
-#undef DECLARE_CHECK_STROP_IMPL
-
-// Helper macro for string comparisons.
-// Don't use this macro directly in your code, use CHECK_STREQ et al below.
-#define CHECK_STROP(func, op, expected, s1, s2) \
-  while (google::CheckOpString _result = \
-         google::Check##func##expected##Impl((s1), (s2), \
-                                     #s1 " " #op " " #s2)) \
-    LOG(FATAL) << *_result.str_
-
-
-// String (char*) equality/inequality checks.
-// CASE versions are case-insensitive.
-//
-// Note that "s1" and "s2" may be temporary strings which are destroyed
-// by the compiler at the end of the current "full expression"
-// (e.g. CHECK_STREQ(Foo().c_str(), Bar().c_str())).
-
-#define CHECK_STREQ(s1, s2) CHECK_STROP(strcmp, ==, true, s1, s2)
-#define CHECK_STRNE(s1, s2) CHECK_STROP(strcmp, !=, false, s1, s2)
-#define CHECK_STRCASEEQ(s1, s2) CHECK_STROP(strcasecmp, ==, true, s1, s2)
-#define CHECK_STRCASENE(s1, s2) CHECK_STROP(strcasecmp, !=, false, s1, s2)
-
-#define CHECK_INDEX(I,A) CHECK(I < (sizeof(A)/sizeof(A[0])))
-#define CHECK_BOUND(B,A) CHECK(B <= (sizeof(A)/sizeof(A[0])))
-
-#define CHECK_DOUBLE_EQ(val1, val2)              \
-  do {                                           \
-    CHECK_LE((val1), (val2)+0.000000000000001L); \
-    CHECK_GE((val1), (val2)-0.000000000000001L); \
-  } while (0)
-
-#define CHECK_NEAR(val1, val2, margin)           \
-  do {                                           \
-    CHECK_LE((val1), (val2)+(margin));           \
-    CHECK_GE((val1), (val2)-(margin));           \
-  } while (0)
-
-// perror()..googly style!
-//
-// PLOG() and PLOG_IF() and PCHECK() behave exactly like their LOG* and
-// CHECK equivalents with the addition that they postpend a description
-// of the current state of errno to their output lines.
-
-#define PLOG(severity) GOOGLE_PLOG(severity, 0).stream()
-
-#define GOOGLE_PLOG(severity, counter)  \
-  google::ErrnoLogMessage( \
-      __FILE__, __LINE__, google::GLOG_ ## severity, counter, \
-      &google::LogMessage::SendToLog)
-
-#define PLOG_IF(severity, condition) \
-  static_cast<void>(0),              \
-  !(condition) ? (void) 0 : google::LogMessageVoidify() & PLOG(severity)
-
-// A CHECK() macro that postpends errno if the condition is false. E.g.
-//
-// if (poll(fds, nfds, timeout) == -1) { PCHECK(errno == EINTR); ... }
-#define PCHECK(condition)  \
-      PLOG_IF(FATAL, GOOGLE_PREDICT_BRANCH_NOT_TAKEN(!(condition))) \
-              << "Check failed: " #condition " "
-
-// A CHECK() macro that lets you assert the success of a function that
-// returns -1 and sets errno in case of an error. E.g.
-//
-// CHECK_ERR(mkdir(path, 0700));
-//
-// or
-//
-// int fd = open(filename, flags); CHECK_ERR(fd) << ": open " << filename;
-#define CHECK_ERR(invocation)                                          \
-PLOG_IF(FATAL, GOOGLE_PREDICT_BRANCH_NOT_TAKEN((invocation) == -1))    \
-        << #invocation
-
-// Use macro expansion to create, for each use of LOG_EVERY_N(), static
-// variables with the __LINE__ expansion as part of the variable name.
-#define LOG_EVERY_N_VARNAME(base, line) LOG_EVERY_N_VARNAME_CONCAT(base, line)
-#define LOG_EVERY_N_VARNAME_CONCAT(base, line) base ## line
-
-#define LOG_OCCURRENCES LOG_EVERY_N_VARNAME(occurrences_, __LINE__)
-#define LOG_OCCURRENCES_MOD_N LOG_EVERY_N_VARNAME(occurrences_mod_n_, __LINE__)
-
-#define SOME_KIND_OF_LOG_EVERY_N(severity, n, what_to_do) \
-  static int LOG_OCCURRENCES = 0, LOG_OCCURRENCES_MOD_N = 0; \
-  ++LOG_OCCURRENCES; \
-  if (++LOG_OCCURRENCES_MOD_N > n) LOG_OCCURRENCES_MOD_N -= n; \
-  if (LOG_OCCURRENCES_MOD_N == 1) \
-    google::LogMessage( \
-        __FILE__, __LINE__, google::GLOG_ ## severity, LOG_OCCURRENCES, \
-        &what_to_do).stream()
-
-#define SOME_KIND_OF_LOG_IF_EVERY_N(severity, condition, n, what_to_do) \
-  static int LOG_OCCURRENCES = 0, LOG_OCCURRENCES_MOD_N = 0; \
-  ++LOG_OCCURRENCES; \
-  if (condition && \
-      ((LOG_OCCURRENCES_MOD_N=(LOG_OCCURRENCES_MOD_N + 1) % n) == (1 % n))) \
-    google::LogMessage( \
-        __FILE__, __LINE__, google::GLOG_ ## severity, LOG_OCCURRENCES, \
-                 &what_to_do).stream()
-
-#define SOME_KIND_OF_PLOG_EVERY_N(severity, n, what_to_do) \
-  static int LOG_OCCURRENCES = 0, LOG_OCCURRENCES_MOD_N = 0; \
-  ++LOG_OCCURRENCES; \
-  if (++LOG_OCCURRENCES_MOD_N > n) LOG_OCCURRENCES_MOD_N -= n; \
-  if (LOG_OCCURRENCES_MOD_N == 1) \
-    google::ErrnoLogMessage( \
-        __FILE__, __LINE__, google::GLOG_ ## severity, LOG_OCCURRENCES, \
-        &what_to_do).stream()
-
-#define SOME_KIND_OF_LOG_FIRST_N(severity, n, what_to_do) \
-  static int LOG_OCCURRENCES = 0; \
-  if (LOG_OCCURRENCES <= n) \
-    ++LOG_OCCURRENCES; \
-  if (LOG_OCCURRENCES <= n) \
-    google::LogMessage( \
-        __FILE__, __LINE__, google::GLOG_ ## severity, LOG_OCCURRENCES, \
-        &what_to_do).stream()
-
-namespace glog_internal_namespace_ {
-template <bool>
-struct CompileAssert {
-};
-struct CrashReason;
-
-// Returns true if FailureSignalHandler is installed.
-// Needs to be exported since it's used by the signalhandler_unittest.
-GOOGLE_GLOG_DLL_DECL bool IsFailureSignalHandlerInstalled();
-}  // namespace glog_internal_namespace_
-
-#define LOG_EVERY_N(severity, n)                                        \
-  SOME_KIND_OF_LOG_EVERY_N(severity, (n), google::LogMessage::SendToLog)
-
-#define SYSLOG_EVERY_N(severity, n) \
-  SOME_KIND_OF_LOG_EVERY_N(severity, (n), google::LogMessage::SendToSyslogAndLog)
-
-#define PLOG_EVERY_N(severity, n) \
-  SOME_KIND_OF_PLOG_EVERY_N(severity, (n), google::LogMessage::SendToLog)
-
-#define LOG_FIRST_N(severity, n) \
-  SOME_KIND_OF_LOG_FIRST_N(severity, (n), google::LogMessage::SendToLog)
-
-#define LOG_IF_EVERY_N(severity, condition, n) \
-  SOME_KIND_OF_LOG_IF_EVERY_N(severity, (condition), (n), google::LogMessage::SendToLog)
-
-// We want the special COUNTER value available for LOG_EVERY_X()'ed messages
-enum PRIVATE_Counter {COUNTER};
-
-#ifdef GLOG_NO_ABBREVIATED_SEVERITIES
-// wingdi.h defines ERROR to be 0. When we call LOG(ERROR), it gets
-// substituted with 0, and it expands to COMPACT_GOOGLE_LOG_0. To allow us
-// to keep using this syntax, we define this macro to do the same thing
-// as COMPACT_GOOGLE_LOG_ERROR.
-#define COMPACT_GOOGLE_LOG_0 COMPACT_GOOGLE_LOG_ERROR
-#define SYSLOG_0 SYSLOG_ERROR
-#define LOG_TO_STRING_0 LOG_TO_STRING_ERROR
-// Needed for LOG_IS_ON(ERROR).
-const LogSeverity GLOG_0 = GLOG_ERROR;
-#else
-// Users may include windows.h after logging.h without
-// GLOG_NO_ABBREVIATED_SEVERITIES nor WIN32_LEAN_AND_MEAN.
-// For this case, we cannot detect if ERROR is defined before users
-// actually use ERROR. Let's make an undefined symbol to warn users.
-# define GLOG_ERROR_MSG ERROR_macro_is_defined_Define_GLOG_NO_ABBREVIATED_SEVERITIES_before_including_logging_h_See_the_document_for_detail
-# define COMPACT_GOOGLE_LOG_0 GLOG_ERROR_MSG
-# define SYSLOG_0 GLOG_ERROR_MSG
-# define LOG_TO_STRING_0 GLOG_ERROR_MSG
-# define GLOG_0 GLOG_ERROR_MSG
-#endif
-
-// Plus some debug-logging macros that get compiled to nothing for production
-
-#if DCHECK_IS_ON()
-
-#define DLOG(severity) LOG(severity)
-#define DVLOG(verboselevel) VLOG(verboselevel)
-#define DLOG_IF(severity, condition) LOG_IF(severity, condition)
-#define DLOG_EVERY_N(severity, n) LOG_EVERY_N(severity, n)
-#define DLOG_IF_EVERY_N(severity, condition, n) \
-  LOG_IF_EVERY_N(severity, condition, n)
-#define DLOG_ASSERT(condition) LOG_ASSERT(condition)
-
-// debug-only checking.  executed if DCHECK_IS_ON().
-#define DCHECK(condition) CHECK(condition)
-#define DCHECK_EQ(val1, val2) CHECK_EQ(val1, val2)
-#define DCHECK_NE(val1, val2) CHECK_NE(val1, val2)
-#define DCHECK_LE(val1, val2) CHECK_LE(val1, val2)
-#define DCHECK_LT(val1, val2) CHECK_LT(val1, val2)
-#define DCHECK_GE(val1, val2) CHECK_GE(val1, val2)
-#define DCHECK_GT(val1, val2) CHECK_GT(val1, val2)
-#define DCHECK_NOTNULL(val) CHECK_NOTNULL(val)
-#define DCHECK_STREQ(str1, str2) CHECK_STREQ(str1, str2)
-#define DCHECK_STRCASEEQ(str1, str2) CHECK_STRCASEEQ(str1, str2)
-#define DCHECK_STRNE(str1, str2) CHECK_STRNE(str1, str2)
-#define DCHECK_STRCASENE(str1, str2) CHECK_STRCASENE(str1, str2)
-
-#else  // !DCHECK_IS_ON()
-
-#define DLOG(severity)  \
-  static_cast<void>(0), \
-  true ? (void) 0 : google::LogMessageVoidify() & LOG(severity)
-
-#define DVLOG(verboselevel)             \
-  static_cast<void>(0),                 \
-  (true || !VLOG_IS_ON(verboselevel)) ? \
-      (void) 0 : google::LogMessageVoidify() & LOG(INFO)
-
-#define DLOG_IF(severity, condition) \
-  static_cast<void>(0),              \
-  (true || !(condition)) ? (void) 0 : google::LogMessageVoidify() & LOG(severity)
-
-#define DLOG_EVERY_N(severity, n) \
-  static_cast<void>(0),           \
-  true ? (void) 0 : google::LogMessageVoidify() & LOG(severity)
-
-#define DLOG_IF_EVERY_N(severity, condition, n) \
-  static_cast<void>(0),                         \
-  (true || !(condition))? (void) 0 : google::LogMessageVoidify() & LOG(severity)
-
-#define DLOG_ASSERT(condition) \
-  static_cast<void>(0),        \
-  true ? (void) 0 : LOG_ASSERT(condition)
-
-// MSVC warning C4127: conditional expression is constant
-#define DCHECK(condition) \
-  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
-  while (false) \
-    GLOG_MSVC_POP_WARNING() CHECK(condition)
-
-#define DCHECK_EQ(val1, val2) \
-  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
-  while (false) \
-    GLOG_MSVC_POP_WARNING() CHECK_EQ(val1, val2)
-
-#define DCHECK_NE(val1, val2) \
-  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
-  while (false) \
-    GLOG_MSVC_POP_WARNING() CHECK_NE(val1, val2)
-
-#define DCHECK_LE(val1, val2) \
-  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
-  while (false) \
-    GLOG_MSVC_POP_WARNING() CHECK_LE(val1, val2)
-
-#define DCHECK_LT(val1, val2) \
-  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
-  while (false) \
-    GLOG_MSVC_POP_WARNING() CHECK_LT(val1, val2)
-
-#define DCHECK_GE(val1, val2) \
-  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
-  while (false) \
-    GLOG_MSVC_POP_WARNING() CHECK_GE(val1, val2)
-
-#define DCHECK_GT(val1, val2) \
-  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
-  while (false) \
-    GLOG_MSVC_POP_WARNING() CHECK_GT(val1, val2)
-
-// You may see warnings in release mode if you don't use the return
-// value of DCHECK_NOTNULL. Please just use DCHECK for such cases.
-#define DCHECK_NOTNULL(val) (val)
-
-#define DCHECK_STREQ(str1, str2) \
-  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
-  while (false) \
-    GLOG_MSVC_POP_WARNING() CHECK_STREQ(str1, str2)
-
-#define DCHECK_STRCASEEQ(str1, str2) \
-  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
-  while (false) \
-    GLOG_MSVC_POP_WARNING() CHECK_STRCASEEQ(str1, str2)
-
-#define DCHECK_STRNE(str1, str2) \
-  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
-  while (false) \
-    GLOG_MSVC_POP_WARNING() CHECK_STRNE(str1, str2)
-
-#define DCHECK_STRCASENE(str1, str2) \
-  GLOG_MSVC_PUSH_DISABLE_WARNING(4127) \
-  while (false) \
-    GLOG_MSVC_POP_WARNING() CHECK_STRCASENE(str1, str2)
-
-#endif  // DCHECK_IS_ON()
-
-// Log only in verbose mode.
-
-#define VLOG(verboselevel) LOG_IF(INFO, VLOG_IS_ON(verboselevel))
-
-#define VLOG_IF(verboselevel, condition) \
-  LOG_IF(INFO, (condition) && VLOG_IS_ON(verboselevel))
-
-#define VLOG_EVERY_N(verboselevel, n) \
-  LOG_IF_EVERY_N(INFO, VLOG_IS_ON(verboselevel), n)
-
-#define VLOG_IF_EVERY_N(verboselevel, condition, n) \
-  LOG_IF_EVERY_N(INFO, (condition) && VLOG_IS_ON(verboselevel), n)
-
-namespace base_logging {
-
-// LogMessage::LogStream is a std::ostream backed by this streambuf.
-// This class ignores overflow and leaves two bytes at the end of the
-// buffer to allow for a '\n' and '\0'.
-class GOOGLE_GLOG_DLL_DECL LogStreamBuf : public std::streambuf {
- public:
-  // REQUIREMENTS: "len" must be >= 2 to account for the '\n' and '\0'.
-  LogStreamBuf(char *buf, int len) {
-    setp(buf, buf + len - 2);
-  }
-
-  // This effectively ignores overflow.
-  virtual int_type overflow(int_type ch) {
-    return ch;
-  }
-
-  // Legacy public ostrstream method.
-  size_t pcount() const { return pptr() - pbase(); }
-  char* pbase() const { return std::streambuf::pbase(); }
-};
-
-}  // namespace base_logging
-
-//
-// This class more or less represents a particular log message.  You
-// create an instance of LogMessage and then stream stuff to it.
-// When you finish streaming to it, ~LogMessage is called and the
-// full message gets streamed to the appropriate destination.
-//
-// You shouldn't actually use LogMessage's constructor to log things,
-// though.  You should use the LOG() macro (and variants thereof)
-// above.
-class GOOGLE_GLOG_DLL_DECL LogMessage {
-public:
-  enum {
-    // Passing kNoLogPrefix for the line number disables the
-    // log-message prefix. Useful for using the LogMessage
-    // infrastructure as a printing utility. See also the --log_prefix
-    // flag for controlling the log-message prefix on an
-    // application-wide basis.
-    kNoLogPrefix = -1
-  };
-
-  // LogStream inherit from non-DLL-exported class (std::ostrstream)
-  // and VC++ produces a warning for this situation.
-  // However, MSDN says "C4275 can be ignored in Microsoft Visual C++
-  // 2005 if you are deriving from a type in the Standard C++ Library"
-  // http://msdn.microsoft.com/en-us/library/3tdb471s(VS.80).aspx
-  // Let's just ignore the warning.
-GLOG_MSVC_PUSH_DISABLE_WARNING(4275)
-  class GOOGLE_GLOG_DLL_DECL LogStream : public std::ostream {
-GLOG_MSVC_POP_WARNING()
-  public:
-    LogStream(char *buf, int len, int ctr)
-        : std::ostream(NULL),
-          streambuf_(buf, len),
-          ctr_(ctr),
-          self_(this) {
-      rdbuf(&streambuf_);
-    }
-
-    int ctr() const { return ctr_; }
-    void set_ctr(int ctr) { ctr_ = ctr; }
-    LogStream* self() const { return self_; }
-
-    // Legacy std::streambuf methods.
-    size_t pcount() const { return streambuf_.pcount(); }
-    char* pbase() const { return streambuf_.pbase(); }
-    char* str() const { return pbase(); }
-
-  private:
-    LogStream(const LogStream&);
-    LogStream& operator=(const LogStream&);
-    base_logging::LogStreamBuf streambuf_;
-    int ctr_;  // Counter hack (for the LOG_EVERY_X() macro)
-    LogStream *self_;  // Consistency check hack
-  };
-
-public:
-  // icc 8 requires this typedef to avoid an internal compiler error.
-  typedef void (LogMessage::*SendMethod)();
-
-  LogMessage(const char* file, int line, LogSeverity severity, int ctr,
-             SendMethod send_method);
-
-  // Two special constructors that generate reduced amounts of code at
-  // LOG call sites for common cases.
-
-  // Used for LOG(INFO): Implied are:
-  // severity = INFO, ctr = 0, send_method = &LogMessage::SendToLog.
-  //
-  // Using this constructor instead of the more complex constructor above
-  // saves 19 bytes per call site.
-  LogMessage(const char* file, int line);
-
-  // Used for LOG(severity) where severity != INFO.  Implied
-  // are: ctr = 0, send_method = &LogMessage::SendToLog
-  //
-  // Using this constructor instead of the more complex constructor above
-  // saves 17 bytes per call site.
-  LogMessage(const char* file, int line, LogSeverity severity);
-
-  // Constructor to log this message to a specified sink (if not NULL).
-  // Implied are: ctr = 0, send_method = &LogMessage::SendToSinkAndLog if
-  // also_send_to_log is true, send_method = &LogMessage::SendToSink otherwise.
-  LogMessage(const char* file, int line, LogSeverity severity, LogSink* sink,
-             bool also_send_to_log);
-
-  // Constructor where we also give a vector<string> pointer
-  // for storing the messages (if the pointer is not NULL).
-  // Implied are: ctr = 0, send_method = &LogMessage::SaveOrSendToLog.
-  LogMessage(const char* file, int line, LogSeverity severity,
-             std::vector<std::string>* outvec);
-
-  // Constructor where we also give a string pointer for storing the
-  // message (if the pointer is not NULL).  Implied are: ctr = 0,
-  // send_method = &LogMessage::WriteToStringAndLog.
-  LogMessage(const char* file, int line, LogSeverity severity,
-             std::string* message);
-
-  // A special constructor used for check failures
-  LogMessage(const char* file, int line, const CheckOpString& result);
-
-  ~LogMessage();
-
-  // Flush a buffered message to the sink set in the constructor.  Always
-  // called by the destructor, it may also be called from elsewhere if
-  // needed.  Only the first call is actioned; any later ones are ignored.
-  void Flush();
-
-  // An arbitrary limit on the length of a single log message.  This
-  // is so that streaming can be done more efficiently.
-  static const size_t kMaxLogMessageLen;
-
-  // Theses should not be called directly outside of logging.*,
-  // only passed as SendMethod arguments to other LogMessage methods:
-  void SendToLog();  // Actually dispatch to the logs
-  void SendToSyslogAndLog();  // Actually dispatch to syslog and the logs
-
-  // Call abort() or similar to perform LOG(FATAL) crash.
-  static void __attribute__((noreturn)) Fail();
-
-  std::ostream& stream();
-
-  int preserved_errno() const;
-
-  // Must be called without the log_mutex held.  (L < log_mutex)
-  static int64 num_messages(int severity);
-
-  struct LogMessageData;
-
-private:
-  // Fully internal SendMethod cases:
-  void SendToSinkAndLog();  // Send to sink if provided and dispatch to the logs
-  void SendToSink();  // Send to sink if provided, do nothing otherwise.
-
-  // Write to string if provided and dispatch to the logs.
-  void WriteToStringAndLog();
-
-  void SaveOrSendToLog();  // Save to stringvec if provided, else to logs
-
-  void Init(const char* file, int line, LogSeverity severity,
-            void (LogMessage::*send_method)());
-
-  // Used to fill in crash information during LOG(FATAL) failures.
-  void RecordCrashReason(glog_internal_namespace_::CrashReason* reason);
-
-  // Counts of messages sent at each priority:
-  static int64 num_messages_[NUM_SEVERITIES];  // under log_mutex
-
-  // We keep the data in a separate struct so that each instance of
-  // LogMessage uses less stack space.
-  LogMessageData* allocated_;
-  LogMessageData* data_;
-
-  friend class LogDestination;
-
-  LogMessage(const LogMessage&);
-  void operator=(const LogMessage&);
-};
-
-// This class happens to be thread-hostile because all instances share
-// a single data buffer, but since it can only be created just before
-// the process dies, we don't worry so much.
-class GOOGLE_GLOG_DLL_DECL LogMessageFatal : public LogMessage {
- public:
-  LogMessageFatal(const char* file, int line);
-  LogMessageFatal(const char* file, int line, const CheckOpString& result);
-  __attribute__((noreturn)) ~LogMessageFatal();
-};
-
-// A non-macro interface to the log facility; (useful
-// when the logging level is not a compile-time constant).
-inline void LogAtLevel(int const severity, std::string const &msg) {
-  LogMessage(__FILE__, __LINE__, severity).stream() << msg;
-}
-
-// A macro alternative of LogAtLevel. New code may want to use this
-// version since there are two advantages: 1. this version outputs the
-// file name and the line number where this macro is put like other
-// LOG macros, 2. this macro can be used as C++ stream.
-#define LOG_AT_LEVEL(severity) google::LogMessage(__FILE__, __LINE__, severity).stream()
-
-// Check if it's compiled in C++11 mode.
-//
-// GXX_EXPERIMENTAL_CXX0X is defined by gcc and clang up to at least
-// gcc-4.7 and clang-3.1 (2011-12-13).  __cplusplus was defined to 1
-// in gcc before 4.7 (Crosstool 16) and clang before 3.1, but is
-// defined according to the language version in effect thereafter.
-// Microsoft Visual Studio 14 (2015) sets __cplusplus==199711 despite
-// reasonably good C++11 support, so we set LANG_CXX for it and
-// newer versions (_MSC_VER >= 1900).
-#if (defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103L || \
-     (defined(_MSC_VER) && _MSC_VER >= 1900))
-// Helper for CHECK_NOTNULL().
-//
-// In C++11, all cases can be handled by a single function. Since the value
-// category of the argument is preserved (also for rvalue references),
-// member initializer lists like the one below will compile correctly:
-//
-//   Foo()
-//     : x_(CHECK_NOTNULL(MethodReturningUniquePtr())) {}
-template <typename T>
-T CheckNotNull(const char* file, int line, const char* names, T&& t) {
- if (t == nullptr) {
-   LogMessageFatal(file, line, new std::string(names));
- }
- return std::forward<T>(t);
-}
-
-#else
-
-// A small helper for CHECK_NOTNULL().
-template <typename T>
-T* CheckNotNull(const char *file, int line, const char *names, T* t) {
-  if (t == NULL) {
-    LogMessageFatal(file, line, new std::string(names));
-  }
-  return t;
-}
-#endif
-
-// Allow folks to put a counter in the LOG_EVERY_X()'ed messages. This
-// only works if ostream is a LogStream. If the ostream is not a
-// LogStream you'll get an assert saying as much at runtime.
-GOOGLE_GLOG_DLL_DECL std::ostream& operator<<(std::ostream &os,
-                                              const PRIVATE_Counter&);
-
-
-// Derived class for PLOG*() above.
-class GOOGLE_GLOG_DLL_DECL ErrnoLogMessage : public LogMessage {
- public:
-
-  ErrnoLogMessage(const char* file, int line, LogSeverity severity, int ctr,
-                  void (LogMessage::*send_method)());
-
-  // Postpends ": strerror(errno) [errno]".
-  ~ErrnoLogMessage();
-
- private:
-  ErrnoLogMessage(const ErrnoLogMessage&);
-  void operator=(const ErrnoLogMessage&);
-};
-
-
-// This class is used to explicitly ignore values in the conditional
-// logging macros.  This avoids compiler warnings like "value computed
-// is not used" and "statement has no effect".
-
-class GOOGLE_GLOG_DLL_DECL LogMessageVoidify {
- public:
-  LogMessageVoidify() { }
-  // This has to be an operator with a precedence lower than << but
-  // higher than ?:
-  void operator&(std::ostream&) { }
-};
-
-
-// Flushes all log files that contains messages that are at least of
-// the specified severity level.  Thread-safe.
-GOOGLE_GLOG_DLL_DECL void FlushLogFiles(LogSeverity min_severity);
-
-// Flushes all log files that contains messages that are at least of
-// the specified severity level. Thread-hostile because it ignores
-// locking -- used for catastrophic failures.
-GOOGLE_GLOG_DLL_DECL void FlushLogFilesUnsafe(LogSeverity min_severity);
-
-//
-// Set the destination to which a particular severity level of log
-// messages is sent.  If base_filename is "", it means "don't log this
-// severity".  Thread-safe.
-//
-GOOGLE_GLOG_DLL_DECL void SetLogDestination(LogSeverity severity,
-                                            const char* base_filename);
-
-//
-// Set the basename of the symlink to the latest log file at a given
-// severity.  If symlink_basename is empty, do not make a symlink.  If
-// you don't call this function, the symlink basename is the
-// invocation name of the program.  Thread-safe.
-//
-GOOGLE_GLOG_DLL_DECL void SetLogSymlink(LogSeverity severity,
-                                        const char* symlink_basename);
-
-//
-// Used to send logs to some other kind of destination
-// Users should subclass LogSink and override send to do whatever they want.
-// Implementations must be thread-safe because a shared instance will
-// be called from whichever thread ran the LOG(XXX) line.
-class GOOGLE_GLOG_DLL_DECL LogSink {
- public:
-  virtual ~LogSink();
-
-  // Sink's logging logic (message_len is such as to exclude '\n' at the end).
-  // This method can't use LOG() or CHECK() as logging system mutex(s) are held
-  // during this call.
-  virtual void send(LogSeverity severity, const char* full_filename,
-                    const char* base_filename, int line,
-                    const struct ::tm* tm_time,
-                    const char* message, size_t message_len) = 0;
-
-  // Redefine this to implement waiting for
-  // the sink's logging logic to complete.
-  // It will be called after each send() returns,
-  // but before that LogMessage exits or crashes.
-  // By default this function does nothing.
-  // Using this function one can implement complex logic for send()
-  // that itself involves logging; and do all this w/o causing deadlocks and
-  // inconsistent rearrangement of log messages.
-  // E.g. if a LogSink has thread-specific actions, the send() method
-  // can simply add the message to a queue and wake up another thread that
-  // handles real logging while itself making some LOG() calls;
-  // WaitTillSent() can be implemented to wait for that logic to complete.
-  // See our unittest for an example.
-  virtual void WaitTillSent();
-
-  // Returns the normal text output of the log message.
-  // Can be useful to implement send().
-  static std::string ToString(LogSeverity severity, const char* file, int line,
-                              const struct ::tm* tm_time,
-                              const char* message, size_t message_len);
-};
-
-// Add or remove a LogSink as a consumer of logging data.  Thread-safe.
-GOOGLE_GLOG_DLL_DECL void AddLogSink(LogSink *destination);
-GOOGLE_GLOG_DLL_DECL void RemoveLogSink(LogSink *destination);
-
-//
-// Specify an "extension" added to the filename specified via
-// SetLogDestination.  This applies to all severity levels.  It's
-// often used to append the port we're listening on to the logfile
-// name.  Thread-safe.
-//
-GOOGLE_GLOG_DLL_DECL void SetLogFilenameExtension(
-    const char* filename_extension);
-
-//
-// Make it so that all log messages of at least a particular severity
-// are logged to stderr (in addition to logging to the usual log
-// file(s)).  Thread-safe.
-//
-GOOGLE_GLOG_DLL_DECL void SetStderrLogging(LogSeverity min_severity);
-
-//
-// Make it so that all log messages go only to stderr.  Thread-safe.
-//
-GOOGLE_GLOG_DLL_DECL void LogToStderr();
-
-//
-// Make it so that all log messages of at least a particular severity are
-// logged via email to a list of addresses (in addition to logging to the
-// usual log file(s)).  The list of addresses is just a string containing
-// the email addresses to send to (separated by spaces, say).  Thread-safe.
-//
-GOOGLE_GLOG_DLL_DECL void SetEmailLogging(LogSeverity min_severity,
-                                          const char* addresses);
-
-// A simple function that sends email. dest is a commma-separated
-// list of addressess.  Thread-safe.
-GOOGLE_GLOG_DLL_DECL bool SendEmail(const char *dest,
-                                    const char *subject, const char *body);
-
-GOOGLE_GLOG_DLL_DECL const std::vector<std::string>& GetLoggingDirectories();
-
-// For tests only:  Clear the internal [cached] list of logging directories to
-// force a refresh the next time GetLoggingDirectories is called.
-// Thread-hostile.
-void TestOnly_ClearLoggingDirectoriesList();
-
-// Returns a set of existing temporary directories, which will be a
-// subset of the directories returned by GetLogginDirectories().
-// Thread-safe.
-GOOGLE_GLOG_DLL_DECL void GetExistingTempDirectories(
-    std::vector<std::string>* list);
-
-// Print any fatal message again -- useful to call from signal handler
-// so that the last thing in the output is the fatal message.
-// Thread-hostile, but a race is unlikely.
-GOOGLE_GLOG_DLL_DECL void ReprintFatalMessage();
-
-// Truncate a log file that may be the append-only output of multiple
-// processes and hence can't simply be renamed/reopened (typically a
-// stdout/stderr).  If the file "path" is > "limit" bytes, copy the
-// last "keep" bytes to offset 0 and truncate the rest. Since we could
-// be racing with other writers, this approach has the potential to
-// lose very small amounts of data. For security, only follow symlinks
-// if the path is /proc/self/fd/*
-GOOGLE_GLOG_DLL_DECL void TruncateLogFile(const char *path,
-                                          int64 limit, int64 keep);
-
-// Truncate stdout and stderr if they are over the value specified by
-// --max_log_size; keep the final 1MB.  This function has the same
-// race condition as TruncateLogFile.
-GOOGLE_GLOG_DLL_DECL void TruncateStdoutStderr();
-
-// Return the string representation of the provided LogSeverity level.
-// Thread-safe.
-GOOGLE_GLOG_DLL_DECL const char* GetLogSeverityName(LogSeverity severity);
-
-// ---------------------------------------------------------------------
-// Implementation details that are not useful to most clients
-// ---------------------------------------------------------------------
-
-// A Logger is the interface used by logging modules to emit entries
-// to a log.  A typical implementation will dump formatted data to a
-// sequence of files.  We also provide interfaces that will forward
-// the data to another thread so that the invoker never blocks.
-// Implementations should be thread-safe since the logging system
-// will write to them from multiple threads.
-
-namespace base {
-
-class GOOGLE_GLOG_DLL_DECL Logger {
- public:
-  virtual ~Logger();
-
-  // Writes "message[0,message_len-1]" corresponding to an event that
-  // occurred at "timestamp".  If "force_flush" is true, the log file
-  // is flushed immediately.
-  //
-  // The input message has already been formatted as deemed
-  // appropriate by the higher level logging facility.  For example,
-  // textual log messages already contain timestamps, and the
-  // file:linenumber header.
-  virtual void Write(bool force_flush,
-                     time_t timestamp,
-                     const char* message,
-                     int message_len) = 0;
-
-  // Flush any buffered messages
-  virtual void Flush() = 0;
-
-  // Get the current LOG file size.
-  // The returned value is approximate since some
-  // logged data may not have been flushed to disk yet.
-  virtual uint32 LogSize() = 0;
-};
-
-// Get the logger for the specified severity level.  The logger
-// remains the property of the logging module and should not be
-// deleted by the caller.  Thread-safe.
-extern GOOGLE_GLOG_DLL_DECL Logger* GetLogger(LogSeverity level);
-
-// Set the logger for the specified severity level.  The logger
-// becomes the property of the logging module and should not
-// be deleted by the caller.  Thread-safe.
-extern GOOGLE_GLOG_DLL_DECL void SetLogger(LogSeverity level, Logger* logger);
-
-}
-
-// glibc has traditionally implemented two incompatible versions of
-// strerror_r(). There is a poorly defined convention for picking the
-// version that we want, but it is not clear whether it even works with
-// all versions of glibc.
-// So, instead, we provide this wrapper that automatically detects the
-// version that is in use, and then implements POSIX semantics.
-// N.B. In addition to what POSIX says, we also guarantee that "buf" will
-// be set to an empty string, if this function failed. This means, in most
-// cases, you do not need to check the error code and you can directly
-// use the value of "buf". It will never have an undefined value.
-// DEPRECATED: Use StrError(int) instead.
-GOOGLE_GLOG_DLL_DECL int posix_strerror_r(int err, char *buf, size_t len);
-
-// A thread-safe replacement for strerror(). Returns a string describing the
-// given POSIX error code.
-GOOGLE_GLOG_DLL_DECL std::string StrError(int err);
-
-// A class for which we define operator<<, which does nothing.
-class GOOGLE_GLOG_DLL_DECL NullStream : public LogMessage::LogStream {
- public:
-  // Initialize the LogStream so the messages can be written somewhere
-  // (they'll never be actually displayed). This will be needed if a
-  // NullStream& is implicitly converted to LogStream&, in which case
-  // the overloaded NullStream::operator<< will not be invoked.
-  NullStream() : LogMessage::LogStream(message_buffer_, 1, 0) { }
-  NullStream(const char* /*file*/, int /*line*/,
-             const CheckOpString& /*result*/) :
-      LogMessage::LogStream(message_buffer_, 1, 0) { }
-  NullStream &stream() { return *this; }
- private:
-  // A very short buffer for messages (which we discard anyway). This
-  // will be needed if NullStream& converted to LogStream& (e.g. as a
-  // result of a conditional expression).
-  char message_buffer_[2];
-};
-
-// Do nothing. This operator is inline, allowing the message to be
-// compiled away. The message will not be compiled away if we do
-// something like (flag ? LOG(INFO) : LOG(ERROR)) << message; when
-// SKIP_LOG=WARNING. In those cases, NullStream will be implicitly
-// converted to LogStream and the message will be computed and then
-// quietly discarded.
-template<class T>
-inline NullStream& operator<<(NullStream &str, const T &) { return str; }
-
-// Similar to NullStream, but aborts the program (without stack
-// trace), like LogMessageFatal.
-class GOOGLE_GLOG_DLL_DECL NullStreamFatal : public NullStream {
- public:
-  NullStreamFatal() { }
-  NullStreamFatal(const char* file, int line, const CheckOpString& result) :
-      NullStream(file, line, result) { }
-  __attribute__((noreturn)) ~NullStreamFatal() throw () { _exit(1); }
-};
-
-// Install a signal handler that will dump signal information and a stack
-// trace when the program crashes on certain signals.  We'll install the
-// signal handler for the following signals.
-//
-// SIGSEGV, SIGILL, SIGFPE, SIGABRT, SIGBUS, and SIGTERM.
-//
-// By default, the signal handler will write the failure dump to the
-// standard error.  You can customize the destination by installing your
-// own writer function by InstallFailureWriter() below.
-//
-// Note on threading:
-//
-// The function should be called before threads are created, if you want
-// to use the failure signal handler for all threads.  The stack trace
-// will be shown only for the thread that receives the signal.  In other
-// words, stack traces of other threads won't be shown.
-GOOGLE_GLOG_DLL_DECL void InstallFailureSignalHandler();
-
-// Installs a function that is used for writing the failure dump.  "data"
-// is the pointer to the beginning of a message to be written, and "size"
-// is the size of the message.  You should not expect the data is
-// terminated with '\0'.
-GOOGLE_GLOG_DLL_DECL void InstallFailureWriter(
-    void (*writer)(const char* data, int size));
-
-}
-
-#endif // _LOGGING_H_
diff --git a/projects/llm_framework/include/glog/raw_logging.h b/projects/llm_framework/include/glog/raw_logging.h
deleted file mode 100644
index cf3f27d9..00000000
--- a/projects/llm_framework/include/glog/raw_logging.h
+++ /dev/null
@@ -1,180 +0,0 @@
-// Copyright (c) 2006, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-//
-// Author: Maxim Lifantsev
-//
-// Thread-safe logging routines that do not allocate any memory or
-// acquire any locks, and can therefore be used by low-level memory
-// allocation and synchronization code.
-
-#ifndef BASE_RAW_LOGGING_H_
-#define BASE_RAW_LOGGING_H_
-
-#include <time.h>
-
-namespace google {
-
-#include "glog/log_severity.h"
-#include "glog/vlog_is_on.h"
-
-// Annoying stuff for windows -- makes sure clients can import these functions
-#ifndef GOOGLE_GLOG_DLL_DECL
-# if defined(_WIN32) && !defined(__CYGWIN__)
-#   define GOOGLE_GLOG_DLL_DECL  __declspec(dllimport)
-# else
-#   define GOOGLE_GLOG_DLL_DECL
-# endif
-#endif
-
-// This is similar to LOG(severity) << format... and VLOG(level) << format..,
-// but
-// * it is to be used ONLY by low-level modules that can't use normal LOG()
-// * it is desiged to be a low-level logger that does not allocate any
-//   memory and does not need any locks, hence:
-// * it logs straight and ONLY to STDERR w/o buffering
-// * it uses an explicit format and arguments list
-// * it will silently chop off really long message strings
-// Usage example:
-//   RAW_LOG(ERROR, "Failed foo with %i: %s", status, error);
-//   RAW_VLOG(3, "status is %i", status);
-// These will print an almost standard log lines like this to stderr only:
-//   E0821 211317 file.cc:123] RAW: Failed foo with 22: bad_file
-//   I0821 211317 file.cc:142] RAW: status is 20
-#define RAW_LOG(severity, ...) \
-  do { \
-    switch (google::GLOG_ ## severity) {  \
-      case 0: \
-        RAW_LOG_INFO(__VA_ARGS__); \
-        break; \
-      case 1: \
-        RAW_LOG_WARNING(__VA_ARGS__); \
-        break; \
-      case 2: \
-        RAW_LOG_ERROR(__VA_ARGS__); \
-        break; \
-      case 3: \
-        RAW_LOG_FATAL(__VA_ARGS__); \
-        break; \
-      default: \
-        break; \
-    } \
-  } while (0)
-
-// The following STRIP_LOG testing is performed in the header file so that it's
-// possible to completely compile out the logging code and the log messages.
-#if STRIP_LOG == 0
-#define RAW_VLOG(verboselevel, ...) \
-  do { \
-    if (VLOG_IS_ON(verboselevel)) { \
-      RAW_LOG_INFO(__VA_ARGS__); \
-    } \
-  } while (0)
-#else
-#define RAW_VLOG(verboselevel, ...) RawLogStub__(0, __VA_ARGS__)
-#endif // STRIP_LOG == 0
-
-#if STRIP_LOG == 0
-#define RAW_LOG_INFO(...) google::RawLog__(google::GLOG_INFO, \
-                                   __FILE__, __LINE__, __VA_ARGS__)
-#else
-#define RAW_LOG_INFO(...) google::RawLogStub__(0, __VA_ARGS__)
-#endif // STRIP_LOG == 0
-
-#if STRIP_LOG <= 1
-#define RAW_LOG_WARNING(...) google::RawLog__(google::GLOG_WARNING,   \
-                                      __FILE__, __LINE__, __VA_ARGS__)
-#else
-#define RAW_LOG_WARNING(...) google::RawLogStub__(0, __VA_ARGS__)
-#endif // STRIP_LOG <= 1
-
-#if STRIP_LOG <= 2
-#define RAW_LOG_ERROR(...) google::RawLog__(google::GLOG_ERROR,       \
-                                    __FILE__, __LINE__, __VA_ARGS__)
-#else
-#define RAW_LOG_ERROR(...) google::RawLogStub__(0, __VA_ARGS__)
-#endif // STRIP_LOG <= 2
-
-#if STRIP_LOG <= 3
-#define RAW_LOG_FATAL(...) google::RawLog__(google::GLOG_FATAL,       \
-                                    __FILE__, __LINE__, __VA_ARGS__)
-#else
-#define RAW_LOG_FATAL(...) \
-  do { \
-    google::RawLogStub__(0, __VA_ARGS__);        \
-    exit(1); \
-  } while (0)
-#endif // STRIP_LOG <= 3
-
-// Similar to CHECK(condition) << message,
-// but for low-level modules: we use only RAW_LOG that does not allocate memory.
-// We do not want to provide args list here to encourage this usage:
-//   if (!cond)  RAW_LOG(FATAL, "foo ...", hard_to_compute_args);
-// so that the args are not computed when not needed.
-#define RAW_CHECK(condition, message)                                   \
-  do {                                                                  \
-    if (!(condition)) {                                                 \
-      RAW_LOG(FATAL, "Check %s failed: %s", #condition, message);       \
-    }                                                                   \
-  } while (0)
-
-// Debug versions of RAW_LOG and RAW_CHECK
-#ifndef NDEBUG
-
-#define RAW_DLOG(severity, ...) RAW_LOG(severity, __VA_ARGS__)
-#define RAW_DCHECK(condition, message) RAW_CHECK(condition, message)
-
-#else  // NDEBUG
-
-#define RAW_DLOG(severity, ...)                                 \
-  while (false)                                                 \
-    RAW_LOG(severity, __VA_ARGS__)
-#define RAW_DCHECK(condition, message) \
-  while (false) \
-    RAW_CHECK(condition, message)
-
-#endif  // NDEBUG
-
-// Stub log function used to work around for unused variable warnings when
-// building with STRIP_LOG > 0.
-static inline void RawLogStub__(int /* ignored */, ...) {
-}
-
-// Helper function to implement RAW_LOG and RAW_VLOG
-// Logs format... at "severity" level, reporting it
-// as called from file:line.
-// This does not allocate memory or acquire locks.
-GOOGLE_GLOG_DLL_DECL void RawLog__(LogSeverity severity,
-                                   const char* file,
-                                   int line,
-                                   const char* format, ...)
-   ;
-
-}
-
-#endif  // BASE_RAW_LOGGING_H_
diff --git a/projects/llm_framework/include/glog/stl_logging.h b/projects/llm_framework/include/glog/stl_logging.h
deleted file mode 100644
index 40a15aa4..00000000
--- a/projects/llm_framework/include/glog/stl_logging.h
+++ /dev/null
@@ -1,220 +0,0 @@
-// Copyright (c) 2003, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-//
-// Stream output operators for STL containers; to be used for logging *only*.
-// Inclusion of this file lets you do:
-//
-// list<string> x;
-// LOG(INFO) << "data: " << x;
-// vector<int> v1, v2;
-// CHECK_EQ(v1, v2);
-//
-// If you want to use this header file with hash maps or slist, you
-// need to define macros before including this file:
-//
-// - GLOG_STL_LOGGING_FOR_UNORDERED     - <unordered_map> and <unordered_set>
-// - GLOG_STL_LOGGING_FOR_TR1_UNORDERED - <tr1/unordered_(map|set)>
-// - GLOG_STL_LOGGING_FOR_EXT_HASH      - <ext/hash_(map|set)>
-// - GLOG_STL_LOGGING_FOR_EXT_SLIST     - <ext/slist>
-//
-
-#ifndef UTIL_GTL_STL_LOGGING_INL_H_
-#define UTIL_GTL_STL_LOGGING_INL_H_
-
-#if !1
-# error We do not support stl_logging for this compiler
-#endif
-
-#include <deque>
-#include <list>
-#include <map>
-#include <ostream>
-#include <set>
-#include <utility>
-#include <vector>
-
-#ifdef GLOG_STL_LOGGING_FOR_UNORDERED
-# include <unordered_map>
-# include <unordered_set>
-#endif
-
-#ifdef GLOG_STL_LOGGING_FOR_TR1_UNORDERED
-# include <tr1/unordered_map>
-# include <tr1/unordered_set>
-#endif
-
-#ifdef GLOG_STL_LOGGING_FOR_EXT_HASH
-# include <ext/hash_set>
-# include <ext/hash_map>
-#endif
-#ifdef GLOG_STL_LOGGING_FOR_EXT_SLIST
-# include <ext/slist>
-#endif
-
-// Forward declare these two, and define them after all the container streams
-// operators so that we can recurse from pair -> container -> container -> pair
-// properly.
-template<class First, class Second>
-std::ostream& operator<<(std::ostream& out, const std::pair<First, Second>& p);
-
-namespace google {
-
-template<class Iter>
-void PrintSequence(std::ostream& out, Iter begin, Iter end);
-
-}
-
-#define OUTPUT_TWO_ARG_CONTAINER(Sequence) \
-template<class T1, class T2> \
-inline std::ostream& operator<<(std::ostream& out, \
-                                const Sequence<T1, T2>& seq) { \
-  google::PrintSequence(out, seq.begin(), seq.end()); \
-  return out; \
-}
-
-OUTPUT_TWO_ARG_CONTAINER(std::vector)
-OUTPUT_TWO_ARG_CONTAINER(std::deque)
-OUTPUT_TWO_ARG_CONTAINER(std::list)
-#ifdef GLOG_STL_LOGGING_FOR_EXT_SLIST
-OUTPUT_TWO_ARG_CONTAINER(__gnu_cxx::slist)
-#endif
-
-#undef OUTPUT_TWO_ARG_CONTAINER
-
-#define OUTPUT_THREE_ARG_CONTAINER(Sequence) \
-template<class T1, class T2, class T3> \
-inline std::ostream& operator<<(std::ostream& out, \
-                                const Sequence<T1, T2, T3>& seq) { \
-  google::PrintSequence(out, seq.begin(), seq.end()); \
-  return out; \
-}
-
-OUTPUT_THREE_ARG_CONTAINER(std::set)
-OUTPUT_THREE_ARG_CONTAINER(std::multiset)
-
-#undef OUTPUT_THREE_ARG_CONTAINER
-
-#define OUTPUT_FOUR_ARG_CONTAINER(Sequence) \
-template<class T1, class T2, class T3, class T4> \
-inline std::ostream& operator<<(std::ostream& out, \
-                                const Sequence<T1, T2, T3, T4>& seq) { \
-  google::PrintSequence(out, seq.begin(), seq.end()); \
-  return out; \
-}
-
-OUTPUT_FOUR_ARG_CONTAINER(std::map)
-OUTPUT_FOUR_ARG_CONTAINER(std::multimap)
-#ifdef GLOG_STL_LOGGING_FOR_UNORDERED
-OUTPUT_FOUR_ARG_CONTAINER(std::unordered_set)
-OUTPUT_FOUR_ARG_CONTAINER(std::unordered_multiset)
-#endif
-#ifdef GLOG_STL_LOGGING_FOR_TR1_UNORDERED
-OUTPUT_FOUR_ARG_CONTAINER(std::tr1::unordered_set)
-OUTPUT_FOUR_ARG_CONTAINER(std::tr1::unordered_multiset)
-#endif
-#ifdef GLOG_STL_LOGGING_FOR_EXT_HASH
-OUTPUT_FOUR_ARG_CONTAINER(__gnu_cxx::hash_set)
-OUTPUT_FOUR_ARG_CONTAINER(__gnu_cxx::hash_multiset)
-#endif
-
-#undef OUTPUT_FOUR_ARG_CONTAINER
-
-#define OUTPUT_FIVE_ARG_CONTAINER(Sequence) \
-template<class T1, class T2, class T3, class T4, class T5> \
-inline std::ostream& operator<<(std::ostream& out, \
-                                const Sequence<T1, T2, T3, T4, T5>& seq) { \
-  google::PrintSequence(out, seq.begin(), seq.end()); \
-  return out; \
-}
-
-#ifdef GLOG_STL_LOGGING_FOR_UNORDERED
-OUTPUT_FIVE_ARG_CONTAINER(std::unordered_map)
-OUTPUT_FIVE_ARG_CONTAINER(std::unordered_multimap)
-#endif
-#ifdef GLOG_STL_LOGGING_FOR_TR1_UNORDERED
-OUTPUT_FIVE_ARG_CONTAINER(std::tr1::unordered_map)
-OUTPUT_FIVE_ARG_CONTAINER(std::tr1::unordered_multimap)
-#endif
-#ifdef GLOG_STL_LOGGING_FOR_EXT_HASH
-OUTPUT_FIVE_ARG_CONTAINER(__gnu_cxx::hash_map)
-OUTPUT_FIVE_ARG_CONTAINER(__gnu_cxx::hash_multimap)
-#endif
-
-#undef OUTPUT_FIVE_ARG_CONTAINER
-
-template<class First, class Second>
-inline std::ostream& operator<<(std::ostream& out,
-                                const std::pair<First, Second>& p) {
-  out << '(' << p.first << ", " << p.second << ')';
-  return out;
-}
-
-namespace google {
-
-template<class Iter>
-inline void PrintSequence(std::ostream& out, Iter begin, Iter end) {
-  // Output at most 100 elements -- appropriate if used for logging.
-  for (int i = 0; begin != end && i < 100; ++i, ++begin) {
-    if (i > 0) out << ' ';
-    out << *begin;
-  }
-  if (begin != end) {
-    out << " ...";
-  }
-}
-
-}
-
-// Note that this is technically undefined behavior! We are adding things into
-// the std namespace for a reason though -- we are providing new operations on
-// types which are themselves defined with this namespace. Without this, these
-// operator overloads cannot be found via ADL. If these definitions are not
-// found via ADL, they must be #included before they're used, which requires
-// this header to be included before apparently independent other headers.
-//
-// For example, base/logging.h defines various template functions to implement
-// CHECK_EQ(x, y) and stream x and y into the log in the event the check fails.
-// It does so via the function template MakeCheckOpValueString:
-//   template<class T>
-//   void MakeCheckOpValueString(strstream* ss, const T& v) {
-//     (*ss) << v;
-//   }
-// Because 'glog/logging.h' is included before 'glog/stl_logging.h',
-// subsequent CHECK_EQ(v1, v2) for vector<...> typed variable v1 and v2 can only
-// find these operator definitions via ADL.
-//
-// Even this solution has problems -- it may pull unintended operators into the
-// namespace as well, allowing them to also be found via ADL, and creating code
-// that only works with a particular order of includes. Long term, we need to
-// move all of the *definitions* into namespace std, bet we need to ensure no
-// one references them first. This lets us take that step. We cannot define them
-// in both because that would create ambiguous overloads when both are found.
-namespace std { using ::operator<<; }
-
-#endif  // UTIL_GTL_STL_LOGGING_INL_H_
diff --git a/projects/llm_framework/include/glog/vlog_is_on.h b/projects/llm_framework/include/glog/vlog_is_on.h
deleted file mode 100644
index 02b0b867..00000000
--- a/projects/llm_framework/include/glog/vlog_is_on.h
+++ /dev/null
@@ -1,129 +0,0 @@
-// Copyright (c) 1999, 2007, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-//
-// Author: Ray Sidney and many others
-//
-// Defines the VLOG_IS_ON macro that controls the variable-verbosity
-// conditional logging.
-//
-// It's used by VLOG and VLOG_IF in logging.h
-// and by RAW_VLOG in raw_logging.h to trigger the logging.
-//
-// It can also be used directly e.g. like this:
-//   if (VLOG_IS_ON(2)) {
-//     // do some logging preparation and logging
-//     // that can't be accomplished e.g. via just VLOG(2) << ...;
-//   }
-//
-// The truth value that VLOG_IS_ON(level) returns is determined by 
-// the three verbosity level flags:
-//   --v=<n>  Gives the default maximal active V-logging level;
-//            0 is the default.
-//            Normally positive values are used for V-logging levels.
-//   --vmodule=<str>  Gives the per-module maximal V-logging levels to override
-//                    the value given by --v.
-//                    E.g. "my_module=2,foo*=3" would change the logging level
-//                    for all code in source files "my_module.*" and "foo*.*"
-//                    ("-inl" suffixes are also disregarded for this matching).
-//
-// SetVLOGLevel helper function is provided to do limited dynamic control over
-// V-logging by overriding the per-module settings given via --vmodule flag.
-//
-// CAVEAT: --vmodule functionality is not available in non gcc compilers.
-//
-
-#ifndef BASE_VLOG_IS_ON_H_
-#define BASE_VLOG_IS_ON_H_
-
-#include "glog/log_severity.h"
-
-// Annoying stuff for windows -- makes sure clients can import these functions
-#ifndef GOOGLE_GLOG_DLL_DECL
-# if defined(_WIN32) && !defined(__CYGWIN__)
-#   define GOOGLE_GLOG_DLL_DECL  __declspec(dllimport)
-# else
-#   define GOOGLE_GLOG_DLL_DECL
-# endif
-#endif
-
-#if defined(__GNUC__)
-// We emit an anonymous static int* variable at every VLOG_IS_ON(n) site.
-// (Normally) the first time every VLOG_IS_ON(n) site is hit,
-// we determine what variable will dynamically control logging at this site:
-// it's either FLAGS_v or an appropriate internal variable
-// matching the current source file that represents results of
-// parsing of --vmodule flag and/or SetVLOGLevel calls.
-#define VLOG_IS_ON(verboselevel)                                \
-  __extension__  \
-  ({ static google::int32* vlocal__ = &google::kLogSiteUninitialized;           \
-     google::int32 verbose_level__ = (verboselevel);                    \
-     (*vlocal__ >= verbose_level__) &&                          \
-     ((vlocal__ != &google::kLogSiteUninitialized) ||                   \
-      (google::InitVLOG3__(&vlocal__, &FLAGS_v,                         \
-                   __FILE__, verbose_level__))); })
-#else
-// GNU extensions not available, so we do not support --vmodule.
-// Dynamic value of FLAGS_v always controls the logging level.
-#define VLOG_IS_ON(verboselevel) (FLAGS_v >= (verboselevel))
-#endif
-
-// Set VLOG(_IS_ON) level for module_pattern to log_level.
-// This lets us dynamically control what is normally set by the --vmodule flag.
-// Returns the level that previously applied to module_pattern.
-// NOTE: To change the log level for VLOG(_IS_ON) sites
-//	 that have already executed after/during InitGoogleLogging,
-//	 one needs to supply the exact --vmodule pattern that applied to them.
-//       (If no --vmodule pattern applied to them
-//       the value of FLAGS_v will continue to control them.)
-extern GOOGLE_GLOG_DLL_DECL int SetVLOGLevel(const char* module_pattern,
-                                             int log_level);
-
-// Various declarations needed for VLOG_IS_ON above: =========================
-
-// Special value used to indicate that a VLOG_IS_ON site has not been
-// initialized.  We make this a large value, so the common-case check
-// of "*vlocal__ >= verbose_level__" in VLOG_IS_ON definition
-// passes in such cases and InitVLOG3__ is then triggered.
-extern google::int32 kLogSiteUninitialized;
-
-// Helper routine which determines the logging info for a particalur VLOG site.
-//   site_flag     is the address of the site-local pointer to the controlling
-//                 verbosity level
-//   site_default  is the default to use for *site_flag
-//   fname         is the current source file name
-//   verbose_level is the argument to VLOG_IS_ON
-// We will return the return value for VLOG_IS_ON
-// and if possible set *site_flag appropriately.
-extern GOOGLE_GLOG_DLL_DECL bool InitVLOG3__(
-    google::int32** site_flag,
-    google::int32* site_default,
-    const char* fname,
-    google::int32 verbose_level);
-
-#endif  // BASE_VLOG_IS_ON_H_
diff --git a/projects/llm_framework/include/libdw.h b/projects/llm_framework/include/libdw.h
deleted file mode 100644
index 77174d28..00000000
--- a/projects/llm_framework/include/libdw.h
+++ /dev/null
@@ -1,1111 +0,0 @@
-/* Interfaces for libdw.
-   Copyright (C) 2002-2010, 2013, 2014, 2016, 2018 Red Hat, Inc.
-   This file is part of elfutils.
-
-   This file is free software; you can redistribute it and/or modify
-   it under the terms of either
-
-     * the GNU Lesser General Public License as published by the Free
-       Software Foundation; either version 3 of the License, or (at
-       your option) any later version
-
-   or
-
-     * the GNU General Public License as published by the Free
-       Software Foundation; either version 2 of the License, or (at
-       your option) any later version
-
-   or both in parallel, as here.
-
-   elfutils is distributed in the hope that it will be useful, but
-   WITHOUT ANY WARRANTY; without even the implied warranty of
-   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
-   General Public License for more details.
-
-   You should have received copies of the GNU General Public License and
-   the GNU Lesser General Public License along with this program.  If
-   not, see <http://www.gnu.org/licenses/>.  */
-
-#ifndef _LIBDW_H
-#define _LIBDW_H	1
-
-#include <gelf.h>
-#include <stdbool.h>
-#include <stddef.h>
-#include <stdint.h>
-
-/* Mode for the session.  */
-typedef enum
-  {
-    DWARF_C_READ,		/* Read .. */
-    DWARF_C_RDWR,		/* Read and write .. */
-    DWARF_C_WRITE,		/* Write .. */
-  }
-Dwarf_Cmd;
-
-
-/* Callback results.  */
-enum
-{
-  DWARF_CB_OK = 0,
-  DWARF_CB_ABORT
-};
-
-
-/* Error values.  */
-enum
-  {
-    DW_TAG_invalid = 0
-#define DW_TAG_invalid	DW_TAG_invalid
-  };
-
-
-/* Type for offset in DWARF file.  */
-typedef GElf_Off Dwarf_Off;
-
-/* Type for address in DWARF file.  */
-typedef GElf_Addr Dwarf_Addr;
-
-/* Integer types.  Big enough to hold any numeric value.  */
-typedef GElf_Xword Dwarf_Word;
-typedef GElf_Sxword Dwarf_Sword;
-/* For the times we know we do not need that much.  */
-typedef GElf_Half Dwarf_Half;
-
-
-/* DWARF abbreviation record.  */
-typedef struct Dwarf_Abbrev Dwarf_Abbrev;
-
-/* Returned to show the last DIE has be returned.  */
-#define DWARF_END_ABBREV ((Dwarf_Abbrev *) -1l)
-
-/* Source code line information for CU.  */
-typedef struct Dwarf_Lines_s Dwarf_Lines;
-
-/* One source code line information.  */
-typedef struct Dwarf_Line_s Dwarf_Line;
-
-/* Source file information.  */
-typedef struct Dwarf_Files_s Dwarf_Files;
-
-/* One address range record.  */
-typedef struct Dwarf_Arange_s Dwarf_Arange;
-
-/* Address ranges of a file.  */
-typedef struct Dwarf_Aranges_s Dwarf_Aranges;
-
-/* CU representation.  */
-struct Dwarf_CU;
-typedef struct Dwarf_CU Dwarf_CU;
-
-/* Macro information.  */
-typedef struct Dwarf_Macro_s Dwarf_Macro;
-
-/* Attribute representation.  */
-typedef struct
-{
-  unsigned int code;
-  unsigned int form;
-  unsigned char *valp;
-  struct Dwarf_CU *cu;
-} Dwarf_Attribute;
-
-
-/* Data block representation.  */
-typedef struct
-{
-  Dwarf_Word length;
-  unsigned char *data;
-} Dwarf_Block;
-
-
-/* DIE information.  */
-typedef struct
-{
-  /* The offset can be computed from the address.  */
-  void *addr;
-  struct Dwarf_CU *cu;
-  Dwarf_Abbrev *abbrev;
-  // XXX We'll see what other information will be needed.
-  long int padding__;
-} Dwarf_Die;
-
-/* Returned to show the last DIE has be returned.  */
-#define DWARF_END_DIE ((Dwarf_Die *) -1l)
-
-
-/* Global symbol information.  */
-typedef struct
-{
-  Dwarf_Off cu_offset;
-  Dwarf_Off die_offset;
-  const char *name;
-} Dwarf_Global;
-
-
-/* One operation in a DWARF location expression.
-   A location expression is an array of these.  */
-typedef struct
-{
-  uint8_t atom;			/* Operation */
-  Dwarf_Word number;		/* Operand */
-  Dwarf_Word number2;		/* Possible second operand */
-  Dwarf_Word offset;		/* Offset in location expression */
-} Dwarf_Op;
-
-
-/* This describes one Common Information Entry read from a CFI section.
-   Pointers here point into the DATA->d_buf block passed to dwarf_next_cfi.  */
-typedef struct
-{
-  Dwarf_Off CIE_id;	 /* Always DW_CIE_ID_64 in Dwarf_CIE structures.  */
-
-  /* Instruction stream describing initial state used by FDEs.  If
-     we did not understand the whole augmentation string and it did
-     not use 'z', then there might be more augmentation data here
-     (and in FDEs) before the actual instructions.  */
-  const uint8_t *initial_instructions;
-  const uint8_t *initial_instructions_end;
-
-  Dwarf_Word code_alignment_factor;
-  Dwarf_Sword data_alignment_factor;
-  Dwarf_Word return_address_register;
-
-  const char *augmentation;	/* Augmentation string.  */
-
-  /* Augmentation data, might be NULL.  The size is correct only if
-     we understood the augmentation string sufficiently.  */
-  const uint8_t *augmentation_data;
-  size_t augmentation_data_size;
-  size_t fde_augmentation_data_size;
-} Dwarf_CIE;
-
-/* This describes one Frame Description Entry read from a CFI section.
-   Pointers here point into the DATA->d_buf block passed to dwarf_next_cfi.  */
-typedef struct
-{
-  /* Section offset of CIE this FDE refers to.  This will never be
-     DW_CIE_ID_64 in an FDE.  If this value is DW_CIE_ID_64, this is
-     actually a Dwarf_CIE structure.  */
-  Dwarf_Off CIE_pointer;
-
-  /* We can't really decode anything further without looking up the CIE
-     and checking its augmentation string.  Here follows the encoded
-     initial_location and address_range, then any augmentation data,
-     then the instruction stream.  This FDE describes PC locations in
-     the byte range [initial_location, initial_location+address_range).
-     When the CIE augmentation string uses 'z', the augmentation data is
-     a DW_FORM_block (self-sized).  Otherwise, when we understand the
-     augmentation string completely, fde_augmentation_data_size gives
-     the number of bytes of augmentation data before the instructions.  */
-  const uint8_t *start;
-  const uint8_t *end;
-} Dwarf_FDE;
-
-/* Each entry in a CFI section is either a CIE described by Dwarf_CIE or
-   an FDE described by Dward_FDE.  Check CIE_id to see which you have.  */
-typedef union
-{
-  Dwarf_Off CIE_id;	 /* Always DW_CIE_ID_64 in Dwarf_CIE structures.  */
-  Dwarf_CIE cie;
-  Dwarf_FDE fde;
-} Dwarf_CFI_Entry;
-
-/* Same as DW_CIE_ID_64 from dwarf.h to keep libdw.h independent.  */
-#define LIBDW_CIE_ID 0xffffffffffffffffULL
-#define dwarf_cfi_cie_p(entry)	((entry)->cie.CIE_id == LIBDW_CIE_ID)
-
-/* Opaque type representing a frame state described by CFI.  */
-typedef struct Dwarf_Frame_s Dwarf_Frame;
-
-/* Opaque type representing a CFI section found in a DWARF or ELF file.  */
-typedef struct Dwarf_CFI_s Dwarf_CFI;
-
-
-/* Handle for debug sessions.  */
-typedef struct Dwarf Dwarf;
-
-
-/* Out-Of-Memory handler.  */
-typedef void (*__noreturn_attribute__ Dwarf_OOM) (void);
-
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* Create a handle for a new debug session.  */
-extern Dwarf *dwarf_begin (int fildes, Dwarf_Cmd cmd);
-
-/* Create a handle for a new debug session for an ELF file.  */
-extern Dwarf *dwarf_begin_elf (Elf *elf, Dwarf_Cmd cmd, Elf_Scn *scngrp);
-
-/* Retrieve ELF descriptor used for DWARF access.  */
-extern Elf *dwarf_getelf (Dwarf *dwarf);
-
-/* Retrieve DWARF descriptor used for a Dwarf_Die or Dwarf_Attribute.
-   A Dwarf_Die or a Dwarf_Attribute is associated with a particular
-   Dwarf_CU handle.  This function returns the DWARF descriptor for
-   that Dwarf_CU.  */
-extern Dwarf *dwarf_cu_getdwarf (Dwarf_CU *cu);
-
-/* Retrieves the DWARF descriptor for debugaltlink data.  Returns NULL
-   if no alternate debug data has been supplied yet.  libdw will try
-   to set the alt file on first use of an alt FORM if not yet explicitly
-   provided by dwarf_setalt.  */
-extern Dwarf *dwarf_getalt (Dwarf *main);
-
-/* Provides the data referenced by the .gnu_debugaltlink section.  The
-   caller should check that MAIN and ALT match (i.e., they have the
-   same build ID).  It is the responsibility of the caller to ensure
-   that the data referenced by ALT stays valid while it is used by
-   MAIN, until dwarf_setalt is called on MAIN with a different
-   descriptor, or dwarf_end.  Must be called before inspecting DIEs
-   that might have alt FORMs.  Otherwise libdw will try to set the
-   alt file itself on first use.  */
-extern void dwarf_setalt (Dwarf *main, Dwarf *alt);
-
-/* Release debugging handling context.  */
-extern int dwarf_end (Dwarf *dwarf);
-
-
-/* Read the header for the DWARF CU.  */
-extern int dwarf_nextcu (Dwarf *dwarf, Dwarf_Off off, Dwarf_Off *next_off,
-			 size_t *header_sizep, Dwarf_Off *abbrev_offsetp,
-			 uint8_t *address_sizep, uint8_t *offset_sizep)
-     __nonnull_attribute__ (3);
-
-/* Read the header of a DWARF CU or type unit.  If TYPE_SIGNATUREP is not
-   null, this reads a type unit from the .debug_types section; otherwise
-   this reads a CU from the .debug_info section.  */
-extern int dwarf_next_unit (Dwarf *dwarf, Dwarf_Off off, Dwarf_Off *next_off,
-			    size_t *header_sizep, Dwarf_Half *versionp,
-			    Dwarf_Off *abbrev_offsetp,
-			    uint8_t *address_sizep, uint8_t *offset_sizep,
-			    uint64_t *type_signaturep, Dwarf_Off *type_offsetp)
-     __nonnull_attribute__ (3);
-
-
-/* Gets the next Dwarf_CU (unit), version, unit type and if available
-   the CU DIE and sub (type) DIE of the unit.  Returns 0 on success,
-   -1 on error or 1 if there are no more units.  To start iterating
-   provide NULL for CU.  If version < 5 the unit type is set from the
-   CU DIE if available (DW_UT_compile for DW_TAG_compile_unit,
-   DW_UT_type for DW_TAG_type_unit or DW_UT_partial for
-   DW_TAG_partial_unit), otherwise it is set to zero.  If unavailable
-   (the version or unit type is unknown) the CU DIE is cleared.
-   Likewise if the sub DIE isn't isn't available (the unit type is not
-   DW_UT_type or DW_UT_split_type) the sub DIE tag is cleared.  */
-extern int dwarf_get_units (Dwarf *dwarf, Dwarf_CU *cu, Dwarf_CU **next_cu,
-			    Dwarf_Half *version, uint8_t *unit_type,
-			    Dwarf_Die *cudie, Dwarf_Die *subdie)
-     __nonnull_attribute__ (3);
-
-/* Provides information and DIEs associated with the given Dwarf_CU
-   unit.  Returns -1 on error, zero on success. Arguments not needed
-   may be NULL.  If they are NULL and aren't known yet, they won't be
-   looked up.  If the subdie doesn't exist for this unit_type it will
-   be cleared.  If there is no unit_id for this unit type it will be
-   set to zero.  */
-extern int dwarf_cu_info (Dwarf_CU *cu,
-			  Dwarf_Half *version, uint8_t *unit_type,
-			  Dwarf_Die *cudie, Dwarf_Die *subdie,
-			  uint64_t *unit_id,
-			  uint8_t *address_size, uint8_t *offset_size);
-
-/* Decode one DWARF CFI entry (CIE or FDE) from the raw section data.
-   The E_IDENT from the originating ELF file indicates the address
-   size and byte order used in the CFI section contained in DATA;
-   EH_FRAME_P should be true for .eh_frame format and false for
-   .debug_frame format.  OFFSET is the byte position in the section
-   to start at; on return *NEXT_OFFSET is filled in with the byte
-   position immediately after this entry.
-
-   On success, returns 0 and fills in *ENTRY; use dwarf_cfi_cie_p to
-   see whether ENTRY->cie or ENTRY->fde is valid.
-
-   On errors, returns -1.  Some format errors will permit safely
-   skipping to the next CFI entry though the current one is unusable.
-   In that case, *NEXT_OFF will be updated before a -1 return.
-
-   If there are no more CFI entries left in the section,
-   returns 1 and sets *NEXT_OFFSET to (Dwarf_Off) -1.  */
-extern int dwarf_next_cfi (const unsigned char e_ident[],
-			   Elf_Data *data, bool eh_frame_p,
-			   Dwarf_Off offset, Dwarf_Off *next_offset,
-			   Dwarf_CFI_Entry *entry)
-  __nonnull_attribute__ (1, 2, 5, 6);
-
-/* Use the CFI in the DWARF .debug_frame section.
-   Returns NULL if there is no such section (not an error).
-   The pointer returned can be used until dwarf_end is called on DWARF,
-   and must not be passed to dwarf_cfi_end.
-   Calling this more than once returns the same pointer.  */
-extern Dwarf_CFI *dwarf_getcfi (Dwarf *dwarf);
-
-/* Use the CFI in the ELF file's exception-handling data.
-   Returns NULL if there is no such data.
-   The pointer returned can be used until elf_end is called on ELF,
-   and must be passed to dwarf_cfi_end before then.
-   Calling this more than once allocates independent data structures.  */
-extern Dwarf_CFI *dwarf_getcfi_elf (Elf *elf);
-
-/* Release resources allocated by dwarf_getcfi_elf.  */
-extern int dwarf_cfi_end (Dwarf_CFI *cache);
-
-
-/* Return DIE at given offset in .debug_info section.  */
-extern Dwarf_Die *dwarf_offdie (Dwarf *dbg, Dwarf_Off offset,
-				Dwarf_Die *result) __nonnull_attribute__ (3);
-
-/* Return DIE at given offset in .debug_types section.  */
-extern Dwarf_Die *dwarf_offdie_types (Dwarf *dbg, Dwarf_Off offset,
-				      Dwarf_Die *result)
-     __nonnull_attribute__ (3);
-
-/* Return offset of DIE.  */
-extern Dwarf_Off dwarf_dieoffset (Dwarf_Die *die);
-
-/* Return offset of DIE in CU.  */
-extern Dwarf_Off dwarf_cuoffset (Dwarf_Die *die);
-
-/* Return CU DIE containing given DIE.  */
-extern Dwarf_Die *dwarf_diecu (Dwarf_Die *die, Dwarf_Die *result,
-			       uint8_t *address_sizep, uint8_t *offset_sizep)
-     __nonnull_attribute__ (2);
-
-/* Given a Dwarf_Die addr returns a (reconstructed) Dwarf_Die, or NULL
-   if the given addr didn't come from a valid Dwarf_Die.  In particular
-   it will make sure that the correct Dwarf_CU pointer is set for the
-   Dwarf_Die, the Dwarf_Abbrev pointer will not be set up yet (it will
-   only be once the Dwarf_Die is used to read attributes, children or
-   siblings).  This functions can be used to keep a reference to a
-   Dwarf_Die which you want to refer to later.  The addr, and the result
-   of this function, is only valid while the associated Dwarf is valid.  */
-extern Dwarf_Die *dwarf_die_addr_die (Dwarf *dbg, void *addr,
-				      Dwarf_Die *result)
-     __nonnull_attribute__ (3);
-
-/* Return the CU DIE and the header info associated with a Dwarf_Die
-   or Dwarf_Attribute.  A Dwarf_Die or a Dwarf_Attribute is associated
-   with a particular Dwarf_CU handle.  This function returns the CU or
-   type unit DIE and header information for that Dwarf_CU.  The
-   returned DIE is either a compile_unit, partial_unit or type_unit.
-   If it is a type_unit, then the type signature and type offset are
-   also provided, otherwise type_offset will be set to zero.  See also
-   dwarf_diecu and dwarf_next_unit.  */
-extern Dwarf_Die *dwarf_cu_die (Dwarf_CU *cu, Dwarf_Die *result,
-				Dwarf_Half *versionp,
-				Dwarf_Off *abbrev_offsetp,
-				uint8_t *address_sizep,
-				uint8_t *offset_sizep,
-				uint64_t *type_signaturep,
-				Dwarf_Off *type_offsetp)
-     __nonnull_attribute__ (2);
-
-/* Return CU DIE containing given address.  */
-extern Dwarf_Die *dwarf_addrdie (Dwarf *dbg, Dwarf_Addr addr,
-				 Dwarf_Die *result) __nonnull_attribute__ (3);
-
-/* Return child of current DIE.  */
-extern int dwarf_child (Dwarf_Die *die, Dwarf_Die *result)
-     __nonnull_attribute__ (2);
-
-/* Locates the first sibling of DIE and places it in RESULT.
-   Returns 0 if a sibling was found, -1 if something went wrong.
-   Returns 1 if no sibling could be found and, if RESULT is not
-   the same as DIE, it sets RESULT->addr to the address of the
-   (non-sibling) DIE that follows this one, or NULL if this DIE
-   was the last one in the compilation unit.  */
-extern int dwarf_siblingof (Dwarf_Die *die, Dwarf_Die *result)
-     __nonnull_attribute__ (2);
-
-/* For type aliases and qualifier type DIEs, which don't modify or
-   change the structural layout of the underlying type, follow the
-   DW_AT_type attribute (recursively) and return the underlying type
-   Dwarf_Die.
-
-   Returns 0 when RESULT contains a Dwarf_Die (possibly equal to the
-   given DIE) that isn't a type alias or qualifier type.  Returns 1
-   when RESULT contains a type alias or qualifier Dwarf_Die that
-   couldn't be peeled further (it doesn't have a DW_TAG_type
-   attribute).  Returns -1 when an error occurred.
-
-   The current DWARF specification defines one type alias tag
-   (DW_TAG_typedef) and seven modifier/qualifier type tags
-   (DW_TAG_const_type, DW_TAG_volatile_type, DW_TAG_restrict_type,
-   DW_TAG_atomic_type, DW_TAG_immutable_type, DW_TAG_packed_type and
-   DW_TAG_shared_type).  This function won't peel modifier type
-   tags that change the way the underlying type is accessed such
-   as the pointer or reference type tags (DW_TAG_pointer_type,
-   DW_TAG_reference_type or DW_TAG_rvalue_reference_type).
-
-   A future version of this function might peel other alias or
-   qualifier type tags if a future DWARF version or GNU extension
-   defines other type aliases or qualifier type tags that don't modify,
-   change the structural layout or the way to access the underlying type.  */
-extern int dwarf_peel_type (Dwarf_Die *die, Dwarf_Die *result)
-    __nonnull_attribute__ (2);
-
-/* Check whether the DIE has children.  */
-extern int dwarf_haschildren (Dwarf_Die *die) __nonnull_attribute__ (1);
-
-/* Walks the attributes of DIE, starting at the one OFFSET bytes in,
-   calling the CALLBACK function for each one.  Stops if the callback
-   function ever returns a value other than DWARF_CB_OK and returns the
-   offset of the offending attribute.  If the end of the attributes
-   is reached 1 is returned.  If something goes wrong -1 is returned and
-   the dwarf error number is set.  */
-extern ptrdiff_t dwarf_getattrs (Dwarf_Die *die,
-				 int (*callback) (Dwarf_Attribute *, void *),
-				 void *arg, ptrdiff_t offset)
-     __nonnull_attribute__ (2);
-
-/* Return tag of given DIE.  */
-extern int dwarf_tag (Dwarf_Die *die) __nonnull_attribute__ (1);
-
-
-/* Return specific attribute of DIE.  */
-extern Dwarf_Attribute *dwarf_attr (Dwarf_Die *die, unsigned int search_name,
-				    Dwarf_Attribute *result)
-     __nonnull_attribute__ (3);
-
-/* Check whether given DIE has specific attribute.  */
-extern int dwarf_hasattr (Dwarf_Die *die, unsigned int search_name);
-
-/* These are the same as dwarf_attr and dwarf_hasattr, respectively,
-   but they resolve an indirect attribute through
-   DW_AT_abstract_origin, DW_AT_specification or, if the DIE is a
-   top-level split CU, the skeleton DIE.  Note that the attribute
-   might come from a DIE in a different CU (possibly from a different
-   Dwarf file).  In that case all attribute information needs to be
-   resolved through the CU associated with the returned
-   Dwarf_Attribute.  The dwarf_form functions already do this
-   automatically.  */
-extern Dwarf_Attribute *dwarf_attr_integrate (Dwarf_Die *die,
-					      unsigned int search_name,
-					      Dwarf_Attribute *result)
-     __nonnull_attribute__ (3);
-extern int dwarf_hasattr_integrate (Dwarf_Die *die, unsigned int search_name);
-
-
-
-
-/* Check whether given attribute has specific form.  */
-extern int dwarf_hasform (Dwarf_Attribute *attr, unsigned int search_form);
-
-/* Return attribute code of given attribute.  */
-extern unsigned int dwarf_whatattr (Dwarf_Attribute *attr);
-
-/* Return form code of given attribute.  */
-extern unsigned int dwarf_whatform (Dwarf_Attribute *attr);
-
-
-/* Return string associated with given attribute.  */
-extern const char *dwarf_formstring (Dwarf_Attribute *attrp);
-
-/* Return unsigned constant represented by attribute.  */
-extern int dwarf_formudata (Dwarf_Attribute *attr, Dwarf_Word *return_uval)
-     __nonnull_attribute__ (2);
-
-/* Return signed constant represented by attribute.  */
-extern int dwarf_formsdata (Dwarf_Attribute *attr, Dwarf_Sword *return_uval)
-     __nonnull_attribute__ (2);
-
-/* Return address represented by attribute.  */
-extern int dwarf_formaddr (Dwarf_Attribute *attr, Dwarf_Addr *return_addr)
-     __nonnull_attribute__ (2);
-
-/* This function is deprecated.  Always use dwarf_formref_die instead.
-   Return reference offset represented by attribute.  */
-extern int dwarf_formref (Dwarf_Attribute *attr, Dwarf_Off *return_offset)
-     __nonnull_attribute__ (2) __deprecated_attribute__;
-
-/* Look up the DIE in a reference-form attribute.  */
-extern Dwarf_Die *dwarf_formref_die (Dwarf_Attribute *attr, Dwarf_Die *die_mem)
-     __nonnull_attribute__ (2);
-
-/* Return block represented by attribute.  */
-extern int dwarf_formblock (Dwarf_Attribute *attr, Dwarf_Block *return_block)
-     __nonnull_attribute__ (2);
-
-/* Return flag represented by attribute.  */
-extern int dwarf_formflag (Dwarf_Attribute *attr, bool *return_bool)
-     __nonnull_attribute__ (2);
-
-
-/* Simplified attribute value access functions.  */
-
-/* Return string in name attribute of DIE.  */
-extern const char *dwarf_diename (Dwarf_Die *die);
-
-/* Return high PC attribute of DIE.  */
-extern int dwarf_highpc (Dwarf_Die *die, Dwarf_Addr *return_addr)
-     __nonnull_attribute__ (2);
-
-/* Return low PC attribute of DIE.  */
-extern int dwarf_lowpc (Dwarf_Die *die, Dwarf_Addr *return_addr)
-     __nonnull_attribute__ (2);
-
-/* Return entry_pc or low_pc attribute of DIE.  */
-extern int dwarf_entrypc (Dwarf_Die *die, Dwarf_Addr *return_addr)
-     __nonnull_attribute__ (2);
-
-/* Return 1 if DIE's lowpc/highpc or ranges attributes match the PC address,
-   0 if not, or -1 for errors.  */
-extern int dwarf_haspc (Dwarf_Die *die, Dwarf_Addr pc);
-
-/* Enumerate the PC address ranges covered by this DIE, covering all
-   addresses where dwarf_haspc returns true.  In the first call OFFSET
-   should be zero and *BASEP need not be initialized.  Returns -1 for
-   errors, zero when there are no more address ranges to report, or a
-   nonzero OFFSET value to pass to the next call.  Each subsequent call
-   must preserve *BASEP from the prior call.  Successful calls fill in
-   *STARTP and *ENDP with a contiguous address range.  */
-extern ptrdiff_t dwarf_ranges (Dwarf_Die *die,
-			       ptrdiff_t offset, Dwarf_Addr *basep,
-			       Dwarf_Addr *startp, Dwarf_Addr *endp);
-
-
-/* Return byte size attribute of DIE.  */
-extern int dwarf_bytesize (Dwarf_Die *die);
-
-/* Return bit size attribute of DIE.  */
-extern int dwarf_bitsize (Dwarf_Die *die);
-
-/* Return bit offset attribute of DIE.  */
-extern int dwarf_bitoffset (Dwarf_Die *die);
-
-/* Return array order attribute of DIE.  */
-extern int dwarf_arrayorder (Dwarf_Die *die);
-
-/* Return source language attribute of DIE.  */
-extern int dwarf_srclang (Dwarf_Die *die);
-
-
-/* Get abbreviation at given offset for given DIE.  */
-extern Dwarf_Abbrev *dwarf_getabbrev (Dwarf_Die *die, Dwarf_Off offset,
-				      size_t *lengthp);
-
-/* Get abbreviation at given offset in .debug_abbrev section.  */
-extern int dwarf_offabbrev (Dwarf *dbg, Dwarf_Off offset, size_t *lengthp,
-			    Dwarf_Abbrev *abbrevp)
-     __nonnull_attribute__ (4);
-
-/* Get abbreviation code.  */
-extern unsigned int dwarf_getabbrevcode (Dwarf_Abbrev *abbrev);
-
-/* Get abbreviation tag.  */
-extern unsigned int dwarf_getabbrevtag (Dwarf_Abbrev *abbrev);
-
-/* Return true if abbreviation is children flag set.  */
-extern int dwarf_abbrevhaschildren (Dwarf_Abbrev *abbrev);
-
-/* Get number of attributes of abbreviation.  */
-extern int dwarf_getattrcnt (Dwarf_Abbrev *abbrev, size_t *attrcntp)
-     __nonnull_attribute__ (2);
-
-/* Get specific attribute of abbreviation.  */
-extern int dwarf_getabbrevattr (Dwarf_Abbrev *abbrev, size_t idx,
-				unsigned int *namep, unsigned int *formp,
-				Dwarf_Off *offset);
-
-/* Get specific attribute of abbreviation and any data encoded with it.
-   Specifically for DW_FORM_implicit_const data will be set to the
-   constant value associated.  */
-extern int dwarf_getabbrevattr_data (Dwarf_Abbrev *abbrev, size_t idx,
-				     unsigned int *namep, unsigned int *formp,
-				     Dwarf_Sword *datap, Dwarf_Off *offset);
-
-/* Get string from-debug_str section.  */
-extern const char *dwarf_getstring (Dwarf *dbg, Dwarf_Off offset,
-				    size_t *lenp);
-
-
-/* Get public symbol information.  */
-extern ptrdiff_t dwarf_getpubnames (Dwarf *dbg,
-				    int (*callback) (Dwarf *, Dwarf_Global *,
-						     void *),
-				    void *arg, ptrdiff_t offset)
-     __nonnull_attribute__ (2);
-
-
-/* Get source file information for CU.  */
-extern int dwarf_getsrclines (Dwarf_Die *cudie, Dwarf_Lines **lines,
-			      size_t *nlines) __nonnull_attribute__ (2, 3);
-
-/* Return one of the source lines of the CU.  */
-extern Dwarf_Line *dwarf_onesrcline (Dwarf_Lines *lines, size_t idx);
-
-/* Get the file source files used in the CU.  */
-extern int dwarf_getsrcfiles (Dwarf_Die *cudie, Dwarf_Files **files,
-			      size_t *nfiles)
-     __nonnull_attribute__ (2);
-
-
-/* Get source for address in CU.  */
-extern Dwarf_Line *dwarf_getsrc_die (Dwarf_Die *cudie, Dwarf_Addr addr);
-
-/* Get source for file and line number.  */
-extern int dwarf_getsrc_file (Dwarf *dbg, const char *fname, int line, int col,
-			      Dwarf_Line ***srcsp, size_t *nsrcs)
-     __nonnull_attribute__ (2, 5, 6);
-
-
-/* Return line address.  */
-extern int dwarf_lineaddr (Dwarf_Line *line, Dwarf_Addr *addrp);
-
-/* Return line VLIW operation index.  */
-extern int dwarf_lineop_index (Dwarf_Line *line, unsigned int *op_indexp);
-
-/* Return line number.  */
-extern int dwarf_lineno (Dwarf_Line *line, int *linep)
-     __nonnull_attribute__ (2);
-
-/* Return column in line.  */
-extern int dwarf_linecol (Dwarf_Line *line, int *colp)
-     __nonnull_attribute__ (2);
-
-/* Return true if record is for beginning of a statement.  */
-extern int dwarf_linebeginstatement (Dwarf_Line *line, bool *flagp)
-     __nonnull_attribute__ (2);
-
-/* Return true if record is for end of sequence.  */
-extern int dwarf_lineendsequence (Dwarf_Line *line, bool *flagp)
-     __nonnull_attribute__ (2);
-
-/* Return true if record is for beginning of a basic block.  */
-extern int dwarf_lineblock (Dwarf_Line *line, bool *flagp)
-     __nonnull_attribute__ (2);
-
-/* Return true if record is for end of prologue.  */
-extern int dwarf_lineprologueend (Dwarf_Line *line, bool *flagp)
-     __nonnull_attribute__ (2);
-
-/* Return true if record is for beginning of epilogue.  */
-extern int dwarf_lineepiloguebegin (Dwarf_Line *line, bool *flagp)
-     __nonnull_attribute__ (2);
-
-/* Return instruction-set architecture in this record.  */
-extern int dwarf_lineisa (Dwarf_Line *line, unsigned int *isap)
-     __nonnull_attribute__ (2);
-
-/* Return code path discriminator in this record.  */
-extern int dwarf_linediscriminator (Dwarf_Line *line, unsigned int *discp)
-     __nonnull_attribute__ (2);
-
-
-/* Find line information for address.  The returned string is NULL when
-   an error occurred, or the file path.  The file path is either absolute
-   or relative to the compilation directory.  See dwarf_decl_file.  */
-extern const char *dwarf_linesrc (Dwarf_Line *line,
-				  Dwarf_Word *mtime, Dwarf_Word *length);
-
-/* Return file information.  The returned string is NULL when
-   an error occurred, or the file path.  The file path is either absolute
-   or relative to the compilation directory.  See dwarf_decl_file.  */
-extern const char *dwarf_filesrc (Dwarf_Files *file, size_t idx,
-				  Dwarf_Word *mtime, Dwarf_Word *length);
-
-/* Return the Dwarf_Files and index associated with the given Dwarf_Line.  */
-extern int dwarf_line_file (Dwarf_Line *line,
-			    Dwarf_Files **files, size_t *idx)
-    __nonnull_attribute__ (2, 3);
-
-/* Return the directory list used in the file information extracted.
-   (*RESULT)[0] is the CU's DW_AT_comp_dir value, and may be null.
-   (*RESULT)[0..*NDIRS-1] are the compile-time include directory path
-   encoded by the compiler.  */
-extern int dwarf_getsrcdirs (Dwarf_Files *files,
-			     const char *const **result, size_t *ndirs)
-  __nonnull_attribute__ (2, 3);
-
-/* Iterates through the debug line units.  Returns 0 on success, -1 on
-   error or 1 if there are no more units.  To start iterating use zero
-   for OFF and set *CU to NULL.  On success NEXT_OFF will be set to
-   the next offset to use.  The *CU will be set if this line table
-   needed a specific CU and needs to be given when calling
-   dwarf_next_lines again (to help dwarf_next_lines quickly find the
-   next CU).  *CU might be set to NULL when it couldn't be found (the
-   compilation directory entry will be the empty string in that case)
-   or for DWARF 5 or later tables, which are self contained.  SRCFILES
-   and SRCLINES may be NULL if the caller is not interested in the
-   actual line or file table.  On success and when not NULL, NFILES
-   and NLINES will be set to the number of files in the file table and
-   number of lines in the line table.  */
-extern int dwarf_next_lines (Dwarf *dwarf, Dwarf_Off off,
-			     Dwarf_Off *next_off, Dwarf_CU **cu,
-			     Dwarf_Files **srcfiles, size_t *nfiles,
-			     Dwarf_Lines **srclines, size_t *nlines)
-  __nonnull_attribute__ (3,4);
-
-/* Return location expression, decoded as a list of operations.  */
-extern int dwarf_getlocation (Dwarf_Attribute *attr, Dwarf_Op **expr,
-			      size_t *exprlen) __nonnull_attribute__ (2, 3);
-
-/* Return location expressions.  If the attribute uses a location list,
-   ADDRESS selects the relevant location expressions from the list.
-   There can be multiple matches, resulting in multiple expressions to
-   return.  EXPRS and EXPRLENS are parallel arrays of NLOCS slots to
-   fill in.  Returns the number of locations filled in, or -1 for
-   errors.  If EXPRS is a null pointer, stores nothing and returns the
-   total number of locations.  A return value of zero means that the
-   location list indicated no value is accessible.  */
-extern int dwarf_getlocation_addr (Dwarf_Attribute *attr, Dwarf_Addr address,
-				   Dwarf_Op **exprs, size_t *exprlens,
-				   size_t nlocs);
-
-/* Enumerate the locations ranges and descriptions covered by the
-   given attribute.  In the first call OFFSET should be zero and
-   *BASEP need not be initialized.  Returns -1 for errors, zero when
-   there are no more locations to report, or a nonzero OFFSET
-   value to pass to the next call.  Each subsequent call must preserve
-   *BASEP from the prior call.  Successful calls fill in *STARTP and
-   *ENDP with a contiguous address range and *EXPR with a pointer to
-   an array of operations with length *EXPRLEN.  If the attribute
-   describes a single location description and not a location list the
-   first call (with OFFSET zero) will return the location description
-   in *EXPR with *STARTP set to zero and *ENDP set to minus one.  */
-extern ptrdiff_t dwarf_getlocations (Dwarf_Attribute *attr,
-				     ptrdiff_t offset, Dwarf_Addr *basep,
-				     Dwarf_Addr *startp, Dwarf_Addr *endp,
-				     Dwarf_Op **expr, size_t *exprlen);
-
-/* Return the block associated with a DW_OP_implicit_value operation.
-   The OP pointer must point into an expression that dwarf_getlocation
-   or dwarf_getlocation_addr has returned given the same ATTR.  */
-extern int dwarf_getlocation_implicit_value (Dwarf_Attribute *attr,
-					     const Dwarf_Op *op,
-					     Dwarf_Block *return_block)
-  __nonnull_attribute__ (2, 3);
-
-/* Return the attribute indicated by a DW_OP_GNU_implicit_pointer operation.
-   The OP pointer must point into an expression that dwarf_getlocation
-   or dwarf_getlocation_addr has returned given the same ATTR.
-   The result is the DW_AT_location or DW_AT_const_value attribute
-   of the OP->number DIE.  */
-extern int dwarf_getlocation_implicit_pointer (Dwarf_Attribute *attr,
-					       const Dwarf_Op *op,
-					       Dwarf_Attribute *result)
-  __nonnull_attribute__ (2, 3);
-
-/* Return the DIE associated with an operation such as
-   DW_OP_GNU_implicit_pointer, DW_OP_GNU_parameter_ref, DW_OP_GNU_convert,
-   DW_OP_GNU_reinterpret, DW_OP_GNU_const_type, DW_OP_GNU_regval_type or
-   DW_OP_GNU_deref_type.  The OP pointer must point into an expression that
-   dwarf_getlocation or dwarf_getlocation_addr has returned given the same
-   ATTR.  The RESULT is a DIE that expresses a type or value needed by the
-   given OP.  */
-extern int dwarf_getlocation_die (Dwarf_Attribute *attr,
-				  const Dwarf_Op *op,
-				  Dwarf_Die *result)
-  __nonnull_attribute__ (2, 3);
-
-/* Return the attribute expressing a value associated with an operation such
-   as DW_OP_implicit_value, DW_OP_GNU_entry_value or DW_OP_GNU_const_type.
-   The OP pointer must point into an expression that dwarf_getlocation
-   or dwarf_getlocation_addr has returned given the same ATTR.
-   The RESULT is a value expressed by an attribute such as DW_AT_location
-   or DW_AT_const_value.  */
-extern int dwarf_getlocation_attr (Dwarf_Attribute *attr,
-				   const Dwarf_Op *op,
-				   Dwarf_Attribute *result)
-  __nonnull_attribute__ (2, 3);
-
-
-/* Compute the byte-size of a type DIE according to DWARF rules.
-   For most types, this is just DW_AT_byte_size.
-   For DW_TAG_array_type it can apply much more complex rules.  */
-extern int dwarf_aggregate_size (Dwarf_Die *die, Dwarf_Word *size);
-
-/* Given a language code, as returned by dwarf_srclan, get the default
-   lower bound for a subrange type without a lower bound attribute.
-   Returns zero on success or -1 on failure when the given language
-   wasn't recognized.  */
-extern int dwarf_default_lower_bound (int lang, Dwarf_Sword *result)
-  __nonnull_attribute__ (2);
-
-/* Return scope DIEs containing PC address.
-   Sets *SCOPES to a malloc'd array of Dwarf_Die structures,
-   and returns the number of elements in the array.
-   (*SCOPES)[0] is the DIE for the innermost scope containing PC,
-   (*SCOPES)[1] is the DIE for the scope containing that scope, and so on.
-   Returns -1 for errors or 0 if no scopes match PC.  */
-extern int dwarf_getscopes (Dwarf_Die *cudie, Dwarf_Addr pc,
-			    Dwarf_Die **scopes);
-
-/* Return scope DIEs containing the given DIE.
-   Sets *SCOPES to a malloc'd array of Dwarf_Die structures,
-   and returns the number of elements in the array.
-   (*SCOPES)[0] is a copy of DIE.
-   (*SCOPES)[1] is the DIE for the scope containing that scope, and so on.
-   Returns -1 for errors or 0 if DIE is not found in any scope entry.  */
-extern int dwarf_getscopes_die (Dwarf_Die *die, Dwarf_Die **scopes);
-
-
-/* Search SCOPES[0..NSCOPES-1] for a variable called NAME.
-   Ignore the first SKIP_SHADOWS scopes that match the name.
-   If MATCH_FILE is not null, accept only declaration in that source file;
-   if MATCH_LINENO or MATCH_LINECOL are also nonzero, accept only declaration
-   at that line and column.
-
-   If successful, fill in *RESULT with the DIE of the variable found,
-   and return N where SCOPES[N] is the scope defining the variable.
-   Return -1 for errors or -2 for no matching variable found.  */
-extern int dwarf_getscopevar (Dwarf_Die *scopes, int nscopes,
-			      const char *name, int skip_shadows,
-			      const char *match_file,
-			      int match_lineno, int match_linecol,
-			      Dwarf_Die *result);
-
-
-
-/* Return list address ranges.  */
-extern int dwarf_getaranges (Dwarf *dbg, Dwarf_Aranges **aranges,
-			     size_t *naranges)
-     __nonnull_attribute__ (2);
-
-/* Return one of the address range entries.  */
-extern Dwarf_Arange *dwarf_onearange (Dwarf_Aranges *aranges, size_t idx);
-
-/* Return information in address range record.  */
-extern int dwarf_getarangeinfo (Dwarf_Arange *arange, Dwarf_Addr *addrp,
-				Dwarf_Word *lengthp, Dwarf_Off *offsetp);
-
-/* Get address range which includes given address.  */
-extern Dwarf_Arange *dwarf_getarange_addr (Dwarf_Aranges *aranges,
-					   Dwarf_Addr addr);
-
-
-
-/* Get functions in CUDIE.  The given callback will be called for all
-   defining DW_TAG_subprograms in the CU DIE tree.  If the callback
-   returns DWARF_CB_ABORT the return value can be used as offset argument
-   to resume the function to find all remaining functions (this is not
-   really recommended, since it needs to rewalk the CU DIE tree first till
-   that offset is found again).  If the callback returns DWARF_CB_OK
-   dwarf_getfuncs will not return but keep calling the callback for each
-   function DIE it finds.  Pass zero for offset on the first call to walk
-   the full CU DIE tree.  If no more functions can be found and the callback
-   returned DWARF_CB_OK then the function returns zero.  */
-extern ptrdiff_t dwarf_getfuncs (Dwarf_Die *cudie,
-				 int (*callback) (Dwarf_Die *, void *),
-				 void *arg, ptrdiff_t offset);
-
-
-/* Return file name containing definition of the given declaration.
-   Of the DECL has an (indirect, see dwarf_attr_integrate) decl_file
-   attribute.  The returned file path is either absolute, or relative
-   to the compilation directory.  Given the decl DIE, the compilation
-   directory can be retrieved through:
-   dwarf_formstring (dwarf_attr (dwarf_diecu (decl, &cudie, NULL, NULL),
-                                 DW_AT_comp_dir, &attr));
-   Returns NULL if no decl_file could be found or an error occurred.  */
-extern const char *dwarf_decl_file (Dwarf_Die *decl);
-
-/* Get line number of beginning of given declaration.  */
-extern int dwarf_decl_line (Dwarf_Die *decl, int *linep)
-     __nonnull_attribute__ (2);
-
-/* Get column number of beginning of given declaration.  */
-extern int dwarf_decl_column (Dwarf_Die *decl, int *colp)
-     __nonnull_attribute__ (2);
-
-
-/* Return nonzero if given function is an abstract inline definition.  */
-extern int dwarf_func_inline (Dwarf_Die *func);
-
-/* Find each concrete inlined instance of the abstract inline definition.  */
-extern int dwarf_func_inline_instances (Dwarf_Die *func,
-					int (*callback) (Dwarf_Die *, void *),
-					void *arg);
-
-
-/* Find the appropriate PC location or locations for function entry
-   breakpoints for the given DW_TAG_subprogram DIE.  Returns -1 for errors.
-   On success, returns the number of breakpoint locations (never zero)
-   and sets *BKPTS to a malloc'd vector of addresses.  */
-extern int dwarf_entry_breakpoints (Dwarf_Die *die, Dwarf_Addr **bkpts);
-
-
-/* Iterate through the macro unit referenced by CUDIE and call
-   CALLBACK for each macro information entry.  To start the iteration,
-   one would pass DWARF_GETMACROS_START for TOKEN.
-
-   The iteration continues while CALLBACK returns DWARF_CB_OK.  If the
-   callback returns DWARF_CB_ABORT, the iteration stops and a
-   continuation token is returned, which can be used to restart the
-   iteration at the point where it ended.  Returns -1 for errors or 0
-   if there are no more macro entries.
-
-   Note that the Dwarf_Macro pointer passed to the callback is only
-   valid for the duration of the callback invocation.
-
-   For backward compatibility, a token of 0 is accepted for starting
-   the iteration as well, but in that case this interface will refuse
-   to serve opcode 0xff from .debug_macro sections.  Such opcode would
-   be considered invalid and would cause dwarf_getmacros to return
-   with error.  */
-#define DWARF_GETMACROS_START PTRDIFF_MIN
-extern ptrdiff_t dwarf_getmacros (Dwarf_Die *cudie,
-				  int (*callback) (Dwarf_Macro *, void *),
-				  void *arg, ptrdiff_t token)
-     __nonnull_attribute__ (2);
-
-/* This is similar in operation to dwarf_getmacros, but selects the
-   unit to iterate through by offset instead of by CU, and always
-   iterates .debug_macro.  This can be used for handling
-   DW_MACRO_GNU_transparent_include's or similar opcodes.
-
-   TOKEN value of DWARF_GETMACROS_START can be used to start the
-   iteration.
-
-   It is not appropriate to obtain macro unit offset by hand from a CU
-   DIE and then request iteration through this interface.  The reason
-   for this is that if a dwarf_macro_getsrcfiles is later called,
-   there would be no way to figure out what DW_AT_comp_dir was present
-   on the CU DIE, and file names referenced in either the macro unit
-   itself, or the .debug_line unit that it references, might be wrong.
-   Use dwarf_getmacros.  */
-extern ptrdiff_t dwarf_getmacros_off (Dwarf *dbg, Dwarf_Off macoff,
-				      int (*callback) (Dwarf_Macro *, void *),
-				      void *arg, ptrdiff_t token)
-  __nonnull_attribute__ (3);
-
-/* Get the source files used by the macro entry.  You shouldn't assume
-   that Dwarf_Files references will remain valid after MACRO becomes
-   invalid.  (Which is to say it's only valid within the
-   dwarf_getmacros* callback.)  Returns 0 for success or a negative
-   value in case of an error.  */
-extern int dwarf_macro_getsrcfiles (Dwarf *dbg, Dwarf_Macro *macro,
-				    Dwarf_Files **files, size_t *nfiles)
-  __nonnull_attribute__ (2, 3, 4);
-
-/* Return macro opcode.  That's a constant that can be either from
-   DW_MACINFO_* domain or DW_MACRO_GNU_* domain.  The two domains have
-   compatible values, so it's OK to use either of them for
-   comparisons.  The only differences is 0xff, which could be either
-   DW_MACINFO_vendor_ext or a vendor-defined DW_MACRO_* constant.  One
-   would need to look if the CU DIE which the iteration was requested
-   for has attribute DW_AT_macro_info, or either of DW_AT_GNU_macros
-   or DW_AT_macros to differentiate the two interpretations.  */
-extern int dwarf_macro_opcode (Dwarf_Macro *macro, unsigned int *opcodep)
-     __nonnull_attribute__ (2);
-
-/* Get number of parameters of MACRO and store it to *PARAMCNTP.  */
-extern int dwarf_macro_getparamcnt (Dwarf_Macro *macro, size_t *paramcntp);
-
-/* Get IDX-th parameter of MACRO (numbered from zero), and stores it
-   to *ATTRIBUTE.  Returns 0 on success or -1 for errors.
-
-   After a successful call, you can query ATTRIBUTE by dwarf_whatform
-   to determine which of the dwarf_formX calls to make to get actual
-   value out of ATTRIBUTE.  Note that calling dwarf_whatattr is not
-   meaningful for pseudo-attributes formed this way.  */
-extern int dwarf_macro_param (Dwarf_Macro *macro, size_t idx,
-			      Dwarf_Attribute *attribute);
-
-/* Return macro parameter with index 0.  This will return -1 if the
-   parameter is not an integral value.  Use dwarf_macro_param for more
-   general access.  */
-extern int dwarf_macro_param1 (Dwarf_Macro *macro, Dwarf_Word *paramp)
-     __nonnull_attribute__ (2);
-
-/* Return macro parameter with index 1.  This will return -1 if the
-   parameter is not an integral or string value.  Use
-   dwarf_macro_param for more general access.  */
-extern int dwarf_macro_param2 (Dwarf_Macro *macro, Dwarf_Word *paramp,
-			       const char **strp);
-
-/* Compute what's known about a call frame when the PC is at ADDRESS.
-   Returns 0 for success or -1 for errors.
-   On success, *FRAME is a malloc'd pointer.  */
-extern int dwarf_cfi_addrframe (Dwarf_CFI *cache,
-				Dwarf_Addr address, Dwarf_Frame **frame)
-  __nonnull_attribute__ (3);
-
-/* Return the DWARF register number used in FRAME to denote
-   the return address in FRAME's caller frame.  The remaining
-   arguments can be non-null to fill in more information.
-
-   Fill [*START, *END) with the PC range to which FRAME's information applies.
-   Fill in *SIGNALP to indicate whether this is a signal-handling frame.
-   If true, this is the implicit call frame that calls a signal handler.
-   This frame's "caller" is actually the interrupted state, not a call;
-   its return address is an exact PC, not a PC after a call instruction.  */
-extern int dwarf_frame_info (Dwarf_Frame *frame,
-			     Dwarf_Addr *start, Dwarf_Addr *end, bool *signalp);
-
-/* Return a DWARF expression that yields the Canonical Frame Address at
-   this frame state.  Returns -1 for errors, or zero for success, with
-   *NOPS set to the number of operations stored at *OPS.  That pointer
-   can be used only as long as FRAME is alive and unchanged.  *NOPS is
-   zero if the CFA cannot be determined here.  Note that if nonempty,
-   *OPS is a DWARF expression, not a location description--append
-   DW_OP_stack_value to a get a location description for the CFA.  */
-extern int dwarf_frame_cfa (Dwarf_Frame *frame, Dwarf_Op **ops, size_t *nops)
-  __nonnull_attribute__ (2);
-
-/* Deliver a DWARF location description that yields the location or
-   value of DWARF register number REGNO in the state described by FRAME.
-
-   Returns -1 for errors or zero for success, setting *NOPS to the
-   number of operations in the array stored at *OPS.  Note the last
-   operation is DW_OP_stack_value if there is no mutable location but
-   only a computable value.
-
-   *NOPS zero with *OPS set to OPS_MEM means CFI says the caller's
-   REGNO is "undefined", i.e. it's call-clobbered and cannot be recovered.
-
-   *NOPS zero with *OPS set to a null pointer means CFI says the
-   caller's REGNO is "same_value", i.e. this frame did not change it;
-   ask the caller frame where to find it.
-
-   For common simple expressions *OPS is OPS_MEM (which is a caller
-   owned array for at least 3 Dwarf_Ops).  For arbitrary DWARF
-   expressions in the CFI, *OPS is an internal pointer that can be
-   used as long as the Dwarf_CFI used to create FRAME remains
-   alive.  */
-extern int dwarf_frame_register (Dwarf_Frame *frame, int regno,
-				 Dwarf_Op ops_mem[3],
-				 Dwarf_Op **ops, size_t *nops)
-  __nonnull_attribute__ (3, 4, 5);
-
-
-/* Return error code of last failing function call.  This value is kept
-   separately for each thread.  */
-extern int dwarf_errno (void);
-
-/* Return error string for ERROR.  If ERROR is zero, return error string
-   for most recent error or NULL is none occurred.  If ERROR is -1 the
-   behaviour is similar to the last case except that not NULL but a legal
-   string is returned.  */
-extern const char *dwarf_errmsg (int err);
-
-
-/* Register new Out-Of-Memory handler.  The old handler is returned.  */
-extern Dwarf_OOM dwarf_new_oom_handler (Dwarf *dbg, Dwarf_OOM handler);
-
-
-/* Inline optimizations.  */
-#ifdef __OPTIMIZE__
-/* Return attribute code of given attribute.  */
-__libdw_extern_inline unsigned int
-dwarf_whatattr (Dwarf_Attribute *attr)
-{
-  return attr == NULL ? 0 : attr->code;
-}
-
-/* Return attribute code of given attribute.  */
-__libdw_extern_inline unsigned int
-dwarf_whatform (Dwarf_Attribute *attr)
-{
-  return attr == NULL ? 0 : attr->form;
-}
-#endif	/* Optimize.  */
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif	/* libdw.h */
diff --git a/projects/llm_framework/include/libelf.h b/projects/llm_framework/include/libelf.h
deleted file mode 100644
index a139e733..00000000
--- a/projects/llm_framework/include/libelf.h
+++ /dev/null
@@ -1,522 +0,0 @@
-/* Interface for libelf.
-   Copyright (C) 1998-2010, 2015 Red Hat, Inc.
-   This file is part of elfutils.
-
-   This file is free software; you can redistribute it and/or modify
-   it under the terms of either
-
-     * the GNU Lesser General Public License as published by the Free
-       Software Foundation; either version 3 of the License, or (at
-       your option) any later version
-
-   or
-
-     * the GNU General Public License as published by the Free
-       Software Foundation; either version 2 of the License, or (at
-       your option) any later version
-
-   or both in parallel, as here.
-
-   elfutils is distributed in the hope that it will be useful, but
-   WITHOUT ANY WARRANTY; without even the implied warranty of
-   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
-   General Public License for more details.
-
-   You should have received copies of the GNU General Public License and
-   the GNU Lesser General Public License along with this program.  If
-   not, see <http://www.gnu.org/licenses/>.  */
-
-#ifndef _LIBELF_H
-#define _LIBELF_H 1
-
-#include <stdint.h>
-#include <sys/types.h>
-
-/* Get the ELF types.  */
-#include <elf.h>
-
-#ifndef SHF_COMPRESSED
- /* Older glibc elf.h might not yet define the ELF compression types.  */
- #define SHF_COMPRESSED      (1 << 11)  /* Section with compressed data. */
-
- /* Section compression header.  Used when SHF_COMPRESSED is set.  */
-
- typedef struct
- {
-   Elf32_Word   ch_type;        /* Compression format.  */
-   Elf32_Word   ch_size;        /* Uncompressed data size.  */
-   Elf32_Word   ch_addralign;   /* Uncompressed data alignment.  */
- } Elf32_Chdr;
-
- typedef struct
- {
-   Elf64_Word   ch_type;        /* Compression format.  */
-   Elf64_Word   ch_reserved;
-   Elf64_Xword  ch_size;        /* Uncompressed data size.  */
-   Elf64_Xword  ch_addralign;   /* Uncompressed data alignment.  */
- } Elf64_Chdr;
-
- /* Legal values for ch_type (compression algorithm).  */
- #define ELFCOMPRESS_ZLIB       1          /* ZLIB/DEFLATE algorithm.  */
- #define ELFCOMPRESS_LOOS       0x60000000 /* Start of OS-specific.  */
- #define ELFCOMPRESS_HIOS       0x6fffffff /* End of OS-specific.  */
- #define ELFCOMPRESS_LOPROC     0x70000000 /* Start of processor-specific.  */
- #define ELFCOMPRESS_HIPROC     0x7fffffff /* End of processor-specific.  */
-#endif
-
-#if __GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 3)
-# define __nonnull_attribute__(...) __attribute__ ((__nonnull__ (__VA_ARGS__)))
-# define __deprecated_attribute__ __attribute__ ((__deprecated__))
-# define __pure_attribute__ __attribute__ ((__pure__))
-# define __const_attribute__ __attribute__ ((__const__))
-#else
-# define __nonnull_attribute__(...)
-# define __deprecated_attribute__
-# define __pure_attribute__
-# define __const_attribute__
-#endif
-
-#if __GNUC__ < 4
-#define __noreturn_attribute__
-#else
-#define __noreturn_attribute__ __attribute__ ((noreturn))
-#endif
-
-#ifdef __GNUC_STDC_INLINE__
-# define __libdw_extern_inline extern __inline __attribute__ ((__gnu_inline__))
-#else
-# define __libdw_extern_inline extern __inline
-#endif
-
-/* Known translation types.  */
-typedef enum
-{
-  ELF_T_BYTE,                   /* unsigned char */
-  ELF_T_ADDR,                   /* Elf32_Addr, Elf64_Addr, ... */
-  ELF_T_DYN,                    /* Dynamic section record.  */
-  ELF_T_EHDR,                   /* ELF header.  */
-  ELF_T_HALF,                   /* Elf32_Half, Elf64_Half, ... */
-  ELF_T_OFF,                    /* Elf32_Off, Elf64_Off, ... */
-  ELF_T_PHDR,                   /* Program header.  */
-  ELF_T_RELA,                   /* Relocation entry with addend.  */
-  ELF_T_REL,                    /* Relocation entry.  */
-  ELF_T_SHDR,                   /* Section header.  */
-  ELF_T_SWORD,                  /* Elf32_Sword, Elf64_Sword, ... */
-  ELF_T_SYM,                    /* Symbol record.  */
-  ELF_T_WORD,                   /* Elf32_Word, Elf64_Word, ... */
-  ELF_T_XWORD,                  /* Elf32_Xword, Elf64_Xword, ... */
-  ELF_T_SXWORD,                 /* Elf32_Sxword, Elf64_Sxword, ... */
-  ELF_T_VDEF,                   /* Elf32_Verdef, Elf64_Verdef, ... */
-  ELF_T_VDAUX,                  /* Elf32_Verdaux, Elf64_Verdaux, ... */
-  ELF_T_VNEED,                  /* Elf32_Verneed, Elf64_Verneed, ... */
-  ELF_T_VNAUX,                  /* Elf32_Vernaux, Elf64_Vernaux, ... */
-  ELF_T_NHDR,                   /* Elf32_Nhdr, Elf64_Nhdr, ... */
-  ELF_T_SYMINFO,		/* Elf32_Syminfo, Elf64_Syminfo, ... */
-  ELF_T_MOVE,			/* Elf32_Move, Elf64_Move, ... */
-  ELF_T_LIB,			/* Elf32_Lib, Elf64_Lib, ... */
-  ELF_T_GNUHASH,		/* GNU-style hash section.  */
-  ELF_T_AUXV,			/* Elf32_auxv_t, Elf64_auxv_t, ... */
-  ELF_T_CHDR,			/* Compressed, Elf32_Chdr, Elf64_Chdr, ... */
-  ELF_T_NHDR8,			/* Special GNU Properties note.  Same as Nhdr,
-				   except padding.  */
-  /* Keep this the last entry.  */
-  ELF_T_NUM
-} Elf_Type;
-
-/* Descriptor for data to be converted to or from memory format.  */
-typedef struct
-{
-  void *d_buf;			/* Pointer to the actual data.  */
-  Elf_Type d_type;		/* Type of this piece of data.  */
-  unsigned int d_version;	/* ELF version.  */
-  size_t d_size;		/* Size in bytes.  */
-  int64_t d_off;		/* Offset into section.  */
-  size_t d_align;		/* Alignment in section.  */
-} Elf_Data;
-
-
-/* Commands for `...'.  */
-typedef enum
-{
-  ELF_C_NULL,			/* Nothing, terminate, or compute only.  */
-  ELF_C_READ,			/* Read .. */
-  ELF_C_RDWR,			/* Read and write .. */
-  ELF_C_WRITE,			/* Write .. */
-  ELF_C_CLR,			/* Clear flag.  */
-  ELF_C_SET,			/* Set flag.  */
-  ELF_C_FDDONE,			/* Signal that file descriptor will not be
-				   used anymore.  */
-  ELF_C_FDREAD,			/* Read rest of data so that file descriptor
-				   is not used anymore.  */
-  /* The following are extensions.  */
-  ELF_C_READ_MMAP,		/* Read, but mmap the file if possible.  */
-  ELF_C_RDWR_MMAP,		/* Read and write, with mmap.  */
-  ELF_C_WRITE_MMAP,		/* Write, with mmap.  */
-  ELF_C_READ_MMAP_PRIVATE,	/* Read, but memory is writable, results are
-				   not written to the file.  */
-  ELF_C_EMPTY,			/* Copy basic file data but not the content. */
-  /* Keep this the last entry.  */
-  ELF_C_NUM
-} Elf_Cmd;
-
-
-/* Flags for the ELF structures.  */
-enum
-{
-  ELF_F_DIRTY = 0x1,
-#define ELF_F_DIRTY		ELF_F_DIRTY
-  ELF_F_LAYOUT = 0x4,
-#define ELF_F_LAYOUT		ELF_F_LAYOUT
-  ELF_F_PERMISSIVE = 0x8
-#define ELF_F_PERMISSIVE	ELF_F_PERMISSIVE
-};
-
-/* Flags for elf_compress[_gnu].  */
-enum
-{
-  ELF_CHF_FORCE = 0x1
-#define ELF_CHF_FORCE ELF_CHF_FORCE
-};
-
-/* Identification values for recognized object files.  */
-typedef enum
-{
-  ELF_K_NONE,			/* Unknown.  */
-  ELF_K_AR,			/* Archive.  */
-  ELF_K_COFF,			/* Stupid old COFF.  */
-  ELF_K_ELF,			/* ELF file.  */
-  /* Keep this the last entry.  */
-  ELF_K_NUM
-} Elf_Kind;
-
-
-/* Archive member header.  */
-typedef struct
-{
-  char *ar_name;		/* Name of archive member.  */
-  time_t ar_date;		/* File date.  */
-  uid_t ar_uid;			/* User ID.  */
-  gid_t ar_gid;			/* Group ID.  */
-  mode_t ar_mode;		/* File mode.  */
-  int64_t ar_size;		/* File size.  */
-  char *ar_rawname;		/* Original name of archive member.  */
-} Elf_Arhdr;
-
-
-/* Archive symbol table entry.  */
-typedef struct
-{
-  char *as_name;		/* Symbol name.  */
-  size_t as_off;		/* Offset for this file in the archive.  */
-  unsigned long int as_hash;	/* Hash value of the name.  */
-} Elf_Arsym;
-
-
-/* Descriptor for the ELF file.  */
-typedef struct Elf Elf;
-
-/* Descriptor for ELF file section.  */
-typedef struct Elf_Scn Elf_Scn;
-
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* Return descriptor for ELF file to work according to CMD.  */
-extern Elf *elf_begin (int __fildes, Elf_Cmd __cmd, Elf *__ref);
-
-/* Create a clone of an existing ELF descriptor.  */
-  extern Elf *elf_clone (Elf *__elf, Elf_Cmd __cmd);
-
-/* Create descriptor for memory region.  */
-extern Elf *elf_memory (char *__image, size_t __size);
-
-/* Advance archive descriptor to next element.  */
-extern Elf_Cmd elf_next (Elf *__elf);
-
-/* Free resources allocated for ELF.  */
-extern int elf_end (Elf *__elf);
-
-/* Update ELF descriptor and write file to disk.  */
-extern int64_t elf_update (Elf *__elf, Elf_Cmd __cmd);
-
-/* Determine what kind of file is associated with ELF.  */
-extern Elf_Kind elf_kind (Elf *__elf) __pure_attribute__;
-
-/* Get the base offset for an object file.  */
-extern int64_t elf_getbase (Elf *__elf);
-
-
-/* Retrieve file identification data.  */
-extern char *elf_getident (Elf *__elf, size_t *__nbytes);
-
-/* Retrieve class-dependent object file header.  */
-extern Elf32_Ehdr *elf32_getehdr (Elf *__elf);
-/* Similar but this time the binary calls is ELFCLASS64.  */
-extern Elf64_Ehdr *elf64_getehdr (Elf *__elf);
-
-/* Create ELF header if none exists.  */
-extern Elf32_Ehdr *elf32_newehdr (Elf *__elf);
-/* Similar but this time the binary calls is ELFCLASS64.  */
-extern Elf64_Ehdr *elf64_newehdr (Elf *__elf);
-
-/* Get the number of program headers in the ELF file.  If the file uses
-   more headers than can be represented in the e_phnum field of the ELF
-   header the information from the sh_info field in the zeroth section
-   header is used.  */
-extern int elf_getphdrnum (Elf *__elf, size_t *__dst);
-
-/* Retrieve class-dependent program header table.  */
-extern Elf32_Phdr *elf32_getphdr (Elf *__elf);
-/* Similar but this time the binary calls is ELFCLASS64.  */
-extern Elf64_Phdr *elf64_getphdr (Elf *__elf);
-
-/* Create ELF program header.  */
-extern Elf32_Phdr *elf32_newphdr (Elf *__elf, size_t __cnt);
-/* Similar but this time the binary calls is ELFCLASS64.  */
-extern Elf64_Phdr *elf64_newphdr (Elf *__elf, size_t __cnt);
-
-
-/* Get section at INDEX.  */
-extern Elf_Scn *elf_getscn (Elf *__elf, size_t __index);
-
-/* Get section at OFFSET.  */
-extern Elf_Scn *elf32_offscn (Elf *__elf, Elf32_Off __offset);
-/* Similar bug this time the binary calls is ELFCLASS64.  */
-extern Elf_Scn *elf64_offscn (Elf *__elf, Elf64_Off __offset);
-
-/* Get index of section.  */
-extern size_t elf_ndxscn (Elf_Scn *__scn);
-
-/* Get section with next section index.  */
-extern Elf_Scn *elf_nextscn (Elf *__elf, Elf_Scn *__scn);
-
-/* Create a new section and append it at the end of the table.  */
-extern Elf_Scn *elf_newscn (Elf *__elf);
-
-/* Get the section index of the extended section index table for the
-   given symbol table.  */
-extern int elf_scnshndx (Elf_Scn *__scn);
-
-/* Get the number of sections in the ELF file.  If the file uses more
-   sections than can be represented in the e_shnum field of the ELF
-   header the information from the sh_size field in the zeroth section
-   header is used.  */
-extern int elf_getshdrnum (Elf *__elf, size_t *__dst);
-/* Sun messed up the implementation of 'elf_getshnum' in their implementation.
-   It was agreed to make the same functionality available under a different
-   name and obsolete the old name.  */
-extern int elf_getshnum (Elf *__elf, size_t *__dst)
-     __deprecated_attribute__;
-
-
-/* Get the section index of the section header string table in the ELF
-   file.  If the index cannot be represented in the e_shstrndx field of
-   the ELF header the information from the sh_link field in the zeroth
-   section header is used.  */
-extern int elf_getshdrstrndx (Elf *__elf, size_t *__dst);
-/* Sun messed up the implementation of 'elf_getshstrndx' in their
-   implementation.  It was agreed to make the same functionality available
-   under a different name and obsolete the old name.  */
-extern int elf_getshstrndx (Elf *__elf, size_t *__dst)
-     __deprecated_attribute__;
-
-
-/* Retrieve section header of ELFCLASS32 binary.  */
-extern Elf32_Shdr *elf32_getshdr (Elf_Scn *__scn);
-/* Similar for ELFCLASS64.  */
-extern Elf64_Shdr *elf64_getshdr (Elf_Scn *__scn);
-
-/* Returns compression header for a section if section data is
-   compressed.  Returns NULL and sets elf_errno if the section isn't
-   compressed or an error occurred.  */
-extern Elf32_Chdr *elf32_getchdr (Elf_Scn *__scn);
-extern Elf64_Chdr *elf64_getchdr (Elf_Scn *__scn);
-
-/* Compress or decompress the data of a section and adjust the section
-   header.
-
-   elf_compress works by setting or clearing the SHF_COMPRESS flag
-   from the section Shdr and will encode or decode a Elf32_Chdr or
-   Elf64_Chdr at the start of the section data.  elf_compress_gnu will
-   encode or decode any section, but is traditionally only used for
-   sections that have a name starting with ".debug" when
-   uncompressed or ".zdebug" when compressed and stores just the
-   uncompressed size.  The GNU compression method is deprecated and
-   should only be used for legacy support.
-
-   elf_compress takes a compression type that should be either zero to
-   decompress or an ELFCOMPRESS algorithm to use for compression.
-   Currently only ELFCOMPRESS_ZLIB is supported.  elf_compress_gnu
-   will compress in the traditional GNU compression format when
-   compress is one and decompress the section data when compress is
-   zero.
-
-   The FLAGS argument can be zero or ELF_CHF_FORCE.  If FLAGS contains
-   ELF_CHF_FORCE then it will always compress the section, even if
-   that would not reduce the size of the data section (including the
-   header).  Otherwise elf_compress and elf_compress_gnu will compress
-   the section only if the total data size is reduced.
-
-   On successful compression or decompression the function returns
-   one.  If (not forced) compression is requested and the data section
-   would not actually reduce in size, the section is not actually
-   compressed and zero is returned.  Otherwise -1 is returned and
-   elf_errno is set.
-
-   It is an error to request compression for a section that already
-   has SHF_COMPRESSED set, or (for elf_compress) to request
-   decompression for an section that doesn't have SHF_COMPRESSED set.
-   If a section has SHF_COMPRESSED set then calling elf_compress_gnu
-   will result in an error.  The section has to be decompressed first
-   using elf_compress.  Calling elf_compress on a section compressed
-   with elf_compress_gnu is fine, but probably useless.
-
-   It is always an error to call these functions on SHT_NOBITS
-   sections or if the section has the SHF_ALLOC flag set.
-   elf_compress_gnu will not check whether the section name starts
-   with ".debug" or .zdebug".  It is the responsibility of the caller
-   to make sure the deprecated GNU compression method is only called
-   on correctly named sections (and to change the name of the section
-   when using elf_compress_gnu).
-
-   All previous returned Shdrs and Elf_Data buffers are invalidated by
-   this call and should no longer be accessed.
-
-   Note that although this changes the header and data returned it
-   doesn't mark the section as dirty.  To keep the changes when
-   calling elf_update the section has to be flagged ELF_F_DIRTY.  */
-extern int elf_compress (Elf_Scn *scn, int type, unsigned int flags);
-extern int elf_compress_gnu (Elf_Scn *scn, int compress, unsigned int flags);
-
-/* Set or clear flags for ELF file.  */
-extern unsigned int elf_flagelf (Elf *__elf, Elf_Cmd __cmd,
-				 unsigned int __flags);
-/* Similarly for the ELF header.  */
-extern unsigned int elf_flagehdr (Elf *__elf, Elf_Cmd __cmd,
-				  unsigned int __flags);
-/* Similarly for the ELF program header.  */
-extern unsigned int elf_flagphdr (Elf *__elf, Elf_Cmd __cmd,
-				  unsigned int __flags);
-/* Similarly for the given ELF section.  */
-extern unsigned int elf_flagscn (Elf_Scn *__scn, Elf_Cmd __cmd,
-				 unsigned int __flags);
-/* Similarly for the given ELF data.  */
-extern unsigned int elf_flagdata (Elf_Data *__data, Elf_Cmd __cmd,
-				  unsigned int __flags);
-/* Similarly for the given ELF section header.  */
-extern unsigned int elf_flagshdr (Elf_Scn *__scn, Elf_Cmd __cmd,
-				  unsigned int __flags);
-
-
-/* Get data from section while translating from file representation to
-   memory representation.  The Elf_Data d_type is set based on the
-   section type if known.  Otherwise d_type is set to ELF_T_BYTE.  If
-   the section contains compressed data then d_type is always set to
-   ELF_T_CHDR.  */
-extern Elf_Data *elf_getdata (Elf_Scn *__scn, Elf_Data *__data);
-
-/* Get uninterpreted section content.  */
-extern Elf_Data *elf_rawdata (Elf_Scn *__scn, Elf_Data *__data);
-
-/* Create new data descriptor for section SCN.  */
-extern Elf_Data *elf_newdata (Elf_Scn *__scn);
-
-/* Get data translated from a chunk of the file contents as section data
-   would be for TYPE.  The resulting Elf_Data pointer is valid until
-   elf_end (ELF) is called.  */
-extern Elf_Data *elf_getdata_rawchunk (Elf *__elf,
-				       int64_t __offset, size_t __size,
-				       Elf_Type __type);
-
-
-/* Return pointer to string at OFFSET in section INDEX.  */
-extern char *elf_strptr (Elf *__elf, size_t __index, size_t __offset);
-
-
-/* Return header of archive.  */
-extern Elf_Arhdr *elf_getarhdr (Elf *__elf);
-
-/* Return offset in archive for current file ELF.  */
-extern int64_t elf_getaroff (Elf *__elf);
-
-/* Select archive element at OFFSET.  */
-extern size_t elf_rand (Elf *__elf, size_t __offset);
-
-/* Get symbol table of archive.  */
-extern Elf_Arsym *elf_getarsym (Elf *__elf, size_t *__narsyms);
-
-
-/* Control ELF descriptor.  */
-extern int elf_cntl (Elf *__elf, Elf_Cmd __cmd);
-
-/* Retrieve uninterpreted file contents.  */
-extern char *elf_rawfile (Elf *__elf, size_t *__nbytes);
-
-
-/* Return size of array of COUNT elements of the type denoted by TYPE
-   in the external representation.  The binary class is taken from ELF.
-   The result is based on version VERSION of the ELF standard.  */
-extern size_t elf32_fsize (Elf_Type __type, size_t __count,
-			   unsigned int __version)
-       __const_attribute__;
-/* Similar but this time the binary calls is ELFCLASS64.  */
-extern size_t elf64_fsize (Elf_Type __type, size_t __count,
-			   unsigned int __version)
-       __const_attribute__;
-
-
-/* Convert data structure from the representation in the file represented
-   by ELF to their memory representation.  */
-extern Elf_Data *elf32_xlatetom (Elf_Data *__dest, const Elf_Data *__src,
-				 unsigned int __encode);
-/* Same for 64 bit class.  */
-extern Elf_Data *elf64_xlatetom (Elf_Data *__dest, const Elf_Data *__src,
-				 unsigned int __encode);
-
-/* Convert data structure from to the representation in memory
-   represented by ELF file representation.  */
-extern Elf_Data *elf32_xlatetof (Elf_Data *__dest, const Elf_Data *__src,
-				 unsigned int __encode);
-/* Same for 64 bit class.  */
-extern Elf_Data *elf64_xlatetof (Elf_Data *__dest, const Elf_Data *__src,
-				 unsigned int __encode);
-
-
-/* Return error code of last failing function call.  This value is kept
-   separately for each thread.  */
-extern int elf_errno (void);
-
-/* Return error string for ERROR.  If ERROR is zero, return error string
-   for most recent error or NULL is none occurred.  If ERROR is -1 the
-   behaviour is similar to the last case except that not NULL but a legal
-   string is returned.  */
-extern const char *elf_errmsg (int __error);
-
-
-/* Coordinate ELF library and application versions.  */
-extern unsigned int elf_version (unsigned int __version);
-
-/* Set fill bytes used to fill holes in data structures.  */
-extern void elf_fill (int __fill);
-
-/* Compute hash value.  */
-extern unsigned long int elf_hash (const char *__string)
-       __pure_attribute__;
-
-/* Compute hash value using the GNU-specific hash function.  */
-extern unsigned long int elf_gnu_hash (const char *__string)
-       __pure_attribute__;
-
-
-/* Compute simple checksum from permanent parts of the ELF file.  */
-extern long int elf32_checksum (Elf *__elf);
-/* Similar but this time the binary calls is ELFCLASS64.  */
-extern long int elf64_checksum (Elf *__elf);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif  /* libelf.h */
diff --git a/projects/llm_framework/include/onnxruntime/core/common/basic_types.h b/projects/llm_framework/include/onnxruntime/core/common/basic_types.h
deleted file mode 100644
index 3eb48693..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/basic_types.h
+++ /dev/null
@@ -1,19 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <cstdint>
-
-namespace onnxruntime {
-
-/** A computed hash value. */
-using HashValue = uint64_t;
-
-/** The type of an argument (input or output).*/
-enum class ArgType : uint8_t {
-  kInput,
-  kOutput,
-};
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/code_location.h b/projects/llm_framework/include/onnxruntime/core/common/code_location.h
deleted file mode 100644
index 88aa6e44..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/code_location.h
+++ /dev/null
@@ -1,58 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <sstream>
-#include <string>
-#include <vector>
-
-namespace onnxruntime {
-/**
-   CodeLocation captures information on where in the source code a message came from.
-*/
-struct CodeLocation {
-  /**
-     @param file_path Usually the value of __FILE__
-     @param line Usually the value of __LINE__
-     @param func Usually the value of __PRETTY_FUNCTION__ or __FUNCTION__
-  */
-  CodeLocation(const char* file_path, const int line, const char* func)
-      : file_and_path{file_path}, line_num{line}, function{func} {
-  }
-
-  /**
-     @param file_path Usually the value of __FILE__
-     @param line Usually the value of __LINE__
-     @param func Usually the value of __PRETTY_FUNCTION__ or __FUNCTION__
-     @param stacktrace Stacktrace from source of message.
-  */
-  CodeLocation(const char* file_path, const int line, const char* func, const std::vector<std::string>& stacktrace)
-      : file_and_path{file_path}, line_num{line}, function{func}, stacktrace(stacktrace) {
-  }
-
-  std::string FileNoPath() const {
-    // assuming we always have work to do, so not trying to avoid creating a new string if
-    // no path was removed.
-    return file_and_path.substr(file_and_path.find_last_of("/\\") + 1);
-  }
-
-  enum Format {
-    kFilename,
-    kFilenameAndPath
-  };
-
-  std::string ToString(Format format = Format::kFilename) const {
-    std::ostringstream out;
-    out << (format == Format::kFilename ? FileNoPath() : file_and_path) << ":" << line_num << " " << function;
-    return out.str();
-  }
-  //utf-8. Because on Windows we compile our code with "/utf-8". And we assume the other platforms only use utf-8.
-  const std::string file_and_path;
-  const int line_num;
-  //utf-8
-  const std::string function;
-  const std::vector<std::string> stacktrace;
-};
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/common.h b/projects/llm_framework/include/onnxruntime/core/common/common.h
deleted file mode 100644
index 44b467cc..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/common.h
+++ /dev/null
@@ -1,287 +0,0 @@
-/**
- * Copyright (c) 2016-present, Facebook, Inc.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-// Portions Copyright (c) Microsoft Corporation
-
-#pragma once
-
-#include <climits>
-#include <cstring>
-#include <algorithm>
-#include <chrono>
-#include <functional>
-#include <memory>
-#include <numeric>
-#include <set>
-#include <sstream>
-#include <string>
-#include <type_traits>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include "core/common/code_location.h"
-#include "core/common/exceptions.h"
-#include "core/common/make_string.h"
-#include "core/common/status.h"
-
-
-namespace onnxruntime {
-
-using TimePoint = std::chrono::high_resolution_clock::time_point;
-
-#ifdef _WIN32
-#define ORT_UNUSED_PARAMETER(x) (x)
-#else
-#define ORT_UNUSED_PARAMETER(x) (void)(x)
-#endif
-
-#ifndef ORT_HAVE_ATTRIBUTE
-#ifdef __has_attribute
-#define ORT_HAVE_ATTRIBUTE(x) __has_attribute(x)
-#else
-#define ORT_HAVE_ATTRIBUTE(x) 0
-#endif
-#endif
-
-// ORT_ATTRIBUTE_UNUSED
-//
-// Prevents the compiler from complaining about or optimizing away variables
-// that appear unused on Linux
-#if ORT_HAVE_ATTRIBUTE(unused) || (defined(__GNUC__) && !defined(__clang__))
-#undef ORT_ATTRIBUTE_UNUSED
-#define ORT_ATTRIBUTE_UNUSED __attribute__((__unused__))
-#else
-#define ORT_ATTRIBUTE_UNUSED
-#endif
-
-#ifdef ORT_NO_EXCEPTIONS
-// Print the given final message, the message must be a null terminated char*
-// ORT will abort after printing the message.
-// For Android, will print to Android system log
-// For other platforms, will print to stderr
-void PrintFinalMessage(const char* msg);
-#endif
-
-// macro to explicitly ignore the return value from a function call so Code Analysis doesn't complain
-#define ORT_IGNORE_RETURN_VALUE(fn) \
-  static_cast<void>(fn)
-
-std::vector<std::string> GetStackTrace();
-// these is a helper function that gets defined by platform/Telemetry
-void LogRuntimeError(uint32_t session_id, const common::Status& status, const char* file,
-                     const char* function, uint32_t line);
-
-// __PRETTY_FUNCTION__ isn't a macro on gcc, so use a check for _MSC_VER
-// so we only define it as one for MSVC
-#if (_MSC_VER && !defined(__PRETTY_FUNCTION__))
-#define __PRETTY_FUNCTION__ __FUNCTION__
-#endif
-
-// Capture where a message is coming from. Use __FUNCTION__ rather than the much longer __PRETTY_FUNCTION__
-#define ORT_WHERE ::onnxruntime::CodeLocation(__FILE__, __LINE__, static_cast<const char*>(__FUNCTION__))
-
-#define ORT_WHERE_WITH_STACK \
-  ::onnxruntime::CodeLocation(__FILE__, __LINE__, static_cast<const char*>(__PRETTY_FUNCTION__), ::onnxruntime::GetStackTrace())
-
-#ifdef ORT_NO_EXCEPTIONS
-
-#define ORT_TRY if (true)
-#define ORT_CATCH(x) else if (false)
-#define ORT_RETHROW
-
-// In order to ignore the catch statement when a specific exception (not ... ) is caught and referred
-// in the body of the catch statements, it is necessary to wrap the body of the catch statement into
-// a lambda function. otherwise the exception referred will be undefined and cause build break
-#define ORT_HANDLE_EXCEPTION(func)
-
-// Throw an exception with optional message.
-// NOTE: The arguments get streamed into a string via ostringstream::operator<<
-// DO NOT use a printf format string, as that will not work as you expect.
-#define ORT_THROW(...)                                                    \
-  do {                                                                    \
-    ::onnxruntime::PrintFinalMessage(                                     \
-        ::onnxruntime::OnnxRuntimeException(                              \
-            ORT_WHERE_WITH_STACK, ::onnxruntime::MakeString(__VA_ARGS__)) \
-            .what());                                                     \
-    abort();                                                              \
-  } while (false)
-
-// Just in order to mark things as not implemented. Do not use in final code.
-#define ORT_NOT_IMPLEMENTED(...)                                                       \
-  do {                                                                                 \
-    ::onnxruntime::PrintFinalMessage(                                                  \
-        ::onnxruntime::NotImplementedException(::onnxruntime::MakeString(__VA_ARGS__)) \
-            .what());                                                                  \
-    abort();                                                                           \
-  } while (false)
-
-// Check condition.
-// NOTE: The arguments get streamed into a string via ostringstream::operator<<
-// DO NOT use a printf format string, as that will not work as you expect.
-#define ORT_ENFORCE(condition, ...)                                                   \
-  do {                                                                                \
-    if (!(condition)) {                                                               \
-      ::onnxruntime::PrintFinalMessage(                                               \
-          ::onnxruntime::OnnxRuntimeException(ORT_WHERE_WITH_STACK, #condition,       \
-                                              ::onnxruntime::MakeString(__VA_ARGS__)) \
-              .what());                                                               \
-      abort();                                                                        \
-    }                                                                                 \
-  } while (false)
-
-#define ORT_THROW_EX(ex, ...)                                                                      \
-  do {                                                                                             \
-    ::onnxruntime::PrintFinalMessage(                                                              \
-        ::onnxruntime::MakeString(#ex, "(", ::onnxruntime::MakeString(__VA_ARGS__), ")").c_str()); \
-    abort();                                                                                       \
-  } while (false)
-
-#else
-
-#define ORT_TRY try
-#define ORT_CATCH(x) catch (x)
-#define ORT_RETHROW throw;
-
-#define ORT_HANDLE_EXCEPTION(func) func()
-
-// Throw an exception with optional message.
-// NOTE: The arguments get streamed into a string via ostringstream::operator<<
-// DO NOT use a printf format string, as that will not work as you expect.
-#define ORT_THROW(...) \
-  throw ::onnxruntime::OnnxRuntimeException(ORT_WHERE_WITH_STACK, ::onnxruntime::MakeString(__VA_ARGS__))
-
-// Just in order to mark things as not implemented. Do not use in final code.
-#define ORT_NOT_IMPLEMENTED(...) \
-  throw ::onnxruntime::NotImplementedException(::onnxruntime::MakeString(__VA_ARGS__))
-
-// Check condition.
-// NOTE: The arguments get streamed into a string via ostringstream::operator<<
-// DO NOT use a printf format string, as that will not work as you expect.
-#define ORT_ENFORCE(condition, ...)                                                      \
-  do {                                                                                   \
-    if (!(condition)) {                                                                  \
-      throw ::onnxruntime::OnnxRuntimeException(ORT_WHERE_WITH_STACK, #condition,        \
-                                                ::onnxruntime::MakeString(__VA_ARGS__)); \
-    }                                                                                    \
-  } while (false)
-
-#define ORT_THROW_EX(ex, ...) \
-  throw ex(__VA_ARGS__)
-
-#endif
-
-#define ORT_MAKE_STATUS(category, code, ...)                     \
-  ::onnxruntime::common::Status(::onnxruntime::common::category, \
-                                ::onnxruntime::common::code,     \
-                                ::onnxruntime::MakeString(__VA_ARGS__))
-
-// Check condition. if met, return status.
-#define ORT_RETURN_IF(condition, ...)                                                                          \
-  do {                                                                                                         \
-    if (condition) {                                                                                           \
-      return ::onnxruntime::common::Status(::onnxruntime::common::ONNXRUNTIME,                                 \
-                                           ::onnxruntime::common::FAIL,                                        \
-                                           ::onnxruntime::MakeString(ORT_WHERE.ToString(), " ", __VA_ARGS__)); \
-    }                                                                                                          \
-  } while (false)
-
-// Check condition. if not met, return status.
-#define ORT_RETURN_IF_NOT(condition, ...) \
-  ORT_RETURN_IF(!(condition), __VA_ARGS__)
-
-// Macros to disable the copy and/or move ctor and assignment methods
-// These are usually placed in the private: declarations for a class.
-
-#define ORT_DISALLOW_COPY(TypeName) TypeName(const TypeName&) = delete
-
-#define ORT_DISALLOW_ASSIGNMENT(TypeName) TypeName& operator=(const TypeName&) = delete
-
-#define ORT_DISALLOW_COPY_AND_ASSIGNMENT(TypeName) \
-  ORT_DISALLOW_COPY(TypeName);                     \
-  ORT_DISALLOW_ASSIGNMENT(TypeName)
-
-#define ORT_DISALLOW_MOVE(TypeName) \
-  TypeName(TypeName&&) = delete;    \
-  TypeName& operator=(TypeName&&) = delete
-
-#define ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(TypeName) \
-  ORT_DISALLOW_COPY_AND_ASSIGNMENT(TypeName);           \
-  ORT_DISALLOW_MOVE(TypeName)
-
-#define ORT_RETURN_IF_ERROR_SESSIONID(expr, session_id)                                                                \
-  do {                                                                                                                 \
-    auto _status = (expr);                                                                                             \
-    if ((!_status.IsOK())) {                                                                                           \
-      ::onnxruntime::LogRuntimeError(session_id, _status, __FILE__, static_cast<const char*>(__FUNCTION__), __LINE__); \
-      return _status;                                                                                                  \
-    }                                                                                                                  \
-  } while (0)
-
-#define ORT_RETURN_IF_ERROR_SESSIONID_(expr) ORT_RETURN_IF_ERROR_SESSIONID(expr, session_id_)
-#define ORT_RETURN_IF_ERROR(expr) ORT_RETURN_IF_ERROR_SESSIONID(expr, 0)
-
-#define ORT_THROW_IF_ERROR(expr)                                                                              \
-  do {                                                                                                        \
-    auto _status = (expr);                                                                                    \
-    if ((!_status.IsOK())) {                                                                                  \
-      ::onnxruntime::LogRuntimeError(0, _status, __FILE__, static_cast<const char*>(__FUNCTION__), __LINE__); \
-      ORT_THROW(_status);                                                                                     \
-    }                                                                                                         \
-  } while (0)
-
-// use this macro when cannot early return
-#define ORT_CHECK_AND_SET_RETVAL(expr) \
-  do {                                 \
-    if (retval.IsOK()) {               \
-      retval = (expr);                 \
-    }                                  \
-  } while (0)
-
-inline long long TimeDiffMicroSeconds(TimePoint start_time) {
-  auto end_time = std::chrono::high_resolution_clock::now();
-  return std::chrono::duration_cast<std::chrono::microseconds>(end_time - start_time).count();
-}
-
-inline long long TimeDiffMicroSeconds(TimePoint start_time, TimePoint end_time) {
-  return std::chrono::duration_cast<std::chrono::microseconds>(end_time - start_time).count();
-}
-
-struct null_type {};
-inline std::string ToUTF8String(const std::string& s) { return s; }
-#ifdef _WIN32
-/**
- * Convert a wide character string to a UTF-8 string
- */
-std::string ToUTF8String(const std::wstring& s);
-
-std::wstring ToWideString(const std::string& s);
-inline std::wstring ToWideString(const std::wstring& s) { return s; }
-#else
-inline std::string ToWideString(const std::string& s) { return s; }
-#endif
-
-constexpr size_t kMaxStrLen = 2048;
-
-// Returns whether `key` is in `container`.
-// Like C++20's map/set contains() member function.
-template <typename Key, typename... OtherContainerArgs,
-          template <typename...> typename AssociativeContainer,
-          typename LookupKey>
-inline bool Contains(const AssociativeContainer<Key, OtherContainerArgs...>& container, LookupKey&& key) {
-  return container.find(std::forward<LookupKey>(key)) != container.end();
-}
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/const_pointer_container.h b/projects/llm_framework/include/onnxruntime/core/common/const_pointer_container.h
deleted file mode 100644
index 1d821ba6..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/const_pointer_container.h
+++ /dev/null
@@ -1,85 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <type_traits>
-
-namespace onnxruntime {
-/**
-   Container has T* entries. e.g. std::vector<T*>, and this class provides const access to those
-   via iterators and direct access, as the standard behavior only makes the pointer constant,
-   and not what is pointed too. i.e. you get a const pointer to T not a pointer to const T without this wrapper.
-   See https://stackoverflow.com/questions/8017036/understanding-const-iterator-with-pointers
-*/
-template <typename Container>
-class ConstPointerContainer {
- public:
-  using T = typename std::remove_pointer<typename Container::value_type>::type;
-
-  class ConstIterator {
-   public:
-    using const_iterator = typename Container::const_iterator;
-    using iterator_category = std::input_iterator_tag;
-    using value_type = T*;
-    using difference_type = std::ptrdiff_t;
-    using pointer = T**;
-    using reference = T*&;
-
-    /** Construct iterator for container that will return const T* entries.*/
-    explicit ConstIterator(const_iterator position) noexcept : current_{position}, item_{nullptr} {}
-    ConstIterator(const ConstIterator& other) = default;
-    ConstIterator& operator=(const ConstIterator& other) = default;
-
-    bool operator==(const ConstIterator& other) const noexcept { return current_ == other.current_; }
-    bool operator!=(const ConstIterator& other) const noexcept { return current_ != other.current_; }
-
-    ConstIterator& operator++() {
-      ++current_;
-      return *this;
-    }
-
-    ConstIterator operator++(int) {
-      ConstIterator tmp{*this};
-      ++(*this);
-      return tmp;
-    }
-
-    const T*& operator*() const {
-      item_ = *current_;
-      return item_;
-    }
-
-    const T** operator->() const { return &(operator*()); };
-
-   private:
-    const_iterator current_;
-    mutable const T* item_;
-  };
-
-  /**
-     Construct wrapper class that will provide const access to the pointers in a container of non-const pointers.
-     @param data Container with non-const pointers. e.g. std::vector<T*>
-  */
-  explicit ConstPointerContainer(const Container& data) noexcept : data_(data) {}
-
-  size_t size() const noexcept { return data_.size(); }
-  bool empty() const noexcept { return data_.empty(); }
-
-  ConstIterator cbegin() const noexcept { return ConstIterator(data_.cbegin()); }
-  ConstIterator cend() const noexcept { return ConstIterator(data_.cend()); }
-
-  ConstIterator begin() const noexcept { return ConstIterator(data_.cbegin()); }
-  ConstIterator end() const noexcept { return ConstIterator(data_.cend()); }
-
-  const T* operator[](size_t index) const { return data_[index]; }
-
-  const T* at(size_t index) const {
-    ORT_ENFORCE(index < data_.size());
-    return data_[index];
-  }
-
- private:
-  const Container& data_;
-};
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/denormal.h b/projects/llm_framework/include/onnxruntime/core/common/denormal.h
deleted file mode 100644
index ca944117..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/denormal.h
+++ /dev/null
@@ -1,12 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-namespace onnxruntime {
-
-// Set or unset flush-to-zero and denormal=as-zero if SSE3 instructions are supported.
-// Return true if SSE3 instruction is supported, otherwise return false.
-bool SetDenormalAsZero(bool on);
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/eigen_common_wrapper.h b/projects/llm_framework/include/onnxruntime/core/common/eigen_common_wrapper.h
deleted file mode 100644
index 4515f80f..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/eigen_common_wrapper.h
+++ /dev/null
@@ -1,62 +0,0 @@
-//-----------------------------------------------------------------------------
-//
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-//
-//-----------------------------------------------------------------------------
-#pragma once
-#include "onnxruntime_config.h"
-// build/external/eigen/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h:162:71:
-// error: ignoring attributes on template argument "Eigen::PacketType<const float, Eigen::DefaultDevice>::type {aka __vector(4) float}" [-Werror=ignored-attributes]
-#if defined(__GNUC__)
-#pragma GCC diagnostic push
-#if __GNUC__ >= 6
-#pragma GCC diagnostic ignored "-Wignored-attributes"
-#endif
-#pragma GCC diagnostic ignored "-Wunused-parameter"
-#pragma GCC diagnostic ignored "-Wunused-result"
-#ifdef HAS_DEPRECATED_COPY
-#pragma GCC diagnostic ignored "-Wdeprecated-copy"
-#endif
-// cmake/external/eigen/unsupported/Eigen/CXX11/../../../Eigen/src/Core/arch/NEON/PacketMath.h:1633:9:
-// error: ‘void* memcpy(void*, const void*, size_t)’ copying an object of non-trivial type ‘Eigen::internal::Packet4c’
-// {aka ‘struct Eigen::internal::eigen_packet_wrapper<int, 2>’} from an array of ‘const int8_t’
-// {aka ‘const signed char’} [-Werror=class-memaccess]
-#ifdef HAS_CLASS_MEMACCESS
-#pragma GCC diagnostic ignored "-Wclass-memaccess"
-#endif
-
-// cmake/external/eigen\Eigen/src/Core/util/Meta.h:454:25:
-// error: 'result_of<Eigen::internal::scalar_product_op<unsigned long long> (const unsigned long long &, const unsigned long long &)>'
-// is deprecated [-Werror,-Wdeprecated-declarations]
-//   typedef typename std::result_of<T>::type type1;
-#ifdef HAS_DEPRECATED_DECLARATIONS
-#pragma GCC diagnostic ignored "-Wdeprecated-declarations"
-#endif
-
-// cmake/external/eigen\Eigen/CXX11/src/Tensor/TensorTrace.h:130:9:
-// error: variable 'num_distinct_reduce_dims' set but not used [-Werror,-Wunused-but-set-variable]
-//   int num_distinct_reduce_dims = 0;
-#ifdef HAS_UNUSED_BUT_SET_VARIABLE
-#pragma GCC diagnostic ignored "-Wunused-but-set-variable"
-#endif
-
-#elif defined(_MSC_VER)
-// build\windows\debug\external\eigen3\unsupported\eigen\cxx11\src/Tensor/Tensor.h(76):
-// warning C4554: '&': check operator precedence for possible error; use parentheses to clarify precedence
-
-// unsupported\eigen\cxx11\src\Tensor\TensorUInt128.h(150,0): Warning C4245: 'initializing': conversion from '__int64'
-// to 'uint64_t', signed/unsigned mismatch
-#pragma warning(push)
-#pragma warning(disable : 4554)
-#pragma warning(disable : 4245)
-#pragma warning(disable : 4127)
-#endif
-
-#include "unsupported/Eigen/CXX11/Tensor"
-
-#if defined(__GNUC__)
-#pragma GCC diagnostic pop
-#elif defined(_MSC_VER)
-#pragma warning(pop)
-#endif
diff --git a/projects/llm_framework/include/onnxruntime/core/common/exceptions.h b/projects/llm_framework/include/onnxruntime/core/common/exceptions.h
deleted file mode 100644
index cbebc88c..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/exceptions.h
+++ /dev/null
@@ -1,71 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <algorithm>
-#include <exception>
-#include <iterator>
-#include <stdexcept>
-#include <string>
-#include <vector>
-
-#include "core/common/common.h"
-#include "core/common/code_location.h"
-
-namespace onnxruntime {
-
-class NotImplementedException : public std::logic_error {
- public:
-  explicit NotImplementedException(const char* _Message = "Function not yet implemented") noexcept : std::logic_error(_Message){};
-  explicit NotImplementedException(const std::string& _Message = "Function not yet implemented") noexcept : std::logic_error(_Message){};
-};
-
-class TypeMismatchException : public std::logic_error {
- public:
-  TypeMismatchException() noexcept : logic_error("Type mismatch"){};
-};
-
-class OnnxRuntimeException : public std::exception {
- public:
-  OnnxRuntimeException(const CodeLocation& location, const std::string& msg) noexcept
-      : OnnxRuntimeException(location, nullptr, msg) {
-  }
-
-  /**
-     Create a new exception that captures the location it was thrown from.
-     @param location Location in the source code the exception is being thrown from
-     @param failed_condition Optional string containing the condition that failed.
-     e.g. "tensor.Size() == input.Size()". May be nullptr.
-     @param msg Message containing additional information about the exception cause.
-  */
-  OnnxRuntimeException(const CodeLocation& location, const char* failed_condition, const std::string& msg)
-      : location_{location} {
-    std::ostringstream ss;
-
-    ss << location.ToString(CodeLocation::kFilenameAndPath);  // output full path in case just the filename is ambiguous
-    if (failed_condition != nullptr) {
-      ss << " " << failed_condition << " was false.";
-    }
-
-    ss << " " << msg << "\n";
-    if (!location.stacktrace.empty()) {
-      ss << "Stacktrace:\n";
-      // skip the first entry in the stacktrace as we have that information from location.ToString()
-      std::copy(++location.stacktrace.begin(), location.stacktrace.end(), std::ostream_iterator<std::string>(ss, "\n"));
-    }
-
-    what_ = ss.str();
-  }
-
-  const char* what() const noexcept override {
-    return what_.c_str();
-  }
-
- private:
-  const CodeLocation location_;
-  const std::vector<std::string> stacktrace_;
-  std::string what_;
-};
-
-}  // namespace onnxruntime
\ No newline at end of file
diff --git a/projects/llm_framework/include/onnxruntime/core/common/gpu_profiler_common.h b/projects/llm_framework/include/onnxruntime/core/common/gpu_profiler_common.h
deleted file mode 100644
index 00d5033e..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/gpu_profiler_common.h
+++ /dev/null
@@ -1,472 +0,0 @@
-#pragma once
-
-#include "core/common/profiler_common.h"
-#include "core/common/inlined_containers.h"
-
-#include <map>
-#include <memory>
-#include <mutex>
-#include <sstream>
-#include <string>
-#include <vector>
-#include <utility>
-
-namespace onnxruntime {
-namespace profiling {
-
-// The classes in this header are implemented as template/inline classes
-// to avoid having to export symbols from the main onnxruntime shared library
-// to ExecutionProvider (EP) shared libraries.
-// More context: The main onnxruntime shared library is optimized for size
-// using --gc-sections during link time to ensure that any unreferenced code
-// is not retained. This poses a problem in using a design pattern where the
-// (abstract) base class is implemented in the main onnxruntime shared library,
-// but (concrete) subclasses are implemented in EP shared libraries. Now, because
-// EP shared libraries are loaded at runtime (as of 11/2022), there will be no
-// references to the base class symbols when the main onnxruntime shared library
-// is compiled. Thus, the base class symbols will not be included in the
-// main onnxruntime shared library. This manifests in being unable to load
-// EP shared libs (because the base class symbols referenced by derived
-// classes are missing).
-// We solve this by implementing base classes that are common to all GPU profilers
-// inline in this header.
-
-class ProfilerActivityBuffer {
- public:
-  ProfilerActivityBuffer() noexcept
-      : data_(nullptr), size_(0) {}
-
-  ProfilerActivityBuffer(const char* data, size_t size) noexcept
-      : data_(std::make_unique<char[]>(size)), size_(size) {
-    memcpy(data_.get(), data, size_);
-  }
-
-  ProfilerActivityBuffer(const ProfilerActivityBuffer& other) noexcept
-      : ProfilerActivityBuffer(other.GetData(), other.GetSize()) {}
-
-  ProfilerActivityBuffer(ProfilerActivityBuffer&& other) noexcept
-      : ProfilerActivityBuffer() {
-    std::swap(data_, other.data_);
-    std::swap(size_, other.size_);
-  }
-
-  ProfilerActivityBuffer& operator=(const ProfilerActivityBuffer& other) noexcept {
-    if (&other == this) {
-      return *this;
-    }
-
-    new (this) ProfilerActivityBuffer{other};
-    return *this;
-  }
-
-  ProfilerActivityBuffer& operator=(ProfilerActivityBuffer&& other) noexcept {
-    if (&other == this) {
-      return *this;
-    }
-
-    new (this) ProfilerActivityBuffer{std::move(other)};
-    return *this;
-  }
-
-  static ProfilerActivityBuffer CreateFromPreallocatedBuffer(std::unique_ptr<char[]>&& buffer_ptr, size_t size) {
-    ProfilerActivityBuffer res{};
-    res.data_ = std::move(buffer_ptr);
-    res.size_ = size;
-    return res;
-  }
-
-  // accessors
-  char* GetData() { return data_.get(); }
-  const char* GetData() const { return data_.get(); }
-  size_t GetSize() const { return size_; }
-
- private:
-  std::unique_ptr<char[]> data_;
-  size_t size_;
-}; /* end class ProfilerActivityBuffer */
-
-template <typename TDerived>
-class GPUTracerManager {
- public:
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(GPUTracerManager);
-  virtual ~GPUTracerManager() {}
-
-  uint64_t RegisterClient() {
-    std::lock_guard<std::mutex> lock(manager_instance_mutex_);
-    auto res = next_client_id_++;
-    per_client_events_by_ext_correlation_.insert({res, {}});
-    ++num_active_clients_;
-    return res;
-  }
-
-  void DeregisterClient(uint64_t client_handle) {
-    std::lock_guard<std::mutex> lock(manager_instance_mutex_);
-    auto it = per_client_events_by_ext_correlation_.find(client_handle);
-    if (it == per_client_events_by_ext_correlation_.end()) {
-      return;
-    }
-    per_client_events_by_ext_correlation_.erase(it);
-    --num_active_clients_;
-    if (num_active_clients_ == 0 && tracing_enabled_) {
-      StopLogging();
-    }
-  }
-
-  void StartLogging() {
-    std::lock_guard<std::mutex> lock(manager_instance_mutex_);
-    if (tracing_enabled_) {
-      return;
-    }
-
-    auto this_as_derived = static_cast<TDerived*>(this);
-    tracing_enabled_ = this_as_derived->OnStartLogging();
-  }
-
-  void Consume(uint64_t client_handle, const TimePoint& start_time, std::map<uint64_t, Events>& events) {
-    auto this_as_derived = static_cast<TDerived*>(this);
-    events.clear();
-    {
-      // Flush any pending activity records before starting
-      // to process the accumulated activity records.
-      std::lock_guard<std::mutex> lock_manager(manager_instance_mutex_);
-      if (!tracing_enabled_) {
-        return;
-      }
-
-      this_as_derived->FlushActivities();
-    }
-
-    std::vector<ProfilerActivityBuffer> activity_buffers;
-    {
-      std::lock_guard<std::mutex> lock(unprocessed_activity_buffers_mutex_);
-      std::swap(unprocessed_activity_buffers_, activity_buffers);
-      unprocessed_activity_buffers_.clear();
-    }
-
-    {
-      // Ensure that at most one thread is working through the activity buffers at any time.
-      std::lock_guard<std::mutex> lock_two(activity_buffer_processor_mutex_);
-      this_as_derived->ProcessActivityBuffers(activity_buffers, start_time);
-      auto it = per_client_events_by_ext_correlation_.find(client_handle);
-      if (it == per_client_events_by_ext_correlation_.end()) {
-        return;
-      }
-      std::swap(events, it->second);
-    }
-  }
-
-  void PushCorrelation(uint64_t client_handle,
-                       uint64_t external_correlation_id,
-                       TimePoint profiling_start_time) {
-    auto this_as_derived = static_cast<TDerived*>(this);
-    std::lock_guard<std::mutex> lock(manager_instance_mutex_);
-    if (!tracing_enabled_) {
-      return;
-    }
-
-    auto it = per_client_events_by_ext_correlation_.find(client_handle);
-    if (it == per_client_events_by_ext_correlation_.end()) {
-      // not a registered client, do nothing
-      return;
-    }
-
-    // external_correlation_id is simply the timestamp of this event,
-    // relative to profiling_start_time. i.e., it was computed as:
-    // external_correlation_id =
-    //      std::chrono::duration_cast<std::chrono::microseconds>(event_start_time - profiling_start_time).count()
-    //
-    // Because of the relative nature of the external_correlation_id, the same
-    // external_correlation_id can be reused across different clients, which then makes it
-    // impossible to recover the client from the external_correlation_id, which in turn
-    // makes it impossible to map events (which are tagged with external_correlation_id) to clients.
-    //
-    // To address these difficulties, we construct a new correlation_id (let's call it unique_cid)
-    // as follows:
-    // unique_cid =
-    //    external_correlation_id +
-    //    std::chrono::duration_cast<std::chrono::microseconds>(profiling_start_time.time_since_epoch()).count()
-    // now, unique_cid is monotonically increasing with time, so it can be used to reliably map events to clients.
-    //
-    // Of course, clients expect lists of events to be returned (on a call to Consume()), that are
-    // still keyed on the external_correlation_id that they've specified here, so we need to remember the
-    // offset to be subtracted
-    uint64_t offset = std::chrono::duration_cast<std::chrono::microseconds>(profiling_start_time.time_since_epoch()).count();
-    auto unique_cid = external_correlation_id + offset;
-    unique_correlation_id_to_client_offset_[unique_cid] = std::make_pair(client_handle, offset);
-    this_as_derived->PushUniqueCorrelation(unique_cid);
-  }
-
-  void PopCorrelation(uint64_t& popped_external_correlation_id) {
-    auto this_as_derived = static_cast<TDerived*>(this);
-    std::lock_guard<std::mutex> lock(manager_instance_mutex_);
-    if (!tracing_enabled_) {
-      return;
-    }
-    uint64_t unique_cid;
-    this_as_derived->PopUniqueCorrelation(unique_cid);
-    // lookup the offset and subtract it before returning popped_external_correlation_id to the client
-    auto client_it = unique_correlation_id_to_client_offset_.find(unique_cid);
-    if (client_it == unique_correlation_id_to_client_offset_.end()) {
-      popped_external_correlation_id = 0;
-      return;
-    }
-    popped_external_correlation_id = unique_cid - client_it->second.second;
-  }
-
-  void PopCorrelation() {
-    uint64_t unused;
-    PopCorrelation(unused);
-  }
-
- protected:
-  GPUTracerManager() {
-    auto this_as_derived = static_cast<TDerived*>(this);
-    uint64_t gpu_ts1, gpu_ts2, cpu_ts;
-
-    // Get the CPU and GPU timestamps to warm up
-    gpu_ts1 = this_as_derived->GetGPUTimestampInNanoseconds();
-    cpu_ts = this->GetCPUTimestampInNanoseconds();
-
-    // Estimate the skew/offset between the CPU and GPU timestamps.
-    gpu_ts1 = this_as_derived->GetGPUTimestampInNanoseconds();
-    cpu_ts = this->GetCPUTimestampInNanoseconds();
-    gpu_ts2 = this_as_derived->GetGPUTimestampInNanoseconds();
-
-    auto gpu_ts = (gpu_ts1 + gpu_ts2) / 2;
-    offset_to_add_to_gpu_timestamps_ = cpu_ts - gpu_ts;
-  }
-
-#if 0
-  // Functional API to be implemented by subclasses
-  // Included here only for documentation purposes
-protected:
-  bool OnStartLogging();
-  void OnStopLogging();
-  void ProcessActivityBuffers(const std::vector<ProfilerActivityBuffer>& buffers,
-                              const TimePoint& start_time);
-  bool PushUniqueCorrelation(uint64_t unique_cid);
-  void PopUniqueCorrelation(uint64_t& popped_unique_cid);
-  void FlushActivities();
-  uint64_t GetGPUTimestampInNanoseconds();
-#endif
-
-  void EnqueueActivityBuffer(ProfilerActivityBuffer&& buffer) {
-    std::lock_guard<std::mutex> lock(unprocessed_activity_buffers_mutex_);
-    unprocessed_activity_buffers_.emplace_back(std::move(buffer));
-  }
-
-  // To be called by subclasses only from ProcessActivityBuffers
-  void MapEventToClient(uint64_t tracer_correlation_id, EventRecord&& event) {
-    auto it = tracer_correlation_to_unique_correlation_.find(tracer_correlation_id);
-    if (it == tracer_correlation_to_unique_correlation_.end()) {
-      // We're yet to receive a mapping to unique_correlation_id for this tracer_correlation_id
-      DeferEventMapping(std::move(event), tracer_correlation_id);
-      return;
-    }
-    auto unique_correlation_id = it->second;
-    auto p_event_list = GetEventListForUniqueCorrelationId(unique_correlation_id);
-    if (p_event_list != nullptr) {
-      p_event_list->emplace_back(std::move(event));
-    }
-  }
-
-  // To be called by subclasses only from ProcessActivityBuffers
-  void NotifyNewCorrelation(uint64_t tracer_correlation_id, uint64_t unique_correlation_id) {
-    tracer_correlation_to_unique_correlation_[tracer_correlation_id] = unique_correlation_id;
-    auto pending_it = events_pending_client_mapping_.find(tracer_correlation_id);
-    if (pending_it == events_pending_client_mapping_.end()) {
-      return;
-    }
-    // Map the pending events to the right client
-    MapEventsToClient(unique_correlation_id, std::move(pending_it->second));
-    events_pending_client_mapping_.erase(pending_it);
-  }
-
-  uint64_t NormalizeGPUTimestampToCPUEpoch(uint64_t gpu_timestamp_in_nanoseconds) {
-    return gpu_timestamp_in_nanoseconds + this->offset_to_add_to_gpu_timestamps_;
-  }
-
- private:
-  // Requires: manager_instance_mutex_ should be held
-  void StopLogging() {
-    auto this_as_derived = static_cast<TDerived*>(this);
-    if (!tracing_enabled_) {
-      return;
-    }
-    this_as_derived->OnStopLogging();
-    tracing_enabled_ = false;
-    Clear();
-  }
-
-  // Requires: manager_instance_mutex_ should be held
-  void Clear() {
-    unprocessed_activity_buffers_.clear();
-    unique_correlation_id_to_client_offset_.clear();
-    per_client_events_by_ext_correlation_.clear();
-    tracer_correlation_to_unique_correlation_.clear();
-    events_pending_client_mapping_.clear();
-  }
-
-  Events* GetEventListForUniqueCorrelationId(uint64_t unique_correlation_id) {
-    auto client_it = unique_correlation_id_to_client_offset_.find(unique_correlation_id);
-    if (client_it == unique_correlation_id_to_client_offset_.end()) {
-      return nullptr;
-    }
-
-    // See the comments on the GetUniqueCorrelationId method for an explanation of
-    // of this offset computation and why it's required.
-    auto const& client_handle_offset = client_it->second;
-    auto external_correlation = unique_correlation_id - client_handle_offset.second;
-    auto& event_list = per_client_events_by_ext_correlation_[client_handle_offset.first][external_correlation];
-    return &event_list;
-  }
-
-  void MapEventsToClient(uint64_t unique_correlation_id, std::vector<EventRecord>&& events) {
-    auto p_event_list = GetEventListForUniqueCorrelationId(unique_correlation_id);
-    if (p_event_list != nullptr) {
-      p_event_list->insert(p_event_list->end(),
-                           std::make_move_iterator(events.begin()),
-                           std::make_move_iterator(events.end()));
-    }
-  }
-
-  void DeferEventMapping(EventRecord&& event, uint64_t tracer_correlation_id) {
-    events_pending_client_mapping_[tracer_correlation_id].emplace_back(std::move(event));
-  }
-
-  uint64_t GetCPUTimestampInNanoseconds() {
-    return std::chrono::duration_cast<std::chrono::nanoseconds>(
-               std::chrono::high_resolution_clock::now().time_since_epoch())
-        .count();
-  }
-
-  std::mutex manager_instance_mutex_;
-  uint64_t next_client_id_ = 1;
-  uint64_t num_active_clients_ = 0;
-  bool tracing_enabled_ = false;
-  std::mutex unprocessed_activity_buffers_mutex_;
-  std::mutex activity_buffer_processor_mutex_;
-
-  // Unprocessed activity buffers
-  std::vector<ProfilerActivityBuffer> unprocessed_activity_buffers_;
-
-  // Keyed on unique_correlation_id -> (client_id/client_handle, offset)
-  // unique_correlation_id - offset == external_correlation_id
-  InlinedHashMap<uint64_t, std::pair<uint64_t, uint64_t>> unique_correlation_id_to_client_offset_;
-
-  // Keyed on tracer_correlation_id -> unique_correlation_id
-  InlinedHashMap<uint64_t, uint64_t> tracer_correlation_to_unique_correlation_;
-
-  // client_id/client_handle -> external_correlation_id -> events
-  InlinedHashMap<uint64_t, std::map<uint64_t, Events>> per_client_events_by_ext_correlation_;
-
-  // Keyed on tracer correlation_id, keeps track of activity records
-  // for which we haven't established the external_correlation_id yet.
-  InlinedHashMap<uint64_t, std::vector<EventRecord>> events_pending_client_mapping_;
-
-  // An offset to add to (the possibly skewed) GPU timestamps
-  // to normalize GPU timestamps with CPU timestamps
-  int64_t offset_to_add_to_gpu_timestamps_;
-}; /* class GPUTracerManager */
-
-// Base class for a GPU profiler
-template <typename TManager>
-class GPUProfilerBase : public EpProfiler {
- protected:
-  GPUProfilerBase() = default;
-  virtual ~GPUProfilerBase() {}
-
-  void MergeEvents(std::map<uint64_t, Events>& events_to_merge, Events& events) {
-    Events merged_events;
-
-    auto event_iter = std::make_move_iterator(events.begin());
-    auto event_end = std::make_move_iterator(events.end());
-    for (auto& map_iter : events_to_merge) {
-      if (map_iter.second.empty()) {
-        continue;
-      }
-
-      auto ts = static_cast<long long>(map_iter.first);
-
-      // find the last occurrence of a matching timestamp,
-      // if one exists
-      while (event_iter != event_end &&
-             (event_iter->ts < ts ||
-              (event_iter->ts == ts &&
-               (event_iter + 1) != event_end &&
-               (event_iter + 1)->ts == ts))) {
-        merged_events.emplace_back(*event_iter);
-        ++event_iter;
-      }
-
-      bool copy_op_names = false;
-      std::string op_name;
-      std::string parent_name;
-
-      if (event_iter != event_end && event_iter->ts == ts) {
-        // We've located a parent event, copy the op_name and set
-        // this event's parent_name property to the name of the parent.
-        copy_op_names = true;
-        op_name = event_iter->args["op_name"];
-        parent_name = event_iter->name;
-        merged_events.emplace_back(*event_iter);
-        ++event_iter;
-      }
-
-      for (auto& evt : map_iter.second) {
-        if (copy_op_names) {
-          // If we have found a matching parent event,
-          // then inherit some names from the parent.
-          evt.args["op_name"] = op_name;
-          evt.args["parent_name"] = parent_name;
-        }
-      }
-
-      merged_events.insert(merged_events.end(),
-                           std::make_move_iterator(map_iter.second.begin()),
-                           std::make_move_iterator(map_iter.second.end()));
-    }
-
-    // move any remaining events
-    merged_events.insert(merged_events.end(), event_iter, event_end);
-    std::swap(events, merged_events);
-  }
-
-  uint64_t client_handle_;
-  TimePoint profiling_start_time_;
-
- public:
-  virtual bool StartProfiling(TimePoint profiling_start_time) override {
-    auto& manager = TManager::GetInstance();
-    manager.StartLogging();
-    profiling_start_time_ = profiling_start_time;
-    return true;
-  }
-
-  virtual void EndProfiling(TimePoint start_time, Events& events) override {
-    auto& manager = TManager::GetInstance();
-    std::map<uint64_t, Events> event_map;
-    manager.Consume(client_handle_, start_time, event_map);
-    MergeEvents(event_map, events);
-  }
-
-  virtual void Start(uint64_t id) override {
-    auto& manager = TManager::GetInstance();
-    manager.PushCorrelation(client_handle_, id, profiling_start_time_);
-  }
-
-  virtual void Stop(uint64_t) override {
-    auto& manager = TManager::GetInstance();
-    manager.PopCorrelation();
-  }
-}; /* class GPUProfilerBase */
-
-// Convert a pointer to a hex string
-static inline std::string PointerToHexString(const void* ptr) {
-  std::ostringstream sstr;
-  sstr << std::hex << ptr;
-  return sstr.str();
-}
-
-} /* end namespace profiling */
-} /* end namespace onnxruntime */
diff --git a/projects/llm_framework/include/onnxruntime/core/common/gsl.h b/projects/llm_framework/include/onnxruntime/core/common/gsl.h
deleted file mode 100644
index 371c5b75..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/gsl.h
+++ /dev/null
@@ -1,6 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include "gsl/gsl"
diff --git a/projects/llm_framework/include/onnxruntime/core/common/hash_combine.h b/projects/llm_framework/include/onnxruntime/core/common/hash_combine.h
deleted file mode 100644
index 5662a329..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/hash_combine.h
+++ /dev/null
@@ -1,21 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-namespace onnxruntime {
-
-// Combine hash value `seed` with hash value `h`, updating `seed` in place.
-// TODO(edgchen1) find a better implementation? e.g., see a more recent version of boost::hash_combine()
-inline void HashCombineWithHashValue(size_t h, size_t& seed) {
-  seed ^= h + 0x9e3779b9 + (seed << 6) + (seed >> 2);
-}
-
-// Combine hash value `seed` with the hash value of `value`, updating `seed` in place.
-// The hash value computation is specified by the `Hash` template parameter.
-template <typename T, typename Hash = std::hash<T>>
-inline void HashCombine(const T& value, size_t& seed) {
-  HashCombineWithHashValue(Hash{}(value), seed);
-}
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/inlined_containers.h b/projects/llm_framework/include/onnxruntime/core/common/inlined_containers.h
deleted file mode 100644
index 3ef9c28e..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/inlined_containers.h
+++ /dev/null
@@ -1,175 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <cmath>
-
-#include "core/common/inlined_containers_fwd.h"
-
-#ifndef DISABLE_ABSEIL
-
-#ifdef _MSC_VER
-#pragma warning(push)
-// C4127: conditional expression is constant
-#pragma warning(disable : 4127)
-// C4324: structure was padded due to alignment specifier
-// Usage of alignas causes some internal padding in places.
-#pragma warning(disable : 4324)
-#endif  // _MSC_VER
-
-#include <absl/container/flat_hash_set.h>
-#include <absl/container/flat_hash_map.h>
-
-#include <absl/container/node_hash_set.h>
-#include <absl/container/node_hash_map.h>
-
-#ifdef _MSC_VER
-#pragma warning(pop)
-#endif  // _MSC_VER
-
-#else  // DISABLE_ABSEIL
-
-#include <unordered_set>
-#include <unordered_map>
-#include <set>
-#include <map>
-
-#endif  // DISABLE_ABSEIL
-
-namespace onnxruntime {
-
-#ifndef DISABLE_ABSEIL
-// InlinedHashSet and InlinedHashMap are preferred
-// hash based containers. They store their values in the
-// buckets array that is allocated in one shot. It eliminates
-// per-node new/delete calls. Always call reserve() on any hash set/map
-// when the number of items is known in advance.
-// This does not allocate a dummy 'end' node on default construction.
-template <typename T, typename Allocator>
-class InlinedHashSet : public absl::flat_hash_set<T,
-                                                  absl::container_internal::hash_default_hash<T>,
-                                                  absl::container_internal::hash_default_eq<T>,
-                                                  Allocator> {
-  using Base = absl::flat_hash_set<T,
-                                   absl::container_internal::hash_default_hash<T>,
-                                   absl::container_internal::hash_default_eq<T>,
-                                   Allocator>;
-
- public:
-  using Base::Base;
-};
-
-template <typename Key, typename Value,
-          typename Allocator>
-class InlinedHashMap : public absl::flat_hash_map<Key, Value,
-                                                  absl::container_internal::hash_default_hash<Key>,
-                                                  absl::container_internal::hash_default_eq<Key>,
-                                                  Allocator> {
-  using Base = absl::flat_hash_map<Key, Value,
-                                   absl::container_internal::hash_default_hash<Key>,
-                                   absl::container_internal::hash_default_eq<Key>,
-                                   Allocator>;
-
- public:
-  using Base::Base;
-};
-
-// Use this hash set/map where pointer stability is required, otherwise use
-// InlinedHashSet and InlinedHashMap
-// This does not allocate a dummy 'end' node on default construction.
-// Use reserve() when the number of elements is known.
-template <typename T, typename Allocator>
-class NodeHashSet : public absl::node_hash_set<T,
-                                               absl::container_internal::hash_default_hash<T>,
-                                               absl::container_internal::hash_default_eq<T>,
-                                               Allocator> {
-  using Base = absl::node_hash_set<T,
-                                   absl::container_internal::hash_default_hash<T>,
-                                   absl::container_internal::hash_default_eq<T>,
-                                   Allocator>;
-
- public:
-  using Base::Base;
-};
-
-template <typename Key, typename Value, typename Allocator>
-class NodeHashMap : public absl::node_hash_map<Key, Value,
-                                               absl::container_internal::hash_default_hash<Key>,
-                                               absl::container_internal::hash_default_eq<Key>,
-                                               Allocator> {
-  using Base = absl::node_hash_map<Key, Value,
-                                   absl::container_internal::hash_default_hash<Key>,
-                                   absl::container_internal::hash_default_eq<Key>,
-                                   Allocator>;
-
- public:
-  using Base::Base;
-};
-
-#else  // DISABLE_ABSEIL
-
-template <typename T, typename Allocator>
-class InlinedHashSet : public std::unordered_set<T,
-                                                 std::hash<T>,
-                                                 std::equal_to<T>,
-                                                 Allocator> {
-  using Base = std::unordered_set<T,
-                                  std::hash<T>,
-                                  std::equal_to<T>,
-                                  Allocator>;
-
- public:
-  using Base::Base;
-};
-
-template <typename Key, typename Value,
-          typename Allocator>
-class InlinedHashMap : public std::unordered_map<Key, Value,
-                                                 std::hash<Key>,
-                                                 std::equal_to<Key>,
-                                                 Allocator> {
-  using Base = std::unordered_map<Key, Value,
-                                  std::hash<Key>,
-                                  std::equal_to<Key>,
-                                  Allocator>;
-
- public:
-  using Base::Base;
-};
-
-// Use this hash set/map where pointer stability is required, otherwise use
-// InlinedHashSet and InlinedHashMap
-// This does not allocate a dummy 'end' node on default construction.
-// Use reserve() when the number of elements is known.
-template <typename T, typename Allocator>
-class NodeHashSet : public std::unordered_set<T,
-                                              std::hash<T>,
-                                              std::equal_to<T>,
-                                              Allocator> {
-  using Base = std::unordered_set<T,
-                                  std::hash<T>,
-                                  std::equal_to<T>,
-                                  Allocator>;
-
- public:
-  using Base::Base;
-};
-
-template <typename Key, typename Value, typename Allocator>
-class NodeHashMap : public std::unordered_map<Key, Value,
-                                       std::hash<Key>,
-                                       std::equal_to<Key>,
-                                       Allocator> {
-  using Base = std::unordered_map<Key, Value,
-                                  std::hash<Key>,
-                                  std::equal_to<Key>,
-                                  Allocator>;
-
- public:
-  using Base::Base;
-};
-
-#endif  // DISABLE_ABSEIL
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/inlined_containers_fwd.h b/projects/llm_framework/include/onnxruntime/core/common/inlined_containers_fwd.h
deleted file mode 100644
index 9e5215c2..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/inlined_containers_fwd.h
+++ /dev/null
@@ -1,147 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <memory>
-#include <utility>
-
-#ifndef DISABLE_ABSEIL
-#ifdef _MSC_VER
-#pragma warning(push)
-// C4127: conditional expression is constant
-#pragma warning(disable : 4127)
-// C4324: structure was padded due to alignment specifier
-// Usage of alignas causes some internal padding in places.
-#pragma warning(disable : 4324)
-#else
-// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102329#c2
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
-#endif
-#endif  // _MSC_VER
-
-#include <absl/container/inlined_vector.h>
-
-#ifdef _MSC_VER
-#pragma warning(pop)
-#else
-#if !defined(__clang__) && defined(__GNUC__)
-#pragma GCC diagnostic pop
-#endif
-#endif  // _MSC_VER
-
-#else
-
-#include <vector>
-
-#endif  // DISABLE_ABSEIL
-
-// Forward declarations for contexts where abseil can not be compiled and
-// not really needed but we want to have it in the headers that are included
-// e.g. CUDA 10 and .CU files
-// InlinedVector seems to be fine with old CUDA
-
-//===- llvm/ADT/SmallVector.h - 'Normally small' vectors --------*- C++ -*-===//
-//
-// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
-// See https://llvm.org/LICENSE.txt for license information.
-// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
-//
-// This file contains code and comments derived from llvm/ADT/SmallVector.h
-//
-// Specifically CalculateInlinedVectorDefaultInlinedElements<T>() template is derived from
-// CalculateSmallVectorDefaultInlinedElements<T>() and its comments.
-
-namespace onnxruntime {
-#ifndef DISABLE_ABSEIL
-/// Inspired by LLVM SmallVector with ONNX Runtime adjustments for abseil.
-///
-/// Helper class for calculating the default number of inline elements for
-/// `InlinedVector<T>`.
-/// This produces the following on MSVC x64
-///    int8_t  -> 41
-//     int16_t -> 21
-//     int32_t -> 11
-//     int64_t -> 6
-//     std::string 40 -> 1
-template <typename T>
-struct CalculateInlinedVectorDefaultInlinedElements {
-  // Parameter controlling the default number of inlined elements
-  // for `InlinedVector<T>`.
-  //
-  // The default number of inlined elements ensures that
-  // 1. There is at least one inlined element.
-  // 2. `sizeof(InlinedVector<T>) <= kPreferredInlinedVectorSizeof` unless
-  // it contradicts 1.
-  static constexpr size_t kPreferredInlinedVectorSizeof = 64;
-
-  // static_assert that sizeof(T) is not "too big".
-  //
-  // Because the InlinedVector must have at least one inlined element, it is possible
-  // for an arbitrarily large inlined element to allocate an arbitrarily large
-  // amount of inline storage. So we want to call attention to these cases and
-  // make sure that users are making an intentional decision if they request a lot of inline storage.
-  //
-  // We want this assertion to trigger in pathological cases, but otherwise
-  // not be too easy to hit. To accomplish that, the cutoff is actually somewhat
-  // larger than kPreferredInlinedVectorSizeof (otherwise,
-  // `InlinedVector<InlinedVector<T>>` would be one easy way to trip it, and that
-  // pattern seems useful in practice).
-  //
-  // One wrinkle is that this assertion is in theory non-portable, since
-  // sizeof(absl::InlinedVector<T, 1>) is in general platform-dependent. However, we don't expect this
-  // to be much of an issue, because most LLVM development happens on 64-bit
-  // hosts, and therefore sizeof(T) is expected to *decrease* when compiled for
-  // 32-bit hosts, dodging the issue. The reverse situation, where development
-  // happens on a 32-bit host and then fails due to sizeof(T) *increasing* on a
-  // 64-bit host, is expected to be very rare.
-  static_assert(
-      sizeof(absl::InlinedVector<T, 1>) <= kPreferredInlinedVectorSizeof,
-      "You are trying to use a default number of inlined elements for "
-      "`InlinedVector<T>` but `sizeof(T)` is really big! Please use an "
-      "explicit number of inlined elements with `InlinedVector<T, N>` to make "
-      "sure you really want that much inline storage.");
-
-  // Discount the size of the header itself when calculating the maximum inline
-  // bytes.
-  static constexpr size_t PreferredInlineBytes =
-      kPreferredInlinedVectorSizeof - (sizeof(absl::InlinedVector<T, 1>) - sizeof(T));
-  static constexpr size_t NumElementsThatFit = PreferredInlineBytes / sizeof(T);
-  static constexpr size_t value =
-      NumElementsThatFit == 0 ? 1 : NumElementsThatFit;
-};
-
-// Use InlinedVector for small arrays that can fit on a stack with a default
-// value pre-calculated.
-// Use TensorShapeVector for shapes.
-template <typename T,
-          size_t N = CalculateInlinedVectorDefaultInlinedElements<T>::value,
-          typename Allocator = std::allocator<T>>
-using InlinedVector = absl::InlinedVector<T, N, Allocator>;
-
-#else
-
-template <typename T,
-          size_t N = 0,
-          typename Allocator = std::allocator<T>>
-using InlinedVector = std::vector<T, Allocator>;
-
-#endif  // DISABLE_ABSEIL
-
-template <typename T,
-          typename Allocator = std::allocator<T>>
-class InlinedHashSet;
-
-template <typename Key, typename Value,
-          typename Allocator = std::allocator<std::pair<const Key, Value>>>
-class InlinedHashMap;
-
-template <typename T, typename Allocator = std::allocator<T>>
-class NodeHashSet;
-
-template <typename Key, typename Value,
-          typename Allocator = std::allocator<std::pair<const Key, Value>>>
-class NodeHashMap;
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/logging/capture.h b/projects/llm_framework/include/onnxruntime/core/common/logging/capture.h
deleted file mode 100644
index 811744d2..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/logging/capture.h
+++ /dev/null
@@ -1,115 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <cstdarg>
-#include "core/common/gsl.h"
-#include "core/common/common.h"
-#include "core/common/code_location.h"
-#include "core/common/logging/severity.h"
-
-namespace onnxruntime {
-namespace logging {
-
-class Logger;
-enum class DataType;
-
-/**
-   Class to capture the details of a log message.
-*/
-class Capture {
- public:
-  /**
-     Initializes a new instance of the Capture class.
-     @param logger The logger.
-     @param severity The severity.
-     @param category The category.
-     @param dataType Type of the data.
-     @param location The file location the log message is coming from.
-  */
-  Capture(const Logger& logger, logging::Severity severity, const char* category,
-          logging::DataType dataType, const CodeLocation& location)
-      : logger_{&logger}, severity_{severity}, category_{category}, data_type_{dataType}, location_{location} {
-  }
-
-  /**
-     The stream that can capture the message via operator<<.
-     @returns Output stream.
-  */
-  std::ostream& Stream() noexcept {
-    return stream_;
-  }
-
-#ifdef _MSC_VER
-// add SAL annotation for printf format string. requires Code Analysis to run to validate usage.
-#define msvc_printf_check _Printf_format_string_
-#define __attribute__(x)  // Disable for MSVC. Supported by GCC and CLang.
-#else
-#define msvc_printf_check
-#endif
-
-  /**
-     Captures a printf style log message.
-     @param name="format">The printf format.
-     @param name="">Arguments to the printf format if needed.
-     @remarks
-     A maximum of 2K of output will be captured currently.
-     Non-static method, so 'this' is implicit first arg, and we use format(printf(2,3)
-  */
-  void CapturePrintf(msvc_printf_check const char* format, ...) __attribute__((format(printf, 2, 3)));
-
-  /**
-     Process a printf style log message.
-     @param format The printf format.
-     @param ... Arguments to the printf format if needed.
-     @remarks
-     A maximum of 2K of output will be captured currently.
-     Note: As va_list is 'char *', we have to disambiguate this from CapturePrintf
-     so that something like "One string: %s", "the string" does not consider "the string"
-     to be the va_list.
-  */
-  void ProcessPrintf(msvc_printf_check const char* format, va_list args);
-
-  logging::Severity Severity() const noexcept {
-    return severity_;
-  }
-
-  char SeverityPrefix() const noexcept {
-    // Carefully setup so severity_ is a valid index
-    GSL_SUPPRESS(bounds .2) {
-      return logging::SEVERITY_PREFIX[static_cast<int>(severity_)];
-    }
-  }
-
-  const char* Category() const noexcept {
-    return category_;
-  }
-
-  logging::DataType DataType() const noexcept {
-    return data_type_;
-  }
-
-  const CodeLocation& Location() const noexcept {
-    return location_;
-  }
-
-  std::string Message() const noexcept {
-    return stream_.str();
-  }
-
-  ~Capture();
-
- private:
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(Capture);
-
-  const Logger* logger_;
-  const logging::Severity severity_;
-  const char* category_;
-  const logging::DataType data_type_;
-  const CodeLocation location_;
-
-  std::ostringstream stream_;
-};
-}  // namespace logging
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/logging/isink.h b/projects/llm_framework/include/onnxruntime/core/common/logging/isink.h
deleted file mode 100644
index a67777d4..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/logging/isink.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <string>
-
-#include "core/common/logging/logging.h"
-
-namespace onnxruntime {
-namespace logging {
-class ISink {
- public:
-  ISink() = default;
-
-  /**
-     Sends the message to the sink.
-     @param timestamp The timestamp.
-     @param logger_id The logger identifier.
-     @param message The captured message.
-  */
-  void Send(const Timestamp& timestamp, const std::string& logger_id, const Capture& message) {
-    SendImpl(timestamp, logger_id, message);
-  }
-
-  /**
-    Sends a Profiling Event Record to the sink.
-    @param Profiling Event Record
-  */
-  virtual void SendProfileEvent(profiling::EventRecord&) const {};
-
-  virtual ~ISink() = default;
-
- private:
-  // Make Code Analysis happy by disabling all for now. Enable as needed.
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(ISink);
-
-  virtual void SendImpl(const Timestamp& timestamp, const std::string& logger_id, const Capture& message) = 0;
-};
-}  // namespace logging
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/logging/logging.h b/projects/llm_framework/include/onnxruntime/core/common/logging/logging.h
deleted file mode 100644
index bea3fa1d..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/logging/logging.h
+++ /dev/null
@@ -1,337 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <atomic>
-#include <chrono>
-#include <climits>
-#include <map>
-#include <memory>
-#include <mutex>
-#include <string>
-
-#include "core/common/common.h"
-#include "core/common/profiler_common.h"
-#include "core/common/logging/capture.h"
-#include "core/common/logging/severity.h"
-
-#include "core/common/logging/macros.h"
-
-/*
-
-  Logging overview and expected usage:
-
-  At program startup:
-  * Create one or more ISink instances. If multiple, combine using composite_sink.
-  * Create a LoggingManager instance with the sink/s with is_default_instance set to true
-  * Only one instance should be created in this way, and it should remain valid for
-  until the program no longer needs to produce log output.
-
-  You can either use the static default Logger which LoggingManager will create when constructed
-  via LoggingManager::DefaultLogger(), or separate Logger instances each with different log ids
-  via LoggingManager::CreateLogger.
-
-  The log id is passed to the ISink instance with the sink determining how the log id is used
-  in the output.
-
-  LoggingManager
-  * creates the Logger instances used by the application
-  * provides a static default logger instance
-  * owns the log sink instance
-  * applies checks on severity and output of user data
-
-  The log macros create a Capture instance to capture the information to log.
-  If the severity and/or user filtering settings would prevent logging, no evaluation
-  of the log arguments will occur, so no performance cost beyond the severity and user
-  filtering check.
-
-  A sink can do further filter as needed.
-
-*/
-
-namespace onnxruntime {
-
-namespace logging {
-
-using Timestamp = std::chrono::time_point<std::chrono::system_clock>;
-
-#ifndef NDEBUG
-ORT_ATTRIBUTE_UNUSED static bool vlog_enabled = true;  // Set directly based on your needs.
-#else
-constexpr bool vlog_enabled = false;  // no VLOG output
-#endif
-
-enum class DataType {
-  SYSTEM = 0,  ///< System data.
-  USER = 1     ///< Contains potentially sensitive user data.
-};
-
-// Internal log categories.
-// Logging interface takes const char* so arbitrary values can also be used.
-struct Category {
-  static const char* onnxruntime;  ///< General output
-  static const char* System;       ///< Log output regarding interactions with the host system
-  // TODO: What other high level categories are meaningful? Model? Optimizer? Execution?
-};
-
-class ISink;
-class Logger;
-class Capture;
-
-/// <summary>
-/// The logging manager.
-/// Owns the log sink and potentially provides a default Logger instance.
-/// Provides filtering based on a minimum LogSeverity level, and of messages with DataType::User if enabled.
-/// </summary>
-class LoggingManager final {
- public:
-  enum InstanceType {
-    Default,  ///< Default instance of LoggingManager that should exist for the lifetime of the program
-    Temporal  ///< Temporal instance. CreateLogger(...) should be used, however DefaultLogger() will NOT be provided via this instance.
-  };
-
-  /**
-     Initializes a new instance of the LoggingManager class.
-     @param sink The sink to write to. Use CompositeSink if you need to write to multiple places.
-     @param default_min_severity The default minimum severity. Messages with lower severity will be ignored unless
-     overridden in CreateLogger.
-     @param default_filter_user_data If set to true ignore messages with DataType::USER unless overridden in CreateLogger.
-     @param instance_type If InstanceType::Default, this is the default instance of the LoggingManager
-     and is expected to exist for the lifetime of the program.
-     It creates and owns the default logger that calls to the static DefaultLogger method return.
-     @param default_logger_id Logger Id to use for the default logger. nullptr/ignored if instance_type == Temporal.
-     @param default_max_vlog_level Default maximum level for VLOG messages to be created unless overridden in CreateLogger.
-     Requires a severity of kVERBOSE for VLOG messages to be logged.
-  */
-  LoggingManager(std::unique_ptr<ISink> sink, Severity default_min_severity, bool default_filter_user_data,
-                 InstanceType instance_type,
-                 const std::string* default_logger_id = nullptr,
-                 int default_max_vlog_level = -1);
-
-  /**
-     Creates a new logger instance which will use the provided logger_id and default severity and vlog levels.
-     @param logger_id The log identifier.
-     @returns A new Logger instance that the caller owns.
-  */
-  std::unique_ptr<Logger> CreateLogger(const std::string& logger_id);
-
-  /**
-     Creates a new logger instance which will use the provided logger_id, severity and vlog levels.
-     @param logger_id The log identifier.
-     @param min_severity The minimum severity. Requests to create messages with lower severity will be ignored.
-     @param filter_user_data If set to true ignore messages with DataType::USER.
-     @param max_vlog_level Maximum level for VLOG messages to be created.
-     @returns A new Logger instance that the caller owns.
-  */
-  std::unique_ptr<Logger> CreateLogger(const std::string& logger_id,
-                                       Severity min_severity, bool filter_user_data, int max_vlog_level = -1);
-
-  /**
-     Gets the default logger instance if set. Throws if no default logger is currently registered.
-     @remarks
-     Creating a LoggingManager instance with is_default_instance == true registers a default logger.
-     Note that the default logger is only valid until the LoggerManager that registered it is destroyed.
-     @returns The default logger if available.
-  */
-  static const Logger& DefaultLogger();
-
-  /**
-    Return a boolean indicating if the default logger has been initialized
-  */
-  static bool HasDefaultLogger() { return nullptr != s_default_logger_; }
-
-  /**
-     Change the minimum severity level for log messages to be output by the default logger.
-     @param severity The severity.
-  */
-  static void SetDefaultLoggerSeverity(Severity severity);
-
-  /**
-     Change the maximum verbosity level for log messages to be output by the default logger.
-     @remarks
-     To activate the verbose log, the logger severity must also be set to kVERBOSE.
-     @param vlog_level The verbosity level.
-  */
-  static void SetDefaultLoggerVerbosity(int vlog_level);
-
-  /**
-     Logs a FATAL level message and creates an exception that can be thrown with error information.
-     @param category The log category.
-     @param location The location the log message was generated.
-     @param format_str The printf format string.
-     @param ... The printf arguments.
-     @returns A new Logger instance that the caller owns.
-  */
-  static std::exception LogFatalAndCreateException(const char* category,
-                                                   const CodeLocation& location,
-                                                   const char* format_str, ...);
-
-  /**
-     Logs the message using the provided logger id.
-     @param logger_id The log identifier.
-     @param message The log message.
-  */
-  void Log(const std::string& logger_id, const Capture& message) const;
-
-  /**
-    Sends a Profiling Event Record to the sink.
-    @param Profiling Event Record
-  */
-  void SendProfileEvent(profiling::EventRecord& eventRecord) const;
-  ~LoggingManager();
-
- private:
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(LoggingManager);
-
-  Timestamp GetTimestamp() const noexcept;
-  void CreateDefaultLogger(const std::string& logger_id);
-
-  std::unique_ptr<ISink> sink_;
-  const Severity default_min_severity_;
-  const bool default_filter_user_data_;
-  const int default_max_vlog_level_;
-  bool owns_default_logger_;
-
-  static Logger* s_default_logger_;
-
-  struct Epochs {
-    const std::chrono::time_point<std::chrono::high_resolution_clock> high_res;
-    const std::chrono::time_point<std::chrono::system_clock> system;
-    const std::chrono::minutes localtime_offset_from_utc;
-  };
-
-  static const Epochs& GetEpochs() noexcept;
-};
-
-/**
-   Logger provides a per-instance log id. Everything else is passed back up to the LoggingManager
-*/
-class Logger {
- public:
-  /**
-     Initializes a new instance of the Logger class.
-     @param loggingManager The logging manager.
-     @param id The identifier for messages coming from this Logger.
-     @param severity Minimum severity for messages to be created and logged.
-     @param filter_user_data Should USER data be filtered from output.
-     @param vlog_level Minimum level for VLOG messages to be created. Note that a severity of kVERBOSE must be provided
-     for VLOG messages to be logged.
-  */
-  Logger(const LoggingManager& loggingManager, std::string id,
-         Severity severity, bool filter_user_data, int vlog_level)
-      : logging_manager_{&loggingManager},
-        id_{id},
-        min_severity_{severity},
-        filter_user_data_{filter_user_data},
-        max_vlog_level_{vlog_level} {
-  }
-
-  /**
-     Get the minimum severity level for log messages to be output.
-     @returns The severity.
-  */
-  Severity GetSeverity() const noexcept { return min_severity_; }
-
-  /**
-     Change the minimum severity level for log messages to be output.
-     @param severity The severity.
-  */
-  void SetSeverity(Severity severity) noexcept { min_severity_ = severity; }
-
-  /**
-     Change the maximum verbosity level for log messages to be output.
-     @remarks
-     To activate the verbose log, the logger severity must also be set to kVERBOSE.
-     @param vlog_level The verbosity.
-  */
-  void SetVerbosity(int vlog_level) noexcept { max_vlog_level_ = vlog_level; }
-
-  /**
-     Check if output is enabled for the provided LogSeverity and DataType values.
-     @param severity The severity.
-     @param data_type Type of the data.
-     @returns True if a message with these values will be logged.
-  */
-  bool OutputIsEnabled(Severity severity, DataType data_type) const noexcept {
-    return (severity >= min_severity_ && (data_type != DataType::USER || !filter_user_data_));
-  }
-
-  /**
-     Return the maximum VLOG level allowed. Disabled unless logging VLOG messages
-  */
-  int VLOGMaxLevel() const noexcept {
-    return min_severity_ > Severity::kVERBOSE ? -1 : max_vlog_level_;
-  }
-
-  /**
-     Logs the captured message.
-     @param message The log message.
-  */
-  void Log(const Capture& message) const {
-    logging_manager_->Log(id_, message);
-  }
-
-  /**
-    Sends a Profiling Event Record to the sink.
-    @param Profiling Event Record
-  */
-  void SendProfileEvent(profiling::EventRecord& eventRecord) const {
-    logging_manager_->SendProfileEvent(eventRecord);
-  }
-
- private:
-  const LoggingManager* logging_manager_;
-  const std::string id_;
-  Severity min_severity_;
-  const bool filter_user_data_;
-  int max_vlog_level_;
-};
-
-inline const Logger& LoggingManager::DefaultLogger() {
-  if (s_default_logger_ == nullptr) {
-    // fail early for attempted misuse. don't use logging macros as we have no logger.
-    ORT_THROW("Attempt to use DefaultLogger but none has been registered.");
-  }
-
-  return *s_default_logger_;
-}
-
-inline void LoggingManager::SetDefaultLoggerSeverity(Severity severity) {
-  if (s_default_logger_ == nullptr) {
-    // fail early for attempted misuse. don't use logging macros as we have no logger.
-    ORT_THROW("Attempt to use DefaultLogger but none has been registered.");
-  }
-
-  s_default_logger_->SetSeverity(severity);
-}
-
-inline void LoggingManager::SetDefaultLoggerVerbosity(int vlog_level) {
-  if (s_default_logger_ == nullptr) {
-    // fail early for attempted misuse. don't use logging macros as we have no logger.
-    ORT_THROW("Attempt to use DefaultLogger but none has been registered.");
-  }
-
-  s_default_logger_->SetVerbosity(vlog_level);
-}
-
-inline Timestamp LoggingManager::GetTimestamp() const noexcept {
-  static const Epochs& epochs = GetEpochs();
-
-  const auto high_res_now = std::chrono::high_resolution_clock::now();
-  return std::chrono::time_point_cast<std::chrono::system_clock::duration>(
-      epochs.system + (high_res_now - epochs.high_res) + epochs.localtime_offset_from_utc);
-}
-
-/**
-   Return the current thread id.
-*/
-unsigned int GetThreadId();
-
-/**
-   Return the current process id.
-*/
-unsigned int GetProcessId();
-
-}  // namespace logging
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/logging/macros.h b/projects/llm_framework/include/onnxruntime/core/common/logging/macros.h
deleted file mode 100644
index 5579f6bd..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/logging/macros.h
+++ /dev/null
@@ -1,278 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-// NOTE: Don't include this file directly. Include logging.h
-
-#define CREATE_MESSAGE(logger, severity, category, datatype) \
-  ::onnxruntime::logging::Capture(logger, ::onnxruntime::logging::Severity::k##severity, category, datatype, ORT_WHERE)
-
-/*
-  Both printf and stream style logging are supported.
-  Not that printf currently has a 2K limit to the message size.
-
-  LOGS_* macros are for stream style
-  LOGF_* macros are for printf style
-
-  The Message class captures the log input, and pushes it through the logger in its destructor.
-
-  Use the *FATAL* macros if you want a Severity::kFatal message to also throw.
-
-  There are a few variants to minimize the length of the macro name required in the calling code.
-  They are optimized so the shortest names are for the (expected) most common usage. This can be
-  tweaked if needed.
-
-  Explicit logger vs LoggingManager::DefaulLogger()
-  Default is for a logger instance to be explicitly passed in.
-  The logger instance provides an identifier so that log messages from different runs can be separated.
-
-  Variants with DEFAULT in the macro name use the default logger provided by logging manager. This is
-  static so accessible from any code, provided a LoggingManager instance created with InstanceType::Default
-  exists somewhere. See logging.h for further explanation of the expected setup.
-
-  DataType
-  Default uses DataType::SYSTEM.
-
-  Variants with USER in the macro name use DataType::USER. This is data that could be PII, and may need to
-  be filtered from output. LoggingManager applies this filtering.
-
-  Category
-  Default category is ::onnxruntime::Logging::Category::onnxruntime.
-
-  If you wish to provide a different category, use variants with CATEGORY in the macro name
-
-*/
-
-/**
- * Note:
- * The stream style logging macros (something like `LOGS() << message`) are designed to be appended to.
- * Normally, we can isolate macro code in a separate scope (e.g., `do {...} while(0)`), but here we need the macro code
- * to interact with subsequent code (i.e., the values to log).
- *
- * When an unisolated conditional is involved, extra care needs to be taken to avoid unexpected parsing behavior.
- * For example:
- *
- * if (enabled)
- *   Capture().Stream()
- *
- * is more direct, but
- *
- * if (!enabled) {
- * } else Capture().Stream()
- *
- * ensures that the `if` does not unintentionally associate with a subsequent `else`.
- */
-
-// Logging with explicit category
-
-// iostream style logging. Capture log info in Message, and push to the logger in ~Message.
-#define LOGS_CATEGORY(logger, severity, category)                              \
-  if (!(logger).OutputIsEnabled(::onnxruntime::logging::Severity::k##severity, \
-                                ::onnxruntime::logging::DataType::SYSTEM)) {   \
-    /* do nothing */                                                           \
-  } else                                                                       \
-    CREATE_MESSAGE(logger, severity, category, ::onnxruntime::logging::DataType::SYSTEM).Stream()
-
-#define LOGS_USER_CATEGORY(logger, severity, category)                         \
-  if (!(logger).OutputIsEnabled(::onnxruntime::logging::Severity::k##severity, \
-                                ::onnxruntime::logging::DataType::USER)) {     \
-    /* do nothing */                                                           \
-  } else                                                                       \
-    CREATE_MESSAGE(logger, severity, category, ::onnxruntime::logging::DataType::USER).Stream()
-
-// printf style logging. Capture log info in Message, and push to the logger in ~Message.
-#define LOGF_CATEGORY(logger, severity, category, format_str, ...)                         \
-  do {                                                                                     \
-    if ((logger).OutputIsEnabled(::onnxruntime::logging::Severity::k##severity,            \
-                                 ::onnxruntime::logging::DataType::SYSTEM))                \
-      CREATE_MESSAGE(logger, severity, category, ::onnxruntime::logging::DataType::SYSTEM) \
-          .CapturePrintf(format_str, ##__VA_ARGS__);                                       \
-  } while (0)
-
-#define LOGF_USER_CATEGORY(logger, severity, category, format_str, ...)                  \
-  do {                                                                                   \
-    if ((logger).OutputIsEnabled(::onnxruntime::logging::Severity::k##severity,          \
-                                 ::onnxruntime::logging::DataType::USER))                \
-      CREATE_MESSAGE(logger, severity, category, ::onnxruntime::logging::DataType::USER) \
-          .CapturePrintf(format_str, ##__VA_ARGS__);                                     \
-  } while (0)
-
-// Logging with category of "onnxruntime"
-
-#define LOGS(logger, severity) \
-  LOGS_CATEGORY(logger, severity, ::onnxruntime::logging::Category::onnxruntime)
-
-#define LOGS_USER(logger, severity) \
-  LOGS_USER_CATEGORY(logger, severity, ::onnxruntime::logging::Category::onnxruntime)
-
-// printf style logging. Capture log info in Message, and push to the logger in ~Message.
-#define LOGF(logger, severity, format_str, ...) \
-  LOGF_CATEGORY(logger, severity, ::onnxruntime::logging::Category::onnxruntime, format_str, ##__VA_ARGS__)
-
-#define LOGF_USER(logger, severity, format_str, ...) \
-  LOGF_USER_CATEGORY(logger, severity, ::onnxruntime::logging::Category::onnxruntime, format_str, ##__VA_ARGS__)
-
-/*
-  Macros that use the default logger.
-  A LoggingManager instance must be currently valid for the default logger to be available.
-*/
-
-// Logging with explicit category
-
-#define LOGS_DEFAULT_CATEGORY(severity, category) \
-  LOGS_CATEGORY(::onnxruntime::logging::LoggingManager::DefaultLogger(), severity, category)
-
-#define LOGS_USER_DEFAULT_CATEGORY(severity, category) \
-  LOGS_USER_CATEGORY(::onnxruntime::logging::LoggingManager::DefaultLogger(), severity, category)
-
-#define LOGF_DEFAULT_CATEGORY(severity, category, format_str, ...) \
-  LOGF_CATEGORY(::onnxruntime::logging::LoggingManager::DefaultLogger(), severity, category, format_str, ##__VA_ARGS__)
-
-#define LOGF_USER_DEFAULT_CATEGORY(severity, category, format_str, ...) \
-  LOGF_USER_CATEGORY(::onnxruntime::logging::LoggingManager::DefaultLogger(), severity, category, format_str, ##__VA_ARGS__)
-
-// Logging with category of "onnxruntime"
-
-#define LOGS_DEFAULT(severity) \
-  LOGS_DEFAULT_CATEGORY(severity, ::onnxruntime::logging::Category::onnxruntime)
-
-#define LOGS_USER_DEFAULT(severity) \
-  LOGS_USER_DEFAULT_CATEGORY(severity, ::onnxruntime::logging::Category::onnxruntime)
-
-#define LOGF_DEFAULT(severity, format_str, ...) \
-  LOGF_DEFAULT_CATEGORY(severity, ::onnxruntime::logging::Category::onnxruntime, format_str, ##__VA_ARGS__)
-
-#define LOGF_USER_DEFAULT(severity, format_str, ...) \
-  LOGF_USER_DEFAULT_CATEGORY(severity, ::onnxruntime::logging::Category::onnxruntime, format_str, ##__VA_ARGS__)
-
-/*
-  Conditional logging
-*/
-
-// Logging with explicit category
-
-#define LOGS_CATEGORY_IF(boolean_expression, logger, severity, category) \
-  if (!((boolean_expression) == true)) {                                 \
-    /* do nothing */                                                     \
-  } else                                                                 \
-    LOGS_CATEGORY(logger, severity, category)
-
-#define LOGS_DEFAULT_CATEGORY_IF(boolean_expression, severity, category) \
-  if (!((boolean_expression) == true)) {                                 \
-    /* do nothing */                                                     \
-  } else                                                                 \
-    LOGS_DEFAULT_CATEGORY(severity, category)
-
-#define LOGS_USER_CATEGORY_IF(boolean_expression, logger, severity, category) \
-  if (!((boolean_expression) == true)) {                                      \
-    /* do nothing */                                                          \
-  } else                                                                      \
-    LOGS_USER_CATEGORY(logger, severity, category)
-
-#define LOGS_USER_DEFAULT_CATEGORY_IF(boolean_expression, severity, category) \
-  if (!((boolean_expression) == true)) {                                      \
-    /* do nothing */                                                          \
-  } else                                                                      \
-    LOGS_USER_DEFAULT_CATEGORY(severity, category)
-
-#define LOGF_CATEGORY_IF(boolean_expression, logger, severity, category, format_str, ...)                   \
-  do {                                                                                                      \
-    if ((boolean_expression) == true) LOGF_CATEGORY(logger, severity, category, format_str, ##__VA_ARGS__); \
-  } while (0)
-
-#define LOGF_DEFAULT_CATEGORY_IF(boolean_expression, severity, category, format_str, ...)                   \
-  do {                                                                                                      \
-    if ((boolean_expression) == true) LOGF_DEFAULT_CATEGORY(severity, category, format_str, ##__VA_ARGS__); \
-  } while (0)
-
-#define LOGF_USER_CATEGORY_IF(boolean_expression, logger, severity, category, format_str, ...)                   \
-  do {                                                                                                           \
-    if ((boolean_expression) == true) LOGF_USER_CATEGORY(logger, severity, category, format_str, ##__VA_ARGS__); \
-  } while (0)
-
-#define LOGF_USER_DEFAULT_CATEGORY_IF(boolean_expression, severity, category, format_str, ...)                   \
-  do {                                                                                                           \
-    if ((boolean_expression) == true) LOGF_USER_DEFAULT_CATEGORY(severity, category, format_str, ##__VA_ARGS__); \
-  } while (0)
-
-// Logging with category of "onnxruntime"
-
-#define LOGS_IF(boolean_expression, logger, severity) \
-  LOGS_CATEGORY_IF(boolean_expression, logger, severity, ::onnxruntime::logging::Category::onnxruntime)
-
-#define LOGS_DEFAULT_IF(boolean_expression, severity) \
-  LOGS_DEFAULT_CATEGORY_IF(boolean_expression, severity, ::onnxruntime::logging::Category::onnxruntime)
-
-#define LOGS_USER_IF(boolean_expression, logger, severity) \
-  LOGS_USER_CATEGORY_IF(boolean_expression, logger, severity, ::onnxruntime::logging::Category::onnxruntime)
-
-#define LOGS_USER_DEFAULT_IF(boolean_expression, severity) \
-  LOGS_USER_DEFAULT_CATEGORY_IF(boolean_expression, severity, ::onnxruntime::logging::Category::onnxruntime)
-
-#define LOGF_IF(boolean_expression, logger, severity, format_str, ...) \
-  LOGF_CATEGORY_IF(boolean_expression, logger, severity, ::onnxruntime::logging::Category::onnxruntime, format_str, ##__VA_ARGS__)
-
-#define LOGF_DEFAULT_IF(boolean_expression, severity, format_str, ...) \
-  LOGF_DEFAULT_CATEGORY_IF(boolean_expression, severity, ::onnxruntime::logging::Category::onnxruntime, format_str, ##__VA_ARGS__)
-
-#define LOGF_USER_IF(boolean_expression, logger, severity, format_str, ...)                                  \
-  LOGF_USER_CATEGORY_IF(boolean_expression, logger, severity, ::onnxruntime::logging::Category::onnxruntime, \
-                        format_str, ##__VA_ARGS__)
-
-#define LOGF_USER_DEFAULT_IF(boolean_expression, severity, format_str, ...)                                  \
-  LOGF_USER_DEFAULT_CATEGORY_IF(boolean_expression, severity, ::onnxruntime::logging::Category::onnxruntime, \
-                                format_str, ##__VA_ARGS__)
-
-/*
-  Debug verbose logging of caller provided level.
-  Disabled in Release builds.
-  Use the _USER variants for VLOG statements involving user data that may need to be filtered.
-*/
-#ifndef NDEBUG
-#define VLOGS(logger, level)                                                         \
-  if (!(::onnxruntime::logging::vlog_enabled && level <= (logger).VLOGMaxLevel())) { \
-    /* do nothing */                                                                 \
-  } else                                                                             \
-    LOGS_CATEGORY(logger, VERBOSE, "VLOG" #level)
-
-#define VLOGS_USER(logger, level)                                                    \
-  if (!(::onnxruntime::logging::vlog_enabled && level <= (logger).VLOGMaxLevel())) { \
-    /* do nothing */                                                                 \
-  } else                                                                             \
-    LOGS_USER_CATEGORY(logger, VERBOSE, "VLOG" #level)
-
-#define VLOGF(logger, level, format_str, ...)                                     \
-  do {                                                                            \
-    if (::onnxruntime::logging::vlog_enabled && level <= (logger).VLOGMaxLevel()) \
-      LOGF_CATEGORY(logger, VERBOSE, "VLOG" #level, format_str, ##__VA_ARGS__);   \
-  } while (0)
-
-#define VLOGF_USER(logger, level, format_str, ...)                                   \
-  do {                                                                               \
-    if (::onnxruntime::logging::vlog_enabled && level <= (logger).VLOGMaxLevel())    \
-      LOGF_USER_CATEGORY(logger, VERBOSE, "VLOG" #level, format_str, ##__VA_ARGS__); \
-  } while (0)
-#else
-// Disabled in Release builds.
-#define VLOGS(logger, level) \
-  if constexpr (true) {} else LOGS_CATEGORY(logger, VERBOSE, "VLOG" #level)
-#define VLOGS_USER(logger, level) \
-  if constexpr (true) {} else LOGS_USER_CATEGORY(logger, VERBOSE, "VLOG" #level)
-#define VLOGF(logger, level, format_str, ...)
-#define VLOGF_USER(logger, level, format_str, ...)
-#endif
-
-
-
-// Default logger variants
-#define VLOGS_DEFAULT(level) \
-  VLOGS(::onnxruntime::logging::LoggingManager::DefaultLogger(), level)
-
-#define VLOGS_USER_DEFAULT(level) \
-  VLOGS_USER(::onnxruntime::logging::LoggingManager::DefaultLogger(), level)
-
-#define VLOGF_DEFAULT(level, format_str, ...) \
-  VLOGF(::onnxruntime::logging::LoggingManager::DefaultLogger(), level, format_str, ##__VA_ARGS__)
-
-#define VLOGF_USER_DEFAULT(level, format_str, ...) \
-  VLOGF_USER(::onnxruntime::logging::LoggingManager::DefaultLogger(), level, format_str, ##__VA_ARGS__)
diff --git a/projects/llm_framework/include/onnxruntime/core/common/logging/severity.h b/projects/llm_framework/include/onnxruntime/core/common/logging/severity.h
deleted file mode 100644
index e43f192e..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/logging/severity.h
+++ /dev/null
@@ -1,22 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-namespace onnxruntime {
-namespace logging {
-// mild violation of naming convention. the 'k' lets us use token concatenation in the macro
-// ::onnxruntime::Logging::Severity::k##severity. It's not legal to have ::onnxruntime::Logging::Severity::##severity
-// the uppercase makes the LOG macro usage look as expected for passing an enum value as it will be LOGS(logger, ERROR)
-enum class Severity {
-  kVERBOSE = 0,
-  kINFO = 1,
-  kWARNING = 2,
-  kERROR = 3,
-  kFATAL = 4
-};
-
-constexpr const char* SEVERITY_PREFIX = "VIWEF";
-
-}  // namespace logging
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/make_string.h b/projects/llm_framework/include/onnxruntime/core/common/make_string.h
deleted file mode 100644
index 826898de..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/make_string.h
+++ /dev/null
@@ -1,126 +0,0 @@
-/**
- * Copyright (c) 2016-present, Facebook, Inc.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-// Portions Copyright (c) Microsoft Corporation
-
-#pragma once
-
-#include <locale>
-#include <sstream>
-#include <type_traits>
-
-namespace onnxruntime {
-
-namespace detail {
-
-inline void MakeStringImpl(std::ostringstream& /*ss*/) noexcept {
-}
-
-template <typename T>
-inline void MakeStringImpl(std::ostringstream& ss, const T& t) noexcept {
-  ss << t;
-}
-
-template <typename T, typename... Args>
-inline void MakeStringImpl(std::ostringstream& ss, const T& t, const Args&... args) noexcept {
-  MakeStringImpl(ss, t);
-  MakeStringImpl(ss, args...);
-}
-
-// see MakeString comments for explanation of why this is necessary
-template <typename... Args>
-inline std::string MakeStringImpl(const Args&... args) noexcept {
-  std::ostringstream ss;
-  MakeStringImpl(ss, args...);
-  return ss.str();
-}
-
-//
-// Infrastructure to convert char[n] to char* to reduce binary size
-//
-
-// default is to leave the type as is
-template <class T>
-struct if_char_array_make_ptr {
-  using type = T;
-};
-
-// specialization that matches an array reference, which is what the char array from a string literal
-// used in a call to MakeString will be.
-// if the type is a char[n] array we 'decay' it to a char* so that the usages can be folded.
-template <class T, size_t N>
-struct if_char_array_make_ptr<T (&)[N]> {
-  // remove a single extent (T[x] -> T, but T[x][y] -> T[y]) so we only match char[x],
-  // and get the type name without the 'const' so both 'const char (&)[n]' and 'char (&)[n]' are matched.
-  using element_type = typename std::remove_const<typename std::remove_extent<T>::type>::type;
-  using type = typename std::conditional<std::is_same<char, element_type>::value, T*, T (&)[N]>::type;
-};
-
-// helper to make usage simpler in MakeString
-template <class T>
-using if_char_array_make_ptr_t = typename if_char_array_make_ptr<T>::type;
-}  // namespace detail
-
-/**
- * Makes a string by concatenating string representations of the arguments.
- * This version uses the current locale.
- */
-template <typename... Args>
-std::string MakeString(const Args&... args) {
-  // We need to update the types from the MakeString template instantiation to decay any char[n] to char*.
-  //   e.g. MakeString("in", "out") goes from MakeString<char[2], char[3]> to MakeStringImpl<char*, char*>
-  //        so that MakeString("out", "in") will also match MakeStringImpl<char*, char*> instead of requiring
-  //        MakeStringImpl<char[3], char[2]>.
-  //
-  // We have to do the type processing before any actual work, so this function purely implements the type processing.
-  // If we do not do it this way we do not get the full binary size reduction.
-  //
-  // See https://stackoverflow.com/a/29418212/684911 for overall details of the approach, but note it does not cover
-  // the need to do the type processing as a separate step.
-
-  return detail::MakeStringImpl(detail::if_char_array_make_ptr_t<Args const&>(args)...);
-}
-
-/**
- * Makes a string by concatenating string representations of the arguments.
- * This version uses std::locale::classic().
- */
-template <typename... Args>
-std::string MakeStringWithClassicLocale(const Args&... args) {
-  std::ostringstream ss;
-  ss.imbue(std::locale::classic());
-  detail::MakeStringImpl(ss, args...);
-  return ss.str();
-}
-
-// MakeString versions for already-a-string types.
-
-inline std::string MakeString(const std::string& str) {
-  return str;
-}
-
-inline std::string MakeString(const char* cstr) {
-  return cstr;
-}
-
-inline std::string MakeStringWithClassicLocale(const std::string& str) {
-  return str;
-}
-
-inline std::string MakeStringWithClassicLocale(const char* cstr) {
-  return cstr;
-}
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/narrow.h b/projects/llm_framework/include/onnxruntime/core/common/narrow.h
deleted file mode 100644
index 15bcf167..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/narrow.h
+++ /dev/null
@@ -1,77 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-// onnxruntime::narrow() is like gsl::narrow() but it is also available when exceptions are disabled.
-
-#if !defined(ORT_NO_EXCEPTIONS)
-
-#include "gsl/narrow"
-
-namespace onnxruntime {
-using gsl::narrow;
-}  // namespace onnxruntime
-
-#else  // ^^ !defined(ORT_NO_EXCEPTIONS) ^^ / vv defined(ORT_NO_EXCEPTIONS) vv
-
-#include <cstdio>     // std::fprintf
-#include <exception>  // std::terminate
-#include <type_traits>
-
-#include "gsl/util"  // gsl::narrow_cast
-
-namespace onnxruntime {
-
-namespace detail {
-[[noreturn]] inline void OnNarrowingError() noexcept {
-  std::fprintf(stderr, "%s", "narrowing error\n");
-  std::terminate();
-}
-}  // namespace detail
-
-// This implementation of onnxruntime::narrow was copied and adapted from:
-// https://github.com/microsoft/GSL/blob/a3534567187d2edc428efd3f13466ff75fe5805c/include/gsl/narrow
-
-// narrow() : a checked version of narrow_cast() that terminates if the cast changed the value
-template <class T, class U, typename std::enable_if<std::is_arithmetic<T>::value>::type* = nullptr>
-// clang-format off
-GSL_SUPPRESS(type.1) // NO-FORMAT: attribute
-                      // clang-format on
-    constexpr T narrow(U u) noexcept {
-  constexpr const bool is_different_signedness =
-      (std::is_signed<T>::value != std::is_signed<U>::value);
-
-  // clang-format off
-GSL_SUPPRESS(es.103) // NO-FORMAT: attribute // don't overflow
-GSL_SUPPRESS(es.104) // NO-FORMAT: attribute // don't underflow
-GSL_SUPPRESS(p.2) // NO-FORMAT: attribute // don't rely on undefined behavior
-  // clang-format on
-  const T t = gsl::narrow_cast<T>(u);  // While this is technically undefined behavior in some cases (i.e., if the source value is of floating-point type
-                                       // and cannot fit into the destination integral type), the resultant behavior is benign on the platforms
-                                       // that we target (i.e., no hardware trap representations are hit).
-
-  if (static_cast<U>(t) != u || (is_different_signedness && ((t < T{}) != (u < U{})))) {
-    detail::OnNarrowingError();
-  }
-
-  return t;
-}
-
-template <class T, class U, typename std::enable_if<!std::is_arithmetic<T>::value>::type* = nullptr>
-// clang-format off
-GSL_SUPPRESS(type.1) // NO-FORMAT: attribute
-                      // clang-format on
-    constexpr T narrow(U u) noexcept {
-  const T t = gsl::narrow_cast<T>(u);
-
-  if (static_cast<U>(t) != u) {
-    detail::OnNarrowingError();
-  }
-
-  return t;
-}
-
-}  // namespace onnxruntime
-
-#endif  // defined(ORT_NO_EXCEPTIONS)
diff --git a/projects/llm_framework/include/onnxruntime/core/common/optional.h b/projects/llm_framework/include/onnxruntime/core/common/optional.h
deleted file mode 100644
index f7106a3b..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/optional.h
+++ /dev/null
@@ -1,23 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-#include <optional>
-
-namespace onnxruntime {
-
-using std::optional;
-
-#ifndef ORT_NO_EXCEPTIONS
-using std::bad_optional_access;
-#endif
-
-using std::nullopt;
-using std::nullopt_t;
-
-using std::in_place;
-using std::in_place_t;
-
-using std::make_optional;
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/parse_string.h b/projects/llm_framework/include/onnxruntime/core/common/parse_string.h
deleted file mode 100644
index 941e3f33..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/parse_string.h
+++ /dev/null
@@ -1,85 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <locale>
-#include <sstream>
-#include <string_view>
-#include <type_traits>
-
-#include "core/common/common.h"
-
-namespace onnxruntime {
-
-/**
- * Tries to parse a value from an entire string.
- */
-template <typename T>
-bool TryParseStringWithClassicLocale(std::string_view str, T& value) {
-  if constexpr (std::is_integral<T>::value && std::is_unsigned<T>::value) {
-    // if T is unsigned integral type, reject negative values which will wrap
-    if (!str.empty() && str[0] == '-') {
-      return false;
-    }
-  }
-
-  // don't allow leading whitespace
-  if (!str.empty() && std::isspace(str[0], std::locale::classic())) {
-    return false;
-  }
-
-  std::istringstream is{std::string{str}};
-  is.imbue(std::locale::classic());
-  T parsed_value{};
-
-  const bool parse_successful =
-      is >> parsed_value &&
-      is.get() == std::istringstream::traits_type::eof();  // don't allow trailing characters
-  if (!parse_successful) {
-    return false;
-  }
-
-  value = std::move(parsed_value);
-  return true;
-}
-
-inline bool TryParseStringWithClassicLocale(std::string_view str, std::string& value) {
-  value = str;
-  return true;
-}
-
-inline bool TryParseStringWithClassicLocale(std::string_view str, bool& value) {
-  if (str == "0" || str == "False" || str == "false") {
-    value = false;
-    return true;
-  }
-
-  if (str == "1" || str == "True" || str == "true") {
-    value = true;
-    return true;
-  }
-
-  return false;
-}
-
-/**
- * Parses a value from an entire string.
- */
-template <typename T>
-Status ParseStringWithClassicLocale(std::string_view s, T& value) {
-  ORT_RETURN_IF_NOT(TryParseStringWithClassicLocale(s, value), "Failed to parse value: \"", value, "\"");
-  return Status::OK();
-}
-
-/**
- * Parses a value from an entire string.
- */
-template <typename T>
-T ParseStringWithClassicLocale(std::string_view s) {
-  T value{};
-  ORT_THROW_IF_ERROR(ParseStringWithClassicLocale(s, value));
-  return value;
-}
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/profiler_common.h b/projects/llm_framework/include/onnxruntime/core/common/profiler_common.h
deleted file mode 100644
index c0898883..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/profiler_common.h
+++ /dev/null
@@ -1,93 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include "core/common/common.h"
-
-#include <string>
-#include <unordered_map>
-
-namespace onnxruntime {
-namespace profiling {
-
-enum EventCategory {
-  SESSION_EVENT = 0,
-  NODE_EVENT,
-  KERNEL_EVENT,
-  API_EVENT,
-  EVENT_CATEGORY_MAX
-};
-
-// Event descriptions for the above session events.
-static constexpr const char* event_category_names_[EVENT_CATEGORY_MAX] = {
-    "Session",
-    "Node",
-    "Kernel",
-    "Api"};
-
-// Timing record for all events.
-struct EventRecord {
-  EventRecord() = default;
-  EventRecord(EventCategory category,
-              int process_id,
-              int thread_id,
-              std::string&& event_name,
-              long long time_stamp,
-              long long duration,
-              std::unordered_map<std::string, std::string>&& event_args)
-      : cat(category),
-        pid(process_id),
-        tid(thread_id),
-        name(std::move(event_name)),
-        ts(time_stamp),
-        dur(duration),
-        args(std::move(event_args)) {}
-
-  EventRecord(EventCategory category,
-              int process_id,
-              int thread_id,
-              const std::string& event_name,
-              long long time_stamp,
-              long long duration,
-              const std::unordered_map<std::string, std::string>& event_args)
-      : cat(category),
-        pid(process_id),
-        tid(thread_id),
-        name(event_name),
-        ts(time_stamp),
-        dur(duration),
-        args(event_args) {}
-
-  EventRecord(const EventRecord& other) = default;
-  EventRecord(EventRecord&& other) noexcept = default;
-  EventRecord& operator=(const EventRecord& other) = default;
-  EventRecord& operator=(EventRecord&& other) = default;
-
-  EventCategory cat = EventCategory::API_EVENT;
-  int pid = -1;
-  int tid = -1;
-  std::string name{};
-  long long ts = 0;
-  long long dur = 0;
-  std::unordered_map<std::string, std::string> args{};
-};
-
-using Events = std::vector<EventRecord>;
-
-//Execution Provider Profiler
-class EpProfiler {
- public:
-  virtual ~EpProfiler() = default;
-  virtual bool StartProfiling(TimePoint profiling_start_time) = 0;      // called when profiling starts
-  virtual void EndProfiling(TimePoint start_time, Events& events) = 0;  // called when profiling ends, save all captures numbers to "events"
-  virtual void Start(uint64_t){};                                       // called before op start, accept an id as argument to identify the op
-  virtual void Stop(uint64_t){};                                        // called after op stop, accept an id as argument to identify the op
-};
-
-// Demangle C++ symbols
-std::string demangle(const char* name);
-std::string demangle(const std::string& name);
-
-}  // namespace profiling
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/span_utils.h b/projects/llm_framework/include/onnxruntime/core/common/span_utils.h
deleted file mode 100644
index 82aa59f0..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/span_utils.h
+++ /dev/null
@@ -1,88 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <algorithm>
-
-#include "core/common/gsl.h"
-
-namespace onnxruntime {
-
-// AsSpan inspired by Fekir's Blog https://fekir.info/post/span-the-missing-constructor/
-// Used under MIT license
-
-// Use AsSpan for less typing on any container including initializer list to create a span
-// (unnamed, untyped initializer list does not automatically convert to gsl::span).
-// {1, 2, 3} as such does not have a type
-// (see https://scottmeyers.blogspot.com/2014/03/if-braced-initializers-have-no-type-why.html)
-//
-//   Example: AsSpan({1, 2, 3}) results in gsl::span<const int>
-//
-// The above would deduce to std::initializer_list<int> and the result is gsl::span<const int>
-//
-// AsSpan<int64_t>({1, 2, 3}) produces gsl::span<const int64_t>
-//
-// We can also do std::array<int64_t, 3>{1, 2, 3} that can be automatically converted to span
-// without memory allocation.
-//
-// If type conversion is not required, then for C++17 std::array template parameters are
-// auto-deduced. Example: std::array{1, 2, 3}.
-// We are aiming at not allocating memory dynamically.
-
-namespace details {
-template <class P>
-constexpr auto AsSpanImpl(P* p, size_t s) {
-  return gsl::span<P>(p, s);
-}
-}  // namespace details
-
-template <class C>
-constexpr auto AsSpan(C& c) {
-  return details::AsSpanImpl(c.data(), c.size());
-}
-
-template <class C>
-constexpr auto AsSpan(const C& c) {
-  return details::AsSpanImpl(c.data(), c.size());
-}
-
-template <class C>
-constexpr auto AsSpan(C&& c) {
-  return details::AsSpanImpl(c.data(), c.size());
-}
-
-template <class T>
-constexpr auto AsSpan(std::initializer_list<T> c) {
-  return details::AsSpanImpl(c.begin(), c.size());
-}
-
-template <class T, size_t N>
-constexpr auto AsSpan(T (&arr)[N]) {
-  return details::AsSpanImpl(arr, N);
-}
-
-template <class T, size_t N>
-constexpr auto AsSpan(const T (&arr)[N]) {
-  return details::AsSpanImpl(arr, N);
-}
-
-template <class T>
-inline gsl::span<const T> EmptySpan() { return gsl::span<const T>(); }
-
-template <class U, class T>
-[[nodiscard]] inline gsl::span<U> ReinterpretAsSpan(gsl::span<T> src) {
-  // adapted from gsl-lite span::as_span():
-  // https://github.com/gsl-lite/gsl-lite/blob/4720a2980a30da085b4ddb4a0ea2a71af7351a48/include/gsl/gsl-lite.hpp#L4102-L4108
-  Expects(src.size_bytes() % sizeof(U) == 0);
-  return gsl::span<U>(reinterpret_cast<U*>(src.data()), src.size_bytes() / sizeof(U));
-}
-
-template <class T1, size_t Extent1, class T2, size_t Extent2>
-[[nodiscard]] inline bool SpanEq(gsl::span<T1, Extent1> a, gsl::span<T2, Extent2> b) {
-  static_assert(std::is_same_v<std::remove_const_t<T1>, std::remove_const_t<T2>>,
-                "T1 and T2 should be the same type except for const qualification");
-  return std::equal(a.begin(), a.end(), b.begin(), b.end());
-}
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/spin_pause.h b/projects/llm_framework/include/onnxruntime/core/common/spin_pause.h
deleted file mode 100644
index 49b71e55..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/spin_pause.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#if defined(_M_AMD64)
-#include <intrin.h>
-#endif
-
-#if defined(__x86_64__)
-#include <xmmintrin.h>
-#endif
-
-namespace onnxruntime {
-
-namespace concurrency {
-
-// Intrinsic to use in spin-loops
-
-inline void SpinPause() {
-#if defined(_M_AMD64) || defined(__x86_64__)
-  _mm_pause();
-#endif
-}
-
-}  // namespace concurrency
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/status.h b/projects/llm_framework/include/onnxruntime/core/common/status.h
deleted file mode 100644
index d6e19929..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/status.h
+++ /dev/null
@@ -1,195 +0,0 @@
-/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-    http://www.apache.org/licenses/LICENSE-2.0
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-// Modifications Copyright (c) Microsoft.
-
-#pragma once
-
-#include <memory>
-#include <ostream>
-#include <string>
-#ifdef _WIN32
-#include <winerror.h>
-#endif
-#include "core/common/gsl.h"
-namespace onnxruntime {
-namespace common {
-
-enum StatusCategory {
-  NONE = 0,
-  SYSTEM = 1,
-  ONNXRUNTIME = 2,
-};
-
-/**
-   Error code for ONNXRuntime.
-*/
-enum StatusCode {
-  OK = 0,
-  FAIL = 1,
-  INVALID_ARGUMENT = 2,
-  NO_SUCHFILE = 3,
-  NO_MODEL = 4,
-  ENGINE_ERROR = 5,
-  RUNTIME_EXCEPTION = 6,
-  INVALID_PROTOBUF = 7,
-  MODEL_LOADED = 8,
-  NOT_IMPLEMENTED = 9,
-  INVALID_GRAPH = 10,
-  EP_FAIL = 11
-};
-
-constexpr const char* StatusCodeToString(StatusCode status) noexcept {
-  switch (status) {
-    case StatusCode::OK:
-      return "SUCCESS";
-    case StatusCode::FAIL:
-      return "FAIL";
-    case StatusCode::INVALID_ARGUMENT:
-      return "INVALID_ARGUMENT";
-    case StatusCode::NO_SUCHFILE:
-      return "NO_SUCHFILE";
-    case StatusCode::NO_MODEL:
-      return "NO_MODEL";
-    case StatusCode::ENGINE_ERROR:
-      return "ENGINE_ERROR";
-    case StatusCode::RUNTIME_EXCEPTION:
-      return "RUNTIME_EXCEPTION";
-    case StatusCode::INVALID_PROTOBUF:
-      return "INVALID_PROTOBUF";
-    case StatusCode::MODEL_LOADED:
-      return "MODEL_LOADED";
-    case StatusCode::NOT_IMPLEMENTED:
-      return "NOT_IMPLEMENTED";
-    case StatusCode::INVALID_GRAPH:
-      return "INVALID_GRAPH";
-    case StatusCode::EP_FAIL:
-      return "EP_FAIL";
-    default:
-      return "GENERAL ERROR";
-  }
-}
-
-#ifdef _WIN32
-constexpr HRESULT StatusCodeToHRESULT(StatusCode status) noexcept {
-  switch (status) {
-    case StatusCode::OK:
-      return S_OK;
-    case StatusCode::FAIL:
-      return E_FAIL;
-    case StatusCode::INVALID_ARGUMENT:
-      return E_INVALIDARG;
-    case StatusCode::NO_SUCHFILE:
-      return HRESULT_FROM_WIN32(ERROR_FILE_NOT_FOUND);
-    case StatusCode::NO_MODEL:
-      return HRESULT_FROM_WIN32(ERROR_FILE_NOT_FOUND);
-    case StatusCode::ENGINE_ERROR:
-      return E_FAIL;
-    case StatusCode::RUNTIME_EXCEPTION:
-      return E_FAIL;
-    case StatusCode::INVALID_PROTOBUF:
-      return HRESULT_FROM_WIN32(ERROR_FILE_CORRUPT);
-    case StatusCode::MODEL_LOADED:
-      return HRESULT_FROM_WIN32(ERROR_INTERNAL_ERROR);
-    case StatusCode::NOT_IMPLEMENTED:
-      return E_NOTIMPL;
-    case StatusCode::INVALID_GRAPH:
-      return HRESULT_FROM_WIN32(ERROR_FILE_CORRUPT);
-    case StatusCode::EP_FAIL:
-      return HRESULT_FROM_WIN32(ERROR_INTERNAL_ERROR);
-    default:
-      return E_FAIL;
-  }
-}
-#endif
-
-class [[nodiscard]] Status {
- public:
-  Status() noexcept = default;
-
-  Status(StatusCategory category, int code, const std::string& msg);
-
-  Status(StatusCategory category, int code, const char* msg);
-
-  Status(StatusCategory category, int code);
-
-  GSL_SUPPRESS(r.11)
-  Status(const Status& other)
-      : state_((other.state_ == nullptr) ? nullptr : new State(*other.state_)) {}
-  GSL_SUPPRESS(r.11)
-  Status& operator=(const Status& other) {
-    if (state_ != other.state_) {
-      if (other.state_ == nullptr) {
-        state_.reset();
-      } else {
-        state_.reset(new State(*other.state_));
-      }
-    }
-    return *this;
-  }
-
-  Status(Status&&) = default;
-  Status& operator=(Status&&) = default;
-  ~Status() = default;
-
-  bool IsOK() const {
-    return (state_ == nullptr);
-  }
-
-  int Code() const noexcept;
-
-  StatusCategory Category() const noexcept;
-
-  const std::string& ErrorMessage() const noexcept;
-
-  std::string ToString() const;
-
-  bool operator==(const Status& other) const {
-    return (this->state_ == other.state_) || (ToString() == other.ToString());
-  }
-
-  bool operator!=(const Status& other) const {
-    return !(*this == other);
-  }
-
-  static Status OK() {
-    return Status();
-  }
-
- private:
-  static const std::string& EmptyString() noexcept;
-
-  struct State {
-    State(StatusCategory cat0, int code0, const std::string& msg0)
-        : category(cat0), code(code0), msg(msg0) {}
-
-    State(StatusCategory cat0, int code0, const char* msg0)
-        : category(cat0), code(code0), msg(msg0) {}
-
-    const StatusCategory category;
-    const int code;
-    const std::string msg;
-  };
-
-  // As long as Code() is OK, state_ == nullptr.
-  std::unique_ptr<State> state_;
-};
-
-inline std::ostream& operator<<(std::ostream& out, const Status& status) {
-  return out << status.ToString();
-}
-
-}  // namespace common
-
-// make Status directly available in the onnxruntime namespace as it is widely used
-using common::Status;
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/common/string_helper.h b/projects/llm_framework/include/onnxruntime/core/common/string_helper.h
deleted file mode 100644
index 13043031..00000000
--- a/projects/llm_framework/include/onnxruntime/core/common/string_helper.h
+++ /dev/null
@@ -1,11 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-#include <string>
-
-// forward declaration
-struct OrtAllocator;
-namespace onnxruntime {
-char* StrDup(const std::string& str, OrtAllocator* allocator);
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/alloc_kind.h b/projects/llm_framework/include/onnxruntime/core/framework/alloc_kind.h
deleted file mode 100644
index c7a953a4..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/alloc_kind.h
+++ /dev/null
@@ -1,36 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-#include <iosfwd>
-
-namespace onnxruntime {
-// The ml-Values fall into the following categories with respect to their
-// memory management:
-//   - inference inputs: owned (allocated and freed) by caller, and is by
-//     default read-only by the runtime.
-//   - inference outputs: allocated by runtime, ownership transferred to
-//     caller. TODO: Make sure this semantics is clear in InferenceSession API.
-//   - weights (constant tensors): can be allocated once (statically), and
-//     reused by all inference calls within an InferenceSession.
-//   - tensor values: The lifetimes of these tensor-values are statically
-//     determined, which is used for memory reuse/sharing optimizations. The
-//     runtime allocates/frees these values at the right time (as determined
-//     by the static allocation plan). Note that this is simplified since we
-//     do not try to optimize for "slice" like ops, where we may be able to
-//     conditionally reuse memory/data in some cases but not others.
-//     Generalizing this is future work.
-
-enum class AllocKind {
-  kNotSet = -1,
-  kAllocate = 0,
-  kReuse = 1,
-  kPreExisting = 2,
-  kAllocateStatically = 3,
-  kAllocateOutput = 4,
-  kShare = 5,
-  kAllocatedExternally = 6
-};
-
-std::ostream& operator<<(std::ostream& out, AllocKind alloc_kind);
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/allocator.h b/projects/llm_framework/include/onnxruntime/core/framework/allocator.h
deleted file mode 100644
index de3b35f3..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/allocator.h
+++ /dev/null
@@ -1,194 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include "core/common/common.h"
-#include "core/framework/allocator_stats.h"
-#include "core/session/onnxruntime_c_api.h"
-#include "ortdevice.h"
-#include "ortmemoryinfo.h"
-
-// This configures the arena based allocator used by ORT
-// See docs/C_API.md for details on what these mean and how to choose these values
-struct OrtArenaCfg {
-  OrtArenaCfg() : max_mem(0),
-                  arena_extend_strategy(-1),
-                  initial_chunk_size_bytes(-1),
-                  max_dead_bytes_per_chunk(-1),
-                  initial_growth_chunk_size_bytes(-1) {}
-  OrtArenaCfg(size_t max_mem, int arena_extend_strategy, int initial_chunk_size_bytes,
-              int max_dead_bytes_per_chunk, int initial_growth_chunk_size_bytes)
-      : max_mem(max_mem),
-        arena_extend_strategy(arena_extend_strategy),
-        initial_chunk_size_bytes(initial_chunk_size_bytes),
-        max_dead_bytes_per_chunk(max_dead_bytes_per_chunk),
-        initial_growth_chunk_size_bytes(initial_growth_chunk_size_bytes) {}
-
-  size_t max_mem;                       // use 0 to allow ORT to choose the default
-  int arena_extend_strategy;            // use -1 to allow ORT to choose the default, 0 = kNextPowerOfTwo, 1 = kSameAsRequested
-  int initial_chunk_size_bytes;         // use -1 to allow ORT to choose the default
-  int max_dead_bytes_per_chunk;         // use -1 to allow ORT to choose the default
-  int initial_growth_chunk_size_bytes;  // use -1 to allow ORT to choose the default
-};
-
-namespace onnxruntime {
-constexpr const char* CPU = "Cpu";
-constexpr const char* CUDA = "Cuda";
-constexpr const char* CUDA_PINNED = "CudaPinned";
-constexpr const char* CANN = "Cann";
-constexpr const char* CANN_PINNED = "CannPinned";
-constexpr const char* DML = "DML";
-constexpr const char* HIP = "Hip";
-constexpr const char* HIP_PINNED = "HipPinned";
-constexpr const char* OpenVINO_CPU = "OpenVINO_CPU";
-constexpr const char* OpenVINO_GPU = "OpenVINO_GPU";
-
-constexpr size_t kAllocAlignment = 256;
-
-class IAllocator;
-class Stream;
-namespace synchronize {
-class Notification;
-}
-using WaitNotificationFn = std::function<void(Stream&, synchronize::Notification&)>;
-void* AllocateBufferWithOptions(IAllocator& allocator, size_t size, bool use_reserve, Stream* stream, WaitNotificationFn wait_fn);
-
-template <typename T>
-using IAllocatorUniquePtr = std::unique_ptr<T, std::function<void(T*)>>;
-
-class IAllocator {
- public:
-  IAllocator(const OrtMemoryInfo& info) : memory_info_(info) {}
-  virtual ~IAllocator() = default;
-  /**
-  @remarks Use SafeInt when calculating the size of memory to allocate using Alloc.
-  */
-  virtual void* Alloc(size_t size) = 0;
-
-  virtual void Free(void* p) = 0;
-
-  // TODO: Find a better name than Reserve() and update in all places.
-  // Reserve() is an interface exposed for an implementation of IAllocator
-  // to optionally implement some allocation logic that by-passes any arena-based
-  // logic that may be housed in the Alloc() implementation.
-  // There are SessionOptions config(s) that allow users to allocate some memory
-  // by-passing arena-based logic.
-  // By default, the base implementation  just calls Alloc().
-  virtual void* Reserve(size_t size) { return Alloc(size); }
-
-  const OrtMemoryInfo& Info() const { return memory_info_; };
-
-  // Each implementation of IAllocator can override and provide their own implementation
-  virtual void GetStats(AllocatorStats* /*stats*/) { return; }
-
-  static bool CalcMemSizeForArray(size_t nmemb, size_t size, size_t* out) noexcept {
-    return CalcMemSizeForArrayWithAlignment(nmemb, size, 0, out);
-  }
-
-  /**
-   * Calculate the memory size for an array. The size is bounds checked using SafeInt.
-   * \tparam alignment must be power of 2
-   * \param nmemb Number of members or elements in the array
-   * \param size Size of each element
-   * \param out Total size required after any alignment is applied
-   * \return true, successful. false, overflow
-   */
-  [[nodiscard]] static bool CalcMemSizeForArrayWithAlignment(size_t nmemb, size_t size, size_t alignment, size_t* out) noexcept;
-
-  /**
-   * https://cwe.mitre.org/data/definitions/190.html
-   * \param alignment must be power of 2
-   * \param nmemb Number of members or elements in the array
-   * \param size Size of each element
-   * \param out Total size required after any alignment is applied
-   * \return true, successful. false, overflow
-   * \remarks This was the original API and was implemented in the header. Replaced with the above version
-   *          implemented in the .cc file so that the SafeInt dependency is internal.
-   */
-  template <size_t alignment>
-  [[nodiscard]] static bool CalcMemSizeForArrayWithAlignment(size_t nmemb, size_t size, size_t* out) noexcept;
-
-  /**
-   * allocate memory for an array which has nmemb items of data, each size bytes long
-   */
-  void* AllocArray(size_t nmemb, size_t size) {
-    size_t len;
-    if (!CalcMemSizeForArray(nmemb, size, &len))
-      return nullptr;
-    return Alloc(len);
-  }
-
-  /**
-   * allocate memory for an array which has nmemb items of data, each size bytes long
-   */
-  template <size_t alignment>
-  void* AllocArrayWithAlignment(size_t nmemb, size_t size) {
-    size_t len;
-    if (!CalcMemSizeForArrayWithAlignment(nmemb, size, alignment, &len))
-      return nullptr;
-    return Alloc(len);
-  }
-
-  /**
-     Create a std::unique_ptr that is allocated and freed by the provided IAllocator.
-     @param allocator The allocator.
-     @param count_or_bytes The exact bytes to allocate if T is void, otherwise the number of elements to allocate.
-     @param use_reserve If true, call Reserve() instead of Alloc() to allocate memory.
-     @param stream Which stream instance allocated chunk will be used with.
-     @param wait_fn If the allocator want to dynamic reuse a chunk from another stream, use this wait_fn to sync on
-                    the target stream to make the reuse safe.
-     @returns std::unique_ptr with allocated memory and deleter.
-  */
-  template <typename T>
-  static IAllocatorUniquePtr<T> MakeUniquePtr(std::shared_ptr<IAllocator> allocator, size_t count_or_bytes,
-                                              bool use_reserve = false,
-                                              Stream* stream = nullptr, WaitNotificationFn wait_fn = nullptr) {
-    if (allocator == nullptr) return nullptr;
-    // for now limit to fundamental types. we could support others, but to do so either we or the caller
-    // needs to call the dtor for the objects, for buffers allocated on device we don't have destructor
-    // static_assert(std::is_fundamental<T>::value, "Fundamental type required as no destructors are called.");
-
-    size_t alloc_size = count_or_bytes;
-
-    // if T is not void, 'count_or_bytes' == number of items so allow for that
-    if constexpr (!std::is_void<T>::value) {
-      // sizeof(void) isn't valid, but the compiler isn't smart enough to ignore that this line isn't
-      // reachable if T is void. use std::conditional to 'use' void* in the sizeof call
-      if (!CalcMemSizeForArray(
-              count_or_bytes, sizeof(typename std::conditional<std::is_void<T>::value, void*, T>::type), &alloc_size)) {
-        return nullptr;
-      }
-    }
-
-    // allocate
-    T* p = static_cast<T*>(AllocateBufferWithOptions(*allocator, alloc_size, use_reserve, stream, std::move(wait_fn)));
-    return IAllocatorUniquePtr<T>{
-        p,
-        [allocator = std::move(allocator)](T* p) { allocator->Free(p); }};
-  }
-
- private:
-  OrtMemoryInfo memory_info_;
-};
-
-template <size_t alignment>
-bool IAllocator::CalcMemSizeForArrayWithAlignment(size_t nmemb, size_t size, size_t* out) noexcept {
-  return CalcMemSizeForArrayWithAlignment(nmemb, size, alignment, out);
-}
-
-class CPUAllocator : public IAllocator {
- public:
-  explicit CPUAllocator(const OrtMemoryInfo& memory_info) : IAllocator(memory_info) {}
-
-  CPUAllocator() : IAllocator(OrtMemoryInfo(CPU, OrtAllocatorType::OrtDeviceAllocator)) {}
-
-  void* Alloc(size_t size) override;
-  void Free(void* p) override;
-};
-
-using AllocatorPtr = std::shared_ptr<IAllocator>;
-
-void* AllocatorDefaultAlloc(size_t size);
-void AllocatorDefaultFree(void* p);
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/buffer_deleter.h b/projects/llm_framework/include/onnxruntime/core/framework/buffer_deleter.h
deleted file mode 100644
index 961eb443..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/buffer_deleter.h
+++ /dev/null
@@ -1,36 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include "core/framework/allocator.h"
-
-namespace onnxruntime {
-
-// TODO: Do we need this class or is IAllocator::MakeUniquePtr sufficient/better
-class BufferDeleter {
- public:
-  BufferDeleter() = default;
-  explicit BufferDeleter(AllocatorPtr alloc)
-      : alloc_(std::move(alloc)) {}
-
-  void operator()(void* p) const {
-    if (alloc_)
-      alloc_->Free(p);
-  }
-
- private:
-  // TODO: we may need consider the lifetime of alloc carefully
-  // The alloc_ here is the allocator that used to allocate the buffer
-  // And need go with the unique_ptr together. If it is using our internal
-  // allocator, it is ok as our allocators are global managed. But if it
-  // is provide by user, user need to be very careful about it.
-  // A weak_ptr may be a choice to reduce the impact, but that require to
-  // change our current allocator mgr to use shared_ptr. Will revisit it
-  // later.
-  AllocatorPtr alloc_{nullptr};
-};
-
-using BufferUniquePtr = std::unique_ptr<void, BufferDeleter>;
-using BufferNakedPtr = void*;
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/customregistry.h b/projects/llm_framework/include/onnxruntime/core/framework/customregistry.h
deleted file mode 100644
index 52f6169e..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/customregistry.h
+++ /dev/null
@@ -1,60 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include "core/common/status.h"
-#include "core/common/logging/logging.h"
-#include "core/framework/op_kernel.h"
-#include "core/framework/kernel_def_builder.h"
-#include "core/framework/kernel_registry.h"
-
-#if !defined(ORT_MINIMAL_BUILD)
-#include "core/graph/schema_registry.h"
-#endif
-
-namespace onnxruntime {
-
-/**
-   Represents a registry that contains both custom kernels and custom schemas.
-*/
-class CustomRegistry final {
- public:
-  CustomRegistry()
-      : kernel_registry_(std::make_shared<KernelRegistry>())
-#if !defined(ORT_MINIMAL_BUILD)
-        ,
-        opschema_registry_(std::make_shared<onnxruntime::OnnxRuntimeOpSchemaRegistry>())
-#endif
-  {
-  }
-
-  /**
-   * Register a kernel definition together with kernel factory method to this session.
-   * If any conflict happened between registered kernel def and built-in kernel def,
-   * registered kernel will have higher priority.
-   * Call this before invoking Initialize().
-   * @return OK if success.
-   */
-  common::Status RegisterCustomKernel(KernelDefBuilder& kernel_def_builder, const KernelCreateFn& kernel_creator);
-
-  common::Status RegisterCustomKernel(KernelCreateInfo&);
-
-  const std::shared_ptr<KernelRegistry>& GetKernelRegistry();
-
-#if !defined(ORT_MINIMAL_BUILD)
-  common::Status RegisterOpSet(std::vector<ONNX_NAMESPACE::OpSchema>& schemas, const std::string& domain,
-                               int baseline_opset_version, int opset_version);
-
-  const std::shared_ptr<onnxruntime::OnnxRuntimeOpSchemaRegistry>& GetOpschemaRegistry();
-#endif
-
- private:
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(CustomRegistry);
-  std::shared_ptr<KernelRegistry> kernel_registry_;
-#if !defined(ORT_MINIMAL_BUILD)
-  std::shared_ptr<onnxruntime::OnnxRuntimeOpSchemaRegistry> opschema_registry_;
-#endif
-};
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/data_types.h b/projects/llm_framework/include/onnxruntime/core/framework/data_types.h
deleted file mode 100644
index 6dc38c27..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/data_types.h
+++ /dev/null
@@ -1,1062 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <cstdint>
-#include <cstring>
-#include <string>
-#include <type_traits>
-#include <map>
-#include <unordered_map>
-#include "core/common/gsl.h"
-#include "core/common/common.h"
-#include "core/common/exceptions.h"
-#include "core/framework/endian.h"
-#include "core/framework/float16.h"
-#include "core/framework/to_tensor_proto_element_type.h"
-#if !defined(ORT_MINIMAL_BUILD)
-#include "onnx/defs/schema.h"
-#else
-#include "onnx/defs/data_type_utils.h"
-#endif
-#include "onnx/onnx_pb.h"
-#include "onnx/onnx-operators_pb.h"
-
-struct OrtValue;
-
-namespace ONNX_NAMESPACE {
-class TypeProto;
-}  // namespace ONNX_NAMESPACE
-
-namespace onnxruntime {
-/// Predefined registered types
-
-#if !defined(DISABLE_ML_OPS)
-
-// maps (only used by ML ops)
-using MapStringToString = std::map<std::string, std::string>;
-using MapStringToInt64 = std::map<std::string, int64_t>;
-using MapStringToFloat = std::map<std::string, float>;
-using MapStringToDouble = std::map<std::string, double>;
-using MapInt64ToString = std::map<int64_t, std::string>;
-using MapInt64ToInt64 = std::map<int64_t, int64_t>;
-using MapInt64ToFloat = std::map<int64_t, float>;
-using MapInt64ToDouble = std::map<int64_t, double>;
-
-// vectors/sequences
-using VectorMapStringToFloat = std::vector<MapStringToFloat>;
-using VectorMapInt64ToFloat = std::vector<MapInt64ToFloat>;
-
-#endif
-
-using VectorString = std::vector<std::string>;
-using VectorInt64 = std::vector<int64_t>;
-
-// Forward declarations
-class DataTypeImpl;
-class TensorTypeBase;
-#if !defined(DISABLE_SPARSE_TENSORS)
-class SparseTensorTypeBase;
-#endif
-class SequenceTensorTypeBase;
-class NonTensorTypeBase;
-#if !defined(DISABLE_OPTIONAL_TYPE)
-class OptionalTypeBase;
-#endif
-class PrimitiveDataTypeBase;
-class Tensor;
-class TensorSeq;
-
-// DataTypeImpl pointer as unique DataTypeImpl identifier.
-using MLDataType = const DataTypeImpl*;
-// be used with class MLValue
-using DeleteFunc = void (*)(void*);
-using CreateFunc = void* (*)();
-
-/**
- * \brief Base class for MLDataType
- *
- */
-class DataTypeImpl {
- public:
-  enum class GeneralType {
-    kInvalid = 0,
-    kNonTensor = 1,
-    kTensor = 2,
-    kTensorSequence = 3,
-    kSparseTensor = 4,
-    kOptional = 5,
-    kPrimitive = 6,
-  };
-
-  const GeneralType type_;
-  const size_t size_;
-
- protected:
-  DataTypeImpl(GeneralType type, size_t size) : type_{type}, size_{size} {}
-
- public:
-  virtual ~DataTypeImpl() = default;
-
-  /**
-   * \brief this API will be used to check type compatibility at runtime
-   *
-   * \param type_proto a TypeProto instance that is constructed for a specific type
-   *        will be checked against a TypeProto instance contained within a corresponding
-   *        MLDataType instance.
-   */
-  virtual bool IsCompatible(const ONNX_NAMESPACE::TypeProto& type_proto) const = 0;
-
-  size_t Size() const { return size_; }
-
-  virtual DeleteFunc GetDeleteFunc() const = 0;
-
-  /**
-   * \brief Retrieves an instance of TypeProto for
-   *        a given MLDataType
-   * \returns optional TypeProto. Only ONNX types
-              has type proto, non-ONNX types will return nullptr.
-   */
-  virtual const ONNX_NAMESPACE::TypeProto* GetTypeProto() const = 0;
-
-  bool IsTensorType() const {
-    return type_ == GeneralType::kTensor;
-  }
-
-  bool IsTensorSequenceType() const {
-    return type_ == GeneralType::kTensorSequence;
-  }
-
-  bool IsSparseTensorType() const {
-    return type_ == GeneralType::kSparseTensor;
-  }
-
-  bool IsOptionalType() const {
-    return type_ == GeneralType::kOptional;
-  }
-
-  bool IsNonTensorType() const {
-    return type_ == GeneralType::kNonTensor;
-  }
-
-  bool IsPrimitiveDataType() const {
-    return type_ == GeneralType::kPrimitive;
-  }
-
-  // Returns this if this is of tensor-type and null otherwise
-  const TensorTypeBase* AsTensorType() const;
-
-  const SequenceTensorTypeBase* AsSequenceTensorType() const;
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-  // Returns this if this is of sparse-tensor-type and null otherwise
-  const SparseTensorTypeBase* AsSparseTensorType() const;
-#endif
-
-#if !defined(DISABLE_OPTIONAL_TYPE)
-  const OptionalTypeBase* AsOptionalType() const;
-#endif
-
-  const NonTensorTypeBase* AsNonTensorType() const;
-
-  // Returns this if this is one of the primitive data types (specialization of PrimitiveDataTypeBase)
-  // and null otherwise
-  const PrimitiveDataTypeBase* AsPrimitiveDataType() const;
-
-  // Return the type meta that we are using in the runtime.
-  template <typename T>
-  static MLDataType GetType();
-
-  // Return the types for a concrete tensor type, like Tensor_Float
-  template <typename elemT>
-  static MLDataType GetTensorType();
-
-  template <typename elemT>
-  static MLDataType GetSequenceTensorType();
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-  // Return the MLDataType for a concrete sparse tensor type.
-  template <typename elemT>
-  static MLDataType GetSparseTensorType();
-#endif
-
-  template <typename T, typename elemT>
-  static MLDataType GetOptionalType();
-
-  /**
-   * Convert an ONNX TypeProto to onnxruntime DataTypeImpl.
-   * However, this conversion is lossy. Don't try to use 'this->GetTypeProto()' converting it back.
-   * Even though GetTypeProto() will not have the original information, it will still have enough to correctly
-   * map to MLDataType.
-   * \param proto
-   */
-  static MLDataType TypeFromProto(const ONNX_NAMESPACE::TypeProto& proto);
-
-  static const TensorTypeBase* TensorTypeFromONNXEnum(int type);
-  static const SequenceTensorTypeBase* SequenceTensorTypeFromONNXEnum(int type);
-#if !defined(DISABLE_SPARSE_TENSORS)
-  static const SparseTensorTypeBase* SparseTensorTypeFromONNXEnum(int type);
-#endif
-
-  static const char* ToString(MLDataType type);
-  static std::vector<std::string> ToString(const std::vector<MLDataType>& types);
-  // Registers ONNX_NAMESPACE::DataType (internalized string) with
-  // MLDataType. DataType is produced by internalizing an instance of
-  // TypeProto contained within MLDataType
-  static void RegisterDataType(MLDataType);
-  static MLDataType GetDataType(const std::string&);
-
-  static const std::vector<MLDataType>& AllTensorTypes();
-  static const std::vector<MLDataType>& AllFixedSizeTensorTypes();
-  static const std::vector<MLDataType>& AllSequenceTensorTypes();
-  static const std::vector<MLDataType>& AllFixedSizeSequenceTensorTypes();
-  static const std::vector<MLDataType>& AllNumericTensorTypes();
-  static const std::vector<MLDataType>& AllIEEEFloatTensorTypes();
-  static const std::vector<MLDataType>& AllFixedSizeTensorExceptHalfTypes();
-  static const std::vector<MLDataType>& AllIEEEFloatTensorExceptHalfTypes();
-  static const std::vector<MLDataType>& AllTensorAndSequenceTensorTypes();
-  static const std::vector<MLDataType>& AllFixedSizeTensorAndSequenceTensorTypes();
-  static const std::vector<MLDataType>& AllOptionalTypes();
-  static const std::vector<MLDataType>& AllTensorAndSequenceTensorAndOptionalTypes();
-};
-
-std::ostream& operator<<(std::ostream& out, MLDataType data_type);
-
-/*
- * Type registration helpers
- */
-namespace data_types_internal {
-/// TensorType helpers
-///
-
-/// Is a given type on the list of types?
-/// Accepts a list of types and the first argument is the type
-/// We are checking if it is listed among those that follow
-template <typename T, typename... Types>
-struct IsAnyOf;
-
-/// Two types remaining, end of the list
-template <typename T, typename Tail>
-struct IsAnyOf<T, Tail> : public std::is_same<T, Tail> {
-};
-
-template <typename T, typename H, typename... Tail>
-struct IsAnyOf<T, H, Tail...> {
-  static constexpr bool value = (std::is_same<T, H>::value ||
-                                 IsAnyOf<T, Tail...>::value);
-};
-
-/// Tells if the specified type is one of fundamental types
-/// that can be contained within a tensor.
-/// We do not have raw fundamental types, rather a subset
-/// of fundamental types is contained within tensors.
-template <typename T>
-struct IsTensorContainedType : public IsAnyOf<T, float, uint8_t, int8_t, uint16_t, int16_t,
-                                              int32_t, int64_t, std::string, bool, MLFloat16,
-                                              double, uint32_t, uint64_t, BFloat16> {
-};
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-/// Use "IsSparseTensorContainedType<T>::value" to test if a type T
-/// is permitted as the element-type of a sparse-tensor.
-
-template <typename T>
-struct IsSparseTensorContainedType : public IsAnyOf<T, float, uint8_t, int8_t, uint16_t, int16_t,
-                                                    int32_t, int64_t, std::string, bool, MLFloat16,
-                                                    double, uint32_t, uint64_t, BFloat16> {
-};
-#endif
-
-#if !defined(DISABLE_OPTIONAL_TYPE)
-/// Tells if the specified type is one of ORT types
-/// that can be contained within an optional struct.
-template <typename T>
-struct IsOptionalOrtType : public IsAnyOf<T, Tensor, TensorSeq> {
-};
-#endif
-
-/// This template's Get() returns a corresponding MLDataType
-/// It dispatches the call to either GetTensorType<>() or
-/// GetType<>()
-template <typename T, bool TensorContainedType>
-struct GetMLDataType;
-
-template <typename T>
-struct GetMLDataType<T, true> {
-  static MLDataType Get() {
-    return DataTypeImpl::GetTensorType<T>();
-  }
-};
-
-template <typename T>
-struct GetMLDataType<T, false> {
-  static MLDataType Get() {
-    return DataTypeImpl::GetType<T>();
-  }
-};
-
-struct TensorTypeHelper {
-  static void Set(ONNX_NAMESPACE::TensorProto_DataType element_type,
-                  ONNX_NAMESPACE::TypeProto& proto) {
-    proto.mutable_tensor_type()->set_elem_type(element_type);
-  }
-};
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-struct SparseTensorTypeHelper {
-  static void Set(ONNX_NAMESPACE::TensorProto_DataType element_type,
-                  ONNX_NAMESPACE::TypeProto& proto) {
-    proto.mutable_sparse_tensor_type()->set_elem_type(element_type);
-  }
-};
-#endif  // !defined(DISABLE_SPARSE_TENSORS)
-
-#if !defined(DISABLE_ML_OPS)
-/// Map helpers
-
-void CopyMutableMapValue(const ONNX_NAMESPACE::TypeProto&,
-                         ONNX_NAMESPACE::TypeProto&);
-
-struct MapTypeHelper {
-  // V can be either a primitive type (in which case it is a tensor)
-  // or other preregistered types
-  template <typename V>
-  static MLDataType GetValueType() {
-    return GetMLDataType<V, IsTensorContainedType<V>::value>::Get();
-  }
-
-  static void Set(ONNX_NAMESPACE::TensorProto_DataType key_type, const ONNX_NAMESPACE::TypeProto* value_proto,
-                  ONNX_NAMESPACE::TypeProto& proto) {
-    ORT_ENFORCE(value_proto != nullptr, "expected a registered ONNX type");
-    proto.mutable_map_type()->set_key_type(key_type);
-    CopyMutableMapValue(*value_proto, proto);
-  }
-};
-#endif
-
-/// Sequence helpers
-
-// Element type is a primitive type so we set it to a tensor<elemT>
-void CopyMutableSeqElement(const ONNX_NAMESPACE::TypeProto&,
-                           ONNX_NAMESPACE::TypeProto&);
-
-// helper to create TypeProto with minimal binary size impact
-struct SequenceTypeHelper {
-  template <typename T>
-  static MLDataType GetElemType() {
-    return GetMLDataType<T, IsTensorContainedType<T>::value>::Get();
-  }
-
-  static void Set(const ONNX_NAMESPACE::TypeProto* elem_proto,
-                  ONNX_NAMESPACE::TypeProto& proto) {
-    ORT_ENFORCE(elem_proto != nullptr, "expected a registered ONNX type");
-    CopyMutableSeqElement(*elem_proto, proto);
-  }
-};
-
-/// Optional helpers
-
-void CopyMutableOptionalElement(const ONNX_NAMESPACE::TypeProto&,
-                                ONNX_NAMESPACE::TypeProto&);
-
-// helper to create TypeProto with minimal binary size impact
-struct OptionalTypeHelper {
-  template <typename T, typename elemT>
-  static MLDataType GetElemType() {
-    if constexpr (std::is_same<T, Tensor>::value) {
-      return DataTypeImpl::GetTensorType<elemT>();
-    } else {
-      static_assert(std::is_same<T, TensorSeq>::value, "Unsupported element type for optional type");
-      return DataTypeImpl::GetSequenceTensorType<elemT>();
-    }
-  }
-
-  static void Set(const onnx::TypeProto* elem_proto, ONNX_NAMESPACE::TypeProto& proto) {
-    ORT_ENFORCE(elem_proto != nullptr, "expected a registered ONNX type");
-    CopyMutableOptionalElement(*elem_proto, proto);
-  }
-};
-
-/// OpaqueTypes helpers
-
-void AssignOpaqueDomainName(const char* domain, const char* name,
-                            ONNX_NAMESPACE::TypeProto& proto);
-
-}  // namespace data_types_internal
-
-//The suppressed warning is: "The type with a virtual function needs either public virtual or protected nonvirtual destructor."
-//However, we do not allocate this type on heap.
-#if defined(_MSC_VER) && !defined(__clang__)
-#pragma warning(push)
-#pragma warning(disable : 26436)
-#endif
-/// All tensors base
-class TensorTypeBase : public DataTypeImpl {
- public:
-  static MLDataType Type();
-
-  /// We first compare type_proto pointers and then
-  /// if they do not match try to account for the case
-  /// where TypeProto was created ad-hoc and not queried from MLDataType
-  bool IsCompatible(const ONNX_NAMESPACE::TypeProto& type_proto) const override;
-
-  DeleteFunc GetDeleteFunc() const override;
-
-  const ONNX_NAMESPACE::TypeProto* GetTypeProto() const override;
-
-  virtual MLDataType GetElementType() const {
-    // should never reach here.
-    ORT_NOT_IMPLEMENTED(__FUNCTION__, " is not implemented");
-  }
-
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(TensorTypeBase);
-
- protected:
-  ONNX_NAMESPACE::TypeProto& MutableTypeProto();
-
-  TensorTypeBase();
-  ~TensorTypeBase() override;
-
- private:
-  struct Impl;
-  Impl* impl_;
-};
-
-/**
- * \brief Tensor type. This type does not have a C++ type associated with
- * it at registration time except the element type. One of the types mentioned
- * above at IsTensorContainedType<> list is acceptable.
- *
- * \details
- *        Usage:
- *        ORT_REGISTER_TENSOR(ELEMENT_TYPE)
- *        Currently all of the Tensors irrespective of the dimensions are mapped to Tensor<type>
- *        type. IsCompatible() currently ignores shape.
- */
-
-template <typename elemT>
-class TensorType : public TensorTypeBase {
- public:
-  static_assert(data_types_internal::IsTensorContainedType<elemT>::value,
-                "Requires one of the tensor fundamental types");
-
-  static MLDataType Type();
-
-  /// Tensors only can contain basic data types
-  /// that have been previously registered with ONNXRuntime
-  MLDataType GetElementType() const override {
-    return DataTypeImpl::GetType<elemT>();
-  }
-
- private:
-  TensorType() {
-    using namespace data_types_internal;
-    TensorTypeHelper::Set(utils::ToTensorProtoElementType<elemT>(), MutableTypeProto());
-  }
-};
-
-#if defined(DISABLE_OPTIONAL_TYPE)
-
-// TODO is this still needed after removing kernel def hashes?
-/// Common base-class for all disabled types. We need DataTypeImpl::ToString to work in a minimal build
-/// with disabled types to keep the ORT format model kernel hashes stable.
-class DisabledTypeBase : public DataTypeImpl {
- public:
-  static MLDataType Type();
-
-  bool IsCompatible(const ONNX_NAMESPACE::TypeProto&) const override {
-    // We always want to return false for the IsCompatible() for a disabled type
-    // because this will ensure that no kernel supporting the disabled type will
-    // be matched to a model node requiring that type and the model load will
-    // result in failure.
-    return false;
-  }
-
-  DeleteFunc GetDeleteFunc() const override {
-    ORT_THROW("Type is disabled in this build.");
-  }
-
-  // This must work
-  const ONNX_NAMESPACE::TypeProto* GetTypeProto() const override;
-
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(DisabledTypeBase);
-
- protected:
-  // This must work
-  ONNX_NAMESPACE::TypeProto& MutableTypeProto();
-
-  DisabledTypeBase(DataTypeImpl::GeneralType type, size_t size);
-  ~DisabledTypeBase() override;
-
- private:
-  struct Impl;
-  Impl* impl_;
-};
-
-#endif
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-/// Common base-class for all sparse-tensors (with different element types).
-class SparseTensorTypeBase : public DataTypeImpl {
- public:
-  static MLDataType Type();
-
-  bool IsCompatible(const ONNX_NAMESPACE::TypeProto& type_proto) const override;
-
-  DeleteFunc GetDeleteFunc() const override;
-
-  const ONNX_NAMESPACE::TypeProto* GetTypeProto() const override;
-
-  virtual MLDataType GetElementType() const {
-    // should never reach here.
-    ORT_NOT_IMPLEMENTED(__FUNCTION__, " is not implemented");
-  }
-
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(SparseTensorTypeBase);
-
- protected:
-  ONNX_NAMESPACE::TypeProto& MutableTypeProto();
-
-  SparseTensorTypeBase();
-  ~SparseTensorTypeBase() override;
-
- private:
-  struct Impl;
-  Impl* impl_;
-};
-
-template <typename elemT>
-class SparseTensorType : public SparseTensorTypeBase {
- public:
-  static_assert(data_types_internal::IsSparseTensorContainedType<elemT>::value,
-                "Requires one of the sparse-tensor fundamental types");
-
-  static MLDataType Type();
-
-  /// Return a MLDataType representing the element-type
-  MLDataType GetElementType() const override {
-    return DataTypeImpl::GetType<elemT>();
-  }
-
- private:
-  SparseTensorType() {
-    using namespace data_types_internal;
-    SparseTensorTypeHelper::Set(utils::ToTensorProtoElementType<elemT>(), MutableTypeProto());
-  }
-};
-
-#endif  // !defined(DISABLE_SPARSE_TENSORS)
-
-/// Common base-class for all optional types.
-
-#if !defined(DISABLE_OPTIONAL_TYPE)
-class OptionalTypeBase : public DataTypeImpl {
- public:
-  static MLDataType Type();
-
-  bool IsCompatible(const ONNX_NAMESPACE::TypeProto& type_proto) const override;
-
-  DeleteFunc GetDeleteFunc() const override {
-    // should never reach here.
-    ORT_NOT_IMPLEMENTED(__FUNCTION__, " is not implemented");
-  }
-
-  const ONNX_NAMESPACE::TypeProto* GetTypeProto() const override;
-
-  virtual MLDataType GetElementType() const {
-    // should never reach here.
-    ORT_NOT_IMPLEMENTED(__FUNCTION__, " is not implemented");
-  }
-
-  OptionalTypeBase(const OptionalTypeBase&) = delete;
-  OptionalTypeBase& operator=(const OptionalTypeBase&) = delete;
-
- protected:
-  ONNX_NAMESPACE::TypeProto& MutableTypeProto();
-
-  OptionalTypeBase();
-  ~OptionalTypeBase() override;
-
- private:
-  struct Impl;
-  Impl* impl_;
-};
-#endif
-
-// Derive from OptionalTypeBase if the Optional type support is enabled,
-// else derive from DisabledTypeBase
-template <typename T, typename elemT>
-class OptionalType :
-#if !defined(DISABLE_OPTIONAL_TYPE)
-    public OptionalTypeBase
-#else
-    public DisabledTypeBase
-#endif
-{
- public:
-  static MLDataType Type();
-
-#if !defined(DISABLE_OPTIONAL_TYPE)
-  static_assert(data_types_internal::IsOptionalOrtType<T>::value,
-                "Requires one of the supported types: Tensor or TensorSeq");
-
-  static_assert(data_types_internal::IsTensorContainedType<elemT>::value,
-                "Requires one of the tensor fundamental types");
-
-  MLDataType GetElementType() const override {
-    return data_types_internal::OptionalTypeHelper::GetElemType<T, elemT>();
-  }
-#endif
-
- private:
-#if !defined(DISABLE_OPTIONAL_TYPE)
-  OptionalType()
-#else
-  OptionalType() : DisabledTypeBase { DataTypeImpl::GeneralType::kOptional, 0 }
-#endif
-  {
-    using namespace data_types_internal;
-    OptionalTypeHelper::Set(OptionalTypeHelper::GetElemType<T, elemT>()->GetTypeProto(), MutableTypeProto());
-  }
-};  // namespace onnxruntime
-
-/**
- * \brief Provide a specialization for your C++ Non-tensor type
- *        so your implementation FromDataTypeContainer/ToDataTypeContainer
- *        functions correctly. Otherwise you get a default implementation
- *        which may not be what you need/want.
- *
- * This class is used to create OrtValue, fetch data from OrtValue via
- * C/C++ APIs
- */
-template <class T>
-struct NonTensorTypeConverter {
-  static void FromContainer(MLDataType /*dtype*/, const void* /*data*/, size_t /*data_size*/, OrtValue& /*output*/) {
-    ORT_THROW("Not implemented");
-  }
-  static void ToContainer(const OrtValue& /*input*/, size_t /*data_size*/, void* /*data*/) {
-    ORT_THROW("Not implemented");
-  }
-};
-
-/**
- * \brief Base type for all non-tensors, maps, sequences and opaques
- */
-class NonTensorTypeBase : public DataTypeImpl {
- public:
-  DeleteFunc GetDeleteFunc() const override = 0;
-
-  virtual CreateFunc GetCreateFunc() const = 0;
-
-  const ONNX_NAMESPACE::TypeProto* GetTypeProto() const override;
-
-  // \brief Override for Non-tensor types to initialize non-tensor CPP
-  // data representation from data. The caller of the interface
-  // should have a shared definition of the data which is used to initialize
-  // CPP data representation. This is used from C API.
-  //
-  // \param data - pointer to a data container structure non_tensor type specific
-  // \param data_size - size of the data container structure, used for rudimentary checks
-  // \param output - reference to a default constructed non-tensor type
-  // \returns OrtValue
-  // \throw if there is an error
-  virtual void FromDataContainer(const void* data, size_t data_size, OrtValue& output) const;
-
-  // \brief Override for Non-tensor types to fetch data from the internal CPP data representation
-  // The caller of the interface should have a shared definition of the data which is used to initialize
-  // CPP data representation. This is used from C API.
-  //
-  // \param input - OrtValue containing data
-  // \param data_size - size of the structure that is being passed for receiving data, used for
-  //                    validation
-  // \param data - pointer to receiving data structure
-  virtual void ToDataContainer(const OrtValue& input, size_t data_size, void* data) const;
-
-  NonTensorTypeBase(const NonTensorTypeBase&) = delete;
-  NonTensorTypeBase& operator=(const NonTensorTypeBase&) = delete;
-
- protected:
-  NonTensorTypeBase(size_t size);
-  ~NonTensorTypeBase() override;
-
-  ONNX_NAMESPACE::TypeProto& MutableTypeProto();
-
-  bool IsMapCompatible(const ONNX_NAMESPACE::TypeProto& type_proto) const;
-
-  bool IsSequenceCompatible(const ONNX_NAMESPACE::TypeProto& type_proto) const;
-
-  bool IsOpaqueCompatible(const ONNX_NAMESPACE::TypeProto& type_proto) const;
-
- private:
-  struct Impl;
-  Impl* impl_;
-};
-
-// This is where T is the actual CPPRuntimeType
-template <typename T>
-class NonTensorType : public NonTensorTypeBase {
- private:
-  static void Delete(void* p) {
-    delete static_cast<T*>(p);
-  }
-
- public:
-  DeleteFunc GetDeleteFunc() const override {
-    return &Delete;
-  }
-
-  CreateFunc GetCreateFunc() const override {
-    return []() -> void* { return new T(); };
-  }
-
- protected:
-  NonTensorType() : NonTensorTypeBase(sizeof(T)) {}
-};
-
-#if !defined(DISABLE_ML_OPS)
-/**
- * \brief MapType. Use this type to register
- * mapping types.
- *
- * \param T - cpp type that you wish to register as runtime MapType
- *
- * \details Usage: ORT_REGISTER_MAP(C++Type)
- *          The type is required to have mapped_type and
- *          key_type defined
- */
-template <typename CPPType>
-class MapType : public NonTensorType<CPPType> {
- public:
-  static_assert(data_types_internal::IsTensorContainedType<typename CPPType::key_type>::value,
-                "Requires one of the tensor fundamental types as key");
-
-  static MLDataType Type();
-
-  bool IsCompatible(const ONNX_NAMESPACE::TypeProto& type_proto) const override {
-    return this->IsMapCompatible(type_proto);
-  }
-
- private:
-  MapType() {
-    using namespace data_types_internal;
-    MapTypeHelper::Set(utils::ToTensorProtoElementType<typename CPPType::key_type>(),
-                       MapTypeHelper::GetValueType<typename CPPType::mapped_type>()->GetTypeProto(),
-                       this->MutableTypeProto());
-  }
-};
-#endif
-
-/**
- * \brief SequenceType. Use to register sequence for non-tensor types.
- *
- *  \param T - CPP type that you wish to register as Sequence
- *             runtime type.
- *
- * \details Usage: ORT_REGISTER_SEQ(C++Type)
- *          The type is required to have value_type defined
- */
-template <typename CPPType>
-class SequenceType : public NonTensorType<CPPType> {
- public:
-  static MLDataType Type();
-
-  bool IsCompatible(const ONNX_NAMESPACE::TypeProto& type_proto) const override {
-    return this->IsSequenceCompatible(type_proto);
-  }
-
- private:
-  SequenceType() {
-    using namespace data_types_internal;
-    SequenceTypeHelper::Set(SequenceTypeHelper::GetElemType<typename CPPType::value_type>()->GetTypeProto(),
-                            this->MutableTypeProto());
-  }
-};
-
-/**
- * \brief SequenceTensorTypeBase serves as a base type class for
- *        Tensor sequences. Akin to TensorTypeBase.
- *        Runtime representation is always TensorSeq.
- */
-class SequenceTensorTypeBase : public DataTypeImpl {
- public:
-  static MLDataType Type();
-
-  bool IsCompatible(const ONNX_NAMESPACE::TypeProto& type_proto) const override;
-
-  virtual MLDataType GetElementType() const {
-    // should never reach here.
-    ORT_NOT_IMPLEMENTED(__FUNCTION__, " is not implemented");
-  }
-
-  DeleteFunc GetDeleteFunc() const override;
-
-  const ONNX_NAMESPACE::TypeProto* GetTypeProto() const override;
-
-  SequenceTensorTypeBase(const SequenceTensorTypeBase&) = delete;
-  SequenceTensorTypeBase& operator=(const SequenceTensorTypeBase&) = delete;
-
- protected:
-  SequenceTensorTypeBase();
-  ~SequenceTensorTypeBase();
-
-  ONNX_NAMESPACE::TypeProto& MutableTypeProto();
-
- private:
-  struct Impl;
-  Impl* impl_;
-};
-#if defined(_MSC_VER) && !defined(__clang__)
-#pragma warning(pop)
-#endif
-/**
- * \brief SequenceTensorType. Use to register sequence for non-tensor types.
- *
- *  \param CPPRuntime - We always use TensorSeq
- *
- *  \param TensorElemType - one of the primitive types
- *
- * \details Usage: ORT_REGISTER_SEQ_TENSOR_TYPE()
- *          The type is required to have value_type defined
- */
-template <typename TensorElemType>
-class SequenceTensorType : public SequenceTensorTypeBase {
- public:
-  static_assert(data_types_internal::IsTensorContainedType<TensorElemType>::value,
-                "Requires one of the tensor fundamental types");
-
-  static MLDataType Type();
-
-  /// Return a MLDataType representing the element-type
-  MLDataType GetElementType() const override {
-    return DataTypeImpl::GetType<TensorElemType>();
-  }
-
- private:
-  SequenceTensorType() {
-    using namespace data_types_internal;
-    SequenceTypeHelper::Set(SequenceTypeHelper::GetElemType<TensorElemType>()->GetTypeProto(),
-                            MutableTypeProto());
-  }
-};
-
-/**
- * \brief OpaqueType
- *
- * \tparam T - cpp runtume that implements the Opaque type
- *
- * \tparam const char D[] - domain must be extern to be unique
- *
- * \tparam const char N[] - name must be extern to be unique
- *
- * \details Only one CPP type can be associated with a particular
- *          OpaqueType registration
- *
- */
-template <typename T, const char D[], const char N[]>
-class OpaqueType : public NonTensorType<T> {
- public:
-  static MLDataType Type();
-
-  bool IsCompatible(const ONNX_NAMESPACE::TypeProto& type_proto) const override {
-    return this->IsOpaqueCompatible(type_proto);
-  }
-
-  void FromDataContainer(const void* data, size_t data_size, OrtValue& output) const override {
-    NonTensorTypeConverter<T>::FromContainer(this, data, data_size, output);
-  }
-
-  void ToDataContainer(const OrtValue& input, size_t data_size, void* data) const override {
-    NonTensorTypeConverter<T>::ToContainer(input, data_size, data);
-  }
-
- private:
-  OpaqueType() {
-    data_types_internal::AssignOpaqueDomainName(D, N, this->MutableTypeProto());
-  }
-};
-
-/**
- * \brief PrimitiveDataTypeBase
- *        Base class for primitive Tensor contained types
- *
- * \details This class contains an integer constant that can be
- *          used for input data type dispatching
- *
- */
-class PrimitiveDataTypeBase : public DataTypeImpl {
- public:
-  bool IsCompatible(const ONNX_NAMESPACE::TypeProto&) const override {
-    return false;
-  }
-
-  const ONNX_NAMESPACE::TypeProto* GetTypeProto() const final {
-    return nullptr;
-  }
-
-  int32_t GetDataType() const {
-    return data_type_;
-  }
-
- protected:
-  PrimitiveDataTypeBase(size_t size, int32_t data_type)
-      : DataTypeImpl{GeneralType::kPrimitive, size}, data_type_{data_type} {}
-
- private:
-  const int32_t data_type_;
-};
-
-/**
- * \brief PrimitiveDataType
- *        Typed specialization for primitive types.
- *        Concrete instances of this class are used by Tensor.
- *
- * \param T - primitive data type
- *
- */
-template <typename T>
-class PrimitiveDataType : public PrimitiveDataTypeBase {
- private:
-  static void Delete(void* p) {
-    delete static_cast<T*>(p);
-  }
-
- public:
-  static MLDataType Type();
-
-  DeleteFunc GetDeleteFunc() const override {
-    return &Delete;
-  }
-
- private:
-  PrimitiveDataType()
-      : PrimitiveDataTypeBase{sizeof(T),
-                              utils::ToTensorProtoElementType<T>()} {
-  }
-};
-
-inline const TensorTypeBase* DataTypeImpl::AsTensorType() const {
-  return IsTensorType() ? static_cast<const TensorTypeBase*>(this) : nullptr;
-}
-
-inline const SequenceTensorTypeBase* DataTypeImpl::AsSequenceTensorType() const {
-  return IsTensorSequenceType() ? static_cast<const SequenceTensorTypeBase*>(this) : nullptr;
-}
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-inline const SparseTensorTypeBase* DataTypeImpl::AsSparseTensorType() const {
-  return IsSparseTensorType() ? static_cast<const SparseTensorTypeBase*>(this) : nullptr;
-}
-#endif
-
-#if !defined(DISABLE_OPTIONAL_TYPE)
-inline const OptionalTypeBase* DataTypeImpl::AsOptionalType() const {
-  return IsOptionalType() ? static_cast<const OptionalTypeBase*>(this) : nullptr;
-}
-#endif
-
-inline const NonTensorTypeBase* DataTypeImpl::AsNonTensorType() const {
-  return IsNonTensorType() ? static_cast<const NonTensorTypeBase*>(this) : nullptr;
-}
-
-inline const PrimitiveDataTypeBase* DataTypeImpl::AsPrimitiveDataType() const {
-  return IsPrimitiveDataType() ? static_cast<const PrimitiveDataTypeBase*>(this) : nullptr;
-}
-
-// Explicit specialization of base class template function
-// is only possible within the enclosing namespace scope,
-// thus a simple way to pre-instantiate a given template
-// at a registration time does not currently work and the macro
-// is needed.
-#define ORT_REGISTER_TENSOR_TYPE(ELEM_TYPE)             \
-  template <>                                           \
-  MLDataType TensorType<ELEM_TYPE>::Type() {            \
-    static TensorType<ELEM_TYPE> tensor_type;           \
-    return &tensor_type;                                \
-  }                                                     \
-  template <>                                           \
-  MLDataType DataTypeImpl::GetTensorType<ELEM_TYPE>() { \
-    return TensorType<ELEM_TYPE>::Type();               \
-  }
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-#define ORT_REGISTER_SPARSE_TENSOR_TYPE(ELEM_TYPE)            \
-  template <>                                                 \
-  MLDataType SparseTensorType<ELEM_TYPE>::Type() {            \
-    static SparseTensorType<ELEM_TYPE> tensor_type;           \
-    return &tensor_type;                                      \
-  }                                                           \
-  template <>                                                 \
-  MLDataType DataTypeImpl::GetSparseTensorType<ELEM_TYPE>() { \
-    return SparseTensorType<ELEM_TYPE>::Type();               \
-  }
-#endif
-
-#define ORT_REGISTER_OPTIONAL_TYPE(ORT_TYPE, TYPE)             \
-  template <>                                                  \
-  MLDataType OptionalType<ORT_TYPE, TYPE>::Type() {            \
-    static OptionalType<ORT_TYPE, TYPE> optional_type;         \
-    return &optional_type;                                     \
-  }                                                            \
-  template <>                                                  \
-  MLDataType DataTypeImpl::GetOptionalType<ORT_TYPE, TYPE>() { \
-    return OptionalType<ORT_TYPE, TYPE>::Type();               \
-  }
-
-#if !defined(DISABLE_ML_OPS)
-#define ORT_REGISTER_MAP(TYPE)               \
-  template <>                                \
-  MLDataType MapType<TYPE>::Type() {         \
-    static MapType<TYPE> map_type;           \
-    return &map_type;                        \
-  }                                          \
-  template <>                                \
-  MLDataType DataTypeImpl::GetType<TYPE>() { \
-    return MapType<TYPE>::Type();            \
-  }
-#endif
-
-#define ORT_REGISTER_SEQ(TYPE)               \
-  template <>                                \
-  MLDataType SequenceType<TYPE>::Type() {    \
-    static SequenceType<TYPE> sequence_type; \
-    return &sequence_type;                   \
-  }                                          \
-  template <>                                \
-  MLDataType DataTypeImpl::GetType<TYPE>() { \
-    return SequenceType<TYPE>::Type();       \
-  }
-
-#define ORT_REGISTER_SEQ_TENSOR_TYPE(ELEM_TYPE)                 \
-  template <>                                                   \
-  MLDataType SequenceTensorType<ELEM_TYPE>::Type() {            \
-    static SequenceTensorType<ELEM_TYPE> sequence_tensor_type;  \
-    return &sequence_tensor_type;                               \
-  }                                                             \
-  template <>                                                   \
-  MLDataType DataTypeImpl::GetSequenceTensorType<ELEM_TYPE>() { \
-    return SequenceTensorType<ELEM_TYPE>::Type();               \
-  }
-
-#define ORT_REGISTER_PRIM_TYPE(TYPE)               \
-  template <>                                      \
-  MLDataType PrimitiveDataType<TYPE>::Type() {     \
-    static PrimitiveDataType<TYPE> prim_data_type; \
-    return &prim_data_type;                        \
-  }                                                \
-  template <>                                      \
-  MLDataType DataTypeImpl::GetType<TYPE>() {       \
-    return PrimitiveDataType<TYPE>::Type();        \
-  }
-
-#define ORT_REGISTER_OPAQUE_TYPE(CPPType, Domain, Name)   \
-  template <>                                             \
-  MLDataType OpaqueType<CPPType, Domain, Name>::Type() {  \
-    static OpaqueType<CPPType, Domain, Name> opaque_type; \
-    return &opaque_type;                                  \
-  }                                                       \
-  template <>                                             \
-  MLDataType DataTypeImpl::GetType<CPPType>() {           \
-    return OpaqueType<CPPType, Domain, Name>::Type();     \
-  }
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/data_types_internal.h b/projects/llm_framework/include/onnxruntime/core/framework/data_types_internal.h
deleted file mode 100644
index 9d791a7f..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/data_types_internal.h
+++ /dev/null
@@ -1,569 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <array>
-#include <cassert>
-#include <cstdint>
-#include <string>
-#include <type_traits>
-#include <vector>
-
-#include "boost/mp11.hpp"
-
-#include "core/common/common.h"
-#include "core/framework/to_tensor_proto_element_type.h"
-#ifndef SHARED_PROVIDER
-#include "core/common/type_list.h"
-#include "core/framework/data_types.h"
-#if !defined(ORT_MINIMAL_BUILD)
-#include "onnx/defs/schema.h"
-#else
-#include "onnx/defs/data_type_utils.h"
-#endif
-#include "onnx/onnx_pb.h"
-#include "onnx/onnx-operators_pb.h"
-#endif
-
-namespace onnxruntime {
-namespace utils {
-
-  // The following primitives are strongly recommended for switching on tensor input datatypes for
-  // kernel implementations.
-  //
-  //  1) If you need to handle all of the primitive tensor contained datatypes, the best choice would be macros
-  //     DispatchOnTensorType or DispatchOnTensorTypeWithReturn. Use inline wrappers so your function can be invoked as function<T>().
-  //  2) if you have a few types, use Tensor.IsDataType<T>()/IsDataTypeString() or use utils::IsPrimitiveDataType<T>()
-  //     if you have a standalone MLDatatType with a sequence of if/else statements.
-  //  3) For something in between, we suggest to use CallDispatcher pattern.
-  //
-  // Invoking DataTypeImpl::GetType<T>() for switching on input types is discouraged and should be avoided.
-  // Every primitive type carries with it an integer constant that can be used for quick switching on types.
-
-#define DispatchOnTensorType(tensor_type, function, ...)          \
-  switch (tensor_type->AsPrimitiveDataType()->GetDataType()) {    \
-    case ONNX_NAMESPACE::TensorProto_DataType_FLOAT:              \
-      function<float>(__VA_ARGS__);                               \
-      break;                                                      \
-    case ONNX_NAMESPACE::TensorProto_DataType_BOOL:               \
-      function<bool>(__VA_ARGS__);                                \
-      break;                                                      \
-    case ONNX_NAMESPACE::TensorProto_DataType_DOUBLE:             \
-      function<double>(__VA_ARGS__);                              \
-      break;                                                      \
-    case ONNX_NAMESPACE::TensorProto_DataType_STRING:             \
-      function<std::string>(__VA_ARGS__);                         \
-      break;                                                      \
-    case ONNX_NAMESPACE::TensorProto_DataType_INT8:               \
-      function<int8_t>(__VA_ARGS__);                              \
-      break;                                                      \
-    case ONNX_NAMESPACE::TensorProto_DataType_UINT8:              \
-      function<uint8_t>(__VA_ARGS__);                             \
-      break;                                                      \
-    case ONNX_NAMESPACE::TensorProto_DataType_INT16:              \
-      function<int16_t>(__VA_ARGS__);                             \
-      break;                                                      \
-    case ONNX_NAMESPACE::TensorProto_DataType_UINT16:             \
-      function<uint16_t>(__VA_ARGS__);                            \
-      break;                                                      \
-    case ONNX_NAMESPACE::TensorProto_DataType_INT32:              \
-      function<int32_t>(__VA_ARGS__);                             \
-      break;                                                      \
-    case ONNX_NAMESPACE::TensorProto_DataType_UINT32:             \
-      function<uint32_t>(__VA_ARGS__);                            \
-      break;                                                      \
-    case ONNX_NAMESPACE::TensorProto_DataType_INT64:              \
-      function<int64_t>(__VA_ARGS__);                             \
-      break;                                                      \
-    case ONNX_NAMESPACE::TensorProto_DataType_UINT64:             \
-      function<uint64_t>(__VA_ARGS__);                            \
-      break;                                                      \
-    case ONNX_NAMESPACE::TensorProto_DataType_FLOAT16:            \
-      function<MLFloat16>(__VA_ARGS__);                           \
-      break;                                                      \
-    case ONNX_NAMESPACE::TensorProto_DataType_BFLOAT16:           \
-      function<BFloat16>(__VA_ARGS__);                            \
-      break;                                                      \
-    default:                                                      \
-      ORT_ENFORCE(false, "Unknown tensor type of ", tensor_type); \
-  }
-
-#define DispatchOnTensorTypeWithReturn(tensor_type, retval, function, ...) \
-  switch (tensor_type->AsPrimitiveDataType()->GetDataType()) {             \
-    case ONNX_NAMESPACE::TensorProto_DataType_FLOAT:                       \
-      retval = function<float>(__VA_ARGS__);                               \
-      break;                                                               \
-    case ONNX_NAMESPACE::TensorProto_DataType_BOOL:                        \
-      retval = function<bool>(__VA_ARGS__);                                \
-      break;                                                               \
-    case ONNX_NAMESPACE::TensorProto_DataType_DOUBLE:                      \
-      retval = function<double>(__VA_ARGS__);                              \
-      break;                                                               \
-    case ONNX_NAMESPACE::TensorProto_DataType_STRING:                      \
-      retval = function<std::string>(__VA_ARGS__);                         \
-      break;                                                               \
-    case ONNX_NAMESPACE::TensorProto_DataType_INT8:                        \
-      retval = function<int8_t>(__VA_ARGS__);                              \
-      break;                                                               \
-    case ONNX_NAMESPACE::TensorProto_DataType_UINT8:                       \
-      retval = function<uint8_t>(__VA_ARGS__);                             \
-      break;                                                               \
-    case ONNX_NAMESPACE::TensorProto_DataType_UINT16:                      \
-      retval = function<uint16_t>(__VA_ARGS__);                            \
-      break;                                                               \
-    case ONNX_NAMESPACE::TensorProto_DataType_INT16:                       \
-      retval = function<int16_t>(__VA_ARGS__);                             \
-      break;                                                               \
-    case ONNX_NAMESPACE::TensorProto_DataType_INT32:                       \
-      retval = function<int32_t>(__VA_ARGS__);                             \
-      break;                                                               \
-    case ONNX_NAMESPACE::TensorProto_DataType_UINT32:                      \
-      retval = function<uint32_t>(__VA_ARGS__);                            \
-      break;                                                               \
-    case ONNX_NAMESPACE::TensorProto_DataType_INT64:                       \
-      retval = function<int64_t>(__VA_ARGS__);                             \
-      break;                                                               \
-    case ONNX_NAMESPACE::TensorProto_DataType_UINT64:                      \
-      retval = function<uint64_t>(__VA_ARGS__);                            \
-      break;                                                               \
-    case ONNX_NAMESPACE::TensorProto_DataType_FLOAT16:                     \
-      retval = function<MLFloat16>(__VA_ARGS__);                           \
-      break;                                                               \
-    case ONNX_NAMESPACE::TensorProto_DataType_BFLOAT16:                    \
-      retval = function<BFloat16>(__VA_ARGS__);                            \
-      break;                                                               \
-    default:                                                               \
-      ORT_ENFORCE(false, "Unknown tensor type of ", tensor_type);          \
-  }
-
-////////////////////////////////////////////////////////////////////////////////
-/// Use the following primitives if you have a few types to switch on so you
-//  can write a short sequence of if/else statements.
-
-// This is a frequently used check so we make a separate utility function.
-inline bool IsDataTypeString(MLDataType dt_type) {
-  auto prim_type = dt_type->AsPrimitiveDataType();
-  return (prim_type != nullptr && prim_type->GetDataType() == ONNX_NAMESPACE::TensorProto_DataType_STRING);
-}
-
-// Test if MLDataType is a concrete type of PrimitiveDataTypeBase
-// and it is T
-template <class T>
-inline bool IsPrimitiveDataType(MLDataType dt_type) {
-  auto prim_type = dt_type->AsPrimitiveDataType();
-  return (prim_type != nullptr && prim_type->GetDataType() == ToTensorProtoElementType<T>());
-}
-
-// Use after AsPrimitiveDataType() is successful
-// Check if PrimitiveDataTypeBase is of type T
-template <class T>
-inline bool IsPrimitiveDataType(const PrimitiveDataTypeBase* prim_type) {
-  assert(prim_type != nullptr);
-  return prim_type->GetDataType() == ToTensorProtoElementType<T>();
-}
-
-// This implementation contains a workaround for GCC bug https://gcc.gnu.org/bugzilla/show_bug.cgi?id=47226
-// GCC until very recently does not support template parameter pack expansion within lambda context.
-namespace mltype_dispatcher_internal {
-
-// T - type handled by this helper
-class CallableDispatchableHelper {
-  int32_t dt_type_;  // Type currently dispatched
-  size_t called_;
-
- public:
-  explicit CallableDispatchableHelper(int32_t dt_type) noexcept : dt_type_(dt_type), called_(0) {}
-
-  // Must return integer to be in a expandable context
-  template <class T, class Fn, class... Args>
-  int Invoke(Fn&& fn, Args&&... args) {
-    if (utils::ToTensorProtoElementType<T>() == dt_type_) {
-      std::forward<Fn>(fn)(std::forward<Args>(args)...);
-      ++called_;
-    }
-    return 0;
-  }
-
-  void CheckCalledOnce() {
-    ORT_ENFORCE(called_ == 1, "Unsupported data type: ", dt_type_);
-  }
-};
-
-// Default policy is to throw an exception.
-// Other policies may set the second result argument accordingly.
-template <class Ret>
-struct UnsupportedTypeDefaultPolicy {
-  void operator()(int32_t dt_type, Ret& /*result*/) const {
-    ORT_THROW("Unsupported data type: ", dt_type);
-  }
-};
-
-// Helper with the result type
-template <class Ret, class UnsupportedPolicy>
-class CallableDispatchableRetHelper {
-  int32_t dt_type_;  // Type currently dispatched
-  size_t called_;
-  Ret result_;
-
- public:
-  explicit CallableDispatchableRetHelper(int32_t dt_type) noexcept : dt_type_(dt_type), called_(0), result_() {}
-
-  Ret Get() {
-    // No type was invoked
-    if (called_ == 0) {
-      UnsupportedPolicy()(dt_type_, result_);
-    }
-    return result_;
-  }
-
-  // Must return integer to be in a expandable context
-  template <class T, class Fn, class... Args>
-  int Invoke(Fn&& fn, Args&&... args) {
-    if (utils::ToTensorProtoElementType<T>() == dt_type_) {
-      result_ = std::forward<Fn>(fn)(std::forward<Args>(args)...);
-      ++called_;
-    }
-    return 0;
-  }
-};
-
-template <typename T>
-using TensorProtoElementTypeConstant =
-    std::integral_constant<ONNX_NAMESPACE::TensorProto_DataType, ToTensorProtoElementType<T>()>;
-
-using UndefinedTensorProtoElementTypeConstant =
-    std::integral_constant<ONNX_NAMESPACE::TensorProto_DataType, ONNX_NAMESPACE::TensorProto_DataType_UNDEFINED>;
-
-}  // namespace mltype_dispatcher_internal
-
-/**
- * This class helps to efficiently dispatch calls to implementation function
- * objects with a tensor element type template argument.
- *
- * The constructor accepts a value corresponding to a tensor element type.
- * For example, it can be obtained from:
- *   input_tensor->GetElementType()
- *
- * The Invoke member functions will instantiate and invoke the provided
- * function object template, Fn. Fn must be default constructible. Fn must also
- * have a tensor element type template argument. This type template argument
- * will be the type that corresponds to the value given in the constructor.
- * These functions accept and forward arbitrary function arguments. They ensure
- * that Fn is called once with the type specified in the constructor.
- *
- * @tparam Types The types supported by the implementation. This should be a
- *         set of ONNX tensor element types that are supported by ORT.
- */
-template <typename... Types>
-class MLTypeCallDispatcher {
-  using SupportedTypeList = TypeList<Types...>;
-  using SupportedTensorProtoElementTypeList =
-      boost::mp11::mp_transform<
-          mltype_dispatcher_internal::TensorProtoElementTypeConstant, SupportedTypeList>;
-
-  static_assert(
-      boost::mp11::mp_and<
-          boost::mp11::mp_is_set<SupportedTensorProtoElementTypeList>,
-          boost::mp11::mp_not<
-              boost::mp11::mp_set_contains<
-                  SupportedTensorProtoElementTypeList,
-                  mltype_dispatcher_internal::UndefinedTensorProtoElementTypeConstant>>>::value,
-      "Types must map to a unique set of ONNX tensor element data types supported by ORT.");
-
-  int32_t dt_type_;
-
- public:
-  /**
-   * Constructor.
-   * @param dt_type The value corresponding to the tensor element type to be
-   *        dispatched to. This can be obtained from
-   *        input_tensor->GetElementType() or
-   *        utils::ToTensorProtoElementType<T>().
-   */
-  explicit MLTypeCallDispatcher(int32_t dt_type) noexcept : dt_type_(dt_type) {}
-
-  /**
-   * Invokes Fn<T> with the specified arguments.
-   *
-   * @tparam Fn The function object template.
-   * @tparam Args The argument types.
-   */
-  template <template <typename...> class Fn, typename... Args>
-  void Invoke(Args&&... args) const {
-    InvokeWithLeadingTemplateArgs<Fn, TypeList<>>(std::forward<Args>(args)...);
-  }
-
-  /**
-   * Invokes Fn<..., T> with leading template arguments and the specified
-   * arguments.
-   *
-   * @tparam Fn The function object template.
-   * @tparam LeadingTemplateArgTypeList A type list of the leading template
-   *         arguments.
-   * @tparam Args The argument types.
-   */
-  template <template <typename...> class Fn, typename LeadingTemplateArgTypeList, typename... Args>
-  void InvokeWithLeadingTemplateArgs(Args&&... args) const {
-    static_assert(
-        boost::mp11::mp_is_list<LeadingTemplateArgTypeList>::value,
-        "LeadingTemplateArgTypeList must be a type list (e.g., onnxruntime::TypeList<T1, T2, ...>).");
-
-    mltype_dispatcher_internal::CallableDispatchableHelper helper(dt_type_);
-
-    // given LeadingTemplateArgTypeList is a type list L<U1, U2, ...>,
-    //   call helper.Invoke() with Fn<U1, U2, ..., T> for each T in Types
-    static_cast<void>(std::array<int, sizeof...(Types)>{
-        helper.template Invoke<Types>(
-            boost::mp11::mp_apply<Fn, boost::mp11::mp_push_back<LeadingTemplateArgTypeList, Types>>(),
-            std::forward<Args>(args)...)...});
-
-    // avoid "unused parameter" warning for the case where Types is empty
-    static_cast<void>(std::array<int, sizeof...(Args)>{(ORT_UNUSED_PARAMETER(args), 0)...});
-
-    helper.CheckCalledOnce();
-  }
-
-  /**
-   * Invokes Fn<T> with the specified arguments and returns the result.
-   *
-   * @tparam Ret The return type. Fn should return a type convertible to Ret.
-   * @tparam Fn The function object template.
-   * @tparam Args The argument types.
-   */
-  template <class Ret, template <typename...> class Fn, typename... Args>
-  Ret InvokeRet(Args&&... args) const {
-    return InvokeRetWithUnsupportedPolicy<
-        Ret, Fn, mltype_dispatcher_internal::UnsupportedTypeDefaultPolicy<Ret>>(
-        std::forward<Args>(args)...);
-  }
-
-  /**
-   * Invokes Fn<T> with the specified arguments and returns the result.
-   *
-   * @tparam Ret The return type. Fn should return a type convertible to Ret.
-   * @tparam Fn The function object template.
-   * @tparam UnsupportedPolicy The policy used to handle unsupported types.
-   *         See mltype_dispatcher_internal::UnsupportedTypeDefaultPolicy
-   *         for an example.
-   * @tparam Args The argument types.
-   */
-  template <class Ret, template <typename...> class Fn, class UnsupportedPolicy, typename... Args>
-  Ret InvokeRetWithUnsupportedPolicy(Args&&... args) const {
-    return InvokeRetWithUnsupportedPolicyAndLeadingTemplateArgs<
-        Ret, Fn, UnsupportedPolicy, TypeList<>>(
-        std::forward<Args>(args)...);
-  }
-
-  /**
-   * Invokes Fn<..., T> with leading template arguments and the specified
-   * arguments and returns the result.
-   *
-   * @tparam Ret The return type. Fn should return a type convertible to Ret.
-   * @tparam Fn The function object template.
-   * @tparam LeadingTemplateArgTypeList A type list of the leading template
-   *         arguments.
-   * @tparam Args The argument types.
-   */
-  template <class Ret, template <typename...> class Fn, typename LeadingTemplateArgTypeList, typename... Args>
-  Ret InvokeRetWithLeadingTemplateArgs(Args&&... args) const {
-    return InvokeRetWithUnsupportedPolicyAndLeadingTemplateArgs<
-        Ret, Fn, mltype_dispatcher_internal::UnsupportedTypeDefaultPolicy<Ret>, LeadingTemplateArgTypeList>(
-        std::forward<Args>(args)...);
-  }
-
-  /**
-   * Invokes Fn<..., T> with leading template arguments and the specified
-   * arguments and returns the result.
-   *
-   * @tparam Ret The return type. Fn should return a type convertible to Ret.
-   * @tparam Fn The function object template.
-   * @tparam UnsupportedPolicy The policy used to handle unsupported types.
-   *         See mltype_dispatcher_internal::UnsupportedTypeDefaultPolicy
-   *         for an example.
-   * @tparam LeadingTemplateArgTypeList A type list of the leading template
-   *         arguments.
-   * @tparam Args The argument types.
-   */
-  template <class Ret,
-            template <typename...> class Fn,
-            class UnsupportedPolicy,
-            typename LeadingTemplateArgTypeList,
-            typename... Args>
-  Ret InvokeRetWithUnsupportedPolicyAndLeadingTemplateArgs(Args&&... args) const {
-    mltype_dispatcher_internal::CallableDispatchableRetHelper<Ret, UnsupportedPolicy> helper(dt_type_);
-
-    // given LeadingTemplateArgTypeList is a type list L<U1, U2, ...>,
-    //   call helper.Invoke() with Fn<U1, U2, ..., T> for each T in Types
-    static_cast<void>(std::array<int, sizeof...(Types)>{
-        helper.template Invoke<Types>(
-            boost::mp11::mp_apply<Fn, boost::mp11::mp_push_back<LeadingTemplateArgTypeList, Types>>(),
-            std::forward<Args>(args)...)...});
-
-    // avoid "unused parameter" warning for the case where Types is empty
-    static_cast<void>(std::array<int, sizeof...(Args)>{(ORT_UNUSED_PARAMETER(args), 0)...});
-
-    return helper.Get();
-  }
-};
-
-// the type MLTypeCallDispatcher<T...> given a type list L<T...>
-template <typename L>
-using MLTypeCallDispatcherFromTypeList = boost::mp11::mp_apply<MLTypeCallDispatcher, L>;
-
-namespace data_types_internal {
-
-enum class ContainerType : uint16_t {
-  kUndefined = 0,
-  kTensor = 1,
-  kMap = 2,
-  kSequence = 3,
-  kOpaque = 4,
-  kOptional = 5
-};
-
-class TypeNode {
-  // type_ is a TypeProto value case enum
-  // that may be a kTypeTensor, kTypeMap, kTypeSequence
-  // prim_type_ is a TypeProto_DataType enum that has meaning
-  // - for Tensor then prim_type_ is the contained type
-  // - for Map prim_type is the key type. Next entry describes map value
-  // - For sequence prim_type_ is not used and has no meaning. Next entry
-  //   describes the value for the sequence
-  // Tensor is always the last entry as it describes a contained primitive type.
-  ContainerType type_;
-  uint16_t prim_type_;
-
- public:
-  TypeNode(ContainerType type, int32_t prim_type) noexcept {
-    type_ = type;
-    prim_type_ = static_cast<uint16_t>(prim_type);
-  }
-
-  bool IsType(ContainerType type) const noexcept {
-    return type_ == type;
-  }
-
-  bool IsPrimType(int32_t prim_type) const noexcept {
-    return prim_type_ == static_cast<uint16_t>(prim_type);
-  }
-};
-
-}  // namespace data_types_internal
-
-////////////////////////////////////////////////////////////////////
-/// Provides generic interface to test whether MLDataType is a Sequence,
-/// Map or an Opaque type including arbitrary recursive definitions
-/// without querying DataTypeImpl::GetType<T> for all known complex types
-
-// T is a sequence contained element type
-// If returns true then we know that the runtime
-// representation is std::vector<T>
-// T itself can be a runtime representation of another
-// sequence, map, opaque type or a tensor
-//
-// That is it can be std::vector or a std::map
-// If T is a primitive type sequence is tested whether it contains
-// tensors of that type
-//
-// If T is an opaque type, then it is only tested to be opaque but not exactly
-// a specific opaque type. To Test for a specific Opaque type use IsOpaqueType() below
-//
-// This class examines the supplied MLDataType and records
-// its information in a vector so any subsequent checks for Sequences and Maps
-// are quick.
-class ContainerChecker {
-  using Cont = std::vector<data_types_internal::TypeNode>;
-  Cont types_;
-
-  // Default IsContainerOfType is for Opaque type
-  template <class T>
-  struct IsContainerOfType {
-    static bool check(const Cont& c, size_t index) {
-      if (index >= c.size()) {
-        return false;
-      }
-      return c[index].IsType(data_types_internal::ContainerType::kOpaque);
-    }
-  };
-
-  // Handles the case where sequence element is also a sequence
-  template <class T>
-  struct IsContainerOfType<std::vector<T>> {
-    static bool check(const Cont& c, size_t index) {
-      if (index >= c.size()) {
-        return false;
-      }
-      if (c[index].IsType(data_types_internal::ContainerType::kSequence)) {
-        ORT_ENFORCE(++index < c.size(), "Sequence is missing type entry for its element");
-        constexpr int32_t prim_type = ToTensorProtoElementType<T>();
-        // Check if this is a primitive type and it matches
-        if constexpr(prim_type != ONNX_NAMESPACE::TensorProto_DataType_UNDEFINED) {
-          return c[index].IsType(data_types_internal::ContainerType::kTensor) &&
-                 c[index].IsPrimType(prim_type);
-        }
-        else {
-          // T is not primitive, check next entry for non-primitive proto
-          return IsContainerOfType<T>::check(c, index);
-        }
-      }
-      return false;
-    }
-  };
-
-  template <class K, class V>
-  struct IsContainerOfType<std::map<K, V>> {
-    static bool check(const Cont& c, size_t index) {
-      static_assert(ToTensorProtoElementType<K>() != ONNX_NAMESPACE::TensorProto_DataType_UNDEFINED,
-                    "Map Key can not be a non-primitive type");
-      if (index >= c.size()) {
-        return false;
-      }
-      if (!c[index].IsType(data_types_internal::ContainerType::kMap)) {
-        return false;
-      }
-      constexpr int32_t key_type = ToTensorProtoElementType<K>();
-      if (!c[index].IsPrimType(key_type)) {
-        return false;
-      }
-      ORT_ENFORCE(++index < c.size(), "Map is missing type entry for its value");
-      constexpr int32_t val_type = ToTensorProtoElementType<V>();
-      if constexpr(val_type != ONNX_NAMESPACE::TensorProto_DataType_UNDEFINED) {
-        return c[index].IsType(data_types_internal::ContainerType::kTensor) &&
-               c[index].IsPrimType(val_type);
-      }
-      else return IsContainerOfType<V>::check(c, index);
-    }
-  };
-
- public:
-  explicit ContainerChecker(MLDataType);
-  ~ContainerChecker() = default;
-
-  bool IsMap() const noexcept {
-    assert(!types_.empty());
-    return types_[0].IsType(data_types_internal::ContainerType::kMap);
-  }
-
-  bool IsSequence() const noexcept {
-    assert(!types_.empty());
-    return types_[0].IsType(data_types_internal::ContainerType::kSequence);
-  }
-
-  template <class T>
-  bool IsSequenceOf() const {
-    assert(!types_.empty());
-    return IsContainerOfType<std::vector<T>>::check(types_, 0);
-  }
-
-  template <class K, class V>
-  bool IsMapOf() const {
-    assert(!types_.empty());
-    return IsContainerOfType<std::map<K, V>>::check(types_, 0);
-  }
-};
-
-bool IsOpaqueType(MLDataType ml_type, const char* domain, const char* name);
-
-}  // namespace utils
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/endian.h b/projects/llm_framework/include/onnxruntime/core/framework/endian.h
deleted file mode 100644
index 629fb78f..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/endian.h
+++ /dev/null
@@ -1,27 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-namespace onnxruntime {
-
-// the semantics of this enum should match std::endian from C++20
-enum class endian {
-#if defined(_WIN32)
-  little = 0,
-  big = 1,
-  native = little,
-#elif defined(__GNUC__) || defined(__clang__)
-  little = __ORDER_LITTLE_ENDIAN__,
-  big = __ORDER_BIG_ENDIAN__,
-  native = __BYTE_ORDER__,
-#else
-#error onnxruntime::endian is not implemented in this environment.
-#endif
-};
-
-static_assert(
-    endian::native == endian::little || endian::native == endian::big,
-    "Only little-endian or big-endian native byte orders are supported.");
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/execution_provider.h b/projects/llm_framework/include/onnxruntime/core/framework/execution_provider.h
deleted file mode 100644
index 09146ccd..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/execution_provider.h
+++ /dev/null
@@ -1,332 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#ifndef SHARED_PROVIDER
-#include <memory>
-#include <unordered_map>
-#include <unordered_set>
-
-#include "core/common/logging/logging.h"
-#include "core/common/status.h"
-#include "core/framework/data_transfer.h"
-#include "core/framework/tensor.h"
-
-namespace onnxruntime {
-class GraphViewer;
-struct ComputeCapability;
-class KernelRegistry;
-struct KernelCreateInfo;
-class Node;
-}  // namespace onnxruntime
-#else
-#include <memory>
-#endif
-
-#include "core/common/basic_types.h"
-#include "core/common/profiler_common.h"
-#include "core/framework/allocatormgr.h"
-#include "core/framework/func_api.h"
-#include "core/framework/provider_options.h"
-#include "core/framework/stream_handles.h"
-
-namespace onnxruntime {
-
-/**
-   Logical device representation.
-*/
-
-// if we are export the fused function to dll, the function will still in the same binary as onnxruntime
-// use std function to give execution provider some chance to capture some state.
-using CreateFunctionStateFunc = std::function<int(ComputeContext*, FunctionState*)>;
-using ComputeFunc = std::function<Status(FunctionState, const OrtApi*, OrtKernelContext*)>;
-using DestroyFunctionStateFunc = std::function<void(FunctionState)>;
-
-struct NodeComputeInfo {
-  CreateFunctionStateFunc create_state_func;
-  ComputeFunc compute_func;
-  DestroyFunctionStateFunc release_state_func;
-};
-
-enum class DataLayout {
-  NCHW,
-  NHWC,
-  NCHWC,
-};
-
-class IExecutionProvider {
- protected:
-  IExecutionProvider(const std::string& type, bool use_metadef_id_creator = false)
-      : type_{type} {
-    if (use_metadef_id_creator) {
-      metadef_id_generator_ = std::make_unique<ModelMetadefIdGenerator>();
-    }
-  }
-
- public:
-  virtual ~IExecutionProvider() = default;
-
-  /**
-     Get all IAllocators for <*this> execution provider.
-  */
-  const std::vector<AllocatorPtr>& GetAllocators() const {
-    return allocator_list_;
-  }
-
-  /**
-   * Get an allocator with specified device id and MemType. Return nullptr if it doesn't exist
-   */
-  virtual AllocatorPtr GetAllocator(int device_id, OrtMemType mem_type) const;
-
-  /**
-   * Returns a data transfer object that implements methods to copy to and
-   * from this device.
-   * If no copy is required for the successful operation of this provider,
-   * return a nullptr.
-   */
-  virtual std::unique_ptr<onnxruntime::IDataTransfer> GetDataTransfer() const {
-    return nullptr;
-  }
-
-  /**
-   * Interface for performing kernel lookup within kernel registries.
-   * Abstracts away lower-level details about kernel registries and kernel matching.
-   */
-  class IKernelLookup {
-   public:
-    /**
-     * Given `node`, try to find a matching kernel for this EP.
-     * The return value is non-null if and only if a matching kernel was found.
-     */
-    virtual const KernelCreateInfo* LookUpKernel(const Node& node) const = 0;
-
-   protected:
-    ~IKernelLookup() = default;
-  };
-
-  /**
-     Get execution provider's capability for the specified <graph>.
-     Return a bunch of IndexedSubGraphs <*this> execution provider can run if
-     the sub-graph contains only one node or can fuse to run if the sub-graph
-     contains more than one node. The node indexes contained in sub-graphs may
-     have overlap, and it's ONNXRuntime's responsibility to do the partition
-     and decide whether a node will be assigned to <*this> execution provider.
-     For kernels registered in a kernel registry, `kernel_lookup` must be used
-     to find a matching kernel for this EP.
-  */
-  virtual std::vector<std::unique_ptr<ComputeCapability>>
-  GetCapability(const onnxruntime::GraphViewer& graph_viewer,
-                const IKernelLookup& kernel_lookup) const;
-
-  /**
-     Get kernel registry per execution provider type.
-     The KernelRegistry share pointer returned is shared across sessions.
-
-     NOTE: this approach was taken to achieve the following goals,
-     1. The execution provider type based kernel registry should be shared
-     across sessions.
-     Only one copy of this kind of kernel registry exists in ONNXRuntime
-     with multiple sessions/models.
-     2. Adding an execution provider into ONNXRuntime does not need to touch ONNXRuntime
-     framework/session code.
-     3. onnxruntime (framework/session) does not depend on any specific
-     execution provider lib.
-  */
-  virtual std::shared_ptr<KernelRegistry> GetKernelRegistry() const { return nullptr; }
-
-  /**
-     Get the device id of current execution provider
-  */
-  virtual int GetDeviceId() const { return 0; };
-
-  /**
-     Get execution provider's configuration options.
-   */
-  virtual ProviderOptions GetProviderOptions() const { return {}; }
-
-  /**
-     Returns an opaque handle whose exact type varies based on the provider
-     and is interpreted accordingly by the corresponding kernel implementation.
-     For Direct3D operator kernels, this may return an IUnknown supporting
-     QueryInterface to ID3D12GraphicsCommandList1.
-  */
-  virtual const void* GetExecutionHandle() const noexcept {
-    return nullptr;
-  }
-
-  /**
-     @return type of the execution provider; should match that set in the node
-     through the SetExecutionProvider API. Example valid return values are:
-     kCpuExecutionProvider, kCudaExecutionProvider
-  */
-  const std::string& Type() const { return type_; }
-
-  /**
-     Blocks until the device has completed all preceding requested tasks.
-     Currently this is primarily used by the IOBinding object to ensure that all
-     inputs have been copied to the device before execution begins.
-  */
-  virtual common::Status Sync() const { return Status::OK(); }
-
-  /**
-     Called when InferenceSession::Run started
-     NOTE that due to async execution in provider, the actual work of previous
-     Run may not be finished on device This function should be regarded as the
-     point after which a new Run would start to submit commands from CPU
-  */
-  virtual common::Status OnRunStart() { return Status::OK(); }
-
-  /**
-     Called when InferenceSession::Run ended
-     NOTE that due to async execution in provider, the actual work of this Run
-     may not be finished on device This function should be regarded as the point
-     that all commands of current Run has been submmited by CPU
-  */
-  virtual common::Status OnRunEnd(bool /*sync_stream*/) { return Status::OK(); }
-
-  /**
-     Indicate whether the graph capturing mode (e.g., cuda graph) is enabled for
-     the provider. Currently only CUDA execution provider supports it.
-   */
-  virtual bool IsGraphCaptureEnabled() const { return false; }
-
-  /**
-     Indicate whether the graph has been captured and instantiated. Currently
-     only CUDA execution provider supports it.
-   */
-  virtual bool IsGraphCaptured() const { return false; }
-
-  /**
-     Run the instantiated graph. Currently only CUDA execution provider supports
-     it.
-   */
-  virtual common::Status ReplayGraph() { return Status::OK(); }
-
-  /**
-     Called when session creation is complete
-     This provides an opportunity for execution providers to optionally synchronize and
-     clean up its temporary resources to reduce memory and ensure the first run is fast.
-  */
-  virtual common::Status OnSessionInitializationEnd() { return Status::OK(); }
-
-  void InsertAllocator(AllocatorPtr allocator);
-  void ReplaceAllocator(AllocatorPtr allocator);
-
-  struct FusedNodeAndGraph {
-    const std::reference_wrapper<onnxruntime::Node> fused_node;
-    // GraphViewer that filters the full graph to the nodes that are covered by 'node'
-    const std::reference_wrapper<GraphViewer> filtered_graph;
-  };
-
-  // Fusion approach that is suppported
-  // !!! The "Function" FusionStyle is deprecated.
-  // !!! If your EP is using this fusion style, please migrate it to "FilteredGraphViewer" style.
-  enum class FusionStyle {
-    // The node fusion will create an onnxruntime::Function based Node that contains a completely new Graph instance
-    // in the Node body. The original nodes and initializers are copied to the new Graph instance in Function::Body().
-    // A GraphProto can be produced from the Node body.
-    Function,
-
-    // The node fusion will create a new Node that defines the inputs and outputs using the IndexedSubGraph
-    // that GetCapability returned. The Node will not be onnxruntime::Function based so will have no Body().
-    // Instead a GraphViewer that filters the full Graph to the fused Nodes will be created.
-    // This is significantly cheaper as it doesn't incur the cost of creating a new Graph instance,
-    // and can be supported in a minimal build.
-    FilteredGraphViewer
-  };
-
-  virtual FusionStyle GetFusionStyle() const {
-    // All the ORT build in EP has migrate to FilteredGraphViewer style.
-    // For newer EPs, please avoid use Function style as it is deprecated.
-    return FusionStyle::FilteredGraphViewer;
-  }
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  /**
-  Given a collection of fused Nodes and the respective GraphViewer instance for the nodes that were fused,
-  return create_state/compute/release_state func for each node.
-  @remarks This is now the default interface when execution provider wants to compile nodes
-           for both minimal build and complete ort build.
-
-           Do NOT cache the GraphViewer in FusedNodeAndGraph.filtered_graph in any of the NodeComputeInfo functions
-           as it is only valid for the duration of the call to Compile.
-  */
-  virtual common::Status Compile(const std::vector<FusedNodeAndGraph>& fused_nodes_and_graphs,
-                                 std::vector<NodeComputeInfo>& node_compute_funcs);
-
-#endif
-
-  void SetLogger(const logging::Logger* logger) {
-    logger_ = logger;
-  }
-
-  const logging::Logger* GetLogger() const {
-    return logger_;
-  }
-
-  /** Generate a unique id that can be used in a MetaDef name. Values are unique for a model instance.
-   The model hash is also returned if you wish to include that in the MetaDef name to ensure uniqueness across models.
-   @param graph_viewer[in] Graph viewer that GetCapability was called with. Can be for the main graph or nested graph.
-   @param model_hash[out] Returns the hash for the main (i.e. top level) graph in the model.
-                          This is created using the model path if available,
-                          or the model input names and the output names from all nodes in the main graph.
-   @remarks e.g. the TensorRT Execution Provider is used in multiple sessions and the underlying infrastructure caches
-            compiled kernels, so the name must be unique and deterministic across models and sessions.
-            NOTE: Ideally this would be a protected method, but to work across the EP bridge it has to be public and
-                  virtual, and ModelMetadefIdGenerator but be defined in the header as well.
-   */
-  virtual int GenerateMetaDefId(const onnxruntime::GraphViewer& graph_viewer, HashValue& model_hash) const;
-
-  /**
-     Register allocators for EP, potentially re-using existing allocators for a device from allocator_manager.
-     If the EP implements this it should generally delay creating any allocators until this is called.
-  */
-  virtual void RegisterAllocator(AllocatorManager& /*allocator_manager*/);
-
-  virtual std::unique_ptr<profiling::EpProfiler> GetProfiler() {
-    return {};
-  }
-
-  virtual DataLayout GetPreferredLayout() const {
-    // NCHW is the default ONNX standard data layout. So default to it.
-    // EPs which prefer a different layout should override to return their preferred layout.
-    return DataLayout::NCHW;
-  }
-
-  virtual void RegisterStreamHandlers(IStreamCommandHandleRegistry& /*stream_handle_registry*/) const {}
-
-  /** Does the EP support concurrent calls to InferenceSession::Run to execute the model.
-   */
-  virtual bool ConcurrentRunSupported() const { return true; }
-
- private:
-  const std::string type_;
-
-  // allocator lookup is done by combining the device id and OrtMemType.
-  // there's also an implicit connection to the underlying OrtDevice involved that is dependent on the EP.
-  // e.g. for a CPU based EP, 'default' memory is a CPU device, and for a GPU based EP 'default' memory is a
-  // GPU device.
-  using AllocatorMap = std::unordered_map<int, AllocatorPtr>;
-  AllocatorMap allocators_;
-
-  // It will be set when this object is registered to a session
-  const logging::Logger* logger_ = nullptr;
-  // convenience list of the allocators so GetAllocatorList doesn't have to build a new vector each time
-  // contains the same instances as allocators_
-  std::vector<AllocatorPtr> allocator_list_;
-
-  // helper to generate ids that are unique to model and deterministic, even if the execution provider is shared across
-  // multiple sessions.
-  class ModelMetadefIdGenerator {
-   public:
-    int GenerateId(const onnxruntime::GraphViewer& graph_viewer, HashValue& model_hash);
-
-   private:
-    std::unordered_map<HashValue, HashValue> main_graph_hash_;  // map graph instance hash to model contents hash
-    std::unordered_map<HashValue, int> model_metadef_id_;       // current unique id for model
-  };
-
-  std::unique_ptr<ModelMetadefIdGenerator> metadef_id_generator_;
-};
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/float16.h b/projects/llm_framework/include/onnxruntime/core/framework/float16.h
deleted file mode 100644
index 290a5e7f..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/float16.h
+++ /dev/null
@@ -1,159 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-#pragma once
-
-#include "endian.h"
-#if defined(CUDA_VERSION) && CUDA_VERSION >= 11000
-#include "cuda_bf16.h"
-#endif
-
-#if !defined(__CUDACC__) && !defined(__HIPCC__)
-#include "core/common/narrow.h"
-#endif
-
-#include "core/common/common.h"
-
-namespace onnxruntime {
-
-#if defined(__CUDACC__) || defined(__HIPCC__)
-#define ORT_HOST_DEVICE __host__ __device__
-#else
-#define ORT_HOST_DEVICE
-#endif
-
-// MLFloat16
-struct MLFloat16 {
-  uint16_t val{0};
-
-  MLFloat16() = default;
-  explicit constexpr MLFloat16(uint16_t x) : val(x) {}
-  explicit MLFloat16(float f);
-
-  float ToFloat() const;
-
-  operator float() const { return ToFloat(); }
-};
-
-inline bool operator==(const MLFloat16& left, const MLFloat16& right) { return left.val == right.val; }
-inline bool operator!=(const MLFloat16& left, const MLFloat16& right) { return left.val != right.val; }
-inline bool operator<(const MLFloat16& left, const MLFloat16& right) { return left.val < right.val; }
-
-// BFloat16
-struct BFloat16 {
-  uint16_t val{0};
-#if defined(__HIP__)
-  ORT_HOST_DEVICE BFloat16() = default;
-#else
-  BFloat16() = default;
-#endif
-
-  struct FromBitsT {};
-  static constexpr ORT_HOST_DEVICE FromBitsT FromBits() { return FromBitsT(); }
-  constexpr ORT_HOST_DEVICE BFloat16(unsigned short bits, FromBitsT) : val(bits) {}
-
-  inline ORT_HOST_DEVICE BFloat16(float v) {
-#if defined(CUDA_VERSION) && CUDA_VERSION >= 11000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
-    val = __bfloat16_as_ushort(__float2bfloat16(v));
-#elif defined(__HIP__)
-    // We should be using memcpy in order to respect the strict aliasing rule but it fails in the HIP environment.
-    if (v != v) {  // isnan
-      val = UINT16_C(0x7FC0);
-    } else {
-      union {
-        uint32_t U32;
-        float F32;
-      };
-
-      F32 = v;
-      uint32_t rounding_bias = ((U32 >> 16) & 1) + UINT32_C(0x7FFF);
-      val = static_cast<uint16_t>((U32 + rounding_bias) >> 16);
-    }
-#else
-    if constexpr(endian::native == endian::little) {
-      std::memcpy(&val, reinterpret_cast<char*>(&v) + sizeof(uint16_t), sizeof(uint16_t));
-    }
-    else {
-      std::memcpy(&val, &v, sizeof(uint16_t));
-    }
-#endif
-  }
-
-  inline ORT_HOST_DEVICE float ToFloat() const {
-#if defined(CUDA_VERSION) && CUDA_VERSION >= 11000
-    return __bfloat162float(*reinterpret_cast<const __nv_bfloat16*>(&val));
-#elif defined(__HIP__)
-    // We should be using memcpy in order to respect the strict aliasing rule but it fails in the HIP environment.
-    float result = 0;
-    uint32_t tmp = val;
-    tmp <<= 16;
-    float* tempRes = reinterpret_cast<float*>(&tmp);
-    result = *tempRes;
-    return result;
-#else
-    float result;
-    char* const first = reinterpret_cast<char*>(&result);
-    char* const second = first + sizeof(uint16_t);
-    if constexpr(endian::native == endian::little) {
-      std::memset(first, 0, sizeof(uint16_t));
-      std::memcpy(second, &val, sizeof(uint16_t));
-    }
-    else {
-      std::memcpy(first, &val, sizeof(uint16_t));
-      std::memset(second, 0, sizeof(uint16_t));
-    }
-    return result;
-#endif
-  }
-
-  inline ORT_HOST_DEVICE operator float() const { return ToFloat(); }
-
-#if defined(CUDA_VERSION) && CUDA_VERSION >= 11000
-  ORT_HOST_DEVICE BFloat16(const __nv_bfloat16& value) { val = *reinterpret_cast<const unsigned short*>(&value); }
-  explicit ORT_HOST_DEVICE operator __nv_bfloat16() const { return *reinterpret_cast<const __nv_bfloat16*>(&val); }
-#endif
-};
-
-inline ORT_HOST_DEVICE bool operator==(const BFloat16& left, const BFloat16& right) { return left.val == right.val; }
-inline ORT_HOST_DEVICE bool operator!=(const BFloat16& left, const BFloat16& right) { return left.val != right.val; }
-inline ORT_HOST_DEVICE bool operator<(const BFloat16& left, const BFloat16& right) { return left.val < right.val; }
-
-
-// User defined suffixes to make it easier to declare
-// initializers with MLFloat16 and BFloat16 from unsigned short
-// E.g 10_f16 or 10_b16
-#if !defined(__CUDACC__) && !defined(__HIPCC__)
-inline MLFloat16 operator"" _f16(unsigned long long int v) {
-  return MLFloat16(narrow<uint16_t>(v));
-}
-
-inline MLFloat16 operator"" _fp16(long double v) {
-  return MLFloat16(static_cast<float>(v));
-}
-
-inline BFloat16 operator"" _b16(unsigned long long int v) {
-  return BFloat16(narrow<uint16_t>(v), BFloat16::FromBits());
-}
-
-inline BFloat16 operator"" _bfp16(long double v) {
-  return BFloat16(static_cast<float>(v));
-}
-
-#endif
-
-inline void BFloat16ToFloat(const BFloat16* blf, float* flt, size_t size) {
-  auto src = blf;
-  auto d = flt;
-  for (; size != 0; ++src, ++d, --size) {
-    *d = src->ToFloat();
-  }
-}
-
-inline void FloatToBFloat16(const float* flt, BFloat16* blf, size_t size) {
-  auto src = flt;
-  auto d = blf;
-  for (; size != 0; ++src, ++d, --size) {
-    new (d) BFloat16(*src);
-  }
-}
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/framework_common.h b/projects/llm_framework/include/onnxruntime/core/framework/framework_common.h
deleted file mode 100644
index dd0dea98..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/framework_common.h
+++ /dev/null
@@ -1,22 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <string>
-#include <unordered_map>
-#include <vector>
-#include "run_options.h"
-
-namespace onnxruntime {  // forward declarations
-class Model;
-class GraphTransformer;
-class NodeArg;
-}  // namespace onnxruntime
-
-namespace onnxruntime {
-using InputDefList = std::vector<const onnxruntime::NodeArg*>;
-using OutputDefList = std::vector<const onnxruntime::NodeArg*>;
-
-using NameMLValMap = std::unordered_map<std::string, OrtValue>;
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/func_api.h b/projects/llm_framework/include/onnxruntime/core/framework/func_api.h
deleted file mode 100644
index 93b52d8a..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/func_api.h
+++ /dev/null
@@ -1,27 +0,0 @@
-#pragma once
-#include "core/common/status.h"
-using onnxruntime::common::Status;  // TODO: Needed by WinML, but shouldn't be put into the global namespace like this
-
-namespace onnxruntime {
-
-// AllocateFunc(void* handle, size_t alignment, size_t size)
-using AllocateFunc = void* (*)(void*, size_t, size_t);
-using DestroyFunc = void (*)(void*, void*);
-using AllocatorHandle = void*;
-
-typedef struct {
-  //right now we only include allocation for host memory
-  AllocateFunc allocate_func;
-  DestroyFunc release_func;
-  AllocatorHandle allocator_handle;
-  const char* node_name;
-} ComputeContext;
-
-using FunctionState = void*;
-// take the ComputeContext, and create a function state.
-using CreateFunctionStateC = int (*)(ComputeContext*, FunctionState*);
-// pass in the function state and input/output tensors, perform compute and return status
-using ComputeFuncC = common::Status (*)(FunctionState, const OrtApi*, OrtKernelContext*);
-// release the function state.
-using DestroyFunctionStateC = void (*)(FunctionState);
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/kernel_def_builder.h b/projects/llm_framework/include/onnxruntime/core/framework/kernel_def_builder.h
deleted file mode 100644
index 2892439f..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/kernel_def_builder.h
+++ /dev/null
@@ -1,353 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <limits.h>
-#include <memory>
-#include <optional>
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-#include "core/common/common.h"
-#include "core/framework/allocator.h"
-#include "core/framework/data_types.h"
-#include "core/graph/basic_types.h"
-
-namespace onnxruntime {
-class KernelDefBuilder;
-
-typedef std::map<size_t, OrtMemType> MemTypeMap;
-
-class KernelDef {
- private:
-  // note that input/output might be on CPU implicitly when the node is from CPU execution provider
-  constexpr static inline bool MemTypeOnCpuExplicitly(OrtMemType mem_type) {
-    return mem_type == OrtMemTypeCPUInput || mem_type == OrtMemTypeCPUOutput;
-  }
-
- public:
-  explicit KernelDef() = default;
-
-  const std::string& OpName() const {
-    return op_name_;
-  }
-
-  const std::string& Domain() const {
-    return op_domain_;
-  }
-
-  void SinceVersion(/*out*/ int* start, /*out*/ int* end) const {
-    *start = op_since_version_start_;
-    *end = op_since_version_end_;
-  }
-
-#ifdef onnxruntime_PYBIND_EXPORT_OPSCHEMA
-  const std::pair<int, int> SinceVersion() const {
-    return std::pair<int, int>(op_since_version_start_, op_since_version_end_);
-  }
-#endif
-
-  onnxruntime::ProviderType Provider() const {
-    return provider_type_;
-  }
-
-  // type constraints with types supported in this build
-  const std::unordered_map<std::string, std::vector<MLDataType>>& TypeConstraints() const {
-    return type_constraints_;
-  }
-
-  const std::vector<std::pair<int, int>>& MayInplace() const {
-    return inplace_map_;
-  }
-
-  const std::vector<std::pair<int, int>>& Alias() const {
-    return alias_map_;
-  }
-
-  const std::optional<std::pair<int, int>>& VariadicAlias() const {
-    return variadic_alias_offsets_;
-  }
-
-  OrtMemType InputMemoryType(size_t input_index) const {
-    auto it = input_memory_type_args_.find(input_index);
-    if (it == input_memory_type_args_.end())
-      return default_inputs_mem_type_;
-    return it->second;
-  }
-
-  bool IsInputOnCpu(size_t input_index) const { return MemTypeOnCpuExplicitly(InputMemoryType(input_index)); }
-
-  bool IsOutputOnCpu(size_t output_index) const { return MemTypeOnCpuExplicitly(OutputMemoryType(output_index)); }
-
-  bool AllocateInputsContiguously() const { return allocate_inputs_contiguously_; }
-
-  bool HasExternalOutputs() const { return external_outputs_; }
-
-#ifdef ENABLE_STRIDED_TENSORS
-  const std::vector<int>& MayStridedInput() const { return may_strided_inputs_; }
-  const std::vector<std::pair<int, int>>& MayStridedOutput() const { return may_strided_output_map_; }
-#endif
-
-  OrtMemType OutputMemoryType(size_t output_index) const {
-    auto it = output_memory_type_args_.find(output_index);
-    if (it == output_memory_type_args_.end())
-      return default_outputs_mem_type_;
-    return it->second;
-  }
-
-  int ExecQueueId() const {
-    return exec_queue_id_;
-  }
-
-  bool IsConflict(const KernelDef& other) const;
-
- private:
-  friend class KernelDefBuilder;
-
-  // The operator name supported by <*this> kernel..
-  std::string op_name_;
-
-  // The operator since_version range supported by <*this> kernel.
-  // A kernel could support an operator definition between <op_since_version_start>
-  // and <op_since_version_end> (inclusive).
-  int op_since_version_start_ = 1;
-  int op_since_version_end_ = INT_MAX;
-
-  // The operator domain supported by <*this> kernel.
-  // Default to 'onnxruntime::kOnnxDomain'.
-  // Please note the behavior of std::string("") and std::string() are different
-  std::string op_domain_;
-
-  // The type of the execution provider.
-  std::string provider_type_;
-
-  // The data types that are supported in this build (enabled) for inputs/outputs.
-  // Key is input/output/type constraint name defined in op schema, Value is supported types.
-  std::unordered_map<std::string, std::vector<MLDataType>> type_constraints_;
-
-  // An element <i, j> means that output j reuses the memory of input i.
-  std::vector<std::pair<int, int>> inplace_map_;
-
-  // An element <i, j> means that output j is an alias of input i.
-  std::vector<std::pair<int, int>> alias_map_;
-
-  // This variable stores <input_offset, output_offset> for the variadic alias mapping
-  // output 'i + output_offset' is an alias of input 'i + input_offset' for all i >= 0
-  std::optional<std::pair<int, int>> variadic_alias_offsets_;
-
-  // Require input tensors to be allocated contiguously.
-  bool allocate_inputs_contiguously_ = false;
-
-  // Whether the outputs are from external.
-  bool external_outputs_ = false;
-
-#ifdef ENABLE_STRIDED_TENSORS
-  // An element i means i-th input can be strided tensor.
-  std::vector<int> may_strided_inputs_;
-
-  // An element <i, j> means j-th output can be a strided tensor, which share the data from i-th input.
-  std::vector<std::pair<int, int>> may_strided_output_map_;
-#endif
-
-  // The memory types of inputs/outputs of this kernel
-  MemTypeMap input_memory_type_args_;
-  MemTypeMap output_memory_type_args_;
-
-  // execution command queue id, 0 for default queue in execution provider
-  int exec_queue_id_ = 0;
-  // Default memory type for all inputs
-  OrtMemType default_inputs_mem_type_{OrtMemTypeDefault};
-  // Default memory type for all outputs
-  OrtMemType default_outputs_mem_type_{OrtMemTypeDefault};
-};
-
-class KernelDefBuilder {
- public:
-  static std::unique_ptr<KernelDefBuilder> Create() { return std::make_unique<KernelDefBuilder>(); }
-
-  explicit KernelDefBuilder()
-      : kernel_def_(std::make_unique<KernelDef>()) {}
-
-  KernelDefBuilder& SetName(const std::string& op_name);
-  KernelDefBuilder& SetName(const char* op_name);
-
-  KernelDefBuilder& SetDomain(const std::string& domain);
-  KernelDefBuilder& SetDomain(const char* domain);
-
-  /**
-     This kernel supports operator definition since <since_version> (to latest).
-  */
-  KernelDefBuilder& SinceVersion(int since_version) {
-    kernel_def_->op_since_version_start_ = since_version;
-    return *this;
-  }
-
-  /**
-     The start and end version should be set accordingly per version range for
-     each domain registered in OpSchemaRegistry::DomainToVersionRange in
-     \onnxruntime\onnxruntime\core\graph\op.h as below.
-     Key: domain. Value: <lowest version, highest version> pair.
-     std::unordered_map<std::string, std::pair<int, int>> map_;
-  */
-  KernelDefBuilder& SinceVersion(int since_version_start, int since_version_end) {
-    kernel_def_->op_since_version_start_ = since_version_start;
-    kernel_def_->op_since_version_end_ = since_version_end;
-    return *this;
-  }
-
-  /**
-     The execution provider type of the kernel.
-  */
-  KernelDefBuilder& Provider(ProviderType provider_type);
-  KernelDefBuilder& Provider(const char* provider_type);
-
-  /**
-     Specify the set of types that this kernel supports. A further restriction
-     of the set of types specified in the op schema.
-
-     @param arg_name The arg name can be either op formal parameter name, say "X", or type
-                     argument name specified in op schema, say "T".
-     @param types The types that are supported in this build.
-  */
-  KernelDefBuilder& TypeConstraint(const std::string& arg_name, std::vector<MLDataType> types);
-  KernelDefBuilder& TypeConstraint(const char* arg_name, std::vector<MLDataType> types);
-
-  /**
-     Like TypeConstraint but supports just a single type.
-  */
-  KernelDefBuilder& TypeConstraint(const std::string& arg_name, MLDataType type);
-  KernelDefBuilder& TypeConstraint(const char* arg_name, MLDataType type);
-
-  /**
-     Inplace mapping from inputs to outputs allowed.
-     It means that uplayer runtime could do memory in-place optimization
-     as it will not impact the correctness of this kernel.
-  */
-  KernelDefBuilder& MayInplace(const std::vector<std::pair<int, int>>& inplaces);
-  KernelDefBuilder& MayInplace(int input_index, int output_index);
-
-  /**
-     Alias mapping from inputs to outputs. Different from Inplace that the
-     content of the tensor is not changed. This is to take care of operators
-     such as Identity and Reshape.
-  */
-  KernelDefBuilder& Alias(const std::vector<std::pair<int, int>>& aliases);
-  KernelDefBuilder& Alias(int input_index, int output_index);
-
-  /**
-     Apply variadic number of alias mapping from inputs to outputs.
-     This is effectively applying Alias(i + input_offset, i + output_offset) for i >= 0
-  */
-  KernelDefBuilder& VariadicAlias(int input_offset, int output_offset);
-
-  /**
-     Specify that this kernel requires input tensors to be allocated
-     contiguously. This allows kernels to execute as a single large
-     computation, rather than numerous smaller computations.
-  */
-  KernelDefBuilder& AllocateInputsContiguously() {
-    kernel_def_->allocate_inputs_contiguously_ = true;
-    return *this;
-  }
-
-  /**
-     Specify that this kernel's output buffers are passed from external,
-     i.e. not created or managed by ORT's memory allocator.
-  */
-  KernelDefBuilder& ExternalOutputs() {
-    kernel_def_->external_outputs_ = true;
-    return *this;
-  }
-
-#ifdef ENABLE_STRIDED_TENSORS
-  /**
-     Specify that the input_index-th input can be strided tensor.
-   */
-  KernelDefBuilder& MayStridedInput(int input_index);
-
-  /**
-     Specify that the output_index-th output can be strided tensor, and share the data
-     from input_index-th input.
-   */
-  KernelDefBuilder& MayStridedOutput(int input_index, int output_index);
-#endif
-
-  /**
-     Specify that this kernel requires an input arg
-     in certain memory type (instead of the default, device memory).
-  */
-  KernelDefBuilder& InputMemoryType(OrtMemType type, int input_index) {
-    kernel_def_->input_memory_type_args_.insert(std::make_pair(input_index, type));
-    return *this;
-  }
-
-  /**
-     Specify that this kernel requires input arguments
-     in certain memory type (instead of the default, device memory).
-  */
-  KernelDefBuilder& InputMemoryType(OrtMemType type, const std::vector<int>& input_indexes) {
-    for (auto input_index : input_indexes) {
-      kernel_def_->input_memory_type_args_.insert(std::make_pair(input_index, type));
-    }
-    return *this;
-  }
-
-  /**
-     Specify that this kernel provides an output arg
-     in certain memory type (instead of the default, device memory).
-  */
-  KernelDefBuilder& OutputMemoryType(OrtMemType type, int output_index) {
-    kernel_def_->output_memory_type_args_.insert(std::make_pair(output_index, type));
-    return *this;
-  }
-
-  /**
-     Specify that this kernel provides an output arguments
-     in certain memory type (instead of the default, device memory).
-  */
-  KernelDefBuilder& OutputMemoryType(OrtMemType type, const std::vector<int>& output_indexes) {
-    for (auto output_index : output_indexes) {
-      kernel_def_->output_memory_type_args_.insert(std::make_pair(output_index, type));
-    }
-    return *this;
-  }
-
-  /**
-     Specify that this kernel runs on which execution queue in the provider
-  */
-  KernelDefBuilder& ExecQueueId(int queue_id) {
-    kernel_def_->exec_queue_id_ = queue_id;
-    return *this;
-  }
-
-  /**
-  Specify the default inputs memory type, if not specified, it is DefaultMemory
-  */
-  KernelDefBuilder& SetDefaultInputsMemoryType(OrtMemType mem_type) {
-    kernel_def_->default_inputs_mem_type_ = mem_type;
-    return *this;
-  }
-
-  /**
-  Specify the default outputs memory type, if not specified, it is DefaultMemory
-  */
-  KernelDefBuilder& SetDefaultOutputMemoryType(OrtMemType mem_type) {
-    kernel_def_->default_outputs_mem_type_ = mem_type;
-    return *this;
-  }
-
-  /**
-     Return the kernel definition, passing ownership of the KernelDef to the caller
-  */
-  std::unique_ptr<KernelDef> Build() {
-    return std::move(kernel_def_);
-  }
-
- private:
-  // we own the KernelDef until Build() is called.
-  std::unique_ptr<KernelDef> kernel_def_;
-};
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/kernel_registry.h b/projects/llm_framework/include/onnxruntime/core/framework/kernel_registry.h
deleted file mode 100644
index 68b610ac..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/kernel_registry.h
+++ /dev/null
@@ -1,91 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <string_view>
-
-#include "core/framework/op_kernel.h"
-
-namespace onnxruntime {
-
-using KernelCreateMap = std::multimap<std::string, KernelCreateInfo>;
-using KernelDefHashes = std::vector<std::pair<std::string, HashValue>>;
-
-class IKernelTypeStrResolver;
-
-/**
- * Each provider has a KernelRegistry. Often, the KernelRegistry only belongs to that specific provider.
- */
-class KernelRegistry {
- public:
-  KernelRegistry() = default;
-
-  // Register a kernel with kernel definition and function to create the kernel.
-  Status Register(KernelDefBuilder& kernel_def_builder, const KernelCreateFn& kernel_creator);
-
-  Status Register(KernelCreateInfo&& create_info);
-
-  // TODO(edgchen1) for TryFindKernel(), consider using `out` != nullptr as indicator of whether kernel was found and
-  // Status as an indication of failure
-
-  // Check if an execution provider can create kernel for a node and return the kernel if so
-  Status TryFindKernel(const Node& node, ProviderType exec_provider,
-                       const IKernelTypeStrResolver& kernel_type_str_resolver,
-                       const KernelCreateInfo** out) const;
-
-  static bool HasImplementationOf(const KernelRegistry& r, const Node& node,
-                                  ProviderType exec_provider,
-                                  const IKernelTypeStrResolver& kernel_type_str_resolver) {
-    const KernelCreateInfo* info;
-    Status st = r.TryFindKernel(node, exec_provider, kernel_type_str_resolver, &info);
-    return st.IsOK();
-  }
-
-#if !defined(ORT_MINIMAL_BUILD)
-  // Find KernelCreateInfo in instant mode
-  Status TryFindKernel(const std::string& op_name, const std::string& domain, const int& version,
-                       const std::unordered_map<std::string, MLDataType>& type_constraints,
-                       ProviderType exec_provider, const KernelCreateInfo** out) const;
-#endif  // !defined(ORT_MINIMAL_BUILD)
-
-  bool IsEmpty() const { return kernel_creator_fn_map_.empty(); }
-
-#ifdef onnxruntime_PYBIND_EXPORT_OPSCHEMA
-  // This is used by the opkernel doc generator to enlist all registered operators for a given provider's opkernel
-  const KernelCreateMap& GetKernelCreateMap() const {
-    return kernel_creator_fn_map_;
-  }
-#endif
-
- private:
-  // Check whether the types of inputs/outputs of the given node match the extra
-  // type-constraints of the given kernel. This serves two purposes: first, to
-  // select the right kernel implementation based on the types of the arguments
-  // when we have multiple kernels, e.g., Clip<float> and Clip<int>; second, to
-  // accommodate (and check) mapping of ONNX (specification) type to the onnxruntime
-  // implementation type (e.g., if we want to implement ONNX's float16 as a regular
-  // float in onnxruntime). (The second, however, requires a globally uniform mapping.)
-  //
-  // Note that this is not intended for type-checking the node against the ONNX
-  // type specification of the corresponding op, which is done before this check.
-  static bool VerifyKernelDef(const Node& node,
-                              const KernelDef& kernel_def,
-                              const IKernelTypeStrResolver& kernel_type_str_resolver,
-                              std::string& error_str);
-
-  static std::string GetMapKey(std::string_view op_name, std::string_view domain, std::string_view provider) {
-    std::string key(op_name);
-    // use the kOnnxDomainAlias of 'ai.onnx' instead of kOnnxDomain's empty string
-    key.append(1, ' ').append(domain.empty() ? kOnnxDomainAlias : domain).append(1, ' ').append(provider);
-    return key;
-  }
-
-  static std::string GetMapKey(const KernelDef& kernel_def) {
-    return GetMapKey(kernel_def.OpName(), kernel_def.Domain(), kernel_def.Provider());
-  }
-  // Kernel create function map from op name to kernel creation info.
-  // key is opname+domain_name+provider_name
-  KernelCreateMap kernel_creator_fn_map_;
-};
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/op_kernel.h b/projects/llm_framework/include/onnxruntime/core/framework/op_kernel.h
deleted file mode 100644
index e7e4ab94..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/op_kernel.h
+++ /dev/null
@@ -1,387 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include "boost/mp11.hpp"
-
-// It is safe to include the below header even if SHARED_PROVIDER macro is enabled
-// as it doesn't include any pb headers.
-#include "core/framework/prepacked_weights_container.h"
-
-#ifndef SHARED_PROVIDER
-#include <functional>
-#include "core/common/exceptions.h"
-#include "core/common/logging/logging.h"
-#include "core/common/status.h"
-#include "core/framework/execution_provider.h"
-#include "core/framework/kernel_def_builder.h"
-#include "core/framework/ort_value.h"
-#include "core/framework/op_kernel_info.h"
-#include "core/framework/op_node_proto_helper.h"
-#include "core/framework/tensor.h"
-#include "core/framework/sparse_tensor.h"
-#include "core/graph/constants.h"
-#include "core/graph/graph_viewer.h"
-#if !defined(ORT_MINIMAL_BUILD)
-#include "onnx/defs/schema.h"
-#else
-#include "onnx/defs/data_type_utils.h"
-#endif
-#include "onnx/onnx_pb.h"
-#include "onnx/onnx-operators_pb.h"
-#include "core/common/gsl.h"
-namespace onnxruntime {
-class OpKernelContext;
-}
-#endif
-
-namespace onnxruntime {
-
-std::unique_ptr<OpKernelInfo> CopyOpKernelInfo(const OpKernelInfo& info);
-
-class OpKernel {
- public:
-  using DoneCallback = std::function<void()>;
-
-  explicit OpKernel(const OpKernelInfo& info) : op_kernel_info_(CopyOpKernelInfo(info)) {}
-  virtual ~OpKernel() = default;
-
-  const onnxruntime::Node& Node() const;
-  const onnxruntime::KernelDef& KernelDef() const;
-
-  [[nodiscard]] virtual Status Compute(_Inout_ OpKernelContext* context) const = 0;
-
-  [[nodiscard]] virtual bool IsAsync() const {
-    // by default all kernels are sync version.
-    return false;
-  }
-
-  [[nodiscard]] virtual Status ComputeAsync(_Inout_ OpKernelContext*, DoneCallback) const {
-    ORT_NOT_IMPLEMENTED(__FUNCTION__, " is not implemented");
-  }
-
-  // Override this function to PrePack initialized constant tensor to the format as needed.
-  // For example, MatMul kernel can pack the input B if it is constant like code below.
-  //   Status PrePack(const Tensor& tensor, int input_idx, /*out*/ bool& is_packed,
-  //                  /*out*/ PrePackedWeights* prepacked_weight_for_caching,
-  //                  AllocatorPtr alloc) override {
-  //     is_packed = false;
-  //     if (input_idx == 1) {
-  //       is_packed = true;
-  //       this.Pack(tensor, this.buffer_, alloc);
-  //       if (prepacked_weight_for_caching) {
-  //           // LOGIC TO CACHE `this.buffer_` SINCE THE KERNEL DOESN"T OWN THE PACKED WEIGHT
-  //       }
-  //     }
-  //     return Status::OK();
-  //   }
-  // Please refer to MatMulIntegerToFloatBase for a complete example
-  // @param tensor: The initialized constant tensor
-  // @param input_idx: The input index of the tensor in this kernel
-  // @param alloc: The kernel's PrePack() method MUST use this allocator for allocating the pre-packed
-  //               weights' buffers. The alloc that the PrePack() method will receive will be either
-  //               the allocator tied to the session if the kernel owns the pre-packed buffer or an
-  //               allocator shared between sessions if the pre-packed buffer is to be shared across sessions
-  //               (i.e.) the kernel does not own the buffer.
-  // @param is_packed: Set it to true if the kernel packed the tensor or to false
-  //                   The kernel is responsible for keeping the packed data and related metadata if is_packed is true,
-  //                   and the original initialized constant tensor will be released and not accessible anymore in
-  //                   the Compute function.
-  // @param prepacked_weights: A PrePackedWeights instance will be provided to the kernel IF the pre-packed weights
-  //                          are meant to be stored in a shared container.
-
-  virtual Status
-  PrePack(const Tensor& /*tensor*/, int /*input_idx*/, AllocatorPtr /*alloc*/,
-          /*out*/ bool& is_packed, /*out*/ PrePackedWeights* /*prepacked_weights*/) {
-    is_packed = false;
-    return Status::OK();
-  }
-
-  // Override this function to use provided pre-packed weight.
-  // Status UseSharedPrePackedBuffers(std::vector<BufferUniquePtr>& prepacked_buffers,
-  //                                 int input_idx,
-  //                                 /*out*/ bool& used_shared_buffers) {
-  //     used_shared_buffers = true;
-  //     this.buffer_ = std::move(prepacked_buffers[0]);
-  //     return Status::OK();
-  //   }
-  // Please refer to MatMulIntegerToFloatBase for a complete example
-  // @param prepacked_buffers: The pre-packed buffers to be used by this kernel for the provided input index
-  //                           (Sometimes a single constant initializer may have multiple pre-packed buffers associated
-  //                            with it and it upto the kernel developer to store it in any order of their choice in PrePack()
-  //                            and must use the same order for retrieval in UseSharedPrePackedBuffers().
-  // @param input_idx: The input index of the tensor in this kernel
-  // @param used_shared_buffers: Boolean flag set by the kernel implementation indicating
-  // that the provided weight has been used by the kernel.
-  virtual Status UseSharedPrePackedBuffers(std::vector<BufferUniquePtr>& /*prepacked_buffers*/,
-                                           int /*input_idx*/,
-                                           /*out*/ bool& used_shared_buffers) {
-    used_shared_buffers = false;
-    return Status::OK();
-  }
-
-  const OrtMemoryInfo& Allocator(int id, OrtMemType mem_type) const;
-  const OpKernelInfo& Info() const {
-    return *op_kernel_info_;
-  }
-
- private:
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(OpKernel);
-  std::unique_ptr<OpKernelInfo> op_kernel_info_;
-};
-class FuncManager;
-using KernelCreateFn = std::function<Status(FuncManager& func_mgr, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out)>;
-using KernelCreatePtrFn = std::add_pointer<Status(FuncManager& func_mgr, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out)>::type;
-
-struct KernelCreateInfo {
-  std::unique_ptr<KernelDef> kernel_def;  // Owned and stored in the global kernel registry.
-  KernelCreateFn kernel_create_func;
-  Status status;
-
-  KernelCreateInfo(std::unique_ptr<KernelDef> definition,
-                   KernelCreateFn create_func)
-      : kernel_def(std::move(definition)),
-        kernel_create_func(create_func) {}
-
-  KernelCreateInfo(KernelCreateInfo&& other) noexcept
-      : kernel_def(std::move(other.kernel_def)),
-        kernel_create_func(std::move(other.kernel_create_func)) {}
-
-  KernelCreateInfo() = default;
-};
-
-// Forward declarations for the non-specialized BuildKernelCreateInfo method.
-template <typename T>
-KernelCreateInfo BuildKernelCreateInfo();
-
-namespace ml {
-template <typename T>
-KernelCreateInfo BuildKernelCreateInfo();
-}  // namespace ml
-
-namespace contrib {
-template <typename T>
-KernelCreateInfo BuildKernelCreateInfo();
-}  // namespace contrib
-
-namespace contrib {
-namespace cuda {
-template <typename T>
-KernelCreateInfo BuildKernelCreateInfo();
-}  // namespace cuda
-}  // namespace contrib
-
-namespace contrib {
-namespace rocm {
-template <typename T>
-KernelCreateInfo BuildKernelCreateInfo();
-}  // namespace rocm
-}  // namespace contrib
-
-namespace contrib {
-namespace snpe {
-template <typename T>
-KernelCreateInfo BuildKernelCreateInfo();
-}  // namespace snpe
-}  // namespace contrib
-
-using BuildKernelCreateInfoFn = KernelCreateInfo (*)();
-
-// Naming convention for operator kernel classes
-#define ONNX_OPERATOR_KERNEL_CLASS_NAME(provider, domain, ver, name) \
-  provider##_##name##_##domain##_ver##ver
-
-#define ONNX_CPU_OPERATOR_KERNEL(name, ver, builder, ...) \
-  ONNX_OPERATOR_KERNEL_EX(name, kOnnxDomain, ver, kCpuExecutionProvider, builder, __VA_ARGS__)
-
-#define ONNX_CPU_OPERATOR_ML_KERNEL(name, ver, builder, ...) \
-  ONNX_OPERATOR_KERNEL_EX(name, kMLDomain, ver, kCpuExecutionProvider, builder, __VA_ARGS__)
-
-#define ONNX_CPU_OPERATOR_MS_KERNEL(name, ver, builder, ...) \
-  ONNX_OPERATOR_KERNEL_EX(name, kMSDomain, ver, kCpuExecutionProvider, builder, __VA_ARGS__)
-
-#define ONNX_OPERATOR_KERNEL_EX(name, domain, ver, provider, builder, ...)                \
-  class ONNX_OPERATOR_KERNEL_CLASS_NAME(provider, domain, ver, name);                     \
-  template <>                                                                             \
-  KernelCreateInfo                                                                        \
-  BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(provider, domain, ver, name)>() { \
-    return KernelCreateInfo(                                                              \
-        builder.SetName(#name)                                                            \
-            .SetDomain(domain)                                                            \
-            .SinceVersion(ver)                                                            \
-            .Provider(provider)                                                           \
-            .Build(),                                                                     \
-        static_cast<KernelCreatePtrFn>(                                                   \
-            [](FuncManager&,                                                              \
-               const OpKernelInfo& info,                                                  \
-               std::unique_ptr<OpKernel>& out) -> Status {                                \
-              out = std::make_unique<__VA_ARGS__>(info);                                  \
-              return Status::OK();                                                        \
-            }));                                                                          \
-  }
-
-#define ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(provider, domain, startver, endver, name) \
-  provider##_##name##_##domain##_ver##startver##_##endver
-
-#define ONNX_CPU_OPERATOR_VERSIONED_KERNEL(name, startver, endver, builder, ...) \
-  ONNX_OPERATOR_VERSIONED_KERNEL_EX(name, kOnnxDomain, startver, endver, kCpuExecutionProvider, builder, __VA_ARGS__)
-
-#define ONNX_CPU_OPERATOR_VERSIONED_ML_KERNEL(name, startver, endver, builder, ...) \
-  ONNX_OPERATOR_VERSIONED_KERNEL_EX(name, kMLDomain, startver, endver, kCpuExecutionProvider, builder, __VA_ARGS__)
-
-#define ONNX_OPERATOR_VERSIONED_KERNEL_EX(name, domain, startver, endver, provider, builder, ...)                                  \
-  class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(provider, domain, startver, endver, name);                                       \
-  template <>                                                                                                                      \
-  KernelCreateInfo                                                                                                                 \
-  BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(provider, domain, startver, endver, name)>() {                   \
-    return KernelCreateInfo(                                                                                                       \
-        builder.SetName(#name)                                                                                                     \
-            .SetDomain(domain)                                                                                                     \
-            .SinceVersion(startver, endver)                                                                                        \
-            .Provider(provider)                                                                                                    \
-            .Build(),                                                                                                              \
-        static_cast<KernelCreatePtrFn>([](FuncManager&, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out) -> Status { out = std::make_unique<__VA_ARGS__>(info); return Status::OK(); })); \
-  }
-
-#define ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(provider, domain, ver, type, name) \
-  provider##_##name##_##domain##_ver##ver##_##type
-
-#define ONNX_CPU_OPERATOR_TYPED_KERNEL(name, ver, type, builder, ...) \
-  ONNX_OPERATOR_TYPED_KERNEL_EX(name, kOnnxDomain, ver, type, kCpuExecutionProvider, builder, __VA_ARGS__)
-
-#define ONNX_CPU_OPERATOR_TYPED_ML_KERNEL(name, ver, type, builder, ...) \
-  ONNX_OPERATOR_TYPED_KERNEL_EX(name, kMLDomain, ver, type, kCpuExecutionProvider, builder, __VA_ARGS__)
-
-#define ONNX_CPU_OPERATOR_TYPED_MS_KERNEL(name, ver, type, builder, ...) \
-  ONNX_OPERATOR_TYPED_KERNEL_EX(name, kMSDomain, ver, type, kCpuExecutionProvider, builder, __VA_ARGS__)
-
-#define ONNX_OPERATOR_TYPED_KERNEL_EX(name, domain, ver, type, provider, builder, ...)                                             \
-  class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(provider, domain, ver, type, name);                                                  \
-  template <>                                                                                                                      \
-  KernelCreateInfo                                                                                                                 \
-  BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(provider, domain, ver, type, name)>() {                              \
-    return KernelCreateInfo(                                                                                                       \
-        builder.SetName(#name)                                                                                                     \
-            .SetDomain(domain)                                                                                                     \
-            .SinceVersion(ver)                                                                                                     \
-            .Provider(provider)                                                                                                    \
-            .Build(),                                                                                                              \
-        static_cast<KernelCreatePtrFn>([](FuncManager&, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out) -> Status { out = std::make_unique<__VA_ARGS__>(info); return Status::OK(); })); \
-  }
-
-#define ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(provider, domain, ver, type1, type2, name) \
-  provider##_##name##_##domain##_ver##ver##_##type1##_##type2
-
-#define ONNX_OPERATOR_TWO_TYPED_KERNEL_EX(name, domain, ver, type1, type2, provider, builder, ...)                                 \
-  class ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(provider, domain, ver, type1, type2, name);                                      \
-  template <>                                                                                                                      \
-  KernelCreateInfo                                                                                                                 \
-  BuildKernelCreateInfo<ONNX_OPERATOR_TWO_TYPED_KERNEL_CLASS_NAME(provider, domain, ver, type1, type2, name)>() {                  \
-    return KernelCreateInfo(                                                                                                       \
-        builder.SetName(#name)                                                                                                     \
-            .SetDomain(domain)                                                                                                     \
-            .SinceVersion(ver)                                                                                                     \
-            .Provider(provider)                                                                                                    \
-            .Build(),                                                                                                              \
-        static_cast<KernelCreatePtrFn>([](FuncManager&, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out) -> Status { out = std::make_unique<__VA_ARGS__>(info); return Status::OK(); })); \
-  }
-
-#define ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(provider, domain, startver, endver, type, name) \
-  provider##_##name##_##domain##_ver##startver##_##endver##_##type
-
-#define ONNX_CPU_OPERATOR_VERSIONED_TYPED_KERNEL(name, startver, endver, type, builder, ...)                         \
-  ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX(name, kOnnxDomain, startver, endver, type, kCpuExecutionProvider, builder, \
-                                          __VA_ARGS__)
-
-#define ONNX_CPU_OPERATOR_VERSIONED_TYPED_ML_KERNEL(name, startver, endver, type, builder, ...)                    \
-  ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX(name, kMLDomain, startver, endver, type, kCpuExecutionProvider, builder, \
-                                          __VA_ARGS__)
-
-#define ONNX_CPU_OPERATOR_VERSIONED_TYPED_MS_KERNEL(name, startver, endver, type, builder, ...)                    \
-  ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX(name, kMSDomain, startver, endver, type, kCpuExecutionProvider, builder, \
-                                          __VA_ARGS__)
-
-#define ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX(name, domain, startver, endver, type, provider, builder, ...)                      \
-  class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(provider, domain, startver, endver, type, name);                           \
-  template <>                                                                                                                      \
-  KernelCreateInfo                                                                                                                 \
-  BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(provider, domain, startver, endver,                        \
-                                                                        type, name)>() {                                           \
-    return KernelCreateInfo(                                                                                                       \
-        builder.SetName(#name)                                                                                                     \
-            .SetDomain(domain)                                                                                                     \
-            .SinceVersion(startver, endver)                                                                                        \
-            .Provider(provider)                                                                                                    \
-            .Build(),                                                                                                              \
-        static_cast<KernelCreatePtrFn>([](FuncManager&, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out) -> Status { out = std::make_unique<__VA_ARGS__>(info); return Status::OK(); })); \
-  }
-
-#define ONNX_OPERATOR_VERSIONED_TWO_TYPED_KERNEL_CLASS_NAME(provider, domain, startver, endver, type1, type2, name) \
-  provider##_##name##_##domain##_ver##startver##_##endver##_##type1##_##type2
-
-#define ONNX_OPERATOR_VERSIONED_TWO_TYPED_KERNEL_EX(name, domain, startver, endver, type1, type2,                                  \
-                                                    provider, builder, ...)                                                        \
-  class ONNX_OPERATOR_VERSIONED_TWO_TYPED_KERNEL_CLASS_NAME(provider, domain, startver, endver, type1, type2, name);               \
-  template <>                                                                                                                      \
-  KernelCreateInfo                                                                                                                 \
-  BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TWO_TYPED_KERNEL_CLASS_NAME(provider, domain, startver, endver,                    \
-                                                                            type1, type2, name)>() {                               \
-    return KernelCreateInfo(                                                                                                       \
-        builder.SetName(#name)                                                                                                     \
-            .SetDomain(domain)                                                                                                     \
-            .SinceVersion(startver, endver)                                                                                        \
-            .Provider(provider)                                                                                                    \
-            .Build(),                                                                                                              \
-        static_cast<KernelCreatePtrFn>([](FuncManager&, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out) -> Status { out = std::make_unique<__VA_ARGS__>(info); return Status::OK(); })); \
-  }
-
-template <typename... Types>
-struct BuildKernelDefConstraintsImpl {
-  std::vector<MLDataType> operator()() const {
-    return {DataTypeImpl::GetTensorType<Types>()...};
-  }
-};
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-template <typename... Types>
-struct BuildKernelDefSparseConstraintsImpl {
-  std::vector<MLDataType> operator()() const {
-    return {DataTypeImpl::GetSparseTensorType<Types>()...};
-  }
-};
-#endif
-
-// Use within macro definitions to create a custom vector of constraints.
-// Example: #define REG_KERNEL(OP, VERSION, KERNEL_CLASS, Type, ...)
-//  .TypeConstraint("T", BuildKernelDefConstraints<Type, __VA_ARGS_>())
-template <typename... Types>
-inline std::vector<MLDataType> BuildKernelDefConstraints() {
-  return BuildKernelDefConstraintsImpl<Types...>{}();
-}
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-template <typename... Types>
-inline std::vector<MLDataType> BuildKernelDefSparseConstraints() {
-  return BuildKernelDefSparseConstraintsImpl<Types...>{}();
-}
-#endif
-
-// version of BuildKernelDefConstraints() which takes a type list
-template <typename L>
-inline std::vector<MLDataType> BuildKernelDefConstraintsFromTypeList() {
-  return boost::mp11::mp_apply<BuildKernelDefConstraintsImpl, L>{}();
-}
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-template <typename L>
-inline std::vector<MLDataType> BuildKernelDefSparseConstraintsFromTypeList() {
-  return boost::mp11::mp_apply<BuildKernelDefSparseConstraintsImpl, L>{}();
-}
-#endif
-
-}  // namespace onnxruntime
-
-#ifndef SHARED_PROVIDER
-#include "core/framework/op_kernel_context.h"
-#endif
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/op_kernel_context.h b/projects/llm_framework/include/onnxruntime/core/framework/op_kernel_context.h
deleted file mode 100644
index 58e2891d..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/op_kernel_context.h
+++ /dev/null
@@ -1,237 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-namespace onnxruntime {
-class IExecutionFrame;
-class Stream;
-namespace concurrency {
-class ThreadPool;
-}
-
-class OpKernelContext {
- public:
-  using ArgMap = std::unordered_map<std::string, size_t>;
-
-  OpKernelContext(_Inout_ IExecutionFrame* frame, _In_ const OpKernel* kernel,
-                  _In_ Stream* stream,
-                  _In_opt_ concurrency::ThreadPool* threadpool, _In_ const logging::Logger& logger);
-
-  virtual ~OpKernelContext() = default;
-
-  /**
-  Return the number of inputs for a variadic argument.
-  @param arg_num The operator argument number.
-  @returns Number of inputs the argument has.
-  */
-  virtual int NumVariadicInputs(size_t arg_num) const;
-
-  virtual MLDataType InputType(int index) const;
-  virtual MLDataType OutputType(int index) const;
-
-  const OrtValue* GetInputOrtValue(int index) const {
-    return GetInputMLValue(index);
-  }
-
-  template <typename T>
-  const T* Input(int index) const {
-    const OrtValue* p_ml_value = GetInputMLValue(index);
-    ORT_TRY {
-      return p_ml_value ? &(p_ml_value->Get<T>()) : nullptr;
-    }
-    ORT_CATCH(const std::exception& /*e*/) {
-      ORT_THROW("Missing Input: " + kernel_->Node().InputDefs()[index]->Name());
-    }
-  }
-
-  // Fetch a required input, enforcing that it is present.
-  template <typename T>
-  const T& RequiredInput(int index) const {
-    const T* input_ptr = Input<T>(index);
-    ORT_ENFORCE(input_ptr, "Required input at index ", index, " is not present.");
-    return *input_ptr;
-  }
-
-  // Fetch output (non-tensor) with specified index.
-  template <typename T>
-  T* Output(int index) {
-    if (index < 0 || index >= OutputCount())
-      return nullptr;
-
-    OrtValue* p_ml_value = GetOrCreateOutputMLValue(index);
-    return p_ml_value ? p_ml_value->GetMutable<T>() : nullptr;
-  }
-
-  // In the case that memory allocation has not been done for an output tensor,
-  // The memory allocation will be done on-the-fly with given tensor shape.
-  // Return nullptr if the output is an unused optional output.
-  Tensor* Output(int index, const TensorShape& shape);
-  Tensor* Output(int index, const std::vector<int64_t>& shape);
-  Tensor* Output(int index, const std::initializer_list<int64_t>& shape);
-
-  // Fetch a required tensor output, enforcing that it is present.
-  Tensor& RequiredOutput(int index, const TensorShape& shape) {
-    Tensor* output_ptr = Output(index, shape);
-    ORT_ENFORCE(output_ptr, "Required output at index ", index, " is not present.");
-    return *output_ptr;
-  }
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-  // Fetch a sparse-tensor output corresponding to the specified index.
-  // shape must specify the shape of the underlying dense-tensor.
-  // Memory allocation for the output may happen when this method is invoked,
-  // unless static optimization pre-allocates it.
-  SparseTensor* OutputSparse(int index, const TensorShape& shape);
-#endif
-
-#if !defined(DISABLE_OPTIONAL_TYPE)
-  // Use this API to output a "None" of a specific type (e.g. Tensor) at specified index
-  template <typename T>
-  void OutputOptionalWithoutData(int index) {
-    auto* output_ort_value = GetOutputMLValue(index);
-
-    auto type = DataTypeImpl::GetType<T>();
-
-    output_ort_value->Init(nullptr,  // This OrtValue is "None" and has no data
-                           type,
-                           type->GetDeleteFunc());
-  }
-#endif
-
-  // Retrieve indexed shape obtained from memory planning before actual
-  // computation. If the indexed shape cannot be inferred, this function returns
-  // false.
-  virtual bool TryGetInferredInputShape(int index, TensorShape& shape) const;
-
-  // Retrieve indexed shape obtained from memory planning before actual
-  // computation. If the indexed shape cannot be inferred, this function returns
-  // false.
-  virtual bool TryGetInferredOutputShape(int index, TensorShape& shape) const;
-
-  const logging::Logger& Logger() const {
-    return *logger_;
-  }
-
-  // always >= 0
-  virtual int InputCount() const {
-    return static_cast<int>(kernel_->Node().InputDefs().size());
-  }
-
-  // always >= 0
-  virtual int ImplicitInputCount() const {
-    return static_cast<int>(kernel_->Node().ImplicitInputDefs().size());
-  }
-
-  // always >= 0
-  virtual int OutputCount() const {
-    return static_cast<int>(kernel_->Node().OutputDefs().size());
-  }
-
-  /**
-   Return an allocator on device 0, with memtype of OrtMemTypeDefault.
-   @remarks Use SafeInt when calculating the size of memory to allocate using AllocatorPtr->Alloc.
-   */
-  [[nodiscard]] virtual Status GetTempSpaceAllocator(AllocatorPtr* output) const;
-
-  /**
-   Return the allocator associated with the CPU EP with memtype of OrtMemTypeDefault.
-   @remarks Use SafeInt when calculating the size of memory to allocate using AllocatorPtr->Alloc.
-   */
-  [[nodiscard]] Status GetTempSpaceCPUAllocator(AllocatorPtr* output) const;
-
-  /**
-  Return the device id that current kernel runs on.
-  */
-  virtual int GetDeviceId() const {
-    return kernel_->Info().GetExecutionProvider()->GetDeviceId();
-  }
-
-  /**
-  Return the compute stream associated with the EP that the kernel is partitioned to.
-  For EPs that do not have a compute stream (e.g. CPU EP), a nullptr is returned.
-  */
-  [[nodiscard]] virtual Stream* GetComputeStream() const {
-    return stream_;
-  }
-
-  /**
-  Returns the opset domain of the underlying kernel
-  **/
-  const std::string& GetOpDomain() const;
-
-  /**
-  Returns the optype of the underlying kernel
-  **/
-  const std::string& GetOpType() const;
-
-  /**
-  Returns the node name of the underlying kernel
-  **/
-  const std::string& GetNodeName() const;
-
-  /**
-  Returns the intra-op threadpool, if available.
-  */
-  _Ret_maybenull_ onnxruntime::concurrency::ThreadPool* GetOperatorThreadPool() const { return threadpool_; }
-
-  /**
-  Returns whether deterministic computation is preferred.
-  */
-  virtual bool GetUseDeterministicCompute() const {
-    return true;
-  }
-
- protected:
-  OpKernelContext(concurrency::ThreadPool* threadpool, const logging::Logger& logger, Stream* stream);
-
-  onnxruntime::NodeIndex GetNodeIndex() const;
-
-  virtual const OrtValue* GetInputMLValue(int index) const;
-  const OrtValue* GetImplicitInputMLValue(int index) const;
-  OrtValue* GetOutputMLValue(int index);
-
-#ifdef ENABLE_ATEN
-  Status SetOutputMLValue(int index, const OrtValue& ort_value);
-#endif
-
-  // Creates the OrtValue* based on the shape, if it does not exist
-  virtual OrtValue* OutputMLValue(int index, const TensorShape& shape);
-
-  virtual OrtValue* GetOrCreateOutputMLValue(int index);
-
- private:
-  ORT_DISALLOW_COPY_AND_ASSIGNMENT(OpKernelContext);
-  int GetInputArgIndex(int index) const;
-  int GetImplicitInputArgIndex(int index) const;
-  int GetOutputArgIndex(int index) const;
-
-  IExecutionFrame* const execution_frame_{};
-  const OpKernel* const kernel_{};
-  concurrency::ThreadPool* const threadpool_{};
-  const logging::Logger* const logger_{};
-
-  // The argument starting index in ExecutionFrame.
-  int node_input_start_index_{-1};
-  int node_implicit_input_start_index_{-1};
-  int node_output_start_index_{-1};
-
-  Stream* stream_;
-};
-
-// Fetching output tensor without shape is not allowed except when it already exists
-template <>
-inline Tensor* OpKernelContext::Output<Tensor>(int index) {
-  OrtValue* p_ml_value = GetOutputMLValue(index);
-  ORT_ENFORCE(p_ml_value, "Please fetch output tensor with specified shape.");
-  return p_ml_value->GetMutable<Tensor>();
-}
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-template <>
-inline SparseTensor* OpKernelContext::Output<SparseTensor>(int index) {
-  OrtValue* p_ml_value = GetOutputMLValue(index);
-  ORT_ENFORCE(p_ml_value, "Please fetch output sparse tensor with specified shape.");
-  return p_ml_value->GetMutable<SparseTensor>();
-}
-#endif
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/op_kernel_info.h b/projects/llm_framework/include/onnxruntime/core/framework/op_kernel_info.h
deleted file mode 100644
index d30f8fea..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/op_kernel_info.h
+++ /dev/null
@@ -1,63 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include "core/framework/execution_provider.h"
-#include "core/framework/kernel_def_builder.h"
-#include "core/framework/ort_value.h"
-#include "core/framework/op_node_proto_helper.h"
-#include "core/graph/graph_viewer.h"
-#include "core/common/gsl.h"
-
-namespace onnxruntime {
-
-class OrtValueNameIdxMap;
-class FuncManager;
-class DataTransferManager;
-struct AllocPlanPerValue;
-
-// A very light-weight class, which works as an aggregated
-// view of all data needed for constructing a Kernel instance.
-// NOTE: it does not own/hold any objects.
-class OpKernelInfo : public OpNodeProtoHelper<ProtoHelperNodeContext> {
- public:
-  explicit OpKernelInfo(const onnxruntime::Node& node,
-                        const KernelDef& kernel_def,
-                        const IExecutionProvider& execution_provider,
-                        const std::unordered_map<int, OrtValue>& constant_initialized_tensors,
-                        const OrtValueNameIdxMap& mlvalue_name_idx_map,
-                        const DataTransferManager& data_transfer_mgr);
-
-  OpKernelInfo(const OpKernelInfo& other);
-
-  const OrtMemoryInfo& GetMemoryInfo(int device_id, OrtMemType mem_type) const;
-
-  AllocatorPtr GetAllocator(int device_id, OrtMemType mem_type) const;
-
-  const KernelDef& GetKernelDef() const;
-
-  const IExecutionProvider* GetExecutionProvider() const noexcept;
-
-  const DataTransferManager& GetDataTransferManager() const noexcept;
-
-  const onnxruntime::Node& node() const noexcept;
-
-  bool TryGetConstantInput(int input_index, const Tensor** constant_input_value) const;
-
- private:
-  ORT_DISALLOW_MOVE(OpKernelInfo);
-  ORT_DISALLOW_ASSIGNMENT(OpKernelInfo);
-
-  const onnxruntime::Node& node_;
-  const KernelDef& kernel_def_;
-  // For non cpu/cuda case, this pointer should be set so that function kernel
-  // will delegate kernel compute call to <execution_provider> compute call.
-  gsl::not_null<const ::onnxruntime::IExecutionProvider*> execution_provider_;
-  const std::unordered_map<int, OrtValue>& constant_initialized_tensors_;
-  const OrtValueNameIdxMap& ort_value_name_idx_map_;
-  const DataTransferManager& data_transfer_mgr_;
-  ProtoHelperNodeContext proto_helper_context_;
-};
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/op_node_proto_helper.h b/projects/llm_framework/include/onnxruntime/core/framework/op_node_proto_helper.h
deleted file mode 100644
index 700e1edc..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/op_node_proto_helper.h
+++ /dev/null
@@ -1,167 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#ifndef SHARED_PROVIDER
-#include "core/common/status.h"
-#include "core/framework/tensor_shape.h"
-#include "core/graph/graph_viewer.h"
-#include "core/common/gsl.h"
-#endif
-
-#ifdef __has_attribute
-#define ORT_HAVE_ATTRIBUTE(x) __has_attribute(x)
-#else
-#define ORT_HAVE_ATTRIBUTE(x) 0
-#endif
-
-#if ORT_HAVE_ATTRIBUTE(nodiscard)
-#define MUST_USE_RESULT [[nodiscard]]
-#elif defined(__clang__) && ORT_HAVE_ATTRIBUTE(warn_unused_result)
-#define MUST_USE_RESULT __attribute__((warn_unused_result))
-#else
-#define MUST_USE_RESULT
-#endif
-
-class IMLOpKernel;
-
-namespace onnxruntime {
-
-/**
-   A set of wrappers with common signatures for use with both OpKernelInfo
-   (as its base class) and InferenceContext.  Used by ABI kernels for both
-   shape / type inference and kernel construction
-*/
-template <class Impl_t>
-class OpNodeProtoHelper {
- public:
-  explicit OpNodeProtoHelper(const Impl_t* impl) : impl_(impl) {}
-
-  /**
-     Get a single attribute
-     Call this function for a required attribute or when a default value for an optional attribute is specified in the op schema
-  */
-  template <typename T>
-  MUST_USE_RESULT Status GetAttr(const std::string& name, T* value) const;
-
-  /**
-     Get a single attribute
-     Call this function only when a default value for an optional attribute isn't specified in the op schema
-  */
-  template <typename T>
-  T GetAttrOrDefault(const std::string& name, const T& default_value) const {
-    T tmp;
-    return GetAttr<T>(name, &tmp).IsOK() ? tmp : default_value;
-  }
-
-  /**
-     Get a single attribute
-     Call this function only when a default value for an optional attribute isn't specified in the op schema
-  */
-  template <typename T>
-  void GetAttrOrDefault(const std::string& name, T* value, const T& default_value) const {
-    if (!GetAttr<T>(name, value).IsOK())
-      *value = default_value;
-  }
-
-  /**
-     Get repeated attributes
-     Call this function only when a default value for an optional attribute isn't specified in the op schema
-  */
-  template <typename T>
-  MUST_USE_RESULT std::vector<T> GetAttrsOrDefault(const std::string& name, const std::vector<T>& default_value = std::vector<T>{}) const {
-    std::vector<T> tmp;
-    return GetAttrs<T>(name, tmp).IsOK() ? tmp : default_value;
-  }
-
-  /// <summary>
-  /// Return a gsl::span that points to an array of primitive types held by AttributeProto
-  /// This function allows to avoid copying big attributes locally into a kernel and operate on
-  /// AttributeProto data directly.
-  ///
-  ///  Does not apply to strings, Tensors and Sparse Tensors that require special treatment.
-  /// </summary>
-  /// <typeparam name="T">Primitive type contained in the array</typeparam>
-  /// <param name="name">Attribute name</param>
-  /// <param name="values">Attribute data in a span, out parameter</param>
-  /// <returns>Status</returns>
-  template <typename T>
-  MUST_USE_RESULT Status GetAttrsAsSpan(const std::string& name, gsl::span<const T>& values) const;
-
-  MUST_USE_RESULT Status GetAttrs(const std::string& name, TensorShapeVector& out) const;
-
-  MUST_USE_RESULT TensorShapeVector GetAttrsOrDefault(const std::string& name, const TensorShapeVector& default_value = TensorShapeVector{}) const {
-    TensorShapeVector tmp;
-    return GetAttrs(name, tmp).IsOK() ? tmp : default_value;
-  }
-
-  /**
-     Get repeated attributes
-  */
-  template <typename T>
-  MUST_USE_RESULT Status GetAttrs(const std::string& name, std::vector<T>& values) const;
-
-  template <typename T>
-  MUST_USE_RESULT Status GetAttrs(const std::string& name, gsl::span<T> values) const;
-
-  MUST_USE_RESULT Status GetAttrsStringRefs(const std::string& name,
-                                            std::vector<std::reference_wrapper<const std::string>>& refs) const;
-
-  uint32_t GetPrimitiveAttrElementCount(ONNX_NAMESPACE::AttributeProto_AttributeType type,
-                                        const std::string& name) const noexcept;
-
-  bool HasPrimitiveAttribute(ONNX_NAMESPACE::AttributeProto_AttributeType type,
-                             const std::string& name) const noexcept;
-
-  uint32_t GetInputCount() const {
-    return gsl::narrow_cast<uint32_t>(impl_->getNumInputs());
-  }
-
-  uint32_t GetOutputCount() const {
-    return gsl::narrow_cast<uint32_t>(impl_->getNumOutputs());
-  }
-
-  const ONNX_NAMESPACE::TypeProto* GetInputType(size_t index) const {
-    return impl_->getInputType(index);
-  }
-
-  const ONNX_NAMESPACE::TypeProto* GetOutputType(size_t index) const {
-    // Work around lack of a const method from the onnx InferenceContext interface
-    return const_cast<Impl_t*>(impl_)->getOutputType(index);
-  }
-
-  // Try to query an attribute, returning nullptr if it doesn't exist
-  const ONNX_NAMESPACE::AttributeProto* TryGetAttribute(const std::string& name) const {
-    return impl_->getAttribute(name);
-  }
-
-  const ONNX_NAMESPACE::AttributeProto* GetAttribute(const std::string& name) const {
-    const ONNX_NAMESPACE::AttributeProto* attr = TryGetAttribute(name);
-    ORT_ENFORCE(attr != nullptr);
-    return attr;
-  }
-
- private:
-  OpNodeProtoHelper() = delete;
-  const Impl_t* impl_ = nullptr;
-};
-
-// The methods on the following class are called by OpNodeProtoHelper, implementing
-// the same signatures as InferenceContext other than const-ness.
-class ProtoHelperNodeContext {
- public:
-  explicit ProtoHelperNodeContext(const onnxruntime::Node& node) : node_(node) {}
-  ProtoHelperNodeContext() = delete;
-
-  const ONNX_NAMESPACE::AttributeProto* getAttribute(const std::string& name) const;
-  size_t getNumInputs() const;
-  const ONNX_NAMESPACE::TypeProto* getInputType(size_t index) const;
-  size_t getNumOutputs() const;
-  const ONNX_NAMESPACE::TypeProto* getOutputType(size_t index) const;
-
- private:
-  const onnxruntime::Node& node_;
-};
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/ort_value.h b/projects/llm_framework/include/onnxruntime/core/framework/ort_value.h
deleted file mode 100644
index 7cbbdcc9..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/ort_value.h
+++ /dev/null
@@ -1,123 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <string>
-#ifndef SHARED_PROVIDER
-#include "core/common/common.h"
-#include "core/common/exceptions.h"
-#include "core/framework/allocator.h"
-#include "core/framework/data_types.h"
-#include "core/framework/tensor.h"
-#include "core/framework/TensorSeq.h"
-
-namespace onnxruntime {
-#if !defined(DISABLE_SPARSE_TENSORS)
-class SparseTensor;
-#endif
-}  // namespace onnxruntime
-
-#endif
-
-/**
-   Represents both tensors and non-tensors.
-*/
-struct OrtValue {
- public:
-  OrtValue() : data_(nullptr) {}
-  ~OrtValue() = default;
-
-  OrtValue(void* pData, onnxruntime::MLDataType type, onnxruntime::DeleteFunc deleter) {
-    Init(pData, type, deleter);
-  }
-
-  void Init(void* pData, onnxruntime::MLDataType type, onnxruntime::DeleteFunc deleter) {
-    data_.reset(pData, deleter);
-    type_ = type;
-  }
-
-  void Init(void* pData, onnxruntime::MLDataType type, const std::function<void(void*)>& deleter) {
-    data_.reset(pData, deleter);
-    type_ = type;
-  }
-
-  bool IsAllocated() const {
-    return data_ && type_;
-  }
-
-  template <typename T>
-  const T& Get() const {
-    ORT_ENFORCE(onnxruntime::DataTypeImpl::GetType<T>() == type_, onnxruntime::DataTypeImpl::GetType<T>(), " != ", type_);
-    return *static_cast<T*>(data_.get());
-  }
-
-  // May return nullptr, if this OrtValue is an optional type and it is "None".
-  template <typename T>
-  T* GetMutable() {
-    ORT_ENFORCE(onnxruntime::DataTypeImpl::GetType<T>() == type_, onnxruntime::DataTypeImpl::GetType<T>(), " != ", type_);
-    return static_cast<T*>(data_.get());
-  }
-
-  bool IsTensor() const noexcept {
-    return (type_ != nullptr && type_->IsTensorType());
-  }
-
-  bool IsTensorSequence() const noexcept {
-    return (type_ != nullptr && type_->IsTensorSequenceType());
-  }
-
-  bool IsSparseTensor() const {
-#if !defined(DISABLE_SPARSE_TENSORS)
-    return (type_ != nullptr && type_->IsSparseTensorType());
-#else
-    ORT_THROW("Sparse tensor is not supported in this build.");
-#endif
-  }
-
-  onnxruntime::MLDataType Type() const {
-    return type_;
-  }
-
- private:
-  std::shared_ptr<void> data_;
-  onnxruntime::MLDataType type_{nullptr};
-};
-
-template <>
-inline const onnxruntime::Tensor& OrtValue::Get<onnxruntime::Tensor>() const {
-  ORT_ENFORCE(IsTensor(), "Trying to get a Tensor, but got: ", onnxruntime::DataTypeImpl::ToString(type_));
-  return *static_cast<onnxruntime::Tensor*>(data_.get());
-}
-
-template <>
-inline onnxruntime::Tensor* OrtValue::GetMutable<onnxruntime::Tensor>() {
-  ORT_ENFORCE(IsTensor(), "Trying to get a Tensor, but got: ", onnxruntime::DataTypeImpl::ToString(type_));
-  return static_cast<onnxruntime::Tensor*>(data_.get());
-}
-
-template <>
-inline const onnxruntime::TensorSeq& OrtValue::Get<onnxruntime::TensorSeq>() const {
-  ORT_ENFORCE(IsTensorSequence(), "Trying to get a TensorSeq, but got: ", onnxruntime::DataTypeImpl::ToString(type_));
-  return *static_cast<onnxruntime::TensorSeq*>(data_.get());
-}
-
-template <>
-inline onnxruntime::TensorSeq* OrtValue::GetMutable<onnxruntime::TensorSeq>() {
-  ORT_ENFORCE(IsTensorSequence(), "Trying to get a TensorSeq, but got: ", onnxruntime::DataTypeImpl::ToString(type_));
-  return static_cast<onnxruntime::TensorSeq*>(data_.get());
-}
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-template <>
-inline const onnxruntime::SparseTensor& OrtValue::Get<onnxruntime::SparseTensor>() const {
-  ORT_ENFORCE(IsSparseTensor(), "Trying to get a SparseTensor, but got: ", onnxruntime::DataTypeImpl::ToString(type_));
-  return *static_cast<onnxruntime::SparseTensor*>(data_.get());
-}
-
-template <>
-inline onnxruntime::SparseTensor* OrtValue::GetMutable<onnxruntime::SparseTensor>() {
-  ORT_ENFORCE(IsSparseTensor(), "Trying to get a SparseTensor, but got: ", onnxruntime::DataTypeImpl::ToString(type_));
-  return static_cast<onnxruntime::SparseTensor*>(data_.get());
-}
-#endif
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/ortdevice.h b/projects/llm_framework/include/onnxruntime/core/framework/ortdevice.h
deleted file mode 100644
index 77f7c3e1..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/ortdevice.h
+++ /dev/null
@@ -1,74 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <sstream>
-
-// Struct to represent a physical device.
-struct OrtDevice {
-  using DeviceType = int8_t;
-  using MemoryType = int8_t;
-  using DeviceId = int16_t;
-
-  // Pre-defined device types.
-  static const DeviceType CPU = 0;
-  static const DeviceType GPU = 1;  // Nvidia or AMD
-  static const DeviceType FPGA = 2;
-  static const DeviceType NPU = 3;  // Ascend
-
-  struct MemType {
-    // Pre-defined memory types.
-    static const MemoryType DEFAULT = 0;
-    static const MemoryType CUDA_PINNED = 1;
-    static const MemoryType HIP_PINNED = 2;
-    static const MemoryType CANN_PINNED = 3;
-  };
-
-  constexpr OrtDevice(DeviceType device_type_, MemoryType memory_type_, DeviceId device_id_)
-      : device_type(device_type_),
-        memory_type(memory_type_),
-        device_id(device_id_) {}
-
-  constexpr OrtDevice() : OrtDevice(CPU, MemType::DEFAULT, 0) {}
-
-  DeviceType Type() const {
-    return device_type;
-  }
-
-  MemoryType MemType() const {
-    return memory_type;
-  }
-
-  DeviceId Id() const {
-    return device_id;
-  }
-
-  std::string ToString() const {
-    std::ostringstream ostr;
-    ostr << "Device:["
-         << "DeviceType:" << static_cast<int>(device_type)
-         << " MemoryType:" << static_cast<int>(memory_type)
-         << " DeviceId:" << device_id
-         << "]";
-    return ostr.str();
-  }
-
- private:
-  // Device type.
-  DeviceType device_type;
-
-  // Memory type.
-  MemoryType memory_type;
-
-  // Device index.
-  DeviceId device_id;
-};
-
-inline bool operator==(const OrtDevice& left, const OrtDevice& other) {
-  return left.Id() == other.Id() && left.MemType() == other.MemType() && left.Type() == other.Type();
-}
-
-inline bool operator!=(const OrtDevice& left, const OrtDevice& other) {
-  return !(left == other);
-}
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/ortmemoryinfo.h b/projects/llm_framework/include/onnxruntime/core/framework/ortmemoryinfo.h
deleted file mode 100644
index 6dd259dd..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/ortmemoryinfo.h
+++ /dev/null
@@ -1,87 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <string_view>
-
-#include "core/common/hash_combine.h"
-
-struct OrtMemoryInfo {
-  OrtMemoryInfo() = default;  // to allow default construction of Tensor
-
-  // use string for name, so we could have customized allocator in execution provider.
-  const char* name = nullptr;
-  int id = -1;
-  OrtMemType mem_type = OrtMemTypeDefault;
-  OrtAllocatorType alloc_type = OrtInvalidAllocator;
-  OrtDevice device;
-
-  constexpr OrtMemoryInfo(const char* name_, OrtAllocatorType type_, OrtDevice device_ = OrtDevice(), int id_ = 0,
-                          OrtMemType mem_type_ = OrtMemTypeDefault)
-#if ((defined(__GNUC__) && __GNUC__ > 4) || defined(__clang__))
-      // this causes a spurious error in CentOS gcc 4.8 build so disable if GCC version < 5
-      __attribute__((nonnull))
-#endif
-      : name(name_),
-        id(id_),
-        mem_type(mem_type_),
-        alloc_type(type_),
-        device(device_) {
-  }
-
-  // To make OrtMemoryInfo become a valid key in std map
-  bool operator<(const OrtMemoryInfo& other) const {
-    if (alloc_type != other.alloc_type)
-      return alloc_type < other.alloc_type;
-    if (mem_type != other.mem_type)
-      return mem_type < other.mem_type;
-    if (id != other.id)
-      return id < other.id;
-
-    return strcmp(name, other.name) < 0;
-  }
-
-  // This is to make OrtMemoryInfo a valid key in hash tables
-  // we ignore device id
-  size_t Hash() const {
-    auto h = std::hash<int>()(alloc_type);
-    onnxruntime::HashCombine(mem_type, h);
-    onnxruntime::HashCombine(id, h);
-    onnxruntime::HashCombine<std::string_view>(name, h);
-    return h;
-  }
-
-  std::string ToString() const {
-    std::ostringstream ostr;
-    ostr << "OrtMemoryInfo:["
-         << "name:" << name
-         << " id:" << id
-         << " OrtMemType:" << mem_type
-         << " OrtAllocatorType:" << alloc_type
-         << " " << device.ToString()
-         << "]";
-    return ostr.str();
-  }
-};
-
-// Required by hash tables
-inline bool operator==(const OrtMemoryInfo& left, const OrtMemoryInfo& other) {
-  return left.mem_type == other.mem_type &&
-         left.alloc_type == other.alloc_type &&
-         left.id == other.id &&
-         strcmp(left.name, other.name) == 0;
-}
-
-inline bool operator!=(const OrtMemoryInfo& lhs, const OrtMemoryInfo& rhs) { return !(lhs == rhs); }
-
-std::ostream& operator<<(std::ostream& out, const OrtMemoryInfo& info);
-
-namespace std {
-template<>
-struct hash<OrtMemoryInfo> {
-  size_t operator()(const OrtMemoryInfo& i) const {
-    return i.Hash();
-  }
-};
-}
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/provider_options.h b/projects/llm_framework/include/onnxruntime/core/framework/provider_options.h
deleted file mode 100644
index aab13e80..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/provider_options.h
+++ /dev/null
@@ -1,18 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-namespace onnxruntime {
-
-// data types for execution provider options
-
-using ProviderOptions = std::unordered_map<std::string, std::string>;
-using ProviderOptionsVector = std::vector<ProviderOptions>;
-using ProviderOptionsMap = std::unordered_map<std::string, ProviderOptions>;
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/provider_options_utils.h b/projects/llm_framework/include/onnxruntime/core/framework/provider_options_utils.h
deleted file mode 100644
index 5967fb91..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/provider_options_utils.h
+++ /dev/null
@@ -1,164 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <algorithm>
-#include <functional>
-#include <type_traits>
-#include <unordered_map>
-#include <vector>
-
-#include "core/common/common.h"
-#include "core/common/parse_string.h"
-#include "core/framework/provider_options.h"
-
-namespace onnxruntime {
-
-template <typename TEnum>
-using EnumNameMapping = std::vector<std::pair<TEnum, std::string>>;
-
-/**
- * Given a mapping and an enumeration value, gets the corresponding name.
- */
-template <typename TEnum>
-Status EnumToName(const EnumNameMapping<TEnum>& mapping, TEnum value, std::string& name) {
-  const auto it = std::find_if(
-      mapping.begin(), mapping.end(),
-      [&value](const std::pair<TEnum, std::string>& entry) {
-        return entry.first == value;
-      });
-  ORT_RETURN_IF(
-      it == mapping.end(),
-      "Failed to map enum value to name: ", static_cast<typename std::underlying_type<TEnum>::type>(value));
-  name = it->second;
-  return Status::OK();
-}
-
-template <typename TEnum>
-std::string EnumToName(const EnumNameMapping<TEnum>& mapping, TEnum value) {
-  std::string name;
-  ORT_THROW_IF_ERROR(EnumToName(mapping, value, name));
-  return name;
-}
-
-/**
- * Given a mapping and a name, gets the corresponding enumeration value.
- */
-template <typename TEnum>
-Status NameToEnum(
-    const EnumNameMapping<TEnum>& mapping, const std::string& name, TEnum& value) {
-  const auto it = std::find_if(
-      mapping.begin(), mapping.end(),
-      [&name](const std::pair<TEnum, std::string>& entry) {
-        return entry.second == name;
-      });
-  ORT_RETURN_IF(
-      it == mapping.end(),
-      "Failed to map enum name to value: ", name);
-  value = it->first;
-  return Status::OK();
-}
-
-template <typename TEnum>
-TEnum NameToEnum(const EnumNameMapping<TEnum>& mapping, const std::string& name) {
-  TEnum value;
-  ORT_THROW_IF_ERROR(NameToEnum(mapping, name, value));
-  return value;
-}
-
-class ProviderOptionsParser {
- public:
-  /**
-   * Adds a parser for a particular provider option value.
-   *
-   * @param name The provider option name.
-   * @param value_parser An object that parses the option value.
-   *        It should be callable with the following signature and return
-   *        whether the parsing was successful:
-   *            Status value_parser(const std::string&)
-   *
-   * @return The current ProviderOptionsParser instance.
-   */
-  template <typename ValueParserType>
-  ProviderOptionsParser& AddValueParser(
-      const std::string& name, ValueParserType value_parser) {
-    ORT_ENFORCE(
-        value_parsers_.emplace(name, ValueParser{value_parser}).second,
-        "Provider option \"", name, "\" already has a value parser.");
-    return *this;
-  }
-
-  /**
-   * Adds a parser for a particular provider option value which converts a
-   * value to the right type and assigns it to the given reference.
-   *
-   * IMPORTANT: This function stores a reference to the destination variable.
-   * The caller must ensure that the reference is valid when Parse() is called!
-   *
-   * @param name The provider option name.
-   * @param dest The destination variable reference.
-   *
-   * @return The current ProviderOptionsParser instance.
-   */
-  template <typename ValueType>
-  ProviderOptionsParser& AddAssignmentToReference(
-      const std::string& name, ValueType& dest) {
-    return AddValueParser(
-        name,
-        [&dest](const std::string& value_str) -> Status {
-          return ParseStringWithClassicLocale(value_str, dest);
-        });
-  }
-
-  /**
-   * Adds a parser for a particular provider option value which maps an
-   * enumeration name to a value and assigns it to the given reference.
-   *
-   * IMPORTANT: This function stores references to the mapping and destination
-   * variables. The caller must ensure that the references are valid when
-   * Parse() is called!
-   *
-   * @param name The provider option name.
-   * @param mapping The enumeration value to name mapping.
-   * @param dest The destination variable reference.
-   *
-   * @return The current ProviderOptionsParser instance.
-   */
-  template <typename EnumType>
-  ProviderOptionsParser& AddAssignmentToEnumReference(
-      const std::string& name, const EnumNameMapping<EnumType>& mapping, EnumType& dest) {
-    return AddValueParser(
-        name,
-        [&mapping, &dest](const std::string& value_str) -> Status {
-          return NameToEnum(mapping, value_str, dest);
-        });
-  }
-
-  /**
-   * Parses the given provider options.
-   */
-  Status Parse(const ProviderOptions& options) const {
-    for (const auto& option : options) {
-      const auto& name = option.first;
-      const auto& value_str = option.second;
-      const auto value_parser_it = value_parsers_.find(name);
-      ORT_RETURN_IF(
-          value_parser_it == value_parsers_.end(),
-          "Unknown provider option: \"", name, "\".");
-
-      const auto parse_status = value_parser_it->second(value_str);
-      ORT_RETURN_IF_NOT(
-          parse_status.IsOK(),
-          "Failed to parse provider option \"", name, "\": ", parse_status.ErrorMessage());
-    }
-
-    return Status::OK();
-  }
-
- private:
-  using ValueParser = std::function<Status(const std::string&)>;
-  std::unordered_map<std::string, ValueParser> value_parsers_;
-};
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/provider_shutdown.h b/projects/llm_framework/include/onnxruntime/core/framework/provider_shutdown.h
deleted file mode 100644
index 7488f12a..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/provider_shutdown.h
+++ /dev/null
@@ -1,8 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-namespace onnxruntime {
-void UnloadSharedProviders();
-}
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/run_options.h b/projects/llm_framework/include/onnxruntime/core/framework/run_options.h
deleted file mode 100644
index 5444c825..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/run_options.h
+++ /dev/null
@@ -1,49 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <string>
-#include <atomic>
-#include "core/session/onnxruntime_c_api.h"
-#include "core/framework/config_options.h"
-
-/**
- * Configuration information for a Run call.
- */
-struct OrtRunOptions {
-  /// Log severity.  See https://github.com/microsoft/onnxruntime/blob/main/include/onnxruntime/core/common/logging/severity.h
-  /// Default = -1 (use the log severity from the InferenceSession that the Run is for).
-  int run_log_severity_level = -1;
-  int run_log_verbosity_level = 0;  ///< VLOG level if debug build and run_log_severity_level is 0 (VERBOSE).
-  std::string run_tag;              ///< A tag for the Run() calls using this.
-
-  // Set to 'true' to ensure the termination of all the outstanding Run() calls
-  // that use this OrtRunOptions instance. Some of the outstanding Run() calls may
-  // be forced to terminate with an error status.
-  bool terminate = false;
-
-  // Set to 'true' to run only the nodes from feeds to required fetches.
-  // So it is possible that only some of the nodes are executed.
-  bool only_execute_path_to_fetches = false;
-
-#ifdef ENABLE_TRAINING
-  // Used by onnxruntime::training::TrainingSession. This class is now deprecated.
-  // Delete training_mode when TrainingSession is deleted.
-  // Set to 'true' to run in training mode.
-  bool training_mode = true;
-#endif
-
-  // Stores the configurations for this run
-  // To add an configuration to this specific run, call OrtApis::AddRunConfigEntry
-  // The configuration keys and value formats are defined in
-  // /include/onnxruntime/core/session/onnxruntime_run_options_config_keys.h
-  onnxruntime::ConfigOptions config_options;
-
-  OrtRunOptions() = default;
-  ~OrtRunOptions() = default;
-};
-
-namespace onnxruntime {
-using RunOptions = OrtRunOptions;
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/sparse_tensor.h b/projects/llm_framework/include/onnxruntime/core/framework/sparse_tensor.h
deleted file mode 100644
index 161c3912..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/sparse_tensor.h
+++ /dev/null
@@ -1,534 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-
-#pragma once
-
-#include "core/framework/data_types.h"
-#include "core/framework/tensor_shape.h"
-#include "core/framework/tensor.h"
-
-struct OrtValue;
-
-namespace onnxruntime {
-
-class IDataTransfer;
-class DataTransferManager;
-
-/**
- * @brief This is a Sparse Format enumeration
- * 
- * 
- */
-enum class SparseFormat : uint32_t {
-  kUndefined = 0x0U,        // For completeness
-  kCoo = 0x1U,              // 1-D or 2-D indices
-  kCsrc = 0x1U << 1,        // Both CSR(C)
-  kBlockSparse = 0x1U << 2  // as in OpenAI
-};
-
-std::ostream& operator<<(std::ostream&, SparseFormat);
-
-/**
- * @brief This class implements SparseTensor. 
- * This class holds sparse non-zero data (values) and sparse format
- * specific indices. There are two main uses for the class (similar to that of Tensor)
- * - one is to re-present model sparse inputs. Such inputs typically reside
- *   in user allocated buffers that are not owned by SparseTensor instance and the instance
- *   serves as a facade to expose user allocated buffers. Such buffers should already
- *   contain proper values and format specific indices.
- *   Use the first constructor
- *   to instantiate SparseTensor and supply values_data pointer. Use*() functions can
- *   be used to supply pointers to format specific indices. These buffers are used as is
- *   and will not be modified or deallocated by the instance. However, the lifespan of the buffers
- *   must eclipse the lifespan of the SparseTensor instance.
- * 
- * - Represent sparse data that is a result of format conversion or a computation result. Use second constructor
- *   to supply a desired allocator. Use Make*() format specific interfaces to supply values and format
- *   specific indices. The specified data will be copied into an internally allocated buffer.
-     Internally, we will represent a SparseTensor as a single contiguous buffer that
- *   contains values followed by format specific indices. We use Tensors to project
- *   values and indices into various parts of buffer.
- */
-
-class SparseTensor final {
- public:
-  /// <summary>
-  /// This constructs an instance that points to user defined buffers.
-  /// Make use of Use* functions to supply format specific indices that
-  /// reside in the user supplied buffers. The instance constructed this way
-  /// will not copy data. The lifespan of supplied buffers is expected to eclipse
-  ///  the lifespan of the sparse tensor instance.
-  /// </summary>
-  /// <param name="elt_type">MlDataType</param>
-  /// <param name="dense_shape">a shape of original tensor in dense form</param>
-  /// <param name="values_shape">shape for user supplied values. Use {0} shape for fully sparse tensors.</param>
-  /// <param name="values_data">a pointer to values. Use nullptr for fully sparse tensors.</param>
-  /// <param name="location">description of the user allocated memory</param>
-  SparseTensor(MLDataType elt_type,
-               const TensorShape& dense_shape,
-               const TensorShape& values_shape,
-               void* values_data,
-               const OrtMemoryInfo& location);
-
-  /// <summary>
-  /// Use this constructor to hold sparse data in the buffer
-  /// allocated with the specified allocator. Use Make*() methods
-  /// to populate the instance with data which will be copied into the
-  /// allocated buffer.
-  /// </summary>
-  /// <param name="elt_type"></param>
-  /// <param name="dense_shape"></param>
-  /// <param name="allocator"></param>
-  SparseTensor(MLDataType elt_type,
-               const TensorShape& dense_shape,
-               std::shared_ptr<IAllocator> allocator);
-
-  SparseTensor() noexcept;
-
-  ~SparseTensor();
-
-  ORT_DISALLOW_COPY_AND_ASSIGNMENT(SparseTensor);
-
-  /// <summary>
-  /// The factory function creates an instance of SparseTensor on the heap
-  /// using appropriate constructor and initializes OrtValue instance wit it.
-  /// </summary>
-  /// <param name="elt_type">element data type</param>
-  /// <param name="dense_shape">dense shape of the sparse tensor</param>
-  /// <param name="values_shape">values shape. Use {0} for fully sparse tensors.</param>
-  /// <param name="values_data">pointer to a user allocated buffer. Use nullptr for fully sparse tensors.</param>
-  /// <param name="location">description of the user allocated buffer</param>
-  /// <param name="ort_value">default constructed input/output ort_value</param>
-  static void InitOrtValue(MLDataType elt_type,
-                           const TensorShape& dense_shape,
-                           const TensorShape& values_shape,
-                           void* values_data,
-                           const OrtMemoryInfo& location,
-                           OrtValue& ort_value);
-
-  /// <summary>
-  /// The factory function creates an instance of SparseTensor on the heap
-  /// using appropriate constructor and initializes OrtValue instance wit it.
-  /// </summary>
-  /// <param name="elt_type">element data type</param>
-  /// <param name="dense_shape">dense shape of the sparse tensor</param>
-  /// <param name="allocator">allocator to use</param>
-  /// <param name="ort_value">default constructed input/output ort_value</param>
-  static void InitOrtValue(MLDataType elt_type,
-                           const TensorShape& dense_shape,
-                           std::shared_ptr<IAllocator> allocator,
-                           OrtValue& ort_value);
-
-  /// <summary>
-  /// The function will check if the OrtValue is allocated
-  /// fetch the containing SparseTensor instance or throw if it
-  /// does not contain one. It will check that the SparseTensor has
-  /// sparse format set (i.e. fully constructed).
-  /// </summary>
-  /// <param name="v">OrtValue instance</param>
-  /// <returns>const SparseTensor Reference</returns>
-  static const SparseTensor& GetSparseTensorFromOrtValue(const OrtValue& v);
-
-  /// <summary>
-  /// /// The function will check if the OrtValue is allocated
-  /// fetch the containing SparseTensor instance or throw if it
-  /// does not contain one. It will check that the SparseTensor does not
-  /// have sparse format set and will return non-const ref to so indices
-  /// can be added to it.
-  /// </summary>
-  /// <param name="v">OrtValue</param>
-  /// <returns>non-const reference to SparseTensor</returns>
-  static SparseTensor& GetSparseTensorFromOrtValue(OrtValue& v);
-
-  /// <summary>
-  // Returns the number of non-zero values (aka "NNZ")
-  // For block sparse formats this may include some zeros in the blocks
-  // are considered non-zero.
-  /// </summary>
-  /// <returns>nnz</returns>
-  size_t NumValues() const { return static_cast<size_t>(values_.Shape().Size()); }
-
-  /// <summary>
-  /// Read only accessor to non-zero values
-  /// </summary>
-  /// <returns></returns>
-  const Tensor& Values() const noexcept {
-    return values_;
-  }
-
-  SparseTensor(SparseTensor&& o) noexcept;
-  SparseTensor& operator=(SparseTensor&& o) noexcept;
-
-  /// <summary>
-  /// Returns SparseFormat that the instance currently holds
-  /// if the value returned in kUndefined, the instance is not populated
-  /// </summary>
-  /// <returns>format enum</returns>
-  SparseFormat Format() const noexcept {
-    return format_;
-  }
-
-  /// <summary>
-  /// Returns a would be dense_shape
-  /// </summary>
-  /// <returns>reference to dense_shape</returns>
-  const TensorShape& DenseShape() const noexcept {
-    return dense_shape_;
-  }
-
-  /// <summary>
-  /// Calculates and returns how much this fully initialized SparseTensor data (would)
-  /// occupy in a contiguous allocation block, or, in fact, occupies if it owns the buffer.
-  /// </summary>
-  /// <returns>required allocation size</returns>
-  int64_t RequiredAllocationSize() const noexcept;
-
-  /// <summary>
-  /// Returns Tensor element type enum.
-  /// Useful for type dispatching
-  /// </summary>
-  /// <returns></returns>
-  int32_t GetElementType() const {
-    return ml_data_type_->GetDataType();
-  }
-
-  /// <summary>
-  /// Return Element MLDataType
-  /// </summary>
-  /// <returns></returns>
-  MLDataType DataType() const noexcept {
-    return ml_data_type_;
-  }
-
-  /// <summary>
-  /// Test for string type
-  /// </summary>
-  /// <returns>true if tensor values are strings</returns>
-  bool IsDataTypeString() const {
-    return utils::IsPrimitiveDataType<std::string>(ml_data_type_);
-  }
-
-  /// <summary>
-  /// Checks if the Tensor contains data type T
-  /// </summary>
-  /// <typeparam name="T"></typeparam>
-  /// <returns>true if tensor contains data of type T</returns>
-  template <class T>
-  bool IsDataType() const {
-    return utils::IsPrimitiveDataType<T>(ml_data_type_);
-  }
-
-  const OrtMemoryInfo& Location() const noexcept { return location_; }
-
-  /// <summary>
-  /// Read only access to Coo indices
-  /// </summary>
-  class CooView {
-   public:
-    explicit CooView(const Tensor& indices) noexcept
-        : indices_(indices) {}
-    const Tensor& Indices() const noexcept { return indices_; }
-
-   private:
-    std::reference_wrapper<const Tensor> indices_;
-  };
-
-  /// <summary>
-  /// Returns Coo index view
-  /// </summary>
-  /// <returns>CooView instance</returns>
-  CooView AsCoo() const;
-
-  /// <summary>
-  /// Uses COO index contained in the user allocated buffer along with the values buffer passed on
-  /// to the constructor. The buffer is used as is and its lifespan must eclipse the lifespan of the sparse
-  /// tensor instance. The OrtMemoryInfo (location) of the index is assumed to be the same as values.
-  ///
-  /// The index size must either exactly match the number of values in which case
-  /// index shape would be 1-D (values_count) or it must be twice the number of values
-  /// in which case its shape would be 2-D (values_count, 2)
-  /// </summary>
-  /// <param name="indices">user allocated buffer span. Use empty span for fully sparse tensors.</param>
-  /// <returns>Status</returns>
-  Status UseCooIndices(gsl::span<int64_t> indices);
-
-  /// <summary>
-  /// The method allocates a single contiguous buffer and copies specified values
-  /// and indices into it using supplied IDataTransfer.
-  ///
-  /// The indices size must either exactly match the number of values in which case
-  /// indices shape would be 1-D (values_count) or it must be twice the number of values
-  /// in which case its shape would be 2-D (values_count, 2).
-  ///
-  /// Values shape is supplied at construction time and its Size() must match values_count.
-  /// </summary>
-  /// <param name="values_count">Use 0 for fully sparse tensors.</param>
-  /// <param name="values_data">pointer to a buffer to be copied. Use nullptr for fully sparse tensors.</param>
-  /// <param name="indices"></param>
-  /// <returns></returns>
-  Status MakeCooData(const IDataTransfer& data_transfer, const OrtMemoryInfo& data_location,
-                     size_t values_count, const void* values_data, gsl::span<const int64_t> indices);
-
-  /// <summary>
-  /// The method allocates a single contiguous buffer and creates instances of std::strings in it, with
-  /// copies of the supplied zero-terminated strings followed by COO indices.
-  /// All data is assumed to be on CPU and the allocator supplied must be
-  /// a CPU based allocator.
-  /// </summary>
-  /// <param name="string_count">use 0 for fully sparse tensors</param>
-  /// <param name="strings">array of char* pointers. use nullptr for fully sparse tensors</param>
-  /// <param name="indices">span of indices. Use empty span for fully sparse tensors.</param>
-  /// <returns>Status</returns>
-  Status MakeCooStrings(size_t string_count, const char* const* strings, gsl::span<const int64_t> indices);
-
-  /// <summary>
-  /// Gives mutable access to Coo buffers so they can be populated
-  /// </summary>
-  class CooMutator {
-   public:
-    CooMutator(Tensor& values, Tensor& indices) noexcept : values_(values), indices_(indices) {}
-    Tensor& Values() noexcept { return values_; }
-    Tensor& Indices() noexcept { return indices_; }
-
-   private:
-    std::reference_wrapper<Tensor> values_;
-    std::reference_wrapper<Tensor> indices_;
-  };
-
-  /// <summary>
-  /// Allocates memory for values and index and returns a mutator so
-  /// data can be copied into the buffer.
-  /// </summary>
-  /// <param name="values_count">use 0 for fully sparse tensors</param>
-  /// <param name="index_count">use 0 for fully sparse tensors</param>
-  /// <returns></returns>
-  CooMutator MakeCooData(size_t values_count, size_t index_count);
-
-  /// <summary>
-  /// Read only access to Csr indices
-  /// </summary>
-  class CsrView {
-   public:
-    CsrView(const Tensor& inner, const Tensor& outer) noexcept
-        : inner_(inner), outer_(outer) {}
-    const Tensor& Inner() const noexcept { return inner_; }
-    const Tensor& Outer() const noexcept { return outer_; }
-
-   private:
-    std::reference_wrapper<const Tensor> inner_;
-    std::reference_wrapper<const Tensor> outer_;
-  };
-
-  /// <summary>
-  /// Returns Csr indices read only view
-  /// </summary>
-  /// <returns></returns>
-  CsrView AsCsr() const;
-
-  /// <summary>
-  /// This function will use Csr indices contained within the user allocated buffers.
-  /// The lifespan of the buffers must eclipse the lifespan of sparse tensor instance.
-  /// </summary>
-  /// <param name="inner_index">User allocated buffer span. use empty span for fully sparse tensors</param>
-  /// <param name="outer_index">User allocated buffer span. Use empty span for fully sparse tensors</param>
-  /// <returns></returns>
-  Status UseCsrIndices(gsl::span<int64_t> inner_index, gsl::span<int64_t> outer_index);
-
-  /// <summary>
-  /// The function will allocate a single contiguous buffer and will copy values
-  /// and indices into it.
-  /// </summary>
-  /// <param name="data_transfer"></param>
-  /// <param name="data_location"></param>
-  /// <param name="values_count">use 0 for fully sparse tensors</param>
-  /// <param name="values_data">pointer to data to be copied. Use nullptr for fully sparse tensors.</param>
-  /// <param name="inner_index">inner index to be copied. Use empty span for fully sparse tensors.</param>
-  /// <param name="outer_index">outer index to be copied. Use empty span for fully sparse tensors.</param>
-  /// <returns></returns>
-  Status MakeCsrData(const IDataTransfer& data_transfer,
-                     const OrtMemoryInfo& data_location,
-                     size_t values_count, const void* values_data,
-                     gsl::span<const int64_t> inner_index,
-                     gsl::span<const int64_t> outer_index);
-
-  /// <summary>
-  /// The method allocates a single contiguous buffer and creates instances of std::strings in it, with
-  /// copies of the supplied zero-terminated strings followed by COO indices.
-  /// All data is assumed to be on CPU and the allocator supplied must be
-  /// a CPU based allocator
-  /// </summary>
-  /// <param name="string_count"></param>
-  /// <param name="strings">array of char* pointers</param>
-  /// <param name="inner_index">inner index to be copied. Use empty span for fully sparse tensors.</param>
-  /// <param name="outer_index">outer index to be copied. Use empty span for fully sparse tensors.</param>
-  /// <returns></returns>
-  Status MakeCsrStrings(size_t string_count, const char* const* strings,
-                        gsl::span<const int64_t> inner_index,
-                        gsl::span<const int64_t> outer_index);
-
-  /// <summary>
-  /// Give writable access to Csr values and indices
-  /// </summary>
-  class CsrMutator {
-   public:
-    CsrMutator(Tensor& values, Tensor& inner, Tensor& outer) noexcept
-        : values_(values), inner_(inner), outer_(outer) {}
-    Tensor& Values() const noexcept { return values_; }
-    Tensor& Inner() const noexcept { return inner_; }
-    Tensor& Outer() const noexcept { return outer_; }
-
-   private:
-    std::reference_wrapper<Tensor> values_;
-    std::reference_wrapper<Tensor> inner_;
-    std::reference_wrapper<Tensor> outer_;
-  };
-
-  /// <summary>
-  /// Allocates memory for values and index and returns mutator so
-  /// data can be populated.
-  /// </summary>
-  /// <param name="values_count">Use 0 for fully sparse tensors.</param>
-  /// <param name="inner_index_count">Use 0 for fully sparse tensors.</param>
-  /// <param name="outer_index_count">Use 0 for fully sparse tensors.</param>
-  /// <returns></returns>
-  CsrMutator MakeCsrData(size_t values_count, size_t inner_index_count, size_t outer_index_count);
-
-  /// <summary>
-  /// Read only access to BlockSparse index
-  /// </summary>
-  class BlockSparseView {
-   public:
-    explicit BlockSparseView(const Tensor& indices) noexcept
-        : indices_(indices) {}
-    const Tensor& Indices() const noexcept { return indices_; }
-
-   private:
-    std::reference_wrapper<const Tensor> indices_;
-  };
-
-  /// <summary>
-  /// Return BlockSparseIndex view
-  /// </summary>
-  /// <returns>an instance of BlockSparseView</returns>
-  BlockSparseView AsBlockSparse() const;
-
-  /// <summary>
-  /// Use blocksparse indices contained in the user allocated buffer. The shape of the index
-  /// must be 2-D and must contain one tuple per each of the value blocks that
-  /// were supplied to the constructor. The supplied buffer lifespan must eclipse the life
-  /// of sparse tensor instance.
-  /// </summary>
-  /// <param name="indices_shape">Use {0} for fully sparse tensors.</param>
-  /// <param name="indices_data">Ptr to user allocated buffer. Use nullptr for fully spare tensors.</param>
-  /// <returns></returns>
-  Status UseBlockSparseIndices(const TensorShape& indices_shape, int32_t* indices_data);
-
-  /// <summary>
-  /// The function allocates a single contiguous buffer and copies values and index
-  /// into it. The shape of the values is expected to be at least 3-D but may contain more
-  /// dimensions. At the very minimum it should be (num_blocks, block_size, block_size).
-  ///
-  // The shape of the index is must be at least 2-D and must contain one tuple per each of
-  // the value blocks that  were supplied to the constructor. Each index tuple is a
-  // (row, col) coordinates of the values block in a dense matrix.
-  /// </summary>
-  /// <param name="data_transfer"></param>
-  /// <param name="data_location"></param>
-  /// <param name="values_shape">The shape is expected to be at least 3-D. However, use {0} for fully sparse tensors.</param>
-  /// <param name="values_data">Pointer to a data to be copied. Use nullptr for fully sparse tensors.</param>
-  /// <param name="indices_shape">The shape is expected to be 2-D. However, you can use {0} for fully sparse tensors.</param>
-  /// <param name="indices_data">Pointer to index data to be copied. Use nullptr for fully sparse tensors.</param>
-  /// <returns></returns>
-  Status MakeBlockSparseData(const IDataTransfer& data_transfer,
-                             const OrtMemoryInfo& data_location,
-                             const TensorShape& values_shape, const void* values_data,
-                             const TensorShape& indices_shape, const int32_t* indices_data);
-
-
-  /// <summary>
-  /// The method allocates a single contiguous buffer and creates instances of std::strings in it, with
-  /// copies of the supplied zero-terminated strings followed by COO indices.
-  /// All data is assumed to be on CPU and the allocator supplied must be
-  /// a CPU based allocator.
-  /// </summary>
-  /// <param name="values_shape">Use {0} shape for fully sparse tensors</param>
-  /// <param name="strings">array of char* ptrs, use nullptr for fully sparse tensor</param>
-  /// <param name="indices_shape">Use {0} for fully sparse tensors</param>
-  /// <param name="indices_data">use nullptr for fully sparse tensors</param>
-  /// <returns></returns>
-  Status MakeBlockSparseStrings(const TensorShape& values_shape, const char* const* strings,
-                                const TensorShape& indices_shape, const int32_t* indices_data);
-
-  /// <summary>
-  /// Mutable data access
-  /// </summary>
-  class BlockSparseMutator {
-   public:
-    BlockSparseMutator(Tensor& values, Tensor& indices) noexcept
-        : values_(values), indices_(indices) {}
-    Tensor& Values() noexcept { return values_; }
-    Tensor& Indices() noexcept { return indices_; }
-
-   private:
-    std::reference_wrapper<Tensor> values_;
-    std::reference_wrapper<Tensor> indices_;
-  };
-
-  /// <summary>
-  /// Allocates memory for values and index and returns mutator so
-  /// data can be populated
-  /// </summary>
-  /// <param name="values_shape">Shape is expected to be 3-D, use {0} for fully sparse tensors</param>
-  /// <param name="indices_shape">Shape is expected to be 2-D, use {0} for fully sparse tensors </param>
-  /// <returns></returns>
-  BlockSparseMutator MakeBlockSparseData(const TensorShape& values_shape, const TensorShape& indices_shape);
-
-  /// <summary>
-  /// X-device copy. Destination tensor must have allocator set.
-  /// </summary>
-  /// <param name="data_transfer_manager"></param>
-  /// <param name="exec_q_id"></param>
-  /// <param name="dst_tensor"></param>
-  /// <returns></returns>
-  Status Copy(const DataTransferManager& data_transfer_manager, SparseTensor& dst_tensor) const;
-
-  /// <summary>
-  /// X-device copy. Destination tensor must have allocator set.
-  /// </summary>
-  /// <param name="dst_tensor"></param>
-  /// <returns></returns>
-  Status Copy(const IDataTransfer& data_transfer, SparseTensor& dst_tensor) const;
-
- private:
-  Status AllocateBuffer(int64_t buffer_size, size_t num_values);
-  void ReleaseBuffer();
-  void* IndicesStart(int64_t values_bytes);
-  const void* IndicesStart(int64_t values_bytes) const;
-  Status ValidateBlockSparseShapes(const TensorShape& values_shape, const TensorShape& index_shape) const;
-
-  std::vector<int64_t> GetCooIndexDims(size_t values_count, size_t index_size) const;
-  void InitCooIndex(const TensorShape& index_shape, int64_t* index_data);
-
-  Status ValidateCsrIndices(size_t values_count, size_t inner_size, size_t outer_size) const;
-  void InitCsrIndices(size_t inner_size, const int64_t* inner, size_t outer_size, const int64_t* outer);
-  void InitBlockSparseIndices(const TensorShape& indices_shape, int32_t* indices_data);
-
-  SparseFormat format_;                        // sparse format enum value
-  TensorShape dense_shape_;                    // a shape of a corresponding dense tensor
-  const PrimitiveDataTypeBase* ml_data_type_;  // MLDataType for contained values
-  AllocatorPtr allocator_;                     // Allocator or nullptr when using user supplied buffers
-  OrtMemoryInfo location_;                     // Memory info where data resides. When allocator is supplied,
-                                               // location_ is obtained from the allocator.
-  void* p_data_;                               // Allocated buffer ptr, or nullptr when using user supplied buffers
-  int64_t buffer_size_;                        // Allocated buffer size or zero when using user supplied buffers.
-  Tensor values_;                              // Tensor instance that holds a values buffer information either user supplied or
-                                               // to a beginning of p_data_, before format specific indices.
-  std::vector<Tensor> format_data_;            // A collection of format specific indices. They contain pointers to either a
-                                               // user supplied buffers or to portions of contiguous buffer p_data_.
-};
-
-}  // namespace onnxruntime
-
-#endif
\ No newline at end of file
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/stream_handles.h b/projects/llm_framework/include/onnxruntime/core/framework/stream_handles.h
deleted file mode 100644
index 72c64a51..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/stream_handles.h
+++ /dev/null
@@ -1,170 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-#pragma once
-
-#include <functional>
-#include <unordered_map>
-#include "core/framework/allocator.h"
-#include "core/framework/ortdevice.h"
-#include "core/common/status.h"
-
-namespace onnxruntime {
-class IExecutionProvider;
-// this opaque handle could be anything the target device generated.
-// it could be a cuda event, or a npu notification implementation
-using NotificationHandle = void*;
-// it can be either a cuda stream, or even nullptr for device doesn't have stream support like cpu.
-using StreamHandle = void*;
-
-namespace synchronize {
-class Notification;
-}
-
-// a stream abstraction which hold an opaque handle, and a reference to which OrtDevice instance this stream belong to.
-// it need to be OrtDevice instance as we might have different stream on different OrtDevice with same type.
-// i.e. different cuda stream on different GPU.
-class Stream {
- public:
-  Stream(StreamHandle h, const OrtDevice& d) : handle_(h), device_(d) {}
-
-  virtual ~Stream() = default;
-  virtual std::unique_ptr<synchronize::Notification> CreateNotification(size_t /*num_consumers*/) {
-    return {};
-  };
-  // block the host thread until all the tasks in the stream finished.
-  virtual void Flush(){};
-  // The framework may reuse the stream instance for multiple iterations.
-  // This is the API that provide a chance to let the device stream cleanup
-  // resource at the end of a iteration.
-  virtual Status CleanUpOnRunEnd() { return Status::OK(); };
-
-  StreamHandle GetHandle() const { return handle_; }
-
-  const OrtDevice& GetDevice() const { return device_; }
-
-  // We use the timestamp based vector clocks to optimize the resource sharing
-  // between different streams.
-  // Each stream maintain following data structure:
-  // 1. Current timestamp
-  // 2. A lookup table that for a given stream, what is its timestamp when the
-  //    last synchronization happened with current stream.
-  // 3. When a notification is activated, it take a snapshot of current stream's
-  //    lookup table.
-  // 4. When synchronization happened (current stream wait on a notification),
-  //    update its lookup table with the table snapshot in notification.
-  // The memory reusing strategy is:
-  // A kernel in current stream is safe to reuse another stream's memory chunk
-  // as long as the reused chunk's timestamp is less than the last synchonized
-  // timestamp recorded in the lookup table.
-
-  // Get the current timestamp
-  uint64_t GetCurrentTimestamp() const { return timestamp_; }
-
-  // return the timestamp when the last synchronization happened between target stream and current stream.
-  // return 0 if no synchonization happened.
-  // if target_stream is nullptr, it means it is a sequence running on device doesn't support Stream (i.e. CPU)
-  // we can safely return 0 in that case to save a lookup.
-  uint64_t GetLastSyncTimestampWithTargetStream(Stream* target_stream) const {
-    if (!target_stream)
-      return 0;
-    auto it = other_stream_clock_.find(target_stream);
-    return it == other_stream_clock_.end() ? 0 : it->second;
-  }
-
-  // make a copy of the current stream lookup table.
-  // this is used to create a snapshot of the stream lookup table in notification.
-  void CloneCurrentStreamSyncTable(std::unordered_map<Stream*, uint64_t>& output) const {
-    output.reserve(other_stream_clock_.size());
-    output.insert(other_stream_clock_.begin(), other_stream_clock_.end());
-  }
-
-  // bump the current timestamp
-  // When a notification get activated, bump the snapshot in its owner.
-  // Stream is not shared across threads, BumpTimeStampAndReturn will only be invoked on the current thread
-  // where the stream is executed on, so there is no race condition.
-  uint64_t BumpTimeStampAndReturn() {
-    return ++timestamp_;
-  }
-
-  // update the stream lookup table with the snapshot saved in notification.
-  void UpdateStreamClock(const std::unordered_map<Stream*, uint64_t>& clock) {
-    for (const auto& kv : clock) {
-      auto ret = other_stream_clock_.insert(kv);
-      if (!ret.second) {
-        ret.first->second = std::max(ret.first->second, kv.second);
-      }
-    }
-  }
-
- private:
-  StreamHandle handle_;
-  const OrtDevice& device_;
-  uint64_t timestamp_{0};
-  // TODO: use inline container.
-  // currently this class is header only, but abseil doesn't compile with nvcc
-  // we need to add new symbol to provider_bridge and hide abseil from the header.
-  std::unordered_map<Stream*, uint64_t> other_stream_clock_{};
-
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(Stream);
-};
-
-namespace synchronize {
-// an object which record the status of the stream, and can be wait on from another stream.
-class Notification {
- public:
-  explicit Notification(Stream& s) : stream_(s) {}
-  virtual ~Notification() = default;
-
-  // this api will perform three operations:
-  // 1. activate the notification on device, for example, record an event on GPU.
-  // 2. take a snapshot of the timestamp lookup table in current stream.
-  // 3. bump the timestamp for current stream.
-  void ActivateAndUpdate() {
-    Activate();
-    stream_.CloneCurrentStreamSyncTable(stream_clock_);
-    stream_clock_[&stream_] = stream_.BumpTimeStampAndReturn();
-  }
-
-  // return the timestamp lookup table saved in the notification.
-  const std::unordered_map<Stream*, uint64_t>& GetStreamSyncTable() {
-    return stream_clock_;
-  }
-
- protected:
-  virtual void Activate() = 0;
-  // which stream create this notification.
-  Stream& stream_;
-  // TODO: use inline container.
-  // currently this class is header only, but abseil doesn't compile with nvcc
-  // we need to add new symbol to provider_bridge and hide abseil from the header.
-  std::unordered_map<Stream*, uint64_t> stream_clock_{};
-};
-}  // namespace synchronize
-
-// the definition for the handle for stream commands
-// EP can register the handle to the executor.
-// in the POC, just use primitive function pointer
-// TODO: use a better way to dispatch handles.
-using CreateStreamFn = std::function<std::unique_ptr<Stream>(const OrtDevice&)>;
-
-// an interface of a simple registry which hold the handles EP registered.
-// make it interface so we can pass it through shared library based execution providers
-class IStreamCommandHandleRegistry {
- public:
-  virtual ~IStreamCommandHandleRegistry() = default;
-  // Wait is a little special as we need to consider the source stream the notification generated, and the stream we are waiting.
-  // i.e., for an cuda event what notify the memory copy, it could be wait on a CPU stream, or on another cuda stream.
-  [[nodiscard]] virtual WaitNotificationFn GetWaitHandle(OrtDevice::DeviceType notification_ower_device_type,
-                                                         OrtDevice::DeviceType executor_device_type) const = 0;
-  // Get the stream creation function registered on the given device type.
-  [[nodiscard]] virtual CreateStreamFn GetCreateStreamFn(OrtDevice::DeviceType execution_device_type) const = 0;
-  // register a wait methond which will be invoked when we wait a notification (created by 'notification_device_type' device) on a stream at 'device_type' device.
-  virtual void RegisterWaitFn(OrtDevice::DeviceType notification_device_type,
-                              OrtDevice::DeviceType device_type,
-                              WaitNotificationFn fn) = 0;
-  // register a handle about how to create stream on given device type.
-  virtual void RegisterCreateStreamFn(OrtDevice::DeviceType device_type, CreateStreamFn f) = 0;
-};
-
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/tensor.h b/projects/llm_framework/include/onnxruntime/core/framework/tensor.h
deleted file mode 100644
index 975793a0..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/tensor.h
+++ /dev/null
@@ -1,323 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <stddef.h>
-#include <iostream>
-#include <string>
-#include <vector>
-
-#include "core/common/gsl.h"
-#include "core/common/common.h"
-#include "core/framework/allocator.h"
-#include "core/framework/tensor_shape.h"
-#include "core/framework/buffer_deleter.h"
-#include "onnxruntime_config.h"
-#include "core/framework/data_types.h"
-#include "core/framework/data_types_internal.h"
-
-struct OrtValue;
-
-namespace onnxruntime {
-
-// TODO:ensure dtype_!=nullptr
-#ifdef __GNUC__
-#pragma GCC diagnostic push
-#ifdef HAS_NULL_DEREFERENCE
-#pragma GCC diagnostic ignored "-Wnull-dereference"
-#endif
-#endif
-/*
-  We want to keep tensor as simple as possible, it is just a placeholder
-  for a piece of memory, with additional shape information.
-  Memory is owned and managed by Executor / Workspace, so Tensor just uses
-  it, and won't do any allocation / release.
-*/
-
-class Tensor final {
- public:
-  // NB! Removing Create() methods returning unique_ptr<Tensor>. Still available in other EPs that are dynamically linked.
-  // Strive not to allocate Tensor with new/delete as it is a shallow class and using it by value is just fine.
-  // Use InitOrtValue() methods to allocate for OrtValue.
-
-  Tensor() = default;  // to allow creating vector<Tensor> to support seq(tensor)
-
-  /**
-   * Create tensor with given type, shape, pre-allocated memory and allocator info.
-   * This function won't check if the preallocated buffer(p_data) has enough room for the shape.
-   * \param p_type Data type of the tensor
-   * \param shape Shape of the tensor
-   * \param p_data A preallocated buffer. Can be NULL if the shape is empty.
-   *              Tensor does not own the data and will not delete it
-   * \param alloc Where the buffer('p_data') was allocated from
-   * \param offset Offset in bytes to start of Tensor within p_data.
-   * \param strides Strides span. Can be empty if the tensor is contiguous.
-   */
-  Tensor(MLDataType p_type, const TensorShape& shape, void* p_data, const OrtMemoryInfo& alloc,
-         ptrdiff_t offset = 0, gsl::span<const int64_t> strides = {});
-
-  /// <summary>
-  /// Creates an instance of Tensor on the heap using the appropriate __ctor and
-  /// initializes OrtValue with it.
-  /// </summary>
-  /// <param name="p_type"></param>
-  /// <param name="shape"></param>
-  /// <param name="p_data"></param>
-  /// <param name="info"></param>
-  /// <param name="offset"></param>
-  /// <param name="strides"></param>
-  static void InitOrtValue(MLDataType p_type, const TensorShape& shape,
-                           void* p_data, const OrtMemoryInfo& location,
-                           OrtValue& ort_value, ptrdiff_t offset = 0,
-                           gsl::span<const int64_t> strides = {});
-
-  /// <summary>
-  /// Creates an instance of Tensor who own the pre-allocated buffer.
-  /// </summary>
-  /// <param name="p_type"></param>
-  /// <param name="shape"></param>
-  /// <param name="p_data"></param>
-  /// <param name="allocator"></param>
-  /// <param name="offset"></param>
-  /// <param name="strides"></param>
-  static void InitOrtValue(MLDataType p_type, const TensorShape& shape,
-                           void* p_data, std::shared_ptr<IAllocator> allocator,
-                           OrtValue& ort_value, ptrdiff_t offset = 0,
-                           gsl::span<const int64_t> strides = {});
-
-  static size_t CalculateTensorStorageSize(MLDataType p_type,
-                                           const TensorShape& shape,
-                                           gsl::span<const int64_t> strides = {});
-
-  /**
-   * Deprecated. The original design is this Tensor class won't do any allocation / release.
-   * However, this function will allocate the buffer for the shape, and do placement new if p_type is string tensor.
-   */
-  Tensor(MLDataType p_type, const TensorShape& shape, std::shared_ptr<IAllocator> allocator,
-         gsl::span<const int64_t> strides = {});
-
-  /// <summary>
-  /// Creates an instance of Tensor on the heap using the appropriate __ctor and
-  /// initializes OrtValue with it.
-  /// </summary>
-  /// <param name="elt_type"></param>
-  /// <param name="shape"></param>
-  /// <param name="allocator"></param>
-  /// <param name="ort_value"></param>
-  /// <param name="strides"></param>
-  static void InitOrtValue(MLDataType elt_type,
-                           const TensorShape& shape,
-                           std::shared_ptr<IAllocator> allocator,
-                           OrtValue& ort_value,
-                           gsl::span<const int64_t> strides = {});
-
-  /**
-   * Create tensor with given type, shape, pre-allocated memory and allocator which will be used to free the pre-allocated memory.
-   * This function won't check if the preallocated buffer(p_data) has enough room for the shape.
-   * However, this function will de-allocate the buffer upon the tensor getting destructed.
-   * \param p_type Data type of the tensor
-   * \param shape Shape of the tensor
-   * \param p_data A preallocated buffer. Can be NULL if the shape is empty.
-   *              Tensor will own the memory and will delete it when the tensor instance is destructed.
-   * \param deleter Allocator used to free the pre-allocated memory
-   * \param offset Offset in bytes to start of Tensor within p_data.
-   * \param strides Strides span. Can be empty if the tensor is contiguous.
-   */
-  Tensor(MLDataType p_type, const TensorShape& shape, void* p_data, std::shared_ptr<IAllocator> deleter,
-         ptrdiff_t offset = 0, gsl::span<const int64_t> strides = {});
-
-  ~Tensor();
-
-  // Move is allowed
-  ORT_DISALLOW_COPY_AND_ASSIGNMENT(Tensor);
-
-  Tensor(Tensor&& other) noexcept;
-
-  Tensor& operator=(Tensor&& other) noexcept;
-
-  /**
-     Returns the data type.
-  */
-  MLDataType DataType() const { return dtype_; }
-
-  /**
-     Returns the data type enum constant
-     @remarks Use utils::ToTensorProtoElementType<T> for comparison.
-  */
-  int32_t GetElementType() const {
-    return dtype_->GetDataType();
-  }
-
-  // Check if contains string data. This is a separate
-  // interface bc it is frequently used.
-  bool IsDataTypeString() const {
-    return utils::IsPrimitiveDataType<std::string>(dtype_);
-  }
-
-  // Checks if the Tensor contains data type T
-  template <class T>
-  bool IsDataType() const {
-    return utils::IsPrimitiveDataType<T>(dtype_);
-  }
-
-  /**
-     Returns the shape of the tensor.
-  */
-  const TensorShape& Shape() const noexcept { return shape_; }
-
-  /**
-     Returns the location of the tensor's memory
-  */
-  const OrtMemoryInfo& Location() const { return alloc_info_; }
-
-  /**
-     May return nullptr if tensor size is zero
-  */
-  template <typename T>
-  T* MutableData() {
-    // Type check
-    ORT_ENFORCE(utils::IsPrimitiveDataType<T>(dtype_), "Tensor type mismatch. ",
-                "T ", "!=", dtype_);
-    return reinterpret_cast<T*>(static_cast<char*>(p_data_) + byte_offset_);
-  }
-
-  /**
-     May return nullptr if tensor size is zero
-  */
-  template <typename T>
-  gsl::span<T> MutableDataAsSpan() {
-    // Type check
-    ORT_ENFORCE(utils::IsPrimitiveDataType<T>(dtype_), "Tensor type mismatch. ",
-                "T ", "!=", dtype_);
-    T* data = reinterpret_cast<T*>(static_cast<char*>(p_data_) + byte_offset_);
-    return gsl::make_span(data, static_cast<size_t>(shape_.Size()));
-  }
-
-  template <typename T>
-  const T* Data() const {
-    // Type check
-    ORT_ENFORCE(utils::IsPrimitiveDataType<T>(dtype_), "Tensor type mismatch. ",
-                "T ", "!=", dtype_);
-    return reinterpret_cast<const T*>(static_cast<char*>(p_data_) + byte_offset_);
-  }
-
-  template <typename T>
-  gsl::span<const T> DataAsSpan() const {
-    // Type check
-    ORT_ENFORCE(utils::IsPrimitiveDataType<T>(dtype_), "Tensor type mismatch. ",
-                "T ", "!=", dtype_);
-    const T* data = reinterpret_cast<const T*>(static_cast<char*>(p_data_) + byte_offset_);
-    return gsl::make_span(data, static_cast<typename gsl::span<T>::size_type>(shape_.Size()));
-  }
-
-  void* MutableDataRaw(MLDataType type) {
-    ORT_ENFORCE(type == dtype_, "Tensor type mismatch.", type, "!=", dtype_);
-    return static_cast<char*>(p_data_) + byte_offset_;
-  }
-
-  const void* DataRaw(MLDataType type) const {
-    ORT_ENFORCE(type == dtype_, "Tensor type mismatch.", type, "!=", dtype_);
-    return static_cast<char*>(p_data_) + byte_offset_;
-  }
-
-  void* MutableDataRaw() noexcept {
-    return static_cast<char*>(p_data_) + byte_offset_;
-  }
-
-  const void* DataRaw() const noexcept {
-    return static_cast<char*>(p_data_) + byte_offset_;
-  }
-
-  bool OwnsBuffer() const noexcept {
-    return buffer_deleter_ != nullptr;
-  }
-
-  /**
-   * Resizes the tensor without touching underlying storage.
-   * This requires the total size of the tensor to remains constant.
-   * @warning this function is NOT thread-safe.
-   */
-  inline void Reshape(const TensorShape& new_shape) {
-    ORT_ENFORCE(shape_.Size() == new_shape.Size(),
-                "Tensor size (" + std::to_string(shape_.Size()) +
-                    ") != new size (" + std::to_string(new_shape.Size()) + ")");
-    shape_ = new_shape;
-  }
-
-  /**
-   * Get the byte offset with respect to the p_data
-   * @warning this is a temporary solution for reusing the buffer bigger than needed.
-   * @warning use with caution - make sure you do boundary check before calling this method (see view.cc)
-   */
-  inline ptrdiff_t ByteOffset() const {
-    return byte_offset_;
-  }
-
-  /**
-   * Set the byte offset with respect to the p_data
-   * @warning this is a temporary solution for reusing the buffer bigger than needed.
-   */
-  inline void SetByteOffset(ptrdiff_t byte_offset) {
-    byte_offset_ = byte_offset;
-  }
-
-  /**
-  The number of bytes of data.
-  */
-  size_t SizeInBytes() const;
-
-#ifdef ENABLE_STRIDED_TENSORS
-  /**
-   * Get the strides of the tensor.
-   */
-  gsl::span<const int64_t> Strides() const;
-
-  /**
-   * Return if the tensor is contiguous.
-   */
-  bool IsContiguous() const noexcept { return is_contiguous_; }
-
-  /**
-   * Set strides.
-   */
-  void SetShapeAndStrides(const TensorShape& new_shape, gsl::span<const int64_t> new_strides);
-#endif
-
-  // More API methods.
- private:
-  void Init(MLDataType p_type,
-            const TensorShape& shape,
-            void* p_raw_data,
-            AllocatorPtr deleter,
-            ptrdiff_t offset = 0,
-            gsl::span<const int64_t> strides = {});
-
-  void ReleaseBuffer();
-
-#ifdef ENABLE_STRIDED_TENSORS
-  bool CheckIsContiguous() const;
-#endif
-
-  void* p_data_;
-  /**
-     if buffer_deleter_ is null, it means tensor does not own the buffer.
-     otherwise tensor will use the deleter to release the buffer when
-     tensor is released.
-  */
-  AllocatorPtr buffer_deleter_;
-
-  TensorShape shape_;
-#ifdef ENABLE_STRIDED_TENSORS
-  mutable TensorShapeVector strides_;
-  bool is_contiguous_ = true;
-#endif
-
-  const PrimitiveDataTypeBase* dtype_;
-  OrtMemoryInfo alloc_info_;
-  ptrdiff_t byte_offset_;
-};
-#ifdef __GNUC__
-#pragma GCC diagnostic pop
-#endif
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/tensor_shape.h b/projects/llm_framework/include/onnxruntime/core/framework/tensor_shape.h
deleted file mode 100644
index b3783696..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/tensor_shape.h
+++ /dev/null
@@ -1,203 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-#include <iosfwd>
-#include <vector>
-#include <algorithm>
-#include <string>
-#include <cstring>
-#include "core/common/gsl.h"
-#include "onnxruntime_config.h"
-
-#ifndef DISABLE_ABSEIL
-// Need to include abseil inlined_vector.h header directly here
-// as hash tables cause CUDA 10.2 compilers to fail. inlined_vector.h is fine.
-#ifdef _MSC_VER
-#pragma warning(push)
-// C4127: conditional expression is constant
-#pragma warning(disable : 4127)
-// C4324: structure was padded due to alignment specifier
-// Usage of alignas causes some internal padding in places.
-#pragma warning(disable : 4324)
-#endif
-
-#include <absl/container/inlined_vector.h>
-
-#ifdef _MSC_VER
-#pragma warning(pop)
-#endif
-#endif  // DISABLE_ABSEIL
-
-#include "core/common/span_utils.h"
-
-namespace onnxruntime {
-#ifdef __GNUC__
-#pragma GCC diagnostic push
-#ifdef HAS_NULL_DEREFERENCE
-#pragma GCC diagnostic ignored "-Wnull-dereference"
-#endif
-#endif
-
-constexpr size_t kTensorShapeSmallBufferElementsSize = 5;
-
-#ifndef DISABLE_ABSEIL
-// Use this type to build a shape and then create TensorShape.
-using TensorShapeVector = absl::InlinedVector<int64_t, kTensorShapeSmallBufferElementsSize>;
-#else
-class TensorShapeVector : public std::vector<int64_t> {
-  using Base = std::vector<int64_t>;
-
- public:
-  using Base::Base;
-};
-
-#endif  // DISABLE_ABSEIL
-
-inline TensorShapeVector ToShapeVector(const gsl::span<const int64_t>& span) {
-  TensorShapeVector out;
-  out.reserve(span.size());
-  out.assign(span.begin(), span.end());
-  return out;
-}
-
-inline gsl::span<const int64_t> ToConstSpan(const TensorShapeVector& vec) {
-  return gsl::make_span(vec);
-}
-
-class TensorShape {
-  // We use negative numbers for unknown symbolic dimension. Each negative
-  // number represents a unique symbolic dimension.
- public:
-  TensorShape() = default;
-
-  TensorShape(const TensorShape& other) : TensorShape(other.GetDims()) {}
-  TensorShape& operator=(const TensorShape& other);
-  TensorShape& operator=(const gsl::span<const int64_t>& dims) {
-    *this = TensorShape(dims);
-    return *this;
-  }
-
-  TensorShape(TensorShape&& other) noexcept { operator=(std::move(other)); }
-  TensorShape& operator=(TensorShape&& other) noexcept;
-
-  TensorShape(gsl::span<const int64_t> dims);
-  TensorShape(const TensorShapeVector& dims) : TensorShape(gsl::make_span(dims)) {}
-  TensorShape(std::initializer_list<int64_t> dims) : TensorShape(gsl::make_span(dims.begin(), dims.end())) {}
-  TensorShape(const int64_t* dimension_sizes, size_t dimension_count) : TensorShape(gsl::span<const int64_t>(dimension_sizes, dimension_count)) {}
-  TensorShape(const std::vector<int64_t>& dims, size_t start, size_t end) : TensorShape(gsl::span<const int64_t>(&dims[start], end - start)) {}
-
-  // Create a TensorShape that points to an existing buffer internally. As no copy is made, 'data' must remain valid for the life of the TensorShape
-  static const TensorShape FromExistingBuffer(const std::vector<int64_t>& data) {
-    return TensorShape(External{}, gsl::span<int64_t>(const_cast<int64_t*>(data.data()), data.size()));
-  }
-
-  /**
-     Return the dimension specified by <idx>.
-  */
-  int64_t operator[](size_t idx) const { return values_[idx]; }
-  int64_t& operator[](size_t idx) { return values_[idx]; }
-
-  bool operator==(const TensorShape& other) const noexcept { return SpanEq(GetDims(), other.GetDims()); }
-  bool operator!=(const TensorShape& other) const noexcept { return !(*this == other); }
-
-  size_t NumDimensions() const noexcept {
-    return values_.size();
-  }
-
-  /**
-     Copy dims into an array with given size
-  */
-  void CopyDims(int64_t* dims, size_t num_dims) const {
-    memcpy(dims, values_.data(), sizeof(int64_t) * std::min(num_dims, NumDimensions()));
-  }
-
-  /**
-     Copy dims from a specific start dim into an array with given size
-     `start_dim` is expected to be in the inclusive range [0, NumDimensions() - 1]
-     and this function does no checks to ensure that
-  */
-  void CopyDims(int64_t* dims, size_t start_dim, size_t num_dims) const {
-    memcpy(dims, values_.data() + start_dim, sizeof(int64_t) * std::min(num_dims, NumDimensions() - start_dim));
-  }
-
-  /**
-     Return underlying vector representation.
-  */
-  gsl::span<const int64_t> GetDims() const { return values_; }
-
-  TensorShapeVector AsShapeVector() const {
-    return ToShapeVector(values_);
-  }
-
-  /**
-   * Return the total number of elements. Returns 1 for an empty (rank 0) TensorShape.
-   *
-   * May return -1
-   */
-  int64_t Size() const;
-
-  /**
-     Return the total number of elements up to the specified dimension.
-     If the dimension interval is empty (dimension == 0), return 1.
-     @param dimension Return size up to this dimension. Value must be between 0 and this->NumDimensions(), inclusive.
-  */
-  int64_t SizeToDimension(size_t dimension) const;
-
-  /**
-     Return the total number of elements from the specified dimension to the end of the tensor shape.
-     If the dimension interval is empty (dimension == this->NumDimensions()), return 1.
-     @param dimension Return size from this dimension to the end. Value must be between 0 and this->NumDimensions(),
-                      inclusive.
-  */
-  int64_t SizeFromDimension(size_t dimension) const;
-
-  /**
-     Return a new TensorShape of the dimensions from dimstart to dimend.
-  */
-  TensorShape Slice(size_t dimstart, size_t dimend) const;
-
-  /**
-     Return a new TensorShape of the dimensions from dimstart to end.
-  */
-  TensorShape Slice(size_t dimstart) const { return Slice(dimstart, values_.size()); }
-
-  /**
-     output dimensions nicely formatted
-  */
-  std::string ToString() const;
-
-  /**
-     Calculate size between start and end.
-     Assumes start and end are between 0 and this->NumDimensions(), inclusive, and that
-     start < end.
-  */
-  int64_t SizeHelper(size_t start, size_t end) const;
-
-  /**
-     empty shape or 1D shape (1) is regarded as scalar tensor
-  */
-  bool IsScalar() const {
-    size_t len = values_.size();
-    return len == 0 || (len == 1 && values_[0] == 1);
-  }
-
- private:
-  struct External {};
-  TensorShape(External, gsl::span<int64_t> buffer) : values_{buffer} {}
-
-  void Allocate(size_t size);
-
-  gsl::span<int64_t> values_;
-  int64_t small_buffer_[kTensorShapeSmallBufferElementsSize]{0};
-  std::unique_ptr<int64_t[]> allocated_buffer_;
-
-  friend struct ProviderHostImpl;  // So that the shared provider interface can access Allocate
-};
-#ifdef __GNUC__
-#pragma GCC diagnostic pop
-#endif
-// operator<< to nicely output to a stream
-std::ostream& operator<<(std::ostream& out, const TensorShape& shape);
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/framework/to_tensor_proto_element_type.h b/projects/llm_framework/include/onnxruntime/core/framework/to_tensor_proto_element_type.h
deleted file mode 100644
index 0be55f00..00000000
--- a/projects/llm_framework/include/onnxruntime/core/framework/to_tensor_proto_element_type.h
+++ /dev/null
@@ -1,79 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <cstdint>
-#include <string>
-
-#ifndef SHARED_PROVIDER
-#include "onnx/onnx_pb.h"
-#endif
-
-#include "core/framework/float16.h"
-
-namespace onnxruntime {
-namespace utils {
-/** Gets the TensorProto_DataType corresponding to the template type `T`. */
-template <typename T>
-constexpr ONNX_NAMESPACE::TensorProto_DataType ToTensorProtoElementType() {
-  return ONNX_NAMESPACE::TensorProto_DataType_UNDEFINED;
-}
-template <>
-constexpr ONNX_NAMESPACE::TensorProto_DataType ToTensorProtoElementType<float>() {
-  return ONNX_NAMESPACE::TensorProto_DataType_FLOAT;
-}
-template <>
-constexpr ONNX_NAMESPACE::TensorProto_DataType ToTensorProtoElementType<uint8_t>() {
-  return ONNX_NAMESPACE::TensorProto_DataType_UINT8;
-}
-template <>
-constexpr ONNX_NAMESPACE::TensorProto_DataType ToTensorProtoElementType<int8_t>() {
-  return ONNX_NAMESPACE::TensorProto_DataType_INT8;
-}
-template <>
-constexpr ONNX_NAMESPACE::TensorProto_DataType ToTensorProtoElementType<uint16_t>() {
-  return ONNX_NAMESPACE::TensorProto_DataType_UINT16;
-}
-template <>
-constexpr ONNX_NAMESPACE::TensorProto_DataType ToTensorProtoElementType<int16_t>() {
-  return ONNX_NAMESPACE::TensorProto_DataType_INT16;
-}
-template <>
-constexpr ONNX_NAMESPACE::TensorProto_DataType ToTensorProtoElementType<int32_t>() {
-  return ONNX_NAMESPACE::TensorProto_DataType_INT32;
-}
-template <>
-constexpr ONNX_NAMESPACE::TensorProto_DataType ToTensorProtoElementType<int64_t>() {
-  return ONNX_NAMESPACE::TensorProto_DataType_INT64;
-}
-template <>
-constexpr ONNX_NAMESPACE::TensorProto_DataType ToTensorProtoElementType<std::string>() {
-  return ONNX_NAMESPACE::TensorProto_DataType_STRING;
-}
-template <>
-constexpr ONNX_NAMESPACE::TensorProto_DataType ToTensorProtoElementType<bool>() {
-  return ONNX_NAMESPACE::TensorProto_DataType_BOOL;
-}
-template <>
-constexpr ONNX_NAMESPACE::TensorProto_DataType ToTensorProtoElementType<MLFloat16>() {
-  return ONNX_NAMESPACE::TensorProto_DataType_FLOAT16;
-}
-template <>
-constexpr ONNX_NAMESPACE::TensorProto_DataType ToTensorProtoElementType<double>() {
-  return ONNX_NAMESPACE::TensorProto_DataType_DOUBLE;
-}
-template <>
-constexpr ONNX_NAMESPACE::TensorProto_DataType ToTensorProtoElementType<uint32_t>() {
-  return ONNX_NAMESPACE::TensorProto_DataType_UINT32;
-}
-template <>
-constexpr ONNX_NAMESPACE::TensorProto_DataType ToTensorProtoElementType<uint64_t>() {
-  return ONNX_NAMESPACE::TensorProto_DataType_UINT64;
-}
-template <>
-constexpr ONNX_NAMESPACE::TensorProto_DataType ToTensorProtoElementType<BFloat16>() {
-  return ONNX_NAMESPACE::TensorProto_DataType_BFLOAT16;
-}
-}  // namespace utils
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/graph/basic_types.h b/projects/llm_framework/include/onnxruntime/core/graph/basic_types.h
deleted file mode 100644
index 36984d04..00000000
--- a/projects/llm_framework/include/onnxruntime/core/graph/basic_types.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <unordered_map>
-#include <unordered_set>
-#include <string>
-#include <cstdint>
-#include <memory>
-#include <functional>
-
-#include "core/common/basic_types.h"
-#include "core/common/status.h"
-
-namespace ONNX_NAMESPACE {
-class ValueInfoProto;
-class TensorProto;
-class SparseTensorProto;
-class TypeProto;
-class AttributeProto;
-// define types that would come from the ONNX library if we were building against it.
-#if defined(ORT_MINIMAL_BUILD)
-using OperatorSetVersion = int;
-#endif
-
-}  // namespace ONNX_NAMESPACE
-
-namespace onnxruntime {
-using NodeIndex = size_t;
-using Version = int64_t;
-using NodeArgInfo = ONNX_NAMESPACE::ValueInfoProto;
-using InitializedTensorSet = std::unordered_map<std::string, const ONNX_NAMESPACE::TensorProto*>;
-using ArgNameToTypeMap = std::unordered_map<std::string, ONNX_NAMESPACE::TypeProto>;
-using ProviderType = const std::string&;
-
-// TODO - Evaluate switching the types below to support transparent comparators and enable
-// lookups based on gsl::cstring_span<> and std::string_view.  This would reduces allocations
-// converting to std::string, but requires conversion to std::map<std::string, foo, std::less<>>
-// instead of std::unordered_map<std::string, foo, [std::less<foo>]>.
-
-using NodeAttributes = std::unordered_map<std::string, ONNX_NAMESPACE::AttributeProto>;
-class IOnnxRuntimeOpSchemaCollection;
-using IOnnxRuntimeOpSchemaCollectionPtr = std::shared_ptr<IOnnxRuntimeOpSchemaCollection>;
-
-class OpKernel;
-class OpKernelInfo;
-class FuncManager;
-using KernelCreateFn = std::function<onnxruntime::common::Status(FuncManager& func_mgr, const OpKernelInfo& info, std::unique_ptr<OpKernel>& out)>;
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/graph/constants.h b/projects/llm_framework/include/onnxruntime/core/graph/constants.h
deleted file mode 100644
index b11c12b8..00000000
--- a/projects/llm_framework/include/onnxruntime/core/graph/constants.h
+++ /dev/null
@@ -1,49 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-namespace onnxruntime {
-
-constexpr const char* kNoOp = "NoOp";
-constexpr const char* kConstant = "Constant";
-constexpr const char* kFunctionOp = "_kFunctionOp";
-constexpr const char* kConstantValue = "value";
-constexpr const char* kOnnxDomain = "";
-// NOTE: Node::Init converts kOnnxDomainAlias to kOnnxDomain, so all Node instances use kOnnxDomain.
-constexpr const char* kOnnxDomainAlias = "ai.onnx";
-constexpr const char* kMLDomain = "ai.onnx.ml";
-constexpr const char* kMSDomain = "com.microsoft";
-constexpr const char* kPytorchAtenDomain = "org.pytorch.aten";
-constexpr const char* kMSExperimentalDomain = "com.microsoft.experimental";
-constexpr const char* kMSNchwcDomain = "com.microsoft.nchwc";
-constexpr const char* kMSInternalNHWCDomain = "com.ms.internal.nhwc";
-constexpr const char* kMSDmlDomain = "com.microsoft.dml";
-constexpr const char* kNGraphDomain = "com.intel.ai";
-constexpr const char* kMIGraphXDomain = "";
-constexpr const char* kVitisAIDomain = "com.xilinx";
-
-constexpr const char* kCpuExecutionProvider = "CPUExecutionProvider";
-constexpr const char* kCudaExecutionProvider = "CUDAExecutionProvider";
-constexpr const char* kDnnlExecutionProvider = "DnnlExecutionProvider";
-constexpr const char* kOpenVINOExecutionProvider = "OpenVINOExecutionProvider";
-constexpr const char* kVitisAIExecutionProvider = "VitisAIExecutionProvider";
-constexpr const char* kTensorrtExecutionProvider = "TensorrtExecutionProvider";
-constexpr const char* kNnapiExecutionProvider = "NnapiExecutionProvider";
-constexpr const char* kRknpuExecutionProvider = "RknpuExecutionProvider";
-constexpr const char* kDmlExecutionProvider = "DmlExecutionProvider";
-constexpr const char* kMIGraphXExecutionProvider = "MIGraphXExecutionProvider";
-constexpr const char* kAclExecutionProvider = "ACLExecutionProvider";
-constexpr const char* kArmNNExecutionProvider = "ArmNNExecutionProvider";
-constexpr const char* kRocmExecutionProvider = "ROCMExecutionProvider";
-constexpr const char* kCoreMLExecutionProvider = "CoreMLExecutionProvider";
-constexpr const char* kSnpeExecutionProvider = "SNPEExecutionProvider";
-constexpr const char* kTvmExecutionProvider = "TvmExecutionProvider";
-constexpr const char* kXnnpackExecutionProvider = "XnnpackExecutionProvider";
-constexpr const char* kCannExecutionProvider = "CANNExecutionProvider";
-constexpr const char* kAzureExecutionProvider = "AzureExecutionProvider";
-
-constexpr const char* kExecutionProviderSharedLibraryPath = "shared_lib_path";
-constexpr const char* kExecutionProviderSharedLibraryEntry = "provider_factory_entry_point";
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/graph/function.h b/projects/llm_framework/include/onnxruntime/core/graph/function.h
deleted file mode 100644
index 6bc03a0a..00000000
--- a/projects/llm_framework/include/onnxruntime/core/graph/function.h
+++ /dev/null
@@ -1,31 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include "core/common/common.h"
-#include "core/graph/indexed_sub_graph.h"
-
-namespace onnxruntime {
-class Graph;
-class Node;
-}  // namespace onnxruntime
-
-namespace onnxruntime {
-
-/** 
-@class Function 
-Class representing a Function.
-*/
-class Function {
- public:
-  virtual ~Function() = default;
-
-  /** Gets the Graph instance for the Function body subgraph. */
-  virtual const onnxruntime::Graph& Body() const = 0;
-
-  /** Gets the Mutable Graph instance for the Function body subgraph. */
-  virtual onnxruntime::Graph& MutableBody() = 0;
-};
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/graph/graph.h b/projects/llm_framework/include/onnxruntime/core/graph/graph.h
deleted file mode 100644
index 0fc2762c..00000000
--- a/projects/llm_framework/include/onnxruntime/core/graph/graph.h
+++ /dev/null
@@ -1,1710 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <limits>
-#include <memory>
-#include <string>
-#include <type_traits>
-#include <unordered_map>
-#include <unordered_set>
-
-#ifdef _WIN32
-#pragma warning(push)
-// disable some warnings from protobuf to pass Windows build
-#pragma warning(disable : 4244)
-#endif
-
-#if !defined(ORT_MINIMAL_BUILD)
-#include "onnx/defs/schema.h"
-#include "core/common/inlined_containers.h"
-#else
-#include "onnx/defs/data_type_utils.h"
-#endif
-#include "onnx/onnx_pb.h"
-#include "onnx/onnx-operators_pb.h"
-
-#ifdef _WIN32
-#pragma warning(pop)
-#endif
-
-#include "core/common/gsl.h"
-
-#include "core/common/common.h"
-#include "core/common/const_pointer_container.h"
-#include "core/common/inlined_containers_fwd.h"
-#include "core/common/path.h"
-#include "core/common/span_utils.h"
-#include "core/common/status.h"
-#include "core/common/logging/logging.h"
-#include "core/graph/basic_types.h"
-#include "core/graph/constants.h"
-#include "core/graph/function.h"
-#if !defined(ORT_MINIMAL_BUILD)
-#include "core/graph/function_template.h"
-#endif
-#include "core/graph/graph_nodes.h"
-#include "core/graph/node_arg.h"
-#include "core/graph/ort_format_load_options.h"
-
-namespace flatbuffers {
-class FlatBufferBuilder;
-template <typename T>
-struct Offset;
-}  // namespace flatbuffers
-
-namespace onnxruntime {
-class Graph;
-struct IndexedSubGraph;
-class Model;
-class OpSignature;
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-class RuntimeOptimizationRecordContainer;
-#endif
-
-namespace fbs {
-struct Graph;
-struct Node;
-struct NodeEdge;
-}  // namespace fbs
-
-/**
-@class Node
-Class representing a node in the graph.
-*/
-class Node {
- public:
-  /** Node types */
-  enum class Type {
-    Primitive = 0,  ///< The node refers to a primitive operator.
-    Fused = 1,      ///< The node refers to a function.
-  };
-
-  explicit Node() = default;
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  Node(std::string_view name,
-       std::string_view op_type,
-       std::string_view description,
-       gsl::span<NodeArg* const> input_args,
-       gsl::span<NodeArg* const> output_args,
-       const NodeAttributes* attributes,
-       std::string_view domain) {
-    Init(std::string{name}, std::string{op_type}, std::string{description},
-         std::vector<NodeArg*>{input_args.begin(), input_args.end()},
-         std::vector<NodeArg*>{output_args.begin(), output_args.end()},
-         attributes, std::string{domain});
-  }
-#endif
-
-  ~Node() = default;
-
-  /**
-  @class EdgeEnd
-  Class representing the end of an edge. It could be an input or output edge end of a node.
-  For the node's input edge end, it's the source end, as the destination end is the node itself.
-  For the node's output edge end, it's the destination end, as the source end is the node itself.
-  */
-  class EdgeEnd {
-   public:
-    /**
-    Construct an EdgeEnd
-    @param node The source node if this is an input edge to the current node,
-    or the destination node if this is an output edge from the current node.
-    @param src_arg_index The node arg index of source node of the edge.
-    @param dst_arg_index The node arg index of destination node of the edge.
-    */
-    EdgeEnd(const Node& node, int src_arg_index, int dst_arg_index) noexcept;
-
-    /** Construct a control edge.
-    @param node The node the edge joins to the current node.
-    */
-    explicit EdgeEnd(const Node& node) noexcept;
-
-    /** Gets the Node that this EdgeEnd refers to. */
-    const Node& GetNode() const noexcept { return *node_; }
-
-    /** Gets the source arg index.
-    @returns the source arg index of <*this> edge.*/
-    int GetSrcArgIndex() const { return src_arg_index_; }
-
-    /** Gets the destination arg index.
-    @returns the destination arg index of <*this> edge.*/
-    int GetDstArgIndex() const { return dst_arg_index_; }
-
-   private:
-    const Node* node_;
-    const int src_arg_index_;
-    const int dst_arg_index_;
-  };
-
-  /** Gets the Node's NodeIndex. */
-  NodeIndex Index() const noexcept { return index_; }
-
-  /** Gets the Node's name. */
-  const std::string& Name() const noexcept { return name_; }
-
-  /** Gets the Node's operator type. */
-  const std::string& OpType() const noexcept { return op_type_; }
-
-  /** Gets the domain of the OperatorSet that specifies the operator returned by #OpType.
-   * @remarks If this is an ONNX operator the value will be kOnnxDomain not kOnnxDomainAlias
-   */
-  const std::string& Domain() const noexcept { return domain_; }
-
-  /** Gets the path of the owning model if any. */
-  const Path& ModelPath() const noexcept;
-
-  /** Gets the Node's execution priority.
-  @remarks Lower value means higher priority  */
-  int Priority() const noexcept { return priority_; };
-
-  /** Sets the execution priority of a node.
-  @remarks Lower value means higher priority  */
-  void SetPriority(int priority) noexcept;
-
-  /** Gets the node description. */
-  const std::string& Description() const noexcept { return description_; }
-
-  /** Gets the Node's Node::Type. */
-  Node::Type NodeType() const noexcept { return node_type_; }
-
-  /** Gets the opset version that the Node's operator was first defined in.
-  @returns Opset version. If -1 the Node's operator has not been set.
-  @remarks Prefer over Op()->SinceVersion() as Op() is disabled in a minimal build
-  */
-  int SinceVersion() const noexcept { return since_version_; }
-
-  /** Sets the since version (opset version that the Node's operator was first defined in.) for this node.
-  @remarks Used during layout transformation for setting since version for layout transformed nodes with
-  domain kMSNHWC.
-  */
-  void SetSinceVersion(int since_version) noexcept { since_version_ = since_version; }
-
-#if !defined(ORT_MINIMAL_BUILD)
-  /** Gets the Node's OpSchema.
-  @remarks The graph containing this node must be resolved, otherwise nullptr will be returned. */
-  const ONNX_NAMESPACE::OpSchema* Op() const noexcept { return op_; }
-
-  /** Create a copy of the called op's FunctionProto if it has one. Returns true if successful. */
-  bool TryGetFunctionProto(ONNX_NAMESPACE::FunctionProto& func_proto) const;
-
-  bool CanBeInlined() const;
-
-  /** Gets the function body if applicable otherwise nullptr. */
-  const Function* GetFunctionBody() const noexcept { return func_body_.get(); }
-#endif
-
-  /**
-  Helper to iterate through the container returned by #InputDefs() or #OutputDefs() and call the provided function.
-  @param node_args Collection of NodeArgs returned by #InputDefs() or #OutputDefs()
-  @param func Function to call for each valid NodeArg in the node_args. The function is called with the NodeArg
-              and the index number in the container.
-  @returns common::Status with success or error information.
-  @remarks Returns immediately on error.
-  */
-  static common::Status ForEachWithIndex(const ConstPointerContainer<std::vector<NodeArg*>>& node_args,
-                                         std::function<common::Status(const NodeArg& arg, size_t index)> func) {
-    for (size_t index = 0; index < node_args.size(); ++index) {
-      auto arg = node_args[index];
-      if (!arg->Exists())
-        continue;
-      ORT_RETURN_IF_ERROR(func(*arg, index));
-    }
-    return common::Status::OK();
-  }
-
-  /** Gets the count of arguments for each of the Node's explicit inputs. */
-  const std::vector<int>& InputArgCount() const noexcept { return definitions_.input_arg_count; }
-
-  /** Gets the Node's input definitions.
-  @remarks requires ConstPointerContainer wrapper to apply const to the NodeArg pointers so access is read-only. */
-  ConstPointerContainer<std::vector<NodeArg*>> InputDefs() const noexcept {
-    return ConstPointerContainer<std::vector<NodeArg*>>(definitions_.input_defs);
-  }
-
-  /** Gets the implicit inputs to this Node.
-  If this Node contains a subgraph, these are the NodeArg's that are implicitly consumed by Nodes within that
-  subgraph. e.g. If and Loop operators.*/
-  ConstPointerContainer<std::vector<NodeArg*>> ImplicitInputDefs() const noexcept {
-    return ConstPointerContainer<std::vector<NodeArg*>>(definitions_.implicit_input_defs);
-  }
-
-  /** Gets the Node's output definitions.
-  @remarks requires ConstPointerContainer wrapper to apply const to the NodeArg pointers so access is read-only. */
-  ConstPointerContainer<std::vector<NodeArg*>> OutputDefs() const noexcept {
-    return ConstPointerContainer<std::vector<NodeArg*>>(definitions_.output_defs);
-  }
-
-#if !defined(ORT_MINIMAL_BUILD)
-  /**
-  Helper to iterate through the container returned by #MutableInputDefs() or #MutableOutputDefs() and call the provided function.
-  @param node_args Collection of NodeArgs returned by #MutableInputDefs() or #MutableOutputDefs()
-  @param func Function to call for each valid NodeArg in the node_args. The function is called with the NodeArg
-              and the index number in the container.
-  @returns common::Status with success or error information.
-  @remarks Returns immediately on error.
-  */
-  static common::Status ForEachMutableWithIndex(std::vector<NodeArg*>& node_args,
-                                                std::function<common::Status(NodeArg& arg, size_t index)> func) {
-    for (size_t index = 0; index < node_args.size(); ++index) {
-      auto arg = node_args[index];
-      if (!arg->Exists())
-        continue;
-      ORT_RETURN_IF_ERROR(func(*arg, index));
-    }
-    return common::Status::OK();
-  }
-
-  /** Gets a modifiable collection of the Node's implicit input definitions. */
-  std::vector<NodeArg*>& MutableImplicitInputDefs() noexcept {
-    return definitions_.implicit_input_defs;
-  }
-#endif  // !defined(ORT_MINIMAL_BUILD)
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  /** Gets a modifiable count of arguments for each of the Node's explicit inputs.
-  @todo This should be removed in favor of a method that updates the input args and the count.
-        Currently these operations are separate which is not a good setup. */
-  std::vector<int>& MutableInputArgsCount() { return definitions_.input_arg_count; }
-
-  /** Gets a modifiable collection of the Node's input definitions. */
-  std::vector<NodeArg*>& MutableInputDefs() noexcept {
-    return definitions_.input_defs;
-  }
-
-  /** Gets a modifiable collection of the Node's output definitions. */
-  std::vector<NodeArg*>& MutableOutputDefs() noexcept {
-    return definitions_.output_defs;
-  }
-#endif  // !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-
-  /** Struct to provide sorting between EdgeEnd instances based on NodeIndex first, and NodeArg::Name second. */
-  struct EdgeEndCompare {
-    bool operator()(const EdgeEnd& lhs, const EdgeEnd& rhs) const {
-      if (lhs.GetNode().Index() == rhs.GetNode().Index()) {
-        if (lhs.GetSrcArgIndex() == rhs.GetSrcArgIndex()) {
-          return lhs.GetDstArgIndex() < rhs.GetDstArgIndex();
-        }
-        return lhs.GetSrcArgIndex() < rhs.GetSrcArgIndex();
-      }
-      return lhs.GetNode().Index() < rhs.GetNode().Index();
-    }
-  };
-
-  using EdgeSet = std::set<EdgeEnd, EdgeEndCompare>;
-  using EdgeConstIterator = EdgeSet::const_iterator;
-
-  /**
-  @class NodeConstIterator
-  Class to provide const access to Node instances iterated via an EdgeConstIterator. */
-  class NodeConstIterator {
-   public:
-    NodeConstIterator(EdgeConstIterator p_iter);
-
-    bool operator==(const NodeConstIterator& p_other) const;
-
-    bool operator!=(const NodeConstIterator& p_other) const;
-
-    void operator++();
-    void operator--();
-
-    const Node& operator*() const;
-    const Node* operator->() const;
-
-   private:
-    EdgeConstIterator m_iter;
-  };
-
-  // Functions defined to traverse a Graph as below.
-
-  /** Gets an iterator to the beginning of the input nodes to this Node. */
-  NodeConstIterator InputNodesBegin() const noexcept { return NodeConstIterator(relationships_.input_edges.cbegin()); };
-  /** Gets an iterator to the end of the input nodes to this Node. */
-  NodeConstIterator InputNodesEnd() const noexcept { return NodeConstIterator(relationships_.input_edges.cend()); }
-
-  /** Gets an iterator to the beginning of the output nodes from this Node. */
-  NodeConstIterator OutputNodesBegin() const noexcept {
-    return NodeConstIterator(relationships_.output_edges.cbegin());
-  }
-
-  /** Gets an iterator to the end of the output nodes from this Node. */
-  NodeConstIterator OutputNodesEnd() const noexcept { return NodeConstIterator(relationships_.output_edges.cend()); }
-
-  /** Gets an iterator to the beginning of the input edges to this Node.
-  @remarks There are no nullptr entries in this collection. */
-  EdgeConstIterator InputEdgesBegin() const noexcept { return relationships_.input_edges.cbegin(); }
-
-  /** Gets an iterator to the end of the input edges to this Node. */
-  EdgeConstIterator InputEdgesEnd() const noexcept { return relationships_.input_edges.cend(); }
-
-  /** Gets an iterator to the beginning of the output edges from this Node.
-  @remarks There are no nullptr entries in this collection. */
-  EdgeConstIterator OutputEdgesBegin() const noexcept { return relationships_.output_edges.cbegin(); }
-
-  /** Gets an iterator to the end of the output edges from this Node. */
-  EdgeConstIterator OutputEdgesEnd() const noexcept { return relationships_.output_edges.cend(); }
-
-  /** Gets the Node's control inputs. */
-  const std::set<std::string>& ControlInputs() const noexcept { return relationships_.control_inputs; }
-
-  /** Gets the number of input edges to this Node */
-  size_t GetInputEdgesCount() const noexcept { return relationships_.input_edges.size(); }
-
-  /** Gets the number of output edges from this Node */
-  size_t GetOutputEdgesCount() const noexcept { return relationships_.output_edges.size(); }
-
-  /** Adds an AttributeProto to this Node.
-  @remarks The attribute name is used as the key in the attribute map. */
-  void AddAttributeProto(ONNX_NAMESPACE::AttributeProto value);
-
-  // keep this signature in sync with ADD_ATTR_SINGLE_INTERFACE below
-  /** Adds an attribute to this Node with the specified attribute name and value. */
-  void AddAttribute(std::string attr_name, int64_t value);
-
-  // keep this signature in sync with ADD_ATTR_LIST_INTERFACE below
-  /** Adds an attribute to this Node with the specified attribute name and values. */
-  void AddAttribute(std::string attr_name, gsl::span<const int64_t> values);
-
-#define ADD_ATTR_SINGLE_INTERFACE(Type) \
-  void AddAttribute(std::string attr_name, Type value)
-
-#define ADD_ATTR_LIST_INTERFACE(Type) \
-  void AddAttribute(std::string attr_name, gsl::span<const Type> values)
-
-#define ADD_ATTR_INTERFACES(Type)  \
-  ADD_ATTR_SINGLE_INTERFACE(Type); \
-  ADD_ATTR_LIST_INTERFACE(Type)
-
-  ADD_ATTR_INTERFACES(float);
-  ADD_ATTR_INTERFACES(std::string);
-  ADD_ATTR_INTERFACES(ONNX_NAMESPACE::TensorProto);
-#if !defined(DISABLE_SPARSE_TENSORS)
-  ADD_ATTR_INTERFACES(ONNX_NAMESPACE::SparseTensorProto);
-#endif
-  ADD_ATTR_INTERFACES(ONNX_NAMESPACE::TypeProto);
-
-  ADD_ATTR_SINGLE_INTERFACE(ONNX_NAMESPACE::GraphProto);
-
-#undef ADD_ATTR_SINGLE_INTERFACE
-#undef ADD_ATTR_LIST_INTERFACE
-#undef ADD_ATTR_INTERFACES
-
-  // The below overload is made so the compiler does not attempt to resolve
-  // string literals with the gsl::span overload
-  template <size_t N>
-  void AddAttribute(std::string attr_name, const char (&value)[N]) {
-    this->AddAttribute(std::move(attr_name), std::string(value, N - 1));
-  }
-
-  /** Gets the Node's attributes. */
-  const NodeAttributes& GetAttributes() const noexcept { return attributes_; }
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  /** Remove the specified attribute from this Node */
-  bool ClearAttribute(const std::string& attr_name);
-#endif  // !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-
-#if !defined(ORT_MINIMAL_BUILD)
-  /** Gets the Node's mutable attributes. */
-  NodeAttributes& GetMutableAttributes() noexcept { return attributes_; }
-
-  /** Gets the Graph instance that is instantiated from a GraphProto attribute during Graph::Resolve.
-  @param attr_name Attribute name for the GraphProto attribute.
-  @returns nullptr if the Graph instance has not been instantiated or attribute does not contain a GraphProto.
-  */
-  const Graph* GetGraphAttribute(const std::string& attr_name) const;
-
-  /** Gets the mutable Graph instance that is instantiated from a GraphProto attribute during Graph::Resolve.
-  @param attr_name Attribute name for the GraphProto attribute.
-  @returns nullptr if the Graph instance has not been instantiated or attribute does not contain a GraphProto.
-  */
-  Graph* GetMutableGraphAttribute(const std::string& attr_name);
-#endif  // !defined(ORT_MINIMAL_BUILD)
-
-  /** Checks if the Node contains at least one subgraph (this is the case for control flow operators, such as If, Scan, Loop).
-  @returns true if the Node contains a subgraph.
-  */
-  bool ContainsSubgraph() const {
-    return !attr_to_subgraph_map_.empty();
-  }
-
-  /** Get the const subgraphs from a node.
-  @remarks Creates a new vector so calling ContainsSubgraphs first is preferred. */
-  std::vector<gsl::not_null<const Graph*>> GetSubgraphs() const;
-
-  /** Gets a map of attribute name to the mutable Graph instances for all subgraphs of the Node.
-  @returns Map of the attribute name that defines the subgraph to the subgraph's Graph instance.
-           nullptr if the Node has no subgraphs.
-  */
-  const std::unordered_map<std::string, gsl::not_null<Graph*>>& GetAttributeNameToMutableSubgraphMap() {
-    return attr_to_subgraph_map_;
-  }
-
-  /** Gets a map of attribute name to the const Graph instances for all subgraphs of the Node.
-  @returns Map of the attribute name that defines the subgraph to the subgraph's Graph instance.
-           nullptr if the Node has no subgraphs.
-  */
-  std::unordered_map<std::string, gsl::not_null<const Graph*>> GetAttributeNameToSubgraphMap() const;
-
-  /** Gets the execution ProviderType that this node will be executed by. */
-  ProviderType GetExecutionProviderType() const noexcept { return execution_provider_type_; }
-
-  /** Sets the execution ProviderType that this Node will be executed by. */
-  void SetExecutionProviderType(ProviderType execution_provider_type) {
-    execution_provider_type_ = execution_provider_type;
-  }
-
-  /** Call the provided function for all explicit inputs, implicit inputs, and outputs of this Node.
-      If the NodeArg is an explicit or implicit input, is_input will be true when func is called.
-      @param include_missing_optional_defs Include NodeArgs that are optional and were not provided
-                                           i.e. NodeArg::Exists() == false.
-      */
-  void ForEachDef(std::function<void(const onnxruntime::NodeArg&, bool is_input)> func,
-                  bool include_missing_optional_defs = false) const;
-
-#if !defined(ORT_MINIMAL_BUILD)
-  /** Replaces any matching definitions in the Node's explicit inputs or explicit outputs.
-  @param replacements Map of current NodeArg to replacement NodeArg.
-  */
-  void ReplaceDefs(const std::map<const onnxruntime::NodeArg*, onnxruntime::NodeArg*>& replacements);
-
-  /** Gets the NodeProto representation of this Node.
-  @param update_subgraphs Update the GraphProto values for any subgraphs in the returned NodeProto.
-                          If graph optimization has been run this is most likely required
-                          to ensure the complete Graph is valid.
-  */
-  void ToProto(ONNX_NAMESPACE::NodeProto& proto, bool update_subgraphs = false) const;
-
-  Status SaveToOrtFormat(flatbuffers::FlatBufferBuilder& builder,
-                         flatbuffers::Offset<onnxruntime::fbs::Node>& fbs_node) const;
-
-  flatbuffers::Offset<onnxruntime::fbs::NodeEdge>
-  SaveEdgesToOrtFormat(flatbuffers::FlatBufferBuilder& builder) const;
-
-  void SetFunctionTemplate(const FunctionTemplate& func_template);
-#endif
-
-  static Status LoadFromOrtFormat(const onnxruntime::fbs::Node& fbs_node, Graph& graph,
-                                  const OrtFormatLoadOptions& load_options,
-                                  const logging::Logger& logger, std::unique_ptr<Node>& node);
-
-  Status LoadFromOrtFormat(const onnxruntime::fbs::Node& fbs_node,
-                           const OrtFormatLoadOptions& load_options,
-                           const logging::Logger& logger);
-  Status LoadEdgesFromOrtFormat(const onnxruntime::fbs::NodeEdge& fbs_node_edgs, const Graph& graph);
-
-  /**
-  @class Definitions
-  The input and output definitions for this Node.
-  */
-  class Definitions {
-   public:
-    Definitions() = default;
-
-    /** The Node's explicit input definitions. */
-    std::vector<NodeArg*> input_defs;
-
-    /**
-    The number of inputs for each argument of the operator or function which this node refers.
-    @remarks For example, #input_defs has 10 elements (inputs), and #input_arg_count is {4, 6}.
-    This means that 4 elements (inputs) of input_defs map to the first argument of the operator or function, and
-    the other 6 map to the second argument.
-    */
-    std::vector<int> input_arg_count;
-
-    /** The Node's output definitions. */
-    std::vector<NodeArg*> output_defs;
-
-    /** The Node's implicit input definitions if the Node contains one or more subgraphs
-    (i.e. GraphProto attributes) and the subgraph/s implicitly consume these values.
-    @remarks For example, a subgraph in an 'If' node gets all its input values via this mechanism rather than
-    there being explicit inputs to the 'If' node that are passed to the subgraph.
-    They are pseudo-inputs to this Node as it has an implicit dependency on them. */
-    std::vector<NodeArg*> implicit_input_defs;
-
-    ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(Definitions);
-
-   private:
-  };
-
-  /**
-  @class Relationships
-  Defines the relationships between this Node and other Nodes in the Graph.
-  */
-  class Relationships {
-   public:
-    Relationships() = default;
-
-    void Clear() noexcept {
-      input_edges.clear();
-      output_edges.clear();
-      control_inputs.clear();
-    }
-
-    /** The edges for Nodes that provide inputs to this Node. */
-    EdgeSet input_edges;
-
-    /** The edges for Nodes that receive outputs from this Node. */
-    EdgeSet output_edges;
-
-    /** The Node names of the control inputs to this Node. */
-    std::set<std::string> control_inputs;
-
-   private:
-    ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(Relationships);
-  };
-
-  // NOTE: This friendship relationship should ONLY be used for calling methods of the Node class and not accessing
-  // the data members directly, so that the Node can maintain its internal invariants.
-  friend class Graph;
-  Node(NodeIndex index, Graph& graph) : index_(index), graph_(&graph) {}
-
- private:
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(Node);
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  void Init(const std::string& name,
-            const std::string& op_type,
-            const std::string& description,
-            const std::vector<NodeArg*>& input_args,
-            const std::vector<NodeArg*>& output_args,
-            const NodeAttributes* attributes,
-            const std::string& domain);
-
-  // internal only method to allow selected classes to directly alter the input/output definitions and arg counts
-  Definitions& MutableDefinitions() noexcept;
-
-  // internal only method to allow selected classes to directly alter the links between nodes.
-  Relationships& MutableRelationships() noexcept;
-
-  void SetNodeType(Node::Type node_type) noexcept { node_type_ = node_type; }
-#endif
-
-  // create a Graph instance for an attribute that contains a GraphProto
-  void CreateSubgraph(const std::string& attr_name);
-
-  const std::vector<std::unique_ptr<Graph>>& MutableSubgraphs() noexcept { return subgraphs_; }
-
-  // validate and update the input arg count
-  common::Status UpdateInputArgCount();
-
-  const Definitions& GetDefinitions() const noexcept { return definitions_; }
-  const Relationships& GetRelationships() const noexcept { return relationships_; }
-
-  // Node index. Default to impossible value rather than 0.
-  NodeIndex index_ = std::numeric_limits<NodeIndex>::max();
-
-  // Node name.
-  std::string name_;
-
-  // Node operator type.
-  std::string op_type_;
-
-  // OperatorSet domain of op_type_.
-  std::string domain_;
-
-#if !defined(ORT_MINIMAL_BUILD)
-  // OperatorSchema that <*this> node refers to.
-  const ONNX_NAMESPACE::OpSchema* op_ = nullptr;
-
-  // Reference to the function template defined in the model.
-  const FunctionTemplate* func_template_ = nullptr;
-#endif
-
-  // Execution priority, lower value for higher priority
-  int priority_ = 0;
-
-  // set from op_->SinceVersion() or via deserialization when OpSchema is not available
-  int since_version_ = -1;
-
-  Node::Type node_type_ = Node::Type::Primitive;
-
-  // The function body is owned by graph_
-  std::unique_ptr<Function> func_body_ = nullptr;
-
-  // Node doc string.
-  std::string description_;
-
-  // input/output defs and arg count
-  Definitions definitions_;
-
-  // Relationships between this node and others in the graph
-  Relationships relationships_;
-
-  // Device.
-  std::string execution_provider_type_;
-
-  // Map from attribute name to attribute.
-  // This allows attribute adding and removing.
-  NodeAttributes attributes_;
-
-  // Graph that contains this Node
-  Graph* graph_ = nullptr;
-
-  // Map of attribute name to the Graph instance created from the GraphProto attribute
-  std::unordered_map<std::string, gsl::not_null<Graph*>> attr_to_subgraph_map_;
-
-  // Graph instances for subgraphs that are owned by this Node
-  std::vector<std::unique_ptr<Graph>> subgraphs_;
-};
-
-/**
-@class Graph
-The Graph representation containing the graph inputs and outputs, the Node instances,
-and the edges connecting the nodes.
-*/
-class Graph {
- public:
-  /** Gets the Graph name. */
-  const std::string& Name() const noexcept;
-
-  /** Gets the Graph description. */
-  const std::string& Description() const noexcept;
-
-  /** Gets the path of the owning model, if any. */
-  const Path& ModelPath() const;
-
-  /** Returns true if this is a subgraph or false if it is a high-level graph. */
-  bool IsSubgraph() const { return parent_graph_ != nullptr; }
-
-  /** Returns the parent graph if this is a subgraph */
-  const Graph* ParentGraph() const { return parent_graph_; }
-
-  /** Returns the mutable parent graph if this is a subgraph */
-  Graph* MutableParentGraph() { return parent_graph_; }
-
-  /** Returns the strict_shape_type_inference that was passed into the constructor. */
-  bool StrictShapeTypeInference() const { return strict_shape_type_inference_; }
-
-#if !defined(ORT_MINIMAL_BUILD)
-  /** Sets the Graph name. */
-  void SetName(const std::string& name);
-
-  /** Gets the Graph description. */
-  void SetDescription(const std::string& description);
-
-  /** Replaces the initializer tensor with the same name as the given initializer tensor.
-  The replacement initializer tensor must have the same type and shape as the existing initializer tensor.
-
-  Note: This currently has linear time complexity. There is room for improvement but it would likely require changes to
-  how initializer tensors are stored and tracked.
-  */
-  common::Status ReplaceInitializedTensor(ONNX_NAMESPACE::TensorProto new_initializer);
-
-#if !defined(DISABLE_EXTERNAL_INITIALIZERS)
-  /** This function takes externally provided data for initializers with external data
-   *    and replaces graph initializers with its content.
-   */
-  common::Status InjectExternalInitializedTensors(const InlinedHashMap<std::string, OrtValue>& external_initializers);
-#endif  // !defined(DISABLE_EXTERNAL_INITIALIZERS)
-
-#endif  // !defined(ORT_MINIMAL_BUILD)
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  /** Add an initializer tensor to the Graph. */
-  void AddInitializedTensor(const ONNX_NAMESPACE::TensorProto& tensor_proto);
-#endif
-
-  /** Remove the initializer tensor with the provided name from the Graph. */
-  void RemoveInitializedTensor(const std::string& tensor_name);
-
-  /** Check if a given name is an initializer tensor's name in this graph. */
-  bool IsInitializedTensor(const std::string& name) const;
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-  /** Check if a given name is a sparse initializer's name in the model
-   * we currently convert sparse_initializer field in the model into dense Tensor instances.
-   * However, we sometimes want to check if this initializer was stored as sparse in the model.
-   */
-  bool IsSparseInitializer(const std::string& name) const;
-#endif
-
-  /** Gets an initializer tensor with the provided name.
-  @param[out] value Set to the TensorProto* if the initializer is found, or nullptr if not.
-  @returns True if found.
-  */
-  bool GetInitializedTensor(const std::string& tensor_name, const ONNX_NAMESPACE::TensorProto*& value) const;
-
-  /** Gets all the initializer tensors in this Graph. */
-  const InitializedTensorSet& GetAllInitializedTensors() const noexcept { return name_to_initial_tensor_; }
-
-  /** Removes all initializer tensors from this Graph and releases the memory they were using. */
-  void CleanAllInitializedTensors() noexcept;
-
-  /** Returns true if an initializer value can be overridden by a graph input with the same name. */
-  bool CanOverrideInitializer() const noexcept { return ir_version_ >= 4; }
-
-  /** returns the initializer's TensorProto if 'name' is an initializer, is constant and
-  cannot be overridden at runtime. If the initializer is not found or is not constant, a nullptr is returned.
-  @param check_outer_scope If true and the graph is a subgraph,
-         check ancestor graph/s for 'name' if not found in 'graph'.
-  @remarks check_outer_scope of true is not supported in a minimal build
-  */
-  const ONNX_NAMESPACE::TensorProto* GetConstantInitializer(const std::string& name, bool check_outer_scope) const;
-
-  /** returns the initializer's TensorProto if 'name' is an initializer (both constant and overridable).
-  If the initializer is not found, a nullptr is returned.
-  @param check_outer_scope If true and the graph is a subgraph,
-         check ancestor graph/s for 'name' if not found in 'graph'.
-  @remarks check_outer_scope of true is not supported in a minimal build
-  */
-  const ONNX_NAMESPACE::TensorProto* GetInitializer(const std::string& name, bool check_outer_scope) const;
-
-  /** Gets the Graph inputs excluding initializers.
-  These are the required inputs to the Graph as the initializers can be optionally overridden via graph inputs.
-  @remarks Contains no nullptr values. */
-  const std::vector<const NodeArg*>& GetInputs() const noexcept { return graph_inputs_excluding_initializers_; }
-
-  /** Gets the Graph inputs including initializers.
-  This is the full set of inputs, in the same order as defined in the GraphProto.
-  @remarks Contains no nullptr values. */
-  const std::vector<const NodeArg*>& GetInputsIncludingInitializers() const noexcept {
-    return graph_inputs_including_initializers_;
-  }
-
-  /** Return true if "node_arg" is a input or an initializer. Otherwise, returns false. */
-  bool IsInputsIncludingInitializers(const NodeArg* node_arg) const noexcept {
-    return std::find(graph_inputs_including_initializers_.begin(),
-                     graph_inputs_including_initializers_.end(), node_arg) != graph_inputs_including_initializers_.end();
-  }
-
-  /** Gets the Graph inputs that are initializers
-  These are overridable initializers. This is a difference between
-  graph_inputs_including_initializers_ and graph_inputs_excluding_initializers_
-  @remarks Contains no nullptr values. */
-  const std::vector<const NodeArg*>& GetOverridableInitializers() const {
-    return graph_overridable_initializers_;
-  }
-
-  /** Gets the Graph outputs.
-  @remarks Contains no nullptr values.*/
-  const std::vector<const NodeArg*>& GetOutputs() const noexcept { return graph_outputs_; }
-
-  bool IsOutput(const NodeArg* node_arg) const noexcept {
-    return std::find(graph_outputs_.begin(), graph_outputs_.end(), node_arg) != graph_outputs_.end();
-  }
-
-  /** Returns true if one or more of the Node outputs are Graph outputs.
-  @remarks Cheaper than calling GetNodeOutputsInGraphOutputs.
-  */
-  bool NodeProducesGraphOutput(const Node& node) const {
-    auto end_outputs = graph_outputs_.cend();
-    for (auto output_def : node.OutputDefs()) {
-      if (std::find(graph_outputs_.cbegin(), end_outputs, output_def) != end_outputs) {
-        return true;
-      }
-    }
-    return false;
-  }
-
-  /** Returns a vector with the indexes of the outputs of the given Node that are also Graph outputs. */
-  std::vector<int> GetNodeOutputsInGraphOutputs(const Node& node) const {
-    int output_idx = 0;
-    std::vector<int> indexes;
-    for (auto output_def : node.OutputDefs()) {
-      if (std::find(GetOutputs().cbegin(), GetOutputs().cend(), output_def) != GetOutputs().cend()) {
-        indexes.push_back(output_idx);
-      }
-
-      ++output_idx;
-    }
-
-    return indexes;
-  }
-
-  /** Gets the NodeArgs that represent value_info instances in the Graph.
-  These are the values that are neither Graph inputs nor outputs.
-  @remarks Contains no nullptr values. */
-  const std::unordered_set<const NodeArg*>& GetValueInfo() const noexcept { return value_info_; }
-
-#if !defined(ORT_MINIMAL_BUILD)
-  void AddValueInfo(const NodeArg* new_value_info);
-#endif
-
-  /** Gets the Node with the specified node index.
-  @returns Node instance if found. nullptr if node_index is invalid or node has been freed.
-  */
-  const Node* GetNode(NodeIndex node_index) const { return NodeAtIndexImpl(node_index); }
-
-  /** Gets the mutable Node with the specified node index.
-  @returns Mutable Node instance if found. nullptr if node_index is invalid or node has been freed.
-  */
-  Node* GetNode(NodeIndex node_index) { return NodeAtIndexImpl(node_index); }
-
-  /** Get a GraphNodes instance that provides mutable access to all valid Nodes in the Graph. */
-  GraphNodes& Nodes() noexcept { return iterable_nodes_; }
-
-  /** Get a GraphNodes instance that provides const access to all valid Nodes in the Graph. */
-  const GraphNodes& Nodes() const noexcept { return iterable_nodes_; }
-
-  /** Get a ConstGraphNodes instance that provides access to a filtered set of valid Nodes in the Graph.
-  @remarks We can't use GraphNodes as that would provide mutable access to the nodes by default, and we can't prevent
-           that by returning a const instance of GraphNodes as we're creating a new instance here due to the filter
-           being something we don't control (i.e. we have to return a new instance so it can't be const).
-  */
-  ConstGraphNodes FilteredNodes(GraphNodes::NodeFilterFunc&& filter_func) const noexcept {
-    return ConstGraphNodes(nodes_, std::move(filter_func));
-  }
-
-  /** Gets the maximum NodeIndex value used in the Graph.
-  WARNING: This actually returns the max index value used + 1.
-  */
-  int MaxNodeIndex() const noexcept { return static_cast<int>(nodes_.size()); }  // assume the casting won't overflow
-
-  /** Gets the number of valid Nodes in the Graph.
-  @remarks This may be smaller than MaxNodeIndex(), as Nodes may be removed during optimization.
-  */
-  int NumberOfNodes() const noexcept { return num_of_nodes_; }
-
-  /** Gets the mutable NodeArg with the provided name.
-  @returns Pointer to NodeArg if found, nullptr if not. */
-  NodeArg* GetNodeArg(const std::string& name) {
-    auto iter = node_args_.find(name);
-    if (iter != node_args_.end()) {
-      return iter->second.get();
-    }
-    return nullptr;
-  }
-
-  /** Gets the const NodeArg with the provided name.
-  @returns Pointer to const NodeArg if found, nullptr if not. */
-  const NodeArg* GetNodeArg(const std::string& name) const {
-    return const_cast<Graph*>(this)->GetNodeArg(name);
-  }
-
-  // search this and up through any parent_graph_ instance for a NodeArg
-  NodeArg* GetNodeArgIncludingParentGraphs(const std::string& node_arg_name);
-
-  /** Gets a mutable NodeArg by name. Creates a new NodeArg that is owned by this Graph if not found.
-  @param name The NodeArg name.
-  @param[in] p_arg_type Optional TypeProto to use if the NodeArg needs to be created.
-  @returns NodeArg reference.
-  */
-  NodeArg& GetOrCreateNodeArg(const std::string& name, const ONNX_NAMESPACE::TypeProto* p_arg_type) {
-    auto iter = node_args_.find(name);
-    if (iter != node_args_.end()) {
-      return *(iter->second);
-    }
-    auto result = node_args_.insert(std::make_pair(name, std::make_unique<NodeArg>(name, p_arg_type)));
-    return *(result.first->second);
-  }
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  /** Generate a unique name in this Graph for a NodeArg */
-  std::string GenerateNodeArgName(const std::string& base_name);
-
-  /** Generate a unique name in this Graph for a Node */
-  std::string GenerateNodeName(const std::string& base_name);
-#endif  // !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-
-#if !defined(ORT_MINIMAL_BUILD)
-  /** Copy a Node and add it to this Graph.
-  @param other Node to copy
-  @returns Reference to the Node that was created and added to this Graph.
-  @remarks Do not call AddNode and Remove Node concurrently as they are not thread-safe.
-  */
-  Node& AddNode(const Node& other);
-#endif
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  /** Add a Node to this Graph.
-  @param name The Node name. Must be unique in this Graph.
-  @param op_type The operator type. e.g. ONNX operator name.
-  @param description Arbitrary description of the Node.
-  @param input_args The explicit inputs to this Node.
-  @param output_args The outputs from this Node.
-  @param attributes Optional NodeAttributes to add.
-  @param domain The domain for the op_type.
-  @returns Reference to the new Node.
-  @remarks Do not call AddNode and Remove Node concurrently as they are not thread-safe.
-  */
-  Node& AddNode(const std::string& name,
-                const std::string& op_type,
-                const std::string& description,
-                gsl::span<NodeArg* const> input_args,
-                gsl::span<NodeArg* const> output_args,
-                const NodeAttributes* attributes = nullptr,
-                const std::string& domain = kOnnxDomain);
-
-  Node& AddNode(const std::string& name,
-                const std::string& op_type,
-                const std::string& description,
-                std::initializer_list<NodeArg*> input_args,
-                std::initializer_list<NodeArg*> output_args,
-                const NodeAttributes* attributes = nullptr,
-                const std::string& domain = kOnnxDomain) {
-    return AddNode(name, op_type, description,
-                   AsSpan(input_args),
-                   AsSpan(output_args),
-                   attributes, domain);
-  }
-
-  Node& AddNode(const std::string& name,
-                const std::string& op_type,
-                const std::string& description,
-                gsl::span<NodeArg* const> input_args,
-                std::initializer_list<NodeArg*> output_args,
-                const NodeAttributes* attributes = nullptr,
-                const std::string& domain = kOnnxDomain) {
-    return AddNode(name, op_type, description,
-                   input_args,
-                   AsSpan(output_args),
-                   attributes, domain);
-  }
-
-  Node& AddNode(const std::string& name,
-                const std::string& op_type,
-                const std::string& description,
-                std::initializer_list<NodeArg*> input_args,
-                gsl::span<NodeArg* const> output_args,
-                const NodeAttributes* attributes = nullptr,
-                const std::string& domain = kOnnxDomain) {
-    return AddNode(name, op_type, description,
-                   AsSpan(input_args),
-                   output_args,
-                   attributes, domain);
-  }
-
-  /** Remove a Node from this Graph and free it.
-  The output edges of this specified node MUST have been removed before removing the node.
-  The input edges of this specified node is removed while removing the node. The process of
-  removing a node from a graph should be,
-  1. Remove out edges of this specified node.
-  2. Remove this specified node.
-  3. Add new input edges connected with all out nodes.
-  @returns true if the node_index was valid
-  @remarks Do not call AddNode and Remove Node concurrently as they are not thread-safe.
-  */
-  bool RemoveNode(NodeIndex node_index);
-
-  /** Add an edge between two Nodes.
-  @param src_node_index NodeIndex of source Node that is providing output to the destination Node.
-  @param dst_node_index NodeIndex of destination Node that is receiving input from the source Node.
-  @param src_arg_index node arg index of source node.
-  @param dst_arg_index node arg index of destination node.
-  */
-  void AddEdge(NodeIndex src_node_index, NodeIndex dst_node_index, int src_arg_index, int dst_arg_index);
-
-  /** Remove an edge between two Nodes.
-  @param src_node_index NodeIndex of source Node to remove an output edge from.
-  @param dst_node_index NodeIndex of destination Node to remove an input edge from.
-  @param src_arg_index node arg index of source node.
-  @param dst_arg_index node arg index of destination node.
-  */
-  void RemoveEdge(NodeIndex src_node_index, NodeIndex dst_node_index, int src_arg_index, int dst_arg_index);
-#endif
-
-#if !defined(ORT_MINIMAL_BUILD)
-  /**
-  Add a control edge between two Nodes in this Graph.
-  The source Node does not produce output that is directly consumed by the destination Node, however the
-  destination Node must execute after the source node. The control edge allows this ordering to occur.
-  */
-  bool AddControlEdge(NodeIndex src_node_index, NodeIndex dst_node_index);
-#endif  // !defined(ORT_MINIMAL_BUILD)
-
-  /** Mark the Graph as needing Resolve() to be called.
-  This should be done after modifying any aspect of the Graph that changes the Nodes or relationships between them. */
-  Graph& SetGraphResolveNeeded() noexcept {
-    graph_resolve_needed_ = true;
-    return *this;
-  }
-
-  /** Gets flag indicating whether Graph::Resolve needs to be called before using the Graph. */
-  bool GraphResolveNeeded() const noexcept {
-    return graph_resolve_needed_;
-  }
-
-  /** Sets flag that Graph::graph_proto_ needs to be updated to reflect changes in the Graph. */
-  Graph& SetGraphProtoSyncNeeded() noexcept {
-    graph_proto_sync_needed_ = true;
-    return *this;
-  }
-
-  /** Gets flag indicating whether Graph::graph_proto_ needs to be synchronized with this Graph instance. */
-  bool GraphProtoSyncNeeded() const noexcept {
-    return graph_proto_sync_needed_;
-  }
-
-  /** Performs a reverse depth-first search (DFS) traversal from a set of nodes, via their inputs,
-  up to their source node/s.
-  @param from NodeIndex values for a set of Nodes to traverse from.
-  @param enter Visit function that will be invoked on a node when it is visited but its parents haven't been.
-  @param leave Visit function invoked on the node after its parents have all been visited.
-  @param comp Comparison function to stabilize the traversal order by making Node ordering deterministic.
-  */
-  void ReverseDFSFrom(gsl::span<NodeIndex const> from,
-                      const std::function<void(const Node*)>& enter,
-                      const std::function<void(const Node*)>& leave,
-                      const std::function<bool(const Node*, const Node*)>& comp = {}) const;
-
-  /** Performs a reverse depth-first search (DFS) traversal from a set of nodes, via their inputs,
-  up to their source node/s.
-  @param from Set of Nodes to traverse from.
-  @param enter Visit function that will be invoked on a node when it is visited but its parents haven't been.
-  @param leave Visit function invoked on the node after its parents have all been visited.
-  @param comp Comparison function to stabilize the traversal order by making Node ordering deterministic.
-  */
-  void ReverseDFSFrom(gsl::span<const Node* const> from,
-                      const std::function<void(const Node*)>& enter,
-                      const std::function<void(const Node*)>& leave,
-                      const std::function<bool(const Node*, const Node*)>& comp = {}) const;
-
-  /** Performs a reverse depth-first search (DFS) traversal from a set of nodes, via their inputs,
-  up to their source node/s.
-  @param from Set of Nodes to traverse from.
-  @param enter Visit function that will be invoked on a node when it is visited but its parents haven't been.
-  @param leave Visit function invoked on the node after its parents have all been visited.
-  @param stop Stop traversal from node n to input node p if stop(n, p) is true.
-  @param comp Comparison function to stabilize the traversal order by making Node ordering deterministic.
-  */
-  void ReverseDFSFrom(gsl::span<const Node* const> from,
-                      const std::function<void(const Node*)>& enter,
-                      const std::function<void(const Node*)>& leave,
-                      const std::function<bool(const Node*, const Node*)>& comp,
-                      const std::function<bool(const Node*, const Node*)>& stop) const;
-
-#if !defined(ORT_MINIMAL_BUILD)
-  /** Performs topological sort with Kahn's algorithm on the graph/s.
-  @param enter Visit function that will be invoked on a node when it is visited.
-  @param comp Comparison function to stabilize the traversal order by making Node ordering deterministic.
-  */
-  void KahnsTopologicalSort(const std::function<void(const Node*)>& enter,
-                            const std::function<bool(const Node*, const Node*)>& comp) const;
-
-#endif
-
-  /** Gets the map of operator domains to their opset versions. */
-  const std::unordered_map<std::string, int>& DomainToVersionMap() const noexcept {
-    return domain_to_version_;
-  }
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  /**
-  Create a single Node that will be the result of the a fusion of multiple nodes in this Graph.
-  @param sub_graph A IndexSubGraph instance with details of the nodes to fuse.
-  @param fused_node_name The name for the new Node.
-  @returns Node with fused subgraph.
-  @remarks As a new Graph instance for the fused nodes is not created, a GraphViewer can be constructed with the
-           IndexedSubGraph information to provide a view of the subgraph. The original nodes are left in place
-           while this is in use.
-           Call FinalizeFuseSubGraph to remove them once the fused replacement node is fully created.
-  */
-  Node& BeginFuseSubGraph(const IndexedSubGraph& sub_graph, const std::string& fused_node_name);
-
-  void FinalizeFuseSubGraph(const IndexedSubGraph& sub_graph, Node& fused_node);
-#endif
-
-#if !defined(ORT_MINIMAL_BUILD)
-  /** Gets the GraphProto representation of this Graph. */
-  const ONNX_NAMESPACE::GraphProto& ToGraphProto();
-  ONNX_NAMESPACE::GraphProto ToGraphProto() const;
-
-  /** Gets the GraphProto representation of this Graph
-  @params external_file_name name of the binary file to use for initializers
-  @param initializer_size_threshold initializers larger or equal to this threshold (in bytes) are saved
-  in the external file. Initializer smaller than this threshold are included in the onnx file.
-  @returns GraphProto serialization of the graph.
-  */
-  ONNX_NAMESPACE::GraphProto ToGraphProtoWithExternalInitializers(const std::string& external_file_name,
-                                                                  size_t initializer_size_threshold) const;
-
-  /** Gets the ISchemaRegistry instances being used with this Graph. */
-  IOnnxRuntimeOpSchemaCollectionPtr GetSchemaRegistry() const;
-
-  /**
-  Looks up the op schema in the schema registry and sets it for the given node.
-  @param node The node to update.
-  @return Whether the node's op schema was set to a valid value.
-  */
-  bool SetOpSchemaFromRegistryForNode(Node& node);
-
-  /**
-  Create a single Function based Node that is the result of the a fusion of multiple nodes in this Graph.
-  A new Graph instance will be created for the fused nodes.
-  @param sub_graph A IndexSubGraph instance with details of the nodes to fuse. Ownership is transferred to the new Node
-  @param fused_node_name The name for the new Node.
-  @returns Function based Node with fused subgraph. The Node body will contain a Function instance.
-  */
-  Node& FuseSubGraph(const IndexedSubGraph& sub_graph, const std::string& fused_node_name);
-
-  /**
-  Directly insert the nodes in the function Node provided into this Graph.
-  @param node Node with Node::Type of Node::Type::Fused
-  @returns Status indicating success or providing an error message.
-  */
-  Status InlineFunction(Node& node);
-
-  /** Mark a NodeArg name as coming from the outer scope when programmatically constructing a Graph that will
-  be used as a GraphProto attribute in another Node..
-  e.g. when creating a Graph instance that will be used as a subgraph in a control flow operator, it is necessary to
-  define placeholder NodeArgs for outer scope values. This prevents these values from becoming explicit graph inputs
-  when the Graph is resolved.
-  */
-  void AddOuterScopeNodeArg(const std::string& name) {
-    ORT_IGNORE_RETURN_VALUE(outer_scope_node_arg_names_.insert(name));
-  }
-
-  /** Explicitly set graph inputs.
-  @param inputs NodeArgs that represent complete graph inputs which need to be explicitly ordered.
-  @remarks Note that the input order matters for subgraphs.
-  */
-  void SetInputs(gsl::span<const NodeArg* const> inputs);
-
-  void SetInputs(std::initializer_list<const NodeArg*> inputs) {
-    SetInputs(AsSpan(inputs));
-  }
-
-  const Model& GetModel() const {
-    return owning_model_;
-  }
-
-  const logging::Logger& GetLogger() const {
-    return logger_;
-  }
-
-  /** Explicitly set graph outputs.
-  @param outputs NodeArgs that represent complete graph outputs which need to be explicitly ordered.
-  @remarks Note that the output order matters for subgraphs.
-  */
-  void SetOutputs(gsl::span<const NodeArg* const> outputs);
-
-  void SetOutputs(std::initializer_list<const NodeArg*> outputs) {
-    SetOutputs(AsSpan(outputs));
-  }
-
-#endif  // !defined(ORT_MINIMAL_BUILD)
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  /** Sets the type of a NodeArg, replacing existing type/shape if any */
-  void SetNodeArgType(NodeArg& arg, const ONNX_NAMESPACE::TypeProto& type_proto);
-
-  const Node* GetProducerNode(const std::string& node_arg_name) const {
-    return GetProducerNodeImpl(*this, node_arg_name);
-  }
-
-  Node* GetMutableProducerNode(const std::string& node_arg_name) {
-    return GetProducerNodeImpl(*this, node_arg_name);
-  }
-
-  void UpdateProducerNode(const std::string& node_arg_name, NodeIndex node_index) {
-    auto iter = node_arg_to_producer_node_.find(node_arg_name);
-
-    if (iter != node_arg_to_producer_node_.end()) {
-      iter->second = node_index;
-    } else {
-      node_arg_to_producer_node_[node_arg_name] = node_index;
-    }
-  }
-
-  std::vector<const Node*> GetConsumerNodes(const std::string& node_arg_name) const {
-    return GetConsumerNodesImpl(*this, node_arg_name);
-  }
-
-  // Without removing the existing consumers, add a consumer to the give node arg name.
-  void AddConsumerNode(const std::string& node_arg_name, Node* consumer) {
-    node_arg_to_consumer_nodes_[node_arg_name].insert(consumer->Index());
-  }
-
-  // Remove a consumer from the set
-  void RemoveConsumerNode(const std::string& node_arg_name, Node* consumer) {
-    node_arg_to_consumer_nodes_[node_arg_name].erase(consumer->Index());
-  }
-#endif  // !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-
-#if !defined(ORT_MINIMAL_BUILD)
-  std::vector<Node*> GetMutableConsumerNodes(const std::string& node_arg_name) {
-    return GetConsumerNodesImpl(*this, node_arg_name);
-  }
-
-  void UpdateConsumerNodes(const std::string& node_arg_name, gsl::span<Node* const> nodes) {
-    // Replace nodes for the arg
-    auto& nodes_for_arg = node_arg_to_consumer_nodes_[node_arg_name];
-    if (!nodes_for_arg.empty()) {
-      nodes_for_arg.clear();
-    }
-
-    nodes_for_arg.reserve(nodes.size());
-    for (Node* node : nodes) {
-      nodes_for_arg.insert(node->Index());
-    }
-  }
-
-  void UpdateConsumerNodes(const std::string& node_arg_name, std::initializer_list<Node*> nodes) {
-    UpdateConsumerNodes(node_arg_name, AsSpan(nodes));
-  }
-
-  /** During constant folding it may become possible to infer the shape for a node.
-      To avoid running a full Resolve allow an individual node to have the shape inferencing re-run.
-  */
-  Status UpdateShapeInference(Node& node);
-
-  // Options to control Graph::Resolve.
-  struct ResolveOptions {
-    // Whether to override existing types with inferred types.
-    bool override_types = false;
-    // Names of initializers to keep even if unused (optional).
-    const std::unordered_set<std::string>* initializer_names_to_preserve = nullptr;
-    // Whether to set that no proto sync is required after resolving.
-    // Useful for resolving right after loading from a GraphProto.
-    bool no_proto_sync_required = false;
-  };
-
-  /**
-  Resolve this Graph to ensure it is completely valid, fully initialized, and able to be executed.
-  1. Run through all validation rules.
-    a. Node name and node output's names should be unique.
-    b. Attribute match between node and op definition.
-    c. Input/Output match between node and op definition.
-    d. Graph is acyclic and sort nodes in topological order.
-  2. Check & Setup inner nodes' dependency.
-  3. Cleanup function definition lists.
-  Note: the weights for training can't be cleaned during resolve.
-  @returns common::Status with success or error information.
-  */
-  common::Status Resolve(const ResolveOptions& options);
-
-  common::Status Resolve() {
-    ResolveOptions default_options;
-    return Resolve(default_options);
-  }
-
-  const std::unordered_set<std::string>& GetOuterScopeNodeArgNames() const noexcept {
-    return outer_scope_node_arg_names_;
-  }
-
-  common::Status SaveToOrtFormat(flatbuffers::FlatBufferBuilder& builder,
-                                 flatbuffers::Offset<onnxruntime::fbs::Graph>& fbs_graph) const;
-
-#endif  // !defined(ORT_MINIMAL_BUILD)
-
-  /** Returns the Node containing the GraphProto for this Graph instance if IsSubgraph is true */
-  const Node* ParentNode() const { return parent_node_; }
-
-  /** Returns true if the name is for a value that is coming from outer scope */
-  bool IsOuterScopeValue(const std::string& name) const {
-    if (!parent_node_) return false;
-    const auto& implicit_input_defs = parent_node_->ImplicitInputDefs();
-    return std::any_of(implicit_input_defs.cbegin(), implicit_input_defs.cend(),
-                       [&name](const NodeArg* implicit_input) {
-                         return implicit_input->Name() == name;
-                       });
-  }
-
-#if !defined(ORT_MINIMAL_BUILD)
-  /** Construct a Graph instance for a subgraph that is created from a GraphProto attribute in a Node.
-  Inherits some properties from the parent graph.
-  @param parent_graph The Graph containing the Node that has the GraphProto attribute.
-  @param parent_node The Node that has the GraphProto attribute.
-  @param subgraph_proto The GraphProto from the Node attribute.
-  */
-  Graph(Graph& parent_graph, const Node& parent_node, ONNX_NAMESPACE::GraphProto& subgraph_proto);
-
-  Graph(const Model& owning_model,
-        IOnnxRuntimeOpSchemaCollectionPtr schema_registry,
-        ONNX_NAMESPACE::GraphProto& subgraph_proto,
-        const std::unordered_map<std::string, int>& domain_version_map,
-        const logging::Logger& logger,
-        bool strict_shape_type_inference);
-#endif
-
-  virtual ~Graph();
-
-  static Status LoadFromOrtFormat(const onnxruntime::fbs::Graph& fbs_graph, const Model& owning_model,
-                                  const std::unordered_map<std::string, int>& domain_to_version,
-#if !defined(ORT_MINIMAL_BUILD)
-                                  IOnnxRuntimeOpSchemaCollectionPtr schema_registry,
-#endif
-                                  const OrtFormatLoadOptions& load_options,
-                                  const logging::Logger& logger, std::unique_ptr<Graph>& graph);
-
-  // deserialize a subgraph
-  static Status LoadFromOrtFormat(const onnxruntime::fbs::Graph& fbs_graph,
-                                  Graph& parent_graph, const Node& parent_node,
-                                  const OrtFormatLoadOptions& load_options,
-                                  const logging::Logger& logger, std::unique_ptr<Graph>& graph);
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  const RuntimeOptimizationRecordContainer& RuntimeOptimizations() const {
-    return runtime_optimizations_;
-  }
-
-  RuntimeOptimizationRecordContainer& MutableRuntimeOptimizations() {
-    return runtime_optimizations_;
-  }
-#endif  // !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-
-  // This friendship relationship should only be used to call Graph::Graph and
-  // Graph::LoadGraph All other access should be via the public API.
-  friend class Model;
-
-  Graph() = delete;
-
-  // Create empty Graph instance to re-create from ORT format serialized data.
-  Graph(const Model& owning_model,
-        const std::unordered_map<std::string, int>& domain_to_version,
-#if !defined(ORT_MINIMAL_BUILD)
-        IOnnxRuntimeOpSchemaCollectionPtr schema_registry,
-#endif
-        Graph* parent_graph, const Node* parent_node,
-        const logging::Logger& logger,
-        bool strict_shape_type_inference);
-
-  // Populate Graph instance from ORT format serialized data.
-  Status LoadFromOrtFormat(const onnxruntime::fbs::Graph& fbs_graph,
-                           const OrtFormatLoadOptions& load_options);
-
-#if !defined(ORT_MINIMAL_BUILD)
-  // Constructor: Given a <GraphProto> loaded from model file, construct
-  // a <Graph> object. Used by Model to create a Graph instance.
-  Graph(const Model& owning_model,
-        ONNX_NAMESPACE::GraphProto* graph_proto,
-        const std::unordered_map<std::string, int>& domain_to_version,
-        Version ir_version,
-        IOnnxRuntimeOpSchemaCollectionPtr schema_registry,
-        const logging::Logger& logger,
-        bool strict_shape_type_inference);
-
-  // internal use by the Graph class only
-  Graph(const Model& owning_model,
-        ONNX_NAMESPACE::GraphProto* graph_proto,
-        const std::unordered_map<std::string, int>& domain_to_version,
-        Version ir_version,
-        IOnnxRuntimeOpSchemaCollectionPtr schema_registry,
-        Graph* parent_graph,
-        const Node* parent_node,
-        const logging::Logger& logger,
-        bool strict_shape_type_inference);
-
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(Graph);
-
- private:
-  void InitializeStateFromModelFileGraphProto();
-
-  // Add node with specified <node_proto>.
-  Node& AddNode(const ONNX_NAMESPACE::NodeProto& node_proto,
-                const ArgNameToTypeMap& name_to_type);
-
-#endif
-
-  Version IrVersion() const noexcept {
-    return ir_version_;
-  }
-
-  Graph& GraphResolveNeeded(bool needed) noexcept {
-    graph_resolve_needed_ = needed;
-    return *this;
-  }
-
-  Graph& GraphProtoSyncNeeded(bool needed) noexcept {
-    graph_proto_sync_needed_ = needed;
-    return *this;
-  }
-
-  // During the Resolve of a Graph it is necessary to recursively descend into subgraphs (created from GraphProto
-  // Node attributes in the Graph) if present.
-  // The ResolveContext holds the collection of values for the current Graph instance, be it the main graph
-  // or a subgraph, so that the various operations that are part of the Resolve can work iteratively or
-  // recursively as needed.
-  struct ResolveContext {
-    ResolveContext(const Graph& owning_graph) : graph{owning_graph} {
-    }
-
-    std::unordered_map<std::string_view, std::pair<Node*, int>> output_args;
-    std::unordered_set<std::string_view> inputs_and_initializers;
-    std::unordered_map<std::string_view, NodeIndex> node_name_to_index;
-    std::unordered_set<Node*> nodes_with_subgraphs;
-
-    // check if the provided name is an input/initialize/node output of this Graph instance during Graph::Resolve.
-    // Graph::node_args_ can have stale entries so we can't rely on that.
-    bool IsLocalValue(const std::string& name) const;
-
-    // check if an ancestor graph has a valid value with the provided name during Graph::Resolve.
-    // Once Graph::Resolve completes Graph::IsOuterScopeValue can be used and is more efficient.
-    bool IsOuterScopeValue(const std::string& name) const;
-
-    void Clear() {
-      output_args.clear();
-      inputs_and_initializers.clear();
-      node_name_to_index.clear();
-      nodes_with_subgraphs.clear();
-    }
-
-   private:
-    bool IsInputInitializerOrOutput(const std::string& name, bool check_ancestors) const;
-
-    const Graph& graph;
-    ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(ResolveContext);
-  };
-
-  // Initialize all the graph inputs, initializers and outputs
-  common::Status InitInputsInitializersOutputs();
-
-  // Initialize overridable initializers container
-  void ComputeOverridableInitializers();
-
-#if !defined(ORT_MINIMAL_BUILD)
-  // Build and verify node connection (edges).
-  // Verify NodeArg name/type/shape matching correctly.
-  common::Status BuildConnections(std::unordered_set<std::string>& outer_scope_node_args_consumed);
-
-  common::Status VerifyNoDuplicateName();
-
-  // Check whether <*this> graph is acyclic while performing a topological sort.
-  // Depth-first going from bottom up through the graph and checking whether there are any back edges.
-  // NodesInTopologicalOrder is updated with the nodes' indexes in topological
-  // order if <Status> returned is "OK", otherwise it's undefined.
-  common::Status PerformTopologicalSortAndCheckIsAcyclic();
-
-  common::Status PerformTypeAndShapeInferencing(const ResolveOptions& options);
-
-  // Recursively find all subgraphs including nested subgraphs
-  void FindAllSubgraphs(std::vector<Graph*>& subgraphs);
-
-  // Iterate this Graph instance and all subgraphs, calling the provided function for each.
-  common::Status ForThisAndAllSubgraphs(const std::vector<Graph*>& subgraphs, std::function<Status(Graph&)> func);
-
-  common::Status InferAndVerifyTypeMatch(Node& node, const ONNX_NAMESPACE::OpSchema& op, const ResolveOptions& options);
-
-  // perform type and shape inferencing on the subgraph and Resolve to validate
-  static common::Status InferAndVerifySubgraphTypes(const Node& node, Graph& subgraph,
-                                                    const std::vector<const ONNX_NAMESPACE::TypeProto*>& input_types,
-                                                    std::vector<const ONNX_NAMESPACE::TypeProto*>& output_types,
-                                                    const Graph::ResolveOptions& options);
-
-  // Apply type-inference and type-checking to all inputs and initializers:
-  common::Status TypeCheckInputsAndInitializers();
-
-  // Compute set of input and initializer names and checking for duplicate names
-  common::Status VerifyInputAndInitializerNames();
-
-  // Infer and set type information across <*this> graph if needed, and verify type/attribute
-  // information matches between node and op.
-
-  common::Status VerifyNodeAndOpMatch(const ResolveOptions& options);
-
-  // Set graph inputs/outputs when resolving a graph..
-  common::Status SetGraphInputsOutputs();
-
-  // recursively accumulate and set the outer scope node args in the resolve context for all subgraphs
-  // so they can be used to resolve outer scope dependencies when running BuildConnections for the subgraphs.
-  common::Status SetOuterScopeNodeArgs(const std::unordered_set<std::string>& outer_scope_node_args);
-
-  // Implementation for initializer replacement
-  Status ReplaceInitializedTensorImpl(ONNX_NAMESPACE::TensorProto new_initializer, bool is_external);
-
-  // Clear all unused initializers and NodeArgs
-  void CleanUnusedInitializersAndNodeArgs(const std::unordered_set<std::string>* initializer_names_to_preserve = nullptr);
-
-  std::vector<NodeArg*> CreateNodeArgs(const google::protobuf::RepeatedPtrField<std::string>& names,
-                                       const ArgNameToTypeMap& name_to_type_map);
-
-  void ToGraphProtoInternal(ONNX_NAMESPACE::GraphProto& graph_proto) const;
-
-#endif  // !defined(ORT_MINIMAL_BUILD)
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  Status PopulateNodeArgToProducerConsumerLookupsFromNodes();
-
-  template <typename TInstance>
-  static auto GetConsumerNodesImpl(
-      TInstance& instance, const std::string& node_arg_name) -> std::vector<decltype(instance.GetNode(0))> {
-    std::vector<decltype(instance.GetNode(0))> results;
-    auto iter = instance.node_arg_to_consumer_nodes_.find(node_arg_name);
-    if (iter != instance.node_arg_to_consumer_nodes_.end()) {
-      results.reserve(iter->second.size());
-      for (auto node_index : iter->second) {
-        results.push_back(instance.GetNode(node_index));
-      }
-    }
-    return results;
-  }
-
-  template <typename TInstance>
-  static auto GetProducerNodeImpl(
-      TInstance& instance, const std::string& node_arg_name) -> decltype(instance.GetNode(0)) {
-    auto iter = instance.node_arg_to_producer_node_.find(node_arg_name);
-    if (iter != instance.node_arg_to_producer_node_.end()) {
-      auto node_index = iter->second;
-      return instance.GetNode(node_index);
-    }
-    return nullptr;
-  }
-
-  gsl::not_null<Node*> AllocateNode();
-
-  // Release the node.
-  // @returns false if node_index was invalid.
-  bool ReleaseNode(NodeIndex node_index);
-
-  Node& CreateFusedSubGraphNode(const IndexedSubGraph& sub_graph, const std::string& fused_node_name);
-#endif  // !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-
-  Node* NodeAtIndexImpl(NodeIndex node_index) const {
-    // if we are trying to access a node that doesn't exist there's (most
-    // likely) either a logic issue or a graph consistency/correctness issue.
-    // use ORT_ENFORCE to prove that or uncover scenarios where we actually
-    // expect attempts to retrieve a non-existent node.
-    ORT_ENFORCE(node_index < nodes_.size(), "Validating no unexpected access using an invalid node_index. Got:",
-                node_index, " Max:", nodes_.size());
-    return nodes_[node_index].get();
-  }
-
-  const Model& owning_model_;
-
-  // GraphProto to store name, version, initializer.
-  // When serializing <*this> Graph to a GraphProto, the nodes and
-  // functions in <Graph> will also be fed into <graph_proto_> so that
-  // it's consistent with <*this> graph.
-  // This pointer is owned by parent model.
-  ONNX_NAMESPACE::GraphProto* graph_proto_;
-
-  // GraphProto that provides storage for the ONNX proto types deserialized from a flexbuffer/flatbuffer
-  ONNX_NAMESPACE::GraphProto deserialized_proto_data_;
-
-  InitializedTensorSet name_to_initial_tensor_;
-
-  std::unordered_set<std::reference_wrapper<const std::string>,
-                     std::hash<std::string>, std::equal_to<std::string>>
-      sparse_tensor_names_;
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  // Runtime optimization storage.
-  // Note: runtime_optimizations_ == *runtime_optimizations_ptr_ and must be initialized
-  std::unique_ptr<RuntimeOptimizationRecordContainer> runtime_optimizations_ptr_;
-  RuntimeOptimizationRecordContainer& runtime_optimizations_;
-#endif  // !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-
-#if !defined(ORT_MINIMAL_BUILD)
-  IOnnxRuntimeOpSchemaCollectionPtr schema_registry_;
-
-  // Currently to make the ORT in-memory graph work, we have to create a temporary op schema
-  // for the fused kernel. I really don't like it. but for short-term solution, let's host
-  // those schemas here.
-  InlinedVector<std::unique_ptr<ONNX_NAMESPACE::OpSchema>> fused_schemas_containers_;
-  // in some case, a fused sub-graph will happens multiple times in one model, we use a map
-  // to store reusable-schema in lookup.
-  InlinedHashMap<std::string, std::reference_wrapper<ONNX_NAMESPACE::OpSchema>> reusable_fused_schema_map_;
-#endif  // !defined(ORT_MINIMAL_BUILD)
-
-  // Graph nodes.
-  // Element in <nodes_> may be nullptr due to graph optimization.
-  std::vector<std::unique_ptr<Node>> nodes_;
-
-  // Wrapper of Graph nodes to provide iteration services that hide nullptr entries
-  GraphNodes iterable_nodes_{nodes_};
-
-  // Number of nodes.
-  // Normally this is smaller than the size of <m_nodes>, as some
-  // elements in <m_nodes> may be removed when doing graph optimization,
-  // or some elements may be merged, etc.
-  int num_of_nodes_ = 0;
-
-  // A flag indicates whether <*this> graph needs to be resolved.
-  bool graph_resolve_needed_ = false;
-
-  bool graph_proto_sync_needed_ = false;
-
-  // The topological order of node index used to do node and op match verification temporarily.
-  std::vector<NodeIndex> nodes_in_topological_order_;
-
-  // Full list of graph inputs. Matches number and order of inputs in the GraphProto.
-  std::vector<const NodeArg*> graph_inputs_including_initializers_;
-  bool graph_inputs_manually_set_ = false;
-
-  // Graph inputs excluding initializers.
-  std::vector<const NodeArg*> graph_inputs_excluding_initializers_;
-
-  // Overridable Initializers. The difference between graph_inputs_including_initializers_
-  // and graph_inputs_excluding_initializers_
-  std::vector<const NodeArg*> graph_overridable_initializers_;
-
-  // Graph outputs.
-  std::vector<const NodeArg*> graph_outputs_;
-  bool graph_outputs_manually_set_ = false;
-
-  // Graph value_info.
-  std::unordered_set<const NodeArg*> value_info_;
-
-  // All node args owned by <*this> graph. Key is node arg name.
-  std::unordered_map<std::string, std::unique_ptr<NodeArg>> node_args_;
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  int name_generator_ = 0;
-
-  // Strings which have been used as node names.
-  // New node name should not conflict with this set.
-  std::unordered_set<std::string> generated_node_names_;
-
-  // Strings which have been used as node_arg names.
-  // New node_arg name should not conflict this this set.
-  std::unordered_set<std::string> generated_node_arg_names_;
-
-  // node arg to its producer node
-  std::unordered_map<std::string, NodeIndex> node_arg_to_producer_node_;
-
-  // node arg to its consumer nodes
-  std::unordered_map<std::string, std::unordered_set<NodeIndex>> node_arg_to_consumer_nodes_;
-#endif  // !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-
-  const std::unordered_map<std::string, int> domain_to_version_;
-
-  // Model IR version.
-  Version ir_version_{ONNX_NAMESPACE::Version::IR_VERSION};
-
-  ResolveContext resolve_context_{*this};
-
-  // the parent graph if this is a subgraph.
-  Graph* parent_graph_;
-  // the node containing the graph if parent_graph_ is not nullptr
-  const Node* parent_node_;
-
-  // NodeArgs that come from outer scope. Used when building a graph so that
-  // these don't get recorded as graph inputs in the GraphProto.
-  std::unordered_set<std::string> outer_scope_node_arg_names_;
-
-  // number of times Resolve has run.
-  int num_resolves_ = 0;
-
-  const logging::Logger& logger_;
-
-  // If true, all inconsistencies encountered during shape and type inference
-  // will be exposed to the caller as failures. If false, in some cases
-  // warnings will be logged but processing will continue and no error will
-  // be returned.
-  const bool strict_shape_type_inference_;
-
-  // distinguishes between graph loaded from model file and graph created from scratch
-  const bool is_loaded_from_model_file_;
-};
-
-#if !defined(ORT_MINIMAL_BUILD)
-// Print NodeArg as
-//  name : type
-// For example,
-//  "110": tensor(float)
-std::ostream& operator<<(std::ostream& out, const NodeArg& node_arg);
-// Print Node as,
-//  (operator's name, operator's type, domain, version) : (input0, input1, ...) -> (output0, output1, ...)
-// For example,
-//  ("Add_14", Add, "", 7) : ("110": tensor(float),"109": tensor(float),) -> ("111": tensor(float),)
-std::ostream& operator<<(std::ostream& out, const Node& node);
-// Print Graph as, for example,
-// Inputs:
-//    "Input": tensor(float)
-// Nodes:
-//    ("add0", Add, "", 7) : ("Input": tensor(float),"Bias": tensor(float),) -> ("add0_out": tensor(float),)
-//    ("matmul", MatMul, "", 9) : ("add0_out": tensor(float),"matmul_weight": tensor(float),) -> ("matmul_out": tensor(float),)
-//    ("add1", Add, "", 7) : ("matmul_out": tensor(float),"add_weight": tensor(float),) -> ("add1_out": tensor(float),)
-//    ("reshape", Reshape, "", 5) : ("add1_out": tensor(float),"concat_out": tensor(int64),) -> ("Result": tensor(float),)
-// Outputs:
-//    "Result": tensor(float)
-// Inputs' and outputs' format is described in document of NodeArg's operator<< above.
-// Node format is described in Node's operator<< above.
-std::ostream& operator<<(std::ostream& out, const Graph& graph);
-#endif
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/graph/graph_nodes.h b/projects/llm_framework/include/onnxruntime/core/graph/graph_nodes.h
deleted file mode 100644
index aec603f7..00000000
--- a/projects/llm_framework/include/onnxruntime/core/graph/graph_nodes.h
+++ /dev/null
@@ -1,182 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <memory>
-#include <type_traits>
-#include <vector>
-
-namespace onnxruntime {
-
-class Node;
-
-/**
-Class to filter out null entries from either a vector of unique_ptr<Node> or a vector of [const] Node* and
-provide an iterator interface that returns [const] Node& for the valid entries.
-*/
-template <typename TNodesContainer>
-class ValidNodes {
- public:
-  template <typename TIterator>
-  class NodeIterator;
-
-  // optional filtering function to return a subset of nodes
-  using NodeFilterFunc = std::function<bool(NodeIndex)>;
-
-  /**
-  Construct a ValidNodes instance to provide iteration over all valid nodes in the TNodesCollection
-  @param[in] nodes Nodes to iterate, skipping invalid entries.
-  */
-  explicit ValidNodes(TNodesContainer& nodes) noexcept : nodes_(&nodes) {}
-
-  explicit ValidNodes(TNodesContainer& nodes, NodeFilterFunc&& filter_node_fn) noexcept
-      : nodes_(&nodes), filter_node_fn_{std::move(filter_node_fn)} {}
-
-  using ConstNodeIterator = NodeIterator<typename TNodesContainer::const_iterator>;
-  using MutableNodeIterator = NodeIterator<typename TNodesContainer::iterator>;
-  using ConstReverseNodeIterator = NodeIterator<typename TNodesContainer::const_reverse_iterator>;
-
-  ConstNodeIterator cbegin() const noexcept {
-    return {nodes_->cbegin(), nodes_->cend(), filter_node_fn_};
-  }
-
-  ConstNodeIterator cend() const noexcept {
-    return {nodes_->cend(), nodes_->cend(), filter_node_fn_};
-  }
-
-  ConstNodeIterator begin() const noexcept {
-    return cbegin();
-  }
-
-  ConstNodeIterator end() const noexcept {
-    return cend();
-  }
-
-  ConstReverseNodeIterator rbegin() const noexcept {
-    return {nodes_->crbegin(), nodes_->crend(), filter_node_fn_};
-  }
-
-  ConstReverseNodeIterator rend() const noexcept {
-    return {nodes_->crend(), nodes_->crend(), filter_node_fn_};
-  }
-
-  // we only allow mutable access if the container is non-const.
-  // we need to templatize the functions for enable_if to work at this level, but mandate T2 being TNodesContainer
-  template <typename T2 = TNodesContainer>
-  typename std::enable_if<!std::is_const<T2>::value, MutableNodeIterator>::type begin() noexcept {
-    static_assert(std::is_same<T2, TNodesContainer>::value, "Explicit specialization is not allowed");
-    return MutableNodeIterator(nodes_->begin(), nodes_->end(), filter_node_fn_);
-  }
-
-  template <typename T2 = TNodesContainer>
-  typename std::enable_if<!std::is_const<T2>::value, MutableNodeIterator>::type end() noexcept {
-    static_assert(std::is_same<T2, TNodesContainer>::value, "Explicit specialization is not allowed");
-    return MutableNodeIterator(nodes_->end(), nodes_->end(), filter_node_fn_);
-  }
-
-  bool empty() const noexcept { return nodes_->empty(); }
-
-  /**
-  @class NodeIterator
-  Iterator to provide const and non-const access to valid Node instances in a Graph.
-  @remarks Skips invalid nodes.
-  */
-  template <typename TIterator>
-  class NodeIterator {
-    // get the type being returned by the iterator. can't use TIterator::value_type as that is always non-const
-    using IterType = typename std::remove_reference<typename std::iterator_traits<TIterator>::reference>::type;
-    // and determine what we will return based on its constness
-    using T = typename std::conditional<std::is_const<IterType>::value,
-                                        const Node,   // return const Node if this is a const iterator
-                                        Node>::type;  // else return Node
-
-   public:
-    using iterator_category = std::input_iterator_tag;
-    using value_type = T;
-    using difference_type = typename TIterator::difference_type;
-    using pointer = T*;
-    using reference = T&;
-    using const_reference = const T&;
-
-    /** Construct a NodeInterator and move to the first valid node. */
-    NodeIterator(const TIterator current, const TIterator end, const NodeFilterFunc& filter_fn) noexcept
-        : current_{current}, end_{end}, apply_filter_{filter_fn != nullptr}, filter_func_{&filter_fn} {
-      // skip to next valid node, stopping at end if none are found
-      while (current_ < end && (*current_ == nullptr ||
-                                (apply_filter_ && (*filter_func_)((*current_)->Index()) == true))) {
-        ++current_;
-      }
-    }
-
-    bool operator==(const NodeIterator<TIterator>& other) const noexcept {
-      return (current_ == other.current_);
-    }
-
-    bool operator!=(const NodeIterator<TIterator>& other) const noexcept {
-      return (current_ != other.current_);
-    }
-
-    void operator++() {
-      if (current_ < end_) {
-        while (++current_ != end_) {
-          if (*current_ != nullptr && (!apply_filter_ || (*filter_func_)((*current_)->Index()) == false))
-            break;
-        }
-      }
-    }
-
-    NodeIterator<TIterator> operator++(int) {
-      NodeIterator<TIterator> tmp{*this};
-      ++(*this);
-
-      return tmp;
-    }
-
-    /** Return the current Node&. This will be const if the iterator was returned from a const GraphNodes instance. */
-    reference operator*() {
-      // if iterator is valid we always have a non-nullptr node
-      // if this is a nullptr we're at end_ and this shouldn't be being called
-      return **current_;
-    }
-
-    pointer operator->() {
-      return current_->get();
-    }
-
-   private:
-    TIterator current_;
-    TIterator end_;
-    bool apply_filter_;                  // store whether filter_func_ is not nullptr and contains a callable
-    const NodeFilterFunc* filter_func_;  // store as pointer so iterator is copyable
-  };
-
- private:
-  gsl::not_null<TNodesContainer*> nodes_;  // always set by ctor
-
-  // no filtering if not set. this instance owns the filter func if set.
-  NodeFilterFunc filter_node_fn_;
-};
-
-/**
-Class that provides iteration over all valid nodes in the Graph.
-*/
-class GraphNodes : public ValidNodes<std::vector<std::unique_ptr<Node>>> {
- public:
-  GraphNodes(std::vector<std::unique_ptr<Node>>& nodes) : ValidNodes(nodes) {
-  }
-};
-
-// Variant that only ever allows const access to nodes and optionally allows filtering of the nodes.
-class ConstGraphNodes : public ValidNodes<const std::vector<std::unique_ptr<Node>>> {
- public:
-  ConstGraphNodes(const std::vector<std::unique_ptr<Node>>& nodes) : ValidNodes(nodes) {
-  }
-
-  ConstGraphNodes(const std::vector<std::unique_ptr<Node>>& nodes,
-                  GraphNodes::NodeFilterFunc&& filter_func)
-      : ValidNodes(nodes, std::move(filter_func)) {
-  }
-};
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/graph/graph_viewer.h b/projects/llm_framework/include/onnxruntime/core/graph/graph_viewer.h
deleted file mode 100644
index 3cdbb070..00000000
--- a/projects/llm_framework/include/onnxruntime/core/graph/graph_viewer.h
+++ /dev/null
@@ -1,222 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include "core/graph/graph.h"
-#include "core/framework/session_options.h"
-#include <unordered_set>
-
-namespace onnxruntime {
-class Function;
-struct IndexedSubGraph;
-}  // namespace onnxruntime
-
-namespace onnxruntime {
-
-// use value-based compare to make sure transformer output order is consistent
-struct NodeCompare {
-  bool operator()(const Node* n1, const Node* n2) const;
-};
-
-/**
-@class GraphViewer
-Class that provides a read-only view of the Graph.
-@remarks If the underlying Graph is changed, GetNodesInTopologicalOrder and GetRootNodes may become invalid.
-*/
-class GraphViewer {
- public:
-  /**
-  Construct a GraphViewer from the provided Graph instance.
-  */
-  explicit GraphViewer(const Graph& graph);
-
-  /**
-  Construct a GraphViewer from the provided Graph instance, filtering to the nodes specified in the IndexedSubGraph
-  */
-  explicit GraphViewer(const Graph& graph, const IndexedSubGraph& filter_info);
-
-  /** Gets the Graph name. */
-  const std::string& Name() const noexcept;
-
-  /** Gets the Graph description. */
-  const std::string& Description() const noexcept;
-
-  /** Gets the path of the owning model if any **/
-  const Path& ModelPath() const noexcept { return graph_->ModelPath(); }
-
-  /**
-  Gets a tensor created from an initializer.
-  @param tensor_name The tensor name
-  @param[out] value Sets the pointer to the TensorProto if found, or nullptr if not.
-  @returns True if found. False if not.
-  */
-  bool GetInitializedTensor(const std::string& tensor_name, const ONNX_NAMESPACE::TensorProto*& value) const;
-
-  /** Returns true if an initializer value can be overridden by a graph input with the same name. */
-  bool CanOverrideInitializer() const noexcept;
-
-  /**
-  Gets the Graph inputs, excluding initializers.
-  @returns Collection of NodeArg pointers for the graph inputs, excluding inputs that have matching initializers.
-  @remarks No nullptr values in the returned collection. The order will be the same as in the GraphProto.
-           Inputs are for filter_info_ if set.
-  */
-  const std::vector<const NodeArg*>& GetInputs() const noexcept;
-
-  /**
-  Gets the Graph inputs, including any initializers.
-  @returns Collection of NodeArg pointers for all the graph inputs.
-  @remarks No nullptr values in the returned collection. The order will be the same as in the GraphProto.
-           Inputs are for filter_info_ if set.
-  */
-  const std::vector<const NodeArg*>& GetInputsIncludingInitializers() const noexcept;
-
-  /**
-  Gets the Graph outputs.
-  @returns Collection of NodeArg pointers for all the graph outputs.
-  @remarks No nullptr values in the returned collection. The order will be the same as in the GraphProto.
-           Outputs are for filter_info_ if set.
-  */
-  const std::vector<const NodeArg*>& GetOutputs() const noexcept;
-
-  /** Returns true if one or more of the Node outputs are Graph outputs.
-   */
-  bool NodeProducesGraphOutput(const Node& node) const;
-
-  /** Gets all ValueInfo NodeArg instances in the Graph.
-  @remarks NOT filtered using filter_info_.
-  */
-  const std::unordered_set<const NodeArg*>& GetValueInfo() const noexcept;
-
-  /**
-  Gets the Node instance at the specified index.
-  @param node_index Index to retrieve Node from.
-  @remarks May return nullptr if index no longer points to a valid node due to the node being freed, or if
-           node is excluded by filter_info_.
-  */
-  const Node* GetNode(NodeIndex node_index) const;
-
-  /**  Gets an iterator over all the valid Nodes in the Graph.
-  @remarks Nodes are filtered using filter_info_ if set.
-  */
-  const ConstGraphNodes& Nodes() const noexcept;
-
-  /** Gets the number of valid nodes in the Graph.
-  @remarks Returns the number of nodes in filter_info_ if set.
-  */
-  int NumberOfNodes() const noexcept;
-
-  /** Gets the maximum NodeIndex value used by Nodes in the Graph. */
-  int MaxNodeIndex() const noexcept;
-
-  /** Gets the NodeIndex values for the Graph nodes, sorted into topological order.
-  @remarks Filtered using filter_info_ if set.
-  */
-  const std::vector<NodeIndex>& GetNodesInTopologicalOrder(ExecutionOrder order = ExecutionOrder::DEFAULT) const;
-
-  /**
-  Gets the NodeIndex values for the root nodes in the Graph.
-  The root nodes are the topmost nodes in the Graph that receive inputs from the Graph inputs
-  and no other nodes in the Graph.
-  @remarks Not supported if filter_info_ is set.
-  */
-  const std::vector<NodeIndex>& GetRootNodes() const;
-
-  /** Gets all tensors created from initializers. */
-  const InitializedTensorSet& GetAllInitializedTensors() const noexcept;
-
-  /**
-  Gets the NodeArg instance for the given name.
-  @returns A NodeArg if found, a nullptr if not.
-  */
-  const NodeArg* GetNodeArg(const std::string& name) const;
-
-  /** Gets the map of operator domains to their opset versions. */
-  const std::unordered_map<std::string, int>& DomainToVersionMap() const noexcept {
-    return graph_->DomainToVersionMap();
-  }
-
-  /** Checks if this is a Subgraph */
-  bool IsSubgraph() const;
-
-  /** Get the internal graph*/
-  const Graph& GetGraph() const { return *graph_; }
-
-#if !defined(ORT_MINIMAL_BUILD)
-  const std::unordered_set<std::string>& GetOuterScopeNodeArgNames() const noexcept;
-#endif
-
-  /**
-  returns true if 'name' is an initializer, and is constant and cannot be overridden at runtime.
-  @param check_outer_scope If true and the 'graph_' is a subgraph, check parent graph/s for 'name'
-                           if the name is not found in 'graph_'.
-  */
-  bool IsConstantInitializer(const std::string& name, bool check_outer_scope) const;
-
-  /** Check if a given name is an initializer tensor's name in this graph. */
-  bool IsInitializedTensor(const std::string& name) const;
-
-  /** returns the initializer's TensorProto if 'name' is an initializer, is constant and
-  cannot be overridden at runtime. If the initializer is not found or is not constant, a nullptr is returned.
-  @param check_outer_scope If true and the graph is a subgraph,
-         check ancestor graph/s for 'name' if not found in 'graph'.
-  @remarks This function will return the result from GetConstantInitializer of the underlying Graph,
-           if a const initializer is part of the underlying Graph but not part of this GraphViewer,
-           it will still be returned instead of nullptr
-  */
-  const ONNX_NAMESPACE::TensorProto* GetConstantInitializer(const std::string& name, bool check_outer_scope) const;
-
-  /** Get the Node containing this Graph if IsSubgraph is true. Returns nullptr otherwise. */
-  const Node* ParentNode() const noexcept { return graph_->ParentNode(); }
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  /** Get the consumer nodes of a node arg */
-  std::vector<const Node*> GetConsumerNodes(const std::string& node_arg_name) const {
-    return graph_->GetConsumerNodes(node_arg_name);
-  }
-
-  /** Get the producer node of a node arg */
-  const Node* GetProducerNode(const std::string& node_arg_name) const {
-    return graph_->GetProducerNode(node_arg_name);
-  }
-#endif
-
-  /** Get the filter info that restricts the graph viewer to a subset of nodes if set.
-  @returns Filter info or nullptr
-  */
-  const IndexedSubGraph* GetFilterInfo() const { return filter_info_; }
-
-#if !defined(ORT_MINIMAL_BUILD)
-  IOnnxRuntimeOpSchemaCollectionPtr GetSchemaRegistry() const { return graph_->GetSchemaRegistry(); }
-#endif
-
- private:
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(GraphViewer);
-  GraphViewer(const Graph& graph, const IndexedSubGraph* filter_info);
-
-  const Graph* graph_;
-  ConstGraphNodes graph_nodes_;
-
-  // The NodeIndex values of the graph nodes sorted in topological order.
-  std::vector<NodeIndex> nodes_in_topological_order_;
-
-#if !defined(ORT_MINIMAL_BUILD)
-  // The NodeIndex values of the graph nodes sorted in topological order with priority.
-  std::vector<NodeIndex> nodes_in_topological_order_with_priority_;
-#endif
-
-  // Graph root nodes.
-  std::vector<NodeIndex> root_nodes_;
-
-  // if we're limiting the view to an IndexedSubGraph we need to create a few pieces of infrastructure that would
-  // usually come from the full graph
-  const IndexedSubGraph* filter_info_{nullptr};
-  using FilteredNodeSet = InlinedHashSet<NodeIndex>;
-  FilteredNodeSet filtered_node_indices_;
-  std::vector<const NodeArg*> filtered_node_inputs_;
-  std::vector<const NodeArg*> filtered_node_inputs_including_initializers_;
-  std::vector<const NodeArg*> filtered_node_outputs_;
-  InitializedTensorSet filtered_initializers_;
-};
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/graph/indexed_sub_graph.h b/projects/llm_framework/include/onnxruntime/core/graph/indexed_sub_graph.h
deleted file mode 100644
index 935ada88..00000000
--- a/projects/llm_framework/include/onnxruntime/core/graph/indexed_sub_graph.h
+++ /dev/null
@@ -1,84 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "core/graph/basic_types.h"
-#if !defined(ORT_MINIMAL_BUILD)
-#include "onnx/defs/schema.h"
-#else
-#include "onnx/defs/data_type_utils.h"
-#endif
-#include "onnx/onnx_pb.h"
-#include "onnx/onnx-operators_pb.h"
-
-namespace onnxruntime {
-
-class OpKernel;
-class OpKernelInfo;
-
-/**
-@class IndexedSubGraph
-
-Class containing information about a subgraph of Nodes from a Graph.
-It contains a NodeIndex array of the Nodes covered by the subgraph,
-and the meta definition needed for representing this subgraph as a FunctionProto,
-which could be serialized/saved to a model file.
-*/
-struct IndexedSubGraph {
-  struct MetaDef {
-    std::string name;    ///< Name of customized SubGraph/FunctionProto
-    std::string domain;  ///< Domain of customized SubGraph/FunctionProto
-    int since_version;   ///< Since version of customized SubGraph/FunctionProto.
-
-    ONNX_NAMESPACE::OperatorStatus status;  ///< Status of customized SubGraph/FunctionProto.
-
-    std::vector<std::string> inputs;                 ///< Inputs of customized SubGraph/FunctionProto.
-    std::vector<std::string> outputs;                ///< Outputs of customized SubGraph/FunctionProto.
-    std::vector<std::string> constant_initializers;  ///< Constant initializers of customized SubGraph/FunctionProto.
-    NodeAttributes attributes;                       ///< Attributes of customized SubGraph/FunctionProto.
-
-    std::string doc_string;  ///< Doc string of customized SubGraph/FunctionProto.
-#if !defined(ORT_MINIMAL_BUILD)
-    /** Type and shape inference function that can optionally be defined for the fused node */
-    std::function<void(ONNX_NAMESPACE::InferenceContext&)> type_and_shape_inference_function;
-#endif
-  };
-
-  /** Nodes covered by this subgraph. The NodeIndex values are from the parent Graph.*/
-  std::vector<onnxruntime::NodeIndex> nodes;
-
-  enum class SourceOfSchema : uint8_t {
-    CREATE,           /// create new schema from info in IndexedSubGraph instance.
-                      /// schema instance will not be re-usable.
-    REUSE_OR_CREATE,  /// re-use existing dynamically created schema with matching domain+name.
-                      /// create re-usable schema if one is not found.
-    EXISTING,         /// use existing statically registered schema.
-                      /// e.g. domain+name matches ONNX or contrib op domain+op_type+opset.
-  };
-  // Either using an existing schema or generating reusable one when fusing nodes using the MetaDef.
-  // MetaDef.domain + MetaDef.name => the domain.op_type that a schema must exist for with a valid since_version.
-  SourceOfSchema schema_source{SourceOfSchema::CREATE};
-
-  /** Set the meta definition needed to represent this subgraph as a FunctionProto
-  It's needed IF AND ONLY IF there are multiple indexes contained in #nodes. */
-  void SetMetaDef(std::unique_ptr<MetaDef>&& meta_def) {
-    meta_def_ = std::move(meta_def);
-  }
-
-  /** Gets the meta definition needed to represent this subgraph as a FunctionProto.
-  @returns MetaDef instance if it has been set. nullptr if not. */
-  const MetaDef* GetMetaDef() const {
-    return meta_def_.get();
-  }
-
- private:
-  // subgraph meta definition.
-  std::unique_ptr<MetaDef> meta_def_;
-};
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/graph/node_arg.h b/projects/llm_framework/include/onnxruntime/core/graph/node_arg.h
deleted file mode 100644
index 921bff59..00000000
--- a/projects/llm_framework/include/onnxruntime/core/graph/node_arg.h
+++ /dev/null
@@ -1,134 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include "onnx/onnx_pb.h"
-
-#include "core/graph/basic_types.h"
-#include "core/common/status.h"
-#include "core/common/logging/logging.h"
-
-namespace onnxruntime {
-
-// Node argument definition, for both input and output,
-// including arg name, arg type (contains both type and shape).
-//
-// Design Question: in my opinion, shape should not be part of type.
-// We may align the protobuf design with our operator registry interface,
-// which has type specified for each operator, but no shape. Well, shape
-// should be inferred with a separate shape inference function given
-// input shapes, or input tensor data sometimes.
-// With shape as part of type (current protobuf design),
-// 1) we'll have to split the "TypeProto" into type and shape in this internal
-// representation interface so that it could be easily used when doing type
-// inference and matching with operator registry.
-// 2) SetType should be always called before SetShape, otherwise, SetShape()
-// will fail. Because shape is located in a TypeProto.
-// Thoughts?
-//
-
-/**
-@class NodeArg
-Class representing a data type that is input or output for a Node, including the shape if it is a Tensor.
-*/
-class NodeArg {
- public:
-  /**
-  Construct a new NodeArg.
-  @param name The name to use.
-  @param p_arg_type Optional TypeProto specifying type and shape information.
-  */
-  NodeArg(const std::string& name,
-          const ONNX_NAMESPACE::TypeProto* p_arg_type);
-
-  NodeArg(NodeArg&&) = default;
-  NodeArg& operator=(NodeArg&& other) = default;
-
-  /** Gets the name. */
-  const std::string& Name() const noexcept;
-
-  /** Gets the data type. */
-  const std::string* Type() const noexcept;
-
-  /** Gets the TypeProto
-  @returns TypeProto if type is set. nullptr otherwise. */
-  const ONNX_NAMESPACE::TypeProto* TypeAsProto() const noexcept;
-
-  /** Gets the shape if NodeArg is for a Tensor.
-  @returns TensorShapeProto if shape is set. nullptr if there's no shape specified. */
-  const ONNX_NAMESPACE::TensorShapeProto* Shape() const;
-
-  /** Return an indicator.
-  @returns true if NodeArg is a normal tensor with a non-empty shape or a scalar with an empty shape. Otherwise, returns false. */
-  bool HasTensorOrScalarShape() const;
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-
-  /** Sets the shape.
-  @remarks Shape can only be set if the TypeProto was provided to the ctor, or #SetType has been called,
-  as the shape information is stored as part of TypeProto. */
-  void SetShape(const ONNX_NAMESPACE::TensorShapeProto& shape);
-
-  /** Clears shape info.
-  @remarks If there is a mismatch during shape inferencing that can't be resolved the shape info may be removed. */
-  void ClearShape();
-
-#endif  // !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-
-#if !defined(ORT_MINIMAL_BUILD)
-
-  /** Override current type from input_type if override_types is set to true, return failure status otherwise.
-  @param input_tensor_elem_type Tensor element type parsed input_type
-  @param current_tensor_elem_type Tensor element type parsed from existing type
-  @param override_types If true, resolve the two inputs or two outputs type when different
-  @returns Success unless there is existing type or shape info that can't be successfully updated. */
-  common::Status OverrideTypesHelper(const ONNX_NAMESPACE::TypeProto& input_type,
-                                     int32_t input_tensor_elem_type,
-                                     int32_t current_tensor_elem_type,
-                                     bool override_types);
-
-  /** Validate and merge type [and shape] info from input_type.
-  @param strict If true, the shape update will fail if there are incompatible values.
-                If false, will be lenient and merge only shape info that can be validly processed.
-  @param override_types If true, resolve the two inputs or two outputs type when different
-  @returns Success unless there is existing type or shape info that can't be successfully updated. */
-  common::Status UpdateTypeAndShape(const ONNX_NAMESPACE::TypeProto& input_type, bool strict, bool override_types, const logging::Logger& logger);
-
-  /** Validate and merge type [and shape] info from node_arg.
-  @param strict If true, the shape update will fail if there are incompatible values.
-                If false, will be lenient and merge only shape info that can be validly processed.
-  @param override_types If true, resolve the two inputs or two outputs type when different
-  @returns Success unless there is existing type or shape info that can't be successfully updated. */
-  common::Status UpdateTypeAndShape(const NodeArg& node_arg, bool strict, bool override_types, const logging::Logger& logger);
-
-#endif  // !defined(ORT_MINIMAL_BUILD)
-
-  /** Gets this NodeArg as a NodeArgInfo, AKA ValueInfoProto. */
-  const NodeArgInfo& ToProto() const noexcept { return node_arg_info_; }
-
-  /** Gets a flag indicating whether this NodeArg exists or not.
-  Optional inputs are allowed in ONNX and an empty #Name represents a non-existent input argument. */
-  bool Exists() const noexcept;
-
-  friend class Graph;
-
-  NodeArg(NodeArgInfo&& node_arg_info);
-
- private:
-  ORT_DISALLOW_COPY_AND_ASSIGNMENT(NodeArg);
-  void SetType(const std::string* p_type);
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-  void SetType(const ONNX_NAMESPACE::TypeProto& type_proto);
-#endif  // !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-
-  // Node arg PType.
-  const std::string* type_;
-
-  // Node arg name, type and shape.
-  NodeArgInfo node_arg_info_;
-
-  // Flag indicates whether <*this> node arg exists or not.
-  bool exists_;
-};
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/graph/schema_registry.h b/projects/llm_framework/include/onnxruntime/core/graph/schema_registry.h
deleted file mode 100644
index 91e91bfe..00000000
--- a/projects/llm_framework/include/onnxruntime/core/graph/schema_registry.h
+++ /dev/null
@@ -1,163 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-#include <mutex>
-#include <deque>
-#include <map>
-#include <sstream>
-
-#include "onnx/onnx_pb.h"
-#include "onnx/onnx-operators_pb.h"
-#include "onnx/defs/schema.h"
-
-#include "core/graph/constants.h"
-#include "core/common/common.h"
-#include "core/common/status.h"
-#include "core/platform/ort_mutex.h"
-
-namespace onnxruntime {
-using OpName_Domain_Version_Schema_Map = std::unordered_map<
-    std::string,
-    std::unordered_map<std::string, std::map<ONNX_NAMESPACE::OperatorSetVersion, ONNX_NAMESPACE::OpSchema>>>;
-
-/** 
-@struct SchemaRegistryVersion
-onnxruntime schema registry is a supplement to the built-in ONNX schema.
-Every schema registry represent a collection of schema deltas from baseline_opset_version to opset_version
-*/
-struct SchemaRegistryVersion {
-  int baseline_opset_version;
-  int opset_version;
-};
-
-using DomainToVersionMap = std::unordered_map<std::string, int>;
-using DomainToVersionRangeMap = std::unordered_map<std::string, SchemaRegistryVersion>;
-
-class IOnnxRuntimeOpSchemaCollection : public ONNX_NAMESPACE::ISchemaRegistry {
- public:
-  virtual DomainToVersionMap GetLatestOpsetVersions(bool is_onnx_only) const = 0;
-
-  using ISchemaRegistry::GetSchema;
-
-  const ONNX_NAMESPACE::OpSchema* GetSchema(const std::string& key, const int maxInclusiveVersion,
-                                            const std::string& domain) const final {
-    const ONNX_NAMESPACE::OpSchema* latest_schema = nullptr;
-    int earliest_opset_where_unchanged = std::numeric_limits<int>::max();
-    GetSchemaAndHistory(key, maxInclusiveVersion, domain, &latest_schema, &earliest_opset_where_unchanged);
-
-    assert(latest_schema == nullptr || (latest_schema->SinceVersion() <= maxInclusiveVersion &&
-                                        earliest_opset_where_unchanged == latest_schema->SinceVersion()));
-
-    return latest_schema;
-  }
-
-  virtual void GetSchemaAndHistory(
-      const std::string& key,
-      int maxInclusiveVersion,
-      const std::string& domain,
-      const ONNX_NAMESPACE::OpSchema** latest_schema,
-      int* earliest_opset_where_unchanged) const = 0;
-};
-
-/** 
-@class OnnxRuntimeOpSchemaRegistry
-
-OnnxRuntimeOpSchemaRegistry is used to provide supplement for built-in ONNX schemas.
-Each OnnxRuntimeOpSchemaRegistry must register complete opsets delta from a baseline version to max opset version.
-(Please notice that baseline opsets are not include in the delta)
-
-For example, ONNXRuntime is build with ONNX 1.2 which is at opset7, to use ONNX opset8 and opset9,
-user could create a OnnxRuntimeOpSchemaRegistry and config it as {baseline_opset_version = 7, opset_version = 9}
-it means this OnnxRuntimeOpSchemaRegistry contains the complete delta from opset7 to opset9.
-*/
-class OnnxRuntimeOpSchemaRegistry : public IOnnxRuntimeOpSchemaCollection {
- public:
-  OnnxRuntimeOpSchemaRegistry() = default;
-
-  common::Status SetBaselineAndOpsetVersionForDomain(
-      const std::string& domain,
-      int baseline_opset_version,
-      int opset_version);
-
-  DomainToVersionMap GetLatestOpsetVersions(bool is_onnx_only) const override;
-
-  // OnnxRuntimeOpSchemaRegistry must register complete delta for a opset.
-  common::Status RegisterOpSet(
-      std::vector<ONNX_NAMESPACE::OpSchema>& schemas,
-      const std::string& domain,
-      int baseline_opset_version,
-      int opset_version);
-
-  using IOnnxRuntimeOpSchemaCollection::GetSchema;
-
-  void GetSchemaAndHistory(const std::string& key, int maxInclusiveVersion, const std::string& domain,
-                           const ONNX_NAMESPACE::OpSchema** latest_schema,
-                           int* earliest_opset_where_unchanged) const override;
-
-  bool empty() const {
-    return map_.empty();
-  }
-
- private:
-  common::Status RegisterOpSchema(ONNX_NAMESPACE::OpSchema&& op_schema);
-
-  common::Status RegisterOpSchemaInternal(ONNX_NAMESPACE::OpSchema&& op_schema);
-
-  OrtMutex mutex_;
-
-  OpName_Domain_Version_Schema_Map map_;
-  DomainToVersionRangeMap domain_version_range_map_;
-};
-
-/**
-@class SchemaRegistryManager 
-
-SchemaRegistryManager provides a view based on built-in ONNX schema and a list of 
-OnnxRuntimeOpSchemaRegistry as supplement. 
-
-The user needs to make sure the customized schema registry is valid, otherwise the behavior is undefined.
-
-@todo We may add more consistency checks later.
-*/
-class SchemaRegistryManager : public onnxruntime::IOnnxRuntimeOpSchemaCollection {
- public:
-  /**
-  Register a new schema registry instance.
-  @remarks The schema registry priority is the reverse of registration order. i.e. the last registry added will be
-  searched first for a matching OpSchema.
-  */
-  void RegisterRegistry(std::shared_ptr<IOnnxRuntimeOpSchemaCollection> registry);
-
-  /** Gets the latest opset versions.
-  @param is_onnx_only If true, return the latest ONNX schemas. If false, return the latest schemas for all domains.
-  */
-  DomainToVersionMap GetLatestOpsetVersions(bool is_onnx_only) const override;
-
-  /** Gets the last released opset versions.
-  @param is_onnx_only If true, return ONNX schemas only. If false, return the schemas for all domains.
-  */
-  DomainToVersionMap GetLastReleasedOpsetVersions(bool is_onnx_only) const ;
-  /**
-  Gets the OpSchema and its history.
-  Searches custom schema registries starting with the last one added. \
-  If the OpSchema is not found the default ONNX schema registry is searched.
-
-  @param key Operator type.
-  @param max_inclusive_version Maximum opset version allowed, inclusive.
-  @param domain The domain of the operator.
-  @param[out] latest_schema Returns the latest OpSchema if found. nullptr otherwise.
-  @param[out] earliest_opset_where_unchanged The earliest opset version preceding max_inclusive_version where the
-  operator is known to be unchanged.
-  */
-  void GetSchemaAndHistory(const std::string& key, int max_inclusive_version, const std::string& domain,
-                           const ONNX_NAMESPACE::OpSchema** latest_schema,
-                           int* earliest_opset_where_unchanged) const override;
-
- private:
-  void GetDomainToVersionMapForRegistries(DomainToVersionMap& domain_version_map, bool is_onnx_only) const;
-
-  std::deque<std::shared_ptr<IOnnxRuntimeOpSchemaCollection>> registries;
-};
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer.h b/projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer.h
deleted file mode 100644
index 579f1445..00000000
--- a/projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer.h
+++ /dev/null
@@ -1,73 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-#include <string>
-
-#include "core/common/common.h"
-#include "core/common/inlined_containers.h"
-#include "core/graph/graph_viewer.h"
-#include "core/optimizer/graph_transformer_level.h"
-
-namespace onnxruntime {
-
-/**
-@class GraphTransformer
-
-The interface for in-place transformation of a Graph.
-*/
-class GraphTransformer {
- public:
-  GraphTransformer(const std::string& name,
-      const InlinedHashSet<std::string_view>& compatible_execution_providers = {}) noexcept
-      : name_(name), compatible_provider_types_(compatible_execution_providers) {
-  }
-
-  virtual ~GraphTransformer() = default;
-
-  /** Gets the name of this graph transformer. */
-  const std::string& Name() const noexcept {
-    return name_;
-  }
-
-  const InlinedHashSet<std::string_view>& GetCompatibleExecutionProviders() const noexcept {
-    return compatible_provider_types_;
-  }
-
-  /** Apply the in-place transformation defined by this transformer to the provided Graph instance.
-  @param[out] modified Set to true if the Graph was modified.
-  @returns Status with success or error information.
-  */
-  common::Status Apply(Graph& graph, bool& modified, const logging::Logger& logger) const;
-
-  virtual bool ShouldOnlyApplyOnce() const { return false; }
-
- protected:
-  /** Helper method to call ApplyImpl on any subgraphs in the Node. */
-  common::Status Recurse(Node& node, bool& modified, int graph_level, const logging::Logger& logger) const {
-    int subgraph_level = ++graph_level;
-    for (auto& entry : node.GetAttributeNameToMutableSubgraphMap()) {
-      auto& subgraph = *entry.second;
-      ORT_RETURN_IF_ERROR(ApplyImpl(subgraph, modified, subgraph_level, logger));
-    }
-
-    return Status::OK();
-  }
-
- private:
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(GraphTransformer);
-
-  // Apply the transform to the graph.
-  // graph_level is 0 for the main graph, and is incremented when descending into the subgraph of a node.
-  // You MUST call Recurse for all valid Nodes in the graph to ensure any subgraphs in control flow nodes
-  // (Scan/If/Loop) are processed as well.
-  // You should avoid calling Graph::Resolve in ApplyImpl unless you are 100% sure it's required. In most cases
-  // the call to Graph::Resolve in Apply prior to ApplyImpl being called, and after ApplyImpl fore the main graph
-  // completes (if 'modified' is true) should suffice.
-  virtual common::Status ApplyImpl(Graph& graph, bool& modified, int graph_level, const logging::Logger& logger)
-      const = 0;
-
-  const std::string name_;
-  const InlinedHashSet<std::string_view> compatible_provider_types_;
-};
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer_config.h b/projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer_config.h
deleted file mode 100644
index c112d9b0..00000000
--- a/projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer_config.h
+++ /dev/null
@@ -1,73 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-#include <string>
-#include <vector>
-#include <type_traits>
-namespace onnxruntime {
-
-struct GraphTransformerConfiguration {
-  struct PropagateCastOpsConfiguration {
-    // Propagate FP16 Cast operations up and FP32 operations down
-    /*
-     * Cast propagation strategy.
-     * One strategy is to insert casts around all the nodes with the allowed opcodes
-     * and reduce, by removing redundent-casts and back-to-back-casts etc., and
-     * the other is to propagate casts using flood-fill approach, expanding float16 regions in the graph
-     * traversing the graph up/down.
-     */
-    enum class Strategy {
-      None = 0,
-      InsertAndReduce = 1,
-      FloodFill = 2, /* Propagate FP16 Cast operations up and FP32 operations down */
-    };
-    using Strategy_t = std::underlying_type<Strategy>::type;
-    friend constexpr Strategy operator|(const Strategy s1, const Strategy s2) {
-      return static_cast<Strategy>(static_cast<Strategy_t>(s1) | static_cast<Strategy_t>(s2));
-    }
-
-    friend Strategy& operator|=(Strategy& s1, Strategy s2) {
-      s1 = s1 | s2;
-      return s1;
-    }
-
-    friend constexpr Strategy operator&(const Strategy s1, const Strategy s2) {
-      return static_cast<Strategy>(static_cast<Strategy_t>(s1) & static_cast<Strategy_t>(s2));
-    }
-
-    friend constexpr Strategy& operator&=(Strategy& s1, Strategy s2) {
-      s1 = s1 & s2;
-      return s1;
-    }
-
-    friend constexpr bool operator==(Strategy s1, Strategy s2) {
-      return static_cast<Strategy_t>(s1) == static_cast<Strategy_t>(s2);
-    }
-
-    friend constexpr bool operator!=(Strategy s1, Strategy s2) {
-      return (s1 == s2) == false;
-    }
-
-    int level{1}; /* -1 => no cast propagation,
-                       0 => use user specified list of opcodes to allow moving cast operations,
-                       1 => use ORT predefined list of level 1 opcodes in addition to the user specified allow opcodes
-                       2 => use ORT predefined list of level 2 opcodes in addition to the user specified allow opcodes
-                    */
-    Strategy strategy = Strategy::FloodFill;
-    // List of allowed opcodes to consider as safe to execute in float16, while moving cast operations
-    std::vector<std::string> allow;
-  };
-
-  PropagateCastOpsConfiguration propagate_cast_ops_config;
-};
-
-// The following declarations are required to refer to these operators in pybind11.
-constexpr GraphTransformerConfiguration::PropagateCastOpsConfiguration::Strategy operator|(GraphTransformerConfiguration::PropagateCastOpsConfiguration::Strategy,
-                                                                                           GraphTransformerConfiguration::PropagateCastOpsConfiguration::Strategy);
-constexpr GraphTransformerConfiguration::PropagateCastOpsConfiguration::Strategy operator&(GraphTransformerConfiguration::PropagateCastOpsConfiguration::Strategy,
-                                                                                           GraphTransformerConfiguration::PropagateCastOpsConfiguration::Strategy);
-constexpr bool operator==(GraphTransformerConfiguration::PropagateCastOpsConfiguration::Strategy, GraphTransformerConfiguration::PropagateCastOpsConfiguration::Strategy);
-constexpr bool operator!=(GraphTransformerConfiguration::PropagateCastOpsConfiguration::Strategy, GraphTransformerConfiguration::PropagateCastOpsConfiguration::Strategy);
-
-}  // namespace onnxruntime
\ No newline at end of file
diff --git a/projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer_level.h b/projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer_level.h
deleted file mode 100644
index 111f38f9..00000000
--- a/projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer_level.h
+++ /dev/null
@@ -1,19 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-namespace onnxruntime {
-
-// Graph transformer level
-// refer to docs/ONNX_Runtime_Graph_Optimizations.md for details
-enum class TransformerLevel : int {
-  Default = 0,  // required transformers only
-  Level1,       // basic optimizations
-  Level2,       // extended optimizations
-  Level3,       // layout optimizations
-  // The max level should always be same as the last level.
-  MaxLevel = Level3
-};
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer_utils.h b/projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer_utils.h
deleted file mode 100644
index e609745b..00000000
--- a/projects/llm_framework/include/onnxruntime/core/optimizer/graph_transformer_utils.h
+++ /dev/null
@@ -1,86 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <memory>
-#include <unordered_set>
-#include <vector>
-
-#include "core/common/inlined_containers.h"
-#include "core/framework/session_options.h"
-#include "core/optimizer/graph_transformer.h"
-
-#if !defined(ORT_MINIMAL_BUILD)
-#include "core/optimizer/rule_based_graph_transformer.h"
-#include "core/optimizer/rewrite_rule.h"
-#endif
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-#include "core/optimizer/selectors_actions/selector_action_transformer_apply_contexts.h"
-#endif
-
-namespace onnxruntime {
-class IExecutionProvider;
-
-namespace optimizer_utils {
-
-#if !defined(ORT_MINIMAL_BUILD)
-
-/** Generates all predefined rules for this level.
-   If rules_to_enable is not empty, it returns the intersection of predefined rules and rules_to_enable.
-   TODO: This is visible for testing at the moment, but we should rather make it private. */
-InlinedVector<std::unique_ptr<RewriteRule>> GenerateRewriteRules(
-    TransformerLevel level,
-    const InlinedHashSet<std::string>& rules_to_disable = {});
-
-/** Given a TransformerLevel, this method generates a name for the rule-based graph transformer of that level. */
-std::string GenerateRuleBasedTransformerName(TransformerLevel level);
-
-/** Generates all rule-based transformers for this level. */
-std::unique_ptr<RuleBasedGraphTransformer> GenerateRuleBasedGraphTransformer(
-    TransformerLevel level,
-    const InlinedHashSet<std::string>& rules_to_disable,
-    const InlinedHashSet<std::string_view>& compatible_execution_providers);
-
-/** Generates all predefined (both rule-based and non-rule-based) transformers for this level.
-    Any transformers or rewrite rules named in rules_and_transformers_to_disable will be excluded. */
-InlinedVector<std::unique_ptr<GraphTransformer>> GenerateTransformers(
-    TransformerLevel level,
-    const SessionOptions& session_options,
-    const IExecutionProvider& execution_provider /*required by constant folding*/,
-    const InlinedHashSet<std::string>& rules_and_transformers_to_disable = {});
-
-#endif  // !defined(ORT_MINIMAL_BUILD)
-
-#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-
-/** Generates all predefined transformers which can be used to provide runtime optimizations for this level
-    in a minimal build.
-    Any transformers or rewrite rules named in rules_and_transformers_to_disable will be excluded.
-
-    This is a distinct function from GenerateTransformers() because:
-    - An ORT format model used in a minimal build will have been pre-optimized to at least level 1 when created, so
-      level 1 transformers are not included.
-    - In a minimal build we have limited optimization/Graph capabilities
-      - Graph::Resolve is not available so the transformer must keep the Graph in a valid state
-      - Limited graph_utils capabilities are included
-    - Only a small subset of transformers support storing/replaying runtime optimizations with an ORT format model
-      - this capability is provided by the SelectionActionTransformer infrastructure
-      - the logic to determine the set of nodes a transformer should modify is captured during creation of the ORT
-        format model
-      - this information is saved in the ORT format model
-      - only the logic to modify the set of nodes is included in the minimal build
-    - The QDQFinalCleanupTransformer and NhwcTransformer transformers are also supported in a minimal build
-*/
-InlinedVector<std::unique_ptr<GraphTransformer>> GenerateTransformersForMinimalBuild(
-    TransformerLevel level,
-    const SessionOptions& session_options,
-    const SatApplyContextVariant& apply_context,
-    const IExecutionProvider& cpu_execution_provider,
-    const InlinedHashSet<std::string>& rules_and_transformers_to_disable = {});
-
-#endif  // !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
-
-}  // namespace optimizer_utils
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/optimizer/rewrite_rule.h b/projects/llm_framework/include/onnxruntime/core/optimizer/rewrite_rule.h
deleted file mode 100644
index 43e7489b..00000000
--- a/projects/llm_framework/include/onnxruntime/core/optimizer/rewrite_rule.h
+++ /dev/null
@@ -1,89 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include "core/common/common.h"
-#include "core/graph/graph_viewer.h"
-
-namespace onnxruntime {
-
-/**
-@class RewriteRule 
-
-The base class for a rewrite rule. A rewrite rule represents a semantics-preserving transformation of a 
-computation graph. It can be used to represent, for example, the elimination of operators that serve as 
-no-ops (e.g., dropout during inference), as well as inlining of "function" definitions or the dual operation 
-of replacing a complex expression by an equivalent function-call). Unlike the more general GraphTransformer, 
-a rewrite rule is a more local transformation that is triggered on a particular node of the graph. 
-
-Each rule has a set of conditions and a body. The conditions have to be satisfied for the body of the rule 
-to be triggered. Therefore, when creating a new rewrite rule, two main functions have to be implemented: 
-- SatisfyCondition defines the condition checks. It is advisable to add the more selective checks first, 
-  because those will lead to discarding fast rules that cannot be applied on a node.
-- Apply is the actual body of the rule that will be executed if SatisfyCondition returns true for a particular
-  node. Note that additional, more complex checks can be included in the Apply if putting them in the
-  SatisfyCondition would lead to duplicate work (e.g., when we make a check on a Node attribute but we need
-  that attribute to execute the rule too).
-In general, simple fast checks are a better fit for SatisfyCondition, whereas more complex ones can be added 
-in the Apply.
-
-In order to avoid evaluating the SatisfyCondition for each rule and each node of the graph, each rewrite rule
-should specify the target op types for which a rule will be evaluated, by overriding the TargetOpTypes() function.
-If the op type of a node is not included in the target op types of a rule, that rule would not be considered at all.
-If the list of op types is left empty, that rule will be triggered for every op type.
-*/
-class RewriteRule {
- public:
-  /**
-  @class RewriteRuleEffect
-
-  Class used to indicate the effect of rule application on a graph's node.
-  */
-  enum class RewriteRuleEffect : uint8_t {
-    kNone,                // The rewrite rule has not modified the graph.
-    kUpdatedCurrentNode,  // The rewrite rule updated (but did not remove) the node on which it was triggered.
-    kRemovedCurrentNode,  // The rewrite rule removed the node on which it was triggered.
-    kModifiedRestOfGraph  // The rewrite rule modified nodes other than the one it was triggered on.
-  };
-
-  RewriteRule(const std::string& name) : name_(name) {}
-
-  virtual ~RewriteRule() = default;
-
-  /** Gets the name of this rewrite rule. */
-  const std::string& Name() const noexcept {
-    return name_;
-  }
-
-  /** Returns the node op types for which this rule will be triggered. If the op type of a node is not included in the
-      target op types of a rule, that rule would not be considered at all. Returning an empty list indicates that we
-      will attempt to trigger the rule for every op type. */
-  virtual std::vector<std::string> TargetOpTypes() const noexcept = 0;
-
-  /** Checks if the condition of the rule is satisfied, and if so applies the body of the rule.
-      @param[in] graph The Graph.
-      @param[in] node The Node to apply the rewrite to.
-      @param[out] rule_effect Enum to indicate if and how the graph was modified as a result of the rule application.
-      @returns Status indicating success or providing error information */
-  common::Status CheckConditionAndApply(Graph& graph, Node& node, RewriteRuleEffect& rule_effect, const logging::Logger& logger) const {
-    return SatisfyCondition(graph, node, logger) ? Apply(graph, node, rule_effect, logger) : Status::OK();
-  }
-
- private:
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(RewriteRule);
-
-  const std::string name_;
-
-  /** Checks if the Node of the given Graph satisfies the conditions of this rule. The body of the rule will be 
-      evaluated if this condition function returns true. This can include a more complex pattern matching (conditions 
-      on the ascending or descending nodes of the node for which this rule was triggered) or some other properties 
-      of the nodes. */
-  virtual bool SatisfyCondition(const Graph& graph, const Node& node, const logging::Logger& logger) const = 0;
-
-  /** This is the actual body of the rule that performs the graph transformation. The transformation happens in-place. 
-      The return-value of node may be different from the input-value due to rewriting.
-      The value of "rule_effect" indicates whether and how the graph was modified by the rule. */
-  virtual common::Status Apply(Graph& graph, Node& node, RewriteRuleEffect& rule_effect, const logging::Logger& logger) const = 0;
-};
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/optimizer/rule_based_graph_transformer.h b/projects/llm_framework/include/onnxruntime/core/optimizer/rule_based_graph_transformer.h
deleted file mode 100644
index 039aed0b..00000000
--- a/projects/llm_framework/include/onnxruntime/core/optimizer/rule_based_graph_transformer.h
+++ /dev/null
@@ -1,82 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include "core/common/common.h"
-#include "core/graph/graph_viewer.h"
-#include "core/optimizer/graph_transformer.h"
-#include "core/optimizer/rewrite_rule.h"
-
-namespace onnxruntime {
-
-/**
-@class RuleBasedGraphTransformer
-
-Rule-based graph transformer that provides an API to register rewrite rules 
-and an API to apply all applicable rules to a Graph.
-
-Represents an IGraphTransformer determined by a set of rewrite rules.
-The transformer will apply all the rewrite rules iteratively as determined by the underlying rewriting strategy.
-Several rewriting-strategies are possible when traversing the graph and applying rewrite rules, 
-each with different trade offs. At the moment, we define one that performs top-down traversal of nodes.
-
-@TODO: Is a bottom-up traversal more efficient?
-@TODO: Is it worth adding the max number of passes a rule should be applied for?
-@TODO: We need to define a contract about whether a rewrite rule is allowed to leave
-       the graph in an inconsistent state (this will determine when and where we will be
-       calling Graph::resolve().
-*/
-class RuleBasedGraphTransformer : public GraphTransformer {
- public:
-  RuleBasedGraphTransformer(const std::string& name,
-                            const InlinedHashSet<std::string_view>& compatible_execution_providers = {})
-      : GraphTransformer(name, compatible_execution_providers) {}
-
-  /** Registers a rewrite rule in this transformer. */
-  Status Register(std::unique_ptr<RewriteRule> rule);
-
-  /** Gets the list of registered rewrite rules that will be triggered on nodes with the given op type 
-      by this rule-based transformer.
-      @returns a pointer to the vector containing all the registered rewrite rules. */
-  const InlinedVector<std::reference_wrapper<const RewriteRule>>* GetRewriteRulesForOpType(const std::string& op_type) const {
-    auto rules = op_type_to_rules_.find(op_type);
-    return (rules != op_type_to_rules_.cend()) ? &rules->second : nullptr;
-  }
-
-  /** Gets the rewrite rules that are evaluated on all nodes irrespective of their op type.
-      @returns a pointer to the vector containing all such rewrite rules or nullptr if no such rule. */
-  const InlinedVector<std::reference_wrapper<const RewriteRule>>* GetAnyOpRewriteRules() const {
-    return &any_op_type_rules_;
-  }
-
-  /** Returns the total number of rules that are registered in this transformer. */
-  size_t RulesCount() const;
-
- protected:
-  /** Applies the given set of rewrite rules on the Node of this Graph.
-      @param[in] graph The Graph.
-      @param[in] node The Node to apply the rules to.
-      @param[in] rules The vector of RewriteRules that will be applied to the Node.
-      @param[out] rule_effect Enum that indicates whether and how the graph was modified as a result of
-      applying rules on this node.
-      @returns Status indicating success or providing error information. */
-  common::Status ApplyRulesOnNode(Graph& graph, Node& node,
-                                  gsl::span<const std::reference_wrapper<const RewriteRule>> rules,
-                                  RewriteRule::RewriteRuleEffect& rule_effect, const logging::Logger& logger) const;
-
- private:
-  using RuleEffect = RewriteRule::RewriteRuleEffect;
-
-  // The list of unique pointers for all rules (so that rules can be registered for several op types).
-  InlinedVector<std::unique_ptr<RewriteRule>> rules_;
-  // Map that associates a node's op type with the vector of rules that are registered to be triggered for that node.
-  InlinedHashMap<std::string, InlinedVector<std::reference_wrapper<const RewriteRule>>> op_type_to_rules_;
-  // Rules that will be evaluated regardless of the op type of the node.
-  InlinedVector<std::reference_wrapper<const RewriteRule>> any_op_type_rules_;
-
-  // Performs a single top-down traversal of the graph and applies all registered rules.
-  common::Status ApplyImpl(Graph& graph, bool& modified, int graph_level, const logging::Logger& logger) const override;
-};
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/providers/cpu/cpu_provider_factory.h b/projects/llm_framework/include/onnxruntime/core/providers/cpu/cpu_provider_factory.h
deleted file mode 100644
index 29267869..00000000
--- a/projects/llm_framework/include/onnxruntime/core/providers/cpu/cpu_provider_factory.h
+++ /dev/null
@@ -1,19 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#include "onnxruntime_c_api.h"
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/**
- * \param use_arena zero: false. non-zero: true.
- */
-ORT_EXPORT
-ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_CPU, _In_ OrtSessionOptions* options, int use_arena)
-ORT_ALL_ARGS_NONNULL;
-
-#ifdef __cplusplus
-}
-#endif
diff --git a/projects/llm_framework/include/onnxruntime/core/session/environment.h b/projects/llm_framework/include/onnxruntime/core/session/environment.h
deleted file mode 100644
index cc511cca..00000000
--- a/projects/llm_framework/include/onnxruntime/core/session/environment.h
+++ /dev/null
@@ -1,96 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <atomic>
-#include <memory>
-#include "core/common/common.h"
-#include "core/common/status.h"
-#include "core/platform/threadpool.h"
-#include "core/common/logging/logging.h"
-#include "core/framework/allocator.h"
-
-struct OrtThreadingOptions;
-namespace onnxruntime {
-/** TODO: remove this class
-   Provides the runtime environment for onnxruntime.
-   Create one instance for the duration of execution.
-*/
-class Environment {
- public:
-  /**
-     Create and initialize the runtime environment.
-    @param logging manager instance that will enable per session logger output using
-    session_options.session_logid as the logger id in messages.
-    If nullptr, the default LoggingManager MUST have been created previously as it will be used
-    for logging. This will use the default logger id in messages.
-    See core/common/logging/logging.h for details, and how LoggingManager::DefaultLogger works.
-    @param tp_options optional set of parameters controlling the number of intra and inter op threads for the global
-    threadpools.
-    @param create_global_thread_pools determine if this function will create the global threadpools or not.
-  */
-  static Status Create(std::unique_ptr<logging::LoggingManager> logging_manager,
-                       std::unique_ptr<Environment>& environment,
-                       const OrtThreadingOptions* tp_options = nullptr,
-                       bool create_global_thread_pools = false);
-
-  logging::LoggingManager* GetLoggingManager() const {
-    return logging_manager_.get();
-  }
-
-  void SetLoggingManager(std::unique_ptr<onnxruntime::logging::LoggingManager> logging_manager) {
-    logging_manager_ = std::move(logging_manager);
-  }
-
-  onnxruntime::concurrency::ThreadPool* GetIntraOpThreadPool() const {
-    return intra_op_thread_pool_.get();
-  }
-
-  onnxruntime::concurrency::ThreadPool* GetInterOpThreadPool() const {
-    return inter_op_thread_pool_.get();
-  }
-
-  bool EnvCreatedWithGlobalThreadPools() const {
-    return create_global_thread_pools_;
-  }
-
-  /**
-   * Registers an allocator for sharing between multiple sessions.
-   * Return an error if an allocator with the same OrtMemoryInfo is already registered.
-   */
-  Status RegisterAllocator(AllocatorPtr allocator);
-
-  /**
-   * Creates and registers an allocator for sharing between multiple sessions.
-   * Return an error if an allocator with the same OrtMemoryInfo is already registered.
-   */
-  Status CreateAndRegisterAllocator(const OrtMemoryInfo& mem_info, const OrtArenaCfg* arena_cfg = nullptr);
-
-  /**
-   * Returns the list of registered allocators in this env.
-   */
-  const std::vector<AllocatorPtr>& GetRegisteredSharedAllocators() const {
-    return shared_allocators_;
-  }
-
-  /**
-   * Removes registered allocator that was previously registered for sharing between multiple sessions.
-   */
-  Status UnregisterAllocator(const OrtMemoryInfo& mem_info);
-
-  Environment() = default;
-
- private:
-  ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(Environment);
-  Status Initialize(std::unique_ptr<logging::LoggingManager> logging_manager,
-                    const OrtThreadingOptions* tp_options = nullptr,
-                    bool create_global_thread_pools = false);
-
-  std::unique_ptr<logging::LoggingManager> logging_manager_;
-  std::unique_ptr<onnxruntime::concurrency::ThreadPool> intra_op_thread_pool_;
-  std::unique_ptr<onnxruntime::concurrency::ThreadPool> inter_op_thread_pool_;
-  bool create_global_thread_pools_{false};
-  std::vector<AllocatorPtr> shared_allocators_;
-};
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/onnxruntime/core/session/experimental_onnxruntime_cxx_api.h b/projects/llm_framework/include/onnxruntime/core/session/experimental_onnxruntime_cxx_api.h
deleted file mode 100644
index 6efb20ad..00000000
--- a/projects/llm_framework/include/onnxruntime/core/session/experimental_onnxruntime_cxx_api.h
+++ /dev/null
@@ -1,69 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-// Summary: The experimental Ort C++ API is a wrapper around the Ort C++ API.
-//
-// This C++ API further simplifies usage and provides support for modern C++ syntax/features
-// at the cost of some overhead (i.e. using std::string over char *).
-//
-// Where applicable, default memory allocator options are used unless explicitly set.
-//
-// Experimental components are designed as drop-in replacements of the regular API, requiring
-// minimal code modifications to use.
-//
-// Example:  Ort::Session -> Ort::Experimental::Session
-//
-// NOTE: Experimental API components are subject to change based on feedback and provide no
-// guarantee of backwards compatibility in future releases.
-
-#pragma once
-#include "onnxruntime_cxx_api.h"
-
-namespace Ort {
-namespace Experimental {
-
-struct Session : Ort::Session {
-  Session(Env& env, std::basic_string<ORTCHAR_T>& model_path, SessionOptions& options)
-      : Ort::Session(env, model_path.data(), options){};
-  Session(Env& env, void* model_data, size_t model_data_length, SessionOptions& options)
-      : Ort::Session(env, model_data, model_data_length, options){};
-
-  // overloaded Run() with sensible defaults
-  std::vector<Ort::Value> Run(const std::vector<std::string>& input_names,
-                              const std::vector<Ort::Value>& input_values,
-                              const std::vector<std::string>& output_names,
-                              const RunOptions& run_options = RunOptions());
-  void Run(const std::vector<std::string>& input_names,
-           const std::vector<Ort::Value>& input_values,
-           const std::vector<std::string>& output_names,
-           std::vector<Ort::Value>& output_values,
-           const RunOptions& run_options = RunOptions());
-
-  // convenience methods that simplify common lower-level API calls
-  std::vector<std::string> GetInputNames() const;
-  std::vector<std::string> GetOutputNames() const;
-  std::vector<std::string> GetOverridableInitializerNames() const;
-
-  // NOTE: shape dimensions may have a negative value to indicate a symbolic/unknown dimension.
-  std::vector<std::vector<int64_t> > GetInputShapes() const;
-  std::vector<std::vector<int64_t> > GetOutputShapes() const;
-  std::vector<std::vector<int64_t> > GetOverridableInitializerShapes() const;
-};
-
-struct Value : Ort::Value {
-  Value(OrtValue* p)
-      : Ort::Value(p){};
-
-  template <typename T>
-  static Ort::Value CreateTensor(T* p_data, size_t p_data_element_count, const std::vector<int64_t>& shape);
-  static Ort::Value CreateTensor(void* p_data, size_t p_data_byte_count, const std::vector<int64_t>& shape, ONNXTensorElementDataType type);
-
-  template <typename T>
-  static Ort::Value CreateTensor(const std::vector<int64_t>& shape);
-  static Ort::Value CreateTensor(const std::vector<int64_t>& shape, ONNXTensorElementDataType type);
-};
-
-}
-}
-
-#include "experimental_onnxruntime_cxx_inline.h"
diff --git a/projects/llm_framework/include/onnxruntime/core/session/experimental_onnxruntime_cxx_inline.h b/projects/llm_framework/include/onnxruntime/core/session/experimental_onnxruntime_cxx_inline.h
deleted file mode 100644
index e754d598..00000000
--- a/projects/llm_framework/include/onnxruntime/core/session/experimental_onnxruntime_cxx_inline.h
+++ /dev/null
@@ -1,108 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-// Do not include this file directly. Please include "experimental_onnxruntime_cxx_api.h" instead.
-//
-// These are the inline implementations of the C++ header APIs. They are in this separate file as to not clutter
-// the main C++ file with implementation details.
-namespace Ort {
-namespace Experimental {
-
-inline std::vector<Ort::Value> Session::Run(const std::vector<std::string>& input_names, const std::vector<Ort::Value>& input_values,
-                                            const std::vector<std::string>& output_names, const RunOptions& run_options) {
-  size_t output_names_count = GetOutputNames().size();
-  std::vector<Ort::Value> output_values;
-  for (size_t i = 0; i < output_names_count; i++) output_values.emplace_back(nullptr);
-  Run(input_names, input_values, output_names, output_values, run_options);
-  return output_values;
-}
-
-inline void Session::Run(const std::vector<std::string>& input_names, const std::vector<Ort::Value>& input_values,
-                         const std::vector<std::string>& output_names, std::vector<Ort::Value>& output_values, const RunOptions& run_options) {
-  size_t input_names_count = input_names.size();
-  size_t output_names_count = output_names.size();
-  std::vector<const char*> input_names_(input_names_count, nullptr);
-  size_t i = 0;
-  for (auto it = input_names.begin(); it != input_names.end(); it++) input_names_[i++] = (*it).c_str();
-  std::vector<const char*> output_names_(output_names_count, nullptr);
-  i = 0;
-  for (auto it = output_names.begin(); it != output_names.end(); it++) output_names_[i++] = (*it).c_str();
-  Ort::Session::Run(run_options, input_names_.data(), input_values.data(), input_names_count, output_names_.data(), output_values.data(), output_names_count);
-}
-
-inline std::vector<std::string> Session::GetInputNames() const {
-  Ort::AllocatorWithDefaultOptions allocator;
-  size_t node_count = GetInputCount();
-  std::vector<std::string> out(node_count);
-  for (size_t i = 0; i < node_count; i++) {
-    auto tmp = GetInputNameAllocated(i, allocator);
-    out[i] = tmp.get();
-  }
-  return out;
-}
-
-inline std::vector<std::string> Session::GetOutputNames() const {
-  Ort::AllocatorWithDefaultOptions allocator;
-  size_t node_count = GetOutputCount();
-  std::vector<std::string> out(node_count);
-  for (size_t i = 0; i < node_count; i++) {
-    auto tmp = GetOutputNameAllocated(i, allocator);
-    out[i] = tmp.get();
-  }
-  return out;
-}
-
-inline std::vector<std::string> Session::GetOverridableInitializerNames() const {
-  Ort::AllocatorWithDefaultOptions allocator;
-  size_t init_count = GetOverridableInitializerCount();
-  std::vector<std::string> out(init_count);
-  for (size_t i = 0; i < init_count; i++) {
-    auto tmp = GetOverridableInitializerNameAllocated(i, allocator);
-    out[i] = tmp.get();
-  }
-  return out;
-}
-
-inline std::vector<std::vector<int64_t>> Session::GetInputShapes() const {
-  size_t node_count = GetInputCount();
-  std::vector<std::vector<int64_t>> out(node_count);
-  for (size_t i = 0; i < node_count; i++) out[i] = GetInputTypeInfo(i).GetTensorTypeAndShapeInfo().GetShape();
-  return out;
-}
-
-inline std::vector<std::vector<int64_t>> Session::GetOutputShapes() const {
-  size_t node_count = GetOutputCount();
-  std::vector<std::vector<int64_t>> out(node_count);
-  for (size_t i = 0; i < node_count; i++) out[i] = GetOutputTypeInfo(i).GetTensorTypeAndShapeInfo().GetShape();
-  return out;
-}
-
-inline std::vector<std::vector<int64_t>> Session::GetOverridableInitializerShapes() const {
-  size_t init_count = GetOverridableInitializerCount();
-  std::vector<std::vector<int64_t>> out(init_count);
-  for (size_t i = 0; i < init_count; i++) out[i] = GetOverridableInitializerTypeInfo(i).GetTensorTypeAndShapeInfo().GetShape();
-  return out;
-}
-
-template <typename T>
-inline Ort::Value Value::CreateTensor(T* p_data, size_t p_data_element_count, const std::vector<int64_t>& shape) {
-  return CreateTensor(p_data, p_data_element_count * sizeof(T), shape, TypeToTensorType<T>::type);
-}
-
-inline Ort::Value Value::CreateTensor(void* p_data, size_t p_data_byte_count, const std::vector<int64_t>& shape, ONNXTensorElementDataType type) {
-  Ort::MemoryInfo info = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
-  return Ort::Value::CreateTensor(info, p_data, p_data_byte_count, shape.data(), shape.size(), type);
-}
-
-template <typename T>
-inline Ort::Value Value::CreateTensor(const std::vector<int64_t>& shape) {
-  return CreateTensor(shape, TypeToTensorType<T>::type);
-}
-
-inline Ort::Value Value::CreateTensor(const std::vector<int64_t>& shape, ONNXTensorElementDataType type) {
-  Ort::AllocatorWithDefaultOptions allocator;
-  return Ort::Value::CreateTensor(allocator, shape.data(), shape.size(), type);
-}
-
-}
-}
diff --git a/projects/llm_framework/include/onnxruntime/core/session/onnxruntime_c_api.h b/projects/llm_framework/include/onnxruntime/core/session/onnxruntime_c_api.h
deleted file mode 100644
index 44875b0f..00000000
--- a/projects/llm_framework/include/onnxruntime/core/session/onnxruntime_c_api.h
+++ /dev/null
@@ -1,3987 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-// See docs\c_cxx\README.md on generating the Doxygen documentation from this file
-
-/** \mainpage C & C++ APIs
- *
- * <h1>C</h1>
- *
- * ::OrtApi - Click here to go to the structure with all C API functions.
- *
- * <h1>C++</h1>
- *
- * ::Ort - Click here to go to the namespace holding all of the C++ wrapper classes
- *
- * It is a set of header only wrapper classes around the C API. The goal is to turn the C style return value error codes into C++ exceptions, and to
- * automate memory management through standard C++ RAII principles.
- *
- * \addtogroup Global
- * ONNX Runtime C API
- * @{
- */
-
-#pragma once
-#include <stdlib.h>
-#include <stdint.h>
-#include <string.h>
-
-/** \brief The API version defined in this header
- *
- * This value is used by some API functions to behave as this version of the header expects.
- */
-#define ORT_API_VERSION 14
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-//! @}
-// SAL2 Definitions
-#ifndef _WIN32
-#define _In_
-#define _In_z_
-#define _In_opt_
-#define _In_opt_z_
-#define _Out_
-#define _Outptr_
-#define _Out_opt_
-#define _Inout_
-#define _Inout_opt_
-#define _Frees_ptr_opt_
-#define _Ret_maybenull_
-#define _Ret_notnull_
-#define _Check_return_
-#define _Outptr_result_maybenull_
-#define _In_reads_(X)
-#define _Inout_updates_all_(X)
-#define _Out_writes_bytes_all_(X)
-#define _Out_writes_all_(X)
-#define _Success_(X)
-#define _Outptr_result_buffer_maybenull_(X)
-#define ORT_ALL_ARGS_NONNULL __attribute__((nonnull))
-#else
-#include <specstrings.h>
-#define ORT_ALL_ARGS_NONNULL
-#endif
-
-#ifdef _WIN32
-// Define ORT_DLL_IMPORT if your program is dynamically linked to Ort.
-// dllexport is not used, we use a .def file.
-#ifdef ORT_DLL_IMPORT
-#define ORT_EXPORT __declspec(dllimport)
-#else
-#define ORT_EXPORT
-#endif
-#define ORT_API_CALL _stdcall
-#define ORT_MUST_USE_RESULT
-#define ORTCHAR_T wchar_t
-#else
-// To make symbols visible on macOS/iOS
-#ifdef __APPLE__
-#define ORT_EXPORT __attribute__((visibility("default")))
-#else
-#define ORT_EXPORT
-#endif
-#define ORT_API_CALL
-#define ORT_MUST_USE_RESULT __attribute__((warn_unused_result))
-#define ORTCHAR_T char
-#endif
-
-#ifndef ORT_TSTR
-#ifdef _WIN32
-#define ORT_TSTR(X) L##X
-#else
-#define ORT_TSTR(X) X
-#endif
-#endif
-
-// Any pointer marked with _In_ or _Out_, cannot be NULL.
-
-// Windows users should use unicode paths when possible to bypass the MAX_PATH limitation
-// Every pointer marked with _In_ or _Out_, cannot be NULL. Caller should ensure that.
-// for ReleaseXXX(...) functions, they can accept NULL pointer.
-
-#ifdef __cplusplus
-// For any compiler with C++11 support, MSVC 2015 and greater, or Clang version supporting noexcept.
-// Such complex condition is needed because compilers set __cplusplus value differently.
-#ifndef __has_feature
-#define __has_feature(x) 0
-#endif
-#if ((__cplusplus >= 201103L) || (_MSC_VER >= 1900) || (defined(__has_feature) && __has_feature(cxx_noexcept)))
-#define NO_EXCEPTION noexcept
-#else
-#define NO_EXCEPTION throw()
-#endif
-#else
-#define NO_EXCEPTION
-#endif
-
-// __VA_ARGS__ on Windows and Linux are different
-#define ORT_API(RETURN_TYPE, NAME, ...) RETURN_TYPE ORT_API_CALL NAME(__VA_ARGS__) NO_EXCEPTION
-
-#define ORT_API_STATUS(NAME, ...)                                                                   \
-  _Success_(return == 0) _Check_return_ _Ret_maybenull_ OrtStatusPtr ORT_API_CALL NAME(__VA_ARGS__) \
-  NO_EXCEPTION ORT_MUST_USE_RESULT
-
-// XXX: Unfortunately, SAL annotations are known to not work with function pointers
-#define ORT_API2_STATUS(NAME, ...) \
-  _Check_return_ _Ret_maybenull_ OrtStatusPtr(ORT_API_CALL* NAME)(__VA_ARGS__) NO_EXCEPTION ORT_MUST_USE_RESULT
-
-// Used in *.cc files. Almost as same as ORT_API_STATUS, except without ORT_MUST_USE_RESULT and ORT_EXPORT
-#define ORT_API_STATUS_IMPL(NAME, ...) \
-  _Success_(return == 0) _Check_return_ _Ret_maybenull_ OrtStatusPtr ORT_API_CALL NAME(__VA_ARGS__) NO_EXCEPTION
-
-#define ORT_CLASS_RELEASE(X) void(ORT_API_CALL * Release##X)(_Frees_ptr_opt_ Ort##X * input)
-
-#ifdef __DOXYGEN__
-#undef ORT_API_STATUS
-#define ORT_API_STATUS(NAME, ...) OrtStatus* NAME(__VA_ARGS__)
-#undef ORT_API2_STATUS
-#define ORT_API2_STATUS(NAME, ...) OrtStatus* NAME(__VA_ARGS__)
-#undef ORT_CLASS_RELEASE
-#define ORT_CLASS_RELEASE(X) void Release##X(Ort##X* input)
-#undef NO_EXCEPTION
-#define NO_EXCEPTION
-#endif
-/** \addtogroup Global
- * ONNX Runtime C API
- * @{
- */
-
-/** Copied from TensorProto::DataType
- * Currently, Ort doesn't support complex64, complex128
- */
-typedef enum ONNXTensorElementDataType {
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_UNDEFINED,
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT,   // maps to c type float
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT8,   // maps to c type uint8_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_INT8,    // maps to c type int8_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT16,  // maps to c type uint16_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_INT16,   // maps to c type int16_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32,   // maps to c type int32_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64,   // maps to c type int64_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING,  // maps to c++ type std::string
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_BOOL,
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT16,
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_DOUBLE,      // maps to c type double
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT32,      // maps to c type uint32_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT64,      // maps to c type uint64_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_COMPLEX64,   // complex with float32 real and imaginary components
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_COMPLEX128,  // complex with float64 real and imaginary components
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_BFLOAT16     // Non-IEEE floating-point format based on IEEE754 single-precision
-} ONNXTensorElementDataType;
-
-// Synced with onnx TypeProto oneof
-typedef enum ONNXType {
-  ONNX_TYPE_UNKNOWN,
-  ONNX_TYPE_TENSOR,
-  ONNX_TYPE_SEQUENCE,
-  ONNX_TYPE_MAP,
-  ONNX_TYPE_OPAQUE,
-  ONNX_TYPE_SPARSETENSOR,
-  ONNX_TYPE_OPTIONAL
-} ONNXType;
-
-// These types are synced with internal
-// SparseFormatFlags
-typedef enum OrtSparseFormat {
-  ORT_SPARSE_UNDEFINED = 0,
-  ORT_SPARSE_COO = 0x1,
-  ORT_SPARSE_CSRC = 0x2,
-  ORT_SPARSE_BLOCK_SPARSE = 0x4
-} OrtSparseFormat;
-
-// Enum allows to query sparse tensor indices
-enum OrtSparseIndicesFormat {
-  ORT_SPARSE_COO_INDICES,
-  ORT_SPARSE_CSR_INNER_INDICES,
-  ORT_SPARSE_CSR_OUTER_INDICES,
-  ORT_SPARSE_BLOCK_SPARSE_INDICES
-};
-
-/** \brief Logging severity levels
- *
- * In typical API usage, specifying a logging severity level specifies the minimum severity of log messages to show.
- */
-typedef enum OrtLoggingLevel {
-  ORT_LOGGING_LEVEL_VERBOSE,  ///< Verbose informational messages (least severe).
-  ORT_LOGGING_LEVEL_INFO,     ///< Informational messages.
-  ORT_LOGGING_LEVEL_WARNING,  ///< Warning messages.
-  ORT_LOGGING_LEVEL_ERROR,    ///< Error messages.
-  ORT_LOGGING_LEVEL_FATAL,    ///< Fatal error messages (most severe).
-} OrtLoggingLevel;
-
-typedef enum OrtErrorCode {
-  ORT_OK,
-  ORT_FAIL,
-  ORT_INVALID_ARGUMENT,
-  ORT_NO_SUCHFILE,
-  ORT_NO_MODEL,
-  ORT_ENGINE_ERROR,
-  ORT_RUNTIME_EXCEPTION,
-  ORT_INVALID_PROTOBUF,
-  ORT_MODEL_LOADED,
-  ORT_NOT_IMPLEMENTED,
-  ORT_INVALID_GRAPH,
-  ORT_EP_FAIL,
-} OrtErrorCode;
-
-typedef enum OrtOpAttrType {
-  ORT_OP_ATTR_UNDEFINED = 0,
-  ORT_OP_ATTR_INT,
-  ORT_OP_ATTR_INTS,
-  ORT_OP_ATTR_FLOAT,
-  ORT_OP_ATTR_FLOATS,
-  ORT_OP_ATTR_STRING,
-  ORT_OP_ATTR_STRINGS,
-} OrtOpAttrType;
-
-//! @}
-#define ORT_RUNTIME_CLASS(X) \
-  struct Ort##X;             \
-  typedef struct Ort##X Ort##X;
-
-/** \addtogroup Global
- * ONNX Runtime C API
- * @{
- */
-// The actual types defined have an Ort prefix
-ORT_RUNTIME_CLASS(Env);
-ORT_RUNTIME_CLASS(Status);  // nullptr for Status* indicates success
-ORT_RUNTIME_CLASS(MemoryInfo);
-ORT_RUNTIME_CLASS(IoBinding);
-ORT_RUNTIME_CLASS(Session);  // Don't call ReleaseSession from Dllmain (because session owns a thread pool)
-ORT_RUNTIME_CLASS(Value);
-ORT_RUNTIME_CLASS(RunOptions);
-ORT_RUNTIME_CLASS(TypeInfo);
-ORT_RUNTIME_CLASS(TensorTypeAndShapeInfo);
-ORT_RUNTIME_CLASS(SessionOptions);
-ORT_RUNTIME_CLASS(CustomOpDomain);
-ORT_RUNTIME_CLASS(MapTypeInfo);
-ORT_RUNTIME_CLASS(SequenceTypeInfo);
-ORT_RUNTIME_CLASS(ModelMetadata);
-ORT_RUNTIME_CLASS(ThreadPoolParams);
-ORT_RUNTIME_CLASS(ThreadingOptions);
-ORT_RUNTIME_CLASS(ArenaCfg);
-ORT_RUNTIME_CLASS(PrepackedWeightsContainer);
-ORT_RUNTIME_CLASS(TensorRTProviderOptionsV2);
-ORT_RUNTIME_CLASS(CUDAProviderOptionsV2);
-ORT_RUNTIME_CLASS(CANNProviderOptions);
-ORT_RUNTIME_CLASS(Op);
-ORT_RUNTIME_CLASS(OpAttr);
-
-#ifdef _WIN32
-typedef _Return_type_success_(return == 0) OrtStatus* OrtStatusPtr;
-#else
-typedef OrtStatus* OrtStatusPtr;
-#endif
-
-/** \brief Memory allocation interface
- *
- * Structure of function pointers that defines a memory allocator. This can be created and filled in by the user for custom allocators.
- *
- * When an allocator is passed to any function, be sure that the allocator object is not destroyed until the last allocated object using it is freed.
- */
-typedef struct OrtAllocator {
-  uint32_t version;                                                                   ///< Must be initialized to ORT_API_VERSION
-  void*(ORT_API_CALL* Alloc)(struct OrtAllocator* this_, size_t size);                ///< Returns a pointer to an allocated block of `size` bytes
-  void(ORT_API_CALL* Free)(struct OrtAllocator* this_, void* p);                      ///< Free a block of memory previously allocated with OrtAllocator::Alloc
-  const struct OrtMemoryInfo*(ORT_API_CALL* Info)(const struct OrtAllocator* this_);  ///< Return a pointer to an ::OrtMemoryInfo that describes this allocator
-} OrtAllocator;
-
-typedef void(ORT_API_CALL* OrtLoggingFunction)(
-    void* param, OrtLoggingLevel severity, const char* category, const char* logid, const char* code_location,
-    const char* message);
-
-/** \brief Graph optimization level
- *
- * Refer to https://www.onnxruntime.ai/docs/resources/graph-optimizations.html
- * for an in-depth understanding of Graph Optimizations
- */
-typedef enum GraphOptimizationLevel {
-  ORT_DISABLE_ALL = 0,
-  ORT_ENABLE_BASIC = 1,
-  ORT_ENABLE_EXTENDED = 2,
-  ORT_ENABLE_ALL = 99
-} GraphOptimizationLevel;
-
-typedef enum ExecutionMode {
-  ORT_SEQUENTIAL = 0,
-  ORT_PARALLEL = 1,
-} ExecutionMode;
-
-/** \brief Language projection identifiers
- * /see OrtApi::SetLanguageProjection
- */
-typedef enum OrtLanguageProjection {
-  ORT_PROJECTION_C = 0,
-  ORT_PROJECTION_CPLUSPLUS = 1,
-  ORT_PROJECTION_CSHARP = 2,
-  ORT_PROJECTION_PYTHON = 3,
-  ORT_PROJECTION_JAVA = 4,
-  ORT_PROJECTION_WINML = 5,
-  ORT_PROJECTION_NODEJS = 6,
-} OrtLanguageProjection;
-
-struct OrtKernelInfo;
-typedef struct OrtKernelInfo OrtKernelInfo;
-struct OrtKernelContext;
-typedef struct OrtKernelContext OrtKernelContext;
-struct OrtCustomOp;
-typedef struct OrtCustomOp OrtCustomOp;
-
-typedef enum OrtAllocatorType {
-  OrtInvalidAllocator = -1,
-  OrtDeviceAllocator = 0,
-  OrtArenaAllocator = 1
-} OrtAllocatorType;
-
-/** \brief Memory types for allocated memory, execution provider specific types should be extended in each provider.
- */
-// Whenever this struct is updated, please also update the MakeKey function in onnxruntime / core / framework / execution_provider.cc
-typedef enum OrtMemType {
-  OrtMemTypeCPUInput = -2,              ///< Any CPU memory used by non-CPU execution provider
-  OrtMemTypeCPUOutput = -1,             ///< CPU accessible memory outputted by non-CPU execution provider, i.e. CUDA_PINNED
-  OrtMemTypeCPU = OrtMemTypeCPUOutput,  ///< Temporary CPU accessible memory allocated by non-CPU execution provider, i.e. CUDA_PINNED
-  OrtMemTypeDefault = 0,                ///< The default allocator for execution provider
-} OrtMemType;
-
-/** \brief This mimics OrtDevice type constants so they can be returned in the API
- */
-typedef enum OrtMemoryInfoDeviceType {
-  OrtMemoryInfoDeviceType_CPU = 0,
-  OrtMemoryInfoDeviceType_GPU = 1,
-  OrtMemoryInfoDeviceType_FPGA = 2
-} OrtMemoryInfoDeviceType;
-
-/** \brief Algorithm to use for cuDNN Convolution Op
- */
-typedef enum OrtCudnnConvAlgoSearch {
-  OrtCudnnConvAlgoSearchExhaustive,  // expensive exhaustive benchmarking using cudnnFindConvolutionForwardAlgorithmEx
-  OrtCudnnConvAlgoSearchHeuristic,   // lightweight heuristic based search using cudnnGetConvolutionForwardAlgorithm_v7
-  OrtCudnnConvAlgoSearchDefault,     // default algorithm using CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM
-} OrtCudnnConvAlgoSearch;
-
-/** \brief CUDA Provider Options
- *
- * \see OrtApi::SessionOptionsAppendExecutionProvider_CUDA
- */
-typedef struct OrtCUDAProviderOptions {
-#ifdef __cplusplus
-  OrtCUDAProviderOptions()
-      : device_id{},
-        cudnn_conv_algo_search{OrtCudnnConvAlgoSearchExhaustive},
-        gpu_mem_limit{SIZE_MAX},
-        arena_extend_strategy{},
-        do_copy_in_default_stream{1},
-        has_user_compute_stream{},
-        user_compute_stream{},
-        default_memory_arena_cfg{},
-        tunable_op_enabled{false} {}
-#endif
-
-  /** \brief CUDA device Id
-   *   Defaults to 0.
-   */
-  int device_id;
-
-  /** \brief CUDA Convolution algorithm search configuration.
-   *   See enum OrtCudnnConvAlgoSearch for more details.
-   *   Defaults to OrtCudnnConvAlgoSearchExhaustive.
-   */
-  OrtCudnnConvAlgoSearch cudnn_conv_algo_search;
-
-  /** \brief CUDA memory limit (To use all possible memory pass in maximum size_t)
-   *   Defaults to SIZE_MAX.
-   *   \note If a ::OrtArenaCfg has been applied, it will override this field
-   */
-  size_t gpu_mem_limit;
-
-  /** \brief Strategy used to grow the memory arena
-   *   0 = kNextPowerOfTwo<br>
-   *   1 = kSameAsRequested<br>
-   *   Defaults to 0.
-   *   \note If a ::OrtArenaCfg has been applied, it will override this field
-   */
-  int arena_extend_strategy;
-
-  /** \brief Flag indicating if copying needs to take place on the same stream as the compute stream in the CUDA EP
-   *   0 = Use separate streams for copying and compute.
-   *   1 = Use the same stream for copying and compute.
-   *   Defaults to 1.
-   *   WARNING: Setting this to 0 may result in data races for some models.
-   *   Please see issue #4829 for more details.
-   */
-  int do_copy_in_default_stream;
-
-  /** \brief Flag indicating if there is a user provided compute stream
-   *   Defaults to 0.
-   */
-  int has_user_compute_stream;
-
-  /** \brief User provided compute stream.
-   *   If provided, please set `has_user_compute_stream` to 1.
-   */
-  void* user_compute_stream;
-
-  /** \brief CUDA memory arena configuration parameters
-   */
-  OrtArenaCfg* default_memory_arena_cfg;
-
-  /** \brief Enable TunableOp.
-   *   Set it to 1 to enable TunableOp. Otherwise, it is disabled by default.
-   *   This option can be superseded by environment variable ORT_CUDA_TUNABLE_OP_ENABLED.
-   */
-  int tunable_op_enabled;
-
-} OrtCUDAProviderOptions;
-
-/** \brief ROCM Provider Options
- *
- * \see OrtApi::SessionOptionsAppendExecutionProvider_ROCM
- */
-typedef struct OrtROCMProviderOptions {
-#ifdef __cplusplus
-  OrtROCMProviderOptions()
-      : device_id{},
-        miopen_conv_exhaustive_search{0},
-        gpu_mem_limit{SIZE_MAX},
-        arena_extend_strategy{},
-        do_copy_in_default_stream{1},
-        has_user_compute_stream{},
-        user_compute_stream{},
-        default_memory_arena_cfg{},
-        tunable_op_enabled{false} {}
-#endif
-
-  /** \brief ROCM device Id
-   *   Defaults to 0.
-   */
-  int device_id;
-
-  /** \brief ROCM MIOpen Convolution algorithm exaustive search option.
-   *   Defaults to 0 (false).
-   */
-  int miopen_conv_exhaustive_search;
-
-  /** \brief ROCM memory limit (To use all possible memory pass in maximum size_t)
-   *   Defaults to SIZE_MAX.
-   *   \note If a ::OrtArenaCfg has been applied, it will override this field
-   */
-  size_t gpu_mem_limit;
-
-  /** \brief Strategy used to grow the memory arena
-   *   0 = kNextPowerOfTwo<br>
-   *   1 = kSameAsRequested<br>
-   *   Defaults to 0.
-   *   \note If a ::OrtArenaCfg has been applied, it will override this field
-   */
-  int arena_extend_strategy;
-
-  /** \brief Flag indicating if copying needs to take place on the same stream as the compute stream in the ROCM EP
-   *   0 = Use separate streams for copying and compute.
-   *   1 = Use the same stream for copying and compute.
-   *   Defaults to 1.
-   *   WARNING: Setting this to 0 may result in data races for some models.
-   *   Please see issue #4829 for more details.
-   */
-  int do_copy_in_default_stream;
-
-  /** \brief Flag indicating if there is a user provided compute stream
-   *   Defaults to 0.
-   */
-  int has_user_compute_stream;
-
-  /** \brief User provided compute stream.
-   *   If provided, please set `has_user_compute_stream` to 1.
-   */
-  void* user_compute_stream;
-
-  /** \brief ROCM memory arena configuration parameters
-   */
-  OrtArenaCfg* default_memory_arena_cfg;
-
-  /** \brief Enable TunableOp.
-   *   Set it to 1 to enable TunableOp. Otherwise, it is disabled by default.
-   *   This option can be superseded by environment variable ORT_ROCM_TUNABLE_OP_ENABLED.
-   */
-  int tunable_op_enabled;
-
-} OrtROCMProviderOptions;
-
-/** \brief TensorRT Provider Options
- *
- * \see OrtApi::SessionOptionsAppendExecutionProvider_TensorRT
- */
-typedef struct OrtTensorRTProviderOptions {
-  int device_id;                                ///< CUDA device id (0 = default device)
-  int has_user_compute_stream;                  // indicator of user specified CUDA compute stream.
-  void* user_compute_stream;                    // user specified CUDA compute stream.
-  int trt_max_partition_iterations;             // maximum iterations for TensorRT parser to get capability
-  int trt_min_subgraph_size;                    // minimum size of TensorRT subgraphs
-  size_t trt_max_workspace_size;                // maximum workspace size for TensorRT.
-  int trt_fp16_enable;                          // enable TensorRT FP16 precision. Default 0 = false, nonzero = true
-  int trt_int8_enable;                          // enable TensorRT INT8 precision. Default 0 = false, nonzero = true
-  const char* trt_int8_calibration_table_name;  // TensorRT INT8 calibration table name.
-  int trt_int8_use_native_calibration_table;    // use native TensorRT generated calibration table. Default 0 = false, nonzero = true
-  int trt_dla_enable;                           // enable DLA. Default 0 = false, nonzero = true
-  int trt_dla_core;                             // DLA core number. Default 0
-  int trt_dump_subgraphs;                       // dump TRT subgraph. Default 0 = false, nonzero = true
-  int trt_engine_cache_enable;                  // enable engine caching. Default 0 = false, nonzero = true
-  const char* trt_engine_cache_path;            // specify engine cache path
-  int trt_engine_decryption_enable;             // enable engine decryption. Default 0 = false, nonzero = true
-  const char* trt_engine_decryption_lib_path;   // specify engine decryption library path
-  int trt_force_sequential_engine_build;        // force building TensorRT engine sequentially. Default 0 = false, nonzero = true
-  // This is the legacy struct and don't add new fields here.
-  // For new field that can be represented by string, please add it in include/onnxruntime/core/providers/tensorrt/tensorrt_provider_options.h
-  // For non-string field, need to create a new separate api to handle it.
-} OrtTensorRTProviderOptions;
-
-/** \brief MIGraphX Provider Options
- *
- * \see OrtApi::SessionOptionsAppendExecutionProvider_MIGraphX
- */
-typedef struct OrtMIGraphXProviderOptions {
-  int device_id;             // hip device id.
-  int migraphx_fp16_enable;  // enable MIGraphX FP16 precision. Default 0 = false, nonzero = true
-  int migraphx_int8_enable;  // enable MIGraphX INT8 precision. Default 0 = false, nonzero = true
-} OrtMIGraphXProviderOptions;
-
-/** \brief OpenVINO Provider Options
- *
- * \see OrtApi::SessionOptionsAppendExecutionProvider_OpenVINO
- */
-typedef struct OrtOpenVINOProviderOptions {
-#ifdef __cplusplus
-  OrtOpenVINOProviderOptions() : device_type{}, enable_vpu_fast_compile{}, device_id{},
-                                 num_of_threads{}, cache_dir{},
-                                 context{}, enable_opencl_throttling{}, enable_dynamic_shapes{} {}
-#endif
-  /** \brief Device type string
-   *
-   * Valid settings are one of: "CPU_FP32", "CPU_FP16", "GPU_FP32", "GPU_FP16", "MYRIAD_FP16", "VAD-M_FP16" or "VAD-F_FP32"
-   */
-  const char* device_type;
-  unsigned char enable_vpu_fast_compile;  ///< 0 = disabled, nonzero = enabled
-  const char* device_id;
-  size_t num_of_threads;               ///< 0 = Use default number of threads
-  const char* cache_dir;          // path is set to empty by default
-  void* context;
-  unsigned char enable_opencl_throttling;  ///< 0 = disabled, nonzero = enabled
-  unsigned char enable_dynamic_shapes;     ///< 0 = disabled, nonzero = enabled
-} OrtOpenVINOProviderOptions;
-
-struct OrtApi;
-typedef struct OrtApi OrtApi;
-
-struct OrtTrainingApi;
-typedef struct OrtTrainingApi OrtTrainingApi;
-
-/** \brief The helper interface to get the right version of OrtApi
- *
- * Get a pointer to this structure through ::OrtGetApiBase
- */
-struct OrtApiBase {
-  /** \brief Get a pointer to the requested version of the ::OrtApi
-   *
-   * \param[in] version Must be ::ORT_API_VERSION
-   * \return The ::OrtApi for the version requested, nullptr will be returned if this version is unsupported, for example when using a runtime
-   *   older than the version created with this header file.
-   */
-  const OrtApi*(ORT_API_CALL* GetApi)(uint32_t version)NO_EXCEPTION;
-  const char*(ORT_API_CALL* GetVersionString)(void)NO_EXCEPTION;  ///< Returns a null terminated string of the version of the Onnxruntime library (eg: "1.8.1")
-};
-typedef struct OrtApiBase OrtApiBase;
-
-/** \brief The Onnxruntime library's entry point to access the C API
- *
- * Call this to get the a pointer to an ::OrtApiBase
- */
-ORT_EXPORT const OrtApiBase* ORT_API_CALL OrtGetApiBase(void) NO_EXCEPTION;
-
-/** \brief Thread work loop function
- *
- * Onnxruntime will provide the working loop on custom thread creation
- * Argument is an onnxruntime built-in type which will be provided when thread pool calls OrtCustomCreateThreadFn
- */
-typedef void (*OrtThreadWorkerFn)(void* ort_worker_fn_param);
-
-typedef const struct OrtCustomHandleType {
-  char __place_holder;
-}* OrtCustomThreadHandle;
-
-/** \brief Ort custom thread creation function
- *
- * The function should return a thread handle to be used in onnxruntime thread pools
- * Onnxruntime will throw exception on return value of nullptr or 0, indicating that the function failed to create a thread
- */
-typedef OrtCustomThreadHandle (*OrtCustomCreateThreadFn)(void* ort_custom_thread_creation_options, OrtThreadWorkerFn ort_thread_worker_fn, void* ort_worker_fn_param);
-
-/** \brief Custom thread join function
- *
- * Onnxruntime thread pool destructor will call the function to join a custom thread.
- * Argument ort_custom_thread_handle is the value returned by OrtCustomCreateThreadFn
- */
-typedef void (*OrtCustomJoinThreadFn)(OrtCustomThreadHandle ort_custom_thread_handle);
-
-typedef OrtStatus*(ORT_API_CALL* RegisterCustomOpsFn)(OrtSessionOptions* options, const OrtApiBase* api);
-
-/** \brief The C API
- *
- * All C API functions are defined inside this structure as pointers to functions.
- * Call OrtApiBase::GetApi to get a pointer to it
- *
- * \nosubgrouping
- */
-struct OrtApi {
-  /// \name OrtStatus
-  /// @{
-
-  /**
-   * \brief Create an OrtStatus from a null terminated string
-   *
-   * \param[in] code
-   * \param[in] msg A null-terminated string. Its contents will be copied.
-   * \return A new OrtStatus object, must be destroyed with OrtApi::ReleaseStatus
-   */
-  OrtStatus*(ORT_API_CALL* CreateStatus)(OrtErrorCode code, _In_ const char* msg)NO_EXCEPTION ORT_ALL_ARGS_NONNULL;
-
-  /** \brief Get OrtErrorCode from OrtStatus
-   *
-   * \param[in] status
-   * \return OrtErrorCode that \p status was created with
-   */
-  OrtErrorCode(ORT_API_CALL* GetErrorCode)(_In_ const OrtStatus* status) NO_EXCEPTION ORT_ALL_ARGS_NONNULL;
-
-  /** \brief Get error string from OrtStatus
-   *
-   * \param[in] status
-   * \return The error message inside the `status`. Do not free the returned value.
-   */
-  const char*(ORT_API_CALL* GetErrorMessage)(_In_ const OrtStatus* status)NO_EXCEPTION ORT_ALL_ARGS_NONNULL;
-
-  /// @}
-  /// \name OrtEnv
-  /// @{
-
-  /** \brief Create an OrtEnv
-   *
-   * \param[in] log_severity_level The log severity level.
-   * \param[in] logid The log identifier.
-   * \param[out] out Returned newly created OrtEnv. Must be freed with OrtApi::ReleaseEnv
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateEnv, OrtLoggingLevel log_severity_level, _In_ const char* logid, _Outptr_ OrtEnv** out);
-
-  /** \brief Create an OrtEnv
-   *
-   * \param[in] logging_function A pointer to a logging function.
-   * \param[in] logger_param A pointer to arbitrary data passed as the ::OrtLoggingFunction `param` parameter to
-   *                         `logging_function`.
-   * \param[in] log_severity_level The log severity level.
-   * \param[in] logid The log identifier.
-   * \param[out] out Returned newly created OrtEnv. Must be freed with OrtApi::ReleaseEnv
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateEnvWithCustomLogger, OrtLoggingFunction logging_function, _In_opt_ void* logger_param,
-                  OrtLoggingLevel log_severity_level, _In_ const char* logid, _Outptr_ OrtEnv** out);
-
-  /** \brief Enable Telemetry
-   *
-   * \note Telemetry events are on by default since they are lightweight
-   * \param[in] env
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(EnableTelemetryEvents, _In_ const OrtEnv* env);
-  /** \brief Disable Telemetry
-   *
-   * \see OrtApi::EnableTelemetryEvents
-   * \param[in] env
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(DisableTelemetryEvents, _In_ const OrtEnv* env);
-
-  /// @}
-  /// \name OrtSession
-  /// @{
-
-  /** \brief Create an OrtSession from a model file
-   *
-   * \param[in] env
-   * \param[in] model_path
-   * \param[in] options
-   * \param[out] out Returned newly created OrtSession. Must be freed with OrtApi::ReleaseSession
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  // TODO: document the path separator convention? '/' vs '\'
-  // TODO: should specify the access characteristics of model_path. Is this read only during the
-  // execution of CreateSession, or does the OrtSession retain a handle to the file/directory
-  // and continue to access throughout the OrtSession lifetime?
-  //  What sort of access is needed to model_path : read or read/write?
-  ORT_API2_STATUS(CreateSession, _In_ const OrtEnv* env, _In_ const ORTCHAR_T* model_path,
-                  _In_ const OrtSessionOptions* options, _Outptr_ OrtSession** out);
-
-  /** \brief Create an OrtSession from memory
-   *
-   * \param[in] env
-   * \param[in] model_data
-   * \param[in] model_data_length
-   * \param[in] options
-   * \param[out] out Returned newly created OrtSession. Must be freed with OrtApi::ReleaseSession
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateSessionFromArray, _In_ const OrtEnv* env, _In_ const void* model_data, size_t model_data_length,
-                  _In_ const OrtSessionOptions* options, _Outptr_ OrtSession** out);
-
-  /** \brief Run the model in an ::OrtSession
-   *
-   * Will not return until the model run has completed. Multiple threads might be used to run the model based on
-   * the options in the ::OrtSession and settings used when creating the ::OrtEnv
-   *
-   * \param[in] session
-   * \param[in] run_options If nullptr, will use a default ::OrtRunOptions
-   * \param[in] input_names Array of null terminated UTF8 encoded strings of the input names
-   * \param[in] inputs Array of ::OrtValue%s of the input values
-   * \param[in] input_len Number of elements in the input_names and inputs arrays
-   * \param[in] output_names Array of null terminated UTF8 encoded strings of the output names
-   * \param[in] output_names_len Number of elements in the output_names and outputs array
-   * \param[out] outputs Array of ::OrtValue%s that the outputs are stored in. This can also be
-   *     an array of nullptr values, in this case ::OrtValue objects will be allocated and pointers
-   *     to them will be set into the `outputs` array.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(Run, _Inout_ OrtSession* session, _In_opt_ const OrtRunOptions* run_options,
-                  _In_reads_(input_len) const char* const* input_names,
-                  _In_reads_(input_len) const OrtValue* const* inputs, size_t input_len,
-                  _In_reads_(output_names_len) const char* const* output_names, size_t output_names_len,
-                  _Inout_updates_all_(output_names_len) OrtValue** outputs);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Create an ::OrtSessionOptions object
-   *
-   * To use additional providers, you must build ORT with the extra providers enabled. Then call one of these
-   * functions to enable them in the session:<br>
-   *   OrtSessionOptionsAppendExecutionProvider_CPU<br>
-   *   OrtSessionOptionsAppendExecutionProvider_CUDA<br>
-   *   OrtSessionOptionsAppendExecutionProvider_(remaining providers...)<br>
-   * The order they are called indicates the preference order as well. In other words call this method
-   * on your most preferred execution provider first followed by the less preferred ones.
-   * If none are called Ort will use its internal CPU execution provider.
-   *
-   * \param[out] options The newly created OrtSessionOptions. Must be freed with OrtApi::ReleaseSessionOptions
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateSessionOptions, _Outptr_ OrtSessionOptions** options);
-
-  /** \brief Set filepath to save optimized model after graph level transformations
-   *
-   * \param[in] options
-   * \param[in] optimized_model_filepath
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetOptimizedModelFilePath, _Inout_ OrtSessionOptions* options,
-                  _In_ const ORTCHAR_T* optimized_model_filepath);
-
-  /** \brief Create a copy of an existing ::OrtSessionOptions
-   *
-   * \param[in] in_options OrtSessionOptions to copy
-   * \param[out] out_options Returned newly created ::OrtSessionOptions. Must be freed with OrtApi::ReleaseSessionOptions
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CloneSessionOptions, _In_ const OrtSessionOptions* in_options,
-                  _Outptr_ OrtSessionOptions** out_options);
-
-  /** \brief Set execution mode
-   *
-   * Controls whether you want to execute operators in your graph sequentially or in parallel. Usually when the model
-   *  has many branches, setting this option to ExecutionMode.ORT_PARALLEL will give you better performance.
-   *  See [docs/ONNX_Runtime_Perf_Tuning.md] for more details.
-   *
-   * \param[in] options
-   * \param[in] execution_mode
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetSessionExecutionMode, _Inout_ OrtSessionOptions* options, ExecutionMode execution_mode);
-
-  /** \brief Enable profiling for a session
-   *
-   * \param[in] options
-   * \param[in] profile_file_prefix
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(EnableProfiling, _Inout_ OrtSessionOptions* options, _In_ const ORTCHAR_T* profile_file_prefix);
-
-  /** \brief Disable profiling for a session
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(DisableProfiling, _Inout_ OrtSessionOptions* options);
-
-  /** \brief Enable the memory pattern optimization
-   *
-   * The idea is if the input shapes are the same, we could trace the internal memory allocation
-   * and generate a memory pattern for future request. So next time we could just do one allocation
-   * with a big chunk for all the internal memory allocation.
-   * \note Memory pattern optimization is only available when Sequential Execution mode is enabled (see OrtApi::SetSessionExecutionMode)
-   *
-   * \see OrtApi::DisableMemPattern
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(EnableMemPattern, _Inout_ OrtSessionOptions* options);
-
-  /** \brief Disable the memory pattern optimization
-   *
-   * \see OrtApi::EnableMemPattern
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(DisableMemPattern, _Inout_ OrtSessionOptions* options);
-
-  /** \brief Enable the memory arena on CPU
-   *
-   * Arena may pre-allocate memory for future usage.
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(EnableCpuMemArena, _Inout_ OrtSessionOptions* options);
-
-  /** \brief Disable the memory arena on CPU
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(DisableCpuMemArena, _Inout_ OrtSessionOptions* options);
-
-  /** \brief Set session log id
-   *
-   * \param[in] options
-   * \param[in] logid The log identifier.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetSessionLogId, _Inout_ OrtSessionOptions* options, const char* logid);
-
-  /** \brief Set session log verbosity level
-   *
-   * Applies to session load, initialization, etc
-   *
-   * \param[in] options
-   * \param[in] session_log_verbosity_level \snippet{doc} snippets.dox Log Verbosity Level
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetSessionLogVerbosityLevel, _Inout_ OrtSessionOptions* options, int session_log_verbosity_level);
-
-  /** \brief Set session log severity level
-   *
-   * \param[in] options
-   * \param[in] session_log_severity_level The log severity level (refer to ::OrtLoggingLevel for possible values).
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetSessionLogSeverityLevel, _Inout_ OrtSessionOptions* options, int session_log_severity_level);
-
-  /** \brief Set the optimization level to apply when loading a graph
-   *
-   * Please see https://www.onnxruntime.ai/docs/resources/graph-optimizations.html for an in-depth explanation
-   * \param[in,out] options The session options object
-   * \param[in] graph_optimization_level The optimization level
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetSessionGraphOptimizationLevel, _Inout_ OrtSessionOptions* options,
-                  GraphOptimizationLevel graph_optimization_level);
-
-  /** \brief Sets the number of threads used to parallelize the execution within nodes
-   *
-   * When running a single node operation, ex. add, this sets the maximum number of threads to use.
-   *
-   * \note If built with OpenMP, this has no effect on the number of threads used. In this case
-   *       use the OpenMP env variables to configure the number of intra op num threads.
-   *
-   * \param[in] options
-   * \param[in] intra_op_num_threads Number of threads to use<br>
-   *   A value of 0 will use the default number of threads<br>
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetIntraOpNumThreads, _Inout_ OrtSessionOptions* options, int intra_op_num_threads);
-
-  /** \brief Sets the number of threads used to parallelize the execution of the graph
-   *
-   * If nodes can be run in parallel, this sets the maximum number of threads to use to run them in parallel.
-   *
-   * \note If sequential execution is enabled this value is ignored, it acts as if it was set to 1.
-   *
-   * \param[in] options
-   * \param[in] inter_op_num_threads Number of threads to use<br>
-   *   A value of 0 will use the default number of threads<br>
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetInterOpNumThreads, _Inout_ OrtSessionOptions* options, int inter_op_num_threads);
-
-  /// @}
-  /// \name OrtCustomOpDomain
-  /// @{
-
-  /** \brief Create a custom op domain
-   *
-   * \param[in] domain
-   * \param[out] out Newly created domain. Must be freed with OrtApi::ReleaseCustomOpDomain
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateCustomOpDomain, _In_ const char* domain, _Outptr_ OrtCustomOpDomain** out);
-
-  /** \brief Add a custom op to a custom op domain
-   *
-   * \note The OrtCustomOp* pointer must remain valid until the ::OrtCustomOpDomain using it is released
-   *
-   * \param[in] custom_op_domain
-   * \param[in] op
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CustomOpDomain_Add, _Inout_ OrtCustomOpDomain* custom_op_domain, _In_ const OrtCustomOp* op);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Add custom op domain to a session options
-   *
-   * \note The OrtCustomOpDomain* must not be deleted until all sessions using it are released
-   *
-   * \param[in] options
-   * \param[in] custom_op_domain
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(AddCustomOpDomain, _Inout_ OrtSessionOptions* options, _In_ OrtCustomOpDomain* custom_op_domain);
-
-  /** \deprecated Use OrtApi::RegisterCustomOpsLibrary_V2.
-   *
-   * Registers custom ops from a shared library.
-   *
-   * Loads a shared library (dll on windows, so on linux, etc) named 'library_path' and looks for this entry point:
-   *		OrtStatus* RegisterCustomOps(OrtSessionOptions * options, const OrtApiBase* api);
-   * It then passes in the provided session options to this function along with the api base.
-   * The handle to the loaded library is returned in library_handle. It can be freed by the caller after all sessions using the passed in
-   * session options are destroyed, or if an error occurs and it is non null.
-   *
-   * \param[in] options
-   * \param[in] library_path
-   * \param[out] library_handle OS specific handle to the loaded library (Use FreeLibrary on Windows, dlclose on Linux, etc.. to unload)
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(RegisterCustomOpsLibrary, _Inout_ OrtSessionOptions* options, _In_ const char* library_path, _Outptr_ void** library_handle);
-
-  /// @}
-  /// \name OrtSession
-  /// @{
-
-  /** \brief Get input count for a session
-   *
-   * This number must also match the number of inputs passed to OrtApi::Run
-   *
-   * \see OrtApi::SessionGetInputTypeInfo, OrtApi::SessionGetInputName, OrtApi::Session
-   *
-   * \param[in] session
-   * \param[out] out Number of inputs
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetInputCount, _In_ const OrtSession* session, _Out_ size_t* out);
-
-  /** \brief Get output count for a session
-   *
-   * This number must also match the number of outputs returned by OrtApi::Run
-   *
-   * \see OrtApi::SessionGetOutputTypeInfo, OrtApi::SessionGetOutputName, OrtApi::Session
-   *
-   * \param[in] session
-   * \param[out] out Number of outputs
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetOutputCount, _In_ const OrtSession* session, _Out_ size_t* out);
-
-  /** \brief Get overridable initializer count
-   *
-   * \see OrtApi::SessionGetOverridableInitializerTypeInfo, OrtApi::SessionGetOverridableInitializerName
-   *
-   * \param[in] session
-   * \param[in] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetOverridableInitializerCount, _In_ const OrtSession* session, _Out_ size_t* out);
-
-  /** \brief Get input type information
-   *
-   * \param[in] session
-   * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetInputCount returns (exclusive)
-   * \param[out] type_info Must be freed with OrtApi::ReleaseTypeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetInputTypeInfo, _In_ const OrtSession* session, size_t index, _Outptr_ OrtTypeInfo** type_info);
-
-  /** \brief Get output type information
-   *
-   * \param[in] session
-   * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetOutputCount returns (exclusive)
-   * \param[out] type_info Must be freed with OrtApi::ReleaseTypeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetOutputTypeInfo, _In_ const OrtSession* session, size_t index, _Outptr_ OrtTypeInfo** type_info);
-
-  /** \brief Get overridable initializer type information
-   *
-   * \param[in] session
-   * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetOverridableInitializerCount returns (exclusive)
-   * \param[out] type_info Must be freed with OrtApi::ReleaseTypeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetOverridableInitializerTypeInfo, _In_ const OrtSession* session, size_t index, _Outptr_ OrtTypeInfo** type_info);
-
-  /** \brief Get input name
-   *
-   * \param[in] session
-   * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetInputCount returns (exclusive)
-   * \param[in] allocator
-   * \param[out] value Set to a null terminated UTF-8 encoded string allocated using `allocator`. Must be freed using `allocator`.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetInputName, _In_ const OrtSession* session, size_t index, _Inout_ OrtAllocator* allocator, _Outptr_ char** value);
-
-  /** \brief Get output name
-   *
-   * \param[in] session
-   * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetOutputCount returns (exclusive)
-   * \param[in] allocator
-   * \param[out] value Set to a null terminated UTF-8 encoded string allocated using `allocator`. Must be freed using `allocator`.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetOutputName, _In_ const OrtSession* session, size_t index, _Inout_ OrtAllocator* allocator, _Outptr_ char** value);
-
-  /** \brief Get overridable initializer name
-   *
-   * \param[in] session
-   * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetOverridableInitializerCount returns (exclusive)
-   * \param[in] allocator
-   * \param[out] value Set to a null terminated UTF-8 encoded string allocated using `allocator`. Must be freed using `allocator`.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetOverridableInitializerName, _In_ const OrtSession* session, size_t index,
-                  _Inout_ OrtAllocator* allocator, _Outptr_ char** value);
-
-  /// @}
-  /// \name OrtRunOptions
-  /// @{
-
-  /** \brief Create an OrtRunOptions
-   *
-   * \param[out] out Returned newly created ::OrtRunOptions. Must be freed with OrtApi::ReleaseRunOptions
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateRunOptions, _Outptr_ OrtRunOptions** out);
-
-  /** \brief Set per-run log verbosity level
-   *
-   * \see OrtApi::RunOptionsGetRunLogVerbosityLevel
-   *
-   * \param[in] options
-   * \param[in] log_verbosity_level \snippet{doc} snippets.dox Log Verbosity Level
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(RunOptionsSetRunLogVerbosityLevel, _Inout_ OrtRunOptions* options, int log_verbosity_level);
-
-  /** \brief Set per-run log severity level
-   *
-   * \see OrtApi::RunOptionsGetRunLogSeverityLevel
-   *
-   * \param[in] options
-   * \param[in] log_severity_level The log severity level (refer to ::OrtLoggingLevel for possible values).
-   */
-  ORT_API2_STATUS(RunOptionsSetRunLogSeverityLevel, _Inout_ OrtRunOptions* options, int log_severity_level);
-
-  /** \brief Set per-run tag
-   *
-   * This is used in a per-run log identifier.
-   *
-   * \see OrtApi::RunOptionsGetRunTag
-   *
-   * \param[in] options
-   * \param[in] run_tag The run tag.
-   */
-  ORT_API2_STATUS(RunOptionsSetRunTag, _Inout_ OrtRunOptions* options, _In_ const char* run_tag);
-
-  /** \brief Get per-run log verbosity level
-   *
-   * \see OrtApi::RunOptionsSetRunLogVerbosityLevel
-   *
-   * \param[in] options
-   * \param[out] log_verbosity_level \snippet{doc} snippets.dox Log Verbosity Level
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(RunOptionsGetRunLogVerbosityLevel, _In_ const OrtRunOptions* options,
-                  _Out_ int* log_verbosity_level);
-
-  /** \brief Get per-run log severity level
-   *
-   * \see OrtApi::RunOptionsSetRunLogSeverityLevel
-   *
-   * \param[in] options
-   * \param[out] log_severity_level The log severity level (refer to ::OrtLoggingLevel for possible values).
-   */
-  ORT_API2_STATUS(RunOptionsGetRunLogSeverityLevel, _In_ const OrtRunOptions* options, _Out_ int* log_severity_level);
-
-  /** \brief Get per-run tag
-   *
-   * This is used in a per-run log identifier.
-   *
-   * \see OrtApi::RunOptionsSetRunTag
-   *
-   * \param[in] options
-   * \param[out] run_tag The run tag.
-   *                     Do not free this value, it is owned by `options`. It will be invalidated if the run tag
-   *                     changes (i.e., with OrtApi::RunOptionsSetRunTag) or `options` is freed.
-   */
-  ORT_API2_STATUS(RunOptionsGetRunTag, _In_ const OrtRunOptions* options, _Out_ const char** run_tag);
-
-  /** \brief Set terminate flag
-   *
-   * If a currently executing session needs to be force terminated, this can be called from another thread to force it to fail with an error.
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(RunOptionsSetTerminate, _Inout_ OrtRunOptions* options);
-
-  /** \brief Clears the terminate flag
-   *
-   * Used so the OrtRunOptions instance can be used in a new OrtApi::Run call without it instantly terminating
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(RunOptionsUnsetTerminate, _Inout_ OrtRunOptions* options);
-
-  /// @}
-  /// \name OrtValue
-  /// @{
-
-  /** \brief Create a tensor
-   *
-   * Create a tensor using a supplied ::OrtAllocator
-   *
-   * \param[in] allocator
-   * \param[in] shape Pointer to the tensor shape dimensions.
-   * \param[in] shape_len The number of tensor shape dimensions.
-   * \param[in] type
-   * \param[out] out Returns newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateTensorAsOrtValue, _Inout_ OrtAllocator* allocator, _In_ const int64_t* shape, size_t shape_len,
-                  ONNXTensorElementDataType type, _Outptr_ OrtValue** out);
-
-  /** \brief Create a tensor backed by a user supplied buffer
-   *
-   * Create a tensor with user's buffer. You can fill the buffer either before calling this function or after.
-   * p_data is owned by caller. ReleaseValue won't release p_data.
-   *
-   * \param[in] info Memory description of where the p_data buffer resides (CPU vs GPU etc).
-   * \param[in] p_data Pointer to the data buffer.
-   * \param[in] p_data_len The number of bytes in the data buffer.
-   * \param[in] shape Pointer to the tensor shape dimensions.
-   * \param[in] shape_len The number of tensor shape dimensions.
-   * \param[in] type The data type.
-   * \param[out] out Returns newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateTensorWithDataAsOrtValue, _In_ const OrtMemoryInfo* info, _Inout_ void* p_data,
-                  size_t p_data_len, _In_ const int64_t* shape, size_t shape_len, ONNXTensorElementDataType type,
-                  _Outptr_ OrtValue** out);
-
-  /** \brief Return if an ::OrtValue is a tensor type
-   *
-   * \param[in] value A tensor type (string tensors are not supported)
-   * \param[out] out Set to 1 iff ::OrtValue is a tensor, 0 otherwise
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(IsTensor, _In_ const OrtValue* value, _Out_ int* out);
-
-  /** \brief Get a pointer to the raw data inside a tensor
-   *
-   * Used to read/write/modify the internal tensor data directly.
-   * \note The returned pointer is valid until the \p value is destroyed.
-   *
-   * \param[in] value A tensor type (string tensors are not supported)
-   * \param[out] out Filled in with a pointer to the internal storage
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetTensorMutableData, _In_ OrtValue* value, _Outptr_ void** out);
-
-  /** \brief Set all strings at once in a string tensor
-   *
-   * \param[in,out] value A tensor of type ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING
-   * \param[in] s An array of strings. Each string in this array must be null terminated.
-   * \param[in] s_len Count of strings in s (Must match the size of \p value's tensor shape)
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(FillStringTensor, _Inout_ OrtValue* value, _In_ const char* const* s, size_t s_len);
-
-  /** \brief Get total byte length for all strings in a string tensor
-   *
-   * Typically used with OrtApi::GetStringTensorContent
-   *
-   * \param[in] value A tensor of type ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING
-   * \param[out] len Total byte length of all strings (does not include trailing nulls)
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetStringTensorDataLength, _In_ const OrtValue* value, _Out_ size_t* len);
-
-  /** \brief Get all strings from a string tensor
-   *
-   * An example of the results:<br>
-   * Given \p value is a string tensor with the strings { "This" "is" "a" "test" }<br>
-   * \p s must have a size of 11 bytes<br>
-   * \p offsets must have 4 elements<br>
-   * After the call, these values will be filled in:<br>
-   * \p s will contain "Thisisatest"<br>
-   * \p offsets will contain { 0, 4, 6, 7 }<br>
-   * The length of the last string is just s_len - offsets[last]
-   *
-   * \param[in] value A tensor of type ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING
-   * \param[in] s Buffer to sequentially write all tensor strings to. Each string is NOT null-terminated.
-   * \param[in] s_len Number of bytes of buffer pointed to by \p s (Get it from OrtApi::GetStringTensorDataLength)
-   * \param[out] offsets Array of start offsets into the strings written to \p s
-   * \param[in] offsets_len Number of elements in offsets
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetStringTensorContent, _In_ const OrtValue* value, _Out_writes_bytes_all_(s_len) void* s,
-                  size_t s_len, _Out_writes_all_(offsets_len) size_t* offsets, size_t offsets_len);
-
-  /// @}
-  /// \name OrtTypeInfo
-  /// @{
-
-  /** \brief Get ::OrtTensorTypeAndShapeInfo from an ::OrtTypeInfo
-   *
-   * \param[in] type_info
-   * \param[out] out Do not free this value, it will be valid until type_info is freed.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CastTypeInfoToTensorInfo, _In_ const OrtTypeInfo* type_info,
-                  _Outptr_result_maybenull_ const OrtTensorTypeAndShapeInfo** out);
-
-  /** \brief Get ::ONNXType from ::OrtTypeInfo
-   *
-   * \param[in] type_info
-   * \param[out] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetOnnxTypeFromTypeInfo, _In_ const OrtTypeInfo* type_info, _Out_ enum ONNXType* out);
-
-  /// @}
-  /// \name OrtTensorTypeAndShapeInfo
-  /// @{
-
-  /** \brief Create an ::OrtTensorTypeAndShapeInfo object
-   *
-   * \param[out] out Returns newly created ::OrtTensorTypeAndShapeInfo. Must be freed with OrtApi::ReleaseTensorTypeAndShapeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateTensorTypeAndShapeInfo, _Outptr_ OrtTensorTypeAndShapeInfo** out);
-
-  /** \brief Set element type in ::OrtTensorTypeAndShapeInfo
-   *
-   * \param[in] info
-   * \param[in] type
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetTensorElementType, _Inout_ OrtTensorTypeAndShapeInfo* info, enum ONNXTensorElementDataType type);
-
-  /** \brief Set shape information in ::OrtTensorTypeAndShapeInfo
-   *
-   * \param[in] info
-   * \param[in] dim_values Array with `dim_count` elements. Can contain negative values.
-   * \param[in] dim_count Number of elements in `dim_values`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetDimensions, OrtTensorTypeAndShapeInfo* info, _In_ const int64_t* dim_values, size_t dim_count);
-
-  /** \brief Get element type in ::OrtTensorTypeAndShapeInfo
-   *
-   * \see OrtApi::SetTensorElementType
-   *
-   * \param[in] info
-   * \param[out] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetTensorElementType, _In_ const OrtTensorTypeAndShapeInfo* info,
-                  _Out_ enum ONNXTensorElementDataType* out);
-
-  /** \brief Get dimension count in ::OrtTensorTypeAndShapeInfo
-   *
-   * \see OrtApi::GetDimensions
-   *
-   * \param[in] info
-   * \param[out] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetDimensionsCount, _In_ const OrtTensorTypeAndShapeInfo* info, _Out_ size_t* out);
-
-  /** \brief Get dimensions in ::OrtTensorTypeAndShapeInfo
-   *
-   * \param[in] info
-   * \param[out] dim_values Array with `dim_values_length` elements. On return, filled with the dimensions stored in the ::OrtTensorTypeAndShapeInfo
-   * \param[in] dim_values_length Number of elements in `dim_values`. Use OrtApi::GetDimensionsCount to get this value
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetDimensions, _In_ const OrtTensorTypeAndShapeInfo* info, _Out_ int64_t* dim_values,
-                  size_t dim_values_length);
-
-  /** \brief Get symbolic dimension names in ::OrtTensorTypeAndShapeInfo
-   *
-   * \param[in] info
-   * \param[in] dim_params Array with `dim_params_length` elements. On return filled with pointers to null terminated strings of the dimension names
-   * \param[in] dim_params_length Number of elements in `dim_params`. Use OrtApi::GetDimensionsCount to get this value
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetSymbolicDimensions, _In_ const OrtTensorTypeAndShapeInfo* info,
-                  _Out_writes_all_(dim_params_length) const char* dim_params[], size_t dim_params_length);
-
-  /** \brief Get total number of elements in a tensor shape from an ::OrtTensorTypeAndShapeInfo
-   *
-   * Return the number of elements specified by the tensor shape (all dimensions multiplied by each other).
-   * For 0 dimensions, 1 is returned. If any dimension is less than 0, the result is always -1.
-   *
-   * Examples:<br>
-   * [] = 1<br>
-   * [1,3,4] = 12<br>
-   * [2,0,4] = 0<br>
-   * [-1,3,4] = -1<br>
-   *
-   * \param[in] info
-   * \param[out] out Number of elements
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetTensorShapeElementCount, _In_ const OrtTensorTypeAndShapeInfo* info, _Out_ size_t* out);
-
-  /// @}
-  /// \name OrtValue
-  /// @{
-
-  /** \brief Get type and shape information from a tensor ::OrtValue
-   *
-   * \param[in] value Must be a tensor (not a map/sequence/etc) or will return failure
-   * \param[out] out Newly created ::OrtTensorTypeAndShapeInfo. Must be freed with OrtApi::ReleaseTensorTypeAndShapeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetTensorTypeAndShape, _In_ const OrtValue* value, _Outptr_ OrtTensorTypeAndShapeInfo** out);
-
-  /** \brief Get type information of an OrtValue
-   *
-   * \param[in] value
-   * \param[out] out Newly created ::OrtTypeInfo. Must be freed with OrtApi::ReleaseTypeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetTypeInfo, _In_ const OrtValue* value, _Outptr_result_maybenull_ OrtTypeInfo** out);
-
-  /** \brief Get ONNXType of an ::OrtValue
-   *
-   * \param[in] value
-   * \param[out] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetValueType, _In_ const OrtValue* value, _Out_ enum ONNXType* out);
-
-  /// @}
-  /// \name OrtMemoryInfo
-  /// @{
-
-  /** \brief Create an ::OrtMemoryInfo
-   *
-   * \param[in] name
-   * \param[in] type
-   * \param[in] id
-   * \param[in] mem_type
-   * \param[out] out Newly created ::OrtMemoryInfo. Must be freed with OrtAPi::ReleaseMemoryInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateMemoryInfo, _In_ const char* name, enum OrtAllocatorType type, int id,
-                  enum OrtMemType mem_type, _Outptr_ OrtMemoryInfo** out);
-
-  /** \brief Create an ::OrtMemoryInfo for CPU memory
-   *
-   * Special case version of OrtApi::CreateMemoryInfo for CPU based memory. Same as using OrtApi::CreateMemoryInfo with name = "Cpu" and id = 0.
-   *
-   * \param[in] type
-   * \param[in] mem_type
-   * \param[out] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateCpuMemoryInfo, enum OrtAllocatorType type, enum OrtMemType mem_type,
-                  _Outptr_ OrtMemoryInfo** out);
-
-  /** \brief Compare ::OrtMemoryInfo objects for equality
-   *
-   * Compares all settings of each ::OrtMemoryInfo for equality
-   *
-   * \param[in] info1
-   * \param[in] info2
-   * \param[out] out Set to 0 if equal, -1 if not equal
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CompareMemoryInfo, _In_ const OrtMemoryInfo* info1, _In_ const OrtMemoryInfo* info2, _Out_ int* out);
-
-  /** \brief Get name from ::OrtMemoryInfo
-   *
-   * \param[in] ptr
-   * \param[out] out Writes null terminated string to this pointer. Do NOT free the returned pointer. It is valid for the lifetime of the ::OrtMemoryInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(MemoryInfoGetName, _In_ const OrtMemoryInfo* ptr, _Out_ const char** out);
-
-  /** \brief Get the id from ::OrtMemoryInfo
-   */
-  ORT_API2_STATUS(MemoryInfoGetId, _In_ const OrtMemoryInfo* ptr, _Out_ int* out);
-
-  /** \brief Get the ::OrtMemType from ::OrtMemoryInfo
-   */
-  ORT_API2_STATUS(MemoryInfoGetMemType, _In_ const OrtMemoryInfo* ptr, _Out_ OrtMemType* out);
-
-  /** \brief Get the ::OrtAllocatorType from ::OrtMemoryInfo
-   */
-  ORT_API2_STATUS(MemoryInfoGetType, _In_ const OrtMemoryInfo* ptr, _Out_ OrtAllocatorType* out);
-
-  /// @}
-  /// \name OrtAllocator
-  /// @{
-
-  /// \brief Calls OrtAllocator::Alloc function
-  ORT_API2_STATUS(AllocatorAlloc, _Inout_ OrtAllocator* ort_allocator, size_t size, _Outptr_ void** out);
-  /// \brief Calls OrtAllocator::Free function
-  ORT_API2_STATUS(AllocatorFree, _Inout_ OrtAllocator* ort_allocator, void* p);
-  /// \brief Calls OrtAllocator::Info function
-  ORT_API2_STATUS(AllocatorGetInfo, _In_ const OrtAllocator* ort_allocator, _Outptr_ const struct OrtMemoryInfo** out);
-
-  /** \brief Get the default allocator
-   *
-   * The default allocator is a CPU based, non-arena. Always returns the same pointer to the same default allocator.
-   *
-   * \param[out] out Returned value should NOT be freed
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetAllocatorWithDefaultOptions, _Outptr_ OrtAllocator** out);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Override session symbolic dimensions
-   *
-   * Override symbolic dimensions (by specific denotation strings) with actual values if known at session initialization time to enable
-   * optimizations that can take advantage of fixed values (such as memory planning, etc)
-   *
-   * \param[in] options
-   * \param[in] dim_denotation
-   * \param[in] dim_value
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(AddFreeDimensionOverride, _Inout_ OrtSessionOptions* options, _In_ const char* dim_denotation,
-                  _In_ int64_t dim_value);
-
-  /// @}
-  /// \name OrtValue
-  /// @{
-
-  /* Internal information (not seen in Doxygen)
-   *
-   * APIs to support non-tensor types - map and sequence.
-   * Currently only the following types are supported
-   * Note: the following types should be kept in sync with data_types.h
-   * Map types
-   * =========
-   * std::map<std::string, std::string>
-   * std::map<std::string, int64_t>
-   * std::map<std::string, float>
-   * std::map<std::string, double>
-   * std::map<int64_t, std::string>
-   * std::map<int64_t, int64_t>
-   * std::map<int64_t, float>
-   * std::map<int64_t, double>
-   *
-   * Sequence types
-   * ==============
-   * std::vector<std::string>
-   * std::vector<int64_t>
-   * std::vector<float>
-   * std::vector<double>
-   * std::vector<std::map<std::string, float>>
-   * std::vector<std::map<int64_t, float>
-   */
-
-  /** \brief Get non tensor data from an ::OrtValue
-   *
-   * If `value` is of type ONNX_TYPE_MAP, you need to retrieve the keys and values
-   * separately. Use index=0 to retrieve keys and index=1 to retrieve values.
-   * If `value` is of type ONNX_TYPE_SEQUENCE, use index to retrieve the index'th element
-   * of the sequence.
-   *
-   * \param[in] value
-   * \param[in] index See above for usage based on `value` type
-   * \param[in] allocator Allocator used to allocate ::OrtValue
-   * \param[out] out Created ::OrtValue that holds the element requested. Must be freed with OrtApi::ReleaseValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetValue, _In_ const OrtValue* value, int index, _Inout_ OrtAllocator* allocator,
-                  _Outptr_ OrtValue** out);
-
-  /** \brief Get non tensor value count from an ::OrtValue
-   *
-   * If `value` is of type ONNX_TYPE_MAP 2 will always be returned. For ONNX_TYPE_SEQUENCE
-   * the number of elements in the sequence will be returned
-   *
-   * \param[in] value
-   * \param[out] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetValueCount, _In_ const OrtValue* value, _Out_ size_t* out);
-
-  /** \brief Create a map or sequence ::OrtValue
-   *
-   * To construct a map (ONNX_TYPE_MAP), use num_values = 2 and `in` should be an array of 2 ::OrtValue%s
-   * representing keys and values.<br>
-   *
-   * To construct a sequence (ONNX_TYPE_SEQUENCE), use num_values = N where N is the number of the elements in the
-   * sequence. 'in' should be an array of N ::OrtValue%s.
-   *
-   * \param[in] in See above for details
-   * \param[in] num_values
-   * \param[in] value_type Must be either ONNX_TYPE_MAP or ONNX_TYPE_SEQUENCE
-   * \param[out] out Newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateValue, _In_reads_(num_values) const OrtValue* const* in, size_t num_values,
-                  enum ONNXType value_type, _Outptr_ OrtValue** out);
-
-  /** \brief Create an opaque (custom user defined type) ::OrtValue
-   *
-   * Constructs an ::OrtValue that contains a value of non-standard type created for
-   * experiments or while awaiting standardization. ::OrtValue in this case would contain
-   * an internal representation of the Opaque type. Opaque types are distinguished from
-   * each other by two strings 1) domain and 2) type name. The combination of the two
-   * must be unique, so the type representation is properly identified internally. The combination
-   * must be properly registered from within ORT at both compile/run time or by another API.
-   *
-   * To construct the ::OrtValue pass domain and type names, also a pointer to a data container
-   * the type of which must be known to both ORT and the client program. That data container may or may
-   * not match the internal representation of the Opaque type. The sizeof(data_container) is passed for
-   * verification purposes.
-   *
-   * \param[in] domain_name Null terminated string of the domain name
-   * \param[in] type_name Null terminated string of the type name
-   * \param[in] data_container User pointer Data to populate ::OrtValue
-   * \param[in] data_container_size Size in bytes of what `data_container` points to
-   * \param[out] out Newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateOpaqueValue, _In_z_ const char* domain_name, _In_z_ const char* type_name,
-                  _In_ const void* data_container, size_t data_container_size, _Outptr_ OrtValue** out);
-
-  /** \brief Get internal data from an opaque (custom user defined type) ::OrtValue
-   *
-   * Copies internal data from an opaque value into a user provided buffer
-   *
-   * \see OrtApi::CreateOpaqueValue
-   *
-   * \param[in] domain_name Null terminated string of the domain name
-   * \param[in] type_name Null terminated string of the type name
-   * \param[in] in The opaque ::OrtValue
-   * \param[out] data_container Buffer to copy data into
-   * \param[out] data_container_size Size in bytes of the buffer pointed to by data_container. Must match the size of the internal buffer.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetOpaqueValue, _In_ const char* domain_name, _In_ const char* type_name, _In_ const OrtValue* in,
-                  _Out_ void* data_container, size_t data_container_size);
-
-  /// @}
-  /// \name OrtKernelInfo
-  /// Custom operator APIs.
-  /// @{
-
-  /** \brief Get a float stored as an attribute in the graph node
-   *
-   * \param[in] info ::OrtKernelInfo instance
-   * \param[in] name Null terminated string of the name of the attribute
-   * \param[out] out Pointer to memory where the attribute will be stored
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(KernelInfoGetAttribute_float, _In_ const OrtKernelInfo* info, _In_ const char* name,
-                  _Out_ float* out);
-
-  /** \brief Fetch a 64-bit int stored as an attribute in the graph node
-   *
-   * \param[in] info ::OrtKernelInfo instance
-   * \param[in] name Null terminated string of the name of the attribute
-   * \param[out] out Pointer to memory where the attribute will be stored
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(KernelInfoGetAttribute_int64, _In_ const OrtKernelInfo* info, _In_ const char* name,
-                  _Out_ int64_t* out);
-
-  /** \brief Fetch a string stored as an attribute in the graph node
-   *
-   * If `out` is nullptr, the value of `size` is set to the true size of the string
-   * attribute, and a success status is returned.
-   *
-   * If the `size` parameter is greater than or equal to the actual string attribute's size,
-   * the value of `size` is set to the true size of the string attribute, the provided memory
-   * is filled with the attribute's contents, and a success status is returned.
-   *
-   * If the `size` parameter is less than the actual string attribute's size and `out`
-   * is not nullptr, the value of `size` is set to the true size of the string attribute
-   * and a failure status is returned.)
-   *
-   * \param[in] info ::OrtKernelInfo instance
-   * \param[in] name Null terminated string of the name of the attribute
-   * \param[out] out Pointer to memory where the attribute will be stored
-   * \param[in,out] size See above comments for details
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(KernelInfoGetAttribute_string, _In_ const OrtKernelInfo* info, _In_ const char* name, _Out_ char* out,
-                  _Inout_ size_t* size);
-
-  /// @}
-  /// \name OrtKernelContext
-  /// Custom operator APIs.
-  /// @{
-
-  /** \brief Used for custom operators, get the input count of a kernel
-   *
-   * \see ::OrtCustomOp
-   */
-  ORT_API2_STATUS(KernelContext_GetInputCount, _In_ const OrtKernelContext* context, _Out_ size_t* out);
-
-  /** \brief Used for custom operators, get the output count of a kernel
-   *
-   * \see ::OrtCustomOp
-   */
-  ORT_API2_STATUS(KernelContext_GetOutputCount, _In_ const OrtKernelContext* context, _Out_ size_t* out);
-
-  /** \brief Used for custom operators, get an input of a kernel
-   *
-   * \see ::OrtCustomOp
-   */
-  ORT_API2_STATUS(KernelContext_GetInput, _In_ const OrtKernelContext* context, _In_ size_t index,
-                  _Out_ const OrtValue** out);
-
-  /** \brief Used for custom operators, get an output of a kernel
-   *
-   * \see ::OrtCustomOp
-   */
-  ORT_API2_STATUS(KernelContext_GetOutput, _Inout_ OrtKernelContext* context, _In_ size_t index,
-                  _In_ const int64_t* dim_values, size_t dim_count, _Outptr_ OrtValue** out);
-
-  /// @}
-  /// \name OrtEnv
-  /// @{
-  ORT_CLASS_RELEASE(Env);
-  /// @}
-  /// \name OrtStatus
-  /// @{
-  ORT_CLASS_RELEASE(Status);
-  /// @}
-  /// \name OrtMemoryInfo
-  /// @{
-  ORT_CLASS_RELEASE(MemoryInfo);
-  /// @}
-  /// \name OrtSession
-  /// @{
-  ORT_CLASS_RELEASE(Session);  // Don't call ReleaseSession from Dllmain (because session owns a thread pool)
-  /// @}
-  /// \name OrtValue
-  /// @{
-  ORT_CLASS_RELEASE(Value);
-  /// @}
-  /// \name OrtRunOptions
-  /// @{
-  ORT_CLASS_RELEASE(RunOptions);
-  /// @}
-  /// \name OrtTypeInfo
-  /// @{
-  ORT_CLASS_RELEASE(TypeInfo);
-  /// @}
-  /// \name OrtTensorTypeAndShapeInfo
-  /// @{
-  ORT_CLASS_RELEASE(TensorTypeAndShapeInfo);
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-  ORT_CLASS_RELEASE(SessionOptions);
-  /// @}
-  /// \name OrtCustomOpDomain
-  /// @{
-  ORT_CLASS_RELEASE(CustomOpDomain);
-
-  /// @}
-  /// \name OrtTypeInfo
-  /// @{
-
-  /** \brief Get denotation from type information
-   *
-   * Augments ::OrtTypeInfo to return denotations on the type.
-   *
-   * This is used by WinML to determine if an input/output is intended to be an Image or a Tensor.
-   *
-   * \param[in] type_info
-   * \param[out] denotation Pointer to the null terminated denotation string is written to this pointer. This pointer is valid until the object is destroyed or the name is changed, do not free.
-   * \param[out] len Length in bytes of the string returned in `denotation`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetDenotationFromTypeInfo, _In_ const OrtTypeInfo* type_info, _Out_ const char** const denotation,
-                  _Out_ size_t* len);
-
-  /** \brief Get detailed map information from an ::OrtTypeInfo
-   *
-   * This augments ::OrtTypeInfo to return an ::OrtMapTypeInfo when the type is a map.
-   * The OrtMapTypeInfo has additional information about the map's key type and value type.
-   *
-   * This is used by WinML to support model reflection APIs.
-   *
-   * \param[out] type_info
-   * \param[out] out A pointer to the ::OrtMapTypeInfo. Do not free this value
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CastTypeInfoToMapTypeInfo, _In_ const OrtTypeInfo* type_info,
-                  _Outptr_result_maybenull_ const OrtMapTypeInfo** out);
-
-  /** \brief Cast ::OrtTypeInfo to an ::OrtSequenceTypeInfo
-   *
-   * This api augments ::OrtTypeInfo to return an ::OrtSequenceTypeInfo when the type is a sequence.
-   * The ::OrtSequenceTypeInfo has additional information about the sequence's element type.
-   *
-   * This is used by WinML to support model reflection APIs.
-   *
-   * \param[in] type_info
-   * \param[out] out A pointer to the OrtSequenceTypeInfo. Do not free this value
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CastTypeInfoToSequenceTypeInfo, _In_ const OrtTypeInfo* type_info,
-                  _Outptr_result_maybenull_ const OrtSequenceTypeInfo** out);
-
-  /// @}
-  /// \name OrtMapTypeInfo
-  /// @{
-
-  /** \brief Get key type from an ::OrtMapTypeInfo
-   *
-   * Key types are restricted to being scalar types.
-   *
-   * This is used by WinML to support model reflection APIs.
-   *
-   * \param[in] map_type_info
-   * \param[out] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetMapKeyType, _In_ const OrtMapTypeInfo* map_type_info, _Out_ enum ONNXTensorElementDataType* out);
-
-  /** \brief Get the value type from an ::OrtMapTypeInfo
-   *
-   * \param[in] map_type_info
-   * \param[out] type_info
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetMapValueType, _In_ const OrtMapTypeInfo* map_type_info, _Outptr_ OrtTypeInfo** type_info);
-
-  /// @}
-  /// \name OrtSequenceTypeInfo
-  /// @{
-
-  /** \brief Get element type from an ::OrtSequenceTypeInfo
-   *
-   * This is used by WinML to support model reflection APIs.
-   *
-   * \param[in] sequence_type_info
-   * \param[out] type_info
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetSequenceElementType, _In_ const OrtSequenceTypeInfo* sequence_type_info,
-                  _Outptr_ OrtTypeInfo** type_info);
-
-  /// @}
-  /// \name OrtMapTypeInfo
-  /// @{
-  ORT_CLASS_RELEASE(MapTypeInfo);
-  /// @}
-  /// \name OrtSequenceTypeInfo
-  /// @{
-  ORT_CLASS_RELEASE(SequenceTypeInfo);
-
-  /// @}
-  /// \name OrtSession
-  /// @{
-
-  /** \brief End profiling and return filename of the profile data
-   *
-   * Profiling is turned on through OrtApi::EnableProfiling
-   *
-   * \param[in] session
-   * \param[in] allocator
-   * \param[out] out Null terminated string of the filename, allocated using `allocator`. Must be freed using `allocator`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionEndProfiling, _In_ OrtSession* session, _Inout_ OrtAllocator* allocator, _Outptr_ char** out);
-
-  /** \brief Get ::OrtModelMetadata from an ::OrtSession
-   *
-   * \param[in] session
-   * \param[out] out Newly created ::OrtModelMetadata. Must be freed using OrtApi::ReleaseModelMetadata
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetModelMetadata, _In_ const OrtSession* session, _Outptr_ OrtModelMetadata** out);
-
-  /// @}
-  /// \name OrtModelMetadata
-  /// @{
-
-  /** \brief Get `producer name` from an ::OrtModelMetadata
-   *
-   * \param[in] model_metadata
-   * \param[in] allocator
-   * \param[out] value Set to a null terminated string allocated using `allocator`. Must be freed using `allocator`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataGetProducerName, _In_ const OrtModelMetadata* model_metadata,
-                  _Inout_ OrtAllocator* allocator, _Outptr_ char** value);
-
-  /** \brief Get `graph name` from an ::OrtModelMetadata
-   *
-   * \param[in] model_metadata
-   * \param[in] allocator
-   * \param[out] value Set to a null terminated string allocated using `allocator`. Must be freed using `allocator`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataGetGraphName, _In_ const OrtModelMetadata* model_metadata,
-                  _Inout_ OrtAllocator* allocator, _Outptr_ char** value);
-
-  /** \brief Get `domain` from an ::OrtModelMetadata
-   *
-   * \param[in] model_metadata
-   * \param[in] allocator
-   * \param[out] value Set to a null terminated string allocated using `allocator`. Must be freed using `allocator`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataGetDomain, _In_ const OrtModelMetadata* model_metadata, _Inout_ OrtAllocator* allocator,
-                  _Outptr_ char** value);
-
-  /** \brief Get `description` from an ::OrtModelMetadata
-   *
-   * \param[in] model_metadata
-   * \param[in] allocator
-   * \param[out] value Set to a null terminated string allocated using `allocator`. Must be freed using `allocator`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataGetDescription, _In_ const OrtModelMetadata* model_metadata,
-                  _Inout_ OrtAllocator* allocator, _Outptr_ char** value);
-
-  /** \brief Return data for a key in the custom metadata map in an ::OrtModelMetadata
-   *
-   * \param[in] model_metadata
-   * \param[in] allocator
-   * \param[in] key Null terminated string
-   * \param[out] value Set to a null terminated string allocated using `allocator`. Must be freed using `allocator`
-   * `value` will be set to nullptr if the given key is not found in the custom metadata map.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataLookupCustomMetadataMap, _In_ const OrtModelMetadata* model_metadata,
-                  _Inout_ OrtAllocator* allocator, _In_ const char* key, _Outptr_result_maybenull_ char** value);
-
-  /** \brief Get version number from an ::OrtModelMetadata
-   *
-   * \param[in] model_metadata
-   * \param[out] value Set to the version number
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataGetVersion, _In_ const OrtModelMetadata* model_metadata, _Out_ int64_t* value);
-
-  ORT_CLASS_RELEASE(ModelMetadata);
-
-  /// @}
-  /// \name OrtEnv
-  /// @{
-
-  /** \brief Create an OrtEnv
-   *
-   * Create an environment with global threadpools that will be shared across sessions.
-   * Use this in conjunction with OrtApi::DisablePerSessionThreads or else the session will use
-   * its own thread pools.
-   *
-   * \param[in] log_severity_level The log severity level.
-   * \param[in] logid The log identifier.
-   * \param[in] tp_options
-   * \param[out] out Returned newly created OrtEnv. Must be freed with OrtApi::ReleaseEnv
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateEnvWithGlobalThreadPools, OrtLoggingLevel log_severity_level, _In_ const char* logid,
-                  _In_ const OrtThreadingOptions* tp_options, _Outptr_ OrtEnv** out);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Use global thread pool on a session
-   *
-   * Disable using per session thread pool and use the shared global threadpool.
-   * This should be used in conjunction with OrtApi::CreateEnvWithGlobalThreadPools.
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(DisablePerSessionThreads, _Inout_ OrtSessionOptions* options);
-
-  /// @}
-  /// \name OrtThreadingOptions
-  /// @{
-
-  /** \brief Create an ::OrtThreadingOptions
-   *
-   * \param[out] out Newly created ::OrtThreadingOptions. Must be freed with OrtApi::ReleaseThreadingOptions
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateThreadingOptions, _Outptr_ OrtThreadingOptions** out);
-
-  ORT_CLASS_RELEASE(ThreadingOptions);
-
-  /// @}
-  /// \name OrtModelMetadata
-  /// @{
-
-  /**
-   *
-   * \param[in] model_metadata
-   * \param[in] allocator
-   * \param[out] keys Array of null terminated strings (array count = num_keys) allocated using `allocator`.
-   *  The strings and the pointer array must be freed using `allocator`
-   *  `keys` will be set to nullptr if the custom metadata map is empty.
-   * \param[out] num_keys Set to the number of elements in the `keys` array
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataGetCustomMetadataMapKeys, _In_ const OrtModelMetadata* model_metadata,
-                  _Inout_ OrtAllocator* allocator, _Outptr_result_buffer_maybenull_(*num_keys) char*** keys, _Out_ int64_t* num_keys);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /**
-   *
-   * Override symbolic dimensions (by specific name strings) with actual values
-   * if known at session initialization time to enable optimizations that can
-   * take advantage of fixed values (such as memory planning, etc)
-   *
-   */
-  ORT_API2_STATUS(AddFreeDimensionOverrideByName,
-                  _Inout_ OrtSessionOptions* options, _In_ const char* dim_name,
-                  _In_ int64_t dim_value);
-
-  /// @}
-  /// \name Misc
-  /// @{
-
-  /** \brief Get the names of all available providers
-   *
-   * \note The providers in the list are not guaranteed to be usable. They may fail to load due to missing system dependencies.
-   *    For example, if the CUDA/cuDNN libraries are not installed, the CUDA provider will report an error when it is added to the session options.
-   *
-   * \param[out] out_ptr Set to a pointer to an array of null terminated strings of the available providers. The entries and the
-   *    array itself must be freed using OrtApi::ReleaseAvailableProviders
-   * \param[out] provider_length Set to the number of entries in the `out_ptr` array
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetAvailableProviders, _Outptr_ char*** out_ptr, _Out_ int* provider_length);
-
-  /** \brief Release data from OrtApi::GetAvailableProviders
-   *
-   * \param[in] ptr The `out_ptr` result from OrtApi::GetAvailableProviders.
-   * \param[in] providers_length The `provider_length` result from OrtApi::GetAvailableProviders
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ReleaseAvailableProviders, _In_ char** ptr,
-                  _In_ int providers_length);
-
-  /// @}
-  /// \name OrtValue
-  /// @{
-
-  /** \brief Get the length of a single string in a string tensor
-   *
-   * \param[in] value A string tensor
-   * \param[in] index Index of the string in the tensor
-   * \param[out] out Set to number of bytes of the string element
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetStringTensorElementLength, _In_ const OrtValue* value, size_t index, _Out_ size_t* out);
-
-  /** \brief Get a single string from a string tensor
-   *
-   * \param[in] value A string tensor
-   * \param[in] s_len Number of bytes in the `s` buffer. Must match the value returned by OrtApi::GetStringTensorElementLength.
-   * \param[in] index Index of the string in the tensor
-   * \param[out] s The string element contents in UTF-8 encoding. The string is NOT null-terminated.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetStringTensorElement, _In_ const OrtValue* value, size_t s_len, size_t index, _Out_writes_bytes_all_(s_len) void* s);
-
-  /** \brief Set a single string in a string tensor
-   *
-   * \param[in] value A string tensor
-   * \param[in] s A null terminated UTF-8 encoded string
-   * \param[in] index Index of the string in the tensor to set
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(FillStringTensorElement, _Inout_ OrtValue* value, _In_ const char* s, size_t index);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Set a session configuration entry as a pair of strings
-   *
-   * If a configuration with same key exists, this will overwrite the configuration with the given config_value.
-   *
-   * The config_key and the format of config_value are defined in onnxruntime_session_options_config_keys.h
-   *
-   * \param[in] options
-   * \param[in] config_key A null terminated string representation of the config key
-   * \param[in] config_value A null terminated string representation of the config value
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(AddSessionConfigEntry, _Inout_ OrtSessionOptions* options,
-                  _In_z_ const char* config_key, _In_z_ const char* config_value);
-
-  /// @}
-  /// \name OrtAllocator
-  /// @{
-
-  /** \brief Create an allocator for an ::OrtSession following an ::OrtMemoryInfo
-   *
-   * \param[in] session
-   * \param[in] mem_info valid ::OrtMemoryInfo instance
-   * \param[out] out Newly created ::OrtAllocator. Must be freed with OrtApi::ReleaseAllocator
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateAllocator, _In_ const OrtSession* session, _In_ const OrtMemoryInfo* mem_info,
-                  _Outptr_ OrtAllocator** out);
-
-  /** \brief Release an ::OrtAllocator obtained from OrtApi::CreateAllocator
-   */
-  ORT_CLASS_RELEASE(Allocator);
-
-  /// @}
-  /// \name OrtSession
-  /// @{
-
-  /** \brief Run a model using Io Bindings for the inputs & outputs
-   *
-   * \see OrtApi::Run
-   *
-   * \param[in] session
-   * \param[in] run_options
-   * \param[in] binding_ptr
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(RunWithBinding, _Inout_ OrtSession* session, _In_ const OrtRunOptions* run_options, _In_ const OrtIoBinding* binding_ptr);
-
-  /** \brief Create an ::OrtIoBinding instance
-   *
-   * An IoBinding object allows one to bind pre-allocated ::OrtValue%s to input names.
-   * Thus if you want to use a raw on device buffer as input or output you can avoid
-   * extra copy during runtime.
-   *
-   * \param[in] session
-   * \param[out] out Newly created ::OrtIoBinding. Must be freed with OrtApi::ReleaseIoBinding
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateIoBinding, _Inout_ OrtSession* session, _Outptr_ OrtIoBinding** out);
-
-  /// @}
-  /// \name OrtIoBinding
-  /// @{
-
-  /** \brief Release an ::OrtIoBinding obtained from OrtApi::CreateIoBinding
-   */
-  ORT_CLASS_RELEASE(IoBinding);
-
-  /** \brief Bind an ::OrtValue to an ::OrtIoBinding input
-   *
-   * When using OrtApi::RunWithBinding this value is used for the named input
-   *
-   * \param[in] binding_ptr
-   * \param[in] name Name for the model input
-   * \param[in] val_ptr ::OrtValue of Tensor type.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(BindInput, _Inout_ OrtIoBinding* binding_ptr, _In_ const char* name, _In_ const OrtValue* val_ptr);
-
-  /** \brief Bind an ::OrtValue to an ::OrtIoBinding output
-   *
-   * When using OrtApi::RunWithBinding this value is used for the named output
-   *
-   * \param[in] binding_ptr
-   * \param[in] name Null terminated string of the model output name
-   * \param[in] val_ptr ::OrtValue of Tensor type.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(BindOutput, _Inout_ OrtIoBinding* binding_ptr, _In_ const char* name, _In_ const OrtValue* val_ptr);
-
-  /** \brief Bind an ::OrtIoBinding output to a device
-   *
-   * Binds the ::OrtValue to a device which is specified by ::OrtMemoryInfo.
-   * You can either create an instance of ::OrtMemoryInfo with a device id or obtain one from the allocator that you have created/are using
-   * This is useful when one or more outputs have dynamic shapes and, it is hard to pre-allocate and bind a chunk of
-   * memory within ::OrtValue ahead of time.
-   *
-   * \see OrtApi::RunWithBinding
-   *
-   * \param[in] binding_ptr
-   * \param[in] name Null terminated string of the device name
-   * \param[in] mem_info_ptr
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(BindOutputToDevice, _Inout_ OrtIoBinding* binding_ptr, _In_ const char* name, _In_ const OrtMemoryInfo* mem_info_ptr);
-
-  /** \brief Get the names of an ::OrtIoBinding's outputs
-   *
-   * Returns the names of the outputs in the order they were bound. This is useful after running the model
-   * with bound outputs because the returned names are in order in which output ::OrtValue are returned. This is useful if
-   * the order of outputs and their names is not known.
-   *
-   * \param[in] binding_ptr
-   * \param[in] allocator Allocator used to allocate continuous buffers for output strings and lengths.
-   * \param[out] buffer Returns an array of non-null terminated UTF-8 strings. The number of strings stored is returned in the count parameter.
-   *   This buffer is allocated using `allocator` and must be freed using it.
-   * \param[out] lengths Returns an array of `count` lengths of the strings returned in `buffer`
-   *   This buffer is allocated using `allocator` and must be freed using it.
-   * \param[out] count Number of strings returned. If `binding_ptr` has no bound outputs, zero is returned,
-   *              no memory allocation is performed and buffer and lengths are set to nullptr.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetBoundOutputNames, _In_ const OrtIoBinding* binding_ptr, _In_ OrtAllocator* allocator,
-                  _Out_ char** buffer, _Out_writes_all_(count) size_t** lengths, _Out_ size_t* count);
-
-  /** \brief Get the output ::OrtValue objects from an ::OrtIoBinding
-   *
-   * Returns an array of pointers to individually allocated ::OrtValue%s that contain results of a model execution with OrtApi::RunWithBinding
-   * The array contains the same number of ::OrtValue%s and they are in the same order as they were bound with OrtApi::BindOutput
-   * or OrtApi::BindOutputToDevice.
-   *
-   * The returned ::OrtValue%s must be released using OrtApi::ReleaseValue after they are no longer needed.
-   * The array is allocated using the specified instance of the allocator and must be freed using the same allocator after
-   * all the ::OrtValue%s contained therein are individually released.
-   *
-   * \param[in] binding_ptr
-   * \param[in] allocator Allocator used to allocate output array
-   * \param[out] output Set to the allocated array of allocated ::OrtValue outputs. Set to nullptr if there are 0 outputs.
-   * \param[out] output_count Set to number of ::OrtValue%s returned
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetBoundOutputValues, _In_ const OrtIoBinding* binding_ptr, _In_ OrtAllocator* allocator,
-                  _Out_writes_all_(output_count) OrtValue*** output, _Out_ size_t* output_count);
-
-  /** \brief Clears any previously set Inputs for an ::OrtIoBinding
-   */
-  void(ORT_API_CALL* ClearBoundInputs)(_Inout_ OrtIoBinding* binding_ptr) NO_EXCEPTION ORT_ALL_ARGS_NONNULL;
-
-  /** \brief Clears any previously set Outputs for an ::OrtIoBinding
-   */
-  void(ORT_API_CALL* ClearBoundOutputs)(_Inout_ OrtIoBinding* binding_ptr) NO_EXCEPTION ORT_ALL_ARGS_NONNULL;
-
-  /// @}
-  /// \name OrtValue
-  /// @{
-
-  /** \brief Direct memory access to a specified tensor element
-   *
-   * For example, given a tensor with shape of [3,224,224], a pointer to the element at location [2,150,128] can be retrieved
-   *
-   * This function only works for numeric type tensors (No strings, etc).
-   * This is a no-copy method whose returned pointer is valid until the passed in ::OrtValue is free'd.
-   *
-   * \param[in] value
-   * \param[in] location_values Pointer to an array of index values that specify an element's location relative to its shape
-   * \param[in] location_values_count Number of elements in location_values. Must match the number of elements in the tensor's shape.
-   * \param[out] out Set to a pointer to the element specified
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(TensorAt, _Inout_ OrtValue* value, const int64_t* location_values, size_t location_values_count, _Outptr_ void** out);
-
-  /// @}
-  /// \name OrtEnv
-  /// @{
-
-  /** \brief Create an allocator and register it with the ::OrtEnv
-   *
-   * Enables sharing the allocator between multiple sessions that use the same env instance.
-   * Lifetime of the created allocator will be valid for the duration of the environment.
-   * Returns an error if an allocator with the same ::OrtMemoryInfo is already registered.
-   *
-   * See https://onnxruntime.ai/docs/reference/api/c-api.html for details.
-   *
-   * \param[in] env ::OrtEnv instance
-   * \param[in] mem_info
-   * \param[in] arena_cfg Pass nullptr for defaults
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateAndRegisterAllocator, _Inout_ OrtEnv* env, _In_ const OrtMemoryInfo* mem_info,
-                  _In_ const OrtArenaCfg* arena_cfg);
-
-  /** \brief Set language projection
-   *
-   * Set the language projection for collecting telemetry data when Env is created.
-   *
-   * The default is ORT_PROJECTION_C, which means it will classify the language not in the list to C also.
-   *
-   * \param[in] ort_env
-   * \param[in] projection
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetLanguageProjection, _In_ const OrtEnv* ort_env, _In_ OrtLanguageProjection projection);
-
-  /// @}
-  /// \name OrtSession
-  /// @{
-
-  /** \brief Return the time that profiling was started
-   *
-   * \note The timer precision varies per platform. On Windows and MacOS, the precision will be ~100ns
-   *
-   * \param[in] session
-   * \param[out] out nanoseconds of profiling's start time
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetProfilingStartTimeNs, _In_ const OrtSession* session, _Outptr_ uint64_t* out);
-
-  /// @}
-  /// \name OrtThreadingOptions
-  /// @{
-
-  /** \brief Set global intra-op thread count
-   *
-   * This configures the global thread pool options to be used in the call to OrtApi::CreateEnvWithGlobalThreadPools
-   *
-   * \param[in] tp_options
-   * \param[in] intra_op_num_threads Number of threads, special values:<br>
-   *    0 = Use default thread count<br>
-   *    1 = The invoking thread will be used; no threads will be created in the thread pool.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetGlobalIntraOpNumThreads, _Inout_ OrtThreadingOptions* tp_options, int intra_op_num_threads);
-
-  /** \brief Set global inter-op thread count
-   *
-   * This configures the global thread pool options to be used in the call to OrtApi::CreateEnvWithGlobalThreadPools
-   *
-   * \param[in] tp_options
-   * \param[in] inter_op_num_threads Number of threads, special values:<br>
-   *    0 = Use default thread count<br>
-   *    1 = The invoking thread will be used; no threads will be created in the thread pool.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetGlobalInterOpNumThreads, _Inout_ OrtThreadingOptions* tp_options, int inter_op_num_threads);
-
-  /** \brief Set global spin control options
-   *
-   * This will configure the global thread pool options to be used in the call to OrtApi::CreateEnvWithGlobalThreadPools.
-   * Allow spinning of thread pools when their queues are empty. This will set the value for both
-   * inter_op and intra_op threadpools.
-   *
-   * \param[in] tp_options
-   * \param[in] allow_spinning Valid values are 0 or 1.<br>
-   *   0 = It won't spin (recommended if CPU usage is high)<br>
-   *   1 = Threadpool will spin to wait for queue to become non-empty
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetGlobalSpinControl, _Inout_ OrtThreadingOptions* tp_options, int allow_spinning);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Add a pre-allocated initializer to a session
-   *
-   * If a model contains an initializer with a name that is same as the name passed to this call,
-   * ORT will use this initializer instance instead of deserializing one from the model file. This
-   * is useful when you want to share the same initializer across sessions.
-   *
-   * \param[in] options
-   * \param[in] name Null terminated string of the initializer name
-   * \param[in] val ::OrtValue containing the initializer. Its lifetime and the underlying initializer buffer must be
-   *   managed by the user (created using the OrtApi::CreateTensorWithDataAsOrtValue) and it must outlive the session object
-   *   to which it is added.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(AddInitializer, _Inout_ OrtSessionOptions* options, _In_z_ const char* name,
-                  _In_ const OrtValue* val);
-
-  /// @}
-  /// \name OrtEnv
-  /// @{
-
-  /**
-   * Create a custom environment with global threadpools and logger that will be shared across sessions.
-   * Use this in conjunction with OrtApi::DisablePerSessionThreads or else the session will use
-   * its own thread pools.
-   *
-   * \param[in] logging_function A pointer to a logging function.
-   * \param[in] logger_param A pointer to arbitrary data passed as the ::OrtLoggingFunction `param` parameter to
-   *                         `logging_function`.
-   * \param[in] log_severity_level The log severity level.
-   * \param[in] logid The log identifier.
-   * \param[in] tp_options
-   * \param[out] out Newly created OrtEnv. Must be freed with OrtApi::ReleaseEnv
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateEnvWithCustomLoggerAndGlobalThreadPools, OrtLoggingFunction logging_function, _In_opt_ void* logger_param, OrtLoggingLevel log_severity_level,
-                  _In_ const char* logid, _In_ const struct OrtThreadingOptions* tp_options, _Outptr_ OrtEnv** out);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Append CUDA provider to session options
-   *
-   * If CUDA is not available (due to a non CUDA enabled build, or if CUDA is not installed on the system), this function will return failure.
-   *
-   * \param[in] options
-   * \param[in] cuda_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_CUDA,
-                  _In_ OrtSessionOptions* options, _In_ const OrtCUDAProviderOptions* cuda_options);
-
-  /** \brief Append ROCM execution provider to the session options
-   *
-   * If ROCM is not available (due to a non ROCM enabled build, or if ROCM is not installed on the system), this function will return failure.
-   *
-   * \param[in] options
-   * \param[in] rocm_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_ROCM,
-                  _In_ OrtSessionOptions* options, _In_ const OrtROCMProviderOptions* rocm_options);
-
-  /** \brief Append OpenVINO execution provider to the session options
-   *
-   * If OpenVINO is not available (due to a non OpenVINO enabled build, or if OpenVINO is not installed on the system), this function will fail.
-   *
-   * \param[in] options
-   * \param[in] provider_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_OpenVINO,
-                  _In_ OrtSessionOptions* options, _In_ const OrtOpenVINOProviderOptions* provider_options);
-
-  /// @}
-  /// \name OrtThreadingOptions
-  /// @{
-
-  /** \brief Set threading flush-to-zero and denormal-as-zero
-   *
-   * Sets global thread pool options to be used in the call to OrtApi::CreateEnvWithGlobalThreadPools.
-   * Flush-to-zero and denormal-as-zero are applied to threads in both intra and inter global thread pool.
-   * \note This option is not needed if the models used have no denormals. Having no denormals is recommended as this option may hurt model accuracy.
-   *
-   * \param[in] tp_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetGlobalDenormalAsZero, _Inout_ OrtThreadingOptions* tp_options);
-
-  /// @}
-  /// \name OrtArenaCfg
-  /// @{
-
-  /** \deprecated Use OrtApi::CreateArenaCfgV2
-   *
-   * This will create the configuration of an arena that can eventually be used to define an arena based allocator's behavior
-   *
-   * \param[in] max_mem Use 0 to allow ORT to choose the default
-   * \param[in] arena_extend_strategy Use -1 to allow ORT to choose the default, 0 = kNextPowerOfTwo, 1 = kSameAsRequested
-   * \param[in] initial_chunk_size_bytes Use -1 to allow ORT to choose the default
-   * \param[in] max_dead_bytes_per_chunk Use -1 to allow ORT to choose the default
-   * \param[in] out A pointer to an OrtArenaCfg instance
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateArenaCfg, _In_ size_t max_mem, int arena_extend_strategy, int initial_chunk_size_bytes,
-                  int max_dead_bytes_per_chunk, _Outptr_ OrtArenaCfg** out);
-
-  ORT_CLASS_RELEASE(ArenaCfg);
-
-  /// @}
-  /// \name OrtModelMetadata
-  /// @{
-
-  /**
-   * Use this to obtain the description of the graph present in the model
-   * (doc_string field of the GraphProto message within the ModelProto message).
-   * If it doesn't exist, an empty string will be returned.
-   *
-   * \param[in] model_metadata An instance of ::OrtModelMetadata
-   * \param[in] allocator Allocator used to allocate the string that will be returned back
-   * \param[out] value Set to a null terminated string allocated using `allocator`.  The caller is responsible for freeing it using `allocator`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataGetGraphDescription, _In_ const OrtModelMetadata* model_metadata,
-                  _Inout_ OrtAllocator* allocator, _Outptr_ char** value);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Append TensorRT provider to session options
-   *
-   * If TensorRT is not available (due to a non TensorRT enabled build, or if TensorRT is not installed on the system), this function will return failure.
-   *
-   * \param[in] options
-   * \param[in] tensorrt_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_TensorRT,
-                  _In_ OrtSessionOptions* options, _In_ const OrtTensorRTProviderOptions* tensorrt_options);
-
-  /// @}
-  /// \name Misc
-  /// @{
-
-  /** \brief Set current GPU device ID
-   *
-   * Set the current device id of the GPU execution provider (CUDA/tensorrt/rocm). The device id should be less
-   * than the total number of devices available. This is only useful when multiple-GPUs are installed and it is
-   * required to restrict execution to a single GPU.
-   *
-   * \param[in] device_id
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetCurrentGpuDeviceId, _In_ int device_id);
-
-  /** \brief Get current GPU device ID
-   *
-   * Get the current device id of the GPU execution provider (CUDA/tensorrt/rocm).
-   *
-   * \see OrtApi::SetCurrentGpuDeviceId
-   *
-   * \param[out] device_id
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetCurrentGpuDeviceId, _In_ int* device_id);
-
-  /// @}
-  /// \name OrtKernelInfo
-  /// Custom operator APIs.
-  /// @{
-
-  /** \brief Fetch an array of int64_t values stored as an attribute in the graph node
-   *
-   *
-   * If `out` is nullptr, the value of `size` is set to the true size of the attribute
-   * array's size, and a success status is returned.
-   *
-   * If the `size` parameter is greater than or equal to the actual attribute array's size,
-   * the value of `size` is set to the true size of the attribute array's size,
-   * the provided memory is filled with the attribute's contents,
-   * and a success status is returned.
-   *
-   * If the `size` parameter is less than the actual attribute array's size and `out`
-   * is not nullptr, the value of `size` is set to the true size of the attribute array's size
-   * and a failure status is returned.)
-   *
-   * \param[in] info instance
-   * \param[in] name name of the attribute to be parsed
-   * \param[out] out pointer to memory where the attribute's contents are to be stored
-   * \param[in, out] size actual size of attribute array
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(KernelInfoGetAttributeArray_float, _In_ const OrtKernelInfo* info, _In_ const char* name,
-                  _Out_ float* out, _Inout_ size_t* size);
-
-  /** \brief Fetch an array of int64_t values stored as an attribute in the graph node
-   *
-   * If `out` is nullptr, the value of `size` is set to the true size of the attribute
-   * array's size, and a success status is returned.
-   *
-   * If the `size` parameter is greater than or equal to the actual attribute array's size,
-   * the value of `size` is set to the true size of the attribute array's size,
-   * the provided memory is filled with the attribute's contents,
-   * and a success status is returned.
-   *
-   * If the `size` parameter is less than the actual attribute array's size and `out`
-   * is not nullptr, the value of `size` is set to the true size of the attribute array's size
-   * and a failure status is returned.)
-   *
-   * \param[in] info instance
-   * \param[in] name name of the attribute to be parsed
-   * \param[out] out pointer to memory where the attribute's contents are to be stored
-   * \param[in, out] size actual size of attribute array
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(KernelInfoGetAttributeArray_int64, _In_ const OrtKernelInfo* info, _In_ const char* name,
-                  _Out_ int64_t* out, _Inout_ size_t* size);
-
-  /// @}
-  /// \name OrtArenaCfg
-  /// @{
-
-  /** \brief Create an ::OrtArenaCfg
-   *
-   * Create the configuration of an arena that can eventually be used to define an arena based allocator's behavior.
-   *
-   * Supported keys are (See https://onnxruntime.ai/docs/reference/api/c-api.html for details on what the
-   * following parameters mean and how to choose these values.):
-   * "max_mem": Maximum memory that can be allocated by the arena based allocator.
-   *  Use 0 for ORT to pick the best value. Default is 0.
-   * "arena_extend_strategy": 0 = kNextPowerOfTwo, 1 = kSameAsRequested.
-   *  Use -1 to allow ORT to choose the default.
-   * "initial_chunk_size_bytes": (Possible) Size of the first allocation in the arena.
-   *  Only relevant if arena strategy is `kNextPowerOfTwo`. Use -1 to allow ORT to choose the default.
-   *  Ultimately, the first allocation size is determined by the allocation memory request.
-   * "max_dead_bytes_per_chunk": Threshold of unused memory in an allocated chunk of arena memory after
-   *  crossing which the current chunk is chunked into 2.
-   * "initial_growth_chunk_size_bytes": (Possible) Size of the second allocation in the arena.
-   *  Only relevant if arena strategy is `kNextPowerOfTwo`. Use -1 to allow ORT to choose the default.
-   *  Ultimately, the allocation size is determined by the allocation memory request.
-   *  Further allocation sizes are governed by the arena extend strategy.
-   *
-   * \param[in] arena_config_keys Keys to configure the arena
-   * \param[in] arena_config_values Values to configure the arena
-   * \param[in] num_keys Number of keys in `arena_config_keys` and `arena_config_values`
-   * \param[out] out Newly created ::OrtArenaCfg. Must be freed with OrtApi::ReleaseArenaCfg
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateArenaCfgV2, _In_reads_(num_keys) const char* const* arena_config_keys,
-                  _In_reads_(num_keys) const size_t* arena_config_values, _In_ size_t num_keys,
-                  _Outptr_ OrtArenaCfg** out);
-
-  /// @}
-  /// \name OrtRunOptions
-  /// @{
-
-  /** \brief Set a single run configuration entry as a pair of strings
-   *
-   * If a configuration with same key exists, this will overwrite the configuration with the given config_value
-   *
-   * The config_key and the format of config_value are defined in onnxruntime_run_options_config_keys.h
-   *
-   * \param[in] options
-   * \param[in] config_key A null terminated string representation of the config key
-   * \param[in] config_value  A null terminated string representation of the config value
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(AddRunConfigEntry, _Inout_ OrtRunOptions* options,
-                  _In_z_ const char* config_key, _In_z_ const char* config_value);
-
-  /// @}
-  /// \name OrtPrepackedWeightsContainer
-  /// @{
-
-  /** \brief Create an ::OrtPrepackedWeightsContainer
-   *
-   * This container will hold pre-packed buffers of shared initializers for sharing between sessions
-   * (i.e.) if there are shared initializers that can be shared between sessions, the pre-packed buffers
-   * of these (if any) may possibly be shared to provide memory footprint savings. Pass this container
-   * to sessions that you would like to share pre-packed buffers of shared initializers at session
-   * creation time.
-   *
-   *  \param[out] out Newly created ::OrtPrepackedWeightsContainer. Must be freed with OrtApi::ReleasePrepackedWeightsContainer
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreatePrepackedWeightsContainer, _Outptr_ OrtPrepackedWeightsContainer** out);
-
-  /** \brief Release OrtPrepackedWeightsContainer instance
-   *
-   * \note instance must not be released until the sessions using it are released
-   */
-  ORT_CLASS_RELEASE(PrepackedWeightsContainer);
-
-  /// @}
-  /// \name OrtSession
-  /// @{
-
-  /** \brief Create session with prepacked weights container
-   *
-   * Same functionality offered by OrtApi::CreateSession except that a container that contains
-   * pre-packed weights' buffers is written into/read from by the created session.
-   * This is useful when used in conjunction with OrtApi::AddInitializer which injects
-   * shared initializer info into sessions. Wherever possible, the pre-packed versions of these
-   * shared initializers are cached in this container so that multiple sessions can just re-use
-   * these instead of duplicating these in memory.
-   *
-   * \param[in] env OrtEnv instance instance
-   * \param[in] model_path Null terminated string of the path (wchar on Windows, char otherwise)
-   * \param[in] options
-   * \param[in] prepacked_weights_container
-   * \param[out] out Newly created ::OrtSession. Must be freed with OrtApi::ReleaseSession
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateSessionWithPrepackedWeightsContainer, _In_ const OrtEnv* env, _In_ const ORTCHAR_T* model_path,
-                  _In_ const OrtSessionOptions* options, _Inout_ OrtPrepackedWeightsContainer* prepacked_weights_container,
-                  _Outptr_ OrtSession** out);
-
-  /** \brief Create session from memory with prepacked weights container
-   *
-   * Same functionality offered by OrtApi::CreateSessionFromArray except that a container that contains
-   * pre-packed weights' buffers is written into/read from by the created session.
-   * This is useful when used in conjunction with OrtApi::AddInitializer which injects
-   * shared initializer info into sessions. Wherever possible, the pre-packed versions of these
-   * shared initializers are cached in this container so that multiple sessions can just re-use
-   * these instead of duplicating these in memory.
-   *
-   * \param[in] env
-   * \param[in] model_data Array of bytes holding the model
-   * \param[in] model_data_length Number of bytes in `model_data_model`
-   * \param[in] options
-   * \param[in] prepacked_weights_container
-   * \param[out] out Newly created ::OrtSession. Must be freed with OrtApi::ReleaseSession
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateSessionFromArrayWithPrepackedWeightsContainer, _In_ const OrtEnv* env,
-                  _In_ const void* model_data, size_t model_data_length,
-                  _In_ const OrtSessionOptions* options, _Inout_ OrtPrepackedWeightsContainer* prepacked_weights_container,
-                  _Outptr_ OrtSession** out);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Append TensorRT execution provider to the session options
-   *
-   * If TensorRT is not available (due to a non TensorRT enabled build), this function will return failure.
-   *
-   * This is slightly different from OrtApi::SessionOptionsAppendExecutionProvider_TensorRT, it takes an
-   * ::OrtTensorRTProviderOptions which is publicly defined. This takes an opaque ::OrtTensorRTProviderOptionsV2
-   * which must be created with OrtApi::CreateTensorRTProviderOptions.
-   *
-   * For OrtApi::SessionOptionsAppendExecutionProvider_TensorRT, the user needs to instantiate ::OrtTensorRTProviderOptions
-   * as well as allocate/release buffers for some members of ::OrtTensorRTProviderOptions.
-   * Here, OrtApi::CreateTensorRTProviderOptions and Ortapi::ReleaseTensorRTProviderOptions will do the memory management for you.
-   *
-   * \param[in] options
-   * \param[in] tensorrt_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_TensorRT_V2,
-                  _In_ OrtSessionOptions* options, _In_ const OrtTensorRTProviderOptionsV2* tensorrt_options);
-
-  /// @}
-  /// \name OrtTensorRTProviderOptionsV2
-  /// @{
-
-  /** \brief Create an OrtTensorRTProviderOptionsV2
-   *
-   * \param[out] out Newly created ::OrtTensorRTProviderOptionsV2. Must be released with OrtApi::ReleaseTensorRTProviderOptions
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateTensorRTProviderOptions, _Outptr_ OrtTensorRTProviderOptionsV2** out);
-
-  /** \brief Set options in a TensorRT Execution Provider.
-   *
-   * Please refer to https://www.onnxruntime.ai/docs/reference/execution-providers/TensorRT-ExecutionProvider.html#c-api-example
-   * to know the available keys and values. Key should be in null terminated string format of the member of ::OrtTensorRTProviderOptionsV2
-   * and value should be its related range.
-   *
-   * For example, key="trt_max_workspace_size" and value="2147483648"
-   *
-   * \param[in] tensorrt_options
-   * \param[in] provider_options_keys Array of UTF-8 null-terminated string for provider options keys
-   * \param[in] provider_options_values Array of UTF-8 null-terminated string for provider options values
-   * \param[in] num_keys Number of elements in the `provider_option_keys` and `provider_options_values` arrays
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(UpdateTensorRTProviderOptions, _Inout_ OrtTensorRTProviderOptionsV2* tensorrt_options,
-                  _In_reads_(num_keys) const char* const* provider_options_keys,
-                  _In_reads_(num_keys) const char* const* provider_options_values,
-                  _In_ size_t num_keys);
-
-  /** \brief Get serialized TensorRT provider options string.
-   *
-   * For example, "trt_max_workspace_size=2147483648;trt_max_partition_iterations=10;trt_int8_enable=1;......"
-   *
-   * \param tensorrt_options - OrTensorRTProviderOptionsV2 instance
-   * \param allocator - a ptr to an instance of OrtAllocator obtained with OrtApi::CreateAllocator or OrtApi::GetAllocatorWithDefaultOptions
-   *                      the specified allocator will be used to allocate continuous buffers for output strings and lengths.
-   * \param ptr - is a UTF-8 null terminated string allocated using 'allocator'. The caller is responsible for using the same allocator to free it.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetTensorRTProviderOptionsAsString, _In_ const OrtTensorRTProviderOptionsV2* tensorrt_options, _Inout_ OrtAllocator* allocator, _Outptr_ char** ptr);
-
-  /** \brief Release an ::OrtTensorRTProviderOptionsV2
-   *
-   * \note This is an exception in the naming convention of other Release* functions, as the name of the method does not have the V2 suffix, but the type does
-   */
-  void(ORT_API_CALL* ReleaseTensorRTProviderOptions)(_Frees_ptr_opt_ OrtTensorRTProviderOptionsV2* input);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Enable custom operators
-   *
-   * See onnxruntime-extensions: https://github.com/microsoft/onnxruntime-extensions.git
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(EnableOrtCustomOps, _Inout_ OrtSessionOptions* options);
-
-  /// @}
-  /// \name OrtAllocator
-  /// @{
-
-  /** \brief Register a custom allocator
-   *
-   * Enables sharing between multiple sessions that use the same env instance.
-   * Returns an error if an allocator with the same ::OrtMemoryInfo is already registered.
-   *
-   * The behavior of this is exactly the same as OrtApi::CreateAndRegisterAllocator except
-   * instead of ORT creating an allocator based on provided info, in this case
-   * ORT uses the user-provided custom allocator.
-   * See https://onnxruntime.ai/docs/reference/api/c-api.html for details.
-   *
-   * \param[in] env
-   * \param[in] allocator User provided allocator
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(RegisterAllocator, _Inout_ OrtEnv* env, _In_ OrtAllocator* allocator);
-
-  /** \brief Unregister a custom allocator
-   *
-   * It is an error if you provide an ::OrtMemoryInfo not corresponding to any
-   * registered allocators for sharing.
-   *
-   * \param[in] env
-   * \param[in] mem_info
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(UnregisterAllocator, _Inout_ OrtEnv* env,
-                  _In_ const OrtMemoryInfo* mem_info);
-
-  /// @}
-  /// \name OrtValue
-  /// @{
-
-  /** \brief Sets *out to 1 iff an ::OrtValue is a SparseTensor, and 0 otherwise
-   *
-   * \param[in] value existing ::OrtValue
-   * \param[out] out unless an error occurs, contains 1 iff the value contains an instance
-   *  of sparse tensor or 0 otherwise.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(IsSparseTensor, _In_ const OrtValue* value, _Out_ int* out);
-
-  /** \brief Create an ::OrtValue with a sparse tensor that is empty.
-   *
-   * Use FillSparseTensor<Format>() functions to populate sparse tensor with non-zero values and
-   * format specific indices data.
-   * Use ReleaseValue to destroy the sparse tensor, this will also release the buffer inside the output value
-   * if any was allocated.
-   * \param[in,out] allocator allocator to use when performing an allocation. Allocation will be performed
-   *   by FillSparseTensor<Format>() APIs. The lifespan of the allocator instance must eclipse the lifespan
-   *   this sparse tensor instance as the same allocator will be used to free memory.
-   * \param[in] dense_shape shape of the original dense tensor
-   * \param[in] dense_shape_len number of shape dimensions being passed
-   * \param[in] type must be one of TENSOR_ELEMENT_DATA_TYPE_xxxx
-   * \param[out] out Should be freed by calling ReleaseValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateSparseTensorAsOrtValue, _Inout_ OrtAllocator* allocator, _In_ const int64_t* dense_shape,
-                  size_t dense_shape_len, ONNXTensorElementDataType type, _Outptr_ OrtValue** out);
-
-  /**
-   * This fills populates an empty tensor that was created using OrtApi::CreateSparseTensorAsOrtValue.
-   * This will allocate required memory and copy the supplied NNZ values and COO indices into that memory allocation.
-   * Memory allocation is performed using the allocator that was specified with OrtApi::CreateSparseTensorAsOrtValue.
-   *
-   * \param[in,out] ort_value ::OrtValue to populate with data
-   * \param[in] data_mem_info serves to identify the location of the data to be copied. If the allocator specified
-   *  at the creation time has memory info that is not the same as mem_info argument to this function a X-device copy will be performed.
-   *  String data is assumed to be on CPU and will only be copied into a CPU allocated buffer.
-   * \param[in] values_shape pointer to values shape array
-   * \param[in] values_shape_len length of the values_shape
-   * \param[in] values pointer to an array of values. For strings, pass const char**.
-   * \param[in] indices_data pointer to a location of COO indices
-   * \param[in] indices_num number of COO indices
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(FillSparseTensorCoo, _Inout_ OrtValue* ort_value, _In_ const OrtMemoryInfo* data_mem_info,
-                  _In_ const int64_t* values_shape, size_t values_shape_len, _In_ const void* values,
-                  _In_ const int64_t* indices_data, size_t indices_num);
-
-  /**
-   * This fills populates an empty tensor that was created using OrtApi::CreateSparseTensorAsOrtValue.
-   * This will allocate required memory and copy the supplied NNZ values and CSR indices into that memory allocation.
-   * Memory allocation is performed using the allocator that was specified with OrtApi::CreateSparseTensorAsOrtValue.
-   *
-   * \param[in,out] ort_value ::OrtValue to populate with data
-   * \param[in] data_mem_info serves to identify the location of the data to be copied. If the allocator specified
-   *  at the creation time has memory info that is not the same as mem_info argument to this function a X-device copy will be performed.
-   *  String data is assumed to be on CPU and will only be copied into a CPU allocated buffer.
-   * \param[in] values_shape pointer to values shape array
-   * \param[in] values_shape_len length of the values_shape
-   * \param[in] values - pointer to an array of values. For strings, pass const char**.
-   * \param[in] inner_indices_data pointer to a location of CSR inner indices
-   * \param[in] inner_indices_num number of CSR inner indices
-   * \param[in] outer_indices_data pointer to a location of CSR outer indices
-   * \param[in] outer_indices_num number of CSR outer indices
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(FillSparseTensorCsr, _Inout_ OrtValue* ort_value, _In_ const OrtMemoryInfo* data_mem_info,
-                  _In_ const int64_t* values_shape, size_t values_shape_len, _In_ const void* values,
-                  _In_ const int64_t* inner_indices_data, size_t inner_indices_num,
-                  _In_ const int64_t* outer_indices_data, size_t outer_indices_num);
-
-  /**
-   * This fills populates an empty tensor that was created using OrtApi::CreateSparseTensorAsOrtValue.
-   * This will allocate required memory and copy the supplied NNZ values and BlockSparse indices into that memory allocation.
-   * Memory allocation is performed using the allocator that was specified with OrtApi::CreateSparseTensorAsOrtValue.
-   *
-   * \param[in,out] ort_value ::OrtValue to populate with data
-   * \param[in] data_mem_info serves to identify the location of the data to be copied. If the allocator specified
-   *  at the creation time has memory info that is not the same as mem_info argument to this function a X-device copy will be performed.
-   *  String data is assumed to be on CPU and will only be copied into a CPU allocated buffer.
-   * \param[in] values_shape
-   * \param[in] values_shape_len
-   * \param[in] values structure with values information
-   * \param[in] indices_shape_data pointer to a location of indices shape
-   * \param[in] indices_shape_len length of the block sparse indices shape
-   * \param[in] indices_data pointer to a location of indices data. Shape will determine the length of the indices data.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(FillSparseTensorBlockSparse, _Inout_ OrtValue* ort_value, _In_ const OrtMemoryInfo* data_mem_info,
-                  _In_ const int64_t* values_shape, size_t values_shape_len, _In_ const void* values,
-                  _In_ const int64_t* indices_shape_data, size_t indices_shape_len,
-                  _In_ const int32_t* indices_data);
-
-  /**
-   * Create an ::OrtValue with a sparse tensor. This is the first step.
-   * Next, use Use<Format>Indices() functions to supply sparse tensor with
-   * format specific indices data and set its sparse format to a specific enum value.
-   * This will not perform memory allocations. It will
-   * use supplied user buffer which should outlive the created sparse tensor.
-   * Use OrtApi::ReleaseValue to destroy the sparse tensor. It would not release the supplied values buffer.
-   * This function can not be used to map strings from the user allocated memory. Strings must always be copied
-   * and have UTF-8 encoding. Therefore, use OrtApi::CreateSparseTensorAsOrtValue above and then fill it with data
-   * using appropriate Make*() function.
-   *
-   * \param[in] info memory info where sparse values reside.
-   * \param[in,out] p_data pointer to a user allocated buffer with values. To create a full sparse tensor with no non-zero
-   *   values, pass nullptr
-   * \param[in] dense_shape shape of the original dense tensor
-   * \param[in] dense_shape_len number of shape dimensions being passed
-   * \param[in] values_shape shape of the values data. To create a fully sparse tensor with no non-zero values,
-   *   pass {0} shape.
-   * \param[in] values_shape_len number of values shape dimensions
-   * \param[in] type must be one of TENSOR_ELEMENT_DATA_TYPE_xxxx
-   * \param[out] out Should be freed by calling ReleaseValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateSparseTensorWithValuesAsOrtValue, _In_ const OrtMemoryInfo* info, _Inout_ void* p_data,
-                  _In_ const int64_t* dense_shape, size_t dense_shape_len,
-                  _In_ const int64_t* values_shape, size_t values_shape_len,
-                  ONNXTensorElementDataType type, _Outptr_ OrtValue** out);
-
-  /**
-   * This assigns Coo format indices to the SparseTensor that was created by
-   * OrtApi::CreateSparseTensorWithValuesAsOrtValue above. It also sets OrtSparseFormat to
-   * ORT_SPARSE_COO. This will not allocate any additional memory for data. The life span of
-   * indices_data buffer should eclipse the life span of this ::OrtValue.
-   *
-   * \param[in,out] ort_value ::OrtValue instance constructed with OrtApi::CreateSparseTensorWithValuesAsOrtValue
-   * \param[in,out] indices_data pointer to a user pre-allocated buffer or nullptr for fully sparse tensors.
-   * \param[in] indices_num  number of COO indices. Should either be 0 for fully sparse tensors, be equal
-   *  to the number of nnz values specified to OrtApi::CreateSparseTensorWithValuesAsOrtValue for 1-D {nnz} indices or
-   *  be twice as number of nnz values for a  2-D indices {nnz, 2}
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(UseCooIndices, _Inout_ OrtValue* ort_value, _Inout_ int64_t* indices_data, size_t indices_num);
-
-  /**
-   * The assigns CSR format indices to the SparseTensor that was created by
-   * OrtApi::CreateSparseTensorWithValuesAsOrtValue above. It also sets OrtSparseFormat to
-   * ORT_SPARSE_CSRC. This will not allocate any additional memory for data. The life spans of
-   * inner_data and outer_data buffers should eclipse the life span of this ::OrtValue.
-   *
-   * \param[in,out] ort_value ::OrtValue instance constructed with OrtApi::CreateSparseTensorWithValuesAsOrtValue
-   * \param[in,out] inner_data pointer to a user pre-allocated buffer or nullptr for fully sparse tensors.
-   * \param[in] inner_num  number of inner CSR indices. Should either be 0 for fully sparse tensors or be equal
-   * to the number of nnz values specified to OrtApi::CreateSparseTensorWithValuesAsOrtValue.
-   * \param[in,out] outer_data pointer to user pre-allocated buffer or nullptr for fully sparse tensors.
-   * \param[in] outer_num number of CSR outer indices. Should either be 0 for fully sparse tensors or
-   * equal to rows + 1 of the dense shape.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(UseCsrIndices, _Inout_ OrtValue* ort_value, _Inout_ int64_t* inner_data, size_t inner_num,
-                  _Inout_ int64_t* outer_data, size_t outer_num);
-
-  /**
-   * The assigns BlockSparse format indices to the SparseTensor that was created by
-   * OrtApi::CreateSparseTensorWithValuesAsOrtValue above. It also sets OrtSparseFormat to
-   * ORT_SPARSE_BLOCK_SPARSE. This will not allocate any additional memory for data. The life span of
-   * indices_data buffer must eclipse the lifespan of this ::OrtValue.
-   *
-   * \param[in,out] ort_value OrtValue instance constructed with OrtApi::CreateSparseTensorWithValuesAsOrtValue
-   * \param[in] indices_shape pointer to indices shape. Use {0} for fully sparse tensors
-   * \param[in] indices_shape_len length of the indices shape
-   * \param[in,out] indices_data pointer to user pre-allocated buffer or nullptr for fully sparse tensors.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(UseBlockSparseIndices, _Inout_ OrtValue* ort_value, const int64_t* indices_shape, size_t indices_shape_len, _Inout_ int32_t* indices_data);
-
-  /** \brief Returns sparse tensor format enum iff a given ort value contains an instance of sparse tensor.
-   *
-   * \param[in] ort_value ::OrtValue that contains an instance of sparse tensor
-   * \param[out] out pointer to out parameter
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetSparseTensorFormat, _In_ const OrtValue* ort_value, _Out_ enum OrtSparseFormat* out);
-
-  /** \brief Returns data type and shape of sparse tensor values (nnz) iff ::OrtValue contains a SparseTensor.
-   *
-   * \param[in] ort_value An ::OrtValue that contains a fully constructed sparse tensor
-   * \param[out] out Must be freed by OrtApi::ReleaseTensorTypeAndShapeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetSparseTensorValuesTypeAndShape, _In_ const OrtValue* ort_value, _Outptr_ OrtTensorTypeAndShapeInfo** out);
-
-  /** \brief Returns numeric data for sparse tensor values (nnz). For string values use GetStringTensor*().
-   *
-   * \param[in] ort_value an instance of ::OrtValue containing sparse tensor
-   * \param[out] out returns a pointer to values data.  Do not attempt to free this ptr.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetSparseTensorValues, _In_ const OrtValue* ort_value, _Outptr_ const void** out);
-
-  /** \brief Returns data type, shape for the type of indices specified by indices_format.
-   *
-   * \param[in] ort_value ::OrtValue containing sparse tensor.
-   * \param[in] indices_format One of the indices formats. It is an error to request a format that the sparse
-   * tensor does not contain.
-   * \param[out] out an instance of ::OrtTensorTypeAndShapeInfo. Must be freed by OrtApi::ReleaseTensorTypeAndShapeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetSparseTensorIndicesTypeShape, _In_ const OrtValue* ort_value, enum OrtSparseIndicesFormat indices_format, _Outptr_ OrtTensorTypeAndShapeInfo** out);
-
-  /** \brief Returns indices data for the type of the indices specified by indices_format
-   *
-   * \param[in] ort_value ::OrtValue containing sparse tensor.
-   * \param[in] indices_format One of the indices formats. It is an error to request a format that the sparse tensor does not contain.
-   * \param[out] num_indices Pointer to where the number of indices entries is returned
-   * \param[out] indices Returned pointer to the indices data. Do not free the returned pointer as it refers to internal data owned by the ::OrtValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetSparseTensorIndices, _In_ const OrtValue* ort_value, enum OrtSparseIndicesFormat indices_format, _Out_ size_t* num_indices, _Outptr_ const void** indices);
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /**
-   * \brief Sets out to 1 iff an optional type OrtValue has an element, 0 otherwise (OrtValue is None)
-   * Use this API to find if the optional type OrtValue is None or not.
-   * If the optional type OrtValue is not None, use the OrtValue just like any other OrtValue.
-   * For example, if you get an OrtValue that corresponds to Optional(tensor) and
-   * if HasValue() returns true, use it as tensor and so on.
-
-   * \param[in] value Input OrtValue.
-   * \param[out] out indicating if the input OrtValue contains data (1) or if it is a None (0)
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(HasValue, _In_ const OrtValue* value, _Out_ int* out);
-
-  /// @}
-  /// \name OrtKernelContext
-  /// Custom operator APIs.
-  /// @{
-
-  /** \brief Used for custom operators, gets the GPU compute stream to use to launch the custom a GPU kernel
-   *   \see ::OrtCustomOp
-   * \param[in]  context OrtKernelContext instance
-   * \param[out] out Returns pointer to a GPU compute stream that can be used to launch the custom GPU kernel.
-   *             If retrieving the GPU compute stream is not relevant (GPU not enabled in the build, kernel partitioned to
-   *             some other EP), then a nullptr is returned as the output param.
-   *             Do not free or mutate the returned pointer as it refers to internal data owned by the underlying session.
-   *             Only use it for custom kernel launching.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(KernelContext_GetGPUComputeStream, _In_ const OrtKernelContext* context, _Outptr_ void** out);
-
-  /// @}
-  /// \name GetTensorMemoryInfo
-  /// @{
-  /** \brief Returns a pointer to the ::OrtMemoryInfo of a Tensor
-   * \param[in] value ::OrtValue containing tensor.
-   * \param[out] mem_info ::OrtMemoryInfo of the tensor. Do NOT free the returned pointer. It is valid for the lifetime of the ::OrtValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetTensorMemoryInfo, _In_ const OrtValue* value, _Out_ const OrtMemoryInfo** mem_info);
-
-  /// @}
-  /// \name GetExecutionProviderApi
-  /// @{
-  /** \brief Get a pointer to the requested version of the Execution Provider specific
-   * API extensions to the OrtApi
-   * \param[in] provider_name The name of the execution provider name. Currently only the following
-   * values are supported: "DML".
-   * \param[in] version Must be ::ORT_API_VERSION.
-   * \param[out] provider_api A void pointer containing a reference to the execution provider versioned api structure.
-   * For example, the provider_api pointer can be cast to the OrtDmlApi* when the provider_name is "DML".
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetExecutionProviderApi, _In_ const char* provider_name, _In_ uint32_t version, _Outptr_ const void** provider_api);
-
-  /// @}
-
-  /// \name SessionOptions
-  /// @{
-  /** \brief Set custom thread creation function
-   *
-   * \param[in] options Session options
-   * \param[in] ort_custom_create_thread_fn Custom thread creation function
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsSetCustomCreateThreadFn, _Inout_ OrtSessionOptions* options, _In_ OrtCustomCreateThreadFn ort_custom_create_thread_fn);
-
-  /** \brief Set creation options for custom thread
-   *
-   * \param[in] options Session options
-   * \param[in] ort_custom_thread_creation_options Custom thread creation options (can be nullptr)
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsSetCustomThreadCreationOptions, _Inout_ OrtSessionOptions* options, _In_ void* ort_custom_thread_creation_options);
-
-  /** \brief Set custom thread join function
-   *
-   * \param[in] options Session options
-   * \param[in] ort_custom_join_thread_fn Custom join thread function, must not be nullptr when ort_custom_create_thread_fn is set
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsSetCustomJoinThreadFn, _Inout_ OrtSessionOptions* options, _In_ OrtCustomJoinThreadFn ort_custom_join_thread_fn);
-  /// @}
-
-  /// \name OrtThreadingOptions
-  /// @{
-  /** \brief Set custom thread creation function for global thread pools
-   *
-   * \param[inout] tp_options
-   * \param[in] ort_custom_create_thread_fn Custom thread creation function
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetGlobalCustomCreateThreadFn, _Inout_ OrtThreadingOptions* tp_options, _In_ OrtCustomCreateThreadFn ort_custom_create_thread_fn);
-
-  /** \brief Set custom thread creation options for global thread pools
-   *
-   * \param[inout] tp_options
-   * \param[in] ort_custom_thread_creation_options Custom thread creation options (can be nullptr)
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetGlobalCustomThreadCreationOptions, _Inout_ OrtThreadingOptions* tp_options, _In_ void* ort_custom_thread_creation_options);
-
-  /** \brief Set custom thread join function for global thread pools
-   *
-   * \param[inout] tp_options
-   * \param[in] ort_custom_join_thread_fn Custom thread join function, must not be nullptr when global ort_custom_create_thread_fn is set
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetGlobalCustomJoinThreadFn, _Inout_ OrtThreadingOptions* tp_options, _In_ OrtCustomJoinThreadFn ort_custom_join_thread_fn);
-  /// @}
-
-  /** \brief Synchronize bound inputs. The call may be necessary for some providers, such as cuda,
-   *   in case the system that allocated bound memory operated on a different stream. However, the
-   *   operation is provider specific and could be a no-op.
-   *
-   * \param[inout] binding_ptr
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SynchronizeBoundInputs, _Inout_ OrtIoBinding* binding_ptr);
-
-  /** \brief Synchronize bound outputs. The call may be necessary for some providers, such as cuda,
-   *   in case the system that allocated bound memory operated on a different stream. However, the
-   *   operation is provider specific and could be a no-op.
-   *
-   * \param[inout] binding_ptr
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SynchronizeBoundOutputs, _Inout_ OrtIoBinding* binding_ptr);
-
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Append CUDA execution provider to the session options
-   *
-   * If CUDA is not available (due to a non CUDA enabled build), this function will return failure.
-   *
-   * This is slightly different from OrtApi::SessionOptionsAppendExecutionProvider_CUDA, it takes an
-   * ::OrtCUDAProviderOptions which is publicly defined. This takes an opaque ::OrtCUDAProviderOptionsV2
-   * which must be created with OrtApi::CreateCUDAProviderOptions.
-   *
-   * For OrtApi::SessionOptionsAppendExecutionProvider_CUDA, the user needs to instantiate ::OrtCUDAProviderOptions
-   * as well as allocate/release buffers for some members of ::OrtCUDAProviderOptions.
-   * Here, OrtApi::CreateCUDAProviderOptions and Ortapi::ReleaseCUDAProviderOptions will do the memory management for you.
-   *
-   * \param[in] options
-   * \param[in] cuda_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.11.
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_CUDA_V2,
-                  _In_ OrtSessionOptions* options, _In_ const OrtCUDAProviderOptionsV2* cuda_options);
-
-  /// @}
-  /// \name OrtCUDAProviderOptionsV2
-  /// @{
-
-  /** \brief Create an OrtCUDAProviderOptionsV2
-   *
-   * \param[out] out Newly created ::OrtCUDAProviderOptionsV2. Must be released with OrtApi::ReleaseCudaProviderOptions
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.11.
-   */
-  ORT_API2_STATUS(CreateCUDAProviderOptions, _Outptr_ OrtCUDAProviderOptionsV2** out);
-
-  /** \brief Set options in a CUDA Execution Provider.
-   *
-   * Please refer to https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html#configuration-options
-   * to know the available keys and values. Key should be in null terminated string format of the member of ::OrtCUDAProviderOptionsV2
-   * and value should be its related range.
-   *
-   * For example, key="device_id" and value="0"
-   *
-   * \param[in] cuda_options
-   * \param[in] provider_options_keys Array of UTF-8 null-terminated string for provider options keys
-   * \param[in] provider_options_values Array of UTF-8 null-terminated string for provider options values
-   * \param[in] num_keys Number of elements in the `provider_option_keys` and `provider_options_values` arrays
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.11.
-   */
-  ORT_API2_STATUS(UpdateCUDAProviderOptions, _Inout_ OrtCUDAProviderOptionsV2* cuda_options,
-                  _In_reads_(num_keys) const char* const* provider_options_keys,
-                  _In_reads_(num_keys) const char* const* provider_options_values,
-                  _In_ size_t num_keys);
-
-  /**
-   * Get serialized CUDA provider options string.
-   *
-   * For example, "device_id=0;arena_extend_strategy=0;......"
-   *
-   * \param cuda_options - OrtCUDAProviderOptionsV2 instance
-   * \param allocator - a ptr to an instance of OrtAllocator obtained with CreateAllocator() or GetAllocatorWithDefaultOptions()
-   *                      the specified allocator will be used to allocate continuous buffers for output strings and lengths.
-   * \param ptr - is a UTF-8 null terminated string allocated using 'allocator'. The caller is responsible for using the same allocator to free it.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.11.
-   */
-  ORT_API2_STATUS(GetCUDAProviderOptionsAsString, _In_ const OrtCUDAProviderOptionsV2* cuda_options, _Inout_ OrtAllocator* allocator, _Outptr_ char** ptr);
-
-  /** \brief Release an ::OrtCUDAProviderOptionsV2
-   *
-   * \note This is an exception in the naming convention of other Release* functions, as the name of the method does not have the V2 suffix, but the type does
-   *
-   * \since Version 1.11.
-   */
-  void(ORT_API_CALL* ReleaseCUDAProviderOptions)(_Frees_ptr_opt_ OrtCUDAProviderOptionsV2* input);
-
-  /// @}
-
-  /** \brief Append MIGraphX provider to session options
-   *
-   * If MIGraphX is not available (due to a non MIGraphX enabled build, or if MIGraphX is not installed on the system), this function will return failure.
-   *
-   * \param[in] options
-   * \param[in] migraphx_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.11.
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_MIGraphX,
-                  _In_ OrtSessionOptions* options, _In_ const OrtMIGraphXProviderOptions* migraphx_options);
-
-  /** \brief Replace initialized Tensors with external data with the data provided in initializers.
-   *
-   * The function will find the initialized TensorProtos with external data in the graph with the provided names and
-   * replace them with the provided tensors. The API verifies that the TensorProto being replaced
-   * has an external data reference and has the same name, dimensions and data type as its replacement. The replacement
-   * will occur before any of the optimizations take place. The data will be copied into the graph
-   * since TensorProto can't refer to the user provided buffers.
-   *
-   * Once the model has been loaded, the OrtValue(s) added to SessionOptions instance will be removed
-   * from the internal SessionOptions copy to save memory, the user provided buffers can then be deallocated
-   * and the SessionOptions instance that refers to them can be destroyed.
-   *
-   * \param[in] options
-   * \param[in] initializer_names Array of null terminated UTF-8 encoded strings of the initializers names.
-   * \param[in] initializers Array of ::OrtValue type
-   * \param[in] initializers_num Number of elements in the initializer_names and initializers
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.12.
-   */
-  ORT_API2_STATUS(AddExternalInitializers, _In_ OrtSessionOptions* options,
-                  _In_reads_(input_len) const char* const* initializer_names,
-                  _In_reads_(input_len) const OrtValue* const* initializers, size_t initializers_num);
-
-  /** \brief: Create attribute of onnxruntime operator
-   *
-   * \param[in] name Name of the attribute
-   * \param[in] data Data content of the attribute
-   * \param[in] len Number of bytes stored in data
-   * \param[in] type Data type
-   * \param[out] op_attr Attribute that has been created, which must be released by OrtApi::ReleaseOpAttr
-   *
-   * \since Version 1.12.
-   */
-  ORT_API2_STATUS(CreateOpAttr,
-                  _In_ const char* name,
-                  _In_ const void* data,
-                  _In_ int len,
-                  _In_ OrtOpAttrType type,
-                  _Outptr_ OrtOpAttr** op_attr);
-
-  /* \brief: Release op attribute
-   *
-   * \param[in] opAttr Attribute created by OrtApi::CreateOpAttr
-   *
-   * \since Version 1.12.
-   */
-  ORT_CLASS_RELEASE(OpAttr);
-
-  /** \brief: Create onnxruntime native operator
-   *
-   * \param[in] info Kernel info
-   * \param[in] op_name Operator name
-   * \param[in] domain Operator domain
-   * \param[in] version Operator opset version
-   * \param[in] type_constraint_names Name of the type contraints, such as "T" or "T1"
-   * \param[in] type_constraint_values Type of each contraints
-   * \param[in] type_constraint_count Number of contraints
-   * \param[in] attr_values Attributes used to initialize the operator
-   * \param[in] attr_count Number of the attributes
-   * \param[in] input_count Number of inputs
-   * \param[in] output_count Number of outputs
-   * \param[out] ort_op Operator that has been created
-   *
-   * \since Version 1.12.
-   */
-  ORT_API2_STATUS(CreateOp,
-                  _In_ const OrtKernelInfo* info,
-                  _In_ const char* op_name,
-                  _In_ const char* domain,
-                  _In_ int version,
-                  _In_opt_ const char** type_constraint_names,
-                  _In_opt_ const ONNXTensorElementDataType* type_constraint_values,
-                  _In_opt_ int type_constraint_count,
-                  _In_opt_ const OrtOpAttr* const* attr_values,
-                  _In_opt_ int attr_count,
-                  _In_ int input_count,
-                  _In_ int output_count,
-                  _Outptr_ OrtOp** ort_op);
-
-  /** \brief: Invoke the operator created by OrtApi::CreateOp
-   * The inputs must follow the order as specified in onnx specification
-   *
-   * \param[in] context Kernel context
-   * \param[in] ort_op Operator that has been created
-   * \param[in] input_values Array of inputs
-   * \param[in] input_count Number of inputs
-   * \param[in] output_values Array of outputs
-   * \param[in] output_count Number of outputs
-   *
-   * \since Version 1.12.
-   */
-  ORT_API2_STATUS(InvokeOp,
-                  _In_ const OrtKernelContext* context,
-                  _In_ const OrtOp* ort_op,
-                  _In_ const OrtValue* const* input_values,
-                  _In_ int input_count,
-                  _Inout_ OrtValue* const* output_values,
-                  _In_ int output_count);
-
-  /* \brief: Release an onnxruntime operator
-   *
-   * \param[in] Op Operator created by OrtApi::CreateOp
-   *
-   * \since Version 1.12.
-   */
-  ORT_CLASS_RELEASE(Op);
-
-  /** \brief: Append execution provider to the session options.
-   * \param[in] options
-   * \param[in] provider_name - provider to add.
-   * \param[in] provider_options_keys - keys to configure the provider options
-   * \param[in] provider_options_values - values to configure the provider options
-   * \param[in] num_keys - number of keys passed in
-   *
-   * Currently supported providers:
-   *   SNPE
-   *   XNNPACK
-   *
-   * Note: If an execution provider has a dedicated SessionOptionsAppendExecutionProvider_<provider name> function
-   *       that should be used to add it.
-   *
-   * SNPE supported keys:
-   *   "runtime": SNPE runtime engine, options: "CPU", "CPU_FLOAT32", "GPU", "GPU_FLOAT32_16_HYBRID", "GPU_FLOAT16",
-   *   "DSP", "DSP_FIXED8_TF", "AIP_FIXED_TF", "AIP_FIXED8_TF".
-   *   Mapping to SNPE Runtime_t definition: CPU, CPU_FLOAT32 => zdl::DlSystem::Runtime_t::CPU;
-   *   GPU, GPU_FLOAT32_16_HYBRID => zdl::DlSystem::Runtime_t::GPU;
-   *   GPU_FLOAT16 => zdl::DlSystem::Runtime_t::GPU_FLOAT16;
-   *   DSP, DSP_FIXED8_TF => zdl::DlSystem::Runtime_t::DSP.
-   *   AIP_FIXED_TF, AIP_FIXED8_TF => zdl::DlSystem::Runtime_t::AIP_FIXED_TF.
-   *   "priority": execution priority, options: "low", "normal".
-   *   "buffer_type": ITensor or user buffers, options: "ITENSOR", user buffer with different types - "TF8", "TF16", "UINT8", "FLOAT".
-   *   "ITENSOR" -- default, ITensor which is float only.
-   *   "TF8" -- quantized model required, "FLOAT" -- for both quantized or non-quantized model
-   *   If SNPE is not available (due to a non Snpe enabled build or its dependencies not being installed), this function will fail.
-   *
-   * XNNPACK supported keys:
-   *   "intra_op_num_threads": number of thread-pool size to use for XNNPACK execution provider.
-   *      default value is 0, which means to use the session thread-pool size.
-   *
-   * \since Version 1.12.
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider, _In_ OrtSessionOptions* options,
-                  _In_ const char* provider_name,
-                  _In_reads_(num_keys) const char* const* provider_options_keys,
-                  _In_reads_(num_keys) const char* const* provider_options_values,
-                  _In_ size_t num_keys);
-
-  /* \brief: Get a copy of kernel info
-   *
-   * \param[in] info Kernel info
-   * \param[out] info_copy Copy of kernel info
-   *
-   * \since Version 1.12.
-   */
-  ORT_API2_STATUS(CopyKernelInfo,
-                  _In_ const OrtKernelInfo* info,
-                  _Outptr_ OrtKernelInfo** info_copy);
-
-  /* \brief: Release kernel info
-   *
-   * \param[in] KernelInfo A copy of kernel info returned by CopyKernelInfo
-   *
-   * \since Version 1.12.
-   */
-  ORT_CLASS_RELEASE(KernelInfo);
-
-  /* \brief: Get the training C Api
-   *
-   * \since Version 1.13
-   */
-  const OrtTrainingApi*(ORT_API_CALL* GetTrainingApi)(uint32_t version)NO_EXCEPTION;
-
-  /** \brief Append CANN provider to session options
-   *
-   * If CANN is not available (due to a non CANN enabled build, or if CANN is not installed on the system), this function will return failure.
-   *
-   * \param[in] options
-   * \param[in] cann_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.13.
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_CANN,
-                  _In_ OrtSessionOptions* options, _In_ const OrtCANNProviderOptions* cann_options);
-
-  /** \brief Create an OrtCANNProviderOptions
-   *
-   * \param[out] out created ::OrtCANNProviderOptions. Must be released with OrtApi::ReleaseCANNProviderOptions
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.13.
-   */
-  ORT_API2_STATUS(CreateCANNProviderOptions, _Outptr_ OrtCANNProviderOptions** out);
-
-  /** \brief Set options in a CANN Execution Provider.
-   *
-   * \param[in] cann_options
-   * \param[in] provider_options_keys Array of UTF-8 null-terminated string for provider options keys
-   * \param[in] provider_options_values Array of UTF-8 null-terminated string for provider options values
-   * \param[in] num_keys Number of elements in the `provider_option_keys` and `provider_options_values` arrays
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.13.
-   */
-  ORT_API2_STATUS(UpdateCANNProviderOptions, _Inout_ OrtCANNProviderOptions* cann_options,
-                  _In_reads_(num_keys) const char* const* provider_options_keys,
-                  _In_reads_(num_keys) const char* const* provider_options_values,
-                  _In_ size_t num_keys);
-
-  /** \brief Get serialized CANN provider options string.
-   *
-   * \param[in] cann_options OrtCANNProviderOptions instance
-   * \param[in] allocator a ptr to an instance of OrtAllocator obtained with CreateAllocator()
-   *                      or GetAllocatorWithDefaultOptions(), the specified allocator will be used to allocate
-   *                      continuous buffers for output strings and lengths.
-   * \param[out] ptr is a UTF-8 null terminated string allocated using 'allocator'.
-   *                 The caller is responsible for using the same allocator to free it.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.13.
-   */
-  ORT_API2_STATUS(GetCANNProviderOptionsAsString, _In_ const OrtCANNProviderOptions* cann_options,
-                  _Inout_ OrtAllocator* allocator, _Outptr_ char** ptr);
-
-  /** \brief Release an OrtCANNProviderOptions
-   *
-   * \param[in] the pointer of OrtCANNProviderOptions which will been deleted
-   *
-   * \since Version 1.13.
-   */
-  void(ORT_API_CALL* ReleaseCANNProviderOptions)(_Frees_ptr_opt_ OrtCANNProviderOptions* input);
-
-  /*  \brief Get OrtDevice type from MemoryInfo
-   *
-   *  \since Version 1.14
-   */
-  void(ORT_API_CALL* MemoryInfoGetDeviceType)(_In_ const OrtMemoryInfo* ptr, _Out_ OrtMemoryInfoDeviceType* out);
-
-  /* \brief Update the OrtEnv instance with custom log severity level
-   *
-   * \param[in] ort_env The OrtEnv instance being used
-   * \param[in] log_severity_level The log severity level.
-   *
-   * \since Version 1.14.
-   */
-  ORT_API2_STATUS(UpdateEnvWithCustomLogLevel, _In_ OrtEnv* ort_env, OrtLoggingLevel log_severity_level);
-
-  /*  \brief Set affinities for intra op threads
-   *
-   * Affinity string follows format:
-   * logical_processor_id,logical_processor_id;logical_processor_id,logical_processor_id
-   * Semicolon isolates configurations among threads, while comma split processors where ith thread expected to attach to.
-   * e.g. 1,2,3;4,5
-   * specifies affinities for two threads, with the 1st thread attach to the 1st, 2nd, and 3rd processor, and 2nd thread to the 4th and 5th.
-   * To ease the configuration, an "interval" is also allowed:
-   * e.g. 1-8;8-16;17-24
-   * orders that the 1st thread runs on first eight processors, 2nd thread runs on next eight processors, and so forth.
-   * Note:
-   * 1. Once set, the number of thread affinities must equal to intra_op_num_threads - 1,
-   *    ort does not set affinity on the main thread which is started and managed by the calling app;
-   * 2. For windows, ort will infer the group id from a logical processor id, for example, assuming there are two groups with each has 64 logical processors,
-   *    an id of 64 will be inferred as the last processor of the 1st group, while 65 will be interpreted as the 1st processor of the second group.
-   *    Hence 64-65 is an invalid configuration, because a windows thread cannot be attached to processors across group boundary.
-   *
-   *  \since Version 1.14
-   */
-  ORT_API2_STATUS(SetGlobalIntraOpThreadAffinity, _Inout_ OrtThreadingOptions* tp_options, const char* affinity_string);
-
-  /** \brief Register custom ops from a shared library.
-   *
-   * Loads a shared library (.dll on windows, .so on linux, etc) named 'library_name' and looks for this entry point:
-   *		OrtStatus* RegisterCustomOps(OrtSessionOptions * options, const OrtApiBase* api);
-   * It then passes in the provided session options to this function along with the api base.
-   *
-   * The handle to the loaded library is automatically released by ORT when the last OrtSession that references the
-   * library handle is released. If no OrtSession is created, then the library handle is released when the provided
-   * OrtSessionOptions is released.
-   *
-   * \param[in] options The session options.
-   * \param[in] library_name The name of the shared library to load and register. Refer to OS-specific dynamic library
-   *                         loading utilities (e.g., LoadLibraryEx on Windows or dlopen on Linux/MacOS) for information
-   *                         on the format of library names and search paths.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(RegisterCustomOpsLibrary_V2, _Inout_ OrtSessionOptions* options, _In_ const ORTCHAR_T* library_name);
-
-  /** \brief Register custom ops by calling a RegisterCustomOpsFn function.
-   *
-   * Searches for registration_func_name and if found calls it.
-   *
-   * The library containing the function must either be linked against or previously loaded by the executable.
-   *
-   * If you want ONNX Runtime to load the library and manage its lifetime, use RegisterCustomOpsLibrary_V2.
-   *
-   * RegisterCustomOpsUsingFunction can be used in scenarios where it may not be possible for ONNX Runtime to load
-   * the library from a path. e.g. mobile platforms where the library must be linked into the app.
-   *
-   * The registration function must have the signature of RegisterCustomOpsFn:
-   *    OrtStatus* (*fn)(OrtSessionOptions* options, const OrtApiBase* api);
-   *
-   * See https://onnxruntime.ai/docs/reference/operators/add-custom-op.html for details on how the registration
-   * function should be implemented.
-   *
-   * \param[in] options OrtSessionOptions that is passed through as the first argument in the call to the
-   *                    registration function.
-   * \param[in] registration_func_name Name of registration function to use.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(RegisterCustomOpsUsingFunction, _Inout_ OrtSessionOptions* options,
-                  _In_ const char* registration_func_name);
-
-  /// @}
-  /// \name OrtKernelInfo
-  /// Custom operator APIs.
-  /// @{
-
-  /** \brief Get the number of inputs from ::OrtKernelInfo.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to query the number of inputs
-   * during kernel/session creation.
-   *
-   * \param[in] info Instance of ::OrtKernelInfo.
-   * \param[out] out Pointer to variable assigned with the result on success.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(KernelInfo_GetInputCount, _In_ const OrtKernelInfo* info, _Out_ size_t* out);
-
-  /** \brief Get the number of outputs from ::OrtKernelInfo.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to query the number of outputs
-   * during kernel/session creation.
-   *
-   * \param[in] info Instance of ::OrtKernelInfo.
-   * \param[out] out Pointer to variable assigned with the result on success.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(KernelInfo_GetOutputCount, _In_ const OrtKernelInfo* info, _Out_ size_t* out);
-
-  /** \brief Get the name of a ::OrtKernelInfo's input.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to query an input's name
-   * during kernel/session creation.
-   *
-   * If `out` is nullptr, the value of `size` is set to the size of the name
-   * string (including null-terminator), and a success status is returned.
-   *
-   * If the `size` parameter is greater than or equal to the name string's size,
-   * the value of `size` is set to the true size of the string (including null-terminator),
-   * the provided memory is filled with the string's contents, and a success status is returned.
-   *
-   * If the `size` parameter is less than the actual string's size and `out`
-   * is not nullptr, the value of `size` is set to the true size of the string
-   * and a failure status is returned.
-   *
-   * \param[in] info An instance of ::OrtKernelInfo.
-   * \param[in] index The index of the input name to get. Returns a failure status if out-of-bounds.
-   * \param[out] out Memory location into which to write the UTF-8 null-terminated string representing the input's name.
-   * \param[in,out] size Pointer to the size of the `out` buffer. See above comments for details.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(KernelInfo_GetInputName, _In_ const OrtKernelInfo* info, size_t index, _Out_ char* out,
-                  _Inout_ size_t* size);
-
-  /** \brief Get the name of a ::OrtKernelInfo's output.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to query an output's name
-   * during kernel/session creation.
-   *
-   * If `out` is nullptr, the value of `size` is set to the size of the name
-   * string (including null-terminator), and a success status is returned.
-   *
-   * If the `size` parameter is greater than or equal to the name string's size,
-   * the value of `size` is set to the true size of the string (including null-terminator),
-   * the provided memory is filled with the string's contents, and a success status is returned.
-   *
-   * If the `size` parameter is less than the actual string's size and `out`
-   * is not nullptr, the value of `size` is set to the true size of the string
-   * and a failure status is returned.
-   *
-   * \param[in] info An instance of ::OrtKernelInfo.
-   * \param[in] index The index of the output name to get. Returns a failure status if out-of-bounds.
-   * \param[out] out Memory location into which to write the UTF-8 null-terminated string representing the output's
-   *                 name.
-   * \param[in,out] size Pointer to the size of the `out` buffer. See above comments for details.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(KernelInfo_GetOutputName, _In_ const OrtKernelInfo* info, size_t index, _Out_ char* out,
-                  _Inout_ size_t* size);
-
-  /** \brief Get the type information for a ::OrtKernelInfo's input.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to query the shape and type information
-   * of an input during kernel/session creation.
-   *
-   * \param[in] info An instance of ::OrtKernelInfo.
-   * \param[out] type_info Pointer set to the resulting ::OrtTypeInfo. Must be freed with OrtApi::ReleaseTypeInfo.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(KernelInfo_GetInputTypeInfo, _In_ const OrtKernelInfo* info, size_t index,
-                  _Outptr_ OrtTypeInfo** type_info);
-
-  /** \brief Get the type information for a ::OrtKernelInfo's output.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to query the shape and type information
-   * of an output during kernel/session creation.
-   *
-   * \param[in] info An instance of ::OrtKernelInfo.
-   * \param[out] type_info Pointer set to the resulting ::OrtTypeInfo. Must be freed with OrtApi::ReleaseTypeInfo.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(KernelInfo_GetOutputTypeInfo, _In_ const OrtKernelInfo* info, size_t index,
-                  _Outptr_ OrtTypeInfo** type_info);
-
-  /** \brief Get a ::OrtValue tensor stored as an attribute in the graph node.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to get a tensor attribute.
-   *
-   * \param[in] info ::OrtKernelInfo instance.
-   * \param[in] name UTF-8 null-terminated string representing the attribute's name.
-   * \param[in] allocator Allocator used to allocate the internal tensor state.
-   * \param[out] out Returns newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue,
-   *                 which will also free internal tensor state allocated with the provided allocator.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(KernelInfoGetAttribute_tensor, _In_ const OrtKernelInfo* info, _In_z_ const char* name,
-                  _Inout_ OrtAllocator* allocator, _Outptr_ OrtValue** out);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// Custom operator APIs
-  /// @{
-
-  /** \brief Checks if the given session configuration entry exists.
-   *
-   * The config_key formats are defined in onnxruntime_session_options_config_keys.h
-   *
-   * Can be used in a custom operator library to check for session configuration entries
-   * that target one or more custom operators in the library. Example: The config entry
-   * custom_op.myop.some_key targets a custom op named "myop".
-   *
-   * \param[in] options The ::OrtSessionOptions instance.
-   * \param[in] config_key A null-terminated UTF-8 string representation of the configuration key.
-   * \param[out] out Pointer set to 1 if the entry exists and 0 otherwise.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(HasSessionConfigEntry, _In_ const OrtSessionOptions* options,
-                  _In_z_ const char* config_key, _Out_ int* out);
-
-  /** \brief Get a session configuration value.
-   *
-   * Returns a failure status if the configuration key does not exist.
-   * The config_key and the format of config_value are defined in onnxruntime_session_options_config_keys.h
-   *
-   * If `config_value` is nullptr, the value of `size` is set to the true size of the string
-   * value (including null-terminator), and a success status is returned.
-   *
-   * If the `size` parameter is greater than or equal to the actual string value's size,
-   * the value of `size` is set to the true size of the string value, the provided memory
-   * is filled with the value's contents, and a success status is returned.
-   *
-   * If the `size` parameter is less than the actual string value's size and `config_value`
-   * is not nullptr, the value of `size` is set to the true size of the string value
-   * and a failure status is returned.
-   *
-   * Can be used in a custom operator library to get session configuration entries
-   * that target one or more custom operators in the library. Example: The config entry
-   * custom_op.myop.some_key targets a custom op named "myop".
-   *
-   * \param[in] options The session options.
-   * \param[in] config_key A null-terminated UTF-8 string representation of the config key.
-   * \param[in] config_value Pointer to memory where the null-terminated UTF-8 string value will be stored.
-   * \param[in,out] size Pointer to the size of the `config_value` buffer. See above comments for details.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(GetSessionConfigEntry, _In_ const OrtSessionOptions* options,
-                  _In_z_ const char* config_key, _Out_ char* config_value, _Inout_ size_t* size);
-
-  /// @}
-
-#ifdef __cplusplus
-  OrtApi(const OrtApi&) = delete;  // Prevent users from accidentally copying the API structure, it should always be passed as a pointer
-#endif
-};
-
-/*
- * Steps to use a custom op:
- *   1 Create an OrtCustomOpDomain with the domain name used by the custom ops
- *   2 Create an OrtCustomOp structure for each op and add them to the domain
- *   3 Call OrtAddCustomOpDomain to add the custom domain of ops to the session options
- */
-
-// Specifies some characteristics of inputs/outputs of custom ops:
-// Specify if the inputs/outputs are one of:
-// 1) Non-optional (input/output must be present in the node)
-// 2) Optional (input/output may be absent in the node)
-// 3) Variadic: A variadic input or output specifies N (i.e., the minimum arity) or more operands.
-//              Only the last input or output of a custom op may be marked as variadic.
-//              The homogeneity of the variadic input or output determines whether all operands must be of the same
-//              tensor element type.
-typedef enum OrtCustomOpInputOutputCharacteristic {
-  INPUT_OUTPUT_REQUIRED = 0,
-  INPUT_OUTPUT_OPTIONAL,
-  INPUT_OUTPUT_VARIADIC,
-} OrtCustomOpInputOutputCharacteristic;
-
-/*
- * The OrtCustomOp structure defines a custom op's schema and its kernel callbacks. The callbacks are filled in by
- * the implementor of the custom op.
- */
-struct OrtCustomOp {
-  uint32_t version;  // Must be initialized to ORT_API_VERSION
-
-  // This callback creates the kernel, which is a user defined parameter that is passed to the Kernel* callbacks below.
-  void*(ORT_API_CALL* CreateKernel)(_In_ const struct OrtCustomOp* op, _In_ const OrtApi* api,
-                                    _In_ const OrtKernelInfo* info);
-
-  // Returns the name of the op
-  const char*(ORT_API_CALL* GetName)(_In_ const struct OrtCustomOp* op);
-
-  // Returns the type of the execution provider, return nullptr to use CPU execution provider
-  const char*(ORT_API_CALL* GetExecutionProviderType)(_In_ const struct OrtCustomOp* op);
-
-  // Returns the count and types of the input & output tensors
-  ONNXTensorElementDataType(ORT_API_CALL* GetInputType)(_In_ const struct OrtCustomOp* op, _In_ size_t index);
-  size_t(ORT_API_CALL* GetInputTypeCount)(_In_ const struct OrtCustomOp* op);
-  ONNXTensorElementDataType(ORT_API_CALL* GetOutputType)(_In_ const struct OrtCustomOp* op, _In_ size_t index);
-  size_t(ORT_API_CALL* GetOutputTypeCount)(_In_ const struct OrtCustomOp* op);
-
-  // Op kernel callbacks
-  void(ORT_API_CALL* KernelCompute)(_In_ void* op_kernel, _In_ OrtKernelContext* context);
-  void(ORT_API_CALL* KernelDestroy)(_In_ void* op_kernel);
-
-  // Returns the characteristics of the input & output tensors
-  OrtCustomOpInputOutputCharacteristic(ORT_API_CALL* GetInputCharacteristic)(_In_ const struct OrtCustomOp* op, _In_ size_t index);
-  OrtCustomOpInputOutputCharacteristic(ORT_API_CALL* GetOutputCharacteristic)(_In_ const struct OrtCustomOp* op, _In_ size_t index);
-
-  // Returns the memory type of the input tensors. This API allows the custom op
-  // to place the inputs on specific devices. By default, it returns
-  // OrtMemTypeDefault, which means the input is placed on the default device for
-  // the execution provider. If the inputs need to be with different memory tyeps,
-  // this function can be overridden to return the specific memory types.
-  OrtMemType(ORT_API_CALL* GetInputMemoryType)(_In_ const struct OrtCustomOp* op, _In_ size_t index);
-
-  // Returns the minimum number of input arguments expected for the variadic input.
-  // Applicable only for custom ops that have a variadic input.
-  int(ORT_API_CALL* GetVariadicInputMinArity)(_In_ const struct OrtCustomOp* op);
-
-  // Returns true (non-zero) if all arguments of a variadic input have to be of the same type (homogeneous),
-  // and false (zero) otherwise.
-  // Applicable only for custom ops that have a variadic input.
-  int(ORT_API_CALL* GetVariadicInputHomogeneity)(_In_ const struct OrtCustomOp* op);
-
-  // Returns the minimum number of output values expected for the variadic output.
-  // Applicable only for custom ops that have a variadic output.
-  int(ORT_API_CALL* GetVariadicOutputMinArity)(_In_ const struct OrtCustomOp* op);
-
-  // Returns true (non-zero) if all outputs values of a variadic output have to be of the same type (homogeneous),
-  // and false (zero) otherwise.
-  // Applicable only for custom ops that have a variadic output.
-  int(ORT_API_CALL* GetVariadicOutputHomogeneity)(_In_ const struct OrtCustomOp* op);
-};
-
-/*
- * This is the old way to add the CUDA provider to the session, please use SessionOptionsAppendExecutionProvider_CUDA above to access the latest functionality
- * This function always exists, but will only succeed if Onnxruntime was built with CUDA support and the CUDA provider shared library exists
- *
- * \param device_id CUDA device id, starts from zero.
- */
-ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_CUDA, _In_ OrtSessionOptions* options, int device_id);
-
-/*
- * This is the old way to add the MIGraphX provider to the session, please use
- * SessionOptionsAppendExecutionProvider_MIGraphX above to access the latest functionality
- * This function always exists, but will only succeed if Onnxruntime was built with
- * HIP support and the MIGraphX provider shared library exists
- *
- * \param device_id HIP device id, starts from zero.
- */
-ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_MIGraphX, _In_ OrtSessionOptions* options, int device_id);
-
-#ifdef __cplusplus
-}
-#endif
-
-//! @}
diff --git a/projects/llm_framework/include/onnxruntime/core/session/onnxruntime_cxx_api.h b/projects/llm_framework/include/onnxruntime/core/session/onnxruntime_cxx_api.h
deleted file mode 100644
index 97b2aa48..00000000
--- a/projects/llm_framework/include/onnxruntime/core/session/onnxruntime_cxx_api.h
+++ /dev/null
@@ -1,1876 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-// Summary: The Ort C++ API is a header only wrapper around the Ort C API.
-//
-// The C++ API simplifies usage by returning values directly instead of error codes, throwing exceptions on errors
-// and automatically releasing resources in the destructors. The primary purpose of C++ API is exception safety so
-// all the resources follow RAII and do not leak memory.
-//
-// Each of the C++ wrapper classes holds only a pointer to the C internal object. Treat them like smart pointers.
-// To create an empty object, pass 'nullptr' to the constructor (for example, Env e{nullptr};). However, you can't use them
-// until you assign an instance that actually holds an underlying object.
-//
-// For Ort objects only move assignment between objects is allowed, there are no copy constructors.
-// Some objects have explicit 'Clone' methods for this purpose.
-//
-// ConstXXXX types are copyable since they do not own the underlying C object, so you can pass them to functions as arguments
-// by value or by reference. ConstXXXX types are restricted to const only interfaces.
-//
-// UnownedXXXX are similar to ConstXXXX but also allow non-const interfaces.
-//
-// The lifetime of the corresponding owning object must eclipse the lifetimes of the ConstXXXX/UnownedXXXX types. They exists so you do not
-// have to fallback to C types and the API with the usual pitfalls. In general, do not use C API from your C++ code.
-
-#pragma once
-#include "onnxruntime_c_api.h"
-#include <cstddef>
-#include <array>
-#include <memory>
-#include <stdexcept>
-#include <string>
-#include <vector>
-#include <unordered_map>
-#include <utility>
-#include <type_traits>
-
-#ifdef ORT_NO_EXCEPTIONS
-#include <iostream>
-#endif
-
-/** \brief All C++ Onnxruntime APIs are defined inside this namespace
- *
- */
-namespace Ort {
-
-/** \brief All C++ methods that can fail will throw an exception of this type
- *
- * If <tt>ORT_NO_EXCEPTIONS</tt> is defined, then any error will result in a call to abort()
- */
-struct Exception : std::exception {
-  Exception(std::string&& string, OrtErrorCode code) : message_{std::move(string)}, code_{code} {}
-
-  OrtErrorCode GetOrtErrorCode() const { return code_; }
-  const char* what() const noexcept override { return message_.c_str(); }
-
- private:
-  std::string message_;
-  OrtErrorCode code_;
-};
-
-#ifdef ORT_NO_EXCEPTIONS
-// The #ifndef is for the very special case where the user of this library wants to define their own way of handling errors.
-// NOTE: This header expects control flow to not continue after calling ORT_CXX_API_THROW
-#ifndef ORT_CXX_API_THROW
-#define ORT_CXX_API_THROW(string, code)       \
-  do {                                        \
-    std::cerr << Ort::Exception(string, code) \
-                     .what()                  \
-              << std::endl;                   \
-    abort();                                  \
-  } while (false)
-#endif
-#else
-#define ORT_CXX_API_THROW(string, code) \
-  throw Ort::Exception(string, code)
-#endif
-
-// This is used internally by the C++ API. This class holds the global variable that points to the OrtApi,
-//  it's in a template so that we can define a global variable in a header and make
-// it transparent to the users of the API.
-template <typename T>
-struct Global {
-  static const OrtApi* api_;
-};
-
-// If macro ORT_API_MANUAL_INIT is defined, no static initialization will be performed. Instead, user must call InitApi() before using it.
-template <typename T>
-#ifdef ORT_API_MANUAL_INIT
-const OrtApi* Global<T>::api_{};
-inline void InitApi() { Global<void>::api_ = OrtGetApiBase()->GetApi(ORT_API_VERSION); }
-
-// Used by custom operator libraries that are not linked to onnxruntime. Sets the global API object, which is
-// required by C++ APIs.
-//
-// Example mycustomop.cc:
-//
-// #define ORT_API_MANUAL_INIT
-// #include <onnxruntime_cxx_api.h>
-// #undef ORT_API_MANUAL_INIT
-//
-// OrtStatus* ORT_API_CALL RegisterCustomOps(OrtSessionOptions* options, const OrtApiBase* api_base) {
-//   Ort::InitApi(api_base->GetApi(ORT_API_VERSION));
-//   // ...
-// }
-//
-inline void InitApi(const OrtApi* api) { Global<void>::api_ = api; }
-#else
-#if defined(_MSC_VER) && !defined(__clang__)
-#pragma warning(push)
-// "Global initializer calls a non-constexpr function." Therefore you can't use ORT APIs in the other global initializers.
-// Please define ORT_API_MANUAL_INIT if it conerns you.
-#pragma warning(disable : 26426)
-#endif
-const OrtApi* Global<T>::api_ = OrtGetApiBase()->GetApi(ORT_API_VERSION);
-#if defined(_MSC_VER) && !defined(__clang__)
-#pragma warning(pop)
-#endif
-#endif
-
-/// This returns a reference to the OrtApi interface in use
-inline const OrtApi& GetApi() { return *Global<void>::api_; }
-
-/// <summary>
-/// This is a C++ wrapper for OrtApi::GetAvailableProviders() and
-/// returns a vector of strings representing the available execution providers.
-/// </summary>
-/// <returns>vector of strings</returns>
-std::vector<std::string> GetAvailableProviders();
-
-/** \brief IEEE 754 half-precision floating point data type
- * \details It is necessary for type dispatching to make use of C++ API
- * The type is implicitly convertible to/from uint16_t.
- * The size of the structure should align with uint16_t and one can freely cast
- * uint16_t buffers to/from Ort::Float16_t to feed and retrieve data.
- *
- * Generally, you can feed any of your types as float16/blfoat16 data to create a tensor
- * on top of it, providing it can form a continuous buffer with 16-bit elements with no padding.
- * And you can also feed a array of uint16_t elements directly. For example,
- *
- * \code{.unparsed}
- * uint16_t values[] = { 15360, 16384, 16896, 17408, 17664};
- * constexpr size_t values_length = sizeof(values) / sizeof(values[0]);
- * std::vector<int64_t> dims = {values_length};  // one dimensional example
- * Ort::MemoryInfo info("Cpu", OrtDeviceAllocator, 0, OrtMemTypeDefault);
- * // Note we are passing bytes count in this api, not number of elements -> sizeof(values)
- * auto float16_tensor = Ort::Value::CreateTensor(info, values, sizeof(values),
- *                                                dims.data(), dims.size(), ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT16);
- * \endcode
- *
- * Here is another example, a little bit more elaborate. Let's assume that you use your own float16 type and you want to use
- * a templated version of the API above so the type is automatically set based on your type. You will need to supply an extra
- * template specialization.
- *
- * \code{.unparsed}
- * namespace yours { struct half {}; } // assume this is your type, define this:
- * namespace Ort {
- * template<>
- * struct TypeToTensorType<yours::half> { static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT16; };
- * } //namespace Ort
- *
- * std::vector<yours::half> values;
- * std::vector<int64_t> dims = {values.size()}; // one dimensional example
- * Ort::MemoryInfo info("Cpu", OrtDeviceAllocator, 0, OrtMemTypeDefault);
- * // Here we are passing element count -> values.size()
- * auto float16_tensor = Ort::Value::CreateTensor<yours::half>(info, values.data(), values.size(), dims.data(), dims.size());
- *
- *  \endcode
- */
-struct Float16_t {
-  uint16_t value;
-  constexpr Float16_t() noexcept : value(0) {}
-  constexpr Float16_t(uint16_t v) noexcept : value(v) {}
-  constexpr operator uint16_t() const noexcept { return value; }
-  constexpr bool operator==(const Float16_t& rhs) const noexcept { return value == rhs.value; };
-  constexpr bool operator!=(const Float16_t& rhs) const noexcept { return value != rhs.value; };
-};
-
-static_assert(sizeof(Float16_t) == sizeof(uint16_t), "Sizes must match");
-
-/** \brief bfloat16 (Brain Floating Point) data type
- * \details It is necessary for type dispatching to make use of C++ API
- * The type is implicitly convertible to/from uint16_t.
- * The size of the structure should align with uint16_t and one can freely cast
- * uint16_t buffers to/from Ort::BFloat16_t to feed and retrieve data.
- *
- * See also code examples for Float16_t above.
- */
-struct BFloat16_t {
-  uint16_t value;
-  constexpr BFloat16_t() noexcept : value(0) {}
-  constexpr BFloat16_t(uint16_t v) noexcept : value(v) {}
-  constexpr operator uint16_t() const noexcept { return value; }
-  constexpr bool operator==(const BFloat16_t& rhs) const noexcept { return value == rhs.value; };
-  constexpr bool operator!=(const BFloat16_t& rhs) const noexcept { return value != rhs.value; };
-};
-
-static_assert(sizeof(BFloat16_t) == sizeof(uint16_t), "Sizes must match");
-
-namespace detail {
-// This is used internally by the C++ API. This macro is to make it easy to generate overloaded methods for all of the various OrtRelease* functions for every Ort* type
-// This can't be done in the C API since C doesn't have function overloading.
-#define ORT_DEFINE_RELEASE(NAME) \
-  inline void OrtRelease(Ort##NAME* ptr) { GetApi().Release##NAME(ptr); }
-
-ORT_DEFINE_RELEASE(Allocator);
-ORT_DEFINE_RELEASE(MemoryInfo);
-ORT_DEFINE_RELEASE(CustomOpDomain);
-ORT_DEFINE_RELEASE(ThreadingOptions);
-ORT_DEFINE_RELEASE(Env);
-ORT_DEFINE_RELEASE(RunOptions);
-ORT_DEFINE_RELEASE(Session);
-ORT_DEFINE_RELEASE(SessionOptions);
-ORT_DEFINE_RELEASE(TensorTypeAndShapeInfo);
-ORT_DEFINE_RELEASE(SequenceTypeInfo);
-ORT_DEFINE_RELEASE(MapTypeInfo);
-ORT_DEFINE_RELEASE(TypeInfo);
-ORT_DEFINE_RELEASE(Value);
-ORT_DEFINE_RELEASE(ModelMetadata);
-ORT_DEFINE_RELEASE(IoBinding);
-ORT_DEFINE_RELEASE(ArenaCfg);
-ORT_DEFINE_RELEASE(Status);
-ORT_DEFINE_RELEASE(OpAttr);
-ORT_DEFINE_RELEASE(Op);
-ORT_DEFINE_RELEASE(KernelInfo);
-
-#undef ORT_DEFINE_RELEASE
-
-/** \brief This is a tagging template type. Use it with Base<T> to indicate that the C++ interface object
- *   has no ownership of the underlying C object.
- */
-template <typename T>
-struct Unowned {
-  using Type = T;
-};
-
-/** \brief Used internally by the C++ API. C++ wrapper types inherit from this.
- *   This is a zero cost abstraction to wrap the C API objects and delete them on destruction.
- *
- * All of the C++ classes
- *  a) serve as containers for pointers to objects that are created by the underlying C API.
- *     Their size is just a pointer size, no need to dynamically allocate them. Use them by value.
- *  b) Each of struct XXXX, XXX instances function as smart pointers to the underlying C API objects.
- *     they would release objects owned automatically when going out of scope, they are move-only.
- *  c) ConstXXXX and UnownedXXX structs function as non-owning, copyable containers for the above pointers.
- *     ConstXXXX allow calling const interfaces only. They give access to objects that are owned by somebody else
- *     such as Onnxruntime or instances of XXXX classes.
- *  d) serve convenient interfaces that return C++ objects and further enhance exception and type safety so they can be used
- *     in C++ code.
- *
- */
-
-/// <summary>
-/// This is a non-const pointer holder that is move-only. Disposes of the pointer on destruction.
-/// </summary>
-template <typename T>
-struct Base {
-  using contained_type = T;
-
-  constexpr Base() = default;
-  constexpr explicit Base(contained_type* p) noexcept : p_{p} {}
-  ~Base() { OrtRelease(p_); }
-
-  Base(const Base&) = delete;
-  Base& operator=(const Base&) = delete;
-
-  Base(Base&& v) noexcept : p_{v.p_} { v.p_ = nullptr; }
-  Base& operator=(Base&& v) noexcept {
-    OrtRelease(p_);
-    p_ = v.release();
-    return *this;
-  }
-
-  constexpr operator contained_type*() const noexcept { return p_; }
-
-  /// \brief Relinquishes ownership of the contained C object pointer
-  /// The underlying object is not destroyed
-  contained_type* release() {
-    T* p = p_;
-    p_ = nullptr;
-    return p;
-  }
-
- protected:
-  contained_type* p_{};
-};
-
-// Undefined. For const types use Base<Unowned<const T>>
-template <typename T>
-struct Base<const T>;
-
-/// <summary>
-/// Covers unowned pointers owned by either the ORT
-/// or some other instance of CPP wrappers.
-/// Used for ConstXXX and UnownedXXXX types that are copyable.
-/// Also convenient to wrap raw OrtXX pointers .
-/// </summary>
-/// <typeparam name="T"></typeparam>
-template <typename T>
-struct Base<Unowned<T>> {
-  using contained_type = typename Unowned<T>::Type;
-
-  constexpr Base() = default;
-  constexpr explicit Base(contained_type* p) noexcept : p_{p} {}
-
-  ~Base() = default;
-
-  Base(const Base&) = default;
-  Base& operator=(const Base&) = default;
-
-  Base(Base&& v) noexcept : p_{v.p_} { v.p_ = nullptr; }
-  Base& operator=(Base&& v) noexcept {
-    p_ = nullptr;
-    std::swap(p_, v.p_);
-    return *this;
-  }
-
-  constexpr operator contained_type*() const noexcept { return p_; }
-
- protected:
-  contained_type* p_{};
-};
-
-// Light functor to release memory with OrtAllocator
-struct AllocatedFree {
-  OrtAllocator* allocator_;
-  explicit AllocatedFree(OrtAllocator* allocator)
-      : allocator_(allocator) {}
-  void operator()(void* ptr) const {
-    if (ptr) allocator_->Free(allocator_, ptr);
-  }
-};
-
-}  // namespace detail
-
-struct AllocatorWithDefaultOptions;
-struct Env;
-struct TypeInfo;
-struct Value;
-struct ModelMetadata;
-
-/** \brief unique_ptr typedef used to own strings allocated by OrtAllocators
- *  and release them at the end of the scope. The lifespan of the given allocator
- *  must eclipse the lifespan of AllocatedStringPtr instance
- */
-using AllocatedStringPtr = std::unique_ptr<char, detail::AllocatedFree>;
-
-/** \brief The Status that holds ownership of OrtStatus received from C API
- *  Use it to safely destroy OrtStatus* returned from the C API. Use appropriate
- *  constructors to construct an instance of a Status object from exceptions.
- */
-struct Status : detail::Base<OrtStatus> {
-  explicit Status(std::nullptr_t) {}       ///< Create an empty object, must be assigned a valid one to be used
-  explicit Status(OrtStatus* status);      ///< Takes ownership of OrtStatus instance returned from the C API. Must be non-null
-  explicit Status(const Exception&);       ///< Creates status instance out of exception
-  explicit Status(const std::exception&);  ///< Creates status instance out of exception
-  std::string GetErrorMessage() const;
-  OrtErrorCode GetErrorCode() const;
-};
-
-/** \brief The ThreadingOptions
- *
- * The ThreadingOptions used for set global threadpools' options of The Env.
- */
-struct ThreadingOptions : detail::Base<OrtThreadingOptions> {
-  /// \brief Wraps OrtApi::CreateThreadingOptions
-  ThreadingOptions();
-
-  /// \brief Wraps OrtApi::SetGlobalIntraOpNumThreads
-  ThreadingOptions& SetGlobalIntraOpNumThreads(int intra_op_num_threads);
-
-  /// \brief Wraps OrtApi::SetGlobalInterOpNumThreads
-  ThreadingOptions& SetGlobalInterOpNumThreads(int inter_op_num_threads);
-
-  /// \brief Wraps OrtApi::SetGlobalSpinControl
-  ThreadingOptions& SetGlobalSpinControl(int allow_spinning);
-
-  /// \brief Wraps OrtApi::SetGlobalDenormalAsZero
-  ThreadingOptions& SetGlobalDenormalAsZero();
-
-  /// \brief Wraps OrtApi::SetGlobalCustomCreateThreadFn
-  ThreadingOptions& SetGlobalCustomCreateThreadFn(OrtCustomCreateThreadFn ort_custom_create_thread_fn);
-
-  /// \brief Wraps OrtApi::SetGlobalCustomThreadCreationOptions
-  ThreadingOptions& SetGlobalCustomThreadCreationOptions(void* ort_custom_thread_creation_options);
-
-  /// \brief Wraps OrtApi::SetGlobalCustomJoinThreadFn
-  ThreadingOptions& SetGlobalCustomJoinThreadFn(OrtCustomJoinThreadFn ort_custom_join_thread_fn);
-};
-
-/** \brief The Env (Environment)
- *
- * The Env holds the logging state used by all other objects.
- * <b>Note:</b> One Env must be created before using any other Onnxruntime functionality
- */
-struct Env : detail::Base<OrtEnv> {
-  explicit Env(std::nullptr_t) {}  ///< Create an empty Env object, must be assigned a valid one to be used
-
-  /// \brief Wraps OrtApi::CreateEnv
-  Env(OrtLoggingLevel logging_level = ORT_LOGGING_LEVEL_WARNING, _In_ const char* logid = "");
-
-  /// \brief Wraps OrtApi::CreateEnvWithCustomLogger
-  Env(OrtLoggingLevel logging_level, const char* logid, OrtLoggingFunction logging_function, void* logger_param);
-
-  /// \brief Wraps OrtApi::CreateEnvWithGlobalThreadPools
-  Env(const OrtThreadingOptions* tp_options, OrtLoggingLevel logging_level = ORT_LOGGING_LEVEL_WARNING, _In_ const char* logid = "");
-
-  /// \brief Wraps OrtApi::CreateEnvWithCustomLoggerAndGlobalThreadPools
-  Env(const OrtThreadingOptions* tp_options, OrtLoggingFunction logging_function, void* logger_param,
-      OrtLoggingLevel logging_level = ORT_LOGGING_LEVEL_WARNING, _In_ const char* logid = "");
-
-  /// \brief C Interop Helper
-  explicit Env(OrtEnv* p) : Base<OrtEnv>{p} {}
-
-  Env& EnableTelemetryEvents();   ///< Wraps OrtApi::EnableTelemetryEvents
-  Env& DisableTelemetryEvents();  ///< Wraps OrtApi::DisableTelemetryEvents
-
-  Env& UpdateEnvWithCustomLogLevel(OrtLoggingLevel log_severity_level);  ///< Wraps OrtApi::UpdateEnvWithCustomLogLevel
-
-  Env& CreateAndRegisterAllocator(const OrtMemoryInfo* mem_info, const OrtArenaCfg* arena_cfg);  ///< Wraps OrtApi::CreateAndRegisterAllocator
-};
-
-/** \brief Custom Op Domain
- *
- */
-struct CustomOpDomain : detail::Base<OrtCustomOpDomain> {
-  explicit CustomOpDomain(std::nullptr_t) {}  ///< Create an empty CustomOpDomain object, must be assigned a valid one to be used
-
-  /// \brief Wraps OrtApi::CreateCustomOpDomain
-  explicit CustomOpDomain(const char* domain);
-
-  // This does not take ownership of the op, simply registers it.
-  void Add(const OrtCustomOp* op);  ///< Wraps CustomOpDomain_Add
-};
-
-/** \brief RunOptions
- *
- */
-struct RunOptions : detail::Base<OrtRunOptions> {
-  explicit RunOptions(std::nullptr_t) {}  ///< Create an empty RunOptions object, must be assigned a valid one to be used
-  RunOptions();                           ///< Wraps OrtApi::CreateRunOptions
-
-  RunOptions& SetRunLogVerbosityLevel(int);  ///< Wraps OrtApi::RunOptionsSetRunLogVerbosityLevel
-  int GetRunLogVerbosityLevel() const;       ///< Wraps OrtApi::RunOptionsGetRunLogVerbosityLevel
-
-  RunOptions& SetRunLogSeverityLevel(int);  ///< Wraps OrtApi::RunOptionsSetRunLogSeverityLevel
-  int GetRunLogSeverityLevel() const;       ///< Wraps OrtApi::RunOptionsGetRunLogSeverityLevel
-
-  RunOptions& SetRunTag(const char* run_tag);  ///< wraps OrtApi::RunOptionsSetRunTag
-  const char* GetRunTag() const;               ///< Wraps OrtApi::RunOptionsGetRunTag
-
-  RunOptions& AddConfigEntry(const char* config_key, const char* config_value);  ///< Wraps OrtApi::AddRunConfigEntry
-
-  /** \brief Terminates all currently executing Session::Run calls that were made using this RunOptions instance
-   *
-   * If a currently executing session needs to be force terminated, this can be called from another thread to force it to fail with an error
-   * Wraps OrtApi::RunOptionsSetTerminate
-   */
-  RunOptions& SetTerminate();
-
-  /** \brief Clears the terminate flag so this RunOptions instance can be used in a new Session::Run call without it instantly terminating
-   *
-   * Wraps OrtApi::RunOptionsUnsetTerminate
-   */
-  RunOptions& UnsetTerminate();
-};
-
-
-namespace detail {
-// Utility function that returns a SessionOption config entry key for a specific custom operator.
-// Ex: custom_op.[custom_op_name].[config]
-std::string MakeCustomOpConfigEntryKey(const char* custom_op_name, const char* config);
-}  // namespace detail
-
-/// <summary>
-/// Class that represents session configuration entries for one or more custom operators.
-///
-/// Example:
-///   Ort::CustomOpConfigs op_configs;
-///   op_configs.AddConfig("my_custom_op", "device_type", "CPU");
-///
-/// Passed to Ort::SessionOptions::RegisterCustomOpsLibrary.
-/// </summary>
-struct CustomOpConfigs {
-  CustomOpConfigs() = default;
-  ~CustomOpConfigs() = default;
-  CustomOpConfigs(const CustomOpConfigs&) = default;
-  CustomOpConfigs& operator=(const CustomOpConfigs&) = default;
-  CustomOpConfigs(CustomOpConfigs&& o) = default;
-  CustomOpConfigs& operator=(CustomOpConfigs&& o) = default;
-
-  /** \brief Adds a session configuration entry/value for a specific custom operator.
-   *
-   * \param custom_op_name The name of the custom operator for which to add a configuration entry.
-   *                       Must match the name returned by the CustomOp's GetName() method.
-   * \param config_key The name of the configuration entry.
-   * \param config_value The value of the configuration entry.
-   * \return A reference to this object to enable call chaining.
-   */
-  CustomOpConfigs& AddConfig(const char* custom_op_name, const char* config_key, const char* config_value);
-
-  /** \brief Returns a flattened map of custom operator configuration entries and their values.
-   *
-   * The keys has been flattened to include both the custom operator name and the configuration entry key name.
-   * For example, a prior call to AddConfig("my_op", "key", "value") corresponds to the flattened key/value pair
-   * {"my_op.key", "value"}.
-   *
-   * \return An unordered map of flattened configurations.
-  */
-  const std::unordered_map<std::string, std::string>& GetFlattenedConfigs() const;
-
- private:
-  std::unordered_map<std::string, std::string> flat_configs_;
-};
-
-/** \brief Options object used when creating a new Session object
- *
- * Wraps ::OrtSessionOptions object and methods
- */
-
-struct SessionOptions;
-
-namespace detail {
-// we separate const-only methods because passing const ptr to non-const methods
-// is only discovered when inline methods are compiled which is counter-intuitive
-template <typename T>
-struct ConstSessionOptionsImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  SessionOptions Clone() const;  ///< Creates and returns a copy of this SessionOptions object. Wraps OrtApi::CloneSessionOptions
-
-  std::string GetConfigEntry(const char* config_key) const;  ///< Wraps OrtApi::GetSessionConfigEntry
-  bool HasConfigEntry(const char* config_key) const;         ///< Wraps OrtApi::HasSessionConfigEntry
-  std::string GetConfigEntryOrDefault(const char* config_key, const std::string& def);
-};
-
-template <typename T>
-struct SessionOptionsImpl : ConstSessionOptionsImpl<T> {
-  using B = ConstSessionOptionsImpl<T>;
-  using B::B;
-
-  SessionOptionsImpl& SetIntraOpNumThreads(int intra_op_num_threads);                              ///< Wraps OrtApi::SetIntraOpNumThreads
-  SessionOptionsImpl& SetInterOpNumThreads(int inter_op_num_threads);                              ///< Wraps OrtApi::SetInterOpNumThreads
-  SessionOptionsImpl& SetGraphOptimizationLevel(GraphOptimizationLevel graph_optimization_level);  ///< Wraps OrtApi::SetSessionGraphOptimizationLevel
-
-  SessionOptionsImpl& EnableCpuMemArena();   ///< Wraps OrtApi::EnableCpuMemArena
-  SessionOptionsImpl& DisableCpuMemArena();  ///< Wraps OrtApi::DisableCpuMemArena
-
-  SessionOptionsImpl& SetOptimizedModelFilePath(const ORTCHAR_T* optimized_model_file);  ///< Wraps OrtApi::SetOptimizedModelFilePath
-
-  SessionOptionsImpl& EnableProfiling(const ORTCHAR_T* profile_file_prefix);  ///< Wraps OrtApi::EnableProfiling
-  SessionOptionsImpl& DisableProfiling();                                     ///< Wraps OrtApi::DisableProfiling
-
-  SessionOptionsImpl& EnableOrtCustomOps();  ///< Wraps OrtApi::EnableOrtCustomOps
-
-  SessionOptionsImpl& EnableMemPattern();   ///< Wraps OrtApi::EnableMemPattern
-  SessionOptionsImpl& DisableMemPattern();  ///< Wraps OrtApi::DisableMemPattern
-
-  SessionOptionsImpl& SetExecutionMode(ExecutionMode execution_mode);  ///< Wraps OrtApi::SetSessionExecutionMode
-
-  SessionOptionsImpl& SetLogId(const char* logid);     ///< Wraps OrtApi::SetSessionLogId
-  SessionOptionsImpl& SetLogSeverityLevel(int level);  ///< Wraps OrtApi::SetSessionLogSeverityLevel
-
-  SessionOptionsImpl& Add(OrtCustomOpDomain* custom_op_domain);  ///< Wraps OrtApi::AddCustomOpDomain
-
-  SessionOptionsImpl& DisablePerSessionThreads();  ///< Wraps OrtApi::DisablePerSessionThreads
-
-  SessionOptionsImpl& AddConfigEntry(const char* config_key, const char* config_value);                        ///< Wraps OrtApi::AddSessionConfigEntry
-
-  SessionOptionsImpl& AddInitializer(const char* name, const OrtValue* ort_val);                                             ///< Wraps OrtApi::AddInitializer
-  SessionOptionsImpl& AddExternalInitializers(const std::vector<std::string>& names, const std::vector<Value>& ort_values);  ///< Wraps OrtApi::AddExternalInitializers
-
-  SessionOptionsImpl& AppendExecutionProvider_CUDA(const OrtCUDAProviderOptions& provider_options);               ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_CUDA
-  SessionOptionsImpl& AppendExecutionProvider_CUDA_V2(const OrtCUDAProviderOptionsV2& provider_options);          ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_CUDA_V2
-  SessionOptionsImpl& AppendExecutionProvider_ROCM(const OrtROCMProviderOptions& provider_options);               ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_ROCM
-  SessionOptionsImpl& AppendExecutionProvider_OpenVINO(const OrtOpenVINOProviderOptions& provider_options);       ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_OpenVINO
-  SessionOptionsImpl& AppendExecutionProvider_TensorRT(const OrtTensorRTProviderOptions& provider_options);       ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_TensorRT
-  SessionOptionsImpl& AppendExecutionProvider_TensorRT_V2(const OrtTensorRTProviderOptionsV2& provider_options);  ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_TensorRT
-  SessionOptionsImpl& AppendExecutionProvider_MIGraphX(const OrtMIGraphXProviderOptions& provider_options);       ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_MIGraphX
-  ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_CANN
-  SessionOptionsImpl& AppendExecutionProvider_CANN(const OrtCANNProviderOptions& provider_options);
-  /// Wraps OrtApi::SessionOptionsAppendExecutionProvider. Currently supports SNPE and XNNPACK.
-  SessionOptionsImpl& AppendExecutionProvider(const std::string& provider_name,
-                                              const std::unordered_map<std::string, std::string>& provider_options = {});
-
-  SessionOptionsImpl& SetCustomCreateThreadFn(OrtCustomCreateThreadFn ort_custom_create_thread_fn);  ///< Wraps OrtApi::SessionOptionsSetCustomCreateThreadFn
-  SessionOptionsImpl& SetCustomThreadCreationOptions(void* ort_custom_thread_creation_options);      ///< Wraps OrtApi::SessionOptionsSetCustomThreadCreationOptions
-  SessionOptionsImpl& SetCustomJoinThreadFn(OrtCustomJoinThreadFn ort_custom_join_thread_fn);        ///< Wraps OrtApi::SessionOptionsSetCustomJoinThreadFn
-
-  ///< Registers the custom operator from the specified shared library via OrtApi::RegisterCustomOpsLibrary_V2.
-  ///< The custom operator configurations are optional. If provided, custom operator configs are set via
-  ///< OrtApi::AddSessionConfigEntry.
-  SessionOptionsImpl& RegisterCustomOpsLibrary(const ORTCHAR_T* library_name, const CustomOpConfigs& custom_op_configs = {});
-
-  SessionOptionsImpl& RegisterCustomOpsUsingFunction(const char* function_name);  ///< Wraps OrtApi::RegisterCustomOpsUsingFunction
-};
-}  // namespace detail
-
-using UnownedSessionOptions = detail::SessionOptionsImpl<detail::Unowned<OrtSessionOptions>>;
-using ConstSessionOptions = detail::ConstSessionOptionsImpl<detail::Unowned<const OrtSessionOptions>>;
-
-/** \brief Wrapper around ::OrtSessionOptions
- *
- */
-struct SessionOptions : detail::SessionOptionsImpl<OrtSessionOptions> {
-  explicit SessionOptions(std::nullptr_t) {}                                                   ///< Create an empty SessionOptions object, must be assigned a valid one to be used
-  SessionOptions();                                                                            ///< Wraps OrtApi::CreateSessionOptions
-  explicit SessionOptions(OrtSessionOptions* p) : SessionOptionsImpl<OrtSessionOptions>{p} {}  ///< Used for interop with the C API
-  UnownedSessionOptions GetUnowned() const { return UnownedSessionOptions{this->p_}; }
-  ConstSessionOptions GetConst() const { return ConstSessionOptions{this->p_}; }
-};
-
-/** \brief Wrapper around ::OrtModelMetadata
- *
- */
-struct ModelMetadata : detail::Base<OrtModelMetadata> {
-  explicit ModelMetadata(std::nullptr_t) {}                                   ///< Create an empty ModelMetadata object, must be assigned a valid one to be used
-  explicit ModelMetadata(OrtModelMetadata* p) : Base<OrtModelMetadata>{p} {}  ///< Used for interop with the C API
-
-  /** \brief Returns a copy of the producer name.
-   *
-   * \param allocator to allocate memory for the copy of the name returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetProducerNameAllocated(OrtAllocator* allocator) const;  ///< Wraps OrtApi::ModelMetadataGetProducerName
-
-  /** \brief Returns a copy of the graph name.
-   *
-   * \param allocator to allocate memory for the copy of the name returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetGraphNameAllocated(OrtAllocator* allocator) const;  ///< Wraps OrtApi::ModelMetadataGetGraphName
-
-  /** \brief Returns a copy of the domain name.
-   *
-   * \param allocator to allocate memory for the copy of the name returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetDomainAllocated(OrtAllocator* allocator) const;  ///< Wraps OrtApi::ModelMetadataGetDomain
-
-  /** \brief Returns a copy of the description.
-   *
-   * \param allocator to allocate memory for the copy of the string returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetDescriptionAllocated(OrtAllocator* allocator) const;  ///< Wraps OrtApi::ModelMetadataGetDescription
-
-  /** \brief Returns a copy of the graph description.
-   *
-   * \param allocator to allocate memory for the copy of the string returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetGraphDescriptionAllocated(OrtAllocator* allocator) const;  ///< Wraps OrtApi::ModelMetadataGetGraphDescription
-
-  /** \brief Returns a vector of copies of the custom metadata keys.
-   *
-   * \param allocator to allocate memory for the copy of the string returned
-   * \return a instance std::vector of smart pointers that would deallocate the buffers when out of scope.
-   *  The OrtAllocator instance must be valid at the point of memory release.
-   */
-  std::vector<AllocatedStringPtr> GetCustomMetadataMapKeysAllocated(OrtAllocator* allocator) const;  ///< Wraps OrtApi::ModelMetadataGetCustomMetadataMapKeys
-
-  /** \brief Looks up a value by a key in the Custom Metadata map
-   *
-   * \param key zero terminated string key to lookup
-   * \param allocator to allocate memory for the copy of the string returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  maybe nullptr if key is not found.
-   *
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr LookupCustomMetadataMapAllocated(const char* key, OrtAllocator* allocator) const;  ///< Wraps OrtApi::ModelMetadataLookupCustomMetadataMap
-
-  int64_t GetVersion() const;  ///< Wraps OrtApi::ModelMetadataGetVersion
-};
-
-struct IoBinding;
-
-namespace detail {
-
-// we separate const-only methods because passing const ptr to non-const methods
-// is only discovered when inline methods are compiled which is counter-intuitive
-template <typename T>
-struct ConstSessionImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  size_t GetInputCount() const;                   ///< Returns the number of model inputs
-  size_t GetOutputCount() const;                  ///< Returns the number of model outputs
-  size_t GetOverridableInitializerCount() const;  ///< Returns the number of inputs that have defaults that can be overridden
-
-  /** \brief Returns a copy of input name at the specified index.
-   *
-   * \param index must less than the value returned by GetInputCount()
-   * \param allocator to allocate memory for the copy of the name returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetInputNameAllocated(size_t index, OrtAllocator* allocator) const;
-
-  /** \brief Returns a copy of output name at then specified index.
-   *
-   * \param index must less than the value returned by GetOutputCount()
-   * \param allocator to allocate memory for the copy of the name returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetOutputNameAllocated(size_t index, OrtAllocator* allocator) const;
-
-  /** \brief Returns a copy of the overridable initializer name at then specified index.
-   *
-   * \param index must less than the value returned by GetOverridableInitializerCount()
-   * \param allocator to allocate memory for the copy of the name returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetOverridableInitializerNameAllocated(size_t index, OrtAllocator* allocator) const;  ///< Wraps OrtApi::SessionGetOverridableInitializerName
-
-  uint64_t GetProfilingStartTimeNs() const;  ///< Wraps OrtApi::SessionGetProfilingStartTimeNs
-  ModelMetadata GetModelMetadata() const;    ///< Wraps OrtApi::SessionGetModelMetadata
-
-  TypeInfo GetInputTypeInfo(size_t index) const;                   ///< Wraps OrtApi::SessionGetInputTypeInfo
-  TypeInfo GetOutputTypeInfo(size_t index) const;                  ///< Wraps OrtApi::SessionGetOutputTypeInfo
-  TypeInfo GetOverridableInitializerTypeInfo(size_t index) const;  ///< Wraps OrtApi::SessionGetOverridableInitializerTypeInfo
-};
-
-template <typename T>
-struct SessionImpl : ConstSessionImpl<T> {
-  using B = ConstSessionImpl<T>;
-  using B::B;
-
-  /** \brief Run the model returning results in an Ort allocated vector.
-   *
-   * Wraps OrtApi::Run
-   *
-   * The caller provides a list of inputs and a list of the desired outputs to return.
-   *
-   * See the output logs for more information on warnings/errors that occur while processing the model.
-   * Common errors are.. (TODO)
-   *
-   * \param[in] run_options
-   * \param[in] input_names Array of null terminated strings of length input_count that is the list of input names
-   * \param[in] input_values Array of Value objects of length input_count that is the list of input values
-   * \param[in] input_count Number of inputs (the size of the input_names & input_values arrays)
-   * \param[in] output_names Array of C style strings of length output_count that is the list of output names
-   * \param[in] output_count Number of outputs (the size of the output_names array)
-   * \return A std::vector of Value objects that directly maps to the output_names array (eg. output_name[0] is the first entry of the returned vector)
-   */
-  std::vector<Value> Run(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count,
-                         const char* const* output_names, size_t output_count);
-
-  /** \brief Run the model returning results in user provided outputs
-   * Same as Run(const RunOptions&, const char* const*, const Value*, size_t,const char* const*, size_t)
-   */
-  void Run(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count,
-           const char* const* output_names, Value* output_values, size_t output_count);
-
-  void Run(const RunOptions& run_options, const IoBinding&);  ///< Wraps OrtApi::RunWithBinding
-
-  /** \brief End profiling and return a copy of the profiling file name.
-   *
-   * \param allocator to allocate memory for the copy of the string returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr EndProfilingAllocated(OrtAllocator* allocator);  ///< Wraps OrtApi::SessionEndProfiling
-};
-
-}  // namespace detail
-
-using ConstSession = detail::ConstSessionImpl<detail::Unowned<const OrtSession>>;
-using UnownedSession = detail::SessionImpl<detail::Unowned<OrtSession>>;
-
-/** \brief Wrapper around ::OrtSession
- *
- */
-struct Session : detail::SessionImpl<OrtSession> {
-  explicit Session(std::nullptr_t) {}                                                   ///< Create an empty Session object, must be assigned a valid one to be used
-  Session(const Env& env, const ORTCHAR_T* model_path, const SessionOptions& options);  ///< Wraps OrtApi::CreateSession
-  Session(const Env& env, const ORTCHAR_T* model_path, const SessionOptions& options,
-          OrtPrepackedWeightsContainer* prepacked_weights_container);                                        ///< Wraps OrtApi::CreateSessionWithPrepackedWeightsContainer
-  Session(const Env& env, const void* model_data, size_t model_data_length, const SessionOptions& options);  ///< Wraps OrtApi::CreateSessionFromArray
-  Session(const Env& env, const void* model_data, size_t model_data_length, const SessionOptions& options,
-          OrtPrepackedWeightsContainer* prepacked_weights_container);  ///< Wraps OrtApi::CreateSessionFromArrayWithPrepackedWeightsContainer
-
-  ConstSession GetConst() const { return ConstSession{this->p_}; }
-  UnownedSession GetUnowned() const { return UnownedSession{this->p_}; }
-};
-
-namespace detail {
-template <typename T>
-struct MemoryInfoImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  std::string GetAllocatorName() const;
-  OrtAllocatorType GetAllocatorType() const;
-  int GetDeviceId() const;
-  OrtMemoryInfoDeviceType GetDeviceType() const;
-  OrtMemType GetMemoryType() const;
-
-  template <typename U>
-  bool operator==(const MemoryInfoImpl<U>& o) const;
-};
-}  // namespace detail
-
-// Const object holder that does not own the underlying object
-using ConstMemoryInfo = detail::MemoryInfoImpl<detail::Unowned<const OrtMemoryInfo>>;
-
-/** \brief Wrapper around ::OrtMemoryInfo
- *
- */
-struct MemoryInfo : detail::MemoryInfoImpl<OrtMemoryInfo> {
-  static MemoryInfo CreateCpu(OrtAllocatorType type, OrtMemType mem_type1);
-  explicit MemoryInfo(std::nullptr_t) {}                                       ///< No instance is created
-  explicit MemoryInfo(OrtMemoryInfo* p) : MemoryInfoImpl<OrtMemoryInfo>{p} {}  ///< Take ownership of a pointer created by C Api
-  MemoryInfo(const char* name, OrtAllocatorType type, int id, OrtMemType mem_type);
-  ConstMemoryInfo GetConst() const { return ConstMemoryInfo{this->p_}; }
-};
-
-namespace detail {
-template <typename T>
-struct TensorTypeAndShapeInfoImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  ONNXTensorElementDataType GetElementType() const;  ///< Wraps OrtApi::GetTensorElementType
-  size_t GetElementCount() const;                    ///< Wraps OrtApi::GetTensorShapeElementCount
-
-  size_t GetDimensionsCount() const;  ///< Wraps OrtApi::GetDimensionsCount
-
-  /** \deprecated use GetShape() returning std::vector
-   * [[deprecated]]
-   * This interface is unsafe to use
-   */
-  [[deprecated("use GetShape()")]] void GetDimensions(int64_t* values, size_t values_count) const;  ///< Wraps OrtApi::GetDimensions
-
-  void GetSymbolicDimensions(const char** values, size_t values_count) const;  ///< Wraps OrtApi::GetSymbolicDimensions
-
-  std::vector<int64_t> GetShape() const;  ///< Uses GetDimensionsCount & GetDimensions to return a std::vector of the shape
-};
-
-}  // namespace detail
-
-using ConstTensorTypeAndShapeInfo = detail::TensorTypeAndShapeInfoImpl<detail::Unowned<const OrtTensorTypeAndShapeInfo>>;
-
-/** \brief Wrapper around ::OrtTensorTypeAndShapeInfo
- *
- */
-struct TensorTypeAndShapeInfo : detail::TensorTypeAndShapeInfoImpl<OrtTensorTypeAndShapeInfo> {
-  explicit TensorTypeAndShapeInfo(std::nullptr_t) {}                                                ///< Create an empty TensorTypeAndShapeInfo object, must be assigned a valid one to be used
-  explicit TensorTypeAndShapeInfo(OrtTensorTypeAndShapeInfo* p) : TensorTypeAndShapeInfoImpl{p} {}  ///< Used for interop with the C API
-  ConstTensorTypeAndShapeInfo GetConst() const { return ConstTensorTypeAndShapeInfo{this->p_}; }
-};
-
-namespace detail {
-template <typename T>
-struct SequenceTypeInfoImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-  TypeInfo GetSequenceElementType() const;  ///< Wraps OrtApi::GetSequenceElementType
-};
-
-}  // namespace detail
-
-using ConstSequenceTypeInfo = detail::SequenceTypeInfoImpl<detail::Unowned<const OrtSequenceTypeInfo>>;
-
-/** \brief Wrapper around ::OrtSequenceTypeInfo
- *
- */
-struct SequenceTypeInfo : detail::SequenceTypeInfoImpl<OrtSequenceTypeInfo> {
-  explicit SequenceTypeInfo(std::nullptr_t) {}                                                         ///< Create an empty SequenceTypeInfo object, must be assigned a valid one to be used
-  explicit SequenceTypeInfo(OrtSequenceTypeInfo* p) : SequenceTypeInfoImpl<OrtSequenceTypeInfo>{p} {}  ///< Used for interop with the C API
-  ConstSequenceTypeInfo GetConst() const { return ConstSequenceTypeInfo{this->p_}; }
-};
-
-namespace detail {
-template <typename T>
-struct MapTypeInfoImpl : detail::Base<T> {
-  using B = Base<T>;
-  using B::B;
-  ONNXTensorElementDataType GetMapKeyType() const;  ///< Wraps OrtApi::GetMapKeyType
-  TypeInfo GetMapValueType() const;                 ///< Wraps OrtApi::GetMapValueType
-};
-
-}  // namespace detail
-
-using ConstMapTypeInfo = detail::MapTypeInfoImpl<detail::Unowned<const OrtMapTypeInfo>>;
-
-/** \brief Wrapper around ::OrtMapTypeInfo
- *
- */
-struct MapTypeInfo : detail::MapTypeInfoImpl<OrtMapTypeInfo> {
-  explicit MapTypeInfo(std::nullptr_t) {}                                          ///< Create an empty MapTypeInfo object, must be assigned a valid one to be used
-  explicit MapTypeInfo(OrtMapTypeInfo* p) : MapTypeInfoImpl<OrtMapTypeInfo>{p} {}  ///< Used for interop with the C API
-  ConstMapTypeInfo GetConst() const { return ConstMapTypeInfo{this->p_}; }
-};
-
-namespace detail {
-template <typename T>
-struct TypeInfoImpl : detail::Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  ConstTensorTypeAndShapeInfo GetTensorTypeAndShapeInfo() const;  ///< Wraps OrtApi::CastTypeInfoToTensorInfo
-  ConstSequenceTypeInfo GetSequenceTypeInfo() const;              ///< Wraps OrtApi::CastTypeInfoToSequenceTypeInfo
-  ConstMapTypeInfo GetMapTypeInfo() const;                        ///< Wraps OrtApi::CastTypeInfoToMapTypeInfo
-
-  ONNXType GetONNXType() const;
-};
-}  // namespace detail
-
-/// <summary>
-/// Contains a constant, unowned OrtTypeInfo that can be copied and passed around by value.
-/// Provides access to const OrtTypeInfo APIs.
-/// </summary>
-using ConstTypeInfo = detail::TypeInfoImpl<detail::Unowned<const OrtTypeInfo>>;
-
-/// <summary>
-/// Type information that may contain either TensorTypeAndShapeInfo or
-/// the information about contained sequence or map depending on the ONNXType.
-/// </summary>
-struct TypeInfo : detail::TypeInfoImpl<OrtTypeInfo> {
-  explicit TypeInfo(std::nullptr_t) {}                                 ///< Create an empty TypeInfo object, must be assigned a valid one to be used
-  explicit TypeInfo(OrtTypeInfo* p) : TypeInfoImpl<OrtTypeInfo>{p} {}  ///< C API Interop
-
-  ConstTypeInfo GetConst() const { return ConstTypeInfo{this->p_}; }
-};
-
-namespace detail {
-// This structure is used to feed  sparse tensor values
-// information for use with FillSparseTensor<Format>() API
-// if the data type for the sparse tensor values is numeric
-// use data.p_data, otherwise, use data.str pointer to feed
-// values. data.str is an array of const char* that are zero terminated.
-// number of strings in the array must match shape size.
-// For fully sparse tensors use shape {0} and set p_data/str
-// to nullptr.
-struct OrtSparseValuesParam {
-  const int64_t* values_shape;
-  size_t values_shape_len;
-  union {
-    const void* p_data;
-    const char** str;
-  } data;
-};
-
-// Provides a way to pass shape in a single
-// argument
-struct Shape {
-  const int64_t* shape;
-  size_t shape_len;
-};
-
-template <typename T>
-struct ConstValueImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  /// <summary>
-  /// Obtains a pointer to a user defined data for experimental purposes
-  /// </summary>
-  template <typename R>
-  void GetOpaqueData(const char* domain, const char* type_name, R&) const;  ///< Wraps OrtApi::GetOpaqueValue
-
-  bool IsTensor() const;  ///< Returns true if Value is a tensor, false for other types like map/sequence/etc
-  bool HasValue() const;  /// < Return true if OrtValue contains data and returns false if the OrtValue is a None
-
-  size_t GetCount() const;  // If a non tensor, returns 2 for map and N for sequence, where N is the number of elements
-  Value GetValue(int index, OrtAllocator* allocator) const;
-
-  /// <summary>
-  /// This API returns a full length of string data contained within either a tensor or a sparse Tensor.
-  /// For sparse tensor it returns a full length of stored non-empty strings (values). The API is useful
-  /// for allocating necessary memory and calling GetStringTensorContent().
-  /// </summary>
-  /// <returns>total length of UTF-8 encoded bytes contained. No zero terminators counted.</returns>
-  size_t GetStringTensorDataLength() const;
-
-  /// <summary>
-  /// The API copies all of the UTF-8 encoded string data contained within a tensor or a sparse tensor
-  /// into a supplied buffer. Use GetStringTensorDataLength() to find out the length of the buffer to allocate.
-  /// The user must also allocate offsets buffer with the number of entries equal to that of the contained
-  /// strings.
-  ///
-  /// Strings are always assumed to be on CPU, no X-device copy.
-  /// </summary>
-  /// <param name="buffer">user allocated buffer</param>
-  /// <param name="buffer_length">length in bytes of the allocated buffer</param>
-  /// <param name="offsets">a pointer to the offsets user allocated buffer</param>
-  /// <param name="offsets_count">count of offsets, must be equal to the number of strings contained.
-  ///   that can be obtained from the shape of the tensor or from GetSparseTensorValuesTypeAndShapeInfo()
-  ///   for sparse tensors</param>
-  void GetStringTensorContent(void* buffer, size_t buffer_length, size_t* offsets, size_t offsets_count) const;
-
-  /// <summary>
-  /// Returns a const typed pointer to the tensor contained data.
-  /// No type checking is performed, the caller must ensure the type matches the tensor type.
-  /// </summary>
-  /// <typeparam name="T"></typeparam>
-  /// <returns>const pointer to data, no copies made</returns>
-  template <typename R>
-  const R* GetTensorData() const;  ///< Wraps OrtApi::GetTensorMutableData   /// <summary>
-
-  /// <summary>
-  /// Returns a non-typed pointer to a tensor contained data.
-  /// </summary>
-  /// <returns>const pointer to data, no copies made</returns>
-  const void* GetTensorRawData() const;
-
-  /// <summary>
-  /// The API returns type information for data contained in a tensor. For sparse
-  /// tensors it returns type information for contained non-zero values.
-  /// It returns dense shape for sparse tensors.
-  /// </summary>
-  /// <returns>TypeInfo</returns>
-  TypeInfo GetTypeInfo() const;
-
-  /// <summary>
-  /// The API returns type information for data contained in a tensor. For sparse
-  /// tensors it returns type information for contained non-zero values.
-  /// It returns dense shape for sparse tensors.
-  /// </summary>
-  /// <returns>TensorTypeAndShapeInfo</returns>
-  TensorTypeAndShapeInfo GetTensorTypeAndShapeInfo() const;
-
-  /// <summary>
-  /// This API returns information about the memory allocation used to hold data.
-  /// </summary>
-  /// <returns>Non owning instance of MemoryInfo</returns>
-  ConstMemoryInfo GetTensorMemoryInfo() const;
-
-  /// <summary>
-  /// The API copies UTF-8 encoded bytes for the requested string element
-  /// contained within a tensor or a sparse tensor into a provided buffer.
-  /// Use GetStringTensorElementLength() to obtain the length of the buffer to allocate.
-  /// </summary>
-  /// <param name="buffer_length"></param>
-  /// <param name="element_index"></param>
-  /// <param name="buffer"></param>
-  void GetStringTensorElement(size_t buffer_length, size_t element_index, void* buffer) const;
-
-  /// <summary>
-  /// The API returns a byte length of UTF-8 encoded string element
-  /// contained in either a tensor or a spare tensor values.
-  /// </summary>
-  /// <param name="element_index"></param>
-  /// <returns>byte length for the specified string element</returns>
-  size_t GetStringTensorElementLength(size_t element_index) const;
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-  /// <summary>
-  /// The API returns the sparse data format this OrtValue holds in a sparse tensor.
-  /// If the sparse tensor was not fully constructed, i.e. Use*() or Fill*() API were not used
-  /// the value returned is ORT_SPARSE_UNDEFINED.
-  /// </summary>
-  /// <returns>Format enum</returns>
-  OrtSparseFormat GetSparseFormat() const;
-
-  /// <summary>
-  /// The API returns type and shape information for stored non-zero values of the
-  /// sparse tensor. Use GetSparseTensorValues() to obtain values buffer pointer.
-  /// </summary>
-  /// <returns>TensorTypeAndShapeInfo values information</returns>
-  TensorTypeAndShapeInfo GetSparseTensorValuesTypeAndShapeInfo() const;
-
-  /// <summary>
-  /// The API returns type and shape information for the specified indices. Each supported
-  /// indices have their own enum values even if a give format has more than one kind of indices.
-  /// Use GetSparseTensorIndicesData() to obtain pointer to indices buffer.
-  /// </summary>
-  /// <param name="format">enum requested</param>
-  /// <returns>type and shape information</returns>
-  TensorTypeAndShapeInfo GetSparseTensorIndicesTypeShapeInfo(OrtSparseIndicesFormat format) const;
-
-  /// <summary>
-  /// The API retrieves a pointer to the internal indices buffer. The API merely performs
-  /// a convenience data type casting on the return type pointer. Make sure you are requesting
-  /// the right type, use GetSparseTensorIndicesTypeShapeInfo();
-  /// </summary>
-  /// <typeparam name="T">type to cast to</typeparam>
-  /// <param name="indices_format">requested indices kind</param>
-  /// <param name="num_indices">number of indices entries</param>
-  /// <returns>Pinter to the internal sparse tensor buffer containing indices. Do not free this pointer.</returns>
-  template <typename R>
-  const R* GetSparseTensorIndicesData(OrtSparseIndicesFormat indices_format, size_t& num_indices) const;
-
-  /// <summary>
-  /// Returns true if the OrtValue contains a sparse tensor
-  /// </summary>
-  /// <returns></returns>
-  bool IsSparseTensor() const;
-
-  /// <summary>
-  /// The API returns a pointer to an internal buffer of the sparse tensor
-  /// containing non-zero values. The API merely does casting. Make sure you
-  /// are requesting the right data type by calling GetSparseTensorValuesTypeAndShapeInfo()
-  /// first.
-  /// </summary>
-  /// <typeparam name="T">numeric data types only. Use GetStringTensor*() to retrieve strings.</typeparam>
-  /// <returns>a pointer to the internal values buffer. Do not free this pointer.</returns>
-  template <typename R>
-  const R* GetSparseTensorValues() const;
-
-#endif
-};
-
-template <typename T>
-struct ValueImpl : ConstValueImpl<T> {
-  using B = ConstValueImpl<T>;
-  using B::B;
-
-  /// <summary>
-  /// Returns a non-const typed pointer to an OrtValue/Tensor contained buffer
-  /// No type checking is performed, the caller must ensure the type matches the tensor type.
-  /// </summary>
-  /// <returns>non-const pointer to data, no copies made</returns>
-  template <typename R>
-  R* GetTensorMutableData();
-
-  /// <summary>
-  /// Returns a non-typed non-const pointer to a tensor contained data.
-  /// </summary>
-  /// <returns>pointer to data, no copies made</returns>
-  void* GetTensorMutableRawData();
-
-  /// <summary>
-  //  Obtain a reference to an element of data at the location specified
-  /// by the vector of dims.
-  /// </summary>
-  /// <typeparam name="R"></typeparam>
-  /// <param name="location">[in] expressed by a vecotr of dimensions offsets</param>
-  /// <returns></returns>
-  template <typename R>
-  R& At(const std::vector<int64_t>& location);
-
-  /// <summary>
-  /// Set all strings at once in a string tensor
-  /// </summary>
-  /// <param name="s">[in] An array of strings. Each string in this array must be null terminated.</param>
-  /// <param name="s_len">[in] Count of strings in s (Must match the size of \p value's tensor shape)</param>
-  void FillStringTensor(const char* const* s, size_t s_len);
-
-  /// <summary>
-  /// Set a single string in a string tensor
-  /// </summary>
-  /// <param name="s">[in] A null terminated UTF-8 encoded string</param>
-  /// <param name="index">[in] Index of the string in the tensor to set</param>
-  void FillStringTensorElement(const char* s, size_t index);
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-  /// <summary>
-  /// Supplies COO format specific indices and marks the contained sparse tensor as being a COO format tensor.
-  /// Values are supplied with a CreateSparseTensor() API. The supplied indices are not copied and the user
-  /// allocated buffers lifespan must eclipse that of the OrtValue.
-  /// The location of the indices is assumed to be the same as specified by OrtMemoryInfo argument at the creation time.
-  /// </summary>
-  /// <param name="indices_data">pointer to the user allocated buffer with indices. Use nullptr for fully sparse tensors.</param>
-  /// <param name="indices_num">number of indices entries. Use 0 for fully sparse tensors</param>
-  void UseCooIndices(int64_t* indices_data, size_t indices_num);
-
-  /// <summary>
-  /// Supplies CSR format specific indices and marks the contained sparse tensor as being a CSR format tensor.
-  /// Values are supplied with a CreateSparseTensor() API. The supplied indices are not copied and the user
-  /// allocated buffers lifespan must eclipse that of the OrtValue.
-  /// The location of the indices is assumed to be the same as specified by OrtMemoryInfo argument at the creation time.
-  /// </summary>
-  /// <param name="inner_data">pointer to the user allocated buffer with inner indices or nullptr for fully sparse tensors</param>
-  /// <param name="inner_num">number of csr inner indices or 0 for fully sparse tensors</param>
-  /// <param name="outer_data">pointer to the user allocated buffer with outer indices or nullptr for fully sparse tensors</param>
-  /// <param name="outer_num">number of csr outer indices or 0 for fully sparse tensors</param>
-  void UseCsrIndices(int64_t* inner_data, size_t inner_num, int64_t* outer_data, size_t outer_num);
-
-  /// <summary>
-  /// Supplies BlockSparse format specific indices and marks the contained sparse tensor as being a BlockSparse format tensor.
-  /// Values are supplied with a CreateSparseTensor() API. The supplied indices are not copied and the user
-  /// allocated buffers lifespan must eclipse that of the OrtValue.
-  /// The location of the indices is assumed to be the same as specified by OrtMemoryInfo argument at the creation time.
-  /// </summary>
-  /// <param name="indices_shape">indices shape or a {0} for fully sparse</param>
-  /// <param name="indices_data">user allocated buffer with indices or nullptr for fully spare tensors</param>
-  void UseBlockSparseIndices(const Shape& indices_shape, int32_t* indices_data);
-
-  /// <summary>
-  /// The API will allocate memory using the allocator instance supplied to the CreateSparseTensor() API
-  /// and copy the values and COO indices into it. If data_mem_info specifies that the data is located
-  /// at difference device than the allocator, a X-device copy will be performed if possible.
-  /// </summary>
-  /// <param name="data_mem_info">specified buffer memory description</param>
-  /// <param name="values_param">values buffer information.</param>
-  /// <param name="indices_data">coo indices buffer or nullptr for fully sparse data</param>
-  /// <param name="indices_num">number of COO indices or 0 for fully sparse data</param>
-  void FillSparseTensorCoo(const OrtMemoryInfo* data_mem_info, const OrtSparseValuesParam& values_param,
-                           const int64_t* indices_data, size_t indices_num);
-
-  /// <summary>
-  /// The API will allocate memory using the allocator instance supplied to the CreateSparseTensor() API
-  /// and copy the values and CSR indices into it. If data_mem_info specifies that the data is located
-  /// at difference device than the allocator, a X-device copy will be performed if possible.
-  /// </summary>
-  /// <param name="data_mem_info">specified buffer memory description</param>
-  /// <param name="values">values buffer information</param>
-  /// <param name="inner_indices_data">csr inner indices pointer or nullptr for fully sparse tensors</param>
-  /// <param name="inner_indices_num">number of csr inner indices or 0 for fully sparse tensors</param>
-  /// <param name="outer_indices_data">pointer to csr indices data or nullptr for fully sparse tensors</param>
-  /// <param name="outer_indices_num">number of csr outer indices or 0</param>
-  void FillSparseTensorCsr(const OrtMemoryInfo* data_mem_info,
-                           const OrtSparseValuesParam& values,
-                           const int64_t* inner_indices_data, size_t inner_indices_num,
-                           const int64_t* outer_indices_data, size_t outer_indices_num);
-
-  /// <summary>
-  /// The API will allocate memory using the allocator instance supplied to the CreateSparseTensor() API
-  /// and copy the values and BlockSparse indices into it. If data_mem_info specifies that the data is located
-  /// at difference device than the allocator, a X-device copy will be performed if possible.
-  /// </summary>
-  /// <param name="data_mem_info">specified buffer memory description</param>
-  /// <param name="values">values buffer information</param>
-  /// <param name="indices_shape">indices shape. use {0} for fully sparse tensors</param>
-  /// <param name="indices_data">pointer to indices data or nullptr for fully sparse tensors</param>
-  void FillSparseTensorBlockSparse(const OrtMemoryInfo* data_mem_info,
-                                   const OrtSparseValuesParam& values,
-                                   const Shape& indices_shape,
-                                   const int32_t* indices_data);
-
-#endif
-};
-
-}  // namespace detail
-
-using ConstValue = detail::ConstValueImpl<detail::Unowned<const OrtValue>>;
-using UnownedValue = detail::ValueImpl<detail::Unowned<OrtValue>>;
-
-/** \brief Wrapper around ::OrtValue
- *
- */
-struct Value : detail::ValueImpl<OrtValue> {
-  using Base = detail::ValueImpl<OrtValue>;
-  using OrtSparseValuesParam = detail::OrtSparseValuesParam;
-  using Shape = detail::Shape;
-
-  explicit Value(std::nullptr_t) {}         ///< Create an empty Value object, must be assigned a valid one to be used
-  explicit Value(OrtValue* p) : Base{p} {}  ///< Used for interop with the C API
-  Value(Value&&) = default;
-  Value& operator=(Value&&) = default;
-
-  ConstValue GetConst() const { return ConstValue{this->p_}; }
-  UnownedValue GetUnowned() const { return UnownedValue{this->p_}; }
-
-  /** \brief Creates a tensor with a user supplied buffer. Wraps OrtApi::CreateTensorWithDataAsOrtValue.
-   * \tparam T The numeric datatype. This API is not suitable for strings.
-   * \param info Memory description of where the p_data buffer resides (CPU vs GPU etc).
-   * \param p_data Pointer to the data buffer.
-   * \param p_data_element_count The number of elements in the data buffer.
-   * \param shape Pointer to the tensor shape dimensions.
-   * \param shape_len The number of tensor shape dimensions.
-   */
-  template <typename T>
-  static Value CreateTensor(const OrtMemoryInfo* info, T* p_data, size_t p_data_element_count, const int64_t* shape, size_t shape_len);
-
-  /** \brief Creates a tensor with a user supplied buffer. Wraps OrtApi::CreateTensorWithDataAsOrtValue.
-   * \param info Memory description of where the p_data buffer resides (CPU vs GPU etc).
-   * \param p_data Pointer to the data buffer.
-   * \param p_data_byte_count The number of bytes in the data buffer.
-   * \param shape Pointer to the tensor shape dimensions.
-   * \param shape_len The number of tensor shape dimensions.
-   * \param type The data type.
-   */
-  static Value CreateTensor(const OrtMemoryInfo* info, void* p_data, size_t p_data_byte_count, const int64_t* shape, size_t shape_len,
-                            ONNXTensorElementDataType type);
-
-  /** \brief Creates a tensor using a supplied OrtAllocator. Wraps OrtApi::CreateTensorAsOrtValue.
-   * \tparam T The numeric datatype. This API is not suitable for strings.
-   * \param allocator The allocator to use.
-   * \param shape Pointer to the tensor shape dimensions.
-   * \param shape_len The number of tensor shape dimensions.
-   */
-  template <typename T>
-  static Value CreateTensor(OrtAllocator* allocator, const int64_t* shape, size_t shape_len);
-
-  /** \brief Creates a tensor using a supplied OrtAllocator. Wraps OrtApi::CreateTensorAsOrtValue.
-   * \param allocator The allocator to use.
-   * \param shape Pointer to the tensor shape dimensions.
-   * \param shape_len The number of tensor shape dimensions.
-   * \param type The data type.
-   */
-  static Value CreateTensor(OrtAllocator* allocator, const int64_t* shape, size_t shape_len, ONNXTensorElementDataType type);
-
-  static Value CreateMap(Value& keys, Value& values);       ///< Wraps OrtApi::CreateValue
-  static Value CreateSequence(std::vector<Value>& values);  ///< Wraps OrtApi::CreateValue
-
-  template <typename T>
-  static Value CreateOpaque(const char* domain, const char* type_name, const T&);  ///< Wraps OrtApi::CreateOpaqueValue
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-  /// <summary>
-  /// This is a simple forwarding method to the other overload that helps deducing
-  /// data type enum value from the type of the buffer.
-  /// </summary>
-  /// <typeparam name="T">numeric datatype. This API is not suitable for strings.</typeparam>
-  /// <param name="info">Memory description where the user buffers reside (CPU vs GPU etc)</param>
-  /// <param name="p_data">pointer to the user supplied buffer, use nullptr for fully sparse tensors</param>
-  /// <param name="dense_shape">a would be dense shape of the tensor</param>
-  /// <param name="values_shape">non zero values shape. Use a single 0 shape for fully sparse tensors.</param>
-  /// <returns></returns>
-  template <typename T>
-  static Value CreateSparseTensor(const OrtMemoryInfo* info, T* p_data, const Shape& dense_shape,
-                                  const Shape& values_shape);
-
-  /// <summary>
-  /// Creates an OrtValue instance containing SparseTensor. This constructs
-  /// a sparse tensor that makes use of user allocated buffers. It does not make copies
-  /// of the user provided data and does not modify it. The lifespan of user provided buffers should
-  /// eclipse the life span of the resulting OrtValue. This call constructs an instance that only contain
-  /// a pointer to non-zero values. To fully populate the sparse tensor call Use<Format>Indices() API below
-  /// to supply a sparse format specific indices.
-  /// This API is not suitable for string data. Use CreateSparseTensor() with allocator specified so strings
-  /// can be properly copied into the allocated buffer.
-  /// </summary>
-  /// <param name="info">Memory description where the user buffers reside (CPU vs GPU etc)</param>
-  /// <param name="p_data">pointer to the user supplied buffer, use nullptr for fully sparse tensors</param>
-  /// <param name="dense_shape">a would be dense shape of the tensor</param>
-  /// <param name="values_shape">non zero values shape. Use a single 0 shape for fully sparse tensors.</param>
-  /// <param name="type">data type</param>
-  /// <returns>Ort::Value instance containing SparseTensor</returns>
-  static Value CreateSparseTensor(const OrtMemoryInfo* info, void* p_data, const Shape& dense_shape,
-                                  const Shape& values_shape, ONNXTensorElementDataType type);
-
-  /// <summary>
-  /// This is a simple forwarding method to the below CreateSparseTensor.
-  /// This helps to specify data type enum in terms of C++ data type.
-  /// Use CreateSparseTensor<T>
-  /// </summary>
-  /// <typeparam name="T">numeric data type only. String data enum must be specified explicitly.</typeparam>
-  /// <param name="allocator">allocator to use</param>
-  /// <param name="dense_shape">a would be dense shape of the tensor</param>
-  /// <returns>Ort::Value</returns>
-  template <typename T>
-  static Value CreateSparseTensor(OrtAllocator* allocator, const Shape& dense_shape);
-
-  /// <summary>
-  /// Creates an instance of OrtValue containing sparse tensor. The created instance has no data.
-  /// The data must be supplied by on of the FillSparseTensor<Format>() methods that take both non-zero values
-  /// and indices. The data will be copied into a buffer that would be allocated using the supplied allocator.
-  /// Use this API to create OrtValues that contain sparse tensors with all supported data types including
-  /// strings.
-  /// </summary>
-  /// <param name="allocator">allocator to use. The allocator lifespan must eclipse that of the resulting OrtValue</param>
-  /// <param name="dense_shape">a would be dense shape of the tensor</param>
-  /// <param name="type">data type</param>
-  /// <returns>an instance of Ort::Value</returns>
-  static Value CreateSparseTensor(OrtAllocator* allocator, const Shape& dense_shape, ONNXTensorElementDataType type);
-
-#endif  // !defined(DISABLE_SPARSE_TENSORS)
-};
-
-/// <summary>
-/// Represents native memory allocation coming from one of the
-/// OrtAllocators registered with OnnxRuntime.
-/// Use it to wrap an allocation made by an allocator
-/// so it can be automatically released when no longer needed.
-/// </summary>
-struct MemoryAllocation {
-  MemoryAllocation(OrtAllocator* allocator, void* p, size_t size);
-  ~MemoryAllocation();
-  MemoryAllocation(const MemoryAllocation&) = delete;
-  MemoryAllocation& operator=(const MemoryAllocation&) = delete;
-  MemoryAllocation(MemoryAllocation&&) noexcept;
-  MemoryAllocation& operator=(MemoryAllocation&&) noexcept;
-
-  void* get() { return p_; }
-  size_t size() const { return size_; }
-
- private:
-  OrtAllocator* allocator_;
-  void* p_;
-  size_t size_;
-};
-
-namespace detail {
-template <typename T>
-struct AllocatorImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  void* Alloc(size_t size);
-  MemoryAllocation GetAllocation(size_t size);
-  void Free(void* p);
-  ConstMemoryInfo GetInfo() const;
-};
-
-}  // namespace detail
-
-/** \brief Wrapper around ::OrtAllocator default instance that is owned by Onnxruntime
- *
- */
-struct AllocatorWithDefaultOptions : detail::AllocatorImpl<detail::Unowned<OrtAllocator>> {
-  explicit AllocatorWithDefaultOptions(std::nullptr_t) {}  ///< Convenience to create a class member and then replace with an instance
-  AllocatorWithDefaultOptions();
-};
-
-/** \brief Wrapper around ::OrtAllocator
- *
- */
-struct Allocator : detail::AllocatorImpl<OrtAllocator> {
-  explicit Allocator(std::nullptr_t) {}  ///< Convenience to create a class member and then replace with an instance
-  Allocator(const Session& session, const OrtMemoryInfo*);
-};
-
-using UnownedAllocator = detail::AllocatorImpl<detail::Unowned<OrtAllocator>>;
-
-namespace detail {
-namespace binding_utils {
-// Bring these out of template
-std::vector<std::string> GetOutputNamesHelper(const OrtIoBinding* binding, OrtAllocator*);
-std::vector<Value> GetOutputValuesHelper(const OrtIoBinding* binding, OrtAllocator*);
-}  // namespace binding_utils
-
-template <typename T>
-struct ConstIoBindingImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  std::vector<std::string> GetOutputNames() const;
-  std::vector<std::string> GetOutputNames(OrtAllocator*) const;
-  std::vector<Value> GetOutputValues() const;
-  std::vector<Value> GetOutputValues(OrtAllocator*) const;
-};
-
-template <typename T>
-struct IoBindingImpl : ConstIoBindingImpl<T> {
-  using B = ConstIoBindingImpl<T>;
-  using B::B;
-
-  void BindInput(const char* name, const Value&);
-  void BindOutput(const char* name, const Value&);
-  void BindOutput(const char* name, const OrtMemoryInfo*);
-  void ClearBoundInputs();
-  void ClearBoundOutputs();
-  void SynchronizeInputs();
-  void SynchronizeOutputs();
-};
-
-}  // namespace detail
-
-using ConstIoBinding = detail::ConstIoBindingImpl<detail::Unowned<const OrtIoBinding>>;
-using UnownedIoBinding = detail::IoBindingImpl<detail::Unowned<OrtIoBinding>>;
-
-/** \brief Wrapper around ::OrtIoBinding
- *
- */
-struct IoBinding : detail::IoBindingImpl<OrtIoBinding> {
-  explicit IoBinding(std::nullptr_t) {}  ///< Create an empty object for convenience. Sometimes, we want to initialize members later.
-  explicit IoBinding(Session& session);
-  ConstIoBinding GetConst() const { return ConstIoBinding{this->p_}; }
-  UnownedIoBinding GetUnowned() const { return UnownedIoBinding{this->p_}; }
-};
-
-/*! \struct Ort::ArenaCfg
- * \brief it is a structure that represents the configuration of an arena based allocator
- * \details Please see docs/C_API.md for details
- */
-struct ArenaCfg : detail::Base<OrtArenaCfg> {
-  explicit ArenaCfg(std::nullptr_t) {}  ///< Create an empty ArenaCfg object, must be assigned a valid one to be used
-  /**
-   * Wraps OrtApi::CreateArenaCfg
-   * \param max_mem - use 0 to allow ORT to choose the default
-   * \param arena_extend_strategy -  use -1 to allow ORT to choose the default, 0 = kNextPowerOfTwo, 1 = kSameAsRequested
-   * \param initial_chunk_size_bytes - use -1 to allow ORT to choose the default
-   * \param max_dead_bytes_per_chunk - use -1 to allow ORT to choose the default
-   * See docs/C_API.md for details on what the following parameters mean and how to choose these values
-   */
-  ArenaCfg(size_t max_mem, int arena_extend_strategy, int initial_chunk_size_bytes, int max_dead_bytes_per_chunk);
-};
-
-//
-// Custom OPs (only needed to implement custom OPs)
-//
-
-/// <summary>
-/// This struct provides life time management for custom op attribute
-/// </summary>
-struct OpAttr : detail::Base<OrtOpAttr> {
-  OpAttr(const char* name, const void* data, int len, OrtOpAttrType type);
-};
-
-/// <summary>
-/// This class wraps a raw pointer OrtKernelContext* that is being passed
-/// to the custom kernel Compute() method. Use it to safely access context
-/// attributes, input and output parameters with exception safety guarantees.
-/// See usage example in onnxruntime/test/testdata/custom_op_library/custom_op_library.cc
-/// </summary>
-struct KernelContext {
-  explicit KernelContext(OrtKernelContext* context);
-  size_t GetInputCount() const;
-  size_t GetOutputCount() const;
-  ConstValue GetInput(size_t index) const;
-  UnownedValue GetOutput(size_t index, const int64_t* dim_values, size_t dim_count) const;
-  UnownedValue GetOutput(size_t index, const std::vector<int64_t>& dims) const;
-  void* GetGPUComputeStream() const;
-
- private:
-  OrtKernelContext* ctx_;
-};
-
-struct KernelInfo;
-
-namespace detail {
-namespace attr_utils {
-void GetAttr(const OrtKernelInfo* p, const char* name, float&);
-void GetAttr(const OrtKernelInfo* p, const char* name, int64_t&);
-void GetAttr(const OrtKernelInfo* p, const char* name, std::string&);
-void GetAttrs(const OrtKernelInfo* p, const char* name, std::vector<float>&);
-void GetAttrs(const OrtKernelInfo* p, const char* name, std::vector<int64_t>&);
-}  // namespace attr_utils
-
-template <typename T>
-struct KernelInfoImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  KernelInfo Copy() const;
-
-  template <typename R>  // R is only implemented for float, int64_t, and string
-  R GetAttribute(const char* name) const {
-    R val;
-    attr_utils::GetAttr(this->p_, name, val);
-    return val;
-  }
-
-  template <typename R>  // R is only implemented for std::vector<float>, std::vector<int64_t>
-  std::vector<R> GetAttributes(const char* name) const {
-    std::vector<R> result;
-    attr_utils::GetAttrs(this->p_, name, result);
-    return result;
-  }
-
-  Value GetTensorAttribute(const char* name, OrtAllocator* allocator) const;
-
-  size_t GetInputCount() const;
-  size_t GetOutputCount() const;
-
-  std::string GetInputName(size_t index) const;
-  std::string GetOutputName(size_t index) const;
-
-  TypeInfo GetInputTypeInfo(size_t index) const;
-  TypeInfo GetOutputTypeInfo(size_t index) const;
-};
-
-}  // namespace detail
-
-using ConstKernelInfo = detail::KernelInfoImpl<detail::Unowned<const OrtKernelInfo>>;
-
-/// <summary>
-/// This struct owns the OrtKernInfo* pointer when a copy is made.
-/// For convenient wrapping of OrtKernelInfo* passed to kernel constructor
-/// and query attributes, warp the pointer with Ort::Unowned<KernelInfo> instance
-/// so it does not destroy the pointer the kernel does not own.
-/// </summary>
-struct KernelInfo : detail::KernelInfoImpl<OrtKernelInfo> {
-  explicit KernelInfo(std::nullptr_t) {}     ///< Create an empty instance to initialize later
-  explicit KernelInfo(OrtKernelInfo* info);  ///< Take ownership of the instance
-  ConstKernelInfo GetConst() const { return ConstKernelInfo{this->p_}; }
-};
-
-/// <summary>
-/// Create and own custom defined operation.
-/// </summary>
-struct Op : detail::Base<OrtOp> {
-  explicit Op(std::nullptr_t) {}  ///< Create an empty Operator object, must be assigned a valid one to be used
-
-  explicit Op(OrtOp*);  ///< Take ownership of the OrtOp
-
-  static Op Create(const OrtKernelInfo* info, const char* op_name, const char* domain,
-                   int version, const char** type_constraint_names,
-                   const ONNXTensorElementDataType* type_constraint_values,
-                   size_t type_constraint_count,
-                   const OpAttr* attr_values,
-                   size_t attr_count,
-                   size_t input_count, size_t output_count);
-
-  void Invoke(const OrtKernelContext* context,
-              const Value* input_values,
-              size_t input_count,
-              Value* output_values,
-              size_t output_count);
-
-  // For easier refactoring
-  void Invoke(const OrtKernelContext* context,
-              const OrtValue* const* input_values,
-              size_t input_count,
-              OrtValue* const* output_values,
-              size_t output_count);
-};
-
-/// <summary>
-/// This entire structure is deprecated, but we not marking
-/// it as a whole yet since we want to preserve for the next release.
-/// </summary>
-struct CustomOpApi {
-  CustomOpApi(const OrtApi& api) : api_(api) {}
-
-  /** \deprecated use Ort::Value::GetTensorTypeAndShape()
-   * [[deprecated]]
-   * This interface produces a pointer that must be released. Not exception safe.
-   */
-  [[deprecated("use Ort::Value::GetTensorTypeAndShape()")]] OrtTensorTypeAndShapeInfo* GetTensorTypeAndShape(_In_ const OrtValue* value);
-
-  /** \deprecated use Ort::TensorTypeAndShapeInfo::GetElementCount()
-   * [[deprecated]]
-   * This interface is redundant.
-   */
-  [[deprecated("use Ort::TensorTypeAndShapeInfo::GetElementCount()")]] size_t GetTensorShapeElementCount(_In_ const OrtTensorTypeAndShapeInfo* info);
-
-  /** \deprecated use Ort::TensorTypeAndShapeInfo::GetElementType()
-   * [[deprecated]]
-   * This interface is redundant.
-   */
-  [[deprecated("use Ort::TensorTypeAndShapeInfo::GetElementType()")]] ONNXTensorElementDataType GetTensorElementType(const OrtTensorTypeAndShapeInfo* info);
-
-  /** \deprecated use Ort::TensorTypeAndShapeInfo::GetDimensionsCount()
-   * [[deprecated]]
-   * This interface is redundant.
-   */
-  [[deprecated("use Ort::TensorTypeAndShapeInfo::GetDimensionsCount()")]] size_t GetDimensionsCount(_In_ const OrtTensorTypeAndShapeInfo* info);
-
-  /** \deprecated use Ort::TensorTypeAndShapeInfo::GetShape()
-   * [[deprecated]]
-   * This interface is redundant.
-   */
-  [[deprecated("use Ort::TensorTypeAndShapeInfo::GetShape()")]] void GetDimensions(_In_ const OrtTensorTypeAndShapeInfo* info, _Out_ int64_t* dim_values, size_t dim_values_length);
-
-  /** \deprecated
-   * [[deprecated]]
-   * This interface sets dimensions to TensorTypeAndShapeInfo, but has no effect on the OrtValue.
-   */
-  [[deprecated("Do not use")]] void SetDimensions(OrtTensorTypeAndShapeInfo* info, _In_ const int64_t* dim_values, size_t dim_count);
-
-  /** \deprecated use Ort::Value::GetTensorMutableData()
-   * [[deprecated]]
-   * This interface is redundant.
-   */
-  template <typename T>
-  [[deprecated("use Ort::Value::GetTensorMutableData()")]] T* GetTensorMutableData(_Inout_ OrtValue* value);
-
-  /** \deprecated use Ort::Value::GetTensorData()
-   * [[deprecated]]
-   * This interface is redundant.
-   */
-  template <typename T>
-  [[deprecated("use Ort::Value::GetTensorData()")]] const T* GetTensorData(_Inout_ const OrtValue* value);
-
-  /** \deprecated use Ort::Value::GetTensorMemoryInfo()
-   * [[deprecated]]
-   * This interface is redundant.
-   */
-  [[deprecated("use Ort::Value::GetTensorMemoryInfo()")]] const OrtMemoryInfo* GetTensorMemoryInfo(_In_ const OrtValue* value);
-
-  /** \deprecated use Ort::TensorTypeAndShapeInfo::GetShape()
-   * [[deprecated]]
-   * This interface is redundant.
-   */
-  [[deprecated("use Ort::TensorTypeAndShapeInfo::GetShape()")]] std::vector<int64_t> GetTensorShape(const OrtTensorTypeAndShapeInfo* info);
-
-  /** \deprecated use TensorTypeAndShapeInfo instances for automatic ownership.
-   * [[deprecated]]
-   * This interface is not exception safe.
-   */
-  [[deprecated("use TensorTypeAndShapeInfo")]] void ReleaseTensorTypeAndShapeInfo(OrtTensorTypeAndShapeInfo* input);
-
-  /** \deprecated use Ort::KernelContext::GetInputCount
-   * [[deprecated]]
-   * This interface is redundant.
-   */
-  [[deprecated("use Ort::KernelContext::GetInputCount")]] size_t KernelContext_GetInputCount(const OrtKernelContext* context);
-
-  /** \deprecated use Ort::KernelContext::GetInput
-   * [[deprecated]]
-   * This interface is redundant.
-   */
-  [[deprecated("use Ort::KernelContext::GetInput")]] const OrtValue* KernelContext_GetInput(const OrtKernelContext* context, _In_ size_t index);
-
-  /** \deprecated use Ort::KernelContext::GetOutputCount
-   * [[deprecated]]
-   * This interface is redundant.
-   */
-  [[deprecated("use Ort::KernelContext::GetOutputCount")]] size_t KernelContext_GetOutputCount(const OrtKernelContext* context);
-
-  /** \deprecated use Ort::KernelContext::GetOutput
-   * [[deprecated]]
-   * This interface is redundant.
-   */
-  [[deprecated("use Ort::KernelContext::GetOutput")]] OrtValue* KernelContext_GetOutput(OrtKernelContext* context, _In_ size_t index, _In_ const int64_t* dim_values, size_t dim_count);
-
-  /** \deprecated use Ort::KernelContext::GetGPUComputeStream
-   * [[deprecated]]
-   * This interface is redundant.
-   */
-  [[deprecated("use Ort::KernelContext::GetGPUComputeStream")]] void* KernelContext_GetGPUComputeStream(const OrtKernelContext* context);
-
-  /** \deprecated use Ort::ThrowOnError()
-   * [[deprecated]]
-   * This interface is redundant.
-   */
-  [[deprecated("use Ort::ThrowOnError()")]] void ThrowOnError(OrtStatus* result);
-
-  /** \deprecated use Ort::OpAttr
-   * [[deprecated]]
-   * This interface is not exception safe.
-   */
-  [[deprecated("use Ort::OpAttr")]] OrtOpAttr* CreateOpAttr(_In_ const char* name,
-                                                            _In_ const void* data,
-                                                            _In_ int len,
-                                                            _In_ OrtOpAttrType type);
-
-  /** \deprecated use Ort::OpAttr
-   * [[deprecated]]
-   * This interface is not exception safe.
-   */
-  [[deprecated("use Ort::OpAttr")]] void ReleaseOpAttr(_Frees_ptr_opt_ OrtOpAttr* op_attr);
-
-  /** \deprecated use Ort::Op
-   * [[deprecated]]
-   * This interface is not exception safe.
-   */
-  [[deprecated("use Ort::Op")]] OrtOp* CreateOp(_In_ const OrtKernelInfo* info,
-                                                _In_ const char* op_name,
-                                                _In_ const char* domain,
-                                                _In_ int version,
-                                                _In_opt_ const char** type_constraint_names,
-                                                _In_opt_ const ONNXTensorElementDataType* type_constraint_values,
-                                                _In_opt_ int type_constraint_count,
-                                                _In_opt_ const OrtOpAttr* const* attr_values,
-                                                _In_opt_ int attr_count,
-                                                _In_ int input_count,
-                                                _In_ int output_count);
-
-  /** \deprecated use Ort::Op::Invoke
-   * [[deprecated]]
-   * This interface is redundant
-   */
-  [[deprecated("use Ort::Op::Invoke")]] void InvokeOp(_In_ const OrtKernelContext* context,
-                                                      _In_ const OrtOp* ort_op,
-                                                      _In_ const OrtValue* const* input_values,
-                                                      _In_ int input_count,
-                                                      _Inout_ OrtValue* const* output_values,
-                                                      _In_ int output_count);
-
-  /** \deprecated use Ort::Op for automatic lifespan management.
-   * [[deprecated]]
-   * This interface is not exception safe.
-   */
-  [[deprecated("use Ort::Op")]] void ReleaseOp(_Frees_ptr_opt_ OrtOp* ort_op);
-
-  /** \deprecated use Ort::KernelInfo for automatic lifespan management or for
-   * querying attributes
-   * [[deprecated]]
-   * This interface is redundant
-   */
-  template <typename T>  // T is only implemented for std::vector<float>, std::vector<int64_t>, float, int64_t, and string
-  [[deprecated("use Ort::KernelInfo::GetAttribute")]] T KernelInfoGetAttribute(_In_ const OrtKernelInfo* info, _In_ const char* name);
-
-  /** \deprecated use Ort::KernelInfo::Copy
-   * querying attributes
-   * [[deprecated]]
-   * This interface is not exception safe
-   */
-  [[deprecated("use Ort::KernelInfo::Copy")]] OrtKernelInfo* CopyKernelInfo(_In_ const OrtKernelInfo* info);
-
-  /** \deprecated use Ort::KernelInfo for lifespan management
-   * querying attributes
-   * [[deprecated]]
-   * This interface is not exception safe
-   */
-  [[deprecated("use Ort::KernelInfo")]] void ReleaseKernelInfo(_Frees_ptr_opt_ OrtKernelInfo* info_copy);
-
- private:
-  const OrtApi& api_;
-};
-
-template <typename TOp, typename TKernel>
-struct CustomOpBase : OrtCustomOp {
-  CustomOpBase() {
-    OrtCustomOp::version = ORT_API_VERSION;
-    OrtCustomOp::CreateKernel = [](const OrtCustomOp* this_, const OrtApi* api, const OrtKernelInfo* info) { return static_cast<const TOp*>(this_)->CreateKernel(*api, info); };
-    OrtCustomOp::GetName = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetName(); };
-
-    OrtCustomOp::GetExecutionProviderType = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetExecutionProviderType(); };
-
-    OrtCustomOp::GetInputTypeCount = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetInputTypeCount(); };
-    OrtCustomOp::GetInputType = [](const OrtCustomOp* this_, size_t index) { return static_cast<const TOp*>(this_)->GetInputType(index); };
-    OrtCustomOp::GetInputMemoryType = [](const OrtCustomOp* this_, size_t index) { return static_cast<const TOp*>(this_)->GetInputMemoryType(index); };
-
-    OrtCustomOp::GetOutputTypeCount = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetOutputTypeCount(); };
-    OrtCustomOp::GetOutputType = [](const OrtCustomOp* this_, size_t index) { return static_cast<const TOp*>(this_)->GetOutputType(index); };
-
-    OrtCustomOp::KernelCompute = [](void* op_kernel, OrtKernelContext* context) { static_cast<TKernel*>(op_kernel)->Compute(context); };
-#if defined(_MSC_VER) && !defined(__clang__)
-#pragma warning(push)
-#pragma warning(disable : 26409)
-#endif
-    OrtCustomOp::KernelDestroy = [](void* op_kernel) { delete static_cast<TKernel*>(op_kernel); };
-#if defined(_MSC_VER) && !defined(__clang__)
-#pragma warning(pop)
-#endif
-    OrtCustomOp::GetInputCharacteristic = [](const OrtCustomOp* this_, size_t index) { return static_cast<const TOp*>(this_)->GetInputCharacteristic(index); };
-    OrtCustomOp::GetOutputCharacteristic = [](const OrtCustomOp* this_, size_t index) { return static_cast<const TOp*>(this_)->GetOutputCharacteristic(index); };
-
-    OrtCustomOp::GetVariadicInputMinArity = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetVariadicInputMinArity(); };
-    OrtCustomOp::GetVariadicInputHomogeneity = [](const OrtCustomOp* this_) { return static_cast<int>(static_cast<const TOp*>(this_)->GetVariadicInputHomogeneity()); };
-    OrtCustomOp::GetVariadicOutputMinArity = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetVariadicOutputMinArity(); };
-    OrtCustomOp::GetVariadicOutputHomogeneity = [](const OrtCustomOp* this_) { return static_cast<int>(static_cast<const TOp*>(this_)->GetVariadicOutputHomogeneity()); };
-  }
-
-  // Default implementation of GetExecutionProviderType that returns nullptr to default to the CPU provider
-  const char* GetExecutionProviderType() const { return nullptr; }
-
-  // Default implementations of GetInputCharacteristic() and GetOutputCharacteristic() below
-  // (inputs and outputs are required by default)
-  OrtCustomOpInputOutputCharacteristic GetInputCharacteristic(size_t /*index*/) const {
-    return OrtCustomOpInputOutputCharacteristic::INPUT_OUTPUT_REQUIRED;
-  }
-
-  OrtCustomOpInputOutputCharacteristic GetOutputCharacteristic(size_t /*index*/) const {
-    return OrtCustomOpInputOutputCharacteristic::INPUT_OUTPUT_REQUIRED;
-  }
-
-  // Default implemention of GetInputMemoryType() that returns OrtMemTypeDefault
-  OrtMemType GetInputMemoryType(size_t /*index*/) const {
-    return OrtMemTypeDefault;
-  }
-
-  // Default implementation of GetVariadicInputMinArity() returns 1 to specify that a variadic input
-  // should expect at least 1 argument.
-  int GetVariadicInputMinArity() const {
-    return 1;
-  }
-
-  // Default implementation of GetVariadicInputHomegeneity() returns true to specify that all arguments
-  // to a variadic input should be of the same type.
-  bool GetVariadicInputHomogeneity() const {
-    return true;
-  }
-
-  // Default implementation of GetVariadicOutputMinArity() returns 1 to specify that a variadic output
-  // should produce at least 1 output value.
-  int GetVariadicOutputMinArity() const {
-    return 1;
-  }
-
-  // Default implementation of GetVariadicOutputHomegeneity() returns true to specify that all output values
-  // produced by a variadic output should be of the same type.
-  bool GetVariadicOutputHomogeneity() const {
-    return true;
-  }
-
-  // Declare list of session config entries used by this Custom Op.
-  // Implement this function in order to get configs from CustomOpBase::GetSessionConfigs().
-  // This default implementation returns an empty vector of config entries.
-  std::vector<std::string> GetSessionConfigKeys() const {
-    return std::vector<std::string>{};
-  }
-
- protected:
-  // Helper function that returns a map of session config entries specified by CustomOpBase::GetSessionConfigKeys.
-  void GetSessionConfigs(std::unordered_map<std::string, std::string>& out, ConstSessionOptions options) const;
-};
-
-}  // namespace Ort
-
-#include "onnxruntime_cxx_inline.h"
diff --git a/projects/llm_framework/include/onnxruntime/core/session/onnxruntime_cxx_inline.h b/projects/llm_framework/include/onnxruntime/core/session/onnxruntime_cxx_inline.h
deleted file mode 100644
index 6d391ad0..00000000
--- a/projects/llm_framework/include/onnxruntime/core/session/onnxruntime_cxx_inline.h
+++ /dev/null
@@ -1,1874 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-// Do not include this file directly. Please include "onnxruntime_cxx_api.h" instead.
-// If interested in trying out features of the new experimental C++ API, include "experimental_onnxruntime_cxx_api.h" instead.
-//
-// These are the inline implementations of the C++ header APIs. They're in this separate file as to not clutter
-// the main C++ file with implementation details.
-
-namespace Ort {
-
-namespace detail {
-inline void ThrowStatus(const Status& st) {
-  std::string error_message = st.GetErrorMessage();
-  OrtErrorCode error_code = st.GetErrorCode();
-  ORT_CXX_API_THROW(std::move(error_message), error_code);
-}
-}  // namespace detail
-
-inline void ThrowOnError(OrtStatus* ort_status) {
-  if (ort_status) {
-    Ort::Status st(ort_status);
-    detail::ThrowStatus(st);
-  }
-}
-
-inline void ThrowOnError(const Status& st) {
-  if (st) {
-    detail::ThrowStatus(st);
-  }
-}
-
-inline Status::Status(OrtStatus* status) : Base<OrtStatus>{status} {
-}
-
-inline Status::Status(const std::exception& e) {
-  p_ = GetApi().CreateStatus(ORT_FAIL, e.what());
-}
-
-inline Status::Status(const Exception& e) {
-  p_ = GetApi().CreateStatus(e.GetOrtErrorCode(), e.what());
-}
-
-inline std::string Status::GetErrorMessage() const {
-  std::string message(GetApi().GetErrorMessage(p_));
-  return message;
-}
-
-inline OrtErrorCode Status::GetErrorCode() const {
-  return GetApi().GetErrorCode(p_);
-}
-
-// This template converts a C++ type into it's ONNXTensorElementDataType
-template <typename T>
-struct TypeToTensorType;
-template <>
-struct TypeToTensorType<float> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT;
-};
-template <>
-struct TypeToTensorType<Float16_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT16;
-};
-template <>
-struct TypeToTensorType<BFloat16_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_BFLOAT16;
-};
-template <>
-struct TypeToTensorType<double> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_DOUBLE;
-};
-template <>
-struct TypeToTensorType<int8_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_INT8;
-};
-template <>
-struct TypeToTensorType<int16_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_INT16;
-};
-template <>
-struct TypeToTensorType<int32_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32;
-};
-template <>
-struct TypeToTensorType<int64_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64;
-};
-template <>
-struct TypeToTensorType<uint8_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT8;
-};
-template <>
-struct TypeToTensorType<uint16_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT16;
-};
-template <>
-struct TypeToTensorType<uint32_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT32;
-};
-template <>
-struct TypeToTensorType<uint64_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT64;
-};
-template <>
-struct TypeToTensorType<bool> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_BOOL;
-};
-
-inline MemoryAllocation::MemoryAllocation(OrtAllocator* allocator, void* p, size_t size)
-    : allocator_(allocator), p_(p), size_(size) {
-}
-
-inline MemoryAllocation::~MemoryAllocation() {
-  if (p_ != nullptr) {
-    // We do not throw out of destructor
-    auto ret = GetApi().AllocatorFree(allocator_, p_);
-    static_cast<void>(ret);
-  }
-}
-
-inline MemoryAllocation::MemoryAllocation(MemoryAllocation&& o) noexcept : allocator_(nullptr), p_(nullptr), size_(0) {
-  *this = std::move(o);
-}
-
-inline MemoryAllocation& MemoryAllocation::operator=(MemoryAllocation&& o) noexcept {
-  OrtAllocator* alloc = nullptr;
-  void* p = nullptr;
-  size_t sz = 0;
-
-  // Swap out this
-  std::swap(alloc, allocator_);
-  std::swap(p, p_);
-  std::swap(sz, size_);
-
-  // Swap with incoming
-  std::swap(allocator_, o.allocator_);
-  std::swap(p_, o.p_);
-  std::swap(size_, o.size_);
-
-  // Destroy this instance if needed
-  MemoryAllocation this_alloc(alloc, p, sz);
-  return *this;
-}
-
-namespace detail {
-
-template <typename T>
-inline void* AllocatorImpl<T>::Alloc(size_t size) {
-  void* out;
-  ThrowOnError(GetApi().AllocatorAlloc(this->p_, size, &out));
-  return out;
-}
-
-template <typename T>
-inline MemoryAllocation AllocatorImpl<T>::GetAllocation(size_t size) {
-  void* out;
-  ThrowOnError(GetApi().AllocatorAlloc(this->p_, size, &out));
-  MemoryAllocation result(this->p_, out, size);
-  return result;
-}
-
-template <typename T>
-inline void AllocatorImpl<T>::Free(void* p) {
-  ThrowOnError(GetApi().AllocatorFree(this->p_, p));
-}
-
-template <typename T>
-inline ConstMemoryInfo AllocatorImpl<T>::GetInfo() const {
-  const OrtMemoryInfo* out;
-  ThrowOnError(GetApi().AllocatorGetInfo(this->p_, &out));
-  return ConstMemoryInfo{out};
-}
-
-}  // namespace detail
-
-inline AllocatorWithDefaultOptions::AllocatorWithDefaultOptions() {
-  ThrowOnError(GetApi().GetAllocatorWithDefaultOptions(&this->p_));
-}
-
-inline Allocator::Allocator(const Session& sess, const OrtMemoryInfo* mem_info) {
-  ThrowOnError(GetApi().CreateAllocator(sess, mem_info, &this->p_));
-}
-
-namespace detail {
-
-template <typename T>
-inline std::string MemoryInfoImpl<T>::GetAllocatorName() const {
-  const char* name = nullptr;
-  ThrowOnError(GetApi().MemoryInfoGetName(this->p_, &name));
-  return std::string(name);
-}
-
-template <typename T>
-inline OrtAllocatorType MemoryInfoImpl<T>::GetAllocatorType() const {
-  OrtAllocatorType type;
-  ThrowOnError(GetApi().MemoryInfoGetType(this->p_, &type));
-  return type;
-}
-
-template <typename T>
-inline int MemoryInfoImpl<T>::GetDeviceId() const {
-  int id = 0;
-  ThrowOnError(GetApi().MemoryInfoGetId(this->p_, &id));
-  return id;
-}
-
-template <typename T>
-inline OrtMemoryInfoDeviceType MemoryInfoImpl<T>::GetDeviceType() const {
-  OrtMemoryInfoDeviceType type;
-  GetApi().MemoryInfoGetDeviceType(this->p_, &type);
-  return type;
-}
-
-template <typename T>
-inline OrtMemType MemoryInfoImpl<T>::GetMemoryType() const {
-  OrtMemType type;
-  ThrowOnError(GetApi().MemoryInfoGetMemType(this->p_, &type));
-  return type;
-}
-
-template <typename T>
-template <typename U>
-inline bool MemoryInfoImpl<T>::operator==(const MemoryInfoImpl<U>& o) const {
-  int comp_result = 0;
-  ThrowOnError(Ort::GetApi().CompareMemoryInfo(this->p_, o, &comp_result));
-  return comp_result == 0;
-}
-
-}  // namespace detail
-
-inline MemoryInfo MemoryInfo::CreateCpu(OrtAllocatorType type, OrtMemType mem_type) {
-  OrtMemoryInfo* p;
-  ThrowOnError(GetApi().CreateCpuMemoryInfo(type, mem_type, &p));
-  return MemoryInfo(p);
-}
-
-inline MemoryInfo::MemoryInfo(const char* name, OrtAllocatorType type, int id, OrtMemType mem_type) {
-  ThrowOnError(GetApi().CreateMemoryInfo(name, type, id, mem_type, &this->p_));
-}
-
-namespace detail {
-template <typename T>
-inline std::vector<std::string> ConstIoBindingImpl<T>::GetOutputNames() const {
-  AllocatorWithDefaultOptions allocator;
-  return binding_utils::GetOutputNamesHelper(this->p_, allocator);
-}
-
-template <typename T>
-inline std::vector<std::string> ConstIoBindingImpl<T>::GetOutputNames(OrtAllocator* allocator) const {
-  return binding_utils::GetOutputNamesHelper(this->p_, allocator);
-}
-
-template <typename T>
-inline std::vector<Value> ConstIoBindingImpl<T>::GetOutputValues() const {
-  AllocatorWithDefaultOptions allocator;
-  return binding_utils::GetOutputValuesHelper(this->p_, allocator);
-}
-
-template <typename T>
-inline std::vector<Value> ConstIoBindingImpl<T>::GetOutputValues(OrtAllocator* allocator) const {
-  return binding_utils::GetOutputValuesHelper(this->p_, allocator);
-}
-
-template <typename T>
-inline void IoBindingImpl<T>::BindInput(const char* name, const Value& value) {
-  ThrowOnError(GetApi().BindInput(this->p_, name, value));
-}
-
-template <typename T>
-inline void IoBindingImpl<T>::BindOutput(const char* name, const Value& value) {
-  ThrowOnError(GetApi().BindOutput(this->p_, name, value));
-}
-
-template <typename T>
-inline void IoBindingImpl<T>::BindOutput(const char* name, const OrtMemoryInfo* mem_info) {
-  ThrowOnError(GetApi().BindOutputToDevice(this->p_, name, mem_info));
-}
-
-template <typename T>
-inline void IoBindingImpl<T>::ClearBoundInputs() {
-  GetApi().ClearBoundInputs(this->p_);
-}
-
-template <typename T>
-inline void IoBindingImpl<T>::ClearBoundOutputs() {
-  GetApi().ClearBoundOutputs(this->p_);
-}
-
-template <typename T>
-inline void IoBindingImpl<T>::SynchronizeInputs() {
-  ThrowOnError(GetApi().SynchronizeBoundInputs(this->p_));
-}
-
-template <typename T>
-inline void IoBindingImpl<T>::SynchronizeOutputs() {
-  ThrowOnError(GetApi().SynchronizeBoundOutputs(this->p_));
-}
-
-namespace binding_utils {
-inline std::vector<std::string> GetOutputNamesHelper(const OrtIoBinding* binding, OrtAllocator* allocator) {
-  std::vector<std::string> result;
-  auto free_fn = detail::AllocatedFree(allocator);
-  using Ptr = std::unique_ptr<void, decltype(free_fn)>;
-
-  char* buffer = nullptr;
-  size_t* lengths = nullptr;
-  size_t count = 0;
-  ThrowOnError(GetApi().GetBoundOutputNames(binding, allocator, &buffer, &lengths, &count));
-
-  if (count == 0) {
-    return result;
-  }
-
-  Ptr buffer_g(buffer, free_fn);
-  Ptr lengths_g(lengths, free_fn);
-
-  result.reserve(count);
-  for (size_t i = 0; i < count; ++i) {
-    auto sz = *lengths;
-    result.emplace_back(buffer, sz);
-    buffer += sz;
-    ++lengths;
-  }
-  return result;
-}
-
-inline std::vector<Value> GetOutputValuesHelper(const OrtIoBinding* binding, OrtAllocator* allocator) {
-  std::vector<Value> result;
-  size_t owned = 0;
-  size_t output_count = 0;
-  // Lambda to release the buffer when no longer needed and
-  // make sure that we destroy all instances on exception
-  auto free_fn = [&owned, &output_count, allocator](OrtValue** buffer) {
-    if (buffer) {
-      while (owned < output_count) {
-        auto* p = buffer + owned++;
-        GetApi().ReleaseValue(*p);
-      }
-      allocator->Free(allocator, buffer);
-    }
-  };
-  using Ptr = std::unique_ptr<OrtValue*, decltype(free_fn)>;
-
-  OrtValue** output_buffer = nullptr;
-  ThrowOnError(GetApi().GetBoundOutputValues(binding, allocator, &output_buffer, &output_count));
-  if (output_count == 0) {
-    return result;
-  }
-
-  Ptr buffer_g(output_buffer, free_fn);
-
-  result.reserve(output_count);
-  for (size_t i = 0; i < output_count; ++i) {
-    result.emplace_back(output_buffer[i]);
-    ++owned;
-  }
-  return result;
-}
-
-}  // namespace binding_utils
-}  // namespace detail
-
-inline IoBinding::IoBinding(Session& session) {
-  ThrowOnError(GetApi().CreateIoBinding(session, &this->p_));
-}
-
-inline ArenaCfg::ArenaCfg(size_t max_mem, int arena_extend_strategy, int initial_chunk_size_bytes, int max_dead_bytes_per_chunk) {
-  ThrowOnError(GetApi().CreateArenaCfg(max_mem, arena_extend_strategy, initial_chunk_size_bytes, max_dead_bytes_per_chunk, &p_));
-}
-
-inline ThreadingOptions::ThreadingOptions() {
-  ThrowOnError(GetApi().CreateThreadingOptions(&p_));
-}
-
-inline ThreadingOptions& ThreadingOptions::SetGlobalIntraOpNumThreads(int intra_op_num_threads) {
-  ThrowOnError(GetApi().SetGlobalIntraOpNumThreads(p_, intra_op_num_threads));
-  return *this;
-}
-
-inline ThreadingOptions& ThreadingOptions::SetGlobalInterOpNumThreads(int inter_op_num_threads) {
-  ThrowOnError(GetApi().SetGlobalInterOpNumThreads(p_, inter_op_num_threads));
-  return *this;
-}
-
-inline ThreadingOptions& ThreadingOptions::SetGlobalSpinControl(int allow_spinning) {
-  ThrowOnError(GetApi().SetGlobalSpinControl(p_, allow_spinning));
-  return *this;
-}
-
-inline ThreadingOptions& ThreadingOptions::SetGlobalDenormalAsZero() {
-  ThrowOnError(GetApi().SetGlobalDenormalAsZero(p_));
-  return *this;
-}
-
-inline ThreadingOptions& ThreadingOptions::SetGlobalCustomCreateThreadFn(OrtCustomCreateThreadFn ort_custom_create_thread_fn) {
-  ThrowOnError(GetApi().SetGlobalCustomCreateThreadFn(p_, ort_custom_create_thread_fn));
-  return *this;
-}
-
-inline ThreadingOptions& ThreadingOptions::SetGlobalCustomThreadCreationOptions(void* ort_custom_thread_creation_options) {
-  ThrowOnError(GetApi().SetGlobalCustomThreadCreationOptions(p_, ort_custom_thread_creation_options));
-  return *this;
-}
-
-inline ThreadingOptions& ThreadingOptions::SetGlobalCustomJoinThreadFn(OrtCustomJoinThreadFn ort_custom_join_thread_fn) {
-  ThrowOnError(GetApi().SetGlobalCustomJoinThreadFn(p_, ort_custom_join_thread_fn));
-  return *this;
-}
-
-inline Env::Env(OrtLoggingLevel logging_level, _In_ const char* logid) {
-  ThrowOnError(GetApi().CreateEnv(logging_level, logid, &p_));
-  if (strcmp(logid, "onnxruntime-node") == 0) {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_NODEJS));
-  } else {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_CPLUSPLUS));
-  }
-}
-
-inline Env::Env(OrtLoggingLevel logging_level, const char* logid, OrtLoggingFunction logging_function, void* logger_param) {
-  ThrowOnError(GetApi().CreateEnvWithCustomLogger(logging_function, logger_param, logging_level, logid, &p_));
-  if (strcmp(logid, "onnxruntime-node") == 0) {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_NODEJS));
-  } else {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_CPLUSPLUS));
-  }
-}
-
-inline Env::Env(const OrtThreadingOptions* tp_options, OrtLoggingLevel logging_level, _In_ const char* logid) {
-  ThrowOnError(GetApi().CreateEnvWithGlobalThreadPools(logging_level, logid, tp_options, &p_));
-  if (strcmp(logid, "onnxruntime-node") == 0) {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_NODEJS));
-  } else {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_CPLUSPLUS));
-  }
-}
-
-inline Env::Env(const OrtThreadingOptions* tp_options, OrtLoggingFunction logging_function, void* logger_param,
-                OrtLoggingLevel logging_level, _In_ const char* logid) {
-  ThrowOnError(GetApi().CreateEnvWithCustomLoggerAndGlobalThreadPools(logging_function, logger_param, logging_level, logid, tp_options, &p_));
-  if (strcmp(logid, "onnxruntime-node") == 0) {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_NODEJS));
-  } else {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_CPLUSPLUS));
-  }
-}
-
-inline Env& Env::EnableTelemetryEvents() {
-  ThrowOnError(GetApi().EnableTelemetryEvents(p_));
-  return *this;
-}
-
-inline Env& Env::DisableTelemetryEvents() {
-  ThrowOnError(GetApi().DisableTelemetryEvents(p_));
-  return *this;
-}
-
-inline Env& Env::UpdateEnvWithCustomLogLevel(OrtLoggingLevel log_severity_level) {
-  ThrowOnError(GetApi().UpdateEnvWithCustomLogLevel(p_, log_severity_level));
-  return *this;
-}
-
-inline Env& Env::CreateAndRegisterAllocator(const OrtMemoryInfo* mem_info, const OrtArenaCfg* arena_cfg) {
-  ThrowOnError(GetApi().CreateAndRegisterAllocator(p_, mem_info, arena_cfg));
-  return *this;
-}
-
-inline CustomOpDomain::CustomOpDomain(const char* domain) {
-  ThrowOnError(GetApi().CreateCustomOpDomain(domain, &p_));
-}
-
-inline void CustomOpDomain::Add(const OrtCustomOp* op) {
-  ThrowOnError(GetApi().CustomOpDomain_Add(p_, op));
-}
-
-inline RunOptions::RunOptions() {
-  ThrowOnError(GetApi().CreateRunOptions(&p_));
-}
-
-inline RunOptions& RunOptions::SetRunLogVerbosityLevel(int level) {
-  ThrowOnError(GetApi().RunOptionsSetRunLogVerbosityLevel(p_, level));
-  return *this;
-}
-
-inline RunOptions& RunOptions::SetRunLogSeverityLevel(int level) {
-  ThrowOnError(GetApi().RunOptionsSetRunLogSeverityLevel(p_, level));
-  return *this;
-}
-
-inline int RunOptions::GetRunLogVerbosityLevel() const {
-  int out;
-  ThrowOnError(GetApi().RunOptionsGetRunLogVerbosityLevel(p_, &out));
-  return out;
-}
-
-inline int RunOptions::GetRunLogSeverityLevel() const {
-  int out;
-  ThrowOnError(GetApi().RunOptionsGetRunLogSeverityLevel(p_, &out));
-  return out;
-}
-
-inline RunOptions& RunOptions::SetRunTag(const char* run_tag) {
-  ThrowOnError(GetApi().RunOptionsSetRunTag(p_, run_tag));
-  return *this;
-}
-
-inline const char* RunOptions::GetRunTag() const {
-  const char* out;
-  ThrowOnError(GetApi().RunOptionsGetRunTag(p_, &out));
-  return out;
-}
-
-inline RunOptions& RunOptions::AddConfigEntry(const char* config_key, const char* config_value) {
-  ThrowOnError(GetApi().AddRunConfigEntry(p_, config_key, config_value));
-  return *this;
-}
-
-inline RunOptions& RunOptions::SetTerminate() {
-  ThrowOnError(GetApi().RunOptionsSetTerminate(p_));
-  return *this;
-}
-
-inline RunOptions& RunOptions::UnsetTerminate() {
-  ThrowOnError(GetApi().RunOptionsUnsetTerminate(p_));
-  return *this;
-}
-
-namespace detail {
-
-template <typename T>
-inline Ort::SessionOptions ConstSessionOptionsImpl<T>::Clone() const {
-  OrtSessionOptions* out;
-  ThrowOnError(GetApi().CloneSessionOptions(this->p_, &out));
-  return SessionOptions{out};
-}
-
-template <typename T>
-inline std::string ConstSessionOptionsImpl<T>::GetConfigEntry(const char* config_key) const {
-  size_t size = 0;
-  // Feed nullptr for the data buffer to query the true size of the string value
-  Ort::ThrowOnError(GetApi().GetSessionConfigEntry(this->p_, config_key, nullptr, &size));
-
-  std::string out;
-  out.resize(size);
-  Ort::ThrowOnError(GetApi().GetSessionConfigEntry(this->p_, config_key, &out[0], &size));
-  out.resize(size - 1);  // remove the terminating character '\0'
-
-  return out;
-}
-
-template <typename T>
-inline bool ConstSessionOptionsImpl<T>::HasConfigEntry(const char* config_key) const {
-  int out = 0;
-  Ort::ThrowOnError(GetApi().HasSessionConfigEntry(this->p_, config_key, &out));
-  return static_cast<bool>(out);
-}
-
-template <typename T>
-inline std::string ConstSessionOptionsImpl<T>::GetConfigEntryOrDefault(const char* config_key, const std::string& def) {
-  if (!this->HasConfigEntry(config_key)) {
-    return def;
-  }
-
-  return this->GetConfigEntry(config_key);
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetIntraOpNumThreads(int intra_op_num_threads) {
-  ThrowOnError(GetApi().SetIntraOpNumThreads(this->p_, intra_op_num_threads));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetInterOpNumThreads(int inter_op_num_threads) {
-  ThrowOnError(GetApi().SetInterOpNumThreads(this->p_, inter_op_num_threads));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetGraphOptimizationLevel(GraphOptimizationLevel graph_optimization_level) {
-  ThrowOnError(GetApi().SetSessionGraphOptimizationLevel(this->p_, graph_optimization_level));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetOptimizedModelFilePath(const ORTCHAR_T* optimized_model_filepath) {
-  ThrowOnError(GetApi().SetOptimizedModelFilePath(this->p_, optimized_model_filepath));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::EnableProfiling(const ORTCHAR_T* profile_file_prefix) {
-  ThrowOnError(GetApi().EnableProfiling(this->p_, profile_file_prefix));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::DisableProfiling() {
-  ThrowOnError(GetApi().DisableProfiling(this->p_));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::EnableOrtCustomOps() {
-  ThrowOnError(GetApi().EnableOrtCustomOps(this->p_));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::EnableMemPattern() {
-  ThrowOnError(GetApi().EnableMemPattern(this->p_));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::DisableMemPattern() {
-  ThrowOnError(GetApi().DisableMemPattern(this->p_));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::EnableCpuMemArena() {
-  ThrowOnError(GetApi().EnableCpuMemArena(this->p_));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::DisableCpuMemArena() {
-  ThrowOnError(GetApi().DisableCpuMemArena(this->p_));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetExecutionMode(ExecutionMode execution_mode) {
-  ThrowOnError(GetApi().SetSessionExecutionMode(this->p_, execution_mode));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetLogId(const char* logid) {
-  ThrowOnError(GetApi().SetSessionLogId(this->p_, logid));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetLogSeverityLevel(int level) {
-  ThrowOnError(GetApi().SetSessionLogSeverityLevel(this->p_, level));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::Add(OrtCustomOpDomain* custom_op_domain) {
-  ThrowOnError(GetApi().AddCustomOpDomain(this->p_, custom_op_domain));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AddConfigEntry(const char* config_key, const char* config_value) {
-  ThrowOnError(GetApi().AddSessionConfigEntry(this->p_, config_key, config_value));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AddInitializer(const char* name, const OrtValue* ort_val) {
-  ThrowOnError(GetApi().AddInitializer(this->p_, name, ort_val));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::DisablePerSessionThreads() {
-  ThrowOnError(GetApi().DisablePerSessionThreads(this->p_));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AddExternalInitializers(const std::vector<std::string>& names,
-                                                                             const std::vector<Value>& ort_values) {
-  const size_t inputs_num = names.size();
-  if (inputs_num != ort_values.size()) {
-    ORT_CXX_API_THROW("Expecting names and ort_values to have the same length", ORT_INVALID_ARGUMENT);
-  }
-  std::vector<const char*> names_ptr;
-  std::vector<const OrtValue*> ort_values_ptrs;
-  names_ptr.reserve(inputs_num);
-  ort_values_ptrs.reserve(inputs_num);
-  for (size_t i = 0; i < inputs_num; ++i) {
-    names_ptr.push_back(names[i].c_str());
-    ort_values_ptrs.push_back(ort_values[i]);
-  }
-  ThrowOnError(GetApi().AddExternalInitializers(this->p_, names_ptr.data(), ort_values_ptrs.data(), inputs_num));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_CUDA(const OrtCUDAProviderOptions& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_CUDA(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_CUDA_V2(const OrtCUDAProviderOptionsV2& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_CUDA_V2(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_ROCM(const OrtROCMProviderOptions& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_ROCM(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_TensorRT(const OrtTensorRTProviderOptions& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_TensorRT(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_TensorRT_V2(const OrtTensorRTProviderOptionsV2& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_TensorRT_V2(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_MIGraphX(const OrtMIGraphXProviderOptions& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_MIGraphX(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_CANN(const OrtCANNProviderOptions& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_CANN(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider(
-    const std::string& provider_name,
-    const std::unordered_map<std::string, std::string>& provider_options) {
-  auto num_entries = provider_options.size();
-  std::vector<const char*> keys, values;
-  if (num_entries > 0) {
-    keys.reserve(num_entries);
-    values.reserve(num_entries);
-
-    for (const auto& entry : provider_options) {
-      keys.push_back(entry.first.c_str());
-      values.push_back(entry.second.c_str());
-    }
-  }
-
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider(this->p_, provider_name.c_str(),
-                                                              keys.data(), values.data(), num_entries));
-
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetCustomCreateThreadFn(OrtCustomCreateThreadFn ort_custom_create_thread_fn) {
-  ThrowOnError(GetApi().SessionOptionsSetCustomCreateThreadFn(this->p_, ort_custom_create_thread_fn));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetCustomThreadCreationOptions(void* ort_custom_thread_creation_options) {
-  ThrowOnError(GetApi().SessionOptionsSetCustomThreadCreationOptions(this->p_, ort_custom_thread_creation_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetCustomJoinThreadFn(OrtCustomJoinThreadFn ort_custom_join_thread_fn) {
-  ThrowOnError(GetApi().SessionOptionsSetCustomJoinThreadFn(this->p_, ort_custom_join_thread_fn));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_OpenVINO(const OrtOpenVINOProviderOptions& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_OpenVINO(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::RegisterCustomOpsLibrary(const ORTCHAR_T* library_name,
-                                                                              const CustomOpConfigs& custom_op_configs) {
-  // Add custom op config entries before registering the custom op library. Otherwise, the config entries _may_ be ignored by
-  // the custom op library.
-  for (const auto& config_iter : custom_op_configs.GetFlattenedConfigs()) {
-    AddConfigEntry(config_iter.first.c_str(), config_iter.second.c_str());
-  }
-
-  ThrowOnError(GetApi().RegisterCustomOpsLibrary_V2(this->p_, library_name));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::RegisterCustomOpsUsingFunction(const char* registration_function_name) {
-  ThrowOnError(GetApi().RegisterCustomOpsUsingFunction(this->p_, registration_function_name));
-  return *this;
-}
-
-/// Session
-template <typename T>
-inline size_t ConstSessionImpl<T>::GetInputCount() const {
-  size_t out;
-  ThrowOnError(GetApi().SessionGetInputCount(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline size_t ConstSessionImpl<T>::GetOutputCount() const {
-  size_t out;
-  ThrowOnError(GetApi().SessionGetOutputCount(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline size_t ConstSessionImpl<T>::GetOverridableInitializerCount() const {
-  size_t out;
-  ThrowOnError(GetApi().SessionGetOverridableInitializerCount(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline AllocatedStringPtr ConstSessionImpl<T>::GetInputNameAllocated(size_t index, OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().SessionGetInputName(this->p_, index, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-template <typename T>
-inline AllocatedStringPtr ConstSessionImpl<T>::GetOutputNameAllocated(size_t index, OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().SessionGetOutputName(this->p_, index, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-template <typename T>
-inline AllocatedStringPtr ConstSessionImpl<T>::GetOverridableInitializerNameAllocated(size_t index, OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().SessionGetOverridableInitializerName(this->p_, index, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-template <typename T>
-inline uint64_t ConstSessionImpl<T>::GetProfilingStartTimeNs() const {
-  uint64_t out;
-  ThrowOnError(GetApi().SessionGetProfilingStartTimeNs(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline ModelMetadata ConstSessionImpl<T>::GetModelMetadata() const {
-  OrtModelMetadata* out;
-  ThrowOnError(GetApi().SessionGetModelMetadata(this->p_, &out));
-  return ModelMetadata{out};
-}
-
-template <typename T>
-inline TypeInfo ConstSessionImpl<T>::GetInputTypeInfo(size_t index) const {
-  OrtTypeInfo* out;
-  ThrowOnError(GetApi().SessionGetInputTypeInfo(this->p_, index, &out));
-  return TypeInfo{out};
-}
-
-template <typename T>
-inline TypeInfo ConstSessionImpl<T>::GetOutputTypeInfo(size_t index) const {
-  OrtTypeInfo* out;
-  ThrowOnError(GetApi().SessionGetOutputTypeInfo(this->p_, index, &out));
-  return TypeInfo{out};
-}
-
-template <typename T>
-inline TypeInfo ConstSessionImpl<T>::GetOverridableInitializerTypeInfo(size_t index) const {
-  OrtTypeInfo* out;
-  ThrowOnError(GetApi().SessionGetOverridableInitializerTypeInfo(this->p_, index, &out));
-  return TypeInfo{out};
-}
-
-template <typename T>
-inline std::vector<Value> SessionImpl<T>::Run(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count,
-                                              const char* const* output_names, size_t output_count) {
-  std::vector<Value> output_values;
-  output_values.reserve(output_count);
-  for (size_t i = 0; i < output_count; i++)
-    output_values.emplace_back(nullptr);
-  Run(run_options, input_names, input_values, input_count, output_names, output_values.data(), output_count);
-  return output_values;
-}
-
-template <typename T>
-inline void SessionImpl<T>::Run(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count,
-                                const char* const* output_names, Value* output_values, size_t output_count) {
-  static_assert(sizeof(Value) == sizeof(OrtValue*), "Value is really just an array of OrtValue* in memory, so we can reinterpret_cast safely");
-  auto ort_input_values = reinterpret_cast<const OrtValue* const*>(input_values);
-  auto ort_output_values = reinterpret_cast<OrtValue**>(output_values);
-  ThrowOnError(GetApi().Run(this->p_, run_options, input_names, ort_input_values, input_count, output_names, output_count, ort_output_values));
-}
-
-template <typename T>
-inline void SessionImpl<T>::Run(const RunOptions& run_options, const IoBinding& io_binding) {
-  ThrowOnError(GetApi().RunWithBinding(this->p_, run_options, io_binding));
-}
-
-template <typename T>
-inline AllocatedStringPtr SessionImpl<T>::EndProfilingAllocated(OrtAllocator* allocator) {
-  char* out = nullptr;
-  ThrowOnError(GetApi().SessionEndProfiling(this->p_, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-}  // namespace detail
-
-inline SessionOptions::SessionOptions() {
-  ThrowOnError(GetApi().CreateSessionOptions(&this->p_));
-}
-
-/// CustomOpConfigs
-inline std::string detail::MakeCustomOpConfigEntryKey(const char* custom_op_name, const char* config) {
-  std::string config_key = "custom_op.";
-
-  config_key += custom_op_name;
-  config_key += ".";
-  config_key += config;
-
-  return config_key;
-}
-
-inline CustomOpConfigs& CustomOpConfigs::AddConfig(const char* custom_op_name, const char* config_key, const char* config_value) {
-  const std::string full_flat_key = detail::MakeCustomOpConfigEntryKey(custom_op_name, config_key);
-  flat_configs_[full_flat_key] = config_value;
-  return *this;
-}
-
-inline const std::unordered_map<std::string, std::string>& CustomOpConfigs::GetFlattenedConfigs() const {
-  return flat_configs_;
-}
-
-inline Session::Session(const Env& env, const ORTCHAR_T* model_path, const SessionOptions& options) {
-  ThrowOnError(GetApi().CreateSession(env, model_path, options, &this->p_));
-}
-
-inline Session::Session(const Env& env, const ORTCHAR_T* model_path, const SessionOptions& options,
-                        OrtPrepackedWeightsContainer* prepacked_weights_container) {
-  ThrowOnError(GetApi().CreateSessionWithPrepackedWeightsContainer(env, model_path, options, prepacked_weights_container, &this->p_));
-}
-
-inline Session::Session(const Env& env, const void* model_data, size_t model_data_length, const SessionOptions& options) {
-  ThrowOnError(GetApi().CreateSessionFromArray(env, model_data, model_data_length, options, &this->p_));
-}
-
-inline Session::Session(const Env& env, const void* model_data, size_t model_data_length,
-                        const SessionOptions& options, OrtPrepackedWeightsContainer* prepacked_weights_container) {
-  ThrowOnError(GetApi().CreateSessionFromArrayWithPrepackedWeightsContainer(env, model_data, model_data_length, options,
-                                                                            prepacked_weights_container, &this->p_));
-}
-
-inline AllocatedStringPtr ModelMetadata::GetProducerNameAllocated(OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().ModelMetadataGetProducerName(p_, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-inline AllocatedStringPtr ModelMetadata::GetGraphNameAllocated(OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().ModelMetadataGetGraphName(p_, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-inline AllocatedStringPtr ModelMetadata::GetDomainAllocated(OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().ModelMetadataGetDomain(p_, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-inline AllocatedStringPtr Ort::ModelMetadata::GetDescriptionAllocated(OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().ModelMetadataGetDescription(p_, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-inline AllocatedStringPtr ModelMetadata::GetGraphDescriptionAllocated(OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().ModelMetadataGetGraphDescription(p_, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-inline AllocatedStringPtr ModelMetadata::LookupCustomMetadataMapAllocated(const char* key, OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().ModelMetadataLookupCustomMetadataMap(p_, allocator, key, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-inline std::vector<AllocatedStringPtr> ModelMetadata::GetCustomMetadataMapKeysAllocated(OrtAllocator* allocator) const {
-  auto deletor = detail::AllocatedFree(allocator);
-  std::vector<AllocatedStringPtr> result;
-
-  char** out = nullptr;
-  int64_t num_keys = 0;
-  ThrowOnError(GetApi().ModelMetadataGetCustomMetadataMapKeys(p_, allocator, &out, &num_keys));
-  if (num_keys <= 0) {
-    return result;
-  }
-
-  // array of pointers will be freed
-  std::unique_ptr<void, decltype(deletor)> array_guard(out, deletor);
-  // reserve may throw
-  auto strings_deletor = [&deletor, num_keys](char** out) { for(int64_t i = 0; i < num_keys; ++i) deletor(out[i]); };
-  std::unique_ptr<char*, decltype(strings_deletor)> strings_guard(out, strings_deletor);
-  result.reserve(static_cast<size_t>(num_keys));
-  strings_guard.release();
-  for (int64_t i = 0; i < num_keys; ++i) {
-    result.push_back(AllocatedStringPtr(out[i], deletor));
-  }
-
-  return result;
-}
-
-inline int64_t ModelMetadata::GetVersion() const {
-  int64_t out;
-  ThrowOnError(GetApi().ModelMetadataGetVersion(p_, &out));
-  return out;
-}
-
-namespace detail {
-
-template <typename T>
-inline ONNXTensorElementDataType TensorTypeAndShapeInfoImpl<T>::GetElementType() const {
-  ONNXTensorElementDataType out;
-  ThrowOnError(GetApi().GetTensorElementType(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline size_t TensorTypeAndShapeInfoImpl<T>::GetElementCount() const {
-  size_t out;
-  ThrowOnError(GetApi().GetTensorShapeElementCount(this->p_, &out));
-  return static_cast<size_t>(out);
-}
-
-template <typename T>
-inline size_t TensorTypeAndShapeInfoImpl<T>::GetDimensionsCount() const {
-  size_t out;
-  ThrowOnError(GetApi().GetDimensionsCount(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline void TensorTypeAndShapeInfoImpl<T>::GetDimensions(int64_t* values, size_t values_count) const {
-  ThrowOnError(GetApi().GetDimensions(this->p_, values, values_count));
-}
-
-template <typename T>
-inline void TensorTypeAndShapeInfoImpl<T>::GetSymbolicDimensions(const char** values, size_t values_count) const {
-  ThrowOnError(GetApi().GetSymbolicDimensions(this->p_, values, values_count));
-}
-
-template <typename T>
-inline std::vector<int64_t> TensorTypeAndShapeInfoImpl<T>::GetShape() const {
-  std::vector<int64_t> out(GetDimensionsCount(), 0);
-  ThrowOnError(GetApi().GetDimensions(this->p_, out.data(), out.size()));
-  return out;
-}
-
-}  // namespace detail
-
-namespace detail {
-template <typename T>
-inline ConstTensorTypeAndShapeInfo TypeInfoImpl<T>::GetTensorTypeAndShapeInfo() const {
-  const OrtTensorTypeAndShapeInfo* out;
-  ThrowOnError(GetApi().CastTypeInfoToTensorInfo(this->p_, &out));
-  return ConstTensorTypeAndShapeInfo{out};
-}
-
-template <typename T>
-inline ConstSequenceTypeInfo TypeInfoImpl<T>::GetSequenceTypeInfo() const {
-  const OrtSequenceTypeInfo* out;
-  ThrowOnError(GetApi().CastTypeInfoToSequenceTypeInfo(this->p_, &out));
-  return ConstSequenceTypeInfo{out};
-}
-
-template <typename T>
-inline ConstMapTypeInfo TypeInfoImpl<T>::GetMapTypeInfo() const {
-  const OrtMapTypeInfo* out;
-  ThrowOnError(GetApi().CastTypeInfoToMapTypeInfo(this->p_, &out));
-  return ConstMapTypeInfo{out};
-}
-
-template <typename T>
-inline ONNXType TypeInfoImpl<T>::GetONNXType() const {
-  ONNXType out;
-  ThrowOnError(GetApi().GetOnnxTypeFromTypeInfo(this->p_, &out));
-  return out;
-}
-
-}  // namespace detail
-
-namespace detail {
-template <typename T>
-inline TypeInfo SequenceTypeInfoImpl<T>::GetSequenceElementType() const {
-  OrtTypeInfo* output;
-  ThrowOnError(GetApi().GetSequenceElementType(this->p_, &output));
-  return TypeInfo{output};
-}
-
-}  // namespace detail
-
-namespace detail {
-template <typename T>
-inline ONNXTensorElementDataType MapTypeInfoImpl<T>::GetMapKeyType() const {
-  ONNXTensorElementDataType out;
-  ThrowOnError(GetApi().GetMapKeyType(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline TypeInfo MapTypeInfoImpl<T>::GetMapValueType() const {
-  OrtTypeInfo* output;
-  ThrowOnError(GetApi().GetMapValueType(this->p_, &output));
-  return TypeInfo{output};
-}
-}  // namespace detail
-
-namespace detail {
-
-template <typename T>
-template <typename R>
-inline void ConstValueImpl<T>::GetOpaqueData(const char* domain, const char* type_name, R& out) const {
-  ThrowOnError(GetApi().GetOpaqueValue(domain, type_name, this->p_, &out, sizeof(R)));
-}
-
-template <typename T>
-inline bool ConstValueImpl<T>::IsTensor() const {
-  int out;
-  ThrowOnError(GetApi().IsTensor(this->p_, &out));
-  return out != 0;
-}
-
-template <typename T>
-inline bool ConstValueImpl<T>::HasValue() const {
-  int out;
-  ThrowOnError(GetApi().HasValue(this->p_, &out));
-  return out != 0;
-}
-
-template <typename T>
-inline size_t ConstValueImpl<T>::GetCount() const {
-  size_t out;
-  ThrowOnError(GetApi().GetValueCount(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline Value ConstValueImpl<T>::GetValue(int index, OrtAllocator* allocator) const {
-  OrtValue* out;
-  ThrowOnError(GetApi().GetValue(this->p_, index, allocator, &out));
-  return Value{out};
-}
-
-template <typename T>
-inline size_t ConstValueImpl<T>::GetStringTensorDataLength() const {
-  size_t out;
-  ThrowOnError(GetApi().GetStringTensorDataLength(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline size_t ConstValueImpl<T>::GetStringTensorElementLength(size_t element_index) const {
-  size_t out;
-  ThrowOnError(GetApi().GetStringTensorElementLength(this->p_, element_index, &out));
-  return out;
-}
-
-template <typename T>
-template <typename R>
-inline const R* ConstValueImpl<T>::GetTensorData() const {
-  R* out;
-  ThrowOnError(GetApi().GetTensorMutableData(const_cast<OrtValue*>(this->p_), (void**)&out));
-  return out;
-}
-
-template <typename T>
-inline const void* ConstValueImpl<T>::GetTensorRawData() const {
-  void* out;
-  ThrowOnError(GetApi().GetTensorMutableData(const_cast<OrtValue*>(this->p_), &out));
-  return out;
-}
-
-template <typename T>
-inline TypeInfo ConstValueImpl<T>::GetTypeInfo() const {
-  OrtTypeInfo* output;
-  ThrowOnError(GetApi().GetTypeInfo(this->p_, &output));
-  return TypeInfo{output};
-}
-
-template <typename T>
-inline TensorTypeAndShapeInfo ConstValueImpl<T>::GetTensorTypeAndShapeInfo() const {
-  OrtTensorTypeAndShapeInfo* output;
-  ThrowOnError(GetApi().GetTensorTypeAndShape(this->p_, &output));
-  return TensorTypeAndShapeInfo{output};
-}
-
-template <typename T>
-inline ConstMemoryInfo ConstValueImpl<T>::GetTensorMemoryInfo() const {
-  const OrtMemoryInfo* mem_info;
-  ThrowOnError(GetApi().GetTensorMemoryInfo(this->p_, &mem_info));
-  return ConstMemoryInfo(mem_info);
-}
-
-template <typename T>
-inline void ConstValueImpl<T>::GetStringTensorElement(size_t buffer_length, size_t element_index, void* buffer) const {
-  ThrowOnError(GetApi().GetStringTensorElement(this->p_, buffer_length, element_index, buffer));
-}
-
-template <typename T>
-inline void ConstValueImpl<T>::GetStringTensorContent(void* buffer, size_t buffer_length, size_t* offsets, size_t offsets_count) const {
-  ThrowOnError(GetApi().GetStringTensorContent(this->p_, buffer, buffer_length, offsets, offsets_count));
-}
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-template <typename T>
-inline OrtSparseFormat ConstValueImpl<T>::GetSparseFormat() const {
-  OrtSparseFormat format;
-  ThrowOnError(GetApi().GetSparseTensorFormat(this->p_, &format));
-  return format;
-}
-
-template <typename T>
-inline TensorTypeAndShapeInfo ConstValueImpl<T>::GetSparseTensorValuesTypeAndShapeInfo() const {
-  OrtTensorTypeAndShapeInfo* output;
-  ThrowOnError(GetApi().GetSparseTensorValuesTypeAndShape(this->p_, &output));
-  return TensorTypeAndShapeInfo{output};
-}
-
-template <typename T>
-inline TensorTypeAndShapeInfo ConstValueImpl<T>::GetSparseTensorIndicesTypeShapeInfo(OrtSparseIndicesFormat indices_format) const {
-  OrtTensorTypeAndShapeInfo* output;
-  ThrowOnError(GetApi().GetSparseTensorIndicesTypeShape(this->p_, indices_format, &output));
-  return TensorTypeAndShapeInfo{output};
-}
-
-template <typename T>
-template <typename R>
-inline const R* ConstValueImpl<T>::GetSparseTensorIndicesData(OrtSparseIndicesFormat indices_format, size_t& num_indices) const {
-  const void* out;
-  ThrowOnError(GetApi().GetSparseTensorIndices(this->p_, indices_format, &num_indices, &out));
-  return reinterpret_cast<const R*>(out);
-}
-
-template <typename T>
-inline bool ConstValueImpl<T>::IsSparseTensor() const {
-  int out;
-  ThrowOnError(GetApi().IsSparseTensor(this->p_, &out));
-  return out != 0;
-}
-
-template <typename T>
-template <typename R>
-inline const R* ConstValueImpl<T>::GetSparseTensorValues() const {
-  const void* out;
-  ThrowOnError(GetApi().GetSparseTensorValues(this->p_, &out));
-  return reinterpret_cast<const R*>(out);
-}
-
-#endif
-
-template <typename T>
-void ValueImpl<T>::FillStringTensor(const char* const* s, size_t s_len) {
-  ThrowOnError(GetApi().FillStringTensor(this->p_, s, s_len));
-}
-
-template <typename T>
-void ValueImpl<T>::FillStringTensorElement(const char* s, size_t index) {
-  ThrowOnError(GetApi().FillStringTensorElement(this->p_, s, index));
-}
-
-template <typename T>
-void* ValueImpl<T>::GetTensorMutableRawData() {
-  void* out;
-  ThrowOnError(GetApi().GetTensorMutableData(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-template <typename R>
-R* ValueImpl<T>::GetTensorMutableData() {
-  R* out;
-  ThrowOnError(GetApi().GetTensorMutableData(this->p_, (void**)&out));
-  return out;
-}
-
-template <typename T>
-template <typename R>
-R& ValueImpl<T>::At(const std::vector<int64_t>& location) {
-  static_assert(!std::is_same<T, std::string>::value, "this api does not support std::string");
-  R* out;
-  ThrowOnError(GetApi().TensorAt(this->p_, location.data(), location.size(), (void**)&out));
-  return *out;
-}
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-template <typename T>
-void ValueImpl<T>::UseCooIndices(int64_t* indices_data, size_t indices_num) {
-  ThrowOnError(GetApi().UseCooIndices(this->p_, indices_data, indices_num));
-}
-
-template <typename T>
-void ValueImpl<T>::UseCsrIndices(int64_t* inner_data, size_t inner_num, int64_t* outer_data, size_t outer_num) {
-  ThrowOnError(GetApi().UseCsrIndices(this->p_, inner_data, inner_num, outer_data, outer_num));
-}
-
-template <typename T>
-void ValueImpl<T>::UseBlockSparseIndices(const Shape& indices_shape, int32_t* indices_data) {
-  ThrowOnError(GetApi().UseBlockSparseIndices(this->p_, indices_shape.shape, indices_shape.shape_len, indices_data));
-}
-
-template <typename T>
-void ValueImpl<T>::FillSparseTensorCoo(const OrtMemoryInfo* mem_info, const OrtSparseValuesParam& values_param,
-                                       const int64_t* indices_data, size_t indices_num) {
-  ThrowOnError(GetApi().FillSparseTensorCoo(this->p_, mem_info, values_param.values_shape,
-                                            values_param.values_shape_len, values_param.data.p_data,
-                                            indices_data, indices_num));
-}
-
-template <typename T>
-void ValueImpl<T>::FillSparseTensorCsr(const OrtMemoryInfo* data_mem_info,
-                                       const OrtSparseValuesParam& values,
-                                       const int64_t* inner_indices_data, size_t inner_indices_num,
-                                       const int64_t* outer_indices_data, size_t outer_indices_num) {
-  ThrowOnError(GetApi().FillSparseTensorCsr(this->p_, data_mem_info, values.values_shape, values.values_shape_len, values.data.p_data,
-                                            inner_indices_data, inner_indices_num,
-                                            outer_indices_data, outer_indices_num));
-}
-
-template <typename T>
-void ValueImpl<T>::FillSparseTensorBlockSparse(const OrtMemoryInfo* data_mem_info,
-                                               const OrtSparseValuesParam& values,
-                                               const Shape& indices_shape,
-                                               const int32_t* indices_data) {
-  ThrowOnError(GetApi().FillSparseTensorBlockSparse(this->p_, data_mem_info, values.values_shape, values.values_shape_len, values.data.p_data,
-                                                    indices_shape.shape, indices_shape.shape_len,
-                                                    indices_data));
-}
-
-#endif  // !defined(DISABLE_SPARSE_TENSORS)
-
-}  // namespace detail
-
-template <typename T>
-inline Value Value::CreateTensor(const OrtMemoryInfo* info, T* p_data, size_t p_data_element_count, const int64_t* shape, size_t shape_len) {
-  return CreateTensor(info, p_data, p_data_element_count * sizeof(T), shape, shape_len, TypeToTensorType<T>::type);
-}
-
-inline Value Value::CreateTensor(const OrtMemoryInfo* info, void* p_data, size_t p_data_byte_count, const int64_t* shape, size_t shape_len,
-                                 ONNXTensorElementDataType type) {
-  OrtValue* out;
-  ThrowOnError(GetApi().CreateTensorWithDataAsOrtValue(info, p_data, p_data_byte_count, shape, shape_len, type, &out));
-  return Value{out};
-}
-
-template <typename T>
-inline Value Value::CreateTensor(OrtAllocator* allocator, const int64_t* shape, size_t shape_len) {
-  return CreateTensor(allocator, shape, shape_len, TypeToTensorType<T>::type);
-}
-
-inline Value Value::CreateTensor(OrtAllocator* allocator, const int64_t* shape, size_t shape_len, ONNXTensorElementDataType type) {
-  OrtValue* out;
-  ThrowOnError(GetApi().CreateTensorAsOrtValue(allocator, shape, shape_len, type, &out));
-  return Value{out};
-}
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-
-template <typename T>
-inline Value Value::CreateSparseTensor(const OrtMemoryInfo* info, T* p_data, const Shape& dense_shape,
-                                       const Shape& values_shape) {
-  return CreateSparseTensor(info, p_data, dense_shape, values_shape, TypeToTensorType<T>::type);
-}
-
-inline Value Value::CreateSparseTensor(const OrtMemoryInfo* info, void* p_data, const Shape& dense_shape,
-                                       const Shape& values_shape, ONNXTensorElementDataType type) {
-  OrtValue* out;
-  ThrowOnError(GetApi().CreateSparseTensorWithValuesAsOrtValue(info, p_data, dense_shape.shape, dense_shape.shape_len,
-                                                               values_shape.shape, values_shape.shape_len, type, &out));
-  return Value{out};
-}
-
-template <typename T>
-inline Value Value::CreateSparseTensor(OrtAllocator* allocator, const Shape& dense_shape) {
-  return CreateSparseTensor(allocator, dense_shape, TypeToTensorType<T>::type);
-}
-
-inline Value Value::CreateSparseTensor(OrtAllocator* allocator, const Shape& dense_shape,
-                                       ONNXTensorElementDataType type) {
-  OrtValue* out;
-  ThrowOnError(GetApi().CreateSparseTensorAsOrtValue(allocator, dense_shape.shape, dense_shape.shape_len, type, &out));
-  return Value{out};
-}
-#endif  // !defined(DISABLE_SPARSE_TENSORS)
-
-inline Value Value::CreateMap(Value& keys, Value& values) {
-  OrtValue* out;
-  OrtValue* inputs[2] = {keys, values};
-  ThrowOnError(GetApi().CreateValue(inputs, 2, ONNX_TYPE_MAP, &out));
-  return Value{out};
-}
-
-inline Value Value::CreateSequence(std::vector<Value>& values) {
-  OrtValue* out;
-  std::vector<OrtValue*> values_ort{values.data(), values.data() + values.size()};
-  ThrowOnError(GetApi().CreateValue(values_ort.data(), values_ort.size(), ONNX_TYPE_SEQUENCE, &out));
-  return Value{out};
-}
-
-template <typename T>
-inline Value Value::CreateOpaque(const char* domain, const char* type_name, const T& data_container) {
-  OrtValue* out;
-  ThrowOnError(GetApi().CreateOpaqueValue(domain, type_name, &data_container, sizeof(T), &out));
-  return Value{out};
-}
-
-//
-// Custom OP Inlines
-//
-inline KernelContext::KernelContext(OrtKernelContext* context) : ctx_(context) {
-}
-
-inline size_t KernelContext::GetInputCount() const {
-  size_t out = 0;
-  Ort::ThrowOnError(GetApi().KernelContext_GetInputCount(ctx_, &out));
-  return out;
-}
-
-inline size_t KernelContext::GetOutputCount() const {
-  size_t out = 0;
-  Ort::ThrowOnError(GetApi().KernelContext_GetOutputCount(ctx_, &out));
-  return out;
-}
-
-inline ConstValue KernelContext::GetInput(size_t index) const {
-  const OrtValue* out = nullptr;
-  Ort::ThrowOnError(GetApi().KernelContext_GetInput(ctx_, index, &out));
-  return ConstValue{out};
-}
-
-inline UnownedValue KernelContext::GetOutput(size_t index, const int64_t* dim_values, size_t dim_count) const {
-  OrtValue* out = nullptr;
-  Ort::ThrowOnError(GetApi().KernelContext_GetOutput(ctx_, index, dim_values, dim_count, &out));
-  return UnownedValue(out);
-}
-
-inline UnownedValue KernelContext::GetOutput(size_t index, const std::vector<int64_t>& dims) const {
-  OrtValue* out = nullptr;
-  Ort::ThrowOnError(GetApi().KernelContext_GetOutput(ctx_, index, dims.data(), dims.size(), &out));
-  return UnownedValue(out);
-}
-
-inline void* KernelContext::GetGPUComputeStream() const {
-  void* out = nullptr;
-  Ort::ThrowOnError(GetApi().KernelContext_GetGPUComputeStream(ctx_, &out));
-  return out;
-}
-
-inline OpAttr::OpAttr(const char* name, const void* data, int len, OrtOpAttrType type) {
-  Ort::ThrowOnError(GetApi().CreateOpAttr(name, data, len, type, &p_));
-}
-
-namespace detail {
-template <typename T>
-inline KernelInfo KernelInfoImpl<T>::Copy() const {
-  OrtKernelInfo* info_copy = nullptr;
-  Ort::ThrowOnError(GetApi().CopyKernelInfo(this->p_, &info_copy));
-  return KernelInfo{info_copy};
-}
-
-template <typename T>
-inline size_t KernelInfoImpl<T>::GetInputCount() const {
-  size_t out = 0;
-  ThrowOnError(GetApi().KernelInfo_GetInputCount(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline size_t KernelInfoImpl<T>::GetOutputCount() const {
-  size_t out = 0;
-  ThrowOnError(GetApi().KernelInfo_GetOutputCount(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline std::string KernelInfoImpl<T>::GetInputName(size_t index) const {
-  size_t size = 0;
-
-  // Feed nullptr for the data buffer to query the true size of the string value
-  Ort::ThrowOnError(GetApi().KernelInfo_GetInputName(this->p_, index, nullptr, &size));
-
-  std::string out;
-  out.resize(size);
-  Ort::ThrowOnError(GetApi().KernelInfo_GetInputName(this->p_, index, &out[0], &size));
-  out.resize(size - 1);  // remove the terminating character '\0'
-
-  return out;
-}
-
-template <typename T>
-inline std::string KernelInfoImpl<T>::GetOutputName(size_t index) const {
-  size_t size = 0;
-
-  // Feed nullptr for the data buffer to query the true size of the string value
-  Ort::ThrowOnError(GetApi().KernelInfo_GetOutputName(this->p_, index, nullptr, &size));
-
-  std::string out;
-  out.resize(size);
-  Ort::ThrowOnError(GetApi().KernelInfo_GetOutputName(this->p_, index, &out[0], &size));
-  out.resize(size - 1);  // remove the terminating character '\0'
-
-  return out;
-}
-
-template <typename T>
-inline TypeInfo KernelInfoImpl<T>::GetInputTypeInfo(size_t index) const {
-  OrtTypeInfo* out = nullptr;
-  ThrowOnError(GetApi().KernelInfo_GetInputTypeInfo(this->p_, index, &out));
-  return TypeInfo{out};
-}
-
-template <typename T>
-inline TypeInfo KernelInfoImpl<T>::GetOutputTypeInfo(size_t index) const {
-  OrtTypeInfo* out = nullptr;
-  ThrowOnError(GetApi().KernelInfo_GetOutputTypeInfo(this->p_, index, &out));
-  return TypeInfo{out};
-}
-
-template <typename T>
-inline Value KernelInfoImpl<T>::GetTensorAttribute(const char* name, OrtAllocator* allocator) const {
-  OrtValue* out = nullptr;
-  ThrowOnError(GetApi().KernelInfoGetAttribute_tensor(this->p_, name, allocator, &out));
-  return Value{out};
-}
-
-inline void attr_utils::GetAttr(const OrtKernelInfo* p, const char* name, float& out) {
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttribute_float(p, name, &out));
-}
-
-inline void attr_utils::GetAttr(const OrtKernelInfo* p, const char* name, int64_t& out) {
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttribute_int64(p, name, &out));
-}
-
-inline void attr_utils::GetAttr(const OrtKernelInfo* p, const char* name, std::string& result) {
-  size_t size = 0;
-  // Feed nullptr for the data buffer to query the true size of the string attribute
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttribute_string(p, name, nullptr, &size));
-
-  std::string out;
-  out.resize(size);
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttribute_string(p, name, &out[0], &size));
-  out.resize(size - 1);  // remove the terminating character '\0'
-  out.swap(result);
-}
-
-inline void attr_utils::GetAttrs(const OrtKernelInfo* p, const char* name, std::vector<float>& result) {
-  size_t size = 0;
-  // Feed nullptr for the data buffer to query the true size of the attribute
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttributeArray_float(p, name, nullptr, &size));
-
-  std::vector<float> out;
-  out.resize(size);
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttributeArray_float(p, name, out.data(), &size));
-  out.swap(result);
-}
-
-inline void attr_utils::GetAttrs(const OrtKernelInfo* p, const char* name, std::vector<int64_t>& result) {
-  size_t size = 0;
-
-  // Feed nullptr for the data buffer to query the true size of the attribute
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttributeArray_int64(p, name, nullptr, &size));
-
-  std::vector<int64_t> out;
-  out.resize(size);
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttributeArray_int64(p, name, out.data(), &size));
-  out.swap(result);
-}
-}  // namespace detail
-
-inline KernelInfo::KernelInfo(OrtKernelInfo* info) : detail::KernelInfoImpl<OrtKernelInfo>{info} {}
-
-inline Op::Op(OrtOp* p) : Base<OrtOp>(p) {}
-
-inline Op Op::Create(const OrtKernelInfo* info, const char* op_name, const char* domain, int version,
-                     const char** type_constraint_names,
-                     const ONNXTensorElementDataType* type_constraint_values,
-                     size_t type_constraint_count,
-                     const OpAttr* attr_values, size_t attr_count,
-                     size_t input_count, size_t output_count) {
-  static_assert(sizeof(OpAttr) == sizeof(OrtOpAttr*),
-                "OpAttr's is expected to be just an array of OrtOpAttr in memory so we can reinterpret safely");
-  auto attr_input_values = reinterpret_cast<const OrtOpAttr* const*>(attr_values);
-  OrtOp* op;
-  Ort::ThrowOnError(GetApi().CreateOp(info, op_name, domain, version, type_constraint_names, type_constraint_values,
-                                      static_cast<int>(type_constraint_count),
-                                      attr_input_values,
-                                      static_cast<int>(attr_count),
-                                      static_cast<int>(input_count),
-                                      static_cast<int>(output_count), &op));
-  return Op{op};
-}
-
-inline void Op::Invoke(const OrtKernelContext* context,
-                       const Value* input_values,
-                       size_t input_count,
-                       Value* output_values,
-                       size_t output_count) {
-  static_assert(sizeof(Value) == sizeof(OrtValue*),
-                "Value is really just an array of OrtValue* in memory, so we can reinterpret_cast safely");
-  auto ort_input_values = reinterpret_cast<const OrtValue* const*>(input_values);
-  auto ort_output_values = reinterpret_cast<OrtValue**>(output_values);
-  Ort::ThrowOnError(GetApi().InvokeOp(context, p_, ort_input_values, static_cast<int>(input_count),
-                                      ort_output_values, static_cast<int>(output_count)));
-}
-
-inline void Op::Invoke(const OrtKernelContext* context,
-                       const OrtValue* const* input_values,
-                       size_t input_count,
-                       OrtValue* const* output_values,
-                       size_t output_count) {
-  Ort::ThrowOnError(GetApi().InvokeOp(context, p_, input_values, static_cast<int>(input_count),
-                                      output_values, static_cast<int>(output_count)));
-}
-
-inline void CustomOpApi::ThrowOnError(OrtStatus* status) {
-  Ort::ThrowOnError(status);
-}
-
-template <>
-inline float CustomOpApi::KernelInfoGetAttribute<float>(_In_ const OrtKernelInfo* info, _In_ const char* name) {
-  float out;
-  Ort::ThrowOnError(api_.KernelInfoGetAttribute_float(info, name, &out));
-  return out;
-}
-
-template <>
-inline int64_t CustomOpApi::KernelInfoGetAttribute<int64_t>(_In_ const OrtKernelInfo* info, _In_ const char* name) {
-  int64_t out;
-  Ort::ThrowOnError(api_.KernelInfoGetAttribute_int64(info, name, &out));
-  return out;
-}
-
-template <>
-inline std::string CustomOpApi::KernelInfoGetAttribute<std::string>(_In_ const OrtKernelInfo* info, _In_ const char* name) {
-  size_t size = 0;
-  std::string out;
-
-  // Feed nullptr for the data buffer to query the true size of the string attribute
-  OrtStatus* status = api_.KernelInfoGetAttribute_string(info, name, nullptr, &size);
-
-  if (status == nullptr) {
-    out.resize(size);
-    Ort::ThrowOnError(api_.KernelInfoGetAttribute_string(info, name, &out[0], &size));
-    out.resize(size - 1);  // remove the terminating character '\0'
-  } else {
-    Ort::ThrowOnError(status);
-  }
-  return out;
-}
-
-template <>
-inline std::vector<float> CustomOpApi::KernelInfoGetAttribute(_In_ const OrtKernelInfo* info, _In_ const char* name) {
-  size_t size = 0;
-  std::vector<float> out;
-
-  // Feed nullptr for the data buffer to query the true size of the attribute
-  OrtStatus* status = api_.KernelInfoGetAttributeArray_float(info, name, nullptr, &size);
-
-  if (status == nullptr) {
-    out.resize(size);
-    Ort::ThrowOnError(api_.KernelInfoGetAttributeArray_float(info, name, out.data(), &size));
-  } else {
-    Ort::ThrowOnError(status);
-  }
-  return out;
-}
-
-template <>
-inline std::vector<int64_t> CustomOpApi::KernelInfoGetAttribute(_In_ const OrtKernelInfo* info, _In_ const char* name) {
-  size_t size = 0;
-  std::vector<int64_t> out;
-
-  // Feed nullptr for the data buffer to query the true size of the attribute
-  OrtStatus* status = api_.KernelInfoGetAttributeArray_int64(info, name, nullptr, &size);
-
-  if (status == nullptr) {
-    out.resize(size);
-    Ort::ThrowOnError(api_.KernelInfoGetAttributeArray_int64(info, name, out.data(), &size));
-  } else {
-    Ort::ThrowOnError(status);
-  }
-  return out;
-}
-inline OrtTensorTypeAndShapeInfo* CustomOpApi::GetTensorTypeAndShape(_In_ const OrtValue* value) {
-  OrtTensorTypeAndShapeInfo* out;
-  Ort::ThrowOnError(api_.GetTensorTypeAndShape(value, &out));
-  return out;
-}
-
-inline size_t CustomOpApi::GetTensorShapeElementCount(_In_ const OrtTensorTypeAndShapeInfo* info) {
-  size_t out;
-  Ort::ThrowOnError(api_.GetTensorShapeElementCount(info, &out));
-  return out;
-}
-
-inline ONNXTensorElementDataType CustomOpApi::GetTensorElementType(const OrtTensorTypeAndShapeInfo* info) {
-  ONNXTensorElementDataType out;
-  Ort::ThrowOnError(api_.GetTensorElementType(info, &out));
-  return out;
-}
-
-inline size_t CustomOpApi::GetDimensionsCount(_In_ const OrtTensorTypeAndShapeInfo* info) {
-  size_t out;
-  Ort::ThrowOnError(api_.GetDimensionsCount(info, &out));
-  return out;
-}
-
-inline void CustomOpApi::GetDimensions(_In_ const OrtTensorTypeAndShapeInfo* info, _Out_ int64_t* dim_values, size_t dim_values_length) {
-  Ort::ThrowOnError(api_.GetDimensions(info, dim_values, dim_values_length));
-}
-
-inline void CustomOpApi::SetDimensions(OrtTensorTypeAndShapeInfo* info, _In_ const int64_t* dim_values, size_t dim_count) {
-  Ort::ThrowOnError(api_.SetDimensions(info, dim_values, dim_count));
-}
-
-template <typename T>
-inline T* CustomOpApi::GetTensorMutableData(_Inout_ OrtValue* value) {
-  T* data;
-  Ort::ThrowOnError(api_.GetTensorMutableData(value, reinterpret_cast<void**>(&data)));
-  return data;
-}
-
-inline const OrtMemoryInfo* CustomOpApi::GetTensorMemoryInfo(_In_ const OrtValue* value) {
-  const OrtMemoryInfo* mem_info;
-  Ort::ThrowOnError(api_.GetTensorMemoryInfo(value, &mem_info));
-  return mem_info;
-}
-
-template <typename T>
-inline const T* CustomOpApi::GetTensorData(_Inout_ const OrtValue* value) {
-  T* data = nullptr;
-  Ort::ThrowOnError(api_.GetTensorMutableData(const_cast<OrtValue*>(value), reinterpret_cast<void**>(&data)));
-  return data;
-}
-
-inline std::vector<int64_t> CustomOpApi::GetTensorShape(const OrtTensorTypeAndShapeInfo* info) {
-  size_t out;
-  Ort::ThrowOnError(api_.GetDimensionsCount(info, &out));
-  std::vector<int64_t> output(out);
-  Ort::ThrowOnError(api_.GetDimensions(info, output.data(), out));
-  return output;
-}
-
-inline void CustomOpApi::ReleaseTensorTypeAndShapeInfo(OrtTensorTypeAndShapeInfo* input) {
-  api_.ReleaseTensorTypeAndShapeInfo(input);
-}
-
-inline size_t CustomOpApi::KernelContext_GetInputCount(const OrtKernelContext* context) {
-  size_t out;
-  Ort::ThrowOnError(api_.KernelContext_GetInputCount(context, &out));
-  return out;
-}
-
-inline const OrtValue* CustomOpApi::KernelContext_GetInput(const OrtKernelContext* context, _In_ size_t index) {
-  const OrtValue* out;
-  Ort::ThrowOnError(api_.KernelContext_GetInput(context, index, &out));
-  return out;
-}
-
-inline size_t CustomOpApi::KernelContext_GetOutputCount(const OrtKernelContext* context) {
-  size_t out;
-  Ort::ThrowOnError(api_.KernelContext_GetOutputCount(context, &out));
-  return out;
-}
-
-inline OrtValue* CustomOpApi::KernelContext_GetOutput(OrtKernelContext* context, _In_ size_t index,
-                                                      _In_ const int64_t* dim_values, size_t dim_count) {
-  OrtValue* out;
-  Ort::ThrowOnError(api_.KernelContext_GetOutput(context, index, dim_values, dim_count, &out));
-  return out;
-}
-
-inline void* CustomOpApi::KernelContext_GetGPUComputeStream(const OrtKernelContext* context) {
-  void* out;
-  Ort::ThrowOnError(api_.KernelContext_GetGPUComputeStream(context, &out));
-  return out;
-}
-
-inline OrtOpAttr* CustomOpApi::CreateOpAttr(_In_ const char* name,
-                                            _In_ const void* data,
-                                            _In_ int len,
-                                            _In_ OrtOpAttrType type) {
-  OrtOpAttr* op_attr{};
-  Ort::ThrowOnError(api_.CreateOpAttr(name, data, len, type, &op_attr));
-  return op_attr;
-}
-
-inline void CustomOpApi::ReleaseOpAttr(_Frees_ptr_opt_ OrtOpAttr* op_attr) {
-  api_.ReleaseOpAttr(op_attr);
-}
-
-inline OrtOp* CustomOpApi::CreateOp(_In_ const OrtKernelInfo* info,
-                                    _In_ const char* op_name,
-                                    _In_ const char* domain,
-                                    _In_ int version,
-                                    _In_opt_ const char** type_constraint_names,
-                                    _In_opt_ const ONNXTensorElementDataType* type_constraint_values,
-                                    _In_opt_ int type_constraint_count,
-                                    _In_opt_ const OrtOpAttr* const* attr_values,
-                                    _In_opt_ int attr_count,
-                                    _In_ int input_count,
-                                    _In_ int output_count) {
-  OrtOp* ort_op{};
-  Ort::ThrowOnError(api_.CreateOp(info, op_name, domain, version, type_constraint_names, type_constraint_values,
-                                  type_constraint_count, attr_values, attr_count, input_count, output_count, &ort_op));
-  return ort_op;
-}
-
-inline void CustomOpApi::InvokeOp(_In_ const OrtKernelContext* context,
-                                  _In_ const OrtOp* ort_op,
-                                  _In_ const OrtValue* const* input_values,
-                                  _In_ int input_count,
-                                  _Inout_ OrtValue* const* output_values,
-                                  _In_ int output_count) {
-  Ort::ThrowOnError(api_.InvokeOp(context, ort_op, input_values, input_count, output_values, output_count));
-}
-
-inline void CustomOpApi::ReleaseOp(_Frees_ptr_opt_ OrtOp* ort_op) {
-  api_.ReleaseOp(ort_op);
-}
-
-inline OrtKernelInfo* CustomOpApi::CopyKernelInfo(_In_ const OrtKernelInfo* info) {
-  OrtKernelInfo* info_copy{};
-  Ort::ThrowOnError(api_.CopyKernelInfo(info, &info_copy));
-  return info_copy;
-}
-
-inline void CustomOpApi::ReleaseKernelInfo(_Frees_ptr_opt_ OrtKernelInfo* info_copy) {
-  api_.ReleaseKernelInfo(info_copy);
-}
-
-inline std::vector<std::string> GetAvailableProviders() {
-  int len;
-  char** providers;
-  ThrowOnError(GetApi().GetAvailableProviders(&providers, &len));
-  std::vector<std::string> available_providers(providers, providers + len);
-  ThrowOnError(GetApi().ReleaseAvailableProviders(providers, len));
-  return available_providers;
-}
-
-SessionOptions& AddInitializer(const char* name, const OrtValue* ort_val);
-
-template <typename TOp, typename TKernel>
-void CustomOpBase<TOp, TKernel>::GetSessionConfigs(std::unordered_map<std::string, std::string>& out,
-                                                   ConstSessionOptions options) const {
-  const TOp* derived = static_cast<const TOp*>(this);
-  std::vector<std::string> keys = derived->GetSessionConfigKeys();
-
-  out.reserve(keys.size());
-
-  std::string config_entry_key = detail::MakeCustomOpConfigEntryKey(derived->GetName(), "");
-  const size_t prefix_size = config_entry_key.length();
-
-  for (const auto& key : keys) {
-    config_entry_key.resize(prefix_size);
-    config_entry_key.append(key);
-    out[key] = options.GetConfigEntryOrDefault(config_entry_key.c_str(), "");
-  }
-}
-
-}  // namespace Ort
diff --git a/projects/llm_framework/include/onnxruntime/core/session/onnxruntime_run_options_config_keys.h b/projects/llm_framework/include/onnxruntime/core/session/onnxruntime_run_options_config_keys.h
deleted file mode 100644
index 1f5fcd50..00000000
--- a/projects/llm_framework/include/onnxruntime/core/session/onnxruntime_run_options_config_keys.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-/*
- * This file defines RunOptions Config Keys and format of the Config Values.
- *
- * The Naming Convention for a RunOptions Config Key,
- * "[Area][.[SubArea1].[SubArea2]...].[Keyname]"
- * Such as "ep.cuda.use_arena"
- * The Config Key cannot be empty
- * The maximum length of the Config Key is 128
- *
- * The string format of a RunOptions Config Value is defined individually for each Config.
- * The maximum length of the Config Value is 1024
- */
-
-// Key for enabling shrinkages of user listed device memory arenas.
-// Expects a list of semi-colon separated key value pairs separated by colon in the following format:
-// "device_0:device_id_0;device_1:device_id_1"
-// No white-spaces allowed in the provided list string.
-// Currently, the only supported devices are : "cpu", "gpu" (case sensitive).
-// If "cpu" is included in the list, DisableCpuMemArena() API must not be called (i.e.) arena for cpu should be enabled.
-// Example usage: "cpu:0;gpu:0" (or) "gpu:0"
-// By default, the value for this key is empty (i.e.) no memory arenas are shrunk
-static const char* const kOrtRunOptionsConfigEnableMemoryArenaShrinkage = "memory.enable_memory_arena_shrinkage";
-
-// Set to '1' to not synchronize execution providers with CPU at the end of session run.
-// Per default it will be set to '0'
-// Taking CUDA EP as an example, it omit triggering cudaStreamSynchronize on the compute stream.
-static const char* const kOrtRunOptionsConfigDisableSynchronizeExecutionProviders = "disable_synchronize_execution_providers";
diff --git a/projects/llm_framework/include/onnxruntime/core/session/onnxruntime_session_options_config_keys.h b/projects/llm_framework/include/onnxruntime/core/session/onnxruntime_session_options_config_keys.h
deleted file mode 100644
index 92482d71..00000000
--- a/projects/llm_framework/include/onnxruntime/core/session/onnxruntime_session_options_config_keys.h
+++ /dev/null
@@ -1,186 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-/*
- * This file defines SessionOptions Config Keys and format of the Config Values.
- *
- * The Naming Convention for a SessionOptions Config Key,
- * "[Area][.[SubArea1].[SubArea2]...].[Keyname]"
- * Such as "ep.cuda.use_arena"
- * The Config Key cannot be empty
- * The maximum length of the Config Key is 128
- *
- * The string format of a SessionOptions Config Value is defined individually for each Config.
- * The maximum length of the Config Value is 1024
- */
-
-// Key for disable PrePacking,
-// If the config value is set to "1" then the prepacking is disabled, otherwise prepacking is enabled (default value)
-static const char* const kOrtSessionOptionsConfigDisablePrepacking = "session.disable_prepacking";
-
-// A value of "1" means allocators registered in the env will be used. "0" means the allocators created in the session
-// will be used. Use this to override the usage of env allocators on a per session level.
-static const char* const kOrtSessionOptionsConfigUseEnvAllocators = "session.use_env_allocators";
-
-// Set to 'ORT' (case sensitive) to load an ORT format model.
-// If unset, model type will default to ONNX unless inferred from filename ('.ort' == ORT format) or bytes to be ORT
-static const char* const kOrtSessionOptionsConfigLoadModelFormat = "session.load_model_format";
-
-// Set to 'ORT' (case sensitive) to save optimized model in ORT format when SessionOptions.optimized_model_path is set.
-// If unset, format will default to ONNX unless optimized_model_filepath ends in '.ort'.
-static const char* const kOrtSessionOptionsConfigSaveModelFormat = "session.save_model_format";
-
-// If a value is "1", flush-to-zero and denormal-as-zero are applied. The default is "0".
-// When multiple sessions are created, a main thread doesn't override changes from succeeding session options,
-// but threads in session thread pools follow option changes.
-// When ORT runs with OpenMP, the same rule is applied, i.e. the first session option to flush-to-zero and
-// denormal-as-zero is only applied to global OpenMP thread pool, which doesn't support per-session thread pool.
-// Note that an alternative way not using this option at runtime is to train and export a model without denormals
-// and that's recommended because turning this option on may hurt model accuracy.
-static const char* const kOrtSessionOptionsConfigSetDenormalAsZero = "session.set_denormal_as_zero";
-
-// It controls to run quantization model in QDQ (QuantizelinearDeQuantizelinear) format or not.
-// "0": enable. ORT does fusion logic for QDQ format.
-// "1": disable. ORT doesn't do fusion logic for QDQ format.
-// Its default value is "0"
-static const char* const kOrtSessionOptionsDisableQuantQDQ = "session.disable_quant_qdq";
-
-// It controls whether to enable Double QDQ remover and Identical Children Consolidation
-// "0": not to disable. ORT does remove the middle 2 Nodes from a Q->(QD->Q)->QD pairs
-// "1": disable. ORT doesn't remove the middle 2 Nodes from a Q->(QD->Q)->QD pairs
-// Its default value is "0"
-static const char* const kOrtSessionOptionsDisableDoubleQDQRemover = "session.disable_double_qdq_remover";
-
-// If set to "1", enables the removal of QuantizeLinear/DequantizeLinear node pairs once all QDQ handling has been
-// completed. e.g. If after all QDQ handling has completed and we have -> FloatOp -> Q -> DQ -> FloatOp -> the
-// Q -> DQ could potentially be removed. This will provide a performance benefit by avoiding going from float to
-// 8-bit and back to float, but could impact accuracy. The impact on accuracy will be model specific and depend on
-// other factors like whether the model was created using Quantization Aware Training or Post Training Quantization.
-// As such, it's best to test to determine if enabling this works well for your scenario.
-// The default value is "0"
-// Available since version 1.11.
-static const char* const kOrtSessionOptionsEnableQuantQDQCleanup = "session.enable_quant_qdq_cleanup";
-
-// Enable or disable gelu approximation in graph optimization. "0": disable; "1": enable. The default is "0".
-// GeluApproximation has side effects which may change the inference results. It is disabled by default due to this.
-static const char* const kOrtSessionOptionsEnableGeluApproximation = "optimization.enable_gelu_approximation";
-
-#ifdef ENABLE_TRAINING
-// Specifies a list of op types for memory footprint reduction.
-// The value should be a ","-delimited list of pair of
-// <subgraph string : optimization strategy : number of subgraph to apply>.
-// For example, "Gelu+Cast+:1:0,Dropout+:1:1".
-//   A valid "subgraph string" should be one subgraph representation output by ORT graph transformations.
-//   "optimization strategy" currently has valid values: 0 - disabled, 1 - recompute.
-//   "number of subgraph to apply" is used to control how many subgraphs to apply optimization, to avoid "oversaving"
-//   the memory.
-static const char* const kOrtSessionOptionsMemoryOptimizerEnabler = "optimization.enable_memory_optimizer";
-
-// Specifies the level for detecting subgraphs for memory footprint reduction.
-// The value should be an integer. The default value is 0.
-static const char* const kOrtSessionOptionsMemoryOptimizerProbeLevel = "optimization.enable_memory_probe_recompute_level";
-#endif
-
-// Enable or disable using device allocator for allocating initialized tensor memory. "1": enable; "0": disable. The default is "0".
-// Using device allocators means the memory allocation is made using malloc/new.
-static const char* const kOrtSessionOptionsUseDeviceAllocatorForInitializers = "session.use_device_allocator_for_initializers";
-
-// Configure whether to allow the inter_op/intra_op threads spinning a number of times before blocking
-// "0": thread will block if found no job to run
-// "1": default, thread will spin a number of times before blocking
-static const char* const kOrtSessionOptionsConfigAllowInterOpSpinning = "session.inter_op.allow_spinning";
-static const char* const kOrtSessionOptionsConfigAllowIntraOpSpinning = "session.intra_op.allow_spinning";
-
-// Key for using model bytes directly for ORT format
-// If a session is created using an input byte array contains the ORT format model data,
-// By default we will copy the model bytes at the time of session creation to ensure the model bytes
-// buffer is valid.
-// Setting this option to "1" will disable copy the model bytes, and use the model bytes directly. The caller
-// has to guarantee that the model bytes are valid until the ORT session using the model bytes is destroyed.
-static const char* const kOrtSessionOptionsConfigUseORTModelBytesDirectly = "session.use_ort_model_bytes_directly";
-
-/// <summary>
-/// Key for using the ORT format model flatbuffer bytes directly for initializers.
-/// This avoids copying the bytes and reduces peak memory usage during model loading and initialization.
-/// Requires `session.use_ort_model_bytes_directly` to be true.
-/// If set, the flatbuffer bytes provided when creating the InferenceSession MUST remain valid for the entire
-/// duration of the InferenceSession.
-/// </summary>
-static const char* const kOrtSessionOptionsConfigUseORTModelBytesForInitializers =
-    "session.use_ort_model_bytes_for_initializers";
-
-// This should only be specified when exporting an ORT format model for use on a different platform.
-// If the ORT format model will be used on ARM platforms set to "1". For other platforms set to "0"
-// Available since version 1.11.
-static const char* const kOrtSessionOptionsQDQIsInt8Allowed = "session.qdqisint8allowed";
-
-// x64 SSE4.1/AVX2/AVX512(with no VNNI) has overflow problem with quantizied matrix multiplication with U8S8.
-// To avoid this we need to use slower U8U8 matrix multiplication instead. This option, if
-// turned on, use slower U8U8 matrix multiplications. Only effective with AVX2 or AVX512
-// platforms.
-static const char* const kOrtSessionOptionsAvx2PrecisionMode = "session.x64quantprecision";
-
-// Specifies how minimal build graph optimizations are handled in a full build.
-// These optimizations are at the extended level or higher.
-// Possible values and their effects are:
-// "save": Save runtime optimizations when saving an ORT format model.
-// "apply": Only apply optimizations available in a minimal build.
-// ""/<unspecified>: Apply optimizations available in a full build.
-// Available since version 1.11.
-static const char* const kOrtSessionOptionsConfigMinimalBuildOptimizations =
-    "optimization.minimal_build_optimizations";
-
-// Note: The options specific to an EP should be specified prior to appending that EP to the session options object in
-// order for them to take effect.
-
-// Specifies a list of stop op types. Nodes of a type in the stop op types and nodes downstream from them will not be
-// run by the NNAPI EP.
-// The value should be a ","-delimited list of op types. For example, "Add,Sub".
-// If not specified, the default set of stop ops is used. To specify an empty stop ops types list and disable stop op
-// exclusion, set the value to "".
-static const char* const kOrtSessionOptionsConfigNnapiEpPartitioningStopOps = "ep.nnapi.partitioning_stop_ops";
-
-// Enabling dynamic block-sizing for multithreading.
-// With a positive value, thread pool will split a task of N iterations to blocks of size starting from:
-// N / (num_of_threads * dynamic_block_base)
-// As execution progresses, the size will decrease according to the diminishing residual of N,
-// meaning the task will be distributed in smaller granularity for better parallelism.
-// For some models, it helps to reduce the variance of E2E inference latency and boost performance.
-// The feature will not function by default, specify any positive integer, e.g. "4", to enable it.
-// Available since version 1.11.
-static const char* const kOrtSessionOptionsConfigDynamicBlockBase = "session.dynamic_block_base";
-
-// This option allows to decrease CPU usage between infrequent
-// requests and forces any TP threads spinning stop immediately when the last of
-// concurrent Run() call returns.
-// Spinning is restarted on the next Run() call.
-// Applies only to internal thread-pools
-static const char* const kOrtSessionOptionsConfigForceSpinningStop = "session.force_spinning_stop";
-
-// "1": all inconsistencies encountered during shape and type inference
-// will result in failures.
-// "0": in some cases warnings will be logged but processing will continue. The default.
-// May be useful to expose bugs in models.
-static const char* const kOrtSessionOptionsConfigStrictShapeTypeInference = "session.strict_shape_type_inference";
-
-// The file saves configuration for partitioning node among logic streams
-static const char* const kNodePartitionConfigFile = "session.node_partition_config_file";
-
-// This Option allows setting affinities for intra op threads.
-// Affinity string follows format:
-// logical_processor_id,logical_processor_id;logical_processor_id,logical_processor_id
-// Semicolon isolates configurations among threads, while comma split processors where ith thread expected to attach to.
-// e.g.1,2,3;4,5
-// specifies affinities for two threads, with the 1st thread attach to the 1st, 2nd, and 3rd processor, and 2nd thread to the 4th and 5th.
-// To ease the configuration, an "interval" is also allowed:
-// e.g. 1-8;8-16;17-24
-// orders that the 1st thread runs on first eight processors, 2nd thread runs on next eight processors, and so forth.
-// Note:
-// 1. Once set, the number of thread affinities must equal to intra_op_num_threads - 1, since ort does not set affinity on the main thread which
-//    is started and managed by the calling app;
-// 2. For windows, ort will infer the group id from a logical processor id, for example, assuming there are two groups with each has 64 logical processors,
-//    an id of 64 will be inferred as the last processor of the 1st group, while 65 will be interpreted as the 1st processor of the second group.
-//    Hence 64-65 is an invalid configuration, because a windows thread cannot be attached to processors across group boundary.
-static const char* const kOrtSessionOptionsConfigIntraOpThreadAffinities = "session.intra_op_thread_affinities";
diff --git a/projects/llm_framework/include/onnxruntime/core/session/snippets.dox b/projects/llm_framework/include/onnxruntime/core/session/snippets.dox
deleted file mode 100644
index 874a8c07..00000000
--- a/projects/llm_framework/include/onnxruntime/core/session/snippets.dox
+++ /dev/null
@@ -1,8 +0,0 @@
-[OrtStatus Return Value]
-<returns>If no error, nullptr will be returned. If there is an error, a pointer to an ::OrtStatus that contains error details will be returned. Use OrtApi::ReleaseStatus to free this pointer.</returns>
-[OrtStatus Return Value]
-
-[Log Verbosity Level]
-The log verbosity level controls the display of ::ORT_LOGGING_LEVEL_VERBOSE severity log messages at a finer granularity.
-It only has an effect in debug builds. The higher the value, the more verbose the log output will be.
-[Log Verbosity Level]
diff --git a/projects/llm_framework/include/opencv4/opencv2/core.hpp b/projects/llm_framework/include/opencv4/opencv2/core.hpp
deleted file mode 100644
index 70ea4f8c..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core.hpp
+++ /dev/null
@@ -1,3344 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2015, Intel Corporation, all rights reserved.
-// Copyright (C) 2009-2011, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2015, OpenCV Foundation, all rights reserved.
-// Copyright (C) 2015, Itseez Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_HPP
-#define OPENCV_CORE_HPP
-
-#ifndef __cplusplus
-#  error core.hpp header must be compiled as C++
-#endif
-
-#include "opencv2/core/cvdef.h"
-#include "opencv2/core/base.hpp"
-#include "opencv2/core/cvstd.hpp"
-#include "opencv2/core/traits.hpp"
-#include "opencv2/core/matx.hpp"
-#include "opencv2/core/types.hpp"
-#include "opencv2/core/mat.hpp"
-#include "opencv2/core/persistence.hpp"
-
-/**
-@defgroup core Core functionality
-@{
-    @defgroup core_basic Basic structures
-    @defgroup core_c C structures and operations
-    @{
-        @defgroup core_c_glue Connections with C++
-    @}
-    @defgroup core_array Operations on arrays
-    @defgroup core_async Asynchronous API
-    @defgroup core_xml XML/YAML Persistence
-    @defgroup core_cluster Clustering
-    @defgroup core_utils Utility and system functions and macros
-    @{
-        @defgroup core_logging Logging facilities
-        @defgroup core_utils_sse SSE utilities
-        @defgroup core_utils_neon NEON utilities
-        @defgroup core_utils_vsx VSX utilities
-        @defgroup core_utils_softfloat Softfloat support
-        @defgroup core_utils_samples Utility functions for OpenCV samples
-    @}
-    @defgroup core_opengl OpenGL interoperability
-    @defgroup core_ipp Intel IPP Asynchronous C/C++ Converters
-    @defgroup core_optim Optimization Algorithms
-    @defgroup core_directx DirectX interoperability
-    @defgroup core_eigen Eigen support
-    @defgroup core_opencl OpenCL support
-    @defgroup core_va_intel Intel VA-API/OpenCL (CL-VA) interoperability
-    @defgroup core_hal Hardware Acceleration Layer
-    @{
-        @defgroup core_hal_functions Functions
-        @defgroup core_hal_interface Interface
-        @defgroup core_hal_intrin Universal intrinsics
-        @{
-            @defgroup core_hal_intrin_impl Private implementation helpers
-        @}
-        @defgroup core_lowlevel_api Low-level API for external libraries / plugins
-    @}
-    @defgroup core_parallel Parallel Processing
-    @{
-        @defgroup core_parallel_backend Parallel backends API
-    @}
-@}
- */
-
-namespace cv {
-
-//! @addtogroup core_utils
-//! @{
-
-/*! @brief Class passed to an error.
-
-This class encapsulates all or almost all necessary
-information about the error happened in the program. The exception is
-usually constructed and thrown implicitly via CV_Error and CV_Error_ macros.
-@see error
- */
-class CV_EXPORTS Exception : public std::exception
-{
-public:
-    /*!
-     Default constructor
-     */
-    Exception();
-    /*!
-     Full constructor. Normally the constructor is not called explicitly.
-     Instead, the macros CV_Error(), CV_Error_() and CV_Assert() are used.
-    */
-    Exception(int _code, const String& _err, const String& _func, const String& _file, int _line);
-    virtual ~Exception() throw();
-
-    /*!
-     \return the error description and the context as a text string.
-    */
-    virtual const char *what() const throw() CV_OVERRIDE;
-    void formatMessage();
-
-    String msg; ///< the formatted error message
-
-    int code; ///< error code @see CVStatus
-    String err; ///< error description
-    String func; ///< function name. Available only when the compiler supports getting it
-    String file; ///< source file name where the error has occurred
-    int line; ///< line number in the source file where the error has occurred
-};
-
-/*! @brief Signals an error and raises the exception.
-
-By default the function prints information about the error to stderr,
-then it either stops if cv::setBreakOnError() had been called before or raises the exception.
-It is possible to alternate error processing by using #redirectError().
-@param exc the exception raisen.
-@deprecated drop this version
- */
-CV_EXPORTS CV_NORETURN void error(const Exception& exc);
-
-enum SortFlags { SORT_EVERY_ROW    = 0, //!< each matrix row is sorted independently
-                 SORT_EVERY_COLUMN = 1, //!< each matrix column is sorted
-                                        //!< independently; this flag and the previous one are
-                                        //!< mutually exclusive.
-                 SORT_ASCENDING    = 0, //!< each matrix row is sorted in the ascending
-                                        //!< order.
-                 SORT_DESCENDING   = 16 //!< each matrix row is sorted in the
-                                        //!< descending order; this flag and the previous one are also
-                                        //!< mutually exclusive.
-               };
-
-//! @} core_utils
-
-//! @addtogroup core
-//! @{
-
-//! Covariation flags
-enum CovarFlags {
-    /** The output covariance matrix is calculated as:
-       \f[\texttt{scale}   \cdot  [  \texttt{vects}  [0]-  \texttt{mean}  , \texttt{vects}  [1]-  \texttt{mean}  ,...]^T  \cdot  [ \texttt{vects}  [0]- \texttt{mean}  , \texttt{vects}  [1]- \texttt{mean}  ,...],\f]
-       The covariance matrix will be nsamples x nsamples. Such an unusual covariance matrix is used
-       for fast PCA of a set of very large vectors (see, for example, the EigenFaces technique for
-       face recognition). Eigenvalues of this "scrambled" matrix match the eigenvalues of the true
-       covariance matrix. The "true" eigenvectors can be easily calculated from the eigenvectors of
-       the "scrambled" covariance matrix. */
-    COVAR_SCRAMBLED = 0,
-    /**The output covariance matrix is calculated as:
-        \f[\texttt{scale}   \cdot  [  \texttt{vects}  [0]-  \texttt{mean}  , \texttt{vects}  [1]-  \texttt{mean}  ,...]  \cdot  [ \texttt{vects}  [0]- \texttt{mean}  , \texttt{vects}  [1]- \texttt{mean}  ,...]^T,\f]
-        covar will be a square matrix of the same size as the total number of elements in each input
-        vector. One and only one of #COVAR_SCRAMBLED and #COVAR_NORMAL must be specified.*/
-    COVAR_NORMAL    = 1,
-    /** If the flag is specified, the function does not calculate mean from
-        the input vectors but, instead, uses the passed mean vector. This is useful if mean has been
-        pre-calculated or known in advance, or if the covariance matrix is calculated by parts. In
-        this case, mean is not a mean vector of the input sub-set of vectors but rather the mean
-        vector of the whole set.*/
-    COVAR_USE_AVG   = 2,
-    /** If the flag is specified, the covariance matrix is scaled. In the
-        "normal" mode, scale is 1./nsamples . In the "scrambled" mode, scale is the reciprocal of the
-        total number of elements in each input vector. By default (if the flag is not specified), the
-        covariance matrix is not scaled ( scale=1 ).*/
-    COVAR_SCALE     = 4,
-    /** If the flag is
-        specified, all the input vectors are stored as rows of the samples matrix. mean should be a
-        single-row vector in this case.*/
-    COVAR_ROWS      = 8,
-    /** If the flag is
-        specified, all the input vectors are stored as columns of the samples matrix. mean should be a
-        single-column vector in this case.*/
-    COVAR_COLS      = 16
-};
-
-//! @addtogroup core_cluster
-//!  @{
-
-//! k-Means flags
-enum KmeansFlags {
-    /** Select random initial centers in each attempt.*/
-    KMEANS_RANDOM_CENTERS     = 0,
-    /** Use kmeans++ center initialization by Arthur and Vassilvitskii [Arthur2007].*/
-    KMEANS_PP_CENTERS         = 2,
-    /** During the first (and possibly the only) attempt, use the
-        user-supplied labels instead of computing them from the initial centers. For the second and
-        further attempts, use the random or semi-random centers. Use one of KMEANS_\*_CENTERS flag
-        to specify the exact method.*/
-    KMEANS_USE_INITIAL_LABELS = 1
-};
-
-//! @} core_cluster
-
-//! @addtogroup core_array
-//! @{
-
-enum ReduceTypes { REDUCE_SUM = 0, //!< the output is the sum of all rows/columns of the matrix.
-                   REDUCE_AVG = 1, //!< the output is the mean vector of all rows/columns of the matrix.
-                   REDUCE_MAX = 2, //!< the output is the maximum (column/row-wise) of all rows/columns of the matrix.
-                   REDUCE_MIN = 3  //!< the output is the minimum (column/row-wise) of all rows/columns of the matrix.
-                 };
-
-//! @} core_array
-
-/** @brief Swaps two matrices
-*/
-CV_EXPORTS void swap(Mat& a, Mat& b);
-/** @overload */
-CV_EXPORTS void swap( UMat& a, UMat& b );
-
-//! @} core
-
-//! @addtogroup core_array
-//! @{
-
-/** @brief Computes the source location of an extrapolated pixel.
-
-The function computes and returns the coordinate of a donor pixel corresponding to the specified
-extrapolated pixel when using the specified extrapolation border mode. For example, if you use
-cv::BORDER_WRAP mode in the horizontal direction, cv::BORDER_REFLECT_101 in the vertical direction and
-want to compute value of the "virtual" pixel Point(-5, 100) in a floating-point image img , it
-looks like:
-@code{.cpp}
-    float val = img.at<float>(borderInterpolate(100, img.rows, cv::BORDER_REFLECT_101),
-                              borderInterpolate(-5, img.cols, cv::BORDER_WRAP));
-@endcode
-Normally, the function is not called directly. It is used inside filtering functions and also in
-copyMakeBorder.
-@param p 0-based coordinate of the extrapolated pixel along one of the axes, likely \<0 or \>= len
-@param len Length of the array along the corresponding axis.
-@param borderType Border type, one of the #BorderTypes, except for #BORDER_TRANSPARENT and
-#BORDER_ISOLATED . When borderType==#BORDER_CONSTANT , the function always returns -1, regardless
-of p and len.
-
-@sa copyMakeBorder
-*/
-CV_EXPORTS_W int borderInterpolate(int p, int len, int borderType);
-
-/** @example samples/cpp/tutorial_code/ImgTrans/copyMakeBorder_demo.cpp
-An example using copyMakeBorder function.
-Check @ref tutorial_copyMakeBorder "the corresponding tutorial" for more details
-*/
-
-/** @brief Forms a border around an image.
-
-The function copies the source image into the middle of the destination image. The areas to the
-left, to the right, above and below the copied source image will be filled with extrapolated
-pixels. This is not what filtering functions based on it do (they extrapolate pixels on-fly), but
-what other more complex functions, including your own, may do to simplify image boundary handling.
-
-The function supports the mode when src is already in the middle of dst . In this case, the
-function does not copy src itself but simply constructs the border, for example:
-
-@code{.cpp}
-    // let border be the same in all directions
-    int border=2;
-    // constructs a larger image to fit both the image and the border
-    Mat gray_buf(rgb.rows + border*2, rgb.cols + border*2, rgb.depth());
-    // select the middle part of it w/o copying data
-    Mat gray(gray_canvas, Rect(border, border, rgb.cols, rgb.rows));
-    // convert image from RGB to grayscale
-    cvtColor(rgb, gray, COLOR_RGB2GRAY);
-    // form a border in-place
-    copyMakeBorder(gray, gray_buf, border, border,
-                   border, border, BORDER_REPLICATE);
-    // now do some custom filtering ...
-    ...
-@endcode
-@note When the source image is a part (ROI) of a bigger image, the function will try to use the
-pixels outside of the ROI to form a border. To disable this feature and always do extrapolation, as
-if src was not a ROI, use borderType | #BORDER_ISOLATED.
-
-@param src Source image.
-@param dst Destination image of the same type as src and the size Size(src.cols+left+right,
-src.rows+top+bottom) .
-@param top the top pixels
-@param bottom the bottom pixels
-@param left the left pixels
-@param right Parameter specifying how many pixels in each direction from the source image rectangle
-to extrapolate. For example, top=1, bottom=1, left=1, right=1 mean that 1 pixel-wide border needs
-to be built.
-@param borderType Border type. See borderInterpolate for details.
-@param value Border value if borderType==BORDER_CONSTANT .
-
-@sa  borderInterpolate
-*/
-CV_EXPORTS_W void copyMakeBorder(InputArray src, OutputArray dst,
-                                 int top, int bottom, int left, int right,
-                                 int borderType, const Scalar& value = Scalar() );
-
-/** @brief Calculates the per-element sum of two arrays or an array and a scalar.
-
-The function add calculates:
-- Sum of two arrays when both input arrays have the same size and the same number of channels:
-\f[\texttt{dst}(I) =  \texttt{saturate} ( \texttt{src1}(I) +  \texttt{src2}(I)) \quad \texttt{if mask}(I) \ne0\f]
-- Sum of an array and a scalar when src2 is constructed from Scalar or has the same number of
-elements as `src1.channels()`:
-\f[\texttt{dst}(I) =  \texttt{saturate} ( \texttt{src1}(I) +  \texttt{src2} ) \quad \texttt{if mask}(I) \ne0\f]
-- Sum of a scalar and an array when src1 is constructed from Scalar or has the same number of
-elements as `src2.channels()`:
-\f[\texttt{dst}(I) =  \texttt{saturate} ( \texttt{src1} +  \texttt{src2}(I) ) \quad \texttt{if mask}(I) \ne0\f]
-where `I` is a multi-dimensional index of array elements. In case of multi-channel arrays, each
-channel is processed independently.
-
-The first function in the list above can be replaced with matrix expressions:
-@code{.cpp}
-    dst = src1 + src2;
-    dst += src1; // equivalent to add(dst, src1, dst);
-@endcode
-The input arrays and the output array can all have the same or different depths. For example, you
-can add a 16-bit unsigned array to a 8-bit signed array and store the sum as a 32-bit
-floating-point array. Depth of the output array is determined by the dtype parameter. In the second
-and third cases above, as well as in the first case, when src1.depth() == src2.depth(), dtype can
-be set to the default -1. In this case, the output array will have the same depth as the input
-array, be it src1, src2 or both.
-@note Saturation is not applied when the output array has the depth CV_32S. You may even get
-result of an incorrect sign in the case of overflow.
-@param src1 first input array or a scalar.
-@param src2 second input array or a scalar.
-@param dst output array that has the same size and number of channels as the input array(s); the
-depth is defined by dtype or src1/src2.
-@param mask optional operation mask - 8-bit single channel array, that specifies elements of the
-output array to be changed.
-@param dtype optional depth of the output array (see the discussion below).
-@sa subtract, addWeighted, scaleAdd, Mat::convertTo
-*/
-CV_EXPORTS_W void add(InputArray src1, InputArray src2, OutputArray dst,
-                      InputArray mask = noArray(), int dtype = -1);
-
-/** @brief Calculates the per-element difference between two arrays or array and a scalar.
-
-The function subtract calculates:
-- Difference between two arrays, when both input arrays have the same size and the same number of
-channels:
-    \f[\texttt{dst}(I) =  \texttt{saturate} ( \texttt{src1}(I) -  \texttt{src2}(I)) \quad \texttt{if mask}(I) \ne0\f]
-- Difference between an array and a scalar, when src2 is constructed from Scalar or has the same
-number of elements as `src1.channels()`:
-    \f[\texttt{dst}(I) =  \texttt{saturate} ( \texttt{src1}(I) -  \texttt{src2} ) \quad \texttt{if mask}(I) \ne0\f]
-- Difference between a scalar and an array, when src1 is constructed from Scalar or has the same
-number of elements as `src2.channels()`:
-    \f[\texttt{dst}(I) =  \texttt{saturate} ( \texttt{src1} -  \texttt{src2}(I) ) \quad \texttt{if mask}(I) \ne0\f]
-- The reverse difference between a scalar and an array in the case of `SubRS`:
-    \f[\texttt{dst}(I) =  \texttt{saturate} ( \texttt{src2} -  \texttt{src1}(I) ) \quad \texttt{if mask}(I) \ne0\f]
-where I is a multi-dimensional index of array elements. In case of multi-channel arrays, each
-channel is processed independently.
-
-The first function in the list above can be replaced with matrix expressions:
-@code{.cpp}
-    dst = src1 - src2;
-    dst -= src1; // equivalent to subtract(dst, src1, dst);
-@endcode
-The input arrays and the output array can all have the same or different depths. For example, you
-can subtract to 8-bit unsigned arrays and store the difference in a 16-bit signed array. Depth of
-the output array is determined by dtype parameter. In the second and third cases above, as well as
-in the first case, when src1.depth() == src2.depth(), dtype can be set to the default -1. In this
-case the output array will have the same depth as the input array, be it src1, src2 or both.
-@note Saturation is not applied when the output array has the depth CV_32S. You may even get
-result of an incorrect sign in the case of overflow.
-@param src1 first input array or a scalar.
-@param src2 second input array or a scalar.
-@param dst output array of the same size and the same number of channels as the input array.
-@param mask optional operation mask; this is an 8-bit single channel array that specifies elements
-of the output array to be changed.
-@param dtype optional depth of the output array
-@sa  add, addWeighted, scaleAdd, Mat::convertTo
-  */
-CV_EXPORTS_W void subtract(InputArray src1, InputArray src2, OutputArray dst,
-                           InputArray mask = noArray(), int dtype = -1);
-
-
-/** @brief Calculates the per-element scaled product of two arrays.
-
-The function multiply calculates the per-element product of two arrays:
-
-\f[\texttt{dst} (I)= \texttt{saturate} ( \texttt{scale} \cdot \texttt{src1} (I)  \cdot \texttt{src2} (I))\f]
-
-There is also a @ref MatrixExpressions -friendly variant of the first function. See Mat::mul .
-
-For a not-per-element matrix product, see gemm .
-
-@note Saturation is not applied when the output array has the depth
-CV_32S. You may even get result of an incorrect sign in the case of
-overflow.
-@param src1 first input array.
-@param src2 second input array of the same size and the same type as src1.
-@param dst output array of the same size and type as src1.
-@param scale optional scale factor.
-@param dtype optional depth of the output array
-@sa add, subtract, divide, scaleAdd, addWeighted, accumulate, accumulateProduct, accumulateSquare,
-Mat::convertTo
-*/
-CV_EXPORTS_W void multiply(InputArray src1, InputArray src2,
-                           OutputArray dst, double scale = 1, int dtype = -1);
-
-/** @brief Performs per-element division of two arrays or a scalar by an array.
-
-The function cv::divide divides one array by another:
-\f[\texttt{dst(I) = saturate(src1(I)*scale/src2(I))}\f]
-or a scalar by an array when there is no src1 :
-\f[\texttt{dst(I) = saturate(scale/src2(I))}\f]
-
-Different channels of multi-channel arrays are processed independently.
-
-For integer types when src2(I) is zero, dst(I) will also be zero.
-
-@note In case of floating point data there is no special defined behavior for zero src2(I) values.
-Regular floating-point division is used.
-Expect correct IEEE-754 behaviour for floating-point data (with NaN, Inf result values).
-
-@note Saturation is not applied when the output array has the depth CV_32S. You may even get
-result of an incorrect sign in the case of overflow.
-@param src1 first input array.
-@param src2 second input array of the same size and type as src1.
-@param scale scalar factor.
-@param dst output array of the same size and type as src2.
-@param dtype optional depth of the output array; if -1, dst will have depth src2.depth(), but in
-case of an array-by-array division, you can only pass -1 when src1.depth()==src2.depth().
-@sa  multiply, add, subtract
-*/
-CV_EXPORTS_W void divide(InputArray src1, InputArray src2, OutputArray dst,
-                         double scale = 1, int dtype = -1);
-
-/** @overload */
-CV_EXPORTS_W void divide(double scale, InputArray src2,
-                         OutputArray dst, int dtype = -1);
-
-/** @brief Calculates the sum of a scaled array and another array.
-
-The function scaleAdd is one of the classical primitive linear algebra operations, known as DAXPY
-or SAXPY in [BLAS](http://en.wikipedia.org/wiki/Basic_Linear_Algebra_Subprograms). It calculates
-the sum of a scaled array and another array:
-\f[\texttt{dst} (I)= \texttt{scale} \cdot \texttt{src1} (I) +  \texttt{src2} (I)\f]
-The function can also be emulated with a matrix expression, for example:
-@code{.cpp}
-    Mat A(3, 3, CV_64F);
-    ...
-    A.row(0) = A.row(1)*2 + A.row(2);
-@endcode
-@param src1 first input array.
-@param alpha scale factor for the first array.
-@param src2 second input array of the same size and type as src1.
-@param dst output array of the same size and type as src1.
-@sa add, addWeighted, subtract, Mat::dot, Mat::convertTo
-*/
-CV_EXPORTS_W void scaleAdd(InputArray src1, double alpha, InputArray src2, OutputArray dst);
-
-/** @example samples/cpp/tutorial_code/HighGUI/AddingImagesTrackbar.cpp
-Check @ref tutorial_trackbar "the corresponding tutorial" for more details
-*/
-
-/** @brief Calculates the weighted sum of two arrays.
-
-The function addWeighted calculates the weighted sum of two arrays as follows:
-\f[\texttt{dst} (I)= \texttt{saturate} ( \texttt{src1} (I)* \texttt{alpha} +  \texttt{src2} (I)* \texttt{beta} +  \texttt{gamma} )\f]
-where I is a multi-dimensional index of array elements. In case of multi-channel arrays, each
-channel is processed independently.
-The function can be replaced with a matrix expression:
-@code{.cpp}
-    dst = src1*alpha + src2*beta + gamma;
-@endcode
-@note Saturation is not applied when the output array has the depth CV_32S. You may even get
-result of an incorrect sign in the case of overflow.
-@param src1 first input array.
-@param alpha weight of the first array elements.
-@param src2 second input array of the same size and channel number as src1.
-@param beta weight of the second array elements.
-@param gamma scalar added to each sum.
-@param dst output array that has the same size and number of channels as the input arrays.
-@param dtype optional depth of the output array; when both input arrays have the same depth, dtype
-can be set to -1, which will be equivalent to src1.depth().
-@sa  add, subtract, scaleAdd, Mat::convertTo
-*/
-CV_EXPORTS_W void addWeighted(InputArray src1, double alpha, InputArray src2,
-                              double beta, double gamma, OutputArray dst, int dtype = -1);
-
-/** @brief Scales, calculates absolute values, and converts the result to 8-bit.
-
-On each element of the input array, the function convertScaleAbs
-performs three operations sequentially: scaling, taking an absolute
-value, conversion to an unsigned 8-bit type:
-\f[\texttt{dst} (I)= \texttt{saturate\_cast<uchar>} (| \texttt{src} (I)* \texttt{alpha} +  \texttt{beta} |)\f]
-In case of multi-channel arrays, the function processes each channel
-independently. When the output is not 8-bit, the operation can be
-emulated by calling the Mat::convertTo method (or by using matrix
-expressions) and then by calculating an absolute value of the result.
-For example:
-@code{.cpp}
-    Mat_<float> A(30,30);
-    randu(A, Scalar(-100), Scalar(100));
-    Mat_<float> B = A*5 + 3;
-    B = abs(B);
-    // Mat_<float> B = abs(A*5+3) will also do the job,
-    // but it will allocate a temporary matrix
-@endcode
-@param src input array.
-@param dst output array.
-@param alpha optional scale factor.
-@param beta optional delta added to the scaled values.
-@sa  Mat::convertTo, cv::abs(const Mat&)
-*/
-CV_EXPORTS_W void convertScaleAbs(InputArray src, OutputArray dst,
-                                  double alpha = 1, double beta = 0);
-
-/** @brief Converts an array to half precision floating number.
-
-This function converts FP32 (single precision floating point) from/to FP16 (half precision floating point). CV_16S format is used to represent FP16 data.
-There are two use modes (src -> dst): CV_32F -> CV_16S and CV_16S -> CV_32F. The input array has to have type of CV_32F or
-CV_16S to represent the bit depth. If the input array is neither of them, the function will raise an error.
-The format of half precision floating point is defined in IEEE 754-2008.
-
-@param src input array.
-@param dst output array.
-*/
-CV_EXPORTS_W void convertFp16(InputArray src, OutputArray dst);
-
-/** @brief Performs a look-up table transform of an array.
-
-The function LUT fills the output array with values from the look-up table. Indices of the entries
-are taken from the input array. That is, the function processes each element of src as follows:
-\f[\texttt{dst} (I)  \leftarrow \texttt{lut(src(I) + d)}\f]
-where
-\f[d =  \fork{0}{if \(\texttt{src}\) has depth \(\texttt{CV_8U}\)}{128}{if \(\texttt{src}\) has depth \(\texttt{CV_8S}\)}\f]
-@param src input array of 8-bit elements.
-@param lut look-up table of 256 elements; in case of multi-channel input array, the table should
-either have a single channel (in this case the same table is used for all channels) or the same
-number of channels as in the input array.
-@param dst output array of the same size and number of channels as src, and the same depth as lut.
-@sa  convertScaleAbs, Mat::convertTo
-*/
-CV_EXPORTS_W void LUT(InputArray src, InputArray lut, OutputArray dst);
-
-/** @brief Calculates the sum of array elements.
-
-The function cv::sum calculates and returns the sum of array elements,
-independently for each channel.
-@param src input array that must have from 1 to 4 channels.
-@sa  countNonZero, mean, meanStdDev, norm, minMaxLoc, reduce
-*/
-CV_EXPORTS_AS(sumElems) Scalar sum(InputArray src);
-
-/** @brief Counts non-zero array elements.
-
-The function returns the number of non-zero elements in src :
-\f[\sum _{I: \; \texttt{src} (I) \ne0 } 1\f]
-@param src single-channel array.
-@sa  mean, meanStdDev, norm, minMaxLoc, calcCovarMatrix
-*/
-CV_EXPORTS_W int countNonZero( InputArray src );
-
-/** @brief Returns the list of locations of non-zero pixels
-
-Given a binary matrix (likely returned from an operation such
-as threshold(), compare(), >, ==, etc, return all of
-the non-zero indices as a cv::Mat or std::vector<cv::Point> (x,y)
-For example:
-@code{.cpp}
-    cv::Mat binaryImage; // input, binary image
-    cv::Mat locations;   // output, locations of non-zero pixels
-    cv::findNonZero(binaryImage, locations);
-
-    // access pixel coordinates
-    Point pnt = locations.at<Point>(i);
-@endcode
-or
-@code{.cpp}
-    cv::Mat binaryImage; // input, binary image
-    vector<Point> locations;   // output, locations of non-zero pixels
-    cv::findNonZero(binaryImage, locations);
-
-    // access pixel coordinates
-    Point pnt = locations[i];
-@endcode
-@param src single-channel array
-@param idx the output array, type of cv::Mat or std::vector<Point>, corresponding to non-zero indices in the input
-*/
-CV_EXPORTS_W void findNonZero( InputArray src, OutputArray idx );
-
-/** @brief Calculates an average (mean) of array elements.
-
-The function cv::mean calculates the mean value M of array elements,
-independently for each channel, and return it:
-\f[\begin{array}{l} N =  \sum _{I: \; \texttt{mask} (I) \ne 0} 1 \\ M_c =  \left ( \sum _{I: \; \texttt{mask} (I) \ne 0}{ \texttt{mtx} (I)_c} \right )/N \end{array}\f]
-When all the mask elements are 0's, the function returns Scalar::all(0)
-@param src input array that should have from 1 to 4 channels so that the result can be stored in
-Scalar_ .
-@param mask optional operation mask.
-@sa  countNonZero, meanStdDev, norm, minMaxLoc
-*/
-CV_EXPORTS_W Scalar mean(InputArray src, InputArray mask = noArray());
-
-/** Calculates a mean and standard deviation of array elements.
-
-The function cv::meanStdDev calculates the mean and the standard deviation M
-of array elements independently for each channel and returns it via the
-output parameters:
-\f[\begin{array}{l} N =  \sum _{I, \texttt{mask} (I)  \ne 0} 1 \\ \texttt{mean} _c =  \frac{\sum_{ I: \; \texttt{mask}(I) \ne 0} \texttt{src} (I)_c}{N} \\ \texttt{stddev} _c =  \sqrt{\frac{\sum_{ I: \; \texttt{mask}(I) \ne 0} \left ( \texttt{src} (I)_c -  \texttt{mean} _c \right )^2}{N}} \end{array}\f]
-When all the mask elements are 0's, the function returns
-mean=stddev=Scalar::all(0).
-@note The calculated standard deviation is only the diagonal of the
-complete normalized covariance matrix. If the full matrix is needed, you
-can reshape the multi-channel array M x N to the single-channel array
-M\*N x mtx.channels() (only possible when the matrix is continuous) and
-then pass the matrix to calcCovarMatrix .
-@param src input array that should have from 1 to 4 channels so that the results can be stored in
-Scalar_ 's.
-@param mean output parameter: calculated mean value.
-@param stddev output parameter: calculated standard deviation.
-@param mask optional operation mask.
-@sa  countNonZero, mean, norm, minMaxLoc, calcCovarMatrix
-*/
-CV_EXPORTS_W void meanStdDev(InputArray src, OutputArray mean, OutputArray stddev,
-                             InputArray mask=noArray());
-
-/** @brief Calculates the  absolute norm of an array.
-
-This version of #norm calculates the absolute norm of src1. The type of norm to calculate is specified using #NormTypes.
-
-As example for one array consider the function \f$r(x)= \begin{pmatrix} x \\ 1-x \end{pmatrix}, x \in [-1;1]\f$.
-The \f$ L_{1}, L_{2} \f$ and \f$ L_{\infty} \f$ norm for the sample value \f$r(-1) = \begin{pmatrix} -1 \\ 2 \end{pmatrix}\f$
-is calculated as follows
-\f{align*}
-    \| r(-1) \|_{L_1} &= |-1| + |2| = 3 \\
-    \| r(-1) \|_{L_2} &= \sqrt{(-1)^{2} + (2)^{2}} = \sqrt{5} \\
-    \| r(-1) \|_{L_\infty} &= \max(|-1|,|2|) = 2
-\f}
-and for \f$r(0.5) = \begin{pmatrix} 0.5 \\ 0.5 \end{pmatrix}\f$ the calculation is
-\f{align*}
-    \| r(0.5) \|_{L_1} &= |0.5| + |0.5| = 1 \\
-    \| r(0.5) \|_{L_2} &= \sqrt{(0.5)^{2} + (0.5)^{2}} = \sqrt{0.5} \\
-    \| r(0.5) \|_{L_\infty} &= \max(|0.5|,|0.5|) = 0.5.
-\f}
-The following graphic shows all values for the three norm functions \f$\| r(x) \|_{L_1}, \| r(x) \|_{L_2}\f$ and \f$\| r(x) \|_{L_\infty}\f$.
-It is notable that the \f$ L_{1} \f$ norm forms the upper and the \f$ L_{\infty} \f$ norm forms the lower border for the example function \f$ r(x) \f$.
-![Graphs for the different norm functions from the above example](pics/NormTypes_OneArray_1-2-INF.png)
-
-When the mask parameter is specified and it is not empty, the norm is
-
-If normType is not specified, #NORM_L2 is used.
-calculated only over the region specified by the mask.
-
-Multi-channel input arrays are treated as single-channel arrays, that is,
-the results for all channels are combined.
-
-Hamming norms can only be calculated with CV_8U depth arrays.
-
-@param src1 first input array.
-@param normType type of the norm (see #NormTypes).
-@param mask optional operation mask; it must have the same size as src1 and CV_8UC1 type.
-*/
-CV_EXPORTS_W double norm(InputArray src1, int normType = NORM_L2, InputArray mask = noArray());
-
-/** @brief Calculates an absolute difference norm or a relative difference norm.
-
-This version of cv::norm calculates the absolute difference norm
-or the relative difference norm of arrays src1 and src2.
-The type of norm to calculate is specified using #NormTypes.
-
-@param src1 first input array.
-@param src2 second input array of the same size and the same type as src1.
-@param normType type of the norm (see #NormTypes).
-@param mask optional operation mask; it must have the same size as src1 and CV_8UC1 type.
-*/
-CV_EXPORTS_W double norm(InputArray src1, InputArray src2,
-                         int normType = NORM_L2, InputArray mask = noArray());
-/** @overload
-@param src first input array.
-@param normType type of the norm (see #NormTypes).
-*/
-CV_EXPORTS double norm( const SparseMat& src, int normType );
-
-/** @brief Computes the Peak Signal-to-Noise Ratio (PSNR) image quality metric.
-
-This function calculates the Peak Signal-to-Noise Ratio (PSNR) image quality metric in decibels (dB),
-between two input arrays src1 and src2. The arrays must have the same type.
-
-The PSNR is calculated as follows:
-
-\f[
-\texttt{PSNR} = 10 \cdot \log_{10}{\left( \frac{R^2}{MSE} \right) }
-\f]
-
-where R is the maximum integer value of depth (e.g. 255 in the case of CV_8U data)
-and MSE is the mean squared error between the two arrays.
-
-@param src1 first input array.
-@param src2 second input array of the same size as src1.
-@param R the maximum pixel value (255 by default)
-
-  */
-CV_EXPORTS_W double PSNR(InputArray src1, InputArray src2, double R=255.);
-
-/** @brief naive nearest neighbor finder
-
-see http://en.wikipedia.org/wiki/Nearest_neighbor_search
-@todo document
-  */
-CV_EXPORTS_W void batchDistance(InputArray src1, InputArray src2,
-                                OutputArray dist, int dtype, OutputArray nidx,
-                                int normType = NORM_L2, int K = 0,
-                                InputArray mask = noArray(), int update = 0,
-                                bool crosscheck = false);
-
-/** @brief Normalizes the norm or value range of an array.
-
-The function cv::normalize normalizes scale and shift the input array elements so that
-\f[\| \texttt{dst} \| _{L_p}= \texttt{alpha}\f]
-(where p=Inf, 1 or 2) when normType=NORM_INF, NORM_L1, or NORM_L2, respectively; or so that
-\f[\min _I  \texttt{dst} (I)= \texttt{alpha} , \, \, \max _I  \texttt{dst} (I)= \texttt{beta}\f]
-
-when normType=NORM_MINMAX (for dense arrays only). The optional mask specifies a sub-array to be
-normalized. This means that the norm or min-n-max are calculated over the sub-array, and then this
-sub-array is modified to be normalized. If you want to only use the mask to calculate the norm or
-min-max but modify the whole array, you can use norm and Mat::convertTo.
-
-In case of sparse matrices, only the non-zero values are analyzed and transformed. Because of this,
-the range transformation for sparse matrices is not allowed since it can shift the zero level.
-
-Possible usage with some positive example data:
-@code{.cpp}
-    vector<double> positiveData = { 2.0, 8.0, 10.0 };
-    vector<double> normalizedData_l1, normalizedData_l2, normalizedData_inf, normalizedData_minmax;
-
-    // Norm to probability (total count)
-    // sum(numbers) = 20.0
-    // 2.0      0.1     (2.0/20.0)
-    // 8.0      0.4     (8.0/20.0)
-    // 10.0     0.5     (10.0/20.0)
-    normalize(positiveData, normalizedData_l1, 1.0, 0.0, NORM_L1);
-
-    // Norm to unit vector: ||positiveData|| = 1.0
-    // 2.0      0.15
-    // 8.0      0.62
-    // 10.0     0.77
-    normalize(positiveData, normalizedData_l2, 1.0, 0.0, NORM_L2);
-
-    // Norm to max element
-    // 2.0      0.2     (2.0/10.0)
-    // 8.0      0.8     (8.0/10.0)
-    // 10.0     1.0     (10.0/10.0)
-    normalize(positiveData, normalizedData_inf, 1.0, 0.0, NORM_INF);
-
-    // Norm to range [0.0;1.0]
-    // 2.0      0.0     (shift to left border)
-    // 8.0      0.75    (6.0/8.0)
-    // 10.0     1.0     (shift to right border)
-    normalize(positiveData, normalizedData_minmax, 1.0, 0.0, NORM_MINMAX);
-@endcode
-
-@param src input array.
-@param dst output array of the same size as src .
-@param alpha norm value to normalize to or the lower range boundary in case of the range
-normalization.
-@param beta upper range boundary in case of the range normalization; it is not used for the norm
-normalization.
-@param norm_type normalization type (see cv::NormTypes).
-@param dtype when negative, the output array has the same type as src; otherwise, it has the same
-number of channels as src and the depth =CV_MAT_DEPTH(dtype).
-@param mask optional operation mask.
-@sa norm, Mat::convertTo, SparseMat::convertTo
-*/
-CV_EXPORTS_W void normalize( InputArray src, InputOutputArray dst, double alpha = 1, double beta = 0,
-                             int norm_type = NORM_L2, int dtype = -1, InputArray mask = noArray());
-
-/** @overload
-@param src input array.
-@param dst output array of the same size as src .
-@param alpha norm value to normalize to or the lower range boundary in case of the range
-normalization.
-@param normType normalization type (see cv::NormTypes).
-*/
-CV_EXPORTS void normalize( const SparseMat& src, SparseMat& dst, double alpha, int normType );
-
-/** @brief Finds the global minimum and maximum in an array.
-
-The function cv::minMaxLoc finds the minimum and maximum element values and their positions. The
-extremums are searched across the whole array or, if mask is not an empty array, in the specified
-array region.
-
-The function do not work with multi-channel arrays. If you need to find minimum or maximum
-elements across all the channels, use Mat::reshape first to reinterpret the array as
-single-channel. Or you may extract the particular channel using either extractImageCOI , or
-mixChannels , or split .
-@param src input single-channel array.
-@param minVal pointer to the returned minimum value; NULL is used if not required.
-@param maxVal pointer to the returned maximum value; NULL is used if not required.
-@param minLoc pointer to the returned minimum location (in 2D case); NULL is used if not required.
-@param maxLoc pointer to the returned maximum location (in 2D case); NULL is used if not required.
-@param mask optional mask used to select a sub-array.
-@sa max, min, reduceArgMin, reduceArgMax, compare, inRange, extractImageCOI, mixChannels, split, Mat::reshape
-*/
-CV_EXPORTS_W void minMaxLoc(InputArray src, CV_OUT double* minVal,
-                            CV_OUT double* maxVal = 0, CV_OUT Point* minLoc = 0,
-                            CV_OUT Point* maxLoc = 0, InputArray mask = noArray());
-
-/**
- * @brief Finds indices of min elements along provided axis
- *
- * @note
- *      - If input or output array is not continuous, this function will create an internal copy.
- *      - NaN handling is left unspecified, see patchNaNs().
- *      - The returned index is always in bounds of input matrix.
- *
- * @param src input single-channel array.
- * @param dst output array of type CV_32SC1 with the same dimensionality as src,
- * except for axis being reduced - it should be set to 1.
- * @param lastIndex whether to get the index of first or last occurrence of min.
- * @param axis axis to reduce along.
- * @sa reduceArgMax, minMaxLoc, min, max, compare, reduce
- */
-CV_EXPORTS_W void reduceArgMin(InputArray src, OutputArray dst, int axis, bool lastIndex = false);
-
-/**
- * @brief Finds indices of max elements along provided axis
- *
- * @note
- *      - If input or output array is not continuous, this function will create an internal copy.
- *      - NaN handling is left unspecified, see patchNaNs().
- *      - The returned index is always in bounds of input matrix.
- *
- * @param src input single-channel array.
- * @param dst output array of type CV_32SC1 with the same dimensionality as src,
- * except for axis being reduced - it should be set to 1.
- * @param lastIndex whether to get the index of first or last occurrence of max.
- * @param axis axis to reduce along.
- * @sa reduceArgMin, minMaxLoc, min, max, compare, reduce
- */
-CV_EXPORTS_W void reduceArgMax(InputArray src, OutputArray dst, int axis, bool lastIndex = false);
-
-/** @brief Finds the global minimum and maximum in an array
-
-The function cv::minMaxIdx finds the minimum and maximum element values and their positions. The
-extremums are searched across the whole array or, if mask is not an empty array, in the specified
-array region. The function does not work with multi-channel arrays. If you need to find minimum or
-maximum elements across all the channels, use Mat::reshape first to reinterpret the array as
-single-channel. Or you may extract the particular channel using either extractImageCOI , or
-mixChannels , or split . In case of a sparse matrix, the minimum is found among non-zero elements
-only.
-@note When minIdx is not NULL, it must have at least 2 elements (as well as maxIdx), even if src is
-a single-row or single-column matrix. In OpenCV (following MATLAB) each array has at least 2
-dimensions, i.e. single-column matrix is Mx1 matrix (and therefore minIdx/maxIdx will be
-(i1,0)/(i2,0)) and single-row matrix is 1xN matrix (and therefore minIdx/maxIdx will be
-(0,j1)/(0,j2)).
-@param src input single-channel array.
-@param minVal pointer to the returned minimum value; NULL is used if not required.
-@param maxVal pointer to the returned maximum value; NULL is used if not required.
-@param minIdx pointer to the returned minimum location (in nD case); NULL is used if not required;
-Otherwise, it must point to an array of src.dims elements, the coordinates of the minimum element
-in each dimension are stored there sequentially.
-@param maxIdx pointer to the returned maximum location (in nD case). NULL is used if not required.
-@param mask specified array region
-*/
-CV_EXPORTS void minMaxIdx(InputArray src, double* minVal, double* maxVal = 0,
-                          int* minIdx = 0, int* maxIdx = 0, InputArray mask = noArray());
-
-/** @overload
-@param a input single-channel array.
-@param minVal pointer to the returned minimum value; NULL is used if not required.
-@param maxVal pointer to the returned maximum value; NULL is used if not required.
-@param minIdx pointer to the returned minimum location (in nD case); NULL is used if not required;
-Otherwise, it must point to an array of src.dims elements, the coordinates of the minimum element
-in each dimension are stored there sequentially.
-@param maxIdx pointer to the returned maximum location (in nD case). NULL is used if not required.
-*/
-CV_EXPORTS void minMaxLoc(const SparseMat& a, double* minVal,
-                          double* maxVal, int* minIdx = 0, int* maxIdx = 0);
-
-/** @brief Reduces a matrix to a vector.
-
-The function #reduce reduces the matrix to a vector by treating the matrix rows/columns as a set of
-1D vectors and performing the specified operation on the vectors until a single row/column is
-obtained. For example, the function can be used to compute horizontal and vertical projections of a
-raster image. In case of #REDUCE_MAX and #REDUCE_MIN , the output image should have the same type as the source one.
-In case of #REDUCE_SUM and #REDUCE_AVG , the output may have a larger element bit-depth to preserve accuracy.
-And multi-channel arrays are also supported in these two reduction modes.
-
-The following code demonstrates its usage for a single channel matrix.
-@snippet snippets/core_reduce.cpp example
-
-And the following code demonstrates its usage for a two-channel matrix.
-@snippet snippets/core_reduce.cpp example2
-
-@param src input 2D matrix.
-@param dst output vector. Its size and type is defined by dim and dtype parameters.
-@param dim dimension index along which the matrix is reduced. 0 means that the matrix is reduced to
-a single row. 1 means that the matrix is reduced to a single column.
-@param rtype reduction operation that could be one of #ReduceTypes
-@param dtype when negative, the output vector will have the same type as the input matrix,
-otherwise, its type will be CV_MAKE_TYPE(CV_MAT_DEPTH(dtype), src.channels()).
-@sa repeat, reduceArgMin, reduceArgMax
-*/
-CV_EXPORTS_W void reduce(InputArray src, OutputArray dst, int dim, int rtype, int dtype = -1);
-
-/** @brief Creates one multi-channel array out of several single-channel ones.
-
-The function cv::merge merges several arrays to make a single multi-channel array. That is, each
-element of the output array will be a concatenation of the elements of the input arrays, where
-elements of i-th input array are treated as mv[i].channels()-element vectors.
-
-The function cv::split does the reverse operation. If you need to shuffle channels in some other
-advanced way, use cv::mixChannels.
-
-The following example shows how to merge 3 single channel matrices into a single 3-channel matrix.
-@snippet snippets/core_merge.cpp example
-
-@param mv input array of matrices to be merged; all the matrices in mv must have the same
-size and the same depth.
-@param count number of input matrices when mv is a plain C array; it must be greater than zero.
-@param dst output array of the same size and the same depth as mv[0]; The number of channels will
-be equal to the parameter count.
-@sa  mixChannels, split, Mat::reshape
-*/
-CV_EXPORTS void merge(const Mat* mv, size_t count, OutputArray dst);
-
-/** @overload
-@param mv input vector of matrices to be merged; all the matrices in mv must have the same
-size and the same depth.
-@param dst output array of the same size and the same depth as mv[0]; The number of channels will
-be the total number of channels in the matrix array.
-  */
-CV_EXPORTS_W void merge(InputArrayOfArrays mv, OutputArray dst);
-
-/** @brief Divides a multi-channel array into several single-channel arrays.
-
-The function cv::split splits a multi-channel array into separate single-channel arrays:
-\f[\texttt{mv} [c](I) =  \texttt{src} (I)_c\f]
-If you need to extract a single channel or do some other sophisticated channel permutation, use
-mixChannels .
-
-The following example demonstrates how to split a 3-channel matrix into 3 single channel matrices.
-@snippet snippets/core_split.cpp example
-
-@param src input multi-channel array.
-@param mvbegin output array; the number of arrays must match src.channels(); the arrays themselves are
-reallocated, if needed.
-@sa merge, mixChannels, cvtColor
-*/
-CV_EXPORTS void split(const Mat& src, Mat* mvbegin);
-
-/** @overload
-@param m input multi-channel array.
-@param mv output vector of arrays; the arrays themselves are reallocated, if needed.
-*/
-CV_EXPORTS_W void split(InputArray m, OutputArrayOfArrays mv);
-
-/** @brief Copies specified channels from input arrays to the specified channels of
-output arrays.
-
-The function cv::mixChannels provides an advanced mechanism for shuffling image channels.
-
-cv::split,cv::merge,cv::extractChannel,cv::insertChannel and some forms of cv::cvtColor are partial cases of cv::mixChannels.
-
-In the example below, the code splits a 4-channel BGRA image into a 3-channel BGR (with B and R
-channels swapped) and a separate alpha-channel image:
-@code{.cpp}
-    Mat bgra( 100, 100, CV_8UC4, Scalar(255,0,0,255) );
-    Mat bgr( bgra.rows, bgra.cols, CV_8UC3 );
-    Mat alpha( bgra.rows, bgra.cols, CV_8UC1 );
-
-    // forming an array of matrices is a quite efficient operation,
-    // because the matrix data is not copied, only the headers
-    Mat out[] = { bgr, alpha };
-    // bgra[0] -> bgr[2], bgra[1] -> bgr[1],
-    // bgra[2] -> bgr[0], bgra[3] -> alpha[0]
-    int from_to[] = { 0,2, 1,1, 2,0, 3,3 };
-    mixChannels( &bgra, 1, out, 2, from_to, 4 );
-@endcode
-@note Unlike many other new-style C++ functions in OpenCV (see the introduction section and
-Mat::create ), cv::mixChannels requires the output arrays to be pre-allocated before calling the
-function.
-@param src input array or vector of matrices; all of the matrices must have the same size and the
-same depth.
-@param nsrcs number of matrices in `src`.
-@param dst output array or vector of matrices; all the matrices **must be allocated**; their size and
-depth must be the same as in `src[0]`.
-@param ndsts number of matrices in `dst`.
-@param fromTo array of index pairs specifying which channels are copied and where; fromTo[k\*2] is
-a 0-based index of the input channel in src, fromTo[k\*2+1] is an index of the output channel in
-dst; the continuous channel numbering is used: the first input image channels are indexed from 0 to
-src[0].channels()-1, the second input image channels are indexed from src[0].channels() to
-src[0].channels() + src[1].channels()-1, and so on, the same scheme is used for the output image
-channels; as a special case, when fromTo[k\*2] is negative, the corresponding output channel is
-filled with zero .
-@param npairs number of index pairs in `fromTo`.
-@sa split, merge, extractChannel, insertChannel, cvtColor
-*/
-CV_EXPORTS void mixChannels(const Mat* src, size_t nsrcs, Mat* dst, size_t ndsts,
-                            const int* fromTo, size_t npairs);
-
-/** @overload
-@param src input array or vector of matrices; all of the matrices must have the same size and the
-same depth.
-@param dst output array or vector of matrices; all the matrices **must be allocated**; their size and
-depth must be the same as in src[0].
-@param fromTo array of index pairs specifying which channels are copied and where; fromTo[k\*2] is
-a 0-based index of the input channel in src, fromTo[k\*2+1] is an index of the output channel in
-dst; the continuous channel numbering is used: the first input image channels are indexed from 0 to
-src[0].channels()-1, the second input image channels are indexed from src[0].channels() to
-src[0].channels() + src[1].channels()-1, and so on, the same scheme is used for the output image
-channels; as a special case, when fromTo[k\*2] is negative, the corresponding output channel is
-filled with zero .
-@param npairs number of index pairs in fromTo.
-*/
-CV_EXPORTS void mixChannels(InputArrayOfArrays src, InputOutputArrayOfArrays dst,
-                            const int* fromTo, size_t npairs);
-
-/** @overload
-@param src input array or vector of matrices; all of the matrices must have the same size and the
-same depth.
-@param dst output array or vector of matrices; all the matrices **must be allocated**; their size and
-depth must be the same as in src[0].
-@param fromTo array of index pairs specifying which channels are copied and where; fromTo[k\*2] is
-a 0-based index of the input channel in src, fromTo[k\*2+1] is an index of the output channel in
-dst; the continuous channel numbering is used: the first input image channels are indexed from 0 to
-src[0].channels()-1, the second input image channels are indexed from src[0].channels() to
-src[0].channels() + src[1].channels()-1, and so on, the same scheme is used for the output image
-channels; as a special case, when fromTo[k\*2] is negative, the corresponding output channel is
-filled with zero .
-*/
-CV_EXPORTS_W void mixChannels(InputArrayOfArrays src, InputOutputArrayOfArrays dst,
-                              const std::vector<int>& fromTo);
-
-/** @brief Extracts a single channel from src (coi is 0-based index)
-@param src input array
-@param dst output array
-@param coi index of channel to extract
-@sa mixChannels, split
-*/
-CV_EXPORTS_W void extractChannel(InputArray src, OutputArray dst, int coi);
-
-/** @brief Inserts a single channel to dst (coi is 0-based index)
-@param src input array
-@param dst output array
-@param coi index of channel for insertion
-@sa mixChannels, merge
-*/
-CV_EXPORTS_W void insertChannel(InputArray src, InputOutputArray dst, int coi);
-
-/** @brief Flips a 2D array around vertical, horizontal, or both axes.
-
-The function cv::flip flips the array in one of three different ways (row
-and column indices are 0-based):
-\f[\texttt{dst} _{ij} =
-\left\{
-\begin{array}{l l}
-\texttt{src} _{\texttt{src.rows}-i-1,j} & if\;  \texttt{flipCode} = 0 \\
-\texttt{src} _{i, \texttt{src.cols} -j-1} & if\;  \texttt{flipCode} > 0 \\
-\texttt{src} _{ \texttt{src.rows} -i-1, \texttt{src.cols} -j-1} & if\; \texttt{flipCode} < 0 \\
-\end{array}
-\right.\f]
-The example scenarios of using the function are the following:
-*   Vertical flipping of the image (flipCode == 0) to switch between
-    top-left and bottom-left image origin. This is a typical operation
-    in video processing on Microsoft Windows\* OS.
-*   Horizontal flipping of the image with the subsequent horizontal
-    shift and absolute difference calculation to check for a
-    vertical-axis symmetry (flipCode \> 0).
-*   Simultaneous horizontal and vertical flipping of the image with
-    the subsequent shift and absolute difference calculation to check
-    for a central symmetry (flipCode \< 0).
-*   Reversing the order of point arrays (flipCode \> 0 or
-    flipCode == 0).
-@param src input array.
-@param dst output array of the same size and type as src.
-@param flipCode a flag to specify how to flip the array; 0 means
-flipping around the x-axis and positive value (for example, 1) means
-flipping around y-axis. Negative value (for example, -1) means flipping
-around both axes.
-@sa transpose , repeat , completeSymm
-*/
-CV_EXPORTS_W void flip(InputArray src, OutputArray dst, int flipCode);
-
-enum RotateFlags {
-    ROTATE_90_CLOCKWISE = 0, //!<Rotate 90 degrees clockwise
-    ROTATE_180 = 1, //!<Rotate 180 degrees clockwise
-    ROTATE_90_COUNTERCLOCKWISE = 2, //!<Rotate 270 degrees clockwise
-};
-/** @brief Rotates a 2D array in multiples of 90 degrees.
-The function cv::rotate rotates the array in one of three different ways:
-*   Rotate by 90 degrees clockwise (rotateCode = ROTATE_90_CLOCKWISE).
-*   Rotate by 180 degrees clockwise (rotateCode = ROTATE_180).
-*   Rotate by 270 degrees clockwise (rotateCode = ROTATE_90_COUNTERCLOCKWISE).
-@param src input array.
-@param dst output array of the same type as src.  The size is the same with ROTATE_180,
-and the rows and cols are switched for ROTATE_90_CLOCKWISE and ROTATE_90_COUNTERCLOCKWISE.
-@param rotateCode an enum to specify how to rotate the array; see the enum #RotateFlags
-@sa transpose , repeat , completeSymm, flip, RotateFlags
-*/
-CV_EXPORTS_W void rotate(InputArray src, OutputArray dst, int rotateCode);
-
-/** @brief Fills the output array with repeated copies of the input array.
-
-The function cv::repeat duplicates the input array one or more times along each of the two axes:
-\f[\texttt{dst} _{ij}= \texttt{src} _{i\mod src.rows, \; j\mod src.cols }\f]
-The second variant of the function is more convenient to use with @ref MatrixExpressions.
-@param src input array to replicate.
-@param ny Flag to specify how many times the `src` is repeated along the
-vertical axis.
-@param nx Flag to specify how many times the `src` is repeated along the
-horizontal axis.
-@param dst output array of the same type as `src`.
-@sa cv::reduce
-*/
-CV_EXPORTS_W void repeat(InputArray src, int ny, int nx, OutputArray dst);
-
-/** @overload
-@param src input array to replicate.
-@param ny Flag to specify how many times the `src` is repeated along the
-vertical axis.
-@param nx Flag to specify how many times the `src` is repeated along the
-horizontal axis.
-  */
-CV_EXPORTS Mat repeat(const Mat& src, int ny, int nx);
-
-/** @brief Applies horizontal concatenation to given matrices.
-
-The function horizontally concatenates two or more cv::Mat matrices (with the same number of rows).
-@code{.cpp}
-    cv::Mat matArray[] = { cv::Mat(4, 1, CV_8UC1, cv::Scalar(1)),
-                           cv::Mat(4, 1, CV_8UC1, cv::Scalar(2)),
-                           cv::Mat(4, 1, CV_8UC1, cv::Scalar(3)),};
-
-    cv::Mat out;
-    cv::hconcat( matArray, 3, out );
-    //out:
-    //[1, 2, 3;
-    // 1, 2, 3;
-    // 1, 2, 3;
-    // 1, 2, 3]
-@endcode
-@param src input array or vector of matrices. all of the matrices must have the same number of rows and the same depth.
-@param nsrc number of matrices in src.
-@param dst output array. It has the same number of rows and depth as the src, and the sum of cols of the src.
-@sa cv::vconcat(const Mat*, size_t, OutputArray), @sa cv::vconcat(InputArrayOfArrays, OutputArray) and @sa cv::vconcat(InputArray, InputArray, OutputArray)
-*/
-CV_EXPORTS void hconcat(const Mat* src, size_t nsrc, OutputArray dst);
-/** @overload
- @code{.cpp}
-    cv::Mat_<float> A = (cv::Mat_<float>(3, 2) << 1, 4,
-                                                  2, 5,
-                                                  3, 6);
-    cv::Mat_<float> B = (cv::Mat_<float>(3, 2) << 7, 10,
-                                                  8, 11,
-                                                  9, 12);
-
-    cv::Mat C;
-    cv::hconcat(A, B, C);
-    //C:
-    //[1, 4, 7, 10;
-    // 2, 5, 8, 11;
-    // 3, 6, 9, 12]
- @endcode
- @param src1 first input array to be considered for horizontal concatenation.
- @param src2 second input array to be considered for horizontal concatenation.
- @param dst output array. It has the same number of rows and depth as the src1 and src2, and the sum of cols of the src1 and src2.
- */
-CV_EXPORTS void hconcat(InputArray src1, InputArray src2, OutputArray dst);
-/** @overload
- @code{.cpp}
-    std::vector<cv::Mat> matrices = { cv::Mat(4, 1, CV_8UC1, cv::Scalar(1)),
-                                      cv::Mat(4, 1, CV_8UC1, cv::Scalar(2)),
-                                      cv::Mat(4, 1, CV_8UC1, cv::Scalar(3)),};
-
-    cv::Mat out;
-    cv::hconcat( matrices, out );
-    //out:
-    //[1, 2, 3;
-    // 1, 2, 3;
-    // 1, 2, 3;
-    // 1, 2, 3]
- @endcode
- @param src input array or vector of matrices. all of the matrices must have the same number of rows and the same depth.
- @param dst output array. It has the same number of rows and depth as the src, and the sum of cols of the src.
-same depth.
- */
-CV_EXPORTS_W void hconcat(InputArrayOfArrays src, OutputArray dst);
-
-/** @brief Applies vertical concatenation to given matrices.
-
-The function vertically concatenates two or more cv::Mat matrices (with the same number of cols).
-@code{.cpp}
-    cv::Mat matArray[] = { cv::Mat(1, 4, CV_8UC1, cv::Scalar(1)),
-                           cv::Mat(1, 4, CV_8UC1, cv::Scalar(2)),
-                           cv::Mat(1, 4, CV_8UC1, cv::Scalar(3)),};
-
-    cv::Mat out;
-    cv::vconcat( matArray, 3, out );
-    //out:
-    //[1,   1,   1,   1;
-    // 2,   2,   2,   2;
-    // 3,   3,   3,   3]
-@endcode
-@param src input array or vector of matrices. all of the matrices must have the same number of cols and the same depth.
-@param nsrc number of matrices in src.
-@param dst output array. It has the same number of cols and depth as the src, and the sum of rows of the src.
-@sa cv::hconcat(const Mat*, size_t, OutputArray), @sa cv::hconcat(InputArrayOfArrays, OutputArray) and @sa cv::hconcat(InputArray, InputArray, OutputArray)
-*/
-CV_EXPORTS void vconcat(const Mat* src, size_t nsrc, OutputArray dst);
-/** @overload
- @code{.cpp}
-    cv::Mat_<float> A = (cv::Mat_<float>(3, 2) << 1, 7,
-                                                  2, 8,
-                                                  3, 9);
-    cv::Mat_<float> B = (cv::Mat_<float>(3, 2) << 4, 10,
-                                                  5, 11,
-                                                  6, 12);
-
-    cv::Mat C;
-    cv::vconcat(A, B, C);
-    //C:
-    //[1, 7;
-    // 2, 8;
-    // 3, 9;
-    // 4, 10;
-    // 5, 11;
-    // 6, 12]
- @endcode
- @param src1 first input array to be considered for vertical concatenation.
- @param src2 second input array to be considered for vertical concatenation.
- @param dst output array. It has the same number of cols and depth as the src1 and src2, and the sum of rows of the src1 and src2.
- */
-CV_EXPORTS void vconcat(InputArray src1, InputArray src2, OutputArray dst);
-/** @overload
- @code{.cpp}
-    std::vector<cv::Mat> matrices = { cv::Mat(1, 4, CV_8UC1, cv::Scalar(1)),
-                                      cv::Mat(1, 4, CV_8UC1, cv::Scalar(2)),
-                                      cv::Mat(1, 4, CV_8UC1, cv::Scalar(3)),};
-
-    cv::Mat out;
-    cv::vconcat( matrices, out );
-    //out:
-    //[1,   1,   1,   1;
-    // 2,   2,   2,   2;
-    // 3,   3,   3,   3]
- @endcode
- @param src input array or vector of matrices. all of the matrices must have the same number of cols and the same depth
- @param dst output array. It has the same number of cols and depth as the src, and the sum of rows of the src.
-same depth.
- */
-CV_EXPORTS_W void vconcat(InputArrayOfArrays src, OutputArray dst);
-
-/** @brief computes bitwise conjunction of the two arrays (dst = src1 & src2)
-Calculates the per-element bit-wise conjunction of two arrays or an
-array and a scalar.
-
-The function cv::bitwise_and calculates the per-element bit-wise logical conjunction for:
-*   Two arrays when src1 and src2 have the same size:
-    \f[\texttt{dst} (I) =  \texttt{src1} (I)  \wedge \texttt{src2} (I) \quad \texttt{if mask} (I) \ne0\f]
-*   An array and a scalar when src2 is constructed from Scalar or has
-    the same number of elements as `src1.channels()`:
-    \f[\texttt{dst} (I) =  \texttt{src1} (I)  \wedge \texttt{src2} \quad \texttt{if mask} (I) \ne0\f]
-*   A scalar and an array when src1 is constructed from Scalar or has
-    the same number of elements as `src2.channels()`:
-    \f[\texttt{dst} (I) =  \texttt{src1}  \wedge \texttt{src2} (I) \quad \texttt{if mask} (I) \ne0\f]
-In case of floating-point arrays, their machine-specific bit
-representations (usually IEEE754-compliant) are used for the operation.
-In case of multi-channel arrays, each channel is processed
-independently. In the second and third cases above, the scalar is first
-converted to the array type.
-@param src1 first input array or a scalar.
-@param src2 second input array or a scalar.
-@param dst output array that has the same size and type as the input
-arrays.
-@param mask optional operation mask, 8-bit single channel array, that
-specifies elements of the output array to be changed.
-*/
-CV_EXPORTS_W void bitwise_and(InputArray src1, InputArray src2,
-                              OutputArray dst, InputArray mask = noArray());
-
-/** @brief Calculates the per-element bit-wise disjunction of two arrays or an
-array and a scalar.
-
-The function cv::bitwise_or calculates the per-element bit-wise logical disjunction for:
-*   Two arrays when src1 and src2 have the same size:
-    \f[\texttt{dst} (I) =  \texttt{src1} (I)  \vee \texttt{src2} (I) \quad \texttt{if mask} (I) \ne0\f]
-*   An array and a scalar when src2 is constructed from Scalar or has
-    the same number of elements as `src1.channels()`:
-    \f[\texttt{dst} (I) =  \texttt{src1} (I)  \vee \texttt{src2} \quad \texttt{if mask} (I) \ne0\f]
-*   A scalar and an array when src1 is constructed from Scalar or has
-    the same number of elements as `src2.channels()`:
-    \f[\texttt{dst} (I) =  \texttt{src1}  \vee \texttt{src2} (I) \quad \texttt{if mask} (I) \ne0\f]
-In case of floating-point arrays, their machine-specific bit
-representations (usually IEEE754-compliant) are used for the operation.
-In case of multi-channel arrays, each channel is processed
-independently. In the second and third cases above, the scalar is first
-converted to the array type.
-@param src1 first input array or a scalar.
-@param src2 second input array or a scalar.
-@param dst output array that has the same size and type as the input
-arrays.
-@param mask optional operation mask, 8-bit single channel array, that
-specifies elements of the output array to be changed.
-*/
-CV_EXPORTS_W void bitwise_or(InputArray src1, InputArray src2,
-                             OutputArray dst, InputArray mask = noArray());
-
-/** @brief Calculates the per-element bit-wise "exclusive or" operation on two
-arrays or an array and a scalar.
-
-The function cv::bitwise_xor calculates the per-element bit-wise logical "exclusive-or"
-operation for:
-*   Two arrays when src1 and src2 have the same size:
-    \f[\texttt{dst} (I) =  \texttt{src1} (I)  \oplus \texttt{src2} (I) \quad \texttt{if mask} (I) \ne0\f]
-*   An array and a scalar when src2 is constructed from Scalar or has
-    the same number of elements as `src1.channels()`:
-    \f[\texttt{dst} (I) =  \texttt{src1} (I)  \oplus \texttt{src2} \quad \texttt{if mask} (I) \ne0\f]
-*   A scalar and an array when src1 is constructed from Scalar or has
-    the same number of elements as `src2.channels()`:
-    \f[\texttt{dst} (I) =  \texttt{src1}  \oplus \texttt{src2} (I) \quad \texttt{if mask} (I) \ne0\f]
-In case of floating-point arrays, their machine-specific bit
-representations (usually IEEE754-compliant) are used for the operation.
-In case of multi-channel arrays, each channel is processed
-independently. In the 2nd and 3rd cases above, the scalar is first
-converted to the array type.
-@param src1 first input array or a scalar.
-@param src2 second input array or a scalar.
-@param dst output array that has the same size and type as the input
-arrays.
-@param mask optional operation mask, 8-bit single channel array, that
-specifies elements of the output array to be changed.
-*/
-CV_EXPORTS_W void bitwise_xor(InputArray src1, InputArray src2,
-                              OutputArray dst, InputArray mask = noArray());
-
-/** @brief  Inverts every bit of an array.
-
-The function cv::bitwise_not calculates per-element bit-wise inversion of the input
-array:
-\f[\texttt{dst} (I) =  \neg \texttt{src} (I)\f]
-In case of a floating-point input array, its machine-specific bit
-representation (usually IEEE754-compliant) is used for the operation. In
-case of multi-channel arrays, each channel is processed independently.
-@param src input array.
-@param dst output array that has the same size and type as the input
-array.
-@param mask optional operation mask, 8-bit single channel array, that
-specifies elements of the output array to be changed.
-*/
-CV_EXPORTS_W void bitwise_not(InputArray src, OutputArray dst,
-                              InputArray mask = noArray());
-
-/** @brief Calculates the per-element absolute difference between two arrays or between an array and a scalar.
-
-The function cv::absdiff calculates:
-*   Absolute difference between two arrays when they have the same
-    size and type:
-    \f[\texttt{dst}(I) =  \texttt{saturate} (| \texttt{src1}(I) -  \texttt{src2}(I)|)\f]
-*   Absolute difference between an array and a scalar when the second
-    array is constructed from Scalar or has as many elements as the
-    number of channels in `src1`:
-    \f[\texttt{dst}(I) =  \texttt{saturate} (| \texttt{src1}(I) -  \texttt{src2} |)\f]
-*   Absolute difference between a scalar and an array when the first
-    array is constructed from Scalar or has as many elements as the
-    number of channels in `src2`:
-    \f[\texttt{dst}(I) =  \texttt{saturate} (| \texttt{src1} -  \texttt{src2}(I) |)\f]
-    where I is a multi-dimensional index of array elements. In case of
-    multi-channel arrays, each channel is processed independently.
-@note Saturation is not applied when the arrays have the depth CV_32S.
-You may even get a negative value in the case of overflow.
-@param src1 first input array or a scalar.
-@param src2 second input array or a scalar.
-@param dst output array that has the same size and type as input arrays.
-@sa cv::abs(const Mat&)
-*/
-CV_EXPORTS_W void absdiff(InputArray src1, InputArray src2, OutputArray dst);
-
-/** @brief  This is an overloaded member function, provided for convenience (python)
-Copies the matrix to another one.
-When the operation mask is specified, if the Mat::create call shown above reallocates the matrix, the newly allocated matrix is initialized with all zeros before copying the data.
-@param src source matrix.
-@param dst Destination matrix. If it does not have a proper size or type before the operation, it is
-reallocated.
-@param mask Operation mask of the same size as \*this. Its non-zero elements indicate which matrix
-elements need to be copied. The mask has to be of type CV_8U and can have 1 or multiple channels.
-*/
-
-void CV_EXPORTS_W copyTo(InputArray src, OutputArray dst, InputArray mask);
-/** @brief  Checks if array elements lie between the elements of two other arrays.
-
-The function checks the range as follows:
--   For every element of a single-channel input array:
-    \f[\texttt{dst} (I)= \texttt{lowerb} (I)_0  \leq \texttt{src} (I)_0 \leq  \texttt{upperb} (I)_0\f]
--   For two-channel arrays:
-    \f[\texttt{dst} (I)= \texttt{lowerb} (I)_0  \leq \texttt{src} (I)_0 \leq  \texttt{upperb} (I)_0  \land \texttt{lowerb} (I)_1  \leq \texttt{src} (I)_1 \leq  \texttt{upperb} (I)_1\f]
--   and so forth.
-
-That is, dst (I) is set to 255 (all 1 -bits) if src (I) is within the
-specified 1D, 2D, 3D, ... box and 0 otherwise.
-
-When the lower and/or upper boundary parameters are scalars, the indexes
-(I) at lowerb and upperb in the above formulas should be omitted.
-@param src first input array.
-@param lowerb inclusive lower boundary array or a scalar.
-@param upperb inclusive upper boundary array or a scalar.
-@param dst output array of the same size as src and CV_8U type.
-*/
-CV_EXPORTS_W void inRange(InputArray src, InputArray lowerb,
-                          InputArray upperb, OutputArray dst);
-
-/** @brief Performs the per-element comparison of two arrays or an array and scalar value.
-
-The function compares:
-*   Elements of two arrays when src1 and src2 have the same size:
-    \f[\texttt{dst} (I) =  \texttt{src1} (I)  \,\texttt{cmpop}\, \texttt{src2} (I)\f]
-*   Elements of src1 with a scalar src2 when src2 is constructed from
-    Scalar or has a single element:
-    \f[\texttt{dst} (I) =  \texttt{src1}(I) \,\texttt{cmpop}\,  \texttt{src2}\f]
-*   src1 with elements of src2 when src1 is constructed from Scalar or
-    has a single element:
-    \f[\texttt{dst} (I) =  \texttt{src1}  \,\texttt{cmpop}\, \texttt{src2} (I)\f]
-When the comparison result is true, the corresponding element of output
-array is set to 255. The comparison operations can be replaced with the
-equivalent matrix expressions:
-@code{.cpp}
-    Mat dst1 = src1 >= src2;
-    Mat dst2 = src1 < 8;
-    ...
-@endcode
-@param src1 first input array or a scalar; when it is an array, it must have a single channel.
-@param src2 second input array or a scalar; when it is an array, it must have a single channel.
-@param dst output array of type ref CV_8U that has the same size and the same number of channels as
-    the input arrays.
-@param cmpop a flag, that specifies correspondence between the arrays (cv::CmpTypes)
-@sa checkRange, min, max, threshold
-*/
-CV_EXPORTS_W void compare(InputArray src1, InputArray src2, OutputArray dst, int cmpop);
-
-/** @brief Calculates per-element minimum of two arrays or an array and a scalar.
-
-The function cv::min calculates the per-element minimum of two arrays:
-\f[\texttt{dst} (I)= \min ( \texttt{src1} (I), \texttt{src2} (I))\f]
-or array and a scalar:
-\f[\texttt{dst} (I)= \min ( \texttt{src1} (I), \texttt{value} )\f]
-@param src1 first input array.
-@param src2 second input array of the same size and type as src1.
-@param dst output array of the same size and type as src1.
-@sa max, compare, inRange, minMaxLoc
-*/
-CV_EXPORTS_W void min(InputArray src1, InputArray src2, OutputArray dst);
-/** @overload
-needed to avoid conflicts with const _Tp& std::min(const _Tp&, const _Tp&, _Compare)
-*/
-CV_EXPORTS void min(const Mat& src1, const Mat& src2, Mat& dst);
-/** @overload
-needed to avoid conflicts with const _Tp& std::min(const _Tp&, const _Tp&, _Compare)
-*/
-CV_EXPORTS void min(const UMat& src1, const UMat& src2, UMat& dst);
-
-/** @brief Calculates per-element maximum of two arrays or an array and a scalar.
-
-The function cv::max calculates the per-element maximum of two arrays:
-\f[\texttt{dst} (I)= \max ( \texttt{src1} (I), \texttt{src2} (I))\f]
-or array and a scalar:
-\f[\texttt{dst} (I)= \max ( \texttt{src1} (I), \texttt{value} )\f]
-@param src1 first input array.
-@param src2 second input array of the same size and type as src1 .
-@param dst output array of the same size and type as src1.
-@sa  min, compare, inRange, minMaxLoc, @ref MatrixExpressions
-*/
-CV_EXPORTS_W void max(InputArray src1, InputArray src2, OutputArray dst);
-/** @overload
-needed to avoid conflicts with const _Tp& std::min(const _Tp&, const _Tp&, _Compare)
-*/
-CV_EXPORTS void max(const Mat& src1, const Mat& src2, Mat& dst);
-/** @overload
-needed to avoid conflicts with const _Tp& std::min(const _Tp&, const _Tp&, _Compare)
-*/
-CV_EXPORTS void max(const UMat& src1, const UMat& src2, UMat& dst);
-
-/** @brief Calculates a square root of array elements.
-
-The function cv::sqrt calculates a square root of each input array element.
-In case of multi-channel arrays, each channel is processed
-independently. The accuracy is approximately the same as of the built-in
-std::sqrt .
-@param src input floating-point array.
-@param dst output array of the same size and type as src.
-*/
-CV_EXPORTS_W void sqrt(InputArray src, OutputArray dst);
-
-/** @brief Raises every array element to a power.
-
-The function cv::pow raises every element of the input array to power :
-\f[\texttt{dst} (I) =  \fork{\texttt{src}(I)^{power}}{if \(\texttt{power}\) is integer}{|\texttt{src}(I)|^{power}}{otherwise}\f]
-
-So, for a non-integer power exponent, the absolute values of input array
-elements are used. However, it is possible to get true values for
-negative values using some extra operations. In the example below,
-computing the 5th root of array src shows:
-@code{.cpp}
-    Mat mask = src < 0;
-    pow(src, 1./5, dst);
-    subtract(Scalar::all(0), dst, dst, mask);
-@endcode
-For some values of power, such as integer values, 0.5 and -0.5,
-specialized faster algorithms are used.
-
-Special values (NaN, Inf) are not handled.
-@param src input array.
-@param power exponent of power.
-@param dst output array of the same size and type as src.
-@sa sqrt, exp, log, cartToPolar, polarToCart
-*/
-CV_EXPORTS_W void pow(InputArray src, double power, OutputArray dst);
-
-/** @brief Calculates the exponent of every array element.
-
-The function cv::exp calculates the exponent of every element of the input
-array:
-\f[\texttt{dst} [I] = e^{ src(I) }\f]
-
-The maximum relative error is about 7e-6 for single-precision input and
-less than 1e-10 for double-precision input. Currently, the function
-converts denormalized values to zeros on output. Special values (NaN,
-Inf) are not handled.
-@param src input array.
-@param dst output array of the same size and type as src.
-@sa log , cartToPolar , polarToCart , phase , pow , sqrt , magnitude
-*/
-CV_EXPORTS_W void exp(InputArray src, OutputArray dst);
-
-/** @brief Calculates the natural logarithm of every array element.
-
-The function cv::log calculates the natural logarithm of every element of the input array:
-\f[\texttt{dst} (I) =  \log (\texttt{src}(I)) \f]
-
-Output on zero, negative and special (NaN, Inf) values is undefined.
-
-@param src input array.
-@param dst output array of the same size and type as src .
-@sa exp, cartToPolar, polarToCart, phase, pow, sqrt, magnitude
-*/
-CV_EXPORTS_W void log(InputArray src, OutputArray dst);
-
-/** @brief Calculates x and y coordinates of 2D vectors from their magnitude and angle.
-
-The function cv::polarToCart calculates the Cartesian coordinates of each 2D
-vector represented by the corresponding elements of magnitude and angle:
-\f[\begin{array}{l} \texttt{x} (I) =  \texttt{magnitude} (I) \cos ( \texttt{angle} (I)) \\ \texttt{y} (I) =  \texttt{magnitude} (I) \sin ( \texttt{angle} (I)) \\ \end{array}\f]
-
-The relative accuracy of the estimated coordinates is about 1e-6.
-@param magnitude input floating-point array of magnitudes of 2D vectors;
-it can be an empty matrix (=Mat()), in this case, the function assumes
-that all the magnitudes are =1; if it is not empty, it must have the
-same size and type as angle.
-@param angle input floating-point array of angles of 2D vectors.
-@param x output array of x-coordinates of 2D vectors; it has the same
-size and type as angle.
-@param y output array of y-coordinates of 2D vectors; it has the same
-size and type as angle.
-@param angleInDegrees when true, the input angles are measured in
-degrees, otherwise, they are measured in radians.
-@sa cartToPolar, magnitude, phase, exp, log, pow, sqrt
-*/
-CV_EXPORTS_W void polarToCart(InputArray magnitude, InputArray angle,
-                              OutputArray x, OutputArray y, bool angleInDegrees = false);
-
-/** @brief Calculates the magnitude and angle of 2D vectors.
-
-The function cv::cartToPolar calculates either the magnitude, angle, or both
-for every 2D vector (x(I),y(I)):
-\f[\begin{array}{l} \texttt{magnitude} (I)= \sqrt{\texttt{x}(I)^2+\texttt{y}(I)^2} , \\ \texttt{angle} (I)= \texttt{atan2} ( \texttt{y} (I), \texttt{x} (I))[ \cdot180 / \pi ] \end{array}\f]
-
-The angles are calculated with accuracy about 0.3 degrees. For the point
-(0,0), the angle is set to 0.
-@param x array of x-coordinates; this must be a single-precision or
-double-precision floating-point array.
-@param y array of y-coordinates, that must have the same size and same type as x.
-@param magnitude output array of magnitudes of the same size and type as x.
-@param angle output array of angles that has the same size and type as
-x; the angles are measured in radians (from 0 to 2\*Pi) or in degrees (0 to 360 degrees).
-@param angleInDegrees a flag, indicating whether the angles are measured
-in radians (which is by default), or in degrees.
-@sa Sobel, Scharr
-*/
-CV_EXPORTS_W void cartToPolar(InputArray x, InputArray y,
-                              OutputArray magnitude, OutputArray angle,
-                              bool angleInDegrees = false);
-
-/** @brief Calculates the rotation angle of 2D vectors.
-
-The function cv::phase calculates the rotation angle of each 2D vector that
-is formed from the corresponding elements of x and y :
-\f[\texttt{angle} (I) =  \texttt{atan2} ( \texttt{y} (I), \texttt{x} (I))\f]
-
-The angle estimation accuracy is about 0.3 degrees. When x(I)=y(I)=0 ,
-the corresponding angle(I) is set to 0.
-@param x input floating-point array of x-coordinates of 2D vectors.
-@param y input array of y-coordinates of 2D vectors; it must have the
-same size and the same type as x.
-@param angle output array of vector angles; it has the same size and
-same type as x .
-@param angleInDegrees when true, the function calculates the angle in
-degrees, otherwise, they are measured in radians.
-*/
-CV_EXPORTS_W void phase(InputArray x, InputArray y, OutputArray angle,
-                        bool angleInDegrees = false);
-
-/** @brief Calculates the magnitude of 2D vectors.
-
-The function cv::magnitude calculates the magnitude of 2D vectors formed
-from the corresponding elements of x and y arrays:
-\f[\texttt{dst} (I) =  \sqrt{\texttt{x}(I)^2 + \texttt{y}(I)^2}\f]
-@param x floating-point array of x-coordinates of the vectors.
-@param y floating-point array of y-coordinates of the vectors; it must
-have the same size as x.
-@param magnitude output array of the same size and type as x.
-@sa cartToPolar, polarToCart, phase, sqrt
-*/
-CV_EXPORTS_W void magnitude(InputArray x, InputArray y, OutputArray magnitude);
-
-/** @brief Checks every element of an input array for invalid values.
-
-The function cv::checkRange checks that every array element is neither NaN nor infinite. When minVal \>
--DBL_MAX and maxVal \< DBL_MAX, the function also checks that each value is between minVal and
-maxVal. In case of multi-channel arrays, each channel is processed independently. If some values
-are out of range, position of the first outlier is stored in pos (when pos != NULL). Then, the
-function either returns false (when quiet=true) or throws an exception.
-@param a input array.
-@param quiet a flag, indicating whether the functions quietly return false when the array elements
-are out of range or they throw an exception.
-@param pos optional output parameter, when not NULL, must be a pointer to array of src.dims
-elements.
-@param minVal inclusive lower boundary of valid values range.
-@param maxVal exclusive upper boundary of valid values range.
-*/
-CV_EXPORTS_W bool checkRange(InputArray a, bool quiet = true, CV_OUT Point* pos = 0,
-                            double minVal = -DBL_MAX, double maxVal = DBL_MAX);
-
-/** @brief converts NaNs to the given number
-@param a input/output matrix (CV_32F type).
-@param val value to convert the NaNs
-*/
-CV_EXPORTS_W void patchNaNs(InputOutputArray a, double val = 0);
-
-/** @brief Performs generalized matrix multiplication.
-
-The function cv::gemm performs generalized matrix multiplication similar to the
-gemm functions in BLAS level 3. For example,
-`gemm(src1, src2, alpha, src3, beta, dst, GEMM_1_T + GEMM_3_T)`
-corresponds to
-\f[\texttt{dst} =  \texttt{alpha} \cdot \texttt{src1} ^T  \cdot \texttt{src2} +  \texttt{beta} \cdot \texttt{src3} ^T\f]
-
-In case of complex (two-channel) data, performed a complex matrix
-multiplication.
-
-The function can be replaced with a matrix expression. For example, the
-above call can be replaced with:
-@code{.cpp}
-    dst = alpha*src1.t()*src2 + beta*src3.t();
-@endcode
-@param src1 first multiplied input matrix that could be real(CV_32FC1,
-CV_64FC1) or complex(CV_32FC2, CV_64FC2).
-@param src2 second multiplied input matrix of the same type as src1.
-@param alpha weight of the matrix product.
-@param src3 third optional delta matrix added to the matrix product; it
-should have the same type as src1 and src2.
-@param beta weight of src3.
-@param dst output matrix; it has the proper size and the same type as
-input matrices.
-@param flags operation flags (cv::GemmFlags)
-@sa mulTransposed , transform
-*/
-CV_EXPORTS_W void gemm(InputArray src1, InputArray src2, double alpha,
-                       InputArray src3, double beta, OutputArray dst, int flags = 0);
-
-/** @brief Calculates the product of a matrix and its transposition.
-
-The function cv::mulTransposed calculates the product of src and its
-transposition:
-\f[\texttt{dst} = \texttt{scale} ( \texttt{src} - \texttt{delta} )^T ( \texttt{src} - \texttt{delta} )\f]
-if aTa=true , and
-\f[\texttt{dst} = \texttt{scale} ( \texttt{src} - \texttt{delta} ) ( \texttt{src} - \texttt{delta} )^T\f]
-otherwise. The function is used to calculate the covariance matrix. With
-zero delta, it can be used as a faster substitute for general matrix
-product A\*B when B=A'
-@param src input single-channel matrix. Note that unlike gemm, the
-function can multiply not only floating-point matrices.
-@param dst output square matrix.
-@param aTa Flag specifying the multiplication ordering. See the
-description below.
-@param delta Optional delta matrix subtracted from src before the
-multiplication. When the matrix is empty ( delta=noArray() ), it is
-assumed to be zero, that is, nothing is subtracted. If it has the same
-size as src , it is simply subtracted. Otherwise, it is "repeated" (see
-repeat ) to cover the full src and then subtracted. Type of the delta
-matrix, when it is not empty, must be the same as the type of created
-output matrix. See the dtype parameter description below.
-@param scale Optional scale factor for the matrix product.
-@param dtype Optional type of the output matrix. When it is negative,
-the output matrix will have the same type as src . Otherwise, it will be
-type=CV_MAT_DEPTH(dtype) that should be either CV_32F or CV_64F .
-@sa calcCovarMatrix, gemm, repeat, reduce
-*/
-CV_EXPORTS_W void mulTransposed( InputArray src, OutputArray dst, bool aTa,
-                                 InputArray delta = noArray(),
-                                 double scale = 1, int dtype = -1 );
-
-/** @brief Transposes a matrix.
-
-The function cv::transpose transposes the matrix src :
-\f[\texttt{dst} (i,j) =  \texttt{src} (j,i)\f]
-@note No complex conjugation is done in case of a complex matrix. It
-should be done separately if needed.
-@param src input array.
-@param dst output array of the same type as src.
-*/
-CV_EXPORTS_W void transpose(InputArray src, OutputArray dst);
-
-/** @brief Performs the matrix transformation of every array element.
-
-The function cv::transform performs the matrix transformation of every
-element of the array src and stores the results in dst :
-\f[\texttt{dst} (I) =  \texttt{m} \cdot \texttt{src} (I)\f]
-(when m.cols=src.channels() ), or
-\f[\texttt{dst} (I) =  \texttt{m} \cdot [ \texttt{src} (I); 1]\f]
-(when m.cols=src.channels()+1 )
-
-Every element of the N -channel array src is interpreted as N -element
-vector that is transformed using the M x N or M x (N+1) matrix m to
-M-element vector - the corresponding element of the output array dst .
-
-The function may be used for geometrical transformation of
-N -dimensional points, arbitrary linear color space transformation (such
-as various kinds of RGB to YUV transforms), shuffling the image
-channels, and so forth.
-@param src input array that must have as many channels (1 to 4) as
-m.cols or m.cols-1.
-@param dst output array of the same size and depth as src; it has as
-many channels as m.rows.
-@param m transformation 2x2 or 2x3 floating-point matrix.
-@sa perspectiveTransform, getAffineTransform, estimateAffine2D, warpAffine, warpPerspective
-*/
-CV_EXPORTS_W void transform(InputArray src, OutputArray dst, InputArray m );
-
-/** @brief Performs the perspective matrix transformation of vectors.
-
-The function cv::perspectiveTransform transforms every element of src by
-treating it as a 2D or 3D vector, in the following way:
-\f[(x, y, z)  \rightarrow (x'/w, y'/w, z'/w)\f]
-where
-\f[(x', y', z', w') =  \texttt{mat} \cdot \begin{bmatrix} x & y & z & 1  \end{bmatrix}\f]
-and
-\f[w =  \fork{w'}{if \(w' \ne 0\)}{\infty}{otherwise}\f]
-
-Here a 3D vector transformation is shown. In case of a 2D vector
-transformation, the z component is omitted.
-
-@note The function transforms a sparse set of 2D or 3D vectors. If you
-want to transform an image using perspective transformation, use
-warpPerspective . If you have an inverse problem, that is, you want to
-compute the most probable perspective transformation out of several
-pairs of corresponding points, you can use getPerspectiveTransform or
-findHomography .
-@param src input two-channel or three-channel floating-point array; each
-element is a 2D/3D vector to be transformed.
-@param dst output array of the same size and type as src.
-@param m 3x3 or 4x4 floating-point transformation matrix.
-@sa  transform, warpPerspective, getPerspectiveTransform, findHomography
-*/
-CV_EXPORTS_W void perspectiveTransform(InputArray src, OutputArray dst, InputArray m );
-
-/** @brief Copies the lower or the upper half of a square matrix to its another half.
-
-The function cv::completeSymm copies the lower or the upper half of a square matrix to
-its another half. The matrix diagonal remains unchanged:
- - \f$\texttt{m}_{ij}=\texttt{m}_{ji}\f$ for \f$i > j\f$ if
-    lowerToUpper=false
- - \f$\texttt{m}_{ij}=\texttt{m}_{ji}\f$ for \f$i < j\f$ if
-    lowerToUpper=true
-
-@param m input-output floating-point square matrix.
-@param lowerToUpper operation flag; if true, the lower half is copied to
-the upper half. Otherwise, the upper half is copied to the lower half.
-@sa flip, transpose
-*/
-CV_EXPORTS_W void completeSymm(InputOutputArray m, bool lowerToUpper = false);
-
-/** @brief Initializes a scaled identity matrix.
-
-The function cv::setIdentity initializes a scaled identity matrix:
-\f[\texttt{mtx} (i,j)= \fork{\texttt{value}}{ if \(i=j\)}{0}{otherwise}\f]
-
-The function can also be emulated using the matrix initializers and the
-matrix expressions:
-@code
-    Mat A = Mat::eye(4, 3, CV_32F)*5;
-    // A will be set to [[5, 0, 0], [0, 5, 0], [0, 0, 5], [0, 0, 0]]
-@endcode
-@param mtx matrix to initialize (not necessarily square).
-@param s value to assign to diagonal elements.
-@sa Mat::zeros, Mat::ones, Mat::setTo, Mat::operator=
-*/
-CV_EXPORTS_W void setIdentity(InputOutputArray mtx, const Scalar& s = Scalar(1));
-
-/** @brief Returns the determinant of a square floating-point matrix.
-
-The function cv::determinant calculates and returns the determinant of the
-specified matrix. For small matrices ( mtx.cols=mtx.rows\<=3 ), the
-direct method is used. For larger matrices, the function uses LU
-factorization with partial pivoting.
-
-For symmetric positively-determined matrices, it is also possible to use
-eigen decomposition to calculate the determinant.
-@param mtx input matrix that must have CV_32FC1 or CV_64FC1 type and
-square size.
-@sa trace, invert, solve, eigen, @ref MatrixExpressions
-*/
-CV_EXPORTS_W double determinant(InputArray mtx);
-
-/** @brief Returns the trace of a matrix.
-
-The function cv::trace returns the sum of the diagonal elements of the
-matrix mtx .
-\f[\mathrm{tr} ( \texttt{mtx} ) =  \sum _i  \texttt{mtx} (i,i)\f]
-@param mtx input matrix.
-*/
-CV_EXPORTS_W Scalar trace(InputArray mtx);
-
-/** @brief Finds the inverse or pseudo-inverse of a matrix.
-
-The function cv::invert inverts the matrix src and stores the result in dst
-. When the matrix src is singular or non-square, the function calculates
-the pseudo-inverse matrix (the dst matrix) so that norm(src\*dst - I) is
-minimal, where I is an identity matrix.
-
-In case of the #DECOMP_LU method, the function returns non-zero value if
-the inverse has been successfully calculated and 0 if src is singular.
-
-In case of the #DECOMP_SVD method, the function returns the inverse
-condition number of src (the ratio of the smallest singular value to the
-largest singular value) and 0 if src is singular. The SVD method
-calculates a pseudo-inverse matrix if src is singular.
-
-Similarly to #DECOMP_LU, the method #DECOMP_CHOLESKY works only with
-non-singular square matrices that should also be symmetrical and
-positively defined. In this case, the function stores the inverted
-matrix in dst and returns non-zero. Otherwise, it returns 0.
-
-@param src input floating-point M x N matrix.
-@param dst output matrix of N x M size and the same type as src.
-@param flags inversion method (cv::DecompTypes)
-@sa solve, SVD
-*/
-CV_EXPORTS_W double invert(InputArray src, OutputArray dst, int flags = DECOMP_LU);
-
-/** @brief Solves one or more linear systems or least-squares problems.
-
-The function cv::solve solves a linear system or least-squares problem (the
-latter is possible with SVD or QR methods, or by specifying the flag
-#DECOMP_NORMAL ):
-\f[\texttt{dst} =  \arg \min _X \| \texttt{src1} \cdot \texttt{X} -  \texttt{src2} \|\f]
-
-If #DECOMP_LU or #DECOMP_CHOLESKY method is used, the function returns 1
-if src1 (or \f$\texttt{src1}^T\texttt{src1}\f$ ) is non-singular. Otherwise,
-it returns 0. In the latter case, dst is not valid. Other methods find a
-pseudo-solution in case of a singular left-hand side part.
-
-@note If you want to find a unity-norm solution of an under-defined
-singular system \f$\texttt{src1}\cdot\texttt{dst}=0\f$ , the function solve
-will not do the work. Use SVD::solveZ instead.
-
-@param src1 input matrix on the left-hand side of the system.
-@param src2 input matrix on the right-hand side of the system.
-@param dst output solution.
-@param flags solution (matrix inversion) method (#DecompTypes)
-@sa invert, SVD, eigen
-*/
-CV_EXPORTS_W bool solve(InputArray src1, InputArray src2,
-                        OutputArray dst, int flags = DECOMP_LU);
-
-/** @brief Sorts each row or each column of a matrix.
-
-The function cv::sort sorts each matrix row or each matrix column in
-ascending or descending order. So you should pass two operation flags to
-get desired behaviour. If you want to sort matrix rows or columns
-lexicographically, you can use STL std::sort generic function with the
-proper comparison predicate.
-
-@param src input single-channel array.
-@param dst output array of the same size and type as src.
-@param flags operation flags, a combination of #SortFlags
-@sa sortIdx, randShuffle
-*/
-CV_EXPORTS_W void sort(InputArray src, OutputArray dst, int flags);
-
-/** @brief Sorts each row or each column of a matrix.
-
-The function cv::sortIdx sorts each matrix row or each matrix column in the
-ascending or descending order. So you should pass two operation flags to
-get desired behaviour. Instead of reordering the elements themselves, it
-stores the indices of sorted elements in the output array. For example:
-@code
-    Mat A = Mat::eye(3,3,CV_32F), B;
-    sortIdx(A, B, SORT_EVERY_ROW + SORT_ASCENDING);
-    // B will probably contain
-    // (because of equal elements in A some permutations are possible):
-    // [[1, 2, 0], [0, 2, 1], [0, 1, 2]]
-@endcode
-@param src input single-channel array.
-@param dst output integer array of the same size as src.
-@param flags operation flags that could be a combination of cv::SortFlags
-@sa sort, randShuffle
-*/
-CV_EXPORTS_W void sortIdx(InputArray src, OutputArray dst, int flags);
-
-/** @brief Finds the real roots of a cubic equation.
-
-The function solveCubic finds the real roots of a cubic equation:
--   if coeffs is a 4-element vector:
-\f[\texttt{coeffs} [0] x^3 +  \texttt{coeffs} [1] x^2 +  \texttt{coeffs} [2] x +  \texttt{coeffs} [3] = 0\f]
--   if coeffs is a 3-element vector:
-\f[x^3 +  \texttt{coeffs} [0] x^2 +  \texttt{coeffs} [1] x +  \texttt{coeffs} [2] = 0\f]
-
-The roots are stored in the roots array.
-@param coeffs equation coefficients, an array of 3 or 4 elements.
-@param roots output array of real roots that has 1 or 3 elements.
-@return number of real roots. It can be 0, 1 or 2.
-*/
-CV_EXPORTS_W int solveCubic(InputArray coeffs, OutputArray roots);
-
-/** @brief Finds the real or complex roots of a polynomial equation.
-
-The function cv::solvePoly finds real and complex roots of a polynomial equation:
-\f[\texttt{coeffs} [n] x^{n} +  \texttt{coeffs} [n-1] x^{n-1} + ... +  \texttt{coeffs} [1] x +  \texttt{coeffs} [0] = 0\f]
-@param coeffs array of polynomial coefficients.
-@param roots output (complex) array of roots.
-@param maxIters maximum number of iterations the algorithm does.
-*/
-CV_EXPORTS_W double solvePoly(InputArray coeffs, OutputArray roots, int maxIters = 300);
-
-/** @brief Calculates eigenvalues and eigenvectors of a symmetric matrix.
-
-The function cv::eigen calculates just eigenvalues, or eigenvalues and eigenvectors of the symmetric
-matrix src:
-@code
-    src*eigenvectors.row(i).t() = eigenvalues.at<srcType>(i)*eigenvectors.row(i).t()
-@endcode
-
-@note Use cv::eigenNonSymmetric for calculation of real eigenvalues and eigenvectors of non-symmetric matrix.
-
-@param src input matrix that must have CV_32FC1 or CV_64FC1 type, square size and be symmetrical
-(src ^T^ == src).
-@param eigenvalues output vector of eigenvalues of the same type as src; the eigenvalues are stored
-in the descending order.
-@param eigenvectors output matrix of eigenvectors; it has the same size and type as src; the
-eigenvectors are stored as subsequent matrix rows, in the same order as the corresponding
-eigenvalues.
-@sa eigenNonSymmetric, completeSymm , PCA
-*/
-CV_EXPORTS_W bool eigen(InputArray src, OutputArray eigenvalues,
-                        OutputArray eigenvectors = noArray());
-
-/** @brief Calculates eigenvalues and eigenvectors of a non-symmetric matrix (real eigenvalues only).
-
-@note Assumes real eigenvalues.
-
-The function calculates eigenvalues and eigenvectors (optional) of the square matrix src:
-@code
-    src*eigenvectors.row(i).t() = eigenvalues.at<srcType>(i)*eigenvectors.row(i).t()
-@endcode
-
-@param src input matrix (CV_32FC1 or CV_64FC1 type).
-@param eigenvalues output vector of eigenvalues (type is the same type as src).
-@param eigenvectors output matrix of eigenvectors (type is the same type as src). The eigenvectors are stored as subsequent matrix rows, in the same order as the corresponding eigenvalues.
-@sa eigen
-*/
-CV_EXPORTS_W void eigenNonSymmetric(InputArray src, OutputArray eigenvalues,
-                                    OutputArray eigenvectors);
-
-/** @brief Calculates the covariance matrix of a set of vectors.
-
-The function cv::calcCovarMatrix calculates the covariance matrix and, optionally, the mean vector of
-the set of input vectors.
-@param samples samples stored as separate matrices
-@param nsamples number of samples
-@param covar output covariance matrix of the type ctype and square size.
-@param mean input or output (depending on the flags) array as the average value of the input vectors.
-@param flags operation flags as a combination of #CovarFlags
-@param ctype type of the matrixl; it equals 'CV_64F' by default.
-@sa PCA, mulTransposed, Mahalanobis
-@todo InputArrayOfArrays
-*/
-CV_EXPORTS void calcCovarMatrix( const Mat* samples, int nsamples, Mat& covar, Mat& mean,
-                                 int flags, int ctype = CV_64F);
-
-/** @overload
-@note use #COVAR_ROWS or #COVAR_COLS flag
-@param samples samples stored as rows/columns of a single matrix.
-@param covar output covariance matrix of the type ctype and square size.
-@param mean input or output (depending on the flags) array as the average value of the input vectors.
-@param flags operation flags as a combination of #CovarFlags
-@param ctype type of the matrixl; it equals 'CV_64F' by default.
-*/
-CV_EXPORTS_W void calcCovarMatrix( InputArray samples, OutputArray covar,
-                                   InputOutputArray mean, int flags, int ctype = CV_64F);
-
-/** wrap PCA::operator() */
-CV_EXPORTS_W void PCACompute(InputArray data, InputOutputArray mean,
-                             OutputArray eigenvectors, int maxComponents = 0);
-
-/** wrap PCA::operator() and add eigenvalues output parameter */
-CV_EXPORTS_AS(PCACompute2) void PCACompute(InputArray data, InputOutputArray mean,
-                                           OutputArray eigenvectors, OutputArray eigenvalues,
-                                           int maxComponents = 0);
-
-/** wrap PCA::operator() */
-CV_EXPORTS_W void PCACompute(InputArray data, InputOutputArray mean,
-                             OutputArray eigenvectors, double retainedVariance);
-
-/** wrap PCA::operator() and add eigenvalues output parameter */
-CV_EXPORTS_AS(PCACompute2) void PCACompute(InputArray data, InputOutputArray mean,
-                                           OutputArray eigenvectors, OutputArray eigenvalues,
-                                           double retainedVariance);
-
-/** wrap PCA::project */
-CV_EXPORTS_W void PCAProject(InputArray data, InputArray mean,
-                             InputArray eigenvectors, OutputArray result);
-
-/** wrap PCA::backProject */
-CV_EXPORTS_W void PCABackProject(InputArray data, InputArray mean,
-                                 InputArray eigenvectors, OutputArray result);
-
-/** wrap SVD::compute */
-CV_EXPORTS_W void SVDecomp( InputArray src, OutputArray w, OutputArray u, OutputArray vt, int flags = 0 );
-
-/** wrap SVD::backSubst */
-CV_EXPORTS_W void SVBackSubst( InputArray w, InputArray u, InputArray vt,
-                               InputArray rhs, OutputArray dst );
-
-/** @brief Calculates the Mahalanobis distance between two vectors.
-
-The function cv::Mahalanobis calculates and returns the weighted distance between two vectors:
-\f[d( \texttt{vec1} , \texttt{vec2} )= \sqrt{\sum_{i,j}{\texttt{icovar(i,j)}\cdot(\texttt{vec1}(I)-\texttt{vec2}(I))\cdot(\texttt{vec1(j)}-\texttt{vec2(j)})} }\f]
-The covariance matrix may be calculated using the #calcCovarMatrix function and then inverted using
-the invert function (preferably using the #DECOMP_SVD method, as the most accurate).
-@param v1 first 1D input vector.
-@param v2 second 1D input vector.
-@param icovar inverse covariance matrix.
-*/
-CV_EXPORTS_W double Mahalanobis(InputArray v1, InputArray v2, InputArray icovar);
-
-/** @brief Performs a forward or inverse Discrete Fourier transform of a 1D or 2D floating-point array.
-
-The function cv::dft performs one of the following:
--   Forward the Fourier transform of a 1D vector of N elements:
-    \f[Y = F^{(N)}  \cdot X,\f]
-    where \f$F^{(N)}_{jk}=\exp(-2\pi i j k/N)\f$ and \f$i=\sqrt{-1}\f$
--   Inverse the Fourier transform of a 1D vector of N elements:
-    \f[\begin{array}{l} X'=  \left (F^{(N)} \right )^{-1}  \cdot Y =  \left (F^{(N)} \right )^*  \cdot y  \\ X = (1/N)  \cdot X, \end{array}\f]
-    where \f$F^*=\left(\textrm{Re}(F^{(N)})-\textrm{Im}(F^{(N)})\right)^T\f$
--   Forward the 2D Fourier transform of a M x N matrix:
-    \f[Y = F^{(M)}  \cdot X  \cdot F^{(N)}\f]
--   Inverse the 2D Fourier transform of a M x N matrix:
-    \f[\begin{array}{l} X'=  \left (F^{(M)} \right )^*  \cdot Y  \cdot \left (F^{(N)} \right )^* \\ X =  \frac{1}{M \cdot N} \cdot X' \end{array}\f]
-
-In case of real (single-channel) data, the output spectrum of the forward Fourier transform or input
-spectrum of the inverse Fourier transform can be represented in a packed format called *CCS*
-(complex-conjugate-symmetrical). It was borrowed from IPL (Intel\* Image Processing Library). Here
-is how 2D *CCS* spectrum looks:
-\f[\begin{bmatrix} Re Y_{0,0} & Re Y_{0,1} & Im Y_{0,1} & Re Y_{0,2} & Im Y_{0,2} &  \cdots & Re Y_{0,N/2-1} & Im Y_{0,N/2-1} & Re Y_{0,N/2}  \\ Re Y_{1,0} & Re Y_{1,1} & Im Y_{1,1} & Re Y_{1,2} & Im Y_{1,2} &  \cdots & Re Y_{1,N/2-1} & Im Y_{1,N/2-1} & Re Y_{1,N/2}  \\ Im Y_{1,0} & Re Y_{2,1} & Im Y_{2,1} & Re Y_{2,2} & Im Y_{2,2} &  \cdots & Re Y_{2,N/2-1} & Im Y_{2,N/2-1} & Im Y_{1,N/2}  \\ \hdotsfor{9} \\ Re Y_{M/2-1,0} &  Re Y_{M-3,1}  & Im Y_{M-3,1} &  \hdotsfor{3} & Re Y_{M-3,N/2-1} & Im Y_{M-3,N/2-1}& Re Y_{M/2-1,N/2}  \\ Im Y_{M/2-1,0} &  Re Y_{M-2,1}  & Im Y_{M-2,1} &  \hdotsfor{3} & Re Y_{M-2,N/2-1} & Im Y_{M-2,N/2-1}& Im Y_{M/2-1,N/2}  \\ Re Y_{M/2,0}  &  Re Y_{M-1,1} &  Im Y_{M-1,1} &  \hdotsfor{3} & Re Y_{M-1,N/2-1} & Im Y_{M-1,N/2-1}& Re Y_{M/2,N/2} \end{bmatrix}\f]
-
-In case of 1D transform of a real vector, the output looks like the first row of the matrix above.
-
-So, the function chooses an operation mode depending on the flags and size of the input array:
--   If #DFT_ROWS is set or the input array has a single row or single column, the function
-    performs a 1D forward or inverse transform of each row of a matrix when #DFT_ROWS is set.
-    Otherwise, it performs a 2D transform.
--   If the input array is real and #DFT_INVERSE is not set, the function performs a forward 1D or
-    2D transform:
-    -   When #DFT_COMPLEX_OUTPUT is set, the output is a complex matrix of the same size as
-        input.
-    -   When #DFT_COMPLEX_OUTPUT is not set, the output is a real matrix of the same size as
-        input. In case of 2D transform, it uses the packed format as shown above. In case of a
-        single 1D transform, it looks like the first row of the matrix above. In case of
-        multiple 1D transforms (when using the #DFT_ROWS flag), each row of the output matrix
-        looks like the first row of the matrix above.
--   If the input array is complex and either #DFT_INVERSE or #DFT_REAL_OUTPUT are not set, the
-    output is a complex array of the same size as input. The function performs a forward or
-    inverse 1D or 2D transform of the whole input array or each row of the input array
-    independently, depending on the flags DFT_INVERSE and DFT_ROWS.
--   When #DFT_INVERSE is set and the input array is real, or it is complex but #DFT_REAL_OUTPUT
-    is set, the output is a real array of the same size as input. The function performs a 1D or 2D
-    inverse transformation of the whole input array or each individual row, depending on the flags
-    #DFT_INVERSE and #DFT_ROWS.
-
-If #DFT_SCALE is set, the scaling is done after the transformation.
-
-Unlike dct , the function supports arrays of arbitrary size. But only those arrays are processed
-efficiently, whose sizes can be factorized in a product of small prime numbers (2, 3, and 5 in the
-current implementation). Such an efficient DFT size can be calculated using the getOptimalDFTSize
-method.
-
-The sample below illustrates how to calculate a DFT-based convolution of two 2D real arrays:
-@code
-    void convolveDFT(InputArray A, InputArray B, OutputArray C)
-    {
-        // reallocate the output array if needed
-        C.create(abs(A.rows - B.rows)+1, abs(A.cols - B.cols)+1, A.type());
-        Size dftSize;
-        // calculate the size of DFT transform
-        dftSize.width = getOptimalDFTSize(A.cols + B.cols - 1);
-        dftSize.height = getOptimalDFTSize(A.rows + B.rows - 1);
-
-        // allocate temporary buffers and initialize them with 0's
-        Mat tempA(dftSize, A.type(), Scalar::all(0));
-        Mat tempB(dftSize, B.type(), Scalar::all(0));
-
-        // copy A and B to the top-left corners of tempA and tempB, respectively
-        Mat roiA(tempA, Rect(0,0,A.cols,A.rows));
-        A.copyTo(roiA);
-        Mat roiB(tempB, Rect(0,0,B.cols,B.rows));
-        B.copyTo(roiB);
-
-        // now transform the padded A & B in-place;
-        // use "nonzeroRows" hint for faster processing
-        dft(tempA, tempA, 0, A.rows);
-        dft(tempB, tempB, 0, B.rows);
-
-        // multiply the spectrums;
-        // the function handles packed spectrum representations well
-        mulSpectrums(tempA, tempB, tempA);
-
-        // transform the product back from the frequency domain.
-        // Even though all the result rows will be non-zero,
-        // you need only the first C.rows of them, and thus you
-        // pass nonzeroRows == C.rows
-        dft(tempA, tempA, DFT_INVERSE + DFT_SCALE, C.rows);
-
-        // now copy the result back to C.
-        tempA(Rect(0, 0, C.cols, C.rows)).copyTo(C);
-
-        // all the temporary buffers will be deallocated automatically
-    }
-@endcode
-To optimize this sample, consider the following approaches:
--   Since nonzeroRows != 0 is passed to the forward transform calls and since A and B are copied to
-    the top-left corners of tempA and tempB, respectively, it is not necessary to clear the whole
-    tempA and tempB. It is only necessary to clear the tempA.cols - A.cols ( tempB.cols - B.cols)
-    rightmost columns of the matrices.
--   This DFT-based convolution does not have to be applied to the whole big arrays, especially if B
-    is significantly smaller than A or vice versa. Instead, you can calculate convolution by parts.
-    To do this, you need to split the output array C into multiple tiles. For each tile, estimate
-    which parts of A and B are required to calculate convolution in this tile. If the tiles in C are
-    too small, the speed will decrease a lot because of repeated work. In the ultimate case, when
-    each tile in C is a single pixel, the algorithm becomes equivalent to the naive convolution
-    algorithm. If the tiles are too big, the temporary arrays tempA and tempB become too big and
-    there is also a slowdown because of bad cache locality. So, there is an optimal tile size
-    somewhere in the middle.
--   If different tiles in C can be calculated in parallel and, thus, the convolution is done by
-    parts, the loop can be threaded.
-
-All of the above improvements have been implemented in #matchTemplate and #filter2D . Therefore, by
-using them, you can get the performance even better than with the above theoretically optimal
-implementation. Though, those two functions actually calculate cross-correlation, not convolution,
-so you need to "flip" the second convolution operand B vertically and horizontally using flip .
-@note
--   An example using the discrete fourier transform can be found at
-    opencv_source_code/samples/cpp/dft.cpp
--   (Python) An example using the dft functionality to perform Wiener deconvolution can be found
-    at opencv_source/samples/python/deconvolution.py
--   (Python) An example rearranging the quadrants of a Fourier image can be found at
-    opencv_source/samples/python/dft.py
-@param src input array that could be real or complex.
-@param dst output array whose size and type depends on the flags .
-@param flags transformation flags, representing a combination of the #DftFlags
-@param nonzeroRows when the parameter is not zero, the function assumes that only the first
-nonzeroRows rows of the input array (#DFT_INVERSE is not set) or only the first nonzeroRows of the
-output array (#DFT_INVERSE is set) contain non-zeros, thus, the function can handle the rest of the
-rows more efficiently and save some time; this technique is very useful for calculating array
-cross-correlation or convolution using DFT.
-@sa dct , getOptimalDFTSize , mulSpectrums, filter2D , matchTemplate , flip , cartToPolar ,
-magnitude , phase
-*/
-CV_EXPORTS_W void dft(InputArray src, OutputArray dst, int flags = 0, int nonzeroRows = 0);
-
-/** @brief Calculates the inverse Discrete Fourier Transform of a 1D or 2D array.
-
-idft(src, dst, flags) is equivalent to dft(src, dst, flags | #DFT_INVERSE) .
-@note None of dft and idft scales the result by default. So, you should pass #DFT_SCALE to one of
-dft or idft explicitly to make these transforms mutually inverse.
-@sa dft, dct, idct, mulSpectrums, getOptimalDFTSize
-@param src input floating-point real or complex array.
-@param dst output array whose size and type depend on the flags.
-@param flags operation flags (see dft and #DftFlags).
-@param nonzeroRows number of dst rows to process; the rest of the rows have undefined content (see
-the convolution sample in dft description.
-*/
-CV_EXPORTS_W void idft(InputArray src, OutputArray dst, int flags = 0, int nonzeroRows = 0);
-
-/** @brief Performs a forward or inverse discrete Cosine transform of 1D or 2D array.
-
-The function cv::dct performs a forward or inverse discrete Cosine transform (DCT) of a 1D or 2D
-floating-point array:
--   Forward Cosine transform of a 1D vector of N elements:
-    \f[Y = C^{(N)}  \cdot X\f]
-    where
-    \f[C^{(N)}_{jk}= \sqrt{\alpha_j/N} \cos \left ( \frac{\pi(2k+1)j}{2N} \right )\f]
-    and
-    \f$\alpha_0=1\f$, \f$\alpha_j=2\f$ for *j \> 0*.
--   Inverse Cosine transform of a 1D vector of N elements:
-    \f[X =  \left (C^{(N)} \right )^{-1}  \cdot Y =  \left (C^{(N)} \right )^T  \cdot Y\f]
-    (since \f$C^{(N)}\f$ is an orthogonal matrix, \f$C^{(N)} \cdot \left(C^{(N)}\right)^T = I\f$ )
--   Forward 2D Cosine transform of M x N matrix:
-    \f[Y = C^{(N)}  \cdot X  \cdot \left (C^{(N)} \right )^T\f]
--   Inverse 2D Cosine transform of M x N matrix:
-    \f[X =  \left (C^{(N)} \right )^T  \cdot X  \cdot C^{(N)}\f]
-
-The function chooses the mode of operation by looking at the flags and size of the input array:
--   If (flags & #DCT_INVERSE) == 0 , the function does a forward 1D or 2D transform. Otherwise, it
-    is an inverse 1D or 2D transform.
--   If (flags & #DCT_ROWS) != 0 , the function performs a 1D transform of each row.
--   If the array is a single column or a single row, the function performs a 1D transform.
--   If none of the above is true, the function performs a 2D transform.
-
-@note Currently dct supports even-size arrays (2, 4, 6 ...). For data analysis and approximation, you
-can pad the array when necessary.
-Also, the function performance depends very much, and not monotonically, on the array size (see
-getOptimalDFTSize ). In the current implementation DCT of a vector of size N is calculated via DFT
-of a vector of size N/2 . Thus, the optimal DCT size N1 \>= N can be calculated as:
-@code
-    size_t getOptimalDCTSize(size_t N) { return 2*getOptimalDFTSize((N+1)/2); }
-    N1 = getOptimalDCTSize(N);
-@endcode
-@param src input floating-point array.
-@param dst output array of the same size and type as src .
-@param flags transformation flags as a combination of cv::DftFlags (DCT_*)
-@sa dft , getOptimalDFTSize , idct
-*/
-CV_EXPORTS_W void dct(InputArray src, OutputArray dst, int flags = 0);
-
-/** @brief Calculates the inverse Discrete Cosine Transform of a 1D or 2D array.
-
-idct(src, dst, flags) is equivalent to dct(src, dst, flags | DCT_INVERSE).
-@param src input floating-point single-channel array.
-@param dst output array of the same size and type as src.
-@param flags operation flags.
-@sa  dct, dft, idft, getOptimalDFTSize
-*/
-CV_EXPORTS_W void idct(InputArray src, OutputArray dst, int flags = 0);
-
-/** @brief Performs the per-element multiplication of two Fourier spectrums.
-
-The function cv::mulSpectrums performs the per-element multiplication of the two CCS-packed or complex
-matrices that are results of a real or complex Fourier transform.
-
-The function, together with dft and idft , may be used to calculate convolution (pass conjB=false )
-or correlation (pass conjB=true ) of two arrays rapidly. When the arrays are complex, they are
-simply multiplied (per element) with an optional conjugation of the second-array elements. When the
-arrays are real, they are assumed to be CCS-packed (see dft for details).
-@param a first input array.
-@param b second input array of the same size and type as src1 .
-@param c output array of the same size and type as src1 .
-@param flags operation flags; currently, the only supported flag is cv::DFT_ROWS, which indicates that
-each row of src1 and src2 is an independent 1D Fourier spectrum. If you do not want to use this flag, then simply add a `0` as value.
-@param conjB optional flag that conjugates the second input array before the multiplication (true)
-or not (false).
-*/
-CV_EXPORTS_W void mulSpectrums(InputArray a, InputArray b, OutputArray c,
-                               int flags, bool conjB = false);
-
-/** @brief Returns the optimal DFT size for a given vector size.
-
-DFT performance is not a monotonic function of a vector size. Therefore, when you calculate
-convolution of two arrays or perform the spectral analysis of an array, it usually makes sense to
-pad the input data with zeros to get a bit larger array that can be transformed much faster than the
-original one. Arrays whose size is a power-of-two (2, 4, 8, 16, 32, ...) are the fastest to process.
-Though, the arrays whose size is a product of 2's, 3's, and 5's (for example, 300 = 5\*5\*3\*2\*2)
-are also processed quite efficiently.
-
-The function cv::getOptimalDFTSize returns the minimum number N that is greater than or equal to vecsize
-so that the DFT of a vector of size N can be processed efficiently. In the current implementation N
-= 2 ^p^ \* 3 ^q^ \* 5 ^r^ for some integer p, q, r.
-
-The function returns a negative number if vecsize is too large (very close to INT_MAX ).
-
-While the function cannot be used directly to estimate the optimal vector size for DCT transform
-(since the current DCT implementation supports only even-size vectors), it can be easily processed
-as getOptimalDFTSize((vecsize+1)/2)\*2.
-@param vecsize vector size.
-@sa dft , dct , idft , idct , mulSpectrums
-*/
-CV_EXPORTS_W int getOptimalDFTSize(int vecsize);
-
-/** @brief Returns the default random number generator.
-
-The function cv::theRNG returns the default random number generator. For each thread, there is a
-separate random number generator, so you can use the function safely in multi-thread environments.
-If you just need to get a single random number using this generator or initialize an array, you can
-use randu or randn instead. But if you are going to generate many random numbers inside a loop, it
-is much faster to use this function to retrieve the generator and then use RNG::operator _Tp() .
-@sa RNG, randu, randn
-*/
-CV_EXPORTS RNG& theRNG();
-
-/** @brief Sets state of default random number generator.
-
-The function cv::setRNGSeed sets state of default random number generator to custom value.
-@param seed new state for default random number generator
-@sa RNG, randu, randn
-*/
-CV_EXPORTS_W void setRNGSeed(int seed);
-
-/** @brief Generates a single uniformly-distributed random number or an array of random numbers.
-
-Non-template variant of the function fills the matrix dst with uniformly-distributed
-random numbers from the specified range:
-\f[\texttt{low} _c  \leq \texttt{dst} (I)_c <  \texttt{high} _c\f]
-@param dst output array of random numbers; the array must be pre-allocated.
-@param low inclusive lower boundary of the generated random numbers.
-@param high exclusive upper boundary of the generated random numbers.
-@sa RNG, randn, theRNG
-*/
-CV_EXPORTS_W void randu(InputOutputArray dst, InputArray low, InputArray high);
-
-/** @brief Fills the array with normally distributed random numbers.
-
-The function cv::randn fills the matrix dst with normally distributed random numbers with the specified
-mean vector and the standard deviation matrix. The generated random numbers are clipped to fit the
-value range of the output array data type.
-@param dst output array of random numbers; the array must be pre-allocated and have 1 to 4 channels.
-@param mean mean value (expectation) of the generated random numbers.
-@param stddev standard deviation of the generated random numbers; it can be either a vector (in
-which case a diagonal standard deviation matrix is assumed) or a square matrix.
-@sa RNG, randu
-*/
-CV_EXPORTS_W void randn(InputOutputArray dst, InputArray mean, InputArray stddev);
-
-/** @brief Shuffles the array elements randomly.
-
-The function cv::randShuffle shuffles the specified 1D array by randomly choosing pairs of elements and
-swapping them. The number of such swap operations will be dst.rows\*dst.cols\*iterFactor .
-@param dst input/output numerical 1D array.
-@param iterFactor scale factor that determines the number of random swap operations (see the details
-below).
-@param rng optional random number generator used for shuffling; if it is zero, theRNG () is used
-instead.
-@sa RNG, sort
-*/
-CV_EXPORTS_W void randShuffle(InputOutputArray dst, double iterFactor = 1., RNG* rng = 0);
-
-/** @brief Principal Component Analysis
-
-The class is used to calculate a special basis for a set of vectors. The
-basis will consist of eigenvectors of the covariance matrix calculated
-from the input set of vectors. The class %PCA can also transform
-vectors to/from the new coordinate space defined by the basis. Usually,
-in this new coordinate system, each vector from the original set (and
-any linear combination of such vectors) can be quite accurately
-approximated by taking its first few components, corresponding to the
-eigenvectors of the largest eigenvalues of the covariance matrix.
-Geometrically it means that you calculate a projection of the vector to
-a subspace formed by a few eigenvectors corresponding to the dominant
-eigenvalues of the covariance matrix. And usually such a projection is
-very close to the original vector. So, you can represent the original
-vector from a high-dimensional space with a much shorter vector
-consisting of the projected vector's coordinates in the subspace. Such a
-transformation is also known as Karhunen-Loeve Transform, or KLT.
-See http://en.wikipedia.org/wiki/Principal_component_analysis
-
-The sample below is the function that takes two matrices. The first
-function stores a set of vectors (a row per vector) that is used to
-calculate PCA. The second function stores another "test" set of vectors
-(a row per vector). First, these vectors are compressed with PCA, then
-reconstructed back, and then the reconstruction error norm is computed
-and printed for each vector. :
-
-@code{.cpp}
-using namespace cv;
-
-PCA compressPCA(const Mat& pcaset, int maxComponents,
-                const Mat& testset, Mat& compressed)
-{
-    PCA pca(pcaset, // pass the data
-            Mat(), // we do not have a pre-computed mean vector,
-                   // so let the PCA engine to compute it
-            PCA::DATA_AS_ROW, // indicate that the vectors
-                                // are stored as matrix rows
-                                // (use PCA::DATA_AS_COL if the vectors are
-                                // the matrix columns)
-            maxComponents // specify, how many principal components to retain
-            );
-    // if there is no test data, just return the computed basis, ready-to-use
-    if( !testset.data )
-        return pca;
-    CV_Assert( testset.cols == pcaset.cols );
-
-    compressed.create(testset.rows, maxComponents, testset.type());
-
-    Mat reconstructed;
-    for( int i = 0; i < testset.rows; i++ )
-    {
-        Mat vec = testset.row(i), coeffs = compressed.row(i), reconstructed;
-        // compress the vector, the result will be stored
-        // in the i-th row of the output matrix
-        pca.project(vec, coeffs);
-        // and then reconstruct it
-        pca.backProject(coeffs, reconstructed);
-        // and measure the error
-        printf("%d. diff = %g\n", i, norm(vec, reconstructed, NORM_L2));
-    }
-    return pca;
-}
-@endcode
-@sa calcCovarMatrix, mulTransposed, SVD, dft, dct
-*/
-class CV_EXPORTS PCA
-{
-public:
-    enum Flags { DATA_AS_ROW = 0, //!< indicates that the input samples are stored as matrix rows
-                 DATA_AS_COL = 1, //!< indicates that the input samples are stored as matrix columns
-                 USE_AVG     = 2  //!
-               };
-
-    /** @brief default constructor
-
-    The default constructor initializes an empty %PCA structure. The other
-    constructors initialize the structure and call PCA::operator()().
-    */
-    PCA();
-
-    /** @overload
-    @param data input samples stored as matrix rows or matrix columns.
-    @param mean optional mean value; if the matrix is empty (@c noArray()),
-    the mean is computed from the data.
-    @param flags operation flags; currently the parameter is only used to
-    specify the data layout (PCA::Flags)
-    @param maxComponents maximum number of components that %PCA should
-    retain; by default, all the components are retained.
-    */
-    PCA(InputArray data, InputArray mean, int flags, int maxComponents = 0);
-
-    /** @overload
-    @param data input samples stored as matrix rows or matrix columns.
-    @param mean optional mean value; if the matrix is empty (noArray()),
-    the mean is computed from the data.
-    @param flags operation flags; currently the parameter is only used to
-    specify the data layout (PCA::Flags)
-    @param retainedVariance Percentage of variance that PCA should retain.
-    Using this parameter will let the PCA decided how many components to
-    retain but it will always keep at least 2.
-    */
-    PCA(InputArray data, InputArray mean, int flags, double retainedVariance);
-
-    /** @brief performs %PCA
-
-    The operator performs %PCA of the supplied dataset. It is safe to reuse
-    the same PCA structure for multiple datasets. That is, if the structure
-    has been previously used with another dataset, the existing internal
-    data is reclaimed and the new @ref eigenvalues, @ref eigenvectors and @ref
-    mean are allocated and computed.
-
-    The computed @ref eigenvalues are sorted from the largest to the smallest and
-    the corresponding @ref eigenvectors are stored as eigenvectors rows.
-
-    @param data input samples stored as the matrix rows or as the matrix
-    columns.
-    @param mean optional mean value; if the matrix is empty (noArray()),
-    the mean is computed from the data.
-    @param flags operation flags; currently the parameter is only used to
-    specify the data layout. (Flags)
-    @param maxComponents maximum number of components that PCA should
-    retain; by default, all the components are retained.
-    */
-    PCA& operator()(InputArray data, InputArray mean, int flags, int maxComponents = 0);
-
-    /** @overload
-    @param data input samples stored as the matrix rows or as the matrix
-    columns.
-    @param mean optional mean value; if the matrix is empty (noArray()),
-    the mean is computed from the data.
-    @param flags operation flags; currently the parameter is only used to
-    specify the data layout. (PCA::Flags)
-    @param retainedVariance Percentage of variance that %PCA should retain.
-    Using this parameter will let the %PCA decided how many components to
-    retain but it will always keep at least 2.
-     */
-    PCA& operator()(InputArray data, InputArray mean, int flags, double retainedVariance);
-
-    /** @brief Projects vector(s) to the principal component subspace.
-
-    The methods project one or more vectors to the principal component
-    subspace, where each vector projection is represented by coefficients in
-    the principal component basis. The first form of the method returns the
-    matrix that the second form writes to the result. So the first form can
-    be used as a part of expression while the second form can be more
-    efficient in a processing loop.
-    @param vec input vector(s); must have the same dimensionality and the
-    same layout as the input data used at %PCA phase, that is, if
-    DATA_AS_ROW are specified, then `vec.cols==data.cols`
-    (vector dimensionality) and `vec.rows` is the number of vectors to
-    project, and the same is true for the PCA::DATA_AS_COL case.
-    */
-    Mat project(InputArray vec) const;
-
-    /** @overload
-    @param vec input vector(s); must have the same dimensionality and the
-    same layout as the input data used at PCA phase, that is, if
-    DATA_AS_ROW are specified, then `vec.cols==data.cols`
-    (vector dimensionality) and `vec.rows` is the number of vectors to
-    project, and the same is true for the PCA::DATA_AS_COL case.
-    @param result output vectors; in case of PCA::DATA_AS_COL, the
-    output matrix has as many columns as the number of input vectors, this
-    means that `result.cols==vec.cols` and the number of rows match the
-    number of principal components (for example, `maxComponents` parameter
-    passed to the constructor).
-     */
-    void project(InputArray vec, OutputArray result) const;
-
-    /** @brief Reconstructs vectors from their PC projections.
-
-    The methods are inverse operations to PCA::project. They take PC
-    coordinates of projected vectors and reconstruct the original vectors.
-    Unless all the principal components have been retained, the
-    reconstructed vectors are different from the originals. But typically,
-    the difference is small if the number of components is large enough (but
-    still much smaller than the original vector dimensionality). As a
-    result, PCA is used.
-    @param vec coordinates of the vectors in the principal component
-    subspace, the layout and size are the same as of PCA::project output
-    vectors.
-     */
-    Mat backProject(InputArray vec) const;
-
-    /** @overload
-    @param vec coordinates of the vectors in the principal component
-    subspace, the layout and size are the same as of PCA::project output
-    vectors.
-    @param result reconstructed vectors; the layout and size are the same as
-    of PCA::project input vectors.
-     */
-    void backProject(InputArray vec, OutputArray result) const;
-
-    /** @brief write PCA objects
-
-    Writes @ref eigenvalues @ref eigenvectors and @ref mean to specified FileStorage
-     */
-    void write(FileStorage& fs) const;
-
-    /** @brief load PCA objects
-
-    Loads @ref eigenvalues @ref eigenvectors and @ref mean from specified FileNode
-     */
-    void read(const FileNode& fn);
-
-    Mat eigenvectors; //!< eigenvectors of the covariation matrix
-    Mat eigenvalues; //!< eigenvalues of the covariation matrix
-    Mat mean; //!< mean value subtracted before the projection and added after the back projection
-};
-
-/** @example samples/cpp/pca.cpp
-An example using %PCA for dimensionality reduction while maintaining an amount of variance
-*/
-
-/** @example samples/cpp/tutorial_code/ml/introduction_to_pca/introduction_to_pca.cpp
-Check @ref tutorial_introduction_to_pca "the corresponding tutorial" for more details
-*/
-
-/**
-@brief Linear Discriminant Analysis
-@todo document this class
-*/
-class CV_EXPORTS LDA
-{
-public:
-    /** @brief constructor
-    Initializes a LDA with num_components (default 0).
-    */
-    explicit LDA(int num_components = 0);
-
-    /** Initializes and performs a Discriminant Analysis with Fisher's
-     Optimization Criterion on given data in src and corresponding labels
-     in labels. If 0 (or less) number of components are given, they are
-     automatically determined for given data in computation.
-    */
-    LDA(InputArrayOfArrays src, InputArray labels, int num_components = 0);
-
-    /** Serializes this object to a given filename.
-      */
-    void save(const String& filename) const;
-
-    /** Deserializes this object from a given filename.
-      */
-    void load(const String& filename);
-
-    /** Serializes this object to a given cv::FileStorage.
-      */
-    void save(FileStorage& fs) const;
-
-    /** Deserializes this object from a given cv::FileStorage.
-      */
-    void load(const FileStorage& node);
-
-    /** destructor
-      */
-    ~LDA();
-
-    /** Compute the discriminants for data in src (row aligned) and labels.
-      */
-    void compute(InputArrayOfArrays src, InputArray labels);
-
-    /** Projects samples into the LDA subspace.
-        src may be one or more row aligned samples.
-      */
-    Mat project(InputArray src);
-
-    /** Reconstructs projections from the LDA subspace.
-        src may be one or more row aligned projections.
-      */
-    Mat reconstruct(InputArray src);
-
-    /** Returns the eigenvectors of this LDA.
-      */
-    Mat eigenvectors() const { return _eigenvectors; }
-
-    /** Returns the eigenvalues of this LDA.
-      */
-    Mat eigenvalues() const { return _eigenvalues; }
-
-    static Mat subspaceProject(InputArray W, InputArray mean, InputArray src);
-    static Mat subspaceReconstruct(InputArray W, InputArray mean, InputArray src);
-
-protected:
-    int _num_components;
-    Mat _eigenvectors;
-    Mat _eigenvalues;
-    void lda(InputArrayOfArrays src, InputArray labels);
-};
-
-/** @brief Singular Value Decomposition
-
-Class for computing Singular Value Decomposition of a floating-point
-matrix. The Singular Value Decomposition is used to solve least-square
-problems, under-determined linear systems, invert matrices, compute
-condition numbers, and so on.
-
-If you want to compute a condition number of a matrix or an absolute value of
-its determinant, you do not need `u` and `vt`. You can pass
-flags=SVD::NO_UV|... . Another flag SVD::FULL_UV indicates that full-size u
-and vt must be computed, which is not necessary most of the time.
-
-@sa invert, solve, eigen, determinant
-*/
-class CV_EXPORTS SVD
-{
-public:
-    enum Flags {
-        /** allow the algorithm to modify the decomposed matrix; it can save space and speed up
-            processing. currently ignored. */
-        MODIFY_A = 1,
-        /** indicates that only a vector of singular values `w` is to be processed, while u and vt
-            will be set to empty matrices */
-        NO_UV    = 2,
-        /** when the matrix is not square, by default the algorithm produces u and vt matrices of
-            sufficiently large size for the further A reconstruction; if, however, FULL_UV flag is
-            specified, u and vt will be full-size square orthogonal matrices.*/
-        FULL_UV  = 4
-    };
-
-    /** @brief the default constructor
-
-    initializes an empty SVD structure
-      */
-    SVD();
-
-    /** @overload
-    initializes an empty SVD structure and then calls SVD::operator()
-    @param src decomposed matrix. The depth has to be CV_32F or CV_64F.
-    @param flags operation flags (SVD::Flags)
-      */
-    SVD( InputArray src, int flags = 0 );
-
-    /** @brief the operator that performs SVD. The previously allocated u, w and vt are released.
-
-    The operator performs the singular value decomposition of the supplied
-    matrix. The u,`vt` , and the vector of singular values w are stored in
-    the structure. The same SVD structure can be reused many times with
-    different matrices. Each time, if needed, the previous u,`vt` , and w
-    are reclaimed and the new matrices are created, which is all handled by
-    Mat::create.
-    @param src decomposed matrix. The depth has to be CV_32F or CV_64F.
-    @param flags operation flags (SVD::Flags)
-      */
-    SVD& operator ()( InputArray src, int flags = 0 );
-
-    /** @brief decomposes matrix and stores the results to user-provided matrices
-
-    The methods/functions perform SVD of matrix. Unlike SVD::SVD constructor
-    and SVD::operator(), they store the results to the user-provided
-    matrices:
-
-    @code{.cpp}
-    Mat A, w, u, vt;
-    SVD::compute(A, w, u, vt);
-    @endcode
-
-    @param src decomposed matrix. The depth has to be CV_32F or CV_64F.
-    @param w calculated singular values
-    @param u calculated left singular vectors
-    @param vt transposed matrix of right singular vectors
-    @param flags operation flags - see SVD::Flags.
-      */
-    static void compute( InputArray src, OutputArray w,
-                         OutputArray u, OutputArray vt, int flags = 0 );
-
-    /** @overload
-    computes singular values of a matrix
-    @param src decomposed matrix. The depth has to be CV_32F or CV_64F.
-    @param w calculated singular values
-    @param flags operation flags - see SVD::Flags.
-      */
-    static void compute( InputArray src, OutputArray w, int flags = 0 );
-
-    /** @brief performs back substitution
-      */
-    static void backSubst( InputArray w, InputArray u,
-                           InputArray vt, InputArray rhs,
-                           OutputArray dst );
-
-    /** @brief solves an under-determined singular linear system
-
-    The method finds a unit-length solution x of a singular linear system
-    A\*x = 0. Depending on the rank of A, there can be no solutions, a
-    single solution or an infinite number of solutions. In general, the
-    algorithm solves the following problem:
-    \f[dst =  \arg \min _{x:  \| x \| =1}  \| src  \cdot x  \|\f]
-    @param src left-hand-side matrix.
-    @param dst found solution.
-      */
-    static void solveZ( InputArray src, OutputArray dst );
-
-    /** @brief performs a singular value back substitution.
-
-    The method calculates a back substitution for the specified right-hand
-    side:
-
-    \f[\texttt{x} =  \texttt{vt} ^T  \cdot diag( \texttt{w} )^{-1}  \cdot \texttt{u} ^T  \cdot \texttt{rhs} \sim \texttt{A} ^{-1}  \cdot \texttt{rhs}\f]
-
-    Using this technique you can either get a very accurate solution of the
-    convenient linear system, or the best (in the least-squares terms)
-    pseudo-solution of an overdetermined linear system.
-
-    @param rhs right-hand side of a linear system (u\*w\*v')\*dst = rhs to
-    be solved, where A has been previously decomposed.
-
-    @param dst found solution of the system.
-
-    @note Explicit SVD with the further back substitution only makes sense
-    if you need to solve many linear systems with the same left-hand side
-    (for example, src ). If all you need is to solve a single system
-    (possibly with multiple rhs immediately available), simply call solve
-    add pass #DECOMP_SVD there. It does absolutely the same thing.
-      */
-    void backSubst( InputArray rhs, OutputArray dst ) const;
-
-    /** @todo document */
-    template<typename _Tp, int m, int n, int nm> static
-    void compute( const Matx<_Tp, m, n>& a, Matx<_Tp, nm, 1>& w, Matx<_Tp, m, nm>& u, Matx<_Tp, n, nm>& vt );
-
-    /** @todo document */
-    template<typename _Tp, int m, int n, int nm> static
-    void compute( const Matx<_Tp, m, n>& a, Matx<_Tp, nm, 1>& w );
-
-    /** @todo document */
-    template<typename _Tp, int m, int n, int nm, int nb> static
-    void backSubst( const Matx<_Tp, nm, 1>& w, const Matx<_Tp, m, nm>& u, const Matx<_Tp, n, nm>& vt, const Matx<_Tp, m, nb>& rhs, Matx<_Tp, n, nb>& dst );
-
-    Mat u, w, vt;
-};
-
-/** @brief Random Number Generator
-
-Random number generator. It encapsulates the state (currently, a 64-bit
-integer) and has methods to return scalar random values and to fill
-arrays with random values. Currently it supports uniform and Gaussian
-(normal) distributions. The generator uses Multiply-With-Carry
-algorithm, introduced by G. Marsaglia (
-<http://en.wikipedia.org/wiki/Multiply-with-carry> ).
-Gaussian-distribution random numbers are generated using the Ziggurat
-algorithm ( <http://en.wikipedia.org/wiki/Ziggurat_algorithm> ),
-introduced by G. Marsaglia and W. W. Tsang.
-*/
-class CV_EXPORTS RNG
-{
-public:
-    enum { UNIFORM = 0,
-           NORMAL  = 1
-         };
-
-    /** @brief constructor
-
-    These are the RNG constructors. The first form sets the state to some
-    pre-defined value, equal to 2\*\*32-1 in the current implementation. The
-    second form sets the state to the specified value. If you passed state=0
-    , the constructor uses the above default value instead to avoid the
-    singular random number sequence, consisting of all zeros.
-    */
-    RNG();
-    /** @overload
-    @param state 64-bit value used to initialize the RNG.
-    */
-    RNG(uint64 state);
-    /**The method updates the state using the MWC algorithm and returns the
-    next 32-bit random number.*/
-    unsigned next();
-
-    /**Each of the methods updates the state using the MWC algorithm and
-    returns the next random number of the specified type. In case of integer
-    types, the returned number is from the available value range for the
-    specified type. In case of floating-point types, the returned value is
-    from [0,1) range.
-    */
-    operator uchar();
-    /** @overload */
-    operator schar();
-    /** @overload */
-    operator ushort();
-    /** @overload */
-    operator short();
-    /** @overload */
-    operator unsigned();
-    /** @overload */
-    operator int();
-    /** @overload */
-    operator float();
-    /** @overload */
-    operator double();
-
-    /** @brief returns a random integer sampled uniformly from [0, N).
-
-    The methods transform the state using the MWC algorithm and return the
-    next random number. The first form is equivalent to RNG::next . The
-    second form returns the random number modulo N , which means that the
-    result is in the range [0, N) .
-    */
-    unsigned operator ()();
-    /** @overload
-    @param N upper non-inclusive boundary of the returned random number.
-    */
-    unsigned operator ()(unsigned N);
-
-    /** @brief returns uniformly distributed integer random number from [a,b) range
-
-    The methods transform the state using the MWC algorithm and return the
-    next uniformly-distributed random number of the specified type, deduced
-    from the input parameter type, from the range [a, b) . There is a nuance
-    illustrated by the following sample:
-
-    @code{.cpp}
-    RNG rng;
-
-    // always produces 0
-    double a = rng.uniform(0, 1);
-
-    // produces double from [0, 1)
-    double a1 = rng.uniform((double)0, (double)1);
-
-    // produces float from [0, 1)
-    float b = rng.uniform(0.f, 1.f);
-
-    // produces double from [0, 1)
-    double c = rng.uniform(0., 1.);
-
-    // may cause compiler error because of ambiguity:
-    //  RNG::uniform(0, (int)0.999999)? or RNG::uniform((double)0, 0.99999)?
-    double d = rng.uniform(0, 0.999999);
-    @endcode
-
-    The compiler does not take into account the type of the variable to
-    which you assign the result of RNG::uniform . The only thing that
-    matters to the compiler is the type of a and b parameters. So, if you
-    want a floating-point random number, but the range boundaries are
-    integer numbers, either put dots in the end, if they are constants, or
-    use explicit type cast operators, as in the a1 initialization above.
-    @param a lower inclusive boundary of the returned random number.
-    @param b upper non-inclusive boundary of the returned random number.
-    */
-    int uniform(int a, int b);
-    /** @overload */
-    float uniform(float a, float b);
-    /** @overload */
-    double uniform(double a, double b);
-
-    /** @brief Fills arrays with random numbers.
-
-    @param mat 2D or N-dimensional matrix; currently matrices with more than
-    4 channels are not supported by the methods, use Mat::reshape as a
-    possible workaround.
-    @param distType distribution type, RNG::UNIFORM or RNG::NORMAL.
-    @param a first distribution parameter; in case of the uniform
-    distribution, this is an inclusive lower boundary, in case of the normal
-    distribution, this is a mean value.
-    @param b second distribution parameter; in case of the uniform
-    distribution, this is a non-inclusive upper boundary, in case of the
-    normal distribution, this is a standard deviation (diagonal of the
-    standard deviation matrix or the full standard deviation matrix).
-    @param saturateRange pre-saturation flag; for uniform distribution only;
-    if true, the method will first convert a and b to the acceptable value
-    range (according to the mat datatype) and then will generate uniformly
-    distributed random numbers within the range [saturate(a), saturate(b)),
-    if saturateRange=false, the method will generate uniformly distributed
-    random numbers in the original range [a, b) and then will saturate them,
-    it means, for example, that
-    <tt>theRNG().fill(mat_8u, RNG::UNIFORM, -DBL_MAX, DBL_MAX)</tt> will likely
-    produce array mostly filled with 0's and 255's, since the range (0, 255)
-    is significantly smaller than [-DBL_MAX, DBL_MAX).
-
-    Each of the methods fills the matrix with the random values from the
-    specified distribution. As the new numbers are generated, the RNG state
-    is updated accordingly. In case of multiple-channel images, every
-    channel is filled independently, which means that RNG cannot generate
-    samples from the multi-dimensional Gaussian distribution with
-    non-diagonal covariance matrix directly. To do that, the method
-    generates samples from multi-dimensional standard Gaussian distribution
-    with zero mean and identity covariation matrix, and then transforms them
-    using transform to get samples from the specified Gaussian distribution.
-    */
-    void fill( InputOutputArray mat, int distType, InputArray a, InputArray b, bool saturateRange = false );
-
-    /** @brief Returns the next random number sampled from the Gaussian distribution
-    @param sigma standard deviation of the distribution.
-
-    The method transforms the state using the MWC algorithm and returns the
-    next random number from the Gaussian distribution N(0,sigma) . That is,
-    the mean value of the returned random numbers is zero and the standard
-    deviation is the specified sigma .
-    */
-    double gaussian(double sigma);
-
-    uint64 state;
-
-    bool operator ==(const RNG& other) const;
-};
-
-/** @brief Mersenne Twister random number generator
-
-Inspired by http://www.math.sci.hiroshima-u.ac.jp/~m-mat/MT/MT2002/CODES/mt19937ar.c
-@todo document
-*/
-class CV_EXPORTS RNG_MT19937
-{
-public:
-    RNG_MT19937();
-    RNG_MT19937(unsigned s);
-    void seed(unsigned s);
-
-    unsigned next();
-
-    operator int();
-    operator unsigned();
-    operator float();
-    operator double();
-
-    unsigned operator ()(unsigned N);
-    unsigned operator ()();
-
-    /** @brief returns uniformly distributed integer random number from [a,b) range*/
-    int uniform(int a, int b);
-    /** @brief returns uniformly distributed floating-point random number from [a,b) range*/
-    float uniform(float a, float b);
-    /** @brief returns uniformly distributed double-precision floating-point random number from [a,b) range*/
-    double uniform(double a, double b);
-
-private:
-    enum PeriodParameters {N = 624, M = 397};
-    unsigned state[N];
-    int mti;
-};
-
-//! @} core_array
-
-//! @addtogroup core_cluster
-//!  @{
-
-/** @example samples/cpp/kmeans.cpp
-An example on K-means clustering
-*/
-
-/** @brief Finds centers of clusters and groups input samples around the clusters.
-
-The function kmeans implements a k-means algorithm that finds the centers of cluster_count clusters
-and groups the input samples around the clusters. As an output, \f$\texttt{bestLabels}_i\f$ contains a
-0-based cluster index for the sample stored in the \f$i^{th}\f$ row of the samples matrix.
-
-@note
--   (Python) An example on K-means clustering can be found at
-    opencv_source_code/samples/python/kmeans.py
-@param data Data for clustering. An array of N-Dimensional points with float coordinates is needed.
-Examples of this array can be:
--   Mat points(count, 2, CV_32F);
--   Mat points(count, 1, CV_32FC2);
--   Mat points(1, count, CV_32FC2);
--   std::vector\<cv::Point2f\> points(sampleCount);
-@param K Number of clusters to split the set by.
-@param bestLabels Input/output integer array that stores the cluster indices for every sample.
-@param criteria The algorithm termination criteria, that is, the maximum number of iterations and/or
-the desired accuracy. The accuracy is specified as criteria.epsilon. As soon as each of the cluster
-centers moves by less than criteria.epsilon on some iteration, the algorithm stops.
-@param attempts Flag to specify the number of times the algorithm is executed using different
-initial labellings. The algorithm returns the labels that yield the best compactness (see the last
-function parameter).
-@param flags Flag that can take values of cv::KmeansFlags
-@param centers Output matrix of the cluster centers, one row per each cluster center.
-@return The function returns the compactness measure that is computed as
-\f[\sum _i  \| \texttt{samples} _i -  \texttt{centers} _{ \texttt{labels} _i} \| ^2\f]
-after every attempt. The best (minimum) value is chosen and the corresponding labels and the
-compactness value are returned by the function. Basically, you can use only the core of the
-function, set the number of attempts to 1, initialize labels each time using a custom algorithm,
-pass them with the ( flags = #KMEANS_USE_INITIAL_LABELS ) flag, and then choose the best
-(most-compact) clustering.
-*/
-CV_EXPORTS_W double kmeans( InputArray data, int K, InputOutputArray bestLabels,
-                            TermCriteria criteria, int attempts,
-                            int flags, OutputArray centers = noArray() );
-
-//! @} core_cluster
-
-//! @addtogroup core_basic
-//! @{
-
-/////////////////////////////// Formatted output of cv::Mat ///////////////////////////
-
-/** @todo document */
-class CV_EXPORTS Formatted
-{
-public:
-    virtual const char* next() = 0;
-    virtual void reset() = 0;
-    virtual ~Formatted();
-};
-
-/** @todo document */
-class CV_EXPORTS Formatter
-{
-public:
-    enum FormatType {
-           FMT_DEFAULT = 0,
-           FMT_MATLAB  = 1,
-           FMT_CSV     = 2,
-           FMT_PYTHON  = 3,
-           FMT_NUMPY   = 4,
-           FMT_C       = 5
-         };
-
-    virtual ~Formatter();
-
-    virtual Ptr<Formatted> format(const Mat& mtx) const = 0;
-
-    virtual void set16fPrecision(int p = 4) = 0;
-    virtual void set32fPrecision(int p = 8) = 0;
-    virtual void set64fPrecision(int p = 16) = 0;
-    virtual void setMultiline(bool ml = true) = 0;
-
-    static Ptr<Formatter> get(Formatter::FormatType fmt = FMT_DEFAULT);
-
-};
-
-static inline
-String& operator << (String& out, Ptr<Formatted> fmtd)
-{
-    fmtd->reset();
-    for(const char* str = fmtd->next(); str; str = fmtd->next())
-        out += cv::String(str);
-    return out;
-}
-
-static inline
-String& operator << (String& out, const Mat& mtx)
-{
-    return out << Formatter::get()->format(mtx);
-}
-
-//////////////////////////////////////// Algorithm ////////////////////////////////////
-
-class CV_EXPORTS Algorithm;
-
-template<typename _Tp, typename _EnumTp = void> struct ParamType {};
-
-
-/** @brief This is a base class for all more or less complex algorithms in OpenCV
-
-especially for classes of algorithms, for which there can be multiple implementations. The examples
-are stereo correspondence (for which there are algorithms like block matching, semi-global block
-matching, graph-cut etc.), background subtraction (which can be done using mixture-of-gaussians
-models, codebook-based algorithm etc.), optical flow (block matching, Lucas-Kanade, Horn-Schunck
-etc.).
-
-Here is example of SimpleBlobDetector use in your application via Algorithm interface:
-@snippet snippets/core_various.cpp Algorithm
-*/
-class CV_EXPORTS_W Algorithm
-{
-public:
-    Algorithm();
-    virtual ~Algorithm();
-
-    /** @brief Clears the algorithm state
-    */
-    CV_WRAP virtual void clear() {}
-
-    /** @brief Stores algorithm parameters in a file storage
-    */
-    virtual void write(FileStorage& fs) const { CV_UNUSED(fs); }
-
-    /** @brief simplified API for language bindings
-    * @overload
-    */
-    CV_WRAP void write(const Ptr<FileStorage>& fs, const String& name = String()) const;
-
-    /** @brief Reads algorithm parameters from a file storage
-    */
-    CV_WRAP virtual void read(const FileNode& fn) { CV_UNUSED(fn); }
-
-    /** @brief Returns true if the Algorithm is empty (e.g. in the very beginning or after unsuccessful read
-    */
-    CV_WRAP virtual bool empty() const { return false; }
-
-    /** @brief Reads algorithm from the file node
-
-    This is static template method of Algorithm. It's usage is following (in the case of SVM):
-    @code
-    cv::FileStorage fsRead("example.xml", FileStorage::READ);
-    Ptr<SVM> svm = Algorithm::read<SVM>(fsRead.root());
-    @endcode
-    In order to make this method work, the derived class must overwrite Algorithm::read(const
-    FileNode& fn) and also have static create() method without parameters
-    (or with all the optional parameters)
-    */
-    template<typename _Tp> static Ptr<_Tp> read(const FileNode& fn)
-    {
-        Ptr<_Tp> obj = _Tp::create();
-        obj->read(fn);
-        return !obj->empty() ? obj : Ptr<_Tp>();
-    }
-
-    /** @brief Loads algorithm from the file
-
-    @param filename Name of the file to read.
-    @param objname The optional name of the node to read (if empty, the first top-level node will be used)
-
-    This is static template method of Algorithm. It's usage is following (in the case of SVM):
-    @code
-    Ptr<SVM> svm = Algorithm::load<SVM>("my_svm_model.xml");
-    @endcode
-    In order to make this method work, the derived class must overwrite Algorithm::read(const
-    FileNode& fn).
-    */
-    template<typename _Tp> static Ptr<_Tp> load(const String& filename, const String& objname=String())
-    {
-        FileStorage fs(filename, FileStorage::READ);
-        CV_Assert(fs.isOpened());
-        FileNode fn = objname.empty() ? fs.getFirstTopLevelNode() : fs[objname];
-        if (fn.empty()) return Ptr<_Tp>();
-        Ptr<_Tp> obj = _Tp::create();
-        obj->read(fn);
-        return !obj->empty() ? obj : Ptr<_Tp>();
-    }
-
-    /** @brief Loads algorithm from a String
-
-    @param strModel The string variable containing the model you want to load.
-    @param objname The optional name of the node to read (if empty, the first top-level node will be used)
-
-    This is static template method of Algorithm. It's usage is following (in the case of SVM):
-    @code
-    Ptr<SVM> svm = Algorithm::loadFromString<SVM>(myStringModel);
-    @endcode
-    */
-    template<typename _Tp> static Ptr<_Tp> loadFromString(const String& strModel, const String& objname=String())
-    {
-        FileStorage fs(strModel, FileStorage::READ + FileStorage::MEMORY);
-        FileNode fn = objname.empty() ? fs.getFirstTopLevelNode() : fs[objname];
-        Ptr<_Tp> obj = _Tp::create();
-        obj->read(fn);
-        return !obj->empty() ? obj : Ptr<_Tp>();
-    }
-
-    /** Saves the algorithm to a file.
-    In order to make this method work, the derived class must implement Algorithm::write(FileStorage& fs). */
-    CV_WRAP virtual void save(const String& filename) const;
-
-    /** Returns the algorithm string identifier.
-    This string is used as top level xml/yml node tag when the object is saved to a file or string. */
-    CV_WRAP virtual String getDefaultName() const;
-
-protected:
-    void writeFormat(FileStorage& fs) const;
-};
-
-enum struct Param {
-    INT=0, BOOLEAN=1, REAL=2, STRING=3, MAT=4, MAT_VECTOR=5, ALGORITHM=6, FLOAT=7,
-    UNSIGNED_INT=8, UINT64=9, UCHAR=11, SCALAR=12
-};
-
-
-
-template<> struct ParamType<bool>
-{
-    typedef bool const_param_type;
-    typedef bool member_type;
-
-    static const Param type = Param::BOOLEAN;
-};
-
-template<> struct ParamType<int>
-{
-    typedef int const_param_type;
-    typedef int member_type;
-
-    static const Param type = Param::INT;
-};
-
-template<> struct ParamType<double>
-{
-    typedef double const_param_type;
-    typedef double member_type;
-
-    static const Param type = Param::REAL;
-};
-
-template<> struct ParamType<String>
-{
-    typedef const String& const_param_type;
-    typedef String member_type;
-
-    static const Param type = Param::STRING;
-};
-
-template<> struct ParamType<Mat>
-{
-    typedef const Mat& const_param_type;
-    typedef Mat member_type;
-
-    static const Param type = Param::MAT;
-};
-
-template<> struct ParamType<std::vector<Mat> >
-{
-    typedef const std::vector<Mat>& const_param_type;
-    typedef std::vector<Mat> member_type;
-
-    static const Param type = Param::MAT_VECTOR;
-};
-
-template<> struct ParamType<Algorithm>
-{
-    typedef const Ptr<Algorithm>& const_param_type;
-    typedef Ptr<Algorithm> member_type;
-
-    static const Param type = Param::ALGORITHM;
-};
-
-template<> struct ParamType<float>
-{
-    typedef float const_param_type;
-    typedef float member_type;
-
-    static const Param type = Param::FLOAT;
-};
-
-template<> struct ParamType<unsigned>
-{
-    typedef unsigned const_param_type;
-    typedef unsigned member_type;
-
-    static const Param type = Param::UNSIGNED_INT;
-};
-
-template<> struct ParamType<uint64>
-{
-    typedef uint64 const_param_type;
-    typedef uint64 member_type;
-
-    static const Param type = Param::UINT64;
-};
-
-template<> struct ParamType<uchar>
-{
-    typedef uchar const_param_type;
-    typedef uchar member_type;
-
-    static const Param type = Param::UCHAR;
-};
-
-template<> struct ParamType<Scalar>
-{
-    typedef const Scalar& const_param_type;
-    typedef Scalar member_type;
-
-    static const Param type = Param::SCALAR;
-};
-
-template<typename _Tp>
-struct ParamType<_Tp, typename std::enable_if< std::is_enum<_Tp>::value >::type>
-{
-    typedef typename std::underlying_type<_Tp>::type const_param_type;
-    typedef typename std::underlying_type<_Tp>::type member_type;
-
-    static const Param type = Param::INT;
-};
-
-//! @} core_basic
-
-} //namespace cv
-
-#include "opencv2/core/operations.hpp"
-#include "opencv2/core/cvstd.inl.hpp"
-#include "opencv2/core/utility.hpp"
-#include "opencv2/core/optim.hpp"
-#include "opencv2/core/ovx.hpp"
-
-#endif /*OPENCV_CORE_HPP*/
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/affine.hpp b/projects/llm_framework/include/opencv4/opencv2/core/affine.hpp
deleted file mode 100644
index 1806382e..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/affine.hpp
+++ /dev/null
@@ -1,678 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_AFFINE3_HPP
-#define OPENCV_CORE_AFFINE3_HPP
-
-#ifdef __cplusplus
-
-#include <opencv2/core.hpp>
-
-namespace cv
-{
-
-//! @addtogroup core
-//! @{
-
-    /** @brief Affine transform
-     *
-     * It represents a 4x4 homogeneous transformation matrix \f$T\f$
-     *
-     *  \f[T =
-     *  \begin{bmatrix}
-     *  R & t\\
-     *  0 & 1\\
-     *  \end{bmatrix}
-     *  \f]
-     *
-     *  where \f$R\f$ is a 3x3 rotation matrix and \f$t\f$ is a 3x1 translation vector.
-     *
-     *  You can specify \f$R\f$ either by a 3x3 rotation matrix or by a 3x1 rotation vector,
-     *  which is converted to a 3x3 rotation matrix by the Rodrigues formula.
-     *
-     *  To construct a matrix \f$T\f$ representing first rotation around the axis \f$r\f$ with rotation
-     *  angle \f$|r|\f$ in radian (right hand rule) and then translation by the vector \f$t\f$, you can use
-     *
-     *  @code
-     *  cv::Vec3f r, t;
-     *  cv::Affine3f T(r, t);
-     *  @endcode
-     *
-     *  If you already have the rotation matrix \f$R\f$, then you can use
-     *
-     *  @code
-     *  cv::Matx33f R;
-     *  cv::Affine3f T(R, t);
-     *  @endcode
-     *
-     *  To extract the rotation matrix \f$R\f$ from \f$T\f$, use
-     *
-     *  @code
-     *  cv::Matx33f R = T.rotation();
-     *  @endcode
-     *
-     *  To extract the translation vector \f$t\f$ from \f$T\f$, use
-     *
-     *  @code
-     *  cv::Vec3f t = T.translation();
-     *  @endcode
-     *
-     *  To extract the rotation vector \f$r\f$ from \f$T\f$, use
-     *
-     *  @code
-     *  cv::Vec3f r = T.rvec();
-     *  @endcode
-     *
-     *  Note that since the mapping from rotation vectors to rotation matrices
-     *  is many to one. The returned rotation vector is not necessarily the one
-     *  you used before to set the matrix.
-     *
-     *  If you have two transformations \f$T = T_1 * T_2\f$, use
-     *
-     *  @code
-     *  cv::Affine3f T, T1, T2;
-     *  T = T2.concatenate(T1);
-     *  @endcode
-     *
-     *  To get the inverse transform of \f$T\f$, use
-     *
-     *  @code
-     *  cv::Affine3f T, T_inv;
-     *  T_inv = T.inv();
-     *  @endcode
-     *
-     */
-    template<typename T>
-    class Affine3
-    {
-    public:
-        typedef T float_type;
-        typedef Matx<float_type, 3, 3> Mat3;
-        typedef Matx<float_type, 4, 4> Mat4;
-        typedef Vec<float_type, 3> Vec3;
-
-       //! Default constructor. It represents a 4x4 identity matrix.
-        Affine3();
-
-        //! Augmented affine matrix
-        Affine3(const Mat4& affine);
-
-        /**
-         *  The resulting 4x4 matrix is
-         *
-         *  \f[
-         *  \begin{bmatrix}
-         *  R & t\\
-         *  0 & 1\\
-         *  \end{bmatrix}
-         *  \f]
-         *
-         * @param R 3x3 rotation matrix.
-         * @param t 3x1 translation vector.
-         */
-        Affine3(const Mat3& R, const Vec3& t = Vec3::all(0));
-
-        /**
-         * Rodrigues vector.
-         *
-         * The last row of the current matrix is set to [0,0,0,1].
-         *
-         * @param rvec 3x1 rotation vector. Its direction indicates the rotation axis and its length
-         *             indicates the rotation angle in radian (using right hand rule).
-         * @param t 3x1 translation vector.
-         */
-        Affine3(const Vec3& rvec, const Vec3& t = Vec3::all(0));
-
-        /**
-         * Combines all constructors above. Supports 4x4, 3x4, 3x3, 1x3, 3x1 sizes of data matrix.
-         *
-         * The last row of the current matrix is set to [0,0,0,1] when data is not 4x4.
-         *
-         * @param data 1-channel matrix.
-         *             when it is 4x4, it is copied to the current matrix and t is not used.
-         *             When it is 3x4, it is copied to the upper part 3x4 of the current matrix and t is not used.
-         *             When it is 3x3, it is copied to the upper left 3x3 part of the current matrix.
-         *             When it is 3x1 or 1x3, it is treated as a rotation vector and the Rodrigues formula is used
-         *                             to compute a 3x3 rotation matrix.
-         * @param t 3x1 translation vector. It is used only when data is neither 4x4 nor 3x4.
-         */
-        explicit Affine3(const Mat& data, const Vec3& t = Vec3::all(0));
-
-        //! From 16-element array
-        explicit Affine3(const float_type* vals);
-
-        //! Create an 4x4 identity transform
-        static Affine3 Identity();
-
-        /**
-         * Rotation matrix.
-         *
-         * Copy the rotation matrix to the upper left 3x3 part of the current matrix.
-         * The remaining elements of the current matrix are not changed.
-         *
-         * @param R 3x3 rotation matrix.
-         *
-         */
-        void rotation(const Mat3& R);
-
-        /**
-         * Rodrigues vector.
-         *
-         * It sets the upper left 3x3 part of the matrix. The remaining part is unaffected.
-         *
-         * @param rvec 3x1 rotation vector. The direction indicates the rotation axis and
-         *             its length indicates the rotation angle in radian (using the right thumb convention).
-         */
-        void rotation(const Vec3& rvec);
-
-        /**
-         * Combines rotation methods above. Supports 3x3, 1x3, 3x1 sizes of data matrix.
-         *
-         * It sets the upper left 3x3 part of the matrix. The remaining part is unaffected.
-         *
-         * @param data 1-channel matrix.
-         *             When it is a 3x3 matrix, it sets the upper left 3x3 part of the current matrix.
-         *             When it is a 1x3 or 3x1 matrix, it is used as a rotation vector. The Rodrigues formula
-         *             is used to compute the rotation matrix and sets the upper left 3x3 part of the current matrix.
-         */
-        void rotation(const Mat& data);
-
-        /**
-         * Copy the 3x3 matrix L to the upper left part of the current matrix
-         *
-         * It sets the upper left 3x3 part of the matrix. The remaining part is unaffected.
-         *
-         * @param L 3x3 matrix.
-         */
-        void linear(const Mat3& L);
-
-        /**
-         * Copy t to the first three elements of the last column of the current matrix
-         *
-         * It sets the upper right 3x1 part of the matrix. The remaining part is unaffected.
-         *
-         * @param t 3x1 translation vector.
-         */
-        void translation(const Vec3& t);
-
-        //! @return the upper left 3x3 part
-        Mat3 rotation() const;
-
-        //! @return the upper left 3x3 part
-        Mat3 linear() const;
-
-        //! @return the upper right 3x1 part
-        Vec3 translation() const;
-
-        //! Rodrigues vector.
-        //! @return a vector representing the upper left 3x3 rotation matrix of the current matrix.
-        //! @warning  Since the mapping between rotation vectors and rotation matrices is many to one,
-        //!           this function returns only one rotation vector that represents the current rotation matrix,
-        //!           which is not necessarily the same one set by `rotation(const Vec3& rvec)`.
-        Vec3 rvec() const;
-
-        //! @return the inverse of the current matrix.
-        Affine3 inv(int method = cv::DECOMP_SVD) const;
-
-        //! a.rotate(R) is equivalent to Affine(R, 0) * a;
-        Affine3 rotate(const Mat3& R) const;
-
-        //! a.rotate(rvec) is equivalent to Affine(rvec, 0) * a;
-        Affine3 rotate(const Vec3& rvec) const;
-
-        //! a.translate(t) is equivalent to Affine(E, t) * a, where E is an identity matrix
-        Affine3 translate(const Vec3& t) const;
-
-        //! a.concatenate(affine) is equivalent to affine * a;
-        Affine3 concatenate(const Affine3& affine) const;
-
-        template <typename Y> operator Affine3<Y>() const;
-
-        template <typename Y> Affine3<Y> cast() const;
-
-        Mat4 matrix;
-
-#if defined EIGEN_WORLD_VERSION && defined EIGEN_GEOMETRY_MODULE_H
-        Affine3(const Eigen::Transform<T, 3, Eigen::Affine, (Eigen::RowMajor)>& affine);
-        Affine3(const Eigen::Transform<T, 3, Eigen::Affine>& affine);
-        operator Eigen::Transform<T, 3, Eigen::Affine, (Eigen::RowMajor)>() const;
-        operator Eigen::Transform<T, 3, Eigen::Affine>() const;
-#endif
-    };
-
-    template<typename T> static
-    Affine3<T> operator*(const Affine3<T>& affine1, const Affine3<T>& affine2);
-
-    //! V is a 3-element vector with member fields x, y and z
-    template<typename T, typename V> static
-    V operator*(const Affine3<T>& affine, const V& vector);
-
-    typedef Affine3<float> Affine3f;
-    typedef Affine3<double> Affine3d;
-
-    static Vec3f operator*(const Affine3f& affine, const Vec3f& vector);
-    static Vec3d operator*(const Affine3d& affine, const Vec3d& vector);
-
-    template<typename _Tp> class DataType< Affine3<_Tp> >
-    {
-    public:
-        typedef Affine3<_Tp>                               value_type;
-        typedef Affine3<typename DataType<_Tp>::work_type> work_type;
-        typedef _Tp                                        channel_type;
-
-        enum { generic_type = 0,
-               channels     = 16,
-               fmt          = traits::SafeFmt<channel_type>::fmt + ((channels - 1) << 8)
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-               ,depth        = DataType<channel_type>::depth
-               ,type         = CV_MAKETYPE(depth, channels)
-#endif
-             };
-
-        typedef Vec<channel_type, channels> vec_type;
-    };
-
-    namespace traits {
-    template<typename _Tp>
-    struct Depth< Affine3<_Tp> > { enum { value = Depth<_Tp>::value }; };
-    template<typename _Tp>
-    struct Type< Affine3<_Tp> > { enum { value = CV_MAKETYPE(Depth<_Tp>::value, 16) }; };
-    } // namespace
-
-//! @} core
-
-}
-
-//! @cond IGNORED
-
-///////////////////////////////////////////////////////////////////////////////////
-// Implementation
-
-template<typename T> inline
-cv::Affine3<T>::Affine3()
-    : matrix(Mat4::eye())
-{}
-
-template<typename T> inline
-cv::Affine3<T>::Affine3(const Mat4& affine)
-    : matrix(affine)
-{}
-
-template<typename T> inline
-cv::Affine3<T>::Affine3(const Mat3& R, const Vec3& t)
-{
-    rotation(R);
-    translation(t);
-    matrix.val[12] = matrix.val[13] = matrix.val[14] = 0;
-    matrix.val[15] = 1;
-}
-
-template<typename T> inline
-cv::Affine3<T>::Affine3(const Vec3& _rvec, const Vec3& t)
-{
-    rotation(_rvec);
-    translation(t);
-    matrix.val[12] = matrix.val[13] = matrix.val[14] = 0;
-    matrix.val[15] = 1;
-}
-
-template<typename T> inline
-cv::Affine3<T>::Affine3(const cv::Mat& data, const Vec3& t)
-{
-    CV_Assert(data.type() == cv::traits::Type<T>::value);
-    CV_Assert(data.channels() == 1);
-
-    if (data.cols == 4 && data.rows == 4)
-    {
-        data.copyTo(matrix);
-        return;
-    }
-    else if (data.cols == 4 && data.rows == 3)
-    {
-        rotation(data(Rect(0, 0, 3, 3)));
-        translation(data(Rect(3, 0, 1, 3)));
-    }
-    else
-    {
-        rotation(data);
-        translation(t);
-    }
-
-    matrix.val[12] = matrix.val[13] = matrix.val[14] = 0;
-    matrix.val[15] = 1;
-}
-
-template<typename T> inline
-cv::Affine3<T>::Affine3(const float_type* vals) : matrix(vals)
-{}
-
-template<typename T> inline
-cv::Affine3<T> cv::Affine3<T>::Identity()
-{
-    return Affine3<T>(cv::Affine3<T>::Mat4::eye());
-}
-
-template<typename T> inline
-void cv::Affine3<T>::rotation(const Mat3& R)
-{
-    linear(R);
-}
-
-template<typename T> inline
-void cv::Affine3<T>::rotation(const Vec3& _rvec)
-{
-    double theta = norm(_rvec);
-
-    if (theta < DBL_EPSILON)
-        rotation(Mat3::eye());
-    else
-    {
-        double c = std::cos(theta);
-        double s = std::sin(theta);
-        double c1 = 1. - c;
-        double itheta = (theta != 0) ? 1./theta : 0.;
-
-        Point3_<T> r = _rvec*itheta;
-
-        Mat3 rrt( r.x*r.x, r.x*r.y, r.x*r.z, r.x*r.y, r.y*r.y, r.y*r.z, r.x*r.z, r.y*r.z, r.z*r.z );
-        Mat3 r_x( 0, -r.z, r.y, r.z, 0, -r.x, -r.y, r.x, 0 );
-
-        // R = cos(theta)*I + (1 - cos(theta))*r*rT + sin(theta)*[r_x]
-        // where [r_x] is [0 -rz ry; rz 0 -rx; -ry rx 0]
-        Mat3 R = c*Mat3::eye() + c1*rrt + s*r_x;
-
-        rotation(R);
-    }
-}
-
-//Combines rotation methods above. Supports 3x3, 1x3, 3x1 sizes of data matrix;
-template<typename T> inline
-void cv::Affine3<T>::rotation(const cv::Mat& data)
-{
-    CV_Assert(data.type() == cv::traits::Type<T>::value);
-    CV_Assert(data.channels() == 1);
-
-    if (data.cols == 3 && data.rows == 3)
-    {
-        Mat3 R;
-        data.copyTo(R);
-        rotation(R);
-    }
-    else if ((data.cols == 3 && data.rows == 1) || (data.cols == 1 && data.rows == 3))
-    {
-        Vec3 _rvec;
-        data.reshape(1, 3).copyTo(_rvec);
-        rotation(_rvec);
-    }
-    else
-        CV_Error(Error::StsError, "Input matrix can only be 3x3, 1x3 or 3x1");
-}
-
-template<typename T> inline
-void cv::Affine3<T>::linear(const Mat3& L)
-{
-    matrix.val[0] = L.val[0]; matrix.val[1] = L.val[1];  matrix.val[ 2] = L.val[2];
-    matrix.val[4] = L.val[3]; matrix.val[5] = L.val[4];  matrix.val[ 6] = L.val[5];
-    matrix.val[8] = L.val[6]; matrix.val[9] = L.val[7];  matrix.val[10] = L.val[8];
-}
-
-template<typename T> inline
-void cv::Affine3<T>::translation(const Vec3& t)
-{
-    matrix.val[3] = t[0]; matrix.val[7] = t[1]; matrix.val[11] = t[2];
-}
-
-template<typename T> inline
-typename cv::Affine3<T>::Mat3 cv::Affine3<T>::rotation() const
-{
-    return linear();
-}
-
-template<typename T> inline
-typename cv::Affine3<T>::Mat3 cv::Affine3<T>::linear() const
-{
-    typename cv::Affine3<T>::Mat3 R;
-    R.val[0] = matrix.val[0];  R.val[1] = matrix.val[1];  R.val[2] = matrix.val[ 2];
-    R.val[3] = matrix.val[4];  R.val[4] = matrix.val[5];  R.val[5] = matrix.val[ 6];
-    R.val[6] = matrix.val[8];  R.val[7] = matrix.val[9];  R.val[8] = matrix.val[10];
-    return R;
-}
-
-template<typename T> inline
-typename cv::Affine3<T>::Vec3 cv::Affine3<T>::translation() const
-{
-    return Vec3(matrix.val[3], matrix.val[7], matrix.val[11]);
-}
-
-template<typename T> inline
-typename cv::Affine3<T>::Vec3 cv::Affine3<T>::rvec() const
-{
-    cv::Vec3d w;
-    cv::Matx33d u, vt, R = rotation();
-    cv::SVD::compute(R, w, u, vt, cv::SVD::FULL_UV + cv::SVD::MODIFY_A);
-    R = u * vt;
-
-    double rx = R.val[7] - R.val[5];
-    double ry = R.val[2] - R.val[6];
-    double rz = R.val[3] - R.val[1];
-
-    double s = std::sqrt((rx*rx + ry*ry + rz*rz)*0.25);
-    double c = (R.val[0] + R.val[4] + R.val[8] - 1) * 0.5;
-    c = c > 1.0 ? 1.0 : c < -1.0 ? -1.0 : c;
-    double theta = std::acos(c);
-
-    if( s < 1e-5 )
-    {
-        if( c > 0 )
-            rx = ry = rz = 0;
-        else
-        {
-            double t;
-            t = (R.val[0] + 1) * 0.5;
-            rx = std::sqrt(std::max(t, 0.0));
-            t = (R.val[4] + 1) * 0.5;
-            ry = std::sqrt(std::max(t, 0.0)) * (R.val[1] < 0 ? -1.0 : 1.0);
-            t = (R.val[8] + 1) * 0.5;
-            rz = std::sqrt(std::max(t, 0.0)) * (R.val[2] < 0 ? -1.0 : 1.0);
-
-            if( fabs(rx) < fabs(ry) && fabs(rx) < fabs(rz) && (R.val[5] > 0) != (ry*rz > 0) )
-                rz = -rz;
-            theta /= std::sqrt(rx*rx + ry*ry + rz*rz);
-            rx *= theta;
-            ry *= theta;
-            rz *= theta;
-        }
-    }
-    else
-    {
-        double vth = 1/(2*s);
-        vth *= theta;
-        rx *= vth; ry *= vth; rz *= vth;
-    }
-
-    return cv::Vec3d(rx, ry, rz);
-}
-
-template<typename T> inline
-cv::Affine3<T> cv::Affine3<T>::inv(int method) const
-{
-    return matrix.inv(method);
-}
-
-template<typename T> inline
-cv::Affine3<T> cv::Affine3<T>::rotate(const Mat3& R) const
-{
-    Mat3 Lc = linear();
-    Vec3 tc = translation();
-    Mat4 result;
-    result.val[12] = result.val[13] = result.val[14] = 0;
-    result.val[15] = 1;
-
-    for(int j = 0; j < 3; ++j)
-    {
-        for(int i = 0; i < 3; ++i)
-        {
-            float_type value = 0;
-            for(int k = 0; k < 3; ++k)
-                value += R(j, k) * Lc(k, i);
-            result(j, i) = value;
-        }
-
-        result(j, 3) = R.row(j).dot(tc.t());
-    }
-    return result;
-}
-
-template<typename T> inline
-cv::Affine3<T> cv::Affine3<T>::rotate(const Vec3& _rvec) const
-{
-    return rotate(Affine3f(_rvec).rotation());
-}
-
-template<typename T> inline
-cv::Affine3<T> cv::Affine3<T>::translate(const Vec3& t) const
-{
-    Mat4 m = matrix;
-    m.val[ 3] += t[0];
-    m.val[ 7] += t[1];
-    m.val[11] += t[2];
-    return m;
-}
-
-template<typename T> inline
-cv::Affine3<T> cv::Affine3<T>::concatenate(const Affine3<T>& affine) const
-{
-    return (*this).rotate(affine.rotation()).translate(affine.translation());
-}
-
-template<typename T> template <typename Y> inline
-cv::Affine3<T>::operator Affine3<Y>() const
-{
-    return Affine3<Y>(matrix);
-}
-
-template<typename T> template <typename Y> inline
-cv::Affine3<Y> cv::Affine3<T>::cast() const
-{
-    return Affine3<Y>(matrix);
-}
-
-template<typename T> inline
-cv::Affine3<T> cv::operator*(const cv::Affine3<T>& affine1, const cv::Affine3<T>& affine2)
-{
-    return affine2.concatenate(affine1);
-}
-
-template<typename T, typename V> inline
-V cv::operator*(const cv::Affine3<T>& affine, const V& v)
-{
-    const typename Affine3<T>::Mat4& m = affine.matrix;
-
-    V r;
-    r.x = m.val[0] * v.x + m.val[1] * v.y + m.val[ 2] * v.z + m.val[ 3];
-    r.y = m.val[4] * v.x + m.val[5] * v.y + m.val[ 6] * v.z + m.val[ 7];
-    r.z = m.val[8] * v.x + m.val[9] * v.y + m.val[10] * v.z + m.val[11];
-    return r;
-}
-
-static inline
-cv::Vec3f cv::operator*(const cv::Affine3f& affine, const cv::Vec3f& v)
-{
-    const cv::Matx44f& m = affine.matrix;
-    cv::Vec3f r;
-    r.val[0] = m.val[0] * v[0] + m.val[1] * v[1] + m.val[ 2] * v[2] + m.val[ 3];
-    r.val[1] = m.val[4] * v[0] + m.val[5] * v[1] + m.val[ 6] * v[2] + m.val[ 7];
-    r.val[2] = m.val[8] * v[0] + m.val[9] * v[1] + m.val[10] * v[2] + m.val[11];
-    return r;
-}
-
-static inline
-cv::Vec3d cv::operator*(const cv::Affine3d& affine, const cv::Vec3d& v)
-{
-    const cv::Matx44d& m = affine.matrix;
-    cv::Vec3d r;
-    r.val[0] = m.val[0] * v[0] + m.val[1] * v[1] + m.val[ 2] * v[2] + m.val[ 3];
-    r.val[1] = m.val[4] * v[0] + m.val[5] * v[1] + m.val[ 6] * v[2] + m.val[ 7];
-    r.val[2] = m.val[8] * v[0] + m.val[9] * v[1] + m.val[10] * v[2] + m.val[11];
-    return r;
-}
-
-
-
-#if defined EIGEN_WORLD_VERSION && defined EIGEN_GEOMETRY_MODULE_H
-
-template<typename T> inline
-cv::Affine3<T>::Affine3(const Eigen::Transform<T, 3, Eigen::Affine, (Eigen::RowMajor)>& affine)
-{
-    cv::Mat(4, 4, cv::traits::Type<T>::value, affine.matrix().data()).copyTo(matrix);
-}
-
-template<typename T> inline
-cv::Affine3<T>::Affine3(const Eigen::Transform<T, 3, Eigen::Affine>& affine)
-{
-    Eigen::Transform<T, 3, Eigen::Affine, (Eigen::RowMajor)> a = affine;
-    cv::Mat(4, 4, cv::traits::Type<T>::value, a.matrix().data()).copyTo(matrix);
-}
-
-template<typename T> inline
-cv::Affine3<T>::operator Eigen::Transform<T, 3, Eigen::Affine, (Eigen::RowMajor)>() const
-{
-    Eigen::Transform<T, 3, Eigen::Affine, (Eigen::RowMajor)> r;
-    cv::Mat hdr(4, 4, cv::traits::Type<T>::value, r.matrix().data());
-    cv::Mat(matrix, false).copyTo(hdr);
-    return r;
-}
-
-template<typename T> inline
-cv::Affine3<T>::operator Eigen::Transform<T, 3, Eigen::Affine>() const
-{
-    return this->operator Eigen::Transform<T, 3, Eigen::Affine, (Eigen::RowMajor)>();
-}
-
-#endif /* defined EIGEN_WORLD_VERSION && defined EIGEN_GEOMETRY_MODULE_H */
-
-//! @endcond
-
-#endif /* __cplusplus */
-
-#endif /* OPENCV_CORE_AFFINE3_HPP */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/async.hpp b/projects/llm_framework/include/opencv4/opencv2/core/async.hpp
deleted file mode 100644
index 54560c7d..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/async.hpp
+++ /dev/null
@@ -1,105 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_CORE_ASYNC_HPP
-#define OPENCV_CORE_ASYNC_HPP
-
-#include <opencv2/core/mat.hpp>
-
-#ifdef CV_CXX11
-//#include <future>
-#include <chrono>
-#endif
-
-namespace cv {
-
-/** @addtogroup core_async
-
-@{
-*/
-
-
-/** @brief Returns result of asynchronous operations
-
-Object has attached asynchronous state.
-Assignment operator doesn't clone asynchronous state (it is shared between all instances).
-
-Result can be fetched via get() method only once.
-
-*/
-class CV_EXPORTS_W AsyncArray
-{
-public:
-    ~AsyncArray() CV_NOEXCEPT;
-    CV_WRAP AsyncArray() CV_NOEXCEPT;
-    AsyncArray(const AsyncArray& o) CV_NOEXCEPT;
-    AsyncArray& operator=(const AsyncArray& o) CV_NOEXCEPT;
-    CV_WRAP void release() CV_NOEXCEPT;
-
-    /** Fetch the result.
-    @param[out] dst destination array
-
-    Waits for result until container has valid result.
-    Throws exception if exception was stored as a result.
-
-    Throws exception on invalid container state.
-
-    @note Result or stored exception can be fetched only once.
-    */
-    CV_WRAP void get(OutputArray dst) const;
-
-    /** Retrieving the result with timeout
-    @param[out] dst destination array
-    @param[in] timeoutNs timeout in nanoseconds, -1 for infinite wait
-
-    @returns true if result is ready, false if the timeout has expired
-
-    @note Result or stored exception can be fetched only once.
-    */
-    bool get(OutputArray dst, int64 timeoutNs) const;
-
-    CV_WRAP inline
-    bool get(OutputArray dst, double timeoutNs) const { return get(dst, (int64)timeoutNs); }
-
-    bool wait_for(int64 timeoutNs) const;
-
-    CV_WRAP inline
-    bool wait_for(double timeoutNs) const { return wait_for((int64)timeoutNs); }
-
-    CV_WRAP bool valid() const CV_NOEXCEPT;
-
-#ifdef CV_CXX11
-    inline AsyncArray(AsyncArray&& o) { p = o.p; o.p = NULL; }
-    inline AsyncArray& operator=(AsyncArray&& o) CV_NOEXCEPT { std::swap(p, o.p); return *this; }
-
-    template<typename _Rep, typename _Period>
-    inline bool get(OutputArray dst, const std::chrono::duration<_Rep, _Period>& timeout)
-    {
-        return get(dst, (int64)(std::chrono::nanoseconds(timeout).count()));
-    }
-
-    template<typename _Rep, typename _Period>
-    inline bool wait_for(const std::chrono::duration<_Rep, _Period>& timeout)
-    {
-        return wait_for((int64)(std::chrono::nanoseconds(timeout).count()));
-    }
-
-#if 0
-    std::future<Mat> getFutureMat() const;
-    std::future<UMat> getFutureUMat() const;
-#endif
-#endif
-
-
-    // PImpl
-    struct Impl; friend struct Impl;
-    inline void* _getImpl() const CV_NOEXCEPT { return p; }
-protected:
-    Impl* p;
-};
-
-
-//! @}
-} // namespace
-#endif // OPENCV_CORE_ASYNC_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/base.hpp b/projects/llm_framework/include/opencv4/opencv2/core/base.hpp
deleted file mode 100644
index 21a61a4e..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/base.hpp
+++ /dev/null
@@ -1,664 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Copyright (C) 2014, Itseez Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_BASE_HPP
-#define OPENCV_CORE_BASE_HPP
-
-#ifndef __cplusplus
-#  error base.hpp header must be compiled as C++
-#endif
-
-#include "opencv2/opencv_modules.hpp"
-
-#include <climits>
-#include <algorithm>
-
-#include "opencv2/core/cvdef.h"
-#include "opencv2/core/cvstd.hpp"
-
-namespace cv
-{
-
-//! @addtogroup core_utils
-//! @{
-
-namespace Error {
-//! error codes
-enum Code {
-    StsOk=                       0,  //!< everything is ok
-    StsBackTrace=               -1,  //!< pseudo error for back trace
-    StsError=                   -2,  //!< unknown /unspecified error
-    StsInternal=                -3,  //!< internal error (bad state)
-    StsNoMem=                   -4,  //!< insufficient memory
-    StsBadArg=                  -5,  //!< function arg/param is bad
-    StsBadFunc=                 -6,  //!< unsupported function
-    StsNoConv=                  -7,  //!< iteration didn't converge
-    StsAutoTrace=               -8,  //!< tracing
-    HeaderIsNull=               -9,  //!< image header is NULL
-    BadImageSize=              -10,  //!< image size is invalid
-    BadOffset=                 -11,  //!< offset is invalid
-    BadDataPtr=                -12,  //!<
-    BadStep=                   -13,  //!< image step is wrong, this may happen for a non-continuous matrix.
-    BadModelOrChSeq=           -14,  //!<
-    BadNumChannels=            -15,  //!< bad number of channels, for example, some functions accept only single channel matrices.
-    BadNumChannel1U=           -16,  //!<
-    BadDepth=                  -17,  //!< input image depth is not supported by the function
-    BadAlphaChannel=           -18,  //!<
-    BadOrder=                  -19,  //!< number of dimensions is out of range
-    BadOrigin=                 -20,  //!< incorrect input origin
-    BadAlign=                  -21,  //!< incorrect input align
-    BadCallBack=               -22,  //!<
-    BadTileSize=               -23,  //!<
-    BadCOI=                    -24,  //!< input COI is not supported
-    BadROISize=                -25,  //!< incorrect input roi
-    MaskIsTiled=               -26,  //!<
-    StsNullPtr=                -27,  //!< null pointer
-    StsVecLengthErr=           -28,  //!< incorrect vector length
-    StsFilterStructContentErr= -29,  //!< incorrect filter structure content
-    StsKernelStructContentErr= -30,  //!< incorrect transform kernel content
-    StsFilterOffsetErr=        -31,  //!< incorrect filter offset value
-    StsBadSize=                -201, //!< the input/output structure size is incorrect
-    StsDivByZero=              -202, //!< division by zero
-    StsInplaceNotSupported=    -203, //!< in-place operation is not supported
-    StsObjectNotFound=         -204, //!< request can't be completed
-    StsUnmatchedFormats=       -205, //!< formats of input/output arrays differ
-    StsBadFlag=                -206, //!< flag is wrong or not supported
-    StsBadPoint=               -207, //!< bad CvPoint
-    StsBadMask=                -208, //!< bad format of mask (neither 8uC1 nor 8sC1)
-    StsUnmatchedSizes=         -209, //!< sizes of input/output structures do not match
-    StsUnsupportedFormat=      -210, //!< the data format/type is not supported by the function
-    StsOutOfRange=             -211, //!< some of parameters are out of range
-    StsParseError=             -212, //!< invalid syntax/structure of the parsed file
-    StsNotImplemented=         -213, //!< the requested function/feature is not implemented
-    StsBadMemBlock=            -214, //!< an allocated block has been corrupted
-    StsAssert=                 -215, //!< assertion failed
-    GpuNotSupported=           -216, //!< no CUDA support
-    GpuApiCallError=           -217, //!< GPU API call error
-    OpenGlNotSupported=        -218, //!< no OpenGL support
-    OpenGlApiCallError=        -219, //!< OpenGL API call error
-    OpenCLApiCallError=        -220, //!< OpenCL API call error
-    OpenCLDoubleNotSupported=  -221,
-    OpenCLInitError=           -222, //!< OpenCL initialization error
-    OpenCLNoAMDBlasFft=        -223
-};
-} //Error
-
-//! @} core_utils
-
-//! @addtogroup core_array
-//! @{
-
-//! matrix decomposition types
-enum DecompTypes {
-    /** Gaussian elimination with the optimal pivot element chosen. */
-    DECOMP_LU       = 0,
-    /** singular value decomposition (SVD) method; the system can be over-defined and/or the matrix
-    src1 can be singular */
-    DECOMP_SVD      = 1,
-    /** eigenvalue decomposition; the matrix src1 must be symmetrical */
-    DECOMP_EIG      = 2,
-    /** Cholesky \f$LL^T\f$ factorization; the matrix src1 must be symmetrical and positively
-    defined */
-    DECOMP_CHOLESKY = 3,
-    /** QR factorization; the system can be over-defined and/or the matrix src1 can be singular */
-    DECOMP_QR       = 4,
-    /** while all the previous flags are mutually exclusive, this flag can be used together with
-    any of the previous; it means that the normal equations
-    \f$\texttt{src1}^T\cdot\texttt{src1}\cdot\texttt{dst}=\texttt{src1}^T\texttt{src2}\f$ are
-    solved instead of the original system
-    \f$\texttt{src1}\cdot\texttt{dst}=\texttt{src2}\f$ */
-    DECOMP_NORMAL   = 16
-};
-
-/** norm types
-
-src1 and src2 denote input arrays.
-*/
-
-enum NormTypes {
-                /**
-                \f[
-                norm =  \forkthree
-                {\|\texttt{src1}\|_{L_{\infty}} =  \max _I | \texttt{src1} (I)|}{if  \(\texttt{normType} = \texttt{NORM_INF}\) }
-                {\|\texttt{src1}-\texttt{src2}\|_{L_{\infty}} =  \max _I | \texttt{src1} (I) -  \texttt{src2} (I)|}{if  \(\texttt{normType} = \texttt{NORM_INF}\) }
-                {\frac{\|\texttt{src1}-\texttt{src2}\|_{L_{\infty}}    }{\|\texttt{src2}\|_{L_{\infty}} }}{if  \(\texttt{normType} = \texttt{NORM_RELATIVE | NORM_INF}\) }
-                \f]
-                */
-                NORM_INF       = 1,
-                /**
-                \f[
-                norm =  \forkthree
-                {\| \texttt{src1} \| _{L_1} =  \sum _I | \texttt{src1} (I)|}{if  \(\texttt{normType} = \texttt{NORM_L1}\)}
-                { \| \texttt{src1} - \texttt{src2} \| _{L_1} =  \sum _I | \texttt{src1} (I) -  \texttt{src2} (I)|}{if  \(\texttt{normType} = \texttt{NORM_L1}\) }
-                { \frac{\|\texttt{src1}-\texttt{src2}\|_{L_1} }{\|\texttt{src2}\|_{L_1}} }{if  \(\texttt{normType} = \texttt{NORM_RELATIVE | NORM_L1}\) }
-                \f]*/
-                 NORM_L1        = 2,
-                 /**
-                 \f[
-                 norm =  \forkthree
-                 { \| \texttt{src1} \| _{L_2} =  \sqrt{\sum_I \texttt{src1}(I)^2} }{if  \(\texttt{normType} = \texttt{NORM_L2}\) }
-                 { \| \texttt{src1} - \texttt{src2} \| _{L_2} =  \sqrt{\sum_I (\texttt{src1}(I) - \texttt{src2}(I))^2} }{if  \(\texttt{normType} = \texttt{NORM_L2}\) }
-                 { \frac{\|\texttt{src1}-\texttt{src2}\|_{L_2} }{\|\texttt{src2}\|_{L_2}} }{if  \(\texttt{normType} = \texttt{NORM_RELATIVE | NORM_L2}\) }
-                 \f]
-                 */
-                 NORM_L2        = 4,
-                 /**
-                 \f[
-                 norm =  \forkthree
-                 { \| \texttt{src1} \| _{L_2} ^{2} = \sum_I \texttt{src1}(I)^2} {if  \(\texttt{normType} = \texttt{NORM_L2SQR}\)}
-                 { \| \texttt{src1} - \texttt{src2} \| _{L_2} ^{2} =  \sum_I (\texttt{src1}(I) - \texttt{src2}(I))^2 }{if  \(\texttt{normType} = \texttt{NORM_L2SQR}\) }
-                 { \left(\frac{\|\texttt{src1}-\texttt{src2}\|_{L_2} }{\|\texttt{src2}\|_{L_2}}\right)^2 }{if  \(\texttt{normType} = \texttt{NORM_RELATIVE | NORM_L2SQR}\) }
-                 \f]
-                 */
-                 NORM_L2SQR     = 5,
-                 /**
-                 In the case of one input array, calculates the Hamming distance of the array from zero,
-                 In the case of two input arrays, calculates the Hamming distance between the arrays.
-                 */
-                 NORM_HAMMING   = 6,
-                 /**
-                 Similar to NORM_HAMMING, but in the calculation, each two bits of the input sequence will
-                 be added and treated as a single bit to be used in the same calculation as NORM_HAMMING.
-                 */
-                 NORM_HAMMING2  = 7,
-                 NORM_TYPE_MASK = 7, //!< bit-mask which can be used to separate norm type from norm flags
-                 NORM_RELATIVE  = 8, //!< flag
-                 NORM_MINMAX    = 32 //!< flag
-               };
-
-//! comparison types
-enum CmpTypes { CMP_EQ = 0, //!< src1 is equal to src2.
-                CMP_GT = 1, //!< src1 is greater than src2.
-                CMP_GE = 2, //!< src1 is greater than or equal to src2.
-                CMP_LT = 3, //!< src1 is less than src2.
-                CMP_LE = 4, //!< src1 is less than or equal to src2.
-                CMP_NE = 5  //!< src1 is unequal to src2.
-              };
-
-//! generalized matrix multiplication flags
-enum GemmFlags { GEMM_1_T = 1, //!< transposes src1
-                 GEMM_2_T = 2, //!< transposes src2
-                 GEMM_3_T = 4 //!< transposes src3
-               };
-
-enum DftFlags {
-    /** performs an inverse 1D or 2D transform instead of the default forward
-        transform. */
-    DFT_INVERSE        = 1,
-    /** scales the result: divide it by the number of array elements. Normally, it is
-        combined with DFT_INVERSE. */
-    DFT_SCALE          = 2,
-    /** performs a forward or inverse transform of every individual row of the input
-        matrix; this flag enables you to transform multiple vectors simultaneously and can be used to
-        decrease the overhead (which is sometimes several times larger than the processing itself) to
-        perform 3D and higher-dimensional transformations and so forth.*/
-    DFT_ROWS           = 4,
-    /** performs a forward transformation of 1D or 2D real array; the result,
-        though being a complex array, has complex-conjugate symmetry (*CCS*, see the function
-        description below for details), and such an array can be packed into a real array of the same
-        size as input, which is the fastest option and which is what the function does by default;
-        however, you may wish to get a full complex array (for simpler spectrum analysis, and so on) -
-        pass the flag to enable the function to produce a full-size complex output array. */
-    DFT_COMPLEX_OUTPUT = 16,
-    /** performs an inverse transformation of a 1D or 2D complex array; the
-        result is normally a complex array of the same size, however, if the input array has
-        conjugate-complex symmetry (for example, it is a result of forward transformation with
-        DFT_COMPLEX_OUTPUT flag), the output is a real array; while the function itself does not
-        check whether the input is symmetrical or not, you can pass the flag and then the function
-        will assume the symmetry and produce the real output array (note that when the input is packed
-        into a real array and inverse transformation is executed, the function treats the input as a
-        packed complex-conjugate symmetrical array, and the output will also be a real array). */
-    DFT_REAL_OUTPUT    = 32,
-    /** specifies that input is complex input. If this flag is set, the input must have 2 channels.
-        On the other hand, for backwards compatibility reason, if input has 2 channels, input is
-        already considered complex. */
-    DFT_COMPLEX_INPUT  = 64,
-    /** performs an inverse 1D or 2D transform instead of the default forward transform. */
-    DCT_INVERSE        = DFT_INVERSE,
-    /** performs a forward or inverse transform of every individual row of the input
-        matrix. This flag enables you to transform multiple vectors simultaneously and can be used to
-        decrease the overhead (which is sometimes several times larger than the processing itself) to
-        perform 3D and higher-dimensional transforms and so forth.*/
-    DCT_ROWS           = DFT_ROWS
-};
-
-//! Various border types, image boundaries are denoted with `|`
-//! @see borderInterpolate, copyMakeBorder
-enum BorderTypes {
-    BORDER_CONSTANT    = 0, //!< `iiiiii|abcdefgh|iiiiiii`  with some specified `i`
-    BORDER_REPLICATE   = 1, //!< `aaaaaa|abcdefgh|hhhhhhh`
-    BORDER_REFLECT     = 2, //!< `fedcba|abcdefgh|hgfedcb`
-    BORDER_WRAP        = 3, //!< `cdefgh|abcdefgh|abcdefg`
-    BORDER_REFLECT_101 = 4, //!< `gfedcb|abcdefgh|gfedcba`
-    BORDER_TRANSPARENT = 5, //!< `uvwxyz|abcdefgh|ijklmno`
-
-    BORDER_REFLECT101  = BORDER_REFLECT_101, //!< same as BORDER_REFLECT_101
-    BORDER_DEFAULT     = BORDER_REFLECT_101, //!< same as BORDER_REFLECT_101
-    BORDER_ISOLATED    = 16 //!< do not look outside of ROI
-};
-
-//! @} core_array
-
-//! @addtogroup core_utils
-//! @{
-
-/*! @brief Signals an error and raises the exception.
-
-By default the function prints information about the error to stderr,
-then it either stops if setBreakOnError() had been called before or raises the exception.
-It is possible to alternate error processing by using redirectError().
-@param _code - error code (Error::Code)
-@param _err - error description
-@param _func - function name. Available only when the compiler supports getting it
-@param _file - source file name where the error has occurred
-@param _line - line number in the source file where the error has occurred
-@see CV_Error, CV_Error_, CV_Assert, CV_DbgAssert
- */
-CV_EXPORTS CV_NORETURN void error(int _code, const String& _err, const char* _func, const char* _file, int _line);
-
-#ifdef CV_STATIC_ANALYSIS
-
-// In practice, some macro are not processed correctly (noreturn is not detected).
-// We need to use simplified definition for them.
-#define CV_Error(code, msg) do { (void)(code); (void)(msg); abort(); } while (0)
-#define CV_Error_(code, args) do { (void)(code); (void)(cv::format args); abort(); } while (0)
-#define CV_Assert( expr ) do { if (!(expr)) abort(); } while (0)
-
-#else // CV_STATIC_ANALYSIS
-
-/** @brief Call the error handler.
-
-Currently, the error handler prints the error code and the error message to the standard
-error stream `stderr`. In the Debug configuration, it then provokes memory access violation, so that
-the execution stack and all the parameters can be analyzed by the debugger. In the Release
-configuration, the exception is thrown.
-
-@param code one of Error::Code
-@param msg error message
-*/
-#define CV_Error( code, msg ) cv::error( code, msg, CV_Func, __FILE__, __LINE__ )
-
-/**  @brief Call the error handler.
-
-This macro can be used to construct an error message on-fly to include some dynamic information,
-for example:
-@code
-    // note the extra parentheses around the formatted text message
-    CV_Error_(Error::StsOutOfRange,
-    ("the value at (%d, %d)=%g is out of range", badPt.x, badPt.y, badValue));
-@endcode
-@param code one of Error::Code
-@param args printf-like formatted error message in parentheses
-*/
-#define CV_Error_( code, args ) cv::error( code, cv::format args, CV_Func, __FILE__, __LINE__ )
-
-/** @brief Checks a condition at runtime and throws exception if it fails
-
-The macros CV_Assert (and CV_DbgAssert(expr)) evaluate the specified expression. If it is 0, the macros
-raise an error (see cv::error). The macro CV_Assert checks the condition in both Debug and Release
-configurations while CV_DbgAssert is only retained in the Debug configuration.
-*/
-#define CV_Assert( expr ) do { if(!!(expr)) ; else cv::error( cv::Error::StsAssert, #expr, CV_Func, __FILE__, __LINE__ ); } while(0)
-
-#endif // CV_STATIC_ANALYSIS
-
-//! @cond IGNORED
-#if !defined(__OPENCV_BUILD)  // TODO: backward compatibility only
-#ifndef CV_ErrorNoReturn
-#define CV_ErrorNoReturn CV_Error
-#endif
-#ifndef CV_ErrorNoReturn_
-#define CV_ErrorNoReturn_ CV_Error_
-#endif
-#endif
-
-#define CV_Assert_1 CV_Assert
-#define CV_Assert_2( expr, ... ) CV_Assert_1(expr); __CV_EXPAND(CV_Assert_1( __VA_ARGS__ ))
-#define CV_Assert_3( expr, ... ) CV_Assert_1(expr); __CV_EXPAND(CV_Assert_2( __VA_ARGS__ ))
-#define CV_Assert_4( expr, ... ) CV_Assert_1(expr); __CV_EXPAND(CV_Assert_3( __VA_ARGS__ ))
-#define CV_Assert_5( expr, ... ) CV_Assert_1(expr); __CV_EXPAND(CV_Assert_4( __VA_ARGS__ ))
-#define CV_Assert_6( expr, ... ) CV_Assert_1(expr); __CV_EXPAND(CV_Assert_5( __VA_ARGS__ ))
-#define CV_Assert_7( expr, ... ) CV_Assert_1(expr); __CV_EXPAND(CV_Assert_6( __VA_ARGS__ ))
-#define CV_Assert_8( expr, ... ) CV_Assert_1(expr); __CV_EXPAND(CV_Assert_7( __VA_ARGS__ ))
-#define CV_Assert_9( expr, ... ) CV_Assert_1(expr); __CV_EXPAND(CV_Assert_8( __VA_ARGS__ ))
-#define CV_Assert_10( expr, ... ) CV_Assert_1(expr); __CV_EXPAND(CV_Assert_9( __VA_ARGS__ ))
-
-#define CV_Assert_N(...) do { __CV_EXPAND(__CV_CAT(CV_Assert_, __CV_VA_NUM_ARGS(__VA_ARGS__)) (__VA_ARGS__)); } while(0)
-
-//! @endcond
-
-#if defined _DEBUG || defined CV_STATIC_ANALYSIS
-#  define CV_DbgAssert(expr) CV_Assert(expr)
-#else
-/** replaced with CV_Assert(expr) in Debug configuration */
-#  define CV_DbgAssert(expr)
-#endif
-
-/*
- * Hamming distance functor - counts the bit differences between two strings - useful for the Brief descriptor
- * bit count of A exclusive XOR'ed with B
- */
-struct CV_EXPORTS Hamming
-{
-    static const NormTypes normType = NORM_HAMMING;
-    typedef unsigned char ValueType;
-    typedef int ResultType;
-
-    /** this will count the bits in a ^ b
-     */
-    ResultType operator()( const unsigned char* a, const unsigned char* b, int size ) const;
-};
-
-typedef Hamming HammingLUT;
-
-/////////////////////////////////// inline norms ////////////////////////////////////
-
-template<typename _Tp> inline _Tp cv_abs(_Tp x) { return std::abs(x); }
-inline int cv_abs(uchar x) { return x; }
-inline int cv_abs(schar x) { return std::abs(x); }
-inline int cv_abs(ushort x) { return x; }
-inline int cv_abs(short x) { return std::abs(x); }
-
-template<typename _Tp, typename _AccTp> static inline
-_AccTp normL2Sqr(const _Tp* a, int n)
-{
-    _AccTp s = 0;
-    int i=0;
-#if CV_ENABLE_UNROLLED
-    for( ; i <= n - 4; i += 4 )
-    {
-        _AccTp v0 = a[i], v1 = a[i+1], v2 = a[i+2], v3 = a[i+3];
-        s += v0*v0 + v1*v1 + v2*v2 + v3*v3;
-    }
-#endif
-    for( ; i < n; i++ )
-    {
-        _AccTp v = a[i];
-        s += v*v;
-    }
-    return s;
-}
-
-template<typename _Tp, typename _AccTp> static inline
-_AccTp normL1(const _Tp* a, int n)
-{
-    _AccTp s = 0;
-    int i = 0;
-#if CV_ENABLE_UNROLLED
-    for(; i <= n - 4; i += 4 )
-    {
-        s += (_AccTp)cv_abs(a[i]) + (_AccTp)cv_abs(a[i+1]) +
-            (_AccTp)cv_abs(a[i+2]) + (_AccTp)cv_abs(a[i+3]);
-    }
-#endif
-    for( ; i < n; i++ )
-        s += cv_abs(a[i]);
-    return s;
-}
-
-template<typename _Tp, typename _AccTp> static inline
-_AccTp normInf(const _Tp* a, int n)
-{
-    _AccTp s = 0;
-    for( int i = 0; i < n; i++ )
-        s = std::max(s, (_AccTp)cv_abs(a[i]));
-    return s;
-}
-
-template<typename _Tp, typename _AccTp> static inline
-_AccTp normL2Sqr(const _Tp* a, const _Tp* b, int n)
-{
-    _AccTp s = 0;
-    int i= 0;
-#if CV_ENABLE_UNROLLED
-    for(; i <= n - 4; i += 4 )
-    {
-        _AccTp v0 = _AccTp(a[i] - b[i]), v1 = _AccTp(a[i+1] - b[i+1]), v2 = _AccTp(a[i+2] - b[i+2]), v3 = _AccTp(a[i+3] - b[i+3]);
-        s += v0*v0 + v1*v1 + v2*v2 + v3*v3;
-    }
-#endif
-    for( ; i < n; i++ )
-    {
-        _AccTp v = _AccTp(a[i] - b[i]);
-        s += v*v;
-    }
-    return s;
-}
-
-static inline float normL2Sqr(const float* a, const float* b, int n)
-{
-    float s = 0.f;
-    for( int i = 0; i < n; i++ )
-    {
-        float v = a[i] - b[i];
-        s += v*v;
-    }
-    return s;
-}
-
-template<typename _Tp, typename _AccTp> static inline
-_AccTp normL1(const _Tp* a, const _Tp* b, int n)
-{
-    _AccTp s = 0;
-    int i= 0;
-#if CV_ENABLE_UNROLLED
-    for(; i <= n - 4; i += 4 )
-    {
-        _AccTp v0 = _AccTp(a[i] - b[i]), v1 = _AccTp(a[i+1] - b[i+1]), v2 = _AccTp(a[i+2] - b[i+2]), v3 = _AccTp(a[i+3] - b[i+3]);
-        s += std::abs(v0) + std::abs(v1) + std::abs(v2) + std::abs(v3);
-    }
-#endif
-    for( ; i < n; i++ )
-    {
-        _AccTp v = _AccTp(a[i] - b[i]);
-        s += std::abs(v);
-    }
-    return s;
-}
-
-inline float normL1(const float* a, const float* b, int n)
-{
-    float s = 0.f;
-    for( int i = 0; i < n; i++ )
-    {
-        s += std::abs(a[i] - b[i]);
-    }
-    return s;
-}
-
-inline int normL1(const uchar* a, const uchar* b, int n)
-{
-    int s = 0;
-    for( int i = 0; i < n; i++ )
-    {
-        s += std::abs(a[i] - b[i]);
-    }
-    return s;
-}
-
-template<typename _Tp, typename _AccTp> static inline
-_AccTp normInf(const _Tp* a, const _Tp* b, int n)
-{
-    _AccTp s = 0;
-    for( int i = 0; i < n; i++ )
-    {
-        _AccTp v0 = a[i] - b[i];
-        s = std::max(s, std::abs(v0));
-    }
-    return s;
-}
-
-/** @brief Computes the cube root of an argument.
-
- The function cubeRoot computes \f$\sqrt[3]{\texttt{val}}\f$. Negative arguments are handled correctly.
- NaN and Inf are not handled. The accuracy approaches the maximum possible accuracy for
- single-precision data.
- @param val A function argument.
- */
-CV_EXPORTS_W float cubeRoot(float val);
-
-/** @overload
-
-cubeRoot with argument of `double` type calls `std::cbrt(double)`
-*/
-static inline
-double cubeRoot(double val)
-{
-    return std::cbrt(val);
-}
-
-/** @brief Calculates the angle of a 2D vector in degrees.
-
- The function fastAtan2 calculates the full-range angle of an input 2D vector. The angle is measured
- in degrees and varies from 0 to 360 degrees. The accuracy is about 0.3 degrees.
- @param x x-coordinate of the vector.
- @param y y-coordinate of the vector.
- */
-CV_EXPORTS_W float fastAtan2(float y, float x);
-
-/** proxy for hal::LU */
-CV_EXPORTS int LU(float* A, size_t astep, int m, float* b, size_t bstep, int n);
-/** proxy for hal::LU */
-CV_EXPORTS int LU(double* A, size_t astep, int m, double* b, size_t bstep, int n);
-/** proxy for hal::Cholesky */
-CV_EXPORTS bool Cholesky(float* A, size_t astep, int m, float* b, size_t bstep, int n);
-/** proxy for hal::Cholesky */
-CV_EXPORTS bool Cholesky(double* A, size_t astep, int m, double* b, size_t bstep, int n);
-
-////////////////// forward declarations for important OpenCV types //////////////////
-
-//! @cond IGNORED
-
-template<typename _Tp, int cn> class Vec;
-template<typename _Tp, int m, int n> class Matx;
-
-template<typename _Tp> class Complex;
-template<typename _Tp> class Point_;
-template<typename _Tp> class Point3_;
-template<typename _Tp> class Size_;
-template<typename _Tp> class Rect_;
-template<typename _Tp> class Scalar_;
-
-class CV_EXPORTS RotatedRect;
-class CV_EXPORTS Range;
-class CV_EXPORTS TermCriteria;
-class CV_EXPORTS KeyPoint;
-class CV_EXPORTS DMatch;
-class CV_EXPORTS RNG;
-
-class CV_EXPORTS Mat;
-class CV_EXPORTS MatExpr;
-
-class CV_EXPORTS UMat;
-
-class CV_EXPORTS SparseMat;
-typedef Mat MatND;
-
-template<typename _Tp> class Mat_;
-template<typename _Tp> class SparseMat_;
-
-class CV_EXPORTS MatConstIterator;
-class CV_EXPORTS SparseMatIterator;
-class CV_EXPORTS SparseMatConstIterator;
-template<typename _Tp> class MatIterator_;
-template<typename _Tp> class MatConstIterator_;
-template<typename _Tp> class SparseMatIterator_;
-template<typename _Tp> class SparseMatConstIterator_;
-
-namespace ogl
-{
-    class CV_EXPORTS Buffer;
-    class CV_EXPORTS Texture2D;
-    class CV_EXPORTS Arrays;
-}
-
-namespace cuda
-{
-    class CV_EXPORTS GpuMat;
-    class CV_EXPORTS HostMem;
-    class CV_EXPORTS Stream;
-    class CV_EXPORTS Event;
-}
-
-namespace cudev
-{
-    template <typename _Tp> class GpuMat_;
-}
-
-namespace ipp
-{
-CV_EXPORTS   unsigned long long getIppFeatures();
-CV_EXPORTS   void setIppStatus(int status, const char * const funcname = NULL, const char * const filename = NULL,
-                             int line = 0);
-CV_EXPORTS   int getIppStatus();
-CV_EXPORTS   String getIppErrorLocation();
-CV_EXPORTS_W bool   useIPP();
-CV_EXPORTS_W void   setUseIPP(bool flag);
-CV_EXPORTS_W String getIppVersion();
-
-// IPP Not-Exact mode. This function may force use of IPP then both IPP and OpenCV provide proper results
-// but have internal accuracy differences which have too much direct or indirect impact on accuracy tests.
-CV_EXPORTS_W bool useIPP_NotExact();
-CV_EXPORTS_W void setUseIPP_NotExact(bool flag);
-#ifndef DISABLE_OPENCV_3_COMPATIBILITY
-static inline bool useIPP_NE() { return useIPP_NotExact(); }
-static inline void setUseIPP_NE(bool flag) { setUseIPP_NotExact(flag); }
-#endif
-
-} // ipp
-
-//! @endcond
-
-//! @} core_utils
-
-
-
-
-} // cv
-
-#include "opencv2/core/neon_utils.hpp"
-#include "opencv2/core/vsx_utils.hpp"
-#include "opencv2/core/check.hpp"
-
-#endif //OPENCV_CORE_BASE_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/bindings_utils.hpp b/projects/llm_framework/include/opencv4/opencv2/core/bindings_utils.hpp
deleted file mode 100644
index 22a86ff9..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/bindings_utils.hpp
+++ /dev/null
@@ -1,248 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_CORE_BINDINGS_UTILS_HPP
-#define OPENCV_CORE_BINDINGS_UTILS_HPP
-
-#include <opencv2/core/async.hpp>
-#include <opencv2/core/detail/async_promise.hpp>
-#include <opencv2/core/utils/logger.hpp>
-
-#include <stdexcept>
-
-namespace cv { namespace utils {
-//! @addtogroup core_utils
-//! @{
-
-CV_EXPORTS_W String dumpInputArray(InputArray argument);
-
-CV_EXPORTS_W String dumpInputArrayOfArrays(InputArrayOfArrays argument);
-
-CV_EXPORTS_W String dumpInputOutputArray(InputOutputArray argument);
-
-CV_EXPORTS_W String dumpInputOutputArrayOfArrays(InputOutputArrayOfArrays argument);
-
-CV_WRAP static inline
-String dumpBool(bool argument)
-{
-    return (argument) ? String("Bool: True") : String("Bool: False");
-}
-
-CV_WRAP static inline
-String dumpInt(int argument)
-{
-    return cv::format("Int: %d", argument);
-}
-
-CV_WRAP static inline
-String dumpSizeT(size_t argument)
-{
-    std::ostringstream oss("size_t: ", std::ios::ate);
-    oss << argument;
-    return oss.str();
-}
-
-CV_WRAP static inline
-String dumpFloat(float argument)
-{
-    return cv::format("Float: %.2f", argument);
-}
-
-CV_WRAP static inline
-String dumpDouble(double argument)
-{
-    return cv::format("Double: %.2f", argument);
-}
-
-CV_WRAP static inline
-String dumpCString(const char* argument)
-{
-    return cv::format("String: %s", argument);
-}
-
-CV_WRAP static inline
-String dumpString(const String& argument)
-{
-    return cv::format("String: %s", argument.c_str());
-}
-
-CV_WRAP static inline
-String testOverloadResolution(int value, const Point& point = Point(42, 24))
-{
-    return format("overload (int=%d, point=(x=%d, y=%d))", value, point.x,
-                  point.y);
-}
-
-CV_WRAP static inline
-String testOverloadResolution(const Rect& rect)
-{
-    return format("overload (rect=(x=%d, y=%d, w=%d, h=%d))", rect.x, rect.y,
-                  rect.width, rect.height);
-}
-
-CV_WRAP static inline
-String dumpRect(const Rect& argument)
-{
-    return format("rect: (x=%d, y=%d, w=%d, h=%d)", argument.x, argument.y,
-                  argument.width, argument.height);
-}
-
-CV_WRAP static inline
-String dumpTermCriteria(const TermCriteria& argument)
-{
-    return format("term_criteria: (type=%d, max_count=%d, epsilon=%lf",
-                  argument.type, argument.maxCount, argument.epsilon);
-}
-
-CV_WRAP static inline
-String dumpRotatedRect(const RotatedRect& argument)
-{
-    return format("rotated_rect: (c_x=%f, c_y=%f, w=%f, h=%f, a=%f)",
-                  argument.center.x, argument.center.y, argument.size.width,
-                  argument.size.height, argument.angle);
-}
-
-CV_WRAP static inline
-RotatedRect testRotatedRect(float x, float y, float w, float h, float angle)
-{
-    return RotatedRect(Point2f(x, y), Size2f(w, h), angle);
-}
-
-CV_WRAP static inline
-std::vector<RotatedRect> testRotatedRectVector(float x, float y, float w, float h, float angle)
-{
-    std::vector<RotatedRect> result;
-    for (int i = 0; i < 10; i++)
-        result.push_back(RotatedRect(Point2f(x + i, y + 2 * i), Size2f(w, h), angle + 10 * i));
-    return result;
-}
-
-CV_WRAP static inline
-String dumpRange(const Range& argument)
-{
-    if (argument == Range::all())
-    {
-        return "range: all";
-    }
-    else
-    {
-        return format("range: (s=%d, e=%d)", argument.start, argument.end);
-    }
-}
-
-CV_WRAP static inline
-int testOverwriteNativeMethod(int argument)
-{
-    return argument;
-}
-
-CV_WRAP static inline
-String testReservedKeywordConversion(int positional_argument, int lambda = 2, int from = 3)
-{
-    return format("arg=%d, lambda=%d, from=%d", positional_argument, lambda, from);
-}
-
-CV_EXPORTS_W String dumpVectorOfInt(const std::vector<int>& vec);
-
-CV_EXPORTS_W String dumpVectorOfDouble(const std::vector<double>& vec);
-
-CV_EXPORTS_W String dumpVectorOfRect(const std::vector<Rect>& vec);
-
-CV_WRAP static inline
-void generateVectorOfRect(size_t len, CV_OUT std::vector<Rect>& vec)
-{
-    vec.resize(len);
-    if (len > 0)
-    {
-        RNG rng(12345);
-        Mat tmp(static_cast<int>(len), 1, CV_32SC4);
-        rng.fill(tmp, RNG::UNIFORM, 10, 20);
-        tmp.copyTo(vec);
-    }
-}
-
-CV_WRAP static inline
-void generateVectorOfInt(size_t len, CV_OUT std::vector<int>& vec)
-{
-    vec.resize(len);
-    if (len > 0)
-    {
-        RNG rng(554433);
-        Mat tmp(static_cast<int>(len), 1, CV_32SC1);
-        rng.fill(tmp, RNG::UNIFORM, -10, 10);
-        tmp.copyTo(vec);
-    }
-}
-
-CV_WRAP static inline
-void generateVectorOfMat(size_t len, int rows, int cols, int dtype, CV_OUT std::vector<Mat>& vec)
-{
-    vec.resize(len);
-    if (len > 0)
-    {
-        RNG rng(65431);
-        for (size_t i = 0; i < len; ++i)
-        {
-            vec[i].create(rows, cols, dtype);
-            rng.fill(vec[i], RNG::UNIFORM, 0, 10);
-        }
-    }
-}
-
-CV_WRAP static inline
-void testRaiseGeneralException()
-{
-    throw std::runtime_error("exception text");
-}
-
-CV_WRAP static inline
-AsyncArray testAsyncArray(InputArray argument)
-{
-    AsyncPromise p;
-    p.setValue(argument);
-    return p.getArrayResult();
-}
-
-CV_WRAP static inline
-AsyncArray testAsyncException()
-{
-    AsyncPromise p;
-    try
-    {
-        CV_Error(Error::StsOk, "Test: Generated async error");
-    }
-    catch (const cv::Exception& e)
-    {
-        p.setException(e);
-    }
-    return p.getArrayResult();
-}
-
-namespace fs {
-    CV_EXPORTS_W cv::String getCacheDirectoryForDownloads();
-} // namespace fs
-
-//! @}  // core_utils
-}  // namespace cv::utils
-
-//! @cond IGNORED
-
-CV_WRAP static inline
-int setLogLevel(int level)
-{
-    // NB: Binding generators doesn't work with enums properly yet, so we define separate overload here
-    return cv::utils::logging::setLogLevel((cv::utils::logging::LogLevel)level);
-}
-
-CV_WRAP static inline
-int getLogLevel()
-{
-    return cv::utils::logging::getLogLevel();
-}
-
-//! @endcond IGNORED
-
-} // namespaces cv /  utils
-
-#endif // OPENCV_CORE_BINDINGS_UTILS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/bufferpool.hpp b/projects/llm_framework/include/opencv4/opencv2/core/bufferpool.hpp
deleted file mode 100644
index 4698e5da..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/bufferpool.hpp
+++ /dev/null
@@ -1,40 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-//
-// Copyright (C) 2014, Advanced Micro Devices, Inc., all rights reserved.
-
-#ifndef OPENCV_CORE_BUFFER_POOL_HPP
-#define OPENCV_CORE_BUFFER_POOL_HPP
-
-#ifdef _MSC_VER
-#pragma warning(push)
-#pragma warning(disable: 4265)
-#endif
-
-namespace cv
-{
-
-//! @addtogroup core
-//! @{
-
-class BufferPoolController
-{
-protected:
-    ~BufferPoolController() { }
-public:
-    virtual size_t getReservedSize() const = 0;
-    virtual size_t getMaxReservedSize() const = 0;
-    virtual void setMaxReservedSize(size_t size) = 0;
-    virtual void freeAllReservedBuffers() = 0;
-};
-
-//! @}
-
-}
-
-#ifdef _MSC_VER
-#pragma warning(pop)
-#endif
-
-#endif // OPENCV_CORE_BUFFER_POOL_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/check.hpp b/projects/llm_framework/include/opencv4/opencv2/core/check.hpp
deleted file mode 100644
index d975223c..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/check.hpp
+++ /dev/null
@@ -1,160 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_CORE_CHECK_HPP
-#define OPENCV_CORE_CHECK_HPP
-
-#include <opencv2/core/base.hpp>
-
-namespace cv {
-
-/** Returns string of cv::Mat depth value: CV_8U -> "CV_8U" or "<invalid depth>" */
-CV_EXPORTS const char* depthToString(int depth);
-
-/** Returns string of cv::Mat depth value: CV_8UC3 -> "CV_8UC3" or "<invalid type>" */
-CV_EXPORTS const String typeToString(int type);
-
-
-//! @cond IGNORED
-namespace detail {
-
-/** Returns string of cv::Mat depth value: CV_8U -> "CV_8U" or NULL */
-CV_EXPORTS const char* depthToString_(int depth);
-
-/** Returns string of cv::Mat depth value: CV_8UC3 -> "CV_8UC3" or cv::String() */
-CV_EXPORTS const cv::String typeToString_(int type);
-
-enum TestOp {
-  TEST_CUSTOM = 0,
-  TEST_EQ = 1,
-  TEST_NE = 2,
-  TEST_LE = 3,
-  TEST_LT = 4,
-  TEST_GE = 5,
-  TEST_GT = 6,
-  CV__LAST_TEST_OP
-};
-
-struct CheckContext {
-    const char* func;
-    const char* file;
-    int line;
-    enum TestOp testOp;
-    const char* message;
-    const char* p1_str;
-    const char* p2_str;
-};
-
-#ifndef CV__CHECK_FILENAME
-# define CV__CHECK_FILENAME __FILE__
-#endif
-
-#ifndef CV__CHECK_FUNCTION
-# if defined _MSC_VER
-#   define CV__CHECK_FUNCTION __FUNCSIG__
-# elif defined __GNUC__
-#   define CV__CHECK_FUNCTION __PRETTY_FUNCTION__
-# else
-#   define CV__CHECK_FUNCTION "<unknown>"
-# endif
-#endif
-
-#define CV__CHECK_LOCATION_VARNAME(id) CVAUX_CONCAT(CVAUX_CONCAT(__cv_check_, id), __LINE__)
-#define CV__DEFINE_CHECK_CONTEXT(id, message, testOp, p1_str, p2_str) \
-    static const cv::detail::CheckContext CV__CHECK_LOCATION_VARNAME(id) = \
-            { CV__CHECK_FUNCTION, CV__CHECK_FILENAME, __LINE__, testOp, "" message, "" p1_str, "" p2_str }
-
-CV_EXPORTS void CV_NORETURN check_failed_auto(const int v1, const int v2, const CheckContext& ctx);
-CV_EXPORTS void CV_NORETURN check_failed_auto(const size_t v1, const size_t v2, const CheckContext& ctx);
-CV_EXPORTS void CV_NORETURN check_failed_auto(const float v1, const float v2, const CheckContext& ctx);
-CV_EXPORTS void CV_NORETURN check_failed_auto(const double v1, const double v2, const CheckContext& ctx);
-CV_EXPORTS void CV_NORETURN check_failed_auto(const Size_<int> v1, const Size_<int> v2, const CheckContext& ctx);
-CV_EXPORTS void CV_NORETURN check_failed_MatDepth(const int v1, const int v2, const CheckContext& ctx);
-CV_EXPORTS void CV_NORETURN check_failed_MatType(const int v1, const int v2, const CheckContext& ctx);
-CV_EXPORTS void CV_NORETURN check_failed_MatChannels(const int v1, const int v2, const CheckContext& ctx);
-
-CV_EXPORTS void CV_NORETURN check_failed_auto(const int v, const CheckContext& ctx);
-CV_EXPORTS void CV_NORETURN check_failed_auto(const size_t v, const CheckContext& ctx);
-CV_EXPORTS void CV_NORETURN check_failed_auto(const float v, const CheckContext& ctx);
-CV_EXPORTS void CV_NORETURN check_failed_auto(const double v, const CheckContext& ctx);
-CV_EXPORTS void CV_NORETURN check_failed_auto(const Size_<int> v, const CheckContext& ctx);
-CV_EXPORTS void CV_NORETURN check_failed_auto(const std::string& v1, const CheckContext& ctx);
-CV_EXPORTS void CV_NORETURN check_failed_MatDepth(const int v, const CheckContext& ctx);
-CV_EXPORTS void CV_NORETURN check_failed_MatType(const int v, const CheckContext& ctx);
-CV_EXPORTS void CV_NORETURN check_failed_MatChannels(const int v, const CheckContext& ctx);
-
-
-#define CV__TEST_EQ(v1, v2) ((v1) == (v2))
-#define CV__TEST_NE(v1, v2) ((v1) != (v2))
-#define CV__TEST_LE(v1, v2) ((v1) <= (v2))
-#define CV__TEST_LT(v1, v2) ((v1) < (v2))
-#define CV__TEST_GE(v1, v2) ((v1) >= (v2))
-#define CV__TEST_GT(v1, v2) ((v1) > (v2))
-
-#define CV__CHECK(id, op, type, v1, v2, v1_str, v2_str, msg_str) do { \
-    if(CV__TEST_##op((v1), (v2))) ; else { \
-        CV__DEFINE_CHECK_CONTEXT(id, msg_str, cv::detail::TEST_ ## op, v1_str, v2_str); \
-        cv::detail::check_failed_ ## type((v1), (v2), CV__CHECK_LOCATION_VARNAME(id)); \
-    } \
-} while (0)
-
-#define CV__CHECK_CUSTOM_TEST(id, type, v, test_expr, v_str, test_expr_str, msg_str) do { \
-    if(!!(test_expr)) ; else { \
-        CV__DEFINE_CHECK_CONTEXT(id, msg_str, cv::detail::TEST_CUSTOM, v_str, test_expr_str); \
-        cv::detail::check_failed_ ## type((v), CV__CHECK_LOCATION_VARNAME(id)); \
-    } \
-} while (0)
-
-} // namespace
-//! @endcond
-
-
-/// Supported values of these types: int, float, double
-#define CV_CheckEQ(v1, v2, msg)  CV__CHECK(_, EQ, auto, v1, v2, #v1, #v2, msg)
-#define CV_CheckNE(v1, v2, msg)  CV__CHECK(_, NE, auto, v1, v2, #v1, #v2, msg)
-#define CV_CheckLE(v1, v2, msg)  CV__CHECK(_, LE, auto, v1, v2, #v1, #v2, msg)
-#define CV_CheckLT(v1, v2, msg)  CV__CHECK(_, LT, auto, v1, v2, #v1, #v2, msg)
-#define CV_CheckGE(v1, v2, msg)  CV__CHECK(_, GE, auto, v1, v2, #v1, #v2, msg)
-#define CV_CheckGT(v1, v2, msg)  CV__CHECK(_, GT, auto, v1, v2, #v1, #v2, msg)
-
-/// Check with additional "decoding" of type values in error message
-#define CV_CheckTypeEQ(t1, t2, msg)  CV__CHECK(_, EQ, MatType, t1, t2, #t1, #t2, msg)
-/// Check with additional "decoding" of depth values in error message
-#define CV_CheckDepthEQ(d1, d2, msg)  CV__CHECK(_, EQ, MatDepth, d1, d2, #d1, #d2, msg)
-
-#define CV_CheckChannelsEQ(c1, c2, msg)  CV__CHECK(_, EQ, MatChannels, c1, c2, #c1, #c2, msg)
-
-/// Example: type == CV_8UC1 || type == CV_8UC3
-#define CV_CheckType(t, test_expr, msg)  CV__CHECK_CUSTOM_TEST(_, MatType, t, (test_expr), #t, #test_expr, msg)
-
-/// Example: depth == CV_32F || depth == CV_64F
-#define CV_CheckDepth(t, test_expr, msg)  CV__CHECK_CUSTOM_TEST(_, MatDepth, t, (test_expr), #t, #test_expr, msg)
-
-/// Example: v == A || v == B
-#define CV_Check(v, test_expr, msg)  CV__CHECK_CUSTOM_TEST(_, auto, v, (test_expr), #v, #test_expr, msg)
-
-/// Some complex conditions: CV_Check(src2, src2.empty() || (src2.type() == src1.type() && src2.size() == src1.size()), "src2 should have same size/type as src1")
-// TODO define pretty-printers
-
-#ifndef NDEBUG
-#define CV_DbgCheck(v, test_expr, msg)  CV__CHECK_CUSTOM_TEST(_, auto, v, (test_expr), #v, #test_expr, msg)
-#define CV_DbgCheckEQ(v1, v2, msg)  CV__CHECK(_, EQ, auto, v1, v2, #v1, #v2, msg)
-#define CV_DbgCheckNE(v1, v2, msg)  CV__CHECK(_, NE, auto, v1, v2, #v1, #v2, msg)
-#define CV_DbgCheckLE(v1, v2, msg)  CV__CHECK(_, LE, auto, v1, v2, #v1, #v2, msg)
-#define CV_DbgCheckLT(v1, v2, msg)  CV__CHECK(_, LT, auto, v1, v2, #v1, #v2, msg)
-#define CV_DbgCheckGE(v1, v2, msg)  CV__CHECK(_, GE, auto, v1, v2, #v1, #v2, msg)
-#define CV_DbgCheckGT(v1, v2, msg)  CV__CHECK(_, GT, auto, v1, v2, #v1, #v2, msg)
-#else
-#define CV_DbgCheck(v, test_expr, msg)  do { } while (0)
-#define CV_DbgCheckEQ(v1, v2, msg)  do { } while (0)
-#define CV_DbgCheckNE(v1, v2, msg)  do { } while (0)
-#define CV_DbgCheckLE(v1, v2, msg)  do { } while (0)
-#define CV_DbgCheckLT(v1, v2, msg)  do { } while (0)
-#define CV_DbgCheckGE(v1, v2, msg)  do { } while (0)
-#define CV_DbgCheckGT(v1, v2, msg)  do { } while (0)
-#endif
-
-} // namespace
-
-#endif // OPENCV_CORE_CHECK_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/core.hpp b/projects/llm_framework/include/opencv4/opencv2/core/core.hpp
deleted file mode 100644
index 43891835..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/core.hpp
+++ /dev/null
@@ -1,48 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifdef __OPENCV_BUILD
-#error this is a compatibility header which should not be used inside the OpenCV library
-#endif
-
-#include "opencv2/core.hpp"
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/core_c.h b/projects/llm_framework/include/opencv4/opencv2/core/core_c.h
deleted file mode 100644
index 09ac1e78..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/core_c.h
+++ /dev/null
@@ -1,3125 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-
-#ifndef OPENCV_CORE_C_H
-#define OPENCV_CORE_C_H
-
-#include "opencv2/core/types_c.h"
-
-#ifdef __cplusplus
-#  ifdef _MSC_VER
-/* disable warning C4190: 'function' has C-linkage specified, but returns UDT 'typename'
-                          which is incompatible with C
-
-   It is OK to disable it because we only extend few plain structures with
-   C++ constructors for simpler interoperability with C++ API of the library
-*/
-#    pragma warning(disable:4190)
-#  elif defined __clang__ && __clang_major__ >= 3
-#    pragma GCC diagnostic ignored "-Wreturn-type-c-linkage"
-#  endif
-#endif
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/** @addtogroup core_c
-    @{
-*/
-
-/****************************************************************************************\
-*          Array allocation, deallocation, initialization and access to elements         *
-\****************************************************************************************/
-
-/** `malloc` wrapper.
-   If there is no enough memory, the function
-   (as well as other OpenCV functions that call cvAlloc)
-   raises an error. */
-CVAPI(void*)  cvAlloc( size_t size );
-
-/** `free` wrapper.
-   Here and further all the memory releasing functions
-   (that all call cvFree) take double pointer in order to
-   to clear pointer to the data after releasing it.
-   Passing pointer to NULL pointer is Ok: nothing happens in this case
-*/
-CVAPI(void)   cvFree_( void* ptr );
-#define cvFree(ptr) (cvFree_(*(ptr)), *(ptr)=0)
-
-/** @brief Creates an image header but does not allocate the image data.
-
-@param size Image width and height
-@param depth Image depth (see cvCreateImage )
-@param channels Number of channels (see cvCreateImage )
- */
-CVAPI(IplImage*)  cvCreateImageHeader( CvSize size, int depth, int channels );
-
-/** @brief Initializes an image header that was previously allocated.
-
-The returned IplImage\* points to the initialized header.
-@param image Image header to initialize
-@param size Image width and height
-@param depth Image depth (see cvCreateImage )
-@param channels Number of channels (see cvCreateImage )
-@param origin Top-left IPL_ORIGIN_TL or bottom-left IPL_ORIGIN_BL
-@param align Alignment for image rows, typically 4 or 8 bytes
- */
-CVAPI(IplImage*) cvInitImageHeader( IplImage* image, CvSize size, int depth,
-                                   int channels, int origin CV_DEFAULT(0),
-                                   int align CV_DEFAULT(4));
-
-/** @brief Creates an image header and allocates the image data.
-
-This function call is equivalent to the following code:
-@code
-    header = cvCreateImageHeader(size, depth, channels);
-    cvCreateData(header);
-@endcode
-@param size Image width and height
-@param depth Bit depth of image elements. See IplImage for valid depths.
-@param channels Number of channels per pixel. See IplImage for details. This function only creates
-images with interleaved channels.
- */
-CVAPI(IplImage*)  cvCreateImage( CvSize size, int depth, int channels );
-
-/** @brief Deallocates an image header.
-
-This call is an analogue of :
-@code
-    if(image )
-    {
-        iplDeallocate(*image, IPL_IMAGE_HEADER | IPL_IMAGE_ROI);
-        *image = 0;
-    }
-@endcode
-but it does not use IPL functions by default (see the CV_TURN_ON_IPL_COMPATIBILITY macro).
-@param image Double pointer to the image header
- */
-CVAPI(void)  cvReleaseImageHeader( IplImage** image );
-
-/** @brief Deallocates the image header and the image data.
-
-This call is a shortened form of :
-@code
-    if(*image )
-    {
-        cvReleaseData(*image);
-        cvReleaseImageHeader(image);
-    }
-@endcode
-@param image Double pointer to the image header
-*/
-CVAPI(void)  cvReleaseImage( IplImage** image );
-
-/** Creates a copy of IPL image (widthStep may differ) */
-CVAPI(IplImage*) cvCloneImage( const IplImage* image );
-
-/** @brief Sets the channel of interest in an IplImage.
-
-If the ROI is set to NULL and the coi is *not* 0, the ROI is allocated. Most OpenCV functions do
-*not* support the COI setting, so to process an individual image/matrix channel one may copy (via
-cvCopy or cvSplit) the channel to a separate image/matrix, process it and then copy the result
-back (via cvCopy or cvMerge) if needed.
-@param image A pointer to the image header
-@param coi The channel of interest. 0 - all channels are selected, 1 - first channel is selected,
-etc. Note that the channel indices become 1-based.
- */
-CVAPI(void)  cvSetImageCOI( IplImage* image, int coi );
-
-/** @brief Returns the index of the channel of interest.
-
-Returns the channel of interest of in an IplImage. Returned values correspond to the coi in
-cvSetImageCOI.
-@param image A pointer to the image header
- */
-CVAPI(int)  cvGetImageCOI( const IplImage* image );
-
-/** @brief Sets an image Region Of Interest (ROI) for a given rectangle.
-
-If the original image ROI was NULL and the rect is not the whole image, the ROI structure is
-allocated.
-
-Most OpenCV functions support the use of ROI and treat the image rectangle as a separate image. For
-example, all of the pixel coordinates are counted from the top-left (or bottom-left) corner of the
-ROI, not the original image.
-@param image A pointer to the image header
-@param rect The ROI rectangle
- */
-CVAPI(void)  cvSetImageROI( IplImage* image, CvRect rect );
-
-/** @brief Resets the image ROI to include the entire image and releases the ROI structure.
-
-This produces a similar result to the following, but in addition it releases the ROI structure. :
-@code
-    cvSetImageROI(image, cvRect(0, 0, image->width, image->height ));
-    cvSetImageCOI(image, 0);
-@endcode
-@param image A pointer to the image header
- */
-CVAPI(void)  cvResetImageROI( IplImage* image );
-
-/** @brief Returns the image ROI.
-
-If there is no ROI set, cvRect(0,0,image-\>width,image-\>height) is returned.
-@param image A pointer to the image header
- */
-CVAPI(CvRect) cvGetImageROI( const IplImage* image );
-
-/** @brief Creates a matrix header but does not allocate the matrix data.
-
-The function allocates a new matrix header and returns a pointer to it. The matrix data can then be
-allocated using cvCreateData or set explicitly to user-allocated data via cvSetData.
-@param rows Number of rows in the matrix
-@param cols Number of columns in the matrix
-@param type Type of the matrix elements, see cvCreateMat
- */
-CVAPI(CvMat*)  cvCreateMatHeader( int rows, int cols, int type );
-
-#define CV_AUTOSTEP  0x7fffffff
-
-/** @brief Initializes a pre-allocated matrix header.
-
-This function is often used to process raw data with OpenCV matrix functions. For example, the
-following code computes the matrix product of two matrices, stored as ordinary arrays:
-@code
-    double a[] = { 1, 2, 3, 4,
-                   5, 6, 7, 8,
-                   9, 10, 11, 12 };
-
-    double b[] = { 1, 5, 9,
-                   2, 6, 10,
-                   3, 7, 11,
-                   4, 8, 12 };
-
-    double c[9];
-    CvMat Ma, Mb, Mc ;
-
-    cvInitMatHeader(&Ma, 3, 4, CV_64FC1, a);
-    cvInitMatHeader(&Mb, 4, 3, CV_64FC1, b);
-    cvInitMatHeader(&Mc, 3, 3, CV_64FC1, c);
-
-    cvMatMulAdd(&Ma, &Mb, 0, &Mc);
-    // the c array now contains the product of a (3x4) and b (4x3)
-@endcode
-@param mat A pointer to the matrix header to be initialized
-@param rows Number of rows in the matrix
-@param cols Number of columns in the matrix
-@param type Type of the matrix elements, see cvCreateMat .
-@param data Optional: data pointer assigned to the matrix header
-@param step Optional: full row width in bytes of the assigned data. By default, the minimal
-possible step is used which assumes there are no gaps between subsequent rows of the matrix.
- */
-CVAPI(CvMat*) cvInitMatHeader( CvMat* mat, int rows, int cols,
-                              int type, void* data CV_DEFAULT(NULL),
-                              int step CV_DEFAULT(CV_AUTOSTEP) );
-
-/** @brief Creates a matrix header and allocates the matrix data.
-
-The function call is equivalent to the following code:
-@code
-    CvMat* mat = cvCreateMatHeader(rows, cols, type);
-    cvCreateData(mat);
-@endcode
-@param rows Number of rows in the matrix
-@param cols Number of columns in the matrix
-@param type The type of the matrix elements in the form
-CV_\<bit depth\>\<S|U|F\>C\<number of channels\> , where S=signed, U=unsigned, F=float. For
-example, CV _ 8UC1 means the elements are 8-bit unsigned and the there is 1 channel, and CV _
-32SC2 means the elements are 32-bit signed and there are 2 channels.
- */
-CVAPI(CvMat*)  cvCreateMat( int rows, int cols, int type );
-
-/** @brief Deallocates a matrix.
-
-The function decrements the matrix data reference counter and deallocates matrix header. If the data
-reference counter is 0, it also deallocates the data. :
-@code
-    if(*mat )
-        cvDecRefData(*mat);
-    cvFree((void**)mat);
-@endcode
-@param mat Double pointer to the matrix
- */
-CVAPI(void)  cvReleaseMat( CvMat** mat );
-
-/** @brief Decrements an array data reference counter.
-
-The function decrements the data reference counter in a CvMat or CvMatND if the reference counter
-
-pointer is not NULL. If the counter reaches zero, the data is deallocated. In the current
-implementation the reference counter is not NULL only if the data was allocated using the
-cvCreateData function. The counter will be NULL in other cases such as: external data was assigned
-to the header using cvSetData, header is part of a larger matrix or image, or the header was
-converted from an image or n-dimensional matrix header.
-@param arr Pointer to an array header
- */
-CV_INLINE  void  cvDecRefData( CvArr* arr )
-{
-    if( CV_IS_MAT( arr ))
-    {
-        CvMat* mat = (CvMat*)arr;
-        mat->data.ptr = NULL;
-        if( mat->refcount != NULL && --*mat->refcount == 0 )
-            cvFree( &mat->refcount );
-        mat->refcount = NULL;
-    }
-    else if( CV_IS_MATND( arr ))
-    {
-        CvMatND* mat = (CvMatND*)arr;
-        mat->data.ptr = NULL;
-        if( mat->refcount != NULL && --*mat->refcount == 0 )
-            cvFree( &mat->refcount );
-        mat->refcount = NULL;
-    }
-}
-
-/** @brief Increments array data reference counter.
-
-The function increments CvMat or CvMatND data reference counter and returns the new counter value if
-the reference counter pointer is not NULL, otherwise it returns zero.
-@param arr Array header
- */
-CV_INLINE  int  cvIncRefData( CvArr* arr )
-{
-    int refcount = 0;
-    if( CV_IS_MAT( arr ))
-    {
-        CvMat* mat = (CvMat*)arr;
-        if( mat->refcount != NULL )
-            refcount = ++*mat->refcount;
-    }
-    else if( CV_IS_MATND( arr ))
-    {
-        CvMatND* mat = (CvMatND*)arr;
-        if( mat->refcount != NULL )
-            refcount = ++*mat->refcount;
-    }
-    return refcount;
-}
-
-
-/** Creates an exact copy of the input matrix (except, may be, step value) */
-CVAPI(CvMat*) cvCloneMat( const CvMat* mat );
-
-
-/** @brief Returns matrix header corresponding to the rectangular sub-array of input image or matrix.
-
-The function returns header, corresponding to a specified rectangle of the input array. In other
-
-words, it allows the user to treat a rectangular part of input array as a stand-alone array. ROI is
-taken into account by the function so the sub-array of ROI is actually extracted.
-@param arr Input array
-@param submat Pointer to the resultant sub-array header
-@param rect Zero-based coordinates of the rectangle of interest
- */
-CVAPI(CvMat*) cvGetSubRect( const CvArr* arr, CvMat* submat, CvRect rect );
-#define cvGetSubArr cvGetSubRect
-
-/** @brief Returns array row or row span.
-
-The function returns the header, corresponding to a specified row/row span of the input array.
-cvGetRow(arr, submat, row) is a shortcut for cvGetRows(arr, submat, row, row+1).
-@param arr Input array
-@param submat Pointer to the resulting sub-array header
-@param start_row Zero-based index of the starting row (inclusive) of the span
-@param end_row Zero-based index of the ending row (exclusive) of the span
-@param delta_row Index step in the row span. That is, the function extracts every delta_row -th
-row from start_row and up to (but not including) end_row .
- */
-CVAPI(CvMat*) cvGetRows( const CvArr* arr, CvMat* submat,
-                        int start_row, int end_row,
-                        int delta_row CV_DEFAULT(1));
-
-/** @overload
-@param arr Input array
-@param submat Pointer to the resulting sub-array header
-@param row Zero-based index of the selected row
-*/
-CV_INLINE  CvMat*  cvGetRow( const CvArr* arr, CvMat* submat, int row )
-{
-    return cvGetRows( arr, submat, row, row + 1, 1 );
-}
-
-
-/** @brief Returns one of more array columns.
-
-The function returns the header, corresponding to a specified column span of the input array. That
-
-is, no data is copied. Therefore, any modifications of the submatrix will affect the original array.
-If you need to copy the columns, use cvCloneMat. cvGetCol(arr, submat, col) is a shortcut for
-cvGetCols(arr, submat, col, col+1).
-@param arr Input array
-@param submat Pointer to the resulting sub-array header
-@param start_col Zero-based index of the starting column (inclusive) of the span
-@param end_col Zero-based index of the ending column (exclusive) of the span
- */
-CVAPI(CvMat*) cvGetCols( const CvArr* arr, CvMat* submat,
-                        int start_col, int end_col );
-
-/** @overload
-@param arr Input array
-@param submat Pointer to the resulting sub-array header
-@param col Zero-based index of the selected column
-*/
-CV_INLINE  CvMat*  cvGetCol( const CvArr* arr, CvMat* submat, int col )
-{
-    return cvGetCols( arr, submat, col, col + 1 );
-}
-
-/** @brief Returns one of array diagonals.
-
-The function returns the header, corresponding to a specified diagonal of the input array.
-@param arr Input array
-@param submat Pointer to the resulting sub-array header
-@param diag Index of the array diagonal. Zero value corresponds to the main diagonal, -1
-corresponds to the diagonal above the main, 1 corresponds to the diagonal below the main, and so
-forth.
- */
-CVAPI(CvMat*) cvGetDiag( const CvArr* arr, CvMat* submat,
-                            int diag CV_DEFAULT(0));
-
-/** low-level scalar <-> raw data conversion functions */
-CVAPI(void) cvScalarToRawData( const CvScalar* scalar, void* data, int type,
-                              int extend_to_12 CV_DEFAULT(0) );
-
-CVAPI(void) cvRawDataToScalar( const void* data, int type, CvScalar* scalar );
-
-/** @brief Creates a new matrix header but does not allocate the matrix data.
-
-The function allocates a header for a multi-dimensional dense array. The array data can further be
-allocated using cvCreateData or set explicitly to user-allocated data via cvSetData.
-@param dims Number of array dimensions
-@param sizes Array of dimension sizes
-@param type Type of array elements, see cvCreateMat
- */
-CVAPI(CvMatND*)  cvCreateMatNDHeader( int dims, const int* sizes, int type );
-
-/** @brief Creates the header and allocates the data for a multi-dimensional dense array.
-
-This function call is equivalent to the following code:
-@code
-    CvMatND* mat = cvCreateMatNDHeader(dims, sizes, type);
-    cvCreateData(mat);
-@endcode
-@param dims Number of array dimensions. This must not exceed CV_MAX_DIM (32 by default, but can be
-changed at build time).
-@param sizes Array of dimension sizes.
-@param type Type of array elements, see cvCreateMat .
- */
-CVAPI(CvMatND*)  cvCreateMatND( int dims, const int* sizes, int type );
-
-/** @brief Initializes a pre-allocated multi-dimensional array header.
-
-@param mat A pointer to the array header to be initialized
-@param dims The number of array dimensions
-@param sizes An array of dimension sizes
-@param type Type of array elements, see cvCreateMat
-@param data Optional data pointer assigned to the matrix header
- */
-CVAPI(CvMatND*)  cvInitMatNDHeader( CvMatND* mat, int dims, const int* sizes,
-                                    int type, void* data CV_DEFAULT(NULL) );
-
-/** @brief Deallocates a multi-dimensional array.
-
-The function decrements the array data reference counter and releases the array header. If the
-reference counter reaches 0, it also deallocates the data. :
-@code
-    if(*mat )
-        cvDecRefData(*mat);
-    cvFree((void**)mat);
-@endcode
-@param mat Double pointer to the array
- */
-CV_INLINE  void  cvReleaseMatND( CvMatND** mat )
-{
-    cvReleaseMat( (CvMat**)mat );
-}
-
-/** Creates a copy of CvMatND (except, may be, steps) */
-CVAPI(CvMatND*) cvCloneMatND( const CvMatND* mat );
-
-/** @brief Creates sparse array.
-
-The function allocates a multi-dimensional sparse array. Initially the array contain no elements,
-that is PtrND and other related functions will return 0 for every index.
-@param dims Number of array dimensions. In contrast to the dense matrix, the number of dimensions is
-practically unlimited (up to \f$2^{16}\f$ ).
-@param sizes Array of dimension sizes
-@param type Type of array elements. The same as for CvMat
- */
-CVAPI(CvSparseMat*)  cvCreateSparseMat( int dims, const int* sizes, int type );
-
-/** @brief Deallocates sparse array.
-
-The function releases the sparse array and clears the array pointer upon exit.
-@param mat Double pointer to the array
- */
-CVAPI(void)  cvReleaseSparseMat( CvSparseMat** mat );
-
-/** Creates a copy of CvSparseMat (except, may be, zero items) */
-CVAPI(CvSparseMat*) cvCloneSparseMat( const CvSparseMat* mat );
-
-/** @brief Initializes sparse array elements iterator.
-
-The function initializes iterator of sparse array elements and returns pointer to the first element,
-or NULL if the array is empty.
-@param mat Input array
-@param mat_iterator Initialized iterator
- */
-CVAPI(CvSparseNode*) cvInitSparseMatIterator( const CvSparseMat* mat,
-                                              CvSparseMatIterator* mat_iterator );
-
-/** @brief Returns the next sparse matrix element
-
-The function moves iterator to the next sparse matrix element and returns pointer to it. In the
-current version there is no any particular order of the elements, because they are stored in the
-hash table. The sample below demonstrates how to iterate through the sparse matrix:
-@code
-    // print all the non-zero sparse matrix elements and compute their sum
-    double sum = 0;
-    int i, dims = cvGetDims(sparsemat);
-    CvSparseMatIterator it;
-    CvSparseNode* node = cvInitSparseMatIterator(sparsemat, &it);
-
-    for(; node != 0; node = cvGetNextSparseNode(&it))
-    {
-        int* idx = CV_NODE_IDX(array, node);
-        float val = *(float*)CV_NODE_VAL(array, node);
-        printf("M");
-        for(i = 0; i < dims; i++ )
-            printf("[%d]", idx[i]);
-        printf("=%g\n", val);
-
-        sum += val;
-    }
-
-    printf("nTotal sum = %g\n", sum);
-@endcode
-@param mat_iterator Sparse array iterator
- */
-CV_INLINE CvSparseNode* cvGetNextSparseNode( CvSparseMatIterator* mat_iterator )
-{
-    if( mat_iterator->node->next )
-        return mat_iterator->node = mat_iterator->node->next;
-    else
-    {
-        int idx;
-        for( idx = ++mat_iterator->curidx; idx < mat_iterator->mat->hashsize; idx++ )
-        {
-            CvSparseNode* node = (CvSparseNode*)mat_iterator->mat->hashtable[idx];
-            if( node )
-            {
-                mat_iterator->curidx = idx;
-                return mat_iterator->node = node;
-            }
-        }
-        return NULL;
-    }
-}
-
-
-#define CV_MAX_ARR 10
-
-/** matrix iterator: used for n-ary operations on dense arrays */
-typedef struct CvNArrayIterator
-{
-    int count; /**< number of arrays */
-    int dims; /**< number of dimensions to iterate */
-    CvSize size; /**< maximal common linear size: { width = size, height = 1 } */
-    uchar* ptr[CV_MAX_ARR]; /**< pointers to the array slices */
-    int stack[CV_MAX_DIM]; /**< for internal use */
-    CvMatND* hdr[CV_MAX_ARR]; /**< pointers to the headers of the
-                                 matrices that are processed */
-}
-CvNArrayIterator;
-
-#define CV_NO_DEPTH_CHECK     1
-#define CV_NO_CN_CHECK        2
-#define CV_NO_SIZE_CHECK      4
-
-/** initializes iterator that traverses through several arrays simultaneously
-   (the function together with cvNextArraySlice is used for
-    N-ari element-wise operations) */
-CVAPI(int) cvInitNArrayIterator( int count, CvArr** arrs,
-                                 const CvArr* mask, CvMatND* stubs,
-                                 CvNArrayIterator* array_iterator,
-                                 int flags CV_DEFAULT(0) );
-
-/** returns zero value if iteration is finished, non-zero (slice length) otherwise */
-CVAPI(int) cvNextNArraySlice( CvNArrayIterator* array_iterator );
-
-
-/** @brief Returns type of array elements.
-
-The function returns type of the array elements. In the case of IplImage the type is converted to
-CvMat-like representation. For example, if the image has been created as:
-@code
-    IplImage* img = cvCreateImage(cvSize(640, 480), IPL_DEPTH_8U, 3);
-@endcode
-The code cvGetElemType(img) will return CV_8UC3.
-@param arr Input array
- */
-CVAPI(int) cvGetElemType( const CvArr* arr );
-
-/** @brief Return number of array dimensions
-
-The function returns the array dimensionality and the array of dimension sizes. In the case of
-IplImage or CvMat it always returns 2 regardless of number of image/matrix rows. For example, the
-following code calculates total number of array elements:
-@code
-    int sizes[CV_MAX_DIM];
-    int i, total = 1;
-    int dims = cvGetDims(arr, size);
-    for(i = 0; i < dims; i++ )
-        total *= sizes[i];
-@endcode
-@param arr Input array
-@param sizes Optional output vector of the array dimension sizes. For 2d arrays the number of rows
-(height) goes first, number of columns (width) next.
- */
-CVAPI(int) cvGetDims( const CvArr* arr, int* sizes CV_DEFAULT(NULL) );
-
-
-/** @brief Returns array size along the specified dimension.
-
-@param arr Input array
-@param index Zero-based dimension index (for matrices 0 means number of rows, 1 means number of
-columns; for images 0 means height, 1 means width)
- */
-CVAPI(int) cvGetDimSize( const CvArr* arr, int index );
-
-
-/** @brief Return pointer to a particular array element.
-
-The functions return a pointer to a specific array element. Number of array dimension should match
-to the number of indices passed to the function except for cvPtr1D function that can be used for
-sequential access to 1D, 2D or nD dense arrays.
-
-The functions can be used for sparse arrays as well - if the requested node does not exist they
-create it and set it to zero.
-
-All these as well as other functions accessing array elements ( cvGetND , cvGetRealND , cvSet
-, cvSetND , cvSetRealND ) raise an error in case if the element index is out of range.
-@param arr Input array
-@param idx0 The first zero-based component of the element index
-@param type Optional output parameter: type of matrix elements
- */
-CVAPI(uchar*) cvPtr1D( const CvArr* arr, int idx0, int* type CV_DEFAULT(NULL));
-/** @overload */
-CVAPI(uchar*) cvPtr2D( const CvArr* arr, int idx0, int idx1, int* type CV_DEFAULT(NULL) );
-/** @overload */
-CVAPI(uchar*) cvPtr3D( const CvArr* arr, int idx0, int idx1, int idx2,
-                      int* type CV_DEFAULT(NULL));
-/** @overload
-@param arr Input array
-@param idx Array of the element indices
-@param type Optional output parameter: type of matrix elements
-@param create_node Optional input parameter for sparse matrices. Non-zero value of the parameter
-means that the requested element is created if it does not exist already.
-@param precalc_hashval Optional input parameter for sparse matrices. If the pointer is not NULL,
-the function does not recalculate the node hash value, but takes it from the specified location.
-It is useful for speeding up pair-wise operations (TODO: provide an example)
-*/
-CVAPI(uchar*) cvPtrND( const CvArr* arr, const int* idx, int* type CV_DEFAULT(NULL),
-                      int create_node CV_DEFAULT(1),
-                      unsigned* precalc_hashval CV_DEFAULT(NULL));
-
-/** @brief Return a specific array element.
-
-The functions return a specific array element. In the case of a sparse array the functions return 0
-if the requested node does not exist (no new node is created by the functions).
-@param arr Input array
-@param idx0 The first zero-based component of the element index
- */
-CVAPI(CvScalar) cvGet1D( const CvArr* arr, int idx0 );
-/** @overload */
-CVAPI(CvScalar) cvGet2D( const CvArr* arr, int idx0, int idx1 );
-/** @overload */
-CVAPI(CvScalar) cvGet3D( const CvArr* arr, int idx0, int idx1, int idx2 );
-/** @overload
-@param arr Input array
-@param idx Array of the element indices
-*/
-CVAPI(CvScalar) cvGetND( const CvArr* arr, const int* idx );
-
-/** @brief Return a specific element of single-channel 1D, 2D, 3D or nD array.
-
-Returns a specific element of a single-channel array. If the array has multiple channels, a runtime
-error is raised. Note that Get?D functions can be used safely for both single-channel and
-multiple-channel arrays though they are a bit slower.
-
-In the case of a sparse array the functions return 0 if the requested node does not exist (no new
-node is created by the functions).
-@param arr Input array. Must have a single channel.
-@param idx0 The first zero-based component of the element index
- */
-CVAPI(double) cvGetReal1D( const CvArr* arr, int idx0 );
-/** @overload */
-CVAPI(double) cvGetReal2D( const CvArr* arr, int idx0, int idx1 );
-/** @overload */
-CVAPI(double) cvGetReal3D( const CvArr* arr, int idx0, int idx1, int idx2 );
-/** @overload
-@param arr Input array. Must have a single channel.
-@param idx Array of the element indices
-*/
-CVAPI(double) cvGetRealND( const CvArr* arr, const int* idx );
-
-/** @brief Change the particular array element.
-
-The functions assign the new value to a particular array element. In the case of a sparse array the
-functions create the node if it does not exist yet.
-@param arr Input array
-@param idx0 The first zero-based component of the element index
-@param value The assigned value
- */
-CVAPI(void) cvSet1D( CvArr* arr, int idx0, CvScalar value );
-/** @overload */
-CVAPI(void) cvSet2D( CvArr* arr, int idx0, int idx1, CvScalar value );
-/** @overload */
-CVAPI(void) cvSet3D( CvArr* arr, int idx0, int idx1, int idx2, CvScalar value );
-/** @overload
-@param arr Input array
-@param idx Array of the element indices
-@param value The assigned value
-*/
-CVAPI(void) cvSetND( CvArr* arr, const int* idx, CvScalar value );
-
-/** @brief Change a specific array element.
-
-The functions assign a new value to a specific element of a single-channel array. If the array has
-multiple channels, a runtime error is raised. Note that the Set\*D function can be used safely for
-both single-channel and multiple-channel arrays, though they are a bit slower.
-
-In the case of a sparse array the functions create the node if it does not yet exist.
-@param arr Input array
-@param idx0 The first zero-based component of the element index
-@param value The assigned value
- */
-CVAPI(void) cvSetReal1D( CvArr* arr, int idx0, double value );
-/** @overload */
-CVAPI(void) cvSetReal2D( CvArr* arr, int idx0, int idx1, double value );
-/** @overload */
-CVAPI(void) cvSetReal3D( CvArr* arr, int idx0,
-                        int idx1, int idx2, double value );
-/** @overload
-@param arr Input array
-@param idx Array of the element indices
-@param value The assigned value
-*/
-CVAPI(void) cvSetRealND( CvArr* arr, const int* idx, double value );
-
-/** clears element of ND dense array,
-   in case of sparse arrays it deletes the specified node */
-CVAPI(void) cvClearND( CvArr* arr, const int* idx );
-
-/** @brief Returns matrix header for arbitrary array.
-
-The function returns a matrix header for the input array that can be a matrix - CvMat, an image -
-IplImage, or a multi-dimensional dense array - CvMatND (the third option is allowed only if
-allowND != 0) . In the case of matrix the function simply returns the input pointer. In the case of
-IplImage\* or CvMatND it initializes the header structure with parameters of the current image ROI
-and returns &header. Because COI is not supported by CvMat, it is returned separately.
-
-The function provides an easy way to handle both types of arrays - IplImage and CvMat using the same
-code. Input array must have non-zero data pointer, otherwise the function will report an error.
-
-@note If the input array is IplImage with planar data layout and COI set, the function returns the
-pointer to the selected plane and COI == 0. This feature allows user to process IplImage structures
-with planar data layout, even though OpenCV does not support such images.
-@param arr Input array
-@param header Pointer to CvMat structure used as a temporary buffer
-@param coi Optional output parameter for storing COI
-@param allowND If non-zero, the function accepts multi-dimensional dense arrays (CvMatND\*) and
-returns 2D matrix (if CvMatND has two dimensions) or 1D matrix (when CvMatND has 1 dimension or
-more than 2 dimensions). The CvMatND array must be continuous.
-@sa cvGetImage, cvarrToMat.
- */
-CVAPI(CvMat*) cvGetMat( const CvArr* arr, CvMat* header,
-                       int* coi CV_DEFAULT(NULL),
-                       int allowND CV_DEFAULT(0));
-
-/** @brief Returns image header for arbitrary array.
-
-The function returns the image header for the input array that can be a matrix (CvMat) or image
-(IplImage). In the case of an image the function simply returns the input pointer. In the case of
-CvMat it initializes an image_header structure with the parameters of the input matrix. Note that
-if we transform IplImage to CvMat using cvGetMat and then transform CvMat back to IplImage using
-this function, we will get different headers if the ROI is set in the original image.
-@param arr Input array
-@param image_header Pointer to IplImage structure used as a temporary buffer
- */
-CVAPI(IplImage*) cvGetImage( const CvArr* arr, IplImage* image_header );
-
-
-/** @brief Changes the shape of a multi-dimensional array without copying the data.
-
-The function is an advanced version of cvReshape that can work with multi-dimensional arrays as
-well (though it can work with ordinary images and matrices) and change the number of dimensions.
-
-Below are the two samples from the cvReshape description rewritten using cvReshapeMatND:
-@code
-    IplImage* color_img = cvCreateImage(cvSize(320,240), IPL_DEPTH_8U, 3);
-    IplImage gray_img_hdr, *gray_img;
-    gray_img = (IplImage*)cvReshapeMatND(color_img, sizeof(gray_img_hdr), &gray_img_hdr, 1, 0, 0);
-    ...
-    int size[] = { 2, 2, 2 };
-    CvMatND* mat = cvCreateMatND(3, size, CV_32F);
-    CvMat row_header, *row;
-    row = (CvMat*)cvReshapeMatND(mat, sizeof(row_header), &row_header, 0, 1, 0);
-@endcode
-In C, the header file for this function includes a convenient macro cvReshapeND that does away with
-the sizeof_header parameter. So, the lines containing the call to cvReshapeMatND in the examples
-may be replaced as follow:
-@code
-    gray_img = (IplImage*)cvReshapeND(color_img, &gray_img_hdr, 1, 0, 0);
-    ...
-    row = (CvMat*)cvReshapeND(mat, &row_header, 0, 1, 0);
-@endcode
-@param arr Input array
-@param sizeof_header Size of output header to distinguish between IplImage, CvMat and CvMatND
-output headers
-@param header Output header to be filled
-@param new_cn New number of channels. new_cn = 0 means that the number of channels remains
-unchanged.
-@param new_dims New number of dimensions. new_dims = 0 means that the number of dimensions
-remains the same.
-@param new_sizes Array of new dimension sizes. Only new_dims-1 values are used, because the
-total number of elements must remain the same. Thus, if new_dims = 1, new_sizes array is not
-used.
- */
-CVAPI(CvArr*) cvReshapeMatND( const CvArr* arr,
-                             int sizeof_header, CvArr* header,
-                             int new_cn, int new_dims, int* new_sizes );
-
-#define cvReshapeND( arr, header, new_cn, new_dims, new_sizes )   \
-      cvReshapeMatND( (arr), sizeof(*(header)), (header),         \
-                      (new_cn), (new_dims), (new_sizes))
-
-/** @brief Changes shape of matrix/image without copying data.
-
-The function initializes the CvMat header so that it points to the same data as the original array
-but has a different shape - different number of channels, different number of rows, or both.
-
-The following example code creates one image buffer and two image headers, the first is for a
-320x240x3 image and the second is for a 960x240x1 image:
-@code
-    IplImage* color_img = cvCreateImage(cvSize(320,240), IPL_DEPTH_8U, 3);
-    CvMat gray_mat_hdr;
-    IplImage gray_img_hdr, *gray_img;
-    cvReshape(color_img, &gray_mat_hdr, 1);
-    gray_img = cvGetImage(&gray_mat_hdr, &gray_img_hdr);
-@endcode
-And the next example converts a 3x3 matrix to a single 1x9 vector:
-@code
-    CvMat* mat = cvCreateMat(3, 3, CV_32F);
-    CvMat row_header, *row;
-    row = cvReshape(mat, &row_header, 0, 1);
-@endcode
-@param arr Input array
-@param header Output header to be filled
-@param new_cn New number of channels. 'new_cn = 0' means that the number of channels remains
-unchanged.
-@param new_rows New number of rows. 'new_rows = 0' means that the number of rows remains
-unchanged unless it needs to be changed according to new_cn value.
-*/
-CVAPI(CvMat*) cvReshape( const CvArr* arr, CvMat* header,
-                        int new_cn, int new_rows CV_DEFAULT(0) );
-
-/** Repeats source 2d array several times in both horizontal and
-   vertical direction to fill destination array */
-CVAPI(void) cvRepeat( const CvArr* src, CvArr* dst );
-
-/** @brief Allocates array data
-
-The function allocates image, matrix or multi-dimensional dense array data. Note that in the case of
-matrix types OpenCV allocation functions are used. In the case of IplImage they are used unless
-CV_TURN_ON_IPL_COMPATIBILITY() has been called before. In the latter case IPL functions are used
-to allocate the data.
-@param arr Array header
- */
-CVAPI(void)  cvCreateData( CvArr* arr );
-
-/** @brief Releases array data.
-
-The function releases the array data. In the case of CvMat or CvMatND it simply calls
-cvDecRefData(), that is the function can not deallocate external data. See also the note to
-cvCreateData .
-@param arr Array header
- */
-CVAPI(void)  cvReleaseData( CvArr* arr );
-
-/** @brief Assigns user data to the array header.
-
-The function assigns user data to the array header. Header should be initialized before using
-cvCreateMatHeader, cvCreateImageHeader, cvCreateMatNDHeader, cvInitMatHeader,
-cvInitImageHeader or cvInitMatNDHeader.
-@param arr Array header
-@param data User data
-@param step Full row length in bytes
- */
-CVAPI(void)  cvSetData( CvArr* arr, void* data, int step );
-
-/** @brief Retrieves low-level information about the array.
-
-The function fills output variables with low-level information about the array data. All output
-
-parameters are optional, so some of the pointers may be set to NULL. If the array is IplImage with
-ROI set, the parameters of ROI are returned.
-
-The following example shows how to get access to array elements. It computes absolute values of the
-array elements :
-@code
-    float* data;
-    int step;
-    CvSize size;
-
-    cvGetRawData(array, (uchar**)&data, &step, &size);
-    step /= sizeof(data[0]);
-
-    for(int y = 0; y < size.height; y++, data += step )
-        for(int x = 0; x < size.width; x++ )
-            data[x] = (float)fabs(data[x]);
-@endcode
-@param arr Array header
-@param data Output pointer to the whole image origin or ROI origin if ROI is set
-@param step Output full row length in bytes
-@param roi_size Output ROI size
- */
-CVAPI(void) cvGetRawData( const CvArr* arr, uchar** data,
-                         int* step CV_DEFAULT(NULL),
-                         CvSize* roi_size CV_DEFAULT(NULL));
-
-/** @brief Returns size of matrix or image ROI.
-
-The function returns number of rows (CvSize::height) and number of columns (CvSize::width) of the
-input matrix or image. In the case of image the size of ROI is returned.
-@param arr array header
- */
-CVAPI(CvSize) cvGetSize( const CvArr* arr );
-
-/** @brief Copies one array to another.
-
-The function copies selected elements from an input array to an output array:
-
-\f[\texttt{dst} (I)= \texttt{src} (I)  \quad \text{if} \quad \texttt{mask} (I)  \ne 0.\f]
-
-If any of the passed arrays is of IplImage type, then its ROI and COI fields are used. Both arrays
-must have the same type, the same number of dimensions, and the same size. The function can also
-copy sparse arrays (mask is not supported in this case).
-@param src The source array
-@param dst The destination array
-@param mask Operation mask, 8-bit single channel array; specifies elements of the destination array
-to be changed
- */
-CVAPI(void)  cvCopy( const CvArr* src, CvArr* dst,
-                     const CvArr* mask CV_DEFAULT(NULL) );
-
-/** @brief Sets every element of an array to a given value.
-
-The function copies the scalar value to every selected element of the destination array:
-\f[\texttt{arr} (I)= \texttt{value} \quad \text{if} \quad \texttt{mask} (I)  \ne 0\f]
-If array arr is of IplImage type, then is ROI used, but COI must not be set.
-@param arr The destination array
-@param value Fill value
-@param mask Operation mask, 8-bit single channel array; specifies elements of the destination
-array to be changed
- */
-CVAPI(void)  cvSet( CvArr* arr, CvScalar value,
-                    const CvArr* mask CV_DEFAULT(NULL) );
-
-/** @brief Clears the array.
-
-The function clears the array. In the case of dense arrays (CvMat, CvMatND or IplImage),
-cvZero(array) is equivalent to cvSet(array,cvScalarAll(0),0). In the case of sparse arrays all the
-elements are removed.
-@param arr Array to be cleared
- */
-CVAPI(void)  cvSetZero( CvArr* arr );
-#define cvZero  cvSetZero
-
-
-/** Splits a multi-channel array into the set of single-channel arrays or
-   extracts particular [color] plane */
-CVAPI(void)  cvSplit( const CvArr* src, CvArr* dst0, CvArr* dst1,
-                      CvArr* dst2, CvArr* dst3 );
-
-/** Merges a set of single-channel arrays into the single multi-channel array
-   or inserts one particular [color] plane to the array */
-CVAPI(void)  cvMerge( const CvArr* src0, const CvArr* src1,
-                      const CvArr* src2, const CvArr* src3,
-                      CvArr* dst );
-
-/** Copies several channels from input arrays to
-   certain channels of output arrays */
-CVAPI(void)  cvMixChannels( const CvArr** src, int src_count,
-                            CvArr** dst, int dst_count,
-                            const int* from_to, int pair_count );
-
-/** @brief Converts one array to another with optional linear transformation.
-
-The function has several different purposes, and thus has several different names. It copies one
-array to another with optional scaling, which is performed first, and/or optional type conversion,
-performed after:
-
-\f[\texttt{dst} (I) =  \texttt{scale} \texttt{src} (I) + ( \texttt{shift} _0, \texttt{shift} _1,...)\f]
-
-All the channels of multi-channel arrays are processed independently.
-
-The type of conversion is done with rounding and saturation, that is if the result of scaling +
-conversion can not be represented exactly by a value of the destination array element type, it is
-set to the nearest representable value on the real axis.
-@param src Source array
-@param dst Destination array
-@param scale Scale factor
-@param shift Value added to the scaled source array elements
- */
-CVAPI(void)  cvConvertScale( const CvArr* src, CvArr* dst,
-                             double scale CV_DEFAULT(1),
-                             double shift CV_DEFAULT(0) );
-#define cvCvtScale cvConvertScale
-#define cvScale  cvConvertScale
-#define cvConvert( src, dst )  cvConvertScale( (src), (dst), 1, 0 )
-
-
-/** Performs linear transformation on every source array element,
-   stores absolute value of the result:
-   dst(x,y,c) = abs(scale*src(x,y,c)+shift).
-   destination array must have 8u type.
-   In other cases one may use cvConvertScale + cvAbsDiffS */
-CVAPI(void)  cvConvertScaleAbs( const CvArr* src, CvArr* dst,
-                                double scale CV_DEFAULT(1),
-                                double shift CV_DEFAULT(0) );
-#define cvCvtScaleAbs  cvConvertScaleAbs
-
-
-/** checks termination criteria validity and
-   sets eps to default_eps (if it is not set),
-   max_iter to default_max_iters (if it is not set)
-*/
-CVAPI(CvTermCriteria) cvCheckTermCriteria( CvTermCriteria criteria,
-                                           double default_eps,
-                                           int default_max_iters );
-
-/****************************************************************************************\
-*                   Arithmetic, logic and comparison operations                          *
-\****************************************************************************************/
-
-/** dst(mask) = src1(mask) + src2(mask) */
-CVAPI(void)  cvAdd( const CvArr* src1, const CvArr* src2, CvArr* dst,
-                    const CvArr* mask CV_DEFAULT(NULL));
-
-/** dst(mask) = src(mask) + value */
-CVAPI(void)  cvAddS( const CvArr* src, CvScalar value, CvArr* dst,
-                     const CvArr* mask CV_DEFAULT(NULL));
-
-/** dst(mask) = src1(mask) - src2(mask) */
-CVAPI(void)  cvSub( const CvArr* src1, const CvArr* src2, CvArr* dst,
-                    const CvArr* mask CV_DEFAULT(NULL));
-
-/** dst(mask) = src(mask) - value = src(mask) + (-value) */
-CV_INLINE  void  cvSubS( const CvArr* src, CvScalar value, CvArr* dst,
-                         const CvArr* mask CV_DEFAULT(NULL))
-{
-    cvAddS( src, cvScalar( -value.val[0], -value.val[1], -value.val[2], -value.val[3]),
-            dst, mask );
-}
-
-/** dst(mask) = value - src(mask) */
-CVAPI(void)  cvSubRS( const CvArr* src, CvScalar value, CvArr* dst,
-                      const CvArr* mask CV_DEFAULT(NULL));
-
-/** dst(idx) = src1(idx) * src2(idx) * scale
-   (scaled element-wise multiplication of 2 arrays) */
-CVAPI(void)  cvMul( const CvArr* src1, const CvArr* src2,
-                    CvArr* dst, double scale CV_DEFAULT(1) );
-
-/** element-wise division/inversion with scaling:
-    dst(idx) = src1(idx) * scale / src2(idx)
-    or dst(idx) = scale / src2(idx) if src1 == 0 */
-CVAPI(void)  cvDiv( const CvArr* src1, const CvArr* src2,
-                    CvArr* dst, double scale CV_DEFAULT(1));
-
-/** dst = src1 * scale + src2 */
-CVAPI(void)  cvScaleAdd( const CvArr* src1, CvScalar scale,
-                         const CvArr* src2, CvArr* dst );
-#define cvAXPY( A, real_scalar, B, C ) cvScaleAdd(A, cvRealScalar(real_scalar), B, C)
-
-/** dst = src1 * alpha + src2 * beta + gamma */
-CVAPI(void)  cvAddWeighted( const CvArr* src1, double alpha,
-                            const CvArr* src2, double beta,
-                            double gamma, CvArr* dst );
-
-/** @brief Calculates the dot product of two arrays in Euclidean metrics.
-
-The function calculates and returns the Euclidean dot product of two arrays.
-
-\f[src1  \bullet src2 =  \sum _I ( \texttt{src1} (I)  \texttt{src2} (I))\f]
-
-In the case of multiple channel arrays, the results for all channels are accumulated. In particular,
-cvDotProduct(a,a) where a is a complex vector, will return \f$||\texttt{a}||^2\f$. The function can
-process multi-dimensional arrays, row by row, layer by layer, and so on.
-@param src1 The first source array
-@param src2 The second source array
- */
-CVAPI(double)  cvDotProduct( const CvArr* src1, const CvArr* src2 );
-
-/** dst(idx) = src1(idx) & src2(idx) */
-CVAPI(void) cvAnd( const CvArr* src1, const CvArr* src2,
-                  CvArr* dst, const CvArr* mask CV_DEFAULT(NULL));
-
-/** dst(idx) = src(idx) & value */
-CVAPI(void) cvAndS( const CvArr* src, CvScalar value,
-                   CvArr* dst, const CvArr* mask CV_DEFAULT(NULL));
-
-/** dst(idx) = src1(idx) | src2(idx) */
-CVAPI(void) cvOr( const CvArr* src1, const CvArr* src2,
-                 CvArr* dst, const CvArr* mask CV_DEFAULT(NULL));
-
-/** dst(idx) = src(idx) | value */
-CVAPI(void) cvOrS( const CvArr* src, CvScalar value,
-                  CvArr* dst, const CvArr* mask CV_DEFAULT(NULL));
-
-/** dst(idx) = src1(idx) ^ src2(idx) */
-CVAPI(void) cvXor( const CvArr* src1, const CvArr* src2,
-                  CvArr* dst, const CvArr* mask CV_DEFAULT(NULL));
-
-/** dst(idx) = src(idx) ^ value */
-CVAPI(void) cvXorS( const CvArr* src, CvScalar value,
-                   CvArr* dst, const CvArr* mask CV_DEFAULT(NULL));
-
-/** dst(idx) = ~src(idx) */
-CVAPI(void) cvNot( const CvArr* src, CvArr* dst );
-
-/** dst(idx) = lower(idx) <= src(idx) < upper(idx) */
-CVAPI(void) cvInRange( const CvArr* src, const CvArr* lower,
-                      const CvArr* upper, CvArr* dst );
-
-/** dst(idx) = lower <= src(idx) < upper */
-CVAPI(void) cvInRangeS( const CvArr* src, CvScalar lower,
-                       CvScalar upper, CvArr* dst );
-
-#define CV_CMP_EQ   0
-#define CV_CMP_GT   1
-#define CV_CMP_GE   2
-#define CV_CMP_LT   3
-#define CV_CMP_LE   4
-#define CV_CMP_NE   5
-
-/** The comparison operation support single-channel arrays only.
-   Destination image should be 8uC1 or 8sC1 */
-
-/** dst(idx) = src1(idx) _cmp_op_ src2(idx) */
-CVAPI(void) cvCmp( const CvArr* src1, const CvArr* src2, CvArr* dst, int cmp_op );
-
-/** dst(idx) = src1(idx) _cmp_op_ value */
-CVAPI(void) cvCmpS( const CvArr* src, double value, CvArr* dst, int cmp_op );
-
-/** dst(idx) = min(src1(idx),src2(idx)) */
-CVAPI(void) cvMin( const CvArr* src1, const CvArr* src2, CvArr* dst );
-
-/** dst(idx) = max(src1(idx),src2(idx)) */
-CVAPI(void) cvMax( const CvArr* src1, const CvArr* src2, CvArr* dst );
-
-/** dst(idx) = min(src(idx),value) */
-CVAPI(void) cvMinS( const CvArr* src, double value, CvArr* dst );
-
-/** dst(idx) = max(src(idx),value) */
-CVAPI(void) cvMaxS( const CvArr* src, double value, CvArr* dst );
-
-/** dst(x,y,c) = abs(src1(x,y,c) - src2(x,y,c)) */
-CVAPI(void) cvAbsDiff( const CvArr* src1, const CvArr* src2, CvArr* dst );
-
-/** dst(x,y,c) = abs(src(x,y,c) - value(c)) */
-CVAPI(void) cvAbsDiffS( const CvArr* src, CvArr* dst, CvScalar value );
-#define cvAbs( src, dst ) cvAbsDiffS( (src), (dst), cvScalarAll(0))
-
-/****************************************************************************************\
-*                                Math operations                                         *
-\****************************************************************************************/
-
-/** Does cartesian->polar coordinates conversion.
-   Either of output components (magnitude or angle) is optional */
-CVAPI(void)  cvCartToPolar( const CvArr* x, const CvArr* y,
-                            CvArr* magnitude, CvArr* angle CV_DEFAULT(NULL),
-                            int angle_in_degrees CV_DEFAULT(0));
-
-/** Does polar->cartesian coordinates conversion.
-   Either of output components (magnitude or angle) is optional.
-   If magnitude is missing it is assumed to be all 1's */
-CVAPI(void)  cvPolarToCart( const CvArr* magnitude, const CvArr* angle,
-                            CvArr* x, CvArr* y,
-                            int angle_in_degrees CV_DEFAULT(0));
-
-/** Does powering: dst(idx) = src(idx)^power */
-CVAPI(void)  cvPow( const CvArr* src, CvArr* dst, double power );
-
-/** Does exponention: dst(idx) = exp(src(idx)).
-   Overflow is not handled yet. Underflow is handled.
-   Maximal relative error is ~7e-6 for single-precision input */
-CVAPI(void)  cvExp( const CvArr* src, CvArr* dst );
-
-/** Calculates natural logarithms: dst(idx) = log(abs(src(idx))).
-   Logarithm of 0 gives large negative number(~-700)
-   Maximal relative error is ~3e-7 for single-precision output
-*/
-CVAPI(void)  cvLog( const CvArr* src, CvArr* dst );
-
-/** Fast arctangent calculation */
-CVAPI(float) cvFastArctan( float y, float x );
-
-/** Fast cubic root calculation */
-CVAPI(float)  cvCbrt( float value );
-
-#define  CV_CHECK_RANGE    1
-#define  CV_CHECK_QUIET    2
-/** Checks array values for NaNs, Infs or simply for too large numbers
-   (if CV_CHECK_RANGE is set). If CV_CHECK_QUIET is set,
-   no runtime errors is raised (function returns zero value in case of "bad" values).
-   Otherwise cvError is called */
-CVAPI(int)  cvCheckArr( const CvArr* arr, int flags CV_DEFAULT(0),
-                        double min_val CV_DEFAULT(0), double max_val CV_DEFAULT(0));
-#define cvCheckArray cvCheckArr
-
-#define CV_RAND_UNI      0
-#define CV_RAND_NORMAL   1
-
-/** @brief Fills an array with random numbers and updates the RNG state.
-
-The function fills the destination array with uniformly or normally distributed random numbers.
-@param rng CvRNG state initialized by cvRNG
-@param arr The destination array
-@param dist_type Distribution type
-> -   **CV_RAND_UNI** uniform distribution
-> -   **CV_RAND_NORMAL** normal or Gaussian distribution
-@param param1 The first parameter of the distribution. In the case of a uniform distribution it is
-the inclusive lower boundary of the random numbers range. In the case of a normal distribution it
-is the mean value of the random numbers.
-@param param2 The second parameter of the distribution. In the case of a uniform distribution it
-is the exclusive upper boundary of the random numbers range. In the case of a normal distribution
-it is the standard deviation of the random numbers.
-@sa randu, randn, RNG::fill.
- */
-CVAPI(void) cvRandArr( CvRNG* rng, CvArr* arr, int dist_type,
-                      CvScalar param1, CvScalar param2 );
-
-CVAPI(void) cvRandShuffle( CvArr* mat, CvRNG* rng,
-                           double iter_factor CV_DEFAULT(1.));
-
-#define CV_SORT_EVERY_ROW 0
-#define CV_SORT_EVERY_COLUMN 1
-#define CV_SORT_ASCENDING 0
-#define CV_SORT_DESCENDING 16
-
-CVAPI(void) cvSort( const CvArr* src, CvArr* dst CV_DEFAULT(NULL),
-                    CvArr* idxmat CV_DEFAULT(NULL),
-                    int flags CV_DEFAULT(0));
-
-/** Finds real roots of a cubic equation */
-CVAPI(int) cvSolveCubic( const CvMat* coeffs, CvMat* roots );
-
-/** Finds all real and complex roots of a polynomial equation */
-CVAPI(void) cvSolvePoly(const CvMat* coeffs, CvMat *roots2,
-      int maxiter CV_DEFAULT(20), int fig CV_DEFAULT(100));
-
-/****************************************************************************************\
-*                                Matrix operations                                       *
-\****************************************************************************************/
-
-/** @brief Calculates the cross product of two 3D vectors.
-
-The function calculates the cross product of two 3D vectors:
-\f[\texttt{dst} =  \texttt{src1} \times \texttt{src2}\f]
-or:
-\f[\begin{array}{l} \texttt{dst} _1 =  \texttt{src1} _2  \texttt{src2} _3 -  \texttt{src1} _3  \texttt{src2} _2 \\ \texttt{dst} _2 =  \texttt{src1} _3  \texttt{src2} _1 -  \texttt{src1} _1  \texttt{src2} _3 \\ \texttt{dst} _3 =  \texttt{src1} _1  \texttt{src2} _2 -  \texttt{src1} _2  \texttt{src2} _1 \end{array}\f]
-@param src1 The first source vector
-@param src2 The second source vector
-@param dst The destination vector
- */
-CVAPI(void)  cvCrossProduct( const CvArr* src1, const CvArr* src2, CvArr* dst );
-
-/** Matrix transform: dst = A*B + C, C is optional */
-#define cvMatMulAdd( src1, src2, src3, dst ) cvGEMM( (src1), (src2), 1., (src3), 1., (dst), 0 )
-#define cvMatMul( src1, src2, dst )  cvMatMulAdd( (src1), (src2), NULL, (dst))
-
-#define CV_GEMM_A_T 1
-#define CV_GEMM_B_T 2
-#define CV_GEMM_C_T 4
-/** Extended matrix transform:
-   dst = alpha*op(A)*op(B) + beta*op(C), where op(X) is X or X^T */
-CVAPI(void)  cvGEMM( const CvArr* src1, const CvArr* src2, double alpha,
-                     const CvArr* src3, double beta, CvArr* dst,
-                     int tABC CV_DEFAULT(0));
-#define cvMatMulAddEx cvGEMM
-
-/** Transforms each element of source array and stores
-   resultant vectors in destination array */
-CVAPI(void)  cvTransform( const CvArr* src, CvArr* dst,
-                          const CvMat* transmat,
-                          const CvMat* shiftvec CV_DEFAULT(NULL));
-#define cvMatMulAddS cvTransform
-
-/** Does perspective transform on every element of input array */
-CVAPI(void)  cvPerspectiveTransform( const CvArr* src, CvArr* dst,
-                                     const CvMat* mat );
-
-/** Calculates (A-delta)*(A-delta)^T (order=0) or (A-delta)^T*(A-delta) (order=1) */
-CVAPI(void) cvMulTransposed( const CvArr* src, CvArr* dst, int order,
-                             const CvArr* delta CV_DEFAULT(NULL),
-                             double scale CV_DEFAULT(1.) );
-
-/** Transposes matrix. Square matrices can be transposed in-place */
-CVAPI(void)  cvTranspose( const CvArr* src, CvArr* dst );
-#define cvT cvTranspose
-
-/** Completes the symmetric matrix from the lower (LtoR=0) or from the upper (LtoR!=0) part */
-CVAPI(void)  cvCompleteSymm( CvMat* matrix, int LtoR CV_DEFAULT(0) );
-
-/** Mirror array data around horizontal (flip=0),
-   vertical (flip=1) or both(flip=-1) axises:
-   cvFlip(src) flips images vertically and sequences horizontally (inplace) */
-CVAPI(void)  cvFlip( const CvArr* src, CvArr* dst CV_DEFAULT(NULL),
-                     int flip_mode CV_DEFAULT(0));
-#define cvMirror cvFlip
-
-
-#define CV_SVD_MODIFY_A   1
-#define CV_SVD_U_T        2
-#define CV_SVD_V_T        4
-
-/** Performs Singular Value Decomposition of a matrix */
-CVAPI(void)   cvSVD( CvArr* A, CvArr* W, CvArr* U CV_DEFAULT(NULL),
-                     CvArr* V CV_DEFAULT(NULL), int flags CV_DEFAULT(0));
-
-/** Performs Singular Value Back Substitution (solves A*X = B):
-   flags must be the same as in cvSVD */
-CVAPI(void)   cvSVBkSb( const CvArr* W, const CvArr* U,
-                        const CvArr* V, const CvArr* B,
-                        CvArr* X, int flags );
-
-#define CV_LU  0
-#define CV_SVD 1
-#define CV_SVD_SYM 2
-#define CV_CHOLESKY 3
-#define CV_QR  4
-#define CV_NORMAL 16
-
-/** Inverts matrix */
-CVAPI(double)  cvInvert( const CvArr* src, CvArr* dst,
-                         int method CV_DEFAULT(CV_LU));
-#define cvInv cvInvert
-
-/** Solves linear system (src1)*(dst) = (src2)
-   (returns 0 if src1 is a singular and CV_LU method is used) */
-CVAPI(int)  cvSolve( const CvArr* src1, const CvArr* src2, CvArr* dst,
-                     int method CV_DEFAULT(CV_LU));
-
-/** Calculates determinant of input matrix */
-CVAPI(double) cvDet( const CvArr* mat );
-
-/** Calculates trace of the matrix (sum of elements on the main diagonal) */
-CVAPI(CvScalar) cvTrace( const CvArr* mat );
-
-/** Finds eigen values and vectors of a symmetric matrix */
-CVAPI(void)  cvEigenVV( CvArr* mat, CvArr* evects, CvArr* evals,
-                        double eps CV_DEFAULT(0),
-                        int lowindex CV_DEFAULT(-1),
-                        int highindex CV_DEFAULT(-1));
-
-///* Finds selected eigen values and vectors of a symmetric matrix */
-//CVAPI(void)  cvSelectedEigenVV( CvArr* mat, CvArr* evects, CvArr* evals,
-//                                int lowindex, int highindex );
-
-/** Makes an identity matrix (mat_ij = i == j) */
-CVAPI(void)  cvSetIdentity( CvArr* mat, CvScalar value CV_DEFAULT(cvRealScalar(1)) );
-
-/** Fills matrix with given range of numbers */
-CVAPI(CvArr*)  cvRange( CvArr* mat, double start, double end );
-
-/**   @anchor core_c_CovarFlags
-@name Flags for cvCalcCovarMatrix
-@see cvCalcCovarMatrix
-  @{
-*/
-
-/** flag for cvCalcCovarMatrix, transpose([v1-avg, v2-avg,...]) * [v1-avg,v2-avg,...] */
-#define CV_COVAR_SCRAMBLED 0
-
-/** flag for cvCalcCovarMatrix, [v1-avg, v2-avg,...] * transpose([v1-avg,v2-avg,...]) */
-#define CV_COVAR_NORMAL    1
-
-/** flag for cvCalcCovarMatrix, do not calc average (i.e. mean vector) - use the input vector instead
-   (useful for calculating covariance matrix by parts) */
-#define CV_COVAR_USE_AVG   2
-
-/** flag for cvCalcCovarMatrix, scale the covariance matrix coefficients by number of the vectors */
-#define CV_COVAR_SCALE     4
-
-/** flag for cvCalcCovarMatrix, all the input vectors are stored in a single matrix, as its rows */
-#define CV_COVAR_ROWS      8
-
-/** flag for cvCalcCovarMatrix, all the input vectors are stored in a single matrix, as its columns */
-#define CV_COVAR_COLS     16
-
-/** @} */
-
-/** Calculates covariation matrix for a set of vectors
-@see @ref core_c_CovarFlags "flags"
-*/
-CVAPI(void)  cvCalcCovarMatrix( const CvArr** vects, int count,
-                                CvArr* cov_mat, CvArr* avg, int flags );
-
-#define CV_PCA_DATA_AS_ROW 0
-#define CV_PCA_DATA_AS_COL 1
-#define CV_PCA_USE_AVG 2
-CVAPI(void)  cvCalcPCA( const CvArr* data, CvArr* mean,
-                        CvArr* eigenvals, CvArr* eigenvects, int flags );
-
-CVAPI(void)  cvProjectPCA( const CvArr* data, const CvArr* mean,
-                           const CvArr* eigenvects, CvArr* result );
-
-CVAPI(void)  cvBackProjectPCA( const CvArr* proj, const CvArr* mean,
-                               const CvArr* eigenvects, CvArr* result );
-
-/** Calculates Mahalanobis(weighted) distance */
-CVAPI(double)  cvMahalanobis( const CvArr* vec1, const CvArr* vec2, const CvArr* mat );
-#define cvMahalonobis  cvMahalanobis
-
-/****************************************************************************************\
-*                                    Array Statistics                                    *
-\****************************************************************************************/
-
-/** Finds sum of array elements */
-CVAPI(CvScalar)  cvSum( const CvArr* arr );
-
-/** Calculates number of non-zero pixels */
-CVAPI(int)  cvCountNonZero( const CvArr* arr );
-
-/** Calculates mean value of array elements */
-CVAPI(CvScalar)  cvAvg( const CvArr* arr, const CvArr* mask CV_DEFAULT(NULL) );
-
-/** Calculates mean and standard deviation of pixel values */
-CVAPI(void)  cvAvgSdv( const CvArr* arr, CvScalar* mean, CvScalar* std_dev,
-                       const CvArr* mask CV_DEFAULT(NULL) );
-
-/** Finds global minimum, maximum and their positions */
-CVAPI(void)  cvMinMaxLoc( const CvArr* arr, double* min_val, double* max_val,
-                          CvPoint* min_loc CV_DEFAULT(NULL),
-                          CvPoint* max_loc CV_DEFAULT(NULL),
-                          const CvArr* mask CV_DEFAULT(NULL) );
-
-/** @anchor core_c_NormFlags
-  @name Flags for cvNorm and cvNormalize
-  @{
-*/
-#define CV_C            1
-#define CV_L1           2
-#define CV_L2           4
-#define CV_NORM_MASK    7
-#define CV_RELATIVE     8
-#define CV_DIFF         16
-#define CV_MINMAX       32
-
-#define CV_DIFF_C       (CV_DIFF | CV_C)
-#define CV_DIFF_L1      (CV_DIFF | CV_L1)
-#define CV_DIFF_L2      (CV_DIFF | CV_L2)
-#define CV_RELATIVE_C   (CV_RELATIVE | CV_C)
-#define CV_RELATIVE_L1  (CV_RELATIVE | CV_L1)
-#define CV_RELATIVE_L2  (CV_RELATIVE | CV_L2)
-/** @} */
-
-/** Finds norm, difference norm or relative difference norm for an array (or two arrays)
-@see ref core_c_NormFlags "flags"
-*/
-CVAPI(double)  cvNorm( const CvArr* arr1, const CvArr* arr2 CV_DEFAULT(NULL),
-                       int norm_type CV_DEFAULT(CV_L2),
-                       const CvArr* mask CV_DEFAULT(NULL) );
-
-/** @see ref core_c_NormFlags "flags" */
-CVAPI(void)  cvNormalize( const CvArr* src, CvArr* dst,
-                          double a CV_DEFAULT(1.), double b CV_DEFAULT(0.),
-                          int norm_type CV_DEFAULT(CV_L2),
-                          const CvArr* mask CV_DEFAULT(NULL) );
-
-/** @anchor core_c_ReduceFlags
-  @name Flags for cvReduce
-  @{
-*/
-#define CV_REDUCE_SUM 0
-#define CV_REDUCE_AVG 1
-#define CV_REDUCE_MAX 2
-#define CV_REDUCE_MIN 3
-/** @} */
-
-/** @see @ref core_c_ReduceFlags "flags" */
-CVAPI(void)  cvReduce( const CvArr* src, CvArr* dst, int dim CV_DEFAULT(-1),
-                       int op CV_DEFAULT(CV_REDUCE_SUM) );
-
-/****************************************************************************************\
-*                      Discrete Linear Transforms and Related Functions                  *
-\****************************************************************************************/
-
-/** @anchor core_c_DftFlags
-  @name Flags for cvDFT, cvDCT and cvMulSpectrums
-  @{
-  */
-#define CV_DXT_FORWARD  0
-#define CV_DXT_INVERSE  1
-#define CV_DXT_SCALE    2 /**< divide result by size of array */
-#define CV_DXT_INV_SCALE (CV_DXT_INVERSE + CV_DXT_SCALE)
-#define CV_DXT_INVERSE_SCALE CV_DXT_INV_SCALE
-#define CV_DXT_ROWS     4 /**< transform each row individually */
-#define CV_DXT_MUL_CONJ 8 /**< conjugate the second argument of cvMulSpectrums */
-/** @} */
-
-/** Discrete Fourier Transform:
-    complex->complex,
-    real->ccs (forward),
-    ccs->real (inverse)
-@see core_c_DftFlags "flags"
-*/
-CVAPI(void)  cvDFT( const CvArr* src, CvArr* dst, int flags,
-                    int nonzero_rows CV_DEFAULT(0) );
-#define cvFFT cvDFT
-
-/** Multiply results of DFTs: DFT(X)*DFT(Y) or DFT(X)*conj(DFT(Y))
-@see core_c_DftFlags "flags"
-*/
-CVAPI(void)  cvMulSpectrums( const CvArr* src1, const CvArr* src2,
-                             CvArr* dst, int flags );
-
-/** Finds optimal DFT vector size >= size0 */
-CVAPI(int)  cvGetOptimalDFTSize( int size0 );
-
-/** Discrete Cosine Transform
-@see core_c_DftFlags "flags"
-*/
-CVAPI(void)  cvDCT( const CvArr* src, CvArr* dst, int flags );
-
-/****************************************************************************************\
-*                              Dynamic data structures                                   *
-\****************************************************************************************/
-
-/** Calculates length of sequence slice (with support of negative indices). */
-CVAPI(int) cvSliceLength( CvSlice slice, const CvSeq* seq );
-
-
-/** Creates new memory storage.
-   block_size == 0 means that default,
-   somewhat optimal size, is used (currently, it is 64K) */
-CVAPI(CvMemStorage*)  cvCreateMemStorage( int block_size CV_DEFAULT(0));
-
-
-/** Creates a memory storage that will borrow memory blocks from parent storage */
-CVAPI(CvMemStorage*)  cvCreateChildMemStorage( CvMemStorage* parent );
-
-
-/** Releases memory storage. All the children of a parent must be released before
-   the parent. A child storage returns all the blocks to parent when it is released */
-CVAPI(void)  cvReleaseMemStorage( CvMemStorage** storage );
-
-
-/** Clears memory storage. This is the only way(!!!) (besides cvRestoreMemStoragePos)
-   to reuse memory allocated for the storage - cvClearSeq,cvClearSet ...
-   do not free any memory.
-   A child storage returns all the blocks to the parent when it is cleared */
-CVAPI(void)  cvClearMemStorage( CvMemStorage* storage );
-
-/** Remember a storage "free memory" position */
-CVAPI(void)  cvSaveMemStoragePos( const CvMemStorage* storage, CvMemStoragePos* pos );
-
-/** Restore a storage "free memory" position */
-CVAPI(void)  cvRestoreMemStoragePos( CvMemStorage* storage, CvMemStoragePos* pos );
-
-/** Allocates continuous buffer of the specified size in the storage */
-CVAPI(void*) cvMemStorageAlloc( CvMemStorage* storage, size_t size );
-
-/** Allocates string in memory storage */
-//CVAPI(CvString) cvMemStorageAllocString( CvMemStorage* storage, const char* ptr,
-//                                         int len CV_DEFAULT(-1) );
-
-/** Creates new empty sequence that will reside in the specified storage */
-CVAPI(CvSeq*)  cvCreateSeq( int seq_flags, size_t header_size,
-                            size_t elem_size, CvMemStorage* storage );
-
-/** Changes default size (granularity) of sequence blocks.
-   The default size is ~1Kbyte */
-CVAPI(void)  cvSetSeqBlockSize( CvSeq* seq, int delta_elems );
-
-
-/** Adds new element to the end of sequence. Returns pointer to the element */
-CVAPI(schar*)  cvSeqPush( CvSeq* seq, const void* element CV_DEFAULT(NULL));
-
-
-/** Adds new element to the beginning of sequence. Returns pointer to it */
-CVAPI(schar*)  cvSeqPushFront( CvSeq* seq, const void* element CV_DEFAULT(NULL));
-
-
-/** Removes the last element from sequence and optionally saves it */
-CVAPI(void)  cvSeqPop( CvSeq* seq, void* element CV_DEFAULT(NULL));
-
-
-/** Removes the first element from sequence and optioanally saves it */
-CVAPI(void)  cvSeqPopFront( CvSeq* seq, void* element CV_DEFAULT(NULL));
-
-
-#define CV_FRONT 1
-#define CV_BACK 0
-/** Adds several new elements to the end of sequence */
-CVAPI(void)  cvSeqPushMulti( CvSeq* seq, const void* elements,
-                             int count, int in_front CV_DEFAULT(0) );
-
-/** Removes several elements from the end of sequence and optionally saves them */
-CVAPI(void)  cvSeqPopMulti( CvSeq* seq, void* elements,
-                            int count, int in_front CV_DEFAULT(0) );
-
-/** Inserts a new element in the middle of sequence.
-   cvSeqInsert(seq,0,elem) == cvSeqPushFront(seq,elem) */
-CVAPI(schar*)  cvSeqInsert( CvSeq* seq, int before_index,
-                            const void* element CV_DEFAULT(NULL));
-
-/** Removes specified sequence element */
-CVAPI(void)  cvSeqRemove( CvSeq* seq, int index );
-
-
-/** Removes all the elements from the sequence. The freed memory
-   can be reused later only by the same sequence unless cvClearMemStorage
-   or cvRestoreMemStoragePos is called */
-CVAPI(void)  cvClearSeq( CvSeq* seq );
-
-
-/** Retrieves pointer to specified sequence element.
-   Negative indices are supported and mean counting from the end
-   (e.g -1 means the last sequence element) */
-CVAPI(schar*)  cvGetSeqElem( const CvSeq* seq, int index );
-
-/** Calculates index of the specified sequence element.
-   Returns -1 if element does not belong to the sequence */
-CVAPI(int)  cvSeqElemIdx( const CvSeq* seq, const void* element,
-                         CvSeqBlock** block CV_DEFAULT(NULL) );
-
-/** Initializes sequence writer. The new elements will be added to the end of sequence */
-CVAPI(void)  cvStartAppendToSeq( CvSeq* seq, CvSeqWriter* writer );
-
-
-/** Combination of cvCreateSeq and cvStartAppendToSeq */
-CVAPI(void)  cvStartWriteSeq( int seq_flags, int header_size,
-                              int elem_size, CvMemStorage* storage,
-                              CvSeqWriter* writer );
-
-/** Closes sequence writer, updates sequence header and returns pointer
-   to the resultant sequence
-   (which may be useful if the sequence was created using cvStartWriteSeq))
-*/
-CVAPI(CvSeq*)  cvEndWriteSeq( CvSeqWriter* writer );
-
-
-/** Updates sequence header. May be useful to get access to some of previously
-   written elements via cvGetSeqElem or sequence reader */
-CVAPI(void)   cvFlushSeqWriter( CvSeqWriter* writer );
-
-
-/** Initializes sequence reader.
-   The sequence can be read in forward or backward direction */
-CVAPI(void) cvStartReadSeq( const CvSeq* seq, CvSeqReader* reader,
-                           int reverse CV_DEFAULT(0) );
-
-
-/** Returns current sequence reader position (currently observed sequence element) */
-CVAPI(int)  cvGetSeqReaderPos( CvSeqReader* reader );
-
-
-/** Changes sequence reader position. It may seek to an absolute or
-   to relative to the current position */
-CVAPI(void)   cvSetSeqReaderPos( CvSeqReader* reader, int index,
-                                 int is_relative CV_DEFAULT(0));
-
-/** Copies sequence content to a continuous piece of memory */
-CVAPI(void*)  cvCvtSeqToArray( const CvSeq* seq, void* elements,
-                               CvSlice slice CV_DEFAULT(CV_WHOLE_SEQ) );
-
-/** Creates sequence header for array.
-   After that all the operations on sequences that do not alter the content
-   can be applied to the resultant sequence */
-CVAPI(CvSeq*) cvMakeSeqHeaderForArray( int seq_type, int header_size,
-                                       int elem_size, void* elements, int total,
-                                       CvSeq* seq, CvSeqBlock* block );
-
-/** Extracts sequence slice (with or without copying sequence elements) */
-CVAPI(CvSeq*) cvSeqSlice( const CvSeq* seq, CvSlice slice,
-                         CvMemStorage* storage CV_DEFAULT(NULL),
-                         int copy_data CV_DEFAULT(0));
-
-CV_INLINE CvSeq* cvCloneSeq( const CvSeq* seq, CvMemStorage* storage CV_DEFAULT(NULL))
-{
-    return cvSeqSlice( seq, CV_WHOLE_SEQ, storage, 1 );
-}
-
-/** Removes sequence slice */
-CVAPI(void)  cvSeqRemoveSlice( CvSeq* seq, CvSlice slice );
-
-/** Inserts a sequence or array into another sequence */
-CVAPI(void)  cvSeqInsertSlice( CvSeq* seq, int before_index, const CvArr* from_arr );
-
-/** a < b ? -1 : a > b ? 1 : 0 */
-typedef int (CV_CDECL* CvCmpFunc)(const void* a, const void* b, void* userdata );
-
-/** Sorts sequence in-place given element comparison function */
-CVAPI(void) cvSeqSort( CvSeq* seq, CvCmpFunc func, void* userdata CV_DEFAULT(NULL) );
-
-/** Finds element in a [sorted] sequence */
-CVAPI(schar*) cvSeqSearch( CvSeq* seq, const void* elem, CvCmpFunc func,
-                           int is_sorted, int* elem_idx,
-                           void* userdata CV_DEFAULT(NULL) );
-
-/** Reverses order of sequence elements in-place */
-CVAPI(void) cvSeqInvert( CvSeq* seq );
-
-/** Splits sequence into one or more equivalence classes using the specified criteria */
-CVAPI(int)  cvSeqPartition( const CvSeq* seq, CvMemStorage* storage,
-                            CvSeq** labels, CvCmpFunc is_equal, void* userdata );
-
-/************ Internal sequence functions ************/
-CVAPI(void)  cvChangeSeqBlock( void* reader, int direction );
-CVAPI(void)  cvCreateSeqBlock( CvSeqWriter* writer );
-
-
-/** Creates a new set */
-CVAPI(CvSet*)  cvCreateSet( int set_flags, int header_size,
-                            int elem_size, CvMemStorage* storage );
-
-/** Adds new element to the set and returns pointer to it */
-CVAPI(int)  cvSetAdd( CvSet* set_header, CvSetElem* elem CV_DEFAULT(NULL),
-                      CvSetElem** inserted_elem CV_DEFAULT(NULL) );
-
-/** Fast variant of cvSetAdd */
-CV_INLINE  CvSetElem* cvSetNew( CvSet* set_header )
-{
-    CvSetElem* elem = set_header->free_elems;
-    if( elem )
-    {
-        set_header->free_elems = elem->next_free;
-        elem->flags = elem->flags & CV_SET_ELEM_IDX_MASK;
-        set_header->active_count++;
-    }
-    else
-        cvSetAdd( set_header, NULL, &elem );
-    return elem;
-}
-
-/** Removes set element given its pointer */
-CV_INLINE  void cvSetRemoveByPtr( CvSet* set_header, void* elem )
-{
-    CvSetElem* _elem = (CvSetElem*)elem;
-    assert( _elem->flags >= 0 /*&& (elem->flags & CV_SET_ELEM_IDX_MASK) < set_header->total*/ );
-    _elem->next_free = set_header->free_elems;
-    _elem->flags = (_elem->flags & CV_SET_ELEM_IDX_MASK) | CV_SET_ELEM_FREE_FLAG;
-    set_header->free_elems = _elem;
-    set_header->active_count--;
-}
-
-/** Removes element from the set by its index  */
-CVAPI(void)   cvSetRemove( CvSet* set_header, int index );
-
-/** Returns a set element by index. If the element doesn't belong to the set,
-   NULL is returned */
-CV_INLINE CvSetElem* cvGetSetElem( const CvSet* set_header, int idx )
-{
-    CvSetElem* elem = (CvSetElem*)(void *)cvGetSeqElem( (CvSeq*)set_header, idx );
-    return elem && CV_IS_SET_ELEM( elem ) ? elem : 0;
-}
-
-/** Removes all the elements from the set */
-CVAPI(void)  cvClearSet( CvSet* set_header );
-
-/** Creates new graph */
-CVAPI(CvGraph*)  cvCreateGraph( int graph_flags, int header_size,
-                                int vtx_size, int edge_size,
-                                CvMemStorage* storage );
-
-/** Adds new vertex to the graph */
-CVAPI(int)  cvGraphAddVtx( CvGraph* graph, const CvGraphVtx* vtx CV_DEFAULT(NULL),
-                           CvGraphVtx** inserted_vtx CV_DEFAULT(NULL) );
-
-
-/** Removes vertex from the graph together with all incident edges */
-CVAPI(int)  cvGraphRemoveVtx( CvGraph* graph, int index );
-CVAPI(int)  cvGraphRemoveVtxByPtr( CvGraph* graph, CvGraphVtx* vtx );
-
-
-/** Link two vertices specified by indices or pointers if they
-   are not connected or return pointer to already existing edge
-   connecting the vertices.
-   Functions return 1 if a new edge was created, 0 otherwise */
-CVAPI(int)  cvGraphAddEdge( CvGraph* graph,
-                            int start_idx, int end_idx,
-                            const CvGraphEdge* edge CV_DEFAULT(NULL),
-                            CvGraphEdge** inserted_edge CV_DEFAULT(NULL) );
-
-CVAPI(int)  cvGraphAddEdgeByPtr( CvGraph* graph,
-                               CvGraphVtx* start_vtx, CvGraphVtx* end_vtx,
-                               const CvGraphEdge* edge CV_DEFAULT(NULL),
-                               CvGraphEdge** inserted_edge CV_DEFAULT(NULL) );
-
-/** Remove edge connecting two vertices */
-CVAPI(void)  cvGraphRemoveEdge( CvGraph* graph, int start_idx, int end_idx );
-CVAPI(void)  cvGraphRemoveEdgeByPtr( CvGraph* graph, CvGraphVtx* start_vtx,
-                                     CvGraphVtx* end_vtx );
-
-/** Find edge connecting two vertices */
-CVAPI(CvGraphEdge*)  cvFindGraphEdge( const CvGraph* graph, int start_idx, int end_idx );
-CVAPI(CvGraphEdge*)  cvFindGraphEdgeByPtr( const CvGraph* graph,
-                                           const CvGraphVtx* start_vtx,
-                                           const CvGraphVtx* end_vtx );
-#define cvGraphFindEdge cvFindGraphEdge
-#define cvGraphFindEdgeByPtr cvFindGraphEdgeByPtr
-
-/** Remove all vertices and edges from the graph */
-CVAPI(void)  cvClearGraph( CvGraph* graph );
-
-
-/** Count number of edges incident to the vertex */
-CVAPI(int)  cvGraphVtxDegree( const CvGraph* graph, int vtx_idx );
-CVAPI(int)  cvGraphVtxDegreeByPtr( const CvGraph* graph, const CvGraphVtx* vtx );
-
-
-/** Retrieves graph vertex by given index */
-#define cvGetGraphVtx( graph, idx ) (CvGraphVtx*)cvGetSetElem((CvSet*)(graph), (idx))
-
-/** Retrieves index of a graph vertex given its pointer */
-#define cvGraphVtxIdx( graph, vtx ) ((vtx)->flags & CV_SET_ELEM_IDX_MASK)
-
-/** Retrieves index of a graph edge given its pointer */
-#define cvGraphEdgeIdx( graph, edge ) ((edge)->flags & CV_SET_ELEM_IDX_MASK)
-
-#define cvGraphGetVtxCount( graph ) ((graph)->active_count)
-#define cvGraphGetEdgeCount( graph ) ((graph)->edges->active_count)
-
-#define  CV_GRAPH_VERTEX        1
-#define  CV_GRAPH_TREE_EDGE     2
-#define  CV_GRAPH_BACK_EDGE     4
-#define  CV_GRAPH_FORWARD_EDGE  8
-#define  CV_GRAPH_CROSS_EDGE    16
-#define  CV_GRAPH_ANY_EDGE      30
-#define  CV_GRAPH_NEW_TREE      32
-#define  CV_GRAPH_BACKTRACKING  64
-#define  CV_GRAPH_OVER          -1
-
-#define  CV_GRAPH_ALL_ITEMS    -1
-
-/** flags for graph vertices and edges */
-#define  CV_GRAPH_ITEM_VISITED_FLAG  (1 << 30)
-#define  CV_IS_GRAPH_VERTEX_VISITED(vtx) \
-    (((CvGraphVtx*)(vtx))->flags & CV_GRAPH_ITEM_VISITED_FLAG)
-#define  CV_IS_GRAPH_EDGE_VISITED(edge) \
-    (((CvGraphEdge*)(edge))->flags & CV_GRAPH_ITEM_VISITED_FLAG)
-#define  CV_GRAPH_SEARCH_TREE_NODE_FLAG   (1 << 29)
-#define  CV_GRAPH_FORWARD_EDGE_FLAG       (1 << 28)
-
-typedef struct CvGraphScanner
-{
-    CvGraphVtx* vtx;       /* current graph vertex (or current edge origin) */
-    CvGraphVtx* dst;       /* current graph edge destination vertex */
-    CvGraphEdge* edge;     /* current edge */
-
-    CvGraph* graph;        /* the graph */
-    CvSeq*   stack;        /* the graph vertex stack */
-    int      index;        /* the lower bound of certainly visited vertices */
-    int      mask;         /* event mask */
-}
-CvGraphScanner;
-
-/** Creates new graph scanner. */
-CVAPI(CvGraphScanner*)  cvCreateGraphScanner( CvGraph* graph,
-                                             CvGraphVtx* vtx CV_DEFAULT(NULL),
-                                             int mask CV_DEFAULT(CV_GRAPH_ALL_ITEMS));
-
-/** Releases graph scanner. */
-CVAPI(void) cvReleaseGraphScanner( CvGraphScanner** scanner );
-
-/** Get next graph element */
-CVAPI(int)  cvNextGraphItem( CvGraphScanner* scanner );
-
-/** Creates a copy of graph */
-CVAPI(CvGraph*) cvCloneGraph( const CvGraph* graph, CvMemStorage* storage );
-
-
-/** Does look-up transformation. Elements of the source array
-   (that should be 8uC1 or 8sC1) are used as indexes in lutarr 256-element table */
-CVAPI(void) cvLUT( const CvArr* src, CvArr* dst, const CvArr* lut );
-
-
-/******************* Iteration through the sequence tree *****************/
-typedef struct CvTreeNodeIterator
-{
-    const void* node;
-    int level;
-    int max_level;
-}
-CvTreeNodeIterator;
-
-CVAPI(void) cvInitTreeNodeIterator( CvTreeNodeIterator* tree_iterator,
-                                   const void* first, int max_level );
-CVAPI(void*) cvNextTreeNode( CvTreeNodeIterator* tree_iterator );
-CVAPI(void*) cvPrevTreeNode( CvTreeNodeIterator* tree_iterator );
-
-/** Inserts sequence into tree with specified "parent" sequence.
-   If parent is equal to frame (e.g. the most external contour),
-   then added contour will have null pointer to parent. */
-CVAPI(void) cvInsertNodeIntoTree( void* node, void* parent, void* frame );
-
-/** Removes contour from tree (together with the contour children). */
-CVAPI(void) cvRemoveNodeFromTree( void* node, void* frame );
-
-/** Gathers pointers to all the sequences,
-   accessible from the `first`, to the single sequence */
-CVAPI(CvSeq*) cvTreeToNodeSeq( const void* first, int header_size,
-                              CvMemStorage* storage );
-
-/** The function implements the K-means algorithm for clustering an array of sample
-   vectors in a specified number of classes */
-#define CV_KMEANS_USE_INITIAL_LABELS    1
-CVAPI(int) cvKMeans2( const CvArr* samples, int cluster_count, CvArr* labels,
-                      CvTermCriteria termcrit, int attempts CV_DEFAULT(1),
-                      CvRNG* rng CV_DEFAULT(0), int flags CV_DEFAULT(0),
-                      CvArr* _centers CV_DEFAULT(0), double* compactness CV_DEFAULT(0) );
-
-/****************************************************************************************\
-*                                    System functions                                    *
-\****************************************************************************************/
-
-/** Loads optimized functions from IPP, MKL etc. or switches back to pure C code */
-CVAPI(int)  cvUseOptimized( int on_off );
-
-typedef IplImage* (CV_STDCALL* Cv_iplCreateImageHeader)
-                            (int,int,int,char*,char*,int,int,int,int,int,
-                            IplROI*,IplImage*,void*,IplTileInfo*);
-typedef void (CV_STDCALL* Cv_iplAllocateImageData)(IplImage*,int,int);
-typedef void (CV_STDCALL* Cv_iplDeallocate)(IplImage*,int);
-typedef IplROI* (CV_STDCALL* Cv_iplCreateROI)(int,int,int,int,int);
-typedef IplImage* (CV_STDCALL* Cv_iplCloneImage)(const IplImage*);
-
-/** @brief Makes OpenCV use IPL functions for allocating IplImage and IplROI structures.
-
-Normally, the function is not called directly. Instead, a simple macro
-CV_TURN_ON_IPL_COMPATIBILITY() is used that calls cvSetIPLAllocators and passes there pointers
-to IPL allocation functions. :
-@code
-    ...
-    CV_TURN_ON_IPL_COMPATIBILITY()
-    ...
-@endcode
-@param create_header pointer to a function, creating IPL image header.
-@param allocate_data pointer to a function, allocating IPL image data.
-@param deallocate pointer to a function, deallocating IPL image.
-@param create_roi pointer to a function, creating IPL image ROI (i.e. Region of Interest).
-@param clone_image pointer to a function, cloning an IPL image.
- */
-CVAPI(void) cvSetIPLAllocators( Cv_iplCreateImageHeader create_header,
-                               Cv_iplAllocateImageData allocate_data,
-                               Cv_iplDeallocate deallocate,
-                               Cv_iplCreateROI create_roi,
-                               Cv_iplCloneImage clone_image );
-
-#define CV_TURN_ON_IPL_COMPATIBILITY()                                  \
-    cvSetIPLAllocators( iplCreateImageHeader, iplAllocateImage,         \
-                        iplDeallocate, iplCreateROI, iplCloneImage )
-
-/****************************************************************************************\
-*                                    Data Persistence                                    *
-\****************************************************************************************/
-
-#if 0
-/********************************** High-level functions ********************************/
-
-/** @brief Opens file storage for reading or writing data.
-
-The function opens file storage for reading or writing data. In the latter case, a new file is
-created or an existing file is rewritten. The type of the read or written file is determined by the
-filename extension: .xml for XML, .yml or .yaml for YAML and .json for JSON.
-
-At the same time, it also supports adding parameters like "example.xml?base64".
-
-The function returns a pointer to the CvFileStorage structure.
-If the file cannot be opened then the function returns NULL.
-@param filename Name of the file associated with the storage
-@param memstorage Memory storage used for temporary data and for
-:   storing dynamic structures, such as CvSeq or CvGraph . If it is NULL, a temporary memory
-    storage is created and used.
-@param flags Can be one of the following:
-> -   **CV_STORAGE_READ** the storage is open for reading
-> -   **CV_STORAGE_WRITE** the storage is open for writing
-      (use **CV_STORAGE_WRITE | CV_STORAGE_WRITE_BASE64** to write rawdata in Base64)
-@param encoding
- */
-CVAPI(CvFileStorage*)  cvOpenFileStorage( const char* filename, CvMemStorage* memstorage,
-                                          int flags, const char* encoding CV_DEFAULT(NULL) );
-
-/** @brief Releases file storage.
-
-The function closes the file associated with the storage and releases all the temporary structures.
-It must be called after all I/O operations with the storage are finished.
-@param fs Double pointer to the released file storage
- */
-CVAPI(void) cvReleaseFileStorage( CvFileStorage** fs );
-
-/** returns attribute value or 0 (NULL) if there is no such attribute */
-CVAPI(const char*) cvAttrValue( const CvAttrList* attr, const char* attr_name );
-
-/** @brief Starts writing a new structure.
-
-The function starts writing a compound structure (collection) that can be a sequence or a map. After
-all the structure fields, which can be scalars or structures, are written, cvEndWriteStruct should
-be called. The function can be used to group some objects or to implement the write function for a
-some user object (see CvTypeInfo).
-@param fs File storage
-@param name Name of the written structure. The structure can be accessed by this name when the
-storage is read.
-@param struct_flags A combination one of the following values:
--   **CV_NODE_SEQ** the written structure is a sequence (see discussion of CvFileStorage ),
-    that is, its elements do not have a name.
--   **CV_NODE_MAP** the written structure is a map (see discussion of CvFileStorage ), that
-    is, all its elements have names.
-One and only one of the two above flags must be specified
--   **CV_NODE_FLOW** the optional flag that makes sense only for YAML streams. It means that
-     the structure is written as a flow (not as a block), which is more compact. It is
-     recommended to use this flag for structures or arrays whose elements are all scalars.
-@param type_name Optional parameter - the object type name. In
-    case of XML it is written as a type_id attribute of the structure opening tag. In the case of
-    YAML it is written after a colon following the structure name (see the example in
-    CvFileStorage description). In case of JSON it is written as a name/value pair.
-    Mainly it is used with user objects. When the storage is read, the
-    encoded type name is used to determine the object type (see CvTypeInfo and cvFindType ).
-@param attributes This parameter is not used in the current implementation
- */
-CVAPI(void) cvStartWriteStruct( CvFileStorage* fs, const char* name,
-                                int struct_flags, const char* type_name CV_DEFAULT(NULL),
-                                CvAttrList attributes CV_DEFAULT(cvAttrList()));
-
-/** @brief Finishes writing to a file node collection.
-@param fs File storage
-@sa cvStartWriteStruct.
- */
-CVAPI(void) cvEndWriteStruct( CvFileStorage* fs );
-
-/** @brief Writes an integer value.
-
-The function writes a single integer value (with or without a name) to the file storage.
-@param fs File storage
-@param name Name of the written value. Should be NULL if and only if the parent structure is a
-sequence.
-@param value The written value
- */
-CVAPI(void) cvWriteInt( CvFileStorage* fs, const char* name, int value );
-
-/** @brief Writes a floating-point value.
-
-The function writes a single floating-point value (with or without a name) to file storage. Special
-values are encoded as follows: NaN (Not A Number) as .NaN, infinity as +.Inf or -.Inf.
-
-The following example shows how to use the low-level writing functions to store custom structures,
-such as termination criteria, without registering a new type. :
-@code
-    void write_termcriteria( CvFileStorage* fs, const char* struct_name,
-                             CvTermCriteria* termcrit )
-    {
-        cvStartWriteStruct( fs, struct_name, CV_NODE_MAP, NULL, cvAttrList(0,0));
-        cvWriteComment( fs, "termination criteria", 1 ); // just a description
-        if( termcrit->type & CV_TERMCRIT_ITER )
-            cvWriteInteger( fs, "max_iterations", termcrit->max_iter );
-        if( termcrit->type & CV_TERMCRIT_EPS )
-            cvWriteReal( fs, "accuracy", termcrit->epsilon );
-        cvEndWriteStruct( fs );
-    }
-@endcode
-@param fs File storage
-@param name Name of the written value. Should be NULL if and only if the parent structure is a
-sequence.
-@param value The written value
-*/
-CVAPI(void) cvWriteReal( CvFileStorage* fs, const char* name, double value );
-
-/** @brief Writes a text string.
-
-The function writes a text string to file storage.
-@param fs File storage
-@param name Name of the written string . Should be NULL if and only if the parent structure is a
-sequence.
-@param str The written text string
-@param quote If non-zero, the written string is put in quotes, regardless of whether they are
-required. Otherwise, if the flag is zero, quotes are used only when they are required (e.g. when
-the string starts with a digit or contains spaces).
- */
-CVAPI(void) cvWriteString( CvFileStorage* fs, const char* name,
-                           const char* str, int quote CV_DEFAULT(0) );
-
-/** @brief Writes a comment.
-
-The function writes a comment into file storage. The comments are skipped when the storage is read.
-@param fs File storage
-@param comment The written comment, single-line or multi-line
-@param eol_comment If non-zero, the function tries to put the comment at the end of current line.
-If the flag is zero, if the comment is multi-line, or if it does not fit at the end of the current
-line, the comment starts a new line.
- */
-CVAPI(void) cvWriteComment( CvFileStorage* fs, const char* comment,
-                            int eol_comment );
-
-/** @brief Writes an object to file storage.
-
-The function writes an object to file storage. First, the appropriate type info is found using
-cvTypeOf. Then, the write method associated with the type info is called.
-
-Attributes are used to customize the writing procedure. The standard types support the following
-attributes (all the dt attributes have the same format as in cvWriteRawData):
-
--# CvSeq
-    -   **header_dt** description of user fields of the sequence header that follow CvSeq, or
-        CvChain (if the sequence is a Freeman chain) or CvContour (if the sequence is a contour or
-        point sequence)
-    -   **dt** description of the sequence elements.
-    -   **recursive** if the attribute is present and is not equal to "0" or "false", the whole
-        tree of sequences (contours) is stored.
--# CvGraph
-    -   **header_dt** description of user fields of the graph header that follows CvGraph;
-    -   **vertex_dt** description of user fields of graph vertices
-    -   **edge_dt** description of user fields of graph edges (note that the edge weight is
-        always written, so there is no need to specify it explicitly)
-
-Below is the code that creates the YAML file shown in the CvFileStorage description:
-@code
-    #include "cxcore.h"
-
-    int main( int argc, char** argv )
-    {
-        CvMat* mat = cvCreateMat( 3, 3, CV_32F );
-        CvFileStorage* fs = cvOpenFileStorage( "example.yml", 0, CV_STORAGE_WRITE );
-
-        cvSetIdentity( mat );
-        cvWrite( fs, "A", mat, cvAttrList(0,0) );
-
-        cvReleaseFileStorage( &fs );
-        cvReleaseMat( &mat );
-        return 0;
-    }
-@endcode
-@param fs File storage
-@param name Name of the written object. Should be NULL if and only if the parent structure is a
-sequence.
-@param ptr Pointer to the object
-@param attributes The attributes of the object. They are specific for each particular type (see
-the discussion below).
- */
-CVAPI(void) cvWrite( CvFileStorage* fs, const char* name, const void* ptr,
-                         CvAttrList attributes CV_DEFAULT(cvAttrList()));
-
-/** @brief Starts the next stream.
-
-The function finishes the currently written stream and starts the next stream. In the case of XML
-the file with multiple streams looks like this:
-@code{.xml}
-    <opencv_storage>
-    <!-- stream #1 data -->
-    </opencv_storage>
-    <opencv_storage>
-    <!-- stream #2 data -->
-    </opencv_storage>
-    ...
-@endcode
-The YAML file will look like this:
-@code{.yaml}
-    %YAML 1.0
-    # stream #1 data
-    ...
-    ---
-    # stream #2 data
-@endcode
-This is useful for concatenating files or for resuming the writing process.
-@param fs File storage
- */
-CVAPI(void) cvStartNextStream( CvFileStorage* fs );
-
-/** @brief Writes multiple numbers.
-
-The function writes an array, whose elements consist of single or multiple numbers. The function
-call can be replaced with a loop containing a few cvWriteInt and cvWriteReal calls, but a single
-call is more efficient. Note that because none of the elements have a name, they should be written
-to a sequence rather than a map.
-@param fs File storage
-@param src Pointer to the written array
-@param len Number of the array elements to write
-@param dt Specification of each array element, see @ref format_spec "format specification"
- */
-CVAPI(void) cvWriteRawData( CvFileStorage* fs, const void* src,
-                                int len, const char* dt );
-
-/** @brief Writes multiple numbers in Base64.
-
-If either CV_STORAGE_WRITE_BASE64 or cv::FileStorage::WRITE_BASE64 is used,
-this function will be the same as cvWriteRawData. If neither, the main
-difference is that it outputs a sequence in Base64 encoding rather than
-in plain text.
-
-This function can only be used to write a sequence with a type "binary".
-
-@param fs File storage
-@param src Pointer to the written array
-@param len Number of the array elements to write
-@param dt Specification of each array element, see @ref format_spec "format specification"
-*/
-CVAPI(void) cvWriteRawDataBase64( CvFileStorage* fs, const void* src,
-                                 int len, const char* dt );
-
-/** @brief Returns a unique pointer for a given name.
-
-The function returns a unique pointer for each particular file node name. This pointer can be then
-passed to the cvGetFileNode function that is faster than cvGetFileNodeByName because it compares
-text strings by comparing pointers rather than the strings' content.
-
-Consider the following example where an array of points is encoded as a sequence of 2-entry maps:
-@code
-    points:
-      - { x: 10, y: 10 }
-      - { x: 20, y: 20 }
-      - { x: 30, y: 30 }
-      # ...
-@endcode
-Then, it is possible to get hashed "x" and "y" pointers to speed up decoding of the points. :
-@code
-    #include "cxcore.h"
-
-    int main( int argc, char** argv )
-    {
-        CvFileStorage* fs = cvOpenFileStorage( "points.yml", 0, CV_STORAGE_READ );
-        CvStringHashNode* x_key = cvGetHashedNode( fs, "x", -1, 1 );
-        CvStringHashNode* y_key = cvGetHashedNode( fs, "y", -1, 1 );
-        CvFileNode* points = cvGetFileNodeByName( fs, 0, "points" );
-
-        if( CV_NODE_IS_SEQ(points->tag) )
-        {
-            CvSeq* seq = points->data.seq;
-            int i, total = seq->total;
-            CvSeqReader reader;
-            cvStartReadSeq( seq, &reader, 0 );
-            for( i = 0; i < total; i++ )
-            {
-                CvFileNode* pt = (CvFileNode*)reader.ptr;
-    #if 1 // faster variant
-                CvFileNode* xnode = cvGetFileNode( fs, pt, x_key, 0 );
-                CvFileNode* ynode = cvGetFileNode( fs, pt, y_key, 0 );
-                assert( xnode && CV_NODE_IS_INT(xnode->tag) &&
-                        ynode && CV_NODE_IS_INT(ynode->tag));
-                int x = xnode->data.i; // or x = cvReadInt( xnode, 0 );
-                int y = ynode->data.i; // or y = cvReadInt( ynode, 0 );
-    #elif 1 // slower variant; does not use x_key & y_key
-                CvFileNode* xnode = cvGetFileNodeByName( fs, pt, "x" );
-                CvFileNode* ynode = cvGetFileNodeByName( fs, pt, "y" );
-                assert( xnode && CV_NODE_IS_INT(xnode->tag) &&
-                        ynode && CV_NODE_IS_INT(ynode->tag));
-                int x = xnode->data.i; // or x = cvReadInt( xnode, 0 );
-                int y = ynode->data.i; // or y = cvReadInt( ynode, 0 );
-    #else // the slowest yet the easiest to use variant
-                int x = cvReadIntByName( fs, pt, "x", 0 );
-                int y = cvReadIntByName( fs, pt, "y", 0 );
-    #endif
-                CV_NEXT_SEQ_ELEM( seq->elem_size, reader );
-                printf("
-            }
-        }
-        cvReleaseFileStorage( &fs );
-        return 0;
-    }
-@endcode
-Please note that whatever method of accessing a map you are using, it is still much slower than
-using plain sequences; for example, in the above example, it is more efficient to encode the points
-as pairs of integers in a single numeric sequence.
-@param fs File storage
-@param name Literal node name
-@param len Length of the name (if it is known apriori), or -1 if it needs to be calculated
-@param create_missing Flag that specifies, whether an absent key should be added into the hash table
-*/
-CVAPI(CvStringHashNode*) cvGetHashedKey( CvFileStorage* fs, const char* name,
-                                        int len CV_DEFAULT(-1),
-                                        int create_missing CV_DEFAULT(0));
-
-/** @brief Retrieves one of the top-level nodes of the file storage.
-
-The function returns one of the top-level file nodes. The top-level nodes do not have a name, they
-correspond to the streams that are stored one after another in the file storage. If the index is out
-of range, the function returns a NULL pointer, so all the top-level nodes can be iterated by
-subsequent calls to the function with stream_index=0,1,..., until the NULL pointer is returned.
-This function can be used as a base for recursive traversal of the file storage.
-@param fs File storage
-@param stream_index Zero-based index of the stream. See cvStartNextStream . In most cases,
-there is only one stream in the file; however, there can be several.
- */
-CVAPI(CvFileNode*) cvGetRootFileNode( const CvFileStorage* fs,
-                                     int stream_index CV_DEFAULT(0) );
-
-/** @brief Finds a node in a map or file storage.
-
-The function finds a file node. It is a faster version of cvGetFileNodeByName (see
-cvGetHashedKey discussion). Also, the function can insert a new node, if it is not in the map yet.
-@param fs File storage
-@param map The parent map. If it is NULL, the function searches a top-level node. If both map and
-key are NULLs, the function returns the root file node - a map that contains top-level nodes.
-@param key Unique pointer to the node name, retrieved with cvGetHashedKey
-@param create_missing Flag that specifies whether an absent node should be added to the map
- */
-CVAPI(CvFileNode*) cvGetFileNode( CvFileStorage* fs, CvFileNode* map,
-                                 const CvStringHashNode* key,
-                                 int create_missing CV_DEFAULT(0) );
-
-/** @brief Finds a node in a map or file storage.
-
-The function finds a file node by name. The node is searched either in map or, if the pointer is
-NULL, among the top-level file storage nodes. Using this function for maps and cvGetSeqElem (or
-sequence reader) for sequences, it is possible to navigate through the file storage. To speed up
-multiple queries for a certain key (e.g., in the case of an array of structures) one may use a
-combination of cvGetHashedKey and cvGetFileNode.
-@param fs File storage
-@param map The parent map. If it is NULL, the function searches in all the top-level nodes
-(streams), starting with the first one.
-@param name The file node name
- */
-CVAPI(CvFileNode*) cvGetFileNodeByName( const CvFileStorage* fs,
-                                       const CvFileNode* map,
-                                       const char* name );
-
-/** @brief Retrieves an integer value from a file node.
-
-The function returns an integer that is represented by the file node. If the file node is NULL, the
-default_value is returned (thus, it is convenient to call the function right after cvGetFileNode
-without checking for a NULL pointer). If the file node has type CV_NODE_INT, then node-\>data.i is
-returned. If the file node has type CV_NODE_REAL, then node-\>data.f is converted to an integer
-and returned. Otherwise the error is reported.
-@param node File node
-@param default_value The value that is returned if node is NULL
- */
-CV_INLINE int cvReadInt( const CvFileNode* node, int default_value CV_DEFAULT(0) )
-{
-    return !node ? default_value :
-        CV_NODE_IS_INT(node->tag) ? node->data.i :
-        CV_NODE_IS_REAL(node->tag) ? cvRound(node->data.f) : 0x7fffffff;
-}
-
-/** @brief Finds a file node and returns its value.
-
-The function is a simple superposition of cvGetFileNodeByName and cvReadInt.
-@param fs File storage
-@param map The parent map. If it is NULL, the function searches a top-level node.
-@param name The node name
-@param default_value The value that is returned if the file node is not found
- */
-CV_INLINE int cvReadIntByName( const CvFileStorage* fs, const CvFileNode* map,
-                         const char* name, int default_value CV_DEFAULT(0) )
-{
-    return cvReadInt( cvGetFileNodeByName( fs, map, name ), default_value );
-}
-
-/** @brief Retrieves a floating-point value from a file node.
-
-The function returns a floating-point value that is represented by the file node. If the file node
-is NULL, the default_value is returned (thus, it is convenient to call the function right after
-cvGetFileNode without checking for a NULL pointer). If the file node has type CV_NODE_REAL ,
-then node-\>data.f is returned. If the file node has type CV_NODE_INT , then node-:math:\>data.f
-is converted to floating-point and returned. Otherwise the result is not determined.
-@param node File node
-@param default_value The value that is returned if node is NULL
- */
-CV_INLINE double cvReadReal( const CvFileNode* node, double default_value CV_DEFAULT(0.) )
-{
-    return !node ? default_value :
-        CV_NODE_IS_INT(node->tag) ? (double)node->data.i :
-        CV_NODE_IS_REAL(node->tag) ? node->data.f : 1e300;
-}
-
-/** @brief Finds a file node and returns its value.
-
-The function is a simple superposition of cvGetFileNodeByName and cvReadReal .
-@param fs File storage
-@param map The parent map. If it is NULL, the function searches a top-level node.
-@param name The node name
-@param default_value The value that is returned if the file node is not found
- */
-CV_INLINE double cvReadRealByName( const CvFileStorage* fs, const CvFileNode* map,
-                        const char* name, double default_value CV_DEFAULT(0.) )
-{
-    return cvReadReal( cvGetFileNodeByName( fs, map, name ), default_value );
-}
-
-/** @brief Retrieves a text string from a file node.
-
-The function returns a text string that is represented by the file node. If the file node is NULL,
-the default_value is returned (thus, it is convenient to call the function right after
-cvGetFileNode without checking for a NULL pointer). If the file node has type CV_NODE_STR , then
-node-:math:\>data.str.ptr is returned. Otherwise the result is not determined.
-@param node File node
-@param default_value The value that is returned if node is NULL
- */
-CV_INLINE const char* cvReadString( const CvFileNode* node,
-                        const char* default_value CV_DEFAULT(NULL) )
-{
-    return !node ? default_value : CV_NODE_IS_STRING(node->tag) ? node->data.str.ptr : 0;
-}
-
-/** @brief Finds a file node by its name and returns its value.
-
-The function is a simple superposition of cvGetFileNodeByName and cvReadString .
-@param fs File storage
-@param map The parent map. If it is NULL, the function searches a top-level node.
-@param name The node name
-@param default_value The value that is returned if the file node is not found
- */
-CV_INLINE const char* cvReadStringByName( const CvFileStorage* fs, const CvFileNode* map,
-                        const char* name, const char* default_value CV_DEFAULT(NULL) )
-{
-    return cvReadString( cvGetFileNodeByName( fs, map, name ), default_value );
-}
-
-
-/** @brief Decodes an object and returns a pointer to it.
-
-The function decodes a user object (creates an object in a native representation from the file
-storage subtree) and returns it. The object to be decoded must be an instance of a registered type
-that supports the read method (see CvTypeInfo). The type of the object is determined by the type
-name that is encoded in the file. If the object is a dynamic structure, it is created either in
-memory storage and passed to cvOpenFileStorage or, if a NULL pointer was passed, in temporary
-memory storage, which is released when cvReleaseFileStorage is called. Otherwise, if the object is
-not a dynamic structure, it is created in a heap and should be released with a specialized function
-or by using the generic cvRelease.
-@param fs File storage
-@param node The root object node
-@param attributes Unused parameter
- */
-CVAPI(void*) cvRead( CvFileStorage* fs, CvFileNode* node,
-                        CvAttrList* attributes CV_DEFAULT(NULL));
-
-/** @brief Finds an object by name and decodes it.
-
-The function is a simple superposition of cvGetFileNodeByName and cvRead.
-@param fs File storage
-@param map The parent map. If it is NULL, the function searches a top-level node.
-@param name The node name
-@param attributes Unused parameter
- */
-CV_INLINE void* cvReadByName( CvFileStorage* fs, const CvFileNode* map,
-                              const char* name, CvAttrList* attributes CV_DEFAULT(NULL) )
-{
-    return cvRead( fs, cvGetFileNodeByName( fs, map, name ), attributes );
-}
-
-
-/** @brief Initializes the file node sequence reader.
-
-The function initializes the sequence reader to read data from a file node. The initialized reader
-can be then passed to cvReadRawDataSlice.
-@param fs File storage
-@param src The file node (a sequence) to read numbers from
-@param reader Pointer to the sequence reader
- */
-CVAPI(void) cvStartReadRawData( const CvFileStorage* fs, const CvFileNode* src,
-                               CvSeqReader* reader );
-
-/** @brief Initializes file node sequence reader.
-
-The function reads one or more elements from the file node, representing a sequence, to a
-user-specified array. The total number of read sequence elements is a product of total and the
-number of components in each array element. For example, if dt=2if, the function will read total\*3
-sequence elements. As with any sequence, some parts of the file node sequence can be skipped or read
-repeatedly by repositioning the reader using cvSetSeqReaderPos.
-@param fs File storage
-@param reader The sequence reader. Initialize it with cvStartReadRawData .
-@param count The number of elements to read
-@param dst Pointer to the destination array
-@param dt Specification of each array element. It has the same format as in cvWriteRawData .
- */
-CVAPI(void) cvReadRawDataSlice( const CvFileStorage* fs, CvSeqReader* reader,
-                               int count, void* dst, const char* dt );
-
-/** @brief Reads multiple numbers.
-
-The function reads elements from a file node that represents a sequence of scalars.
-@param fs File storage
-@param src The file node (a sequence) to read numbers from
-@param dst Pointer to the destination array
-@param dt Specification of each array element. It has the same format as in cvWriteRawData .
- */
-CVAPI(void) cvReadRawData( const CvFileStorage* fs, const CvFileNode* src,
-                          void* dst, const char* dt );
-
-/** @brief Writes a file node to another file storage.
-
-The function writes a copy of a file node to file storage. Possible applications of the function are
-merging several file storages into one and conversion between XML, YAML and JSON formats.
-@param fs Destination file storage
-@param new_node_name New name of the file node in the destination file storage. To keep the
-existing name, use cvcvGetFileNodeName
-@param node The written node
-@param embed If the written node is a collection and this parameter is not zero, no extra level of
-hierarchy is created. Instead, all the elements of node are written into the currently written
-structure. Of course, map elements can only be embedded into another map, and sequence elements
-can only be embedded into another sequence.
- */
-CVAPI(void) cvWriteFileNode( CvFileStorage* fs, const char* new_node_name,
-                            const CvFileNode* node, int embed );
-
-/** @brief Returns the name of a file node.
-
-The function returns the name of a file node or NULL, if the file node does not have a name or if
-node is NULL.
-@param node File node
- */
-CVAPI(const char*) cvGetFileNodeName( const CvFileNode* node );
-
-/*********************************** Adding own types ***********************************/
-
-/** @brief Registers a new type.
-
-The function registers a new type, which is described by info . The function creates a copy of the
-structure, so the user should delete it after calling the function.
-@param info Type info structure
- */
-CVAPI(void) cvRegisterType( const CvTypeInfo* info );
-
-/** @brief Unregisters the type.
-
-The function unregisters a type with a specified name. If the name is unknown, it is possible to
-locate the type info by an instance of the type using cvTypeOf or by iterating the type list,
-starting from cvFirstType, and then calling cvUnregisterType(info-\>typeName).
-@param type_name Name of an unregistered type
- */
-CVAPI(void) cvUnregisterType( const char* type_name );
-
-/** @brief Returns the beginning of a type list.
-
-The function returns the first type in the list of registered types. Navigation through the list can
-be done via the prev and next fields of the CvTypeInfo structure.
- */
-CVAPI(CvTypeInfo*) cvFirstType(void);
-
-/** @brief Finds a type by its name.
-
-The function finds a registered type by its name. It returns NULL if there is no type with the
-specified name.
-@param type_name Type name
- */
-CVAPI(CvTypeInfo*) cvFindType( const char* type_name );
-
-/** @brief Returns the type of an object.
-
-The function finds the type of a given object. It iterates through the list of registered types and
-calls the is_instance function/method for every type info structure with that object until one of
-them returns non-zero or until the whole list has been traversed. In the latter case, the function
-returns NULL.
-@param struct_ptr The object pointer
- */
-CVAPI(CvTypeInfo*) cvTypeOf( const void* struct_ptr );
-
-#endif
-
-/** @brief Releases an object.
-
- The function finds the type of a given object and calls release with the double pointer.
- @param struct_ptr Double pointer to the object
- */
-CVAPI(void) cvRelease( void** struct_ptr );
-
-/** @brief Makes a clone of an object.
-
-The function finds the type of a given object and calls clone with the passed object. Of course, if
-you know the object type, for example, struct_ptr is CvMat\*, it is faster to call the specific
-function, like cvCloneMat.
-@param struct_ptr The object to clone
- */
-CVAPI(void*) cvClone( const void* struct_ptr );
-
-/*********************************** Measuring Execution Time ***************************/
-
-/** helper functions for RNG initialization and accurate time measurement:
-   uses internal clock counter on x86 */
-CVAPI(int64)  cvGetTickCount( void );
-CVAPI(double) cvGetTickFrequency( void );
-
-/*********************************** CPU capabilities ***********************************/
-
-CVAPI(int) cvCheckHardwareSupport(int feature);
-
-/*********************************** Multi-Threading ************************************/
-
-/** retrieve/set the number of threads used in OpenMP implementations */
-CVAPI(int)  cvGetNumThreads( void );
-CVAPI(void) cvSetNumThreads( int threads CV_DEFAULT(0) );
-/** get index of the thread being executed */
-CVAPI(int)  cvGetThreadNum( void );
-
-
-/********************************** Error Handling **************************************/
-
-/** Get current OpenCV error status */
-CVAPI(int) cvGetErrStatus( void );
-
-/** Sets error status silently */
-CVAPI(void) cvSetErrStatus( int status );
-
-#define CV_ErrModeLeaf     0   /* Print error and exit program */
-#define CV_ErrModeParent   1   /* Print error and continue */
-#define CV_ErrModeSilent   2   /* Don't print and continue */
-
-/** Retrieves current error processing mode */
-CVAPI(int)  cvGetErrMode( void );
-
-/** Sets error processing mode, returns previously used mode */
-CVAPI(int) cvSetErrMode( int mode );
-
-/** Sets error status and performs some additional actions (displaying message box,
- writing message to stderr, terminating application etc.)
- depending on the current error mode */
-CVAPI(void) cvError( int status, const char* func_name,
-                    const char* err_msg, const char* file_name, int line );
-
-/** Retrieves textual description of the error given its code */
-CVAPI(const char*) cvErrorStr( int status );
-
-/** Retrieves detailed information about the last error occurred */
-CVAPI(int) cvGetErrInfo( const char** errcode_desc, const char** description,
-                        const char** filename, int* line );
-
-/** Maps IPP error codes to the counterparts from OpenCV */
-CVAPI(int) cvErrorFromIppStatus( int ipp_status );
-
-typedef int (CV_CDECL *CvErrorCallback)( int status, const char* func_name,
-                                        const char* err_msg, const char* file_name, int line, void* userdata );
-
-/** Assigns a new error-handling function */
-CVAPI(CvErrorCallback) cvRedirectError( CvErrorCallback error_handler,
-                                       void* userdata CV_DEFAULT(NULL),
-                                       void** prev_userdata CV_DEFAULT(NULL) );
-
-/** Output nothing */
-CVAPI(int) cvNulDevReport( int status, const char* func_name, const char* err_msg,
-                          const char* file_name, int line, void* userdata );
-
-/** Output to console(fprintf(stderr,...)) */
-CVAPI(int) cvStdErrReport( int status, const char* func_name, const char* err_msg,
-                          const char* file_name, int line, void* userdata );
-
-/** Output to MessageBox(WIN32) */
-CVAPI(int) cvGuiBoxReport( int status, const char* func_name, const char* err_msg,
-                          const char* file_name, int line, void* userdata );
-
-#define OPENCV_ERROR(status,func,context)                           \
-cvError((status),(func),(context),__FILE__,__LINE__)
-
-#define OPENCV_ASSERT(expr,func,context)                            \
-{if (! (expr))                                      \
-{OPENCV_ERROR(CV_StsInternal,(func),(context));}}
-
-#define OPENCV_CALL( Func )                                         \
-{                                                                   \
-Func;                                                           \
-}
-
-
-/** CV_FUNCNAME macro defines icvFuncName constant which is used by CV_ERROR macro */
-#ifdef CV_NO_FUNC_NAMES
-#define CV_FUNCNAME( Name )
-#define cvFuncName ""
-#else
-#define CV_FUNCNAME( Name )  \
-static char cvFuncName[] = Name
-#endif
-
-
-/**
- CV_ERROR macro unconditionally raises error with passed code and message.
- After raising error, control will be transferred to the exit label.
- */
-#define CV_ERROR( Code, Msg )                                       \
-{                                                                   \
-    cvError( (Code), cvFuncName, Msg, __FILE__, __LINE__ );        \
-    __CV_EXIT__;                                                   \
-}
-
-/**
- CV_CHECK macro checks error status after CV (or IPL)
- function call. If error detected, control will be transferred to the exit
- label.
- */
-#define CV_CHECK()                                                  \
-{                                                                   \
-    if( cvGetErrStatus() < 0 )                                      \
-        CV_ERROR( CV_StsBackTrace, "Inner function failed." );      \
-}
-
-
-/**
- CV_CALL macro calls CV (or IPL) function, checks error status and
- signals a error if the function failed. Useful in "parent node"
- error processing mode
- */
-#define CV_CALL( Func )                                             \
-{                                                                   \
-    Func;                                                           \
-    CV_CHECK();                                                     \
-}
-
-
-/** Runtime assertion macro */
-#define CV_ASSERT( Condition )                                          \
-{                                                                       \
-    if( !(Condition) )                                                  \
-        CV_ERROR( CV_StsInternal, "Assertion: " #Condition " failed" ); \
-}
-
-#define __CV_BEGIN__       {
-#define __CV_END__         goto exit; exit: ; }
-#define __CV_EXIT__        goto exit
-
-/** @} core_c */
-
-#ifdef __cplusplus
-} // extern "C"
-#endif
-
-#ifdef __cplusplus
-
-#include "opencv2/core/utility.hpp"
-
-namespace cv
-{
-
-//! @addtogroup core_c_glue
-//! @{
-
-/////////////////////////////////////////// glue ///////////////////////////////////////////
-
-//! converts array (CvMat or IplImage) to cv::Mat
-CV_EXPORTS Mat cvarrToMat(const CvArr* arr, bool copyData=false,
-                          bool allowND=true, int coiMode=0,
-                          AutoBuffer<double>* buf=0);
-
-static inline Mat cvarrToMatND(const CvArr* arr, bool copyData=false, int coiMode=0)
-{
-    return cvarrToMat(arr, copyData, true, coiMode);
-}
-
-
-//! extracts Channel of Interest from CvMat or IplImage and makes cv::Mat out of it.
-CV_EXPORTS void extractImageCOI(const CvArr* arr, OutputArray coiimg, int coi=-1);
-//! inserts single-channel cv::Mat into a multi-channel CvMat or IplImage
-CV_EXPORTS void insertImageCOI(InputArray coiimg, CvArr* arr, int coi=-1);
-
-
-
-////// specialized implementations of DefaultDeleter::operator() for classic OpenCV types //////
-
-template<> struct DefaultDeleter<CvMat>{ CV_EXPORTS void operator ()(CvMat* obj) const; };
-template<> struct DefaultDeleter<IplImage>{ CV_EXPORTS void operator ()(IplImage* obj) const; };
-template<> struct DefaultDeleter<CvMatND>{ CV_EXPORTS void operator ()(CvMatND* obj) const; };
-template<> struct DefaultDeleter<CvSparseMat>{ CV_EXPORTS void operator ()(CvSparseMat* obj) const; };
-template<> struct DefaultDeleter<CvMemStorage>{ CV_EXPORTS void operator ()(CvMemStorage* obj) const; };
-
-////////////// convenient wrappers for operating old-style dynamic structures //////////////
-
-template<typename _Tp> class SeqIterator;
-
-typedef Ptr<CvMemStorage> MemStorage;
-
-/*!
- Template Sequence Class derived from CvSeq
-
- The class provides more convenient access to sequence elements,
- STL-style operations and iterators.
-
- \note The class is targeted for simple data types,
-    i.e. no constructors or destructors
-    are called for the sequence elements.
-*/
-template<typename _Tp> class Seq
-{
-public:
-    typedef SeqIterator<_Tp> iterator;
-    typedef SeqIterator<_Tp> const_iterator;
-
-    //! the default constructor
-    Seq();
-    //! the constructor for wrapping CvSeq structure. The real element type in CvSeq should match _Tp.
-    Seq(const CvSeq* seq);
-    //! creates the empty sequence that resides in the specified storage
-    Seq(MemStorage& storage, int headerSize = sizeof(CvSeq));
-    //! returns read-write reference to the specified element
-    _Tp& operator [](int idx);
-    //! returns read-only reference to the specified element
-    const _Tp& operator[](int idx) const;
-    //! returns iterator pointing to the beginning of the sequence
-    SeqIterator<_Tp> begin() const;
-    //! returns iterator pointing to the element following the last sequence element
-    SeqIterator<_Tp> end() const;
-    //! returns the number of elements in the sequence
-    size_t size() const;
-    //! returns the type of sequence elements (CV_8UC1 ... CV_64FC(CV_CN_MAX) ...)
-    int type() const;
-    //! returns the depth of sequence elements (CV_8U ... CV_64F)
-    int depth() const;
-    //! returns the number of channels in each sequence element
-    int channels() const;
-    //! returns the size of each sequence element
-    size_t elemSize() const;
-    //! returns index of the specified sequence element
-    size_t index(const _Tp& elem) const;
-    //! appends the specified element to the end of the sequence
-    void push_back(const _Tp& elem);
-    //! appends the specified element to the front of the sequence
-    void push_front(const _Tp& elem);
-    //! appends zero or more elements to the end of the sequence
-    void push_back(const _Tp* elems, size_t count);
-    //! appends zero or more elements to the front of the sequence
-    void push_front(const _Tp* elems, size_t count);
-    //! inserts the specified element to the specified position
-    void insert(int idx, const _Tp& elem);
-    //! inserts zero or more elements to the specified position
-    void insert(int idx, const _Tp* elems, size_t count);
-    //! removes element at the specified position
-    void remove(int idx);
-    //! removes the specified subsequence
-    void remove(const Range& r);
-
-    //! returns reference to the first sequence element
-    _Tp& front();
-    //! returns read-only reference to the first sequence element
-    const _Tp& front() const;
-    //! returns reference to the last sequence element
-    _Tp& back();
-    //! returns read-only reference to the last sequence element
-    const _Tp& back() const;
-    //! returns true iff the sequence contains no elements
-    bool empty() const;
-
-    //! removes all the elements from the sequence
-    void clear();
-    //! removes the first element from the sequence
-    void pop_front();
-    //! removes the last element from the sequence
-    void pop_back();
-    //! removes zero or more elements from the beginning of the sequence
-    void pop_front(_Tp* elems, size_t count);
-    //! removes zero or more elements from the end of the sequence
-    void pop_back(_Tp* elems, size_t count);
-
-    //! copies the whole sequence or the sequence slice to the specified vector
-    void copyTo(std::vector<_Tp>& vec, const Range& range=Range::all()) const;
-    //! returns the vector containing all the sequence elements
-    operator std::vector<_Tp>() const;
-
-    CvSeq* seq;
-};
-
-
-/*!
- STL-style Sequence Iterator inherited from the CvSeqReader structure
-*/
-template<typename _Tp> class SeqIterator : public CvSeqReader
-{
-public:
-    //! the default constructor
-    SeqIterator();
-    //! the constructor setting the iterator to the beginning or to the end of the sequence
-    SeqIterator(const Seq<_Tp>& seq, bool seekEnd=false);
-    //! positions the iterator within the sequence
-    void seek(size_t pos);
-    //! reports the current iterator position
-    size_t tell() const;
-    //! returns reference to the current sequence element
-    _Tp& operator *();
-    //! returns read-only reference to the current sequence element
-    const _Tp& operator *() const;
-    //! moves iterator to the next sequence element
-    SeqIterator& operator ++();
-    //! moves iterator to the next sequence element
-    SeqIterator operator ++(int) const;
-    //! moves iterator to the previous sequence element
-    SeqIterator& operator --();
-    //! moves iterator to the previous sequence element
-    SeqIterator operator --(int) const;
-
-    //! moves iterator forward by the specified offset (possibly negative)
-    SeqIterator& operator +=(int);
-    //! moves iterator backward by the specified offset (possibly negative)
-    SeqIterator& operator -=(int);
-
-    // this is index of the current element module seq->total*2
-    // (to distinguish between 0 and seq->total)
-    int index;
-};
-
-
-
-// bridge C++ => C Seq API
-CV_EXPORTS schar*  seqPush( CvSeq* seq, const void* element=0);
-CV_EXPORTS schar*  seqPushFront( CvSeq* seq, const void* element=0);
-CV_EXPORTS void  seqPop( CvSeq* seq, void* element=0);
-CV_EXPORTS void  seqPopFront( CvSeq* seq, void* element=0);
-CV_EXPORTS void  seqPopMulti( CvSeq* seq, void* elements,
-                              int count, int in_front=0 );
-CV_EXPORTS void  seqRemove( CvSeq* seq, int index );
-CV_EXPORTS void  clearSeq( CvSeq* seq );
-CV_EXPORTS schar*  getSeqElem( const CvSeq* seq, int index );
-CV_EXPORTS void  seqRemoveSlice( CvSeq* seq, CvSlice slice );
-CV_EXPORTS void  seqInsertSlice( CvSeq* seq, int before_index, const CvArr* from_arr );
-
-template<typename _Tp> inline Seq<_Tp>::Seq() : seq(0) {}
-template<typename _Tp> inline Seq<_Tp>::Seq( const CvSeq* _seq ) : seq((CvSeq*)_seq)
-{
-    CV_Assert(!_seq || _seq->elem_size == sizeof(_Tp));
-}
-
-template<typename _Tp> inline Seq<_Tp>::Seq( MemStorage& storage,
-                                             int headerSize )
-{
-    CV_Assert(headerSize >= (int)sizeof(CvSeq));
-    seq = cvCreateSeq(DataType<_Tp>::type, headerSize, sizeof(_Tp), storage);
-}
-
-template<typename _Tp> inline _Tp& Seq<_Tp>::operator [](int idx)
-{ return *(_Tp*)getSeqElem(seq, idx); }
-
-template<typename _Tp> inline const _Tp& Seq<_Tp>::operator [](int idx) const
-{ return *(_Tp*)getSeqElem(seq, idx); }
-
-template<typename _Tp> inline SeqIterator<_Tp> Seq<_Tp>::begin() const
-{ return SeqIterator<_Tp>(*this); }
-
-template<typename _Tp> inline SeqIterator<_Tp> Seq<_Tp>::end() const
-{ return SeqIterator<_Tp>(*this, true); }
-
-template<typename _Tp> inline size_t Seq<_Tp>::size() const
-{ return seq ? seq->total : 0; }
-
-template<typename _Tp> inline int Seq<_Tp>::type() const
-{ return seq ? CV_MAT_TYPE(seq->flags) : 0; }
-
-template<typename _Tp> inline int Seq<_Tp>::depth() const
-{ return seq ? CV_MAT_DEPTH(seq->flags) : 0; }
-
-template<typename _Tp> inline int Seq<_Tp>::channels() const
-{ return seq ? CV_MAT_CN(seq->flags) : 0; }
-
-template<typename _Tp> inline size_t Seq<_Tp>::elemSize() const
-{ return seq ? seq->elem_size : 0; }
-
-template<typename _Tp> inline size_t Seq<_Tp>::index(const _Tp& elem) const
-{ return cvSeqElemIdx(seq, &elem); }
-
-template<typename _Tp> inline void Seq<_Tp>::push_back(const _Tp& elem)
-{ cvSeqPush(seq, &elem); }
-
-template<typename _Tp> inline void Seq<_Tp>::push_front(const _Tp& elem)
-{ cvSeqPushFront(seq, &elem); }
-
-template<typename _Tp> inline void Seq<_Tp>::push_back(const _Tp* elem, size_t count)
-{ cvSeqPushMulti(seq, elem, (int)count, 0); }
-
-template<typename _Tp> inline void Seq<_Tp>::push_front(const _Tp* elem, size_t count)
-{ cvSeqPushMulti(seq, elem, (int)count, 1); }
-
-template<typename _Tp> inline _Tp& Seq<_Tp>::back()
-{ return *(_Tp*)getSeqElem(seq, -1); }
-
-template<typename _Tp> inline const _Tp& Seq<_Tp>::back() const
-{ return *(const _Tp*)getSeqElem(seq, -1); }
-
-template<typename _Tp> inline _Tp& Seq<_Tp>::front()
-{ return *(_Tp*)getSeqElem(seq, 0); }
-
-template<typename _Tp> inline const _Tp& Seq<_Tp>::front() const
-{ return *(const _Tp*)getSeqElem(seq, 0); }
-
-template<typename _Tp> inline bool Seq<_Tp>::empty() const
-{ return !seq || seq->total == 0; }
-
-template<typename _Tp> inline void Seq<_Tp>::clear()
-{ if(seq) clearSeq(seq); }
-
-template<typename _Tp> inline void Seq<_Tp>::pop_back()
-{ seqPop(seq); }
-
-template<typename _Tp> inline void Seq<_Tp>::pop_front()
-{ seqPopFront(seq); }
-
-template<typename _Tp> inline void Seq<_Tp>::pop_back(_Tp* elem, size_t count)
-{ seqPopMulti(seq, elem, (int)count, 0); }
-
-template<typename _Tp> inline void Seq<_Tp>::pop_front(_Tp* elem, size_t count)
-{ seqPopMulti(seq, elem, (int)count, 1); }
-
-template<typename _Tp> inline void Seq<_Tp>::insert(int idx, const _Tp& elem)
-{ seqInsert(seq, idx, &elem); }
-
-template<typename _Tp> inline void Seq<_Tp>::insert(int idx, const _Tp* elems, size_t count)
-{
-    CvMat m = cvMat(1, count, DataType<_Tp>::type, elems);
-    seqInsertSlice(seq, idx, &m);
-}
-
-template<typename _Tp> inline void Seq<_Tp>::remove(int idx)
-{ seqRemove(seq, idx); }
-
-template<typename _Tp> inline void Seq<_Tp>::remove(const Range& r)
-{ seqRemoveSlice(seq, cvSlice(r.start, r.end)); }
-
-template<typename _Tp> inline void Seq<_Tp>::copyTo(std::vector<_Tp>& vec, const Range& range) const
-{
-    size_t len = !seq ? 0 : range == Range::all() ? seq->total : range.end - range.start;
-    vec.resize(len);
-    if( seq && len )
-        cvCvtSeqToArray(seq, &vec[0], cvSlice(range));
-}
-
-template<typename _Tp> inline Seq<_Tp>::operator std::vector<_Tp>() const
-{
-    std::vector<_Tp> vec;
-    copyTo(vec);
-    return vec;
-}
-
-template<typename _Tp> inline SeqIterator<_Tp>::SeqIterator()
-{ memset(this, 0, sizeof(*this)); }
-
-template<typename _Tp> inline SeqIterator<_Tp>::SeqIterator(const Seq<_Tp>& _seq, bool seekEnd)
-{
-    cvStartReadSeq(_seq.seq, this);
-    index = seekEnd ? _seq.seq->total : 0;
-}
-
-template<typename _Tp> inline void SeqIterator<_Tp>::seek(size_t pos)
-{
-    cvSetSeqReaderPos(this, (int)pos, false);
-    index = pos;
-}
-
-template<typename _Tp> inline size_t SeqIterator<_Tp>::tell() const
-{ return index; }
-
-template<typename _Tp> inline _Tp& SeqIterator<_Tp>::operator *()
-{ return *(_Tp*)ptr; }
-
-template<typename _Tp> inline const _Tp& SeqIterator<_Tp>::operator *() const
-{ return *(const _Tp*)ptr; }
-
-template<typename _Tp> inline SeqIterator<_Tp>& SeqIterator<_Tp>::operator ++()
-{
-    CV_NEXT_SEQ_ELEM(sizeof(_Tp), *this);
-    if( ++index >= seq->total*2 )
-        index = 0;
-    return *this;
-}
-
-template<typename _Tp> inline SeqIterator<_Tp> SeqIterator<_Tp>::operator ++(int) const
-{
-    SeqIterator<_Tp> it = *this;
-    ++*this;
-    return it;
-}
-
-template<typename _Tp> inline SeqIterator<_Tp>& SeqIterator<_Tp>::operator --()
-{
-    CV_PREV_SEQ_ELEM(sizeof(_Tp), *this);
-    if( --index < 0 )
-        index = seq->total*2-1;
-    return *this;
-}
-
-template<typename _Tp> inline SeqIterator<_Tp> SeqIterator<_Tp>::operator --(int) const
-{
-    SeqIterator<_Tp> it = *this;
-    --*this;
-    return it;
-}
-
-template<typename _Tp> inline SeqIterator<_Tp>& SeqIterator<_Tp>::operator +=(int delta)
-{
-    cvSetSeqReaderPos(this, delta, 1);
-    index += delta;
-    int n = seq->total*2;
-    if( index < 0 )
-        index += n;
-    if( index >= n )
-        index -= n;
-    return *this;
-}
-
-template<typename _Tp> inline SeqIterator<_Tp>& SeqIterator<_Tp>::operator -=(int delta)
-{
-    return (*this += -delta);
-}
-
-template<typename _Tp> inline ptrdiff_t operator - (const SeqIterator<_Tp>& a,
-                                                    const SeqIterator<_Tp>& b)
-{
-    ptrdiff_t delta = a.index - b.index, n = a.seq->total;
-    if( delta > n || delta < -n )
-        delta += delta < 0 ? n : -n;
-    return delta;
-}
-
-template<typename _Tp> inline bool operator == (const SeqIterator<_Tp>& a,
-                                                const SeqIterator<_Tp>& b)
-{
-    return a.seq == b.seq && a.index == b.index;
-}
-
-template<typename _Tp> inline bool operator != (const SeqIterator<_Tp>& a,
-                                                const SeqIterator<_Tp>& b)
-{
-    return !(a == b);
-}
-
-//! @}
-
-} // cv
-
-#endif
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda.hpp
deleted file mode 100644
index 716b8bf2..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda.hpp
+++ /dev/null
@@ -1,1270 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_CUDA_HPP
-#define OPENCV_CORE_CUDA_HPP
-
-#ifndef __cplusplus
-#  error cuda.hpp header must be compiled as C++
-#endif
-
-#include "opencv2/core.hpp"
-#include "opencv2/core/cuda_types.hpp"
-
-/**
-  @defgroup cuda CUDA-accelerated Computer Vision
-  @{
-    @defgroup cudacore Core part
-    @{
-      @defgroup cudacore_init Initialization and Information
-      @defgroup cudacore_struct Data Structures
-    @}
-  @}
- */
-
-namespace cv { namespace cuda {
-
-//! @addtogroup cudacore_struct
-//! @{
-
-//===================================================================================
-// GpuMat
-//===================================================================================
-
-/** @brief Base storage class for GPU memory with reference counting.
-
-Its interface matches the Mat interface with the following limitations:
-
--   no arbitrary dimensions support (only 2D)
--   no functions that return references to their data (because references on GPU are not valid for
-    CPU)
--   no expression templates technique support
-
-Beware that the latter limitation may lead to overloaded matrix operators that cause memory
-allocations. The GpuMat class is convertible to cuda::PtrStepSz and cuda::PtrStep so it can be
-passed directly to the kernel.
-
-@note In contrast with Mat, in most cases GpuMat::isContinuous() == false . This means that rows are
-aligned to a size depending on the hardware. Single-row GpuMat is always a continuous matrix.
-
-@note You are not recommended to leave static or global GpuMat variables allocated, that is, to rely
-on its destructor. The destruction order of such variables and CUDA context is undefined. GPU memory
-release function returns error if the CUDA context has been destroyed before.
-
-Some member functions are described as a "Blocking Call" while some are described as a
-"Non-Blocking Call". Blocking functions are synchronous to host. It is guaranteed that the GPU
-operation is finished when the function returns. However, non-blocking functions are asynchronous to
-host. Those functions may return even if the GPU operation is not finished.
-
-Compared to their blocking counterpart, non-blocking functions accept Stream as an additional
-argument. If a non-default stream is passed, the GPU operation may overlap with operations in other
-streams.
-
-@sa Mat
- */
-class CV_EXPORTS_W GpuMat
-{
-public:
-    class CV_EXPORTS_W Allocator
-    {
-    public:
-        virtual ~Allocator() {}
-
-        // allocator must fill data, step and refcount fields
-        virtual bool allocate(GpuMat* mat, int rows, int cols, size_t elemSize) = 0;
-        virtual void free(GpuMat* mat) = 0;
-    };
-
-    //! default allocator
-    CV_WRAP static GpuMat::Allocator* defaultAllocator();
-    CV_WRAP static void setDefaultAllocator(GpuMat::Allocator* allocator);
-
-    //! default constructor
-    CV_WRAP explicit GpuMat(GpuMat::Allocator* allocator = GpuMat::defaultAllocator());
-
-    //! constructs GpuMat of the specified size and type
-    CV_WRAP GpuMat(int rows, int cols, int type, GpuMat::Allocator* allocator = GpuMat::defaultAllocator());
-    CV_WRAP GpuMat(Size size, int type, GpuMat::Allocator* allocator = GpuMat::defaultAllocator());
-
-    //! constructs GpuMat and fills it with the specified value _s
-    CV_WRAP GpuMat(int rows, int cols, int type, Scalar s, GpuMat::Allocator* allocator = GpuMat::defaultAllocator());
-    CV_WRAP GpuMat(Size size, int type, Scalar s, GpuMat::Allocator* allocator = GpuMat::defaultAllocator());
-
-    //! copy constructor
-    CV_WRAP GpuMat(const GpuMat& m);
-
-    //! constructor for GpuMat headers pointing to user-allocated data
-    GpuMat(int rows, int cols, int type, void* data, size_t step = Mat::AUTO_STEP);
-    GpuMat(Size size, int type, void* data, size_t step = Mat::AUTO_STEP);
-
-    //! creates a GpuMat header for a part of the bigger matrix
-    CV_WRAP GpuMat(const GpuMat& m, Range rowRange, Range colRange);
-    CV_WRAP GpuMat(const GpuMat& m, Rect roi);
-
-    //! builds GpuMat from host memory (Blocking call)
-    CV_WRAP explicit GpuMat(InputArray arr, GpuMat::Allocator* allocator = GpuMat::defaultAllocator());
-
-    //! destructor - calls release()
-    ~GpuMat();
-
-    //! assignment operators
-    GpuMat& operator =(const GpuMat& m);
-
-    //! allocates new GpuMat data unless the GpuMat already has specified size and type
-    CV_WRAP void create(int rows, int cols, int type);
-    CV_WRAP void create(Size size, int type);
-
-    //! decreases reference counter, deallocate the data when reference counter reaches 0
-    void release();
-
-    //! swaps with other smart pointer
-    CV_WRAP void swap(GpuMat& mat);
-
-    /** @brief Performs data upload to GpuMat (Blocking call)
-
-    This function copies data from host memory to device memory. As being a blocking call, it is
-    guaranteed that the copy operation is finished when this function returns.
-    */
-    CV_WRAP void upload(InputArray arr);
-
-    /** @brief Performs data upload to GpuMat (Non-Blocking call)
-
-    This function copies data from host memory to device memory. As being a non-blocking call, this
-    function may return even if the copy operation is not finished.
-
-    The copy operation may be overlapped with operations in other non-default streams if \p stream is
-    not the default stream and \p dst is HostMem allocated with HostMem::PAGE_LOCKED option.
-    */
-    CV_WRAP void upload(InputArray arr, Stream& stream);
-
-    /** @brief Performs data download from GpuMat (Blocking call)
-
-    This function copies data from device memory to host memory. As being a blocking call, it is
-    guaranteed that the copy operation is finished when this function returns.
-    */
-    CV_WRAP void download(OutputArray dst) const;
-
-    /** @brief Performs data download from GpuMat (Non-Blocking call)
-
-    This function copies data from device memory to host memory. As being a non-blocking call, this
-    function may return even if the copy operation is not finished.
-
-    The copy operation may be overlapped with operations in other non-default streams if \p stream is
-    not the default stream and \p dst is HostMem allocated with HostMem::PAGE_LOCKED option.
-    */
-    CV_WRAP void download(OutputArray dst, Stream& stream) const;
-
-    //! returns deep copy of the GpuMat, i.e. the data is copied
-    CV_WRAP GpuMat clone() const;
-
-    //! copies the GpuMat content to device memory (Blocking call)
-    CV_WRAP void copyTo(OutputArray dst) const;
-
-    //! copies the GpuMat content to device memory (Non-Blocking call)
-    CV_WRAP void copyTo(OutputArray dst, Stream& stream) const;
-
-    //! copies those GpuMat elements to "m" that are marked with non-zero mask elements (Blocking call)
-    CV_WRAP void copyTo(OutputArray dst, InputArray mask) const;
-
-    //! copies those GpuMat elements to "m" that are marked with non-zero mask elements (Non-Blocking call)
-    CV_WRAP void copyTo(OutputArray dst, InputArray mask, Stream& stream) const;
-
-    //! sets some of the GpuMat elements to s (Blocking call)
-    CV_WRAP GpuMat& setTo(Scalar s);
-
-    //! sets some of the GpuMat elements to s (Non-Blocking call)
-    CV_WRAP GpuMat& setTo(Scalar s, Stream& stream);
-
-    //! sets some of the GpuMat elements to s, according to the mask (Blocking call)
-    CV_WRAP GpuMat& setTo(Scalar s, InputArray mask);
-
-    //! sets some of the GpuMat elements to s, according to the mask (Non-Blocking call)
-    CV_WRAP GpuMat& setTo(Scalar s, InputArray mask, Stream& stream);
-
-    //! converts GpuMat to another datatype (Blocking call)
-    CV_WRAP void convertTo(OutputArray dst, int rtype) const;
-
-    //! converts GpuMat to another datatype (Non-Blocking call)
-    CV_WRAP void convertTo(OutputArray dst, int rtype, Stream& stream) const;
-
-    //! converts GpuMat to another datatype with scaling (Blocking call)
-    CV_WRAP void convertTo(OutputArray dst, int rtype, double alpha, double beta = 0.0) const;
-
-    //! converts GpuMat to another datatype with scaling (Non-Blocking call)
-    CV_WRAP void convertTo(OutputArray dst, int rtype, double alpha, Stream& stream) const;
-
-    //! converts GpuMat to another datatype with scaling (Non-Blocking call)
-    CV_WRAP void convertTo(OutputArray dst, int rtype, double alpha, double beta, Stream& stream) const;
-
-    CV_WRAP void assignTo(GpuMat& m, int type = -1) const;
-
-    //! returns pointer to y-th row
-    uchar* ptr(int y = 0);
-    const uchar* ptr(int y = 0) const;
-
-    //! template version of the above method
-    template<typename _Tp> _Tp* ptr(int y = 0);
-    template<typename _Tp> const _Tp* ptr(int y = 0) const;
-
-    template <typename _Tp> operator PtrStepSz<_Tp>() const;
-    template <typename _Tp> operator PtrStep<_Tp>() const;
-
-    //! returns a new GpuMat header for the specified row
-    CV_WRAP GpuMat row(int y) const;
-
-    //! returns a new GpuMat header for the specified column
-    CV_WRAP GpuMat col(int x) const;
-
-    //! ... for the specified row span
-    CV_WRAP GpuMat rowRange(int startrow, int endrow) const;
-    CV_WRAP GpuMat rowRange(Range r) const;
-
-    //! ... for the specified column span
-    CV_WRAP GpuMat colRange(int startcol, int endcol) const;
-    CV_WRAP GpuMat colRange(Range r) const;
-
-    //! extracts a rectangular sub-GpuMat (this is a generalized form of row, rowRange etc.)
-    GpuMat operator ()(Range rowRange, Range colRange) const;
-    GpuMat operator ()(Rect roi) const;
-
-    //! creates alternative GpuMat header for the same data, with different
-    //! number of channels and/or different number of rows
-    CV_WRAP GpuMat reshape(int cn, int rows = 0) const;
-
-    //! locates GpuMat header within a parent GpuMat
-    CV_WRAP void locateROI(Size& wholeSize, Point& ofs) const;
-
-    //! moves/resizes the current GpuMat ROI inside the parent GpuMat
-    CV_WRAP GpuMat& adjustROI(int dtop, int dbottom, int dleft, int dright);
-
-    //! returns true iff the GpuMat data is continuous
-    //! (i.e. when there are no gaps between successive rows)
-    CV_WRAP bool isContinuous() const;
-
-    //! returns element size in bytes
-    CV_WRAP size_t elemSize() const;
-
-    //! returns the size of element channel in bytes
-    CV_WRAP size_t elemSize1() const;
-
-    //! returns element type
-    CV_WRAP int type() const;
-
-    //! returns element type
-    CV_WRAP int depth() const;
-
-    //! returns number of channels
-    CV_WRAP int channels() const;
-
-    //! returns step/elemSize1()
-    CV_WRAP size_t step1() const;
-
-    //! returns GpuMat size : width == number of columns, height == number of rows
-    CV_WRAP Size size() const;
-
-    //! returns true if GpuMat data is NULL
-    CV_WRAP bool empty() const;
-
-    // returns pointer to cuda memory
-    CV_WRAP void* cudaPtr() const;
-
-    //! internal use method: updates the continuity flag
-    CV_WRAP void updateContinuityFlag();
-
-    /*! includes several bit-fields:
-    - the magic signature
-    - continuity flag
-    - depth
-    - number of channels
-    */
-    int flags;
-
-    //! the number of rows and columns
-    int rows, cols;
-
-    //! a distance between successive rows in bytes; includes the gap if any
-    CV_PROP size_t step;
-
-    //! pointer to the data
-    uchar* data;
-
-    //! pointer to the reference counter;
-    //! when GpuMat points to user-allocated data, the pointer is NULL
-    int* refcount;
-
-    //! helper fields used in locateROI and adjustROI
-    uchar* datastart;
-    const uchar* dataend;
-
-    //! allocator
-    Allocator* allocator;
-};
-
-struct CV_EXPORTS_W GpuData
-{
-    explicit GpuData(size_t _size);
-     ~GpuData();
-
-    GpuData(const GpuData&) = delete;
-    GpuData& operator=(const GpuData&) = delete;
-
-    GpuData(GpuData&&) = delete;
-    GpuData& operator=(GpuData&&) = delete;
-
-    uchar* data;
-    size_t size;
-};
-
-class CV_EXPORTS_W GpuMatND
-{
-public:
-    using SizeArray = std::vector<int>;
-    using StepArray = std::vector<size_t>;
-    using IndexArray = std::vector<int>;
-
-    //! destructor
-    ~GpuMatND();
-
-    //! default constructor
-    GpuMatND();
-
-    /** @overload
-    @param size Array of integers specifying an n-dimensional array shape.
-    @param type Array type. Use CV_8UC1, ..., CV_16FC4 to create 1-4 channel matrices, or
-    CV_8UC(n), ..., CV_64FC(n) to create multi-channel (up to CV_CN_MAX channels) matrices.
-    */
-    GpuMatND(SizeArray size, int type);
-
-    /** @overload
-    @param size Array of integers specifying an n-dimensional array shape.
-    @param type Array type. Use CV_8UC1, ..., CV_16FC4 to create 1-4 channel matrices, or
-    CV_8UC(n), ..., CV_64FC(n) to create multi-channel (up to CV_CN_MAX channels) matrices.
-    @param data Pointer to the user data. Matrix constructors that take data and step parameters do not
-    allocate matrix data. Instead, they just initialize the matrix header that points to the specified
-    data, which means that no data is copied. This operation is very efficient and can be used to
-    process external data using OpenCV functions. The external data is not automatically deallocated, so
-    you should take care of it.
-    @param step Array of _size.size()-1 steps in case of a multi-dimensional array (the last step is always
-    set to the element size). If not specified, the matrix is assumed to be continuous.
-    */
-    GpuMatND(SizeArray size, int type, void* data, StepArray step = StepArray());
-
-    /** @brief Allocates GPU memory.
-    Suppose there is some GPU memory already allocated. In that case, this method may choose to reuse that
-    GPU memory under the specific condition: it must be of the same size and type, not externally allocated,
-    the GPU memory is continuous(i.e., isContinuous() is true), and is not a sub-matrix of another GpuMatND
-    (i.e., isSubmatrix() is false). In other words, this method guarantees that the GPU memory allocated by
-    this method is always continuous and is not a sub-region of another GpuMatND.
-    */
-    void create(SizeArray size, int type);
-
-    void release();
-
-    void swap(GpuMatND& m) noexcept;
-
-    /** @brief Creates a full copy of the array and the underlying data.
-    The method creates a full copy of the array. It mimics the behavior of Mat::clone(), i.e.
-    the original step is not taken into account. So, the array copy is a continuous array
-    occupying total()\*elemSize() bytes.
-    */
-    GpuMatND clone() const;
-
-    /** @overload
-    This overload is non-blocking, so it may return even if the copy operation is not finished.
-    */
-    GpuMatND clone(Stream& stream) const;
-
-    /** @brief Extracts a sub-matrix.
-    The operator makes a new header for the specified sub-array of \*this.
-    The operator is an O(1) operation, that is, no matrix data is copied.
-    @param ranges Array of selected ranges along each dimension.
-    */
-    GpuMatND operator()(const std::vector<Range>& ranges) const;
-
-    /** @brief Creates a GpuMat header for a 2D plane part of an n-dim matrix.
-    @note The returned GpuMat is constructed with the constructor for user-allocated data.
-    That is, It does not perform reference counting.
-    @note This function does not increment this GpuMatND's reference counter.
-    */
-    GpuMat createGpuMatHeader(IndexArray idx, Range rowRange, Range colRange) const;
-
-    /** @overload
-    Creates a GpuMat header if this GpuMatND is effectively 2D.
-    @note The returned GpuMat is constructed with the constructor for user-allocated data.
-    That is, It does not perform reference counting.
-    @note This function does not increment this GpuMatND's reference counter.
-    */
-    GpuMat createGpuMatHeader() const;
-
-    /** @brief Extracts a 2D plane part of an n-dim matrix.
-    It differs from createGpuMatHeader(IndexArray, Range, Range) in that it clones a part of this
-    GpuMatND to the returned GpuMat.
-    @note This operator does not increment this GpuMatND's reference counter;
-    */
-    GpuMat operator()(IndexArray idx, Range rowRange, Range colRange) const;
-
-    /** @brief Extracts a 2D plane part of an n-dim matrix if this GpuMatND is effectively 2D.
-    It differs from createGpuMatHeader() in that it clones a part of this GpuMatND.
-    @note This operator does not increment this GpuMatND's reference counter;
-    */
-    operator GpuMat() const;
-
-    GpuMatND(const GpuMatND&) = default;
-    GpuMatND& operator=(const GpuMatND&) = default;
-
-#if defined(__GNUC__) && __GNUC__ < 5
-    // error: function '...' defaulted on its first declaration with an exception-specification
-    // that differs from the implicit declaration '...'
-
-    GpuMatND(GpuMatND&&) = default;
-    GpuMatND& operator=(GpuMatND&&) = default;
-#else
-    GpuMatND(GpuMatND&&) noexcept = default;
-    GpuMatND& operator=(GpuMatND&&) noexcept = default;
-#endif
-
-    void upload(InputArray src);
-    void upload(InputArray src, Stream& stream);
-    void download(OutputArray dst) const;
-    void download(OutputArray dst, Stream& stream) const;
-
-    //! returns true iff the GpuMatND data is continuous
-    //! (i.e. when there are no gaps between successive rows)
-    bool isContinuous() const;
-
-    //! returns true if the matrix is a sub-matrix of another matrix
-    bool isSubmatrix() const;
-
-    //! returns element size in bytes
-    size_t elemSize() const;
-
-    //! returns the size of element channel in bytes
-    size_t elemSize1() const;
-
-    //! returns true if data is null
-    bool empty() const;
-
-    //! returns true if not empty and points to external(user-allocated) gpu memory
-    bool external() const;
-
-    //! returns pointer to the first byte of the GPU memory
-    uchar* getDevicePtr() const;
-
-    //! returns the total number of array elements
-    size_t total() const;
-
-    //! returns the size of underlying memory in bytes
-    size_t totalMemSize() const;
-
-    //! returns element type
-    int type() const;
-
-private:
-    //! internal use
-    void setFields(SizeArray size, int type, StepArray step = StepArray());
-
-public:
-    /*! includes several bit-fields:
-    - the magic signature
-    - continuity flag
-    - depth
-    - number of channels
-    */
-    int flags;
-
-    //! matrix dimensionality
-    int dims;
-
-    //! shape of this array
-    SizeArray size;
-
-    /*! step values
-    Their semantics is identical to the semantics of step for Mat.
-    */
-    StepArray step;
-
-private:
-    /*! internal use
-    If this GpuMatND holds external memory, this is empty.
-    */
-    std::shared_ptr<GpuData> data_;
-
-    /*! internal use
-    If this GpuMatND manages memory with reference counting, this value is
-    always equal to data_->data. If this GpuMatND holds external memory,
-    data_ is empty and data points to the external memory.
-    */
-    uchar* data;
-
-    /*! internal use
-    If this GpuMatND is a sub-matrix of a larger matrix, this value is the
-    difference of the first byte between the sub-matrix and the whole matrix.
-    */
-    size_t offset;
-};
-
-/** @brief Creates a continuous matrix.
-
-@param rows Row count.
-@param cols Column count.
-@param type Type of the matrix.
-@param arr Destination matrix. This parameter changes only if it has a proper type and area (
-\f$\texttt{rows} \times \texttt{cols}\f$ ).
-
-Matrix is called continuous if its elements are stored continuously, that is, without gaps at the
-end of each row.
- */
-CV_EXPORTS_W void createContinuous(int rows, int cols, int type, OutputArray arr);
-
-/** @brief Ensures that the size of a matrix is big enough and the matrix has a proper type.
-
-@param rows Minimum desired number of rows.
-@param cols Minimum desired number of columns.
-@param type Desired matrix type.
-@param arr Destination matrix.
-
-The function does not reallocate memory if the matrix has proper attributes already.
- */
-CV_EXPORTS_W void ensureSizeIsEnough(int rows, int cols, int type, OutputArray arr);
-
-/** @brief BufferPool for use with CUDA streams
-
-BufferPool utilizes Stream's allocator to create new buffers for GpuMat's. It is
-only useful when enabled with #setBufferPoolUsage.
-
-@code
-    setBufferPoolUsage(true);
-@endcode
-
-@note #setBufferPoolUsage must be called \em before any Stream declaration.
-
-Users may specify custom allocator for Stream and may implement their own stream based
-functions utilizing the same underlying GPU memory management.
-
-If custom allocator is not specified, BufferPool utilizes StackAllocator by
-default. StackAllocator allocates a chunk of GPU device memory beforehand,
-and when GpuMat is declared later on, it is given the pre-allocated memory.
-This kind of strategy reduces the number of calls for memory allocating APIs
-such as cudaMalloc or cudaMallocPitch.
-
-Below is an example that utilizes BufferPool with StackAllocator:
-
-@code
-    #include <opencv2/opencv.hpp>
-
-    using namespace cv;
-    using namespace cv::cuda
-
-    int main()
-    {
-        setBufferPoolUsage(true);                               // Tell OpenCV that we are going to utilize BufferPool
-        setBufferPoolConfig(getDevice(), 1024 * 1024 * 64, 2);  // Allocate 64 MB, 2 stacks (default is 10 MB, 5 stacks)
-
-        Stream stream1, stream2;                                // Each stream uses 1 stack
-        BufferPool pool1(stream1), pool2(stream2);
-
-        GpuMat d_src1 = pool1.getBuffer(4096, 4096, CV_8UC1);   // 16MB
-        GpuMat d_dst1 = pool1.getBuffer(4096, 4096, CV_8UC3);   // 48MB, pool1 is now full
-
-        GpuMat d_src2 = pool2.getBuffer(1024, 1024, CV_8UC1);   // 1MB
-        GpuMat d_dst2 = pool2.getBuffer(1024, 1024, CV_8UC3);   // 3MB
-
-        cvtColor(d_src1, d_dst1, CV_GRAY2BGR, 0, stream1);
-        cvtColor(d_src2, d_dst2, CV_GRAY2BGR, 0, stream2);
-    }
-@endcode
-
-If we allocate another GpuMat on pool1 in the above example, it will be carried out by
-the DefaultAllocator since the stack for pool1 is full.
-
-@code
-    GpuMat d_add1 = pool1.getBuffer(1024, 1024, CV_8UC1);   // Stack for pool1 is full, memory is allocated with DefaultAllocator
-@endcode
-
-If a third stream is declared in the above example, allocating with #getBuffer
-within that stream will also be carried out by the DefaultAllocator because we've run out of
-stacks.
-
-@code
-    Stream stream3;                                         // Only 2 stacks were allocated, we've run out of stacks
-    BufferPool pool3(stream3);
-    GpuMat d_src3 = pool3.getBuffer(1024, 1024, CV_8UC1);   // Memory is allocated with DefaultAllocator
-@endcode
-
-@warning When utilizing StackAllocator, deallocation order is important.
-
-Just like a stack, deallocation must be done in LIFO order. Below is an example of
-erroneous usage that violates LIFO rule. If OpenCV is compiled in Debug mode, this
-sample code will emit CV_Assert error.
-
-@code
-    int main()
-    {
-        setBufferPoolUsage(true);                               // Tell OpenCV that we are going to utilize BufferPool
-        Stream stream;                                          // A default size (10 MB) stack is allocated to this stream
-        BufferPool pool(stream);
-
-        GpuMat mat1 = pool.getBuffer(1024, 1024, CV_8UC1);      // Allocate mat1 (1MB)
-        GpuMat mat2 = pool.getBuffer(1024, 1024, CV_8UC1);      // Allocate mat2 (1MB)
-
-        mat1.release();                                         // erroneous usage : mat2 must be deallocated before mat1
-    }
-@endcode
-
-Since C++ local variables are destroyed in the reverse order of construction,
-the code sample below satisfies the LIFO rule. Local GpuMat's are deallocated
-and the corresponding memory is automatically returned to the pool for later usage.
-
-@code
-    int main()
-    {
-        setBufferPoolUsage(true);                               // Tell OpenCV that we are going to utilize BufferPool
-        setBufferPoolConfig(getDevice(), 1024 * 1024 * 64, 2);  // Allocate 64 MB, 2 stacks (default is 10 MB, 5 stacks)
-
-        Stream stream1, stream2;                                // Each stream uses 1 stack
-        BufferPool pool1(stream1), pool2(stream2);
-
-        for (int i = 0; i < 10; i++)
-        {
-            GpuMat d_src1 = pool1.getBuffer(4096, 4096, CV_8UC1);   // 16MB
-            GpuMat d_dst1 = pool1.getBuffer(4096, 4096, CV_8UC3);   // 48MB, pool1 is now full
-
-            GpuMat d_src2 = pool2.getBuffer(1024, 1024, CV_8UC1);   // 1MB
-            GpuMat d_dst2 = pool2.getBuffer(1024, 1024, CV_8UC3);   // 3MB
-
-            d_src1.setTo(Scalar(i), stream1);
-            d_src2.setTo(Scalar(i), stream2);
-
-            cvtColor(d_src1, d_dst1, CV_GRAY2BGR, 0, stream1);
-            cvtColor(d_src2, d_dst2, CV_GRAY2BGR, 0, stream2);
-                                                                    // The order of destruction of the local variables is:
-                                                                    //   d_dst2 => d_src2 => d_dst1 => d_src1
-                                                                    // LIFO rule is satisfied, this code runs without error
-        }
-    }
-@endcode
- */
-class CV_EXPORTS_W BufferPool
-{
-public:
-
-    //! Gets the BufferPool for the given stream.
-    explicit BufferPool(Stream& stream);
-
-    //! Allocates a new GpuMat of given size and type.
-    CV_WRAP GpuMat getBuffer(int rows, int cols, int type);
-
-    //! Allocates a new GpuMat of given size and type.
-    CV_WRAP GpuMat getBuffer(Size size, int type) { return getBuffer(size.height, size.width, type); }
-
-    //! Returns the allocator associated with the stream.
-    CV_WRAP Ptr<GpuMat::Allocator> getAllocator() const { return allocator_; }
-
-private:
-    Ptr<GpuMat::Allocator> allocator_;
-};
-
-//! BufferPool management (must be called before Stream creation)
-CV_EXPORTS_W void setBufferPoolUsage(bool on);
-CV_EXPORTS_W void setBufferPoolConfig(int deviceId, size_t stackSize, int stackCount);
-
-//===================================================================================
-// HostMem
-//===================================================================================
-
-/** @brief Class with reference counting wrapping special memory type allocation functions from CUDA.
-
-Its interface is also Mat-like but with additional memory type parameters.
-
--   **PAGE_LOCKED** sets a page locked memory type used commonly for fast and asynchronous
-    uploading/downloading data from/to GPU.
--   **SHARED** specifies a zero copy memory allocation that enables mapping the host memory to GPU
-    address space, if supported.
--   **WRITE_COMBINED** sets the write combined buffer that is not cached by CPU. Such buffers are
-    used to supply GPU with data when GPU only reads it. The advantage is a better CPU cache
-    utilization.
-
-@note Allocation size of such memory types is usually limited. For more details, see *CUDA 2.2
-Pinned Memory APIs* document or *CUDA C Programming Guide*.
- */
-class CV_EXPORTS_W HostMem
-{
-public:
-    enum AllocType { PAGE_LOCKED = 1, SHARED = 2, WRITE_COMBINED = 4 };
-
-    static MatAllocator* getAllocator(HostMem::AllocType alloc_type = HostMem::AllocType::PAGE_LOCKED);
-
-    CV_WRAP explicit HostMem(HostMem::AllocType alloc_type = HostMem::AllocType::PAGE_LOCKED);
-
-    HostMem(const HostMem& m);
-
-    CV_WRAP HostMem(int rows, int cols, int type, HostMem::AllocType alloc_type = HostMem::AllocType::PAGE_LOCKED);
-    CV_WRAP HostMem(Size size, int type, HostMem::AllocType alloc_type = HostMem::AllocType::PAGE_LOCKED);
-
-    //! creates from host memory with coping data
-    CV_WRAP explicit HostMem(InputArray arr, HostMem::AllocType alloc_type = HostMem::AllocType::PAGE_LOCKED);
-
-    ~HostMem();
-
-    HostMem& operator =(const HostMem& m);
-
-    //! swaps with other smart pointer
-    CV_WRAP void swap(HostMem& b);
-
-    //! returns deep copy of the matrix, i.e. the data is copied
-    CV_WRAP HostMem clone() const;
-
-    //! allocates new matrix data unless the matrix already has specified size and type.
-    CV_WRAP void create(int rows, int cols, int type);
-    void create(Size size, int type);
-
-    //! creates alternative HostMem header for the same data, with different
-    //! number of channels and/or different number of rows
-    CV_WRAP HostMem reshape(int cn, int rows = 0) const;
-
-    //! decrements reference counter and released memory if needed.
-    void release();
-
-    //! returns matrix header with disabled reference counting for HostMem data.
-    CV_WRAP Mat createMatHeader() const;
-
-    /** @brief Maps CPU memory to GPU address space and creates the cuda::GpuMat header without reference counting
-    for it.
-
-    This can be done only if memory was allocated with the SHARED flag and if it is supported by the
-    hardware. Laptops often share video and CPU memory, so address spaces can be mapped, which
-    eliminates an extra copy.
-     */
-    GpuMat createGpuMatHeader() const;
-
-    // Please see cv::Mat for descriptions
-    CV_WRAP bool isContinuous() const;
-    CV_WRAP size_t elemSize() const;
-    CV_WRAP size_t elemSize1() const;
-    CV_WRAP int type() const;
-    CV_WRAP int depth() const;
-    CV_WRAP int channels() const;
-    CV_WRAP size_t step1() const;
-    CV_WRAP Size size() const;
-    CV_WRAP bool empty() const;
-
-    // Please see cv::Mat for descriptions
-    int flags;
-    int rows, cols;
-    CV_PROP size_t step;
-
-    uchar* data;
-    int* refcount;
-
-    uchar* datastart;
-    const uchar* dataend;
-
-    AllocType alloc_type;
-};
-
-/** @brief Page-locks the memory of matrix and maps it for the device(s).
-
-@param m Input matrix.
- */
-CV_EXPORTS_W void registerPageLocked(Mat& m);
-
-/** @brief Unmaps the memory of matrix and makes it pageable again.
-
-@param m Input matrix.
- */
-CV_EXPORTS_W void unregisterPageLocked(Mat& m);
-
-//===================================================================================
-// Stream
-//===================================================================================
-
-/** @brief This class encapsulates a queue of asynchronous calls.
-
-@note Currently, you may face problems if an operation is enqueued twice with different data. Some
-functions use the constant GPU memory, and next call may update the memory before the previous one
-has been finished. But calling different operations asynchronously is safe because each operation
-has its own constant buffer. Memory copy/upload/download/set operations to the buffers you hold are
-also safe.
-
-@note The Stream class is not thread-safe. Please use different Stream objects for different CPU threads.
-
-@code
-void thread1()
-{
-    cv::cuda::Stream stream1;
-    cv::cuda::func1(..., stream1);
-}
-
-void thread2()
-{
-    cv::cuda::Stream stream2;
-    cv::cuda::func2(..., stream2);
-}
-@endcode
-
-@note By default all CUDA routines are launched in Stream::Null() object, if the stream is not specified by user.
-In multi-threading environment the stream objects must be passed explicitly (see previous note).
- */
-class CV_EXPORTS_W Stream
-{
-    typedef void (Stream::*bool_type)() const;
-    void this_type_does_not_support_comparisons() const {}
-
-public:
-    typedef void (*StreamCallback)(int status, void* userData);
-
-    //! creates a new asynchronous stream
-    CV_WRAP Stream();
-
-    //! creates a new asynchronous stream with custom allocator
-    CV_WRAP Stream(const Ptr<GpuMat::Allocator>& allocator);
-
-    /** @brief creates a new Stream using the cudaFlags argument to determine the behaviors of the stream
-
-    @note The cudaFlags parameter is passed to the underlying api cudaStreamCreateWithFlags() and
-    supports the same parameter values.
-    @code
-        // creates an OpenCV cuda::Stream that manages an asynchronous, non-blocking,
-        // non-default CUDA stream
-        cv::cuda::Stream cvStream(cudaStreamNonBlocking);
-    @endcode
-     */
-    CV_WRAP Stream(const size_t cudaFlags);
-
-    /** @brief Returns true if the current stream queue is finished. Otherwise, it returns false.
-    */
-    CV_WRAP bool queryIfComplete() const;
-
-    /** @brief Blocks the current CPU thread until all operations in the stream are complete.
-    */
-    CV_WRAP void waitForCompletion();
-
-    /** @brief Makes a compute stream wait on an event.
-    */
-    CV_WRAP void waitEvent(const Event& event);
-
-    /** @brief Adds a callback to be called on the host after all currently enqueued items in the stream have
-    completed.
-
-    @note Callbacks must not make any CUDA API calls. Callbacks must not perform any synchronization
-    that may depend on outstanding device work or other callbacks that are not mandated to run earlier.
-    Callbacks without a mandated order (in independent streams) execute in undefined order and may be
-    serialized.
-     */
-    void enqueueHostCallback(StreamCallback callback, void* userData);
-
-    //! return Stream object for default CUDA stream
-    CV_WRAP static Stream& Null();
-
-    //! returns true if stream object is not default (!= 0)
-    operator bool_type() const;
-
-    //! return Pointer to CUDA stream
-    CV_WRAP void* cudaPtr() const;
-
-    class Impl;
-
-private:
-    Ptr<Impl> impl_;
-    Stream(const Ptr<Impl>& impl);
-
-    friend struct StreamAccessor;
-    friend class BufferPool;
-    friend class DefaultDeviceInitializer;
-};
-
-class CV_EXPORTS_W Event
-{
-public:
-    enum CreateFlags
-    {
-        DEFAULT        = 0x00,  /**< Default event flag */
-        BLOCKING_SYNC  = 0x01,  /**< Event uses blocking synchronization */
-        DISABLE_TIMING = 0x02,  /**< Event will not record timing data */
-        INTERPROCESS   = 0x04   /**< Event is suitable for interprocess use. DisableTiming must be set */
-    };
-
-    CV_WRAP explicit Event(Event::CreateFlags flags = Event::CreateFlags::DEFAULT);
-
-    //! records an event
-    CV_WRAP void record(Stream& stream = Stream::Null());
-
-    //! queries an event's status
-    CV_WRAP bool queryIfComplete() const;
-
-    //! waits for an event to complete
-    CV_WRAP void waitForCompletion();
-
-    //! computes the elapsed time between events
-    CV_WRAP static float elapsedTime(const Event& start, const Event& end);
-
-    class Impl;
-
-private:
-    Ptr<Impl> impl_;
-    Event(const Ptr<Impl>& impl);
-
-    friend struct EventAccessor;
-};
-
-//! @} cudacore_struct
-
-//===================================================================================
-// Initialization & Info
-//===================================================================================
-
-//! @addtogroup cudacore_init
-//! @{
-
-/** @brief Returns the number of installed CUDA-enabled devices.
-
-Use this function before any other CUDA functions calls. If OpenCV is compiled without CUDA support,
-this function returns 0. If the CUDA driver is not installed, or is incompatible, this function
-returns -1.
- */
-CV_EXPORTS_W int getCudaEnabledDeviceCount();
-
-/** @brief Sets a device and initializes it for the current thread.
-
-@param device System index of a CUDA device starting with 0.
-
-If the call of this function is omitted, a default device is initialized at the fist CUDA usage.
- */
-CV_EXPORTS_W void setDevice(int device);
-
-/** @brief Returns the current device index set by cuda::setDevice or initialized by default.
- */
-CV_EXPORTS_W int getDevice();
-
-/** @brief Explicitly destroys and cleans up all resources associated with the current device in the current
-process.
-
-Any subsequent API call to this device will reinitialize the device.
- */
-CV_EXPORTS_W void resetDevice();
-
-/** @brief Enumeration providing CUDA computing features.
- */
-enum FeatureSet
-{
-    FEATURE_SET_COMPUTE_10 = 10,
-    FEATURE_SET_COMPUTE_11 = 11,
-    FEATURE_SET_COMPUTE_12 = 12,
-    FEATURE_SET_COMPUTE_13 = 13,
-    FEATURE_SET_COMPUTE_20 = 20,
-    FEATURE_SET_COMPUTE_21 = 21,
-    FEATURE_SET_COMPUTE_30 = 30,
-    FEATURE_SET_COMPUTE_32 = 32,
-    FEATURE_SET_COMPUTE_35 = 35,
-    FEATURE_SET_COMPUTE_50 = 50,
-
-    GLOBAL_ATOMICS = FEATURE_SET_COMPUTE_11,
-    SHARED_ATOMICS = FEATURE_SET_COMPUTE_12,
-    NATIVE_DOUBLE = FEATURE_SET_COMPUTE_13,
-    WARP_SHUFFLE_FUNCTIONS = FEATURE_SET_COMPUTE_30,
-    DYNAMIC_PARALLELISM = FEATURE_SET_COMPUTE_35
-};
-
-//! checks whether current device supports the given feature
-CV_EXPORTS bool deviceSupports(FeatureSet feature_set);
-
-/** @brief Class providing a set of static methods to check what NVIDIA\* card architecture the CUDA module was
-built for.
-
-According to the CUDA C Programming Guide Version 3.2: "PTX code produced for some specific compute
-capability can always be compiled to binary code of greater or equal compute capability".
- */
-class CV_EXPORTS_W TargetArchs
-{
-public:
-    /** @brief The following method checks whether the module was built with the support of the given feature:
-
-    @param feature_set Features to be checked. See :ocvcuda::FeatureSet.
-     */
-    static bool builtWith(FeatureSet feature_set);
-
-    /** @brief There is a set of methods to check whether the module contains intermediate (PTX) or binary CUDA
-    code for the given architecture(s):
-
-    @param major Major compute capability version.
-    @param minor Minor compute capability version.
-     */
-    CV_WRAP static bool has(int major, int minor);
-    CV_WRAP static bool hasPtx(int major, int minor);
-    CV_WRAP static bool hasBin(int major, int minor);
-
-    CV_WRAP static bool hasEqualOrLessPtx(int major, int minor);
-    CV_WRAP static bool hasEqualOrGreater(int major, int minor);
-    CV_WRAP static bool hasEqualOrGreaterPtx(int major, int minor);
-    CV_WRAP static bool hasEqualOrGreaterBin(int major, int minor);
-};
-
-/** @brief Class providing functionality for querying the specified GPU properties.
- */
-class CV_EXPORTS_W DeviceInfo
-{
-public:
-    //! creates DeviceInfo object for the current GPU
-    CV_WRAP DeviceInfo();
-
-    /** @brief The constructors.
-
-    @param device_id System index of the CUDA device starting with 0.
-
-    Constructs the DeviceInfo object for the specified device. If device_id parameter is missed, it
-    constructs an object for the current device.
-     */
-    CV_WRAP DeviceInfo(int device_id);
-
-    /** @brief Returns system index of the CUDA device starting with 0.
-    */
-    CV_WRAP int deviceID() const;
-
-    //! ASCII string identifying device
-    const char* name() const;
-
-    //! global memory available on device in bytes
-    CV_WRAP size_t totalGlobalMem() const;
-
-    //! shared memory available per block in bytes
-    CV_WRAP size_t sharedMemPerBlock() const;
-
-    //! 32-bit registers available per block
-    CV_WRAP int regsPerBlock() const;
-
-    //! warp size in threads
-    CV_WRAP int warpSize() const;
-
-    //! maximum pitch in bytes allowed by memory copies
-    CV_WRAP size_t memPitch() const;
-
-    //! maximum number of threads per block
-    CV_WRAP int maxThreadsPerBlock() const;
-
-    //! maximum size of each dimension of a block
-    CV_WRAP Vec3i maxThreadsDim() const;
-
-    //! maximum size of each dimension of a grid
-    CV_WRAP Vec3i maxGridSize() const;
-
-    //! clock frequency in kilohertz
-    CV_WRAP int clockRate() const;
-
-    //! constant memory available on device in bytes
-    CV_WRAP size_t totalConstMem() const;
-
-    //! major compute capability
-    CV_WRAP int majorVersion() const;
-
-    //! minor compute capability
-    CV_WRAP int minorVersion() const;
-
-    //! alignment requirement for textures
-    CV_WRAP size_t textureAlignment() const;
-
-    //! pitch alignment requirement for texture references bound to pitched memory
-    CV_WRAP size_t texturePitchAlignment() const;
-
-    //! number of multiprocessors on device
-    CV_WRAP int multiProcessorCount() const;
-
-    //! specified whether there is a run time limit on kernels
-    CV_WRAP bool kernelExecTimeoutEnabled() const;
-
-    //! device is integrated as opposed to discrete
-    CV_WRAP bool integrated() const;
-
-    //! device can map host memory with cudaHostAlloc/cudaHostGetDevicePointer
-    CV_WRAP bool canMapHostMemory() const;
-
-    enum ComputeMode
-    {
-        ComputeModeDefault,         /**< default compute mode (Multiple threads can use cudaSetDevice with this device) */
-        ComputeModeExclusive,       /**< compute-exclusive-thread mode (Only one thread in one process will be able to use cudaSetDevice with this device) */
-        ComputeModeProhibited,      /**< compute-prohibited mode (No threads can use cudaSetDevice with this device) */
-        ComputeModeExclusiveProcess /**< compute-exclusive-process mode (Many threads in one process will be able to use cudaSetDevice with this device) */
-    };
-
-    //! compute mode
-    CV_WRAP DeviceInfo::ComputeMode computeMode() const;
-
-    //! maximum 1D texture size
-    CV_WRAP int maxTexture1D() const;
-
-    //! maximum 1D mipmapped texture size
-    CV_WRAP int maxTexture1DMipmap() const;
-
-    //! maximum size for 1D textures bound to linear memory
-    CV_WRAP int maxTexture1DLinear() const;
-
-    //! maximum 2D texture dimensions
-    CV_WRAP Vec2i maxTexture2D() const;
-
-    //! maximum 2D mipmapped texture dimensions
-    CV_WRAP Vec2i maxTexture2DMipmap() const;
-
-    //! maximum dimensions (width, height, pitch) for 2D textures bound to pitched memory
-    CV_WRAP Vec3i maxTexture2DLinear() const;
-
-    //! maximum 2D texture dimensions if texture gather operations have to be performed
-    CV_WRAP Vec2i maxTexture2DGather() const;
-
-    //! maximum 3D texture dimensions
-    CV_WRAP Vec3i maxTexture3D() const;
-
-    //! maximum Cubemap texture dimensions
-    CV_WRAP int maxTextureCubemap() const;
-
-    //! maximum 1D layered texture dimensions
-    CV_WRAP Vec2i maxTexture1DLayered() const;
-
-    //! maximum 2D layered texture dimensions
-    CV_WRAP Vec3i maxTexture2DLayered() const;
-
-    //! maximum Cubemap layered texture dimensions
-    CV_WRAP Vec2i maxTextureCubemapLayered() const;
-
-    //! maximum 1D surface size
-    CV_WRAP int maxSurface1D() const;
-
-    //! maximum 2D surface dimensions
-    CV_WRAP Vec2i maxSurface2D() const;
-
-    //! maximum 3D surface dimensions
-    CV_WRAP Vec3i maxSurface3D() const;
-
-    //! maximum 1D layered surface dimensions
-    CV_WRAP Vec2i maxSurface1DLayered() const;
-
-    //! maximum 2D layered surface dimensions
-    CV_WRAP Vec3i maxSurface2DLayered() const;
-
-    //! maximum Cubemap surface dimensions
-    CV_WRAP int maxSurfaceCubemap() const;
-
-    //! maximum Cubemap layered surface dimensions
-    CV_WRAP Vec2i maxSurfaceCubemapLayered() const;
-
-    //! alignment requirements for surfaces
-    CV_WRAP size_t surfaceAlignment() const;
-
-    //! device can possibly execute multiple kernels concurrently
-    CV_WRAP bool concurrentKernels() const;
-
-    //! device has ECC support enabled
-    CV_WRAP bool ECCEnabled() const;
-
-    //! PCI bus ID of the device
-    CV_WRAP int pciBusID() const;
-
-    //! PCI device ID of the device
-    CV_WRAP int pciDeviceID() const;
-
-    //! PCI domain ID of the device
-    CV_WRAP int pciDomainID() const;
-
-    //! true if device is a Tesla device using TCC driver, false otherwise
-    CV_WRAP bool tccDriver() const;
-
-    //! number of asynchronous engines
-    CV_WRAP int asyncEngineCount() const;
-
-    //! device shares a unified address space with the host
-    CV_WRAP bool unifiedAddressing() const;
-
-    //! peak memory clock frequency in kilohertz
-    CV_WRAP int memoryClockRate() const;
-
-    //! global memory bus width in bits
-    CV_WRAP int memoryBusWidth() const;
-
-    //! size of L2 cache in bytes
-    CV_WRAP int l2CacheSize() const;
-
-    //! maximum resident threads per multiprocessor
-    CV_WRAP int maxThreadsPerMultiProcessor() const;
-
-    //! gets free and total device memory
-    CV_WRAP void queryMemory(size_t& totalMemory, size_t& freeMemory) const;
-    CV_WRAP size_t freeMemory() const;
-    CV_WRAP size_t totalMemory() const;
-
-    /** @brief Provides information on CUDA feature support.
-
-    @param feature_set Features to be checked. See cuda::FeatureSet.
-
-    This function returns true if the device has the specified CUDA feature. Otherwise, it returns false
-     */
-    bool supports(FeatureSet feature_set) const;
-
-    /** @brief Checks the CUDA module and device compatibility.
-
-    This function returns true if the CUDA module can be run on the specified device. Otherwise, it
-    returns false .
-     */
-    CV_WRAP bool isCompatible() const;
-
-private:
-    int device_id_;
-};
-
-CV_EXPORTS_W void printCudaDeviceInfo(int device);
-CV_EXPORTS_W void printShortCudaDeviceInfo(int device);
-
-/** @brief Converts an array to half precision floating number.
-
-@param _src input array.
-@param _dst output array.
-@param stream Stream for the asynchronous version.
-@sa convertFp16
-*/
-CV_EXPORTS void convertFp16(InputArray _src, OutputArray _dst, Stream& stream = Stream::Null());
-
-//! @} cudacore_init
-
-}} // namespace cv { namespace cuda {
-
-
-#include "opencv2/core/cuda.inl.hpp"
-
-#endif /* OPENCV_CORE_CUDA_HPP */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda.inl.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda.inl.hpp
deleted file mode 100644
index 3f2a0c72..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda.inl.hpp
+++ /dev/null
@@ -1,723 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_CUDAINL_HPP
-#define OPENCV_CORE_CUDAINL_HPP
-
-#include "opencv2/core/cuda.hpp"
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda {
-
-//===================================================================================
-// GpuMat
-//===================================================================================
-
-inline
-GpuMat::GpuMat(Allocator* allocator_)
-    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), allocator(allocator_)
-{}
-
-inline
-GpuMat::GpuMat(int rows_, int cols_, int type_, Allocator* allocator_)
-    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), allocator(allocator_)
-{
-    if (rows_ > 0 && cols_ > 0)
-        create(rows_, cols_, type_);
-}
-
-inline
-GpuMat::GpuMat(Size size_, int type_, Allocator* allocator_)
-    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), allocator(allocator_)
-{
-    if (size_.height > 0 && size_.width > 0)
-        create(size_.height, size_.width, type_);
-}
-
-inline
-GpuMat::GpuMat(int rows_, int cols_, int type_, Scalar s_, Allocator* allocator_)
-    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), allocator(allocator_)
-{
-    if (rows_ > 0 && cols_ > 0)
-    {
-        create(rows_, cols_, type_);
-        setTo(s_);
-    }
-}
-
-inline
-GpuMat::GpuMat(Size size_, int type_, Scalar s_, Allocator* allocator_)
-    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), allocator(allocator_)
-{
-    if (size_.height > 0 && size_.width > 0)
-    {
-        create(size_.height, size_.width, type_);
-        setTo(s_);
-    }
-}
-
-inline
-GpuMat::GpuMat(const GpuMat& m)
-    : flags(m.flags), rows(m.rows), cols(m.cols), step(m.step), data(m.data), refcount(m.refcount), datastart(m.datastart), dataend(m.dataend), allocator(m.allocator)
-{
-    if (refcount)
-        CV_XADD(refcount, 1);
-}
-
-inline
-GpuMat::GpuMat(InputArray arr, Allocator* allocator_) :
-    flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), allocator(allocator_)
-{
-    upload(arr);
-}
-
-inline
-GpuMat::~GpuMat()
-{
-    release();
-}
-
-inline
-GpuMat& GpuMat::operator =(const GpuMat& m)
-{
-    if (this != &m)
-    {
-        GpuMat temp(m);
-        swap(temp);
-    }
-
-    return *this;
-}
-
-inline
-void GpuMat::create(Size size_, int type_)
-{
-    create(size_.height, size_.width, type_);
-}
-
-inline
-void GpuMat::swap(GpuMat& b)
-{
-    std::swap(flags, b.flags);
-    std::swap(rows, b.rows);
-    std::swap(cols, b.cols);
-    std::swap(step, b.step);
-    std::swap(data, b.data);
-    std::swap(datastart, b.datastart);
-    std::swap(dataend, b.dataend);
-    std::swap(refcount, b.refcount);
-    std::swap(allocator, b.allocator);
-}
-
-inline
-GpuMat GpuMat::clone() const
-{
-    GpuMat m;
-    copyTo(m);
-    return m;
-}
-
-inline
-void GpuMat::copyTo(OutputArray dst, InputArray mask) const
-{
-    copyTo(dst, mask, Stream::Null());
-}
-
-inline
-GpuMat& GpuMat::setTo(Scalar s)
-{
-    return setTo(s, Stream::Null());
-}
-
-inline
-GpuMat& GpuMat::setTo(Scalar s, InputArray mask)
-{
-    return setTo(s, mask, Stream::Null());
-}
-
-inline
-void GpuMat::convertTo(OutputArray dst, int rtype) const
-{
-    convertTo(dst, rtype, Stream::Null());
-}
-
-inline
-void GpuMat::convertTo(OutputArray dst, int rtype, double alpha, double beta) const
-{
-    convertTo(dst, rtype, alpha, beta, Stream::Null());
-}
-
-inline
-void GpuMat::convertTo(OutputArray dst, int rtype, double alpha, Stream& stream) const
-{
-    convertTo(dst, rtype, alpha, 0.0, stream);
-}
-
-inline
-void GpuMat::assignTo(GpuMat& m, int _type) const
-{
-    if (_type < 0)
-        m = *this;
-    else
-        convertTo(m, _type);
-}
-
-inline
-uchar* GpuMat::ptr(int y)
-{
-    CV_DbgAssert( (unsigned)y < (unsigned)rows );
-    return data + step * y;
-}
-
-inline
-const uchar* GpuMat::ptr(int y) const
-{
-    CV_DbgAssert( (unsigned)y < (unsigned)rows );
-    return data + step * y;
-}
-
-template<typename _Tp> inline
-_Tp* GpuMat::ptr(int y)
-{
-    return (_Tp*)ptr(y);
-}
-
-template<typename _Tp> inline
-const _Tp* GpuMat::ptr(int y) const
-{
-    return (const _Tp*)ptr(y);
-}
-
-template <class T> inline
-GpuMat::operator PtrStepSz<T>() const
-{
-    return PtrStepSz<T>(rows, cols, (T*)data, step);
-}
-
-template <class T> inline
-GpuMat::operator PtrStep<T>() const
-{
-    return PtrStep<T>((T*)data, step);
-}
-
-inline
-GpuMat GpuMat::row(int y) const
-{
-    return GpuMat(*this, Range(y, y+1), Range::all());
-}
-
-inline
-GpuMat GpuMat::col(int x) const
-{
-    return GpuMat(*this, Range::all(), Range(x, x+1));
-}
-
-inline
-GpuMat GpuMat::rowRange(int startrow, int endrow) const
-{
-    return GpuMat(*this, Range(startrow, endrow), Range::all());
-}
-
-inline
-GpuMat GpuMat::rowRange(Range r) const
-{
-    return GpuMat(*this, r, Range::all());
-}
-
-inline
-GpuMat GpuMat::colRange(int startcol, int endcol) const
-{
-    return GpuMat(*this, Range::all(), Range(startcol, endcol));
-}
-
-inline
-GpuMat GpuMat::colRange(Range r) const
-{
-    return GpuMat(*this, Range::all(), r);
-}
-
-inline
-GpuMat GpuMat::operator ()(Range rowRange_, Range colRange_) const
-{
-    return GpuMat(*this, rowRange_, colRange_);
-}
-
-inline
-GpuMat GpuMat::operator ()(Rect roi) const
-{
-    return GpuMat(*this, roi);
-}
-
-inline
-bool GpuMat::isContinuous() const
-{
-    return (flags & Mat::CONTINUOUS_FLAG) != 0;
-}
-
-inline
-size_t GpuMat::elemSize() const
-{
-    return CV_ELEM_SIZE(flags);
-}
-
-inline
-size_t GpuMat::elemSize1() const
-{
-    return CV_ELEM_SIZE1(flags);
-}
-
-inline
-int GpuMat::type() const
-{
-    return CV_MAT_TYPE(flags);
-}
-
-inline
-int GpuMat::depth() const
-{
-    return CV_MAT_DEPTH(flags);
-}
-
-inline
-int GpuMat::channels() const
-{
-    return CV_MAT_CN(flags);
-}
-
-inline
-size_t GpuMat::step1() const
-{
-    return step / elemSize1();
-}
-
-inline
-Size GpuMat::size() const
-{
-    return Size(cols, rows);
-}
-
-inline
-bool GpuMat::empty() const
-{
-    return data == 0;
-}
-
-inline
-void* GpuMat::cudaPtr() const
-{
-    return data;
-}
-
-static inline
-GpuMat createContinuous(int rows, int cols, int type)
-{
-    GpuMat m;
-    createContinuous(rows, cols, type, m);
-    return m;
-}
-
-static inline
-void createContinuous(Size size, int type, OutputArray arr)
-{
-    createContinuous(size.height, size.width, type, arr);
-}
-
-static inline
-GpuMat createContinuous(Size size, int type)
-{
-    GpuMat m;
-    createContinuous(size, type, m);
-    return m;
-}
-
-static inline
-void ensureSizeIsEnough(Size size, int type, OutputArray arr)
-{
-    ensureSizeIsEnough(size.height, size.width, type, arr);
-}
-
-static inline
-void swap(GpuMat& a, GpuMat& b)
-{
-    a.swap(b);
-}
-
-//===================================================================================
-// GpuMatND
-//===================================================================================
-
-inline
-GpuMatND::GpuMatND() :
-    flags(0), dims(0), data(nullptr), offset(0)
-{
-}
-
-inline
-GpuMatND::GpuMatND(SizeArray _size, int _type) :
-    flags(0), dims(0), data(nullptr), offset(0)
-{
-    create(std::move(_size), _type);
-}
-
-inline
-void GpuMatND::swap(GpuMatND& m) noexcept
-{
-    std::swap(*this, m);
-}
-
-inline
-bool GpuMatND::isContinuous() const
-{
-    return (flags & Mat::CONTINUOUS_FLAG) != 0;
-}
-
-inline
-bool GpuMatND::isSubmatrix() const
-{
-    return (flags & Mat::SUBMATRIX_FLAG) != 0;
-}
-
-inline
-size_t GpuMatND::elemSize() const
-{
-    return CV_ELEM_SIZE(flags);
-}
-
-inline
-size_t GpuMatND::elemSize1() const
-{
-    return CV_ELEM_SIZE1(flags);
-}
-
-inline
-bool GpuMatND::empty() const
-{
-    return data == nullptr;
-}
-
-inline
-bool GpuMatND::external() const
-{
-    return !empty() && data_.use_count() == 0;
-}
-
-inline
-uchar* GpuMatND::getDevicePtr() const
-{
-    return data + offset;
-}
-
-inline
-size_t GpuMatND::total() const
-{
-    size_t p = 1;
-    for(auto s : size)
-        p *= s;
-    return p;
-}
-
-inline
-size_t GpuMatND::totalMemSize() const
-{
-    return size[0] * step[0];
-}
-
-inline
-int GpuMatND::type() const
-{
-    return CV_MAT_TYPE(flags);
-}
-
-//===================================================================================
-// HostMem
-//===================================================================================
-
-inline
-HostMem::HostMem(AllocType alloc_type_)
-    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), alloc_type(alloc_type_)
-{
-}
-
-inline
-HostMem::HostMem(const HostMem& m)
-    : flags(m.flags), rows(m.rows), cols(m.cols), step(m.step), data(m.data), refcount(m.refcount), datastart(m.datastart), dataend(m.dataend), alloc_type(m.alloc_type)
-{
-    if( refcount )
-        CV_XADD(refcount, 1);
-}
-
-inline
-HostMem::HostMem(int rows_, int cols_, int type_, AllocType alloc_type_)
-    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), alloc_type(alloc_type_)
-{
-    if (rows_ > 0 && cols_ > 0)
-        create(rows_, cols_, type_);
-}
-
-inline
-HostMem::HostMem(Size size_, int type_, AllocType alloc_type_)
-    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), alloc_type(alloc_type_)
-{
-    if (size_.height > 0 && size_.width > 0)
-        create(size_.height, size_.width, type_);
-}
-
-inline
-HostMem::HostMem(InputArray arr, AllocType alloc_type_)
-    : flags(0), rows(0), cols(0), step(0), data(0), refcount(0), datastart(0), dataend(0), alloc_type(alloc_type_)
-{
-    arr.getMat().copyTo(*this);
-}
-
-inline
-HostMem::~HostMem()
-{
-    release();
-}
-
-inline
-HostMem& HostMem::operator =(const HostMem& m)
-{
-    if (this != &m)
-    {
-        HostMem temp(m);
-        swap(temp);
-    }
-
-    return *this;
-}
-
-inline
-void HostMem::swap(HostMem& b)
-{
-    std::swap(flags, b.flags);
-    std::swap(rows, b.rows);
-    std::swap(cols, b.cols);
-    std::swap(step, b.step);
-    std::swap(data, b.data);
-    std::swap(datastart, b.datastart);
-    std::swap(dataend, b.dataend);
-    std::swap(refcount, b.refcount);
-    std::swap(alloc_type, b.alloc_type);
-}
-
-inline
-HostMem HostMem::clone() const
-{
-    HostMem m(size(), type(), alloc_type);
-    createMatHeader().copyTo(m);
-    return m;
-}
-
-inline
-void HostMem::create(Size size_, int type_)
-{
-    create(size_.height, size_.width, type_);
-}
-
-inline
-Mat HostMem::createMatHeader() const
-{
-    return Mat(size(), type(), data, step);
-}
-
-inline
-bool HostMem::isContinuous() const
-{
-    return (flags & Mat::CONTINUOUS_FLAG) != 0;
-}
-
-inline
-size_t HostMem::elemSize() const
-{
-    return CV_ELEM_SIZE(flags);
-}
-
-inline
-size_t HostMem::elemSize1() const
-{
-    return CV_ELEM_SIZE1(flags);
-}
-
-inline
-int HostMem::type() const
-{
-    return CV_MAT_TYPE(flags);
-}
-
-inline
-int HostMem::depth() const
-{
-    return CV_MAT_DEPTH(flags);
-}
-
-inline
-int HostMem::channels() const
-{
-    return CV_MAT_CN(flags);
-}
-
-inline
-size_t HostMem::step1() const
-{
-    return step / elemSize1();
-}
-
-inline
-Size HostMem::size() const
-{
-    return Size(cols, rows);
-}
-
-inline
-bool HostMem::empty() const
-{
-    return data == 0;
-}
-
-static inline
-void swap(HostMem& a, HostMem& b)
-{
-    a.swap(b);
-}
-
-//===================================================================================
-// Stream
-//===================================================================================
-
-inline
-Stream::Stream(const Ptr<Impl>& impl)
-    : impl_(impl)
-{
-}
-
-//===================================================================================
-// Event
-//===================================================================================
-
-inline
-Event::Event(const Ptr<Impl>& impl)
-    : impl_(impl)
-{
-}
-
-//===================================================================================
-// Initialization & Info
-//===================================================================================
-
-inline
-bool TargetArchs::has(int major, int minor)
-{
-    return hasPtx(major, minor) || hasBin(major, minor);
-}
-
-inline
-bool TargetArchs::hasEqualOrGreater(int major, int minor)
-{
-    return hasEqualOrGreaterPtx(major, minor) || hasEqualOrGreaterBin(major, minor);
-}
-
-inline
-DeviceInfo::DeviceInfo()
-{
-    device_id_ = getDevice();
-}
-
-inline
-DeviceInfo::DeviceInfo(int device_id)
-{
-    CV_Assert( device_id >= 0 && device_id < getCudaEnabledDeviceCount() );
-    device_id_ = device_id;
-}
-
-inline
-int DeviceInfo::deviceID() const
-{
-    return device_id_;
-}
-
-inline
-size_t DeviceInfo::freeMemory() const
-{
-    size_t _totalMemory = 0, _freeMemory = 0;
-    queryMemory(_totalMemory, _freeMemory);
-    return _freeMemory;
-}
-
-inline
-size_t DeviceInfo::totalMemory() const
-{
-    size_t _totalMemory = 0, _freeMemory = 0;
-    queryMemory(_totalMemory, _freeMemory);
-    return _totalMemory;
-}
-
-inline
-bool DeviceInfo::supports(FeatureSet feature_set) const
-{
-    int version = majorVersion() * 10 + minorVersion();
-    return version >= feature_set;
-}
-
-
-}} // namespace cv { namespace cuda {
-
-//===================================================================================
-// Mat
-//===================================================================================
-
-namespace cv {
-
-inline
-Mat::Mat(const cuda::GpuMat& m)
-    : flags(0), dims(0), rows(0), cols(0), data(0), datastart(0), dataend(0), datalimit(0), allocator(0), u(0), size(&rows)
-{
-    m.download(*this);
-}
-
-}
-
-//! @endcond
-
-#endif // OPENCV_CORE_CUDAINL_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/block.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/block.hpp
deleted file mode 100644
index c277f0ea..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/block.hpp
+++ /dev/null
@@ -1,211 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_DEVICE_BLOCK_HPP
-#define OPENCV_CUDA_DEVICE_BLOCK_HPP
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    struct Block
-    {
-        static __device__ __forceinline__ unsigned int id()
-        {
-            return blockIdx.x;
-        }
-
-        static __device__ __forceinline__ unsigned int stride()
-        {
-            return blockDim.x * blockDim.y * blockDim.z;
-        }
-
-        static __device__ __forceinline__ void sync()
-        {
-            __syncthreads();
-        }
-
-        static __device__ __forceinline__ int flattenedThreadId()
-        {
-            return threadIdx.z * blockDim.x * blockDim.y + threadIdx.y * blockDim.x + threadIdx.x;
-        }
-
-        template<typename It, typename T>
-        static __device__ __forceinline__ void fill(It beg, It end, const T& value)
-        {
-            int STRIDE = stride();
-            It t = beg + flattenedThreadId();
-
-            for(; t < end; t += STRIDE)
-                *t = value;
-        }
-
-        template<typename OutIt, typename T>
-        static __device__ __forceinline__ void yota(OutIt beg, OutIt end, T value)
-        {
-            int STRIDE = stride();
-            int tid = flattenedThreadId();
-            value += tid;
-
-            for(OutIt t = beg + tid; t < end; t += STRIDE, value += STRIDE)
-                *t = value;
-        }
-
-        template<typename InIt, typename OutIt>
-        static __device__ __forceinline__ void copy(InIt beg, InIt end, OutIt out)
-        {
-            int STRIDE = stride();
-            InIt  t = beg + flattenedThreadId();
-            OutIt o = out + (t - beg);
-
-            for(; t < end; t += STRIDE, o += STRIDE)
-                *o = *t;
-        }
-
-        template<typename InIt, typename OutIt, class UnOp>
-        static __device__ __forceinline__ void transform(InIt beg, InIt end, OutIt out, UnOp op)
-        {
-            int STRIDE = stride();
-            InIt  t = beg + flattenedThreadId();
-            OutIt o = out + (t - beg);
-
-            for(; t < end; t += STRIDE, o += STRIDE)
-                *o = op(*t);
-        }
-
-        template<typename InIt1, typename InIt2, typename OutIt, class BinOp>
-        static __device__ __forceinline__ void transform(InIt1 beg1, InIt1 end1, InIt2 beg2, OutIt out, BinOp op)
-        {
-            int STRIDE = stride();
-            InIt1 t1 = beg1 + flattenedThreadId();
-            InIt2 t2 = beg2 + flattenedThreadId();
-            OutIt o  = out + (t1 - beg1);
-
-            for(; t1 < end1; t1 += STRIDE, t2 += STRIDE, o += STRIDE)
-                *o = op(*t1, *t2);
-        }
-
-        template<int CTA_SIZE, typename T, class BinOp>
-        static __device__ __forceinline__ void reduce(volatile T* buffer, BinOp op)
-        {
-            int tid = flattenedThreadId();
-            T val =  buffer[tid];
-
-            if (CTA_SIZE >= 1024) { if (tid < 512) buffer[tid] = val = op(val, buffer[tid + 512]); __syncthreads(); }
-            if (CTA_SIZE >=  512) { if (tid < 256) buffer[tid] = val = op(val, buffer[tid + 256]); __syncthreads(); }
-            if (CTA_SIZE >=  256) { if (tid < 128) buffer[tid] = val = op(val, buffer[tid + 128]); __syncthreads(); }
-            if (CTA_SIZE >=  128) { if (tid <  64) buffer[tid] = val = op(val, buffer[tid +  64]); __syncthreads(); }
-
-            if (tid < 32)
-            {
-                if (CTA_SIZE >=   64) { buffer[tid] = val = op(val, buffer[tid +  32]); }
-                if (CTA_SIZE >=   32) { buffer[tid] = val = op(val, buffer[tid +  16]); }
-                if (CTA_SIZE >=   16) { buffer[tid] = val = op(val, buffer[tid +   8]); }
-                if (CTA_SIZE >=    8) { buffer[tid] = val = op(val, buffer[tid +   4]); }
-                if (CTA_SIZE >=    4) { buffer[tid] = val = op(val, buffer[tid +   2]); }
-                if (CTA_SIZE >=    2) { buffer[tid] = val = op(val, buffer[tid +   1]); }
-            }
-        }
-
-        template<int CTA_SIZE, typename T, class BinOp>
-        static __device__ __forceinline__ T reduce(volatile T* buffer, T init, BinOp op)
-        {
-            int tid = flattenedThreadId();
-            T val =  buffer[tid] = init;
-            __syncthreads();
-
-            if (CTA_SIZE >= 1024) { if (tid < 512) buffer[tid] = val = op(val, buffer[tid + 512]); __syncthreads(); }
-            if (CTA_SIZE >=  512) { if (tid < 256) buffer[tid] = val = op(val, buffer[tid + 256]); __syncthreads(); }
-            if (CTA_SIZE >=  256) { if (tid < 128) buffer[tid] = val = op(val, buffer[tid + 128]); __syncthreads(); }
-            if (CTA_SIZE >=  128) { if (tid <  64) buffer[tid] = val = op(val, buffer[tid +  64]); __syncthreads(); }
-
-            if (tid < 32)
-            {
-                if (CTA_SIZE >=   64) { buffer[tid] = val = op(val, buffer[tid +  32]); }
-                if (CTA_SIZE >=   32) { buffer[tid] = val = op(val, buffer[tid +  16]); }
-                if (CTA_SIZE >=   16) { buffer[tid] = val = op(val, buffer[tid +   8]); }
-                if (CTA_SIZE >=    8) { buffer[tid] = val = op(val, buffer[tid +   4]); }
-                if (CTA_SIZE >=    4) { buffer[tid] = val = op(val, buffer[tid +   2]); }
-                if (CTA_SIZE >=    2) { buffer[tid] = val = op(val, buffer[tid +   1]); }
-            }
-            __syncthreads();
-            return buffer[0];
-        }
-
-        template <typename T, class BinOp>
-        static __device__ __forceinline__ void reduce_n(T* data, unsigned int n, BinOp op)
-        {
-            int ftid = flattenedThreadId();
-            int sft = stride();
-
-            if (sft < n)
-            {
-                for (unsigned int i = sft + ftid; i < n; i += sft)
-                    data[ftid] = op(data[ftid], data[i]);
-
-                __syncthreads();
-
-                n = sft;
-            }
-
-            while (n > 1)
-            {
-                unsigned int half = n/2;
-
-                if (ftid < half)
-                    data[ftid] = op(data[ftid], data[n - ftid - 1]);
-
-                __syncthreads();
-
-                n = n - half;
-            }
-        }
-    };
-}}}
-
-//! @endcond
-
-#endif /* OPENCV_CUDA_DEVICE_BLOCK_HPP */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/border_interpolate.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/border_interpolate.hpp
deleted file mode 100644
index 874f705b..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/border_interpolate.hpp
+++ /dev/null
@@ -1,722 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_BORDER_INTERPOLATE_HPP
-#define OPENCV_CUDA_BORDER_INTERPOLATE_HPP
-
-#include "saturate_cast.hpp"
-#include "vec_traits.hpp"
-#include "vec_math.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    //////////////////////////////////////////////////////////////
-    // BrdConstant
-
-    template <typename D> struct BrdRowConstant
-    {
-        typedef D result_type;
-
-        explicit __host__ __device__ __forceinline__ BrdRowConstant(int width_, const D& val_ = VecTraits<D>::all(0)) : width(width_), val(val_) {}
-
-        template <typename T> __device__ __forceinline__ D at_low(int x, const T* data) const
-        {
-            return x >= 0 ? saturate_cast<D>(data[x]) : val;
-        }
-
-        template <typename T> __device__ __forceinline__ D at_high(int x, const T* data) const
-        {
-            return x < width ? saturate_cast<D>(data[x]) : val;
-        }
-
-        template <typename T> __device__ __forceinline__ D at(int x, const T* data) const
-        {
-            return (x >= 0 && x < width) ? saturate_cast<D>(data[x]) : val;
-        }
-
-        int width;
-        D val;
-    };
-
-    template <typename D> struct BrdColConstant
-    {
-        typedef D result_type;
-
-        explicit __host__ __device__ __forceinline__ BrdColConstant(int height_, const D& val_ = VecTraits<D>::all(0)) : height(height_), val(val_) {}
-
-        template <typename T> __device__ __forceinline__ D at_low(int y, const T* data, size_t step) const
-        {
-            return y >= 0 ? saturate_cast<D>(*(const T*)((const char*)data + y * step)) : val;
-        }
-
-        template <typename T> __device__ __forceinline__ D at_high(int y, const T* data, size_t step) const
-        {
-            return y < height ? saturate_cast<D>(*(const T*)((const char*)data + y * step)) : val;
-        }
-
-        template <typename T> __device__ __forceinline__ D at(int y, const T* data, size_t step) const
-        {
-            return (y >= 0 && y < height) ? saturate_cast<D>(*(const T*)((const char*)data + y * step)) : val;
-        }
-
-        int height;
-        D val;
-    };
-
-    template <typename D> struct BrdConstant
-    {
-        typedef D result_type;
-
-        __host__ __device__ __forceinline__ BrdConstant(int height_, int width_, const D& val_ = VecTraits<D>::all(0)) : height(height_), width(width_), val(val_)
-        {
-        }
-
-        template <typename T> __device__ __forceinline__ D at(int y, int x, const T* data, size_t step) const
-        {
-            return (x >= 0 && x < width && y >= 0 && y < height) ? saturate_cast<D>(((const T*)((const uchar*)data + y * step))[x]) : val;
-        }
-
-        template <typename Ptr2D> __device__ __forceinline__ D at(typename Ptr2D::index_type y, typename Ptr2D::index_type x, const Ptr2D& src) const
-        {
-            return (x >= 0 && x < width && y >= 0 && y < height) ? saturate_cast<D>(src(y, x)) : val;
-        }
-
-        int height;
-        int width;
-        D val;
-    };
-
-    //////////////////////////////////////////////////////////////
-    // BrdReplicate
-
-    template <typename D> struct BrdRowReplicate
-    {
-        typedef D result_type;
-
-        explicit __host__ __device__ __forceinline__ BrdRowReplicate(int width) : last_col(width - 1) {}
-        template <typename U> __host__ __device__ __forceinline__ BrdRowReplicate(int width, U) : last_col(width - 1) {}
-
-        __device__ __forceinline__ int idx_col_low(int x) const
-        {
-            return ::max(x, 0);
-        }
-
-        __device__ __forceinline__ int idx_col_high(int x) const
-        {
-            return ::min(x, last_col);
-        }
-
-        __device__ __forceinline__ int idx_col(int x) const
-        {
-            return idx_col_low(idx_col_high(x));
-        }
-
-        template <typename T> __device__ __forceinline__ D at_low(int x, const T* data) const
-        {
-            return saturate_cast<D>(data[idx_col_low(x)]);
-        }
-
-        template <typename T> __device__ __forceinline__ D at_high(int x, const T* data) const
-        {
-            return saturate_cast<D>(data[idx_col_high(x)]);
-        }
-
-        template <typename T> __device__ __forceinline__ D at(int x, const T* data) const
-        {
-            return saturate_cast<D>(data[idx_col(x)]);
-        }
-
-        int last_col;
-    };
-
-    template <typename D> struct BrdColReplicate
-    {
-        typedef D result_type;
-
-        explicit __host__ __device__ __forceinline__ BrdColReplicate(int height) : last_row(height - 1) {}
-        template <typename U> __host__ __device__ __forceinline__ BrdColReplicate(int height, U) : last_row(height - 1) {}
-
-        __device__ __forceinline__ int idx_row_low(int y) const
-        {
-            return ::max(y, 0);
-        }
-
-        __device__ __forceinline__ int idx_row_high(int y) const
-        {
-            return ::min(y, last_row);
-        }
-
-        __device__ __forceinline__ int idx_row(int y) const
-        {
-            return idx_row_low(idx_row_high(y));
-        }
-
-        template <typename T> __device__ __forceinline__ D at_low(int y, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(*(const T*)((const char*)data + idx_row_low(y) * step));
-        }
-
-        template <typename T> __device__ __forceinline__ D at_high(int y, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(*(const T*)((const char*)data + idx_row_high(y) * step));
-        }
-
-        template <typename T> __device__ __forceinline__ D at(int y, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(*(const T*)((const char*)data + idx_row(y) * step));
-        }
-
-        int last_row;
-    };
-
-    template <typename D> struct BrdReplicate
-    {
-        typedef D result_type;
-
-        __host__ __device__ __forceinline__ BrdReplicate(int height, int width) : last_row(height - 1), last_col(width - 1) {}
-        template <typename U> __host__ __device__ __forceinline__ BrdReplicate(int height, int width, U) : last_row(height - 1), last_col(width - 1) {}
-
-        __device__ __forceinline__ int idx_row_low(int y) const
-        {
-            return ::max(y, 0);
-        }
-
-        __device__ __forceinline__ int idx_row_high(int y) const
-        {
-            return ::min(y, last_row);
-        }
-
-        __device__ __forceinline__ int idx_row(int y) const
-        {
-            return idx_row_low(idx_row_high(y));
-        }
-
-        __device__ __forceinline__ int idx_col_low(int x) const
-        {
-            return ::max(x, 0);
-        }
-
-        __device__ __forceinline__ int idx_col_high(int x) const
-        {
-            return ::min(x, last_col);
-        }
-
-        __device__ __forceinline__ int idx_col(int x) const
-        {
-            return idx_col_low(idx_col_high(x));
-        }
-
-        template <typename T> __device__ __forceinline__ D at(int y, int x, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(((const T*)((const char*)data + idx_row(y) * step))[idx_col(x)]);
-        }
-
-        template <typename Ptr2D> __device__ __forceinline__ D at(typename Ptr2D::index_type y, typename Ptr2D::index_type x, const Ptr2D& src) const
-        {
-            return saturate_cast<D>(src(idx_row(y), idx_col(x)));
-        }
-
-        int last_row;
-        int last_col;
-    };
-
-    //////////////////////////////////////////////////////////////
-    // BrdReflect101
-
-    template <typename D> struct BrdRowReflect101
-    {
-        typedef D result_type;
-
-        explicit __host__ __device__ __forceinline__ BrdRowReflect101(int width) : last_col(width - 1) {}
-        template <typename U> __host__ __device__ __forceinline__ BrdRowReflect101(int width, U) : last_col(width - 1) {}
-
-        __device__ __forceinline__ int idx_col_low(int x) const
-        {
-            return ::abs(x) % (last_col + 1);
-        }
-
-        __device__ __forceinline__ int idx_col_high(int x) const
-        {
-            return ::abs(last_col - ::abs(last_col - x)) % (last_col + 1);
-        }
-
-        __device__ __forceinline__ int idx_col(int x) const
-        {
-            return idx_col_low(idx_col_high(x));
-        }
-
-        template <typename T> __device__ __forceinline__ D at_low(int x, const T* data) const
-        {
-            return saturate_cast<D>(data[idx_col_low(x)]);
-        }
-
-        template <typename T> __device__ __forceinline__ D at_high(int x, const T* data) const
-        {
-            return saturate_cast<D>(data[idx_col_high(x)]);
-        }
-
-        template <typename T> __device__ __forceinline__ D at(int x, const T* data) const
-        {
-            return saturate_cast<D>(data[idx_col(x)]);
-        }
-
-        int last_col;
-    };
-
-    template <typename D> struct BrdColReflect101
-    {
-        typedef D result_type;
-
-        explicit __host__ __device__ __forceinline__ BrdColReflect101(int height) : last_row(height - 1) {}
-        template <typename U> __host__ __device__ __forceinline__ BrdColReflect101(int height, U) : last_row(height - 1) {}
-
-        __device__ __forceinline__ int idx_row_low(int y) const
-        {
-            return ::abs(y) % (last_row + 1);
-        }
-
-        __device__ __forceinline__ int idx_row_high(int y) const
-        {
-            return ::abs(last_row - ::abs(last_row - y)) % (last_row + 1);
-        }
-
-        __device__ __forceinline__ int idx_row(int y) const
-        {
-            return idx_row_low(idx_row_high(y));
-        }
-
-        template <typename T> __device__ __forceinline__ D at_low(int y, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(*(const D*)((const char*)data + idx_row_low(y) * step));
-        }
-
-        template <typename T> __device__ __forceinline__ D at_high(int y, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(*(const D*)((const char*)data + idx_row_high(y) * step));
-        }
-
-        template <typename T> __device__ __forceinline__ D at(int y, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(*(const D*)((const char*)data + idx_row(y) * step));
-        }
-
-        int last_row;
-    };
-
-    template <typename D> struct BrdReflect101
-    {
-        typedef D result_type;
-
-        __host__ __device__ __forceinline__ BrdReflect101(int height, int width) : last_row(height - 1), last_col(width - 1) {}
-        template <typename U> __host__ __device__ __forceinline__ BrdReflect101(int height, int width, U) : last_row(height - 1), last_col(width - 1) {}
-
-        __device__ __forceinline__ int idx_row_low(int y) const
-        {
-            return ::abs(y) % (last_row + 1);
-        }
-
-        __device__ __forceinline__ int idx_row_high(int y) const
-        {
-            return ::abs(last_row - ::abs(last_row - y)) % (last_row + 1);
-        }
-
-        __device__ __forceinline__ int idx_row(int y) const
-        {
-            return idx_row_low(idx_row_high(y));
-        }
-
-        __device__ __forceinline__ int idx_col_low(int x) const
-        {
-            return ::abs(x) % (last_col + 1);
-        }
-
-        __device__ __forceinline__ int idx_col_high(int x) const
-        {
-            return ::abs(last_col - ::abs(last_col - x)) % (last_col + 1);
-        }
-
-        __device__ __forceinline__ int idx_col(int x) const
-        {
-            return idx_col_low(idx_col_high(x));
-        }
-
-        template <typename T> __device__ __forceinline__ D at(int y, int x, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(((const T*)((const char*)data + idx_row(y) * step))[idx_col(x)]);
-        }
-
-        template <typename Ptr2D> __device__ __forceinline__ D at(typename Ptr2D::index_type y, typename Ptr2D::index_type x, const Ptr2D& src) const
-        {
-            return saturate_cast<D>(src(idx_row(y), idx_col(x)));
-        }
-
-        int last_row;
-        int last_col;
-    };
-
-    //////////////////////////////////////////////////////////////
-    // BrdReflect
-
-    template <typename D> struct BrdRowReflect
-    {
-        typedef D result_type;
-
-        explicit __host__ __device__ __forceinline__ BrdRowReflect(int width) : last_col(width - 1) {}
-        template <typename U> __host__ __device__ __forceinline__ BrdRowReflect(int width, U) : last_col(width - 1) {}
-
-        __device__ __forceinline__ int idx_col_low(int x) const
-        {
-            return (::abs(x) - (x < 0)) % (last_col + 1);
-        }
-
-        __device__ __forceinline__ int idx_col_high(int x) const
-        {
-            return ::abs(last_col - ::abs(last_col - x) + (x > last_col)) % (last_col + 1);
-        }
-
-        __device__ __forceinline__ int idx_col(int x) const
-        {
-            return idx_col_high(::abs(x) - (x < 0));
-        }
-
-        template <typename T> __device__ __forceinline__ D at_low(int x, const T* data) const
-        {
-            return saturate_cast<D>(data[idx_col_low(x)]);
-        }
-
-        template <typename T> __device__ __forceinline__ D at_high(int x, const T* data) const
-        {
-            return saturate_cast<D>(data[idx_col_high(x)]);
-        }
-
-        template <typename T> __device__ __forceinline__ D at(int x, const T* data) const
-        {
-            return saturate_cast<D>(data[idx_col(x)]);
-        }
-
-        int last_col;
-    };
-
-    template <typename D> struct BrdColReflect
-    {
-        typedef D result_type;
-
-        explicit __host__ __device__ __forceinline__ BrdColReflect(int height) : last_row(height - 1) {}
-        template <typename U> __host__ __device__ __forceinline__ BrdColReflect(int height, U) : last_row(height - 1) {}
-
-        __device__ __forceinline__ int idx_row_low(int y) const
-        {
-            return (::abs(y) - (y < 0)) % (last_row + 1);
-        }
-
-        __device__ __forceinline__ int idx_row_high(int y) const
-        {
-            return ::abs(last_row - ::abs(last_row - y) + (y > last_row)) % (last_row + 1);
-        }
-
-        __device__ __forceinline__ int idx_row(int y) const
-        {
-            return idx_row_high(::abs(y) - (y < 0));
-        }
-
-        template <typename T> __device__ __forceinline__ D at_low(int y, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(*(const D*)((const char*)data + idx_row_low(y) * step));
-        }
-
-        template <typename T> __device__ __forceinline__ D at_high(int y, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(*(const D*)((const char*)data + idx_row_high(y) * step));
-        }
-
-        template <typename T> __device__ __forceinline__ D at(int y, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(*(const D*)((const char*)data + idx_row(y) * step));
-        }
-
-        int last_row;
-    };
-
-    template <typename D> struct BrdReflect
-    {
-        typedef D result_type;
-
-        __host__ __device__ __forceinline__ BrdReflect(int height, int width) : last_row(height - 1), last_col(width - 1) {}
-        template <typename U> __host__ __device__ __forceinline__ BrdReflect(int height, int width, U) : last_row(height - 1), last_col(width - 1) {}
-
-        __device__ __forceinline__ int idx_row_low(int y) const
-        {
-            return (::abs(y) - (y < 0)) % (last_row + 1);
-        }
-
-        __device__ __forceinline__ int idx_row_high(int y) const
-        {
-            return /*::abs*/(last_row - ::abs(last_row - y) + (y > last_row)) /*% (last_row + 1)*/;
-        }
-
-        __device__ __forceinline__ int idx_row(int y) const
-        {
-            return idx_row_low(idx_row_high(y));
-        }
-
-        __device__ __forceinline__ int idx_col_low(int x) const
-        {
-            return (::abs(x) - (x < 0)) % (last_col + 1);
-        }
-
-        __device__ __forceinline__ int idx_col_high(int x) const
-        {
-            return (last_col - ::abs(last_col - x) + (x > last_col));
-        }
-
-        __device__ __forceinline__ int idx_col(int x) const
-        {
-            return idx_col_low(idx_col_high(x));
-        }
-
-        template <typename T> __device__ __forceinline__ D at(int y, int x, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(((const T*)((const char*)data + idx_row(y) * step))[idx_col(x)]);
-        }
-
-        template <typename Ptr2D> __device__ __forceinline__ D at(typename Ptr2D::index_type y, typename Ptr2D::index_type x, const Ptr2D& src) const
-        {
-            return saturate_cast<D>(src(idx_row(y), idx_col(x)));
-        }
-
-        int last_row;
-        int last_col;
-    };
-
-    //////////////////////////////////////////////////////////////
-    // BrdWrap
-
-    template <typename D> struct BrdRowWrap
-    {
-        typedef D result_type;
-
-        explicit __host__ __device__ __forceinline__ BrdRowWrap(int width_) : width(width_) {}
-        template <typename U> __host__ __device__ __forceinline__ BrdRowWrap(int width_, U) : width(width_) {}
-
-        __device__ __forceinline__ int idx_col_low(int x) const
-        {
-            return (x >= 0) * x + (x < 0) * (x - ((x - width + 1) / width) * width);
-        }
-
-        __device__ __forceinline__ int idx_col_high(int x) const
-        {
-            return (x < width) * x + (x >= width) * (x % width);
-        }
-
-        __device__ __forceinline__ int idx_col(int x) const
-        {
-            return idx_col_high(idx_col_low(x));
-        }
-
-        template <typename T> __device__ __forceinline__ D at_low(int x, const T* data) const
-        {
-            return saturate_cast<D>(data[idx_col_low(x)]);
-        }
-
-        template <typename T> __device__ __forceinline__ D at_high(int x, const T* data) const
-        {
-            return saturate_cast<D>(data[idx_col_high(x)]);
-        }
-
-        template <typename T> __device__ __forceinline__ D at(int x, const T* data) const
-        {
-            return saturate_cast<D>(data[idx_col(x)]);
-        }
-
-        int width;
-    };
-
-    template <typename D> struct BrdColWrap
-    {
-        typedef D result_type;
-
-        explicit __host__ __device__ __forceinline__ BrdColWrap(int height_) : height(height_) {}
-        template <typename U> __host__ __device__ __forceinline__ BrdColWrap(int height_, U) : height(height_) {}
-
-        __device__ __forceinline__ int idx_row_low(int y) const
-        {
-            return (y >= 0) * y + (y < 0) * (y - ((y - height + 1) / height) * height);
-        }
-
-        __device__ __forceinline__ int idx_row_high(int y) const
-        {
-            return (y < height) * y + (y >= height) * (y % height);
-        }
-
-        __device__ __forceinline__ int idx_row(int y) const
-        {
-            return idx_row_high(idx_row_low(y));
-        }
-
-        template <typename T> __device__ __forceinline__ D at_low(int y, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(*(const D*)((const char*)data + idx_row_low(y) * step));
-        }
-
-        template <typename T> __device__ __forceinline__ D at_high(int y, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(*(const D*)((const char*)data + idx_row_high(y) * step));
-        }
-
-        template <typename T> __device__ __forceinline__ D at(int y, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(*(const D*)((const char*)data + idx_row(y) * step));
-        }
-
-        int height;
-    };
-
-    template <typename D> struct BrdWrap
-    {
-        typedef D result_type;
-
-        __host__ __device__ __forceinline__ BrdWrap(int height_, int width_) :
-            height(height_), width(width_)
-        {
-        }
-        template <typename U>
-        __host__ __device__ __forceinline__ BrdWrap(int height_, int width_, U) :
-            height(height_), width(width_)
-        {
-        }
-
-        __device__ __forceinline__ int idx_row_low(int y) const
-        {
-            return (y >= 0) ? y : (y - ((y - height + 1) / height) * height);
-        }
-
-        __device__ __forceinline__ int idx_row_high(int y) const
-        {
-            return (y < height) ? y : (y % height);
-        }
-
-        __device__ __forceinline__ int idx_row(int y) const
-        {
-            return idx_row_high(idx_row_low(y));
-        }
-
-        __device__ __forceinline__ int idx_col_low(int x) const
-        {
-            return (x >= 0) ? x : (x - ((x - width + 1) / width) * width);
-        }
-
-        __device__ __forceinline__ int idx_col_high(int x) const
-        {
-            return (x < width) ? x : (x % width);
-        }
-
-        __device__ __forceinline__ int idx_col(int x) const
-        {
-            return idx_col_high(idx_col_low(x));
-        }
-
-        template <typename T> __device__ __forceinline__ D at(int y, int x, const T* data, size_t step) const
-        {
-            return saturate_cast<D>(((const T*)((const char*)data + idx_row(y) * step))[idx_col(x)]);
-        }
-
-        template <typename Ptr2D> __device__ __forceinline__ D at(typename Ptr2D::index_type y, typename Ptr2D::index_type x, const Ptr2D& src) const
-        {
-            return saturate_cast<D>(src(idx_row(y), idx_col(x)));
-        }
-
-        int height;
-        int width;
-    };
-
-    //////////////////////////////////////////////////////////////
-    // BorderReader
-
-    template <typename Ptr2D, typename B> struct BorderReader
-    {
-        typedef typename B::result_type elem_type;
-        typedef typename Ptr2D::index_type index_type;
-
-        __host__ __device__ __forceinline__ BorderReader(const Ptr2D& ptr_, const B& b_) : ptr(ptr_), b(b_) {}
-
-        __device__ __forceinline__ elem_type operator ()(index_type y, index_type x) const
-        {
-            return b.at(y, x, ptr);
-        }
-
-        Ptr2D ptr;
-        B b;
-    };
-
-    // under win32 there is some bug with templated types that passed as kernel parameters
-    // with this specialization all works fine
-    template <typename Ptr2D, typename D> struct BorderReader< Ptr2D, BrdConstant<D> >
-    {
-        typedef typename BrdConstant<D>::result_type elem_type;
-        typedef typename Ptr2D::index_type index_type;
-
-        __host__ __device__ __forceinline__ BorderReader(const Ptr2D& src_, const BrdConstant<D>& b) :
-            src(src_), height(b.height), width(b.width), val(b.val)
-        {
-        }
-
-        __device__ __forceinline__ D operator ()(index_type y, index_type x) const
-        {
-            return (x >= 0 && x < width && y >= 0 && y < height) ? saturate_cast<D>(src(y, x)) : val;
-        }
-
-        Ptr2D src;
-        int height;
-        int width;
-        D val;
-    };
-}}} // namespace cv { namespace cuda { namespace cudev
-
-//! @endcond
-
-#endif // OPENCV_CUDA_BORDER_INTERPOLATE_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/color.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/color.hpp
deleted file mode 100644
index dcce2802..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/color.hpp
+++ /dev/null
@@ -1,309 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_COLOR_HPP
-#define OPENCV_CUDA_COLOR_HPP
-
-#include "detail/color_detail.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    // All OPENCV_CUDA_IMPLEMENT_*_TRAITS(ColorSpace1_to_ColorSpace2, ...) macros implements
-    // template <typename T> class ColorSpace1_to_ColorSpace2_traits
-    // {
-    //     typedef ... functor_type;
-    //     static __host__ __device__ functor_type create_functor();
-    // };
-
-    OPENCV_CUDA_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_rgb, 3, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_bgra, 3, 4, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2RGB_TRAITS(bgr_to_rgba, 3, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_bgr, 4, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_rgb, 4, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2RGB_TRAITS(bgra_to_rgba, 4, 4, 2)
-
-    #undef OPENCV_CUDA_IMPLEMENT_RGB2RGB_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(bgr_to_bgr555, 3, 0, 5)
-    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(bgr_to_bgr565, 3, 0, 6)
-    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(rgb_to_bgr555, 3, 2, 5)
-    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(rgb_to_bgr565, 3, 2, 6)
-    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(bgra_to_bgr555, 4, 0, 5)
-    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(bgra_to_bgr565, 4, 0, 6)
-    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(rgba_to_bgr555, 4, 2, 5)
-    OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(rgba_to_bgr565, 4, 2, 6)
-
-    #undef OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_rgb, 3, 2, 5)
-    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_rgb, 3, 2, 6)
-    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_bgr, 3, 0, 5)
-    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_bgr, 3, 0, 6)
-    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_rgba, 4, 2, 5)
-    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_rgba, 4, 2, 6)
-    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr555_to_bgra, 4, 0, 5)
-    OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(bgr565_to_bgra, 4, 0, 6)
-
-    #undef OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_GRAY2RGB_TRAITS(gray_to_bgr, 3)
-    OPENCV_CUDA_IMPLEMENT_GRAY2RGB_TRAITS(gray_to_bgra, 4)
-
-    #undef OPENCV_CUDA_IMPLEMENT_GRAY2RGB_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_GRAY2RGB5x5_TRAITS(gray_to_bgr555, 5)
-    OPENCV_CUDA_IMPLEMENT_GRAY2RGB5x5_TRAITS(gray_to_bgr565, 6)
-
-    #undef OPENCV_CUDA_IMPLEMENT_GRAY2RGB5x5_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_RGB5x52GRAY_TRAITS(bgr555_to_gray, 5)
-    OPENCV_CUDA_IMPLEMENT_RGB5x52GRAY_TRAITS(bgr565_to_gray, 6)
-
-    #undef OPENCV_CUDA_IMPLEMENT_RGB5x52GRAY_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_RGB2GRAY_TRAITS(rgb_to_gray, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2GRAY_TRAITS(bgr_to_gray, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2GRAY_TRAITS(rgba_to_gray, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2GRAY_TRAITS(bgra_to_gray, 4, 0)
-
-    #undef OPENCV_CUDA_IMPLEMENT_RGB2GRAY_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(rgb_to_yuv, 3, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(rgba_to_yuv, 4, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(rgb_to_yuv4, 3, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(rgba_to_yuv4, 4, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(bgr_to_yuv, 3, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(bgra_to_yuv, 4, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(bgr_to_yuv4, 3, 4, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(bgra_to_yuv4, 4, 4, 0)
-
-    #undef OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_rgb, 3, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_rgba, 3, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_rgb, 4, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_rgba, 4, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_bgr, 3, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv_to_bgra, 3, 4, 0)
-    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_bgr, 4, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(yuv4_to_bgra, 4, 4, 0)
-
-    #undef OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(rgb_to_YCrCb, 3, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(rgba_to_YCrCb, 4, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(rgb_to_YCrCb4, 3, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(rgba_to_YCrCb4, 4, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(bgr_to_YCrCb, 3, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(bgra_to_YCrCb, 4, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(bgr_to_YCrCb4, 3, 4, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(bgra_to_YCrCb4, 4, 4, 0)
-
-    #undef OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_rgb, 3, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_rgba, 3, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_rgb, 4, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_rgba, 4, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_bgr, 3, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb_to_bgra, 3, 4, 0)
-    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_bgr, 4, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(YCrCb4_to_bgra, 4, 4, 0)
-
-    #undef OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(rgb_to_xyz, 3, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(rgba_to_xyz, 4, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(rgb_to_xyz4, 3, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(rgba_to_xyz4, 4, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(bgr_to_xyz, 3, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(bgra_to_xyz, 4, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(bgr_to_xyz4, 3, 4, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(bgra_to_xyz4, 4, 4, 0)
-
-    #undef OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_rgb, 3, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_rgb, 4, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_rgba, 3, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_rgba, 4, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_bgr, 3, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_bgr, 4, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz_to_bgra, 3, 4, 0)
-    OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(xyz4_to_bgra, 4, 4, 0)
-
-    #undef OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(rgb_to_hsv, 3, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(rgba_to_hsv, 4, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(rgb_to_hsv4, 3, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(rgba_to_hsv4, 4, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(bgr_to_hsv, 3, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(bgra_to_hsv, 4, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(bgr_to_hsv4, 3, 4, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(bgra_to_hsv4, 4, 4, 0)
-
-    #undef OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_rgb, 3, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_rgba, 3, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_rgb, 4, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_rgba, 4, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_bgr, 3, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv_to_bgra, 3, 4, 0)
-    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_bgr, 4, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(hsv4_to_bgra, 4, 4, 0)
-
-    #undef OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(rgb_to_hls, 3, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(rgba_to_hls, 4, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(rgb_to_hls4, 3, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(rgba_to_hls4, 4, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(bgr_to_hls, 3, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(bgra_to_hls, 4, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(bgr_to_hls4, 3, 4, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(bgra_to_hls4, 4, 4, 0)
-
-    #undef OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls_to_rgb, 3, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls_to_rgba, 3, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_rgb, 4, 3, 2)
-    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_rgba, 4, 4, 2)
-    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls_to_bgr, 3, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls_to_bgra, 3, 4, 0)
-    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_bgr, 4, 3, 0)
-    OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(hls4_to_bgra, 4, 4, 0)
-
-    #undef OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(rgb_to_lab, 3, 3, true, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(rgba_to_lab, 4, 3, true, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(rgb_to_lab4, 3, 4, true, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(rgba_to_lab4, 4, 4, true, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(bgr_to_lab, 3, 3, true, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(bgra_to_lab, 4, 3, true, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(bgr_to_lab4, 3, 4, true, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(bgra_to_lab4, 4, 4, true, 0)
-
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lrgb_to_lab, 3, 3, false, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lrgba_to_lab, 4, 3, false, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lrgb_to_lab4, 3, 4, false, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lrgba_to_lab4, 4, 4, false, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lbgr_to_lab, 3, 3, false, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lbgra_to_lab, 4, 3, false, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lbgr_to_lab4, 3, 4, false, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(lbgra_to_lab4, 4, 4, false, 0)
-
-    #undef OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_rgb, 3, 3, true, 2)
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_rgb, 4, 3, true, 2)
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_rgba, 3, 4, true, 2)
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_rgba, 4, 4, true, 2)
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_bgr, 3, 3, true, 0)
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_bgr, 4, 3, true, 0)
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_bgra, 3, 4, true, 0)
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_bgra, 4, 4, true, 0)
-
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_lrgb, 3, 3, false, 2)
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_lrgb, 4, 3, false, 2)
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_lrgba, 3, 4, false, 2)
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_lrgba, 4, 4, false, 2)
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_lbgr, 3, 3, false, 0)
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_lbgr, 4, 3, false, 0)
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab_to_lbgra, 3, 4, false, 0)
-    OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(lab4_to_lbgra, 4, 4, false, 0)
-
-    #undef OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(rgb_to_luv, 3, 3, true, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(rgba_to_luv, 4, 3, true, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(rgb_to_luv4, 3, 4, true, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(rgba_to_luv4, 4, 4, true, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(bgr_to_luv, 3, 3, true, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(bgra_to_luv, 4, 3, true, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(bgr_to_luv4, 3, 4, true, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(bgra_to_luv4, 4, 4, true, 0)
-
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lrgb_to_luv, 3, 3, false, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lrgba_to_luv, 4, 3, false, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lrgb_to_luv4, 3, 4, false, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lrgba_to_luv4, 4, 4, false, 2)
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lbgr_to_luv, 3, 3, false, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lbgra_to_luv, 4, 3, false, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lbgr_to_luv4, 3, 4, false, 0)
-    OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(lbgra_to_luv4, 4, 4, false, 0)
-
-    #undef OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS
-
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_rgb, 3, 3, true, 2)
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_rgb, 4, 3, true, 2)
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_rgba, 3, 4, true, 2)
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_rgba, 4, 4, true, 2)
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_bgr, 3, 3, true, 0)
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_bgr, 4, 3, true, 0)
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_bgra, 3, 4, true, 0)
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_bgra, 4, 4, true, 0)
-
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_lrgb, 3, 3, false, 2)
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_lrgb, 4, 3, false, 2)
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_lrgba, 3, 4, false, 2)
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_lrgba, 4, 4, false, 2)
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_lbgr, 3, 3, false, 0)
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_lbgr, 4, 3, false, 0)
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv_to_lbgra, 3, 4, false, 0)
-    OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(luv4_to_lbgra, 4, 4, false, 0)
-
-    #undef OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS
-}}} // namespace cv { namespace cuda { namespace cudev
-
-//! @endcond
-
-#endif // OPENCV_CUDA_COLOR_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/common.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/common.hpp
deleted file mode 100644
index 80b2ff08..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/common.hpp
+++ /dev/null
@@ -1,123 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_COMMON_HPP
-#define OPENCV_CUDA_COMMON_HPP
-
-#include <cuda_runtime.h>
-#include "opencv2/core/cuda_types.hpp"
-#include "opencv2/core/cvdef.h"
-#include "opencv2/core/base.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-#ifndef CV_PI_F
-    #ifndef CV_PI
-        #define CV_PI_F 3.14159265f
-    #else
-        #define CV_PI_F ((float)CV_PI)
-    #endif
-#endif
-
-namespace cv { namespace cuda {
-    static inline void checkCudaError(cudaError_t err, const char* file, const int line, const char* func)
-    {
-        if (cudaSuccess != err)
-            cv::error(cv::Error::GpuApiCallError, cudaGetErrorString(err), func, file, line);
-    }
-}}
-
-#ifndef cudaSafeCall
-    #define cudaSafeCall(expr)  cv::cuda::checkCudaError(expr, __FILE__, __LINE__, CV_Func)
-#endif
-
-namespace cv { namespace cuda
-{
-    template <typename T> static inline bool isAligned(const T* ptr, size_t size)
-    {
-        return reinterpret_cast<size_t>(ptr) % size == 0;
-    }
-
-    static inline bool isAligned(size_t step, size_t size)
-    {
-        return step % size == 0;
-    }
-}}
-
-namespace cv { namespace cuda
-{
-    namespace device
-    {
-        __host__ __device__ __forceinline__ int divUp(int total, int grain)
-        {
-            return (total + grain - 1) / grain;
-        }
-
-        template<class T> inline void bindTexture(const textureReference* tex, const PtrStepSz<T>& img)
-        {
-            cudaChannelFormatDesc desc = cudaCreateChannelDesc<T>();
-            cudaSafeCall( cudaBindTexture2D(0, tex, img.ptr(), &desc, img.cols, img.rows, img.step) );
-        }
-
-        template<class T> inline void createTextureObjectPitch2D(cudaTextureObject_t* tex, PtrStepSz<T>& img, const cudaTextureDesc& texDesc)
-        {
-            cudaResourceDesc resDesc;
-            memset(&resDesc, 0, sizeof(resDesc));
-            resDesc.resType = cudaResourceTypePitch2D;
-            resDesc.res.pitch2D.devPtr = static_cast<void*>(img.ptr());
-            resDesc.res.pitch2D.height = img.rows;
-            resDesc.res.pitch2D.width = img.cols;
-            resDesc.res.pitch2D.pitchInBytes = img.step;
-            resDesc.res.pitch2D.desc = cudaCreateChannelDesc<T>();
-
-            cudaSafeCall( cudaCreateTextureObject(tex, &resDesc, &texDesc, NULL) );
-        }
-    }
-}}
-
-//! @endcond
-
-#endif // OPENCV_CUDA_COMMON_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/datamov_utils.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/datamov_utils.hpp
deleted file mode 100644
index 6820d0fd..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/datamov_utils.hpp
+++ /dev/null
@@ -1,113 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_DATAMOV_UTILS_HPP
-#define OPENCV_CUDA_DATAMOV_UTILS_HPP
-
-#include "common.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 200
-
-        // for Fermi memory space is detected automatically
-        template <typename T> struct ForceGlob
-        {
-            __device__ __forceinline__ static void Load(const T* ptr, int offset, T& val)  { val = ptr[offset];  }
-        };
-
-    #else // __CUDA_ARCH__ >= 200
-
-        #if defined(_WIN64) || defined(__LP64__)
-            // 64-bit register modifier for inlined asm
-            #define OPENCV_CUDA_ASM_PTR "l"
-        #else
-            // 32-bit register modifier for inlined asm
-            #define OPENCV_CUDA_ASM_PTR "r"
-        #endif
-
-        template<class T> struct ForceGlob;
-
-        #define OPENCV_CUDA_DEFINE_FORCE_GLOB(base_type, ptx_type, reg_mod) \
-            template <> struct ForceGlob<base_type> \
-            { \
-                __device__ __forceinline__ static void Load(const base_type* ptr, int offset, base_type& val) \
-                { \
-                    asm("ld.global."#ptx_type" %0, [%1];" : "="#reg_mod(val) : OPENCV_CUDA_ASM_PTR(ptr + offset)); \
-                } \
-            };
-
-        #define OPENCV_CUDA_DEFINE_FORCE_GLOB_B(base_type, ptx_type) \
-            template <> struct ForceGlob<base_type> \
-            { \
-                __device__ __forceinline__ static void Load(const base_type* ptr, int offset, base_type& val) \
-                { \
-                    asm("ld.global."#ptx_type" %0, [%1];" : "=r"(*reinterpret_cast<uint*>(&val)) : OPENCV_CUDA_ASM_PTR(ptr + offset)); \
-                } \
-            };
-
-            OPENCV_CUDA_DEFINE_FORCE_GLOB_B(uchar,  u8)
-            OPENCV_CUDA_DEFINE_FORCE_GLOB_B(schar,  s8)
-            OPENCV_CUDA_DEFINE_FORCE_GLOB_B(char,   b8)
-            OPENCV_CUDA_DEFINE_FORCE_GLOB  (ushort, u16, h)
-            OPENCV_CUDA_DEFINE_FORCE_GLOB  (short,  s16, h)
-            OPENCV_CUDA_DEFINE_FORCE_GLOB  (uint,   u32, r)
-            OPENCV_CUDA_DEFINE_FORCE_GLOB  (int,    s32, r)
-            OPENCV_CUDA_DEFINE_FORCE_GLOB  (float,  f32, f)
-            OPENCV_CUDA_DEFINE_FORCE_GLOB  (double, f64, d)
-
-        #undef OPENCV_CUDA_DEFINE_FORCE_GLOB
-        #undef OPENCV_CUDA_DEFINE_FORCE_GLOB_B
-        #undef OPENCV_CUDA_ASM_PTR
-
-    #endif // __CUDA_ARCH__ >= 200
-}}} // namespace cv { namespace cuda { namespace cudev
-
-//! @endcond
-
-#endif // OPENCV_CUDA_DATAMOV_UTILS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/color_detail.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/color_detail.hpp
deleted file mode 100644
index f4b47965..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/color_detail.hpp
+++ /dev/null
@@ -1,2018 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_COLOR_DETAIL_HPP
-#define OPENCV_CUDA_COLOR_DETAIL_HPP
-
-#include "../common.hpp"
-#include "../vec_traits.hpp"
-#include "../saturate_cast.hpp"
-#include "../limits.hpp"
-#include "../functional.hpp"
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    #ifndef CV_DESCALE
-        #define CV_DESCALE(x, n) (((x) + (1 << ((n)-1))) >> (n))
-    #endif
-
-    namespace color_detail
-    {
-        template<typename T> struct ColorChannel
-        {
-            typedef float worktype_f;
-            static __device__ __forceinline__ T max() { return numeric_limits<T>::max(); }
-            static __device__ __forceinline__ T half() { return (T)(max()/2 + 1); }
-        };
-
-        template<> struct ColorChannel<float>
-        {
-            typedef float worktype_f;
-            static __device__ __forceinline__ float max() { return 1.f; }
-            static __device__ __forceinline__ float half() { return 0.5f; }
-        };
-
-        template <typename T> static __device__ __forceinline__ void setAlpha(typename TypeVec<T, 3>::vec_type& vec, T val)
-        {
-        }
-
-        template <typename T> static __device__ __forceinline__ void setAlpha(typename TypeVec<T, 4>::vec_type& vec, T val)
-        {
-            vec.w = val;
-        }
-
-        template <typename T> static __device__ __forceinline__ T getAlpha(const typename TypeVec<T, 3>::vec_type& vec)
-        {
-            return ColorChannel<T>::max();
-        }
-
-        template <typename T> static __device__ __forceinline__ T getAlpha(const typename TypeVec<T, 4>::vec_type& vec)
-        {
-            return vec.w;
-        }
-
-        //constants for conversion from/to RGB and Gray, YUV, YCrCb according to BT.601
-        constexpr float B2YF = 0.114f;
-        constexpr float G2YF = 0.587f;
-        constexpr float R2YF = 0.299f;
-
-        //to YCbCr
-        constexpr float YCBF = 0.564f; // == 1/2/(1-B2YF)
-        constexpr float YCRF = 0.713f; // == 1/2/(1-R2YF)
-        const     int   YCBI = 9241;  // == YCBF*16384
-        const     int   YCRI = 11682; // == YCRF*16384
-        //to YUV
-        constexpr float B2UF = 0.492f;
-        constexpr float R2VF = 0.877f;
-        const     int   B2UI = 8061;  // == B2UF*16384
-        const     int   R2VI = 14369; // == R2VF*16384
-        //from YUV
-        constexpr float U2BF = 2.032f;
-        constexpr float U2GF = -0.395f;
-        constexpr float V2GF = -0.581f;
-        constexpr float V2RF = 1.140f;
-        const     int   U2BI = 33292;
-        const     int   U2GI = -6472;
-        const     int   V2GI = -9519;
-        const     int   V2RI = 18678;
-        //from YCrCb
-        constexpr float CB2BF = 1.773f;
-        constexpr float CB2GF = -0.344f;
-        constexpr float CR2GF = -0.714f;
-        constexpr float CR2RF = 1.403f;
-        const     int   CB2BI = 29049;
-        const     int   CB2GI = -5636;
-        const     int   CR2GI = -11698;
-        const     int   CR2RI = 22987;
-
-        enum
-        {
-            yuv_shift  = 14,
-            xyz_shift  = 12,
-            gray_shift = 15,
-            R2Y        = 4899,
-            G2Y        = 9617,
-            B2Y        = 1868,
-            RY15 =  9798, // == R2YF*32768 + 0.5
-            GY15 = 19235, // == G2YF*32768 + 0.5
-            BY15 =  3735, // == B2YF*32768 + 0.5
-            BLOCK_SIZE = 256
-        };
-    }
-
-////////////////// Various 3/4-channel to 3/4-channel RGB transformations /////////////////
-
-    namespace color_detail
-    {
-        template <typename T, int scn, int dcn, int bidx> struct RGB2RGB
-            : unary_function<typename TypeVec<T, scn>::vec_type, typename TypeVec<T, dcn>::vec_type>
-        {
-            __device__ typename TypeVec<T, dcn>::vec_type operator()(const typename TypeVec<T, scn>::vec_type& src) const
-            {
-                typename TypeVec<T, dcn>::vec_type dst;
-
-                dst.x = (&src.x)[bidx];
-                dst.y = src.y;
-                dst.z = (&src.x)[bidx^2];
-                setAlpha(dst, getAlpha<T>(src));
-
-                return dst;
-            }
-
-            __host__ __device__ __forceinline__ RGB2RGB() {}
-            __host__ __device__ __forceinline__ RGB2RGB(const RGB2RGB&) {}
-        };
-
-        template <> struct RGB2RGB<uchar, 4, 4, 2> : unary_function<uint, uint>
-        {
-            __device__ uint operator()(uint src) const
-            {
-                uint dst = 0;
-
-                dst |= (0xffu & (src >> 16));
-                dst |= (0xffu & (src >> 8)) << 8;
-                dst |= (0xffu & (src)) << 16;
-                dst |= (0xffu & (src >> 24)) << 24;
-
-                return dst;
-            }
-
-            __host__ __device__ __forceinline__ RGB2RGB() {}
-            __host__ __device__ __forceinline__ RGB2RGB(const RGB2RGB&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_RGB2RGB_TRAITS(name, scn, dcn, bidx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2RGB<T, scn, dcn, bidx> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-/////////// Transforming 16-bit (565 or 555) RGB to/from 24/32-bit (888[8]) RGB //////////
-
-    namespace color_detail
-    {
-        template <int green_bits, int bidx> struct RGB2RGB5x5Converter;
-        template<int bidx> struct RGB2RGB5x5Converter<6, bidx>
-        {
-            static __device__ __forceinline__ ushort cvt(const uchar3& src)
-            {
-                return (ushort)(((&src.x)[bidx] >> 3) | ((src.y & ~3) << 3) | (((&src.x)[bidx^2] & ~7) << 8));
-            }
-
-            static __device__ __forceinline__ ushort cvt(uint src)
-            {
-                uint b = 0xffu & (src >> (bidx * 8));
-                uint g = 0xffu & (src >> 8);
-                uint r = 0xffu & (src >> ((bidx ^ 2) * 8));
-                return (ushort)((b >> 3) | ((g & ~3) << 3) | ((r & ~7) << 8));
-            }
-        };
-
-        template<int bidx> struct RGB2RGB5x5Converter<5, bidx>
-        {
-            static __device__ __forceinline__ ushort cvt(const uchar3& src)
-            {
-                return (ushort)(((&src.x)[bidx] >> 3) | ((src.y & ~7) << 2) | (((&src.x)[bidx^2] & ~7) << 7));
-            }
-
-            static __device__ __forceinline__ ushort cvt(uint src)
-            {
-                uint b = 0xffu & (src >> (bidx * 8));
-                uint g = 0xffu & (src >> 8);
-                uint r = 0xffu & (src >> ((bidx ^ 2) * 8));
-                uint a = 0xffu & (src >> 24);
-                return (ushort)((b >> 3) | ((g & ~7) << 2) | ((r & ~7) << 7) | (a * 0x8000));
-            }
-        };
-
-        template<int scn, int bidx, int green_bits> struct RGB2RGB5x5;
-
-        template<int bidx, int green_bits> struct RGB2RGB5x5<3, bidx,green_bits> : unary_function<uchar3, ushort>
-        {
-            __device__ __forceinline__ ushort operator()(const uchar3& src) const
-            {
-                return RGB2RGB5x5Converter<green_bits, bidx>::cvt(src);
-            }
-
-            __host__ __device__ __forceinline__ RGB2RGB5x5() {}
-            __host__ __device__ __forceinline__ RGB2RGB5x5(const RGB2RGB5x5&) {}
-        };
-
-        template<int bidx, int green_bits> struct RGB2RGB5x5<4, bidx,green_bits> : unary_function<uint, ushort>
-        {
-            __device__ __forceinline__ ushort operator()(uint src) const
-            {
-                return RGB2RGB5x5Converter<green_bits, bidx>::cvt(src);
-            }
-
-            __host__ __device__ __forceinline__ RGB2RGB5x5() {}
-            __host__ __device__ __forceinline__ RGB2RGB5x5(const RGB2RGB5x5&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_RGB2RGB5x5_TRAITS(name, scn, bidx, green_bits) \
-    struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2RGB5x5<scn, bidx, green_bits> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-    namespace color_detail
-    {
-        template <int green_bits, int bidx> struct RGB5x52RGBConverter;
-
-        template <int bidx> struct RGB5x52RGBConverter<5, bidx>
-        {
-            static __device__ __forceinline__ void cvt(uint src, uchar3& dst)
-            {
-                (&dst.x)[bidx] = src << 3;
-                dst.y = (src >> 2) & ~7;
-                (&dst.x)[bidx ^ 2] = (src >> 7) & ~7;
-            }
-
-            static __device__ __forceinline__ void cvt(uint src, uint& dst)
-            {
-                dst = 0;
-
-                dst |= (0xffu & (src << 3)) << (bidx * 8);
-                dst |= (0xffu & ((src >> 2) & ~7)) << 8;
-                dst |= (0xffu & ((src >> 7) & ~7)) << ((bidx ^ 2) * 8);
-                dst |= ((src & 0x8000) * 0xffu) << 24;
-            }
-        };
-
-        template <int bidx> struct RGB5x52RGBConverter<6, bidx>
-        {
-            static __device__ __forceinline__ void cvt(uint src, uchar3& dst)
-            {
-                (&dst.x)[bidx] = src << 3;
-                dst.y = (src >> 3) & ~3;
-                (&dst.x)[bidx ^ 2] = (src >> 8) & ~7;
-            }
-
-            static __device__ __forceinline__ void cvt(uint src, uint& dst)
-            {
-                dst = 0xffu << 24;
-
-                dst |= (0xffu & (src << 3)) << (bidx * 8);
-                dst |= (0xffu &((src >> 3) & ~3)) << 8;
-                dst |= (0xffu & ((src >> 8) & ~7)) << ((bidx ^ 2) * 8);
-            }
-        };
-
-        template <int dcn, int bidx, int green_bits> struct RGB5x52RGB;
-
-        template <int bidx, int green_bits> struct RGB5x52RGB<3, bidx, green_bits> : unary_function<ushort, uchar3>
-        {
-            __device__ __forceinline__ uchar3 operator()(ushort src) const
-            {
-                uchar3 dst;
-                RGB5x52RGBConverter<green_bits, bidx>::cvt(src, dst);
-                return dst;
-            }
-            __host__ __device__ __forceinline__ RGB5x52RGB() {}
-            __host__ __device__ __forceinline__ RGB5x52RGB(const RGB5x52RGB&) {}
-
-        };
-
-        template <int bidx, int green_bits> struct RGB5x52RGB<4, bidx, green_bits> : unary_function<ushort, uint>
-        {
-            __device__ __forceinline__ uint operator()(ushort src) const
-            {
-                uint dst;
-                RGB5x52RGBConverter<green_bits, bidx>::cvt(src, dst);
-                return dst;
-            }
-            __host__ __device__ __forceinline__ RGB5x52RGB() {}
-            __host__ __device__ __forceinline__ RGB5x52RGB(const RGB5x52RGB&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_RGB5x52RGB_TRAITS(name, dcn, bidx, green_bits) \
-    struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB5x52RGB<dcn, bidx, green_bits> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-///////////////////////////////// Grayscale to Color ////////////////////////////////
-
-    namespace color_detail
-    {
-        template <typename T, int dcn> struct Gray2RGB : unary_function<T, typename TypeVec<T, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<T, dcn>::vec_type operator()(T src) const
-            {
-                typename TypeVec<T, dcn>::vec_type dst;
-
-                dst.z = dst.y = dst.x = src;
-                setAlpha(dst, ColorChannel<T>::max());
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ Gray2RGB() {}
-            __host__ __device__ __forceinline__ Gray2RGB(const Gray2RGB&) {}
-        };
-
-        template <> struct Gray2RGB<uchar, 4> : unary_function<uchar, uint>
-        {
-            __device__ __forceinline__ uint operator()(uint src) const
-            {
-                uint dst = 0xffu << 24;
-
-                dst |= src;
-                dst |= src << 8;
-                dst |= src << 16;
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ Gray2RGB() {}
-            __host__ __device__ __forceinline__ Gray2RGB(const Gray2RGB&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_GRAY2RGB_TRAITS(name, dcn) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::Gray2RGB<T, dcn> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-    namespace color_detail
-    {
-        template <int green_bits> struct Gray2RGB5x5Converter;
-        template<> struct Gray2RGB5x5Converter<6>
-        {
-            static __device__ __forceinline__ ushort cvt(uint t)
-            {
-                return (ushort)((t >> 3) | ((t & ~3) << 3) | ((t & ~7) << 8));
-            }
-        };
-
-        template<> struct Gray2RGB5x5Converter<5>
-        {
-            static __device__ __forceinline__ ushort cvt(uint t)
-            {
-                t >>= 3;
-                return (ushort)(t | (t << 5) | (t << 10));
-            }
-        };
-
-        template<int green_bits> struct Gray2RGB5x5 : unary_function<uchar, ushort>
-        {
-            __device__ __forceinline__ ushort operator()(uint src) const
-            {
-                return Gray2RGB5x5Converter<green_bits>::cvt(src);
-            }
-
-            __host__ __device__ __forceinline__ Gray2RGB5x5() {}
-            __host__ __device__ __forceinline__ Gray2RGB5x5(const Gray2RGB5x5&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_GRAY2RGB5x5_TRAITS(name, green_bits) \
-    struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::Gray2RGB5x5<green_bits> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-///////////////////////////////// Color to Grayscale ////////////////////////////////
-
-    namespace color_detail
-    {
-        template <int green_bits> struct RGB5x52GrayConverter;
-        template <> struct RGB5x52GrayConverter<6>
-        {
-            static __device__ __forceinline__ uchar cvt(uint t)
-            {
-                return (uchar)CV_DESCALE(((t << 3) & 0xf8) * BY15 + ((t >> 3) & 0xfc) * GY15 + ((t >> 8) & 0xf8) * RY15, gray_shift);
-            }
-        };
-
-        template <> struct RGB5x52GrayConverter<5>
-        {
-            static __device__ __forceinline__ uchar cvt(uint t)
-            {
-                return (uchar)CV_DESCALE(((t << 3) & 0xf8) * BY15 + ((t >> 2) & 0xf8) * GY15 + ((t >> 7) & 0xf8) * RY15, gray_shift);
-            }
-        };
-
-        template<int green_bits> struct RGB5x52Gray : unary_function<ushort, uchar>
-        {
-            __device__ __forceinline__ uchar operator()(uint src) const
-            {
-                return RGB5x52GrayConverter<green_bits>::cvt(src);
-            }
-            __host__ __device__ __forceinline__ RGB5x52Gray() {}
-            __host__ __device__ __forceinline__ RGB5x52Gray(const RGB5x52Gray&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_RGB5x52GRAY_TRAITS(name, green_bits) \
-    struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB5x52Gray<green_bits> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-    namespace color_detail
-    {
-        template <int bidx, typename T> static __device__ __forceinline__ T RGB2GrayConvert(const T* src)
-        {
-            return (T)CV_DESCALE((unsigned)(src[bidx] * BY15 + src[1] * GY15 + src[bidx^2] * RY15), gray_shift);
-        }
-
-        template <int bidx> static __device__ __forceinline__ uchar RGB2GrayConvert(uint src)
-        {
-            uint b = 0xffu & (src >> (bidx * 8));
-            uint g = 0xffu & (src >> 8);
-            uint r = 0xffu & (src >> ((bidx ^ 2) * 8));
-            return CV_DESCALE((uint)(b * BY15 + g * GY15 + r * RY15), gray_shift);
-        }
-
-        template <int bidx> static __device__ __forceinline__ float RGB2GrayConvert(const float* src)
-        {
-            return src[bidx] * B2YF + src[1] * G2YF + src[bidx^2] * R2YF;
-        }
-
-        template <typename T, int scn, int bidx> struct RGB2Gray : unary_function<typename TypeVec<T, scn>::vec_type, T>
-        {
-            __device__ __forceinline__ T operator()(const typename TypeVec<T, scn>::vec_type& src) const
-            {
-                return RGB2GrayConvert<bidx>(&src.x);
-            }
-            __host__ __device__ __forceinline__ RGB2Gray() {}
-            __host__ __device__ __forceinline__ RGB2Gray(const RGB2Gray&) {}
-        };
-
-        template <int bidx> struct RGB2Gray<uchar, 4, bidx> : unary_function<uint, uchar>
-        {
-            __device__ __forceinline__ uchar operator()(uint src) const
-            {
-                return RGB2GrayConvert<bidx>(src);
-            }
-            __host__ __device__ __forceinline__ RGB2Gray() {}
-            __host__ __device__ __forceinline__ RGB2Gray(const RGB2Gray&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_RGB2GRAY_TRAITS(name, scn, bidx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2Gray<T, scn, bidx> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-///////////////////////////////////// RGB <-> YUV //////////////////////////////////////
-
-    namespace color_detail
-    {
-        __constant__ float c_RGB2YUVCoeffs_f[5] = { B2YF, G2YF, R2YF, B2UF, R2VF };
-        __constant__ int   c_RGB2YUVCoeffs_i[5] = { B2Y, G2Y, R2Y, B2UI, R2VI };
-
-        template <int bidx, typename T, typename D> static __device__ void RGB2YUVConvert(const T* src, D& dst)
-        {
-            const int delta = ColorChannel<T>::half() * (1 << yuv_shift);
-
-            const int Y = CV_DESCALE(src[0] * c_RGB2YUVCoeffs_i[bidx^2] + src[1] * c_RGB2YUVCoeffs_i[1] + src[2] * c_RGB2YUVCoeffs_i[bidx], yuv_shift);
-            const int Cr = CV_DESCALE((src[bidx^2] - Y) * c_RGB2YUVCoeffs_i[3] + delta, yuv_shift);
-            const int Cb = CV_DESCALE((src[bidx] - Y) * c_RGB2YUVCoeffs_i[4] + delta, yuv_shift);
-
-            dst.x = saturate_cast<T>(Y);
-            dst.y = saturate_cast<T>(Cr);
-            dst.z = saturate_cast<T>(Cb);
-        }
-
-        template <int bidx, typename D> static __device__ __forceinline__ void RGB2YUVConvert(const float* src, D& dst)
-        {
-            dst.x = src[0] * c_RGB2YUVCoeffs_f[bidx^2] + src[1] * c_RGB2YUVCoeffs_f[1] + src[2] * c_RGB2YUVCoeffs_f[bidx];
-            dst.y = (src[bidx^2] - dst.x) * c_RGB2YUVCoeffs_f[3] + ColorChannel<float>::half();
-            dst.z = (src[bidx] - dst.x) * c_RGB2YUVCoeffs_f[4] + ColorChannel<float>::half();
-        }
-
-        template <typename T, int scn, int dcn, int bidx> struct RGB2YUV
-            : unary_function<typename TypeVec<T, scn>::vec_type, typename TypeVec<T, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<T, dcn>::vec_type operator ()(const typename TypeVec<T, scn>::vec_type& src) const
-            {
-                typename TypeVec<T, dcn>::vec_type dst;
-                RGB2YUVConvert<bidx>(&src.x, dst);
-                return dst;
-            }
-            __host__ __device__ __forceinline__ RGB2YUV() {}
-            __host__ __device__ __forceinline__ RGB2YUV(const RGB2YUV&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_RGB2YUV_TRAITS(name, scn, dcn, bidx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2YUV<T, scn, dcn, bidx> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-    namespace color_detail
-    {
-        __constant__ float c_YUV2RGBCoeffs_f[5] = { U2BF, U2GF, V2GF, V2RF };
-        __constant__ int   c_YUV2RGBCoeffs_i[5] = { U2BI, U2GI, V2GI, V2RI };
-
-        template <int bidx, typename T, typename D> static __device__ void YUV2RGBConvert(const T& src, D* dst)
-        {
-            const int b = src.x + CV_DESCALE((src.z - ColorChannel<D>::half()) * c_YUV2RGBCoeffs_i[3], yuv_shift);
-
-            const int g = src.x + CV_DESCALE((src.z - ColorChannel<D>::half()) * c_YUV2RGBCoeffs_i[2]
-                                             + (src.y - ColorChannel<D>::half()) * c_YUV2RGBCoeffs_i[1], yuv_shift);
-
-            const int r = src.x + CV_DESCALE((src.y - ColorChannel<D>::half()) * c_YUV2RGBCoeffs_i[0], yuv_shift);
-
-            dst[bidx] = saturate_cast<D>(b);
-            dst[1] = saturate_cast<D>(g);
-            dst[bidx^2] = saturate_cast<D>(r);
-        }
-
-        template <int bidx> static __device__ uint YUV2RGBConvert(uint src)
-        {
-            const int x = 0xff & (src);
-            const int y = 0xff & (src >> 8);
-            const int z = 0xff & (src >> 16);
-
-            const int b = x + CV_DESCALE((z - ColorChannel<uchar>::half()) * c_YUV2RGBCoeffs_i[3], yuv_shift);
-
-            const int g = x + CV_DESCALE((z - ColorChannel<uchar>::half()) * c_YUV2RGBCoeffs_i[2]
-                                         + (y - ColorChannel<uchar>::half()) * c_YUV2RGBCoeffs_i[1], yuv_shift);
-
-            const int r = x + CV_DESCALE((y - ColorChannel<uchar>::half()) * c_YUV2RGBCoeffs_i[0], yuv_shift);
-
-            uint dst = 0xffu << 24;
-
-            dst |= saturate_cast<uchar>(b) << (bidx * 8);
-            dst |= saturate_cast<uchar>(g) << 8;
-            dst |= saturate_cast<uchar>(r) << ((bidx ^ 2) * 8);
-
-            return dst;
-        }
-
-        template <int bidx, typename T> static __device__ __forceinline__ void YUV2RGBConvert(const T& src, float* dst)
-        {
-            dst[bidx] = src.x + (src.z - ColorChannel<float>::half()) * c_YUV2RGBCoeffs_f[3];
-
-            dst[1] = src.x + (src.z - ColorChannel<float>::half()) * c_YUV2RGBCoeffs_f[2]
-                     + (src.y - ColorChannel<float>::half()) * c_YUV2RGBCoeffs_f[1];
-
-            dst[bidx^2] = src.x + (src.y - ColorChannel<float>::half()) * c_YUV2RGBCoeffs_f[0];
-        }
-
-        template <typename T, int scn, int dcn, int bidx> struct YUV2RGB
-            : unary_function<typename TypeVec<T, scn>::vec_type, typename TypeVec<T, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<T, dcn>::vec_type operator ()(const typename TypeVec<T, scn>::vec_type& src) const
-            {
-                typename TypeVec<T, dcn>::vec_type dst;
-
-                YUV2RGBConvert<bidx>(src, &dst.x);
-                setAlpha(dst, ColorChannel<T>::max());
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ YUV2RGB() {}
-            __host__ __device__ __forceinline__ YUV2RGB(const YUV2RGB&) {}
-        };
-
-        template <int bidx> struct YUV2RGB<uchar, 4, 4, bidx> : unary_function<uint, uint>
-        {
-            __device__ __forceinline__ uint operator ()(uint src) const
-            {
-                return YUV2RGBConvert<bidx>(src);
-            }
-            __host__ __device__ __forceinline__ YUV2RGB() {}
-            __host__ __device__ __forceinline__ YUV2RGB(const YUV2RGB&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_YUV2RGB_TRAITS(name, scn, dcn, bidx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::YUV2RGB<T, scn, dcn, bidx> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-///////////////////////////////////// RGB <-> YCrCb //////////////////////////////////////
-
-    namespace color_detail
-    {
-        __constant__ float c_RGB2YCrCbCoeffs_f[5] = {R2YF, G2YF, B2YF, YCRF, YCBF};
-        __constant__ int   c_RGB2YCrCbCoeffs_i[5] = {R2Y, G2Y, B2Y, YCRI, YCBI};
-
-        template <int bidx, typename T, typename D> static __device__ void RGB2YCrCbConvert(const T* src, D& dst)
-        {
-            const int delta = ColorChannel<T>::half() * (1 << yuv_shift);
-
-            const int Y = CV_DESCALE(src[0] * c_RGB2YCrCbCoeffs_i[bidx^2] + src[1] * c_RGB2YCrCbCoeffs_i[1] + src[2] * c_RGB2YCrCbCoeffs_i[bidx], yuv_shift);
-            const int Cr = CV_DESCALE((src[bidx^2] - Y) * c_RGB2YCrCbCoeffs_i[3] + delta, yuv_shift);
-            const int Cb = CV_DESCALE((src[bidx] - Y) * c_RGB2YCrCbCoeffs_i[4] + delta, yuv_shift);
-
-            dst.x = saturate_cast<T>(Y);
-            dst.y = saturate_cast<T>(Cr);
-            dst.z = saturate_cast<T>(Cb);
-        }
-
-        template <int bidx> static __device__ uint RGB2YCrCbConvert(uint src)
-        {
-            const int delta = ColorChannel<uchar>::half() * (1 << yuv_shift);
-
-            const int Y = CV_DESCALE((0xffu & src) * c_RGB2YCrCbCoeffs_i[bidx^2] + (0xffu & (src >> 8)) * c_RGB2YCrCbCoeffs_i[1] + (0xffu & (src >> 16)) * c_RGB2YCrCbCoeffs_i[bidx], yuv_shift);
-            const int Cr = CV_DESCALE(((0xffu & (src >> ((bidx ^ 2) * 8))) - Y) * c_RGB2YCrCbCoeffs_i[3] + delta, yuv_shift);
-            const int Cb = CV_DESCALE(((0xffu & (src >> (bidx * 8))) - Y) * c_RGB2YCrCbCoeffs_i[4] + delta, yuv_shift);
-
-            uint dst = 0;
-
-            dst |= saturate_cast<uchar>(Y);
-            dst |= saturate_cast<uchar>(Cr) << 8;
-            dst |= saturate_cast<uchar>(Cb) << 16;
-
-            return dst;
-        }
-
-        template <int bidx, typename D> static __device__ __forceinline__ void RGB2YCrCbConvert(const float* src, D& dst)
-        {
-            dst.x = src[0] * c_RGB2YCrCbCoeffs_f[bidx^2] + src[1] * c_RGB2YCrCbCoeffs_f[1] + src[2] * c_RGB2YCrCbCoeffs_f[bidx];
-            dst.y = (src[bidx^2] - dst.x) * c_RGB2YCrCbCoeffs_f[3] + ColorChannel<float>::half();
-            dst.z = (src[bidx] - dst.x) * c_RGB2YCrCbCoeffs_f[4] + ColorChannel<float>::half();
-        }
-
-        template <typename T, int scn, int dcn, int bidx> struct RGB2YCrCb
-            : unary_function<typename TypeVec<T, scn>::vec_type, typename TypeVec<T, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<T, dcn>::vec_type operator ()(const typename TypeVec<T, scn>::vec_type& src) const
-            {
-                typename TypeVec<T, dcn>::vec_type dst;
-                RGB2YCrCbConvert<bidx>(&src.x, dst);
-                return dst;
-            }
-            __host__ __device__ __forceinline__ RGB2YCrCb() {}
-            __host__ __device__ __forceinline__ RGB2YCrCb(const RGB2YCrCb&) {}
-        };
-
-        template <int bidx> struct RGB2YCrCb<uchar, 4, 4, bidx> : unary_function<uint, uint>
-        {
-            __device__ __forceinline__ uint operator ()(uint src) const
-            {
-                return RGB2YCrCbConvert<bidx>(src);
-            }
-
-            __host__ __device__ __forceinline__ RGB2YCrCb() {}
-            __host__ __device__ __forceinline__ RGB2YCrCb(const RGB2YCrCb&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_RGB2YCrCb_TRAITS(name, scn, dcn, bidx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2YCrCb<T, scn, dcn, bidx> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-    namespace color_detail
-    {
-        __constant__ float c_YCrCb2RGBCoeffs_f[5] = {CR2RF, CR2GF, CB2GF, CB2BF};
-        __constant__ int   c_YCrCb2RGBCoeffs_i[5] = {CR2RI, CR2GI, CB2GI, CB2BI};
-
-        template <int bidx, typename T, typename D> static __device__ void YCrCb2RGBConvert(const T& src, D* dst)
-        {
-            const int b = src.x + CV_DESCALE((src.z - ColorChannel<D>::half()) * c_YCrCb2RGBCoeffs_i[3], yuv_shift);
-            const int g = src.x + CV_DESCALE((src.z - ColorChannel<D>::half()) * c_YCrCb2RGBCoeffs_i[2] + (src.y - ColorChannel<D>::half()) * c_YCrCb2RGBCoeffs_i[1], yuv_shift);
-            const int r = src.x + CV_DESCALE((src.y - ColorChannel<D>::half()) * c_YCrCb2RGBCoeffs_i[0], yuv_shift);
-
-            dst[bidx] = saturate_cast<D>(b);
-            dst[1] = saturate_cast<D>(g);
-            dst[bidx^2] = saturate_cast<D>(r);
-        }
-
-        template <int bidx> static __device__ uint YCrCb2RGBConvert(uint src)
-        {
-            const int x = 0xff & (src);
-            const int y = 0xff & (src >> 8);
-            const int z = 0xff & (src >> 16);
-
-            const int b = x + CV_DESCALE((z - ColorChannel<uchar>::half()) * c_YCrCb2RGBCoeffs_i[3], yuv_shift);
-            const int g = x + CV_DESCALE((z - ColorChannel<uchar>::half()) * c_YCrCb2RGBCoeffs_i[2] + (y - ColorChannel<uchar>::half()) * c_YCrCb2RGBCoeffs_i[1], yuv_shift);
-            const int r = x + CV_DESCALE((y - ColorChannel<uchar>::half()) * c_YCrCb2RGBCoeffs_i[0], yuv_shift);
-
-            uint dst = 0xffu << 24;
-
-            dst |= saturate_cast<uchar>(b) << (bidx * 8);
-            dst |= saturate_cast<uchar>(g) << 8;
-            dst |= saturate_cast<uchar>(r) << ((bidx ^ 2) * 8);
-
-            return dst;
-        }
-
-        template <int bidx, typename T> __device__ __forceinline__ void YCrCb2RGBConvert(const T& src, float* dst)
-        {
-            dst[bidx] = src.x + (src.z - ColorChannel<float>::half()) * c_YCrCb2RGBCoeffs_f[3];
-            dst[1] = src.x + (src.z - ColorChannel<float>::half()) * c_YCrCb2RGBCoeffs_f[2] + (src.y - ColorChannel<float>::half()) * c_YCrCb2RGBCoeffs_f[1];
-            dst[bidx^2] = src.x + (src.y - ColorChannel<float>::half()) * c_YCrCb2RGBCoeffs_f[0];
-        }
-
-        template <typename T, int scn, int dcn, int bidx> struct YCrCb2RGB
-            : unary_function<typename TypeVec<T, scn>::vec_type, typename TypeVec<T, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<T, dcn>::vec_type operator ()(const typename TypeVec<T, scn>::vec_type& src) const
-            {
-                typename TypeVec<T, dcn>::vec_type dst;
-
-                YCrCb2RGBConvert<bidx>(src, &dst.x);
-                setAlpha(dst, ColorChannel<T>::max());
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ YCrCb2RGB() {}
-            __host__ __device__ __forceinline__ YCrCb2RGB(const YCrCb2RGB&) {}
-        };
-
-        template <int bidx> struct YCrCb2RGB<uchar, 4, 4, bidx> : unary_function<uint, uint>
-        {
-            __device__ __forceinline__ uint operator ()(uint src) const
-            {
-                return YCrCb2RGBConvert<bidx>(src);
-            }
-            __host__ __device__ __forceinline__ YCrCb2RGB() {}
-            __host__ __device__ __forceinline__ YCrCb2RGB(const YCrCb2RGB&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_YCrCb2RGB_TRAITS(name, scn, dcn, bidx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::YCrCb2RGB<T, scn, dcn, bidx> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-////////////////////////////////////// RGB <-> XYZ ///////////////////////////////////////
-
-    namespace color_detail
-    {
-        __constant__ float c_RGB2XYZ_D65f[9] = { 0.412453f, 0.357580f, 0.180423f, 0.212671f, 0.715160f, 0.072169f, 0.019334f, 0.119193f, 0.950227f };
-        __constant__ int   c_RGB2XYZ_D65i[9] = { 1689, 1465, 739, 871, 2929, 296, 79, 488, 3892 };
-
-        template <int bidx, typename T, typename D> static __device__ __forceinline__ void RGB2XYZConvert(const T* src, D& dst)
-        {
-            dst.z = saturate_cast<T>(CV_DESCALE(src[bidx^2] * c_RGB2XYZ_D65i[6] + src[1] * c_RGB2XYZ_D65i[7] + src[bidx] * c_RGB2XYZ_D65i[8], xyz_shift));
-            dst.x = saturate_cast<T>(CV_DESCALE(src[bidx^2] * c_RGB2XYZ_D65i[0] + src[1] * c_RGB2XYZ_D65i[1] + src[bidx] * c_RGB2XYZ_D65i[2], xyz_shift));
-            dst.y = saturate_cast<T>(CV_DESCALE(src[bidx^2] * c_RGB2XYZ_D65i[3] + src[1] * c_RGB2XYZ_D65i[4] + src[bidx] * c_RGB2XYZ_D65i[5], xyz_shift));
-        }
-
-        template <int bidx> static __device__ __forceinline__ uint RGB2XYZConvert(uint src)
-        {
-            const uint b = 0xffu & (src >> (bidx * 8));
-            const uint g = 0xffu & (src >> 8);
-            const uint r = 0xffu & (src >> ((bidx ^ 2) * 8));
-
-            const uint x = saturate_cast<uchar>(CV_DESCALE(r * c_RGB2XYZ_D65i[0] + g * c_RGB2XYZ_D65i[1] + b * c_RGB2XYZ_D65i[2], xyz_shift));
-            const uint y = saturate_cast<uchar>(CV_DESCALE(r * c_RGB2XYZ_D65i[3] + g * c_RGB2XYZ_D65i[4] + b * c_RGB2XYZ_D65i[5], xyz_shift));
-            const uint z = saturate_cast<uchar>(CV_DESCALE(r * c_RGB2XYZ_D65i[6] + g * c_RGB2XYZ_D65i[7] + b * c_RGB2XYZ_D65i[8], xyz_shift));
-
-            uint dst = 0;
-
-            dst |= x;
-            dst |= y << 8;
-            dst |= z << 16;
-
-            return dst;
-        }
-
-        template <int bidx, typename D> static __device__ __forceinline__ void RGB2XYZConvert(const float* src, D& dst)
-        {
-            dst.x = src[bidx^2] * c_RGB2XYZ_D65f[0] + src[1] * c_RGB2XYZ_D65f[1] + src[bidx] * c_RGB2XYZ_D65f[2];
-            dst.y = src[bidx^2] * c_RGB2XYZ_D65f[3] + src[1] * c_RGB2XYZ_D65f[4] + src[bidx] * c_RGB2XYZ_D65f[5];
-            dst.z = src[bidx^2] * c_RGB2XYZ_D65f[6] + src[1] * c_RGB2XYZ_D65f[7] + src[bidx] * c_RGB2XYZ_D65f[8];
-        }
-
-        template <typename T, int scn, int dcn, int bidx> struct RGB2XYZ
-            : unary_function<typename TypeVec<T, scn>::vec_type, typename TypeVec<T, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<T, dcn>::vec_type operator()(const typename TypeVec<T, scn>::vec_type& src) const
-            {
-                typename TypeVec<T, dcn>::vec_type dst;
-
-                RGB2XYZConvert<bidx>(&src.x, dst);
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ RGB2XYZ() {}
-            __host__ __device__ __forceinline__ RGB2XYZ(const RGB2XYZ&) {}
-        };
-
-        template <int bidx> struct RGB2XYZ<uchar, 4, 4, bidx> : unary_function<uint, uint>
-        {
-            __device__ __forceinline__ uint operator()(uint src) const
-            {
-                return RGB2XYZConvert<bidx>(src);
-            }
-            __host__ __device__ __forceinline__ RGB2XYZ() {}
-            __host__ __device__ __forceinline__ RGB2XYZ(const RGB2XYZ&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_RGB2XYZ_TRAITS(name, scn, dcn, bidx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2XYZ<T, scn, dcn, bidx> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-    namespace color_detail
-    {
-        __constant__ float c_XYZ2sRGB_D65f[9] = { 3.240479f, -1.53715f, -0.498535f, -0.969256f, 1.875991f, 0.041556f, 0.055648f, -0.204043f, 1.057311f };
-        __constant__ int   c_XYZ2sRGB_D65i[9] = { 13273, -6296, -2042, -3970, 7684, 170, 228, -836, 4331 };
-
-        template <int bidx, typename T, typename D> static __device__ __forceinline__ void XYZ2RGBConvert(const T& src, D* dst)
-        {
-            dst[bidx^2] = saturate_cast<D>(CV_DESCALE(src.x * c_XYZ2sRGB_D65i[0] + src.y * c_XYZ2sRGB_D65i[1] + src.z * c_XYZ2sRGB_D65i[2], xyz_shift));
-            dst[1]      = saturate_cast<D>(CV_DESCALE(src.x * c_XYZ2sRGB_D65i[3] + src.y * c_XYZ2sRGB_D65i[4] + src.z * c_XYZ2sRGB_D65i[5], xyz_shift));
-            dst[bidx]   = saturate_cast<D>(CV_DESCALE(src.x * c_XYZ2sRGB_D65i[6] + src.y * c_XYZ2sRGB_D65i[7] + src.z * c_XYZ2sRGB_D65i[8], xyz_shift));
-        }
-
-        template <int bidx> static __device__ __forceinline__ uint XYZ2RGBConvert(uint src)
-        {
-            const int x = 0xff & src;
-            const int y = 0xff & (src >> 8);
-            const int z = 0xff & (src >> 16);
-
-            const uint r = saturate_cast<uchar>(CV_DESCALE(x * c_XYZ2sRGB_D65i[0] + y * c_XYZ2sRGB_D65i[1] + z * c_XYZ2sRGB_D65i[2], xyz_shift));
-            const uint g = saturate_cast<uchar>(CV_DESCALE(x * c_XYZ2sRGB_D65i[3] + y * c_XYZ2sRGB_D65i[4] + z * c_XYZ2sRGB_D65i[5], xyz_shift));
-            const uint b = saturate_cast<uchar>(CV_DESCALE(x * c_XYZ2sRGB_D65i[6] + y * c_XYZ2sRGB_D65i[7] + z * c_XYZ2sRGB_D65i[8], xyz_shift));
-
-            uint dst = 0xffu << 24;
-
-            dst |= b << (bidx * 8);
-            dst |= g << 8;
-            dst |= r << ((bidx ^ 2) * 8);
-
-            return dst;
-        }
-
-        template <int bidx, typename T> static __device__ __forceinline__ void XYZ2RGBConvert(const T& src, float* dst)
-        {
-            dst[bidx^2] = src.x * c_XYZ2sRGB_D65f[0] + src.y * c_XYZ2sRGB_D65f[1] + src.z * c_XYZ2sRGB_D65f[2];
-            dst[1]      = src.x * c_XYZ2sRGB_D65f[3] + src.y * c_XYZ2sRGB_D65f[4] + src.z * c_XYZ2sRGB_D65f[5];
-            dst[bidx]   = src.x * c_XYZ2sRGB_D65f[6] + src.y * c_XYZ2sRGB_D65f[7] + src.z * c_XYZ2sRGB_D65f[8];
-        }
-
-        template <typename T, int scn, int dcn, int bidx> struct XYZ2RGB
-            : unary_function<typename TypeVec<T, scn>::vec_type, typename TypeVec<T, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<T, dcn>::vec_type operator()(const typename TypeVec<T, scn>::vec_type& src) const
-            {
-                typename TypeVec<T, dcn>::vec_type dst;
-
-                XYZ2RGBConvert<bidx>(src, &dst.x);
-                setAlpha(dst, ColorChannel<T>::max());
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ XYZ2RGB() {}
-            __host__ __device__ __forceinline__ XYZ2RGB(const XYZ2RGB&) {}
-        };
-
-        template <int bidx> struct XYZ2RGB<uchar, 4, 4, bidx> : unary_function<uint, uint>
-        {
-            __device__ __forceinline__ uint operator()(uint src) const
-            {
-                return XYZ2RGBConvert<bidx>(src);
-            }
-            __host__ __device__ __forceinline__ XYZ2RGB() {}
-            __host__ __device__ __forceinline__ XYZ2RGB(const XYZ2RGB&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_XYZ2RGB_TRAITS(name, scn, dcn, bidx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::XYZ2RGB<T, scn, dcn, bidx> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-////////////////////////////////////// RGB <-> HSV ///////////////////////////////////////
-
-    namespace color_detail
-    {
-        __constant__ int c_HsvDivTable   [256] = {0, 1044480, 522240, 348160, 261120, 208896, 174080, 149211, 130560, 116053, 104448, 94953, 87040, 80345, 74606, 69632, 65280, 61440, 58027, 54973, 52224, 49737, 47476, 45412, 43520, 41779, 40172, 38684, 37303, 36017, 34816, 33693, 32640, 31651, 30720, 29842, 29013, 28229, 27486, 26782, 26112, 25475, 24869, 24290, 23738, 23211, 22706, 22223, 21760, 21316, 20890, 20480, 20086, 19707, 19342, 18991, 18651, 18324, 18008, 17703, 17408, 17123, 16846, 16579, 16320, 16069, 15825, 15589, 15360, 15137, 14921, 14711, 14507, 14308, 14115, 13926, 13743, 13565, 13391, 13221, 13056, 12895, 12738, 12584, 12434, 12288, 12145, 12006, 11869, 11736, 11605, 11478, 11353, 11231, 11111, 10995, 10880, 10768, 10658, 10550, 10445, 10341, 10240, 10141, 10043, 9947, 9854, 9761, 9671, 9582, 9495, 9410, 9326, 9243, 9162, 9082, 9004, 8927, 8852, 8777, 8704, 8632, 8561, 8492, 8423, 8356, 8290, 8224, 8160, 8097, 8034, 7973, 7913, 7853, 7795, 7737, 7680, 7624, 7569, 7514, 7461, 7408, 7355, 7304, 7253, 7203, 7154, 7105, 7057, 7010, 6963, 6917, 6872, 6827, 6782, 6739, 6695, 6653, 6611, 6569, 6528, 6487, 6447, 6408, 6369, 6330, 6292, 6254, 6217, 6180, 6144, 6108, 6073, 6037, 6003, 5968, 5935, 5901, 5868, 5835, 5803, 5771, 5739, 5708, 5677, 5646, 5615, 5585, 5556, 5526, 5497, 5468, 5440, 5412, 5384, 5356, 5329, 5302, 5275, 5249, 5222, 5196, 5171, 5145, 5120, 5095, 5070, 5046, 5022, 4998, 4974, 4950, 4927, 4904, 4881, 4858, 4836, 4813, 4791, 4769, 4748, 4726, 4705, 4684, 4663, 4642, 4622, 4601, 4581, 4561, 4541, 4522, 4502, 4483, 4464, 4445, 4426, 4407, 4389, 4370, 4352, 4334, 4316, 4298, 4281, 4263, 4246, 4229, 4212, 4195, 4178, 4161, 4145, 4128, 4112, 4096};
-        __constant__ int c_HsvDivTable180[256] = {0, 122880, 61440, 40960, 30720, 24576, 20480, 17554, 15360, 13653, 12288, 11171, 10240, 9452, 8777, 8192, 7680, 7228, 6827, 6467, 6144, 5851, 5585, 5343, 5120, 4915, 4726, 4551, 4389, 4237, 4096, 3964, 3840, 3724, 3614, 3511, 3413, 3321, 3234, 3151, 3072, 2997, 2926, 2858, 2793, 2731, 2671, 2614, 2560, 2508, 2458, 2409, 2363, 2318, 2276, 2234, 2194, 2156, 2119, 2083, 2048, 2014, 1982, 1950, 1920, 1890, 1862, 1834, 1807, 1781, 1755, 1731, 1707, 1683, 1661, 1638, 1617, 1596, 1575, 1555, 1536, 1517, 1499, 1480, 1463, 1446, 1429, 1412, 1396, 1381, 1365, 1350, 1336, 1321, 1307, 1293, 1280, 1267, 1254, 1241, 1229, 1217, 1205, 1193, 1182, 1170, 1159, 1148, 1138, 1127, 1117, 1107, 1097, 1087, 1078, 1069, 1059, 1050, 1041, 1033, 1024, 1016, 1007, 999, 991, 983, 975, 968, 960, 953, 945, 938, 931, 924, 917, 910, 904, 897, 890, 884, 878, 871, 865, 859, 853, 847, 842, 836, 830, 825, 819, 814, 808, 803, 798, 793, 788, 783, 778, 773, 768, 763, 759, 754, 749, 745, 740, 736, 731, 727, 723, 719, 714, 710, 706, 702, 698, 694, 690, 686, 683, 679, 675, 671, 668, 664, 661, 657, 654, 650, 647, 643, 640, 637, 633, 630, 627, 624, 621, 617, 614, 611, 608, 605, 602, 599, 597, 594, 591, 588, 585, 582, 580, 577, 574, 572, 569, 566, 564, 561, 559, 556, 554, 551, 549, 546, 544, 541, 539, 537, 534, 532, 530, 527, 525, 523, 521, 518, 516, 514, 512, 510, 508, 506, 504, 502, 500, 497, 495, 493, 492, 490, 488, 486, 484, 482};
-        __constant__ int c_HsvDivTable256[256] = {0, 174763, 87381, 58254, 43691, 34953, 29127, 24966, 21845, 19418, 17476, 15888, 14564, 13443, 12483, 11651, 10923, 10280, 9709, 9198, 8738, 8322, 7944, 7598, 7282, 6991, 6722, 6473, 6242, 6026, 5825, 5638, 5461, 5296, 5140, 4993, 4855, 4723, 4599, 4481, 4369, 4263, 4161, 4064, 3972, 3884, 3799, 3718, 3641, 3567, 3495, 3427, 3361, 3297, 3236, 3178, 3121, 3066, 3013, 2962, 2913, 2865, 2819, 2774, 2731, 2689, 2648, 2608, 2570, 2533, 2497, 2461, 2427, 2394, 2362, 2330, 2300, 2270, 2241, 2212, 2185, 2158, 2131, 2106, 2081, 2056, 2032, 2009, 1986, 1964, 1942, 1920, 1900, 1879, 1859, 1840, 1820, 1802, 1783, 1765, 1748, 1730, 1713, 1697, 1680, 1664, 1649, 1633, 1618, 1603, 1589, 1574, 1560, 1547, 1533, 1520, 1507, 1494, 1481, 1469, 1456, 1444, 1432, 1421, 1409, 1398, 1387, 1376, 1365, 1355, 1344, 1334, 1324, 1314, 1304, 1295, 1285, 1276, 1266, 1257, 1248, 1239, 1231, 1222, 1214, 1205, 1197, 1189, 1181, 1173, 1165, 1157, 1150, 1142, 1135, 1128, 1120, 1113, 1106, 1099, 1092, 1085, 1079, 1072, 1066, 1059, 1053, 1046, 1040, 1034, 1028, 1022, 1016, 1010, 1004, 999, 993, 987, 982, 976, 971, 966, 960, 955, 950, 945, 940, 935, 930, 925, 920, 915, 910, 906, 901, 896, 892, 887, 883, 878, 874, 869, 865, 861, 857, 853, 848, 844, 840, 836, 832, 828, 824, 820, 817, 813, 809, 805, 802, 798, 794, 791, 787, 784, 780, 777, 773, 770, 767, 763, 760, 757, 753, 750, 747, 744, 741, 737, 734, 731, 728, 725, 722, 719, 716, 713, 710, 708, 705, 702, 699, 696, 694, 691, 688, 685};
-
-        template <int bidx, int hr, typename D> static __device__ void RGB2HSVConvert(const uchar* src, D& dst)
-        {
-            const int hsv_shift = 12;
-            const int* hdiv_table = hr == 180 ? c_HsvDivTable180 : c_HsvDivTable256;
-
-            int b = src[bidx], g = src[1], r = src[bidx^2];
-            int h, s, v = b;
-            int vmin = b, diff;
-            int vr, vg;
-
-            v = ::max(v, g);
-            v = ::max(v, r);
-            vmin = ::min(vmin, g);
-            vmin = ::min(vmin, r);
-
-            diff = v - vmin;
-            vr = (v == r) * -1;
-            vg = (v == g) * -1;
-
-            s = (diff * c_HsvDivTable[v] + (1 << (hsv_shift-1))) >> hsv_shift;
-            h = (vr & (g - b)) + (~vr & ((vg & (b - r + 2 * diff)) + ((~vg) & (r - g + 4 * diff))));
-            h = (h * hdiv_table[diff] + (1 << (hsv_shift-1))) >> hsv_shift;
-            h += (h < 0) * hr;
-
-            dst.x = saturate_cast<uchar>(h);
-            dst.y = (uchar)s;
-            dst.z = (uchar)v;
-        }
-
-        template <int bidx, int hr> static __device__ uint RGB2HSVConvert(uint src)
-        {
-            const int hsv_shift = 12;
-            const int* hdiv_table = hr == 180 ? c_HsvDivTable180 : c_HsvDivTable256;
-
-            const int b = 0xff & (src >> (bidx * 8));
-            const int g = 0xff & (src >> 8);
-            const int r = 0xff & (src >> ((bidx ^ 2) * 8));
-
-            int h, s, v = b;
-            int vmin = b, diff;
-            int vr, vg;
-
-            v = ::max(v, g);
-            v = ::max(v, r);
-            vmin = ::min(vmin, g);
-            vmin = ::min(vmin, r);
-
-            diff = v - vmin;
-            vr = (v == r) * -1;
-            vg = (v == g) * -1;
-
-            s = (diff * c_HsvDivTable[v] + (1 << (hsv_shift-1))) >> hsv_shift;
-            h = (vr & (g - b)) + (~vr & ((vg & (b - r + 2 * diff)) + ((~vg) & (r - g + 4 * diff))));
-            h = (h * hdiv_table[diff] + (1 << (hsv_shift-1))) >> hsv_shift;
-            h += (h < 0) * hr;
-
-            uint dst = 0;
-
-            dst |= saturate_cast<uchar>(h);
-            dst |= (0xffu & s) << 8;
-            dst |= (0xffu & v) << 16;
-
-            return dst;
-        }
-
-        template <int bidx, int hr, typename D> static __device__ void RGB2HSVConvert(const float* src, D& dst)
-        {
-            const float hscale = hr * (1.f / 360.f);
-
-            float b = src[bidx], g = src[1], r = src[bidx^2];
-            float h, s, v;
-
-            float vmin, diff;
-
-            v = vmin = r;
-            v = fmax(v, g);
-            v = fmax(v, b);
-            vmin = fmin(vmin, g);
-            vmin = fmin(vmin, b);
-
-            diff = v - vmin;
-            s = diff / (float)(::fabs(v) + numeric_limits<float>::epsilon());
-            diff = (float)(60. / (diff + numeric_limits<float>::epsilon()));
-
-            h  = (v == r) * (g - b) * diff;
-            h += (v != r && v == g) * ((b - r) * diff + 120.f);
-            h += (v != r && v != g) * ((r - g) * diff + 240.f);
-            h += (h < 0) * 360.f;
-
-            dst.x = h * hscale;
-            dst.y = s;
-            dst.z = v;
-        }
-
-        template <typename T, int scn, int dcn, int bidx, int hr> struct RGB2HSV
-            : unary_function<typename TypeVec<T, scn>::vec_type, typename TypeVec<T, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<T, dcn>::vec_type operator()(const typename TypeVec<T, scn>::vec_type& src) const
-            {
-                typename TypeVec<T, dcn>::vec_type dst;
-
-                RGB2HSVConvert<bidx, hr>(&src.x, dst);
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ RGB2HSV() {}
-            __host__ __device__ __forceinline__ RGB2HSV(const RGB2HSV&) {}
-        };
-
-        template <int bidx, int hr> struct RGB2HSV<uchar, 4, 4, bidx, hr> : unary_function<uint, uint>
-        {
-            __device__ __forceinline__ uint operator()(uint src) const
-            {
-                return RGB2HSVConvert<bidx, hr>(src);
-            }
-            __host__ __device__ __forceinline__ RGB2HSV() {}
-            __host__ __device__ __forceinline__ RGB2HSV(const RGB2HSV&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_RGB2HSV_TRAITS(name, scn, dcn, bidx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2HSV<T, scn, dcn, bidx, 180> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    }; \
-    template <typename T> struct name ## _full_traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2HSV<T, scn, dcn, bidx, 256> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    }; \
-    template <> struct name ## _traits<float> \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2HSV<float, scn, dcn, bidx, 360> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    }; \
-    template <> struct name ## _full_traits<float> \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2HSV<float, scn, dcn, bidx, 360> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-    namespace color_detail
-    {
-        __constant__ int c_HsvSectorData[6][3] = { {1,3,0}, {1,0,2}, {3,0,1}, {0,2,1}, {0,1,3}, {2,1,0} };
-
-        template <int bidx, int hr, typename T> static __device__ void HSV2RGBConvert(const T& src, float* dst)
-        {
-            const float hscale = 6.f / hr;
-
-            float h = src.x, s = src.y, v = src.z;
-            float b = v, g = v, r = v;
-
-            if (s != 0)
-            {
-                h *= hscale;
-
-                if( h < 0 )
-                    do h += 6; while( h < 0 );
-                else if( h >= 6 )
-                    do h -= 6; while( h >= 6 );
-
-                int sector = __float2int_rd(h);
-                h -= sector;
-
-                if ( (unsigned)sector >= 6u )
-                {
-                    sector = 0;
-                    h = 0.f;
-                }
-
-                float tab[4];
-                tab[0] = v;
-                tab[1] = v * (1.f - s);
-                tab[2] = v * (1.f - s * h);
-                tab[3] = v * (1.f - s * (1.f - h));
-
-                b = tab[c_HsvSectorData[sector][0]];
-                g = tab[c_HsvSectorData[sector][1]];
-                r = tab[c_HsvSectorData[sector][2]];
-            }
-
-            dst[bidx] = b;
-            dst[1] = g;
-            dst[bidx^2] = r;
-        }
-
-        template <int bidx, int HR, typename T> static __device__ void HSV2RGBConvert(const T& src, uchar* dst)
-        {
-            float3 buf;
-
-            buf.x = src.x;
-            buf.y = src.y * (1.f / 255.f);
-            buf.z = src.z * (1.f / 255.f);
-
-            HSV2RGBConvert<bidx, HR>(buf, &buf.x);
-
-            dst[0] = saturate_cast<uchar>(buf.x * 255.f);
-            dst[1] = saturate_cast<uchar>(buf.y * 255.f);
-            dst[2] = saturate_cast<uchar>(buf.z * 255.f);
-        }
-
-        template <int bidx, int hr> static __device__ uint HSV2RGBConvert(uint src)
-        {
-            float3 buf;
-
-            buf.x = src & 0xff;
-            buf.y = ((src >> 8) & 0xff) * (1.f/255.f);
-            buf.z = ((src >> 16) & 0xff) * (1.f/255.f);
-
-            HSV2RGBConvert<bidx, hr>(buf, &buf.x);
-
-            uint dst = 0xffu << 24;
-
-            dst |= saturate_cast<uchar>(buf.x * 255.f);
-            dst |= saturate_cast<uchar>(buf.y * 255.f) << 8;
-            dst |= saturate_cast<uchar>(buf.z * 255.f) << 16;
-
-            return dst;
-        }
-
-        template <typename T, int scn, int dcn, int bidx, int hr> struct HSV2RGB
-            : unary_function<typename TypeVec<T, scn>::vec_type, typename TypeVec<T, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<T, dcn>::vec_type operator()(const typename TypeVec<T, scn>::vec_type& src) const
-            {
-                typename TypeVec<T, dcn>::vec_type dst;
-
-                HSV2RGBConvert<bidx, hr>(src, &dst.x);
-                setAlpha(dst, ColorChannel<T>::max());
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ HSV2RGB() {}
-            __host__ __device__ __forceinline__ HSV2RGB(const HSV2RGB&) {}
-        };
-
-        template <int bidx, int hr> struct HSV2RGB<uchar, 4, 4, bidx, hr> : unary_function<uint, uint>
-        {
-            __device__ __forceinline__ uint operator()(uint src) const
-            {
-                return HSV2RGBConvert<bidx, hr>(src);
-            }
-            __host__ __device__ __forceinline__ HSV2RGB() {}
-            __host__ __device__ __forceinline__ HSV2RGB(const HSV2RGB&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_HSV2RGB_TRAITS(name, scn, dcn, bidx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::HSV2RGB<T, scn, dcn, bidx, 180> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    }; \
-    template <typename T> struct name ## _full_traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::HSV2RGB<T, scn, dcn, bidx, 255> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    }; \
-    template <> struct name ## _traits<float> \
-    { \
-        typedef ::cv::cuda::device::color_detail::HSV2RGB<float, scn, dcn, bidx, 360> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    }; \
-    template <> struct name ## _full_traits<float> \
-    { \
-        typedef ::cv::cuda::device::color_detail::HSV2RGB<float, scn, dcn, bidx, 360> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-/////////////////////////////////////// RGB <-> HLS ////////////////////////////////////////
-
-    namespace color_detail
-    {
-        template <int bidx, int hr, typename D> static __device__ void RGB2HLSConvert(const float* src, D& dst)
-        {
-            const float hscale = hr * (1.f / 360.f);
-
-            float b = src[bidx], g = src[1], r = src[bidx^2];
-            float h = 0.f, s = 0.f, l;
-            float vmin, vmax, diff;
-
-            vmax = vmin = r;
-            vmax = fmax(vmax, g);
-            vmax = fmax(vmax, b);
-            vmin = fmin(vmin, g);
-            vmin = fmin(vmin, b);
-
-            diff = vmax - vmin;
-            l = (vmax + vmin) * 0.5f;
-
-            if (diff > numeric_limits<float>::epsilon())
-            {
-                s = (l < 0.5f) * diff / (vmax + vmin);
-                s += (l >= 0.5f) * diff / (2.0f - vmax - vmin);
-
-                diff = 60.f / diff;
-
-                h  = (vmax == r) * (g - b) * diff;
-                h += (vmax != r && vmax == g) * ((b - r) * diff + 120.f);
-                h += (vmax != r && vmax != g) * ((r - g) * diff + 240.f);
-                h += (h < 0.f) * 360.f;
-            }
-
-            dst.x = h * hscale;
-            dst.y = l;
-            dst.z = s;
-        }
-
-        template <int bidx, int hr, typename D> static __device__ void RGB2HLSConvert(const uchar* src, D& dst)
-        {
-            float3 buf;
-
-            buf.x = src[0] * (1.f / 255.f);
-            buf.y = src[1] * (1.f / 255.f);
-            buf.z = src[2] * (1.f / 255.f);
-
-            RGB2HLSConvert<bidx, hr>(&buf.x, buf);
-
-            dst.x = saturate_cast<uchar>(buf.x);
-            dst.y = saturate_cast<uchar>(buf.y*255.f);
-            dst.z = saturate_cast<uchar>(buf.z*255.f);
-        }
-
-        template <int bidx, int hr> static __device__ uint RGB2HLSConvert(uint src)
-        {
-            float3 buf;
-
-            buf.x = (0xff & src) * (1.f / 255.f);
-            buf.y = (0xff & (src >> 8)) * (1.f / 255.f);
-            buf.z = (0xff & (src >> 16)) * (1.f / 255.f);
-
-            RGB2HLSConvert<bidx, hr>(&buf.x, buf);
-
-            uint dst = 0xffu << 24;
-
-            dst |= saturate_cast<uchar>(buf.x);
-            dst |= saturate_cast<uchar>(buf.y * 255.f) << 8;
-            dst |= saturate_cast<uchar>(buf.z * 255.f) << 16;
-
-            return dst;
-        }
-
-        template <typename T, int scn, int dcn, int bidx, int hr> struct RGB2HLS
-            : unary_function<typename TypeVec<T, scn>::vec_type, typename TypeVec<T, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<T, dcn>::vec_type operator()(const typename TypeVec<T, scn>::vec_type& src) const
-            {
-                typename TypeVec<T, dcn>::vec_type dst;
-
-                RGB2HLSConvert<bidx, hr>(&src.x, dst);
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ RGB2HLS() {}
-            __host__ __device__ __forceinline__ RGB2HLS(const RGB2HLS&) {}
-        };
-
-        template <int bidx, int hr> struct RGB2HLS<uchar, 4, 4, bidx, hr> : unary_function<uint, uint>
-        {
-            __device__ __forceinline__ uint operator()(uint src) const
-            {
-                return RGB2HLSConvert<bidx, hr>(src);
-            }
-            __host__ __device__ __forceinline__ RGB2HLS() {}
-            __host__ __device__ __forceinline__ RGB2HLS(const RGB2HLS&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_RGB2HLS_TRAITS(name, scn, dcn, bidx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2HLS<T, scn, dcn, bidx, 180> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    }; \
-    template <typename T> struct name ## _full_traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2HLS<T, scn, dcn, bidx, 256> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    }; \
-    template <> struct name ## _traits<float> \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2HLS<float, scn, dcn, bidx, 360> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    }; \
-    template <> struct name ## _full_traits<float> \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2HLS<float, scn, dcn, bidx, 360> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-    namespace color_detail
-    {
-        __constant__ int c_HlsSectorData[6][3] = { {1,3,0}, {1,0,2}, {3,0,1}, {0,2,1}, {0,1,3}, {2,1,0} };
-
-        template <int bidx, int hr, typename T> static __device__ void HLS2RGBConvert(const T& src, float* dst)
-        {
-            const float hscale = 6.0f / hr;
-
-            float h = src.x, l = src.y, s = src.z;
-            float b = l, g = l, r = l;
-
-            if (s != 0)
-            {
-                float p2  = (l <= 0.5f) * l * (1 + s);
-                      p2 += (l > 0.5f) * (l + s - l * s);
-                float p1 = 2 * l - p2;
-
-                h *= hscale;
-
-                if( h < 0 )
-                    do h += 6; while( h < 0 );
-                else if( h >= 6 )
-                    do h -= 6; while( h >= 6 );
-
-                int sector;
-                sector = __float2int_rd(h);
-
-                h -= sector;
-
-                float tab[4];
-                tab[0] = p2;
-                tab[1] = p1;
-                tab[2] = p1 + (p2 - p1) * (1 - h);
-                tab[3] = p1 + (p2 - p1) * h;
-
-                b = tab[c_HlsSectorData[sector][0]];
-                g = tab[c_HlsSectorData[sector][1]];
-                r = tab[c_HlsSectorData[sector][2]];
-            }
-
-            dst[bidx] = b;
-            dst[1] = g;
-            dst[bidx^2] = r;
-        }
-
-        template <int bidx, int hr, typename T> static __device__ void HLS2RGBConvert(const T& src, uchar* dst)
-        {
-            float3 buf;
-
-            buf.x = src.x;
-            buf.y = src.y * (1.f / 255.f);
-            buf.z = src.z * (1.f / 255.f);
-
-            HLS2RGBConvert<bidx, hr>(buf, &buf.x);
-
-            dst[0] = saturate_cast<uchar>(buf.x * 255.f);
-            dst[1] = saturate_cast<uchar>(buf.y * 255.f);
-            dst[2] = saturate_cast<uchar>(buf.z * 255.f);
-        }
-
-        template <int bidx, int hr> static __device__ uint HLS2RGBConvert(uint src)
-        {
-            float3 buf;
-
-            buf.x = 0xff & src;
-            buf.y = (0xff & (src >> 8)) * (1.f / 255.f);
-            buf.z = (0xff & (src >> 16)) * (1.f / 255.f);
-
-            HLS2RGBConvert<bidx, hr>(buf, &buf.x);
-
-            uint dst = 0xffu << 24;
-
-            dst |= saturate_cast<uchar>(buf.x * 255.f);
-            dst |= saturate_cast<uchar>(buf.y * 255.f) << 8;
-            dst |= saturate_cast<uchar>(buf.z * 255.f) << 16;
-
-            return dst;
-        }
-
-        template <typename T, int scn, int dcn, int bidx, int hr> struct HLS2RGB
-            : unary_function<typename TypeVec<T, scn>::vec_type, typename TypeVec<T, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<T, dcn>::vec_type operator()(const typename TypeVec<T, scn>::vec_type& src) const
-            {
-                typename TypeVec<T, dcn>::vec_type dst;
-
-                HLS2RGBConvert<bidx, hr>(src, &dst.x);
-                setAlpha(dst, ColorChannel<T>::max());
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ HLS2RGB() {}
-            __host__ __device__ __forceinline__ HLS2RGB(const HLS2RGB&) {}
-        };
-
-        template <int bidx, int hr> struct HLS2RGB<uchar, 4, 4, bidx, hr> : unary_function<uint, uint>
-        {
-            __device__ __forceinline__ uint operator()(uint src) const
-            {
-                return HLS2RGBConvert<bidx, hr>(src);
-            }
-            __host__ __device__ __forceinline__ HLS2RGB() {}
-            __host__ __device__ __forceinline__ HLS2RGB(const HLS2RGB&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_HLS2RGB_TRAITS(name, scn, dcn, bidx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::HLS2RGB<T, scn, dcn, bidx, 180> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    }; \
-    template <typename T> struct name ## _full_traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::HLS2RGB<T, scn, dcn, bidx, 255> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    }; \
-    template <> struct name ## _traits<float> \
-    { \
-        typedef ::cv::cuda::device::color_detail::HLS2RGB<float, scn, dcn, bidx, 360> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    }; \
-    template <> struct name ## _full_traits<float> \
-    { \
-        typedef ::cv::cuda::device::color_detail::HLS2RGB<float, scn, dcn, bidx, 360> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-///////////////////////////////////// RGB <-> Lab /////////////////////////////////////
-
-    namespace color_detail
-    {
-        enum
-        {
-            LAB_CBRT_TAB_SIZE = 1024,
-            GAMMA_TAB_SIZE = 1024,
-            lab_shift = xyz_shift,
-            gamma_shift = 3,
-            lab_shift2 = (lab_shift + gamma_shift),
-            LAB_CBRT_TAB_SIZE_B = (256 * 3 / 2 * (1 << gamma_shift))
-        };
-
-        __constant__ ushort c_sRGBGammaTab_b[] = {0,1,1,2,2,3,4,4,5,6,6,7,8,8,9,10,11,11,12,13,14,15,16,17,19,20,21,22,24,25,26,28,29,31,33,34,36,38,40,41,43,45,47,49,51,54,56,58,60,63,65,68,70,73,75,78,81,83,86,89,92,95,98,101,105,108,111,115,118,121,125,129,132,136,140,144,147,151,155,160,164,168,172,176,181,185,190,194,199,204,209,213,218,223,228,233,239,244,249,255,260,265,271,277,282,288,294,300,306,312,318,324,331,337,343,350,356,363,370,376,383,390,397,404,411,418,426,433,440,448,455,463,471,478,486,494,502,510,518,527,535,543,552,560,569,578,586,595,604,613,622,631,641,650,659,669,678,688,698,707,717,727,737,747,757,768,778,788,799,809,820,831,842,852,863,875,886,897,908,920,931,943,954,966,978,990,1002,1014,1026,1038,1050,1063,1075,1088,1101,1113,1126,1139,1152,1165,1178,1192,1205,1218,1232,1245,1259,1273,1287,1301,1315,1329,1343,1357,1372,1386,1401,1415,1430,1445,1460,1475,1490,1505,1521,1536,1551,1567,1583,1598,1614,1630,1646,1662,1678,1695,1711,1728,1744,1761,1778,1794,1811,1828,1846,1863,1880,1897,1915,1933,1950,1968,1986,2004,2022,2040};
-
-        __device__ __forceinline__ int LabCbrt_b(int i)
-        {
-            float x = i * (1.f / (255.f * (1 << gamma_shift)));
-            return (1 << lab_shift2) * (x < 0.008856f ? x * 7.787f + 0.13793103448275862f : ::cbrtf(x));
-        }
-
-        template <bool srgb, int blueIdx, typename T, typename D>
-        __device__ __forceinline__ void RGB2LabConvert_b(const T& src, D& dst)
-        {
-            const int Lscale = (116 * 255 + 50) / 100;
-            const int Lshift = -((16 * 255 * (1 << lab_shift2) + 50) / 100);
-
-            int B = blueIdx == 0 ? src.x : src.z;
-            int G = src.y;
-            int R = blueIdx == 0 ? src.z : src.x;
-
-            if (srgb)
-            {
-                B = c_sRGBGammaTab_b[B];
-                G = c_sRGBGammaTab_b[G];
-                R = c_sRGBGammaTab_b[R];
-            }
-            else
-            {
-                B <<= 3;
-                G <<= 3;
-                R <<= 3;
-            }
-
-            int fX = LabCbrt_b(CV_DESCALE(B * 778 + G * 1541 + R * 1777, lab_shift));
-            int fY = LabCbrt_b(CV_DESCALE(B * 296 + G * 2929 + R * 871, lab_shift));
-            int fZ = LabCbrt_b(CV_DESCALE(B * 3575 + G * 448 + R * 73, lab_shift));
-
-            int L = CV_DESCALE(Lscale * fY + Lshift, lab_shift2);
-            int a = CV_DESCALE(500 * (fX - fY) + 128 * (1 << lab_shift2), lab_shift2);
-            int b = CV_DESCALE(200 * (fY - fZ) + 128 * (1 << lab_shift2), lab_shift2);
-
-            dst.x = saturate_cast<uchar>(L);
-            dst.y = saturate_cast<uchar>(a);
-            dst.z = saturate_cast<uchar>(b);
-        }
-
-        __device__ __forceinline__ float splineInterpolate(float x, const float* tab, int n)
-        {
-            int ix = ::min(::max(int(x), 0), n-1);
-            x -= ix;
-            tab += ix * 4;
-            return ((tab[3] * x + tab[2]) * x + tab[1]) * x + tab[0];
-        }
-
-        __constant__ float c_sRGBGammaTab[] = {0,7.55853e-05,0.,-7.51331e-13,7.55853e-05,7.55853e-05,-2.25399e-12,3.75665e-12,0.000151171,7.55853e-05,9.01597e-12,-6.99932e-12,0.000226756,7.55853e-05,-1.1982e-11,2.41277e-12,0.000302341,7.55853e-05,-4.74369e-12,1.19001e-11,0.000377927,7.55853e-05,3.09568e-11,-2.09095e-11,0.000453512,7.55853e-05,-3.17718e-11,1.35303e-11,0.000529097,7.55853e-05,8.81905e-12,-4.10782e-12,0.000604683,7.55853e-05,-3.50439e-12,2.90097e-12,0.000680268,7.55853e-05,5.19852e-12,-7.49607e-12,0.000755853,7.55853e-05,-1.72897e-11,2.70833e-11,0.000831439,7.55854e-05,6.39602e-11,-4.26295e-11,0.000907024,7.55854e-05,-6.39282e-11,2.70193e-11,0.000982609,7.55853e-05,1.71298e-11,-7.24017e-12,0.00105819,7.55853e-05,-4.59077e-12,1.94137e-12,0.00113378,7.55853e-05,1.23333e-12,-5.25291e-13,0.00120937,7.55853e-05,-3.42545e-13,1.59799e-13,0.00128495,7.55853e-05,1.36852e-13,-1.13904e-13,0.00136054,7.55853e-05,-2.04861e-13,2.95818e-13,0.00143612,7.55853e-05,6.82594e-13,-1.06937e-12,0.00151171,7.55853e-05,-2.52551e-12,3.98166e-12,0.00158729,7.55853e-05,9.41946e-12,-1.48573e-11,0.00166288,7.55853e-05,-3.51523e-11,5.54474e-11,0.00173846,7.55854e-05,1.3119e-10,-9.0517e-11,0.00181405,7.55854e-05,-1.40361e-10,7.37899e-11,0.00188963,7.55853e-05,8.10085e-11,-8.82272e-11,0.00196522,7.55852e-05,-1.83673e-10,1.62704e-10,0.0020408,7.55853e-05,3.04438e-10,-2.13341e-10,0.00211639,7.55853e-05,-3.35586e-10,2.25e-10,0.00219197,7.55853e-05,3.39414e-10,-2.20997e-10,0.00226756,7.55853e-05,-3.23576e-10,1.93326e-10,0.00234315,7.55853e-05,2.564e-10,-8.66446e-11,0.00241873,7.55855e-05,-3.53328e-12,-7.9578e-11,0.00249432,7.55853e-05,-2.42267e-10,1.72126e-10,0.0025699,7.55853e-05,2.74111e-10,-1.43265e-10,0.00264549,7.55854e-05,-1.55683e-10,-6.47292e-11,0.00272107,7.55849e-05,-3.4987e-10,8.67842e-10,0.00279666,7.55868e-05,2.25366e-09,-3.8723e-09,0.00287224,7.55797e-05,-9.36325e-09,1.5087e-08,0.00294783,7.56063e-05,3.58978e-08,-5.69415e-08,0.00302341,7.55072e-05,-1.34927e-07,2.13144e-07,0.003099,7.58768e-05,5.04507e-07,1.38713e-07,0.00317552,7.7302e-05,9.20646e-07,-1.55186e-07,0.00325359,7.86777e-05,4.55087e-07,4.26813e-08,0.00333276,7.97159e-05,5.83131e-07,-1.06495e-08,0.00341305,8.08502e-05,5.51182e-07,3.87467e-09,0.00349446,8.19642e-05,5.62806e-07,-1.92586e-10,0.00357698,8.30892e-05,5.62228e-07,1.0866e-09,0.00366063,8.4217e-05,5.65488e-07,5.02818e-10,0.00374542,8.53494e-05,5.66997e-07,8.60211e-10,0.00383133,8.6486e-05,5.69577e-07,7.13044e-10,0.00391839,8.76273e-05,5.71716e-07,4.78527e-10,0.00400659,8.87722e-05,5.73152e-07,1.09818e-09,0.00409594,8.99218e-05,5.76447e-07,2.50964e-10,0.00418644,9.10754e-05,5.772e-07,1.15762e-09,0.00427809,9.22333e-05,5.80672e-07,2.40865e-10,0.0043709,9.33954e-05,5.81395e-07,1.13854e-09,0.00446488,9.45616e-05,5.84811e-07,3.27267e-10,0.00456003,9.57322e-05,5.85792e-07,8.1197e-10,0.00465635,9.69062e-05,5.88228e-07,6.15823e-10,0.00475384,9.80845e-05,5.90076e-07,9.15747e-10,0.00485252,9.92674e-05,5.92823e-07,3.778e-10,0.00495238,0.000100454,5.93956e-07,8.32623e-10,0.00505343,0.000101645,5.96454e-07,4.82695e-10,0.00515567,0.000102839,5.97902e-07,9.61904e-10,0.00525911,0.000104038,6.00788e-07,3.26281e-10,0.00536375,0.00010524,6.01767e-07,9.926e-10,0.00546959,0.000106447,6.04745e-07,3.59933e-10,0.00557664,0.000107657,6.05824e-07,8.2728e-10,0.0056849,0.000108871,6.08306e-07,5.21898e-10,0.00579438,0.00011009,6.09872e-07,8.10492e-10,0.00590508,0.000111312,6.12303e-07,4.27046e-10,0.00601701,0.000112538,6.13585e-07,7.40878e-10,0.00613016,0.000113767,6.15807e-07,8.00469e-10,0.00624454,0.000115001,6.18209e-07,2.48178e-10,0.00636016,0.000116238,6.18953e-07,1.00073e-09,0.00647702,0.000117479,6.21955e-07,4.05654e-10,0.00659512,0.000118724,6.23172e-07,6.36192e-10,0.00671447,0.000119973,6.25081e-07,7.74927e-10,0.00683507,0.000121225,6.27406e-07,4.54975e-10,0.00695692,0.000122481,6.28771e-07,6.64841e-10,0.00708003,0.000123741,6.30765e-07,6.10972e-10,0.00720441,0.000125004,6.32598e-07,6.16543e-10,0.00733004,0.000126271,6.34448e-07,6.48204e-10,0.00745695,0.000127542,6.36392e-07,5.15835e-10,0.00758513,0.000128816,6.3794e-07,5.48103e-10,0.00771458,0.000130094,6.39584e-07,1.01706e-09,0.00784532,0.000131376,6.42635e-07,4.0283e-11,0.00797734,0.000132661,6.42756e-07,6.84471e-10,0.00811064,0.000133949,6.4481e-07,9.47144e-10,0.00824524,0.000135241,6.47651e-07,1.83472e-10,0.00838112,0.000136537,6.48201e-07,1.11296e-09,0.00851831,0.000137837,6.5154e-07,2.13163e-11,0.0086568,0.00013914,6.51604e-07,6.64462e-10,0.00879659,0.000140445,6.53598e-07,1.04613e-09,0.00893769,0.000141756,6.56736e-07,-1.92377e-10,0.0090801,0.000143069,6.56159e-07,1.58601e-09,0.00922383,0.000144386,6.60917e-07,-5.63754e-10,0.00936888,0.000145706,6.59226e-07,1.60033e-09,0.00951524,0.000147029,6.64027e-07,-2.49543e-10,0.00966294,0.000148356,6.63278e-07,1.26043e-09,0.00981196,0.000149687,6.67059e-07,-1.35572e-10,0.00996231,0.00015102,6.66653e-07,1.14458e-09,0.010114,0.000152357,6.70086e-07,2.13864e-10,0.010267,0.000153698,6.70728e-07,7.93856e-10,0.0104214,0.000155042,6.73109e-07,3.36077e-10,0.0105771,0.000156389,6.74118e-07,6.55765e-10,0.0107342,0.000157739,6.76085e-07,7.66211e-10,0.0108926,0.000159094,6.78384e-07,4.66116e-12,0.0110524,0.000160451,6.78398e-07,1.07775e-09,0.0112135,0.000161811,6.81631e-07,3.41023e-10,0.011376,0.000163175,6.82654e-07,3.5205e-10,0.0115398,0.000164541,6.8371e-07,1.04473e-09,0.0117051,0.000165912,6.86844e-07,1.25757e-10,0.0118717,0.000167286,6.87222e-07,3.14818e-10,0.0120396,0.000168661,6.88166e-07,1.40886e-09,0.012209,0.000170042,6.92393e-07,-3.62244e-10,0.0123797,0.000171425,6.91306e-07,9.71397e-10,0.0125518,0.000172811,6.9422e-07,2.02003e-10,0.0127253,0.0001742,6.94826e-07,1.01448e-09,0.0129002,0.000175593,6.97869e-07,3.96653e-10,0.0130765,0.00017699,6.99059e-07,1.92927e-10,0.0132542,0.000178388,6.99638e-07,6.94305e-10,0.0134333,0.00017979,7.01721e-07,7.55108e-10,0.0136138,0.000181195,7.03986e-07,1.05918e-11,0.0137957,0.000182603,7.04018e-07,1.06513e-09,0.013979,0.000184015,7.07214e-07,3.85512e-10,0.0141637,0.00018543,7.0837e-07,1.86769e-10,0.0143499,0.000186848,7.0893e-07,7.30116e-10,0.0145374,0.000188268,7.11121e-07,6.17983e-10,0.0147264,0.000189692,7.12975e-07,5.23282e-10,0.0149168,0.000191119,7.14545e-07,8.28398e-11,0.0151087,0.000192549,7.14793e-07,1.0081e-09,0.0153019,0.000193981,7.17817e-07,5.41244e-10,0.0154966,0.000195418,7.19441e-07,-3.7907e-10,0.0156928,0.000196856,7.18304e-07,1.90641e-09,0.0158903,0.000198298,7.24023e-07,-7.27387e-10,0.0160893,0.000199744,7.21841e-07,1.00317e-09,0.0162898,0.000201191,7.24851e-07,4.39949e-10,0.0164917,0.000202642,7.2617e-07,9.6234e-10,0.0166951,0.000204097,7.29057e-07,-5.64019e-10,0.0168999,0.000205554,7.27365e-07,1.29374e-09,0.0171062,0.000207012,7.31247e-07,9.77025e-10,0.017314,0.000208478,7.34178e-07,-1.47651e-09,0.0175232,0.000209942,7.29748e-07,3.06636e-09,0.0177338,0.00021141,7.38947e-07,-1.47573e-09,0.017946,0.000212884,7.3452e-07,9.7386e-10,0.0181596,0.000214356,7.37442e-07,1.30562e-09,0.0183747,0.000215835,7.41358e-07,-6.08376e-10,0.0185913,0.000217315,7.39533e-07,1.12785e-09,0.0188093,0.000218798,7.42917e-07,-1.77711e-10,0.0190289,0.000220283,7.42384e-07,1.44562e-09,0.0192499,0.000221772,7.46721e-07,-1.68825e-11,0.0194724,0.000223266,7.4667e-07,4.84533e-10,0.0196964,0.000224761,7.48124e-07,-5.85298e-11,0.0199219,0.000226257,7.47948e-07,1.61217e-09,0.0201489,0.000227757,7.52785e-07,-8.02136e-10,0.0203775,0.00022926,7.50378e-07,1.59637e-09,0.0206075,0.000230766,7.55167e-07,4.47168e-12,0.020839,0.000232276,7.55181e-07,2.48387e-10,0.021072,0.000233787,7.55926e-07,8.6474e-10,0.0213066,0.000235302,7.5852e-07,1.78299e-11,0.0215426,0.000236819,7.58573e-07,9.26567e-10,0.0217802,0.000238339,7.61353e-07,1.34529e-12,0.0220193,0.000239862,7.61357e-07,9.30659e-10,0.0222599,0.000241387,7.64149e-07,1.34529e-12,0.0225021,0.000242915,7.64153e-07,9.26567e-10,0.0227458,0.000244447,7.66933e-07,1.76215e-11,0.022991,0.00024598,7.66986e-07,8.65536e-10,0.0232377,0.000247517,7.69582e-07,2.45677e-10,0.023486,0.000249057,7.70319e-07,1.44193e-11,0.0237358,0.000250598,7.70363e-07,1.55918e-09,0.0239872,0.000252143,7.7504e-07,-6.63173e-10,0.0242401,0.000253691,7.73051e-07,1.09357e-09,0.0244946,0.000255241,7.76331e-07,1.41919e-11,0.0247506,0.000256793,7.76374e-07,7.12248e-10,0.0250082,0.000258348,7.78511e-07,8.62049e-10,0.0252673,0.000259908,7.81097e-07,-4.35061e-10,0.025528,0.000261469,7.79792e-07,8.7825e-10,0.0257902,0.000263031,7.82426e-07,6.47181e-10,0.0260541,0.000264598,7.84368e-07,2.58448e-10,0.0263194,0.000266167,7.85143e-07,1.81558e-10,0.0265864,0.000267738,7.85688e-07,8.78041e-10,0.0268549,0.000269312,7.88322e-07,3.15102e-11,0.027125,0.000270889,7.88417e-07,8.58525e-10,0.0273967,0.000272468,7.90992e-07,2.59812e-10,0.02767,0.000274051,7.91772e-07,-3.5224e-11,0.0279448,0.000275634,7.91666e-07,1.74377e-09,0.0282212,0.000277223,7.96897e-07,-1.35196e-09,0.0284992,0.000278813,7.92841e-07,1.80141e-09,0.0287788,0.000280404,7.98246e-07,-2.65629e-10,0.0290601,0.000281999,7.97449e-07,1.12374e-09,0.0293428,0.000283598,8.0082e-07,-5.04106e-10,0.0296272,0.000285198,7.99308e-07,8.92764e-10,0.0299132,0.000286799,8.01986e-07,6.58379e-10,0.0302008,0.000288405,8.03961e-07,1.98971e-10,0.0304901,0.000290014,8.04558e-07,4.08382e-10,0.0307809,0.000291624,8.05783e-07,3.01839e-11,0.0310733,0.000293236,8.05874e-07,1.33343e-09,0.0313673,0.000294851,8.09874e-07,2.2419e-10,0.031663,0.000296472,8.10547e-07,-3.67606e-10,0.0319603,0.000298092,8.09444e-07,1.24624e-09,0.0322592,0.000299714,8.13182e-07,-8.92025e-10,0.0325597,0.000301338,8.10506e-07,2.32183e-09,0.0328619,0.000302966,8.17472e-07,-9.44719e-10,0.0331657,0.000304598,8.14638e-07,1.45703e-09,0.0334711,0.000306232,8.19009e-07,-1.15805e-09,0.0337781,0.000307866,8.15535e-07,3.17507e-09,0.0340868,0.000309507,8.2506e-07,-4.09161e-09,0.0343971,0.000311145,8.12785e-07,5.74079e-09,0.0347091,0.000312788,8.30007e-07,-3.97034e-09,0.0350227,0.000314436,8.18096e-07,2.68985e-09,0.035338,0.00031608,8.26166e-07,6.61676e-10,0.0356549,0.000317734,8.28151e-07,-1.61123e-09,0.0359734,0.000319386,8.23317e-07,2.05786e-09,0.0362936,0.000321038,8.29491e-07,8.30388e-10,0.0366155,0.0003227,8.31982e-07,-1.65424e-09,0.036939,0.000324359,8.27019e-07,2.06129e-09,0.0372642,0.000326019,8.33203e-07,8.59719e-10,0.0375911,0.000327688,8.35782e-07,-1.77488e-09,0.0379196,0.000329354,8.30458e-07,2.51464e-09,0.0382498,0.000331023,8.38002e-07,-8.33135e-10,0.0385817,0.000332696,8.35502e-07,8.17825e-10,0.0389152,0.00033437,8.37956e-07,1.28718e-09,0.0392504,0.00033605,8.41817e-07,-2.2413e-09,0.0395873,0.000337727,8.35093e-07,3.95265e-09,0.0399258,0.000339409,8.46951e-07,-2.39332e-09,0.0402661,0.000341095,8.39771e-07,1.89533e-09,0.040608,0.000342781,8.45457e-07,-1.46271e-09,0.0409517,0.000344467,8.41069e-07,3.95554e-09,0.041297,0.000346161,8.52936e-07,-3.18369e-09,0.041644,0.000347857,8.43385e-07,1.32873e-09,0.0419927,0.000349548,8.47371e-07,1.59402e-09,0.0423431,0.000351248,8.52153e-07,-2.54336e-10,0.0426952,0.000352951,8.5139e-07,-5.76676e-10,0.043049,0.000354652,8.4966e-07,2.56114e-09,0.0434045,0.000356359,8.57343e-07,-2.21744e-09,0.0437617,0.000358067,8.50691e-07,2.58344e-09,0.0441206,0.000359776,8.58441e-07,-6.65826e-10,0.0444813,0.000361491,8.56444e-07,7.99218e-11,0.0448436,0.000363204,8.56684e-07,3.46063e-10,0.0452077,0.000364919,8.57722e-07,2.26116e-09,0.0455734,0.000366641,8.64505e-07,-1.94005e-09,0.045941,0.000368364,8.58685e-07,1.77384e-09,0.0463102,0.000370087,8.64007e-07,-1.43005e-09,0.0466811,0.000371811,8.59717e-07,3.94634e-09,0.0470538,0.000373542,8.71556e-07,-3.17946e-09,0.0474282,0.000375276,8.62017e-07,1.32104e-09,0.0478043,0.000377003,8.6598e-07,1.62045e-09,0.0481822,0.00037874,8.70842e-07,-3.52297e-10,0.0485618,0.000380481,8.69785e-07,-2.11211e-10,0.0489432,0.00038222,8.69151e-07,1.19716e-09,0.0493263,0.000383962,8.72743e-07,-8.52026e-10,0.0497111,0.000385705,8.70187e-07,2.21092e-09,0.0500977,0.000387452,8.76819e-07,-5.41339e-10,0.050486,0.000389204,8.75195e-07,-4.5361e-11,0.0508761,0.000390954,8.75059e-07,7.22669e-10,0.0512679,0.000392706,8.77227e-07,8.79936e-10,0.0516615,0.000394463,8.79867e-07,-5.17048e-10,0.0520568,0.000396222,8.78316e-07,1.18833e-09,0.0524539,0.000397982,8.81881e-07,-5.11022e-10,0.0528528,0.000399744,8.80348e-07,8.55683e-10,0.0532534,0.000401507,8.82915e-07,8.13562e-10,0.0536558,0.000403276,8.85356e-07,-3.84603e-10,0.05406,0.000405045,8.84202e-07,7.24962e-10,0.0544659,0.000406816,8.86377e-07,1.20986e-09,0.0548736,0.000408592,8.90006e-07,-1.83896e-09,0.0552831,0.000410367,8.84489e-07,2.42071e-09,0.0556944,0.000412143,8.91751e-07,-3.93413e-10,0.0561074,0.000413925,8.90571e-07,-8.46967e-10,0.0565222,0.000415704,8.8803e-07,3.78122e-09,0.0569388,0.000417491,8.99374e-07,-3.1021e-09,0.0573572,0.000419281,8.90068e-07,1.17658e-09,0.0577774,0.000421064,8.93597e-07,2.12117e-09,0.0581993,0.000422858,8.99961e-07,-2.21068e-09,0.0586231,0.000424651,8.93329e-07,2.9961e-09,0.0590486,0.000426447,9.02317e-07,-2.32311e-09,0.059476,0.000428244,8.95348e-07,2.57122e-09,0.0599051,0.000430043,9.03062e-07,-5.11098e-10,0.0603361,0.000431847,9.01528e-07,-5.27166e-10,0.0607688,0.000433649,8.99947e-07,2.61984e-09,0.0612034,0.000435457,9.07806e-07,-2.50141e-09,0.0616397,0.000437265,9.00302e-07,3.66045e-09,0.0620779,0.000439076,9.11283e-07,-4.68977e-09,0.0625179,0.000440885,8.97214e-07,7.64783e-09,0.0629597,0.000442702,9.20158e-07,-7.27499e-09,0.0634033,0.000444521,8.98333e-07,6.55113e-09,0.0638487,0.000446337,9.17986e-07,-4.02844e-09,0.0642959,0.000448161,9.05901e-07,2.11196e-09,0.064745,0.000449979,9.12236e-07,3.03125e-09,0.0651959,0.000451813,9.2133e-07,-6.78648e-09,0.0656486,0.000453635,9.00971e-07,9.21375e-09,0.0661032,0.000455464,9.28612e-07,-7.71684e-09,0.0665596,0.000457299,9.05462e-07,6.7522e-09,0.0670178,0.00045913,9.25718e-07,-4.3907e-09,0.0674778,0.000460968,9.12546e-07,3.36e-09,0.0679397,0.000462803,9.22626e-07,-1.59876e-09,0.0684034,0.000464644,9.1783e-07,3.0351e-09,0.068869,0.000466488,9.26935e-07,-3.09101e-09,0.0693364,0.000468333,9.17662e-07,1.8785e-09,0.0698057,0.000470174,9.23298e-07,3.02733e-09,0.0702768,0.00047203,9.3238e-07,-6.53722e-09,0.0707497,0.000473875,9.12768e-07,8.22054e-09,0.0712245,0.000475725,9.37429e-07,-3.99325e-09,0.0717012,0.000477588,9.2545e-07,3.01839e-10,0.0721797,0.00047944,9.26355e-07,2.78597e-09,0.0726601,0.000481301,9.34713e-07,-3.99507e-09,0.0731423,0.000483158,9.22728e-07,5.7435e-09,0.0736264,0.000485021,9.39958e-07,-4.07776e-09,0.0741123,0.000486888,9.27725e-07,3.11695e-09,0.0746002,0.000488753,9.37076e-07,-9.39394e-10,0.0750898,0.000490625,9.34258e-07,6.4055e-10,0.0755814,0.000492495,9.3618e-07,-1.62265e-09,0.0760748,0.000494363,9.31312e-07,5.84995e-09,0.0765701,0.000496243,9.48861e-07,-6.87601e-09,0.0770673,0.00049812,9.28233e-07,6.75296e-09,0.0775664,0.000499997,9.48492e-07,-5.23467e-09,0.0780673,0.000501878,9.32788e-07,6.73523e-09,0.0785701,0.000503764,9.52994e-07,-6.80514e-09,0.0790748,0.000505649,9.32578e-07,5.5842e-09,0.0795814,0.000507531,9.49331e-07,-6.30583e-10,0.0800899,0.000509428,9.47439e-07,-3.0618e-09,0.0806003,0.000511314,9.38254e-07,5.4273e-09,0.0811125,0.000513206,9.54536e-07,-3.74627e-09,0.0816267,0.000515104,9.43297e-07,2.10713e-09,0.0821427,0.000516997,9.49618e-07,2.76839e-09,0.0826607,0.000518905,9.57924e-07,-5.73006e-09,0.0831805,0.000520803,9.40733e-07,5.25072e-09,0.0837023,0.0005227,9.56486e-07,-3.71718e-10,0.084226,0.000524612,9.5537e-07,-3.76404e-09,0.0847515,0.000526512,9.44078e-07,7.97735e-09,0.085279,0.000528424,9.6801e-07,-5.79367e-09,0.0858084,0.000530343,9.50629e-07,2.96268e-10,0.0863397,0.000532245,9.51518e-07,4.6086e-09,0.0868729,0.000534162,9.65344e-07,-3.82947e-09,0.087408,0.000536081,9.53856e-07,3.25861e-09,0.087945,0.000537998,9.63631e-07,-1.7543e-09,0.088484,0.00053992,9.58368e-07,3.75849e-09,0.0890249,0.000541848,9.69644e-07,-5.82891e-09,0.0895677,0.00054377,9.52157e-07,4.65593e-09,0.0901124,0.000545688,9.66125e-07,2.10643e-09,0.0906591,0.000547627,9.72444e-07,-5.63099e-09,0.0912077,0.000549555,9.55551e-07,5.51627e-09,0.0917582,0.000551483,9.721e-07,-1.53292e-09,0.0923106,0.000553422,9.67501e-07,6.15311e-10,0.092865,0.000555359,9.69347e-07,-9.28291e-10,0.0934213,0.000557295,9.66562e-07,3.09774e-09,0.0939796,0.000559237,9.75856e-07,-4.01186e-09,0.0945398,0.000561177,9.6382e-07,5.49892e-09,0.095102,0.000563121,9.80317e-07,-3.08258e-09,0.0956661,0.000565073,9.71069e-07,-6.19176e-10,0.0962321,0.000567013,9.69212e-07,5.55932e-09,0.0968001,0.000568968,9.8589e-07,-6.71704e-09,0.09737,0.00057092,9.65738e-07,6.40762e-09,0.0979419,0.00057287,9.84961e-07,-4.0122e-09,0.0985158,0.000574828,9.72925e-07,2.19059e-09,0.0990916,0.000576781,9.79496e-07,2.70048e-09,0.0996693,0.000578748,9.87598e-07,-5.54193e-09,0.100249,0.000580706,9.70972e-07,4.56597e-09,0.100831,0.000582662,9.8467e-07,2.17923e-09,0.101414,0.000584638,9.91208e-07,-5.83232e-09,0.102,0.000586603,9.73711e-07,6.24884e-09,0.102588,0.000588569,9.92457e-07,-4.26178e-09,0.103177,0.000590541,9.79672e-07,3.34781e-09,0.103769,0.00059251,9.89715e-07,-1.67904e-09,0.104362,0.000594485,9.84678e-07,3.36839e-09,0.104958,0.000596464,9.94783e-07,-4.34397e-09,0.105555,0.000598441,9.81751e-07,6.55696e-09,0.106155,0.000600424,1.00142e-06,-6.98272e-09,0.106756,0.000602406,9.80474e-07,6.4728e-09,0.107359,0.000604386,9.99893e-07,-4.00742e-09,0.107965,0.000606374,9.8787e-07,2.10654e-09,0.108572,0.000608356,9.9419e-07,3.0318e-09,0.109181,0.000610353,1.00329e-06,-6.7832e-09,0.109793,0.00061234,9.82936e-07,9.1998e-09,0.110406,0.000614333,1.01054e-06,-7.6642e-09,0.111021,0.000616331,9.87543e-07,6.55579e-09,0.111639,0.000618326,1.00721e-06,-3.65791e-09,0.112258,0.000620329,9.96236e-07,6.25467e-10,0.112879,0.000622324,9.98113e-07,1.15593e-09,0.113503,0.000624323,1.00158e-06,2.20158e-09,0.114128,0.000626333,1.00819e-06,-2.51191e-09,0.114755,0.000628342,1.00065e-06,3.95517e-10,0.115385,0.000630345,1.00184e-06,9.29807e-10,0.116016,0.000632351,1.00463e-06,3.33599e-09,0.116649,0.00063437,1.01463e-06,-6.82329e-09,0.117285,0.000636379,9.94163e-07,9.05595e-09,0.117922,0.000638395,1.02133e-06,-7.04862e-09,0.118562,0.000640416,1.00019e-06,4.23737e-09,0.119203,0.000642429,1.0129e-06,-2.45033e-09,0.119847,0.000644448,1.00555e-06,5.56395e-09,0.120492,0.000646475,1.02224e-06,-4.9043e-09,0.121139,0.000648505,1.00753e-06,-8.47952e-10,0.121789,0.000650518,1.00498e-06,8.29622e-09,0.122441,0.000652553,1.02987e-06,-9.98538e-09,0.123094,0.000654582,9.99914e-07,9.2936e-09,0.12375,0.00065661,1.02779e-06,-4.83707e-09,0.124407,0.000658651,1.01328e-06,2.60411e-09,0.125067,0.000660685,1.0211e-06,-5.57945e-09,0.125729,0.000662711,1.00436e-06,1.22631e-08,0.126392,0.000664756,1.04115e-06,-1.36704e-08,0.127058,0.000666798,1.00014e-06,1.26161e-08,0.127726,0.000668836,1.03798e-06,-6.99155e-09,0.128396,0.000670891,1.01701e-06,4.48836e-10,0.129068,0.000672926,1.01836e-06,5.19606e-09,0.129742,0.000674978,1.03394e-06,-6.3319e-09,0.130418,0.000677027,1.01495e-06,5.2305e-09,0.131096,0.000679073,1.03064e-06,3.11123e-10,0.131776,0.000681135,1.03157e-06,-6.47511e-09,0.132458,0.000683179,1.01215e-06,1.06882e-08,0.133142,0.000685235,1.04421e-06,-6.47519e-09,0.133829,0.000687304,1.02479e-06,3.11237e-10,0.134517,0.000689355,1.02572e-06,5.23035e-09,0.135207,0.000691422,1.04141e-06,-6.3316e-09,0.1359,0.000693486,1.02242e-06,5.19484e-09,0.136594,0.000695546,1.038e-06,4.53497e-10,0.137291,0.000697623,1.03936e-06,-7.00891e-09,0.137989,0.000699681,1.01834e-06,1.2681e-08,0.13869,0.000701756,1.05638e-06,-1.39128e-08,0.139393,0.000703827,1.01464e-06,1.31679e-08,0.140098,0.000705896,1.05414e-06,-8.95659e-09,0.140805,0.000707977,1.02727e-06,7.75742e-09,0.141514,0.000710055,1.05055e-06,-7.17182e-09,0.142225,0.000712135,1.02903e-06,6.02862e-09,0.142938,0.000714211,1.04712e-06,-2.04163e-09,0.143653,0.000716299,1.04099e-06,2.13792e-09,0.144371,0.000718387,1.04741e-06,-6.51009e-09,0.14509,0.000720462,1.02787e-06,9.00123e-09,0.145812,0.000722545,1.05488e-06,3.07523e-10,0.146535,0.000724656,1.0558e-06,-1.02312e-08,0.147261,0.000726737,1.02511e-06,1.0815e-08,0.147989,0.000728819,1.05755e-06,-3.22681e-09,0.148719,0.000730925,1.04787e-06,2.09244e-09,0.14945,0.000733027,1.05415e-06,-5.143e-09,0.150185,0.00073512,1.03872e-06,3.57844e-09,0.150921,0.000737208,1.04946e-06,5.73027e-09,0.151659,0.000739324,1.06665e-06,-1.15983e-08,0.152399,0.000741423,1.03185e-06,1.08605e-08,0.153142,0.000743519,1.06443e-06,-2.04106e-09,0.153886,0.000745642,1.05831e-06,-2.69642e-09,0.154633,0.00074775,1.05022e-06,-2.07425e-09,0.155382,0.000749844,1.044e-06,1.09934e-08,0.156133,0.000751965,1.07698e-06,-1.20972e-08,0.156886,0.000754083,1.04069e-06,7.59288e-09,0.157641,0.000756187,1.06347e-06,-3.37305e-09,0.158398,0.000758304,1.05335e-06,5.89921e-09,0.159158,0.000760428,1.07104e-06,-5.32248e-09,0.159919,0.000762554,1.05508e-06,4.8927e-10,0.160683,0.000764666,1.05654e-06,3.36547e-09,0.161448,0.000766789,1.06664e-06,9.50081e-10,0.162216,0.000768925,1.06949e-06,-7.16568e-09,0.162986,0.000771043,1.04799e-06,1.28114e-08,0.163758,0.000773177,1.08643e-06,-1.42774e-08,0.164533,0.000775307,1.0436e-06,1.44956e-08,0.165309,0.000777438,1.08708e-06,-1.39025e-08,0.166087,0.00077957,1.04538e-06,1.13118e-08,0.166868,0.000781695,1.07931e-06,-1.54224e-09,0.167651,0.000783849,1.07468e-06,-5.14312e-09,0.168436,0.000785983,1.05925e-06,7.21381e-09,0.169223,0.000788123,1.0809e-06,-8.81096e-09,0.170012,0.000790259,1.05446e-06,1.31289e-08,0.170803,0.000792407,1.09385e-06,-1.39022e-08,0.171597,0.000794553,1.05214e-06,1.26775e-08,0.172392,0.000796695,1.09018e-06,-7.00557e-09,0.17319,0.000798855,1.06916e-06,4.43796e-10,0.17399,0.000800994,1.07049e-06,5.23031e-09,0.174792,0.000803151,1.08618e-06,-6.46397e-09,0.175596,0.000805304,1.06679e-06,5.72444e-09,0.176403,0.000807455,1.08396e-06,-1.53254e-09,0.177211,0.000809618,1.07937e-06,4.05673e-10,0.178022,0.000811778,1.08058e-06,-9.01916e-11,0.178835,0.000813939,1.08031e-06,-4.49821e-11,0.17965,0.000816099,1.08018e-06,2.70234e-10,0.180467,0.00081826,1.08099e-06,-1.03603e-09,0.181286,0.000820419,1.07788e-06,3.87392e-09,0.182108,0.000822587,1.0895e-06,4.41522e-10,0.182932,0.000824767,1.09083e-06,-5.63997e-09,0.183758,0.000826932,1.07391e-06,7.21707e-09,0.184586,0.000829101,1.09556e-06,-8.32718e-09,0.185416,0.000831267,1.07058e-06,1.11907e-08,0.186248,0.000833442,1.10415e-06,-6.63336e-09,0.187083,0.00083563,1.08425e-06,4.41484e-10,0.187919,0.0008378,1.08557e-06,4.86754e-09,0.188758,0.000839986,1.10017e-06,-5.01041e-09,0.189599,0.000842171,1.08514e-06,2.72811e-10,0.190443,0.000844342,1.08596e-06,3.91916e-09,0.191288,0.000846526,1.09772e-06,-1.04819e-09,0.192136,0.000848718,1.09457e-06,2.73531e-10,0.192985,0.000850908,1.0954e-06,-4.58916e-11,0.193837,0.000853099,1.09526e-06,-9.01158e-11,0.194692,0.000855289,1.09499e-06,4.06506e-10,0.195548,0.00085748,1.09621e-06,-1.53595e-09,0.196407,0.000859668,1.0916e-06,5.73717e-09,0.197267,0.000861869,1.10881e-06,-6.51164e-09,0.19813,0.000864067,1.08928e-06,5.40831e-09,0.198995,0.000866261,1.1055e-06,-2.20401e-10,0.199863,0.000868472,1.10484e-06,-4.52652e-09,0.200732,0.000870668,1.09126e-06,3.42508e-09,0.201604,0.000872861,1.10153e-06,5.72762e-09,0.202478,0.000875081,1.11872e-06,-1.14344e-08,0.203354,0.000877284,1.08441e-06,1.02076e-08,0.204233,0.000879484,1.11504e-06,4.06355e-10,0.205113,0.000881715,1.11626e-06,-1.18329e-08,0.205996,0.000883912,1.08076e-06,1.71227e-08,0.206881,0.000886125,1.13213e-06,-1.19546e-08,0.207768,0.000888353,1.09626e-06,8.93465e-10,0.208658,0.000890548,1.09894e-06,8.38062e-09,0.209549,0.000892771,1.12408e-06,-4.61353e-09,0.210443,0.000895006,1.11024e-06,-4.82756e-09,0.211339,0.000897212,1.09576e-06,9.02245e-09,0.212238,0.00089943,1.12283e-06,-1.45997e-09,0.213138,0.000901672,1.11845e-06,-3.18255e-09,0.214041,0.000903899,1.1089e-06,-7.11073e-10,0.214946,0.000906115,1.10677e-06,6.02692e-09,0.215853,0.000908346,1.12485e-06,-8.49548e-09,0.216763,0.00091057,1.09936e-06,1.30537e-08,0.217675,0.000912808,1.13852e-06,-1.3917e-08,0.218588,0.000915044,1.09677e-06,1.28121e-08,0.219505,0.000917276,1.13521e-06,-7.5288e-09,0.220423,0.000919523,1.11262e-06,2.40205e-09,0.221344,0.000921756,1.11983e-06,-2.07941e-09,0.222267,0.000923989,1.11359e-06,5.91551e-09,0.223192,0.000926234,1.13134e-06,-6.68149e-09,0.224119,0.000928477,1.11129e-06,5.90929e-09,0.225049,0.000930717,1.12902e-06,-2.05436e-09,0.22598,0.000932969,1.12286e-06,2.30807e-09,0.226915,0.000935222,1.12978e-06,-7.17796e-09,0.227851,0.00093746,1.10825e-06,1.15028e-08,0.228789,0.000939711,1.14276e-06,-9.03083e-09,0.22973,0.000941969,1.11566e-06,9.71932e-09,0.230673,0.00094423,1.14482e-06,-1.49452e-08,0.231619,0.000946474,1.09998e-06,2.02591e-08,0.232566,0.000948735,1.16076e-06,-2.13879e-08,0.233516,0.000950993,1.0966e-06,2.05888e-08,0.234468,0.000953247,1.15837e-06,-1.62642e-08,0.235423,0.000955515,1.10957e-06,1.46658e-08,0.236379,0.000957779,1.15357e-06,-1.25966e-08,0.237338,0.000960048,1.11578e-06,5.91793e-09,0.238299,0.000962297,1.13353e-06,3.82602e-09,0.239263,0.000964576,1.14501e-06,-6.3208e-09,0.240229,0.000966847,1.12605e-06,6.55613e-09,0.241197,0.000969119,1.14572e-06,-5.00268e-09,0.242167,0.000971395,1.13071e-06,-1.44659e-09,0.243139,0.000973652,1.12637e-06,1.07891e-08,0.244114,0.000975937,1.15874e-06,-1.19073e-08,0.245091,0.000978219,1.12302e-06,7.03782e-09,0.246071,0.000980486,1.14413e-06,-1.34276e-09,0.247052,0.00098277,1.1401e-06,-1.66669e-09,0.248036,0.000985046,1.1351e-06,8.00935e-09,0.249022,0.00098734,1.15913e-06,-1.54694e-08,0.250011,0.000989612,1.11272e-06,2.4066e-08,0.251002,0.000991909,1.18492e-06,-2.11901e-08,0.251995,0.000994215,1.12135e-06,1.08973e-09,0.25299,0.000996461,1.12462e-06,1.68311e-08,0.253988,0.000998761,1.17511e-06,-8.8094e-09,0.254987,0.00100109,1.14868e-06,-1.13958e-08,0.25599,0.00100335,1.1145e-06,2.45902e-08,0.256994,0.00100565,1.18827e-06,-2.73603e-08,0.258001,0.00100795,1.10618e-06,2.52464e-08,0.25901,0.00101023,1.18192e-06,-1.40207e-08,0.260021,0.00101256,1.13986e-06,1.03387e-09,0.261035,0.00101484,1.14296e-06,9.8853e-09,0.262051,0.00101715,1.17262e-06,-1.07726e-08,0.263069,0.00101947,1.1403e-06,3.40272e-09,0.26409,0.00102176,1.15051e-06,-2.83827e-09,0.265113,0.00102405,1.142e-06,7.95039e-09,0.266138,0.00102636,1.16585e-06,8.39047e-10,0.267166,0.00102869,1.16836e-06,-1.13066e-08,0.268196,0.00103099,1.13444e-06,1.4585e-08,0.269228,0.00103331,1.1782e-06,-1.72314e-08,0.270262,0.00103561,1.1265e-06,2.45382e-08,0.271299,0.00103794,1.20012e-06,-2.13166e-08,0.272338,0.00104028,1.13617e-06,1.12364e-09,0.273379,0.00104255,1.13954e-06,1.68221e-08,0.274423,0.00104488,1.19001e-06,-8.80736e-09,0.275469,0.00104723,1.16358e-06,-1.13948e-08,0.276518,0.00104953,1.1294e-06,2.45839e-08,0.277568,0.00105186,1.20315e-06,-2.73361e-08,0.278621,0.00105418,1.12114e-06,2.51559e-08,0.279677,0.0010565,1.19661e-06,-1.36832e-08,0.280734,0.00105885,1.15556e-06,-2.25706e-10,0.281794,0.00106116,1.15488e-06,1.45862e-08,0.282857,0.00106352,1.19864e-06,-2.83167e-08,0.283921,0.00106583,1.11369e-06,3.90759e-08,0.284988,0.00106817,1.23092e-06,-3.85801e-08,0.286058,0.00107052,1.11518e-06,2.58375e-08,0.287129,0.00107283,1.19269e-06,-5.16498e-09,0.288203,0.0010752,1.1772e-06,-5.17768e-09,0.28928,0.00107754,1.16167e-06,-3.92671e-09,0.290358,0.00107985,1.14988e-06,2.08846e-08,0.29144,0.00108221,1.21254e-06,-2.00072e-08,0.292523,0.00108458,1.15252e-06,-4.60659e-10,0.293609,0.00108688,1.15114e-06,2.18499e-08,0.294697,0.00108925,1.21669e-06,-2.73343e-08,0.295787,0.0010916,1.13468e-06,2.78826e-08,0.29688,0.00109395,1.21833e-06,-2.45915e-08,0.297975,0.00109632,1.14456e-06,1.08787e-08,0.299073,0.00109864,1.17719e-06,1.08788e-08,0.300172,0.00110102,1.20983e-06,-2.45915e-08,0.301275,0.00110337,1.13605e-06,2.78828e-08,0.302379,0.00110573,1.2197e-06,-2.73348e-08,0.303486,0.00110808,1.1377e-06,2.18518e-08,0.304595,0.00111042,1.20325e-06,-4.67556e-10,0.305707,0.00111283,1.20185e-06,-1.99816e-08,0.306821,0.00111517,1.14191e-06,2.07891e-08,0.307937,0.00111752,1.20427e-06,-3.57026e-09,0.309056,0.00111992,1.19356e-06,-6.50797e-09,0.310177,0.00112228,1.17404e-06,-2.00165e-10,0.3113,0.00112463,1.17344e-06,7.30874e-09,0.312426,0.001127,1.19536e-06,7.67424e-10,0.313554,0.00112939,1.19767e-06,-1.03784e-08,0.314685,0.00113176,1.16653e-06,1.09437e-08,0.315818,0.00113412,1.19936e-06,-3.59406e-09,0.316953,0.00113651,1.18858e-06,3.43251e-09,0.318091,0.0011389,1.19888e-06,-1.0136e-08,0.319231,0.00114127,1.16847e-06,7.30915e-09,0.320374,0.00114363,1.1904e-06,1.07018e-08,0.321518,0.00114604,1.2225e-06,-2.03137e-08,0.322666,0.00114842,1.16156e-06,1.09484e-08,0.323815,0.00115078,1.19441e-06,6.32224e-09,0.324967,0.00115319,1.21337e-06,-6.43509e-09,0.326122,0.00115559,1.19407e-06,-1.03842e-08,0.327278,0.00115795,1.16291e-06,1.81697e-08,0.328438,0.00116033,1.21742e-06,-2.6901e-09,0.329599,0.00116276,1.20935e-06,-7.40939e-09,0.330763,0.00116515,1.18713e-06,2.52533e-09,0.331929,0.00116754,1.1947e-06,-2.69191e-09,0.333098,0.00116992,1.18663e-06,8.24218e-09,0.334269,0.00117232,1.21135e-06,-4.74377e-10,0.335443,0.00117474,1.20993e-06,-6.34471e-09,0.336619,0.00117714,1.1909e-06,-3.94922e-09,0.337797,0.00117951,1.17905e-06,2.21417e-08,0.338978,0.00118193,1.24547e-06,-2.50128e-08,0.340161,0.00118435,1.17043e-06,1.8305e-08,0.341346,0.00118674,1.22535e-06,-1.84048e-08,0.342534,0.00118914,1.17013e-06,2.55121e-08,0.343725,0.00119156,1.24667e-06,-2.40389e-08,0.344917,0.00119398,1.17455e-06,1.10389e-08,0.346113,0.00119636,1.20767e-06,9.68574e-09,0.34731,0.0011988,1.23673e-06,-1.99797e-08,0.34851,0.00120122,1.17679e-06,1.06284e-08,0.349713,0.0012036,1.20867e-06,7.26868e-09,0.350917,0.00120604,1.23048e-06,-9.90072e-09,0.352125,0.00120847,1.20078e-06,2.53177e-09,0.353334,0.00121088,1.20837e-06,-2.26199e-10,0.354546,0.0012133,1.20769e-06,-1.62705e-09,0.355761,0.00121571,1.20281e-06,6.73435e-09,0.356978,0.00121813,1.22302e-06,4.49207e-09,0.358197,0.00122059,1.23649e-06,-2.47027e-08,0.359419,0.00122299,1.16238e-06,3.47142e-08,0.360643,0.00122542,1.26653e-06,-2.47472e-08,0.36187,0.00122788,1.19229e-06,4.66965e-09,0.363099,0.00123028,1.20629e-06,6.06872e-09,0.36433,0.00123271,1.2245e-06,8.57729e-10,0.365564,0.00123516,1.22707e-06,-9.49952e-09,0.366801,0.00123759,1.19858e-06,7.33792e-09,0.36804,0.00124001,1.22059e-06,9.95025e-09,0.369281,0.00124248,1.25044e-06,-1.73366e-08,0.370525,0.00124493,1.19843e-06,-2.08464e-10,0.371771,0.00124732,1.1978e-06,1.81704e-08,0.373019,0.00124977,1.25232e-06,-1.28683e-08,0.37427,0.00125224,1.21371e-06,3.50042e-09,0.375524,0.00125468,1.22421e-06,-1.1335e-09,0.37678,0.00125712,1.22081e-06,1.03345e-09,0.378038,0.00125957,1.22391e-06,-3.00023e-09,0.379299,0.00126201,1.21491e-06,1.09676e-08,0.380562,0.00126447,1.24781e-06,-1.10676e-08,0.381828,0.00126693,1.21461e-06,3.50042e-09,0.383096,0.00126937,1.22511e-06,-2.93403e-09,0.384366,0.00127181,1.21631e-06,8.23574e-09,0.385639,0.00127427,1.24102e-06,-2.06607e-10,0.386915,0.00127675,1.2404e-06,-7.40935e-09,0.388193,0.00127921,1.21817e-06,4.1761e-11,0.389473,0.00128165,1.21829e-06,7.24223e-09,0.390756,0.0012841,1.24002e-06,7.91564e-10,0.392042,0.00128659,1.2424e-06,-1.04086e-08,0.393329,0.00128904,1.21117e-06,1.10405e-08,0.39462,0.0012915,1.24429e-06,-3.951e-09,0.395912,0.00129397,1.23244e-06,4.7634e-09,0.397208,0.00129645,1.24673e-06,-1.51025e-08,0.398505,0.0012989,1.20142e-06,2.58443e-08,0.399805,0.00130138,1.27895e-06,-2.86702e-08,0.401108,0.00130385,1.19294e-06,2.92318e-08,0.402413,0.00130632,1.28064e-06,-2.86524e-08,0.403721,0.0013088,1.19468e-06,2.57731e-08,0.405031,0.00131127,1.272e-06,-1.48355e-08,0.406343,0.00131377,1.2275e-06,3.76652e-09,0.407658,0.00131623,1.23879e-06,-2.30784e-10,0.408976,0.00131871,1.2381e-06,-2.84331e-09,0.410296,0.00132118,1.22957e-06,1.16041e-08,0.411618,0.00132367,1.26438e-06,-1.37708e-08,0.412943,0.00132616,1.22307e-06,1.36768e-08,0.41427,0.00132865,1.2641e-06,-1.1134e-08,0.4156,0.00133114,1.2307e-06,1.05714e-09,0.416933,0.00133361,1.23387e-06,6.90538e-09,0.418267,0.00133609,1.25459e-06,1.12372e-09,0.419605,0.00133861,1.25796e-06,-1.14002e-08,0.420945,0.00134109,1.22376e-06,1.46747e-08,0.422287,0.00134358,1.26778e-06,-1.7496e-08,0.423632,0.00134606,1.21529e-06,2.5507e-08,0.424979,0.00134857,1.29182e-06,-2.49272e-08,0.426329,0.00135108,1.21703e-06,1.45972e-08,0.427681,0.00135356,1.26083e-06,-3.65935e-09,0.429036,0.00135607,1.24985e-06,4.00178e-11,0.430393,0.00135857,1.24997e-06,3.49917e-09,0.431753,0.00136108,1.26047e-06,-1.40366e-08,0.433116,0.00136356,1.21836e-06,2.28448e-08,0.43448,0.00136606,1.28689e-06,-1.77378e-08,0.435848,0.00136858,1.23368e-06,1.83043e-08,0.437218,0.0013711,1.28859e-06,-2.56769e-08,0.43859,0.0013736,1.21156e-06,2.47987e-08,0.439965,0.0013761,1.28595e-06,-1.39133e-08,0.441342,0.00137863,1.24421e-06,1.05202e-09,0.442722,0.00138112,1.24737e-06,9.70507e-09,0.444104,0.00138365,1.27649e-06,-1.00698e-08,0.445489,0.00138617,1.24628e-06,7.72123e-10,0.446877,0.00138867,1.24859e-06,6.98132e-09,0.448267,0.00139118,1.26954e-06,1.10477e-09,0.449659,0.00139373,1.27285e-06,-1.14003e-08,0.451054,0.00139624,1.23865e-06,1.4694e-08,0.452452,0.00139876,1.28273e-06,-1.75734e-08,0.453852,0.00140127,1.23001e-06,2.5797e-08,0.455254,0.00140381,1.3074e-06,-2.60097e-08,0.456659,0.00140635,1.22937e-06,1.86371e-08,0.458067,0.00140886,1.28529e-06,-1.8736e-08,0.459477,0.00141137,1.22908e-06,2.65048e-08,0.46089,0.00141391,1.30859e-06,-2.76784e-08,0.462305,0.00141645,1.22556e-06,2.46043e-08,0.463722,0.00141897,1.29937e-06,-1.11341e-08,0.465143,0.00142154,1.26597e-06,-9.87033e-09,0.466565,0.00142404,1.23636e-06,2.08131e-08,0.467991,0.00142657,1.2988e-06,-1.37773e-08,0.469419,0.00142913,1.25746e-06,4.49378e-09,0.470849,0.00143166,1.27094e-06,-4.19781e-09,0.472282,0.00143419,1.25835e-06,1.22975e-08,0.473717,0.00143674,1.29524e-06,-1.51902e-08,0.475155,0.00143929,1.24967e-06,1.86608e-08,0.476596,0.00144184,1.30566e-06,-2.96506e-08,0.478039,0.00144436,1.2167e-06,4.03368e-08,0.479485,0.00144692,1.33771e-06,-4.22896e-08,0.480933,0.00144947,1.21085e-06,3.94148e-08,0.482384,0.00145201,1.32909e-06,-2.59626e-08,0.483837,0.00145459,1.2512e-06,4.83124e-09,0.485293,0.0014571,1.2657e-06,6.63757e-09,0.486751,0.00145966,1.28561e-06,-1.57911e-09,0.488212,0.00146222,1.28087e-06,-3.21468e-10,0.489676,0.00146478,1.27991e-06,2.86517e-09,0.491142,0.00146735,1.2885e-06,-1.11392e-08,0.49261,0.00146989,1.25508e-06,1.18893e-08,0.494081,0.00147244,1.29075e-06,-6.61574e-09,0.495555,0.001475,1.27091e-06,1.45736e-08,0.497031,0.00147759,1.31463e-06,-2.18759e-08,0.49851,0.00148015,1.249e-06,1.33252e-08,0.499992,0.00148269,1.28897e-06,-1.62277e-09,0.501476,0.00148526,1.28411e-06,-6.83421e-09,0.502962,0.00148781,1.2636e-06,2.89596e-08,0.504451,0.00149042,1.35048e-06,-4.93997e-08,0.505943,0.00149298,1.20228e-06,4.94299e-08,0.507437,0.00149553,1.35057e-06,-2.91107e-08,0.508934,0.00149814,1.26324e-06,7.40848e-09,0.510434,0.00150069,1.28547e-06,-5.23187e-10,0.511936,0.00150326,1.2839e-06,-5.31585e-09,0.51344,0.00150581,1.26795e-06,2.17866e-08,0.514947,0.00150841,1.33331e-06,-2.22257e-08,0.516457,0.00151101,1.26663e-06,7.51178e-09,0.517969,0.00151357,1.28917e-06,-7.82128e-09,0.519484,0.00151613,1.2657e-06,2.37733e-08,0.521002,0.00151873,1.33702e-06,-2.76674e-08,0.522522,0.00152132,1.25402e-06,2.72917e-08,0.524044,0.00152391,1.3359e-06,-2.18949e-08,0.525569,0.00152652,1.27021e-06,6.83372e-10,0.527097,0.00152906,1.27226e-06,1.91613e-08,0.528628,0.00153166,1.32974e-06,-1.77241e-08,0.53016,0.00153427,1.27657e-06,-7.86963e-09,0.531696,0.0015368,1.25296e-06,4.92027e-08,0.533234,0.00153945,1.40057e-06,-6.9732e-08,0.534775,0.00154204,1.19138e-06,5.09114e-08,0.536318,0.00154458,1.34411e-06,-1.4704e-08,0.537864,0.00154722,1.3e-06,7.9048e-09,0.539413,0.00154984,1.32371e-06,-1.69152e-08,0.540964,0.00155244,1.27297e-06,1.51355e-10,0.542517,0.00155499,1.27342e-06,1.63099e-08,0.544074,0.00155758,1.32235e-06,-5.78647e-09,0.545633,0.00156021,1.30499e-06,6.83599e-09,0.547194,0.00156284,1.3255e-06,-2.15575e-08,0.548758,0.00156543,1.26083e-06,1.97892e-08,0.550325,0.00156801,1.32019e-06,2.00525e-09,0.551894,0.00157065,1.32621e-06,-2.78103e-08,0.553466,0.00157322,1.24278e-06,4.96314e-08,0.555041,0.00157586,1.39167e-06,-5.1506e-08,0.556618,0.00157849,1.23716e-06,3.71835e-08,0.558198,0.00158107,1.34871e-06,-3.76233e-08,0.55978,0.00158366,1.23584e-06,5.37052e-08,0.561365,0.00158629,1.39695e-06,-5.79884e-08,0.562953,0.00158891,1.22299e-06,5.90392e-08,0.564543,0.00159153,1.4001e-06,-5.89592e-08,0.566136,0.00159416,1.22323e-06,5.7588e-08,0.567731,0.00159678,1.39599e-06,-5.21835e-08,0.569329,0.00159941,1.23944e-06,3.19369e-08,0.57093,0.00160199,1.33525e-06,-1.59594e-08,0.572533,0.00160461,1.28737e-06,3.19006e-08,0.574139,0.00160728,1.38307e-06,-5.20383e-08,0.575748,0.00160989,1.22696e-06,5.70431e-08,0.577359,0.00161251,1.39809e-06,-5.69247e-08,0.578973,0.00161514,1.22731e-06,5.14463e-08,0.580589,0.00161775,1.38165e-06,-2.9651e-08,0.582208,0.00162042,1.2927e-06,7.55339e-09,0.58383,0.00162303,1.31536e-06,-5.62636e-10,0.585455,0.00162566,1.31367e-06,-5.30281e-09,0.587081,0.00162827,1.29776e-06,2.17738e-08,0.588711,0.00163093,1.36309e-06,-2.21875e-08,0.590343,0.00163359,1.29652e-06,7.37164e-09,0.591978,0.00163621,1.31864e-06,-7.29907e-09,0.593616,0.00163882,1.29674e-06,2.18247e-08,0.595256,0.00164148,1.36221e-06,-2.03952e-08,0.596899,0.00164414,1.30103e-06,1.51241e-10,0.598544,0.00164675,1.30148e-06,1.97902e-08,0.600192,0.00164941,1.36085e-06,-1.97074e-08,0.601843,0.00165207,1.30173e-06,-5.65175e-10,0.603496,0.00165467,1.30004e-06,2.1968e-08,0.605152,0.00165734,1.36594e-06,-2.77024e-08,0.606811,0.00165999,1.28283e-06,2.92369e-08,0.608472,0.00166264,1.37054e-06,-2.96407e-08,0.610136,0.00166529,1.28162e-06,2.97215e-08,0.611803,0.00166795,1.37079e-06,-2.96408e-08,0.613472,0.0016706,1.28186e-06,2.92371e-08,0.615144,0.00167325,1.36957e-06,-2.77031e-08,0.616819,0.00167591,1.28647e-06,2.19708e-08,0.618496,0.00167855,1.35238e-06,-5.75407e-10,0.620176,0.00168125,1.35065e-06,-1.9669e-08,0.621858,0.00168389,1.29164e-06,1.96468e-08,0.623544,0.00168653,1.35058e-06,6.86403e-10,0.625232,0.00168924,1.35264e-06,-2.23924e-08,0.626922,0.00169187,1.28547e-06,2.92788e-08,0.628615,0.00169453,1.3733e-06,-3.51181e-08,0.630311,0.00169717,1.26795e-06,5.15889e-08,0.63201,0.00169987,1.42272e-06,-5.2028e-08,0.633711,0.00170255,1.26663e-06,3.73139e-08,0.635415,0.0017052,1.37857e-06,-3.76227e-08,0.637121,0.00170784,1.2657e-06,5.35722e-08,0.63883,0.00171054,1.42642e-06,-5.74567e-08,0.640542,0.00171322,1.25405e-06,5.70456e-08,0.642257,0.0017159,1.42519e-06,-5.15163e-08,0.643974,0.00171859,1.27064e-06,2.98103e-08,0.645694,0.00172122,1.36007e-06,-8.12016e-09,0.647417,0.00172392,1.33571e-06,2.67039e-09,0.649142,0.0017266,1.34372e-06,-2.56152e-09,0.65087,0.00172928,1.33604e-06,7.57571e-09,0.6526,0.00173197,1.35876e-06,-2.77413e-08,0.654334,0.00173461,1.27554e-06,4.3785e-08,0.65607,0.00173729,1.40689e-06,-2.81896e-08,0.657808,0.00174002,1.32233e-06,9.36893e-09,0.65955,0.00174269,1.35043e-06,-9.28617e-09,0.661294,0.00174536,1.32257e-06,2.77757e-08,0.66304,0.00174809,1.4059e-06,-4.2212e-08,0.66479,0.00175078,1.27926e-06,2.1863e-08,0.666542,0.0017534,1.34485e-06,1.43648e-08,0.668297,0.00175613,1.38795e-06,-1.97177e-08,0.670054,0.00175885,1.3288e-06,4.90115e-09,0.671814,0.00176152,1.3435e-06,1.13232e-10,0.673577,0.00176421,1.34384e-06,-5.3542e-09,0.675343,0.00176688,1.32778e-06,2.13035e-08,0.677111,0.0017696,1.39169e-06,-2.02553e-08,0.678882,0.00177232,1.33092e-06,1.13005e-10,0.680656,0.00177499,1.33126e-06,1.98031e-08,0.682432,0.00177771,1.39067e-06,-1.97211e-08,0.684211,0.00178043,1.33151e-06,-5.2349e-10,0.685993,0.00178309,1.32994e-06,2.18151e-08,0.687777,0.00178582,1.39538e-06,-2.71325e-08,0.689564,0.00178853,1.31398e-06,2.71101e-08,0.691354,0.00179124,1.39531e-06,-2.17035e-08,0.693147,0.00179396,1.3302e-06,9.92865e-11,0.694942,0.00179662,1.3305e-06,2.13063e-08,0.69674,0.00179935,1.39442e-06,-2.57198e-08,0.698541,0.00180206,1.31726e-06,2.19682e-08,0.700344,0.00180476,1.38317e-06,-2.54852e-09,0.70215,0.00180752,1.37552e-06,-1.17741e-08,0.703959,0.00181023,1.3402e-06,-9.95999e-09,0.705771,0.00181288,1.31032e-06,5.16141e-08,0.707585,0.00181566,1.46516e-06,-7.72869e-08,0.709402,0.00181836,1.2333e-06,7.87197e-08,0.711222,0.00182106,1.46946e-06,-5.87781e-08,0.713044,0.00182382,1.29312e-06,3.71834e-08,0.714869,0.00182652,1.40467e-06,-3.03511e-08,0.716697,0.00182924,1.31362e-06,2.46161e-08,0.718528,0.00183194,1.38747e-06,-8.5087e-09,0.720361,0.00183469,1.36194e-06,9.41892e-09,0.722197,0.00183744,1.3902e-06,-2.91671e-08,0.724036,0.00184014,1.3027e-06,4.76448e-08,0.725878,0.00184288,1.44563e-06,-4.22028e-08,0.727722,0.00184565,1.31902e-06,1.95682e-09,0.729569,0.00184829,1.3249e-06,3.43754e-08,0.731419,0.00185104,1.42802e-06,-2.0249e-08,0.733271,0.00185384,1.36727e-06,-1.29838e-08,0.735126,0.00185654,1.32832e-06,1.25794e-08,0.736984,0.00185923,1.36606e-06,2.22711e-08,0.738845,0.00186203,1.43287e-06,-4.20594e-08,0.740708,0.00186477,1.3067e-06,2.67571e-08,0.742574,0.00186746,1.38697e-06,-5.36424e-09,0.744443,0.00187022,1.37087e-06,-5.30023e-09,0.746315,0.00187295,1.35497e-06,2.65653e-08,0.748189,0.00187574,1.43467e-06,-4.13564e-08,0.750066,0.00187848,1.3106e-06,1.9651e-08,0.751946,0.00188116,1.36955e-06,2.23572e-08,0.753828,0.00188397,1.43663e-06,-4.9475e-08,0.755714,0.00188669,1.2882e-06,5.63335e-08,0.757602,0.00188944,1.4572e-06,-5.66499e-08,0.759493,0.00189218,1.28725e-06,5.10567e-08,0.761386,0.00189491,1.44042e-06,-2.83677e-08,0.763283,0.00189771,1.35532e-06,2.80962e-09,0.765182,0.00190042,1.36375e-06,1.71293e-08,0.767083,0.0019032,1.41513e-06,-1.17221e-08,0.768988,0.001906,1.37997e-06,-2.98453e-08,0.770895,0.00190867,1.29043e-06,7.14987e-08,0.772805,0.00191146,1.50493e-06,-7.73354e-08,0.774718,0.00191424,1.27292e-06,5.90292e-08,0.776634,0.00191697,1.45001e-06,-3.9572e-08,0.778552,0.00191975,1.33129e-06,3.9654e-08,0.780473,0.00192253,1.45026e-06,-5.94395e-08,0.782397,0.00192525,1.27194e-06,7.88945e-08,0.784324,0.00192803,1.50862e-06,-7.73249e-08,0.786253,0.00193082,1.27665e-06,5.15913e-08,0.788185,0.00193352,1.43142e-06,-9.83099e-09,0.79012,0.00193636,1.40193e-06,-1.22672e-08,0.792058,0.00193912,1.36513e-06,-7.05275e-10,0.793999,0.00194185,1.36301e-06,1.50883e-08,0.795942,0.00194462,1.40828e-06,-4.33147e-11,0.797888,0.00194744,1.40815e-06,-1.49151e-08,0.799837,0.00195021,1.3634e-06,9.93244e-11,0.801788,0.00195294,1.3637e-06,1.45179e-08,0.803743,0.00195571,1.40725e-06,1.43363e-09,0.8057,0.00195853,1.41155e-06,-2.02525e-08,0.80766,0.00196129,1.35079e-06,1.99718e-08,0.809622,0.00196405,1.41071e-06,-3.01649e-11,0.811588,0.00196687,1.41062e-06,-1.9851e-08,0.813556,0.00196964,1.35107e-06,1.98296e-08,0.815527,0.0019724,1.41056e-06,1.37485e-10,0.817501,0.00197522,1.41097e-06,-2.03796e-08,0.819477,0.00197798,1.34983e-06,2.17763e-08,0.821457,0.00198074,1.41516e-06,-7.12085e-09,0.823439,0.00198355,1.3938e-06,6.70707e-09,0.825424,0.00198636,1.41392e-06,-1.97074e-08,0.827412,0.00198913,1.35479e-06,1.25179e-08,0.829402,0.00199188,1.39235e-06,2.92405e-08,0.831396,0.00199475,1.48007e-06,-6.98755e-08,0.833392,0.0019975,1.27044e-06,7.14477e-08,0.835391,0.00200026,1.48479e-06,-3.71014e-08,0.837392,0.00200311,1.37348e-06,1.73533e-08,0.839397,0.00200591,1.42554e-06,-3.23118e-08,0.841404,0.00200867,1.32861e-06,5.2289e-08,0.843414,0.00201148,1.48547e-06,-5.76348e-08,0.845427,0.00201428,1.31257e-06,5.9041e-08,0.847443,0.00201708,1.48969e-06,-5.93197e-08,0.849461,0.00201988,1.31173e-06,5.90289e-08,0.851482,0.00202268,1.48882e-06,-5.75864e-08,0.853507,0.00202549,1.31606e-06,5.21075e-08,0.855533,0.00202828,1.47238e-06,-3.16344e-08,0.857563,0.00203113,1.37748e-06,1.48257e-08,0.859596,0.00203393,1.42196e-06,-2.76684e-08,0.861631,0.00203669,1.33895e-06,3.62433e-08,0.863669,0.00203947,1.44768e-06,1.90463e-09,0.86571,0.00204237,1.45339e-06,-4.38617e-08,0.867754,0.00204515,1.32181e-06,5.43328e-08,0.8698,0.00204796,1.48481e-06,-5.42603e-08,0.87185,0.00205076,1.32203e-06,4.34989e-08,0.873902,0.00205354,1.45252e-06,-5.26029e-10,0.875957,0.00205644,1.45095e-06,-4.13949e-08,0.878015,0.00205922,1.32676e-06,4.68962e-08,0.880075,0.00206201,1.46745e-06,-2.69807e-08,0.882139,0.00206487,1.38651e-06,1.42181e-09,0.884205,0.00206764,1.39077e-06,2.12935e-08,0.886274,0.00207049,1.45465e-06,-2.69912e-08,0.888346,0.00207332,1.37368e-06,2.70664e-08,0.890421,0.00207615,1.45488e-06,-2.16698e-08,0.892498,0.00207899,1.38987e-06,8.14756e-12,0.894579,0.00208177,1.38989e-06,2.16371e-08,0.896662,0.00208462,1.45481e-06,-2.6952e-08,0.898748,0.00208744,1.37395e-06,2.65663e-08,0.900837,0.00209027,1.45365e-06,-1.97084e-08,0.902928,0.00209312,1.39452e-06,-7.33731e-09,0.905023,0.00209589,1.37251e-06,4.90578e-08,0.90712,0.00209878,1.51968e-06,-6.96845e-08,0.90922,0.00210161,1.31063e-06,5.08664e-08,0.911323,0.00210438,1.46323e-06,-1.45717e-08,0.913429,0.00210727,1.41952e-06,7.42038e-09,0.915538,0.00211013,1.44178e-06,-1.51097e-08,0.917649,0.00211297,1.39645e-06,-6.58618e-09,0.919764,0.00211574,1.37669e-06,4.14545e-08,0.921881,0.00211862,1.50105e-06,-4.00222e-08,0.924001,0.0021215,1.38099e-06,-5.7518e-10,0.926124,0.00212426,1.37926e-06,4.23229e-08,0.92825,0.00212714,1.50623e-06,-4.9507e-08,0.930378,0.00213001,1.35771e-06,3.64958e-08,0.93251,0.00213283,1.4672e-06,-3.68713e-08,0.934644,0.00213566,1.35658e-06,5.13848e-08,0.936781,0.00213852,1.51074e-06,-4.94585e-08,0.938921,0.0021414,1.36236e-06,2.72399e-08,0.941064,0.0021442,1.44408e-06,1.0372e-10,0.943209,0.00214709,1.44439e-06,-2.76547e-08,0.945358,0.0021499,1.36143e-06,5.09106e-08,0.947509,0.00215277,1.51416e-06,-5.67784e-08,0.949663,0.00215563,1.34382e-06,5.69935e-08,0.95182,0.00215849,1.5148e-06,-5.19861e-08,0.95398,0.00216136,1.35885e-06,3.17417e-08,0.956143,0.00216418,1.45407e-06,-1.53758e-08,0.958309,0.00216704,1.40794e-06,2.97615e-08,0.960477,0.00216994,1.49723e-06,-4.40657e-08,0.962649,0.00217281,1.36503e-06,2.72919e-08,0.964823,0.00217562,1.44691e-06,-5.49729e-09,0.967,0.0021785,1.43041e-06,-5.30273e-09,0.96918,0.00218134,1.41451e-06,2.67084e-08,0.971363,0.00218425,1.49463e-06,-4.19265e-08,0.973548,0.00218711,1.36885e-06,2.17881e-08,0.975737,0.00218992,1.43422e-06,1.43789e-08,0.977928,0.00219283,1.47735e-06,-1.96989e-08,0.980122,0.00219572,1.41826e-06,4.81221e-09,0.98232,0.00219857,1.43269e-06,4.50048e-10,0.98452,0.00220144,1.43404e-06,-6.61237e-09,0.986722,0.00220429,1.41421e-06,2.59993e-08,0.988928,0.0022072,1.4922e-06,-3.77803e-08,0.991137,0.00221007,1.37886e-06,5.9127e-09,0.993348,0.00221284,1.3966e-06,1.33339e-07,0.995563,0.00221604,1.79662e-06,-5.98872e-07,0.99778,0.00222015,0.,0.};
-
-        template <bool srgb, int blueIdx, typename T, typename D>
-        __device__ __forceinline__ void RGB2LabConvert_f(const T& src, D& dst)
-        {
-            const float _1_3 = 1.0f / 3.0f;
-            const float _a = 16.0f / 116.0f;
-
-            float B = blueIdx == 0 ? src.x : src.z;
-            float G = src.y;
-            float R = blueIdx == 0 ? src.z : src.x;
-
-            if (srgb)
-            {
-                B = splineInterpolate(B * GAMMA_TAB_SIZE, c_sRGBGammaTab, GAMMA_TAB_SIZE);
-                G = splineInterpolate(G * GAMMA_TAB_SIZE, c_sRGBGammaTab, GAMMA_TAB_SIZE);
-                R = splineInterpolate(R * GAMMA_TAB_SIZE, c_sRGBGammaTab, GAMMA_TAB_SIZE);
-            }
-
-            float X = B * 0.189828f + G * 0.376219f + R * 0.433953f;
-            float Y = B * 0.072169f + G * 0.715160f + R * 0.212671f;
-            float Z = B * 0.872766f + G * 0.109477f + R * 0.017758f;
-
-            float FX = X > 0.008856f ? ::powf(X, _1_3) : (7.787f * X + _a);
-            float FY = Y > 0.008856f ? ::powf(Y, _1_3) : (7.787f * Y + _a);
-            float FZ = Z > 0.008856f ? ::powf(Z, _1_3) : (7.787f * Z + _a);
-
-            float L = Y > 0.008856f ? (116.f * FY - 16.f) : (903.3f * Y);
-            float a = 500.f * (FX - FY);
-            float b = 200.f * (FY - FZ);
-
-            dst.x = L;
-            dst.y = a;
-            dst.z = b;
-        }
-
-        template <typename T, int scn, int dcn, bool srgb, int blueIdx> struct RGB2Lab;
-        template <int scn, int dcn, bool srgb, int blueIdx>
-        struct RGB2Lab<uchar, scn, dcn, srgb, blueIdx>
-            : unary_function<typename TypeVec<uchar, scn>::vec_type, typename TypeVec<uchar, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<uchar, dcn>::vec_type operator ()(const typename TypeVec<uchar, scn>::vec_type& src) const
-            {
-                typename TypeVec<uchar, dcn>::vec_type dst;
-
-                RGB2LabConvert_b<srgb, blueIdx>(src, dst);
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ RGB2Lab() {}
-            __host__ __device__ __forceinline__ RGB2Lab(const RGB2Lab&) {}
-        };
-        template <int scn, int dcn, bool srgb, int blueIdx>
-        struct RGB2Lab<float, scn, dcn, srgb, blueIdx>
-            : unary_function<typename TypeVec<float, scn>::vec_type, typename TypeVec<float, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<float, dcn>::vec_type operator ()(const typename TypeVec<float, scn>::vec_type& src) const
-            {
-                typename TypeVec<float, dcn>::vec_type dst;
-
-                RGB2LabConvert_f<srgb, blueIdx>(src, dst);
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ RGB2Lab() {}
-            __host__ __device__ __forceinline__ RGB2Lab(const RGB2Lab&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_RGB2Lab_TRAITS(name, scn, dcn, srgb, blueIdx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2Lab<T, scn, dcn, srgb, blueIdx> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-    namespace color_detail
-    {
-        __constant__ float c_sRGBInvGammaTab[] = {0,0.0126255,0.,-8.33961e-06,0.0126172,0.0126005,-2.50188e-05,4.1698e-05,0.0252344,0.0126756,0.000100075,-0.000158451,0.0378516,0.0124004,-0.000375277,-0.000207393,0.0496693,0.0110276,-0.000997456,0.00016837,0.0598678,0.00953783,-0.000492346,2.07235e-05,0.068934,0.00861531,-0.000430176,3.62876e-05,0.0771554,0.00786382,-0.000321313,1.87625e-05,0.0847167,0.00727748,-0.000265025,1.53594e-05,0.0917445,0.00679351,-0.000218947,1.10545e-05,0.0983301,0.00638877,-0.000185784,8.66984e-06,0.104542,0.00604322,-0.000159774,6.82996e-06,0.110432,0.00574416,-0.000139284,5.51008e-06,0.116042,0.00548212,-0.000122754,4.52322e-06,0.121406,0.00525018,-0.000109184,3.75557e-06,0.126551,0.00504308,-9.79177e-05,3.17134e-06,0.131499,0.00485676,-8.84037e-05,2.68469e-06,0.13627,0.004688,-8.03496e-05,2.31725e-06,0.14088,0.00453426,-7.33978e-05,2.00868e-06,0.145343,0.00439349,-6.73718e-05,1.74775e-06,0.149671,0.00426399,-6.21286e-05,1.53547e-06,0.153875,0.00414434,-5.75222e-05,1.364e-06,0.157963,0.00403338,-5.34301e-05,1.20416e-06,0.161944,0.00393014,-4.98177e-05,1.09114e-06,0.165825,0.00383377,-4.65443e-05,9.57987e-07,0.169613,0.00374356,-4.36703e-05,8.88359e-07,0.173314,0.00365888,-4.10052e-05,7.7849e-07,0.176933,0.00357921,-3.86697e-05,7.36254e-07,0.180474,0.00350408,-3.6461e-05,6.42534e-07,0.183942,0.00343308,-3.45334e-05,6.12614e-07,0.187342,0.00336586,-3.26955e-05,5.42894e-07,0.190675,0.00330209,-3.10669e-05,5.08967e-07,0.193947,0.00324149,-2.954e-05,4.75977e-07,0.197159,0.00318383,-2.8112e-05,4.18343e-07,0.200315,0.00312887,-2.6857e-05,4.13651e-07,0.203418,0.00307639,-2.5616e-05,3.70847e-07,0.206469,0.00302627,-2.45035e-05,3.3813e-07,0.209471,0.00297828,-2.34891e-05,3.32999e-07,0.212426,0.0029323,-2.24901e-05,2.96826e-07,0.215336,0.00288821,-2.15996e-05,2.82736e-07,0.218203,0.00284586,-2.07514e-05,2.70961e-07,0.221029,0.00280517,-1.99385e-05,2.42744e-07,0.223814,0.00276602,-1.92103e-05,2.33277e-07,0.226561,0.0027283,-1.85105e-05,2.2486e-07,0.229271,0.00269195,-1.78359e-05,2.08383e-07,0.231945,0.00265691,-1.72108e-05,1.93305e-07,0.234585,0.00262307,-1.66308e-05,1.80687e-07,0.237192,0.00259035,-1.60888e-05,1.86632e-07,0.239766,0.00255873,-1.55289e-05,1.60569e-07,0.24231,0.00252815,-1.50472e-05,1.54566e-07,0.244823,0.00249852,-1.45835e-05,1.59939e-07,0.247307,0.00246983,-1.41037e-05,1.29549e-07,0.249763,0.00244202,-1.3715e-05,1.41429e-07,0.252191,0.00241501,-1.32907e-05,1.39198e-07,0.254593,0.00238885,-1.28731e-05,1.06444e-07,0.256969,0.00236342,-1.25538e-05,1.2048e-07,0.25932,0.00233867,-1.21924e-05,1.26892e-07,0.261647,0.00231467,-1.18117e-05,8.72084e-08,0.26395,0.00229131,-1.15501e-05,1.20323e-07,0.26623,0.00226857,-1.11891e-05,8.71514e-08,0.268487,0.00224645,-1.09276e-05,9.73165e-08,0.270723,0.00222489,-1.06357e-05,8.98259e-08,0.272937,0.00220389,-1.03662e-05,7.98218e-08,0.275131,0.00218339,-1.01267e-05,9.75254e-08,0.277304,0.00216343,-9.83416e-06,6.65195e-08,0.279458,0.00214396,-9.63461e-06,8.34313e-08,0.281592,0.00212494,-9.38431e-06,7.65919e-08,0.283708,0.00210641,-9.15454e-06,5.7236e-08,0.285805,0.00208827,-8.98283e-06,8.18939e-08,0.287885,0.00207055,-8.73715e-06,6.2224e-08,0.289946,0.00205326,-8.55047e-06,5.66388e-08,0.291991,0.00203633,-8.38056e-06,6.88491e-08,0.294019,0.00201978,-8.17401e-06,5.53955e-08,0.296031,0.00200359,-8.00782e-06,6.71971e-08,0.298027,0.00198778,-7.80623e-06,3.34439e-08,0.300007,0.00197227,-7.7059e-06,6.7248e-08,0.301971,0.00195706,-7.50416e-06,5.51915e-08,0.303921,0.00194221,-7.33858e-06,3.98124e-08,0.305856,0.00192766,-7.21915e-06,5.37795e-08,0.307776,0.00191338,-7.05781e-06,4.30919e-08,0.309683,0.00189939,-6.92853e-06,4.20744e-08,0.311575,0.00188566,-6.80231e-06,5.68321e-08,0.313454,0.00187223,-6.63181e-06,2.86195e-08,0.31532,0.00185905,-6.54595e-06,3.73075e-08,0.317172,0.00184607,-6.43403e-06,6.05684e-08,0.319012,0.00183338,-6.25233e-06,1.84426e-08,0.320839,0.00182094,-6.197e-06,4.44757e-08,0.322654,0.00180867,-6.06357e-06,4.20729e-08,0.324456,0.00179667,-5.93735e-06,2.56511e-08,0.326247,0.00178488,-5.8604e-06,3.41368e-08,0.328026,0.00177326,-5.75799e-06,4.64177e-08,0.329794,0.00176188,-5.61874e-06,1.86107e-08,0.33155,0.0017507,-5.5629e-06,2.81511e-08,0.333295,0.00173966,-5.47845e-06,4.75987e-08,0.335029,0.00172884,-5.33565e-06,1.98726e-08,0.336753,0.00171823,-5.27604e-06,2.19226e-08,0.338466,0.00170775,-5.21027e-06,4.14483e-08,0.340169,0.00169745,-5.08592e-06,2.09017e-08,0.341861,0.00168734,-5.02322e-06,2.39561e-08,0.343543,0.00167737,-4.95135e-06,3.22852e-08,0.345216,0.00166756,-4.85449e-06,2.57173e-08,0.346878,0.00165793,-4.77734e-06,1.38569e-08,0.348532,0.00164841,-4.73577e-06,3.80634e-08,0.350175,0.00163906,-4.62158e-06,1.27043e-08,0.35181,0.00162985,-4.58347e-06,3.03279e-08,0.353435,0.00162078,-4.49249e-06,1.49961e-08,0.355051,0.00161184,-4.4475e-06,2.88977e-08,0.356659,0.00160303,-4.3608e-06,1.84241e-08,0.358257,0.00159436,-4.30553e-06,1.6616e-08,0.359848,0.0015858,-4.25568e-06,3.43218e-08,0.361429,0.00157739,-4.15272e-06,-4.89172e-09,0.363002,0.00156907,-4.16739e-06,4.48498e-08,0.364567,0.00156087,-4.03284e-06,4.30676e-09,0.366124,0.00155282,-4.01992e-06,2.73303e-08,0.367673,0.00154486,-3.93793e-06,5.58036e-09,0.369214,0.001537,-3.92119e-06,3.97554e-08,0.370747,0.00152928,-3.80193e-06,-1.55904e-08,0.372272,0.00152163,-3.8487e-06,5.24081e-08,0.37379,0.00151409,-3.69147e-06,-1.52272e-08,0.375301,0.00150666,-3.73715e-06,3.83028e-08,0.376804,0.0014993,-3.62225e-06,1.10278e-08,0.378299,0.00149209,-3.58916e-06,6.99326e-09,0.379788,0.00148493,-3.56818e-06,2.06038e-08,0.381269,0.00147786,-3.50637e-06,2.98009e-08,0.382744,0.00147093,-3.41697e-06,-2.05978e-08,0.384211,0.00146404,-3.47876e-06,5.25899e-08,0.385672,0.00145724,-3.32099e-06,-1.09471e-08,0.387126,0.00145056,-3.35383e-06,2.10009e-08,0.388573,0.00144392,-3.29083e-06,1.63501e-08,0.390014,0.00143739,-3.24178e-06,3.00641e-09,0.391448,0.00143091,-3.23276e-06,3.12282e-08,0.392875,0.00142454,-3.13908e-06,-8.70932e-09,0.394297,0.00141824,-3.16521e-06,3.34114e-08,0.395712,0.00141201,-3.06497e-06,-5.72754e-09,0.397121,0.00140586,-3.08215e-06,1.9301e-08,0.398524,0.00139975,-3.02425e-06,1.7931e-08,0.39992,0.00139376,-2.97046e-06,-1.61822e-09,0.401311,0.00138781,-2.97531e-06,1.83442e-08,0.402696,0.00138192,-2.92028e-06,1.76485e-08,0.404075,0.00137613,-2.86733e-06,4.68617e-10,0.405448,0.00137039,-2.86593e-06,1.02794e-08,0.406816,0.00136469,-2.83509e-06,1.80179e-08,0.408178,0.00135908,-2.78104e-06,7.05594e-09,0.409534,0.00135354,-2.75987e-06,1.33633e-08,0.410885,0.00134806,-2.71978e-06,-9.04568e-10,0.41223,0.00134261,-2.72249e-06,2.0057e-08,0.41357,0.00133723,-2.66232e-06,1.00841e-08,0.414905,0.00133194,-2.63207e-06,-7.88835e-10,0.416234,0.00132667,-2.63444e-06,2.28734e-08,0.417558,0.00132147,-2.56582e-06,-1.29785e-09,0.418877,0.00131633,-2.56971e-06,1.21205e-08,0.420191,0.00131123,-2.53335e-06,1.24202e-08,0.421499,0.0013062,-2.49609e-06,-2.19681e-09,0.422803,0.0013012,-2.50268e-06,2.61696e-08,0.424102,0.00129628,-2.42417e-06,-1.30747e-08,0.425396,0.00129139,-2.46339e-06,2.6129e-08,0.426685,0.00128654,-2.38501e-06,-2.03454e-09,0.427969,0.00128176,-2.39111e-06,1.18115e-08,0.429248,0.00127702,-2.35567e-06,1.43932e-08,0.430523,0.00127235,-2.31249e-06,-9.77965e-09,0.431793,0.00126769,-2.34183e-06,2.47253e-08,0.433058,0.00126308,-2.26766e-06,2.85278e-10,0.434319,0.00125855,-2.2668e-06,3.93614e-09,0.435575,0.00125403,-2.25499e-06,1.37722e-08,0.436827,0.00124956,-2.21368e-06,5.79803e-10,0.438074,0.00124513,-2.21194e-06,1.37112e-08,0.439317,0.00124075,-2.1708e-06,4.17973e-09,0.440556,0.00123642,-2.15826e-06,-6.27703e-10,0.44179,0.0012321,-2.16015e-06,2.81332e-08,0.44302,0.00122787,-2.07575e-06,-2.24985e-08,0.444246,0.00122365,-2.14324e-06,3.20586e-08,0.445467,0.00121946,-2.04707e-06,-1.6329e-08,0.446685,0.00121532,-2.09605e-06,3.32573e-08,0.447898,0.00121122,-1.99628e-06,-2.72927e-08,0.449107,0.00120715,-2.07816e-06,4.6111e-08,0.450312,0.00120313,-1.93983e-06,-3.79416e-08,0.451514,0.00119914,-2.05365e-06,4.60507e-08,0.452711,0.00119517,-1.9155e-06,-2.7052e-08,0.453904,0.00119126,-1.99666e-06,3.23551e-08,0.455093,0.00118736,-1.89959e-06,-1.29613e-08,0.456279,0.00118352,-1.93848e-06,1.94905e-08,0.45746,0.0011797,-1.88e-06,-5.39588e-09,0.458638,0.00117593,-1.89619e-06,2.09282e-09,0.459812,0.00117214,-1.88991e-06,2.68267e-08,0.460982,0.00116844,-1.80943e-06,-1.99925e-08,0.462149,0.00116476,-1.86941e-06,2.3341e-08,0.463312,0.00116109,-1.79939e-06,-1.37674e-08,0.464471,0.00115745,-1.84069e-06,3.17287e-08,0.465627,0.00115387,-1.7455e-06,-2.37407e-08,0.466779,0.00115031,-1.81673e-06,3.34315e-08,0.467927,0.00114677,-1.71643e-06,-2.05786e-08,0.469073,0.00114328,-1.77817e-06,1.90802e-08,0.470214,0.00113978,-1.72093e-06,3.86247e-09,0.471352,0.00113635,-1.70934e-06,-4.72759e-09,0.472487,0.00113292,-1.72352e-06,1.50478e-08,0.473618,0.00112951,-1.67838e-06,4.14108e-09,0.474746,0.00112617,-1.66595e-06,-1.80986e-09,0.47587,0.00112283,-1.67138e-06,3.09816e-09,0.476991,0.0011195,-1.66209e-06,1.92198e-08,0.478109,0.00111623,-1.60443e-06,-2.03726e-08,0.479224,0.00111296,-1.66555e-06,3.2468e-08,0.480335,0.00110973,-1.56814e-06,-2.00922e-08,0.481443,0.00110653,-1.62842e-06,1.80983e-08,0.482548,0.00110333,-1.57413e-06,7.30362e-09,0.48365,0.0011002,-1.55221e-06,-1.75107e-08,0.484749,0.00109705,-1.60475e-06,3.29373e-08,0.485844,0.00109393,-1.50594e-06,-2.48315e-08,0.486937,0.00109085,-1.58043e-06,3.65865e-08,0.488026,0.0010878,-1.47067e-06,-3.21078e-08,0.489112,0.00108476,-1.56699e-06,3.22397e-08,0.490195,0.00108172,-1.47027e-06,-7.44391e-09,0.491276,0.00107876,-1.49261e-06,-2.46428e-09,0.492353,0.00107577,-1.5e-06,1.73011e-08,0.493427,0.00107282,-1.4481e-06,-7.13552e-09,0.494499,0.0010699,-1.4695e-06,1.1241e-08,0.495567,0.001067,-1.43578e-06,-8.02637e-09,0.496633,0.0010641,-1.45986e-06,2.08645e-08,0.497695,0.00106124,-1.39726e-06,-1.58271e-08,0.498755,0.0010584,-1.44475e-06,1.26415e-08,0.499812,0.00105555,-1.40682e-06,2.48655e-08,0.500866,0.00105281,-1.33222e-06,-5.24988e-08,0.501918,0.00104999,-1.48972e-06,6.59206e-08,0.502966,0.00104721,-1.29196e-06,-3.237e-08,0.504012,0.00104453,-1.38907e-06,3.95479e-09,0.505055,0.00104176,-1.3772e-06,1.65509e-08,0.506096,0.00103905,-1.32755e-06,-1.05539e-08,0.507133,0.00103637,-1.35921e-06,2.56648e-08,0.508168,0.00103373,-1.28222e-06,-3.25007e-08,0.509201,0.00103106,-1.37972e-06,4.47336e-08,0.51023,0.00102844,-1.24552e-06,-2.72245e-08,0.511258,0.00102587,-1.32719e-06,4.55952e-09,0.512282,0.00102323,-1.31352e-06,8.98645e-09,0.513304,0.00102063,-1.28656e-06,1.90992e-08,0.514323,0.00101811,-1.22926e-06,-2.57786e-08,0.51534,0.00101557,-1.30659e-06,2.44104e-08,0.516355,0.00101303,-1.23336e-06,-1.22581e-08,0.517366,0.00101053,-1.27014e-06,2.4622e-08,0.518376,0.00100806,-1.19627e-06,-2.66253e-08,0.519383,0.00100559,-1.27615e-06,2.22744e-08,0.520387,0.00100311,-1.20932e-06,-2.8679e-09,0.521389,0.00100068,-1.21793e-06,-1.08029e-08,0.522388,0.000998211,-1.25034e-06,4.60795e-08,0.523385,0.000995849,-1.1121e-06,-5.4306e-08,0.52438,0.000993462,-1.27502e-06,5.19354e-08,0.525372,0.000991067,-1.11921e-06,-3.42262e-08,0.526362,0.000988726,-1.22189e-06,2.53646e-08,0.52735,0.000986359,-1.14579e-06,-7.62782e-09,0.528335,0.000984044,-1.16868e-06,5.14668e-09,0.529318,0.000981722,-1.15324e-06,-1.29589e-08,0.530298,0.000979377,-1.19211e-06,4.66888e-08,0.531276,0.000977133,-1.05205e-06,-5.45868e-08,0.532252,0.000974865,-1.21581e-06,5.24495e-08,0.533226,0.000972591,-1.05846e-06,-3.60019e-08,0.534198,0.000970366,-1.16647e-06,3.19537e-08,0.535167,0.000968129,-1.07061e-06,-3.2208e-08,0.536134,0.000965891,-1.16723e-06,3.72738e-08,0.537099,0.000963668,-1.05541e-06,2.32205e-09,0.538061,0.000961564,-1.04844e-06,-4.65618e-08,0.539022,0.000959328,-1.18813e-06,6.47159e-08,0.53998,0.000957146,-9.93979e-07,-3.3488e-08,0.540936,0.000955057,-1.09444e-06,9.63166e-09,0.54189,0.000952897,-1.06555e-06,-5.03871e-09,0.542842,0.000950751,-1.08066e-06,1.05232e-08,0.543792,0.000948621,-1.04909e-06,2.25503e-08,0.544739,0.000946591,-9.81444e-07,-4.11195e-08,0.545685,0.000944504,-1.1048e-06,2.27182e-08,0.546628,0.000942363,-1.03665e-06,9.85146e-09,0.54757,0.000940319,-1.00709e-06,-2.51938e-09,0.548509,0.000938297,-1.01465e-06,2.25858e-10,0.549446,0.000936269,-1.01397e-06,1.61598e-09,0.550381,0.000934246,-1.00913e-06,-6.68983e-09,0.551315,0.000932207,-1.0292e-06,2.51434e-08,0.552246,0.000930224,-9.53765e-07,-3.42793e-08,0.553175,0.000928214,-1.0566e-06,5.23688e-08,0.554102,0.000926258,-8.99497e-07,-5.59865e-08,0.555028,0.000924291,-1.06746e-06,5.23679e-08,0.555951,0.000922313,-9.10352e-07,-3.42763e-08,0.556872,0.00092039,-1.01318e-06,2.51326e-08,0.557792,0.000918439,-9.37783e-07,-6.64954e-09,0.558709,0.000916543,-9.57732e-07,1.46554e-09,0.559625,0.000914632,-9.53335e-07,7.87281e-10,0.560538,0.000912728,-9.50973e-07,-4.61466e-09,0.56145,0.000910812,-9.64817e-07,1.76713e-08,0.56236,0.000908935,-9.11804e-07,-6.46564e-09,0.563268,0.000907092,-9.312e-07,8.19121e-09,0.564174,0.000905255,-9.06627e-07,-2.62992e-08,0.565078,0.000903362,-9.85524e-07,3.74007e-08,0.565981,0.000901504,-8.73322e-07,-4.0942e-09,0.566882,0.000899745,-8.85605e-07,-2.1024e-08,0.56778,0.00089791,-9.48677e-07,2.85854e-08,0.568677,0.000896099,-8.62921e-07,-3.3713e-08,0.569573,0.000894272,-9.64059e-07,4.6662e-08,0.570466,0.000892484,-8.24073e-07,-3.37258e-08,0.571358,0.000890734,-9.25251e-07,2.86365e-08,0.572247,0.00088897,-8.39341e-07,-2.12155e-08,0.573135,0.000887227,-9.02988e-07,-3.37913e-09,0.574022,0.000885411,-9.13125e-07,3.47319e-08,0.574906,0.000883689,-8.08929e-07,-1.63394e-08,0.575789,0.000882022,-8.57947e-07,-2.8979e-08,0.57667,0.00088022,-9.44885e-07,7.26509e-08,0.57755,0.000878548,-7.26932e-07,-8.28106e-08,0.578427,0.000876845,-9.75364e-07,7.97774e-08,0.579303,0.000875134,-7.36032e-07,-5.74849e-08,0.580178,0.00087349,-9.08486e-07,3.09529e-08,0.58105,0.000871765,-8.15628e-07,-6.72206e-09,0.581921,0.000870114,-8.35794e-07,-4.06451e-09,0.582791,0.00086843,-8.47987e-07,2.29799e-08,0.583658,0.000866803,-7.79048e-07,-2.82503e-08,0.584524,0.00086516,-8.63799e-07,3.04167e-08,0.585388,0.000863524,-7.72548e-07,-3.38119e-08,0.586251,0.000861877,-8.73984e-07,4.52264e-08,0.587112,0.000860265,-7.38305e-07,-2.78842e-08,0.587972,0.000858705,-8.21958e-07,6.70567e-09,0.58883,0.000857081,-8.01841e-07,1.06161e-09,0.589686,0.000855481,-7.98656e-07,-1.09521e-08,0.590541,0.00085385,-8.31512e-07,4.27468e-08,0.591394,0.000852316,-7.03272e-07,-4.08257e-08,0.592245,0.000850787,-8.25749e-07,1.34677e-09,0.593095,0.000849139,-8.21709e-07,3.54387e-08,0.593944,0.000847602,-7.15393e-07,-2.38924e-08,0.59479,0.0008461,-7.8707e-07,5.26143e-10,0.595636,0.000844527,-7.85491e-07,2.17879e-08,0.596479,0.000843021,-7.20127e-07,-2.80733e-08,0.597322,0.000841497,-8.04347e-07,3.09005e-08,0.598162,0.000839981,-7.11646e-07,-3.5924e-08,0.599002,0.00083845,-8.19418e-07,5.3191e-08,0.599839,0.000836971,-6.59845e-07,-5.76307e-08,0.600676,0.000835478,-8.32737e-07,5.81227e-08,0.60151,0.000833987,-6.58369e-07,-5.56507e-08,0.602344,0.000832503,-8.25321e-07,4.52706e-08,0.603175,0.000830988,-6.89509e-07,-6.22236e-09,0.604006,0.000829591,-7.08176e-07,-2.03811e-08,0.604834,0.000828113,-7.6932e-07,2.8142e-08,0.605662,0.000826659,-6.84894e-07,-3.25822e-08,0.606488,0.000825191,-7.8264e-07,4.25823e-08,0.607312,0.000823754,-6.54893e-07,-1.85376e-08,0.608135,0.000822389,-7.10506e-07,-2.80365e-08,0.608957,0.000820883,-7.94616e-07,7.1079e-08,0.609777,0.000819507,-5.81379e-07,-7.74655e-08,0.610596,0.000818112,-8.13775e-07,5.9969e-08,0.611413,0.000816665,-6.33868e-07,-4.32013e-08,0.612229,0.000815267,-7.63472e-07,5.32313e-08,0.613044,0.0008139,-6.03778e-07,-5.05148e-08,0.613857,0.000812541,-7.55323e-07,2.96187e-08,0.614669,0.000811119,-6.66466e-07,-8.35545e-09,0.615479,0.000809761,-6.91533e-07,3.80301e-09,0.616288,0.00080839,-6.80124e-07,-6.85666e-09,0.617096,0.000807009,-7.00694e-07,2.36237e-08,0.617903,0.000805678,-6.29822e-07,-2.80336e-08,0.618708,0.000804334,-7.13923e-07,2.8906e-08,0.619511,0.000802993,-6.27205e-07,-2.79859e-08,0.620314,0.000801655,-7.11163e-07,2.34329e-08,0.621114,0.000800303,-6.40864e-07,-6.14108e-09,0.621914,0.000799003,-6.59287e-07,1.13151e-09,0.622712,0.000797688,-6.55893e-07,1.61507e-09,0.62351,0.000796381,-6.51048e-07,-7.59186e-09,0.624305,0.000795056,-6.73823e-07,2.87524e-08,0.6251,0.000793794,-5.87566e-07,-4.7813e-08,0.625893,0.000792476,-7.31005e-07,4.32901e-08,0.626685,0.000791144,-6.01135e-07,-6.13814e-09,0.627475,0.000789923,-6.19549e-07,-1.87376e-08,0.628264,0.000788628,-6.75762e-07,2.14837e-08,0.629052,0.000787341,-6.11311e-07,-7.59265e-09,0.629839,0.000786095,-6.34089e-07,8.88692e-09,0.630625,0.000784854,-6.07428e-07,-2.7955e-08,0.631409,0.000783555,-6.91293e-07,4.33285e-08,0.632192,0.000782302,-5.61307e-07,-2.61497e-08,0.632973,0.000781101,-6.39757e-07,1.6658e-09,0.633754,0.000779827,-6.34759e-07,1.94866e-08,0.634533,0.000778616,-5.76299e-07,-2.00076e-08,0.635311,0.000777403,-6.36322e-07,9.39091e-10,0.636088,0.000776133,-6.33505e-07,1.62512e-08,0.636863,0.000774915,-5.84751e-07,-6.33937e-09,0.637638,0.000773726,-6.03769e-07,9.10609e-09,0.638411,0.000772546,-5.76451e-07,-3.00849e-08,0.639183,0.000771303,-6.66706e-07,5.1629e-08,0.639953,0.000770125,-5.11819e-07,-5.7222e-08,0.640723,0.000768929,-6.83485e-07,5.80497e-08,0.641491,0.000767736,-5.09336e-07,-5.57674e-08,0.642259,0.000766551,-6.76638e-07,4.58105e-08,0.643024,0.000765335,-5.39206e-07,-8.26541e-09,0.643789,0.000764231,-5.64002e-07,-1.27488e-08,0.644553,0.000763065,-6.02249e-07,-3.44168e-10,0.645315,0.00076186,-6.03281e-07,1.41254e-08,0.646077,0.000760695,-5.60905e-07,3.44727e-09,0.646837,0.000759584,-5.50563e-07,-2.79144e-08,0.647596,0.000758399,-6.34307e-07,4.86057e-08,0.648354,0.000757276,-4.88489e-07,-4.72989e-08,0.64911,0.000756158,-6.30386e-07,2.13807e-08,0.649866,0.000754961,-5.66244e-07,2.13808e-08,0.65062,0.000753893,-5.02102e-07,-4.7299e-08,0.651374,0.000752746,-6.43999e-07,4.86059e-08,0.652126,0.000751604,-4.98181e-07,-2.79154e-08,0.652877,0.000750524,-5.81927e-07,3.45089e-09,0.653627,0.000749371,-5.71575e-07,1.41119e-08,0.654376,0.00074827,-5.29239e-07,-2.93748e-10,0.655123,0.00074721,-5.3012e-07,-1.29368e-08,0.65587,0.000746111,-5.68931e-07,-7.56355e-09,0.656616,0.000744951,-5.91621e-07,4.3191e-08,0.65736,0.000743897,-4.62048e-07,-4.59911e-08,0.658103,0.000742835,-6.00022e-07,2.15642e-08,0.658846,0.0007417,-5.35329e-07,1.93389e-08,0.659587,0.000740687,-4.77312e-07,-3.93152e-08,0.660327,0.000739615,-5.95258e-07,1.87126e-08,0.661066,0.00073848,-5.3912e-07,2.40695e-08,0.661804,0.000737474,-4.66912e-07,-5.53859e-08,0.662541,0.000736374,-6.33069e-07,7.82648e-08,0.663277,0.000735343,-3.98275e-07,-7.88593e-08,0.664012,0.00073431,-6.34853e-07,5.83585e-08,0.664745,0.000733215,-4.59777e-07,-3.53656e-08,0.665478,0.000732189,-5.65874e-07,2.34994e-08,0.66621,0.000731128,-4.95376e-07,9.72743e-10,0.66694,0.00073014,-4.92458e-07,-2.73903e-08,0.66767,0.000729073,-5.74629e-07,4.89839e-08,0.668398,0.000728071,-4.27677e-07,-4.93359e-08,0.669126,0.000727068,-5.75685e-07,2.91504e-08,0.669853,0.000726004,-4.88234e-07,-7.66109e-09,0.670578,0.000725004,-5.11217e-07,1.49392e-09,0.671303,0.000723986,-5.06735e-07,1.68533e-09,0.672026,0.000722978,-5.01679e-07,-8.23525e-09,0.672749,0.00072195,-5.26385e-07,3.12556e-08,0.67347,0.000720991,-4.32618e-07,-5.71825e-08,0.674191,0.000719954,-6.04166e-07,7.8265e-08,0.67491,0.00071898,-3.69371e-07,-7.70634e-08,0.675628,0.00071801,-6.00561e-07,5.11747e-08,0.676346,0.000716963,-4.47037e-07,-8.42615e-09,0.677062,0.000716044,-4.72315e-07,-1.747e-08,0.677778,0.000715046,-5.24725e-07,1.87015e-08,0.678493,0.000714053,-4.68621e-07,2.26856e-09,0.679206,0.000713123,-4.61815e-07,-2.77758e-08,0.679919,0.000712116,-5.45142e-07,4.92298e-08,0.68063,0.000711173,-3.97453e-07,-4.99339e-08,0.681341,0.000710228,-5.47255e-07,3.12967e-08,0.682051,0.000709228,-4.53365e-07,-1.56481e-08,0.68276,0.000708274,-5.00309e-07,3.12958e-08,0.683467,0.000707367,-4.06422e-07,-4.99303e-08,0.684174,0.000706405,-5.56213e-07,4.9216e-08,0.68488,0.00070544,-4.08565e-07,-2.77245e-08,0.685585,0.00070454,-4.91738e-07,2.07748e-09,0.686289,0.000703562,-4.85506e-07,1.94146e-08,0.686992,0.00070265,-4.27262e-07,-2.01314e-08,0.687695,0.000701735,-4.87656e-07,1.50616e-09,0.688396,0.000700764,-4.83137e-07,1.41067e-08,0.689096,0.00069984,-4.40817e-07,1.67168e-09,0.689795,0.000698963,-4.35802e-07,-2.07934e-08,0.690494,0.000698029,-4.98182e-07,2.18972e-08,0.691192,0.000697099,-4.32491e-07,-7.19092e-09,0.691888,0.000696212,-4.54064e-07,6.86642e-09,0.692584,0.000695325,-4.33464e-07,-2.02747e-08,0.693279,0.000694397,-4.94288e-07,1.46279e-08,0.693973,0.000693452,-4.50405e-07,2.13678e-08,0.694666,0.000692616,-3.86301e-07,-4.04945e-08,0.695358,0.000691721,-5.07785e-07,2.14009e-08,0.696049,0.00069077,-4.43582e-07,1.44955e-08,0.69674,0.000689926,-4.00096e-07,-1.97783e-08,0.697429,0.000689067,-4.5943e-07,5.01296e-09,0.698118,0.000688163,-4.44392e-07,-2.73521e-10,0.698805,0.000687273,-4.45212e-07,-3.91893e-09,0.699492,0.000686371,-4.56969e-07,1.59493e-08,0.700178,0.000685505,-4.09121e-07,-2.73351e-10,0.700863,0.000684686,-4.09941e-07,-1.4856e-08,0.701548,0.000683822,-4.54509e-07,9.25979e-11,0.702231,0.000682913,-4.54231e-07,1.44855e-08,0.702913,0.000682048,-4.10775e-07,1.56992e-09,0.703595,0.000681231,-4.06065e-07,-2.07652e-08,0.704276,0.000680357,-4.68361e-07,2.18864e-08,0.704956,0.000679486,-4.02701e-07,-7.17595e-09,0.705635,0.000678659,-4.24229e-07,6.81748e-09,0.706313,0.000677831,-4.03777e-07,-2.0094e-08,0.70699,0.000676963,-4.64059e-07,1.39538e-08,0.707667,0.000676077,-4.22197e-07,2.38835e-08,0.708343,0.000675304,-3.50547e-07,-4.98831e-08,0.709018,0.000674453,-5.00196e-07,5.64395e-08,0.709692,0.000673622,-3.30878e-07,-5.66657e-08,0.710365,0.00067279,-5.00875e-07,5.1014e-08,0.711037,0.000671942,-3.47833e-07,-2.81809e-08,0.711709,0.000671161,-4.32376e-07,2.10513e-09,0.712379,0.000670303,-4.2606e-07,1.97604e-08,0.713049,0.00066951,-3.66779e-07,-2.15422e-08,0.713718,0.000668712,-4.31406e-07,6.8038e-09,0.714387,0.000667869,-4.10994e-07,-5.67295e-09,0.715054,0.00066703,-4.28013e-07,1.5888e-08,0.715721,0.000666222,-3.80349e-07,1.72576e-09,0.716387,0.000665467,-3.75172e-07,-2.27911e-08,0.717052,0.000664648,-4.43545e-07,2.9834e-08,0.717716,0.00066385,-3.54043e-07,-3.69401e-08,0.718379,0.000663031,-4.64864e-07,5.83219e-08,0.719042,0.000662277,-2.89898e-07,-7.71382e-08,0.719704,0.000661465,-5.21313e-07,7.14171e-08,0.720365,0.000660637,-3.07061e-07,-2.97161e-08,0.721025,0.000659934,-3.96209e-07,-1.21575e-08,0.721685,0.000659105,-4.32682e-07,1.87412e-08,0.722343,0.000658296,-3.76458e-07,-3.2029e-09,0.723001,0.000657533,-3.86067e-07,-5.9296e-09,0.723659,0.000656743,-4.03856e-07,2.69213e-08,0.724315,0.000656016,-3.23092e-07,-4.21511e-08,0.724971,0.000655244,-4.49545e-07,2.24737e-08,0.725625,0.000654412,-3.82124e-07,1.18611e-08,0.726279,0.000653683,-3.46541e-07,-1.03132e-08,0.726933,0.000652959,-3.7748e-07,-3.02128e-08,0.727585,0.000652114,-4.68119e-07,7.15597e-08,0.728237,0.000651392,-2.5344e-07,-7.72119e-08,0.728888,0.000650654,-4.85075e-07,5.8474e-08,0.729538,0.000649859,-3.09654e-07,-3.74746e-08,0.730188,0.000649127,-4.22077e-07,3.18197e-08,0.730837,0.000648379,-3.26618e-07,-3.01997e-08,0.731485,0.000647635,-4.17217e-07,2.93747e-08,0.732132,0.000646888,-3.29093e-07,-2.76943e-08,0.732778,0.000646147,-4.12176e-07,2.17979e-08,0.733424,0.000645388,-3.46783e-07,1.07292e-10,0.734069,0.000644695,-3.46461e-07,-2.22271e-08,0.734713,0.000643935,-4.13142e-07,2.91963e-08,0.735357,0.000643197,-3.25553e-07,-3.49536e-08,0.736,0.000642441,-4.30414e-07,5.10133e-08,0.736642,0.000641733,-2.77374e-07,-4.98904e-08,0.737283,0.000641028,-4.27045e-07,2.93392e-08,0.737924,0.000640262,-3.39028e-07,-7.86156e-09,0.738564,0.000639561,-3.62612e-07,2.10703e-09,0.739203,0.000638842,-3.56291e-07,-5.6653e-10,0.739842,0.000638128,-3.57991e-07,1.59086e-10,0.740479,0.000637412,-3.57513e-07,-6.98321e-11,0.741116,0.000636697,-3.57723e-07,1.20214e-10,0.741753,0.000635982,-3.57362e-07,-4.10987e-10,0.742388,0.000635266,-3.58595e-07,1.5237e-09,0.743023,0.000634553,-3.54024e-07,-5.68376e-09,0.743657,0.000633828,-3.71075e-07,2.12113e-08,0.744291,0.00063315,-3.07441e-07,-1.95569e-08,0.744924,0.000632476,-3.66112e-07,-2.58816e-09,0.745556,0.000631736,-3.73877e-07,2.99096e-08,0.746187,0.000631078,-2.84148e-07,-5.74454e-08,0.746818,0.000630337,-4.56484e-07,8.06629e-08,0.747448,0.000629666,-2.14496e-07,-8.63922e-08,0.748077,0.000628978,-4.73672e-07,8.60918e-08,0.748706,0.000628289,-2.15397e-07,-7.91613e-08,0.749334,0.000627621,-4.5288e-07,5.17393e-08,0.749961,0.00062687,-2.97663e-07,-8.58662e-09,0.750588,0.000626249,-3.23422e-07,-1.73928e-08,0.751214,0.00062555,-3.75601e-07,1.85532e-08,0.751839,0.000624855,-3.19941e-07,2.78479e-09,0.752463,0.000624223,-3.11587e-07,-2.96923e-08,0.753087,0.000623511,-4.00664e-07,5.63799e-08,0.75371,0.000622879,-2.31524e-07,-7.66179e-08,0.754333,0.000622186,-4.61378e-07,7.12778e-08,0.754955,0.000621477,-2.47545e-07,-2.96794e-08,0.755576,0.000620893,-3.36583e-07,-1.21648e-08,0.756196,0.000620183,-3.73077e-07,1.87339e-08,0.756816,0.000619493,-3.16875e-07,-3.16622e-09,0.757435,0.00061885,-3.26374e-07,-6.0691e-09,0.758054,0.000618179,-3.44581e-07,2.74426e-08,0.758672,0.000617572,-2.62254e-07,-4.40968e-08,0.759289,0.000616915,-3.94544e-07,2.97352e-08,0.759906,0.000616215,-3.05338e-07,-1.52393e-08,0.760522,0.000615559,-3.51056e-07,3.12221e-08,0.761137,0.000614951,-2.5739e-07,-5.00443e-08,0.761751,0.000614286,-4.07523e-07,4.9746e-08,0.762365,0.00061362,-2.58285e-07,-2.97303e-08,0.762979,0.000613014,-3.47476e-07,9.57079e-09,0.763591,0.000612348,-3.18764e-07,-8.55287e-09,0.764203,0.000611685,-3.44422e-07,2.46407e-08,0.764815,0.00061107,-2.705e-07,-3.04053e-08,0.765426,0.000610437,-3.61716e-07,3.73759e-08,0.766036,0.000609826,-2.49589e-07,-5.94935e-08,0.766645,0.000609149,-4.28069e-07,8.13889e-08,0.767254,0.000608537,-1.83902e-07,-8.72483e-08,0.767862,0.000607907,-4.45647e-07,8.87901e-08,0.76847,0.000607282,-1.79277e-07,-8.90983e-08,0.769077,0.000606656,-4.46572e-07,8.87892e-08,0.769683,0.000606029,-1.80204e-07,-8.72446e-08,0.770289,0.000605407,-4.41938e-07,8.13752e-08,0.770894,0.000604768,-1.97812e-07,-5.94423e-08,0.771498,0.000604194,-3.76139e-07,3.71848e-08,0.772102,0.000603553,-2.64585e-07,-2.96922e-08,0.772705,0.000602935,-3.53661e-07,2.19793e-08,0.773308,0.000602293,-2.87723e-07,1.37955e-09,0.77391,0.000601722,-2.83585e-07,-2.74976e-08,0.774512,0.000601072,-3.66077e-07,4.9006e-08,0.775112,0.000600487,-2.19059e-07,-4.93171e-08,0.775712,0.000599901,-3.67011e-07,2.90531e-08,0.776312,0.000599254,-2.79851e-07,-7.29081e-09,0.776911,0.000598673,-3.01724e-07,1.10077e-10,0.777509,0.00059807,-3.01393e-07,6.85053e-09,0.778107,0.000597487,-2.80842e-07,-2.75123e-08,0.778704,0.000596843,-3.63379e-07,4.35939e-08,0.779301,0.000596247,-2.32597e-07,-2.7654e-08,0.779897,0.000595699,-3.15559e-07,7.41741e-09,0.780492,0.00059509,-2.93307e-07,-2.01562e-09,0.781087,0.000594497,-2.99354e-07,6.45059e-10,0.781681,0.000593901,-2.97418e-07,-5.64635e-10,0.782275,0.000593304,-2.99112e-07,1.61347e-09,0.782868,0.000592711,-2.94272e-07,-5.88926e-09,0.78346,0.000592105,-3.1194e-07,2.19436e-08,0.784052,0.000591546,-2.46109e-07,-2.22805e-08,0.784643,0.000590987,-3.1295e-07,7.57368e-09,0.785234,0.000590384,-2.90229e-07,-8.01428e-09,0.785824,0.00058978,-3.14272e-07,2.44834e-08,0.786414,0.000589225,-2.40822e-07,-3.03148e-08,0.787003,0.000588652,-3.31766e-07,3.7171e-08,0.787591,0.0005881,-2.20253e-07,-5.87646e-08,0.788179,0.000587483,-3.96547e-07,7.86782e-08,0.788766,0.000586926,-1.60512e-07,-7.71342e-08,0.789353,0.000586374,-3.91915e-07,5.10444e-08,0.789939,0.000585743,-2.38782e-07,-7.83422e-09,0.790524,0.000585242,-2.62284e-07,-1.97076e-08,0.791109,0.000584658,-3.21407e-07,2.70598e-08,0.791693,0.000584097,-2.40228e-07,-2.89269e-08,0.792277,0.000583529,-3.27008e-07,2.90431e-08,0.792861,0.000582963,-2.39879e-07,-2.76409e-08,0.793443,0.0005824,-3.22802e-07,2.1916e-08,0.794025,0.00058182,-2.57054e-07,-4.18368e-10,0.794607,0.000581305,-2.58309e-07,-2.02425e-08,0.795188,0.000580727,-3.19036e-07,2.17838e-08,0.795768,0.000580155,-2.53685e-07,-7.28814e-09,0.796348,0.000579625,-2.75549e-07,7.36871e-09,0.796928,0.000579096,-2.53443e-07,-2.21867e-08,0.797506,0.000578523,-3.20003e-07,2.17736e-08,0.798085,0.000577948,-2.54683e-07,-5.30296e-09,0.798662,0.000577423,-2.70592e-07,-5.61698e-10,0.799239,0.00057688,-2.72277e-07,7.54977e-09,0.799816,0.000576358,-2.49627e-07,-2.96374e-08,0.800392,0.00057577,-3.38539e-07,5.1395e-08,0.800968,0.000575247,-1.84354e-07,-5.67335e-08,0.801543,0.000574708,-3.54555e-07,5.63297e-08,0.802117,0.000574168,-1.85566e-07,-4.93759e-08,0.802691,0.000573649,-3.33693e-07,2.19646e-08,0.803264,0.000573047,-2.678e-07,2.1122e-08,0.803837,0.000572575,-2.04433e-07,-4.68482e-08,0.804409,0.000572026,-3.44978e-07,4.70613e-08,0.804981,0.000571477,-2.03794e-07,-2.21877e-08,0.805552,0.000571003,-2.70357e-07,-1.79153e-08,0.806123,0.000570408,-3.24103e-07,3.42443e-08,0.806693,0.000569863,-2.2137e-07,1.47556e-10,0.807263,0.000569421,-2.20928e-07,-3.48345e-08,0.807832,0.000568874,-3.25431e-07,1.99812e-08,0.808401,0.000568283,-2.65487e-07,1.45143e-08,0.808969,0.000567796,-2.21945e-07,-1.84338e-08,0.809536,0.000567297,-2.77246e-07,-3.83608e-10,0.810103,0.000566741,-2.78397e-07,1.99683e-08,0.81067,0.000566244,-2.18492e-07,-1.98848e-08,0.811236,0.000565747,-2.78146e-07,-3.38976e-11,0.811801,0.000565191,-2.78248e-07,2.00204e-08,0.812366,0.000564695,-2.18187e-07,-2.04429e-08,0.812931,0.000564197,-2.79516e-07,2.1467e-09,0.813495,0.000563644,-2.73076e-07,1.18561e-08,0.814058,0.000563134,-2.37507e-07,1.00334e-08,0.814621,0.000562689,-2.07407e-07,-5.19898e-08,0.815183,0.000562118,-3.63376e-07,7.87163e-08,0.815745,0.000561627,-1.27227e-07,-8.40616e-08,0.816306,0.000561121,-3.79412e-07,7.87163e-08,0.816867,0.000560598,-1.43263e-07,-5.19898e-08,0.817428,0.000560156,-2.99233e-07,1.00335e-08,0.817988,0.000559587,-2.69132e-07,1.18559e-08,0.818547,0.000559085,-2.33564e-07,2.14764e-09,0.819106,0.000558624,-2.27122e-07,-2.04464e-08,0.819664,0.000558108,-2.88461e-07,2.00334e-08,0.820222,0.000557591,-2.28361e-07,-8.24277e-11,0.820779,0.000557135,-2.28608e-07,-1.97037e-08,0.821336,0.000556618,-2.87719e-07,1.92925e-08,0.821893,0.000556101,-2.29841e-07,2.13831e-09,0.822448,0.000555647,-2.23427e-07,-2.78458e-08,0.823004,0.000555117,-3.06964e-07,4.96402e-08,0.823559,0.000554652,-1.58043e-07,-5.15058e-08,0.824113,0.000554181,-3.12561e-07,3.71737e-08,0.824667,0.000553668,-2.0104e-07,-3.75844e-08,0.82522,0.000553153,-3.13793e-07,5.35592e-08,0.825773,0.000552686,-1.53115e-07,-5.74431e-08,0.826326,0.000552207,-3.25444e-07,5.7004e-08,0.826878,0.000551728,-1.54433e-07,-5.13635e-08,0.827429,0.000551265,-3.08523e-07,2.92406e-08,0.82798,0.000550735,-2.20801e-07,-5.99424e-09,0.828531,0.000550276,-2.38784e-07,-5.26363e-09,0.829081,0.000549782,-2.54575e-07,2.70488e-08,0.82963,0.000549354,-1.73429e-07,-4.33268e-08,0.83018,0.000548878,-3.03409e-07,2.7049e-08,0.830728,0.000548352,-2.22262e-07,-5.26461e-09,0.831276,0.000547892,-2.38056e-07,-5.99057e-09,0.831824,0.000547397,-2.56027e-07,2.92269e-08,0.832371,0.000546973,-1.68347e-07,-5.13125e-08,0.832918,0.000546482,-3.22284e-07,5.68139e-08,0.833464,0.000546008,-1.51843e-07,-5.67336e-08,0.83401,0.000545534,-3.22043e-07,5.09113e-08,0.834555,0.000545043,-1.6931e-07,-2.77022e-08,0.8351,0.000544621,-2.52416e-07,2.92924e-10,0.835644,0.000544117,-2.51537e-07,2.65305e-08,0.836188,0.000543694,-1.71946e-07,-4.68105e-08,0.836732,0.00054321,-3.12377e-07,4.15021e-08,0.837275,0.000542709,-1.87871e-07,1.13355e-11,0.837817,0.000542334,-1.87837e-07,-4.15474e-08,0.838359,0.000541833,-3.12479e-07,4.69691e-08,0.838901,0.000541349,-1.71572e-07,-2.71196e-08,0.839442,0.000540925,-2.52931e-07,1.90462e-09,0.839983,0.000540425,-2.47217e-07,1.95011e-08,0.840523,0.000539989,-1.88713e-07,-2.03045e-08,0.841063,0.00053955,-2.49627e-07,2.11216e-09,0.841602,0.000539057,-2.4329e-07,1.18558e-08,0.842141,0.000538606,-2.07723e-07,1.00691e-08,0.842679,0.000538221,-1.77516e-07,-5.21324e-08,0.843217,0.00053771,-3.33913e-07,7.92513e-08,0.843755,0.00053728,-9.6159e-08,-8.60587e-08,0.844292,0.000536829,-3.54335e-07,8.61696e-08,0.844828,0.000536379,-9.58263e-08,-7.98057e-08,0.845364,0.000535948,-3.35243e-07,5.42394e-08,0.8459,0.00053544,-1.72525e-07,-1.79426e-08,0.846435,0.000535041,-2.26353e-07,1.75308e-08,0.84697,0.000534641,-1.73761e-07,-5.21806e-08,0.847505,0.000534137,-3.30302e-07,7.19824e-08,0.848038,0.000533692,-1.14355e-07,-5.69349e-08,0.848572,0.000533293,-2.8516e-07,3.65479e-08,0.849105,0.000532832,-1.75516e-07,-2.96519e-08,0.849638,0.000532392,-2.64472e-07,2.2455e-08,0.85017,0.000531931,-1.97107e-07,-5.63451e-10,0.850702,0.000531535,-1.98797e-07,-2.02011e-08,0.851233,0.000531077,-2.59401e-07,2.17634e-08,0.851764,0.000530623,-1.94111e-07,-7.24794e-09,0.852294,0.000530213,-2.15854e-07,7.22832e-09,0.852824,0.000529803,-1.94169e-07,-2.16653e-08,0.853354,0.00052935,-2.59165e-07,1.98283e-08,0.853883,0.000528891,-1.9968e-07,1.95678e-09,0.854412,0.000528497,-1.9381e-07,-2.76554e-08,0.85494,0.000528027,-2.76776e-07,4.90603e-08,0.855468,0.00052762,-1.29596e-07,-4.93764e-08,0.855995,0.000527213,-2.77725e-07,2.92361e-08,0.856522,0.000526745,-1.90016e-07,-7.96341e-09,0.857049,0.000526341,-2.13907e-07,2.61752e-09,0.857575,0.000525922,-2.06054e-07,-2.50665e-09,0.8581,0.000525502,-2.13574e-07,7.40906e-09,0.858626,0.000525097,-1.91347e-07,-2.71296e-08,0.859151,0.000524633,-2.72736e-07,4.15048e-08,0.859675,0.000524212,-1.48221e-07,-1.96802e-08,0.860199,0.000523856,-2.07262e-07,-2.23886e-08,0.860723,0.000523375,-2.74428e-07,4.96299e-08,0.861246,0.000522975,-1.25538e-07,-5.69216e-08,0.861769,0.000522553,-2.96303e-07,5.88473e-08,0.862291,0.000522137,-1.19761e-07,-5.92584e-08,0.862813,0.00052172,-2.97536e-07,5.8977e-08,0.863334,0.000521301,-1.20605e-07,-5.74403e-08,0.863855,0.000520888,-2.92926e-07,5.15751e-08,0.864376,0.000520457,-1.38201e-07,-2.96506e-08,0.864896,0.000520091,-2.27153e-07,7.42277e-09,0.865416,0.000519659,-2.04885e-07,-4.05057e-11,0.865936,0.00051925,-2.05006e-07,-7.26074e-09,0.866455,0.000518818,-2.26788e-07,2.90835e-08,0.866973,0.000518451,-1.39538e-07,-4.94686e-08,0.867492,0.000518024,-2.87944e-07,4.95814e-08,0.868009,0.000517597,-1.39199e-07,-2.96479e-08,0.868527,0.000517229,-2.28143e-07,9.40539e-09,0.869044,0.000516801,-1.99927e-07,-7.9737e-09,0.86956,0.000516378,-2.23848e-07,2.24894e-08,0.870077,0.000515997,-1.5638e-07,-2.23793e-08,0.870592,0.000515617,-2.23517e-07,7.42302e-09,0.871108,0.000515193,-2.01248e-07,-7.31283e-09,0.871623,0.000514768,-2.23187e-07,2.18283e-08,0.872137,0.000514387,-1.57702e-07,-2.03959e-08,0.872652,0.000514011,-2.1889e-07,1.50711e-10,0.873165,0.000513573,-2.18437e-07,1.97931e-08,0.873679,0.000513196,-1.59058e-07,-1.97183e-08,0.874192,0.000512819,-2.18213e-07,-5.24324e-10,0.874704,0.000512381,-2.19786e-07,2.18156e-08,0.875217,0.000512007,-1.54339e-07,-2.71336e-08,0.875728,0.000511616,-2.3574e-07,2.71141e-08,0.87624,0.000511226,-1.54398e-07,-2.17182e-08,0.876751,0.000510852,-2.19552e-07,1.54131e-10,0.877262,0.000510414,-2.1909e-07,2.11017e-08,0.877772,0.000510039,-1.55785e-07,-2.49562e-08,0.878282,0.000509652,-2.30654e-07,1.91183e-08,0.878791,0.000509248,-1.73299e-07,8.08751e-09,0.8793,0.000508926,-1.49036e-07,-5.14684e-08,0.879809,0.000508474,-3.03441e-07,7.85766e-08,0.880317,0.000508103,-6.77112e-08,-8.40242e-08,0.880825,0.000507715,-3.19784e-07,7.87063e-08,0.881333,0.000507312,-8.36649e-08,-5.19871e-08,0.88184,0.000506988,-2.39626e-07,1.00327e-08,0.882346,0.000506539,-2.09528e-07,1.18562e-08,0.882853,0.000506156,-1.73959e-07,2.14703e-09,0.883359,0.000505814,-1.67518e-07,-2.04444e-08,0.883864,0.000505418,-2.28851e-07,2.00258e-08,0.88437,0.00050502,-1.68774e-07,-5.42855e-11,0.884874,0.000504682,-1.68937e-07,-1.98087e-08,0.885379,0.000504285,-2.28363e-07,1.96842e-08,0.885883,0.000503887,-1.6931e-07,6.76342e-10,0.886387,0.000503551,-1.67281e-07,-2.23896e-08,0.88689,0.000503149,-2.3445e-07,2.92774e-08,0.887393,0.000502768,-1.46618e-07,-3.51152e-08,0.887896,0.00050237,-2.51963e-07,5.15787e-08,0.888398,0.00050202,-9.72271e-08,-5.19903e-08,0.8889,0.00050167,-2.53198e-07,3.71732e-08,0.889401,0.000501275,-1.41678e-07,-3.70978e-08,0.889902,0.00050088,-2.52972e-07,5.16132e-08,0.890403,0.000500529,-9.81321e-08,-5.01459e-08,0.890903,0.000500183,-2.4857e-07,2.9761e-08,0.891403,0.000499775,-1.59287e-07,-9.29351e-09,0.891903,0.000499428,-1.87167e-07,7.41301e-09,0.892402,0.000499076,-1.64928e-07,-2.03585e-08,0.892901,0.000498685,-2.26004e-07,1.44165e-08,0.893399,0.000498276,-1.82754e-07,2.22974e-08,0.893898,0.000497978,-1.15862e-07,-4.40013e-08,0.894395,0.000497614,-2.47866e-07,3.44985e-08,0.894893,0.000497222,-1.44371e-07,-3.43882e-08,0.89539,0.00049683,-2.47535e-07,4.34497e-08,0.895886,0.000496465,-1.17186e-07,-2.02012e-08,0.896383,0.00049617,-1.7779e-07,-2.22497e-08,0.896879,0.000495748,-2.44539e-07,4.95952e-08,0.897374,0.000495408,-9.57532e-08,-5.69217e-08,0.89787,0.000495045,-2.66518e-07,5.88823e-08,0.898364,0.000494689,-8.98713e-08,-5.93983e-08,0.898859,0.000494331,-2.68066e-07,5.95017e-08,0.899353,0.000493973,-8.95613e-08,-5.9399e-08,0.899847,0.000493616,-2.67758e-07,5.8885e-08,0.90034,0.000493257,-9.11033e-08,-5.69317e-08,0.900833,0.000492904,-2.61898e-07,4.96326e-08,0.901326,0.000492529,-1.13001e-07,-2.23893e-08,0.901819,0.000492236,-1.80169e-07,-1.968e-08,0.902311,0.000491817,-2.39209e-07,4.15047e-08,0.902802,0.000491463,-1.14694e-07,-2.71296e-08,0.903293,0.000491152,-1.96083e-07,7.409e-09,0.903784,0.000490782,-1.73856e-07,-2.50645e-09,0.904275,0.000490427,-1.81376e-07,2.61679e-09,0.904765,0.000490072,-1.73525e-07,-7.96072e-09,0.905255,0.000489701,-1.97407e-07,2.92261e-08,0.905745,0.000489394,-1.09729e-07,-4.93389e-08,0.906234,0.000489027,-2.57746e-07,4.89204e-08,0.906723,0.000488658,-1.10985e-07,-2.71333e-08,0.907211,0.000488354,-1.92385e-07,8.30861e-12,0.907699,0.00048797,-1.9236e-07,2.71001e-08,0.908187,0.000487666,-1.1106e-07,-4.88041e-08,0.908675,0.000487298,-2.57472e-07,4.89069e-08,0.909162,0.000486929,-1.10751e-07,-2.76143e-08,0.909649,0.000486625,-1.93594e-07,1.9457e-09,0.910135,0.000486244,-1.87757e-07,1.98315e-08,0.910621,0.000485928,-1.28262e-07,-2.16671e-08,0.911107,0.000485606,-1.93264e-07,7.23216e-09,0.911592,0.000485241,-1.71567e-07,-7.26152e-09,0.912077,0.000484877,-1.93352e-07,2.18139e-08,0.912562,0.000484555,-1.2791e-07,-2.03895e-08,0.913047,0.000484238,-1.89078e-07,1.39494e-10,0.913531,0.000483861,-1.8866e-07,1.98315e-08,0.914014,0.000483543,-1.29165e-07,-1.98609e-08,0.914498,0.000483225,-1.88748e-07,7.39912e-12,0.914981,0.000482847,-1.88726e-07,1.98313e-08,0.915463,0.000482529,-1.29232e-07,-1.9728e-08,0.915946,0.000482212,-1.88416e-07,-5.24035e-10,0.916428,0.000481833,-1.89988e-07,2.18241e-08,0.916909,0.000481519,-1.24516e-07,-2.71679e-08,0.917391,0.000481188,-2.06019e-07,2.72427e-08,0.917872,0.000480858,-1.24291e-07,-2.21985e-08,0.918353,0.000480543,-1.90886e-07,1.94644e-09,0.918833,0.000480167,-1.85047e-07,1.44127e-08,0.919313,0.00047984,-1.41809e-07,7.39438e-12,0.919793,0.000479556,-1.41787e-07,-1.44423e-08,0.920272,0.000479229,-1.85114e-07,-1.84291e-09,0.920751,0.000478854,-1.90642e-07,2.18139e-08,0.92123,0.000478538,-1.25201e-07,-2.58081e-08,0.921708,0.00047821,-2.02625e-07,2.18139e-08,0.922186,0.00047787,-1.37183e-07,-1.84291e-09,0.922664,0.00047759,-1.42712e-07,-1.44423e-08,0.923141,0.000477262,-1.86039e-07,7.34701e-12,0.923618,0.00047689,-1.86017e-07,1.44129e-08,0.924095,0.000476561,-1.42778e-07,1.94572e-09,0.924572,0.000476281,-1.36941e-07,-2.21958e-08,0.925048,0.000475941,-2.03528e-07,2.72327e-08,0.925523,0.000475615,-1.2183e-07,-2.71304e-08,0.925999,0.00047529,-2.03221e-07,2.16843e-08,0.926474,0.000474949,-1.38168e-07,-2.16005e-12,0.926949,0.000474672,-1.38175e-07,-2.16756e-08,0.927423,0.000474331,-2.03202e-07,2.71001e-08,0.927897,0.000474006,-1.21902e-07,-2.71201e-08,0.928371,0.000473681,-2.03262e-07,2.17757e-08,0.928845,0.00047334,-1.37935e-07,-3.78028e-10,0.929318,0.000473063,-1.39069e-07,-2.02636e-08,0.929791,0.000472724,-1.9986e-07,2.18276e-08,0.930263,0.000472389,-1.34377e-07,-7.44231e-09,0.930736,0.000472098,-1.56704e-07,7.94165e-09,0.931208,0.000471809,-1.32879e-07,-2.43243e-08,0.931679,0.00047147,-2.05851e-07,2.97508e-08,0.932151,0.000471148,-1.16599e-07,-3.50742e-08,0.932622,0.000470809,-2.21822e-07,5.09414e-08,0.933092,0.000470518,-6.89976e-08,-4.94821e-08,0.933563,0.000470232,-2.17444e-07,2.77775e-08,0.934033,0.00046988,-1.34111e-07,-2.02351e-09,0.934502,0.000469606,-1.40182e-07,-1.96835e-08,0.934972,0.000469267,-1.99232e-07,2.11529e-08,0.935441,0.000468932,-1.35774e-07,-5.32332e-09,0.93591,0.000468644,-1.51743e-07,1.40413e-10,0.936378,0.000468341,-1.51322e-07,4.76166e-09,0.936846,0.000468053,-1.37037e-07,-1.9187e-08,0.937314,0.000467721,-1.94598e-07,1.23819e-08,0.937782,0.000467369,-1.57453e-07,2.92642e-08,0.938249,0.000467142,-6.96601e-08,-6.98342e-08,0.938716,0.000466793,-2.79163e-07,7.12586e-08,0.939183,0.000466449,-6.53869e-08,-3.63863e-08,0.939649,0.000466209,-1.74546e-07,1.46818e-08,0.940115,0.000465904,-1.305e-07,-2.2341e-08,0.940581,0.000465576,-1.97523e-07,1.50774e-08,0.941046,0.000465226,-1.52291e-07,2.16359e-08,0.941511,0.000464986,-8.73832e-08,-4.20162e-08,0.941976,0.000464685,-2.13432e-07,2.72198e-08,0.942441,0.00046434,-1.31773e-07,-7.2581e-09,0.942905,0.000464055,-1.53547e-07,1.81263e-09,0.943369,0.000463753,-1.48109e-07,7.58386e-12,0.943832,0.000463457,-1.48086e-07,-1.84298e-09,0.944296,0.000463155,-1.53615e-07,7.36433e-09,0.944759,0.00046287,-1.31522e-07,-2.76143e-08,0.945221,0.000462524,-2.14365e-07,4.34883e-08,0.945684,0.000462226,-8.39003e-08,-2.71297e-08,0.946146,0.000461977,-1.65289e-07,5.42595e-09,0.946608,0.000461662,-1.49012e-07,5.42593e-09,0.947069,0.000461381,-1.32734e-07,-2.71297e-08,0.94753,0.000461034,-2.14123e-07,4.34881e-08,0.947991,0.000460736,-8.36585e-08,-2.76134e-08,0.948452,0.000460486,-1.66499e-07,7.36083e-09,0.948912,0.000460175,-1.44416e-07,-1.82993e-09,0.949372,0.000459881,-1.49906e-07,-4.11073e-11,0.949832,0.000459581,-1.50029e-07,1.99434e-09,0.950291,0.000459287,-1.44046e-07,-7.93627e-09,0.950751,0.000458975,-1.67855e-07,2.97507e-08,0.951209,0.000458728,-7.86029e-08,-5.1462e-08,0.951668,0.000458417,-2.32989e-07,5.6888e-08,0.952126,0.000458121,-6.2325e-08,-5.68806e-08,0.952584,0.000457826,-2.32967e-07,5.14251e-08,0.953042,0.000457514,-7.86914e-08,-2.96107e-08,0.953499,0.000457268,-1.67523e-07,7.41296e-09,0.953956,0.000456955,-1.45285e-07,-4.11262e-11,0.954413,0.000456665,-1.45408e-07,-7.24847e-09,0.95487,0.000456352,-1.67153e-07,2.9035e-08,0.955326,0.000456105,-8.00484e-08,-4.92869e-08,0.955782,0.000455797,-2.27909e-07,4.89032e-08,0.956238,0.000455488,-8.11994e-08,-2.71166e-08,0.956693,0.000455244,-1.62549e-07,-4.13678e-11,0.957148,0.000454919,-1.62673e-07,2.72821e-08,0.957603,0.000454675,-8.0827e-08,-4.94824e-08,0.958057,0.000454365,-2.29274e-07,5.14382e-08,0.958512,0.000454061,-7.49597e-08,-3.7061e-08,0.958965,0.0004538,-1.86143e-07,3.72013e-08,0.959419,0.000453539,-7.45389e-08,-5.21396e-08,0.959873,0.000453234,-2.30958e-07,5.21476e-08,0.960326,0.000452928,-7.45146e-08,-3.72416e-08,0.960778,0.000452667,-1.8624e-07,3.72143e-08,0.961231,0.000452407,-7.45967e-08,-5.20109e-08,0.961683,0.000452101,-2.30629e-07,5.16199e-08,0.962135,0.000451795,-7.57696e-08,-3.52595e-08,0.962587,0.000451538,-1.81548e-07,2.98133e-08,0.963038,0.000451264,-9.2108e-08,-2.43892e-08,0.963489,0.000451007,-1.65276e-07,8.13892e-09,0.96394,0.000450701,-1.40859e-07,-8.16647e-09,0.964391,0.000450394,-1.65358e-07,2.45269e-08,0.964841,0.000450137,-9.17775e-08,-3.03367e-08,0.965291,0.000449863,-1.82787e-07,3.7215e-08,0.965741,0.000449609,-7.11424e-08,-5.89188e-08,0.96619,0.00044929,-2.47899e-07,7.92509e-08,0.966639,0.000449032,-1.01462e-08,-7.92707e-08,0.967088,0.000448773,-2.47958e-07,5.90181e-08,0.967537,0.000448455,-7.0904e-08,-3.75925e-08,0.967985,0.0004482,-1.83681e-07,3.17471e-08,0.968433,0.000447928,-8.84401e-08,-2.97913e-08,0.968881,0.000447662,-1.77814e-07,2.78133e-08,0.969329,0.000447389,-9.4374e-08,-2.18572e-08,0.969776,0.000447135,-1.59946e-07,1.10134e-11,0.970223,0.000446815,-1.59913e-07,2.18132e-08,0.97067,0.000446561,-9.44732e-08,-2.76591e-08,0.971116,0.000446289,-1.7745e-07,2.92185e-08,0.971562,0.000446022,-8.97948e-08,-2.96104e-08,0.972008,0.000445753,-1.78626e-07,2.96185e-08,0.972454,0.000445485,-8.97706e-08,-2.92588e-08,0.972899,0.000445218,-1.77547e-07,2.78123e-08,0.973344,0.000444946,-9.41103e-08,-2.23856e-08,0.973789,0.000444691,-1.61267e-07,2.12559e-09,0.974233,0.000444374,-1.5489e-07,1.38833e-08,0.974678,0.000444106,-1.13241e-07,1.94591e-09,0.975122,0.000443886,-1.07403e-07,-2.16669e-08,0.975565,0.000443606,-1.72404e-07,2.5117e-08,0.976009,0.000443336,-9.70526e-08,-1.91963e-08,0.976452,0.000443085,-1.54642e-07,-7.93627e-09,0.976895,0.000442752,-1.7845e-07,5.09414e-08,0.977338,0.000442548,-2.56262e-08,-7.66201e-08,0.97778,0.000442266,-2.55486e-07,7.67249e-08,0.978222,0.000441986,-2.53118e-08,-5.14655e-08,0.978664,0.000441781,-1.79708e-07,9.92773e-09,0.979106,0.000441451,-1.49925e-07,1.17546e-08,0.979547,0.000441186,-1.14661e-07,2.65868e-09,0.979988,0.000440965,-1.06685e-07,-2.23893e-08,0.980429,0.000440684,-1.73853e-07,2.72939e-08,0.980869,0.000440419,-9.19716e-08,-2.71816e-08,0.98131,0.000440153,-1.73516e-07,2.18278e-08,0.98175,0.000439872,-1.08033e-07,-5.24833e-10,0.982189,0.000439654,-1.09607e-07,-1.97284e-08,0.982629,0.000439376,-1.68793e-07,1.98339e-08,0.983068,0.000439097,-1.09291e-07,-2.62901e-12,0.983507,0.000438879,-1.09299e-07,-1.98234e-08,0.983946,0.000438601,-1.68769e-07,1.96916e-08,0.984384,0.000438322,-1.09694e-07,6.6157e-10,0.984823,0.000438105,-1.0771e-07,-2.23379e-08,0.985261,0.000437823,-1.74723e-07,2.90855e-08,0.985698,0.00043756,-8.74669e-08,-3.43992e-08,0.986136,0.000437282,-1.90665e-07,4.89068e-08,0.986573,0.000437048,-4.39442e-08,-4.20188e-08,0.98701,0.000436834,-1.7e-07,-4.11073e-11,0.987446,0.000436494,-1.70124e-07,4.21832e-08,0.987883,0.00043628,-4.35742e-08,-4.94824e-08,0.988319,0.000436044,-1.92021e-07,3.6537e-08,0.988755,0.00043577,-8.24102e-08,-3.70611e-08,0.989191,0.000435494,-1.93593e-07,5.21026e-08,0.989626,0.000435263,-3.72855e-08,-5.21402e-08,0.990061,0.000435032,-1.93706e-07,3.7249e-08,0.990496,0.000434756,-8.19592e-08,-3.72512e-08,0.990931,0.000434481,-1.93713e-07,5.21511e-08,0.991365,0.00043425,-3.72595e-08,-5.21439e-08,0.991799,0.000434019,-1.93691e-07,3.72152e-08,0.992233,0.000433743,-8.20456e-08,-3.71123e-08,0.992667,0.000433468,-1.93382e-07,5.16292e-08,0.9931,0.000433236,-3.84947e-08,-5.01953e-08,0.993533,0.000433008,-1.89081e-07,2.99427e-08,0.993966,0.00043272,-9.92525e-08,-9.9708e-09,0.994399,0.000432491,-1.29165e-07,9.94051e-09,0.994831,0.000432263,-9.93434e-08,-2.97912e-08,0.995263,0.000431975,-1.88717e-07,4.96198e-08,0.995695,0.000431746,-3.98578e-08,-4.94785e-08,0.996127,0.000431518,-1.88293e-07,2.9085e-08,0.996558,0.000431229,-1.01038e-07,-7.25675e-09,0.996989,0.000431005,-1.22809e-07,-5.79945e-11,0.99742,0.000430759,-1.22983e-07,7.48873e-09,0.997851,0.000430536,-1.00516e-07,-2.98969e-08,0.998281,0.000430245,-1.90207e-07,5.24942e-08,0.998711,0.000430022,-3.27246e-08,-6.08706e-08,0.999141,0.000429774,-2.15336e-07,7.17788e-08,0.999571,0.000429392,0.,0.};
-
-        template <bool srgb, int blueIdx, typename T, typename D>
-        __device__ __forceinline__ void Lab2RGBConvert_f(const T& src, D& dst)
-        {
-            const float lThresh = 0.008856f * 903.3f;
-            const float fThresh = 7.787f * 0.008856f + 16.0f / 116.0f;
-
-            float Y, fy;
-
-            if (src.x <= lThresh)
-            {
-                Y = src.x / 903.3f;
-                fy = 7.787f * Y + 16.0f / 116.0f;
-            }
-            else
-            {
-                fy = (src.x + 16.0f) / 116.0f;
-                Y = fy * fy * fy;
-            }
-
-            float X = src.y / 500.0f + fy;
-            float Z = fy - src.z / 200.0f;
-
-            if (X <= fThresh)
-                X = (X - 16.0f / 116.0f) / 7.787f;
-            else
-                X = X * X * X;
-
-            if (Z <= fThresh)
-                Z = (Z - 16.0f / 116.0f) / 7.787f;
-            else
-                Z = Z * Z * Z;
-
-            float B = 0.052891f * X - 0.204043f * Y + 1.151152f * Z;
-            float G = -0.921235f * X + 1.875991f * Y + 0.045244f * Z;
-            float R = 3.079933f * X - 1.537150f * Y - 0.542782f * Z;
-
-            if (srgb)
-            {
-                B = splineInterpolate(B * GAMMA_TAB_SIZE, c_sRGBInvGammaTab, GAMMA_TAB_SIZE);
-                G = splineInterpolate(G * GAMMA_TAB_SIZE, c_sRGBInvGammaTab, GAMMA_TAB_SIZE);
-                R = splineInterpolate(R * GAMMA_TAB_SIZE, c_sRGBInvGammaTab, GAMMA_TAB_SIZE);
-            }
-
-            dst.x = blueIdx == 0 ? B : R;
-            dst.y = G;
-            dst.z = blueIdx == 0 ? R : B;
-            setAlpha(dst, ColorChannel<float>::max());
-        }
-
-        template <bool srgb, int blueIdx, typename T, typename D>
-        __device__ __forceinline__ void Lab2RGBConvert_b(const T& src, D& dst)
-        {
-            float3 srcf, dstf;
-
-            srcf.x = src.x * (100.f / 255.f);
-            srcf.y = src.y - 128;
-            srcf.z = src.z - 128;
-
-            Lab2RGBConvert_f<srgb, blueIdx>(srcf, dstf);
-
-            dst.x = saturate_cast<uchar>(dstf.x * 255.f);
-            dst.y = saturate_cast<uchar>(dstf.y * 255.f);
-            dst.z = saturate_cast<uchar>(dstf.z * 255.f);
-            setAlpha(dst, ColorChannel<uchar>::max());
-        }
-
-        template <typename T, int scn, int dcn, bool srgb, int blueIdx> struct Lab2RGB;
-        template <int scn, int dcn, bool srgb, int blueIdx>
-        struct Lab2RGB<uchar, scn, dcn, srgb, blueIdx>
-            : unary_function<typename TypeVec<uchar, scn>::vec_type, typename TypeVec<uchar, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<uchar, dcn>::vec_type operator ()(const typename TypeVec<uchar, scn>::vec_type& src) const
-            {
-                typename TypeVec<uchar, dcn>::vec_type dst;
-
-                Lab2RGBConvert_b<srgb, blueIdx>(src, dst);
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ Lab2RGB() {}
-            __host__ __device__ __forceinline__ Lab2RGB(const Lab2RGB&) {}
-        };
-        template <int scn, int dcn, bool srgb, int blueIdx>
-        struct Lab2RGB<float, scn, dcn, srgb, blueIdx>
-            : unary_function<typename TypeVec<float, scn>::vec_type, typename TypeVec<float, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<float, dcn>::vec_type operator ()(const typename TypeVec<float, scn>::vec_type& src) const
-            {
-                typename TypeVec<float, dcn>::vec_type dst;
-
-                Lab2RGBConvert_f<srgb, blueIdx>(src, dst);
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ Lab2RGB() {}
-            __host__ __device__ __forceinline__ Lab2RGB(const Lab2RGB&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_Lab2RGB_TRAITS(name, scn, dcn, srgb, blueIdx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::Lab2RGB<T, scn, dcn, srgb, blueIdx> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-///////////////////////////////////// RGB <-> Luv /////////////////////////////////////
-
-    namespace color_detail
-    {
-        __constant__ float c_LabCbrtTab[] = {0.137931,0.0114066,0.,1.18859e-07,0.149338,0.011407,3.56578e-07,-5.79396e-07,0.160745,0.0114059,-1.38161e-06,2.16892e-06,0.172151,0.0114097,5.12516e-06,-8.0814e-06,0.183558,0.0113957,-1.9119e-05,3.01567e-05,0.194965,0.0114479,7.13509e-05,-0.000112545,0.206371,0.011253,-0.000266285,-0.000106493,0.217252,0.0104009,-0.000585765,7.32149e-05,0.22714,0.00944906,-0.00036612,1.21917e-05,0.236235,0.0087534,-0.000329545,2.01753e-05,0.244679,0.00815483,-0.000269019,1.24435e-05,0.252577,0.00765412,-0.000231689,1.05618e-05,0.26001,0.00722243,-0.000200003,8.26662e-06,0.267041,0.00684723,-0.000175203,6.76746e-06,0.27372,0.00651712,-0.000154901,5.61192e-06,0.280088,0.00622416,-0.000138065,4.67009e-06,0.286179,0.00596204,-0.000124055,3.99012e-06,0.292021,0.0057259,-0.000112085,3.36032e-06,0.297638,0.00551181,-0.000102004,2.95338e-06,0.30305,0.00531666,-9.31435e-05,2.52875e-06,0.308277,0.00513796,-8.55572e-05,2.22022e-06,0.313331,0.00497351,-7.88966e-05,1.97163e-06,0.318228,0.00482163,-7.29817e-05,1.7248e-06,0.322978,0.00468084,-6.78073e-05,1.55998e-06,0.327593,0.0045499,-6.31274e-05,1.36343e-06,0.332081,0.00442774,-5.90371e-05,1.27136e-06,0.336451,0.00431348,-5.5223e-05,1.09111e-06,0.34071,0.00420631,-5.19496e-05,1.0399e-06,0.344866,0.00410553,-4.88299e-05,9.18347e-07,0.348923,0.00401062,-4.60749e-05,8.29942e-07,0.352889,0.00392096,-4.35851e-05,7.98478e-07,0.356767,0.00383619,-4.11896e-05,6.84917e-07,0.360562,0.00375586,-3.91349e-05,6.63976e-07,0.36428,0.00367959,-3.7143e-05,5.93086e-07,0.367923,0.00360708,-3.53637e-05,5.6976e-07,0.371495,0.00353806,-3.36544e-05,4.95533e-07,0.375,0.00347224,-3.21678e-05,4.87951e-07,0.378441,0.00340937,-3.0704e-05,4.4349e-07,0.38182,0.00334929,-2.93735e-05,4.20297e-07,0.38514,0.0032918,-2.81126e-05,3.7872e-07,0.388404,0.00323671,-2.69764e-05,3.596e-07,0.391614,0.00318384,-2.58976e-05,3.5845e-07,0.394772,0.00313312,-2.48223e-05,2.92765e-07,0.397881,0.00308435,-2.3944e-05,3.18232e-07,0.400942,0.00303742,-2.29893e-05,2.82046e-07,0.403957,0.00299229,-2.21432e-05,2.52315e-07,0.406927,0.00294876,-2.13862e-05,2.58416e-07,0.409855,0.00290676,-2.0611e-05,2.33939e-07,0.412741,0.00286624,-1.99092e-05,2.36342e-07,0.415587,0.00282713,-1.92001e-05,1.916e-07,0.418396,0.00278931,-1.86253e-05,2.1915e-07,0.421167,0.00275271,-1.79679e-05,1.83498e-07,0.423901,0.00271733,-1.74174e-05,1.79343e-07,0.426602,0.00268303,-1.68794e-05,1.72013e-07,0.429268,0.00264979,-1.63633e-05,1.75686e-07,0.431901,0.00261759,-1.58363e-05,1.3852e-07,0.434503,0.00258633,-1.54207e-05,1.64304e-07,0.437074,0.00255598,-1.49278e-05,1.28136e-07,0.439616,0.00252651,-1.45434e-05,1.57618e-07,0.442128,0.0024979,-1.40705e-05,1.0566e-07,0.444612,0.00247007,-1.37535e-05,1.34998e-07,0.447068,0.00244297,-1.33485e-05,1.29207e-07,0.449498,0.00241666,-1.29609e-05,9.32347e-08,0.451902,0.00239102,-1.26812e-05,1.23703e-07,0.45428,0.00236603,-1.23101e-05,9.74072e-08,0.456634,0.0023417,-1.20179e-05,1.12518e-07,0.458964,0.002318,-1.16803e-05,7.83681e-08,0.46127,0.00229488,-1.14452e-05,1.10452e-07,0.463554,0.00227232,-1.11139e-05,7.58719e-08,0.465815,0.00225032,-1.08863e-05,9.2699e-08,0.468055,0.00222882,-1.06082e-05,8.97738e-08,0.470273,0.00220788,-1.03388e-05,5.4845e-08,0.47247,0.00218736,-1.01743e-05,1.0808e-07,0.474648,0.00216734,-9.85007e-06,4.9277e-08,0.476805,0.00214779,-9.70224e-06,8.22408e-08,0.478943,0.00212863,-9.45551e-06,6.87942e-08,0.481063,0.00210993,-9.24913e-06,5.98144e-08,0.483163,0.00209161,-9.06969e-06,7.93789e-08,0.485246,0.00207371,-8.83155e-06,3.99032e-08,0.487311,0.00205616,-8.71184e-06,8.88325e-08,0.489358,0.002039,-8.44534e-06,2.20004e-08,0.491389,0.00202218,-8.37934e-06,9.13872e-08,0.493403,0.0020057,-8.10518e-06,2.96829e-08,0.495401,0.00198957,-8.01613e-06,5.81028e-08,0.497382,0.00197372,-7.84183e-06,6.5731e-08,0.499348,0.00195823,-7.64463e-06,3.66019e-08,0.501299,0.00194305,-7.53483e-06,2.62811e-08,0.503234,0.00192806,-7.45598e-06,9.66907e-08,0.505155,0.00191344,-7.16591e-06,4.18928e-09,0.507061,0.00189912,-7.15334e-06,6.53665e-08,0.508953,0.00188501,-6.95724e-06,3.23686e-08,0.510831,0.00187119,-6.86014e-06,4.35774e-08,0.512696,0.0018576,-6.72941e-06,3.17406e-08,0.514547,0.00184424,-6.63418e-06,6.78785e-08,0.516384,0.00183117,-6.43055e-06,-5.23126e-09,0.518209,0.0018183,-6.44624e-06,7.22562e-08,0.520021,0.00180562,-6.22947e-06,1.42292e-08,0.52182,0.0017932,-6.18679e-06,4.9641e-08,0.523607,0.00178098,-6.03786e-06,2.56259e-08,0.525382,0.00176898,-5.96099e-06,2.66696e-08,0.527145,0.00175714,-5.88098e-06,4.65094e-08,0.528897,0.00174552,-5.74145e-06,2.57114e-08,0.530637,0.00173411,-5.66431e-06,2.94588e-08,0.532365,0.00172287,-5.57594e-06,3.52667e-08,0.534082,0.00171182,-5.47014e-06,8.28868e-09,0.535789,0.00170091,-5.44527e-06,5.07871e-08,0.537484,0.00169017,-5.29291e-06,2.69817e-08,0.539169,0.00167967,-5.21197e-06,2.01009e-08,0.540844,0.0016693,-5.15166e-06,1.18237e-08,0.542508,0.00165903,-5.11619e-06,5.18135e-08,0.544162,0.00164896,-4.96075e-06,1.9341e-08,0.545806,0.00163909,-4.90273e-06,-9.96867e-09,0.54744,0.00162926,-4.93263e-06,8.01382e-08,0.549064,0.00161963,-4.69222e-06,-1.25601e-08,0.550679,0.00161021,-4.7299e-06,2.97067e-08,0.552285,0.00160084,-4.64078e-06,1.29426e-08,0.553881,0.0015916,-4.60195e-06,3.77327e-08,0.555468,0.00158251,-4.48875e-06,1.49412e-08,0.557046,0.00157357,-4.44393e-06,2.17118e-08,0.558615,0.00156475,-4.3788e-06,1.74206e-08,0.560176,0.00155605,-4.32653e-06,2.78152e-08,0.561727,0.00154748,-4.24309e-06,-9.47239e-09,0.563271,0.00153896,-4.27151e-06,6.9679e-08,0.564805,0.00153063,-4.06247e-06,-3.08246e-08,0.566332,0.00152241,-4.15494e-06,5.36188e-08,0.56785,0.00151426,-3.99409e-06,-4.83594e-09,0.56936,0.00150626,-4.00859e-06,2.53293e-08,0.570863,0.00149832,-3.93261e-06,2.27286e-08,0.572357,0.00149052,-3.86442e-06,2.96541e-09,0.573844,0.0014828,-3.85552e-06,2.50147e-08,0.575323,0.00147516,-3.78048e-06,1.61842e-08,0.576794,0.00146765,-3.73193e-06,2.94582e-08,0.578258,0.00146028,-3.64355e-06,-1.48076e-08,0.579715,0.00145295,-3.68798e-06,2.97724e-08,0.581164,0.00144566,-3.59866e-06,1.49272e-08,0.582606,0.00143851,-3.55388e-06,2.97285e-08,0.584041,0.00143149,-3.46469e-06,-1.46323e-08,0.585469,0.00142451,-3.50859e-06,2.88004e-08,0.58689,0.00141758,-3.42219e-06,1.864e-08,0.588304,0.00141079,-3.36627e-06,1.58482e-08,0.589712,0.00140411,-3.31872e-06,-2.24279e-08,0.591112,0.00139741,-3.38601e-06,7.38639e-08,0.592507,0.00139085,-3.16441e-06,-3.46088e-08,0.593894,0.00138442,-3.26824e-06,4.96675e-09,0.595275,0.0013779,-3.25334e-06,7.4346e-08,0.59665,0.00137162,-3.0303e-06,-6.39319e-08,0.598019,0.00136536,-3.2221e-06,6.21725e-08,0.599381,0.00135911,-3.03558e-06,-5.94423e-09,0.600737,0.00135302,-3.05341e-06,2.12091e-08,0.602087,0.00134697,-2.98979e-06,-1.92876e-08,0.603431,0.00134094,-3.04765e-06,5.5941e-08,0.604769,0.00133501,-2.87983e-06,-2.56622e-08,0.606101,0.00132917,-2.95681e-06,4.67078e-08,0.607427,0.0013234,-2.81669e-06,-4.19592e-08,0.608748,0.00131764,-2.94257e-06,6.15243e-08,0.610062,0.00131194,-2.75799e-06,-2.53244e-08,0.611372,0.00130635,-2.83397e-06,3.97739e-08,0.612675,0.0013008,-2.71465e-06,-1.45618e-08,0.613973,0.00129533,-2.75833e-06,1.84733e-08,0.615266,0.00128986,-2.70291e-06,2.73606e-10,0.616553,0.00128446,-2.70209e-06,4.00367e-08,0.617835,0.00127918,-2.58198e-06,-4.12113e-08,0.619111,0.00127389,-2.70561e-06,6.52039e-08,0.620383,0.00126867,-2.51e-06,-4.07901e-08,0.621649,0.00126353,-2.63237e-06,3.83516e-08,0.62291,0.00125838,-2.51732e-06,6.59315e-09,0.624166,0.00125337,-2.49754e-06,-5.11939e-09,0.625416,0.00124836,-2.5129e-06,1.38846e-08,0.626662,0.00124337,-2.47124e-06,9.18514e-09,0.627903,0.00123846,-2.44369e-06,8.97952e-09,0.629139,0.0012336,-2.41675e-06,1.45012e-08,0.63037,0.00122881,-2.37325e-06,-7.37949e-09,0.631597,0.00122404,-2.39538e-06,1.50169e-08,0.632818,0.00121929,-2.35033e-06,6.91648e-09,0.634035,0.00121461,-2.32958e-06,1.69219e-08,0.635248,0.00121,-2.27882e-06,-1.49997e-08,0.636455,0.0012054,-2.32382e-06,4.30769e-08,0.637659,0.00120088,-2.19459e-06,-3.80986e-08,0.638857,0.00119638,-2.30888e-06,4.97134e-08,0.640051,0.00119191,-2.15974e-06,-4.15463e-08,0.641241,0.00118747,-2.28438e-06,5.68667e-08,0.642426,0.00118307,-2.11378e-06,-7.10641e-09,0.643607,0.00117882,-2.1351e-06,-2.8441e-08,0.644784,0.00117446,-2.22042e-06,6.12658e-08,0.645956,0.00117021,-2.03663e-06,-3.78083e-08,0.647124,0.00116602,-2.15005e-06,3.03627e-08,0.648288,0.00116181,-2.05896e-06,-2.40379e-08,0.649448,0.00115762,-2.13108e-06,6.57887e-08,0.650603,0.00115356,-1.93371e-06,-6.03028e-08,0.651755,0.00114951,-2.11462e-06,5.62134e-08,0.652902,0.00114545,-1.94598e-06,-4.53417e-08,0.654046,0.00114142,-2.082e-06,6.55489e-08,0.655185,0.00113745,-1.88536e-06,-3.80396e-08,0.656321,0.00113357,-1.99948e-06,2.70049e-08,0.657452,0.00112965,-1.91846e-06,-1.03755e-08,0.65858,0.00112578,-1.94959e-06,1.44973e-08,0.659704,0.00112192,-1.9061e-06,1.1991e-08,0.660824,0.00111815,-1.87012e-06,-2.85634e-09,0.66194,0.0011144,-1.87869e-06,-5.65782e-10,0.663053,0.00111064,-1.88039e-06,5.11947e-09,0.664162,0.0011069,-1.86503e-06,3.96924e-08,0.665267,0.00110328,-1.74595e-06,-4.46795e-08,0.666368,0.00109966,-1.87999e-06,1.98161e-08,0.667466,0.00109596,-1.82054e-06,2.502e-08,0.66856,0.00109239,-1.74548e-06,-6.86593e-10,0.669651,0.0010889,-1.74754e-06,-2.22739e-08,0.670738,0.00108534,-1.81437e-06,3.01776e-08,0.671821,0.0010818,-1.72383e-06,2.07732e-08,0.672902,0.00107841,-1.66151e-06,-5.36658e-08,0.673978,0.00107493,-1.82251e-06,7.46802e-08,0.675051,0.00107151,-1.59847e-06,-6.62411e-08,0.676121,0.00106811,-1.79719e-06,7.10748e-08,0.677188,0.00106473,-1.58397e-06,-3.92441e-08,0.678251,0.00106145,-1.7017e-06,2.62973e-08,0.679311,0.00105812,-1.62281e-06,-6.34035e-09,0.680367,0.00105486,-1.64183e-06,-9.36249e-10,0.68142,0.00105157,-1.64464e-06,1.00854e-08,0.68247,0.00104831,-1.61438e-06,2.01995e-08,0.683517,0.00104514,-1.55378e-06,-3.1279e-08,0.68456,0.00104194,-1.64762e-06,4.53114e-08,0.685601,0.00103878,-1.51169e-06,-3.07573e-08,0.686638,0.00103567,-1.60396e-06,1.81133e-08,0.687672,0.00103251,-1.54962e-06,1.79085e-08,0.688703,0.00102947,-1.49589e-06,-3.01428e-08,0.689731,0.00102639,-1.58632e-06,4.30583e-08,0.690756,0.00102334,-1.45715e-06,-2.28814e-08,0.691778,0.00102036,-1.52579e-06,-1.11373e-08,0.692797,0.00101727,-1.5592e-06,6.74305e-08,0.693812,0.00101436,-1.35691e-06,-7.97709e-08,0.694825,0.0010114,-1.59622e-06,7.28391e-08,0.695835,0.00100843,-1.37771e-06,-3.27715e-08,0.696842,0.00100558,-1.47602e-06,-1.35807e-09,0.697846,0.00100262,-1.48009e-06,3.82037e-08,0.698847,0.000999775,-1.36548e-06,-3.22474e-08,0.699846,0.000996948,-1.46223e-06,3.11809e-08,0.700841,0.000994117,-1.36868e-06,-3.28714e-08,0.701834,0.000991281,-1.4673e-06,4.07001e-08,0.702824,0.000988468,-1.3452e-06,-1.07197e-08,0.703811,0.000985746,-1.37736e-06,2.17866e-09,0.704795,0.000982998,-1.37082e-06,2.00521e-09,0.705777,0.000980262,-1.3648e-06,-1.01996e-08,0.706756,0.000977502,-1.3954e-06,3.87931e-08,0.707732,0.000974827,-1.27902e-06,-2.57632e-08,0.708706,0.000972192,-1.35631e-06,4.65513e-09,0.709676,0.000969493,-1.34235e-06,7.14257e-09,0.710645,0.00096683,-1.32092e-06,2.63791e-08,0.71161,0.000964267,-1.24178e-06,-5.30543e-08,0.712573,0.000961625,-1.40095e-06,6.66289e-08,0.713533,0.000959023,-1.20106e-06,-3.46474e-08,0.714491,0.000956517,-1.305e-06,1.23559e-08,0.715446,0.000953944,-1.26793e-06,-1.47763e-08,0.716399,0.000951364,-1.31226e-06,4.67494e-08,0.717349,0.000948879,-1.17201e-06,-5.3012e-08,0.718297,0.000946376,-1.33105e-06,4.60894e-08,0.719242,0.000943852,-1.19278e-06,-1.21366e-08,0.720185,0.00094143,-1.22919e-06,2.45673e-09,0.721125,0.000938979,-1.22182e-06,2.30966e-09,0.722063,0.000936543,-1.21489e-06,-1.16954e-08,0.722998,0.000934078,-1.24998e-06,4.44718e-08,0.723931,0.000931711,-1.11656e-06,-4.69823e-08,0.724861,0.000929337,-1.25751e-06,2.4248e-08,0.725789,0.000926895,-1.18477e-06,9.5949e-09,0.726715,0.000924554,-1.15598e-06,-3.02286e-09,0.727638,0.000922233,-1.16505e-06,2.49649e-09,0.72856,0.00091991,-1.15756e-06,-6.96321e-09,0.729478,0.000917575,-1.17845e-06,2.53564e-08,0.730395,0.000915294,-1.10238e-06,-3.48578e-08,0.731309,0.000912984,-1.20695e-06,5.44704e-08,0.732221,0.000910734,-1.04354e-06,-6.38144e-08,0.73313,0.000908455,-1.23499e-06,8.15781e-08,0.734038,0.00090623,-9.90253e-07,-8.3684e-08,0.734943,0.000903999,-1.2413e-06,7.43441e-08,0.735846,0.000901739,-1.01827e-06,-3.48787e-08,0.736746,0.000899598,-1.12291e-06,5.56596e-09,0.737645,0.000897369,-1.10621e-06,1.26148e-08,0.738541,0.000895194,-1.06837e-06,3.57935e-09,0.739435,0.000893068,-1.05763e-06,-2.69322e-08,0.740327,0.000890872,-1.13842e-06,4.45448e-08,0.741217,0.000888729,-1.00479e-06,-3.20376e-08,0.742105,0.000886623,-1.1009e-06,2.40011e-08,0.74299,0.000884493,-1.0289e-06,-4.36209e-09,0.743874,0.000882422,-1.04199e-06,-6.55268e-09,0.744755,0.000880319,-1.06164e-06,3.05728e-08,0.745634,0.000878287,-9.69926e-07,-5.61338e-08,0.746512,0.000876179,-1.13833e-06,7.4753e-08,0.747387,0.000874127,-9.14068e-07,-6.40644e-08,0.74826,0.000872106,-1.10626e-06,6.22955e-08,0.749131,0.000870081,-9.19375e-07,-6.59083e-08,0.75,0.000868044,-1.1171e-06,8.21284e-08,0.750867,0.000866056,-8.70714e-07,-8.37915e-08,0.751732,0.000864064,-1.12209e-06,7.42237e-08,0.752595,0.000862042,-8.99418e-07,-3.42894e-08,0.753456,0.00086014,-1.00229e-06,3.32955e-09,0.754315,0.000858146,-9.92297e-07,2.09712e-08,0.755173,0.000856224,-9.29384e-07,-2.76096e-08,0.756028,0.000854282,-1.01221e-06,2.98627e-08,0.756881,0.000852348,-9.22625e-07,-3.22365e-08,0.757733,0.000850406,-1.01933e-06,3.94786e-08,0.758582,0.000848485,-9.00898e-07,-6.46833e-09,0.75943,0.000846664,-9.20303e-07,-1.36052e-08,0.760275,0.000844783,-9.61119e-07,1.28447e-09,0.761119,0.000842864,-9.57266e-07,8.4674e-09,0.761961,0.000840975,-9.31864e-07,2.44506e-08,0.762801,0.000839185,-8.58512e-07,-4.6665e-08,0.763639,0.000837328,-9.98507e-07,4.30001e-08,0.764476,0.00083546,-8.69507e-07,-6.12609e-09,0.76531,0.000833703,-8.87885e-07,-1.84959e-08,0.766143,0.000831871,-9.43372e-07,2.05052e-08,0.766974,0.000830046,-8.81857e-07,-3.92026e-09,0.767803,0.000828271,-8.93618e-07,-4.82426e-09,0.768631,0.000826469,-9.0809e-07,2.32172e-08,0.769456,0.000824722,-8.38439e-07,-2.84401e-08,0.77028,0.00082296,-9.23759e-07,3.09386e-08,0.771102,0.000821205,-8.30943e-07,-3.57099e-08,0.771922,0.000819436,-9.38073e-07,5.22963e-08,0.772741,0.000817717,-7.81184e-07,-5.42658e-08,0.773558,0.000815992,-9.43981e-07,4.55579e-08,0.774373,0.000814241,-8.07308e-07,-8.75656e-09,0.775186,0.0008126,-8.33578e-07,-1.05315e-08,0.775998,0.000810901,-8.65172e-07,-8.72188e-09,0.776808,0.000809145,-8.91338e-07,4.54191e-08,0.777616,0.000807498,-7.5508e-07,-5.37454e-08,0.778423,0.000805827,-9.16317e-07,5.03532e-08,0.779228,0.000804145,-7.65257e-07,-2.84584e-08,0.780031,0.000802529,-8.50632e-07,3.87579e-09,0.780833,0.00080084,-8.39005e-07,1.29552e-08,0.781633,0.0007992,-8.00139e-07,3.90804e-09,0.782432,0.000797612,-7.88415e-07,-2.85874e-08,0.783228,0.000795949,-8.74177e-07,5.0837e-08,0.784023,0.000794353,-7.21666e-07,-5.55513e-08,0.784817,0.000792743,-8.8832e-07,5.21587e-08,0.785609,0.000791123,-7.31844e-07,-3.38744e-08,0.786399,0.000789558,-8.33467e-07,2.37342e-08,0.787188,0.000787962,-7.62264e-07,-1.45775e-09,0.787975,0.000786433,-7.66638e-07,-1.79034e-08,0.788761,0.000784846,-8.20348e-07,1.34665e-08,0.789545,0.000783246,-7.79948e-07,2.3642e-08,0.790327,0.000781757,-7.09022e-07,-4.84297e-08,0.791108,0.000780194,-8.54311e-07,5.08674e-08,0.791888,0.000778638,-7.01709e-07,-3.58303e-08,0.792666,0.000777127,-8.092e-07,3.28493e-08,0.793442,0.000775607,-7.10652e-07,-3.59624e-08,0.794217,0.000774078,-8.1854e-07,5.13959e-08,0.79499,0.000772595,-6.64352e-07,-5.04121e-08,0.795762,0.000771115,-8.15588e-07,3.10431e-08,0.796532,0.000769577,-7.22459e-07,-1.41557e-08,0.797301,0.00076809,-7.64926e-07,2.55795e-08,0.798069,0.000766636,-6.88187e-07,-2.85578e-08,0.798835,0.000765174,-7.73861e-07,2.90472e-08,0.799599,0.000763714,-6.86719e-07,-2.80262e-08,0.800362,0.000762256,-7.70798e-07,2.34531e-08,0.801123,0.000760785,-7.00438e-07,-6.18144e-09,0.801884,0.000759366,-7.18983e-07,1.27263e-09,0.802642,0.000757931,-7.15165e-07,1.09101e-09,0.803399,0.000756504,-7.11892e-07,-5.63675e-09,0.804155,0.000755064,-7.28802e-07,2.14559e-08,0.80491,0.00075367,-6.64434e-07,-2.05821e-08,0.805663,0.00075228,-7.26181e-07,1.26812e-09,0.806414,0.000750831,-7.22377e-07,1.55097e-08,0.807164,0.000749433,-6.75848e-07,-3.70216e-09,0.807913,0.00074807,-6.86954e-07,-7.0105e-10,0.80866,0.000746694,-6.89057e-07,6.5063e-09,0.809406,0.000745336,-6.69538e-07,-2.53242e-08,0.810151,0.000743921,-7.45511e-07,3.51858e-08,0.810894,0.000742535,-6.39953e-07,3.79034e-09,0.811636,0.000741267,-6.28582e-07,-5.03471e-08,0.812377,0.000739858,-7.79624e-07,7.83886e-08,0.813116,0.000738534,-5.44458e-07,-8.43935e-08,0.813854,0.000737192,-7.97638e-07,8.03714e-08,0.81459,0.000735838,-5.56524e-07,-5.82784e-08,0.815325,0.00073455,-7.31359e-07,3.35329e-08,0.816059,0.000733188,-6.3076e-07,-1.62486e-08,0.816792,0.000731878,-6.79506e-07,3.14614e-08,0.817523,0.000730613,-5.85122e-07,-4.99925e-08,0.818253,0.000729293,-7.35099e-07,4.92994e-08,0.818982,0.000727971,-5.87201e-07,-2.79959e-08,0.819709,0.000726712,-6.71189e-07,3.07959e-09,0.820435,0.000725379,-6.6195e-07,1.56777e-08,0.82116,0.000724102,-6.14917e-07,-6.18564e-09,0.821883,0.000722854,-6.33474e-07,9.06488e-09,0.822606,0.000721614,-6.06279e-07,-3.00739e-08,0.823327,0.000720311,-6.96501e-07,5.16262e-08,0.824046,0.000719073,-5.41623e-07,-5.72214e-08,0.824765,0.000717818,-7.13287e-07,5.80503e-08,0.825482,0.000716566,-5.39136e-07,-5.57703e-08,0.826198,0.00071532,-7.06447e-07,4.58215e-08,0.826912,0.000714045,-5.68983e-07,-8.30636e-09,0.827626,0.000712882,-5.93902e-07,-1.25961e-08,0.828338,0.000711656,-6.3169e-07,-9.13985e-10,0.829049,0.00071039,-6.34432e-07,1.62519e-08,0.829759,0.00070917,-5.85676e-07,-4.48904e-09,0.830468,0.000707985,-5.99143e-07,1.70418e-09,0.831175,0.000706792,-5.9403e-07,-2.32768e-09,0.831881,0.000705597,-6.01014e-07,7.60648e-09,0.832586,0.000704418,-5.78194e-07,-2.80982e-08,0.83329,0.000703177,-6.62489e-07,4.51817e-08,0.833993,0.000701988,-5.26944e-07,-3.34192e-08,0.834694,0.000700834,-6.27201e-07,2.88904e-08,0.835394,0.000699666,-5.4053e-07,-2.25378e-08,0.836093,0.000698517,-6.08143e-07,1.65589e-09,0.836791,0.000697306,-6.03176e-07,1.59142e-08,0.837488,0.000696147,-5.55433e-07,-5.70801e-09,0.838184,0.000695019,-5.72557e-07,6.91792e-09,0.838878,0.000693895,-5.51803e-07,-2.19637e-08,0.839571,0.000692725,-6.17694e-07,2.13321e-08,0.840263,0.000691554,-5.53698e-07,-3.75996e-09,0.840954,0.000690435,-5.64978e-07,-6.29219e-09,0.841644,0.000689287,-5.83855e-07,2.89287e-08,0.842333,0.000688206,-4.97068e-07,-4.98181e-08,0.843021,0.000687062,-6.46523e-07,5.11344e-08,0.843707,0.000685922,-4.9312e-07,-3.55102e-08,0.844393,0.00068483,-5.9965e-07,3.13019e-08,0.845077,0.000683724,-5.05745e-07,-3.00925e-08,0.84576,0.000682622,-5.96022e-07,2.94636e-08,0.846442,0.000681519,-5.07631e-07,-2.81572e-08,0.847123,0.000680419,-5.92103e-07,2.35606e-08,0.847803,0.000679306,-5.21421e-07,-6.48045e-09,0.848482,0.000678243,-5.40863e-07,2.36124e-09,0.849159,0.000677169,-5.33779e-07,-2.96461e-09,0.849836,0.000676092,-5.42673e-07,9.49728e-09,0.850512,0.000675035,-5.14181e-07,-3.50245e-08,0.851186,0.000673902,-6.19254e-07,7.09959e-08,0.851859,0.000672876,-4.06267e-07,-7.01453e-08,0.852532,0.000671853,-6.16703e-07,3.07714e-08,0.853203,0.000670712,-5.24388e-07,6.66423e-09,0.853873,0.000669684,-5.04396e-07,2.17629e-09,0.854542,0.000668681,-4.97867e-07,-1.53693e-08,0.855211,0.000667639,-5.43975e-07,-3.03752e-10,0.855878,0.000666551,-5.44886e-07,1.65844e-08,0.856544,0.000665511,-4.95133e-07,-6.42907e-09,0.857209,0.000664501,-5.1442e-07,9.13195e-09,0.857873,0.0006635,-4.87024e-07,-3.00987e-08,0.858536,0.000662435,-5.7732e-07,5.16584e-08,0.859198,0.000661436,-4.22345e-07,-5.73255e-08,0.859859,0.000660419,-5.94322e-07,5.84343e-08,0.860518,0.000659406,-4.19019e-07,-5.72022e-08,0.861177,0.000658396,-5.90626e-07,5.11653e-08,0.861835,0.000657368,-4.3713e-07,-2.82495e-08,0.862492,0.000656409,-5.21878e-07,2.22788e-09,0.863148,0.000655372,-5.15195e-07,1.9338e-08,0.863803,0.0006544,-4.5718e-07,-1.99754e-08,0.864457,0.000653425,-5.17107e-07,9.59024e-10,0.86511,0.000652394,-5.1423e-07,1.61393e-08,0.865762,0.000651414,-4.65812e-07,-5.91149e-09,0.866413,0.000650465,-4.83546e-07,7.50665e-09,0.867063,0.00064952,-4.61026e-07,-2.4115e-08,0.867712,0.000648526,-5.33371e-07,2.93486e-08,0.86836,0.000647547,-4.45325e-07,-3.36748e-08,0.869007,0.000646555,-5.4635e-07,4.57461e-08,0.869653,0.0006456,-4.09112e-07,-3.01002e-08,0.870298,0.000644691,-4.99412e-07,1.50501e-08,0.870942,0.000643738,-4.54262e-07,-3.01002e-08,0.871585,0.000642739,-5.44563e-07,4.57461e-08,0.872228,0.000641787,-4.07324e-07,-3.36748e-08,0.872869,0.000640871,-5.08349e-07,2.93486e-08,0.873509,0.000639943,-4.20303e-07,-2.4115e-08,0.874149,0.00063903,-4.92648e-07,7.50655e-09,0.874787,0.000638067,-4.70128e-07,-5.91126e-09,0.875425,0.000637109,-4.87862e-07,1.61385e-08,0.876062,0.000636182,-4.39447e-07,9.61961e-10,0.876697,0.000635306,-4.36561e-07,-1.99863e-08,0.877332,0.000634373,-4.9652e-07,1.93785e-08,0.877966,0.000633438,-4.38384e-07,2.07697e-09,0.878599,0.000632567,-4.32153e-07,-2.76864e-08,0.879231,0.00063162,-5.15212e-07,4.90641e-08,0.879862,0.000630737,-3.6802e-07,-4.93606e-08,0.880493,0.000629852,-5.16102e-07,2.9169e-08,0.881122,0.000628908,-4.28595e-07,-7.71083e-09,0.881751,0.000628027,-4.51727e-07,1.6744e-09,0.882378,0.000627129,-4.46704e-07,1.01317e-09,0.883005,0.000626239,-4.43665e-07,-5.72703e-09,0.883631,0.000625334,-4.60846e-07,2.1895e-08,0.884255,0.000624478,-3.95161e-07,-2.22481e-08,0.88488,0.000623621,-4.61905e-07,7.4928e-09,0.885503,0.00062272,-4.39427e-07,-7.72306e-09,0.886125,0.000621818,-4.62596e-07,2.33995e-08,0.886746,0.000620963,-3.92398e-07,-2.62704e-08,0.887367,0.000620099,-4.71209e-07,2.20775e-08,0.887987,0.000619223,-4.04976e-07,-2.43496e-09,0.888605,0.000618406,-4.12281e-07,-1.23377e-08,0.889223,0.000617544,-4.49294e-07,-7.81876e-09,0.88984,0.000616622,-4.72751e-07,4.36128e-08,0.890457,0.000615807,-3.41912e-07,-4.7423e-08,0.891072,0.000614981,-4.84181e-07,2.68698e-08,0.891687,0.000614093,-4.03572e-07,-4.51384e-10,0.8923,0.000613285,-4.04926e-07,-2.50643e-08,0.892913,0.0006124,-4.80119e-07,4.11038e-08,0.893525,0.000611563,-3.56808e-07,-2.01414e-08,0.894136,0.000610789,-4.17232e-07,-2.01426e-08,0.894747,0.000609894,-4.7766e-07,4.11073e-08,0.895356,0.000609062,-3.54338e-07,-2.50773e-08,0.895965,0.000608278,-4.2957e-07,-4.02954e-10,0.896573,0.000607418,-4.30779e-07,2.66891e-08,0.89718,0.000606636,-3.50711e-07,-4.67489e-08,0.897786,0.000605795,-4.90958e-07,4.10972e-08,0.898391,0.000604936,-3.67666e-07,1.56948e-09,0.898996,0.000604205,-3.62958e-07,-4.73751e-08,0.8996,0.000603337,-5.05083e-07,6.87214e-08,0.900202,0.000602533,-2.98919e-07,-4.86966e-08,0.900805,0.000601789,-4.45009e-07,6.85589e-09,0.901406,0.00060092,-4.24441e-07,2.1273e-08,0.902007,0.000600135,-3.60622e-07,-3.23434e-08,0.902606,0.000599317,-4.57652e-07,4.84959e-08,0.903205,0.000598547,-3.12164e-07,-4.24309e-08,0.903803,0.000597795,-4.39457e-07,2.01844e-09,0.904401,0.000596922,-4.33402e-07,3.43571e-08,0.904997,0.000596159,-3.30331e-07,-2.02374e-08,0.905593,0.000595437,-3.91043e-07,-1.30123e-08,0.906188,0.000594616,-4.3008e-07,1.26819e-08,0.906782,0.000593794,-3.92034e-07,2.18894e-08,0.907376,0.000593076,-3.26366e-07,-4.06349e-08,0.907968,0.000592301,-4.4827e-07,2.1441e-08,0.90856,0.000591469,-3.83947e-07,1.44754e-08,0.909151,0.000590744,-3.40521e-07,-1.97379e-08,0.909742,0.000590004,-3.99735e-07,4.87161e-09,0.910331,0.000589219,-3.8512e-07,2.51532e-10,0.91092,0.00058845,-3.84366e-07,-5.87776e-09,0.911508,0.000587663,-4.01999e-07,2.32595e-08,0.912096,0.000586929,-3.3222e-07,-2.75554e-08,0.912682,0.000586182,-4.14887e-07,2.73573e-08,0.913268,0.000585434,-3.32815e-07,-2.22692e-08,0.913853,0.000584702,-3.99622e-07,2.11486e-09,0.914437,0.000583909,-3.93278e-07,1.38098e-08,0.915021,0.000583164,-3.51848e-07,2.25042e-09,0.915604,0.000582467,-3.45097e-07,-2.28115e-08,0.916186,0.000581708,-4.13531e-07,2.93911e-08,0.916767,0.000580969,-3.25358e-07,-3.51481e-08,0.917348,0.000580213,-4.30803e-07,5.15967e-08,0.917928,0.000579506,-2.76012e-07,-5.20296e-08,0.918507,0.000578798,-4.32101e-07,3.73124e-08,0.919085,0.000578046,-3.20164e-07,-3.76154e-08,0.919663,0.000577293,-4.3301e-07,5.35447e-08,0.92024,0.000576587,-2.72376e-07,-5.7354e-08,0.920816,0.000575871,-4.44438e-07,5.66621e-08,0.921391,0.000575152,-2.74452e-07,-5.00851e-08,0.921966,0.000574453,-4.24707e-07,2.4469e-08,0.92254,0.000573677,-3.513e-07,1.18138e-08,0.923114,0.000573009,-3.15859e-07,-1.21195e-08,0.923686,0.000572341,-3.52217e-07,-2.29403e-08,0.924258,0.000571568,-4.21038e-07,4.4276e-08,0.924829,0.000570859,-2.8821e-07,-3.49546e-08,0.9254,0.000570178,-3.93074e-07,3.59377e-08,0.92597,0.000569499,-2.85261e-07,-4.91915e-08,0.926539,0.000568781,-4.32835e-07,4.16189e-08,0.927107,0.00056804,-3.07979e-07,1.92523e-09,0.927675,0.00056743,-3.02203e-07,-4.93198e-08,0.928242,0.000566678,-4.50162e-07,7.61447e-08,0.928809,0.000566006,-2.21728e-07,-7.6445e-08,0.929374,0.000565333,-4.51063e-07,5.08216e-08,0.929939,0.000564583,-2.98599e-07,-7.63212e-09,0.930503,0.000563963,-3.21495e-07,-2.02931e-08,0.931067,0.000563259,-3.82374e-07,2.92001e-08,0.93163,0.000562582,-2.94774e-07,-3.69025e-08,0.932192,0.000561882,-4.05482e-07,5.88053e-08,0.932754,0.000561247,-2.29066e-07,-7.91094e-08,0.933315,0.000560552,-4.66394e-07,7.88184e-08,0.933875,0.000559856,-2.29939e-07,-5.73501e-08,0.934434,0.000559224,-4.01989e-07,3.13727e-08,0.934993,0.000558514,-3.07871e-07,-8.53611e-09,0.935551,0.000557873,-3.33479e-07,2.77175e-09,0.936109,0.000557214,-3.25164e-07,-2.55091e-09,0.936666,0.000556556,-3.32817e-07,7.43188e-09,0.937222,0.000555913,-3.10521e-07,-2.71766e-08,0.937778,0.00055521,-3.92051e-07,4.167e-08,0.938333,0.000554551,-2.67041e-07,-2.02941e-08,0.938887,0.000553956,-3.27923e-07,-2.00984e-08,0.93944,0.00055324,-3.88218e-07,4.10828e-08,0.939993,0.000552587,-2.6497e-07,-2.50237e-08,0.940546,0.000551982,-3.40041e-07,-5.92583e-10,0.941097,0.0005513,-3.41819e-07,2.7394e-08,0.941648,0.000550698,-2.59637e-07,-4.93788e-08,0.942199,0.000550031,-4.07773e-07,5.09119e-08,0.942748,0.000549368,-2.55038e-07,-3.50595e-08,0.943297,0.000548753,-3.60216e-07,2.97214e-08,0.943846,0.000548122,-2.71052e-07,-2.42215e-08,0.944394,0.000547507,-3.43716e-07,7.55985e-09,0.944941,0.000546842,-3.21037e-07,-6.01796e-09,0.945487,0.000546182,-3.3909e-07,1.65119e-08,0.946033,0.000545553,-2.89555e-07,-4.2498e-10,0.946578,0.000544973,-2.9083e-07,-1.4812e-08,0.947123,0.000544347,-3.35266e-07,6.83068e-11,0.947667,0.000543676,-3.35061e-07,1.45388e-08,0.94821,0.00054305,-2.91444e-07,1.38123e-09,0.948753,0.000542471,-2.87301e-07,-2.00637e-08,0.949295,0.000541836,-3.47492e-07,1.92688e-08,0.949837,0.000541199,-2.89685e-07,2.59298e-09,0.950378,0.000540628,-2.81906e-07,-2.96407e-08,0.950918,0.000539975,-3.70829e-07,5.63652e-08,0.951458,0.000539402,-2.01733e-07,-7.66107e-08,0.951997,0.000538769,-4.31565e-07,7.12638e-08,0.952535,0.00053812,-2.17774e-07,-2.96305e-08,0.953073,0.000537595,-3.06665e-07,-1.23464e-08,0.95361,0.000536945,-3.43704e-07,1.94114e-08,0.954147,0.000536316,-2.8547e-07,-5.69451e-09,0.954683,0.000535728,-3.02554e-07,3.36666e-09,0.955219,0.000535133,-2.92454e-07,-7.77208e-09,0.955753,0.000534525,-3.1577e-07,2.77216e-08,0.956288,0.000533976,-2.32605e-07,-4.35097e-08,0.956821,0.00053338,-3.63134e-07,2.7108e-08,0.957354,0.000532735,-2.8181e-07,-5.31772e-09,0.957887,0.000532156,-2.97764e-07,-5.83718e-09,0.958419,0.000531543,-3.15275e-07,2.86664e-08,0.95895,0.000530998,-2.29276e-07,-4.9224e-08,0.959481,0.000530392,-3.76948e-07,4.90201e-08,0.960011,0.000529785,-2.29887e-07,-2.76471e-08,0.96054,0.000529243,-3.12829e-07,1.96385e-09,0.961069,0.000528623,-3.06937e-07,1.97917e-08,0.961598,0.000528068,-2.47562e-07,-2.15261e-08,0.962125,0.000527508,-3.1214e-07,6.70795e-09,0.962653,0.000526904,-2.92016e-07,-5.30573e-09,0.963179,0.000526304,-3.07934e-07,1.4515e-08,0.963705,0.000525732,-2.64389e-07,6.85048e-09,0.964231,0.000525224,-2.43837e-07,-4.19169e-08,0.964756,0.00052461,-3.69588e-07,4.1608e-08,0.96528,0.000523996,-2.44764e-07,-5.30598e-09,0.965804,0.000523491,-2.60682e-07,-2.03841e-08,0.966327,0.000522908,-3.21834e-07,2.72378e-08,0.966849,0.000522346,-2.40121e-07,-2.89625e-08,0.967371,0.000521779,-3.27008e-07,2.90075e-08,0.967893,0.000521212,-2.39986e-07,-2.74629e-08,0.968414,0.00052065,-3.22374e-07,2.12396e-08,0.968934,0.000520069,-2.58656e-07,2.10922e-09,0.969454,0.000519558,-2.52328e-07,-2.96765e-08,0.969973,0.000518964,-3.41357e-07,5.6992e-08,0.970492,0.000518452,-1.70382e-07,-7.90821e-08,0.97101,0.000517874,-4.07628e-07,8.05224e-08,0.971528,0.000517301,-1.66061e-07,-6.41937e-08,0.972045,0.000516776,-3.58642e-07,5.70429e-08,0.972561,0.00051623,-1.87513e-07,-4.47686e-08,0.973077,0.00051572,-3.21819e-07,2.82237e-09,0.973593,0.000515085,-3.13352e-07,3.34792e-08,0.974108,0.000514559,-2.12914e-07,-1.75298e-08,0.974622,0.000514081,-2.65503e-07,-2.29648e-08,0.975136,0.000513481,-3.34398e-07,4.97843e-08,0.975649,0.000512961,-1.85045e-07,-5.6963e-08,0.976162,0.00051242,-3.55934e-07,5.88585e-08,0.976674,0.000511885,-1.79359e-07,-5.92616e-08,0.977185,0.000511348,-3.57143e-07,5.89785e-08,0.977696,0.000510811,-1.80208e-07,-5.74433e-08,0.978207,0.000510278,-3.52538e-07,5.15854e-08,0.978717,0.000509728,-1.97781e-07,-2.9689e-08,0.979226,0.000509243,-2.86848e-07,7.56591e-09,0.979735,0.000508692,-2.64151e-07,-5.74649e-10,0.980244,0.000508162,-2.65875e-07,-5.26732e-09,0.980752,0.000507615,-2.81677e-07,2.16439e-08,0.981259,0.000507116,-2.16745e-07,-2.17037e-08,0.981766,0.000506618,-2.81856e-07,5.56636e-09,0.982272,0.000506071,-2.65157e-07,-5.61689e-10,0.982778,0.000505539,-2.66842e-07,-3.31963e-09,0.983283,0.000504995,-2.76801e-07,1.38402e-08,0.983788,0.000504483,-2.3528e-07,7.56339e-09,0.984292,0.000504035,-2.1259e-07,-4.40938e-08,0.984796,0.000503478,-3.44871e-07,4.96026e-08,0.985299,0.000502937,-1.96064e-07,-3.51071e-08,0.985802,0.000502439,-3.01385e-07,3.12212e-08,0.986304,0.00050193,-2.07721e-07,-3.0173e-08,0.986806,0.000501424,-2.9824e-07,2.9866e-08,0.987307,0.000500917,-2.08642e-07,-2.96865e-08,0.987808,0.000500411,-2.97702e-07,2.92753e-08,0.988308,0.000499903,-2.09876e-07,-2.78101e-08,0.988807,0.0004994,-2.93306e-07,2.23604e-08,0.989307,0.000498881,-2.26225e-07,-2.02681e-09,0.989805,0.000498422,-2.32305e-07,-1.42531e-08,0.990303,0.000497915,-2.75065e-07,-5.65232e-10,0.990801,0.000497363,-2.76761e-07,1.65141e-08,0.991298,0.000496859,-2.27218e-07,-5.88639e-09,0.991795,0.000496387,-2.44878e-07,7.0315e-09,0.992291,0.000495918,-2.23783e-07,-2.22396e-08,0.992787,0.000495404,-2.90502e-07,2.23224e-08,0.993282,0.00049489,-2.23535e-07,-7.44543e-09,0.993776,0.000494421,-2.45871e-07,7.45924e-09,0.994271,0.000493951,-2.23493e-07,-2.23915e-08,0.994764,0.000493437,-2.90668e-07,2.25021e-08,0.995257,0.000492923,-2.23161e-07,-8.01218e-09,0.99575,0.000492453,-2.47198e-07,9.54669e-09,0.996242,0.000491987,-2.18558e-07,-3.01746e-08,0.996734,0.000491459,-3.09082e-07,5.1547e-08,0.997225,0.000490996,-1.54441e-07,-5.68039e-08,0.997716,0.000490517,-3.24853e-07,5.64594e-08,0.998206,0.000490036,-1.55474e-07,-4.98245e-08,0.998696,0.000489576,-3.04948e-07,2.36292e-08,0.999186,0.000489037,-2.3406e-07,1.49121e-08,0.999674,0.000488613,-1.89324e-07,-2.3673e-08,1.00016,0.000488164,-2.60343e-07,2.01754e-08,1.00065,0.000487704,-1.99816e-07,-5.70288e-08,1.00114,0.000487133,-3.70903e-07,8.87303e-08,1.00162,0.000486657,-1.04712e-07,-5.94737e-08,1.00211,0.000486269,-2.83133e-07,2.99553e-08,1.0026,0.000485793,-1.93267e-07,-6.03474e-08,1.00308,0.000485225,-3.74309e-07,9.2225e-08,1.00357,0.000484754,-9.76345e-08,-7.0134e-08,1.00405,0.000484348,-3.08036e-07,6.91016e-08,1.00454,0.000483939,-1.00731e-07,-8.70633e-08,1.00502,0.000483476,-3.61921e-07,4.07328e-08,1.0055,0.000482875,-2.39723e-07,4.33413e-08,1.00599,0.000482525,-1.09699e-07,-9.48886e-08,1.00647,0.000482021,-3.94365e-07,9.77947e-08,1.00695,0.000481526,-1.00981e-07,-5.78713e-08,1.00743,0.00048115,-2.74595e-07,1.44814e-08,1.00791,0.000480645,-2.31151e-07,-5.42665e-11,1.00839,0.000480182,-2.31314e-07,-1.42643e-08,1.00887,0.000479677,-2.74106e-07,5.71115e-08,1.00935,0.0004793,-1.02772e-07,-9.49724e-08,1.00983,0.000478809,-3.87689e-07,8.43596e-08,1.01031,0.000478287,-1.3461e-07,-4.04755e-09,1.01079,0.000478006,-1.46753e-07,-6.81694e-08,1.01127,0.000477508,-3.51261e-07,3.83067e-08,1.01174,0.00047692,-2.36341e-07,3.41521e-08,1.01222,0.00047655,-1.33885e-07,-5.57058e-08,1.0127,0.000476115,-3.01002e-07,6.94616e-08,1.01317,0.000475721,-9.26174e-08,-1.02931e-07,1.01365,0.000475227,-4.01412e-07,1.03846e-07,1.01412,0.000474736,-8.98751e-08,-7.40321e-08,1.0146,0.000474334,-3.11971e-07,7.30735e-08,1.01507,0.00047393,-9.27508e-08,-9.90527e-08,1.01554,0.000473447,-3.89909e-07,8.47188e-08,1.01602,0.000472921,-1.35753e-07,-1.40381e-09,1.01649,0.000472645,-1.39964e-07,-7.91035e-08,1.01696,0.000472128,-3.77275e-07,7.93993e-08,1.01744,0.000471612,-1.39077e-07,-7.52607e-11,1.01791,0.000471334,-1.39302e-07,-7.90983e-08,1.01838,0.000470818,-3.76597e-07,7.80499e-08,1.01885,0.000470299,-1.42448e-07,5.31733e-09,1.01932,0.00047003,-1.26496e-07,-9.93193e-08,1.01979,0.000469479,-4.24453e-07,1.53541e-07,1.02026,0.00046909,3.617e-08,-1.57217e-07,1.02073,0.000468691,-4.35482e-07,1.177e-07,1.02119,0.000468173,-8.23808e-08,-7.51659e-08,1.02166,0.000467783,-3.07878e-07,6.37538e-08,1.02213,0.000467358,-1.16617e-07,-6.064e-08,1.0226,0.000466943,-2.98537e-07,5.9597e-08,1.02306,0.000466525,-1.19746e-07,-5.85386e-08,1.02353,0.00046611,-2.95362e-07,5.53482e-08,1.024,0.000465685,-1.29317e-07,-4.36449e-08,1.02446,0.000465296,-2.60252e-07,2.20268e-11,1.02493,0.000464775,-2.60186e-07,4.35568e-08,1.02539,0.000464386,-1.29516e-07,-5.50398e-08,1.02586,0.000463961,-2.94635e-07,5.73932e-08,1.02632,0.000463544,-1.22456e-07,-5.53236e-08,1.02678,0.000463133,-2.88426e-07,4.46921e-08,1.02725,0.000462691,-1.5435e-07,-4.23534e-09,1.02771,0.000462369,-1.67056e-07,-2.77507e-08,1.02817,0.000461952,-2.50308e-07,-3.97101e-09,1.02863,0.000461439,-2.62221e-07,4.36348e-08,1.02909,0.000461046,-1.31317e-07,-5.13589e-08,1.02955,0.000460629,-2.85394e-07,4.25913e-08,1.03001,0.000460186,-1.5762e-07,2.0285e-10,1.03047,0.000459871,-1.57011e-07,-4.34027e-08,1.03093,0.000459427,-2.87219e-07,5.41987e-08,1.03139,0.000459015,-1.24623e-07,-5.4183e-08,1.03185,0.000458604,-2.87172e-07,4.33239e-08,1.03231,0.000458159,-1.572e-07,9.65817e-11,1.03277,0.000457845,-1.56911e-07,-4.37103e-08,1.03323,0.0004574,-2.88041e-07,5.55351e-08,1.03368,0.000456991,-1.21436e-07,-5.9221e-08,1.03414,0.00045657,-2.99099e-07,6.21394e-08,1.0346,0.000456158,-1.1268e-07,-7.01275e-08,1.03505,0.000455723,-3.23063e-07,9.91614e-08,1.03551,0.000455374,-2.55788e-08,-8.80996e-08,1.03596,0.000455058,-2.89878e-07,1.48184e-08,1.03642,0.000454523,-2.45422e-07,2.88258e-08,1.03687,0.000454119,-1.58945e-07,-1.09125e-08,1.03733,0.000453768,-1.91682e-07,1.48241e-08,1.03778,0.000453429,-1.4721e-07,-4.83838e-08,1.03823,0.00045299,-2.92361e-07,5.95019e-08,1.03869,0.000452584,-1.13856e-07,-7.04146e-08,1.03914,0.000452145,-3.25099e-07,1.02947e-07,1.03959,0.000451803,-1.62583e-08,-1.02955e-07,1.04004,0.000451462,-3.25123e-07,7.04544e-08,1.04049,0.000451023,-1.1376e-07,-5.96534e-08,1.04094,0.000450616,-2.9272e-07,4.89499e-08,1.04139,0.000450178,-1.45871e-07,-1.69369e-08,1.04184,0.000449835,-1.96681e-07,1.87977e-08,1.04229,0.000449498,-1.40288e-07,-5.82539e-08,1.04274,0.000449043,-3.1505e-07,9.50087e-08,1.04319,0.000448698,-3.00238e-08,-8.33623e-08,1.04364,0.000448388,-2.80111e-07,2.20363e-11,1.04409,0.000447828,-2.80045e-07,8.32742e-08,1.04454,0.000447517,-3.02221e-08,-9.47002e-08,1.04498,0.000447173,-3.14323e-07,5.7108e-08,1.04543,0.000446716,-1.42999e-07,-1.45225e-08,1.04588,0.000446386,-1.86566e-07,9.82022e-10,1.04632,0.000446016,-1.8362e-07,1.05944e-08,1.04677,0.00044568,-1.51837e-07,-4.33597e-08,1.04721,0.000445247,-2.81916e-07,4.36352e-08,1.04766,0.000444814,-1.51011e-07,-1.19717e-08,1.0481,0.000444476,-1.86926e-07,4.25158e-09,1.04855,0.000444115,-1.74171e-07,-5.03461e-09,1.04899,0.000443751,-1.89275e-07,1.58868e-08,1.04944,0.00044342,-1.41614e-07,-5.85127e-08,1.04988,0.000442961,-3.17152e-07,9.89548e-08,1.05032,0.000442624,-2.0288e-08,-9.88878e-08,1.05076,0.000442287,-3.16951e-07,5.81779e-08,1.05121,0.000441827,-1.42418e-07,-1.46144e-08,1.05165,0.000441499,-1.86261e-07,2.79892e-10,1.05209,0.000441127,-1.85421e-07,1.34949e-08,1.05253,0.000440797,-1.44937e-07,-5.42594e-08,1.05297,0.000440344,-3.07715e-07,8.43335e-08,1.05341,0.000439982,-5.47146e-08,-4.46558e-08,1.05385,0.000439738,-1.88682e-07,-2.49193e-08,1.05429,0.000439286,-2.6344e-07,2.5124e-08,1.05473,0.000438835,-1.88068e-07,4.36328e-08,1.05517,0.000438589,-5.71699e-08,-8.04459e-08,1.05561,0.000438234,-2.98508e-07,3.97324e-08,1.05605,0.000437756,-1.79311e-07,4.07258e-08,1.05648,0.000437519,-5.71332e-08,-8.34263e-08,1.05692,0.000437155,-3.07412e-07,5.45608e-08,1.05736,0.000436704,-1.4373e-07,-1.56078e-08,1.05779,0.000436369,-1.90553e-07,7.87043e-09,1.05823,0.000436012,-1.66942e-07,-1.58739e-08,1.05867,0.00043563,-2.14563e-07,5.56251e-08,1.0591,0.000435368,-4.76881e-08,-8.74172e-08,1.05954,0.000435011,-3.0994e-07,5.56251e-08,1.05997,0.000434558,-1.43064e-07,-1.58739e-08,1.06041,0.000434224,-1.90686e-07,7.87042e-09,1.06084,0.000433866,-1.67075e-07,-1.56078e-08,1.06127,0.000433485,-2.13898e-07,5.45609e-08,1.06171,0.000433221,-5.02157e-08,-8.34263e-08,1.06214,0.00043287,-3.00495e-07,4.07258e-08,1.06257,0.000432391,-1.78317e-07,3.97325e-08,1.063,0.000432154,-5.91198e-08,-8.04464e-08,1.06344,0.000431794,-3.00459e-07,4.36347e-08,1.06387,0.000431324,-1.69555e-07,2.5117e-08,1.0643,0.000431061,-9.42041e-08,-2.48934e-08,1.06473,0.000430798,-1.68884e-07,-4.47527e-08,1.06516,0.000430326,-3.03142e-07,8.46951e-08,1.06559,0.000429973,-4.90573e-08,-5.56089e-08,1.06602,0.000429708,-2.15884e-07,1.85314e-08,1.06645,0.000429332,-1.6029e-07,-1.85166e-08,1.06688,0.000428956,-2.1584e-07,5.5535e-08,1.06731,0.000428691,-4.92347e-08,-8.44142e-08,1.06774,0.000428339,-3.02477e-07,4.37032e-08,1.06816,0.000427865,-1.71368e-07,2.88107e-08,1.06859,0.000427609,-8.49356e-08,-3.97367e-08,1.06902,0.00042732,-2.04146e-07,1.09267e-08,1.06945,0.000426945,-1.71365e-07,-3.97023e-09,1.06987,0.00042659,-1.83276e-07,4.9542e-09,1.0703,0.000426238,-1.68414e-07,-1.58466e-08,1.07073,0.000425854,-2.15953e-07,5.84321e-08,1.07115,0.000425597,-4.0657e-08,-9.86725e-08,1.07158,0.00042522,-3.36674e-07,9.78392e-08,1.072,0.00042484,-4.31568e-08,-5.42658e-08,1.07243,0.000424591,-2.05954e-07,1.45377e-11,1.07285,0.000424179,-2.0591e-07,5.42076e-08,1.07328,0.00042393,-4.32877e-08,-9.76357e-08,1.0737,0.00042355,-3.36195e-07,9.79165e-08,1.07412,0.000423172,-4.24451e-08,-5.56118e-08,1.07455,0.00042292,-2.09281e-07,5.32143e-09,1.07497,0.000422518,-1.93316e-07,3.43261e-08,1.07539,0.000422234,-9.0338e-08,-2.34165e-08,1.07581,0.000421983,-1.60588e-07,-5.98692e-08,1.07623,0.000421482,-3.40195e-07,1.43684e-07,1.07666,0.000421233,9.08574e-08,-1.5724e-07,1.07708,0.000420943,-3.80862e-07,1.27647e-07,1.0775,0.000420564,2.0791e-09,-1.1493e-07,1.07792,0.000420223,-3.4271e-07,9.36534e-08,1.07834,0.000419819,-6.17499e-08,-2.12653e-08,1.07876,0.000419632,-1.25546e-07,-8.59219e-09,1.07918,0.000419355,-1.51322e-07,-6.35752e-08,1.0796,0.000418861,-3.42048e-07,1.43684e-07,1.08002,0.000418608,8.90034e-08,-1.53532e-07,1.08043,0.000418326,-3.71593e-07,1.12817e-07,1.08085,0.000417921,-3.31414e-08,-5.93184e-08,1.08127,0.000417677,-2.11097e-07,5.24697e-09,1.08169,0.00041727,-1.95356e-07,3.83305e-08,1.0821,0.000416995,-8.03642e-08,-3.93597e-08,1.08252,0.000416716,-1.98443e-07,-1.0094e-10,1.08294,0.000416319,-1.98746e-07,3.97635e-08,1.08335,0.00041604,-7.94557e-08,-3.97437e-08,1.08377,0.000415762,-1.98687e-07,1.94215e-12,1.08419,0.000415365,-1.98681e-07,3.97359e-08,1.0846,0.000415087,-7.94732e-08,-3.97362e-08,1.08502,0.000414809,-1.98682e-07,-4.31063e-13,1.08543,0.000414411,-1.98683e-07,3.97379e-08,1.08584,0.000414133,-7.94694e-08,-3.97418e-08,1.08626,0.000413855,-1.98695e-07,2.00563e-11,1.08667,0.000413458,-1.98635e-07,3.96616e-08,1.08709,0.000413179,-7.965e-08,-3.9457e-08,1.0875,0.000412902,-1.98021e-07,-1.04281e-09,1.08791,0.000412502,-2.01149e-07,4.36282e-08,1.08832,0.000412231,-7.02648e-08,-5.42608e-08,1.08874,0.000411928,-2.33047e-07,5.42057e-08,1.08915,0.000411624,-7.04301e-08,-4.33527e-08,1.08956,0.000411353,-2.00488e-07,-4.07378e-12,1.08997,0.000410952,-2.005e-07,4.3369e-08,1.09038,0.000410681,-7.03934e-08,-5.42627e-08,1.09079,0.000410378,-2.33182e-07,5.44726e-08,1.0912,0.000410075,-6.97637e-08,-4.44186e-08,1.09161,0.000409802,-2.03019e-07,3.99235e-09,1.09202,0.000409408,-1.91042e-07,2.84491e-08,1.09243,0.000409111,-1.05695e-07,1.42043e-09,1.09284,0.000408904,-1.01434e-07,-3.41308e-08,1.09325,0.000408599,-2.03826e-07,1.58937e-08,1.09366,0.000408239,-1.56145e-07,-2.94438e-08,1.09406,0.000407838,-2.44476e-07,1.01881e-07,1.09447,0.000407655,6.11676e-08,-1.39663e-07,1.09488,0.000407358,-3.57822e-07,9.91432e-08,1.09529,0.00040694,-6.03921e-08,-1.84912e-08,1.09569,0.000406764,-1.15866e-07,-2.51785e-08,1.0961,0.000406457,-1.91401e-07,-4.03115e-12,1.09651,0.000406074,-1.91413e-07,2.51947e-08,1.09691,0.000405767,-1.15829e-07,1.84346e-08,1.09732,0.00040559,-6.05254e-08,-9.89332e-08,1.09772,0.000405172,-3.57325e-07,1.3888e-07,1.09813,0.000404874,5.93136e-08,-9.8957e-08,1.09853,0.000404696,-2.37557e-07,1.853e-08,1.09894,0.000404277,-1.81968e-07,2.48372e-08,1.09934,0.000403987,-1.07456e-07,1.33047e-09,1.09975,0.000403776,-1.03465e-07,-3.01591e-08,1.10015,0.000403479,-1.93942e-07,9.66054e-11,1.10055,0.000403091,-1.93652e-07,2.97727e-08,1.10096,0.000402793,-1.04334e-07,2.19273e-11,1.10136,0.000402585,-1.04268e-07,-2.98604e-08,1.10176,0.000402287,-1.93849e-07,2.10325e-10,1.10216,0.0004019,-1.93218e-07,2.90191e-08,1.10256,0.0004016,-1.06161e-07,2.92264e-09,1.10297,0.000401397,-9.73931e-08,-4.07096e-08,1.10337,0.00040108,-2.19522e-07,4.07067e-08,1.10377,0.000400763,-9.7402e-08,-2.90783e-09,1.10417,0.000400559,-1.06126e-07,-2.90754e-08,1.10457,0.00040026,-1.93352e-07,9.00021e-14,1.10497,0.000399873,-1.93351e-07,2.9075e-08,1.10537,0.000399574,-1.06126e-07,2.90902e-09,1.10577,0.00039937,-9.73992e-08,-4.07111e-08,1.10617,0.000399053,-2.19533e-07,4.07262e-08,1.10657,0.000398736,-9.73541e-08,-2.98424e-09,1.10697,0.000398533,-1.06307e-07,-2.87892e-08,1.10736,0.000398234,-1.92674e-07,-1.06824e-09,1.10776,0.000397845,-1.95879e-07,3.30622e-08,1.10816,0.000397552,-9.66926e-08,-1.19712e-08,1.10856,0.000397323,-1.32606e-07,1.48225e-08,1.10895,0.000397102,-8.81387e-08,-4.73187e-08,1.10935,0.000396784,-2.30095e-07,5.52429e-08,1.10975,0.00039649,-6.4366e-08,-5.44437e-08,1.11014,0.000396198,-2.27697e-07,4.33226e-08,1.11054,0.000395872,-9.77293e-08,3.62656e-10,1.11094,0.000395678,-9.66414e-08,-4.47732e-08,1.11133,0.00039535,-2.30961e-07,5.95208e-08,1.11173,0.000395067,-5.23985e-08,-7.41008e-08,1.11212,0.00039474,-2.74701e-07,1.17673e-07,1.11252,0.000394543,7.83181e-08,-1.58172e-07,1.11291,0.000394225,-3.96199e-07,1.57389e-07,1.1133,0.000393905,7.59679e-08,-1.13756e-07,1.1137,0.000393716,-2.653e-07,5.92165e-08,1.11409,0.000393363,-8.76507e-08,-3.90074e-09,1.11449,0.000393176,-9.93529e-08,-4.36136e-08,1.11488,0.000392846,-2.30194e-07,5.91457e-08,1.11527,0.000392563,-5.27564e-08,-7.376e-08,1.11566,0.000392237,-2.74037e-07,1.16685e-07,1.11606,0.000392039,7.60189e-08,-1.54562e-07,1.11645,0.000391727,-3.87667e-07,1.43935e-07,1.11684,0.000391384,4.4137e-08,-6.35487e-08,1.11723,0.000391281,-1.46509e-07,-8.94896e-09,1.11762,0.000390961,-1.73356e-07,-1.98647e-08,1.11801,0.000390555,-2.3295e-07,8.8408e-08,1.1184,0.000390354,3.22736e-08,-9.53486e-08,1.11879,0.000390133,-2.53772e-07,5.45677e-08,1.11918,0.000389789,-9.0069e-08,-3.71296e-09,1.11957,0.000389598,-1.01208e-07,-3.97159e-08,1.11996,0.000389276,-2.20355e-07,4.33671e-08,1.12035,0.000388966,-9.02542e-08,-1.45431e-08,1.12074,0.000388741,-1.33883e-07,1.48052e-08,1.12113,0.000388518,-8.94678e-08,-4.46778e-08,1.12152,0.000388205,-2.23501e-07,4.46966e-08,1.12191,0.000387892,-8.94114e-08,-1.48992e-08,1.12229,0.000387669,-1.34109e-07,1.49003e-08,1.12268,0.000387445,-8.94082e-08,-4.47019e-08,1.12307,0.000387132,-2.23514e-07,4.4698e-08,1.12345,0.000386819,-8.942e-08,-1.48806e-08,1.12384,0.000386596,-1.34062e-07,1.48245e-08,1.12423,0.000386372,-8.95885e-08,-4.44172e-08,1.12461,0.00038606,-2.2284e-07,4.36351e-08,1.125,0.000385745,-9.19348e-08,-1.09139e-08,1.12539,0.000385528,-1.24677e-07,2.05584e-11,1.12577,0.000385279,-1.24615e-07,1.08317e-08,1.12616,0.000385062,-9.21198e-08,-4.33473e-08,1.12654,0.000384748,-2.22162e-07,4.33481e-08,1.12693,0.000384434,-9.21174e-08,-1.08356e-08,1.12731,0.000384217,-1.24624e-07,-5.50907e-12,1.12769,0.000383968,-1.24641e-07,1.08577e-08,1.12808,0.000383751,-9.20679e-08,-4.34252e-08,1.12846,0.000383437,-2.22343e-07,4.36337e-08,1.12884,0.000383123,-9.14422e-08,-1.19005e-08,1.12923,0.000382904,-1.27144e-07,3.96813e-09,1.12961,0.000382662,-1.15239e-07,-3.97207e-09,1.12999,0.000382419,-1.27155e-07,1.19201e-08,1.13038,0.000382201,-9.1395e-08,-4.37085e-08,1.13076,0.000381887,-2.2252e-07,4.37046e-08,1.13114,0.000381573,-9.14068e-08,-1.19005e-08,1.13152,0.000381355,-1.27108e-07,3.89734e-09,1.1319,0.000381112,-1.15416e-07,-3.68887e-09,1.13228,0.00038087,-1.26483e-07,1.08582e-08,1.13266,0.00038065,-9.39083e-08,-3.97438e-08,1.13304,0.000380343,-2.1314e-07,2.89076e-08,1.13342,0.000380003,-1.26417e-07,4.33225e-08,1.1338,0.00037988,3.55072e-09,-8.29883e-08,1.13418,0.000379638,-2.45414e-07,5.0212e-08,1.13456,0.000379298,-9.47781e-08,1.34964e-09,1.13494,0.000379113,-9.07292e-08,-5.56105e-08,1.13532,0.000378764,-2.57561e-07,1.01883e-07,1.1357,0.000378555,4.80889e-08,-1.13504e-07,1.13608,0.000378311,-2.92423e-07,1.13713e-07,1.13646,0.000378067,4.87176e-08,-1.02931e-07,1.13683,0.000377856,-2.60076e-07,5.95923e-08,1.13721,0.000377514,-8.12988e-08,-1.62288e-08,1.13759,0.000377303,-1.29985e-07,5.32278e-09,1.13797,0.000377059,-1.14017e-07,-5.06237e-09,1.13834,0.000376816,-1.29204e-07,1.49267e-08,1.13872,0.000376602,-8.44237e-08,-5.46444e-08,1.1391,0.000376269,-2.48357e-07,8.44417e-08,1.13947,0.000376026,4.96815e-09,-4.47039e-08,1.13985,0.000375902,-1.29143e-07,-2.48355e-08,1.14023,0.000375569,-2.0365e-07,2.48368e-08,1.1406,0.000375236,-1.2914e-07,4.46977e-08,1.14098,0.000375112,4.95341e-09,-8.44184e-08,1.14135,0.000374869,-2.48302e-07,5.45572e-08,1.14173,0.000374536,-8.463e-08,-1.46013e-08,1.1421,0.000374323,-1.28434e-07,3.8478e-09,1.14247,0.000374077,-1.1689e-07,-7.89941e-10,1.14285,0.000373841,-1.1926e-07,-6.88042e-10,1.14322,0.0003736,-1.21324e-07,3.54213e-09,1.1436,0.000373368,-1.10698e-07,-1.34805e-08,1.14397,0.000373107,-1.51139e-07,5.03798e-08,1.14434,0.000372767,0.,0.};
-
-        template <bool srgb, int blueIdx, typename T, typename D>
-        __device__ __forceinline__ void RGB2LuvConvert_f(const T& src, D& dst)
-        {
-            const float _d = 1.f / (0.950456f + 15 + 1.088754f * 3);
-            const float _un = 13 * (4 * 0.950456f * _d);
-            const float _vn = 13 * (9 * _d);
-
-            float B = blueIdx == 0 ? src.x : src.z;
-            float G = src.y;
-            float R = blueIdx == 0 ? src.z : src.x;
-
-            if (srgb)
-            {
-                B = splineInterpolate(B * GAMMA_TAB_SIZE, c_sRGBGammaTab, GAMMA_TAB_SIZE);
-                G = splineInterpolate(G * GAMMA_TAB_SIZE, c_sRGBGammaTab, GAMMA_TAB_SIZE);
-                R = splineInterpolate(R * GAMMA_TAB_SIZE, c_sRGBGammaTab, GAMMA_TAB_SIZE);
-            }
-
-            float X = R * 0.412453f + G * 0.357580f + B * 0.180423f;
-            float Y = R * 0.212671f + G * 0.715160f + B * 0.072169f;
-            float Z = R * 0.019334f + G * 0.119193f + B * 0.950227f;
-
-            float L = splineInterpolate(Y * (LAB_CBRT_TAB_SIZE / 1.5f), c_LabCbrtTab, LAB_CBRT_TAB_SIZE);
-            L = 116.f * L - 16.f;
-
-            const float d = (4 * 13) / ::fmaxf(X + 15 * Y + 3 * Z, numeric_limits<float>::epsilon());
-            float u = L * (X * d - _un);
-            float v = L * ((9 * 0.25f) * Y * d - _vn);
-
-            dst.x = L;
-            dst.y = u;
-            dst.z = v;
-        }
-
-        template <bool srgb, int blueIdx, typename T, typename D>
-        __device__ __forceinline__ void RGB2LuvConvert_b(const T& src, D& dst)
-        {
-            float3 srcf, dstf;
-
-            srcf.x = src.x * (1.f / 255.f);
-            srcf.y = src.y * (1.f / 255.f);
-            srcf.z = src.z * (1.f / 255.f);
-
-            RGB2LuvConvert_f<srgb, blueIdx>(srcf, dstf);
-
-            dst.x = saturate_cast<uchar>(dstf.x * 2.55f);
-            dst.y = saturate_cast<uchar>(dstf.y * 0.72033898305084743f + 96.525423728813564f);
-            dst.z = saturate_cast<uchar>(dstf.z * 0.9732824427480916f + 136.259541984732824f);
-        }
-
-        template <typename T, int scn, int dcn, bool srgb, int blueIdx> struct RGB2Luv;
-        template <int scn, int dcn, bool srgb, int blueIdx>
-        struct RGB2Luv<uchar, scn, dcn, srgb, blueIdx>
-            : unary_function<typename TypeVec<uchar, scn>::vec_type, typename TypeVec<uchar, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<uchar, dcn>::vec_type operator ()(const typename TypeVec<uchar, scn>::vec_type& src) const
-            {
-                typename TypeVec<uchar, dcn>::vec_type dst;
-
-                RGB2LuvConvert_b<srgb, blueIdx>(src, dst);
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ RGB2Luv() {}
-            __host__ __device__ __forceinline__ RGB2Luv(const RGB2Luv&) {}
-        };
-        template <int scn, int dcn, bool srgb, int blueIdx>
-        struct RGB2Luv<float, scn, dcn, srgb, blueIdx>
-            : unary_function<typename TypeVec<float, scn>::vec_type, typename TypeVec<float, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<float, dcn>::vec_type operator ()(const typename TypeVec<float, scn>::vec_type& src) const
-            {
-                typename TypeVec<float, dcn>::vec_type dst;
-
-                RGB2LuvConvert_f<srgb, blueIdx>(src, dst);
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ RGB2Luv() {}
-            __host__ __device__ __forceinline__ RGB2Luv(const RGB2Luv&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_RGB2Luv_TRAITS(name, scn, dcn, srgb, blueIdx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::RGB2Luv<T, scn, dcn, srgb, blueIdx> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-    namespace color_detail
-    {
-        template <bool srgb, int blueIdx, typename T, typename D>
-        __device__ __forceinline__ void Luv2RGBConvert_f(const T& src, D& dst)
-        {
-            const float _d = 1.f / (0.950456f + 15 + 1.088754f * 3);
-            const float _un = 4 * 0.950456f * _d;
-            const float _vn = 9 * _d;
-
-            float L = src.x;
-            float u = src.y;
-            float v = src.z;
-
-            float Y = (L + 16.f) * (1.f / 116.f);
-            Y = Y * Y * Y;
-
-            float d = (1.f / 13.f) / L;
-            u = u * d + _un;
-            v = v * d + _vn;
-
-            float iv = 1.f / v;
-            float X = 2.25f * u * Y * iv;
-            float Z = (12 - 3 * u - 20 * v) * Y * 0.25f * iv;
-
-            float B = 0.055648f * X - 0.204043f * Y + 1.057311f * Z;
-            float G = -0.969256f * X + 1.875991f * Y + 0.041556f * Z;
-            float R = 3.240479f * X - 1.537150f * Y - 0.498535f * Z;
-
-            if (srgb)
-            {
-                B = splineInterpolate(B * GAMMA_TAB_SIZE, c_sRGBInvGammaTab, GAMMA_TAB_SIZE);
-                G = splineInterpolate(G * GAMMA_TAB_SIZE, c_sRGBInvGammaTab, GAMMA_TAB_SIZE);
-                R = splineInterpolate(R * GAMMA_TAB_SIZE, c_sRGBInvGammaTab, GAMMA_TAB_SIZE);
-            }
-
-            dst.x = blueIdx == 0 ? B : R;
-            dst.y = G;
-            dst.z = blueIdx == 0 ? R : B;
-            setAlpha(dst, ColorChannel<float>::max());
-        }
-
-        template <bool srgb, int blueIdx, typename T, typename D>
-        __device__ __forceinline__ void Luv2RGBConvert_b(const T& src, D& dst)
-        {
-            float3 srcf, dstf;
-
-            srcf.x = src.x * (100.f / 255.f);
-            srcf.y = src.y * 1.388235294117647f - 134.f;
-            srcf.z = src.z * 1.027450980392157f - 140.f;
-
-            Luv2RGBConvert_f<srgb, blueIdx>(srcf, dstf);
-
-            dst.x = saturate_cast<uchar>(dstf.x * 255.f);
-            dst.y = saturate_cast<uchar>(dstf.y * 255.f);
-            dst.z = saturate_cast<uchar>(dstf.z * 255.f);
-            setAlpha(dst, ColorChannel<uchar>::max());
-        }
-
-        template <typename T, int scn, int dcn, bool srgb, int blueIdx> struct Luv2RGB;
-        template <int scn, int dcn, bool srgb, int blueIdx>
-        struct Luv2RGB<uchar, scn, dcn, srgb, blueIdx>
-            : unary_function<typename TypeVec<uchar, scn>::vec_type, typename TypeVec<uchar, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<uchar, dcn>::vec_type operator ()(const typename TypeVec<uchar, scn>::vec_type& src) const
-            {
-                typename TypeVec<uchar, dcn>::vec_type dst;
-
-                Luv2RGBConvert_b<srgb, blueIdx>(src, dst);
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ Luv2RGB() {}
-            __host__ __device__ __forceinline__ Luv2RGB(const Luv2RGB&) {}
-        };
-        template <int scn, int dcn, bool srgb, int blueIdx>
-        struct Luv2RGB<float, scn, dcn, srgb, blueIdx>
-            : unary_function<typename TypeVec<float, scn>::vec_type, typename TypeVec<float, dcn>::vec_type>
-        {
-            __device__ __forceinline__ typename TypeVec<float, dcn>::vec_type operator ()(const typename TypeVec<float, scn>::vec_type& src) const
-            {
-                typename TypeVec<float, dcn>::vec_type dst;
-
-                Luv2RGBConvert_f<srgb, blueIdx>(src, dst);
-
-                return dst;
-            }
-            __host__ __device__ __forceinline__ Luv2RGB() {}
-            __host__ __device__ __forceinline__ Luv2RGB(const Luv2RGB&) {}
-        };
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_Luv2RGB_TRAITS(name, scn, dcn, srgb, blueIdx) \
-    template <typename T> struct name ## _traits \
-    { \
-        typedef ::cv::cuda::device::color_detail::Luv2RGB<T, scn, dcn, srgb, blueIdx> functor_type; \
-        static __host__ __device__ __forceinline__ functor_type create_functor() \
-        { \
-            return functor_type(); \
-        } \
-    };
-
-    #undef CV_DESCALE
-
-}}} // namespace cv { namespace cuda { namespace cudev
-
-//! @endcond
-
-#endif // OPENCV_CUDA_COLOR_DETAIL_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/reduce.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/reduce.hpp
deleted file mode 100644
index 8af20b0d..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/reduce.hpp
+++ /dev/null
@@ -1,365 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_REDUCE_DETAIL_HPP
-#define OPENCV_CUDA_REDUCE_DETAIL_HPP
-
-#include <thrust/tuple.h>
-#include "../warp.hpp"
-#include "../warp_shuffle.hpp"
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    namespace reduce_detail
-    {
-        template <typename T> struct GetType;
-        template <typename T> struct GetType<T*>
-        {
-            typedef T type;
-        };
-        template <typename T> struct GetType<volatile T*>
-        {
-            typedef T type;
-        };
-        template <typename T> struct GetType<T&>
-        {
-            typedef T type;
-        };
-
-        template <unsigned int I, unsigned int N>
-        struct For
-        {
-            template <class PointerTuple, class ValTuple>
-            static __device__ void loadToSmem(const PointerTuple& smem, const ValTuple& val, unsigned int tid)
-            {
-                thrust::get<I>(smem)[tid] = thrust::get<I>(val);
-
-                For<I + 1, N>::loadToSmem(smem, val, tid);
-            }
-            template <class PointerTuple, class ValTuple>
-            static __device__ void loadFromSmem(const PointerTuple& smem, const ValTuple& val, unsigned int tid)
-            {
-                thrust::get<I>(val) = thrust::get<I>(smem)[tid];
-
-                For<I + 1, N>::loadFromSmem(smem, val, tid);
-            }
-
-            template <class PointerTuple, class ValTuple, class OpTuple>
-            static __device__ void merge(const PointerTuple& smem, const ValTuple& val, unsigned int tid, unsigned int delta, const OpTuple& op)
-            {
-                typename GetType<typename thrust::tuple_element<I, PointerTuple>::type>::type reg = thrust::get<I>(smem)[tid + delta];
-                thrust::get<I>(smem)[tid] = thrust::get<I>(val) = thrust::get<I>(op)(thrust::get<I>(val), reg);
-
-                For<I + 1, N>::merge(smem, val, tid, delta, op);
-            }
-            template <class ValTuple, class OpTuple>
-            static __device__ void mergeShfl(const ValTuple& val, unsigned int delta, unsigned int width, const OpTuple& op)
-            {
-                typename GetType<typename thrust::tuple_element<I, ValTuple>::type>::type reg = shfl_down(thrust::get<I>(val), delta, width);
-                thrust::get<I>(val) = thrust::get<I>(op)(thrust::get<I>(val), reg);
-
-                For<I + 1, N>::mergeShfl(val, delta, width, op);
-            }
-        };
-        template <unsigned int N>
-        struct For<N, N>
-        {
-            template <class PointerTuple, class ValTuple>
-            static __device__ void loadToSmem(const PointerTuple&, const ValTuple&, unsigned int)
-            {
-            }
-            template <class PointerTuple, class ValTuple>
-            static __device__ void loadFromSmem(const PointerTuple&, const ValTuple&, unsigned int)
-            {
-            }
-
-            template <class PointerTuple, class ValTuple, class OpTuple>
-            static __device__ void merge(const PointerTuple&, const ValTuple&, unsigned int, unsigned int, const OpTuple&)
-            {
-            }
-            template <class ValTuple, class OpTuple>
-            static __device__ void mergeShfl(const ValTuple&, unsigned int, unsigned int, const OpTuple&)
-            {
-            }
-        };
-
-        template <typename T>
-        __device__ __forceinline__ void loadToSmem(volatile T* smem, T& val, unsigned int tid)
-        {
-            smem[tid] = val;
-        }
-        template <typename T>
-        __device__ __forceinline__ void loadFromSmem(volatile T* smem, T& val, unsigned int tid)
-        {
-            val = smem[tid];
-        }
-        template <typename P0, typename P1, typename P2, typename P3, typename P4, typename P5, typename P6, typename P7, typename P8, typename P9,
-                  typename R0, typename R1, typename R2, typename R3, typename R4, typename R5, typename R6, typename R7, typename R8, typename R9>
-        __device__ __forceinline__ void loadToSmem(const thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>& smem,
-                                                       const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>& val,
-                                                       unsigned int tid)
-        {
-            For<0, thrust::tuple_size<thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9> >::value>::loadToSmem(smem, val, tid);
-        }
-        template <typename P0, typename P1, typename P2, typename P3, typename P4, typename P5, typename P6, typename P7, typename P8, typename P9,
-                  typename R0, typename R1, typename R2, typename R3, typename R4, typename R5, typename R6, typename R7, typename R8, typename R9>
-        __device__ __forceinline__ void loadFromSmem(const thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>& smem,
-                                                         const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>& val,
-                                                         unsigned int tid)
-        {
-            For<0, thrust::tuple_size<thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9> >::value>::loadFromSmem(smem, val, tid);
-        }
-
-        template <typename T, class Op>
-        __device__ __forceinline__ void merge(volatile T* smem, T& val, unsigned int tid, unsigned int delta, const Op& op)
-        {
-            T reg = smem[tid + delta];
-            smem[tid] = val = op(val, reg);
-        }
-        template <typename T, class Op>
-        __device__ __forceinline__ void mergeShfl(T& val, unsigned int delta, unsigned int width, const Op& op)
-        {
-            T reg = shfl_down(val, delta, width);
-            val = op(val, reg);
-        }
-        template <typename P0, typename P1, typename P2, typename P3, typename P4, typename P5, typename P6, typename P7, typename P8, typename P9,
-                  typename R0, typename R1, typename R2, typename R3, typename R4, typename R5, typename R6, typename R7, typename R8, typename R9,
-                  class Op0, class Op1, class Op2, class Op3, class Op4, class Op5, class Op6, class Op7, class Op8, class Op9>
-        __device__ __forceinline__ void merge(const thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>& smem,
-                                              const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>& val,
-                                              unsigned int tid,
-                                              unsigned int delta,
-                                              const thrust::tuple<Op0, Op1, Op2, Op3, Op4, Op5, Op6, Op7, Op8, Op9>& op)
-        {
-            For<0, thrust::tuple_size<thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9> >::value>::merge(smem, val, tid, delta, op);
-        }
-        template <typename R0, typename R1, typename R2, typename R3, typename R4, typename R5, typename R6, typename R7, typename R8, typename R9,
-                  class Op0, class Op1, class Op2, class Op3, class Op4, class Op5, class Op6, class Op7, class Op8, class Op9>
-        __device__ __forceinline__ void mergeShfl(const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>& val,
-                                                  unsigned int delta,
-                                                  unsigned int width,
-                                                  const thrust::tuple<Op0, Op1, Op2, Op3, Op4, Op5, Op6, Op7, Op8, Op9>& op)
-        {
-            For<0, thrust::tuple_size<thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9> >::value>::mergeShfl(val, delta, width, op);
-        }
-
-        template <unsigned int N> struct Generic
-        {
-            template <typename Pointer, typename Reference, class Op>
-            static __device__ void reduce(Pointer smem, Reference val, unsigned int tid, Op op)
-            {
-                loadToSmem(smem, val, tid);
-                if (N >= 32)
-                    __syncthreads();
-
-                if (N >= 2048)
-                {
-                    if (tid < 1024)
-                        merge(smem, val, tid, 1024, op);
-
-                    __syncthreads();
-                }
-                if (N >= 1024)
-                {
-                    if (tid < 512)
-                        merge(smem, val, tid, 512, op);
-
-                    __syncthreads();
-                }
-                if (N >= 512)
-                {
-                    if (tid < 256)
-                        merge(smem, val, tid, 256, op);
-
-                    __syncthreads();
-                }
-                if (N >= 256)
-                {
-                    if (tid < 128)
-                        merge(smem, val, tid, 128, op);
-
-                    __syncthreads();
-                }
-                if (N >= 128)
-                {
-                    if (tid < 64)
-                        merge(smem, val, tid, 64, op);
-
-                    __syncthreads();
-                }
-                if (N >= 64)
-                {
-                    if (tid < 32)
-                        merge(smem, val, tid, 32, op);
-                }
-
-                if (tid < 16)
-                {
-                    merge(smem, val, tid, 16, op);
-                    merge(smem, val, tid, 8, op);
-                    merge(smem, val, tid, 4, op);
-                    merge(smem, val, tid, 2, op);
-                    merge(smem, val, tid, 1, op);
-                }
-            }
-        };
-
-        template <unsigned int I, typename Pointer, typename Reference, class Op>
-        struct Unroll
-        {
-            static __device__ void loopShfl(Reference val, Op op, unsigned int N)
-            {
-                mergeShfl(val, I, N, op);
-                Unroll<I / 2, Pointer, Reference, Op>::loopShfl(val, op, N);
-            }
-            static __device__ void loop(Pointer smem, Reference val, unsigned int tid, Op op)
-            {
-                merge(smem, val, tid, I, op);
-                Unroll<I / 2, Pointer, Reference, Op>::loop(smem, val, tid, op);
-            }
-        };
-        template <typename Pointer, typename Reference, class Op>
-        struct Unroll<0, Pointer, Reference, Op>
-        {
-            static __device__ void loopShfl(Reference, Op, unsigned int)
-            {
-            }
-            static __device__ void loop(Pointer, Reference, unsigned int, Op)
-            {
-            }
-        };
-
-        template <unsigned int N> struct WarpOptimized
-        {
-            template <typename Pointer, typename Reference, class Op>
-            static __device__ void reduce(Pointer smem, Reference val, unsigned int tid, Op op)
-            {
-            #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
-                CV_UNUSED(smem);
-                CV_UNUSED(tid);
-
-                Unroll<N / 2, Pointer, Reference, Op>::loopShfl(val, op, N);
-            #else
-                loadToSmem(smem, val, tid);
-
-                if (tid < N / 2)
-                    Unroll<N / 2, Pointer, Reference, Op>::loop(smem, val, tid, op);
-            #endif
-            }
-        };
-
-        template <unsigned int N> struct GenericOptimized32
-        {
-            enum { M = N / 32 };
-
-            template <typename Pointer, typename Reference, class Op>
-            static __device__ void reduce(Pointer smem, Reference val, unsigned int tid, Op op)
-            {
-                const unsigned int laneId = Warp::laneId();
-
-            #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
-                Unroll<16, Pointer, Reference, Op>::loopShfl(val, op, warpSize);
-
-                if (laneId == 0)
-                    loadToSmem(smem, val, tid / 32);
-            #else
-                loadToSmem(smem, val, tid);
-
-                if (laneId < 16)
-                    Unroll<16, Pointer, Reference, Op>::loop(smem, val, tid, op);
-
-                __syncthreads();
-
-                if (laneId == 0)
-                    loadToSmem(smem, val, tid / 32);
-            #endif
-
-                __syncthreads();
-
-                loadFromSmem(smem, val, tid);
-
-                if (tid < 32)
-                {
-                #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
-                    Unroll<M / 2, Pointer, Reference, Op>::loopShfl(val, op, M);
-                #else
-                    Unroll<M / 2, Pointer, Reference, Op>::loop(smem, val, tid, op);
-                #endif
-                }
-            }
-        };
-
-        template <bool val, class T1, class T2> struct StaticIf;
-        template <class T1, class T2> struct StaticIf<true, T1, T2>
-        {
-            typedef T1 type;
-        };
-        template <class T1, class T2> struct StaticIf<false, T1, T2>
-        {
-            typedef T2 type;
-        };
-
-        template <unsigned int N> struct IsPowerOf2
-        {
-            enum { value = ((N != 0) && !(N & (N - 1))) };
-        };
-
-        template <unsigned int N> struct Dispatcher
-        {
-            typedef typename StaticIf<
-                (N <= 32) && IsPowerOf2<N>::value,
-                WarpOptimized<N>,
-                typename StaticIf<
-                    (N <= 1024) && IsPowerOf2<N>::value,
-                    GenericOptimized32<N>,
-                    Generic<N>
-                >::type
-            >::type reductor;
-        };
-    }
-}}}
-
-//! @endcond
-
-#endif // OPENCV_CUDA_REDUCE_DETAIL_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/reduce_key_val.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/reduce_key_val.hpp
deleted file mode 100644
index df37c173..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/reduce_key_val.hpp
+++ /dev/null
@@ -1,502 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_PRED_VAL_REDUCE_DETAIL_HPP
-#define OPENCV_CUDA_PRED_VAL_REDUCE_DETAIL_HPP
-
-#include <thrust/tuple.h>
-#include "../warp.hpp"
-#include "../warp_shuffle.hpp"
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    namespace reduce_key_val_detail
-    {
-        template <typename T> struct GetType;
-        template <typename T> struct GetType<T*>
-        {
-            typedef T type;
-        };
-        template <typename T> struct GetType<volatile T*>
-        {
-            typedef T type;
-        };
-        template <typename T> struct GetType<T&>
-        {
-            typedef T type;
-        };
-
-        template <unsigned int I, unsigned int N>
-        struct For
-        {
-            template <class PointerTuple, class ReferenceTuple>
-            static __device__ void loadToSmem(const PointerTuple& smem, const ReferenceTuple& data, unsigned int tid)
-            {
-                thrust::get<I>(smem)[tid] = thrust::get<I>(data);
-
-                For<I + 1, N>::loadToSmem(smem, data, tid);
-            }
-            template <class PointerTuple, class ReferenceTuple>
-            static __device__ void loadFromSmem(const PointerTuple& smem, const ReferenceTuple& data, unsigned int tid)
-            {
-                thrust::get<I>(data) = thrust::get<I>(smem)[tid];
-
-                For<I + 1, N>::loadFromSmem(smem, data, tid);
-            }
-
-            template <class ReferenceTuple>
-            static __device__ void copyShfl(const ReferenceTuple& val, unsigned int delta, int width)
-            {
-                thrust::get<I>(val) = shfl_down(thrust::get<I>(val), delta, width);
-
-                For<I + 1, N>::copyShfl(val, delta, width);
-            }
-            template <class PointerTuple, class ReferenceTuple>
-            static __device__ void copy(const PointerTuple& svals, const ReferenceTuple& val, unsigned int tid, unsigned int delta)
-            {
-                thrust::get<I>(svals)[tid] = thrust::get<I>(val) = thrust::get<I>(svals)[tid + delta];
-
-                For<I + 1, N>::copy(svals, val, tid, delta);
-            }
-
-            template <class KeyReferenceTuple, class ValReferenceTuple, class CmpTuple>
-            static __device__ void mergeShfl(const KeyReferenceTuple& key, const ValReferenceTuple& val, const CmpTuple& cmp, unsigned int delta, int width)
-            {
-                typename GetType<typename thrust::tuple_element<I, KeyReferenceTuple>::type>::type reg = shfl_down(thrust::get<I>(key), delta, width);
-
-                if (thrust::get<I>(cmp)(reg, thrust::get<I>(key)))
-                {
-                    thrust::get<I>(key) = reg;
-                    thrust::get<I>(val) = shfl_down(thrust::get<I>(val), delta, width);
-                }
-
-                For<I + 1, N>::mergeShfl(key, val, cmp, delta, width);
-            }
-            template <class KeyPointerTuple, class KeyReferenceTuple, class ValPointerTuple, class ValReferenceTuple, class CmpTuple>
-            static __device__ void merge(const KeyPointerTuple& skeys, const KeyReferenceTuple& key,
-                                         const ValPointerTuple& svals, const ValReferenceTuple& val,
-                                         const CmpTuple& cmp,
-                                         unsigned int tid, unsigned int delta)
-            {
-                typename GetType<typename thrust::tuple_element<I, KeyPointerTuple>::type>::type reg = thrust::get<I>(skeys)[tid + delta];
-
-                if (thrust::get<I>(cmp)(reg, thrust::get<I>(key)))
-                {
-                    thrust::get<I>(skeys)[tid] = thrust::get<I>(key) = reg;
-                    thrust::get<I>(svals)[tid] = thrust::get<I>(val) = thrust::get<I>(svals)[tid + delta];
-                }
-
-                For<I + 1, N>::merge(skeys, key, svals, val, cmp, tid, delta);
-            }
-        };
-        template <unsigned int N>
-        struct For<N, N>
-        {
-            template <class PointerTuple, class ReferenceTuple>
-            static __device__ void loadToSmem(const PointerTuple&, const ReferenceTuple&, unsigned int)
-            {
-            }
-            template <class PointerTuple, class ReferenceTuple>
-            static __device__ void loadFromSmem(const PointerTuple&, const ReferenceTuple&, unsigned int)
-            {
-            }
-
-            template <class ReferenceTuple>
-            static __device__ void copyShfl(const ReferenceTuple&, unsigned int, int)
-            {
-            }
-            template <class PointerTuple, class ReferenceTuple>
-            static __device__ void copy(const PointerTuple&, const ReferenceTuple&, unsigned int, unsigned int)
-            {
-            }
-
-            template <class KeyReferenceTuple, class ValReferenceTuple, class CmpTuple>
-            static __device__ void mergeShfl(const KeyReferenceTuple&, const ValReferenceTuple&, const CmpTuple&, unsigned int, int)
-            {
-            }
-            template <class KeyPointerTuple, class KeyReferenceTuple, class ValPointerTuple, class ValReferenceTuple, class CmpTuple>
-            static __device__ void merge(const KeyPointerTuple&, const KeyReferenceTuple&,
-                                         const ValPointerTuple&, const ValReferenceTuple&,
-                                         const CmpTuple&,
-                                         unsigned int, unsigned int)
-            {
-            }
-        };
-
-        //////////////////////////////////////////////////////
-        // loadToSmem
-
-        template <typename T>
-        __device__ __forceinline__ void loadToSmem(volatile T* smem, T& data, unsigned int tid)
-        {
-            smem[tid] = data;
-        }
-        template <typename T>
-        __device__ __forceinline__ void loadFromSmem(volatile T* smem, T& data, unsigned int tid)
-        {
-            data = smem[tid];
-        }
-        template <typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
-                  typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9>
-        __device__ __forceinline__ void loadToSmem(const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& smem,
-                                                   const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& data,
-                                                   unsigned int tid)
-        {
-            For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::loadToSmem(smem, data, tid);
-        }
-        template <typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
-                  typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9>
-        __device__ __forceinline__ void loadFromSmem(const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& smem,
-                                                     const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& data,
-                                                     unsigned int tid)
-        {
-            For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::loadFromSmem(smem, data, tid);
-        }
-
-        //////////////////////////////////////////////////////
-        // copyVals
-
-        template <typename V>
-        __device__ __forceinline__ void copyValsShfl(V& val, unsigned int delta, int width)
-        {
-            val = shfl_down(val, delta, width);
-        }
-        template <typename V>
-        __device__ __forceinline__ void copyVals(volatile V* svals, V& val, unsigned int tid, unsigned int delta)
-        {
-            svals[tid] = val = svals[tid + delta];
-        }
-        template <typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9>
-        __device__ __forceinline__ void copyValsShfl(const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
-                                                     unsigned int delta,
-                                                     int width)
-        {
-            For<0, thrust::tuple_size<thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9> >::value>::copyShfl(val, delta, width);
-        }
-        template <typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
-                  typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9>
-        __device__ __forceinline__ void copyVals(const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals,
-                                                 const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
-                                                 unsigned int tid, unsigned int delta)
-        {
-            For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::copy(svals, val, tid, delta);
-        }
-
-        //////////////////////////////////////////////////////
-        // merge
-
-        template <typename K, typename V, class Cmp>
-        __device__ __forceinline__ void mergeShfl(K& key, V& val, const Cmp& cmp, unsigned int delta, int width)
-        {
-            K reg = shfl_down(key, delta, width);
-
-            if (cmp(reg, key))
-            {
-                key = reg;
-                copyValsShfl(val, delta, width);
-            }
-        }
-        template <typename K, typename V, class Cmp>
-        __device__ __forceinline__ void merge(volatile K* skeys, K& key, volatile V* svals, V& val, const Cmp& cmp, unsigned int tid, unsigned int delta)
-        {
-            K reg = skeys[tid + delta];
-
-            if (cmp(reg, key))
-            {
-                skeys[tid] = key = reg;
-                copyVals(svals, val, tid, delta);
-            }
-        }
-        template <typename K,
-                  typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
-                  class Cmp>
-        __device__ __forceinline__ void mergeShfl(K& key,
-                                                  const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
-                                                  const Cmp& cmp,
-                                                  unsigned int delta, int width)
-        {
-            K reg = shfl_down(key, delta, width);
-
-            if (cmp(reg, key))
-            {
-                key = reg;
-                copyValsShfl(val, delta, width);
-            }
-        }
-        template <typename K,
-                  typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
-                  typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
-                  class Cmp>
-        __device__ __forceinline__ void merge(volatile K* skeys, K& key,
-                                              const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals,
-                                              const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
-                                              const Cmp& cmp, unsigned int tid, unsigned int delta)
-        {
-            K reg = skeys[tid + delta];
-
-            if (cmp(reg, key))
-            {
-                skeys[tid] = key = reg;
-                copyVals(svals, val, tid, delta);
-            }
-        }
-        template <typename KR0, typename KR1, typename KR2, typename KR3, typename KR4, typename KR5, typename KR6, typename KR7, typename KR8, typename KR9,
-                  typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
-                  class Cmp0, class Cmp1, class Cmp2, class Cmp3, class Cmp4, class Cmp5, class Cmp6, class Cmp7, class Cmp8, class Cmp9>
-        __device__ __forceinline__ void mergeShfl(const thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9>& key,
-                                                  const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
-                                                  const thrust::tuple<Cmp0, Cmp1, Cmp2, Cmp3, Cmp4, Cmp5, Cmp6, Cmp7, Cmp8, Cmp9>& cmp,
-                                                  unsigned int delta, int width)
-        {
-            For<0, thrust::tuple_size<thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9> >::value>::mergeShfl(key, val, cmp, delta, width);
-        }
-        template <typename KP0, typename KP1, typename KP2, typename KP3, typename KP4, typename KP5, typename KP6, typename KP7, typename KP8, typename KP9,
-                  typename KR0, typename KR1, typename KR2, typename KR3, typename KR4, typename KR5, typename KR6, typename KR7, typename KR8, typename KR9,
-                  typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
-                  typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
-                  class Cmp0, class Cmp1, class Cmp2, class Cmp3, class Cmp4, class Cmp5, class Cmp6, class Cmp7, class Cmp8, class Cmp9>
-        __device__ __forceinline__ void merge(const thrust::tuple<KP0, KP1, KP2, KP3, KP4, KP5, KP6, KP7, KP8, KP9>& skeys,
-                                              const thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9>& key,
-                                              const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals,
-                                              const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
-                                              const thrust::tuple<Cmp0, Cmp1, Cmp2, Cmp3, Cmp4, Cmp5, Cmp6, Cmp7, Cmp8, Cmp9>& cmp,
-                                              unsigned int tid, unsigned int delta)
-        {
-            For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::merge(skeys, key, svals, val, cmp, tid, delta);
-        }
-
-        //////////////////////////////////////////////////////
-        // Generic
-
-        template <unsigned int N> struct Generic
-        {
-            template <class KP, class KR, class VP, class VR, class Cmp>
-            static __device__ void reduce(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp)
-            {
-                loadToSmem(skeys, key, tid);
-                loadValsToSmem(svals, val, tid);
-                if (N >= 32)
-                    __syncthreads();
-
-                if (N >= 2048)
-                {
-                    if (tid < 1024)
-                        merge(skeys, key, svals, val, cmp, tid, 1024);
-
-                    __syncthreads();
-                }
-                if (N >= 1024)
-                {
-                    if (tid < 512)
-                        merge(skeys, key, svals, val, cmp, tid, 512);
-
-                    __syncthreads();
-                }
-                if (N >= 512)
-                {
-                    if (tid < 256)
-                        merge(skeys, key, svals, val, cmp, tid, 256);
-
-                    __syncthreads();
-                }
-                if (N >= 256)
-                {
-                    if (tid < 128)
-                        merge(skeys, key, svals, val, cmp, tid, 128);
-
-                    __syncthreads();
-                }
-                if (N >= 128)
-                {
-                    if (tid < 64)
-                        merge(skeys, key, svals, val, cmp, tid, 64);
-
-                    __syncthreads();
-                }
-                if (N >= 64)
-                {
-                    if (tid < 32)
-                        merge(skeys, key, svals, val, cmp, tid, 32);
-                }
-
-                if (tid < 16)
-                {
-                    merge(skeys, key, svals, val, cmp, tid, 16);
-                    merge(skeys, key, svals, val, cmp, tid, 8);
-                    merge(skeys, key, svals, val, cmp, tid, 4);
-                    merge(skeys, key, svals, val, cmp, tid, 2);
-                    merge(skeys, key, svals, val, cmp, tid, 1);
-                }
-            }
-        };
-
-        template <unsigned int I, class KP, class KR, class VP, class VR, class Cmp>
-        struct Unroll
-        {
-            static __device__ void loopShfl(KR key, VR val, Cmp cmp, unsigned int N)
-            {
-                mergeShfl(key, val, cmp, I, N);
-                Unroll<I / 2, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, N);
-            }
-            static __device__ void loop(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp)
-            {
-                merge(skeys, key, svals, val, cmp, tid, I);
-                Unroll<I / 2, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp);
-            }
-        };
-        template <class KP, class KR, class VP, class VR, class Cmp>
-        struct Unroll<0, KP, KR, VP, VR, Cmp>
-        {
-            static __device__ void loopShfl(KR, VR, Cmp, unsigned int)
-            {
-            }
-            static __device__ void loop(KP, KR, VP, VR, unsigned int, Cmp)
-            {
-            }
-        };
-
-        template <unsigned int N> struct WarpOptimized
-        {
-            template <class KP, class KR, class VP, class VR, class Cmp>
-            static __device__ void reduce(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp)
-            {
-            #if 0 // __CUDA_ARCH__ >= 300
-                CV_UNUSED(skeys);
-                CV_UNUSED(svals);
-                CV_UNUSED(tid);
-
-                Unroll<N / 2, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, N);
-            #else
-                loadToSmem(skeys, key, tid);
-                loadToSmem(svals, val, tid);
-
-                if (tid < N / 2)
-                    Unroll<N / 2, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp);
-            #endif
-            }
-        };
-
-        template <unsigned int N> struct GenericOptimized32
-        {
-            enum { M = N / 32 };
-
-            template <class KP, class KR, class VP, class VR, class Cmp>
-            static __device__ void reduce(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp)
-            {
-                const unsigned int laneId = Warp::laneId();
-
-            #if 0 // __CUDA_ARCH__ >= 300
-                Unroll<16, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, warpSize);
-
-                if (laneId == 0)
-                {
-                    loadToSmem(skeys, key, tid / 32);
-                    loadToSmem(svals, val, tid / 32);
-                }
-            #else
-                loadToSmem(skeys, key, tid);
-                loadToSmem(svals, val, tid);
-
-                if (laneId < 16)
-                    Unroll<16, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp);
-
-                __syncthreads();
-
-                if (laneId == 0)
-                {
-                    loadToSmem(skeys, key, tid / 32);
-                    loadToSmem(svals, val, tid / 32);
-                }
-            #endif
-
-                __syncthreads();
-
-                loadFromSmem(skeys, key, tid);
-
-                if (tid < 32)
-                {
-                #if 0 // __CUDA_ARCH__ >= 300
-                    loadFromSmem(svals, val, tid);
-
-                    Unroll<M / 2, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, M);
-                #else
-                    Unroll<M / 2, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp);
-                #endif
-                }
-            }
-        };
-
-        template <bool val, class T1, class T2> struct StaticIf;
-        template <class T1, class T2> struct StaticIf<true, T1, T2>
-        {
-            typedef T1 type;
-        };
-        template <class T1, class T2> struct StaticIf<false, T1, T2>
-        {
-            typedef T2 type;
-        };
-
-        template <unsigned int N> struct IsPowerOf2
-        {
-            enum { value = ((N != 0) && !(N & (N - 1))) };
-        };
-
-        template <unsigned int N> struct Dispatcher
-        {
-            typedef typename StaticIf<
-                (N <= 32) && IsPowerOf2<N>::value,
-                WarpOptimized<N>,
-                typename StaticIf<
-                    (N <= 1024) && IsPowerOf2<N>::value,
-                    GenericOptimized32<N>,
-                    Generic<N>
-                >::type
-            >::type reductor;
-        };
-    }
-}}}
-
-//! @endcond
-
-#endif // OPENCV_CUDA_PRED_VAL_REDUCE_DETAIL_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/transform_detail.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/transform_detail.hpp
deleted file mode 100644
index 19198488..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/transform_detail.hpp
+++ /dev/null
@@ -1,392 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_TRANSFORM_DETAIL_HPP
-#define OPENCV_CUDA_TRANSFORM_DETAIL_HPP
-
-#include "../common.hpp"
-#include "../vec_traits.hpp"
-#include "../functional.hpp"
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    namespace transform_detail
-    {
-        //! Read Write Traits
-
-        template <typename T, typename D, int shift> struct UnaryReadWriteTraits
-        {
-            typedef typename TypeVec<T, shift>::vec_type read_type;
-            typedef typename TypeVec<D, shift>::vec_type write_type;
-        };
-
-        template <typename T1, typename T2, typename D, int shift> struct BinaryReadWriteTraits
-        {
-            typedef typename TypeVec<T1, shift>::vec_type read_type1;
-            typedef typename TypeVec<T2, shift>::vec_type read_type2;
-            typedef typename TypeVec<D, shift>::vec_type write_type;
-        };
-
-        //! Transform kernels
-
-        template <int shift> struct OpUnroller;
-        template <> struct OpUnroller<1>
-        {
-            template <typename T, typename D, typename UnOp, typename Mask>
-            static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, UnOp& op, int x_shifted, int y)
-            {
-                if (mask(y, x_shifted))
-                    dst.x = op(src.x);
-            }
-
-            template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
-            static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, BinOp& op, int x_shifted, int y)
-            {
-                if (mask(y, x_shifted))
-                    dst.x = op(src1.x, src2.x);
-            }
-        };
-        template <> struct OpUnroller<2>
-        {
-            template <typename T, typename D, typename UnOp, typename Mask>
-            static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, UnOp& op, int x_shifted, int y)
-            {
-                if (mask(y, x_shifted))
-                    dst.x = op(src.x);
-                if (mask(y, x_shifted + 1))
-                    dst.y = op(src.y);
-            }
-
-            template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
-            static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, BinOp& op, int x_shifted, int y)
-            {
-                if (mask(y, x_shifted))
-                    dst.x = op(src1.x, src2.x);
-                if (mask(y, x_shifted + 1))
-                    dst.y = op(src1.y, src2.y);
-            }
-        };
-        template <> struct OpUnroller<3>
-        {
-            template <typename T, typename D, typename UnOp, typename Mask>
-            static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
-            {
-                if (mask(y, x_shifted))
-                    dst.x = op(src.x);
-                if (mask(y, x_shifted + 1))
-                    dst.y = op(src.y);
-                if (mask(y, x_shifted + 2))
-                    dst.z = op(src.z);
-            }
-
-            template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
-            static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
-            {
-                if (mask(y, x_shifted))
-                    dst.x = op(src1.x, src2.x);
-                if (mask(y, x_shifted + 1))
-                    dst.y = op(src1.y, src2.y);
-                if (mask(y, x_shifted + 2))
-                    dst.z = op(src1.z, src2.z);
-            }
-        };
-        template <> struct OpUnroller<4>
-        {
-            template <typename T, typename D, typename UnOp, typename Mask>
-            static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
-            {
-                if (mask(y, x_shifted))
-                    dst.x = op(src.x);
-                if (mask(y, x_shifted + 1))
-                    dst.y = op(src.y);
-                if (mask(y, x_shifted + 2))
-                    dst.z = op(src.z);
-                if (mask(y, x_shifted + 3))
-                    dst.w = op(src.w);
-            }
-
-            template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
-            static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
-            {
-                if (mask(y, x_shifted))
-                    dst.x = op(src1.x, src2.x);
-                if (mask(y, x_shifted + 1))
-                    dst.y = op(src1.y, src2.y);
-                if (mask(y, x_shifted + 2))
-                    dst.z = op(src1.z, src2.z);
-                if (mask(y, x_shifted + 3))
-                    dst.w = op(src1.w, src2.w);
-            }
-        };
-        template <> struct OpUnroller<8>
-        {
-            template <typename T, typename D, typename UnOp, typename Mask>
-            static __device__ __forceinline__ void unroll(const T& src, D& dst, const Mask& mask, const UnOp& op, int x_shifted, int y)
-            {
-                if (mask(y, x_shifted))
-                    dst.a0 = op(src.a0);
-                if (mask(y, x_shifted + 1))
-                    dst.a1 = op(src.a1);
-                if (mask(y, x_shifted + 2))
-                    dst.a2 = op(src.a2);
-                if (mask(y, x_shifted + 3))
-                    dst.a3 = op(src.a3);
-                if (mask(y, x_shifted + 4))
-                    dst.a4 = op(src.a4);
-                if (mask(y, x_shifted + 5))
-                    dst.a5 = op(src.a5);
-                if (mask(y, x_shifted + 6))
-                    dst.a6 = op(src.a6);
-                if (mask(y, x_shifted + 7))
-                    dst.a7 = op(src.a7);
-            }
-
-            template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
-            static __device__ __forceinline__ void unroll(const T1& src1, const T2& src2, D& dst, const Mask& mask, const BinOp& op, int x_shifted, int y)
-            {
-                if (mask(y, x_shifted))
-                    dst.a0 = op(src1.a0, src2.a0);
-                if (mask(y, x_shifted + 1))
-                    dst.a1 = op(src1.a1, src2.a1);
-                if (mask(y, x_shifted + 2))
-                    dst.a2 = op(src1.a2, src2.a2);
-                if (mask(y, x_shifted + 3))
-                    dst.a3 = op(src1.a3, src2.a3);
-                if (mask(y, x_shifted + 4))
-                    dst.a4 = op(src1.a4, src2.a4);
-                if (mask(y, x_shifted + 5))
-                    dst.a5 = op(src1.a5, src2.a5);
-                if (mask(y, x_shifted + 6))
-                    dst.a6 = op(src1.a6, src2.a6);
-                if (mask(y, x_shifted + 7))
-                    dst.a7 = op(src1.a7, src2.a7);
-            }
-        };
-
-        template <typename T, typename D, typename UnOp, typename Mask>
-        static __global__ void transformSmart(const PtrStepSz<T> src_, PtrStep<D> dst_, const Mask mask, const UnOp op)
-        {
-            typedef TransformFunctorTraits<UnOp> ft;
-            typedef typename UnaryReadWriteTraits<T, D, ft::smart_shift>::read_type read_type;
-            typedef typename UnaryReadWriteTraits<T, D, ft::smart_shift>::write_type write_type;
-
-            const int x = threadIdx.x + blockIdx.x * blockDim.x;
-            const int y = threadIdx.y + blockIdx.y * blockDim.y;
-            const int x_shifted = x * ft::smart_shift;
-
-            if (y < src_.rows)
-            {
-                const T* src = src_.ptr(y);
-                D* dst = dst_.ptr(y);
-
-                if (x_shifted + ft::smart_shift - 1 < src_.cols)
-                {
-                    const read_type src_n_el = ((const read_type*)src)[x];
-                    OpUnroller<ft::smart_shift>::unroll(src_n_el, ((write_type*)dst)[x], mask, op, x_shifted, y);
-                }
-                else
-                {
-                    for (int real_x = x_shifted; real_x < src_.cols; ++real_x)
-                    {
-                        if (mask(y, real_x))
-                            dst[real_x] = op(src[real_x]);
-                    }
-                }
-            }
-        }
-
-        template <typename T, typename D, typename UnOp, typename Mask>
-        __global__ static void transformSimple(const PtrStepSz<T> src, PtrStep<D> dst, const Mask mask, const UnOp op)
-        {
-            const int x = blockDim.x * blockIdx.x + threadIdx.x;
-            const int y = blockDim.y * blockIdx.y + threadIdx.y;
-
-            if (x < src.cols && y < src.rows && mask(y, x))
-            {
-                dst.ptr(y)[x] = op(src.ptr(y)[x]);
-            }
-        }
-
-        template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
-        static __global__ void transformSmart(const PtrStepSz<T1> src1_, const PtrStep<T2> src2_, PtrStep<D> dst_,
-            const Mask mask, const BinOp op)
-        {
-            typedef TransformFunctorTraits<BinOp> ft;
-            typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::read_type1 read_type1;
-            typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::read_type2 read_type2;
-            typedef typename BinaryReadWriteTraits<T1, T2, D, ft::smart_shift>::write_type write_type;
-
-            const int x = threadIdx.x + blockIdx.x * blockDim.x;
-            const int y = threadIdx.y + blockIdx.y * blockDim.y;
-            const int x_shifted = x * ft::smart_shift;
-
-            if (y < src1_.rows)
-            {
-                const T1* src1 = src1_.ptr(y);
-                const T2* src2 = src2_.ptr(y);
-                D* dst = dst_.ptr(y);
-
-                if (x_shifted + ft::smart_shift - 1 < src1_.cols)
-                {
-                    const read_type1 src1_n_el = ((const read_type1*)src1)[x];
-                    const read_type2 src2_n_el = ((const read_type2*)src2)[x];
-
-                    OpUnroller<ft::smart_shift>::unroll(src1_n_el, src2_n_el, ((write_type*)dst)[x], mask, op, x_shifted, y);
-                }
-                else
-                {
-                    for (int real_x = x_shifted; real_x < src1_.cols; ++real_x)
-                    {
-                        if (mask(y, real_x))
-                            dst[real_x] = op(src1[real_x], src2[real_x]);
-                    }
-                }
-            }
-        }
-
-        template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
-        static __global__ void transformSimple(const PtrStepSz<T1> src1, const PtrStep<T2> src2, PtrStep<D> dst,
-            const Mask mask, const BinOp op)
-        {
-            const int x = blockDim.x * blockIdx.x + threadIdx.x;
-            const int y = blockDim.y * blockIdx.y + threadIdx.y;
-
-            if (x < src1.cols && y < src1.rows && mask(y, x))
-            {
-                const T1 src1_data = src1.ptr(y)[x];
-                const T2 src2_data = src2.ptr(y)[x];
-                dst.ptr(y)[x] = op(src1_data, src2_data);
-            }
-        }
-
-        template <bool UseSmart> struct TransformDispatcher;
-        template<> struct TransformDispatcher<false>
-        {
-            template <typename T, typename D, typename UnOp, typename Mask>
-            static void call(PtrStepSz<T> src, PtrStepSz<D> dst, UnOp op, Mask mask, cudaStream_t stream)
-            {
-                typedef TransformFunctorTraits<UnOp> ft;
-
-                const dim3 threads(ft::simple_block_dim_x, ft::simple_block_dim_y, 1);
-                const dim3 grid(divUp(src.cols, threads.x), divUp(src.rows, threads.y), 1);
-
-                transformSimple<T, D><<<grid, threads, 0, stream>>>(src, dst, mask, op);
-                cudaSafeCall( cudaGetLastError() );
-
-                if (stream == 0)
-                    cudaSafeCall( cudaDeviceSynchronize() );
-            }
-
-            template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
-            static void call(PtrStepSz<T1> src1, PtrStepSz<T2> src2, PtrStepSz<D> dst, BinOp op, Mask mask, cudaStream_t stream)
-            {
-                typedef TransformFunctorTraits<BinOp> ft;
-
-                const dim3 threads(ft::simple_block_dim_x, ft::simple_block_dim_y, 1);
-                const dim3 grid(divUp(src1.cols, threads.x), divUp(src1.rows, threads.y), 1);
-
-                transformSimple<T1, T2, D><<<grid, threads, 0, stream>>>(src1, src2, dst, mask, op);
-                cudaSafeCall( cudaGetLastError() );
-
-                if (stream == 0)
-                    cudaSafeCall( cudaDeviceSynchronize() );
-            }
-        };
-        template<> struct TransformDispatcher<true>
-        {
-            template <typename T, typename D, typename UnOp, typename Mask>
-            static void call(PtrStepSz<T> src, PtrStepSz<D> dst, UnOp op, Mask mask, cudaStream_t stream)
-            {
-                typedef TransformFunctorTraits<UnOp> ft;
-
-                CV_StaticAssert(ft::smart_shift != 1, "");
-
-                if (!isAligned(src.data, ft::smart_shift * sizeof(T)) || !isAligned(src.step, ft::smart_shift * sizeof(T)) ||
-                    !isAligned(dst.data, ft::smart_shift * sizeof(D)) || !isAligned(dst.step, ft::smart_shift * sizeof(D)))
-                {
-                    TransformDispatcher<false>::call(src, dst, op, mask, stream);
-                    return;
-                }
-
-                const dim3 threads(ft::smart_block_dim_x, ft::smart_block_dim_y, 1);
-                const dim3 grid(divUp(src.cols, threads.x * ft::smart_shift), divUp(src.rows, threads.y), 1);
-
-                transformSmart<T, D><<<grid, threads, 0, stream>>>(src, dst, mask, op);
-                cudaSafeCall( cudaGetLastError() );
-
-                if (stream == 0)
-                    cudaSafeCall( cudaDeviceSynchronize() );
-            }
-
-            template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
-            static void call(PtrStepSz<T1> src1, PtrStepSz<T2> src2, PtrStepSz<D> dst, BinOp op, Mask mask, cudaStream_t stream)
-            {
-                typedef TransformFunctorTraits<BinOp> ft;
-
-                CV_StaticAssert(ft::smart_shift != 1, "");
-
-                if (!isAligned(src1.data, ft::smart_shift * sizeof(T1)) || !isAligned(src1.step, ft::smart_shift * sizeof(T1)) ||
-                    !isAligned(src2.data, ft::smart_shift * sizeof(T2)) || !isAligned(src2.step, ft::smart_shift * sizeof(T2)) ||
-                    !isAligned(dst.data, ft::smart_shift * sizeof(D)) || !isAligned(dst.step, ft::smart_shift * sizeof(D)))
-                {
-                    TransformDispatcher<false>::call(src1, src2, dst, op, mask, stream);
-                    return;
-                }
-
-                const dim3 threads(ft::smart_block_dim_x, ft::smart_block_dim_y, 1);
-                const dim3 grid(divUp(src1.cols, threads.x * ft::smart_shift), divUp(src1.rows, threads.y), 1);
-
-                transformSmart<T1, T2, D><<<grid, threads, 0, stream>>>(src1, src2, dst, mask, op);
-                cudaSafeCall( cudaGetLastError() );
-
-                if (stream == 0)
-                    cudaSafeCall( cudaDeviceSynchronize() );
-            }
-        };
-    } // namespace transform_detail
-}}} // namespace cv { namespace cuda { namespace cudev
-
-//! @endcond
-
-#endif // OPENCV_CUDA_TRANSFORM_DETAIL_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/type_traits_detail.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/type_traits_detail.hpp
deleted file mode 100644
index a78bd2c0..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/type_traits_detail.hpp
+++ /dev/null
@@ -1,191 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_TYPE_TRAITS_DETAIL_HPP
-#define OPENCV_CUDA_TYPE_TRAITS_DETAIL_HPP
-
-#include "../common.hpp"
-#include "../vec_traits.hpp"
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    namespace type_traits_detail
-    {
-        template <bool, typename T1, typename T2> struct Select { typedef T1 type; };
-        template <typename T1, typename T2> struct Select<false, T1, T2> { typedef T2 type; };
-
-        template <typename T> struct IsSignedIntergral { enum {value = 0}; };
-        template <> struct IsSignedIntergral<schar> { enum {value = 1}; };
-        template <> struct IsSignedIntergral<char1> { enum {value = 1}; };
-        template <> struct IsSignedIntergral<short> { enum {value = 1}; };
-        template <> struct IsSignedIntergral<short1> { enum {value = 1}; };
-        template <> struct IsSignedIntergral<int> { enum {value = 1}; };
-        template <> struct IsSignedIntergral<int1> { enum {value = 1}; };
-
-        template <typename T> struct IsUnsignedIntegral { enum {value = 0}; };
-        template <> struct IsUnsignedIntegral<uchar> { enum {value = 1}; };
-        template <> struct IsUnsignedIntegral<uchar1> { enum {value = 1}; };
-        template <> struct IsUnsignedIntegral<ushort> { enum {value = 1}; };
-        template <> struct IsUnsignedIntegral<ushort1> { enum {value = 1}; };
-        template <> struct IsUnsignedIntegral<uint> { enum {value = 1}; };
-        template <> struct IsUnsignedIntegral<uint1> { enum {value = 1}; };
-
-        template <typename T> struct IsIntegral { enum {value = IsSignedIntergral<T>::value || IsUnsignedIntegral<T>::value}; };
-        template <> struct IsIntegral<char> { enum {value = 1}; };
-        template <> struct IsIntegral<bool> { enum {value = 1}; };
-
-        template <typename T> struct IsFloat { enum {value = 0}; };
-        template <> struct IsFloat<float> { enum {value = 1}; };
-        template <> struct IsFloat<double> { enum {value = 1}; };
-
-        template <typename T> struct IsVec { enum {value = 0}; };
-        template <> struct IsVec<uchar1> { enum {value = 1}; };
-        template <> struct IsVec<uchar2> { enum {value = 1}; };
-        template <> struct IsVec<uchar3> { enum {value = 1}; };
-        template <> struct IsVec<uchar4> { enum {value = 1}; };
-        template <> struct IsVec<uchar8> { enum {value = 1}; };
-        template <> struct IsVec<char1> { enum {value = 1}; };
-        template <> struct IsVec<char2> { enum {value = 1}; };
-        template <> struct IsVec<char3> { enum {value = 1}; };
-        template <> struct IsVec<char4> { enum {value = 1}; };
-        template <> struct IsVec<char8> { enum {value = 1}; };
-        template <> struct IsVec<ushort1> { enum {value = 1}; };
-        template <> struct IsVec<ushort2> { enum {value = 1}; };
-        template <> struct IsVec<ushort3> { enum {value = 1}; };
-        template <> struct IsVec<ushort4> { enum {value = 1}; };
-        template <> struct IsVec<ushort8> { enum {value = 1}; };
-        template <> struct IsVec<short1> { enum {value = 1}; };
-        template <> struct IsVec<short2> { enum {value = 1}; };
-        template <> struct IsVec<short3> { enum {value = 1}; };
-        template <> struct IsVec<short4> { enum {value = 1}; };
-        template <> struct IsVec<short8> { enum {value = 1}; };
-        template <> struct IsVec<uint1> { enum {value = 1}; };
-        template <> struct IsVec<uint2> { enum {value = 1}; };
-        template <> struct IsVec<uint3> { enum {value = 1}; };
-        template <> struct IsVec<uint4> { enum {value = 1}; };
-        template <> struct IsVec<uint8> { enum {value = 1}; };
-        template <> struct IsVec<int1> { enum {value = 1}; };
-        template <> struct IsVec<int2> { enum {value = 1}; };
-        template <> struct IsVec<int3> { enum {value = 1}; };
-        template <> struct IsVec<int4> { enum {value = 1}; };
-        template <> struct IsVec<int8> { enum {value = 1}; };
-        template <> struct IsVec<float1> { enum {value = 1}; };
-        template <> struct IsVec<float2> { enum {value = 1}; };
-        template <> struct IsVec<float3> { enum {value = 1}; };
-        template <> struct IsVec<float4> { enum {value = 1}; };
-        template <> struct IsVec<float8> { enum {value = 1}; };
-        template <> struct IsVec<double1> { enum {value = 1}; };
-        template <> struct IsVec<double2> { enum {value = 1}; };
-        template <> struct IsVec<double3> { enum {value = 1}; };
-        template <> struct IsVec<double4> { enum {value = 1}; };
-        template <> struct IsVec<double8> { enum {value = 1}; };
-
-        template <class U> struct AddParameterType { typedef const U& type; };
-        template <class U> struct AddParameterType<U&> { typedef U& type; };
-        template <> struct AddParameterType<void> { typedef void type; };
-
-        template <class U> struct ReferenceTraits
-        {
-            enum { value = false };
-            typedef U type;
-        };
-        template <class U> struct ReferenceTraits<U&>
-        {
-            enum { value = true };
-            typedef U type;
-        };
-
-        template <class U> struct PointerTraits
-        {
-            enum { value = false };
-            typedef void type;
-        };
-        template <class U> struct PointerTraits<U*>
-        {
-            enum { value = true };
-            typedef U type;
-        };
-        template <class U> struct PointerTraits<U*&>
-        {
-            enum { value = true };
-            typedef U type;
-        };
-
-        template <class U> struct UnConst
-        {
-            typedef U type;
-            enum { value = 0 };
-        };
-        template <class U> struct UnConst<const U>
-        {
-            typedef U type;
-            enum { value = 1 };
-        };
-        template <class U> struct UnConst<const U&>
-        {
-            typedef U& type;
-            enum { value = 1 };
-        };
-
-        template <class U> struct UnVolatile
-        {
-            typedef U type;
-            enum { value = 0 };
-        };
-        template <class U> struct UnVolatile<volatile U>
-        {
-            typedef U type;
-            enum { value = 1 };
-        };
-        template <class U> struct UnVolatile<volatile U&>
-        {
-            typedef U& type;
-            enum { value = 1 };
-        };
-    } // namespace type_traits_detail
-}}} // namespace cv { namespace cuda { namespace cudev
-
-//! @endcond
-
-#endif // OPENCV_CUDA_TYPE_TRAITS_DETAIL_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/vec_distance_detail.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/vec_distance_detail.hpp
deleted file mode 100644
index 8283a995..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/detail/vec_distance_detail.hpp
+++ /dev/null
@@ -1,121 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_VEC_DISTANCE_DETAIL_HPP
-#define OPENCV_CUDA_VEC_DISTANCE_DETAIL_HPP
-
-#include "../datamov_utils.hpp"
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    namespace vec_distance_detail
-    {
-        template <int THREAD_DIM, int N> struct UnrollVecDiffCached
-        {
-            template <typename Dist, typename T1, typename T2>
-            static __device__ void calcCheck(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int ind)
-            {
-                if (ind < len)
-                {
-                    T1 val1 = *vecCached++;
-
-                    T2 val2;
-                    ForceGlob<T2>::Load(vecGlob, ind, val2);
-
-                    dist.reduceIter(val1, val2);
-
-                    UnrollVecDiffCached<THREAD_DIM, N - 1>::calcCheck(vecCached, vecGlob, len, dist, ind + THREAD_DIM);
-                }
-            }
-
-            template <typename Dist, typename T1, typename T2>
-            static __device__ void calcWithoutCheck(const T1* vecCached, const T2* vecGlob, Dist& dist)
-            {
-                T1 val1 = *vecCached++;
-
-                T2 val2;
-                ForceGlob<T2>::Load(vecGlob, 0, val2);
-                vecGlob += THREAD_DIM;
-
-                dist.reduceIter(val1, val2);
-
-                UnrollVecDiffCached<THREAD_DIM, N - 1>::calcWithoutCheck(vecCached, vecGlob, dist);
-            }
-        };
-        template <int THREAD_DIM> struct UnrollVecDiffCached<THREAD_DIM, 0>
-        {
-            template <typename Dist, typename T1, typename T2>
-            static __device__ __forceinline__ void calcCheck(const T1*, const T2*, int, Dist&, int)
-            {
-            }
-
-            template <typename Dist, typename T1, typename T2>
-            static __device__ __forceinline__ void calcWithoutCheck(const T1*, const T2*, Dist&)
-            {
-            }
-        };
-
-        template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN> struct VecDiffCachedCalculator;
-        template <int THREAD_DIM, int MAX_LEN> struct VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, false>
-        {
-            template <typename Dist, typename T1, typename T2>
-            static __device__ __forceinline__ void calc(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int tid)
-            {
-                UnrollVecDiffCached<THREAD_DIM, MAX_LEN / THREAD_DIM>::calcCheck(vecCached, vecGlob, len, dist, tid);
-            }
-        };
-        template <int THREAD_DIM, int MAX_LEN> struct VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, true>
-        {
-            template <typename Dist, typename T1, typename T2>
-            static __device__ __forceinline__ void calc(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, int tid)
-            {
-                UnrollVecDiffCached<THREAD_DIM, MAX_LEN / THREAD_DIM>::calcWithoutCheck(vecCached, vecGlob + tid, dist);
-            }
-        };
-    } // namespace vec_distance_detail
-}}} // namespace cv { namespace cuda { namespace cudev
-
-//! @endcond
-
-#endif // OPENCV_CUDA_VEC_DISTANCE_DETAIL_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/dynamic_smem.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/dynamic_smem.hpp
deleted file mode 100644
index 42570c68..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/dynamic_smem.hpp
+++ /dev/null
@@ -1,88 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_DYNAMIC_SMEM_HPP
-#define OPENCV_CUDA_DYNAMIC_SMEM_HPP
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    template<class T> struct DynamicSharedMem
-    {
-        __device__ __forceinline__ operator T*()
-        {
-            extern __shared__ int __smem[];
-            return (T*)__smem;
-        }
-
-        __device__ __forceinline__ operator const T*() const
-        {
-            extern __shared__ int __smem[];
-            return (T*)__smem;
-        }
-    };
-
-    // specialize for double to avoid unaligned memory access compile errors
-    template<> struct DynamicSharedMem<double>
-    {
-        __device__ __forceinline__ operator double*()
-        {
-            extern __shared__ double __smem_d[];
-            return (double*)__smem_d;
-        }
-
-        __device__ __forceinline__ operator const double*() const
-        {
-            extern __shared__ double __smem_d[];
-            return (double*)__smem_d;
-        }
-    };
-}}}
-
-//! @endcond
-
-#endif // OPENCV_CUDA_DYNAMIC_SMEM_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/emulation.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/emulation.hpp
deleted file mode 100644
index 17dc1171..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/emulation.hpp
+++ /dev/null
@@ -1,269 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_EMULATION_HPP_
-#define OPENCV_CUDA_EMULATION_HPP_
-
-#include "common.hpp"
-#include "warp_reduce.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    struct Emulation
-    {
-
-        static __device__ __forceinline__ int syncthreadsOr(int pred)
-        {
-#if defined (__CUDA_ARCH__) && (__CUDA_ARCH__ < 200)
-                // just campilation stab
-                return 0;
-#else
-                return __syncthreads_or(pred);
-#endif
-        }
-
-        template<int CTA_SIZE>
-        static __forceinline__ __device__ int Ballot(int predicate)
-        {
-#if defined (__CUDA_ARCH__) && (__CUDA_ARCH__ >= 200)
-            return __ballot(predicate);
-#else
-            __shared__ volatile int cta_buffer[CTA_SIZE];
-
-            int tid = threadIdx.x;
-            cta_buffer[tid] = predicate ? (1 << (tid & 31)) : 0;
-            return warp_reduce(cta_buffer);
-#endif
-        }
-
-        struct smem
-        {
-            enum { TAG_MASK = (1U << ( (sizeof(unsigned int) << 3) - 5U)) - 1U };
-
-            template<typename T>
-            static __device__ __forceinline__ T atomicInc(T* address, T val)
-            {
-#if defined (__CUDA_ARCH__) && (__CUDA_ARCH__ < 120)
-                T count;
-                unsigned int tag = threadIdx.x << ( (sizeof(unsigned int) << 3) - 5U);
-                do
-                {
-                    count = *address & TAG_MASK;
-                    count = tag | (count + 1);
-                    *address = count;
-                } while (*address != count);
-
-                return (count & TAG_MASK) - 1;
-#else
-                return ::atomicInc(address, val);
-#endif
-            }
-
-            template<typename T>
-            static __device__ __forceinline__ T atomicAdd(T* address, T val)
-            {
-#if defined (__CUDA_ARCH__) && (__CUDA_ARCH__ < 120)
-                T count;
-                unsigned int tag = threadIdx.x << ( (sizeof(unsigned int) << 3) - 5U);
-                do
-                {
-                    count = *address & TAG_MASK;
-                    count = tag | (count + val);
-                    *address = count;
-                } while (*address != count);
-
-                return (count & TAG_MASK) - val;
-#else
-                return ::atomicAdd(address, val);
-#endif
-            }
-
-            template<typename T>
-            static __device__ __forceinline__ T atomicMin(T* address, T val)
-            {
-#if defined (__CUDA_ARCH__) && (__CUDA_ARCH__ < 120)
-                T count = ::min(*address, val);
-                do
-                {
-                    *address = count;
-                } while (*address > count);
-
-                return count;
-#else
-                return ::atomicMin(address, val);
-#endif
-            }
-        }; // struct cmem
-
-        struct glob
-        {
-            static __device__ __forceinline__ int atomicAdd(int* address, int val)
-            {
-                return ::atomicAdd(address, val);
-            }
-            static __device__ __forceinline__ unsigned int atomicAdd(unsigned int* address, unsigned int val)
-            {
-                return ::atomicAdd(address, val);
-            }
-            static __device__ __forceinline__ float atomicAdd(float* address, float val)
-            {
-            #if __CUDA_ARCH__ >= 200
-                return ::atomicAdd(address, val);
-            #else
-                int* address_as_i = (int*) address;
-                int old = *address_as_i, assumed;
-                do {
-                    assumed = old;
-                    old = ::atomicCAS(address_as_i, assumed,
-                        __float_as_int(val + __int_as_float(assumed)));
-                } while (assumed != old);
-                return __int_as_float(old);
-            #endif
-            }
-            static __device__ __forceinline__ double atomicAdd(double* address, double val)
-            {
-            #if __CUDA_ARCH__ >= 130
-                unsigned long long int* address_as_ull = (unsigned long long int*) address;
-                unsigned long long int old = *address_as_ull, assumed;
-                do {
-                    assumed = old;
-                    old = ::atomicCAS(address_as_ull, assumed,
-                        __double_as_longlong(val + __longlong_as_double(assumed)));
-                } while (assumed != old);
-                return __longlong_as_double(old);
-            #else
-                CV_UNUSED(address);
-                CV_UNUSED(val);
-                return 0.0;
-            #endif
-            }
-
-            static __device__ __forceinline__ int atomicMin(int* address, int val)
-            {
-                return ::atomicMin(address, val);
-            }
-            static __device__ __forceinline__ float atomicMin(float* address, float val)
-            {
-            #if __CUDA_ARCH__ >= 120
-                int* address_as_i = (int*) address;
-                int old = *address_as_i, assumed;
-                do {
-                    assumed = old;
-                    old = ::atomicCAS(address_as_i, assumed,
-                        __float_as_int(::fminf(val, __int_as_float(assumed))));
-                } while (assumed != old);
-                return __int_as_float(old);
-            #else
-                CV_UNUSED(address);
-                CV_UNUSED(val);
-                return 0.0f;
-            #endif
-            }
-            static __device__ __forceinline__ double atomicMin(double* address, double val)
-            {
-            #if __CUDA_ARCH__ >= 130
-                unsigned long long int* address_as_ull = (unsigned long long int*) address;
-                unsigned long long int old = *address_as_ull, assumed;
-                do {
-                    assumed = old;
-                    old = ::atomicCAS(address_as_ull, assumed,
-                        __double_as_longlong(::fmin(val, __longlong_as_double(assumed))));
-                } while (assumed != old);
-                return __longlong_as_double(old);
-            #else
-                CV_UNUSED(address);
-                CV_UNUSED(val);
-                return 0.0;
-            #endif
-            }
-
-            static __device__ __forceinline__ int atomicMax(int* address, int val)
-            {
-                return ::atomicMax(address, val);
-            }
-            static __device__ __forceinline__ float atomicMax(float* address, float val)
-            {
-            #if __CUDA_ARCH__ >= 120
-                int* address_as_i = (int*) address;
-                int old = *address_as_i, assumed;
-                do {
-                    assumed = old;
-                    old = ::atomicCAS(address_as_i, assumed,
-                        __float_as_int(::fmaxf(val, __int_as_float(assumed))));
-                } while (assumed != old);
-                return __int_as_float(old);
-            #else
-                CV_UNUSED(address);
-                CV_UNUSED(val);
-                return 0.0f;
-            #endif
-            }
-            static __device__ __forceinline__ double atomicMax(double* address, double val)
-            {
-            #if __CUDA_ARCH__ >= 130
-                unsigned long long int* address_as_ull = (unsigned long long int*) address;
-                unsigned long long int old = *address_as_ull, assumed;
-                do {
-                    assumed = old;
-                    old = ::atomicCAS(address_as_ull, assumed,
-                        __double_as_longlong(::fmax(val, __longlong_as_double(assumed))));
-                } while (assumed != old);
-                return __longlong_as_double(old);
-            #else
-                CV_UNUSED(address);
-                CV_UNUSED(val);
-                return 0.0;
-            #endif
-            }
-        };
-    }; //struct Emulation
-}}} // namespace cv { namespace cuda { namespace cudev
-
-//! @endcond
-
-#endif /* OPENCV_CUDA_EMULATION_HPP_ */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/filters.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/filters.hpp
deleted file mode 100644
index bf3147ed..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/filters.hpp
+++ /dev/null
@@ -1,293 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_FILTERS_HPP
-#define OPENCV_CUDA_FILTERS_HPP
-
-#include "saturate_cast.hpp"
-#include "vec_traits.hpp"
-#include "vec_math.hpp"
-#include "type_traits.hpp"
-#include "nppdefs.h"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    template <typename Ptr2D> struct PointFilter
-    {
-        typedef typename Ptr2D::elem_type elem_type;
-        typedef float index_type;
-
-        explicit __host__ __device__ __forceinline__ PointFilter(const Ptr2D& src_, float fx = 0.f, float fy = 0.f)
-        : src(src_)
-        {
-            CV_UNUSED(fx);
-            CV_UNUSED(fy);
-        }
-
-        __device__ __forceinline__ elem_type operator ()(float y, float x) const
-        {
-            return src(__float2int_rz(y), __float2int_rz(x));
-        }
-
-        Ptr2D src;
-    };
-
-    template <typename Ptr2D> struct LinearFilter
-    {
-        typedef typename Ptr2D::elem_type elem_type;
-        typedef float index_type;
-
-        explicit __host__ __device__ __forceinline__ LinearFilter(const Ptr2D& src_, float fx = 0.f, float fy = 0.f)
-        : src(src_)
-        {
-            CV_UNUSED(fx);
-            CV_UNUSED(fy);
-        }
-        __device__ __forceinline__ elem_type operator ()(float y, float x) const
-        {
-            typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
-
-            work_type out = VecTraits<work_type>::all(0);
-
-            const int x1 = __float2int_rd(x);
-            const int y1 = __float2int_rd(y);
-            if (x1 <= NPP_MIN_32S || x1 >= NPP_MAX_32S || y1 <= NPP_MIN_32S || y1 >= NPP_MAX_32S)
-            {
-                elem_type src_reg = src(y1, x1);
-                out = out + src_reg * 1.0f;
-                return saturate_cast<elem_type>(out);
-            }
-            const int x2 = x1 + 1;
-            const int y2 = y1 + 1;
-
-            elem_type src_reg = src(y1, x1);
-            out = out + src_reg * ((x2 - x) * (y2 - y));
-
-            src_reg = src(y1, x2);
-            out = out + src_reg * ((x - x1) * (y2 - y));
-
-            src_reg = src(y2, x1);
-            out = out + src_reg * ((x2 - x) * (y - y1));
-
-            src_reg = src(y2, x2);
-            out = out + src_reg * ((x - x1) * (y - y1));
-
-            return saturate_cast<elem_type>(out);
-        }
-
-        Ptr2D src;
-    };
-
-    template <typename Ptr2D> struct CubicFilter
-    {
-        typedef typename Ptr2D::elem_type elem_type;
-        typedef float index_type;
-        typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
-
-        explicit __host__ __device__ __forceinline__ CubicFilter(const Ptr2D& src_, float fx = 0.f, float fy = 0.f)
-        : src(src_)
-        {
-            CV_UNUSED(fx);
-            CV_UNUSED(fy);
-        }
-
-        static __device__ __forceinline__ float bicubicCoeff(float x_)
-        {
-            float x = fabsf(x_);
-            if (x <= 1.0f)
-            {
-                return x * x * (1.5f * x - 2.5f) + 1.0f;
-            }
-            else if (x < 2.0f)
-            {
-                return x * (x * (-0.5f * x + 2.5f) - 4.0f) + 2.0f;
-            }
-            else
-            {
-                return 0.0f;
-            }
-        }
-
-        __device__ elem_type operator ()(float y, float x) const
-        {
-            const float xmin = ::ceilf(x - 2.0f);
-            const float xmax = ::floorf(x + 2.0f);
-
-            const float ymin = ::ceilf(y - 2.0f);
-            const float ymax = ::floorf(y + 2.0f);
-
-            work_type sum = VecTraits<work_type>::all(0);
-            float wsum = 0.0f;
-
-            for (float cy = ymin; cy <= ymax; cy += 1.0f)
-            {
-                for (float cx = xmin; cx <= xmax; cx += 1.0f)
-                {
-                    const float w = bicubicCoeff(x - cx) * bicubicCoeff(y - cy);
-                    sum = sum + w * src(__float2int_rd(cy), __float2int_rd(cx));
-                    wsum += w;
-                }
-            }
-
-            work_type res = (!wsum)? VecTraits<work_type>::all(0) : sum / wsum;
-
-            return saturate_cast<elem_type>(res);
-        }
-
-        Ptr2D src;
-    };
-    // for integer scaling
-    template <typename Ptr2D> struct IntegerAreaFilter
-    {
-        typedef typename Ptr2D::elem_type elem_type;
-        typedef float index_type;
-
-        explicit __host__ __device__ __forceinline__ IntegerAreaFilter(const Ptr2D& src_, float scale_x_, float scale_y_)
-            : src(src_), scale_x(scale_x_), scale_y(scale_y_), scale(1.f / (scale_x * scale_y)) {}
-
-        __device__ __forceinline__ elem_type operator ()(float y, float x) const
-        {
-            float fsx1 = x * scale_x;
-            float fsx2 = fsx1 + scale_x;
-
-            int sx1 = __float2int_ru(fsx1);
-            int sx2 = __float2int_rd(fsx2);
-
-            float fsy1 = y * scale_y;
-            float fsy2 = fsy1 + scale_y;
-
-            int sy1 = __float2int_ru(fsy1);
-            int sy2 = __float2int_rd(fsy2);
-
-            typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
-            work_type out = VecTraits<work_type>::all(0.f);
-
-            for(int dy = sy1; dy < sy2; ++dy)
-                for(int dx = sx1; dx < sx2; ++dx)
-                {
-                    out = out + src(dy, dx) * scale;
-                }
-
-            return saturate_cast<elem_type>(out);
-        }
-
-        Ptr2D src;
-        float scale_x, scale_y ,scale;
-    };
-
-    template <typename Ptr2D> struct AreaFilter
-    {
-        typedef typename Ptr2D::elem_type elem_type;
-        typedef float index_type;
-
-        explicit __host__ __device__ __forceinline__ AreaFilter(const Ptr2D& src_, float scale_x_, float scale_y_)
-            : src(src_), scale_x(scale_x_), scale_y(scale_y_){}
-
-        __device__ __forceinline__ elem_type operator ()(float y, float x) const
-        {
-            float fsx1 = x * scale_x;
-            float fsx2 = fsx1 + scale_x;
-
-            int sx1 = __float2int_ru(fsx1);
-            int sx2 = __float2int_rd(fsx2);
-
-            float fsy1 = y * scale_y;
-            float fsy2 = fsy1 + scale_y;
-
-            int sy1 = __float2int_ru(fsy1);
-            int sy2 = __float2int_rd(fsy2);
-
-            float scale = 1.f / (fminf(scale_x, src.width - fsx1) * fminf(scale_y, src.height - fsy1));
-
-            typedef typename TypeVec<float, VecTraits<elem_type>::cn>::vec_type work_type;
-            work_type out = VecTraits<work_type>::all(0.f);
-
-            for (int dy = sy1; dy < sy2; ++dy)
-            {
-                for (int dx = sx1; dx < sx2; ++dx)
-                    out = out + src(dy, dx) * scale;
-
-                if (sx1 > fsx1)
-                    out = out + src(dy, (sx1 -1) ) * ((sx1 - fsx1) * scale);
-
-                if (sx2 < fsx2)
-                    out = out + src(dy, sx2) * ((fsx2 -sx2) * scale);
-            }
-
-            if (sy1 > fsy1)
-                for (int dx = sx1; dx < sx2; ++dx)
-                    out = out + src( (sy1 - 1) , dx) * ((sy1 -fsy1) * scale);
-
-            if (sy2 < fsy2)
-                for (int dx = sx1; dx < sx2; ++dx)
-                    out = out + src(sy2, dx) * ((fsy2 -sy2) * scale);
-
-            if ((sy1 > fsy1) &&  (sx1 > fsx1))
-                out = out + src( (sy1 - 1) , (sx1 - 1)) * ((sy1 -fsy1) * (sx1 -fsx1) * scale);
-
-            if ((sy1 > fsy1) &&  (sx2 < fsx2))
-                out = out + src( (sy1 - 1) , sx2) * ((sy1 -fsy1) * (fsx2 -sx2) * scale);
-
-            if ((sy2 < fsy2) &&  (sx2 < fsx2))
-                out = out + src(sy2, sx2) * ((fsy2 -sy2) * (fsx2 -sx2) * scale);
-
-            if ((sy2 < fsy2) &&  (sx1 > fsx1))
-                out = out + src(sy2, (sx1 - 1)) * ((fsy2 -sy2) * (sx1 -fsx1) * scale);
-
-            return saturate_cast<elem_type>(out);
-        }
-
-        Ptr2D src;
-        float scale_x, scale_y;
-        int width, haight;
-    };
-}}} // namespace cv { namespace cuda { namespace cudev
-
-//! @endcond
-
-#endif // OPENCV_CUDA_FILTERS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/funcattrib.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/funcattrib.hpp
deleted file mode 100644
index f5820804..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/funcattrib.hpp
+++ /dev/null
@@ -1,79 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_DEVICE_FUNCATTRIB_HPP
-#define OPENCV_CUDA_DEVICE_FUNCATTRIB_HPP
-
-#include <cstdio>
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    template<class Func>
-    void printFuncAttrib(Func& func)
-    {
-
-        cudaFuncAttributes attrs;
-        cudaFuncGetAttributes(&attrs, func);
-
-        printf("=== Function stats ===\n");
-        printf("Name: \n");
-        printf("sharedSizeBytes    = %d\n", attrs.sharedSizeBytes);
-        printf("constSizeBytes     = %d\n", attrs.constSizeBytes);
-        printf("localSizeBytes     = %d\n", attrs.localSizeBytes);
-        printf("maxThreadsPerBlock = %d\n", attrs.maxThreadsPerBlock);
-        printf("numRegs            = %d\n", attrs.numRegs);
-        printf("ptxVersion         = %d\n", attrs.ptxVersion);
-        printf("binaryVersion      = %d\n", attrs.binaryVersion);
-        printf("\n");
-        fflush(stdout);
-    }
-}}} // namespace cv { namespace cuda { namespace cudev
-
-//! @endcond
-
-#endif  /* OPENCV_CUDA_DEVICE_FUNCATTRIB_HPP */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/functional.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/functional.hpp
deleted file mode 100644
index 9f53d875..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/functional.hpp
+++ /dev/null
@@ -1,805 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_FUNCTIONAL_HPP
-#define OPENCV_CUDA_FUNCTIONAL_HPP
-
-#include <functional>
-#include "saturate_cast.hpp"
-#include "vec_traits.hpp"
-#include "type_traits.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    // Function Objects
-    template<typename Argument, typename Result> struct unary_function
-    {
-        typedef Argument argument_type;
-        typedef Result result_type;
-    };
-    template<typename Argument1, typename Argument2, typename Result> struct binary_function
-    {
-        typedef Argument1 first_argument_type;
-        typedef Argument2 second_argument_type;
-        typedef Result result_type;
-    };
-
-    // Arithmetic Operations
-    template <typename T> struct plus : binary_function<T, T, T>
-    {
-        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
-                                                 typename TypeTraits<T>::ParameterType b) const
-        {
-            return a + b;
-        }
-        __host__ __device__ __forceinline__ plus() {}
-        __host__ __device__ __forceinline__ plus(const plus&) {}
-    };
-
-    template <typename T> struct minus : binary_function<T, T, T>
-    {
-        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
-                                                 typename TypeTraits<T>::ParameterType b) const
-        {
-            return a - b;
-        }
-        __host__ __device__ __forceinline__ minus() {}
-        __host__ __device__ __forceinline__ minus(const minus&) {}
-    };
-
-    template <typename T> struct multiplies : binary_function<T, T, T>
-    {
-        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
-                                                 typename TypeTraits<T>::ParameterType b) const
-        {
-            return a * b;
-        }
-        __host__ __device__ __forceinline__ multiplies() {}
-        __host__ __device__ __forceinline__ multiplies(const multiplies&) {}
-    };
-
-    template <typename T> struct divides : binary_function<T, T, T>
-    {
-        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
-                                                 typename TypeTraits<T>::ParameterType b) const
-        {
-            return a / b;
-        }
-        __host__ __device__ __forceinline__ divides() {}
-        __host__ __device__ __forceinline__ divides(const divides&) {}
-    };
-
-    template <typename T> struct modulus : binary_function<T, T, T>
-    {
-        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
-                                                 typename TypeTraits<T>::ParameterType b) const
-        {
-            return a % b;
-        }
-        __host__ __device__ __forceinline__ modulus() {}
-        __host__ __device__ __forceinline__ modulus(const modulus&) {}
-    };
-
-    template <typename T> struct negate : unary_function<T, T>
-    {
-        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a) const
-        {
-            return -a;
-        }
-        __host__ __device__ __forceinline__ negate() {}
-        __host__ __device__ __forceinline__ negate(const negate&) {}
-    };
-
-    // Comparison Operations
-    template <typename T> struct equal_to : binary_function<T, T, bool>
-    {
-        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
-                                                    typename TypeTraits<T>::ParameterType b) const
-        {
-            return a == b;
-        }
-        __host__ __device__ __forceinline__ equal_to() {}
-        __host__ __device__ __forceinline__ equal_to(const equal_to&) {}
-    };
-
-    template <typename T> struct not_equal_to : binary_function<T, T, bool>
-    {
-        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
-                                                    typename TypeTraits<T>::ParameterType b) const
-        {
-            return a != b;
-        }
-        __host__ __device__ __forceinline__ not_equal_to() {}
-        __host__ __device__ __forceinline__ not_equal_to(const not_equal_to&) {}
-    };
-
-    template <typename T> struct greater : binary_function<T, T, bool>
-    {
-        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
-                                                    typename TypeTraits<T>::ParameterType b) const
-        {
-            return a > b;
-        }
-        __host__ __device__ __forceinline__ greater() {}
-        __host__ __device__ __forceinline__ greater(const greater&) {}
-    };
-
-    template <typename T> struct less : binary_function<T, T, bool>
-    {
-        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
-                                                    typename TypeTraits<T>::ParameterType b) const
-        {
-            return a < b;
-        }
-        __host__ __device__ __forceinline__ less() {}
-        __host__ __device__ __forceinline__ less(const less&) {}
-    };
-
-    template <typename T> struct greater_equal : binary_function<T, T, bool>
-    {
-        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
-                                                    typename TypeTraits<T>::ParameterType b) const
-        {
-            return a >= b;
-        }
-        __host__ __device__ __forceinline__ greater_equal() {}
-        __host__ __device__ __forceinline__ greater_equal(const greater_equal&) {}
-    };
-
-    template <typename T> struct less_equal : binary_function<T, T, bool>
-    {
-        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
-                                                    typename TypeTraits<T>::ParameterType b) const
-        {
-            return a <= b;
-        }
-        __host__ __device__ __forceinline__ less_equal() {}
-        __host__ __device__ __forceinline__ less_equal(const less_equal&) {}
-    };
-
-    // Logical Operations
-    template <typename T> struct logical_and : binary_function<T, T, bool>
-    {
-        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
-                                                    typename TypeTraits<T>::ParameterType b) const
-        {
-            return a && b;
-        }
-        __host__ __device__ __forceinline__ logical_and() {}
-        __host__ __device__ __forceinline__ logical_and(const logical_and&) {}
-    };
-
-    template <typename T> struct logical_or : binary_function<T, T, bool>
-    {
-        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a,
-                                                    typename TypeTraits<T>::ParameterType b) const
-        {
-            return a || b;
-        }
-        __host__ __device__ __forceinline__ logical_or() {}
-        __host__ __device__ __forceinline__ logical_or(const logical_or&) {}
-    };
-
-    template <typename T> struct logical_not : unary_function<T, bool>
-    {
-        __device__ __forceinline__ bool operator ()(typename TypeTraits<T>::ParameterType a) const
-        {
-            return !a;
-        }
-        __host__ __device__ __forceinline__ logical_not() {}
-        __host__ __device__ __forceinline__ logical_not(const logical_not&) {}
-    };
-
-    // Bitwise Operations
-    template <typename T> struct bit_and : binary_function<T, T, T>
-    {
-        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
-                                                 typename TypeTraits<T>::ParameterType b) const
-        {
-            return a & b;
-        }
-        __host__ __device__ __forceinline__ bit_and() {}
-        __host__ __device__ __forceinline__ bit_and(const bit_and&) {}
-    };
-
-    template <typename T> struct bit_or : binary_function<T, T, T>
-    {
-        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
-                                                 typename TypeTraits<T>::ParameterType b) const
-        {
-            return a | b;
-        }
-        __host__ __device__ __forceinline__ bit_or() {}
-        __host__ __device__ __forceinline__ bit_or(const bit_or&) {}
-    };
-
-    template <typename T> struct bit_xor : binary_function<T, T, T>
-    {
-        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType a,
-                                                 typename TypeTraits<T>::ParameterType b) const
-        {
-            return a ^ b;
-        }
-        __host__ __device__ __forceinline__ bit_xor() {}
-        __host__ __device__ __forceinline__ bit_xor(const bit_xor&) {}
-    };
-
-    template <typename T> struct bit_not : unary_function<T, T>
-    {
-        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType v) const
-        {
-            return ~v;
-        }
-        __host__ __device__ __forceinline__ bit_not() {}
-        __host__ __device__ __forceinline__ bit_not(const bit_not&) {}
-    };
-
-    // Generalized Identity Operations
-    template <typename T> struct identity : unary_function<T, T>
-    {
-        __device__ __forceinline__ typename TypeTraits<T>::ParameterType operator()(typename TypeTraits<T>::ParameterType x) const
-        {
-            return x;
-        }
-        __host__ __device__ __forceinline__ identity() {}
-        __host__ __device__ __forceinline__ identity(const identity&) {}
-    };
-
-    template <typename T1, typename T2> struct project1st : binary_function<T1, T2, T1>
-    {
-        __device__ __forceinline__ typename TypeTraits<T1>::ParameterType operator()(typename TypeTraits<T1>::ParameterType lhs, typename TypeTraits<T2>::ParameterType rhs) const
-        {
-            return lhs;
-        }
-        __host__ __device__ __forceinline__ project1st() {}
-        __host__ __device__ __forceinline__ project1st(const project1st&) {}
-    };
-
-    template <typename T1, typename T2> struct project2nd : binary_function<T1, T2, T2>
-    {
-        __device__ __forceinline__ typename TypeTraits<T2>::ParameterType operator()(typename TypeTraits<T1>::ParameterType lhs, typename TypeTraits<T2>::ParameterType rhs) const
-        {
-            return rhs;
-        }
-        __host__ __device__ __forceinline__ project2nd() {}
-        __host__ __device__ __forceinline__ project2nd(const project2nd&) {}
-    };
-
-    // Min/Max Operations
-
-#define OPENCV_CUDA_IMPLEMENT_MINMAX(name, type, op) \
-    template <> struct name<type> : binary_function<type, type, type> \
-    { \
-        __device__ __forceinline__ type operator()(type lhs, type rhs) const {return op(lhs, rhs);} \
-        __host__ __device__ __forceinline__ name() {}\
-        __host__ __device__ __forceinline__ name(const name&) {}\
-    };
-
-    template <typename T> struct maximum : binary_function<T, T, T>
-    {
-        __device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType lhs, typename TypeTraits<T>::ParameterType rhs) const
-        {
-            return max(lhs, rhs);
-        }
-        __host__ __device__ __forceinline__ maximum() {}
-        __host__ __device__ __forceinline__ maximum(const maximum&) {}
-    };
-
-    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, uchar, ::max)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, schar, ::max)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, char, ::max)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, ushort, ::max)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, short, ::max)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, int, ::max)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, uint, ::max)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, float, ::fmax)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(maximum, double, ::fmax)
-
-    template <typename T> struct minimum : binary_function<T, T, T>
-    {
-        __device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType lhs, typename TypeTraits<T>::ParameterType rhs) const
-        {
-            return min(lhs, rhs);
-        }
-        __host__ __device__ __forceinline__ minimum() {}
-        __host__ __device__ __forceinline__ minimum(const minimum&) {}
-    };
-
-    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, uchar, ::min)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, schar, ::min)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, char, ::min)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, ushort, ::min)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, short, ::min)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, int, ::min)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, uint, ::min)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, float, ::fmin)
-    OPENCV_CUDA_IMPLEMENT_MINMAX(minimum, double, ::fmin)
-
-#undef OPENCV_CUDA_IMPLEMENT_MINMAX
-
-    // Math functions
-
-    template <typename T> struct abs_func : unary_function<T, T>
-    {
-        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType x) const
-        {
-            return abs(x);
-        }
-
-        __host__ __device__ __forceinline__ abs_func() {}
-        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
-    };
-    template <> struct abs_func<unsigned char> : unary_function<unsigned char, unsigned char>
-    {
-        __device__ __forceinline__ unsigned char operator ()(unsigned char x) const
-        {
-            return x;
-        }
-
-        __host__ __device__ __forceinline__ abs_func() {}
-        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
-    };
-    template <> struct abs_func<signed char> : unary_function<signed char, signed char>
-    {
-        __device__ __forceinline__ signed char operator ()(signed char x) const
-        {
-            return ::abs((int)x);
-        }
-
-        __host__ __device__ __forceinline__ abs_func() {}
-        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
-    };
-    template <> struct abs_func<char> : unary_function<char, char>
-    {
-        __device__ __forceinline__ char operator ()(char x) const
-        {
-            return ::abs((int)x);
-        }
-
-        __host__ __device__ __forceinline__ abs_func() {}
-        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
-    };
-    template <> struct abs_func<unsigned short> : unary_function<unsigned short, unsigned short>
-    {
-        __device__ __forceinline__ unsigned short operator ()(unsigned short x) const
-        {
-            return x;
-        }
-
-        __host__ __device__ __forceinline__ abs_func() {}
-        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
-    };
-    template <> struct abs_func<short> : unary_function<short, short>
-    {
-        __device__ __forceinline__ short operator ()(short x) const
-        {
-            return ::abs((int)x);
-        }
-
-        __host__ __device__ __forceinline__ abs_func() {}
-        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
-    };
-    template <> struct abs_func<unsigned int> : unary_function<unsigned int, unsigned int>
-    {
-        __device__ __forceinline__ unsigned int operator ()(unsigned int x) const
-        {
-            return x;
-        }
-
-        __host__ __device__ __forceinline__ abs_func() {}
-        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
-    };
-    template <> struct abs_func<int> : unary_function<int, int>
-    {
-        __device__ __forceinline__ int operator ()(int x) const
-        {
-            return ::abs(x);
-        }
-
-        __host__ __device__ __forceinline__ abs_func() {}
-        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
-    };
-    template <> struct abs_func<float> : unary_function<float, float>
-    {
-        __device__ __forceinline__ float operator ()(float x) const
-        {
-            return ::fabsf(x);
-        }
-
-        __host__ __device__ __forceinline__ abs_func() {}
-        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
-    };
-    template <> struct abs_func<double> : unary_function<double, double>
-    {
-        __device__ __forceinline__ double operator ()(double x) const
-        {
-            return ::fabs(x);
-        }
-
-        __host__ __device__ __forceinline__ abs_func() {}
-        __host__ __device__ __forceinline__ abs_func(const abs_func&) {}
-    };
-
-#define OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(name, func) \
-    template <typename T> struct name ## _func : unary_function<T, float> \
-    { \
-        __device__ __forceinline__ float operator ()(typename TypeTraits<T>::ParameterType v) const \
-        { \
-            return func ## f(v); \
-        } \
-        __host__ __device__ __forceinline__ name ## _func() {} \
-        __host__ __device__ __forceinline__ name ## _func(const name ## _func&) {} \
-    }; \
-    template <> struct name ## _func<double> : unary_function<double, double> \
-    { \
-        __device__ __forceinline__ double operator ()(double v) const \
-        { \
-            return func(v); \
-        } \
-        __host__ __device__ __forceinline__ name ## _func() {} \
-        __host__ __device__ __forceinline__ name ## _func(const name ## _func&) {} \
-    };
-
-#define OPENCV_CUDA_IMPLEMENT_BIN_FUNCTOR(name, func) \
-    template <typename T> struct name ## _func : binary_function<T, T, float> \
-    { \
-        __device__ __forceinline__ float operator ()(typename TypeTraits<T>::ParameterType v1, typename TypeTraits<T>::ParameterType v2) const \
-        { \
-            return func ## f(v1, v2); \
-        } \
-        __host__ __device__ __forceinline__ name ## _func() {} \
-        __host__ __device__ __forceinline__ name ## _func(const name ## _func&) {} \
-    }; \
-    template <> struct name ## _func<double> : binary_function<double, double, double> \
-    { \
-        __device__ __forceinline__ double operator ()(double v1, double v2) const \
-        { \
-            return func(v1, v2); \
-        } \
-        __host__ __device__ __forceinline__ name ## _func() {} \
-        __host__ __device__ __forceinline__ name ## _func(const name ## _func&) {} \
-    };
-
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(sqrt, ::sqrt)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(exp, ::exp)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(exp2, ::exp2)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(exp10, ::exp10)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(log, ::log)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(log2, ::log2)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(log10, ::log10)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(sin, ::sin)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(cos, ::cos)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(tan, ::tan)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(asin, ::asin)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(acos, ::acos)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(atan, ::atan)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(sinh, ::sinh)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(cosh, ::cosh)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(tanh, ::tanh)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(asinh, ::asinh)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(acosh, ::acosh)
-    OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR(atanh, ::atanh)
-
-    OPENCV_CUDA_IMPLEMENT_BIN_FUNCTOR(hypot, ::hypot)
-    OPENCV_CUDA_IMPLEMENT_BIN_FUNCTOR(atan2, ::atan2)
-    OPENCV_CUDA_IMPLEMENT_BIN_FUNCTOR(pow, ::pow)
-
-    #undef OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR
-    #undef OPENCV_CUDA_IMPLEMENT_UN_FUNCTOR_NO_DOUBLE
-    #undef OPENCV_CUDA_IMPLEMENT_BIN_FUNCTOR
-
-    template<typename T> struct hypot_sqr_func : binary_function<T, T, float>
-    {
-        __device__ __forceinline__ T operator ()(typename TypeTraits<T>::ParameterType src1, typename TypeTraits<T>::ParameterType src2) const
-        {
-            return src1 * src1 + src2 * src2;
-        }
-        __host__ __device__ __forceinline__ hypot_sqr_func() {}
-        __host__ __device__ __forceinline__ hypot_sqr_func(const hypot_sqr_func&) {}
-    };
-
-    // Saturate Cast Functor
-    template <typename T, typename D> struct saturate_cast_func : unary_function<T, D>
-    {
-        __device__ __forceinline__ D operator ()(typename TypeTraits<T>::ParameterType v) const
-        {
-            return saturate_cast<D>(v);
-        }
-        __host__ __device__ __forceinline__ saturate_cast_func() {}
-        __host__ __device__ __forceinline__ saturate_cast_func(const saturate_cast_func&) {}
-    };
-
-    // Threshold Functors
-    template <typename T> struct thresh_binary_func : unary_function<T, T>
-    {
-        __host__ __device__ __forceinline__ thresh_binary_func(T thresh_, T maxVal_) : thresh(thresh_), maxVal(maxVal_) {}
-
-        __device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
-        {
-            return (src > thresh) * maxVal;
-        }
-
-        __host__ __device__ __forceinline__ thresh_binary_func() {}
-        __host__ __device__ __forceinline__ thresh_binary_func(const thresh_binary_func& other)
-            : thresh(other.thresh), maxVal(other.maxVal) {}
-
-        T thresh;
-        T maxVal;
-    };
-
-    template <typename T> struct thresh_binary_inv_func : unary_function<T, T>
-    {
-        __host__ __device__ __forceinline__ thresh_binary_inv_func(T thresh_, T maxVal_) : thresh(thresh_), maxVal(maxVal_) {}
-
-        __device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
-        {
-            return (src <= thresh) * maxVal;
-        }
-
-        __host__ __device__ __forceinline__ thresh_binary_inv_func() {}
-        __host__ __device__ __forceinline__ thresh_binary_inv_func(const thresh_binary_inv_func& other)
-            : thresh(other.thresh), maxVal(other.maxVal) {}
-
-        T thresh;
-        T maxVal;
-    };
-
-    template <typename T> struct thresh_trunc_func : unary_function<T, T>
-    {
-        explicit __host__ __device__ __forceinline__ thresh_trunc_func(T thresh_, T maxVal_ = 0) : thresh(thresh_) {CV_UNUSED(maxVal_);}
-
-        __device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
-        {
-            return minimum<T>()(src, thresh);
-        }
-
-        __host__ __device__ __forceinline__ thresh_trunc_func() {}
-        __host__ __device__ __forceinline__ thresh_trunc_func(const thresh_trunc_func& other)
-            : thresh(other.thresh) {}
-
-        T thresh;
-    };
-
-    template <typename T> struct thresh_to_zero_func : unary_function<T, T>
-    {
-        explicit __host__ __device__ __forceinline__ thresh_to_zero_func(T thresh_, T maxVal_ = 0) : thresh(thresh_) {CV_UNUSED(maxVal_);}
-
-        __device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
-        {
-            return (src > thresh) * src;
-        }
-
-        __host__ __device__ __forceinline__ thresh_to_zero_func() {}
-       __host__  __device__ __forceinline__ thresh_to_zero_func(const thresh_to_zero_func& other)
-            : thresh(other.thresh) {}
-
-        T thresh;
-    };
-
-    template <typename T> struct thresh_to_zero_inv_func : unary_function<T, T>
-    {
-        explicit __host__ __device__ __forceinline__ thresh_to_zero_inv_func(T thresh_, T maxVal_ = 0) : thresh(thresh_) {CV_UNUSED(maxVal_);}
-
-        __device__ __forceinline__ T operator()(typename TypeTraits<T>::ParameterType src) const
-        {
-            return (src <= thresh) * src;
-        }
-
-        __host__ __device__ __forceinline__ thresh_to_zero_inv_func() {}
-        __host__ __device__ __forceinline__ thresh_to_zero_inv_func(const thresh_to_zero_inv_func& other)
-            : thresh(other.thresh) {}
-
-        T thresh;
-    };
-
-    // Function Object Adaptors
-    template <typename Predicate> struct unary_negate : unary_function<typename Predicate::argument_type, bool>
-    {
-      explicit __host__ __device__ __forceinline__ unary_negate(const Predicate& p) : pred(p) {}
-
-      __device__ __forceinline__ bool operator()(typename TypeTraits<typename Predicate::argument_type>::ParameterType x) const
-      {
-          return !pred(x);
-      }
-
-      __host__ __device__ __forceinline__ unary_negate() {}
-      __host__ __device__ __forceinline__ unary_negate(const unary_negate& other) : pred(other.pred) {}
-
-      Predicate pred;
-    };
-
-    template <typename Predicate> __host__ __device__ __forceinline__ unary_negate<Predicate> not1(const Predicate& pred)
-    {
-        return unary_negate<Predicate>(pred);
-    }
-
-    template <typename Predicate> struct binary_negate : binary_function<typename Predicate::first_argument_type, typename Predicate::second_argument_type, bool>
-    {
-        explicit __host__ __device__ __forceinline__ binary_negate(const Predicate& p) : pred(p) {}
-
-        __device__ __forceinline__ bool operator()(typename TypeTraits<typename Predicate::first_argument_type>::ParameterType x,
-                                                   typename TypeTraits<typename Predicate::second_argument_type>::ParameterType y) const
-        {
-            return !pred(x,y);
-        }
-
-        __host__ __device__ __forceinline__ binary_negate() {}
-        __host__ __device__ __forceinline__ binary_negate(const binary_negate& other) : pred(other.pred) {}
-
-        Predicate pred;
-    };
-
-    template <typename BinaryPredicate> __host__ __device__ __forceinline__ binary_negate<BinaryPredicate> not2(const BinaryPredicate& pred)
-    {
-        return binary_negate<BinaryPredicate>(pred);
-    }
-
-    template <typename Op> struct binder1st : unary_function<typename Op::second_argument_type, typename Op::result_type>
-    {
-        __host__ __device__ __forceinline__ binder1st(const Op& op_, const typename Op::first_argument_type& arg1_) : op(op_), arg1(arg1_) {}
-
-        __device__ __forceinline__ typename Op::result_type operator ()(typename TypeTraits<typename Op::second_argument_type>::ParameterType a) const
-        {
-            return op(arg1, a);
-        }
-
-        __host__ __device__ __forceinline__ binder1st() {}
-        __host__ __device__ __forceinline__ binder1st(const binder1st& other) : op(other.op), arg1(other.arg1) {}
-
-        Op op;
-        typename Op::first_argument_type arg1;
-    };
-
-    template <typename Op, typename T> __host__ __device__ __forceinline__ binder1st<Op> bind1st(const Op& op, const T& x)
-    {
-        return binder1st<Op>(op, typename Op::first_argument_type(x));
-    }
-
-    template <typename Op> struct binder2nd : unary_function<typename Op::first_argument_type, typename Op::result_type>
-    {
-        __host__ __device__ __forceinline__ binder2nd(const Op& op_, const typename Op::second_argument_type& arg2_) : op(op_), arg2(arg2_) {}
-
-        __forceinline__ __device__ typename Op::result_type operator ()(typename TypeTraits<typename Op::first_argument_type>::ParameterType a) const
-        {
-            return op(a, arg2);
-        }
-
-        __host__ __device__ __forceinline__ binder2nd() {}
-        __host__ __device__ __forceinline__ binder2nd(const binder2nd& other) : op(other.op), arg2(other.arg2) {}
-
-        Op op;
-        typename Op::second_argument_type arg2;
-    };
-
-    template <typename Op, typename T> __host__ __device__ __forceinline__ binder2nd<Op> bind2nd(const Op& op, const T& x)
-    {
-        return binder2nd<Op>(op, typename Op::second_argument_type(x));
-    }
-
-    // Functor Traits
-    template <typename F> struct IsUnaryFunction
-    {
-        typedef char Yes;
-        struct No {Yes a[2];};
-
-        template <typename T, typename D> static Yes check(unary_function<T, D>);
-        static No check(...);
-
-        static F makeF();
-
-        enum { value = (sizeof(check(makeF())) == sizeof(Yes)) };
-    };
-
-    template <typename F> struct IsBinaryFunction
-    {
-        typedef char Yes;
-        struct No {Yes a[2];};
-
-        template <typename T1, typename T2, typename D> static Yes check(binary_function<T1, T2, D>);
-        static No check(...);
-
-        static F makeF();
-
-        enum { value = (sizeof(check(makeF())) == sizeof(Yes)) };
-    };
-
-    namespace functional_detail
-    {
-        template <size_t src_elem_size, size_t dst_elem_size> struct UnOpShift { enum { shift = 1 }; };
-        template <size_t src_elem_size> struct UnOpShift<src_elem_size, 1> { enum { shift = 4 }; };
-        template <size_t src_elem_size> struct UnOpShift<src_elem_size, 2> { enum { shift = 2 }; };
-
-        template <typename T, typename D> struct DefaultUnaryShift
-        {
-            enum { shift = UnOpShift<sizeof(T), sizeof(D)>::shift };
-        };
-
-        template <size_t src_elem_size1, size_t src_elem_size2, size_t dst_elem_size> struct BinOpShift { enum { shift = 1 }; };
-        template <size_t src_elem_size1, size_t src_elem_size2> struct BinOpShift<src_elem_size1, src_elem_size2, 1> { enum { shift = 4 }; };
-        template <size_t src_elem_size1, size_t src_elem_size2> struct BinOpShift<src_elem_size1, src_elem_size2, 2> { enum { shift = 2 }; };
-
-        template <typename T1, typename T2, typename D> struct DefaultBinaryShift
-        {
-            enum { shift = BinOpShift<sizeof(T1), sizeof(T2), sizeof(D)>::shift };
-        };
-
-        template <typename Func, bool unary = IsUnaryFunction<Func>::value> struct ShiftDispatcher;
-        template <typename Func> struct ShiftDispatcher<Func, true>
-        {
-            enum { shift = DefaultUnaryShift<typename Func::argument_type, typename Func::result_type>::shift };
-        };
-        template <typename Func> struct ShiftDispatcher<Func, false>
-        {
-            enum { shift = DefaultBinaryShift<typename Func::first_argument_type, typename Func::second_argument_type, typename Func::result_type>::shift };
-        };
-    }
-
-    template <typename Func> struct DefaultTransformShift
-    {
-        enum { shift = functional_detail::ShiftDispatcher<Func>::shift };
-    };
-
-    template <typename Func> struct DefaultTransformFunctorTraits
-    {
-        enum { simple_block_dim_x = 16 };
-        enum { simple_block_dim_y = 16 };
-
-        enum { smart_block_dim_x = 16 };
-        enum { smart_block_dim_y = 16 };
-        enum { smart_shift = DefaultTransformShift<Func>::shift };
-    };
-
-    template <typename Func> struct TransformFunctorTraits : DefaultTransformFunctorTraits<Func> {};
-
-#define OPENCV_CUDA_TRANSFORM_FUNCTOR_TRAITS(type) \
-    template <> struct TransformFunctorTraits< type > : DefaultTransformFunctorTraits< type >
-}}} // namespace cv { namespace cuda { namespace cudev
-
-//! @endcond
-
-#endif // OPENCV_CUDA_FUNCTIONAL_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/limits.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/limits.hpp
deleted file mode 100644
index 7e15ed62..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/limits.hpp
+++ /dev/null
@@ -1,128 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_LIMITS_HPP
-#define OPENCV_CUDA_LIMITS_HPP
-
-#include <limits.h>
-#include <float.h>
-#include "common.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-template <class T> struct numeric_limits;
-
-template <> struct numeric_limits<bool>
-{
-    __device__ __forceinline__ static bool min() { return false; }
-    __device__ __forceinline__ static bool max() { return true;  }
-    static const bool is_signed = false;
-};
-
-template <> struct numeric_limits<signed char>
-{
-    __device__ __forceinline__ static signed char min() { return SCHAR_MIN; }
-    __device__ __forceinline__ static signed char max() { return SCHAR_MAX; }
-    static const bool is_signed = true;
-};
-
-template <> struct numeric_limits<unsigned char>
-{
-    __device__ __forceinline__ static unsigned char min() { return 0; }
-    __device__ __forceinline__ static unsigned char max() { return UCHAR_MAX; }
-    static const bool is_signed = false;
-};
-
-template <> struct numeric_limits<short>
-{
-    __device__ __forceinline__ static short min() { return SHRT_MIN; }
-    __device__ __forceinline__ static short max() { return SHRT_MAX; }
-    static const bool is_signed = true;
-};
-
-template <> struct numeric_limits<unsigned short>
-{
-    __device__ __forceinline__ static unsigned short min() { return 0; }
-    __device__ __forceinline__ static unsigned short max() { return USHRT_MAX; }
-    static const bool is_signed = false;
-};
-
-template <> struct numeric_limits<int>
-{
-    __device__ __forceinline__ static int min() { return INT_MIN; }
-    __device__ __forceinline__ static int max() { return INT_MAX; }
-    static const bool is_signed = true;
-};
-
-template <> struct numeric_limits<unsigned int>
-{
-    __device__ __forceinline__ static unsigned int min() { return 0; }
-    __device__ __forceinline__ static unsigned int max() { return UINT_MAX; }
-    static const bool is_signed = false;
-};
-
-template <> struct numeric_limits<float>
-{
-    __device__ __forceinline__ static float min() { return FLT_MIN; }
-    __device__ __forceinline__ static float max() { return FLT_MAX; }
-    __device__ __forceinline__ static float epsilon() { return FLT_EPSILON; }
-    static const bool is_signed = true;
-};
-
-template <> struct numeric_limits<double>
-{
-    __device__ __forceinline__ static double min() { return DBL_MIN; }
-    __device__ __forceinline__ static double max() { return DBL_MAX; }
-    __device__ __forceinline__ static double epsilon() { return DBL_EPSILON; }
-    static const bool is_signed = true;
-};
-}}} // namespace cv { namespace cuda { namespace cudev {
-
-//! @endcond
-
-#endif // OPENCV_CUDA_LIMITS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/reduce.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/reduce.hpp
deleted file mode 100644
index 5de36508..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/reduce.hpp
+++ /dev/null
@@ -1,209 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_REDUCE_HPP
-#define OPENCV_CUDA_REDUCE_HPP
-
-#ifndef THRUST_DEBUG // eliminate -Wundef warning
-#define THRUST_DEBUG 0
-#endif
-
-#include <thrust/tuple.h>
-#include "detail/reduce.hpp"
-#include "detail/reduce_key_val.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    template <int N, typename T, class Op>
-    __device__ __forceinline__ void reduce(volatile T* smem, T& val, unsigned int tid, const Op& op)
-    {
-        reduce_detail::Dispatcher<N>::reductor::template reduce<volatile T*, T&, const Op&>(smem, val, tid, op);
-    }
-    template <int N,
-              typename P0, typename P1, typename P2, typename P3, typename P4, typename P5, typename P6, typename P7, typename P8, typename P9,
-              typename R0, typename R1, typename R2, typename R3, typename R4, typename R5, typename R6, typename R7, typename R8, typename R9,
-              class Op0, class Op1, class Op2, class Op3, class Op4, class Op5, class Op6, class Op7, class Op8, class Op9>
-    __device__ __forceinline__ void reduce(const thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>& smem,
-                                           const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>& val,
-                                           unsigned int tid,
-                                           const thrust::tuple<Op0, Op1, Op2, Op3, Op4, Op5, Op6, Op7, Op8, Op9>& op)
-    {
-        reduce_detail::Dispatcher<N>::reductor::template reduce<
-                const thrust::tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>&,
-                const thrust::tuple<R0, R1, R2, R3, R4, R5, R6, R7, R8, R9>&,
-                const thrust::tuple<Op0, Op1, Op2, Op3, Op4, Op5, Op6, Op7, Op8, Op9>&>(smem, val, tid, op);
-    }
-
-    template <unsigned int N, typename K, typename V, class Cmp>
-    __device__ __forceinline__ void reduceKeyVal(volatile K* skeys, K& key, volatile V* svals, V& val, unsigned int tid, const Cmp& cmp)
-    {
-        reduce_key_val_detail::Dispatcher<N>::reductor::template reduce<volatile K*, K&, volatile V*, V&, const Cmp&>(skeys, key, svals, val, tid, cmp);
-    }
-    template <unsigned int N,
-              typename K,
-              typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
-              typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
-              class Cmp>
-    __device__ __forceinline__ void reduceKeyVal(volatile K* skeys, K& key,
-                                                 const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals,
-                                                 const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
-                                                 unsigned int tid, const Cmp& cmp)
-    {
-        reduce_key_val_detail::Dispatcher<N>::reductor::template reduce<volatile K*, K&,
-                const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>&,
-                const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>&,
-                const Cmp&>(skeys, key, svals, val, tid, cmp);
-    }
-    template <unsigned int N,
-              typename KP0, typename KP1, typename KP2, typename KP3, typename KP4, typename KP5, typename KP6, typename KP7, typename KP8, typename KP9,
-              typename KR0, typename KR1, typename KR2, typename KR3, typename KR4, typename KR5, typename KR6, typename KR7, typename KR8, typename KR9,
-              typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9,
-              typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9,
-              class Cmp0, class Cmp1, class Cmp2, class Cmp3, class Cmp4, class Cmp5, class Cmp6, class Cmp7, class Cmp8, class Cmp9>
-    __device__ __forceinline__ void reduceKeyVal(const thrust::tuple<KP0, KP1, KP2, KP3, KP4, KP5, KP6, KP7, KP8, KP9>& skeys,
-                                                 const thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9>& key,
-                                                 const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals,
-                                                 const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val,
-                                                 unsigned int tid,
-                                                 const thrust::tuple<Cmp0, Cmp1, Cmp2, Cmp3, Cmp4, Cmp5, Cmp6, Cmp7, Cmp8, Cmp9>& cmp)
-    {
-        reduce_key_val_detail::Dispatcher<N>::reductor::template reduce<
-                const thrust::tuple<KP0, KP1, KP2, KP3, KP4, KP5, KP6, KP7, KP8, KP9>&,
-                const thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9>&,
-                const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>&,
-                const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>&,
-                const thrust::tuple<Cmp0, Cmp1, Cmp2, Cmp3, Cmp4, Cmp5, Cmp6, Cmp7, Cmp8, Cmp9>&
-                >(skeys, key, svals, val, tid, cmp);
-    }
-
-    // smem_tuple
-
-    template <typename T0>
-    __device__ __forceinline__
-    thrust::tuple<volatile T0*>
-    smem_tuple(T0* t0)
-    {
-        return thrust::make_tuple((volatile T0*) t0);
-    }
-
-    template <typename T0, typename T1>
-    __device__ __forceinline__
-    thrust::tuple<volatile T0*, volatile T1*>
-    smem_tuple(T0* t0, T1* t1)
-    {
-        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1);
-    }
-
-    template <typename T0, typename T1, typename T2>
-    __device__ __forceinline__
-    thrust::tuple<volatile T0*, volatile T1*, volatile T2*>
-    smem_tuple(T0* t0, T1* t1, T2* t2)
-    {
-        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2);
-    }
-
-    template <typename T0, typename T1, typename T2, typename T3>
-    __device__ __forceinline__
-    thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*>
-    smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3)
-    {
-        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3);
-    }
-
-    template <typename T0, typename T1, typename T2, typename T3, typename T4>
-    __device__ __forceinline__
-    thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*>
-    smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4)
-    {
-        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4);
-    }
-
-    template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5>
-    __device__ __forceinline__
-    thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*, volatile T5*>
-    smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4, T5* t5)
-    {
-        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4, (volatile T5*) t5);
-    }
-
-    template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5, typename T6>
-    __device__ __forceinline__
-    thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*, volatile T5*, volatile T6*>
-    smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4, T5* t5, T6* t6)
-    {
-        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4, (volatile T5*) t5, (volatile T6*) t6);
-    }
-
-    template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5, typename T6, typename T7>
-    __device__ __forceinline__
-    thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*, volatile T5*, volatile T6*, volatile T7*>
-    smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4, T5* t5, T6* t6, T7* t7)
-    {
-        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4, (volatile T5*) t5, (volatile T6*) t6, (volatile T7*) t7);
-    }
-
-    template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5, typename T6, typename T7, typename T8>
-    __device__ __forceinline__
-    thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*, volatile T5*, volatile T6*, volatile T7*, volatile T8*>
-    smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4, T5* t5, T6* t6, T7* t7, T8* t8)
-    {
-        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4, (volatile T5*) t5, (volatile T6*) t6, (volatile T7*) t7, (volatile T8*) t8);
-    }
-
-    template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5, typename T6, typename T7, typename T8, typename T9>
-    __device__ __forceinline__
-    thrust::tuple<volatile T0*, volatile T1*, volatile T2*, volatile T3*, volatile T4*, volatile T5*, volatile T6*, volatile T7*, volatile T8*, volatile T9*>
-    smem_tuple(T0* t0, T1* t1, T2* t2, T3* t3, T4* t4, T5* t5, T6* t6, T7* t7, T8* t8, T9* t9)
-    {
-        return thrust::make_tuple((volatile T0*) t0, (volatile T1*) t1, (volatile T2*) t2, (volatile T3*) t3, (volatile T4*) t4, (volatile T5*) t5, (volatile T6*) t6, (volatile T7*) t7, (volatile T8*) t8, (volatile T9*) t9);
-    }
-}}}
-
-//! @endcond
-
-#endif // OPENCV_CUDA_REDUCE_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/saturate_cast.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/saturate_cast.hpp
deleted file mode 100644
index c3a3d1cb..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/saturate_cast.hpp
+++ /dev/null
@@ -1,292 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_SATURATE_CAST_HPP
-#define OPENCV_CUDA_SATURATE_CAST_HPP
-
-#include "common.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(uchar v) { return _Tp(v); }
-    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(schar v) { return _Tp(v); }
-    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(ushort v) { return _Tp(v); }
-    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(short v) { return _Tp(v); }
-    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(uint v) { return _Tp(v); }
-    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(int v) { return _Tp(v); }
-    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(float v) { return _Tp(v); }
-    template<typename _Tp> __device__ __forceinline__ _Tp saturate_cast(double v) { return _Tp(v); }
-
-    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(schar v)
-    {
-        uint res = 0;
-        int vi = v;
-        asm("cvt.sat.u8.s8 %0, %1;" : "=r"(res) : "r"(vi));
-        return res;
-    }
-    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(short v)
-    {
-        uint res = 0;
-        asm("cvt.sat.u8.s16 %0, %1;" : "=r"(res) : "h"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(ushort v)
-    {
-        uint res = 0;
-        asm("cvt.sat.u8.u16 %0, %1;" : "=r"(res) : "h"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(int v)
-    {
-        uint res = 0;
-        asm("cvt.sat.u8.s32 %0, %1;" : "=r"(res) : "r"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(uint v)
-    {
-        uint res = 0;
-        asm("cvt.sat.u8.u32 %0, %1;" : "=r"(res) : "r"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(float v)
-    {
-        uint res = 0;
-        asm("cvt.rni.sat.u8.f32 %0, %1;" : "=r"(res) : "f"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ uchar saturate_cast<uchar>(double v)
-    {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
-        uint res = 0;
-        asm("cvt.rni.sat.u8.f64 %0, %1;" : "=r"(res) : "d"(v));
-        return res;
-    #else
-        return saturate_cast<uchar>((float)v);
-    #endif
-    }
-
-    template<> __device__ __forceinline__ schar saturate_cast<schar>(uchar v)
-    {
-        uint res = 0;
-        uint vi = v;
-        asm("cvt.sat.s8.u8 %0, %1;" : "=r"(res) : "r"(vi));
-        return res;
-    }
-    template<> __device__ __forceinline__ schar saturate_cast<schar>(short v)
-    {
-        uint res = 0;
-        asm("cvt.sat.s8.s16 %0, %1;" : "=r"(res) : "h"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ schar saturate_cast<schar>(ushort v)
-    {
-        uint res = 0;
-        asm("cvt.sat.s8.u16 %0, %1;" : "=r"(res) : "h"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ schar saturate_cast<schar>(int v)
-    {
-        uint res = 0;
-        asm("cvt.sat.s8.s32 %0, %1;" : "=r"(res) : "r"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ schar saturate_cast<schar>(uint v)
-    {
-        uint res = 0;
-        asm("cvt.sat.s8.u32 %0, %1;" : "=r"(res) : "r"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ schar saturate_cast<schar>(float v)
-    {
-        uint res = 0;
-        asm("cvt.rni.sat.s8.f32 %0, %1;" : "=r"(res) : "f"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ schar saturate_cast<schar>(double v)
-    {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
-        uint res = 0;
-        asm("cvt.rni.sat.s8.f64 %0, %1;" : "=r"(res) : "d"(v));
-        return res;
-    #else
-        return saturate_cast<schar>((float)v);
-    #endif
-    }
-
-    template<> __device__ __forceinline__ ushort saturate_cast<ushort>(schar v)
-    {
-        ushort res = 0;
-        int vi = v;
-        asm("cvt.sat.u16.s8 %0, %1;" : "=h"(res) : "r"(vi));
-        return res;
-    }
-    template<> __device__ __forceinline__ ushort saturate_cast<ushort>(short v)
-    {
-        ushort res = 0;
-        asm("cvt.sat.u16.s16 %0, %1;" : "=h"(res) : "h"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ ushort saturate_cast<ushort>(int v)
-    {
-        ushort res = 0;
-        asm("cvt.sat.u16.s32 %0, %1;" : "=h"(res) : "r"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ ushort saturate_cast<ushort>(uint v)
-    {
-        ushort res = 0;
-        asm("cvt.sat.u16.u32 %0, %1;" : "=h"(res) : "r"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ ushort saturate_cast<ushort>(float v)
-    {
-        ushort res = 0;
-        asm("cvt.rni.sat.u16.f32 %0, %1;" : "=h"(res) : "f"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ ushort saturate_cast<ushort>(double v)
-    {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
-        ushort res = 0;
-        asm("cvt.rni.sat.u16.f64 %0, %1;" : "=h"(res) : "d"(v));
-        return res;
-    #else
-        return saturate_cast<ushort>((float)v);
-    #endif
-    }
-
-    template<> __device__ __forceinline__ short saturate_cast<short>(ushort v)
-    {
-        short res = 0;
-        asm("cvt.sat.s16.u16 %0, %1;" : "=h"(res) : "h"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ short saturate_cast<short>(int v)
-    {
-        short res = 0;
-        asm("cvt.sat.s16.s32 %0, %1;" : "=h"(res) : "r"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ short saturate_cast<short>(uint v)
-    {
-        short res = 0;
-        asm("cvt.sat.s16.u32 %0, %1;" : "=h"(res) : "r"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ short saturate_cast<short>(float v)
-    {
-        short res = 0;
-        asm("cvt.rni.sat.s16.f32 %0, %1;" : "=h"(res) : "f"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ short saturate_cast<short>(double v)
-    {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
-        short res = 0;
-        asm("cvt.rni.sat.s16.f64 %0, %1;" : "=h"(res) : "d"(v));
-        return res;
-    #else
-        return saturate_cast<short>((float)v);
-    #endif
-    }
-
-    template<> __device__ __forceinline__ int saturate_cast<int>(uint v)
-    {
-        int res = 0;
-        asm("cvt.sat.s32.u32 %0, %1;" : "=r"(res) : "r"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ int saturate_cast<int>(float v)
-    {
-        return __float2int_rn(v);
-    }
-    template<> __device__ __forceinline__ int saturate_cast<int>(double v)
-    {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
-        return __double2int_rn(v);
-    #else
-        return saturate_cast<int>((float)v);
-    #endif
-    }
-
-    template<> __device__ __forceinline__ uint saturate_cast<uint>(schar v)
-    {
-        uint res = 0;
-        int vi = v;
-        asm("cvt.sat.u32.s8 %0, %1;" : "=r"(res) : "r"(vi));
-        return res;
-    }
-    template<> __device__ __forceinline__ uint saturate_cast<uint>(short v)
-    {
-        uint res = 0;
-        asm("cvt.sat.u32.s16 %0, %1;" : "=r"(res) : "h"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ uint saturate_cast<uint>(int v)
-    {
-        uint res = 0;
-        asm("cvt.sat.u32.s32 %0, %1;" : "=r"(res) : "r"(v));
-        return res;
-    }
-    template<> __device__ __forceinline__ uint saturate_cast<uint>(float v)
-    {
-        return __float2uint_rn(v);
-    }
-    template<> __device__ __forceinline__ uint saturate_cast<uint>(double v)
-    {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 130
-        return __double2uint_rn(v);
-    #else
-        return saturate_cast<uint>((float)v);
-    #endif
-    }
-}}}
-
-//! @endcond
-
-#endif /* OPENCV_CUDA_SATURATE_CAST_HPP */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/scan.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/scan.hpp
deleted file mode 100644
index e128fb09..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/scan.hpp
+++ /dev/null
@@ -1,258 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_SCAN_HPP
-#define OPENCV_CUDA_SCAN_HPP
-
-#include "opencv2/core/cuda/common.hpp"
-#include "opencv2/core/cuda/utility.hpp"
-#include "opencv2/core/cuda/warp.hpp"
-#include "opencv2/core/cuda/warp_shuffle.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    enum ScanKind { EXCLUSIVE = 0,  INCLUSIVE = 1 };
-
-    template <ScanKind Kind, typename T, typename F> struct WarpScan
-    {
-        __device__ __forceinline__ WarpScan() {}
-        __device__ __forceinline__ WarpScan(const WarpScan& other) { CV_UNUSED(other); }
-
-        __device__ __forceinline__ T operator()( volatile T *ptr , const unsigned int idx)
-        {
-            const unsigned int lane = idx & 31;
-            F op;
-
-            if ( lane >=  1) ptr [idx ] = op(ptr [idx -  1], ptr [idx]);
-            if ( lane >=  2) ptr [idx ] = op(ptr [idx -  2], ptr [idx]);
-            if ( lane >=  4) ptr [idx ] = op(ptr [idx -  4], ptr [idx]);
-            if ( lane >=  8) ptr [idx ] = op(ptr [idx -  8], ptr [idx]);
-            if ( lane >= 16) ptr [idx ] = op(ptr [idx - 16], ptr [idx]);
-
-            if( Kind == INCLUSIVE )
-                return ptr [idx];
-            else
-                return (lane > 0) ? ptr [idx - 1] : 0;
-        }
-
-        __device__ __forceinline__ unsigned int index(const unsigned int tid)
-        {
-            return tid;
-        }
-
-        __device__ __forceinline__ void init(volatile T *ptr){}
-
-        static const int warp_offset      = 0;
-
-        typedef WarpScan<INCLUSIVE, T, F>  merge;
-    };
-
-    template <ScanKind Kind , typename T, typename F> struct WarpScanNoComp
-    {
-        __device__ __forceinline__ WarpScanNoComp() {}
-        __device__ __forceinline__ WarpScanNoComp(const WarpScanNoComp& other) { CV_UNUSED(other); }
-
-        __device__ __forceinline__ T operator()( volatile T *ptr , const unsigned int idx)
-        {
-            const unsigned int lane = threadIdx.x & 31;
-            F op;
-
-            ptr [idx ] = op(ptr [idx -  1], ptr [idx]);
-            ptr [idx ] = op(ptr [idx -  2], ptr [idx]);
-            ptr [idx ] = op(ptr [idx -  4], ptr [idx]);
-            ptr [idx ] = op(ptr [idx -  8], ptr [idx]);
-            ptr [idx ] = op(ptr [idx - 16], ptr [idx]);
-
-            if( Kind == INCLUSIVE )
-                return ptr [idx];
-            else
-                return (lane > 0) ? ptr [idx - 1] : 0;
-        }
-
-        __device__ __forceinline__ unsigned int index(const unsigned int tid)
-        {
-            return (tid >> warp_log) * warp_smem_stride + 16 + (tid & warp_mask);
-        }
-
-        __device__ __forceinline__ void init(volatile T *ptr)
-        {
-            ptr[threadIdx.x] = 0;
-        }
-
-        static const int warp_smem_stride = 32 + 16 + 1;
-        static const int warp_offset      = 16;
-        static const int warp_log         = 5;
-        static const int warp_mask        = 31;
-
-        typedef WarpScanNoComp<INCLUSIVE, T, F> merge;
-    };
-
-    template <ScanKind Kind , typename T, typename Sc, typename F> struct BlockScan
-    {
-        __device__ __forceinline__ BlockScan() {}
-        __device__ __forceinline__ BlockScan(const BlockScan& other) { CV_UNUSED(other); }
-
-        __device__ __forceinline__ T operator()(volatile T *ptr)
-        {
-            const unsigned int tid  = threadIdx.x;
-            const unsigned int lane = tid & warp_mask;
-            const unsigned int warp = tid >> warp_log;
-
-            Sc scan;
-            typename Sc::merge merge_scan;
-            const unsigned int idx = scan.index(tid);
-
-            T val = scan(ptr, idx);
-            __syncthreads ();
-
-            if( warp == 0)
-                scan.init(ptr);
-            __syncthreads ();
-
-            if( lane == 31 )
-                ptr [scan.warp_offset + warp ] = (Kind == INCLUSIVE) ? val : ptr [idx];
-            __syncthreads ();
-
-            if( warp == 0 )
-                merge_scan(ptr, idx);
-            __syncthreads();
-
-            if ( warp > 0)
-                val = ptr [scan.warp_offset + warp - 1] + val;
-            __syncthreads ();
-
-            ptr[idx] = val;
-            __syncthreads ();
-
-            return val ;
-        }
-
-        static const int warp_log  = 5;
-        static const int warp_mask = 31;
-    };
-
-    template <typename T>
-    __device__ T warpScanInclusive(T idata, volatile T* s_Data, unsigned int tid)
-    {
-    #if __CUDA_ARCH__ >= 300
-        const unsigned int laneId = cv::cuda::device::Warp::laneId();
-
-        // scan on shuffl functions
-        #pragma unroll
-        for (int i = 1; i <= (OPENCV_CUDA_WARP_SIZE / 2); i *= 2)
-        {
-            const T n = cv::cuda::device::shfl_up(idata, i);
-            if (laneId >= i)
-                  idata += n;
-        }
-
-        return idata;
-    #else
-        unsigned int pos = 2 * tid - (tid & (OPENCV_CUDA_WARP_SIZE - 1));
-        s_Data[pos] = 0;
-        pos += OPENCV_CUDA_WARP_SIZE;
-        s_Data[pos] = idata;
-
-        s_Data[pos] += s_Data[pos - 1];
-        s_Data[pos] += s_Data[pos - 2];
-        s_Data[pos] += s_Data[pos - 4];
-        s_Data[pos] += s_Data[pos - 8];
-        s_Data[pos] += s_Data[pos - 16];
-
-        return s_Data[pos];
-    #endif
-    }
-
-    template <typename T>
-    __device__ __forceinline__ T warpScanExclusive(T idata, volatile T* s_Data, unsigned int tid)
-    {
-        return warpScanInclusive(idata, s_Data, tid) - idata;
-    }
-
-    template <int tiNumScanThreads, typename T>
-    __device__ T blockScanInclusive(T idata, volatile T* s_Data, unsigned int tid)
-    {
-        if (tiNumScanThreads > OPENCV_CUDA_WARP_SIZE)
-        {
-            //Bottom-level inclusive warp scan
-            T warpResult = warpScanInclusive(idata, s_Data, tid);
-
-            //Save top elements of each warp for exclusive warp scan
-            //sync to wait for warp scans to complete (because s_Data is being overwritten)
-            __syncthreads();
-            if ((tid & (OPENCV_CUDA_WARP_SIZE - 1)) == (OPENCV_CUDA_WARP_SIZE - 1))
-            {
-                s_Data[tid >> OPENCV_CUDA_LOG_WARP_SIZE] = warpResult;
-            }
-
-            //wait for warp scans to complete
-            __syncthreads();
-
-            if (tid < (tiNumScanThreads / OPENCV_CUDA_WARP_SIZE) )
-            {
-                //grab top warp elements
-                T val = s_Data[tid];
-                //calculate exclusive scan and write back to shared memory
-                s_Data[tid] = warpScanExclusive(val, s_Data, tid);
-            }
-
-            //return updated warp scans with exclusive scan results
-            __syncthreads();
-
-            return warpResult + s_Data[tid >> OPENCV_CUDA_LOG_WARP_SIZE];
-        }
-        else
-        {
-            return warpScanInclusive(idata, s_Data, tid);
-        }
-    }
-}}}
-
-//! @endcond
-
-#endif // OPENCV_CUDA_SCAN_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/simd_functions.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/simd_functions.hpp
deleted file mode 100644
index 3d8c2e0d..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/simd_functions.hpp
+++ /dev/null
@@ -1,869 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-/*
- * Copyright (c) 2013 NVIDIA Corporation. All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *
- *   Redistributions of source code must retain the above copyright notice,
- *   this list of conditions and the following disclaimer.
- *
- *   Redistributions in binary form must reproduce the above copyright notice,
- *   this list of conditions and the following disclaimer in the documentation
- *   and/or other materials provided with the distribution.
- *
- *   Neither the name of NVIDIA Corporation nor the names of its contributors
- *   may be used to endorse or promote products derived from this software
- *   without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
- * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
- * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
- * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
- * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
- * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
- * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
- * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
- * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
- * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
- * POSSIBILITY OF SUCH DAMAGE.
- */
-
-#ifndef OPENCV_CUDA_SIMD_FUNCTIONS_HPP
-#define OPENCV_CUDA_SIMD_FUNCTIONS_HPP
-
-#include "common.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    // 2
-
-    static __device__ __forceinline__ unsigned int vadd2(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vadd2.u32.u32.u32.sat %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #elif __CUDA_ARCH__ >= 200
-        asm("vadd.u32.u32.u32.sat %0.h0, %1.h0, %2.h0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vadd.u32.u32.u32.sat %0.h1, %1.h1, %2.h1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int s;
-        s = a ^ b;          // sum bits
-        r = a + b;          // actual sum
-        s = s ^ r;          // determine carry-ins for each bit position
-        s = s & 0x00010000; // carry-in to high word (= carry-out from low word)
-        r = r - s;          // subtract out carry-out from low word
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vsub2(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vsub2.u32.u32.u32.sat %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #elif __CUDA_ARCH__ >= 200
-        asm("vsub.u32.u32.u32.sat %0.h0, %1.h0, %2.h0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vsub.u32.u32.u32.sat %0.h1, %1.h1, %2.h1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int s;
-        s = a ^ b;          // sum bits
-        r = a - b;          // actual sum
-        s = s ^ r;          // determine carry-ins for each bit position
-        s = s & 0x00010000; // borrow to high word
-        r = r + s;          // compensate for borrow from low word
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vabsdiff2(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vabsdiff2.u32.u32.u32.sat %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #elif __CUDA_ARCH__ >= 200
-        asm("vabsdiff.u32.u32.u32.sat %0.h0, %1.h0, %2.h0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vabsdiff.u32.u32.u32.sat %0.h1, %1.h1, %2.h1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int s, t, u, v;
-        s = a & 0x0000ffff; // extract low halfword
-        r = b & 0x0000ffff; // extract low halfword
-        u = ::max(r, s);    // maximum of low halfwords
-        v = ::min(r, s);    // minimum of low halfwords
-        s = a & 0xffff0000; // extract high halfword
-        r = b & 0xffff0000; // extract high halfword
-        t = ::max(r, s);    // maximum of high halfwords
-        s = ::min(r, s);    // minimum of high halfwords
-        r = u | t;          // maximum of both halfwords
-        s = v | s;          // minimum of both halfwords
-        r = r - s;          // |a - b| = max(a,b) - min(a,b);
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vavg2(unsigned int a, unsigned int b)
-    {
-        unsigned int r, s;
-
-        // HAKMEM #23: a + b = 2 * (a & b) + (a ^ b) ==>
-        // (a + b) / 2 = (a & b) + ((a ^ b) >> 1)
-        s = a ^ b;
-        r = a & b;
-        s = s & 0xfffefffe; // ensure shift doesn't cross halfword boundaries
-        s = s >> 1;
-        s = r + s;
-
-        return s;
-    }
-
-    static __device__ __forceinline__ unsigned int vavrg2(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vavrg2.u32.u32.u32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        // HAKMEM #23: a + b = 2 * (a | b) - (a ^ b) ==>
-        // (a + b + 1) / 2 = (a | b) - ((a ^ b) >> 1)
-        unsigned int s;
-        s = a ^ b;
-        r = a | b;
-        s = s & 0xfffefffe; // ensure shift doesn't cross half-word boundaries
-        s = s >> 1;
-        r = r - s;
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vseteq2(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vset2.u32.u32.eq %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        // inspired by Alan Mycroft's null-byte detection algorithm:
-        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
-        unsigned int c;
-        r = a ^ b;          // 0x0000 if a == b
-        c = r | 0x80008000; // set msbs, to catch carry out
-        r = r ^ c;          // extract msbs, msb = 1 if r < 0x8000
-        c = c - 0x00010001; // msb = 0, if r was 0x0000 or 0x8000
-        c = r & ~c;         // msb = 1, if r was 0x0000
-        r = c >> 15;        // convert to bool
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vcmpeq2(unsigned int a, unsigned int b)
-    {
-        unsigned int r, c;
-
-    #if __CUDA_ARCH__ >= 300
-        r = vseteq2(a, b);
-        c = r << 16;        // convert bool
-        r = c - r;          //  into mask
-    #else
-        // inspired by Alan Mycroft's null-byte detection algorithm:
-        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
-        r = a ^ b;          // 0x0000 if a == b
-        c = r | 0x80008000; // set msbs, to catch carry out
-        r = r ^ c;          // extract msbs, msb = 1 if r < 0x8000
-        c = c - 0x00010001; // msb = 0, if r was 0x0000 or 0x8000
-        c = r & ~c;         // msb = 1, if r was 0x0000
-        r = c >> 15;        // convert
-        r = c - r;          //  msbs to
-        r = c | r;          //   mask
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vsetge2(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vset2.u32.u32.ge %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int c;
-        asm("not.b32 %0, %0;" : "+r"(b));
-        c = vavrg2(a, b);   // (a + ~b + 1) / 2 = (a - b) / 2
-        c = c & 0x80008000; // msb = carry-outs
-        r = c >> 15;        // convert to bool
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vcmpge2(unsigned int a, unsigned int b)
-    {
-        unsigned int r, c;
-
-    #if __CUDA_ARCH__ >= 300
-        r = vsetge2(a, b);
-        c = r << 16;        // convert bool
-        r = c - r;          //  into mask
-    #else
-        asm("not.b32 %0, %0;" : "+r"(b));
-        c = vavrg2(a, b);   // (a + ~b + 1) / 2 = (a - b) / 2
-        c = c & 0x80008000; // msb = carry-outs
-        r = c >> 15;        // convert
-        r = c - r;          //  msbs to
-        r = c | r;          //   mask
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vsetgt2(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vset2.u32.u32.gt %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int c;
-        asm("not.b32 %0, %0;" : "+r"(b));
-        c = vavg2(a, b);    // (a + ~b) / 2 = (a - b) / 2 [rounded down]
-        c = c & 0x80008000; // msbs = carry-outs
-        r = c >> 15;        // convert to bool
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vcmpgt2(unsigned int a, unsigned int b)
-    {
-        unsigned int r, c;
-
-    #if __CUDA_ARCH__ >= 300
-        r = vsetgt2(a, b);
-        c = r << 16;        // convert bool
-        r = c - r;          //  into mask
-    #else
-        asm("not.b32 %0, %0;" : "+r"(b));
-        c = vavg2(a, b);    // (a + ~b) / 2 = (a - b) / 2 [rounded down]
-        c = c & 0x80008000; // msbs = carry-outs
-        r = c >> 15;        // convert
-        r = c - r;          //  msbs to
-        r = c | r;          //   mask
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vsetle2(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vset2.u32.u32.le %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int c;
-        asm("not.b32 %0, %0;" : "+r"(a));
-        c = vavrg2(a, b);   // (b + ~a + 1) / 2 = (b - a) / 2
-        c = c & 0x80008000; // msb = carry-outs
-        r = c >> 15;        // convert to bool
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vcmple2(unsigned int a, unsigned int b)
-    {
-        unsigned int r, c;
-
-    #if __CUDA_ARCH__ >= 300
-        r = vsetle2(a, b);
-        c = r << 16;        // convert bool
-        r = c - r;          //  into mask
-    #else
-        asm("not.b32 %0, %0;" : "+r"(a));
-        c = vavrg2(a, b);   // (b + ~a + 1) / 2 = (b - a) / 2
-        c = c & 0x80008000; // msb = carry-outs
-        r = c >> 15;        // convert
-        r = c - r;          //  msbs to
-        r = c | r;          //   mask
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vsetlt2(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vset2.u32.u32.lt %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int c;
-        asm("not.b32 %0, %0;" : "+r"(a));
-        c = vavg2(a, b);    // (b + ~a) / 2 = (b - a) / 2 [rounded down]
-        c = c & 0x80008000; // msb = carry-outs
-        r = c >> 15;        // convert to bool
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vcmplt2(unsigned int a, unsigned int b)
-    {
-        unsigned int r, c;
-
-    #if __CUDA_ARCH__ >= 300
-        r = vsetlt2(a, b);
-        c = r << 16;        // convert bool
-        r = c - r;          //  into mask
-    #else
-        asm("not.b32 %0, %0;" : "+r"(a));
-        c = vavg2(a, b);    // (b + ~a) / 2 = (b - a) / 2 [rounded down]
-        c = c & 0x80008000; // msb = carry-outs
-        r = c >> 15;        // convert
-        r = c - r;          //  msbs to
-        r = c | r;          //   mask
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vsetne2(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm ("vset2.u32.u32.ne %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        // inspired by Alan Mycroft's null-byte detection algorithm:
-        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
-        unsigned int c;
-        r = a ^ b;          // 0x0000 if a == b
-        c = r | 0x80008000; // set msbs, to catch carry out
-        c = c - 0x00010001; // msb = 0, if r was 0x0000 or 0x8000
-        c = r | c;          // msb = 1, if r was not 0x0000
-        c = c & 0x80008000; // extract msbs
-        r = c >> 15;        // convert to bool
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vcmpne2(unsigned int a, unsigned int b)
-    {
-        unsigned int r, c;
-
-    #if __CUDA_ARCH__ >= 300
-        r = vsetne2(a, b);
-        c = r << 16;        // convert bool
-        r = c - r;          //  into mask
-    #else
-        // inspired by Alan Mycroft's null-byte detection algorithm:
-        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
-        r = a ^ b;          // 0x0000 if a == b
-        c = r | 0x80008000; // set msbs, to catch carry out
-        c = c - 0x00010001; // msb = 0, if r was 0x0000 or 0x8000
-        c = r | c;          // msb = 1, if r was not 0x0000
-        c = c & 0x80008000; // extract msbs
-        r = c >> 15;        // convert
-        r = c - r;          //  msbs to
-        r = c | r;          //   mask
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vmax2(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vmax2.u32.u32.u32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #elif __CUDA_ARCH__ >= 200
-        asm("vmax.u32.u32.u32 %0.h0, %1.h0, %2.h0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vmax.u32.u32.u32 %0.h1, %1.h1, %2.h1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int s, t, u;
-        r = a & 0x0000ffff; // extract low halfword
-        s = b & 0x0000ffff; // extract low halfword
-        t = ::max(r, s);    // maximum of low halfwords
-        r = a & 0xffff0000; // extract high halfword
-        s = b & 0xffff0000; // extract high halfword
-        u = ::max(r, s);    // maximum of high halfwords
-        r = t | u;          // combine halfword maximums
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vmin2(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vmin2.u32.u32.u32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #elif __CUDA_ARCH__ >= 200
-        asm("vmin.u32.u32.u32 %0.h0, %1.h0, %2.h0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vmin.u32.u32.u32 %0.h1, %1.h1, %2.h1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int s, t, u;
-        r = a & 0x0000ffff; // extract low halfword
-        s = b & 0x0000ffff; // extract low halfword
-        t = ::min(r, s);    // minimum of low halfwords
-        r = a & 0xffff0000; // extract high halfword
-        s = b & 0xffff0000; // extract high halfword
-        u = ::min(r, s);    // minimum of high halfwords
-        r = t | u;          // combine halfword minimums
-    #endif
-
-        return r;
-    }
-
-    // 4
-
-    static __device__ __forceinline__ unsigned int vadd4(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vadd4.u32.u32.u32.sat %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #elif __CUDA_ARCH__ >= 200
-        asm("vadd.u32.u32.u32.sat %0.b0, %1.b0, %2.b0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vadd.u32.u32.u32.sat %0.b1, %1.b1, %2.b1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vadd.u32.u32.u32.sat %0.b2, %1.b2, %2.b2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vadd.u32.u32.u32.sat %0.b3, %1.b3, %2.b3, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int s, t;
-        s = a ^ b;          // sum bits
-        r = a & 0x7f7f7f7f; // clear msbs
-        t = b & 0x7f7f7f7f; // clear msbs
-        s = s & 0x80808080; // msb sum bits
-        r = r + t;          // add without msbs, record carry-out in msbs
-        r = r ^ s;          // sum of msb sum and carry-in bits, w/o carry-out
-    #endif /* __CUDA_ARCH__ >= 300 */
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vsub4(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vsub4.u32.u32.u32.sat %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #elif __CUDA_ARCH__ >= 200
-        asm("vsub.u32.u32.u32.sat %0.b0, %1.b0, %2.b0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vsub.u32.u32.u32.sat %0.b1, %1.b1, %2.b1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vsub.u32.u32.u32.sat %0.b2, %1.b2, %2.b2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vsub.u32.u32.u32.sat %0.b3, %1.b3, %2.b3, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int s, t;
-        s = a ^ ~b;         // inverted sum bits
-        r = a | 0x80808080; // set msbs
-        t = b & 0x7f7f7f7f; // clear msbs
-        s = s & 0x80808080; // inverted msb sum bits
-        r = r - t;          // subtract w/o msbs, record inverted borrows in msb
-        r = r ^ s;          // combine inverted msb sum bits and borrows
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vavg4(unsigned int a, unsigned int b)
-    {
-        unsigned int r, s;
-
-        // HAKMEM #23: a + b = 2 * (a & b) + (a ^ b) ==>
-        // (a + b) / 2 = (a & b) + ((a ^ b) >> 1)
-        s = a ^ b;
-        r = a & b;
-        s = s & 0xfefefefe; // ensure following shift doesn't cross byte boundaries
-        s = s >> 1;
-        s = r + s;
-
-        return s;
-    }
-
-    static __device__ __forceinline__ unsigned int vavrg4(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vavrg4.u32.u32.u32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        // HAKMEM #23: a + b = 2 * (a | b) - (a ^ b) ==>
-        // (a + b + 1) / 2 = (a | b) - ((a ^ b) >> 1)
-        unsigned int c;
-        c = a ^ b;
-        r = a | b;
-        c = c & 0xfefefefe; // ensure following shift doesn't cross byte boundaries
-        c = c >> 1;
-        r = r - c;
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vseteq4(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vset4.u32.u32.eq %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        // inspired by Alan Mycroft's null-byte detection algorithm:
-        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
-        unsigned int c;
-        r = a ^ b;          // 0x00 if a == b
-        c = r | 0x80808080; // set msbs, to catch carry out
-        r = r ^ c;          // extract msbs, msb = 1 if r < 0x80
-        c = c - 0x01010101; // msb = 0, if r was 0x00 or 0x80
-        c = r & ~c;         // msb = 1, if r was 0x00
-        r = c >> 7;         // convert to bool
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vcmpeq4(unsigned int a, unsigned int b)
-    {
-        unsigned int r, t;
-
-    #if __CUDA_ARCH__ >= 300
-        r = vseteq4(a, b);
-        t = r << 8;         // convert bool
-        r = t - r;          //  to mask
-    #else
-        // inspired by Alan Mycroft's null-byte detection algorithm:
-        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
-        t = a ^ b;          // 0x00 if a == b
-        r = t | 0x80808080; // set msbs, to catch carry out
-        t = t ^ r;          // extract msbs, msb = 1 if t < 0x80
-        r = r - 0x01010101; // msb = 0, if t was 0x00 or 0x80
-        r = t & ~r;         // msb = 1, if t was 0x00
-        t = r >> 7;         // build mask
-        t = r - t;          //  from
-        r = t | r;          //   msbs
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vsetle4(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vset4.u32.u32.le %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int c;
-        asm("not.b32 %0, %0;" : "+r"(a));
-        c = vavrg4(a, b);   // (b + ~a + 1) / 2 = (b - a) / 2
-        c = c & 0x80808080; // msb = carry-outs
-        r = c >> 7;         // convert to bool
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vcmple4(unsigned int a, unsigned int b)
-    {
-        unsigned int r, c;
-
-    #if __CUDA_ARCH__ >= 300
-        r = vsetle4(a, b);
-        c = r << 8;         // convert bool
-        r = c - r;          //  to mask
-    #else
-        asm("not.b32 %0, %0;" : "+r"(a));
-        c = vavrg4(a, b);   // (b + ~a + 1) / 2 = (b - a) / 2
-        c = c & 0x80808080; // msbs = carry-outs
-        r = c >> 7;         // convert
-        r = c - r;          //  msbs to
-        r = c | r;          //   mask
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vsetlt4(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vset4.u32.u32.lt %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int c;
-        asm("not.b32 %0, %0;" : "+r"(a));
-        c = vavg4(a, b);    // (b + ~a) / 2 = (b - a) / 2 [rounded down]
-        c = c & 0x80808080; // msb = carry-outs
-        r = c >> 7;         // convert to bool
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vcmplt4(unsigned int a, unsigned int b)
-    {
-        unsigned int r, c;
-
-    #if __CUDA_ARCH__ >= 300
-        r = vsetlt4(a, b);
-        c = r << 8;         // convert bool
-        r = c - r;          //  to mask
-    #else
-        asm("not.b32 %0, %0;" : "+r"(a));
-        c = vavg4(a, b);    // (b + ~a) / 2 = (b - a) / 2 [rounded down]
-        c = c & 0x80808080; // msbs = carry-outs
-        r = c >> 7;         // convert
-        r = c - r;          //  msbs to
-        r = c | r;          //   mask
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vsetge4(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vset4.u32.u32.ge %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int c;
-        asm("not.b32 %0, %0;" : "+r"(b));
-        c = vavrg4(a, b);   // (a + ~b + 1) / 2 = (a - b) / 2
-        c = c & 0x80808080; // msb = carry-outs
-        r = c >> 7;         // convert to bool
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vcmpge4(unsigned int a, unsigned int b)
-    {
-        unsigned int r, s;
-
-    #if __CUDA_ARCH__ >= 300
-        r = vsetge4(a, b);
-        s = r << 8;         // convert bool
-        r = s - r;          //  to mask
-    #else
-        asm ("not.b32 %0,%0;" : "+r"(b));
-        r = vavrg4 (a, b);  // (a + ~b + 1) / 2 = (a - b) / 2
-        r = r & 0x80808080; // msb = carry-outs
-        s = r >> 7;         // build mask
-        s = r - s;          //  from
-        r = s | r;          //   msbs
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vsetgt4(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vset4.u32.u32.gt %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int c;
-        asm("not.b32 %0, %0;" : "+r"(b));
-        c = vavg4(a, b);    // (a + ~b) / 2 = (a - b) / 2 [rounded down]
-        c = c & 0x80808080; // msb = carry-outs
-        r = c >> 7;         // convert to bool
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vcmpgt4(unsigned int a, unsigned int b)
-    {
-        unsigned int r, c;
-
-    #if __CUDA_ARCH__ >= 300
-        r = vsetgt4(a, b);
-        c = r << 8;         // convert bool
-        r = c - r;          //  to mask
-    #else
-        asm("not.b32 %0, %0;" : "+r"(b));
-        c = vavg4(a, b);    // (a + ~b) / 2 = (a - b) / 2 [rounded down]
-        c = c & 0x80808080; // msb = carry-outs
-        r = c >> 7;         // convert
-        r = c - r;          //  msbs to
-        r = c | r;          //   mask
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vsetne4(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vset4.u32.u32.ne %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        // inspired by Alan Mycroft's null-byte detection algorithm:
-        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
-        unsigned int c;
-        r = a ^ b;          // 0x00 if a == b
-        c = r | 0x80808080; // set msbs, to catch carry out
-        c = c - 0x01010101; // msb = 0, if r was 0x00 or 0x80
-        c = r | c;          // msb = 1, if r was not 0x00
-        c = c & 0x80808080; // extract msbs
-        r = c >> 7;         // convert to bool
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vcmpne4(unsigned int a, unsigned int b)
-    {
-        unsigned int r, c;
-
-    #if __CUDA_ARCH__ >= 300
-        r = vsetne4(a, b);
-        c = r << 8;         // convert bool
-        r = c - r;          //  to mask
-    #else
-        // inspired by Alan Mycroft's null-byte detection algorithm:
-        // null_byte(x) = ((x - 0x01010101) & (~x & 0x80808080))
-        r = a ^ b;          // 0x00 if a == b
-        c = r | 0x80808080; // set msbs, to catch carry out
-        c = c - 0x01010101; // msb = 0, if r was 0x00 or 0x80
-        c = r | c;          // msb = 1, if r was not 0x00
-        c = c & 0x80808080; // extract msbs
-        r = c >> 7;         // convert
-        r = c - r;          //  msbs to
-        r = c | r;          //   mask
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vabsdiff4(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vabsdiff4.u32.u32.u32.sat %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #elif __CUDA_ARCH__ >= 200
-        asm("vabsdiff.u32.u32.u32.sat %0.b0, %1.b0, %2.b0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vabsdiff.u32.u32.u32.sat %0.b1, %1.b1, %2.b1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vabsdiff.u32.u32.u32.sat %0.b2, %1.b2, %2.b2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vabsdiff.u32.u32.u32.sat %0.b3, %1.b3, %2.b3, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int s;
-        s = vcmpge4(a, b);  // mask = 0xff if a >= b
-        r = a ^ b;          //
-        s = (r &  s) ^ b;   // select a when a >= b, else select b => max(a,b)
-        r = s ^ r;          // select a when b >= a, else select b => min(a,b)
-        r = s - r;          // |a - b| = max(a,b) - min(a,b);
-    #endif
-
-        return r;
-    }
-
-    static __device__ __forceinline__ unsigned int vmax4(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vmax4.u32.u32.u32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #elif __CUDA_ARCH__ >= 200
-        asm("vmax.u32.u32.u32 %0.b0, %1.b0, %2.b0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vmax.u32.u32.u32 %0.b1, %1.b1, %2.b1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vmax.u32.u32.u32 %0.b2, %1.b2, %2.b2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vmax.u32.u32.u32 %0.b3, %1.b3, %2.b3, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int s;
-        s = vcmpge4(a, b);  // mask = 0xff if a >= b
-        r = a & s;          // select a when b >= a
-        s = b & ~s;         // select b when b < a
-        r = r | s;          // combine byte selections
-    #endif
-
-        return r;           // byte-wise unsigned maximum
-    }
-
-    static __device__ __forceinline__ unsigned int vmin4(unsigned int a, unsigned int b)
-    {
-        unsigned int r = 0;
-
-    #if __CUDA_ARCH__ >= 300
-        asm("vmin4.u32.u32.u32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #elif __CUDA_ARCH__ >= 200
-        asm("vmin.u32.u32.u32 %0.b0, %1.b0, %2.b0, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vmin.u32.u32.u32 %0.b1, %1.b1, %2.b1, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vmin.u32.u32.u32 %0.b2, %1.b2, %2.b2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-        asm("vmin.u32.u32.u32 %0.b3, %1.b3, %2.b3, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(r));
-    #else
-        unsigned int s;
-        s = vcmpge4(b, a);  // mask = 0xff if a >= b
-        r = a & s;          // select a when b >= a
-        s = b & ~s;         // select b when b < a
-        r = r | s;          // combine byte selections
-    #endif
-
-        return r;
-    }
-}}}
-
-//! @endcond
-
-#endif // OPENCV_CUDA_SIMD_FUNCTIONS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/transform.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/transform.hpp
deleted file mode 100644
index 42aa6ea1..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/transform.hpp
+++ /dev/null
@@ -1,75 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_TRANSFORM_HPP
-#define OPENCV_CUDA_TRANSFORM_HPP
-
-#include "common.hpp"
-#include "utility.hpp"
-#include "detail/transform_detail.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    template <typename T, typename D, typename UnOp, typename Mask>
-    static inline void transform(PtrStepSz<T> src, PtrStepSz<D> dst, UnOp op, const Mask& mask, cudaStream_t stream)
-    {
-        typedef TransformFunctorTraits<UnOp> ft;
-        transform_detail::TransformDispatcher<VecTraits<T>::cn == 1 && VecTraits<D>::cn == 1 && ft::smart_shift != 1>::call(src, dst, op, mask, stream);
-    }
-
-    template <typename T1, typename T2, typename D, typename BinOp, typename Mask>
-    static inline void transform(PtrStepSz<T1> src1, PtrStepSz<T2> src2, PtrStepSz<D> dst, BinOp op, const Mask& mask, cudaStream_t stream)
-    {
-        typedef TransformFunctorTraits<BinOp> ft;
-        transform_detail::TransformDispatcher<VecTraits<T1>::cn == 1 && VecTraits<T2>::cn == 1 && VecTraits<D>::cn == 1 && ft::smart_shift != 1>::call(src1, src2, dst, op, mask, stream);
-    }
-}}}
-
-//! @endcond
-
-#endif // OPENCV_CUDA_TRANSFORM_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/type_traits.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/type_traits.hpp
deleted file mode 100644
index 8b7a3fd1..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/type_traits.hpp
+++ /dev/null
@@ -1,90 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_TYPE_TRAITS_HPP
-#define OPENCV_CUDA_TYPE_TRAITS_HPP
-
-#include "detail/type_traits_detail.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    template <typename T> struct IsSimpleParameter
-    {
-        enum {value = type_traits_detail::IsIntegral<T>::value || type_traits_detail::IsFloat<T>::value ||
-            type_traits_detail::PointerTraits<typename type_traits_detail::ReferenceTraits<T>::type>::value};
-    };
-
-    template <typename T> struct TypeTraits
-    {
-        typedef typename type_traits_detail::UnConst<T>::type                                                NonConstType;
-        typedef typename type_traits_detail::UnVolatile<T>::type                                             NonVolatileType;
-        typedef typename type_traits_detail::UnVolatile<typename type_traits_detail::UnConst<T>::type>::type UnqualifiedType;
-        typedef typename type_traits_detail::PointerTraits<UnqualifiedType>::type                            PointeeType;
-        typedef typename type_traits_detail::ReferenceTraits<T>::type                                        ReferredType;
-
-        enum { isConst          = type_traits_detail::UnConst<T>::value };
-        enum { isVolatile       = type_traits_detail::UnVolatile<T>::value };
-
-        enum { isReference      = type_traits_detail::ReferenceTraits<UnqualifiedType>::value };
-        enum { isPointer        = type_traits_detail::PointerTraits<typename type_traits_detail::ReferenceTraits<UnqualifiedType>::type>::value };
-
-        enum { isUnsignedInt    = type_traits_detail::IsUnsignedIntegral<UnqualifiedType>::value };
-        enum { isSignedInt      = type_traits_detail::IsSignedIntergral<UnqualifiedType>::value };
-        enum { isIntegral       = type_traits_detail::IsIntegral<UnqualifiedType>::value };
-        enum { isFloat          = type_traits_detail::IsFloat<UnqualifiedType>::value };
-        enum { isArith          = isIntegral || isFloat };
-        enum { isVec            = type_traits_detail::IsVec<UnqualifiedType>::value };
-
-        typedef typename type_traits_detail::Select<IsSimpleParameter<UnqualifiedType>::value,
-            T, typename type_traits_detail::AddParameterType<T>::type>::type ParameterType;
-    };
-}}}
-
-//! @endcond
-
-#endif // OPENCV_CUDA_TYPE_TRAITS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/utility.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/utility.hpp
deleted file mode 100644
index 7f5db48a..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/utility.hpp
+++ /dev/null
@@ -1,230 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_UTILITY_HPP
-#define OPENCV_CUDA_UTILITY_HPP
-
-#include "saturate_cast.hpp"
-#include "datamov_utils.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    struct CV_EXPORTS ThrustAllocator
-    {
-        typedef uchar value_type;
-        virtual ~ThrustAllocator();
-        virtual __device__ __host__ uchar* allocate(size_t numBytes) = 0;
-        virtual __device__ __host__ void deallocate(uchar* ptr, size_t numBytes) = 0;
-        static ThrustAllocator& getAllocator();
-        static void setAllocator(ThrustAllocator* allocator);
-    };
-    #define OPENCV_CUDA_LOG_WARP_SIZE        (5)
-    #define OPENCV_CUDA_WARP_SIZE            (1 << OPENCV_CUDA_LOG_WARP_SIZE)
-    #define OPENCV_CUDA_LOG_MEM_BANKS        ((__CUDA_ARCH__ >= 200) ? 5 : 4) // 32 banks on fermi, 16 on tesla
-    #define OPENCV_CUDA_MEM_BANKS            (1 << OPENCV_CUDA_LOG_MEM_BANKS)
-
-    ///////////////////////////////////////////////////////////////////////////////
-    // swap
-
-    template <typename T> void __device__ __host__ __forceinline__ swap(T& a, T& b)
-    {
-        const T temp = a;
-        a = b;
-        b = temp;
-    }
-
-    ///////////////////////////////////////////////////////////////////////////////
-    // Mask Reader
-
-    struct SingleMask
-    {
-        explicit __host__ __device__ __forceinline__ SingleMask(PtrStepb mask_) : mask(mask_) {}
-        __host__ __device__ __forceinline__ SingleMask(const SingleMask& mask_): mask(mask_.mask){}
-
-        __device__ __forceinline__ bool operator()(int y, int x) const
-        {
-            return mask.ptr(y)[x] != 0;
-        }
-
-        PtrStepb mask;
-    };
-
-    struct SingleMaskChannels
-    {
-        __host__ __device__ __forceinline__ SingleMaskChannels(PtrStepb mask_, int channels_)
-        : mask(mask_), channels(channels_) {}
-        __host__ __device__ __forceinline__ SingleMaskChannels(const SingleMaskChannels& mask_)
-            :mask(mask_.mask), channels(mask_.channels){}
-
-        __device__ __forceinline__ bool operator()(int y, int x) const
-        {
-            return mask.ptr(y)[x / channels] != 0;
-        }
-
-        PtrStepb mask;
-        int channels;
-    };
-
-    struct MaskCollection
-    {
-        explicit __host__ __device__ __forceinline__ MaskCollection(PtrStepb* maskCollection_)
-            : maskCollection(maskCollection_) {}
-
-        __device__ __forceinline__ MaskCollection(const MaskCollection& masks_)
-            : maskCollection(masks_.maskCollection), curMask(masks_.curMask){}
-
-        __device__ __forceinline__ void next()
-        {
-            curMask = *maskCollection++;
-        }
-        __device__ __forceinline__ void setMask(int z)
-        {
-            curMask = maskCollection[z];
-        }
-
-        __device__ __forceinline__ bool operator()(int y, int x) const
-        {
-            uchar val;
-            return curMask.data == 0 || (ForceGlob<uchar>::Load(curMask.ptr(y), x, val), (val != 0));
-        }
-
-        const PtrStepb* maskCollection;
-        PtrStepb curMask;
-    };
-
-    struct WithOutMask
-    {
-        __host__ __device__ __forceinline__ WithOutMask(){}
-        __host__ __device__ __forceinline__ WithOutMask(const WithOutMask&){}
-
-        __device__ __forceinline__ void next() const
-        {
-        }
-        __device__ __forceinline__ void setMask(int) const
-        {
-        }
-
-        __device__ __forceinline__ bool operator()(int, int) const
-        {
-            return true;
-        }
-
-        __device__ __forceinline__ bool operator()(int, int, int) const
-        {
-            return true;
-        }
-
-        static __device__ __forceinline__ bool check(int, int)
-        {
-            return true;
-        }
-
-        static __device__ __forceinline__ bool check(int, int, int)
-        {
-            return true;
-        }
-    };
-
-    ///////////////////////////////////////////////////////////////////////////////
-    // Solve linear system
-
-    // solve 2x2 linear system Ax=b
-    template <typename T> __device__ __forceinline__ bool solve2x2(const T A[2][2], const T b[2], T x[2])
-    {
-        T det = A[0][0] * A[1][1] - A[1][0] * A[0][1];
-
-        if (det != 0)
-        {
-            double invdet = 1.0 / det;
-
-            x[0] = saturate_cast<T>(invdet * (b[0] * A[1][1] - b[1] * A[0][1]));
-
-            x[1] = saturate_cast<T>(invdet * (A[0][0] * b[1] - A[1][0] * b[0]));
-
-            return true;
-        }
-
-        return false;
-    }
-
-    // solve 3x3 linear system Ax=b
-    template <typename T> __device__ __forceinline__ bool solve3x3(const T A[3][3], const T b[3], T x[3])
-    {
-        T det = A[0][0] * (A[1][1] * A[2][2] - A[1][2] * A[2][1])
-              - A[0][1] * (A[1][0] * A[2][2] - A[1][2] * A[2][0])
-              + A[0][2] * (A[1][0] * A[2][1] - A[1][1] * A[2][0]);
-
-        if (det != 0)
-        {
-            double invdet = 1.0 / det;
-
-            x[0] = saturate_cast<T>(invdet *
-                (b[0]    * (A[1][1] * A[2][2] - A[1][2] * A[2][1]) -
-                 A[0][1] * (b[1]    * A[2][2] - A[1][2] * b[2]   ) +
-                 A[0][2] * (b[1]    * A[2][1] - A[1][1] * b[2]   )));
-
-            x[1] = saturate_cast<T>(invdet *
-                (A[0][0] * (b[1]    * A[2][2] - A[1][2] * b[2]   ) -
-                 b[0]    * (A[1][0] * A[2][2] - A[1][2] * A[2][0]) +
-                 A[0][2] * (A[1][0] * b[2]    - b[1]    * A[2][0])));
-
-            x[2] = saturate_cast<T>(invdet *
-                (A[0][0] * (A[1][1] * b[2]    - b[1]    * A[2][1]) -
-                 A[0][1] * (A[1][0] * b[2]    - b[1]    * A[2][0]) +
-                 b[0]    * (A[1][0] * A[2][1] - A[1][1] * A[2][0])));
-
-            return true;
-        }
-
-        return false;
-    }
-}}} // namespace cv { namespace cuda { namespace cudev
-
-//! @endcond
-
-#endif // OPENCV_CUDA_UTILITY_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/vec_distance.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/vec_distance.hpp
deleted file mode 100644
index ef6e5108..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/vec_distance.hpp
+++ /dev/null
@@ -1,232 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_VEC_DISTANCE_HPP
-#define OPENCV_CUDA_VEC_DISTANCE_HPP
-
-#include "reduce.hpp"
-#include "functional.hpp"
-#include "detail/vec_distance_detail.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    template <typename T> struct L1Dist
-    {
-        typedef int value_type;
-        typedef int result_type;
-
-        __device__ __forceinline__ L1Dist() : mySum(0) {}
-
-        __device__ __forceinline__ void reduceIter(int val1, int val2)
-        {
-            mySum = __sad(val1, val2, mySum);
-        }
-
-        template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
-        {
-            reduce<THREAD_DIM>(smem, mySum, tid, plus<int>());
-        }
-
-        __device__ __forceinline__ operator int() const
-        {
-            return mySum;
-        }
-
-        int mySum;
-    };
-    template <> struct L1Dist<float>
-    {
-        typedef float value_type;
-        typedef float result_type;
-
-        __device__ __forceinline__ L1Dist() : mySum(0.0f) {}
-
-        __device__ __forceinline__ void reduceIter(float val1, float val2)
-        {
-            mySum += ::fabs(val1 - val2);
-        }
-
-        template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
-        {
-            reduce<THREAD_DIM>(smem, mySum, tid, plus<float>());
-        }
-
-        __device__ __forceinline__ operator float() const
-        {
-            return mySum;
-        }
-
-        float mySum;
-    };
-
-    struct L2Dist
-    {
-        typedef float value_type;
-        typedef float result_type;
-
-        __device__ __forceinline__ L2Dist() : mySum(0.0f) {}
-
-        __device__ __forceinline__ void reduceIter(float val1, float val2)
-        {
-            float reg = val1 - val2;
-            mySum += reg * reg;
-        }
-
-        template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(float* smem, int tid)
-        {
-            reduce<THREAD_DIM>(smem, mySum, tid, plus<float>());
-        }
-
-        __device__ __forceinline__ operator float() const
-        {
-            return sqrtf(mySum);
-        }
-
-        float mySum;
-    };
-
-    struct HammingDist
-    {
-        typedef int value_type;
-        typedef int result_type;
-
-        __device__ __forceinline__ HammingDist() : mySum(0) {}
-
-        __device__ __forceinline__ void reduceIter(int val1, int val2)
-        {
-            mySum += __popc(val1 ^ val2);
-        }
-
-        template <int THREAD_DIM> __device__ __forceinline__ void reduceAll(int* smem, int tid)
-        {
-            reduce<THREAD_DIM>(smem, mySum, tid, plus<int>());
-        }
-
-        __device__ __forceinline__ operator int() const
-        {
-            return mySum;
-        }
-
-        int mySum;
-    };
-
-    // calc distance between two vectors in global memory
-    template <int THREAD_DIM, typename Dist, typename T1, typename T2>
-    __device__ void calcVecDiffGlobal(const T1* vec1, const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid)
-    {
-        for (int i = tid; i < len; i += THREAD_DIM)
-        {
-            T1 val1;
-            ForceGlob<T1>::Load(vec1, i, val1);
-
-            T2 val2;
-            ForceGlob<T2>::Load(vec2, i, val2);
-
-            dist.reduceIter(val1, val2);
-        }
-
-        dist.reduceAll<THREAD_DIM>(smem, tid);
-    }
-
-    // calc distance between two vectors, first vector is cached in register or shared memory, second vector is in global memory
-    template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN, typename Dist, typename T1, typename T2>
-    __device__ __forceinline__ void calcVecDiffCached(const T1* vecCached, const T2* vecGlob, int len, Dist& dist, typename Dist::result_type* smem, int tid)
-    {
-        vec_distance_detail::VecDiffCachedCalculator<THREAD_DIM, MAX_LEN, LEN_EQ_MAX_LEN>::calc(vecCached, vecGlob, len, dist, tid);
-
-        dist.reduceAll<THREAD_DIM>(smem, tid);
-    }
-
-    // calc distance between two vectors in global memory
-    template <int THREAD_DIM, typename T1> struct VecDiffGlobal
-    {
-        explicit __device__ __forceinline__ VecDiffGlobal(const T1* vec1_, int = 0, void* = 0, int = 0, int = 0)
-        {
-            vec1 = vec1_;
-        }
-
-        template <typename T2, typename Dist>
-        __device__ __forceinline__ void calc(const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid) const
-        {
-            calcVecDiffGlobal<THREAD_DIM>(vec1, vec2, len, dist, smem, tid);
-        }
-
-        const T1* vec1;
-    };
-
-    // calc distance between two vectors, first vector is cached in register memory, second vector is in global memory
-    template <int THREAD_DIM, int MAX_LEN, bool LEN_EQ_MAX_LEN, typename U> struct VecDiffCachedRegister
-    {
-        template <typename T1> __device__ __forceinline__ VecDiffCachedRegister(const T1* vec1, int len, U* smem, int glob_tid, int tid)
-        {
-            if (glob_tid < len)
-                smem[glob_tid] = vec1[glob_tid];
-            __syncthreads();
-
-            U* vec1ValsPtr = vec1Vals;
-
-            #pragma unroll
-            for (int i = tid; i < MAX_LEN; i += THREAD_DIM)
-                *vec1ValsPtr++ = smem[i];
-
-            __syncthreads();
-        }
-
-        template <typename T2, typename Dist>
-        __device__ __forceinline__ void calc(const T2* vec2, int len, Dist& dist, typename Dist::result_type* smem, int tid) const
-        {
-            calcVecDiffCached<THREAD_DIM, MAX_LEN, LEN_EQ_MAX_LEN>(vec1Vals, vec2, len, dist, smem, tid);
-        }
-
-        U vec1Vals[MAX_LEN / THREAD_DIM];
-    };
-}}} // namespace cv { namespace cuda { namespace cudev
-
-//! @endcond
-
-#endif // OPENCV_CUDA_VEC_DISTANCE_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/vec_math.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/vec_math.hpp
deleted file mode 100644
index 80b13036..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/vec_math.hpp
+++ /dev/null
@@ -1,923 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_VECMATH_HPP
-#define OPENCV_CUDA_VECMATH_HPP
-
-#include "vec_traits.hpp"
-#include "saturate_cast.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-
-// saturate_cast
-
-namespace vec_math_detail
-{
-    template <int cn, typename VecD> struct SatCastHelper;
-    template <typename VecD> struct SatCastHelper<1, VecD>
-    {
-        template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
-        {
-            typedef typename VecTraits<VecD>::elem_type D;
-            return VecTraits<VecD>::make(saturate_cast<D>(v.x));
-        }
-    };
-    template <typename VecD> struct SatCastHelper<2, VecD>
-    {
-        template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
-        {
-            typedef typename VecTraits<VecD>::elem_type D;
-            return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y));
-        }
-    };
-    template <typename VecD> struct SatCastHelper<3, VecD>
-    {
-        template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
-        {
-            typedef typename VecTraits<VecD>::elem_type D;
-            return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y), saturate_cast<D>(v.z));
-        }
-    };
-    template <typename VecD> struct SatCastHelper<4, VecD>
-    {
-        template <typename VecS> static __device__ __forceinline__ VecD cast(const VecS& v)
-        {
-            typedef typename VecTraits<VecD>::elem_type D;
-            return VecTraits<VecD>::make(saturate_cast<D>(v.x), saturate_cast<D>(v.y), saturate_cast<D>(v.z), saturate_cast<D>(v.w));
-        }
-    };
-
-    template <typename VecD, typename VecS> static __device__ __forceinline__ VecD saturate_cast_helper(const VecS& v)
-    {
-        return SatCastHelper<VecTraits<VecD>::cn, VecD>::cast(v);
-    }
-}
-
-template<typename T> static __device__ __forceinline__ T saturate_cast(const uchar1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const char1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const ushort1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const short1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const uint1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const int1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const float1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const double1& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-
-template<typename T> static __device__ __forceinline__ T saturate_cast(const uchar2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const char2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const ushort2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const short2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const uint2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const int2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const float2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const double2& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-
-template<typename T> static __device__ __forceinline__ T saturate_cast(const uchar3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const char3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const ushort3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const short3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const uint3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const int3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const float3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const double3& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-
-template<typename T> static __device__ __forceinline__ T saturate_cast(const uchar4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const char4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const ushort4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const short4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const uint4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const int4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const float4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-template<typename T> static __device__ __forceinline__ T saturate_cast(const double4& v) {return vec_math_detail::saturate_cast_helper<T>(v);}
-
-// unary operators
-
-#define CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(op, input_type, output_type) \
-    __device__ __forceinline__ output_type ## 1 operator op(const input_type ## 1 & a) \
-    { \
-        return VecTraits<output_type ## 1>::make(op (a.x)); \
-    } \
-    __device__ __forceinline__ output_type ## 2 operator op(const input_type ## 2 & a) \
-    { \
-        return VecTraits<output_type ## 2>::make(op (a.x), op (a.y)); \
-    } \
-    __device__ __forceinline__ output_type ## 3 operator op(const input_type ## 3 & a) \
-    { \
-        return VecTraits<output_type ## 3>::make(op (a.x), op (a.y), op (a.z)); \
-    } \
-    __device__ __forceinline__ output_type ## 4 operator op(const input_type ## 4 & a) \
-    { \
-        return VecTraits<output_type ## 4>::make(op (a.x), op (a.y), op (a.z), op (a.w)); \
-    }
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(-, char, char)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(-, short, short)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(-, int, int)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(-, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(-, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, uchar, uchar)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, char, uchar)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, ushort, uchar)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, short, uchar)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, int, uchar)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, uint, uchar)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, float, uchar)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(!, double, uchar)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(~, uchar, uchar)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(~, char, char)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(~, ushort, ushort)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(~, short, short)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(~, int, int)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_OP(~, uint, uint)
-
-#undef CV_CUDEV_IMPLEMENT_VEC_UNARY_OP
-
-// unary functions
-
-#define CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(func_name, func, input_type, output_type) \
-    __device__ __forceinline__ output_type ## 1 func_name(const input_type ## 1 & a) \
-    { \
-        return VecTraits<output_type ## 1>::make(func (a.x)); \
-    } \
-    __device__ __forceinline__ output_type ## 2 func_name(const input_type ## 2 & a) \
-    { \
-        return VecTraits<output_type ## 2>::make(func (a.x), func (a.y)); \
-    } \
-    __device__ __forceinline__ output_type ## 3 func_name(const input_type ## 3 & a) \
-    { \
-        return VecTraits<output_type ## 3>::make(func (a.x), func (a.y), func (a.z)); \
-    } \
-    __device__ __forceinline__ output_type ## 4 func_name(const input_type ## 4 & a) \
-    { \
-        return VecTraits<output_type ## 4>::make(func (a.x), func (a.y), func (a.z), func (a.w)); \
-    }
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(abs, ::fabsf, float, float)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrtf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrtf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrtf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrtf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrtf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrtf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrtf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sqrt, ::sqrt, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::expf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::expf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::expf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::expf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::expf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::expf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::expf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp, ::exp, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2f, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2f, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2f, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2f, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2f, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2f, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2f, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp2, ::exp2, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10f, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10f, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10f, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10f, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10f, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10f, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10f, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(exp10, ::exp10, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::logf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::logf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::logf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::logf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::logf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::logf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::logf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log, ::log, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2f, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2f, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2f, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2f, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2f, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2f, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2f, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log2, ::log2, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10f, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10f, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10f, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10f, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10f, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10f, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10f, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(log10, ::log10, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sinf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sinf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sinf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sinf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sinf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sinf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sinf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sin, ::sin, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cosf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cosf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cosf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cosf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cosf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cosf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cosf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cos, ::cos, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tanf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tanf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tanf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tanf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tanf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tanf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tanf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tan, ::tan, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asinf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asinf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asinf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asinf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asinf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asinf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asinf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asin, ::asin, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acosf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acosf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acosf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acosf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acosf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acosf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acosf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acos, ::acos, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atanf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atanf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atanf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atanf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atanf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atanf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atanf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atan, ::atan, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinhf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinhf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinhf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinhf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinhf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinhf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinhf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(sinh, ::sinh, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::coshf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::coshf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::coshf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::coshf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::coshf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::coshf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::coshf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(cosh, ::cosh, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanhf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanhf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanhf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanhf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanhf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanhf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanhf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(tanh, ::tanh, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinhf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinhf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinhf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinhf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinhf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinhf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinhf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(asinh, ::asinh, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acoshf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acoshf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acoshf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acoshf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acoshf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acoshf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acoshf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(acosh, ::acosh, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanhf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanhf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanhf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanhf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanhf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanhf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanhf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC(atanh, ::atanh, double, double)
-
-#undef CV_CUDEV_IMPLEMENT_VEC_UNARY_FUNC
-
-// binary operators (vec & vec)
-
-#define CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(op, input_type, output_type) \
-    __device__ __forceinline__ output_type ## 1 operator op(const input_type ## 1 & a, const input_type ## 1 & b) \
-    { \
-        return VecTraits<output_type ## 1>::make(a.x op b.x); \
-    } \
-    __device__ __forceinline__ output_type ## 2 operator op(const input_type ## 2 & a, const input_type ## 2 & b) \
-    { \
-        return VecTraits<output_type ## 2>::make(a.x op b.x, a.y op b.y); \
-    } \
-    __device__ __forceinline__ output_type ## 3 operator op(const input_type ## 3 & a, const input_type ## 3 & b) \
-    { \
-        return VecTraits<output_type ## 3>::make(a.x op b.x, a.y op b.y, a.z op b.z); \
-    } \
-    __device__ __forceinline__ output_type ## 4 operator op(const input_type ## 4 & a, const input_type ## 4 & b) \
-    { \
-        return VecTraits<output_type ## 4>::make(a.x op b.x, a.y op b.y, a.z op b.z, a.w op b.w); \
-    }
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, uchar, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, char, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, ushort, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, short, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, int, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, uint, uint)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, float, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(+, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, uchar, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, char, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, ushort, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, short, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, int, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, uint, uint)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, float, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(-, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, uchar, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, char, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, ushort, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, short, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, int, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, uint, uint)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, float, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(*, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, uchar, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, char, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, ushort, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, short, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, int, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, uint, uint)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, float, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(/, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, uchar, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, char, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, ushort, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, short, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, int, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, uint, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, float, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(==, double, uchar)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, uchar, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, char, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, ushort, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, short, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, int, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, uint, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, float, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(!=, double, uchar)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, uchar, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, char, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, ushort, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, short, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, int, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, uint, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, float, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>, double, uchar)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, uchar, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, char, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, ushort, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, short, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, int, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, uint, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, float, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<, double, uchar)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, uchar, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, char, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, ushort, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, short, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, int, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, uint, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, float, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(>=, double, uchar)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, uchar, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, char, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, ushort, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, short, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, int, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, uint, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, float, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(<=, double, uchar)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, uchar, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, char, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, ushort, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, short, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, int, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, uint, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, float, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&&, double, uchar)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, uchar, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, char, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, ushort, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, short, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, int, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, uint, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, float, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(||, double, uchar)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&, uchar, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&, char, char)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&, ushort, ushort)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&, short, short)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&, int, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(&, uint, uint)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(|, uchar, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(|, char, char)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(|, ushort, ushort)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(|, short, short)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(|, int, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(|, uint, uint)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(^, uchar, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(^, char, char)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(^, ushort, ushort)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(^, short, short)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(^, int, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_OP(^, uint, uint)
-
-#undef CV_CUDEV_IMPLEMENT_VEC_BINARY_OP
-
-// binary operators (vec & scalar)
-
-#define CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(op, input_type, scalar_type, output_type) \
-    __device__ __forceinline__ output_type ## 1 operator op(const input_type ## 1 & a, scalar_type s) \
-    { \
-        return VecTraits<output_type ## 1>::make(a.x op s); \
-    } \
-    __device__ __forceinline__ output_type ## 1 operator op(scalar_type s, const input_type ## 1 & b) \
-    { \
-        return VecTraits<output_type ## 1>::make(s op b.x); \
-    } \
-    __device__ __forceinline__ output_type ## 2 operator op(const input_type ## 2 & a, scalar_type s) \
-    { \
-        return VecTraits<output_type ## 2>::make(a.x op s, a.y op s); \
-    } \
-    __device__ __forceinline__ output_type ## 2 operator op(scalar_type s, const input_type ## 2 & b) \
-    { \
-        return VecTraits<output_type ## 2>::make(s op b.x, s op b.y); \
-    } \
-    __device__ __forceinline__ output_type ## 3 operator op(const input_type ## 3 & a, scalar_type s) \
-    { \
-        return VecTraits<output_type ## 3>::make(a.x op s, a.y op s, a.z op s); \
-    } \
-    __device__ __forceinline__ output_type ## 3 operator op(scalar_type s, const input_type ## 3 & b) \
-    { \
-        return VecTraits<output_type ## 3>::make(s op b.x, s op b.y, s op b.z); \
-    } \
-    __device__ __forceinline__ output_type ## 4 operator op(const input_type ## 4 & a, scalar_type s) \
-    { \
-        return VecTraits<output_type ## 4>::make(a.x op s, a.y op s, a.z op s, a.w op s); \
-    } \
-    __device__ __forceinline__ output_type ## 4 operator op(scalar_type s, const input_type ## 4 & b) \
-    { \
-        return VecTraits<output_type ## 4>::make(s op b.x, s op b.y, s op b.z, s op b.w); \
-    }
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, uchar, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, uchar, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, uchar, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, char, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, char, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, char, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, ushort, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, ushort, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, ushort, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, short, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, short, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, short, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, int, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, int, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, int, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, uint, uint, uint)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, uint, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, uint, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, float, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, float, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(+, double, double, double)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, uchar, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, uchar, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, uchar, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, char, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, char, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, char, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, ushort, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, ushort, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, ushort, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, short, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, short, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, short, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, int, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, int, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, int, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, uint, uint, uint)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, uint, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, uint, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, float, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, float, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(-, double, double, double)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, uchar, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, uchar, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, uchar, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, char, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, char, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, char, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, ushort, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, ushort, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, ushort, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, short, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, short, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, short, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, int, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, int, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, int, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, uint, uint, uint)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, uint, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, uint, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, float, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, float, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(*, double, double, double)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, uchar, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, uchar, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, uchar, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, char, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, char, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, char, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, ushort, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, ushort, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, ushort, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, short, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, short, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, short, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, int, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, int, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, int, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, uint, uint, uint)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, uint, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, uint, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, float, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, float, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(/, double, double, double)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, uchar, uchar, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, char, char, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, ushort, ushort, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, short, short, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, int, int, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, uint, uint, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, float, float, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(==, double, double, uchar)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, uchar, uchar, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, char, char, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, ushort, ushort, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, short, short, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, int, int, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, uint, uint, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, float, float, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(!=, double, double, uchar)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, uchar, uchar, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, char, char, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, ushort, ushort, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, short, short, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, int, int, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, uint, uint, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, float, float, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>, double, double, uchar)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, uchar, uchar, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, char, char, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, ushort, ushort, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, short, short, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, int, int, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, uint, uint, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, float, float, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<, double, double, uchar)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, uchar, uchar, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, char, char, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, ushort, ushort, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, short, short, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, int, int, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, uint, uint, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, float, float, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(>=, double, double, uchar)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, uchar, uchar, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, char, char, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, ushort, ushort, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, short, short, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, int, int, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, uint, uint, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, float, float, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(<=, double, double, uchar)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, uchar, uchar, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, char, char, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, ushort, ushort, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, short, short, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, int, int, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, uint, uint, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, float, float, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&&, double, double, uchar)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, uchar, uchar, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, char, char, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, ushort, ushort, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, short, short, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, int, int, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, uint, uint, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, float, float, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(||, double, double, uchar)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&, uchar, uchar, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&, char, char, char)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&, ushort, ushort, ushort)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&, short, short, short)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&, int, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(&, uint, uint, uint)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(|, uchar, uchar, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(|, char, char, char)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(|, ushort, ushort, ushort)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(|, short, short, short)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(|, int, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(|, uint, uint, uint)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(^, uchar, uchar, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(^, char, char, char)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(^, ushort, ushort, ushort)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(^, short, short, short)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(^, int, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP(^, uint, uint, uint)
-
-#undef CV_CUDEV_IMPLEMENT_SCALAR_BINARY_OP
-
-// binary function (vec & vec)
-
-#define CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(func_name, func, input_type, output_type) \
-    __device__ __forceinline__ output_type ## 1 func_name(const input_type ## 1 & a, const input_type ## 1 & b) \
-    { \
-        return VecTraits<output_type ## 1>::make(func (a.x, b.x)); \
-    } \
-    __device__ __forceinline__ output_type ## 2 func_name(const input_type ## 2 & a, const input_type ## 2 & b) \
-    { \
-        return VecTraits<output_type ## 2>::make(func (a.x, b.x), func (a.y, b.y)); \
-    } \
-    __device__ __forceinline__ output_type ## 3 func_name(const input_type ## 3 & a, const input_type ## 3 & b) \
-    { \
-        return VecTraits<output_type ## 3>::make(func (a.x, b.x), func (a.y, b.y), func (a.z, b.z)); \
-    } \
-    __device__ __forceinline__ output_type ## 4 func_name(const input_type ## 4 & a, const input_type ## 4 & b) \
-    { \
-        return VecTraits<output_type ## 4>::make(func (a.x, b.x), func (a.y, b.y), func (a.z, b.z), func (a.w, b.w)); \
-    }
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::max, uchar, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::max, char, char)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::max, ushort, ushort)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::max, short, short)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::max, uint, uint)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::max, int, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::fmaxf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(max, ::fmax, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::min, uchar, uchar)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::min, char, char)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::min, ushort, ushort)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::min, short, short)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::min, uint, uint)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::min, int, int)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::fminf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(min, ::fmin, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypotf, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypotf, char, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypotf, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypotf, short, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypotf, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypotf, int, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypotf, float, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(hypot, ::hypot, double, double)
-
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2f, uchar, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2f, char, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2f, ushort, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2f, short, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2f, uint, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2f, int, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2f, float, float)
-CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC(atan2, ::atan2, double, double)
-
-#undef CV_CUDEV_IMPLEMENT_VEC_BINARY_FUNC
-
-// binary function (vec & scalar)
-
-#define CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(func_name, func, input_type, scalar_type, output_type) \
-    __device__ __forceinline__ output_type ## 1 func_name(const input_type ## 1 & a, scalar_type s) \
-    { \
-        return VecTraits<output_type ## 1>::make(func ((output_type) a.x, (output_type) s)); \
-    } \
-    __device__ __forceinline__ output_type ## 1 func_name(scalar_type s, const input_type ## 1 & b) \
-    { \
-        return VecTraits<output_type ## 1>::make(func ((output_type) s, (output_type) b.x)); \
-    } \
-    __device__ __forceinline__ output_type ## 2 func_name(const input_type ## 2 & a, scalar_type s) \
-    { \
-        return VecTraits<output_type ## 2>::make(func ((output_type) a.x, (output_type) s), func ((output_type) a.y, (output_type) s)); \
-    } \
-    __device__ __forceinline__ output_type ## 2 func_name(scalar_type s, const input_type ## 2 & b) \
-    { \
-        return VecTraits<output_type ## 2>::make(func ((output_type) s, (output_type) b.x), func ((output_type) s, (output_type) b.y)); \
-    } \
-    __device__ __forceinline__ output_type ## 3 func_name(const input_type ## 3 & a, scalar_type s) \
-    { \
-        return VecTraits<output_type ## 3>::make(func ((output_type) a.x, (output_type) s), func ((output_type) a.y, (output_type) s), func ((output_type) a.z, (output_type) s)); \
-    } \
-    __device__ __forceinline__ output_type ## 3 func_name(scalar_type s, const input_type ## 3 & b) \
-    { \
-        return VecTraits<output_type ## 3>::make(func ((output_type) s, (output_type) b.x), func ((output_type) s, (output_type) b.y), func ((output_type) s, (output_type) b.z)); \
-    } \
-    __device__ __forceinline__ output_type ## 4 func_name(const input_type ## 4 & a, scalar_type s) \
-    { \
-        return VecTraits<output_type ## 4>::make(func ((output_type) a.x, (output_type) s), func ((output_type) a.y, (output_type) s), func ((output_type) a.z, (output_type) s), func ((output_type) a.w, (output_type) s)); \
-    } \
-    __device__ __forceinline__ output_type ## 4 func_name(scalar_type s, const input_type ## 4 & b) \
-    { \
-        return VecTraits<output_type ## 4>::make(func ((output_type) s, (output_type) b.x), func ((output_type) s, (output_type) b.y), func ((output_type) s, (output_type) b.z), func ((output_type) s, (output_type) b.w)); \
-    }
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::max, uchar, uchar, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmaxf, uchar, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, uchar, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::max, char, char, char)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmaxf, char, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, char, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::max, ushort, ushort, ushort)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmaxf, ushort, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, ushort, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::max, short, short, short)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmaxf, short, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, short, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::max, uint, uint, uint)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmaxf, uint, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, uint, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::max, int, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmaxf, int, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, int, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmaxf, float, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, float, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(max, ::fmax, double, double, double)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::min, uchar, uchar, uchar)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fminf, uchar, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, uchar, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::min, char, char, char)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fminf, char, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, char, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::min, ushort, ushort, ushort)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fminf, ushort, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, ushort, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::min, short, short, short)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fminf, short, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, short, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::min, uint, uint, uint)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fminf, uint, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, uint, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::min, int, int, int)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fminf, int, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, int, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fminf, float, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, float, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(min, ::fmin, double, double, double)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypotf, uchar, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, uchar, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypotf, char, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, char, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypotf, ushort, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, ushort, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypotf, short, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, short, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypotf, uint, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, uint, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypotf, int, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, int, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypotf, float, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, float, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(hypot, ::hypot, double, double, double)
-
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2f, uchar, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, uchar, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2f, char, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, char, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2f, ushort, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, ushort, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2f, short, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, short, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2f, uint, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, uint, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2f, int, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, int, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2f, float, float, float)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, float, double, double)
-CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC(atan2, ::atan2, double, double, double)
-
-#undef CV_CUDEV_IMPLEMENT_SCALAR_BINARY_FUNC
-
-}}} // namespace cv { namespace cuda { namespace device
-
-//! @endcond
-
-#endif // OPENCV_CUDA_VECMATH_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/vec_traits.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/vec_traits.hpp
deleted file mode 100644
index b5ff281a..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/vec_traits.hpp
+++ /dev/null
@@ -1,288 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_VEC_TRAITS_HPP
-#define OPENCV_CUDA_VEC_TRAITS_HPP
-
-#include "common.hpp"
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    template<typename T, int N> struct TypeVec;
-
-    struct __align__(8) uchar8
-    {
-        uchar a0, a1, a2, a3, a4, a5, a6, a7;
-    };
-    static __host__ __device__ __forceinline__ uchar8 make_uchar8(uchar a0, uchar a1, uchar a2, uchar a3, uchar a4, uchar a5, uchar a6, uchar a7)
-    {
-        uchar8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
-        return val;
-    }
-    struct __align__(8) char8
-    {
-        schar a0, a1, a2, a3, a4, a5, a6, a7;
-    };
-    static __host__ __device__ __forceinline__ char8 make_char8(schar a0, schar a1, schar a2, schar a3, schar a4, schar a5, schar a6, schar a7)
-    {
-        char8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
-        return val;
-    }
-    struct __align__(16) ushort8
-    {
-        ushort a0, a1, a2, a3, a4, a5, a6, a7;
-    };
-    static __host__ __device__ __forceinline__ ushort8 make_ushort8(ushort a0, ushort a1, ushort a2, ushort a3, ushort a4, ushort a5, ushort a6, ushort a7)
-    {
-        ushort8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
-        return val;
-    }
-    struct __align__(16) short8
-    {
-        short a0, a1, a2, a3, a4, a5, a6, a7;
-    };
-    static __host__ __device__ __forceinline__ short8 make_short8(short a0, short a1, short a2, short a3, short a4, short a5, short a6, short a7)
-    {
-        short8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
-        return val;
-    }
-    struct __align__(32) uint8
-    {
-        uint a0, a1, a2, a3, a4, a5, a6, a7;
-    };
-    static __host__ __device__ __forceinline__ uint8 make_uint8(uint a0, uint a1, uint a2, uint a3, uint a4, uint a5, uint a6, uint a7)
-    {
-        uint8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
-        return val;
-    }
-    struct __align__(32) int8
-    {
-        int a0, a1, a2, a3, a4, a5, a6, a7;
-    };
-    static __host__ __device__ __forceinline__ int8 make_int8(int a0, int a1, int a2, int a3, int a4, int a5, int a6, int a7)
-    {
-        int8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
-        return val;
-    }
-    struct __align__(32) float8
-    {
-        float a0, a1, a2, a3, a4, a5, a6, a7;
-    };
-    static __host__ __device__ __forceinline__ float8 make_float8(float a0, float a1, float a2, float a3, float a4, float a5, float a6, float a7)
-    {
-        float8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
-        return val;
-    }
-    struct double8
-    {
-        double a0, a1, a2, a3, a4, a5, a6, a7;
-    };
-    static __host__ __device__ __forceinline__ double8 make_double8(double a0, double a1, double a2, double a3, double a4, double a5, double a6, double a7)
-    {
-        double8 val = {a0, a1, a2, a3, a4, a5, a6, a7};
-        return val;
-    }
-
-#define OPENCV_CUDA_IMPLEMENT_TYPE_VEC(type) \
-    template<> struct TypeVec<type, 1> { typedef type vec_type; }; \
-    template<> struct TypeVec<type ## 1, 1> { typedef type ## 1 vec_type; }; \
-    template<> struct TypeVec<type, 2> { typedef type ## 2 vec_type; }; \
-    template<> struct TypeVec<type ## 2, 2> { typedef type ## 2 vec_type; }; \
-    template<> struct TypeVec<type, 3> { typedef type ## 3 vec_type; }; \
-    template<> struct TypeVec<type ## 3, 3> { typedef type ## 3 vec_type; }; \
-    template<> struct TypeVec<type, 4> { typedef type ## 4 vec_type; }; \
-    template<> struct TypeVec<type ## 4, 4> { typedef type ## 4 vec_type; }; \
-    template<> struct TypeVec<type, 8> { typedef type ## 8 vec_type; }; \
-    template<> struct TypeVec<type ## 8, 8> { typedef type ## 8 vec_type; };
-
-    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(uchar)
-    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(char)
-    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(ushort)
-    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(short)
-    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(int)
-    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(uint)
-    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(float)
-    OPENCV_CUDA_IMPLEMENT_TYPE_VEC(double)
-
-    #undef OPENCV_CUDA_IMPLEMENT_TYPE_VEC
-
-    template<> struct TypeVec<schar, 1> { typedef schar vec_type; };
-    template<> struct TypeVec<schar, 2> { typedef char2 vec_type; };
-    template<> struct TypeVec<schar, 3> { typedef char3 vec_type; };
-    template<> struct TypeVec<schar, 4> { typedef char4 vec_type; };
-    template<> struct TypeVec<schar, 8> { typedef char8 vec_type; };
-
-    template<> struct TypeVec<bool, 1> { typedef uchar vec_type; };
-    template<> struct TypeVec<bool, 2> { typedef uchar2 vec_type; };
-    template<> struct TypeVec<bool, 3> { typedef uchar3 vec_type; };
-    template<> struct TypeVec<bool, 4> { typedef uchar4 vec_type; };
-    template<> struct TypeVec<bool, 8> { typedef uchar8 vec_type; };
-
-    template<typename T> struct VecTraits;
-
-#define OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(type) \
-    template<> struct VecTraits<type> \
-    { \
-        typedef type elem_type; \
-        enum {cn=1}; \
-        static __device__ __host__ __forceinline__ type all(type v) {return v;} \
-        static __device__ __host__ __forceinline__ type make(type x) {return x;} \
-        static __device__ __host__ __forceinline__ type make(const type* v) {return *v;} \
-    }; \
-    template<> struct VecTraits<type ## 1> \
-    { \
-        typedef type elem_type; \
-        enum {cn=1}; \
-        static __device__ __host__ __forceinline__ type ## 1 all(type v) {return make_ ## type ## 1(v);} \
-        static __device__ __host__ __forceinline__ type ## 1 make(type x) {return make_ ## type ## 1(x);} \
-        static __device__ __host__ __forceinline__ type ## 1 make(const type* v) {return make_ ## type ## 1(*v);} \
-    }; \
-    template<> struct VecTraits<type ## 2> \
-    { \
-        typedef type elem_type; \
-        enum {cn=2}; \
-        static __device__ __host__ __forceinline__ type ## 2 all(type v) {return make_ ## type ## 2(v, v);} \
-        static __device__ __host__ __forceinline__ type ## 2 make(type x, type y) {return make_ ## type ## 2(x, y);} \
-        static __device__ __host__ __forceinline__ type ## 2 make(const type* v) {return make_ ## type ## 2(v[0], v[1]);} \
-    }; \
-    template<> struct VecTraits<type ## 3> \
-    { \
-        typedef type elem_type; \
-        enum {cn=3}; \
-        static __device__ __host__ __forceinline__ type ## 3 all(type v) {return make_ ## type ## 3(v, v, v);} \
-        static __device__ __host__ __forceinline__ type ## 3 make(type x, type y, type z) {return make_ ## type ## 3(x, y, z);} \
-        static __device__ __host__ __forceinline__ type ## 3 make(const type* v) {return make_ ## type ## 3(v[0], v[1], v[2]);} \
-    }; \
-    template<> struct VecTraits<type ## 4> \
-    { \
-        typedef type elem_type; \
-        enum {cn=4}; \
-        static __device__ __host__ __forceinline__ type ## 4 all(type v) {return make_ ## type ## 4(v, v, v, v);} \
-        static __device__ __host__ __forceinline__ type ## 4 make(type x, type y, type z, type w) {return make_ ## type ## 4(x, y, z, w);} \
-        static __device__ __host__ __forceinline__ type ## 4 make(const type* v) {return make_ ## type ## 4(v[0], v[1], v[2], v[3]);} \
-    }; \
-    template<> struct VecTraits<type ## 8> \
-    { \
-        typedef type elem_type; \
-        enum {cn=8}; \
-        static __device__ __host__ __forceinline__ type ## 8 all(type v) {return make_ ## type ## 8(v, v, v, v, v, v, v, v);} \
-        static __device__ __host__ __forceinline__ type ## 8 make(type a0, type a1, type a2, type a3, type a4, type a5, type a6, type a7) {return make_ ## type ## 8(a0, a1, a2, a3, a4, a5, a6, a7);} \
-        static __device__ __host__ __forceinline__ type ## 8 make(const type* v) {return make_ ## type ## 8(v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7]);} \
-    };
-
-    OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(uchar)
-    OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(ushort)
-    OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(short)
-    OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(int)
-    OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(uint)
-    OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(float)
-    OPENCV_CUDA_IMPLEMENT_VEC_TRAITS(double)
-
-    #undef OPENCV_CUDA_IMPLEMENT_VEC_TRAITS
-
-    template<> struct VecTraits<char>
-    {
-        typedef char elem_type;
-        enum {cn=1};
-        static __device__ __host__ __forceinline__ char all(char v) {return v;}
-        static __device__ __host__ __forceinline__ char make(char x) {return x;}
-        static __device__ __host__ __forceinline__ char make(const char* x) {return *x;}
-    };
-    template<> struct VecTraits<schar>
-    {
-        typedef schar elem_type;
-        enum {cn=1};
-        static __device__ __host__ __forceinline__ schar all(schar v) {return v;}
-        static __device__ __host__ __forceinline__ schar make(schar x) {return x;}
-        static __device__ __host__ __forceinline__ schar make(const schar* x) {return *x;}
-    };
-    template<> struct VecTraits<char1>
-    {
-        typedef schar elem_type;
-        enum {cn=1};
-        static __device__ __host__ __forceinline__ char1 all(schar v) {return make_char1(v);}
-        static __device__ __host__ __forceinline__ char1 make(schar x) {return make_char1(x);}
-        static __device__ __host__ __forceinline__ char1 make(const schar* v) {return make_char1(v[0]);}
-    };
-    template<> struct VecTraits<char2>
-    {
-        typedef schar elem_type;
-        enum {cn=2};
-        static __device__ __host__ __forceinline__ char2 all(schar v) {return make_char2(v, v);}
-        static __device__ __host__ __forceinline__ char2 make(schar x, schar y) {return make_char2(x, y);}
-        static __device__ __host__ __forceinline__ char2 make(const schar* v) {return make_char2(v[0], v[1]);}
-    };
-    template<> struct VecTraits<char3>
-    {
-        typedef schar elem_type;
-        enum {cn=3};
-        static __device__ __host__ __forceinline__ char3 all(schar v) {return make_char3(v, v, v);}
-        static __device__ __host__ __forceinline__ char3 make(schar x, schar y, schar z) {return make_char3(x, y, z);}
-        static __device__ __host__ __forceinline__ char3 make(const schar* v) {return make_char3(v[0], v[1], v[2]);}
-    };
-    template<> struct VecTraits<char4>
-    {
-        typedef schar elem_type;
-        enum {cn=4};
-        static __device__ __host__ __forceinline__ char4 all(schar v) {return make_char4(v, v, v, v);}
-        static __device__ __host__ __forceinline__ char4 make(schar x, schar y, schar z, schar w) {return make_char4(x, y, z, w);}
-        static __device__ __host__ __forceinline__ char4 make(const schar* v) {return make_char4(v[0], v[1], v[2], v[3]);}
-    };
-    template<> struct VecTraits<char8>
-    {
-        typedef schar elem_type;
-        enum {cn=8};
-        static __device__ __host__ __forceinline__ char8 all(schar v) {return make_char8(v, v, v, v, v, v, v, v);}
-        static __device__ __host__ __forceinline__ char8 make(schar a0, schar a1, schar a2, schar a3, schar a4, schar a5, schar a6, schar a7) {return make_char8(a0, a1, a2, a3, a4, a5, a6, a7);}
-        static __device__ __host__ __forceinline__ char8 make(const schar* v) {return make_char8(v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7]);}
-    };
-}}} // namespace cv { namespace cuda { namespace cudev
-
-//! @endcond
-
-#endif // OPENCV_CUDA_VEC_TRAITS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/warp.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/warp.hpp
deleted file mode 100644
index 8af7e6a2..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/warp.hpp
+++ /dev/null
@@ -1,139 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_DEVICE_WARP_HPP
-#define OPENCV_CUDA_DEVICE_WARP_HPP
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    struct Warp
-    {
-        enum
-        {
-            LOG_WARP_SIZE = 5,
-            WARP_SIZE     = 1 << LOG_WARP_SIZE,
-            STRIDE        = WARP_SIZE
-        };
-
-        /** \brief Returns the warp lane ID of the calling thread. */
-        static __device__ __forceinline__ unsigned int laneId()
-        {
-            unsigned int ret;
-            asm("mov.u32 %0, %%laneid;" : "=r"(ret) );
-            return ret;
-        }
-
-        template<typename It, typename T>
-        static __device__ __forceinline__ void fill(It beg, It end, const T& value)
-        {
-            for(It t = beg + laneId(); t < end; t += STRIDE)
-                *t = value;
-        }
-
-        template<typename InIt, typename OutIt>
-        static __device__ __forceinline__ OutIt copy(InIt beg, InIt end, OutIt out)
-        {
-            for(InIt t = beg + laneId(); t < end; t += STRIDE, out += STRIDE)
-                *out = *t;
-            return out;
-        }
-
-        template<typename InIt, typename OutIt, class UnOp>
-        static __device__ __forceinline__ OutIt transform(InIt beg, InIt end, OutIt out, UnOp op)
-        {
-            for(InIt t = beg + laneId(); t < end; t += STRIDE, out += STRIDE)
-                *out = op(*t);
-            return out;
-        }
-
-        template<typename InIt1, typename InIt2, typename OutIt, class BinOp>
-        static __device__ __forceinline__ OutIt transform(InIt1 beg1, InIt1 end1, InIt2 beg2, OutIt out, BinOp op)
-        {
-            unsigned int lane = laneId();
-
-            InIt1 t1 = beg1 + lane;
-            InIt2 t2 = beg2 + lane;
-            for(; t1 < end1; t1 += STRIDE, t2 += STRIDE, out += STRIDE)
-                *out = op(*t1, *t2);
-            return out;
-        }
-
-        template <class T, class BinOp>
-        static __device__ __forceinline__ T reduce(volatile T *ptr, BinOp op)
-        {
-            const unsigned int lane = laneId();
-
-            if (lane < 16)
-            {
-                T partial = ptr[lane];
-
-                ptr[lane] = partial = op(partial, ptr[lane + 16]);
-                ptr[lane] = partial = op(partial, ptr[lane + 8]);
-                ptr[lane] = partial = op(partial, ptr[lane + 4]);
-                ptr[lane] = partial = op(partial, ptr[lane + 2]);
-                ptr[lane] = partial = op(partial, ptr[lane + 1]);
-            }
-
-            return *ptr;
-        }
-
-        template<typename OutIt, typename T>
-        static __device__ __forceinline__ void yota(OutIt beg, OutIt end, T value)
-        {
-            unsigned int lane = laneId();
-            value += lane;
-
-            for(OutIt t = beg + lane; t < end; t += STRIDE, value += STRIDE)
-                *t = value;
-        }
-    };
-}}} // namespace cv { namespace cuda { namespace cudev
-
-//! @endcond
-
-#endif /* OPENCV_CUDA_DEVICE_WARP_HPP */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/warp_reduce.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/warp_reduce.hpp
deleted file mode 100644
index 530303d2..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/warp_reduce.hpp
+++ /dev/null
@@ -1,76 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_WARP_REDUCE_HPP__
-#define OPENCV_CUDA_WARP_REDUCE_HPP__
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-    template <class T>
-    __device__ __forceinline__ T warp_reduce(volatile T *ptr , const unsigned int tid = threadIdx.x)
-    {
-        const unsigned int lane = tid & 31; // index of thread in warp (0..31)
-
-        if (lane < 16)
-        {
-            T partial = ptr[tid];
-
-            ptr[tid] = partial = partial + ptr[tid + 16];
-            ptr[tid] = partial = partial + ptr[tid + 8];
-            ptr[tid] = partial = partial + ptr[tid + 4];
-            ptr[tid] = partial = partial + ptr[tid + 2];
-            ptr[tid] = partial = partial + ptr[tid + 1];
-        }
-
-        return ptr[tid - lane];
-    }
-}}} // namespace cv { namespace cuda { namespace cudev {
-
-//! @endcond
-
-#endif /* OPENCV_CUDA_WARP_REDUCE_HPP__ */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda/warp_shuffle.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda/warp_shuffle.hpp
deleted file mode 100644
index 0da54aee..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda/warp_shuffle.hpp
+++ /dev/null
@@ -1,162 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CUDA_WARP_SHUFFLE_HPP
-#define OPENCV_CUDA_WARP_SHUFFLE_HPP
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-namespace cv { namespace cuda { namespace device
-{
-#if __CUDACC_VER_MAJOR__ >= 9
-#  define __shfl(x, y, z) __shfl_sync(0xFFFFFFFFU, x, y, z)
-#  define __shfl_up(x, y, z) __shfl_up_sync(0xFFFFFFFFU, x, y, z)
-#  define __shfl_down(x, y, z) __shfl_down_sync(0xFFFFFFFFU, x, y, z)
-#endif
-    template <typename T>
-    __device__ __forceinline__ T shfl(T val, int srcLane, int width = warpSize)
-    {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
-        return __shfl(val, srcLane, width);
-    #else
-        return T();
-    #endif
-    }
-    __device__ __forceinline__ unsigned int shfl(unsigned int val, int srcLane, int width = warpSize)
-    {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
-        return (unsigned int) __shfl((int) val, srcLane, width);
-    #else
-        return 0;
-    #endif
-    }
-    __device__ __forceinline__ double shfl(double val, int srcLane, int width = warpSize)
-    {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
-        int lo = __double2loint(val);
-        int hi = __double2hiint(val);
-
-        lo = __shfl(lo, srcLane, width);
-        hi = __shfl(hi, srcLane, width);
-
-        return __hiloint2double(hi, lo);
-    #else
-        return 0.0;
-    #endif
-    }
-
-    template <typename T>
-    __device__ __forceinline__ T shfl_down(T val, unsigned int delta, int width = warpSize)
-    {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
-        return __shfl_down(val, delta, width);
-    #else
-        return T();
-    #endif
-    }
-    __device__ __forceinline__ unsigned int shfl_down(unsigned int val, unsigned int delta, int width = warpSize)
-    {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
-        return (unsigned int) __shfl_down((int) val, delta, width);
-    #else
-        return 0;
-    #endif
-    }
-    __device__ __forceinline__ double shfl_down(double val, unsigned int delta, int width = warpSize)
-    {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
-        int lo = __double2loint(val);
-        int hi = __double2hiint(val);
-
-        lo = __shfl_down(lo, delta, width);
-        hi = __shfl_down(hi, delta, width);
-
-        return __hiloint2double(hi, lo);
-    #else
-        return 0.0;
-    #endif
-    }
-
-    template <typename T>
-    __device__ __forceinline__ T shfl_up(T val, unsigned int delta, int width = warpSize)
-    {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
-        return __shfl_up(val, delta, width);
-    #else
-        return T();
-    #endif
-    }
-    __device__ __forceinline__ unsigned int shfl_up(unsigned int val, unsigned int delta, int width = warpSize)
-    {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
-        return (unsigned int) __shfl_up((int) val, delta, width);
-    #else
-        return 0;
-    #endif
-    }
-    __device__ __forceinline__ double shfl_up(double val, unsigned int delta, int width = warpSize)
-    {
-    #if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
-        int lo = __double2loint(val);
-        int hi = __double2hiint(val);
-
-        lo = __shfl_up(lo, delta, width);
-        hi = __shfl_up(hi, delta, width);
-
-        return __hiloint2double(hi, lo);
-    #else
-        return 0.0;
-    #endif
-    }
-}}}
-
-#  undef __shfl
-#  undef __shfl_up
-#  undef __shfl_down
-
-//! @endcond
-
-#endif // OPENCV_CUDA_WARP_SHUFFLE_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda_stream_accessor.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda_stream_accessor.hpp
deleted file mode 100644
index deaf356f..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda_stream_accessor.hpp
+++ /dev/null
@@ -1,86 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_CUDA_STREAM_ACCESSOR_HPP
-#define OPENCV_CORE_CUDA_STREAM_ACCESSOR_HPP
-
-#ifndef __cplusplus
-#  error cuda_stream_accessor.hpp header must be compiled as C++
-#endif
-
-/** @file cuda_stream_accessor.hpp
- * This is only header file that depends on CUDA Runtime API. All other headers are independent.
- */
-
-#include <cuda_runtime.h>
-#include "opencv2/core/cuda.hpp"
-
-namespace cv
-{
-    namespace cuda
-    {
-
-//! @addtogroup cudacore_struct
-//! @{
-
-        /** @brief Class that enables getting cudaStream_t from cuda::Stream
-         */
-        struct StreamAccessor
-        {
-            CV_EXPORTS static cudaStream_t getStream(const Stream& stream);
-            CV_EXPORTS static Stream wrapStream(cudaStream_t stream);
-        };
-
-        /** @brief Class that enables getting cudaEvent_t from cuda::Event
-         */
-        struct EventAccessor
-        {
-            CV_EXPORTS static cudaEvent_t getEvent(const Event& event);
-            CV_EXPORTS static Event wrapEvent(cudaEvent_t event);
-        };
-
-//! @}
-
-    }
-}
-
-#endif /* OPENCV_CORE_CUDA_STREAM_ACCESSOR_HPP */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cuda_types.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cuda_types.hpp
deleted file mode 100644
index b33f0617..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cuda_types.hpp
+++ /dev/null
@@ -1,144 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_CUDA_TYPES_HPP
-#define OPENCV_CORE_CUDA_TYPES_HPP
-
-#ifndef __cplusplus
-#  error cuda_types.hpp header must be compiled as C++
-#endif
-
-#if defined(__OPENCV_BUILD) && defined(__clang__)
-#pragma clang diagnostic ignored "-Winconsistent-missing-override"
-#endif
-#if defined(__OPENCV_BUILD) && defined(__GNUC__) && __GNUC__ >= 5
-#pragma GCC diagnostic ignored "-Wsuggest-override"
-#endif
-
-/** @file
- * @deprecated Use @ref cudev instead.
- */
-
-//! @cond IGNORED
-
-#ifdef __CUDACC__
-    #define __CV_CUDA_HOST_DEVICE__ __host__ __device__ __forceinline__
-#else
-    #define __CV_CUDA_HOST_DEVICE__
-#endif
-
-namespace cv
-{
-    namespace cuda
-    {
-
-        // Simple lightweight structures that encapsulates information about an image on device.
-        // It is intended to pass to nvcc-compiled code. GpuMat depends on headers that nvcc can't compile
-
-        template <typename T> struct DevPtr
-        {
-            typedef T elem_type;
-            typedef int index_type;
-
-            enum { elem_size = sizeof(elem_type) };
-
-            T* data;
-
-            __CV_CUDA_HOST_DEVICE__ DevPtr() : data(0) {}
-            __CV_CUDA_HOST_DEVICE__ DevPtr(T* data_) : data(data_) {}
-
-            __CV_CUDA_HOST_DEVICE__ size_t elemSize() const { return elem_size; }
-            __CV_CUDA_HOST_DEVICE__ operator       T*()       { return data; }
-            __CV_CUDA_HOST_DEVICE__ operator const T*() const { return data; }
-        };
-
-        template <typename T> struct PtrSz : public DevPtr<T>
-        {
-            __CV_CUDA_HOST_DEVICE__ PtrSz() : size(0) {}
-            __CV_CUDA_HOST_DEVICE__ PtrSz(T* data_, size_t size_) : DevPtr<T>(data_), size(size_) {}
-
-            size_t size;
-        };
-
-        template <typename T> struct PtrStep : public DevPtr<T>
-        {
-            __CV_CUDA_HOST_DEVICE__ PtrStep() : step(0) {}
-            __CV_CUDA_HOST_DEVICE__ PtrStep(T* data_, size_t step_) : DevPtr<T>(data_), step(step_) {}
-
-            size_t step;
-
-            __CV_CUDA_HOST_DEVICE__       T* ptr(int y = 0)       { return (      T*)( (      char*)(((DevPtr<T>*)this)->data) + y * step); }
-            __CV_CUDA_HOST_DEVICE__ const T* ptr(int y = 0) const { return (const T*)( (const char*)(((DevPtr<T>*)this)->data) + y * step); }
-
-            __CV_CUDA_HOST_DEVICE__       T& operator ()(int y, int x)       { return ptr(y)[x]; }
-            __CV_CUDA_HOST_DEVICE__ const T& operator ()(int y, int x) const { return ptr(y)[x]; }
-        };
-
-        template <typename T> struct PtrStepSz : public PtrStep<T>
-        {
-            __CV_CUDA_HOST_DEVICE__ PtrStepSz() : cols(0), rows(0) {}
-            __CV_CUDA_HOST_DEVICE__ PtrStepSz(int rows_, int cols_, T* data_, size_t step_)
-                : PtrStep<T>(data_, step_), cols(cols_), rows(rows_) {}
-
-            template <typename U>
-            explicit PtrStepSz(const PtrStepSz<U>& d) : PtrStep<T>((T*)d.data, d.step), cols(d.cols), rows(d.rows){}
-
-            int cols;
-            int rows;
-        };
-
-        typedef PtrStepSz<unsigned char> PtrStepSzb;
-        typedef PtrStepSz<unsigned short> PtrStepSzus;
-        typedef PtrStepSz<float> PtrStepSzf;
-        typedef PtrStepSz<int> PtrStepSzi;
-
-        typedef PtrStep<unsigned char> PtrStepb;
-        typedef PtrStep<unsigned short> PtrStepus;
-        typedef PtrStep<float> PtrStepf;
-        typedef PtrStep<int> PtrStepi;
-
-    }
-}
-
-//! @endcond
-
-#endif /* OPENCV_CORE_CUDA_TYPES_HPP */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cv_cpu_dispatch.h b/projects/llm_framework/include/opencv4/opencv2/core/cv_cpu_dispatch.h
deleted file mode 100644
index ab5a67d4..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cv_cpu_dispatch.h
+++ /dev/null
@@ -1,368 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#if defined __OPENCV_BUILD \
-
-#include "cv_cpu_config.h"
-#include "cv_cpu_helper.h"
-
-#ifdef CV_CPU_DISPATCH_MODE
-#define CV_CPU_OPTIMIZATION_NAMESPACE __CV_CAT(opt_, CV_CPU_DISPATCH_MODE)
-#define CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN namespace __CV_CAT(opt_, CV_CPU_DISPATCH_MODE) {
-#define CV_CPU_OPTIMIZATION_NAMESPACE_END }
-#else
-#define CV_CPU_OPTIMIZATION_NAMESPACE cpu_baseline
-#define CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN namespace cpu_baseline {
-#define CV_CPU_OPTIMIZATION_NAMESPACE_END }
-#define CV_CPU_BASELINE_MODE 1
-#endif
-
-
-#define __CV_CPU_DISPATCH_CHAIN_END(fn, args, mode, ...)  /* done */
-#define __CV_CPU_DISPATCH(fn, args, mode, ...) __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-#define __CV_CPU_DISPATCH_EXPAND(fn, args, ...) __CV_EXPAND(__CV_CPU_DISPATCH(fn, args, __VA_ARGS__))
-#define CV_CPU_DISPATCH(fn, args, ...) __CV_CPU_DISPATCH_EXPAND(fn, args, __VA_ARGS__, END) // expand macros
-
-
-#if defined CV_ENABLE_INTRINSICS \
-    && !defined CV_DISABLE_OPTIMIZATION \
-    && !defined __CUDACC__ /* do not include SSE/AVX/NEON headers for NVCC compiler */ \
-
-#ifdef CV_CPU_COMPILE_SSE2
-#  include <emmintrin.h>
-#  define CV_MMX 1
-#  define CV_SSE 1
-#  define CV_SSE2 1
-#endif
-#ifdef CV_CPU_COMPILE_SSE3
-#  include <pmmintrin.h>
-#  define CV_SSE3 1
-#endif
-#ifdef CV_CPU_COMPILE_SSSE3
-#  include <tmmintrin.h>
-#  define CV_SSSE3 1
-#endif
-#ifdef CV_CPU_COMPILE_SSE4_1
-#  include <smmintrin.h>
-#  define CV_SSE4_1 1
-#endif
-#ifdef CV_CPU_COMPILE_SSE4_2
-#  include <nmmintrin.h>
-#  define CV_SSE4_2 1
-#endif
-#ifdef CV_CPU_COMPILE_POPCNT
-#  ifdef _MSC_VER
-#    include <nmmintrin.h>
-#    if defined(_M_X64)
-#      define CV_POPCNT_U64 (int)_mm_popcnt_u64
-#    endif
-#    define CV_POPCNT_U32 _mm_popcnt_u32
-#  else
-#    include <popcntintrin.h>
-#    if defined(__x86_64__)
-#      define CV_POPCNT_U64 __builtin_popcountll
-#    endif
-#    define CV_POPCNT_U32 __builtin_popcount
-#  endif
-#  define CV_POPCNT 1
-#endif
-#ifdef CV_CPU_COMPILE_AVX
-#  include <immintrin.h>
-#  define CV_AVX 1
-#endif
-#ifdef CV_CPU_COMPILE_FP16
-#  if defined(__arm__) || defined(__aarch64__) || defined(_M_ARM) || defined(_M_ARM64)
-#    include <arm_neon.h>
-#  else
-#    include <immintrin.h>
-#  endif
-#  define CV_FP16 1
-#endif
-#ifdef CV_CPU_COMPILE_AVX2
-#  include <immintrin.h>
-#  define CV_AVX2 1
-#endif
-#ifdef CV_CPU_COMPILE_AVX_512F
-#  include <immintrin.h>
-#  define CV_AVX_512F 1
-#endif
-#ifdef CV_CPU_COMPILE_AVX512_COMMON
-#  define CV_AVX512_COMMON 1
-#  define CV_AVX_512CD 1
-#endif
-#ifdef CV_CPU_COMPILE_AVX512_KNL
-#  define CV_AVX512_KNL 1
-#  define CV_AVX_512ER 1
-#  define CV_AVX_512PF 1
-#endif
-#ifdef CV_CPU_COMPILE_AVX512_KNM
-#  define CV_AVX512_KNM 1
-#  define CV_AVX_5124FMAPS 1
-#  define CV_AVX_5124VNNIW 1
-#  define CV_AVX_512VPOPCNTDQ 1
-#endif
-#ifdef CV_CPU_COMPILE_AVX512_SKX
-#  define CV_AVX512_SKX 1
-#  define CV_AVX_512VL 1
-#  define CV_AVX_512BW 1
-#  define CV_AVX_512DQ 1
-#endif
-#ifdef CV_CPU_COMPILE_AVX512_CNL
-#  define CV_AVX512_CNL 1
-#  define CV_AVX_512IFMA 1
-#  define CV_AVX_512VBMI 1
-#endif
-#ifdef CV_CPU_COMPILE_AVX512_CLX
-#  define CV_AVX512_CLX 1
-#  define CV_AVX_512VNNI 1
-#endif
-#ifdef CV_CPU_COMPILE_AVX512_ICL
-#  define CV_AVX512_ICL 1
-#  undef CV_AVX_512IFMA
-#  define CV_AVX_512IFMA 1
-#  undef CV_AVX_512VBMI
-#  define CV_AVX_512VBMI 1
-#  undef CV_AVX_512VNNI
-#  define CV_AVX_512VNNI 1
-#  define CV_AVX_512VBMI2 1
-#  define CV_AVX_512BITALG 1
-#  define CV_AVX_512VPOPCNTDQ 1
-#endif
-#ifdef CV_CPU_COMPILE_FMA3
-#  define CV_FMA3 1
-#endif
-
-#if defined _WIN32 && (defined(_M_ARM) || defined(_M_ARM64)) && (defined(CV_CPU_COMPILE_NEON) || !defined(_MSC_VER))
-# include <Intrin.h>
-# include <arm_neon.h>
-# define CV_NEON 1
-#elif defined(__ARM_NEON__) || (defined (__ARM_NEON) && defined(__aarch64__))
-#  include <arm_neon.h>
-#  define CV_NEON 1
-#endif
-
-#if defined(__riscv) && defined(__riscv_vector) && defined(__riscv_vector_071)
-# include<riscv-vector.h>
-# define CV_RVV071 1
-#endif
-
-#if defined(__ARM_NEON__) || defined(__aarch64__)
-#  include <arm_neon.h>
-#endif
-
-#ifdef CV_CPU_COMPILE_VSX
-#  include <altivec.h>
-#  undef vector
-#  undef pixel
-#  undef bool
-#  define CV_VSX 1
-#endif
-
-#ifdef CV_CPU_COMPILE_VSX3
-#  define CV_VSX3 1
-#endif
-
-#ifdef CV_CPU_COMPILE_MSA
-#  include "hal/msa_macros.h"
-#  define CV_MSA 1
-#endif
-
-#ifdef __EMSCRIPTEN__
-#  define CV_WASM_SIMD 1
-#  include <wasm_simd128.h>
-#endif
-
-#if defined CV_CPU_COMPILE_RVV
-#  define CV_RVV 1
-#  include <riscv_vector.h>
-#endif
-
-#endif // CV_ENABLE_INTRINSICS && !CV_DISABLE_OPTIMIZATION && !__CUDACC__
-
-#if defined CV_CPU_COMPILE_AVX && !defined CV_CPU_BASELINE_COMPILE_AVX
-struct VZeroUpperGuard {
-#ifdef __GNUC__
-    __attribute__((always_inline))
-#endif
-    inline VZeroUpperGuard() { _mm256_zeroupper(); }
-#ifdef __GNUC__
-    __attribute__((always_inline))
-#endif
-    inline ~VZeroUpperGuard() { _mm256_zeroupper(); }
-};
-#define __CV_AVX_GUARD VZeroUpperGuard __vzeroupper_guard; CV_UNUSED(__vzeroupper_guard);
-#endif
-
-#ifdef __CV_AVX_GUARD
-#define CV_AVX_GUARD __CV_AVX_GUARD
-#else
-#define CV_AVX_GUARD
-#endif
-
-#endif // __OPENCV_BUILD
-
-
-
-#if !defined __OPENCV_BUILD /* Compatibility code */ \
-    && !defined __CUDACC__ /* do not include SSE/AVX/NEON headers for NVCC compiler */
-#if defined __SSE2__ || defined _M_X64 || (defined _M_IX86_FP && _M_IX86_FP >= 2)
-#  include <emmintrin.h>
-#  define CV_MMX 1
-#  define CV_SSE 1
-#  define CV_SSE2 1
-#elif defined _WIN32 && (defined(_M_ARM) || defined(_M_ARM64)) && (defined(CV_CPU_COMPILE_NEON) || !defined(_MSC_VER))
-# include <Intrin.h>
-# include <arm_neon.h>
-# define CV_NEON 1
-#elif defined(__ARM_NEON__) || (defined (__ARM_NEON) && defined(__aarch64__))
-#  include <arm_neon.h>
-#  define CV_NEON 1
-#elif defined(__VSX__) && defined(__PPC64__) && defined(__LITTLE_ENDIAN__)
-#  include <altivec.h>
-#  undef vector
-#  undef pixel
-#  undef bool
-#  define CV_VSX 1
-#endif
-
-#ifdef __F16C__
-#  include <immintrin.h>
-#  define CV_FP16 1
-#endif
-
-#endif // !__OPENCV_BUILD && !__CUDACC (Compatibility code)
-
-
-
-#ifndef CV_MMX
-#  define CV_MMX 0
-#endif
-#ifndef CV_SSE
-#  define CV_SSE 0
-#endif
-#ifndef CV_SSE2
-#  define CV_SSE2 0
-#endif
-#ifndef CV_SSE3
-#  define CV_SSE3 0
-#endif
-#ifndef CV_SSSE3
-#  define CV_SSSE3 0
-#endif
-#ifndef CV_SSE4_1
-#  define CV_SSE4_1 0
-#endif
-#ifndef CV_SSE4_2
-#  define CV_SSE4_2 0
-#endif
-#ifndef CV_POPCNT
-#  define CV_POPCNT 0
-#endif
-#ifndef CV_AVX
-#  define CV_AVX 0
-#endif
-#ifndef CV_FP16
-#  define CV_FP16 0
-#endif
-#ifndef CV_AVX2
-#  define CV_AVX2 0
-#endif
-#ifndef CV_FMA3
-#  define CV_FMA3 0
-#endif
-#ifndef CV_AVX_512F
-#  define CV_AVX_512F 0
-#endif
-#ifndef CV_AVX_512BW
-#  define CV_AVX_512BW 0
-#endif
-#ifndef CV_AVX_512CD
-#  define CV_AVX_512CD 0
-#endif
-#ifndef CV_AVX_512DQ
-#  define CV_AVX_512DQ 0
-#endif
-#ifndef CV_AVX_512ER
-#  define CV_AVX_512ER 0
-#endif
-#ifndef CV_AVX_512IFMA
-#  define CV_AVX_512IFMA 0
-#endif
-#define CV_AVX_512IFMA512 CV_AVX_512IFMA // deprecated
-#ifndef CV_AVX_512PF
-#  define CV_AVX_512PF 0
-#endif
-#ifndef CV_AVX_512VBMI
-#  define CV_AVX_512VBMI 0
-#endif
-#ifndef CV_AVX_512VL
-#  define CV_AVX_512VL 0
-#endif
-#ifndef CV_AVX_5124FMAPS
-#  define CV_AVX_5124FMAPS 0
-#endif
-#ifndef CV_AVX_5124VNNIW
-#  define CV_AVX_5124VNNIW 0
-#endif
-#ifndef CV_AVX_512VPOPCNTDQ
-#  define CV_AVX_512VPOPCNTDQ 0
-#endif
-#ifndef CV_AVX_512VNNI
-#  define CV_AVX_512VNNI 0
-#endif
-#ifndef CV_AVX_512VBMI2
-#  define CV_AVX_512VBMI2 0
-#endif
-#ifndef CV_AVX_512BITALG
-#  define CV_AVX_512BITALG 0
-#endif
-#ifndef CV_AVX512_COMMON
-#  define CV_AVX512_COMMON 0
-#endif
-#ifndef CV_AVX512_KNL
-#  define CV_AVX512_KNL 0
-#endif
-#ifndef CV_AVX512_KNM
-#  define CV_AVX512_KNM 0
-#endif
-#ifndef CV_AVX512_SKX
-#  define CV_AVX512_SKX 0
-#endif
-#ifndef CV_AVX512_CNL
-#  define CV_AVX512_CNL 0
-#endif
-#ifndef CV_AVX512_CLX
-#  define CV_AVX512_CLX 0
-#endif
-#ifndef CV_AVX512_ICL
-#  define CV_AVX512_ICL 0
-#endif
-
-#ifndef CV_NEON
-#  define CV_NEON 0
-#endif
-
-#ifndef CV_RVV071
-#  define CV_RVV071 0
-#endif
-
-#ifndef CV_VSX
-#  define CV_VSX 0
-#endif
-
-#ifndef CV_VSX3
-#  define CV_VSX3 0
-#endif
-
-#ifndef CV_MSA
-#  define CV_MSA 0
-#endif
-
-#ifndef CV_WASM_SIMD
-#  define CV_WASM_SIMD 0
-#endif
-
-#ifndef CV_RVV
-#  define CV_RVV 0
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cv_cpu_helper.h b/projects/llm_framework/include/opencv4/opencv2/core/cv_cpu_helper.h
deleted file mode 100644
index 39ae0b91..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cv_cpu_helper.h
+++ /dev/null
@@ -1,508 +0,0 @@
-// AUTOGENERATED, DO NOT EDIT
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_SSE
-#  define CV_TRY_SSE 1
-#  define CV_CPU_FORCE_SSE 1
-#  define CV_CPU_HAS_SUPPORT_SSE 1
-#  define CV_CPU_CALL_SSE(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_SSE_(fn, args) return (opt_SSE::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_SSE
-#  define CV_TRY_SSE 1
-#  define CV_CPU_FORCE_SSE 0
-#  define CV_CPU_HAS_SUPPORT_SSE (cv::checkHardwareSupport(CV_CPU_SSE))
-#  define CV_CPU_CALL_SSE(fn, args) if (CV_CPU_HAS_SUPPORT_SSE) return (opt_SSE::fn args)
-#  define CV_CPU_CALL_SSE_(fn, args) if (CV_CPU_HAS_SUPPORT_SSE) return (opt_SSE::fn args)
-#else
-#  define CV_TRY_SSE 0
-#  define CV_CPU_FORCE_SSE 0
-#  define CV_CPU_HAS_SUPPORT_SSE 0
-#  define CV_CPU_CALL_SSE(fn, args)
-#  define CV_CPU_CALL_SSE_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_SSE(fn, args, mode, ...)  CV_CPU_CALL_SSE(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_SSE2
-#  define CV_TRY_SSE2 1
-#  define CV_CPU_FORCE_SSE2 1
-#  define CV_CPU_HAS_SUPPORT_SSE2 1
-#  define CV_CPU_CALL_SSE2(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_SSE2_(fn, args) return (opt_SSE2::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_SSE2
-#  define CV_TRY_SSE2 1
-#  define CV_CPU_FORCE_SSE2 0
-#  define CV_CPU_HAS_SUPPORT_SSE2 (cv::checkHardwareSupport(CV_CPU_SSE2))
-#  define CV_CPU_CALL_SSE2(fn, args) if (CV_CPU_HAS_SUPPORT_SSE2) return (opt_SSE2::fn args)
-#  define CV_CPU_CALL_SSE2_(fn, args) if (CV_CPU_HAS_SUPPORT_SSE2) return (opt_SSE2::fn args)
-#else
-#  define CV_TRY_SSE2 0
-#  define CV_CPU_FORCE_SSE2 0
-#  define CV_CPU_HAS_SUPPORT_SSE2 0
-#  define CV_CPU_CALL_SSE2(fn, args)
-#  define CV_CPU_CALL_SSE2_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_SSE2(fn, args, mode, ...)  CV_CPU_CALL_SSE2(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_SSE3
-#  define CV_TRY_SSE3 1
-#  define CV_CPU_FORCE_SSE3 1
-#  define CV_CPU_HAS_SUPPORT_SSE3 1
-#  define CV_CPU_CALL_SSE3(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_SSE3_(fn, args) return (opt_SSE3::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_SSE3
-#  define CV_TRY_SSE3 1
-#  define CV_CPU_FORCE_SSE3 0
-#  define CV_CPU_HAS_SUPPORT_SSE3 (cv::checkHardwareSupport(CV_CPU_SSE3))
-#  define CV_CPU_CALL_SSE3(fn, args) if (CV_CPU_HAS_SUPPORT_SSE3) return (opt_SSE3::fn args)
-#  define CV_CPU_CALL_SSE3_(fn, args) if (CV_CPU_HAS_SUPPORT_SSE3) return (opt_SSE3::fn args)
-#else
-#  define CV_TRY_SSE3 0
-#  define CV_CPU_FORCE_SSE3 0
-#  define CV_CPU_HAS_SUPPORT_SSE3 0
-#  define CV_CPU_CALL_SSE3(fn, args)
-#  define CV_CPU_CALL_SSE3_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_SSE3(fn, args, mode, ...)  CV_CPU_CALL_SSE3(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_SSSE3
-#  define CV_TRY_SSSE3 1
-#  define CV_CPU_FORCE_SSSE3 1
-#  define CV_CPU_HAS_SUPPORT_SSSE3 1
-#  define CV_CPU_CALL_SSSE3(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_SSSE3_(fn, args) return (opt_SSSE3::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_SSSE3
-#  define CV_TRY_SSSE3 1
-#  define CV_CPU_FORCE_SSSE3 0
-#  define CV_CPU_HAS_SUPPORT_SSSE3 (cv::checkHardwareSupport(CV_CPU_SSSE3))
-#  define CV_CPU_CALL_SSSE3(fn, args) if (CV_CPU_HAS_SUPPORT_SSSE3) return (opt_SSSE3::fn args)
-#  define CV_CPU_CALL_SSSE3_(fn, args) if (CV_CPU_HAS_SUPPORT_SSSE3) return (opt_SSSE3::fn args)
-#else
-#  define CV_TRY_SSSE3 0
-#  define CV_CPU_FORCE_SSSE3 0
-#  define CV_CPU_HAS_SUPPORT_SSSE3 0
-#  define CV_CPU_CALL_SSSE3(fn, args)
-#  define CV_CPU_CALL_SSSE3_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_SSSE3(fn, args, mode, ...)  CV_CPU_CALL_SSSE3(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_SSE4_1
-#  define CV_TRY_SSE4_1 1
-#  define CV_CPU_FORCE_SSE4_1 1
-#  define CV_CPU_HAS_SUPPORT_SSE4_1 1
-#  define CV_CPU_CALL_SSE4_1(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_SSE4_1_(fn, args) return (opt_SSE4_1::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_SSE4_1
-#  define CV_TRY_SSE4_1 1
-#  define CV_CPU_FORCE_SSE4_1 0
-#  define CV_CPU_HAS_SUPPORT_SSE4_1 (cv::checkHardwareSupport(CV_CPU_SSE4_1))
-#  define CV_CPU_CALL_SSE4_1(fn, args) if (CV_CPU_HAS_SUPPORT_SSE4_1) return (opt_SSE4_1::fn args)
-#  define CV_CPU_CALL_SSE4_1_(fn, args) if (CV_CPU_HAS_SUPPORT_SSE4_1) return (opt_SSE4_1::fn args)
-#else
-#  define CV_TRY_SSE4_1 0
-#  define CV_CPU_FORCE_SSE4_1 0
-#  define CV_CPU_HAS_SUPPORT_SSE4_1 0
-#  define CV_CPU_CALL_SSE4_1(fn, args)
-#  define CV_CPU_CALL_SSE4_1_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_SSE4_1(fn, args, mode, ...)  CV_CPU_CALL_SSE4_1(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_SSE4_2
-#  define CV_TRY_SSE4_2 1
-#  define CV_CPU_FORCE_SSE4_2 1
-#  define CV_CPU_HAS_SUPPORT_SSE4_2 1
-#  define CV_CPU_CALL_SSE4_2(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_SSE4_2_(fn, args) return (opt_SSE4_2::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_SSE4_2
-#  define CV_TRY_SSE4_2 1
-#  define CV_CPU_FORCE_SSE4_2 0
-#  define CV_CPU_HAS_SUPPORT_SSE4_2 (cv::checkHardwareSupport(CV_CPU_SSE4_2))
-#  define CV_CPU_CALL_SSE4_2(fn, args) if (CV_CPU_HAS_SUPPORT_SSE4_2) return (opt_SSE4_2::fn args)
-#  define CV_CPU_CALL_SSE4_2_(fn, args) if (CV_CPU_HAS_SUPPORT_SSE4_2) return (opt_SSE4_2::fn args)
-#else
-#  define CV_TRY_SSE4_2 0
-#  define CV_CPU_FORCE_SSE4_2 0
-#  define CV_CPU_HAS_SUPPORT_SSE4_2 0
-#  define CV_CPU_CALL_SSE4_2(fn, args)
-#  define CV_CPU_CALL_SSE4_2_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_SSE4_2(fn, args, mode, ...)  CV_CPU_CALL_SSE4_2(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_POPCNT
-#  define CV_TRY_POPCNT 1
-#  define CV_CPU_FORCE_POPCNT 1
-#  define CV_CPU_HAS_SUPPORT_POPCNT 1
-#  define CV_CPU_CALL_POPCNT(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_POPCNT_(fn, args) return (opt_POPCNT::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_POPCNT
-#  define CV_TRY_POPCNT 1
-#  define CV_CPU_FORCE_POPCNT 0
-#  define CV_CPU_HAS_SUPPORT_POPCNT (cv::checkHardwareSupport(CV_CPU_POPCNT))
-#  define CV_CPU_CALL_POPCNT(fn, args) if (CV_CPU_HAS_SUPPORT_POPCNT) return (opt_POPCNT::fn args)
-#  define CV_CPU_CALL_POPCNT_(fn, args) if (CV_CPU_HAS_SUPPORT_POPCNT) return (opt_POPCNT::fn args)
-#else
-#  define CV_TRY_POPCNT 0
-#  define CV_CPU_FORCE_POPCNT 0
-#  define CV_CPU_HAS_SUPPORT_POPCNT 0
-#  define CV_CPU_CALL_POPCNT(fn, args)
-#  define CV_CPU_CALL_POPCNT_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_POPCNT(fn, args, mode, ...)  CV_CPU_CALL_POPCNT(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX
-#  define CV_TRY_AVX 1
-#  define CV_CPU_FORCE_AVX 1
-#  define CV_CPU_HAS_SUPPORT_AVX 1
-#  define CV_CPU_CALL_AVX(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_AVX_(fn, args) return (opt_AVX::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX
-#  define CV_TRY_AVX 1
-#  define CV_CPU_FORCE_AVX 0
-#  define CV_CPU_HAS_SUPPORT_AVX (cv::checkHardwareSupport(CV_CPU_AVX))
-#  define CV_CPU_CALL_AVX(fn, args) if (CV_CPU_HAS_SUPPORT_AVX) return (opt_AVX::fn args)
-#  define CV_CPU_CALL_AVX_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX) return (opt_AVX::fn args)
-#else
-#  define CV_TRY_AVX 0
-#  define CV_CPU_FORCE_AVX 0
-#  define CV_CPU_HAS_SUPPORT_AVX 0
-#  define CV_CPU_CALL_AVX(fn, args)
-#  define CV_CPU_CALL_AVX_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_AVX(fn, args, mode, ...)  CV_CPU_CALL_AVX(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_FP16
-#  define CV_TRY_FP16 1
-#  define CV_CPU_FORCE_FP16 1
-#  define CV_CPU_HAS_SUPPORT_FP16 1
-#  define CV_CPU_CALL_FP16(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_FP16_(fn, args) return (opt_FP16::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_FP16
-#  define CV_TRY_FP16 1
-#  define CV_CPU_FORCE_FP16 0
-#  define CV_CPU_HAS_SUPPORT_FP16 (cv::checkHardwareSupport(CV_CPU_FP16))
-#  define CV_CPU_CALL_FP16(fn, args) if (CV_CPU_HAS_SUPPORT_FP16) return (opt_FP16::fn args)
-#  define CV_CPU_CALL_FP16_(fn, args) if (CV_CPU_HAS_SUPPORT_FP16) return (opt_FP16::fn args)
-#else
-#  define CV_TRY_FP16 0
-#  define CV_CPU_FORCE_FP16 0
-#  define CV_CPU_HAS_SUPPORT_FP16 0
-#  define CV_CPU_CALL_FP16(fn, args)
-#  define CV_CPU_CALL_FP16_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_FP16(fn, args, mode, ...)  CV_CPU_CALL_FP16(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX2
-#  define CV_TRY_AVX2 1
-#  define CV_CPU_FORCE_AVX2 1
-#  define CV_CPU_HAS_SUPPORT_AVX2 1
-#  define CV_CPU_CALL_AVX2(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_AVX2_(fn, args) return (opt_AVX2::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX2
-#  define CV_TRY_AVX2 1
-#  define CV_CPU_FORCE_AVX2 0
-#  define CV_CPU_HAS_SUPPORT_AVX2 (cv::checkHardwareSupport(CV_CPU_AVX2))
-#  define CV_CPU_CALL_AVX2(fn, args) if (CV_CPU_HAS_SUPPORT_AVX2) return (opt_AVX2::fn args)
-#  define CV_CPU_CALL_AVX2_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX2) return (opt_AVX2::fn args)
-#else
-#  define CV_TRY_AVX2 0
-#  define CV_CPU_FORCE_AVX2 0
-#  define CV_CPU_HAS_SUPPORT_AVX2 0
-#  define CV_CPU_CALL_AVX2(fn, args)
-#  define CV_CPU_CALL_AVX2_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_AVX2(fn, args, mode, ...)  CV_CPU_CALL_AVX2(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_FMA3
-#  define CV_TRY_FMA3 1
-#  define CV_CPU_FORCE_FMA3 1
-#  define CV_CPU_HAS_SUPPORT_FMA3 1
-#  define CV_CPU_CALL_FMA3(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_FMA3_(fn, args) return (opt_FMA3::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_FMA3
-#  define CV_TRY_FMA3 1
-#  define CV_CPU_FORCE_FMA3 0
-#  define CV_CPU_HAS_SUPPORT_FMA3 (cv::checkHardwareSupport(CV_CPU_FMA3))
-#  define CV_CPU_CALL_FMA3(fn, args) if (CV_CPU_HAS_SUPPORT_FMA3) return (opt_FMA3::fn args)
-#  define CV_CPU_CALL_FMA3_(fn, args) if (CV_CPU_HAS_SUPPORT_FMA3) return (opt_FMA3::fn args)
-#else
-#  define CV_TRY_FMA3 0
-#  define CV_CPU_FORCE_FMA3 0
-#  define CV_CPU_HAS_SUPPORT_FMA3 0
-#  define CV_CPU_CALL_FMA3(fn, args)
-#  define CV_CPU_CALL_FMA3_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_FMA3(fn, args, mode, ...)  CV_CPU_CALL_FMA3(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX_512F
-#  define CV_TRY_AVX_512F 1
-#  define CV_CPU_FORCE_AVX_512F 1
-#  define CV_CPU_HAS_SUPPORT_AVX_512F 1
-#  define CV_CPU_CALL_AVX_512F(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_AVX_512F_(fn, args) return (opt_AVX_512F::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX_512F
-#  define CV_TRY_AVX_512F 1
-#  define CV_CPU_FORCE_AVX_512F 0
-#  define CV_CPU_HAS_SUPPORT_AVX_512F (cv::checkHardwareSupport(CV_CPU_AVX_512F))
-#  define CV_CPU_CALL_AVX_512F(fn, args) if (CV_CPU_HAS_SUPPORT_AVX_512F) return (opt_AVX_512F::fn args)
-#  define CV_CPU_CALL_AVX_512F_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX_512F) return (opt_AVX_512F::fn args)
-#else
-#  define CV_TRY_AVX_512F 0
-#  define CV_CPU_FORCE_AVX_512F 0
-#  define CV_CPU_HAS_SUPPORT_AVX_512F 0
-#  define CV_CPU_CALL_AVX_512F(fn, args)
-#  define CV_CPU_CALL_AVX_512F_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_AVX_512F(fn, args, mode, ...)  CV_CPU_CALL_AVX_512F(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX512_COMMON
-#  define CV_TRY_AVX512_COMMON 1
-#  define CV_CPU_FORCE_AVX512_COMMON 1
-#  define CV_CPU_HAS_SUPPORT_AVX512_COMMON 1
-#  define CV_CPU_CALL_AVX512_COMMON(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_AVX512_COMMON_(fn, args) return (opt_AVX512_COMMON::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX512_COMMON
-#  define CV_TRY_AVX512_COMMON 1
-#  define CV_CPU_FORCE_AVX512_COMMON 0
-#  define CV_CPU_HAS_SUPPORT_AVX512_COMMON (cv::checkHardwareSupport(CV_CPU_AVX512_COMMON))
-#  define CV_CPU_CALL_AVX512_COMMON(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_COMMON) return (opt_AVX512_COMMON::fn args)
-#  define CV_CPU_CALL_AVX512_COMMON_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_COMMON) return (opt_AVX512_COMMON::fn args)
-#else
-#  define CV_TRY_AVX512_COMMON 0
-#  define CV_CPU_FORCE_AVX512_COMMON 0
-#  define CV_CPU_HAS_SUPPORT_AVX512_COMMON 0
-#  define CV_CPU_CALL_AVX512_COMMON(fn, args)
-#  define CV_CPU_CALL_AVX512_COMMON_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_AVX512_COMMON(fn, args, mode, ...)  CV_CPU_CALL_AVX512_COMMON(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX512_KNL
-#  define CV_TRY_AVX512_KNL 1
-#  define CV_CPU_FORCE_AVX512_KNL 1
-#  define CV_CPU_HAS_SUPPORT_AVX512_KNL 1
-#  define CV_CPU_CALL_AVX512_KNL(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_AVX512_KNL_(fn, args) return (opt_AVX512_KNL::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX512_KNL
-#  define CV_TRY_AVX512_KNL 1
-#  define CV_CPU_FORCE_AVX512_KNL 0
-#  define CV_CPU_HAS_SUPPORT_AVX512_KNL (cv::checkHardwareSupport(CV_CPU_AVX512_KNL))
-#  define CV_CPU_CALL_AVX512_KNL(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_KNL) return (opt_AVX512_KNL::fn args)
-#  define CV_CPU_CALL_AVX512_KNL_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_KNL) return (opt_AVX512_KNL::fn args)
-#else
-#  define CV_TRY_AVX512_KNL 0
-#  define CV_CPU_FORCE_AVX512_KNL 0
-#  define CV_CPU_HAS_SUPPORT_AVX512_KNL 0
-#  define CV_CPU_CALL_AVX512_KNL(fn, args)
-#  define CV_CPU_CALL_AVX512_KNL_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_AVX512_KNL(fn, args, mode, ...)  CV_CPU_CALL_AVX512_KNL(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX512_KNM
-#  define CV_TRY_AVX512_KNM 1
-#  define CV_CPU_FORCE_AVX512_KNM 1
-#  define CV_CPU_HAS_SUPPORT_AVX512_KNM 1
-#  define CV_CPU_CALL_AVX512_KNM(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_AVX512_KNM_(fn, args) return (opt_AVX512_KNM::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX512_KNM
-#  define CV_TRY_AVX512_KNM 1
-#  define CV_CPU_FORCE_AVX512_KNM 0
-#  define CV_CPU_HAS_SUPPORT_AVX512_KNM (cv::checkHardwareSupport(CV_CPU_AVX512_KNM))
-#  define CV_CPU_CALL_AVX512_KNM(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_KNM) return (opt_AVX512_KNM::fn args)
-#  define CV_CPU_CALL_AVX512_KNM_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_KNM) return (opt_AVX512_KNM::fn args)
-#else
-#  define CV_TRY_AVX512_KNM 0
-#  define CV_CPU_FORCE_AVX512_KNM 0
-#  define CV_CPU_HAS_SUPPORT_AVX512_KNM 0
-#  define CV_CPU_CALL_AVX512_KNM(fn, args)
-#  define CV_CPU_CALL_AVX512_KNM_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_AVX512_KNM(fn, args, mode, ...)  CV_CPU_CALL_AVX512_KNM(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX512_SKX
-#  define CV_TRY_AVX512_SKX 1
-#  define CV_CPU_FORCE_AVX512_SKX 1
-#  define CV_CPU_HAS_SUPPORT_AVX512_SKX 1
-#  define CV_CPU_CALL_AVX512_SKX(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_AVX512_SKX_(fn, args) return (opt_AVX512_SKX::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX512_SKX
-#  define CV_TRY_AVX512_SKX 1
-#  define CV_CPU_FORCE_AVX512_SKX 0
-#  define CV_CPU_HAS_SUPPORT_AVX512_SKX (cv::checkHardwareSupport(CV_CPU_AVX512_SKX))
-#  define CV_CPU_CALL_AVX512_SKX(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_SKX) return (opt_AVX512_SKX::fn args)
-#  define CV_CPU_CALL_AVX512_SKX_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_SKX) return (opt_AVX512_SKX::fn args)
-#else
-#  define CV_TRY_AVX512_SKX 0
-#  define CV_CPU_FORCE_AVX512_SKX 0
-#  define CV_CPU_HAS_SUPPORT_AVX512_SKX 0
-#  define CV_CPU_CALL_AVX512_SKX(fn, args)
-#  define CV_CPU_CALL_AVX512_SKX_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_AVX512_SKX(fn, args, mode, ...)  CV_CPU_CALL_AVX512_SKX(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX512_CNL
-#  define CV_TRY_AVX512_CNL 1
-#  define CV_CPU_FORCE_AVX512_CNL 1
-#  define CV_CPU_HAS_SUPPORT_AVX512_CNL 1
-#  define CV_CPU_CALL_AVX512_CNL(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_AVX512_CNL_(fn, args) return (opt_AVX512_CNL::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX512_CNL
-#  define CV_TRY_AVX512_CNL 1
-#  define CV_CPU_FORCE_AVX512_CNL 0
-#  define CV_CPU_HAS_SUPPORT_AVX512_CNL (cv::checkHardwareSupport(CV_CPU_AVX512_CNL))
-#  define CV_CPU_CALL_AVX512_CNL(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_CNL) return (opt_AVX512_CNL::fn args)
-#  define CV_CPU_CALL_AVX512_CNL_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_CNL) return (opt_AVX512_CNL::fn args)
-#else
-#  define CV_TRY_AVX512_CNL 0
-#  define CV_CPU_FORCE_AVX512_CNL 0
-#  define CV_CPU_HAS_SUPPORT_AVX512_CNL 0
-#  define CV_CPU_CALL_AVX512_CNL(fn, args)
-#  define CV_CPU_CALL_AVX512_CNL_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_AVX512_CNL(fn, args, mode, ...)  CV_CPU_CALL_AVX512_CNL(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX512_CLX
-#  define CV_TRY_AVX512_CLX 1
-#  define CV_CPU_FORCE_AVX512_CLX 1
-#  define CV_CPU_HAS_SUPPORT_AVX512_CLX 1
-#  define CV_CPU_CALL_AVX512_CLX(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_AVX512_CLX_(fn, args) return (opt_AVX512_CLX::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX512_CLX
-#  define CV_TRY_AVX512_CLX 1
-#  define CV_CPU_FORCE_AVX512_CLX 0
-#  define CV_CPU_HAS_SUPPORT_AVX512_CLX (cv::checkHardwareSupport(CV_CPU_AVX512_CLX))
-#  define CV_CPU_CALL_AVX512_CLX(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_CLX) return (opt_AVX512_CLX::fn args)
-#  define CV_CPU_CALL_AVX512_CLX_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_CLX) return (opt_AVX512_CLX::fn args)
-#else
-#  define CV_TRY_AVX512_CLX 0
-#  define CV_CPU_FORCE_AVX512_CLX 0
-#  define CV_CPU_HAS_SUPPORT_AVX512_CLX 0
-#  define CV_CPU_CALL_AVX512_CLX(fn, args)
-#  define CV_CPU_CALL_AVX512_CLX_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_AVX512_CLX(fn, args, mode, ...)  CV_CPU_CALL_AVX512_CLX(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX512_ICL
-#  define CV_TRY_AVX512_ICL 1
-#  define CV_CPU_FORCE_AVX512_ICL 1
-#  define CV_CPU_HAS_SUPPORT_AVX512_ICL 1
-#  define CV_CPU_CALL_AVX512_ICL(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_AVX512_ICL_(fn, args) return (opt_AVX512_ICL::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX512_ICL
-#  define CV_TRY_AVX512_ICL 1
-#  define CV_CPU_FORCE_AVX512_ICL 0
-#  define CV_CPU_HAS_SUPPORT_AVX512_ICL (cv::checkHardwareSupport(CV_CPU_AVX512_ICL))
-#  define CV_CPU_CALL_AVX512_ICL(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_ICL) return (opt_AVX512_ICL::fn args)
-#  define CV_CPU_CALL_AVX512_ICL_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_ICL) return (opt_AVX512_ICL::fn args)
-#else
-#  define CV_TRY_AVX512_ICL 0
-#  define CV_CPU_FORCE_AVX512_ICL 0
-#  define CV_CPU_HAS_SUPPORT_AVX512_ICL 0
-#  define CV_CPU_CALL_AVX512_ICL(fn, args)
-#  define CV_CPU_CALL_AVX512_ICL_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_AVX512_ICL(fn, args, mode, ...)  CV_CPU_CALL_AVX512_ICL(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_NEON
-#  define CV_TRY_NEON 1
-#  define CV_CPU_FORCE_NEON 1
-#  define CV_CPU_HAS_SUPPORT_NEON 1
-#  define CV_CPU_CALL_NEON(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_NEON_(fn, args) return (opt_NEON::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_NEON
-#  define CV_TRY_NEON 1
-#  define CV_CPU_FORCE_NEON 0
-#  define CV_CPU_HAS_SUPPORT_NEON (cv::checkHardwareSupport(CV_CPU_NEON))
-#  define CV_CPU_CALL_NEON(fn, args) if (CV_CPU_HAS_SUPPORT_NEON) return (opt_NEON::fn args)
-#  define CV_CPU_CALL_NEON_(fn, args) if (CV_CPU_HAS_SUPPORT_NEON) return (opt_NEON::fn args)
-#else
-#  define CV_TRY_NEON 0
-#  define CV_CPU_FORCE_NEON 0
-#  define CV_CPU_HAS_SUPPORT_NEON 0
-#  define CV_CPU_CALL_NEON(fn, args)
-#  define CV_CPU_CALL_NEON_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_NEON(fn, args, mode, ...)  CV_CPU_CALL_NEON(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_MSA
-#  define CV_TRY_MSA 1
-#  define CV_CPU_FORCE_MSA 1
-#  define CV_CPU_HAS_SUPPORT_MSA 1
-#  define CV_CPU_CALL_MSA(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_MSA_(fn, args) return (opt_MSA::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_MSA
-#  define CV_TRY_MSA 1
-#  define CV_CPU_FORCE_MSA 0
-#  define CV_CPU_HAS_SUPPORT_MSA (cv::checkHardwareSupport(CV_CPU_MSA))
-#  define CV_CPU_CALL_MSA(fn, args) if (CV_CPU_HAS_SUPPORT_MSA) return (opt_MSA::fn args)
-#  define CV_CPU_CALL_MSA_(fn, args) if (CV_CPU_HAS_SUPPORT_MSA) return (opt_MSA::fn args)
-#else
-#  define CV_TRY_MSA 0
-#  define CV_CPU_FORCE_MSA 0
-#  define CV_CPU_HAS_SUPPORT_MSA 0
-#  define CV_CPU_CALL_MSA(fn, args)
-#  define CV_CPU_CALL_MSA_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_MSA(fn, args, mode, ...)  CV_CPU_CALL_MSA(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_VSX
-#  define CV_TRY_VSX 1
-#  define CV_CPU_FORCE_VSX 1
-#  define CV_CPU_HAS_SUPPORT_VSX 1
-#  define CV_CPU_CALL_VSX(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_VSX_(fn, args) return (opt_VSX::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_VSX
-#  define CV_TRY_VSX 1
-#  define CV_CPU_FORCE_VSX 0
-#  define CV_CPU_HAS_SUPPORT_VSX (cv::checkHardwareSupport(CV_CPU_VSX))
-#  define CV_CPU_CALL_VSX(fn, args) if (CV_CPU_HAS_SUPPORT_VSX) return (opt_VSX::fn args)
-#  define CV_CPU_CALL_VSX_(fn, args) if (CV_CPU_HAS_SUPPORT_VSX) return (opt_VSX::fn args)
-#else
-#  define CV_TRY_VSX 0
-#  define CV_CPU_FORCE_VSX 0
-#  define CV_CPU_HAS_SUPPORT_VSX 0
-#  define CV_CPU_CALL_VSX(fn, args)
-#  define CV_CPU_CALL_VSX_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_VSX(fn, args, mode, ...)  CV_CPU_CALL_VSX(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_VSX3
-#  define CV_TRY_VSX3 1
-#  define CV_CPU_FORCE_VSX3 1
-#  define CV_CPU_HAS_SUPPORT_VSX3 1
-#  define CV_CPU_CALL_VSX3(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_VSX3_(fn, args) return (opt_VSX3::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_VSX3
-#  define CV_TRY_VSX3 1
-#  define CV_CPU_FORCE_VSX3 0
-#  define CV_CPU_HAS_SUPPORT_VSX3 (cv::checkHardwareSupport(CV_CPU_VSX3))
-#  define CV_CPU_CALL_VSX3(fn, args) if (CV_CPU_HAS_SUPPORT_VSX3) return (opt_VSX3::fn args)
-#  define CV_CPU_CALL_VSX3_(fn, args) if (CV_CPU_HAS_SUPPORT_VSX3) return (opt_VSX3::fn args)
-#else
-#  define CV_TRY_VSX3 0
-#  define CV_CPU_FORCE_VSX3 0
-#  define CV_CPU_HAS_SUPPORT_VSX3 0
-#  define CV_CPU_CALL_VSX3(fn, args)
-#  define CV_CPU_CALL_VSX3_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_VSX3(fn, args, mode, ...)  CV_CPU_CALL_VSX3(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_RVV
-#  define CV_TRY_RVV 1
-#  define CV_CPU_FORCE_RVV 1
-#  define CV_CPU_HAS_SUPPORT_RVV 1
-#  define CV_CPU_CALL_RVV(fn, args) return (cpu_baseline::fn args)
-#  define CV_CPU_CALL_RVV_(fn, args) return (opt_RVV::fn args)
-#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_RVV
-#  define CV_TRY_RVV 1
-#  define CV_CPU_FORCE_RVV 0
-#  define CV_CPU_HAS_SUPPORT_RVV (cv::checkHardwareSupport(CV_CPU_RVV))
-#  define CV_CPU_CALL_RVV(fn, args) if (CV_CPU_HAS_SUPPORT_RVV) return (opt_RVV::fn args)
-#  define CV_CPU_CALL_RVV_(fn, args) if (CV_CPU_HAS_SUPPORT_RVV) return (opt_RVV::fn args)
-#else
-#  define CV_TRY_RVV 0
-#  define CV_CPU_FORCE_RVV 0
-#  define CV_CPU_HAS_SUPPORT_RVV 0
-#  define CV_CPU_CALL_RVV(fn, args)
-#  define CV_CPU_CALL_RVV_(fn, args)
-#endif
-#define __CV_CPU_DISPATCH_CHAIN_RVV(fn, args, mode, ...)  CV_CPU_CALL_RVV(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
-
-#define CV_CPU_CALL_BASELINE(fn, args) return (cpu_baseline::fn args)
-#define __CV_CPU_DISPATCH_CHAIN_BASELINE(fn, args, mode, ...)  CV_CPU_CALL_BASELINE(fn, args) /* last in sequence */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cvdef.h b/projects/llm_framework/include/opencv4/opencv2/core/cvdef.h
deleted file mode 100644
index f785f32b..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cvdef.h
+++ /dev/null
@@ -1,967 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Copyright (C) 2015, Itseez Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_CVDEF_H
-#define OPENCV_CORE_CVDEF_H
-
-#include "opencv2/core/version.hpp"
-
-//! @addtogroup core_utils
-//! @{
-
-#ifdef OPENCV_INCLUDE_PORT_FILE  // User-provided header file with custom platform configuration
-#include OPENCV_INCLUDE_PORT_FILE
-#endif
-
-#if !defined CV_DOXYGEN && !defined CV_IGNORE_DEBUG_BUILD_GUARD
-#if (defined(_MSC_VER) && (defined(DEBUG) || defined(_DEBUG))) || \
-    (defined(_GLIBCXX_DEBUG) || defined(_GLIBCXX_DEBUG_PEDANTIC))
-// Guard to prevent using of binary incompatible binaries / runtimes
-// https://github.com/opencv/opencv/pull/9161
-#define CV__DEBUG_NS_BEGIN namespace debug_build_guard {
-#define CV__DEBUG_NS_END }
-namespace cv { namespace debug_build_guard { } using namespace debug_build_guard; }
-#endif
-#endif
-
-#ifndef CV__DEBUG_NS_BEGIN
-#define CV__DEBUG_NS_BEGIN
-#define CV__DEBUG_NS_END
-#endif
-
-
-#ifdef __OPENCV_BUILD
-#include "cvconfig.h"
-#endif
-
-#ifndef __CV_EXPAND
-#define __CV_EXPAND(x) x
-#endif
-
-#ifndef __CV_CAT
-#define __CV_CAT__(x, y) x ## y
-#define __CV_CAT_(x, y) __CV_CAT__(x, y)
-#define __CV_CAT(x, y) __CV_CAT_(x, y)
-#endif
-
-#define __CV_VA_NUM_ARGS_HELPER(_1, _2, _3, _4, _5, _6, _7, _8, _9, _10, N, ...) N
-#define __CV_VA_NUM_ARGS(...) __CV_EXPAND(__CV_VA_NUM_ARGS_HELPER(__VA_ARGS__, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0))
-
-#ifdef CV_Func
-// keep current value (through OpenCV port file)
-#elif defined __GNUC__ || (defined (__cpluscplus) && (__cpluscplus >= 201103))
-#define CV_Func __func__
-#elif defined __clang__ && (__clang_minor__ * 100 + __clang_major__ >= 305)
-#define CV_Func __func__
-#elif defined(__STDC_VERSION__) && (__STDC_VERSION >= 199901)
-#define CV_Func __func__
-#elif defined _MSC_VER
-#define CV_Func __FUNCTION__
-#elif defined(__INTEL_COMPILER) && (_INTEL_COMPILER >= 600)
-#define CV_Func __FUNCTION__
-#elif defined __IBMCPP__ && __IBMCPP__ >=500
-#define CV_Func __FUNCTION__
-#elif defined __BORLAND__ && (__BORLANDC__ >= 0x550)
-#define CV_Func __FUNC__
-#else
-#define CV_Func "<unknown>"
-#endif
-
-//! @cond IGNORED
-
-//////////////// static assert /////////////////
-#define CVAUX_CONCAT_EXP(a, b) a##b
-#define CVAUX_CONCAT(a, b) CVAUX_CONCAT_EXP(a,b)
-
-#if defined(__clang__)
-#  ifndef __has_extension
-#    define __has_extension __has_feature /* compatibility, for older versions of clang */
-#  endif
-#  if __has_extension(cxx_static_assert)
-#    define CV_StaticAssert(condition, reason)    static_assert((condition), reason " " #condition)
-#  elif __has_extension(c_static_assert)
-#    define CV_StaticAssert(condition, reason)    _Static_assert((condition), reason " " #condition)
-#  endif
-#elif defined(__GNUC__)
-#  if (defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103L)
-#    define CV_StaticAssert(condition, reason)    static_assert((condition), reason " " #condition)
-#  endif
-#elif defined(_MSC_VER)
-#  if _MSC_VER >= 1600 /* MSVC 10 */
-#    define CV_StaticAssert(condition, reason)    static_assert((condition), reason " " #condition)
-#  endif
-#endif
-#ifndef CV_StaticAssert
-#  if !defined(__clang__) && defined(__GNUC__) && (__GNUC__*100 + __GNUC_MINOR__ > 302)
-#    define CV_StaticAssert(condition, reason) ({ extern int __attribute__((error("CV_StaticAssert: " reason " " #condition))) CV_StaticAssert(); ((condition) ? 0 : CV_StaticAssert()); })
-#  else
-namespace cv {
-     template <bool x> struct CV_StaticAssert_failed;
-     template <> struct CV_StaticAssert_failed<true> { enum { val = 1 }; };
-     template<int x> struct CV_StaticAssert_test {};
-}
-#    define CV_StaticAssert(condition, reason)\
-       typedef cv::CV_StaticAssert_test< sizeof(cv::CV_StaticAssert_failed< static_cast<bool>(condition) >) > CVAUX_CONCAT(CV_StaticAssert_failed_at_, __LINE__)
-#  endif
-#endif
-
-// Suppress warning "-Wdeprecated-declarations" / C4996
-#if defined(_MSC_VER)
-    #define CV_DO_PRAGMA(x) __pragma(x)
-#elif defined(__GNUC__)
-    #define CV_DO_PRAGMA(x) _Pragma (#x)
-#else
-    #define CV_DO_PRAGMA(x)
-#endif
-
-#ifdef _MSC_VER
-#define CV_SUPPRESS_DEPRECATED_START \
-    CV_DO_PRAGMA(warning(push)) \
-    CV_DO_PRAGMA(warning(disable: 4996))
-#define CV_SUPPRESS_DEPRECATED_END CV_DO_PRAGMA(warning(pop))
-#elif defined (__clang__) || ((__GNUC__)  && (__GNUC__*100 + __GNUC_MINOR__ > 405))
-#define CV_SUPPRESS_DEPRECATED_START \
-    CV_DO_PRAGMA(GCC diagnostic push) \
-    CV_DO_PRAGMA(GCC diagnostic ignored "-Wdeprecated-declarations")
-#define CV_SUPPRESS_DEPRECATED_END CV_DO_PRAGMA(GCC diagnostic pop)
-#else
-#define CV_SUPPRESS_DEPRECATED_START
-#define CV_SUPPRESS_DEPRECATED_END
-#endif
-
-#define CV_UNUSED(name) (void)name
-
-//! @endcond
-
-// undef problematic defines sometimes defined by system headers (windows.h in particular)
-#undef small
-#undef min
-#undef max
-#undef abs
-#undef Complex
-
-#if defined __cplusplus
-#include <limits>
-#else
-#include <limits.h>
-#endif
-
-#include "opencv2/core/hal/interface.h"
-
-#if defined __ICL
-#  define CV_ICC   __ICL
-#elif defined __ICC
-#  define CV_ICC   __ICC
-#elif defined __ECL
-#  define CV_ICC   __ECL
-#elif defined __ECC
-#  define CV_ICC   __ECC
-#elif defined __INTEL_COMPILER
-#  define CV_ICC   __INTEL_COMPILER
-#endif
-
-#ifndef CV_INLINE
-#  if defined __cplusplus
-#    define CV_INLINE static inline
-#  elif defined _MSC_VER
-#    define CV_INLINE __inline
-#  else
-#    define CV_INLINE static
-#  endif
-#endif
-
-#ifndef CV_ALWAYS_INLINE
-#if defined(__GNUC__) && (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 1))
-#define CV_ALWAYS_INLINE inline __attribute__((always_inline))
-#elif defined(_MSC_VER)
-#define CV_ALWAYS_INLINE __forceinline
-#else
-#define CV_ALWAYS_INLINE inline
-#endif
-#endif
-
-#if defined CV_DISABLE_OPTIMIZATION || (defined CV_ICC && !defined CV_ENABLE_UNROLLED)
-#  define CV_ENABLE_UNROLLED 0
-#else
-#  define CV_ENABLE_UNROLLED 1
-#endif
-
-#ifdef __GNUC__
-#  define CV_DECL_ALIGNED(x) __attribute__ ((aligned (x)))
-#elif defined _MSC_VER
-#  define CV_DECL_ALIGNED(x) __declspec(align(x))
-#else
-#  define CV_DECL_ALIGNED(x)
-#endif
-
-/* CPU features and intrinsics support */
-#define CV_CPU_NONE             0
-#define CV_CPU_MMX              1
-#define CV_CPU_SSE              2
-#define CV_CPU_SSE2             3
-#define CV_CPU_SSE3             4
-#define CV_CPU_SSSE3            5
-#define CV_CPU_SSE4_1           6
-#define CV_CPU_SSE4_2           7
-#define CV_CPU_POPCNT           8
-#define CV_CPU_FP16             9
-#define CV_CPU_AVX              10
-#define CV_CPU_AVX2             11
-#define CV_CPU_FMA3             12
-
-#define CV_CPU_AVX_512F         13
-#define CV_CPU_AVX_512BW        14
-#define CV_CPU_AVX_512CD        15
-#define CV_CPU_AVX_512DQ        16
-#define CV_CPU_AVX_512ER        17
-#define CV_CPU_AVX_512IFMA512   18 // deprecated
-#define CV_CPU_AVX_512IFMA      18
-#define CV_CPU_AVX_512PF        19
-#define CV_CPU_AVX_512VBMI      20
-#define CV_CPU_AVX_512VL        21
-#define CV_CPU_AVX_512VBMI2     22
-#define CV_CPU_AVX_512VNNI      23
-#define CV_CPU_AVX_512BITALG    24
-#define CV_CPU_AVX_512VPOPCNTDQ 25
-#define CV_CPU_AVX_5124VNNIW    26
-#define CV_CPU_AVX_5124FMAPS    27
-
-#define CV_CPU_NEON             100
-
-#define CV_CPU_MSA              150
-
-#define CV_CPU_RISCVV           170
-
-#define CV_CPU_VSX              200
-#define CV_CPU_VSX3             201
-
-#define CV_CPU_RVV              210
-
-// CPU features groups
-#define CV_CPU_AVX512_SKX       256
-#define CV_CPU_AVX512_COMMON    257
-#define CV_CPU_AVX512_KNL       258
-#define CV_CPU_AVX512_KNM       259
-#define CV_CPU_AVX512_CNL       260
-#define CV_CPU_AVX512_CLX       261
-#define CV_CPU_AVX512_ICL       262
-
-// when adding to this list remember to update the following enum
-#define CV_HARDWARE_MAX_FEATURE 512
-
-/** @brief Available CPU features.
-*/
-enum CpuFeatures {
-    CPU_MMX             = 1,
-    CPU_SSE             = 2,
-    CPU_SSE2            = 3,
-    CPU_SSE3            = 4,
-    CPU_SSSE3           = 5,
-    CPU_SSE4_1          = 6,
-    CPU_SSE4_2          = 7,
-    CPU_POPCNT          = 8,
-    CPU_FP16            = 9,
-    CPU_AVX             = 10,
-    CPU_AVX2            = 11,
-    CPU_FMA3            = 12,
-
-    CPU_AVX_512F        = 13,
-    CPU_AVX_512BW       = 14,
-    CPU_AVX_512CD       = 15,
-    CPU_AVX_512DQ       = 16,
-    CPU_AVX_512ER       = 17,
-    CPU_AVX_512IFMA512  = 18, // deprecated
-    CPU_AVX_512IFMA     = 18,
-    CPU_AVX_512PF       = 19,
-    CPU_AVX_512VBMI     = 20,
-    CPU_AVX_512VL       = 21,
-    CPU_AVX_512VBMI2    = 22,
-    CPU_AVX_512VNNI     = 23,
-    CPU_AVX_512BITALG   = 24,
-    CPU_AVX_512VPOPCNTDQ= 25,
-    CPU_AVX_5124VNNIW   = 26,
-    CPU_AVX_5124FMAPS   = 27,
-
-    CPU_NEON            = 100,
-
-    CPU_MSA             = 150,
-
-    CPU_RISCVV          = 170,
-
-    CPU_VSX             = 200,
-    CPU_VSX3            = 201,
-
-    CPU_RVV             = 210,
-
-    CPU_AVX512_SKX      = 256, //!< Skylake-X with AVX-512F/CD/BW/DQ/VL
-    CPU_AVX512_COMMON   = 257, //!< Common instructions AVX-512F/CD for all CPUs that support AVX-512
-    CPU_AVX512_KNL      = 258, //!< Knights Landing with AVX-512F/CD/ER/PF
-    CPU_AVX512_KNM      = 259, //!< Knights Mill with AVX-512F/CD/ER/PF/4FMAPS/4VNNIW/VPOPCNTDQ
-    CPU_AVX512_CNL      = 260, //!< Cannon Lake with AVX-512F/CD/BW/DQ/VL/IFMA/VBMI
-    CPU_AVX512_CLX      = 261, //!< Cascade Lake with AVX-512F/CD/BW/DQ/VL/VNNI
-    CPU_AVX512_ICL      = 262, //!< Ice Lake with AVX-512F/CD/BW/DQ/VL/IFMA/VBMI/VNNI/VBMI2/BITALG/VPOPCNTDQ
-
-    CPU_MAX_FEATURE     = 512  // see CV_HARDWARE_MAX_FEATURE
-};
-
-
-#include "cv_cpu_dispatch.h"
-
-#if !defined(CV_STRONG_ALIGNMENT) && defined(__arm__) && !(defined(__aarch64__) || defined(_M_ARM64))
-// int*, int64* should be propertly aligned pointers on ARMv7
-#define CV_STRONG_ALIGNMENT 1
-#endif
-#if !defined(CV_STRONG_ALIGNMENT)
-#define CV_STRONG_ALIGNMENT 0
-#endif
-
-/* fundamental constants */
-#define CV_PI   3.1415926535897932384626433832795
-#define CV_2PI  6.283185307179586476925286766559
-#define CV_LOG2 0.69314718055994530941723212145818
-
-#if defined __ARM_FP16_FORMAT_IEEE \
-    && !defined __CUDACC__
-#  define CV_FP16_TYPE 1
-#else
-#  define CV_FP16_TYPE 0
-#endif
-
-typedef union Cv16suf
-{
-    short i;
-    ushort u;
-#if CV_FP16_TYPE
-    __fp16 h;
-#endif
-}
-Cv16suf;
-
-typedef union Cv32suf
-{
-    int i;
-    unsigned u;
-    float f;
-}
-Cv32suf;
-
-typedef union Cv64suf
-{
-    int64 i;
-    uint64 u;
-    double f;
-}
-Cv64suf;
-
-#ifndef OPENCV_ABI_COMPATIBILITY
-#define OPENCV_ABI_COMPATIBILITY 400
-#endif
-
-#ifdef __OPENCV_BUILD
-#  define DISABLE_OPENCV_3_COMPATIBILITY
-#  define OPENCV_DISABLE_DEPRECATED_COMPATIBILITY
-#endif
-
-#ifndef CV_EXPORTS
-# if (defined _WIN32 || defined WINCE || defined __CYGWIN__) && defined(CVAPI_EXPORTS)
-#   define CV_EXPORTS __declspec(dllexport)
-# elif defined __GNUC__ && __GNUC__ >= 4 && (defined(CVAPI_EXPORTS) || defined(__APPLE__))
-#   define CV_EXPORTS __attribute__ ((visibility ("default")))
-# endif
-#endif
-
-#ifndef CV_EXPORTS
-# define CV_EXPORTS
-#endif
-
-#ifdef _MSC_VER
-#   define CV_EXPORTS_TEMPLATE
-#else
-#   define CV_EXPORTS_TEMPLATE CV_EXPORTS
-#endif
-
-#ifndef CV_DEPRECATED
-#  if defined(__GNUC__)
-#    define CV_DEPRECATED __attribute__ ((deprecated))
-#  elif defined(_MSC_VER)
-#    define CV_DEPRECATED __declspec(deprecated)
-#  else
-#    define CV_DEPRECATED
-#  endif
-#endif
-
-#ifndef CV_DEPRECATED_EXTERNAL
-#  if defined(__OPENCV_BUILD)
-#    define CV_DEPRECATED_EXTERNAL /* nothing */
-#  else
-#    define CV_DEPRECATED_EXTERNAL CV_DEPRECATED
-#  endif
-#endif
-
-
-#ifndef CV_EXTERN_C
-#  ifdef __cplusplus
-#    define CV_EXTERN_C extern "C"
-#  else
-#    define CV_EXTERN_C
-#  endif
-#endif
-
-/* special informative macros for wrapper generators */
-#define CV_EXPORTS_W CV_EXPORTS
-#define CV_EXPORTS_W_SIMPLE CV_EXPORTS
-#define CV_EXPORTS_AS(synonym) CV_EXPORTS
-#define CV_EXPORTS_W_MAP CV_EXPORTS
-#define CV_IN_OUT
-#define CV_OUT
-#define CV_PROP
-#define CV_PROP_RW
-#define CV_WRAP
-#define CV_WRAP_AS(synonym)
-#define CV_WRAP_MAPPABLE(mappable)
-#define CV_WRAP_PHANTOM(phantom_header)
-#define CV_WRAP_DEFAULT(val)
-
-/****************************************************************************************\
-*                                  Matrix type (Mat)                                     *
-\****************************************************************************************/
-
-#define CV_MAT_CN_MASK          ((CV_CN_MAX - 1) << CV_CN_SHIFT)
-#define CV_MAT_CN(flags)        ((((flags) & CV_MAT_CN_MASK) >> CV_CN_SHIFT) + 1)
-#define CV_MAT_TYPE_MASK        (CV_DEPTH_MAX*CV_CN_MAX - 1)
-#define CV_MAT_TYPE(flags)      ((flags) & CV_MAT_TYPE_MASK)
-#define CV_MAT_CONT_FLAG_SHIFT  14
-#define CV_MAT_CONT_FLAG        (1 << CV_MAT_CONT_FLAG_SHIFT)
-#define CV_IS_MAT_CONT(flags)   ((flags) & CV_MAT_CONT_FLAG)
-#define CV_IS_CONT_MAT          CV_IS_MAT_CONT
-#define CV_SUBMAT_FLAG_SHIFT    15
-#define CV_SUBMAT_FLAG          (1 << CV_SUBMAT_FLAG_SHIFT)
-#define CV_IS_SUBMAT(flags)     ((flags) & CV_MAT_SUBMAT_FLAG)
-
-/** Size of each channel item,
-   0x28442211 = 0010 1000 0100 0100 0010 0010 0001 0001 ~ array of sizeof(arr_type_elem) */
-#define CV_ELEM_SIZE1(type) ((0x28442211 >> CV_MAT_DEPTH(type)*4) & 15)
-
-#define CV_ELEM_SIZE(type) (CV_MAT_CN(type)*CV_ELEM_SIZE1(type))
-
-#ifndef MIN
-#  define MIN(a,b)  ((a) > (b) ? (b) : (a))
-#endif
-
-#ifndef MAX
-#  define MAX(a,b)  ((a) < (b) ? (b) : (a))
-#endif
-
-///////////////////////////////////////// Enum operators ///////////////////////////////////////
-
-/**
-
-Provides compatibility operators for both classical and C++11 enum classes,
-as well as exposing the C++11 enum class members for backwards compatibility
-
-@code
-    // Provides operators required for flag enums
-    CV_ENUM_FLAGS(AccessFlag)
-
-    // Exposes the listed members of the enum class AccessFlag to the current namespace
-    CV_ENUM_CLASS_EXPOSE(AccessFlag, ACCESS_READ [, ACCESS_WRITE [, ...] ]);
-@endcode
-*/
-
-#define __CV_ENUM_CLASS_EXPOSE_1(EnumType, MEMBER_CONST)                                              \
-static const EnumType MEMBER_CONST = EnumType::MEMBER_CONST;                                          \
-
-#define __CV_ENUM_CLASS_EXPOSE_2(EnumType, MEMBER_CONST, ...)                                         \
-__CV_ENUM_CLASS_EXPOSE_1(EnumType, MEMBER_CONST);                                                     \
-__CV_EXPAND(__CV_ENUM_CLASS_EXPOSE_1(EnumType, __VA_ARGS__));                                         \
-
-#define __CV_ENUM_CLASS_EXPOSE_3(EnumType, MEMBER_CONST, ...)                                         \
-__CV_ENUM_CLASS_EXPOSE_1(EnumType, MEMBER_CONST);                                                     \
-__CV_EXPAND(__CV_ENUM_CLASS_EXPOSE_2(EnumType, __VA_ARGS__));                                         \
-
-#define __CV_ENUM_CLASS_EXPOSE_4(EnumType, MEMBER_CONST, ...)                                         \
-__CV_ENUM_CLASS_EXPOSE_1(EnumType, MEMBER_CONST);                                                     \
-__CV_EXPAND(__CV_ENUM_CLASS_EXPOSE_3(EnumType, __VA_ARGS__));                                         \
-
-#define __CV_ENUM_CLASS_EXPOSE_5(EnumType, MEMBER_CONST, ...)                                         \
-__CV_ENUM_CLASS_EXPOSE_1(EnumType, MEMBER_CONST);                                                     \
-__CV_EXPAND(__CV_ENUM_CLASS_EXPOSE_4(EnumType, __VA_ARGS__));                                         \
-
-#define __CV_ENUM_CLASS_EXPOSE_6(EnumType, MEMBER_CONST, ...)                                         \
-__CV_ENUM_CLASS_EXPOSE_1(EnumType, MEMBER_CONST);                                                     \
-__CV_EXPAND(__CV_ENUM_CLASS_EXPOSE_5(EnumType, __VA_ARGS__));                                         \
-
-#define __CV_ENUM_CLASS_EXPOSE_7(EnumType, MEMBER_CONST, ...)                                         \
-__CV_ENUM_CLASS_EXPOSE_1(EnumType, MEMBER_CONST);                                                     \
-__CV_EXPAND(__CV_ENUM_CLASS_EXPOSE_6(EnumType, __VA_ARGS__));                                         \
-
-#define __CV_ENUM_CLASS_EXPOSE_8(EnumType, MEMBER_CONST, ...)                                         \
-__CV_ENUM_CLASS_EXPOSE_1(EnumType, MEMBER_CONST);                                                     \
-__CV_EXPAND(__CV_ENUM_CLASS_EXPOSE_7(EnumType, __VA_ARGS__));                                         \
-
-#define __CV_ENUM_CLASS_EXPOSE_9(EnumType, MEMBER_CONST, ...)                                         \
-__CV_ENUM_CLASS_EXPOSE_1(EnumType, MEMBER_CONST);                                                     \
-__CV_EXPAND(__CV_ENUM_CLASS_EXPOSE_8(EnumType, __VA_ARGS__));                                         \
-
-#define __CV_ENUM_FLAGS_LOGICAL_NOT(EnumType)                                                         \
-static inline bool operator!(const EnumType& val)                                                     \
-{                                                                                                     \
-    typedef std::underlying_type<EnumType>::type UnderlyingType;                                      \
-    return !static_cast<UnderlyingType>(val);                                                         \
-}                                                                                                     \
-
-#define __CV_ENUM_FLAGS_LOGICAL_NOT_EQ(Arg1Type, Arg2Type)                                            \
-static inline bool operator!=(const Arg1Type& a, const Arg2Type& b)                                   \
-{                                                                                                     \
-    return static_cast<int>(a) != static_cast<int>(b);                                                \
-}                                                                                                     \
-
-#define __CV_ENUM_FLAGS_LOGICAL_EQ(Arg1Type, Arg2Type)                                                \
-static inline bool operator==(const Arg1Type& a, const Arg2Type& b)                                   \
-{                                                                                                     \
-    return static_cast<int>(a) == static_cast<int>(b);                                                \
-}                                                                                                     \
-
-#define __CV_ENUM_FLAGS_BITWISE_NOT(EnumType)                                                         \
-static inline EnumType operator~(const EnumType& val)                                                 \
-{                                                                                                     \
-    typedef std::underlying_type<EnumType>::type UnderlyingType;                                      \
-    return static_cast<EnumType>(~static_cast<UnderlyingType>(val));                                  \
-}                                                                                                     \
-
-#define __CV_ENUM_FLAGS_BITWISE_OR(EnumType, Arg1Type, Arg2Type)                                      \
-static inline EnumType operator|(const Arg1Type& a, const Arg2Type& b)                                \
-{                                                                                                     \
-    typedef std::underlying_type<EnumType>::type UnderlyingType;                                      \
-    return static_cast<EnumType>(static_cast<UnderlyingType>(a) | static_cast<UnderlyingType>(b));    \
-}                                                                                                     \
-
-#define __CV_ENUM_FLAGS_BITWISE_AND(EnumType, Arg1Type, Arg2Type)                                     \
-static inline EnumType operator&(const Arg1Type& a, const Arg2Type& b)                                \
-{                                                                                                     \
-    typedef std::underlying_type<EnumType>::type UnderlyingType;                                      \
-    return static_cast<EnumType>(static_cast<UnderlyingType>(a) & static_cast<UnderlyingType>(b));    \
-}                                                                                                     \
-
-#define __CV_ENUM_FLAGS_BITWISE_XOR(EnumType, Arg1Type, Arg2Type)                                     \
-static inline EnumType operator^(const Arg1Type& a, const Arg2Type& b)                                \
-{                                                                                                     \
-    typedef std::underlying_type<EnumType>::type UnderlyingType;                                      \
-    return static_cast<EnumType>(static_cast<UnderlyingType>(a) ^ static_cast<UnderlyingType>(b));    \
-}                                                                                                     \
-
-#define __CV_ENUM_FLAGS_BITWISE_OR_EQ(EnumType, Arg1Type)                                             \
-static inline EnumType& operator|=(EnumType& _this, const Arg1Type& val)                              \
-{                                                                                                     \
-    _this = static_cast<EnumType>(static_cast<int>(_this) | static_cast<int>(val));                   \
-    return _this;                                                                                     \
-}                                                                                                     \
-
-#define __CV_ENUM_FLAGS_BITWISE_AND_EQ(EnumType, Arg1Type)                                            \
-static inline EnumType& operator&=(EnumType& _this, const Arg1Type& val)                              \
-{                                                                                                     \
-    _this = static_cast<EnumType>(static_cast<int>(_this) & static_cast<int>(val));                   \
-    return _this;                                                                                     \
-}                                                                                                     \
-
-#define __CV_ENUM_FLAGS_BITWISE_XOR_EQ(EnumType, Arg1Type)                                            \
-static inline EnumType& operator^=(EnumType& _this, const Arg1Type& val)                              \
-{                                                                                                     \
-    _this = static_cast<EnumType>(static_cast<int>(_this) ^ static_cast<int>(val));                   \
-    return _this;                                                                                     \
-}                                                                                                     \
-
-#define CV_ENUM_CLASS_EXPOSE(EnumType, ...)                                                           \
-__CV_EXPAND(__CV_CAT(__CV_ENUM_CLASS_EXPOSE_, __CV_VA_NUM_ARGS(__VA_ARGS__))(EnumType, __VA_ARGS__)); \
-
-#define CV_ENUM_FLAGS(EnumType)                                                                       \
-__CV_ENUM_FLAGS_LOGICAL_NOT      (EnumType)                                                           \
-__CV_ENUM_FLAGS_LOGICAL_EQ       (EnumType, int)                                                      \
-__CV_ENUM_FLAGS_LOGICAL_NOT_EQ   (EnumType, int)                                                      \
-                                                                                                      \
-__CV_ENUM_FLAGS_BITWISE_NOT      (EnumType)                                                           \
-__CV_ENUM_FLAGS_BITWISE_OR       (EnumType, EnumType, EnumType)                                       \
-__CV_ENUM_FLAGS_BITWISE_AND      (EnumType, EnumType, EnumType)                                       \
-__CV_ENUM_FLAGS_BITWISE_XOR      (EnumType, EnumType, EnumType)                                       \
-                                                                                                      \
-__CV_ENUM_FLAGS_BITWISE_OR_EQ    (EnumType, EnumType)                                                 \
-__CV_ENUM_FLAGS_BITWISE_AND_EQ   (EnumType, EnumType)                                                 \
-__CV_ENUM_FLAGS_BITWISE_XOR_EQ   (EnumType, EnumType)                                                 \
-
-/****************************************************************************************\
-*                                    static analysys                                     *
-\****************************************************************************************/
-
-// In practice, some macro are not processed correctly (noreturn is not detected).
-// We need to use simplified definition for them.
-#ifndef CV_STATIC_ANALYSIS
-# if defined(__KLOCWORK__) || defined(__clang_analyzer__) || defined(__COVERITY__)
-#   define CV_STATIC_ANALYSIS 1
-# endif
-#else
-# if defined(CV_STATIC_ANALYSIS) && !(__CV_CAT(1, CV_STATIC_ANALYSIS) == 1)  // defined and not empty
-#   if 0 == CV_STATIC_ANALYSIS
-#     undef CV_STATIC_ANALYSIS
-#   endif
-# endif
-#endif
-
-/****************************************************************************************\
-*                                    Thread sanitizer                                    *
-\****************************************************************************************/
-#ifndef CV_THREAD_SANITIZER
-# if defined(__has_feature)
-#   if __has_feature(thread_sanitizer)
-#     define CV_THREAD_SANITIZER
-#   endif
-# endif
-#endif
-
-/****************************************************************************************\
-*          exchange-add operation for atomic operations on reference counters            *
-\****************************************************************************************/
-
-#ifdef CV_XADD
-  // allow to use user-defined macro
-#elif defined __GNUC__ || defined __clang__
-#  if defined __clang__ && __clang_major__ >= 3 && !defined __ANDROID__ && !defined __EMSCRIPTEN__ && !defined(__CUDACC__)  && !defined __INTEL_COMPILER
-#    ifdef __ATOMIC_ACQ_REL
-#      define CV_XADD(addr, delta) __c11_atomic_fetch_add((_Atomic(int)*)(addr), delta, __ATOMIC_ACQ_REL)
-#    else
-#      define CV_XADD(addr, delta) __atomic_fetch_add((_Atomic(int)*)(addr), delta, 4)
-#    endif
-#  else
-#    if defined __ATOMIC_ACQ_REL && !defined __clang__
-       // version for gcc >= 4.7
-#      define CV_XADD(addr, delta) (int)__atomic_fetch_add((unsigned*)(addr), (unsigned)(delta), __ATOMIC_ACQ_REL)
-#    else
-#      define CV_XADD(addr, delta) (int)__sync_fetch_and_add((unsigned*)(addr), (unsigned)(delta))
-#    endif
-#  endif
-#elif defined _MSC_VER && !defined RC_INVOKED
-#  include <intrin.h>
-#  define CV_XADD(addr, delta) (int)_InterlockedExchangeAdd((long volatile*)addr, delta)
-#else
-  #ifdef OPENCV_FORCE_UNSAFE_XADD
-    CV_INLINE int CV_XADD(int* addr, int delta) { int tmp = *addr; *addr += delta; return tmp; }
-  #else
-    #error "OpenCV: can't define safe CV_XADD macro for current platform (unsupported). Define CV_XADD macro through custom port header (see OPENCV_INCLUDE_PORT_FILE)"
-  #endif
-#endif
-
-
-/****************************************************************************************\
-*                                  CV_NORETURN attribute                                 *
-\****************************************************************************************/
-
-#ifndef CV_NORETURN
-#  if defined(__GNUC__)
-#    define CV_NORETURN __attribute__((__noreturn__))
-#  elif defined(_MSC_VER) && (_MSC_VER >= 1300)
-#    define CV_NORETURN __declspec(noreturn)
-#  else
-#    define CV_NORETURN /* nothing by default */
-#  endif
-#endif
-
-/****************************************************************************************\
-*                       CV_NODISCARD_STD attribute (C++17)                               *
-* encourages the compiler to issue a warning if the return value is discarded            *
-\****************************************************************************************/
-#ifndef CV_NODISCARD_STD
-#  ifndef __has_cpp_attribute
-//   workaround preprocessor non-compliance https://reviews.llvm.org/D57851
-#    define __has_cpp_attribute(__x) 0
-#  endif
-#  if __has_cpp_attribute(nodiscard)
-#    define CV_NODISCARD_STD [[nodiscard]]
-#  elif __cplusplus >= 201703L
-//   available when compiler is C++17 compliant
-#    define CV_NODISCARD_STD [[nodiscard]]
-#  elif defined(__INTEL_COMPILER)
-     // see above, available when C++17 is enabled
-#  elif defined(_MSC_VER) && _MSC_VER >= 1911 && _MSVC_LANG >= 201703L
-//   available with VS2017 v15.3+ with /std:c++17 or higher; works on functions and classes
-#    define CV_NODISCARD_STD [[nodiscard]]
-#  elif defined(__GNUC__) && (((__GNUC__ * 100) + __GNUC_MINOR__) >= 700) && (__cplusplus >= 201103L)
-//   available with GCC 7.0+; works on functions, works or silently fails on classes
-#    define CV_NODISCARD_STD [[nodiscard]]
-#  elif defined(__GNUC__) && (((__GNUC__ * 100) + __GNUC_MINOR__) >= 408) && (__cplusplus >= 201103L)
-//   available with GCC 4.8+ but it usually does nothing and can fail noisily -- therefore not used
-//   define CV_NODISCARD_STD [[gnu::warn_unused_result]]
-#  endif
-#endif
-#ifndef CV_NODISCARD_STD
-#  define CV_NODISCARD_STD /* nothing by default */
-#endif
-
-
-/****************************************************************************************\
-*                      CV_NODISCARD attribute (deprecated, GCC only)                     *
-* DONT USE: use instead the standard CV_NODISCARD_STD macro above                        *
-*           this legacy method silently fails to issue warning until some version        *
-*           after gcc 6.3.0. Yet with gcc 7+ you can use the above standard method       *
-*           which makes this method useless. Don't use it.                               *
-* @deprecated use instead CV_NODISCARD_STD                                               *
-\****************************************************************************************/
-#ifndef CV_NODISCARD
-#  if defined(__GNUC__)
-#    define CV_NODISCARD __attribute__((__warn_unused_result__))
-#  elif defined(__clang__) && defined(__has_attribute)
-#    if __has_attribute(__warn_unused_result__)
-#      define CV_NODISCARD __attribute__((__warn_unused_result__))
-#    endif
-#  endif
-#endif
-#ifndef CV_NODISCARD
-#  define CV_NODISCARD /* nothing by default */
-#endif
-
-
-/****************************************************************************************\
-*                                    C++ 11                                              *
-\****************************************************************************************/
-#ifndef CV_CXX11
-#  if __cplusplus >= 201103L || (defined(_MSC_VER) && _MSC_VER >= 1800)
-#    define CV_CXX11 1
-#  endif
-#else
-#  if CV_CXX11 == 0
-#    undef CV_CXX11
-#  endif
-#endif
-#ifndef CV_CXX11
-#  error "OpenCV 4.x+ requires enabled C++11 support"
-#endif
-
-#define CV_CXX_MOVE_SEMANTICS 1
-#define CV_CXX_MOVE(x) std::move(x)
-#define CV_CXX_STD_ARRAY 1
-#include <array>
-#ifndef CV_OVERRIDE
-#  define CV_OVERRIDE override
-#endif
-#ifndef CV_FINAL
-#  define CV_FINAL final
-#endif
-
-#ifndef CV_NOEXCEPT
-#  if __cplusplus >= 201103L || (defined(_MSC_VER) && _MSC_VER >= 1900/*MSVS 2015*/)
-#    define CV_NOEXCEPT noexcept
-#  endif
-#endif
-#ifndef CV_NOEXCEPT
-#  define CV_NOEXCEPT
-#endif
-
-#ifndef CV_CONSTEXPR
-#  if __cplusplus >= 201103L || (defined(_MSC_VER) && _MSC_VER >= 1900/*MSVS 2015*/)
-#    define CV_CONSTEXPR constexpr
-#  endif
-#endif
-#ifndef CV_CONSTEXPR
-#  define CV_CONSTEXPR
-#endif
-
-// Integer types portatibility
-#ifdef OPENCV_STDINT_HEADER
-#include OPENCV_STDINT_HEADER
-#elif defined(__cplusplus)
-#if defined(_MSC_VER) && _MSC_VER < 1600 /* MSVS 2010 */
-namespace cv {
-typedef signed char int8_t;
-typedef unsigned char uint8_t;
-typedef signed short int16_t;
-typedef unsigned short uint16_t;
-typedef signed int int32_t;
-typedef unsigned int uint32_t;
-typedef signed __int64 int64_t;
-typedef unsigned __int64 uint64_t;
-}
-#elif defined(_MSC_VER) || __cplusplus >= 201103L
-#include <cstdint>
-namespace cv {
-using std::int8_t;
-using std::uint8_t;
-using std::int16_t;
-using std::uint16_t;
-using std::int32_t;
-using std::uint32_t;
-using std::int64_t;
-using std::uint64_t;
-}
-#else
-#include <stdint.h>
-namespace cv {
-typedef ::int8_t int8_t;
-typedef ::uint8_t uint8_t;
-typedef ::int16_t int16_t;
-typedef ::uint16_t uint16_t;
-typedef ::int32_t int32_t;
-typedef ::uint32_t uint32_t;
-typedef ::int64_t int64_t;
-typedef ::uint64_t uint64_t;
-}
-#endif
-#else // pure C
-#include <stdint.h>
-#endif
-
-#ifdef __cplusplus
-namespace cv
-{
-
-class float16_t
-{
-public:
-#if CV_FP16_TYPE
-
-    float16_t() : h(0) {}
-    explicit float16_t(float x) { h = (__fp16)x; }
-    operator float() const { return (float)h; }
-    static float16_t fromBits(ushort w)
-    {
-        Cv16suf u;
-        u.u = w;
-        float16_t result;
-        result.h = u.h;
-        return result;
-    }
-    static float16_t zero()
-    {
-        float16_t result;
-        result.h = (__fp16)0;
-        return result;
-    }
-    ushort bits() const
-    {
-        Cv16suf u;
-        u.h = h;
-        return u.u;
-    }
-protected:
-    __fp16 h;
-
-#else
-    float16_t() : w(0) {}
-    explicit float16_t(float x)
-    {
-    #if CV_FP16
-        __m128 v = _mm_load_ss(&x);
-        w = (ushort)_mm_cvtsi128_si32(_mm_cvtps_ph(v, 0));
-    #else
-        Cv32suf in;
-        in.f = x;
-        unsigned sign = in.u & 0x80000000;
-        in.u ^= sign;
-
-        if( in.u >= 0x47800000 )
-            w = (ushort)(in.u > 0x7f800000 ? 0x7e00 : 0x7c00);
-        else
-        {
-            if (in.u < 0x38800000)
-            {
-                in.f += 0.5f;
-                w = (ushort)(in.u - 0x3f000000);
-            }
-            else
-            {
-                unsigned t = in.u + 0xc8000fff;
-                w = (ushort)((t + ((in.u >> 13) & 1)) >> 13);
-            }
-        }
-
-        w = (ushort)(w | (sign >> 16));
-    #endif
-    }
-
-    operator float() const
-    {
-    #if CV_FP16
-        float f;
-        _mm_store_ss(&f, _mm_cvtph_ps(_mm_cvtsi32_si128(w)));
-        return f;
-    #else
-        Cv32suf out;
-
-        unsigned t = ((w & 0x7fff) << 13) + 0x38000000;
-        unsigned sign = (w & 0x8000) << 16;
-        unsigned e = w & 0x7c00;
-
-        out.u = t + (1 << 23);
-        out.u = (e >= 0x7c00 ? t + 0x38000000 :
-                 e == 0 ? (static_cast<void>(out.f -= 6.103515625e-05f), out.u) : t) | sign;
-        return out.f;
-    #endif
-    }
-
-    static float16_t fromBits(ushort b)
-    {
-        float16_t result;
-        result.w = b;
-        return result;
-    }
-    static float16_t zero()
-    {
-        float16_t result;
-        result.w = (ushort)0;
-        return result;
-    }
-    ushort bits() const { return w; }
-protected:
-    ushort w;
-
-#endif
-};
-
-}
-#endif
-
-//! @}
-
-#ifndef __cplusplus
-#include "opencv2/core/fast_math.hpp" // define cvRound(double)
-#endif
-
-#endif // OPENCV_CORE_CVDEF_H
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cvstd.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cvstd.hpp
deleted file mode 100644
index 6ce9e4b0..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cvstd.hpp
+++ /dev/null
@@ -1,190 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_CVSTD_HPP
-#define OPENCV_CORE_CVSTD_HPP
-
-#ifndef __cplusplus
-#  error cvstd.hpp header must be compiled as C++
-#endif
-
-#include "opencv2/core/cvdef.h"
-#include <cstddef>
-#include <cstring>
-#include <cctype>
-
-#include <string>
-
-// import useful primitives from stl
-#  include <algorithm>
-#  include <utility>
-#  include <cstdlib> //for abs(int)
-#  include <cmath>
-
-namespace cv
-{
-    static inline uchar abs(uchar a) { return a; }
-    static inline ushort abs(ushort a) { return a; }
-    static inline unsigned abs(unsigned a) { return a; }
-    static inline uint64 abs(uint64 a) { return a; }
-
-    using std::min;
-    using std::max;
-    using std::abs;
-    using std::swap;
-    using std::sqrt;
-    using std::exp;
-    using std::pow;
-    using std::log;
-}
-
-#include "cvstd_wrapper.hpp"
-
-namespace cv {
-
-//! @addtogroup core_utils
-//! @{
-
-//////////////////////////// memory management functions ////////////////////////////
-
-/** @brief Allocates an aligned memory buffer.
-
-The function allocates the buffer of the specified size and returns it. When the buffer size is 16
-bytes or more, the returned buffer is aligned to 16 bytes.
-@param bufSize Allocated buffer size.
- */
-CV_EXPORTS void* fastMalloc(size_t bufSize);
-
-/** @brief Deallocates a memory buffer.
-
-The function deallocates the buffer allocated with fastMalloc . If NULL pointer is passed, the
-function does nothing. C version of the function clears the pointer *pptr* to avoid problems with
-double memory deallocation.
-@param ptr Pointer to the allocated buffer.
- */
-CV_EXPORTS void fastFree(void* ptr);
-
-/*!
-  The STL-compliant memory Allocator based on cv::fastMalloc() and cv::fastFree()
-*/
-template<typename _Tp> class Allocator
-{
-public:
-    typedef _Tp value_type;
-    typedef value_type* pointer;
-    typedef const value_type* const_pointer;
-    typedef value_type& reference;
-    typedef const value_type& const_reference;
-    typedef size_t size_type;
-    typedef ptrdiff_t difference_type;
-    template<typename U> class rebind { typedef Allocator<U> other; };
-
-    explicit Allocator() {}
-    ~Allocator() {}
-    explicit Allocator(Allocator const&) {}
-    template<typename U>
-    explicit Allocator(Allocator<U> const&) {}
-
-    // address
-    pointer address(reference r) { return &r; }
-    const_pointer address(const_reference r) { return &r; }
-
-    pointer allocate(size_type count, const void* =0) { return reinterpret_cast<pointer>(fastMalloc(count * sizeof (_Tp))); }
-    void deallocate(pointer p, size_type) { fastFree(p); }
-
-    void construct(pointer p, const _Tp& v) { new(static_cast<void*>(p)) _Tp(v); }
-    void destroy(pointer p) { p->~_Tp(); }
-
-    size_type max_size() const { return cv::max(static_cast<_Tp>(-1)/sizeof(_Tp), 1); }
-};
-
-//! @} core_utils
-
-//! @endcond
-
-//! @addtogroup core_basic
-//! @{
-
-//////////////////////////////// string class ////////////////////////////////
-
-class CV_EXPORTS FileNode; //for string constructor from FileNode
-
-typedef std::string String;
-
-#ifndef OPENCV_DISABLE_STRING_LOWER_UPPER_CONVERSIONS
-
-//! @cond IGNORED
-namespace details {
-// std::tolower is int->int
-static inline char char_tolower(char ch)
-{
-    return (char)std::tolower((int)ch);
-}
-// std::toupper is int->int
-static inline char char_toupper(char ch)
-{
-    return (char)std::toupper((int)ch);
-}
-} // namespace details
-//! @endcond
-
-static inline std::string toLowerCase(const std::string& str)
-{
-    std::string result(str);
-    std::transform(result.begin(), result.end(), result.begin(), details::char_tolower);
-    return result;
-}
-
-static inline std::string toUpperCase(const std::string& str)
-{
-    std::string result(str);
-    std::transform(result.begin(), result.end(), result.begin(), details::char_toupper);
-    return result;
-}
-
-#endif // OPENCV_DISABLE_STRING_LOWER_UPPER_CONVERSIONS
-
-//! @} core_basic
-} // cv
-
-#endif //OPENCV_CORE_CVSTD_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cvstd.inl.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cvstd.inl.hpp
deleted file mode 100644
index 37ad1e69..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cvstd.inl.hpp
+++ /dev/null
@@ -1,197 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_CVSTDINL_HPP
-#define OPENCV_CORE_CVSTDINL_HPP
-
-#include <complex>
-#include <ostream>
-#include <sstream>
-
-//! @cond IGNORED
-
-#ifdef _MSC_VER
-#pragma warning( push )
-#pragma warning( disable: 4127 )
-#endif
-
-namespace cv
-{
-
-template<typename _Tp> class DataType< std::complex<_Tp> >
-{
-public:
-    typedef std::complex<_Tp>  value_type;
-    typedef value_type         work_type;
-    typedef _Tp                channel_type;
-
-    enum { generic_type = 0,
-           depth        = DataType<channel_type>::depth,
-           channels     = 2,
-           fmt          = DataType<channel_type>::fmt + ((channels - 1) << 8),
-           type         = CV_MAKETYPE(depth, channels) };
-
-    typedef Vec<channel_type, channels> vec_type;
-};
-
-static inline
-std::ostream& operator << (std::ostream& out, Ptr<Formatted> fmtd)
-{
-    fmtd->reset();
-    for(const char* str = fmtd->next(); str; str = fmtd->next())
-        out << str;
-    return out;
-}
-
-static inline
-std::ostream& operator << (std::ostream& out, const Mat& mtx)
-{
-    return out << Formatter::get()->format(mtx);
-}
-
-static inline
-std::ostream& operator << (std::ostream& out, const UMat& m)
-{
-    return out << m.getMat(ACCESS_READ);
-}
-
-template<typename _Tp> static inline
-std::ostream& operator << (std::ostream& out, const Complex<_Tp>& c)
-{
-    return out << "(" << c.re << "," << c.im << ")";
-}
-
-template<typename _Tp> static inline
-std::ostream& operator << (std::ostream& out, const std::vector<Point_<_Tp> >& vec)
-{
-    return out << Formatter::get()->format(Mat(vec));
-}
-
-
-template<typename _Tp> static inline
-std::ostream& operator << (std::ostream& out, const std::vector<Point3_<_Tp> >& vec)
-{
-    return out << Formatter::get()->format(Mat(vec));
-}
-
-
-template<typename _Tp, int m, int n> static inline
-std::ostream& operator << (std::ostream& out, const Matx<_Tp, m, n>& matx)
-{
-    return out << Formatter::get()->format(Mat(matx));
-}
-
-template<typename _Tp> static inline
-std::ostream& operator << (std::ostream& out, const Point_<_Tp>& p)
-{
-    out << "[" << p.x << ", " << p.y << "]";
-    return out;
-}
-
-template<typename _Tp> static inline
-std::ostream& operator << (std::ostream& out, const Point3_<_Tp>& p)
-{
-    out << "[" << p.x << ", " << p.y << ", " << p.z << "]";
-    return out;
-}
-
-template<typename _Tp, int n> static inline
-std::ostream& operator << (std::ostream& out, const Vec<_Tp, n>& vec)
-{
-    out << "[";
-    if (cv::traits::Depth<_Tp>::value <= CV_32S)
-    {
-        for (int i = 0; i < n - 1; ++i) {
-            out << (int)vec[i] << ", ";
-        }
-        out << (int)vec[n-1] << "]";
-    }
-    else
-    {
-        for (int i = 0; i < n - 1; ++i) {
-            out << vec[i] << ", ";
-        }
-        out << vec[n-1] << "]";
-    }
-
-    return out;
-}
-
-template<typename _Tp> static inline
-std::ostream& operator << (std::ostream& out, const Size_<_Tp>& size)
-{
-    return out << "[" << size.width << " x " << size.height << "]";
-}
-
-template<typename _Tp> static inline
-std::ostream& operator << (std::ostream& out, const Rect_<_Tp>& rect)
-{
-    return out << "[" << rect.width << " x " << rect.height << " from (" << rect.x << ", " << rect.y << ")]";
-}
-
-static inline std::ostream& operator << (std::ostream& out, const MatSize& msize)
-{
-    int i, dims = msize.dims();
-    for( i = 0; i < dims; i++ )
-    {
-        out << msize[i];
-        if( i < dims-1 )
-            out << " x ";
-    }
-    return out;
-}
-
-static inline std::ostream &operator<< (std::ostream &s, cv::Range &r)
-{
-    return s << "[" << r.start << " : " << r.end << ")";
-}
-
-} // cv
-
-#ifdef _MSC_VER
-#pragma warning( pop )
-#endif
-
-//! @endcond
-
-#endif // OPENCV_CORE_CVSTDINL_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/cvstd_wrapper.hpp b/projects/llm_framework/include/opencv4/opencv2/core/cvstd_wrapper.hpp
deleted file mode 100644
index 25e0041f..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/cvstd_wrapper.hpp
+++ /dev/null
@@ -1,154 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_CORE_CVSTD_WRAPPER_HPP
-#define OPENCV_CORE_CVSTD_WRAPPER_HPP
-
-#include "opencv2/core/cvdef.h"
-
-#include <string>
-#include <memory>  // std::shared_ptr
-#include <type_traits>  // std::enable_if
-
-namespace cv {
-
-using std::nullptr_t;
-
-//! @addtogroup core_basic
-//! @{
-
-#ifdef CV_DOXYGEN
-
-template <typename _Tp> using Ptr = std::shared_ptr<_Tp>;  // In ideal world it should look like this, but we need some compatibility workarounds below
-
-template<typename _Tp, typename ... A1> static inline
-Ptr<_Tp> makePtr(const A1&... a1) { return std::make_shared<_Tp>(a1...); }
-
-#else  // cv::Ptr with compatibility workarounds
-
-// It should be defined for C-API types only.
-// C++ types should use regular "delete" operator.
-template<typename Y> struct DefaultDeleter;
-#if 0
-{
-    void operator()(Y* p) const;
-};
-#endif
-
-namespace sfinae {
-template<typename C, typename Ret, typename... Args>
-struct has_parenthesis_operator
-{
-private:
-    template<typename T>
-    static CV_CONSTEXPR std::true_type has_parenthesis_operator_check(typename std::is_same<typename std::decay<decltype(std::declval<T>().operator()(std::declval<Args>()...))>::type, Ret>::type*);
-
-    template<typename> static CV_CONSTEXPR std::false_type has_parenthesis_operator_check(...);
-
-    typedef decltype(has_parenthesis_operator_check<C>(0)) type;
-
-public:
-#if __cplusplus >= 201103L || (defined(_MSC_VER) && _MSC_VER >= 1900/*MSVS 2015*/)
-    static CV_CONSTEXPR bool value = type::value;
-#else
-    // support MSVS 2013
-    static const int value = type::value;
-#endif
-};
-} // namespace sfinae
-
-template <typename T, typename = void>
-struct has_custom_delete
-        : public std::false_type {};
-
-// Force has_custom_delete to std::false_type when NVCC is compiling CUDA source files
-#ifndef __CUDACC__
-template <typename T>
-struct has_custom_delete<T, typename std::enable_if< sfinae::has_parenthesis_operator<DefaultDeleter<T>, void, T*>::value >::type >
-        : public std::true_type {};
-#endif
-
-template<typename T>
-struct Ptr : public std::shared_ptr<T>
-{
-#if 0
-    using std::shared_ptr<T>::shared_ptr;  // GCC 5.x can't handle this
-#else
-    inline Ptr() CV_NOEXCEPT : std::shared_ptr<T>() {}
-    inline Ptr(nullptr_t) CV_NOEXCEPT : std::shared_ptr<T>(nullptr) {}
-    template<typename Y, typename D> inline Ptr(Y* p, D d) : std::shared_ptr<T>(p, d) {}
-    template<typename D> inline Ptr(nullptr_t, D d) : std::shared_ptr<T>(nullptr, d) {}
-
-    template<typename Y> inline Ptr(const Ptr<Y>& r, T* ptr) CV_NOEXCEPT : std::shared_ptr<T>(r, ptr) {}
-
-    inline Ptr(const Ptr<T>& o) CV_NOEXCEPT : std::shared_ptr<T>(o) {}
-    inline Ptr(Ptr<T>&& o) CV_NOEXCEPT : std::shared_ptr<T>(std::move(o)) {}
-
-    template<typename Y> inline Ptr(const Ptr<Y>& o) CV_NOEXCEPT : std::shared_ptr<T>(o) {}
-    template<typename Y> inline Ptr(Ptr<Y>&& o) CV_NOEXCEPT : std::shared_ptr<T>(std::move(o)) {}
-#endif
-    inline Ptr(const std::shared_ptr<T>& o) CV_NOEXCEPT : std::shared_ptr<T>(o) {}
-    inline Ptr(std::shared_ptr<T>&& o) CV_NOEXCEPT : std::shared_ptr<T>(std::move(o)) {}
-
-    // Overload with custom DefaultDeleter: Ptr<IplImage>(...)
-    template<typename Y>
-    inline Ptr(const std::true_type&, Y* ptr) : std::shared_ptr<T>(ptr, DefaultDeleter<Y>()) {}
-
-    // Overload without custom deleter: Ptr<std::string>(...);
-    template<typename Y>
-    inline Ptr(const std::false_type&, Y* ptr) : std::shared_ptr<T>(ptr) {}
-
-    template<typename Y = T>
-    inline Ptr(Y* ptr) : Ptr(has_custom_delete<Y>(), ptr) {}
-
-    // Overload with custom DefaultDeleter: Ptr<IplImage>(...)
-    template<typename Y>
-    inline void reset(const std::true_type&, Y* ptr) { std::shared_ptr<T>::reset(ptr, DefaultDeleter<Y>()); }
-
-    // Overload without custom deleter: Ptr<std::string>(...);
-    template<typename Y>
-    inline void reset(const std::false_type&, Y* ptr) { std::shared_ptr<T>::reset(ptr); }
-
-    template<typename Y>
-    inline void reset(Y* ptr) { Ptr<T>::reset(has_custom_delete<Y>(), ptr); }
-
-    template<class Y, class Deleter>
-    void reset(Y* ptr, Deleter d) { std::shared_ptr<T>::reset(ptr, d); }
-
-    void reset() CV_NOEXCEPT { std::shared_ptr<T>::reset(); }
-
-    Ptr& operator=(const Ptr& o) { std::shared_ptr<T>::operator =(o); return *this; }
-    template<typename Y> inline Ptr& operator=(const Ptr<Y>& o) { std::shared_ptr<T>::operator =(o); return *this; }
-
-    T* operator->() const CV_NOEXCEPT { return std::shared_ptr<T>::get();}
-    typename std::add_lvalue_reference<T>::type operator*() const CV_NOEXCEPT { return *std::shared_ptr<T>::get(); }
-
-    // OpenCV 3.x methods (not a part of standard C++ library)
-    inline void release() { std::shared_ptr<T>::reset(); }
-    inline operator T* () const { return std::shared_ptr<T>::get(); }
-    inline bool empty() const { return std::shared_ptr<T>::get() == nullptr; }
-
-    template<typename Y> inline
-    Ptr<Y> staticCast() const CV_NOEXCEPT { return std::static_pointer_cast<Y>(*this); }
-
-    template<typename Y> inline
-    Ptr<Y> constCast() const CV_NOEXCEPT { return std::const_pointer_cast<Y>(*this); }
-
-    template<typename Y> inline
-    Ptr<Y> dynamicCast() const CV_NOEXCEPT { return std::dynamic_pointer_cast<Y>(*this); }
-};
-
-template<typename _Tp, typename ... A1> static inline
-Ptr<_Tp> makePtr(const A1&... a1)
-{
-    static_assert( !has_custom_delete<_Tp>::value, "Can't use this makePtr with custom DefaultDeleter");
-    return (Ptr<_Tp>)std::make_shared<_Tp>(a1...);
-}
-
-#endif // CV_DOXYGEN
-
-//! @} core_basic
-} // cv
-
-#endif //OPENCV_CORE_CVSTD_WRAPPER_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/detail/async_promise.hpp b/projects/llm_framework/include/opencv4/opencv2/core/detail/async_promise.hpp
deleted file mode 100644
index 6eb3fb52..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/detail/async_promise.hpp
+++ /dev/null
@@ -1,71 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_CORE_ASYNC_PROMISE_HPP
-#define OPENCV_CORE_ASYNC_PROMISE_HPP
-
-#include "../async.hpp"
-
-#include "exception_ptr.hpp"
-
-namespace cv {
-
-/** @addtogroup core_async
-@{
-*/
-
-
-/** @brief Provides result of asynchronous operations
-
-*/
-class CV_EXPORTS AsyncPromise
-{
-public:
-    ~AsyncPromise() CV_NOEXCEPT;
-    AsyncPromise() CV_NOEXCEPT;
-    explicit AsyncPromise(const AsyncPromise& o) CV_NOEXCEPT;
-    AsyncPromise& operator=(const AsyncPromise& o) CV_NOEXCEPT;
-    void release() CV_NOEXCEPT;
-
-    /** Returns associated AsyncArray
-    @note Can be called once
-    */
-    AsyncArray getArrayResult();
-
-    /** Stores asynchronous result.
-    @param[in] value result
-    */
-    void setValue(InputArray value);
-
-    // TODO "move" setters
-
-#if CV__EXCEPTION_PTR
-    /** Stores exception.
-    @param[in] exception exception to be raised in AsyncArray
-    */
-    void setException(std::exception_ptr exception);
-#endif
-
-    /** Stores exception.
-    @param[in] exception exception to be raised in AsyncArray
-    */
-    void setException(const cv::Exception& exception);
-
-#ifdef CV_CXX11
-    explicit AsyncPromise(AsyncPromise&& o) { p = o.p; o.p = NULL; }
-    AsyncPromise& operator=(AsyncPromise&& o) CV_NOEXCEPT { std::swap(p, o.p); return *this; }
-#endif
-
-
-    // PImpl
-    typedef struct AsyncArray::Impl Impl; friend struct AsyncArray::Impl;
-    inline void* _getImpl() const CV_NOEXCEPT { return p; }
-protected:
-    Impl* p;
-};
-
-
-//! @}
-} // namespace
-#endif // OPENCV_CORE_ASYNC_PROMISE_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/detail/dispatch_helper.impl.hpp b/projects/llm_framework/include/opencv4/opencv2/core/detail/dispatch_helper.impl.hpp
deleted file mode 100644
index d6ec6769..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/detail/dispatch_helper.impl.hpp
+++ /dev/null
@@ -1,49 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_CORE_DETAIL_DISPATCH_HELPER_IMPL_HPP
-#define OPENCV_CORE_DETAIL_DISPATCH_HELPER_IMPL_HPP
-
-//! @cond IGNORED
-
-namespace cv {
-namespace detail {
-
-template<template<typename> class Functor, typename... Args>
-static inline void depthDispatch(const int depth, Args&&... args)
-{
-    switch (depth)
-    {
-        case CV_8U:
-            Functor<uint8_t>{}(std::forward<Args>(args)...);
-            break;
-        case CV_8S:
-            Functor<int8_t>{}(std::forward<Args>(args)...);
-            break;
-        case CV_16U:
-            Functor<uint16_t>{}(std::forward<Args>(args)...);
-            break;
-        case CV_16S:
-            Functor<int16_t>{}(std::forward<Args>(args)...);
-            break;
-        case CV_32S:
-            Functor<int32_t>{}(std::forward<Args>(args)...);
-            break;
-        case CV_32F:
-            Functor<float>{}(std::forward<Args>(args)...);
-            break;
-        case CV_64F:
-            Functor<double>{}(std::forward<Args>(args)...);
-            break;
-        case CV_16F:
-        default:
-            CV_Error(cv::Error::BadDepth, "Unsupported matrix type.");
-    };
-}
-
-}}
-
-//! @endcond
-
-#endif //OPENCV_CORE_DETAIL_DISPATCH_HELPER_IMPL_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/detail/exception_ptr.hpp b/projects/llm_framework/include/opencv4/opencv2/core/detail/exception_ptr.hpp
deleted file mode 100644
index d98ffc40..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/detail/exception_ptr.hpp
+++ /dev/null
@@ -1,27 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_CORE_DETAILS_EXCEPTION_PTR_H
-#define OPENCV_CORE_DETAILS_EXCEPTION_PTR_H
-
-#ifndef CV__EXCEPTION_PTR
-#  if defined(__ANDROID__) && defined(ATOMIC_INT_LOCK_FREE) && ATOMIC_INT_LOCK_FREE < 2
-#    define CV__EXCEPTION_PTR 0  // Not supported, details: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58938
-#  elif defined(CV_CXX11)
-#    define CV__EXCEPTION_PTR 1
-#  elif defined(_MSC_VER)
-#    define CV__EXCEPTION_PTR (_MSC_VER >= 1600)
-#  elif defined(__clang__)
-#    define CV__EXCEPTION_PTR 0  // C++11 only (see above)
-#  elif defined(__GNUC__) && defined(__GXX_EXPERIMENTAL_CXX0X__)
-#    define CV__EXCEPTION_PTR (__GXX_EXPERIMENTAL_CXX0X__ > 0)
-#  endif
-#endif
-#ifndef CV__EXCEPTION_PTR
-#  define CV__EXCEPTION_PTR 0
-#elif CV__EXCEPTION_PTR
-#  include <exception>  // std::exception_ptr
-#endif
-
-#endif // OPENCV_CORE_DETAILS_EXCEPTION_PTR_H
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/directx.hpp b/projects/llm_framework/include/opencv4/opencv2/core/directx.hpp
deleted file mode 100644
index 056a85a1..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/directx.hpp
+++ /dev/null
@@ -1,184 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2010-2013, Advanced Micro Devices, Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors as is and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the copyright holders or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_DIRECTX_HPP
-#define OPENCV_CORE_DIRECTX_HPP
-
-#include "mat.hpp"
-#include "ocl.hpp"
-
-#if !defined(__d3d11_h__)
-struct ID3D11Device;
-struct ID3D11Texture2D;
-#endif
-
-#if !defined(__d3d10_h__)
-struct ID3D10Device;
-struct ID3D10Texture2D;
-#endif
-
-#if !defined(_D3D9_H_)
-struct IDirect3DDevice9;
-struct IDirect3DDevice9Ex;
-struct IDirect3DSurface9;
-#endif
-
-
-namespace cv { namespace directx {
-
-namespace ocl {
-using namespace cv::ocl;
-
-//! @addtogroup core_directx
-// This section describes OpenCL and DirectX interoperability.
-//
-// To enable DirectX support, configure OpenCV using CMake with WITH_DIRECTX=ON . Note, DirectX is
-// supported only on Windows.
-//
-// To use OpenCL functionality you should first initialize OpenCL context from DirectX resource.
-//
-//! @{
-
-// TODO static functions in the Context class
-//! @brief Creates OpenCL context from D3D11 device
-//
-//! @param pD3D11Device - pointer to D3D11 device
-//! @return Returns reference to OpenCL Context
-CV_EXPORTS Context& initializeContextFromD3D11Device(ID3D11Device* pD3D11Device);
-
-//! @brief Creates OpenCL context from D3D10 device
-//
-//! @param pD3D10Device - pointer to D3D10 device
-//! @return Returns reference to OpenCL Context
-CV_EXPORTS Context& initializeContextFromD3D10Device(ID3D10Device* pD3D10Device);
-
-//! @brief Creates OpenCL context from Direct3DDevice9Ex device
-//
-//! @param pDirect3DDevice9Ex - pointer to Direct3DDevice9Ex device
-//! @return Returns reference to OpenCL Context
-CV_EXPORTS Context& initializeContextFromDirect3DDevice9Ex(IDirect3DDevice9Ex* pDirect3DDevice9Ex);
-
-//! @brief Creates OpenCL context from Direct3DDevice9 device
-//
-//! @param pDirect3DDevice9 - pointer to Direct3Device9 device
-//! @return Returns reference to OpenCL Context
-CV_EXPORTS Context& initializeContextFromDirect3DDevice9(IDirect3DDevice9* pDirect3DDevice9);
-
-//! @}
-
-} // namespace cv::directx::ocl
-
-//! @addtogroup core_directx
-//! @{
-
-//! @brief Converts InputArray to ID3D11Texture2D. If destination texture format is DXGI_FORMAT_NV12 then
-//!        input UMat expected to be in BGR format and data will be downsampled and color-converted to NV12.
-//
-//! @note Note: Destination texture must be allocated by application. Function does memory copy from src to
-//!             pD3D11Texture2D
-//
-//! @param src - source InputArray
-//! @param pD3D11Texture2D - destination D3D11 texture
-CV_EXPORTS void convertToD3D11Texture2D(InputArray src, ID3D11Texture2D* pD3D11Texture2D);
-
-//! @brief Converts ID3D11Texture2D to OutputArray. If input texture format is DXGI_FORMAT_NV12 then
-//!        data will be upsampled and color-converted to BGR format.
-//
-//! @note Note: Destination matrix will be re-allocated if it has not enough memory to match texture size.
-//!             function does memory copy from pD3D11Texture2D to dst
-//
-//! @param pD3D11Texture2D - source D3D11 texture
-//! @param dst             - destination OutputArray
-CV_EXPORTS void convertFromD3D11Texture2D(ID3D11Texture2D* pD3D11Texture2D, OutputArray dst);
-
-//! @brief Converts InputArray to ID3D10Texture2D
-//
-//! @note Note: function does memory copy from src to
-//!             pD3D10Texture2D
-//
-//! @param src             - source InputArray
-//! @param pD3D10Texture2D - destination D3D10 texture
-CV_EXPORTS void convertToD3D10Texture2D(InputArray src, ID3D10Texture2D* pD3D10Texture2D);
-
-//! @brief Converts ID3D10Texture2D to OutputArray
-//
-//! @note Note: function does memory copy from pD3D10Texture2D
-//!             to dst
-//
-//! @param pD3D10Texture2D - source D3D10 texture
-//! @param dst             - destination OutputArray
-CV_EXPORTS void convertFromD3D10Texture2D(ID3D10Texture2D* pD3D10Texture2D, OutputArray dst);
-
-//! @brief Converts InputArray to IDirect3DSurface9
-//
-//! @note Note: function does memory copy from src to
-//!             pDirect3DSurface9
-//
-//! @param src                 - source InputArray
-//! @param pDirect3DSurface9   - destination D3D10 texture
-//! @param surfaceSharedHandle - shared handle
-CV_EXPORTS void convertToDirect3DSurface9(InputArray src, IDirect3DSurface9* pDirect3DSurface9, void* surfaceSharedHandle = NULL);
-
-//! @brief Converts IDirect3DSurface9 to OutputArray
-//
-//! @note Note: function does memory copy from pDirect3DSurface9
-//!             to dst
-//
-//! @param pDirect3DSurface9   - source D3D10 texture
-//! @param dst                 - destination OutputArray
-//! @param surfaceSharedHandle - shared handle
-CV_EXPORTS void convertFromDirect3DSurface9(IDirect3DSurface9* pDirect3DSurface9, OutputArray dst, void* surfaceSharedHandle = NULL);
-
-//! @brief Get OpenCV type from DirectX type
-//! @param iDXGI_FORMAT - enum DXGI_FORMAT for D3D10/D3D11
-//! @return OpenCV type or -1 if there is no equivalent
-CV_EXPORTS int getTypeFromDXGI_FORMAT(const int iDXGI_FORMAT); // enum DXGI_FORMAT for D3D10/D3D11
-
-//! @brief Get OpenCV type from DirectX type
-//! @param iD3DFORMAT - enum D3DTYPE for D3D9
-//! @return OpenCV type or -1 if there is no equivalent
-CV_EXPORTS int getTypeFromD3DFORMAT(const int iD3DFORMAT); // enum D3DTYPE for D3D9
-
-//! @}
-
-} } // namespace cv::directx
-
-#endif // OPENCV_CORE_DIRECTX_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/dualquaternion.hpp b/projects/llm_framework/include/opencv4/opencv2/core/dualquaternion.hpp
deleted file mode 100644
index 1f644e9d..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/dualquaternion.hpp
+++ /dev/null
@@ -1,979 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2020, Huawei Technologies Co., Ltd. All rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//       http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Author: Liangqian Kong <kongliangqian@huawei.com>
-//         Longbu Wang <wanglongbu@huawei.com>
-#ifndef OPENCV_CORE_DUALQUATERNION_HPP
-#define OPENCV_CORE_DUALQUATERNION_HPP
-
-#include <opencv2/core/quaternion.hpp>
-#include <opencv2/core/affine.hpp>
-
-namespace cv{
-//! @addtogroup core
-//! @{
-
-template <typename _Tp> class DualQuat;
-template <typename _Tp> std::ostream& operator<<(std::ostream&, const DualQuat<_Tp>&);
-
-/**
- * Dual quaternions were introduced to describe rotation together with translation while ordinary
- * quaternions can only describe rotation. It can be used for shortest path pose interpolation,
- * local pose optimization or volumetric deformation. More details can be found
- * - https://en.wikipedia.org/wiki/Dual_quaternion
- * - ["A beginners guide to dual-quaternions: what they are, how they work, and how to use them for 3D character hierarchies", Ben Kenwright, 2012](https://borodust.org/public/shared/beginner_dual_quats.pdf)
- * - ["Dual Quaternions", Yan-Bin Jia, 2013](http://web.cs.iastate.edu/~cs577/handouts/dual-quaternion.pdf)
- * - ["Geometric Skinning with Approximate Dual Quaternion Blending", Kavan, 2008](https://www.cs.utah.edu/~ladislav/kavan08geometric/kavan08geometric)
- * - http://rodolphe-vaillant.fr/?e=29
- *
- * A unit dual quaternion can be classically represented as:
- * \f[
- * \begin{equation}
- * \begin{split}
- * \sigma &= \left(r+\frac{\epsilon}{2}tr\right)\\
- * &= [w, x, y, z, w\_, x\_, y\_, z\_]
- * \end{split}
- * \end{equation}
- * \f]
- * where \f$r, t\f$ represents the rotation (ordinary unit quaternion) and translation (pure ordinary quaternion) respectively.
- *
- * A general dual quaternions which consist of two quaternions is usually represented in form of:
- * \f[
- * \sigma = p + \epsilon q
- * \f]
- * where the introduced dual unit \f$\epsilon\f$ satisfies \f$\epsilon^2 = \epsilon^3 =...=0\f$, and \f$p, q\f$ are quaternions.
- *
- * Alternatively, dual quaternions can also be interpreted as four components which are all [dual numbers](https://www.cs.utah.edu/~ladislav/kavan08geometric/kavan08geometric):
- * \f[
- * \sigma = \hat{q}_w + \hat{q}_xi + \hat{q}_yj + \hat{q}_zk
- * \f]
- * If we set \f$\hat{q}_x, \hat{q}_y\f$ and \f$\hat{q}_z\f$ equal to 0, a dual quaternion is transformed to a dual number. see normalize().
- *
- * If you want to create a dual quaternion, you can use:
- *
- * ```
- * using namespace cv;
- * double angle = CV_PI;
- *
- * // create from eight number
- * DualQuatd dq1(1, 2, 3, 4, 5, 6, 7, 8); //p = [1,2,3,4]. q=[5,6,7,8]
- *
- * // create from Vec
- * Vec<double, 8> v{1,2,3,4,5,6,7,8};
- * DualQuatd dq_v{v};
- *
- * // create from two quaternion
- * Quatd p(1, 2, 3, 4);
- * Quatd q(5, 6, 7, 8);
- * DualQuatd dq2 = DualQuatd::createFromQuat(p, q);
- *
- * // create from an angle, an axis and a translation
- * Vec3d axis{0, 0, 1};
- * Vec3d trans{3, 4, 5};
- * DualQuatd dq3 = DualQuatd::createFromAngleAxisTrans(angle, axis, trans);
- *
- * // If you already have an instance of class Affine3, then you can use
- * Affine3d R = dq3.toAffine3();
- * DualQuatd dq4 = DualQuatd::createFromAffine3(R);
- *
- * // or create directly by affine transformation matrix Rt
- * // see createFromMat() in detail for the form of Rt
- * Matx44d Rt = dq3.toMat();
- * DualQuatd dq5 = DualQuatd::createFromMat(Rt);
- *
- * // Any rotation + translation movement can
- * // be expressed as a rotation + translation around the same line in space (expressed by Plucker
- * // coords), and here's a way to represent it this way.
- * Vec3d axis{1, 1, 1}; // axis will be normalized in createFromPitch
- * Vec3d trans{3, 4 ,5};
- * axis = axis / std::sqrt(axis.dot(axis));// The formula for computing moment that I use below requires a normalized axis
- * Vec3d moment = 1.0 / 2 * (trans.cross(axis) + axis.cross(trans.cross(axis)) *
- *                            std::cos(rotation_angle / 2) / std::sin(rotation_angle / 2));
- * double d = trans.dot(qaxis);
- * DualQuatd dq6 = DualQuatd::createFromPitch(angle, d, axis, moment);
- * ```
- *
- * A point \f$v=(x, y, z)\f$ in form of dual quaternion is \f$[1+\epsilon v]=[1,0,0,0,0,x,y,z]\f$.
- * The transformation of a point \f$v_1\f$ to another point \f$v_2\f$ under the dual quaternion \f$\sigma\f$ is
- * \f[
- * 1 + \epsilon v_2 = \sigma * (1 + \epsilon v_1) * \sigma^{\star}
- * \f]
- * where \f$\sigma^{\star}=p^*-\epsilon q^*.\f$
- *
- * A line in the \f$Pl\ddot{u}cker\f$ coordinates \f$(\hat{l}, m)\f$ defined by the dual quaternion \f$l=\hat{l}+\epsilon m\f$.
- * To transform a line, \f[l_2 = \sigma * l_1 * \sigma^*,\f] where \f$\sigma=r+\frac{\epsilon}{2}rt\f$ and
- * \f$\sigma^*=p^*+\epsilon q^*\f$.
- *
- * To extract the Vec<double, 8> or Vec<float, 8>, see toVec();
- *
- * To extract the affine transformation matrix, see toMat();
- *
- * To extract the instance of Affine3, see toAffine3();
- *
- * If two quaternions \f$q_0, q_1\f$ are needed to be interpolated, you can use sclerp()
- * ```
- * DualQuatd::sclerp(q0, q1, t)
- * ```
- * or dqblend().
- * ```
- * DualQuatd::dqblend(q0, q1, t)
- * ```
- * With more than two dual quaternions to be blended, you can use generalize linear dual quaternion blending
- * with the corresponding weights, i.e. gdqblend().
- *
- */
-template <typename _Tp>
-class CV_EXPORTS DualQuat{
-    static_assert(std::is_floating_point<_Tp>::value, "Dual quaternion only make sense with type of float or double");
-    using value_type = _Tp;
-
-public:
-    static constexpr _Tp CV_DUAL_QUAT_EPS = (_Tp)1.e-6;
-
-    DualQuat();
-
-    /**
-     * @brief create from eight same type numbers.
-     */
-    DualQuat(const _Tp w, const _Tp x, const _Tp y, const _Tp z, const _Tp w_, const _Tp x_, const _Tp y_, const _Tp z_);
-
-    /**
-     * @brief create from a double or float vector.
-     */
-    DualQuat(const Vec<_Tp, 8> &q);
-
-    _Tp w, x, y, z, w_, x_, y_, z_;
-
-    /**
-     * @brief create Dual Quaternion from two same type quaternions p and q.
-     * A Dual Quaternion \f$\sigma\f$ has the form:
-     * \f[\sigma = p + \epsilon q\f]
-     * where p and q are defined as follows:
-     * \f[\begin{equation}
-     *    \begin{split}
-     *    p &= w + x\boldsymbol{i} + y\boldsymbol{j} + z\boldsymbol{k}\\
-     *    q &= w\_ + x\_\boldsymbol{i} + y\_\boldsymbol{j} + z\_\boldsymbol{k}.
-     *    \end{split}
-     *   \end{equation}
-     * \f]
-     * The p and q are the real part and dual part respectively.
-     * @param realPart a quaternion, real part of dual quaternion.
-     * @param dualPart a quaternion, dual part of dual quaternion.
-     * @sa Quat
-    */
-    static DualQuat<_Tp> createFromQuat(const Quat<_Tp> &realPart, const Quat<_Tp> &dualPart);
-
-    /**
-     * @brief create a dual quaternion from a rotation angle \f$\theta\f$, a rotation axis
-     * \f$\boldsymbol{u}\f$ and a translation \f$\boldsymbol{t}\f$.
-     * It generates a dual quaternion \f$\sigma\f$ in the form of
-     * \f[\begin{equation}
-     *    \begin{split}
-     *    \sigma &= r + \frac{\epsilon}{2}\boldsymbol{t}r \\
-     *           &= [\cos(\frac{\theta}{2}), \boldsymbol{u}\sin(\frac{\theta}{2})]
-     *           + \frac{\epsilon}{2}[0, \boldsymbol{t}][[\cos(\frac{\theta}{2}),
-     *           \boldsymbol{u}\sin(\frac{\theta}{2})]]\\
-     *           &= \cos(\frac{\theta}{2}) + \boldsymbol{u}\sin(\frac{\theta}{2})
-     *           + \frac{\epsilon}{2}(-(\boldsymbol{t} \cdot \boldsymbol{u})\sin(\frac{\theta}{2})
-     *           + \boldsymbol{t}\cos(\frac{\theta}{2}) + \boldsymbol{u} \times \boldsymbol{t} \sin(\frac{\theta}{2})).
-     *    \end{split}
-     *    \end{equation}\f]
-     * @param angle rotation angle.
-     * @param axis rotation axis.
-     * @param translation a vector of length 3.
-     * @note Axis will be normalized in this function. And translation is applied
-     * after the rotation. Use @ref createFromQuat(r, r * t / 2) to create a dual quaternion
-     * which translation is applied before rotation.
-     * @sa Quat
-     */
-    static DualQuat<_Tp> createFromAngleAxisTrans(const _Tp angle, const Vec<_Tp, 3> &axis, const Vec<_Tp, 3> &translation);
-
-    /**
-     * @brief Transform this dual quaternion to an affine transformation matrix \f$M\f$.
-     * Dual quaternion consists of a rotation \f$r=[a,b,c,d]\f$ and a translation \f$t=[\Delta x,\Delta y,\Delta z]\f$. The
-     * affine transformation matrix \f$M\f$ has the form
-     * \f[
-     * \begin{bmatrix}
-     * 1-2(e_2^2 +e_3^2) &2(e_1e_2-e_0e_3) &2(e_0e_2+e_1e_3) &\Delta x\\
-     * 2(e_0e_3+e_1e_2)  &1-2(e_1^2+e_3^2) &2(e_2e_3-e_0e_1) &\Delta y\\
-     * 2(e_1e_3-e_0e_2)  &2(e_0e_1+e_2e_3) &1-2(e_1^2-e_2^2) &\Delta z\\
-     * 0&0&0&1
-     * \end{bmatrix}
-     * \f]
-     *  if A is a matrix consisting of  n points to be transformed, this could be achieved by
-     * \f[
-     *  new\_A = M * A
-     * \f]
-     * where A has the form
-     * \f[
-     * \begin{bmatrix}
-     * x_0& x_1& x_2&...&x_n\\
-     * y_0& y_1& y_2&...&y_n\\
-     * z_0& z_1& z_2&...&z_n\\
-     * 1&1&1&...&1
-     * \end{bmatrix}
-     * \f]
-     * where the same subscript represent the same point. The size of A should be \f$[4,n]\f$.
-     * and the same size for matrix new_A.
-     * @param _R 4x4 matrix that represents rotations and translation.
-     * @note Translation is applied after the rotation. Use createFromQuat(r, r * t / 2) to create
-     * a dual quaternion which translation is applied before rotation.
-     */
-    static DualQuat<_Tp> createFromMat(InputArray _R);
-
-    /**
-     * @brief create dual quaternion from an affine matrix. The definition of affine matrix can refer to  createFromMat()
-     */
-    static DualQuat<_Tp> createFromAffine3(const Affine3<_Tp> &R);
-
-    /**
-     * @brief A dual quaternion is a vector in form of
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * \sigma &=\boldsymbol{p} + \epsilon \boldsymbol{q}\\
-     * &= \cos\hat{\frac{\theta}{2}}+\overline{\hat{l}}\sin\frac{\hat{\theta}}{2}
-     * \end{split}
-     * \end{equation}
-     * \f]
-     * where \f$\hat{\theta}\f$ is dual angle and \f$\overline{\hat{l}}\f$ is dual axis:
-     * \f[
-     * \hat{\theta}=\theta + \epsilon d,\\
-     * \overline{\hat{l}}= \hat{l} +\epsilon m.
-     * \f]
-     * In this representation, \f$\theta\f$ is rotation angle and \f$(\hat{l},m)\f$ is the screw axis, d is the translation distance along the axis.
-     *
-     * @param angle rotation angle.
-     * @param d translation along the rotation axis.
-     * @param axis rotation axis represented by quaternion with w = 0.
-     * @param moment the moment of line, and it should be orthogonal to axis.
-     * @note Translation is applied after the rotation. Use createFromQuat(r, r * t / 2) to create
-     * a dual quaternion which translation is applied before rotation.
-     */
-    static DualQuat<_Tp> createFromPitch(const _Tp angle, const _Tp d, const Vec<_Tp, 3> &axis, const Vec<_Tp, 3> &moment);
-
-    /**
-     * @brief return a quaternion which represent the real part of dual quaternion.
-     * The definition of real part is in createFromQuat().
-     * @sa createFromQuat, getDualPart
-     */
-    Quat<_Tp> getRealPart() const;
-
-    /**
-     * @brief return a quaternion which represent the dual part of dual quaternion.
-     * The definition of dual part is in createFromQuat().
-     * @sa createFromQuat, getRealPart
-     */
-    Quat<_Tp> getDualPart() const;
-
-    /**
-     * @brief return the conjugate of a dual quaternion.
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * \sigma^* &= (p + \epsilon q)^*
-     *          &= (p^* + \epsilon q^*)
-     * \end{split}
-     * \end{equation}
-     * \f]
-     * @param dq a dual quaternion.
-     */
-    template <typename T>
-    friend DualQuat<T> conjugate(const DualQuat<T> &dq);
-
-    /**
-     * @brief return the conjugate of a dual quaternion.
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * \sigma^* &= (p + \epsilon q)^*
-     *          &= (p^* + \epsilon q^*)
-     * \end{split}
-     * \end{equation}
-     * \f]
-     */
-    DualQuat<_Tp> conjugate() const;
-
-    /**
-     * @brief return the rotation in quaternion form.
-     */
-    Quat<_Tp> getRotation(QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief return the translation vector.
-     * The rotation \f$r\f$ in this dual quaternion \f$\sigma\f$ is applied before translation \f$t\f$.
-     * The dual quaternion \f$\sigma\f$ is defined as
-     * \f[\begin{equation}
-     * \begin{split}
-     * \sigma &= p + \epsilon q \\
-     *        &= r + \frac{\epsilon}{2}{t}r.
-     * \end{split}
-     * \end{equation}\f]
-     * Thus, the translation can be obtained as follows
-     * \f[t = 2qp^*.\f]
-     * @param assumeUnit if @ref QUAT_ASSUME_UNIT, this dual quaternion assume to be a unit dual quaternion
-     * and this function will save some computations.
-     * @note This dual quaternion's translation is applied after the rotation.
-     */
-    Vec<_Tp, 3> getTranslation(QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief return the norm \f$||\sigma||\f$ of dual quaternion \f$\sigma = p + \epsilon q\f$.
-     * \f[
-     *  \begin{equation}
-     *  \begin{split}
-     *  ||\sigma|| &= \sqrt{\sigma * \sigma^*} \\
-     *        &= ||p|| + \epsilon \frac{p \cdot q}{||p||}.
-     *  \end{split}
-     *  \end{equation}
-     *  \f]
-     * Generally speaking, the norm of a not unit dual
-     * quaternion is a dual number. For convenience, we return it in the form of a dual quaternion
-     * , i.e.
-     * \f[ ||\sigma|| = [||p||, 0, 0, 0, \frac{p \cdot q}{||p||}, 0, 0, 0].\f]
-     *
-     * @note The data type of dual number is dual quaternion.
-     */
-    DualQuat<_Tp> norm() const;
-
-    /**
-     * @brief return a normalized dual quaternion.
-     * A dual quaternion can be expressed as
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * \sigma &= p + \epsilon q\\
-     * &=||\sigma||\left(r+\frac{1}{2}tr\right)
-     * \end{split}
-     * \end{equation}
-     * \f]
-     * where \f$r, t\f$ represents the rotation (ordinary quaternion) and translation (pure ordinary quaternion) respectively,
-     * and \f$||\sigma||\f$ is the norm of dual quaternion(a dual number).
-     * A dual quaternion is unit if and only if
-     * \f[
-     * ||p||=1, p \cdot q=0
-     * \f]
-     * where \f$\cdot\f$ means dot product.
-     * The process of normalization is
-     * \f[
-     * \sigma_{u}=\frac{\sigma}{||\sigma||}
-     * \f]
-     * Next, we simply proof \f$\sigma_u\f$ is a unit dual quaternion:
-     * \f[
-     * \renewcommand{\Im}{\operatorname{Im}}
-     * \begin{equation}
-     * \begin{split}
-     * \sigma_{u}=\frac{\sigma}{||\sigma||}&=\frac{p + \epsilon q}{||p||+\epsilon\frac{p\cdot q}{||p||}}\\
-     * &=\frac{p}{||p||}+\epsilon\left(\frac{q}{||p||}-p\frac{p\cdot q}{||p||^3}\right)\\
-     * &=\frac{p}{||p||}+\epsilon\frac{1}{||p||^2}\left(qp^{*}-p\cdot q\right)\frac{p}{||p||}\\
-     * &=\frac{p}{||p||}+\epsilon\frac{1}{||p||^2}\Im(qp^*)\frac{p}{||p||}.\\
-     * \end{split}
-     * \end{equation}
-     * \f]
-     * As expected, the real part is a rotation and dual part is a pure quaternion.
-     */
-    DualQuat<_Tp> normalize() const;
-
-    /**
-     * @brief if \f$\sigma = p + \epsilon q\f$ is a dual quaternion, p is not zero,
-     * the inverse dual quaternion is
-     * \f[\sigma^{-1} = \frac{\sigma^*}{||\sigma||^2}, \f]
-     * or equivalentlly,
-     * \f[\sigma^{-1} = p^{-1} - \epsilon p^{-1}qp^{-1}.\f]
-     * @param dq a dual quaternion.
-     * @param assumeUnit if @ref QUAT_ASSUME_UNIT, dual quaternion dq assume to be a unit dual quaternion
-     * and this function will save some computations.
-     */
-    template <typename T>
-    friend DualQuat<T> inv(const DualQuat<T> &dq, QuatAssumeType assumeUnit);
-
-    /**
-     * @brief if \f$\sigma = p + \epsilon q\f$ is a dual quaternion, p is not zero,
-     * the inverse dual quaternion is
-     * \f[\sigma^{-1} = \frac{\sigma^*}{||\sigma||^2}, \f]
-     * or equivalentlly,
-     * \f[\sigma^{-1} = p^{-1} - \epsilon p^{-1}qp^{-1}.\f]
-     * @param assumeUnit if @ref QUAT_ASSUME_UNIT, this dual quaternion assume to be a unit dual quaternion
-     * and this function will save some computations.
-     */
-    DualQuat<_Tp> inv(QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief return the dot product of two dual quaternion.
-     * @param p other dual quaternion.
-     */
-    _Tp dot(DualQuat<_Tp> p) const;
-
-    /**
-     ** @brief return the value of \f$p^t\f$ where p is a dual quaternion.
-     * This could be calculated as:
-     * \f[
-     * p^t = \exp(t\ln p)
-     * \f]
-     * @param dq a dual quaternion.
-     * @param t index of power function.
-     * @param assumeUnit if @ref QUAT_ASSUME_UNIT, dual quaternion dq assume to be a unit dual quaternion
-     * and this function will save some computations.
-     */
-    template <typename T>
-    friend DualQuat<T> power(const DualQuat<T> &dq, const T t, QuatAssumeType assumeUnit);
-
-    /**
-     ** @brief return the value of \f$p^t\f$ where p is a dual quaternion.
-     * This could be calculated as:
-     * \f[
-     * p^t = \exp(t\ln p)
-     * \f]
-     *
-     * @param t index of power function.
-     * @param assumeUnit if @ref QUAT_ASSUME_UNIT, this dual quaternion assume to be a unit dual quaternion
-     * and this function will save some computations.
-     */
-    DualQuat<_Tp> power(const _Tp t, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief return the value of \f$p^q\f$ where p and q are dual quaternions.
-     * This could be calculated as:
-     * \f[
-     * p^q = \exp(q\ln p)
-     * \f]
-     * @param p a dual quaternion.
-     * @param q a dual quaternion.
-     * @param assumeUnit if @ref QUAT_ASSUME_UNIT, dual quaternion p assume to be a dual unit quaternion
-     * and this function will save some computations.
-     */
-    template <typename T>
-    friend DualQuat<T> power(const DualQuat<T>& p, const DualQuat<T>& q, QuatAssumeType assumeUnit);
-
-    /**
-     * @brief return the value of \f$p^q\f$ where p and q are dual quaternions.
-     * This could be calculated as:
-     * \f[
-     * p^q = \exp(q\ln p)
-     * \f]
-     *
-     * @param q a dual quaternion
-     * @param assumeUnit if @ref QUAT_ASSUME_UNIT, this dual quaternion assume to be a dual unit quaternion
-     * and this function will save some computations.
-     */
-    DualQuat<_Tp> power(const DualQuat<_Tp>& q, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief return the value of exponential function value
-     * @param dq a dual quaternion.
-     */
-    template <typename T>
-    friend DualQuat<T> exp(const DualQuat<T> &dq);
-
-    /**
-     * @brief return the value of exponential function value
-     */
-    DualQuat<_Tp> exp() const;
-
-    /**
-     * @brief return the value of logarithm function value
-     *
-     * @param dq a dual quaternion.
-     * @param assumeUnit if @ref QUAT_ASSUME_UNIT, dual quaternion dq assume to be a unit dual quaternion
-     * and this function will save some computations.
-     */
-    template <typename T>
-    friend DualQuat<T> log(const DualQuat<T> &dq, QuatAssumeType assumeUnit);
-
-    /**
-     * @brief return the value of logarithm function value
-     * @param assumeUnit if @ref QUAT_ASSUME_UNIT, this dual quaternion assume to be a unit dual quaternion
-     * and this function will save some computations.
-     */
-    DualQuat<_Tp> log(QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief Transform this dual quaternion to a vector.
-     */
-    Vec<_Tp, 8> toVec() const;
-
-    /**
-     * @brief Transform this dual quaternion to a affine transformation matrix
-     * the form of matrix, see createFromMat().
-     */
-    Matx<_Tp, 4, 4> toMat(QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-      * @brief Transform this dual quaternion to a instance of Affine3.
-      */
-    Affine3<_Tp> toAffine3(QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief The screw linear interpolation(ScLERP) is an extension of spherical linear interpolation of dual quaternion.
-     * If \f$\sigma_1\f$ and \f$\sigma_2\f$ are two dual quaternions representing the initial and final pose.
-     * The interpolation of ScLERP function can be defined as:
-     * \f[
-     * ScLERP(t;\sigma_1,\sigma_2) = \sigma_1 * (\sigma_1^{-1} * \sigma_2)^t, t\in[0,1]
-     * \f]
-     *
-     * @param q1 a dual quaternion represents a initial pose.
-     * @param q2 a dual quaternion represents a final pose.
-     * @param t interpolation parameter
-     * @param directChange if true, it always return the shortest path.
-     * @param assumeUnit if @ref QUAT_ASSUME_UNIT, this dual quaternion assume to be a unit dual quaternion
-     * and this function will save some computations.
-     *
-     * For example
-     * ```
-     * double angle1 = CV_PI / 2;
-     * Vec3d axis{0, 0, 1};
-     * Vec3d t(0, 0, 3);
-     * DualQuatd initial = DualQuatd::createFromAngleAxisTrans(angle1, axis, t);
-     * double angle2 = CV_PI;
-     * DualQuatd final = DualQuatd::createFromAngleAxisTrans(angle2, axis, t);
-     * DualQuatd inter = DualQuatd::sclerp(initial, final, 0.5);
-     * ```
-     */
-    static DualQuat<_Tp> sclerp(const DualQuat<_Tp> &q1, const DualQuat<_Tp> &q2, const _Tp t,
-                                bool directChange=true, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT);
-    /**
-     * @brief The method of Dual Quaternion linear Blending(DQB) is to compute a transformation between dual quaternion
-     * \f$q_1\f$ and \f$q_2\f$ and can be defined as:
-     * \f[
-     * DQB(t;{\boldsymbol{q}}_1,{\boldsymbol{q}}_2)=
-     * \frac{(1-t){\boldsymbol{q}}_1+t{\boldsymbol{q}}_2}{||(1-t){\boldsymbol{q}}_1+t{\boldsymbol{q}}_2||}.
-     * \f]
-     * where \f$q_1\f$ and \f$q_2\f$ are unit dual quaternions representing the input transformations.
-     * If you want to use DQB that works for more than two rigid transformations, see @ref gdqblend
-     *
-     * @param q1 a unit dual quaternion representing the input transformations.
-     * @param q2 a unit dual quaternion representing the input transformations.
-     * @param t parameter \f$t\in[0,1]\f$.
-     * @param assumeUnit if @ref QUAT_ASSUME_UNIT, this dual quaternion assume to be a unit dual quaternion
-     * and this function will save some computations.
-     *
-     * @sa gdqblend
-     */
-    static DualQuat<_Tp> dqblend(const DualQuat<_Tp> &q1, const DualQuat<_Tp> &q2, const _Tp t,
-                                   QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT);
-
-    /**
-     * @brief The generalized Dual Quaternion linear Blending works for more than two rigid transformations.
-     * If these transformations are expressed as unit dual quaternions \f$q_1,...,q_n\f$ with convex weights
-     * \f$w = (w_1,...,w_n)\f$, the generalized DQB is simply
-     * \f[
-     * gDQB(\boldsymbol{w};{\boldsymbol{q}}_1,...,{\boldsymbol{q}}_n)=\frac{w_1{\boldsymbol{q}}_1+...+w_n{\boldsymbol{q}}_n}
-     * {||w_1{\boldsymbol{q}}_1+...+w_n{\boldsymbol{q}}_n||}.
-     * \f]
-     * @param dualquat vector of dual quaternions
-     * @param weights vector of weights, the size of weights should be the same as dualquat, and the weights should
-     * satisfy \f$\sum_0^n w_{i} = 1\f$ and \f$w_i>0\f$.
-     * @param assumeUnit if @ref QUAT_ASSUME_UNIT, these dual quaternions assume to be unit quaternions
-     * and this function will save some computations.
-     * @note the type of weights' element should be the same as the date type of dual quaternion inside the dualquat.
-     */
-    template <int cn>
-    static DualQuat<_Tp> gdqblend(const Vec<DualQuat<_Tp>, cn> &dualquat, InputArray weights,
-                                QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT);
-
-    /**
-     * @brief The generalized Dual Quaternion linear Blending works for more than two rigid transformations.
-     * If these transformations are expressed as unit dual quaternions \f$q_1,...,q_n\f$ with convex weights
-     * \f$w = (w_1,...,w_n)\f$, the generalized DQB is simply
-     * \f[
-     * gDQB(\boldsymbol{w};{\boldsymbol{q}}_1,...,{\boldsymbol{q}}_n)=\frac{w_1{\boldsymbol{q}}_1+...+w_n{\boldsymbol{q}}_n}
-     * {||w_1{\boldsymbol{q}}_1+...+w_n{\boldsymbol{q}}_n||}.
-     * \f]
-     * @param dualquat The dual quaternions which have 8 channels and 1 row or 1 col.
-     * @param weights vector of weights, the size of weights should be the same as dualquat, and the weights should
-     * satisfy \f$\sum_0^n w_{i} = 1\f$ and \f$w_i>0\f$.
-     * @param assumeUnit if @ref QUAT_ASSUME_UNIT, these dual quaternions assume to be unit quaternions
-     * and this function will save some computations.
-     * @note the type of weights' element should be the same as the date type of dual quaternion inside the dualquat.
-     */
-    static DualQuat<_Tp> gdqblend(InputArray dualquat, InputArray weights,
-                                QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT);
-
-    /**
-     * @brief Return opposite dual quaternion \f$-p\f$
-     * which satisfies \f$p + (-p) = 0.\f$
-     *
-     * For example
-     * ```
-     * DualQuatd q{1, 2, 3, 4, 5, 6, 7, 8};
-     * std::cout << -q << std::endl; // [-1, -2, -3, -4, -5, -6, -7, -8]
-     * ```
-     */
-    DualQuat<_Tp> operator-() const;
-
-    /**
-     * @brief return true if two dual quaternions p and q are nearly equal, i.e. when the absolute
-     * value of each \f$p_i\f$ and \f$q_i\f$ is less than CV_DUAL_QUAT_EPS.
-     */
-    bool operator==(const DualQuat<_Tp>&) const;
-
-    /**
-     * @brief Subtraction operator of two dual quaternions p and q.
-     * It returns a new dual quaternion that each value is the sum of \f$p_i\f$ and \f$-q_i\f$.
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * DualQuatd q{5, 6, 7, 8, 9, 10, 11, 12};
-     * std::cout << p - q << std::endl; //[-4, -4, -4, -4, 4, -4, -4, -4]
-     * ```
-     */
-    DualQuat<_Tp> operator-(const DualQuat<_Tp>&) const;
-
-    /**
-     * @brief Subtraction assignment operator of two dual quaternions p and q.
-     * It subtracts right operand from the left operand and assign the result to left operand.
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * DualQuatd q{5, 6, 7, 8, 9, 10, 11, 12};
-     * p -= q; // equivalent to p = p - q
-     * std::cout << p << std::endl; //[-4, -4, -4, -4, 4, -4, -4, -4]
-     *
-     * ```
-     */
-    DualQuat<_Tp>& operator-=(const DualQuat<_Tp>&);
-
-    /**
-     * @brief Addition operator of two dual quaternions p and q.
-     * It returns a new dual quaternion that each value is the sum of \f$p_i\f$ and \f$q_i\f$.
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * DualQuatd q{5, 6, 7, 8, 9, 10, 11, 12};
-     * std::cout << p + q << std::endl; //[6, 8, 10, 12, 14, 16, 18, 20]
-     * ```
-     */
-    DualQuat<_Tp> operator+(const DualQuat<_Tp>&) const;
-
-    /**
-     * @brief Addition assignment operator of two dual quaternions p and q.
-     * It adds right operand to the left operand and assign the result to left operand.
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * DualQuatd q{5, 6, 7, 8, 9, 10, 11, 12};
-     * p += q; // equivalent to p = p + q
-     * std::cout << p << std::endl; //[6, 8, 10, 12, 14, 16, 18, 20]
-     *
-     * ```
-     */
-    DualQuat<_Tp>& operator+=(const DualQuat<_Tp>&);
-
-    /**
-     * @brief Multiplication assignment operator of two quaternions.
-     * It multiplies right operand with the left operand and assign the result to left operand.
-     *
-     * Rule of dual quaternion multiplication:
-     * The dual quaternion can be written as an ordered pair of quaternions [A, B]. Thus
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p * q &= [A, B][C, D]\\
-     * &=[AC, AD + BC]
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * DualQuatd q{5, 6, 7, 8, 9, 10, 11, 12};
-     * p *= q;
-     * std::cout << p << std::endl; //[-60, 12, 30, 24, -216, 80, 124, 120]
-     * ```
-     */
-    DualQuat<_Tp>& operator*=(const DualQuat<_Tp>&);
-
-    /**
-     * @brief Multiplication assignment operator of a quaternions and a scalar.
-     * It multiplies right operand with the left operand and assign the result to left operand.
-     *
-     * Rule of dual quaternion multiplication with a scalar:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p * s &= [w, x, y, z, w\_, x\_, y\_, z\_] * s\\
-     *  &=[w   s, x   s, y   s, z   s, w\_  \space  s, x\_  \space  s, y\_ \space  s, z\_ \space  s].
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * double s = 2.0;
-     * p *= s;
-     * std::cout << p << std::endl; //[2, 4, 6, 8, 10, 12, 14, 16]
-     * ```
-     * @note the type of scalar should be equal to the dual quaternion.
-     */
-    DualQuat<_Tp> operator*=(const _Tp s);
-
-
-    /**
-     * @brief Multiplication operator of two dual quaternions q and p.
-     * Multiplies values on either side of the operator.
-     *
-     * Rule of dual quaternion multiplication:
-     * The dual quaternion can be written as an ordered pair of quaternions [A, B]. Thus
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p * q &= [A, B][C, D]\\
-     * &=[AC, AD + BC]
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * DualQuatd q{5, 6, 7, 8, 9, 10, 11, 12};
-     * std::cout << p * q << std::endl; //[-60, 12, 30, 24, -216, 80, 124, 120]
-     * ```
-     */
-    DualQuat<_Tp> operator*(const DualQuat<_Tp>&) const;
-
-    /**
-     * @brief Division operator of a dual quaternions and a scalar.
-     * It divides left operand with the right operand and assign the result to left operand.
-     *
-     * Rule of dual quaternion division with a scalar:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p / s &= [w, x, y, z, w\_, x\_, y\_, z\_] / s\\
-     * &=[w/s, x/s, y/s, z/s, w\_/s, x\_/s, y\_/s, z\_/s].
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * double s = 2.0;
-     * p /= s; // equivalent to p = p / s
-     * std::cout << p << std::endl; //[0.5, 1, 1.5, 2, 2.5, 3, 3.5, 4]
-     * ```
-     * @note the type of scalar should be equal to this dual quaternion.
-     */
-    DualQuat<_Tp> operator/(const _Tp s) const;
-
-    /**
-     * @brief Division operator of two dual quaternions p and q.
-     * Divides left hand operand by right hand operand.
-     *
-     * Rule of dual quaternion division with a dual quaternion:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p / q &= p * q.inv()\\
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * DualQuatd q{5, 6, 7, 8, 9, 10, 11, 12};
-     * std::cout << p / q << std::endl; // equivalent to p * q.inv()
-     * ```
-     */
-    DualQuat<_Tp> operator/(const DualQuat<_Tp>&) const;
-
-    /**
-     * @brief Division assignment operator of two dual quaternions p and q;
-     * It divides left operand with the right operand and assign the result to left operand.
-     *
-     * Rule of dual quaternion division with a quaternion:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p / q&= p * q.inv()\\
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * DualQuatd q{5, 6, 7, 8, 9, 10, 11, 12};
-     * p /= q; // equivalent to p = p * q.inv()
-     * std::cout << p << std::endl;
-     * ```
-     */
-    DualQuat<_Tp>& operator/=(const DualQuat<_Tp>&);
-
-    /**
-     * @brief Division assignment operator of a dual quaternions and a scalar.
-     * It divides left operand with the right operand and assign the result to left operand.
-     *
-     * Rule of dual quaternion division with a scalar:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p / s &= [w, x, y, z, w\_, x\_, y\_ ,z\_] / s\\
-     * &=[w / s, x / s, y / s, z / s, w\_ / \space s, x\_ / \space s, y\_ / \space s, z\_ / \space s].
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * double s = 2.0;;
-     * p /= s; // equivalent to p = p / s
-     * std::cout << p << std::endl; //[0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0]
-     * ```
-     * @note the type of scalar should be equal to the dual quaternion.
-     */
-    Quat<_Tp>& operator/=(const _Tp s);
-
-    /**
-     * @brief Addition operator of a scalar and a dual quaternions.
-     * Adds right hand operand from left hand operand.
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * double scalar = 2.0;
-     * std::cout << scalar + p << std::endl; //[3.0, 2, 3, 4, 5, 6, 7, 8]
-     * ```
-     * @note the type of scalar should be equal to the dual quaternion.
-     */
-    template <typename T>
-    friend DualQuat<T> cv::operator+(const T s, const DualQuat<T>&);
-
-    /**
-     * @brief Addition operator of a dual quaternions and a scalar.
-     * Adds right hand operand from left hand operand.
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * double scalar = 2.0;
-     * std::cout << p + scalar << std::endl; //[3.0, 2, 3, 4, 5, 6, 7, 8]
-     * ```
-     * @note the type of scalar should be equal to the dual quaternion.
-     */
-    template <typename T>
-    friend DualQuat<T> cv::operator+(const DualQuat<T>&, const T s);
-
-    /**
-     * @brief Multiplication operator of a scalar and a dual quaternions.
-     * It multiplies right operand with the left operand and assign the result to left operand.
-     *
-     * Rule of dual quaternion multiplication with a scalar:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p * s &= [w, x, y, z, w\_, x\_, y\_, z\_] * s\\
-     * &=[w s, x s, y s, z s, w\_ \space s, x\_ \space s, y\_ \space s, z\_ \space s].
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * double s = 2.0;
-     * std::cout << s * p << std::endl; //[2, 4, 6, 8, 10, 12, 14, 16]
-     * ```
-     * @note the type of scalar should be equal to the dual quaternion.
-     */
-    template <typename T>
-    friend DualQuat<T> cv::operator*(const T s, const DualQuat<T>&);
-
-    /**
-     * @brief Subtraction operator of a dual quaternion and a scalar.
-     * Subtracts right hand operand from left hand operand.
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * double scalar = 2.0;
-     * std::cout << p - scalar << std::endl; //[-1, 2, 3, 4, 5, 6, 7, 8]
-     * ```
-     * @note the type of scalar should be equal to the dual quaternion.
-     */
-    template <typename T>
-    friend DualQuat<T> cv::operator-(const DualQuat<T>&, const T s);
-
-    /**
-     * @brief Subtraction operator of a scalar and a dual quaternions.
-     * Subtracts right hand operand from left hand operand.
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * double scalar = 2.0;
-     * std::cout << scalar - p << std::endl; //[1.0, -2, -3, -4, -5, -6, -7, -8]
-     * ```
-     * @note the type of scalar should be equal to the dual quaternion.
-     */
-    template <typename T>
-    friend DualQuat<T> cv::operator-(const T s, const DualQuat<T>&);
-
-    /**
-     * @brief Multiplication operator of a dual quaternions and a scalar.
-     * It multiplies right operand with the left operand and assign the result to left operand.
-     *
-     * Rule of dual quaternion multiplication with a scalar:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p * s &= [w, x, y, z, w\_, x\_, y\_, z\_] * s\\
-     * &=[w s, x s, y s, z s, w\_ \space s, x\_ \space s, y\_ \space s, z\_ \space s].
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * DualQuatd p{1, 2, 3, 4, 5, 6, 7, 8};
-     * double s = 2.0;
-     * std::cout << p * s << std::endl; //[2, 4, 6, 8, 10, 12, 14, 16]
-     * ```
-     * @note the type of scalar should be equal to the dual quaternion.
-     */
-    template <typename T>
-    friend DualQuat<T> cv::operator*(const DualQuat<T>&, const T s);
-
-    template <typename S>
-    friend std::ostream& cv::operator<<(std::ostream&, const DualQuat<S>&);
-
-};
-
-using DualQuatd = DualQuat<double>;
-using DualQuatf = DualQuat<float>;
-
-//! @} core
-}//namespace
-
-#include "dualquaternion.inl.hpp"
-
-#endif /* OPENCV_CORE_QUATERNION_HPP */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/dualquaternion.inl.hpp b/projects/llm_framework/include/opencv4/opencv2/core/dualquaternion.inl.hpp
deleted file mode 100644
index 6abb1592..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/dualquaternion.inl.hpp
+++ /dev/null
@@ -1,487 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2020, Huawei Technologies Co., Ltd. All rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//       http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Author: Liangqian Kong <kongliangqian@huawei.com>
-//         Longbu Wang <wanglongbu@huawei.com>
-
-#ifndef OPENCV_CORE_DUALQUATERNION_INL_HPP
-#define OPENCV_CORE_DUALQUATERNION_INL_HPP
-
-#ifndef OPENCV_CORE_DUALQUATERNION_HPP
-#error This is not a standalone header. Include dualquaternion.hpp instead.
-#endif
-
-///////////////////////////////////////////////////////////////////////////////////////
-//Implementation
-namespace cv {
-
-template <typename T>
-DualQuat<T>::DualQuat():w(0), x(0), y(0), z(0), w_(0), x_(0), y_(0), z_(0){};
-
-template <typename T>
-DualQuat<T>::DualQuat(const T vw, const T vx, const T vy, const T vz, const T _w, const T _x, const T _y, const T _z):
-                      w(vw), x(vx), y(vy), z(vz), w_(_w), x_(_x), y_(_y), z_(_z){};
-
-template <typename T>
-DualQuat<T>::DualQuat(const Vec<T, 8> &q):w(q[0]), x(q[1]), y(q[2]), z(q[3]),
-                                          w_(q[4]), x_(q[5]), y_(q[6]), z_(q[7]){};
-
-template <typename T>
-DualQuat<T> DualQuat<T>::createFromQuat(const Quat<T> &realPart, const Quat<T> &dualPart)
-{
-    T w = realPart.w;
-    T x = realPart.x;
-    T y = realPart.y;
-    T z = realPart.z;
-    T w_ = dualPart.w;
-    T x_ = dualPart.x;
-    T y_ = dualPart.y;
-    T z_ = dualPart.z;
-    return DualQuat<T>(w, x, y, z, w_, x_, y_, z_);
-}
-
-template <typename T>
-DualQuat<T> DualQuat<T>::createFromAngleAxisTrans(const T angle, const Vec<T, 3> &axis, const Vec<T, 3> &trans)
-{
-    Quat<T> r = Quat<T>::createFromAngleAxis(angle, axis);
-    Quat<T> t{0, trans[0], trans[1], trans[2]};
-    return createFromQuat(r, t * r * T(0.5));
-}
-
-template <typename T>
-DualQuat<T> DualQuat<T>::createFromMat(InputArray _R)
-{
-    CV_CheckTypeEQ(_R.type(), cv::traits::Type<T>::value, "");
-    if (_R.size() != Size(4, 4))
-    {
-        CV_Error(Error::StsBadArg, "The input matrix must have 4 columns and 4 rows");
-    }
-    Mat R = _R.getMat();
-    Quat<T> r = Quat<T>::createFromRotMat(R.colRange(0, 3).rowRange(0, 3));
-    Quat<T> trans(0, R.at<T>(0, 3), R.at<T>(1, 3), R.at<T>(2, 3));
-    return createFromQuat(r, trans * r * T(0.5));
-}
-
-template <typename T>
-DualQuat<T> DualQuat<T>::createFromAffine3(const Affine3<T> &R)
-{
-    return createFromMat(R.matrix);
-}
-
-template <typename T>
-DualQuat<T> DualQuat<T>::createFromPitch(const T angle, const T d, const Vec<T, 3> &axis, const Vec<T, 3> &moment)
-{
-    T half_angle = angle * T(0.5), half_d = d * T(0.5);
-    Quat<T> qaxis = Quat<T>(0, axis[0], axis[1], axis[2]).normalize();
-    Quat<T> qmoment = Quat<T>(0, moment[0], moment[1], moment[2]);
-    qmoment -= qaxis * axis.dot(moment);
-    Quat<T> dual = -half_d * std::sin(half_angle) + std::sin(half_angle) * qmoment +
-        half_d * std::cos(half_angle) * qaxis;
-    return createFromQuat(Quat<T>::createFromAngleAxis(angle, axis), dual);
-}
-
-template <typename T>
-inline bool DualQuat<T>::operator==(const DualQuat<T> &q) const
-{
-    return (abs(w - q.w) < CV_DUAL_QUAT_EPS && abs(x - q.x) < CV_DUAL_QUAT_EPS &&
-            abs(y - q.y) < CV_DUAL_QUAT_EPS && abs(z - q.z) < CV_DUAL_QUAT_EPS &&
-            abs(w_ - q.w_) < CV_DUAL_QUAT_EPS && abs(x_ - q.x_) < CV_DUAL_QUAT_EPS &&
-            abs(y_ - q.y_) < CV_DUAL_QUAT_EPS && abs(z_ - q.z_) < CV_DUAL_QUAT_EPS);
-}
-
-template <typename T>
-inline Quat<T> DualQuat<T>::getRealPart() const
-{
-    return Quat<T>(w, x, y, z);
-}
-
-template <typename T>
-inline Quat<T> DualQuat<T>::getDualPart() const
-{
-    return Quat<T>(w_, x_, y_, z_);
-}
-
-template <typename T>
-inline DualQuat<T> conjugate(const DualQuat<T> &dq)
-{
-    return dq.conjugate();
-}
-
-template <typename T>
-inline DualQuat<T> DualQuat<T>::conjugate() const
-{
-    return DualQuat<T>(w, -x, -y, -z, w_, -x_, -y_, -z_);
-}
-
-template <typename T>
-DualQuat<T> DualQuat<T>::norm() const
-{
-    Quat<T> real = getRealPart();
-    T realNorm = real.norm();
-    Quat<T> dual = getDualPart();
-    if (realNorm < CV_DUAL_QUAT_EPS){
-        return DualQuat<T>(0, 0, 0, 0, 0, 0, 0, 0);
-    }
-    return DualQuat<T>(realNorm, 0, 0, 0, real.dot(dual) / realNorm, 0, 0, 0);
-}
-
-template <typename T>
-inline Quat<T> DualQuat<T>::getRotation(QuatAssumeType assumeUnit) const
-{
-    if (assumeUnit)
-    {
-        return getRealPart();
-    }
-    return getRealPart().normalize();
-}
-
-template <typename T>
-inline Vec<T, 3> DualQuat<T>::getTranslation(QuatAssumeType assumeUnit) const
-{
-    Quat<T> trans = T(2.0) * (getDualPart() * getRealPart().inv(assumeUnit));
-    return Vec<T, 3>{trans[1], trans[2], trans[3]};
-}
-
-template <typename T>
-DualQuat<T> DualQuat<T>::normalize() const
-{
-    Quat<T> p = getRealPart();
-    Quat<T> q = getDualPart();
-    T p_norm = p.norm();
-    if (p_norm < CV_DUAL_QUAT_EPS)
-    {
-        CV_Error(Error::StsBadArg, "Cannot normalize this dual quaternion: the norm is too small.");
-    }
-    Quat<T> p_nr = p / p_norm;
-    Quat<T> q_nr = q / p_norm;
-    return createFromQuat(p_nr, q_nr - p_nr * p_nr.dot(q_nr));
-}
-
-template <typename T>
-inline T DualQuat<T>::dot(DualQuat<T> q) const
-{
-    return q.w * w + q.x * x + q.y * y + q.z * z + q.w_ * w_ + q.x_ * x_ + q.y_ * y_ + q.z_ * z_;
-}
-
-template <typename T>
-inline DualQuat<T> inv(const DualQuat<T> &dq, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT)
-{
-    return dq.inv(assumeUnit);
-}
-
-template <typename T>
-inline DualQuat<T> DualQuat<T>::inv(QuatAssumeType assumeUnit) const
-{
-    Quat<T> real = getRealPart();
-    Quat<T> dual = getDualPart();
-    return createFromQuat(real.inv(assumeUnit), -real.inv(assumeUnit) * dual * real.inv(assumeUnit));
-}
-
-template <typename T>
-inline DualQuat<T> DualQuat<T>::operator-(const DualQuat<T> &q) const
-{
-    return DualQuat<T>(w - q.w, x - q.x, y - q.y, z - q.z, w_ - q.w_, x_ - q.x_, y_ - q.y_, z_ - q.z_);
-}
-
-template <typename T>
-inline DualQuat<T> DualQuat<T>::operator-() const
-{
-    return DualQuat<T>(-w, -x, -y, -z, -w_, -x_, -y_, -z_);
-}
-
-template <typename T>
-inline DualQuat<T> DualQuat<T>::operator+(const DualQuat<T> &q) const
-{
-    return DualQuat<T>(w + q.w, x + q.x, y + q.y, z + q.z, w_ + q.w_, x_ + q.x_, y_ + q.y_, z_ + q.z_);
-}
-
-template <typename T>
-inline DualQuat<T>& DualQuat<T>::operator+=(const DualQuat<T> &q)
-{
-    *this = *this + q;
-    return *this;
-}
-
-template <typename T>
-inline DualQuat<T> DualQuat<T>::operator*(const DualQuat<T> &q) const
-{
-    Quat<T> A = getRealPart();
-    Quat<T> B = getDualPart();
-    Quat<T> C = q.getRealPart();
-    Quat<T> D = q.getDualPart();
-    return DualQuat<T>::createFromQuat(A * C, A * D + B * C);
-}
-
-template <typename T>
-inline DualQuat<T>& DualQuat<T>::operator*=(const DualQuat<T> &q)
-{
-    *this = *this * q;
-    return *this;
-}
-
-template <typename T>
-inline DualQuat<T> operator+(const T a, const DualQuat<T> &q)
-{
-    return DualQuat<T>(a + q.w, q.x, q.y, q.z, q.w_, q.x_, q.y_, q.z_);
-}
-
-template <typename T>
-inline DualQuat<T> operator+(const DualQuat<T> &q, const T a)
-{
-    return DualQuat<T>(a + q.w, q.x, q.y, q.z, q.w_, q.x_, q.y_, q.z_);
-}
-
-template <typename T>
-inline DualQuat<T> operator-(const DualQuat<T> &q, const T a)
-{
-    return DualQuat<T>(q.w - a, q.x, q.y, q.z, q.w_, q.x_, q.y_, q.z_);
-}
-
-template <typename T>
-inline DualQuat<T>& DualQuat<T>::operator-=(const DualQuat<T> &q)
-{
-    *this = *this - q;
-    return *this;
-}
-
-template <typename T>
-inline DualQuat<T> operator-(const T a, const DualQuat<T> &q)
-{
-    return DualQuat<T>(a - q.w, -q.x, -q.y, -q.z, -q.w_, -q.x_, -q.y_, -q.z_);
-}
-
-template <typename T>
-inline DualQuat<T> operator*(const T a, const DualQuat<T> &q)
-{
-    return DualQuat<T>(q.w * a, q.x * a, q.y * a, q.z * a, q.w_ * a, q.x_ * a, q.y_ * a, q.z_ * a);
-}
-
-template <typename T>
-inline DualQuat<T> operator*(const DualQuat<T> &q, const T a)
-{
-    return DualQuat<T>(q.w * a, q.x * a, q.y * a, q.z * a, q.w_ * a, q.x_ * a, q.y_ * a, q.z_ * a);
-}
-
-template <typename T>
-inline DualQuat<T> DualQuat<T>::operator/(const T a) const
-{
-    return DualQuat<T>(w / a, x / a, y / a, z / a, w_ / a, x_ / a, y_ / a, z_ / a);
-}
-
-template <typename T>
-inline DualQuat<T> DualQuat<T>::operator/(const DualQuat<T> &q) const
-{
-    return *this * q.inv();
-}
-
-template <typename T>
-inline DualQuat<T>& DualQuat<T>::operator/=(const DualQuat<T> &q)
-{
-    *this = *this / q;
-    return *this;
-}
-
-template <typename T>
-std::ostream & operator<<(std::ostream &os, const DualQuat<T> &q)
-{
-    os << "DualQuat " << Vec<T, 8>{q.w, q.x, q.y, q.z, q.w_, q.x_, q.y_, q.z_};
-    return os;
-}
-
-template <typename T>
-inline DualQuat<T> exp(const DualQuat<T> &dq)
-{
-    return dq.exp();
-}
-
-namespace detail {
-
-template <typename _Tp>
-Matx<_Tp, 4, 4> jacob_exp(const Quat<_Tp> &q)
-{
-    _Tp nv = std::sqrt(q.x * q.x + q.y * q.y + q.z * q.z);
-    _Tp sinc_nv = abs(nv) < cv::DualQuat<_Tp>::CV_DUAL_QUAT_EPS ? _Tp(1.0) - nv * nv * _Tp(1.0/6.0) : std::sin(nv) / nv;
-    _Tp csiii_nv = abs(nv) < cv::DualQuat<_Tp>::CV_DUAL_QUAT_EPS ? -_Tp(1.0/3.0) : (std::cos(nv) - sinc_nv) / nv / nv;
-    Matx<_Tp, 4, 4> J_exp_quat {
-        std::cos(nv), -sinc_nv * q.x,  -sinc_nv * q.y,  -sinc_nv * q.z,
-        sinc_nv * q.x, csiii_nv * q.x * q.x + sinc_nv, csiii_nv * q.x * q.y, csiii_nv * q.x * q.z,
-        sinc_nv * q.y, csiii_nv * q.y * q.x, csiii_nv * q.y * q.y + sinc_nv, csiii_nv * q.y * q.z,
-        sinc_nv * q.z, csiii_nv * q.z * q.x, csiii_nv * q.z * q.y, csiii_nv * q.z * q.z + sinc_nv
-    };
-    return std::exp(q.w) * J_exp_quat;
-}
-
-} // namespace detail
-
-template <typename T>
-DualQuat<T> DualQuat<T>::exp() const
-{
-    Quat<T> real = getRealPart();
-    return createFromQuat(real.exp(), Quat<T>(detail::jacob_exp(real) * getDualPart().toVec()));
-}
-
-template <typename T>
-DualQuat<T> log(const DualQuat<T> &dq, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT)
-{
-    return dq.log(assumeUnit);
-}
-
-template <typename T>
-DualQuat<T> DualQuat<T>::log(QuatAssumeType assumeUnit) const
-{
-    Quat<T> plog = getRealPart().log(assumeUnit);
-    Matx<T, 4, 4> jacob = detail::jacob_exp(plog);
-    return createFromQuat(plog, Quat<T>(jacob.inv() * getDualPart().toVec()));
-}
-
-template <typename T>
-inline DualQuat<T> power(const DualQuat<T> &dq, const T t, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT)
-{
-    return dq.power(t, assumeUnit);
-}
-
-template <typename T>
-inline DualQuat<T> DualQuat<T>::power(const T t, QuatAssumeType assumeUnit) const
-{
-    return (t * log(assumeUnit)).exp();
-}
-
-template <typename T>
-inline DualQuat<T> power(const DualQuat<T> &p, const DualQuat<T> &q, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT)
-{
-    return p.power(q, assumeUnit);
-}
-
-template <typename T>
-inline DualQuat<T> DualQuat<T>::power(const DualQuat<T> &q, QuatAssumeType assumeUnit) const
-{
-    return (q * log(assumeUnit)).exp();
-}
-
-template <typename T>
-inline Vec<T, 8> DualQuat<T>::toVec() const
-{
-   return Vec<T, 8>(w, x, y, z, w_, x_, y_, z_);
-}
-
-template <typename T>
-Affine3<T> DualQuat<T>::toAffine3(QuatAssumeType assumeUnit) const
-{
-    return Affine3<T>(toMat(assumeUnit));
-}
-
-template <typename T>
-Matx<T, 4, 4> DualQuat<T>::toMat(QuatAssumeType assumeUnit) const
-{
-    Matx<T, 4, 4> rot44 = getRotation(assumeUnit).toRotMat4x4();
-    Vec<T, 3> translation = getTranslation(assumeUnit);
-    rot44(0, 3) = translation[0];
-    rot44(1, 3) = translation[1];
-    rot44(2, 3) = translation[2];
-    return rot44;
-}
-
-template <typename T>
-DualQuat<T> DualQuat<T>::sclerp(const DualQuat<T> &q0, const DualQuat<T> &q1, const T t, bool directChange, QuatAssumeType assumeUnit)
-{
-    DualQuat<T> v0(q0), v1(q1);
-    if (!assumeUnit)
-    {
-        v0 = v0.normalize();
-        v1 = v1.normalize();
-    }
-    Quat<T> v0Real = v0.getRealPart();
-    Quat<T> v1Real = v1.getRealPart();
-    if (directChange && v1Real.dot(v0Real) < 0)
-    {
-        v0 = -v0;
-    }
-    DualQuat<T> v0inv1 = v0.inv() * v1;
-    return v0 * v0inv1.power(t, QUAT_ASSUME_UNIT);
-}
-
-template <typename T>
-DualQuat<T> DualQuat<T>::dqblend(const DualQuat<T> &q1, const DualQuat<T> &q2, const T t, QuatAssumeType assumeUnit)
-{
-    DualQuat<T> v1(q1), v2(q2);
-    if (!assumeUnit)
-    {
-        v1 = v1.normalize();
-        v2 = v2.normalize();
-    }
-    if (v1.getRotation(assumeUnit).dot(v2.getRotation(assumeUnit)) < 0)
-    {
-        return ((1 - t) * v1 - t * v2).normalize();
-    }
-    return ((1 - t) * v1 + t * v2).normalize();
-}
-
-template <typename T>
-DualQuat<T> DualQuat<T>::gdqblend(InputArray _dualquat, InputArray _weight, QuatAssumeType assumeUnit)
-{
-    CV_CheckTypeEQ(_weight.type(), cv::traits::Type<T>::value, "");
-    CV_CheckTypeEQ(_dualquat.type(), CV_MAKETYPE(CV_MAT_DEPTH(cv::traits::Type<T>::value), 8), "");
-    Size dq_s = _dualquat.size();
-    if (dq_s != _weight.size() || (dq_s.height != 1 && dq_s.width != 1))
-    {
-        CV_Error(Error::StsBadArg, "The size of weight must be the same as dualquat, both of them should be (1, n) or (n, 1)");
-    }
-    Mat dualquat = _dualquat.getMat(), weight = _weight.getMat();
-    const int cn = std::max(dq_s.width, dq_s.height);
-    if (!assumeUnit)
-    {
-        for (int i = 0; i < cn; ++i)
-        {
-            dualquat.at<Vec<T, 8>>(i) = DualQuat<T>{dualquat.at<Vec<T, 8>>(i)}.normalize().toVec();
-        }
-    }
-    Vec<T, 8> dq_blend = dualquat.at<Vec<T, 8>>(0) * weight.at<T>(0);
-    Quat<T> q0 = DualQuat<T> {dualquat.at<Vec<T, 8>>(0)}.getRotation(assumeUnit);
-    for (int i = 1; i < cn; ++i)
-    {
-        T k = q0.dot(DualQuat<T>{dualquat.at<Vec<T, 8>>(i)}.getRotation(assumeUnit)) < 0 ? -1: 1;
-        dq_blend = dq_blend + dualquat.at<Vec<T, 8>>(i) * k * weight.at<T>(i);
-    }
-    return DualQuat<T>{dq_blend}.normalize();
-}
-
-template <typename T>
-template <int cn>
-DualQuat<T> DualQuat<T>::gdqblend(const Vec<DualQuat<T>, cn> &_dualquat, InputArray _weight, QuatAssumeType assumeUnit)
-{
-    Vec<DualQuat<T>, cn> dualquat(_dualquat);
-    if (cn == 0)
-    {
-        return DualQuat<T>(1, 0, 0, 0, 0, 0, 0, 0);
-    }
-    Mat dualquat_mat(cn, 1, CV_64FC(8));
-    for (int i = 0; i < cn ; ++i)
-    {
-        dualquat_mat.at<Vec<T, 8>>(i) = dualquat[i].toVec();
-    }
-    return gdqblend(dualquat_mat, _weight, assumeUnit);
-}
-
-} //namespace cv
-
-#endif /*OPENCV_CORE_DUALQUATERNION_INL_HPP*/
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/eigen.hpp b/projects/llm_framework/include/opencv4/opencv2/core/eigen.hpp
deleted file mode 100644
index 51f41474..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/eigen.hpp
+++ /dev/null
@@ -1,402 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-
-#ifndef OPENCV_CORE_EIGEN_HPP
-#define OPENCV_CORE_EIGEN_HPP
-
-#ifndef EIGEN_WORLD_VERSION
-#error "Wrong usage of OpenCV's Eigen utility header. Include Eigen's headers first. See https://github.com/opencv/opencv/issues/17366"
-#endif
-
-#include "opencv2/core.hpp"
-
-#if defined _MSC_VER && _MSC_VER >= 1200
-#define NOMINMAX // fix https://github.com/opencv/opencv/issues/17548
-#pragma warning( disable: 4714 ) //__forceinline is not inlined
-#pragma warning( disable: 4127 ) //conditional expression is constant
-#pragma warning( disable: 4244 ) //conversion from '__int64' to 'int', possible loss of data
-#endif
-
-#if !defined(OPENCV_DISABLE_EIGEN_TENSOR_SUPPORT)
-#if EIGEN_WORLD_VERSION == 3 && EIGEN_MAJOR_VERSION >= 3 \
-    && defined(CV_CXX11) && defined(CV_CXX_STD_ARRAY)
-#include <unsupported/Eigen/CXX11/Tensor>
-#define OPENCV_EIGEN_TENSOR_SUPPORT 1
-#endif  // EIGEN_WORLD_VERSION == 3 && EIGEN_MAJOR_VERSION >= 3
-#endif  // !defined(OPENCV_DISABLE_EIGEN_TENSOR_SUPPORT)
-
-namespace cv
-{
-
-/** @addtogroup core_eigen
-These functions are provided for OpenCV-Eigen interoperability. They convert `Mat`
-objects to corresponding `Eigen::Matrix` objects and vice-versa. Consult the [Eigen
-documentation](https://eigen.tuxfamily.org/dox/group__TutorialMatrixClass.html) for
-information about the `Matrix` template type.
-
-@note Using these functions requires the `Eigen/Dense` or similar header to be
-included before this header.
-*/
-//! @{
-
-#if defined(OPENCV_EIGEN_TENSOR_SUPPORT) || defined(CV_DOXYGEN)
-/** @brief Converts an Eigen::Tensor to a cv::Mat.
-
-The method converts an Eigen::Tensor with shape (H x W x C) to a cv::Mat where:
- H = number of rows
- W = number of columns
- C = number of channels
-
-Usage:
-\code
-Eigen::Tensor<float, 3, Eigen::RowMajor> a_tensor(...);
-// populate tensor with values
-Mat a_mat;
-eigen2cv(a_tensor, a_mat);
-\endcode
-*/
-template <typename _Tp, int _layout> static inline
-void eigen2cv( const Eigen::Tensor<_Tp, 3, _layout> &src, OutputArray dst )
-{
-    if( !(_layout & Eigen::RowMajorBit) )
-    {
-        const std::array<int, 3> shuffle{2, 1, 0};
-        Eigen::Tensor<_Tp, 3, !_layout> row_major_tensor = src.swap_layout().shuffle(shuffle);
-        Mat _src(src.dimension(0), src.dimension(1), CV_MAKETYPE(DataType<_Tp>::type, src.dimension(2)), row_major_tensor.data());
-        _src.copyTo(dst);
-    }
-    else
-    {
-        Mat _src(src.dimension(0), src.dimension(1), CV_MAKETYPE(DataType<_Tp>::type, src.dimension(2)), (void *)src.data());
-        _src.copyTo(dst);
-    }
-}
-
-/** @brief Converts a cv::Mat to an Eigen::Tensor.
-
-The method converts a cv::Mat to an Eigen Tensor with shape (H x W x C) where:
- H = number of rows
- W = number of columns
- C = number of channels
-
-Usage:
-\code
-Mat a_mat(...);
-// populate Mat with values
-Eigen::Tensor<float, 3, Eigen::RowMajor> a_tensor(...);
-cv2eigen(a_mat, a_tensor);
-\endcode
-*/
-template <typename _Tp, int _layout> static inline
-void cv2eigen( const Mat &src, Eigen::Tensor<_Tp, 3, _layout> &dst )
-{
-    if( !(_layout & Eigen::RowMajorBit) )
-    {
-        Eigen::Tensor<_Tp, 3, !_layout> row_major_tensor(src.rows, src.cols, src.channels());
-        Mat _dst(src.rows, src.cols, CV_MAKETYPE(DataType<_Tp>::type, src.channels()), row_major_tensor.data());
-        if (src.type() == _dst.type())
-            src.copyTo(_dst);
-        else
-            src.convertTo(_dst, _dst.type());
-        const std::array<int, 3> shuffle{2, 1, 0};
-        dst = row_major_tensor.swap_layout().shuffle(shuffle);
-    }
-    else
-    {
-        dst.resize(src.rows, src.cols, src.channels());
-        Mat _dst(src.rows, src.cols, CV_MAKETYPE(DataType<_Tp>::type, src.channels()), dst.data());
-        if (src.type() == _dst.type())
-            src.copyTo(_dst);
-        else
-            src.convertTo(_dst, _dst.type());
-    }
-}
-
-/** @brief Maps cv::Mat data to an Eigen::TensorMap.
-
-The method wraps an existing Mat data array with an Eigen TensorMap of shape (H x W x C) where:
- H = number of rows
- W = number of columns
- C = number of channels
-
-Explicit instantiation of the return type is required.
-
-@note Caller should be aware of the lifetime of the cv::Mat instance and take appropriate safety measures.
-The cv::Mat instance will retain ownership of the data and the Eigen::TensorMap will lose access when the cv::Mat data is deallocated.
-
-The example below initializes a cv::Mat and produces an Eigen::TensorMap:
-\code
-float arr[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11};
-Mat a_mat(2, 2, CV_32FC3, arr);
-Eigen::TensorMap<Eigen::Tensor<float, 3, Eigen::RowMajor>> a_tensormap = cv2eigen_tensormap<float>(a_mat);
-\endcode
-*/
-template <typename _Tp> static inline
-Eigen::TensorMap<Eigen::Tensor<_Tp, 3, Eigen::RowMajor>> cv2eigen_tensormap(InputArray src)
-{
-    Mat mat = src.getMat();
-    CV_CheckTypeEQ(mat.type(), CV_MAKETYPE(traits::Type<_Tp>::value, mat.channels()), "");
-    return Eigen::TensorMap<Eigen::Tensor<_Tp, 3, Eigen::RowMajor>>((_Tp *)mat.data, mat.rows, mat.cols, mat.channels());
-}
-#endif // OPENCV_EIGEN_TENSOR_SUPPORT
-
-template<typename _Tp, int _rows, int _cols, int _options, int _maxRows, int _maxCols> static inline
-void eigen2cv( const Eigen::Matrix<_Tp, _rows, _cols, _options, _maxRows, _maxCols>& src, OutputArray dst )
-{
-    if( !(src.Flags & Eigen::RowMajorBit) )
-    {
-        Mat _src(src.cols(), src.rows(), traits::Type<_Tp>::value,
-              (void*)src.data(), src.outerStride()*sizeof(_Tp));
-        transpose(_src, dst);
-    }
-    else
-    {
-        Mat _src(src.rows(), src.cols(), traits::Type<_Tp>::value,
-                 (void*)src.data(), src.outerStride()*sizeof(_Tp));
-        _src.copyTo(dst);
-    }
-}
-
-// Matx case
-template<typename _Tp, int _rows, int _cols, int _options, int _maxRows, int _maxCols> static inline
-void eigen2cv( const Eigen::Matrix<_Tp, _rows, _cols, _options, _maxRows, _maxCols>& src,
-               Matx<_Tp, _rows, _cols>& dst )
-{
-    if( !(src.Flags & Eigen::RowMajorBit) )
-    {
-        dst = Matx<_Tp, _cols, _rows>(static_cast<const _Tp*>(src.data())).t();
-    }
-    else
-    {
-        dst = Matx<_Tp, _rows, _cols>(static_cast<const _Tp*>(src.data()));
-    }
-}
-
-template<typename _Tp, int _rows, int _cols, int _options, int _maxRows, int _maxCols> static inline
-void cv2eigen( const Mat& src,
-               Eigen::Matrix<_Tp, _rows, _cols, _options, _maxRows, _maxCols>& dst )
-{
-    CV_DbgAssert(src.rows == _rows && src.cols == _cols);
-    if( !(dst.Flags & Eigen::RowMajorBit) )
-    {
-        const Mat _dst(src.cols, src.rows, traits::Type<_Tp>::value,
-                 dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        if( src.type() == _dst.type() )
-            transpose(src, _dst);
-        else if( src.cols == src.rows )
-        {
-            src.convertTo(_dst, _dst.type());
-            transpose(_dst, _dst);
-        }
-        else
-            Mat(src.t()).convertTo(_dst, _dst.type());
-    }
-    else
-    {
-        const Mat _dst(src.rows, src.cols, traits::Type<_Tp>::value,
-                 dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        src.convertTo(_dst, _dst.type());
-    }
-}
-
-// Matx case
-template<typename _Tp, int _rows, int _cols, int _options, int _maxRows, int _maxCols> static inline
-void cv2eigen( const Matx<_Tp, _rows, _cols>& src,
-               Eigen::Matrix<_Tp, _rows, _cols, _options, _maxRows, _maxCols>& dst )
-{
-    if( !(dst.Flags & Eigen::RowMajorBit) )
-    {
-        const Mat _dst(_cols, _rows, traits::Type<_Tp>::value,
-                 dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        transpose(src, _dst);
-    }
-    else
-    {
-        const Mat _dst(_rows, _cols, traits::Type<_Tp>::value,
-                 dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        Mat(src).copyTo(_dst);
-    }
-}
-
-template<typename _Tp>  static inline
-void cv2eigen( const Mat& src,
-               Eigen::Matrix<_Tp, Eigen::Dynamic, Eigen::Dynamic>& dst )
-{
-    dst.resize(src.rows, src.cols);
-    if( !(dst.Flags & Eigen::RowMajorBit) )
-    {
-        const Mat _dst(src.cols, src.rows, traits::Type<_Tp>::value,
-             dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        if( src.type() == _dst.type() )
-            transpose(src, _dst);
-        else if( src.cols == src.rows )
-        {
-            src.convertTo(_dst, _dst.type());
-            transpose(_dst, _dst);
-        }
-        else
-            Mat(src.t()).convertTo(_dst, _dst.type());
-    }
-    else
-    {
-        const Mat _dst(src.rows, src.cols, traits::Type<_Tp>::value,
-                 dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        src.convertTo(_dst, _dst.type());
-    }
-}
-
-// Matx case
-template<typename _Tp, int _rows, int _cols> static inline
-void cv2eigen( const Matx<_Tp, _rows, _cols>& src,
-               Eigen::Matrix<_Tp, Eigen::Dynamic, Eigen::Dynamic>& dst )
-{
-    dst.resize(_rows, _cols);
-    if( !(dst.Flags & Eigen::RowMajorBit) )
-    {
-        const Mat _dst(_cols, _rows, traits::Type<_Tp>::value,
-             dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        transpose(src, _dst);
-    }
-    else
-    {
-        const Mat _dst(_rows, _cols, traits::Type<_Tp>::value,
-                 dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        Mat(src).copyTo(_dst);
-    }
-}
-
-template<typename _Tp> static inline
-void cv2eigen( const Mat& src,
-               Eigen::Matrix<_Tp, Eigen::Dynamic, 1>& dst )
-{
-    CV_Assert(src.cols == 1);
-    dst.resize(src.rows);
-
-    if( !(dst.Flags & Eigen::RowMajorBit) )
-    {
-        const Mat _dst(src.cols, src.rows, traits::Type<_Tp>::value,
-                 dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        if( src.type() == _dst.type() )
-            transpose(src, _dst);
-        else
-            Mat(src.t()).convertTo(_dst, _dst.type());
-    }
-    else
-    {
-        const Mat _dst(src.rows, src.cols, traits::Type<_Tp>::value,
-                 dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        src.convertTo(_dst, _dst.type());
-    }
-}
-
-// Matx case
-template<typename _Tp, int _rows> static inline
-void cv2eigen( const Matx<_Tp, _rows, 1>& src,
-               Eigen::Matrix<_Tp, Eigen::Dynamic, 1>& dst )
-{
-    dst.resize(_rows);
-
-    if( !(dst.Flags & Eigen::RowMajorBit) )
-    {
-        const Mat _dst(1, _rows, traits::Type<_Tp>::value,
-                 dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        transpose(src, _dst);
-    }
-    else
-    {
-        const Mat _dst(_rows, 1, traits::Type<_Tp>::value,
-                 dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        src.copyTo(_dst);
-    }
-}
-
-
-template<typename _Tp> static inline
-void cv2eigen( const Mat& src,
-               Eigen::Matrix<_Tp, 1, Eigen::Dynamic>& dst )
-{
-    CV_Assert(src.rows == 1);
-    dst.resize(src.cols);
-    if( !(dst.Flags & Eigen::RowMajorBit) )
-    {
-        const Mat _dst(src.cols, src.rows, traits::Type<_Tp>::value,
-                 dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        if( src.type() == _dst.type() )
-            transpose(src, _dst);
-        else
-            Mat(src.t()).convertTo(_dst, _dst.type());
-    }
-    else
-    {
-        const Mat _dst(src.rows, src.cols, traits::Type<_Tp>::value,
-                 dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        src.convertTo(_dst, _dst.type());
-    }
-}
-
-//Matx
-template<typename _Tp, int _cols> static inline
-void cv2eigen( const Matx<_Tp, 1, _cols>& src,
-               Eigen::Matrix<_Tp, 1, Eigen::Dynamic>& dst )
-{
-    dst.resize(_cols);
-    if( !(dst.Flags & Eigen::RowMajorBit) )
-    {
-        const Mat _dst(_cols, 1, traits::Type<_Tp>::value,
-                 dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        transpose(src, _dst);
-    }
-    else
-    {
-        const Mat _dst(1, _cols, traits::Type<_Tp>::value,
-                 dst.data(), (size_t)(dst.outerStride()*sizeof(_Tp)));
-        Mat(src).copyTo(_dst);
-    }
-}
-
-//! @}
-
-} // cv
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/fast_math.hpp b/projects/llm_framework/include/opencv4/opencv2/core/fast_math.hpp
deleted file mode 100644
index eb4fbe21..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/fast_math.hpp
+++ /dev/null
@@ -1,411 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Copyright (C) 2015, Itseez Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_FAST_MATH_HPP
-#define OPENCV_CORE_FAST_MATH_HPP
-
-#include "opencv2/core/cvdef.h"
-
-//! @addtogroup core_utils
-//! @{
-
-/****************************************************************************************\
-*                                      fast math                                         *
-\****************************************************************************************/
-
-#ifdef __cplusplus
-#  include <cmath>
-#else
-#  ifdef __BORLANDC__
-#    include <fastmath.h>
-#  else
-#    include <math.h>
-#  endif
-#endif
-
-#if defined(__CUDACC__)
-  // nothing, intrinsics/asm code is not supported
-#else
-  #if ((defined _MSC_VER && defined _M_X64) \
-      || (defined __GNUC__ && defined __x86_64__ && defined __SSE2__)) \
-      && !defined(OPENCV_SKIP_INCLUDE_EMMINTRIN_H)
-    #include <emmintrin.h>
-  #endif
-
-  #if defined __PPC64__ && defined __GNUC__ && defined _ARCH_PWR8 \
-      && !defined(OPENCV_SKIP_INCLUDE_ALTIVEC_H)
-    #include <altivec.h>
-    #undef vector
-    #undef bool
-    #undef pixel
-  #endif
-
-  #if defined(CV_INLINE_ROUND_FLT)
-    // user-specified version
-    // CV_INLINE_ROUND_DBL should be defined too
-  #elif defined __GNUC__ && defined __arm__ && (defined __ARM_PCS_VFP || defined __ARM_VFPV3__ || defined __ARM_NEON__) && !defined __SOFTFP__
-    // 1. general scheme
-    #define ARM_ROUND(_value, _asm_string) \
-        int res; \
-        float temp; \
-        CV_UNUSED(temp); \
-        __asm__(_asm_string : [res] "=r" (res), [temp] "=w" (temp) : [value] "w" (_value)); \
-        return res
-    // 2. version for double
-    #ifdef __clang__
-        #define CV_INLINE_ROUND_DBL(value) ARM_ROUND(value, "vcvtr.s32.f64 %[temp], %[value] \n vmov %[res], %[temp]")
-    #else
-        #define CV_INLINE_ROUND_DBL(value) ARM_ROUND(value, "vcvtr.s32.f64 %[temp], %P[value] \n vmov %[res], %[temp]")
-    #endif
-    // 3. version for float
-    #define CV_INLINE_ROUND_FLT(value) ARM_ROUND(value, "vcvtr.s32.f32 %[temp], %[value]\n vmov %[res], %[temp]")
-  #elif defined __PPC64__ && defined __GNUC__ && defined _ARCH_PWR8
-    // P8 and newer machines can convert fp32/64 to int quickly.
-    #define CV_INLINE_ROUND_DBL(value) \
-        int out; \
-        double temp; \
-        __asm__( "fctiw %[temp],%[in]\n\tmfvsrwz %[out],%[temp]\n\t" : [out] "=r" (out), [temp] "=d" (temp) : [in] "d" ((double)(value)) : ); \
-        return out;
-
-    // FP32 also works with FP64 routine above
-    #define CV_INLINE_ROUND_FLT(value) CV_INLINE_ROUND_DBL(value)
-  #endif
-
-  #ifdef CV_INLINE_ISINF_FLT
-    // user-specified version
-    // CV_INLINE_ISINF_DBL should be defined too
-  #elif defined __PPC64__ && defined _ARCH_PWR9 && defined(scalar_test_data_class)
-    #define CV_INLINE_ISINF_DBL(value) return scalar_test_data_class(value, 0x30);
-    #define CV_INLINE_ISINF_FLT(value) CV_INLINE_ISINF_DBL(value)
-  #endif
-
-  #ifdef CV_INLINE_ISNAN_FLT
-    // user-specified version
-    // CV_INLINE_ISNAN_DBL should be defined too
-  #elif defined __PPC64__ && defined _ARCH_PWR9 && defined(scalar_test_data_class)
-    #define CV_INLINE_ISNAN_DBL(value) return scalar_test_data_class(value, 0x40);
-    #define CV_INLINE_ISNAN_FLT(value) CV_INLINE_ISNAN_DBL(value)
-  #endif
-
-  #if !defined(OPENCV_USE_FASTMATH_BUILTINS) \
-    && ( \
-        defined(__x86_64__) || defined(__i686__) \
-        || defined(__arm__) \
-        || defined(__PPC64__) \
-    )
-    /* Let builtin C math functions when available. Dedicated hardware is available to
-       round and convert FP values. */
-    #define OPENCV_USE_FASTMATH_BUILTINS 1
-  #endif
-
-  /* Enable builtin math functions if possible, desired, and available.
-     Note, not all math functions inline equally. E.g lrint will not inline
-     without the -fno-math-errno option. */
-  #if defined(CV_ICC)
-    // nothing
-  #elif defined(OPENCV_USE_FASTMATH_BUILTINS) && OPENCV_USE_FASTMATH_BUILTINS
-    #if defined(__clang__)
-      #define CV__FASTMATH_ENABLE_CLANG_MATH_BUILTINS
-      #if !defined(CV_INLINE_ISNAN_DBL) && __has_builtin(__builtin_isnan)
-        #define CV_INLINE_ISNAN_DBL(value) return __builtin_isnan(value);
-      #endif
-      #if !defined(CV_INLINE_ISNAN_FLT) && __has_builtin(__builtin_isnan)
-        #define CV_INLINE_ISNAN_FLT(value) return __builtin_isnan(value);
-      #endif
-      #if !defined(CV_INLINE_ISINF_DBL) && __has_builtin(__builtin_isinf)
-        #define CV_INLINE_ISINF_DBL(value) return __builtin_isinf(value);
-      #endif
-      #if !defined(CV_INLINE_ISINF_FLT) && __has_builtin(__builtin_isinf)
-        #define CV_INLINE_ISINF_FLT(value) return __builtin_isinf(value);
-      #endif
-    #elif defined(__GNUC__)
-      #define CV__FASTMATH_ENABLE_GCC_MATH_BUILTINS
-      #if !defined(CV_INLINE_ISNAN_DBL)
-        #define CV_INLINE_ISNAN_DBL(value) return __builtin_isnan(value);
-      #endif
-      #if !defined(CV_INLINE_ISNAN_FLT)
-        #define CV_INLINE_ISNAN_FLT(value) return __builtin_isnanf(value);
-      #endif
-      #if !defined(CV_INLINE_ISINF_DBL)
-        #define CV_INLINE_ISINF_DBL(value) return __builtin_isinf(value);
-      #endif
-      #if !defined(CV_INLINE_ISINF_FLT)
-        #define CV_INLINE_ISINF_FLT(value) return __builtin_isinff(value);
-      #endif
-    #elif defined(_MSC_VER)
-      #if !defined(CV_INLINE_ISNAN_DBL)
-        #define CV_INLINE_ISNAN_DBL(value) return isnan(value);
-      #endif
-      #if !defined(CV_INLINE_ISNAN_FLT)
-        #define CV_INLINE_ISNAN_FLT(value) return isnan(value);
-      #endif
-      #if !defined(CV_INLINE_ISINF_DBL)
-        #define CV_INLINE_ISINF_DBL(value) return isinf(value);
-      #endif
-      #if !defined(CV_INLINE_ISINF_FLT)
-        #define CV_INLINE_ISINF_FLT(value) return isinf(value);
-      #endif
-    #endif
-  #endif
-
-#endif // defined(__CUDACC__)
-
-/** @brief Rounds floating-point number to the nearest integer
-
- @param value floating-point number. If the value is outside of INT_MIN ... INT_MAX range, the
- result is not defined.
- */
-CV_INLINE int
-cvRound( double value )
-{
-#if defined CV_INLINE_ROUND_DBL
-    CV_INLINE_ROUND_DBL(value);
-#elif ((defined _MSC_VER && defined _M_X64) || (defined __GNUC__ && defined __x86_64__ \
-    && defined __SSE2__ && !defined __APPLE__) || CV_SSE2) \
-    && !defined(__CUDACC__)
-    __m128d t = _mm_set_sd( value );
-    return _mm_cvtsd_si32(t);
-#elif defined _MSC_VER && defined _M_IX86
-    int t;
-    __asm
-    {
-        fld value;
-        fistp t;
-    }
-    return t;
-#elif defined CV_ICC || defined __GNUC__
-    return (int)(lrint(value));
-#else
-    /* it's ok if round does not comply with IEEE754 standard;
-       the tests should allow +/-1 difference when the tested functions use round */
-    return (int)(value + (value >= 0 ? 0.5 : -0.5));
-#endif
-}
-
-
-/** @brief Rounds floating-point number to the nearest integer not larger than the original.
-
- The function computes an integer i such that:
- \f[i \le \texttt{value} < i+1\f]
- @param value floating-point number. If the value is outside of INT_MIN ... INT_MAX range, the
- result is not defined.
- */
-CV_INLINE int cvFloor( double value )
-{
-#if (defined CV__FASTMATH_ENABLE_GCC_MATH_BUILTINS || defined CV__FASTMATH_ENABLE_CLANG_MATH_BUILTINS) \
-    && ( \
-        defined(__PPC64__) \
-    )
-    return __builtin_floor(value);
-#else
-    int i = (int)value;
-    return i - (i > value);
-#endif
-}
-
-/** @brief Rounds floating-point number to the nearest integer not smaller than the original.
-
- The function computes an integer i such that:
- \f[i \le \texttt{value} < i+1\f]
- @param value floating-point number. If the value is outside of INT_MIN ... INT_MAX range, the
- result is not defined.
- */
-CV_INLINE int cvCeil( double value )
-{
-#if (defined CV__FASTMATH_ENABLE_GCC_MATH_BUILTINS || defined CV__FASTMATH_ENABLE_CLANG_MATH_BUILTINS) \
-    && ( \
-        defined(__PPC64__) \
-    )
-    return __builtin_ceil(value);
-#else
-    int i = (int)value;
-    return i + (i < value);
-#endif
-}
-
-/** @brief Determines if the argument is Not A Number.
-
- @param value The input floating-point value
-
- The function returns 1 if the argument is Not A Number (as defined by IEEE754 standard), 0
- otherwise. */
-CV_INLINE int cvIsNaN( double value )
-{
-#if defined CV_INLINE_ISNAN_DBL
-    CV_INLINE_ISNAN_DBL(value);
-#else
-    Cv64suf ieee754;
-    ieee754.f = value;
-    return ((unsigned)(ieee754.u >> 32) & 0x7fffffff) +
-           ((unsigned)ieee754.u != 0) > 0x7ff00000;
-#endif
-}
-
-/** @brief Determines if the argument is Infinity.
-
- @param value The input floating-point value
-
- The function returns 1 if the argument is a plus or minus infinity (as defined by IEEE754 standard)
- and 0 otherwise. */
-CV_INLINE int cvIsInf( double value )
-{
-#if defined CV_INLINE_ISINF_DBL
-    CV_INLINE_ISINF_DBL(value);
-#elif defined(__x86_64__) || defined(_M_X64) || defined(__aarch64__) || defined(_M_ARM64) || defined(__PPC64__)
-    Cv64suf ieee754;
-    ieee754.f = value;
-    return (ieee754.u & 0x7fffffff00000000) ==
-                        0x7ff0000000000000;
-#else
-    Cv64suf ieee754;
-    ieee754.f = value;
-    return ((unsigned)(ieee754.u >> 32) & 0x7fffffff) == 0x7ff00000 &&
-            (unsigned)ieee754.u == 0;
-#endif
-}
-
-#ifdef __cplusplus
-
-/** @overload */
-CV_INLINE int cvRound(float value)
-{
-#if defined CV_INLINE_ROUND_FLT
-    CV_INLINE_ROUND_FLT(value);
-#elif ((defined _MSC_VER && defined _M_X64) || (defined __GNUC__ && defined __x86_64__ \
-    && defined __SSE2__ && !defined __APPLE__) || CV_SSE2) \
-    && !defined(__CUDACC__)
-    __m128 t = _mm_set_ss( value );
-    return _mm_cvtss_si32(t);
-#elif defined _MSC_VER && defined _M_IX86
-    int t;
-    __asm
-    {
-        fld value;
-        fistp t;
-    }
-    return t;
-#elif defined CV_ICC || defined __GNUC__
-    return (int)(lrintf(value));
-#else
-    /* it's ok if round does not comply with IEEE754 standard;
-     the tests should allow +/-1 difference when the tested functions use round */
-    return (int)(value + (value >= 0 ? 0.5f : -0.5f));
-#endif
-}
-
-/** @overload */
-CV_INLINE int cvRound( int value )
-{
-    return value;
-}
-
-/** @overload */
-CV_INLINE int cvFloor( float value )
-{
-#if (defined CV__FASTMATH_ENABLE_GCC_MATH_BUILTINS || defined CV__FASTMATH_ENABLE_CLANG_MATH_BUILTINS) \
-    && ( \
-        defined(__PPC64__) \
-    )
-    return __builtin_floorf(value);
-#else
-    int i = (int)value;
-    return i - (i > value);
-#endif
-}
-
-/** @overload */
-CV_INLINE int cvFloor( int value )
-{
-    return value;
-}
-
-/** @overload */
-CV_INLINE int cvCeil( float value )
-{
-#if (defined CV__FASTMATH_ENABLE_GCC_MATH_BUILTINS || defined CV__FASTMATH_ENABLE_CLANG_MATH_BUILTINS) \
-    && ( \
-        defined(__PPC64__) \
-    )
-    return __builtin_ceilf(value);
-#else
-    int i = (int)value;
-    return i + (i < value);
-#endif
-}
-
-/** @overload */
-CV_INLINE int cvCeil( int value )
-{
-    return value;
-}
-
-/** @overload */
-CV_INLINE int cvIsNaN( float value )
-{
-#if defined CV_INLINE_ISNAN_FLT
-    CV_INLINE_ISNAN_FLT(value);
-#else
-    Cv32suf ieee754;
-    ieee754.f = value;
-    return (ieee754.u & 0x7fffffff) > 0x7f800000;
-#endif
-}
-
-/** @overload */
-CV_INLINE int cvIsInf( float value )
-{
-#if defined CV_INLINE_ISINF_FLT
-    CV_INLINE_ISINF_FLT(value);
-#else
-    Cv32suf ieee754;
-    ieee754.f = value;
-    return (ieee754.u & 0x7fffffff) == 0x7f800000;
-#endif
-}
-
-#endif // __cplusplus
-
-//! @} core_utils
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/hal.hpp b/projects/llm_framework/include/opencv4/opencv2/core/hal/hal.hpp
deleted file mode 100644
index 0d68078d..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/hal.hpp
+++ /dev/null
@@ -1,256 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Copyright (C) 2015, Itseez Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_HAL_HPP
-#define OPENCV_HAL_HPP
-
-#include "opencv2/core/cvdef.h"
-#include "opencv2/core/cvstd.hpp"
-#include "opencv2/core/hal/interface.h"
-
-namespace cv { namespace hal {
-
-//! @addtogroup core_hal_functions
-//! @{
-
-CV_EXPORTS int normHamming(const uchar* a, int n);
-CV_EXPORTS int normHamming(const uchar* a, const uchar* b, int n);
-
-CV_EXPORTS int normHamming(const uchar* a, int n, int cellSize);
-CV_EXPORTS int normHamming(const uchar* a, const uchar* b, int n, int cellSize);
-
-CV_EXPORTS int LU32f(float* A, size_t astep, int m, float* b, size_t bstep, int n);
-CV_EXPORTS int LU64f(double* A, size_t astep, int m, double* b, size_t bstep, int n);
-CV_EXPORTS bool Cholesky32f(float* A, size_t astep, int m, float* b, size_t bstep, int n);
-CV_EXPORTS bool Cholesky64f(double* A, size_t astep, int m, double* b, size_t bstep, int n);
-CV_EXPORTS void SVD32f(float* At, size_t astep, float* W, float* U, size_t ustep, float* Vt, size_t vstep, int m, int n, int flags);
-CV_EXPORTS void SVD64f(double* At, size_t astep, double* W, double* U, size_t ustep, double* Vt, size_t vstep, int m, int n, int flags);
-CV_EXPORTS int QR32f(float* A, size_t astep, int m, int n, int k, float* b, size_t bstep, float* hFactors);
-CV_EXPORTS int QR64f(double* A, size_t astep, int m, int n, int k, double* b, size_t bstep, double* hFactors);
-
-CV_EXPORTS void gemm32f(const float* src1, size_t src1_step, const float* src2, size_t src2_step,
-                        float alpha, const float* src3, size_t src3_step, float beta, float* dst, size_t dst_step,
-                        int m_a, int n_a, int n_d, int flags);
-CV_EXPORTS void gemm64f(const double* src1, size_t src1_step, const double* src2, size_t src2_step,
-                        double alpha, const double* src3, size_t src3_step, double beta, double* dst, size_t dst_step,
-                        int m_a, int n_a, int n_d, int flags);
-CV_EXPORTS void gemm32fc(const float* src1, size_t src1_step, const float* src2, size_t src2_step,
-                        float alpha, const float* src3, size_t src3_step, float beta, float* dst, size_t dst_step,
-                        int m_a, int n_a, int n_d, int flags);
-CV_EXPORTS void gemm64fc(const double* src1, size_t src1_step, const double* src2, size_t src2_step,
-                        double alpha, const double* src3, size_t src3_step, double beta, double* dst, size_t dst_step,
-                        int m_a, int n_a, int n_d, int flags);
-
-CV_EXPORTS int normL1_(const uchar* a, const uchar* b, int n);
-CV_EXPORTS float normL1_(const float* a, const float* b, int n);
-CV_EXPORTS float normL2Sqr_(const float* a, const float* b, int n);
-
-CV_EXPORTS void exp32f(const float* src, float* dst, int n);
-CV_EXPORTS void exp64f(const double* src, double* dst, int n);
-CV_EXPORTS void log32f(const float* src, float* dst, int n);
-CV_EXPORTS void log64f(const double* src, double* dst, int n);
-
-CV_EXPORTS void fastAtan32f(const float* y, const float* x, float* dst, int n, bool angleInDegrees);
-CV_EXPORTS void fastAtan64f(const double* y, const double* x, double* dst, int n, bool angleInDegrees);
-CV_EXPORTS void magnitude32f(const float* x, const float* y, float* dst, int n);
-CV_EXPORTS void magnitude64f(const double* x, const double* y, double* dst, int n);
-CV_EXPORTS void sqrt32f(const float* src, float* dst, int len);
-CV_EXPORTS void sqrt64f(const double* src, double* dst, int len);
-CV_EXPORTS void invSqrt32f(const float* src, float* dst, int len);
-CV_EXPORTS void invSqrt64f(const double* src, double* dst, int len);
-
-CV_EXPORTS void split8u(const uchar* src, uchar** dst, int len, int cn );
-CV_EXPORTS void split16u(const ushort* src, ushort** dst, int len, int cn );
-CV_EXPORTS void split32s(const int* src, int** dst, int len, int cn );
-CV_EXPORTS void split64s(const int64* src, int64** dst, int len, int cn );
-
-CV_EXPORTS void merge8u(const uchar** src, uchar* dst, int len, int cn );
-CV_EXPORTS void merge16u(const ushort** src, ushort* dst, int len, int cn );
-CV_EXPORTS void merge32s(const int** src, int* dst, int len, int cn );
-CV_EXPORTS void merge64s(const int64** src, int64* dst, int len, int cn );
-
-CV_EXPORTS void add8u( const uchar* src1, size_t step1, const uchar* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void add8s( const schar* src1, size_t step1, const schar* src2, size_t step2, schar* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void add16u( const ushort* src1, size_t step1, const ushort* src2, size_t step2, ushort* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void add16s( const short* src1, size_t step1, const short* src2, size_t step2, short* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void add32s( const int* src1, size_t step1, const int* src2, size_t step2, int* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void add32f( const float* src1, size_t step1, const float* src2, size_t step2, float* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void add64f( const double* src1, size_t step1, const double* src2, size_t step2, double* dst, size_t step, int width, int height, void* );
-
-CV_EXPORTS void sub8u( const uchar* src1, size_t step1, const uchar* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void sub8s( const schar* src1, size_t step1, const schar* src2, size_t step2, schar* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void sub16u( const ushort* src1, size_t step1, const ushort* src2, size_t step2, ushort* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void sub16s( const short* src1, size_t step1, const short* src2, size_t step2, short* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void sub32s( const int* src1, size_t step1, const int* src2, size_t step2, int* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void sub32f( const float* src1, size_t step1, const float* src2, size_t step2, float* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void sub64f( const double* src1, size_t step1, const double* src2, size_t step2, double* dst, size_t step, int width, int height, void* );
-
-CV_EXPORTS void max8u( const uchar* src1, size_t step1, const uchar* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void max8s( const schar* src1, size_t step1, const schar* src2, size_t step2, schar* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void max16u( const ushort* src1, size_t step1, const ushort* src2, size_t step2, ushort* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void max16s( const short* src1, size_t step1, const short* src2, size_t step2, short* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void max32s( const int* src1, size_t step1, const int* src2, size_t step2, int* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void max32f( const float* src1, size_t step1, const float* src2, size_t step2, float* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void max64f( const double* src1, size_t step1, const double* src2, size_t step2, double* dst, size_t step, int width, int height, void* );
-
-CV_EXPORTS void min8u( const uchar* src1, size_t step1, const uchar* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void min8s( const schar* src1, size_t step1, const schar* src2, size_t step2, schar* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void min16u( const ushort* src1, size_t step1, const ushort* src2, size_t step2, ushort* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void min16s( const short* src1, size_t step1, const short* src2, size_t step2, short* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void min32s( const int* src1, size_t step1, const int* src2, size_t step2, int* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void min32f( const float* src1, size_t step1, const float* src2, size_t step2, float* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void min64f( const double* src1, size_t step1, const double* src2, size_t step2, double* dst, size_t step, int width, int height, void* );
-
-CV_EXPORTS void absdiff8u( const uchar* src1, size_t step1, const uchar* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void absdiff8s( const schar* src1, size_t step1, const schar* src2, size_t step2, schar* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void absdiff16u( const ushort* src1, size_t step1, const ushort* src2, size_t step2, ushort* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void absdiff16s( const short* src1, size_t step1, const short* src2, size_t step2, short* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void absdiff32s( const int* src1, size_t step1, const int* src2, size_t step2, int* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void absdiff32f( const float* src1, size_t step1, const float* src2, size_t step2, float* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void absdiff64f( const double* src1, size_t step1, const double* src2, size_t step2, double* dst, size_t step, int width, int height, void* );
-
-CV_EXPORTS void and8u( const uchar* src1, size_t step1, const uchar* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void or8u( const uchar* src1, size_t step1, const uchar* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void xor8u( const uchar* src1, size_t step1, const uchar* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* );
-CV_EXPORTS void not8u( const uchar* src1, size_t step1, const uchar* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* );
-
-CV_EXPORTS void cmp8u(const uchar* src1, size_t step1, const uchar* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* _cmpop);
-CV_EXPORTS void cmp8s(const schar* src1, size_t step1, const schar* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* _cmpop);
-CV_EXPORTS void cmp16u(const ushort* src1, size_t step1, const ushort* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* _cmpop);
-CV_EXPORTS void cmp16s(const short* src1, size_t step1, const short* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* _cmpop);
-CV_EXPORTS void cmp32s(const int* src1, size_t step1, const int* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* _cmpop);
-CV_EXPORTS void cmp32f(const float* src1, size_t step1, const float* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* _cmpop);
-CV_EXPORTS void cmp64f(const double* src1, size_t step1, const double* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* _cmpop);
-
-CV_EXPORTS void mul8u( const uchar* src1, size_t step1, const uchar* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void mul8s( const schar* src1, size_t step1, const schar* src2, size_t step2, schar* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void mul16u( const ushort* src1, size_t step1, const ushort* src2, size_t step2, ushort* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void mul16s( const short* src1, size_t step1, const short* src2, size_t step2, short* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void mul32s( const int* src1, size_t step1, const int* src2, size_t step2, int* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void mul32f( const float* src1, size_t step1, const float* src2, size_t step2, float* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void mul64f( const double* src1, size_t step1, const double* src2, size_t step2, double* dst, size_t step, int width, int height, void* scale);
-
-CV_EXPORTS void div8u( const uchar* src1, size_t step1, const uchar* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void div8s( const schar* src1, size_t step1, const schar* src2, size_t step2, schar* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void div16u( const ushort* src1, size_t step1, const ushort* src2, size_t step2, ushort* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void div16s( const short* src1, size_t step1, const short* src2, size_t step2, short* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void div32s( const int* src1, size_t step1, const int* src2, size_t step2, int* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void div32f( const float* src1, size_t step1, const float* src2, size_t step2, float* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void div64f( const double* src1, size_t step1, const double* src2, size_t step2, double* dst, size_t step, int width, int height, void* scale);
-
-CV_EXPORTS void recip8u( const uchar *, size_t, const uchar * src2, size_t step2, uchar* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void recip8s( const schar *, size_t, const schar * src2, size_t step2, schar* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void recip16u( const ushort *, size_t, const ushort * src2, size_t step2, ushort* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void recip16s( const short *, size_t, const short * src2, size_t step2, short* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void recip32s( const int *, size_t, const int * src2, size_t step2, int* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void recip32f( const float *, size_t, const float * src2, size_t step2, float* dst, size_t step, int width, int height, void* scale);
-CV_EXPORTS void recip64f( const double *, size_t, const double * src2, size_t step2, double* dst, size_t step, int width, int height, void* scale);
-
-CV_EXPORTS void addWeighted8u( const uchar* src1, size_t step1, const uchar* src2, size_t step2, uchar* dst, size_t step, int width, int height, void* _scalars );
-CV_EXPORTS void addWeighted8s( const schar* src1, size_t step1, const schar* src2, size_t step2, schar* dst, size_t step, int width, int height, void* scalars );
-CV_EXPORTS void addWeighted16u( const ushort* src1, size_t step1, const ushort* src2, size_t step2, ushort* dst, size_t step, int width, int height, void* scalars );
-CV_EXPORTS void addWeighted16s( const short* src1, size_t step1, const short* src2, size_t step2, short* dst, size_t step, int width, int height, void* scalars );
-CV_EXPORTS void addWeighted32s( const int* src1, size_t step1, const int* src2, size_t step2, int* dst, size_t step, int width, int height, void* scalars );
-CV_EXPORTS void addWeighted32f( const float* src1, size_t step1, const float* src2, size_t step2, float* dst, size_t step, int width, int height, void* scalars );
-CV_EXPORTS void addWeighted64f( const double* src1, size_t step1, const double* src2, size_t step2, double* dst, size_t step, int width, int height, void* scalars );
-
-CV_EXPORTS void cvt16f32f( const float16_t* src, float* dst, int len );
-CV_EXPORTS void cvt32f16f( const float* src, float16_t* dst, int len );
-
-CV_EXPORTS void addRNGBias32f( float* arr, const float* scaleBiasPairs, int len );
-CV_EXPORTS void addRNGBias64f( double* arr, const double* scaleBiasPairs, int len );
-
-struct CV_EXPORTS DFT1D
-{
-    static Ptr<DFT1D> create(int len, int count, int depth, int flags, bool * useBuffer = 0);
-    virtual void apply(const uchar *src, uchar *dst) = 0;
-    virtual ~DFT1D() {}
-};
-
-struct CV_EXPORTS DFT2D
-{
-    static Ptr<DFT2D> create(int width, int height, int depth,
-                             int src_channels, int dst_channels,
-                             int flags, int nonzero_rows = 0);
-    virtual void apply(const uchar *src_data, size_t src_step, uchar *dst_data, size_t dst_step) = 0;
-    virtual ~DFT2D() {}
-};
-
-struct CV_EXPORTS DCT2D
-{
-    static Ptr<DCT2D> create(int width, int height, int depth, int flags);
-    virtual void apply(const uchar *src_data, size_t src_step, uchar *dst_data, size_t dst_step) = 0;
-    virtual ~DCT2D() {}
-};
-
-//! @} core_hal
-
-//=============================================================================
-// for binary compatibility with 3.0
-
-//! @cond IGNORED
-
-CV_EXPORTS int LU(float* A, size_t astep, int m, float* b, size_t bstep, int n);
-CV_EXPORTS int LU(double* A, size_t astep, int m, double* b, size_t bstep, int n);
-CV_EXPORTS bool Cholesky(float* A, size_t astep, int m, float* b, size_t bstep, int n);
-CV_EXPORTS bool Cholesky(double* A, size_t astep, int m, double* b, size_t bstep, int n);
-
-CV_EXPORTS void exp(const float* src, float* dst, int n);
-CV_EXPORTS void exp(const double* src, double* dst, int n);
-CV_EXPORTS void log(const float* src, float* dst, int n);
-CV_EXPORTS void log(const double* src, double* dst, int n);
-
-CV_EXPORTS void fastAtan2(const float* y, const float* x, float* dst, int n, bool angleInDegrees);
-CV_EXPORTS void magnitude(const float* x, const float* y, float* dst, int n);
-CV_EXPORTS void magnitude(const double* x, const double* y, double* dst, int n);
-CV_EXPORTS void sqrt(const float* src, float* dst, int len);
-CV_EXPORTS void sqrt(const double* src, double* dst, int len);
-CV_EXPORTS void invSqrt(const float* src, float* dst, int len);
-CV_EXPORTS void invSqrt(const double* src, double* dst, int len);
-
-//! @endcond
-
-}} //cv::hal
-
-#endif //OPENCV_HAL_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/interface.h b/projects/llm_framework/include/opencv4/opencv2/core/hal/interface.h
deleted file mode 100644
index 6f0a83d3..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/interface.h
+++ /dev/null
@@ -1,190 +0,0 @@
-#ifndef OPENCV_CORE_HAL_INTERFACE_H
-#define OPENCV_CORE_HAL_INTERFACE_H
-
-//! @addtogroup core_hal_interface
-//! @{
-
-//! @name Return codes
-//! @{
-#define CV_HAL_ERROR_OK 0
-#define CV_HAL_ERROR_NOT_IMPLEMENTED 1
-#define CV_HAL_ERROR_UNKNOWN -1
-//! @}
-
-#ifdef __cplusplus
-#include <cstddef>
-#else
-#include <stddef.h>
-#include <stdbool.h>
-#endif
-
-//! @name Data types
-//! primitive types
-//! - schar  - signed 1 byte integer
-//! - uchar  - unsigned 1 byte integer
-//! - short  - signed 2 byte integer
-//! - ushort - unsigned 2 byte integer
-//! - int    - signed 4 byte integer
-//! - uint   - unsigned 4 byte integer
-//! - int64  - signed 8 byte integer
-//! - uint64 - unsigned 8 byte integer
-//! @{
-#if !defined _MSC_VER && !defined __BORLANDC__
-#  if defined __cplusplus && __cplusplus >= 201103L && !defined __APPLE__
-#    include <cstdint>
-#    ifdef __NEWLIB__
-        typedef unsigned int uint;
-#    else
-        typedef std::uint32_t uint;
-#    endif
-#  else
-#    include <stdint.h>
-     typedef uint32_t uint;
-#  endif
-#else
-   typedef unsigned uint;
-#endif
-
-typedef signed char schar;
-
-#ifndef __IPL_H__
-   typedef unsigned char uchar;
-   typedef unsigned short ushort;
-#endif
-
-#if defined _MSC_VER || defined __BORLANDC__
-   typedef __int64 int64;
-   typedef unsigned __int64 uint64;
-#  define CV_BIG_INT(n)   n##I64
-#  define CV_BIG_UINT(n)  n##UI64
-#else
-   typedef int64_t int64;
-   typedef uint64_t uint64;
-#  define CV_BIG_INT(n)   n##LL
-#  define CV_BIG_UINT(n)  n##ULL
-#endif
-
-#define CV_USRTYPE1 (void)"CV_USRTYPE1 support has been dropped in OpenCV 4.0"
-
-#define CV_CN_MAX     512
-#define CV_CN_SHIFT   3
-#define CV_DEPTH_MAX  (1 << CV_CN_SHIFT)
-
-#define CV_8U   0
-#define CV_8S   1
-#define CV_16U  2
-#define CV_16S  3
-#define CV_32S  4
-#define CV_32F  5
-#define CV_64F  6
-#define CV_16F  7
-
-#define CV_MAT_DEPTH_MASK       (CV_DEPTH_MAX - 1)
-#define CV_MAT_DEPTH(flags)     ((flags) & CV_MAT_DEPTH_MASK)
-
-#define CV_MAKETYPE(depth,cn) (CV_MAT_DEPTH(depth) + (((cn)-1) << CV_CN_SHIFT))
-#define CV_MAKE_TYPE CV_MAKETYPE
-
-#define CV_8UC1 CV_MAKETYPE(CV_8U,1)
-#define CV_8UC2 CV_MAKETYPE(CV_8U,2)
-#define CV_8UC3 CV_MAKETYPE(CV_8U,3)
-#define CV_8UC4 CV_MAKETYPE(CV_8U,4)
-#define CV_8UC(n) CV_MAKETYPE(CV_8U,(n))
-
-#define CV_8SC1 CV_MAKETYPE(CV_8S,1)
-#define CV_8SC2 CV_MAKETYPE(CV_8S,2)
-#define CV_8SC3 CV_MAKETYPE(CV_8S,3)
-#define CV_8SC4 CV_MAKETYPE(CV_8S,4)
-#define CV_8SC(n) CV_MAKETYPE(CV_8S,(n))
-
-#define CV_16UC1 CV_MAKETYPE(CV_16U,1)
-#define CV_16UC2 CV_MAKETYPE(CV_16U,2)
-#define CV_16UC3 CV_MAKETYPE(CV_16U,3)
-#define CV_16UC4 CV_MAKETYPE(CV_16U,4)
-#define CV_16UC(n) CV_MAKETYPE(CV_16U,(n))
-
-#define CV_16SC1 CV_MAKETYPE(CV_16S,1)
-#define CV_16SC2 CV_MAKETYPE(CV_16S,2)
-#define CV_16SC3 CV_MAKETYPE(CV_16S,3)
-#define CV_16SC4 CV_MAKETYPE(CV_16S,4)
-#define CV_16SC(n) CV_MAKETYPE(CV_16S,(n))
-
-#define CV_32SC1 CV_MAKETYPE(CV_32S,1)
-#define CV_32SC2 CV_MAKETYPE(CV_32S,2)
-#define CV_32SC3 CV_MAKETYPE(CV_32S,3)
-#define CV_32SC4 CV_MAKETYPE(CV_32S,4)
-#define CV_32SC(n) CV_MAKETYPE(CV_32S,(n))
-
-#define CV_32FC1 CV_MAKETYPE(CV_32F,1)
-#define CV_32FC2 CV_MAKETYPE(CV_32F,2)
-#define CV_32FC3 CV_MAKETYPE(CV_32F,3)
-#define CV_32FC4 CV_MAKETYPE(CV_32F,4)
-#define CV_32FC(n) CV_MAKETYPE(CV_32F,(n))
-
-#define CV_64FC1 CV_MAKETYPE(CV_64F,1)
-#define CV_64FC2 CV_MAKETYPE(CV_64F,2)
-#define CV_64FC3 CV_MAKETYPE(CV_64F,3)
-#define CV_64FC4 CV_MAKETYPE(CV_64F,4)
-#define CV_64FC(n) CV_MAKETYPE(CV_64F,(n))
-
-#define CV_16FC1 CV_MAKETYPE(CV_16F,1)
-#define CV_16FC2 CV_MAKETYPE(CV_16F,2)
-#define CV_16FC3 CV_MAKETYPE(CV_16F,3)
-#define CV_16FC4 CV_MAKETYPE(CV_16F,4)
-#define CV_16FC(n) CV_MAKETYPE(CV_16F,(n))
-//! @}
-
-//! @name Comparison operation
-//! @sa cv::CmpTypes
-//! @{
-#define CV_HAL_CMP_EQ 0
-#define CV_HAL_CMP_GT 1
-#define CV_HAL_CMP_GE 2
-#define CV_HAL_CMP_LT 3
-#define CV_HAL_CMP_LE 4
-#define CV_HAL_CMP_NE 5
-//! @}
-
-//! @name Border processing modes
-//! @sa cv::BorderTypes
-//! @{
-#define CV_HAL_BORDER_CONSTANT 0
-#define CV_HAL_BORDER_REPLICATE 1
-#define CV_HAL_BORDER_REFLECT 2
-#define CV_HAL_BORDER_WRAP 3
-#define CV_HAL_BORDER_REFLECT_101 4
-#define CV_HAL_BORDER_TRANSPARENT 5
-#define CV_HAL_BORDER_ISOLATED 16
-//! @}
-
-//! @name DFT flags
-//! @{
-#define CV_HAL_DFT_INVERSE        1
-#define CV_HAL_DFT_SCALE          2
-#define CV_HAL_DFT_ROWS           4
-#define CV_HAL_DFT_COMPLEX_OUTPUT 16
-#define CV_HAL_DFT_REAL_OUTPUT    32
-#define CV_HAL_DFT_TWO_STAGE      64
-#define CV_HAL_DFT_STAGE_COLS    128
-#define CV_HAL_DFT_IS_CONTINUOUS 512
-#define CV_HAL_DFT_IS_INPLACE 1024
-//! @}
-
-//! @name SVD flags
-//! @{
-#define CV_HAL_SVD_NO_UV    1
-#define CV_HAL_SVD_SHORT_UV 2
-#define CV_HAL_SVD_MODIFY_A 4
-#define CV_HAL_SVD_FULL_UV  8
-//! @}
-
-//! @name Gemm flags
-//! @{
-#define CV_HAL_GEMM_1_T 1
-#define CV_HAL_GEMM_2_T 2
-#define CV_HAL_GEMM_3_T 4
-//! @}
-
-//! @}
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin.hpp b/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin.hpp
deleted file mode 100644
index ac331f21..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin.hpp
+++ /dev/null
@@ -1,706 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Copyright (C) 2015, Itseez Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_HAL_INTRIN_HPP
-#define OPENCV_HAL_INTRIN_HPP
-
-#include <cmath>
-#include <float.h>
-#include <stdlib.h>
-#include "opencv2/core/cvdef.h"
-
-#define OPENCV_HAL_ADD(a, b) ((a) + (b))
-#define OPENCV_HAL_AND(a, b) ((a) & (b))
-#define OPENCV_HAL_NOP(a) (a)
-#define OPENCV_HAL_1ST(a, b) (a)
-
-namespace {
-inline unsigned int trailingZeros32(unsigned int value) {
-#if defined(_MSC_VER)
-#if (_MSC_VER < 1700) || defined(_M_ARM) || defined(_M_ARM64)
-    unsigned long index = 0;
-    _BitScanForward(&index, value);
-    return (unsigned int)index;
-#elif defined(__clang__)
-    // clang-cl doesn't export _tzcnt_u32 for non BMI systems
-    return value ? __builtin_ctz(value) : 32;
-#else
-    return _tzcnt_u32(value);
-#endif
-#elif defined(__GNUC__) || defined(__GNUG__)
-    return __builtin_ctz(value);
-#elif defined(__ICC) || defined(__INTEL_COMPILER)
-    return _bit_scan_forward(value);
-#elif defined(__clang__)
-    return llvm.cttz.i32(value, true);
-#else
-    static const int MultiplyDeBruijnBitPosition[32] = {
-        0, 1, 28, 2, 29, 14, 24, 3, 30, 22, 20, 15, 25, 17, 4, 8,
-        31, 27, 13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9 };
-    return MultiplyDeBruijnBitPosition[((uint32_t)((value & -value) * 0x077CB531U)) >> 27];
-#endif
-}
-}
-
-// unlike HAL API, which is in cv::hal,
-// we put intrinsics into cv namespace to make its
-// access from within opencv code more accessible
-namespace cv {
-
-namespace hal {
-
-enum StoreMode
-{
-    STORE_UNALIGNED = 0,
-    STORE_ALIGNED = 1,
-    STORE_ALIGNED_NOCACHE = 2
-};
-
-}
-
-// TODO FIXIT: Don't use "God" traits. Split on separate cases.
-template<typename _Tp> struct V_TypeTraits
-{
-};
-
-#define CV_INTRIN_DEF_TYPE_TRAITS(type, int_type_, uint_type_, abs_type_, w_type_, q_type_, sum_type_) \
-    template<> struct V_TypeTraits<type> \
-    { \
-        typedef type value_type; \
-        typedef int_type_ int_type; \
-        typedef abs_type_ abs_type; \
-        typedef uint_type_ uint_type; \
-        typedef w_type_ w_type; \
-        typedef q_type_ q_type; \
-        typedef sum_type_ sum_type; \
-    \
-        static inline int_type reinterpret_int(type x) \
-        { \
-            union { type l; int_type i; } v; \
-            v.l = x; \
-            return v.i; \
-        } \
-    \
-        static inline type reinterpret_from_int(int_type x) \
-        { \
-            union { type l; int_type i; } v; \
-            v.i = x; \
-            return v.l; \
-        } \
-    }
-
-#define CV_INTRIN_DEF_TYPE_TRAITS_NO_Q_TYPE(type, int_type_, uint_type_, abs_type_, w_type_, sum_type_) \
-    template<> struct V_TypeTraits<type> \
-    { \
-        typedef type value_type; \
-        typedef int_type_ int_type; \
-        typedef abs_type_ abs_type; \
-        typedef uint_type_ uint_type; \
-        typedef w_type_ w_type; \
-        typedef sum_type_ sum_type; \
-    \
-        static inline int_type reinterpret_int(type x) \
-        { \
-            union { type l; int_type i; } v; \
-            v.l = x; \
-            return v.i; \
-        } \
-    \
-        static inline type reinterpret_from_int(int_type x) \
-        { \
-            union { type l; int_type i; } v; \
-            v.i = x; \
-            return v.l; \
-        } \
-    }
-
-CV_INTRIN_DEF_TYPE_TRAITS(uchar, schar, uchar, uchar, ushort, unsigned, unsigned);
-CV_INTRIN_DEF_TYPE_TRAITS(schar, schar, uchar, uchar, short, int, int);
-CV_INTRIN_DEF_TYPE_TRAITS(ushort, short, ushort, ushort, unsigned, uint64, unsigned);
-CV_INTRIN_DEF_TYPE_TRAITS(short, short, ushort, ushort, int, int64, int);
-CV_INTRIN_DEF_TYPE_TRAITS_NO_Q_TYPE(unsigned, int, unsigned, unsigned, uint64, unsigned);
-CV_INTRIN_DEF_TYPE_TRAITS_NO_Q_TYPE(int, int, unsigned, unsigned, int64, int);
-CV_INTRIN_DEF_TYPE_TRAITS_NO_Q_TYPE(float, int, unsigned, float, double, float);
-CV_INTRIN_DEF_TYPE_TRAITS_NO_Q_TYPE(uint64, int64, uint64, uint64, void, uint64);
-CV_INTRIN_DEF_TYPE_TRAITS_NO_Q_TYPE(int64, int64, uint64, uint64, void, int64);
-CV_INTRIN_DEF_TYPE_TRAITS_NO_Q_TYPE(double, int64, uint64, double, void, double);
-
-#ifndef CV_DOXYGEN
-
-#ifndef CV_CPU_OPTIMIZATION_HAL_NAMESPACE
-#ifdef CV_FORCE_SIMD128_CPP
-    #define CV_CPU_OPTIMIZATION_HAL_NAMESPACE hal_EMULATOR_CPP
-    #define CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN namespace hal_EMULATOR_CPP {
-    #define CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END }
-#elif defined(CV_CPU_DISPATCH_MODE)
-    #define CV_CPU_OPTIMIZATION_HAL_NAMESPACE __CV_CAT(hal_, CV_CPU_DISPATCH_MODE)
-    #define CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN namespace __CV_CAT(hal_, CV_CPU_DISPATCH_MODE) {
-    #define CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END }
-#else
-    #define CV_CPU_OPTIMIZATION_HAL_NAMESPACE hal_baseline
-    #define CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN namespace hal_baseline {
-    #define CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END }
-#endif
-#endif // CV_CPU_OPTIMIZATION_HAL_NAMESPACE
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END
-using namespace CV_CPU_OPTIMIZATION_HAL_NAMESPACE;
-#endif
-}
-
-#ifdef CV_DOXYGEN
-#   undef CV_AVX2
-#   undef CV_SSE2
-#   undef CV_NEON
-#   undef CV_VSX
-#   undef CV_FP16
-#   undef CV_MSA
-#   undef CV_RVV
-#endif
-
-#if (CV_SSE2 || CV_NEON || CV_VSX || CV_MSA || CV_WASM_SIMD || CV_RVV071 || CV_RVV) && !defined(CV_FORCE_SIMD128_CPP)
-#define CV__SIMD_FORWARD 128
-#include "opencv2/core/hal/intrin_forward.hpp"
-#endif
-
-#if CV_SSE2 && !defined(CV_FORCE_SIMD128_CPP)
-
-#include "opencv2/core/hal/intrin_sse_em.hpp"
-#include "opencv2/core/hal/intrin_sse.hpp"
-
-#elif CV_NEON && !defined(CV_FORCE_SIMD128_CPP)
-
-#include "opencv2/core/hal/intrin_neon.hpp"
-
-#elif CV_RVV071 && !defined(CV_FORCE_SIMD128_CPP)
-#define CV_SIMD128_CPP 0
-#include "opencv2/core/hal/intrin_rvv071.hpp"
-
-#elif CV_VSX && !defined(CV_FORCE_SIMD128_CPP)
-
-#include "opencv2/core/hal/intrin_vsx.hpp"
-
-#elif CV_MSA && !defined(CV_FORCE_SIMD128_CPP)
-
-#include "opencv2/core/hal/intrin_msa.hpp"
-
-#elif CV_WASM_SIMD && !defined(CV_FORCE_SIMD128_CPP)
-#include "opencv2/core/hal/intrin_wasm.hpp"
-
-#elif CV_RVV && !defined(CV_FORCE_SIMD128_CPP)
-#include "opencv2/core/hal/intrin_rvv.hpp"
-
-#else
-
-#include "opencv2/core/hal/intrin_cpp.hpp"
-
-#endif
-
-// AVX2 can be used together with SSE2, so
-// we define those two sets of intrinsics at once.
-// Most of the intrinsics do not conflict (the proper overloaded variant is
-// resolved by the argument types, e.g. v_float32x4 ~ SSE2, v_float32x8 ~ AVX2),
-// but some of AVX2 intrinsics get v256_ prefix instead of v_, e.g. v256_load() vs v_load().
-// Correspondingly, the wide intrinsics (which are mapped to the "widest"
-// available instruction set) will get vx_ prefix
-// (and will be mapped to v256_ counterparts) (e.g. vx_load() => v256_load())
-#if CV_AVX2
-
-#define CV__SIMD_FORWARD 256
-#include "opencv2/core/hal/intrin_forward.hpp"
-#include "opencv2/core/hal/intrin_avx.hpp"
-
-#endif
-
-// AVX512 can be used together with SSE2 and AVX2, so
-// we define those sets of intrinsics at once.
-// For some of AVX512 intrinsics get v512_ prefix instead of v_, e.g. v512_load() vs v_load().
-// Wide intrinsics will be mapped to v512_ counterparts in this case(e.g. vx_load() => v512_load())
-#if CV_AVX512_SKX
-
-#define CV__SIMD_FORWARD 512
-#include "opencv2/core/hal/intrin_forward.hpp"
-#include "opencv2/core/hal/intrin_avx512.hpp"
-
-#endif
-
-//! @cond IGNORED
-
-namespace cv {
-
-#ifndef CV_DOXYGEN
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
-#endif
-
-#ifndef CV_SIMD128
-#define CV_SIMD128 0
-#endif
-
-#ifndef CV_SIMD128_CPP
-#define CV_SIMD128_CPP 0
-#endif
-
-#ifndef CV_SIMD128_64F
-#define CV_SIMD128_64F 0
-#endif
-
-#ifndef CV_SIMD256
-#define CV_SIMD256 0
-#endif
-
-#ifndef CV_SIMD256_64F
-#define CV_SIMD256_64F 0
-#endif
-
-#ifndef CV_SIMD512
-#define CV_SIMD512 0
-#endif
-
-#ifndef CV_SIMD512_64F
-#define CV_SIMD512_64F 0
-#endif
-
-#ifndef CV_SIMD128_FP16
-#define CV_SIMD128_FP16 0
-#endif
-
-#ifndef CV_SIMD256_FP16
-#define CV_SIMD256_FP16 0
-#endif
-
-#ifndef CV_SIMD512_FP16
-#define CV_SIMD512_FP16 0
-#endif
-
-//==================================================================================================
-
-template<typename _Tp> struct V_RegTraits
-{
-};
-
-#define CV_DEF_REG_TRAITS(prefix, _reg, lane_type, suffix, _u_reg, _w_reg, _q_reg, _int_reg, _round_reg) \
-    template<> struct V_RegTraits<_reg> \
-    { \
-        typedef _reg reg; \
-        typedef _u_reg u_reg; \
-        typedef _w_reg w_reg; \
-        typedef _q_reg q_reg; \
-        typedef _int_reg int_reg; \
-        typedef _round_reg round_reg; \
-    }
-
-#if CV_SIMD128 || CV_SIMD128_CPP
-    CV_DEF_REG_TRAITS(v, v_uint8x16, uchar, u8, v_uint8x16, v_uint16x8, v_uint32x4, v_int8x16, void);
-    CV_DEF_REG_TRAITS(v, v_int8x16, schar, s8, v_uint8x16, v_int16x8, v_int32x4, v_int8x16, void);
-    CV_DEF_REG_TRAITS(v, v_uint16x8, ushort, u16, v_uint16x8, v_uint32x4, v_uint64x2, v_int16x8, void);
-    CV_DEF_REG_TRAITS(v, v_int16x8, short, s16, v_uint16x8, v_int32x4, v_int64x2, v_int16x8, void);
-    CV_DEF_REG_TRAITS(v, v_uint32x4, unsigned, u32, v_uint32x4, v_uint64x2, void, v_int32x4, void);
-    CV_DEF_REG_TRAITS(v, v_int32x4, int, s32, v_uint32x4, v_int64x2, void, v_int32x4, void);
-#if CV_SIMD128_64F || CV_SIMD128_CPP
-    CV_DEF_REG_TRAITS(v, v_float32x4, float, f32, v_float32x4, v_float64x2, void, v_int32x4, v_int32x4);
-#else
-    CV_DEF_REG_TRAITS(v, v_float32x4, float, f32, v_float32x4, void, void, v_int32x4, v_int32x4);
-#endif
-    CV_DEF_REG_TRAITS(v, v_uint64x2, uint64, u64, v_uint64x2, void, void, v_int64x2, void);
-    CV_DEF_REG_TRAITS(v, v_int64x2, int64, s64, v_uint64x2, void, void, v_int64x2, void);
-#if CV_SIMD128_64F
-    CV_DEF_REG_TRAITS(v, v_float64x2, double, f64, v_float64x2, void, void, v_int64x2, v_int32x4);
-#endif
-#endif
-
-#if CV_SIMD256
-    CV_DEF_REG_TRAITS(v256, v_uint8x32, uchar, u8, v_uint8x32, v_uint16x16, v_uint32x8, v_int8x32, void);
-    CV_DEF_REG_TRAITS(v256, v_int8x32, schar, s8, v_uint8x32, v_int16x16, v_int32x8, v_int8x32, void);
-    CV_DEF_REG_TRAITS(v256, v_uint16x16, ushort, u16, v_uint16x16, v_uint32x8, v_uint64x4, v_int16x16, void);
-    CV_DEF_REG_TRAITS(v256, v_int16x16, short, s16, v_uint16x16, v_int32x8, v_int64x4, v_int16x16, void);
-    CV_DEF_REG_TRAITS(v256, v_uint32x8, unsigned, u32, v_uint32x8, v_uint64x4, void, v_int32x8, void);
-    CV_DEF_REG_TRAITS(v256, v_int32x8, int, s32, v_uint32x8, v_int64x4, void, v_int32x8, void);
-    CV_DEF_REG_TRAITS(v256, v_float32x8, float, f32, v_float32x8, v_float64x4, void, v_int32x8, v_int32x8);
-    CV_DEF_REG_TRAITS(v256, v_uint64x4, uint64, u64, v_uint64x4, void, void, v_int64x4, void);
-    CV_DEF_REG_TRAITS(v256, v_int64x4, int64, s64, v_uint64x4, void, void, v_int64x4, void);
-    CV_DEF_REG_TRAITS(v256, v_float64x4, double, f64, v_float64x4, void, void, v_int64x4, v_int32x8);
-#endif
-
-#if CV_SIMD512
-    CV_DEF_REG_TRAITS(v512, v_uint8x64, uchar, u8, v_uint8x64, v_uint16x32, v_uint32x16, v_int8x64, void);
-    CV_DEF_REG_TRAITS(v512, v_int8x64, schar, s8, v_uint8x64, v_int16x32, v_int32x16, v_int8x64, void);
-    CV_DEF_REG_TRAITS(v512, v_uint16x32, ushort, u16, v_uint16x32, v_uint32x16, v_uint64x8, v_int16x32, void);
-    CV_DEF_REG_TRAITS(v512, v_int16x32, short, s16, v_uint16x32, v_int32x16, v_int64x8, v_int16x32, void);
-    CV_DEF_REG_TRAITS(v512, v_uint32x16, unsigned, u32, v_uint32x16, v_uint64x8, void, v_int32x16, void);
-    CV_DEF_REG_TRAITS(v512, v_int32x16, int, s32, v_uint32x16, v_int64x8, void, v_int32x16, void);
-    CV_DEF_REG_TRAITS(v512, v_float32x16, float, f32, v_float32x16, v_float64x8, void, v_int32x16, v_int32x16);
-    CV_DEF_REG_TRAITS(v512, v_uint64x8, uint64, u64, v_uint64x8, void, void, v_int64x8, void);
-    CV_DEF_REG_TRAITS(v512, v_int64x8, int64, s64, v_uint64x8, void, void, v_int64x8, void);
-    CV_DEF_REG_TRAITS(v512, v_float64x8, double, f64, v_float64x8, void, void, v_int64x8, v_int32x16);
-#endif
-//! @endcond
-
-#if CV_SIMD512 && (!defined(CV__SIMD_FORCE_WIDTH) || CV__SIMD_FORCE_WIDTH == 512)
-#define CV__SIMD_NAMESPACE simd512
-namespace CV__SIMD_NAMESPACE {
-    #define CV_SIMD 1
-    #define CV_SIMD_64F CV_SIMD512_64F
-    #define CV_SIMD_FP16 CV_SIMD512_FP16
-    #define CV_SIMD_WIDTH 64
-//! @addtogroup core_hal_intrin
-//! @{
-    //! @brief Maximum available vector register capacity 8-bit unsigned integer values
-    typedef v_uint8x64    v_uint8;
-    //! @brief Maximum available vector register capacity 8-bit signed integer values
-    typedef v_int8x64     v_int8;
-    //! @brief Maximum available vector register capacity 16-bit unsigned integer values
-    typedef v_uint16x32   v_uint16;
-    //! @brief Maximum available vector register capacity 16-bit signed integer values
-    typedef v_int16x32    v_int16;
-    //! @brief Maximum available vector register capacity 32-bit unsigned integer values
-    typedef v_uint32x16   v_uint32;
-    //! @brief Maximum available vector register capacity 32-bit signed integer values
-    typedef v_int32x16    v_int32;
-    //! @brief Maximum available vector register capacity 64-bit unsigned integer values
-    typedef v_uint64x8    v_uint64;
-    //! @brief Maximum available vector register capacity 64-bit signed integer values
-    typedef v_int64x8     v_int64;
-    //! @brief Maximum available vector register capacity 32-bit floating point values (single precision)
-    typedef v_float32x16  v_float32;
-    #if CV_SIMD512_64F
-    //! @brief Maximum available vector register capacity 64-bit floating point values (double precision)
-    typedef v_float64x8   v_float64;
-    #endif
-//! @}
-
-    #define VXPREFIX(func) v512##func
-} // namespace
-using namespace CV__SIMD_NAMESPACE;
-#elif CV_SIMD256 && (!defined(CV__SIMD_FORCE_WIDTH) || CV__SIMD_FORCE_WIDTH == 256)
-#define CV__SIMD_NAMESPACE simd256
-namespace CV__SIMD_NAMESPACE {
-    #define CV_SIMD 1
-    #define CV_SIMD_64F CV_SIMD256_64F
-    #define CV_SIMD_FP16 CV_SIMD256_FP16
-    #define CV_SIMD_WIDTH 32
-//! @addtogroup core_hal_intrin
-//! @{
-    //! @brief Maximum available vector register capacity 8-bit unsigned integer values
-    typedef v_uint8x32   v_uint8;
-    //! @brief Maximum available vector register capacity 8-bit signed integer values
-    typedef v_int8x32    v_int8;
-    //! @brief Maximum available vector register capacity 16-bit unsigned integer values
-    typedef v_uint16x16  v_uint16;
-    //! @brief Maximum available vector register capacity 16-bit signed integer values
-    typedef v_int16x16   v_int16;
-    //! @brief Maximum available vector register capacity 32-bit unsigned integer values
-    typedef v_uint32x8   v_uint32;
-    //! @brief Maximum available vector register capacity 32-bit signed integer values
-    typedef v_int32x8    v_int32;
-    //! @brief Maximum available vector register capacity 64-bit unsigned integer values
-    typedef v_uint64x4   v_uint64;
-    //! @brief Maximum available vector register capacity 64-bit signed integer values
-    typedef v_int64x4    v_int64;
-    //! @brief Maximum available vector register capacity 32-bit floating point values (single precision)
-    typedef v_float32x8  v_float32;
-    #if CV_SIMD256_64F
-    //! @brief Maximum available vector register capacity 64-bit floating point values (double precision)
-    typedef v_float64x4  v_float64;
-    #endif
-//! @}
-
-    #define VXPREFIX(func) v256##func
-} // namespace
-using namespace CV__SIMD_NAMESPACE;
-#elif (CV_SIMD128 || CV_SIMD128_CPP) && (!defined(CV__SIMD_FORCE_WIDTH) || CV__SIMD_FORCE_WIDTH == 128)
-#if defined CV_SIMD128_CPP
-#define CV__SIMD_NAMESPACE simd128_cpp
-#else
-#define CV__SIMD_NAMESPACE simd128
-#endif
-namespace CV__SIMD_NAMESPACE {
-    #define CV_SIMD CV_SIMD128
-    #define CV_SIMD_64F CV_SIMD128_64F
-    #define CV_SIMD_WIDTH 16
-//! @addtogroup core_hal_intrin
-//! @{
-    //! @brief Maximum available vector register capacity 8-bit unsigned integer values
-    typedef v_uint8x16  v_uint8;
-    //! @brief Maximum available vector register capacity 8-bit signed integer values
-    typedef v_int8x16   v_int8;
-    //! @brief Maximum available vector register capacity 16-bit unsigned integer values
-    typedef v_uint16x8  v_uint16;
-    //! @brief Maximum available vector register capacity 16-bit signed integer values
-    typedef v_int16x8   v_int16;
-    //! @brief Maximum available vector register capacity 32-bit unsigned integer values
-    typedef v_uint32x4  v_uint32;
-    //! @brief Maximum available vector register capacity 32-bit signed integer values
-    typedef v_int32x4   v_int32;
-    //! @brief Maximum available vector register capacity 64-bit unsigned integer values
-    typedef v_uint64x2  v_uint64;
-    //! @brief Maximum available vector register capacity 64-bit signed integer values
-    typedef v_int64x2   v_int64;
-    //! @brief Maximum available vector register capacity 32-bit floating point values (single precision)
-    typedef v_float32x4 v_float32;
-    #if CV_SIMD128_64F
-    //! @brief Maximum available vector register capacity 64-bit floating point values (double precision)
-    typedef v_float64x2 v_float64;
-    #endif
-//! @}
-
-    #define VXPREFIX(func) v##func
-} // namespace
-using namespace CV__SIMD_NAMESPACE;
-#endif
-
-namespace CV__SIMD_NAMESPACE {
-//! @addtogroup core_hal_intrin
-//! @{
-    //! @name Wide init with value
-    //! @{
-    //! @brief Create maximum available capacity vector with elements set to a specific value
-    inline v_uint8 vx_setall_u8(uchar v) { return VXPREFIX(_setall_u8)(v); }
-    inline v_int8 vx_setall_s8(schar v) { return VXPREFIX(_setall_s8)(v); }
-    inline v_uint16 vx_setall_u16(ushort v) { return VXPREFIX(_setall_u16)(v); }
-    inline v_int16 vx_setall_s16(short v) { return VXPREFIX(_setall_s16)(v); }
-    inline v_int32 vx_setall_s32(int v) { return VXPREFIX(_setall_s32)(v); }
-    inline v_uint32 vx_setall_u32(unsigned v) { return VXPREFIX(_setall_u32)(v); }
-    inline v_float32 vx_setall_f32(float v) { return VXPREFIX(_setall_f32)(v); }
-    inline v_int64 vx_setall_s64(int64 v) { return VXPREFIX(_setall_s64)(v); }
-    inline v_uint64 vx_setall_u64(uint64 v) { return VXPREFIX(_setall_u64)(v); }
-#if CV_SIMD_64F
-    inline v_float64 vx_setall_f64(double v) { return VXPREFIX(_setall_f64)(v); }
-#endif
-    //! @}
-
-    //! @name Wide init with zero
-    //! @{
-    //! @brief Create maximum available capacity vector with elements set to zero
-    inline v_uint8 vx_setzero_u8() { return VXPREFIX(_setzero_u8)(); }
-    inline v_int8 vx_setzero_s8() { return VXPREFIX(_setzero_s8)(); }
-    inline v_uint16 vx_setzero_u16() { return VXPREFIX(_setzero_u16)(); }
-    inline v_int16 vx_setzero_s16() { return VXPREFIX(_setzero_s16)(); }
-    inline v_int32 vx_setzero_s32() { return VXPREFIX(_setzero_s32)(); }
-    inline v_uint32 vx_setzero_u32() { return VXPREFIX(_setzero_u32)(); }
-    inline v_float32 vx_setzero_f32() { return VXPREFIX(_setzero_f32)(); }
-    inline v_int64 vx_setzero_s64() { return VXPREFIX(_setzero_s64)(); }
-    inline v_uint64 vx_setzero_u64() { return VXPREFIX(_setzero_u64)(); }
-#if CV_SIMD_64F
-    inline v_float64 vx_setzero_f64() { return VXPREFIX(_setzero_f64)(); }
-#endif
-    //! @}
-
-    //! @name Wide load from memory
-    //! @{
-    //! @brief Load maximum available capacity register contents from memory
-    inline v_uint8 vx_load(const uchar * ptr) { return VXPREFIX(_load)(ptr); }
-    inline v_int8 vx_load(const schar * ptr) { return VXPREFIX(_load)(ptr); }
-    inline v_uint16 vx_load(const ushort * ptr) { return VXPREFIX(_load)(ptr); }
-    inline v_int16 vx_load(const short * ptr) { return VXPREFIX(_load)(ptr); }
-    inline v_int32 vx_load(const int * ptr) { return VXPREFIX(_load)(ptr); }
-    inline v_uint32 vx_load(const unsigned * ptr) { return VXPREFIX(_load)(ptr); }
-    inline v_float32 vx_load(const float * ptr) { return VXPREFIX(_load)(ptr); }
-    inline v_int64 vx_load(const int64 * ptr) { return VXPREFIX(_load)(ptr); }
-    inline v_uint64 vx_load(const uint64 * ptr) { return VXPREFIX(_load)(ptr); }
-#if CV_SIMD_64F
-    inline v_float64 vx_load(const double * ptr) { return VXPREFIX(_load)(ptr); }
-#endif
-    //! @}
-
-    //! @name Wide load from memory(aligned)
-    //! @{
-    //! @brief Load maximum available capacity register contents from memory(aligned)
-    inline v_uint8 vx_load_aligned(const uchar * ptr) { return VXPREFIX(_load_aligned)(ptr); }
-    inline v_int8 vx_load_aligned(const schar * ptr) { return VXPREFIX(_load_aligned)(ptr); }
-    inline v_uint16 vx_load_aligned(const ushort * ptr) { return VXPREFIX(_load_aligned)(ptr); }
-    inline v_int16 vx_load_aligned(const short * ptr) { return VXPREFIX(_load_aligned)(ptr); }
-    inline v_int32 vx_load_aligned(const int * ptr) { return VXPREFIX(_load_aligned)(ptr); }
-    inline v_uint32 vx_load_aligned(const unsigned * ptr) { return VXPREFIX(_load_aligned)(ptr); }
-    inline v_float32 vx_load_aligned(const float * ptr) { return VXPREFIX(_load_aligned)(ptr); }
-    inline v_int64 vx_load_aligned(const int64 * ptr) { return VXPREFIX(_load_aligned)(ptr); }
-    inline v_uint64 vx_load_aligned(const uint64 * ptr) { return VXPREFIX(_load_aligned)(ptr); }
-#if CV_SIMD_64F
-    inline v_float64 vx_load_aligned(const double * ptr) { return VXPREFIX(_load_aligned)(ptr); }
-#endif
-    //! @}
-
-    //! @name Wide load lower half from memory
-    //! @{
-    //! @brief Load lower half of maximum available capacity register from memory
-    inline v_uint8 vx_load_low(const uchar * ptr) { return VXPREFIX(_load_low)(ptr); }
-    inline v_int8 vx_load_low(const schar * ptr) { return VXPREFIX(_load_low)(ptr); }
-    inline v_uint16 vx_load_low(const ushort * ptr) { return VXPREFIX(_load_low)(ptr); }
-    inline v_int16 vx_load_low(const short * ptr) { return VXPREFIX(_load_low)(ptr); }
-    inline v_int32 vx_load_low(const int * ptr) { return VXPREFIX(_load_low)(ptr); }
-    inline v_uint32 vx_load_low(const unsigned * ptr) { return VXPREFIX(_load_low)(ptr); }
-    inline v_float32 vx_load_low(const float * ptr) { return VXPREFIX(_load_low)(ptr); }
-    inline v_int64 vx_load_low(const int64 * ptr) { return VXPREFIX(_load_low)(ptr); }
-    inline v_uint64 vx_load_low(const uint64 * ptr) { return VXPREFIX(_load_low)(ptr); }
-#if CV_SIMD_64F
-    inline v_float64 vx_load_low(const double * ptr) { return VXPREFIX(_load_low)(ptr); }
-#endif
-    //! @}
-
-    //! @name Wide load halfs from memory
-    //! @{
-    //! @brief Load maximum available capacity register contents from two memory blocks
-    inline v_uint8 vx_load_halves(const uchar * ptr0, const uchar * ptr1) { return VXPREFIX(_load_halves)(ptr0, ptr1); }
-    inline v_int8 vx_load_halves(const schar * ptr0, const schar * ptr1) { return VXPREFIX(_load_halves)(ptr0, ptr1); }
-    inline v_uint16 vx_load_halves(const ushort * ptr0, const ushort * ptr1) { return VXPREFIX(_load_halves)(ptr0, ptr1); }
-    inline v_int16 vx_load_halves(const short * ptr0, const short * ptr1) { return VXPREFIX(_load_halves)(ptr0, ptr1); }
-    inline v_int32 vx_load_halves(const int * ptr0, const int * ptr1) { return VXPREFIX(_load_halves)(ptr0, ptr1); }
-    inline v_uint32 vx_load_halves(const unsigned * ptr0, const unsigned * ptr1) { return VXPREFIX(_load_halves)(ptr0, ptr1); }
-    inline v_float32 vx_load_halves(const float * ptr0, const float * ptr1) { return VXPREFIX(_load_halves)(ptr0, ptr1); }
-    inline v_int64 vx_load_halves(const int64 * ptr0, const int64 * ptr1) { return VXPREFIX(_load_halves)(ptr0, ptr1); }
-    inline v_uint64 vx_load_halves(const uint64 * ptr0, const uint64 * ptr1) { return VXPREFIX(_load_halves)(ptr0, ptr1); }
-#if CV_SIMD_64F
-    inline v_float64 vx_load_halves(const double * ptr0, const double * ptr1) { return VXPREFIX(_load_halves)(ptr0, ptr1); }
-#endif
-    //! @}
-
-    //! @name Wide LUT of elements
-    //! @{
-    //! @brief Load maximum available capacity register contents with array elements by provided indexes
-    inline v_uint8 vx_lut(const uchar * ptr, const int* idx) { return VXPREFIX(_lut)(ptr, idx); }
-    inline v_int8 vx_lut(const schar * ptr, const int* idx) { return VXPREFIX(_lut)(ptr, idx); }
-    inline v_uint16 vx_lut(const ushort * ptr, const int* idx) { return VXPREFIX(_lut)(ptr, idx); }
-    inline v_int16 vx_lut(const short* ptr, const int* idx) { return VXPREFIX(_lut)(ptr, idx); }
-    inline v_int32 vx_lut(const int* ptr, const int* idx) { return VXPREFIX(_lut)(ptr, idx); }
-    inline v_uint32 vx_lut(const unsigned* ptr, const int* idx) { return VXPREFIX(_lut)(ptr, idx); }
-    inline v_float32 vx_lut(const float* ptr, const int* idx) { return VXPREFIX(_lut)(ptr, idx); }
-    inline v_int64 vx_lut(const int64 * ptr, const int* idx) { return VXPREFIX(_lut)(ptr, idx); }
-    inline v_uint64 vx_lut(const uint64 * ptr, const int* idx) { return VXPREFIX(_lut)(ptr, idx); }
-#if CV_SIMD_64F
-    inline v_float64 vx_lut(const double* ptr, const int* idx) { return VXPREFIX(_lut)(ptr, idx); }
-#endif
-    //! @}
-
-    //! @name Wide LUT of element pairs
-    //! @{
-    //! @brief Load maximum available capacity register contents with array element pairs by provided indexes
-    inline v_uint8 vx_lut_pairs(const uchar * ptr, const int* idx) { return VXPREFIX(_lut_pairs)(ptr, idx); }
-    inline v_int8 vx_lut_pairs(const schar * ptr, const int* idx) { return VXPREFIX(_lut_pairs)(ptr, idx); }
-    inline v_uint16 vx_lut_pairs(const ushort * ptr, const int* idx) { return VXPREFIX(_lut_pairs)(ptr, idx); }
-    inline v_int16 vx_lut_pairs(const short* ptr, const int* idx) { return VXPREFIX(_lut_pairs)(ptr, idx); }
-    inline v_int32 vx_lut_pairs(const int* ptr, const int* idx) { return VXPREFIX(_lut_pairs)(ptr, idx); }
-    inline v_uint32 vx_lut_pairs(const unsigned* ptr, const int* idx) { return VXPREFIX(_lut_pairs)(ptr, idx); }
-    inline v_float32 vx_lut_pairs(const float* ptr, const int* idx) { return VXPREFIX(_lut_pairs)(ptr, idx); }
-    inline v_int64 vx_lut_pairs(const int64 * ptr, const int* idx) { return VXPREFIX(_lut_pairs)(ptr, idx); }
-    inline v_uint64 vx_lut_pairs(const uint64 * ptr, const int* idx) { return VXPREFIX(_lut_pairs)(ptr, idx); }
-#if CV_SIMD_64F
-    inline v_float64 vx_lut_pairs(const double* ptr, const int* idx) { return VXPREFIX(_lut_pairs)(ptr, idx); }
-#endif
-    //! @}
-
-    //! @name Wide LUT of element quads
-    //! @{
-    //! @brief Load maximum available capacity register contents with array element quads by provided indexes
-    inline v_uint8 vx_lut_quads(const uchar* ptr, const int* idx) { return VXPREFIX(_lut_quads)(ptr, idx); }
-    inline v_int8 vx_lut_quads(const schar* ptr, const int* idx) { return VXPREFIX(_lut_quads)(ptr, idx); }
-    inline v_uint16 vx_lut_quads(const ushort* ptr, const int* idx) { return VXPREFIX(_lut_quads)(ptr, idx); }
-    inline v_int16 vx_lut_quads(const short* ptr, const int* idx) { return VXPREFIX(_lut_quads)(ptr, idx); }
-    inline v_int32 vx_lut_quads(const int* ptr, const int* idx) { return VXPREFIX(_lut_quads)(ptr, idx); }
-    inline v_uint32 vx_lut_quads(const unsigned* ptr, const int* idx) { return VXPREFIX(_lut_quads)(ptr, idx); }
-    inline v_float32 vx_lut_quads(const float* ptr, const int* idx) { return VXPREFIX(_lut_quads)(ptr, idx); }
-    //! @}
-
-    //! @name Wide load with double expansion
-    //! @{
-    //! @brief Load maximum available capacity register contents from memory with double expand
-    inline v_uint16 vx_load_expand(const uchar * ptr) { return VXPREFIX(_load_expand)(ptr); }
-    inline v_int16 vx_load_expand(const schar * ptr) { return VXPREFIX(_load_expand)(ptr); }
-    inline v_uint32 vx_load_expand(const ushort * ptr) { return VXPREFIX(_load_expand)(ptr); }
-    inline v_int32 vx_load_expand(const short* ptr) { return VXPREFIX(_load_expand)(ptr); }
-    inline v_int64 vx_load_expand(const int* ptr) { return VXPREFIX(_load_expand)(ptr); }
-    inline v_uint64 vx_load_expand(const unsigned* ptr) { return VXPREFIX(_load_expand)(ptr); }
-    inline v_float32 vx_load_expand(const float16_t * ptr) { return VXPREFIX(_load_expand)(ptr); }
-    //! @}
-
-    //! @name Wide load with quad expansion
-    //! @{
-    //! @brief Load maximum available capacity register contents from memory with quad expand
-    inline v_uint32 vx_load_expand_q(const uchar * ptr) { return VXPREFIX(_load_expand_q)(ptr); }
-    inline v_int32 vx_load_expand_q(const schar * ptr) { return VXPREFIX(_load_expand_q)(ptr); }
-    //! @}
-
-    /** @brief SIMD processing state cleanup call */
-    inline void vx_cleanup() { VXPREFIX(_cleanup)(); }
-
-
-//! @cond IGNORED
-
-    // backward compatibility
-    template<typename _Tp, typename _Tvec> static inline
-    void vx_store(_Tp* dst, const _Tvec& v) { return v_store(dst, v); }
-    // backward compatibility
-    template<typename _Tp, typename _Tvec> static inline
-    void vx_store_aligned(_Tp* dst, const _Tvec& v) { return v_store_aligned(dst, v); }
-
-//! @endcond
-
-
-//! @}
-    #undef VXPREFIX
-} // namespace
-
-//! @cond IGNORED
-#ifndef CV_SIMD_64F
-#define CV_SIMD_64F 0
-#endif
-
-#ifndef CV_SIMD_FP16
-#define CV_SIMD_FP16 0  //!< Defined to 1 on native support of operations with float16x8_t / float16x16_t (SIMD256) types
-#endif
-
-#ifndef CV_SIMD
-#define CV_SIMD 0
-#endif
-
-#include "simd_utils.impl.hpp"
-
-#ifndef CV_DOXYGEN
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END
-#endif
-
-} // cv::
-
-//! @endcond
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_avx.hpp b/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_avx.hpp
deleted file mode 100644
index 979b6163..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_avx.hpp
+++ /dev/null
@@ -1,3177 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html
-
-#ifndef OPENCV_HAL_INTRIN_AVX_HPP
-#define OPENCV_HAL_INTRIN_AVX_HPP
-
-#define CV_SIMD256 1
-#define CV_SIMD256_64F 1
-#define CV_SIMD256_FP16 0  // no native operations with FP16 type. Only load/store from float32x8 are available (if CV_FP16 == 1)
-
-namespace cv
-{
-
-//! @cond IGNORED
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
-
-///////// Utils ////////////
-
-inline __m256i _v256_combine(const __m128i& lo, const __m128i& hi)
-{ return _mm256_inserti128_si256(_mm256_castsi128_si256(lo), hi, 1); }
-
-inline __m256 _v256_combine(const __m128& lo, const __m128& hi)
-{ return _mm256_insertf128_ps(_mm256_castps128_ps256(lo), hi, 1); }
-
-inline __m256d _v256_combine(const __m128d& lo, const __m128d& hi)
-{ return _mm256_insertf128_pd(_mm256_castpd128_pd256(lo), hi, 1); }
-
-inline int _v_cvtsi256_si32(const __m256i& a)
-{ return _mm_cvtsi128_si32(_mm256_castsi256_si128(a)); }
-
-inline __m256i _v256_shuffle_odd_64(const __m256i& v)
-{ return _mm256_permute4x64_epi64(v, _MM_SHUFFLE(3, 1, 2, 0)); }
-
-inline __m256d _v256_shuffle_odd_64(const __m256d& v)
-{ return _mm256_permute4x64_pd(v, _MM_SHUFFLE(3, 1, 2, 0)); }
-
-template<int imm>
-inline __m256i _v256_permute2x128(const __m256i& a, const __m256i& b)
-{ return _mm256_permute2x128_si256(a, b, imm); }
-
-template<int imm>
-inline __m256 _v256_permute2x128(const __m256& a, const __m256& b)
-{ return _mm256_permute2f128_ps(a, b, imm); }
-
-template<int imm>
-inline __m256d _v256_permute2x128(const __m256d& a, const __m256d& b)
-{ return _mm256_permute2f128_pd(a, b, imm); }
-
-template<int imm, typename _Tpvec>
-inline _Tpvec v256_permute2x128(const _Tpvec& a, const _Tpvec& b)
-{ return _Tpvec(_v256_permute2x128<imm>(a.val, b.val)); }
-
-template<int imm>
-inline __m256i _v256_permute4x64(const __m256i& a)
-{ return _mm256_permute4x64_epi64(a, imm); }
-
-template<int imm>
-inline __m256d _v256_permute4x64(const __m256d& a)
-{ return _mm256_permute4x64_pd(a, imm); }
-
-template<int imm, typename _Tpvec>
-inline _Tpvec v256_permute4x64(const _Tpvec& a)
-{ return _Tpvec(_v256_permute4x64<imm>(a.val)); }
-
-inline __m128i _v256_extract_high(const __m256i& v)
-{ return _mm256_extracti128_si256(v, 1); }
-
-inline __m128  _v256_extract_high(const __m256& v)
-{ return _mm256_extractf128_ps(v, 1); }
-
-inline __m128d _v256_extract_high(const __m256d& v)
-{ return _mm256_extractf128_pd(v, 1); }
-
-inline __m128i _v256_extract_low(const __m256i& v)
-{ return _mm256_castsi256_si128(v); }
-
-inline __m128  _v256_extract_low(const __m256& v)
-{ return _mm256_castps256_ps128(v); }
-
-inline __m128d _v256_extract_low(const __m256d& v)
-{ return _mm256_castpd256_pd128(v); }
-
-inline __m256i _v256_packs_epu32(const __m256i& a, const __m256i& b)
-{
-    const __m256i m = _mm256_set1_epi32(65535);
-    __m256i am = _mm256_min_epu32(a, m);
-    __m256i bm = _mm256_min_epu32(b, m);
-    return _mm256_packus_epi32(am, bm);
-}
-
-template<int i>
-inline int _v256_extract_epi8(const __m256i& a)
-{
-#if defined(CV__SIMD_HAVE_mm256_extract_epi8) || (CV_AVX2 && (!defined(_MSC_VER) || _MSC_VER >= 1910/*MSVS 2017*/))
-    return _mm256_extract_epi8(a, i);
-#else
-    __m128i b = _mm256_extractf128_si256(a, ((i) >> 4));
-    return _mm_extract_epi8(b, i & 15);  // SSE4.1
-#endif
-}
-
-template<int i>
-inline int _v256_extract_epi16(const __m256i& a)
-{
-#if defined(CV__SIMD_HAVE_mm256_extract_epi8) || (CV_AVX2 && (!defined(_MSC_VER) || _MSC_VER >= 1910/*MSVS 2017*/))
-    return _mm256_extract_epi16(a, i);
-#else
-    __m128i b = _mm256_extractf128_si256(a, ((i) >> 3));
-    return _mm_extract_epi16(b, i & 7);  // SSE2
-#endif
-}
-
-template<int i>
-inline int _v256_extract_epi32(const __m256i& a)
-{
-#if defined(CV__SIMD_HAVE_mm256_extract_epi8) || (CV_AVX2 && (!defined(_MSC_VER) || _MSC_VER >= 1910/*MSVS 2017*/))
-    return _mm256_extract_epi32(a, i);
-#else
-    __m128i b = _mm256_extractf128_si256(a, ((i) >> 2));
-    return _mm_extract_epi32(b, i & 3);  // SSE4.1
-#endif
-}
-
-template<int i>
-inline int64 _v256_extract_epi64(const __m256i& a)
-{
-#if defined(CV__SIMD_HAVE_mm256_extract_epi8) || (CV_AVX2 && (!defined(_MSC_VER) || _MSC_VER >= 1910/*MSVS 2017*/))
-    return _mm256_extract_epi64(a, i);
-#else
-    __m128i b = _mm256_extractf128_si256(a, ((i) >> 1));
-    return _mm_extract_epi64(b, i & 1);  // SSE4.1
-#endif
-}
-
-///////// Types ////////////
-
-struct v_uint8x32
-{
-    typedef uchar lane_type;
-    enum { nlanes = 32 };
-    __m256i val;
-
-    explicit v_uint8x32(__m256i v) : val(v) {}
-    v_uint8x32(uchar v0,  uchar v1,  uchar v2,  uchar v3,
-               uchar v4,  uchar v5,  uchar v6,  uchar v7,
-               uchar v8,  uchar v9,  uchar v10, uchar v11,
-               uchar v12, uchar v13, uchar v14, uchar v15,
-               uchar v16, uchar v17, uchar v18, uchar v19,
-               uchar v20, uchar v21, uchar v22, uchar v23,
-               uchar v24, uchar v25, uchar v26, uchar v27,
-               uchar v28, uchar v29, uchar v30, uchar v31)
-    {
-        val = _mm256_setr_epi8((char)v0, (char)v1, (char)v2, (char)v3,
-            (char)v4,  (char)v5,  (char)v6 , (char)v7,  (char)v8,  (char)v9,
-            (char)v10, (char)v11, (char)v12, (char)v13, (char)v14, (char)v15,
-            (char)v16, (char)v17, (char)v18, (char)v19, (char)v20, (char)v21,
-            (char)v22, (char)v23, (char)v24, (char)v25, (char)v26, (char)v27,
-            (char)v28, (char)v29, (char)v30, (char)v31);
-    }
-    /* coverity[uninit_ctor]: suppress warning */
-    v_uint8x32() {}
-
-    uchar get0() const { return (uchar)_v_cvtsi256_si32(val); }
-};
-
-struct v_int8x32
-{
-    typedef schar lane_type;
-    enum { nlanes = 32 };
-    __m256i val;
-
-    explicit v_int8x32(__m256i v) : val(v) {}
-    v_int8x32(schar v0,  schar v1,  schar v2,  schar v3,
-              schar v4,  schar v5,  schar v6,  schar v7,
-              schar v8,  schar v9,  schar v10, schar v11,
-              schar v12, schar v13, schar v14, schar v15,
-              schar v16, schar v17, schar v18, schar v19,
-              schar v20, schar v21, schar v22, schar v23,
-              schar v24, schar v25, schar v26, schar v27,
-              schar v28, schar v29, schar v30, schar v31)
-    {
-        val = _mm256_setr_epi8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9,
-            v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20,
-            v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31);
-    }
-    /* coverity[uninit_ctor]: suppress warning */
-    v_int8x32() {}
-
-    schar get0() const { return (schar)_v_cvtsi256_si32(val); }
-};
-
-struct v_uint16x16
-{
-    typedef ushort lane_type;
-    enum { nlanes = 16 };
-    __m256i val;
-
-    explicit v_uint16x16(__m256i v) : val(v) {}
-    v_uint16x16(ushort v0,  ushort v1,  ushort v2,  ushort v3,
-                ushort v4,  ushort v5,  ushort v6,  ushort v7,
-                ushort v8,  ushort v9,  ushort v10, ushort v11,
-                ushort v12, ushort v13, ushort v14, ushort v15)
-    {
-        val = _mm256_setr_epi16((short)v0, (short)v1, (short)v2, (short)v3,
-            (short)v4,  (short)v5,  (short)v6,  (short)v7,  (short)v8,  (short)v9,
-            (short)v10, (short)v11, (short)v12, (short)v13, (short)v14, (short)v15);
-    }
-    /* coverity[uninit_ctor]: suppress warning */
-    v_uint16x16() {}
-
-    ushort get0() const { return (ushort)_v_cvtsi256_si32(val); }
-};
-
-struct v_int16x16
-{
-    typedef short lane_type;
-    enum { nlanes = 16 };
-    __m256i val;
-
-    explicit v_int16x16(__m256i v) : val(v) {}
-    v_int16x16(short v0,  short v1,  short v2,  short v3,
-               short v4,  short v5,  short v6,  short v7,
-               short v8,  short v9,  short v10, short v11,
-               short v12, short v13, short v14, short v15)
-    {
-        val = _mm256_setr_epi16(v0, v1, v2, v3, v4, v5, v6, v7,
-            v8, v9, v10, v11, v12, v13, v14, v15);
-    }
-    /* coverity[uninit_ctor]: suppress warning */
-    v_int16x16() {}
-
-    short get0() const { return (short)_v_cvtsi256_si32(val); }
-};
-
-struct v_uint32x8
-{
-    typedef unsigned lane_type;
-    enum { nlanes = 8 };
-    __m256i val;
-
-    explicit v_uint32x8(__m256i v) : val(v) {}
-    v_uint32x8(unsigned v0, unsigned v1, unsigned v2, unsigned v3,
-               unsigned v4, unsigned v5, unsigned v6, unsigned v7)
-    {
-        val = _mm256_setr_epi32((unsigned)v0, (unsigned)v1, (unsigned)v2,
-            (unsigned)v3, (unsigned)v4, (unsigned)v5, (unsigned)v6, (unsigned)v7);
-    }
-    /* coverity[uninit_ctor]: suppress warning */
-    v_uint32x8() {}
-
-    unsigned get0() const { return (unsigned)_v_cvtsi256_si32(val); }
-};
-
-struct v_int32x8
-{
-    typedef int lane_type;
-    enum { nlanes = 8 };
-    __m256i val;
-
-    explicit v_int32x8(__m256i v) : val(v) {}
-    v_int32x8(int v0, int v1, int v2, int v3,
-              int v4, int v5, int v6, int v7)
-    {
-        val = _mm256_setr_epi32(v0, v1, v2, v3, v4, v5, v6, v7);
-    }
-    /* coverity[uninit_ctor]: suppress warning */
-    v_int32x8() {}
-
-    int get0() const { return _v_cvtsi256_si32(val); }
-};
-
-struct v_float32x8
-{
-    typedef float lane_type;
-    enum { nlanes = 8 };
-    __m256 val;
-
-    explicit v_float32x8(__m256 v) : val(v) {}
-    v_float32x8(float v0, float v1, float v2, float v3,
-                float v4, float v5, float v6, float v7)
-    {
-        val = _mm256_setr_ps(v0, v1, v2, v3, v4, v5, v6, v7);
-    }
-    /* coverity[uninit_ctor]: suppress warning */
-    v_float32x8() {}
-
-    float get0() const { return _mm_cvtss_f32(_mm256_castps256_ps128(val)); }
-};
-
-struct v_uint64x4
-{
-    typedef uint64 lane_type;
-    enum { nlanes = 4 };
-    __m256i val;
-
-    explicit v_uint64x4(__m256i v) : val(v) {}
-    v_uint64x4(uint64 v0, uint64 v1, uint64 v2, uint64 v3)
-    { val = _mm256_setr_epi64x((int64)v0, (int64)v1, (int64)v2, (int64)v3); }
-    /* coverity[uninit_ctor]: suppress warning */
-    v_uint64x4() {}
-
-    uint64 get0() const
-    {
-    #if defined __x86_64__ || defined _M_X64
-        return (uint64)_mm_cvtsi128_si64(_mm256_castsi256_si128(val));
-    #else
-        int a = _mm_cvtsi128_si32(_mm256_castsi256_si128(val));
-        int b = _mm_cvtsi128_si32(_mm256_castsi256_si128(_mm256_srli_epi64(val, 32)));
-        return (unsigned)a | ((uint64)(unsigned)b << 32);
-    #endif
-    }
-};
-
-struct v_int64x4
-{
-    typedef int64 lane_type;
-    enum { nlanes = 4 };
-    __m256i val;
-
-    explicit v_int64x4(__m256i v) : val(v) {}
-    v_int64x4(int64 v0, int64 v1, int64 v2, int64 v3)
-    { val = _mm256_setr_epi64x(v0, v1, v2, v3); }
-    /* coverity[uninit_ctor]: suppress warning */
-    v_int64x4() {}
-
-    int64 get0() const
-    {
-    #if defined __x86_64__ || defined _M_X64
-        return (int64)_mm_cvtsi128_si64(_mm256_castsi256_si128(val));
-    #else
-        int a = _mm_cvtsi128_si32(_mm256_castsi256_si128(val));
-        int b = _mm_cvtsi128_si32(_mm256_castsi256_si128(_mm256_srli_epi64(val, 32)));
-        return (int64)((unsigned)a | ((uint64)(unsigned)b << 32));
-    #endif
-    }
-};
-
-struct v_float64x4
-{
-    typedef double lane_type;
-    enum { nlanes = 4 };
-    __m256d val;
-
-    explicit v_float64x4(__m256d v) : val(v) {}
-    v_float64x4(double v0, double v1, double v2, double v3)
-    { val = _mm256_setr_pd(v0, v1, v2, v3); }
-    /* coverity[uninit_ctor]: suppress warning */
-    v_float64x4() {}
-
-    double get0() const { return _mm_cvtsd_f64(_mm256_castpd256_pd128(val)); }
-};
-
-//////////////// Load and store operations ///////////////
-
-#define OPENCV_HAL_IMPL_AVX_LOADSTORE(_Tpvec, _Tp)                    \
-    inline _Tpvec v256_load(const _Tp* ptr)                           \
-    { return _Tpvec(_mm256_loadu_si256((const __m256i*)ptr)); }       \
-    inline _Tpvec v256_load_aligned(const _Tp* ptr)                   \
-    { return _Tpvec(_mm256_load_si256((const __m256i*)ptr)); }        \
-    inline _Tpvec v256_load_low(const _Tp* ptr)                       \
-    {                                                                 \
-        __m128i v128 = _mm_loadu_si128((const __m128i*)ptr);          \
-        return _Tpvec(_mm256_castsi128_si256(v128));                  \
-    }                                                                 \
-    inline _Tpvec v256_load_halves(const _Tp* ptr0, const _Tp* ptr1)  \
-    {                                                                 \
-        __m128i vlo = _mm_loadu_si128((const __m128i*)ptr0);          \
-        __m128i vhi = _mm_loadu_si128((const __m128i*)ptr1);          \
-        return _Tpvec(_v256_combine(vlo, vhi));                       \
-    }                                                                 \
-    inline void v_store(_Tp* ptr, const _Tpvec& a)                    \
-    { _mm256_storeu_si256((__m256i*)ptr, a.val); }                    \
-    inline void v_store_aligned(_Tp* ptr, const _Tpvec& a)            \
-    { _mm256_store_si256((__m256i*)ptr, a.val); }                     \
-    inline void v_store_aligned_nocache(_Tp* ptr, const _Tpvec& a)    \
-    { _mm256_stream_si256((__m256i*)ptr, a.val); }                    \
-    inline void v_store(_Tp* ptr, const _Tpvec& a, hal::StoreMode mode) \
-    { \
-        if( mode == hal::STORE_UNALIGNED ) \
-            _mm256_storeu_si256((__m256i*)ptr, a.val); \
-        else if( mode == hal::STORE_ALIGNED_NOCACHE )  \
-            _mm256_stream_si256((__m256i*)ptr, a.val); \
-        else \
-            _mm256_store_si256((__m256i*)ptr, a.val); \
-    } \
-    inline void v_store_low(_Tp* ptr, const _Tpvec& a)                \
-    { _mm_storeu_si128((__m128i*)ptr, _v256_extract_low(a.val)); }    \
-    inline void v_store_high(_Tp* ptr, const _Tpvec& a)               \
-    { _mm_storeu_si128((__m128i*)ptr, _v256_extract_high(a.val)); }
-
-OPENCV_HAL_IMPL_AVX_LOADSTORE(v_uint8x32,  uchar)
-OPENCV_HAL_IMPL_AVX_LOADSTORE(v_int8x32,   schar)
-OPENCV_HAL_IMPL_AVX_LOADSTORE(v_uint16x16, ushort)
-OPENCV_HAL_IMPL_AVX_LOADSTORE(v_int16x16,  short)
-OPENCV_HAL_IMPL_AVX_LOADSTORE(v_uint32x8,  unsigned)
-OPENCV_HAL_IMPL_AVX_LOADSTORE(v_int32x8,   int)
-OPENCV_HAL_IMPL_AVX_LOADSTORE(v_uint64x4,  uint64)
-OPENCV_HAL_IMPL_AVX_LOADSTORE(v_int64x4,   int64)
-
-#define OPENCV_HAL_IMPL_AVX_LOADSTORE_FLT(_Tpvec, _Tp, suffix, halfreg)   \
-    inline _Tpvec v256_load(const _Tp* ptr)                               \
-    { return _Tpvec(_mm256_loadu_##suffix(ptr)); }                        \
-    inline _Tpvec v256_load_aligned(const _Tp* ptr)                       \
-    { return _Tpvec(_mm256_load_##suffix(ptr)); }                         \
-    inline _Tpvec v256_load_low(const _Tp* ptr)                           \
-    {                                                                     \
-        return _Tpvec(_mm256_cast##suffix##128_##suffix##256              \
-                     (_mm_loadu_##suffix(ptr)));                          \
-    }                                                                     \
-    inline _Tpvec v256_load_halves(const _Tp* ptr0, const _Tp* ptr1)      \
-    {                                                                     \
-        halfreg vlo = _mm_loadu_##suffix(ptr0);                           \
-        halfreg vhi = _mm_loadu_##suffix(ptr1);                           \
-        return _Tpvec(_v256_combine(vlo, vhi));                           \
-    }                                                                     \
-    inline void v_store(_Tp* ptr, const _Tpvec& a)                        \
-    { _mm256_storeu_##suffix(ptr, a.val); }                               \
-    inline void v_store_aligned(_Tp* ptr, const _Tpvec& a)                \
-    { _mm256_store_##suffix(ptr, a.val); }                                \
-    inline void v_store_aligned_nocache(_Tp* ptr, const _Tpvec& a)        \
-    { _mm256_stream_##suffix(ptr, a.val); }                               \
-    inline void v_store(_Tp* ptr, const _Tpvec& a, hal::StoreMode mode) \
-    { \
-        if( mode == hal::STORE_UNALIGNED ) \
-            _mm256_storeu_##suffix(ptr, a.val); \
-        else if( mode == hal::STORE_ALIGNED_NOCACHE )  \
-            _mm256_stream_##suffix(ptr, a.val); \
-        else \
-            _mm256_store_##suffix(ptr, a.val); \
-    } \
-    inline void v_store_low(_Tp* ptr, const _Tpvec& a)                    \
-    { _mm_storeu_##suffix(ptr, _v256_extract_low(a.val)); }               \
-    inline void v_store_high(_Tp* ptr, const _Tpvec& a)                   \
-    { _mm_storeu_##suffix(ptr, _v256_extract_high(a.val)); }
-
-OPENCV_HAL_IMPL_AVX_LOADSTORE_FLT(v_float32x8, float,  ps, __m128)
-OPENCV_HAL_IMPL_AVX_LOADSTORE_FLT(v_float64x4, double, pd, __m128d)
-
-#define OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, _Tpvecf, suffix, cast) \
-    inline _Tpvec v_reinterpret_as_##suffix(const _Tpvecf& a)   \
-    { return _Tpvec(cast(a.val)); }
-
-#define OPENCV_HAL_IMPL_AVX_INIT(_Tpvec, _Tp, suffix, ssuffix, ctype_s)          \
-    inline _Tpvec v256_setzero_##suffix()                                        \
-    { return _Tpvec(_mm256_setzero_si256()); }                                   \
-    inline _Tpvec v256_setall_##suffix(_Tp v)                                    \
-    { return _Tpvec(_mm256_set1_##ssuffix((ctype_s)v)); }                        \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_uint8x32,  suffix, OPENCV_HAL_NOP)        \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_int8x32,   suffix, OPENCV_HAL_NOP)        \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_uint16x16, suffix, OPENCV_HAL_NOP)        \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_int16x16,  suffix, OPENCV_HAL_NOP)        \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_uint32x8,  suffix, OPENCV_HAL_NOP)        \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_int32x8,   suffix, OPENCV_HAL_NOP)        \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_uint64x4,  suffix, OPENCV_HAL_NOP)        \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_int64x4,   suffix, OPENCV_HAL_NOP)        \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_float32x8, suffix, _mm256_castps_si256)   \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_float64x4, suffix, _mm256_castpd_si256)
-
-OPENCV_HAL_IMPL_AVX_INIT(v_uint8x32,  uchar,    u8,  epi8,   char)
-OPENCV_HAL_IMPL_AVX_INIT(v_int8x32,   schar,    s8,  epi8,   char)
-OPENCV_HAL_IMPL_AVX_INIT(v_uint16x16, ushort,   u16, epi16,  short)
-OPENCV_HAL_IMPL_AVX_INIT(v_int16x16,  short,    s16, epi16,  short)
-OPENCV_HAL_IMPL_AVX_INIT(v_uint32x8,  unsigned, u32, epi32,  int)
-OPENCV_HAL_IMPL_AVX_INIT(v_int32x8,   int,      s32, epi32,  int)
-OPENCV_HAL_IMPL_AVX_INIT(v_uint64x4,  uint64,   u64, epi64x, int64)
-OPENCV_HAL_IMPL_AVX_INIT(v_int64x4,   int64,    s64, epi64x, int64)
-
-#define OPENCV_HAL_IMPL_AVX_INIT_FLT(_Tpvec, _Tp, suffix, zsuffix, cast) \
-    inline _Tpvec v256_setzero_##suffix()                                \
-    { return _Tpvec(_mm256_setzero_##zsuffix()); }                       \
-    inline _Tpvec v256_setall_##suffix(_Tp v)                            \
-    { return _Tpvec(_mm256_set1_##zsuffix(v)); }                         \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_uint8x32,  suffix, cast)          \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_int8x32,   suffix, cast)          \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_uint16x16, suffix, cast)          \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_int16x16,  suffix, cast)          \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_uint32x8,  suffix, cast)          \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_int32x8,   suffix, cast)          \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_uint64x4,  suffix, cast)          \
-    OPENCV_HAL_IMPL_AVX_CAST(_Tpvec, v_int64x4,   suffix, cast)
-
-OPENCV_HAL_IMPL_AVX_INIT_FLT(v_float32x8, float,  f32, ps, _mm256_castsi256_ps)
-OPENCV_HAL_IMPL_AVX_INIT_FLT(v_float64x4, double, f64, pd, _mm256_castsi256_pd)
-
-inline v_float32x8 v_reinterpret_as_f32(const v_float32x8& a)
-{ return a; }
-inline v_float32x8 v_reinterpret_as_f32(const v_float64x4& a)
-{ return v_float32x8(_mm256_castpd_ps(a.val)); }
-
-inline v_float64x4 v_reinterpret_as_f64(const v_float64x4& a)
-{ return a; }
-inline v_float64x4 v_reinterpret_as_f64(const v_float32x8& a)
-{ return v_float64x4(_mm256_castps_pd(a.val)); }
-
-/* Recombine */
-/*#define OPENCV_HAL_IMPL_AVX_COMBINE(_Tpvec, perm)                    \
-    inline _Tpvec v_combine_low(const _Tpvec& a, const _Tpvec& b)    \
-    { return _Tpvec(perm(a.val, b.val, 0x20)); }                     \
-    inline _Tpvec v_combine_high(const _Tpvec& a, const _Tpvec& b)   \
-    { return _Tpvec(perm(a.val, b.val, 0x31)); }                     \
-    inline void v_recombine(const _Tpvec& a, const _Tpvec& b,        \
-                             _Tpvec& c, _Tpvec& d)                   \
-    { c = v_combine_low(a, b); d = v_combine_high(a, b); }
-
-#define OPENCV_HAL_IMPL_AVX_UNPACKS(_Tpvec, suffix)                  \
-    OPENCV_HAL_IMPL_AVX_COMBINE(_Tpvec, _mm256_permute2x128_si256)   \
-    inline void v_zip(const _Tpvec& a0, const _Tpvec& a1,            \
-                             _Tpvec& b0, _Tpvec& b1)                 \
-    {                                                                \
-        __m256i v0 = _v256_shuffle_odd_64(a0.val);                   \
-        __m256i v1 = _v256_shuffle_odd_64(a1.val);                   \
-        b0.val = _mm256_unpacklo_##suffix(v0, v1);                   \
-        b1.val = _mm256_unpackhi_##suffix(v0, v1);                   \
-    }
-
-OPENCV_HAL_IMPL_AVX_UNPACKS(v_uint8x32,  epi8)
-OPENCV_HAL_IMPL_AVX_UNPACKS(v_int8x32,   epi8)
-OPENCV_HAL_IMPL_AVX_UNPACKS(v_uint16x16, epi16)
-OPENCV_HAL_IMPL_AVX_UNPACKS(v_int16x16,  epi16)
-OPENCV_HAL_IMPL_AVX_UNPACKS(v_uint32x8,  epi32)
-OPENCV_HAL_IMPL_AVX_UNPACKS(v_int32x8,   epi32)
-OPENCV_HAL_IMPL_AVX_UNPACKS(v_uint64x4,  epi64)
-OPENCV_HAL_IMPL_AVX_UNPACKS(v_int64x4,   epi64)
-OPENCV_HAL_IMPL_AVX_COMBINE(v_float32x8, _mm256_permute2f128_ps)
-OPENCV_HAL_IMPL_AVX_COMBINE(v_float64x4, _mm256_permute2f128_pd)
-
-inline void v_zip(const v_float32x8& a0, const v_float32x8& a1, v_float32x8& b0, v_float32x8& b1)
-{
-    __m256 v0 = _mm256_unpacklo_ps(a0.val, a1.val);
-    __m256 v1 = _mm256_unpackhi_ps(a0.val, a1.val);
-    v_recombine(v_float32x8(v0), v_float32x8(v1), b0, b1);
-}
-
-inline void v_zip(const v_float64x4& a0, const v_float64x4& a1, v_float64x4& b0, v_float64x4& b1)
-{
-    __m256d v0 = _v_shuffle_odd_64(a0.val);
-    __m256d v1 = _v_shuffle_odd_64(a1.val);
-    b0.val = _mm256_unpacklo_pd(v0, v1);
-    b1.val = _mm256_unpackhi_pd(v0, v1);
-}*/
-
-//////////////// Variant Value reordering ///////////////
-
-// unpacks
-#define OPENCV_HAL_IMPL_AVX_UNPACK(_Tpvec, suffix)                 \
-    inline _Tpvec v256_unpacklo(const _Tpvec& a, const _Tpvec& b)  \
-    { return _Tpvec(_mm256_unpacklo_##suffix(a.val, b.val)); }     \
-    inline _Tpvec v256_unpackhi(const _Tpvec& a, const _Tpvec& b)  \
-    { return _Tpvec(_mm256_unpackhi_##suffix(a.val, b.val)); }
-
-OPENCV_HAL_IMPL_AVX_UNPACK(v_uint8x32,  epi8)
-OPENCV_HAL_IMPL_AVX_UNPACK(v_int8x32,   epi8)
-OPENCV_HAL_IMPL_AVX_UNPACK(v_uint16x16, epi16)
-OPENCV_HAL_IMPL_AVX_UNPACK(v_int16x16,  epi16)
-OPENCV_HAL_IMPL_AVX_UNPACK(v_uint32x8,  epi32)
-OPENCV_HAL_IMPL_AVX_UNPACK(v_int32x8,   epi32)
-OPENCV_HAL_IMPL_AVX_UNPACK(v_uint64x4,  epi64)
-OPENCV_HAL_IMPL_AVX_UNPACK(v_int64x4,   epi64)
-OPENCV_HAL_IMPL_AVX_UNPACK(v_float32x8, ps)
-OPENCV_HAL_IMPL_AVX_UNPACK(v_float64x4, pd)
-
-// blend
-#define OPENCV_HAL_IMPL_AVX_BLEND(_Tpvec, suffix)               \
-    template<int m>                                             \
-    inline _Tpvec v256_blend(const _Tpvec& a, const _Tpvec& b)  \
-    { return _Tpvec(_mm256_blend_##suffix(a.val, b.val, m)); }
-
-OPENCV_HAL_IMPL_AVX_BLEND(v_uint16x16, epi16)
-OPENCV_HAL_IMPL_AVX_BLEND(v_int16x16,  epi16)
-OPENCV_HAL_IMPL_AVX_BLEND(v_uint32x8,  epi32)
-OPENCV_HAL_IMPL_AVX_BLEND(v_int32x8,   epi32)
-OPENCV_HAL_IMPL_AVX_BLEND(v_float32x8, ps)
-OPENCV_HAL_IMPL_AVX_BLEND(v_float64x4, pd)
-
-template<int m>
-inline v_uint64x4 v256_blend(const v_uint64x4& a, const v_uint64x4& b)
-{
-    enum {M0 = m};
-    enum {M1 = (M0 | (M0 << 2)) & 0x33};
-    enum {M2 = (M1 | (M1 << 1)) & 0x55};
-    enum {MM =  M2 | (M2 << 1)};
-    return v_uint64x4(_mm256_blend_epi32(a.val, b.val, MM));
-}
-template<int m>
-inline v_int64x4 v256_blend(const v_int64x4& a, const v_int64x4& b)
-{ return v_int64x4(v256_blend<m>(v_uint64x4(a.val), v_uint64x4(b.val)).val); }
-
-// shuffle
-// todo: emulate 64bit
-#define OPENCV_HAL_IMPL_AVX_SHUFFLE(_Tpvec, intrin)  \
-    template<int m>                                  \
-    inline _Tpvec v256_shuffle(const _Tpvec& a)      \
-    { return _Tpvec(_mm256_##intrin(a.val, m)); }
-
-OPENCV_HAL_IMPL_AVX_SHUFFLE(v_uint32x8,  shuffle_epi32)
-OPENCV_HAL_IMPL_AVX_SHUFFLE(v_int32x8,   shuffle_epi32)
-OPENCV_HAL_IMPL_AVX_SHUFFLE(v_float32x8, permute_ps)
-OPENCV_HAL_IMPL_AVX_SHUFFLE(v_float64x4, permute_pd)
-
-template<typename _Tpvec>
-inline void v256_zip(const _Tpvec& a, const _Tpvec& b, _Tpvec& ab0, _Tpvec& ab1)
-{
-    ab0 = v256_unpacklo(a, b);
-    ab1 = v256_unpackhi(a, b);
-}
-
-template<typename _Tpvec>
-inline _Tpvec v256_combine_diagonal(const _Tpvec& a, const _Tpvec& b)
-{ return _Tpvec(_mm256_blend_epi32(a.val, b.val, 0xf0)); }
-
-inline v_float32x8 v256_combine_diagonal(const v_float32x8& a, const v_float32x8& b)
-{ return v256_blend<0xf0>(a, b); }
-
-inline v_float64x4 v256_combine_diagonal(const v_float64x4& a, const v_float64x4& b)
-{ return v256_blend<0xc>(a, b); }
-
-template<typename _Tpvec>
-inline _Tpvec v256_alignr_128(const _Tpvec& a, const _Tpvec& b)
-{ return v256_permute2x128<0x21>(a, b); }
-
-template<typename _Tpvec>
-inline _Tpvec v256_alignr_64(const _Tpvec& a, const _Tpvec& b)
-{ return _Tpvec(_mm256_alignr_epi8(a.val, b.val, 8)); }
-inline v_float64x4 v256_alignr_64(const v_float64x4& a, const v_float64x4& b)
-{ return v_float64x4(_mm256_shuffle_pd(b.val, a.val, _MM_SHUFFLE(0, 0, 1, 1))); }
-// todo: emulate float32
-
-template<typename _Tpvec>
-inline _Tpvec v256_swap_halves(const _Tpvec& a)
-{ return v256_permute2x128<1>(a, a); }
-
-template<typename _Tpvec>
-inline _Tpvec v256_reverse_64(const _Tpvec& a)
-{ return v256_permute4x64<_MM_SHUFFLE(0, 1, 2, 3)>(a); }
-
-// ZIP
-#define OPENCV_HAL_IMPL_AVX_ZIP(_Tpvec)                              \
-    inline _Tpvec v_combine_low(const _Tpvec& a, const _Tpvec& b)    \
-    { return v256_permute2x128<0x20>(a, b); }                        \
-    inline _Tpvec v_combine_high(const _Tpvec& a, const _Tpvec& b)   \
-    { return v256_permute2x128<0x31>(a, b); }                        \
-    inline void v_recombine(const _Tpvec& a, const _Tpvec& b,        \
-                             _Tpvec& c, _Tpvec& d)                   \
-    {                                                                \
-        _Tpvec a1b0 = v256_alignr_128(a, b);                         \
-        c = v256_combine_diagonal(a, a1b0);                          \
-        d = v256_combine_diagonal(a1b0, b);                          \
-    }                                                                \
-    inline void v_zip(const _Tpvec& a, const _Tpvec& b,              \
-                      _Tpvec& ab0, _Tpvec& ab1)                      \
-    {                                                                \
-        _Tpvec ab0ab2, ab1ab3;                                       \
-        v256_zip(a, b, ab0ab2, ab1ab3);                              \
-        v_recombine(ab0ab2, ab1ab3, ab0, ab1);                       \
-    }
-
-OPENCV_HAL_IMPL_AVX_ZIP(v_uint8x32)
-OPENCV_HAL_IMPL_AVX_ZIP(v_int8x32)
-OPENCV_HAL_IMPL_AVX_ZIP(v_uint16x16)
-OPENCV_HAL_IMPL_AVX_ZIP(v_int16x16)
-OPENCV_HAL_IMPL_AVX_ZIP(v_uint32x8)
-OPENCV_HAL_IMPL_AVX_ZIP(v_int32x8)
-OPENCV_HAL_IMPL_AVX_ZIP(v_uint64x4)
-OPENCV_HAL_IMPL_AVX_ZIP(v_int64x4)
-OPENCV_HAL_IMPL_AVX_ZIP(v_float32x8)
-OPENCV_HAL_IMPL_AVX_ZIP(v_float64x4)
-
-////////// Arithmetic, bitwise and comparison operations /////////
-
-/* Element-wise binary and unary operations */
-
-/** Arithmetics **/
-#define OPENCV_HAL_IMPL_AVX_BIN_OP(bin_op, _Tpvec, intrin)            \
-    inline _Tpvec operator bin_op (const _Tpvec& a, const _Tpvec& b)  \
-    { return _Tpvec(intrin(a.val, b.val)); }                          \
-    inline _Tpvec& operator bin_op##= (_Tpvec& a, const _Tpvec& b)    \
-    { a.val = intrin(a.val, b.val); return a; }
-
-OPENCV_HAL_IMPL_AVX_BIN_OP(+, v_uint8x32,  _mm256_adds_epu8)
-OPENCV_HAL_IMPL_AVX_BIN_OP(-, v_uint8x32,  _mm256_subs_epu8)
-OPENCV_HAL_IMPL_AVX_BIN_OP(+, v_int8x32,   _mm256_adds_epi8)
-OPENCV_HAL_IMPL_AVX_BIN_OP(-, v_int8x32,   _mm256_subs_epi8)
-OPENCV_HAL_IMPL_AVX_BIN_OP(+, v_uint16x16, _mm256_adds_epu16)
-OPENCV_HAL_IMPL_AVX_BIN_OP(-, v_uint16x16, _mm256_subs_epu16)
-OPENCV_HAL_IMPL_AVX_BIN_OP(+, v_int16x16,  _mm256_adds_epi16)
-OPENCV_HAL_IMPL_AVX_BIN_OP(-, v_int16x16,  _mm256_subs_epi16)
-OPENCV_HAL_IMPL_AVX_BIN_OP(+, v_uint32x8,  _mm256_add_epi32)
-OPENCV_HAL_IMPL_AVX_BIN_OP(-, v_uint32x8,  _mm256_sub_epi32)
-OPENCV_HAL_IMPL_AVX_BIN_OP(*, v_uint32x8,  _mm256_mullo_epi32)
-OPENCV_HAL_IMPL_AVX_BIN_OP(+, v_int32x8,   _mm256_add_epi32)
-OPENCV_HAL_IMPL_AVX_BIN_OP(-, v_int32x8,   _mm256_sub_epi32)
-OPENCV_HAL_IMPL_AVX_BIN_OP(*, v_int32x8,   _mm256_mullo_epi32)
-OPENCV_HAL_IMPL_AVX_BIN_OP(+, v_uint64x4,  _mm256_add_epi64)
-OPENCV_HAL_IMPL_AVX_BIN_OP(-, v_uint64x4,  _mm256_sub_epi64)
-OPENCV_HAL_IMPL_AVX_BIN_OP(+, v_int64x4,   _mm256_add_epi64)
-OPENCV_HAL_IMPL_AVX_BIN_OP(-, v_int64x4,   _mm256_sub_epi64)
-
-OPENCV_HAL_IMPL_AVX_BIN_OP(+, v_float32x8, _mm256_add_ps)
-OPENCV_HAL_IMPL_AVX_BIN_OP(-, v_float32x8, _mm256_sub_ps)
-OPENCV_HAL_IMPL_AVX_BIN_OP(*, v_float32x8, _mm256_mul_ps)
-OPENCV_HAL_IMPL_AVX_BIN_OP(/, v_float32x8, _mm256_div_ps)
-OPENCV_HAL_IMPL_AVX_BIN_OP(+, v_float64x4, _mm256_add_pd)
-OPENCV_HAL_IMPL_AVX_BIN_OP(-, v_float64x4, _mm256_sub_pd)
-OPENCV_HAL_IMPL_AVX_BIN_OP(*, v_float64x4, _mm256_mul_pd)
-OPENCV_HAL_IMPL_AVX_BIN_OP(/, v_float64x4, _mm256_div_pd)
-
-// saturating multiply 8-bit, 16-bit
-inline v_uint8x32 operator * (const v_uint8x32& a, const v_uint8x32& b)
-{
-    v_uint16x16 c, d;
-    v_mul_expand(a, b, c, d);
-    return v_pack(c, d);
-}
-inline v_int8x32 operator * (const v_int8x32& a, const v_int8x32& b)
-{
-    v_int16x16 c, d;
-    v_mul_expand(a, b, c, d);
-    return v_pack(c, d);
-}
-inline v_uint16x16 operator * (const v_uint16x16& a, const v_uint16x16& b)
-{
-    __m256i pl = _mm256_mullo_epi16(a.val, b.val);
-    __m256i ph = _mm256_mulhi_epu16(a.val, b.val);
-    __m256i p0 = _mm256_unpacklo_epi16(pl, ph);
-    __m256i p1 = _mm256_unpackhi_epi16(pl, ph);
-    return v_uint16x16(_v256_packs_epu32(p0, p1));
-}
-inline v_int16x16 operator * (const v_int16x16& a, const v_int16x16& b)
-{
-    __m256i pl = _mm256_mullo_epi16(a.val, b.val);
-    __m256i ph = _mm256_mulhi_epi16(a.val, b.val);
-    __m256i p0 = _mm256_unpacklo_epi16(pl, ph);
-    __m256i p1 = _mm256_unpackhi_epi16(pl, ph);
-    return v_int16x16(_mm256_packs_epi32(p0, p1));
-}
-inline v_uint8x32& operator *= (v_uint8x32& a, const v_uint8x32& b)
-{ a = a * b; return a; }
-inline v_int8x32& operator *= (v_int8x32& a, const v_int8x32& b)
-{ a = a * b; return a; }
-inline v_uint16x16& operator *= (v_uint16x16& a, const v_uint16x16& b)
-{ a = a * b; return a; }
-inline v_int16x16& operator *= (v_int16x16& a, const v_int16x16& b)
-{ a = a * b; return a; }
-
-/** Non-saturating arithmetics **/
-#define OPENCV_HAL_IMPL_AVX_BIN_FUNC(func, _Tpvec, intrin) \
-    inline _Tpvec func(const _Tpvec& a, const _Tpvec& b)   \
-    { return _Tpvec(intrin(a.val, b.val)); }
-
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_add_wrap, v_uint8x32,  _mm256_add_epi8)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_add_wrap, v_int8x32,   _mm256_add_epi8)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_add_wrap, v_uint16x16, _mm256_add_epi16)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_add_wrap, v_int16x16,  _mm256_add_epi16)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_sub_wrap, v_uint8x32,  _mm256_sub_epi8)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_sub_wrap, v_int8x32,   _mm256_sub_epi8)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_sub_wrap, v_uint16x16, _mm256_sub_epi16)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_sub_wrap, v_int16x16,  _mm256_sub_epi16)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_mul_wrap, v_uint16x16, _mm256_mullo_epi16)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_mul_wrap, v_int16x16,  _mm256_mullo_epi16)
-
-inline v_uint8x32 v_mul_wrap(const v_uint8x32& a, const v_uint8x32& b)
-{
-    __m256i ad = _mm256_srai_epi16(a.val, 8);
-    __m256i bd = _mm256_srai_epi16(b.val, 8);
-    __m256i p0 = _mm256_mullo_epi16(a.val, b.val); // even
-    __m256i p1 = _mm256_slli_epi16(_mm256_mullo_epi16(ad, bd), 8); // odd
-
-    const __m256i b01 = _mm256_set1_epi32(0xFF00FF00);
-    return v_uint8x32(_mm256_blendv_epi8(p0, p1, b01));
-}
-inline v_int8x32 v_mul_wrap(const v_int8x32& a, const v_int8x32& b)
-{
-    return v_reinterpret_as_s8(v_mul_wrap(v_reinterpret_as_u8(a), v_reinterpret_as_u8(b)));
-}
-
-//  Multiply and expand
-inline void v_mul_expand(const v_uint8x32& a, const v_uint8x32& b,
-                         v_uint16x16& c, v_uint16x16& d)
-{
-    v_uint16x16 a0, a1, b0, b1;
-    v_expand(a, a0, a1);
-    v_expand(b, b0, b1);
-    c = v_mul_wrap(a0, b0);
-    d = v_mul_wrap(a1, b1);
-}
-
-inline void v_mul_expand(const v_int8x32& a, const v_int8x32& b,
-                         v_int16x16& c, v_int16x16& d)
-{
-    v_int16x16 a0, a1, b0, b1;
-    v_expand(a, a0, a1);
-    v_expand(b, b0, b1);
-    c = v_mul_wrap(a0, b0);
-    d = v_mul_wrap(a1, b1);
-}
-
-inline void v_mul_expand(const v_int16x16& a, const v_int16x16& b,
-                         v_int32x8& c, v_int32x8& d)
-{
-    v_int16x16 vhi = v_int16x16(_mm256_mulhi_epi16(a.val, b.val));
-
-    v_int16x16 v0, v1;
-    v_zip(v_mul_wrap(a, b), vhi, v0, v1);
-
-    c = v_reinterpret_as_s32(v0);
-    d = v_reinterpret_as_s32(v1);
-}
-
-inline void v_mul_expand(const v_uint16x16& a, const v_uint16x16& b,
-                         v_uint32x8& c, v_uint32x8& d)
-{
-    v_uint16x16 vhi = v_uint16x16(_mm256_mulhi_epu16(a.val, b.val));
-
-    v_uint16x16 v0, v1;
-    v_zip(v_mul_wrap(a, b), vhi, v0, v1);
-
-    c = v_reinterpret_as_u32(v0);
-    d = v_reinterpret_as_u32(v1);
-}
-
-inline void v_mul_expand(const v_uint32x8& a, const v_uint32x8& b,
-                         v_uint64x4& c, v_uint64x4& d)
-{
-    __m256i v0 = _mm256_mul_epu32(a.val, b.val);
-    __m256i v1 = _mm256_mul_epu32(_mm256_srli_epi64(a.val, 32), _mm256_srli_epi64(b.val, 32));
-    v_zip(v_uint64x4(v0), v_uint64x4(v1), c, d);
-}
-
-inline v_int16x16 v_mul_hi(const v_int16x16& a, const v_int16x16& b) { return v_int16x16(_mm256_mulhi_epi16(a.val, b.val)); }
-inline v_uint16x16 v_mul_hi(const v_uint16x16& a, const v_uint16x16& b) { return v_uint16x16(_mm256_mulhi_epu16(a.val, b.val)); }
-
-/** Bitwise shifts **/
-#define OPENCV_HAL_IMPL_AVX_SHIFT_OP(_Tpuvec, _Tpsvec, suffix, srai)  \
-    inline _Tpuvec operator << (const _Tpuvec& a, int imm)            \
-    { return _Tpuvec(_mm256_slli_##suffix(a.val, imm)); }             \
-    inline _Tpsvec operator << (const _Tpsvec& a, int imm)            \
-    { return _Tpsvec(_mm256_slli_##suffix(a.val, imm)); }             \
-    inline _Tpuvec operator >> (const _Tpuvec& a, int imm)            \
-    { return _Tpuvec(_mm256_srli_##suffix(a.val, imm)); }             \
-    inline _Tpsvec operator >> (const _Tpsvec& a, int imm)            \
-    { return _Tpsvec(srai(a.val, imm)); }                             \
-    template<int imm>                                                 \
-    inline _Tpuvec v_shl(const _Tpuvec& a)                            \
-    { return _Tpuvec(_mm256_slli_##suffix(a.val, imm)); }             \
-    template<int imm>                                                 \
-    inline _Tpsvec v_shl(const _Tpsvec& a)                            \
-    { return _Tpsvec(_mm256_slli_##suffix(a.val, imm)); }             \
-    template<int imm>                                                 \
-    inline _Tpuvec v_shr(const _Tpuvec& a)                            \
-    { return _Tpuvec(_mm256_srli_##suffix(a.val, imm)); }             \
-    template<int imm>                                                 \
-    inline _Tpsvec v_shr(const _Tpsvec& a)                            \
-    { return _Tpsvec(srai(a.val, imm)); }
-
-OPENCV_HAL_IMPL_AVX_SHIFT_OP(v_uint16x16, v_int16x16, epi16, _mm256_srai_epi16)
-OPENCV_HAL_IMPL_AVX_SHIFT_OP(v_uint32x8,  v_int32x8,  epi32, _mm256_srai_epi32)
-
-inline __m256i _mm256_srai_epi64xx(const __m256i a, int imm)
-{
-    __m256i d = _mm256_set1_epi64x((int64)1 << 63);
-    __m256i r = _mm256_srli_epi64(_mm256_add_epi64(a, d), imm);
-    return _mm256_sub_epi64(r, _mm256_srli_epi64(d, imm));
-}
-OPENCV_HAL_IMPL_AVX_SHIFT_OP(v_uint64x4,  v_int64x4,  epi64, _mm256_srai_epi64xx)
-
-
-/** Bitwise logic **/
-#define OPENCV_HAL_IMPL_AVX_LOGIC_OP(_Tpvec, suffix, not_const)  \
-    OPENCV_HAL_IMPL_AVX_BIN_OP(&, _Tpvec, _mm256_and_##suffix)   \
-    OPENCV_HAL_IMPL_AVX_BIN_OP(|, _Tpvec, _mm256_or_##suffix)    \
-    OPENCV_HAL_IMPL_AVX_BIN_OP(^, _Tpvec, _mm256_xor_##suffix)   \
-    inline _Tpvec operator ~ (const _Tpvec& a)                   \
-    { return _Tpvec(_mm256_xor_##suffix(a.val, not_const)); }
-
-OPENCV_HAL_IMPL_AVX_LOGIC_OP(v_uint8x32,   si256, _mm256_set1_epi32(-1))
-OPENCV_HAL_IMPL_AVX_LOGIC_OP(v_int8x32,    si256, _mm256_set1_epi32(-1))
-OPENCV_HAL_IMPL_AVX_LOGIC_OP(v_uint16x16,  si256, _mm256_set1_epi32(-1))
-OPENCV_HAL_IMPL_AVX_LOGIC_OP(v_int16x16,   si256, _mm256_set1_epi32(-1))
-OPENCV_HAL_IMPL_AVX_LOGIC_OP(v_uint32x8,   si256, _mm256_set1_epi32(-1))
-OPENCV_HAL_IMPL_AVX_LOGIC_OP(v_int32x8,    si256, _mm256_set1_epi32(-1))
-OPENCV_HAL_IMPL_AVX_LOGIC_OP(v_uint64x4,   si256, _mm256_set1_epi64x(-1))
-OPENCV_HAL_IMPL_AVX_LOGIC_OP(v_int64x4,    si256, _mm256_set1_epi64x(-1))
-OPENCV_HAL_IMPL_AVX_LOGIC_OP(v_float32x8,  ps,    _mm256_castsi256_ps(_mm256_set1_epi32(-1)))
-OPENCV_HAL_IMPL_AVX_LOGIC_OP(v_float64x4,  pd,    _mm256_castsi256_pd(_mm256_set1_epi32(-1)))
-
-/** Select **/
-#define OPENCV_HAL_IMPL_AVX_SELECT(_Tpvec, suffix)                               \
-    inline _Tpvec v_select(const _Tpvec& mask, const _Tpvec& a, const _Tpvec& b) \
-    { return _Tpvec(_mm256_blendv_##suffix(b.val, a.val, mask.val)); }
-
-OPENCV_HAL_IMPL_AVX_SELECT(v_uint8x32,  epi8)
-OPENCV_HAL_IMPL_AVX_SELECT(v_int8x32,   epi8)
-OPENCV_HAL_IMPL_AVX_SELECT(v_uint16x16, epi8)
-OPENCV_HAL_IMPL_AVX_SELECT(v_int16x16,  epi8)
-OPENCV_HAL_IMPL_AVX_SELECT(v_uint32x8,  epi8)
-OPENCV_HAL_IMPL_AVX_SELECT(v_int32x8,   epi8)
-OPENCV_HAL_IMPL_AVX_SELECT(v_float32x8, ps)
-OPENCV_HAL_IMPL_AVX_SELECT(v_float64x4, pd)
-
-/** Comparison **/
-#define OPENCV_HAL_IMPL_AVX_CMP_OP_OV(_Tpvec)                     \
-    inline _Tpvec operator != (const _Tpvec& a, const _Tpvec& b)  \
-    { return ~(a == b); }                                         \
-    inline _Tpvec operator <  (const _Tpvec& a, const _Tpvec& b)  \
-    { return b > a; }                                             \
-    inline _Tpvec operator >= (const _Tpvec& a, const _Tpvec& b)  \
-    { return ~(a < b); }                                          \
-    inline _Tpvec operator <= (const _Tpvec& a, const _Tpvec& b)  \
-    { return b >= a; }
-
-#define OPENCV_HAL_IMPL_AVX_CMP_OP_INT(_Tpuvec, _Tpsvec, suffix, sbit)   \
-    inline _Tpuvec operator == (const _Tpuvec& a, const _Tpuvec& b)      \
-    { return _Tpuvec(_mm256_cmpeq_##suffix(a.val, b.val)); }             \
-    inline _Tpuvec operator > (const _Tpuvec& a, const _Tpuvec& b)       \
-    {                                                                    \
-        __m256i smask = _mm256_set1_##suffix(sbit);                      \
-        return _Tpuvec(_mm256_cmpgt_##suffix(                            \
-                       _mm256_xor_si256(a.val, smask),                   \
-                       _mm256_xor_si256(b.val, smask)));                 \
-    }                                                                    \
-    inline _Tpsvec operator == (const _Tpsvec& a, const _Tpsvec& b)      \
-    { return _Tpsvec(_mm256_cmpeq_##suffix(a.val, b.val)); }             \
-    inline _Tpsvec operator > (const _Tpsvec& a, const _Tpsvec& b)       \
-    { return _Tpsvec(_mm256_cmpgt_##suffix(a.val, b.val)); }             \
-    OPENCV_HAL_IMPL_AVX_CMP_OP_OV(_Tpuvec)                               \
-    OPENCV_HAL_IMPL_AVX_CMP_OP_OV(_Tpsvec)
-
-OPENCV_HAL_IMPL_AVX_CMP_OP_INT(v_uint8x32,  v_int8x32,  epi8,  (char)-128)
-OPENCV_HAL_IMPL_AVX_CMP_OP_INT(v_uint16x16, v_int16x16, epi16, (short)-32768)
-OPENCV_HAL_IMPL_AVX_CMP_OP_INT(v_uint32x8,  v_int32x8,  epi32, (int)0x80000000)
-
-#define OPENCV_HAL_IMPL_AVX_CMP_OP_64BIT(_Tpvec)                 \
-    inline _Tpvec operator == (const _Tpvec& a, const _Tpvec& b) \
-    { return _Tpvec(_mm256_cmpeq_epi64(a.val, b.val)); }         \
-    inline _Tpvec operator != (const _Tpvec& a, const _Tpvec& b) \
-    { return ~(a == b); }
-
-OPENCV_HAL_IMPL_AVX_CMP_OP_64BIT(v_uint64x4)
-OPENCV_HAL_IMPL_AVX_CMP_OP_64BIT(v_int64x4)
-
-#define OPENCV_HAL_IMPL_AVX_CMP_FLT(bin_op, imm8, _Tpvec, suffix)    \
-    inline _Tpvec operator bin_op (const _Tpvec& a, const _Tpvec& b) \
-    { return _Tpvec(_mm256_cmp_##suffix(a.val, b.val, imm8)); }
-
-#define OPENCV_HAL_IMPL_AVX_CMP_OP_FLT(_Tpvec, suffix)               \
-    OPENCV_HAL_IMPL_AVX_CMP_FLT(==, _CMP_EQ_OQ,  _Tpvec, suffix)     \
-    OPENCV_HAL_IMPL_AVX_CMP_FLT(!=, _CMP_NEQ_OQ, _Tpvec, suffix)     \
-    OPENCV_HAL_IMPL_AVX_CMP_FLT(<,  _CMP_LT_OQ,  _Tpvec, suffix)     \
-    OPENCV_HAL_IMPL_AVX_CMP_FLT(>,  _CMP_GT_OQ,  _Tpvec, suffix)     \
-    OPENCV_HAL_IMPL_AVX_CMP_FLT(<=, _CMP_LE_OQ,  _Tpvec, suffix)     \
-    OPENCV_HAL_IMPL_AVX_CMP_FLT(>=, _CMP_GE_OQ,  _Tpvec, suffix)
-
-OPENCV_HAL_IMPL_AVX_CMP_OP_FLT(v_float32x8, ps)
-OPENCV_HAL_IMPL_AVX_CMP_OP_FLT(v_float64x4, pd)
-
-inline v_float32x8 v_not_nan(const v_float32x8& a)
-{ return v_float32x8(_mm256_cmp_ps(a.val, a.val, _CMP_ORD_Q)); }
-inline v_float64x4 v_not_nan(const v_float64x4& a)
-{ return v_float64x4(_mm256_cmp_pd(a.val, a.val, _CMP_ORD_Q)); }
-
-/** min/max **/
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_min, v_uint8x32,  _mm256_min_epu8)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_max, v_uint8x32,  _mm256_max_epu8)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_min, v_int8x32,   _mm256_min_epi8)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_max, v_int8x32,   _mm256_max_epi8)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_min, v_uint16x16, _mm256_min_epu16)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_max, v_uint16x16, _mm256_max_epu16)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_min, v_int16x16,  _mm256_min_epi16)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_max, v_int16x16,  _mm256_max_epi16)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_min, v_uint32x8,  _mm256_min_epu32)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_max, v_uint32x8,  _mm256_max_epu32)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_min, v_int32x8,   _mm256_min_epi32)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_max, v_int32x8,   _mm256_max_epi32)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_min, v_float32x8, _mm256_min_ps)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_max, v_float32x8, _mm256_max_ps)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_min, v_float64x4, _mm256_min_pd)
-OPENCV_HAL_IMPL_AVX_BIN_FUNC(v_max, v_float64x4, _mm256_max_pd)
-
-/** Rotate **/
-template<int imm>
-inline v_uint8x32 v_rotate_left(const v_uint8x32& a, const v_uint8x32& b)
-{
-    enum {IMM_R = (16 - imm) & 0xFF};
-    enum {IMM_R2 = (32 - imm) & 0xFF};
-
-    if (imm == 0)  return a;
-    if (imm == 32) return b;
-    if (imm > 32)  return v_uint8x32();
-
-    __m256i swap = _mm256_permute2x128_si256(a.val, b.val, 0x03);
-    if (imm == 16) return v_uint8x32(swap);
-    if (imm < 16)  return v_uint8x32(_mm256_alignr_epi8(a.val, swap, IMM_R));
-    return v_uint8x32(_mm256_alignr_epi8(swap, b.val, IMM_R2)); // imm < 32
-}
-
-template<int imm>
-inline v_uint8x32 v_rotate_right(const v_uint8x32& a, const v_uint8x32& b)
-{
-    enum {IMM_L = (imm - 16) & 0xFF};
-
-    if (imm == 0)  return a;
-    if (imm == 32) return b;
-    if (imm > 32)  return v_uint8x32();
-
-    __m256i swap = _mm256_permute2x128_si256(a.val, b.val, 0x21);
-    if (imm == 16) return v_uint8x32(swap);
-    if (imm < 16)  return v_uint8x32(_mm256_alignr_epi8(swap, a.val, imm));
-    return v_uint8x32(_mm256_alignr_epi8(b.val, swap, IMM_L));
-}
-
-template<int imm>
-inline v_uint8x32 v_rotate_left(const v_uint8x32& a)
-{
-    enum {IMM_L = (imm - 16) & 0xFF};
-    enum {IMM_R = (16 - imm) & 0xFF};
-
-    if (imm == 0) return a;
-    if (imm > 32) return v_uint8x32();
-
-    // ESAC control[3] ? [127:0] = 0
-    __m256i swapz = _mm256_permute2x128_si256(a.val, a.val, _MM_SHUFFLE(0, 0, 2, 0));
-    if (imm == 16) return v_uint8x32(swapz);
-    if (imm < 16)  return v_uint8x32(_mm256_alignr_epi8(a.val, swapz, IMM_R));
-    return v_uint8x32(_mm256_slli_si256(swapz, IMM_L));
-}
-
-template<int imm>
-inline v_uint8x32 v_rotate_right(const v_uint8x32& a)
-{
-    enum {IMM_L = (imm - 16) & 0xFF};
-
-    if (imm == 0) return a;
-    if (imm > 32) return v_uint8x32();
-
-    // ESAC control[3] ? [127:0] = 0
-    __m256i swapz = _mm256_permute2x128_si256(a.val, a.val, _MM_SHUFFLE(2, 0, 0, 1));
-    if (imm == 16) return v_uint8x32(swapz);
-    if (imm < 16)  return v_uint8x32(_mm256_alignr_epi8(swapz, a.val, imm));
-    return v_uint8x32(_mm256_srli_si256(swapz, IMM_L));
-}
-
-#define OPENCV_HAL_IMPL_AVX_ROTATE_CAST(intrin, _Tpvec, cast)     \
-    template<int imm>                                             \
-    inline _Tpvec intrin(const _Tpvec& a, const _Tpvec& b)        \
-    {                                                             \
-        enum {IMMxW = imm * sizeof(typename _Tpvec::lane_type)};  \
-        v_uint8x32 ret = intrin<IMMxW>(v_reinterpret_as_u8(a),    \
-                                       v_reinterpret_as_u8(b));   \
-        return _Tpvec(cast(ret.val));                             \
-    }                                                             \
-    template<int imm>                                             \
-    inline _Tpvec intrin(const _Tpvec& a)                         \
-    {                                                             \
-        enum {IMMxW = imm * sizeof(typename _Tpvec::lane_type)};  \
-        v_uint8x32 ret = intrin<IMMxW>(v_reinterpret_as_u8(a));   \
-        return _Tpvec(cast(ret.val));                             \
-    }
-
-#define OPENCV_HAL_IMPL_AVX_ROTATE(_Tpvec)                                  \
-    OPENCV_HAL_IMPL_AVX_ROTATE_CAST(v_rotate_left,  _Tpvec, OPENCV_HAL_NOP) \
-    OPENCV_HAL_IMPL_AVX_ROTATE_CAST(v_rotate_right, _Tpvec, OPENCV_HAL_NOP)
-
-OPENCV_HAL_IMPL_AVX_ROTATE(v_int8x32)
-OPENCV_HAL_IMPL_AVX_ROTATE(v_uint16x16)
-OPENCV_HAL_IMPL_AVX_ROTATE(v_int16x16)
-OPENCV_HAL_IMPL_AVX_ROTATE(v_uint32x8)
-OPENCV_HAL_IMPL_AVX_ROTATE(v_int32x8)
-OPENCV_HAL_IMPL_AVX_ROTATE(v_uint64x4)
-OPENCV_HAL_IMPL_AVX_ROTATE(v_int64x4)
-
-OPENCV_HAL_IMPL_AVX_ROTATE_CAST(v_rotate_left,  v_float32x8, _mm256_castsi256_ps)
-OPENCV_HAL_IMPL_AVX_ROTATE_CAST(v_rotate_right, v_float32x8, _mm256_castsi256_ps)
-OPENCV_HAL_IMPL_AVX_ROTATE_CAST(v_rotate_left,  v_float64x4, _mm256_castsi256_pd)
-OPENCV_HAL_IMPL_AVX_ROTATE_CAST(v_rotate_right, v_float64x4, _mm256_castsi256_pd)
-
-/** Reverse **/
-inline v_uint8x32 v_reverse(const v_uint8x32 &a)
-{
-    static const __m256i perm = _mm256_setr_epi8(
-            15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0,
-            15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
-    __m256i vec = _mm256_shuffle_epi8(a.val, perm);
-    return v_uint8x32(_mm256_permute2x128_si256(vec, vec, 1));
-}
-
-inline v_int8x32 v_reverse(const v_int8x32 &a)
-{ return v_reinterpret_as_s8(v_reverse(v_reinterpret_as_u8(a))); }
-
-inline v_uint16x16 v_reverse(const v_uint16x16 &a)
-{
-    static const __m256i perm = _mm256_setr_epi8(
-            14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1,
-            14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1);
-    __m256i vec = _mm256_shuffle_epi8(a.val, perm);
-    return v_uint16x16(_mm256_permute2x128_si256(vec, vec, 1));
-}
-
-inline v_int16x16 v_reverse(const v_int16x16 &a)
-{ return v_reinterpret_as_s16(v_reverse(v_reinterpret_as_u16(a))); }
-
-inline v_uint32x8 v_reverse(const v_uint32x8 &a)
-{
-    static const __m256i perm = _mm256_setr_epi32(7, 6, 5, 4, 3, 2, 1, 0);
-    return v_uint32x8(_mm256_permutevar8x32_epi32(a.val, perm));
-}
-
-inline v_int32x8 v_reverse(const v_int32x8 &a)
-{ return v_reinterpret_as_s32(v_reverse(v_reinterpret_as_u32(a))); }
-
-inline v_float32x8 v_reverse(const v_float32x8 &a)
-{ return v_reinterpret_as_f32(v_reverse(v_reinterpret_as_u32(a))); }
-
-inline v_uint64x4 v_reverse(const v_uint64x4 &a)
-{
-    return v_uint64x4(_mm256_permute4x64_epi64(a.val, _MM_SHUFFLE(0, 1, 2, 3)));
-}
-
-inline v_int64x4 v_reverse(const v_int64x4 &a)
-{ return v_reinterpret_as_s64(v_reverse(v_reinterpret_as_u64(a))); }
-
-inline v_float64x4 v_reverse(const v_float64x4 &a)
-{ return v_reinterpret_as_f64(v_reverse(v_reinterpret_as_u64(a))); }
-
-////////// Reduce and mask /////////
-
-/** Reduce **/
-inline unsigned v_reduce_sum(const v_uint8x32& a)
-{
-    __m256i half = _mm256_sad_epu8(a.val, _mm256_setzero_si256());
-    __m128i quarter = _mm_add_epi32(_v256_extract_low(half), _v256_extract_high(half));
-    return (unsigned)_mm_cvtsi128_si32(_mm_add_epi32(quarter, _mm_unpackhi_epi64(quarter, quarter)));
-}
-inline int v_reduce_sum(const v_int8x32& a)
-{
-    __m256i half = _mm256_sad_epu8(_mm256_xor_si256(a.val, _mm256_set1_epi8((schar)-128)), _mm256_setzero_si256());
-    __m128i quarter = _mm_add_epi32(_v256_extract_low(half), _v256_extract_high(half));
-    return (unsigned)_mm_cvtsi128_si32(_mm_add_epi32(quarter, _mm_unpackhi_epi64(quarter, quarter))) - 4096;
-}
-#define OPENCV_HAL_IMPL_AVX_REDUCE_32(_Tpvec, sctype, func, intrin) \
-    inline sctype v_reduce_##func(const _Tpvec& a) \
-    { \
-        __m128i val = intrin(_v256_extract_low(a.val), _v256_extract_high(a.val)); \
-        val = intrin(val, _mm_srli_si128(val,8)); \
-        val = intrin(val, _mm_srli_si128(val,4)); \
-        val = intrin(val, _mm_srli_si128(val,2)); \
-        val = intrin(val, _mm_srli_si128(val,1)); \
-        return (sctype)_mm_cvtsi128_si32(val); \
-    }
-
-OPENCV_HAL_IMPL_AVX_REDUCE_32(v_uint8x32, uchar, min, _mm_min_epu8)
-OPENCV_HAL_IMPL_AVX_REDUCE_32(v_int8x32,  schar, min, _mm_min_epi8)
-OPENCV_HAL_IMPL_AVX_REDUCE_32(v_uint8x32, uchar, max, _mm_max_epu8)
-OPENCV_HAL_IMPL_AVX_REDUCE_32(v_int8x32,  schar, max, _mm_max_epi8)
-
-#define OPENCV_HAL_IMPL_AVX_REDUCE_16(_Tpvec, sctype, func, intrin) \
-    inline sctype v_reduce_##func(const _Tpvec& a)                  \
-    {                                                               \
-        __m128i v0 = _v256_extract_low(a.val);                      \
-        __m128i v1 = _v256_extract_high(a.val);                     \
-        v0 = intrin(v0, v1);                                        \
-        v0 = intrin(v0, _mm_srli_si128(v0, 8));                     \
-        v0 = intrin(v0, _mm_srli_si128(v0, 4));                     \
-        v0 = intrin(v0, _mm_srli_si128(v0, 2));                     \
-        return (sctype) _mm_cvtsi128_si32(v0);                      \
-    }
-
-OPENCV_HAL_IMPL_AVX_REDUCE_16(v_uint16x16, ushort, min, _mm_min_epu16)
-OPENCV_HAL_IMPL_AVX_REDUCE_16(v_int16x16,  short,  min, _mm_min_epi16)
-OPENCV_HAL_IMPL_AVX_REDUCE_16(v_uint16x16, ushort, max, _mm_max_epu16)
-OPENCV_HAL_IMPL_AVX_REDUCE_16(v_int16x16,  short,  max, _mm_max_epi16)
-
-#define OPENCV_HAL_IMPL_AVX_REDUCE_8(_Tpvec, sctype, func, intrin) \
-    inline sctype v_reduce_##func(const _Tpvec& a)                 \
-    {                                                              \
-        __m128i v0 = _v256_extract_low(a.val);                     \
-        __m128i v1 = _v256_extract_high(a.val);                    \
-        v0 = intrin(v0, v1);                                       \
-        v0 = intrin(v0, _mm_srli_si128(v0, 8));                    \
-        v0 = intrin(v0, _mm_srli_si128(v0, 4));                    \
-        return (sctype) _mm_cvtsi128_si32(v0);                     \
-    }
-
-OPENCV_HAL_IMPL_AVX_REDUCE_8(v_uint32x8, unsigned, min, _mm_min_epu32)
-OPENCV_HAL_IMPL_AVX_REDUCE_8(v_int32x8,  int,      min, _mm_min_epi32)
-OPENCV_HAL_IMPL_AVX_REDUCE_8(v_uint32x8, unsigned, max, _mm_max_epu32)
-OPENCV_HAL_IMPL_AVX_REDUCE_8(v_int32x8,  int,      max, _mm_max_epi32)
-
-#define OPENCV_HAL_IMPL_AVX_REDUCE_FLT(func, intrin)                  \
-    inline float v_reduce_##func(const v_float32x8& a)                \
-    {                                                                 \
-        __m128 v0 = _v256_extract_low(a.val);                         \
-        __m128 v1 = _v256_extract_high(a.val);                        \
-        v0 = intrin(v0, v1);                                          \
-        v0 = intrin(v0, _mm_permute_ps(v0, _MM_SHUFFLE(0, 0, 3, 2))); \
-        v0 = intrin(v0, _mm_permute_ps(v0, _MM_SHUFFLE(0, 0, 0, 1))); \
-        return _mm_cvtss_f32(v0);                                     \
-    }
-
-OPENCV_HAL_IMPL_AVX_REDUCE_FLT(min, _mm_min_ps)
-OPENCV_HAL_IMPL_AVX_REDUCE_FLT(max, _mm_max_ps)
-
-inline int v_reduce_sum(const v_int32x8& a)
-{
-    __m256i s0 = _mm256_hadd_epi32(a.val, a.val);
-            s0 = _mm256_hadd_epi32(s0, s0);
-
-    __m128i s1 = _v256_extract_high(s0);
-            s1 = _mm_add_epi32(_v256_extract_low(s0), s1);
-
-    return _mm_cvtsi128_si32(s1);
-}
-
-inline unsigned v_reduce_sum(const v_uint32x8& a)
-{ return v_reduce_sum(v_reinterpret_as_s32(a)); }
-
-inline int v_reduce_sum(const v_int16x16& a)
-{ return v_reduce_sum(v_expand_low(a) + v_expand_high(a)); }
-inline unsigned v_reduce_sum(const v_uint16x16& a)
-{ return v_reduce_sum(v_expand_low(a) + v_expand_high(a)); }
-
-inline float v_reduce_sum(const v_float32x8& a)
-{
-    __m256 s0 = _mm256_hadd_ps(a.val, a.val);
-           s0 = _mm256_hadd_ps(s0, s0);
-
-    __m128 s1 = _v256_extract_high(s0);
-           s1 = _mm_add_ps(_v256_extract_low(s0), s1);
-
-    return _mm_cvtss_f32(s1);
-}
-
-inline uint64 v_reduce_sum(const v_uint64x4& a)
-{
-    uint64 CV_DECL_ALIGNED(32) idx[2];
-    _mm_store_si128((__m128i*)idx, _mm_add_epi64(_v256_extract_low(a.val), _v256_extract_high(a.val)));
-    return idx[0] + idx[1];
-}
-inline int64 v_reduce_sum(const v_int64x4& a)
-{
-    int64 CV_DECL_ALIGNED(32) idx[2];
-    _mm_store_si128((__m128i*)idx, _mm_add_epi64(_v256_extract_low(a.val), _v256_extract_high(a.val)));
-    return idx[0] + idx[1];
-}
-inline double v_reduce_sum(const v_float64x4& a)
-{
-    __m256d s0 = _mm256_hadd_pd(a.val, a.val);
-    return _mm_cvtsd_f64(_mm_add_pd(_v256_extract_low(s0), _v256_extract_high(s0)));
-}
-
-inline v_float32x8 v_reduce_sum4(const v_float32x8& a, const v_float32x8& b,
-                                 const v_float32x8& c, const v_float32x8& d)
-{
-    __m256 ab = _mm256_hadd_ps(a.val, b.val);
-    __m256 cd = _mm256_hadd_ps(c.val, d.val);
-    return v_float32x8(_mm256_hadd_ps(ab, cd));
-}
-
-inline unsigned v_reduce_sad(const v_uint8x32& a, const v_uint8x32& b)
-{
-    __m256i half = _mm256_sad_epu8(a.val, b.val);
-    __m128i quarter = _mm_add_epi32(_v256_extract_low(half), _v256_extract_high(half));
-    return (unsigned)_mm_cvtsi128_si32(_mm_add_epi32(quarter, _mm_unpackhi_epi64(quarter, quarter)));
-}
-inline unsigned v_reduce_sad(const v_int8x32& a, const v_int8x32& b)
-{
-    __m256i half = _mm256_set1_epi8(0x7f);
-    half = _mm256_sad_epu8(_mm256_add_epi8(a.val, half), _mm256_add_epi8(b.val, half));
-    __m128i quarter = _mm_add_epi32(_v256_extract_low(half), _v256_extract_high(half));
-    return (unsigned)_mm_cvtsi128_si32(_mm_add_epi32(quarter, _mm_unpackhi_epi64(quarter, quarter)));
-}
-inline unsigned v_reduce_sad(const v_uint16x16& a, const v_uint16x16& b)
-{
-    v_uint32x8 l, h;
-    v_expand(v_add_wrap(a - b, b - a), l, h);
-    return v_reduce_sum(l + h);
-}
-inline unsigned v_reduce_sad(const v_int16x16& a, const v_int16x16& b)
-{
-    v_uint32x8 l, h;
-    v_expand(v_reinterpret_as_u16(v_sub_wrap(v_max(a, b), v_min(a, b))), l, h);
-    return v_reduce_sum(l + h);
-}
-inline unsigned v_reduce_sad(const v_uint32x8& a, const v_uint32x8& b)
-{
-    return v_reduce_sum(v_max(a, b) - v_min(a, b));
-}
-inline unsigned v_reduce_sad(const v_int32x8& a, const v_int32x8& b)
-{
-    v_int32x8 m = a < b;
-    return v_reduce_sum(v_reinterpret_as_u32(((a - b) ^ m) - m));
-}
-inline float v_reduce_sad(const v_float32x8& a, const v_float32x8& b)
-{
-    return v_reduce_sum((a - b) & v_float32x8(_mm256_castsi256_ps(_mm256_set1_epi32(0x7fffffff))));
-}
-
-/** Popcount **/
-inline v_uint8x32 v_popcount(const v_uint8x32& a)
-{
-    __m256i _popcnt_table = _mm256_setr_epi8(0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4,
-                                             0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);
-    __m256i _popcnt_mask = _mm256_set1_epi8(0x0F);
-    return v_uint8x32(_mm256_add_epi8(_mm256_shuffle_epi8(_popcnt_table, _mm256_and_si256(                  a.val    , _popcnt_mask)),
-                                      _mm256_shuffle_epi8(_popcnt_table, _mm256_and_si256(_mm256_srli_epi16(a.val, 4), _popcnt_mask))));
-}
-inline v_uint16x16 v_popcount(const v_uint16x16& a)
-{
-    v_uint8x32 p = v_popcount(v_reinterpret_as_u8(a));
-    p += v_rotate_right<1>(p);
-    return v_reinterpret_as_u16(p) & v256_setall_u16(0x00ff);
-}
-inline v_uint32x8 v_popcount(const v_uint32x8& a)
-{
-    v_uint8x32 p = v_popcount(v_reinterpret_as_u8(a));
-    p += v_rotate_right<1>(p);
-    p += v_rotate_right<2>(p);
-    return v_reinterpret_as_u32(p) & v256_setall_u32(0x000000ff);
-}
-inline v_uint64x4 v_popcount(const v_uint64x4& a)
-{
-    return v_uint64x4(_mm256_sad_epu8(v_popcount(v_reinterpret_as_u8(a)).val, _mm256_setzero_si256()));
-}
-inline v_uint8x32 v_popcount(const v_int8x32& a)
-{ return v_popcount(v_reinterpret_as_u8(a)); }
-inline v_uint16x16 v_popcount(const v_int16x16& a)
-{ return v_popcount(v_reinterpret_as_u16(a)); }
-inline v_uint32x8 v_popcount(const v_int32x8& a)
-{ return v_popcount(v_reinterpret_as_u32(a)); }
-inline v_uint64x4 v_popcount(const v_int64x4& a)
-{ return v_popcount(v_reinterpret_as_u64(a)); }
-
-/** Mask **/
-inline int v_signmask(const v_int8x32& a)
-{ return _mm256_movemask_epi8(a.val); }
-inline int v_signmask(const v_uint8x32& a)
-{ return v_signmask(v_reinterpret_as_s8(a)); }
-
-inline int v_signmask(const v_int16x16& a)
-{ return v_signmask(v_pack(a, a)) & 0xFFFF; }
-inline int v_signmask(const v_uint16x16& a)
-{ return v_signmask(v_reinterpret_as_s16(a)); }
-
-inline int v_signmask(const v_float32x8& a)
-{ return _mm256_movemask_ps(a.val); }
-inline int v_signmask(const v_float64x4& a)
-{ return _mm256_movemask_pd(a.val); }
-
-inline int v_signmask(const v_int32x8& a)
-{ return v_signmask(v_reinterpret_as_f32(a)); }
-inline int v_signmask(const v_uint32x8& a)
-{ return v_signmask(v_reinterpret_as_f32(a)); }
-
-inline int v_signmask(const v_int64x4& a)
-{ return v_signmask(v_reinterpret_as_f64(a)); }
-inline int v_signmask(const v_uint64x4& a)
-{ return v_signmask(v_reinterpret_as_f64(a)); }
-
-inline int v_scan_forward(const v_int8x32& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))); }
-inline int v_scan_forward(const v_uint8x32& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))); }
-inline int v_scan_forward(const v_int16x16& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 2; }
-inline int v_scan_forward(const v_uint16x16& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 2; }
-inline int v_scan_forward(const v_int32x8& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 4; }
-inline int v_scan_forward(const v_uint32x8& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 4; }
-inline int v_scan_forward(const v_float32x8& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 4; }
-inline int v_scan_forward(const v_int64x4& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 8; }
-inline int v_scan_forward(const v_uint64x4& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 8; }
-inline int v_scan_forward(const v_float64x4& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 8; }
-
-/** Checks **/
-#define OPENCV_HAL_IMPL_AVX_CHECK(_Tpvec, allmask) \
-    inline bool v_check_all(const _Tpvec& a) { return v_signmask(a) == allmask; } \
-    inline bool v_check_any(const _Tpvec& a) { return v_signmask(a) != 0; }
-OPENCV_HAL_IMPL_AVX_CHECK(v_uint8x32, -1)
-OPENCV_HAL_IMPL_AVX_CHECK(v_int8x32, -1)
-OPENCV_HAL_IMPL_AVX_CHECK(v_uint32x8, 255)
-OPENCV_HAL_IMPL_AVX_CHECK(v_int32x8, 255)
-OPENCV_HAL_IMPL_AVX_CHECK(v_uint64x4, 15)
-OPENCV_HAL_IMPL_AVX_CHECK(v_int64x4, 15)
-OPENCV_HAL_IMPL_AVX_CHECK(v_float32x8, 255)
-OPENCV_HAL_IMPL_AVX_CHECK(v_float64x4, 15)
-
-#define OPENCV_HAL_IMPL_AVX_CHECK_SHORT(_Tpvec)  \
-    inline bool v_check_all(const _Tpvec& a) { return (v_signmask(v_reinterpret_as_s8(a)) & 0xaaaaaaaa) == 0xaaaaaaaa; } \
-    inline bool v_check_any(const _Tpvec& a) { return (v_signmask(v_reinterpret_as_s8(a)) & 0xaaaaaaaa) != 0; }
-OPENCV_HAL_IMPL_AVX_CHECK_SHORT(v_uint16x16)
-OPENCV_HAL_IMPL_AVX_CHECK_SHORT(v_int16x16)
-
-////////// Other math /////////
-
-/** Some frequent operations **/
-#if CV_FMA3
-#define OPENCV_HAL_IMPL_AVX_MULADD(_Tpvec, suffix)                            \
-    inline _Tpvec v_fma(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c)    \
-    { return _Tpvec(_mm256_fmadd_##suffix(a.val, b.val, c.val)); }            \
-    inline _Tpvec v_muladd(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c) \
-    { return _Tpvec(_mm256_fmadd_##suffix(a.val, b.val, c.val)); }
-#else
-#define OPENCV_HAL_IMPL_AVX_MULADD(_Tpvec, suffix)                                    \
-    inline _Tpvec v_fma(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c)            \
-    { return _Tpvec(_mm256_add_##suffix(_mm256_mul_##suffix(a.val, b.val), c.val)); } \
-    inline _Tpvec v_muladd(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c)         \
-    { return _Tpvec(_mm256_add_##suffix(_mm256_mul_##suffix(a.val, b.val), c.val)); }
-#endif
-
-#define OPENCV_HAL_IMPL_AVX_MISC(_Tpvec, suffix)                              \
-    inline _Tpvec v_sqrt(const _Tpvec& x)                                     \
-    { return _Tpvec(_mm256_sqrt_##suffix(x.val)); }                           \
-    inline _Tpvec v_sqr_magnitude(const _Tpvec& a, const _Tpvec& b)           \
-    { return v_fma(a, a, b * b); }                                            \
-    inline _Tpvec v_magnitude(const _Tpvec& a, const _Tpvec& b)               \
-    { return v_sqrt(v_fma(a, a, b*b)); }
-
-OPENCV_HAL_IMPL_AVX_MULADD(v_float32x8, ps)
-OPENCV_HAL_IMPL_AVX_MULADD(v_float64x4, pd)
-OPENCV_HAL_IMPL_AVX_MISC(v_float32x8, ps)
-OPENCV_HAL_IMPL_AVX_MISC(v_float64x4, pd)
-
-inline v_int32x8 v_fma(const v_int32x8& a, const v_int32x8& b, const v_int32x8& c)
-{
-    return a * b + c;
-}
-
-inline v_int32x8 v_muladd(const v_int32x8& a, const v_int32x8& b, const v_int32x8& c)
-{
-    return v_fma(a, b, c);
-}
-
-inline v_float32x8 v_invsqrt(const v_float32x8& x)
-{
-    v_float32x8 half = x * v256_setall_f32(0.5);
-    v_float32x8 t  = v_float32x8(_mm256_rsqrt_ps(x.val));
-    // todo: _mm256_fnmsub_ps
-    t *= v256_setall_f32(1.5) - ((t * t) * half);
-    return t;
-}
-
-inline v_float64x4 v_invsqrt(const v_float64x4& x)
-{
-    return v256_setall_f64(1.) / v_sqrt(x);
-}
-
-/** Absolute values **/
-#define OPENCV_HAL_IMPL_AVX_ABS(_Tpvec, suffix)         \
-    inline v_u##_Tpvec v_abs(const v_##_Tpvec& x)       \
-    { return v_u##_Tpvec(_mm256_abs_##suffix(x.val)); }
-
-OPENCV_HAL_IMPL_AVX_ABS(int8x32,  epi8)
-OPENCV_HAL_IMPL_AVX_ABS(int16x16, epi16)
-OPENCV_HAL_IMPL_AVX_ABS(int32x8,  epi32)
-
-inline v_float32x8 v_abs(const v_float32x8& x)
-{ return x & v_float32x8(_mm256_castsi256_ps(_mm256_set1_epi32(0x7fffffff))); }
-inline v_float64x4 v_abs(const v_float64x4& x)
-{ return x & v_float64x4(_mm256_castsi256_pd(_mm256_srli_epi64(_mm256_set1_epi64x(-1), 1))); }
-
-/** Absolute difference **/
-inline v_uint8x32 v_absdiff(const v_uint8x32& a, const v_uint8x32& b)
-{ return v_add_wrap(a - b,  b - a); }
-inline v_uint16x16 v_absdiff(const v_uint16x16& a, const v_uint16x16& b)
-{ return v_add_wrap(a - b,  b - a); }
-inline v_uint32x8 v_absdiff(const v_uint32x8& a, const v_uint32x8& b)
-{ return v_max(a, b) - v_min(a, b); }
-
-inline v_uint8x32 v_absdiff(const v_int8x32& a, const v_int8x32& b)
-{
-    v_int8x32 d = v_sub_wrap(a, b);
-    v_int8x32 m = a < b;
-    return v_reinterpret_as_u8(v_sub_wrap(d ^ m, m));
-}
-
-inline v_uint16x16 v_absdiff(const v_int16x16& a, const v_int16x16& b)
-{ return v_reinterpret_as_u16(v_sub_wrap(v_max(a, b), v_min(a, b))); }
-
-inline v_uint32x8 v_absdiff(const v_int32x8& a, const v_int32x8& b)
-{
-    v_int32x8 d = a - b;
-    v_int32x8 m = a < b;
-    return v_reinterpret_as_u32((d ^ m) - m);
-}
-
-inline v_float32x8 v_absdiff(const v_float32x8& a, const v_float32x8& b)
-{ return v_abs(a - b); }
-
-inline v_float64x4 v_absdiff(const v_float64x4& a, const v_float64x4& b)
-{ return v_abs(a - b); }
-
-/** Saturating absolute difference **/
-inline v_int8x32 v_absdiffs(const v_int8x32& a, const v_int8x32& b)
-{
-    v_int8x32 d = a - b;
-    v_int8x32 m = a < b;
-    return (d ^ m) - m;
-}
-inline v_int16x16 v_absdiffs(const v_int16x16& a, const v_int16x16& b)
-{ return v_max(a, b) - v_min(a, b); }
-
-////////// Conversions /////////
-
-/** Rounding **/
-inline v_int32x8 v_round(const v_float32x8& a)
-{ return v_int32x8(_mm256_cvtps_epi32(a.val)); }
-
-inline v_int32x8 v_round(const v_float64x4& a)
-{ return v_int32x8(_mm256_castsi128_si256(_mm256_cvtpd_epi32(a.val))); }
-
-inline v_int32x8 v_round(const v_float64x4& a, const v_float64x4& b)
-{
-    __m128i ai = _mm256_cvtpd_epi32(a.val), bi = _mm256_cvtpd_epi32(b.val);
-    return v_int32x8(_v256_combine(ai, bi));
-}
-
-inline v_int32x8 v_trunc(const v_float32x8& a)
-{ return v_int32x8(_mm256_cvttps_epi32(a.val)); }
-
-inline v_int32x8 v_trunc(const v_float64x4& a)
-{ return v_int32x8(_mm256_castsi128_si256(_mm256_cvttpd_epi32(a.val))); }
-
-inline v_int32x8 v_floor(const v_float32x8& a)
-{ return v_int32x8(_mm256_cvttps_epi32(_mm256_floor_ps(a.val))); }
-
-inline v_int32x8 v_floor(const v_float64x4& a)
-{ return v_trunc(v_float64x4(_mm256_floor_pd(a.val))); }
-
-inline v_int32x8 v_ceil(const v_float32x8& a)
-{ return v_int32x8(_mm256_cvttps_epi32(_mm256_ceil_ps(a.val))); }
-
-inline v_int32x8 v_ceil(const v_float64x4& a)
-{ return v_trunc(v_float64x4(_mm256_ceil_pd(a.val))); }
-
-/** To float **/
-inline v_float32x8 v_cvt_f32(const v_int32x8& a)
-{ return v_float32x8(_mm256_cvtepi32_ps(a.val)); }
-
-inline v_float32x8 v_cvt_f32(const v_float64x4& a)
-{ return v_float32x8(_mm256_castps128_ps256(_mm256_cvtpd_ps(a.val))); }
-
-inline v_float32x8 v_cvt_f32(const v_float64x4& a, const v_float64x4& b)
-{
-    __m128 af = _mm256_cvtpd_ps(a.val), bf = _mm256_cvtpd_ps(b.val);
-    return v_float32x8(_v256_combine(af, bf));
-}
-
-inline v_float64x4 v_cvt_f64(const v_int32x8& a)
-{ return v_float64x4(_mm256_cvtepi32_pd(_v256_extract_low(a.val))); }
-
-inline v_float64x4 v_cvt_f64_high(const v_int32x8& a)
-{ return v_float64x4(_mm256_cvtepi32_pd(_v256_extract_high(a.val))); }
-
-inline v_float64x4 v_cvt_f64(const v_float32x8& a)
-{ return v_float64x4(_mm256_cvtps_pd(_v256_extract_low(a.val))); }
-
-inline v_float64x4 v_cvt_f64_high(const v_float32x8& a)
-{ return v_float64x4(_mm256_cvtps_pd(_v256_extract_high(a.val))); }
-
-// from (Mysticial and wim) https://stackoverflow.com/q/41144668
-inline v_float64x4 v_cvt_f64(const v_int64x4& v)
-{
-    // constants encoded as floating-point
-    __m256i magic_i_lo   = _mm256_set1_epi64x(0x4330000000000000); // 2^52
-    __m256i magic_i_hi32 = _mm256_set1_epi64x(0x4530000080000000); // 2^84 + 2^63
-    __m256i magic_i_all  = _mm256_set1_epi64x(0x4530000080100000); // 2^84 + 2^63 + 2^52
-    __m256d magic_d_all  = _mm256_castsi256_pd(magic_i_all);
-
-    // Blend the 32 lowest significant bits of v with magic_int_lo
-    __m256i v_lo         = _mm256_blend_epi32(magic_i_lo, v.val, 0x55);
-    // Extract the 32 most significant bits of v
-    __m256i v_hi         = _mm256_srli_epi64(v.val, 32);
-    // Flip the msb of v_hi and blend with 0x45300000
-            v_hi         = _mm256_xor_si256(v_hi, magic_i_hi32);
-    // Compute in double precision
-    __m256d v_hi_dbl     = _mm256_sub_pd(_mm256_castsi256_pd(v_hi), magic_d_all);
-    // (v_hi - magic_d_all) + v_lo  Do not assume associativity of floating point addition
-    __m256d result       = _mm256_add_pd(v_hi_dbl, _mm256_castsi256_pd(v_lo));
-    return v_float64x4(result);
-}
-
-////////////// Lookup table access ////////////////////
-
-inline v_int8x32 v256_lut(const schar* tab, const int* idx)
-{
-    return v_int8x32(_mm256_setr_epi8(tab[idx[ 0]], tab[idx[ 1]], tab[idx[ 2]], tab[idx[ 3]], tab[idx[ 4]], tab[idx[ 5]], tab[idx[ 6]], tab[idx[ 7]],
-                                      tab[idx[ 8]], tab[idx[ 9]], tab[idx[10]], tab[idx[11]], tab[idx[12]], tab[idx[13]], tab[idx[14]], tab[idx[15]],
-                                      tab[idx[16]], tab[idx[17]], tab[idx[18]], tab[idx[19]], tab[idx[20]], tab[idx[21]], tab[idx[22]], tab[idx[23]],
-                                      tab[idx[24]], tab[idx[25]], tab[idx[26]], tab[idx[27]], tab[idx[28]], tab[idx[29]], tab[idx[30]], tab[idx[31]]));
-}
-inline v_int8x32 v256_lut_pairs(const schar* tab, const int* idx)
-{
-    return v_int8x32(_mm256_setr_epi16(*(const short*)(tab + idx[ 0]), *(const short*)(tab + idx[ 1]), *(const short*)(tab + idx[ 2]), *(const short*)(tab + idx[ 3]),
-                                       *(const short*)(tab + idx[ 4]), *(const short*)(tab + idx[ 5]), *(const short*)(tab + idx[ 6]), *(const short*)(tab + idx[ 7]),
-                                       *(const short*)(tab + idx[ 8]), *(const short*)(tab + idx[ 9]), *(const short*)(tab + idx[10]), *(const short*)(tab + idx[11]),
-                                       *(const short*)(tab + idx[12]), *(const short*)(tab + idx[13]), *(const short*)(tab + idx[14]), *(const short*)(tab + idx[15])));
-}
-inline v_int8x32 v256_lut_quads(const schar* tab, const int* idx)
-{
-    return v_int8x32(_mm256_i32gather_epi32((const int*)tab, _mm256_loadu_si256((const __m256i*)idx), 1));
-}
-inline v_uint8x32 v256_lut(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v256_lut((const schar *)tab, idx)); }
-inline v_uint8x32 v256_lut_pairs(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v256_lut_pairs((const schar *)tab, idx)); }
-inline v_uint8x32 v256_lut_quads(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v256_lut_quads((const schar *)tab, idx)); }
-
-inline v_int16x16 v256_lut(const short* tab, const int* idx)
-{
-    return v_int16x16(_mm256_setr_epi16(tab[idx[0]], tab[idx[1]], tab[idx[ 2]], tab[idx[ 3]], tab[idx[ 4]], tab[idx[ 5]], tab[idx[ 6]], tab[idx[ 7]],
-                                        tab[idx[8]], tab[idx[9]], tab[idx[10]], tab[idx[11]], tab[idx[12]], tab[idx[13]], tab[idx[14]], tab[idx[15]]));
-}
-inline v_int16x16 v256_lut_pairs(const short* tab, const int* idx)
-{
-    return v_int16x16(_mm256_i32gather_epi32((const int*)tab, _mm256_loadu_si256((const __m256i*)idx), 2));
-}
-inline v_int16x16 v256_lut_quads(const short* tab, const int* idx)
-{
-#if defined(__GNUC__)
-    return v_int16x16(_mm256_i32gather_epi64((const long long int*)tab, _mm_loadu_si128((const __m128i*)idx), 2));//Looks like intrinsic has wrong definition
-#else
-    return v_int16x16(_mm256_i32gather_epi64((const int64*)tab, _mm_loadu_si128((const __m128i*)idx), 2));
-#endif
-}
-inline v_uint16x16 v256_lut(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v256_lut((const short *)tab, idx)); }
-inline v_uint16x16 v256_lut_pairs(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v256_lut_pairs((const short *)tab, idx)); }
-inline v_uint16x16 v256_lut_quads(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v256_lut_quads((const short *)tab, idx)); }
-
-inline v_int32x8 v256_lut(const int* tab, const int* idx)
-{
-    return v_int32x8(_mm256_i32gather_epi32(tab, _mm256_loadu_si256((const __m256i*)idx), 4));
-}
-inline v_int32x8 v256_lut_pairs(const int* tab, const int* idx)
-{
-#if defined(__GNUC__)
-    return v_int32x8(_mm256_i32gather_epi64((const long long int*)tab, _mm_loadu_si128((const __m128i*)idx), 4));
-#else
-    return v_int32x8(_mm256_i32gather_epi64((const int64*)tab, _mm_loadu_si128((const __m128i*)idx), 4));
-#endif
-}
-inline v_int32x8 v256_lut_quads(const int* tab, const int* idx)
-{
-    return v_int32x8(_v256_combine(_mm_loadu_si128((const __m128i*)(tab + idx[0])), _mm_loadu_si128((const __m128i*)(tab + idx[1]))));
-}
-inline v_uint32x8 v256_lut(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v256_lut((const int *)tab, idx)); }
-inline v_uint32x8 v256_lut_pairs(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v256_lut_pairs((const int *)tab, idx)); }
-inline v_uint32x8 v256_lut_quads(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v256_lut_quads((const int *)tab, idx)); }
-
-inline v_int64x4 v256_lut(const int64* tab, const int* idx)
-{
-#if defined(__GNUC__)
-    return v_int64x4(_mm256_i32gather_epi64((const long long int*)tab, _mm_loadu_si128((const __m128i*)idx), 8));
-#else
-    return v_int64x4(_mm256_i32gather_epi64(tab, _mm_loadu_si128((const __m128i*)idx), 8));
-#endif
-}
-inline v_int64x4 v256_lut_pairs(const int64* tab, const int* idx)
-{
-    return v_int64x4(_v256_combine(_mm_loadu_si128((const __m128i*)(tab + idx[0])), _mm_loadu_si128((const __m128i*)(tab + idx[1]))));
-}
-inline v_uint64x4 v256_lut(const uint64* tab, const int* idx) { return v_reinterpret_as_u64(v256_lut((const int64 *)tab, idx)); }
-inline v_uint64x4 v256_lut_pairs(const uint64* tab, const int* idx) { return v_reinterpret_as_u64(v256_lut_pairs((const int64 *)tab, idx)); }
-
-inline v_float32x8 v256_lut(const float* tab, const int* idx)
-{
-    return v_float32x8(_mm256_i32gather_ps(tab, _mm256_loadu_si256((const __m256i*)idx), 4));
-}
-inline v_float32x8 v256_lut_pairs(const float* tab, const int* idx) { return v_reinterpret_as_f32(v256_lut_pairs((const int *)tab, idx)); }
-inline v_float32x8 v256_lut_quads(const float* tab, const int* idx) { return v_reinterpret_as_f32(v256_lut_quads((const int *)tab, idx)); }
-
-inline v_float64x4 v256_lut(const double* tab, const int* idx)
-{
-    return v_float64x4(_mm256_i32gather_pd(tab, _mm_loadu_si128((const __m128i*)idx), 8));
-}
-inline v_float64x4 v256_lut_pairs(const double* tab, const int* idx) { return v_float64x4(_v256_combine(_mm_loadu_pd(tab + idx[0]), _mm_loadu_pd(tab + idx[1]))); }
-
-inline v_int32x8 v_lut(const int* tab, const v_int32x8& idxvec)
-{
-    return v_int32x8(_mm256_i32gather_epi32(tab, idxvec.val, 4));
-}
-
-inline v_uint32x8 v_lut(const unsigned* tab, const v_int32x8& idxvec)
-{
-    return v_reinterpret_as_u32(v_lut((const int *)tab, idxvec));
-}
-
-inline v_float32x8 v_lut(const float* tab, const v_int32x8& idxvec)
-{
-    return v_float32x8(_mm256_i32gather_ps(tab, idxvec.val, 4));
-}
-
-inline v_float64x4 v_lut(const double* tab, const v_int32x8& idxvec)
-{
-    return v_float64x4(_mm256_i32gather_pd(tab, _mm256_castsi256_si128(idxvec.val), 8));
-}
-
-inline void v_lut_deinterleave(const float* tab, const v_int32x8& idxvec, v_float32x8& x, v_float32x8& y)
-{
-    int CV_DECL_ALIGNED(32) idx[8];
-    v_store_aligned(idx, idxvec);
-    __m128 z = _mm_setzero_ps();
-    __m128 xy01, xy45, xy23, xy67;
-    xy01 = _mm_loadl_pi(z, (const __m64*)(tab + idx[0]));
-    xy01 = _mm_loadh_pi(xy01, (const __m64*)(tab + idx[1]));
-    xy45 = _mm_loadl_pi(z, (const __m64*)(tab + idx[4]));
-    xy45 = _mm_loadh_pi(xy45, (const __m64*)(tab + idx[5]));
-    __m256 xy0145 = _v256_combine(xy01, xy45);
-    xy23 = _mm_loadl_pi(z, (const __m64*)(tab + idx[2]));
-    xy23 = _mm_loadh_pi(xy23, (const __m64*)(tab + idx[3]));
-    xy67 = _mm_loadl_pi(z, (const __m64*)(tab + idx[6]));
-    xy67 = _mm_loadh_pi(xy67, (const __m64*)(tab + idx[7]));
-    __m256 xy2367 = _v256_combine(xy23, xy67);
-
-    __m256 xxyy0145 = _mm256_unpacklo_ps(xy0145, xy2367);
-    __m256 xxyy2367 = _mm256_unpackhi_ps(xy0145, xy2367);
-
-    x = v_float32x8(_mm256_unpacklo_ps(xxyy0145, xxyy2367));
-    y = v_float32x8(_mm256_unpackhi_ps(xxyy0145, xxyy2367));
-}
-
-inline void v_lut_deinterleave(const double* tab, const v_int32x8& idxvec, v_float64x4& x, v_float64x4& y)
-{
-    int CV_DECL_ALIGNED(32) idx[4];
-    v_store_low(idx, idxvec);
-    __m128d xy0 = _mm_loadu_pd(tab + idx[0]);
-    __m128d xy2 = _mm_loadu_pd(tab + idx[2]);
-    __m128d xy1 = _mm_loadu_pd(tab + idx[1]);
-    __m128d xy3 = _mm_loadu_pd(tab + idx[3]);
-    __m256d xy02 = _v256_combine(xy0, xy2);
-    __m256d xy13 = _v256_combine(xy1, xy3);
-
-    x = v_float64x4(_mm256_unpacklo_pd(xy02, xy13));
-    y = v_float64x4(_mm256_unpackhi_pd(xy02, xy13));
-}
-
-inline v_int8x32 v_interleave_pairs(const v_int8x32& vec)
-{
-    return v_int8x32(_mm256_shuffle_epi8(vec.val, _mm256_set_epi64x(0x0f0d0e0c0b090a08, 0x0705060403010200, 0x0f0d0e0c0b090a08, 0x0705060403010200)));
-}
-inline v_uint8x32 v_interleave_pairs(const v_uint8x32& vec) { return v_reinterpret_as_u8(v_interleave_pairs(v_reinterpret_as_s8(vec))); }
-inline v_int8x32 v_interleave_quads(const v_int8x32& vec)
-{
-    return v_int8x32(_mm256_shuffle_epi8(vec.val, _mm256_set_epi64x(0x0f0b0e0a0d090c08, 0x0703060205010400, 0x0f0b0e0a0d090c08, 0x0703060205010400)));
-}
-inline v_uint8x32 v_interleave_quads(const v_uint8x32& vec) { return v_reinterpret_as_u8(v_interleave_quads(v_reinterpret_as_s8(vec))); }
-
-inline v_int16x16 v_interleave_pairs(const v_int16x16& vec)
-{
-    return v_int16x16(_mm256_shuffle_epi8(vec.val, _mm256_set_epi64x(0x0f0e0b0a0d0c0908, 0x0706030205040100, 0x0f0e0b0a0d0c0908, 0x0706030205040100)));
-}
-inline v_uint16x16 v_interleave_pairs(const v_uint16x16& vec) { return v_reinterpret_as_u16(v_interleave_pairs(v_reinterpret_as_s16(vec))); }
-inline v_int16x16 v_interleave_quads(const v_int16x16& vec)
-{
-    return v_int16x16(_mm256_shuffle_epi8(vec.val, _mm256_set_epi64x(0x0f0e07060d0c0504, 0x0b0a030209080100, 0x0f0e07060d0c0504, 0x0b0a030209080100)));
-}
-inline v_uint16x16 v_interleave_quads(const v_uint16x16& vec) { return v_reinterpret_as_u16(v_interleave_quads(v_reinterpret_as_s16(vec))); }
-
-inline v_int32x8 v_interleave_pairs(const v_int32x8& vec)
-{
-    return v_int32x8(_mm256_shuffle_epi32(vec.val, _MM_SHUFFLE(3, 1, 2, 0)));
-}
-inline v_uint32x8 v_interleave_pairs(const v_uint32x8& vec) { return v_reinterpret_as_u32(v_interleave_pairs(v_reinterpret_as_s32(vec))); }
-inline v_float32x8 v_interleave_pairs(const v_float32x8& vec) { return v_reinterpret_as_f32(v_interleave_pairs(v_reinterpret_as_s32(vec))); }
-
-inline v_int8x32 v_pack_triplets(const v_int8x32& vec)
-{
-    return v_int8x32(_mm256_permutevar8x32_epi32(_mm256_shuffle_epi8(vec.val, _mm256_broadcastsi128_si256(_mm_set_epi64x(0xffffff0f0e0d0c0a, 0x0908060504020100))),
-                                                 _mm256_set_epi64x(0x0000000700000007, 0x0000000600000005, 0x0000000400000002, 0x0000000100000000)));
-}
-inline v_uint8x32 v_pack_triplets(const v_uint8x32& vec) { return v_reinterpret_as_u8(v_pack_triplets(v_reinterpret_as_s8(vec))); }
-
-inline v_int16x16 v_pack_triplets(const v_int16x16& vec)
-{
-    return v_int16x16(_mm256_permutevar8x32_epi32(_mm256_shuffle_epi8(vec.val, _mm256_broadcastsi128_si256(_mm_set_epi64x(0xffff0f0e0d0c0b0a, 0x0908050403020100))),
-                                                  _mm256_set_epi64x(0x0000000700000007, 0x0000000600000005, 0x0000000400000002, 0x0000000100000000)));
-}
-inline v_uint16x16 v_pack_triplets(const v_uint16x16& vec) { return v_reinterpret_as_u16(v_pack_triplets(v_reinterpret_as_s16(vec))); }
-
-inline v_int32x8 v_pack_triplets(const v_int32x8& vec)
-{
-    return v_int32x8(_mm256_permutevar8x32_epi32(vec.val, _mm256_set_epi64x(0x0000000700000007, 0x0000000600000005, 0x0000000400000002, 0x0000000100000000)));
-}
-inline v_uint32x8 v_pack_triplets(const v_uint32x8& vec) { return v_reinterpret_as_u32(v_pack_triplets(v_reinterpret_as_s32(vec))); }
-inline v_float32x8 v_pack_triplets(const v_float32x8& vec)
-{
-    return v_float32x8(_mm256_permutevar8x32_ps(vec.val, _mm256_set_epi64x(0x0000000700000007, 0x0000000600000005, 0x0000000400000002, 0x0000000100000000)));
-}
-
-////////// Matrix operations /////////
-
-//////// Dot Product ////////
-
-// 16 >> 32
-inline v_int32x8 v_dotprod(const v_int16x16& a, const v_int16x16& b)
-{ return v_int32x8(_mm256_madd_epi16(a.val, b.val)); }
-inline v_int32x8 v_dotprod(const v_int16x16& a, const v_int16x16& b, const v_int32x8& c)
-{ return v_dotprod(a, b) + c; }
-
-// 32 >> 64
-inline v_int64x4 v_dotprod(const v_int32x8& a, const v_int32x8& b)
-{
-    __m256i even = _mm256_mul_epi32(a.val, b.val);
-    __m256i odd = _mm256_mul_epi32(_mm256_srli_epi64(a.val, 32), _mm256_srli_epi64(b.val, 32));
-    return v_int64x4(_mm256_add_epi64(even, odd));
-}
-inline v_int64x4 v_dotprod(const v_int32x8& a, const v_int32x8& b, const v_int64x4& c)
-{ return v_dotprod(a, b) + c; }
-
-// 8 >> 32
-inline v_uint32x8 v_dotprod_expand(const v_uint8x32& a, const v_uint8x32& b)
-{
-    __m256i even_m = _mm256_set1_epi32(0xFF00FF00);
-    __m256i even_a = _mm256_blendv_epi8(a.val, _mm256_setzero_si256(), even_m);
-    __m256i odd_a  = _mm256_srli_epi16(a.val, 8);
-
-    __m256i even_b = _mm256_blendv_epi8(b.val, _mm256_setzero_si256(), even_m);
-    __m256i odd_b  = _mm256_srli_epi16(b.val, 8);
-
-    __m256i prod0  = _mm256_madd_epi16(even_a, even_b);
-    __m256i prod1  = _mm256_madd_epi16(odd_a, odd_b);
-    return v_uint32x8(_mm256_add_epi32(prod0, prod1));
-}
-inline v_uint32x8 v_dotprod_expand(const v_uint8x32& a, const v_uint8x32& b, const v_uint32x8& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-inline v_int32x8 v_dotprod_expand(const v_int8x32& a, const v_int8x32& b)
-{
-    __m256i even_a = _mm256_srai_epi16(_mm256_bslli_epi128(a.val, 1), 8);
-    __m256i odd_a  = _mm256_srai_epi16(a.val, 8);
-
-    __m256i even_b = _mm256_srai_epi16(_mm256_bslli_epi128(b.val, 1), 8);
-    __m256i odd_b  = _mm256_srai_epi16(b.val, 8);
-
-    __m256i prod0  = _mm256_madd_epi16(even_a, even_b);
-    __m256i prod1  = _mm256_madd_epi16(odd_a, odd_b);
-    return v_int32x8(_mm256_add_epi32(prod0, prod1));
-}
-inline v_int32x8 v_dotprod_expand(const v_int8x32& a, const v_int8x32& b, const v_int32x8& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-// 16 >> 64
-inline v_uint64x4 v_dotprod_expand(const v_uint16x16& a, const v_uint16x16& b)
-{
-    __m256i mullo = _mm256_mullo_epi16(a.val, b.val);
-    __m256i mulhi = _mm256_mulhi_epu16(a.val, b.val);
-    __m256i mul0  = _mm256_unpacklo_epi16(mullo, mulhi);
-    __m256i mul1  = _mm256_unpackhi_epi16(mullo, mulhi);
-
-    __m256i p02   = _mm256_blend_epi32(mul0, _mm256_setzero_si256(), 0xAA);
-    __m256i p13   = _mm256_srli_epi64(mul0, 32);
-    __m256i p46   = _mm256_blend_epi32(mul1, _mm256_setzero_si256(), 0xAA);
-    __m256i p57   = _mm256_srli_epi64(mul1, 32);
-
-    __m256i p15_  = _mm256_add_epi64(p02, p13);
-    __m256i p9d_  = _mm256_add_epi64(p46, p57);
-
-    return v_uint64x4(_mm256_add_epi64(
-        _mm256_unpacklo_epi64(p15_, p9d_),
-        _mm256_unpackhi_epi64(p15_, p9d_)
-    ));
-}
-inline v_uint64x4 v_dotprod_expand(const v_uint16x16& a, const v_uint16x16& b, const v_uint64x4& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-inline v_int64x4 v_dotprod_expand(const v_int16x16& a, const v_int16x16& b)
-{
-    __m256i prod = _mm256_madd_epi16(a.val, b.val);
-    __m256i sign = _mm256_srai_epi32(prod, 31);
-
-    __m256i lo = _mm256_unpacklo_epi32(prod, sign);
-    __m256i hi = _mm256_unpackhi_epi32(prod, sign);
-
-    return v_int64x4(_mm256_add_epi64(
-        _mm256_unpacklo_epi64(lo, hi),
-        _mm256_unpackhi_epi64(lo, hi)
-    ));
-}
-inline v_int64x4 v_dotprod_expand(const v_int16x16& a, const v_int16x16& b, const v_int64x4& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-// 32 >> 64f
-inline v_float64x4 v_dotprod_expand(const v_int32x8& a, const v_int32x8& b)
-{ return v_cvt_f64(v_dotprod(a, b)); }
-inline v_float64x4 v_dotprod_expand(const v_int32x8& a, const v_int32x8& b, const v_float64x4& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-//////// Fast Dot Product ////////
-
-// 16 >> 32
-inline v_int32x8 v_dotprod_fast(const v_int16x16& a, const v_int16x16& b)
-{ return v_dotprod(a, b); }
-inline v_int32x8 v_dotprod_fast(const v_int16x16& a, const v_int16x16& b, const v_int32x8& c)
-{ return v_dotprod(a, b, c); }
-
-// 32 >> 64
-inline v_int64x4 v_dotprod_fast(const v_int32x8& a, const v_int32x8& b)
-{ return v_dotprod(a, b); }
-inline v_int64x4 v_dotprod_fast(const v_int32x8& a, const v_int32x8& b, const v_int64x4& c)
-{ return v_dotprod(a, b, c); }
-
-// 8 >> 32
-inline v_uint32x8 v_dotprod_expand_fast(const v_uint8x32& a, const v_uint8x32& b)
-{ return v_dotprod_expand(a, b); }
-inline v_uint32x8 v_dotprod_expand_fast(const v_uint8x32& a, const v_uint8x32& b, const v_uint32x8& c)
-{ return v_dotprod_expand(a, b, c); }
-
-inline v_int32x8 v_dotprod_expand_fast(const v_int8x32& a, const v_int8x32& b)
-{ return v_dotprod_expand(a, b); }
-inline v_int32x8 v_dotprod_expand_fast(const v_int8x32& a, const v_int8x32& b, const v_int32x8& c)
-{ return v_dotprod_expand(a, b, c); }
-
-// 16 >> 64
-inline v_uint64x4 v_dotprod_expand_fast(const v_uint16x16& a, const v_uint16x16& b)
-{
-    __m256i mullo = _mm256_mullo_epi16(a.val, b.val);
-    __m256i mulhi = _mm256_mulhi_epu16(a.val, b.val);
-    __m256i mul0  = _mm256_unpacklo_epi16(mullo, mulhi);
-    __m256i mul1  = _mm256_unpackhi_epi16(mullo, mulhi);
-
-    __m256i p02   = _mm256_blend_epi32(mul0, _mm256_setzero_si256(), 0xAA);
-    __m256i p13   = _mm256_srli_epi64(mul0, 32);
-    __m256i p46   = _mm256_blend_epi32(mul1, _mm256_setzero_si256(), 0xAA);
-    __m256i p57   = _mm256_srli_epi64(mul1, 32);
-
-    __m256i p15_  = _mm256_add_epi64(p02, p13);
-    __m256i p9d_  = _mm256_add_epi64(p46, p57);
-
-    return v_uint64x4(_mm256_add_epi64(p15_, p9d_));
-}
-inline v_uint64x4 v_dotprod_expand_fast(const v_uint16x16& a, const v_uint16x16& b, const v_uint64x4& c)
-{ return v_dotprod_expand_fast(a, b) + c; }
-
-inline v_int64x4 v_dotprod_expand_fast(const v_int16x16& a, const v_int16x16& b)
-{
-    __m256i prod = _mm256_madd_epi16(a.val, b.val);
-    __m256i sign = _mm256_srai_epi32(prod, 31);
-    __m256i lo = _mm256_unpacklo_epi32(prod, sign);
-    __m256i hi = _mm256_unpackhi_epi32(prod, sign);
-    return v_int64x4(_mm256_add_epi64(lo, hi));
-}
-inline v_int64x4 v_dotprod_expand_fast(const v_int16x16& a, const v_int16x16& b, const v_int64x4& c)
-{ return v_dotprod_expand_fast(a, b) + c; }
-
-// 32 >> 64f
-inline v_float64x4 v_dotprod_expand_fast(const v_int32x8& a, const v_int32x8& b)
-{ return v_dotprod_expand(a, b); }
-inline v_float64x4 v_dotprod_expand_fast(const v_int32x8& a, const v_int32x8& b, const v_float64x4& c)
-{ return v_dotprod_expand(a, b, c); }
-
-#define OPENCV_HAL_AVX_SPLAT2_PS(a, im) \
-    v_float32x8(_mm256_permute_ps(a.val, _MM_SHUFFLE(im, im, im, im)))
-
-inline v_float32x8 v_matmul(const v_float32x8& v, const v_float32x8& m0,
-                            const v_float32x8& m1, const v_float32x8& m2,
-                            const v_float32x8& m3)
-{
-    v_float32x8 v04 = OPENCV_HAL_AVX_SPLAT2_PS(v, 0);
-    v_float32x8 v15 = OPENCV_HAL_AVX_SPLAT2_PS(v, 1);
-    v_float32x8 v26 = OPENCV_HAL_AVX_SPLAT2_PS(v, 2);
-    v_float32x8 v37 = OPENCV_HAL_AVX_SPLAT2_PS(v, 3);
-    return v_fma(v04, m0, v_fma(v15, m1, v_fma(v26, m2, v37 * m3)));
-}
-
-inline v_float32x8 v_matmuladd(const v_float32x8& v, const v_float32x8& m0,
-                               const v_float32x8& m1, const v_float32x8& m2,
-                               const v_float32x8& a)
-{
-    v_float32x8 v04 = OPENCV_HAL_AVX_SPLAT2_PS(v, 0);
-    v_float32x8 v15 = OPENCV_HAL_AVX_SPLAT2_PS(v, 1);
-    v_float32x8 v26 = OPENCV_HAL_AVX_SPLAT2_PS(v, 2);
-    return v_fma(v04, m0, v_fma(v15, m1, v_fma(v26, m2, a)));
-}
-
-#define OPENCV_HAL_IMPL_AVX_TRANSPOSE4x4(_Tpvec, suffix, cast_from, cast_to)    \
-    inline void v_transpose4x4(const _Tpvec& a0, const _Tpvec& a1,              \
-                               const _Tpvec& a2, const _Tpvec& a3,              \
-                               _Tpvec& b0, _Tpvec& b1, _Tpvec& b2, _Tpvec& b3)  \
-    {                                                                           \
-        __m256i t0 = cast_from(_mm256_unpacklo_##suffix(a0.val, a1.val));       \
-        __m256i t1 = cast_from(_mm256_unpacklo_##suffix(a2.val, a3.val));       \
-        __m256i t2 = cast_from(_mm256_unpackhi_##suffix(a0.val, a1.val));       \
-        __m256i t3 = cast_from(_mm256_unpackhi_##suffix(a2.val, a3.val));       \
-        b0.val = cast_to(_mm256_unpacklo_epi64(t0, t1));                        \
-        b1.val = cast_to(_mm256_unpackhi_epi64(t0, t1));                        \
-        b2.val = cast_to(_mm256_unpacklo_epi64(t2, t3));                        \
-        b3.val = cast_to(_mm256_unpackhi_epi64(t2, t3));                        \
-    }
-
-OPENCV_HAL_IMPL_AVX_TRANSPOSE4x4(v_uint32x8,  epi32, OPENCV_HAL_NOP, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_AVX_TRANSPOSE4x4(v_int32x8,   epi32, OPENCV_HAL_NOP, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_AVX_TRANSPOSE4x4(v_float32x8, ps, _mm256_castps_si256, _mm256_castsi256_ps)
-
-//////////////// Value reordering ///////////////
-
-/* Expand */
-#define OPENCV_HAL_IMPL_AVX_EXPAND(_Tpvec, _Tpwvec, _Tp, intrin)    \
-    inline void v_expand(const _Tpvec& a, _Tpwvec& b0, _Tpwvec& b1) \
-    {                                                               \
-        b0.val = intrin(_v256_extract_low(a.val));                  \
-        b1.val = intrin(_v256_extract_high(a.val));                 \
-    }                                                               \
-    inline _Tpwvec v_expand_low(const _Tpvec& a)                    \
-    { return _Tpwvec(intrin(_v256_extract_low(a.val))); }           \
-    inline _Tpwvec v_expand_high(const _Tpvec& a)                   \
-    { return _Tpwvec(intrin(_v256_extract_high(a.val))); }          \
-    inline _Tpwvec v256_load_expand(const _Tp* ptr)                 \
-    {                                                               \
-        __m128i a = _mm_loadu_si128((const __m128i*)ptr);           \
-        return _Tpwvec(intrin(a));                                  \
-    }
-
-OPENCV_HAL_IMPL_AVX_EXPAND(v_uint8x32,  v_uint16x16, uchar,    _mm256_cvtepu8_epi16)
-OPENCV_HAL_IMPL_AVX_EXPAND(v_int8x32,   v_int16x16,  schar,    _mm256_cvtepi8_epi16)
-OPENCV_HAL_IMPL_AVX_EXPAND(v_uint16x16, v_uint32x8,  ushort,   _mm256_cvtepu16_epi32)
-OPENCV_HAL_IMPL_AVX_EXPAND(v_int16x16,  v_int32x8,   short,    _mm256_cvtepi16_epi32)
-OPENCV_HAL_IMPL_AVX_EXPAND(v_uint32x8,  v_uint64x4,  unsigned, _mm256_cvtepu32_epi64)
-OPENCV_HAL_IMPL_AVX_EXPAND(v_int32x8,   v_int64x4,   int,      _mm256_cvtepi32_epi64)
-
-#define OPENCV_HAL_IMPL_AVX_EXPAND_Q(_Tpvec, _Tp, intrin)   \
-    inline _Tpvec v256_load_expand_q(const _Tp* ptr)        \
-    {                                                       \
-        __m128i a = _mm_loadl_epi64((const __m128i*)ptr);   \
-        return _Tpvec(intrin(a));                           \
-    }
-
-OPENCV_HAL_IMPL_AVX_EXPAND_Q(v_uint32x8, uchar, _mm256_cvtepu8_epi32)
-OPENCV_HAL_IMPL_AVX_EXPAND_Q(v_int32x8,  schar, _mm256_cvtepi8_epi32)
-
-/* pack */
-// 16
-inline v_int8x32 v_pack(const v_int16x16& a, const v_int16x16& b)
-{ return v_int8x32(_v256_shuffle_odd_64(_mm256_packs_epi16(a.val, b.val))); }
-
-inline v_uint8x32 v_pack(const v_uint16x16& a, const v_uint16x16& b)
-{
-    __m256i t = _mm256_set1_epi16(255);
-    __m256i a1 = _mm256_min_epu16(a.val, t);
-    __m256i b1 = _mm256_min_epu16(b.val, t);
-    return v_uint8x32(_v256_shuffle_odd_64(_mm256_packus_epi16(a1, b1)));
-}
-
-inline v_uint8x32 v_pack_u(const v_int16x16& a, const v_int16x16& b)
-{
-    return v_uint8x32(_v256_shuffle_odd_64(_mm256_packus_epi16(a.val, b.val)));
-}
-
-inline void v_pack_store(schar* ptr, const v_int16x16& a)
-{ v_store_low(ptr, v_pack(a, a)); }
-
-inline void v_pack_store(uchar* ptr, const v_uint16x16& a)
-{
-    const __m256i m = _mm256_set1_epi16(255);
-    __m256i am = _mm256_min_epu16(a.val, m);
-            am =  _v256_shuffle_odd_64(_mm256_packus_epi16(am, am));
-    v_store_low(ptr, v_uint8x32(am));
-}
-
-inline void v_pack_u_store(uchar* ptr, const v_int16x16& a)
-{ v_store_low(ptr, v_pack_u(a, a)); }
-
-template<int n> inline
-v_uint8x32 v_rshr_pack(const v_uint16x16& a, const v_uint16x16& b)
-{
-    // we assume that n > 0, and so the shifted 16-bit values can be treated as signed numbers.
-    v_uint16x16 delta = v256_setall_u16((short)(1 << (n-1)));
-    return v_pack_u(v_reinterpret_as_s16((a + delta) >> n),
-                    v_reinterpret_as_s16((b + delta) >> n));
-}
-
-template<int n> inline
-void v_rshr_pack_store(uchar* ptr, const v_uint16x16& a)
-{
-    v_uint16x16 delta = v256_setall_u16((short)(1 << (n-1)));
-    v_pack_u_store(ptr, v_reinterpret_as_s16((a + delta) >> n));
-}
-
-template<int n> inline
-v_uint8x32 v_rshr_pack_u(const v_int16x16& a, const v_int16x16& b)
-{
-    v_int16x16 delta = v256_setall_s16((short)(1 << (n-1)));
-    return v_pack_u((a + delta) >> n, (b + delta) >> n);
-}
-
-template<int n> inline
-void v_rshr_pack_u_store(uchar* ptr, const v_int16x16& a)
-{
-    v_int16x16 delta = v256_setall_s16((short)(1 << (n-1)));
-    v_pack_u_store(ptr, (a + delta) >> n);
-}
-
-template<int n> inline
-v_int8x32 v_rshr_pack(const v_int16x16& a, const v_int16x16& b)
-{
-    v_int16x16 delta = v256_setall_s16((short)(1 << (n-1)));
-    return v_pack((a + delta) >> n, (b + delta) >> n);
-}
-
-template<int n> inline
-void v_rshr_pack_store(schar* ptr, const v_int16x16& a)
-{
-    v_int16x16 delta = v256_setall_s16((short)(1 << (n-1)));
-    v_pack_store(ptr, (a + delta) >> n);
-}
-
-// 32
-inline v_int16x16 v_pack(const v_int32x8& a, const v_int32x8& b)
-{ return v_int16x16(_v256_shuffle_odd_64(_mm256_packs_epi32(a.val, b.val))); }
-
-inline v_uint16x16 v_pack(const v_uint32x8& a, const v_uint32x8& b)
-{ return v_uint16x16(_v256_shuffle_odd_64(_v256_packs_epu32(a.val, b.val))); }
-
-inline v_uint16x16 v_pack_u(const v_int32x8& a, const v_int32x8& b)
-{ return v_uint16x16(_v256_shuffle_odd_64(_mm256_packus_epi32(a.val, b.val))); }
-
-inline void v_pack_store(short* ptr, const v_int32x8& a)
-{ v_store_low(ptr, v_pack(a, a)); }
-
-inline void v_pack_store(ushort* ptr, const v_uint32x8& a)
-{
-    const __m256i m = _mm256_set1_epi32(65535);
-    __m256i am = _mm256_min_epu32(a.val, m);
-            am = _v256_shuffle_odd_64(_mm256_packus_epi32(am, am));
-    v_store_low(ptr, v_uint16x16(am));
-}
-
-inline void v_pack_u_store(ushort* ptr, const v_int32x8& a)
-{ v_store_low(ptr, v_pack_u(a, a)); }
-
-
-template<int n> inline
-v_uint16x16 v_rshr_pack(const v_uint32x8& a, const v_uint32x8& b)
-{
-    // we assume that n > 0, and so the shifted 32-bit values can be treated as signed numbers.
-    v_uint32x8 delta = v256_setall_u32(1 << (n-1));
-    return v_pack_u(v_reinterpret_as_s32((a + delta) >> n),
-                    v_reinterpret_as_s32((b + delta) >> n));
-}
-
-template<int n> inline
-void v_rshr_pack_store(ushort* ptr, const v_uint32x8& a)
-{
-    v_uint32x8 delta = v256_setall_u32(1 << (n-1));
-    v_pack_u_store(ptr, v_reinterpret_as_s32((a + delta) >> n));
-}
-
-template<int n> inline
-v_uint16x16 v_rshr_pack_u(const v_int32x8& a, const v_int32x8& b)
-{
-    v_int32x8 delta = v256_setall_s32(1 << (n-1));
-    return v_pack_u((a + delta) >> n, (b + delta) >> n);
-}
-
-template<int n> inline
-void v_rshr_pack_u_store(ushort* ptr, const v_int32x8& a)
-{
-    v_int32x8 delta = v256_setall_s32(1 << (n-1));
-    v_pack_u_store(ptr, (a + delta) >> n);
-}
-
-template<int n> inline
-v_int16x16 v_rshr_pack(const v_int32x8& a, const v_int32x8& b)
-{
-    v_int32x8 delta = v256_setall_s32(1 << (n-1));
-    return v_pack((a + delta) >> n, (b + delta) >> n);
-}
-
-template<int n> inline
-void v_rshr_pack_store(short* ptr, const v_int32x8& a)
-{
-    v_int32x8 delta = v256_setall_s32(1 << (n-1));
-    v_pack_store(ptr, (a + delta) >> n);
-}
-
-// 64
-// Non-saturating pack
-inline v_uint32x8 v_pack(const v_uint64x4& a, const v_uint64x4& b)
-{
-    __m256i a0 = _mm256_shuffle_epi32(a.val, _MM_SHUFFLE(0, 0, 2, 0));
-    __m256i b0 = _mm256_shuffle_epi32(b.val, _MM_SHUFFLE(0, 0, 2, 0));
-    __m256i ab = _mm256_unpacklo_epi64(a0, b0); // a0, a1, b0, b1, a2, a3, b2, b3
-    return v_uint32x8(_v256_shuffle_odd_64(ab));
-}
-
-inline v_int32x8 v_pack(const v_int64x4& a, const v_int64x4& b)
-{ return v_reinterpret_as_s32(v_pack(v_reinterpret_as_u64(a), v_reinterpret_as_u64(b))); }
-
-inline void v_pack_store(unsigned* ptr, const v_uint64x4& a)
-{
-    __m256i a0 = _mm256_shuffle_epi32(a.val, _MM_SHUFFLE(0, 0, 2, 0));
-    v_store_low(ptr, v_uint32x8(_v256_shuffle_odd_64(a0)));
-}
-
-inline void v_pack_store(int* ptr, const v_int64x4& b)
-{ v_pack_store((unsigned*)ptr, v_reinterpret_as_u64(b)); }
-
-template<int n> inline
-v_uint32x8 v_rshr_pack(const v_uint64x4& a, const v_uint64x4& b)
-{
-    v_uint64x4 delta = v256_setall_u64((uint64)1 << (n-1));
-    return v_pack((a + delta) >> n, (b + delta) >> n);
-}
-
-template<int n> inline
-void v_rshr_pack_store(unsigned* ptr, const v_uint64x4& a)
-{
-    v_uint64x4 delta = v256_setall_u64((uint64)1 << (n-1));
-    v_pack_store(ptr, (a + delta) >> n);
-}
-
-template<int n> inline
-v_int32x8 v_rshr_pack(const v_int64x4& a, const v_int64x4& b)
-{
-    v_int64x4 delta = v256_setall_s64((int64)1 << (n-1));
-    return v_pack((a + delta) >> n, (b + delta) >> n);
-}
-
-template<int n> inline
-void v_rshr_pack_store(int* ptr, const v_int64x4& a)
-{
-    v_int64x4 delta = v256_setall_s64((int64)1 << (n-1));
-    v_pack_store(ptr, (a + delta) >> n);
-}
-
-// pack boolean
-inline v_uint8x32 v_pack_b(const v_uint16x16& a, const v_uint16x16& b)
-{
-    __m256i ab = _mm256_packs_epi16(a.val, b.val);
-    return v_uint8x32(_v256_shuffle_odd_64(ab));
-}
-
-inline v_uint8x32 v_pack_b(const v_uint32x8& a, const v_uint32x8& b,
-                           const v_uint32x8& c, const v_uint32x8& d)
-{
-    __m256i ab = _mm256_packs_epi32(a.val, b.val);
-    __m256i cd = _mm256_packs_epi32(c.val, d.val);
-
-    __m256i abcd = _v256_shuffle_odd_64(_mm256_packs_epi16(ab, cd));
-    return v_uint8x32(_mm256_shuffle_epi32(abcd, _MM_SHUFFLE(3, 1, 2, 0)));
-}
-
-inline v_uint8x32 v_pack_b(const v_uint64x4& a, const v_uint64x4& b, const v_uint64x4& c,
-                           const v_uint64x4& d, const v_uint64x4& e, const v_uint64x4& f,
-                           const v_uint64x4& g, const v_uint64x4& h)
-{
-    __m256i ab = _mm256_packs_epi32(a.val, b.val);
-    __m256i cd = _mm256_packs_epi32(c.val, d.val);
-    __m256i ef = _mm256_packs_epi32(e.val, f.val);
-    __m256i gh = _mm256_packs_epi32(g.val, h.val);
-
-    __m256i abcd = _mm256_packs_epi32(ab, cd);
-    __m256i efgh = _mm256_packs_epi32(ef, gh);
-    __m256i pkall = _v256_shuffle_odd_64(_mm256_packs_epi16(abcd, efgh));
-
-    __m256i rev = _mm256_alignr_epi8(pkall, pkall, 8);
-    return v_uint8x32(_mm256_unpacklo_epi16(pkall, rev));
-}
-
-/* Recombine */
-// its up there with load and store operations
-
-/* Extract */
-#define OPENCV_HAL_IMPL_AVX_EXTRACT(_Tpvec)                    \
-    template<int s>                                            \
-    inline _Tpvec v_extract(const _Tpvec& a, const _Tpvec& b)  \
-    { return v_rotate_right<s>(a, b); }
-
-OPENCV_HAL_IMPL_AVX_EXTRACT(v_uint8x32)
-OPENCV_HAL_IMPL_AVX_EXTRACT(v_int8x32)
-OPENCV_HAL_IMPL_AVX_EXTRACT(v_uint16x16)
-OPENCV_HAL_IMPL_AVX_EXTRACT(v_int16x16)
-OPENCV_HAL_IMPL_AVX_EXTRACT(v_uint32x8)
-OPENCV_HAL_IMPL_AVX_EXTRACT(v_int32x8)
-OPENCV_HAL_IMPL_AVX_EXTRACT(v_uint64x4)
-OPENCV_HAL_IMPL_AVX_EXTRACT(v_int64x4)
-OPENCV_HAL_IMPL_AVX_EXTRACT(v_float32x8)
-OPENCV_HAL_IMPL_AVX_EXTRACT(v_float64x4)
-
-template<int i>
-inline uchar v_extract_n(v_uint8x32 a)
-{
-    return (uchar)_v256_extract_epi8<i>(a.val);
-}
-
-template<int i>
-inline schar v_extract_n(v_int8x32 a)
-{
-    return (schar)v_extract_n<i>(v_reinterpret_as_u8(a));
-}
-
-template<int i>
-inline ushort v_extract_n(v_uint16x16 a)
-{
-    return (ushort)_v256_extract_epi16<i>(a.val);
-}
-
-template<int i>
-inline short v_extract_n(v_int16x16 a)
-{
-    return (short)v_extract_n<i>(v_reinterpret_as_u16(a));
-}
-
-template<int i>
-inline uint v_extract_n(v_uint32x8 a)
-{
-    return (uint)_v256_extract_epi32<i>(a.val);
-}
-
-template<int i>
-inline int v_extract_n(v_int32x8 a)
-{
-    return (int)v_extract_n<i>(v_reinterpret_as_u32(a));
-}
-
-template<int i>
-inline uint64 v_extract_n(v_uint64x4 a)
-{
-    return (uint64)_v256_extract_epi64<i>(a.val);
-}
-
-template<int i>
-inline int64 v_extract_n(v_int64x4 v)
-{
-    return (int64)v_extract_n<i>(v_reinterpret_as_u64(v));
-}
-
-template<int i>
-inline float v_extract_n(v_float32x8 v)
-{
-    union { uint iv; float fv; } d;
-    d.iv = v_extract_n<i>(v_reinterpret_as_u32(v));
-    return d.fv;
-}
-
-template<int i>
-inline double v_extract_n(v_float64x4 v)
-{
-    union { uint64 iv; double dv; } d;
-    d.iv = v_extract_n<i>(v_reinterpret_as_u64(v));
-    return d.dv;
-}
-
-template<int i>
-inline v_uint32x8 v_broadcast_element(v_uint32x8 a)
-{
-    static const __m256i perm = _mm256_set1_epi32((char)i);
-    return v_uint32x8(_mm256_permutevar8x32_epi32(a.val, perm));
-}
-
-template<int i>
-inline v_int32x8 v_broadcast_element(const v_int32x8 &a)
-{ return v_reinterpret_as_s32(v_broadcast_element<i>(v_reinterpret_as_u32(a))); }
-
-template<int i>
-inline v_float32x8 v_broadcast_element(const v_float32x8 &a)
-{ return v_reinterpret_as_f32(v_broadcast_element<i>(v_reinterpret_as_u32(a))); }
-
-
-///////////////////// load deinterleave /////////////////////////////
-
-inline void v_load_deinterleave( const uchar* ptr, v_uint8x32& a, v_uint8x32& b )
-{
-    __m256i ab0 = _mm256_loadu_si256((const __m256i*)ptr);
-    __m256i ab1 = _mm256_loadu_si256((const __m256i*)(ptr + 32));
-
-    const __m256i sh = _mm256_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15,
-                                               0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15);
-    __m256i p0 = _mm256_shuffle_epi8(ab0, sh);
-    __m256i p1 = _mm256_shuffle_epi8(ab1, sh);
-    __m256i pl = _mm256_permute2x128_si256(p0, p1, 0 + 2*16);
-    __m256i ph = _mm256_permute2x128_si256(p0, p1, 1 + 3*16);
-    __m256i a0 = _mm256_unpacklo_epi64(pl, ph);
-    __m256i b0 = _mm256_unpackhi_epi64(pl, ph);
-    a = v_uint8x32(a0);
-    b = v_uint8x32(b0);
-}
-
-inline void v_load_deinterleave( const ushort* ptr, v_uint16x16& a, v_uint16x16& b )
-{
-    __m256i ab0 = _mm256_loadu_si256((const __m256i*)ptr);
-    __m256i ab1 = _mm256_loadu_si256((const __m256i*)(ptr + 16));
-
-    const __m256i sh = _mm256_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15,
-                                               0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15);
-    __m256i p0 = _mm256_shuffle_epi8(ab0, sh);
-    __m256i p1 = _mm256_shuffle_epi8(ab1, sh);
-    __m256i pl = _mm256_permute2x128_si256(p0, p1, 0 + 2*16);
-    __m256i ph = _mm256_permute2x128_si256(p0, p1, 1 + 3*16);
-    __m256i a0 = _mm256_unpacklo_epi64(pl, ph);
-    __m256i b0 = _mm256_unpackhi_epi64(pl, ph);
-    a = v_uint16x16(a0);
-    b = v_uint16x16(b0);
-}
-
-inline void v_load_deinterleave( const unsigned* ptr, v_uint32x8& a, v_uint32x8& b )
-{
-    __m256i ab0 = _mm256_loadu_si256((const __m256i*)ptr);
-    __m256i ab1 = _mm256_loadu_si256((const __m256i*)(ptr + 8));
-
-    enum { sh = 0+2*4+1*16+3*64 };
-    __m256i p0 = _mm256_shuffle_epi32(ab0, sh);
-    __m256i p1 = _mm256_shuffle_epi32(ab1, sh);
-    __m256i pl = _mm256_permute2x128_si256(p0, p1, 0 + 2*16);
-    __m256i ph = _mm256_permute2x128_si256(p0, p1, 1 + 3*16);
-    __m256i a0 = _mm256_unpacklo_epi64(pl, ph);
-    __m256i b0 = _mm256_unpackhi_epi64(pl, ph);
-    a = v_uint32x8(a0);
-    b = v_uint32x8(b0);
-}
-
-inline void v_load_deinterleave( const uint64* ptr, v_uint64x4& a, v_uint64x4& b )
-{
-    __m256i ab0 = _mm256_loadu_si256((const __m256i*)ptr);
-    __m256i ab1 = _mm256_loadu_si256((const __m256i*)(ptr + 4));
-
-    __m256i pl = _mm256_permute2x128_si256(ab0, ab1, 0 + 2*16);
-    __m256i ph = _mm256_permute2x128_si256(ab0, ab1, 1 + 3*16);
-    __m256i a0 = _mm256_unpacklo_epi64(pl, ph);
-    __m256i b0 = _mm256_unpackhi_epi64(pl, ph);
-    a = v_uint64x4(a0);
-    b = v_uint64x4(b0);
-}
-
-inline void v_load_deinterleave( const uchar* ptr, v_uint8x32& a, v_uint8x32& b, v_uint8x32& c )
-{
-    __m256i bgr0 = _mm256_loadu_si256((const __m256i*)ptr);
-    __m256i bgr1 = _mm256_loadu_si256((const __m256i*)(ptr + 32));
-    __m256i bgr2 = _mm256_loadu_si256((const __m256i*)(ptr + 64));
-
-    __m256i s02_low = _mm256_permute2x128_si256(bgr0, bgr2, 0 + 2*16);
-    __m256i s02_high = _mm256_permute2x128_si256(bgr0, bgr2, 1 + 3*16);
-
-    const __m256i m0 = _mm256_setr_epi8(0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0,
-                                               0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0);
-    const __m256i m1 = _mm256_setr_epi8(0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0,
-                                               -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1);
-
-    __m256i b0 = _mm256_blendv_epi8(_mm256_blendv_epi8(s02_low, s02_high, m0), bgr1, m1);
-    __m256i g0 = _mm256_blendv_epi8(_mm256_blendv_epi8(s02_high, s02_low, m1), bgr1, m0);
-    __m256i r0 = _mm256_blendv_epi8(_mm256_blendv_epi8(bgr1, s02_low, m0), s02_high, m1);
-
-    const __m256i
-    sh_b = _mm256_setr_epi8(0, 3, 6, 9, 12, 15, 2, 5, 8, 11, 14, 1, 4, 7, 10, 13,
-                            0, 3, 6, 9, 12, 15, 2, 5, 8, 11, 14, 1, 4, 7, 10, 13),
-    sh_g = _mm256_setr_epi8(1, 4, 7, 10, 13, 0, 3, 6, 9, 12, 15, 2, 5, 8, 11, 14,
-                            1, 4, 7, 10, 13, 0, 3, 6, 9, 12, 15, 2, 5, 8, 11, 14),
-    sh_r = _mm256_setr_epi8(2, 5, 8, 11, 14, 1, 4, 7, 10, 13, 0, 3, 6, 9, 12, 15,
-                            2, 5, 8, 11, 14, 1, 4, 7, 10, 13, 0, 3, 6, 9, 12, 15);
-    b0 = _mm256_shuffle_epi8(b0, sh_b);
-    g0 = _mm256_shuffle_epi8(g0, sh_g);
-    r0 = _mm256_shuffle_epi8(r0, sh_r);
-
-    a = v_uint8x32(b0);
-    b = v_uint8x32(g0);
-    c = v_uint8x32(r0);
-}
-
-inline void v_load_deinterleave( const ushort* ptr, v_uint16x16& a, v_uint16x16& b, v_uint16x16& c )
-{
-    __m256i bgr0 = _mm256_loadu_si256((const __m256i*)ptr);
-    __m256i bgr1 = _mm256_loadu_si256((const __m256i*)(ptr + 16));
-    __m256i bgr2 = _mm256_loadu_si256((const __m256i*)(ptr + 32));
-
-    __m256i s02_low = _mm256_permute2x128_si256(bgr0, bgr2, 0 + 2*16);
-    __m256i s02_high = _mm256_permute2x128_si256(bgr0, bgr2, 1 + 3*16);
-
-    const __m256i m0 = _mm256_setr_epi8(0, 0, -1, -1, 0, 0, 0, 0, -1, -1, 0, 0, 0, 0, -1, -1,
-                                               0, 0, 0, 0, -1, -1, 0, 0, 0, 0, -1, -1, 0, 0, 0, 0);
-    const __m256i m1 = _mm256_setr_epi8(0, 0, 0, 0, -1, -1, 0, 0, 0, 0, -1, -1, 0, 0, 0, 0,
-                                               -1, -1, 0, 0, 0, 0, -1, -1, 0, 0, 0, 0, -1, -1, 0, 0);
-    __m256i b0 = _mm256_blendv_epi8(_mm256_blendv_epi8(s02_low, s02_high, m0), bgr1, m1);
-    __m256i g0 = _mm256_blendv_epi8(_mm256_blendv_epi8(bgr1, s02_low, m0), s02_high, m1);
-    __m256i r0 = _mm256_blendv_epi8(_mm256_blendv_epi8(s02_high, s02_low, m1), bgr1, m0);
-    const __m256i sh_b = _mm256_setr_epi8(0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15, 4, 5, 10, 11,
-                                                 0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15, 4, 5, 10, 11);
-    const __m256i sh_g = _mm256_setr_epi8(2, 3, 8, 9, 14, 15, 4, 5, 10, 11, 0, 1, 6, 7, 12, 13,
-                                                 2, 3, 8, 9, 14, 15, 4, 5, 10, 11, 0, 1, 6, 7, 12, 13);
-    const __m256i sh_r = _mm256_setr_epi8(4, 5, 10, 11, 0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15,
-                                                 4, 5, 10, 11, 0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15);
-    b0 = _mm256_shuffle_epi8(b0, sh_b);
-    g0 = _mm256_shuffle_epi8(g0, sh_g);
-    r0 = _mm256_shuffle_epi8(r0, sh_r);
-
-    a = v_uint16x16(b0);
-    b = v_uint16x16(g0);
-    c = v_uint16x16(r0);
-}
-
-inline void v_load_deinterleave( const unsigned* ptr, v_uint32x8& a, v_uint32x8& b, v_uint32x8& c )
-{
-    __m256i bgr0 = _mm256_loadu_si256((const __m256i*)ptr);
-    __m256i bgr1 = _mm256_loadu_si256((const __m256i*)(ptr + 8));
-    __m256i bgr2 = _mm256_loadu_si256((const __m256i*)(ptr + 16));
-
-    __m256i s02_low = _mm256_permute2x128_si256(bgr0, bgr2, 0 + 2*16);
-    __m256i s02_high = _mm256_permute2x128_si256(bgr0, bgr2, 1 + 3*16);
-
-    __m256i b0 = _mm256_blend_epi32(_mm256_blend_epi32(s02_low, s02_high, 0x24), bgr1, 0x92);
-    __m256i g0 = _mm256_blend_epi32(_mm256_blend_epi32(s02_high, s02_low, 0x92), bgr1, 0x24);
-    __m256i r0 = _mm256_blend_epi32(_mm256_blend_epi32(bgr1, s02_low, 0x24), s02_high, 0x92);
-
-    b0 = _mm256_shuffle_epi32(b0, 0x6c);
-    g0 = _mm256_shuffle_epi32(g0, 0xb1);
-    r0 = _mm256_shuffle_epi32(r0, 0xc6);
-
-    a = v_uint32x8(b0);
-    b = v_uint32x8(g0);
-    c = v_uint32x8(r0);
-}
-
-inline void v_load_deinterleave( const uint64* ptr, v_uint64x4& a, v_uint64x4& b, v_uint64x4& c )
-{
-    __m256i bgr0 = _mm256_loadu_si256((const __m256i*)ptr);
-    __m256i bgr1 = _mm256_loadu_si256((const __m256i*)(ptr + 4));
-    __m256i bgr2 = _mm256_loadu_si256((const __m256i*)(ptr + 8));
-
-    __m256i s01 = _mm256_blend_epi32(bgr0, bgr1, 0xf0);
-    __m256i s12 = _mm256_blend_epi32(bgr1, bgr2, 0xf0);
-    __m256i s20r = _mm256_permute4x64_epi64(_mm256_blend_epi32(bgr2, bgr0, 0xf0), 0x1b);
-    __m256i b0 = _mm256_unpacklo_epi64(s01, s20r);
-    __m256i g0 = _mm256_alignr_epi8(s12, s01, 8);
-    __m256i r0 = _mm256_unpackhi_epi64(s20r, s12);
-
-    a = v_uint64x4(b0);
-    b = v_uint64x4(g0);
-    c = v_uint64x4(r0);
-}
-
-inline void v_load_deinterleave( const uchar* ptr, v_uint8x32& a, v_uint8x32& b, v_uint8x32& c, v_uint8x32& d )
-{
-    __m256i bgr0 = _mm256_loadu_si256((const __m256i*)ptr);
-    __m256i bgr1 = _mm256_loadu_si256((const __m256i*)(ptr + 32));
-    __m256i bgr2 = _mm256_loadu_si256((const __m256i*)(ptr + 64));
-    __m256i bgr3 = _mm256_loadu_si256((const __m256i*)(ptr + 96));
-    const __m256i sh = _mm256_setr_epi8(0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15,
-                                               0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15);
-
-    __m256i p0 = _mm256_shuffle_epi8(bgr0, sh);
-    __m256i p1 = _mm256_shuffle_epi8(bgr1, sh);
-    __m256i p2 = _mm256_shuffle_epi8(bgr2, sh);
-    __m256i p3 = _mm256_shuffle_epi8(bgr3, sh);
-
-    __m256i p01l = _mm256_unpacklo_epi32(p0, p1);
-    __m256i p01h = _mm256_unpackhi_epi32(p0, p1);
-    __m256i p23l = _mm256_unpacklo_epi32(p2, p3);
-    __m256i p23h = _mm256_unpackhi_epi32(p2, p3);
-
-    __m256i pll = _mm256_permute2x128_si256(p01l, p23l, 0 + 2*16);
-    __m256i plh = _mm256_permute2x128_si256(p01l, p23l, 1 + 3*16);
-    __m256i phl = _mm256_permute2x128_si256(p01h, p23h, 0 + 2*16);
-    __m256i phh = _mm256_permute2x128_si256(p01h, p23h, 1 + 3*16);
-
-    __m256i b0 = _mm256_unpacklo_epi32(pll, plh);
-    __m256i g0 = _mm256_unpackhi_epi32(pll, plh);
-    __m256i r0 = _mm256_unpacklo_epi32(phl, phh);
-    __m256i a0 = _mm256_unpackhi_epi32(phl, phh);
-
-    a = v_uint8x32(b0);
-    b = v_uint8x32(g0);
-    c = v_uint8x32(r0);
-    d = v_uint8x32(a0);
-}
-
-inline void v_load_deinterleave( const ushort* ptr, v_uint16x16& a, v_uint16x16& b, v_uint16x16& c, v_uint16x16& d )
-{
-    __m256i bgr0 = _mm256_loadu_si256((const __m256i*)ptr);
-    __m256i bgr1 = _mm256_loadu_si256((const __m256i*)(ptr + 16));
-    __m256i bgr2 = _mm256_loadu_si256((const __m256i*)(ptr + 32));
-    __m256i bgr3 = _mm256_loadu_si256((const __m256i*)(ptr + 48));
-    const __m256i sh = _mm256_setr_epi8(0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
-                                               0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15);
-    __m256i p0 = _mm256_shuffle_epi8(bgr0, sh);
-    __m256i p1 = _mm256_shuffle_epi8(bgr1, sh);
-    __m256i p2 = _mm256_shuffle_epi8(bgr2, sh);
-    __m256i p3 = _mm256_shuffle_epi8(bgr3, sh);
-
-    __m256i p01l = _mm256_unpacklo_epi32(p0, p1);
-    __m256i p01h = _mm256_unpackhi_epi32(p0, p1);
-    __m256i p23l = _mm256_unpacklo_epi32(p2, p3);
-    __m256i p23h = _mm256_unpackhi_epi32(p2, p3);
-
-    __m256i pll = _mm256_permute2x128_si256(p01l, p23l, 0 + 2*16);
-    __m256i plh = _mm256_permute2x128_si256(p01l, p23l, 1 + 3*16);
-    __m256i phl = _mm256_permute2x128_si256(p01h, p23h, 0 + 2*16);
-    __m256i phh = _mm256_permute2x128_si256(p01h, p23h, 1 + 3*16);
-
-    __m256i b0 = _mm256_unpacklo_epi32(pll, plh);
-    __m256i g0 = _mm256_unpackhi_epi32(pll, plh);
-    __m256i r0 = _mm256_unpacklo_epi32(phl, phh);
-    __m256i a0 = _mm256_unpackhi_epi32(phl, phh);
-
-    a = v_uint16x16(b0);
-    b = v_uint16x16(g0);
-    c = v_uint16x16(r0);
-    d = v_uint16x16(a0);
-}
-
-inline void v_load_deinterleave( const unsigned* ptr, v_uint32x8& a, v_uint32x8& b, v_uint32x8& c, v_uint32x8& d )
-{
-    __m256i p0 = _mm256_loadu_si256((const __m256i*)ptr);
-    __m256i p1 = _mm256_loadu_si256((const __m256i*)(ptr + 8));
-    __m256i p2 = _mm256_loadu_si256((const __m256i*)(ptr + 16));
-    __m256i p3 = _mm256_loadu_si256((const __m256i*)(ptr + 24));
-
-    __m256i p01l = _mm256_unpacklo_epi32(p0, p1);
-    __m256i p01h = _mm256_unpackhi_epi32(p0, p1);
-    __m256i p23l = _mm256_unpacklo_epi32(p2, p3);
-    __m256i p23h = _mm256_unpackhi_epi32(p2, p3);
-
-    __m256i pll = _mm256_permute2x128_si256(p01l, p23l, 0 + 2*16);
-    __m256i plh = _mm256_permute2x128_si256(p01l, p23l, 1 + 3*16);
-    __m256i phl = _mm256_permute2x128_si256(p01h, p23h, 0 + 2*16);
-    __m256i phh = _mm256_permute2x128_si256(p01h, p23h, 1 + 3*16);
-
-    __m256i b0 = _mm256_unpacklo_epi32(pll, plh);
-    __m256i g0 = _mm256_unpackhi_epi32(pll, plh);
-    __m256i r0 = _mm256_unpacklo_epi32(phl, phh);
-    __m256i a0 = _mm256_unpackhi_epi32(phl, phh);
-
-    a = v_uint32x8(b0);
-    b = v_uint32x8(g0);
-    c = v_uint32x8(r0);
-    d = v_uint32x8(a0);
-}
-
-inline void v_load_deinterleave( const uint64* ptr, v_uint64x4& a, v_uint64x4& b, v_uint64x4& c, v_uint64x4& d )
-{
-    __m256i bgra0 = _mm256_loadu_si256((const __m256i*)ptr);
-    __m256i bgra1 = _mm256_loadu_si256((const __m256i*)(ptr + 4));
-    __m256i bgra2 = _mm256_loadu_si256((const __m256i*)(ptr + 8));
-    __m256i bgra3 = _mm256_loadu_si256((const __m256i*)(ptr + 12));
-
-    __m256i l02 = _mm256_permute2x128_si256(bgra0, bgra2, 0 + 2*16);
-    __m256i h02 = _mm256_permute2x128_si256(bgra0, bgra2, 1 + 3*16);
-    __m256i l13 = _mm256_permute2x128_si256(bgra1, bgra3, 0 + 2*16);
-    __m256i h13 = _mm256_permute2x128_si256(bgra1, bgra3, 1 + 3*16);
-
-    __m256i b0 = _mm256_unpacklo_epi64(l02, l13);
-    __m256i g0 = _mm256_unpackhi_epi64(l02, l13);
-    __m256i r0 = _mm256_unpacklo_epi64(h02, h13);
-    __m256i a0 = _mm256_unpackhi_epi64(h02, h13);
-
-    a = v_uint64x4(b0);
-    b = v_uint64x4(g0);
-    c = v_uint64x4(r0);
-    d = v_uint64x4(a0);
-}
-
-///////////////////////////// store interleave /////////////////////////////////////
-
-inline void v_store_interleave( uchar* ptr, const v_uint8x32& x, const v_uint8x32& y,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    __m256i xy_l = _mm256_unpacklo_epi8(x.val, y.val);
-    __m256i xy_h = _mm256_unpackhi_epi8(x.val, y.val);
-
-    __m256i xy0 = _mm256_permute2x128_si256(xy_l, xy_h, 0 + 2*16);
-    __m256i xy1 = _mm256_permute2x128_si256(xy_l, xy_h, 1 + 3*16);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm256_stream_si256((__m256i*)ptr, xy0);
-        _mm256_stream_si256((__m256i*)(ptr + 32), xy1);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm256_store_si256((__m256i*)ptr, xy0);
-        _mm256_store_si256((__m256i*)(ptr + 32), xy1);
-    }
-    else
-    {
-        _mm256_storeu_si256((__m256i*)ptr, xy0);
-        _mm256_storeu_si256((__m256i*)(ptr + 32), xy1);
-    }
-}
-
-inline void v_store_interleave( ushort* ptr, const v_uint16x16& x, const v_uint16x16& y,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    __m256i xy_l = _mm256_unpacklo_epi16(x.val, y.val);
-    __m256i xy_h = _mm256_unpackhi_epi16(x.val, y.val);
-
-    __m256i xy0 = _mm256_permute2x128_si256(xy_l, xy_h, 0 + 2*16);
-    __m256i xy1 = _mm256_permute2x128_si256(xy_l, xy_h, 1 + 3*16);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm256_stream_si256((__m256i*)ptr, xy0);
-        _mm256_stream_si256((__m256i*)(ptr + 16), xy1);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm256_store_si256((__m256i*)ptr, xy0);
-        _mm256_store_si256((__m256i*)(ptr + 16), xy1);
-    }
-    else
-    {
-        _mm256_storeu_si256((__m256i*)ptr, xy0);
-        _mm256_storeu_si256((__m256i*)(ptr + 16), xy1);
-    }
-}
-
-inline void v_store_interleave( unsigned* ptr, const v_uint32x8& x, const v_uint32x8& y,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    __m256i xy_l = _mm256_unpacklo_epi32(x.val, y.val);
-    __m256i xy_h = _mm256_unpackhi_epi32(x.val, y.val);
-
-    __m256i xy0 = _mm256_permute2x128_si256(xy_l, xy_h, 0 + 2*16);
-    __m256i xy1 = _mm256_permute2x128_si256(xy_l, xy_h, 1 + 3*16);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm256_stream_si256((__m256i*)ptr, xy0);
-        _mm256_stream_si256((__m256i*)(ptr + 8), xy1);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm256_store_si256((__m256i*)ptr, xy0);
-        _mm256_store_si256((__m256i*)(ptr + 8), xy1);
-    }
-    else
-    {
-        _mm256_storeu_si256((__m256i*)ptr, xy0);
-        _mm256_storeu_si256((__m256i*)(ptr + 8), xy1);
-    }
-}
-
-inline void v_store_interleave( uint64* ptr, const v_uint64x4& x, const v_uint64x4& y,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    __m256i xy_l = _mm256_unpacklo_epi64(x.val, y.val);
-    __m256i xy_h = _mm256_unpackhi_epi64(x.val, y.val);
-
-    __m256i xy0 = _mm256_permute2x128_si256(xy_l, xy_h, 0 + 2*16);
-    __m256i xy1 = _mm256_permute2x128_si256(xy_l, xy_h, 1 + 3*16);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm256_stream_si256((__m256i*)ptr, xy0);
-        _mm256_stream_si256((__m256i*)(ptr + 4), xy1);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm256_store_si256((__m256i*)ptr, xy0);
-        _mm256_store_si256((__m256i*)(ptr + 4), xy1);
-    }
-    else
-    {
-        _mm256_storeu_si256((__m256i*)ptr, xy0);
-        _mm256_storeu_si256((__m256i*)(ptr + 4), xy1);
-    }
-}
-
-inline void v_store_interleave( uchar* ptr, const v_uint8x32& a, const v_uint8x32& b, const v_uint8x32& c,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    const __m256i sh_b = _mm256_setr_epi8(
-            0, 11, 6, 1, 12, 7, 2, 13, 8, 3, 14, 9, 4, 15, 10, 5,
-            0, 11, 6, 1, 12, 7, 2, 13, 8, 3, 14, 9, 4, 15, 10, 5);
-    const __m256i sh_g = _mm256_setr_epi8(
-            5, 0, 11, 6, 1, 12, 7, 2, 13, 8, 3, 14, 9, 4, 15, 10,
-            5, 0, 11, 6, 1, 12, 7, 2, 13, 8, 3, 14, 9, 4, 15, 10);
-    const __m256i sh_r = _mm256_setr_epi8(
-            10, 5, 0, 11, 6, 1, 12, 7, 2, 13, 8, 3, 14, 9, 4, 15,
-            10, 5, 0, 11, 6, 1, 12, 7, 2, 13, 8, 3, 14, 9, 4, 15);
-
-    __m256i b0 = _mm256_shuffle_epi8(a.val, sh_b);
-    __m256i g0 = _mm256_shuffle_epi8(b.val, sh_g);
-    __m256i r0 = _mm256_shuffle_epi8(c.val, sh_r);
-
-    const __m256i m0 = _mm256_setr_epi8(0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0,
-                                               0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0);
-    const __m256i m1 = _mm256_setr_epi8(0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0,
-                                               0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0);
-
-    __m256i p0 = _mm256_blendv_epi8(_mm256_blendv_epi8(b0, g0, m0), r0, m1);
-    __m256i p1 = _mm256_blendv_epi8(_mm256_blendv_epi8(g0, r0, m0), b0, m1);
-    __m256i p2 = _mm256_blendv_epi8(_mm256_blendv_epi8(r0, b0, m0), g0, m1);
-
-    __m256i bgr0 = _mm256_permute2x128_si256(p0, p1, 0 + 2*16);
-    __m256i bgr1 = _mm256_permute2x128_si256(p2, p0, 0 + 3*16);
-    __m256i bgr2 = _mm256_permute2x128_si256(p1, p2, 1 + 3*16);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm256_stream_si256((__m256i*)ptr, bgr0);
-        _mm256_stream_si256((__m256i*)(ptr + 32), bgr1);
-        _mm256_stream_si256((__m256i*)(ptr + 64), bgr2);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm256_store_si256((__m256i*)ptr, bgr0);
-        _mm256_store_si256((__m256i*)(ptr + 32), bgr1);
-        _mm256_store_si256((__m256i*)(ptr + 64), bgr2);
-    }
-    else
-    {
-        _mm256_storeu_si256((__m256i*)ptr, bgr0);
-        _mm256_storeu_si256((__m256i*)(ptr + 32), bgr1);
-        _mm256_storeu_si256((__m256i*)(ptr + 64), bgr2);
-    }
-}
-
-inline void v_store_interleave( ushort* ptr, const v_uint16x16& a, const v_uint16x16& b, const v_uint16x16& c,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    const __m256i sh_b = _mm256_setr_epi8(
-         0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15, 4, 5, 10, 11,
-         0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15, 4, 5, 10, 11);
-    const __m256i sh_g = _mm256_setr_epi8(
-         10, 11, 0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15, 4, 5,
-         10, 11, 0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15, 4, 5);
-    const __m256i sh_r = _mm256_setr_epi8(
-         4, 5, 10, 11, 0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15,
-         4, 5, 10, 11, 0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15);
-
-    __m256i b0 = _mm256_shuffle_epi8(a.val, sh_b);
-    __m256i g0 = _mm256_shuffle_epi8(b.val, sh_g);
-    __m256i r0 = _mm256_shuffle_epi8(c.val, sh_r);
-
-    const __m256i m0 = _mm256_setr_epi8(0, 0, -1, -1, 0, 0, 0, 0, -1, -1, 0, 0, 0, 0, -1, -1,
-                                               0, 0, 0, 0, -1, -1, 0, 0, 0, 0, -1, -1, 0, 0, 0, 0);
-    const __m256i m1 = _mm256_setr_epi8(0, 0, 0, 0, -1, -1, 0, 0, 0, 0, -1, -1, 0, 0, 0, 0,
-                                               -1, -1, 0, 0, 0, 0, -1, -1, 0, 0, 0, 0, -1, -1, 0, 0);
-
-    __m256i p0 = _mm256_blendv_epi8(_mm256_blendv_epi8(b0, g0, m0), r0, m1);
-    __m256i p1 = _mm256_blendv_epi8(_mm256_blendv_epi8(g0, r0, m0), b0, m1);
-    __m256i p2 = _mm256_blendv_epi8(_mm256_blendv_epi8(r0, b0, m0), g0, m1);
-
-    __m256i bgr0 = _mm256_permute2x128_si256(p0, p2, 0 + 2*16);
-    //__m256i bgr1 = p1;
-    __m256i bgr2 = _mm256_permute2x128_si256(p0, p2, 1 + 3*16);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm256_stream_si256((__m256i*)ptr, bgr0);
-        _mm256_stream_si256((__m256i*)(ptr + 16), p1);
-        _mm256_stream_si256((__m256i*)(ptr + 32), bgr2);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm256_store_si256((__m256i*)ptr, bgr0);
-        _mm256_store_si256((__m256i*)(ptr + 16), p1);
-        _mm256_store_si256((__m256i*)(ptr + 32), bgr2);
-    }
-    else
-    {
-        _mm256_storeu_si256((__m256i*)ptr, bgr0);
-        _mm256_storeu_si256((__m256i*)(ptr + 16), p1);
-        _mm256_storeu_si256((__m256i*)(ptr + 32), bgr2);
-    }
-}
-
-inline void v_store_interleave( unsigned* ptr, const v_uint32x8& a, const v_uint32x8& b, const v_uint32x8& c,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    __m256i b0 = _mm256_shuffle_epi32(a.val, 0x6c);
-    __m256i g0 = _mm256_shuffle_epi32(b.val, 0xb1);
-    __m256i r0 = _mm256_shuffle_epi32(c.val, 0xc6);
-
-    __m256i p0 = _mm256_blend_epi32(_mm256_blend_epi32(b0, g0, 0x92), r0, 0x24);
-    __m256i p1 = _mm256_blend_epi32(_mm256_blend_epi32(g0, r0, 0x92), b0, 0x24);
-    __m256i p2 = _mm256_blend_epi32(_mm256_blend_epi32(r0, b0, 0x92), g0, 0x24);
-
-    __m256i bgr0 = _mm256_permute2x128_si256(p0, p1, 0 + 2*16);
-    //__m256i bgr1 = p2;
-    __m256i bgr2 = _mm256_permute2x128_si256(p0, p1, 1 + 3*16);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm256_stream_si256((__m256i*)ptr, bgr0);
-        _mm256_stream_si256((__m256i*)(ptr + 8), p2);
-        _mm256_stream_si256((__m256i*)(ptr + 16), bgr2);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm256_store_si256((__m256i*)ptr, bgr0);
-        _mm256_store_si256((__m256i*)(ptr + 8), p2);
-        _mm256_store_si256((__m256i*)(ptr + 16), bgr2);
-    }
-    else
-    {
-        _mm256_storeu_si256((__m256i*)ptr, bgr0);
-        _mm256_storeu_si256((__m256i*)(ptr + 8), p2);
-        _mm256_storeu_si256((__m256i*)(ptr + 16), bgr2);
-    }
-}
-
-inline void v_store_interleave( uint64* ptr, const v_uint64x4& a, const v_uint64x4& b, const v_uint64x4& c,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    __m256i s01 = _mm256_unpacklo_epi64(a.val, b.val);
-    __m256i s12 = _mm256_unpackhi_epi64(b.val, c.val);
-    __m256i s20 = _mm256_blend_epi32(c.val, a.val, 0xcc);
-
-    __m256i bgr0 = _mm256_permute2x128_si256(s01, s20, 0 + 2*16);
-    __m256i bgr1 = _mm256_blend_epi32(s01, s12, 0x0f);
-    __m256i bgr2 = _mm256_permute2x128_si256(s20, s12, 1 + 3*16);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm256_stream_si256((__m256i*)ptr, bgr0);
-        _mm256_stream_si256((__m256i*)(ptr + 4), bgr1);
-        _mm256_stream_si256((__m256i*)(ptr + 8), bgr2);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm256_store_si256((__m256i*)ptr, bgr0);
-        _mm256_store_si256((__m256i*)(ptr + 4), bgr1);
-        _mm256_store_si256((__m256i*)(ptr + 8), bgr2);
-    }
-    else
-    {
-        _mm256_storeu_si256((__m256i*)ptr, bgr0);
-        _mm256_storeu_si256((__m256i*)(ptr + 4), bgr1);
-        _mm256_storeu_si256((__m256i*)(ptr + 8), bgr2);
-    }
-}
-
-inline void v_store_interleave( uchar* ptr, const v_uint8x32& a, const v_uint8x32& b,
-                                const v_uint8x32& c, const v_uint8x32& d,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    __m256i bg0 = _mm256_unpacklo_epi8(a.val, b.val);
-    __m256i bg1 = _mm256_unpackhi_epi8(a.val, b.val);
-    __m256i ra0 = _mm256_unpacklo_epi8(c.val, d.val);
-    __m256i ra1 = _mm256_unpackhi_epi8(c.val, d.val);
-
-    __m256i bgra0_ = _mm256_unpacklo_epi16(bg0, ra0);
-    __m256i bgra1_ = _mm256_unpackhi_epi16(bg0, ra0);
-    __m256i bgra2_ = _mm256_unpacklo_epi16(bg1, ra1);
-    __m256i bgra3_ = _mm256_unpackhi_epi16(bg1, ra1);
-
-    __m256i bgra0 = _mm256_permute2x128_si256(bgra0_, bgra1_, 0 + 2*16);
-    __m256i bgra2 = _mm256_permute2x128_si256(bgra0_, bgra1_, 1 + 3*16);
-    __m256i bgra1 = _mm256_permute2x128_si256(bgra2_, bgra3_, 0 + 2*16);
-    __m256i bgra3 = _mm256_permute2x128_si256(bgra2_, bgra3_, 1 + 3*16);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm256_stream_si256((__m256i*)ptr, bgra0);
-        _mm256_stream_si256((__m256i*)(ptr + 32), bgra1);
-        _mm256_stream_si256((__m256i*)(ptr + 64), bgra2);
-        _mm256_stream_si256((__m256i*)(ptr + 96), bgra3);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm256_store_si256((__m256i*)ptr, bgra0);
-        _mm256_store_si256((__m256i*)(ptr + 32), bgra1);
-        _mm256_store_si256((__m256i*)(ptr + 64), bgra2);
-        _mm256_store_si256((__m256i*)(ptr + 96), bgra3);
-    }
-    else
-    {
-        _mm256_storeu_si256((__m256i*)ptr, bgra0);
-        _mm256_storeu_si256((__m256i*)(ptr + 32), bgra1);
-        _mm256_storeu_si256((__m256i*)(ptr + 64), bgra2);
-        _mm256_storeu_si256((__m256i*)(ptr + 96), bgra3);
-    }
-}
-
-inline void v_store_interleave( ushort* ptr, const v_uint16x16& a, const v_uint16x16& b,
-                                const v_uint16x16& c, const v_uint16x16& d,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    __m256i bg0 = _mm256_unpacklo_epi16(a.val, b.val);
-    __m256i bg1 = _mm256_unpackhi_epi16(a.val, b.val);
-    __m256i ra0 = _mm256_unpacklo_epi16(c.val, d.val);
-    __m256i ra1 = _mm256_unpackhi_epi16(c.val, d.val);
-
-    __m256i bgra0_ = _mm256_unpacklo_epi32(bg0, ra0);
-    __m256i bgra1_ = _mm256_unpackhi_epi32(bg0, ra0);
-    __m256i bgra2_ = _mm256_unpacklo_epi32(bg1, ra1);
-    __m256i bgra3_ = _mm256_unpackhi_epi32(bg1, ra1);
-
-    __m256i bgra0 = _mm256_permute2x128_si256(bgra0_, bgra1_, 0 + 2*16);
-    __m256i bgra2 = _mm256_permute2x128_si256(bgra0_, bgra1_, 1 + 3*16);
-    __m256i bgra1 = _mm256_permute2x128_si256(bgra2_, bgra3_, 0 + 2*16);
-    __m256i bgra3 = _mm256_permute2x128_si256(bgra2_, bgra3_, 1 + 3*16);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm256_stream_si256((__m256i*)ptr, bgra0);
-        _mm256_stream_si256((__m256i*)(ptr + 16), bgra1);
-        _mm256_stream_si256((__m256i*)(ptr + 32), bgra2);
-        _mm256_stream_si256((__m256i*)(ptr + 48), bgra3);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm256_store_si256((__m256i*)ptr, bgra0);
-        _mm256_store_si256((__m256i*)(ptr + 16), bgra1);
-        _mm256_store_si256((__m256i*)(ptr + 32), bgra2);
-        _mm256_store_si256((__m256i*)(ptr + 48), bgra3);
-    }
-    else
-    {
-        _mm256_storeu_si256((__m256i*)ptr, bgra0);
-        _mm256_storeu_si256((__m256i*)(ptr + 16), bgra1);
-        _mm256_storeu_si256((__m256i*)(ptr + 32), bgra2);
-        _mm256_storeu_si256((__m256i*)(ptr + 48), bgra3);
-    }
-}
-
-inline void v_store_interleave( unsigned* ptr, const v_uint32x8& a, const v_uint32x8& b,
-                                const v_uint32x8& c, const v_uint32x8& d,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    __m256i bg0 = _mm256_unpacklo_epi32(a.val, b.val);
-    __m256i bg1 = _mm256_unpackhi_epi32(a.val, b.val);
-    __m256i ra0 = _mm256_unpacklo_epi32(c.val, d.val);
-    __m256i ra1 = _mm256_unpackhi_epi32(c.val, d.val);
-
-    __m256i bgra0_ = _mm256_unpacklo_epi64(bg0, ra0);
-    __m256i bgra1_ = _mm256_unpackhi_epi64(bg0, ra0);
-    __m256i bgra2_ = _mm256_unpacklo_epi64(bg1, ra1);
-    __m256i bgra3_ = _mm256_unpackhi_epi64(bg1, ra1);
-
-    __m256i bgra0 = _mm256_permute2x128_si256(bgra0_, bgra1_, 0 + 2*16);
-    __m256i bgra2 = _mm256_permute2x128_si256(bgra0_, bgra1_, 1 + 3*16);
-    __m256i bgra1 = _mm256_permute2x128_si256(bgra2_, bgra3_, 0 + 2*16);
-    __m256i bgra3 = _mm256_permute2x128_si256(bgra2_, bgra3_, 1 + 3*16);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm256_stream_si256((__m256i*)ptr, bgra0);
-        _mm256_stream_si256((__m256i*)(ptr + 8), bgra1);
-        _mm256_stream_si256((__m256i*)(ptr + 16), bgra2);
-        _mm256_stream_si256((__m256i*)(ptr + 24), bgra3);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm256_store_si256((__m256i*)ptr, bgra0);
-        _mm256_store_si256((__m256i*)(ptr + 8), bgra1);
-        _mm256_store_si256((__m256i*)(ptr + 16), bgra2);
-        _mm256_store_si256((__m256i*)(ptr + 24), bgra3);
-    }
-    else
-    {
-        _mm256_storeu_si256((__m256i*)ptr, bgra0);
-        _mm256_storeu_si256((__m256i*)(ptr + 8), bgra1);
-        _mm256_storeu_si256((__m256i*)(ptr + 16), bgra2);
-        _mm256_storeu_si256((__m256i*)(ptr + 24), bgra3);
-    }
-}
-
-inline void v_store_interleave( uint64* ptr, const v_uint64x4& a, const v_uint64x4& b,
-                                const v_uint64x4& c, const v_uint64x4& d,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    __m256i bg0 = _mm256_unpacklo_epi64(a.val, b.val);
-    __m256i bg1 = _mm256_unpackhi_epi64(a.val, b.val);
-    __m256i ra0 = _mm256_unpacklo_epi64(c.val, d.val);
-    __m256i ra1 = _mm256_unpackhi_epi64(c.val, d.val);
-
-    __m256i bgra0 = _mm256_permute2x128_si256(bg0, ra0, 0 + 2*16);
-    __m256i bgra1 = _mm256_permute2x128_si256(bg1, ra1, 0 + 2*16);
-    __m256i bgra2 = _mm256_permute2x128_si256(bg0, ra0, 1 + 3*16);
-    __m256i bgra3 = _mm256_permute2x128_si256(bg1, ra1, 1 + 3*16);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm256_stream_si256((__m256i*)ptr, bgra0);
-        _mm256_stream_si256((__m256i*)(ptr + 4), bgra1);
-        _mm256_stream_si256((__m256i*)(ptr + 8), bgra2);
-        _mm256_stream_si256((__m256i*)(ptr + 12), bgra3);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm256_store_si256((__m256i*)ptr, bgra0);
-        _mm256_store_si256((__m256i*)(ptr + 4), bgra1);
-        _mm256_store_si256((__m256i*)(ptr + 8), bgra2);
-        _mm256_store_si256((__m256i*)(ptr + 12), bgra3);
-    }
-    else
-    {
-        _mm256_storeu_si256((__m256i*)ptr, bgra0);
-        _mm256_storeu_si256((__m256i*)(ptr + 4), bgra1);
-        _mm256_storeu_si256((__m256i*)(ptr + 8), bgra2);
-        _mm256_storeu_si256((__m256i*)(ptr + 12), bgra3);
-    }
-}
-
-#define OPENCV_HAL_IMPL_AVX_LOADSTORE_INTERLEAVE(_Tpvec0, _Tp0, suffix0, _Tpvec1, _Tp1, suffix1) \
-inline void v_load_deinterleave( const _Tp0* ptr, _Tpvec0& a0, _Tpvec0& b0 ) \
-{ \
-    _Tpvec1 a1, b1; \
-    v_load_deinterleave((const _Tp1*)ptr, a1, b1); \
-    a0 = v_reinterpret_as_##suffix0(a1); \
-    b0 = v_reinterpret_as_##suffix0(b1); \
-} \
-inline void v_load_deinterleave( const _Tp0* ptr, _Tpvec0& a0, _Tpvec0& b0, _Tpvec0& c0 ) \
-{ \
-    _Tpvec1 a1, b1, c1; \
-    v_load_deinterleave((const _Tp1*)ptr, a1, b1, c1); \
-    a0 = v_reinterpret_as_##suffix0(a1); \
-    b0 = v_reinterpret_as_##suffix0(b1); \
-    c0 = v_reinterpret_as_##suffix0(c1); \
-} \
-inline void v_load_deinterleave( const _Tp0* ptr, _Tpvec0& a0, _Tpvec0& b0, _Tpvec0& c0, _Tpvec0& d0 ) \
-{ \
-    _Tpvec1 a1, b1, c1, d1; \
-    v_load_deinterleave((const _Tp1*)ptr, a1, b1, c1, d1); \
-    a0 = v_reinterpret_as_##suffix0(a1); \
-    b0 = v_reinterpret_as_##suffix0(b1); \
-    c0 = v_reinterpret_as_##suffix0(c1); \
-    d0 = v_reinterpret_as_##suffix0(d1); \
-} \
-inline void v_store_interleave( _Tp0* ptr, const _Tpvec0& a0, const _Tpvec0& b0, \
-                                hal::StoreMode mode=hal::STORE_UNALIGNED ) \
-{ \
-    _Tpvec1 a1 = v_reinterpret_as_##suffix1(a0); \
-    _Tpvec1 b1 = v_reinterpret_as_##suffix1(b0); \
-    v_store_interleave((_Tp1*)ptr, a1, b1, mode);      \
-} \
-inline void v_store_interleave( _Tp0* ptr, const _Tpvec0& a0, const _Tpvec0& b0, const _Tpvec0& c0, \
-                                hal::StoreMode mode=hal::STORE_UNALIGNED ) \
-{ \
-    _Tpvec1 a1 = v_reinterpret_as_##suffix1(a0); \
-    _Tpvec1 b1 = v_reinterpret_as_##suffix1(b0); \
-    _Tpvec1 c1 = v_reinterpret_as_##suffix1(c0); \
-    v_store_interleave((_Tp1*)ptr, a1, b1, c1, mode);  \
-} \
-inline void v_store_interleave( _Tp0* ptr, const _Tpvec0& a0, const _Tpvec0& b0, \
-                                const _Tpvec0& c0, const _Tpvec0& d0, \
-                                hal::StoreMode mode=hal::STORE_UNALIGNED ) \
-{ \
-    _Tpvec1 a1 = v_reinterpret_as_##suffix1(a0); \
-    _Tpvec1 b1 = v_reinterpret_as_##suffix1(b0); \
-    _Tpvec1 c1 = v_reinterpret_as_##suffix1(c0); \
-    _Tpvec1 d1 = v_reinterpret_as_##suffix1(d0); \
-    v_store_interleave((_Tp1*)ptr, a1, b1, c1, d1, mode); \
-}
-
-OPENCV_HAL_IMPL_AVX_LOADSTORE_INTERLEAVE(v_int8x32, schar, s8, v_uint8x32, uchar, u8)
-OPENCV_HAL_IMPL_AVX_LOADSTORE_INTERLEAVE(v_int16x16, short, s16, v_uint16x16, ushort, u16)
-OPENCV_HAL_IMPL_AVX_LOADSTORE_INTERLEAVE(v_int32x8, int, s32, v_uint32x8, unsigned, u32)
-OPENCV_HAL_IMPL_AVX_LOADSTORE_INTERLEAVE(v_float32x8, float, f32, v_uint32x8, unsigned, u32)
-OPENCV_HAL_IMPL_AVX_LOADSTORE_INTERLEAVE(v_int64x4, int64, s64, v_uint64x4, uint64, u64)
-OPENCV_HAL_IMPL_AVX_LOADSTORE_INTERLEAVE(v_float64x4, double, f64, v_uint64x4, uint64, u64)
-
-//
-// FP16
-//
-
-inline v_float32x8 v256_load_expand(const float16_t* ptr)
-{
-#if CV_FP16
-    return v_float32x8(_mm256_cvtph_ps(_mm_loadu_si128((const __m128i*)ptr)));
-#else
-    float CV_DECL_ALIGNED(32) buf[8];
-    for (int i = 0; i < 8; i++)
-        buf[i] = (float)ptr[i];
-    return v256_load_aligned(buf);
-#endif
-}
-
-inline void v_pack_store(float16_t* ptr, const v_float32x8& a)
-{
-#if CV_FP16
-    __m128i ah = _mm256_cvtps_ph(a.val, 0);
-    _mm_storeu_si128((__m128i*)ptr, ah);
-#else
-    float CV_DECL_ALIGNED(32) buf[8];
-    v_store_aligned(buf, a);
-    for (int i = 0; i < 8; i++)
-        ptr[i] = float16_t(buf[i]);
-#endif
-}
-
-//
-// end of FP16
-//
-
-inline void v256_cleanup() { _mm256_zeroall(); }
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END
-
-//! @endcond
-
-} // cv::
-
-#endif // OPENCV_HAL_INTRIN_AVX_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_avx512.hpp b/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_avx512.hpp
deleted file mode 100644
index d20d6dd1..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_avx512.hpp
+++ /dev/null
@@ -1,3090 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html
-
-#ifndef OPENCV_HAL_INTRIN_AVX512_HPP
-#define OPENCV_HAL_INTRIN_AVX512_HPP
-
-#if defined(_MSC_VER) && (_MSC_VER < 1920/*MSVS2019*/)
-# pragma warning(disable:4146)  // unary minus operator applied to unsigned type, result still unsigned
-# pragma warning(disable:4309)  // 'argument': truncation of constant value
-# pragma warning(disable:4310)  // cast truncates constant value
-#endif
-
-#define CVT_ROUND_MODES_IMPLEMENTED 0
-
-#define CV_SIMD512 1
-#define CV_SIMD512_64F 1
-#define CV_SIMD512_FP16 0  // no native operations with FP16 type. Only load/store from float32x8 are available (if CV_FP16 == 1)
-
-#define _v512_set_epu64(a7, a6, a5, a4, a3, a2, a1, a0) _mm512_set_epi64((int64)(a7),(int64)(a6),(int64)(a5),(int64)(a4),(int64)(a3),(int64)(a2),(int64)(a1),(int64)(a0))
-#define _v512_set_epu32(a15, a14, a13, a12, a11, a10,  a9,  a8,  a7,  a6,  a5,  a4,  a3,  a2,  a1,  a0) \
-        _mm512_set_epi64(((int64)(a15)<<32)|(int64)(a14), ((int64)(a13)<<32)|(int64)(a12), ((int64)(a11)<<32)|(int64)(a10), ((int64)( a9)<<32)|(int64)( a8), \
-                         ((int64)( a7)<<32)|(int64)( a6), ((int64)( a5)<<32)|(int64)( a4), ((int64)( a3)<<32)|(int64)( a2), ((int64)( a1)<<32)|(int64)( a0))
-#define _v512_set_epu16(a31, a30, a29, a28, a27, a26, a25, a24, a23, a22, a21, a20, a19, a18, a17, a16, \
-                        a15, a14, a13, a12, a11, a10,  a9,  a8,  a7,  a6,  a5,  a4,  a3,  a2,  a1,  a0) \
-        _v512_set_epu32(((unsigned)(a31)<<16)|(unsigned)(a30), ((unsigned)(a29)<<16)|(unsigned)(a28), ((unsigned)(a27)<<16)|(unsigned)(a26), ((unsigned)(a25)<<16)|(unsigned)(a24), \
-                        ((unsigned)(a23)<<16)|(unsigned)(a22), ((unsigned)(a21)<<16)|(unsigned)(a20), ((unsigned)(a19)<<16)|(unsigned)(a18), ((unsigned)(a17)<<16)|(unsigned)(a16), \
-                        ((unsigned)(a15)<<16)|(unsigned)(a14), ((unsigned)(a13)<<16)|(unsigned)(a12), ((unsigned)(a11)<<16)|(unsigned)(a10), ((unsigned)( a9)<<16)|(unsigned)( a8), \
-                        ((unsigned)( a7)<<16)|(unsigned)( a6), ((unsigned)( a5)<<16)|(unsigned)( a4), ((unsigned)( a3)<<16)|(unsigned)( a2), ((unsigned)( a1)<<16)|(unsigned)( a0))
-#define _v512_set_epu8(a63, a62, a61, a60, a59, a58, a57, a56, a55, a54, a53, a52, a51, a50, a49, a48, \
-                       a47, a46, a45, a44, a43, a42, a41, a40, a39, a38, a37, a36, a35, a34, a33, a32, \
-                       a31, a30, a29, a28, a27, a26, a25, a24, a23, a22, a21, a20, a19, a18, a17, a16, \
-                       a15, a14, a13, a12, a11, a10,  a9,  a8,  a7,  a6,  a5,  a4,  a3,  a2,  a1,  a0) \
-        _v512_set_epu32(((unsigned)(a63)<<24)|((unsigned)(a62)<<16)|((unsigned)(a61)<<8)|(unsigned)(a60),((unsigned)(a59)<<24)|((unsigned)(a58)<<16)|((unsigned)(a57)<<8)|(unsigned)(a56), \
-                        ((unsigned)(a55)<<24)|((unsigned)(a54)<<16)|((unsigned)(a53)<<8)|(unsigned)(a52),((unsigned)(a51)<<24)|((unsigned)(a50)<<16)|((unsigned)(a49)<<8)|(unsigned)(a48), \
-                        ((unsigned)(a47)<<24)|((unsigned)(a46)<<16)|((unsigned)(a45)<<8)|(unsigned)(a44),((unsigned)(a43)<<24)|((unsigned)(a42)<<16)|((unsigned)(a41)<<8)|(unsigned)(a40), \
-                        ((unsigned)(a39)<<24)|((unsigned)(a38)<<16)|((unsigned)(a37)<<8)|(unsigned)(a36),((unsigned)(a35)<<24)|((unsigned)(a34)<<16)|((unsigned)(a33)<<8)|(unsigned)(a32), \
-                        ((unsigned)(a31)<<24)|((unsigned)(a30)<<16)|((unsigned)(a29)<<8)|(unsigned)(a28),((unsigned)(a27)<<24)|((unsigned)(a26)<<16)|((unsigned)(a25)<<8)|(unsigned)(a24), \
-                        ((unsigned)(a23)<<24)|((unsigned)(a22)<<16)|((unsigned)(a21)<<8)|(unsigned)(a20),((unsigned)(a19)<<24)|((unsigned)(a18)<<16)|((unsigned)(a17)<<8)|(unsigned)(a16), \
-                        ((unsigned)(a15)<<24)|((unsigned)(a14)<<16)|((unsigned)(a13)<<8)|(unsigned)(a12),((unsigned)(a11)<<24)|((unsigned)(a10)<<16)|((unsigned)( a9)<<8)|(unsigned)( a8), \
-                        ((unsigned)( a7)<<24)|((unsigned)( a6)<<16)|((unsigned)( a5)<<8)|(unsigned)( a4),((unsigned)( a3)<<24)|((unsigned)( a2)<<16)|((unsigned)( a1)<<8)|(unsigned)( a0))
-#define _v512_set_epi8(a63, a62, a61, a60, a59, a58, a57, a56, a55, a54, a53, a52, a51, a50, a49, a48, \
-                       a47, a46, a45, a44, a43, a42, a41, a40, a39, a38, a37, a36, a35, a34, a33, a32, \
-                       a31, a30, a29, a28, a27, a26, a25, a24, a23, a22, a21, a20, a19, a18, a17, a16, \
-                       a15, a14, a13, a12, a11, a10,  a9,  a8,  a7,  a6,  a5,  a4,  a3,  a2,  a1,  a0) \
-        _v512_set_epu8((uchar)(a63), (uchar)(a62), (uchar)(a61), (uchar)(a60), (uchar)(a59), (uchar)(a58), (uchar)(a57), (uchar)(a56), \
-                       (uchar)(a55), (uchar)(a54), (uchar)(a53), (uchar)(a52), (uchar)(a51), (uchar)(a50), (uchar)(a49), (uchar)(a48), \
-                       (uchar)(a47), (uchar)(a46), (uchar)(a45), (uchar)(a44), (uchar)(a43), (uchar)(a42), (uchar)(a41), (uchar)(a40), \
-                       (uchar)(a39), (uchar)(a38), (uchar)(a37), (uchar)(a36), (uchar)(a35), (uchar)(a34), (uchar)(a33), (uchar)(a32), \
-                       (uchar)(a31), (uchar)(a30), (uchar)(a29), (uchar)(a28), (uchar)(a27), (uchar)(a26), (uchar)(a25), (uchar)(a24), \
-                       (uchar)(a23), (uchar)(a22), (uchar)(a21), (uchar)(a20), (uchar)(a19), (uchar)(a18), (uchar)(a17), (uchar)(a16), \
-                       (uchar)(a15), (uchar)(a14), (uchar)(a13), (uchar)(a12), (uchar)(a11), (uchar)(a10), (uchar)( a9), (uchar)( a8), \
-                       (uchar)( a7), (uchar)( a6), (uchar)( a5), (uchar)( a4), (uchar)( a3), (uchar)( a2), (uchar)( a1), (uchar)( a0))
-
-#ifndef _mm512_cvtpd_pslo
-#ifdef _mm512_zextsi256_si512
-#define _mm512_cvtpd_pslo(a) _mm512_zextps256_ps512(_mm512_cvtpd_ps(a))
-#else
-//if preferred way to extend with zeros is unavailable
-#define _mm512_cvtpd_pslo(a) _mm512_castps256_ps512(_mm512_cvtpd_ps(a))
-#endif
-#endif
-///////// Utils ////////////
-
-namespace
-{
-
-inline __m512i _v512_combine(const __m256i& lo, const __m256i& hi)
-{ return _mm512_inserti32x8(_mm512_castsi256_si512(lo), hi, 1); }
-
-inline __m512 _v512_combine(const __m256& lo, const __m256& hi)
-{ return _mm512_insertf32x8(_mm512_castps256_ps512(lo), hi, 1); }
-
-inline __m512d _v512_combine(const __m256d& lo, const __m256d& hi)
-{ return _mm512_insertf64x4(_mm512_castpd256_pd512(lo), hi, 1); }
-
-inline int _v_cvtsi512_si32(const __m512i& a)
-{ return _mm_cvtsi128_si32(_mm512_castsi512_si128(a)); }
-
-inline __m256i _v512_extract_high(const __m512i& v)
-{ return _mm512_extracti32x8_epi32(v, 1); }
-
-inline __m256  _v512_extract_high(const __m512& v)
-{ return _mm512_extractf32x8_ps(v, 1); }
-
-inline __m256d _v512_extract_high(const __m512d& v)
-{ return _mm512_extractf64x4_pd(v, 1); }
-
-inline __m256i _v512_extract_low(const __m512i& v)
-{ return _mm512_castsi512_si256(v); }
-
-inline __m256  _v512_extract_low(const __m512& v)
-{ return _mm512_castps512_ps256(v); }
-
-inline __m256d _v512_extract_low(const __m512d& v)
-{ return _mm512_castpd512_pd256(v); }
-
-inline __m512i _v512_insert(const __m512i& a, const __m256i& b)
-{ return _mm512_inserti32x8(a, b, 0); }
-
-inline __m512 _v512_insert(const __m512& a, const __m256& b)
-{ return _mm512_insertf32x8(a, b, 0); }
-
-inline __m512d _v512_insert(const __m512d& a, const __m256d& b)
-{ return _mm512_insertf64x4(a, b, 0); }
-
-}
-
-namespace cv
-{
-
-//! @cond IGNORED
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
-
-///////// Types ////////////
-
-struct v_uint8x64
-{
-    typedef uchar lane_type;
-    enum { nlanes = 64 };
-    __m512i val;
-
-    explicit v_uint8x64(__m512i v) : val(v) {}
-    v_uint8x64(uchar v0,  uchar v1,  uchar v2,  uchar v3,
-               uchar v4,  uchar v5,  uchar v6,  uchar v7,
-               uchar v8,  uchar v9,  uchar v10, uchar v11,
-               uchar v12, uchar v13, uchar v14, uchar v15,
-               uchar v16, uchar v17, uchar v18, uchar v19,
-               uchar v20, uchar v21, uchar v22, uchar v23,
-               uchar v24, uchar v25, uchar v26, uchar v27,
-               uchar v28, uchar v29, uchar v30, uchar v31,
-               uchar v32, uchar v33, uchar v34, uchar v35,
-               uchar v36, uchar v37, uchar v38, uchar v39,
-               uchar v40, uchar v41, uchar v42, uchar v43,
-               uchar v44, uchar v45, uchar v46, uchar v47,
-               uchar v48, uchar v49, uchar v50, uchar v51,
-               uchar v52, uchar v53, uchar v54, uchar v55,
-               uchar v56, uchar v57, uchar v58, uchar v59,
-               uchar v60, uchar v61, uchar v62, uchar v63)
-    {
-        val = _v512_set_epu8(v63, v62, v61, v60, v59, v58, v57, v56, v55, v54, v53, v52, v51, v50, v49, v48,
-                             v47, v46, v45, v44, v43, v42, v41, v40, v39, v38, v37, v36, v35, v34, v33, v32,
-                             v31, v30, v29, v28, v27, v26, v25, v24, v23, v22, v21, v20, v19, v18, v17, v16,
-                             v15, v14, v13, v12, v11, v10, v9,  v8,  v7,  v6,  v5,  v4,  v3,  v2,  v1,  v0);
-    }
-    v_uint8x64() {}
-
-    static inline v_uint8x64 zero() { return v_uint8x64(_mm512_setzero_si512()); }
-
-    uchar get0() const { return (uchar)_v_cvtsi512_si32(val); }
-};
-
-struct v_int8x64
-{
-    typedef schar lane_type;
-    enum { nlanes = 64 };
-    __m512i val;
-
-    explicit v_int8x64(__m512i v) : val(v) {}
-    v_int8x64(schar v0,  schar v1,  schar v2,  schar v3,
-              schar v4,  schar v5,  schar v6,  schar v7,
-              schar v8,  schar v9,  schar v10, schar v11,
-              schar v12, schar v13, schar v14, schar v15,
-              schar v16, schar v17, schar v18, schar v19,
-              schar v20, schar v21, schar v22, schar v23,
-              schar v24, schar v25, schar v26, schar v27,
-              schar v28, schar v29, schar v30, schar v31,
-              schar v32, schar v33, schar v34, schar v35,
-              schar v36, schar v37, schar v38, schar v39,
-              schar v40, schar v41, schar v42, schar v43,
-              schar v44, schar v45, schar v46, schar v47,
-              schar v48, schar v49, schar v50, schar v51,
-              schar v52, schar v53, schar v54, schar v55,
-              schar v56, schar v57, schar v58, schar v59,
-              schar v60, schar v61, schar v62, schar v63)
-    {
-        val = _v512_set_epi8(v63, v62, v61, v60, v59, v58, v57, v56, v55, v54, v53, v52, v51, v50, v49, v48,
-                             v47, v46, v45, v44, v43, v42, v41, v40, v39, v38, v37, v36, v35, v34, v33, v32,
-                             v31, v30, v29, v28, v27, v26, v25, v24, v23, v22, v21, v20, v19, v18, v17, v16,
-                             v15, v14, v13, v12, v11, v10, v9,  v8,  v7,  v6,  v5,  v4,  v3,  v2,  v1,  v0);
-    }
-    v_int8x64() {}
-
-    static inline v_int8x64 zero() { return v_int8x64(_mm512_setzero_si512()); }
-
-    schar get0() const { return (schar)_v_cvtsi512_si32(val); }
-};
-
-struct v_uint16x32
-{
-    typedef ushort lane_type;
-    enum { nlanes = 32 };
-    __m512i val;
-
-    explicit v_uint16x32(__m512i v) : val(v) {}
-    v_uint16x32(ushort v0,  ushort v1,  ushort v2,  ushort v3,
-                ushort v4,  ushort v5,  ushort v6,  ushort v7,
-                ushort v8,  ushort v9,  ushort v10, ushort v11,
-                ushort v12, ushort v13, ushort v14, ushort v15,
-                ushort v16, ushort v17, ushort v18, ushort v19,
-                ushort v20, ushort v21, ushort v22, ushort v23,
-                ushort v24, ushort v25, ushort v26, ushort v27,
-                ushort v28, ushort v29, ushort v30, ushort v31)
-    {
-        val = _v512_set_epu16(v31, v30, v29, v28, v27, v26, v25, v24, v23, v22, v21, v20, v19, v18, v17, v16,
-                              v15, v14, v13, v12, v11, v10, v9,  v8,  v7,  v6,  v5,  v4,  v3,  v2,  v1,  v0);
-    }
-    v_uint16x32() {}
-
-    static inline v_uint16x32 zero() { return v_uint16x32(_mm512_setzero_si512()); }
-
-    ushort get0() const { return (ushort)_v_cvtsi512_si32(val); }
-};
-
-struct v_int16x32
-{
-    typedef short lane_type;
-    enum { nlanes = 32 };
-    __m512i val;
-
-    explicit v_int16x32(__m512i v) : val(v) {}
-    v_int16x32(short v0,  short v1,  short v2,  short v3,  short v4,  short v5,  short v6,  short v7,
-               short v8,  short v9,  short v10, short v11, short v12, short v13, short v14, short v15,
-               short v16, short v17, short v18, short v19, short v20, short v21, short v22, short v23,
-               short v24, short v25, short v26, short v27, short v28, short v29, short v30, short v31)
-    {
-        val = _v512_set_epu16((ushort)v31, (ushort)v30, (ushort)v29, (ushort)v28, (ushort)v27, (ushort)v26, (ushort)v25, (ushort)v24,
-                              (ushort)v23, (ushort)v22, (ushort)v21, (ushort)v20, (ushort)v19, (ushort)v18, (ushort)v17, (ushort)v16,
-                              (ushort)v15, (ushort)v14, (ushort)v13, (ushort)v12, (ushort)v11, (ushort)v10, (ushort)v9 , (ushort)v8,
-                              (ushort)v7 , (ushort)v6 , (ushort)v5 , (ushort)v4 , (ushort)v3 , (ushort)v2 , (ushort)v1 , (ushort)v0);
-    }
-    v_int16x32() {}
-
-    static inline v_int16x32 zero() { return v_int16x32(_mm512_setzero_si512()); }
-
-    short get0() const { return (short)_v_cvtsi512_si32(val); }
-};
-
-struct v_uint32x16
-{
-    typedef unsigned lane_type;
-    enum { nlanes = 16 };
-    __m512i val;
-
-    explicit v_uint32x16(__m512i v) : val(v) {}
-    v_uint32x16(unsigned v0,  unsigned v1,  unsigned v2,  unsigned v3,
-                unsigned v4,  unsigned v5,  unsigned v6,  unsigned v7,
-                unsigned v8,  unsigned v9,  unsigned v10, unsigned v11,
-                unsigned v12, unsigned v13, unsigned v14, unsigned v15)
-    {
-        val = _mm512_setr_epi32((int)v0,  (int)v1,  (int)v2,  (int)v3, (int)v4,  (int)v5,  (int)v6,  (int)v7,
-                                (int)v8,  (int)v9,  (int)v10, (int)v11, (int)v12, (int)v13, (int)v14, (int)v15);
-    }
-    v_uint32x16() {}
-
-    static inline v_uint32x16 zero() { return v_uint32x16(_mm512_setzero_si512()); }
-
-    unsigned get0() const { return (unsigned)_v_cvtsi512_si32(val); }
-};
-
-struct v_int32x16
-{
-    typedef int lane_type;
-    enum { nlanes = 16 };
-    __m512i val;
-
-    explicit v_int32x16(__m512i v) : val(v) {}
-    v_int32x16(int v0, int v1, int v2,  int v3,  int v4,  int v5,  int v6,  int v7,
-               int v8, int v9, int v10, int v11, int v12, int v13, int v14, int v15)
-    {
-        val = _mm512_setr_epi32(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15);
-    }
-    v_int32x16() {}
-
-    static inline v_int32x16 zero() { return v_int32x16(_mm512_setzero_si512()); }
-
-    int get0() const { return _v_cvtsi512_si32(val); }
-};
-
-struct v_float32x16
-{
-    typedef float lane_type;
-    enum { nlanes = 16 };
-    __m512 val;
-
-    explicit v_float32x16(__m512 v) : val(v) {}
-    v_float32x16(float v0, float v1, float v2,  float v3,  float v4,  float v5,  float v6,  float v7,
-                 float v8, float v9, float v10, float v11, float v12, float v13, float v14, float v15)
-    {
-        val = _mm512_setr_ps(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15);
-    }
-    v_float32x16() {}
-
-    static inline v_float32x16 zero() { return v_float32x16(_mm512_setzero_ps()); }
-
-    float get0() const { return _mm_cvtss_f32(_mm512_castps512_ps128(val)); }
-};
-
-struct v_uint64x8
-{
-    typedef uint64 lane_type;
-    enum { nlanes = 8 };
-    __m512i val;
-
-    explicit v_uint64x8(__m512i v) : val(v) {}
-    v_uint64x8(uint64 v0, uint64 v1, uint64 v2, uint64 v3, uint64 v4, uint64 v5, uint64 v6, uint64 v7)
-    { val = _mm512_setr_epi64((int64)v0, (int64)v1, (int64)v2, (int64)v3, (int64)v4, (int64)v5, (int64)v6, (int64)v7); }
-    v_uint64x8() {}
-
-    static inline v_uint64x8 zero() { return v_uint64x8(_mm512_setzero_si512()); }
-
-    uint64 get0() const
-    {
-    #if defined __x86_64__ || defined _M_X64
-        return (uint64)_mm_cvtsi128_si64(_mm512_castsi512_si128(val));
-    #else
-        int a = _mm_cvtsi128_si32(_mm512_castsi512_si128(val));
-        int b = _mm_cvtsi128_si32(_mm512_castsi512_si128(_mm512_srli_epi64(val, 32)));
-        return (unsigned)a | ((uint64)(unsigned)b << 32);
-    #endif
-    }
-};
-
-struct v_int64x8
-{
-    typedef int64 lane_type;
-    enum { nlanes = 8 };
-    __m512i val;
-
-    explicit v_int64x8(__m512i v) : val(v) {}
-    v_int64x8(int64 v0, int64 v1, int64 v2, int64 v3, int64 v4, int64 v5, int64 v6, int64 v7)
-    { val = _mm512_setr_epi64(v0, v1, v2, v3, v4, v5, v6, v7); }
-    v_int64x8() {}
-
-    static inline v_int64x8 zero() { return v_int64x8(_mm512_setzero_si512()); }
-
-    int64 get0() const
-    {
-    #if defined __x86_64__ || defined _M_X64
-        return (int64)_mm_cvtsi128_si64(_mm512_castsi512_si128(val));
-    #else
-        int a = _mm_cvtsi128_si32(_mm512_castsi512_si128(val));
-        int b = _mm_cvtsi128_si32(_mm512_castsi512_si128(_mm512_srli_epi64(val, 32)));
-        return (int64)((unsigned)a | ((uint64)(unsigned)b << 32));
-    #endif
-    }
-};
-
-struct v_float64x8
-{
-    typedef double lane_type;
-    enum { nlanes = 8 };
-    __m512d val;
-
-    explicit v_float64x8(__m512d v) : val(v) {}
-    v_float64x8(double v0, double v1, double v2, double v3, double v4, double v5, double v6, double v7)
-    { val = _mm512_setr_pd(v0, v1, v2, v3, v4, v5, v6, v7); }
-    v_float64x8() {}
-
-    static inline v_float64x8 zero() { return v_float64x8(_mm512_setzero_pd()); }
-
-    double get0() const { return _mm_cvtsd_f64(_mm512_castpd512_pd128(val)); }
-};
-
-//////////////// Load and store operations ///////////////
-
-#define OPENCV_HAL_IMPL_AVX512_LOADSTORE(_Tpvec, _Tp)                    \
-    inline _Tpvec v512_load(const _Tp* ptr)                           \
-    { return _Tpvec(_mm512_loadu_si512((const __m512i*)ptr)); }       \
-    inline _Tpvec v512_load_aligned(const _Tp* ptr)                   \
-    { return _Tpvec(_mm512_load_si512((const __m512i*)ptr)); }        \
-    inline _Tpvec v512_load_low(const _Tp* ptr)                       \
-    {                                                                 \
-        __m256i v256 = _mm256_loadu_si256((const __m256i*)ptr);       \
-        return _Tpvec(_mm512_castsi256_si512(v256));                  \
-    }                                                                 \
-    inline _Tpvec v512_load_halves(const _Tp* ptr0, const _Tp* ptr1)  \
-    {                                                                 \
-        __m256i vlo = _mm256_loadu_si256((const __m256i*)ptr0);       \
-        __m256i vhi = _mm256_loadu_si256((const __m256i*)ptr1);       \
-        return _Tpvec(_v512_combine(vlo, vhi));                       \
-    }                                                                 \
-    inline void v_store(_Tp* ptr, const _Tpvec& a)                    \
-    { _mm512_storeu_si512((__m512i*)ptr, a.val); }                    \
-    inline void v_store_aligned(_Tp* ptr, const _Tpvec& a)            \
-    { _mm512_store_si512((__m512i*)ptr, a.val); }                     \
-    inline void v_store_aligned_nocache(_Tp* ptr, const _Tpvec& a)    \
-    { _mm512_stream_si512((__m512i*)ptr, a.val); }                    \
-    inline void v_store(_Tp* ptr, const _Tpvec& a, hal::StoreMode mode) \
-    { \
-        if( mode == hal::STORE_UNALIGNED ) \
-            _mm512_storeu_si512((__m512i*)ptr, a.val); \
-        else if( mode == hal::STORE_ALIGNED_NOCACHE )  \
-            _mm512_stream_si512((__m512i*)ptr, a.val); \
-        else \
-            _mm512_store_si512((__m512i*)ptr, a.val); \
-    } \
-    inline void v_store_low(_Tp* ptr, const _Tpvec& a)                \
-    { _mm256_storeu_si256((__m256i*)ptr, _v512_extract_low(a.val)); }    \
-    inline void v_store_high(_Tp* ptr, const _Tpvec& a)               \
-    { _mm256_storeu_si256((__m256i*)ptr, _v512_extract_high(a.val)); }
-
-OPENCV_HAL_IMPL_AVX512_LOADSTORE(v_uint8x64,  uchar)
-OPENCV_HAL_IMPL_AVX512_LOADSTORE(v_int8x64,   schar)
-OPENCV_HAL_IMPL_AVX512_LOADSTORE(v_uint16x32, ushort)
-OPENCV_HAL_IMPL_AVX512_LOADSTORE(v_int16x32,  short)
-OPENCV_HAL_IMPL_AVX512_LOADSTORE(v_uint32x16,  unsigned)
-OPENCV_HAL_IMPL_AVX512_LOADSTORE(v_int32x16,   int)
-OPENCV_HAL_IMPL_AVX512_LOADSTORE(v_uint64x8,  uint64)
-OPENCV_HAL_IMPL_AVX512_LOADSTORE(v_int64x8,   int64)
-
-#define OPENCV_HAL_IMPL_AVX512_LOADSTORE_FLT(_Tpvec, _Tp, suffix, halfreg)   \
-    inline _Tpvec v512_load(const _Tp* ptr)                               \
-    { return _Tpvec(_mm512_loadu_##suffix(ptr)); }                        \
-    inline _Tpvec v512_load_aligned(const _Tp* ptr)                       \
-    { return _Tpvec(_mm512_load_##suffix(ptr)); }                         \
-    inline _Tpvec v512_load_low(const _Tp* ptr)                           \
-    {                                                                     \
-        return _Tpvec(_mm512_cast##suffix##256_##suffix##512              \
-                     (_mm256_loadu_##suffix(ptr)));                       \
-    }                                                                     \
-    inline _Tpvec v512_load_halves(const _Tp* ptr0, const _Tp* ptr1)      \
-    {                                                                     \
-        halfreg vlo = _mm256_loadu_##suffix(ptr0);                        \
-        halfreg vhi = _mm256_loadu_##suffix(ptr1);                        \
-        return _Tpvec(_v512_combine(vlo, vhi));                           \
-    }                                                                     \
-    inline void v_store(_Tp* ptr, const _Tpvec& a)                        \
-    { _mm512_storeu_##suffix(ptr, a.val); }                               \
-    inline void v_store_aligned(_Tp* ptr, const _Tpvec& a)                \
-    { _mm512_store_##suffix(ptr, a.val); }                                \
-    inline void v_store_aligned_nocache(_Tp* ptr, const _Tpvec& a)        \
-    { _mm512_stream_##suffix(ptr, a.val); }                               \
-    inline void v_store(_Tp* ptr, const _Tpvec& a, hal::StoreMode mode) \
-    { \
-        if( mode == hal::STORE_UNALIGNED ) \
-            _mm512_storeu_##suffix(ptr, a.val); \
-        else if( mode == hal::STORE_ALIGNED_NOCACHE )  \
-            _mm512_stream_##suffix(ptr, a.val); \
-        else \
-            _mm512_store_##suffix(ptr, a.val); \
-    } \
-    inline void v_store_low(_Tp* ptr, const _Tpvec& a)                    \
-    { _mm256_storeu_##suffix(ptr, _v512_extract_low(a.val)); }            \
-    inline void v_store_high(_Tp* ptr, const _Tpvec& a)                   \
-    { _mm256_storeu_##suffix(ptr, _v512_extract_high(a.val)); }
-
-OPENCV_HAL_IMPL_AVX512_LOADSTORE_FLT(v_float32x16, float,  ps, __m256)
-OPENCV_HAL_IMPL_AVX512_LOADSTORE_FLT(v_float64x8, double, pd, __m256d)
-
-#define OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, _Tpvecf, suffix, cast) \
-    inline _Tpvec v_reinterpret_as_##suffix(const _Tpvecf& a)   \
-    { return _Tpvec(cast(a.val)); }
-
-#define OPENCV_HAL_IMPL_AVX512_INIT(_Tpvec, _Tp, suffix, ssuffix, ctype_s)         \
-    inline _Tpvec v512_setzero_##suffix()                                          \
-    { return _Tpvec(_mm512_setzero_si512()); }                                     \
-    inline _Tpvec v512_setall_##suffix(_Tp v)                                      \
-    { return _Tpvec(_mm512_set1_##ssuffix((ctype_s)v)); }                          \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_uint8x64,   suffix, OPENCV_HAL_NOP)      \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_int8x64,    suffix, OPENCV_HAL_NOP)      \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_uint16x32,  suffix, OPENCV_HAL_NOP)      \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_int16x32,   suffix, OPENCV_HAL_NOP)      \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_uint32x16,  suffix, OPENCV_HAL_NOP)      \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_int32x16,   suffix, OPENCV_HAL_NOP)      \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_uint64x8,   suffix, OPENCV_HAL_NOP)      \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_int64x8,    suffix, OPENCV_HAL_NOP)      \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_float32x16, suffix, _mm512_castps_si512) \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_float64x8,  suffix, _mm512_castpd_si512)
-
-OPENCV_HAL_IMPL_AVX512_INIT(v_uint8x64,  uchar,    u8,  epi8,   char)
-OPENCV_HAL_IMPL_AVX512_INIT(v_int8x64,   schar,    s8,  epi8,   char)
-OPENCV_HAL_IMPL_AVX512_INIT(v_uint16x32, ushort,   u16, epi16,  short)
-OPENCV_HAL_IMPL_AVX512_INIT(v_int16x32,  short,    s16, epi16,  short)
-OPENCV_HAL_IMPL_AVX512_INIT(v_uint32x16, unsigned, u32, epi32,  int)
-OPENCV_HAL_IMPL_AVX512_INIT(v_int32x16,  int,      s32, epi32,  int)
-OPENCV_HAL_IMPL_AVX512_INIT(v_uint64x8,  uint64,   u64, epi64,  int64)
-OPENCV_HAL_IMPL_AVX512_INIT(v_int64x8,   int64,    s64, epi64,  int64)
-
-#define OPENCV_HAL_IMPL_AVX512_INIT_FLT(_Tpvec, _Tp, suffix, zsuffix, cast) \
-    inline _Tpvec v512_setzero_##suffix()                                   \
-    { return _Tpvec(_mm512_setzero_##zsuffix()); }                          \
-    inline _Tpvec v512_setall_##suffix(_Tp v)                               \
-    { return _Tpvec(_mm512_set1_##zsuffix(v)); }                            \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_uint8x64,  suffix, cast)          \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_int8x64,   suffix, cast)          \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_uint16x32, suffix, cast)          \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_int16x32,  suffix, cast)          \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_uint32x16, suffix, cast)          \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_int32x16,  suffix, cast)          \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_uint64x8,  suffix, cast)          \
-    OPENCV_HAL_IMPL_AVX512_CAST(_Tpvec, v_int64x8,   suffix, cast)
-
-OPENCV_HAL_IMPL_AVX512_INIT_FLT(v_float32x16, float,  f32, ps, _mm512_castsi512_ps)
-OPENCV_HAL_IMPL_AVX512_INIT_FLT(v_float64x8,  double, f64, pd, _mm512_castsi512_pd)
-
-inline v_float32x16 v_reinterpret_as_f32(const v_float32x16& a)
-{ return a; }
-inline v_float32x16 v_reinterpret_as_f32(const v_float64x8& a)
-{ return v_float32x16(_mm512_castpd_ps(a.val)); }
-
-inline v_float64x8 v_reinterpret_as_f64(const v_float64x8& a)
-{ return a; }
-inline v_float64x8 v_reinterpret_as_f64(const v_float32x16& a)
-{ return v_float64x8(_mm512_castps_pd(a.val)); }
-
-// FP16
-inline v_float32x16 v512_load_expand(const float16_t* ptr)
-{
-    return v_float32x16(_mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*)ptr)));
-}
-
-inline void v_pack_store(float16_t* ptr, const v_float32x16& a)
-{
-    __m256i ah = _mm512_cvtps_ph(a.val, 0);
-    _mm256_storeu_si256((__m256i*)ptr, ah);
-}
-
-/* Recombine & ZIP */
-inline void v_zip(const v_int8x64& a, const v_int8x64& b, v_int8x64& ab0, v_int8x64& ab1)
-{
-#if CV_AVX_512VBMI
-    __m512i mask0 = _v512_set_epu8( 95,  31,  94,  30,  93,  29,  92,  28,  91,  27,  90,  26,  89,  25,  88,  24,
-                                    87,  23,  86,  22,  85,  21,  84,  20,  83,  19,  82,  18,  81,  17,  80,  16,
-                                    79,  15,  78,  14,  77,  13,  76,  12,  75,  11,  74,  10,  73,   9,  72,   8,
-                                    71,   7,  70,   6,  69,   5,  68,   4,  67,   3,  66,   2,  65,   1,  64,   0);
-    ab0 = v_int8x64(_mm512_permutex2var_epi8(a.val, mask0, b.val));
-    __m512i mask1 = _v512_set_epu8(127,  63, 126,  62, 125,  61, 124,  60, 123,  59, 122,  58, 121,  57, 120,  56,
-                                   119,  55, 118,  54, 117,  53, 116,  52, 115,  51, 114,  50, 113,  49, 112,  48,
-                                   111,  47, 110,  46, 109,  45, 108,  44, 107,  43, 106,  42, 105,  41, 104,  40,
-                                   103,  39, 102,  38, 101,  37, 100,  36,  99,  35,  98,  34,  97,  33,  96,  32);
-    ab1 = v_int8x64(_mm512_permutex2var_epi8(a.val, mask1, b.val));
-#else
-    __m512i low  = _mm512_unpacklo_epi8(a.val, b.val);
-    __m512i high = _mm512_unpackhi_epi8(a.val, b.val);
-    ab0 = v_int8x64(_mm512_permutex2var_epi64(low, _v512_set_epu64(11, 10, 3, 2,  9,  8, 1, 0), high));
-    ab1 = v_int8x64(_mm512_permutex2var_epi64(low, _v512_set_epu64(15, 14, 7, 6, 13, 12, 5, 4), high));
-#endif
-}
-inline void v_zip(const v_int16x32& a, const v_int16x32& b, v_int16x32& ab0, v_int16x32& ab1)
-{
-    __m512i mask0 = _v512_set_epu16(47, 15, 46, 14, 45, 13, 44, 12, 43, 11, 42, 10, 41,  9, 40,  8,
-                                    39,  7, 38,  6, 37,  5, 36,  4, 35,  3, 34,  2, 33,  1, 32,  0);
-    ab0 = v_int16x32(_mm512_permutex2var_epi16(a.val, mask0, b.val));
-    __m512i mask1 = _v512_set_epu16(63, 31, 62, 30, 61, 29, 60, 28, 59, 27, 58, 26, 57, 25, 56, 24,
-                                    55, 23, 54, 22, 53, 21, 52, 20, 51, 19, 50, 18, 49, 17, 48, 16);
-    ab1 = v_int16x32(_mm512_permutex2var_epi16(a.val, mask1, b.val));
-}
-inline void v_zip(const v_int32x16& a, const v_int32x16& b, v_int32x16& ab0, v_int32x16& ab1)
-{
-    __m512i mask0 = _v512_set_epu32(23,  7, 22,  6, 21,  5, 20,  4, 19,  3, 18,  2, 17, 1, 16, 0);
-    ab0 = v_int32x16(_mm512_permutex2var_epi32(a.val, mask0, b.val));
-    __m512i mask1 = _v512_set_epu32(31, 15, 30, 14, 29, 13, 28, 12, 27, 11, 26, 10, 25, 9, 24, 8);
-    ab1 = v_int32x16(_mm512_permutex2var_epi32(a.val, mask1, b.val));
-}
-inline void v_zip(const v_int64x8& a, const v_int64x8& b, v_int64x8& ab0, v_int64x8& ab1)
-{
-    __m512i mask0 = _v512_set_epu64(11, 3, 10, 2,  9, 1,  8, 0);
-    ab0 = v_int64x8(_mm512_permutex2var_epi64(a.val, mask0, b.val));
-    __m512i mask1 = _v512_set_epu64(15, 7, 14, 6, 13, 5, 12, 4);
-    ab1 = v_int64x8(_mm512_permutex2var_epi64(a.val, mask1, b.val));
-}
-
-inline void v_zip(const v_uint8x64&  a, const v_uint8x64&  b, v_uint8x64& ab0, v_uint8x64& ab1)
-{
-    v_int8x64 i0, i1;
-    v_zip(v_reinterpret_as_s8(a), v_reinterpret_as_s8(b), i0, i1);
-    ab0 = v_reinterpret_as_u8(i0);
-    ab1 = v_reinterpret_as_u8(i1);
-}
-inline void v_zip(const v_uint16x32&  a, const v_uint16x32&  b, v_uint16x32& ab0, v_uint16x32& ab1)
-{
-    v_int16x32 i0, i1;
-    v_zip(v_reinterpret_as_s16(a), v_reinterpret_as_s16(b), i0, i1);
-    ab0 = v_reinterpret_as_u16(i0);
-    ab1 = v_reinterpret_as_u16(i1);
-}
-inline void v_zip(const v_uint32x16&  a, const v_uint32x16&  b, v_uint32x16& ab0, v_uint32x16& ab1)
-{
-    v_int32x16 i0, i1;
-    v_zip(v_reinterpret_as_s32(a), v_reinterpret_as_s32(b), i0, i1);
-    ab0 = v_reinterpret_as_u32(i0);
-    ab1 = v_reinterpret_as_u32(i1);
-}
-inline void v_zip(const v_uint64x8&  a, const v_uint64x8&  b, v_uint64x8& ab0, v_uint64x8& ab1)
-{
-    v_int64x8 i0, i1;
-    v_zip(v_reinterpret_as_s64(a), v_reinterpret_as_s64(b), i0, i1);
-    ab0 = v_reinterpret_as_u64(i0);
-    ab1 = v_reinterpret_as_u64(i1);
-}
-inline void v_zip(const v_float32x16&  a, const v_float32x16&  b, v_float32x16& ab0, v_float32x16& ab1)
-{
-    v_int32x16 i0, i1;
-    v_zip(v_reinterpret_as_s32(a), v_reinterpret_as_s32(b), i0, i1);
-    ab0 = v_reinterpret_as_f32(i0);
-    ab1 = v_reinterpret_as_f32(i1);
-}
-inline void v_zip(const v_float64x8&  a, const v_float64x8&  b, v_float64x8& ab0, v_float64x8& ab1)
-{
-    v_int64x8 i0, i1;
-    v_zip(v_reinterpret_as_s64(a), v_reinterpret_as_s64(b), i0, i1);
-    ab0 = v_reinterpret_as_f64(i0);
-    ab1 = v_reinterpret_as_f64(i1);
-}
-
-#define OPENCV_HAL_IMPL_AVX512_COMBINE(_Tpvec, suffix)                                    \
-    inline _Tpvec v_combine_low(const _Tpvec& a, const _Tpvec& b)                         \
-    { return _Tpvec(_v512_combine(_v512_extract_low(a.val), _v512_extract_low(b.val))); } \
-    inline _Tpvec v_combine_high(const _Tpvec& a, const _Tpvec& b)                        \
-    { return _Tpvec(_v512_insert(b.val, _v512_extract_high(a.val))); }                    \
-    inline void v_recombine(const _Tpvec& a, const _Tpvec& b,                             \
-                                  _Tpvec& c, _Tpvec& d)                                   \
-    {                                                                                     \
-        c.val = _v512_combine(_v512_extract_low(a.val),_v512_extract_low(b.val));         \
-        d.val = _v512_insert(b.val,_v512_extract_high(a.val));                            \
-    }
-
-
-OPENCV_HAL_IMPL_AVX512_COMBINE(v_uint8x64,   epi8)
-OPENCV_HAL_IMPL_AVX512_COMBINE(v_int8x64,    epi8)
-OPENCV_HAL_IMPL_AVX512_COMBINE(v_uint16x32,  epi16)
-OPENCV_HAL_IMPL_AVX512_COMBINE(v_int16x32,   epi16)
-OPENCV_HAL_IMPL_AVX512_COMBINE(v_uint32x16,  epi32)
-OPENCV_HAL_IMPL_AVX512_COMBINE(v_int32x16,   epi32)
-OPENCV_HAL_IMPL_AVX512_COMBINE(v_uint64x8,   epi64)
-OPENCV_HAL_IMPL_AVX512_COMBINE(v_int64x8,    epi64)
-OPENCV_HAL_IMPL_AVX512_COMBINE(v_float32x16, ps)
-OPENCV_HAL_IMPL_AVX512_COMBINE(v_float64x8,  pd)
-
-////////// Arithmetic, bitwise and comparison operations /////////
-
-/* Element-wise binary and unary operations */
-
-/** Non-saturating arithmetics **/
-#define OPENCV_HAL_IMPL_AVX512_BIN_FUNC(func, _Tpvec, intrin) \
-    inline _Tpvec func(const _Tpvec& a, const _Tpvec& b)      \
-    { return _Tpvec(intrin(a.val, b.val)); }
-
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_add_wrap, v_uint8x64, _mm512_add_epi8)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_add_wrap, v_int8x64, _mm512_add_epi8)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_add_wrap, v_uint16x32, _mm512_add_epi16)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_add_wrap, v_int16x32, _mm512_add_epi16)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_sub_wrap, v_uint8x64, _mm512_sub_epi8)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_sub_wrap, v_int8x64, _mm512_sub_epi8)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_sub_wrap, v_uint16x32, _mm512_sub_epi16)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_sub_wrap, v_int16x32, _mm512_sub_epi16)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_mul_wrap, v_uint16x32, _mm512_mullo_epi16)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_mul_wrap, v_int16x32, _mm512_mullo_epi16)
-
-inline v_uint8x64 v_mul_wrap(const v_uint8x64& a, const v_uint8x64& b)
-{
-    __m512i ad = _mm512_srai_epi16(a.val, 8);
-    __m512i bd = _mm512_srai_epi16(b.val, 8);
-    __m512i p0 = _mm512_mullo_epi16(a.val, b.val); // even
-    __m512i p1 = _mm512_slli_epi16(_mm512_mullo_epi16(ad, bd), 8); // odd
-    return v_uint8x64(_mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, p0, p1));
-}
-inline v_int8x64 v_mul_wrap(const v_int8x64& a, const v_int8x64& b)
-{
-    return v_reinterpret_as_s8(v_mul_wrap(v_reinterpret_as_u8(a), v_reinterpret_as_u8(b)));
-}
-
-#define OPENCV_HAL_IMPL_AVX512_BIN_OP(bin_op, _Tpvec, intrin)            \
-    inline _Tpvec operator bin_op (const _Tpvec& a, const _Tpvec& b)     \
-    { return _Tpvec(intrin(a.val, b.val)); }                             \
-    inline _Tpvec& operator bin_op##= (_Tpvec& a, const _Tpvec& b)       \
-    { a.val = intrin(a.val, b.val); return a; }
-
-OPENCV_HAL_IMPL_AVX512_BIN_OP(+, v_uint32x16, _mm512_add_epi32)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(-, v_uint32x16, _mm512_sub_epi32)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(+, v_int32x16, _mm512_add_epi32)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(-, v_int32x16, _mm512_sub_epi32)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(+, v_uint64x8, _mm512_add_epi64)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(-, v_uint64x8, _mm512_sub_epi64)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(+, v_int64x8, _mm512_add_epi64)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(-, v_int64x8, _mm512_sub_epi64)
-
-OPENCV_HAL_IMPL_AVX512_BIN_OP(*, v_uint32x16, _mm512_mullo_epi32)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(*, v_int32x16, _mm512_mullo_epi32)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(*, v_uint64x8, _mm512_mullo_epi64)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(*, v_int64x8, _mm512_mullo_epi64)
-
-/** Saturating arithmetics **/
-OPENCV_HAL_IMPL_AVX512_BIN_OP(+, v_uint8x64,  _mm512_adds_epu8)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(-, v_uint8x64,  _mm512_subs_epu8)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(+, v_int8x64,   _mm512_adds_epi8)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(-, v_int8x64,   _mm512_subs_epi8)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(+, v_uint16x32, _mm512_adds_epu16)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(-, v_uint16x32, _mm512_subs_epu16)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(+, v_int16x32,  _mm512_adds_epi16)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(-, v_int16x32,  _mm512_subs_epi16)
-
-OPENCV_HAL_IMPL_AVX512_BIN_OP(+, v_float32x16, _mm512_add_ps)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(-, v_float32x16, _mm512_sub_ps)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(*, v_float32x16, _mm512_mul_ps)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(/, v_float32x16, _mm512_div_ps)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(+, v_float64x8, _mm512_add_pd)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(-, v_float64x8, _mm512_sub_pd)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(*, v_float64x8, _mm512_mul_pd)
-OPENCV_HAL_IMPL_AVX512_BIN_OP(/, v_float64x8, _mm512_div_pd)
-
-// saturating multiply
-inline v_uint8x64 operator * (const v_uint8x64& a, const v_uint8x64& b)
-{
-    v_uint16x32 c, d;
-    v_mul_expand(a, b, c, d);
-    return v_pack(c, d);
-}
-inline v_int8x64 operator * (const v_int8x64& a, const v_int8x64& b)
-{
-    v_int16x32 c, d;
-    v_mul_expand(a, b, c, d);
-    return v_pack(c, d);
-}
-inline v_uint16x32 operator * (const v_uint16x32& a, const v_uint16x32& b)
-{
-    __m512i pl = _mm512_mullo_epi16(a.val, b.val);
-    __m512i ph = _mm512_mulhi_epu16(a.val, b.val);
-    __m512i p0 = _mm512_unpacklo_epi16(pl, ph);
-    __m512i p1 = _mm512_unpackhi_epi16(pl, ph);
-
-    const __m512i m = _mm512_set1_epi32(65535);
-    return v_uint16x32(_mm512_packus_epi32(_mm512_min_epu32(p0, m), _mm512_min_epu32(p1, m)));
-}
-inline v_int16x32 operator * (const v_int16x32& a, const v_int16x32& b)
-{
-    __m512i pl = _mm512_mullo_epi16(a.val, b.val);
-    __m512i ph = _mm512_mulhi_epi16(a.val, b.val);
-    __m512i p0 = _mm512_unpacklo_epi16(pl, ph);
-    __m512i p1 = _mm512_unpackhi_epi16(pl, ph);
-    return v_int16x32(_mm512_packs_epi32(p0, p1));
-}
-
-inline v_uint8x64& operator *= (v_uint8x64& a, const v_uint8x64& b)
-{ a = a * b; return a; }
-inline v_int8x64& operator *= (v_int8x64& a, const v_int8x64& b)
-{ a = a * b; return a; }
-inline v_uint16x32& operator *= (v_uint16x32& a, const v_uint16x32& b)
-{ a = a * b; return a; }
-inline v_int16x32& operator *= (v_int16x32& a, const v_int16x32& b)
-{ a = a * b; return a; }
-
-inline v_int16x32 v_mul_hi(const v_int16x32& a, const v_int16x32& b) { return v_int16x32(_mm512_mulhi_epi16(a.val, b.val)); }
-inline v_uint16x32 v_mul_hi(const v_uint16x32& a, const v_uint16x32& b) { return v_uint16x32(_mm512_mulhi_epu16(a.val, b.val)); }
-
-//  Multiply and expand
-inline void v_mul_expand(const v_uint8x64& a, const v_uint8x64& b,
-                         v_uint16x32& c, v_uint16x32& d)
-{
-    v_uint16x32 a0, a1, b0, b1;
-    v_expand(a, a0, a1);
-    v_expand(b, b0, b1);
-    c = v_mul_wrap(a0, b0);
-    d = v_mul_wrap(a1, b1);
-}
-
-inline void v_mul_expand(const v_int8x64& a, const v_int8x64& b,
-                         v_int16x32& c, v_int16x32& d)
-{
-    v_int16x32 a0, a1, b0, b1;
-    v_expand(a, a0, a1);
-    v_expand(b, b0, b1);
-    c = v_mul_wrap(a0, b0);
-    d = v_mul_wrap(a1, b1);
-}
-
-inline void v_mul_expand(const v_int16x32& a, const v_int16x32& b,
-                         v_int32x16& c, v_int32x16& d)
-{
-    v_int16x32 v0, v1;
-    v_zip(v_mul_wrap(a, b), v_mul_hi(a, b), v0, v1);
-
-    c = v_reinterpret_as_s32(v0);
-    d = v_reinterpret_as_s32(v1);
-}
-
-inline void v_mul_expand(const v_uint16x32& a, const v_uint16x32& b,
-                         v_uint32x16& c, v_uint32x16& d)
-{
-    v_uint16x32 v0, v1;
-    v_zip(v_mul_wrap(a, b), v_mul_hi(a, b), v0, v1);
-
-    c = v_reinterpret_as_u32(v0);
-    d = v_reinterpret_as_u32(v1);
-}
-
-inline void v_mul_expand(const v_uint32x16& a, const v_uint32x16& b,
-                         v_uint64x8& c, v_uint64x8& d)
-{
-    v_zip(v_uint64x8(_mm512_mul_epu32(a.val, b.val)),
-          v_uint64x8(_mm512_mul_epu32(_mm512_srli_epi64(a.val, 32), _mm512_srli_epi64(b.val, 32))), c, d);
-}
-
-inline void v_mul_expand(const v_int32x16& a, const v_int32x16& b,
-    v_int64x8& c, v_int64x8& d)
-{
-    v_zip(v_int64x8(_mm512_mul_epi32(a.val, b.val)),
-          v_int64x8(_mm512_mul_epi32(_mm512_srli_epi64(a.val, 32), _mm512_srli_epi64(b.val, 32))), c, d);
-}
-
-/** Bitwise shifts **/
-#define OPENCV_HAL_IMPL_AVX512_SHIFT_OP(_Tpuvec, _Tpsvec, suffix) \
-    inline _Tpuvec operator << (const _Tpuvec& a, int imm)        \
-    { return _Tpuvec(_mm512_slli_##suffix(a.val, imm)); }         \
-    inline _Tpsvec operator << (const _Tpsvec& a, int imm)        \
-    { return _Tpsvec(_mm512_slli_##suffix(a.val, imm)); }         \
-    inline _Tpuvec operator >> (const _Tpuvec& a, int imm)        \
-    { return _Tpuvec(_mm512_srli_##suffix(a.val, imm)); }         \
-    inline _Tpsvec operator >> (const _Tpsvec& a, int imm)        \
-    { return _Tpsvec(_mm512_srai_##suffix(a.val, imm)); }         \
-    template<int imm>                                             \
-    inline _Tpuvec v_shl(const _Tpuvec& a)                        \
-    { return _Tpuvec(_mm512_slli_##suffix(a.val, imm)); }         \
-    template<int imm>                                             \
-    inline _Tpsvec v_shl(const _Tpsvec& a)                        \
-    { return _Tpsvec(_mm512_slli_##suffix(a.val, imm)); }         \
-    template<int imm>                                             \
-    inline _Tpuvec v_shr(const _Tpuvec& a)                        \
-    { return _Tpuvec(_mm512_srli_##suffix(a.val, imm)); }         \
-    template<int imm>                                             \
-    inline _Tpsvec v_shr(const _Tpsvec& a)                        \
-    { return _Tpsvec(_mm512_srai_##suffix(a.val, imm)); }
-
-OPENCV_HAL_IMPL_AVX512_SHIFT_OP(v_uint16x32, v_int16x32, epi16)
-OPENCV_HAL_IMPL_AVX512_SHIFT_OP(v_uint32x16, v_int32x16, epi32)
-OPENCV_HAL_IMPL_AVX512_SHIFT_OP(v_uint64x8,  v_int64x8,  epi64)
-
-
-/** Bitwise logic **/
-#define OPENCV_HAL_IMPL_AVX512_LOGIC_OP(_Tpvec, suffix, not_const) \
-    OPENCV_HAL_IMPL_AVX512_BIN_OP(&, _Tpvec, _mm512_and_##suffix)  \
-    OPENCV_HAL_IMPL_AVX512_BIN_OP(|, _Tpvec, _mm512_or_##suffix)   \
-    OPENCV_HAL_IMPL_AVX512_BIN_OP(^, _Tpvec, _mm512_xor_##suffix)  \
-    inline _Tpvec operator ~ (const _Tpvec& a)                     \
-    { return _Tpvec(_mm512_xor_##suffix(a.val, not_const)); }
-
-OPENCV_HAL_IMPL_AVX512_LOGIC_OP(v_uint8x64,   si512, _mm512_set1_epi32(-1))
-OPENCV_HAL_IMPL_AVX512_LOGIC_OP(v_int8x64,    si512, _mm512_set1_epi32(-1))
-OPENCV_HAL_IMPL_AVX512_LOGIC_OP(v_uint16x32,  si512, _mm512_set1_epi32(-1))
-OPENCV_HAL_IMPL_AVX512_LOGIC_OP(v_int16x32,   si512, _mm512_set1_epi32(-1))
-OPENCV_HAL_IMPL_AVX512_LOGIC_OP(v_uint32x16,  si512, _mm512_set1_epi32(-1))
-OPENCV_HAL_IMPL_AVX512_LOGIC_OP(v_int32x16,   si512, _mm512_set1_epi32(-1))
-OPENCV_HAL_IMPL_AVX512_LOGIC_OP(v_uint64x8,   si512, _mm512_set1_epi64(-1))
-OPENCV_HAL_IMPL_AVX512_LOGIC_OP(v_int64x8,    si512, _mm512_set1_epi64(-1))
-OPENCV_HAL_IMPL_AVX512_LOGIC_OP(v_float32x16, ps,    _mm512_castsi512_ps(_mm512_set1_epi32(-1)))
-OPENCV_HAL_IMPL_AVX512_LOGIC_OP(v_float64x8,  pd,    _mm512_castsi512_pd(_mm512_set1_epi32(-1)))
-
-/** Select **/
-#define OPENCV_HAL_IMPL_AVX512_SELECT(_Tpvec, suffix, zsuf)                      \
-    inline _Tpvec v_select(const _Tpvec& mask, const _Tpvec& a, const _Tpvec& b) \
-    { return _Tpvec(_mm512_mask_blend_##suffix(_mm512_cmp_##suffix##_mask(mask.val, _mm512_setzero_##zsuf(), _MM_CMPINT_EQ), a.val, b.val)); }
-
-OPENCV_HAL_IMPL_AVX512_SELECT(v_uint8x64,   epi8, si512)
-OPENCV_HAL_IMPL_AVX512_SELECT(v_int8x64,    epi8, si512)
-OPENCV_HAL_IMPL_AVX512_SELECT(v_uint16x32, epi16, si512)
-OPENCV_HAL_IMPL_AVX512_SELECT(v_int16x32,  epi16, si512)
-OPENCV_HAL_IMPL_AVX512_SELECT(v_uint32x16, epi32, si512)
-OPENCV_HAL_IMPL_AVX512_SELECT(v_int32x16,  epi32, si512)
-OPENCV_HAL_IMPL_AVX512_SELECT(v_uint64x8,  epi64, si512)
-OPENCV_HAL_IMPL_AVX512_SELECT(v_int64x8,   epi64, si512)
-OPENCV_HAL_IMPL_AVX512_SELECT(v_float32x16,   ps,    ps)
-OPENCV_HAL_IMPL_AVX512_SELECT(v_float64x8,    pd,    pd)
-
-/** Comparison **/
-#define OPENCV_HAL_IMPL_AVX512_CMP_INT(bin_op, imm8, _Tpvec, sufcmp, sufset, tval) \
-    inline _Tpvec operator bin_op (const _Tpvec& a, const _Tpvec& b)               \
-    { return _Tpvec(_mm512_maskz_set1_##sufset(_mm512_cmp_##sufcmp##_mask(a.val, b.val, imm8), tval)); }
-
-#define OPENCV_HAL_IMPL_AVX512_CMP_OP_INT(_Tpvec, sufcmp, sufset, tval)              \
-    OPENCV_HAL_IMPL_AVX512_CMP_INT(==, _MM_CMPINT_EQ,  _Tpvec, sufcmp, sufset, tval) \
-    OPENCV_HAL_IMPL_AVX512_CMP_INT(!=, _MM_CMPINT_NE,  _Tpvec, sufcmp, sufset, tval) \
-    OPENCV_HAL_IMPL_AVX512_CMP_INT(<,  _MM_CMPINT_LT,  _Tpvec, sufcmp, sufset, tval) \
-    OPENCV_HAL_IMPL_AVX512_CMP_INT(>,  _MM_CMPINT_NLE, _Tpvec, sufcmp, sufset, tval) \
-    OPENCV_HAL_IMPL_AVX512_CMP_INT(<=, _MM_CMPINT_LE,  _Tpvec, sufcmp, sufset, tval) \
-    OPENCV_HAL_IMPL_AVX512_CMP_INT(>=, _MM_CMPINT_NLT, _Tpvec, sufcmp, sufset, tval)
-
-OPENCV_HAL_IMPL_AVX512_CMP_OP_INT(v_uint8x64,   epu8,  epi8, (char)-1)
-OPENCV_HAL_IMPL_AVX512_CMP_OP_INT(v_int8x64,    epi8,  epi8, (char)-1)
-OPENCV_HAL_IMPL_AVX512_CMP_OP_INT(v_uint16x32, epu16, epi16, (short)-1)
-OPENCV_HAL_IMPL_AVX512_CMP_OP_INT(v_int16x32,  epi16, epi16, (short)-1)
-OPENCV_HAL_IMPL_AVX512_CMP_OP_INT(v_uint32x16, epu32, epi32, (int)-1)
-OPENCV_HAL_IMPL_AVX512_CMP_OP_INT(v_int32x16,  epi32, epi32, (int)-1)
-OPENCV_HAL_IMPL_AVX512_CMP_OP_INT(v_uint64x8,  epu64, epi64, (int64)-1)
-OPENCV_HAL_IMPL_AVX512_CMP_OP_INT(v_int64x8,   epi64, epi64, (int64)-1)
-
-#define OPENCV_HAL_IMPL_AVX512_CMP_FLT(bin_op, imm8, _Tpvec, sufcmp, sufset, tval) \
-    inline _Tpvec operator bin_op (const _Tpvec& a, const _Tpvec& b)               \
-    { return _Tpvec(_mm512_castsi512_##sufcmp(_mm512_maskz_set1_##sufset(_mm512_cmp_##sufcmp##_mask(a.val, b.val, imm8), tval))); }
-
-#define OPENCV_HAL_IMPL_AVX512_CMP_OP_FLT(_Tpvec, sufcmp, sufset, tval)           \
-    OPENCV_HAL_IMPL_AVX512_CMP_FLT(==, _CMP_EQ_OQ,  _Tpvec, sufcmp, sufset, tval) \
-    OPENCV_HAL_IMPL_AVX512_CMP_FLT(!=, _CMP_NEQ_OQ, _Tpvec, sufcmp, sufset, tval) \
-    OPENCV_HAL_IMPL_AVX512_CMP_FLT(<,  _CMP_LT_OQ,  _Tpvec, sufcmp, sufset, tval) \
-    OPENCV_HAL_IMPL_AVX512_CMP_FLT(>,  _CMP_GT_OQ,  _Tpvec, sufcmp, sufset, tval) \
-    OPENCV_HAL_IMPL_AVX512_CMP_FLT(<=, _CMP_LE_OQ,  _Tpvec, sufcmp, sufset, tval) \
-    OPENCV_HAL_IMPL_AVX512_CMP_FLT(>=, _CMP_GE_OQ,  _Tpvec, sufcmp, sufset, tval)
-
-OPENCV_HAL_IMPL_AVX512_CMP_OP_FLT(v_float32x16, ps, epi32, (int)-1)
-OPENCV_HAL_IMPL_AVX512_CMP_OP_FLT(v_float64x8,  pd, epi64, (int64)-1)
-
-inline v_float32x16 v_not_nan(const v_float32x16& a)
-{ return v_float32x16(_mm512_castsi512_ps(_mm512_maskz_set1_epi32(_mm512_cmp_ps_mask(a.val, a.val, _CMP_ORD_Q), (int)-1))); }
-inline v_float64x8 v_not_nan(const v_float64x8& a)
-{ return v_float64x8(_mm512_castsi512_pd(_mm512_maskz_set1_epi64(_mm512_cmp_pd_mask(a.val, a.val, _CMP_ORD_Q), (int64)-1))); }
-
-/** min/max **/
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_min, v_uint8x64,   _mm512_min_epu8)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_max, v_uint8x64,   _mm512_max_epu8)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_min, v_int8x64,    _mm512_min_epi8)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_max, v_int8x64,    _mm512_max_epi8)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_min, v_uint16x32,  _mm512_min_epu16)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_max, v_uint16x32,  _mm512_max_epu16)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_min, v_int16x32,   _mm512_min_epi16)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_max, v_int16x32,   _mm512_max_epi16)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_min, v_uint32x16,  _mm512_min_epu32)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_max, v_uint32x16,  _mm512_max_epu32)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_min, v_int32x16,   _mm512_min_epi32)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_max, v_int32x16,   _mm512_max_epi32)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_min, v_uint64x8,   _mm512_min_epu64)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_max, v_uint64x8,   _mm512_max_epu64)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_min, v_int64x8,    _mm512_min_epi64)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_max, v_int64x8,    _mm512_max_epi64)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_min, v_float32x16, _mm512_min_ps)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_max, v_float32x16, _mm512_max_ps)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_min, v_float64x8,  _mm512_min_pd)
-OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_max, v_float64x8,  _mm512_max_pd)
-
-/** Rotate **/
-namespace {
-    template<bool prec, int imm4, bool part, int imm32>
-    struct _v_rotate_right { static inline v_int8x64 eval(const v_int8x64&, const v_int8x64&) { return v_int8x64(); }};
-    template<int imm4, int imm32>
-    struct _v_rotate_right<true, imm4, false, imm32> { static inline v_int8x64 eval(const v_int8x64& a, const v_int8x64& b)
-    {
-        return v_int8x64(_mm512_or_si512(_mm512_srli_epi32(_mm512_alignr_epi32(b.val, a.val, imm32    ),    imm4 *8),
-                                         _mm512_slli_epi32(_mm512_alignr_epi32(b.val, a.val, imm32 + 1), (4-imm4)*8)));
-    }};
-    template<int imm4>
-    struct _v_rotate_right<true, imm4, false, 15> { static inline v_int8x64 eval(const v_int8x64& a, const v_int8x64& b)
-    {
-        return v_int8x64(_mm512_or_si512(_mm512_srli_epi32(_mm512_alignr_epi32(b.val, a.val, 15),    imm4 *8),
-                                         _mm512_slli_epi32(                                b.val, (4-imm4)*8)));
-    }};
-    template<int imm4, int imm32>
-    struct _v_rotate_right<true, imm4, true, imm32> { static inline v_int8x64 eval(const v_int8x64&, const v_int8x64& b)
-    {
-        return v_int8x64(_mm512_or_si512(_mm512_srli_epi32(_mm512_alignr_epi32(_mm512_setzero_si512(), b.val, imm32 - 16),    imm4 *8),
-                                         _mm512_slli_epi32(_mm512_alignr_epi32(_mm512_setzero_si512(), b.val, imm32 - 15), (4-imm4)*8)));
-    }};
-    template<int imm4>
-    struct _v_rotate_right<true, imm4, true, 31> { static inline v_int8x64 eval(const v_int8x64&, const v_int8x64& b)
-    { return v_int8x64(_mm512_srli_epi32(_mm512_alignr_epi32(_mm512_setzero_si512(), b.val, 15), imm4*8)); }};
-    template<int imm32>
-    struct _v_rotate_right<false, 0, false, imm32> { static inline v_int8x64 eval(const v_int8x64& a, const v_int8x64& b)
-    { return v_int8x64(_mm512_alignr_epi32(b.val, a.val, imm32)); }};
-    template<>
-    struct _v_rotate_right<false, 0, false, 0> { static inline v_int8x64 eval(const v_int8x64& a, const v_int8x64&) { return a; }};
-    template<int imm32>
-    struct _v_rotate_right<false, 0, true, imm32> { static inline v_int8x64 eval(const v_int8x64&, const v_int8x64& b)
-    { return v_int8x64(_mm512_alignr_epi32(_mm512_setzero_si512(), b.val, imm32 - 16)); }};
-    template<>
-    struct _v_rotate_right<false, 0, true, 16> { static inline v_int8x64 eval(const v_int8x64&, const v_int8x64& b) { return b; }};
-    template<>
-    struct _v_rotate_right<false, 0, true, 32> { static inline v_int8x64 eval(const v_int8x64&, const v_int8x64&) { return v_int8x64(); }};
-}
-template<int imm> inline v_int8x64 v_rotate_right(const v_int8x64& a, const v_int8x64& b)
-{
-    return imm >= 128 ? v_int8x64() :
-#if CV_AVX_512VBMI
-    v_int8x64(_mm512_permutex2var_epi8(a.val,
-    _v512_set_epu8(0x3f + imm, 0x3e + imm, 0x3d + imm, 0x3c + imm, 0x3b + imm, 0x3a + imm, 0x39 + imm, 0x38 + imm,
-                   0x37 + imm, 0x36 + imm, 0x35 + imm, 0x34 + imm, 0x33 + imm, 0x32 + imm, 0x31 + imm, 0x30 + imm,
-                   0x2f + imm, 0x2e + imm, 0x2d + imm, 0x2c + imm, 0x2b + imm, 0x2a + imm, 0x29 + imm, 0x28 + imm,
-                   0x27 + imm, 0x26 + imm, 0x25 + imm, 0x24 + imm, 0x23 + imm, 0x22 + imm, 0x21 + imm, 0x20 + imm,
-                   0x1f + imm, 0x1e + imm, 0x1d + imm, 0x1c + imm, 0x1b + imm, 0x1a + imm, 0x19 + imm, 0x18 + imm,
-                   0x17 + imm, 0x16 + imm, 0x15 + imm, 0x14 + imm, 0x13 + imm, 0x12 + imm, 0x11 + imm, 0x10 + imm,
-                   0x0f + imm, 0x0e + imm, 0x0d + imm, 0x0c + imm, 0x0b + imm, 0x0a + imm, 0x09 + imm, 0x08 + imm,
-                   0x07 + imm, 0x06 + imm, 0x05 + imm, 0x04 + imm, 0x03 + imm, 0x02 + imm, 0x01 + imm, 0x00 + imm), b.val));
-#else
-    _v_rotate_right<imm%4!=0, imm%4, (imm/4 > 15), imm/4>::eval(a, b);
-#endif
-}
-template<int imm>
-inline v_int8x64 v_rotate_left(const v_int8x64& a, const v_int8x64& b)
-{
-    if (imm == 0) return a;
-    if (imm == 64) return b;
-    if (imm >= 128) return v_int8x64();
-#if CV_AVX_512VBMI
-    return v_int8x64(_mm512_permutex2var_epi8(b.val,
-           _v512_set_epi8(0x7f - imm,0x7e - imm,0x7d - imm,0x7c - imm,0x7b - imm,0x7a - imm,0x79 - imm,0x78 - imm,
-                          0x77 - imm,0x76 - imm,0x75 - imm,0x74 - imm,0x73 - imm,0x72 - imm,0x71 - imm,0x70 - imm,
-                          0x6f - imm,0x6e - imm,0x6d - imm,0x6c - imm,0x6b - imm,0x6a - imm,0x69 - imm,0x68 - imm,
-                          0x67 - imm,0x66 - imm,0x65 - imm,0x64 - imm,0x63 - imm,0x62 - imm,0x61 - imm,0x60 - imm,
-                          0x5f - imm,0x5e - imm,0x5d - imm,0x5c - imm,0x5b - imm,0x5a - imm,0x59 - imm,0x58 - imm,
-                          0x57 - imm,0x56 - imm,0x55 - imm,0x54 - imm,0x53 - imm,0x52 - imm,0x51 - imm,0x50 - imm,
-                          0x4f - imm,0x4e - imm,0x4d - imm,0x4c - imm,0x4b - imm,0x4a - imm,0x49 - imm,0x48 - imm,
-                          0x47 - imm,0x46 - imm,0x45 - imm,0x44 - imm,0x43 - imm,0x42 - imm,0x41 - imm,0x40 - imm), a.val));
-#else
-    return imm < 64 ? v_rotate_right<64 - imm>(b, a) : v_rotate_right<128 - imm>(v512_setzero_s8(), b);
-#endif
-}
-template<int imm>
-inline v_int8x64 v_rotate_right(const v_int8x64& a)
-{
-    if (imm == 0) return a;
-    if (imm >= 64) return v_int8x64();
-#if CV_AVX_512VBMI
-    return v_int8x64(_mm512_maskz_permutexvar_epi8(0xFFFFFFFFFFFFFFFF >> imm,
-           _v512_set_epu8(0x3f + imm,0x3e + imm,0x3d + imm,0x3c + imm,0x3b + imm,0x3a + imm,0x39 + imm,0x38 + imm,
-                          0x37 + imm,0x36 + imm,0x35 + imm,0x34 + imm,0x33 + imm,0x32 + imm,0x31 + imm,0x30 + imm,
-                          0x2f + imm,0x2e + imm,0x2d + imm,0x2c + imm,0x2b + imm,0x2a + imm,0x29 + imm,0x28 + imm,
-                          0x27 + imm,0x26 + imm,0x25 + imm,0x24 + imm,0x23 + imm,0x22 + imm,0x21 + imm,0x20 + imm,
-                          0x1f + imm,0x1e + imm,0x1d + imm,0x1c + imm,0x1b + imm,0x1a + imm,0x19 + imm,0x18 + imm,
-                          0x17 + imm,0x16 + imm,0x15 + imm,0x14 + imm,0x13 + imm,0x12 + imm,0x11 + imm,0x10 + imm,
-                          0x0f + imm,0x0e + imm,0x0d + imm,0x0c + imm,0x0b + imm,0x0a + imm,0x09 + imm,0x08 + imm,
-                          0x07 + imm,0x06 + imm,0x05 + imm,0x04 + imm,0x03 + imm,0x02 + imm,0x01 + imm,0x00 + imm), a.val));
-#else
-    return v_rotate_right<imm>(a, v512_setzero_s8());
-#endif
-}
-template<int imm>
-inline v_int8x64 v_rotate_left(const v_int8x64& a)
-{
-    if (imm == 0) return a;
-    if (imm >= 64) return v_int8x64();
-#if CV_AVX_512VBMI
-    return v_int8x64(_mm512_maskz_permutexvar_epi8(0xFFFFFFFFFFFFFFFF << imm,
-           _v512_set_epi8(0x3f - imm,0x3e - imm,0x3d - imm,0x3c - imm,0x3b - imm,0x3a - imm,0x39 - imm,0x38 - imm,
-                          0x37 - imm,0x36 - imm,0x35 - imm,0x34 - imm,0x33 - imm,0x32 - imm,0x31 - imm,0x30 - imm,
-                          0x2f - imm,0x2e - imm,0x2d - imm,0x2c - imm,0x2b - imm,0x2a - imm,0x29 - imm,0x28 - imm,
-                          0x27 - imm,0x26 - imm,0x25 - imm,0x24 - imm,0x23 - imm,0x22 - imm,0x21 - imm,0x20 - imm,
-                          0x1f - imm,0x1e - imm,0x1d - imm,0x1c - imm,0x1b - imm,0x1a - imm,0x19 - imm,0x18 - imm,
-                          0x17 - imm,0x16 - imm,0x15 - imm,0x14 - imm,0x13 - imm,0x12 - imm,0x11 - imm,0x10 - imm,
-                          0x0f - imm,0x0e - imm,0x0d - imm,0x0c - imm,0x0b - imm,0x0a - imm,0x09 - imm,0x08 - imm,
-                          0x07 - imm,0x06 - imm,0x05 - imm,0x04 - imm,0x03 - imm,0x02 - imm,0x01 - imm,0x00 - imm), a.val));
-#else
-    return v_rotate_right<64 - imm>(v512_setzero_s8(), a);
-#endif
-}
-
-#define OPENCV_HAL_IMPL_AVX512_ROTATE_PM(_Tpvec, suffix)                                                                                   \
-template<int imm> inline _Tpvec v_rotate_left(const _Tpvec& a, const _Tpvec& b)                                                            \
-{ return v_reinterpret_as_##suffix(v_rotate_left<imm * sizeof(_Tpvec::lane_type)>(v_reinterpret_as_s8(a), v_reinterpret_as_s8(b))); }      \
-template<int imm> inline _Tpvec v_rotate_right(const _Tpvec& a, const _Tpvec& b)                                                           \
-{ return v_reinterpret_as_##suffix(v_rotate_right<imm * sizeof(_Tpvec::lane_type)>(v_reinterpret_as_s8(a), v_reinterpret_as_s8(b))); }     \
-template<int imm> inline _Tpvec v_rotate_left(const _Tpvec& a)                                                                             \
-{ return v_reinterpret_as_##suffix(v_rotate_left<imm * sizeof(_Tpvec::lane_type)>(v_reinterpret_as_s8(a))); }                              \
-template<int imm> inline _Tpvec v_rotate_right(const _Tpvec& a)                                                                            \
-{ return v_reinterpret_as_##suffix(v_rotate_right<imm * sizeof(_Tpvec::lane_type)>(v_reinterpret_as_s8(a))); }
-
-#define OPENCV_HAL_IMPL_AVX512_ROTATE_EC(_Tpvec, suffix)                                                                                   \
-template<int imm>                                                                                                                          \
-inline _Tpvec v_rotate_left(const _Tpvec& a, const _Tpvec& b)                                                                              \
-{                                                                                                                                          \
-    enum { SHIFT2 = (_Tpvec::nlanes - imm) };                                                                                              \
-    enum { MASK = ((1 << _Tpvec::nlanes) - 1) };                                                                                           \
-    if (imm == 0) return a;                                                                                                                \
-    if (imm == _Tpvec::nlanes) return b;                                                                                                   \
-    if (imm >= 2*_Tpvec::nlanes) return _Tpvec::zero();                                                                                    \
-    return _Tpvec(_mm512_mask_expand_##suffix(_mm512_maskz_compress_##suffix((MASK << SHIFT2)&MASK, b.val), (MASK << (imm))&MASK, a.val)); \
-}                                                                                                                                          \
-template<int imm>                                                                                                                          \
-inline _Tpvec v_rotate_right(const _Tpvec& a, const _Tpvec& b)                                                                             \
-{                                                                                                                                          \
-    enum { SHIFT2 = (_Tpvec::nlanes - imm) };                                                                                              \
-    enum { MASK = ((1 << _Tpvec::nlanes) - 1) };                                                                                           \
-    if (imm == 0) return a;                                                                                                                \
-    if (imm == _Tpvec::nlanes) return b;                                                                                                   \
-    if (imm >= 2*_Tpvec::nlanes) return _Tpvec::zero();                                                                                    \
-    return _Tpvec(_mm512_mask_expand_##suffix(_mm512_maskz_compress_##suffix((MASK << (imm))&MASK, a.val), (MASK << SHIFT2)&MASK, b.val)); \
-}                                                                                                                                          \
-template<int imm>                                                                                                                          \
-inline _Tpvec v_rotate_left(const _Tpvec& a)                                                                                               \
-{                                                                                                                                          \
-    if (imm == 0) return a;                                                                                                                \
-    if (imm >= _Tpvec::nlanes) return _Tpvec::zero();                                                                                      \
-    return _Tpvec(_mm512_maskz_expand_##suffix((1 << _Tpvec::nlanes) - (1 << (imm)), a.val));                                              \
-}                                                                                                                                          \
-template<int imm>                                                                                                                          \
-inline _Tpvec v_rotate_right(const _Tpvec& a)                                                                                              \
-{                                                                                                                                          \
-    if (imm == 0) return a;                                                                                                                \
-    if (imm >= _Tpvec::nlanes) return _Tpvec::zero();                                                                                      \
-    return _Tpvec(_mm512_maskz_compress_##suffix((1 << _Tpvec::nlanes) - (1 << (imm)), a.val));                                            \
-}
-
-OPENCV_HAL_IMPL_AVX512_ROTATE_PM(v_uint8x64,   u8)
-OPENCV_HAL_IMPL_AVX512_ROTATE_PM(v_uint16x32,  u16)
-OPENCV_HAL_IMPL_AVX512_ROTATE_PM(v_int16x32,   s16)
-OPENCV_HAL_IMPL_AVX512_ROTATE_EC(v_uint32x16,  epi32)
-OPENCV_HAL_IMPL_AVX512_ROTATE_EC(v_int32x16,   epi32)
-OPENCV_HAL_IMPL_AVX512_ROTATE_EC(v_uint64x8,   epi64)
-OPENCV_HAL_IMPL_AVX512_ROTATE_EC(v_int64x8,    epi64)
-OPENCV_HAL_IMPL_AVX512_ROTATE_EC(v_float32x16, ps)
-OPENCV_HAL_IMPL_AVX512_ROTATE_EC(v_float64x8,  pd)
-
-/** Reverse **/
-inline v_uint8x64 v_reverse(const v_uint8x64 &a)
-{
-#if CV_AVX_512VBMI
-    static const __m512i perm = _mm512_set_epi32(
-            0x00010203, 0x04050607, 0x08090a0b, 0x0c0d0e0f,
-            0x10111213, 0x14151617, 0x18191a1b, 0x1c1d1e1f,
-            0x20212223, 0x24252627, 0x28292a2b, 0x2c2d2e2f,
-            0x30313233, 0x34353637, 0x38393a3b, 0x3c3d3e3f);
-    return v_uint8x64(_mm512_permutexvar_epi8(perm, a.val));
-#else
-    static const __m512i shuf = _mm512_set_epi32(
-            0x00010203, 0x04050607, 0x08090a0b, 0x0c0d0e0f,
-            0x00010203, 0x04050607, 0x08090a0b, 0x0c0d0e0f,
-            0x00010203, 0x04050607, 0x08090a0b, 0x0c0d0e0f,
-            0x00010203, 0x04050607, 0x08090a0b, 0x0c0d0e0f);
-    static const __m512i perm = _mm512_set_epi64(1, 0, 3, 2, 5, 4, 7, 6);
-    __m512i vec = _mm512_shuffle_epi8(a.val, shuf);
-    return v_uint8x64(_mm512_permutexvar_epi64(perm, vec));
-#endif
-}
-
-inline v_int8x64 v_reverse(const v_int8x64 &a)
-{ return v_reinterpret_as_s8(v_reverse(v_reinterpret_as_u8(a))); }
-
-inline v_uint16x32 v_reverse(const v_uint16x32 &a)
-{
-#if CV_AVX_512VBMI
-    static const __m512i perm = _mm512_set_epi32(
-            0x00000001, 0x00020003, 0x00040005, 0x00060007,
-            0x00080009, 0x000a000b, 0x000c000d, 0x000e000f,
-            0x00100011, 0x00120013, 0x00140015, 0x00160017,
-            0x00180019, 0x001a001b, 0x001c001d, 0x001e001f);
-    return v_uint16x32(_mm512_permutexvar_epi16(perm, a.val));
-#else
-    static const __m512i shuf = _mm512_set_epi32(
-            0x01000302, 0x05040706, 0x09080b0a, 0x0d0c0f0e,
-            0x01000302, 0x05040706, 0x09080b0a, 0x0d0c0f0e,
-            0x01000302, 0x05040706, 0x09080b0a, 0x0d0c0f0e,
-            0x01000302, 0x05040706, 0x09080b0a, 0x0d0c0f0e);
-    static const __m512i perm = _mm512_set_epi64(1, 0, 3, 2, 5, 4, 7, 6);
-    __m512i vec = _mm512_shuffle_epi8(a.val, shuf);
-    return v_uint16x32(_mm512_permutexvar_epi64(perm, vec));
-#endif
-}
-
-inline v_int16x32 v_reverse(const v_int16x32 &a)
-{ return v_reinterpret_as_s16(v_reverse(v_reinterpret_as_u16(a))); }
-
-inline v_uint32x16 v_reverse(const v_uint32x16 &a)
-{
-    static const __m512i perm = _mm512_set_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13,14, 15);
-    return v_uint32x16(_mm512_permutexvar_epi32(perm, a.val));
-}
-
-inline v_int32x16 v_reverse(const v_int32x16 &a)
-{ return v_reinterpret_as_s32(v_reverse(v_reinterpret_as_u32(a))); }
-
-inline v_float32x16 v_reverse(const v_float32x16 &a)
-{ return v_reinterpret_as_f32(v_reverse(v_reinterpret_as_u32(a))); }
-
-inline v_uint64x8 v_reverse(const v_uint64x8 &a)
-{
-    static const __m512i perm = _mm512_set_epi64(0, 1, 2, 3, 4, 5, 6, 7);
-    return v_uint64x8(_mm512_permutexvar_epi64(perm, a.val));
-}
-
-inline v_int64x8 v_reverse(const v_int64x8 &a)
-{ return v_reinterpret_as_s64(v_reverse(v_reinterpret_as_u64(a))); }
-
-inline v_float64x8 v_reverse(const v_float64x8 &a)
-{ return v_reinterpret_as_f64(v_reverse(v_reinterpret_as_u64(a))); }
-
-////////// Reduce /////////
-
-/** Reduce **/
-#define OPENCV_HAL_IMPL_AVX512_REDUCE_ADD64(a, b) a + b
-#define OPENCV_HAL_IMPL_AVX512_REDUCE_8(sctype, func, _Tpvec, ifunc, scop)                                          \
-    inline sctype v_reduce_##func(const _Tpvec& a)                                                                  \
-    { __m256i half = _mm256_##ifunc(_v512_extract_low(a.val), _v512_extract_high(a.val));                           \
-      sctype CV_DECL_ALIGNED(64) idx[2];                                                                            \
-      _mm_store_si128((__m128i*)idx, _mm_##ifunc(_mm256_castsi256_si128(half), _mm256_extracti128_si256(half, 1))); \
-      return scop(idx[0], idx[1]); }
-OPENCV_HAL_IMPL_AVX512_REDUCE_8(uint64, min, v_uint64x8, min_epu64, min)
-OPENCV_HAL_IMPL_AVX512_REDUCE_8(uint64, max, v_uint64x8, max_epu64, max)
-OPENCV_HAL_IMPL_AVX512_REDUCE_8(uint64, sum, v_uint64x8, add_epi64, OPENCV_HAL_IMPL_AVX512_REDUCE_ADD64)
-OPENCV_HAL_IMPL_AVX512_REDUCE_8(int64,  min, v_int64x8,  min_epi64, min)
-OPENCV_HAL_IMPL_AVX512_REDUCE_8(int64,  max, v_int64x8,  max_epi64, max)
-OPENCV_HAL_IMPL_AVX512_REDUCE_8(int64,  sum, v_int64x8,  add_epi64, OPENCV_HAL_IMPL_AVX512_REDUCE_ADD64)
-
-#define OPENCV_HAL_IMPL_AVX512_REDUCE_8F(func, ifunc, scop)                                         \
-    inline double v_reduce_##func(const v_float64x8& a)                                             \
-    { __m256d half = _mm256_##ifunc(_v512_extract_low(a.val), _v512_extract_high(a.val));           \
-      double CV_DECL_ALIGNED(64) idx[2];                                                            \
-      _mm_store_pd(idx, _mm_##ifunc(_mm256_castpd256_pd128(half), _mm256_extractf128_pd(half, 1))); \
-      return scop(idx[0], idx[1]); }
-OPENCV_HAL_IMPL_AVX512_REDUCE_8F(min, min_pd, min)
-OPENCV_HAL_IMPL_AVX512_REDUCE_8F(max, max_pd, max)
-OPENCV_HAL_IMPL_AVX512_REDUCE_8F(sum, add_pd, OPENCV_HAL_IMPL_AVX512_REDUCE_ADD64)
-
-#define OPENCV_HAL_IMPL_AVX512_REDUCE_16(sctype, func, _Tpvec, ifunc)                                 \
-    inline sctype v_reduce_##func(const _Tpvec& a)                                                    \
-    { __m256i half = _mm256_##ifunc(_v512_extract_low(a.val), _v512_extract_high(a.val));             \
-      __m128i quarter = _mm_##ifunc(_mm256_castsi256_si128(half), _mm256_extracti128_si256(half, 1)); \
-      quarter = _mm_##ifunc(quarter, _mm_srli_si128(quarter, 8));                                     \
-      quarter = _mm_##ifunc(quarter, _mm_srli_si128(quarter, 4));                                     \
-      return (sctype)_mm_cvtsi128_si32(quarter); }
-OPENCV_HAL_IMPL_AVX512_REDUCE_16(uint, min, v_uint32x16, min_epu32)
-OPENCV_HAL_IMPL_AVX512_REDUCE_16(uint, max, v_uint32x16, max_epu32)
-OPENCV_HAL_IMPL_AVX512_REDUCE_16(int,  min, v_int32x16,  min_epi32)
-OPENCV_HAL_IMPL_AVX512_REDUCE_16(int,  max, v_int32x16,  max_epi32)
-
-#define OPENCV_HAL_IMPL_AVX512_REDUCE_16F(func, ifunc)                                            \
-    inline float v_reduce_##func(const v_float32x16& a)                                           \
-    { __m256 half = _mm256_##ifunc(_v512_extract_low(a.val), _v512_extract_high(a.val));          \
-      __m128 quarter = _mm_##ifunc(_mm256_castps256_ps128(half), _mm256_extractf128_ps(half, 1)); \
-      quarter = _mm_##ifunc(quarter, _mm_permute_ps(quarter, _MM_SHUFFLE(0, 0, 3, 2)));           \
-      quarter = _mm_##ifunc(quarter, _mm_permute_ps(quarter, _MM_SHUFFLE(0, 0, 0, 1)));           \
-      return _mm_cvtss_f32(quarter); }
-OPENCV_HAL_IMPL_AVX512_REDUCE_16F(min, min_ps)
-OPENCV_HAL_IMPL_AVX512_REDUCE_16F(max, max_ps)
-
-inline float v_reduce_sum(const v_float32x16& a)
-{
-    __m256 half = _mm256_add_ps(_v512_extract_low(a.val), _v512_extract_high(a.val));
-    __m128 quarter = _mm_add_ps(_mm256_castps256_ps128(half), _mm256_extractf128_ps(half, 1));
-    quarter = _mm_hadd_ps(quarter, quarter);
-    return _mm_cvtss_f32(_mm_hadd_ps(quarter, quarter));
-}
-inline int v_reduce_sum(const v_int32x16& a)
-{
-    __m256i half = _mm256_add_epi32(_v512_extract_low(a.val), _v512_extract_high(a.val));
-    __m128i quarter = _mm_add_epi32(_mm256_castsi256_si128(half), _mm256_extracti128_si256(half, 1));
-    quarter = _mm_hadd_epi32(quarter, quarter);
-    return _mm_cvtsi128_si32(_mm_hadd_epi32(quarter, quarter));
-}
-inline uint v_reduce_sum(const v_uint32x16& a)
-{ return (uint)v_reduce_sum(v_reinterpret_as_s32(a)); }
-
-#define OPENCV_HAL_IMPL_AVX512_REDUCE_32(sctype, func, _Tpvec, ifunc)                                 \
-    inline sctype v_reduce_##func(const _Tpvec& a)                                                    \
-    { __m256i half = _mm256_##ifunc(_v512_extract_low(a.val), _v512_extract_high(a.val));             \
-      __m128i quarter = _mm_##ifunc(_mm256_castsi256_si128(half), _mm256_extracti128_si256(half, 1)); \
-      quarter = _mm_##ifunc(quarter, _mm_srli_si128(quarter, 8));                                     \
-      quarter = _mm_##ifunc(quarter, _mm_srli_si128(quarter, 4));                                     \
-      quarter = _mm_##ifunc(quarter, _mm_srli_si128(quarter, 2));                                     \
-      return (sctype)_mm_cvtsi128_si32(quarter); }
-OPENCV_HAL_IMPL_AVX512_REDUCE_32(ushort, min, v_uint16x32, min_epu16)
-OPENCV_HAL_IMPL_AVX512_REDUCE_32(ushort, max, v_uint16x32, max_epu16)
-OPENCV_HAL_IMPL_AVX512_REDUCE_32(short,  min, v_int16x32,  min_epi16)
-OPENCV_HAL_IMPL_AVX512_REDUCE_32(short,  max, v_int16x32,  max_epi16)
-
-inline int v_reduce_sum(const v_int16x32& a)
-{ return v_reduce_sum(v_expand_low(a) + v_expand_high(a)); }
-inline uint v_reduce_sum(const v_uint16x32& a)
-{ return v_reduce_sum(v_expand_low(a) + v_expand_high(a)); }
-
-#define OPENCV_HAL_IMPL_AVX512_REDUCE_64(sctype, func, _Tpvec, ifunc)                                 \
-    inline sctype v_reduce_##func(const _Tpvec& a)                                                    \
-    { __m256i half = _mm256_##ifunc(_v512_extract_low(a.val), _v512_extract_high(a.val));             \
-      __m128i quarter = _mm_##ifunc(_mm256_castsi256_si128(half), _mm256_extracti128_si256(half, 1)); \
-      quarter = _mm_##ifunc(quarter, _mm_srli_si128(quarter, 8));                                     \
-      quarter = _mm_##ifunc(quarter, _mm_srli_si128(quarter, 4));                                     \
-      quarter = _mm_##ifunc(quarter, _mm_srli_si128(quarter, 2));                                     \
-      quarter = _mm_##ifunc(quarter, _mm_srli_si128(quarter, 1));                                     \
-      return (sctype)_mm_cvtsi128_si32(quarter); }
-OPENCV_HAL_IMPL_AVX512_REDUCE_64(uchar, min, v_uint8x64, min_epu8)
-OPENCV_HAL_IMPL_AVX512_REDUCE_64(uchar, max, v_uint8x64, max_epu8)
-OPENCV_HAL_IMPL_AVX512_REDUCE_64(schar, min, v_int8x64,  min_epi8)
-OPENCV_HAL_IMPL_AVX512_REDUCE_64(schar, max, v_int8x64,  max_epi8)
-
-#define OPENCV_HAL_IMPL_AVX512_REDUCE_64_SUM(sctype, _Tpvec, suffix)                                    \
-    inline sctype v_reduce_sum(const _Tpvec& a)                                                         \
-    {   __m512i a16 = _mm512_add_epi16(_mm512_cvt##suffix##_epi16(_v512_extract_low(a.val)),            \
-                                       _mm512_cvt##suffix##_epi16(_v512_extract_high(a.val)));          \
-        a16 = _mm512_cvtepi16_epi32(_mm256_add_epi16(_v512_extract_low(a16), _v512_extract_high(a16))); \
-        __m256i a8 = _mm256_add_epi32(_v512_extract_low(a16), _v512_extract_high(a16));                 \
-        __m128i a4 = _mm_add_epi32(_mm256_castsi256_si128(a8), _mm256_extracti128_si256(a8, 1));        \
-        a4 = _mm_hadd_epi32(a4, a4);                                                                    \
-        return (sctype)_mm_cvtsi128_si32(_mm_hadd_epi32(a4, a4)); }
-OPENCV_HAL_IMPL_AVX512_REDUCE_64_SUM(uint, v_uint8x64, epu8)
-OPENCV_HAL_IMPL_AVX512_REDUCE_64_SUM(int,  v_int8x64,  epi8)
-
-inline v_float32x16 v_reduce_sum4(const v_float32x16& a, const v_float32x16& b,
-                                  const v_float32x16& c, const v_float32x16& d)
-{
-    __m256 abl = _mm256_hadd_ps(_v512_extract_low(a.val), _v512_extract_low(b.val));
-    __m256 abh = _mm256_hadd_ps(_v512_extract_high(a.val), _v512_extract_high(b.val));
-    __m256 cdl = _mm256_hadd_ps(_v512_extract_low(c.val), _v512_extract_low(d.val));
-    __m256 cdh = _mm256_hadd_ps(_v512_extract_high(c.val), _v512_extract_high(d.val));
-    return v_float32x16(_v512_combine(_mm256_hadd_ps(abl, cdl), _mm256_hadd_ps(abh, cdh)));
-}
-
-inline unsigned v_reduce_sad(const v_uint8x64& a, const v_uint8x64& b)
-{
-    __m512i val = _mm512_sad_epu8(a.val, b.val);
-    __m256i half = _mm256_add_epi32(_v512_extract_low(val), _v512_extract_high(val));
-    __m128i quarter = _mm_add_epi32(_mm256_castsi256_si128(half), _mm256_extracti128_si256(half, 1));
-    return (unsigned)_mm_cvtsi128_si32(_mm_add_epi32(quarter, _mm_unpackhi_epi64(quarter, quarter)));
-}
-inline unsigned v_reduce_sad(const v_int8x64& a, const v_int8x64& b)
-{
-    __m512i val = _mm512_set1_epi8(-128);
-    val = _mm512_sad_epu8(_mm512_add_epi8(a.val, val), _mm512_add_epi8(b.val, val));
-    __m256i half = _mm256_add_epi32(_v512_extract_low(val), _v512_extract_high(val));
-    __m128i quarter = _mm_add_epi32(_mm256_castsi256_si128(half), _mm256_extracti128_si256(half, 1));
-    return (unsigned)_mm_cvtsi128_si32(_mm_add_epi32(quarter, _mm_unpackhi_epi64(quarter, quarter)));
-}
-inline unsigned v_reduce_sad(const v_uint16x32& a, const v_uint16x32& b)
-{ return v_reduce_sum(v_add_wrap(a - b, b - a)); }
-inline unsigned v_reduce_sad(const v_int16x32& a, const v_int16x32& b)
-{ return v_reduce_sum(v_reinterpret_as_u16(v_sub_wrap(v_max(a, b), v_min(a, b)))); }
-inline unsigned v_reduce_sad(const v_uint32x16& a, const v_uint32x16& b)
-{ return v_reduce_sum(v_max(a, b) - v_min(a, b)); }
-inline unsigned v_reduce_sad(const v_int32x16& a, const v_int32x16& b)
-{ return v_reduce_sum(v_reinterpret_as_u32(v_max(a, b) - v_min(a, b))); }
-inline float v_reduce_sad(const v_float32x16& a, const v_float32x16& b)
-{ return v_reduce_sum((a - b) & v_float32x16(_mm512_castsi512_ps(_mm512_set1_epi32(0x7fffffff)))); }
-inline double v_reduce_sad(const v_float64x8& a, const v_float64x8& b)
-{ return v_reduce_sum((a - b) & v_float64x8(_mm512_castsi512_pd(_mm512_set1_epi64(0x7fffffffffffffff)))); }
-
-/** Popcount **/
-inline v_uint8x64 v_popcount(const v_int8x64& a)
-{
-#if CV_AVX_512BITALG
-    return v_uint8x64(_mm512_popcnt_epi8(a.val));
-#elif CV_AVX_512VBMI
-    __m512i _popcnt_table0 = _v512_set_epu8(7, 6, 6, 5, 6, 5, 5, 4, 6, 5, 5, 4, 5, 4, 4, 3,
-                                            5, 4, 4, 3, 4, 3, 3, 2, 4, 3, 3, 2, 3, 2, 2, 1,
-                                            5, 4, 4, 3, 4, 3, 3, 2, 4, 3, 3, 2, 3, 2, 2, 1,
-                                            4, 3, 3, 2, 3, 2, 2, 1, 3, 2, 2, 1, 2, 1, 1, 0);
-    __m512i _popcnt_table1 = _v512_set_epu8(7, 6, 6, 5, 6, 5, 5, 4, 6, 5, 5, 4, 5, 4, 4, 3,
-                                            6, 5, 5, 4, 5, 4, 4, 3, 5, 4, 4, 3, 4, 3, 3, 2,
-                                            6, 5, 5, 4, 5, 4, 4, 3, 5, 4, 4, 3, 4, 3, 3, 2,
-                                            5, 4, 4, 3, 4, 3, 3, 2, 4, 3, 3, 2, 3, 2, 2, 1);
-    return v_uint8x64(_mm512_sub_epi8(_mm512_permutex2var_epi8(_popcnt_table0, a.val, _popcnt_table1), _mm512_movm_epi8(_mm512_movepi8_mask(a.val))));
-#else
-    __m512i _popcnt_table = _mm512_set4_epi32(0x04030302, 0x03020201, 0x03020201, 0x02010100);
-    __m512i _popcnt_mask = _mm512_set1_epi8(0x0F);
-
-    return v_uint8x64(_mm512_add_epi8(_mm512_shuffle_epi8(_popcnt_table, _mm512_and_si512(                  a.val,     _popcnt_mask)),
-                                      _mm512_shuffle_epi8(_popcnt_table, _mm512_and_si512(_mm512_srli_epi16(a.val, 4), _popcnt_mask))));
-#endif
-}
-inline v_uint16x32 v_popcount(const v_int16x32& a)
-{
-#if CV_AVX_512BITALG
-    return v_uint16x32(_mm512_popcnt_epi16(a.val));
-#elif CV_AVX_512VPOPCNTDQ
-    __m512i zero = _mm512_setzero_si512();
-    return v_uint16x32(_mm512_packs_epi32(_mm512_popcnt_epi32(_mm512_unpacklo_epi16(a.val, zero)),
-                                          _mm512_popcnt_epi32(_mm512_unpackhi_epi16(a.val, zero))));
-#else
-    v_uint8x64 p = v_popcount(v_reinterpret_as_s8(a));
-    p += v_rotate_right<1>(p);
-    return v_reinterpret_as_u16(p) & v512_setall_u16(0x00ff);
-#endif
-}
-inline v_uint32x16 v_popcount(const v_int32x16& a)
-{
-#if CV_AVX_512VPOPCNTDQ
-    return v_uint32x16(_mm512_popcnt_epi32(a.val));
-#else
-    v_uint8x64 p = v_popcount(v_reinterpret_as_s8(a));
-    p += v_rotate_right<1>(p);
-    p += v_rotate_right<2>(p);
-    return v_reinterpret_as_u32(p) & v512_setall_u32(0x000000ff);
-#endif
-}
-inline v_uint64x8 v_popcount(const v_int64x8& a)
-{
-#if CV_AVX_512VPOPCNTDQ
-    return v_uint64x8(_mm512_popcnt_epi64(a.val));
-#else
-    return v_uint64x8(_mm512_sad_epu8(v_popcount(v_reinterpret_as_s8(a)).val, _mm512_setzero_si512()));
-#endif
-}
-
-
-inline v_uint8x64  v_popcount(const v_uint8x64&  a) { return v_popcount(v_reinterpret_as_s8 (a)); }
-inline v_uint16x32 v_popcount(const v_uint16x32& a) { return v_popcount(v_reinterpret_as_s16(a)); }
-inline v_uint32x16 v_popcount(const v_uint32x16& a) { return v_popcount(v_reinterpret_as_s32(a)); }
-inline v_uint64x8  v_popcount(const v_uint64x8&  a) { return v_popcount(v_reinterpret_as_s64(a)); }
-
-
-////////// Other math /////////
-
-/** Some frequent operations **/
-#if CV_FMA3
-#define OPENCV_HAL_IMPL_AVX512_MULADD(_Tpvec, suffix)                         \
-    inline _Tpvec v_fma(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c)    \
-    { return _Tpvec(_mm512_fmadd_##suffix(a.val, b.val, c.val)); }            \
-    inline _Tpvec v_muladd(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c) \
-    { return _Tpvec(_mm512_fmadd_##suffix(a.val, b.val, c.val)); }
-#else
-#define OPENCV_HAL_IMPL_AVX512_MULADD(_Tpvec, suffix)                                 \
-    inline _Tpvec v_fma(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c)            \
-    { return _Tpvec(_mm512_add_##suffix(_mm512_mul_##suffix(a.val, b.val), c.val)); } \
-    inline _Tpvec v_muladd(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c)         \
-    { return _Tpvec(_mm512_add_##suffix(_mm512_mul_##suffix(a.val, b.val), c.val)); }
-#endif
-
-#define OPENCV_HAL_IMPL_AVX512_MISC(_Tpvec, suffix)                           \
-    inline _Tpvec v_sqrt(const _Tpvec& x)                                     \
-    { return _Tpvec(_mm512_sqrt_##suffix(x.val)); }                           \
-    inline _Tpvec v_sqr_magnitude(const _Tpvec& a, const _Tpvec& b)           \
-    { return v_fma(a, a, b * b); }                                            \
-    inline _Tpvec v_magnitude(const _Tpvec& a, const _Tpvec& b)               \
-    { return v_sqrt(v_fma(a, a, b * b)); }
-
-OPENCV_HAL_IMPL_AVX512_MULADD(v_float32x16, ps)
-OPENCV_HAL_IMPL_AVX512_MULADD(v_float64x8,  pd)
-OPENCV_HAL_IMPL_AVX512_MISC(v_float32x16, ps)
-OPENCV_HAL_IMPL_AVX512_MISC(v_float64x8,  pd)
-
-inline v_int32x16 v_fma(const v_int32x16& a, const v_int32x16& b, const v_int32x16& c)
-{ return a * b + c; }
-inline v_int32x16 v_muladd(const v_int32x16& a, const v_int32x16& b, const v_int32x16& c)
-{ return v_fma(a, b, c); }
-
-inline v_float32x16 v_invsqrt(const v_float32x16& x)
-{
-#if CV_AVX_512ER
-    return v_float32x16(_mm512_rsqrt28_ps(x.val));
-#else
-    v_float32x16 half = x * v512_setall_f32(0.5);
-    v_float32x16 t  = v_float32x16(_mm512_rsqrt14_ps(x.val));
-    t *= v512_setall_f32(1.5) - ((t * t) * half);
-    return t;
-#endif
-}
-
-inline v_float64x8 v_invsqrt(const v_float64x8& x)
-{
-#if CV_AVX_512ER
-    return v_float64x8(_mm512_rsqrt28_pd(x.val));
-#else
-    return v512_setall_f64(1.) / v_sqrt(x);
-//    v_float64x8 half = x * v512_setall_f64(0.5);
-//    v_float64x8 t = v_float64x8(_mm512_rsqrt14_pd(x.val));
-//    t *= v512_setall_f64(1.5) - ((t * t) * half);
-//    t *= v512_setall_f64(1.5) - ((t * t) * half);
-//    return t;
-#endif
-}
-
-/** Absolute values **/
-#define OPENCV_HAL_IMPL_AVX512_ABS(_Tpvec, _Tpuvec, suffix) \
-    inline _Tpuvec v_abs(const _Tpvec& x)                   \
-    { return _Tpuvec(_mm512_abs_##suffix(x.val)); }
-
-OPENCV_HAL_IMPL_AVX512_ABS(v_int8x64,    v_uint8x64,    epi8)
-OPENCV_HAL_IMPL_AVX512_ABS(v_int16x32,   v_uint16x32,  epi16)
-OPENCV_HAL_IMPL_AVX512_ABS(v_int32x16,   v_uint32x16,  epi32)
-OPENCV_HAL_IMPL_AVX512_ABS(v_int64x8,    v_uint64x8,   epi64)
-
-inline v_float32x16 v_abs(const v_float32x16& x)
-{
-#ifdef _mm512_abs_pd
-    return v_float32x16(_mm512_abs_ps(x.val));
-#else
-    return v_float32x16(_mm512_castsi512_ps(_mm512_and_si512(_mm512_castps_si512(x.val),
-                        _v512_set_epu64(0x7FFFFFFF7FFFFFFF, 0x7FFFFFFF7FFFFFFF, 0x7FFFFFFF7FFFFFFF, 0x7FFFFFFF7FFFFFFF,
-                                        0x7FFFFFFF7FFFFFFF, 0x7FFFFFFF7FFFFFFF, 0x7FFFFFFF7FFFFFFF, 0x7FFFFFFF7FFFFFFF))));
-#endif
-}
-
-inline v_float64x8 v_abs(const v_float64x8& x)
-{
-#ifdef _mm512_abs_pd
-    #if defined __GNUC__ && (__GNUC__ < 7 || (__GNUC__ == 7 && __GNUC_MINOR__ <= 3) || (__GNUC__ == 8 && __GNUC_MINOR__ <= 2))
-        // Workaround for https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87476
-        return v_float64x8(_mm512_abs_pd(_mm512_castpd_ps(x.val)));
-    #else
-        return v_float64x8(_mm512_abs_pd(x.val));
-    #endif
-#else
-    return v_float64x8(_mm512_castsi512_pd(_mm512_and_si512(_mm512_castpd_si512(x.val),
-                       _v512_set_epu64(0x7FFFFFFFFFFFFFFF, 0x7FFFFFFFFFFFFFFF, 0x7FFFFFFFFFFFFFFF, 0x7FFFFFFFFFFFFFFF,
-                                       0x7FFFFFFFFFFFFFFF, 0x7FFFFFFFFFFFFFFF, 0x7FFFFFFFFFFFFFFF, 0x7FFFFFFFFFFFFFFF))));
-#endif
-}
-
-/** Absolute difference **/
-inline v_uint8x64 v_absdiff(const v_uint8x64& a, const v_uint8x64& b)
-{ return v_add_wrap(a - b,  b - a); }
-inline v_uint16x32 v_absdiff(const v_uint16x32& a, const v_uint16x32& b)
-{ return v_add_wrap(a - b,  b - a); }
-inline v_uint32x16 v_absdiff(const v_uint32x16& a, const v_uint32x16& b)
-{ return v_max(a, b) - v_min(a, b); }
-
-inline v_uint8x64 v_absdiff(const v_int8x64& a, const v_int8x64& b)
-{
-    v_int8x64 d = v_sub_wrap(a, b);
-    v_int8x64 m = a < b;
-    return v_reinterpret_as_u8(v_sub_wrap(d ^ m, m));
-}
-
-inline v_uint16x32 v_absdiff(const v_int16x32& a, const v_int16x32& b)
-{ return v_reinterpret_as_u16(v_sub_wrap(v_max(a, b), v_min(a, b))); }
-
-inline v_uint32x16 v_absdiff(const v_int32x16& a, const v_int32x16& b)
-{
-    v_int32x16 d = a - b;
-    v_int32x16 m = a < b;
-    return v_reinterpret_as_u32((d ^ m) - m);
-}
-
-inline v_float32x16 v_absdiff(const v_float32x16& a, const v_float32x16& b)
-{ return v_abs(a - b); }
-
-inline v_float64x8 v_absdiff(const v_float64x8& a, const v_float64x8& b)
-{ return v_abs(a - b); }
-
-/** Saturating absolute difference **/
-inline v_int8x64 v_absdiffs(const v_int8x64& a, const v_int8x64& b)
-{
-    v_int8x64 d = a - b;
-    v_int8x64 m = a < b;
-    return (d ^ m) - m;
-}
-inline v_int16x32 v_absdiffs(const v_int16x32& a, const v_int16x32& b)
-{ return v_max(a, b) - v_min(a, b); }
-
-////////// Conversions /////////
-
-/** Rounding **/
-inline v_int32x16 v_round(const v_float32x16& a)
-{ return v_int32x16(_mm512_cvtps_epi32(a.val)); }
-
-inline v_int32x16 v_round(const v_float64x8& a)
-{ return v_int32x16(_mm512_castsi256_si512(_mm512_cvtpd_epi32(a.val))); }
-
-inline v_int32x16 v_round(const v_float64x8& a, const v_float64x8& b)
-{ return v_int32x16(_v512_combine(_mm512_cvtpd_epi32(a.val), _mm512_cvtpd_epi32(b.val))); }
-
-inline v_int32x16 v_trunc(const v_float32x16& a)
-{ return v_int32x16(_mm512_cvttps_epi32(a.val)); }
-
-inline v_int32x16 v_trunc(const v_float64x8& a)
-{ return v_int32x16(_mm512_castsi256_si512(_mm512_cvttpd_epi32(a.val))); }
-
-#if CVT_ROUND_MODES_IMPLEMENTED
-inline v_int32x16 v_floor(const v_float32x16& a)
-{ return v_int32x16(_mm512_cvt_roundps_epi32(a.val, _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC)); }
-
-inline v_int32x16 v_floor(const v_float64x8& a)
-{ return v_int32x16(_mm512_castsi256_si512(_mm512_cvt_roundpd_epi32(a.val, _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC))); }
-
-inline v_int32x16 v_ceil(const v_float32x16& a)
-{ return v_int32x16(_mm512_cvt_roundps_epi32(a.val, _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC)); }
-
-inline v_int32x16 v_ceil(const v_float64x8& a)
-{ return v_int32x16(_mm512_castsi256_si512(_mm512_cvt_roundpd_epi32(a.val, _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC))); }
-#else
-inline v_int32x16 v_floor(const v_float32x16& a)
-{ return v_int32x16(_mm512_cvtps_epi32(_mm512_roundscale_ps(a.val, 1))); }
-
-inline v_int32x16 v_floor(const v_float64x8& a)
-{ return v_int32x16(_mm512_castsi256_si512(_mm512_cvtpd_epi32(_mm512_roundscale_pd(a.val, 1)))); }
-
-inline v_int32x16 v_ceil(const v_float32x16& a)
-{ return v_int32x16(_mm512_cvtps_epi32(_mm512_roundscale_ps(a.val, 2))); }
-
-inline v_int32x16 v_ceil(const v_float64x8& a)
-{ return v_int32x16(_mm512_castsi256_si512(_mm512_cvtpd_epi32(_mm512_roundscale_pd(a.val, 2)))); }
-#endif
-
-/** To float **/
-inline v_float32x16 v_cvt_f32(const v_int32x16& a)
-{ return v_float32x16(_mm512_cvtepi32_ps(a.val)); }
-
-inline v_float32x16 v_cvt_f32(const v_float64x8& a)
-{ return v_float32x16(_mm512_cvtpd_pslo(a.val)); }
-
-inline v_float32x16 v_cvt_f32(const v_float64x8& a, const v_float64x8& b)
-{ return v_float32x16(_v512_combine(_mm512_cvtpd_ps(a.val), _mm512_cvtpd_ps(b.val))); }
-
-inline v_float64x8 v_cvt_f64(const v_int32x16& a)
-{ return v_float64x8(_mm512_cvtepi32_pd(_v512_extract_low(a.val))); }
-
-inline v_float64x8 v_cvt_f64_high(const v_int32x16& a)
-{ return v_float64x8(_mm512_cvtepi32_pd(_v512_extract_high(a.val))); }
-
-inline v_float64x8 v_cvt_f64(const v_float32x16& a)
-{ return v_float64x8(_mm512_cvtps_pd(_v512_extract_low(a.val))); }
-
-inline v_float64x8 v_cvt_f64_high(const v_float32x16& a)
-{ return v_float64x8(_mm512_cvtps_pd(_v512_extract_high(a.val))); }
-
-// from (Mysticial and wim) https://stackoverflow.com/q/41144668
-inline v_float64x8 v_cvt_f64(const v_int64x8& v)
-{
-#if CV_AVX_512DQ
-    return v_float64x8(_mm512_cvtepi64_pd(v.val));
-#else
-    // constants encoded as floating-point
-    __m512i magic_i_lo   = _mm512_set1_epi64(0x4330000000000000); // 2^52
-    __m512i magic_i_hi32 = _mm512_set1_epi64(0x4530000080000000); // 2^84 + 2^63
-    __m512i magic_i_all  = _mm512_set1_epi64(0x4530000080100000); // 2^84 + 2^63 + 2^52
-    __m512d magic_d_all  = _mm512_castsi512_pd(magic_i_all);
-
-    // Blend the 32 lowest significant bits of v with magic_int_lo
-    __m512i v_lo         = _mm512_mask_blend_epi32(0x5555, magic_i_lo, v.val);
-    // Extract the 32 most significant bits of v
-    __m512i v_hi         = _mm512_srli_epi64(v.val, 32);
-    // Flip the msb of v_hi and blend with 0x45300000
-            v_hi         = _mm512_xor_si512(v_hi, magic_i_hi32);
-    // Compute in double precision
-    __m512d v_hi_dbl     = _mm512_sub_pd(_mm512_castsi512_pd(v_hi), magic_d_all);
-    // (v_hi - magic_d_all) + v_lo  Do not assume associativity of floating point addition
-    __m512d result       = _mm512_add_pd(v_hi_dbl, _mm512_castsi512_pd(v_lo));
-    return v_float64x8(result);
-#endif
-}
-
-////////////// Lookup table access ////////////////////
-
-inline v_int8x64 v512_lut(const schar* tab, const int* idx)
-{
-    __m128i p0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_mm512_loadu_si512((const __m512i*)idx    ), (const int *)tab, 1));
-    __m128i p1 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_mm512_loadu_si512((const __m512i*)idx + 1), (const int *)tab, 1));
-    __m128i p2 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_mm512_loadu_si512((const __m512i*)idx + 2), (const int *)tab, 1));
-    __m128i p3 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_mm512_loadu_si512((const __m512i*)idx + 3), (const int *)tab, 1));
-    return v_int8x64(_mm512_inserti32x4(_mm512_inserti32x4(_mm512_inserti32x4(_mm512_castsi128_si512(p0), p1, 1), p2, 2), p3, 3));
-}
-inline v_int8x64 v512_lut_pairs(const schar* tab, const int* idx)
-{
-    __m256i p0 = _mm512_cvtepi32_epi16(_mm512_i32gather_epi32(_mm512_loadu_si512((const __m512i*)idx    ), (const int *)tab, 1));
-    __m256i p1 = _mm512_cvtepi32_epi16(_mm512_i32gather_epi32(_mm512_loadu_si512((const __m512i*)idx + 1), (const int *)tab, 1));
-    return v_int8x64(_v512_combine(p0, p1));
-}
-inline v_int8x64 v512_lut_quads(const schar* tab, const int* idx)
-{
-    return v_int8x64(_mm512_i32gather_epi32(_mm512_loadu_si512((const __m512i*)idx), (const int *)tab, 1));
-}
-inline v_uint8x64 v512_lut(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v512_lut((const schar *)tab, idx)); }
-inline v_uint8x64 v512_lut_pairs(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v512_lut_pairs((const schar *)tab, idx)); }
-inline v_uint8x64 v512_lut_quads(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v512_lut_quads((const schar *)tab, idx)); }
-
-inline v_int16x32 v512_lut(const short* tab, const int* idx)
-{
-    __m256i p0 = _mm512_cvtepi32_epi16(_mm512_i32gather_epi32(_mm512_loadu_si512((const __m512i*)idx    ), (const int *)tab, 2));
-    __m256i p1 = _mm512_cvtepi32_epi16(_mm512_i32gather_epi32(_mm512_loadu_si512((const __m512i*)idx + 1), (const int *)tab, 2));
-    return v_int16x32(_v512_combine(p0, p1));
-}
-inline v_int16x32 v512_lut_pairs(const short* tab, const int* idx)
-{
-    return v_int16x32(_mm512_i32gather_epi32(_mm512_loadu_si512((const __m512i*)idx), (const int *)tab, 2));
-}
-inline v_int16x32 v512_lut_quads(const short* tab, const int* idx)
-{
-#if defined(__GNUC__)
-    return v_int16x32(_mm512_i32gather_epi64(_mm256_loadu_si256((const __m256i*)idx), (const long long int*)tab, 2));
-#else
-    return v_int16x32(_mm512_i32gather_epi64(_mm256_loadu_si256((const __m256i*)idx), (const int64*)tab, 2));
-#endif
-}
-inline v_uint16x32 v512_lut(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v512_lut((const short *)tab, idx)); }
-inline v_uint16x32 v512_lut_pairs(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v512_lut_pairs((const short *)tab, idx)); }
-inline v_uint16x32 v512_lut_quads(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v512_lut_quads((const short *)tab, idx)); }
-
-inline v_int32x16 v512_lut(const int* tab, const int* idx)
-{
-    return v_int32x16(_mm512_i32gather_epi32(_mm512_loadu_si512((const __m512i*)idx), tab, 4));
-}
-inline v_int32x16 v512_lut_pairs(const int* tab, const int* idx)
-{
-#if defined(__GNUC__)
-    return v_int32x16(_mm512_i32gather_epi64(_mm256_loadu_si256((const __m256i*)idx), (const long long int*)tab, 4));
-#else
-    return v_int32x16(_mm512_i32gather_epi64(_mm256_loadu_si256((const __m256i*)idx), (const int64*)tab, 4));
-#endif
-}
-inline v_int32x16 v512_lut_quads(const int* tab, const int* idx)
-{
-    return v_int32x16(_mm512_inserti32x4(_mm512_inserti32x4(_mm512_inserti32x4(_mm512_castsi128_si512(
-                          _mm_loadu_si128((const __m128i*)(tab + idx[0]))),
-                          _mm_loadu_si128((const __m128i*)(tab + idx[1])), 1),
-                          _mm_loadu_si128((const __m128i*)(tab + idx[2])), 2),
-                          _mm_loadu_si128((const __m128i*)(tab + idx[3])), 3));
-}
-inline v_uint32x16 v512_lut(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v512_lut((const int *)tab, idx)); }
-inline v_uint32x16 v512_lut_pairs(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v512_lut_pairs((const int *)tab, idx)); }
-inline v_uint32x16 v512_lut_quads(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v512_lut_quads((const int *)tab, idx)); }
-
-inline v_int64x8 v512_lut(const int64* tab, const int* idx)
-{
-#if defined(__GNUC__)
-    return v_int64x8(_mm512_i32gather_epi64(_mm256_loadu_si256((const __m256i*)idx), (const long long int*)tab, 8));
-#else
-    return v_int64x8(_mm512_i32gather_epi64(_mm256_loadu_si256((const __m256i*)idx), tab , 8));
-#endif
-}
-inline v_int64x8 v512_lut_pairs(const int64* tab, const int* idx)
-{
-    return v_int64x8(_mm512_inserti32x4(_mm512_inserti32x4(_mm512_inserti32x4(_mm512_castsi128_si512(
-                         _mm_loadu_si128((const __m128i*)(tab + idx[0]))),
-                         _mm_loadu_si128((const __m128i*)(tab + idx[1])), 1),
-                         _mm_loadu_si128((const __m128i*)(tab + idx[2])), 2),
-                         _mm_loadu_si128((const __m128i*)(tab + idx[3])), 3));
-}
-inline v_uint64x8 v512_lut(const uint64* tab, const int* idx) { return v_reinterpret_as_u64(v512_lut((const int64 *)tab, idx)); }
-inline v_uint64x8 v512_lut_pairs(const uint64* tab, const int* idx) { return v_reinterpret_as_u64(v512_lut_pairs((const int64 *)tab, idx)); }
-
-inline v_float32x16 v512_lut(const float* tab, const int* idx)
-{
-    return v_float32x16(_mm512_i32gather_ps(_mm512_loadu_si512((const __m512i*)idx), tab, 4));
-}
-inline v_float32x16 v512_lut_pairs(const float* tab, const int* idx) { return v_reinterpret_as_f32(v512_lut_pairs((const int *)tab, idx)); }
-inline v_float32x16 v512_lut_quads(const float* tab, const int* idx) { return v_reinterpret_as_f32(v512_lut_quads((const int *)tab, idx)); }
-
-inline v_float64x8 v512_lut(const double* tab, const int* idx)
-{
-    return v_float64x8(_mm512_i32gather_pd(_mm256_loadu_si256((const __m256i*)idx), tab, 8));
-}
-inline v_float64x8 v512_lut_pairs(const double* tab, const int* idx)
-{
-        return v_float64x8(_mm512_insertf64x2(_mm512_insertf64x2(_mm512_insertf64x2(_mm512_castpd128_pd512(
-                               _mm_loadu_pd(tab + idx[0])),
-                               _mm_loadu_pd(tab + idx[1]), 1),
-                               _mm_loadu_pd(tab + idx[2]), 2),
-                               _mm_loadu_pd(tab + idx[3]), 3));
-}
-
-inline v_int32x16 v_lut(const int* tab, const v_int32x16& idxvec)
-{
-    return v_int32x16(_mm512_i32gather_epi32(idxvec.val, tab, 4));
-}
-
-inline v_uint32x16 v_lut(const unsigned* tab, const v_int32x16& idxvec)
-{
-    return v_reinterpret_as_u32(v_lut((const int *)tab, idxvec));
-}
-
-inline v_float32x16 v_lut(const float* tab, const v_int32x16& idxvec)
-{
-    return v_float32x16(_mm512_i32gather_ps(idxvec.val, tab, 4));
-}
-
-inline v_float64x8 v_lut(const double* tab, const v_int32x16& idxvec)
-{
-    return v_float64x8(_mm512_i32gather_pd(_v512_extract_low(idxvec.val), tab, 8));
-}
-
-inline void v_lut_deinterleave(const float* tab, const v_int32x16& idxvec, v_float32x16& x, v_float32x16& y)
-{
-    x.val = _mm512_i32gather_ps(idxvec.val, tab, 4);
-    y.val = _mm512_i32gather_ps(idxvec.val, &tab[1], 4);
-}
-
-inline void v_lut_deinterleave(const double* tab, const v_int32x16& idxvec, v_float64x8& x, v_float64x8& y)
-{
-    x.val = _mm512_i32gather_pd(_v512_extract_low(idxvec.val), tab, 8);
-    y.val = _mm512_i32gather_pd(_v512_extract_low(idxvec.val), &tab[1], 8);
-}
-
-inline v_int8x64 v_interleave_pairs(const v_int8x64& vec)
-{
-    return v_int8x64(_mm512_shuffle_epi8(vec.val, _mm512_set4_epi32(0x0f0d0e0c, 0x0b090a08, 0x07050604, 0x03010200)));
-}
-inline v_uint8x64 v_interleave_pairs(const v_uint8x64& vec) { return v_reinterpret_as_u8(v_interleave_pairs(v_reinterpret_as_s8(vec))); }
-inline v_int8x64 v_interleave_quads(const v_int8x64& vec)
-{
-    return v_int8x64(_mm512_shuffle_epi8(vec.val, _mm512_set4_epi32(0x0f0b0e0a, 0x0d090c08, 0x07030602, 0x05010400)));
-}
-inline v_uint8x64 v_interleave_quads(const v_uint8x64& vec) { return v_reinterpret_as_u8(v_interleave_quads(v_reinterpret_as_s8(vec))); }
-
-inline v_int16x32 v_interleave_pairs(const v_int16x32& vec)
-{
-    return v_int16x32(_mm512_shuffle_epi8(vec.val, _mm512_set4_epi32(0x0f0e0b0a, 0x0d0c0908, 0x07060302, 0x05040100)));
-}
-inline v_uint16x32 v_interleave_pairs(const v_uint16x32& vec) { return v_reinterpret_as_u16(v_interleave_pairs(v_reinterpret_as_s16(vec))); }
-inline v_int16x32 v_interleave_quads(const v_int16x32& vec)
-{
-    return v_int16x32(_mm512_shuffle_epi8(vec.val, _mm512_set4_epi32(0x0f0e0706, 0x0d0c0504, 0x0b0a0302, 0x09080100)));
-}
-inline v_uint16x32 v_interleave_quads(const v_uint16x32& vec) { return v_reinterpret_as_u16(v_interleave_quads(v_reinterpret_as_s16(vec))); }
-
-inline v_int32x16 v_interleave_pairs(const v_int32x16& vec)
-{
-    return v_int32x16(_mm512_shuffle_epi32(vec.val, _MM_PERM_ACBD));
-}
-inline v_uint32x16 v_interleave_pairs(const v_uint32x16& vec) { return v_reinterpret_as_u32(v_interleave_pairs(v_reinterpret_as_s32(vec))); }
-inline v_float32x16 v_interleave_pairs(const v_float32x16& vec) { return v_reinterpret_as_f32(v_interleave_pairs(v_reinterpret_as_s32(vec))); }
-
-inline v_int8x64 v_pack_triplets(const v_int8x64& vec)
-{
-    return v_int8x64(_mm512_permutexvar_epi32(_v512_set_epu64(0x0000000f0000000f, 0x0000000f0000000f, 0x0000000e0000000d, 0x0000000c0000000a,
-                                                              0x0000000900000008, 0x0000000600000005, 0x0000000400000002, 0x0000000100000000),
-                                              _mm512_shuffle_epi8(vec.val, _mm512_set4_epi32(0xffffff0f, 0x0e0d0c0a, 0x09080605, 0x04020100))));
-}
-inline v_uint8x64 v_pack_triplets(const v_uint8x64& vec) { return v_reinterpret_as_u8(v_pack_triplets(v_reinterpret_as_s8(vec))); }
-
-inline v_int16x32 v_pack_triplets(const v_int16x32& vec)
-{
-    return v_int16x32(_mm512_permutexvar_epi16(_v512_set_epu64(0x001f001f001f001f, 0x001f001f001f001f, 0x001e001d001c001a, 0x0019001800160015,
-                                                               0x0014001200110010, 0x000e000d000c000a, 0x0009000800060005, 0x0004000200010000), vec.val));
-}
-inline v_uint16x32 v_pack_triplets(const v_uint16x32& vec) { return v_reinterpret_as_u16(v_pack_triplets(v_reinterpret_as_s16(vec))); }
-
-inline v_int32x16 v_pack_triplets(const v_int32x16& vec)
-{
-    return v_int32x16(_mm512_permutexvar_epi32(_v512_set_epu64(0x0000000f0000000f, 0x0000000f0000000f, 0x0000000e0000000d, 0x0000000c0000000a,
-                                                               0x0000000900000008, 0x0000000600000005, 0x0000000400000002, 0x0000000100000000), vec.val));
-}
-inline v_uint32x16 v_pack_triplets(const v_uint32x16& vec) { return v_reinterpret_as_u32(v_pack_triplets(v_reinterpret_as_s32(vec))); }
-inline v_float32x16 v_pack_triplets(const v_float32x16& vec)
-{
-    return v_float32x16(_mm512_permutexvar_ps(_v512_set_epu64(0x0000000f0000000f, 0x0000000f0000000f, 0x0000000e0000000d, 0x0000000c0000000a,
-                                                              0x0000000900000008, 0x0000000600000005, 0x0000000400000002, 0x0000000100000000), vec.val));
-}
-
-////////// Matrix operations /////////
-
-//////// Dot Product ////////
-
-// 16 >> 32
-inline v_int32x16 v_dotprod(const v_int16x32& a, const v_int16x32& b)
-{ return v_int32x16(_mm512_madd_epi16(a.val, b.val)); }
-inline v_int32x16 v_dotprod(const v_int16x32& a, const v_int16x32& b, const v_int32x16& c)
-{ return v_dotprod(a, b) + c; }
-
-// 32 >> 64
-inline v_int64x8 v_dotprod(const v_int32x16& a, const v_int32x16& b)
-{
-    __m512i even = _mm512_mul_epi32(a.val, b.val);
-    __m512i odd = _mm512_mul_epi32(_mm512_srli_epi64(a.val, 32), _mm512_srli_epi64(b.val, 32));
-    return v_int64x8(_mm512_add_epi64(even, odd));
-}
-inline v_int64x8 v_dotprod(const v_int32x16& a, const v_int32x16& b, const v_int64x8& c)
-{ return v_dotprod(a, b) + c; }
-
-// 8 >> 32
-inline v_uint32x16 v_dotprod_expand(const v_uint8x64& a, const v_uint8x64& b)
-{
-    __m512i even_a = _mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, a.val, _mm512_setzero_si512());
-    __m512i odd_a  = _mm512_srli_epi16(a.val, 8);
-
-    __m512i even_b = _mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, b.val, _mm512_setzero_si512());
-    __m512i odd_b  = _mm512_srli_epi16(b.val, 8);
-
-    __m512i prod0  = _mm512_madd_epi16(even_a, even_b);
-    __m512i prod1  = _mm512_madd_epi16(odd_a, odd_b);
-    return v_uint32x16(_mm512_add_epi32(prod0, prod1));
-}
-inline v_uint32x16 v_dotprod_expand(const v_uint8x64& a, const v_uint8x64& b, const v_uint32x16& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-inline v_int32x16 v_dotprod_expand(const v_int8x64& a, const v_int8x64& b)
-{
-    __m512i even_a = _mm512_srai_epi16(_mm512_bslli_epi128(a.val, 1), 8);
-    __m512i odd_a  = _mm512_srai_epi16(a.val, 8);
-
-    __m512i even_b = _mm512_srai_epi16(_mm512_bslli_epi128(b.val, 1), 8);
-    __m512i odd_b  = _mm512_srai_epi16(b.val, 8);
-
-    __m512i prod0  = _mm512_madd_epi16(even_a, even_b);
-    __m512i prod1  = _mm512_madd_epi16(odd_a, odd_b);
-    return v_int32x16(_mm512_add_epi32(prod0, prod1));
-}
-inline v_int32x16 v_dotprod_expand(const v_int8x64& a, const v_int8x64& b, const v_int32x16& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-// 16 >> 64
-inline v_uint64x8 v_dotprod_expand(const v_uint16x32& a, const v_uint16x32& b)
-{
-    __m512i mullo = _mm512_mullo_epi16(a.val, b.val);
-    __m512i mulhi = _mm512_mulhi_epu16(a.val, b.val);
-    __m512i mul0  = _mm512_unpacklo_epi16(mullo, mulhi);
-    __m512i mul1  = _mm512_unpackhi_epi16(mullo, mulhi);
-
-    __m512i p02   = _mm512_mask_blend_epi32(0xAAAA, mul0, _mm512_setzero_si512());
-    __m512i p13   = _mm512_srli_epi64(mul0, 32);
-    __m512i p46   = _mm512_mask_blend_epi32(0xAAAA, mul1, _mm512_setzero_si512());
-    __m512i p57   = _mm512_srli_epi64(mul1, 32);
-
-    __m512i p15_  = _mm512_add_epi64(p02, p13);
-    __m512i p9d_  = _mm512_add_epi64(p46, p57);
-
-    return v_uint64x8(_mm512_add_epi64(
-        _mm512_unpacklo_epi64(p15_, p9d_),
-        _mm512_unpackhi_epi64(p15_, p9d_)
-    ));
-}
-inline v_uint64x8 v_dotprod_expand(const v_uint16x32& a, const v_uint16x32& b, const v_uint64x8& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-inline v_int64x8 v_dotprod_expand(const v_int16x32& a, const v_int16x32& b)
-{
-    __m512i prod = _mm512_madd_epi16(a.val, b.val);
-    __m512i even = _mm512_srai_epi64(_mm512_bslli_epi128(prod, 4), 32);
-    __m512i odd  = _mm512_srai_epi64(prod, 32);
-    return v_int64x8(_mm512_add_epi64(even, odd));
-}
-inline v_int64x8 v_dotprod_expand(const v_int16x32& a, const v_int16x32& b, const v_int64x8& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-// 32 >> 64f
-inline v_float64x8 v_dotprod_expand(const v_int32x16& a, const v_int32x16& b)
-{ return v_cvt_f64(v_dotprod(a, b)); }
-inline v_float64x8 v_dotprod_expand(const v_int32x16& a, const v_int32x16& b, const v_float64x8& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-//////// Fast Dot Product ////////
-
-// 16 >> 32
-inline v_int32x16 v_dotprod_fast(const v_int16x32& a, const v_int16x32& b)
-{ return v_dotprod(a, b); }
-inline v_int32x16 v_dotprod_fast(const v_int16x32& a, const v_int16x32& b, const v_int32x16& c)
-{ return v_dotprod(a, b, c); }
-
-// 32 >> 64
-inline v_int64x8 v_dotprod_fast(const v_int32x16& a, const v_int32x16& b)
-{ return v_dotprod(a, b); }
-inline v_int64x8 v_dotprod_fast(const v_int32x16& a, const v_int32x16& b, const v_int64x8& c)
-{ return v_dotprod(a, b, c); }
-
-// 8 >> 32
-inline v_uint32x16 v_dotprod_expand_fast(const v_uint8x64& a, const v_uint8x64& b)
-{ return v_dotprod_expand(a, b); }
-inline v_uint32x16 v_dotprod_expand_fast(const v_uint8x64& a, const v_uint8x64& b, const v_uint32x16& c)
-{ return v_dotprod_expand(a, b, c); }
-
-inline v_int32x16 v_dotprod_expand_fast(const v_int8x64& a, const v_int8x64& b)
-{ return v_dotprod_expand(a, b); }
-inline v_int32x16 v_dotprod_expand_fast(const v_int8x64& a, const v_int8x64& b, const v_int32x16& c)
-{ return v_dotprod_expand(a, b, c); }
-
-// 16 >> 64
-inline v_uint64x8 v_dotprod_expand_fast(const v_uint16x32& a, const v_uint16x32& b)
-{
-    __m512i mullo = _mm512_mullo_epi16(a.val, b.val);
-    __m512i mulhi = _mm512_mulhi_epu16(a.val, b.val);
-    __m512i mul0  = _mm512_unpacklo_epi16(mullo, mulhi);
-    __m512i mul1  = _mm512_unpackhi_epi16(mullo, mulhi);
-
-    __m512i p02   = _mm512_mask_blend_epi32(0xAAAA, mul0, _mm512_setzero_si512());
-    __m512i p13   = _mm512_srli_epi64(mul0, 32);
-    __m512i p46   = _mm512_mask_blend_epi32(0xAAAA, mul1, _mm512_setzero_si512());
-    __m512i p57   = _mm512_srli_epi64(mul1, 32);
-
-    __m512i p15_  = _mm512_add_epi64(p02, p13);
-    __m512i p9d_  = _mm512_add_epi64(p46, p57);
-    return v_uint64x8(_mm512_add_epi64(p15_, p9d_));
-}
-inline v_uint64x8 v_dotprod_expand_fast(const v_uint16x32& a, const v_uint16x32& b, const v_uint64x8& c)
-{ return v_dotprod_expand_fast(a, b) + c; }
-
-inline v_int64x8 v_dotprod_expand_fast(const v_int16x32& a, const v_int16x32& b)
-{ return v_dotprod_expand(a, b); }
-inline v_int64x8 v_dotprod_expand_fast(const v_int16x32& a, const v_int16x32& b, const v_int64x8& c)
-{ return v_dotprod_expand(a, b, c); }
-
-// 32 >> 64f
-inline v_float64x8 v_dotprod_expand_fast(const v_int32x16& a, const v_int32x16& b)
-{ return v_dotprod_expand(a, b); }
-inline v_float64x8 v_dotprod_expand_fast(const v_int32x16& a, const v_int32x16& b, const v_float64x8& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-
-#define OPENCV_HAL_AVX512_SPLAT2_PS(a, im) \
-    v_float32x16(_mm512_permute_ps(a.val, _MM_SHUFFLE(im, im, im, im)))
-
-inline v_float32x16 v_matmul(const v_float32x16& v,
-                             const v_float32x16& m0, const v_float32x16& m1,
-                             const v_float32x16& m2, const v_float32x16& m3)
-{
-    v_float32x16 v04 = OPENCV_HAL_AVX512_SPLAT2_PS(v, 0);
-    v_float32x16 v15 = OPENCV_HAL_AVX512_SPLAT2_PS(v, 1);
-    v_float32x16 v26 = OPENCV_HAL_AVX512_SPLAT2_PS(v, 2);
-    v_float32x16 v37 = OPENCV_HAL_AVX512_SPLAT2_PS(v, 3);
-    return v_fma(v04, m0, v_fma(v15, m1, v_fma(v26, m2, v37 * m3)));
-}
-
-inline v_float32x16 v_matmuladd(const v_float32x16& v,
-                                const v_float32x16& m0, const v_float32x16& m1,
-                                const v_float32x16& m2, const v_float32x16& a)
-{
-    v_float32x16 v04 = OPENCV_HAL_AVX512_SPLAT2_PS(v, 0);
-    v_float32x16 v15 = OPENCV_HAL_AVX512_SPLAT2_PS(v, 1);
-    v_float32x16 v26 = OPENCV_HAL_AVX512_SPLAT2_PS(v, 2);
-    return v_fma(v04, m0, v_fma(v15, m1, v_fma(v26, m2, a)));
-}
-
-#define OPENCV_HAL_IMPL_AVX512_TRANSPOSE4x4(_Tpvec, suffix, cast_from, cast_to) \
-    inline void v_transpose4x4(const _Tpvec& a0, const _Tpvec& a1,              \
-                               const _Tpvec& a2, const _Tpvec& a3,              \
-                               _Tpvec& b0, _Tpvec& b1, _Tpvec& b2, _Tpvec& b3)  \
-    {                                                                           \
-        __m512i t0 = cast_from(_mm512_unpacklo_##suffix(a0.val, a1.val));       \
-        __m512i t1 = cast_from(_mm512_unpacklo_##suffix(a2.val, a3.val));       \
-        __m512i t2 = cast_from(_mm512_unpackhi_##suffix(a0.val, a1.val));       \
-        __m512i t3 = cast_from(_mm512_unpackhi_##suffix(a2.val, a3.val));       \
-        b0.val = cast_to(_mm512_unpacklo_epi64(t0, t1));                        \
-        b1.val = cast_to(_mm512_unpackhi_epi64(t0, t1));                        \
-        b2.val = cast_to(_mm512_unpacklo_epi64(t2, t3));                        \
-        b3.val = cast_to(_mm512_unpackhi_epi64(t2, t3));                        \
-    }
-
-OPENCV_HAL_IMPL_AVX512_TRANSPOSE4x4(v_uint32x16,  epi32, OPENCV_HAL_NOP, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_AVX512_TRANSPOSE4x4(v_int32x16,   epi32, OPENCV_HAL_NOP, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_AVX512_TRANSPOSE4x4(v_float32x16, ps, _mm512_castps_si512, _mm512_castsi512_ps)
-
-//////////////// Value reordering ///////////////
-
-/* Expand */
-#define OPENCV_HAL_IMPL_AVX512_EXPAND(_Tpvec, _Tpwvec, _Tp, intrin) \
-    inline void v_expand(const _Tpvec& a, _Tpwvec& b0, _Tpwvec& b1) \
-    {                                                               \
-        b0.val = intrin(_v512_extract_low(a.val));                  \
-        b1.val = intrin(_v512_extract_high(a.val));                 \
-    }                                                               \
-    inline _Tpwvec v_expand_low(const _Tpvec& a)                    \
-    { return _Tpwvec(intrin(_v512_extract_low(a.val))); }           \
-    inline _Tpwvec v_expand_high(const _Tpvec& a)                   \
-    { return _Tpwvec(intrin(_v512_extract_high(a.val))); }          \
-    inline _Tpwvec v512_load_expand(const _Tp* ptr)                 \
-    {                                                               \
-        __m256i a = _mm256_loadu_si256((const __m256i*)ptr);        \
-        return _Tpwvec(intrin(a));                                  \
-    }
-
-OPENCV_HAL_IMPL_AVX512_EXPAND(v_uint8x64,  v_uint16x32, uchar,    _mm512_cvtepu8_epi16)
-OPENCV_HAL_IMPL_AVX512_EXPAND(v_int8x64,   v_int16x32,  schar,    _mm512_cvtepi8_epi16)
-OPENCV_HAL_IMPL_AVX512_EXPAND(v_uint16x32, v_uint32x16, ushort,   _mm512_cvtepu16_epi32)
-OPENCV_HAL_IMPL_AVX512_EXPAND(v_int16x32,  v_int32x16,  short,    _mm512_cvtepi16_epi32)
-OPENCV_HAL_IMPL_AVX512_EXPAND(v_uint32x16, v_uint64x8,  unsigned, _mm512_cvtepu32_epi64)
-OPENCV_HAL_IMPL_AVX512_EXPAND(v_int32x16,  v_int64x8,   int,      _mm512_cvtepi32_epi64)
-
-#define OPENCV_HAL_IMPL_AVX512_EXPAND_Q(_Tpvec, _Tp, intrin) \
-    inline _Tpvec v512_load_expand_q(const _Tp* ptr)         \
-    {                                                        \
-        __m128i a = _mm_loadu_si128((const __m128i*)ptr);    \
-        return _Tpvec(intrin(a));                            \
-    }
-
-OPENCV_HAL_IMPL_AVX512_EXPAND_Q(v_uint32x16, uchar, _mm512_cvtepu8_epi32)
-OPENCV_HAL_IMPL_AVX512_EXPAND_Q(v_int32x16,  schar, _mm512_cvtepi8_epi32)
-
-/* pack */
-// 16
-inline v_int8x64 v_pack(const v_int16x32& a, const v_int16x32& b)
-{ return v_int8x64(_mm512_permutexvar_epi64(_v512_set_epu64(7, 5, 3, 1, 6, 4, 2, 0), _mm512_packs_epi16(a.val, b.val))); }
-
-inline v_uint8x64 v_pack(const v_uint16x32& a, const v_uint16x32& b)
-{
-    const __m512i t = _mm512_set1_epi16(255);
-    return v_uint8x64(_v512_combine(_mm512_cvtepi16_epi8(_mm512_min_epu16(a.val, t)), _mm512_cvtepi16_epi8(_mm512_min_epu16(b.val, t))));
-}
-
-inline v_uint8x64 v_pack_u(const v_int16x32& a, const v_int16x32& b)
-{
-    return v_uint8x64(_mm512_permutexvar_epi64(_v512_set_epu64(7, 5, 3, 1, 6, 4, 2, 0), _mm512_packus_epi16(a.val, b.val)));
-}
-
-inline void v_pack_store(schar* ptr, const v_int16x32& a)
-{ v_store_low(ptr, v_pack(a, a)); }
-
-inline void v_pack_store(uchar* ptr, const v_uint16x32& a)
-{
-    const __m512i m = _mm512_set1_epi16(255);
-    _mm256_storeu_si256((__m256i*)ptr, _mm512_cvtepi16_epi8(_mm512_min_epu16(a.val, m)));
-}
-
-inline void v_pack_u_store(uchar* ptr, const v_int16x32& a)
-{ v_store_low(ptr, v_pack_u(a, a)); }
-
-template<int n> inline
-v_uint8x64 v_rshr_pack(const v_uint16x32& a, const v_uint16x32& b)
-{
-    // we assume that n > 0, and so the shifted 16-bit values can be treated as signed numbers.
-    v_uint16x32 delta = v512_setall_u16((short)(1 << (n-1)));
-    return v_pack_u(v_reinterpret_as_s16((a + delta) >> n),
-                    v_reinterpret_as_s16((b + delta) >> n));
-}
-
-template<int n> inline
-void v_rshr_pack_store(uchar* ptr, const v_uint16x32& a)
-{
-    v_uint16x32 delta = v512_setall_u16((short)(1 << (n-1)));
-    v_pack_u_store(ptr, v_reinterpret_as_s16((a + delta) >> n));
-}
-
-template<int n> inline
-v_uint8x64 v_rshr_pack_u(const v_int16x32& a, const v_int16x32& b)
-{
-    v_int16x32 delta = v512_setall_s16((short)(1 << (n-1)));
-    return v_pack_u((a + delta) >> n, (b + delta) >> n);
-}
-
-template<int n> inline
-void v_rshr_pack_u_store(uchar* ptr, const v_int16x32& a)
-{
-    v_int16x32 delta = v512_setall_s16((short)(1 << (n-1)));
-    v_pack_u_store(ptr, (a + delta) >> n);
-}
-
-template<int n> inline
-v_int8x64 v_rshr_pack(const v_int16x32& a, const v_int16x32& b)
-{
-    v_int16x32 delta = v512_setall_s16((short)(1 << (n-1)));
-    return v_pack((a + delta) >> n, (b + delta) >> n);
-}
-
-template<int n> inline
-void v_rshr_pack_store(schar* ptr, const v_int16x32& a)
-{
-    v_int16x32 delta = v512_setall_s16((short)(1 << (n-1)));
-    v_pack_store(ptr, (a + delta) >> n);
-}
-
-// 32
-inline v_int16x32 v_pack(const v_int32x16& a, const v_int32x16& b)
-{ return v_int16x32(_mm512_permutexvar_epi64(_v512_set_epu64(7, 5, 3, 1, 6, 4, 2, 0), _mm512_packs_epi32(a.val, b.val))); }
-
-inline v_uint16x32 v_pack(const v_uint32x16& a, const v_uint32x16& b)
-{
-    const __m512i m = _mm512_set1_epi32(65535);
-    return v_uint16x32(_v512_combine(_mm512_cvtepi32_epi16(_mm512_min_epu32(a.val, m)), _mm512_cvtepi32_epi16(_mm512_min_epu32(b.val, m))));
-}
-
-inline v_uint16x32 v_pack_u(const v_int32x16& a, const v_int32x16& b)
-{ return v_uint16x32(_mm512_permutexvar_epi64(_v512_set_epu64(7, 5, 3, 1, 6, 4, 2, 0), _mm512_packus_epi32(a.val, b.val))); }
-
-inline void v_pack_store(short* ptr, const v_int32x16& a)
-{ v_store_low(ptr, v_pack(a, a)); }
-
-inline void v_pack_store(ushort* ptr, const v_uint32x16& a)
-{
-    const __m512i m = _mm512_set1_epi32(65535);
-    _mm256_storeu_si256((__m256i*)ptr, _mm512_cvtepi32_epi16(_mm512_min_epu32(a.val, m)));
-}
-
-inline void v_pack_u_store(ushort* ptr, const v_int32x16& a)
-{ v_store_low(ptr, v_pack_u(a, a)); }
-
-
-template<int n> inline
-v_uint16x32 v_rshr_pack(const v_uint32x16& a, const v_uint32x16& b)
-{
-    v_uint32x16 delta = v512_setall_u32(1 << (n-1));
-    return v_pack_u(v_reinterpret_as_s32((a + delta) >> n),
-                    v_reinterpret_as_s32((b + delta) >> n));
-}
-
-template<int n> inline
-void v_rshr_pack_store(ushort* ptr, const v_uint32x16& a)
-{
-    v_uint32x16 delta = v512_setall_u32(1 << (n-1));
-    v_pack_u_store(ptr, v_reinterpret_as_s32((a + delta) >> n));
-}
-
-template<int n> inline
-v_uint16x32 v_rshr_pack_u(const v_int32x16& a, const v_int32x16& b)
-{
-    v_int32x16 delta = v512_setall_s32(1 << (n-1));
-    return v_pack_u((a + delta) >> n, (b + delta) >> n);
-}
-
-template<int n> inline
-void v_rshr_pack_u_store(ushort* ptr, const v_int32x16& a)
-{
-    v_int32x16 delta = v512_setall_s32(1 << (n-1));
-    v_pack_u_store(ptr, (a + delta) >> n);
-}
-
-template<int n> inline
-v_int16x32 v_rshr_pack(const v_int32x16& a, const v_int32x16& b)
-{
-    v_int32x16 delta = v512_setall_s32(1 << (n-1));
-    return v_pack((a + delta) >> n, (b + delta) >> n);
-}
-
-template<int n> inline
-void v_rshr_pack_store(short* ptr, const v_int32x16& a)
-{
-    v_int32x16 delta = v512_setall_s32(1 << (n-1));
-    v_pack_store(ptr, (a + delta) >> n);
-}
-
-// 64
-// Non-saturating pack
-inline v_uint32x16 v_pack(const v_uint64x8& a, const v_uint64x8& b)
-{ return v_uint32x16(_v512_combine(_mm512_cvtepi64_epi32(a.val), _mm512_cvtepi64_epi32(b.val))); }
-
-inline v_int32x16 v_pack(const v_int64x8& a, const v_int64x8& b)
-{ return v_reinterpret_as_s32(v_pack(v_reinterpret_as_u64(a), v_reinterpret_as_u64(b))); }
-
-inline void v_pack_store(unsigned* ptr, const v_uint64x8& a)
-{ _mm256_storeu_si256((__m256i*)ptr, _mm512_cvtepi64_epi32(a.val)); }
-
-inline void v_pack_store(int* ptr, const v_int64x8& b)
-{ v_pack_store((unsigned*)ptr, v_reinterpret_as_u64(b)); }
-
-template<int n> inline
-v_uint32x16 v_rshr_pack(const v_uint64x8& a, const v_uint64x8& b)
-{
-    v_uint64x8 delta = v512_setall_u64((uint64)1 << (n-1));
-    return v_pack((a + delta) >> n, (b + delta) >> n);
-}
-
-template<int n> inline
-void v_rshr_pack_store(unsigned* ptr, const v_uint64x8& a)
-{
-    v_uint64x8 delta = v512_setall_u64((uint64)1 << (n-1));
-    v_pack_store(ptr, (a + delta) >> n);
-}
-
-template<int n> inline
-v_int32x16 v_rshr_pack(const v_int64x8& a, const v_int64x8& b)
-{
-    v_int64x8 delta = v512_setall_s64((int64)1 << (n-1));
-    return v_pack((a + delta) >> n, (b + delta) >> n);
-}
-
-template<int n> inline
-void v_rshr_pack_store(int* ptr, const v_int64x8& a)
-{
-    v_int64x8 delta = v512_setall_s64((int64)1 << (n-1));
-    v_pack_store(ptr, (a + delta) >> n);
-}
-
-// pack boolean
-inline v_uint8x64 v_pack_b(const v_uint16x32& a, const v_uint16x32& b)
-{ return v_uint8x64(_mm512_permutexvar_epi64(_v512_set_epu64(7, 5, 3, 1, 6, 4, 2, 0), _mm512_packs_epi16(a.val, b.val))); }
-
-inline v_uint8x64 v_pack_b(const v_uint32x16& a, const v_uint32x16& b,
-                           const v_uint32x16& c, const v_uint32x16& d)
-{
-    __m512i ab = _mm512_packs_epi32(a.val, b.val);
-    __m512i cd = _mm512_packs_epi32(c.val, d.val);
-
-    return v_uint8x64(_mm512_permutexvar_epi32(_v512_set_epu32(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0), _mm512_packs_epi16(ab, cd)));
-}
-
-inline v_uint8x64 v_pack_b(const v_uint64x8& a, const v_uint64x8& b, const v_uint64x8& c,
-                           const v_uint64x8& d, const v_uint64x8& e, const v_uint64x8& f,
-                           const v_uint64x8& g, const v_uint64x8& h)
-{
-    __m512i ab = _mm512_packs_epi32(a.val, b.val);
-    __m512i cd = _mm512_packs_epi32(c.val, d.val);
-    __m512i ef = _mm512_packs_epi32(e.val, f.val);
-    __m512i gh = _mm512_packs_epi32(g.val, h.val);
-
-    __m512i abcd = _mm512_packs_epi32(ab, cd);
-    __m512i efgh = _mm512_packs_epi32(ef, gh);
-
-    return v_uint8x64(_mm512_permutexvar_epi16(_v512_set_epu16(31, 23, 15, 7, 30, 22, 14, 6, 29, 21, 13, 5, 28, 20, 12, 4,
-                                                               27, 19, 11, 3, 26, 18, 10, 2, 25, 17,  9, 1, 24, 16,  8, 0), _mm512_packs_epi16(abcd, efgh)));
-}
-
-/* Recombine */
-// its up there with load and store operations
-
-/* Extract */
-#define OPENCV_HAL_IMPL_AVX512_EXTRACT(_Tpvec)                \
-    template<int s>                                           \
-    inline _Tpvec v_extract(const _Tpvec& a, const _Tpvec& b) \
-    { return v_rotate_right<s>(a, b); }
-
-OPENCV_HAL_IMPL_AVX512_EXTRACT(v_uint8x64)
-OPENCV_HAL_IMPL_AVX512_EXTRACT(v_int8x64)
-OPENCV_HAL_IMPL_AVX512_EXTRACT(v_uint16x32)
-OPENCV_HAL_IMPL_AVX512_EXTRACT(v_int16x32)
-OPENCV_HAL_IMPL_AVX512_EXTRACT(v_uint32x16)
-OPENCV_HAL_IMPL_AVX512_EXTRACT(v_int32x16)
-OPENCV_HAL_IMPL_AVX512_EXTRACT(v_uint64x8)
-OPENCV_HAL_IMPL_AVX512_EXTRACT(v_int64x8)
-OPENCV_HAL_IMPL_AVX512_EXTRACT(v_float32x16)
-OPENCV_HAL_IMPL_AVX512_EXTRACT(v_float64x8)
-
-#define OPENCV_HAL_IMPL_AVX512_EXTRACT_N(_Tpvec, _Tp) \
-template<int i> inline _Tp v_extract_n(_Tpvec v) { return v_rotate_right<i>(v).get0(); }
-
-OPENCV_HAL_IMPL_AVX512_EXTRACT_N(v_uint8x64, uchar)
-OPENCV_HAL_IMPL_AVX512_EXTRACT_N(v_int8x64, schar)
-OPENCV_HAL_IMPL_AVX512_EXTRACT_N(v_uint16x32, ushort)
-OPENCV_HAL_IMPL_AVX512_EXTRACT_N(v_int16x32, short)
-OPENCV_HAL_IMPL_AVX512_EXTRACT_N(v_uint32x16, uint)
-OPENCV_HAL_IMPL_AVX512_EXTRACT_N(v_int32x16, int)
-OPENCV_HAL_IMPL_AVX512_EXTRACT_N(v_uint64x8, uint64)
-OPENCV_HAL_IMPL_AVX512_EXTRACT_N(v_int64x8, int64)
-OPENCV_HAL_IMPL_AVX512_EXTRACT_N(v_float32x16, float)
-OPENCV_HAL_IMPL_AVX512_EXTRACT_N(v_float64x8, double)
-
-template<int i>
-inline v_uint32x16 v_broadcast_element(v_uint32x16 a)
-{
-    static const __m512i perm = _mm512_set1_epi32((char)i);
-    return v_uint32x16(_mm512_permutexvar_epi32(perm, a.val));
-}
-
-template<int i>
-inline v_int32x16 v_broadcast_element(const v_int32x16 &a)
-{ return v_reinterpret_as_s32(v_broadcast_element<i>(v_reinterpret_as_u32(a))); }
-
-template<int i>
-inline v_float32x16 v_broadcast_element(const v_float32x16 &a)
-{ return v_reinterpret_as_f32(v_broadcast_element<i>(v_reinterpret_as_u32(a))); }
-
-
-///////////////////// load deinterleave /////////////////////////////
-
-inline void v_load_deinterleave( const uchar* ptr, v_uint8x64& a, v_uint8x64& b )
-{
-    __m512i ab0 = _mm512_loadu_si512((const __m512i*)ptr);
-    __m512i ab1 = _mm512_loadu_si512((const __m512i*)(ptr + 64));
-#if CV_AVX_512VBMI
-    __m512i mask0 = _v512_set_epu8(126, 124, 122, 120, 118, 116, 114, 112, 110, 108, 106, 104, 102, 100, 98, 96,
-                                    94,  92,  90,  88,  86,  84,  82,  80,  78,  76,  74,  72,  70,  68, 66, 64,
-                                    62,  60,  58,  56,  54,  52,  50,  48,  46,  44,  42,  40,  38,  36, 34, 32,
-                                    30,  28,  26,  24,  22,  20,  18,  16,  14,  12,  10,   8,   6,   4,  2,  0);
-    __m512i mask1 = _v512_set_epu8(127, 125, 123, 121, 119, 117, 115, 113, 111, 109, 107, 105, 103, 101, 99, 97,
-                                    95,  93,  91,  89,  87,  85,  83,  81,  79,  77,  75,  73,  71,  69, 67, 65,
-                                    63,  61,  59,  57,  55,  53,  51,  49,  47,  45,  43,  41,  39,  37, 35, 33,
-                                    31,  29,  27,  25,  23,  21,  19,  17,  15,  13,  11,   9,   7,   5,  3,  1);
-    a = v_uint8x64(_mm512_permutex2var_epi8(ab0, mask0, ab1));
-    b = v_uint8x64(_mm512_permutex2var_epi8(ab0, mask1, ab1));
-#else
-    __m512i mask0 = _mm512_set4_epi32(0x0f0d0b09, 0x07050301, 0x0e0c0a08, 0x06040200);
-    __m512i a0b0 = _mm512_shuffle_epi8(ab0, mask0);
-    __m512i a1b1 = _mm512_shuffle_epi8(ab1, mask0);
-    __m512i mask1 = _v512_set_epu64(14, 12, 10, 8, 6, 4, 2, 0);
-    __m512i mask2 = _v512_set_epu64(15, 13, 11, 9, 7, 5, 3, 1);
-    a = v_uint8x64(_mm512_permutex2var_epi64(a0b0, mask1, a1b1));
-    b = v_uint8x64(_mm512_permutex2var_epi64(a0b0, mask2, a1b1));
-#endif
-}
-
-inline void v_load_deinterleave( const ushort* ptr, v_uint16x32& a, v_uint16x32& b )
-{
-    __m512i ab0 = _mm512_loadu_si512((const __m512i*)ptr);
-    __m512i ab1 = _mm512_loadu_si512((const __m512i*)(ptr + 32));
-    __m512i mask0 = _v512_set_epu16(62, 60, 58, 56, 54, 52, 50, 48, 46, 44, 42, 40, 38, 36, 34, 32,
-                                    30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10,  8,  6,  4,  2,  0);
-    __m512i mask1 = _v512_set_epu16(63, 61, 59, 57, 55, 53, 51, 49, 47, 45, 43, 41, 39, 37, 35, 33,
-                                    31, 29, 27, 25, 23, 21, 19, 17, 15, 13, 11,  9,  7,  5,  3,  1);
-    a = v_uint16x32(_mm512_permutex2var_epi16(ab0, mask0, ab1));
-    b = v_uint16x32(_mm512_permutex2var_epi16(ab0, mask1, ab1));
-}
-
-inline void v_load_deinterleave( const unsigned* ptr, v_uint32x16& a, v_uint32x16& b )
-{
-    __m512i ab0 = _mm512_loadu_si512((const __m512i*)ptr);
-    __m512i ab1 = _mm512_loadu_si512((const __m512i*)(ptr + 16));
-    __m512i mask0 = _v512_set_epu32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
-    __m512i mask1 = _v512_set_epu32(31, 29, 27, 25, 23, 21, 19, 17, 15, 13, 11, 9, 7, 5, 3, 1);
-    a = v_uint32x16(_mm512_permutex2var_epi32(ab0, mask0, ab1));
-    b = v_uint32x16(_mm512_permutex2var_epi32(ab0, mask1, ab1));
-}
-
-inline void v_load_deinterleave( const uint64* ptr, v_uint64x8& a, v_uint64x8& b )
-{
-    __m512i ab0 = _mm512_loadu_si512((const __m512i*)ptr);
-    __m512i ab1 = _mm512_loadu_si512((const __m512i*)(ptr + 8));
-    __m512i mask0 = _v512_set_epu64(14, 12, 10, 8, 6, 4, 2, 0);
-    __m512i mask1 = _v512_set_epu64(15, 13, 11, 9, 7, 5, 3, 1);
-    a = v_uint64x8(_mm512_permutex2var_epi64(ab0, mask0, ab1));
-    b = v_uint64x8(_mm512_permutex2var_epi64(ab0, mask1, ab1));
-}
-
-inline void v_load_deinterleave( const uchar* ptr, v_uint8x64& a, v_uint8x64& b, v_uint8x64& c )
-{
-    __m512i bgr0 = _mm512_loadu_si512((const __m512i*)ptr);
-    __m512i bgr1 = _mm512_loadu_si512((const __m512i*)(ptr + 64));
-    __m512i bgr2 = _mm512_loadu_si512((const __m512i*)(ptr + 128));
-
-#if CV_AVX_512VBMI2
-    __m512i mask0 = _v512_set_epu8(126, 123, 120, 117, 114, 111, 108, 105, 102,  99,  96,  93,  90,  87,  84, 81,
-                                    78,  75,  72,  69,  66,  63,  60,  57,  54,  51,  48,  45,  42,  39,  36, 33,
-                                    30,  27,  24,  21,  18,  15,  12,   9,   6,   3,   0,  62,  59,  56,  53, 50,
-                                    47,  44,  41,  38,  35,  32,  29,  26,  23,  20,  17,  14,  11,   8,   5,  2);
-    __m512i r0b01 = _mm512_permutex2var_epi8(bgr0, mask0, bgr1);
-    __m512i b1g12 = _mm512_permutex2var_epi8(bgr1, mask0, bgr2);
-    __m512i r12b2 = _mm512_permutex2var_epi8(bgr1,
-                    _v512_set_epu8(125, 122, 119, 116, 113, 110, 107, 104, 101,  98,  95,  92,  89,  86,  83, 80,
-                                    77,  74,  71,  68,  65, 127, 124, 121, 118, 115, 112, 109, 106, 103, 100, 97,
-                                    94,  91,  88,  85,  82,  79,  76,  73,  70,  67,  64,  61,  58,  55,  52, 49,
-                                    46,  43,  40,  37,  34,  31,  28,  25,  22,  19,  16,  13,  10,   7,   4,  1), bgr2);
-    a = v_uint8x64(_mm512_mask_compress_epi8(r12b2, 0xffffffffffe00000, r0b01));
-    b = v_uint8x64(_mm512_mask_compress_epi8(b1g12, 0x2492492492492492, bgr0));
-    c = v_uint8x64(_mm512_mask_expand_epi8(r0b01, 0xffffffffffe00000, r12b2));
-#elif CV_AVX_512VBMI
-    __m512i b0g0b1 = _mm512_mask_blend_epi8(0xb6db6db6db6db6db, bgr1, bgr0);
-    __m512i g1r1g2 = _mm512_mask_blend_epi8(0xb6db6db6db6db6db, bgr2, bgr1);
-    __m512i r2b2r0 = _mm512_mask_blend_epi8(0xb6db6db6db6db6db, bgr0, bgr2);
-    a = v_uint8x64(_mm512_permutex2var_epi8(b0g0b1, _v512_set_epu8(125, 122, 119, 116, 113, 110, 107, 104, 101,  98,  95,  92,  89,  86,  83,  80,
-                                                                    77,  74,  71,  68,  65,  63,  61,  60,  58,  57,  55,  54,  52,  51,  49,  48,
-                                                                    46,  45,  43,  42,  40,  39,  37,  36,  34,  33,  31,  30,  28,  27,  25,  24,
-                                                                    23,  21,  20,  18,  17,  15,  14,  12,  11,   9,   8,   6,   5,   3,   2,   0), bgr2));
-    b = v_uint8x64(_mm512_permutex2var_epi8(g1r1g2, _v512_set_epu8( 63,  61,  60,  58,  57,  55,  54,  52,  51,  49,  48,  46,  45,  43,  42,  40,
-                                                                    39,  37,  36,  34,  33,  31,  30,  28,  27,  25,  24,  23,  21,  20,  18,  17,
-                                                                    15,  14,  12,  11,   9,   8,   6,   5,   3,   2,   0, 126, 123, 120, 117, 114,
-                                                                   111, 108, 105, 102,  99,  96,  93,  90,  87,  84,  81,  78,  75,  72,  69,  66), bgr0));
-    c = v_uint8x64(_mm512_permutex2var_epi8(r2b2r0, _v512_set_epu8( 63,  60,  57,  54,  51,  48,  45,  42,  39,  36,  33,  30,  27,  24,  21,  18,
-                                                                    15,  12,   9,   6,   3,   0, 125, 122, 119, 116, 113, 110, 107, 104, 101,  98,
-                                                                    95,  92,  89,  86,  83,  80,  77,  74,  71,  68,  65,  62,  59,  56,  53,  50,
-                                                                    47,  44,  41,  38,  35,  32,  29,  26,  23,  20,  17,  14,  11,   8,   5,   2), bgr1));
-#else
-    __m512i mask0 = _v512_set_epu16(61, 58, 55, 52, 49, 46, 43, 40, 37, 34, 63, 60, 57, 54, 51, 48,
-                                    45, 42, 39, 36, 33, 30, 27, 24, 21, 18, 15, 12,  9,  6,  3,  0);
-    __m512i b01g1 = _mm512_permutex2var_epi16(bgr0, mask0, bgr1);
-    __m512i r12b2 = _mm512_permutex2var_epi16(bgr1, mask0, bgr2);
-    __m512i g20r0 = _mm512_permutex2var_epi16(bgr2, mask0, bgr0);
-
-    __m512i b0g0 = _mm512_mask_blend_epi32(0xf800, b01g1, r12b2);
-    __m512i r0b1 = _mm512_permutex2var_epi16(bgr1, _v512_set_epu16(42, 41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 29, 26, 23, 20, 17,
-                                                                   14, 11,  8,  5,  2, 53, 52, 51, 50, 49, 48, 47, 46, 45, 44, 43), g20r0);
-    __m512i g1r1 = _mm512_alignr_epi32(r12b2, g20r0, 11);
-    a = v_uint8x64(_mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, b0g0, r0b1));
-    c = v_uint8x64(_mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, r0b1, g1r1));
-    b = v_uint8x64(_mm512_shuffle_epi8(_mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, g1r1, b0g0), _mm512_set4_epi32(0x0e0f0c0d, 0x0a0b0809, 0x06070405, 0x02030001)));
-#endif
-}
-
-inline void v_load_deinterleave( const ushort* ptr, v_uint16x32& a, v_uint16x32& b, v_uint16x32& c )
-{
-    __m512i bgr0 = _mm512_loadu_si512((const __m512i*)ptr);
-    __m512i bgr1 = _mm512_loadu_si512((const __m512i*)(ptr + 32));
-    __m512i bgr2 = _mm512_loadu_si512((const __m512i*)(ptr + 64));
-
-    __m512i mask0 = _v512_set_epu16(61, 58, 55, 52, 49, 46, 43, 40, 37, 34, 63, 60, 57, 54, 51, 48,
-                                    45, 42, 39, 36, 33, 30, 27, 24, 21, 18, 15, 12,  9,  6,  3,  0);
-    __m512i b01g1 = _mm512_permutex2var_epi16(bgr0, mask0, bgr1);
-    __m512i r12b2 = _mm512_permutex2var_epi16(bgr1, mask0, bgr2);
-    __m512i g20r0 = _mm512_permutex2var_epi16(bgr2, mask0, bgr0);
-
-    a = v_uint16x32(_mm512_mask_blend_epi32(0xf800, b01g1, r12b2));
-    b = v_uint16x32(_mm512_permutex2var_epi16(bgr1, _v512_set_epu16(42, 41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 29, 26, 23, 20, 17,
-                                                                    14, 11,  8,  5,  2, 53, 52, 51, 50, 49, 48, 47, 46, 45, 44, 43), g20r0));
-    c = v_uint16x32(_mm512_alignr_epi32(r12b2, g20r0, 11));
-}
-
-inline void v_load_deinterleave( const unsigned* ptr, v_uint32x16& a, v_uint32x16& b, v_uint32x16& c )
-{
-    __m512i bgr0 = _mm512_loadu_si512((const __m512i*)ptr);
-    __m512i bgr1 = _mm512_loadu_si512((const __m512i*)(ptr + 16));
-    __m512i bgr2 = _mm512_loadu_si512((const __m512i*)(ptr + 32));
-
-    __m512i mask0 = _v512_set_epu32(29, 26, 23, 20, 17, 30, 27, 24, 21, 18, 15, 12, 9, 6, 3, 0);
-    __m512i b01r1 = _mm512_permutex2var_epi32(bgr0, mask0, bgr1);
-    __m512i g12b2 = _mm512_permutex2var_epi32(bgr1, mask0, bgr2);
-    __m512i r20g0 = _mm512_permutex2var_epi32(bgr2, mask0, bgr0);
-
-    a = v_uint32x16(_mm512_mask_blend_epi32(0xf800, b01r1, g12b2));
-    b = v_uint32x16(_mm512_alignr_epi32(g12b2, r20g0, 11));
-    c = v_uint32x16(_mm512_permutex2var_epi32(bgr1, _v512_set_epu32(21, 20, 19, 18, 17, 16, 13, 10, 7, 4, 1, 26, 25, 24, 23, 22), r20g0));
-}
-
-inline void v_load_deinterleave( const uint64* ptr, v_uint64x8& a, v_uint64x8& b, v_uint64x8& c )
-{
-    __m512i bgr0 = _mm512_loadu_si512((const __m512i*)ptr);
-    __m512i bgr1 = _mm512_loadu_si512((const __m512i*)(ptr + 8));
-    __m512i bgr2 = _mm512_loadu_si512((const __m512i*)(ptr + 16));
-
-    __m512i mask0 = _v512_set_epu64(13, 10, 15, 12, 9, 6, 3, 0);
-    __m512i b01g1 = _mm512_permutex2var_epi64(bgr0, mask0, bgr1);
-    __m512i r12b2 = _mm512_permutex2var_epi64(bgr1, mask0, bgr2);
-    __m512i g20r0 = _mm512_permutex2var_epi64(bgr2, mask0, bgr0);
-
-    a = v_uint64x8(_mm512_mask_blend_epi64(0xc0, b01g1, r12b2));
-    c = v_uint64x8(_mm512_alignr_epi64(r12b2, g20r0, 6));
-    b = v_uint64x8(_mm512_permutex2var_epi64(bgr1, _v512_set_epu64(10, 9, 8, 5, 2, 13, 12, 11), g20r0));
-}
-
-inline void v_load_deinterleave( const uchar* ptr, v_uint8x64& a, v_uint8x64& b, v_uint8x64& c, v_uint8x64& d )
-{
-    __m512i bgra0 = _mm512_loadu_si512((const __m512i*)ptr);
-    __m512i bgra1 = _mm512_loadu_si512((const __m512i*)(ptr + 64));
-    __m512i bgra2 = _mm512_loadu_si512((const __m512i*)(ptr + 128));
-    __m512i bgra3 = _mm512_loadu_si512((const __m512i*)(ptr + 192));
-
-#if CV_AVX_512VBMI
-    __m512i mask0 = _v512_set_epu8(126, 124, 122, 120, 118, 116, 114, 112, 110, 108, 106, 104, 102, 100, 98, 96,
-                                    94,  92,  90,  88,  86,  84,  82,  80,  78,  76,  74,  72,  70,  68, 66, 64,
-                                    62,  60,  58,  56,  54,  52,  50,  48,  46,  44,  42,  40,  38,  36, 34, 32,
-                                    30,  28,  26,  24,  22,  20,  18,  16,  14,  12,  10,   8,   6,   4,  2,  0);
-    __m512i mask1 = _v512_set_epu8(127, 125, 123, 121, 119, 117, 115, 113, 111, 109, 107, 105, 103, 101, 99, 97,
-                                    95,  93,  91,  89,  87,  85,  83,  81,  79,  77,  75,  73,  71,  69, 67, 65,
-                                    63,  61,  59,  57,  55,  53,  51,  49,  47,  45,  43,  41,  39,  37, 35, 33,
-                                    31,  29,  27,  25,  23,  21,  19,  17,  15,  13,  11,   9,   7,   5,  3,  1);
-
-    __m512i br01 = _mm512_permutex2var_epi8(bgra0, mask0, bgra1);
-    __m512i ga01 = _mm512_permutex2var_epi8(bgra0, mask1, bgra1);
-    __m512i br23 = _mm512_permutex2var_epi8(bgra2, mask0, bgra3);
-    __m512i ga23 = _mm512_permutex2var_epi8(bgra2, mask1, bgra3);
-
-    a = v_uint8x64(_mm512_permutex2var_epi8(br01, mask0, br23));
-    c = v_uint8x64(_mm512_permutex2var_epi8(br01, mask1, br23));
-    b = v_uint8x64(_mm512_permutex2var_epi8(ga01, mask0, ga23));
-    d = v_uint8x64(_mm512_permutex2var_epi8(ga01, mask1, ga23));
-#else
-    __m512i mask = _mm512_set4_epi32(0x0f0b0703, 0x0e0a0602, 0x0d090501, 0x0c080400);
-    __m512i b0g0r0a0 = _mm512_shuffle_epi8(bgra0, mask);
-    __m512i b1g1r1a1 = _mm512_shuffle_epi8(bgra1, mask);
-    __m512i b2g2r2a2 = _mm512_shuffle_epi8(bgra2, mask);
-    __m512i b3g3r3a3 = _mm512_shuffle_epi8(bgra3, mask);
-
-    __m512i mask0 = _v512_set_epu32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
-    __m512i mask1 = _v512_set_epu32(31, 29, 27, 25, 23, 21, 19, 17, 15, 13, 11, 9, 7, 5, 3, 1);
-
-    __m512i br01 = _mm512_permutex2var_epi32(b0g0r0a0, mask0, b1g1r1a1);
-    __m512i ga01 = _mm512_permutex2var_epi32(b0g0r0a0, mask1, b1g1r1a1);
-    __m512i br23 = _mm512_permutex2var_epi32(b2g2r2a2, mask0, b3g3r3a3);
-    __m512i ga23 = _mm512_permutex2var_epi32(b2g2r2a2, mask1, b3g3r3a3);
-
-    a = v_uint8x64(_mm512_permutex2var_epi32(br01, mask0, br23));
-    c = v_uint8x64(_mm512_permutex2var_epi32(br01, mask1, br23));
-    b = v_uint8x64(_mm512_permutex2var_epi32(ga01, mask0, ga23));
-    d = v_uint8x64(_mm512_permutex2var_epi32(ga01, mask1, ga23));
-#endif
-}
-
-inline void v_load_deinterleave( const ushort* ptr, v_uint16x32& a, v_uint16x32& b, v_uint16x32& c, v_uint16x32& d )
-{
-    __m512i bgra0 = _mm512_loadu_si512((const __m512i*)ptr);
-    __m512i bgra1 = _mm512_loadu_si512((const __m512i*)(ptr + 32));
-    __m512i bgra2 = _mm512_loadu_si512((const __m512i*)(ptr + 64));
-    __m512i bgra3 = _mm512_loadu_si512((const __m512i*)(ptr + 96));
-
-    __m512i mask0 = _v512_set_epu16(62, 60, 58, 56, 54, 52, 50, 48, 46, 44, 42, 40, 38, 36, 34, 32,
-                                    30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10,  8,  6,  4,  2,  0);
-    __m512i mask1 = _v512_set_epu16(63, 61, 59, 57, 55, 53, 51, 49, 47, 45, 43, 41, 39, 37, 35, 33,
-                                    31, 29, 27, 25, 23, 21, 19, 17, 15, 13, 11,  9,  7,  5,  3,  1);
-
-    __m512i br01 = _mm512_permutex2var_epi16(bgra0, mask0, bgra1);
-    __m512i ga01 = _mm512_permutex2var_epi16(bgra0, mask1, bgra1);
-    __m512i br23 = _mm512_permutex2var_epi16(bgra2, mask0, bgra3);
-    __m512i ga23 = _mm512_permutex2var_epi16(bgra2, mask1, bgra3);
-
-    a = v_uint16x32(_mm512_permutex2var_epi16(br01, mask0, br23));
-    c = v_uint16x32(_mm512_permutex2var_epi16(br01, mask1, br23));
-    b = v_uint16x32(_mm512_permutex2var_epi16(ga01, mask0, ga23));
-    d = v_uint16x32(_mm512_permutex2var_epi16(ga01, mask1, ga23));
-}
-
-inline void v_load_deinterleave( const unsigned* ptr, v_uint32x16& a, v_uint32x16& b, v_uint32x16& c, v_uint32x16& d )
-{
-    __m512i bgra0 = _mm512_loadu_si512((const __m512i*)ptr);
-    __m512i bgra1 = _mm512_loadu_si512((const __m512i*)(ptr + 16));
-    __m512i bgra2 = _mm512_loadu_si512((const __m512i*)(ptr + 32));
-    __m512i bgra3 = _mm512_loadu_si512((const __m512i*)(ptr + 48));
-
-    __m512i mask0 = _v512_set_epu32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
-    __m512i mask1 = _v512_set_epu32(31, 29, 27, 25, 23, 21, 19, 17, 15, 13, 11, 9, 7, 5, 3, 1);
-
-    __m512i br01 = _mm512_permutex2var_epi32(bgra0, mask0, bgra1);
-    __m512i ga01 = _mm512_permutex2var_epi32(bgra0, mask1, bgra1);
-    __m512i br23 = _mm512_permutex2var_epi32(bgra2, mask0, bgra3);
-    __m512i ga23 = _mm512_permutex2var_epi32(bgra2, mask1, bgra3);
-
-    a = v_uint32x16(_mm512_permutex2var_epi32(br01, mask0, br23));
-    c = v_uint32x16(_mm512_permutex2var_epi32(br01, mask1, br23));
-    b = v_uint32x16(_mm512_permutex2var_epi32(ga01, mask0, ga23));
-    d = v_uint32x16(_mm512_permutex2var_epi32(ga01, mask1, ga23));
-}
-
-inline void v_load_deinterleave( const uint64* ptr, v_uint64x8& a, v_uint64x8& b, v_uint64x8& c, v_uint64x8& d )
-{
-    __m512i bgra0 = _mm512_loadu_si512((const __m512i*)ptr);
-    __m512i bgra1 = _mm512_loadu_si512((const __m512i*)(ptr + 8));
-    __m512i bgra2 = _mm512_loadu_si512((const __m512i*)(ptr + 16));
-    __m512i bgra3 = _mm512_loadu_si512((const __m512i*)(ptr + 24));
-
-    __m512i mask0 = _v512_set_epu64(14, 12, 10, 8, 6, 4, 2, 0);
-    __m512i mask1 = _v512_set_epu64(15, 13, 11, 9, 7, 5, 3, 1);
-
-    __m512i br01 = _mm512_permutex2var_epi64(bgra0, mask0, bgra1);
-    __m512i ga01 = _mm512_permutex2var_epi64(bgra0, mask1, bgra1);
-    __m512i br23 = _mm512_permutex2var_epi64(bgra2, mask0, bgra3);
-    __m512i ga23 = _mm512_permutex2var_epi64(bgra2, mask1, bgra3);
-
-    a = v_uint64x8(_mm512_permutex2var_epi64(br01, mask0, br23));
-    c = v_uint64x8(_mm512_permutex2var_epi64(br01, mask1, br23));
-    b = v_uint64x8(_mm512_permutex2var_epi64(ga01, mask0, ga23));
-    d = v_uint64x8(_mm512_permutex2var_epi64(ga01, mask1, ga23));
-}
-
-///////////////////////////// store interleave /////////////////////////////////////
-
-inline void v_store_interleave( uchar* ptr, const v_uint8x64& x, const v_uint8x64& y,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    v_uint8x64 low, high;
-    v_zip(x, y, low, high);
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm512_stream_si512((__m512i*)ptr, low.val);
-        _mm512_stream_si512((__m512i*)(ptr + 64), high.val);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm512_store_si512((__m512i*)ptr, low.val);
-        _mm512_store_si512((__m512i*)(ptr + 64), high.val);
-    }
-    else
-    {
-        _mm512_storeu_si512((__m512i*)ptr, low.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 64), high.val);
-    }
-}
-
-inline void v_store_interleave( ushort* ptr, const v_uint16x32& x, const v_uint16x32& y,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    v_uint16x32 low, high;
-    v_zip(x, y, low, high);
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm512_stream_si512((__m512i*)ptr, low.val);
-        _mm512_stream_si512((__m512i*)(ptr + 32), high.val);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm512_store_si512((__m512i*)ptr, low.val);
-        _mm512_store_si512((__m512i*)(ptr + 32), high.val);
-    }
-    else
-    {
-        _mm512_storeu_si512((__m512i*)ptr, low.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 32), high.val);
-    }
-}
-
-inline void v_store_interleave( unsigned* ptr, const v_uint32x16& x, const v_uint32x16& y,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    v_uint32x16 low, high;
-    v_zip(x, y, low, high);
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm512_stream_si512((__m512i*)ptr, low.val);
-        _mm512_stream_si512((__m512i*)(ptr + 16), high.val);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm512_store_si512((__m512i*)ptr, low.val);
-        _mm512_store_si512((__m512i*)(ptr + 16), high.val);
-    }
-    else
-    {
-        _mm512_storeu_si512((__m512i*)ptr, low.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 16), high.val);
-    }
-}
-
-inline void v_store_interleave( uint64* ptr, const v_uint64x8& x, const v_uint64x8& y,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    v_uint64x8 low, high;
-    v_zip(x, y, low, high);
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm512_stream_si512((__m512i*)ptr, low.val);
-        _mm512_stream_si512((__m512i*)(ptr + 8), high.val);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm512_store_si512((__m512i*)ptr, low.val);
-        _mm512_store_si512((__m512i*)(ptr + 8), high.val);
-    }
-    else
-    {
-        _mm512_storeu_si512((__m512i*)ptr, low.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 8), high.val);
-    }
-}
-
-inline void v_store_interleave( uchar* ptr, const v_uint8x64& a, const v_uint8x64& b, const v_uint8x64& c,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-#if CV_AVX_512VBMI
-    __m512i mask0 = _v512_set_epu8(127,  84,  20, 126,  83,  19, 125,  82,  18, 124,  81,  17, 123,  80,  16, 122,
-                                    79,  15, 121,  78,  14, 120,  77,  13, 119,  76,  12, 118,  75,  11, 117,  74,
-                                    10, 116,  73,   9, 115,  72,   8, 114,  71,   7, 113,  70,   6, 112,  69,   5,
-                                   111,  68,   4, 110,  67,   3, 109,  66,   2, 108,  65,   1, 107,  64,   0, 106);
-    __m512i mask1 = _v512_set_epu8( 21,  42, 105,  20,  41, 104,  19,  40, 103,  18,  39, 102,  17,  38, 101,  16,
-                                    37, 100,  15,  36,  99,  14,  35,  98,  13,  34,  97,  12,  33,  96,  11,  32,
-                                    95,  10,  31,  94,   9,  30,  93,   8,  29,  92,   7,  28,  91,   6,  27,  90,
-                                     5,  26,  89,   4,  25,  88,   3,  24,  87,   2,  23,  86,   1,  22,  85,   0);
-    __m512i mask2 = _v512_set_epu8(106, 127,  63, 105, 126,  62, 104, 125,  61, 103, 124,  60, 102, 123,  59, 101,
-                                   122,  58, 100, 121,  57,  99, 120,  56,  98, 119,  55,  97, 118,  54,  96, 117,
-                                    53,  95, 116,  52,  94, 115,  51,  93, 114,  50,  92, 113,  49,  91, 112,  48,
-                                    90, 111,  47,  89, 110,  46,  88, 109,  45,  87, 108,  44,  86, 107,  43,  85);
-    __m512i r2g0r0 = _mm512_permutex2var_epi8(b.val, mask0, c.val);
-    __m512i b0r1b1 = _mm512_permutex2var_epi8(a.val, mask1, c.val);
-    __m512i g1b2g2 = _mm512_permutex2var_epi8(a.val, mask2, b.val);
-
-    __m512i bgr0 = _mm512_mask_blend_epi8(0x9249249249249249, r2g0r0, b0r1b1);
-    __m512i bgr1 = _mm512_mask_blend_epi8(0x9249249249249249, b0r1b1, g1b2g2);
-    __m512i bgr2 = _mm512_mask_blend_epi8(0x9249249249249249, g1b2g2, r2g0r0);
-#else
-    __m512i g1g0 = _mm512_shuffle_epi8(b.val, _mm512_set4_epi32(0x0e0f0c0d, 0x0a0b0809, 0x06070405, 0x02030001));
-    __m512i b0g0 = _mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, a.val, g1g0);
-    __m512i r0b1 = _mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, c.val, a.val);
-    __m512i g1r1 = _mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, g1g0, c.val);
-
-    __m512i mask0 = _v512_set_epu16(42, 10, 31, 41,  9, 30, 40,  8, 29, 39,  7, 28, 38,  6, 27, 37,
-                                     5, 26, 36,  4, 25, 35,  3, 24, 34,  2, 23, 33,  1, 22, 32,  0);
-    __m512i mask1 = _v512_set_epu16(21, 52, 41, 20, 51, 40, 19, 50, 39, 18, 49, 38, 17, 48, 37, 16,
-                                    47, 36, 15, 46, 35, 14, 45, 34, 13, 44, 33, 12, 43, 32, 11, 42);
-    __m512i mask2 = _v512_set_epu16(63, 31, 20, 62, 30, 19, 61, 29, 18, 60, 28, 17, 59, 27, 16, 58,
-                                    26, 15, 57, 25, 14, 56, 24, 13, 55, 23, 12, 54, 22, 11, 53, 21);
-    __m512i b0g0b2 = _mm512_permutex2var_epi16(b0g0, mask0, r0b1);
-    __m512i r1b1r0 = _mm512_permutex2var_epi16(b0g0, mask1, g1r1);
-    __m512i g2r2g1 = _mm512_permutex2var_epi16(r0b1, mask2, g1r1);
-
-    __m512i bgr0 = _mm512_mask_blend_epi16(0x24924924, b0g0b2, r1b1r0);
-    __m512i bgr1 = _mm512_mask_blend_epi16(0x24924924, r1b1r0, g2r2g1);
-    __m512i bgr2 = _mm512_mask_blend_epi16(0x24924924, g2r2g1, b0g0b2);
-#endif
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm512_stream_si512((__m512i*)ptr, bgr0);
-        _mm512_stream_si512((__m512i*)(ptr + 64), bgr1);
-        _mm512_stream_si512((__m512i*)(ptr + 128), bgr2);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm512_store_si512((__m512i*)ptr, bgr0);
-        _mm512_store_si512((__m512i*)(ptr + 64), bgr1);
-        _mm512_store_si512((__m512i*)(ptr + 128), bgr2);
-    }
-    else
-    {
-        _mm512_storeu_si512((__m512i*)ptr, bgr0);
-        _mm512_storeu_si512((__m512i*)(ptr + 64), bgr1);
-        _mm512_storeu_si512((__m512i*)(ptr + 128), bgr2);
-    }
-}
-
-inline void v_store_interleave( ushort* ptr, const v_uint16x32& a, const v_uint16x32& b, const v_uint16x32& c,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    __m512i mask0 = _v512_set_epu16(42, 10, 31, 41,  9, 30, 40,  8, 29, 39,  7, 28, 38,  6, 27, 37,
-                                     5, 26, 36,  4, 25, 35,  3, 24, 34,  2, 23, 33,  1, 22, 32,  0);
-    __m512i mask1 = _v512_set_epu16(21, 52, 41, 20, 51, 40, 19, 50, 39, 18, 49, 38, 17, 48, 37, 16,
-                                    47, 36, 15, 46, 35, 14, 45, 34, 13, 44, 33, 12, 43, 32, 11, 42);
-    __m512i mask2 = _v512_set_epu16(63, 31, 20, 62, 30, 19, 61, 29, 18, 60, 28, 17, 59, 27, 16, 58,
-                                    26, 15, 57, 25, 14, 56, 24, 13, 55, 23, 12, 54, 22, 11, 53, 21);
-    __m512i b0g0b2 = _mm512_permutex2var_epi16(a.val, mask0, b.val);
-    __m512i r1b1r0 = _mm512_permutex2var_epi16(a.val, mask1, c.val);
-    __m512i g2r2g1 = _mm512_permutex2var_epi16(b.val, mask2, c.val);
-
-    __m512i bgr0 = _mm512_mask_blend_epi16(0x24924924, b0g0b2, r1b1r0);
-    __m512i bgr1 = _mm512_mask_blend_epi16(0x24924924, r1b1r0, g2r2g1);
-    __m512i bgr2 = _mm512_mask_blend_epi16(0x24924924, g2r2g1, b0g0b2);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm512_stream_si512((__m512i*)ptr, bgr0);
-        _mm512_stream_si512((__m512i*)(ptr + 32), bgr1);
-        _mm512_stream_si512((__m512i*)(ptr + 64), bgr2);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm512_store_si512((__m512i*)ptr, bgr0);
-        _mm512_store_si512((__m512i*)(ptr + 32), bgr1);
-        _mm512_store_si512((__m512i*)(ptr + 64), bgr2);
-    }
-    else
-    {
-        _mm512_storeu_si512((__m512i*)ptr, bgr0);
-        _mm512_storeu_si512((__m512i*)(ptr + 32), bgr1);
-        _mm512_storeu_si512((__m512i*)(ptr + 64), bgr2);
-    }
-}
-
-inline void v_store_interleave( unsigned* ptr, const v_uint32x16& a, const v_uint32x16& b, const v_uint32x16& c,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    __m512i mask0 = _v512_set_epu32(26, 31, 15, 25, 30, 14, 24, 29, 13, 23, 28, 12, 22, 27, 11, 21);
-    __m512i mask1 = _v512_set_epu32(31, 10, 25, 30,  9, 24, 29,  8, 23, 28,  7, 22, 27,  6, 21, 26);
-    __m512i g1b2g2 = _mm512_permutex2var_epi32(a.val, mask0, b.val);
-    __m512i r2r1b1 = _mm512_permutex2var_epi32(a.val, mask1, c.val);
-
-    __m512i bgr0 = _mm512_mask_expand_epi32(_mm512_mask_expand_epi32(_mm512_maskz_expand_epi32(0x9249, a.val), 0x2492, b.val), 0x4924, c.val);
-    __m512i bgr1 = _mm512_mask_blend_epi32(0x9249, r2r1b1, g1b2g2);
-    __m512i bgr2 = _mm512_mask_blend_epi32(0x9249, g1b2g2, r2r1b1);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm512_stream_si512((__m512i*)ptr, bgr0);
-        _mm512_stream_si512((__m512i*)(ptr + 16), bgr1);
-        _mm512_stream_si512((__m512i*)(ptr + 32), bgr2);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm512_store_si512((__m512i*)ptr, bgr0);
-        _mm512_store_si512((__m512i*)(ptr + 16), bgr1);
-        _mm512_store_si512((__m512i*)(ptr + 32), bgr2);
-    }
-    else
-    {
-        _mm512_storeu_si512((__m512i*)ptr, bgr0);
-        _mm512_storeu_si512((__m512i*)(ptr + 16), bgr1);
-        _mm512_storeu_si512((__m512i*)(ptr + 32), bgr2);
-    }
-}
-
-inline void v_store_interleave( uint64* ptr, const v_uint64x8& a, const v_uint64x8& b, const v_uint64x8& c,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    __m512i mask0 = _v512_set_epu64( 5, 12,  7,  4, 11,  6,  3, 10);
-    __m512i mask1 = _v512_set_epu64(15,  7,  4, 14,  6,  3, 13,  5);
-    __m512i r1b1b2 = _mm512_permutex2var_epi64(a.val, mask0, c.val);
-    __m512i g2r2g1 = _mm512_permutex2var_epi64(b.val, mask1, c.val);
-
-    __m512i bgr0 = _mm512_mask_expand_epi64(_mm512_mask_expand_epi64(_mm512_maskz_expand_epi64(0x49, a.val), 0x92, b.val), 0x24, c.val);
-    __m512i bgr1 = _mm512_mask_blend_epi64(0xdb, g2r2g1, r1b1b2);
-    __m512i bgr2 = _mm512_mask_blend_epi64(0xdb, r1b1b2, g2r2g1);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm512_stream_si512((__m512i*)ptr, bgr0);
-        _mm512_stream_si512((__m512i*)(ptr + 8), bgr1);
-        _mm512_stream_si512((__m512i*)(ptr + 16), bgr2);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm512_store_si512((__m512i*)ptr, bgr0);
-        _mm512_store_si512((__m512i*)(ptr + 8), bgr1);
-        _mm512_store_si512((__m512i*)(ptr + 16), bgr2);
-    }
-    else
-    {
-        _mm512_storeu_si512((__m512i*)ptr, bgr0);
-        _mm512_storeu_si512((__m512i*)(ptr + 8), bgr1);
-        _mm512_storeu_si512((__m512i*)(ptr + 16), bgr2);
-    }
-}
-
-inline void v_store_interleave( uchar* ptr, const v_uint8x64& a, const v_uint8x64& b,
-                                const v_uint8x64& c, const v_uint8x64& d,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    v_uint8x64 br01, br23, ga01, ga23;
-    v_zip(a, c, br01, br23);
-    v_zip(b, d, ga01, ga23);
-    v_uint8x64 bgra0, bgra1, bgra2, bgra3;
-    v_zip(br01, ga01, bgra0, bgra1);
-    v_zip(br23, ga23, bgra2, bgra3);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm512_stream_si512((__m512i*)ptr, bgra0.val);
-        _mm512_stream_si512((__m512i*)(ptr + 64), bgra1.val);
-        _mm512_stream_si512((__m512i*)(ptr + 128), bgra2.val);
-        _mm512_stream_si512((__m512i*)(ptr + 192), bgra3.val);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm512_store_si512((__m512i*)ptr, bgra0.val);
-        _mm512_store_si512((__m512i*)(ptr + 64), bgra1.val);
-        _mm512_store_si512((__m512i*)(ptr + 128), bgra2.val);
-        _mm512_store_si512((__m512i*)(ptr + 192), bgra3.val);
-    }
-    else
-    {
-        _mm512_storeu_si512((__m512i*)ptr, bgra0.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 64), bgra1.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 128), bgra2.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 192), bgra3.val);
-    }
-}
-
-inline void v_store_interleave( ushort* ptr, const v_uint16x32& a, const v_uint16x32& b,
-                                const v_uint16x32& c, const v_uint16x32& d,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    v_uint16x32 br01, br23, ga01, ga23;
-    v_zip(a, c, br01, br23);
-    v_zip(b, d, ga01, ga23);
-    v_uint16x32 bgra0, bgra1, bgra2, bgra3;
-    v_zip(br01, ga01, bgra0, bgra1);
-    v_zip(br23, ga23, bgra2, bgra3);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm512_stream_si512((__m512i*)ptr, bgra0.val);
-        _mm512_stream_si512((__m512i*)(ptr + 32), bgra1.val);
-        _mm512_stream_si512((__m512i*)(ptr + 64), bgra2.val);
-        _mm512_stream_si512((__m512i*)(ptr + 96), bgra3.val);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm512_store_si512((__m512i*)ptr, bgra0.val);
-        _mm512_store_si512((__m512i*)(ptr + 32), bgra1.val);
-        _mm512_store_si512((__m512i*)(ptr + 64), bgra2.val);
-        _mm512_store_si512((__m512i*)(ptr + 96), bgra3.val);
-    }
-    else
-    {
-        _mm512_storeu_si512((__m512i*)ptr, bgra0.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 32), bgra1.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 64), bgra2.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 96), bgra3.val);
-    }
-}
-
-inline void v_store_interleave( unsigned* ptr, const v_uint32x16& a, const v_uint32x16& b,
-                                const v_uint32x16& c, const v_uint32x16& d,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    v_uint32x16 br01, br23, ga01, ga23;
-    v_zip(a, c, br01, br23);
-    v_zip(b, d, ga01, ga23);
-    v_uint32x16 bgra0, bgra1, bgra2, bgra3;
-    v_zip(br01, ga01, bgra0, bgra1);
-    v_zip(br23, ga23, bgra2, bgra3);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm512_stream_si512((__m512i*)ptr, bgra0.val);
-        _mm512_stream_si512((__m512i*)(ptr + 16), bgra1.val);
-        _mm512_stream_si512((__m512i*)(ptr + 32), bgra2.val);
-        _mm512_stream_si512((__m512i*)(ptr + 48), bgra3.val);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm512_store_si512((__m512i*)ptr, bgra0.val);
-        _mm512_store_si512((__m512i*)(ptr + 16), bgra1.val);
-        _mm512_store_si512((__m512i*)(ptr + 32), bgra2.val);
-        _mm512_store_si512((__m512i*)(ptr + 48), bgra3.val);
-    }
-    else
-    {
-        _mm512_storeu_si512((__m512i*)ptr, bgra0.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 16), bgra1.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 32), bgra2.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 48), bgra3.val);
-    }
-}
-
-inline void v_store_interleave( uint64* ptr, const v_uint64x8& a, const v_uint64x8& b,
-                                const v_uint64x8& c, const v_uint64x8& d,
-                                hal::StoreMode mode=hal::STORE_UNALIGNED )
-{
-    v_uint64x8 br01, br23, ga01, ga23;
-    v_zip(a, c, br01, br23);
-    v_zip(b, d, ga01, ga23);
-    v_uint64x8 bgra0, bgra1, bgra2, bgra3;
-    v_zip(br01, ga01, bgra0, bgra1);
-    v_zip(br23, ga23, bgra2, bgra3);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm512_stream_si512((__m512i*)ptr, bgra0.val);
-        _mm512_stream_si512((__m512i*)(ptr + 8), bgra1.val);
-        _mm512_stream_si512((__m512i*)(ptr + 16), bgra2.val);
-        _mm512_stream_si512((__m512i*)(ptr + 24), bgra3.val);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm512_store_si512((__m512i*)ptr, bgra0.val);
-        _mm512_store_si512((__m512i*)(ptr + 8), bgra1.val);
-        _mm512_store_si512((__m512i*)(ptr + 16), bgra2.val);
-        _mm512_store_si512((__m512i*)(ptr + 24), bgra3.val);
-    }
-    else
-    {
-        _mm512_storeu_si512((__m512i*)ptr, bgra0.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 8), bgra1.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 16), bgra2.val);
-        _mm512_storeu_si512((__m512i*)(ptr + 24), bgra3.val);
-    }
-}
-
-#define OPENCV_HAL_IMPL_AVX512_LOADSTORE_INTERLEAVE(_Tpvec0, _Tp0, suffix0, _Tpvec1, _Tp1, suffix1) \
-inline void v_load_deinterleave( const _Tp0* ptr, _Tpvec0& a0, _Tpvec0& b0 ) \
-{ \
-    _Tpvec1 a1, b1; \
-    v_load_deinterleave((const _Tp1*)ptr, a1, b1); \
-    a0 = v_reinterpret_as_##suffix0(a1); \
-    b0 = v_reinterpret_as_##suffix0(b1); \
-} \
-inline void v_load_deinterleave( const _Tp0* ptr, _Tpvec0& a0, _Tpvec0& b0, _Tpvec0& c0 ) \
-{ \
-    _Tpvec1 a1, b1, c1; \
-    v_load_deinterleave((const _Tp1*)ptr, a1, b1, c1); \
-    a0 = v_reinterpret_as_##suffix0(a1); \
-    b0 = v_reinterpret_as_##suffix0(b1); \
-    c0 = v_reinterpret_as_##suffix0(c1); \
-} \
-inline void v_load_deinterleave( const _Tp0* ptr, _Tpvec0& a0, _Tpvec0& b0, _Tpvec0& c0, _Tpvec0& d0 ) \
-{ \
-    _Tpvec1 a1, b1, c1, d1; \
-    v_load_deinterleave((const _Tp1*)ptr, a1, b1, c1, d1); \
-    a0 = v_reinterpret_as_##suffix0(a1); \
-    b0 = v_reinterpret_as_##suffix0(b1); \
-    c0 = v_reinterpret_as_##suffix0(c1); \
-    d0 = v_reinterpret_as_##suffix0(d1); \
-} \
-inline void v_store_interleave( _Tp0* ptr, const _Tpvec0& a0, const _Tpvec0& b0, \
-                                hal::StoreMode mode=hal::STORE_UNALIGNED ) \
-{ \
-    _Tpvec1 a1 = v_reinterpret_as_##suffix1(a0); \
-    _Tpvec1 b1 = v_reinterpret_as_##suffix1(b0); \
-    v_store_interleave((_Tp1*)ptr, a1, b1, mode);      \
-} \
-inline void v_store_interleave( _Tp0* ptr, const _Tpvec0& a0, const _Tpvec0& b0, const _Tpvec0& c0, \
-                                hal::StoreMode mode=hal::STORE_UNALIGNED ) \
-{ \
-    _Tpvec1 a1 = v_reinterpret_as_##suffix1(a0); \
-    _Tpvec1 b1 = v_reinterpret_as_##suffix1(b0); \
-    _Tpvec1 c1 = v_reinterpret_as_##suffix1(c0); \
-    v_store_interleave((_Tp1*)ptr, a1, b1, c1, mode);  \
-} \
-inline void v_store_interleave( _Tp0* ptr, const _Tpvec0& a0, const _Tpvec0& b0, \
-                                const _Tpvec0& c0, const _Tpvec0& d0, \
-                                hal::StoreMode mode=hal::STORE_UNALIGNED ) \
-{ \
-    _Tpvec1 a1 = v_reinterpret_as_##suffix1(a0); \
-    _Tpvec1 b1 = v_reinterpret_as_##suffix1(b0); \
-    _Tpvec1 c1 = v_reinterpret_as_##suffix1(c0); \
-    _Tpvec1 d1 = v_reinterpret_as_##suffix1(d0); \
-    v_store_interleave((_Tp1*)ptr, a1, b1, c1, d1, mode); \
-}
-
-OPENCV_HAL_IMPL_AVX512_LOADSTORE_INTERLEAVE(v_int8x64, schar, s8, v_uint8x64, uchar, u8)
-OPENCV_HAL_IMPL_AVX512_LOADSTORE_INTERLEAVE(v_int16x32, short, s16, v_uint16x32, ushort, u16)
-OPENCV_HAL_IMPL_AVX512_LOADSTORE_INTERLEAVE(v_int32x16, int, s32, v_uint32x16, unsigned, u32)
-OPENCV_HAL_IMPL_AVX512_LOADSTORE_INTERLEAVE(v_float32x16, float, f32, v_uint32x16, unsigned, u32)
-OPENCV_HAL_IMPL_AVX512_LOADSTORE_INTERLEAVE(v_int64x8, int64, s64, v_uint64x8, uint64, u64)
-OPENCV_HAL_IMPL_AVX512_LOADSTORE_INTERLEAVE(v_float64x8, double, f64, v_uint64x8, uint64, u64)
-
-////////// Mask and checks /////////
-
-/** Mask **/
-inline int64 v_signmask(const v_int8x64& a) { return (int64)_mm512_movepi8_mask(a.val); }
-inline int v_signmask(const v_int16x32& a) { return (int)_mm512_cmp_epi16_mask(a.val, _mm512_setzero_si512(), _MM_CMPINT_LT); }
-inline int v_signmask(const v_int32x16& a) { return (int)_mm512_cmp_epi32_mask(a.val, _mm512_setzero_si512(), _MM_CMPINT_LT); }
-inline int v_signmask(const v_int64x8& a) { return (int)_mm512_cmp_epi64_mask(a.val, _mm512_setzero_si512(), _MM_CMPINT_LT); }
-
-inline int64 v_signmask(const v_uint8x64& a) { return v_signmask(v_reinterpret_as_s8(a)); }
-inline int v_signmask(const v_uint16x32& a) { return v_signmask(v_reinterpret_as_s16(a)); }
-inline int v_signmask(const v_uint32x16& a) { return v_signmask(v_reinterpret_as_s32(a)); }
-inline int v_signmask(const v_uint64x8& a) { return v_signmask(v_reinterpret_as_s64(a)); }
-inline int v_signmask(const v_float32x16& a) { return v_signmask(v_reinterpret_as_s32(a)); }
-inline int v_signmask(const v_float64x8& a) { return v_signmask(v_reinterpret_as_s64(a)); }
-
-/** Checks **/
-inline bool v_check_all(const v_int8x64& a) { return !(bool)_mm512_cmp_epi8_mask(a.val, _mm512_setzero_si512(), _MM_CMPINT_NLT); }
-inline bool v_check_any(const v_int8x64& a) { return (bool)_mm512_movepi8_mask(a.val); }
-inline bool v_check_all(const v_int16x32& a) { return !(bool)_mm512_cmp_epi16_mask(a.val, _mm512_setzero_si512(), _MM_CMPINT_NLT); }
-inline bool v_check_any(const v_int16x32& a) { return (bool)_mm512_cmp_epi16_mask(a.val, _mm512_setzero_si512(), _MM_CMPINT_LT); }
-inline bool v_check_all(const v_int32x16& a) { return !(bool)_mm512_cmp_epi32_mask(a.val, _mm512_setzero_si512(), _MM_CMPINT_NLT); }
-inline bool v_check_any(const v_int32x16& a) { return (bool)_mm512_cmp_epi32_mask(a.val, _mm512_setzero_si512(), _MM_CMPINT_LT); }
-inline bool v_check_all(const v_int64x8& a) { return !(bool)_mm512_cmp_epi64_mask(a.val, _mm512_setzero_si512(), _MM_CMPINT_NLT); }
-inline bool v_check_any(const v_int64x8& a) { return (bool)_mm512_cmp_epi64_mask(a.val, _mm512_setzero_si512(), _MM_CMPINT_LT); }
-
-inline bool v_check_all(const v_float32x16& a) { return v_check_all(v_reinterpret_as_s32(a)); }
-inline bool v_check_any(const v_float32x16& a) { return v_check_any(v_reinterpret_as_s32(a)); }
-inline bool v_check_all(const v_float64x8& a) { return v_check_all(v_reinterpret_as_s64(a)); }
-inline bool v_check_any(const v_float64x8& a) { return v_check_any(v_reinterpret_as_s64(a)); }
-inline bool v_check_all(const v_uint8x64& a) { return v_check_all(v_reinterpret_as_s8(a)); }
-inline bool v_check_all(const v_uint16x32& a) { return v_check_all(v_reinterpret_as_s16(a)); }
-inline bool v_check_all(const v_uint32x16& a) { return v_check_all(v_reinterpret_as_s32(a)); }
-inline bool v_check_all(const v_uint64x8& a) { return v_check_all(v_reinterpret_as_s64(a)); }
-inline bool v_check_any(const v_uint8x64& a) { return v_check_any(v_reinterpret_as_s8(a)); }
-inline bool v_check_any(const v_uint16x32& a) { return v_check_any(v_reinterpret_as_s16(a)); }
-inline bool v_check_any(const v_uint32x16& a) { return v_check_any(v_reinterpret_as_s32(a)); }
-inline bool v_check_any(const v_uint64x8& a) { return v_check_any(v_reinterpret_as_s64(a)); }
-
-inline int v_scan_forward(const v_int8x64& a)
-{
-    int64 mask = _mm512_movepi8_mask(a.val);
-    int mask32 = (int)mask;
-    return mask != 0 ? mask32 != 0 ? trailingZeros32(mask32) : 32 + trailingZeros32((int)(mask >> 32)) : 0;
-}
-inline int v_scan_forward(const v_uint8x64& a) { return v_scan_forward(v_reinterpret_as_s8(a)); }
-inline int v_scan_forward(const v_int16x32& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s16(a))); }
-inline int v_scan_forward(const v_uint16x32& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s16(a))); }
-inline int v_scan_forward(const v_int32x16& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s16(a))) / 2; }
-inline int v_scan_forward(const v_uint32x16& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s16(a))) / 2; }
-inline int v_scan_forward(const v_float32x16& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s16(a))) / 2; }
-inline int v_scan_forward(const v_int64x8& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s16(a))) / 4; }
-inline int v_scan_forward(const v_uint64x8& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s16(a))) / 4; }
-inline int v_scan_forward(const v_float64x8& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s16(a))) / 4; }
-
-inline void v512_cleanup() { _mm256_zeroall(); }
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END
-
-//! @endcond
-
-} // cv::
-
-#endif // OPENCV_HAL_INTRIN_AVX_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_cpp.hpp b/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_cpp.hpp
deleted file mode 100644
index 46222140..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_cpp.hpp
+++ /dev/null
@@ -1,3320 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Copyright (C) 2015, Itseez Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_HAL_INTRIN_CPP_HPP
-#define OPENCV_HAL_INTRIN_CPP_HPP
-
-#include <limits>
-#include <cstring>
-#include <algorithm>
-#include "opencv2/core/saturate.hpp"
-
-//! @cond IGNORED
-#define CV_SIMD128_CPP 1
-#if defined(CV_FORCE_SIMD128_CPP)
-#define CV_SIMD128 1
-#define CV_SIMD128_64F 1
-#endif
-#if defined(CV_DOXYGEN)
-#define CV_SIMD128 1
-#define CV_SIMD128_64F 1
-#define CV_SIMD256 1
-#define CV_SIMD256_64F 1
-#define CV_SIMD512 1
-#define CV_SIMD512_64F 1
-#else
-#define CV_SIMD256 0 // Explicitly disable SIMD256 and SIMD512 support for scalar intrinsic implementation
-#define CV_SIMD512 0 // to avoid warnings during compilation
-#endif
-//! @endcond
-
-namespace cv
-{
-
-#ifndef CV_DOXYGEN
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
-#endif
-
-/** @addtogroup core_hal_intrin
-
-"Universal intrinsics" is a types and functions set intended to simplify vectorization of code on
-different platforms. Currently a few different SIMD extensions on different architectures are supported.
-128 bit registers of various types support is implemented for a wide range of architectures
-including x86(__SSE/SSE2/SSE4.2__), ARM(__NEON__), PowerPC(__VSX__), MIPS(__MSA__).
-256 bit long registers are supported on x86(__AVX2__) and 512 bit long registers are supported on x86(__AVX512__).
-In case when there is no SIMD extension available during compilation, fallback C++ implementation of intrinsics
-will be chosen and code will work as expected although it could be slower.
-
-### Types
-
-There are several types representing packed values vector registers, each type is
-implemented as a structure based on a one SIMD register.
-
-- cv::v_uint8 and cv::v_int8: 8-bit integer values (unsigned/signed) - char
-- cv::v_uint16 and cv::v_int16: 16-bit integer values (unsigned/signed) - short
-- cv::v_uint32 and cv::v_int32: 32-bit integer values (unsigned/signed) - int
-- cv::v_uint64 and cv::v_int64: 64-bit integer values (unsigned/signed) - int64
-- cv::v_float32: 32-bit floating point values (signed) - float
-- cv::v_float64: 64-bit floating point values (signed) - double
-
-Exact bit length(and value quantity) of listed types is compile time deduced and depends on architecture SIMD
-capabilities chosen as available during compilation of the library. All the types contains __nlanes__ enumeration
-to check for exact value quantity of the type.
-
-In case the exact bit length of the type is important it is possible to use specific fixed length register types.
-
-There are several types representing 128-bit registers.
-
-- cv::v_uint8x16 and cv::v_int8x16: sixteen 8-bit integer values (unsigned/signed) - char
-- cv::v_uint16x8 and cv::v_int16x8: eight 16-bit integer values (unsigned/signed) - short
-- cv::v_uint32x4 and cv::v_int32x4: four 32-bit integer values (unsigned/signed) - int
-- cv::v_uint64x2 and cv::v_int64x2: two 64-bit integer values (unsigned/signed) - int64
-- cv::v_float32x4: four 32-bit floating point values (signed) - float
-- cv::v_float64x2: two 64-bit floating point values (signed) - double
-
-There are several types representing 256-bit registers.
-
-- cv::v_uint8x32 and cv::v_int8x32: thirty two 8-bit integer values (unsigned/signed) - char
-- cv::v_uint16x16 and cv::v_int16x16: sixteen 16-bit integer values (unsigned/signed) - short
-- cv::v_uint32x8 and cv::v_int32x8: eight 32-bit integer values (unsigned/signed) - int
-- cv::v_uint64x4 and cv::v_int64x4: four 64-bit integer values (unsigned/signed) - int64
-- cv::v_float32x8: eight 32-bit floating point values (signed) - float
-- cv::v_float64x4: four 64-bit floating point values (signed) - double
-
-@note
-256 bit registers at the moment implemented for AVX2 SIMD extension only, if you want to use this type directly,
-don't forget to check the CV_SIMD256 preprocessor definition:
-@code
-#if CV_SIMD256
-//...
-#endif
-@endcode
-
-There are several types representing 512-bit registers.
-
-- cv::v_uint8x64 and cv::v_int8x64: sixty four 8-bit integer values (unsigned/signed) - char
-- cv::v_uint16x32 and cv::v_int16x32: thirty two 16-bit integer values (unsigned/signed) - short
-- cv::v_uint32x16 and cv::v_int32x16: sixteen 32-bit integer values (unsigned/signed) - int
-- cv::v_uint64x8 and cv::v_int64x8: eight 64-bit integer values (unsigned/signed) - int64
-- cv::v_float32x16: sixteen 32-bit floating point values (signed) - float
-- cv::v_float64x8: eight 64-bit floating point values (signed) - double
-@note
-512 bit registers at the moment implemented for AVX512 SIMD extension only, if you want to use this type directly,
-don't forget to check the CV_SIMD512 preprocessor definition.
-
-@note
-cv::v_float64x2 is not implemented in NEON variant, if you want to use this type, don't forget to
-check the CV_SIMD128_64F preprocessor definition.
-
-### Load and store operations
-
-These operations allow to set contents of the register explicitly or by loading it from some memory
-block and to save contents of the register to memory block.
-
-There are variable size register load operations that provide result of maximum available size
-depending on chosen platform capabilities.
-- Constructors:
-@ref v_reg::v_reg(const _Tp *ptr) "from memory",
-- Other create methods:
-vx_setall_s8, vx_setall_u8, ...,
-vx_setzero_u8, vx_setzero_s8, ...
-- Memory load operations:
-vx_load, vx_load_aligned, vx_load_low, vx_load_halves,
-- Memory operations with expansion of values:
-vx_load_expand, vx_load_expand_q
-
-Also there are fixed size register load/store operations.
-
-For 128 bit registers
-- Constructors:
-@ref v_reg::v_reg(const _Tp *ptr) "from memory",
-@ref v_reg::v_reg(_Tp s0, _Tp s1) "from two values", ...
-- Other create methods:
-@ref v_setall_s8, @ref v_setall_u8, ...,
-@ref v_setzero_u8, @ref v_setzero_s8, ...
-- Memory load operations:
-@ref v_load, @ref v_load_aligned, @ref v_load_low, @ref v_load_halves,
-- Memory operations with expansion of values:
-@ref v_load_expand, @ref v_load_expand_q
-
-For 256 bit registers(check CV_SIMD256 preprocessor definition)
-- Constructors:
-@ref v_reg::v_reg(const _Tp *ptr) "from memory",
-@ref v_reg::v_reg(_Tp s0, _Tp s1, _Tp s2, _Tp s3) "from four values", ...
-- Other create methods:
-@ref v256_setall_s8, @ref v256_setall_u8, ...,
-@ref v256_setzero_u8, @ref v256_setzero_s8, ...
-- Memory load operations:
-@ref v256_load, @ref v256_load_aligned, @ref v256_load_low, @ref v256_load_halves,
-- Memory operations with expansion of values:
-@ref v256_load_expand, @ref v256_load_expand_q
-
-For 512 bit registers(check CV_SIMD512 preprocessor definition)
-- Constructors:
-@ref v_reg::v_reg(const _Tp *ptr) "from memory",
-@ref v_reg::v_reg(_Tp s0, _Tp s1, _Tp s2, _Tp s3, _Tp s4, _Tp s5, _Tp s6, _Tp s7) "from eight values", ...
-- Other create methods:
-@ref v512_setall_s8, @ref v512_setall_u8, ...,
-@ref v512_setzero_u8, @ref v512_setzero_s8, ...
-- Memory load operations:
-@ref v512_load, @ref v512_load_aligned, @ref v512_load_low, @ref v512_load_halves,
-- Memory operations with expansion of values:
-@ref v512_load_expand, @ref v512_load_expand_q
-
-Store to memory operations are similar across different platform capabilities:
-@ref v_store, @ref v_store_aligned,
-@ref v_store_high, @ref v_store_low
-
-### Value reordering
-
-These operations allow to reorder or recombine elements in one or multiple vectors.
-
-- Interleave, deinterleave (2, 3 and 4 channels): @ref v_load_deinterleave, @ref v_store_interleave
-- Expand: @ref v_expand, @ref v_expand_low, @ref v_expand_high
-- Pack: @ref v_pack, @ref v_pack_u, @ref v_pack_b, @ref v_rshr_pack, @ref v_rshr_pack_u,
-@ref v_pack_store, @ref v_pack_u_store, @ref v_rshr_pack_store, @ref v_rshr_pack_u_store
-- Recombine: @ref v_zip, @ref v_recombine, @ref v_combine_low, @ref v_combine_high
-- Reverse: @ref v_reverse
-- Extract: @ref v_extract
-
-
-### Arithmetic, bitwise and comparison operations
-
-Element-wise binary and unary operations.
-
-- Arithmetics:
-@ref operator +(const v_reg &a, const v_reg &b) "+",
-@ref operator -(const v_reg &a, const v_reg &b) "-",
-@ref operator *(const v_reg &a, const v_reg &b) "*",
-@ref operator /(const v_reg &a, const v_reg &b) "/",
-@ref v_mul_expand
-
-- Non-saturating arithmetics: @ref v_add_wrap, @ref v_sub_wrap
-
-- Bitwise shifts:
-@ref operator <<(const v_reg &a, int s) "<<",
-@ref operator >>(const v_reg &a, int s) ">>",
-@ref v_shl, @ref v_shr
-
-- Bitwise logic:
-@ref operator &(const v_reg &a, const v_reg &b) "&",
-@ref operator |(const v_reg &a, const v_reg &b) "|",
-@ref operator ^(const v_reg &a, const v_reg &b) "^",
-@ref operator ~(const v_reg &a) "~"
-
-- Comparison:
-@ref operator >(const v_reg &a, const v_reg &b) ">",
-@ref operator >=(const v_reg &a, const v_reg &b) ">=",
-@ref operator <(const v_reg &a, const v_reg &b) "<",
-@ref operator <=(const v_reg &a, const v_reg &b) "<=",
-@ref operator ==(const v_reg &a, const v_reg &b) "==",
-@ref operator !=(const v_reg &a, const v_reg &b) "!="
-
-- min/max: @ref v_min, @ref v_max
-
-### Reduce and mask
-
-Most of these operations return only one value.
-
-- Reduce: @ref v_reduce_min, @ref v_reduce_max, @ref v_reduce_sum, @ref v_popcount
-- Mask: @ref v_signmask, @ref v_check_all, @ref v_check_any, @ref v_select
-
-### Other math
-
-- Some frequent operations: @ref v_sqrt, @ref v_invsqrt, @ref v_magnitude, @ref v_sqr_magnitude
-- Absolute values: @ref v_abs, @ref v_absdiff, @ref v_absdiffs
-
-### Conversions
-
-Different type conversions and casts:
-
-- Rounding: @ref v_round, @ref v_floor, @ref v_ceil, @ref v_trunc,
-- To float: @ref v_cvt_f32, @ref v_cvt_f64
-- Reinterpret: @ref v_reinterpret_as_u8, @ref v_reinterpret_as_s8, ...
-
-### Matrix operations
-
-In these operations vectors represent matrix rows/columns: @ref v_dotprod, @ref v_dotprod_fast,
-@ref v_dotprod_expand, @ref v_dotprod_expand_fast, @ref v_matmul, @ref v_transpose4x4
-
-### Usability
-
-Most operations are implemented only for some subset of the available types, following matrices
-shows the applicability of different operations to the types.
-
-Regular integers:
-
-| Operations\\Types | uint 8 | int 8 | uint 16 | int 16 | uint 32 | int 32 |
-|-------------------|:-:|:-:|:-:|:-:|:-:|:-:|
-|load, store        | x | x | x | x | x | x |
-|interleave         | x | x | x | x | x | x |
-|expand             | x | x | x | x | x | x |
-|expand_low         | x | x | x | x | x | x |
-|expand_high        | x | x | x | x | x | x |
-|expand_q           | x | x |   |   |   |   |
-|add, sub           | x | x | x | x | x | x |
-|add_wrap, sub_wrap | x | x | x | x |   |   |
-|mul_wrap           | x | x | x | x |   |   |
-|mul                | x | x | x | x | x | x |
-|mul_expand         | x | x | x | x | x |   |
-|compare            | x | x | x | x | x | x |
-|shift              |   |   | x | x | x | x |
-|dotprod            |   |   |   | x |   | x |
-|dotprod_fast       |   |   |   | x |   | x |
-|dotprod_expand     | x | x | x | x |   | x |
-|dotprod_expand_fast| x | x | x | x |   | x |
-|logical            | x | x | x | x | x | x |
-|min, max           | x | x | x | x | x | x |
-|absdiff            | x | x | x | x | x | x |
-|absdiffs           |   | x |   | x |   |   |
-|reduce             | x | x | x | x | x | x |
-|mask               | x | x | x | x | x | x |
-|pack               | x | x | x | x | x | x |
-|pack_u             | x |   | x |   |   |   |
-|pack_b             | x |   |   |   |   |   |
-|unpack             | x | x | x | x | x | x |
-|extract            | x | x | x | x | x | x |
-|rotate (lanes)     | x | x | x | x | x | x |
-|cvt_flt32          |   |   |   |   |   | x |
-|cvt_flt64          |   |   |   |   |   | x |
-|transpose4x4       |   |   |   |   | x | x |
-|reverse            | x | x | x | x | x | x |
-|extract_n          | x | x | x | x | x | x |
-|broadcast_element  |   |   |   |   | x | x |
-
-Big integers:
-
-| Operations\\Types | uint 64 | int 64 |
-|-------------------|:-:|:-:|
-|load, store        | x | x |
-|add, sub           | x | x |
-|shift              | x | x |
-|logical            | x | x |
-|reverse            | x | x |
-|extract            | x | x |
-|rotate (lanes)     | x | x |
-|cvt_flt64          |   | x |
-|extract_n          | x | x |
-
-Floating point:
-
-| Operations\\Types | float 32 | float 64 |
-|-------------------|:-:|:-:|
-|load, store        | x | x |
-|interleave         | x |   |
-|add, sub           | x | x |
-|mul                | x | x |
-|div                | x | x |
-|compare            | x | x |
-|min, max           | x | x |
-|absdiff            | x | x |
-|reduce             | x |   |
-|mask               | x | x |
-|unpack             | x | x |
-|cvt_flt32          |   | x |
-|cvt_flt64          | x |   |
-|sqrt, abs          | x | x |
-|float math         | x | x |
-|transpose4x4       | x |   |
-|extract            | x | x |
-|rotate (lanes)     | x | x |
-|reverse            | x | x |
-|extract_n          | x | x |
-|broadcast_element  | x |   |
-
- @{ */
-
-template<typename _Tp, int n> struct v_reg
-{
-//! @cond IGNORED
-    typedef _Tp lane_type;
-    enum { nlanes = n };
-// !@endcond
-
-    /** @brief Constructor
-
-    Initializes register with data from memory
-    @param ptr pointer to memory block with data for register */
-    explicit v_reg(const _Tp* ptr) { for( int i = 0; i < n; i++ ) s[i] = ptr[i]; }
-
-    /** @brief Constructor
-
-    Initializes register with two 64-bit values */
-    v_reg(_Tp s0, _Tp s1) { s[0] = s0; s[1] = s1; }
-
-    /** @brief Constructor
-
-    Initializes register with four 32-bit values */
-    v_reg(_Tp s0, _Tp s1, _Tp s2, _Tp s3) { s[0] = s0; s[1] = s1; s[2] = s2; s[3] = s3; }
-
-    /** @brief Constructor
-
-    Initializes register with eight 16-bit values */
-    v_reg(_Tp s0, _Tp s1, _Tp s2, _Tp s3,
-           _Tp s4, _Tp s5, _Tp s6, _Tp s7)
-    {
-        s[0] = s0; s[1] = s1; s[2] = s2; s[3] = s3;
-        s[4] = s4; s[5] = s5; s[6] = s6; s[7] = s7;
-    }
-
-    /** @brief Constructor
-
-    Initializes register with sixteen 8-bit values */
-    v_reg(_Tp s0, _Tp s1, _Tp s2, _Tp s3,
-           _Tp s4, _Tp s5, _Tp s6, _Tp s7,
-           _Tp s8, _Tp s9, _Tp s10, _Tp s11,
-           _Tp s12, _Tp s13, _Tp s14, _Tp s15)
-    {
-        s[0] = s0; s[1] = s1; s[2] = s2; s[3] = s3;
-        s[4] = s4; s[5] = s5; s[6] = s6; s[7] = s7;
-        s[8] = s8; s[9] = s9; s[10] = s10; s[11] = s11;
-        s[12] = s12; s[13] = s13; s[14] = s14; s[15] = s15;
-    }
-
-    /** @brief Default constructor
-
-    Does not initialize anything*/
-    v_reg() {}
-
-    /** @brief Copy constructor */
-    v_reg(const v_reg<_Tp, n> & r)
-    {
-        for( int i = 0; i < n; i++ )
-            s[i] = r.s[i];
-    }
-    /** @brief Access first value
-
-    Returns value of the first lane according to register type, for example:
-    @code{.cpp}
-    v_int32x4 r(1, 2, 3, 4);
-    int v = r.get0(); // returns 1
-    v_uint64x2 r(1, 2);
-    uint64_t v = r.get0(); // returns 1
-    @endcode
-    */
-    _Tp get0() const { return s[0]; }
-
-//! @cond IGNORED
-    _Tp get(const int i) const { return s[i]; }
-    v_reg<_Tp, n> high() const
-    {
-        v_reg<_Tp, n> c;
-        int i;
-        for( i = 0; i < n/2; i++ )
-        {
-            c.s[i] = s[i+(n/2)];
-            c.s[i+(n/2)] = 0;
-        }
-        return c;
-    }
-
-    static v_reg<_Tp, n> zero()
-    {
-        v_reg<_Tp, n> c;
-        for( int i = 0; i < n; i++ )
-            c.s[i] = (_Tp)0;
-        return c;
-    }
-
-    static v_reg<_Tp, n> all(_Tp s)
-    {
-        v_reg<_Tp, n> c;
-        for( int i = 0; i < n; i++ )
-            c.s[i] = s;
-        return c;
-    }
-
-    template<typename _Tp2, int n2> v_reg<_Tp2, n2> reinterpret_as() const
-    {
-        size_t bytes = std::min(sizeof(_Tp2)*n2, sizeof(_Tp)*n);
-        v_reg<_Tp2, n2> c;
-        std::memcpy(&c.s[0], &s[0], bytes);
-        return c;
-    }
-
-    v_reg& operator=(const v_reg<_Tp, n> & r)
-    {
-        for( int i = 0; i < n; i++ )
-            s[i] = r.s[i];
-        return *this;
-    }
-
-    _Tp s[n];
-//! @endcond
-};
-
-/** @brief Sixteen 8-bit unsigned integer values */
-typedef v_reg<uchar, 16> v_uint8x16;
-/** @brief Sixteen 8-bit signed integer values */
-typedef v_reg<schar, 16> v_int8x16;
-/** @brief Eight 16-bit unsigned integer values */
-typedef v_reg<ushort, 8> v_uint16x8;
-/** @brief Eight 16-bit signed integer values */
-typedef v_reg<short, 8> v_int16x8;
-/** @brief Four 32-bit unsigned integer values */
-typedef v_reg<unsigned, 4> v_uint32x4;
-/** @brief Four 32-bit signed integer values */
-typedef v_reg<int, 4> v_int32x4;
-/** @brief Four 32-bit floating point values (single precision) */
-typedef v_reg<float, 4> v_float32x4;
-/** @brief Two 64-bit floating point values (double precision) */
-typedef v_reg<double, 2> v_float64x2;
-/** @brief Two 64-bit unsigned integer values */
-typedef v_reg<uint64, 2> v_uint64x2;
-/** @brief Two 64-bit signed integer values */
-typedef v_reg<int64, 2> v_int64x2;
-
-#if CV_SIMD256
-/** @brief Thirty two 8-bit unsigned integer values */
-typedef v_reg<uchar, 32> v_uint8x32;
-/** @brief Thirty two 8-bit signed integer values */
-typedef v_reg<schar, 32> v_int8x32;
-/** @brief Sixteen 16-bit unsigned integer values */
-typedef v_reg<ushort, 16> v_uint16x16;
-/** @brief Sixteen 16-bit signed integer values */
-typedef v_reg<short, 16> v_int16x16;
-/** @brief Eight 32-bit unsigned integer values */
-typedef v_reg<unsigned, 8> v_uint32x8;
-/** @brief Eight 32-bit signed integer values */
-typedef v_reg<int, 8> v_int32x8;
-/** @brief Eight 32-bit floating point values (single precision) */
-typedef v_reg<float, 8> v_float32x8;
-/** @brief Four 64-bit floating point values (double precision) */
-typedef v_reg<double, 4> v_float64x4;
-/** @brief Four 64-bit unsigned integer values */
-typedef v_reg<uint64, 4> v_uint64x4;
-/** @brief Four 64-bit signed integer values */
-typedef v_reg<int64, 4> v_int64x4;
-#endif
-
-#if CV_SIMD512
-/** @brief Sixty four 8-bit unsigned integer values */
-typedef v_reg<uchar, 64> v_uint8x64;
-/** @brief Sixty four 8-bit signed integer values */
-typedef v_reg<schar, 64> v_int8x64;
-/** @brief Thirty two 16-bit unsigned integer values */
-typedef v_reg<ushort, 32> v_uint16x32;
-/** @brief Thirty two 16-bit signed integer values */
-typedef v_reg<short, 32> v_int16x32;
-/** @brief Sixteen 32-bit unsigned integer values */
-typedef v_reg<unsigned, 16> v_uint32x16;
-/** @brief Sixteen 32-bit signed integer values */
-typedef v_reg<int, 16> v_int32x16;
-/** @brief Sixteen 32-bit floating point values (single precision) */
-typedef v_reg<float, 16> v_float32x16;
-/** @brief Eight 64-bit floating point values (double precision) */
-typedef v_reg<double, 8> v_float64x8;
-/** @brief Eight 64-bit unsigned integer values */
-typedef v_reg<uint64, 8> v_uint64x8;
-/** @brief Eight 64-bit signed integer values */
-typedef v_reg<int64, 8> v_int64x8;
-#endif
-
-enum {
-    simd128_width = 16,
-#if CV_SIMD256
-    simd256_width = 32,
-#endif
-#if CV_SIMD512
-    simd512_width = 64,
-    simdmax_width = simd512_width
-#elif CV_SIMD256
-    simdmax_width = simd256_width
-#else
-    simdmax_width = simd128_width
-#endif
-};
-
-/** @brief Add values
-
-For all types. */
-template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n> operator+(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
-template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n>& operator+=(v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
-
-/** @brief Subtract values
-
-For all types. */
-template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n> operator-(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
-template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n>& operator-=(v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
-
-/** @brief Multiply values
-
-For 16- and 32-bit integer types and floating types. */
-template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n> operator*(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
-template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n>& operator*=(v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
-
-/** @brief Divide values
-
-For floating types only. */
-template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n> operator/(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
-template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n>& operator/=(v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
-
-
-/** @brief Bitwise AND
-
-Only for integer types. */
-template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n> operator&(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
-template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n>& operator&=(v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
-
-/** @brief Bitwise OR
-
-Only for integer types. */
-template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n> operator|(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
-template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n>& operator|=(v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
-
-/** @brief Bitwise XOR
-
-Only for integer types.*/
-template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n> operator^(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
-template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n>& operator^=(v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b);
-
-/** @brief Bitwise NOT
-
-Only for integer types.*/
-template<typename _Tp, int n> CV_INLINE v_reg<_Tp, n> operator~(const v_reg<_Tp, n>& a);
-
-
-#ifndef CV_DOXYGEN
-
-#define CV__HAL_INTRIN_EXPAND_WITH_INTEGER_TYPES(macro_name, ...) \
-__CV_EXPAND(macro_name(uchar, __VA_ARGS__)) \
-__CV_EXPAND(macro_name(schar, __VA_ARGS__)) \
-__CV_EXPAND(macro_name(ushort, __VA_ARGS__)) \
-__CV_EXPAND(macro_name(short, __VA_ARGS__)) \
-__CV_EXPAND(macro_name(unsigned, __VA_ARGS__)) \
-__CV_EXPAND(macro_name(int, __VA_ARGS__)) \
-__CV_EXPAND(macro_name(uint64, __VA_ARGS__)) \
-__CV_EXPAND(macro_name(int64, __VA_ARGS__)) \
-
-#define CV__HAL_INTRIN_EXPAND_WITH_FP_TYPES(macro_name, ...) \
-__CV_EXPAND(macro_name(float, __VA_ARGS__)) \
-__CV_EXPAND(macro_name(double, __VA_ARGS__)) \
-
-#define CV__HAL_INTRIN_EXPAND_WITH_ALL_TYPES(macro_name, ...) \
-CV__HAL_INTRIN_EXPAND_WITH_INTEGER_TYPES(macro_name, __VA_ARGS__) \
-CV__HAL_INTRIN_EXPAND_WITH_FP_TYPES(macro_name, __VA_ARGS__) \
-
-#define CV__HAL_INTRIN_IMPL_BIN_OP_(_Tp, bin_op) \
-template<int n> inline \
-v_reg<_Tp, n> operator bin_op (const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b) \
-{ \
-    v_reg<_Tp, n> c; \
-    for( int i = 0; i < n; i++ ) \
-        c.s[i] = saturate_cast<_Tp>(a.s[i] bin_op b.s[i]); \
-    return c; \
-} \
-template<int n> inline \
-v_reg<_Tp, n>& operator bin_op##= (v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b) \
-{ \
-    for( int i = 0; i < n; i++ ) \
-        a.s[i] = saturate_cast<_Tp>(a.s[i] bin_op b.s[i]); \
-    return a; \
-}
-
-#define CV__HAL_INTRIN_IMPL_BIN_OP(bin_op) CV__HAL_INTRIN_EXPAND_WITH_ALL_TYPES(CV__HAL_INTRIN_IMPL_BIN_OP_, bin_op)
-
-CV__HAL_INTRIN_IMPL_BIN_OP(+)
-CV__HAL_INTRIN_IMPL_BIN_OP(-)
-CV__HAL_INTRIN_IMPL_BIN_OP(*)
-CV__HAL_INTRIN_EXPAND_WITH_FP_TYPES(CV__HAL_INTRIN_IMPL_BIN_OP_, /)
-
-#define CV__HAL_INTRIN_IMPL_BIT_OP_(_Tp, bit_op) \
-template<int n> CV_INLINE \
-v_reg<_Tp, n> operator bit_op (const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b) \
-{ \
-    v_reg<_Tp, n> c; \
-    typedef typename V_TypeTraits<_Tp>::int_type itype; \
-    for( int i = 0; i < n; i++ ) \
-        c.s[i] = V_TypeTraits<_Tp>::reinterpret_from_int((itype)(V_TypeTraits<_Tp>::reinterpret_int(a.s[i]) bit_op \
-                                                        V_TypeTraits<_Tp>::reinterpret_int(b.s[i]))); \
-    return c; \
-} \
-template<int n> CV_INLINE \
-v_reg<_Tp, n>& operator bit_op##= (v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b) \
-{ \
-    typedef typename V_TypeTraits<_Tp>::int_type itype; \
-    for( int i = 0; i < n; i++ ) \
-        a.s[i] = V_TypeTraits<_Tp>::reinterpret_from_int((itype)(V_TypeTraits<_Tp>::reinterpret_int(a.s[i]) bit_op \
-                                                        V_TypeTraits<_Tp>::reinterpret_int(b.s[i]))); \
-    return a; \
-}
-
-#define CV__HAL_INTRIN_IMPL_BIT_OP(bit_op) \
-CV__HAL_INTRIN_EXPAND_WITH_INTEGER_TYPES(CV__HAL_INTRIN_IMPL_BIT_OP_, bit_op) \
-CV__HAL_INTRIN_EXPAND_WITH_FP_TYPES(CV__HAL_INTRIN_IMPL_BIT_OP_, bit_op) /* TODO: FIXIT remove this after masks refactoring */
-
-
-CV__HAL_INTRIN_IMPL_BIT_OP(&)
-CV__HAL_INTRIN_IMPL_BIT_OP(|)
-CV__HAL_INTRIN_IMPL_BIT_OP(^)
-
-#define CV__HAL_INTRIN_IMPL_BITWISE_NOT_(_Tp, dummy) \
-template<int n> CV_INLINE \
-v_reg<_Tp, n> operator ~ (const v_reg<_Tp, n>& a) \
-{ \
-    v_reg<_Tp, n> c; \
-    for( int i = 0; i < n; i++ ) \
-        c.s[i] = V_TypeTraits<_Tp>::reinterpret_from_int(~V_TypeTraits<_Tp>::reinterpret_int(a.s[i])); \
-    return c; \
-} \
-
-CV__HAL_INTRIN_EXPAND_WITH_INTEGER_TYPES(CV__HAL_INTRIN_IMPL_BITWISE_NOT_, ~)
-
-#endif  // !CV_DOXYGEN
-
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_MATH_FUNC(func, cfunc, _Tp2) \
-template<typename _Tp, int n> inline v_reg<_Tp2, n> func(const v_reg<_Tp, n>& a) \
-{ \
-    v_reg<_Tp2, n> c; \
-    for( int i = 0; i < n; i++ ) \
-        c.s[i] = cfunc(a.s[i]); \
-    return c; \
-}
-
-/** @brief Square root of elements
-
-Only for floating point types.*/
-OPENCV_HAL_IMPL_MATH_FUNC(v_sqrt, std::sqrt, _Tp)
-
-//! @cond IGNORED
-OPENCV_HAL_IMPL_MATH_FUNC(v_sin, std::sin, _Tp)
-OPENCV_HAL_IMPL_MATH_FUNC(v_cos, std::cos, _Tp)
-OPENCV_HAL_IMPL_MATH_FUNC(v_exp, std::exp, _Tp)
-OPENCV_HAL_IMPL_MATH_FUNC(v_log, std::log, _Tp)
-//! @endcond
-
-/** @brief Absolute value of elements
-
-Only for floating point types.*/
-OPENCV_HAL_IMPL_MATH_FUNC(v_abs, (typename V_TypeTraits<_Tp>::abs_type)std::abs,
-                          typename V_TypeTraits<_Tp>::abs_type)
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_MINMAX_FUNC(func, cfunc) \
-template<typename _Tp, int n> inline v_reg<_Tp, n> func(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b) \
-{ \
-    v_reg<_Tp, n> c; \
-    for( int i = 0; i < n; i++ ) \
-        c.s[i] = cfunc(a.s[i], b.s[i]); \
-    return c; \
-}
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_REDUCE_MINMAX_FUNC(func, cfunc) \
-template<typename _Tp, int n> inline _Tp func(const v_reg<_Tp, n>& a) \
-{ \
-    _Tp c = a.s[0]; \
-    for( int i = 1; i < n; i++ ) \
-        c = cfunc(c, a.s[i]); \
-    return c; \
-}
-
-/** @brief Choose min values for each pair
-
-Scheme:
-@code
-{A1 A2 ...}
-{B1 B2 ...}
---------------
-{min(A1,B1) min(A2,B2) ...}
-@endcode
-For all types except 64-bit integer. */
-OPENCV_HAL_IMPL_MINMAX_FUNC(v_min, std::min)
-
-/** @brief Choose max values for each pair
-
-Scheme:
-@code
-{A1 A2 ...}
-{B1 B2 ...}
---------------
-{max(A1,B1) max(A2,B2) ...}
-@endcode
-For all types except 64-bit integer. */
-OPENCV_HAL_IMPL_MINMAX_FUNC(v_max, std::max)
-
-/** @brief Find one min value
-
-Scheme:
-@code
-{A1 A2 A3 ...} => min(A1,A2,A3,...)
-@endcode
-For all types except 64-bit integer and 64-bit floating point types. */
-OPENCV_HAL_IMPL_REDUCE_MINMAX_FUNC(v_reduce_min, std::min)
-
-/** @brief Find one max value
-
-Scheme:
-@code
-{A1 A2 A3 ...} => max(A1,A2,A3,...)
-@endcode
-For all types except 64-bit integer and 64-bit floating point types. */
-OPENCV_HAL_IMPL_REDUCE_MINMAX_FUNC(v_reduce_max, std::max)
-
-static const unsigned char popCountTable[] =
-{
-    0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    4, 5, 5, 6, 5, 6, 6, 7, 5, 6, 6, 7, 6, 7, 7, 8,
-};
-/** @brief Count the 1 bits in the vector lanes and return result as corresponding unsigned type
-
-Scheme:
-@code
-{A1 A2 A3 ...} => {popcount(A1), popcount(A2), popcount(A3), ...}
-@endcode
-For all integer types. */
-template<typename _Tp, int n>
-inline v_reg<typename V_TypeTraits<_Tp>::abs_type, n> v_popcount(const v_reg<_Tp, n>& a)
-{
-    v_reg<typename V_TypeTraits<_Tp>::abs_type, n> b = v_reg<typename V_TypeTraits<_Tp>::abs_type, n>::zero();
-    for (int i = 0; i < n*(int)sizeof(_Tp); i++)
-        b.s[i/sizeof(_Tp)] += popCountTable[v_reinterpret_as_u8(a).s[i]];
-    return b;
-}
-
-
-//! @cond IGNORED
-template<typename _Tp, int n>
-inline void v_minmax( const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b,
-                      v_reg<_Tp, n>& minval, v_reg<_Tp, n>& maxval )
-{
-    for( int i = 0; i < n; i++ )
-    {
-        minval.s[i] = std::min(a.s[i], b.s[i]);
-        maxval.s[i] = std::max(a.s[i], b.s[i]);
-    }
-}
-//! @endcond
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_CMP_OP(cmp_op) \
-template<typename _Tp, int n> \
-inline v_reg<_Tp, n> operator cmp_op(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b) \
-{ \
-    typedef typename V_TypeTraits<_Tp>::int_type itype; \
-    v_reg<_Tp, n> c; \
-    for( int i = 0; i < n; i++ ) \
-        c.s[i] = V_TypeTraits<_Tp>::reinterpret_from_int((itype)-(int)(a.s[i] cmp_op b.s[i])); \
-    return c; \
-}
-
-/** @brief Less-than comparison
-
-For all types except 64-bit integer values. */
-OPENCV_HAL_IMPL_CMP_OP(<)
-
-/** @brief Greater-than comparison
-
-For all types except 64-bit integer values. */
-OPENCV_HAL_IMPL_CMP_OP(>)
-
-/** @brief Less-than or equal comparison
-
-For all types except 64-bit integer values. */
-OPENCV_HAL_IMPL_CMP_OP(<=)
-
-/** @brief Greater-than or equal comparison
-
-For all types except 64-bit integer values. */
-OPENCV_HAL_IMPL_CMP_OP(>=)
-
-/** @brief Equal comparison
-
-For all types except 64-bit integer values. */
-OPENCV_HAL_IMPL_CMP_OP(==)
-
-/** @brief Not equal comparison
-
-For all types except 64-bit integer values. */
-OPENCV_HAL_IMPL_CMP_OP(!=)
-
-template<int n>
-inline v_reg<float, n> v_not_nan(const v_reg<float, n>& a)
-{
-    typedef typename V_TypeTraits<float>::int_type itype;
-    v_reg<float, n> c;
-    for (int i = 0; i < n; i++)
-        c.s[i] = V_TypeTraits<float>::reinterpret_from_int((itype)-(int)(a.s[i] == a.s[i]));
-    return c;
-}
-template<int n>
-inline v_reg<double, n> v_not_nan(const v_reg<double, n>& a)
-{
-    typedef typename V_TypeTraits<double>::int_type itype;
-    v_reg<double, n> c;
-    for (int i = 0; i < n; i++)
-        c.s[i] = V_TypeTraits<double>::reinterpret_from_int((itype)-(int)(a.s[i] == a.s[i]));
-    return c;
-}
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_ARITHM_OP(func, bin_op, cast_op, _Tp2) \
-template<typename _Tp, int n> \
-inline v_reg<_Tp2, n> func(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b) \
-{ \
-    typedef _Tp2 rtype; \
-    v_reg<rtype, n> c; \
-    for( int i = 0; i < n; i++ ) \
-        c.s[i] = cast_op(a.s[i] bin_op b.s[i]); \
-    return c; \
-}
-
-/** @brief Add values without saturation
-
-For 8- and 16-bit integer values. */
-OPENCV_HAL_IMPL_ARITHM_OP(v_add_wrap, +, (_Tp), _Tp)
-
-/** @brief Subtract values without saturation
-
-For 8- and 16-bit integer values. */
-OPENCV_HAL_IMPL_ARITHM_OP(v_sub_wrap, -, (_Tp), _Tp)
-
-/** @brief Multiply values without saturation
-
-For 8- and 16-bit integer values. */
-OPENCV_HAL_IMPL_ARITHM_OP(v_mul_wrap, *, (_Tp), _Tp)
-
-//! @cond IGNORED
-template<typename T> inline T _absdiff(T a, T b)
-{
-    return a > b ? a - b : b - a;
-}
-//! @endcond
-
-/** @brief Absolute difference
-
-Returns \f$ |a - b| \f$ converted to corresponding unsigned type.
-Example:
-@code{.cpp}
-v_int32x4 a, b; // {1, 2, 3, 4} and {4, 3, 2, 1}
-v_uint32x4 c = v_absdiff(a, b); // result is {3, 1, 1, 3}
-@endcode
-For 8-, 16-, 32-bit integer source types. */
-template<typename _Tp, int n>
-inline v_reg<typename V_TypeTraits<_Tp>::abs_type, n> v_absdiff(const v_reg<_Tp, n>& a, const v_reg<_Tp, n> & b)
-{
-    typedef typename V_TypeTraits<_Tp>::abs_type rtype;
-    v_reg<rtype, n> c;
-    const rtype mask = (rtype)(std::numeric_limits<_Tp>::is_signed ? (1 << (sizeof(rtype)*8 - 1)) : 0);
-    for( int i = 0; i < n; i++ )
-    {
-        rtype ua = a.s[i] ^ mask;
-        rtype ub = b.s[i] ^ mask;
-        c.s[i] = _absdiff(ua, ub);
-    }
-    return c;
-}
-
-/** @overload
-
-For 32-bit floating point values */
-template<int n> inline v_reg<float, n> v_absdiff(const v_reg<float, n>& a, const v_reg<float, n>& b)
-{
-    v_reg<float, n> c;
-    for( int i = 0; i < c.nlanes; i++ )
-        c.s[i] = _absdiff(a.s[i], b.s[i]);
-    return c;
-}
-
-/** @overload
-
-For 64-bit floating point values */
-template<int n> inline v_reg<double, n> v_absdiff(const v_reg<double, n>& a, const v_reg<double, n>& b)
-{
-    v_reg<double, n> c;
-    for( int i = 0; i < c.nlanes; i++ )
-        c.s[i] = _absdiff(a.s[i], b.s[i]);
-    return c;
-}
-
-/** @brief Saturating absolute difference
-
-Returns \f$ saturate(|a - b|) \f$ .
-For 8-, 16-bit signed integer source types. */
-template<typename _Tp, int n>
-inline v_reg<_Tp, n> v_absdiffs(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b)
-{
-    v_reg<_Tp, n> c;
-    for( int i = 0; i < n; i++)
-        c.s[i] = saturate_cast<_Tp>(std::abs(a.s[i] - b.s[i]));
-    return c;
-}
-
-/** @brief Inversed square root
-
-Returns \f$ 1/sqrt(a) \f$
-For floating point types only. */
-template<typename _Tp, int n>
-inline v_reg<_Tp, n> v_invsqrt(const v_reg<_Tp, n>& a)
-{
-    v_reg<_Tp, n> c;
-    for( int i = 0; i < n; i++ )
-        c.s[i] = 1.f/std::sqrt(a.s[i]);
-    return c;
-}
-
-/** @brief Magnitude
-
-Returns \f$ sqrt(a^2 + b^2) \f$
-For floating point types only. */
-template<typename _Tp, int n>
-inline v_reg<_Tp, n> v_magnitude(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b)
-{
-    v_reg<_Tp, n> c;
-    for( int i = 0; i < n; i++ )
-        c.s[i] = std::sqrt(a.s[i]*a.s[i] + b.s[i]*b.s[i]);
-    return c;
-}
-
-/** @brief Square of the magnitude
-
-Returns \f$ a^2 + b^2 \f$
-For floating point types only. */
-template<typename _Tp, int n>
-inline v_reg<_Tp, n> v_sqr_magnitude(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b)
-{
-    v_reg<_Tp, n> c;
-    for( int i = 0; i < n; i++ )
-        c.s[i] = a.s[i]*a.s[i] + b.s[i]*b.s[i];
-    return c;
-}
-
-/** @brief Multiply and add
-
- Returns \f$ a*b + c \f$
- For floating point types and signed 32bit int only. */
-template<typename _Tp, int n>
-inline v_reg<_Tp, n> v_fma(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b,
-                           const v_reg<_Tp, n>& c)
-{
-    v_reg<_Tp, n> d;
-    for( int i = 0; i < n; i++ )
-        d.s[i] = a.s[i]*b.s[i] + c.s[i];
-    return d;
-}
-
-/** @brief A synonym for v_fma */
-template<typename _Tp, int n>
-inline v_reg<_Tp, n> v_muladd(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b,
-                              const v_reg<_Tp, n>& c)
-{
-    return v_fma(a, b, c);
-}
-
-/** @brief Dot product of elements
-
-Multiply values in two registers and sum adjacent result pairs.
-
-Scheme:
-@code
-  {A1 A2 ...} // 16-bit
-x {B1 B2 ...} // 16-bit
--------------
-{A1B1+A2B2 ...} // 32-bit
-
-@endcode
-*/
-template<typename _Tp, int n> inline v_reg<typename V_TypeTraits<_Tp>::w_type, n/2>
-v_dotprod(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b)
-{
-    typedef typename V_TypeTraits<_Tp>::w_type w_type;
-    v_reg<w_type, n/2> c;
-    for( int i = 0; i < (n/2); i++ )
-        c.s[i] = (w_type)a.s[i*2]*b.s[i*2] + (w_type)a.s[i*2+1]*b.s[i*2+1];
-    return c;
-}
-
-/** @brief Dot product of elements
-
-Same as cv::v_dotprod, but add a third element to the sum of adjacent pairs.
-Scheme:
-@code
-  {A1 A2 ...} // 16-bit
-x {B1 B2 ...} // 16-bit
--------------
-  {A1B1+A2B2+C1 ...} // 32-bit
-@endcode
-*/
-template<typename _Tp, int n> inline v_reg<typename V_TypeTraits<_Tp>::w_type, n/2>
-v_dotprod(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b,
-          const v_reg<typename V_TypeTraits<_Tp>::w_type, n / 2>& c)
-{
-    typedef typename V_TypeTraits<_Tp>::w_type w_type;
-    v_reg<w_type, n/2> s;
-    for( int i = 0; i < (n/2); i++ )
-        s.s[i] = (w_type)a.s[i*2]*b.s[i*2] + (w_type)a.s[i*2+1]*b.s[i*2+1] + c.s[i];
-    return s;
-}
-
-/** @brief Fast Dot product of elements
-
-Same as cv::v_dotprod, but it may perform unorder sum between result pairs in some platforms,
-this intrinsic can be used if the sum among all lanes is only matters
-and also it should be yielding better performance on the affected platforms.
-
-*/
-template<typename _Tp, int n> inline v_reg<typename V_TypeTraits<_Tp>::w_type, n/2>
-v_dotprod_fast(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b)
-{ return v_dotprod(a, b); }
-
-/** @brief Fast Dot product of elements
-
-Same as cv::v_dotprod_fast, but add a third element to the sum of adjacent pairs.
-*/
-template<typename _Tp, int n> inline v_reg<typename V_TypeTraits<_Tp>::w_type, n/2>
-v_dotprod_fast(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b,
-               const v_reg<typename V_TypeTraits<_Tp>::w_type, n / 2>& c)
-{ return v_dotprod(a, b, c); }
-
-/** @brief Dot product of elements and expand
-
-Multiply values in two registers and expand the sum of adjacent result pairs.
-
-Scheme:
-@code
-  {A1 A2 A3 A4 ...} // 8-bit
-x {B1 B2 B3 B4 ...} // 8-bit
--------------
-  {A1B1+A2B2+A3B3+A4B4 ...} // 32-bit
-
-@endcode
-*/
-template<typename _Tp, int n> inline v_reg<typename V_TypeTraits<_Tp>::q_type, n/4>
-v_dotprod_expand(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b)
-{
-    typedef typename V_TypeTraits<_Tp>::q_type q_type;
-    v_reg<q_type, n/4> s;
-    for( int i = 0; i < (n/4); i++ )
-        s.s[i] = (q_type)a.s[i*4    ]*b.s[i*4    ] + (q_type)a.s[i*4 + 1]*b.s[i*4 + 1] +
-                 (q_type)a.s[i*4 + 2]*b.s[i*4 + 2] + (q_type)a.s[i*4 + 3]*b.s[i*4 + 3];
-    return s;
-}
-
-/** @brief Dot product of elements
-
-Same as cv::v_dotprod_expand, but add a third element to the sum of adjacent pairs.
-Scheme:
-@code
-  {A1 A2 A3 A4 ...} // 8-bit
-x {B1 B2 B3 B4 ...} // 8-bit
--------------
-  {A1B1+A2B2+A3B3+A4B4+C1 ...} // 32-bit
-@endcode
-*/
-template<typename _Tp, int n> inline v_reg<typename V_TypeTraits<_Tp>::q_type, n/4>
-v_dotprod_expand(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b,
-                 const v_reg<typename V_TypeTraits<_Tp>::q_type, n / 4>& c)
-{
-    typedef typename V_TypeTraits<_Tp>::q_type q_type;
-    v_reg<q_type, n/4> s;
-    for( int i = 0; i < (n/4); i++ )
-        s.s[i] = (q_type)a.s[i*4    ]*b.s[i*4    ] + (q_type)a.s[i*4 + 1]*b.s[i*4 + 1] +
-                 (q_type)a.s[i*4 + 2]*b.s[i*4 + 2] + (q_type)a.s[i*4 + 3]*b.s[i*4 + 3] + c.s[i];
-    return s;
-}
-
-/** @brief Fast Dot product of elements and expand
-
-Multiply values in two registers and expand the sum of adjacent result pairs.
-
-Same as cv::v_dotprod_expand, but it may perform unorder sum between result pairs in some platforms,
-this intrinsic can be used if the sum among all lanes is only matters
-and also it should be yielding better performance on the affected platforms.
-
-*/
-template<typename _Tp, int n> inline v_reg<typename V_TypeTraits<_Tp>::q_type, n/4>
-v_dotprod_expand_fast(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b)
-{ return v_dotprod_expand(a, b); }
-
-/** @brief Fast Dot product of elements
-
-Same as cv::v_dotprod_expand_fast, but add a third element to the sum of adjacent pairs.
-*/
-template<typename _Tp, int n> inline v_reg<typename V_TypeTraits<_Tp>::q_type, n/4>
-v_dotprod_expand_fast(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b,
-                      const v_reg<typename V_TypeTraits<_Tp>::q_type, n / 4>& c)
-{ return v_dotprod_expand(a, b, c); }
-
-/** @brief Multiply and expand
-
-Multiply values two registers and store results in two registers with wider pack type.
-Scheme:
-@code
-  {A B C D} // 32-bit
-x {E F G H} // 32-bit
----------------
-{AE BF}         // 64-bit
-        {CG DH} // 64-bit
-@endcode
-Example:
-@code{.cpp}
-v_uint32x4 a, b; // {1,2,3,4} and {2,2,2,2}
-v_uint64x2 c, d; // results
-v_mul_expand(a, b, c, d); // c, d = {2,4}, {6, 8}
-@endcode
-Implemented only for 16- and unsigned 32-bit source types (v_int16x8, v_uint16x8, v_uint32x4).
-*/
-template<typename _Tp, int n> inline void v_mul_expand(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b,
-                                                       v_reg<typename V_TypeTraits<_Tp>::w_type, n/2>& c,
-                                                       v_reg<typename V_TypeTraits<_Tp>::w_type, n/2>& d)
-{
-    typedef typename V_TypeTraits<_Tp>::w_type w_type;
-    for( int i = 0; i < (n/2); i++ )
-    {
-        c.s[i] = (w_type)a.s[i]*b.s[i];
-        d.s[i] = (w_type)a.s[i+(n/2)]*b.s[i+(n/2)];
-    }
-}
-
-/** @brief Multiply and extract high part
-
-Multiply values two registers and store high part of the results.
-Implemented only for 16-bit source types (v_int16x8, v_uint16x8). Returns \f$ a*b >> 16 \f$
-*/
-template<typename _Tp, int n> inline v_reg<_Tp, n> v_mul_hi(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b)
-{
-    typedef typename V_TypeTraits<_Tp>::w_type w_type;
-    v_reg<_Tp, n> c;
-    for (int i = 0; i < n; i++)
-        c.s[i] = (_Tp)(((w_type)a.s[i] * b.s[i]) >> sizeof(_Tp)*8);
-    return c;
-}
-
-//! @cond IGNORED
-template<typename _Tp, int n> inline void v_hsum(const v_reg<_Tp, n>& a,
-                                                 v_reg<typename V_TypeTraits<_Tp>::w_type, n/2>& c)
-{
-    typedef typename V_TypeTraits<_Tp>::w_type w_type;
-    for( int i = 0; i < (n/2); i++ )
-    {
-        c.s[i] = (w_type)a.s[i*2] + a.s[i*2+1];
-    }
-}
-//! @endcond
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_SHIFT_OP(shift_op) \
-template<typename _Tp, int n> inline v_reg<_Tp, n> operator shift_op(const v_reg<_Tp, n>& a, int imm) \
-{ \
-    v_reg<_Tp, n> c; \
-    for( int i = 0; i < n; i++ ) \
-        c.s[i] = (_Tp)(a.s[i] shift_op imm); \
-    return c; \
-}
-
-/** @brief Bitwise shift left
-
-For 16-, 32- and 64-bit integer values. */
-OPENCV_HAL_IMPL_SHIFT_OP(<< )
-
-/** @brief Bitwise shift right
-
-For 16-, 32- and 64-bit integer values. */
-OPENCV_HAL_IMPL_SHIFT_OP(>> )
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_ROTATE_SHIFT_OP(suffix,opA,opB) \
-template<int imm, typename _Tp, int n> inline v_reg<_Tp, n> v_rotate_##suffix(const v_reg<_Tp, n>& a) \
-{ \
-    v_reg<_Tp, n> b; \
-    for (int i = 0; i < n; i++) \
-    { \
-        int sIndex = i opA imm; \
-        if (0 <= sIndex && sIndex < n) \
-        { \
-            b.s[i] = a.s[sIndex]; \
-        } \
-        else \
-        { \
-            b.s[i] = 0; \
-        } \
-    } \
-    return b; \
-} \
-template<int imm, typename _Tp, int n> inline v_reg<_Tp, n> v_rotate_##suffix(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b) \
-{ \
-    v_reg<_Tp, n> c; \
-    for (int i = 0; i < n; i++) \
-    { \
-        int aIndex = i opA imm; \
-        int bIndex = i opA imm opB n; \
-        if (0 <= bIndex && bIndex < n) \
-        { \
-            c.s[i] = b.s[bIndex]; \
-        } \
-        else if (0 <= aIndex && aIndex < n) \
-        { \
-            c.s[i] = a.s[aIndex]; \
-        } \
-        else \
-        { \
-            c.s[i] = 0; \
-        } \
-    } \
-    return c; \
-}
-
-/** @brief Element shift left among vector
-
-For all type */
-OPENCV_HAL_IMPL_ROTATE_SHIFT_OP(left,  -, +)
-
-/** @brief Element shift right among vector
-
-For all type */
-OPENCV_HAL_IMPL_ROTATE_SHIFT_OP(right, +, -)
-
-/** @brief Sum packed values
-
-Scheme:
-@code
-{A1 A2 A3 ...} => sum{A1,A2,A3,...}
-@endcode
-*/
-template<typename _Tp, int n> inline typename V_TypeTraits<_Tp>::sum_type v_reduce_sum(const v_reg<_Tp, n>& a)
-{
-    typename V_TypeTraits<_Tp>::sum_type c = a.s[0];
-    for( int i = 1; i < n; i++ )
-        c += a.s[i];
-    return c;
-}
-
-/** @brief Sums all elements of each input vector, returns the vector of sums
-
- Scheme:
- @code
- result[0] = a[0] + a[1] + a[2] + a[3]
- result[1] = b[0] + b[1] + b[2] + b[3]
- result[2] = c[0] + c[1] + c[2] + c[3]
- result[3] = d[0] + d[1] + d[2] + d[3]
- @endcode
-*/
-template<int n> inline v_reg<float, n> v_reduce_sum4(const v_reg<float, n>& a, const v_reg<float, n>& b,
-    const v_reg<float, n>& c, const v_reg<float, n>& d)
-{
-    v_reg<float, n> r;
-    for(int i = 0; i < (n/4); i++)
-    {
-        r.s[i*4 + 0] = a.s[i*4 + 0] + a.s[i*4 + 1] + a.s[i*4 + 2] + a.s[i*4 + 3];
-        r.s[i*4 + 1] = b.s[i*4 + 0] + b.s[i*4 + 1] + b.s[i*4 + 2] + b.s[i*4 + 3];
-        r.s[i*4 + 2] = c.s[i*4 + 0] + c.s[i*4 + 1] + c.s[i*4 + 2] + c.s[i*4 + 3];
-        r.s[i*4 + 3] = d.s[i*4 + 0] + d.s[i*4 + 1] + d.s[i*4 + 2] + d.s[i*4 + 3];
-    }
-    return r;
-}
-
-/** @brief Sum absolute differences of values
-
-Scheme:
-@code
-{A1 A2 A3 ...} {B1 B2 B3 ...} => sum{ABS(A1-B1),abs(A2-B2),abs(A3-B3),...}
-@endcode
-For all types except 64-bit types.*/
-template<typename _Tp, int n> inline typename V_TypeTraits< typename V_TypeTraits<_Tp>::abs_type >::sum_type v_reduce_sad(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b)
-{
-    typename V_TypeTraits< typename V_TypeTraits<_Tp>::abs_type >::sum_type c = _absdiff(a.s[0], b.s[0]);
-    for (int i = 1; i < n; i++)
-        c += _absdiff(a.s[i], b.s[i]);
-    return c;
-}
-
-/** @brief Get negative values mask
-@deprecated v_signmask depends on a lane count heavily and therefore isn't universal enough
-
-Returned value is a bit mask with bits set to 1 on places corresponding to negative packed values indexes.
-Example:
-@code{.cpp}
-v_int32x4 r; // set to {-1, -1, 1, 1}
-int mask = v_signmask(r); // mask = 3 <== 00000000 00000000 00000000 00000011
-@endcode
-*/
-template<typename _Tp, int n> inline int v_signmask(const v_reg<_Tp, n>& a)
-{
-    int mask = 0;
-    for( int i = 0; i < n; i++ )
-        mask |= (V_TypeTraits<_Tp>::reinterpret_int(a.s[i]) < 0) << i;
-    return mask;
-}
-
-/** @brief Get first negative lane index
-
-Returned value is an index of first negative lane (undefined for input of all positive values)
-Example:
-@code{.cpp}
-v_int32x4 r; // set to {0, 0, -1, -1}
-int idx = v_heading_zeros(r); // idx = 2
-@endcode
-*/
-template <typename _Tp, int n> inline int v_scan_forward(const v_reg<_Tp, n>& a)
-{
-    for (int i = 0; i < n; i++)
-        if(V_TypeTraits<_Tp>::reinterpret_int(a.s[i]) < 0)
-            return i;
-    return 0;
-}
-
-/** @brief Check if all packed values are less than zero
-
-Unsigned values will be casted to signed: `uchar 254 => char -2`.
-*/
-template<typename _Tp, int n> inline bool v_check_all(const v_reg<_Tp, n>& a)
-{
-    for( int i = 0; i < n; i++ )
-        if( V_TypeTraits<_Tp>::reinterpret_int(a.s[i]) >= 0 )
-            return false;
-    return true;
-}
-
-/** @brief Check if any of packed values is less than zero
-
-Unsigned values will be casted to signed: `uchar 254 => char -2`.
-*/
-template<typename _Tp, int n> inline bool v_check_any(const v_reg<_Tp, n>& a)
-{
-    for( int i = 0; i < n; i++ )
-        if( V_TypeTraits<_Tp>::reinterpret_int(a.s[i]) < 0 )
-            return true;
-    return false;
-}
-
-/** @brief Per-element select (blend operation)
-
-Return value will be built by combining values _a_ and _b_ using the following scheme:
-    result[i] = mask[i] ? a[i] : b[i];
-
-@note: _mask_ element values are restricted to these values:
-- 0: select element from _b_
-- 0xff/0xffff/etc: select element from _a_
-(fully compatible with bitwise-based operator)
-*/
-template<typename _Tp, int n> inline v_reg<_Tp, n> v_select(const v_reg<_Tp, n>& mask,
-                                                           const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b)
-{
-    typedef V_TypeTraits<_Tp> Traits;
-    typedef typename Traits::int_type int_type;
-    v_reg<_Tp, n> c;
-    for( int i = 0; i < n; i++ )
-    {
-        int_type m = Traits::reinterpret_int(mask.s[i]);
-        CV_DbgAssert(m == 0 || m == (~(int_type)0));  // restrict mask values: 0 or 0xff/0xffff/etc
-        c.s[i] = m ? a.s[i] : b.s[i];
-    }
-    return c;
-}
-
-/** @brief Expand values to the wider pack type
-
-Copy contents of register to two registers with 2x wider pack type.
-Scheme:
-@code
- int32x4     int64x2 int64x2
-{A B C D} ==> {A B} , {C D}
-@endcode */
-template<typename _Tp, int n> inline void v_expand(const v_reg<_Tp, n>& a,
-                            v_reg<typename V_TypeTraits<_Tp>::w_type, n/2>& b0,
-                            v_reg<typename V_TypeTraits<_Tp>::w_type, n/2>& b1)
-{
-    for( int i = 0; i < (n/2); i++ )
-    {
-        b0.s[i] = a.s[i];
-        b1.s[i] = a.s[i+(n/2)];
-    }
-}
-
-/** @brief Expand lower values to the wider pack type
-
-Same as cv::v_expand, but return lower half of the vector.
-
-Scheme:
-@code
- int32x4     int64x2
-{A B C D} ==> {A B}
-@endcode */
-template<typename _Tp, int n>
-inline v_reg<typename V_TypeTraits<_Tp>::w_type, n/2>
-v_expand_low(const v_reg<_Tp, n>& a)
-{
-    v_reg<typename V_TypeTraits<_Tp>::w_type, n/2> b;
-    for( int i = 0; i < (n/2); i++ )
-        b.s[i] = a.s[i];
-    return b;
-}
-
-/** @brief Expand higher values to the wider pack type
-
-Same as cv::v_expand_low, but expand higher half of the vector instead.
-
-Scheme:
-@code
- int32x4     int64x2
-{A B C D} ==> {C D}
-@endcode */
-template<typename _Tp, int n>
-inline v_reg<typename V_TypeTraits<_Tp>::w_type, n/2>
-v_expand_high(const v_reg<_Tp, n>& a)
-{
-    v_reg<typename V_TypeTraits<_Tp>::w_type, n/2> b;
-    for( int i = 0; i < (n/2); i++ )
-        b.s[i] = a.s[i+(n/2)];
-    return b;
-}
-
-//! @cond IGNORED
-template<typename _Tp, int n> inline v_reg<typename V_TypeTraits<_Tp>::int_type, n>
-    v_reinterpret_as_int(const v_reg<_Tp, n>& a)
-{
-    v_reg<typename V_TypeTraits<_Tp>::int_type, n> c;
-    for( int i = 0; i < n; i++ )
-        c.s[i] = V_TypeTraits<_Tp>::reinterpret_int(a.s[i]);
-    return c;
-}
-
-template<typename _Tp, int n> inline v_reg<typename V_TypeTraits<_Tp>::uint_type, n>
-    v_reinterpret_as_uint(const v_reg<_Tp, n>& a)
-{
-    v_reg<typename V_TypeTraits<_Tp>::uint_type, n> c;
-    for( int i = 0; i < n; i++ )
-        c.s[i] = V_TypeTraits<_Tp>::reinterpret_uint(a.s[i]);
-    return c;
-}
-//! @endcond
-
-/** @brief Interleave two vectors
-
-Scheme:
-@code
-  {A1 A2 A3 A4}
-  {B1 B2 B3 B4}
----------------
-  {A1 B1 A2 B2} and {A3 B3 A4 B4}
-@endcode
-For all types except 64-bit.
-*/
-template<typename _Tp, int n> inline void v_zip( const v_reg<_Tp, n>& a0, const v_reg<_Tp, n>& a1,
-                                               v_reg<_Tp, n>& b0, v_reg<_Tp, n>& b1 )
-{
-    int i;
-    for( i = 0; i < n/2; i++ )
-    {
-        b0.s[i*2] = a0.s[i];
-        b0.s[i*2+1] = a1.s[i];
-    }
-    for( ; i < n; i++ )
-    {
-        b1.s[i*2-n] = a0.s[i];
-        b1.s[i*2-n+1] = a1.s[i];
-    }
-}
-
-/** @brief Load register contents from memory
-
-@param ptr pointer to memory block with data
-@return register object
-
-@note Returned type will be detected from passed pointer type, for example uchar ==> cv::v_uint8x16, int ==> cv::v_int32x4, etc.
-
-@note Use vx_load version to get maximum available register length result
-
-@note Alignment requirement:
-if CV_STRONG_ALIGNMENT=1 then passed pointer must be aligned (`sizeof(lane type)` should be enough).
-Do not cast pointer types without runtime check for pointer alignment (like `uchar*` => `int*`).
- */
-template<typename _Tp>
-inline v_reg<_Tp, simd128_width / sizeof(_Tp)> v_load(const _Tp* ptr)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    return v_reg<_Tp, simd128_width / sizeof(_Tp)>(ptr);
-}
-
-#if CV_SIMD256
-/** @brief Load 256-bit length register contents from memory
-
-@param ptr pointer to memory block with data
-@return register object
-
-@note Returned type will be detected from passed pointer type, for example uchar ==> cv::v_uint8x32, int ==> cv::v_int32x8, etc.
-
-@note Check CV_SIMD256 preprocessor definition prior to use.
-Use vx_load version to get maximum available register length result
-
-@note Alignment requirement:
-if CV_STRONG_ALIGNMENT=1 then passed pointer must be aligned (`sizeof(lane type)` should be enough).
-Do not cast pointer types without runtime check for pointer alignment (like `uchar*` => `int*`).
- */
-template<typename _Tp>
-inline v_reg<_Tp, simd256_width / sizeof(_Tp)> v256_load(const _Tp* ptr)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    return v_reg<_Tp, simd256_width / sizeof(_Tp)>(ptr);
-}
-#endif
-
-#if CV_SIMD512
-/** @brief Load 512-bit length register contents from memory
-
-@param ptr pointer to memory block with data
-@return register object
-
-@note Returned type will be detected from passed pointer type, for example uchar ==> cv::v_uint8x64, int ==> cv::v_int32x16, etc.
-
-@note Check CV_SIMD512 preprocessor definition prior to use.
-Use vx_load version to get maximum available register length result
-
-@note Alignment requirement:
-if CV_STRONG_ALIGNMENT=1 then passed pointer must be aligned (`sizeof(lane type)` should be enough).
-Do not cast pointer types without runtime check for pointer alignment (like `uchar*` => `int*`).
- */
-template<typename _Tp>
-inline v_reg<_Tp, simd512_width / sizeof(_Tp)> v512_load(const _Tp* ptr)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    return v_reg<_Tp, simd512_width / sizeof(_Tp)>(ptr);
-}
-#endif
-
-/** @brief Load register contents from memory (aligned)
-
-similar to cv::v_load, but source memory block should be aligned (to 16-byte boundary in case of SIMD128, 32-byte - SIMD256, etc)
-
-@note Use vx_load_aligned version to get maximum available register length result
-*/
-template<typename _Tp>
-inline v_reg<_Tp, simd128_width / sizeof(_Tp)> v_load_aligned(const _Tp* ptr)
-{
-    CV_Assert(isAligned<sizeof(v_reg<_Tp, simd128_width / sizeof(_Tp)>)>(ptr));
-    return v_reg<_Tp, simd128_width / sizeof(_Tp)>(ptr);
-}
-
-#if CV_SIMD256
-/** @brief Load register contents from memory (aligned)
-
-similar to cv::v256_load, but source memory block should be aligned (to 32-byte boundary in case of SIMD256, 64-byte - SIMD512, etc)
-
-@note Check CV_SIMD256 preprocessor definition prior to use.
-Use vx_load_aligned version to get maximum available register length result
-*/
-template<typename _Tp>
-inline v_reg<_Tp, simd256_width / sizeof(_Tp)> v256_load_aligned(const _Tp* ptr)
-{
-    CV_Assert(isAligned<sizeof(v_reg<_Tp, simd256_width / sizeof(_Tp)>)>(ptr));
-    return v_reg<_Tp, simd256_width / sizeof(_Tp)>(ptr);
-}
-#endif
-
-#if CV_SIMD512
-/** @brief Load register contents from memory (aligned)
-
-similar to cv::v512_load, but source memory block should be aligned (to 64-byte boundary in case of SIMD512, etc)
-
-@note Check CV_SIMD512 preprocessor definition prior to use.
-Use vx_load_aligned version to get maximum available register length result
-*/
-template<typename _Tp>
-inline v_reg<_Tp, simd512_width / sizeof(_Tp)> v512_load_aligned(const _Tp* ptr)
-{
-    CV_Assert(isAligned<sizeof(v_reg<_Tp, simd512_width / sizeof(_Tp)>)>(ptr));
-    return v_reg<_Tp, simd512_width / sizeof(_Tp)>(ptr);
-}
-#endif
-
-/** @brief Load 64-bits of data to lower part (high part is undefined).
-
-@param ptr memory block containing data for first half (0..n/2)
-
-@code{.cpp}
-int lo[2] = { 1, 2 };
-v_int32x4 r = v_load_low(lo);
-@endcode
-
-@note Use vx_load_low version to get maximum available register length result
-*/
-template<typename _Tp>
-inline v_reg<_Tp, simd128_width / sizeof(_Tp)> v_load_low(const _Tp* ptr)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    v_reg<_Tp, simd128_width / sizeof(_Tp)> c;
-    for( int i = 0; i < c.nlanes/2; i++ )
-    {
-        c.s[i] = ptr[i];
-    }
-    return c;
-}
-
-#if CV_SIMD256
-/** @brief Load 128-bits of data to lower part (high part is undefined).
-
-@param ptr memory block containing data for first half (0..n/2)
-
-@code{.cpp}
-int lo[4] = { 1, 2, 3, 4 };
-v_int32x8 r = v256_load_low(lo);
-@endcode
-
-@note Check CV_SIMD256 preprocessor definition prior to use.
-Use vx_load_low version to get maximum available register length result
-*/
-template<typename _Tp>
-inline v_reg<_Tp, simd256_width / sizeof(_Tp)> v256_load_low(const _Tp* ptr)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    v_reg<_Tp, simd256_width / sizeof(_Tp)> c;
-    for (int i = 0; i < c.nlanes / 2; i++)
-    {
-        c.s[i] = ptr[i];
-    }
-    return c;
-}
-#endif
-
-#if CV_SIMD512
-/** @brief Load 256-bits of data to lower part (high part is undefined).
-
-@param ptr memory block containing data for first half (0..n/2)
-
-@code{.cpp}
-int lo[8] = { 1, 2, 3, 4, 5, 6, 7, 8 };
-v_int32x16 r = v512_load_low(lo);
-@endcode
-
-@note Check CV_SIMD512 preprocessor definition prior to use.
-Use vx_load_low version to get maximum available register length result
-*/
-template<typename _Tp>
-inline v_reg<_Tp, simd512_width / sizeof(_Tp)> v512_load_low(const _Tp* ptr)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    v_reg<_Tp, simd512_width / sizeof(_Tp)> c;
-    for (int i = 0; i < c.nlanes / 2; i++)
-    {
-        c.s[i] = ptr[i];
-    }
-    return c;
-}
-#endif
-
-/** @brief Load register contents from two memory blocks
-
-@param loptr memory block containing data for first half (0..n/2)
-@param hiptr memory block containing data for second half (n/2..n)
-
-@code{.cpp}
-int lo[2] = { 1, 2 }, hi[2] = { 3, 4 };
-v_int32x4 r = v_load_halves(lo, hi);
-@endcode
-
-@note Use vx_load_halves version to get maximum available register length result
-*/
-template<typename _Tp>
-inline v_reg<_Tp, simd128_width / sizeof(_Tp)> v_load_halves(const _Tp* loptr, const _Tp* hiptr)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(loptr));
-    CV_Assert(isAligned<sizeof(_Tp)>(hiptr));
-#endif
-    v_reg<_Tp, simd128_width / sizeof(_Tp)> c;
-    for( int i = 0; i < c.nlanes/2; i++ )
-    {
-        c.s[i] = loptr[i];
-        c.s[i+c.nlanes/2] = hiptr[i];
-    }
-    return c;
-}
-
-#if CV_SIMD256
-/** @brief Load register contents from two memory blocks
-
-@param loptr memory block containing data for first half (0..n/2)
-@param hiptr memory block containing data for second half (n/2..n)
-
-@code{.cpp}
-int lo[4] = { 1, 2, 3, 4 }, hi[4] = { 5, 6, 7, 8 };
-v_int32x8 r = v256_load_halves(lo, hi);
-@endcode
-
-@note Check CV_SIMD256 preprocessor definition prior to use.
-Use vx_load_halves version to get maximum available register length result
-*/
-template<typename _Tp>
-inline v_reg<_Tp, simd256_width / sizeof(_Tp)> v256_load_halves(const _Tp* loptr, const _Tp* hiptr)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(loptr));
-    CV_Assert(isAligned<sizeof(_Tp)>(hiptr));
-#endif
-    v_reg<_Tp, simd256_width / sizeof(_Tp)> c;
-    for (int i = 0; i < c.nlanes / 2; i++)
-    {
-        c.s[i] = loptr[i];
-        c.s[i + c.nlanes / 2] = hiptr[i];
-    }
-    return c;
-}
-#endif
-
-#if CV_SIMD512
-/** @brief Load register contents from two memory blocks
-
-@param loptr memory block containing data for first half (0..n/2)
-@param hiptr memory block containing data for second half (n/2..n)
-
-@code{.cpp}
-int lo[4] = { 1, 2, 3, 4, 5, 6, 7, 8 }, hi[4] = { 9, 10, 11, 12, 13, 14, 15, 16 };
-v_int32x16 r = v512_load_halves(lo, hi);
-@endcode
-
-@note Check CV_SIMD512 preprocessor definition prior to use.
-Use vx_load_halves version to get maximum available register length result
-*/
-template<typename _Tp>
-inline v_reg<_Tp, simd512_width / sizeof(_Tp)> v512_load_halves(const _Tp* loptr, const _Tp* hiptr)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(loptr));
-    CV_Assert(isAligned<sizeof(_Tp)>(hiptr));
-#endif
-    v_reg<_Tp, simd512_width / sizeof(_Tp)> c;
-    for (int i = 0; i < c.nlanes / 2; i++)
-    {
-        c.s[i] = loptr[i];
-        c.s[i + c.nlanes / 2] = hiptr[i];
-    }
-    return c;
-}
-#endif
-
-/** @brief Load register contents from memory with double expand
-
-Same as cv::v_load, but result pack type will be 2x wider than memory type.
-
-@code{.cpp}
-short buf[4] = {1, 2, 3, 4}; // type is int16
-v_int32x4 r = v_load_expand(buf); // r = {1, 2, 3, 4} - type is int32
-@endcode
-For 8-, 16-, 32-bit integer source types.
-
-@note Use vx_load_expand version to get maximum available register length result
-*/
-template<typename _Tp>
-inline v_reg<typename V_TypeTraits<_Tp>::w_type, simd128_width / sizeof(typename V_TypeTraits<_Tp>::w_type)>
-v_load_expand(const _Tp* ptr)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    typedef typename V_TypeTraits<_Tp>::w_type w_type;
-    v_reg<w_type, simd128_width / sizeof(w_type)> c;
-    for( int i = 0; i < c.nlanes; i++ )
-    {
-        c.s[i] = ptr[i];
-    }
-    return c;
-}
-
-#if CV_SIMD256
-/** @brief Load register contents from memory with double expand
-
-Same as cv::v256_load, but result pack type will be 2x wider than memory type.
-
-@code{.cpp}
-short buf[8] = {1, 2, 3, 4, 5, 6, 7, 8}; // type is int16
-v_int32x8 r = v256_load_expand(buf); // r = {1, 2, 3, 4, 5, 6, 7, 8} - type is int32
-@endcode
-For 8-, 16-, 32-bit integer source types.
-
-@note Check CV_SIMD256 preprocessor definition prior to use.
-Use vx_load_expand version to get maximum available register length result
-*/
-template<typename _Tp>
-inline v_reg<typename V_TypeTraits<_Tp>::w_type, simd256_width / sizeof(typename V_TypeTraits<_Tp>::w_type)>
-v256_load_expand(const _Tp* ptr)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    typedef typename V_TypeTraits<_Tp>::w_type w_type;
-    v_reg<w_type, simd256_width / sizeof(w_type)> c;
-    for (int i = 0; i < c.nlanes; i++)
-    {
-        c.s[i] = ptr[i];
-    }
-    return c;
-}
-#endif
-
-#if CV_SIMD512
-/** @brief Load register contents from memory with double expand
-
-Same as cv::v512_load, but result pack type will be 2x wider than memory type.
-
-@code{.cpp}
-short buf[8] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16}; // type is int16
-v_int32x16 r = v512_load_expand(buf); // r = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16} - type is int32
-@endcode
-For 8-, 16-, 32-bit integer source types.
-
-@note Check CV_SIMD512 preprocessor definition prior to use.
-Use vx_load_expand version to get maximum available register length result
-*/
-template<typename _Tp>
-inline v_reg<typename V_TypeTraits<_Tp>::w_type, simd512_width / sizeof(typename V_TypeTraits<_Tp>::w_type)>
-v512_load_expand(const _Tp* ptr)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    typedef typename V_TypeTraits<_Tp>::w_type w_type;
-    v_reg<w_type, simd512_width / sizeof(w_type)> c;
-    for (int i = 0; i < c.nlanes; i++)
-    {
-        c.s[i] = ptr[i];
-    }
-    return c;
-}
-#endif
-
-/** @brief Load register contents from memory with quad expand
-
-Same as cv::v_load_expand, but result type is 4 times wider than source.
-@code{.cpp}
-char buf[4] = {1, 2, 3, 4}; // type is int8
-v_int32x4 r = v_load_expand_q(buf); // r = {1, 2, 3, 4} - type is int32
-@endcode
-For 8-bit integer source types.
-
-@note Use vx_load_expand_q version to get maximum available register length result
-*/
-template<typename _Tp>
-inline v_reg<typename V_TypeTraits<_Tp>::q_type, simd128_width / sizeof(typename V_TypeTraits<_Tp>::q_type)>
-v_load_expand_q(const _Tp* ptr)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    typedef typename V_TypeTraits<_Tp>::q_type q_type;
-    v_reg<q_type, simd128_width / sizeof(q_type)> c;
-    for( int i = 0; i < c.nlanes; i++ )
-    {
-        c.s[i] = ptr[i];
-    }
-    return c;
-}
-
-#if CV_SIMD256
-/** @brief Load register contents from memory with quad expand
-
-Same as cv::v256_load_expand, but result type is 4 times wider than source.
-@code{.cpp}
-char buf[8] = {1, 2, 3, 4, 5, 6, 7, 8}; // type is int8
-v_int32x8 r = v256_load_expand_q(buf); // r = {1, 2, 3, 4, 5, 6, 7, 8} - type is int32
-@endcode
-For 8-bit integer source types.
-
-@note Check CV_SIMD256 preprocessor definition prior to use.
-Use vx_load_expand_q version to get maximum available register length result
-*/
-template<typename _Tp>
-inline v_reg<typename V_TypeTraits<_Tp>::q_type, simd256_width / sizeof(typename V_TypeTraits<_Tp>::q_type)>
-v256_load_expand_q(const _Tp* ptr)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    typedef typename V_TypeTraits<_Tp>::q_type q_type;
-    v_reg<q_type, simd256_width / sizeof(q_type)> c;
-    for (int i = 0; i < c.nlanes; i++)
-    {
-        c.s[i] = ptr[i];
-    }
-    return c;
-}
-#endif
-
-#if CV_SIMD512
-/** @brief Load register contents from memory with quad expand
-
-Same as cv::v512_load_expand, but result type is 4 times wider than source.
-@code{.cpp}
-char buf[16] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16}; // type is int8
-v_int32x16 r = v512_load_expand_q(buf); // r = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16} - type is int32
-@endcode
-For 8-bit integer source types.
-
-@note Check CV_SIMD512 preprocessor definition prior to use.
-Use vx_load_expand_q version to get maximum available register length result
-*/
-template<typename _Tp>
-inline v_reg<typename V_TypeTraits<_Tp>::q_type, simd512_width / sizeof(typename V_TypeTraits<_Tp>::q_type)>
-v512_load_expand_q(const _Tp* ptr)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    typedef typename V_TypeTraits<_Tp>::q_type q_type;
-    v_reg<q_type, simd512_width / sizeof(q_type)> c;
-    for (int i = 0; i < c.nlanes; i++)
-    {
-        c.s[i] = ptr[i];
-    }
-    return c;
-}
-#endif
-
-/** @brief Load and deinterleave (2 channels)
-
-Load data from memory deinterleave and store to 2 registers.
-Scheme:
-@code
-{A1 B1 A2 B2 ...} ==> {A1 A2 ...}, {B1 B2 ...}
-@endcode
-For all types except 64-bit. */
-template<typename _Tp, int n> inline void v_load_deinterleave(const _Tp* ptr, v_reg<_Tp, n>& a,
-                                                            v_reg<_Tp, n>& b)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    int i, i2;
-    for( i = i2 = 0; i < n; i++, i2 += 2 )
-    {
-        a.s[i] = ptr[i2];
-        b.s[i] = ptr[i2+1];
-    }
-}
-
-/** @brief Load and deinterleave (3 channels)
-
-Load data from memory deinterleave and store to 3 registers.
-Scheme:
-@code
-{A1 B1 C1 A2 B2 C2 ...} ==> {A1 A2 ...}, {B1 B2 ...}, {C1 C2 ...}
-@endcode
-For all types except 64-bit. */
-template<typename _Tp, int n> inline void v_load_deinterleave(const _Tp* ptr, v_reg<_Tp, n>& a,
-                                                            v_reg<_Tp, n>& b, v_reg<_Tp, n>& c)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    int i, i3;
-    for( i = i3 = 0; i < n; i++, i3 += 3 )
-    {
-        a.s[i] = ptr[i3];
-        b.s[i] = ptr[i3+1];
-        c.s[i] = ptr[i3+2];
-    }
-}
-
-/** @brief Load and deinterleave (4 channels)
-
-Load data from memory deinterleave and store to 4 registers.
-Scheme:
-@code
-{A1 B1 C1 D1 A2 B2 C2 D2 ...} ==> {A1 A2 ...}, {B1 B2 ...}, {C1 C2 ...}, {D1 D2 ...}
-@endcode
-For all types except 64-bit. */
-template<typename _Tp, int n>
-inline void v_load_deinterleave(const _Tp* ptr, v_reg<_Tp, n>& a,
-                                v_reg<_Tp, n>& b, v_reg<_Tp, n>& c,
-                                v_reg<_Tp, n>& d)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    int i, i4;
-    for( i = i4 = 0; i < n; i++, i4 += 4 )
-    {
-        a.s[i] = ptr[i4];
-        b.s[i] = ptr[i4+1];
-        c.s[i] = ptr[i4+2];
-        d.s[i] = ptr[i4+3];
-    }
-}
-
-/** @brief Interleave and store (2 channels)
-
-Interleave and store data from 2 registers to memory.
-Scheme:
-@code
-{A1 A2 ...}, {B1 B2 ...} ==> {A1 B1 A2 B2 ...}
-@endcode
-For all types except 64-bit. */
-template<typename _Tp, int n>
-inline void v_store_interleave( _Tp* ptr, const v_reg<_Tp, n>& a,
-                               const v_reg<_Tp, n>& b,
-                               hal::StoreMode /*mode*/=hal::STORE_UNALIGNED)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    int i, i2;
-    for( i = i2 = 0; i < n; i++, i2 += 2 )
-    {
-        ptr[i2] = a.s[i];
-        ptr[i2+1] = b.s[i];
-    }
-}
-
-/** @brief Interleave and store (3 channels)
-
-Interleave and store data from 3 registers to memory.
-Scheme:
-@code
-{A1 A2 ...}, {B1 B2 ...}, {C1 C2 ...} ==> {A1 B1 C1 A2 B2 C2 ...}
-@endcode
-For all types except 64-bit. */
-template<typename _Tp, int n>
-inline void v_store_interleave( _Tp* ptr, const v_reg<_Tp, n>& a,
-                                const v_reg<_Tp, n>& b, const v_reg<_Tp, n>& c,
-                                hal::StoreMode /*mode*/=hal::STORE_UNALIGNED)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    int i, i3;
-    for( i = i3 = 0; i < n; i++, i3 += 3 )
-    {
-        ptr[i3] = a.s[i];
-        ptr[i3+1] = b.s[i];
-        ptr[i3+2] = c.s[i];
-    }
-}
-
-/** @brief Interleave and store (4 channels)
-
-Interleave and store data from 4 registers to memory.
-Scheme:
-@code
-{A1 A2 ...}, {B1 B2 ...}, {C1 C2 ...}, {D1 D2 ...} ==> {A1 B1 C1 D1 A2 B2 C2 D2 ...}
-@endcode
-For all types except 64-bit. */
-template<typename _Tp, int n> inline void v_store_interleave( _Tp* ptr, const v_reg<_Tp, n>& a,
-                                                            const v_reg<_Tp, n>& b, const v_reg<_Tp, n>& c,
-                                                            const v_reg<_Tp, n>& d,
-                                                            hal::StoreMode /*mode*/=hal::STORE_UNALIGNED)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    int i, i4;
-    for( i = i4 = 0; i < n; i++, i4 += 4 )
-    {
-        ptr[i4] = a.s[i];
-        ptr[i4+1] = b.s[i];
-        ptr[i4+2] = c.s[i];
-        ptr[i4+3] = d.s[i];
-    }
-}
-
-/** @brief Store data to memory
-
-Store register contents to memory.
-Scheme:
-@code
-  REG {A B C D} ==> MEM {A B C D}
-@endcode
-Pointer can be unaligned. */
-template<typename _Tp, int n>
-inline void v_store(_Tp* ptr, const v_reg<_Tp, n>& a)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    for( int i = 0; i < n; i++ )
-        ptr[i] = a.s[i];
-}
-
-template<typename _Tp, int n>
-inline void v_store(_Tp* ptr, const v_reg<_Tp, n>& a, hal::StoreMode /*mode*/)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    v_store(ptr, a);
-}
-
-/** @brief Store data to memory (lower half)
-
-Store lower half of register contents to memory.
-Scheme:
-@code
-  REG {A B C D} ==> MEM {A B}
-@endcode */
-template<typename _Tp, int n>
-inline void v_store_low(_Tp* ptr, const v_reg<_Tp, n>& a)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    for( int i = 0; i < (n/2); i++ )
-        ptr[i] = a.s[i];
-}
-
-/** @brief Store data to memory (higher half)
-
-Store higher half of register contents to memory.
-Scheme:
-@code
-  REG {A B C D} ==> MEM {C D}
-@endcode */
-template<typename _Tp, int n>
-inline void v_store_high(_Tp* ptr, const v_reg<_Tp, n>& a)
-{
-#if CV_STRONG_ALIGNMENT
-    CV_Assert(isAligned<sizeof(_Tp)>(ptr));
-#endif
-    for( int i = 0; i < (n/2); i++ )
-        ptr[i] = a.s[i+(n/2)];
-}
-
-/** @brief Store data to memory (aligned)
-
-Store register contents to memory.
-Scheme:
-@code
-  REG {A B C D} ==> MEM {A B C D}
-@endcode
-Pointer __should__ be aligned by 16-byte boundary. */
-template<typename _Tp, int n>
-inline void v_store_aligned(_Tp* ptr, const v_reg<_Tp, n>& a)
-{
-    CV_Assert(isAligned<sizeof(v_reg<_Tp, n>)>(ptr));
-    v_store(ptr, a);
-}
-
-template<typename _Tp, int n>
-inline void v_store_aligned_nocache(_Tp* ptr, const v_reg<_Tp, n>& a)
-{
-    CV_Assert(isAligned<sizeof(v_reg<_Tp, n>)>(ptr));
-    v_store(ptr, a);
-}
-
-template<typename _Tp, int n>
-inline void v_store_aligned(_Tp* ptr, const v_reg<_Tp, n>& a, hal::StoreMode /*mode*/)
-{
-    CV_Assert(isAligned<sizeof(v_reg<_Tp, n>)>(ptr));
-    v_store(ptr, a);
-}
-
-/** @brief Combine vector from first elements of two vectors
-
-Scheme:
-@code
-  {A1 A2 A3 A4}
-  {B1 B2 B3 B4}
----------------
-  {A1 A2 B1 B2}
-@endcode
-For all types except 64-bit. */
-template<typename _Tp, int n>
-inline v_reg<_Tp, n> v_combine_low(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b)
-{
-    v_reg<_Tp, n> c;
-    for( int i = 0; i < (n/2); i++ )
-    {
-        c.s[i] = a.s[i];
-        c.s[i+(n/2)] = b.s[i];
-    }
-    return c;
-}
-
-/** @brief Combine vector from last elements of two vectors
-
-Scheme:
-@code
-  {A1 A2 A3 A4}
-  {B1 B2 B3 B4}
----------------
-  {A3 A4 B3 B4}
-@endcode
-For all types except 64-bit. */
-template<typename _Tp, int n>
-inline v_reg<_Tp, n> v_combine_high(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b)
-{
-    v_reg<_Tp, n> c;
-    for( int i = 0; i < (n/2); i++ )
-    {
-        c.s[i] = a.s[i+(n/2)];
-        c.s[i+(n/2)] = b.s[i+(n/2)];
-    }
-    return c;
-}
-
-/** @brief Combine two vectors from lower and higher parts of two other vectors
-
-@code{.cpp}
-low = cv::v_combine_low(a, b);
-high = cv::v_combine_high(a, b);
-@endcode */
-template<typename _Tp, int n>
-inline void v_recombine(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b,
-                        v_reg<_Tp, n>& low, v_reg<_Tp, n>& high)
-{
-    for( int i = 0; i < (n/2); i++ )
-    {
-        low.s[i] = a.s[i];
-        low.s[i+(n/2)] = b.s[i];
-        high.s[i] = a.s[i+(n/2)];
-        high.s[i+(n/2)] = b.s[i+(n/2)];
-    }
-}
-
-/** @brief Vector reverse order
-
-Reverse the order of the vector
-Scheme:
-@code
-  REG {A1 ... An} ==> REG {An ... A1}
-@endcode
-For all types. */
-template<typename _Tp, int n>
-inline v_reg<_Tp, n> v_reverse(const v_reg<_Tp, n>& a)
-{
-    v_reg<_Tp, n> c;
-    for( int i = 0; i < n; i++ )
-        c.s[i] = a.s[n-i-1];
-    return c;
-}
-
-/** @brief Vector extract
-
-Scheme:
-@code
-  {A1 A2 A3 A4}
-  {B1 B2 B3 B4}
-========================
-shift = 1  {A2 A3 A4 B1}
-shift = 2  {A3 A4 B1 B2}
-shift = 3  {A4 B1 B2 B3}
-@endcode
-Restriction: 0 <= shift < nlanes
-
-Usage:
-@code
-v_int32x4 a, b, c;
-c = v_extract<2>(a, b);
-@endcode
-For all types. */
-template<int s, typename _Tp, int n>
-inline v_reg<_Tp, n> v_extract(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b)
-{
-    v_reg<_Tp, n> r;
-    const int shift = n - s;
-    int i = 0;
-    for (; i < shift; ++i)
-        r.s[i] = a.s[i+s];
-    for (; i < n; ++i)
-        r.s[i] = b.s[i-shift];
-    return r;
-}
-
-/** @brief Vector extract
-
-Scheme:
-Return the s-th element of v.
-Restriction: 0 <= s < nlanes
-
-Usage:
-@code
-v_int32x4 a;
-int r;
-r = v_extract_n<2>(a);
-@endcode
-For all types. */
-template<int s, typename _Tp, int n>
-inline _Tp v_extract_n(const v_reg<_Tp, n>& v)
-{
-    CV_DbgAssert(s >= 0 && s < n);
-    return v.s[s];
-}
-
-/** @brief Broadcast i-th element of vector
-
-Scheme:
-@code
-{ v[0] v[1] v[2] ... v[SZ] } => { v[i], v[i], v[i] ... v[i] }
-@endcode
-Restriction: 0 <= i < nlanes
-Supported types: 32-bit integers and floats (s32/u32/f32)
- */
-template<int i, typename _Tp, int n>
-inline v_reg<_Tp, n> v_broadcast_element(const v_reg<_Tp, n>& a)
-{
-    CV_DbgAssert(i >= 0 && i < n);
-    return v_reg<_Tp, n>::all(a.s[i]);
-}
-
-/** @brief Round elements
-
-Rounds each value. Input type is float vector ==> output type is int vector.
-@note Only for floating point types.
-*/
-template<int n> inline v_reg<int, n> v_round(const v_reg<float, n>& a)
-{
-    v_reg<int, n> c;
-    for( int i = 0; i < n; i++ )
-        c.s[i] = cvRound(a.s[i]);
-    return c;
-}
-
-/** @overload */
-template<int n> inline v_reg<int, n*2> v_round(const v_reg<double, n>& a, const v_reg<double, n>& b)
-{
-    v_reg<int, n*2> c;
-    for( int i = 0; i < n; i++ )
-    {
-        c.s[i] = cvRound(a.s[i]);
-        c.s[i+n] = cvRound(b.s[i]);
-    }
-    return c;
-}
-
-/** @brief Floor elements
-
-Floor each value. Input type is float vector ==> output type is int vector.
-@note Only for floating point types.
-*/
-template<int n> inline v_reg<int, n> v_floor(const v_reg<float, n>& a)
-{
-    v_reg<int, n> c;
-    for( int i = 0; i < n; i++ )
-        c.s[i] = cvFloor(a.s[i]);
-    return c;
-}
-
-/** @brief Ceil elements
-
-Ceil each value. Input type is float vector ==> output type is int vector.
-@note Only for floating point types.
-*/
-template<int n> inline v_reg<int, n> v_ceil(const v_reg<float, n>& a)
-{
-    v_reg<int, n> c;
-    for( int i = 0; i < n; i++ )
-        c.s[i] = cvCeil(a.s[i]);
-    return c;
-}
-
-/** @brief Truncate elements
-
-Truncate each value. Input type is float vector ==> output type is int vector.
-@note Only for floating point types.
-*/
-template<int n> inline v_reg<int, n> v_trunc(const v_reg<float, n>& a)
-{
-    v_reg<int, n> c;
-    for( int i = 0; i < n; i++ )
-        c.s[i] = (int)(a.s[i]);
-    return c;
-}
-
-/** @overload */
-template<int n> inline v_reg<int, n*2> v_round(const v_reg<double, n>& a)
-{
-    v_reg<int, n*2> c;
-    for( int i = 0; i < n; i++ )
-    {
-        c.s[i] = cvRound(a.s[i]);
-        c.s[i+n] = 0;
-    }
-    return c;
-}
-
-/** @overload */
-template<int n> inline v_reg<int, n*2> v_floor(const v_reg<double, n>& a)
-{
-    v_reg<int, n*2> c;
-    for( int i = 0; i < n; i++ )
-    {
-        c.s[i] = cvFloor(a.s[i]);
-        c.s[i+n] = 0;
-    }
-    return c;
-}
-
-/** @overload */
-template<int n> inline v_reg<int, n*2> v_ceil(const v_reg<double, n>& a)
-{
-    v_reg<int, n*2> c;
-    for( int i = 0; i < n; i++ )
-    {
-        c.s[i] = cvCeil(a.s[i]);
-        c.s[i+n] = 0;
-    }
-    return c;
-}
-
-/** @overload */
-template<int n> inline v_reg<int, n*2> v_trunc(const v_reg<double, n>& a)
-{
-    v_reg<int, n*2> c;
-    for( int i = 0; i < n; i++ )
-    {
-        c.s[i] = (int)(a.s[i]);
-        c.s[i+n] = 0;
-    }
-    return c;
-}
-
-/** @brief Convert to float
-
-Supported input type is cv::v_int32. */
-template<int n> inline v_reg<float, n> v_cvt_f32(const v_reg<int, n>& a)
-{
-    v_reg<float, n> c;
-    for( int i = 0; i < n; i++ )
-        c.s[i] = (float)a.s[i];
-    return c;
-}
-
-/** @brief Convert lower half to float
-
-Supported input type is cv::v_float64. */
-template<int n> inline v_reg<float, n*2> v_cvt_f32(const v_reg<double, n>& a)
-{
-    v_reg<float, n*2> c;
-    for( int i = 0; i < n; i++ )
-    {
-        c.s[i] = (float)a.s[i];
-        c.s[i+n] = 0;
-    }
-    return c;
-}
-
-/** @brief Convert to float
-
-Supported input type is cv::v_float64. */
-template<int n> inline v_reg<float, n*2> v_cvt_f32(const v_reg<double, n>& a, const v_reg<double, n>& b)
-{
-    v_reg<float, n*2> c;
-    for( int i = 0; i < n; i++ )
-    {
-        c.s[i] = (float)a.s[i];
-        c.s[i+n] = (float)b.s[i];
-    }
-    return c;
-}
-
-/** @brief Convert lower half to double
-
-Supported input type is cv::v_int32. */
-template<int n> CV_INLINE v_reg<double, n/2> v_cvt_f64(const v_reg<int, n>& a)
-{
-    v_reg<double, (n/2)> c;
-    for( int i = 0; i < (n/2); i++ )
-        c.s[i] = (double)a.s[i];
-    return c;
-}
-
-/** @brief Convert to double high part of vector
-
-Supported input type is cv::v_int32. */
-template<int n> CV_INLINE v_reg<double, (n/2)> v_cvt_f64_high(const v_reg<int, n>& a)
-{
-    v_reg<double, (n/2)> c;
-    for( int i = 0; i < (n/2); i++ )
-        c.s[i] = (double)a.s[i + (n/2)];
-    return c;
-}
-
-/** @brief Convert lower half to double
-
-Supported input type is cv::v_float32. */
-template<int n> CV_INLINE v_reg<double, (n/2)> v_cvt_f64(const v_reg<float, n>& a)
-{
-    v_reg<double, (n/2)> c;
-    for( int i = 0; i < (n/2); i++ )
-        c.s[i] = (double)a.s[i];
-    return c;
-}
-
-/** @brief Convert to double high part of vector
-
-Supported input type is cv::v_float32. */
-template<int n> CV_INLINE v_reg<double, (n/2)> v_cvt_f64_high(const v_reg<float, n>& a)
-{
-    v_reg<double, (n/2)> c;
-    for( int i = 0; i < (n/2); i++ )
-        c.s[i] = (double)a.s[i + (n/2)];
-    return c;
-}
-
-/** @brief Convert to double
-
-Supported input type is cv::v_int64. */
-template<int n> CV_INLINE v_reg<double, n> v_cvt_f64(const v_reg<int64, n>& a)
-{
-    v_reg<double, n> c;
-    for( int i = 0; i < n; i++ )
-        c.s[i] = (double)a.s[i];
-    return c;
-}
-
-
-template<typename _Tp> inline v_reg<_Tp, simd128_width / sizeof(_Tp)> v_lut(const _Tp* tab, const int* idx)
-{
-    v_reg<_Tp, simd128_width / sizeof(_Tp)> c;
-    for (int i = 0; i < c.nlanes; i++)
-        c.s[i] = tab[idx[i]];
-    return c;
-}
-template<typename _Tp> inline v_reg<_Tp, simd128_width / sizeof(_Tp)> v_lut_pairs(const _Tp* tab, const int* idx)
-{
-    v_reg<_Tp, simd128_width / sizeof(_Tp)> c;
-    for (int i = 0; i < c.nlanes; i++)
-        c.s[i] = tab[idx[i / 2] + i % 2];
-    return c;
-}
-template<typename _Tp> inline v_reg<_Tp, simd128_width / sizeof(_Tp)> v_lut_quads(const _Tp* tab, const int* idx)
-{
-    v_reg<_Tp, simd128_width / sizeof(_Tp)> c;
-    for (int i = 0; i < c.nlanes; i++)
-        c.s[i] = tab[idx[i / 4] + i % 4];
-    return c;
-}
-
-template<int n> inline v_reg<int, n> v_lut(const int* tab, const v_reg<int, n>& idx)
-{
-    v_reg<int, n> c;
-    for( int i = 0; i < n; i++ )
-        c.s[i] = tab[idx.s[i]];
-    return c;
-}
-
-template<int n> inline v_reg<unsigned, n> v_lut(const unsigned* tab, const v_reg<int, n>& idx)
-{
-    v_reg<int, n> c;
-    for (int i = 0; i < n; i++)
-        c.s[i] = tab[idx.s[i]];
-    return c;
-}
-
-template<int n> inline v_reg<float, n> v_lut(const float* tab, const v_reg<int, n>& idx)
-{
-    v_reg<float, n> c;
-    for( int i = 0; i < n; i++ )
-        c.s[i] = tab[idx.s[i]];
-    return c;
-}
-
-template<int n> inline v_reg<double, n/2> v_lut(const double* tab, const v_reg<int, n>& idx)
-{
-    v_reg<double, n/2> c;
-    for( int i = 0; i < n/2; i++ )
-        c.s[i] = tab[idx.s[i]];
-    return c;
-}
-
-
-template<int n> inline void v_lut_deinterleave(const float* tab, const v_reg<int, n>& idx,
-                                               v_reg<float, n>& x, v_reg<float, n>& y)
-{
-    for( int i = 0; i < n; i++ )
-    {
-        int j = idx.s[i];
-        x.s[i] = tab[j];
-        y.s[i] = tab[j+1];
-    }
-}
-
-template<int n> inline void v_lut_deinterleave(const double* tab, const v_reg<int, n*2>& idx,
-                                               v_reg<double, n>& x, v_reg<double, n>& y)
-{
-    for( int i = 0; i < n; i++ )
-    {
-        int j = idx.s[i];
-        x.s[i] = tab[j];
-        y.s[i] = tab[j+1];
-    }
-}
-
-template<typename _Tp, int n> inline v_reg<_Tp, n> v_interleave_pairs(const v_reg<_Tp, n>& vec)
-{
-    v_reg<_Tp, n> c;
-    for (int i = 0; i < n/4; i++)
-    {
-        c.s[4*i  ] = vec.s[4*i  ];
-        c.s[4*i+1] = vec.s[4*i+2];
-        c.s[4*i+2] = vec.s[4*i+1];
-        c.s[4*i+3] = vec.s[4*i+3];
-    }
-    return c;
-}
-
-template<typename _Tp, int n> inline v_reg<_Tp, n> v_interleave_quads(const v_reg<_Tp, n>& vec)
-{
-    v_reg<_Tp, n> c;
-    for (int i = 0; i < n/8; i++)
-    {
-        c.s[8*i  ] = vec.s[8*i  ];
-        c.s[8*i+1] = vec.s[8*i+4];
-        c.s[8*i+2] = vec.s[8*i+1];
-        c.s[8*i+3] = vec.s[8*i+5];
-        c.s[8*i+4] = vec.s[8*i+2];
-        c.s[8*i+5] = vec.s[8*i+6];
-        c.s[8*i+6] = vec.s[8*i+3];
-        c.s[8*i+7] = vec.s[8*i+7];
-    }
-    return c;
-}
-
-template<typename _Tp, int n> inline v_reg<_Tp, n> v_pack_triplets(const v_reg<_Tp, n>& vec)
-{
-    v_reg<_Tp, n> c;
-    for (int i = 0; i < n/4; i++)
-    {
-        c.s[3*i  ] = vec.s[4*i  ];
-        c.s[3*i+1] = vec.s[4*i+1];
-        c.s[3*i+2] = vec.s[4*i+2];
-    }
-    return c;
-}
-
-/** @brief Transpose 4x4 matrix
-
-Scheme:
-@code
-a0  {A1 A2 A3 A4}
-a1  {B1 B2 B3 B4}
-a2  {C1 C2 C3 C4}
-a3  {D1 D2 D3 D4}
-===============
-b0  {A1 B1 C1 D1}
-b1  {A2 B2 C2 D2}
-b2  {A3 B3 C3 D3}
-b3  {A4 B4 C4 D4}
-@endcode
-*/
-template<typename _Tp, int n>
-inline void v_transpose4x4( v_reg<_Tp, n>& a0, const v_reg<_Tp, n>& a1,
-                            const v_reg<_Tp, n>& a2, const v_reg<_Tp, n>& a3,
-                            v_reg<_Tp, n>& b0, v_reg<_Tp, n>& b1,
-                            v_reg<_Tp, n>& b2, v_reg<_Tp, n>& b3 )
-{
-    for (int i = 0; i < n / 4; i++)
-    {
-        b0.s[0 + i*4] = a0.s[0 + i*4]; b0.s[1 + i*4] = a1.s[0 + i*4];
-        b0.s[2 + i*4] = a2.s[0 + i*4]; b0.s[3 + i*4] = a3.s[0 + i*4];
-        b1.s[0 + i*4] = a0.s[1 + i*4]; b1.s[1 + i*4] = a1.s[1 + i*4];
-        b1.s[2 + i*4] = a2.s[1 + i*4]; b1.s[3 + i*4] = a3.s[1 + i*4];
-        b2.s[0 + i*4] = a0.s[2 + i*4]; b2.s[1 + i*4] = a1.s[2 + i*4];
-        b2.s[2 + i*4] = a2.s[2 + i*4]; b2.s[3 + i*4] = a3.s[2 + i*4];
-        b3.s[0 + i*4] = a0.s[3 + i*4]; b3.s[1 + i*4] = a1.s[3 + i*4];
-        b3.s[2 + i*4] = a2.s[3 + i*4]; b3.s[3 + i*4] = a3.s[3 + i*4];
-    }
-}
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_C_INIT_ZERO(_Tpvec, prefix, suffix) \
-inline _Tpvec prefix##_setzero_##suffix() { return _Tpvec::zero(); }
-
-//! @name Init with zero
-//! @{
-//! @brief Create new vector with zero elements
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_uint8x16, v, u8)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_int8x16, v, s8)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_uint16x8, v, u16)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_int16x8, v, s16)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_uint32x4, v, u32)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_int32x4, v, s32)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_float32x4, v, f32)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_float64x2, v, f64)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_uint64x2, v, u64)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_int64x2, v, s64)
-
-#if CV_SIMD256
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_uint8x32, v256, u8)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_int8x32, v256, s8)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_uint16x16, v256, u16)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_int16x16, v256, s16)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_uint32x8, v256, u32)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_int32x8, v256, s32)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_float32x8, v256, f32)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_float64x4, v256, f64)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_uint64x4, v256, u64)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_int64x4, v256, s64)
-#endif
-
-#if CV_SIMD512
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_uint8x64, v512, u8)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_int8x64, v512, s8)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_uint16x32, v512, u16)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_int16x32, v512, s16)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_uint32x16, v512, u32)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_int32x16, v512, s32)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_float32x16, v512, f32)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_float64x8, v512, f64)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_uint64x8, v512, u64)
-OPENCV_HAL_IMPL_C_INIT_ZERO(v_int64x8, v512, s64)
-#endif
-//! @}
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_C_INIT_VAL(_Tpvec, _Tp, prefix, suffix) \
-inline _Tpvec prefix##_setall_##suffix(_Tp val) { return _Tpvec::all(val); }
-
-//! @name Init with value
-//! @{
-//! @brief Create new vector with elements set to a specific value
-OPENCV_HAL_IMPL_C_INIT_VAL(v_uint8x16, uchar, v, u8)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_int8x16, schar, v, s8)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_uint16x8, ushort, v, u16)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_int16x8, short, v, s16)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_uint32x4, unsigned, v, u32)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_int32x4, int, v, s32)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_float32x4, float, v, f32)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_float64x2, double, v, f64)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_uint64x2, uint64, v, u64)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_int64x2, int64, v, s64)
-
-#if CV_SIMD256
-OPENCV_HAL_IMPL_C_INIT_VAL(v_uint8x32, uchar, v256, u8)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_int8x32, schar, v256, s8)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_uint16x16, ushort, v256, u16)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_int16x16, short, v256, s16)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_uint32x8, unsigned, v256, u32)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_int32x8, int, v256, s32)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_float32x8, float, v256, f32)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_float64x4, double, v256, f64)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_uint64x4, uint64, v256, u64)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_int64x4, int64, v256, s64)
-#endif
-
-#if CV_SIMD512
-OPENCV_HAL_IMPL_C_INIT_VAL(v_uint8x64, uchar, v512, u8)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_int8x64, schar, v512, s8)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_uint16x32, ushort, v512, u16)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_int16x32, short, v512, s16)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_uint32x16, unsigned, v512, u32)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_int32x16, int, v512, s32)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_float32x16, float, v512, f32)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_float64x8, double, v512, f64)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_uint64x8, uint64, v512, u64)
-OPENCV_HAL_IMPL_C_INIT_VAL(v_int64x8, int64, v512, s64)
-#endif
-//! @}
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_C_REINTERPRET(_Tp, suffix) \
-template<typename _Tp0, int n0> inline v_reg<_Tp, n0*sizeof(_Tp0)/sizeof(_Tp)> \
-    v_reinterpret_as_##suffix(const v_reg<_Tp0, n0>& a) \
-{ return a.template reinterpret_as<_Tp, n0*sizeof(_Tp0)/sizeof(_Tp)>(); }
-
-//! @name Reinterpret
-//! @{
-//! @brief Convert vector to different type without modifying underlying data.
-OPENCV_HAL_IMPL_C_REINTERPRET(uchar, u8)
-OPENCV_HAL_IMPL_C_REINTERPRET(schar, s8)
-OPENCV_HAL_IMPL_C_REINTERPRET(ushort, u16)
-OPENCV_HAL_IMPL_C_REINTERPRET(short, s16)
-OPENCV_HAL_IMPL_C_REINTERPRET(unsigned, u32)
-OPENCV_HAL_IMPL_C_REINTERPRET(int, s32)
-OPENCV_HAL_IMPL_C_REINTERPRET(float, f32)
-OPENCV_HAL_IMPL_C_REINTERPRET(double, f64)
-OPENCV_HAL_IMPL_C_REINTERPRET(uint64, u64)
-OPENCV_HAL_IMPL_C_REINTERPRET(int64, s64)
-//! @}
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_C_SHIFTL(_Tp) \
-template<int shift, int n> inline v_reg<_Tp, n> v_shl(const v_reg<_Tp, n>& a) \
-{ return a << shift; }
-
-//! @name Left shift
-//! @{
-//! @brief Shift left
-OPENCV_HAL_IMPL_C_SHIFTL(ushort)
-OPENCV_HAL_IMPL_C_SHIFTL(short)
-OPENCV_HAL_IMPL_C_SHIFTL(unsigned)
-OPENCV_HAL_IMPL_C_SHIFTL(int)
-OPENCV_HAL_IMPL_C_SHIFTL(uint64)
-OPENCV_HAL_IMPL_C_SHIFTL(int64)
-//! @}
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_C_SHIFTR(_Tp) \
-template<int shift, int n> inline v_reg<_Tp, n> v_shr(const v_reg<_Tp, n>& a) \
-{ return a >> shift; }
-
-//! @name Right shift
-//! @{
-//! @brief Shift right
-OPENCV_HAL_IMPL_C_SHIFTR(ushort)
-OPENCV_HAL_IMPL_C_SHIFTR(short)
-OPENCV_HAL_IMPL_C_SHIFTR(unsigned)
-OPENCV_HAL_IMPL_C_SHIFTR(int)
-OPENCV_HAL_IMPL_C_SHIFTR(uint64)
-OPENCV_HAL_IMPL_C_SHIFTR(int64)
-//! @}
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_C_RSHIFTR(_Tp) \
-template<int shift, int n> inline v_reg<_Tp, n> v_rshr(const v_reg<_Tp, n>& a) \
-{ \
-    v_reg<_Tp, n> c; \
-    for( int i = 0; i < n; i++ ) \
-        c.s[i] = (_Tp)((a.s[i] + ((_Tp)1 << (shift - 1))) >> shift); \
-    return c; \
-}
-
-//! @name Rounding shift
-//! @{
-//! @brief Rounding shift right
-OPENCV_HAL_IMPL_C_RSHIFTR(ushort)
-OPENCV_HAL_IMPL_C_RSHIFTR(short)
-OPENCV_HAL_IMPL_C_RSHIFTR(unsigned)
-OPENCV_HAL_IMPL_C_RSHIFTR(int)
-OPENCV_HAL_IMPL_C_RSHIFTR(uint64)
-OPENCV_HAL_IMPL_C_RSHIFTR(int64)
-//! @}
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_C_PACK(_Tp, _Tpn, pack_suffix, cast) \
-template<int n> inline v_reg<_Tpn, 2*n> v_##pack_suffix(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b) \
-{ \
-    v_reg<_Tpn, 2*n> c; \
-    for( int i = 0; i < n; i++ ) \
-    { \
-        c.s[i] = cast<_Tpn>(a.s[i]); \
-        c.s[i+n] = cast<_Tpn>(b.s[i]); \
-    } \
-    return c; \
-}
-
-//! @name Pack
-//! @{
-//! @brief Pack values from two vectors to one
-//!
-//! Return vector type have twice more elements than input vector types. Variant with _u_ suffix also
-//! converts to corresponding unsigned type.
-//!
-//! - pack: for 16-, 32- and 64-bit integer input types
-//! - pack_u: for 16- and 32-bit signed integer input types
-//!
-//! @note All variants except 64-bit use saturation.
-OPENCV_HAL_IMPL_C_PACK(ushort, uchar, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_PACK(short, schar, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_PACK(unsigned, ushort, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_PACK(int, short, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_PACK(uint64, unsigned, pack, static_cast)
-OPENCV_HAL_IMPL_C_PACK(int64, int, pack, static_cast)
-OPENCV_HAL_IMPL_C_PACK(short, uchar, pack_u, saturate_cast)
-OPENCV_HAL_IMPL_C_PACK(int, ushort, pack_u, saturate_cast)
-//! @}
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_C_RSHR_PACK(_Tp, _Tpn, pack_suffix, cast) \
-template<int shift, int n> inline v_reg<_Tpn, 2*n> v_rshr_##pack_suffix(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b) \
-{ \
-    v_reg<_Tpn, 2*n> c; \
-    for( int i = 0; i < n; i++ ) \
-    { \
-        c.s[i] = cast<_Tpn>((a.s[i] + ((_Tp)1 << (shift - 1))) >> shift); \
-        c.s[i+n] = cast<_Tpn>((b.s[i] + ((_Tp)1 << (shift - 1))) >> shift); \
-    } \
-    return c; \
-}
-
-//! @name Pack with rounding shift
-//! @{
-//! @brief Pack values from two vectors to one with rounding shift
-//!
-//! Values from the input vectors will be shifted right by _n_ bits with rounding, converted to narrower
-//! type and returned in the result vector. Variant with _u_ suffix converts to unsigned type.
-//!
-//! - pack: for 16-, 32- and 64-bit integer input types
-//! - pack_u: for 16- and 32-bit signed integer input types
-//!
-//! @note All variants except 64-bit use saturation.
-OPENCV_HAL_IMPL_C_RSHR_PACK(ushort, uchar, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_RSHR_PACK(short, schar, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_RSHR_PACK(unsigned, ushort, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_RSHR_PACK(int, short, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_RSHR_PACK(uint64, unsigned, pack, static_cast)
-OPENCV_HAL_IMPL_C_RSHR_PACK(int64, int, pack, static_cast)
-OPENCV_HAL_IMPL_C_RSHR_PACK(short, uchar, pack_u, saturate_cast)
-OPENCV_HAL_IMPL_C_RSHR_PACK(int, ushort, pack_u, saturate_cast)
-//! @}
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_C_PACK_STORE(_Tp, _Tpn, pack_suffix, cast) \
-template<int n> inline void v_##pack_suffix##_store(_Tpn* ptr, const v_reg<_Tp, n>& a) \
-{ \
-    for( int i = 0; i < n; i++ ) \
-        ptr[i] = cast<_Tpn>(a.s[i]); \
-}
-
-//! @name Pack and store
-//! @{
-//! @brief Store values from the input vector into memory with pack
-//!
-//! Values will be stored into memory with conversion to narrower type.
-//! Variant with _u_ suffix converts to corresponding unsigned type.
-//!
-//! - pack: for 16-, 32- and 64-bit integer input types
-//! - pack_u: for 16- and 32-bit signed integer input types
-//!
-//! @note All variants except 64-bit use saturation.
-OPENCV_HAL_IMPL_C_PACK_STORE(ushort, uchar, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_PACK_STORE(short, schar, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_PACK_STORE(unsigned, ushort, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_PACK_STORE(int, short, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_PACK_STORE(uint64, unsigned, pack, static_cast)
-OPENCV_HAL_IMPL_C_PACK_STORE(int64, int, pack, static_cast)
-OPENCV_HAL_IMPL_C_PACK_STORE(short, uchar, pack_u, saturate_cast)
-OPENCV_HAL_IMPL_C_PACK_STORE(int, ushort, pack_u, saturate_cast)
-//! @}
-
-//! @brief Helper macro
-//! @ingroup core_hal_intrin_impl
-#define OPENCV_HAL_IMPL_C_RSHR_PACK_STORE(_Tp, _Tpn, pack_suffix, cast) \
-template<int shift, int n> inline void v_rshr_##pack_suffix##_store(_Tpn* ptr, const v_reg<_Tp, n>& a) \
-{ \
-    for( int i = 0; i < n; i++ ) \
-        ptr[i] = cast<_Tpn>((a.s[i] + ((_Tp)1 << (shift - 1))) >> shift); \
-}
-
-//! @name Pack and store with rounding shift
-//! @{
-//! @brief Store values from the input vector into memory with pack
-//!
-//! Values will be shifted _n_ bits right with rounding, converted to narrower type and stored into
-//! memory. Variant with _u_ suffix converts to unsigned type.
-//!
-//! - pack: for 16-, 32- and 64-bit integer input types
-//! - pack_u: for 16- and 32-bit signed integer input types
-//!
-//! @note All variants except 64-bit use saturation.
-OPENCV_HAL_IMPL_C_RSHR_PACK_STORE(ushort, uchar, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_RSHR_PACK_STORE(short, schar, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_RSHR_PACK_STORE(unsigned, ushort, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_RSHR_PACK_STORE(int, short, pack, saturate_cast)
-OPENCV_HAL_IMPL_C_RSHR_PACK_STORE(uint64, unsigned, pack, static_cast)
-OPENCV_HAL_IMPL_C_RSHR_PACK_STORE(int64, int, pack, static_cast)
-OPENCV_HAL_IMPL_C_RSHR_PACK_STORE(short, uchar, pack_u, saturate_cast)
-OPENCV_HAL_IMPL_C_RSHR_PACK_STORE(int, ushort, pack_u, saturate_cast)
-//! @}
-
-//! @cond IGNORED
-template<typename _Tpm, typename _Tp, int n>
-inline void _pack_b(_Tpm* mptr, const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b)
-{
-    for (int i = 0; i < n; ++i)
-    {
-        mptr[i] = (_Tpm)a.s[i];
-        mptr[i + n] = (_Tpm)b.s[i];
-    }
-}
-//! @endcond
-
-//! @name Pack boolean values
-//! @{
-//! @brief Pack boolean values from multiple vectors to one unsigned 8-bit integer vector
-//!
-//! @note Must provide valid boolean values to guarantee same result for all architectures.
-
-/** @brief
-//! For 16-bit boolean values
-
-Scheme:
-@code
-a  {0xFFFF 0 0 0xFFFF 0 0xFFFF 0xFFFF 0}
-b  {0xFFFF 0 0xFFFF 0 0 0xFFFF 0 0xFFFF}
-===============
-{
-   0xFF 0 0 0xFF 0 0xFF 0xFF 0
-   0xFF 0 0xFF 0 0 0xFF 0 0xFF
-}
-@endcode */
-
-template<int n> inline v_reg<uchar, 2*n> v_pack_b(const v_reg<ushort, n>& a, const v_reg<ushort, n>& b)
-{
-    v_reg<uchar, 2*n> mask;
-    _pack_b(mask.s, a, b);
-    return mask;
-}
-
-/** @overload
-For 32-bit boolean values
-
-Scheme:
-@code
-a  {0xFFFF.. 0 0 0xFFFF..}
-b  {0 0xFFFF.. 0xFFFF.. 0}
-c  {0xFFFF.. 0 0xFFFF.. 0}
-d  {0 0xFFFF.. 0 0xFFFF..}
-===============
-{
-   0xFF 0 0 0xFF 0 0xFF 0xFF 0
-   0xFF 0 0xFF 0 0 0xFF 0 0xFF
-}
-@endcode */
-
-template<int n> inline v_reg<uchar, 4*n> v_pack_b(const v_reg<unsigned, n>& a, const v_reg<unsigned, n>& b,
-                                                  const v_reg<unsigned, n>& c, const v_reg<unsigned, n>& d)
-{
-    v_reg<uchar, 4*n> mask;
-    _pack_b(mask.s, a, b);
-    _pack_b(mask.s + 2*n, c, d);
-    return mask;
-}
-
-/** @overload
-For 64-bit boolean values
-
-Scheme:
-@code
-a  {0xFFFF.. 0}
-b  {0 0xFFFF..}
-c  {0xFFFF.. 0}
-d  {0 0xFFFF..}
-
-e  {0xFFFF.. 0}
-f  {0xFFFF.. 0}
-g  {0 0xFFFF..}
-h  {0 0xFFFF..}
-===============
-{
-   0xFF 0 0 0xFF 0xFF 0 0 0xFF
-   0xFF 0 0xFF 0 0 0xFF 0 0xFF
-}
-@endcode */
-template<int n> inline v_reg<uchar, 8*n> v_pack_b(const v_reg<uint64, n>& a, const v_reg<uint64, n>& b,
-                                                  const v_reg<uint64, n>& c, const v_reg<uint64, n>& d,
-                                                  const v_reg<uint64, n>& e, const v_reg<uint64, n>& f,
-                                                  const v_reg<uint64, n>& g, const v_reg<uint64, n>& h)
-{
-    v_reg<uchar, 8*n> mask;
-    _pack_b(mask.s, a, b);
-    _pack_b(mask.s + 2*n, c, d);
-    _pack_b(mask.s + 4*n, e, f);
-    _pack_b(mask.s + 6*n, g, h);
-    return mask;
-}
-//! @}
-
-/** @brief Matrix multiplication
-
-Scheme:
-@code
-{A0 A1 A2 A3}   |V0|
-{B0 B1 B2 B3}   |V1|
-{C0 C1 C2 C3}   |V2|
-{D0 D1 D2 D3} x |V3|
-====================
-{R0 R1 R2 R3}, where:
-R0 = A0V0 + B0V1 + C0V2 + D0V3,
-R1 = A1V0 + B1V1 + C1V2 + D1V3
-...
-@endcode
-*/
-template<int n>
-inline v_reg<float, n> v_matmul(const v_reg<float, n>& v,
-                                const v_reg<float, n>& a, const v_reg<float, n>& b,
-                                const v_reg<float, n>& c, const v_reg<float, n>& d)
-{
-    v_reg<float, n> res;
-    for (int i = 0; i < n / 4; i++)
-    {
-        res.s[0 + i*4] = v.s[0 + i*4] * a.s[0 + i*4] + v.s[1 + i*4] * b.s[0 + i*4] + v.s[2 + i*4] * c.s[0 + i*4] + v.s[3 + i*4] * d.s[0 + i*4];
-        res.s[1 + i*4] = v.s[0 + i*4] * a.s[1 + i*4] + v.s[1 + i*4] * b.s[1 + i*4] + v.s[2 + i*4] * c.s[1 + i*4] + v.s[3 + i*4] * d.s[1 + i*4];
-        res.s[2 + i*4] = v.s[0 + i*4] * a.s[2 + i*4] + v.s[1 + i*4] * b.s[2 + i*4] + v.s[2 + i*4] * c.s[2 + i*4] + v.s[3 + i*4] * d.s[2 + i*4];
-        res.s[3 + i*4] = v.s[0 + i*4] * a.s[3 + i*4] + v.s[1 + i*4] * b.s[3 + i*4] + v.s[2 + i*4] * c.s[3 + i*4] + v.s[3 + i*4] * d.s[3 + i*4];
-    }
-    return res;
-}
-
-/** @brief Matrix multiplication and add
-
-Scheme:
-@code
-{A0 A1 A2 A3}   |V0|   |D0|
-{B0 B1 B2 B3}   |V1|   |D1|
-{C0 C1 C2 C3} x |V2| + |D2|
-====================   |D3|
-{R0 R1 R2 R3}, where:
-R0 = A0V0 + B0V1 + C0V2 + D0,
-R1 = A1V0 + B1V1 + C1V2 + D1
-...
-@endcode
-*/
-template<int n>
-inline v_reg<float, n> v_matmuladd(const v_reg<float, n>& v,
-                                   const v_reg<float, n>& a, const v_reg<float, n>& b,
-                                   const v_reg<float, n>& c, const v_reg<float, n>& d)
-{
-    v_reg<float, n> res;
-    for (int i = 0; i < n / 4; i++)
-    {
-        res.s[0 + i * 4] = v.s[0 + i * 4] * a.s[0 + i * 4] + v.s[1 + i * 4] * b.s[0 + i * 4] + v.s[2 + i * 4] * c.s[0 + i * 4] + d.s[0 + i * 4];
-        res.s[1 + i * 4] = v.s[0 + i * 4] * a.s[1 + i * 4] + v.s[1 + i * 4] * b.s[1 + i * 4] + v.s[2 + i * 4] * c.s[1 + i * 4] + d.s[1 + i * 4];
-        res.s[2 + i * 4] = v.s[0 + i * 4] * a.s[2 + i * 4] + v.s[1 + i * 4] * b.s[2 + i * 4] + v.s[2 + i * 4] * c.s[2 + i * 4] + d.s[2 + i * 4];
-        res.s[3 + i * 4] = v.s[0 + i * 4] * a.s[3 + i * 4] + v.s[1 + i * 4] * b.s[3 + i * 4] + v.s[2 + i * 4] * c.s[3 + i * 4] + d.s[3 + i * 4];
-    }
-    return res;
-}
-
-
-template<int n> inline v_reg<double, n/2> v_dotprod_expand(const v_reg<int, n>& a, const v_reg<int, n>& b)
-{ return v_fma(v_cvt_f64(a), v_cvt_f64(b), v_cvt_f64_high(a) * v_cvt_f64_high(b)); }
-template<int n> inline v_reg<double, n/2> v_dotprod_expand(const v_reg<int, n>& a, const v_reg<int, n>& b,
-                                                           const v_reg<double, n/2>& c)
-{ return v_fma(v_cvt_f64(a), v_cvt_f64(b), v_fma(v_cvt_f64_high(a), v_cvt_f64_high(b), c)); }
-
-template<int n> inline v_reg<double, n/2> v_dotprod_expand_fast(const v_reg<int, n>& a, const v_reg<int, n>& b)
-{ return v_dotprod_expand(a, b); }
-template<int n> inline v_reg<double, n/2> v_dotprod_expand_fast(const v_reg<int, n>& a, const v_reg<int, n>& b,
-                                                                const v_reg<double, n/2>& c)
-{ return v_dotprod_expand(a, b, c); }
-
-////// FP16 support ///////
-
-inline v_reg<float, simd128_width / sizeof(float)>
-v_load_expand(const float16_t* ptr)
-{
-    v_reg<float, simd128_width / sizeof(float)> v;
-    for( int i = 0; i < v.nlanes; i++ )
-    {
-        v.s[i] = ptr[i];
-    }
-    return v;
-}
-#if CV_SIMD256
-inline v_reg<float, simd256_width / sizeof(float)>
-v256_load_expand(const float16_t* ptr)
-{
-    v_reg<float, simd256_width / sizeof(float)> v;
-    for (int i = 0; i < v.nlanes; i++)
-    {
-        v.s[i] = ptr[i];
-    }
-    return v;
-}
-#endif
-#if CV_SIMD512
-inline v_reg<float, simd512_width / sizeof(float)>
-v512_load_expand(const float16_t* ptr)
-{
-    v_reg<float, simd512_width / sizeof(float)> v;
-    for (int i = 0; i < v.nlanes; i++)
-    {
-        v.s[i] = ptr[i];
-    }
-    return v;
-}
-#endif
-
-template<int n> inline void
-v_pack_store(float16_t* ptr, const v_reg<float, n>& v)
-{
-    for( int i = 0; i < v.nlanes; i++ )
-    {
-        ptr[i] = float16_t(v.s[i]);
-    }
-}
-
-inline void v_cleanup() {}
-#if CV_SIMD256
-inline void v256_cleanup() {}
-#endif
-#if CV_SIMD512
-inline void v512_cleanup() {}
-#endif
-
-//! @}
-
-#ifndef CV_DOXYGEN
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END
-#endif
-}
-
-#if !defined(CV_DOXYGEN)
-#undef CV_SIMD256
-#undef CV_SIMD512
-#endif
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_forward.hpp b/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_forward.hpp
deleted file mode 100644
index 979f15a2..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_forward.hpp
+++ /dev/null
@@ -1,191 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html
-
-#ifndef CV__SIMD_FORWARD
-#error "Need to pre-define forward width"
-#endif
-
-namespace cv
-{
-
-//! @cond IGNORED
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
-
-/** Types **/
-#if CV__SIMD_FORWARD == 1024
-// [todo] 1024
-#error "1024-long ops not implemented yet"
-#elif CV__SIMD_FORWARD == 512
-// 512
-#define __CV_VX(fun)   v512_##fun
-#define __CV_V_UINT8   v_uint8x64
-#define __CV_V_INT8    v_int8x64
-#define __CV_V_UINT16  v_uint16x32
-#define __CV_V_INT16   v_int16x32
-#define __CV_V_UINT32  v_uint32x16
-#define __CV_V_INT32   v_int32x16
-#define __CV_V_UINT64  v_uint64x8
-#define __CV_V_INT64   v_int64x8
-#define __CV_V_FLOAT32 v_float32x16
-#define __CV_V_FLOAT64 v_float64x8
-struct v_uint8x64;
-struct v_int8x64;
-struct v_uint16x32;
-struct v_int16x32;
-struct v_uint32x16;
-struct v_int32x16;
-struct v_uint64x8;
-struct v_int64x8;
-struct v_float32x16;
-struct v_float64x8;
-#elif CV__SIMD_FORWARD == 256
-// 256
-#define __CV_VX(fun)   v256_##fun
-#define __CV_V_UINT8   v_uint8x32
-#define __CV_V_INT8    v_int8x32
-#define __CV_V_UINT16  v_uint16x16
-#define __CV_V_INT16   v_int16x16
-#define __CV_V_UINT32  v_uint32x8
-#define __CV_V_INT32   v_int32x8
-#define __CV_V_UINT64  v_uint64x4
-#define __CV_V_INT64   v_int64x4
-#define __CV_V_FLOAT32 v_float32x8
-#define __CV_V_FLOAT64 v_float64x4
-struct v_uint8x32;
-struct v_int8x32;
-struct v_uint16x16;
-struct v_int16x16;
-struct v_uint32x8;
-struct v_int32x8;
-struct v_uint64x4;
-struct v_int64x4;
-struct v_float32x8;
-struct v_float64x4;
-#else
-// 128
-#define __CV_VX(fun)   v_##fun
-#define __CV_V_UINT8   v_uint8x16
-#define __CV_V_INT8    v_int8x16
-#define __CV_V_UINT16  v_uint16x8
-#define __CV_V_INT16   v_int16x8
-#define __CV_V_UINT32  v_uint32x4
-#define __CV_V_INT32   v_int32x4
-#define __CV_V_UINT64  v_uint64x2
-#define __CV_V_INT64   v_int64x2
-#define __CV_V_FLOAT32 v_float32x4
-#define __CV_V_FLOAT64 v_float64x2
-struct v_uint8x16;
-struct v_int8x16;
-struct v_uint16x8;
-struct v_int16x8;
-struct v_uint32x4;
-struct v_int32x4;
-struct v_uint64x2;
-struct v_int64x2;
-struct v_float32x4;
-struct v_float64x2;
-#endif
-
-/** Value reordering **/
-
-// Expansion
-void v_expand(const __CV_V_UINT8&,  __CV_V_UINT16&, __CV_V_UINT16&);
-void v_expand(const __CV_V_INT8&,   __CV_V_INT16&,  __CV_V_INT16&);
-void v_expand(const __CV_V_UINT16&, __CV_V_UINT32&, __CV_V_UINT32&);
-void v_expand(const __CV_V_INT16&,  __CV_V_INT32&,  __CV_V_INT32&);
-void v_expand(const __CV_V_UINT32&, __CV_V_UINT64&, __CV_V_UINT64&);
-void v_expand(const __CV_V_INT32&,  __CV_V_INT64&,  __CV_V_INT64&);
-// Low Expansion
-__CV_V_UINT16 v_expand_low(const __CV_V_UINT8&);
-__CV_V_INT16  v_expand_low(const __CV_V_INT8&);
-__CV_V_UINT32 v_expand_low(const __CV_V_UINT16&);
-__CV_V_INT32  v_expand_low(const __CV_V_INT16&);
-__CV_V_UINT64 v_expand_low(const __CV_V_UINT32&);
-__CV_V_INT64  v_expand_low(const __CV_V_INT32&);
-// High Expansion
-__CV_V_UINT16 v_expand_high(const __CV_V_UINT8&);
-__CV_V_INT16  v_expand_high(const __CV_V_INT8&);
-__CV_V_UINT32 v_expand_high(const __CV_V_UINT16&);
-__CV_V_INT32  v_expand_high(const __CV_V_INT16&);
-__CV_V_UINT64 v_expand_high(const __CV_V_UINT32&);
-__CV_V_INT64  v_expand_high(const __CV_V_INT32&);
-// Load & Low Expansion
-__CV_V_UINT16 __CV_VX(load_expand)(const uchar*);
-__CV_V_INT16  __CV_VX(load_expand)(const schar*);
-__CV_V_UINT32 __CV_VX(load_expand)(const ushort*);
-__CV_V_INT32  __CV_VX(load_expand)(const short*);
-__CV_V_UINT64 __CV_VX(load_expand)(const uint*);
-__CV_V_INT64  __CV_VX(load_expand)(const int*);
-// Load lower 8-bit and expand into 32-bit
-__CV_V_UINT32 __CV_VX(load_expand_q)(const uchar*);
-__CV_V_INT32  __CV_VX(load_expand_q)(const schar*);
-
-// Saturating Pack
-__CV_V_UINT8  v_pack(const __CV_V_UINT16&, const __CV_V_UINT16&);
-__CV_V_INT8   v_pack(const __CV_V_INT16&,  const __CV_V_INT16&);
-__CV_V_UINT16 v_pack(const __CV_V_UINT32&, const __CV_V_UINT32&);
-__CV_V_INT16  v_pack(const __CV_V_INT32&,  const __CV_V_INT32&);
-// Non-saturating Pack
-__CV_V_UINT32 v_pack(const __CV_V_UINT64&, const __CV_V_UINT64&);
-__CV_V_INT32  v_pack(const __CV_V_INT64&,  const __CV_V_INT64&);
-// Pack signed integers with unsigned saturation
-__CV_V_UINT8  v_pack_u(const __CV_V_INT16&, const __CV_V_INT16&);
-__CV_V_UINT16 v_pack_u(const __CV_V_INT32&, const __CV_V_INT32&);
-
-/** Arithmetic, bitwise and comparison operations **/
-
-// Non-saturating multiply
-#if CV_VSX
-template<typename Tvec>
-Tvec v_mul_wrap(const Tvec& a, const Tvec& b);
-#else
-__CV_V_UINT8  v_mul_wrap(const __CV_V_UINT8&,  const __CV_V_UINT8&);
-__CV_V_INT8   v_mul_wrap(const __CV_V_INT8&,   const __CV_V_INT8&);
-__CV_V_UINT16 v_mul_wrap(const __CV_V_UINT16&, const __CV_V_UINT16&);
-__CV_V_INT16  v_mul_wrap(const __CV_V_INT16&,  const __CV_V_INT16&);
-#endif
-
-//  Multiply and expand
-#if CV_VSX
-template<typename Tvec, typename Twvec>
-void v_mul_expand(const Tvec& a, const Tvec& b, Twvec& c, Twvec& d);
-#else
-void v_mul_expand(const __CV_V_UINT8&,  const __CV_V_UINT8&,  __CV_V_UINT16&, __CV_V_UINT16&);
-void v_mul_expand(const __CV_V_INT8&,   const __CV_V_INT8&,   __CV_V_INT16&,  __CV_V_INT16&);
-void v_mul_expand(const __CV_V_UINT16&, const __CV_V_UINT16&, __CV_V_UINT32&, __CV_V_UINT32&);
-void v_mul_expand(const __CV_V_INT16&,  const __CV_V_INT16&,  __CV_V_INT32&,  __CV_V_INT32&);
-void v_mul_expand(const __CV_V_UINT32&, const __CV_V_UINT32&, __CV_V_UINT64&, __CV_V_UINT64&);
-void v_mul_expand(const __CV_V_INT32&,  const __CV_V_INT32&,  __CV_V_INT64&,  __CV_V_INT64&);
-#endif
-
-// Conversions
-__CV_V_FLOAT32 v_cvt_f32(const __CV_V_INT32& a);
-__CV_V_FLOAT32 v_cvt_f32(const __CV_V_FLOAT64& a);
-__CV_V_FLOAT32 v_cvt_f32(const __CV_V_FLOAT64& a, const __CV_V_FLOAT64& b);
-__CV_V_FLOAT64 v_cvt_f64(const __CV_V_INT32& a);
-__CV_V_FLOAT64 v_cvt_f64_high(const __CV_V_INT32& a);
-__CV_V_FLOAT64 v_cvt_f64(const __CV_V_FLOAT32& a);
-__CV_V_FLOAT64 v_cvt_f64_high(const __CV_V_FLOAT32& a);
-__CV_V_FLOAT64 v_cvt_f64(const __CV_V_INT64& a);
-
-/** Cleanup **/
-#undef CV__SIMD_FORWARD
-#undef __CV_VX
-#undef __CV_V_UINT8
-#undef __CV_V_INT8
-#undef __CV_V_UINT16
-#undef __CV_V_INT16
-#undef __CV_V_UINT32
-#undef __CV_V_INT32
-#undef __CV_V_UINT64
-#undef __CV_V_INT64
-#undef __CV_V_FLOAT32
-#undef __CV_V_FLOAT64
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END
-
-//! @endcond
-
-} // cv::
\ No newline at end of file
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_msa.hpp b/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_msa.hpp
deleted file mode 100644
index a1fbb093..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_msa.hpp
+++ /dev/null
@@ -1,1887 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_HAL_INTRIN_MSA_HPP
-#define OPENCV_HAL_INTRIN_MSA_HPP
-
-#include <algorithm>
-#include "opencv2/core/utility.hpp"
-
-namespace cv
-{
-
-//! @cond IGNORED
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
-
-#define CV_SIMD128 1
-
-//MSA implements 128-bit wide vector registers shared with the 64-bit wide floating-point unit registers.
-//MSA and FPU can not be both present, unless the FPU has 64-bit floating-point registers.
-#define CV_SIMD128_64F 1
-
-struct v_uint8x16
-{
-    typedef uchar lane_type;
-    enum { nlanes = 16 };
-
-    v_uint8x16() {}
-    explicit v_uint8x16(v16u8 v) : val(v) {}
-    v_uint8x16(uchar v0, uchar v1, uchar v2, uchar v3, uchar v4, uchar v5, uchar v6, uchar v7,
-               uchar v8, uchar v9, uchar v10, uchar v11, uchar v12, uchar v13, uchar v14, uchar v15)
-    {
-        uchar v[] = {v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15};
-        val = msa_ld1q_u8(v);
-    }
-
-    uchar get0() const
-    {
-        return msa_getq_lane_u8(val, 0);
-    }
-
-    v16u8 val;
-};
-
-struct v_int8x16
-{
-    typedef schar lane_type;
-    enum { nlanes = 16 };
-
-    v_int8x16() {}
-    explicit v_int8x16(v16i8 v) : val(v) {}
-    v_int8x16(schar v0, schar v1, schar v2, schar v3, schar v4, schar v5, schar v6, schar v7,
-               schar v8, schar v9, schar v10, schar v11, schar v12, schar v13, schar v14, schar v15)
-    {
-        schar v[] = {v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15};
-        val = msa_ld1q_s8(v);
-    }
-
-    schar get0() const
-    {
-        return msa_getq_lane_s8(val, 0);
-    }
-
-    v16i8 val;
-};
-
-struct v_uint16x8
-{
-    typedef ushort lane_type;
-    enum { nlanes = 8 };
-
-    v_uint16x8() {}
-    explicit v_uint16x8(v8u16 v) : val(v) {}
-    v_uint16x8(ushort v0, ushort v1, ushort v2, ushort v3, ushort v4, ushort v5, ushort v6, ushort v7)
-    {
-        ushort v[] = {v0, v1, v2, v3, v4, v5, v6, v7};
-        val = msa_ld1q_u16(v);
-    }
-
-    ushort get0() const
-    {
-        return msa_getq_lane_u16(val, 0);
-    }
-
-    v8u16 val;
-};
-
-struct v_int16x8
-{
-    typedef short lane_type;
-    enum { nlanes = 8 };
-
-    v_int16x8() {}
-    explicit v_int16x8(v8i16 v) : val(v) {}
-    v_int16x8(short v0, short v1, short v2, short v3, short v4, short v5, short v6, short v7)
-    {
-        short v[] = {v0, v1, v2, v3, v4, v5, v6, v7};
-        val = msa_ld1q_s16(v);
-    }
-
-    short get0() const
-    {
-        return msa_getq_lane_s16(val, 0);
-    }
-
-    v8i16 val;
-};
-
-struct v_uint32x4
-{
-    typedef unsigned int lane_type;
-    enum { nlanes = 4 };
-
-    v_uint32x4() {}
-    explicit v_uint32x4(v4u32 v) : val(v) {}
-    v_uint32x4(unsigned int v0, unsigned int v1, unsigned int v2, unsigned int v3)
-    {
-        unsigned int v[] = {v0, v1, v2, v3};
-        val = msa_ld1q_u32(v);
-    }
-
-    unsigned int get0() const
-    {
-        return msa_getq_lane_u32(val, 0);
-    }
-
-    v4u32 val;
-};
-
-struct v_int32x4
-{
-    typedef int lane_type;
-    enum { nlanes = 4 };
-
-    v_int32x4() {}
-    explicit v_int32x4(v4i32 v) : val(v) {}
-    v_int32x4(int v0, int v1, int v2, int v3)
-    {
-        int v[] = {v0, v1, v2, v3};
-        val = msa_ld1q_s32(v);
-    }
-
-    int get0() const
-    {
-        return msa_getq_lane_s32(val, 0);
-    }
-
-    v4i32 val;
-};
-
-struct v_float32x4
-{
-    typedef float lane_type;
-    enum { nlanes = 4 };
-
-    v_float32x4() {}
-    explicit v_float32x4(v4f32 v) : val(v) {}
-    v_float32x4(float v0, float v1, float v2, float v3)
-    {
-        float v[] = {v0, v1, v2, v3};
-        val = msa_ld1q_f32(v);
-    }
-
-    float get0() const
-    {
-        return msa_getq_lane_f32(val, 0);
-    }
-
-    v4f32 val;
-};
-
-struct v_uint64x2
-{
-    typedef uint64 lane_type;
-    enum { nlanes = 2 };
-
-    v_uint64x2() {}
-    explicit v_uint64x2(v2u64 v) : val(v) {}
-    v_uint64x2(uint64 v0, uint64 v1)
-    {
-        uint64 v[] = {v0, v1};
-        val = msa_ld1q_u64(v);
-    }
-
-    uint64 get0() const
-    {
-        return msa_getq_lane_u64(val, 0);
-    }
-
-    v2u64 val;
-};
-
-struct v_int64x2
-{
-    typedef int64 lane_type;
-    enum { nlanes = 2 };
-
-    v_int64x2() {}
-    explicit v_int64x2(v2i64 v) : val(v) {}
-    v_int64x2(int64 v0, int64 v1)
-    {
-        int64 v[] = {v0, v1};
-        val = msa_ld1q_s64(v);
-    }
-
-    int64 get0() const
-    {
-        return msa_getq_lane_s64(val, 0);
-    }
-
-    v2i64 val;
-};
-
-struct v_float64x2
-{
-    typedef double lane_type;
-    enum { nlanes = 2 };
-
-    v_float64x2() {}
-    explicit v_float64x2(v2f64 v) : val(v) {}
-    v_float64x2(double v0, double v1)
-    {
-        double v[] = {v0, v1};
-        val = msa_ld1q_f64(v);
-    }
-
-    double get0() const
-    {
-        return msa_getq_lane_f64(val, 0);
-    }
-
-    v2f64 val;
-};
-
-#define OPENCV_HAL_IMPL_MSA_INIT(_Tpv, _Tp, suffix) \
-inline v_##_Tpv v_setzero_##suffix() { return v_##_Tpv(msa_dupq_n_##suffix((_Tp)0)); } \
-inline v_##_Tpv v_setall_##suffix(_Tp v) { return v_##_Tpv(msa_dupq_n_##suffix(v)); } \
-inline v_uint8x16 v_reinterpret_as_u8(const v_##_Tpv& v) { return v_uint8x16(MSA_TPV_REINTERPRET(v16u8, v.val)); } \
-inline v_int8x16 v_reinterpret_as_s8(const v_##_Tpv& v) { return v_int8x16(MSA_TPV_REINTERPRET(v16i8, v.val)); } \
-inline v_uint16x8 v_reinterpret_as_u16(const v_##_Tpv& v) { return v_uint16x8(MSA_TPV_REINTERPRET(v8u16, v.val)); } \
-inline v_int16x8 v_reinterpret_as_s16(const v_##_Tpv& v) { return v_int16x8(MSA_TPV_REINTERPRET(v8i16, v.val)); } \
-inline v_uint32x4 v_reinterpret_as_u32(const v_##_Tpv& v) { return v_uint32x4(MSA_TPV_REINTERPRET(v4u32, v.val)); } \
-inline v_int32x4 v_reinterpret_as_s32(const v_##_Tpv& v) { return v_int32x4(MSA_TPV_REINTERPRET(v4i32, v.val)); } \
-inline v_uint64x2 v_reinterpret_as_u64(const v_##_Tpv& v) { return v_uint64x2(MSA_TPV_REINTERPRET(v2u64, v.val)); } \
-inline v_int64x2 v_reinterpret_as_s64(const v_##_Tpv& v) { return v_int64x2(MSA_TPV_REINTERPRET(v2i64, v.val)); } \
-inline v_float32x4 v_reinterpret_as_f32(const v_##_Tpv& v) { return v_float32x4(MSA_TPV_REINTERPRET(v4f32, v.val)); } \
-inline v_float64x2 v_reinterpret_as_f64(const v_##_Tpv& v) { return v_float64x2(MSA_TPV_REINTERPRET(v2f64, v.val)); }
-
-OPENCV_HAL_IMPL_MSA_INIT(uint8x16, uchar, u8)
-OPENCV_HAL_IMPL_MSA_INIT(int8x16, schar, s8)
-OPENCV_HAL_IMPL_MSA_INIT(uint16x8, ushort, u16)
-OPENCV_HAL_IMPL_MSA_INIT(int16x8, short, s16)
-OPENCV_HAL_IMPL_MSA_INIT(uint32x4, unsigned int, u32)
-OPENCV_HAL_IMPL_MSA_INIT(int32x4, int, s32)
-OPENCV_HAL_IMPL_MSA_INIT(uint64x2, uint64, u64)
-OPENCV_HAL_IMPL_MSA_INIT(int64x2, int64, s64)
-OPENCV_HAL_IMPL_MSA_INIT(float32x4, float, f32)
-OPENCV_HAL_IMPL_MSA_INIT(float64x2, double, f64)
-
-#define OPENCV_HAL_IMPL_MSA_PACK(_Tpvec, _Tpwvec, pack, mov, rshr) \
-inline _Tpvec v_##pack(const _Tpwvec& a, const _Tpwvec& b) \
-{ \
-    return _Tpvec(mov(a.val, b.val)); \
-} \
-template<int n> inline \
-_Tpvec v_rshr_##pack(const _Tpwvec& a, const _Tpwvec& b) \
-{ \
-    return _Tpvec(rshr(a.val, b.val, n)); \
-}
-
-OPENCV_HAL_IMPL_MSA_PACK(v_uint8x16, v_uint16x8, pack, msa_qpack_u16, msa_qrpackr_u16)
-OPENCV_HAL_IMPL_MSA_PACK(v_int8x16, v_int16x8, pack, msa_qpack_s16, msa_qrpackr_s16)
-OPENCV_HAL_IMPL_MSA_PACK(v_uint16x8, v_uint32x4, pack, msa_qpack_u32, msa_qrpackr_u32)
-OPENCV_HAL_IMPL_MSA_PACK(v_int16x8, v_int32x4, pack, msa_qpack_s32, msa_qrpackr_s32)
-OPENCV_HAL_IMPL_MSA_PACK(v_uint32x4, v_uint64x2, pack, msa_pack_u64, msa_rpackr_u64)
-OPENCV_HAL_IMPL_MSA_PACK(v_int32x4, v_int64x2, pack, msa_pack_s64, msa_rpackr_s64)
-OPENCV_HAL_IMPL_MSA_PACK(v_uint8x16, v_int16x8, pack_u, msa_qpacku_s16, msa_qrpackru_s16)
-OPENCV_HAL_IMPL_MSA_PACK(v_uint16x8, v_int32x4, pack_u, msa_qpacku_s32, msa_qrpackru_s32)
-
-#define OPENCV_HAL_IMPL_MSA_PACK_STORE(_Tpvec, _Tp, hreg, suffix, _Tpwvec, pack, mov, rshr) \
-inline void v_##pack##_store(_Tp* ptr, const _Tpwvec& a) \
-{ \
-    hreg a1 = mov(a.val); \
-    msa_st1_##suffix(ptr, a1); \
-} \
-template<int n> inline \
-void v_rshr_##pack##_store(_Tp* ptr, const _Tpwvec& a) \
-{ \
-    hreg a1 = rshr(a.val, n); \
-    msa_st1_##suffix(ptr, a1); \
-}
-
-OPENCV_HAL_IMPL_MSA_PACK_STORE(v_uint8x16, uchar, v8u8, u8, v_uint16x8, pack, msa_qmovn_u16, msa_qrshrn_n_u16)
-OPENCV_HAL_IMPL_MSA_PACK_STORE(v_int8x16, schar, v8i8, s8, v_int16x8, pack, msa_qmovn_s16, msa_qrshrn_n_s16)
-OPENCV_HAL_IMPL_MSA_PACK_STORE(v_uint16x8, ushort, v4u16, u16, v_uint32x4, pack, msa_qmovn_u32, msa_qrshrn_n_u32)
-OPENCV_HAL_IMPL_MSA_PACK_STORE(v_int16x8, short, v4i16, s16, v_int32x4, pack, msa_qmovn_s32, msa_qrshrn_n_s32)
-OPENCV_HAL_IMPL_MSA_PACK_STORE(v_uint32x4, unsigned, v2u32, u32, v_uint64x2, pack, msa_movn_u64, msa_rshrn_n_u64)
-OPENCV_HAL_IMPL_MSA_PACK_STORE(v_int32x4, int, v2i32, s32, v_int64x2, pack, msa_movn_s64, msa_rshrn_n_s64)
-OPENCV_HAL_IMPL_MSA_PACK_STORE(v_uint8x16, uchar, v8u8, u8, v_int16x8, pack_u, msa_qmovun_s16, msa_qrshrun_n_s16)
-OPENCV_HAL_IMPL_MSA_PACK_STORE(v_uint16x8, ushort, v4u16, u16, v_int32x4, pack_u, msa_qmovun_s32, msa_qrshrun_n_s32)
-
-// pack boolean
-inline v_uint8x16 v_pack_b(const v_uint16x8& a, const v_uint16x8& b)
-{
-    return v_uint8x16(msa_pack_u16(a.val, b.val));
-}
-
-inline v_uint8x16 v_pack_b(const v_uint32x4& a, const v_uint32x4& b,
-                           const v_uint32x4& c, const v_uint32x4& d)
-{
-    return v_uint8x16(msa_pack_u16(msa_pack_u32(a.val, b.val), msa_pack_u32(c.val, d.val)));
-}
-
-inline v_uint8x16 v_pack_b(const v_uint64x2& a, const v_uint64x2& b, const v_uint64x2& c,
-                           const v_uint64x2& d, const v_uint64x2& e, const v_uint64x2& f,
-                           const v_uint64x2& g, const v_uint64x2& h)
-{
-    v8u16 abcd = msa_pack_u32(msa_pack_u64(a.val, b.val), msa_pack_u64(c.val, d.val));
-    v8u16 efgh = msa_pack_u32(msa_pack_u64(e.val, f.val), msa_pack_u64(g.val, h.val));
-    return v_uint8x16(msa_pack_u16(abcd, efgh));
-}
-
-inline v_float32x4 v_matmul(const v_float32x4& v, const v_float32x4& m0,
-                            const v_float32x4& m1, const v_float32x4& m2,
-                            const v_float32x4& m3)
-{
-    v4f32 v0 = v.val;
-    v4f32 res = msa_mulq_lane_f32(m0.val, v0, 0);
-    res = msa_mlaq_lane_f32(res, m1.val, v0, 1);
-    res = msa_mlaq_lane_f32(res, m2.val, v0, 2);
-    res = msa_mlaq_lane_f32(res, m3.val, v0, 3);
-    return v_float32x4(res);
-}
-
-inline v_float32x4 v_matmuladd(const v_float32x4& v, const v_float32x4& m0,
-                               const v_float32x4& m1, const v_float32x4& m2,
-                               const v_float32x4& a)
-{
-    v4f32 v0 = v.val;
-    v4f32 res = msa_mulq_lane_f32(m0.val, v0, 0);
-    res = msa_mlaq_lane_f32(res, m1.val, v0, 1);
-    res = msa_mlaq_lane_f32(res, m2.val, v0, 2);
-    res = msa_addq_f32(res, a.val);
-    return v_float32x4(res);
-}
-
-#define OPENCV_HAL_IMPL_MSA_BIN_OP(bin_op, _Tpvec, intrin) \
-inline _Tpvec operator bin_op (const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(intrin(a.val, b.val)); \
-} \
-inline _Tpvec& operator bin_op##= (_Tpvec& a, const _Tpvec& b) \
-{ \
-    a.val = intrin(a.val, b.val); \
-    return a; \
-}
-
-OPENCV_HAL_IMPL_MSA_BIN_OP(+, v_uint8x16, msa_qaddq_u8)
-OPENCV_HAL_IMPL_MSA_BIN_OP(-, v_uint8x16, msa_qsubq_u8)
-OPENCV_HAL_IMPL_MSA_BIN_OP(+, v_int8x16, msa_qaddq_s8)
-OPENCV_HAL_IMPL_MSA_BIN_OP(-, v_int8x16, msa_qsubq_s8)
-OPENCV_HAL_IMPL_MSA_BIN_OP(+, v_uint16x8, msa_qaddq_u16)
-OPENCV_HAL_IMPL_MSA_BIN_OP(-, v_uint16x8, msa_qsubq_u16)
-OPENCV_HAL_IMPL_MSA_BIN_OP(+, v_int16x8, msa_qaddq_s16)
-OPENCV_HAL_IMPL_MSA_BIN_OP(-, v_int16x8, msa_qsubq_s16)
-OPENCV_HAL_IMPL_MSA_BIN_OP(+, v_int32x4, msa_addq_s32)
-OPENCV_HAL_IMPL_MSA_BIN_OP(-, v_int32x4, msa_subq_s32)
-OPENCV_HAL_IMPL_MSA_BIN_OP(*, v_int32x4, msa_mulq_s32)
-OPENCV_HAL_IMPL_MSA_BIN_OP(+, v_uint32x4, msa_addq_u32)
-OPENCV_HAL_IMPL_MSA_BIN_OP(-, v_uint32x4, msa_subq_u32)
-OPENCV_HAL_IMPL_MSA_BIN_OP(*, v_uint32x4, msa_mulq_u32)
-OPENCV_HAL_IMPL_MSA_BIN_OP(+, v_float32x4, msa_addq_f32)
-OPENCV_HAL_IMPL_MSA_BIN_OP(-, v_float32x4, msa_subq_f32)
-OPENCV_HAL_IMPL_MSA_BIN_OP(*, v_float32x4, msa_mulq_f32)
-OPENCV_HAL_IMPL_MSA_BIN_OP(+, v_int64x2, msa_addq_s64)
-OPENCV_HAL_IMPL_MSA_BIN_OP(-, v_int64x2, msa_subq_s64)
-OPENCV_HAL_IMPL_MSA_BIN_OP(+, v_uint64x2, msa_addq_u64)
-OPENCV_HAL_IMPL_MSA_BIN_OP(-, v_uint64x2, msa_subq_u64)
-OPENCV_HAL_IMPL_MSA_BIN_OP(/, v_float32x4, msa_divq_f32)
-OPENCV_HAL_IMPL_MSA_BIN_OP(+, v_float64x2, msa_addq_f64)
-OPENCV_HAL_IMPL_MSA_BIN_OP(-, v_float64x2, msa_subq_f64)
-OPENCV_HAL_IMPL_MSA_BIN_OP(*, v_float64x2, msa_mulq_f64)
-OPENCV_HAL_IMPL_MSA_BIN_OP(/, v_float64x2, msa_divq_f64)
-
-// saturating multiply 8-bit, 16-bit
-#define OPENCV_HAL_IMPL_MSA_MUL_SAT(_Tpvec, _Tpwvec)         \
-inline _Tpvec operator * (const _Tpvec& a, const _Tpvec& b)  \
-{                                                            \
-    _Tpwvec c, d;                                            \
-    v_mul_expand(a, b, c, d);                                \
-    return v_pack(c, d);                                     \
-}                                                            \
-inline _Tpvec& operator *= (_Tpvec& a, const _Tpvec& b)      \
-{a = a * b; return a; }
-
-OPENCV_HAL_IMPL_MSA_MUL_SAT(v_int8x16,  v_int16x8)
-OPENCV_HAL_IMPL_MSA_MUL_SAT(v_uint8x16, v_uint16x8)
-OPENCV_HAL_IMPL_MSA_MUL_SAT(v_int16x8,  v_int32x4)
-OPENCV_HAL_IMPL_MSA_MUL_SAT(v_uint16x8, v_uint32x4)
-
-//  Multiply and expand
-inline void v_mul_expand(const v_int8x16& a, const v_int8x16& b,
-                         v_int16x8& c, v_int16x8& d)
-{
-    v16i8 a_lo, a_hi, b_lo, b_hi;
-
-    ILVRL_B2_SB(a.val, msa_dupq_n_s8(0), a_lo, a_hi);
-    ILVRL_B2_SB(b.val, msa_dupq_n_s8(0), b_lo, b_hi);
-    c.val = msa_mulq_s16(msa_paddlq_s8(a_lo), msa_paddlq_s8(b_lo));
-    d.val = msa_mulq_s16(msa_paddlq_s8(a_hi), msa_paddlq_s8(b_hi));
-}
-
-inline void v_mul_expand(const v_uint8x16& a, const v_uint8x16& b,
-                         v_uint16x8& c, v_uint16x8& d)
-{
-    v16u8 a_lo, a_hi, b_lo, b_hi;
-
-    ILVRL_B2_UB(a.val, msa_dupq_n_u8(0), a_lo, a_hi);
-    ILVRL_B2_UB(b.val, msa_dupq_n_u8(0), b_lo, b_hi);
-    c.val = msa_mulq_u16(msa_paddlq_u8(a_lo), msa_paddlq_u8(b_lo));
-    d.val = msa_mulq_u16(msa_paddlq_u8(a_hi), msa_paddlq_u8(b_hi));
-}
-
-inline void v_mul_expand(const v_int16x8& a, const v_int16x8& b,
-                         v_int32x4& c, v_int32x4& d)
-{
-    v8i16 a_lo, a_hi, b_lo, b_hi;
-
-    ILVRL_H2_SH(a.val, msa_dupq_n_s16(0), a_lo, a_hi);
-    ILVRL_H2_SH(b.val, msa_dupq_n_s16(0), b_lo, b_hi);
-    c.val = msa_mulq_s32(msa_paddlq_s16(a_lo), msa_paddlq_s16(b_lo));
-    d.val = msa_mulq_s32(msa_paddlq_s16(a_hi), msa_paddlq_s16(b_hi));
-}
-
-inline void v_mul_expand(const v_uint16x8& a, const v_uint16x8& b,
-                         v_uint32x4& c, v_uint32x4& d)
-{
-    v8u16 a_lo, a_hi, b_lo, b_hi;
-
-    ILVRL_H2_UH(a.val, msa_dupq_n_u16(0), a_lo, a_hi);
-    ILVRL_H2_UH(b.val, msa_dupq_n_u16(0), b_lo, b_hi);
-    c.val = msa_mulq_u32(msa_paddlq_u16(a_lo), msa_paddlq_u16(b_lo));
-    d.val = msa_mulq_u32(msa_paddlq_u16(a_hi), msa_paddlq_u16(b_hi));
-}
-
-inline void v_mul_expand(const v_uint32x4& a, const v_uint32x4& b,
-                         v_uint64x2& c, v_uint64x2& d)
-{
-    v4u32 a_lo, a_hi, b_lo, b_hi;
-
-    ILVRL_W2_UW(a.val, msa_dupq_n_u32(0), a_lo, a_hi);
-    ILVRL_W2_UW(b.val, msa_dupq_n_u32(0), b_lo, b_hi);
-    c.val = msa_mulq_u64(msa_paddlq_u32(a_lo), msa_paddlq_u32(b_lo));
-    d.val = msa_mulq_u64(msa_paddlq_u32(a_hi), msa_paddlq_u32(b_hi));
-}
-
-inline v_int16x8 v_mul_hi(const v_int16x8& a, const v_int16x8& b)
-{
-    v8i16 a_lo, a_hi, b_lo, b_hi;
-
-    ILVRL_H2_SH(a.val, msa_dupq_n_s16(0), a_lo, a_hi);
-    ILVRL_H2_SH(b.val, msa_dupq_n_s16(0), b_lo, b_hi);
-
-    return v_int16x8(msa_packr_s32(msa_mulq_s32(msa_paddlq_s16(a_lo), msa_paddlq_s16(b_lo)),
-                                   msa_mulq_s32(msa_paddlq_s16(a_hi), msa_paddlq_s16(b_hi)), 16));
-}
-
-inline v_uint16x8 v_mul_hi(const v_uint16x8& a, const v_uint16x8& b)
-{
-    v8u16 a_lo, a_hi, b_lo, b_hi;
-
-    ILVRL_H2_UH(a.val, msa_dupq_n_u16(0), a_lo, a_hi);
-    ILVRL_H2_UH(b.val, msa_dupq_n_u16(0), b_lo, b_hi);
-
-    return v_uint16x8(msa_packr_u32(msa_mulq_u32(msa_paddlq_u16(a_lo), msa_paddlq_u16(b_lo)),
-                                    msa_mulq_u32(msa_paddlq_u16(a_hi), msa_paddlq_u16(b_hi)), 16));
-}
-
-//////// Dot Product ////////
-
-// 16 >> 32
-inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b)
-{ return v_int32x4(msa_dotp_s_w(a.val, b.val)); }
-inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b, const v_int32x4& c)
-{ return v_int32x4(msa_dpadd_s_w(c.val , a.val, b.val)); }
-
-// 32 >> 64
-inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b)
-{ return v_int64x2(msa_dotp_s_d(a.val, b.val)); }
-inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b, const v_int64x2& c)
-{ return v_int64x2(msa_dpadd_s_d(c.val , a.val, b.val)); }
-
-// 8 >> 32
-inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b)
-{
-    v8u16 even_a = msa_shrq_n_u16(msa_shlq_n_u16(MSA_TPV_REINTERPRET(v8u16, a.val), 8), 8);
-    v8u16 odd_a  = msa_shrq_n_u16(MSA_TPV_REINTERPRET(v8u16, a.val), 8);
-    v8u16 even_b = msa_shrq_n_u16(msa_shlq_n_u16(MSA_TPV_REINTERPRET(v8u16, b.val), 8), 8);
-    v8u16 odd_b  = msa_shrq_n_u16(MSA_TPV_REINTERPRET(v8u16, b.val), 8);
-    v4u32 prod   = msa_dotp_u_w(even_a, even_b);
-    return v_uint32x4(msa_dpadd_u_w(prod, odd_a, odd_b));
-}
-inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b, const v_uint32x4& c)
-{
-    v8u16 even_a = msa_shrq_n_u16(msa_shlq_n_u16(MSA_TPV_REINTERPRET(v8u16, a.val), 8), 8);
-    v8u16 odd_a  = msa_shrq_n_u16(MSA_TPV_REINTERPRET(v8u16, a.val), 8);
-    v8u16 even_b = msa_shrq_n_u16(msa_shlq_n_u16(MSA_TPV_REINTERPRET(v8u16, b.val), 8), 8);
-    v8u16 odd_b  = msa_shrq_n_u16(MSA_TPV_REINTERPRET(v8u16, b.val), 8);
-    v4u32 prod   = msa_dpadd_u_w(c.val, even_a, even_b);
-    return v_uint32x4(msa_dpadd_u_w(prod, odd_a, odd_b));
-}
-
-inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b)
-{
-    v8i16 prod = msa_dotp_s_h(a.val, b.val);
-    return v_int32x4(msa_hadd_s32(prod, prod));
-}
-inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b,
-                                  const v_int32x4& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-// 16 >> 64
-inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b)
-{
-    v4u32 even_a = msa_shrq_n_u32(msa_shlq_n_u32(MSA_TPV_REINTERPRET(v4u32, a.val), 16), 16);
-    v4u32 odd_a  = msa_shrq_n_u32(MSA_TPV_REINTERPRET(v4u32, a.val), 16);
-    v4u32 even_b = msa_shrq_n_u32(msa_shlq_n_u32(MSA_TPV_REINTERPRET(v4u32, b.val), 16), 16);
-    v4u32 odd_b  = msa_shrq_n_u32(MSA_TPV_REINTERPRET(v4u32, b.val), 16);
-    v2u64 prod   = msa_dotp_u_d(even_a, even_b);
-    return v_uint64x2(msa_dpadd_u_d(prod, odd_a, odd_b));
-}
-inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b,
-                                   const v_uint64x2& c)
-{
-    v4u32 even_a = msa_shrq_n_u32(msa_shlq_n_u32(MSA_TPV_REINTERPRET(v4u32, a.val), 16), 16);
-    v4u32 odd_a  = msa_shrq_n_u32(MSA_TPV_REINTERPRET(v4u32, a.val), 16);
-    v4u32 even_b = msa_shrq_n_u32(msa_shlq_n_u32(MSA_TPV_REINTERPRET(v4u32, b.val), 16), 16);
-    v4u32 odd_b  = msa_shrq_n_u32(MSA_TPV_REINTERPRET(v4u32, b.val), 16);
-    v2u64 prod   = msa_dpadd_u_d(c.val, even_a, even_b);
-    return v_uint64x2(msa_dpadd_u_d(prod, odd_a, odd_b));
-}
-
-inline v_int64x2 v_dotprod_expand(const v_int16x8& a, const v_int16x8& b)
-{
-    v4i32 prod = msa_dotp_s_w(a.val, b.val);
-    return v_int64x2(msa_hadd_s64(prod, prod));
-}
-inline v_int64x2 v_dotprod_expand(const v_int16x8& a, const v_int16x8& b, const v_int64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-// 32 >> 64f
-inline v_float64x2 v_dotprod_expand(const v_int32x4& a, const v_int32x4& b)
-{ return v_cvt_f64(v_dotprod(a, b)); }
-inline v_float64x2 v_dotprod_expand(const v_int32x4& a, const v_int32x4& b, const v_float64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-
-//////// Fast Dot Product ////////
-
-// 16 >> 32
-inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b)
-{ return v_dotprod(a, b); }
-inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b, const v_int32x4& c)
-{ return v_dotprod(a, b, c); }
-
-// 32 >> 64
-inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b)
-{ return v_dotprod(a, b); }
-inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b, const v_int64x2& c)
-{ return v_dotprod(a, b, c); }
-
-// 8 >> 32
-inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b)
-{ return v_dotprod_expand(a, b); }
-inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b, const v_uint32x4& c)
-{ return v_dotprod_expand(a, b, c); }
-inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b)
-{ return v_dotprod_expand(a, b); }
-inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b, const v_int32x4& c)
-{ return v_dotprod_expand(a, b, c); }
-
-// 16 >> 64
-inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b)
-{ return v_dotprod_expand(a, b); }
-inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b, const v_uint64x2& c)
-{ return v_dotprod_expand(a, b, c); }
-inline v_int64x2 v_dotprod_expand_fast(const v_int16x8& a, const v_int16x8& b)
-{ return v_dotprod_expand(a, b); }
-inline v_int64x2 v_dotprod_expand_fast(const v_int16x8& a, const v_int16x8& b, const v_int64x2& c)
-{ return v_dotprod_expand(a, b, c); }
-
-// 32 >> 64f
-inline v_float64x2 v_dotprod_expand_fast(const v_int32x4& a, const v_int32x4& b)
-{ return v_dotprod_expand(a, b); }
-inline v_float64x2 v_dotprod_expand_fast(const v_int32x4& a, const v_int32x4& b, const v_float64x2& c)
-{ return v_dotprod_expand(a, b, c); }
-
-#define OPENCV_HAL_IMPL_MSA_LOGIC_OP(_Tpvec, _Tpv, suffix) \
-OPENCV_HAL_IMPL_MSA_BIN_OP(&, _Tpvec, msa_andq_##suffix)   \
-OPENCV_HAL_IMPL_MSA_BIN_OP(|, _Tpvec, msa_orrq_##suffix)   \
-OPENCV_HAL_IMPL_MSA_BIN_OP(^, _Tpvec, msa_eorq_##suffix)   \
-inline _Tpvec operator ~ (const _Tpvec& a) \
-{ \
-    return _Tpvec(MSA_TPV_REINTERPRET(_Tpv, msa_mvnq_u8(MSA_TPV_REINTERPRET(v16u8, a.val)))); \
-}
-
-OPENCV_HAL_IMPL_MSA_LOGIC_OP(v_uint8x16, v16u8, u8)
-OPENCV_HAL_IMPL_MSA_LOGIC_OP(v_int8x16, v16i8, s8)
-OPENCV_HAL_IMPL_MSA_LOGIC_OP(v_uint16x8, v8u16, u16)
-OPENCV_HAL_IMPL_MSA_LOGIC_OP(v_int16x8, v8i16, s16)
-OPENCV_HAL_IMPL_MSA_LOGIC_OP(v_uint32x4, v4u32, u32)
-OPENCV_HAL_IMPL_MSA_LOGIC_OP(v_int32x4, v4i32, s32)
-OPENCV_HAL_IMPL_MSA_LOGIC_OP(v_uint64x2, v2u64, u64)
-OPENCV_HAL_IMPL_MSA_LOGIC_OP(v_int64x2, v2i64, s64)
-
-#define OPENCV_HAL_IMPL_MSA_FLT_BIT_OP(bin_op, intrin) \
-inline v_float32x4 operator bin_op (const v_float32x4& a, const v_float32x4& b) \
-{ \
-    return v_float32x4(MSA_TPV_REINTERPRET(v4f32, intrin(MSA_TPV_REINTERPRET(v4i32, a.val), MSA_TPV_REINTERPRET(v4i32, b.val)))); \
-} \
-inline v_float32x4& operator bin_op##= (v_float32x4& a, const v_float32x4& b) \
-{ \
-    a.val = MSA_TPV_REINTERPRET(v4f32, intrin(MSA_TPV_REINTERPRET(v4i32, a.val), MSA_TPV_REINTERPRET(v4i32, b.val))); \
-    return a; \
-}
-
-OPENCV_HAL_IMPL_MSA_FLT_BIT_OP(&, msa_andq_s32)
-OPENCV_HAL_IMPL_MSA_FLT_BIT_OP(|, msa_orrq_s32)
-OPENCV_HAL_IMPL_MSA_FLT_BIT_OP(^, msa_eorq_s32)
-
-inline v_float32x4 operator ~ (const v_float32x4& a)
-{
-    return v_float32x4(MSA_TPV_REINTERPRET(v4f32, msa_mvnq_s32(MSA_TPV_REINTERPRET(v4i32, a.val))));
-}
-
-/* v_abs */
-#define OPENCV_HAL_IMPL_MSA_ABS(_Tpuvec, _Tpsvec, usuffix, ssuffix) \
-inline _Tpuvec v_abs(const _Tpsvec& a) \
-{ \
-    return v_reinterpret_as_##usuffix(_Tpsvec(msa_absq_##ssuffix(a.val))); \
-}
-
-OPENCV_HAL_IMPL_MSA_ABS(v_uint8x16, v_int8x16, u8, s8)
-OPENCV_HAL_IMPL_MSA_ABS(v_uint16x8, v_int16x8, u16, s16)
-OPENCV_HAL_IMPL_MSA_ABS(v_uint32x4, v_int32x4, u32, s32)
-
-/* v_abs(float), v_sqrt, v_invsqrt */
-#define OPENCV_HAL_IMPL_MSA_BASIC_FUNC(_Tpvec, func, intrin) \
-inline _Tpvec func(const _Tpvec& a) \
-{ \
-    return _Tpvec(intrin(a.val)); \
-}
-
-OPENCV_HAL_IMPL_MSA_BASIC_FUNC(v_float32x4, v_abs, msa_absq_f32)
-OPENCV_HAL_IMPL_MSA_BASIC_FUNC(v_float64x2, v_abs, msa_absq_f64)
-OPENCV_HAL_IMPL_MSA_BASIC_FUNC(v_float32x4, v_sqrt, msa_sqrtq_f32)
-OPENCV_HAL_IMPL_MSA_BASIC_FUNC(v_float32x4, v_invsqrt, msa_rsqrtq_f32)
-OPENCV_HAL_IMPL_MSA_BASIC_FUNC(v_float64x2, v_sqrt, msa_sqrtq_f64)
-OPENCV_HAL_IMPL_MSA_BASIC_FUNC(v_float64x2, v_invsqrt, msa_rsqrtq_f64)
-
-#define OPENCV_HAL_IMPL_MSA_DBL_BIT_OP(bin_op, intrin) \
-inline v_float64x2 operator bin_op (const v_float64x2& a, const v_float64x2& b) \
-{ \
-    return v_float64x2(MSA_TPV_REINTERPRET(v2f64, intrin(MSA_TPV_REINTERPRET(v2i64, a.val), MSA_TPV_REINTERPRET(v2i64, b.val)))); \
-} \
-inline v_float64x2& operator bin_op##= (v_float64x2& a, const v_float64x2& b) \
-{ \
-    a.val = MSA_TPV_REINTERPRET(v2f64, intrin(MSA_TPV_REINTERPRET(v2i64, a.val), MSA_TPV_REINTERPRET(v2i64, b.val))); \
-    return a; \
-}
-
-OPENCV_HAL_IMPL_MSA_DBL_BIT_OP(&, msa_andq_s64)
-OPENCV_HAL_IMPL_MSA_DBL_BIT_OP(|, msa_orrq_s64)
-OPENCV_HAL_IMPL_MSA_DBL_BIT_OP(^, msa_eorq_s64)
-
-inline v_float64x2 operator ~ (const v_float64x2& a)
-{
-    return v_float64x2(MSA_TPV_REINTERPRET(v2f64, msa_mvnq_s32(MSA_TPV_REINTERPRET(v4i32, a.val))));
-}
-
-// TODO: exp, log, sin, cos
-
-#define OPENCV_HAL_IMPL_MSA_BIN_FUNC(_Tpvec, func, intrin) \
-inline _Tpvec func(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(intrin(a.val, b.val)); \
-}
-
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_uint8x16, v_min, msa_minq_u8)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_uint8x16, v_max, msa_maxq_u8)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_int8x16, v_min, msa_minq_s8)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_int8x16, v_max, msa_maxq_s8)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_uint16x8, v_min, msa_minq_u16)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_uint16x8, v_max, msa_maxq_u16)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_int16x8, v_min, msa_minq_s16)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_int16x8, v_max, msa_maxq_s16)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_uint32x4, v_min, msa_minq_u32)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_uint32x4, v_max, msa_maxq_u32)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_int32x4, v_min, msa_minq_s32)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_int32x4, v_max, msa_maxq_s32)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_float32x4, v_min, msa_minq_f32)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_float32x4, v_max, msa_maxq_f32)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_float64x2, v_min, msa_minq_f64)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_float64x2, v_max, msa_maxq_f64)
-
-#define OPENCV_HAL_IMPL_MSA_INT_CMP_OP(_Tpvec, _Tpv, suffix, not_suffix) \
-inline _Tpvec operator == (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(MSA_TPV_REINTERPRET(_Tpv, msa_ceqq_##suffix(a.val, b.val))); } \
-inline _Tpvec operator != (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(MSA_TPV_REINTERPRET(_Tpv, msa_mvnq_##not_suffix(msa_ceqq_##suffix(a.val, b.val)))); } \
-inline _Tpvec operator < (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(MSA_TPV_REINTERPRET(_Tpv, msa_cltq_##suffix(a.val, b.val))); } \
-inline _Tpvec operator > (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(MSA_TPV_REINTERPRET(_Tpv, msa_cgtq_##suffix(a.val, b.val))); } \
-inline _Tpvec operator <= (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(MSA_TPV_REINTERPRET(_Tpv, msa_cleq_##suffix(a.val, b.val))); } \
-inline _Tpvec operator >= (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(MSA_TPV_REINTERPRET(_Tpv, msa_cgeq_##suffix(a.val, b.val))); }
-
-OPENCV_HAL_IMPL_MSA_INT_CMP_OP(v_uint8x16, v16u8, u8, u8)
-OPENCV_HAL_IMPL_MSA_INT_CMP_OP(v_int8x16, v16i8, s8, u8)
-OPENCV_HAL_IMPL_MSA_INT_CMP_OP(v_uint16x8, v8u16, u16, u16)
-OPENCV_HAL_IMPL_MSA_INT_CMP_OP(v_int16x8, v8i16, s16, u16)
-OPENCV_HAL_IMPL_MSA_INT_CMP_OP(v_uint32x4, v4u32, u32, u32)
-OPENCV_HAL_IMPL_MSA_INT_CMP_OP(v_int32x4, v4i32, s32, u32)
-OPENCV_HAL_IMPL_MSA_INT_CMP_OP(v_float32x4, v4f32, f32, u32)
-OPENCV_HAL_IMPL_MSA_INT_CMP_OP(v_uint64x2, v2u64, u64, u64)
-OPENCV_HAL_IMPL_MSA_INT_CMP_OP(v_int64x2, v2i64, s64, u64)
-OPENCV_HAL_IMPL_MSA_INT_CMP_OP(v_float64x2, v2f64, f64, u64)
-
-inline v_float32x4 v_not_nan(const v_float32x4& a)
-{ return v_float32x4(MSA_TPV_REINTERPRET(v4f32, msa_ceqq_f32(a.val, a.val))); }
-inline v_float64x2 v_not_nan(const v_float64x2& a)
-{ return v_float64x2(MSA_TPV_REINTERPRET(v2f64, msa_ceqq_f64(a.val, a.val))); }
-
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_uint8x16, v_add_wrap, msa_addq_u8)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_int8x16, v_add_wrap, msa_addq_s8)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_uint16x8, v_add_wrap, msa_addq_u16)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_int16x8, v_add_wrap, msa_addq_s16)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_uint8x16, v_sub_wrap, msa_subq_u8)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_int8x16, v_sub_wrap, msa_subq_s8)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_uint16x8, v_sub_wrap, msa_subq_u16)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_int16x8, v_sub_wrap, msa_subq_s16)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_uint8x16, v_mul_wrap, msa_mulq_u8)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_int8x16, v_mul_wrap, msa_mulq_s8)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_uint16x8, v_mul_wrap, msa_mulq_u16)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_int16x8, v_mul_wrap, msa_mulq_s16)
-
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_uint8x16, v_absdiff, msa_abdq_u8)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_uint16x8, v_absdiff, msa_abdq_u16)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_uint32x4, v_absdiff, msa_abdq_u32)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_float32x4, v_absdiff, msa_abdq_f32)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_float64x2, v_absdiff, msa_abdq_f64)
-
-/** Saturating absolute difference **/
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_int8x16, v_absdiffs, msa_qabdq_s8)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC(v_int16x8, v_absdiffs, msa_qabdq_s16)
-
-#define OPENCV_HAL_IMPL_MSA_BIN_FUNC2(_Tpvec, _Tpvec2, _Tpv, func, intrin) \
-inline _Tpvec2 func(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec2(MSA_TPV_REINTERPRET(_Tpv, intrin(a.val, b.val))); \
-}
-
-OPENCV_HAL_IMPL_MSA_BIN_FUNC2(v_int8x16, v_uint8x16, v16u8, v_absdiff, msa_abdq_s8)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC2(v_int16x8, v_uint16x8, v8u16, v_absdiff, msa_abdq_s16)
-OPENCV_HAL_IMPL_MSA_BIN_FUNC2(v_int32x4, v_uint32x4, v4u32, v_absdiff, msa_abdq_s32)
-
-/* v_magnitude, v_sqr_magnitude, v_fma, v_muladd */
-inline v_float32x4 v_magnitude(const v_float32x4& a, const v_float32x4& b)
-{
-    v_float32x4 x(msa_mlaq_f32(msa_mulq_f32(a.val, a.val), b.val, b.val));
-    return v_sqrt(x);
-}
-
-inline v_float32x4 v_sqr_magnitude(const v_float32x4& a, const v_float32x4& b)
-{
-    return v_float32x4(msa_mlaq_f32(msa_mulq_f32(a.val, a.val), b.val, b.val));
-}
-
-inline v_float32x4 v_fma(const v_float32x4& a, const v_float32x4& b, const v_float32x4& c)
-{
-    return v_float32x4(msa_mlaq_f32(c.val, a.val, b.val));
-}
-
-inline v_int32x4 v_fma(const v_int32x4& a, const v_int32x4& b, const v_int32x4& c)
-{
-    return v_int32x4(msa_mlaq_s32(c.val, a.val, b.val));
-}
-
-inline v_float32x4 v_muladd(const v_float32x4& a, const v_float32x4& b, const v_float32x4& c)
-{
-    return v_fma(a, b, c);
-}
-
-inline v_int32x4 v_muladd(const v_int32x4& a, const v_int32x4& b, const v_int32x4& c)
-{
-    return v_fma(a, b, c);
-}
-
-inline v_float64x2 v_magnitude(const v_float64x2& a, const v_float64x2& b)
-{
-    v_float64x2 x(msa_mlaq_f64(msa_mulq_f64(a.val, a.val), b.val, b.val));
-    return v_sqrt(x);
-}
-
-inline v_float64x2 v_sqr_magnitude(const v_float64x2& a, const v_float64x2& b)
-{
-    return v_float64x2(msa_mlaq_f64(msa_mulq_f64(a.val, a.val), b.val, b.val));
-}
-
-inline v_float64x2 v_fma(const v_float64x2& a, const v_float64x2& b, const v_float64x2& c)
-{
-    return v_float64x2(msa_mlaq_f64(c.val, a.val, b.val));
-}
-
-inline v_float64x2 v_muladd(const v_float64x2& a, const v_float64x2& b, const v_float64x2& c)
-{
-    return v_fma(a, b, c);
-}
-
-// trade efficiency for convenience
-#define OPENCV_HAL_IMPL_MSA_SHIFT_OP(_Tpvec, suffix, _Tps, ssuffix) \
-inline _Tpvec operator << (const _Tpvec& a, int n) \
-{ return _Tpvec(msa_shlq_##suffix(a.val, msa_dupq_n_##ssuffix((_Tps)n))); } \
-inline _Tpvec operator >> (const _Tpvec& a, int n) \
-{ return _Tpvec(msa_shrq_##suffix(a.val, msa_dupq_n_##ssuffix((_Tps)n))); } \
-template<int n> inline _Tpvec v_shl(const _Tpvec& a) \
-{ return _Tpvec(msa_shlq_n_##suffix(a.val, n)); } \
-template<int n> inline _Tpvec v_shr(const _Tpvec& a) \
-{ return _Tpvec(msa_shrq_n_##suffix(a.val, n)); } \
-template<int n> inline _Tpvec v_rshr(const _Tpvec& a) \
-{ return _Tpvec(msa_rshrq_n_##suffix(a.val, n)); }
-
-OPENCV_HAL_IMPL_MSA_SHIFT_OP(v_uint8x16, u8, schar, s8)
-OPENCV_HAL_IMPL_MSA_SHIFT_OP(v_int8x16, s8, schar, s8)
-OPENCV_HAL_IMPL_MSA_SHIFT_OP(v_uint16x8, u16, short, s16)
-OPENCV_HAL_IMPL_MSA_SHIFT_OP(v_int16x8, s16, short, s16)
-OPENCV_HAL_IMPL_MSA_SHIFT_OP(v_uint32x4, u32, int, s32)
-OPENCV_HAL_IMPL_MSA_SHIFT_OP(v_int32x4, s32, int, s32)
-OPENCV_HAL_IMPL_MSA_SHIFT_OP(v_uint64x2, u64, int64, s64)
-OPENCV_HAL_IMPL_MSA_SHIFT_OP(v_int64x2, s64, int64, s64)
-
-/* v_rotate_right, v_rotate_left */
-#define OPENCV_HAL_IMPL_MSA_ROTATE_OP(_Tpvec, _Tpv, _Tpvs, suffix) \
-template<int n> inline _Tpvec v_rotate_right(const _Tpvec& a) \
-{ \
-    return _Tpvec(MSA_TPV_REINTERPRET(_Tpv, msa_extq_##suffix(MSA_TPV_REINTERPRET(_Tpvs, a.val), msa_dupq_n_##suffix(0), n))); \
-} \
-template<int n> inline _Tpvec v_rotate_left(const _Tpvec& a) \
-{ \
-    return _Tpvec(MSA_TPV_REINTERPRET(_Tpv, msa_extq_##suffix(msa_dupq_n_##suffix(0), MSA_TPV_REINTERPRET(_Tpvs, a.val), _Tpvec::nlanes - n))); \
-} \
-template<> inline _Tpvec v_rotate_left<0>(const _Tpvec& a) \
-{ \
-    return a; \
-} \
-template<int n> inline _Tpvec v_rotate_right(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(MSA_TPV_REINTERPRET(_Tpv, msa_extq_##suffix(MSA_TPV_REINTERPRET(_Tpvs, a.val), MSA_TPV_REINTERPRET(_Tpvs, b.val), n))); \
-} \
-template<int n> inline _Tpvec v_rotate_left(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(MSA_TPV_REINTERPRET(_Tpv, msa_extq_##suffix(MSA_TPV_REINTERPRET(_Tpvs, b.val), MSA_TPV_REINTERPRET(_Tpvs, a.val), _Tpvec::nlanes - n))); \
-} \
-template<> inline _Tpvec v_rotate_left<0>(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    CV_UNUSED(b); \
-    return a; \
-}
-
-OPENCV_HAL_IMPL_MSA_ROTATE_OP(v_uint8x16, v16u8, v16i8, s8)
-OPENCV_HAL_IMPL_MSA_ROTATE_OP(v_int8x16, v16i8, v16i8, s8)
-OPENCV_HAL_IMPL_MSA_ROTATE_OP(v_uint16x8, v8u16, v8i16, s16)
-OPENCV_HAL_IMPL_MSA_ROTATE_OP(v_int16x8, v8i16, v8i16, s16)
-OPENCV_HAL_IMPL_MSA_ROTATE_OP(v_uint32x4, v4u32, v4i32, s32)
-OPENCV_HAL_IMPL_MSA_ROTATE_OP(v_int32x4, v4i32, v4i32, s32)
-OPENCV_HAL_IMPL_MSA_ROTATE_OP(v_float32x4, v4f32, v4i32, s32)
-OPENCV_HAL_IMPL_MSA_ROTATE_OP(v_uint64x2, v2u64, v2i64, s64)
-OPENCV_HAL_IMPL_MSA_ROTATE_OP(v_int64x2, v2i64, v2i64, s64)
-OPENCV_HAL_IMPL_MSA_ROTATE_OP(v_float64x2, v2f64, v2i64, s64)
-
-#define OPENCV_HAL_IMPL_MSA_LOADSTORE_OP(_Tpvec, _Tp, suffix) \
-inline _Tpvec v_load(const _Tp* ptr) \
-{ return _Tpvec(msa_ld1q_##suffix(ptr)); } \
-inline _Tpvec v_load_aligned(const _Tp* ptr) \
-{ return _Tpvec(msa_ld1q_##suffix(ptr)); } \
-inline _Tpvec v_load_low(const _Tp* ptr) \
-{ return _Tpvec(msa_combine_##suffix(msa_ld1_##suffix(ptr), msa_dup_n_##suffix((_Tp)0))); } \
-inline _Tpvec v_load_halves(const _Tp* ptr0, const _Tp* ptr1) \
-{ return _Tpvec(msa_combine_##suffix(msa_ld1_##suffix(ptr0), msa_ld1_##suffix(ptr1))); } \
-inline void v_store(_Tp* ptr, const _Tpvec& a) \
-{ msa_st1q_##suffix(ptr, a.val); } \
-inline void v_store_aligned(_Tp* ptr, const _Tpvec& a) \
-{ msa_st1q_##suffix(ptr, a.val); } \
-inline void v_store_aligned_nocache(_Tp* ptr, const _Tpvec& a) \
-{ msa_st1q_##suffix(ptr, a.val); } \
-inline void v_store(_Tp* ptr, const _Tpvec& a, hal::StoreMode /*mode*/) \
-{ msa_st1q_##suffix(ptr, a.val); } \
-inline void v_store_low(_Tp* ptr, const _Tpvec& a) \
-{ \
-    int n  = _Tpvec::nlanes; \
-    for( int i = 0; i < (n/2); i++ ) \
-        ptr[i] = a.val[i]; \
-} \
-inline void v_store_high(_Tp* ptr, const _Tpvec& a) \
-{ \
-    int n  = _Tpvec::nlanes; \
-    for( int i = 0; i < (n/2); i++ ) \
-        ptr[i] = a.val[i+(n/2)]; \
-}
-
-OPENCV_HAL_IMPL_MSA_LOADSTORE_OP(v_uint8x16, uchar, u8)
-OPENCV_HAL_IMPL_MSA_LOADSTORE_OP(v_int8x16, schar, s8)
-OPENCV_HAL_IMPL_MSA_LOADSTORE_OP(v_uint16x8, ushort, u16)
-OPENCV_HAL_IMPL_MSA_LOADSTORE_OP(v_int16x8, short, s16)
-OPENCV_HAL_IMPL_MSA_LOADSTORE_OP(v_uint32x4, unsigned, u32)
-OPENCV_HAL_IMPL_MSA_LOADSTORE_OP(v_int32x4, int, s32)
-OPENCV_HAL_IMPL_MSA_LOADSTORE_OP(v_uint64x2, uint64, u64)
-OPENCV_HAL_IMPL_MSA_LOADSTORE_OP(v_int64x2, int64, s64)
-OPENCV_HAL_IMPL_MSA_LOADSTORE_OP(v_float32x4, float, f32)
-OPENCV_HAL_IMPL_MSA_LOADSTORE_OP(v_float64x2, double, f64)
-
-
-/** Reverse **/
-inline v_uint8x16 v_reverse(const v_uint8x16 &a)
-{
-    v_uint8x16 c = v_uint8x16((v16u8)__builtin_msa_vshf_b((v16i8)((v2i64){0x08090A0B0C0D0E0F, 0x0001020304050607}), msa_dupq_n_s8(0), (v16i8)a.val));
-    return c;
-}
-
-inline v_int8x16 v_reverse(const v_int8x16 &a)
-{ return v_reinterpret_as_s8(v_reverse(v_reinterpret_as_u8(a))); }
-
-inline v_uint16x8 v_reverse(const v_uint16x8 &a)
-{
-    v_uint16x8 c = v_uint16x8((v8u16)__builtin_msa_vshf_h((v8i16)((v2i64){0x0004000500060007, 0x0000000100020003}), msa_dupq_n_s16(0), (v8i16)a.val));
-    return c;
-}
-
-inline v_int16x8 v_reverse(const v_int16x8 &a)
-{ return v_reinterpret_as_s16(v_reverse(v_reinterpret_as_u16(a))); }
-
-inline v_uint32x4 v_reverse(const v_uint32x4 &a)
-{
-    v_uint32x4 c;
-    c.val[0] = a.val[3];
-    c.val[1] = a.val[2];
-    c.val[2] = a.val[1];
-    c.val[3] = a.val[0];
-    return c;
-}
-
-inline v_int32x4 v_reverse(const v_int32x4 &a)
-{ return v_reinterpret_as_s32(v_reverse(v_reinterpret_as_u32(a))); }
-
-inline v_float32x4 v_reverse(const v_float32x4 &a)
-{ return v_reinterpret_as_f32(v_reverse(v_reinterpret_as_u32(a))); }
-
-inline v_uint64x2 v_reverse(const v_uint64x2 &a)
-{
-    v_uint64x2 c;
-    c.val[0] = a.val[1];
-    c.val[1] = a.val[0];
-    return c;
-}
-
-inline v_int64x2 v_reverse(const v_int64x2 &a)
-{ return v_reinterpret_as_s64(v_reverse(v_reinterpret_as_u64(a))); }
-
-inline v_float64x2 v_reverse(const v_float64x2 &a)
-{ return v_reinterpret_as_f64(v_reverse(v_reinterpret_as_u64(a))); }
-
-
-#define OPENCV_HAL_IMPL_MSA_REDUCE_OP_8U(func, cfunc) \
-inline unsigned short v_reduce_##func(const v_uint16x8& a) \
-{ \
-    v8u16 a_lo, a_hi; \
-    ILVRL_H2_UH(a.val, msa_dupq_n_u16(0), a_lo, a_hi); \
-    v4u32 b = msa_##func##q_u32(msa_paddlq_u16(a_lo), msa_paddlq_u16(a_hi)); \
-    v4u32 b_lo, b_hi; \
-    ILVRL_W2_UW(b, msa_dupq_n_u32(0), b_lo, b_hi); \
-    v2u64 c = msa_##func##q_u64(msa_paddlq_u32(b_lo), msa_paddlq_u32(b_hi)); \
-    return (unsigned short)cfunc(c[0], c[1]); \
-}
-
-OPENCV_HAL_IMPL_MSA_REDUCE_OP_8U(max, std::max)
-OPENCV_HAL_IMPL_MSA_REDUCE_OP_8U(min, std::min)
-
-#define OPENCV_HAL_IMPL_MSA_REDUCE_OP_8S(func, cfunc) \
-inline short v_reduce_##func(const v_int16x8& a) \
-{ \
-    v8i16 a_lo, a_hi; \
-    ILVRL_H2_SH(a.val, msa_dupq_n_s16(0), a_lo, a_hi); \
-    v4i32 b = msa_##func##q_s32(msa_paddlq_s16(a_lo), msa_paddlq_s16(a_hi)); \
-    v4i32 b_lo, b_hi; \
-    ILVRL_W2_SW(b, msa_dupq_n_s32(0), b_lo, b_hi); \
-    v2i64 c = msa_##func##q_s64(msa_paddlq_s32(b_lo), msa_paddlq_s32(b_hi)); \
-    return (short)cfunc(c[0], c[1]); \
-}
-
-OPENCV_HAL_IMPL_MSA_REDUCE_OP_8S(max, std::max)
-OPENCV_HAL_IMPL_MSA_REDUCE_OP_8S(min, std::min)
-
-#define OPENCV_HAL_IMPL_MSA_REDUCE_OP_4(_Tpvec, scalartype, func, cfunc) \
-inline scalartype v_reduce_##func(const _Tpvec& a) \
-{ \
-    return (scalartype)cfunc(cfunc(a.val[0], a.val[1]), cfunc(a.val[2], a.val[3])); \
-}
-
-OPENCV_HAL_IMPL_MSA_REDUCE_OP_4(v_uint32x4, unsigned, max, std::max)
-OPENCV_HAL_IMPL_MSA_REDUCE_OP_4(v_uint32x4, unsigned, min, std::min)
-OPENCV_HAL_IMPL_MSA_REDUCE_OP_4(v_int32x4, int, max, std::max)
-OPENCV_HAL_IMPL_MSA_REDUCE_OP_4(v_int32x4, int, min, std::min)
-OPENCV_HAL_IMPL_MSA_REDUCE_OP_4(v_float32x4, float, max, std::max)
-OPENCV_HAL_IMPL_MSA_REDUCE_OP_4(v_float32x4, float, min, std::min)
-
-
-#define OPENCV_HAL_IMPL_MSA_REDUCE_OP_16(_Tpvec, scalartype, _Tpvec2, func) \
-inline scalartype v_reduce_##func(const _Tpvec& a) \
-{ \
-    _Tpvec2 a1, a2; \
-    v_expand(a, a1, a2); \
-    return (scalartype)v_reduce_##func(v_##func(a1, a2)); \
-}
-
-OPENCV_HAL_IMPL_MSA_REDUCE_OP_16(v_uint8x16, uchar, v_uint16x8, min)
-OPENCV_HAL_IMPL_MSA_REDUCE_OP_16(v_uint8x16, uchar, v_uint16x8, max)
-OPENCV_HAL_IMPL_MSA_REDUCE_OP_16(v_int8x16, char, v_int16x8, min)
-OPENCV_HAL_IMPL_MSA_REDUCE_OP_16(v_int8x16, char, v_int16x8, max)
-
-
-
-#define OPENCV_HAL_IMPL_MSA_REDUCE_SUM(_Tpvec, scalartype, suffix) \
-inline scalartype v_reduce_sum(const _Tpvec& a) \
-{ \
-    return (scalartype)msa_sum_##suffix(a.val); \
-}
-
-OPENCV_HAL_IMPL_MSA_REDUCE_SUM(v_uint8x16, unsigned char, u8)
-OPENCV_HAL_IMPL_MSA_REDUCE_SUM(v_int8x16, char, s8)
-OPENCV_HAL_IMPL_MSA_REDUCE_SUM(v_uint16x8, unsigned short, u16)
-OPENCV_HAL_IMPL_MSA_REDUCE_SUM(v_int16x8, short, s16)
-OPENCV_HAL_IMPL_MSA_REDUCE_SUM(v_uint32x4, unsigned, u32)
-OPENCV_HAL_IMPL_MSA_REDUCE_SUM(v_int32x4, int, s32)
-OPENCV_HAL_IMPL_MSA_REDUCE_SUM(v_float32x4, float, f32)
-
-inline uint64 v_reduce_sum(const v_uint64x2& a)
-{ return (uint64)(msa_getq_lane_u64(a.val, 0) + msa_getq_lane_u64(a.val, 1)); }
-inline int64 v_reduce_sum(const v_int64x2& a)
-{ return (int64)(msa_getq_lane_s64(a.val, 0) + msa_getq_lane_s64(a.val, 1)); }
-inline double v_reduce_sum(const v_float64x2& a)
-{
-    return msa_getq_lane_f64(a.val, 0) + msa_getq_lane_f64(a.val, 1);
-}
-
-/* v_reduce_sum4, v_reduce_sad */
-inline v_float32x4 v_reduce_sum4(const v_float32x4& a, const v_float32x4& b,
-                                 const v_float32x4& c, const v_float32x4& d)
-{
-    v4f32 u0 = msa_addq_f32(MSA_TPV_REINTERPRET(v4f32, msa_ilvevq_s32(MSA_TPV_REINTERPRET(v4i32, b.val), MSA_TPV_REINTERPRET(v4i32, a.val))),
-                            MSA_TPV_REINTERPRET(v4f32, msa_ilvodq_s32(MSA_TPV_REINTERPRET(v4i32, b.val), MSA_TPV_REINTERPRET(v4i32, a.val)))); // a0+a1 b0+b1 a2+a3 b2+b3
-    v4f32 u1 = msa_addq_f32(MSA_TPV_REINTERPRET(v4f32, msa_ilvevq_s32(MSA_TPV_REINTERPRET(v4i32, d.val), MSA_TPV_REINTERPRET(v4i32, c.val))),
-                            MSA_TPV_REINTERPRET(v4f32, msa_ilvodq_s32(MSA_TPV_REINTERPRET(v4i32, d.val), MSA_TPV_REINTERPRET(v4i32, c.val)))); // c0+c1 d0+d1 c2+c3 d2+d3
-
-    return v_float32x4(msa_addq_f32(MSA_TPV_REINTERPRET(v4f32, msa_ilvrq_s64(MSA_TPV_REINTERPRET(v2i64, u1), MSA_TPV_REINTERPRET(v2i64, u0))),
-                                    MSA_TPV_REINTERPRET(v4f32, msa_ilvlq_s64(MSA_TPV_REINTERPRET(v2i64, u1), MSA_TPV_REINTERPRET(v2i64, u0)))));
-}
-
-inline unsigned v_reduce_sad(const v_uint8x16& a, const v_uint8x16& b)
-{
-    v16u8 t0 = msa_abdq_u8(a.val, b.val);
-    v8u16 t1 = msa_paddlq_u8(t0);
-    v4u32 t2 = msa_paddlq_u16(t1);
-    return msa_sum_u32(t2);
-}
-inline unsigned v_reduce_sad(const v_int8x16& a, const v_int8x16& b)
-{
-    v16u8 t0 = MSA_TPV_REINTERPRET(v16u8, msa_abdq_s8(a.val, b.val));
-    v8u16 t1 = msa_paddlq_u8(t0);
-    v4u32 t2 = msa_paddlq_u16(t1);
-    return msa_sum_u32(t2);
-}
-inline unsigned v_reduce_sad(const v_uint16x8& a, const v_uint16x8& b)
-{
-    v8u16 t0 = msa_abdq_u16(a.val, b.val);
-    v4u32 t1 = msa_paddlq_u16(t0);
-    return msa_sum_u32(t1);
-}
-inline unsigned v_reduce_sad(const v_int16x8& a, const v_int16x8& b)
-{
-    v8u16 t0 = MSA_TPV_REINTERPRET(v8u16, msa_abdq_s16(a.val, b.val));
-    v4u32 t1 = msa_paddlq_u16(t0);
-    return msa_sum_u32(t1);
-}
-inline unsigned v_reduce_sad(const v_uint32x4& a, const v_uint32x4& b)
-{
-    v4u32 t0 = msa_abdq_u32(a.val, b.val);
-    return msa_sum_u32(t0);
-}
-inline unsigned v_reduce_sad(const v_int32x4& a, const v_int32x4& b)
-{
-    v4u32 t0 = MSA_TPV_REINTERPRET(v4u32, msa_abdq_s32(a.val, b.val));
-    return msa_sum_u32(t0);
-}
-inline float v_reduce_sad(const v_float32x4& a, const v_float32x4& b)
-{
-    v4f32 t0 = msa_abdq_f32(a.val, b.val);
-    return msa_sum_f32(t0);
-}
-
-/* v_popcount */
-#define OPENCV_HAL_IMPL_MSA_POPCOUNT_SIZE8(_Tpvec) \
-inline v_uint8x16 v_popcount(const _Tpvec& a) \
-{ \
-    v16u8 t = MSA_TPV_REINTERPRET(v16u8, msa_cntq_s8(MSA_TPV_REINTERPRET(v16i8, a.val))); \
-    return v_uint8x16(t); \
-}
-OPENCV_HAL_IMPL_MSA_POPCOUNT_SIZE8(v_uint8x16)
-OPENCV_HAL_IMPL_MSA_POPCOUNT_SIZE8(v_int8x16)
-
-#define OPENCV_HAL_IMPL_MSA_POPCOUNT_SIZE16(_Tpvec) \
-inline v_uint16x8 v_popcount(const _Tpvec& a) \
-{ \
-    v8u16 t = MSA_TPV_REINTERPRET(v8u16, msa_cntq_s16(MSA_TPV_REINTERPRET(v8i16, a.val))); \
-    return v_uint16x8(t); \
-}
-OPENCV_HAL_IMPL_MSA_POPCOUNT_SIZE16(v_uint16x8)
-OPENCV_HAL_IMPL_MSA_POPCOUNT_SIZE16(v_int16x8)
-
-#define OPENCV_HAL_IMPL_MSA_POPCOUNT_SIZE32(_Tpvec) \
-inline v_uint32x4 v_popcount(const _Tpvec& a) \
-{ \
-    v4u32 t = MSA_TPV_REINTERPRET(v4u32, msa_cntq_s32(MSA_TPV_REINTERPRET(v4i32, a.val))); \
-    return v_uint32x4(t); \
-}
-OPENCV_HAL_IMPL_MSA_POPCOUNT_SIZE32(v_uint32x4)
-OPENCV_HAL_IMPL_MSA_POPCOUNT_SIZE32(v_int32x4)
-
-#define OPENCV_HAL_IMPL_MSA_POPCOUNT_SIZE64(_Tpvec) \
-inline v_uint64x2 v_popcount(const _Tpvec& a) \
-{ \
-    v2u64 t = MSA_TPV_REINTERPRET(v2u64, msa_cntq_s64(MSA_TPV_REINTERPRET(v2i64, a.val))); \
-    return v_uint64x2(t); \
-}
-OPENCV_HAL_IMPL_MSA_POPCOUNT_SIZE64(v_uint64x2)
-OPENCV_HAL_IMPL_MSA_POPCOUNT_SIZE64(v_int64x2)
-
-inline int v_signmask(const v_uint8x16& a)
-{
-    v8i8 m0 = msa_create_s8(CV_BIG_UINT(0x0706050403020100));
-    v16u8 v0 = msa_shlq_u8(msa_shrq_n_u8(a.val, 7), msa_combine_s8(m0, m0));
-    v8u16 v1 = msa_paddlq_u8(v0);
-    v4u32 v2 = msa_paddlq_u16(v1);
-    v2u64 v3 = msa_paddlq_u32(v2);
-    return (int)msa_getq_lane_u64(v3, 0) + ((int)msa_getq_lane_u64(v3, 1) << 8);
-}
-inline int v_signmask(const v_int8x16& a)
-{ return v_signmask(v_reinterpret_as_u8(a)); }
-
-inline int v_signmask(const v_uint16x8& a)
-{
-    v4i16 m0 = msa_create_s16(CV_BIG_UINT(0x0003000200010000));
-    v8u16 v0 = msa_shlq_u16(msa_shrq_n_u16(a.val, 15), msa_combine_s16(m0, m0));
-    v4u32 v1 = msa_paddlq_u16(v0);
-    v2u64 v2 = msa_paddlq_u32(v1);
-    return (int)msa_getq_lane_u64(v2, 0) + ((int)msa_getq_lane_u64(v2, 1) << 4);
-}
-inline int v_signmask(const v_int16x8& a)
-{ return v_signmask(v_reinterpret_as_u16(a)); }
-
-inline int v_signmask(const v_uint32x4& a)
-{
-    v2i32 m0 = msa_create_s32(CV_BIG_UINT(0x0000000100000000));
-    v4u32 v0 = msa_shlq_u32(msa_shrq_n_u32(a.val, 31), msa_combine_s32(m0, m0));
-    v2u64 v1 = msa_paddlq_u32(v0);
-    return (int)msa_getq_lane_u64(v1, 0) + ((int)msa_getq_lane_u64(v1, 1) << 2);
-}
-inline int v_signmask(const v_int32x4& a)
-{ return v_signmask(v_reinterpret_as_u32(a)); }
-inline int v_signmask(const v_float32x4& a)
-{ return v_signmask(v_reinterpret_as_u32(a)); }
-
-inline int v_signmask(const v_uint64x2& a)
-{
-    v2u64 v0 = msa_shrq_n_u64(a.val, 63);
-    return (int)msa_getq_lane_u64(v0, 0) + ((int)msa_getq_lane_u64(v0, 1) << 1);
-}
-inline int v_signmask(const v_int64x2& a)
-{ return v_signmask(v_reinterpret_as_u64(a)); }
-inline int v_signmask(const v_float64x2& a)
-{ return v_signmask(v_reinterpret_as_u64(a)); }
-
-inline int v_scan_forward(const v_int8x16& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_uint8x16& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_int16x8& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_uint16x8& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_int32x4& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_uint32x4& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_float32x4& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_int64x2& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_uint64x2& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_float64x2& a) { return trailingZeros32(v_signmask(a)); }
-
-#define OPENCV_HAL_IMPL_MSA_CHECK_ALLANY(_Tpvec, _Tpvec2, suffix, shift) \
-inline bool v_check_all(const v_##_Tpvec& a) \
-{ \
-    _Tpvec2 v0 = msa_shrq_n_##suffix(msa_mvnq_##suffix(a.val), shift); \
-    v2u64 v1 = MSA_TPV_REINTERPRET(v2u64, v0); \
-    return (msa_getq_lane_u64(v1, 0) | msa_getq_lane_u64(v1, 1)) == 0; \
-} \
-inline bool v_check_any(const v_##_Tpvec& a) \
-{ \
-    _Tpvec2 v0 = msa_shrq_n_##suffix(a.val, shift); \
-    v2u64 v1 = MSA_TPV_REINTERPRET(v2u64, v0); \
-    return (msa_getq_lane_u64(v1, 0) | msa_getq_lane_u64(v1, 1)) != 0; \
-}
-
-OPENCV_HAL_IMPL_MSA_CHECK_ALLANY(uint8x16, v16u8, u8, 7)
-OPENCV_HAL_IMPL_MSA_CHECK_ALLANY(uint16x8, v8u16, u16, 15)
-OPENCV_HAL_IMPL_MSA_CHECK_ALLANY(uint32x4, v4u32, u32, 31)
-OPENCV_HAL_IMPL_MSA_CHECK_ALLANY(uint64x2, v2u64, u64, 63)
-
-inline bool v_check_all(const v_int8x16& a)
-{ return v_check_all(v_reinterpret_as_u8(a)); }
-inline bool v_check_all(const v_int16x8& a)
-{ return v_check_all(v_reinterpret_as_u16(a)); }
-inline bool v_check_all(const v_int32x4& a)
-{ return v_check_all(v_reinterpret_as_u32(a)); }
-inline bool v_check_all(const v_float32x4& a)
-{ return v_check_all(v_reinterpret_as_u32(a)); }
-
-inline bool v_check_any(const v_int8x16& a)
-{ return v_check_any(v_reinterpret_as_u8(a)); }
-inline bool v_check_any(const v_int16x8& a)
-{ return v_check_any(v_reinterpret_as_u16(a)); }
-inline bool v_check_any(const v_int32x4& a)
-{ return v_check_any(v_reinterpret_as_u32(a)); }
-inline bool v_check_any(const v_float32x4& a)
-{ return v_check_any(v_reinterpret_as_u32(a)); }
-
-inline bool v_check_all(const v_int64x2& a)
-{ return v_check_all(v_reinterpret_as_u64(a)); }
-inline bool v_check_all(const v_float64x2& a)
-{ return v_check_all(v_reinterpret_as_u64(a)); }
-inline bool v_check_any(const v_int64x2& a)
-{ return v_check_any(v_reinterpret_as_u64(a)); }
-inline bool v_check_any(const v_float64x2& a)
-{ return v_check_any(v_reinterpret_as_u64(a)); }
-
-/* v_select */
-#define OPENCV_HAL_IMPL_MSA_SELECT(_Tpvec, _Tpv, _Tpvu) \
-inline _Tpvec v_select(const _Tpvec& mask, const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(MSA_TPV_REINTERPRET(_Tpv, msa_bslq_u8(MSA_TPV_REINTERPRET(_Tpvu, mask.val), \
-                  MSA_TPV_REINTERPRET(_Tpvu, b.val), MSA_TPV_REINTERPRET(_Tpvu, a.val)))); \
-}
-
-OPENCV_HAL_IMPL_MSA_SELECT(v_uint8x16, v16u8, v16u8)
-OPENCV_HAL_IMPL_MSA_SELECT(v_int8x16, v16i8, v16u8)
-OPENCV_HAL_IMPL_MSA_SELECT(v_uint16x8, v8u16, v16u8)
-OPENCV_HAL_IMPL_MSA_SELECT(v_int16x8, v8i16, v16u8)
-OPENCV_HAL_IMPL_MSA_SELECT(v_uint32x4, v4u32, v16u8)
-OPENCV_HAL_IMPL_MSA_SELECT(v_int32x4, v4i32, v16u8)
-OPENCV_HAL_IMPL_MSA_SELECT(v_float32x4, v4f32, v16u8)
-OPENCV_HAL_IMPL_MSA_SELECT(v_float64x2, v2f64, v16u8)
-
-#define OPENCV_HAL_IMPL_MSA_EXPAND(_Tpvec, _Tpwvec, _Tp, suffix, ssuffix, _Tpv, _Tpvs) \
-inline void v_expand(const _Tpvec& a, _Tpwvec& b0, _Tpwvec& b1) \
-{ \
-    _Tpv a_lo = MSA_TPV_REINTERPRET(_Tpv, msa_ilvrq_##ssuffix(MSA_TPV_REINTERPRET(_Tpvs, a.val), msa_dupq_n_##ssuffix(0))); \
-    _Tpv a_hi = MSA_TPV_REINTERPRET(_Tpv, msa_ilvlq_##ssuffix(MSA_TPV_REINTERPRET(_Tpvs, a.val), msa_dupq_n_##ssuffix(0))); \
-    b0.val = msa_paddlq_##suffix(a_lo); \
-    b1.val = msa_paddlq_##suffix(a_hi); \
-} \
-inline _Tpwvec v_expand_low(const _Tpvec& a) \
-{ \
-    _Tpv a_lo = MSA_TPV_REINTERPRET(_Tpv, msa_ilvrq_##ssuffix(MSA_TPV_REINTERPRET(_Tpvs, a.val), msa_dupq_n_##ssuffix(0))); \
-    return _Tpwvec(msa_paddlq_##suffix(a_lo)); \
-} \
-inline _Tpwvec v_expand_high(const _Tpvec& a) \
-{ \
-    _Tpv a_hi = MSA_TPV_REINTERPRET(_Tpv, msa_ilvlq_##ssuffix(MSA_TPV_REINTERPRET(_Tpvs, a.val), msa_dupq_n_##ssuffix(0))); \
-    return _Tpwvec(msa_paddlq_##suffix(a_hi)); \
-} \
-inline _Tpwvec v_load_expand(const _Tp* ptr) \
-{ \
-    return _Tpwvec(msa_movl_##suffix(msa_ld1_##suffix(ptr))); \
-}
-
-OPENCV_HAL_IMPL_MSA_EXPAND(v_uint8x16, v_uint16x8, uchar, u8, s8, v16u8, v16i8)
-OPENCV_HAL_IMPL_MSA_EXPAND(v_int8x16, v_int16x8, schar, s8, s8, v16i8, v16i8)
-OPENCV_HAL_IMPL_MSA_EXPAND(v_uint16x8, v_uint32x4, ushort, u16, s16, v8u16, v8i16)
-OPENCV_HAL_IMPL_MSA_EXPAND(v_int16x8, v_int32x4, short, s16, s16, v8i16, v8i16)
-OPENCV_HAL_IMPL_MSA_EXPAND(v_uint32x4, v_uint64x2, uint, u32, s32, v4u32, v4i32)
-OPENCV_HAL_IMPL_MSA_EXPAND(v_int32x4, v_int64x2, int, s32, s32, v4i32, v4i32)
-
-inline v_uint32x4 v_load_expand_q(const uchar* ptr)
-{
-    return v_uint32x4((v4u32){ptr[0], ptr[1], ptr[2], ptr[3]});
-}
-
-inline v_int32x4 v_load_expand_q(const schar* ptr)
-{
-    return v_int32x4((v4i32){ptr[0], ptr[1], ptr[2], ptr[3]});
-}
-
-/* v_zip, v_combine_low, v_combine_high, v_recombine */
-#define OPENCV_HAL_IMPL_MSA_UNPACKS(_Tpvec, _Tpv, _Tpvs, ssuffix) \
-inline void v_zip(const _Tpvec& a0, const _Tpvec& a1, _Tpvec& b0, _Tpvec& b1) \
-{ \
-    b0.val = MSA_TPV_REINTERPRET(_Tpv, msa_ilvrq_##ssuffix(MSA_TPV_REINTERPRET(_Tpvs, a1.val), MSA_TPV_REINTERPRET(_Tpvs, a0.val))); \
-    b1.val = MSA_TPV_REINTERPRET(_Tpv, msa_ilvlq_##ssuffix(MSA_TPV_REINTERPRET(_Tpvs, a1.val), MSA_TPV_REINTERPRET(_Tpvs, a0.val))); \
-} \
-inline _Tpvec v_combine_low(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(MSA_TPV_REINTERPRET(_Tpv, msa_ilvrq_s64(MSA_TPV_REINTERPRET(v2i64, b.val), MSA_TPV_REINTERPRET(v2i64, a.val)))); \
-} \
-inline _Tpvec v_combine_high(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(MSA_TPV_REINTERPRET(_Tpv, msa_ilvlq_s64(MSA_TPV_REINTERPRET(v2i64, b.val), MSA_TPV_REINTERPRET(v2i64, a.val)))); \
-} \
-inline void v_recombine(const _Tpvec& a, const _Tpvec& b, _Tpvec& c, _Tpvec& d) \
-{ \
-    c.val = MSA_TPV_REINTERPRET(_Tpv, msa_ilvrq_s64(MSA_TPV_REINTERPRET(v2i64, b.val), MSA_TPV_REINTERPRET(v2i64, a.val))); \
-    d.val = MSA_TPV_REINTERPRET(_Tpv, msa_ilvlq_s64(MSA_TPV_REINTERPRET(v2i64, b.val), MSA_TPV_REINTERPRET(v2i64, a.val))); \
-}
-
-OPENCV_HAL_IMPL_MSA_UNPACKS(v_uint8x16, v16u8, v16i8, s8)
-OPENCV_HAL_IMPL_MSA_UNPACKS(v_int8x16, v16i8, v16i8, s8)
-OPENCV_HAL_IMPL_MSA_UNPACKS(v_uint16x8, v8u16, v8i16, s16)
-OPENCV_HAL_IMPL_MSA_UNPACKS(v_int16x8, v8i16, v8i16, s16)
-OPENCV_HAL_IMPL_MSA_UNPACKS(v_uint32x4, v4u32, v4i32, s32)
-OPENCV_HAL_IMPL_MSA_UNPACKS(v_int32x4, v4i32, v4i32, s32)
-OPENCV_HAL_IMPL_MSA_UNPACKS(v_float32x4, v4f32, v4i32, s32)
-OPENCV_HAL_IMPL_MSA_UNPACKS(v_float64x2, v2f64, v2i64, s64)
-
-/* v_extract */
-#define OPENCV_HAL_IMPL_MSA_EXTRACT(_Tpvec, _Tpv, _Tpvs, suffix) \
-template <int s> \
-inline _Tpvec v_extract(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(MSA_TPV_REINTERPRET(_Tpv, msa_extq_##suffix(MSA_TPV_REINTERPRET(_Tpvs, a.val), MSA_TPV_REINTERPRET(_Tpvs, b.val), s))); \
-}
-
-OPENCV_HAL_IMPL_MSA_EXTRACT(v_uint8x16, v16u8, v16i8, s8)
-OPENCV_HAL_IMPL_MSA_EXTRACT(v_int8x16, v16i8, v16i8, s8)
-OPENCV_HAL_IMPL_MSA_EXTRACT(v_uint16x8, v8u16, v8i16, s16)
-OPENCV_HAL_IMPL_MSA_EXTRACT(v_int16x8, v8i16, v8i16, s16)
-OPENCV_HAL_IMPL_MSA_EXTRACT(v_uint32x4, v4u32, v4i32, s32)
-OPENCV_HAL_IMPL_MSA_EXTRACT(v_int32x4, v4i32, v4i32, s32)
-OPENCV_HAL_IMPL_MSA_EXTRACT(v_uint64x2, v2u64, v2i64, s64)
-OPENCV_HAL_IMPL_MSA_EXTRACT(v_int64x2, v2i64, v2i64, s64)
-OPENCV_HAL_IMPL_MSA_EXTRACT(v_float32x4, v4f32, v4i32, s32)
-OPENCV_HAL_IMPL_MSA_EXTRACT(v_float64x2, v2f64, v2i64, s64)
-
-/* v_round, v_floor, v_ceil, v_trunc */
-inline v_int32x4 v_round(const v_float32x4& a)
-{
-    return v_int32x4(msa_cvttintq_s32_f32(a.val));
-}
-
-inline v_int32x4 v_floor(const v_float32x4& a)
-{
-    v4i32 a1 = msa_cvttintq_s32_f32(a.val);
-    return v_int32x4(msa_addq_s32(a1, MSA_TPV_REINTERPRET(v4i32, msa_cgtq_f32(msa_cvtfintq_f32_s32(a1), a.val))));
-}
-
-inline v_int32x4 v_ceil(const v_float32x4& a)
-{
-    v4i32 a1 = msa_cvttintq_s32_f32(a.val);
-    return v_int32x4(msa_subq_s32(a1, MSA_TPV_REINTERPRET(v4i32, msa_cgtq_f32(a.val, msa_cvtfintq_f32_s32(a1)))));
-}
-
-inline v_int32x4 v_trunc(const v_float32x4& a)
-{
-    return v_int32x4(msa_cvttruncq_s32_f32(a.val));
-}
-
-inline v_int32x4 v_round(const v_float64x2& a)
-{
-    return v_int32x4(msa_pack_s64(msa_cvttintq_s64_f64(a.val), msa_dupq_n_s64(0)));
-}
-
-inline v_int32x4 v_round(const v_float64x2& a, const v_float64x2& b)
-{
-    return v_int32x4(msa_pack_s64(msa_cvttintq_s64_f64(a.val), msa_cvttintq_s64_f64(b.val)));
-}
-
-inline v_int32x4 v_floor(const v_float64x2& a)
-{
-    v2f64 a1 = msa_cvtrintq_f64(a.val);
-    return v_int32x4(msa_pack_s64(msa_addq_s64(msa_cvttruncq_s64_f64(a1), MSA_TPV_REINTERPRET(v2i64, msa_cgtq_f64(a1, a.val))), msa_dupq_n_s64(0)));
-}
-
-inline v_int32x4 v_ceil(const v_float64x2& a)
-{
-    v2f64 a1 = msa_cvtrintq_f64(a.val);
-    return v_int32x4(msa_pack_s64(msa_subq_s64(msa_cvttruncq_s64_f64(a1), MSA_TPV_REINTERPRET(v2i64, msa_cgtq_f64(a.val, a1))), msa_dupq_n_s64(0)));
-}
-
-inline v_int32x4 v_trunc(const v_float64x2& a)
-{
-    return v_int32x4(msa_pack_s64(msa_cvttruncq_s64_f64(a.val), msa_dupq_n_s64(0)));
-}
-
-#define OPENCV_HAL_IMPL_MSA_TRANSPOSE4x4(_Tpvec, _Tpv, _Tpvs, ssuffix) \
-inline void v_transpose4x4(const _Tpvec& a0, const _Tpvec& a1, \
-                           const _Tpvec& a2, const _Tpvec& a3, \
-                           _Tpvec& b0, _Tpvec& b1, \
-                           _Tpvec& b2, _Tpvec& b3) \
-{ \
-    _Tpv t00 = MSA_TPV_REINTERPRET(_Tpv, msa_ilvrq_##ssuffix(MSA_TPV_REINTERPRET(_Tpvs, a1.val), MSA_TPV_REINTERPRET(_Tpvs, a0.val))); \
-    _Tpv t01 = MSA_TPV_REINTERPRET(_Tpv, msa_ilvlq_##ssuffix(MSA_TPV_REINTERPRET(_Tpvs, a1.val), MSA_TPV_REINTERPRET(_Tpvs, a0.val))); \
-    _Tpv t10 = MSA_TPV_REINTERPRET(_Tpv, msa_ilvrq_##ssuffix(MSA_TPV_REINTERPRET(_Tpvs, a3.val), MSA_TPV_REINTERPRET(_Tpvs, a2.val))); \
-    _Tpv t11 = MSA_TPV_REINTERPRET(_Tpv, msa_ilvlq_##ssuffix(MSA_TPV_REINTERPRET(_Tpvs, a3.val), MSA_TPV_REINTERPRET(_Tpvs, a2.val))); \
-    b0.val = MSA_TPV_REINTERPRET(_Tpv, msa_ilvrq_s64(MSA_TPV_REINTERPRET(v2i64, t10), MSA_TPV_REINTERPRET(v2i64, t00))); \
-    b1.val = MSA_TPV_REINTERPRET(_Tpv, msa_ilvlq_s64(MSA_TPV_REINTERPRET(v2i64, t10), MSA_TPV_REINTERPRET(v2i64, t00))); \
-    b2.val = MSA_TPV_REINTERPRET(_Tpv, msa_ilvrq_s64(MSA_TPV_REINTERPRET(v2i64, t11), MSA_TPV_REINTERPRET(v2i64, t01))); \
-    b3.val = MSA_TPV_REINTERPRET(_Tpv, msa_ilvlq_s64(MSA_TPV_REINTERPRET(v2i64, t11), MSA_TPV_REINTERPRET(v2i64, t01))); \
-}
-
-OPENCV_HAL_IMPL_MSA_TRANSPOSE4x4(v_uint32x4, v4u32, v4i32, s32)
-OPENCV_HAL_IMPL_MSA_TRANSPOSE4x4(v_int32x4, v4i32, v4i32, s32)
-OPENCV_HAL_IMPL_MSA_TRANSPOSE4x4(v_float32x4, v4f32, v4i32, s32)
-
-#define OPENCV_HAL_IMPL_MSA_INTERLEAVED(_Tpvec, _Tp, suffix) \
-inline void v_load_deinterleave(const _Tp* ptr, v_##_Tpvec& a, v_##_Tpvec& b) \
-{ \
-    msa_ld2q_##suffix(ptr, &a.val, &b.val); \
-} \
-inline void v_load_deinterleave(const _Tp* ptr, v_##_Tpvec& a, v_##_Tpvec& b, v_##_Tpvec& c) \
-{ \
-    msa_ld3q_##suffix(ptr, &a.val, &b.val, &c.val); \
-} \
-inline void v_load_deinterleave(const _Tp* ptr, v_##_Tpvec& a, v_##_Tpvec& b, \
-                                v_##_Tpvec& c, v_##_Tpvec& d) \
-{ \
-    msa_ld4q_##suffix(ptr, &a.val, &b.val, &c.val, &d.val); \
-} \
-inline void v_store_interleave( _Tp* ptr, const v_##_Tpvec& a, const v_##_Tpvec& b, \
-                                hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ \
-    msa_st2q_##suffix(ptr, a.val, b.val); \
-} \
-inline void v_store_interleave( _Tp* ptr, const v_##_Tpvec& a, const v_##_Tpvec& b, \
-                                const v_##_Tpvec& c, hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ \
-    msa_st3q_##suffix(ptr, a.val, b.val, c.val); \
-} \
-inline void v_store_interleave( _Tp* ptr, const v_##_Tpvec& a, const v_##_Tpvec& b, \
-                                const v_##_Tpvec& c, const v_##_Tpvec& d, \
-                                hal::StoreMode /*mode*/=hal::STORE_UNALIGNED ) \
-{ \
-    msa_st4q_##suffix(ptr, a.val, b.val, c.val, d.val); \
-}
-
-OPENCV_HAL_IMPL_MSA_INTERLEAVED(uint8x16, uchar, u8)
-OPENCV_HAL_IMPL_MSA_INTERLEAVED(int8x16, schar, s8)
-OPENCV_HAL_IMPL_MSA_INTERLEAVED(uint16x8, ushort, u16)
-OPENCV_HAL_IMPL_MSA_INTERLEAVED(int16x8, short, s16)
-OPENCV_HAL_IMPL_MSA_INTERLEAVED(uint32x4, unsigned, u32)
-OPENCV_HAL_IMPL_MSA_INTERLEAVED(int32x4, int, s32)
-OPENCV_HAL_IMPL_MSA_INTERLEAVED(float32x4, float, f32)
-OPENCV_HAL_IMPL_MSA_INTERLEAVED(uint64x2, uint64, u64)
-OPENCV_HAL_IMPL_MSA_INTERLEAVED(int64x2, int64, s64)
-OPENCV_HAL_IMPL_MSA_INTERLEAVED(float64x2, double, f64)
-
-/* v_cvt_f32, v_cvt_f64, v_cvt_f64_high */
-inline v_float32x4 v_cvt_f32(const v_int32x4& a)
-{
-    return v_float32x4(msa_cvtfintq_f32_s32(a.val));
-}
-
-inline v_float32x4 v_cvt_f32(const v_float64x2& a)
-{
-    return v_float32x4(msa_cvtfq_f32_f64(a.val, msa_dupq_n_f64(0.0f)));
-}
-
-inline v_float32x4 v_cvt_f32(const v_float64x2& a, const v_float64x2& b)
-{
-    return v_float32x4(msa_cvtfq_f32_f64(a.val, b.val));
-}
-
-inline v_float64x2 v_cvt_f64(const v_int32x4& a)
-{
-    return v_float64x2(msa_cvtflq_f64_f32(msa_cvtfintq_f32_s32(a.val)));
-}
-
-inline v_float64x2 v_cvt_f64_high(const v_int32x4& a)
-{
-    return v_float64x2(msa_cvtfhq_f64_f32(msa_cvtfintq_f32_s32(a.val)));
-}
-
-inline v_float64x2 v_cvt_f64(const v_float32x4& a)
-{
-    return v_float64x2(msa_cvtflq_f64_f32(a.val));
-}
-
-inline v_float64x2 v_cvt_f64_high(const v_float32x4& a)
-{
-    return v_float64x2(msa_cvtfhq_f64_f32(a.val));
-}
-
-inline v_float64x2 v_cvt_f64(const v_int64x2& a)
-{
-    return v_float64x2(msa_cvtfintq_f64_s64(a.val));
-}
-
-////////////// Lookup table access ////////////////////
-inline v_int8x16 v_lut(const schar* tab, const int* idx)
-{
-    schar CV_DECL_ALIGNED(32) elems[16] =
-    {
-        tab[idx[ 0]],
-        tab[idx[ 1]],
-        tab[idx[ 2]],
-        tab[idx[ 3]],
-        tab[idx[ 4]],
-        tab[idx[ 5]],
-        tab[idx[ 6]],
-        tab[idx[ 7]],
-        tab[idx[ 8]],
-        tab[idx[ 9]],
-        tab[idx[10]],
-        tab[idx[11]],
-        tab[idx[12]],
-        tab[idx[13]],
-        tab[idx[14]],
-        tab[idx[15]]
-    };
-    return v_int8x16(msa_ld1q_s8(elems));
-}
-inline v_int8x16 v_lut_pairs(const schar* tab, const int* idx)
-{
-    schar CV_DECL_ALIGNED(32) elems[16] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[1]],
-        tab[idx[1] + 1],
-        tab[idx[2]],
-        tab[idx[2] + 1],
-        tab[idx[3]],
-        tab[idx[3] + 1],
-        tab[idx[4]],
-        tab[idx[4] + 1],
-        tab[idx[5]],
-        tab[idx[5] + 1],
-        tab[idx[6]],
-        tab[idx[6] + 1],
-        tab[idx[7]],
-        tab[idx[7] + 1]
-    };
-    return v_int8x16(msa_ld1q_s8(elems));
-}
-inline v_int8x16 v_lut_quads(const schar* tab, const int* idx)
-{
-    schar CV_DECL_ALIGNED(32) elems[16] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[0] + 2],
-        tab[idx[0] + 3],
-        tab[idx[1]],
-        tab[idx[1] + 1],
-        tab[idx[1] + 2],
-        tab[idx[1] + 3],
-        tab[idx[2]],
-        tab[idx[2] + 1],
-        tab[idx[2] + 2],
-        tab[idx[2] + 3],
-        tab[idx[3]],
-        tab[idx[3] + 1],
-        tab[idx[3] + 2],
-        tab[idx[3] + 3]
-    };
-    return v_int8x16(msa_ld1q_s8(elems));
-}
-inline v_uint8x16 v_lut(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut((schar*)tab, idx)); }
-inline v_uint8x16 v_lut_pairs(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut_pairs((schar*)tab, idx)); }
-inline v_uint8x16 v_lut_quads(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut_quads((schar*)tab, idx)); }
-
-
-inline v_int16x8 v_lut(const short* tab, const int* idx)
-{
-    short CV_DECL_ALIGNED(32) elems[8] =
-    {
-        tab[idx[0]],
-        tab[idx[1]],
-        tab[idx[2]],
-        tab[idx[3]],
-        tab[idx[4]],
-        tab[idx[5]],
-        tab[idx[6]],
-        tab[idx[7]]
-    };
-    return v_int16x8(msa_ld1q_s16(elems));
-}
-inline v_int16x8 v_lut_pairs(const short* tab, const int* idx)
-{
-    short CV_DECL_ALIGNED(32) elems[8] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[1]],
-        tab[idx[1] + 1],
-        tab[idx[2]],
-        tab[idx[2] + 1],
-        tab[idx[3]],
-        tab[idx[3] + 1]
-    };
-    return v_int16x8(msa_ld1q_s16(elems));
-}
-inline v_int16x8 v_lut_quads(const short* tab, const int* idx)
-{
-    return v_int16x8(msa_combine_s16(msa_ld1_s16(tab + idx[0]), msa_ld1_s16(tab + idx[1])));
-}
-inline v_uint16x8 v_lut(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut((short*)tab, idx)); }
-inline v_uint16x8 v_lut_pairs(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut_pairs((short*)tab, idx)); }
-inline v_uint16x8 v_lut_quads(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut_quads((short*)tab, idx)); }
-
-inline v_int32x4 v_lut(const int* tab, const int* idx)
-{
-    int CV_DECL_ALIGNED(32) elems[4] =
-    {
-        tab[idx[0]],
-        tab[idx[1]],
-        tab[idx[2]],
-        tab[idx[3]]
-    };
-    return v_int32x4(msa_ld1q_s32(elems));
-}
-inline v_int32x4 v_lut_pairs(const int* tab, const int* idx)
-{
-    return v_int32x4(msa_combine_s32(msa_ld1_s32(tab + idx[0]), msa_ld1_s32(tab + idx[1])));
-}
-inline v_int32x4 v_lut_quads(const int* tab, const int* idx)
-{
-    return v_int32x4(msa_ld1q_s32(tab + idx[0]));
-}
-inline v_uint32x4 v_lut(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut((int*)tab, idx)); }
-inline v_uint32x4 v_lut_pairs(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut_pairs((int*)tab, idx)); }
-inline v_uint32x4 v_lut_quads(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut_quads((int*)tab, idx)); }
-
-inline v_int64x2 v_lut(const int64_t* tab, const int* idx)
-{
-    return v_int64x2(msa_combine_s64(msa_create_s64(tab[idx[0]]), msa_create_s64(tab[idx[1]])));
-}
-inline v_int64x2 v_lut_pairs(const int64_t* tab, const int* idx)
-{
-    return v_int64x2(msa_ld1q_s64(tab + idx[0]));
-}
-inline v_uint64x2 v_lut(const uint64_t* tab, const int* idx) { return v_reinterpret_as_u64(v_lut((const int64_t *)tab, idx)); }
-inline v_uint64x2 v_lut_pairs(const uint64_t* tab, const int* idx) { return v_reinterpret_as_u64(v_lut_pairs((const int64_t *)tab, idx)); }
-
-inline v_float32x4 v_lut(const float* tab, const int* idx)
-{
-    float CV_DECL_ALIGNED(32) elems[4] =
-    {
-        tab[idx[0]],
-        tab[idx[1]],
-        tab[idx[2]],
-        tab[idx[3]]
-    };
-    return v_float32x4(msa_ld1q_f32(elems));
-}
-inline v_float32x4 v_lut_pairs(const float* tab, const int* idx)
-{
-    uint64 CV_DECL_ALIGNED(32) elems[2] =
-    {
-        *(uint64*)(tab + idx[0]),
-        *(uint64*)(tab + idx[1])
-    };
-    return v_float32x4(MSA_TPV_REINTERPRET(v4f32, msa_ld1q_u64(elems)));
-}
-inline v_float32x4 v_lut_quads(const float* tab, const int* idx)
-{
-    return v_float32x4(msa_ld1q_f32(tab + idx[0]));
-}
-
-inline v_int32x4 v_lut(const int* tab, const v_int32x4& idxvec)
-{
-    int CV_DECL_ALIGNED(32) idx[4];
-    v_store_aligned(idx, idxvec);
-
-    return v_int32x4(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]]);
-}
-
-inline v_uint32x4 v_lut(const unsigned* tab, const v_int32x4& idxvec)
-{
-    unsigned CV_DECL_ALIGNED(32) elems[4] =
-    {
-        tab[msa_getq_lane_s32(idxvec.val, 0)],
-        tab[msa_getq_lane_s32(idxvec.val, 1)],
-        tab[msa_getq_lane_s32(idxvec.val, 2)],
-        tab[msa_getq_lane_s32(idxvec.val, 3)]
-    };
-    return v_uint32x4(msa_ld1q_u32(elems));
-}
-
-inline v_float32x4 v_lut(const float* tab, const v_int32x4& idxvec)
-{
-    int CV_DECL_ALIGNED(32) idx[4];
-    v_store_aligned(idx, idxvec);
-
-    return v_float32x4(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]]);
-}
-
-inline void v_lut_deinterleave(const float* tab, const v_int32x4& idxvec, v_float32x4& x, v_float32x4& y)
-{
-    int CV_DECL_ALIGNED(32) idx[4];
-    v_store_aligned(idx, idxvec);
-
-    v4f32 xy02 = msa_combine_f32(msa_ld1_f32(tab + idx[0]), msa_ld1_f32(tab + idx[2]));
-    v4f32 xy13 = msa_combine_f32(msa_ld1_f32(tab + idx[1]), msa_ld1_f32(tab + idx[3]));
-    x = v_float32x4(MSA_TPV_REINTERPRET(v4f32, msa_ilvevq_s32(MSA_TPV_REINTERPRET(v4i32, xy13), MSA_TPV_REINTERPRET(v4i32, xy02))));
-    y = v_float32x4(MSA_TPV_REINTERPRET(v4f32, msa_ilvodq_s32(MSA_TPV_REINTERPRET(v4i32, xy13), MSA_TPV_REINTERPRET(v4i32, xy02))));
-}
-
-inline v_int8x16 v_interleave_pairs(const v_int8x16& vec)
-{
-    v_int8x16 c = v_int8x16(__builtin_msa_vshf_b((v16i8)((v2i64){0x0705060403010200, 0x0F0D0E0C0B090A08}), msa_dupq_n_s8(0), vec.val));
-    return c;
-}
-inline v_uint8x16 v_interleave_pairs(const v_uint8x16& vec)
-{ return v_reinterpret_as_u8(v_interleave_pairs(v_reinterpret_as_s8(vec))); }
-inline v_int8x16 v_interleave_quads(const v_int8x16& vec)
-{
-    v_int8x16 c = v_int8x16(__builtin_msa_vshf_b((v16i8)((v2i64){0x0703060205010400, 0x0F0B0E0A0D090C08}), msa_dupq_n_s8(0), vec.val));
-    return c;
-}
-inline v_uint8x16 v_interleave_quads(const v_uint8x16& vec) { return v_reinterpret_as_u8(v_interleave_quads(v_reinterpret_as_s8(vec))); }
-
-inline v_int16x8 v_interleave_pairs(const v_int16x8& vec)
-{
-    v_int16x8 c = v_int16x8(__builtin_msa_vshf_h((v8i16)((v2i64){0x0003000100020000, 0x0007000500060004}), msa_dupq_n_s16(0), vec.val));
-    return c;
-}
-
-inline v_uint16x8 v_interleave_pairs(const v_uint16x8& vec) { return v_reinterpret_as_u16(v_interleave_pairs(v_reinterpret_as_s16(vec))); }
-
-inline v_int16x8 v_interleave_quads(const v_int16x8& vec)
-{
-    v_int16x8 c = v_int16x8(__builtin_msa_vshf_h((v8i16)((v2i64){0x0005000100040000, 0x0007000300060002}), msa_dupq_n_s16(0), vec.val));
-    return c;
-}
-
-inline v_uint16x8 v_interleave_quads(const v_uint16x8& vec) { return v_reinterpret_as_u16(v_interleave_quads(v_reinterpret_as_s16(vec))); }
-
-inline v_int32x4 v_interleave_pairs(const v_int32x4& vec)
-{
-    v_int32x4 c;
-    c.val[0] = vec.val[0];
-    c.val[1] = vec.val[2];
-    c.val[2] = vec.val[1];
-    c.val[3] = vec.val[3];
-    return c;
-}
-
-inline v_uint32x4 v_interleave_pairs(const v_uint32x4& vec) { return v_reinterpret_as_u32(v_interleave_pairs(v_reinterpret_as_s32(vec))); }
-inline v_float32x4 v_interleave_pairs(const v_float32x4& vec) { return v_reinterpret_as_f32(v_interleave_pairs(v_reinterpret_as_s32(vec))); }
-
-inline v_int8x16 v_pack_triplets(const v_int8x16& vec)
-{
-    v_int8x16 c = v_int8x16(__builtin_msa_vshf_b((v16i8)((v2i64){0x0908060504020100, 0x131211100E0D0C0A}), msa_dupq_n_s8(0), vec.val));
-    return c;
-}
-
-inline v_uint8x16 v_pack_triplets(const v_uint8x16& vec) { return v_reinterpret_as_u8(v_pack_triplets(v_reinterpret_as_s8(vec))); }
-
-inline v_int16x8 v_pack_triplets(const v_int16x8& vec)
-{
-    v_int16x8 c = v_int16x8(__builtin_msa_vshf_h((v8i16)((v2i64){0x0004000200010000, 0x0009000800060005}), msa_dupq_n_s16(0), vec.val));
-    return c;
-}
-
-inline v_uint16x8 v_pack_triplets(const v_uint16x8& vec) { return v_reinterpret_as_u16(v_pack_triplets(v_reinterpret_as_s16(vec))); }
-inline v_int32x4 v_pack_triplets(const v_int32x4& vec) { return vec; }
-inline v_uint32x4 v_pack_triplets(const v_uint32x4& vec) { return vec; }
-inline v_float32x4 v_pack_triplets(const v_float32x4& vec) { return vec; }
-
-inline v_float64x2 v_lut(const double* tab, const int* idx)
-{
-    double CV_DECL_ALIGNED(32) elems[2] =
-    {
-        tab[idx[0]],
-        tab[idx[1]]
-    };
-    return v_float64x2(msa_ld1q_f64(elems));
-}
-
-inline v_float64x2 v_lut_pairs(const double* tab, const int* idx)
-{
-    return v_float64x2(msa_ld1q_f64(tab + idx[0]));
-}
-
-inline v_float64x2 v_lut(const double* tab, const v_int32x4& idxvec)
-{
-    int CV_DECL_ALIGNED(32) idx[4];
-    v_store_aligned(idx, idxvec);
-
-    return v_float64x2(tab[idx[0]], tab[idx[1]]);
-}
-
-inline void v_lut_deinterleave(const double* tab, const v_int32x4& idxvec, v_float64x2& x, v_float64x2& y)
-{
-    int CV_DECL_ALIGNED(32) idx[4];
-    v_store_aligned(idx, idxvec);
-
-    v2f64 xy0 = msa_ld1q_f64(tab + idx[0]);
-    v2f64 xy1 = msa_ld1q_f64(tab + idx[1]);
-    x = v_float64x2(MSA_TPV_REINTERPRET(v2f64, msa_ilvevq_s64(MSA_TPV_REINTERPRET(v2i64, xy1), MSA_TPV_REINTERPRET(v2i64, xy0))));
-    y = v_float64x2(MSA_TPV_REINTERPRET(v2f64, msa_ilvodq_s64(MSA_TPV_REINTERPRET(v2i64, xy1), MSA_TPV_REINTERPRET(v2i64, xy0))));
-}
-
-template<int i, typename _Tp>
-inline typename _Tp::lane_type v_extract_n(const _Tp& a)
-{
-    return v_rotate_right<i>(a).get0();
-}
-
-template<int i>
-inline v_uint32x4 v_broadcast_element(const v_uint32x4& a)
-{
-    return v_setall_u32(v_extract_n<i>(a));
-}
-template<int i>
-inline v_int32x4 v_broadcast_element(const v_int32x4& a)
-{
-    return v_setall_s32(v_extract_n<i>(a));
-}
-template<int i>
-inline v_float32x4 v_broadcast_element(const v_float32x4& a)
-{
-    return v_setall_f32(v_extract_n<i>(a));
-}
-
-////// FP16 support ///////
-#if CV_FP16
-inline v_float32x4 v_load_expand(const float16_t* ptr)
-{
-#ifndef msa_ld1_f16
-    v4f16 v = (v4f16)msa_ld1_s16((const short*)ptr);
-#else
-    v4f16 v = msa_ld1_f16((const __fp16*)ptr);
-#endif
-    return v_float32x4(msa_cvt_f32_f16(v));
-}
-
-inline void v_pack_store(float16_t* ptr, const v_float32x4& v)
-{
-    v4f16 hv = msa_cvt_f16_f32(v.val);
-
-#ifndef msa_st1_f16
-    msa_st1_s16((short*)ptr, (int16x4_t)hv);
-#else
-    msa_st1_f16((__fp16*)ptr, hv);
-#endif
-}
-#else
-inline v_float32x4 v_load_expand(const float16_t* ptr)
-{
-    float buf[4];
-    for( int i = 0; i < 4; i++ )
-        buf[i] = (float)ptr[i];
-    return v_load(buf);
-}
-
-inline void v_pack_store(float16_t* ptr, const v_float32x4& v)
-{
-    float buf[4];
-    v_store(buf, v);
-    for( int i = 0; i < 4; i++ )
-        ptr[i] = (float16_t)buf[i];
-}
-#endif
-
-inline void v_cleanup() {}
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END
-
-//! @endcond
-
-}
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_neon.hpp b/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_neon.hpp
deleted file mode 100644
index e17972a3..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_neon.hpp
+++ /dev/null
@@ -1,2613 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Copyright (C) 2015, Itseez Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_HAL_INTRIN_NEON_HPP
-#define OPENCV_HAL_INTRIN_NEON_HPP
-
-#include <algorithm>
-#include "opencv2/core/utility.hpp"
-
-namespace cv
-{
-
-//! @cond IGNORED
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
-
-#define CV_SIMD128 1
-#if defined(__aarch64__) || defined(_M_ARM64)
-#define CV_SIMD128_64F 1
-#else
-#define CV_SIMD128_64F 0
-#endif
-
-// The following macro checks if the code is being compiled for the
-// AArch64 execution state of Armv8, to enable the 128-bit
-// intrinsics. The macro `__ARM_64BIT_STATE` is the one recommended by
-// the Arm C Language Extension (ACLE) specifications [1] to check the
-// availability of 128-bit intrinsics, and it is supporrted by clang
-// and gcc. The macro `_M_ARM64` is the equivalent one for Microsoft
-// Visual Studio [2] .
-//
-// [1] https://developer.arm.com/documentation/101028/0012/13--Advanced-SIMD--Neon--intrinsics
-// [2] https://docs.microsoft.com/en-us/cpp/preprocessor/predefined-macros
-#if defined(__ARM_64BIT_STATE) || defined(_M_ARM64)
-#define CV_NEON_AARCH64 1
-#else
-#define CV_NEON_AARCH64 0
-#endif
-
-// TODO
-#define CV_NEON_DOT 0
-
-//////////// Utils ////////////
-
-#if CV_SIMD128_64F
-#define OPENCV_HAL_IMPL_NEON_UNZIP(_Tpv, _Tpvx2, suffix) \
-    inline void _v128_unzip(const _Tpv& a, const _Tpv& b, _Tpv& c, _Tpv& d) \
-    { c = vuzp1q_##suffix(a, b); d = vuzp2q_##suffix(a, b); }
-#define OPENCV_HAL_IMPL_NEON_UNZIP_L(_Tpv, _Tpvx2, suffix) \
-    inline void _v128_unzip(const _Tpv&a, const _Tpv&b, _Tpv& c, _Tpv& d) \
-    { c = vuzp1_##suffix(a, b); d = vuzp2_##suffix(a, b); }
-#else
-#define OPENCV_HAL_IMPL_NEON_UNZIP(_Tpv, _Tpvx2, suffix) \
-    inline void _v128_unzip(const _Tpv& a, const _Tpv& b, _Tpv& c, _Tpv& d) \
-    { _Tpvx2 ab = vuzpq_##suffix(a, b); c = ab.val[0]; d = ab.val[1]; }
-#define OPENCV_HAL_IMPL_NEON_UNZIP_L(_Tpv, _Tpvx2, suffix) \
-    inline void _v128_unzip(const _Tpv& a, const _Tpv& b, _Tpv& c, _Tpv& d) \
-    { _Tpvx2 ab = vuzp_##suffix(a, b); c = ab.val[0]; d = ab.val[1]; }
-#endif
-
-#if CV_SIMD128_64F
-#define OPENCV_HAL_IMPL_NEON_REINTERPRET(_Tpv, suffix) \
-    template <typename T> static inline \
-    _Tpv vreinterpretq_##suffix##_f64(T a) { return (_Tpv) a; } \
-    template <typename T> static inline \
-    float64x2_t vreinterpretq_f64_##suffix(T a) { return (float64x2_t) a; }
-#else
-#define OPENCV_HAL_IMPL_NEON_REINTERPRET(_Tpv, suffix)
-#endif
-
-#define OPENCV_HAL_IMPL_NEON_UTILS_SUFFIX(_Tpv, _Tpvl, suffix) \
-    OPENCV_HAL_IMPL_NEON_UNZIP(_Tpv##_t, _Tpv##x2_t, suffix) \
-    OPENCV_HAL_IMPL_NEON_UNZIP_L(_Tpvl##_t, _Tpvl##x2_t, suffix) \
-    OPENCV_HAL_IMPL_NEON_REINTERPRET(_Tpv##_t, suffix)
-
-#define OPENCV_HAL_IMPL_NEON_UTILS_SUFFIX_I64(_Tpv, _Tpvl, suffix) \
-    OPENCV_HAL_IMPL_NEON_REINTERPRET(_Tpv##_t, suffix)
-
-#define OPENCV_HAL_IMPL_NEON_UTILS_SUFFIX_F64(_Tpv, _Tpvl, suffix) \
-    OPENCV_HAL_IMPL_NEON_UNZIP(_Tpv##_t, _Tpv##x2_t, suffix)
-
-OPENCV_HAL_IMPL_NEON_UTILS_SUFFIX(uint8x16, uint8x8,  u8)
-OPENCV_HAL_IMPL_NEON_UTILS_SUFFIX(int8x16,  int8x8,   s8)
-OPENCV_HAL_IMPL_NEON_UTILS_SUFFIX(uint16x8, uint16x4, u16)
-OPENCV_HAL_IMPL_NEON_UTILS_SUFFIX(int16x8,  int16x4,  s16)
-OPENCV_HAL_IMPL_NEON_UTILS_SUFFIX(uint32x4, uint32x2, u32)
-OPENCV_HAL_IMPL_NEON_UTILS_SUFFIX(int32x4,  int32x2,  s32)
-OPENCV_HAL_IMPL_NEON_UTILS_SUFFIX(float32x4, float32x2, f32)
-OPENCV_HAL_IMPL_NEON_UTILS_SUFFIX_I64(uint64x2, uint64x1, u64)
-OPENCV_HAL_IMPL_NEON_UTILS_SUFFIX_I64(int64x2,  int64x1,  s64)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_NEON_UTILS_SUFFIX_F64(float64x2, float64x1,f64)
-#endif
-
-//////////// Types ////////////
-
-struct v_uint8x16
-{
-    typedef uchar lane_type;
-    enum { nlanes = 16 };
-
-    v_uint8x16() {}
-    explicit v_uint8x16(uint8x16_t v) : val(v) {}
-    v_uint8x16(uchar v0, uchar v1, uchar v2, uchar v3, uchar v4, uchar v5, uchar v6, uchar v7,
-               uchar v8, uchar v9, uchar v10, uchar v11, uchar v12, uchar v13, uchar v14, uchar v15)
-    {
-        uchar v[] = {v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15};
-        val = vld1q_u8(v);
-    }
-    uchar get0() const
-    {
-        return vgetq_lane_u8(val, 0);
-    }
-
-    uint8x16_t val;
-};
-
-struct v_int8x16
-{
-    typedef schar lane_type;
-    enum { nlanes = 16 };
-
-    v_int8x16() {}
-    explicit v_int8x16(int8x16_t v) : val(v) {}
-    v_int8x16(schar v0, schar v1, schar v2, schar v3, schar v4, schar v5, schar v6, schar v7,
-               schar v8, schar v9, schar v10, schar v11, schar v12, schar v13, schar v14, schar v15)
-    {
-        schar v[] = {v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15};
-        val = vld1q_s8(v);
-    }
-    schar get0() const
-    {
-        return vgetq_lane_s8(val, 0);
-    }
-
-    int8x16_t val;
-};
-
-struct v_uint16x8
-{
-    typedef ushort lane_type;
-    enum { nlanes = 8 };
-
-    v_uint16x8() {}
-    explicit v_uint16x8(uint16x8_t v) : val(v) {}
-    v_uint16x8(ushort v0, ushort v1, ushort v2, ushort v3, ushort v4, ushort v5, ushort v6, ushort v7)
-    {
-        ushort v[] = {v0, v1, v2, v3, v4, v5, v6, v7};
-        val = vld1q_u16(v);
-    }
-    ushort get0() const
-    {
-        return vgetq_lane_u16(val, 0);
-    }
-
-    uint16x8_t val;
-};
-
-struct v_int16x8
-{
-    typedef short lane_type;
-    enum { nlanes = 8 };
-
-    v_int16x8() {}
-    explicit v_int16x8(int16x8_t v) : val(v) {}
-    v_int16x8(short v0, short v1, short v2, short v3, short v4, short v5, short v6, short v7)
-    {
-        short v[] = {v0, v1, v2, v3, v4, v5, v6, v7};
-        val = vld1q_s16(v);
-    }
-    short get0() const
-    {
-        return vgetq_lane_s16(val, 0);
-    }
-
-    int16x8_t val;
-};
-
-struct v_uint32x4
-{
-    typedef unsigned lane_type;
-    enum { nlanes = 4 };
-
-    v_uint32x4() {}
-    explicit v_uint32x4(uint32x4_t v) : val(v) {}
-    v_uint32x4(unsigned v0, unsigned v1, unsigned v2, unsigned v3)
-    {
-        unsigned v[] = {v0, v1, v2, v3};
-        val = vld1q_u32(v);
-    }
-    unsigned get0() const
-    {
-        return vgetq_lane_u32(val, 0);
-    }
-
-    uint32x4_t val;
-};
-
-struct v_int32x4
-{
-    typedef int lane_type;
-    enum { nlanes = 4 };
-
-    v_int32x4() {}
-    explicit v_int32x4(int32x4_t v) : val(v) {}
-    v_int32x4(int v0, int v1, int v2, int v3)
-    {
-        int v[] = {v0, v1, v2, v3};
-        val = vld1q_s32(v);
-    }
-    int get0() const
-    {
-        return vgetq_lane_s32(val, 0);
-    }
-    int32x4_t val;
-};
-
-struct v_float32x4
-{
-    typedef float lane_type;
-    enum { nlanes = 4 };
-
-    v_float32x4() {}
-    explicit v_float32x4(float32x4_t v) : val(v) {}
-    v_float32x4(float v0, float v1, float v2, float v3)
-    {
-        float v[] = {v0, v1, v2, v3};
-        val = vld1q_f32(v);
-    }
-    float get0() const
-    {
-        return vgetq_lane_f32(val, 0);
-    }
-    float32x4_t val;
-};
-
-struct v_uint64x2
-{
-    typedef uint64 lane_type;
-    enum { nlanes = 2 };
-
-    v_uint64x2() {}
-    explicit v_uint64x2(uint64x2_t v) : val(v) {}
-    v_uint64x2(uint64 v0, uint64 v1)
-    {
-        uint64 v[] = {v0, v1};
-        val = vld1q_u64(v);
-    }
-    uint64 get0() const
-    {
-        return vgetq_lane_u64(val, 0);
-    }
-    uint64x2_t val;
-};
-
-struct v_int64x2
-{
-    typedef int64 lane_type;
-    enum { nlanes = 2 };
-
-    v_int64x2() {}
-    explicit v_int64x2(int64x2_t v) : val(v) {}
-    v_int64x2(int64 v0, int64 v1)
-    {
-        int64 v[] = {v0, v1};
-        val = vld1q_s64(v);
-    }
-    int64 get0() const
-    {
-        return vgetq_lane_s64(val, 0);
-    }
-    int64x2_t val;
-};
-
-#if CV_SIMD128_64F
-struct v_float64x2
-{
-    typedef double lane_type;
-    enum { nlanes = 2 };
-
-    v_float64x2() {}
-    explicit v_float64x2(float64x2_t v) : val(v) {}
-    v_float64x2(double v0, double v1)
-    {
-        double v[] = {v0, v1};
-        val = vld1q_f64(v);
-    }
-    double get0() const
-    {
-        return vgetq_lane_f64(val, 0);
-    }
-    float64x2_t val;
-};
-#endif
-
-#define OPENCV_HAL_IMPL_NEON_INIT(_Tpv, _Tp, suffix) \
-inline v_##_Tpv v_setzero_##suffix() { return v_##_Tpv(vdupq_n_##suffix((_Tp)0)); } \
-inline v_##_Tpv v_setall_##suffix(_Tp v) { return v_##_Tpv(vdupq_n_##suffix(v)); } \
-inline _Tpv##_t vreinterpretq_##suffix##_##suffix(_Tpv##_t v) { return v; } \
-inline v_uint8x16 v_reinterpret_as_u8(const v_##_Tpv& v) { return v_uint8x16(vreinterpretq_u8_##suffix(v.val)); } \
-inline v_int8x16 v_reinterpret_as_s8(const v_##_Tpv& v) { return v_int8x16(vreinterpretq_s8_##suffix(v.val)); } \
-inline v_uint16x8 v_reinterpret_as_u16(const v_##_Tpv& v) { return v_uint16x8(vreinterpretq_u16_##suffix(v.val)); } \
-inline v_int16x8 v_reinterpret_as_s16(const v_##_Tpv& v) { return v_int16x8(vreinterpretq_s16_##suffix(v.val)); } \
-inline v_uint32x4 v_reinterpret_as_u32(const v_##_Tpv& v) { return v_uint32x4(vreinterpretq_u32_##suffix(v.val)); } \
-inline v_int32x4 v_reinterpret_as_s32(const v_##_Tpv& v) { return v_int32x4(vreinterpretq_s32_##suffix(v.val)); } \
-inline v_uint64x2 v_reinterpret_as_u64(const v_##_Tpv& v) { return v_uint64x2(vreinterpretq_u64_##suffix(v.val)); } \
-inline v_int64x2 v_reinterpret_as_s64(const v_##_Tpv& v) { return v_int64x2(vreinterpretq_s64_##suffix(v.val)); } \
-inline v_float32x4 v_reinterpret_as_f32(const v_##_Tpv& v) { return v_float32x4(vreinterpretq_f32_##suffix(v.val)); }
-
-OPENCV_HAL_IMPL_NEON_INIT(uint8x16, uchar, u8)
-OPENCV_HAL_IMPL_NEON_INIT(int8x16, schar, s8)
-OPENCV_HAL_IMPL_NEON_INIT(uint16x8, ushort, u16)
-OPENCV_HAL_IMPL_NEON_INIT(int16x8, short, s16)
-OPENCV_HAL_IMPL_NEON_INIT(uint32x4, unsigned, u32)
-OPENCV_HAL_IMPL_NEON_INIT(int32x4, int, s32)
-OPENCV_HAL_IMPL_NEON_INIT(uint64x2, uint64, u64)
-OPENCV_HAL_IMPL_NEON_INIT(int64x2, int64, s64)
-OPENCV_HAL_IMPL_NEON_INIT(float32x4, float, f32)
-#if CV_SIMD128_64F
-#define OPENCV_HAL_IMPL_NEON_INIT_64(_Tpv, suffix) \
-inline v_float64x2 v_reinterpret_as_f64(const v_##_Tpv& v) { return v_float64x2(vreinterpretq_f64_##suffix(v.val)); }
-OPENCV_HAL_IMPL_NEON_INIT(float64x2, double, f64)
-OPENCV_HAL_IMPL_NEON_INIT_64(uint8x16, u8)
-OPENCV_HAL_IMPL_NEON_INIT_64(int8x16, s8)
-OPENCV_HAL_IMPL_NEON_INIT_64(uint16x8, u16)
-OPENCV_HAL_IMPL_NEON_INIT_64(int16x8, s16)
-OPENCV_HAL_IMPL_NEON_INIT_64(uint32x4, u32)
-OPENCV_HAL_IMPL_NEON_INIT_64(int32x4, s32)
-OPENCV_HAL_IMPL_NEON_INIT_64(uint64x2, u64)
-OPENCV_HAL_IMPL_NEON_INIT_64(int64x2, s64)
-OPENCV_HAL_IMPL_NEON_INIT_64(float32x4, f32)
-OPENCV_HAL_IMPL_NEON_INIT_64(float64x2, f64)
-#endif
-
-#define OPENCV_HAL_IMPL_NEON_PACK(_Tpvec, _Tp, hreg, suffix, _Tpwvec, pack, mov, rshr) \
-inline _Tpvec v_##pack(const _Tpwvec& a, const _Tpwvec& b) \
-{ \
-    hreg a1 = mov(a.val), b1 = mov(b.val); \
-    return _Tpvec(vcombine_##suffix(a1, b1)); \
-} \
-inline void v_##pack##_store(_Tp* ptr, const _Tpwvec& a) \
-{ \
-    hreg a1 = mov(a.val); \
-    vst1_##suffix(ptr, a1); \
-} \
-template<int n> inline \
-_Tpvec v_rshr_##pack(const _Tpwvec& a, const _Tpwvec& b) \
-{ \
-    hreg a1 = rshr(a.val, n); \
-    hreg b1 = rshr(b.val, n); \
-    return _Tpvec(vcombine_##suffix(a1, b1)); \
-} \
-template<int n> inline \
-void v_rshr_##pack##_store(_Tp* ptr, const _Tpwvec& a) \
-{ \
-    hreg a1 = rshr(a.val, n); \
-    vst1_##suffix(ptr, a1); \
-}
-
-OPENCV_HAL_IMPL_NEON_PACK(v_uint8x16, uchar, uint8x8_t, u8, v_uint16x8, pack, vqmovn_u16, vqrshrn_n_u16)
-OPENCV_HAL_IMPL_NEON_PACK(v_int8x16, schar, int8x8_t, s8, v_int16x8, pack, vqmovn_s16, vqrshrn_n_s16)
-OPENCV_HAL_IMPL_NEON_PACK(v_uint16x8, ushort, uint16x4_t, u16, v_uint32x4, pack, vqmovn_u32, vqrshrn_n_u32)
-OPENCV_HAL_IMPL_NEON_PACK(v_int16x8, short, int16x4_t, s16, v_int32x4, pack, vqmovn_s32, vqrshrn_n_s32)
-OPENCV_HAL_IMPL_NEON_PACK(v_uint32x4, unsigned, uint32x2_t, u32, v_uint64x2, pack, vmovn_u64, vrshrn_n_u64)
-OPENCV_HAL_IMPL_NEON_PACK(v_int32x4, int, int32x2_t, s32, v_int64x2, pack, vmovn_s64, vrshrn_n_s64)
-
-OPENCV_HAL_IMPL_NEON_PACK(v_uint8x16, uchar, uint8x8_t, u8, v_int16x8, pack_u, vqmovun_s16, vqrshrun_n_s16)
-OPENCV_HAL_IMPL_NEON_PACK(v_uint16x8, ushort, uint16x4_t, u16, v_int32x4, pack_u, vqmovun_s32, vqrshrun_n_s32)
-
-// pack boolean
-inline v_uint8x16 v_pack_b(const v_uint16x8& a, const v_uint16x8& b)
-{
-    uint8x16_t ab = vcombine_u8(vmovn_u16(a.val), vmovn_u16(b.val));
-    return v_uint8x16(ab);
-}
-
-inline v_uint8x16 v_pack_b(const v_uint32x4& a, const v_uint32x4& b,
-                           const v_uint32x4& c, const v_uint32x4& d)
-{
-    uint16x8_t nab = vcombine_u16(vmovn_u32(a.val), vmovn_u32(b.val));
-    uint16x8_t ncd = vcombine_u16(vmovn_u32(c.val), vmovn_u32(d.val));
-    return v_uint8x16(vcombine_u8(vmovn_u16(nab), vmovn_u16(ncd)));
-}
-
-inline v_uint8x16 v_pack_b(const v_uint64x2& a, const v_uint64x2& b, const v_uint64x2& c,
-                           const v_uint64x2& d, const v_uint64x2& e, const v_uint64x2& f,
-                           const v_uint64x2& g, const v_uint64x2& h)
-{
-    uint32x4_t ab = vcombine_u32(vmovn_u64(a.val), vmovn_u64(b.val));
-    uint32x4_t cd = vcombine_u32(vmovn_u64(c.val), vmovn_u64(d.val));
-    uint32x4_t ef = vcombine_u32(vmovn_u64(e.val), vmovn_u64(f.val));
-    uint32x4_t gh = vcombine_u32(vmovn_u64(g.val), vmovn_u64(h.val));
-
-    uint16x8_t abcd = vcombine_u16(vmovn_u32(ab), vmovn_u32(cd));
-    uint16x8_t efgh = vcombine_u16(vmovn_u32(ef), vmovn_u32(gh));
-    return v_uint8x16(vcombine_u8(vmovn_u16(abcd), vmovn_u16(efgh)));
-}
-
-inline v_float32x4 v_matmul(const v_float32x4& v, const v_float32x4& m0,
-                            const v_float32x4& m1, const v_float32x4& m2,
-                            const v_float32x4& m3)
-{
-    float32x2_t vl = vget_low_f32(v.val), vh = vget_high_f32(v.val);
-    float32x4_t res = vmulq_lane_f32(m0.val, vl, 0);
-    res = vmlaq_lane_f32(res, m1.val, vl, 1);
-    res = vmlaq_lane_f32(res, m2.val, vh, 0);
-    res = vmlaq_lane_f32(res, m3.val, vh, 1);
-    return v_float32x4(res);
-}
-
-inline v_float32x4 v_matmuladd(const v_float32x4& v, const v_float32x4& m0,
-                               const v_float32x4& m1, const v_float32x4& m2,
-                               const v_float32x4& a)
-{
-    float32x2_t vl = vget_low_f32(v.val), vh = vget_high_f32(v.val);
-    float32x4_t res = vmulq_lane_f32(m0.val, vl, 0);
-    res = vmlaq_lane_f32(res, m1.val, vl, 1);
-    res = vmlaq_lane_f32(res, m2.val, vh, 0);
-    res = vaddq_f32(res, a.val);
-    return v_float32x4(res);
-}
-
-#define OPENCV_HAL_IMPL_NEON_BIN_OP(bin_op, _Tpvec, intrin) \
-inline _Tpvec operator bin_op (const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(intrin(a.val, b.val)); \
-} \
-inline _Tpvec& operator bin_op##= (_Tpvec& a, const _Tpvec& b) \
-{ \
-    a.val = intrin(a.val, b.val); \
-    return a; \
-}
-
-OPENCV_HAL_IMPL_NEON_BIN_OP(+, v_uint8x16, vqaddq_u8)
-OPENCV_HAL_IMPL_NEON_BIN_OP(-, v_uint8x16, vqsubq_u8)
-OPENCV_HAL_IMPL_NEON_BIN_OP(+, v_int8x16, vqaddq_s8)
-OPENCV_HAL_IMPL_NEON_BIN_OP(-, v_int8x16, vqsubq_s8)
-OPENCV_HAL_IMPL_NEON_BIN_OP(+, v_uint16x8, vqaddq_u16)
-OPENCV_HAL_IMPL_NEON_BIN_OP(-, v_uint16x8, vqsubq_u16)
-OPENCV_HAL_IMPL_NEON_BIN_OP(+, v_int16x8, vqaddq_s16)
-OPENCV_HAL_IMPL_NEON_BIN_OP(-, v_int16x8, vqsubq_s16)
-OPENCV_HAL_IMPL_NEON_BIN_OP(+, v_int32x4, vaddq_s32)
-OPENCV_HAL_IMPL_NEON_BIN_OP(-, v_int32x4, vsubq_s32)
-OPENCV_HAL_IMPL_NEON_BIN_OP(*, v_int32x4, vmulq_s32)
-OPENCV_HAL_IMPL_NEON_BIN_OP(+, v_uint32x4, vaddq_u32)
-OPENCV_HAL_IMPL_NEON_BIN_OP(-, v_uint32x4, vsubq_u32)
-OPENCV_HAL_IMPL_NEON_BIN_OP(*, v_uint32x4, vmulq_u32)
-OPENCV_HAL_IMPL_NEON_BIN_OP(+, v_float32x4, vaddq_f32)
-OPENCV_HAL_IMPL_NEON_BIN_OP(-, v_float32x4, vsubq_f32)
-OPENCV_HAL_IMPL_NEON_BIN_OP(*, v_float32x4, vmulq_f32)
-OPENCV_HAL_IMPL_NEON_BIN_OP(+, v_int64x2, vaddq_s64)
-OPENCV_HAL_IMPL_NEON_BIN_OP(-, v_int64x2, vsubq_s64)
-OPENCV_HAL_IMPL_NEON_BIN_OP(+, v_uint64x2, vaddq_u64)
-OPENCV_HAL_IMPL_NEON_BIN_OP(-, v_uint64x2, vsubq_u64)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_NEON_BIN_OP(/, v_float32x4, vdivq_f32)
-OPENCV_HAL_IMPL_NEON_BIN_OP(+, v_float64x2, vaddq_f64)
-OPENCV_HAL_IMPL_NEON_BIN_OP(-, v_float64x2, vsubq_f64)
-OPENCV_HAL_IMPL_NEON_BIN_OP(*, v_float64x2, vmulq_f64)
-OPENCV_HAL_IMPL_NEON_BIN_OP(/, v_float64x2, vdivq_f64)
-#else
-inline v_float32x4 operator / (const v_float32x4& a, const v_float32x4& b)
-{
-    float32x4_t reciprocal = vrecpeq_f32(b.val);
-    reciprocal = vmulq_f32(vrecpsq_f32(b.val, reciprocal), reciprocal);
-    reciprocal = vmulq_f32(vrecpsq_f32(b.val, reciprocal), reciprocal);
-    return v_float32x4(vmulq_f32(a.val, reciprocal));
-}
-inline v_float32x4& operator /= (v_float32x4& a, const v_float32x4& b)
-{
-    float32x4_t reciprocal = vrecpeq_f32(b.val);
-    reciprocal = vmulq_f32(vrecpsq_f32(b.val, reciprocal), reciprocal);
-    reciprocal = vmulq_f32(vrecpsq_f32(b.val, reciprocal), reciprocal);
-    a.val = vmulq_f32(a.val, reciprocal);
-    return a;
-}
-#endif
-
-// saturating multiply 8-bit, 16-bit
-#define OPENCV_HAL_IMPL_NEON_MUL_SAT(_Tpvec, _Tpwvec)            \
-    inline _Tpvec operator * (const _Tpvec& a, const _Tpvec& b)  \
-    {                                                            \
-        _Tpwvec c, d;                                            \
-        v_mul_expand(a, b, c, d);                                \
-        return v_pack(c, d);                                     \
-    }                                                            \
-    inline _Tpvec& operator *= (_Tpvec& a, const _Tpvec& b)      \
-    { a = a * b; return a; }
-
-OPENCV_HAL_IMPL_NEON_MUL_SAT(v_int8x16,  v_int16x8)
-OPENCV_HAL_IMPL_NEON_MUL_SAT(v_uint8x16, v_uint16x8)
-OPENCV_HAL_IMPL_NEON_MUL_SAT(v_int16x8,  v_int32x4)
-OPENCV_HAL_IMPL_NEON_MUL_SAT(v_uint16x8, v_uint32x4)
-
-//  Multiply and expand
-inline void v_mul_expand(const v_int8x16& a, const v_int8x16& b,
-                         v_int16x8& c, v_int16x8& d)
-{
-    c.val = vmull_s8(vget_low_s8(a.val), vget_low_s8(b.val));
-#if CV_NEON_AARCH64
-    d.val = vmull_high_s8(a.val, b.val);
-#else // #if CV_NEON_AARCH64
-    d.val = vmull_s8(vget_high_s8(a.val), vget_high_s8(b.val));
-#endif // #if CV_NEON_AARCH64
-}
-
-inline void v_mul_expand(const v_uint8x16& a, const v_uint8x16& b,
-                         v_uint16x8& c, v_uint16x8& d)
-{
-    c.val = vmull_u8(vget_low_u8(a.val), vget_low_u8(b.val));
-#if CV_NEON_AARCH64
-    d.val = vmull_high_u8(a.val, b.val);
-#else // #if CV_NEON_AARCH64
-    d.val = vmull_u8(vget_high_u8(a.val), vget_high_u8(b.val));
-#endif // #if CV_NEON_AARCH64
-}
-
-inline void v_mul_expand(const v_int16x8& a, const v_int16x8& b,
-                         v_int32x4& c, v_int32x4& d)
-{
-    c.val = vmull_s16(vget_low_s16(a.val), vget_low_s16(b.val));
-#if CV_NEON_AARCH64
-    d.val = vmull_high_s16(a.val, b.val);
-#else // #if CV_NEON_AARCH64
-    d.val = vmull_s16(vget_high_s16(a.val), vget_high_s16(b.val));
-#endif // #if CV_NEON_AARCH64
-}
-
-inline void v_mul_expand(const v_uint16x8& a, const v_uint16x8& b,
-                         v_uint32x4& c, v_uint32x4& d)
-{
-    c.val = vmull_u16(vget_low_u16(a.val), vget_low_u16(b.val));
-#if CV_NEON_AARCH64
-    d.val = vmull_high_u16(a.val, b.val);
-#else // #if CV_NEON_AARCH64
-    d.val = vmull_u16(vget_high_u16(a.val), vget_high_u16(b.val));
-#endif // #if CV_NEON_AARCH64
-}
-
-inline void v_mul_expand(const v_uint32x4& a, const v_uint32x4& b,
-                         v_uint64x2& c, v_uint64x2& d)
-{
-    c.val = vmull_u32(vget_low_u32(a.val), vget_low_u32(b.val));
-#if CV_NEON_AARCH64
-    d.val = vmull_high_u32(a.val, b.val);
-#else // #if CV_NEON_AARCH64
-    d.val = vmull_u32(vget_high_u32(a.val), vget_high_u32(b.val));
-#endif // #if CV_NEON_AARCH64
-}
-
-inline v_int16x8 v_mul_hi(const v_int16x8& a, const v_int16x8& b)
-{
-    return v_int16x8(vcombine_s16(
-                                  vshrn_n_s32(vmull_s16( vget_low_s16(a.val),  vget_low_s16(b.val)), 16),
-                                  vshrn_n_s32(
-#if CV_NEON_AARCH64
-                                    vmull_high_s16(a.val, b.val)
-#else // #if CV_NEON_AARCH64
-                                    vmull_s16(vget_high_s16(a.val), vget_high_s16(b.val))
-#endif // #if CV_NEON_AARCH64
-                                    , 16)
-                                 ));
-}
-inline v_uint16x8 v_mul_hi(const v_uint16x8& a, const v_uint16x8& b)
-{
-    return v_uint16x8(vcombine_u16(
-                                   vshrn_n_u32(vmull_u16( vget_low_u16(a.val),  vget_low_u16(b.val)), 16),
-                                   vshrn_n_u32(
-#if CV_NEON_AARCH64
-                                    vmull_high_u16(a.val, b.val)
-#else // #if CV_NEON_AARCH64
-                                    vmull_u16(vget_high_u16(a.val), vget_high_u16(b.val))
-#endif // #if CV_NEON_AARCH64
-                                    , 16)
-                                  ));
-}
-
-//////// Dot Product ////////
-
-// 16 >> 32
-inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b)
-{
-    int16x8_t uzp1, uzp2;
-    _v128_unzip(a.val, b.val, uzp1, uzp2);
-    int16x4_t a0 = vget_low_s16(uzp1);
-    int16x4_t b0 = vget_high_s16(uzp1);
-    int16x4_t a1 = vget_low_s16(uzp2);
-    int16x4_t b1 = vget_high_s16(uzp2);
-    int32x4_t p = vmull_s16(a0, b0);
-    return v_int32x4(vmlal_s16(p, a1, b1));
-}
-inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b, const v_int32x4& c)
-{
-    int16x8_t uzp1, uzp2;
-    _v128_unzip(a.val, b.val, uzp1, uzp2);
-    int16x4_t a0 = vget_low_s16(uzp1);
-    int16x4_t b0 = vget_high_s16(uzp1);
-    int16x4_t a1 = vget_low_s16(uzp2);
-    int16x4_t b1 = vget_high_s16(uzp2);
-    int32x4_t p = vmlal_s16(c.val, a0, b0);
-    return v_int32x4(vmlal_s16(p, a1, b1));
-}
-
-// 32 >> 64
-inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b)
-{
-    int32x4_t uzp1, uzp2;
-    _v128_unzip(a.val, b.val, uzp1, uzp2);
-    int32x2_t a0 = vget_low_s32(uzp1);
-    int32x2_t b0 = vget_high_s32(uzp1);
-    int32x2_t a1 = vget_low_s32(uzp2);
-    int32x2_t b1 = vget_high_s32(uzp2);
-    int64x2_t p = vmull_s32(a0, b0);
-    return v_int64x2(vmlal_s32(p, a1, b1));
-}
-inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b, const v_int64x2& c)
-{
-    int32x4_t uzp1, uzp2;
-    _v128_unzip(a.val, b.val, uzp1, uzp2);
-    int32x2_t a0 = vget_low_s32(uzp1);
-    int32x2_t b0 = vget_high_s32(uzp1);
-    int32x2_t a1 = vget_low_s32(uzp2);
-    int32x2_t b1 = vget_high_s32(uzp2);
-    int64x2_t p = vmlal_s32(c.val, a0, b0);
-    return v_int64x2(vmlal_s32(p, a1, b1));
-}
-
-// 8 >> 32
-inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b)
-{
-#if CV_NEON_DOT
-    return v_uint32x4(vdotq_u32(vdupq_n_u32(0), a.val, b.val));
-#else
-    const uint8x16_t zero   = vreinterpretq_u8_u32(vdupq_n_u32(0));
-    const uint8x16_t mask   = vreinterpretq_u8_u32(vdupq_n_u32(0x00FF00FF));
-    const uint16x8_t zero32 = vreinterpretq_u16_u32(vdupq_n_u32(0));
-    const uint16x8_t mask32 = vreinterpretq_u16_u32(vdupq_n_u32(0x0000FFFF));
-
-    uint16x8_t even = vmulq_u16(vreinterpretq_u16_u8(vbslq_u8(mask, a.val, zero)),
-                                vreinterpretq_u16_u8(vbslq_u8(mask, b.val, zero)));
-    uint16x8_t odd  = vmulq_u16(vshrq_n_u16(vreinterpretq_u16_u8(a.val), 8),
-                                vshrq_n_u16(vreinterpretq_u16_u8(b.val), 8));
-
-    uint32x4_t s0 = vaddq_u32(vreinterpretq_u32_u16(vbslq_u16(mask32, even, zero32)),
-                              vreinterpretq_u32_u16(vbslq_u16(mask32, odd,  zero32)));
-    uint32x4_t s1 = vaddq_u32(vshrq_n_u32(vreinterpretq_u32_u16(even), 16),
-                              vshrq_n_u32(vreinterpretq_u32_u16(odd),  16));
-    return v_uint32x4(vaddq_u32(s0, s1));
-#endif
-}
-inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b,
-                                   const v_uint32x4& c)
-{
-#if CV_NEON_DOT
-    return v_uint32x4(vdotq_u32(c.val, a.val, b.val));
-#else
-    return v_dotprod_expand(a, b) + c;
-#endif
-}
-
-inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b)
-{
-#if CV_NEON_DOT
-    return v_int32x4(vdotq_s32(vdupq_n_s32(0), a.val, b.val));
-#else
-    int16x8_t p0  = vmull_s8(vget_low_s8(a.val), vget_low_s8(b.val));
-    int16x8_t p1  = vmull_s8(vget_high_s8(a.val), vget_high_s8(b.val));
-    int16x8_t uzp1, uzp2;
-    _v128_unzip(p0, p1, uzp1, uzp2);
-    int16x8_t sum = vaddq_s16(uzp1, uzp2);
-    int16x4_t uzpl1, uzpl2;
-    _v128_unzip(vget_low_s16(sum), vget_high_s16(sum), uzpl1, uzpl2);
-    return v_int32x4(vaddl_s16(uzpl1, uzpl2));
-#endif
-}
-inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b,
-                                  const v_int32x4& c)
-{
-#if CV_NEON_DOT
-    return v_int32x4(vdotq_s32(c.val, a.val, b.val));
-#else
-    return v_dotprod_expand(a, b) + c;
-#endif
-}
-
-// 16 >> 64
-inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b)
-{
-    const uint16x8_t zero = vreinterpretq_u16_u32(vdupq_n_u32(0));
-    const uint16x8_t mask = vreinterpretq_u16_u32(vdupq_n_u32(0x0000FFFF));
-
-    uint32x4_t even = vmulq_u32(vreinterpretq_u32_u16(vbslq_u16(mask, a.val, zero)),
-                                vreinterpretq_u32_u16(vbslq_u16(mask, b.val, zero)));
-    uint32x4_t odd  = vmulq_u32(vshrq_n_u32(vreinterpretq_u32_u16(a.val), 16),
-                                vshrq_n_u32(vreinterpretq_u32_u16(b.val), 16));
-    uint32x4_t uzp1, uzp2;
-    _v128_unzip(even, odd, uzp1, uzp2);
-    uint64x2_t s0  = vaddl_u32(vget_low_u32(uzp1), vget_high_u32(uzp1));
-    uint64x2_t s1  = vaddl_u32(vget_low_u32(uzp2), vget_high_u32(uzp2));
-    return v_uint64x2(vaddq_u64(s0, s1));
-}
-inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b, const v_uint64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-inline v_int64x2 v_dotprod_expand(const v_int16x8& a, const v_int16x8& b)
-{
-    int32x4_t p0  = vmull_s16(vget_low_s16(a.val),  vget_low_s16(b.val));
-    int32x4_t p1  = vmull_s16(vget_high_s16(a.val), vget_high_s16(b.val));
-
-    int32x4_t uzp1, uzp2;
-    _v128_unzip(p0, p1, uzp1, uzp2);
-    int32x4_t sum = vaddq_s32(uzp1, uzp2);
-
-    int32x2_t uzpl1, uzpl2;
-    _v128_unzip(vget_low_s32(sum), vget_high_s32(sum), uzpl1, uzpl2);
-    return v_int64x2(vaddl_s32(uzpl1, uzpl2));
-}
-inline v_int64x2 v_dotprod_expand(const v_int16x8& a, const v_int16x8& b,
-                                  const v_int64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-// 32 >> 64f
-#if CV_SIMD128_64F
-inline v_float64x2 v_dotprod_expand(const v_int32x4& a, const v_int32x4& b)
-{ return v_cvt_f64(v_dotprod(a, b)); }
-inline v_float64x2 v_dotprod_expand(const v_int32x4& a,   const v_int32x4& b,
-                                    const v_float64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-#endif
-
-//////// Fast Dot Product ////////
-
-// 16 >> 32
-inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b)
-{
-#if CV_NEON_AARCH64
-    int32x4_t p = vmull_s16(vget_low_s16(a.val), vget_low_s16(b.val));
-    return v_int32x4(vmlal_high_s16(p, a.val, b.val));
-#else
-    int16x4_t a0 = vget_low_s16(a.val);
-    int16x4_t a1 = vget_high_s16(a.val);
-    int16x4_t b0 = vget_low_s16(b.val);
-    int16x4_t b1 = vget_high_s16(b.val);
-    int32x4_t p = vmull_s16(a0, b0);
-    return v_int32x4(vmlal_s16(p, a1, b1));
-#endif
-}
-inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b, const v_int32x4& c)
-{
-#if CV_NEON_AARCH64
-    int32x4_t p = vmlal_s16(c.val, vget_low_s16(a.val), vget_low_s16(b.val));
-    return v_int32x4(vmlal_high_s16(p, a.val, b.val));
-#else
-    int16x4_t a0 = vget_low_s16(a.val);
-    int16x4_t a1 = vget_high_s16(a.val);
-    int16x4_t b0 = vget_low_s16(b.val);
-    int16x4_t b1 = vget_high_s16(b.val);
-    int32x4_t p = vmlal_s16(c.val, a0, b0);
-    return v_int32x4(vmlal_s16(p, a1, b1));
-#endif
-}
-
-// 32 >> 64
-inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b)
-{
-#if CV_NEON_AARCH64
-    int64x2_t p = vmull_s32(vget_low_s32(a.val), vget_low_s32(b.val));
-    return v_int64x2(vmlal_high_s32(p, a.val, b.val));
-#else
-    int32x2_t a0 = vget_low_s32(a.val);
-    int32x2_t a1 = vget_high_s32(a.val);
-    int32x2_t b0 = vget_low_s32(b.val);
-    int32x2_t b1 = vget_high_s32(b.val);
-    int64x2_t p = vmull_s32(a0, b0);
-    return v_int64x2(vmlal_s32(p, a1, b1));
-#endif
-}
-inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b, const v_int64x2& c)
-{
-#if CV_NEON_AARCH64
-    int64x2_t p = vmlal_s32(c.val, vget_low_s32(a.val), vget_low_s32(b.val));
-    return v_int64x2(vmlal_high_s32(p, a.val, b.val));
-#else
-    int32x2_t a0 = vget_low_s32(a.val);
-    int32x2_t a1 = vget_high_s32(a.val);
-    int32x2_t b0 = vget_low_s32(b.val);
-    int32x2_t b1 = vget_high_s32(b.val);
-    int64x2_t p = vmlal_s32(c.val, a0, b0);
-    return v_int64x2(vmlal_s32(p, a1, b1));
-#endif
-}
-
-// 8 >> 32
-inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b)
-{
-#if CV_NEON_DOT
-    return v_uint32x4(vdotq_u32(vdupq_n_u32(0), a.val, b.val));
-#else
-    uint16x8_t p0 = vmull_u8(vget_low_u8(a.val), vget_low_u8(b.val));
-    uint16x8_t p1 = vmull_u8(vget_high_u8(a.val), vget_high_u8(b.val));
-    uint32x4_t s0 = vaddl_u16(vget_low_u16(p0), vget_low_u16(p1));
-    uint32x4_t s1 = vaddl_u16(vget_high_u16(p0), vget_high_u16(p1));
-    return v_uint32x4(vaddq_u32(s0, s1));
-#endif
-}
-inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b, const v_uint32x4& c)
-{
-#if CV_NEON_DOT
-    return v_uint32x4(vdotq_u32(c.val, a.val, b.val));
-#else
-    return v_dotprod_expand_fast(a, b) + c;
-#endif
-}
-
-inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b)
-{
-#if CV_NEON_DOT
-    return v_int32x4(vdotq_s32(vdupq_n_s32(0), a.val, b.val));
-#else
-    int16x8_t prod = vmull_s8(vget_low_s8(a.val), vget_low_s8(b.val));
-    prod = vmlal_s8(prod, vget_high_s8(a.val), vget_high_s8(b.val));
-    return v_int32x4(vaddl_s16(vget_low_s16(prod), vget_high_s16(prod)));
-#endif
-}
-inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b, const v_int32x4& c)
-{
-#if CV_NEON_DOT
-    return v_int32x4(vdotq_s32(c.val, a.val, b.val));
-#else
-    return v_dotprod_expand_fast(a, b) + c;
-#endif
-}
-
-// 16 >> 64
-inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b)
-{
-    uint32x4_t p0  = vmull_u16(vget_low_u16(a.val),  vget_low_u16(b.val));
-    uint32x4_t p1  = vmull_u16(vget_high_u16(a.val), vget_high_u16(b.val));
-    uint64x2_t s0  = vaddl_u32(vget_low_u32(p0), vget_high_u32(p0));
-    uint64x2_t s1  = vaddl_u32(vget_low_u32(p1), vget_high_u32(p1));
-    return v_uint64x2(vaddq_u64(s0, s1));
-}
-inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b, const v_uint64x2& c)
-{ return v_dotprod_expand_fast(a, b) + c; }
-
-inline v_int64x2 v_dotprod_expand_fast(const v_int16x8& a, const v_int16x8& b)
-{
-    int32x4_t prod = vmull_s16(vget_low_s16(a.val), vget_low_s16(b.val));
-    prod = vmlal_s16(prod, vget_high_s16(a.val), vget_high_s16(b.val));
-    return v_int64x2(vaddl_s32(vget_low_s32(prod), vget_high_s32(prod)));
-}
-inline v_int64x2 v_dotprod_expand_fast(const v_int16x8& a, const v_int16x8& b, const v_int64x2& c)
-{ return v_dotprod_expand_fast(a, b) + c; }
-
-// 32 >> 64f
-#if CV_SIMD128_64F
-inline v_float64x2 v_dotprod_expand_fast(const v_int32x4& a, const v_int32x4& b)
-{ return v_cvt_f64(v_dotprod_fast(a, b)); }
-inline v_float64x2 v_dotprod_expand_fast(const v_int32x4& a, const v_int32x4& b, const v_float64x2& c)
-{ return v_dotprod_expand_fast(a, b) + c; }
-#endif
-
-
-#define OPENCV_HAL_IMPL_NEON_LOGIC_OP(_Tpvec, suffix) \
-    OPENCV_HAL_IMPL_NEON_BIN_OP(&, _Tpvec, vandq_##suffix) \
-    OPENCV_HAL_IMPL_NEON_BIN_OP(|, _Tpvec, vorrq_##suffix) \
-    OPENCV_HAL_IMPL_NEON_BIN_OP(^, _Tpvec, veorq_##suffix) \
-    inline _Tpvec operator ~ (const _Tpvec& a) \
-    { \
-        return _Tpvec(vreinterpretq_##suffix##_u8(vmvnq_u8(vreinterpretq_u8_##suffix(a.val)))); \
-    }
-
-OPENCV_HAL_IMPL_NEON_LOGIC_OP(v_uint8x16, u8)
-OPENCV_HAL_IMPL_NEON_LOGIC_OP(v_int8x16, s8)
-OPENCV_HAL_IMPL_NEON_LOGIC_OP(v_uint16x8, u16)
-OPENCV_HAL_IMPL_NEON_LOGIC_OP(v_int16x8, s16)
-OPENCV_HAL_IMPL_NEON_LOGIC_OP(v_uint32x4, u32)
-OPENCV_HAL_IMPL_NEON_LOGIC_OP(v_int32x4, s32)
-OPENCV_HAL_IMPL_NEON_LOGIC_OP(v_uint64x2, u64)
-OPENCV_HAL_IMPL_NEON_LOGIC_OP(v_int64x2, s64)
-
-#define OPENCV_HAL_IMPL_NEON_FLT_BIT_OP(bin_op, intrin) \
-inline v_float32x4 operator bin_op (const v_float32x4& a, const v_float32x4& b) \
-{ \
-    return v_float32x4(vreinterpretq_f32_s32(intrin(vreinterpretq_s32_f32(a.val), vreinterpretq_s32_f32(b.val)))); \
-} \
-inline v_float32x4& operator bin_op##= (v_float32x4& a, const v_float32x4& b) \
-{ \
-    a.val = vreinterpretq_f32_s32(intrin(vreinterpretq_s32_f32(a.val), vreinterpretq_s32_f32(b.val))); \
-    return a; \
-}
-
-OPENCV_HAL_IMPL_NEON_FLT_BIT_OP(&, vandq_s32)
-OPENCV_HAL_IMPL_NEON_FLT_BIT_OP(|, vorrq_s32)
-OPENCV_HAL_IMPL_NEON_FLT_BIT_OP(^, veorq_s32)
-
-inline v_float32x4 operator ~ (const v_float32x4& a)
-{
-    return v_float32x4(vreinterpretq_f32_s32(vmvnq_s32(vreinterpretq_s32_f32(a.val))));
-}
-
-#if CV_SIMD128_64F
-inline v_float32x4 v_sqrt(const v_float32x4& x)
-{
-    return v_float32x4(vsqrtq_f32(x.val));
-}
-
-inline v_float32x4 v_invsqrt(const v_float32x4& x)
-{
-    v_float32x4 one = v_setall_f32(1.0f);
-    return one / v_sqrt(x);
-}
-#else
-inline v_float32x4 v_sqrt(const v_float32x4& x)
-{
-    float32x4_t x1 = vmaxq_f32(x.val, vdupq_n_f32(FLT_MIN));
-    float32x4_t e = vrsqrteq_f32(x1);
-    e = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x1, e), e), e);
-    e = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x1, e), e), e);
-    return v_float32x4(vmulq_f32(x.val, e));
-}
-
-inline v_float32x4 v_invsqrt(const v_float32x4& x)
-{
-    float32x4_t e = vrsqrteq_f32(x.val);
-    e = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x.val, e), e), e);
-    e = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x.val, e), e), e);
-    return v_float32x4(e);
-}
-#endif
-
-#define OPENCV_HAL_IMPL_NEON_ABS(_Tpuvec, _Tpsvec, usuffix, ssuffix) \
-inline _Tpuvec v_abs(const _Tpsvec& a) { return v_reinterpret_as_##usuffix(_Tpsvec(vabsq_##ssuffix(a.val))); }
-
-OPENCV_HAL_IMPL_NEON_ABS(v_uint8x16, v_int8x16, u8, s8)
-OPENCV_HAL_IMPL_NEON_ABS(v_uint16x8, v_int16x8, u16, s16)
-OPENCV_HAL_IMPL_NEON_ABS(v_uint32x4, v_int32x4, u32, s32)
-
-inline v_float32x4 v_abs(v_float32x4 x)
-{ return v_float32x4(vabsq_f32(x.val)); }
-
-#if CV_SIMD128_64F
-#define OPENCV_HAL_IMPL_NEON_DBL_BIT_OP(bin_op, intrin) \
-inline v_float64x2 operator bin_op (const v_float64x2& a, const v_float64x2& b) \
-{ \
-    return v_float64x2(vreinterpretq_f64_s64(intrin(vreinterpretq_s64_f64(a.val), vreinterpretq_s64_f64(b.val)))); \
-} \
-inline v_float64x2& operator bin_op##= (v_float64x2& a, const v_float64x2& b) \
-{ \
-    a.val = vreinterpretq_f64_s64(intrin(vreinterpretq_s64_f64(a.val), vreinterpretq_s64_f64(b.val))); \
-    return a; \
-}
-
-OPENCV_HAL_IMPL_NEON_DBL_BIT_OP(&, vandq_s64)
-OPENCV_HAL_IMPL_NEON_DBL_BIT_OP(|, vorrq_s64)
-OPENCV_HAL_IMPL_NEON_DBL_BIT_OP(^, veorq_s64)
-
-inline v_float64x2 operator ~ (const v_float64x2& a)
-{
-    return v_float64x2(vreinterpretq_f64_s32(vmvnq_s32(vreinterpretq_s32_f64(a.val))));
-}
-
-inline v_float64x2 v_sqrt(const v_float64x2& x)
-{
-    return v_float64x2(vsqrtq_f64(x.val));
-}
-
-inline v_float64x2 v_invsqrt(const v_float64x2& x)
-{
-    v_float64x2 one = v_setall_f64(1.0f);
-    return one / v_sqrt(x);
-}
-
-inline v_float64x2 v_abs(v_float64x2 x)
-{ return v_float64x2(vabsq_f64(x.val)); }
-#endif
-
-// TODO: exp, log, sin, cos
-
-#define OPENCV_HAL_IMPL_NEON_BIN_FUNC(_Tpvec, func, intrin) \
-inline _Tpvec func(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(intrin(a.val, b.val)); \
-}
-
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint8x16, v_min, vminq_u8)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint8x16, v_max, vmaxq_u8)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_int8x16, v_min, vminq_s8)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_int8x16, v_max, vmaxq_s8)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint16x8, v_min, vminq_u16)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint16x8, v_max, vmaxq_u16)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_int16x8, v_min, vminq_s16)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_int16x8, v_max, vmaxq_s16)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint32x4, v_min, vminq_u32)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint32x4, v_max, vmaxq_u32)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_int32x4, v_min, vminq_s32)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_int32x4, v_max, vmaxq_s32)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_float32x4, v_min, vminq_f32)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_float32x4, v_max, vmaxq_f32)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_float64x2, v_min, vminq_f64)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_float64x2, v_max, vmaxq_f64)
-#endif
-
-#if CV_SIMD128_64F
-inline int64x2_t vmvnq_s64(int64x2_t a)
-{
-    int64x2_t vx = vreinterpretq_s64_u32(vdupq_n_u32(0xFFFFFFFF));
-    return veorq_s64(a, vx);
-}
-inline uint64x2_t vmvnq_u64(uint64x2_t a)
-{
-    uint64x2_t vx = vreinterpretq_u64_u32(vdupq_n_u32(0xFFFFFFFF));
-    return veorq_u64(a, vx);
-}
-#endif
-#define OPENCV_HAL_IMPL_NEON_INT_CMP_OP(_Tpvec, cast, suffix, not_suffix) \
-inline _Tpvec operator == (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(cast(vceqq_##suffix(a.val, b.val))); } \
-inline _Tpvec operator != (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(cast(vmvnq_##not_suffix(vceqq_##suffix(a.val, b.val)))); } \
-inline _Tpvec operator < (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(cast(vcltq_##suffix(a.val, b.val))); } \
-inline _Tpvec operator > (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(cast(vcgtq_##suffix(a.val, b.val))); } \
-inline _Tpvec operator <= (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(cast(vcleq_##suffix(a.val, b.val))); } \
-inline _Tpvec operator >= (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(cast(vcgeq_##suffix(a.val, b.val))); }
-
-OPENCV_HAL_IMPL_NEON_INT_CMP_OP(v_uint8x16, OPENCV_HAL_NOP, u8, u8)
-OPENCV_HAL_IMPL_NEON_INT_CMP_OP(v_int8x16, vreinterpretq_s8_u8, s8, u8)
-OPENCV_HAL_IMPL_NEON_INT_CMP_OP(v_uint16x8, OPENCV_HAL_NOP, u16, u16)
-OPENCV_HAL_IMPL_NEON_INT_CMP_OP(v_int16x8, vreinterpretq_s16_u16, s16, u16)
-OPENCV_HAL_IMPL_NEON_INT_CMP_OP(v_uint32x4, OPENCV_HAL_NOP, u32, u32)
-OPENCV_HAL_IMPL_NEON_INT_CMP_OP(v_int32x4, vreinterpretq_s32_u32, s32, u32)
-OPENCV_HAL_IMPL_NEON_INT_CMP_OP(v_float32x4, vreinterpretq_f32_u32, f32, u32)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_NEON_INT_CMP_OP(v_uint64x2, OPENCV_HAL_NOP, u64, u64)
-OPENCV_HAL_IMPL_NEON_INT_CMP_OP(v_int64x2, vreinterpretq_s64_u64, s64, u64)
-OPENCV_HAL_IMPL_NEON_INT_CMP_OP(v_float64x2, vreinterpretq_f64_u64, f64, u64)
-#endif
-
-inline v_float32x4 v_not_nan(const v_float32x4& a)
-{ return v_float32x4(vreinterpretq_f32_u32(vceqq_f32(a.val, a.val))); }
-#if CV_SIMD128_64F
-inline v_float64x2 v_not_nan(const v_float64x2& a)
-{ return v_float64x2(vreinterpretq_f64_u64(vceqq_f64(a.val, a.val))); }
-#endif
-
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint8x16, v_add_wrap, vaddq_u8)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_int8x16, v_add_wrap, vaddq_s8)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint16x8, v_add_wrap, vaddq_u16)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_int16x8, v_add_wrap, vaddq_s16)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint8x16, v_sub_wrap, vsubq_u8)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_int8x16, v_sub_wrap, vsubq_s8)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint16x8, v_sub_wrap, vsubq_u16)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_int16x8, v_sub_wrap, vsubq_s16)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint8x16, v_mul_wrap, vmulq_u8)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_int8x16, v_mul_wrap, vmulq_s8)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint16x8, v_mul_wrap, vmulq_u16)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_int16x8, v_mul_wrap, vmulq_s16)
-
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint8x16, v_absdiff, vabdq_u8)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint16x8, v_absdiff, vabdq_u16)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint32x4, v_absdiff, vabdq_u32)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_float32x4, v_absdiff, vabdq_f32)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_float64x2, v_absdiff, vabdq_f64)
-#endif
-
-/** Saturating absolute difference **/
-inline v_int8x16 v_absdiffs(const v_int8x16& a, const v_int8x16& b)
-{ return v_int8x16(vqabsq_s8(vqsubq_s8(a.val, b.val))); }
-inline v_int16x8 v_absdiffs(const v_int16x8& a, const v_int16x8& b)
-{ return v_int16x8(vqabsq_s16(vqsubq_s16(a.val, b.val))); }
-
-#define OPENCV_HAL_IMPL_NEON_BIN_FUNC2(_Tpvec, _Tpvec2, cast, func, intrin) \
-inline _Tpvec2 func(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec2(cast(intrin(a.val, b.val))); \
-}
-
-OPENCV_HAL_IMPL_NEON_BIN_FUNC2(v_int8x16, v_uint8x16, vreinterpretq_u8_s8, v_absdiff, vabdq_s8)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC2(v_int16x8, v_uint16x8, vreinterpretq_u16_s16, v_absdiff, vabdq_s16)
-OPENCV_HAL_IMPL_NEON_BIN_FUNC2(v_int32x4, v_uint32x4, vreinterpretq_u32_s32, v_absdiff, vabdq_s32)
-
-inline v_float32x4 v_magnitude(const v_float32x4& a, const v_float32x4& b)
-{
-    v_float32x4 x(vmlaq_f32(vmulq_f32(a.val, a.val), b.val, b.val));
-    return v_sqrt(x);
-}
-
-inline v_float32x4 v_sqr_magnitude(const v_float32x4& a, const v_float32x4& b)
-{
-    return v_float32x4(vmlaq_f32(vmulq_f32(a.val, a.val), b.val, b.val));
-}
-
-inline v_float32x4 v_fma(const v_float32x4& a, const v_float32x4& b, const v_float32x4& c)
-{
-#if CV_SIMD128_64F
-    // ARMv8, which adds support for 64-bit floating-point (so CV_SIMD128_64F is defined),
-    // also adds FMA support both for single- and double-precision floating-point vectors
-    return v_float32x4(vfmaq_f32(c.val, a.val, b.val));
-#else
-    return v_float32x4(vmlaq_f32(c.val, a.val, b.val));
-#endif
-}
-
-inline v_int32x4 v_fma(const v_int32x4& a, const v_int32x4& b, const v_int32x4& c)
-{
-    return v_int32x4(vmlaq_s32(c.val, a.val, b.val));
-}
-
-inline v_float32x4 v_muladd(const v_float32x4& a, const v_float32x4& b, const v_float32x4& c)
-{
-    return v_fma(a, b, c);
-}
-
-inline v_int32x4 v_muladd(const v_int32x4& a, const v_int32x4& b, const v_int32x4& c)
-{
-    return v_fma(a, b, c);
-}
-
-#if CV_SIMD128_64F
-inline v_float64x2 v_magnitude(const v_float64x2& a, const v_float64x2& b)
-{
-    v_float64x2 x(vaddq_f64(vmulq_f64(a.val, a.val), vmulq_f64(b.val, b.val)));
-    return v_sqrt(x);
-}
-
-inline v_float64x2 v_sqr_magnitude(const v_float64x2& a, const v_float64x2& b)
-{
-    return v_float64x2(vaddq_f64(vmulq_f64(a.val, a.val), vmulq_f64(b.val, b.val)));
-}
-
-inline v_float64x2 v_fma(const v_float64x2& a, const v_float64x2& b, const v_float64x2& c)
-{
-    return v_float64x2(vfmaq_f64(c.val, a.val, b.val));
-}
-
-inline v_float64x2 v_muladd(const v_float64x2& a, const v_float64x2& b, const v_float64x2& c)
-{
-    return v_fma(a, b, c);
-}
-#endif
-
-// trade efficiency for convenience
-#define OPENCV_HAL_IMPL_NEON_SHIFT_OP(_Tpvec, suffix, _Tps, ssuffix) \
-inline _Tpvec operator << (const _Tpvec& a, int n) \
-{ return _Tpvec(vshlq_##suffix(a.val, vdupq_n_##ssuffix((_Tps)n))); } \
-inline _Tpvec operator >> (const _Tpvec& a, int n) \
-{ return _Tpvec(vshlq_##suffix(a.val, vdupq_n_##ssuffix((_Tps)-n))); } \
-template<int n> inline _Tpvec v_shl(const _Tpvec& a) \
-{ return _Tpvec(vshlq_n_##suffix(a.val, n)); } \
-template<int n> inline _Tpvec v_shr(const _Tpvec& a) \
-{ return _Tpvec(vshrq_n_##suffix(a.val, n)); } \
-template<int n> inline _Tpvec v_rshr(const _Tpvec& a) \
-{ return _Tpvec(vrshrq_n_##suffix(a.val, n)); }
-
-OPENCV_HAL_IMPL_NEON_SHIFT_OP(v_uint8x16, u8, schar, s8)
-OPENCV_HAL_IMPL_NEON_SHIFT_OP(v_int8x16, s8, schar, s8)
-OPENCV_HAL_IMPL_NEON_SHIFT_OP(v_uint16x8, u16, short, s16)
-OPENCV_HAL_IMPL_NEON_SHIFT_OP(v_int16x8, s16, short, s16)
-OPENCV_HAL_IMPL_NEON_SHIFT_OP(v_uint32x4, u32, int, s32)
-OPENCV_HAL_IMPL_NEON_SHIFT_OP(v_int32x4, s32, int, s32)
-OPENCV_HAL_IMPL_NEON_SHIFT_OP(v_uint64x2, u64, int64, s64)
-OPENCV_HAL_IMPL_NEON_SHIFT_OP(v_int64x2, s64, int64, s64)
-
-#define OPENCV_HAL_IMPL_NEON_ROTATE_OP(_Tpvec, suffix) \
-template<int n> inline _Tpvec v_rotate_right(const _Tpvec& a) \
-{ return _Tpvec(vextq_##suffix(a.val, vdupq_n_##suffix(0), n)); } \
-template<int n> inline _Tpvec v_rotate_left(const _Tpvec& a) \
-{ return _Tpvec(vextq_##suffix(vdupq_n_##suffix(0), a.val, _Tpvec::nlanes - n)); } \
-template<> inline _Tpvec v_rotate_left<0>(const _Tpvec& a) \
-{ return a; } \
-template<int n> inline _Tpvec v_rotate_right(const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(vextq_##suffix(a.val, b.val, n)); } \
-template<int n> inline _Tpvec v_rotate_left(const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(vextq_##suffix(b.val, a.val, _Tpvec::nlanes - n)); } \
-template<> inline _Tpvec v_rotate_left<0>(const _Tpvec& a, const _Tpvec& b) \
-{ CV_UNUSED(b); return a; }
-
-OPENCV_HAL_IMPL_NEON_ROTATE_OP(v_uint8x16, u8)
-OPENCV_HAL_IMPL_NEON_ROTATE_OP(v_int8x16, s8)
-OPENCV_HAL_IMPL_NEON_ROTATE_OP(v_uint16x8, u16)
-OPENCV_HAL_IMPL_NEON_ROTATE_OP(v_int16x8, s16)
-OPENCV_HAL_IMPL_NEON_ROTATE_OP(v_uint32x4, u32)
-OPENCV_HAL_IMPL_NEON_ROTATE_OP(v_int32x4, s32)
-OPENCV_HAL_IMPL_NEON_ROTATE_OP(v_float32x4, f32)
-OPENCV_HAL_IMPL_NEON_ROTATE_OP(v_uint64x2, u64)
-OPENCV_HAL_IMPL_NEON_ROTATE_OP(v_int64x2, s64)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_NEON_ROTATE_OP(v_float64x2, f64)
-#endif
-
-#if defined(__clang__) && defined(__aarch64__)
-// avoid LD2 instruction. details: https://github.com/opencv/opencv/issues/14863
-#define OPENCV_HAL_IMPL_NEON_LOAD_LOW_OP(_Tpvec, _Tp, suffix) \
-inline _Tpvec v_load_low(const _Tp* ptr) \
-{ \
-typedef uint64 CV_DECL_ALIGNED(1) unaligned_uint64; \
-uint64 v = *(unaligned_uint64*)ptr; \
-return _Tpvec(v_reinterpret_as_##suffix(v_uint64x2(v, (uint64)123456))); \
-}
-#else
-#define OPENCV_HAL_IMPL_NEON_LOAD_LOW_OP(_Tpvec, _Tp, suffix) \
-inline _Tpvec v_load_low(const _Tp* ptr) \
-{ return _Tpvec(vcombine_##suffix(vld1_##suffix(ptr), vdup_n_##suffix((_Tp)0))); }
-#endif
-
-#define OPENCV_HAL_IMPL_NEON_LOADSTORE_OP(_Tpvec, _Tp, suffix) \
-inline _Tpvec v_load(const _Tp* ptr) \
-{ return _Tpvec(vld1q_##suffix(ptr)); } \
-inline _Tpvec v_load_aligned(const _Tp* ptr) \
-{ return _Tpvec(vld1q_##suffix(ptr)); } \
-OPENCV_HAL_IMPL_NEON_LOAD_LOW_OP(_Tpvec, _Tp, suffix) \
-inline _Tpvec v_load_halves(const _Tp* ptr0, const _Tp* ptr1) \
-{ return _Tpvec(vcombine_##suffix(vld1_##suffix(ptr0), vld1_##suffix(ptr1))); } \
-inline void v_store(_Tp* ptr, const _Tpvec& a) \
-{ vst1q_##suffix(ptr, a.val); } \
-inline void v_store_aligned(_Tp* ptr, const _Tpvec& a) \
-{ vst1q_##suffix(ptr, a.val); } \
-inline void v_store_aligned_nocache(_Tp* ptr, const _Tpvec& a) \
-{ vst1q_##suffix(ptr, a.val); } \
-inline void v_store(_Tp* ptr, const _Tpvec& a, hal::StoreMode /*mode*/) \
-{ vst1q_##suffix(ptr, a.val); } \
-inline void v_store_low(_Tp* ptr, const _Tpvec& a) \
-{ vst1_##suffix(ptr, vget_low_##suffix(a.val)); } \
-inline void v_store_high(_Tp* ptr, const _Tpvec& a) \
-{ vst1_##suffix(ptr, vget_high_##suffix(a.val)); }
-
-OPENCV_HAL_IMPL_NEON_LOADSTORE_OP(v_uint8x16, uchar, u8)
-OPENCV_HAL_IMPL_NEON_LOADSTORE_OP(v_int8x16, schar, s8)
-OPENCV_HAL_IMPL_NEON_LOADSTORE_OP(v_uint16x8, ushort, u16)
-OPENCV_HAL_IMPL_NEON_LOADSTORE_OP(v_int16x8, short, s16)
-OPENCV_HAL_IMPL_NEON_LOADSTORE_OP(v_uint32x4, unsigned, u32)
-OPENCV_HAL_IMPL_NEON_LOADSTORE_OP(v_int32x4, int, s32)
-OPENCV_HAL_IMPL_NEON_LOADSTORE_OP(v_uint64x2, uint64, u64)
-OPENCV_HAL_IMPL_NEON_LOADSTORE_OP(v_int64x2, int64, s64)
-OPENCV_HAL_IMPL_NEON_LOADSTORE_OP(v_float32x4, float, f32)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_NEON_LOADSTORE_OP(v_float64x2, double, f64)
-#endif
-
-inline unsigned v_reduce_sum(const v_uint8x16& a)
-{
-#if CV_NEON_AARCH64
-    uint16_t t0 = vaddlvq_u8(a.val);
-    return t0;
-#else // #if CV_NEON_AARCH64
-    uint32x4_t t0 = vpaddlq_u16(vpaddlq_u8(a.val));
-    uint32x2_t t1 = vpadd_u32(vget_low_u32(t0), vget_high_u32(t0));
-    return vget_lane_u32(vpadd_u32(t1, t1), 0);
-#endif // #if CV_NEON_AARCH64
-}
-inline int v_reduce_sum(const v_int8x16& a)
-{
-#if CV_NEON_AARCH64
-    int16_t t0 = vaddlvq_s8(a.val);
-    return t0;
-#else // #if CV_NEON_AARCH64
-    int32x4_t t0 = vpaddlq_s16(vpaddlq_s8(a.val));
-    int32x2_t t1 = vpadd_s32(vget_low_s32(t0), vget_high_s32(t0));
-    return vget_lane_s32(vpadd_s32(t1, t1), 0);
-#endif // #if CV_NEON_AARCH64
-}
-inline unsigned v_reduce_sum(const v_uint16x8& a)
-{
-#if CV_NEON_AARCH64
-    uint32_t t0 = vaddlvq_u16(a.val);
-    return t0;
-#else // #if CV_NEON_AARCH64
-    uint32x4_t t0 = vpaddlq_u16(a.val);
-    uint32x2_t t1 = vpadd_u32(vget_low_u32(t0), vget_high_u32(t0));
-    return vget_lane_u32(vpadd_u32(t1, t1), 0);
-#endif // #if CV_NEON_AARCH64
-}
-inline int v_reduce_sum(const v_int16x8& a)
-{
-#if CV_NEON_AARCH64
-    int32_t t0 = vaddlvq_s16(a.val);
-    return t0;
-#else // #if CV_NEON_AARCH64
-    int32x4_t t0 = vpaddlq_s16(a.val);
-    int32x2_t t1 = vpadd_s32(vget_low_s32(t0), vget_high_s32(t0));
-    return vget_lane_s32(vpadd_s32(t1, t1), 0);
-#endif // #if CV_NEON_AARCH64
-}
-
-#if CV_NEON_AARCH64
-#define OPENCV_HAL_IMPL_NEON_REDUCE_OP_16(_Tpvec, _Tpnvec, scalartype, func, vectorfunc, suffix) \
-inline scalartype v_reduce_##func(const _Tpvec& a) \
-{ \
-    return v##vectorfunc##vq_##suffix(a.val); \
-}
-#else // #if CV_NEON_AARCH64
-#define OPENCV_HAL_IMPL_NEON_REDUCE_OP_16(_Tpvec, _Tpnvec, scalartype, func, vectorfunc, suffix) \
-inline scalartype v_reduce_##func(const _Tpvec& a) \
-{ \
-    _Tpnvec##_t a0 = vp##vectorfunc##_##suffix(vget_low_##suffix(a.val), vget_high_##suffix(a.val)); \
-    a0 = vp##vectorfunc##_##suffix(a0, a0); \
-    a0 = vp##vectorfunc##_##suffix(a0, a0); \
-    return (scalartype)vget_lane_##suffix(vp##vectorfunc##_##suffix(a0, a0),0); \
-}
-#endif // #if CV_NEON_AARCH64
-
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_16(v_uint8x16, uint8x8, uchar, max, max, u8)
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_16(v_uint8x16, uint8x8, uchar, min, min, u8)
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_16(v_int8x16, int8x8, schar, max, max, s8)
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_16(v_int8x16, int8x8, schar, min, min, s8)
-
-#if CV_NEON_AARCH64
-#define OPENCV_HAL_IMPL_NEON_REDUCE_OP_8(_Tpvec, _Tpnvec, scalartype, func, vectorfunc, suffix) \
-inline scalartype v_reduce_##func(const _Tpvec& a) \
-{ \
-    return v##vectorfunc##vq_##suffix(a.val); \
-}
-#else // #if CV_NEON_AARCH64
-#define OPENCV_HAL_IMPL_NEON_REDUCE_OP_8(_Tpvec, _Tpnvec, scalartype, func, vectorfunc, suffix) \
-inline scalartype v_reduce_##func(const _Tpvec& a) \
-{ \
-    _Tpnvec##_t a0 = vp##vectorfunc##_##suffix(vget_low_##suffix(a.val), vget_high_##suffix(a.val)); \
-    a0 = vp##vectorfunc##_##suffix(a0, a0); \
-    return (scalartype)vget_lane_##suffix(vp##vectorfunc##_##suffix(a0, a0),0); \
-}
-#endif // #if CV_NEON_AARCH64
-
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_8(v_uint16x8, uint16x4, ushort, max, max, u16)
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_8(v_uint16x8, uint16x4, ushort, min, min, u16)
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_8(v_int16x8, int16x4, short, max, max, s16)
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_8(v_int16x8, int16x4, short, min, min, s16)
-
-#if CV_NEON_AARCH64
-#define OPENCV_HAL_IMPL_NEON_REDUCE_OP_4(_Tpvec, _Tpnvec, scalartype, func, vectorfunc, suffix) \
-inline scalartype v_reduce_##func(const _Tpvec& a) \
-{ \
-    return v##vectorfunc##vq_##suffix(a.val); \
-}
-#else // #if CV_NEON_AARCH64
-#define OPENCV_HAL_IMPL_NEON_REDUCE_OP_4(_Tpvec, _Tpnvec, scalartype, func, vectorfunc, suffix) \
-inline scalartype v_reduce_##func(const _Tpvec& a) \
-{ \
-    _Tpnvec##_t a0 = vp##vectorfunc##_##suffix(vget_low_##suffix(a.val), vget_high_##suffix(a.val)); \
-    return (scalartype)vget_lane_##suffix(vp##vectorfunc##_##suffix(a0, vget_high_##suffix(a.val)),0); \
-}
-#endif // #if CV_NEON_AARCH64
-
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_4(v_uint32x4, uint32x2, unsigned, sum, add, u32)
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_4(v_uint32x4, uint32x2, unsigned, max, max, u32)
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_4(v_uint32x4, uint32x2, unsigned, min, min, u32)
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_4(v_int32x4, int32x2, int, sum, add, s32)
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_4(v_int32x4, int32x2, int, max, max, s32)
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_4(v_int32x4, int32x2, int, min, min, s32)
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_4(v_float32x4, float32x2, float, sum, add, f32)
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_4(v_float32x4, float32x2, float, max, max, f32)
-OPENCV_HAL_IMPL_NEON_REDUCE_OP_4(v_float32x4, float32x2, float, min, min, f32)
-
-inline uint64 v_reduce_sum(const v_uint64x2& a)
-{
-#if CV_NEON_AARCH64
-    return vaddvq_u64(a.val);
-#else // #if CV_NEON_AARCH64
-    return vget_lane_u64(vadd_u64(vget_low_u64(a.val), vget_high_u64(a.val)),0);
-#endif // #if CV_NEON_AARCH64
-}
-inline int64 v_reduce_sum(const v_int64x2& a)
-{
-#if CV_NEON_AARCH64
-    return vaddvq_s64(a.val);
-#else // #if CV_NEON_AARCH64
-    return vget_lane_s64(vadd_s64(vget_low_s64(a.val), vget_high_s64(a.val)),0);
-#endif // #if CV_NEON_AARCH64
-}
-#if CV_SIMD128_64F
-inline double v_reduce_sum(const v_float64x2& a)
-{
-    return vaddvq_f64(a.val);
-}
-#endif
-
-inline v_float32x4 v_reduce_sum4(const v_float32x4& a, const v_float32x4& b,
-                                 const v_float32x4& c, const v_float32x4& d)
-{
-#if CV_NEON_AARCH64
-    float32x4_t ab = vpaddq_f32(a.val, b.val); // a0+a1 a2+a3 b0+b1 b2+b3
-    float32x4_t cd = vpaddq_f32(c.val, d.val); // c0+c1 d0+d1 c2+c3 d2+d3
-    return v_float32x4(vpaddq_f32(ab, cd));  // sumA sumB sumC sumD
-#else // #if CV_NEON_AARCH64
-    float32x4x2_t ab = vtrnq_f32(a.val, b.val);
-    float32x4x2_t cd = vtrnq_f32(c.val, d.val);
-
-    float32x4_t u0 = vaddq_f32(ab.val[0], ab.val[1]); // a0+a1 b0+b1 a2+a3 b2+b3
-    float32x4_t u1 = vaddq_f32(cd.val[0], cd.val[1]); // c0+c1 d0+d1 c2+c3 d2+d3
-
-    float32x4_t v0 = vcombine_f32(vget_low_f32(u0), vget_low_f32(u1));
-    float32x4_t v1 = vcombine_f32(vget_high_f32(u0), vget_high_f32(u1));
-
-    return v_float32x4(vaddq_f32(v0, v1));
-#endif // #if CV_NEON_AARCH64
-}
-
-inline unsigned v_reduce_sad(const v_uint8x16& a, const v_uint8x16& b)
-{
-#if CV_NEON_AARCH64
-    uint8x16_t t0 = vabdq_u8(a.val, b.val);
-    uint16_t t1 = vaddlvq_u8(t0);
-    return t1;
-#else // #if CV_NEON_AARCH64
-    uint32x4_t t0 = vpaddlq_u16(vpaddlq_u8(vabdq_u8(a.val, b.val)));
-    uint32x2_t t1 = vpadd_u32(vget_low_u32(t0), vget_high_u32(t0));
-    return vget_lane_u32(vpadd_u32(t1, t1), 0);
-#endif // #if CV_NEON_AARCH64
-}
-inline unsigned v_reduce_sad(const v_int8x16& a, const v_int8x16& b)
-{
-#if CV_NEON_AARCH64
-    uint8x16_t t0 = vreinterpretq_u8_s8(vabdq_s8(a.val, b.val));
-    uint16_t t1 = vaddlvq_u8(t0);
-    return t1;
-#else // #if CV_NEON_AARCH64
-    uint32x4_t t0 = vpaddlq_u16(vpaddlq_u8(vreinterpretq_u8_s8(vabdq_s8(a.val, b.val))));
-    uint32x2_t t1 = vpadd_u32(vget_low_u32(t0), vget_high_u32(t0));
-    return vget_lane_u32(vpadd_u32(t1, t1), 0);
-#endif // #if CV_NEON_AARCH64
-}
-inline unsigned v_reduce_sad(const v_uint16x8& a, const v_uint16x8& b)
-{
-#if CV_NEON_AARCH64
-    uint16x8_t t0 = vabdq_u16(a.val, b.val);
-    uint32_t t1 = vaddlvq_u16(t0);
-    return t1;
-#else // #if CV_NEON_AARCH64
-    uint32x4_t t0 = vpaddlq_u16(vabdq_u16(a.val, b.val));
-    uint32x2_t t1 = vpadd_u32(vget_low_u32(t0), vget_high_u32(t0));
-    return vget_lane_u32(vpadd_u32(t1, t1), 0);
-#endif // #if CV_NEON_AARCH64
-}
-inline unsigned v_reduce_sad(const v_int16x8& a, const v_int16x8& b)
-{
-#if CV_NEON_AARCH64
-    uint16x8_t t0 = vreinterpretq_u16_s16(vabdq_s16(a.val, b.val));
-    uint32_t t1 = vaddlvq_u16(t0);
-    return t1;
-#else // #if CV_NEON_AARCH64
-    uint32x4_t t0 = vpaddlq_u16(vreinterpretq_u16_s16(vabdq_s16(a.val, b.val)));
-    uint32x2_t t1 = vpadd_u32(vget_low_u32(t0), vget_high_u32(t0));
-    return vget_lane_u32(vpadd_u32(t1, t1), 0);
-#endif // #if CV_NEON_AARCH64
-}
-inline unsigned v_reduce_sad(const v_uint32x4& a, const v_uint32x4& b)
-{
-#if CV_NEON_AARCH64
-    uint32x4_t t0 = vabdq_u32(a.val, b.val);
-    uint32_t t1 = vaddvq_u32(t0);
-    return t1;
-#else // #if CV_NEON_AARCH64
-    uint32x4_t t0 = vabdq_u32(a.val, b.val);
-    uint32x2_t t1 = vpadd_u32(vget_low_u32(t0), vget_high_u32(t0));
-    return vget_lane_u32(vpadd_u32(t1, t1), 0);
-#endif // #if CV_NEON_AARCH64
-}
-inline unsigned v_reduce_sad(const v_int32x4& a, const v_int32x4& b)
-{
-#if CV_NEON_AARCH64
-    uint32x4_t t0 = vreinterpretq_u32_s32(vabdq_s32(a.val, b.val));
-    uint32_t t1 = vaddvq_u32(t0);
-    return t1;
-#else // #if CV_NEON_AARCH64
-    uint32x4_t t0 = vreinterpretq_u32_s32(vabdq_s32(a.val, b.val));
-    uint32x2_t t1 = vpadd_u32(vget_low_u32(t0), vget_high_u32(t0));
-    return vget_lane_u32(vpadd_u32(t1, t1), 0);
-#endif // #if CV_NEON_AARCH64
-}
-inline float v_reduce_sad(const v_float32x4& a, const v_float32x4& b)
-{
-#if CV_NEON_AARCH64
-    float32x4_t t0 = vabdq_f32(a.val, b.val);
-    return vaddvq_f32(t0);
-#else // #if CV_NEON_AARCH64
-    float32x4_t t0 = vabdq_f32(a.val, b.val);
-    float32x2_t t1 = vpadd_f32(vget_low_f32(t0), vget_high_f32(t0));
-    return vget_lane_f32(vpadd_f32(t1, t1), 0);
-#endif // #if CV_NEON_AARCH64
-}
-
-inline v_uint8x16 v_popcount(const v_uint8x16& a)
-{ return v_uint8x16(vcntq_u8(a.val)); }
-inline v_uint8x16 v_popcount(const v_int8x16& a)
-{ return v_uint8x16(vcntq_u8(vreinterpretq_u8_s8(a.val))); }
-inline v_uint16x8 v_popcount(const v_uint16x8& a)
-{ return v_uint16x8(vpaddlq_u8(vcntq_u8(vreinterpretq_u8_u16(a.val)))); }
-inline v_uint16x8 v_popcount(const v_int16x8& a)
-{ return v_uint16x8(vpaddlq_u8(vcntq_u8(vreinterpretq_u8_s16(a.val)))); }
-inline v_uint32x4 v_popcount(const v_uint32x4& a)
-{ return v_uint32x4(vpaddlq_u16(vpaddlq_u8(vcntq_u8(vreinterpretq_u8_u32(a.val))))); }
-inline v_uint32x4 v_popcount(const v_int32x4& a)
-{ return v_uint32x4(vpaddlq_u16(vpaddlq_u8(vcntq_u8(vreinterpretq_u8_s32(a.val))))); }
-inline v_uint64x2 v_popcount(const v_uint64x2& a)
-{ return v_uint64x2(vpaddlq_u32(vpaddlq_u16(vpaddlq_u8(vcntq_u8(vreinterpretq_u8_u64(a.val)))))); }
-inline v_uint64x2 v_popcount(const v_int64x2& a)
-{ return v_uint64x2(vpaddlq_u32(vpaddlq_u16(vpaddlq_u8(vcntq_u8(vreinterpretq_u8_s64(a.val)))))); }
-
-inline int v_signmask(const v_uint8x16& a)
-{
-#if CV_NEON_AARCH64
-    const int8x16_t signPosition = {0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7};
-    const uint8x16_t byteOrder = {0,8,1,9,2,10,3,11,4,12,5,13,6,14,7,15};
-    uint8x16_t v0 = vshlq_u8(vshrq_n_u8(a.val, 7), signPosition);
-    uint8x16_t v1 = vqtbl1q_u8(v0, byteOrder);
-    uint32_t t0 = vaddlvq_u16(vreinterpretq_u16_u8(v1));
-    return t0;
-#else // #if CV_NEON_AARCH64
-    int8x8_t m0 = vcreate_s8(CV_BIG_UINT(0x0706050403020100));
-    uint8x16_t v0 = vshlq_u8(vshrq_n_u8(a.val, 7), vcombine_s8(m0, m0));
-    uint64x2_t v1 = vpaddlq_u32(vpaddlq_u16(vpaddlq_u8(v0)));
-    return (int)vgetq_lane_u64(v1, 0) + ((int)vgetq_lane_u64(v1, 1) << 8);
-#endif // #if CV_NEON_AARCH64
-}
-
-inline int v_signmask(const v_int8x16& a)
-{ return v_signmask(v_reinterpret_as_u8(a)); }
-
-inline int v_signmask(const v_uint16x8& a)
-{
-#if CV_NEON_AARCH64
-    const int16x8_t signPosition = {0,1,2,3,4,5,6,7};
-    uint16x8_t v0 = vshlq_u16(vshrq_n_u16(a.val, 15), signPosition);
-    uint32_t t0 = vaddlvq_u16(v0);
-    return t0;
-#else // #if CV_NEON_AARCH64
-    int16x4_t m0 = vcreate_s16(CV_BIG_UINT(0x0003000200010000));
-    uint16x8_t v0 = vshlq_u16(vshrq_n_u16(a.val, 15), vcombine_s16(m0, m0));
-    uint64x2_t v1 = vpaddlq_u32(vpaddlq_u16(v0));
-    return (int)vgetq_lane_u64(v1, 0) + ((int)vgetq_lane_u64(v1, 1) << 4);
-#endif // #if CV_NEON_AARCH64
-}
-inline int v_signmask(const v_int16x8& a)
-{ return v_signmask(v_reinterpret_as_u16(a)); }
-
-inline int v_signmask(const v_uint32x4& a)
-{
-#if CV_NEON_AARCH64
-    const int32x4_t signPosition = {0,1,2,3};
-    uint32x4_t v0 = vshlq_u32(vshrq_n_u32(a.val, 31), signPosition);
-    uint32_t t0 = vaddvq_u32(v0);
-    return t0;
-#else // #if CV_NEON_AARCH64
-    int32x2_t m0 = vcreate_s32(CV_BIG_UINT(0x0000000100000000));
-    uint32x4_t v0 = vshlq_u32(vshrq_n_u32(a.val, 31), vcombine_s32(m0, m0));
-    uint64x2_t v1 = vpaddlq_u32(v0);
-    return (int)vgetq_lane_u64(v1, 0) + ((int)vgetq_lane_u64(v1, 1) << 2);
-#endif // #if CV_NEON_AARCH64
-}
-inline int v_signmask(const v_int32x4& a)
-{ return v_signmask(v_reinterpret_as_u32(a)); }
-inline int v_signmask(const v_float32x4& a)
-{ return v_signmask(v_reinterpret_as_u32(a)); }
-inline int v_signmask(const v_uint64x2& a)
-{
-#if CV_NEON_AARCH64
-    const int64x2_t signPosition = {0,1};
-    uint64x2_t v0 = vshlq_u64(vshrq_n_u64(a.val, 63), signPosition);
-    uint64_t t0 = vaddvq_u64(v0);
-    return t0;
-#else // #if CV_NEON_AARCH64
-    int64x1_t m0 = vdup_n_s64(0);
-    uint64x2_t v0 = vshlq_u64(vshrq_n_u64(a.val, 63), vcombine_s64(m0, m0));
-    return (int)vgetq_lane_u64(v0, 0) + ((int)vgetq_lane_u64(v0, 1) << 1);
-#endif // #if CV_NEON_AARCH64
-}
-inline int v_signmask(const v_int64x2& a)
-{ return v_signmask(v_reinterpret_as_u64(a)); }
-#if CV_SIMD128_64F
-inline int v_signmask(const v_float64x2& a)
-{ return v_signmask(v_reinterpret_as_u64(a)); }
-#endif
-
-inline int v_scan_forward(const v_int8x16& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_uint8x16& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_int16x8& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_uint16x8& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_int32x4& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_uint32x4& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_float32x4& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_int64x2& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_uint64x2& a) { return trailingZeros32(v_signmask(a)); }
-#if CV_SIMD128_64F
-inline int v_scan_forward(const v_float64x2& a) { return trailingZeros32(v_signmask(a)); }
-#endif
-
-#if CV_NEON_AARCH64
-    #define OPENCV_HAL_IMPL_NEON_CHECK_ALLANY(_Tpvec, suffix, shift) \
-    inline bool v_check_all(const v_##_Tpvec& a) \
-    { \
-        return (vminvq_##suffix(a.val) >> shift) != 0; \
-    } \
-    inline bool v_check_any(const v_##_Tpvec& a) \
-    { \
-        return (vmaxvq_##suffix(a.val) >> shift) != 0; \
-    }
-#else // #if CV_NEON_AARCH64
-    #define OPENCV_HAL_IMPL_NEON_CHECK_ALLANY(_Tpvec, suffix, shift) \
-    inline bool v_check_all(const v_##_Tpvec& a) \
-    { \
-        _Tpvec##_t v0 = vshrq_n_##suffix(vmvnq_##suffix(a.val), shift); \
-        uint64x2_t v1 = vreinterpretq_u64_##suffix(v0); \
-        return (vgetq_lane_u64(v1, 0) | vgetq_lane_u64(v1, 1)) == 0; \
-    } \
-    inline bool v_check_any(const v_##_Tpvec& a) \
-    { \
-        _Tpvec##_t v0 = vshrq_n_##suffix(a.val, shift); \
-        uint64x2_t v1 = vreinterpretq_u64_##suffix(v0); \
-        return (vgetq_lane_u64(v1, 0) | vgetq_lane_u64(v1, 1)) != 0; \
-    }
-#endif // #if CV_NEON_AARCH64
-
-OPENCV_HAL_IMPL_NEON_CHECK_ALLANY(uint8x16, u8, 7)
-OPENCV_HAL_IMPL_NEON_CHECK_ALLANY(uint16x8, u16, 15)
-OPENCV_HAL_IMPL_NEON_CHECK_ALLANY(uint32x4, u32, 31)
-
-inline bool v_check_all(const v_uint64x2& a)
-{
-    uint64x2_t v0 = vshrq_n_u64(a.val, 63);
-    return (vgetq_lane_u64(v0, 0) & vgetq_lane_u64(v0, 1)) == 1;
-}
-inline bool v_check_any(const v_uint64x2& a)
-{
-    uint64x2_t v0 = vshrq_n_u64(a.val, 63);
-    return (vgetq_lane_u64(v0, 0) | vgetq_lane_u64(v0, 1)) != 0;
-}
-
-inline bool v_check_all(const v_int8x16& a)
-{ return v_check_all(v_reinterpret_as_u8(a)); }
-inline bool v_check_all(const v_int16x8& a)
-{ return v_check_all(v_reinterpret_as_u16(a)); }
-inline bool v_check_all(const v_int32x4& a)
-{ return v_check_all(v_reinterpret_as_u32(a)); }
-inline bool v_check_all(const v_float32x4& a)
-{ return v_check_all(v_reinterpret_as_u32(a)); }
-
-inline bool v_check_any(const v_int8x16& a)
-{ return v_check_any(v_reinterpret_as_u8(a)); }
-inline bool v_check_any(const v_int16x8& a)
-{ return v_check_any(v_reinterpret_as_u16(a)); }
-inline bool v_check_any(const v_int32x4& a)
-{ return v_check_any(v_reinterpret_as_u32(a)); }
-inline bool v_check_any(const v_float32x4& a)
-{ return v_check_any(v_reinterpret_as_u32(a)); }
-
-inline bool v_check_all(const v_int64x2& a)
-{ return v_check_all(v_reinterpret_as_u64(a)); }
-inline bool v_check_any(const v_int64x2& a)
-{ return v_check_any(v_reinterpret_as_u64(a)); }
-#if CV_SIMD128_64F
-inline bool v_check_all(const v_float64x2& a)
-{ return v_check_all(v_reinterpret_as_u64(a)); }
-inline bool v_check_any(const v_float64x2& a)
-{ return v_check_any(v_reinterpret_as_u64(a)); }
-#endif
-
-#define OPENCV_HAL_IMPL_NEON_SELECT(_Tpvec, suffix, usuffix) \
-inline _Tpvec v_select(const _Tpvec& mask, const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(vbslq_##suffix(vreinterpretq_##usuffix##_##suffix(mask.val), a.val, b.val)); \
-}
-
-OPENCV_HAL_IMPL_NEON_SELECT(v_uint8x16, u8, u8)
-OPENCV_HAL_IMPL_NEON_SELECT(v_int8x16, s8, u8)
-OPENCV_HAL_IMPL_NEON_SELECT(v_uint16x8, u16, u16)
-OPENCV_HAL_IMPL_NEON_SELECT(v_int16x8, s16, u16)
-OPENCV_HAL_IMPL_NEON_SELECT(v_uint32x4, u32, u32)
-OPENCV_HAL_IMPL_NEON_SELECT(v_int32x4, s32, u32)
-OPENCV_HAL_IMPL_NEON_SELECT(v_float32x4, f32, u32)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_NEON_SELECT(v_float64x2, f64, u64)
-#endif
-
-#if CV_NEON_AARCH64
-#define OPENCV_HAL_IMPL_NEON_EXPAND(_Tpvec, _Tpwvec, _Tp, suffix) \
-inline void v_expand(const _Tpvec& a, _Tpwvec& b0, _Tpwvec& b1) \
-{ \
-    b0.val = vmovl_##suffix(vget_low_##suffix(a.val)); \
-    b1.val = vmovl_high_##suffix(a.val); \
-} \
-inline _Tpwvec v_expand_low(const _Tpvec& a) \
-{ \
-    return _Tpwvec(vmovl_##suffix(vget_low_##suffix(a.val))); \
-} \
-inline _Tpwvec v_expand_high(const _Tpvec& a) \
-{ \
-    return _Tpwvec(vmovl_high_##suffix(a.val)); \
-} \
-inline _Tpwvec v_load_expand(const _Tp* ptr) \
-{ \
-    return _Tpwvec(vmovl_##suffix(vld1_##suffix(ptr))); \
-}
-#else
-#define OPENCV_HAL_IMPL_NEON_EXPAND(_Tpvec, _Tpwvec, _Tp, suffix) \
-inline void v_expand(const _Tpvec& a, _Tpwvec& b0, _Tpwvec& b1) \
-{ \
-    b0.val = vmovl_##suffix(vget_low_##suffix(a.val)); \
-    b1.val = vmovl_##suffix(vget_high_##suffix(a.val)); \
-} \
-inline _Tpwvec v_expand_low(const _Tpvec& a) \
-{ \
-    return _Tpwvec(vmovl_##suffix(vget_low_##suffix(a.val))); \
-} \
-inline _Tpwvec v_expand_high(const _Tpvec& a) \
-{ \
-    return _Tpwvec(vmovl_##suffix(vget_high_##suffix(a.val))); \
-} \
-inline _Tpwvec v_load_expand(const _Tp* ptr) \
-{ \
-    return _Tpwvec(vmovl_##suffix(vld1_##suffix(ptr))); \
-}
-#endif
-
-OPENCV_HAL_IMPL_NEON_EXPAND(v_uint8x16, v_uint16x8, uchar, u8)
-OPENCV_HAL_IMPL_NEON_EXPAND(v_int8x16, v_int16x8, schar, s8)
-OPENCV_HAL_IMPL_NEON_EXPAND(v_uint16x8, v_uint32x4, ushort, u16)
-OPENCV_HAL_IMPL_NEON_EXPAND(v_int16x8, v_int32x4, short, s16)
-OPENCV_HAL_IMPL_NEON_EXPAND(v_uint32x4, v_uint64x2, uint, u32)
-OPENCV_HAL_IMPL_NEON_EXPAND(v_int32x4, v_int64x2, int, s32)
-
-inline v_uint32x4 v_load_expand_q(const uchar* ptr)
-{
-    typedef unsigned int CV_DECL_ALIGNED(1) unaligned_uint;
-    uint8x8_t v0 = vcreate_u8(*(unaligned_uint*)ptr);
-    uint16x4_t v1 = vget_low_u16(vmovl_u8(v0));
-    return v_uint32x4(vmovl_u16(v1));
-}
-
-inline v_int32x4 v_load_expand_q(const schar* ptr)
-{
-    typedef unsigned int CV_DECL_ALIGNED(1) unaligned_uint;
-    int8x8_t v0 = vcreate_s8(*(unaligned_uint*)ptr);
-    int16x4_t v1 = vget_low_s16(vmovl_s8(v0));
-    return v_int32x4(vmovl_s16(v1));
-}
-
-#if defined(__aarch64__) || defined(_M_ARM64)
-#define OPENCV_HAL_IMPL_NEON_UNPACKS(_Tpvec, suffix) \
-inline void v_zip(const v_##_Tpvec& a0, const v_##_Tpvec& a1, v_##_Tpvec& b0, v_##_Tpvec& b1) \
-{ \
-    b0.val = vzip1q_##suffix(a0.val, a1.val); \
-    b1.val = vzip2q_##suffix(a0.val, a1.val); \
-} \
-inline v_##_Tpvec v_combine_low(const v_##_Tpvec& a, const v_##_Tpvec& b) \
-{ \
-    return v_##_Tpvec(vcombine_##suffix(vget_low_##suffix(a.val), vget_low_##suffix(b.val))); \
-} \
-inline v_##_Tpvec v_combine_high(const v_##_Tpvec& a, const v_##_Tpvec& b) \
-{ \
-    return v_##_Tpvec(vcombine_##suffix(vget_high_##suffix(a.val), vget_high_##suffix(b.val))); \
-} \
-inline void v_recombine(const v_##_Tpvec& a, const v_##_Tpvec& b, v_##_Tpvec& c, v_##_Tpvec& d) \
-{ \
-    c.val = vcombine_##suffix(vget_low_##suffix(a.val), vget_low_##suffix(b.val)); \
-    d.val = vcombine_##suffix(vget_high_##suffix(a.val), vget_high_##suffix(b.val)); \
-}
-#else
-#define OPENCV_HAL_IMPL_NEON_UNPACKS(_Tpvec, suffix) \
-inline void v_zip(const v_##_Tpvec& a0, const v_##_Tpvec& a1, v_##_Tpvec& b0, v_##_Tpvec& b1) \
-{ \
-    _Tpvec##x2_t p = vzipq_##suffix(a0.val, a1.val); \
-    b0.val = p.val[0]; \
-    b1.val = p.val[1]; \
-} \
-inline v_##_Tpvec v_combine_low(const v_##_Tpvec& a, const v_##_Tpvec& b) \
-{ \
-    return v_##_Tpvec(vcombine_##suffix(vget_low_##suffix(a.val), vget_low_##suffix(b.val))); \
-} \
-inline v_##_Tpvec v_combine_high(const v_##_Tpvec& a, const v_##_Tpvec& b) \
-{ \
-    return v_##_Tpvec(vcombine_##suffix(vget_high_##suffix(a.val), vget_high_##suffix(b.val))); \
-} \
-inline void v_recombine(const v_##_Tpvec& a, const v_##_Tpvec& b, v_##_Tpvec& c, v_##_Tpvec& d) \
-{ \
-    c.val = vcombine_##suffix(vget_low_##suffix(a.val), vget_low_##suffix(b.val)); \
-    d.val = vcombine_##suffix(vget_high_##suffix(a.val), vget_high_##suffix(b.val)); \
-}
-#endif
-
-OPENCV_HAL_IMPL_NEON_UNPACKS(uint8x16, u8)
-OPENCV_HAL_IMPL_NEON_UNPACKS(int8x16, s8)
-OPENCV_HAL_IMPL_NEON_UNPACKS(uint16x8, u16)
-OPENCV_HAL_IMPL_NEON_UNPACKS(int16x8, s16)
-OPENCV_HAL_IMPL_NEON_UNPACKS(uint32x4, u32)
-OPENCV_HAL_IMPL_NEON_UNPACKS(int32x4, s32)
-OPENCV_HAL_IMPL_NEON_UNPACKS(float32x4, f32)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_NEON_UNPACKS(float64x2, f64)
-#endif
-
-inline v_uint8x16 v_reverse(const v_uint8x16 &a)
-{
-    uint8x16_t vec = vrev64q_u8(a.val);
-    return v_uint8x16(vextq_u8(vec, vec, 8));
-}
-
-inline v_int8x16 v_reverse(const v_int8x16 &a)
-{ return v_reinterpret_as_s8(v_reverse(v_reinterpret_as_u8(a))); }
-
-inline v_uint16x8 v_reverse(const v_uint16x8 &a)
-{
-    uint16x8_t vec = vrev64q_u16(a.val);
-    return v_uint16x8(vextq_u16(vec, vec, 4));
-}
-
-inline v_int16x8 v_reverse(const v_int16x8 &a)
-{ return v_reinterpret_as_s16(v_reverse(v_reinterpret_as_u16(a))); }
-
-inline v_uint32x4 v_reverse(const v_uint32x4 &a)
-{
-    uint32x4_t vec = vrev64q_u32(a.val);
-    return v_uint32x4(vextq_u32(vec, vec, 2));
-}
-
-inline v_int32x4 v_reverse(const v_int32x4 &a)
-{ return v_reinterpret_as_s32(v_reverse(v_reinterpret_as_u32(a))); }
-
-inline v_float32x4 v_reverse(const v_float32x4 &a)
-{ return v_reinterpret_as_f32(v_reverse(v_reinterpret_as_u32(a))); }
-
-inline v_uint64x2 v_reverse(const v_uint64x2 &a)
-{
-    uint64x2_t vec = a.val;
-    uint64x1_t vec_lo = vget_low_u64(vec);
-    uint64x1_t vec_hi = vget_high_u64(vec);
-    return v_uint64x2(vcombine_u64(vec_hi, vec_lo));
-}
-
-inline v_int64x2 v_reverse(const v_int64x2 &a)
-{ return v_reinterpret_as_s64(v_reverse(v_reinterpret_as_u64(a))); }
-
-#if CV_SIMD128_64F
-inline v_float64x2 v_reverse(const v_float64x2 &a)
-{ return v_reinterpret_as_f64(v_reverse(v_reinterpret_as_u64(a))); }
-#endif
-
-#define OPENCV_HAL_IMPL_NEON_EXTRACT(_Tpvec, suffix) \
-template <int s> \
-inline v_##_Tpvec v_extract(const v_##_Tpvec& a, const v_##_Tpvec& b) \
-{ \
-    return v_##_Tpvec(vextq_##suffix(a.val, b.val, s)); \
-}
-
-OPENCV_HAL_IMPL_NEON_EXTRACT(uint8x16, u8)
-OPENCV_HAL_IMPL_NEON_EXTRACT(int8x16, s8)
-OPENCV_HAL_IMPL_NEON_EXTRACT(uint16x8, u16)
-OPENCV_HAL_IMPL_NEON_EXTRACT(int16x8, s16)
-OPENCV_HAL_IMPL_NEON_EXTRACT(uint32x4, u32)
-OPENCV_HAL_IMPL_NEON_EXTRACT(int32x4, s32)
-OPENCV_HAL_IMPL_NEON_EXTRACT(uint64x2, u64)
-OPENCV_HAL_IMPL_NEON_EXTRACT(int64x2, s64)
-OPENCV_HAL_IMPL_NEON_EXTRACT(float32x4, f32)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_NEON_EXTRACT(float64x2, f64)
-#endif
-
-#define OPENCV_HAL_IMPL_NEON_EXTRACT_N(_Tpvec, _Tp, suffix) \
-template<int i> inline _Tp v_extract_n(_Tpvec v) { return vgetq_lane_##suffix(v.val, i); }
-
-OPENCV_HAL_IMPL_NEON_EXTRACT_N(v_uint8x16, uchar, u8)
-OPENCV_HAL_IMPL_NEON_EXTRACT_N(v_int8x16, schar, s8)
-OPENCV_HAL_IMPL_NEON_EXTRACT_N(v_uint16x8, ushort, u16)
-OPENCV_HAL_IMPL_NEON_EXTRACT_N(v_int16x8, short, s16)
-OPENCV_HAL_IMPL_NEON_EXTRACT_N(v_uint32x4, uint, u32)
-OPENCV_HAL_IMPL_NEON_EXTRACT_N(v_int32x4, int, s32)
-OPENCV_HAL_IMPL_NEON_EXTRACT_N(v_uint64x2, uint64, u64)
-OPENCV_HAL_IMPL_NEON_EXTRACT_N(v_int64x2, int64, s64)
-OPENCV_HAL_IMPL_NEON_EXTRACT_N(v_float32x4, float, f32)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_NEON_EXTRACT_N(v_float64x2, double, f64)
-#endif
-
-#define OPENCV_HAL_IMPL_NEON_BROADCAST(_Tpvec, _Tp, suffix) \
-template<int i> inline _Tpvec v_broadcast_element(_Tpvec v) { _Tp t = v_extract_n<i>(v); return v_setall_##suffix(t); }
-
-OPENCV_HAL_IMPL_NEON_BROADCAST(v_uint8x16, uchar, u8)
-OPENCV_HAL_IMPL_NEON_BROADCAST(v_int8x16, schar, s8)
-OPENCV_HAL_IMPL_NEON_BROADCAST(v_uint16x8, ushort, u16)
-OPENCV_HAL_IMPL_NEON_BROADCAST(v_int16x8, short, s16)
-OPENCV_HAL_IMPL_NEON_BROADCAST(v_uint32x4, uint, u32)
-OPENCV_HAL_IMPL_NEON_BROADCAST(v_int32x4, int, s32)
-OPENCV_HAL_IMPL_NEON_BROADCAST(v_uint64x2, uint64, u64)
-OPENCV_HAL_IMPL_NEON_BROADCAST(v_int64x2, int64, s64)
-OPENCV_HAL_IMPL_NEON_BROADCAST(v_float32x4, float, f32)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_NEON_BROADCAST(v_float64x2, double, f64)
-#endif
-
-#if CV_SIMD128_64F
-inline v_int32x4 v_round(const v_float32x4& a)
-{
-    float32x4_t a_ = a.val;
-    int32x4_t result;
-    __asm__ ("fcvtns %0.4s, %1.4s"
-             : "=w"(result)
-             : "w"(a_)
-             : /* No clobbers */);
-    return v_int32x4(result);
-}
-#else
-inline v_int32x4 v_round(const v_float32x4& a)
-{
-    static const int32x4_t v_sign = vdupq_n_s32(1 << 31),
-        v_05 = vreinterpretq_s32_f32(vdupq_n_f32(0.5f));
-
-    int32x4_t v_addition = vorrq_s32(v_05, vandq_s32(v_sign, vreinterpretq_s32_f32(a.val)));
-    return v_int32x4(vcvtq_s32_f32(vaddq_f32(a.val, vreinterpretq_f32_s32(v_addition))));
-}
-#endif
-inline v_int32x4 v_floor(const v_float32x4& a)
-{
-    int32x4_t a1 = vcvtq_s32_f32(a.val);
-    uint32x4_t mask = vcgtq_f32(vcvtq_f32_s32(a1), a.val);
-    return v_int32x4(vaddq_s32(a1, vreinterpretq_s32_u32(mask)));
-}
-
-inline v_int32x4 v_ceil(const v_float32x4& a)
-{
-    int32x4_t a1 = vcvtq_s32_f32(a.val);
-    uint32x4_t mask = vcgtq_f32(a.val, vcvtq_f32_s32(a1));
-    return v_int32x4(vsubq_s32(a1, vreinterpretq_s32_u32(mask)));
-}
-
-inline v_int32x4 v_trunc(const v_float32x4& a)
-{ return v_int32x4(vcvtq_s32_f32(a.val)); }
-
-#if CV_SIMD128_64F
-inline v_int32x4 v_round(const v_float64x2& a)
-{
-    static const int32x2_t zero = vdup_n_s32(0);
-    return v_int32x4(vcombine_s32(vmovn_s64(vcvtaq_s64_f64(a.val)), zero));
-}
-
-inline v_int32x4 v_round(const v_float64x2& a, const v_float64x2& b)
-{
-    return v_int32x4(vcombine_s32(vmovn_s64(vcvtaq_s64_f64(a.val)), vmovn_s64(vcvtaq_s64_f64(b.val))));
-}
-
-inline v_int32x4 v_floor(const v_float64x2& a)
-{
-    static const int32x2_t zero = vdup_n_s32(0);
-    int64x2_t a1 = vcvtq_s64_f64(a.val);
-    uint64x2_t mask = vcgtq_f64(vcvtq_f64_s64(a1), a.val);
-    a1 = vaddq_s64(a1, vreinterpretq_s64_u64(mask));
-    return v_int32x4(vcombine_s32(vmovn_s64(a1), zero));
-}
-
-inline v_int32x4 v_ceil(const v_float64x2& a)
-{
-    static const int32x2_t zero = vdup_n_s32(0);
-    int64x2_t a1 = vcvtq_s64_f64(a.val);
-    uint64x2_t mask = vcgtq_f64(a.val, vcvtq_f64_s64(a1));
-    a1 = vsubq_s64(a1, vreinterpretq_s64_u64(mask));
-    return v_int32x4(vcombine_s32(vmovn_s64(a1), zero));
-}
-
-inline v_int32x4 v_trunc(const v_float64x2& a)
-{
-    static const int32x2_t zero = vdup_n_s32(0);
-    return v_int32x4(vcombine_s32(vmovn_s64(vcvtaq_s64_f64(a.val)), zero));
-}
-#endif
-
-#if CV_NEON_AARCH64
-#define OPENCV_HAL_IMPL_NEON_TRANSPOSE4x4(_Tpvec, suffix) \
-inline void v_transpose4x4(const v_##_Tpvec& a0, const v_##_Tpvec& a1, \
-                         const v_##_Tpvec& a2, const v_##_Tpvec& a3, \
-                         v_##_Tpvec& b0, v_##_Tpvec& b1, \
-                         v_##_Tpvec& b2, v_##_Tpvec& b3) \
-{ \
-    /* -- Pass 1: 64b transpose */ \
-    _Tpvec##_t t0 = vreinterpretq_##suffix##32_##suffix##64( \
-                        vtrn1q_##suffix##64(vreinterpretq_##suffix##64_##suffix##32(a0.val), \
-                                            vreinterpretq_##suffix##64_##suffix##32(a2.val))); \
-    _Tpvec##_t t1 = vreinterpretq_##suffix##32_##suffix##64( \
-                        vtrn1q_##suffix##64(vreinterpretq_##suffix##64_##suffix##32(a1.val), \
-                                            vreinterpretq_##suffix##64_##suffix##32(a3.val))); \
-    _Tpvec##_t t2 = vreinterpretq_##suffix##32_##suffix##64( \
-                        vtrn2q_##suffix##64(vreinterpretq_##suffix##64_##suffix##32(a0.val), \
-                                            vreinterpretq_##suffix##64_##suffix##32(a2.val))); \
-    _Tpvec##_t t3 = vreinterpretq_##suffix##32_##suffix##64( \
-                        vtrn2q_##suffix##64(vreinterpretq_##suffix##64_##suffix##32(a1.val), \
-                                            vreinterpretq_##suffix##64_##suffix##32(a3.val))); \
-    /* -- Pass 2: 32b transpose */ \
-    b0.val = vtrn1q_##suffix##32(t0, t1); \
-    b1.val = vtrn2q_##suffix##32(t0, t1); \
-    b2.val = vtrn1q_##suffix##32(t2, t3); \
-    b3.val = vtrn2q_##suffix##32(t2, t3); \
-}
-
-OPENCV_HAL_IMPL_NEON_TRANSPOSE4x4(uint32x4, u)
-OPENCV_HAL_IMPL_NEON_TRANSPOSE4x4(int32x4, s)
-OPENCV_HAL_IMPL_NEON_TRANSPOSE4x4(float32x4, f)
-#else // #if CV_NEON_AARCH64
-#define OPENCV_HAL_IMPL_NEON_TRANSPOSE4x4(_Tpvec, suffix) \
-inline void v_transpose4x4(const v_##_Tpvec& a0, const v_##_Tpvec& a1, \
-                         const v_##_Tpvec& a2, const v_##_Tpvec& a3, \
-                         v_##_Tpvec& b0, v_##_Tpvec& b1, \
-                         v_##_Tpvec& b2, v_##_Tpvec& b3) \
-{ \
-    /* m00 m01 m02 m03 */ \
-    /* m10 m11 m12 m13 */ \
-    /* m20 m21 m22 m23 */ \
-    /* m30 m31 m32 m33 */ \
-    _Tpvec##x2_t t0 = vtrnq_##suffix(a0.val, a1.val); \
-    _Tpvec##x2_t t1 = vtrnq_##suffix(a2.val, a3.val); \
-    /* m00 m10 m02 m12 */ \
-    /* m01 m11 m03 m13 */ \
-    /* m20 m30 m22 m32 */ \
-    /* m21 m31 m23 m33 */ \
-    b0.val = vcombine_##suffix(vget_low_##suffix(t0.val[0]), vget_low_##suffix(t1.val[0])); \
-    b1.val = vcombine_##suffix(vget_low_##suffix(t0.val[1]), vget_low_##suffix(t1.val[1])); \
-    b2.val = vcombine_##suffix(vget_high_##suffix(t0.val[0]), vget_high_##suffix(t1.val[0])); \
-    b3.val = vcombine_##suffix(vget_high_##suffix(t0.val[1]), vget_high_##suffix(t1.val[1])); \
-}
-
-OPENCV_HAL_IMPL_NEON_TRANSPOSE4x4(uint32x4, u32)
-OPENCV_HAL_IMPL_NEON_TRANSPOSE4x4(int32x4, s32)
-OPENCV_HAL_IMPL_NEON_TRANSPOSE4x4(float32x4, f32)
-#endif // #if CV_NEON_AARCH64
-
-#define OPENCV_HAL_IMPL_NEON_INTERLEAVED(_Tpvec, _Tp, suffix) \
-inline void v_load_deinterleave(const _Tp* ptr, v_##_Tpvec& a, v_##_Tpvec& b) \
-{ \
-    _Tpvec##x2_t v = vld2q_##suffix(ptr); \
-    a.val = v.val[0]; \
-    b.val = v.val[1]; \
-} \
-inline void v_load_deinterleave(const _Tp* ptr, v_##_Tpvec& a, v_##_Tpvec& b, v_##_Tpvec& c) \
-{ \
-    _Tpvec##x3_t v = vld3q_##suffix(ptr); \
-    a.val = v.val[0]; \
-    b.val = v.val[1]; \
-    c.val = v.val[2]; \
-} \
-inline void v_load_deinterleave(const _Tp* ptr, v_##_Tpvec& a, v_##_Tpvec& b, \
-                                v_##_Tpvec& c, v_##_Tpvec& d) \
-{ \
-    _Tpvec##x4_t v = vld4q_##suffix(ptr); \
-    a.val = v.val[0]; \
-    b.val = v.val[1]; \
-    c.val = v.val[2]; \
-    d.val = v.val[3]; \
-} \
-inline void v_store_interleave( _Tp* ptr, const v_##_Tpvec& a, const v_##_Tpvec& b, \
-                                hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ \
-    _Tpvec##x2_t v; \
-    v.val[0] = a.val; \
-    v.val[1] = b.val; \
-    vst2q_##suffix(ptr, v); \
-} \
-inline void v_store_interleave( _Tp* ptr, const v_##_Tpvec& a, const v_##_Tpvec& b, \
-                                const v_##_Tpvec& c, hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ \
-    _Tpvec##x3_t v; \
-    v.val[0] = a.val; \
-    v.val[1] = b.val; \
-    v.val[2] = c.val; \
-    vst3q_##suffix(ptr, v); \
-} \
-inline void v_store_interleave( _Tp* ptr, const v_##_Tpvec& a, const v_##_Tpvec& b, \
-                                const v_##_Tpvec& c, const v_##_Tpvec& d, \
-                                hal::StoreMode /*mode*/=hal::STORE_UNALIGNED ) \
-{ \
-    _Tpvec##x4_t v; \
-    v.val[0] = a.val; \
-    v.val[1] = b.val; \
-    v.val[2] = c.val; \
-    v.val[3] = d.val; \
-    vst4q_##suffix(ptr, v); \
-}
-
-#define OPENCV_HAL_IMPL_NEON_INTERLEAVED_INT64(tp, suffix) \
-inline void v_load_deinterleave( const tp* ptr, v_##tp##x2& a, v_##tp##x2& b ) \
-{ \
-    tp##x1_t a0 = vld1_##suffix(ptr); \
-    tp##x1_t b0 = vld1_##suffix(ptr + 1); \
-    tp##x1_t a1 = vld1_##suffix(ptr + 2); \
-    tp##x1_t b1 = vld1_##suffix(ptr + 3); \
-    a = v_##tp##x2(vcombine_##suffix(a0, a1)); \
-    b = v_##tp##x2(vcombine_##suffix(b0, b1)); \
-} \
- \
-inline void v_load_deinterleave( const tp* ptr, v_##tp##x2& a, \
-                                 v_##tp##x2& b, v_##tp##x2& c ) \
-{ \
-    tp##x1_t a0 = vld1_##suffix(ptr); \
-    tp##x1_t b0 = vld1_##suffix(ptr + 1); \
-    tp##x1_t c0 = vld1_##suffix(ptr + 2); \
-    tp##x1_t a1 = vld1_##suffix(ptr + 3); \
-    tp##x1_t b1 = vld1_##suffix(ptr + 4); \
-    tp##x1_t c1 = vld1_##suffix(ptr + 5); \
-    a = v_##tp##x2(vcombine_##suffix(a0, a1)); \
-    b = v_##tp##x2(vcombine_##suffix(b0, b1)); \
-    c = v_##tp##x2(vcombine_##suffix(c0, c1)); \
-} \
- \
-inline void v_load_deinterleave( const tp* ptr, v_##tp##x2& a, v_##tp##x2& b, \
-                                 v_##tp##x2& c, v_##tp##x2& d ) \
-{ \
-    tp##x1_t a0 = vld1_##suffix(ptr); \
-    tp##x1_t b0 = vld1_##suffix(ptr + 1); \
-    tp##x1_t c0 = vld1_##suffix(ptr + 2); \
-    tp##x1_t d0 = vld1_##suffix(ptr + 3); \
-    tp##x1_t a1 = vld1_##suffix(ptr + 4); \
-    tp##x1_t b1 = vld1_##suffix(ptr + 5); \
-    tp##x1_t c1 = vld1_##suffix(ptr + 6); \
-    tp##x1_t d1 = vld1_##suffix(ptr + 7); \
-    a = v_##tp##x2(vcombine_##suffix(a0, a1)); \
-    b = v_##tp##x2(vcombine_##suffix(b0, b1)); \
-    c = v_##tp##x2(vcombine_##suffix(c0, c1)); \
-    d = v_##tp##x2(vcombine_##suffix(d0, d1)); \
-} \
- \
-inline void v_store_interleave( tp* ptr, const v_##tp##x2& a, const v_##tp##x2& b, \
-                                hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ \
-    vst1_##suffix(ptr, vget_low_##suffix(a.val)); \
-    vst1_##suffix(ptr + 1, vget_low_##suffix(b.val)); \
-    vst1_##suffix(ptr + 2, vget_high_##suffix(a.val)); \
-    vst1_##suffix(ptr + 3, vget_high_##suffix(b.val)); \
-} \
- \
-inline void v_store_interleave( tp* ptr, const v_##tp##x2& a, \
-                                const v_##tp##x2& b, const v_##tp##x2& c, \
-                                hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ \
-    vst1_##suffix(ptr, vget_low_##suffix(a.val)); \
-    vst1_##suffix(ptr + 1, vget_low_##suffix(b.val)); \
-    vst1_##suffix(ptr + 2, vget_low_##suffix(c.val)); \
-    vst1_##suffix(ptr + 3, vget_high_##suffix(a.val)); \
-    vst1_##suffix(ptr + 4, vget_high_##suffix(b.val)); \
-    vst1_##suffix(ptr + 5, vget_high_##suffix(c.val)); \
-} \
- \
-inline void v_store_interleave( tp* ptr, const v_##tp##x2& a, const v_##tp##x2& b, \
-                                const v_##tp##x2& c, const v_##tp##x2& d, \
-                                hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ \
-    vst1_##suffix(ptr, vget_low_##suffix(a.val)); \
-    vst1_##suffix(ptr + 1, vget_low_##suffix(b.val)); \
-    vst1_##suffix(ptr + 2, vget_low_##suffix(c.val)); \
-    vst1_##suffix(ptr + 3, vget_low_##suffix(d.val)); \
-    vst1_##suffix(ptr + 4, vget_high_##suffix(a.val)); \
-    vst1_##suffix(ptr + 5, vget_high_##suffix(b.val)); \
-    vst1_##suffix(ptr + 6, vget_high_##suffix(c.val)); \
-    vst1_##suffix(ptr + 7, vget_high_##suffix(d.val)); \
-}
-
-OPENCV_HAL_IMPL_NEON_INTERLEAVED(uint8x16, uchar, u8)
-OPENCV_HAL_IMPL_NEON_INTERLEAVED(int8x16, schar, s8)
-OPENCV_HAL_IMPL_NEON_INTERLEAVED(uint16x8, ushort, u16)
-OPENCV_HAL_IMPL_NEON_INTERLEAVED(int16x8, short, s16)
-OPENCV_HAL_IMPL_NEON_INTERLEAVED(uint32x4, unsigned, u32)
-OPENCV_HAL_IMPL_NEON_INTERLEAVED(int32x4, int, s32)
-OPENCV_HAL_IMPL_NEON_INTERLEAVED(float32x4, float, f32)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_NEON_INTERLEAVED(float64x2, double, f64)
-#endif
-
-OPENCV_HAL_IMPL_NEON_INTERLEAVED_INT64(int64, s64)
-OPENCV_HAL_IMPL_NEON_INTERLEAVED_INT64(uint64, u64)
-
-inline v_float32x4 v_cvt_f32(const v_int32x4& a)
-{
-    return v_float32x4(vcvtq_f32_s32(a.val));
-}
-
-#if CV_SIMD128_64F
-inline v_float32x4 v_cvt_f32(const v_float64x2& a)
-{
-    float32x2_t zero = vdup_n_f32(0.0f);
-    return v_float32x4(vcombine_f32(vcvt_f32_f64(a.val), zero));
-}
-
-inline v_float32x4 v_cvt_f32(const v_float64x2& a, const v_float64x2& b)
-{
-    return v_float32x4(vcombine_f32(vcvt_f32_f64(a.val), vcvt_f32_f64(b.val)));
-}
-
-inline v_float64x2 v_cvt_f64(const v_int32x4& a)
-{
-    return v_float64x2(vcvt_f64_f32(vcvt_f32_s32(vget_low_s32(a.val))));
-}
-
-inline v_float64x2 v_cvt_f64_high(const v_int32x4& a)
-{
-    return v_float64x2(vcvt_f64_f32(vcvt_f32_s32(vget_high_s32(a.val))));
-}
-
-inline v_float64x2 v_cvt_f64(const v_float32x4& a)
-{
-    return v_float64x2(vcvt_f64_f32(vget_low_f32(a.val)));
-}
-
-inline v_float64x2 v_cvt_f64_high(const v_float32x4& a)
-{
-    return v_float64x2(vcvt_f64_f32(vget_high_f32(a.val)));
-}
-
-inline v_float64x2 v_cvt_f64(const v_int64x2& a)
-{  return v_float64x2(vcvtq_f64_s64(a.val)); }
-
-#endif
-
-////////////// Lookup table access ////////////////////
-
-inline v_int8x16 v_lut(const schar* tab, const int* idx)
-{
-    schar CV_DECL_ALIGNED(32) elems[16] =
-    {
-        tab[idx[ 0]],
-        tab[idx[ 1]],
-        tab[idx[ 2]],
-        tab[idx[ 3]],
-        tab[idx[ 4]],
-        tab[idx[ 5]],
-        tab[idx[ 6]],
-        tab[idx[ 7]],
-        tab[idx[ 8]],
-        tab[idx[ 9]],
-        tab[idx[10]],
-        tab[idx[11]],
-        tab[idx[12]],
-        tab[idx[13]],
-        tab[idx[14]],
-        tab[idx[15]]
-    };
-    return v_int8x16(vld1q_s8(elems));
-}
-inline v_int8x16 v_lut_pairs(const schar* tab, const int* idx)
-{
-    schar CV_DECL_ALIGNED(32) elems[16] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[1]],
-        tab[idx[1] + 1],
-        tab[idx[2]],
-        tab[idx[2] + 1],
-        tab[idx[3]],
-        tab[idx[3] + 1],
-        tab[idx[4]],
-        tab[idx[4] + 1],
-        tab[idx[5]],
-        tab[idx[5] + 1],
-        tab[idx[6]],
-        tab[idx[6] + 1],
-        tab[idx[7]],
-        tab[idx[7] + 1]
-    };
-    return v_int8x16(vld1q_s8(elems));
-}
-inline v_int8x16 v_lut_quads(const schar* tab, const int* idx)
-{
-    schar CV_DECL_ALIGNED(32) elems[16] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[0] + 2],
-        tab[idx[0] + 3],
-        tab[idx[1]],
-        tab[idx[1] + 1],
-        tab[idx[1] + 2],
-        tab[idx[1] + 3],
-        tab[idx[2]],
-        tab[idx[2] + 1],
-        tab[idx[2] + 2],
-        tab[idx[2] + 3],
-        tab[idx[3]],
-        tab[idx[3] + 1],
-        tab[idx[3] + 2],
-        tab[idx[3] + 3]
-    };
-    return v_int8x16(vld1q_s8(elems));
-}
-inline v_uint8x16 v_lut(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut((schar*)tab, idx)); }
-inline v_uint8x16 v_lut_pairs(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut_pairs((schar*)tab, idx)); }
-inline v_uint8x16 v_lut_quads(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut_quads((schar*)tab, idx)); }
-
-inline v_int16x8 v_lut(const short* tab, const int* idx)
-{
-    short CV_DECL_ALIGNED(32) elems[8] =
-    {
-        tab[idx[0]],
-        tab[idx[1]],
-        tab[idx[2]],
-        tab[idx[3]],
-        tab[idx[4]],
-        tab[idx[5]],
-        tab[idx[6]],
-        tab[idx[7]]
-    };
-    return v_int16x8(vld1q_s16(elems));
-}
-inline v_int16x8 v_lut_pairs(const short* tab, const int* idx)
-{
-    short CV_DECL_ALIGNED(32) elems[8] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[1]],
-        tab[idx[1] + 1],
-        tab[idx[2]],
-        tab[idx[2] + 1],
-        tab[idx[3]],
-        tab[idx[3] + 1]
-    };
-    return v_int16x8(vld1q_s16(elems));
-}
-inline v_int16x8 v_lut_quads(const short* tab, const int* idx)
-{
-    return v_int16x8(vcombine_s16(vld1_s16(tab + idx[0]), vld1_s16(tab + idx[1])));
-}
-inline v_uint16x8 v_lut(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut((short*)tab, idx)); }
-inline v_uint16x8 v_lut_pairs(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut_pairs((short*)tab, idx)); }
-inline v_uint16x8 v_lut_quads(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut_quads((short*)tab, idx)); }
-
-inline v_int32x4 v_lut(const int* tab, const int* idx)
-{
-    int CV_DECL_ALIGNED(32) elems[4] =
-    {
-        tab[idx[0]],
-        tab[idx[1]],
-        tab[idx[2]],
-        tab[idx[3]]
-    };
-    return v_int32x4(vld1q_s32(elems));
-}
-inline v_int32x4 v_lut_pairs(const int* tab, const int* idx)
-{
-    return v_int32x4(vcombine_s32(vld1_s32(tab + idx[0]), vld1_s32(tab + idx[1])));
-}
-inline v_int32x4 v_lut_quads(const int* tab, const int* idx)
-{
-    return v_int32x4(vld1q_s32(tab + idx[0]));
-}
-inline v_uint32x4 v_lut(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut((int*)tab, idx)); }
-inline v_uint32x4 v_lut_pairs(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut_pairs((int*)tab, idx)); }
-inline v_uint32x4 v_lut_quads(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut_quads((int*)tab, idx)); }
-
-inline v_int64x2 v_lut(const int64_t* tab, const int* idx)
-{
-    return v_int64x2(vcombine_s64(vcreate_s64(tab[idx[0]]), vcreate_s64(tab[idx[1]])));
-}
-inline v_int64x2 v_lut_pairs(const int64_t* tab, const int* idx)
-{
-    return v_int64x2(vld1q_s64(tab + idx[0]));
-}
-inline v_uint64x2 v_lut(const uint64_t* tab, const int* idx) { return v_reinterpret_as_u64(v_lut((const int64_t *)tab, idx)); }
-inline v_uint64x2 v_lut_pairs(const uint64_t* tab, const int* idx) { return v_reinterpret_as_u64(v_lut_pairs((const int64_t *)tab, idx)); }
-
-inline v_float32x4 v_lut(const float* tab, const int* idx)
-{
-    float CV_DECL_ALIGNED(32) elems[4] =
-    {
-        tab[idx[0]],
-        tab[idx[1]],
-        tab[idx[2]],
-        tab[idx[3]]
-    };
-    return v_float32x4(vld1q_f32(elems));
-}
-inline v_float32x4 v_lut_pairs(const float* tab, const int* idx)
-{
-    typedef uint64 CV_DECL_ALIGNED(1) unaligned_uint64;
-
-    uint64 CV_DECL_ALIGNED(32) elems[2] =
-    {
-        *(unaligned_uint64*)(tab + idx[0]),
-        *(unaligned_uint64*)(tab + idx[1])
-    };
-    return v_float32x4(vreinterpretq_f32_u64(vld1q_u64(elems)));
-}
-inline v_float32x4 v_lut_quads(const float* tab, const int* idx)
-{
-    return v_float32x4(vld1q_f32(tab + idx[0]));
-}
-
-inline v_int32x4 v_lut(const int* tab, const v_int32x4& idxvec)
-{
-    int CV_DECL_ALIGNED(32) elems[4] =
-    {
-        tab[vgetq_lane_s32(idxvec.val, 0)],
-        tab[vgetq_lane_s32(idxvec.val, 1)],
-        tab[vgetq_lane_s32(idxvec.val, 2)],
-        tab[vgetq_lane_s32(idxvec.val, 3)]
-    };
-    return v_int32x4(vld1q_s32(elems));
-}
-
-inline v_uint32x4 v_lut(const unsigned* tab, const v_int32x4& idxvec)
-{
-    unsigned CV_DECL_ALIGNED(32) elems[4] =
-    {
-        tab[vgetq_lane_s32(idxvec.val, 0)],
-        tab[vgetq_lane_s32(idxvec.val, 1)],
-        tab[vgetq_lane_s32(idxvec.val, 2)],
-        tab[vgetq_lane_s32(idxvec.val, 3)]
-    };
-    return v_uint32x4(vld1q_u32(elems));
-}
-
-inline v_float32x4 v_lut(const float* tab, const v_int32x4& idxvec)
-{
-    float CV_DECL_ALIGNED(32) elems[4] =
-    {
-        tab[vgetq_lane_s32(idxvec.val, 0)],
-        tab[vgetq_lane_s32(idxvec.val, 1)],
-        tab[vgetq_lane_s32(idxvec.val, 2)],
-        tab[vgetq_lane_s32(idxvec.val, 3)]
-    };
-    return v_float32x4(vld1q_f32(elems));
-}
-
-inline void v_lut_deinterleave(const float* tab, const v_int32x4& idxvec, v_float32x4& x, v_float32x4& y)
-{
-    /*int CV_DECL_ALIGNED(32) idx[4];
-    v_store(idx, idxvec);
-
-    float32x4_t xy02 = vcombine_f32(vld1_f32(tab + idx[0]), vld1_f32(tab + idx[2]));
-    float32x4_t xy13 = vcombine_f32(vld1_f32(tab + idx[1]), vld1_f32(tab + idx[3]));
-
-    float32x4x2_t xxyy = vuzpq_f32(xy02, xy13);
-    x = v_float32x4(xxyy.val[0]);
-    y = v_float32x4(xxyy.val[1]);*/
-    int CV_DECL_ALIGNED(32) idx[4];
-    v_store_aligned(idx, idxvec);
-
-    x = v_float32x4(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]]);
-    y = v_float32x4(tab[idx[0]+1], tab[idx[1]+1], tab[idx[2]+1], tab[idx[3]+1]);
-}
-
-inline v_int8x16 v_interleave_pairs(const v_int8x16& vec)
-{
-    return v_int8x16(vcombine_s8(vtbl1_s8(vget_low_s8(vec.val), vcreate_s8(0x0705060403010200)), vtbl1_s8(vget_high_s8(vec.val), vcreate_s8(0x0705060403010200))));
-}
-inline v_uint8x16 v_interleave_pairs(const v_uint8x16& vec) { return v_reinterpret_as_u8(v_interleave_pairs(v_reinterpret_as_s8(vec))); }
-inline v_int8x16 v_interleave_quads(const v_int8x16& vec)
-{
-    return v_int8x16(vcombine_s8(vtbl1_s8(vget_low_s8(vec.val), vcreate_s8(0x0703060205010400)), vtbl1_s8(vget_high_s8(vec.val), vcreate_s8(0x0703060205010400))));
-}
-inline v_uint8x16 v_interleave_quads(const v_uint8x16& vec) { return v_reinterpret_as_u8(v_interleave_quads(v_reinterpret_as_s8(vec))); }
-
-inline v_int16x8 v_interleave_pairs(const v_int16x8& vec)
-{
-    return v_int16x8(vreinterpretq_s16_s8(vcombine_s8(vtbl1_s8(vget_low_s8(vreinterpretq_s8_s16(vec.val)), vcreate_s8(0x0706030205040100)), vtbl1_s8(vget_high_s8(vreinterpretq_s8_s16(vec.val)), vcreate_s8(0x0706030205040100)))));
-}
-inline v_uint16x8 v_interleave_pairs(const v_uint16x8& vec) { return v_reinterpret_as_u16(v_interleave_pairs(v_reinterpret_as_s16(vec))); }
-inline v_int16x8 v_interleave_quads(const v_int16x8& vec)
-{
-    int16x4x2_t res = vzip_s16(vget_low_s16(vec.val), vget_high_s16(vec.val));
-    return v_int16x8(vcombine_s16(res.val[0], res.val[1]));
-}
-inline v_uint16x8 v_interleave_quads(const v_uint16x8& vec) { return v_reinterpret_as_u16(v_interleave_quads(v_reinterpret_as_s16(vec))); }
-
-inline v_int32x4 v_interleave_pairs(const v_int32x4& vec)
-{
-    int32x2x2_t res = vzip_s32(vget_low_s32(vec.val), vget_high_s32(vec.val));
-    return v_int32x4(vcombine_s32(res.val[0], res.val[1]));
-}
-inline v_uint32x4 v_interleave_pairs(const v_uint32x4& vec) { return v_reinterpret_as_u32(v_interleave_pairs(v_reinterpret_as_s32(vec))); }
-inline v_float32x4 v_interleave_pairs(const v_float32x4& vec) { return v_reinterpret_as_f32(v_interleave_pairs(v_reinterpret_as_s32(vec))); }
-
-inline v_int8x16 v_pack_triplets(const v_int8x16& vec)
-{
-    return v_int8x16(vextq_s8(vcombine_s8(vtbl1_s8(vget_low_s8(vec.val), vcreate_s8(0x0605040201000000)), vtbl1_s8(vget_high_s8(vec.val), vcreate_s8(0x0807060504020100))), vdupq_n_s8(0), 2));
-}
-inline v_uint8x16 v_pack_triplets(const v_uint8x16& vec) { return v_reinterpret_as_u8(v_pack_triplets(v_reinterpret_as_s8(vec))); }
-
-inline v_int16x8 v_pack_triplets(const v_int16x8& vec)
-{
-    return v_int16x8(vreinterpretq_s16_s8(vextq_s8(vcombine_s8(vtbl1_s8(vget_low_s8(vreinterpretq_s8_s16(vec.val)), vcreate_s8(0x0504030201000000)), vget_high_s8(vreinterpretq_s8_s16(vec.val))), vdupq_n_s8(0), 2)));
-}
-inline v_uint16x8 v_pack_triplets(const v_uint16x8& vec) { return v_reinterpret_as_u16(v_pack_triplets(v_reinterpret_as_s16(vec))); }
-
-inline v_int32x4 v_pack_triplets(const v_int32x4& vec) { return vec; }
-inline v_uint32x4 v_pack_triplets(const v_uint32x4& vec) { return vec; }
-inline v_float32x4 v_pack_triplets(const v_float32x4& vec) { return vec; }
-
-#if CV_SIMD128_64F
-inline v_float64x2 v_lut(const double* tab, const int* idx)
-{
-    double CV_DECL_ALIGNED(32) elems[2] =
-    {
-        tab[idx[0]],
-        tab[idx[1]]
-    };
-    return v_float64x2(vld1q_f64(elems));
-}
-
-inline v_float64x2 v_lut_pairs(const double* tab, const int* idx)
-{
-    return v_float64x2(vld1q_f64(tab + idx[0]));
-}
-
-inline v_float64x2 v_lut(const double* tab, const v_int32x4& idxvec)
-{
-    double CV_DECL_ALIGNED(32) elems[2] =
-    {
-        tab[vgetq_lane_s32(idxvec.val, 0)],
-        tab[vgetq_lane_s32(idxvec.val, 1)],
-    };
-    return v_float64x2(vld1q_f64(elems));
-}
-
-inline void v_lut_deinterleave(const double* tab, const v_int32x4& idxvec, v_float64x2& x, v_float64x2& y)
-{
-    int CV_DECL_ALIGNED(32) idx[4];
-    v_store_aligned(idx, idxvec);
-
-    x = v_float64x2(tab[idx[0]], tab[idx[1]]);
-    y = v_float64x2(tab[idx[0]+1], tab[idx[1]+1]);
-}
-#endif
-
-////// FP16 support ///////
-#if CV_FP16
-inline v_float32x4 v_load_expand(const float16_t* ptr)
-{
-    float16x4_t v =
-    #ifndef vld1_f16 // APPLE compiler defines vld1_f16 as macro
-        (float16x4_t)vld1_s16((const short*)ptr);
-    #else
-        vld1_f16((const __fp16*)ptr);
-    #endif
-    return v_float32x4(vcvt_f32_f16(v));
-}
-
-inline void v_pack_store(float16_t* ptr, const v_float32x4& v)
-{
-    float16x4_t hv = vcvt_f16_f32(v.val);
-
-    #ifndef vst1_f16 // APPLE compiler defines vst1_f16 as macro
-        vst1_s16((short*)ptr, (int16x4_t)hv);
-    #else
-        vst1_f16((__fp16*)ptr, hv);
-    #endif
-}
-#else
-inline v_float32x4 v_load_expand(const float16_t* ptr)
-{
-    const int N = 4;
-    float buf[N];
-    for( int i = 0; i < N; i++ ) buf[i] = (float)ptr[i];
-    return v_load(buf);
-}
-
-inline void v_pack_store(float16_t* ptr, const v_float32x4& v)
-{
-    const int N = 4;
-    float buf[N];
-    v_store(buf, v);
-    for( int i = 0; i < N; i++ ) ptr[i] = float16_t(buf[i]);
-}
-#endif
-
-inline void v_cleanup() {}
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END
-
-//! @endcond
-
-}
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_rvv.hpp b/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_rvv.hpp
deleted file mode 100644
index fe6c0776..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_rvv.hpp
+++ /dev/null
@@ -1,2887 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-// The original implementation has been contributed by Yin Zhang.
-// Copyright (C) 2020, Institute of Software, Chinese Academy of Sciences.
-
-#ifndef OPENCV_HAL_INTRIN_RVV_HPP
-#define OPENCV_HAL_INTRIN_RVV_HPP
-
-#include <algorithm>
-
-namespace cv
-{
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
-
-#define CV_SIMD128 1
-#define CV_SIMD128_64F 1
-
-//////////// Unsupported native intrinsics in C++ ////////////
-// The following types have been defined in clang, but not in GCC yet.
-#ifndef __clang__
-
-struct vuint8mf2_t
-{
-    uchar val[8] = {0};
-    vuint8mf2_t() {}
-    vuint8mf2_t(const uchar* ptr)
-    {
-        for (int i = 0; i < 8; ++i)
-        {
-            val[i] = ptr[i];
-        }
-    }
-};
-struct vint8mf2_t
-{
-    schar val[8] = {0};
-    vint8mf2_t() {}
-    vint8mf2_t(const schar* ptr)
-    {
-        for (int i = 0; i < 8; ++i)
-        {
-            val[i] = ptr[i];
-        }
-    }
-};
-struct vuint16mf2_t
-{
-    ushort val[4] = {0};
-    vuint16mf2_t() {}
-    vuint16mf2_t(const ushort* ptr)
-    {
-        for (int i = 0; i < 4; ++i)
-        {
-            val[i] = ptr[i];
-        }
-    }
-};
-struct vint16mf2_t
-{
-    short val[4] = {0};
-    vint16mf2_t() {}
-    vint16mf2_t(const short* ptr)
-    {
-        for (int i = 0; i < 4; ++i)
-        {
-            val[i] = ptr[i];
-        }
-    }
-};
-struct vuint32mf2_t
-{
-    unsigned val[2] = {0};
-    vuint32mf2_t() {}
-    vuint32mf2_t(const unsigned* ptr)
-    {
-        val[0] = ptr[0];
-        val[1] = ptr[1];
-    }
-};
-struct vint32mf2_t
-{
-    int val[2] = {0};
-    vint32mf2_t() {}
-    vint32mf2_t(const int* ptr)
-    {
-        val[0] = ptr[0];
-        val[1] = ptr[1];
-    }
-};
-struct vfloat32mf2_t
-{
-    float val[2] = {0};
-    vfloat32mf2_t() {}
-    vfloat32mf2_t(const float* ptr)
-    {
-        val[0] = ptr[0];
-        val[1] = ptr[1];
-    }
-};
-struct vuint64mf2_t
-{
-    uint64 val[1] = {0};
-    vuint64mf2_t() {}
-    vuint64mf2_t(const uint64* ptr)
-    {
-        val[0] = ptr[0];
-    }
-};
-struct vint64mf2_t
-{
-    int64 val[1] = {0};
-    vint64mf2_t() {}
-    vint64mf2_t(const int64* ptr)
-    {
-        val[0] = ptr[0];
-    }
-};
-struct vfloat64mf2_t
-{
-    double val[1] = {0};
-    vfloat64mf2_t() {}
-    vfloat64mf2_t(const double* ptr)
-    {
-        val[0] = ptr[0];
-    }
-};
-struct vuint8mf4_t
-{
-    uchar val[4] = {0};
-    vuint8mf4_t() {}
-    vuint8mf4_t(const uchar* ptr)
-    {
-        for (int i = 0; i < 4; ++i)
-        {
-            val[i] = ptr[i];
-        }
-    }
-};
-struct vint8mf4_t
-{
-    schar val[4] = {0};
-    vint8mf4_t() {}
-    vint8mf4_t(const schar* ptr)
-    {
-        for (int i = 0; i < 4; ++i)
-        {
-            val[i] = ptr[i];
-        }
-    }
-};
-
-#define OPENCV_HAL_IMPL_RVV_NATIVE_LOADSTORE_MF2(_Tpvec, _Tp, suffix, width, n) \
-inline _Tpvec vle##width##_v_##suffix##mf2(const _Tp* ptr, size_t vl) \
-{ \
-    CV_UNUSED(vl); \
-    return _Tpvec(ptr); \
-} \
-inline void vse##width##_v_##suffix##mf2(_Tp* ptr, _Tpvec v, size_t vl) \
-{ \
-    CV_UNUSED(vl); \
-    for (int i = 0; i < n; ++i) \
-    { \
-            ptr[i] = v.val[i]; \
-    } \
-}
-
-OPENCV_HAL_IMPL_RVV_NATIVE_LOADSTORE_MF2(vuint8mf2_t, uint8_t, u8, 8, 8)
-OPENCV_HAL_IMPL_RVV_NATIVE_LOADSTORE_MF2(vint8mf2_t, int8_t, i8, 8, 8)
-OPENCV_HAL_IMPL_RVV_NATIVE_LOADSTORE_MF2(vuint16mf2_t, uint16_t, u16, 16, 4)
-OPENCV_HAL_IMPL_RVV_NATIVE_LOADSTORE_MF2(vint16mf2_t, int16_t, i16, 16, 4)
-OPENCV_HAL_IMPL_RVV_NATIVE_LOADSTORE_MF2(vuint32mf2_t, uint32_t, u32, 32, 2)
-OPENCV_HAL_IMPL_RVV_NATIVE_LOADSTORE_MF2(vint32mf2_t, int32_t, i32, 32, 2)
-OPENCV_HAL_IMPL_RVV_NATIVE_LOADSTORE_MF2(vfloat32mf2_t, float32_t, f32, 32, 2)
-OPENCV_HAL_IMPL_RVV_NATIVE_LOADSTORE_MF2(vuint64mf2_t, uint64_t, u64, 64, 1)
-OPENCV_HAL_IMPL_RVV_NATIVE_LOADSTORE_MF2(vint64mf2_t, int64_t, i64, 64, 1)
-OPENCV_HAL_IMPL_RVV_NATIVE_LOADSTORE_MF2(vfloat64mf2_t, float64_t, f64, 64, 1)
-
-
-#define OPENCV_HAL_IMPL_RVV_NATIVE_WCVT(_Tpwvec, _Tpvec, _wTp, wcvt, suffix, width, n) \
-inline _Tpwvec wcvt (_Tpvec v, size_t vl) \
-{ \
-    _wTp tmp[n]; \
-    for (int i = 0; i < n; ++i) \
-    { \
-            tmp[i] = (_wTp)v.val[i]; \
-    } \
-    return vle##width##_v_##suffix##m1(tmp, vl); \
-}
-
-OPENCV_HAL_IMPL_RVV_NATIVE_WCVT(vuint16m1_t, vuint8mf2_t, ushort, vwcvtu_x_x_v_u16m1, u16, 16, 8)
-OPENCV_HAL_IMPL_RVV_NATIVE_WCVT(vint16m1_t, vint8mf2_t, short, vwcvt_x_x_v_i16m1, i16, 16, 8)
-OPENCV_HAL_IMPL_RVV_NATIVE_WCVT(vuint32m1_t, vuint16mf2_t, unsigned, vwcvtu_x_x_v_u32m1, u32, 32, 4)
-OPENCV_HAL_IMPL_RVV_NATIVE_WCVT(vint32m1_t, vint16mf2_t, int, vwcvt_x_x_v_i32m1, i32, 32, 4)
-OPENCV_HAL_IMPL_RVV_NATIVE_WCVT(vuint64m1_t, vuint32mf2_t, uint64, vwcvtu_x_x_v_u64m1, u64, 64, 2)
-OPENCV_HAL_IMPL_RVV_NATIVE_WCVT(vint64m1_t, vint32mf2_t, int64, vwcvt_x_x_v_i64m1, i64, 64, 2)
-
-inline vuint8mf4_t vle8_v_u8mf4 (const uint8_t *base, size_t vl)
-{
-    CV_UNUSED(vl);
-    return vuint8mf4_t(base);
-}
-inline vint8mf4_t vle8_v_i8mf4 (const int8_t *base, size_t vl)
-{
-    CV_UNUSED(vl);
-    return vint8mf4_t(base);
-}
-
-inline vuint16mf2_t vwcvtu_x_x_v_u16mf2 (vuint8mf4_t src, size_t vl)
-{
-    ushort tmp[4];
-    for (int i = 0; i < 4; ++i)
-    {
-            tmp[i] = (ushort)src.val[i];
-    }
-    return vle16_v_u16mf2(tmp, vl);
-}
-inline vint16mf2_t vwcvt_x_x_v_i16mf2 (vint8mf4_t src, size_t vl)
-{
-    short tmp[4];
-    for (int i = 0; i < 4; ++i)
-    {
-            tmp[i] = (short)src.val[i];
-    }
-    return vle16_v_i16mf2(tmp, vl);
-}
-#endif
-
-//////////// Types ////////////
-
-struct v_uint8x16
-{
-    typedef uchar lane_type;
-    enum { nlanes = 16 };
-
-    v_uint8x16() {}
-    explicit v_uint8x16(vuint8m1_t v)
-    {
-        vse8_v_u8m1(val, v, nlanes);
-    }
-    v_uint8x16(uchar v0, uchar v1, uchar v2, uchar v3, uchar v4, uchar v5, uchar v6, uchar v7,
-               uchar v8, uchar v9, uchar v10, uchar v11, uchar v12, uchar v13, uchar v14, uchar v15)
-    {
-        uchar v[] = {v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15};
-        for (int i = 0; i < nlanes; ++i)
-        {
-            val[i] = v[i];
-        }
-    }
-    operator vuint8m1_t() const
-    {
-        return vle8_v_u8m1(val, nlanes);
-    }
-    uchar get0() const
-    {
-        return val[0];
-    }
-
-    uchar val[16];
-};
-
-struct v_int8x16
-{
-    typedef schar lane_type;
-    enum { nlanes = 16 };
-
-    v_int8x16() {}
-    explicit v_int8x16(vint8m1_t v)
-    {
-        vse8_v_i8m1(val, v, nlanes);
-    }
-    v_int8x16(schar v0, schar v1, schar v2, schar v3, schar v4, schar v5, schar v6, schar v7,
-               schar v8, schar v9, schar v10, schar v11, schar v12, schar v13, schar v14, schar v15)
-    {
-        schar v[] = {v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15};
-        for (int i = 0; i < nlanes; ++i)
-        {
-            val[i] = v[i];
-        }
-    }
-    operator vint8m1_t() const
-    {
-        return vle8_v_i8m1(val, nlanes);
-    }
-    schar get0() const
-    {
-        return val[0];
-    }
-
-    schar val[16];
-};
-
-struct v_uint16x8
-{
-    typedef ushort lane_type;
-    enum { nlanes = 8 };
-
-    v_uint16x8() {}
-    explicit v_uint16x8(vuint16m1_t v)
-    {
-        vse16_v_u16m1(val, v, nlanes);
-    }
-    v_uint16x8(ushort v0, ushort v1, ushort v2, ushort v3, ushort v4, ushort v5, ushort v6, ushort v7)
-    {
-        ushort v[] = {v0, v1, v2, v3, v4, v5, v6, v7};
-        for (int i = 0; i < nlanes; ++i)
-        {
-            val[i] = v[i];
-        }
-    }
-    operator vuint16m1_t() const
-    {
-        return vle16_v_u16m1(val, nlanes);
-    }
-    ushort get0() const
-    {
-        return val[0];
-    }
-
-    ushort val[8];
-};
-
-struct v_int16x8
-{
-    typedef short lane_type;
-    enum { nlanes = 8 };
-
-    v_int16x8() {}
-    explicit v_int16x8(vint16m1_t v)
-    {
-        vse16_v_i16m1(val, v, nlanes);
-    }
-    v_int16x8(short v0, short v1, short v2, short v3, short v4, short v5, short v6, short v7)
-    {
-        short v[] = {v0, v1, v2, v3, v4, v5, v6, v7};
-        for (int i = 0; i < nlanes; ++i)
-        {
-            val[i] = v[i];
-        }
-    }
-    operator vint16m1_t() const
-    {
-        return vle16_v_i16m1(val, nlanes);
-    }
-    short get0() const
-    {
-        return val[0];
-    }
-
-    short val[8];
-};
-
-struct v_uint32x4
-{
-    typedef unsigned lane_type;
-    enum { nlanes = 4 };
-
-    v_uint32x4() {}
-    explicit v_uint32x4(vuint32m1_t v)
-    {
-        vse32_v_u32m1(val, v, nlanes);
-    }
-    v_uint32x4(unsigned v0, unsigned v1, unsigned v2, unsigned v3)
-    {
-        unsigned v[] = {v0, v1, v2, v3};
-        for (int i = 0; i < nlanes; ++i)
-        {
-            val[i] = v[i];
-        }
-    }
-    operator vuint32m1_t() const
-    {
-        return vle32_v_u32m1(val, nlanes);
-    }
-    unsigned get0() const
-    {
-        return val[0];
-    }
-
-    unsigned val[4];
-};
-
-struct v_int32x4
-{
-    typedef int lane_type;
-    enum { nlanes = 4 };
-
-    v_int32x4() {}
-    explicit v_int32x4(vint32m1_t v)
-    {
-        vse32_v_i32m1(val, v, nlanes);
-    }
-    v_int32x4(int v0, int v1, int v2, int v3)
-    {
-        int v[] = {v0, v1, v2, v3};
-        for (int i = 0; i < nlanes; ++i)
-        {
-            val[i] = v[i];
-        }
-    }
-    operator vint32m1_t() const
-    {
-        return vle32_v_i32m1(val, nlanes);
-    }
-    int get0() const
-    {
-        return val[0];
-    }
-    int val[4];
-};
-
-struct v_float32x4
-{
-    typedef float lane_type;
-    enum { nlanes = 4 };
-
-    v_float32x4() {}
-    explicit v_float32x4(vfloat32m1_t v)
-    {
-        vse32_v_f32m1(val, v, nlanes);
-    }
-    v_float32x4(float v0, float v1, float v2, float v3)
-    {
-        float v[] = {v0, v1, v2, v3};
-        for (int i = 0; i < nlanes; ++i)
-        {
-            val[i] = v[i];
-        }
-    }
-    operator vfloat32m1_t() const
-    {
-        return vle32_v_f32m1(val, nlanes);
-    }
-    float get0() const
-    {
-        return val[0];
-    }
-    float val[4];
-};
-
-struct v_uint64x2
-{
-    typedef uint64 lane_type;
-    enum { nlanes = 2 };
-
-    v_uint64x2() {}
-    explicit v_uint64x2(vuint64m1_t v)
-    {
-        vse64_v_u64m1(val, v, nlanes);
-    }
-    v_uint64x2(uint64 v0, uint64 v1)
-    {
-        uint64 v[] = {v0, v1};
-        for (int i = 0; i < nlanes; ++i)
-        {
-            val[i] = v[i];
-        }
-    }
-    operator vuint64m1_t() const
-    {
-        return vle64_v_u64m1(val, nlanes);
-    }
-    uint64 get0() const
-    {
-        return val[0];
-    }
-
-    uint64 val[2];
-};
-
-struct v_int64x2
-{
-    typedef int64 lane_type;
-    enum { nlanes = 2 };
-
-    v_int64x2() {}
-    explicit v_int64x2(vint64m1_t v)
-    {
-        vse64_v_i64m1(val, v, nlanes);
-    }
-    v_int64x2(int64 v0, int64 v1)
-    {
-        int64 v[] = {v0, v1};
-        for (int i = 0; i < nlanes; ++i)
-        {
-            val[i] = v[i];
-        }
-    }
-    operator vint64m1_t() const
-    {
-        return vle64_v_i64m1(val, nlanes);
-    }
-    int64 get0() const
-    {
-        return val[0];
-    }
-
-    int64 val[2];
-};
-
-#if CV_SIMD128_64F
-struct v_float64x2
-{
-    typedef double lane_type;
-    enum { nlanes = 2 };
-
-    v_float64x2() {}
-    explicit v_float64x2(vfloat64m1_t v)
-    {
-        vse64_v_f64m1(val, v, nlanes);
-    }
-    v_float64x2(double v0, double v1)
-    {
-        double v[] = {v0, v1};
-        for (int i = 0; i < nlanes; ++i)
-        {
-            val[i] = v[i];
-        }
-    }
-    operator vfloat64m1_t() const
-    {
-        return vle64_v_f64m1(val, nlanes);
-    }
-    double get0() const
-    {
-        return val[0];
-    }
-
-    double val[2];
-};
-#endif
-
-
-//////////// Initial ////////////
-
-#define OPENCV_HAL_IMPL_RVV_INIT_INTEGER(_Tpvec, _Tp, suffix1, suffix2, vl) \
-inline v_##_Tpvec v_setzero_##suffix1() \
-{ \
-    return v_##_Tpvec(vmv_v_x_##suffix2##m1(0, vl)); \
-} \
-inline v_##_Tpvec v_setall_##suffix1(_Tp v) \
-{ \
-    return v_##_Tpvec(vmv_v_x_##suffix2##m1(v, vl)); \
-}
-
-OPENCV_HAL_IMPL_RVV_INIT_INTEGER(uint8x16, uchar, u8, u8, 16)
-OPENCV_HAL_IMPL_RVV_INIT_INTEGER(int8x16, schar, s8, i8, 16)
-OPENCV_HAL_IMPL_RVV_INIT_INTEGER(uint16x8, ushort, u16, u16, 8)
-OPENCV_HAL_IMPL_RVV_INIT_INTEGER(int16x8, short, s16, i16, 8)
-OPENCV_HAL_IMPL_RVV_INIT_INTEGER(uint32x4, unsigned, u32, u32, 4)
-OPENCV_HAL_IMPL_RVV_INIT_INTEGER(int32x4, int, s32, i32, 4)
-OPENCV_HAL_IMPL_RVV_INIT_INTEGER(uint64x2, uint64, u64, u64, 2)
-OPENCV_HAL_IMPL_RVV_INIT_INTEGER(int64x2, int64, s64, i64, 2)
-
-#define OPENCV_HAL_IMPL_RVV_INIT_FP(_Tpv, _Tp, suffix, vl) \
-inline v_##_Tpv v_setzero_##suffix() \
-{ \
-    return v_##_Tpv(vfmv_v_f_##suffix##m1(0, vl)); \
-} \
-inline v_##_Tpv v_setall_##suffix(_Tp v) \
-{ \
-    return v_##_Tpv(vfmv_v_f_##suffix##m1(v, vl)); \
-}
-
-OPENCV_HAL_IMPL_RVV_INIT_FP(float32x4, float, f32, 4)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_INIT_FP(float64x2, double, f64, 2)
-#endif
-
-//////////// Reinterpret ////////////
-
-#define OPENCV_HAL_IMPL_RVV_SELF_REINTERPRET(_Tpvec, suffix) \
-inline v_##_Tpvec v_reinterpret_as_##suffix(const v_##_Tpvec& v) { return v; }
-
-OPENCV_HAL_IMPL_RVV_SELF_REINTERPRET(uint8x16, u8)
-OPENCV_HAL_IMPL_RVV_SELF_REINTERPRET(int8x16, s8)
-OPENCV_HAL_IMPL_RVV_SELF_REINTERPRET(uint16x8, u16)
-OPENCV_HAL_IMPL_RVV_SELF_REINTERPRET(int16x8, s16)
-OPENCV_HAL_IMPL_RVV_SELF_REINTERPRET(uint32x4, u32)
-OPENCV_HAL_IMPL_RVV_SELF_REINTERPRET(int32x4, s32)
-OPENCV_HAL_IMPL_RVV_SELF_REINTERPRET(float32x4, f32)
-OPENCV_HAL_IMPL_RVV_SELF_REINTERPRET(uint64x2, u64)
-OPENCV_HAL_IMPL_RVV_SELF_REINTERPRET(int64x2, s64)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_SELF_REINTERPRET(float64x2, f64)
-#endif
-
-#define OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(_Tpvec1, _Tpvec2, suffix1, suffix2, nsuffix1, nsuffix2) \
-inline v_##_Tpvec1 v_reinterpret_as_##suffix1(const v_##_Tpvec2& v) \
-{ \
-    return v_##_Tpvec1(vreinterpret_v_##nsuffix2##m1_##nsuffix1##m1(v));\
-} \
-inline v_##_Tpvec2 v_reinterpret_as_##suffix2(const v_##_Tpvec1& v) \
-{ \
-    return v_##_Tpvec2(vreinterpret_v_##nsuffix1##m1_##nsuffix2##m1(v));\
-}
-
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(uint8x16, int8x16, u8, s8, u8, i8)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(uint16x8, int16x8, u16, s16, u16, i16)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(uint32x4, int32x4, u32, s32, u32, i32)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(uint32x4, float32x4, u32, f32, u32, f32)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(int32x4, float32x4, s32, f32, i32, f32)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(uint64x2, int64x2, u64, s64, u64, i64)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(uint64x2, float64x2, u64, f64, u64, f64)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(int64x2, float64x2, s64, f64, i64, f64)
-#endif
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(uint8x16, uint16x8, u8, u16, u8, u16)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(uint8x16, uint32x4, u8, u32, u8, u32)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(uint8x16, uint64x2, u8, u64, u8, u64)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(uint16x8, uint32x4, u16, u32, u16, u32)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(uint16x8, uint64x2, u16, u64, u16, u64)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(uint32x4, uint64x2, u32, u64, u32, u64)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(int8x16, int16x8, s8, s16, i8, i16)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(int8x16, int32x4, s8, s32, i8, i32)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(int8x16, int64x2, s8, s64, i8, i64)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(int16x8, int32x4, s16, s32, i16, i32)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(int16x8, int64x2, s16, s64, i16, i64)
-OPENCV_HAL_IMPL_RVV_NATIVE_REINTERPRET(int32x4, int64x2, s32, s64, i32, i64)
-
-
-#define OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(_Tpvec1, _Tpvec2, suffix1, suffix2, nsuffix1, nsuffix2, width1, width2) \
-inline v_##_Tpvec1 v_reinterpret_as_##suffix1(const v_##_Tpvec2& v) \
-{ \
-    return v_##_Tpvec1(vreinterpret_v_##nsuffix1##width2##m1_##nsuffix1##width1##m1(vreinterpret_v_##nsuffix2##width2##m1_##nsuffix1##width2##m1(v)));\
-} \
-inline v_##_Tpvec2 v_reinterpret_as_##suffix2(const v_##_Tpvec1& v) \
-{ \
-    return v_##_Tpvec2(vreinterpret_v_##nsuffix1##width2##m1_##nsuffix2##width2##m1(vreinterpret_v_##nsuffix1##width1##m1_##nsuffix1##width2##m1(v)));\
-}
-
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint8x16, int16x8, u8, s16, u, i, 8, 16)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint8x16, int32x4, u8, s32, u, i, 8, 32)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint8x16, int64x2, u8, s64, u, i, 8, 64)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint16x8, int8x16, u16, s8, u, i, 16, 8)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint16x8, int32x4, u16, s32, u, i, 16, 32)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint16x8, int64x2, u16, s64, u, i, 16, 64)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint32x4, int8x16, u32, s8, u, i, 32, 8)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint32x4, int16x8, u32, s16, u, i, 32, 16)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint32x4, int64x2, u32, s64, u, i, 32, 64)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint64x2, int8x16, u64, s8, u, i, 64, 8)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint64x2, int16x8, u64, s16, u, i, 64, 16)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint64x2, int32x4, u64, s32, u, i, 64, 32)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint8x16, float32x4, u8, f32, u, f, 8, 32)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint16x8, float32x4, u16, f32, u, f, 16, 32)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint64x2, float32x4, u64, f32, u, f, 64, 32)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(int8x16, float32x4, s8, f32, i, f, 8, 32)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(int16x8, float32x4, s16, f32, i, f, 16, 32)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(int64x2, float32x4, s64, f32, i, f, 64, 32)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint8x16, float64x2, u8, f64, u, f, 8, 64)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint16x8, float64x2, u16, f64, u, f, 16, 64)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint32x4, float64x2, u32, f64, u, f, 32, 64)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(int8x16, float64x2, s8, f64, i, f, 8, 64)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(int16x8, float64x2, s16, f64, i, f, 16, 64)
-OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(int32x4, float64x2, s32, f64, i, f, 32, 64)
-
-// Three times reinterpret
-inline v_float32x4 v_reinterpret_as_f32(const v_float64x2& v) \
-{ \
-    return v_float32x4(vreinterpret_v_u32m1_f32m1(vreinterpret_v_u64m1_u32m1(vreinterpret_v_f64m1_u64m1(v))));\
-} \
-inline v_float64x2 v_reinterpret_as_f64(const v_float32x4& v) \
-{ \
-    return v_float64x2(vreinterpret_v_u64m1_f64m1(vreinterpret_v_u32m1_u64m1(vreinterpret_v_f32m1_u32m1(v))));\
-}
-
-////////////// Extract //////////////
-
-#define OPENCV_HAL_IMPL_RVV_EXTRACT_INTEGER(_Tpvec, _Tp, suffix, vmv, vl) \
-template <int s> \
-inline _Tpvec v_extract(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(vslideup_vx_##suffix##m1(vslidedown_vx_##suffix##m1(vmv_v_x_##suffix##m1(0, vl), a, s, vl), b, _Tpvec::nlanes - s, vl)); \
-} \
-template<int i> inline _Tp v_extract_n(_Tpvec v) \
-{ \
-    return _Tp(vmv(vslidedown_vx_##suffix##m1(vmv_v_x_##suffix##m1(0, vl), v, i, vl))); \
-}
-
-
-OPENCV_HAL_IMPL_RVV_EXTRACT_INTEGER(v_uint8x16, uchar, u8, vmv_x_s_u8m1_u8, 16)
-OPENCV_HAL_IMPL_RVV_EXTRACT_INTEGER(v_int8x16, schar, i8, vmv_x_s_i8m1_i8, 16)
-OPENCV_HAL_IMPL_RVV_EXTRACT_INTEGER(v_uint16x8, ushort, u16, vmv_x_s_u16m1_u16, 8)
-OPENCV_HAL_IMPL_RVV_EXTRACT_INTEGER(v_int16x8, short, i16, vmv_x_s_i16m1_i16, 8)
-OPENCV_HAL_IMPL_RVV_EXTRACT_INTEGER(v_uint32x4, uint, u32, vmv_x_s_u32m1_u32, 4)
-OPENCV_HAL_IMPL_RVV_EXTRACT_INTEGER(v_int32x4, int, i32, vmv_x_s_i32m1_i32, 4)
-OPENCV_HAL_IMPL_RVV_EXTRACT_INTEGER(v_uint64x2, uint64, u64, vmv_x_s_u64m1_u64, 2)
-OPENCV_HAL_IMPL_RVV_EXTRACT_INTEGER(v_int64x2, int64, i64, vmv_x_s_i64m1_i64, 2)
-
-#define OPENCV_HAL_IMPL_RVV_EXTRACT_FP(_Tpvec, _Tp, suffix, vmv, vl) \
-template <int s> \
-inline _Tpvec v_extract(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(vslideup_vx_##suffix##m1(vslidedown_vx_##suffix##m1(vfmv_v_f_##suffix##m1(0, vl), a, s, vl), b, _Tpvec::nlanes - s, vl)); \
-} \
-template<int i> inline _Tp v_extract_n(_Tpvec v) \
-{ \
-    return _Tp(vmv(vslidedown_vx_##suffix##m1(vfmv_v_f_##suffix##m1(0, vl), v, i, vl))); \
-}
-
-OPENCV_HAL_IMPL_RVV_EXTRACT_FP(v_float32x4, float, f32, vfmv_f_s_f32m1_f32, 4)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_EXTRACT_FP(v_float64x2, double, f64, vfmv_f_s_f64m1_f64, 2)
-#endif
-
-////////////// Load/Store //////////////
-
-#define OPENCV_HAL_IMPL_RVV_LOADSTORE_OP(_Tpvec, _nTpvec, _Tp, hvl, vl, width, suffix, vmv) \
-inline _Tpvec v_load(const _Tp* ptr) \
-{ \
-    return _Tpvec(vle##width##_v_##suffix##m1(ptr, vl)); \
-} \
-inline _Tpvec v_load_aligned(const _Tp* ptr) \
-{ \
-    return _Tpvec(vle##width##_v_##suffix##m1(ptr, vl)); \
-} \
-inline _Tpvec v_load_low(const _Tp* ptr) \
-{ \
-    _Tpvec res = _Tpvec(vle##width##_v_##suffix##m1(ptr, hvl)); \
-    return res; \
-} \
-inline void v_store(_Tp* ptr, const _Tpvec& a) \
-{ \
-    vse##width##_v_##suffix##m1(ptr, a, vl); \
-} \
-inline void v_store_aligned(_Tp* ptr, const _Tpvec& a) \
-{ \
-    vse##width##_v_##suffix##m1(ptr, a, vl); \
-} \
-inline void v_store_aligned_nocache(_Tp* ptr, const _Tpvec& a) \
-{ \
-    vse##width##_v_##suffix##m1(ptr, a, vl); \
-} \
-inline void v_store(_Tp* ptr, const _Tpvec& a, hal::StoreMode /*mode*/) \
-{ \
-    vse##width##_v_##suffix##m1(ptr, a, vl); \
-} \
-inline void v_store_low(_Tp* ptr, const _Tpvec& a) \
-{ \
-    vse##width##_v_##suffix##m1(ptr, a, hvl); \
-} \
-inline void v_store_high(_Tp* ptr, const _Tpvec& a) \
-{ \
-    vse##width##_v_##suffix##m1(ptr, vslidedown_vx_##suffix##m1(vmv(0, vl), a, hvl, vl), hvl); \
-}
-
-OPENCV_HAL_IMPL_RVV_LOADSTORE_OP(v_uint8x16, vuint8m1_t, uchar, 8, 16, 8, u8, vmv_v_x_u8m1)
-OPENCV_HAL_IMPL_RVV_LOADSTORE_OP(v_int8x16, vint8m1_t, schar, 8, 16, 8, i8, vmv_v_x_i8m1)
-OPENCV_HAL_IMPL_RVV_LOADSTORE_OP(v_uint16x8, vuint16m1_t, ushort, 4, 8, 16, u16, vmv_v_x_u16m1)
-OPENCV_HAL_IMPL_RVV_LOADSTORE_OP(v_int16x8, vint16m1_t, short, 4, 8, 16, i16, vmv_v_x_i16m1)
-OPENCV_HAL_IMPL_RVV_LOADSTORE_OP(v_uint32x4, vuint32m1_t, unsigned, 2, 4, 32, u32, vmv_v_x_u32m1)
-OPENCV_HAL_IMPL_RVV_LOADSTORE_OP(v_int32x4, vint32m1_t, int, 2, 4, 32, i32, vmv_v_x_i32m1)
-OPENCV_HAL_IMPL_RVV_LOADSTORE_OP(v_uint64x2, vuint64m1_t, uint64, 1, 2, 64, u64, vmv_v_x_u64m1)
-OPENCV_HAL_IMPL_RVV_LOADSTORE_OP(v_int64x2, vint64m1_t, int64, 1, 2, 64, i64, vmv_v_x_i64m1)
-OPENCV_HAL_IMPL_RVV_LOADSTORE_OP(v_float32x4, vfloat32m1_t, float, 2, 4, 32, f32, vfmv_v_f_f32m1)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_LOADSTORE_OP(v_float64x2, vfloat64m1_t, double, 1, 2, 64, f64, vfmv_v_f_f64m1)
-#endif
-
-inline v_int8x16 v_load_halves(const schar* ptr0, const schar* ptr1)
-{
-    schar elems[16] =
-    {
-        ptr0[0], ptr0[1], ptr0[2], ptr0[3], ptr0[4], ptr0[5], ptr0[6], ptr0[7],
-        ptr1[0], ptr1[1], ptr1[2], ptr1[3], ptr1[4], ptr1[5], ptr1[6], ptr1[7]
-    };
-    return v_int8x16(vle8_v_i8m1(elems, 16));
-}
-inline v_uint8x16 v_load_halves(const uchar* ptr0, const uchar* ptr1) { return v_reinterpret_as_u8(v_load_halves((schar*)ptr0, (schar*)ptr1)); }
-
-inline v_int16x8 v_load_halves(const short* ptr0, const short* ptr1)
-{
-    short elems[8] =
-    {
-        ptr0[0], ptr0[1], ptr0[2], ptr0[3], ptr1[0], ptr1[1], ptr1[2], ptr1[3]
-    };
-    return v_int16x8(vle16_v_i16m1(elems, 8));
-}
-inline v_uint16x8 v_load_halves(const ushort* ptr0, const ushort* ptr1) { return v_reinterpret_as_u16(v_load_halves((short*)ptr0, (short*)ptr1)); }
-
-inline v_int32x4 v_load_halves(const int* ptr0, const int* ptr1)
-{
-    int elems[4] =
-    {
-        ptr0[0], ptr0[1], ptr1[0], ptr1[1]
-    };
-    return v_int32x4(vle32_v_i32m1(elems, 4));
-}
-inline v_float32x4 v_load_halves(const float* ptr0, const float* ptr1)
-{
-    float elems[4] =
-    {
-        ptr0[0], ptr0[1], ptr1[0], ptr1[1]
-    };
-    return v_float32x4(vle32_v_f32m1(elems, 4));
-}
-inline v_uint32x4 v_load_halves(const unsigned* ptr0, const unsigned* ptr1) { return v_reinterpret_as_u32(v_load_halves((int*)ptr0, (int*)ptr1)); }
-
-inline v_int64x2 v_load_halves(const int64* ptr0, const int64* ptr1)
-{
-    int64 elems[2] =
-    {
-        ptr0[0], ptr1[0]
-    };
-    return v_int64x2(vle64_v_i64m1(elems, 2));
-}
-inline v_uint64x2 v_load_halves(const uint64* ptr0, const uint64* ptr1) { return v_reinterpret_as_u64(v_load_halves((int64*)ptr0, (int64*)ptr1)); }
-
-#if CV_SIMD128_64F
-inline v_float64x2 v_load_halves(const double* ptr0, const double* ptr1)
-{
-    double elems[2] =
-    {
-        ptr0[0], ptr1[0]
-    };
-    return v_float64x2(vle64_v_f64m1(elems, 2));
-}
-#endif
-
-
-////////////// Lookup table access ////////////////////
-
-inline v_int8x16 v_lut(const schar* tab, const int* idx)
-{
-    schar elems[16] =
-    {
-        tab[idx[ 0]],
-        tab[idx[ 1]],
-        tab[idx[ 2]],
-        tab[idx[ 3]],
-        tab[idx[ 4]],
-        tab[idx[ 5]],
-        tab[idx[ 6]],
-        tab[idx[ 7]],
-        tab[idx[ 8]],
-        tab[idx[ 9]],
-        tab[idx[10]],
-        tab[idx[11]],
-        tab[idx[12]],
-        tab[idx[13]],
-        tab[idx[14]],
-        tab[idx[15]]
-    };
-    return v_int8x16(vle8_v_i8m1(elems, 16));
-}
-inline v_int8x16 v_lut_pairs(const schar* tab, const int* idx)
-{
-    schar elems[16] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[1]],
-        tab[idx[1] + 1],
-        tab[idx[2]],
-        tab[idx[2] + 1],
-        tab[idx[3]],
-        tab[idx[3] + 1],
-        tab[idx[4]],
-        tab[idx[4] + 1],
-        tab[idx[5]],
-        tab[idx[5] + 1],
-        tab[idx[6]],
-        tab[idx[6] + 1],
-        tab[idx[7]],
-        tab[idx[7] + 1]
-    };
-    return v_int8x16(vle8_v_i8m1(elems, 16));
-}
-inline v_int8x16 v_lut_quads(const schar* tab, const int* idx)
-{
-    schar elems[16] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[0] + 2],
-        tab[idx[0] + 3],
-        tab[idx[1]],
-        tab[idx[1] + 1],
-        tab[idx[1] + 2],
-        tab[idx[1] + 3],
-        tab[idx[2]],
-        tab[idx[2] + 1],
-        tab[idx[2] + 2],
-        tab[idx[2] + 3],
-        tab[idx[3]],
-        tab[idx[3] + 1],
-        tab[idx[3] + 2],
-        tab[idx[3] + 3]
-    };
-    return v_int8x16(vle8_v_i8m1(elems, 16));
-}
-inline v_uint8x16 v_lut(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut((schar*)tab, idx)); }
-inline v_uint8x16 v_lut_pairs(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut_pairs((schar*)tab, idx)); }
-inline v_uint8x16 v_lut_quads(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut_quads((schar*)tab, idx)); }
-
-inline v_int16x8 v_lut(const short* tab, const int* idx)
-{
-    short elems[8] =
-    {
-        tab[idx[0]],
-        tab[idx[1]],
-        tab[idx[2]],
-        tab[idx[3]],
-        tab[idx[4]],
-        tab[idx[5]],
-        tab[idx[6]],
-        tab[idx[7]]
-    };
-    return v_int16x8(vle16_v_i16m1(elems, 8));
-}
-inline v_int16x8 v_lut_pairs(const short* tab, const int* idx)
-{
-    short elems[8] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[1]],
-        tab[idx[1] + 1],
-        tab[idx[2]],
-        tab[idx[2] + 1],
-        tab[idx[3]],
-        tab[idx[3] + 1]
-    };
-    return v_int16x8(vle16_v_i16m1(elems, 8));
-}
-inline v_int16x8 v_lut_quads(const short* tab, const int* idx)
-{
-    short elems[8] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[0] + 2],
-        tab[idx[0] + 3],
-        tab[idx[1]],
-        tab[idx[1] + 1],
-        tab[idx[1] + 2],
-        tab[idx[1] + 3]
-    };
-    return v_int16x8(vle16_v_i16m1(elems, 8));
-}
-inline v_uint16x8 v_lut(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut((short*)tab, idx)); }
-inline v_uint16x8 v_lut_pairs(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut_pairs((short*)tab, idx)); }
-inline v_uint16x8 v_lut_quads(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut_quads((short*)tab, idx)); }
-
-inline v_int32x4 v_lut(const int* tab, const int* idx)
-{
-    int elems[4] =
-    {
-        tab[idx[0]],
-        tab[idx[1]],
-        tab[idx[2]],
-        tab[idx[3]]
-    };
-    return v_int32x4(vle32_v_i32m1(elems, 4));
-}
-inline v_int32x4 v_lut_pairs(const int* tab, const int* idx)
-{
-    int elems[4] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[1]],
-        tab[idx[1] + 1]
-    };
-    return v_int32x4(vle32_v_i32m1(elems, 4));
-}
-inline v_int32x4 v_lut_quads(const int* tab, const int* idx)
-{
-    return v_int32x4(vle32_v_i32m1(tab + idx[0], 4));
-}
-
-inline v_uint32x4 v_lut(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut((int*)tab, idx)); }
-inline v_uint32x4 v_lut_pairs(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut_pairs((int*)tab, idx)); }
-inline v_uint32x4 v_lut_quads(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut_quads((int*)tab, idx)); }
-
-inline v_int64x2 v_lut(const int64_t* tab, const int* idx)
-{
-    int64_t elems[2] =
-    {
-        tab[idx[0]],
-        tab[idx[1]]
-    };
-    return v_int64x2(vle64_v_i64m1(elems, 2));
-}
-inline v_int64x2 v_lut_pairs(const int64* tab, const int* idx)
-{
-    return v_int64x2(vle64_v_i64m1(tab + idx[0], 2));
-}
-inline v_uint64x2 v_lut(const uint64* tab, const int* idx) { return v_reinterpret_as_u64(v_lut((const int64_t *)tab, idx)); }
-inline v_uint64x2 v_lut_pairs(const uint64* tab, const int* idx) { return v_reinterpret_as_u64(v_lut_pairs((const int64_t *)tab, idx)); }
-
-inline v_float32x4 v_lut(const float* tab, const int* idx)
-{
-    float elems[4] =
-    {
-        tab[idx[0]],
-        tab[idx[1]],
-        tab[idx[2]],
-        tab[idx[3]]
-    };
-    return v_float32x4(vle32_v_f32m1(elems, 4));
-}
-inline v_float32x4 v_lut_pairs(const float* tab, const int* idx)
-{
-    float elems[4] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[1]],
-        tab[idx[1] + 1]
-    };
-    return v_float32x4(vle32_v_f32m1(elems, 4));
-}
-inline v_float32x4 v_lut_quads(const float* tab, const int* idx)
-{
-    return v_float32x4(vle32_v_f32m1(tab + idx[0], 4));
-}
-
-inline v_int32x4 v_lut(const int* tab, const v_int32x4& idxvec)
-{
-    int elems[4] =
-    {
-        tab[v_extract_n<0>(idxvec)],
-        tab[v_extract_n<1>(idxvec)],
-        tab[v_extract_n<2>(idxvec)],
-        tab[v_extract_n<3>(idxvec)]
-    };
-    return v_int32x4(vle32_v_i32m1(elems, 4));
-}
-
-inline v_uint32x4 v_lut(const unsigned* tab, const v_int32x4& idxvec)
-{
-    unsigned elems[4] =
-    {
-        tab[v_extract_n<0>(idxvec)],
-        tab[v_extract_n<1>(idxvec)],
-        tab[v_extract_n<2>(idxvec)],
-        tab[v_extract_n<3>(idxvec)]
-    };
-    return v_uint32x4(vle32_v_u32m1(elems, 4));
-}
-
-inline v_float32x4 v_lut(const float* tab, const v_int32x4& idxvec)
-{
-    float elems[4] =
-    {
-        tab[v_extract_n<0>(idxvec)],
-        tab[v_extract_n<1>(idxvec)],
-        tab[v_extract_n<2>(idxvec)],
-        tab[v_extract_n<3>(idxvec)]
-    };
-    return v_float32x4(vle32_v_f32m1(elems, 4));
-}
-
-inline void v_lut_deinterleave(const float* tab, const v_int32x4& idxvec, v_float32x4& x, v_float32x4& y)
-{
-    int idx[4];
-    v_store_aligned(idx, idxvec);
-
-    x = v_float32x4(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]]);
-    y = v_float32x4(tab[idx[0]+1], tab[idx[1]+1], tab[idx[2]+1], tab[idx[3]+1]);
-}
-
-#if CV_SIMD128_64F
-inline v_float64x2 v_lut(const double* tab, const int* idx)
-{
-    double elems[2] =
-    {
-        tab[idx[0]],
-        tab[idx[1]]
-    };
-    return v_float64x2(vle64_v_f64m1(elems, 2));
-}
-
-inline v_float64x2 v_lut_pairs(const double* tab, const int* idx)
-{
-    return v_float64x2(vle64_v_f64m1(tab + idx[0], 2));
-}
-
-inline v_float64x2 v_lut(const double* tab, const v_int32x4& idxvec)
-{
-    double elems[2] =
-    {
-        tab[v_extract_n<0>(idxvec)],
-        tab[v_extract_n<1>(idxvec)]
-    };
-    return v_float64x2(vle64_v_f64m1(elems, 2));
-}
-
-inline void v_lut_deinterleave(const double* tab, const v_int32x4& idxvec, v_float64x2& x, v_float64x2& y)
-{
-    int idx[4] = {0};
-    v_store_aligned(idx, idxvec);
-
-    x = v_float64x2(tab[idx[0]], tab[idx[1]]);
-    y = v_float64x2(tab[idx[0]+1], tab[idx[1]+1]);
-}
-#endif
-
-////////////// Pack boolean ////////////////////
-
-inline v_uint8x16 v_pack_b(const v_uint16x8& a, const v_uint16x8& b)
-{
-    ushort ptr[16] = {0};
-    v_store(ptr, a);
-    v_store(ptr + 8, b);
-    return v_uint8x16(vnsrl_wx_u8m1(vle16_v_u16m2(ptr, 16), 0, 16));
-}
-
-inline v_uint8x16 v_pack_b(const v_uint32x4& a, const v_uint32x4& b,
-                           const v_uint32x4& c, const v_uint32x4& d)
-{
-    unsigned ptr[16] = {0};
-    v_store(ptr, a);
-    v_store(ptr + 4, b);
-    v_store(ptr + 8, c);
-    v_store(ptr + 12, d);
-    return v_uint8x16(vnsrl_wx_u8m1(vnsrl_wx_u16m2(vle32_v_u32m4(ptr, 16), 0, 16), 0, 16));
-}
-
-inline v_uint8x16 v_pack_b(const v_uint64x2& a, const v_uint64x2& b, const v_uint64x2& c,
-                           const v_uint64x2& d, const v_uint64x2& e, const v_uint64x2& f,
-                           const v_uint64x2& g, const v_uint64x2& h)
-{
-    uint64 ptr[16] = {0};
-    v_store(ptr, a);
-    v_store(ptr + 2, b);
-    v_store(ptr + 4, c);
-    v_store(ptr + 6, d);
-    v_store(ptr + 8, e);
-    v_store(ptr + 10, f);
-    v_store(ptr + 12, g);
-    v_store(ptr + 14, h);
-    return v_uint8x16(vnsrl_wx_u8m1(vnsrl_wx_u16m2(vnsrl_wx_u32m4(vle64_v_u64m8(ptr, 16), 0, 16), 0, 16), 0, 16));
-}
-
-////////////// Arithmetics //////////////
-#define OPENCV_HAL_IMPL_RVV_BIN_OP(bin_op, _Tpvec, intrin, vl) \
-inline _Tpvec operator bin_op (const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(intrin(a, b, vl)); \
-} \
-inline _Tpvec& operator bin_op##= (_Tpvec& a, const _Tpvec& b) \
-{ \
-    a = _Tpvec(intrin(a, b, vl)); \
-    return a; \
-}
-
-OPENCV_HAL_IMPL_RVV_BIN_OP(+, v_uint8x16, vsaddu_vv_u8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_OP(-, v_uint8x16, vssubu_vv_u8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_OP(/, v_uint8x16, vdivu_vv_u8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_OP(+, v_int8x16, vsadd_vv_i8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_OP(-, v_int8x16, vssub_vv_i8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_OP(/, v_int8x16, vdiv_vv_i8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_OP(+, v_uint16x8, vsaddu_vv_u16m1, 8)
-OPENCV_HAL_IMPL_RVV_BIN_OP(-, v_uint16x8, vssubu_vv_u16m1, 8)
-OPENCV_HAL_IMPL_RVV_BIN_OP(/, v_uint16x8, vdivu_vv_u16m1, 8)
-OPENCV_HAL_IMPL_RVV_BIN_OP(+, v_int16x8, vsadd_vv_i16m1, 8)
-OPENCV_HAL_IMPL_RVV_BIN_OP(-, v_int16x8, vssub_vv_i16m1, 8)
-OPENCV_HAL_IMPL_RVV_BIN_OP(/, v_int16x8, vdiv_vv_i16m1, 8)
-OPENCV_HAL_IMPL_RVV_BIN_OP(+, v_uint32x4, vadd_vv_u32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_OP(-, v_uint32x4, vsub_vv_u32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_OP(*, v_uint32x4, vmul_vv_u32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_OP(/, v_uint32x4, vdivu_vv_u32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_OP(+, v_int32x4, vadd_vv_i32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_OP(-, v_int32x4, vsub_vv_i32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_OP(*, v_int32x4, vmul_vv_i32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_OP(/, v_int32x4, vdiv_vv_i32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_OP(+, v_float32x4, vfadd_vv_f32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_OP(-, v_float32x4, vfsub_vv_f32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_OP(*, v_float32x4, vfmul_vv_f32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_OP(/, v_float32x4, vfdiv_vv_f32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_OP(+, v_uint64x2, vadd_vv_u64m1, 2)
-OPENCV_HAL_IMPL_RVV_BIN_OP(-, v_uint64x2, vsub_vv_u64m1, 2)
-OPENCV_HAL_IMPL_RVV_BIN_OP(*, v_uint64x2, vmul_vv_u64m1, 2)
-OPENCV_HAL_IMPL_RVV_BIN_OP(/, v_uint64x2, vdivu_vv_u64m1, 2)
-OPENCV_HAL_IMPL_RVV_BIN_OP(+, v_int64x2, vadd_vv_i64m1, 2)
-OPENCV_HAL_IMPL_RVV_BIN_OP(-, v_int64x2, vsub_vv_i64m1, 2)
-OPENCV_HAL_IMPL_RVV_BIN_OP(*, v_int64x2, vmul_vv_i64m1, 2)
-OPENCV_HAL_IMPL_RVV_BIN_OP(/, v_int64x2, vdiv_vv_i64m1, 2)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_BIN_OP(+, v_float64x2, vfadd_vv_f64m1, 2)
-OPENCV_HAL_IMPL_RVV_BIN_OP(-, v_float64x2, vfsub_vv_f64m1, 2)
-OPENCV_HAL_IMPL_RVV_BIN_OP(*, v_float64x2, vfmul_vv_f64m1, 2)
-OPENCV_HAL_IMPL_RVV_BIN_OP(/, v_float64x2, vfdiv_vv_f64m1, 2)
-#endif
-
-
-////////////// Bitwise logic //////////////
-
-#define OPENCV_HAL_IMPL_RVV_LOGIC_OP(_Tpvec, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_BIN_OP(&, _Tpvec, vand_vv_##suffix##m1, vl) \
-OPENCV_HAL_IMPL_RVV_BIN_OP(|, _Tpvec, vor_vv_##suffix##m1, vl) \
-OPENCV_HAL_IMPL_RVV_BIN_OP(^, _Tpvec, vxor_vv_##suffix##m1, vl) \
-inline _Tpvec operator ~ (const _Tpvec& a) \
-{ \
-    return _Tpvec(vnot_v_##suffix##m1(a, vl)); \
-}
-
-OPENCV_HAL_IMPL_RVV_LOGIC_OP(v_uint8x16, u8, 16)
-OPENCV_HAL_IMPL_RVV_LOGIC_OP(v_int8x16, i8, 16)
-OPENCV_HAL_IMPL_RVV_LOGIC_OP(v_uint16x8, u16, 8)
-OPENCV_HAL_IMPL_RVV_LOGIC_OP(v_int16x8, i16, 8)
-OPENCV_HAL_IMPL_RVV_LOGIC_OP(v_uint32x4, u32, 4)
-OPENCV_HAL_IMPL_RVV_LOGIC_OP(v_int32x4, i32, 4)
-OPENCV_HAL_IMPL_RVV_LOGIC_OP(v_uint64x2, u64, 2)
-OPENCV_HAL_IMPL_RVV_LOGIC_OP(v_int64x2, i64, 2)
-
-#define OPENCV_HAL_IMPL_RVV_FLT_BIT_OP(bin_op, intrin) \
-inline v_float32x4 operator bin_op (const v_float32x4& a, const v_float32x4& b) \
-{ \
-    return v_float32x4(vreinterpret_v_i32m1_f32m1(intrin(vreinterpret_v_f32m1_i32m1(a), vreinterpret_v_f32m1_i32m1(b), 4))); \
-} \
-inline v_float32x4& operator bin_op##= (v_float32x4& a, const v_float32x4& b) \
-{ \
-    a = v_float32x4(vreinterpret_v_i32m1_f32m1(intrin(vreinterpret_v_f32m1_i32m1(a), vreinterpret_v_f32m1_i32m1(b), 4))); \
-    return a; \
-}
-
-OPENCV_HAL_IMPL_RVV_FLT_BIT_OP(&, vand_vv_i32m1)
-OPENCV_HAL_IMPL_RVV_FLT_BIT_OP(|, vor_vv_i32m1)
-OPENCV_HAL_IMPL_RVV_FLT_BIT_OP(^, vxor_vv_i32m1)
-
-inline v_float32x4 operator ~ (const v_float32x4& a)
-{
-    return v_float32x4(vreinterpret_v_i32m1_f32m1(vnot_v_i32m1(vreinterpret_v_f32m1_i32m1(a), 4)));
-}
-
-#if CV_SIMD128_64F
-#define OPENCV_HAL_IMPL_RVV_FLT64_BIT_OP(bin_op, intrin) \
-inline v_float64x2 operator bin_op (const v_float64x2& a, const v_float64x2& b) \
-{ \
-    return v_float64x2(vreinterpret_v_i64m1_f64m1(intrin(vreinterpret_v_f64m1_i64m1(a), vreinterpret_v_f64m1_i64m1(b), 2))); \
-} \
-inline v_float64x2& operator bin_op##= (v_float64x2& a, const v_float64x2& b) \
-{ \
-    a = v_float64x2(vreinterpret_v_i64m1_f64m1(intrin(vreinterpret_v_f64m1_i64m1(a), vreinterpret_v_f64m1_i64m1(b), 2))); \
-    return a; \
-}
-
-OPENCV_HAL_IMPL_RVV_FLT64_BIT_OP(&, vand_vv_i64m1)
-OPENCV_HAL_IMPL_RVV_FLT64_BIT_OP(|, vor_vv_i64m1)
-OPENCV_HAL_IMPL_RVV_FLT64_BIT_OP(^, vxor_vv_i64m1)
-
-inline v_float64x2 operator ~ (const v_float64x2& a)
-{
-    return v_float64x2(vreinterpret_v_i64m1_f64m1(vnot_v_i64m1(vreinterpret_v_f64m1_i64m1(a), 2)));
-}
-#endif
-
-////////////// Bitwise shifts //////////////
-
-#define OPENCV_HAL_IMPL_RVV_UNSIGNED_SHIFT_OP(_Tpvec, suffix, vl) \
-inline _Tpvec operator << (const _Tpvec& a, int n) \
-{ \
-    return _Tpvec(vsll_vx_##suffix##m1(a, uint8_t(n), vl)); \
-} \
-inline _Tpvec operator >> (const _Tpvec& a, int n) \
-{ \
-    return _Tpvec(vsrl_vx_##suffix##m1(a, uint8_t(n), vl)); \
-} \
-template<int n> inline _Tpvec v_shl(const _Tpvec& a) \
-{ \
-    return _Tpvec(vsll_vx_##suffix##m1(a, uint8_t(n), vl)); \
-} \
-template<int n> inline _Tpvec v_shr(const _Tpvec& a) \
-{ \
-    return _Tpvec(vsrl_vx_##suffix##m1(a, uint8_t(n), vl)); \
-}
-
-#define OPENCV_HAL_IMPL_RVV_SIGNED_SHIFT_OP(_Tpvec, suffix, vl) \
-inline _Tpvec operator << (const _Tpvec& a, int n) \
-{ \
-    return _Tpvec(vsll_vx_##suffix##m1(a, uint8_t(n), vl)); \
-} \
-inline _Tpvec operator >> (const _Tpvec& a, int n) \
-{ \
-    return _Tpvec(vsra_vx_##suffix##m1(a, uint8_t(n), vl)); \
-} \
-template<int n> inline _Tpvec v_shl(const _Tpvec& a) \
-{ \
-    return _Tpvec(vsll_vx_##suffix##m1(a, uint8_t(n), vl)); \
-} \
-template<int n> inline _Tpvec v_shr(const _Tpvec& a) \
-{ \
-    return _Tpvec(vsra_vx_##suffix##m1(a, uint8_t(n), vl)); \
-}
-
-OPENCV_HAL_IMPL_RVV_UNSIGNED_SHIFT_OP(v_uint8x16, u8, 16)
-OPENCV_HAL_IMPL_RVV_UNSIGNED_SHIFT_OP(v_uint16x8, u16, 8)
-OPENCV_HAL_IMPL_RVV_UNSIGNED_SHIFT_OP(v_uint32x4, u32, 4)
-OPENCV_HAL_IMPL_RVV_UNSIGNED_SHIFT_OP(v_uint64x2, u64, 2)
-OPENCV_HAL_IMPL_RVV_SIGNED_SHIFT_OP(v_int8x16, i8, 16)
-OPENCV_HAL_IMPL_RVV_SIGNED_SHIFT_OP(v_int16x8, i16, 8)
-OPENCV_HAL_IMPL_RVV_SIGNED_SHIFT_OP(v_int32x4, i32, 4)
-OPENCV_HAL_IMPL_RVV_SIGNED_SHIFT_OP(v_int64x2, i64, 2)
-
-
-////////////// Comparison //////////////
-
-#define OPENCV_HAL_IMPL_RVV_INT_CMP_OP(_Tpvec, op, intrin, suffix, vl) \
-inline _Tpvec operator op (const _Tpvec& a, const _Tpvec& b) \
-{ \
-    uint64_t ones = -1; \
-    return _Tpvec(vmerge_vxm_##suffix##m1(intrin(a, b, vl), vmv_v_x_##suffix##m1(0, vl), ones, vl)); \
-}
-
-#define OPENCV_HAL_IMPL_RVV_FLOAT_CMP_OP(_Tpvec, op, intrin, suffix, vl) \
-inline _Tpvec operator op (const _Tpvec& a, const _Tpvec& b) \
-{ \
-    union { uint64 u; double d; } ones; ones.u = -1; \
-    return _Tpvec(vfmerge_vfm_##suffix##m1(intrin(a, b, vl), vfmv_v_f_##suffix##m1(0, vl), ones.d, vl)); \
-}
-
-#define OPENCV_HAL_IMPL_RVV_UNSIGNED_CMP(_Tpvec, suffix, width, vl) \
-OPENCV_HAL_IMPL_RVV_INT_CMP_OP(_Tpvec, ==, vmseq_vv_##suffix##m1_b##width, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_INT_CMP_OP(_Tpvec, !=, vmsne_vv_##suffix##m1_b##width, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_INT_CMP_OP(_Tpvec, <, vmsltu_vv_##suffix##m1_b##width, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_INT_CMP_OP(_Tpvec, >, vmsgtu_vv_##suffix##m1_b##width, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_INT_CMP_OP(_Tpvec, <=, vmsleu_vv_##suffix##m1_b##width, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_INT_CMP_OP(_Tpvec, >=, vmsgeu_vv_##suffix##m1_b##width, suffix, vl)
-
-#define OPENCV_HAL_IMPL_RVV_SIGNED_CMP(_Tpvec, suffix, width, vl) \
-OPENCV_HAL_IMPL_RVV_INT_CMP_OP(_Tpvec, ==, vmseq_vv_##suffix##m1_b##width, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_INT_CMP_OP(_Tpvec, !=, vmsne_vv_##suffix##m1_b##width, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_INT_CMP_OP(_Tpvec, <, vmslt_vv_##suffix##m1_b##width, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_INT_CMP_OP(_Tpvec, >, vmsgt_vv_##suffix##m1_b##width, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_INT_CMP_OP(_Tpvec, <=, vmsle_vv_##suffix##m1_b##width, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_INT_CMP_OP(_Tpvec, >=, vmsge_vv_##suffix##m1_b##width, suffix, vl)
-
-#define OPENCV_HAL_IMPL_RVV_FLOAT_CMP(_Tpvec, suffix, width, vl) \
-OPENCV_HAL_IMPL_RVV_FLOAT_CMP_OP(_Tpvec, ==, vmfeq_vv_##suffix##m1_b##width, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_FLOAT_CMP_OP(_Tpvec, !=, vmfne_vv_##suffix##m1_b##width, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_FLOAT_CMP_OP(_Tpvec, <, vmflt_vv_##suffix##m1_b##width, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_FLOAT_CMP_OP(_Tpvec, >, vmfgt_vv_##suffix##m1_b##width, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_FLOAT_CMP_OP(_Tpvec, <=, vmfle_vv_##suffix##m1_b##width, suffix, vl) \
-OPENCV_HAL_IMPL_RVV_FLOAT_CMP_OP(_Tpvec, >=, vmfge_vv_##suffix##m1_b##width, suffix, vl)
-
-
-OPENCV_HAL_IMPL_RVV_UNSIGNED_CMP(v_uint8x16, u8, 8, 16)
-OPENCV_HAL_IMPL_RVV_UNSIGNED_CMP(v_uint16x8, u16, 16, 8)
-OPENCV_HAL_IMPL_RVV_UNSIGNED_CMP(v_uint32x4, u32, 32, 4)
-OPENCV_HAL_IMPL_RVV_UNSIGNED_CMP(v_uint64x2, u64, 64, 2)
-OPENCV_HAL_IMPL_RVV_SIGNED_CMP(v_int8x16, i8, 8, 16)
-OPENCV_HAL_IMPL_RVV_SIGNED_CMP(v_int16x8, i16, 16, 8)
-OPENCV_HAL_IMPL_RVV_SIGNED_CMP(v_int32x4, i32, 32, 4)
-OPENCV_HAL_IMPL_RVV_SIGNED_CMP(v_int64x2, i64, 64, 2)
-OPENCV_HAL_IMPL_RVV_FLOAT_CMP(v_float32x4, f32, 32, 4)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_FLOAT_CMP(v_float64x2, f64, 64, 2)
-#endif
-
-inline v_float32x4 v_not_nan(const v_float32x4& a)
-{ return a == a; }
-
-#if CV_SIMD128_64F
-inline v_float64x2 v_not_nan(const v_float64x2& a)
-{ return a == a; }
-#endif
-
-////////////// Min/Max //////////////
-
-#define OPENCV_HAL_IMPL_RVV_BIN_FUNC(_Tpvec, func, intrin, vl) \
-inline _Tpvec func(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(intrin(a, b, vl)); \
-}
-
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_uint8x16, v_min, vminu_vv_u8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_uint8x16, v_max, vmaxu_vv_u8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_int8x16, v_min, vmin_vv_i8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_int8x16, v_max, vmax_vv_i8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_uint16x8, v_min, vminu_vv_u16m1, 8)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_uint16x8, v_max, vmaxu_vv_u16m1, 8)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_int16x8, v_min, vmin_vv_i16m1, 8)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_int16x8, v_max, vmax_vv_i16m1, 8)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_uint32x4, v_min, vminu_vv_u32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_uint32x4, v_max, vmaxu_vv_u32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_int32x4, v_min, vmin_vv_i32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_int32x4, v_max, vmax_vv_i32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_float32x4, v_min, vfmin_vv_f32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_float32x4, v_max, vfmax_vv_f32m1, 4)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_uint64x2, v_min, vminu_vv_u64m1, 2)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_uint64x2, v_max, vmaxu_vv_u64m1, 2)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_int64x2, v_min, vmin_vv_i64m1, 2)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_int64x2, v_max, vmax_vv_i64m1, 2)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_float64x2, v_min, vfmin_vv_f64m1, 2)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_float64x2, v_max, vfmax_vv_f64m1, 2)
-#endif
-
-////////////// Arithmetics wrap //////////////
-
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_uint8x16, v_add_wrap, vadd_vv_u8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_int8x16, v_add_wrap, vadd_vv_i8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_uint16x8, v_add_wrap, vadd_vv_u16m1, 8)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_int16x8, v_add_wrap, vadd_vv_i16m1, 8)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_uint8x16, v_sub_wrap, vsub_vv_u8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_int8x16, v_sub_wrap, vsub_vv_i8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_uint16x8, v_sub_wrap, vsub_vv_u16m1, 8)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_int16x8, v_sub_wrap, vsub_vv_i16m1, 8)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_uint8x16, v_mul_wrap, vmul_vv_u8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_int8x16, v_mul_wrap, vmul_vv_i8m1, 16)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_uint16x8, v_mul_wrap, vmul_vv_u16m1, 8)
-OPENCV_HAL_IMPL_RVV_BIN_FUNC(v_int16x8, v_mul_wrap, vmul_vv_i16m1, 8)
-
-////////////// Reduce //////////////
-
-#define OPENCV_HAL_IMPL_RVV_REDUCE_SUM(_Tpvec, _wTpvec, _nwTpvec, scalartype, suffix, wsuffix, vl, red) \
-inline scalartype v_reduce_sum(const _Tpvec& a)  \
-{ \
-    _nwTpvec zero = vmv_v_x_##wsuffix##m1(0, vl); \
-    _nwTpvec res = vmv_v_x_##wsuffix##m1(0, vl); \
-    res = v##red##_vs_##suffix##m1_##wsuffix##m1(res, a, zero, vl); \
-    return (scalartype)(_wTpvec(res).get0()); \
-}
-
-OPENCV_HAL_IMPL_RVV_REDUCE_SUM(v_uint8x16, v_uint16x8, vuint16m1_t, unsigned, u8, u16, 16, wredsumu)
-OPENCV_HAL_IMPL_RVV_REDUCE_SUM(v_int8x16, v_int16x8, vint16m1_t, int, i8, i16, 16, wredsum)
-OPENCV_HAL_IMPL_RVV_REDUCE_SUM(v_uint16x8, v_uint32x4, vuint32m1_t, unsigned, u16, u32, 8, wredsumu)
-OPENCV_HAL_IMPL_RVV_REDUCE_SUM(v_int16x8, v_int32x4, vint32m1_t, int, i16, i32, 8, wredsum)
-OPENCV_HAL_IMPL_RVV_REDUCE_SUM(v_uint32x4, v_uint64x2, vuint64m1_t, unsigned, u32, u64, 4, wredsumu)
-OPENCV_HAL_IMPL_RVV_REDUCE_SUM(v_int32x4, v_int64x2, vint64m1_t, int, i32, i64, 4, wredsum)
-OPENCV_HAL_IMPL_RVV_REDUCE_SUM(v_uint64x2, v_uint64x2, vuint64m1_t, uint64, u64, u64, 2, redsum)
-OPENCV_HAL_IMPL_RVV_REDUCE_SUM(v_int64x2, v_int64x2, vint64m1_t, int64, i64, i64, 2, redsum)
-
-#define OPENCV_HAL_IMPL_RVV_REDUCE_SUM_FP(_Tpvec, _wTpvec, _nwTpvec, scalartype, suffix, wsuffix, vl, red) \
-inline scalartype v_reduce_sum(const _Tpvec& a)  \
-{ \
-    _nwTpvec zero = vfmv_v_f_##wsuffix##m1(0, vl); \
-    _nwTpvec res = vfmv_v_f_##wsuffix##m1(0, vl); \
-    res = v##red##_vs_##suffix##m1_##wsuffix##m1(res, a, zero, vl); \
-    return (scalartype)(_wTpvec(res).get0()); \
-}
-
-// vfredsum for float has renamed to fredosum, also updated in GNU.
-OPENCV_HAL_IMPL_RVV_REDUCE_SUM_FP(v_float32x4, v_float32x4, vfloat32m1_t, float, f32, f32, 4, fredosum)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_REDUCE_SUM_FP(v_float64x2, v_float64x2, vfloat64m1_t, double, f64, f64, 2, fredosum)
-#endif
-
-
-#define OPENCV_HAL_IMPL_RVV_REDUCE(_Tpvec, func, scalartype, suffix, vl, red) \
-inline scalartype v_reduce_##func(const _Tpvec& a)  \
-{ \
-    _Tpvec res = _Tpvec(v##red##_vs_##suffix##m1_##suffix##m1(a, a, a, vl)); \
-    return scalartype(res.get0()); \
-}
-
-OPENCV_HAL_IMPL_RVV_REDUCE(v_uint8x16, min, uchar, u8, 16, redminu)
-OPENCV_HAL_IMPL_RVV_REDUCE(v_int8x16, min, schar, i8, 16, redmin)
-OPENCV_HAL_IMPL_RVV_REDUCE(v_uint16x8, min, ushort, u16, 8, redminu)
-OPENCV_HAL_IMPL_RVV_REDUCE(v_int16x8, min, short, i16, 8, redmin)
-OPENCV_HAL_IMPL_RVV_REDUCE(v_uint32x4, min, unsigned, u32, 4, redminu)
-OPENCV_HAL_IMPL_RVV_REDUCE(v_int32x4, min, int, i32, 4, redmin)
-OPENCV_HAL_IMPL_RVV_REDUCE(v_float32x4, min, float, f32, 4, fredmin)
-OPENCV_HAL_IMPL_RVV_REDUCE(v_uint8x16, max, uchar, u8, 16, redmaxu)
-OPENCV_HAL_IMPL_RVV_REDUCE(v_int8x16, max, schar, i8, 16, redmax)
-OPENCV_HAL_IMPL_RVV_REDUCE(v_uint16x8, max, ushort, u16, 8, redmaxu)
-OPENCV_HAL_IMPL_RVV_REDUCE(v_int16x8, max, short, i16, 8, redmax)
-OPENCV_HAL_IMPL_RVV_REDUCE(v_uint32x4, max, unsigned, u32, 4, redmaxu)
-OPENCV_HAL_IMPL_RVV_REDUCE(v_int32x4, max, int, i32, 4, redmax)
-OPENCV_HAL_IMPL_RVV_REDUCE(v_float32x4, max, float, f32, 4, fredmax)
-
-
-inline v_float32x4 v_reduce_sum4(const v_float32x4& a, const v_float32x4& b,
-                                 const v_float32x4& c, const v_float32x4& d)
-{
-    float elems[4] =
-    {
-        v_reduce_sum(a),
-        v_reduce_sum(b),
-        v_reduce_sum(c),
-        v_reduce_sum(d)
-    };
-    return v_float32x4(vle32_v_f32m1(elems, 4));
-}
-
-////////////// Square-Root //////////////
-
-inline v_float32x4 v_sqrt(const v_float32x4& x)
-{
-    return v_float32x4(vfsqrt_v_f32m1(x, 4));
-}
-
-inline v_float32x4 v_invsqrt(const v_float32x4& x)
-{
-    v_float32x4 one = v_setall_f32(1.0f);
-    return one / v_sqrt(x);
-}
-
-#if CV_SIMD128_64F
-inline v_float64x2 v_sqrt(const v_float64x2& x)
-{
-    return v_float64x2(vfsqrt_v_f64m1(x, 4));
-}
-
-inline v_float64x2 v_invsqrt(const v_float64x2& x)
-{
-    v_float64x2 one = v_setall_f64(1.0f);
-    return one / v_sqrt(x);
-}
-#endif
-
-inline v_float32x4 v_magnitude(const v_float32x4& a, const v_float32x4& b)
-{
-    v_float32x4 x(vfmacc_vv_f32m1(vfmul_vv_f32m1(a, a, 4), b, b, 4));
-    return v_sqrt(x);
-}
-
-inline v_float32x4 v_sqr_magnitude(const v_float32x4& a, const v_float32x4& b)
-{
-    return v_float32x4(vfmacc_vv_f32m1(vfmul_vv_f32m1(a, a, 4), b, b, 4));
-}
-
-#if CV_SIMD128_64F
-inline v_float64x2 v_magnitude(const v_float64x2& a, const v_float64x2& b)
-{
-    v_float64x2 x(vfmacc_vv_f64m1(vfmul_vv_f64m1(a, a, 2), b, b, 2));
-    return v_sqrt(x);
-}
-
-inline v_float64x2 v_sqr_magnitude(const v_float64x2& a, const v_float64x2& b)
-{
-    return v_float64x2(vfmacc_vv_f64m1(vfmul_vv_f64m1(a, a, 2), b, b, 2));
-}
-#endif
-
-////////////// Multiply-Add //////////////
-
-inline v_float32x4 v_fma(const v_float32x4& a, const v_float32x4& b, const v_float32x4& c)
-{
-    return v_float32x4(vfmacc_vv_f32m1(c, a, b, 4));
-}
-inline v_int32x4 v_fma(const v_int32x4& a, const v_int32x4& b, const v_int32x4& c)
-{
-    return v_int32x4(vmacc_vv_i32m1(c, a, b, 4));
-}
-
-inline v_float32x4 v_muladd(const v_float32x4& a, const v_float32x4& b, const v_float32x4& c)
-{
-    return v_fma(a, b, c);
-}
-
-inline v_int32x4 v_muladd(const v_int32x4& a, const v_int32x4& b, const v_int32x4& c)
-{
-    return v_fma(a, b, c);
-}
-
-#if CV_SIMD128_64F
-inline v_float64x2 v_fma(const v_float64x2& a, const v_float64x2& b, const v_float64x2& c)
-{
-    return v_float64x2(vfmacc_vv_f64m1(c, a, b, 2));
-}
-
-inline v_float64x2 v_muladd(const v_float64x2& a, const v_float64x2& b, const v_float64x2& c)
-{
-    return v_fma(a, b, c);
-}
-#endif
-
-////////////// Check all/any //////////////
-
-// use overloaded vcpop in clang, no casting like (vuint64m1_t) is needed.
-#ifndef __clang__
-#define OPENCV_HAL_IMPL_RVV_CHECK_ALLANY(_Tpvec, suffix, shift, vl) \
-inline bool v_check_all(const _Tpvec& a) \
-{ \
-    v_uint64x2 v = v_uint64x2((vuint64m1_t)vsrl_vx_##suffix##m1(vnot_v_##suffix##m1(a, vl), shift, vl)); \
-    return (v.val[0] | v.val[1]) == 0; \
-} \
-inline bool v_check_any(const _Tpvec& a) \
-{ \
-    v_uint64x2 v = v_uint64x2((vuint64m1_t)vsrl_vx_##suffix##m1(a, shift, vl)); \
-    return (v.val[0] | v.val[1]) != 0; \
-}
-
-OPENCV_HAL_IMPL_RVV_CHECK_ALLANY(v_uint8x16, u8, 7, 16)
-OPENCV_HAL_IMPL_RVV_CHECK_ALLANY(v_uint16x8, u16, 15, 8)
-OPENCV_HAL_IMPL_RVV_CHECK_ALLANY(v_uint32x4, u32, 31, 4)
-OPENCV_HAL_IMPL_RVV_CHECK_ALLANY(v_uint64x2, u64, 63, 2)
-
-
-inline bool v_check_all(const v_int8x16& a)
-{ return v_check_all(v_reinterpret_as_u8(a)); }
-inline bool v_check_any(const v_int8x16& a)
-{ return v_check_any(v_reinterpret_as_u8(a)); }
-
-inline bool v_check_all(const v_int16x8& a)
-{ return v_check_all(v_reinterpret_as_u16(a)); }
-inline bool v_check_any(const v_int16x8& a)
-{ return v_check_any(v_reinterpret_as_u16(a)); }
-
-inline bool v_check_all(const v_int32x4& a)
-{ return v_check_all(v_reinterpret_as_u32(a)); }
-inline bool v_check_any(const v_int32x4& a)
-{ return v_check_any(v_reinterpret_as_u32(a)); }
-
-inline bool v_check_all(const v_float32x4& a)
-{ return v_check_all(v_reinterpret_as_u32(a)); }
-inline bool v_check_any(const v_float32x4& a)
-{ return v_check_any(v_reinterpret_as_u32(a)); }
-
-inline bool v_check_all(const v_int64x2& a)
-{ return v_check_all(v_reinterpret_as_u64(a)); }
-inline bool v_check_any(const v_int64x2& a)
-{ return v_check_any(v_reinterpret_as_u64(a)); }
-
-#if CV_SIMD128_64F
-inline bool v_check_all(const v_float64x2& a)
-{ return v_check_all(v_reinterpret_as_u64(a)); }
-inline bool v_check_any(const v_float64x2& a)
-{ return v_check_any(v_reinterpret_as_u64(a)); }
-#endif
-#else
-#define OPENCV_HAL_IMPL_RVV_CHECK_ALLANY(_Tpvec, vl) \
-inline bool v_check_all(const _Tpvec& a) \
-{ \
-    return vcpop(vmslt(a, 0, vl), vl) == vl; \
-} \
-inline bool v_check_any(const _Tpvec& a) \
-{ \
-    return vcpop(vmslt(a, 0, vl), vl) != 0; \
-}
-
-OPENCV_HAL_IMPL_RVV_CHECK_ALLANY(v_int8x16, 16)
-OPENCV_HAL_IMPL_RVV_CHECK_ALLANY(v_int16x8, 8)
-OPENCV_HAL_IMPL_RVV_CHECK_ALLANY(v_int32x4, 4)
-OPENCV_HAL_IMPL_RVV_CHECK_ALLANY(v_int64x2, 2)
-
-
-inline bool v_check_all(const v_uint8x16& a)
-{ return v_check_all(v_reinterpret_as_s8(a)); }
-inline bool v_check_any(const v_uint8x16& a)
-{ return v_check_any(v_reinterpret_as_s8(a)); }
-
-inline bool v_check_all(const v_uint16x8& a)
-{ return v_check_all(v_reinterpret_as_s16(a)); }
-inline bool v_check_any(const v_uint16x8& a)
-{ return v_check_any(v_reinterpret_as_s16(a)); }
-
-inline bool v_check_all(const v_uint32x4& a)
-{ return v_check_all(v_reinterpret_as_s32(a)); }
-inline bool v_check_any(const v_uint32x4& a)
-{ return v_check_any(v_reinterpret_as_s32(a)); }
-
-inline bool v_check_all(const v_float32x4& a)
-{ return v_check_all(v_reinterpret_as_s32(a)); }
-inline bool v_check_any(const v_float32x4& a)
-{ return v_check_any(v_reinterpret_as_s32(a)); }
-
-inline bool v_check_all(const v_uint64x2& a)
-{ return v_check_all(v_reinterpret_as_s64(a)); }
-inline bool v_check_any(const v_uint64x2& a)
-{ return v_check_any(v_reinterpret_as_s64(a)); }
-
-#if CV_SIMD128_64F
-inline bool v_check_all(const v_float64x2& a)
-{ return v_check_all(v_reinterpret_as_s64(a)); }
-inline bool v_check_any(const v_float64x2& a)
-{ return v_check_any(v_reinterpret_as_s64(a)); }
-#endif
-#endif
-////////////// abs //////////////
-
-#define OPENCV_HAL_IMPL_RVV_ABSDIFF(_Tpvec, abs) \
-inline _Tpvec v_##abs(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return v_max(a, b) - v_min(a, b); \
-}
-
-OPENCV_HAL_IMPL_RVV_ABSDIFF(v_uint8x16, absdiff)
-OPENCV_HAL_IMPL_RVV_ABSDIFF(v_uint16x8, absdiff)
-OPENCV_HAL_IMPL_RVV_ABSDIFF(v_uint32x4, absdiff)
-OPENCV_HAL_IMPL_RVV_ABSDIFF(v_float32x4, absdiff)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_ABSDIFF(v_float64x2, absdiff)
-#endif
-OPENCV_HAL_IMPL_RVV_ABSDIFF(v_int8x16, absdiffs)
-OPENCV_HAL_IMPL_RVV_ABSDIFF(v_int16x8, absdiffs)
-
-// use reinterpret instead of c-style casting.
-#ifndef __clang__
-#define OPENCV_HAL_IMPL_RVV_ABSDIFF_S(_Tpvec, _rTpvec, _nwTpvec, sub, rshr, vl) \
-inline _rTpvec v_absdiff(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _rTpvec(rshr((_nwTpvec)sub(v_max(a, b), v_min(a, b), vl), 0, vl)); \
-}
-
-OPENCV_HAL_IMPL_RVV_ABSDIFF_S(v_int8x16, v_uint8x16, vuint16m2_t, vwsub_vv_i16m2, vnclipu_wx_u8m1, 16)
-OPENCV_HAL_IMPL_RVV_ABSDIFF_S(v_int16x8, v_uint16x8, vuint32m2_t, vwsub_vv_i32m2, vnclipu_wx_u16m1, 8)
-OPENCV_HAL_IMPL_RVV_ABSDIFF_S(v_int32x4, v_uint32x4, vuint64m2_t, vwsub_vv_i64m2, vnclipu_wx_u32m1, 4)
-#else
-#define OPENCV_HAL_IMPL_RVV_ABSDIFF_S(_Tpvec, _rTpvec, _nwTpvec, sub, rshr, width, vl) \
-inline _rTpvec v_absdiff(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _rTpvec(rshr(vreinterpret_u##width##m2(sub(v_max(a, b), v_min(a, b), vl)), 0, vl)); \
-}
-
-OPENCV_HAL_IMPL_RVV_ABSDIFF_S(v_int8x16, v_uint8x16, vuint16m2_t, vwsub_vv_i16m2, vnclipu_wx_u8m1, 16, 16)
-OPENCV_HAL_IMPL_RVV_ABSDIFF_S(v_int16x8, v_uint16x8, vuint32m2_t, vwsub_vv_i32m2, vnclipu_wx_u16m1, 32, 8)
-OPENCV_HAL_IMPL_RVV_ABSDIFF_S(v_int32x4, v_uint32x4, vuint64m2_t, vwsub_vv_i64m2, vnclipu_wx_u32m1, 64, 4)
-#endif
-#define OPENCV_HAL_IMPL_RVV_ABS(_Tprvec, _Tpvec, suffix) \
-inline _Tprvec v_abs(const _Tpvec& a) \
-{ \
-    return v_absdiff(a, v_setzero_##suffix()); \
-}
-
-OPENCV_HAL_IMPL_RVV_ABS(v_uint8x16, v_int8x16, s8)
-OPENCV_HAL_IMPL_RVV_ABS(v_uint16x8, v_int16x8, s16)
-OPENCV_HAL_IMPL_RVV_ABS(v_uint32x4, v_int32x4, s32)
-OPENCV_HAL_IMPL_RVV_ABS(v_float32x4, v_float32x4, f32)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_ABS(v_float64x2, v_float64x2, f64)
-#endif
-
-
-#define OPENCV_HAL_IMPL_RVV_REDUCE_SAD(_Tpvec, scalartype) \
-inline scalartype v_reduce_sad(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return v_reduce_sum(v_absdiff(a, b)); \
-}
-
-OPENCV_HAL_IMPL_RVV_REDUCE_SAD(v_uint8x16, unsigned)
-OPENCV_HAL_IMPL_RVV_REDUCE_SAD(v_int8x16, unsigned)
-OPENCV_HAL_IMPL_RVV_REDUCE_SAD(v_uint16x8, unsigned)
-OPENCV_HAL_IMPL_RVV_REDUCE_SAD(v_int16x8, unsigned)
-OPENCV_HAL_IMPL_RVV_REDUCE_SAD(v_uint32x4, unsigned)
-OPENCV_HAL_IMPL_RVV_REDUCE_SAD(v_int32x4, unsigned)
-OPENCV_HAL_IMPL_RVV_REDUCE_SAD(v_float32x4, float)
-
-////////////// Select //////////////
-
-#define OPENCV_HAL_IMPL_RVV_SELECT(_Tpvec, merge, ne, vl) \
-inline _Tpvec v_select(const _Tpvec& mask, const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(merge(ne(mask, 0, vl), b, a, vl)); \
-}
-
-OPENCV_HAL_IMPL_RVV_SELECT(v_uint8x16, vmerge_vvm_u8m1, vmsne_vx_u8m1_b8, 16)
-OPENCV_HAL_IMPL_RVV_SELECT(v_int8x16, vmerge_vvm_i8m1, vmsne_vx_i8m1_b8, 16)
-OPENCV_HAL_IMPL_RVV_SELECT(v_uint16x8, vmerge_vvm_u16m1, vmsne_vx_u16m1_b16, 8)
-OPENCV_HAL_IMPL_RVV_SELECT(v_int16x8, vmerge_vvm_i16m1, vmsne_vx_i16m1_b16, 8)
-OPENCV_HAL_IMPL_RVV_SELECT(v_uint32x4, vmerge_vvm_u32m1, vmsne_vx_u32m1_b32, 4)
-OPENCV_HAL_IMPL_RVV_SELECT(v_int32x4, vmerge_vvm_i32m1, vmsne_vx_i32m1_b32, 4)
-OPENCV_HAL_IMPL_RVV_SELECT(v_float32x4, vmerge_vvm_f32m1, vmfne_vf_f32m1_b32, 4)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_SELECT(v_float64x2, vmerge_vvm_f64m1, vmfne_vf_f64m1_b64, 2)
-#endif
-
-////////////// Rotate shift //////////////
-
-#define OPENCV_HAL_IMPL_RVV_ROTATE_INTEGER(_Tpvec, suffix, vl) \
-template<int n> inline _Tpvec v_rotate_right(const _Tpvec& a) \
-{ \
-    return _Tpvec(vslidedown_vx_##suffix##m1(vmv_v_x_##suffix##m1(0, vl), a, n, vl)); \
-} \
-template<int n> inline _Tpvec v_rotate_left(const _Tpvec& a) \
-{ \
-    return _Tpvec(vslideup_vx_##suffix##m1(vmv_v_x_##suffix##m1(0, vl), a, n, vl)); \
-} \
-template<> inline _Tpvec v_rotate_left<0>(const _Tpvec& a) \
-{ return a; } \
-template<int n> inline _Tpvec v_rotate_right(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(vslideup_vx_##suffix##m1(vslidedown_vx_##suffix##m1(vmv_v_x_##suffix##m1(0, vl), a, n, vl), b, _Tpvec::nlanes - n, vl)); \
-} \
-template<int n> inline _Tpvec v_rotate_left(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(vslideup_vx_##suffix##m1(vslidedown_vx_##suffix##m1(vmv_v_x_##suffix##m1(0, vl), b, _Tpvec::nlanes - n, vl), a, n, vl)); \
-} \
-template<> inline _Tpvec v_rotate_left<0>(const _Tpvec& a, const _Tpvec& b) \
-{ CV_UNUSED(b); return a; }
-
-OPENCV_HAL_IMPL_RVV_ROTATE_INTEGER(v_uint8x16, u8, 16)
-OPENCV_HAL_IMPL_RVV_ROTATE_INTEGER(v_int8x16, i8, 16)
-OPENCV_HAL_IMPL_RVV_ROTATE_INTEGER(v_uint16x8, u16, 8)
-OPENCV_HAL_IMPL_RVV_ROTATE_INTEGER(v_int16x8, i16, 8)
-OPENCV_HAL_IMPL_RVV_ROTATE_INTEGER(v_uint32x4, u32, 4)
-OPENCV_HAL_IMPL_RVV_ROTATE_INTEGER(v_int32x4, i32, 4)
-OPENCV_HAL_IMPL_RVV_ROTATE_INTEGER(v_uint64x2, u64, 2)
-OPENCV_HAL_IMPL_RVV_ROTATE_INTEGER(v_int64x2, i64, 2)
-
-#define OPENCV_HAL_IMPL_RVV_ROTATE_FP(_Tpvec, suffix, vl) \
-template<int n> inline _Tpvec v_rotate_right(const _Tpvec& a) \
-{ \
-    return _Tpvec(vslidedown_vx_##suffix##m1(vfmv_v_f_##suffix##m1(0, vl), a, n, vl)); \
-} \
-template<int n> inline _Tpvec v_rotate_left(const _Tpvec& a) \
-{ \
-    return _Tpvec(vslideup_vx_##suffix##m1(vfmv_v_f_##suffix##m1(0, vl), a, n, vl)); \
-} \
-template<> inline _Tpvec v_rotate_left<0>(const _Tpvec& a) \
-{ return a; } \
-template<int n> inline _Tpvec v_rotate_right(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(vslideup_vx_##suffix##m1(vslidedown_vx_##suffix##m1(vfmv_v_f_##suffix##m1(0, vl), a, n, vl), b, _Tpvec::nlanes - n, vl)); \
-} \
-template<int n> inline _Tpvec v_rotate_left(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(vslideup_vx_##suffix##m1(vslidedown_vx_##suffix##m1(vfmv_v_f_##suffix##m1(0, vl), b, _Tpvec::nlanes - n, vl), a, n, vl)); \
-} \
-template<> inline _Tpvec v_rotate_left<0>(const _Tpvec& a, const _Tpvec& b) \
-{ CV_UNUSED(b); return a; }
-
-OPENCV_HAL_IMPL_RVV_ROTATE_FP(v_float32x4, f32, 4)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_ROTATE_FP(v_float64x2, f64, 2)
-#endif
-
-////////////// Convert to float //////////////
-
-inline v_float32x4 v_cvt_f32(const v_int32x4& a)
-{
-    return v_float32x4(vfcvt_f_x_v_f32m1(a, 4));
-}
-
-#if CV_SIMD128_64F
-inline v_float32x4 v_cvt_f32(const v_float64x2& a)
-{
-    double arr[4] = {a.val[0], a.val[1], 0, 0};
-    vfloat64m2_t tmp = vle64_v_f64m2(arr, 4);
-    return v_float32x4(vfncvt_f_f_w_f32m1(tmp, 4));
-}
-
-inline v_float32x4 v_cvt_f32(const v_float64x2& a, const v_float64x2& b)
-{
-    double arr[4] = {a.val[0], a.val[1], b.val[0], b.val[1]};
-    vfloat64m2_t tmp = vle64_v_f64m2(arr, 4);
-    return v_float32x4(vfncvt_f_f_w_f32m1(tmp, 4));
-}
-
-inline v_float64x2 v_cvt_f64(const v_int32x4& a)
-{
-    double ptr[4] = {0};
-    vse64_v_f64m2(ptr, vfwcvt_f_x_v_f64m2(a, 4), 4);
-    double elems[2] =
-    {
-        ptr[0], ptr[1]
-    };
-    return v_float64x2(vle64_v_f64m1(elems, 2));
-}
-
-inline v_float64x2 v_cvt_f64_high(const v_int32x4& a)
-{
-    double ptr[4] = {0};
-    vse64_v_f64m2(ptr, vfwcvt_f_x_v_f64m2(a, 4), 4);
-    double elems[2] =
-    {
-        ptr[2], ptr[3]
-    };
-    return v_float64x2(vle64_v_f64m1(elems, 2));
-}
-
-inline v_float64x2 v_cvt_f64(const v_float32x4& a)
-{
-    double ptr[4] = {0};
-    vse64_v_f64m2(ptr, vfwcvt_f_f_v_f64m2(a, 4), 4);
-    double elems[2] =
-    {
-        ptr[0], ptr[1]
-    };
-    return v_float64x2(vle64_v_f64m1(elems, 2));
-}
-
-inline v_float64x2 v_cvt_f64_high(const v_float32x4& a)
-{
-    double ptr[4] = {0};
-    vse64_v_f64m2(ptr, vfwcvt_f_f_v_f64m2(a, 4), 4);
-    double elems[2] =
-    {
-        ptr[2], ptr[3]
-    };
-    return v_float64x2(vle64_v_f64m1(elems, 2));
-}
-
-inline v_float64x2 v_cvt_f64(const v_int64x2& a)
-{
-    return v_float64x2(vfcvt_f_x_v_f64m1(a, 2));
-}
-#endif
-
-////////////// Broadcast //////////////
-
-#define OPENCV_HAL_IMPL_RVV_BROADCAST(_Tpvec, suffix) \
-template<int i> inline _Tpvec v_broadcast_element(_Tpvec v) \
-{ \
-    return v_setall_##suffix(v_extract_n<i>(v)); \
-}
-
-OPENCV_HAL_IMPL_RVV_BROADCAST(v_uint8x16, u8)
-OPENCV_HAL_IMPL_RVV_BROADCAST(v_int8x16, s8)
-OPENCV_HAL_IMPL_RVV_BROADCAST(v_uint16x8, u16)
-OPENCV_HAL_IMPL_RVV_BROADCAST(v_int16x8, s16)
-OPENCV_HAL_IMPL_RVV_BROADCAST(v_uint32x4, u32)
-OPENCV_HAL_IMPL_RVV_BROADCAST(v_int32x4, s32)
-OPENCV_HAL_IMPL_RVV_BROADCAST(v_uint64x2, u64)
-OPENCV_HAL_IMPL_RVV_BROADCAST(v_int64x2, s64)
-OPENCV_HAL_IMPL_RVV_BROADCAST(v_float32x4, f32)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_BROADCAST(v_float64x2, f64)
-#endif
-
-////////////// Transpose4x4 //////////////
-
-#define OPENCV_HAL_IMPL_RVV_TRANSPOSE4x4(_Tpvec, _Tp, suffix) \
-inline void v_transpose4x4(const v_##_Tpvec& a0, const v_##_Tpvec& a1, \
-                         const v_##_Tpvec& a2, const v_##_Tpvec& a3, \
-                         v_##_Tpvec& b0, v_##_Tpvec& b1, \
-                         v_##_Tpvec& b2, v_##_Tpvec& b3) \
-{ \
-    _Tp elems0[4] = \
-    { \
-        v_extract_n<0>(a0), \
-        v_extract_n<0>(a1), \
-        v_extract_n<0>(a2), \
-        v_extract_n<0>(a3) \
-    }; \
-    b0 = v_load(elems0); \
-    _Tp elems1[4] = \
-    { \
-        v_extract_n<1>(a0), \
-        v_extract_n<1>(a1), \
-        v_extract_n<1>(a2), \
-        v_extract_n<1>(a3) \
-    }; \
-    b1 = v_load(elems1); \
-    _Tp elems2[4] = \
-    { \
-        v_extract_n<2>(a0), \
-        v_extract_n<2>(a1), \
-        v_extract_n<2>(a2), \
-        v_extract_n<2>(a3) \
-    }; \
-    b2 = v_load(elems2); \
-    _Tp elems3[4] = \
-    { \
-        v_extract_n<3>(a0), \
-        v_extract_n<3>(a1), \
-        v_extract_n<3>(a2), \
-        v_extract_n<3>(a3) \
-    }; \
-    b3 = v_load(elems3); \
-}
-
-OPENCV_HAL_IMPL_RVV_TRANSPOSE4x4(uint32x4, unsigned, u32)
-OPENCV_HAL_IMPL_RVV_TRANSPOSE4x4(int32x4, int, i32)
-OPENCV_HAL_IMPL_RVV_TRANSPOSE4x4(float32x4, float, f32)
-
-////////////// Reverse //////////////
-
-#define OPENCV_HAL_IMPL_RVV_REVERSE(_Tpvec, _Tp, suffix) \
-inline _Tpvec v_reverse(const _Tpvec& a)  \
-{ \
-    _Tp ptr[_Tpvec::nlanes] = {0}; \
-    _Tp ptra[_Tpvec::nlanes] = {0}; \
-    v_store(ptra, a); \
-    for (int i = 0; i < _Tpvec::nlanes; i++) \
-    { \
-        ptr[i] = ptra[_Tpvec::nlanes-i-1]; \
-    } \
-    return v_load(ptr); \
-}
-
-OPENCV_HAL_IMPL_RVV_REVERSE(v_uint8x16, uchar, u8)
-OPENCV_HAL_IMPL_RVV_REVERSE(v_int8x16, schar, i8)
-OPENCV_HAL_IMPL_RVV_REVERSE(v_uint16x8, ushort, u16)
-OPENCV_HAL_IMPL_RVV_REVERSE(v_int16x8, short, i16)
-OPENCV_HAL_IMPL_RVV_REVERSE(v_uint32x4, unsigned, u32)
-OPENCV_HAL_IMPL_RVV_REVERSE(v_int32x4, int, i32)
-OPENCV_HAL_IMPL_RVV_REVERSE(v_float32x4, float, f32)
-OPENCV_HAL_IMPL_RVV_REVERSE(v_uint64x2, uint64, u64)
-OPENCV_HAL_IMPL_RVV_REVERSE(v_int64x2, int64, i64)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_REVERSE(v_float64x2, double, f64)
-#endif
-
-//////////// Value reordering ////////////
-
-#define OPENCV_HAL_IMPL_RVV_EXPAND(_Tpwvec, _Tp, _Tpvec, width, suffix, wcvt, vl) \
-inline void v_expand(const _Tpvec& a, _Tpwvec& b0, _Tpwvec& b1) \
-{ \
-    _Tp lptr[_Tpvec::nlanes/2] = {0}; \
-    _Tp hptr[_Tpvec::nlanes/2] = {0}; \
-    v_store_low(lptr, a); \
-    v_store_high(hptr, a); \
-    b0 = _Tpwvec(wcvt(vle##width##_v_##suffix##mf2(lptr, vl), vl)); \
-    b1 = _Tpwvec(wcvt(vle##width##_v_##suffix##mf2(hptr, vl), vl)); \
-} \
-inline _Tpwvec v_expand_low(const _Tpvec& a) \
-{ \
-    _Tp lptr[_Tpvec::nlanes/2] = {0}; \
-    v_store_low(lptr, a); \
-    return _Tpwvec(wcvt(vle##width##_v_##suffix##mf2(lptr, vl), vl)); \
-} \
-inline _Tpwvec v_expand_high(const _Tpvec& a) \
-{ \
-    _Tp hptr[_Tpvec::nlanes/2] = {0}; \
-    v_store_high(hptr, a); \
-    return _Tpwvec(wcvt(vle##width##_v_##suffix##mf2(hptr, vl), vl)); \
-} \
-inline _Tpwvec v_load_expand(const _Tp* ptr) \
-{ \
-    return _Tpwvec(wcvt(vle##width##_v_##suffix##mf2(ptr, vl), vl)); \
-}
-
-OPENCV_HAL_IMPL_RVV_EXPAND(v_uint16x8, uchar, v_uint8x16, 8, u8, vwcvtu_x_x_v_u16m1, 8)
-OPENCV_HAL_IMPL_RVV_EXPAND(v_int16x8, schar, v_int8x16, 8, i8, vwcvt_x_x_v_i16m1, 8)
-OPENCV_HAL_IMPL_RVV_EXPAND(v_uint32x4, ushort, v_uint16x8, 16, u16, vwcvtu_x_x_v_u32m1, 4)
-OPENCV_HAL_IMPL_RVV_EXPAND(v_int32x4, short, v_int16x8, 16, i16, vwcvt_x_x_v_i32m1, 4)
-OPENCV_HAL_IMPL_RVV_EXPAND(v_uint64x2, uint, v_uint32x4, 32, u32, vwcvtu_x_x_v_u64m1, 2)
-OPENCV_HAL_IMPL_RVV_EXPAND(v_int64x2, int, v_int32x4, 32, i32, vwcvt_x_x_v_i64m1, 2)
-
-inline v_uint32x4 v_load_expand_q(const uchar* ptr)
-{
-    return v_uint32x4(vwcvtu_x_x_v_u32m1(vwcvtu_x_x_v_u16mf2(vle8_v_u8mf4(ptr, 4), 4), 4));
-}
-
-inline v_int32x4 v_load_expand_q(const schar* ptr)
-{
-    return v_int32x4(vwcvt_x_x_v_i32m1(vwcvt_x_x_v_i16mf2(vle8_v_i8mf4(ptr, 4), 4), 4));
-}
-
-
-#define OPENCV_HAL_IMPL_RVV_PACK(_Tpvec, _Tp, _wTpvec, _wTp, hwidth, width, hsuffix, suffix, rshr, shr, hvl, vl) \
-inline _Tpvec v_pack(const _wTpvec& a, const _wTpvec& b) \
-{ \
-    _wTp arr[_Tpvec::nlanes] = {0}; \
-    v_store(arr, a); \
-    v_store(arr + _wTpvec::nlanes, b); \
-    return _Tpvec(shr(vle##width##_v_##suffix##m2(arr, vl), 0, vl)); \
-} \
-inline void v_pack_store(_Tp* ptr, const _wTpvec& a) \
-{ \
-    _wTp arr[_Tpvec::nlanes] = {0}; \
-    v_store(arr, a); \
-    v_store(arr + _wTpvec::nlanes, _wTpvec(vmv_v_x_##suffix##m1(0, hvl))); \
-    vse##hwidth##_v_##hsuffix##m1(ptr, shr(vle##width##_v_##suffix##m2(arr, vl), 0, vl), hvl); \
-} \
-template<int n> inline \
-_Tpvec v_rshr_pack(const _wTpvec& a, const _wTpvec& b) \
-{ \
-    _wTp arr[_Tpvec::nlanes] = {0}; \
-    v_store(arr, a); \
-    v_store(arr + _wTpvec::nlanes, b); \
-    return _Tpvec(rshr(vle##width##_v_##suffix##m2(arr, vl), n, vl)); \
-} \
-template<int n> inline \
-void v_rshr_pack_store(_Tp* ptr, const _wTpvec& a) \
-{ \
-    _wTp arr[_Tpvec::nlanes] = {0}; \
-    v_store(arr, a); \
-    v_store(arr + _wTpvec::nlanes, _wTpvec(vmv_v_x_##suffix##m1(0, hvl))); \
-    vse##hwidth##_v_##hsuffix##m1(ptr, _Tpvec(rshr(vle##width##_v_##suffix##m2(arr, vl), n, vl)), hvl); \
-}
-
-OPENCV_HAL_IMPL_RVV_PACK(v_uint8x16, uchar, v_uint16x8, ushort, 8, 16, u8, u16, vnclipu_wx_u8m1, vnclipu_wx_u8m1, 8, 16)
-OPENCV_HAL_IMPL_RVV_PACK(v_int8x16, schar, v_int16x8, short, 8, 16, i8, i16, vnclip_wx_i8m1, vnclip_wx_i8m1, 8, 16)
-OPENCV_HAL_IMPL_RVV_PACK(v_uint16x8, ushort, v_uint32x4, unsigned, 16, 32, u16, u32, vnclipu_wx_u16m1, vnclipu_wx_u16m1, 4, 8)
-OPENCV_HAL_IMPL_RVV_PACK(v_int16x8, short, v_int32x4, int, 16, 32, i16, i32, vnclip_wx_i16m1, vnclip_wx_i16m1, 4, 8)
-OPENCV_HAL_IMPL_RVV_PACK(v_uint32x4, unsigned, v_uint64x2, uint64, 32, 64, u32, u64, vnclipu_wx_u32m1, vnsrl_wx_u32m1, 2, 4)
-OPENCV_HAL_IMPL_RVV_PACK(v_int32x4, int, v_int64x2, int64, 32, 64, i32, i64, vnclip_wx_i32m1, vnsra_wx_i32m1, 2, 4)
-
-
-#define OPENCV_HAL_IMPL_RVV_PACK_U(_Tpvec, _Tp, _wTpvec, _wTp, hwidth, width, hsuffix, suffix, rshr, cast, hvl, vl) \
-inline _Tpvec v_pack_u(const _wTpvec& a, const _wTpvec& b) \
-{ \
-    _wTp arr[_Tpvec::nlanes] = {0}; \
-    v_store(arr, a); \
-    v_store(arr + _wTpvec::nlanes, b); \
-    return _Tpvec(rshr(cast(vmax_vx_##suffix##m2(vle##width##_v_##suffix##m2(arr, vl), 0, vl)), 0, vl)); \
-} \
-inline void v_pack_u_store(_Tp* ptr, const _wTpvec& a) \
-{ \
-    _wTp arr[_Tpvec::nlanes] = {0}; \
-    v_store(arr, a); \
-    v_store(arr + _wTpvec::nlanes, _wTpvec(vmv_v_x_##suffix##m1(0, hvl))); \
-    vse##hwidth##_v_##hsuffix##m1(ptr, rshr(cast(vmax_vx_##suffix##m2(vle##width##_v_##suffix##m2(arr, vl), 0, vl)), 0, vl), hvl); \
-} \
-template<int n> inline \
-_Tpvec v_rshr_pack_u(const _wTpvec& a, const _wTpvec& b) \
-{ \
-    _wTp arr[_Tpvec::nlanes] = {0}; \
-    v_store(arr, a); \
-    v_store(arr + _wTpvec::nlanes, b); \
-    return _Tpvec(rshr(cast(vmax_vx_##suffix##m2(vle##width##_v_##suffix##m2(arr, vl), 0, vl)), n, vl)); \
-} \
-template<int n> inline \
-void v_rshr_pack_u_store(_Tp* ptr, const _wTpvec& a) \
-{ \
-    _wTp arr[_Tpvec::nlanes] = {0}; \
-    v_store(arr, a); \
-    v_store(arr + _wTpvec::nlanes, _wTpvec(vmv_v_x_##suffix##m1(0, hvl))); \
-    v_store(ptr, _Tpvec(rshr(cast(vmax_vx_##suffix##m2(vle##width##_v_##suffix##m2(arr, vl), 0, vl)), n, vl))); \
-}
-
-OPENCV_HAL_IMPL_RVV_PACK_U(v_uint8x16, uchar, v_int16x8, short, 8, 16, u8, i16, vnclipu_wx_u8m1, vreinterpret_v_i16m2_u16m2, 8, 16)
-OPENCV_HAL_IMPL_RVV_PACK_U(v_uint16x8, ushort, v_int32x4, int, 16, 32, u16, i32, vnclipu_wx_u16m1, vreinterpret_v_i32m2_u32m2, 4, 8)
-
-
-#define OPENCV_HAL_IMPL_RVV_UNPACKS(_Tpvec, _Tp, suffix) \
-inline void v_zip(const v_##_Tpvec& a0, const v_##_Tpvec& a1, v_##_Tpvec& b0, v_##_Tpvec& b1) \
-{ \
-    _Tp ptra0[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptra1[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrb0[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrb1[v_##_Tpvec::nlanes] = {0}; \
-    v_store(ptra0, a0); \
-    v_store(ptra1, a1); \
-    int i; \
-    for( i = 0; i < v_##_Tpvec::nlanes/2; i++ ) \
-    { \
-        ptrb0[i*2] = ptra0[i]; \
-        ptrb0[i*2+1] = ptra1[i]; \
-    } \
-    for( ; i < v_##_Tpvec::nlanes; i++ ) \
-    { \
-        ptrb1[i*2-v_##_Tpvec::nlanes] = ptra0[i]; \
-        ptrb1[i*2-v_##_Tpvec::nlanes+1] = ptra1[i]; \
-    } \
-    b0 = v_load(ptrb0); \
-    b1 = v_load(ptrb1); \
-} \
-inline v_##_Tpvec v_combine_low(const v_##_Tpvec& a, const v_##_Tpvec& b) \
-{ \
-    _Tp ptra[v_##_Tpvec::nlanes/2] = {0}; \
-    _Tp ptrb[v_##_Tpvec::nlanes/2] = {0}; \
-    v_store_low(ptra, a); \
-    v_store_low(ptrb, b); \
-    return v_load_halves(ptra, ptrb); \
-} \
-inline v_##_Tpvec v_combine_high(const v_##_Tpvec& a, const v_##_Tpvec& b) \
-{ \
-    _Tp ptra[v_##_Tpvec::nlanes/2] = {0}; \
-    _Tp ptrb[v_##_Tpvec::nlanes/2] = {0}; \
-    v_store_high(ptra, a); \
-    v_store_high(ptrb, b); \
-    return v_load_halves(ptra, ptrb); \
-} \
-inline void v_recombine(const v_##_Tpvec& a, const v_##_Tpvec& b, v_##_Tpvec& c, v_##_Tpvec& d) \
-{ \
-    c = v_combine_low(a, b); \
-    d = v_combine_high(a, b); \
-}
-
-OPENCV_HAL_IMPL_RVV_UNPACKS(uint8x16, uchar, u8)
-OPENCV_HAL_IMPL_RVV_UNPACKS(int8x16, schar, i8)
-OPENCV_HAL_IMPL_RVV_UNPACKS(uint16x8, ushort, u16)
-OPENCV_HAL_IMPL_RVV_UNPACKS(int16x8, short, i16)
-OPENCV_HAL_IMPL_RVV_UNPACKS(uint32x4, unsigned, u32)
-OPENCV_HAL_IMPL_RVV_UNPACKS(int32x4, int, i32)
-OPENCV_HAL_IMPL_RVV_UNPACKS(float32x4, float, f32)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_UNPACKS(float64x2, double, f64)
-#endif
-
-
-#define OPENCV_HAL_IMPL_RVV_INTERLEAVED(_Tpvec, _Tp) \
-inline void v_load_deinterleave(const _Tp* ptr, v_##_Tpvec& a, v_##_Tpvec& b) \
-{ \
-    _Tp ptra[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrb[v_##_Tpvec::nlanes] = {0}; \
-    int i, i2; \
-    for( i = i2 = 0; i < v_##_Tpvec::nlanes; i++, i2 += 2 ) \
-    { \
-        ptra[i] = ptr[i2]; \
-        ptrb[i] = ptr[i2+1]; \
-    } \
-    a = v_load(ptra); \
-    b = v_load(ptrb); \
-} \
-inline void v_load_deinterleave(const _Tp* ptr, v_##_Tpvec& a, v_##_Tpvec& b, v_##_Tpvec& c) \
-{ \
-    _Tp ptra[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrb[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrc[v_##_Tpvec::nlanes] = {0}; \
-    int i, i3; \
-    for( i = i3 = 0; i < v_##_Tpvec::nlanes; i++, i3 += 3 ) \
-    { \
-        ptra[i] = ptr[i3]; \
-        ptrb[i] = ptr[i3+1]; \
-        ptrc[i] = ptr[i3+2]; \
-    } \
-    a = v_load(ptra); \
-    b = v_load(ptrb); \
-    c = v_load(ptrc); \
-} \
-inline void v_load_deinterleave(const _Tp* ptr, v_##_Tpvec& a, v_##_Tpvec& b, \
-                                v_##_Tpvec& c, v_##_Tpvec& d) \
-{ \
-    _Tp ptra[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrb[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrc[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrd[v_##_Tpvec::nlanes] = {0}; \
-    int i, i4; \
-    for( i = i4 = 0; i < v_##_Tpvec::nlanes; i++, i4 += 4 ) \
-    { \
-        ptra[i] = ptr[i4]; \
-        ptrb[i] = ptr[i4+1]; \
-        ptrc[i] = ptr[i4+2]; \
-        ptrd[i] = ptr[i4+3]; \
-    } \
-    a = v_load(ptra); \
-    b = v_load(ptrb); \
-    c = v_load(ptrc); \
-    d = v_load(ptrd); \
-} \
-inline void v_store_interleave( _Tp* ptr, const v_##_Tpvec& a, const v_##_Tpvec& b, \
-                                hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ \
-    int i, i2; \
-    _Tp ptra[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrb[v_##_Tpvec::nlanes] = {0}; \
-    v_store(ptra, a); \
-    v_store(ptrb, b); \
-    for( i = i2 = 0; i < v_##_Tpvec::nlanes; i++, i2 += 2 ) \
-    { \
-        ptr[i2] = ptra[i]; \
-        ptr[i2+1] = ptrb[i]; \
-    } \
-} \
-inline void v_store_interleave( _Tp* ptr, const v_##_Tpvec& a, const v_##_Tpvec& b, \
-                                const v_##_Tpvec& c, hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ \
-    int i, i3; \
-    _Tp ptra[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrb[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrc[v_##_Tpvec::nlanes] = {0}; \
-    v_store(ptra, a); \
-    v_store(ptrb, b); \
-    v_store(ptrc, c); \
-    for( i = i3 = 0; i < v_##_Tpvec::nlanes; i++, i3 += 3 ) \
-    { \
-        ptr[i3] = ptra[i]; \
-        ptr[i3+1] = ptrb[i]; \
-        ptr[i3+2] = ptrc[i]; \
-    } \
-} \
-inline void v_store_interleave( _Tp* ptr, const v_##_Tpvec& a, const v_##_Tpvec& b, \
-                                const v_##_Tpvec& c, const v_##_Tpvec& d, \
-                                hal::StoreMode /*mode*/=hal::STORE_UNALIGNED ) \
-{ \
-    int i, i4; \
-    _Tp ptra[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrb[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrc[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrd[v_##_Tpvec::nlanes] = {0}; \
-    v_store(ptra, a); \
-    v_store(ptrb, b); \
-    v_store(ptrc, c); \
-    v_store(ptrd, d); \
-    for( i = i4 = 0; i < v_##_Tpvec::nlanes; i++, i4 += 4 ) \
-    { \
-        ptr[i4] = ptra[i]; \
-        ptr[i4+1] = ptrb[i]; \
-        ptr[i4+2] = ptrc[i]; \
-        ptr[i4+3] = ptrd[i]; \
-    } \
-} \
-inline v_##_Tpvec v_interleave_pairs(const v_##_Tpvec& vec) \
-{ \
-    _Tp ptr[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrvec[v_##_Tpvec::nlanes] = {0}; \
-    v_store(ptrvec, vec); \
-    for (int i = 0; i < v_##_Tpvec::nlanes/4; i++) \
-    { \
-        ptr[4*i  ] = ptrvec[4*i  ]; \
-        ptr[4*i+1] = ptrvec[4*i+2]; \
-        ptr[4*i+2] = ptrvec[4*i+1]; \
-        ptr[4*i+3] = ptrvec[4*i+3]; \
-    } \
-    return v_load(ptr); \
-} \
-inline v_##_Tpvec v_interleave_quads(const v_##_Tpvec& vec) \
-{ \
-    _Tp ptr[v_##_Tpvec::nlanes] = {0}; \
-    _Tp ptrvec[v_##_Tpvec::nlanes] = {0}; \
-    v_store(ptrvec, vec); \
-    for (int i = 0; i < v_##_Tpvec::nlanes/8; i++) \
-    { \
-        ptr[8*i  ] = ptrvec[4*i  ]; \
-        ptr[8*i+1] = ptrvec[4*i+4]; \
-        ptr[8*i+2] = ptrvec[4*i+1]; \
-        ptr[8*i+3] = ptrvec[4*i+5]; \
-        ptr[8*i+4] = ptrvec[4*i+2]; \
-        ptr[8*i+5] = ptrvec[4*i+6]; \
-        ptr[8*i+6] = ptrvec[4*i+3]; \
-        ptr[8*i+7] = ptrvec[4*i+7]; \
-    } \
-    return v_load(ptr); \
-}
-
-OPENCV_HAL_IMPL_RVV_INTERLEAVED(uint8x16, uchar)
-OPENCV_HAL_IMPL_RVV_INTERLEAVED(int8x16, schar)
-OPENCV_HAL_IMPL_RVV_INTERLEAVED(uint16x8, ushort)
-OPENCV_HAL_IMPL_RVV_INTERLEAVED(int16x8, short)
-OPENCV_HAL_IMPL_RVV_INTERLEAVED(uint32x4, unsigned)
-OPENCV_HAL_IMPL_RVV_INTERLEAVED(int32x4, int)
-OPENCV_HAL_IMPL_RVV_INTERLEAVED(float32x4, float)
-OPENCV_HAL_IMPL_RVV_INTERLEAVED(uint64x2, uint64)
-OPENCV_HAL_IMPL_RVV_INTERLEAVED(int64x2, int64)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_INTERLEAVED(float64x2, double)
-#endif
-
-//////////// PopCount ////////////
-
-static const unsigned char popCountTable[] =
-{
-    0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    4, 5, 5, 6, 5, 6, 6, 7, 5, 6, 6, 7, 6, 7, 7, 8,
-};
-
-#define OPENCV_HAL_IMPL_RVV_POPCOUNT_OP(_rTpvec, _Tpvec, _rTp, _Tp, suffix) \
-inline _rTpvec v_popcount(const _Tpvec& a) \
-{ \
-    uchar ptra[16] = {0}; \
-    v_store(ptra, v_reinterpret_as_u8(a)); \
-    _rTp ptr[_Tpvec::nlanes] = {0}; \
-    v_store(ptr, v_setzero_##suffix()); \
-    for (int i = 0; i < _Tpvec::nlanes*(int)sizeof(_Tp); i++) \
-        ptr[i/sizeof(_Tp)] += popCountTable[ptra[i]]; \
-    return v_load(ptr); \
-}
-
-OPENCV_HAL_IMPL_RVV_POPCOUNT_OP(v_uint8x16, v_uint8x16, uchar, uchar, u8)
-OPENCV_HAL_IMPL_RVV_POPCOUNT_OP(v_uint8x16, v_int8x16, uchar, schar, u8)
-OPENCV_HAL_IMPL_RVV_POPCOUNT_OP(v_uint16x8, v_uint16x8, ushort, ushort, u16)
-OPENCV_HAL_IMPL_RVV_POPCOUNT_OP(v_uint16x8, v_int16x8, ushort, short, u16)
-OPENCV_HAL_IMPL_RVV_POPCOUNT_OP(v_uint32x4, v_uint32x4, unsigned, unsigned, u32)
-OPENCV_HAL_IMPL_RVV_POPCOUNT_OP(v_uint32x4, v_int32x4, unsigned, int, u32)
-OPENCV_HAL_IMPL_RVV_POPCOUNT_OP(v_uint64x2, v_uint64x2, uint64, uint64, u64)
-OPENCV_HAL_IMPL_RVV_POPCOUNT_OP(v_uint64x2, v_int64x2, uint64, int64, u64)
-
-//////////// SignMask ////////////
-
-#define OPENCV_HAL_IMPL_RVV_SIGNMASK_OP(_Tpvec, _Tp, suffix, vl, shift) \
-inline int v_signmask(const _Tpvec& a) \
-{ \
-    int mask = 0; \
-    _Tpvec tmp = _Tpvec(vsrl_vx_##suffix##m1(a, shift, vl)); \
-    for( int i = 0; i < _Tpvec::nlanes; i++ ) \
-        mask |= (int)(tmp.val[i]) << i; \
-    return mask; \
-}
-
-OPENCV_HAL_IMPL_RVV_SIGNMASK_OP(v_uint8x16, uchar, u8, 16, 7)
-OPENCV_HAL_IMPL_RVV_SIGNMASK_OP(v_uint16x8, ushort, u16, 8, 15)
-OPENCV_HAL_IMPL_RVV_SIGNMASK_OP(v_uint32x4, unsigned, u32, 4, 31)
-OPENCV_HAL_IMPL_RVV_SIGNMASK_OP(v_uint64x2, uint64, u64, 2, 63)
-
-inline int v_signmask(const v_int8x16& a)
-{ return v_signmask(v_reinterpret_as_u8(a)); }
-inline int v_signmask(const v_int16x8& a)
-{ return v_signmask(v_reinterpret_as_u16(a)); }
-inline int v_signmask(const v_int32x4& a)
-{ return v_signmask(v_reinterpret_as_u32(a)); }
-inline int v_signmask(const v_float32x4& a)
-{ return v_signmask(v_reinterpret_as_u32(a)); }
-inline int v_signmask(const v_int64x2& a)
-{ return v_signmask(v_reinterpret_as_u64(a)); }
-#if CV_SIMD128_64F
-inline int v_signmask(const v_float64x2& a)
-{ return v_signmask(v_reinterpret_as_u64(a)); }
-#endif
-
-
-//////////// Scan forward ////////////
-
-#define OPENCV_HAL_IMPL_RVV_SCAN_FORWOARD_OP(_Tpvec, _Tp, suffix) \
-inline int v_scan_forward(const _Tpvec& a) \
-{ \
-    _Tp ptr[_Tpvec::nlanes] = {0}; \
-    v_store(ptr, v_reinterpret_as_##suffix(a)); \
-    for (int i = 0; i < _Tpvec::nlanes; i++) \
-        if(int(ptr[i]) < 0) \
-            return i; \
-    return 0; \
-}
-
-OPENCV_HAL_IMPL_RVV_SCAN_FORWOARD_OP(v_uint8x16, uchar, u8)
-OPENCV_HAL_IMPL_RVV_SCAN_FORWOARD_OP(v_int8x16, schar, s8)
-OPENCV_HAL_IMPL_RVV_SCAN_FORWOARD_OP(v_uint16x8, ushort, u16)
-OPENCV_HAL_IMPL_RVV_SCAN_FORWOARD_OP(v_int16x8, short, s16)
-OPENCV_HAL_IMPL_RVV_SCAN_FORWOARD_OP(v_uint32x4, unsigned, u32)
-OPENCV_HAL_IMPL_RVV_SCAN_FORWOARD_OP(v_int32x4, int, s32)
-OPENCV_HAL_IMPL_RVV_SCAN_FORWOARD_OP(v_float32x4, float, f32)
-OPENCV_HAL_IMPL_RVV_SCAN_FORWOARD_OP(v_uint64x2, uint64, u64)
-OPENCV_HAL_IMPL_RVV_SCAN_FORWOARD_OP(v_int64x2, int64, s64)
-#if CV_SIMD128_64F
-OPENCV_HAL_IMPL_RVV_SCAN_FORWOARD_OP(v_float64x2, double, f64)
-#endif
-
-//////////// Pack triplets ////////////
-
-// use reinterpret instead of c-style casting.
-#ifndef __clang__
-inline v_int8x16 v_pack_triplets(const v_int8x16& vec)
-{
-    uint64 ptr[2] = {0x0908060504020100, 0xFFFFFFFF0E0D0C0A};
-    return v_int8x16((vint8m1_t)vrgather_vv_u8m1((vuint8m1_t)vint8m1_t(vec), (vuint8m1_t)vle64_v_u64m1(ptr, 2), 16));
-}
-inline v_uint8x16 v_pack_triplets(const v_uint8x16& vec)
-{
-    return v_reinterpret_as_u8(v_pack_triplets(v_reinterpret_as_s8(vec)));
-}
-
-inline v_int16x8 v_pack_triplets(const v_int16x8& vec)
-{
-    uint64 ptr[2] = {0x0908060504020100, 0xFFFFFFFF0E0D0C0A};
-    return v_int16x8((vint16m1_t)vrgather_vv_u8m1((vuint8m1_t)vint16m1_t(vec), (vuint8m1_t)vle64_v_u64m1(ptr, 2), 16));
-}
-inline v_uint16x8 v_pack_triplets(const v_uint16x8& vec)
-{
-    return v_reinterpret_as_u16(v_pack_triplets(v_reinterpret_as_s16(vec)));
-}
-
-inline v_int32x4 v_pack_triplets(const v_int32x4& vec) { return vec; }
-inline v_uint32x4 v_pack_triplets(const v_uint32x4& vec) { return vec; }
-inline v_float32x4 v_pack_triplets(const v_float32x4& vec) { return vec; }
-
-#else
-
-inline v_int8x16 v_pack_triplets(const v_int8x16& vec)
-{
-    uint64 ptr[2] = {0x0908060504020100, 0xFFFFFFFF0E0D0C0A};
-    return v_int8x16(vreinterpret_i8m1(vrgather_vv_u8m1(v_reinterpret_as_u8(vec), vreinterpret_u8m1(vle64_v_u64m1(ptr, 2)), 16)));
-}
-inline v_uint8x16 v_pack_triplets(const v_uint8x16& vec)
-{
-    return v_reinterpret_as_u8(v_pack_triplets(v_reinterpret_as_s8(vec)));
-}
-
-inline v_int16x8 v_pack_triplets(const v_int16x8& vec)
-{
-    uint64 ptr[2] = {0x0908060504020100, 0xFFFFFFFF0E0D0C0A};
-    return v_int16x8(v_reinterpret_as_s16(v_uint8x16(vrgather_vv_u8m1(v_reinterpret_as_u8(vec), vreinterpret_u8m1(vle64_v_u64m1(ptr, 2)), 16))));
-}
-inline v_uint16x8 v_pack_triplets(const v_uint16x8& vec)
-{
-    return v_reinterpret_as_u16(v_pack_triplets(v_reinterpret_as_s16(vec)));
-}
-
-inline v_int32x4 v_pack_triplets(const v_int32x4& vec) { return vec; }
-inline v_uint32x4 v_pack_triplets(const v_uint32x4& vec) { return vec; }
-inline v_float32x4 v_pack_triplets(const v_float32x4& vec) { return vec; }
-
-#endif
-
-////// FP16 support ///////
-
-#if CV_FP16
-inline v_float32x4 v_load_expand(const float16_t* ptr)
-{
-    return v_float32x4(vfwcvt_f_f_v_f32m1(vle16_v_f16mf2(ptr, 4), 4));
-}
-
-inline void v_pack_store(float16_t* ptr, const v_float32x4& v)
-{
-    vse16_v_f16mf2(ptr, vfncvt_f_f_w_f16mf2(v, 4), 4);
-}
-#else
-inline v_float32x4 v_load_expand(const float16_t* ptr)
-{
-    const int N = 4;
-    float buf[N];
-    for( int i = 0; i < N; i++ ) buf[i] = (float)ptr[i];
-    return v_load(buf);
-}
-
-inline void v_pack_store(float16_t* ptr, const v_float32x4& v)
-{
-    const int N = 4;
-    float buf[N];
-    v_store(buf, v);
-    for( int i = 0; i < N; i++ ) ptr[i] = float16_t(buf[i]);
-}
-#endif
-
-////////////// Rounding //////////////
-
-inline v_int32x4 v_round(const v_float32x4& a)
-{
-    return v_int32x4(vfcvt_x_f_v_i32m1(a, 4));
-}
-
-inline v_int32x4 v_floor(const v_float32x4& a)
-{
-    v_float32x4 ZP5 = v_setall_f32(0.5f);
-    v_float32x4 t = a - ZP5;
-    return v_int32x4(vfcvt_x_f_v_i32m1(t, 4));
-}
-
-inline v_int32x4 v_ceil(const v_float32x4& a)
-{
-    v_float32x4 ZP5 = v_setall_f32(0.5f);
-    v_float32x4 t = a + ZP5;
-    return v_int32x4(vfcvt_x_f_v_i32m1(t, 4));
-}
-
-inline v_int32x4 v_trunc(const v_float32x4& a)
-{
-    return v_int32x4(vfcvt_rtz_x_f_v_i32m1(a, 4));
-}
-#if CV_SIMD128_64F
-inline v_int32x4 v_round(const v_float64x2& a)
-{
-    double arr[4] = {a.val[0], a.val[1], 0, 0};
-    vfloat64m2_t tmp = vle64_v_f64m2(arr, 4);
-    return v_int32x4(vfncvt_x_f_w_i32m1(tmp, 4));
-}
-
-inline v_int32x4 v_round(const v_float64x2& a, const v_float64x2& b)
-{
-    double arr[4] = {a.val[0], a.val[1], b.val[0], b.val[1]};
-    vfloat64m2_t tmp = vle64_v_f64m2(arr, 4);
-    return v_int32x4(vfncvt_x_f_w_i32m1(tmp, 4));
-}
-
-inline v_int32x4 v_floor(const v_float64x2& a)
-{
-    double arr[4] = {a.val[0]-0.5f, a.val[1]-0.5f, 0, 0};
-    vfloat64m2_t tmp = vle64_v_f64m2(arr, 4);
-    return v_int32x4(vfncvt_x_f_w_i32m1(tmp, 4));
-}
-
-inline v_int32x4 v_ceil(const v_float64x2& a)
-{
-    double arr[4] = {a.val[0]+0.5f, a.val[1]+0.5f, 0, 0};
-    vfloat64m2_t tmp = vle64_v_f64m2(arr, 4);
-    return v_int32x4(vfncvt_x_f_w_i32m1(tmp, 4));
-}
-
-inline v_int32x4 v_trunc(const v_float64x2& a)
-{
-    double arr[4] = {a.val[0], a.val[1], 0, 0};
-    vfloat64m2_t tmp = vle64_v_f64m2(arr, 4);
-    return v_int32x4(vfncvt_rtz_x_f_w_i32m1(tmp, 4));
-}
-#endif
-
-
-//////// Dot Product ////////
-
-// 16 >> 32
-inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b)
-{
-    int ptr[8] = {0};
-    v_int32x4 t1, t2;
-    vse32_v_i32m2(ptr, vwmul_vv_i32m2(a, b, 8), 8);
-    v_load_deinterleave(ptr, t1, t2);
-    return t1 + t2;
-}
-inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b, const v_int32x4& c)
-{
-    int ptr[8] = {0};
-    v_int32x4 t1, t2;
-    vse32_v_i32m2(ptr, vwmul_vv_i32m2(a, b, 8), 8);
-    v_load_deinterleave(ptr, t1, t2);
-    return t1 + t2 + c;
-}
-
-// 32 >> 64
-inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b)
-{
-    int64 ptr[4] = {0};
-    v_int64x2 t1, t2;
-    vse64_v_i64m2(ptr, vwmul_vv_i64m2(a, b, 4), 4);
-    v_load_deinterleave(ptr, t1, t2);
-    return t1 + t2;
-}
-inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b, const v_int64x2& c)
-{
-    int64 ptr[4] = {0};
-    v_int64x2 t1, t2;
-    vse64_v_i64m2(ptr, vwmul_vv_i64m2(a, b, 4), 4);
-    v_load_deinterleave(ptr, t1, t2);
-    return t1 + t2 + c;
-}
-
-// 8 >> 32
-inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b)
-{
-    unsigned ptr[16] = {0};
-    v_uint32x4 t1, t2, t3, t4;
-    vse32_v_u32m4(ptr, vwcvtu_x_x_v_u32m4(vwmulu_vv_u16m2(a, b, 16), 16), 16);
-    v_load_deinterleave(ptr, t1, t2, t3, t4);
-    return t1 + t2 + t3 + t4;
-}
-inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b,
-                                   const v_uint32x4& c)
-{
-    unsigned ptr[16] = {0};
-    v_uint32x4 t1, t2, t3, t4;
-    vse32_v_u32m4(ptr, vwcvtu_x_x_v_u32m4(vwmulu_vv_u16m2(a, b, 16), 16), 16);
-    v_load_deinterleave(ptr, t1, t2, t3, t4);
-    return t1 + t2 + t3 + t4 + c;
-}
-
-inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b)
-{
-    int ptr[16] = {0};
-    v_int32x4 t1, t2, t3, t4;
-    vse32_v_i32m4(ptr, vwcvt_x_x_v_i32m4(vwmul_vv_i16m2(a, b, 16), 16), 16);
-    v_load_deinterleave(ptr, t1, t2, t3, t4);
-    return t1 + t2 + t3 + t4;
-}
-inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b,
-                                  const v_int32x4& c)
-{
-    int ptr[16] = {0};
-    v_int32x4 t1, t2, t3, t4;
-    vse32_v_i32m4(ptr, vwcvt_x_x_v_i32m4(vwmul_vv_i16m2(a, b, 16), 16), 16);
-    v_load_deinterleave(ptr, t1, t2, t3, t4);
-    return t1 + t2 + t3 + t4 + c;
-}
-
-// 16 >> 64
-inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b)
-{
-    uint64 ptr[8] = {0};
-    v_uint64x2 t1, t2, t3, t4;
-    vse64_v_u64m4(ptr, vwcvtu_x_x_v_u64m4(vwmulu_vv_u32m2(a, b, 8), 8), 8);
-    v_load_deinterleave(ptr, t1, t2, t3, t4);
-    return t1 + t2 + t3 + t4;
-}
-inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b, const v_uint64x2& c)
-{
-    uint64 ptr[8] = {0};
-    v_uint64x2 t1, t2, t3, t4;
-    vse64_v_u64m4(ptr, vwcvtu_x_x_v_u64m4(vwmulu_vv_u32m2(a, b, 8), 8), 8);
-    v_load_deinterleave(ptr, t1, t2, t3, t4);
-    return t1 + t2 + t3 + t4 + c;
-}
-
-inline v_int64x2 v_dotprod_expand(const v_int16x8& a, const v_int16x8& b)
-{
-    int64 ptr[8] = {0};
-    v_int64x2 t1, t2, t3, t4;
-    vse64_v_i64m4(ptr, vwcvt_x_x_v_i64m4(vwmul_vv_i32m2(a, b, 8), 8), 8);
-    v_load_deinterleave(ptr, t1, t2, t3, t4);
-    return t1 + t2 + t3 + t4;
-}
-inline v_int64x2 v_dotprod_expand(const v_int16x8& a, const v_int16x8& b,
-                                  const v_int64x2& c)
-{
-    int64 ptr[8] = {0};
-    v_int64x2 t1, t2, t3, t4;
-    vse64_v_i64m4(ptr, vwcvt_x_x_v_i64m4(vwmul_vv_i32m2(a, b, 8), 8), 8);
-    v_load_deinterleave(ptr, t1, t2, t3, t4);
-    return t1 + t2 + t3 + t4 + c;
-}
-
-// 32 >> 64f
-#if CV_SIMD128_64F
-inline v_float64x2 v_dotprod_expand(const v_int32x4& a, const v_int32x4& b)
-{ return v_cvt_f64(v_dotprod(a, b)); }
-inline v_float64x2 v_dotprod_expand(const v_int32x4& a,   const v_int32x4& b,
-                                    const v_float64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-#endif
-
-//////// Fast Dot Product ////////
-
-// 16 >> 32
-inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b)
-{
-    int ptr[8] = {0};
-    vse32_v_i32m2(ptr, vwmul_vv_i32m2(a, b, 8), 8);
-    v_int32x4 t1 = v_load(ptr);
-    v_int32x4 t2 = v_load(ptr+4);
-    return t1 + t2;
-}
-inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b, const v_int32x4& c)
-{
-    int ptr[8] = {0};
-    vse32_v_i32m2(ptr, vwmul_vv_i32m2(a, b, 8), 8);
-    v_int32x4 t1 = v_load(ptr);
-    v_int32x4 t2 = v_load(ptr+4);
-    return t1 + t2 + c;
-}
-
-// 32 >> 64
-inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b)
-{
-    int64 ptr[4] = {0};
-    vse64_v_i64m2(ptr, vwmul_vv_i64m2(a, b, 4), 4);
-    v_int64x2 t1 = v_load(ptr);
-    v_int64x2 t2 = v_load(ptr+2);
-    return t1 + t2;
-}
-inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b, const v_int64x2& c)
-{
-    int64 ptr[4] = {0};
-    vse64_v_i64m2(ptr, vwmul_vv_i64m2(a, b, 4), 4);
-    v_int64x2 t1 = v_load(ptr);
-    v_int64x2 t2 = v_load(ptr+2);
-    return t1 + t2 + c;
-}
-
-
-// 8 >> 32
-inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b)
-{
-    unsigned ptr[16] = {0};
-    vse32_v_u32m4(ptr, vwcvtu_x_x_v_u32m4(vwmulu_vv_u16m2(a, b, 16), 16), 16);
-    v_uint32x4 t1 = v_load(ptr);
-    v_uint32x4 t2 = v_load(ptr+4);
-    v_uint32x4 t3 = v_load(ptr+8);
-    v_uint32x4 t4 = v_load(ptr+12);
-    return t1 + t2 + t3 + t4;
-}
-inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b, const v_uint32x4& c)
-{
-    unsigned ptr[16] = {0};
-    vse32_v_u32m4(ptr, vwcvtu_x_x_v_u32m4(vwmulu_vv_u16m2(a, b, 16), 16), 16);
-    v_uint32x4 t1 = v_load(ptr);
-    v_uint32x4 t2 = v_load(ptr+4);
-    v_uint32x4 t3 = v_load(ptr+8);
-    v_uint32x4 t4 = v_load(ptr+12);
-    return t1 + t2 + t3 + t4 + c;
-}
-inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b)
-{
-    int ptr[16] = {0};
-    vse32_v_i32m4(ptr, vwcvt_x_x_v_i32m4(vwmul_vv_i16m2(a, b, 16), 16), 16);
-    v_int32x4 t1 = v_load(ptr);
-    v_int32x4 t2 = v_load(ptr+4);
-    v_int32x4 t3 = v_load(ptr+8);
-    v_int32x4 t4 = v_load(ptr+12);
-    return t1 + t2 + t3 + t4;
-}
-inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b, const v_int32x4& c)
-{
-    int ptr[16] = {0};
-    vse32_v_i32m4(ptr, vwcvt_x_x_v_i32m4(vwmul_vv_i16m2(a, b, 16), 16), 16);
-    v_int32x4 t1 = v_load(ptr);
-    v_int32x4 t2 = v_load(ptr+4);
-    v_int32x4 t3 = v_load(ptr+8);
-    v_int32x4 t4 = v_load(ptr+12);
-    return t1 + t2 + t3 + t4 + c;
-}
-
-// 16 >> 64
-inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b)
-{
-    uint64 ptr[8] = {0};
-    vse64_v_u64m4(ptr, vwcvtu_x_x_v_u64m4(vwmulu_vv_u32m2(a, b, 8), 8), 8);
-    v_uint64x2 t1 = v_load(ptr);
-    v_uint64x2 t2 = v_load(ptr+2);
-    v_uint64x2 t3 = v_load(ptr+4);
-    v_uint64x2 t4 = v_load(ptr+6);
-    return t1 + t2 + t3 + t4;
-}
-inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b, const v_uint64x2& c)
-{
-    uint64 ptr[8] = {0};
-    vse64_v_u64m4(ptr, vwcvtu_x_x_v_u64m4(vwmulu_vv_u32m2(a, b, 8), 8), 8);
-    v_uint64x2 t1 = v_load(ptr);
-    v_uint64x2 t2 = v_load(ptr+2);
-    v_uint64x2 t3 = v_load(ptr+4);
-    v_uint64x2 t4 = v_load(ptr+6);
-    return t1 + t2 + t3 + t4 + c;
-}
-inline v_int64x2 v_dotprod_expand_fast(const v_int16x8& a, const v_int16x8& b)
-{
-    int64 ptr[8] = {0};
-    vse64_v_i64m4(ptr, vwcvt_x_x_v_i64m4(vwmul_vv_i32m2(a, b, 8), 8), 8);
-    v_int64x2 t1 = v_load(ptr);
-    v_int64x2 t2 = v_load(ptr+2);
-    v_int64x2 t3 = v_load(ptr+4);
-    v_int64x2 t4 = v_load(ptr+6);
-    return t1 + t2 + t3 + t4;
-}
-inline v_int64x2 v_dotprod_expand_fast(const v_int16x8& a, const v_int16x8& b, const v_int64x2& c)
-{
-    int64 ptr[8] = {0};
-    vse64_v_i64m4(ptr, vwcvt_x_x_v_i64m4(vwmul_vv_i32m2(a, b, 8), 8), 8);
-    v_int64x2 t1 = v_load(ptr);
-    v_int64x2 t2 = v_load(ptr+2);
-    v_int64x2 t3 = v_load(ptr+4);
-    v_int64x2 t4 = v_load(ptr+6);
-    return t1 + t2 + t3 + t4 + c;
-}
-
-// 32 >> 64f
-#if CV_SIMD128_64F
-inline v_float64x2 v_dotprod_expand_fast(const v_int32x4& a, const v_int32x4& b)
-{ return v_cvt_f64(v_dotprod_fast(a, b)); }
-inline v_float64x2 v_dotprod_expand_fast(const v_int32x4& a, const v_int32x4& b, const v_float64x2& c)
-{ return v_dotprod_expand_fast(a, b) + c; }
-#endif
-
-
-inline v_float32x4 v_matmul(const v_float32x4& v, const v_float32x4& m0,
-                            const v_float32x4& m1, const v_float32x4& m2,
-                            const v_float32x4& m3)
-{
-    vfloat32m1_t res = vfmul_vf_f32m1(m0, v_extract_n<0>(v), 4);
-    res = vfmacc_vf_f32m1(res, v_extract_n<1>(v), m1, 4);
-    res = vfmacc_vf_f32m1(res, v_extract_n<2>(v), m2, 4);
-    res = vfmacc_vf_f32m1(res, v_extract_n<3>(v), m3, 4);
-    return v_float32x4(res);
-}
-
-inline v_float32x4 v_matmuladd(const v_float32x4& v, const v_float32x4& m0,
-                               const v_float32x4& m1, const v_float32x4& m2,
-                               const v_float32x4& a)
-{
-    vfloat32m1_t res = vfmul_vf_f32m1(m0, v_extract_n<0>(v), 4);
-    res = vfmacc_vf_f32m1(res, v_extract_n<1>(v), m1, 4);
-    res = vfmacc_vf_f32m1(res, v_extract_n<2>(v), m2, 4);
-    return v_float32x4(res) + a;
-}
-
-#define OPENCV_HAL_IMPL_RVV_MUL_EXPAND(_Tpvec, _Tpwvec, _Tpw, suffix, wmul, width, vl, hvl) \
-inline void v_mul_expand(const _Tpvec& a, const _Tpvec& b, _Tpwvec& c, _Tpwvec& d) \
-{ \
-    _Tpw ptr[_Tpwvec::nlanes*2] = {0}; \
-    vse##width##_v_##suffix##m2(ptr, wmul(a, b, vl), vl); \
-    c = _Tpwvec(vle##width##_v_##suffix##m1(ptr, hvl)); \
-    d = _Tpwvec(vle##width##_v_##suffix##m1(ptr+_Tpwvec::nlanes, hvl)); \
-}
-
-OPENCV_HAL_IMPL_RVV_MUL_EXPAND(v_uint8x16, v_uint16x8, ushort, u16, vwmulu_vv_u16m2, 16, 16, 8)
-OPENCV_HAL_IMPL_RVV_MUL_EXPAND(v_int8x16, v_int16x8, short, i16, vwmul_vv_i16m2, 16, 16, 8)
-OPENCV_HAL_IMPL_RVV_MUL_EXPAND(v_uint16x8, v_uint32x4, unsigned, u32, vwmulu_vv_u32m2, 32, 8, 4)
-OPENCV_HAL_IMPL_RVV_MUL_EXPAND(v_int16x8, v_int32x4, int, i32, vwmul_vv_i32m2, 32, 8, 4)
-OPENCV_HAL_IMPL_RVV_MUL_EXPAND(v_uint32x4, v_uint64x2, uint64, u64, vwmulu_vv_u64m2, 64, 4, 2)
-
-
-inline v_int16x8 v_mul_hi(const v_int16x8& a, const v_int16x8& b)
-{
-    return v_int16x8(vnsra_wx_i16m1(vwmul_vv_i32m2(a, b, 8), 16, 8));
-}
-inline v_uint16x8 v_mul_hi(const v_uint16x8& a, const v_uint16x8& b)
-{
-    return v_uint16x8(vnsrl_wx_u16m1(vwmulu_vv_u32m2(a, b, 8), 16, 8));
-}
-
-
-//////// Saturating Multiply ////////
-
-#define OPENCV_HAL_IMPL_RVV_MUL_SAT(_Tpvec, _wTpvec) \
-inline _Tpvec operator * (const _Tpvec& a, const _Tpvec& b) \
-{ \
-    _wTpvec c, d; \
-    v_mul_expand(a, b, c, d); \
-    return v_pack(c, d); \
-} \
-inline _Tpvec& operator *= (_Tpvec& a, const _Tpvec& b) \
-{ \
-    a = a * b; \
-    return a; \
-}
-
-OPENCV_HAL_IMPL_RVV_MUL_SAT(v_uint8x16, v_uint16x8)
-OPENCV_HAL_IMPL_RVV_MUL_SAT(v_int8x16, v_int16x8)
-OPENCV_HAL_IMPL_RVV_MUL_SAT(v_uint16x8, v_uint32x4)
-OPENCV_HAL_IMPL_RVV_MUL_SAT(v_int16x8, v_int32x4)
-
-
-inline void v_cleanup() {}
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END
-
-
-}
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_rvv071.hpp b/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_rvv071.hpp
deleted file mode 100644
index 2bdc622f..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_rvv071.hpp
+++ /dev/null
@@ -1,2545 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html
-
-// Copyright (C) 2015, PingTouGe Semiconductor Co., Ltd., all rights reserved.
-
-#ifndef OPENCV_HAL_INTRIN_RISCVV_HPP
-#define OPENCV_HAL_INTRIN_RISCVV_HPP
-
-#include <float.h>
-#include <algorithm>
-#include "opencv2/core/utility.hpp"
-
-namespace cv
-{
-
-//! @cond IGNORED
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
-
-#define CV_SIMD128 1
-#define CV_SIMD128_64F 1
-//////////// Types ////////////
-struct v_uint8x16
-{
-    typedef uchar lane_type;
-    enum { nlanes = 16 };
-
-    v_uint8x16() {}
-    explicit v_uint8x16(vuint8m1_t v) : val(v) {}
-    v_uint8x16(uchar v0, uchar v1, uchar v2, uchar v3, uchar v4, uchar v5, uchar v6, uchar v7,
-               uchar v8, uchar v9, uchar v10, uchar v11, uchar v12, uchar v13, uchar v14, uchar v15)
-    {
-        uchar v[] = {v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15};
-        val = (vuint8m1_t)vle_v_u8m1((unsigned char*)v, 16);
-    }
-    uchar get0() const
-    {
-        return vmv_x_s_u8m1_u8(val, 16);
-    }
-
-    vuint8m1_t val;
-};
-
-struct v_int8x16
-{
-    typedef schar lane_type;
-    enum { nlanes = 16 };
-
-    v_int8x16() {}
-    explicit v_int8x16(vint8m1_t v) : val(v) {}
-    v_int8x16(schar v0, schar v1, schar v2, schar v3, schar v4, schar v5, schar v6, schar v7,
-               schar v8, schar v9, schar v10, schar v11, schar v12, schar v13, schar v14, schar v15)
-    {
-        schar v[] = {v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15};
-        val = (vint8m1_t)vle_v_i8m1((schar*)v, 16);
-    }
-    schar get0() const
-    {
-        return vmv_x_s_i8m1_i8(val, 16);
-    }
-
-    vint8m1_t val;
-};
-
-struct v_uint16x8
-{
-    typedef ushort lane_type;
-    enum { nlanes = 8 };
-
-    v_uint16x8() {}
-    explicit v_uint16x8(vuint16m1_t v) : val(v) {}
-    v_uint16x8(ushort v0, ushort v1, ushort v2, ushort v3, ushort v4, ushort v5, ushort v6, ushort v7)
-    {
-        ushort v[] = {v0, v1, v2, v3, v4, v5, v6, v7};
-        val = (vuint16m1_t)vle_v_u16m1((unsigned short*)v, 8);
-    }
-    ushort get0() const
-    {
-        return vmv_x_s_u16m1_u16(val, 8);
-    }
-
-    vuint16m1_t val;
-};
-
-struct v_int16x8
-{
-    typedef short lane_type;
-    enum { nlanes = 8 };
-
-    v_int16x8() {}
-    explicit v_int16x8(vint16m1_t v) : val(v) {}
-    v_int16x8(short v0, short v1, short v2, short v3, short v4, short v5, short v6, short v7)
-    {
-        short v[] = {v0, v1, v2, v3, v4, v5, v6, v7};
-        val = (vint16m1_t)vle_v_i16m1((signed short*)v, 8);
-    }
-    short get0() const
-    {
-        return vmv_x_s_i16m1_i16(val, 8);
-    }
-
-    vint16m1_t val;
-};
-
-struct v_uint32x4
-{
-    typedef unsigned lane_type;
-    enum { nlanes = 4 };
-
-    v_uint32x4() {}
-    explicit v_uint32x4(vuint32m1_t v) : val(v) {}
-    v_uint32x4(unsigned v0, unsigned v1, unsigned v2, unsigned v3)
-    {
-        unsigned v[] = {v0, v1, v2, v3};
-        val = (vuint32m1_t)vle_v_u32m1((unsigned int*)v, 4);
-    }
-    unsigned get0() const
-    {
-        return vmv_x_s_u32m1_u32(val, 4);
-    }
-
-    vuint32m1_t val;
-};
-
-struct v_int32x4
-{
-    typedef int lane_type;
-    enum { nlanes = 4 };
-
-    v_int32x4() {}
-    explicit v_int32x4(vint32m1_t v) : val(v) {}
-    v_int32x4(int v0, int v1, int v2, int v3)
-    {
-        int v[] = {v0, v1, v2, v3};
-        val = (vint32m1_t)vle_v_i32m1((signed int*)v, 4);
-    }
-    int get0() const
-    {
-        return vmv_x_s_i32m1_i32(val, 4);
-    }
-    vint32m1_t val;
-};
-
-struct v_float32x4
-{
-    typedef float lane_type;
-    enum { nlanes = 4 };
-
-    v_float32x4() {}
-    explicit v_float32x4(vfloat32m1_t v) : val(v) {}
-    v_float32x4(float v0, float v1, float v2, float v3)
-    {
-        float v[] = {v0, v1, v2, v3};
-        val = (vfloat32m1_t)vle_v_f32m1((float*)v, 4);
-    }
-    float get0() const
-    {
-        return vfmv_f_s_f32m1_f32(val, 4);
-    }
-    vfloat32m1_t val;
-};
-
-struct v_uint64x2
-{
-    typedef uint64 lane_type;
-    enum { nlanes = 2 };
-
-    v_uint64x2() {}
-    explicit v_uint64x2(vuint64m1_t v) : val(v) {}
-    v_uint64x2(uint64 v0, uint64 v1)
-    {
-        uint64 v[] = {v0, v1};
-        val = (vuint64m1_t)vle_v_u64m1((unsigned long*)v, 2);
-    }
-    uint64 get0() const
-    {
-        return vmv_x_s_u64m1_u64(val, 2);
-    }
-    vuint64m1_t val;
-};
-
-struct v_int64x2
-{
-    typedef int64 lane_type;
-    enum { nlanes = 2 };
-
-    v_int64x2() {}
-    explicit v_int64x2(vint64m1_t v) : val(v) {}
-    v_int64x2(int64 v0, int64 v1)
-    {
-        int64 v[] = {v0, v1};
-        val = (vint64m1_t)vle_v_i64m1((long*)v, 2);
-    }
-    int64 get0() const
-    {
-        return vmv_x_s_i64m1_i64(val, 2);
-    }
-    vint64m1_t val;
-};
-
-struct v_float64x2
-{
-    typedef double lane_type;
-    enum { nlanes = 2 };
-
-    v_float64x2() {}
-    explicit v_float64x2(vfloat64m1_t v) : val(v) {}
-    v_float64x2(double v0, double v1)
-    {
-        double v[] = {v0, v1};
-        val = (vfloat64m1_t)vle_v_f64m1((double*)v, 2);
-    }
-    double get0() const
-    {
-        return vfmv_f_s_f64m1_f64(val, 2);
-    }
-    vfloat64m1_t val;
-};
-
-#define OPENCV_HAL_IMPL_RISCVV_INIT(_Tpv, _Tp, suffix) \
-inline _Tp##m1_t vreinterpretq_##suffix##_##suffix(_Tp##m1_t v) { return v; } \
-inline v_uint8x16 v_reinterpret_as_u8(const v_##_Tpv& v) { return v_uint8x16((vuint8m1_t)(v.val)); } \
-inline v_int8x16 v_reinterpret_as_s8(const v_##_Tpv& v) { return v_int8x16((vint8m1_t)(v.val)); } \
-inline v_uint16x8 v_reinterpret_as_u16(const v_##_Tpv& v) { return v_uint16x8((vuint16m1_t)(v.val)); } \
-inline v_int16x8 v_reinterpret_as_s16(const v_##_Tpv& v) { return v_int16x8((vint16m1_t)(v.val)); } \
-inline v_uint32x4 v_reinterpret_as_u32(const v_##_Tpv& v) { return v_uint32x4((vuint32m1_t)(v.val)); } \
-inline v_int32x4 v_reinterpret_as_s32(const v_##_Tpv& v) { return v_int32x4((vint32m1_t)(v.val)); } \
-inline v_uint64x2 v_reinterpret_as_u64(const v_##_Tpv& v) { return v_uint64x2((vuint64m1_t)(v.val)); } \
-inline v_int64x2 v_reinterpret_as_s64(const v_##_Tpv& v) { return v_int64x2((vint64m1_t)(v.val)); } \
-inline v_float32x4 v_reinterpret_as_f32(const v_##_Tpv& v) { return v_float32x4((vfloat32m1_t)(v.val)); }\
-inline v_float64x2 v_reinterpret_as_f64(const v_##_Tpv& v) { return v_float64x2((vfloat64m1_t)(v.val)); }
-
-
-OPENCV_HAL_IMPL_RISCVV_INIT(uint8x16, vuint8, u8)
-OPENCV_HAL_IMPL_RISCVV_INIT(int8x16, vint8, s8)
-OPENCV_HAL_IMPL_RISCVV_INIT(uint16x8, vuint16, u16)
-OPENCV_HAL_IMPL_RISCVV_INIT(int16x8, vint16, s16)
-OPENCV_HAL_IMPL_RISCVV_INIT(uint32x4, vuint32, u32)
-OPENCV_HAL_IMPL_RISCVV_INIT(int32x4, vint32, s32)
-OPENCV_HAL_IMPL_RISCVV_INIT(uint64x2, vuint64, u64)
-OPENCV_HAL_IMPL_RISCVV_INIT(int64x2, vint64, s64)
-OPENCV_HAL_IMPL_RISCVV_INIT(float64x2, vfloat64, f64)
-OPENCV_HAL_IMPL_RISCVV_INIT(float32x4, vfloat32, f32)
-#define OPENCV_HAL_IMPL_RISCVV_INIT_SET(__Tp, _Tp, suffix, len, num) \
-inline v_##_Tp##x##num v_setzero_##suffix() { return v_##_Tp##x##num((v##_Tp##m1_t){0}); }     \
-inline v_##_Tp##x##num v_setall_##suffix(__Tp v) { return v_##_Tp##x##num(vmv_v_x_##len##m1(v, num)); }
-
-OPENCV_HAL_IMPL_RISCVV_INIT_SET(uchar, uint8, u8, u8, 16)
-OPENCV_HAL_IMPL_RISCVV_INIT_SET(char, int8, s8, i8, 16)
-OPENCV_HAL_IMPL_RISCVV_INIT_SET(ushort, uint16, u16, u16, 8)
-OPENCV_HAL_IMPL_RISCVV_INIT_SET(short, int16, s16, i16, 8)
-OPENCV_HAL_IMPL_RISCVV_INIT_SET(unsigned int, uint32, u32, u32, 4)
-OPENCV_HAL_IMPL_RISCVV_INIT_SET(int, int32, s32, i32, 4)
-OPENCV_HAL_IMPL_RISCVV_INIT_SET(unsigned long, uint64, u64, u64, 2)
-OPENCV_HAL_IMPL_RISCVV_INIT_SET(long, int64, s64, i64, 2)
-inline v_float32x4 v_setzero_f32() { return v_float32x4((vfloat32m1_t){0}); }
-inline v_float32x4 v_setall_f32(float v) { return v_float32x4(vfmv_v_f_f32m1(v, 4)); }
-
-inline v_float64x2 v_setzero_f64() { return v_float64x2(vfmv_v_f_f64m1(0, 2)); }
-inline v_float64x2 v_setall_f64(double v) { return v_float64x2(vfmv_v_f_f64m1(v, 2)); }
-
-
-#define OPENCV_HAL_IMPL_RISCVV_BIN_OP(bin_op, _Tpvec, intrin) \
-inline _Tpvec operator bin_op (const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(intrin(a.val, b.val)); \
-} \
-inline _Tpvec& operator bin_op##= (_Tpvec& a, const _Tpvec& b) \
-{ \
-    a.val = intrin(a.val, b.val); \
-    return a; \
-}
-
-#define OPENCV_HAL_IMPL_RISCVV_BIN_OPN(bin_op, _Tpvec, intrin, num) \
-inline _Tpvec operator bin_op (const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(intrin(a.val, b.val, num)); \
-} \
-inline _Tpvec& operator bin_op##= (_Tpvec& a, const _Tpvec& b) \
-{ \
-    a.val = intrin(a.val, b.val, num); \
-    return a; \
-}
-
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(+, v_uint8x16, vsaddu_vv_u8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(-, v_uint8x16, vssubu_vv_u8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(+, v_int8x16, vsadd_vv_i8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(-, v_int8x16, vssub_vv_i8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(+, v_uint16x8, vsaddu_vv_u16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(-, v_uint16x8, vssubu_vv_u16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(+, v_int16x8, vsadd_vv_i16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(-, v_int16x8, vssub_vv_i16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(+, v_int32x4, vsadd_vv_i32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(-, v_int32x4, vssub_vv_i32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(*, v_int32x4, vmul_vv_i32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(+, v_uint32x4, vadd_vv_u32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(-, v_uint32x4, vsub_vv_u32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(*, v_uint32x4, vmul_vv_u32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(+, v_int64x2, vsadd_vv_i64m1, 2)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(-, v_int64x2, vssub_vv_i64m1, 2)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(+, v_uint64x2, vadd_vv_u64m1, 2)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(-, v_uint64x2, vsub_vv_u64m1, 2)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(+, v_float32x4, vfadd_vv_f32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(-, v_float32x4, vfsub_vv_f32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(*, v_float32x4, vfmul_vv_f32m1, 4)
-inline v_float32x4 operator / (const v_float32x4& a, const v_float32x4& b)
-{
-    return v_float32x4(vfdiv_vv_f32m1(a.val, b.val, 4));
-}
-inline v_float32x4& operator /= (v_float32x4& a, const v_float32x4& b)
-{
-    a.val = vfdiv_vv_f32m1(a.val, b.val, 4);
-    return a;
-}
-
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(+, v_float64x2, vfadd_vv_f64m1, 2)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(-, v_float64x2, vfsub_vv_f64m1, 2)
-OPENCV_HAL_IMPL_RISCVV_BIN_OPN(*, v_float64x2, vfmul_vv_f64m1, 2)
-inline v_float64x2 operator / (const v_float64x2& a, const v_float64x2& b)
-{
-    return v_float64x2(vfdiv_vv_f64m1(a.val, b.val, 2));
-}
-inline v_float64x2& operator /= (v_float64x2& a, const v_float64x2& b)
-{
-    a.val = vfdiv_vv_f64m1(a.val, b.val, 2);
-    return a;
-}
-// TODO: exp, log, sin, cos
-
-#define OPENCV_HAL_IMPL_RISCVV_BIN_FUNC(_Tpvec, func, intrin) \
-inline _Tpvec func(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(intrin(a.val, b.val)); \
-}
-
-#define OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(_Tpvec, func, intrin, num) \
-inline _Tpvec func(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(intrin(a.val, b.val, num)); \
-}
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_uint8x16, v_min, vminu_vv_u8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_uint8x16, v_max, vmaxu_vv_u8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_int8x16, v_min, vmin_vv_i8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_int8x16, v_max, vmax_vv_i8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_uint16x8, v_min, vminu_vv_u16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_uint16x8, v_max, vmaxu_vv_u16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_int16x8, v_min, vmin_vv_i16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_int16x8, v_max, vmax_vv_i16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_uint32x4, v_min, vminu_vv_u32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_uint32x4, v_max, vmaxu_vv_u32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_int32x4, v_min, vmin_vv_i32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_int32x4, v_max, vmax_vv_i32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_float32x4, v_min, vfmin_vv_f32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_float32x4, v_max, vfmax_vv_f32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_float64x2, v_min, vfmin_vv_f64m1, 2)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_float64x2, v_max, vfmax_vv_f64m1, 2)
-
-inline v_float32x4 v_sqrt(const v_float32x4& x)
-{
-    return v_float32x4(vfsqrt_v_f32m1(x.val, 4));
-}
-
-inline v_float32x4 v_invsqrt(const v_float32x4& x)
-{
-    return v_float32x4(vfrdiv_vf_f32m1(vfsqrt_v_f32m1(x.val, 4), 1, 4));
-}
-
-inline v_float32x4 v_magnitude(const v_float32x4& a, const v_float32x4& b)
-{
-    v_float32x4 x(vfmacc_vv_f32m1(vfmul_vv_f32m1(a.val, a.val, 4), b.val, b.val, 4));
-    return v_sqrt(x);
-}
-
-inline v_float32x4 v_sqr_magnitude(const v_float32x4& a, const v_float32x4& b)
-{
-    return v_float32x4(vfmacc_vv_f32m1(vfmul_vv_f32m1(a.val, a.val, 4), b.val, b.val, 4));
-}
-
-inline v_float32x4 v_fma(const v_float32x4& a, const v_float32x4& b, const v_float32x4& c)
-{
-    return v_float32x4(vfmacc_vv_f32m1(c.val, a.val, b.val, 4));
-}
-
-inline v_int32x4 v_fma(const v_int32x4& a, const v_int32x4& b, const v_int32x4& c)
-{
-    return v_int32x4(vmacc_vv_i32m1(c.val, a.val, b.val, 4));
-}
-
-inline v_float32x4 v_muladd(const v_float32x4& a, const v_float32x4& b, const v_float32x4& c)
-{
-    return v_fma(a, b, c);
-}
-
-inline v_int32x4 v_muladd(const v_int32x4& a, const v_int32x4& b, const v_int32x4& c)
-{
-    return v_fma(a, b, c);
-}
-
-inline v_float32x4 v_matmul(const v_float32x4& v, const v_float32x4& m0,
-                            const v_float32x4& m1, const v_float32x4& m2,
-                            const v_float32x4& m3)
-{
-    vfloat32m1_t res = vfmul_vf_f32m1(m0.val, v.val[0], 4);//vmuli_f32(m0.val, v.val, 0);
-    res = vfmacc_vf_f32m1(res, v.val[1], m1.val, 4);//vmulai_f32(res, m1.val, v.val, 1);
-    res = vfmacc_vf_f32m1(res, v.val[2], m2.val, 4);//vmulai_f32(res, m1.val, v.val, 1);
-    res = vfmacc_vf_f32m1(res, v.val[3], m3.val, 4);//vmulai_f32(res, m1.val, v.val, 1);
-    return v_float32x4(res);
-}
-
-inline v_float32x4 v_matmuladd(const v_float32x4& v, const v_float32x4& m0,
-                               const v_float32x4& m1, const v_float32x4& m2,
-                               const v_float32x4& a)
-{
-    vfloat32m1_t res = vfmul_vf_f32m1(m0.val, v.val[0], 4);//vmuli_f32(m0.val, v.val, 0);
-    res = vfmacc_vf_f32m1(res, v.val[1], m1.val, 4);//vmulai_f32(res, m1.val, v.val, 1);
-    res = vfmacc_vf_f32m1(res, v.val[2], m2.val, 4);//vmulai_f32(res, m1.val, v.val, 1);
-    res = vfadd_vv_f32m1(res, a.val, 4);//vmulai_f32(res, m1.val, v.val, 1);
-    return v_float32x4(res);
-}
-
-inline v_float64x2 v_sqrt(const v_float64x2& x)
-{
-    return v_float64x2(vfsqrt_v_f64m1(x.val, 2));
-}
-
-inline v_float64x2 v_invsqrt(const v_float64x2& x)
-{
-    return v_float64x2(vfrdiv_vf_f64m1(vfsqrt_v_f64m1(x.val, 2), 1, 2));
-}
-
-inline v_float64x2 v_magnitude(const v_float64x2& a, const v_float64x2& b)
-{
-    v_float64x2 x(vfmacc_vv_f64m1(vfmul_vv_f64m1(a.val, a.val, 2), b.val, b.val, 2));
-    return v_sqrt(x);
-}
-
-inline v_float64x2 v_sqr_magnitude(const v_float64x2& a, const v_float64x2& b)
-{
-    return v_float64x2(vfmacc_vv_f64m1(vfmul_vv_f64m1(a.val, a.val, 2), b.val, b.val, 2));
-}
-
-inline v_float64x2 v_fma(const v_float64x2& a, const v_float64x2& b, const v_float64x2& c)
-{
-    return v_float64x2(vfmacc_vv_f64m1(c.val, a.val, b.val, 2));
-}
-
-inline v_float64x2 v_muladd(const v_float64x2& a, const v_float64x2& b, const v_float64x2& c)
-{
-    return v_fma(a, b, c);
-}
-
-#define OPENCV_HAL_IMPL_RISCVV_LOGIC_OPN(_Tpvec, suffix, num) \
-    OPENCV_HAL_IMPL_RISCVV_BIN_OPN(&, _Tpvec, vand_vv_##suffix, num) \
-    OPENCV_HAL_IMPL_RISCVV_BIN_OPN(|, _Tpvec, vor_vv_##suffix, num) \
-    OPENCV_HAL_IMPL_RISCVV_BIN_OPN(^, _Tpvec, vxor_vv_##suffix, num) \
-    inline _Tpvec operator ~ (const _Tpvec & a) \
-    { \
-        return _Tpvec(vnot_v_##suffix(a.val, num)); \
-    }
-
-OPENCV_HAL_IMPL_RISCVV_LOGIC_OPN(v_uint8x16, u8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_LOGIC_OPN(v_uint16x8, u16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_LOGIC_OPN(v_uint32x4, u32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_LOGIC_OPN(v_uint64x2, u64m1, 2)
-OPENCV_HAL_IMPL_RISCVV_LOGIC_OPN(v_int8x16,  i8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_LOGIC_OPN(v_int16x8,  i16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_LOGIC_OPN(v_int32x4,  i32m1, 4)
-OPENCV_HAL_IMPL_RISCVV_LOGIC_OPN(v_int64x2,  i64m1, 2)
-
-#define OPENCV_HAL_IMPL_RISCVV_FLT_BIT_OP(bin_op, intrin) \
-inline v_float32x4 operator bin_op (const v_float32x4& a, const v_float32x4& b) \
-{ \
-    return v_float32x4(vfloat32m1_t(intrin(vint32m1_t(a.val), vint32m1_t(b.val), 4))); \
-} \
-inline v_float32x4& operator bin_op##= (v_float32x4& a, const v_float32x4& b) \
-{ \
-    a.val = vfloat32m1_t(intrin(vint32m1_t(a.val), vint32m1_t(b.val), 4)); \
-    return a; \
-}
-
-OPENCV_HAL_IMPL_RISCVV_FLT_BIT_OP(&, vand_vv_i32m1)
-OPENCV_HAL_IMPL_RISCVV_FLT_BIT_OP(|, vor_vv_i32m1)
-OPENCV_HAL_IMPL_RISCVV_FLT_BIT_OP(^, vxor_vv_i32m1)
-
-inline v_float32x4 operator ~ (const v_float32x4& a)
-{
-    return v_float32x4((vfloat32m1_t)(vnot_v_i32m1((vint32m1_t)(a.val), 4)));
-}
-
-#define OPENCV_HAL_IMPL_RISCVV_FLT_64BIT_OP(bin_op, intrin) \
-inline v_float64x2 operator bin_op (const v_float64x2& a, const v_float64x2& b) \
-{ \
-    return v_float64x2(vfloat64m1_t(intrin(vint64m1_t(a.val), vint64m1_t(b.val), 2))); \
-} \
-inline v_float64x2& operator bin_op##= (v_float64x2& a, const v_float64x2& b) \
-{ \
-    a.val = vfloat64m1_t(intrin(vint64m1_t(a.val), vint64m1_t(b.val), 2)); \
-    return a; \
-}
-
-OPENCV_HAL_IMPL_RISCVV_FLT_64BIT_OP(&, vand_vv_i64m1)
-OPENCV_HAL_IMPL_RISCVV_FLT_64BIT_OP(|, vor_vv_i64m1)
-OPENCV_HAL_IMPL_RISCVV_FLT_64BIT_OP(^, vxor_vv_i64m1)
-
-inline v_float64x2 operator ~ (const v_float64x2& a)
-{
-    return v_float64x2((vfloat64m1_t)(vnot_v_i64m1((vint64m1_t)(a.val), 2)));
-}
-inline v_int16x8 v_mul_hi(const v_int16x8& a, const v_int16x8& b)
-{
-    return v_int16x8(vmulh_vv_i16m1(a.val, b.val, 8));
-}
-inline v_uint16x8 v_mul_hi(const v_uint16x8& a, const v_uint16x8& b)
-{
-    return v_uint16x8(vmulhu_vv_u16m1(a.val, b.val, 8));
-}
-
-//#define OPENCV_HAL_IMPL_RISCVV_ABS(_Tpuvec, _Tpsvec, usuffix, ssuffix) \
-//inline _Tpuvec v_abs(const _Tpsvec& a) {    \
-//    E##xm1_t mask=vmflt_vf_e32xm1_f32m1(x.val, 0.0, 4);
-
-//OPENCV_HAL_IMPL_RISCVV_ABS(v_uint8x16, v_int8x16, u8, s8)
-//OPENCV_HAL_IMPL_RISCVV_ABS(v_uint16x8, v_int16x8, u16, s16)
-//OPENCV_HAL_IMPL_RISCVV_ABS(v_uint32x4, v_int32x4, u32, s32)
-
-inline v_uint32x4 v_abs(v_int32x4 x)
-{
-    vbool32_t mask=vmslt_vx_i32m1_b32(x.val, 0, 4);
-    return v_uint32x4((vuint32m1_t)vrsub_vx_i32m1_m(mask, x.val, x.val, 0, 4));
-}
-
-inline v_uint16x8 v_abs(v_int16x8 x)
-{
-    vbool16_t mask=vmslt_vx_i16m1_b16(x.val, 0, 8);
-    return v_uint16x8((vuint16m1_t)vrsub_vx_i16m1_m(mask, x.val, x.val, 0, 8));
-}
-
-inline v_uint8x16 v_abs(v_int8x16 x)
-{
-    vbool8_t mask=vmslt_vx_i8m1_b8(x.val, 0, 16);
-    return v_uint8x16((vuint8m1_t)vrsub_vx_i8m1_m(mask, x.val, x.val, 0, 16));
-}
-
-inline v_float32x4 v_abs(v_float32x4 x)
-{
-    return (v_float32x4)vfsgnjx_vv_f32m1(x.val, x.val, 4);
-}
-
-inline v_float64x2 v_abs(v_float64x2 x)
-{
-    return (v_float64x2)vfsgnjx_vv_f64m1(x.val, x.val, 2);
-}
-
-inline v_float32x4 v_absdiff(const v_float32x4& a, const v_float32x4& b)
-{
-    vfloat32m1_t ret = vfsub_vv_f32m1(a.val, b.val, 4);
-    return (v_float32x4)vfsgnjx_vv_f32m1(ret, ret, 4);
-}
-
-inline v_float64x2 v_absdiff(const v_float64x2& a, const v_float64x2& b)
-{
-    vfloat64m1_t ret = vfsub_vv_f64m1(a.val, b.val, 2);
-    return (v_float64x2)vfsgnjx_vv_f64m1(ret, ret, 2);
-}
-
-#define OPENCV_HAL_IMPL_RISCVV_ABSDIFF_U(bit, num) \
-inline v_uint##bit##x##num v_absdiff(v_uint##bit##x##num a, v_uint##bit##x##num b){    \
-    vuint##bit##m1_t vmax = vmaxu_vv_u##bit##m1(a.val, b.val, num);    \
-    vuint##bit##m1_t vmin = vminu_vv_u##bit##m1(a.val, b.val, num);    \
-    return v_uint##bit##x##num(vsub_vv_u##bit##m1(vmax, vmin, num));\
-}
-
-OPENCV_HAL_IMPL_RISCVV_ABSDIFF_U(8, 16)
-OPENCV_HAL_IMPL_RISCVV_ABSDIFF_U(16, 8)
-OPENCV_HAL_IMPL_RISCVV_ABSDIFF_U(32, 4)
-
-/** Saturating absolute difference **/
-inline v_int8x16 v_absdiffs(v_int8x16 a, v_int8x16 b){
-    vint8m1_t vmax = vmax_vv_i8m1(a.val, b.val, 16);
-    vint8m1_t vmin = vmin_vv_i8m1(a.val, b.val, 16);
-    return v_int8x16(vssub_vv_i8m1(vmax, vmin, 16));
-}
-inline v_int16x8 v_absdiffs(v_int16x8 a, v_int16x8 b){
-    vint16m1_t vmax = vmax_vv_i16m1(a.val, b.val, 8);
-    vint16m1_t vmin = vmin_vv_i16m1(a.val, b.val, 8);
-    return v_int16x8(vssub_vv_i16m1(vmax, vmin, 8));
-}
-
-#define OPENCV_HAL_IMPL_RISCVV_ABSDIFF(_Tpvec, _Tpv, num) \
-inline v_uint##_Tpvec v_absdiff(v_int##_Tpvec a, v_int##_Tpvec b){    \
-     vint##_Tpv##_t max = vmax_vv_i##_Tpv(a.val, b.val, num);\
-     vint##_Tpv##_t min = vmin_vv_i##_Tpv(a.val, b.val, num);\
-    return v_uint##_Tpvec((vuint##_Tpv##_t)vsub_vv_i##_Tpv(max, min, num));    \
-}
-
-OPENCV_HAL_IMPL_RISCVV_ABSDIFF(8x16, 8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_ABSDIFF(16x8, 16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_ABSDIFF(32x4, 32m1, 4)
-
-//  Multiply and expand
-inline void v_mul_expand(const v_int8x16& a, const v_int8x16& b,
-                         v_int16x8& c, v_int16x8& d)
-{
-    vint16m2_t res = vundefined_i16m2();
-    res = vwmul_vv_i16m2(a.val, b.val, 16);
-    c.val = vget_i16m2_i16m1(res, 0);
-    d.val = vget_i16m2_i16m1(res, 1);
-}
-
-inline void v_mul_expand(const v_uint8x16& a, const v_uint8x16& b,
-                         v_uint16x8& c, v_uint16x8& d)
-{
-    vuint16m2_t res = vundefined_u16m2();
-    res = vwmulu_vv_u16m2(a.val, b.val, 16);
-    c.val = vget_u16m2_u16m1(res, 0);
-    d.val = vget_u16m2_u16m1(res, 1);
-}
-
-inline void v_mul_expand(const v_int16x8& a, const v_int16x8& b,
-                         v_int32x4& c, v_int32x4& d)
-{
-    vint32m2_t res = vundefined_i32m2();
-    res = vwmul_vv_i32m2(a.val, b.val, 8);
-    c.val = vget_i32m2_i32m1(res, 0);
-    d.val = vget_i32m2_i32m1(res, 1);
-}
-
-inline void v_mul_expand(const v_uint16x8& a, const v_uint16x8& b,
-                         v_uint32x4& c, v_uint32x4& d)
-{
-    vuint32m2_t res = vundefined_u32m2();
-    res = vwmulu_vv_u32m2(a.val, b.val, 8);
-    c.val = vget_u32m2_u32m1(res, 0);
-    d.val = vget_u32m2_u32m1(res, 1);
-}
-
-inline void v_mul_expand(const v_int32x4& a, const v_int32x4& b,
-                         v_int64x2& c, v_int64x2& d)
-{
-    vint64m2_t res = vundefined_i64m2();
-    res = vwmul_vv_i64m2(a.val, b.val, 4);
-    c.val = vget_i64m2_i64m1(res, 0);
-    d.val = vget_i64m2_i64m1(res, 1);
-}
-
-inline void v_mul_expand(const v_uint32x4& a, const v_uint32x4& b,
-                         v_uint64x2& c, v_uint64x2& d)
-{
-    vuint64m2_t res = vundefined_u64m2();
-    res = vwmulu_vv_u64m2(a.val, b.val, 4);
-    c.val = vget_u64m2_u64m1(res, 0);
-    d.val = vget_u64m2_u64m1(res, 1);
-}
-
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_uint8x16, v_add_wrap, vadd_vv_u8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_int8x16, v_add_wrap, vadd_vv_i8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_uint16x8, v_add_wrap, vadd_vv_u16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_int16x8, v_add_wrap, vadd_vv_i16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_uint8x16, v_sub_wrap, vsub_vv_u8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_int8x16, v_sub_wrap, vsub_vv_i8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_uint16x8, v_sub_wrap, vsub_vv_u16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_int16x8, v_sub_wrap, vsub_vv_i16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_uint8x16, v_mul_wrap, vmul_vv_u8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_int8x16, v_mul_wrap, vmul_vv_i8m1, 16)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_uint16x8, v_mul_wrap, vmul_vv_u16m1, 8)
-OPENCV_HAL_IMPL_RISCVV_BINN_FUNC(v_int16x8, v_mul_wrap, vmul_vv_i16m1, 8)
-//////// Dot Product ////////
-// 16 >> 32
-inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b)
-{
-    vint32m2_t res = vundefined_i32m2();
-    res = vwmul_vv_i32m2(a.val, b.val, 8);
-    res = vrgather_vv_i32m2(res, (vuint32m2_t){0, 2, 4, 6, 1, 3, 5, 7}, 8);
-    return v_int32x4(vadd_vv_i32m1(vget_i32m2_i32m1(res, 0), vget_i32m2_i32m1(res, 1), 4));
-}
-inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b, const v_int32x4& c)
-{
-    vint32m2_t res = vundefined_i32m2();
-    res = vwmul_vv_i32m2(a.val, b.val, 8);
-    res = vrgather_vv_i32m2(res, (vuint32m2_t){0, 2, 4, 6, 1, 3, 5, 7}, 8);
-    return v_int32x4(vadd_vv_i32m1(vadd_vv_i32m1(vget_i32m2_i32m1(res, 0),vget_i32m2_i32m1(res, 1), 4), c.val, 4));
-}
-
-// 32 >> 64
-inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b)
-{
-    vint64m2_t res = vundefined_i64m2();
-    res = vwmul_vv_i64m2(a.val, b.val, 4);
-    res = vrgather_vv_i64m2(res, (vuint64m2_t){0, 2, 1, 3}, 4);
-    return v_int64x2(vadd_vv_i64m1(vget_i64m2_i64m1(res, 0), vget_i64m2_i64m1(res, 1), 2));
-}
-inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b, const v_int64x2& c)
-{
-    vint64m2_t res = vundefined_i64m2();
-    res = vwmul_vv_i64m2(a.val, b.val, 4);
-    res = vrgather_vv_i64m2(res, (vuint64m2_t){0, 2, 1, 3}, 4);
-    return v_int64x2(vadd_vv_i64m1(vadd_vv_i64m1(vget_i64m2_i64m1(res, 0), vget_i64m2_i64m1(res, 1), 2), c.val, 2));
-}
-
-// 8 >> 32
-inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b)
-{
-    vuint16m2_t v1 = vundefined_u16m2();
-    vuint32m2_t v2 = vundefined_u32m2();
-    v1 = vwmulu_vv_u16m2(a.val, b.val, 16);
-    v1 = vrgather_vv_u16m2(v1, (vuint16m2_t){0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15}, 16);
-    v2 = vwaddu_vv_u32m2(vget_u16m2_u16m1(v1, 0), vget_u16m2_u16m1(v1, 1), 8);
-    return v_uint32x4(vadd_vv_u32m1(vget_u32m2_u32m1(v2, 0), vget_u32m2_u32m1(v2, 1), 4));
-}
-
-inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b,
-                                   const v_uint32x4& c)
-{
-    vuint16m2_t v1 = vundefined_u16m2();
-    vuint32m2_t v2 = vundefined_u32m2();
-    v1 = vwmulu_vv_u16m2(a.val, b.val, 16);
-    v1 = vrgather_vv_u16m2(v1, (vuint16m2_t){0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15}, 16);
-    v2 = vwaddu_vv_u32m2(vget_u16m2_u16m1(v1, 0), vget_u16m2_u16m1(v1, 1), 8);
-    return v_uint32x4(vadd_vv_u32m1(vadd_vv_u32m1(vget_u32m2_u32m1(v2, 0), vget_u32m2_u32m1(v2, 1), 4), c.val, 4));
-}
-
-inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b)
-{
-    vint16m2_t v1 = vundefined_i16m2();
-    vint32m2_t v2 = vundefined_i32m2();
-    v1 = vwmul_vv_i16m2(a.val, b.val, 16);
-    v1 = vrgather_vv_i16m2(v1, (vuint16m2_t){0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15}, 16);
-    v2 = vwadd_vv_i32m2(vget_i16m2_i16m1(v1, 0), vget_i16m2_i16m1(v1, 1), 8);
-    return v_int32x4(vadd_vv_i32m1(vget_i32m2_i32m1(v2, 0), vget_i32m2_i32m1(v2, 1), 4));
-}
-
-inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b,
-                                   const v_int32x4& c)
-{
-    vint16m2_t v1 = vundefined_i16m2();
-    vint32m2_t v2 = vundefined_i32m2();
-    v1 = vwmul_vv_i16m2(a.val, b.val, 16);
-    v1 = vrgather_vv_i16m2(v1, (vuint16m2_t){0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15}, 16);
-    v2 = vwadd_vv_i32m2(vget_i16m2_i16m1(v1, 0), vget_i16m2_i16m1(v1, 1), 8);
-    return v_int32x4(vadd_vv_i32m1(vadd_vv_i32m1(vget_i32m2_i32m1(v2, 0), vget_i32m2_i32m1(v2, 1), 4), c.val, 4));
-}
-
-inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b)
-{
-    vuint32m2_t v1 = vundefined_u32m2();
-    vuint64m2_t v2 = vundefined_u64m2();
-    v1 = vwmulu_vv_u32m2(a.val, b.val, 8);
-    v1 = vrgather_vv_u32m2(v1, (vuint32m2_t){0, 4, 1, 5, 2, 6, 3, 7}, 8);
-    v2 = vwaddu_vv_u64m2(vget_u32m2_u32m1(v1, 0), vget_u32m2_u32m1(v1, 1), 4);
-    return v_uint64x2(vadd_vv_u64m1(vget_u64m2_u64m1(v2, 0), vget_u64m2_u64m1(v2, 1), 2));
-}
-
-inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b,
-                                   const v_uint64x2& c)
-{
-    vuint32m2_t v1 = vundefined_u32m2();
-    vuint64m2_t v2 = vundefined_u64m2();
-    v1 = vwmulu_vv_u32m2(a.val, b.val, 8);
-    v1 = vrgather_vv_u32m2(v1, (vuint32m2_t){0, 4, 1, 5, 2, 6, 3, 7}, 8);
-    v2 = vwaddu_vv_u64m2(vget_u32m2_u32m1(v1, 0), vget_u32m2_u32m1(v1, 1), 4);
-    return v_uint64x2(vadd_vv_u64m1(vadd_vv_u64m1(vget_u64m2_u64m1(v2, 0), vget_u64m2_u64m1(v2, 1), 2), c.val, 2));
-}
-
-inline v_int64x2 v_dotprod_expand(const v_int16x8& a, const v_int16x8& b)
-{
-    vint32m2_t v1 = vundefined_i32m2();
-    vint64m2_t v2 = vundefined_i64m2();
-    v1 = vwmul_vv_i32m2(a.val, b.val, 8);
-    v1 = vrgather_vv_i32m2(v1, (vuint32m2_t){0, 4, 1, 5, 2, 6, 3, 7}, 8);
-    v2 = vwadd_vv_i64m2(vget_i32m2_i32m1(v1, 0), vget_i32m2_i32m1(v1, 1), 4);
-    return v_int64x2(vadd_vv_i64m1(vget_i64m2_i64m1(v2, 0), vget_i64m2_i64m1(v2, 1), 2));
-}
-
-inline v_int64x2 v_dotprod_expand(const v_int16x8& a, const v_int16x8& b,
-                                   const v_int64x2& c)
-{
-    vint32m2_t v1 = vundefined_i32m2();
-    vint64m2_t v2 = vundefined_i64m2();
-    v1 = vwmul_vv_i32m2(a.val, b.val, 8);
-    v1 = vrgather_vv_i32m2(v1, (vuint32m2_t){0, 4, 1, 5, 2, 6, 3, 7}, 8);
-    v2 = vwadd_vv_i64m2(vget_i32m2_i32m1(v1, 0), vget_i32m2_i32m1(v1, 1), 4);
-    return v_int64x2(vadd_vv_i64m1(vadd_vv_i64m1(vget_i64m2_i64m1(v2, 0), vget_i64m2_i64m1(v2, 1), 2), c.val, 2));
-}
-
-//////// Fast Dot Product ////////
-// 16 >> 32
-inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b)
-{
-    vint32m2_t v1 = vundefined_i32m2();
-    v1 = vwmul_vv_i32m2(a.val, b.val, 8);
-    return v_int32x4(vadd_vv_i32m1(vget_i32m2_i32m1(v1, 0), vget_i32m2_i32m1(v1, 1), 4));
-}
-
-inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b, const v_int32x4& c)
-{
-    vint32m2_t v1 = vundefined_i32m2();
-    v1 = vwmul_vv_i32m2(a.val, b.val, 8);
-    return v_int32x4(vadd_vv_i32m1(vadd_vv_i32m1(vget_i32m2_i32m1(v1, 0), vget_i32m2_i32m1(v1, 1), 4), c.val, 4));
-}
-
-// 32 >> 64
-inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b)
-{
-    vint64m2_t v1 = vundefined_i64m2();
-    v1 = vwmul_vv_i64m2(a.val, b.val, 4);
-    return v_int64x2(vadd_vv_i64m1(vget_i64m2_i64m1(v1, 0), vget_i64m2_i64m1(v1, 1), 2));
-}
-inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b, const v_int64x2& c)
-{
-    vint64m2_t v1 = vundefined_i64m2();
-    v1 = vwmul_vv_i64m2(a.val, b.val, 8);
-    return v_int64x2(vadd_vv_i64m1(vadd_vv_i64m1(vget_i64m2_i64m1(v1, 0), vget_i64m2_i64m1(v1, 1), 4), c.val, 4));
-}
-
-// 8 >> 32
-inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b)
-{
-    vuint16m2_t v1 = vundefined_u16m2();
-    vuint32m2_t v2 = vundefined_u32m2();
-    v1 = vwmulu_vv_u16m2(a.val, b.val, 16);
-    v2 = vwaddu_vv_u32m2(vget_u16m2_u16m1(v1, 0), vget_u16m2_u16m1(v1, 1), 8);
-    return v_uint32x4(vadd_vv_u32m1(vget_u32m2_u32m1(v2, 0), vget_u32m2_u32m1(v2, 1), 4));
-}
-
-inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b, const v_uint32x4& c)
-{
-    vuint16m2_t v1 = vundefined_u16m2();
-    vuint32m2_t v2 = vundefined_u32m2();
-    v1 = vwmulu_vv_u16m2(a.val, b.val, 16);
-    v2 = vwaddu_vv_u32m2(vget_u16m2_u16m1(v1, 0), vget_u16m2_u16m1(v1, 1), 8);
-    return v_uint32x4(vadd_vv_u32m1(vadd_vv_u32m1(vget_u32m2_u32m1(v2, 0), vget_u32m2_u32m1(v2, 1), 4), c.val, 4));
-}
-
-inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b)
-{
-    vint16m2_t v1 = vundefined_i16m2();
-    vint32m2_t v2 = vundefined_i32m2();
-    v1 = vwmul_vv_i16m2(a.val, b.val, 16);
-    v2 = vwadd_vv_i32m2(vget_i16m2_i16m1(v1, 0), vget_i16m2_i16m1(v1, 1), 8);
-    return v_int32x4(vadd_vv_i32m1(vget_i32m2_i32m1(v2, 0), vget_i32m2_i32m1(v2, 1), 4));
-}
-inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b, const v_int32x4& c)
-{
-    vint16m2_t v1 = vundefined_i16m2();
-    vint32m2_t v2 = vundefined_i32m2();
-    v1 = vwmul_vv_i16m2(a.val, b.val, 16);
-    v2 = vwadd_vv_i32m2(vget_i16m2_i16m1(v1, 0), vget_i16m2_i16m1(v1, 1), 8);
-    return v_int32x4(vadd_vv_i32m1(vadd_vv_i32m1(vget_i32m2_i32m1(v2, 0), vget_i32m2_i32m1(v2, 1), 4), c.val, 4));
-}
-
-// 16 >> 64
-inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b)
-{
-    vuint32m2_t v1 = vundefined_u32m2();
-    vuint64m2_t v2 = vundefined_u64m2();
-    v1 = vwmulu_vv_u32m2(a.val, b.val, 8);
-    v2 = vwaddu_vv_u64m2(vget_u32m2_u32m1(v1, 0), vget_u32m2_u32m1(v1, 1), 4);
-    return v_uint64x2(vadd_vv_u64m1(vget_u64m2_u64m1(v2, 0), vget_u64m2_u64m1(v2, 1), 2));
-}
-inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b, const v_uint64x2& c)
-{
-    vuint32m2_t v1 = vundefined_u32m2();
-    vuint64m2_t v2 = vundefined_u64m2();
-    v1 = vwmulu_vv_u32m2(a.val, b.val, 8);
-    v2 = vwaddu_vv_u64m2(vget_u32m2_u32m1(v1, 0), vget_u32m2_u32m1(v1, 1), 4);
-    return v_uint64x2(vadd_vv_u64m1(vadd_vv_u64m1(vget_u64m2_u64m1(v2, 0), vget_u64m2_u64m1(v2, 1), 2), c.val, 2));
-}
-
-inline v_int64x2 v_dotprod_expand_fast(const v_int16x8& a, const v_int16x8& b)
-{
-    vint32m2_t v1 = vundefined_i32m2();
-    vint64m2_t v2 = vundefined_i64m2();
-    v1 = vwmul_vv_i32m2(a.val, b.val, 8);
-    v2 = vwadd_vv_i64m2(vget_i32m2_i32m1(v1, 0), vget_i32m2_i32m1(v1, 1), 4);
-    return v_int64x2(vadd_vv_i64m1(vget_i64m2_i64m1(v2, 0), vget_i64m2_i64m1(v2, 1), 2));
-}
-inline v_int64x2 v_dotprod_expand_fast(const v_int16x8& a, const v_int16x8& b, const v_int64x2& c)
-{
-    vint32m2_t v1 = vundefined_i32m2();
-    vint64m2_t v2 = vundefined_i64m2();
-    v1 = vwmul_vv_i32m2(a.val, b.val, 8);
-    v2 = vwadd_vv_i64m2(vget_i32m2_i32m1(v1, 0), vget_i32m2_i32m1(v1, 1), 4);
-    return v_int64x2(vadd_vv_i64m1(vadd_vv_i64m1(vget_i64m2_i64m1(v2, 0), vget_i64m2_i64m1(v2, 1), 2), c.val, 2));
-}
-
-
-#define OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_W(_Tpvec, _Tpvec2, len, scalartype, func, intrin, num) \
-inline scalartype v_reduce_##func(const v_##_Tpvec##x##num& a) \
-{\
-    v##_Tpvec2##m1_t val = vmv_v_x_##len##m1(0, num); \
-    val = intrin(val, a.val, val, num);    \
-    return vmv_x_s_##len##m1_##len(val, num);    \
-}
-
-
-#define OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_(_Tpvec, _Tpvec2, scalartype, func, funcu, num) \
-inline scalartype v_reduce_##func(const v_##_Tpvec##x##num& a) \
-{\
-    v##_Tpvec##m1_t val = (v##_Tpvec##m1_t)vmv_v_x_i8m1(0, num); \
-    val = v##funcu##_vs_##_Tpvec2##m1_##_Tpvec2##m1(val, a.val, a.val, num);    \
-    return val[0];    \
-}
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_W(int8, int16, i16, int, sum, vwredsum_vs_i8m1_i16m1, 16)
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_W(int16, int32, i32, int, sum, vwredsum_vs_i16m1_i32m1, 8)
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_W(int32, int64, i64, int, sum, vwredsum_vs_i32m1_i64m1, 4)
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_W(uint8, uint16, u16, unsigned, sum, vwredsumu_vs_u8m1_u16m1, 16)
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_W(uint16, uint32, u32, unsigned, sum, vwredsumu_vs_u16m1_u32m1, 8)
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_W(uint32, uint64, u64, unsigned, sum, vwredsumu_vs_u32m1_u64m1, 4)
-inline float v_reduce_sum(const v_float32x4& a) \
-{\
-    vfloat32m1_t val = vfmv_v_f_f32m1(0.0, 4); \
-    val = vfredsum_vs_f32m1_f32m1(val, a.val, val, 4);    \
-    return vfmv_f_s_f32m1_f32(val, 4);    \
-}
-inline double v_reduce_sum(const v_float64x2& a) \
-{\
-    vfloat64m1_t val = vfmv_v_f_f64m1(0.0, 2); \
-    val = vfredsum_vs_f64m1_f64m1(val, a.val, val, 2);    \
-    return vfmv_f_s_f64m1_f64(val, 2);    \
-}
-inline uint64 v_reduce_sum(const v_uint64x2& a)
-{ return vext_x_v_u64m1_u64((vuint64m1_t)a.val, 0, 2)+vext_x_v_u64m1_u64((vuint64m1_t)a.val, 1, 2); }
-
-inline int64 v_reduce_sum(const v_int64x2& a)
-{ return vext_x_v_i64m1_i64((vint64m1_t)a.val, 0, 2)+vext_x_v_i64m1_i64((vint64m1_t)a.val, 1, 2); }
-
-#define OPENCV_HAL_IMPL_RISCVV_REDUCE_OP(func)    \
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_(int8,  i8, int, func, red##func, 16)    \
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_(int16, i16, int, func, red##func, 8)    \
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_(int32, i32, int, func, red##func, 4)    \
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_(int64, i64, int, func, red##func, 2)    \
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_(uint8,  u8, unsigned, func, red##func##u, 16)    \
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_(uint16, u16, unsigned, func, red##func##u, 8)    \
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_(uint32, u32, unsigned, func, red##func##u, 4)    \
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP_(float32, f32, float, func, fred##func, 4)
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP(max)
-OPENCV_HAL_IMPL_RISCVV_REDUCE_OP(min)
-
-inline v_float32x4 v_reduce_sum4(const v_float32x4& a, const v_float32x4& b,
-                                 const v_float32x4& c, const v_float32x4& d)
-{
-    vfloat32m1_t a0 = vfmv_v_f_f32m1(0.0, 4);
-    vfloat32m1_t b0 = vfmv_v_f_f32m1(0.0, 4);
-    vfloat32m1_t c0 = vfmv_v_f_f32m1(0.0, 4);
-    vfloat32m1_t d0 = vfmv_v_f_f32m1(0.0, 4);
-    a0 = vfredsum_vs_f32m1_f32m1(a0, a.val, a0, 4);
-    b0 = vfredsum_vs_f32m1_f32m1(b0, b.val, b0, 4);
-    c0 = vfredsum_vs_f32m1_f32m1(c0, c.val, c0, 4);
-    d0 = vfredsum_vs_f32m1_f32m1(d0, d.val, d0, 4);
-    return v_float32x4(a0[0], b0[0], c0[0], d0[0]);
-}
-
-inline float v_reduce_sad(const v_float32x4& a, const v_float32x4& b)
-{
-    vfloat32m1_t a0 = vfmv_v_f_f32m1(0.0, 4);
-    vfloat32m1_t x = vfsub_vv_f32m1(a.val, b.val, 4);
-    vbool32_t mask=vmflt_vf_f32m1_b32(x, 0, 4);
-    vfloat32m1_t val = vfrsub_vf_f32m1_m(mask, x, x, 0, 4);
-    a0 = vfredsum_vs_f32m1_f32m1(a0, val, a0, 4);
-    return a0[0];
-}
-
-#define OPENCV_HAL_IMPL_RISCVV_REDUCE_SAD(_Tpvec, _Tpvec2) \
-inline unsigned v_reduce_sad(const _Tpvec& a, const _Tpvec&b){    \
-    _Tpvec2 x = v_absdiff(a, b);    \
-    return v_reduce_sum(x);    \
-}
-
-OPENCV_HAL_IMPL_RISCVV_REDUCE_SAD(v_int8x16, v_uint8x16)
-OPENCV_HAL_IMPL_RISCVV_REDUCE_SAD(v_uint8x16, v_uint8x16)
-OPENCV_HAL_IMPL_RISCVV_REDUCE_SAD(v_int16x8, v_uint16x8)
-OPENCV_HAL_IMPL_RISCVV_REDUCE_SAD(v_uint16x8, v_uint16x8)
-OPENCV_HAL_IMPL_RISCVV_REDUCE_SAD(v_int32x4, v_uint32x4)
-OPENCV_HAL_IMPL_RISCVV_REDUCE_SAD(v_uint32x4, v_uint32x4)
-
-#define OPENCV_HAL_IMPL_RISCVV_INT_CMP_OP(_Tpvec, _Tp, _T, num, uv) \
-inline _Tpvec operator == (const _Tpvec& a, const _Tpvec& b) \
-{ \
-    vbool##_T##_t mask = vmseq_vv_##_Tp##_b##_T(a.val, b.val, num);    \
-    return _Tpvec(vmerge_vxm_##_Tp(mask, vmv_v_x_##_Tp(0, num), -1, num));    \
-} \
-inline _Tpvec operator != (const _Tpvec& a, const _Tpvec& b) \
-{ \
-    vbool##_T##_t mask = vmsne_vv_##_Tp##_b##_T(a.val, b.val, num);    \
-    return _Tpvec(vmerge_vxm_##_Tp(mask, vmv_v_x_##_Tp(0, num), -1, num));    \
-} \
-inline _Tpvec operator < (const _Tpvec& a, const _Tpvec& b) \
-{ \
-    vbool##_T##_t mask = vmslt##uv##_Tp##_b##_T(a.val, b.val, num);    \
-    return _Tpvec(vmerge_vxm_##_Tp(mask, vmv_v_x_##_Tp(0, num), -1, num));    \
-} \
-inline _Tpvec operator > (const _Tpvec& a, const _Tpvec& b) \
-{ \
-    vbool##_T##_t mask = vmslt##uv##_Tp##_b##_T(b.val, a.val, num);    \
-    return _Tpvec(vmerge_vxm_##_Tp(mask, vmv_v_x_##_Tp(0, num), -1, num));    \
-} \
-inline _Tpvec operator <= (const _Tpvec& a, const _Tpvec& b) \
-{ \
-    vbool##_T##_t mask = vmsle##uv##_Tp##_b##_T(a.val, b.val, num);    \
-    return _Tpvec(vmerge_vxm_##_Tp(mask, vmv_v_x_##_Tp(0, num), -1, num));    \
-} \
-inline _Tpvec operator >= (const _Tpvec& a, const _Tpvec& b) \
-{ \
-    vbool##_T##_t mask = vmsle##uv##_Tp##_b##_T(b.val, a.val, num);    \
-    return _Tpvec(vmerge_vxm_##_Tp(mask, vmv_v_x_##_Tp(0, num), -1, num));    \
-} \
-
-OPENCV_HAL_IMPL_RISCVV_INT_CMP_OP(v_int8x16, i8m1,  8, 16, _vv_)
-OPENCV_HAL_IMPL_RISCVV_INT_CMP_OP(v_int16x8, i16m1, 16, 8, _vv_)
-OPENCV_HAL_IMPL_RISCVV_INT_CMP_OP(v_int32x4, i32m1, 32, 4, _vv_)
-OPENCV_HAL_IMPL_RISCVV_INT_CMP_OP(v_int64x2, i64m1, 64, 2, _vv_)
-OPENCV_HAL_IMPL_RISCVV_INT_CMP_OP(v_uint8x16, u8m1, 8, 16, u_vv_)
-OPENCV_HAL_IMPL_RISCVV_INT_CMP_OP(v_uint16x8, u16m1, 16, 8, u_vv_)
-OPENCV_HAL_IMPL_RISCVV_INT_CMP_OP(v_uint32x4, u32m1, 32, 4, u_vv_)
-OPENCV_HAL_IMPL_RISCVV_INT_CMP_OP(v_uint64x2, u64m1, 64, 2, u_vv_)
-
-//TODO: ==
-inline v_float32x4 operator == (const v_float32x4& a, const v_float32x4& b)
-{
-    vbool32_t mask = vmfeq_vv_f32m1_b32(a.val, b.val, 4);
-    vint32m1_t res = vmerge_vxm_i32m1(mask, vmv_v_x_i32m1(0.0, 4), -1, 4);
-    return v_float32x4((vfloat32m1_t)res);
-}
-inline v_float32x4 operator != (const v_float32x4& a, const v_float32x4& b)
-{
-    vbool32_t mask = vmfne_vv_f32m1_b32(a.val, b.val, 4);
-    vint32m1_t res = vmerge_vxm_i32m1(mask, vmv_v_x_i32m1(0.0, 4), -1, 4);
-    return v_float32x4((vfloat32m1_t)res);
-}
-inline v_float32x4 operator < (const v_float32x4& a, const v_float32x4& b)
-{
-    vbool32_t mask = vmflt_vv_f32m1_b32(a.val, b.val, 4);
-    vint32m1_t res = vmerge_vxm_i32m1(mask, vmv_v_x_i32m1(0.0, 4), -1, 4);
-    return v_float32x4((vfloat32m1_t)res);
-}
-inline v_float32x4 operator <= (const v_float32x4& a, const v_float32x4& b)
-{
-    vbool32_t mask = vmfle_vv_f32m1_b32(a.val, b.val, 4);
-    vint32m1_t res = vmerge_vxm_i32m1(mask, vmv_v_x_i32m1(0.0, 4), -1, 4);
-    return v_float32x4((vfloat32m1_t)res);
-}
-inline v_float32x4 operator > (const v_float32x4& a, const v_float32x4& b)
-{
-    vbool32_t mask = vmfgt_vv_f32m1_b32(a.val, b.val, 4);
-    vint32m1_t res = vmerge_vxm_i32m1(mask, vmv_v_x_i32m1(0.0, 4), -1, 4);
-    return v_float32x4((vfloat32m1_t)res);
-}
-inline v_float32x4 operator >= (const v_float32x4& a, const v_float32x4& b)
-{
-    vbool32_t mask = vmfge_vv_f32m1_b32(a.val, b.val, 4);
-    vint32m1_t res = vmerge_vxm_i32m1(mask, vmv_v_x_i32m1(0.0, 4), -1, 4);
-    return v_float32x4((vfloat32m1_t)res);
-}
-inline v_float32x4 v_not_nan(const v_float32x4& a)
-{
-    vbool32_t mask = vmford_vv_f32m1_b32(a.val, a.val, 4);
-    vint32m1_t res = vmerge_vxm_i32m1(mask, vmv_v_x_i32m1(0.0, 4), -1, 4);
-    return v_float32x4((vfloat32m1_t)res);
-}
-
-//TODO: ==
-inline v_float64x2 operator == (const v_float64x2& a, const v_float64x2& b)
-{
-    vbool64_t mask = vmfeq_vv_f64m1_b64(a.val, b.val, 2);
-    vint64m1_t res = vmerge_vxm_i64m1(mask, vmv_v_x_i64m1(0.0, 2), -1, 2);
-    return v_float64x2((vfloat64m1_t)res);
-}
-inline v_float64x2 operator != (const v_float64x2& a, const v_float64x2& b)
-{
-    vbool64_t mask = vmfne_vv_f64m1_b64(a.val, b.val, 2);
-    vint64m1_t res = vmerge_vxm_i64m1(mask, vmv_v_x_i64m1(0.0, 2), -1, 2);
-    return v_float64x2((vfloat64m1_t)res);
-}
-inline v_float64x2 operator < (const v_float64x2& a, const v_float64x2& b)
-{
-    vbool64_t mask = vmflt_vv_f64m1_b64(a.val, b.val, 2);
-    vint64m1_t res = vmerge_vxm_i64m1(mask, vmv_v_x_i64m1(0.0, 2), -1, 2);
-    return v_float64x2((vfloat64m1_t)res);
-}
-inline v_float64x2 operator <= (const v_float64x2& a, const v_float64x2& b)
-{
-    vbool64_t mask = vmfle_vv_f64m1_b64(a.val, b.val, 2);
-    vint64m1_t res = vmerge_vxm_i64m1(mask, vmv_v_x_i64m1(0.0, 2), -1, 2);
-    return v_float64x2((vfloat64m1_t)res);
-}
-inline v_float64x2 operator > (const v_float64x2& a, const v_float64x2& b)
-{
-    vbool64_t mask = vmfgt_vv_f64m1_b64(a.val, b.val, 2);
-    vint64m1_t res = vmerge_vxm_i64m1(mask, vmv_v_x_i64m1(0.0, 2), -1, 2);
-    return v_float64x2((vfloat64m1_t)res);
-}
-inline v_float64x2 operator >= (const v_float64x2& a, const v_float64x2& b)
-{
-    vbool64_t mask = vmfge_vv_f64m1_b64(a.val, b.val, 2);
-    vint64m1_t res = vmerge_vxm_i64m1(mask, vmv_v_x_i64m1(0.0, 2), -1, 2);
-    return v_float64x2((vfloat64m1_t)res);
-}
-inline v_float64x2 v_not_nan(const v_float64x2& a)
-{
-    vbool64_t mask = vmford_vv_f64m1_b64(a.val, a.val, 2);
-    vint64m1_t res = vmerge_vxm_i64m1(mask, vmv_v_x_i64m1(0.0, 2), -1, 2);
-    return v_float64x2((vfloat64m1_t)res);
-}
-#define OPENCV_HAL_IMPL_RISCVV_TRANSPOSE4x4(_Tp, _T) \
-inline void v_transpose4x4(const v_##_Tp##32x4& a0, const v_##_Tp##32x4& a1, \
-                         const v_##_Tp##32x4& a2, const v_##_Tp##32x4& a3, \
-                         v_##_Tp##32x4& b0, v_##_Tp##32x4& b1, \
-                         v_##_Tp##32x4& b2, v_##_Tp##32x4& b3) \
-{ \
-    v##_Tp##32m4_t val = vundefined_##_T##m4();    \
-    val = vset_##_T##m4(val, 0, a0.val);    \
-    val = vset_##_T##m4(val, 1, a1.val);    \
-    val = vset_##_T##m4(val, 2, a2.val);    \
-    val = vset_##_T##m4(val, 3, a3.val);   \
-    val = vrgather_vv_##_T##m4(val, (vuint32m4_t){0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15}, 16);    \
-    b0.val = vget_##_T##m4_##_T##m1(val, 0);   \
-    b1.val = vget_##_T##m4_##_T##m1(val, 1);   \
-    b2.val = vget_##_T##m4_##_T##m1(val, 2);   \
-    b3.val = vget_##_T##m4_##_T##m1(val, 3);   \
-}
-OPENCV_HAL_IMPL_RISCVV_TRANSPOSE4x4(uint, u32)
-OPENCV_HAL_IMPL_RISCVV_TRANSPOSE4x4(int, i32)
-OPENCV_HAL_IMPL_RISCVV_TRANSPOSE4x4(float, f32)
-
-
-#define OPENCV_HAL_IMPL_RISCVV_SHIFT_LEFT(_Tpvec, suffix, _T, num) \
-inline _Tpvec operator << (const _Tpvec& a, int n) \
-{ return _Tpvec((vsll_vx_##_T##m1(a.val, n, num))); } \
-template<int n> inline _Tpvec v_shl(const _Tpvec& a) \
-{ return _Tpvec((vsll_vx_##_T##m1(a.val, n, num))); }
-
-#define OPENCV_HAL_IMPL_RISCVV_SHIFT_RIGHT(_Tpvec, suffix, _T, num, intric) \
-inline _Tpvec operator >> (const _Tpvec& a, int n) \
-{ return _Tpvec((v##intric##_vx_##_T##m1(a.val, n, num))); } \
-template<int n> inline _Tpvec v_shr(const _Tpvec& a) \
-{ return _Tpvec((v##intric##_vx_##_T##m1(a.val, n, num))); }\
-template<int n> inline _Tpvec v_rshr(const _Tpvec& a) \
-{ return _Tpvec((v##intric##_vx_##_T##m1(vadd_vx_##_T##m1(a.val, 1<<(n-1), num), n, num))); }
-
-// trade efficiency for convenience
-#define OPENCV_HAL_IMPL_RISCVV_SHIFT_OP(suffix, _T, num, intrin) \
-OPENCV_HAL_IMPL_RISCVV_SHIFT_LEFT(v_##suffix##x##num, suffix, _T, num) \
-OPENCV_HAL_IMPL_RISCVV_SHIFT_RIGHT(v_##suffix##x##num, suffix, _T, num, intrin)
-
-OPENCV_HAL_IMPL_RISCVV_SHIFT_OP(uint8, u8, 16, srl)
-OPENCV_HAL_IMPL_RISCVV_SHIFT_OP(uint16, u16, 8, srl)
-OPENCV_HAL_IMPL_RISCVV_SHIFT_OP(uint32, u32, 4, srl)
-OPENCV_HAL_IMPL_RISCVV_SHIFT_OP(uint64, u64, 2, srl)
-OPENCV_HAL_IMPL_RISCVV_SHIFT_OP(int8, i8, 16, sra)
-OPENCV_HAL_IMPL_RISCVV_SHIFT_OP(int16, i16, 8, sra)
-OPENCV_HAL_IMPL_RISCVV_SHIFT_OP(int32, i32, 4, sra)
-OPENCV_HAL_IMPL_RISCVV_SHIFT_OP(int64, i64, 2, sra)
-
-#if 0
-#define VUP4(n) {0, 1, 2, 3}
-#define VUP8(n) {0, 1, 2, 3, 4, 5, 6, 7}
-#define VUP16(n) {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15}
-#define VUP2(n) {0, 1}
-#endif
-#define OPENCV_HAL_IMPL_RISCVV_ROTATE_OP(_Tpvec, suffix, _T, num, num2, vmv, len) \
-template<int n> inline _Tpvec v_rotate_left(const _Tpvec& a) \
-{    \
-    suffix##m1_t tmp = vmv##_##_T##m1(0, num);\
-        tmp = vslideup_vx_##_T##m1_m(vmset_m_##len(num), tmp, a.val, n, num);\
-        return _Tpvec(tmp);\
-} \
-template<int n> inline _Tpvec v_rotate_right(const _Tpvec& a) \
-{     \
-        return _Tpvec(vslidedown_vx_##_T##m1(a.val, n, num));\
-} \
-template<> inline _Tpvec v_rotate_left<0>(const _Tpvec& a) \
-{ return a; } \
-template<int n> inline _Tpvec v_rotate_right(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    suffix##m2_t tmp = vundefined_##_T##m2();    \
-    tmp = vset_##_T##m2(tmp, 0, a.val);          \
-    tmp = vset_##_T##m2(tmp, 1, b.val);          \
-        tmp = vslidedown_vx_##_T##m2(tmp, n, num2);\
-        return _Tpvec(vget_##_T##m2_##_T##m1(tmp, 0));\
-} \
-template<int n> inline _Tpvec v_rotate_left(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    suffix##m2_t tmp = vundefined_##_T##m2();    \
-    tmp = vset_##_T##m2(tmp, 0, b.val);    \
-    tmp = vset_##_T##m2(tmp, 1, a.val);    \
-        tmp = vslideup_vx_##_T##m2(tmp, n, num2);\
-        return _Tpvec(vget_##_T##m2_##_T##m1(tmp, 1));\
-} \
-template<> inline _Tpvec v_rotate_left<0>(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    CV_UNUSED(b); return a; \
-}
-
-OPENCV_HAL_IMPL_RISCVV_ROTATE_OP(v_uint8x16, vuint8, u8, 16, 32, vmv_v_x, b8)
-OPENCV_HAL_IMPL_RISCVV_ROTATE_OP(v_int8x16, vint8, i8, 16, 32, vmv_v_x, b8)
-OPENCV_HAL_IMPL_RISCVV_ROTATE_OP(v_uint16x8, vuint16, u16, 8, 16, vmv_v_x, b16)
-OPENCV_HAL_IMPL_RISCVV_ROTATE_OP(v_int16x8, vint16, i16, 8, 16, vmv_v_x, b16)
-OPENCV_HAL_IMPL_RISCVV_ROTATE_OP(v_uint32x4, vuint32, u32, 4, 8, vmv_v_x, b32)
-OPENCV_HAL_IMPL_RISCVV_ROTATE_OP(v_int32x4, vint32, i32, 4, 8, vmv_v_x, b32)
-OPENCV_HAL_IMPL_RISCVV_ROTATE_OP(v_uint64x2, vuint64, u64, 2, 4, vmv_v_x, b64)
-OPENCV_HAL_IMPL_RISCVV_ROTATE_OP(v_int64x2, vint64, i64, 2, 4, vmv_v_x, b64)
-OPENCV_HAL_IMPL_RISCVV_ROTATE_OP(v_float32x4, vfloat32, f32, 4, 8, vfmv_v_f, b32)
-OPENCV_HAL_IMPL_RISCVV_ROTATE_OP(v_float64x2, vfloat64, f64, 2, 4, vfmv_v_f, b64)
-
-#define OPENCV_HAL_IMPL_RISCVV_LOADSTORE_OP(_Tpvec, _Tp, _Tp2, len, hnum, num) \
-inline _Tpvec v_load_halves(const _Tp* ptr0, const _Tp* ptr1) \
-{ \
-  typedef uint64 CV_DECL_ALIGNED(1) unaligned_uint64; \
-  vuint64m1_t tmp = {*(unaligned_uint64*)ptr0, *(unaligned_uint64*)ptr1};\
-    return _Tpvec(_Tp2##_t(tmp)); } \
-inline _Tpvec v_load_low(const _Tp* ptr) \
-{ return _Tpvec(vle_v_##len(ptr, hnum)); }\
-inline _Tpvec v_load_aligned(const _Tp* ptr) \
-{ return _Tpvec(vle_v_##len(ptr, num)); } \
-inline _Tpvec v_load(const _Tp* ptr) \
-{ return _Tpvec((_Tp2##_t)vle_v_##len((const _Tp *)ptr, num)); } \
-inline void v_store_low(_Tp* ptr, const _Tpvec& a) \
-{ vse_v_##len(ptr, a.val, hnum);}\
-inline void v_store_high(_Tp* ptr, const _Tpvec& a) \
-{ \
-  _Tp2##_t a0 = vslidedown_vx_##len(a.val, hnum, num);    \
-  vse_v_##len(ptr, a0, hnum);}\
-inline void v_store(_Tp* ptr, const _Tpvec& a) \
-{ vse_v_##len(ptr, a.val, num); } \
-inline void v_store_aligned(_Tp* ptr, const _Tpvec& a) \
-{ vse_v_##len(ptr, a.val, num); } \
-inline void v_store_aligned_nocache(_Tp* ptr, const _Tpvec& a) \
-{ vse_v_##len(ptr, a.val, num); } \
-inline void v_store(_Tp* ptr, const _Tpvec& a, hal::StoreMode /*mode*/) \
-{ vse_v_##len(ptr, a.val, num); }
-
-OPENCV_HAL_IMPL_RISCVV_LOADSTORE_OP(v_uint8x16, uchar, vuint8m1, u8m1, 8, 16)
-OPENCV_HAL_IMPL_RISCVV_LOADSTORE_OP(v_int8x16,  schar, vint8m1, i8m1, 8, 16)
-OPENCV_HAL_IMPL_RISCVV_LOADSTORE_OP(v_uint16x8, ushort, vuint16m1, u16m1, 4, 8)
-OPENCV_HAL_IMPL_RISCVV_LOADSTORE_OP(v_int16x8,  short,  vint16m1, i16m1, 4, 8)
-OPENCV_HAL_IMPL_RISCVV_LOADSTORE_OP(v_uint32x4, unsigned, vuint32m1, u32m1, 2, 4)
-OPENCV_HAL_IMPL_RISCVV_LOADSTORE_OP(v_int32x4,  int,     vint32m1, i32m1, 2, 4)
-OPENCV_HAL_IMPL_RISCVV_LOADSTORE_OP(v_uint64x2, unsigned long, vuint64m1, u64m1, 1, 2)
-OPENCV_HAL_IMPL_RISCVV_LOADSTORE_OP(v_int64x2,  long,     vint64m1, i64m1, 1, 2)
-OPENCV_HAL_IMPL_RISCVV_LOADSTORE_OP(v_float32x4, float, vfloat32m1, f32m1, 2, 4)
-OPENCV_HAL_IMPL_RISCVV_LOADSTORE_OP(v_float64x2, double, vfloat64m1, f64m1, 1, 2)
-
-
-////////////// Lookup table access ////////////////////
-
-inline v_int8x16 v_lut(const schar* tab, const int* idx)
-{
-#if 1
-    schar CV_DECL_ALIGNED(32) elems[16] =
-    {
-        tab[idx[ 0]],
-        tab[idx[ 1]],
-        tab[idx[ 2]],
-        tab[idx[ 3]],
-        tab[idx[ 4]],
-        tab[idx[ 5]],
-        tab[idx[ 6]],
-        tab[idx[ 7]],
-        tab[idx[ 8]],
-        tab[idx[ 9]],
-        tab[idx[10]],
-        tab[idx[11]],
-        tab[idx[12]],
-        tab[idx[13]],
-        tab[idx[14]],
-        tab[idx[15]]
-    };
-    return v_int8x16(vle_v_i8m1(elems, 16));
-#else
-    int32xm4_t index32 = vlev_int32xm4(idx, 16);
-    vint16m2_t index16 = vnsra_vx_i16m2_int32xm4(index32, 0, 16);
-    vint8m1_t index = vnsra_vx_i8m1_i16m2(index16, 0, 16);
-    return v_int8x16(vlxbv_i8m1(tab, index, 16));
-#endif
-}
-
-inline v_int8x16 v_lut_pairs(const schar* tab, const int* idx){
-    schar CV_DECL_ALIGNED(32) elems[16] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[1]],
-        tab[idx[1] + 1],
-        tab[idx[2]],
-        tab[idx[2] + 1],
-        tab[idx[3]],
-        tab[idx[3] + 1],
-        tab[idx[4]],
-        tab[idx[4] + 1],
-        tab[idx[5]],
-        tab[idx[5] + 1],
-        tab[idx[6]],
-        tab[idx[6] + 1],
-        tab[idx[7]],
-        tab[idx[7] + 1]
-    };
-    return v_int8x16(vle_v_i8m1(elems, 16));
-}
-inline v_int8x16 v_lut_quads(const schar* tab, const int* idx)
-{
-    schar CV_DECL_ALIGNED(32) elems[16] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[0] + 2],
-        tab[idx[0] + 3],
-        tab[idx[1]],
-        tab[idx[1] + 1],
-        tab[idx[1] + 2],
-        tab[idx[1] + 3],
-        tab[idx[2]],
-        tab[idx[2] + 1],
-        tab[idx[2] + 2],
-        tab[idx[2] + 3],
-        tab[idx[3]],
-        tab[idx[3] + 1],
-        tab[idx[3] + 2],
-        tab[idx[3] + 3]
-    };
-    return v_int8x16(vle_v_i8m1(elems, 16));
-}
-
-inline v_uint8x16 v_lut(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut((schar*)tab, idx)); }
-inline v_uint8x16 v_lut_pairs(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut_pairs((schar*)tab, idx)); }
-inline v_uint8x16 v_lut_quads(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut_quads((schar*)tab, idx)); }
-
-inline v_int16x8 v_lut(const short* tab, const int* idx)
-{
-    short CV_DECL_ALIGNED(32) elems[8] =
-    {
-        tab[idx[0]],
-        tab[idx[1]],
-        tab[idx[2]],
-        tab[idx[3]],
-        tab[idx[4]],
-        tab[idx[5]],
-        tab[idx[6]],
-        tab[idx[7]]
-    };
-    return v_int16x8(vle_v_i16m1(elems, 8));
-}
-inline v_int16x8 v_lut_pairs(const short* tab, const int* idx)
-{
-    short CV_DECL_ALIGNED(32) elems[8] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[1]],
-        tab[idx[1] + 1],
-        tab[idx[2]],
-        tab[idx[2] + 1],
-        tab[idx[3]],
-        tab[idx[3] + 1]
-    };
-    return v_int16x8(vle_v_i16m1(elems, 8));
-}
-inline v_int16x8 v_lut_quads(const short* tab, const int* idx)
-{
-    short CV_DECL_ALIGNED(32) elems[8] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[0] + 2],
-        tab[idx[0] + 3],
-        tab[idx[1]],
-        tab[idx[1] + 1],
-        tab[idx[1] + 2],
-        tab[idx[1] + 3]
-    };
-    return v_int16x8(vle_v_i16m1(elems, 8));
-}
-inline v_uint16x8 v_lut(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut((short*)tab, idx)); }
-inline v_uint16x8 v_lut_pairs(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut_pairs((short*)tab, idx)); }
-inline v_uint16x8 v_lut_quads(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut_quads((short*)tab, idx)); }
-
-inline v_int32x4 v_lut(const int* tab, const int* idx)
-{
-    int CV_DECL_ALIGNED(32) elems[4] =
-    {
-        tab[idx[0]],
-        tab[idx[1]],
-        tab[idx[2]],
-        tab[idx[3]]
-    };
-    return v_int32x4(vle_v_i32m1(elems, 4));
-}
-inline v_int32x4 v_lut_pairs(const int* tab, const int* idx)
-{
-    int CV_DECL_ALIGNED(32) elems[4] =
-    {
-        tab[idx[0]],
-        tab[idx[0] + 1],
-        tab[idx[1]],
-        tab[idx[1] + 1]
-    };
-    return v_int32x4(vle_v_i32m1(elems, 4));
-}
-inline v_int32x4 v_lut_quads(const int* tab, const int* idx)
-{
-    return v_int32x4(vle_v_i32m1(tab+idx[0], 4));
-}
-inline v_uint32x4 v_lut(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut((int*)tab, idx)); }
-inline v_uint32x4 v_lut_pairs(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut_pairs((int*)tab, idx)); }
-inline v_uint32x4 v_lut_quads(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut_quads((int*)tab, idx)); }
-
-inline v_int64x2 v_lut(const int64_t* tab, const int* idx)
-{
-    vint64m1_t res = {tab[idx[0]], tab[idx[1]]};
-    return v_int64x2(res);
-}
-inline v_int64x2 v_lut_pairs(const int64_t* tab, const int* idx)
-{
-    return v_int64x2(vle_v_i64m1(tab+idx[0], 2));
-}
-
-inline v_uint64x2 v_lut(const uint64_t* tab, const int* idx)
-{
-    vuint64m1_t res = {tab[idx[0]], tab[idx[1]]};
-    return v_uint64x2(res);
-}
-inline v_uint64x2 v_lut_pairs(const uint64_t* tab, const int* idx)
-{
-    return v_uint64x2(vle_v_u64m1(tab+idx[0], 2));
-}
-
-inline v_float32x4 v_lut(const float* tab, const int* idx)
-{
-    float CV_DECL_ALIGNED(32) elems[4] =
-    {
-        tab[idx[0]],
-        tab[idx[1]],
-        tab[idx[2]],
-        tab[idx[3]]
-    };
-    return v_float32x4(vle_v_f32m1(elems, 4));
-}
-inline v_float32x4 v_lut_pairs(const float* tab, const int* idx)
-{
-    float CV_DECL_ALIGNED(32) elems[4] =
-    {
-        tab[idx[0]],
-        tab[idx[0]+1],
-        tab[idx[1]],
-        tab[idx[1]+1]
-    };
-    return v_float32x4(vle_v_f32m1(elems, 4));
-}
-inline v_float32x4 v_lut_quads(const float* tab, const int* idx)
-{
-    return v_float32x4(vle_v_f32m1(tab + idx[0], 4));
-}
-inline v_float64x2 v_lut(const double* tab, const int* idx)
-{
-    vfloat64m1_t res = {tab[idx[0]], tab[idx[1]]};
-    return v_float64x2(res);
-}
-inline v_float64x2 v_lut_pairs(const double* tab, const int* idx)
-{
-    return v_float64x2(vle_v_f64m1(tab+idx[0], 2));
-}
-
-inline v_int32x4 v_lut(const int* tab, const v_int32x4& idxvec)
-{
-    int CV_DECL_ALIGNED(32) elems[4] =
-    {
-        tab[idxvec.val[0]],
-        tab[idxvec.val[1]],
-        tab[idxvec.val[2]],
-        tab[idxvec.val[3]]
-    };
-    return v_int32x4(vle_v_i32m1(elems, 4));
-}
-
-inline v_uint32x4 v_lut(const unsigned* tab, const v_int32x4& idxvec)
-{
-    unsigned CV_DECL_ALIGNED(32) elems[4] =
-    {
-        tab[idxvec.val[0]],
-        tab[idxvec.val[1]],
-        tab[idxvec.val[2]],
-        tab[idxvec.val[3]]
-    };
-    return v_uint32x4(vle_v_u32m1(elems, 4));
-}
-
-inline v_float32x4 v_lut(const float* tab, const v_int32x4& idxvec)
-{
-    float CV_DECL_ALIGNED(32) elems[4] =
-    {
-        tab[idxvec.val[0]],
-        tab[idxvec.val[1]],
-        tab[idxvec.val[2]],
-        tab[idxvec.val[3]]
-    };
-    return v_float32x4(vle_v_f32m1(elems, 4));
-}
-inline v_float64x2 v_lut(const double* tab, const v_int32x4& idxvec)
-{
-    vfloat64m1_t res = {tab[idxvec.val[0]], tab[idxvec.val[1]]};
-    return v_float64x2(res);
-}
-inline void v_lut_deinterleave(const float* tab, const v_int32x4& idxvec, v_float32x4& x, v_float32x4& y)
-{
-    vint32m1_t index_x = vmul_vx_i32m1(idxvec.val, 4, 4);
-    vint32m1_t index_y = vadd_vx_i32m1(index_x, 4, 4);
-
-    x.val = vlxe_v_f32m1(tab, index_x, 4);
-    y.val = vlxe_v_f32m1(tab, index_y, 4);
-}
-
-inline void v_lut_deinterleave(const double* tab, const v_int32x4& idxvec, v_float64x2& x, v_float64x2& y)
-{
-    int CV_DECL_ALIGNED(32) idx[4];
-    v_store_aligned(idx, idxvec);
-
-    x = v_float64x2(tab[idx[0]], tab[idx[1]]);
-    y = v_float64x2(tab[idx[0]+1], tab[idx[1]+1]);
-}
-
-#define OPENCV_HAL_IMPL_RISCVV_PACKS(_Tp, _Tp2, _T2, num2, _T1, num, intrin, shr, _Type) \
-inline v_##_Tp##x##num v_pack(const v_##_Tp2##x##num2& a, const v_##_Tp2##x##num2& b) \
-{ \
-    v##_Tp2##m2_t  tmp = vundefined_##_T2##m2();    \
-    tmp = vset_##_T2##m2(tmp, 0, a.val);    \
-    tmp = vset_##_T2##m2(tmp, 1, b.val);    \
-    return v_##_Tp##x##num(shr##_##_T1##m1(tmp, 0, num)); \
-}\
-template<int n> inline \
-v_##_Tp##x##num v_rshr_pack(const v_##_Tp2##x##num2& a, const v_##_Tp2##x##num2& b) \
-{ \
-    v##_Tp2##m2_t  tmp = vundefined_##_T2##m2();    \
-    tmp = vset_##_T2##m2(tmp, 0, a.val);    \
-    tmp = vset_##_T2##m2(tmp, 1, b.val);    \
-    return v_##_Tp##x##num(intrin##_##_T1##m1(tmp, n, num)); \
-}\
-inline void v_pack_store(_Type* ptr, const v_##_Tp2##x##num2& a) \
-{ \
-    v##_Tp2##m2_t tmp = vundefined_##_T2##m2();    \
-    tmp = vset_##_T2##m2(tmp, 0, a.val);    \
-    tmp = vset_##_T2##m2(tmp, 1, vmv_v_x_##_T2##m1(0, num2));    \
-    asm("" ::: "memory");                                       \
-    vse_v_##_T1##m1(ptr, shr##_##_T1##m1(tmp, 0, num), num2); \
-}\
-template<int n> inline \
-void v_rshr_pack_store(_Type* ptr, const v_##_Tp2##x##num2& a) \
-{ \
-    v##_Tp2##m2_t tmp = vundefined_##_T2##m2();    \
-    tmp = vset_##_T2##m2(tmp, 0, a.val);    \
-    tmp = vset_##_T2##m2(tmp, 1, vmv_v_x_##_T2##m1(0, num2));    \
-    vse_v_##_T1##m1(ptr, intrin##_##_T1##m1(tmp, n, num), num2); \
-}
-OPENCV_HAL_IMPL_RISCVV_PACKS(int8, int16, i16, 8, i8, 16, vnclip_vx, vnclip_vx, signed char)
-OPENCV_HAL_IMPL_RISCVV_PACKS(int16, int32, i32, 4, i16, 8, vnclip_vx, vnclip_vx, signed short)
-OPENCV_HAL_IMPL_RISCVV_PACKS(int32, int64, i64, 2, i32, 4, vnclip_vx, vnsra_vx, int)
-OPENCV_HAL_IMPL_RISCVV_PACKS(uint8, uint16, u16, 8, u8, 16, vnclipu_vx, vnclipu_vx, unsigned char)
-OPENCV_HAL_IMPL_RISCVV_PACKS(uint16, uint32, u32, 4, u16, 8, vnclipu_vx, vnclipu_vx, unsigned short)
-OPENCV_HAL_IMPL_RISCVV_PACKS(uint32, uint64, u64, 2, u32, 4, vnclipu_vx, vnsrl_vx, unsigned int)
-
-// pack boolean
-inline v_uint8x16 v_pack_b(const v_uint16x8& a, const v_uint16x8& b)
-{
-    vuint16m2_t tmp = vundefined_u16m2();    \
-    tmp = vset_u16m2(tmp, 0, a.val);    \
-    tmp = vset_u16m2(tmp, 1, b.val);    \
-    return v_uint8x16(vnsrl_vx_u8m1(tmp, 0, 16));
-}
-
-inline v_uint8x16 v_pack_b(const v_uint32x4& a, const v_uint32x4& b,
-                           const v_uint32x4& c, const v_uint32x4& d)
-{
-    vuint32m4_t vabcd = vundefined_u32m4();    \
-    vuint16m2_t v16 = vundefined_u16m2();    \
-    vabcd = vset_u32m4(vabcd, 0, a.val);    \
-    vabcd = vset_u32m4(vabcd, 1, b.val);    \
-    vabcd = vset_u32m4(vabcd, 2, c.val);    \
-    vabcd = vset_u32m4(vabcd, 3, d.val);    \
-    v16 = vnsrl_vx_u16m2(vabcd, 0, 16);
-    return v_uint8x16(vnsrl_vx_u8m1(v16, 0, 16));
-}
-
-inline v_uint8x16 v_pack_b(const v_uint64x2& a, const v_uint64x2& b, const v_uint64x2& c,
-                           const v_uint64x2& d, const v_uint64x2& e, const v_uint64x2& f,
-                           const v_uint64x2& g, const v_uint64x2& h)
-{
-    vuint64m8_t v64 = vundefined_u64m8();    \
-    vuint32m4_t v32 = vundefined_u32m4();    \
-    vuint16m2_t v16 = vundefined_u16m2();    \
-    v64 = vset_u64m8(v64, 0, a.val);    \
-    v64 = vset_u64m8(v64, 1, b.val);    \
-    v64 = vset_u64m8(v64, 2, c.val);    \
-    v64 = vset_u64m8(v64, 3, d.val);    \
-    v64 = vset_u64m8(v64, 4, e.val);    \
-    v64 = vset_u64m8(v64, 5, f.val);    \
-    v64 = vset_u64m8(v64, 6, g.val);    \
-    v64 = vset_u64m8(v64, 7, h.val);    \
-    v32 = vnsrl_vx_u32m4(v64, 0, 16);
-    v16 = vnsrl_vx_u16m2(v32, 0, 16);
-    return v_uint8x16(vnsrl_vx_u8m1(v16, 0, 16));
-}
-
-//inline v_uint8x16 v_pack_u(const v_int16x8& a, const v_int16x8& b) \
-//{ \
-//    int16xm2_u tmp;    \
-//    tmp.m1[0] = (vint16m1_t)a.val;    \
-//    tmp.m1[1] = (vint16m1_t)b.val;    \
-//    e8xm1_t mask = (e8xm1_t)vmsge_vx_e16xm2_i16m2(tmp.v, 0, 16);\
-//    return v_uint8x16(vnclipuvi_mask_u8m1_u16m2(vmv_v_x_u8m1(0, 16), (vuint16m2_t)tmp.v, 0, mask, 16));
-//}
-
-#define OPENCV_HAL_IMPL_RISCVV_PACK_U(tp1, num1, tp2, num2, _Tp) \
-inline v_uint##tp1##x##num1 v_pack_u(const v_int##tp2##x##num2& a, const v_int##tp2##x##num2& b) \
-{ \
-    vint##tp2##m2_t tmp = vundefined_##i##tp2##m2();    \
-    tmp = vset_##i##tp2##m2(tmp, 0, a.val);    \
-    tmp = vset_##i##tp2##m2(tmp, 1, b.val);    \
-    vint##tp2##m2_t val = vmax_vx_i##tp2##m2(tmp, 0, num1);\
-    return v_uint##tp1##x##num1(vnclipu_vx_u##tp1##m1((vuint##tp2##m2_t)val, 0, num1));    \
-} \
-inline void v_pack_u_store(_Tp* ptr, const v_int##tp2##x##num2& a) \
-{ \
-    vint##tp2##m2_t tmp = vundefined_##i##tp2##m2();    \
-    tmp = vset_##i##tp2##m2(tmp, 0, a.val);    \
-    vint##tp2##m2_t val = vmax_vx_i##tp2##m2(tmp, 0, num1);\
-    return vse_v_u##tp1##m1(ptr, vnclipu_vx_u##tp1##m1((vuint##tp2##m2_t)val, 0, num1), num2);    \
-} \
-template<int n> inline \
-v_uint##tp1##x##num1 v_rshr_pack_u(const v_int##tp2##x##num2& a, const v_int##tp2##x##num2& b) \
-{ \
-    vint##tp2##m2_t tmp = vundefined_##i##tp2##m2();    \
-    tmp = vset_##i##tp2##m2(tmp, 0, a.val);    \
-    tmp = vset_##i##tp2##m2(tmp, 1, b.val);    \
-    vint##tp2##m2_t val = vmax_vx_i##tp2##m2(tmp, 0, num1);\
-    return v_uint##tp1##x##num1(vnclipu_vx_u##tp1##m1((vuint##tp2##m2_t)val, n, num1));    \
-} \
-template<int n> inline \
-void v_rshr_pack_u_store(_Tp* ptr, const v_int##tp2##x##num2& a) \
-{ \
-    vint##tp2##m2_t tmp = vundefined_##i##tp2##m2();    \
-    tmp = vset_##i##tp2##m2(tmp, 0, a.val);    \
-    vint##tp2##m2_t val_ = vmax_vx_i##tp2##m2(tmp, 0, num1);\
-    vuint##tp1##m1_t val = vnclipu_vx_u##tp1##m1((vuint##tp2##m2_t)val_, n, num1);    \
-    return vse_v_u##tp1##m1(ptr, val, num2);\
-}
-OPENCV_HAL_IMPL_RISCVV_PACK_U(8, 16, 16, 8, unsigned char )
-OPENCV_HAL_IMPL_RISCVV_PACK_U(16, 8, 32, 4, unsigned short)
-
-#ifdef __GNUC__
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wuninitialized"
-#endif
-
-// saturating multiply 8-bit, 16-bit
-#define OPENCV_HAL_IMPL_RISCVV_MUL_SAT(_Tpvec, _Tpwvec)            \
-    inline _Tpvec operator * (const _Tpvec& a, const _Tpvec& b)  \
-    {                                                            \
-        _Tpwvec c, d;                                            \
-        v_mul_expand(a, b, c, d);                                \
-        return v_pack(c, d);                                     \
-    }                                                            \
-    inline _Tpvec& operator *= (_Tpvec& a, const _Tpvec& b)      \
-    { a = a * b; return a; }
-
-OPENCV_HAL_IMPL_RISCVV_MUL_SAT(v_int8x16,  v_int16x8)
-OPENCV_HAL_IMPL_RISCVV_MUL_SAT(v_uint8x16, v_uint16x8)
-OPENCV_HAL_IMPL_RISCVV_MUL_SAT(v_int16x8,  v_int32x4)
-OPENCV_HAL_IMPL_RISCVV_MUL_SAT(v_uint16x8, v_uint32x4)
-
-#ifdef __GNUC__
-#pragma GCC diagnostic pop
-#endif
-static const signed char popCountTable[256] =
-{
-    0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    4, 5, 5, 6, 5, 6, 6, 7, 5, 6, 6, 7, 6, 7, 7, 8,
-};
-
-inline vuint8m1_t vcnt_u8(vuint8m1_t val){
-    vuint8m1_t v0 = val & 1;
-    return vlxe_v_u8m1((unsigned char*)popCountTable, val >> 1, 16)+v0;
-}
-
-inline v_uint8x16
-v_popcount(const v_uint8x16& a)
-{
-    return v_uint8x16(vcnt_u8(a.val));
-}
-
-inline v_uint8x16
-v_popcount(const v_int8x16& a)
-{
-    return v_uint8x16(vcnt_u8((vuint8m1_t)a.val));
-}
-
-inline v_uint16x8
-v_popcount(const v_uint16x8& a)
-{
-    vuint8m2_t tmp = vundefined_u8m2();
-    tmp = vset_u8m2(tmp, 0, vcnt_u8((vuint8m1_t)a.val));
-    vuint64m2_t mask = (vuint64m2_t){0x0E0C0A0806040200, 0, 0x0F0D0B0907050301, 0};
-    tmp = vrgather_vv_u8m2(tmp, (vuint8m2_t)mask, 32);    \
-    vuint16m2_t res = vwaddu_vv_u16m2(vget_u8m2_u8m1(tmp, 0), vget_u8m2_u8m1(tmp, 1), 8);
-    return v_uint16x8(vget_u16m2_u16m1(res, 0));
-}
-
-inline v_uint16x8
-v_popcount(const v_int16x8& a)
-{
-    vuint8m2_t tmp = vundefined_u8m2();
-    tmp = vset_u8m2(tmp, 0, vcnt_u8((vuint8m1_t)a.val));
-    vuint64m2_t mask = (vuint64m2_t){0x0E0C0A0806040200, 0, 0x0F0D0B0907050301, 0};
-    tmp = vrgather_vv_u8m2(tmp, (vuint8m2_t)mask, 32);    \
-    vuint16m2_t res = vwaddu_vv_u16m2(vget_u8m2_u8m1(tmp, 0), vget_u8m2_u8m1(tmp, 1), 8);
-    return v_uint16x8(vget_u16m2_u16m1(res, 0));
-}
-
-inline v_uint32x4
-v_popcount(const v_uint32x4& a)
-{
-    vuint8m2_t tmp = vundefined_u8m2();
-    tmp = vset_u8m2(tmp, 0, vcnt_u8((vuint8m1_t)a.val));
-    vuint64m2_t mask = (vuint64m2_t){0xFFFFFFFF0C080400, 0xFFFFFFFF0D090501,
-                     0xFFFFFFFF0E0A0602, 0xFFFFFFFF0F0B0703};
-    tmp = vrgather_vv_u8m2(tmp, (vuint8m2_t)mask, 32);    \
-    vuint16m2_t res_ = vwaddu_vv_u16m2(vget_u8m2_u8m1(tmp, 0), vget_u8m2_u8m1(tmp, 1), 16);
-    vuint32m2_t res  = vwaddu_vv_u32m2(vget_u16m2_u16m1(res_, 0), vget_u16m2_u16m1(res_, 1), 8);
-    return v_uint32x4(vget_u32m2_u32m1(res, 0));
-}
-
-inline v_uint32x4
-v_popcount(const v_int32x4& a)
-{
-    vuint8m2_t tmp = vundefined_u8m2();
-    tmp = vset_u8m2(tmp, 0, vcnt_u8((vuint8m1_t)a.val));
-    vuint64m2_t mask = (vuint64m2_t){0xFFFFFFFF0C080400, 0xFFFFFFFF0D090501,
-                     0xFFFFFFFF0E0A0602, 0xFFFFFFFF0F0B0703};
-    tmp = vrgather_vv_u8m2(tmp, (vuint8m2_t)mask, 32);    \
-    vuint16m2_t res_ = vwaddu_vv_u16m2(vget_u8m2_u8m1(tmp, 0), vget_u8m2_u8m1(tmp, 1), 16);
-    vuint32m2_t res  = vwaddu_vv_u32m2(vget_u16m2_u16m1(res_, 0), vget_u16m2_u16m1(res_, 1), 8);
-    return v_uint32x4(vget_u32m2_u32m1(res, 0));
-}
-
-inline v_uint64x2
-v_popcount(const v_uint64x2& a)
-{
-    vuint8m2_t tmp = vundefined_u8m2();
-    tmp = vset_u8m2(tmp, 0, vcnt_u8((vuint8m1_t)a.val));
-    vuint64m2_t mask = (vuint64m2_t){0x0706050403020100, 0x0000000000000000,
-                     0x0F0E0D0C0B0A0908, 0x0000000000000000};
-    tmp = vrgather_vv_u8m2(tmp, (vuint8m2_t)mask, 32);    \
-    vuint8m1_t zero = vmv_v_x_u8m1(0, 16);
-    vuint8m1_t res1 = zero;
-    vuint8m1_t res2 = zero;
-    res1 = vredsum_vs_u8m1_u8m1(res1, vget_u8m2_u8m1(tmp, 0), zero, 8);
-    res2 = vredsum_vs_u8m1_u8m1(res2, vget_u8m2_u8m1(tmp, 1), zero, 8);
-
-    return v_uint64x2((unsigned long)vmv_x_s_u8m1_u8(res1, 8), (unsigned long)vmv_x_s_u8m1_u8(res2, 8));
-}
-
-inline v_uint64x2
-v_popcount(const v_int64x2& a)
-{
-    vuint8m2_t tmp = vundefined_u8m2();
-    tmp = vset_u8m2(tmp, 0, vcnt_u8((vuint8m1_t)a.val));
-    vuint64m2_t mask = (vuint64m2_t){0x0706050403020100, 0x0000000000000000,
-                     0x0F0E0D0C0B0A0908, 0x0000000000000000};
-    tmp = vrgather_vv_u8m2(tmp, (vuint8m2_t)mask, 32);    \
-    vuint8m1_t zero = vmv_v_x_u8m1(0, 16);
-    vuint8m1_t res1 = zero;
-    vuint8m1_t res2 = zero;
-    res1 = vredsum_vs_u8m1_u8m1(res1, vget_u8m2_u8m1(tmp, 0), zero, 8);
-    res2 = vredsum_vs_u8m1_u8m1(res2, vget_u8m2_u8m1(tmp, 1), zero, 8);
-
-    return v_uint64x2((unsigned long)vmv_x_s_u8m1_u8(res1, 8), (unsigned long)vmv_x_s_u8m1_u8(res2, 8));
-}
-
-#define SMASK 1, 2, 4, 8, 16, 32, 64, 128
-inline int v_signmask(const v_uint8x16& a)
-{
-    vuint8m1_t t0  = vsrl_vx_u8m1(a.val, 7, 16);
-    vuint8m1_t m1  = (vuint8m1_t){SMASK, SMASK};
-    vuint16m2_t t1 = vwmulu_vv_u16m2(t0, m1, 16);
-    vuint32m1_t res = vmv_v_x_u32m1(0, 4);
-    vuint32m2_t t2 = vwmulu_vx_u32m2(vget_u16m2_u16m1(t1, 1), 256, 8);
-    res = vredsum_vs_u32m2_u32m1(res, t2, res, 8);
-    res = vwredsumu_vs_u16m1_u32m1(res, vget_u16m2_u16m1(t1, 0), res, 8);
-    return vmv_x_s_u32m1_u32(res, 8);
-}
-inline int v_signmask(const v_int8x16& a)
-{
-    vuint8m1_t t0 = vsrl_vx_u8m1((vuint8m1_t)a.val, 7, 16);
-    vuint8m1_t m1 = (vuint8m1_t){SMASK, SMASK};
-    vint16m2_t t1 = (vint16m2_t)vwmulu_vv_u16m2(t0, m1, 16);
-    vint32m1_t res = vmv_v_x_i32m1(0, 4);
-    vint32m2_t t2 = vwmul_vx_i32m2(vget_i16m2_i16m1(t1, 1), 256, 8);
-    res = vredsum_vs_i32m2_i32m1(res, t2, res, 8);
-    res = vwredsum_vs_i16m1_i32m1(res, vget_i16m2_i16m1(t1, 0), res, 8);
-    return vmv_x_s_i32m1_i32(res, 8);
-}
-
-inline int v_signmask(const v_int16x8& a)
-{
-    vint16m1_t t0 = (vint16m1_t)vsrl_vx_u16m1((vuint16m1_t)a.val, 15, 8);
-    vint16m1_t m1 = (vint16m1_t){SMASK};
-    vint16m1_t t1 = vmul_vv_i16m1(t0, m1, 8);
-    vint16m1_t res = vmv_v_x_i16m1(0, 8);
-    res = vredsum_vs_i16m1_i16m1(res, t1, res, 8);
-    return vmv_x_s_i16m1_i16(res, 8);
-}
-inline int v_signmask(const v_uint16x8& a)
-{
-    vint16m1_t t0 = (vint16m1_t)vsrl_vx_u16m1((vuint16m1_t)a.val, 15, 8);
-    vint16m1_t m1 = (vint16m1_t){SMASK};
-    vint16m1_t t1 = vmul_vv_i16m1(t0, m1, 8);
-    vint16m1_t res = vmv_v_x_i16m1(0, 8);
-    res = vredsum_vs_i16m1_i16m1(res, t1, res, 8);
-    return vmv_x_s_i16m1_i16(res, 8);
-}
-inline int v_signmask(const v_int32x4& a)
-{
-    vint32m1_t t0 = (vint32m1_t)vsrl_vx_u32m1((vuint32m1_t)a.val, 31, 4);
-    vint32m1_t m1 = (vint32m1_t){1, 2, 4, 8};
-    vint32m1_t res = vmv_v_x_i32m1(0, 4);
-    vint32m1_t t1 = vmul_vv_i32m1(t0, m1, 4);
-    res = vredsum_vs_i32m1_i32m1(res, t1, res, 4);
-    return vmv_x_s_i32m1_i32(res, 4);
-}
-inline int v_signmask(const v_uint32x4& a)
-{
-    vint32m1_t t0 = (vint32m1_t)vsrl_vx_u32m1(a.val, 31, 4);
-    vint32m1_t m1 = (vint32m1_t){1, 2, 4, 8};
-    vint32m1_t res = vmv_v_x_i32m1(0, 4);
-    vint32m1_t t1 = vmul_vv_i32m1(t0, m1, 4);
-    res = vredsum_vs_i32m1_i32m1(res, t1, res, 4);
-    return vmv_x_s_i32m1_i32(res, 4);
-}
-inline int v_signmask(const v_uint64x2& a)
-{
-    vuint64m1_t v0 = vsrl_vx_u64m1(a.val, 63, 2);
-    int res = (int)vext_x_v_u64m1_u64(v0, 0, 2) + ((int)vext_x_v_u64m1_u64(v0, 1, 2) << 1);
-    return res;
-}
-inline int v_signmask(const v_int64x2& a)
-{ return v_signmask(v_reinterpret_as_u64(a)); }
-inline int v_signmask(const v_float64x2& a)
-{ return v_signmask(v_reinterpret_as_u64(a)); }
-inline int v_signmask(const v_float32x4& a)
-{
-    vint32m1_t t0 = (vint32m1_t)vsrl_vx_u32m1((vuint32m1_t)a.val, 31, 4);
-    vint32m1_t m1 = (vint32m1_t){1, 2, 4, 8};
-    vint32m1_t res = vmv_v_x_i32m1(0, 4);
-    vint32m1_t t1 = vmul_vv_i32m1(t0, m1, 4);
-    res = vredsum_vs_i32m1_i32m1(res, t1, res, 4);
-    return vmv_x_s_i32m1_i32(res, 4);
-}
-
-inline int v_scan_forward(const v_int8x16& a) {
-int val = v_signmask(a);
-if(val==0) return 0;
-else return trailingZeros32(val); }
-inline int v_scan_forward(const v_uint8x16& a) {
-int val = v_signmask(a);
-if(val==0) return 0;
-else return trailingZeros32(val); }
-inline int v_scan_forward(const v_int16x8& a) {
-int val = v_signmask(a);
-if(val==0) return 0;
-else return trailingZeros32(val); }
-inline int v_scan_forward(const v_uint16x8& a) {
-int val = v_signmask(a);
-if(val==0) return 0;
-else return trailingZeros32(val); }
-inline int v_scan_forward(const v_int32x4& a) {
-int val = v_signmask(a);
-if(val==0) return 0;
-else return trailingZeros32(val); }
-inline int v_scan_forward(const v_uint32x4& a) {
-int val = v_signmask(a);
-if(val==0) return 0;
-else return trailingZeros32(val); }
-inline int v_scan_forward(const v_float32x4& a) {
-int val = v_signmask(a);
-if(val==0) return 0;
-else return trailingZeros32(val); }
-inline int v_scan_forward(const v_int64x2& a) {
-int val = v_signmask(a);
-if(val==0) return 0;
-else return trailingZeros32(val); }
-inline int v_scan_forward(const v_uint64x2& a) {
-int val = v_signmask(a);
-if(val==0) return 0;
-else return trailingZeros32(val); }
-
-#define OPENCV_HAL_IMPL_RISCVV_CHECK_ALLANY(_Tpvec, suffix, _T, shift, num) \
-inline bool v_check_all(const v_##_Tpvec& a) \
-{ \
-    suffix##m1_t v0 = vsrl_vx_##_T(vnot_v_##_T(a.val, num), shift, num); \
-    vuint64m1_t v1 = vuint64m1_t(v0); \
-    return (v1[0] | v1[1]) == 0; \
-} \
-inline bool v_check_any(const v_##_Tpvec& a) \
-{ \
-    suffix##m1_t v0 = vsrl_vx_##_T(a.val, shift, num); \
-    vuint64m1_t v1 = vuint64m1_t(v0); \
-    return (v1[0] | v1[1]) != 0; \
-}
-
-OPENCV_HAL_IMPL_RISCVV_CHECK_ALLANY(uint8x16, vuint8,  u8m1, 7, 16)
-OPENCV_HAL_IMPL_RISCVV_CHECK_ALLANY(uint16x8, vuint16, u16m1, 15, 8)
-OPENCV_HAL_IMPL_RISCVV_CHECK_ALLANY(uint32x4, vuint32, u32m1, 31, 4)
-OPENCV_HAL_IMPL_RISCVV_CHECK_ALLANY(uint64x2, vuint64, u64m1, 63, 2)
-
-inline bool v_check_all(const v_int8x16& a)
-{ return v_check_all(v_reinterpret_as_u8(a)); }
-inline bool v_check_all(const v_int16x8& a)
-{ return v_check_all(v_reinterpret_as_u16(a)); }
-inline bool v_check_all(const v_int32x4& a)
-{ return v_check_all(v_reinterpret_as_u32(a)); }
-inline bool v_check_all(const v_float32x4& a)
-{ return v_check_all(v_reinterpret_as_u32(a)); }
-inline bool v_check_all(const v_int64x2& a)
-{ return v_check_all(v_reinterpret_as_u64(a)); }
-inline bool v_check_all(const v_float64x2& a)
-{ return v_check_all(v_reinterpret_as_u64(a)); }
-
-inline bool v_check_any(const v_int8x16& a)
-{ return v_check_any(v_reinterpret_as_u8(a)); }
-inline bool v_check_any(const v_int16x8& a)
-{ return v_check_any(v_reinterpret_as_u16(a)); }
-inline bool v_check_any(const v_int32x4& a)
-{ return v_check_any(v_reinterpret_as_u32(a)); }
-inline bool v_check_any(const v_float32x4& a)
-{ return v_check_any(v_reinterpret_as_u32(a)); }
-inline bool v_check_any(const v_int64x2& a)
-{ return v_check_any(v_reinterpret_as_u64(a)); }
-inline bool v_check_any(const v_float64x2& a)
-{ return v_check_any(v_reinterpret_as_u64(a)); }
-
-#define OPENCV_HAL_IMPL_RISCVV_SELECT(_Tpvec, suffix, _Tpvec2, num) \
-inline _Tpvec v_select(const _Tpvec& mask, const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(vmerge_vvm_##suffix(_Tpvec2(mask.val), b.val, a.val, num)); \
-}
-
-OPENCV_HAL_IMPL_RISCVV_SELECT(v_int8x16,  i8m1, vbool8_t, 16)
-OPENCV_HAL_IMPL_RISCVV_SELECT(v_int16x8,  i16m1, vbool16_t, 8)
-OPENCV_HAL_IMPL_RISCVV_SELECT(v_int32x4,  i32m1, vbool32_t, 4)
-OPENCV_HAL_IMPL_RISCVV_SELECT(v_uint8x16, u8m1, vbool8_t, 16)
-OPENCV_HAL_IMPL_RISCVV_SELECT(v_uint16x8, u16m1, vbool16_t, 8)
-OPENCV_HAL_IMPL_RISCVV_SELECT(v_uint32x4, u32m1, vbool32_t, 4)
-inline v_float32x4 v_select(const v_float32x4& mask, const v_float32x4& a, const v_float32x4& b)
-{
-    return v_float32x4((vfloat32m1_t)vmerge_vvm_u32m1((vbool32_t)mask.val, (vuint32m1_t)b.val, (vuint32m1_t)a.val, 4));
-}
-inline v_float64x2 v_select(const v_float64x2& mask, const v_float64x2& a, const v_float64x2& b)
-{
-    return v_float64x2((vfloat64m1_t)vmerge_vvm_u64m1((vbool64_t)mask.val, (vuint64m1_t)b.val, (vuint64m1_t)a.val, 2));
-}
-
-#define OPENCV_HAL_IMPL_RISCVV_EXPAND(add, _Tpvec, _Tpwvec, _Tp, _Tp1, num1, _Tp2, num2, _T1, _T2) \
-inline void v_expand(const _Tpvec& a, v_##_Tpwvec& b0, v_##_Tpwvec& b1) \
-{ \
-    _T1##_t b = vw##add##_vv_##_Tp2##m2(a.val, vmv_v_x_##_Tp1(0, num1), num1);    \
-    b0.val = vget_##_Tp2##m2_##_Tp2##m1(b, 0);  \
-    b1.val = vget_##_Tp2##m2_##_Tp2##m1(b, 1);  \
-} \
-inline v_##_Tpwvec v_expand_low(const _Tpvec& a) \
-{ \
-    _T1##_t b = vw##add##_vv_##_Tp2##m2(a.val, vmv_v_x_##_Tp1(0, num2), num2);    \
-    return v_##_Tpwvec(vget_##_Tp2##m2_##_Tp2##m1(b, 0)); \
-} \
-inline v_##_Tpwvec v_expand_high(const _Tpvec& a) \
-{ \
-    _T1##_t b = vw##add##_vv_##_Tp2##m2(a.val, vmv_v_x_##_Tp1(0, num1), num1);    \
-    return v_##_Tpwvec(vget_##_Tp2##m2_##_Tp2##m1(b, 1)); \
-} \
-inline v_##_Tpwvec v_load_expand(const _Tp* ptr) \
-{ \
-    _T2##_t val = vle##_v_##_Tp1(ptr, num2);    \
-    _T1##_t b = vw##add##_vv_##_Tp2##m2(val, vmv_v_x_##_Tp1(0, num2), num2);    \
-    return v_##_Tpwvec(vget_##_Tp2##m2_##_Tp2##m1(b, 0)); \
-}
-
-OPENCV_HAL_IMPL_RISCVV_EXPAND(addu, v_uint8x16, uint16x8, uchar, u8m1, 16, u16, 8, vuint16m2, vuint8m1)
-OPENCV_HAL_IMPL_RISCVV_EXPAND(addu, v_uint16x8, uint32x4, ushort,  u16m1, 8, u32, 4, vuint32m2, vuint16m1)
-OPENCV_HAL_IMPL_RISCVV_EXPAND(addu, v_uint32x4, uint64x2, uint,  u32m1, 4, u64, 2, vuint64m2, vuint32m1)
-OPENCV_HAL_IMPL_RISCVV_EXPAND(add, v_int8x16, int16x8, schar,  i8m1, 16, i16, 8, vint16m2, vint8m1)
-OPENCV_HAL_IMPL_RISCVV_EXPAND(add, v_int16x8, int32x4, short,  i16m1, 8, i32, 4, vint32m2, vint16m1)
-OPENCV_HAL_IMPL_RISCVV_EXPAND(add, v_int32x4, int64x2, int,  i32m1, 4, i64, 2, vint64m2, vint32m1)
-
-inline v_uint32x4 v_load_expand_q(const uchar* ptr)
-{
-    vuint16m2_t b = vundefined_u16m2();
-    vuint32m2_t c = vundefined_u32m2();
-    vuint8m1_t val = vle_v_u8m1(ptr, 4);    \
-    b = vwaddu_vv_u16m2(val, vmv_v_x_u8m1(0, 4), 4);    \
-    c = vwaddu_vv_u32m2(vget_u16m2_u16m1(b, 0), vmv_v_x_u16m1(0, 4), 4);    \
-    return v_uint32x4(vget_u32m2_u32m1(c, 0));
-}
-
-inline v_int32x4 v_load_expand_q(const schar* ptr)
-{
-    vint16m2_t b = vundefined_i16m2();
-    vint32m2_t c = vundefined_i32m2();
-    vint8m1_t val = vle_v_i8m1(ptr, 4);    \
-    b = vwadd_vv_i16m2(val, vmv_v_x_i8m1(0, 4), 4);    \
-    c = vwadd_vv_i32m2(vget_i16m2_i16m1(b, 0), vmv_v_x_i16m1(0, 4), 4);    \
-    return v_int32x4(vget_i32m2_i32m1(c, 0));
-}
-#define VITL_16 (vuint64m2_t){0x1303120211011000, 0x1707160615051404, 0x1B0B1A0A19091808, 0x1F0F1E0E1D0D1C0C}
-#define VITL_8 (vuint64m2_t){0x0009000100080000, 0x000B0003000A0002, 0x000D0005000C0004, 0x000F0007000E0006}
-#define VITL_4 (vuint64m2_t){0x0000000400000000, 0x0000000500000001, 0x0000000600000002, 0x0000000700000003}
-#define VITL_2 (vuint64m2_t){0, 2, 1, 3}
-#define LOW_4  0x0000000100000000, 0x0000000500000004
-#define LOW_8  0x0003000200010000, 0x000B000A00090008
-#define LOW_16 0x0706050403020100, 0x1716151413121110
-#define HIGH_4  0x0000000300000002, 0x0000000700000006
-#define HIGH_8  0x0007000600050004, 0x000F000E000D000C
-#define HIGH_16 0x0F0E0D0C0B0A0908,  0x1F1E1D1C1B1A1918
-#define OPENCV_HAL_IMPL_RISCVV_UNPACKS(_Tpvec, _Tp, _T, _UTp, _UT, num, num2, len, numh) \
-inline void v_zip(const v_##_Tpvec& a0, const v_##_Tpvec& a1, v_##_Tpvec& b0, v_##_Tpvec& b1) \
-{ \
-    v##_Tp##m2_t tmp = vundefined_##_T##m2();\
-    tmp = vset_##_T##m2(tmp, 0, a0.val); \
-    tmp = vset_##_T##m2(tmp, 1, a1.val); \
-    vuint64m2_t mask = VITL_##num;    \
-    tmp = (v##_Tp##m2_t)vrgather_vv_##_T##m2((v##_Tp##m2_t)tmp, (v##_UTp##m2_t)mask, num2);    \
-    b0.val = vget_##_T##m2_##_T##m1(tmp, 0); \
-    b1.val = vget_##_T##m2_##_T##m1(tmp, 1); \
-} \
-inline v_##_Tpvec v_combine_low(const v_##_Tpvec& a, const v_##_Tpvec& b) \
-{ \
-    v##_Tp##m1_t b0 = vslideup_vx_##_T##m1_m(vmset_m_##len(num), a.val, b.val, numh, num);    \
-    return v_##_Tpvec(b0);\
-} \
-inline v_##_Tpvec v_combine_high(const v_##_Tpvec& a, const v_##_Tpvec& b) \
-{ \
-    v##_Tp##m1_t b0 = vslidedown_vx_##_T##m1(b.val, numh, num);    \
-    v##_Tp##m1_t a0 = vslidedown_vx_##_T##m1(a.val, numh, num);    \
-    v##_Tp##m1_t b1 = vslideup_vx_##_T##m1_m(vmset_m_##len(num), a0, b0, numh, num);    \
-    return v_##_Tpvec(b1);\
-} \
-inline void v_recombine(const v_##_Tpvec& a, const v_##_Tpvec& b, v_##_Tpvec& c, v_##_Tpvec& d) \
-{ \
-    c.val = vslideup_vx_##_T##m1_m(vmset_m_##len(num), a.val, b.val, numh, num);    \
-    v##_Tp##m1_t b0 = vslidedown_vx_##_T##m1(b.val, numh, num);    \
-    v##_Tp##m1_t a0 = vslidedown_vx_##_T##m1(a.val, numh, num);    \
-    d.val = vslideup_vx_##_T##m1_m(vmset_m_##len(num), a0, b0, numh, num);    \
-}
-
-OPENCV_HAL_IMPL_RISCVV_UNPACKS(uint8x16, uint8, u8, uint8, u8, 16, 32, b8, 8)
-OPENCV_HAL_IMPL_RISCVV_UNPACKS(int8x16, int8, i8, uint8, u8, 16, 32, b8, 8)
-OPENCV_HAL_IMPL_RISCVV_UNPACKS(uint16x8, uint16, u16, uint16, u16, 8, 16, b16, 4)
-OPENCV_HAL_IMPL_RISCVV_UNPACKS(int16x8, int16, i16, uint16, u16, 8, 16, b16, 4)
-OPENCV_HAL_IMPL_RISCVV_UNPACKS(uint32x4, uint32, u32, uint32, u32, 4, 8, b32, 2)
-OPENCV_HAL_IMPL_RISCVV_UNPACKS(int32x4, int32, i32, uint32, u32, 4, 8, b32, 2)
-OPENCV_HAL_IMPL_RISCVV_UNPACKS(float32x4, float32, f32, uint32, u32, 4, 8, b32, 2)
-OPENCV_HAL_IMPL_RISCVV_UNPACKS(float64x2, float64, f64, uint64, u64, 2, 4, b64, 1)
-
-inline v_uint8x16 v_reverse(const v_uint8x16 &a)
-{
-    vuint64m1_t mask = (vuint64m1_t){0x08090A0B0C0D0E0F, 0x0001020304050607};
-    return v_uint8x16(vrgather_vv_u8m1(a.val, (vuint8m1_t)mask, 16));
-}
-inline v_int8x16 v_reverse(const v_int8x16 &a)
-{
-    vint64m1_t mask = (vint64m1_t){0x08090A0B0C0D0E0F, 0x0001020304050607};
-    return v_int8x16(vrgather_vv_i8m1(a.val, (vuint8m1_t)mask, 16));
-}
-
-inline v_uint16x8 v_reverse(const v_uint16x8 &a)
-{
-    vuint64m1_t mask = (vuint64m1_t){0x0004000500060007, 0x000000100020003};
-    return v_uint16x8(vrgather_vv_u16m1(a.val, (vuint16m1_t)mask, 8));
-}
-
-inline v_int16x8 v_reverse(const v_int16x8 &a)
-{
-    vint64m1_t mask = (vint64m1_t){0x0004000500060007, 0x000000100020003};
-    return v_int16x8(vrgather_vv_i16m1(a.val, (vuint16m1_t)mask, 8));
-}
-inline v_uint32x4 v_reverse(const v_uint32x4 &a)
-{
-    return v_uint32x4(vrgather_vv_u32m1(a.val, (vuint32m1_t){3, 2, 1, 0}, 4));
-}
-
-inline v_int32x4 v_reverse(const v_int32x4 &a)
-{
-    return v_int32x4(vrgather_vv_i32m1(a.val, (vuint32m1_t){3, 2, 1, 0}, 4));
-}
-
-inline v_float32x4 v_reverse(const v_float32x4 &a)
-{ return v_reinterpret_as_f32(v_reverse(v_reinterpret_as_u32(a))); }
-
-inline v_uint64x2 v_reverse(const v_uint64x2 &a)
-{
-    return v_uint64x2(a.val[1], a.val[0]);
-}
-
-inline v_int64x2 v_reverse(const v_int64x2 &a)
-{
-    return v_int64x2(a.val[1], a.val[0]);
-}
-
-inline v_float64x2 v_reverse(const v_float64x2 &a)
-{
-    return v_float64x2(a.val[1], a.val[0]);
-}
-
-#define OPENCV_HAL_IMPL_RISCVV_EXTRACT(_Tpvec, suffix, size) \
-template <int n> \
-inline _Tpvec v_extract(const _Tpvec& a, const _Tpvec& b) \
-{ return v_rotate_right<n>(a, b);}
-OPENCV_HAL_IMPL_RISCVV_EXTRACT(v_uint8x16, u8, 0)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT(v_int8x16, s8, 0)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT(v_uint16x8, u16, 1)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT(v_int16x8, s16, 1)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT(v_uint32x4, u32, 2)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT(v_int32x4, s32, 2)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT(v_uint64x2, u64, 3)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT(v_int64x2, s64, 3)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT(v_float32x4, f32, 2)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT(v_float64x2, f64, 3)
-
-
-#define OPENCV_HAL_IMPL_RISCVV_EXTRACT_N(_Tpvec, _Tp, suffix) \
-template<int i> inline _Tp v_extract_n(_Tpvec v) { return v.val[i]; }
-
-OPENCV_HAL_IMPL_RISCVV_EXTRACT_N(v_uint8x16, uchar, u8)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT_N(v_int8x16, schar, s8)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT_N(v_uint16x8, ushort, u16)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT_N(v_int16x8, short, s16)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT_N(v_uint32x4, uint, u32)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT_N(v_int32x4, int, s32)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT_N(v_uint64x2, uint64, u64)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT_N(v_int64x2, int64, s64)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT_N(v_float32x4, float, f32)
-OPENCV_HAL_IMPL_RISCVV_EXTRACT_N(v_float64x2, double, f64)
-
-#define OPENCV_HAL_IMPL_RISCVV_BROADCAST(_Tpvec, _Tp, num) \
-template<int i> inline _Tpvec v_broadcast_element(_Tpvec v) { return _Tpvec(vrgather_vx_##_Tp##m1(v.val, i, num)); }
-
-OPENCV_HAL_IMPL_RISCVV_BROADCAST(v_uint8x16, u8, 16)
-OPENCV_HAL_IMPL_RISCVV_BROADCAST(v_int8x16, i8, 16)
-OPENCV_HAL_IMPL_RISCVV_BROADCAST(v_uint16x8, u16, 8)
-OPENCV_HAL_IMPL_RISCVV_BROADCAST(v_int16x8, i16, 8)
-OPENCV_HAL_IMPL_RISCVV_BROADCAST(v_uint32x4, u32, 4)
-OPENCV_HAL_IMPL_RISCVV_BROADCAST(v_int32x4, i32, 4)
-OPENCV_HAL_IMPL_RISCVV_BROADCAST(v_uint64x2, u64, 2)
-OPENCV_HAL_IMPL_RISCVV_BROADCAST(v_int64x2, i64, 2)
-OPENCV_HAL_IMPL_RISCVV_BROADCAST(v_float32x4, f32, 4)
-inline v_int32x4 v_round(const v_float32x4& a)
-{
-    __builtin_riscv_fsrm(0);
-    vint32m1_t nan = vand_vx_i32m1((vint32m1_t)a.val, 0x7f800000, 4);
-    vbool32_t mask = vmsne_vx_i32m1_b32(nan, 0x7f800000, 4);
-    vint32m1_t val = vfcvt_x_f_v_i32m1_m(mask, vmv_v_x_i32m1(0, 4), a.val, 4);
-    __builtin_riscv_fsrm(0);
-    return v_int32x4(val);
-}
-inline v_int32x4 v_floor(const v_float32x4& a)
-{
-    __builtin_riscv_fsrm(2);
-    vint32m1_t nan = vand_vx_i32m1((vint32m1_t)a.val, 0x7f800000, 4);
-    vbool32_t mask = vmsne_vx_i32m1_b32(nan, 0x7f800000, 4);
-    vint32m1_t val = vfcvt_x_f_v_i32m1_m(mask, vmv_v_x_i32m1(0, 4), a.val, 4);
-    __builtin_riscv_fsrm(0);
-    return v_int32x4(val);
-}
-
-inline v_int32x4 v_ceil(const v_float32x4& a)
-{
-    __builtin_riscv_fsrm(3);
-    vint32m1_t nan = vand_vx_i32m1((vint32m1_t)a.val, 0x7f800000, 4);
-    vbool32_t mask = vmsne_vx_i32m1_b32(nan, 0x7f800000, 4);
-    vint32m1_t val = vfcvt_x_f_v_i32m1_m(mask, vmv_v_x_i32m1(0, 4), a.val, 4);
-    __builtin_riscv_fsrm(0);
-    return v_int32x4(val);
-}
-
-inline v_int32x4 v_trunc(const v_float32x4& a)
-{
-    __builtin_riscv_fsrm(1);
-    vint32m1_t nan = vand_vx_i32m1((vint32m1_t)a.val, 0x7f800000, 4);
-    vbool32_t mask = vmsne_vx_i32m1_b32(nan, 0x7f800000, 4);
-    vint32m1_t val = vfcvt_x_f_v_i32m1_m(mask, vmv_v_x_i32m1(0, 4), a.val, 4);
-    __builtin_riscv_fsrm(0);
-    return v_int32x4(val);
-}
-
-inline v_int32x4 v_round(const v_float64x2& a)
-{
-    __builtin_riscv_fsrm(0);
-    vfloat64m2_t _val = vundefined_f64m2();
-    _val = vset_f64m2(_val, 0, a.val);
-    //_val = vset_f64m2(_val, 1, a.val);
-    _val = vset_f64m2(_val, 1, vfmv_v_f_f64m1(0, 2));
-    vint32m1_t val = vfncvt_x_f_v_i32m1(_val, 4);
-    __builtin_riscv_fsrm(0);
-    return v_int32x4(val);
-}
-inline v_int32x4 v_round(const v_float64x2& a, const v_float64x2& b)
-{
-    __builtin_riscv_fsrm(0);
-    vfloat64m2_t _val = vundefined_f64m2();
-    _val = vset_f64m2(_val, 0, a.val);
-    _val = vset_f64m2(_val, 1, b.val);
-    vint32m1_t val = vfncvt_x_f_v_i32m1(_val, 4);
-    __builtin_riscv_fsrm(0);
-    return v_int32x4(val);
-}
-inline v_int32x4 v_floor(const v_float64x2& a)
-{
-    __builtin_riscv_fsrm(2);
-    vfloat64m2_t _val = vundefined_f64m2();
-    _val = vset_f64m2(_val, 0, a.val);
-    vfloat32m1_t aval = vfncvt_f_f_v_f32m1(_val, 2);
-
-    vint32m1_t nan = vand_vx_i32m1((vint32m1_t)aval, 0x7f800000, 4);
-    vbool32_t mask = vmsne_vx_i32m1_b32(nan, 0x7f800000, 4);
-    vint32m1_t val = vfcvt_x_f_v_i32m1_m(mask, vmv_v_x_i32m1(0, 4), aval, 4);
-    __builtin_riscv_fsrm(0);
-    return v_int32x4(val);
-}
-
-inline v_int32x4 v_ceil(const v_float64x2& a)
-{
-    __builtin_riscv_fsrm(3);
-    vfloat64m2_t _val = vundefined_f64m2();
-    _val = vset_f64m2(_val, 0, a.val);
-    vfloat32m1_t aval = vfncvt_f_f_v_f32m1(_val, 2);
-
-    vint32m1_t nan = vand_vx_i32m1((vint32m1_t)aval, 0x7f800000, 4);
-    vbool32_t mask = vmsne_vx_i32m1_b32(nan, 0x7f800000, 4);
-    vint32m1_t val = vfcvt_x_f_v_i32m1_m(mask, vmv_v_x_i32m1(0, 4), aval, 4);
-    __builtin_riscv_fsrm(0);
-    return v_int32x4(val);
-}
-
-inline v_int32x4 v_trunc(const v_float64x2& a)
-{
-    __builtin_riscv_fsrm(1);
-    vfloat64m2_t _val = vundefined_f64m2();
-    _val = vset_f64m2(_val, 0, a.val);
-    vfloat32m1_t aval = vfncvt_f_f_v_f32m1(_val, 2);
-
-    vint32m1_t nan = vand_vx_i32m1((vint32m1_t)aval, 0x7f800000, 4);
-    vbool32_t mask = vmsne_vx_i32m1_b32(nan, 0x7f800000, 4);
-    vint32m1_t val = vfcvt_x_f_v_i32m1_m(mask, vmv_v_x_i32m1(0, 4), aval, 4);
-    __builtin_riscv_fsrm(0);
-    return v_int32x4(val);
-}
-
-#define OPENCV_HAL_IMPL_RISCVV_LOAD_DEINTERLEAVED(intrin, _Tpvec, num, _Tp, _T)    \
-inline void v_load_deinterleave(const _Tp* ptr, v_##_Tpvec##x##num& a, v_##_Tpvec##x##num& b) \
-{ \
-    v##_Tpvec##m1x2_t ret = intrin##2e_v_##_T##m1x2(ptr, num);\
-    a.val = vget_##_T##m1x2_##_T##m1(ret, 0);  \
-    b.val = vget_##_T##m1x2_##_T##m1(ret, 1);  \
-} \
-inline void v_load_deinterleave(const _Tp* ptr, v_##_Tpvec##x##num& a, v_##_Tpvec##x##num& b, v_##_Tpvec##x##num& c) \
-{ \
-    v##_Tpvec##m1x3_t ret = intrin##3e_v_##_T##m1x3(ptr, num);\
-    a.val = vget_##_T##m1x3_##_T##m1(ret, 0);  \
-    b.val = vget_##_T##m1x3_##_T##m1(ret, 1);  \
-    c.val = vget_##_T##m1x3_##_T##m1(ret, 2);  \
-}\
-inline void v_load_deinterleave(const _Tp* ptr, v_##_Tpvec##x##num& a, v_##_Tpvec##x##num& b, \
-                                v_##_Tpvec##x##num& c, v_##_Tpvec##x##num& d) \
-{ \
-    v##_Tpvec##m1x4_t ret = intrin##4e_v_##_T##m1x4(ptr, num);\
-    a.val = vget_##_T##m1x4_##_T##m1(ret, 0);  \
-    b.val = vget_##_T##m1x4_##_T##m1(ret, 1);  \
-    c.val = vget_##_T##m1x4_##_T##m1(ret, 2);  \
-    d.val = vget_##_T##m1x4_##_T##m1(ret, 3);  \
-} \
-
-#define OPENCV_HAL_IMPL_RISCVV_STORE_INTERLEAVED(intrin, _Tpvec, num, _Tp, _T)    \
-inline void v_store_interleave( _Tp* ptr, const v_##_Tpvec##x##num& a, const v_##_Tpvec##x##num& b, \
-                                hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ \
-    v##_Tpvec##m1x2_t ret = vundefined_##_T##m1x2();      \
-    ret = vset_##_T##m1x2(ret, 0, a.val);  \
-    ret = vset_##_T##m1x2(ret, 1, b.val);  \
-    intrin##2e_v_##_T##m1x2(ptr, ret, num); \
-} \
-inline void v_store_interleave( _Tp* ptr, const v_##_Tpvec##x##num& a, const v_##_Tpvec##x##num& b, \
-                                const v_##_Tpvec##x##num& c, hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ \
-    v##_Tpvec##m1x3_t ret = vundefined_##_T##m1x3();       \
-    ret = vset_##_T##m1x3(ret, 0, a.val);  \
-    ret = vset_##_T##m1x3(ret, 1, b.val);  \
-    ret = vset_##_T##m1x3(ret, 2, c.val);  \
-    intrin##3e_v_##_T##m1x3(ptr, ret, num); \
-} \
-inline void v_store_interleave( _Tp* ptr, const v_##_Tpvec##x##num& a, const v_##_Tpvec##x##num& b, \
-                                const v_##_Tpvec##x##num& c, const v_##_Tpvec##x##num& d, \
-                                hal::StoreMode /*mode*/=hal::STORE_UNALIGNED ) \
-{ \
-    v##_Tpvec##m1x4_t ret = vundefined_##_T##m1x4();             \
-    ret = vset_##_T##m1x4(ret, 0, a.val);  \
-    ret = vset_##_T##m1x4(ret, 1, b.val);  \
-    ret = vset_##_T##m1x4(ret, 2, c.val);  \
-    ret = vset_##_T##m1x4(ret, 3, d.val);  \
-    intrin##4e_v_##_T##m1x4(ptr, ret, num); \
-}
-
-#define OPENCV_HAL_IMPL_RISCVV_INTERLEAVED(_Tpvec, _Tp, num, ld, st, _T) \
-OPENCV_HAL_IMPL_RISCVV_LOAD_DEINTERLEAVED(ld, _Tpvec, num, _Tp, _T)    \
-OPENCV_HAL_IMPL_RISCVV_STORE_INTERLEAVED(st, _Tpvec, num, _Tp, _T)
-
-//OPENCV_HAL_IMPL_RISCVV_INTERLEAVED(uint8, uchar, )
-OPENCV_HAL_IMPL_RISCVV_INTERLEAVED(int8, schar, 16, vlseg, vsseg, i8)
-OPENCV_HAL_IMPL_RISCVV_INTERLEAVED(int16, short, 8, vlseg, vsseg, i16)
-OPENCV_HAL_IMPL_RISCVV_INTERLEAVED(int32, int, 4, vlseg, vsseg, i32)
-
-OPENCV_HAL_IMPL_RISCVV_INTERLEAVED(uint8, unsigned char, 16, vlseg, vsseg, u8)
-OPENCV_HAL_IMPL_RISCVV_INTERLEAVED(uint16, unsigned short, 8, vlseg, vsseg, u16)
-OPENCV_HAL_IMPL_RISCVV_INTERLEAVED(uint32, unsigned int, 4, vlseg, vsseg, u32)
-
-#define OPENCV_HAL_IMPL_RISCVV_INTERLEAVED_(_Tpvec, _Tp, num, _T) \
-inline void v_load_deinterleave(const _Tp* ptr, v_##_Tpvec##x##num& a, v_##_Tpvec##x##num& b) \
-{ \
-    v##_Tpvec##m1x2_t ret = vlseg2e_v_##_T##m1x2(ptr, num); \
-    a.val = vget_##_T##m1x2_##_T##m1(ret, 0);  \
-    b.val = vget_##_T##m1x2_##_T##m1(ret, 1);  \
-} \
-inline void v_load_deinterleave(const _Tp* ptr, v_##_Tpvec##x##num& a, v_##_Tpvec##x##num& b, v_##_Tpvec##x##num& c) \
-{ \
-    v##_Tpvec##m1x3_t ret = vlseg3e_v_##_T##m1x3(ptr, num);    \
-    a.val = vget_##_T##m1x3_##_T##m1(ret, 0);  \
-    b.val = vget_##_T##m1x3_##_T##m1(ret, 1);  \
-    c.val = vget_##_T##m1x3_##_T##m1(ret, 2);  \
-}\
-inline void v_load_deinterleave(const _Tp* ptr, v_##_Tpvec##x##num& a, v_##_Tpvec##x##num& b, \
-                                v_##_Tpvec##x##num& c, v_##_Tpvec##x##num& d) \
-{ \
-    v##_Tpvec##m1x4_t ret = vlseg4e_v_##_T##m1x4(ptr, num);    \
-    a.val = vget_##_T##m1x4_##_T##m1(ret, 0);  \
-    b.val = vget_##_T##m1x4_##_T##m1(ret, 1);  \
-    c.val = vget_##_T##m1x4_##_T##m1(ret, 2);  \
-    d.val = vget_##_T##m1x4_##_T##m1(ret, 3);  \
-} \
-inline void v_store_interleave( _Tp* ptr, const v_##_Tpvec##x##num& a, const v_##_Tpvec##x##num& b, \
-                                hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ \
-    v##_Tpvec##m1x2_t ret = vundefined_##_T##m1x2();    \
-    ret = vset_##_T##m1x2(ret, 0, a.val);  \
-    ret = vset_##_T##m1x2(ret, 1, b.val);  \
-    vsseg2e_v_##_T##m1x2(ptr, ret, num);    \
-} \
-inline void v_store_interleave( _Tp* ptr, const v_##_Tpvec##x##num& a, const v_##_Tpvec##x##num& b, \
-                                const v_##_Tpvec##x##num& c, hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ \
-    v##_Tpvec##m1x3_t ret = vundefined_##_T##m1x3();    \
-    ret = vset_##_T##m1x3(ret, 0, a.val);  \
-    ret = vset_##_T##m1x3(ret, 1, b.val);  \
-    ret = vset_##_T##m1x3(ret, 2, c.val);  \
-    vsseg3e_v_##_T##m1x3(ptr, ret, num);    \
-} \
-inline void v_store_interleave( _Tp* ptr, const v_##_Tpvec##x##num& a, const v_##_Tpvec##x##num& b, \
-                                const v_##_Tpvec##x##num& c, const v_##_Tpvec##x##num& d, \
-                                hal::StoreMode /*mode*/=hal::STORE_UNALIGNED ) \
-{ \
-    v##_Tpvec##m1x4_t ret = vundefined_##_T##m1x4();    \
-    ret = vset_##_T##m1x4(ret, 0, a.val);  \
-    ret = vset_##_T##m1x4(ret, 1, b.val);  \
-    ret = vset_##_T##m1x4(ret, 2, c.val);  \
-    ret = vset_##_T##m1x4(ret, 3, d.val);  \
-    vsseg4e_v_##_T##m1x4(ptr, ret, num);    \
-}
-OPENCV_HAL_IMPL_RISCVV_INTERLEAVED_(float32, float, 4, f32)
-OPENCV_HAL_IMPL_RISCVV_INTERLEAVED_(float64, double, 2, f64)
-
-OPENCV_HAL_IMPL_RISCVV_INTERLEAVED_(uint64, unsigned long, 2, u64)
-OPENCV_HAL_IMPL_RISCVV_INTERLEAVED_(int64, long, 2, i64)
-
-inline v_float32x4 v_cvt_f32(const v_int32x4& a)
-{
-    return v_float32x4(vfcvt_f_x_v_f32m1(a.val, 4));
-}
-
-#if CV_SIMD128_64F
-inline v_float32x4 v_cvt_f32(const v_float64x2& a)
-{
-    vfloat64m2_t _val = vundefined_f64m2();
-    _val = vset_f64m2(_val, 0, a.val);
-    vfloat32m1_t aval = vfncvt_f_f_v_f32m1(_val, 2);
-    return v_float32x4(aval);
-}
-
-inline v_float32x4 v_cvt_f32(const v_float64x2& a, const v_float64x2& b)
-{
-    vfloat64m2_t _val = vundefined_f64m2();
-    _val = vset_f64m2(_val, 0, a.val);
-    _val = vset_f64m2(_val, 1, b.val);
-    vfloat32m1_t aval = vfncvt_f_f_v_f32m1(_val, 4);
-    return v_float32x4(aval);
-}
-
-inline v_float64x2 v_cvt_f64(const v_int32x4& a)
-{
-    vfloat32m1_t val = vfcvt_f_x_v_f32m1(a.val, 4);
-    vfloat64m2_t _val = vfwcvt_f_f_v_f64m2(val, 4);
-    return v_float64x2(vget_f64m2_f64m1(_val, 0));
-}
-
-inline v_float64x2 v_cvt_f64_high(const v_int32x4& a)
-{
-    vfloat32m1_t val = vfcvt_f_x_v_f32m1(a.val, 4);
-    vfloat64m2_t _val = vfwcvt_f_f_v_f64m2(val, 4);
-    return v_float64x2(vget_f64m2_f64m1(_val, 1));
-}
-
-inline v_float64x2 v_cvt_f64(const v_float32x4& a)
-{
-    vfloat64m2_t _val  = vfwcvt_f_f_v_f64m2(a.val, 4);
-    return v_float64x2(vget_f64m2_f64m1(_val, 0));
-}
-
-inline v_float64x2 v_cvt_f64_high(const v_float32x4& a)
-{
-    vfloat64m2_t _val  = vfwcvt_f_f_v_f64m2(a.val, 4);
-    return v_float64x2(vget_f64m2_f64m1(_val, 1));
-}
-
-inline v_float64x2 v_cvt_f64(const v_int64x2& a)
-{
-    return v_float64x2(vfcvt_f_x_v_f64m1(a.val, 2));
-}
-
-#endif
-inline v_int8x16 v_interleave_pairs(const v_int8x16& vec)
-{
-    vuint64m1_t m0 = {0x0705060403010200, 0x0F0D0E0C0B090A08};
-    return v_int8x16(vrgather_vv_i8m1(vec.val, (vuint8m1_t)m0, 16));
-}
-inline v_uint8x16 v_interleave_pairs(const v_uint8x16& vec)
-{
-    return v_reinterpret_as_u8(v_interleave_pairs(v_reinterpret_as_s8(vec)));
-}
-
-inline v_int8x16 v_interleave_quads(const v_int8x16& vec)
-{
-    vuint64m1_t m0 = {0x0703060205010400, 0x0F0B0E0A0D090C08};
-    return v_int8x16(vrgather_vv_i8m1(vec.val, (vuint8m1_t)m0, 16));
-}
-inline v_uint8x16 v_interleave_quads(const v_uint8x16& vec)
-{
-    return v_reinterpret_as_u8(v_interleave_quads(v_reinterpret_as_s8(vec)));
-}
-
-inline v_int16x8 v_interleave_pairs(const v_int16x8& vec)
-{
-    vuint64m1_t m0 = {0x0706030205040100, 0x0F0E0B0A0D0C0908};
-    return v_int16x8((vint16m1_t)vrgather_vv_u8m1((vuint8m1_t)vec.val, (vuint8m1_t)m0, 16));
-}
-inline v_uint16x8 v_interleave_pairs(const v_uint16x8& vec) { return v_reinterpret_as_u16(v_interleave_pairs(v_reinterpret_as_s16(vec))); }
-inline v_int16x8 v_interleave_quads(const v_int16x8& vec)
-{
-    vuint64m1_t m0 = {0x0B0A030209080100, 0x0F0E07060D0C0504};
-    return v_int16x8((vint16m1_t)vrgather_vv_u8m1((vuint8m1_t)(vec.val), (vuint8m1_t)m0, 16));
-}
-inline v_uint16x8 v_interleave_quads(const v_uint16x8& vec) { return v_reinterpret_as_u16(v_interleave_quads(v_reinterpret_as_s16(vec))); }
-
-inline v_int32x4 v_interleave_pairs(const v_int32x4& vec)
-{
-    vuint64m1_t m0 = {0x0B0A090803020100, 0x0F0E0D0C07060504};
-    return v_int32x4((vint32m1_t)vrgather_vv_u8m1((vuint8m1_t)(vec.val), (vuint8m1_t)m0, 16));
-}
-inline v_uint32x4 v_interleave_pairs(const v_uint32x4& vec) { return v_reinterpret_as_u32(v_interleave_pairs(v_reinterpret_as_s32(vec))); }
-inline v_float32x4 v_interleave_pairs(const v_float32x4& vec) { return v_reinterpret_as_f32(v_interleave_pairs(v_reinterpret_as_s32(vec))); }
-inline v_int8x16 v_pack_triplets(const v_int8x16& vec)
-{
-    vuint64m1_t m0 = {0x0908060504020100, 0xFFFFFFFF0E0D0C0A};
-    return v_int8x16((vint8m1_t)vrgather_vv_u8m1((vuint8m1_t)(vec.val), (vuint8m1_t)m0, 16));
-}
-inline v_uint8x16 v_pack_triplets(const v_uint8x16& vec) { return v_reinterpret_as_u8(v_pack_triplets(v_reinterpret_as_s8(vec))); }
-
-inline v_int16x8 v_pack_triplets(const v_int16x8& vec)
-{
-    vuint64m1_t m0 = {0x0908050403020100, 0xFFFFFFFF0D0C0B0A};
-    return v_int16x8((vint16m1_t)vrgather_vv_u8m1((vuint8m1_t)(vec.val), (vuint8m1_t)m0, 16));
-}
-inline v_uint16x8 v_pack_triplets(const v_uint16x8& vec) { return v_reinterpret_as_u16(v_pack_triplets(v_reinterpret_as_s16(vec))); }
-
-inline v_int32x4 v_pack_triplets(const v_int32x4& vec) { return vec; }
-inline v_uint32x4 v_pack_triplets(const v_uint32x4& vec) { return vec; }
-inline v_float32x4 v_pack_triplets(const v_float32x4& vec) { return vec; }
-
-#if CV_SIMD128_64F
-inline v_float64x2 v_dotprod_expand(const v_int32x4& a, const v_int32x4& b)
-{ return v_cvt_f64(v_dotprod(a, b)); }
-inline v_float64x2 v_dotprod_expand(const v_int32x4& a,   const v_int32x4& b,
-                                    const v_float64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-inline v_float64x2 v_dotprod_expand_fast(const v_int32x4& a, const v_int32x4& b)
-{
-    vint64m2_t v1 = vwmul_vv_i64m2(a.val, b.val, 4);
-    vfloat64m1_t res = vfcvt_f_x_v_f64m1(vadd_vv_i64m1(vget_i64m2_i64m1(v1, 0), vget_i64m2_i64m1(v1, 1), 2), 2);
-    return v_float64x2(res);
-}
-inline v_float64x2 v_dotprod_expand_fast(const v_int32x4& a, const v_int32x4& b, const v_float64x2& c)
-{ v_float64x2 res = v_dotprod_expand_fast(a, b);
-  return res + c; }
-#endif
-////// FP16 support ///////
-inline v_float32x4 v_load_expand(const float16_t* ptr)
-{
-    vfloat16m1_t v = vle_v_f16m1((__fp16*)ptr, 4);
-    vfloat32m2_t v32 = vfwcvt_f_f_v_f32m2(v, 4);
-    return v_float32x4(vget_f32m2_f32m1(v32, 0));
-}
-
-inline void v_pack_store(float16_t* ptr, const v_float32x4& v)
-{
-    vfloat32m2_t v32 = vundefined_f32m2();
-    v32 = vset_f32m2(v32, 0, v.val);
-    vfloat16m1_t hv = vfncvt_f_f_v_f16m1(v32, 4);
-    vse_v_f16m1((__fp16*)ptr, hv, 4);
-}
-
-
-inline void v_cleanup() {}
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END
-
-//! @endcond
-
-}
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_sse.hpp b/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_sse.hpp
deleted file mode 100644
index 443ee160..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_sse.hpp
+++ /dev/null
@@ -1,3467 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Copyright (C) 2015, Itseez Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_HAL_SSE_HPP
-#define OPENCV_HAL_SSE_HPP
-
-#include <algorithm>
-#include "opencv2/core/utility.hpp"
-
-#define CV_SIMD128 1
-#define CV_SIMD128_64F 1
-#define CV_SIMD128_FP16 0  // no native operations with FP16 type.
-
-namespace cv
-{
-
-//! @cond IGNORED
-
-//
-// Compilation troubleshooting:
-// - MSVC: error C2719: 'a': formal parameter with requested alignment of 16 won't be aligned
-//   Replace parameter declaration to const reference:
-//   -v_int32x4 a
-//   +const v_int32x4& a
-//
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
-
-///////// Types ////////////
-
-struct v_uint8x16
-{
-    typedef uchar lane_type;
-    typedef __m128i vector_type;
-    enum { nlanes = 16 };
-
-    /* coverity[uninit_ctor]: suppress warning */
-    v_uint8x16() {}
-    explicit v_uint8x16(__m128i v) : val(v) {}
-    v_uint8x16(uchar v0, uchar v1, uchar v2, uchar v3, uchar v4, uchar v5, uchar v6, uchar v7,
-               uchar v8, uchar v9, uchar v10, uchar v11, uchar v12, uchar v13, uchar v14, uchar v15)
-    {
-        val = _mm_setr_epi8((char)v0, (char)v1, (char)v2, (char)v3,
-                            (char)v4, (char)v5, (char)v6, (char)v7,
-                            (char)v8, (char)v9, (char)v10, (char)v11,
-                            (char)v12, (char)v13, (char)v14, (char)v15);
-    }
-
-    uchar get0() const
-    {
-        return (uchar)_mm_cvtsi128_si32(val);
-    }
-
-    __m128i val;
-};
-
-struct v_int8x16
-{
-    typedef schar lane_type;
-    typedef __m128i vector_type;
-    enum { nlanes = 16 };
-
-    /* coverity[uninit_ctor]: suppress warning */
-    v_int8x16() {}
-    explicit v_int8x16(__m128i v) : val(v) {}
-    v_int8x16(schar v0, schar v1, schar v2, schar v3, schar v4, schar v5, schar v6, schar v7,
-              schar v8, schar v9, schar v10, schar v11, schar v12, schar v13, schar v14, schar v15)
-    {
-        val = _mm_setr_epi8((char)v0, (char)v1, (char)v2, (char)v3,
-                            (char)v4, (char)v5, (char)v6, (char)v7,
-                            (char)v8, (char)v9, (char)v10, (char)v11,
-                            (char)v12, (char)v13, (char)v14, (char)v15);
-    }
-
-    schar get0() const
-    {
-        return (schar)_mm_cvtsi128_si32(val);
-    }
-
-    __m128i val;
-};
-
-struct v_uint16x8
-{
-    typedef ushort lane_type;
-    typedef __m128i vector_type;
-    enum { nlanes = 8 };
-
-    /* coverity[uninit_ctor]: suppress warning */
-    v_uint16x8() {}
-    explicit v_uint16x8(__m128i v) : val(v) {}
-    v_uint16x8(ushort v0, ushort v1, ushort v2, ushort v3, ushort v4, ushort v5, ushort v6, ushort v7)
-    {
-        val = _mm_setr_epi16((short)v0, (short)v1, (short)v2, (short)v3,
-                             (short)v4, (short)v5, (short)v6, (short)v7);
-    }
-
-    ushort get0() const
-    {
-        return (ushort)_mm_cvtsi128_si32(val);
-    }
-
-    __m128i val;
-};
-
-struct v_int16x8
-{
-    typedef short lane_type;
-    typedef __m128i vector_type;
-    enum { nlanes = 8 };
-
-    /* coverity[uninit_ctor]: suppress warning */
-    v_int16x8() {}
-    explicit v_int16x8(__m128i v) : val(v) {}
-    v_int16x8(short v0, short v1, short v2, short v3, short v4, short v5, short v6, short v7)
-    {
-        val = _mm_setr_epi16((short)v0, (short)v1, (short)v2, (short)v3,
-                             (short)v4, (short)v5, (short)v6, (short)v7);
-    }
-
-    short get0() const
-    {
-        return (short)_mm_cvtsi128_si32(val);
-    }
-
-    __m128i val;
-};
-
-struct v_uint32x4
-{
-    typedef unsigned lane_type;
-    typedef __m128i vector_type;
-    enum { nlanes = 4 };
-
-    /* coverity[uninit_ctor]: suppress warning */
-    v_uint32x4() {}
-    explicit v_uint32x4(__m128i v) : val(v) {}
-    v_uint32x4(unsigned v0, unsigned v1, unsigned v2, unsigned v3)
-    {
-        val = _mm_setr_epi32((int)v0, (int)v1, (int)v2, (int)v3);
-    }
-
-    unsigned get0() const
-    {
-        return (unsigned)_mm_cvtsi128_si32(val);
-    }
-
-    __m128i val;
-};
-
-struct v_int32x4
-{
-    typedef int lane_type;
-    typedef __m128i vector_type;
-    enum { nlanes = 4 };
-
-    /* coverity[uninit_ctor]: suppress warning */
-    v_int32x4() {}
-    explicit v_int32x4(__m128i v) : val(v) {}
-    v_int32x4(int v0, int v1, int v2, int v3)
-    {
-        val = _mm_setr_epi32(v0, v1, v2, v3);
-    }
-
-    int get0() const
-    {
-        return _mm_cvtsi128_si32(val);
-    }
-
-    __m128i val;
-};
-
-struct v_float32x4
-{
-    typedef float lane_type;
-    typedef __m128 vector_type;
-    enum { nlanes = 4 };
-
-    /* coverity[uninit_ctor]: suppress warning */
-    v_float32x4() {}
-    explicit v_float32x4(__m128 v) : val(v) {}
-    v_float32x4(float v0, float v1, float v2, float v3)
-    {
-        val = _mm_setr_ps(v0, v1, v2, v3);
-    }
-
-    float get0() const
-    {
-        return _mm_cvtss_f32(val);
-    }
-
-    __m128 val;
-};
-
-struct v_uint64x2
-{
-    typedef uint64 lane_type;
-    typedef __m128i vector_type;
-    enum { nlanes = 2 };
-
-    /* coverity[uninit_ctor]: suppress warning */
-    v_uint64x2() {}
-    explicit v_uint64x2(__m128i v) : val(v) {}
-    v_uint64x2(uint64 v0, uint64 v1)
-    {
-#if defined(_MSC_VER) && _MSC_VER >= 1920/*MSVS 2019*/ && defined(_M_X64) && !defined(__clang__)
-        val = _mm_setr_epi64x((int64_t)v0, (int64_t)v1);
-#elif defined(__GNUC__)
-        val = _mm_setr_epi64((__m64)v0, (__m64)v1);
-#else
-        val = _mm_setr_epi32((int)v0, (int)(v0 >> 32), (int)v1, (int)(v1 >> 32));
-#endif
-    }
-
-    uint64 get0() const
-    {
-    #if !defined(__x86_64__) && !defined(_M_X64)
-        int a = _mm_cvtsi128_si32(val);
-        int b = _mm_cvtsi128_si32(_mm_srli_epi64(val, 32));
-        return (unsigned)a | ((uint64)(unsigned)b << 32);
-    #else
-        return (uint64)_mm_cvtsi128_si64(val);
-    #endif
-    }
-
-    __m128i val;
-};
-
-struct v_int64x2
-{
-    typedef int64 lane_type;
-    typedef __m128i vector_type;
-    enum { nlanes = 2 };
-
-    /* coverity[uninit_ctor]: suppress warning */
-    v_int64x2() {}
-    explicit v_int64x2(__m128i v) : val(v) {}
-    v_int64x2(int64 v0, int64 v1)
-    {
-#if defined(_MSC_VER) && _MSC_VER >= 1920/*MSVS 2019*/ && defined(_M_X64) && !defined(__clang__)
-        val = _mm_setr_epi64x((int64_t)v0, (int64_t)v1);
-#elif defined(__GNUC__)
-        val = _mm_setr_epi64((__m64)v0, (__m64)v1);
-#else
-        val = _mm_setr_epi32((int)v0, (int)(v0 >> 32), (int)v1, (int)(v1 >> 32));
-#endif
-    }
-
-    int64 get0() const
-    {
-    #if !defined(__x86_64__) && !defined(_M_X64)
-        int a = _mm_cvtsi128_si32(val);
-        int b = _mm_cvtsi128_si32(_mm_srli_epi64(val, 32));
-        return (int64)((unsigned)a | ((uint64)(unsigned)b << 32));
-    #else
-        return _mm_cvtsi128_si64(val);
-    #endif
-    }
-
-    __m128i val;
-};
-
-struct v_float64x2
-{
-    typedef double lane_type;
-    typedef __m128d vector_type;
-    enum { nlanes = 2 };
-
-    /* coverity[uninit_ctor]: suppress warning */
-    v_float64x2() {}
-    explicit v_float64x2(__m128d v) : val(v) {}
-    v_float64x2(double v0, double v1)
-    {
-        val = _mm_setr_pd(v0, v1);
-    }
-
-    double get0() const
-    {
-        return _mm_cvtsd_f64(val);
-    }
-
-    __m128d val;
-};
-
-namespace hal_sse_internal
-{
-    template <typename to_sse_type, typename from_sse_type>
-    to_sse_type v_sse_reinterpret_as(const from_sse_type& val);
-
-#define OPENCV_HAL_IMPL_SSE_REINTERPRET_RAW(to_sse_type, from_sse_type, sse_cast_intrin) \
-    template<> inline \
-    to_sse_type v_sse_reinterpret_as(const from_sse_type& a) \
-    { return sse_cast_intrin(a); }
-
-    OPENCV_HAL_IMPL_SSE_REINTERPRET_RAW(__m128i, __m128i, OPENCV_HAL_NOP)
-    OPENCV_HAL_IMPL_SSE_REINTERPRET_RAW(__m128i, __m128, _mm_castps_si128)
-    OPENCV_HAL_IMPL_SSE_REINTERPRET_RAW(__m128i, __m128d, _mm_castpd_si128)
-    OPENCV_HAL_IMPL_SSE_REINTERPRET_RAW(__m128, __m128i, _mm_castsi128_ps)
-    OPENCV_HAL_IMPL_SSE_REINTERPRET_RAW(__m128, __m128, OPENCV_HAL_NOP)
-    OPENCV_HAL_IMPL_SSE_REINTERPRET_RAW(__m128, __m128d, _mm_castpd_ps)
-    OPENCV_HAL_IMPL_SSE_REINTERPRET_RAW(__m128d, __m128i, _mm_castsi128_pd)
-    OPENCV_HAL_IMPL_SSE_REINTERPRET_RAW(__m128d, __m128, _mm_castps_pd)
-    OPENCV_HAL_IMPL_SSE_REINTERPRET_RAW(__m128d, __m128d, OPENCV_HAL_NOP)
-}
-
-#define OPENCV_HAL_IMPL_SSE_INITVEC(_Tpvec, _Tp, suffix, zsuffix, ssuffix, _Tps, cast) \
-inline _Tpvec v_setzero_##suffix() { return _Tpvec(_mm_setzero_##zsuffix()); } \
-inline _Tpvec v_setall_##suffix(_Tp v) { return _Tpvec(_mm_set1_##ssuffix((_Tps)v)); } \
-template<typename _Tpvec0> inline _Tpvec v_reinterpret_as_##suffix(const _Tpvec0& a) \
-{ return _Tpvec(cast(a.val)); }
-
-OPENCV_HAL_IMPL_SSE_INITVEC(v_uint8x16, uchar, u8, si128, epi8, schar, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_SSE_INITVEC(v_int8x16, schar, s8, si128, epi8, schar, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_SSE_INITVEC(v_uint16x8, ushort, u16, si128, epi16, short, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_SSE_INITVEC(v_int16x8, short, s16, si128, epi16, short, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_SSE_INITVEC(v_uint32x4, unsigned, u32, si128, epi32, int, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_SSE_INITVEC(v_int32x4, int, s32, si128, epi32, int, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_SSE_INITVEC(v_float32x4, float, f32, ps, ps, float, _mm_castsi128_ps)
-OPENCV_HAL_IMPL_SSE_INITVEC(v_float64x2, double, f64, pd, pd, double, _mm_castsi128_pd)
-
-inline v_uint64x2 v_setzero_u64() { return v_uint64x2(_mm_setzero_si128()); }
-inline v_int64x2 v_setzero_s64() { return v_int64x2(_mm_setzero_si128()); }
-inline v_uint64x2 v_setall_u64(uint64 val) { return v_uint64x2(val, val); }
-inline v_int64x2 v_setall_s64(int64 val) { return v_int64x2(val, val); }
-
-template<typename _Tpvec> inline
-v_uint64x2 v_reinterpret_as_u64(const _Tpvec& a) { return v_uint64x2(a.val); }
-template<typename _Tpvec> inline
-v_int64x2 v_reinterpret_as_s64(const _Tpvec& a) { return v_int64x2(a.val); }
-inline v_float32x4 v_reinterpret_as_f32(const v_uint64x2& a)
-{ return v_float32x4(_mm_castsi128_ps(a.val)); }
-inline v_float32x4 v_reinterpret_as_f32(const v_int64x2& a)
-{ return v_float32x4(_mm_castsi128_ps(a.val)); }
-inline v_float64x2 v_reinterpret_as_f64(const v_uint64x2& a)
-{ return v_float64x2(_mm_castsi128_pd(a.val)); }
-inline v_float64x2 v_reinterpret_as_f64(const v_int64x2& a)
-{ return v_float64x2(_mm_castsi128_pd(a.val)); }
-
-#define OPENCV_HAL_IMPL_SSE_INIT_FROM_FLT(_Tpvec, suffix) \
-inline _Tpvec v_reinterpret_as_##suffix(const v_float32x4& a) \
-{ return _Tpvec(_mm_castps_si128(a.val)); } \
-inline _Tpvec v_reinterpret_as_##suffix(const v_float64x2& a) \
-{ return _Tpvec(_mm_castpd_si128(a.val)); }
-
-OPENCV_HAL_IMPL_SSE_INIT_FROM_FLT(v_uint8x16, u8)
-OPENCV_HAL_IMPL_SSE_INIT_FROM_FLT(v_int8x16, s8)
-OPENCV_HAL_IMPL_SSE_INIT_FROM_FLT(v_uint16x8, u16)
-OPENCV_HAL_IMPL_SSE_INIT_FROM_FLT(v_int16x8, s16)
-OPENCV_HAL_IMPL_SSE_INIT_FROM_FLT(v_uint32x4, u32)
-OPENCV_HAL_IMPL_SSE_INIT_FROM_FLT(v_int32x4, s32)
-OPENCV_HAL_IMPL_SSE_INIT_FROM_FLT(v_uint64x2, u64)
-OPENCV_HAL_IMPL_SSE_INIT_FROM_FLT(v_int64x2, s64)
-
-inline v_float32x4 v_reinterpret_as_f32(const v_float32x4& a) {return a; }
-inline v_float64x2 v_reinterpret_as_f64(const v_float64x2& a) {return a; }
-inline v_float32x4 v_reinterpret_as_f32(const v_float64x2& a) {return v_float32x4(_mm_castpd_ps(a.val)); }
-inline v_float64x2 v_reinterpret_as_f64(const v_float32x4& a) {return v_float64x2(_mm_castps_pd(a.val)); }
-
-//////////////// PACK ///////////////
-inline v_uint8x16 v_pack(const v_uint16x8& a, const v_uint16x8& b)
-{
-    __m128i delta = _mm_set1_epi16(255);
-    return v_uint8x16(_mm_packus_epi16(_mm_subs_epu16(a.val, _mm_subs_epu16(a.val, delta)),
-                                       _mm_subs_epu16(b.val, _mm_subs_epu16(b.val, delta))));
-}
-
-inline void v_pack_store(uchar* ptr, const v_uint16x8& a)
-{
-    __m128i delta = _mm_set1_epi16(255);
-    __m128i a1 = _mm_subs_epu16(a.val, _mm_subs_epu16(a.val, delta));
-    _mm_storel_epi64((__m128i*)ptr, _mm_packus_epi16(a1, a1));
-}
-
-inline v_uint8x16 v_pack_u(const v_int16x8& a, const v_int16x8& b)
-{ return v_uint8x16(_mm_packus_epi16(a.val, b.val)); }
-
-inline void v_pack_u_store(uchar* ptr, const v_int16x8& a)
-{ _mm_storel_epi64((__m128i*)ptr, _mm_packus_epi16(a.val, a.val)); }
-
-template<int n> inline
-v_uint8x16 v_rshr_pack(const v_uint16x8& a, const v_uint16x8& b)
-{
-    // we assume that n > 0, and so the shifted 16-bit values can be treated as signed numbers.
-    __m128i delta = _mm_set1_epi16((short)(1 << (n-1)));
-    return v_uint8x16(_mm_packus_epi16(_mm_srli_epi16(_mm_adds_epu16(a.val, delta), n),
-                                       _mm_srli_epi16(_mm_adds_epu16(b.val, delta), n)));
-}
-
-template<int n> inline
-void v_rshr_pack_store(uchar* ptr, const v_uint16x8& a)
-{
-    __m128i delta = _mm_set1_epi16((short)(1 << (n-1)));
-    __m128i a1 = _mm_srli_epi16(_mm_adds_epu16(a.val, delta), n);
-    _mm_storel_epi64((__m128i*)ptr, _mm_packus_epi16(a1, a1));
-}
-
-template<int n> inline
-v_uint8x16 v_rshr_pack_u(const v_int16x8& a, const v_int16x8& b)
-{
-    __m128i delta = _mm_set1_epi16((short)(1 << (n-1)));
-    return v_uint8x16(_mm_packus_epi16(_mm_srai_epi16(_mm_adds_epi16(a.val, delta), n),
-                                       _mm_srai_epi16(_mm_adds_epi16(b.val, delta), n)));
-}
-
-template<int n> inline
-void v_rshr_pack_u_store(uchar* ptr, const v_int16x8& a)
-{
-    __m128i delta = _mm_set1_epi16((short)(1 << (n-1)));
-    __m128i a1 = _mm_srai_epi16(_mm_adds_epi16(a.val, delta), n);
-    _mm_storel_epi64((__m128i*)ptr, _mm_packus_epi16(a1, a1));
-}
-
-inline v_int8x16 v_pack(const v_int16x8& a, const v_int16x8& b)
-{ return v_int8x16(_mm_packs_epi16(a.val, b.val)); }
-
-inline void v_pack_store(schar* ptr, const v_int16x8& a)
-{ _mm_storel_epi64((__m128i*)ptr, _mm_packs_epi16(a.val, a.val)); }
-
-template<int n> inline
-v_int8x16 v_rshr_pack(const v_int16x8& a, const v_int16x8& b)
-{
-    // we assume that n > 0, and so the shifted 16-bit values can be treated as signed numbers.
-    __m128i delta = _mm_set1_epi16((short)(1 << (n-1)));
-    return v_int8x16(_mm_packs_epi16(_mm_srai_epi16(_mm_adds_epi16(a.val, delta), n),
-                                     _mm_srai_epi16(_mm_adds_epi16(b.val, delta), n)));
-}
-template<int n> inline
-void v_rshr_pack_store(schar* ptr, const v_int16x8& a)
-{
-    // we assume that n > 0, and so the shifted 16-bit values can be treated as signed numbers.
-    __m128i delta = _mm_set1_epi16((short)(1 << (n-1)));
-    __m128i a1 = _mm_srai_epi16(_mm_adds_epi16(a.val, delta), n);
-    _mm_storel_epi64((__m128i*)ptr, _mm_packs_epi16(a1, a1));
-}
-
-
-// byte-wise "mask ? a : b"
-inline __m128i v_select_si128(__m128i mask, __m128i a, __m128i b)
-{
-#if CV_SSE4_1
-    return _mm_blendv_epi8(b, a, mask);
-#else
-    return _mm_xor_si128(b, _mm_and_si128(_mm_xor_si128(a, b), mask));
-#endif
-}
-
-inline v_uint16x8 v_pack(const v_uint32x4& a, const v_uint32x4& b)
-{ return v_uint16x8(_v128_packs_epu32(a.val, b.val)); }
-
-inline void v_pack_store(ushort* ptr, const v_uint32x4& a)
-{
-    __m128i z = _mm_setzero_si128(), maxval32 = _mm_set1_epi32(65535), delta32 = _mm_set1_epi32(32768);
-    __m128i a1 = _mm_sub_epi32(v_select_si128(_mm_cmpgt_epi32(z, a.val), maxval32, a.val), delta32);
-    __m128i r = _mm_packs_epi32(a1, a1);
-    _mm_storel_epi64((__m128i*)ptr, _mm_sub_epi16(r, _mm_set1_epi16(-32768)));
-}
-
-template<int n> inline
-v_uint16x8 v_rshr_pack(const v_uint32x4& a, const v_uint32x4& b)
-{
-    __m128i delta = _mm_set1_epi32(1 << (n-1)), delta32 = _mm_set1_epi32(32768);
-    __m128i a1 = _mm_sub_epi32(_mm_srli_epi32(_mm_add_epi32(a.val, delta), n), delta32);
-    __m128i b1 = _mm_sub_epi32(_mm_srli_epi32(_mm_add_epi32(b.val, delta), n), delta32);
-    return v_uint16x8(_mm_sub_epi16(_mm_packs_epi32(a1, b1), _mm_set1_epi16(-32768)));
-}
-
-template<int n> inline
-void v_rshr_pack_store(ushort* ptr, const v_uint32x4& a)
-{
-    __m128i delta = _mm_set1_epi32(1 << (n-1)), delta32 = _mm_set1_epi32(32768);
-    __m128i a1 = _mm_sub_epi32(_mm_srli_epi32(_mm_add_epi32(a.val, delta), n), delta32);
-    __m128i a2 = _mm_sub_epi16(_mm_packs_epi32(a1, a1), _mm_set1_epi16(-32768));
-    _mm_storel_epi64((__m128i*)ptr, a2);
-}
-
-inline v_uint16x8 v_pack_u(const v_int32x4& a, const v_int32x4& b)
-{
-#if CV_SSE4_1
-    return v_uint16x8(_mm_packus_epi32(a.val, b.val));
-#else
-    __m128i delta32 = _mm_set1_epi32(32768);
-
-    // preliminary saturate negative values to zero
-    __m128i a1 = _mm_and_si128(a.val, _mm_cmpgt_epi32(a.val, _mm_set1_epi32(0)));
-    __m128i b1 = _mm_and_si128(b.val, _mm_cmpgt_epi32(b.val, _mm_set1_epi32(0)));
-
-    __m128i r = _mm_packs_epi32(_mm_sub_epi32(a1, delta32), _mm_sub_epi32(b1, delta32));
-    return v_uint16x8(_mm_sub_epi16(r, _mm_set1_epi16(-32768)));
-#endif
-}
-
-inline void v_pack_u_store(ushort* ptr, const v_int32x4& a)
-{
-#if CV_SSE4_1
-    _mm_storel_epi64((__m128i*)ptr, _mm_packus_epi32(a.val, a.val));
-#else
-    __m128i delta32 = _mm_set1_epi32(32768);
-    __m128i a1 = _mm_sub_epi32(a.val, delta32);
-    __m128i r = _mm_sub_epi16(_mm_packs_epi32(a1, a1), _mm_set1_epi16(-32768));
-    _mm_storel_epi64((__m128i*)ptr, r);
-#endif
-}
-
-template<int n> inline
-v_uint16x8 v_rshr_pack_u(const v_int32x4& a, const v_int32x4& b)
-{
-#if CV_SSE4_1
-    __m128i delta = _mm_set1_epi32(1 << (n - 1));
-    return v_uint16x8(_mm_packus_epi32(_mm_srai_epi32(_mm_add_epi32(a.val, delta), n),
-                                       _mm_srai_epi32(_mm_add_epi32(b.val, delta), n)));
-#else
-    __m128i delta = _mm_set1_epi32(1 << (n-1)), delta32 = _mm_set1_epi32(32768);
-    __m128i a1 = _mm_sub_epi32(_mm_srai_epi32(_mm_add_epi32(a.val, delta), n), delta32);
-    __m128i a2 = _mm_sub_epi16(_mm_packs_epi32(a1, a1), _mm_set1_epi16(-32768));
-    __m128i b1 = _mm_sub_epi32(_mm_srai_epi32(_mm_add_epi32(b.val, delta), n), delta32);
-    __m128i b2 = _mm_sub_epi16(_mm_packs_epi32(b1, b1), _mm_set1_epi16(-32768));
-    return v_uint16x8(_mm_unpacklo_epi64(a2, b2));
-#endif
-}
-
-template<int n> inline
-void v_rshr_pack_u_store(ushort* ptr, const v_int32x4& a)
-{
-#if CV_SSE4_1
-    __m128i delta = _mm_set1_epi32(1 << (n - 1));
-    __m128i a1 = _mm_srai_epi32(_mm_add_epi32(a.val, delta), n);
-    _mm_storel_epi64((__m128i*)ptr, _mm_packus_epi32(a1, a1));
-#else
-    __m128i delta = _mm_set1_epi32(1 << (n-1)), delta32 = _mm_set1_epi32(32768);
-    __m128i a1 = _mm_sub_epi32(_mm_srai_epi32(_mm_add_epi32(a.val, delta), n), delta32);
-    __m128i a2 = _mm_sub_epi16(_mm_packs_epi32(a1, a1), _mm_set1_epi16(-32768));
-    _mm_storel_epi64((__m128i*)ptr, a2);
-#endif
-}
-
-inline v_int16x8 v_pack(const v_int32x4& a, const v_int32x4& b)
-{ return v_int16x8(_mm_packs_epi32(a.val, b.val)); }
-
-inline void v_pack_store(short* ptr, const v_int32x4& a)
-{
-    _mm_storel_epi64((__m128i*)ptr, _mm_packs_epi32(a.val, a.val));
-}
-
-template<int n> inline
-v_int16x8 v_rshr_pack(const v_int32x4& a, const v_int32x4& b)
-{
-    __m128i delta = _mm_set1_epi32(1 << (n-1));
-    return v_int16x8(_mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(a.val, delta), n),
-                                     _mm_srai_epi32(_mm_add_epi32(b.val, delta), n)));
-}
-
-template<int n> inline
-void v_rshr_pack_store(short* ptr, const v_int32x4& a)
-{
-    __m128i delta = _mm_set1_epi32(1 << (n-1));
-    __m128i a1 = _mm_srai_epi32(_mm_add_epi32(a.val, delta), n);
-    _mm_storel_epi64((__m128i*)ptr, _mm_packs_epi32(a1, a1));
-}
-
-
-// [a0 0 | b0 0]  [a1 0 | b1 0]
-inline v_uint32x4 v_pack(const v_uint64x2& a, const v_uint64x2& b)
-{
-    __m128i v0 = _mm_unpacklo_epi32(a.val, b.val); // a0 a1 0 0
-    __m128i v1 = _mm_unpackhi_epi32(a.val, b.val); // b0 b1 0 0
-    return v_uint32x4(_mm_unpacklo_epi32(v0, v1));
-}
-
-inline void v_pack_store(unsigned* ptr, const v_uint64x2& a)
-{
-    __m128i a1 = _mm_shuffle_epi32(a.val, _MM_SHUFFLE(0, 2, 2, 0));
-    _mm_storel_epi64((__m128i*)ptr, a1);
-}
-
-// [a0 0 | b0 0]  [a1 0 | b1 0]
-inline v_int32x4 v_pack(const v_int64x2& a, const v_int64x2& b)
-{
-    __m128i v0 = _mm_unpacklo_epi32(a.val, b.val); // a0 a1 0 0
-    __m128i v1 = _mm_unpackhi_epi32(a.val, b.val); // b0 b1 0 0
-    return v_int32x4(_mm_unpacklo_epi32(v0, v1));
-}
-
-inline void v_pack_store(int* ptr, const v_int64x2& a)
-{
-    __m128i a1 = _mm_shuffle_epi32(a.val, _MM_SHUFFLE(0, 2, 2, 0));
-    _mm_storel_epi64((__m128i*)ptr, a1);
-}
-
-template<int n> inline
-v_uint32x4 v_rshr_pack(const v_uint64x2& a, const v_uint64x2& b)
-{
-    uint64 delta = (uint64)1 << (n-1);
-    v_uint64x2 delta2(delta, delta);
-    __m128i a1 = _mm_srli_epi64(_mm_add_epi64(a.val, delta2.val), n);
-    __m128i b1 = _mm_srli_epi64(_mm_add_epi64(b.val, delta2.val), n);
-    __m128i v0 = _mm_unpacklo_epi32(a1, b1); // a0 a1 0 0
-    __m128i v1 = _mm_unpackhi_epi32(a1, b1); // b0 b1 0 0
-    return v_uint32x4(_mm_unpacklo_epi32(v0, v1));
-}
-
-template<int n> inline
-void v_rshr_pack_store(unsigned* ptr, const v_uint64x2& a)
-{
-    uint64 delta = (uint64)1 << (n-1);
-    v_uint64x2 delta2(delta, delta);
-    __m128i a1 = _mm_srli_epi64(_mm_add_epi64(a.val, delta2.val), n);
-    __m128i a2 = _mm_shuffle_epi32(a1, _MM_SHUFFLE(0, 2, 2, 0));
-    _mm_storel_epi64((__m128i*)ptr, a2);
-}
-
-inline __m128i v_sign_epi64(__m128i a)
-{
-    return _mm_shuffle_epi32(_mm_srai_epi32(a, 31), _MM_SHUFFLE(3, 3, 1, 1)); // x m0 | x m1
-}
-
-inline __m128i v_srai_epi64(__m128i a, int imm)
-{
-    __m128i smask = v_sign_epi64(a);
-    return _mm_xor_si128(_mm_srli_epi64(_mm_xor_si128(a, smask), imm), smask);
-}
-
-template<int n> inline
-v_int32x4 v_rshr_pack(const v_int64x2& a, const v_int64x2& b)
-{
-    int64 delta = (int64)1 << (n-1);
-    v_int64x2 delta2(delta, delta);
-    __m128i a1 = v_srai_epi64(_mm_add_epi64(a.val, delta2.val), n);
-    __m128i b1 = v_srai_epi64(_mm_add_epi64(b.val, delta2.val), n);
-    __m128i v0 = _mm_unpacklo_epi32(a1, b1); // a0 a1 0 0
-    __m128i v1 = _mm_unpackhi_epi32(a1, b1); // b0 b1 0 0
-    return v_int32x4(_mm_unpacklo_epi32(v0, v1));
-}
-
-template<int n> inline
-void v_rshr_pack_store(int* ptr, const v_int64x2& a)
-{
-    int64 delta = (int64)1 << (n-1);
-    v_int64x2 delta2(delta, delta);
-    __m128i a1 = v_srai_epi64(_mm_add_epi64(a.val, delta2.val), n);
-    __m128i a2 = _mm_shuffle_epi32(a1, _MM_SHUFFLE(0, 2, 2, 0));
-    _mm_storel_epi64((__m128i*)ptr, a2);
-}
-
-// pack boolean
-inline v_uint8x16 v_pack_b(const v_uint16x8& a, const v_uint16x8& b)
-{
-    __m128i ab = _mm_packs_epi16(a.val, b.val);
-    return v_uint8x16(ab);
-}
-
-inline v_uint8x16 v_pack_b(const v_uint32x4& a, const v_uint32x4& b,
-                           const v_uint32x4& c, const v_uint32x4& d)
-{
-    __m128i ab = _mm_packs_epi32(a.val, b.val);
-    __m128i cd = _mm_packs_epi32(c.val, d.val);
-    return v_uint8x16(_mm_packs_epi16(ab, cd));
-}
-
-inline v_uint8x16 v_pack_b(const v_uint64x2& a, const v_uint64x2& b, const v_uint64x2& c,
-                           const v_uint64x2& d, const v_uint64x2& e, const v_uint64x2& f,
-                           const v_uint64x2& g, const v_uint64x2& h)
-{
-    __m128i ab = _mm_packs_epi32(a.val, b.val);
-    __m128i cd = _mm_packs_epi32(c.val, d.val);
-    __m128i ef = _mm_packs_epi32(e.val, f.val);
-    __m128i gh = _mm_packs_epi32(g.val, h.val);
-
-    __m128i abcd = _mm_packs_epi32(ab, cd);
-    __m128i efgh = _mm_packs_epi32(ef, gh);
-    return v_uint8x16(_mm_packs_epi16(abcd, efgh));
-}
-
-inline v_float32x4 v_matmul(const v_float32x4& v, const v_float32x4& m0,
-                            const v_float32x4& m1, const v_float32x4& m2,
-                            const v_float32x4& m3)
-{
-    __m128 v0 = _mm_mul_ps(_mm_shuffle_ps(v.val, v.val, _MM_SHUFFLE(0, 0, 0, 0)), m0.val);
-    __m128 v1 = _mm_mul_ps(_mm_shuffle_ps(v.val, v.val, _MM_SHUFFLE(1, 1, 1, 1)), m1.val);
-    __m128 v2 = _mm_mul_ps(_mm_shuffle_ps(v.val, v.val, _MM_SHUFFLE(2, 2, 2, 2)), m2.val);
-    __m128 v3 = _mm_mul_ps(_mm_shuffle_ps(v.val, v.val, _MM_SHUFFLE(3, 3, 3, 3)), m3.val);
-
-    return v_float32x4(_mm_add_ps(_mm_add_ps(v0, v1), _mm_add_ps(v2, v3)));
-}
-
-inline v_float32x4 v_matmuladd(const v_float32x4& v, const v_float32x4& m0,
-                               const v_float32x4& m1, const v_float32x4& m2,
-                               const v_float32x4& a)
-{
-    __m128 v0 = _mm_mul_ps(_mm_shuffle_ps(v.val, v.val, _MM_SHUFFLE(0, 0, 0, 0)), m0.val);
-    __m128 v1 = _mm_mul_ps(_mm_shuffle_ps(v.val, v.val, _MM_SHUFFLE(1, 1, 1, 1)), m1.val);
-    __m128 v2 = _mm_mul_ps(_mm_shuffle_ps(v.val, v.val, _MM_SHUFFLE(2, 2, 2, 2)), m2.val);
-
-    return v_float32x4(_mm_add_ps(_mm_add_ps(v0, v1), _mm_add_ps(v2, a.val)));
-}
-
-#define OPENCV_HAL_IMPL_SSE_BIN_OP(bin_op, _Tpvec, intrin) \
-    inline _Tpvec operator bin_op (const _Tpvec& a, const _Tpvec& b) \
-    { \
-        return _Tpvec(intrin(a.val, b.val)); \
-    } \
-    inline _Tpvec& operator bin_op##= (_Tpvec& a, const _Tpvec& b) \
-    { \
-        a.val = intrin(a.val, b.val); \
-        return a; \
-    }
-
-OPENCV_HAL_IMPL_SSE_BIN_OP(+, v_uint8x16, _mm_adds_epu8)
-OPENCV_HAL_IMPL_SSE_BIN_OP(-, v_uint8x16, _mm_subs_epu8)
-OPENCV_HAL_IMPL_SSE_BIN_OP(+, v_int8x16, _mm_adds_epi8)
-OPENCV_HAL_IMPL_SSE_BIN_OP(-, v_int8x16, _mm_subs_epi8)
-OPENCV_HAL_IMPL_SSE_BIN_OP(+, v_uint16x8, _mm_adds_epu16)
-OPENCV_HAL_IMPL_SSE_BIN_OP(-, v_uint16x8, _mm_subs_epu16)
-OPENCV_HAL_IMPL_SSE_BIN_OP(+, v_int16x8, _mm_adds_epi16)
-OPENCV_HAL_IMPL_SSE_BIN_OP(-, v_int16x8, _mm_subs_epi16)
-OPENCV_HAL_IMPL_SSE_BIN_OP(+, v_uint32x4, _mm_add_epi32)
-OPENCV_HAL_IMPL_SSE_BIN_OP(-, v_uint32x4, _mm_sub_epi32)
-OPENCV_HAL_IMPL_SSE_BIN_OP(*, v_uint32x4, _v128_mullo_epi32)
-OPENCV_HAL_IMPL_SSE_BIN_OP(+, v_int32x4, _mm_add_epi32)
-OPENCV_HAL_IMPL_SSE_BIN_OP(-, v_int32x4, _mm_sub_epi32)
-OPENCV_HAL_IMPL_SSE_BIN_OP(*, v_int32x4, _v128_mullo_epi32)
-OPENCV_HAL_IMPL_SSE_BIN_OP(+, v_float32x4, _mm_add_ps)
-OPENCV_HAL_IMPL_SSE_BIN_OP(-, v_float32x4, _mm_sub_ps)
-OPENCV_HAL_IMPL_SSE_BIN_OP(*, v_float32x4, _mm_mul_ps)
-OPENCV_HAL_IMPL_SSE_BIN_OP(/, v_float32x4, _mm_div_ps)
-OPENCV_HAL_IMPL_SSE_BIN_OP(+, v_float64x2, _mm_add_pd)
-OPENCV_HAL_IMPL_SSE_BIN_OP(-, v_float64x2, _mm_sub_pd)
-OPENCV_HAL_IMPL_SSE_BIN_OP(*, v_float64x2, _mm_mul_pd)
-OPENCV_HAL_IMPL_SSE_BIN_OP(/, v_float64x2, _mm_div_pd)
-OPENCV_HAL_IMPL_SSE_BIN_OP(+, v_uint64x2, _mm_add_epi64)
-OPENCV_HAL_IMPL_SSE_BIN_OP(-, v_uint64x2, _mm_sub_epi64)
-OPENCV_HAL_IMPL_SSE_BIN_OP(+, v_int64x2, _mm_add_epi64)
-OPENCV_HAL_IMPL_SSE_BIN_OP(-, v_int64x2, _mm_sub_epi64)
-
-// saturating multiply 8-bit, 16-bit
-#define OPENCV_HAL_IMPL_SSE_MUL_SAT(_Tpvec, _Tpwvec)             \
-    inline _Tpvec operator * (const _Tpvec& a, const _Tpvec& b)  \
-    {                                                            \
-        _Tpwvec c, d;                                            \
-        v_mul_expand(a, b, c, d);                                \
-        return v_pack(c, d);                                     \
-    }                                                            \
-    inline _Tpvec& operator *= (_Tpvec& a, const _Tpvec& b)      \
-    { a = a * b; return a; }
-
-OPENCV_HAL_IMPL_SSE_MUL_SAT(v_uint8x16, v_uint16x8)
-OPENCV_HAL_IMPL_SSE_MUL_SAT(v_int8x16,  v_int16x8)
-OPENCV_HAL_IMPL_SSE_MUL_SAT(v_uint16x8, v_uint32x4)
-OPENCV_HAL_IMPL_SSE_MUL_SAT(v_int16x8,  v_int32x4)
-
-//  Multiply and expand
-inline void v_mul_expand(const v_uint8x16& a, const v_uint8x16& b,
-                         v_uint16x8& c, v_uint16x8& d)
-{
-    v_uint16x8 a0, a1, b0, b1;
-    v_expand(a, a0, a1);
-    v_expand(b, b0, b1);
-    c = v_mul_wrap(a0, b0);
-    d = v_mul_wrap(a1, b1);
-}
-
-inline void v_mul_expand(const v_int8x16& a, const v_int8x16& b,
-                         v_int16x8& c, v_int16x8& d)
-{
-    v_int16x8 a0, a1, b0, b1;
-    v_expand(a, a0, a1);
-    v_expand(b, b0, b1);
-    c = v_mul_wrap(a0, b0);
-    d = v_mul_wrap(a1, b1);
-}
-
-inline void v_mul_expand(const v_int16x8& a, const v_int16x8& b,
-                         v_int32x4& c, v_int32x4& d)
-{
-    __m128i v0 = _mm_mullo_epi16(a.val, b.val);
-    __m128i v1 = _mm_mulhi_epi16(a.val, b.val);
-    c.val = _mm_unpacklo_epi16(v0, v1);
-    d.val = _mm_unpackhi_epi16(v0, v1);
-}
-
-inline void v_mul_expand(const v_uint16x8& a, const v_uint16x8& b,
-                         v_uint32x4& c, v_uint32x4& d)
-{
-    __m128i v0 = _mm_mullo_epi16(a.val, b.val);
-    __m128i v1 = _mm_mulhi_epu16(a.val, b.val);
-    c.val = _mm_unpacklo_epi16(v0, v1);
-    d.val = _mm_unpackhi_epi16(v0, v1);
-}
-
-inline void v_mul_expand(const v_uint32x4& a, const v_uint32x4& b,
-                         v_uint64x2& c, v_uint64x2& d)
-{
-    __m128i c0 = _mm_mul_epu32(a.val, b.val);
-    __m128i c1 = _mm_mul_epu32(_mm_srli_epi64(a.val, 32), _mm_srli_epi64(b.val, 32));
-    c.val = _mm_unpacklo_epi64(c0, c1);
-    d.val = _mm_unpackhi_epi64(c0, c1);
-}
-
-inline v_int16x8 v_mul_hi(const v_int16x8& a, const v_int16x8& b) { return v_int16x8(_mm_mulhi_epi16(a.val, b.val)); }
-inline v_uint16x8 v_mul_hi(const v_uint16x8& a, const v_uint16x8& b) { return v_uint16x8(_mm_mulhi_epu16(a.val, b.val)); }
-
-//////// Dot Product ////////
-
-// 16 >> 32
-inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b)
-{ return v_int32x4(_mm_madd_epi16(a.val, b.val)); }
-inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b, const v_int32x4& c)
-{ return v_dotprod(a, b) + c; }
-
-// 32 >> 64
-inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b)
-{
-#if CV_SSE4_1
-    __m128i even = _mm_mul_epi32(a.val, b.val);
-    __m128i odd = _mm_mul_epi32(_mm_srli_epi64(a.val, 32), _mm_srli_epi64(b.val, 32));
-    return v_int64x2(_mm_add_epi64(even, odd));
-#else
-    __m128i even_u = _mm_mul_epu32(a.val, b.val);
-    __m128i odd_u = _mm_mul_epu32(_mm_srli_epi64(a.val, 32), _mm_srli_epi64(b.val, 32));
-    // convert unsigned to signed high multiplication (from: Agner Fog(veclib) and H S Warren: Hacker's delight, 2003, p. 132)
-    __m128i a_sign = _mm_srai_epi32(a.val, 31);
-    __m128i b_sign = _mm_srai_epi32(b.val, 31);
-    // |x * sign of x
-    __m128i axb  = _mm_and_si128(a.val, b_sign);
-    __m128i bxa  = _mm_and_si128(b.val, a_sign);
-    // sum of sign corrections
-    __m128i ssum = _mm_add_epi32(bxa, axb);
-    __m128i even_ssum = _mm_slli_epi64(ssum, 32);
-    __m128i odd_ssum = _mm_and_si128(ssum, _mm_set_epi32(-1, 0, -1, 0));
-    // convert to signed and prod
-    return v_int64x2(_mm_add_epi64(_mm_sub_epi64(even_u, even_ssum), _mm_sub_epi64(odd_u, odd_ssum)));
-#endif
-}
-inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b, const v_int64x2& c)
-{ return v_dotprod(a, b) + c; }
-
-// 8 >> 32
-inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b)
-{
-    __m128i a0 = _mm_srli_epi16(_mm_slli_si128(a.val, 1), 8); // even
-    __m128i a1 = _mm_srli_epi16(a.val, 8); // odd
-    __m128i b0 = _mm_srli_epi16(_mm_slli_si128(b.val, 1), 8);
-    __m128i b1 = _mm_srli_epi16(b.val, 8);
-    __m128i p0 = _mm_madd_epi16(a0, b0);
-    __m128i p1 = _mm_madd_epi16(a1, b1);
-    return v_uint32x4(_mm_add_epi32(p0, p1));
-}
-inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b, const v_uint32x4& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b)
-{
-    __m128i a0 = _mm_srai_epi16(_mm_slli_si128(a.val, 1), 8); // even
-    __m128i a1 = _mm_srai_epi16(a.val, 8); // odd
-    __m128i b0 = _mm_srai_epi16(_mm_slli_si128(b.val, 1), 8);
-    __m128i b1 = _mm_srai_epi16(b.val, 8);
-    __m128i p0 = _mm_madd_epi16(a0, b0);
-    __m128i p1 = _mm_madd_epi16(a1, b1);
-    return v_int32x4(_mm_add_epi32(p0, p1));
-}
-inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b, const v_int32x4& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-// 16 >> 64
-inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b)
-{
-    v_uint32x4 c, d;
-    v_mul_expand(a, b, c, d);
-
-    v_uint64x2 c0, c1, d0, d1;
-    v_expand(c, c0, c1);
-    v_expand(d, d0, d1);
-
-    c0 += c1; d0 += d1;
-    return v_uint64x2(_mm_add_epi64(
-        _mm_unpacklo_epi64(c0.val, d0.val),
-        _mm_unpackhi_epi64(c0.val, d0.val)
-    ));
-}
-inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b, const v_uint64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-inline v_int64x2 v_dotprod_expand(const v_int16x8& a, const v_int16x8& b)
-{
-    v_int32x4 prod = v_dotprod(a, b);
-    v_int64x2 c, d;
-    v_expand(prod, c, d);
-    return v_int64x2(_mm_add_epi64(
-        _mm_unpacklo_epi64(c.val, d.val),
-        _mm_unpackhi_epi64(c.val, d.val)
-    ));
-}
-inline v_int64x2 v_dotprod_expand(const v_int16x8& a, const v_int16x8& b, const v_int64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-// 32 >> 64f
-inline v_float64x2 v_dotprod_expand(const v_int32x4& a, const v_int32x4& b)
-{
-#if CV_SSE4_1
-    return v_cvt_f64(v_dotprod(a, b));
-#else
-    v_float64x2 c = v_cvt_f64(a) * v_cvt_f64(b);
-    v_float64x2 d = v_cvt_f64_high(a) * v_cvt_f64_high(b);
-
-    return v_float64x2(_mm_add_pd(
-        _mm_unpacklo_pd(c.val, d.val),
-        _mm_unpackhi_pd(c.val, d.val)
-    ));
-#endif
-}
-inline v_float64x2 v_dotprod_expand(const v_int32x4& a, const v_int32x4& b, const v_float64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-//////// Fast Dot Product ////////
-
-// 16 >> 32
-inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b)
-{ return v_dotprod(a, b); }
-inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b, const v_int32x4& c)
-{ return v_dotprod(a, b) + c; }
-
-// 32 >> 64
-inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b)
-{ return v_dotprod(a, b); }
-inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b, const v_int64x2& c)
-{ return v_dotprod_fast(a, b) + c; }
-
-// 8 >> 32
-inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b)
-{
-    __m128i a0 = v_expand_low(a).val;
-    __m128i a1 = v_expand_high(a).val;
-    __m128i b0 = v_expand_low(b).val;
-    __m128i b1 = v_expand_high(b).val;
-    __m128i p0 = _mm_madd_epi16(a0, b0);
-    __m128i p1 = _mm_madd_epi16(a1, b1);
-    return v_uint32x4(_mm_add_epi32(p0, p1));
-}
-inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b, const v_uint32x4& c)
-{ return v_dotprod_expand_fast(a, b) + c; }
-
-inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b)
-{
-#if CV_SSE4_1
-    __m128i a0 = _mm_cvtepi8_epi16(a.val);
-    __m128i a1 = v_expand_high(a).val;
-    __m128i b0 = _mm_cvtepi8_epi16(b.val);
-    __m128i b1 = v_expand_high(b).val;
-    __m128i p0 = _mm_madd_epi16(a0, b0);
-    __m128i p1 = _mm_madd_epi16(a1, b1);
-    return v_int32x4(_mm_add_epi32(p0, p1));
-#else
-    return v_dotprod_expand(a, b);
-#endif
-}
-inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b, const v_int32x4& c)
-{ return v_dotprod_expand_fast(a, b) + c; }
-
-// 16 >> 64
-inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b)
-{
-    v_uint32x4 c, d;
-    v_mul_expand(a, b, c, d);
-
-    v_uint64x2 c0, c1, d0, d1;
-    v_expand(c, c0, c1);
-    v_expand(d, d0, d1);
-
-    c0 += c1; d0 += d1;
-    return c0 + d0;
-}
-inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b, const v_uint64x2& c)
-{ return v_dotprod_expand_fast(a, b) + c; }
-
-inline v_int64x2 v_dotprod_expand_fast(const v_int16x8& a, const v_int16x8& b)
-{
-    v_int32x4 prod = v_dotprod(a, b);
-    v_int64x2 c, d;
-    v_expand(prod, c, d);
-    return c + d;
-}
-inline v_int64x2 v_dotprod_expand_fast(const v_int16x8& a, const v_int16x8& b, const v_int64x2& c)
-{ return v_dotprod_expand_fast(a, b) + c; }
-
-// 32 >> 64f
-v_float64x2 v_fma(const v_float64x2& a, const v_float64x2& b, const v_float64x2& c);
-inline v_float64x2 v_dotprod_expand_fast(const v_int32x4& a, const v_int32x4& b)
-{ return v_fma(v_cvt_f64(a), v_cvt_f64(b), v_cvt_f64_high(a) * v_cvt_f64_high(b)); }
-inline v_float64x2 v_dotprod_expand_fast(const v_int32x4& a,   const v_int32x4& b, const v_float64x2& c)
-{ return v_fma(v_cvt_f64(a), v_cvt_f64(b), v_fma(v_cvt_f64_high(a), v_cvt_f64_high(b), c)); }
-
-#define OPENCV_HAL_IMPL_SSE_LOGIC_OP(_Tpvec, suffix, not_const) \
-    OPENCV_HAL_IMPL_SSE_BIN_OP(&, _Tpvec, _mm_and_##suffix) \
-    OPENCV_HAL_IMPL_SSE_BIN_OP(|, _Tpvec, _mm_or_##suffix) \
-    OPENCV_HAL_IMPL_SSE_BIN_OP(^, _Tpvec, _mm_xor_##suffix) \
-    inline _Tpvec operator ~ (const _Tpvec& a) \
-    { \
-        return _Tpvec(_mm_xor_##suffix(a.val, not_const)); \
-    }
-
-OPENCV_HAL_IMPL_SSE_LOGIC_OP(v_uint8x16, si128, _mm_set1_epi32(-1))
-OPENCV_HAL_IMPL_SSE_LOGIC_OP(v_int8x16, si128, _mm_set1_epi32(-1))
-OPENCV_HAL_IMPL_SSE_LOGIC_OP(v_uint16x8, si128, _mm_set1_epi32(-1))
-OPENCV_HAL_IMPL_SSE_LOGIC_OP(v_int16x8, si128, _mm_set1_epi32(-1))
-OPENCV_HAL_IMPL_SSE_LOGIC_OP(v_uint32x4, si128, _mm_set1_epi32(-1))
-OPENCV_HAL_IMPL_SSE_LOGIC_OP(v_int32x4, si128, _mm_set1_epi32(-1))
-OPENCV_HAL_IMPL_SSE_LOGIC_OP(v_uint64x2, si128, _mm_set1_epi32(-1))
-OPENCV_HAL_IMPL_SSE_LOGIC_OP(v_int64x2, si128, _mm_set1_epi32(-1))
-OPENCV_HAL_IMPL_SSE_LOGIC_OP(v_float32x4, ps, _mm_castsi128_ps(_mm_set1_epi32(-1)))
-OPENCV_HAL_IMPL_SSE_LOGIC_OP(v_float64x2, pd, _mm_castsi128_pd(_mm_set1_epi32(-1)))
-
-inline v_float32x4 v_sqrt(const v_float32x4& x)
-{ return v_float32x4(_mm_sqrt_ps(x.val)); }
-
-inline v_float32x4 v_invsqrt(const v_float32x4& x)
-{
-    const __m128 _0_5 = _mm_set1_ps(0.5f), _1_5 = _mm_set1_ps(1.5f);
-    __m128 t = x.val;
-    __m128 h = _mm_mul_ps(t, _0_5);
-    t = _mm_rsqrt_ps(t);
-    t = _mm_mul_ps(t, _mm_sub_ps(_1_5, _mm_mul_ps(_mm_mul_ps(t, t), h)));
-    return v_float32x4(t);
-}
-
-inline v_float64x2 v_sqrt(const v_float64x2& x)
-{ return v_float64x2(_mm_sqrt_pd(x.val)); }
-
-inline v_float64x2 v_invsqrt(const v_float64x2& x)
-{
-    const __m128d v_1 = _mm_set1_pd(1.);
-    return v_float64x2(_mm_div_pd(v_1, _mm_sqrt_pd(x.val)));
-}
-
-#define OPENCV_HAL_IMPL_SSE_ABS_INT_FUNC(_Tpuvec, _Tpsvec, func, suffix, subWidth) \
-inline _Tpuvec v_abs(const _Tpsvec& x) \
-{ return _Tpuvec(_mm_##func##_ep##suffix(x.val, _mm_sub_ep##subWidth(_mm_setzero_si128(), x.val))); }
-
-OPENCV_HAL_IMPL_SSE_ABS_INT_FUNC(v_uint8x16, v_int8x16, min, u8, i8)
-OPENCV_HAL_IMPL_SSE_ABS_INT_FUNC(v_uint16x8, v_int16x8, max, i16, i16)
-inline v_uint32x4 v_abs(const v_int32x4& x)
-{
-    __m128i s = _mm_srli_epi32(x.val, 31);
-    __m128i f = _mm_srai_epi32(x.val, 31);
-    return v_uint32x4(_mm_add_epi32(_mm_xor_si128(x.val, f), s));
-}
-inline v_float32x4 v_abs(const v_float32x4& x)
-{ return v_float32x4(_mm_and_ps(x.val, _mm_castsi128_ps(_mm_set1_epi32(0x7fffffff)))); }
-inline v_float64x2 v_abs(const v_float64x2& x)
-{
-    return v_float64x2(_mm_and_pd(x.val,
-        _mm_castsi128_pd(_mm_srli_epi64(_mm_set1_epi32(-1), 1))));
-}
-
-// TODO: exp, log, sin, cos
-
-#define OPENCV_HAL_IMPL_SSE_BIN_FUNC(_Tpvec, func, intrin) \
-inline _Tpvec func(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(intrin(a.val, b.val)); \
-}
-
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_uint8x16, v_min, _mm_min_epu8)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_uint8x16, v_max, _mm_max_epu8)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_int16x8, v_min, _mm_min_epi16)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_int16x8, v_max, _mm_max_epi16)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_float32x4, v_min, _mm_min_ps)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_float32x4, v_max, _mm_max_ps)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_float64x2, v_min, _mm_min_pd)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_float64x2, v_max, _mm_max_pd)
-
-inline v_int8x16 v_min(const v_int8x16& a, const v_int8x16& b)
-{
-#if CV_SSE4_1
-    return v_int8x16(_mm_min_epi8(a.val, b.val));
-#else
-    __m128i delta = _mm_set1_epi8((char)-128);
-    return v_int8x16(_mm_xor_si128(delta, _mm_min_epu8(_mm_xor_si128(a.val, delta),
-                                                       _mm_xor_si128(b.val, delta))));
-#endif
-}
-inline v_int8x16 v_max(const v_int8x16& a, const v_int8x16& b)
-{
-#if CV_SSE4_1
-    return v_int8x16(_mm_max_epi8(a.val, b.val));
-#else
-    __m128i delta = _mm_set1_epi8((char)-128);
-    return v_int8x16(_mm_xor_si128(delta, _mm_max_epu8(_mm_xor_si128(a.val, delta),
-                                                       _mm_xor_si128(b.val, delta))));
-#endif
-}
-inline v_uint16x8 v_min(const v_uint16x8& a, const v_uint16x8& b)
-{
-#if CV_SSE4_1
-    return v_uint16x8(_mm_min_epu16(a.val, b.val));
-#else
-    return v_uint16x8(_mm_subs_epu16(a.val, _mm_subs_epu16(a.val, b.val)));
-#endif
-}
-inline v_uint16x8 v_max(const v_uint16x8& a, const v_uint16x8& b)
-{
-#if CV_SSE4_1
-    return v_uint16x8(_mm_max_epu16(a.val, b.val));
-#else
-    return v_uint16x8(_mm_adds_epu16(_mm_subs_epu16(a.val, b.val), b.val));
-#endif
-}
-inline v_uint32x4 v_min(const v_uint32x4& a, const v_uint32x4& b)
-{
-#if CV_SSE4_1
-    return v_uint32x4(_mm_min_epu32(a.val, b.val));
-#else
-    __m128i delta = _mm_set1_epi32((int)0x80000000);
-    __m128i mask = _mm_cmpgt_epi32(_mm_xor_si128(a.val, delta), _mm_xor_si128(b.val, delta));
-    return v_uint32x4(v_select_si128(mask, b.val, a.val));
-#endif
-}
-inline v_uint32x4 v_max(const v_uint32x4& a, const v_uint32x4& b)
-{
-#if CV_SSE4_1
-    return v_uint32x4(_mm_max_epu32(a.val, b.val));
-#else
-    __m128i delta = _mm_set1_epi32((int)0x80000000);
-    __m128i mask = _mm_cmpgt_epi32(_mm_xor_si128(a.val, delta), _mm_xor_si128(b.val, delta));
-    return v_uint32x4(v_select_si128(mask, a.val, b.val));
-#endif
-}
-inline v_int32x4 v_min(const v_int32x4& a, const v_int32x4& b)
-{
-#if CV_SSE4_1
-    return v_int32x4(_mm_min_epi32(a.val, b.val));
-#else
-    return v_int32x4(v_select_si128(_mm_cmpgt_epi32(a.val, b.val), b.val, a.val));
-#endif
-}
-inline v_int32x4 v_max(const v_int32x4& a, const v_int32x4& b)
-{
-#if CV_SSE4_1
-    return v_int32x4(_mm_max_epi32(a.val, b.val));
-#else
-    return v_int32x4(v_select_si128(_mm_cmpgt_epi32(a.val, b.val), a.val, b.val));
-#endif
-}
-
-#define OPENCV_HAL_IMPL_SSE_INT_CMP_OP(_Tpuvec, _Tpsvec, suffix, sbit) \
-inline _Tpuvec operator == (const _Tpuvec& a, const _Tpuvec& b) \
-{ return _Tpuvec(_mm_cmpeq_##suffix(a.val, b.val)); } \
-inline _Tpuvec operator != (const _Tpuvec& a, const _Tpuvec& b) \
-{ \
-    __m128i not_mask = _mm_set1_epi32(-1); \
-    return _Tpuvec(_mm_xor_si128(_mm_cmpeq_##suffix(a.val, b.val), not_mask)); \
-} \
-inline _Tpsvec operator == (const _Tpsvec& a, const _Tpsvec& b) \
-{ return _Tpsvec(_mm_cmpeq_##suffix(a.val, b.val)); } \
-inline _Tpsvec operator != (const _Tpsvec& a, const _Tpsvec& b) \
-{ \
-    __m128i not_mask = _mm_set1_epi32(-1); \
-    return _Tpsvec(_mm_xor_si128(_mm_cmpeq_##suffix(a.val, b.val), not_mask)); \
-} \
-inline _Tpuvec operator < (const _Tpuvec& a, const _Tpuvec& b) \
-{ \
-    __m128i smask = _mm_set1_##suffix(sbit); \
-    return _Tpuvec(_mm_cmpgt_##suffix(_mm_xor_si128(b.val, smask), _mm_xor_si128(a.val, smask))); \
-} \
-inline _Tpuvec operator > (const _Tpuvec& a, const _Tpuvec& b) \
-{ \
-    __m128i smask = _mm_set1_##suffix(sbit); \
-    return _Tpuvec(_mm_cmpgt_##suffix(_mm_xor_si128(a.val, smask), _mm_xor_si128(b.val, smask))); \
-} \
-inline _Tpuvec operator <= (const _Tpuvec& a, const _Tpuvec& b) \
-{ \
-    __m128i smask = _mm_set1_##suffix(sbit); \
-    __m128i not_mask = _mm_set1_epi32(-1); \
-    __m128i res = _mm_cmpgt_##suffix(_mm_xor_si128(a.val, smask), _mm_xor_si128(b.val, smask)); \
-    return _Tpuvec(_mm_xor_si128(res, not_mask)); \
-} \
-inline _Tpuvec operator >= (const _Tpuvec& a, const _Tpuvec& b) \
-{ \
-    __m128i smask = _mm_set1_##suffix(sbit); \
-    __m128i not_mask = _mm_set1_epi32(-1); \
-    __m128i res = _mm_cmpgt_##suffix(_mm_xor_si128(b.val, smask), _mm_xor_si128(a.val, smask)); \
-    return _Tpuvec(_mm_xor_si128(res, not_mask)); \
-} \
-inline _Tpsvec operator < (const _Tpsvec& a, const _Tpsvec& b) \
-{ \
-    return _Tpsvec(_mm_cmpgt_##suffix(b.val, a.val)); \
-} \
-inline _Tpsvec operator > (const _Tpsvec& a, const _Tpsvec& b) \
-{ \
-    return _Tpsvec(_mm_cmpgt_##suffix(a.val, b.val)); \
-} \
-inline _Tpsvec operator <= (const _Tpsvec& a, const _Tpsvec& b) \
-{ \
-    __m128i not_mask = _mm_set1_epi32(-1); \
-    return _Tpsvec(_mm_xor_si128(_mm_cmpgt_##suffix(a.val, b.val), not_mask)); \
-} \
-inline _Tpsvec operator >= (const _Tpsvec& a, const _Tpsvec& b) \
-{ \
-    __m128i not_mask = _mm_set1_epi32(-1); \
-    return _Tpsvec(_mm_xor_si128(_mm_cmpgt_##suffix(b.val, a.val), not_mask)); \
-}
-
-OPENCV_HAL_IMPL_SSE_INT_CMP_OP(v_uint8x16, v_int8x16, epi8, (char)-128)
-OPENCV_HAL_IMPL_SSE_INT_CMP_OP(v_uint16x8, v_int16x8, epi16, (short)-32768)
-OPENCV_HAL_IMPL_SSE_INT_CMP_OP(v_uint32x4, v_int32x4, epi32, (int)0x80000000)
-
-#define OPENCV_HAL_IMPL_SSE_FLT_CMP_OP(_Tpvec, suffix) \
-inline _Tpvec operator == (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(_mm_cmpeq_##suffix(a.val, b.val)); } \
-inline _Tpvec operator != (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(_mm_cmpneq_##suffix(a.val, b.val)); } \
-inline _Tpvec operator < (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(_mm_cmplt_##suffix(a.val, b.val)); } \
-inline _Tpvec operator > (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(_mm_cmpgt_##suffix(a.val, b.val)); } \
-inline _Tpvec operator <= (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(_mm_cmple_##suffix(a.val, b.val)); } \
-inline _Tpvec operator >= (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(_mm_cmpge_##suffix(a.val, b.val)); }
-
-OPENCV_HAL_IMPL_SSE_FLT_CMP_OP(v_float32x4, ps)
-OPENCV_HAL_IMPL_SSE_FLT_CMP_OP(v_float64x2, pd)
-
-#if CV_SSE4_1
-#define OPENCV_HAL_IMPL_SSE_64BIT_CMP_OP(_Tpvec) \
-inline _Tpvec operator == (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(_mm_cmpeq_epi64(a.val, b.val)); } \
-inline _Tpvec operator != (const _Tpvec& a, const _Tpvec& b) \
-{ return ~(a == b); }
-#else
-#define OPENCV_HAL_IMPL_SSE_64BIT_CMP_OP(_Tpvec) \
-inline _Tpvec operator == (const _Tpvec& a, const _Tpvec& b) \
-{ __m128i cmp = _mm_cmpeq_epi32(a.val, b.val); \
-  return _Tpvec(_mm_and_si128(cmp, _mm_shuffle_epi32(cmp, _MM_SHUFFLE(2, 3, 0, 1)))); } \
-inline _Tpvec operator != (const _Tpvec& a, const _Tpvec& b) \
-{ return ~(a == b); }
-#endif
-
-OPENCV_HAL_IMPL_SSE_64BIT_CMP_OP(v_uint64x2)
-OPENCV_HAL_IMPL_SSE_64BIT_CMP_OP(v_int64x2)
-
-inline v_float32x4 v_not_nan(const v_float32x4& a)
-{ return v_float32x4(_mm_cmpord_ps(a.val, a.val)); }
-inline v_float64x2 v_not_nan(const v_float64x2& a)
-{ return v_float64x2(_mm_cmpord_pd(a.val, a.val)); }
-
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_uint8x16, v_add_wrap, _mm_add_epi8)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_int8x16, v_add_wrap, _mm_add_epi8)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_uint16x8, v_add_wrap, _mm_add_epi16)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_int16x8, v_add_wrap, _mm_add_epi16)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_uint8x16, v_sub_wrap, _mm_sub_epi8)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_int8x16, v_sub_wrap, _mm_sub_epi8)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_uint16x8, v_sub_wrap, _mm_sub_epi16)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_int16x8, v_sub_wrap, _mm_sub_epi16)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_uint16x8, v_mul_wrap, _mm_mullo_epi16)
-OPENCV_HAL_IMPL_SSE_BIN_FUNC(v_int16x8, v_mul_wrap, _mm_mullo_epi16)
-
-inline v_uint8x16 v_mul_wrap(const v_uint8x16& a, const v_uint8x16& b)
-{
-    __m128i ad = _mm_srai_epi16(a.val, 8);
-    __m128i bd = _mm_srai_epi16(b.val, 8);
-    __m128i p0 = _mm_mullo_epi16(a.val, b.val); // even
-    __m128i p1 = _mm_slli_epi16(_mm_mullo_epi16(ad, bd), 8); // odd
-    const __m128i b01 = _mm_set1_epi32(0xFF00FF00);
-    return v_uint8x16(_v128_blendv_epi8(p0, p1, b01));
-}
-inline v_int8x16 v_mul_wrap(const v_int8x16& a, const v_int8x16& b)
-{
-    return v_reinterpret_as_s8(v_mul_wrap(v_reinterpret_as_u8(a), v_reinterpret_as_u8(b)));
-}
-
-/** Absolute difference **/
-
-inline v_uint8x16 v_absdiff(const v_uint8x16& a, const v_uint8x16& b)
-{ return v_add_wrap(a - b,  b - a); }
-inline v_uint16x8 v_absdiff(const v_uint16x8& a, const v_uint16x8& b)
-{ return v_add_wrap(a - b,  b - a); }
-inline v_uint32x4 v_absdiff(const v_uint32x4& a, const v_uint32x4& b)
-{ return v_max(a, b) - v_min(a, b); }
-
-inline v_uint8x16 v_absdiff(const v_int8x16& a, const v_int8x16& b)
-{
-    v_int8x16 d = v_sub_wrap(a, b);
-    v_int8x16 m = a < b;
-    return v_reinterpret_as_u8(v_sub_wrap(d ^ m, m));
-}
-inline v_uint16x8 v_absdiff(const v_int16x8& a, const v_int16x8& b)
-{
-    return v_reinterpret_as_u16(v_sub_wrap(v_max(a, b), v_min(a, b)));
-}
-inline v_uint32x4 v_absdiff(const v_int32x4& a, const v_int32x4& b)
-{
-    v_int32x4 d = a - b;
-    v_int32x4 m = a < b;
-    return v_reinterpret_as_u32((d ^ m) - m);
-}
-
-/** Saturating absolute difference **/
-inline v_int8x16 v_absdiffs(const v_int8x16& a, const v_int8x16& b)
-{
-    v_int8x16 d = a - b;
-    v_int8x16 m = a < b;
-    return (d ^ m) - m;
- }
-inline v_int16x8 v_absdiffs(const v_int16x8& a, const v_int16x8& b)
-{ return v_max(a, b) - v_min(a, b); }
-
-
-inline v_int32x4 v_fma(const v_int32x4& a, const v_int32x4& b, const v_int32x4& c)
-{
-    return a * b + c;
-}
-
-inline v_int32x4 v_muladd(const v_int32x4& a, const v_int32x4& b, const v_int32x4& c)
-{
-    return v_fma(a, b, c);
-}
-
-inline v_float32x4 v_fma(const v_float32x4& a, const v_float32x4& b, const v_float32x4& c)
-{
-#if CV_FMA3
-    return v_float32x4(_mm_fmadd_ps(a.val, b.val, c.val));
-#else
-    return v_float32x4(_mm_add_ps(_mm_mul_ps(a.val, b.val), c.val));
-#endif
-}
-
-inline v_float64x2 v_fma(const v_float64x2& a, const v_float64x2& b, const v_float64x2& c)
-{
-#if CV_FMA3
-    return v_float64x2(_mm_fmadd_pd(a.val, b.val, c.val));
-#else
-    return v_float64x2(_mm_add_pd(_mm_mul_pd(a.val, b.val), c.val));
-#endif
-}
-
-#define OPENCV_HAL_IMPL_SSE_MISC_FLT_OP(_Tpvec, _Tp, _Tpreg, suffix, absmask_vec) \
-inline _Tpvec v_absdiff(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    _Tpreg absmask = _mm_castsi128_##suffix(absmask_vec); \
-    return _Tpvec(_mm_and_##suffix(_mm_sub_##suffix(a.val, b.val), absmask)); \
-} \
-inline _Tpvec v_magnitude(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    _Tpvec res = v_fma(a, a, b*b); \
-    return _Tpvec(_mm_sqrt_##suffix(res.val)); \
-} \
-inline _Tpvec v_sqr_magnitude(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return v_fma(a, a, b*b); \
-} \
-inline _Tpvec v_muladd(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c) \
-{ \
-    return v_fma(a, b, c); \
-}
-
-OPENCV_HAL_IMPL_SSE_MISC_FLT_OP(v_float32x4, float, __m128, ps, _mm_set1_epi32((int)0x7fffffff))
-OPENCV_HAL_IMPL_SSE_MISC_FLT_OP(v_float64x2, double, __m128d, pd, _mm_srli_epi64(_mm_set1_epi32(-1), 1))
-
-#define OPENCV_HAL_IMPL_SSE_SHIFT_OP(_Tpuvec, _Tpsvec, suffix, srai) \
-inline _Tpuvec operator << (const _Tpuvec& a, int imm) \
-{ \
-    return _Tpuvec(_mm_slli_##suffix(a.val, imm)); \
-} \
-inline _Tpsvec operator << (const _Tpsvec& a, int imm) \
-{ \
-    return _Tpsvec(_mm_slli_##suffix(a.val, imm)); \
-} \
-inline _Tpuvec operator >> (const _Tpuvec& a, int imm) \
-{ \
-    return _Tpuvec(_mm_srli_##suffix(a.val, imm)); \
-} \
-inline _Tpsvec operator >> (const _Tpsvec& a, int imm) \
-{ \
-    return _Tpsvec(srai(a.val, imm)); \
-} \
-template<int imm> \
-inline _Tpuvec v_shl(const _Tpuvec& a) \
-{ \
-    return _Tpuvec(_mm_slli_##suffix(a.val, imm)); \
-} \
-template<int imm> \
-inline _Tpsvec v_shl(const _Tpsvec& a) \
-{ \
-    return _Tpsvec(_mm_slli_##suffix(a.val, imm)); \
-} \
-template<int imm> \
-inline _Tpuvec v_shr(const _Tpuvec& a) \
-{ \
-    return _Tpuvec(_mm_srli_##suffix(a.val, imm)); \
-} \
-template<int imm> \
-inline _Tpsvec v_shr(const _Tpsvec& a) \
-{ \
-    return _Tpsvec(srai(a.val, imm)); \
-}
-
-OPENCV_HAL_IMPL_SSE_SHIFT_OP(v_uint16x8, v_int16x8, epi16, _mm_srai_epi16)
-OPENCV_HAL_IMPL_SSE_SHIFT_OP(v_uint32x4, v_int32x4, epi32, _mm_srai_epi32)
-OPENCV_HAL_IMPL_SSE_SHIFT_OP(v_uint64x2, v_int64x2, epi64, v_srai_epi64)
-
-namespace hal_sse_internal
-{
-    template <int imm,
-        bool is_invalid = ((imm < 0) || (imm > 16)),
-        bool is_first = (imm == 0),
-        bool is_half = (imm == 8),
-        bool is_second = (imm == 16),
-        bool is_other = (((imm > 0) && (imm < 8)) || ((imm > 8) && (imm < 16)))>
-    class v_sse_palignr_u8_class;
-
-    template <int imm>
-    class v_sse_palignr_u8_class<imm, true, false, false, false, false>;
-
-    template <int imm>
-    class v_sse_palignr_u8_class<imm, false, true, false, false, false>
-    {
-    public:
-        inline __m128i operator()(const __m128i& a, const __m128i&) const
-        {
-            return a;
-        }
-    };
-
-    template <int imm>
-    class v_sse_palignr_u8_class<imm, false, false, true, false, false>
-    {
-    public:
-        inline __m128i operator()(const __m128i& a, const __m128i& b) const
-        {
-            return _mm_unpacklo_epi64(_mm_unpackhi_epi64(a, a), b);
-        }
-    };
-
-    template <int imm>
-    class v_sse_palignr_u8_class<imm, false, false, false, true, false>
-    {
-    public:
-        inline __m128i operator()(const __m128i&, const __m128i& b) const
-        {
-            return b;
-        }
-    };
-
-    template <int imm>
-    class v_sse_palignr_u8_class<imm, false, false, false, false, true>
-    {
-#if CV_SSSE3
-    public:
-        inline __m128i operator()(const __m128i& a, const __m128i& b) const
-        {
-            return _mm_alignr_epi8(b, a, imm);
-        }
-#else
-    public:
-        inline __m128i operator()(const __m128i& a, const __m128i& b) const
-        {
-            enum { imm2 = (sizeof(__m128i) - imm) };
-            return _mm_or_si128(_mm_srli_si128(a, imm), _mm_slli_si128(b, imm2));
-        }
-#endif
-    };
-
-    template <int imm>
-    inline __m128i v_sse_palignr_u8(const __m128i& a, const __m128i& b)
-    {
-        CV_StaticAssert((imm >= 0) && (imm <= 16), "Invalid imm for v_sse_palignr_u8.");
-        return v_sse_palignr_u8_class<imm>()(a, b);
-    }
-}
-
-template<int imm, typename _Tpvec>
-inline _Tpvec v_rotate_right(const _Tpvec &a)
-{
-    using namespace hal_sse_internal;
-    enum { imm2 = (imm * sizeof(typename _Tpvec::lane_type)) };
-    return _Tpvec(v_sse_reinterpret_as<typename _Tpvec::vector_type>(
-        _mm_srli_si128(
-            v_sse_reinterpret_as<__m128i>(a.val), imm2)));
-}
-
-template<int imm, typename _Tpvec>
-inline _Tpvec v_rotate_left(const _Tpvec &a)
-{
-    using namespace hal_sse_internal;
-    enum { imm2 = (imm * sizeof(typename _Tpvec::lane_type)) };
-    return _Tpvec(v_sse_reinterpret_as<typename _Tpvec::vector_type>(
-        _mm_slli_si128(
-            v_sse_reinterpret_as<__m128i>(a.val), imm2)));
-}
-
-template<int imm, typename _Tpvec>
-inline _Tpvec v_rotate_right(const _Tpvec &a, const _Tpvec &b)
-{
-    using namespace hal_sse_internal;
-    enum { imm2 = (imm * sizeof(typename _Tpvec::lane_type)) };
-    return _Tpvec(v_sse_reinterpret_as<typename _Tpvec::vector_type>(
-        v_sse_palignr_u8<imm2>(
-            v_sse_reinterpret_as<__m128i>(a.val),
-            v_sse_reinterpret_as<__m128i>(b.val))));
-}
-
-template<int imm, typename _Tpvec>
-inline _Tpvec v_rotate_left(const _Tpvec &a, const _Tpvec &b)
-{
-    using namespace hal_sse_internal;
-    enum { imm2 = ((_Tpvec::nlanes - imm) * sizeof(typename _Tpvec::lane_type)) };
-    return _Tpvec(v_sse_reinterpret_as<typename _Tpvec::vector_type>(
-        v_sse_palignr_u8<imm2>(
-            v_sse_reinterpret_as<__m128i>(b.val),
-            v_sse_reinterpret_as<__m128i>(a.val))));
-}
-
-#define OPENCV_HAL_IMPL_SSE_LOADSTORE_INT_OP(_Tpvec, _Tp) \
-inline _Tpvec v_load(const _Tp* ptr) \
-{ return _Tpvec(_mm_loadu_si128((const __m128i*)ptr)); } \
-inline _Tpvec v_load_aligned(const _Tp* ptr) \
-{ return _Tpvec(_mm_load_si128((const __m128i*)ptr)); } \
-inline _Tpvec v_load_low(const _Tp* ptr) \
-{ return _Tpvec(_mm_loadl_epi64((const __m128i*)ptr)); } \
-inline _Tpvec v_load_halves(const _Tp* ptr0, const _Tp* ptr1) \
-{ \
-    return _Tpvec(_mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i*)ptr0), \
-                                     _mm_loadl_epi64((const __m128i*)ptr1))); \
-} \
-inline void v_store(_Tp* ptr, const _Tpvec& a) \
-{ _mm_storeu_si128((__m128i*)ptr, a.val); } \
-inline void v_store_aligned(_Tp* ptr, const _Tpvec& a) \
-{ _mm_store_si128((__m128i*)ptr, a.val); } \
-inline void v_store_aligned_nocache(_Tp* ptr, const _Tpvec& a) \
-{ _mm_stream_si128((__m128i*)ptr, a.val); } \
-inline void v_store(_Tp* ptr, const _Tpvec& a, hal::StoreMode mode) \
-{ \
-    if( mode == hal::STORE_UNALIGNED ) \
-        _mm_storeu_si128((__m128i*)ptr, a.val); \
-    else if( mode == hal::STORE_ALIGNED_NOCACHE )  \
-        _mm_stream_si128((__m128i*)ptr, a.val); \
-    else \
-        _mm_store_si128((__m128i*)ptr, a.val); \
-} \
-inline void v_store_low(_Tp* ptr, const _Tpvec& a) \
-{ _mm_storel_epi64((__m128i*)ptr, a.val); } \
-inline void v_store_high(_Tp* ptr, const _Tpvec& a) \
-{ _mm_storel_epi64((__m128i*)ptr, _mm_unpackhi_epi64(a.val, a.val)); }
-
-OPENCV_HAL_IMPL_SSE_LOADSTORE_INT_OP(v_uint8x16, uchar)
-OPENCV_HAL_IMPL_SSE_LOADSTORE_INT_OP(v_int8x16, schar)
-OPENCV_HAL_IMPL_SSE_LOADSTORE_INT_OP(v_uint16x8, ushort)
-OPENCV_HAL_IMPL_SSE_LOADSTORE_INT_OP(v_int16x8, short)
-OPENCV_HAL_IMPL_SSE_LOADSTORE_INT_OP(v_uint32x4, unsigned)
-OPENCV_HAL_IMPL_SSE_LOADSTORE_INT_OP(v_int32x4, int)
-OPENCV_HAL_IMPL_SSE_LOADSTORE_INT_OP(v_uint64x2, uint64)
-OPENCV_HAL_IMPL_SSE_LOADSTORE_INT_OP(v_int64x2, int64)
-
-#define OPENCV_HAL_IMPL_SSE_LOADSTORE_FLT_OP(_Tpvec, _Tp, suffix) \
-inline _Tpvec v_load(const _Tp* ptr) \
-{ return _Tpvec(_mm_loadu_##suffix(ptr)); } \
-inline _Tpvec v_load_aligned(const _Tp* ptr) \
-{ return _Tpvec(_mm_load_##suffix(ptr)); } \
-inline _Tpvec v_load_low(const _Tp* ptr) \
-{ return _Tpvec(_mm_castsi128_##suffix(_mm_loadl_epi64((const __m128i*)ptr))); } \
-inline _Tpvec v_load_halves(const _Tp* ptr0, const _Tp* ptr1) \
-{ \
-    return _Tpvec(_mm_castsi128_##suffix( \
-        _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i*)ptr0), \
-                           _mm_loadl_epi64((const __m128i*)ptr1)))); \
-} \
-inline void v_store(_Tp* ptr, const _Tpvec& a) \
-{ _mm_storeu_##suffix(ptr, a.val); } \
-inline void v_store_aligned(_Tp* ptr, const _Tpvec& a) \
-{ _mm_store_##suffix(ptr, a.val); } \
-inline void v_store_aligned_nocache(_Tp* ptr, const _Tpvec& a) \
-{ _mm_stream_##suffix(ptr, a.val); } \
-inline void v_store(_Tp* ptr, const _Tpvec& a, hal::StoreMode mode) \
-{ \
-    if( mode == hal::STORE_UNALIGNED ) \
-        _mm_storeu_##suffix(ptr, a.val); \
-    else if( mode == hal::STORE_ALIGNED_NOCACHE )  \
-        _mm_stream_##suffix(ptr, a.val); \
-    else \
-        _mm_store_##suffix(ptr, a.val); \
-} \
-inline void v_store_low(_Tp* ptr, const _Tpvec& a) \
-{ _mm_storel_epi64((__m128i*)ptr, _mm_cast##suffix##_si128(a.val)); } \
-inline void v_store_high(_Tp* ptr, const _Tpvec& a) \
-{ \
-    __m128i a1 = _mm_cast##suffix##_si128(a.val); \
-    _mm_storel_epi64((__m128i*)ptr, _mm_unpackhi_epi64(a1, a1)); \
-}
-
-OPENCV_HAL_IMPL_SSE_LOADSTORE_FLT_OP(v_float32x4, float, ps)
-OPENCV_HAL_IMPL_SSE_LOADSTORE_FLT_OP(v_float64x2, double, pd)
-
-inline unsigned v_reduce_sum(const v_uint8x16& a)
-{
-    __m128i half = _mm_sad_epu8(a.val, _mm_setzero_si128());
-    return (unsigned)_mm_cvtsi128_si32(_mm_add_epi32(half, _mm_unpackhi_epi64(half, half)));
-}
-inline int v_reduce_sum(const v_int8x16& a)
-{
-    __m128i half = _mm_set1_epi8((schar)-128);
-    half = _mm_sad_epu8(_mm_xor_si128(a.val, half), _mm_setzero_si128());
-    return _mm_cvtsi128_si32(_mm_add_epi32(half, _mm_unpackhi_epi64(half, half))) - 2048;
-}
-#define OPENCV_HAL_IMPL_SSE_REDUCE_OP_16(func) \
-inline schar v_reduce_##func(const v_int8x16& a) \
-{ \
-    __m128i val = a.val; \
-    __m128i smask = _mm_set1_epi8((schar)-128); \
-    val = _mm_xor_si128(val, smask); \
-    val = _mm_##func##_epu8(val, _mm_srli_si128(val,8)); \
-    val = _mm_##func##_epu8(val, _mm_srli_si128(val,4)); \
-    val = _mm_##func##_epu8(val, _mm_srli_si128(val,2)); \
-    val = _mm_##func##_epu8(val, _mm_srli_si128(val,1)); \
-    return (schar)_mm_cvtsi128_si32(val) ^ (schar)-128; \
-} \
-inline uchar v_reduce_##func(const v_uint8x16& a) \
-{ \
-    __m128i val = a.val; \
-    val = _mm_##func##_epu8(val, _mm_srli_si128(val,8)); \
-    val = _mm_##func##_epu8(val, _mm_srli_si128(val,4)); \
-    val = _mm_##func##_epu8(val, _mm_srli_si128(val,2)); \
-    val = _mm_##func##_epu8(val, _mm_srli_si128(val,1)); \
-    return (uchar)_mm_cvtsi128_si32(val); \
-}
-OPENCV_HAL_IMPL_SSE_REDUCE_OP_16(max)
-OPENCV_HAL_IMPL_SSE_REDUCE_OP_16(min)
-
-#define OPENCV_HAL_IMPL_SSE_REDUCE_OP_8(_Tpvec, scalartype, func, suffix, sbit) \
-inline scalartype v_reduce_##func(const v_##_Tpvec& a) \
-{ \
-    __m128i val = a.val; \
-    val = _mm_##func##_##suffix(val, _mm_srli_si128(val,8)); \
-    val = _mm_##func##_##suffix(val, _mm_srli_si128(val,4)); \
-    val = _mm_##func##_##suffix(val, _mm_srli_si128(val,2)); \
-    return (scalartype)_mm_cvtsi128_si32(val); \
-} \
-inline unsigned scalartype v_reduce_##func(const v_u##_Tpvec& a) \
-{ \
-    __m128i val = a.val; \
-    __m128i smask = _mm_set1_epi16(sbit); \
-    val = _mm_xor_si128(val, smask); \
-    val = _mm_##func##_##suffix(val, _mm_srli_si128(val,8)); \
-    val = _mm_##func##_##suffix(val, _mm_srli_si128(val,4)); \
-    val = _mm_##func##_##suffix(val, _mm_srli_si128(val,2)); \
-    return (unsigned scalartype)(_mm_cvtsi128_si32(val) ^  sbit); \
-}
-OPENCV_HAL_IMPL_SSE_REDUCE_OP_8(int16x8, short, max, epi16, (short)-32768)
-OPENCV_HAL_IMPL_SSE_REDUCE_OP_8(int16x8, short, min, epi16, (short)-32768)
-
-#define OPENCV_HAL_IMPL_SSE_REDUCE_OP_4_SUM(_Tpvec, scalartype, regtype, suffix, cast_from, cast_to, extract) \
-inline scalartype v_reduce_sum(const _Tpvec& a) \
-{ \
-    regtype val = a.val; \
-    val = _mm_add_##suffix(val, cast_to(_mm_srli_si128(cast_from(val), 8))); \
-    val = _mm_add_##suffix(val, cast_to(_mm_srli_si128(cast_from(val), 4))); \
-    return (scalartype)_mm_cvt##extract(val); \
-}
-
-#define OPENCV_HAL_IMPL_SSE_REDUCE_OP_4(_Tpvec, scalartype, func, scalar_func) \
-inline scalartype v_reduce_##func(const _Tpvec& a) \
-{ \
-    scalartype CV_DECL_ALIGNED(16) buf[4]; \
-    v_store_aligned(buf, a); \
-    scalartype s0 = scalar_func(buf[0], buf[1]); \
-    scalartype s1 = scalar_func(buf[2], buf[3]); \
-    return scalar_func(s0, s1); \
-}
-
-OPENCV_HAL_IMPL_SSE_REDUCE_OP_4_SUM(v_uint32x4, unsigned, __m128i, epi32, OPENCV_HAL_NOP, OPENCV_HAL_NOP, si128_si32)
-OPENCV_HAL_IMPL_SSE_REDUCE_OP_4_SUM(v_int32x4, int, __m128i, epi32, OPENCV_HAL_NOP, OPENCV_HAL_NOP, si128_si32)
-OPENCV_HAL_IMPL_SSE_REDUCE_OP_4_SUM(v_float32x4, float, __m128, ps, _mm_castps_si128, _mm_castsi128_ps, ss_f32)
-
-inline int v_reduce_sum(const v_int16x8& a)
-{ return v_reduce_sum(v_expand_low(a) + v_expand_high(a)); }
-inline unsigned v_reduce_sum(const v_uint16x8& a)
-{ return v_reduce_sum(v_expand_low(a) + v_expand_high(a)); }
-
-inline uint64 v_reduce_sum(const v_uint64x2& a)
-{
-    uint64 CV_DECL_ALIGNED(32) idx[2];
-    v_store_aligned(idx, a);
-    return idx[0] + idx[1];
-}
-inline int64 v_reduce_sum(const v_int64x2& a)
-{
-    int64 CV_DECL_ALIGNED(32) idx[2];
-    v_store_aligned(idx, a);
-    return idx[0] + idx[1];
-}
-inline double v_reduce_sum(const v_float64x2& a)
-{
-    double CV_DECL_ALIGNED(32) idx[2];
-    v_store_aligned(idx, a);
-    return idx[0] + idx[1];
-}
-
-inline v_float32x4 v_reduce_sum4(const v_float32x4& a, const v_float32x4& b,
-                                 const v_float32x4& c, const v_float32x4& d)
-{
-#if CV_SSE3
-    __m128 ab = _mm_hadd_ps(a.val, b.val);
-    __m128 cd = _mm_hadd_ps(c.val, d.val);
-    return v_float32x4(_mm_hadd_ps(ab, cd));
-#else
-    __m128 ac = _mm_add_ps(_mm_unpacklo_ps(a.val, c.val), _mm_unpackhi_ps(a.val, c.val));
-    __m128 bd = _mm_add_ps(_mm_unpacklo_ps(b.val, d.val), _mm_unpackhi_ps(b.val, d.val));
-    return v_float32x4(_mm_add_ps(_mm_unpacklo_ps(ac, bd), _mm_unpackhi_ps(ac, bd)));
-#endif
-}
-
-OPENCV_HAL_IMPL_SSE_REDUCE_OP_4(v_uint32x4, unsigned, max, std::max)
-OPENCV_HAL_IMPL_SSE_REDUCE_OP_4(v_uint32x4, unsigned, min, std::min)
-OPENCV_HAL_IMPL_SSE_REDUCE_OP_4(v_int32x4, int, max, std::max)
-OPENCV_HAL_IMPL_SSE_REDUCE_OP_4(v_int32x4, int, min, std::min)
-OPENCV_HAL_IMPL_SSE_REDUCE_OP_4(v_float32x4, float, max, std::max)
-OPENCV_HAL_IMPL_SSE_REDUCE_OP_4(v_float32x4, float, min, std::min)
-
-inline unsigned v_reduce_sad(const v_uint8x16& a, const v_uint8x16& b)
-{
-    __m128i half = _mm_sad_epu8(a.val, b.val);
-    return (unsigned)_mm_cvtsi128_si32(_mm_add_epi32(half, _mm_unpackhi_epi64(half, half)));
-}
-inline unsigned v_reduce_sad(const v_int8x16& a, const v_int8x16& b)
-{
-    __m128i half = _mm_set1_epi8(0x7f);
-    half = _mm_sad_epu8(_mm_add_epi8(a.val, half), _mm_add_epi8(b.val, half));
-    return (unsigned)_mm_cvtsi128_si32(_mm_add_epi32(half, _mm_unpackhi_epi64(half, half)));
-}
-inline unsigned v_reduce_sad(const v_uint16x8& a, const v_uint16x8& b)
-{
-    v_uint32x4 l, h;
-    v_expand(v_absdiff(a, b), l, h);
-    return v_reduce_sum(l + h);
-}
-inline unsigned v_reduce_sad(const v_int16x8& a, const v_int16x8& b)
-{
-    v_uint32x4 l, h;
-    v_expand(v_absdiff(a, b), l, h);
-    return v_reduce_sum(l + h);
-}
-inline unsigned v_reduce_sad(const v_uint32x4& a, const v_uint32x4& b)
-{
-    return v_reduce_sum(v_absdiff(a, b));
-}
-inline unsigned v_reduce_sad(const v_int32x4& a, const v_int32x4& b)
-{
-    return v_reduce_sum(v_absdiff(a, b));
-}
-inline float v_reduce_sad(const v_float32x4& a, const v_float32x4& b)
-{
-    return v_reduce_sum(v_absdiff(a, b));
-}
-
-inline v_uint8x16 v_popcount(const v_uint8x16& a)
-{
-    __m128i m1 = _mm_set1_epi32(0x55555555);
-    __m128i m2 = _mm_set1_epi32(0x33333333);
-    __m128i m4 = _mm_set1_epi32(0x0f0f0f0f);
-    __m128i p = a.val;
-    p = _mm_add_epi32(_mm_and_si128(_mm_srli_epi32(p, 1), m1), _mm_and_si128(p, m1));
-    p = _mm_add_epi32(_mm_and_si128(_mm_srli_epi32(p, 2), m2), _mm_and_si128(p, m2));
-    p = _mm_add_epi32(_mm_and_si128(_mm_srli_epi32(p, 4), m4), _mm_and_si128(p, m4));
-    return v_uint8x16(p);
-}
-inline v_uint16x8 v_popcount(const v_uint16x8& a)
-{
-    v_uint8x16 p = v_popcount(v_reinterpret_as_u8(a));
-    p += v_rotate_right<1>(p);
-    return v_reinterpret_as_u16(p) & v_setall_u16(0x00ff);
-}
-inline v_uint32x4 v_popcount(const v_uint32x4& a)
-{
-    v_uint8x16 p = v_popcount(v_reinterpret_as_u8(a));
-    p += v_rotate_right<1>(p);
-    p += v_rotate_right<2>(p);
-    return v_reinterpret_as_u32(p) & v_setall_u32(0x000000ff);
-}
-inline v_uint64x2 v_popcount(const v_uint64x2& a)
-{
-    return v_uint64x2(_mm_sad_epu8(v_popcount(v_reinterpret_as_u8(a)).val, _mm_setzero_si128()));
-}
-inline v_uint8x16 v_popcount(const v_int8x16& a)
-{ return v_popcount(v_reinterpret_as_u8(a)); }
-inline v_uint16x8 v_popcount(const v_int16x8& a)
-{ return v_popcount(v_reinterpret_as_u16(a)); }
-inline v_uint32x4 v_popcount(const v_int32x4& a)
-{ return v_popcount(v_reinterpret_as_u32(a)); }
-inline v_uint64x2 v_popcount(const v_int64x2& a)
-{ return v_popcount(v_reinterpret_as_u64(a)); }
-
-#define OPENCV_HAL_IMPL_SSE_CHECK_SIGNS(_Tpvec, suffix, cast_op, allmask) \
-inline int v_signmask(const _Tpvec& a)   { return _mm_movemask_##suffix(cast_op(a.val)); } \
-inline bool v_check_all(const _Tpvec& a) { return _mm_movemask_##suffix(cast_op(a.val)) == allmask; } \
-inline bool v_check_any(const _Tpvec& a) { return _mm_movemask_##suffix(cast_op(a.val)) != 0; }
-OPENCV_HAL_IMPL_SSE_CHECK_SIGNS(v_uint8x16, epi8, OPENCV_HAL_NOP, 65535)
-OPENCV_HAL_IMPL_SSE_CHECK_SIGNS(v_int8x16, epi8, OPENCV_HAL_NOP, 65535)
-OPENCV_HAL_IMPL_SSE_CHECK_SIGNS(v_uint32x4, ps, _mm_castsi128_ps, 15)
-OPENCV_HAL_IMPL_SSE_CHECK_SIGNS(v_int32x4, ps, _mm_castsi128_ps, 15)
-OPENCV_HAL_IMPL_SSE_CHECK_SIGNS(v_uint64x2, pd, _mm_castsi128_pd, 3)
-OPENCV_HAL_IMPL_SSE_CHECK_SIGNS(v_int64x2, pd, _mm_castsi128_pd, 3)
-OPENCV_HAL_IMPL_SSE_CHECK_SIGNS(v_float32x4, ps, OPENCV_HAL_NOP, 15)
-OPENCV_HAL_IMPL_SSE_CHECK_SIGNS(v_float64x2, pd, OPENCV_HAL_NOP, 3)
-
-#define OPENCV_HAL_IMPL_SSE_CHECK_SIGNS_SHORT(_Tpvec) \
-inline int v_signmask(const _Tpvec& a) { return _mm_movemask_epi8(_mm_packs_epi16(a.val, a.val)) & 255; } \
-inline bool v_check_all(const _Tpvec& a) { return (_mm_movemask_epi8(a.val) & 0xaaaa) == 0xaaaa; } \
-inline bool v_check_any(const _Tpvec& a) { return (_mm_movemask_epi8(a.val) & 0xaaaa) != 0; }
-OPENCV_HAL_IMPL_SSE_CHECK_SIGNS_SHORT(v_uint16x8)
-OPENCV_HAL_IMPL_SSE_CHECK_SIGNS_SHORT(v_int16x8)
-
-inline int v_scan_forward(const v_int8x16& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))); }
-inline int v_scan_forward(const v_uint8x16& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))); }
-inline int v_scan_forward(const v_int16x8& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 2; }
-inline int v_scan_forward(const v_uint16x8& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 2; }
-inline int v_scan_forward(const v_int32x4& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 4; }
-inline int v_scan_forward(const v_uint32x4& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 4; }
-inline int v_scan_forward(const v_float32x4& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 4; }
-inline int v_scan_forward(const v_int64x2& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 8; }
-inline int v_scan_forward(const v_uint64x2& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 8; }
-inline int v_scan_forward(const v_float64x2& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 8; }
-
-#if CV_SSE4_1
-#define OPENCV_HAL_IMPL_SSE_SELECT(_Tpvec, cast_ret, cast, suffix) \
-inline _Tpvec v_select(const _Tpvec& mask, const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(cast_ret(_mm_blendv_##suffix(cast(b.val), cast(a.val), cast(mask.val)))); \
-}
-
-OPENCV_HAL_IMPL_SSE_SELECT(v_uint8x16, OPENCV_HAL_NOP, OPENCV_HAL_NOP, epi8)
-OPENCV_HAL_IMPL_SSE_SELECT(v_int8x16, OPENCV_HAL_NOP, OPENCV_HAL_NOP, epi8)
-OPENCV_HAL_IMPL_SSE_SELECT(v_uint16x8, OPENCV_HAL_NOP, OPENCV_HAL_NOP, epi8)
-OPENCV_HAL_IMPL_SSE_SELECT(v_int16x8, OPENCV_HAL_NOP, OPENCV_HAL_NOP, epi8)
-OPENCV_HAL_IMPL_SSE_SELECT(v_uint32x4, _mm_castps_si128, _mm_castsi128_ps, ps)
-OPENCV_HAL_IMPL_SSE_SELECT(v_int32x4, _mm_castps_si128, _mm_castsi128_ps, ps)
-// OPENCV_HAL_IMPL_SSE_SELECT(v_uint64x2, TBD, TBD, pd)
-// OPENCV_HAL_IMPL_SSE_SELECT(v_int64x2, TBD, TBD, ps)
-OPENCV_HAL_IMPL_SSE_SELECT(v_float32x4, OPENCV_HAL_NOP, OPENCV_HAL_NOP, ps)
-OPENCV_HAL_IMPL_SSE_SELECT(v_float64x2, OPENCV_HAL_NOP, OPENCV_HAL_NOP, pd)
-
-#else // CV_SSE4_1
-
-#define OPENCV_HAL_IMPL_SSE_SELECT(_Tpvec, suffix) \
-inline _Tpvec v_select(const _Tpvec& mask, const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(_mm_xor_##suffix(b.val, _mm_and_##suffix(_mm_xor_##suffix(b.val, a.val), mask.val))); \
-}
-
-OPENCV_HAL_IMPL_SSE_SELECT(v_uint8x16, si128)
-OPENCV_HAL_IMPL_SSE_SELECT(v_int8x16, si128)
-OPENCV_HAL_IMPL_SSE_SELECT(v_uint16x8, si128)
-OPENCV_HAL_IMPL_SSE_SELECT(v_int16x8, si128)
-OPENCV_HAL_IMPL_SSE_SELECT(v_uint32x4, si128)
-OPENCV_HAL_IMPL_SSE_SELECT(v_int32x4, si128)
-// OPENCV_HAL_IMPL_SSE_SELECT(v_uint64x2, si128)
-// OPENCV_HAL_IMPL_SSE_SELECT(v_int64x2, si128)
-OPENCV_HAL_IMPL_SSE_SELECT(v_float32x4, ps)
-OPENCV_HAL_IMPL_SSE_SELECT(v_float64x2, pd)
-#endif
-
-/* Expand */
-#define OPENCV_HAL_IMPL_SSE_EXPAND(_Tpvec, _Tpwvec, _Tp, intrin)    \
-    inline void v_expand(const _Tpvec& a, _Tpwvec& b0, _Tpwvec& b1) \
-    {                                                               \
-        b0.val = intrin(a.val);                                     \
-        b1.val = __CV_CAT(intrin, _high)(a.val);                    \
-    }                                                               \
-    inline _Tpwvec v_expand_low(const _Tpvec& a)                    \
-    { return _Tpwvec(intrin(a.val)); }                              \
-    inline _Tpwvec v_expand_high(const _Tpvec& a)                   \
-    { return _Tpwvec(__CV_CAT(intrin, _high)(a.val)); }             \
-    inline _Tpwvec v_load_expand(const _Tp* ptr)                    \
-    {                                                               \
-        __m128i a = _mm_loadl_epi64((const __m128i*)ptr);           \
-        return _Tpwvec(intrin(a));                                  \
-    }
-
-OPENCV_HAL_IMPL_SSE_EXPAND(v_uint8x16, v_uint16x8,  uchar,    _v128_cvtepu8_epi16)
-OPENCV_HAL_IMPL_SSE_EXPAND(v_int8x16,  v_int16x8,   schar,    _v128_cvtepi8_epi16)
-OPENCV_HAL_IMPL_SSE_EXPAND(v_uint16x8, v_uint32x4,  ushort,   _v128_cvtepu16_epi32)
-OPENCV_HAL_IMPL_SSE_EXPAND(v_int16x8,  v_int32x4,   short,    _v128_cvtepi16_epi32)
-OPENCV_HAL_IMPL_SSE_EXPAND(v_uint32x4, v_uint64x2,  unsigned, _v128_cvtepu32_epi64)
-OPENCV_HAL_IMPL_SSE_EXPAND(v_int32x4,  v_int64x2,   int,      _v128_cvtepi32_epi64)
-
-#define OPENCV_HAL_IMPL_SSE_EXPAND_Q(_Tpvec, _Tp, intrin)  \
-    inline _Tpvec v_load_expand_q(const _Tp* ptr)          \
-    {                                                      \
-        __m128i a = _mm_cvtsi32_si128(*(const int*)ptr);   \
-        return _Tpvec(intrin(a));                          \
-    }
-
-OPENCV_HAL_IMPL_SSE_EXPAND_Q(v_uint32x4, uchar, _v128_cvtepu8_epi32)
-OPENCV_HAL_IMPL_SSE_EXPAND_Q(v_int32x4,  schar, _v128_cvtepi8_epi32)
-
-#define OPENCV_HAL_IMPL_SSE_UNPACKS(_Tpvec, suffix, cast_from, cast_to) \
-inline void v_zip(const _Tpvec& a0, const _Tpvec& a1, _Tpvec& b0, _Tpvec& b1) \
-{ \
-    b0.val = _mm_unpacklo_##suffix(a0.val, a1.val); \
-    b1.val = _mm_unpackhi_##suffix(a0.val, a1.val); \
-} \
-inline _Tpvec v_combine_low(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    __m128i a1 = cast_from(a.val), b1 = cast_from(b.val); \
-    return _Tpvec(cast_to(_mm_unpacklo_epi64(a1, b1))); \
-} \
-inline _Tpvec v_combine_high(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    __m128i a1 = cast_from(a.val), b1 = cast_from(b.val); \
-    return _Tpvec(cast_to(_mm_unpackhi_epi64(a1, b1))); \
-} \
-inline void v_recombine(const _Tpvec& a, const _Tpvec& b, _Tpvec& c, _Tpvec& d) \
-{ \
-    __m128i a1 = cast_from(a.val), b1 = cast_from(b.val); \
-    c.val = cast_to(_mm_unpacklo_epi64(a1, b1)); \
-    d.val = cast_to(_mm_unpackhi_epi64(a1, b1)); \
-}
-
-OPENCV_HAL_IMPL_SSE_UNPACKS(v_uint8x16, epi8, OPENCV_HAL_NOP, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_SSE_UNPACKS(v_int8x16, epi8, OPENCV_HAL_NOP, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_SSE_UNPACKS(v_uint16x8, epi16, OPENCV_HAL_NOP, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_SSE_UNPACKS(v_int16x8, epi16, OPENCV_HAL_NOP, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_SSE_UNPACKS(v_uint32x4, epi32, OPENCV_HAL_NOP, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_SSE_UNPACKS(v_int32x4, epi32, OPENCV_HAL_NOP, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_SSE_UNPACKS(v_float32x4, ps, _mm_castps_si128, _mm_castsi128_ps)
-OPENCV_HAL_IMPL_SSE_UNPACKS(v_float64x2, pd, _mm_castpd_si128, _mm_castsi128_pd)
-
-inline v_uint8x16 v_reverse(const v_uint8x16 &a)
-{
-#if CV_SSSE3
-    static const __m128i perm = _mm_setr_epi8(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
-    return v_uint8x16(_mm_shuffle_epi8(a.val, perm));
-#else
-    uchar CV_DECL_ALIGNED(32) d[16];
-    v_store_aligned(d, a);
-    return v_uint8x16(d[15], d[14], d[13], d[12], d[11], d[10], d[9], d[8], d[7], d[6], d[5], d[4], d[3], d[2], d[1], d[0]);
-#endif
-}
-
-inline v_int8x16 v_reverse(const v_int8x16 &a)
-{ return v_reinterpret_as_s8(v_reverse(v_reinterpret_as_u8(a))); }
-
-inline v_uint16x8 v_reverse(const v_uint16x8 &a)
-{
-#if CV_SSSE3
-    static const __m128i perm = _mm_setr_epi8(14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1);
-    return v_uint16x8(_mm_shuffle_epi8(a.val, perm));
-#else
-    __m128i r = _mm_shuffle_epi32(a.val, _MM_SHUFFLE(0, 1, 2, 3));
-    r = _mm_shufflelo_epi16(r, _MM_SHUFFLE(2, 3, 0, 1));
-    r = _mm_shufflehi_epi16(r, _MM_SHUFFLE(2, 3, 0, 1));
-    return v_uint16x8(r);
-#endif
-}
-
-inline v_int16x8 v_reverse(const v_int16x8 &a)
-{ return v_reinterpret_as_s16(v_reverse(v_reinterpret_as_u16(a))); }
-
-inline v_uint32x4 v_reverse(const v_uint32x4 &a)
-{
-    return v_uint32x4(_mm_shuffle_epi32(a.val, _MM_SHUFFLE(0, 1, 2, 3)));
-}
-
-inline v_int32x4 v_reverse(const v_int32x4 &a)
-{ return v_reinterpret_as_s32(v_reverse(v_reinterpret_as_u32(a))); }
-
-inline v_float32x4 v_reverse(const v_float32x4 &a)
-{ return v_reinterpret_as_f32(v_reverse(v_reinterpret_as_u32(a))); }
-
-inline v_uint64x2 v_reverse(const v_uint64x2 &a)
-{
-    return v_uint64x2(_mm_shuffle_epi32(a.val, _MM_SHUFFLE(1, 0, 3, 2)));
-}
-
-inline v_int64x2 v_reverse(const v_int64x2 &a)
-{ return v_reinterpret_as_s64(v_reverse(v_reinterpret_as_u64(a))); }
-
-inline v_float64x2 v_reverse(const v_float64x2 &a)
-{ return v_reinterpret_as_f64(v_reverse(v_reinterpret_as_u64(a))); }
-
-template<int s, typename _Tpvec>
-inline _Tpvec v_extract(const _Tpvec& a, const _Tpvec& b)
-{
-    return v_rotate_right<s>(a, b);
-}
-
-inline v_int32x4 v_round(const v_float32x4& a)
-{ return v_int32x4(_mm_cvtps_epi32(a.val)); }
-
-inline v_int32x4 v_floor(const v_float32x4& a)
-{
-    __m128i a1 = _mm_cvtps_epi32(a.val);
-    __m128i mask = _mm_castps_si128(_mm_cmpgt_ps(_mm_cvtepi32_ps(a1), a.val));
-    return v_int32x4(_mm_add_epi32(a1, mask));
-}
-
-inline v_int32x4 v_ceil(const v_float32x4& a)
-{
-    __m128i a1 = _mm_cvtps_epi32(a.val);
-    __m128i mask = _mm_castps_si128(_mm_cmpgt_ps(a.val, _mm_cvtepi32_ps(a1)));
-    return v_int32x4(_mm_sub_epi32(a1, mask));
-}
-
-inline v_int32x4 v_trunc(const v_float32x4& a)
-{ return v_int32x4(_mm_cvttps_epi32(a.val)); }
-
-inline v_int32x4 v_round(const v_float64x2& a)
-{ return v_int32x4(_mm_cvtpd_epi32(a.val)); }
-
-inline v_int32x4 v_round(const v_float64x2& a, const v_float64x2& b)
-{
-    __m128i ai = _mm_cvtpd_epi32(a.val), bi = _mm_cvtpd_epi32(b.val);
-    return v_int32x4(_mm_unpacklo_epi64(ai, bi));
-}
-
-inline v_int32x4 v_floor(const v_float64x2& a)
-{
-    __m128i a1 = _mm_cvtpd_epi32(a.val);
-    __m128i mask = _mm_castpd_si128(_mm_cmpgt_pd(_mm_cvtepi32_pd(a1), a.val));
-    mask = _mm_srli_si128(_mm_slli_si128(mask, 4), 8); // m0 m0 m1 m1 => m0 m1 0 0
-    return v_int32x4(_mm_add_epi32(a1, mask));
-}
-
-inline v_int32x4 v_ceil(const v_float64x2& a)
-{
-    __m128i a1 = _mm_cvtpd_epi32(a.val);
-    __m128i mask = _mm_castpd_si128(_mm_cmpgt_pd(a.val, _mm_cvtepi32_pd(a1)));
-    mask = _mm_srli_si128(_mm_slli_si128(mask, 4), 8); // m0 m0 m1 m1 => m0 m1 0 0
-    return v_int32x4(_mm_sub_epi32(a1, mask));
-}
-
-inline v_int32x4 v_trunc(const v_float64x2& a)
-{ return v_int32x4(_mm_cvttpd_epi32(a.val)); }
-
-#define OPENCV_HAL_IMPL_SSE_TRANSPOSE4x4(_Tpvec, suffix, cast_from, cast_to) \
-inline void v_transpose4x4(const _Tpvec& a0, const _Tpvec& a1, \
-                           const _Tpvec& a2, const _Tpvec& a3, \
-                           _Tpvec& b0, _Tpvec& b1, \
-                           _Tpvec& b2, _Tpvec& b3) \
-{ \
-    __m128i t0 = cast_from(_mm_unpacklo_##suffix(a0.val, a1.val)); \
-    __m128i t1 = cast_from(_mm_unpacklo_##suffix(a2.val, a3.val)); \
-    __m128i t2 = cast_from(_mm_unpackhi_##suffix(a0.val, a1.val)); \
-    __m128i t3 = cast_from(_mm_unpackhi_##suffix(a2.val, a3.val)); \
-\
-    b0.val = cast_to(_mm_unpacklo_epi64(t0, t1)); \
-    b1.val = cast_to(_mm_unpackhi_epi64(t0, t1)); \
-    b2.val = cast_to(_mm_unpacklo_epi64(t2, t3)); \
-    b3.val = cast_to(_mm_unpackhi_epi64(t2, t3)); \
-}
-
-OPENCV_HAL_IMPL_SSE_TRANSPOSE4x4(v_uint32x4, epi32, OPENCV_HAL_NOP, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_SSE_TRANSPOSE4x4(v_int32x4, epi32, OPENCV_HAL_NOP, OPENCV_HAL_NOP)
-OPENCV_HAL_IMPL_SSE_TRANSPOSE4x4(v_float32x4, ps, _mm_castps_si128, _mm_castsi128_ps)
-
-// load deinterleave
-inline void v_load_deinterleave(const uchar* ptr, v_uint8x16& a, v_uint8x16& b)
-{
-    __m128i t00 = _mm_loadu_si128((const __m128i*)ptr);
-    __m128i t01 = _mm_loadu_si128((const __m128i*)(ptr + 16));
-
-    __m128i t10 = _mm_unpacklo_epi8(t00, t01);
-    __m128i t11 = _mm_unpackhi_epi8(t00, t01);
-
-    __m128i t20 = _mm_unpacklo_epi8(t10, t11);
-    __m128i t21 = _mm_unpackhi_epi8(t10, t11);
-
-    __m128i t30 = _mm_unpacklo_epi8(t20, t21);
-    __m128i t31 = _mm_unpackhi_epi8(t20, t21);
-
-    a.val = _mm_unpacklo_epi8(t30, t31);
-    b.val = _mm_unpackhi_epi8(t30, t31);
-}
-
-inline void v_load_deinterleave(const uchar* ptr, v_uint8x16& a, v_uint8x16& b, v_uint8x16& c)
-{
-#if CV_SSE4_1
-    const __m128i m0 = _mm_setr_epi8(0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0);
-    const __m128i m1 = _mm_setr_epi8(0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0);
-    __m128i s0 = _mm_loadu_si128((const __m128i*)ptr);
-    __m128i s1 = _mm_loadu_si128((const __m128i*)(ptr + 16));
-    __m128i s2 = _mm_loadu_si128((const __m128i*)(ptr + 32));
-    __m128i a0 = _mm_blendv_epi8(_mm_blendv_epi8(s0, s1, m0), s2, m1);
-    __m128i b0 = _mm_blendv_epi8(_mm_blendv_epi8(s1, s2, m0), s0, m1);
-    __m128i c0 = _mm_blendv_epi8(_mm_blendv_epi8(s2, s0, m0), s1, m1);
-    const __m128i sh_b = _mm_setr_epi8(0, 3, 6, 9, 12, 15, 2, 5, 8, 11, 14, 1, 4, 7, 10, 13);
-    const __m128i sh_g = _mm_setr_epi8(1, 4, 7, 10, 13, 0, 3, 6, 9, 12, 15, 2, 5, 8, 11, 14);
-    const __m128i sh_r = _mm_setr_epi8(2, 5, 8, 11, 14, 1, 4, 7, 10, 13, 0, 3, 6, 9, 12, 15);
-    a0 = _mm_shuffle_epi8(a0, sh_b);
-    b0 = _mm_shuffle_epi8(b0, sh_g);
-    c0 = _mm_shuffle_epi8(c0, sh_r);
-    a.val = a0;
-    b.val = b0;
-    c.val = c0;
-#elif CV_SSSE3
-    const __m128i m0 = _mm_setr_epi8(0, 3, 6, 9, 12, 15, 1, 4, 7, 10, 13, 2, 5, 8, 11, 14);
-    const __m128i m1 = _mm_alignr_epi8(m0, m0, 11);
-    const __m128i m2 = _mm_alignr_epi8(m0, m0, 6);
-
-    __m128i t0 = _mm_loadu_si128((const __m128i*)ptr);
-    __m128i t1 = _mm_loadu_si128((const __m128i*)(ptr + 16));
-    __m128i t2 = _mm_loadu_si128((const __m128i*)(ptr + 32));
-
-    __m128i s0 = _mm_shuffle_epi8(t0, m0);
-    __m128i s1 = _mm_shuffle_epi8(t1, m1);
-    __m128i s2 = _mm_shuffle_epi8(t2, m2);
-
-    t0 = _mm_alignr_epi8(s1, _mm_slli_si128(s0, 10), 5);
-    a.val = _mm_alignr_epi8(s2, t0, 5);
-
-    t1 = _mm_alignr_epi8(_mm_srli_si128(s1, 5), _mm_slli_si128(s0, 5), 6);
-    b.val = _mm_alignr_epi8(_mm_srli_si128(s2, 5), t1, 5);
-
-    t2 = _mm_alignr_epi8(_mm_srli_si128(s2, 10), s1, 11);
-    c.val = _mm_alignr_epi8(t2, s0, 11);
-#else
-    __m128i t00 = _mm_loadu_si128((const __m128i*)ptr);
-    __m128i t01 = _mm_loadu_si128((const __m128i*)(ptr + 16));
-    __m128i t02 = _mm_loadu_si128((const __m128i*)(ptr + 32));
-
-    __m128i t10 = _mm_unpacklo_epi8(t00, _mm_unpackhi_epi64(t01, t01));
-    __m128i t11 = _mm_unpacklo_epi8(_mm_unpackhi_epi64(t00, t00), t02);
-    __m128i t12 = _mm_unpacklo_epi8(t01, _mm_unpackhi_epi64(t02, t02));
-
-    __m128i t20 = _mm_unpacklo_epi8(t10, _mm_unpackhi_epi64(t11, t11));
-    __m128i t21 = _mm_unpacklo_epi8(_mm_unpackhi_epi64(t10, t10), t12);
-    __m128i t22 = _mm_unpacklo_epi8(t11, _mm_unpackhi_epi64(t12, t12));
-
-    __m128i t30 = _mm_unpacklo_epi8(t20, _mm_unpackhi_epi64(t21, t21));
-    __m128i t31 = _mm_unpacklo_epi8(_mm_unpackhi_epi64(t20, t20), t22);
-    __m128i t32 = _mm_unpacklo_epi8(t21, _mm_unpackhi_epi64(t22, t22));
-
-    a.val = _mm_unpacklo_epi8(t30, _mm_unpackhi_epi64(t31, t31));
-    b.val = _mm_unpacklo_epi8(_mm_unpackhi_epi64(t30, t30), t32);
-    c.val = _mm_unpacklo_epi8(t31, _mm_unpackhi_epi64(t32, t32));
-#endif
-}
-
-inline void v_load_deinterleave(const uchar* ptr, v_uint8x16& a, v_uint8x16& b, v_uint8x16& c, v_uint8x16& d)
-{
-    __m128i u0 = _mm_loadu_si128((const __m128i*)ptr); // a0 b0 c0 d0 a1 b1 c1 d1 ...
-    __m128i u1 = _mm_loadu_si128((const __m128i*)(ptr + 16)); // a4 b4 c4 d4 ...
-    __m128i u2 = _mm_loadu_si128((const __m128i*)(ptr + 32)); // a8 b8 c8 d8 ...
-    __m128i u3 = _mm_loadu_si128((const __m128i*)(ptr + 48)); // a12 b12 c12 d12 ...
-
-    __m128i v0 = _mm_unpacklo_epi8(u0, u2); // a0 a8 b0 b8 ...
-    __m128i v1 = _mm_unpackhi_epi8(u0, u2); // a2 a10 b2 b10 ...
-    __m128i v2 = _mm_unpacklo_epi8(u1, u3); // a4 a12 b4 b12 ...
-    __m128i v3 = _mm_unpackhi_epi8(u1, u3); // a6 a14 b6 b14 ...
-
-    u0 = _mm_unpacklo_epi8(v0, v2); // a0 a4 a8 a12 ...
-    u1 = _mm_unpacklo_epi8(v1, v3); // a2 a6 a10 a14 ...
-    u2 = _mm_unpackhi_epi8(v0, v2); // a1 a5 a9 a13 ...
-    u3 = _mm_unpackhi_epi8(v1, v3); // a3 a7 a11 a15 ...
-
-    v0 = _mm_unpacklo_epi8(u0, u1); // a0 a2 a4 a6 ...
-    v1 = _mm_unpacklo_epi8(u2, u3); // a1 a3 a5 a7 ...
-    v2 = _mm_unpackhi_epi8(u0, u1); // c0 c2 c4 c6 ...
-    v3 = _mm_unpackhi_epi8(u2, u3); // c1 c3 c5 c7 ...
-
-    a.val = _mm_unpacklo_epi8(v0, v1);
-    b.val = _mm_unpackhi_epi8(v0, v1);
-    c.val = _mm_unpacklo_epi8(v2, v3);
-    d.val = _mm_unpackhi_epi8(v2, v3);
-}
-
-inline void v_load_deinterleave(const ushort* ptr, v_uint16x8& a, v_uint16x8& b)
-{
-    __m128i v0 = _mm_loadu_si128((__m128i*)(ptr));     // a0 b0 a1 b1 a2 b2 a3 b3
-    __m128i v1 = _mm_loadu_si128((__m128i*)(ptr + 8)); // a4 b4 a5 b5 a6 b6 a7 b7
-
-    __m128i v2 = _mm_unpacklo_epi16(v0, v1); // a0 a4 b0 b4 a1 a5 b1 b5
-    __m128i v3 = _mm_unpackhi_epi16(v0, v1); // a2 a6 b2 b6 a3 a7 b3 b7
-    __m128i v4 = _mm_unpacklo_epi16(v2, v3); // a0 a2 a4 a6 b0 b2 b4 b6
-    __m128i v5 = _mm_unpackhi_epi16(v2, v3); // a1 a3 a5 a7 b1 b3 b5 b7
-
-    a.val = _mm_unpacklo_epi16(v4, v5); // a0 a1 a2 a3 a4 a5 a6 a7
-    b.val = _mm_unpackhi_epi16(v4, v5); // b0 b1 ab b3 b4 b5 b6 b7
-}
-
-inline void v_load_deinterleave(const ushort* ptr, v_uint16x8& a, v_uint16x8& b, v_uint16x8& c)
-{
-#if CV_SSE4_1
-    __m128i v0 = _mm_loadu_si128((__m128i*)(ptr));
-    __m128i v1 = _mm_loadu_si128((__m128i*)(ptr + 8));
-    __m128i v2 = _mm_loadu_si128((__m128i*)(ptr + 16));
-    __m128i a0 = _mm_blend_epi16(_mm_blend_epi16(v0, v1, 0x92), v2, 0x24);
-    __m128i b0 = _mm_blend_epi16(_mm_blend_epi16(v2, v0, 0x92), v1, 0x24);
-    __m128i c0 = _mm_blend_epi16(_mm_blend_epi16(v1, v2, 0x92), v0, 0x24);
-
-    const __m128i sh_a = _mm_setr_epi8(0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15, 4, 5, 10, 11);
-    const __m128i sh_b = _mm_setr_epi8(2, 3, 8, 9, 14, 15, 4, 5, 10, 11, 0, 1, 6, 7, 12, 13);
-    const __m128i sh_c = _mm_setr_epi8(4, 5, 10, 11, 0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15);
-    a0 = _mm_shuffle_epi8(a0, sh_a);
-    b0 = _mm_shuffle_epi8(b0, sh_b);
-    c0 = _mm_shuffle_epi8(c0, sh_c);
-
-    a.val = a0;
-    b.val = b0;
-    c.val = c0;
-#else
-    __m128i t00 = _mm_loadu_si128((const __m128i*)ptr);
-    __m128i t01 = _mm_loadu_si128((const __m128i*)(ptr + 8));
-    __m128i t02 = _mm_loadu_si128((const __m128i*)(ptr + 16));
-
-    __m128i t10 = _mm_unpacklo_epi16(t00, _mm_unpackhi_epi64(t01, t01));
-    __m128i t11 = _mm_unpacklo_epi16(_mm_unpackhi_epi64(t00, t00), t02);
-    __m128i t12 = _mm_unpacklo_epi16(t01, _mm_unpackhi_epi64(t02, t02));
-
-    __m128i t20 = _mm_unpacklo_epi16(t10, _mm_unpackhi_epi64(t11, t11));
-    __m128i t21 = _mm_unpacklo_epi16(_mm_unpackhi_epi64(t10, t10), t12);
-    __m128i t22 = _mm_unpacklo_epi16(t11, _mm_unpackhi_epi64(t12, t12));
-
-    a.val = _mm_unpacklo_epi16(t20, _mm_unpackhi_epi64(t21, t21));
-    b.val = _mm_unpacklo_epi16(_mm_unpackhi_epi64(t20, t20), t22);
-    c.val = _mm_unpacklo_epi16(t21, _mm_unpackhi_epi64(t22, t22));
-#endif
-}
-
-inline void v_load_deinterleave(const ushort* ptr, v_uint16x8& a, v_uint16x8& b, v_uint16x8& c, v_uint16x8& d)
-{
-    __m128i u0 = _mm_loadu_si128((const __m128i*)ptr); // a0 b0 c0 d0 a1 b1 c1 d1
-    __m128i u1 = _mm_loadu_si128((const __m128i*)(ptr + 8)); // a2 b2 c2 d2 ...
-    __m128i u2 = _mm_loadu_si128((const __m128i*)(ptr + 16)); // a4 b4 c4 d4 ...
-    __m128i u3 = _mm_loadu_si128((const __m128i*)(ptr + 24)); // a6 b6 c6 d6 ...
-
-    __m128i v0 = _mm_unpacklo_epi16(u0, u2); // a0 a4 b0 b4 ...
-    __m128i v1 = _mm_unpackhi_epi16(u0, u2); // a1 a5 b1 b5 ...
-    __m128i v2 = _mm_unpacklo_epi16(u1, u3); // a2 a6 b2 b6 ...
-    __m128i v3 = _mm_unpackhi_epi16(u1, u3); // a3 a7 b3 b7 ...
-
-    u0 = _mm_unpacklo_epi16(v0, v2); // a0 a2 a4 a6 ...
-    u1 = _mm_unpacklo_epi16(v1, v3); // a1 a3 a5 a7 ...
-    u2 = _mm_unpackhi_epi16(v0, v2); // c0 c2 c4 c6 ...
-    u3 = _mm_unpackhi_epi16(v1, v3); // c1 c3 c5 c7 ...
-
-    a.val = _mm_unpacklo_epi16(u0, u1);
-    b.val = _mm_unpackhi_epi16(u0, u1);
-    c.val = _mm_unpacklo_epi16(u2, u3);
-    d.val = _mm_unpackhi_epi16(u2, u3);
-}
-
-inline void v_load_deinterleave(const unsigned* ptr, v_uint32x4& a, v_uint32x4& b)
-{
-    __m128i v0 = _mm_loadu_si128((__m128i*)(ptr));     // a0 b0 a1 b1
-    __m128i v1 = _mm_loadu_si128((__m128i*)(ptr + 4)); // a2 b2 a3 b3
-
-    __m128i v2 = _mm_unpacklo_epi32(v0, v1); // a0 a2 b0 b2
-    __m128i v3 = _mm_unpackhi_epi32(v0, v1); // a1 a3 b1 b3
-
-    a.val = _mm_unpacklo_epi32(v2, v3); // a0 a1 a2 a3
-    b.val = _mm_unpackhi_epi32(v2, v3); // b0 b1 ab b3
-}
-
-inline void v_load_deinterleave(const unsigned* ptr, v_uint32x4& a, v_uint32x4& b, v_uint32x4& c)
-{
-    __m128i t00 = _mm_loadu_si128((const __m128i*)ptr);
-    __m128i t01 = _mm_loadu_si128((const __m128i*)(ptr + 4));
-    __m128i t02 = _mm_loadu_si128((const __m128i*)(ptr + 8));
-
-    __m128i t10 = _mm_unpacklo_epi32(t00, _mm_unpackhi_epi64(t01, t01));
-    __m128i t11 = _mm_unpacklo_epi32(_mm_unpackhi_epi64(t00, t00), t02);
-    __m128i t12 = _mm_unpacklo_epi32(t01, _mm_unpackhi_epi64(t02, t02));
-
-    a.val = _mm_unpacklo_epi32(t10, _mm_unpackhi_epi64(t11, t11));
-    b.val = _mm_unpacklo_epi32(_mm_unpackhi_epi64(t10, t10), t12);
-    c.val = _mm_unpacklo_epi32(t11, _mm_unpackhi_epi64(t12, t12));
-}
-
-inline void v_load_deinterleave(const unsigned* ptr, v_uint32x4& a, v_uint32x4& b, v_uint32x4& c, v_uint32x4& d)
-{
-    v_uint32x4 s0(_mm_loadu_si128((const __m128i*)ptr));        // a0 b0 c0 d0
-    v_uint32x4 s1(_mm_loadu_si128((const __m128i*)(ptr + 4)));  // a1 b1 c1 d1
-    v_uint32x4 s2(_mm_loadu_si128((const __m128i*)(ptr + 8)));  // a2 b2 c2 d2
-    v_uint32x4 s3(_mm_loadu_si128((const __m128i*)(ptr + 12))); // a3 b3 c3 d3
-
-    v_transpose4x4(s0, s1, s2, s3, a, b, c, d);
-}
-
-inline void v_load_deinterleave(const float* ptr, v_float32x4& a, v_float32x4& b)
-{
-    __m128 u0 = _mm_loadu_ps(ptr);       // a0 b0 a1 b1
-    __m128 u1 = _mm_loadu_ps((ptr + 4)); // a2 b2 a3 b3
-
-    a.val = _mm_shuffle_ps(u0, u1, _MM_SHUFFLE(2, 0, 2, 0)); // a0 a1 a2 a3
-    b.val = _mm_shuffle_ps(u0, u1, _MM_SHUFFLE(3, 1, 3, 1)); // b0 b1 ab b3
-}
-
-inline void v_load_deinterleave(const float* ptr, v_float32x4& a, v_float32x4& b, v_float32x4& c)
-{
-    __m128 t0 = _mm_loadu_ps(ptr + 0);
-    __m128 t1 = _mm_loadu_ps(ptr + 4);
-    __m128 t2 = _mm_loadu_ps(ptr + 8);
-
-    __m128 at12 = _mm_shuffle_ps(t1, t2, _MM_SHUFFLE(0, 1, 0, 2));
-    a.val = _mm_shuffle_ps(t0, at12, _MM_SHUFFLE(2, 0, 3, 0));
-
-    __m128 bt01 = _mm_shuffle_ps(t0, t1, _MM_SHUFFLE(0, 0, 0, 1));
-    __m128 bt12 = _mm_shuffle_ps(t1, t2, _MM_SHUFFLE(0, 2, 0, 3));
-    b.val = _mm_shuffle_ps(bt01, bt12, _MM_SHUFFLE(2, 0, 2, 0));
-
-    __m128 ct01 = _mm_shuffle_ps(t0, t1, _MM_SHUFFLE(0, 1, 0, 2));
-    c.val = _mm_shuffle_ps(ct01, t2, _MM_SHUFFLE(3, 0, 2, 0));
-}
-
-inline void v_load_deinterleave(const float* ptr, v_float32x4& a, v_float32x4& b, v_float32x4& c, v_float32x4& d)
-{
-    __m128 t0 = _mm_loadu_ps(ptr +  0);
-    __m128 t1 = _mm_loadu_ps(ptr +  4);
-    __m128 t2 = _mm_loadu_ps(ptr +  8);
-    __m128 t3 = _mm_loadu_ps(ptr + 12);
-    __m128 t02lo = _mm_unpacklo_ps(t0, t2);
-    __m128 t13lo = _mm_unpacklo_ps(t1, t3);
-    __m128 t02hi = _mm_unpackhi_ps(t0, t2);
-    __m128 t13hi = _mm_unpackhi_ps(t1, t3);
-    a.val = _mm_unpacklo_ps(t02lo, t13lo);
-    b.val = _mm_unpackhi_ps(t02lo, t13lo);
-    c.val = _mm_unpacklo_ps(t02hi, t13hi);
-    d.val = _mm_unpackhi_ps(t02hi, t13hi);
-}
-
-inline void v_load_deinterleave(const uint64 *ptr, v_uint64x2& a, v_uint64x2& b)
-{
-    __m128i t0 = _mm_loadu_si128((const __m128i*)ptr);
-    __m128i t1 = _mm_loadu_si128((const __m128i*)(ptr + 2));
-
-    a = v_uint64x2(_mm_unpacklo_epi64(t0, t1));
-    b = v_uint64x2(_mm_unpackhi_epi64(t0, t1));
-}
-
-inline void v_load_deinterleave(const uint64 *ptr, v_uint64x2& a, v_uint64x2& b, v_uint64x2& c)
-{
-    __m128i t0 = _mm_loadu_si128((const __m128i*)ptr); // a0, b0
-    __m128i t1 = _mm_loadu_si128((const __m128i*)(ptr + 2)); // c0, a1
-    __m128i t2 = _mm_loadu_si128((const __m128i*)(ptr + 4)); // b1, c1
-
-    t1 = _mm_shuffle_epi32(t1, 0x4e); // a1, c0
-
-    a = v_uint64x2(_mm_unpacklo_epi64(t0, t1));
-    b = v_uint64x2(_mm_unpacklo_epi64(_mm_unpackhi_epi64(t0, t0), t2));
-    c = v_uint64x2(_mm_unpackhi_epi64(t1, t2));
-}
-
-inline void v_load_deinterleave(const uint64 *ptr, v_uint64x2& a,
-                                v_uint64x2& b, v_uint64x2& c, v_uint64x2& d)
-{
-    __m128i t0 = _mm_loadu_si128((const __m128i*)ptr); // a0 b0
-    __m128i t1 = _mm_loadu_si128((const __m128i*)(ptr + 2)); // c0 d0
-    __m128i t2 = _mm_loadu_si128((const __m128i*)(ptr + 4)); // a1 b1
-    __m128i t3 = _mm_loadu_si128((const __m128i*)(ptr + 6)); // c1 d1
-
-    a = v_uint64x2(_mm_unpacklo_epi64(t0, t2));
-    b = v_uint64x2(_mm_unpackhi_epi64(t0, t2));
-    c = v_uint64x2(_mm_unpacklo_epi64(t1, t3));
-    d = v_uint64x2(_mm_unpackhi_epi64(t1, t3));
-}
-
-// store interleave
-
-inline void v_store_interleave( uchar* ptr, const v_uint8x16& a, const v_uint8x16& b,
-                                hal::StoreMode mode = hal::STORE_UNALIGNED)
-{
-    __m128i v0 = _mm_unpacklo_epi8(a.val, b.val);
-    __m128i v1 = _mm_unpackhi_epi8(a.val, b.val);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm_stream_si128((__m128i*)(ptr), v0);
-        _mm_stream_si128((__m128i*)(ptr + 16), v1);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm_store_si128((__m128i*)(ptr), v0);
-        _mm_store_si128((__m128i*)(ptr + 16), v1);
-    }
-    else
-    {
-        _mm_storeu_si128((__m128i*)(ptr), v0);
-        _mm_storeu_si128((__m128i*)(ptr + 16), v1);
-    }
-}
-
-inline void v_store_interleave( uchar* ptr, const v_uint8x16& a, const v_uint8x16& b,
-                                const v_uint8x16& c, hal::StoreMode mode = hal::STORE_UNALIGNED)
-{
-#if CV_SSE4_1
-    const __m128i sh_a = _mm_setr_epi8(0, 11, 6, 1, 12, 7, 2, 13, 8, 3, 14, 9, 4, 15, 10, 5);
-    const __m128i sh_b = _mm_setr_epi8(5, 0, 11, 6, 1, 12, 7, 2, 13, 8, 3, 14, 9, 4, 15, 10);
-    const __m128i sh_c = _mm_setr_epi8(10, 5, 0, 11, 6, 1, 12, 7, 2, 13, 8, 3, 14, 9, 4, 15);
-    __m128i a0 = _mm_shuffle_epi8(a.val, sh_a);
-    __m128i b0 = _mm_shuffle_epi8(b.val, sh_b);
-    __m128i c0 = _mm_shuffle_epi8(c.val, sh_c);
-
-    const __m128i m0 = _mm_setr_epi8(0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0);
-    const __m128i m1 = _mm_setr_epi8(0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0, -1, 0, 0);
-    __m128i v0 = _mm_blendv_epi8(_mm_blendv_epi8(a0, b0, m1), c0, m0);
-    __m128i v1 = _mm_blendv_epi8(_mm_blendv_epi8(b0, c0, m1), a0, m0);
-    __m128i v2 = _mm_blendv_epi8(_mm_blendv_epi8(c0, a0, m1), b0, m0);
-#elif CV_SSSE3
-    const __m128i m0 = _mm_setr_epi8(0, 6, 11, 1, 7, 12, 2, 8, 13, 3, 9, 14, 4, 10, 15, 5);
-    const __m128i m1 = _mm_setr_epi8(5, 11, 0, 6, 12, 1, 7, 13, 2, 8, 14, 3, 9, 15, 4, 10);
-    const __m128i m2 = _mm_setr_epi8(10, 0, 5, 11, 1, 6, 12, 2, 7, 13, 3, 8, 14, 4, 9, 15);
-
-    __m128i t0 = _mm_alignr_epi8(b.val, _mm_slli_si128(a.val, 10), 5);
-    t0 = _mm_alignr_epi8(c.val, t0, 5);
-    __m128i v0 = _mm_shuffle_epi8(t0, m0);
-
-    __m128i t1 = _mm_alignr_epi8(_mm_srli_si128(b.val, 5), _mm_slli_si128(a.val, 5), 6);
-    t1 = _mm_alignr_epi8(_mm_srli_si128(c.val, 5), t1, 5);
-    __m128i v1 = _mm_shuffle_epi8(t1, m1);
-
-    __m128i t2 = _mm_alignr_epi8(_mm_srli_si128(c.val, 10), b.val, 11);
-    t2 = _mm_alignr_epi8(t2, a.val, 11);
-    __m128i v2 = _mm_shuffle_epi8(t2, m2);
-#else
-    __m128i z = _mm_setzero_si128();
-    __m128i ab0 = _mm_unpacklo_epi8(a.val, b.val);
-    __m128i ab1 = _mm_unpackhi_epi8(a.val, b.val);
-    __m128i c0 = _mm_unpacklo_epi8(c.val, z);
-    __m128i c1 = _mm_unpackhi_epi8(c.val, z);
-
-    __m128i p00 = _mm_unpacklo_epi16(ab0, c0);
-    __m128i p01 = _mm_unpackhi_epi16(ab0, c0);
-    __m128i p02 = _mm_unpacklo_epi16(ab1, c1);
-    __m128i p03 = _mm_unpackhi_epi16(ab1, c1);
-
-    __m128i p10 = _mm_unpacklo_epi32(p00, p01);
-    __m128i p11 = _mm_unpackhi_epi32(p00, p01);
-    __m128i p12 = _mm_unpacklo_epi32(p02, p03);
-    __m128i p13 = _mm_unpackhi_epi32(p02, p03);
-
-    __m128i p20 = _mm_unpacklo_epi64(p10, p11);
-    __m128i p21 = _mm_unpackhi_epi64(p10, p11);
-    __m128i p22 = _mm_unpacklo_epi64(p12, p13);
-    __m128i p23 = _mm_unpackhi_epi64(p12, p13);
-
-    p20 = _mm_slli_si128(p20, 1);
-    p22 = _mm_slli_si128(p22, 1);
-
-    __m128i p30 = _mm_slli_epi64(_mm_unpacklo_epi32(p20, p21), 8);
-    __m128i p31 = _mm_srli_epi64(_mm_unpackhi_epi32(p20, p21), 8);
-    __m128i p32 = _mm_slli_epi64(_mm_unpacklo_epi32(p22, p23), 8);
-    __m128i p33 = _mm_srli_epi64(_mm_unpackhi_epi32(p22, p23), 8);
-
-    __m128i p40 = _mm_unpacklo_epi64(p30, p31);
-    __m128i p41 = _mm_unpackhi_epi64(p30, p31);
-    __m128i p42 = _mm_unpacklo_epi64(p32, p33);
-    __m128i p43 = _mm_unpackhi_epi64(p32, p33);
-
-    __m128i v0 = _mm_or_si128(_mm_srli_si128(p40, 2), _mm_slli_si128(p41, 10));
-    __m128i v1 = _mm_or_si128(_mm_srli_si128(p41, 6), _mm_slli_si128(p42, 6));
-    __m128i v2 = _mm_or_si128(_mm_srli_si128(p42, 10), _mm_slli_si128(p43, 2));
-#endif
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm_stream_si128((__m128i*)(ptr), v0);
-        _mm_stream_si128((__m128i*)(ptr + 16), v1);
-        _mm_stream_si128((__m128i*)(ptr + 32), v2);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm_store_si128((__m128i*)(ptr), v0);
-        _mm_store_si128((__m128i*)(ptr + 16), v1);
-        _mm_store_si128((__m128i*)(ptr + 32), v2);
-    }
-    else
-    {
-        _mm_storeu_si128((__m128i*)(ptr), v0);
-        _mm_storeu_si128((__m128i*)(ptr + 16), v1);
-        _mm_storeu_si128((__m128i*)(ptr + 32), v2);
-    }
-}
-
-inline void v_store_interleave( uchar* ptr, const v_uint8x16& a, const v_uint8x16& b,
-                                const v_uint8x16& c, const v_uint8x16& d,
-                                hal::StoreMode mode = hal::STORE_UNALIGNED)
-{
-    // a0 a1 a2 a3 ....
-    // b0 b1 b2 b3 ....
-    // c0 c1 c2 c3 ....
-    // d0 d1 d2 d3 ....
-    __m128i u0 = _mm_unpacklo_epi8(a.val, c.val); // a0 c0 a1 c1 ...
-    __m128i u1 = _mm_unpackhi_epi8(a.val, c.val); // a8 c8 a9 c9 ...
-    __m128i u2 = _mm_unpacklo_epi8(b.val, d.val); // b0 d0 b1 d1 ...
-    __m128i u3 = _mm_unpackhi_epi8(b.val, d.val); // b8 d8 b9 d9 ...
-
-    __m128i v0 = _mm_unpacklo_epi8(u0, u2); // a0 b0 c0 d0 ...
-    __m128i v1 = _mm_unpackhi_epi8(u0, u2); // a4 b4 c4 d4 ...
-    __m128i v2 = _mm_unpacklo_epi8(u1, u3); // a8 b8 c8 d8 ...
-    __m128i v3 = _mm_unpackhi_epi8(u1, u3); // a12 b12 c12 d12 ...
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm_stream_si128((__m128i*)(ptr), v0);
-        _mm_stream_si128((__m128i*)(ptr + 16), v1);
-        _mm_stream_si128((__m128i*)(ptr + 32), v2);
-        _mm_stream_si128((__m128i*)(ptr + 48), v3);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm_store_si128((__m128i*)(ptr), v0);
-        _mm_store_si128((__m128i*)(ptr + 16), v1);
-        _mm_store_si128((__m128i*)(ptr + 32), v2);
-        _mm_store_si128((__m128i*)(ptr + 48), v3);
-    }
-    else
-    {
-        _mm_storeu_si128((__m128i*)(ptr), v0);
-        _mm_storeu_si128((__m128i*)(ptr + 16), v1);
-        _mm_storeu_si128((__m128i*)(ptr + 32), v2);
-        _mm_storeu_si128((__m128i*)(ptr + 48), v3);
-    }
-}
-
-inline void v_store_interleave( ushort* ptr, const v_uint16x8& a, const v_uint16x8& b,
-                                hal::StoreMode mode = hal::STORE_UNALIGNED)
-{
-    __m128i v0 = _mm_unpacklo_epi16(a.val, b.val);
-    __m128i v1 = _mm_unpackhi_epi16(a.val, b.val);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm_stream_si128((__m128i*)(ptr), v0);
-        _mm_stream_si128((__m128i*)(ptr + 8), v1);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm_store_si128((__m128i*)(ptr), v0);
-        _mm_store_si128((__m128i*)(ptr + 8), v1);
-    }
-    else
-    {
-        _mm_storeu_si128((__m128i*)(ptr), v0);
-        _mm_storeu_si128((__m128i*)(ptr + 8), v1);
-    }
-}
-
-inline void v_store_interleave( ushort* ptr, const v_uint16x8& a,
-                                const v_uint16x8& b, const v_uint16x8& c,
-                                hal::StoreMode mode = hal::STORE_UNALIGNED)
-{
-#if CV_SSE4_1
-    const __m128i sh_a = _mm_setr_epi8(0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15, 4, 5, 10, 11);
-    const __m128i sh_b = _mm_setr_epi8(10, 11, 0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15, 4, 5);
-    const __m128i sh_c = _mm_setr_epi8(4, 5, 10, 11, 0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15);
-    __m128i a0 = _mm_shuffle_epi8(a.val, sh_a);
-    __m128i b0 = _mm_shuffle_epi8(b.val, sh_b);
-    __m128i c0 = _mm_shuffle_epi8(c.val, sh_c);
-
-    __m128i v0 = _mm_blend_epi16(_mm_blend_epi16(a0, b0, 0x92), c0, 0x24);
-    __m128i v1 = _mm_blend_epi16(_mm_blend_epi16(c0, a0, 0x92), b0, 0x24);
-    __m128i v2 = _mm_blend_epi16(_mm_blend_epi16(b0, c0, 0x92), a0, 0x24);
-#else
-    __m128i z = _mm_setzero_si128();
-    __m128i ab0 = _mm_unpacklo_epi16(a.val, b.val);
-    __m128i ab1 = _mm_unpackhi_epi16(a.val, b.val);
-    __m128i c0 = _mm_unpacklo_epi16(c.val, z);
-    __m128i c1 = _mm_unpackhi_epi16(c.val, z);
-
-    __m128i p10 = _mm_unpacklo_epi32(ab0, c0);
-    __m128i p11 = _mm_unpackhi_epi32(ab0, c0);
-    __m128i p12 = _mm_unpacklo_epi32(ab1, c1);
-    __m128i p13 = _mm_unpackhi_epi32(ab1, c1);
-
-    __m128i p20 = _mm_unpacklo_epi64(p10, p11);
-    __m128i p21 = _mm_unpackhi_epi64(p10, p11);
-    __m128i p22 = _mm_unpacklo_epi64(p12, p13);
-    __m128i p23 = _mm_unpackhi_epi64(p12, p13);
-
-    p20 = _mm_slli_si128(p20, 2);
-    p22 = _mm_slli_si128(p22, 2);
-
-    __m128i p30 = _mm_unpacklo_epi64(p20, p21);
-    __m128i p31 = _mm_unpackhi_epi64(p20, p21);
-    __m128i p32 = _mm_unpacklo_epi64(p22, p23);
-    __m128i p33 = _mm_unpackhi_epi64(p22, p23);
-
-    __m128i v0 = _mm_or_si128(_mm_srli_si128(p30, 2), _mm_slli_si128(p31, 10));
-    __m128i v1 = _mm_or_si128(_mm_srli_si128(p31, 6), _mm_slli_si128(p32, 6));
-    __m128i v2 = _mm_or_si128(_mm_srli_si128(p32, 10), _mm_slli_si128(p33, 2));
-#endif
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm_stream_si128((__m128i*)(ptr), v0);
-        _mm_stream_si128((__m128i*)(ptr + 8), v1);
-        _mm_stream_si128((__m128i*)(ptr + 16), v2);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm_store_si128((__m128i*)(ptr), v0);
-        _mm_store_si128((__m128i*)(ptr + 8), v1);
-        _mm_store_si128((__m128i*)(ptr + 16), v2);
-    }
-    else
-    {
-        _mm_storeu_si128((__m128i*)(ptr), v0);
-        _mm_storeu_si128((__m128i*)(ptr + 8), v1);
-        _mm_storeu_si128((__m128i*)(ptr + 16), v2);
-    }
-}
-
-inline void v_store_interleave( ushort* ptr, const v_uint16x8& a, const v_uint16x8& b,
-                                const v_uint16x8& c, const v_uint16x8& d,
-                                hal::StoreMode mode = hal::STORE_UNALIGNED)
-{
-    // a0 a1 a2 a3 ....
-    // b0 b1 b2 b3 ....
-    // c0 c1 c2 c3 ....
-    // d0 d1 d2 d3 ....
-    __m128i u0 = _mm_unpacklo_epi16(a.val, c.val); // a0 c0 a1 c1 ...
-    __m128i u1 = _mm_unpackhi_epi16(a.val, c.val); // a4 c4 a5 c5 ...
-    __m128i u2 = _mm_unpacklo_epi16(b.val, d.val); // b0 d0 b1 d1 ...
-    __m128i u3 = _mm_unpackhi_epi16(b.val, d.val); // b4 d4 b5 d5 ...
-
-    __m128i v0 = _mm_unpacklo_epi16(u0, u2); // a0 b0 c0 d0 ...
-    __m128i v1 = _mm_unpackhi_epi16(u0, u2); // a2 b2 c2 d2 ...
-    __m128i v2 = _mm_unpacklo_epi16(u1, u3); // a4 b4 c4 d4 ...
-    __m128i v3 = _mm_unpackhi_epi16(u1, u3); // a6 b6 c6 d6 ...
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm_stream_si128((__m128i*)(ptr), v0);
-        _mm_stream_si128((__m128i*)(ptr + 8), v1);
-        _mm_stream_si128((__m128i*)(ptr + 16), v2);
-        _mm_stream_si128((__m128i*)(ptr + 24), v3);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm_store_si128((__m128i*)(ptr), v0);
-        _mm_store_si128((__m128i*)(ptr + 8), v1);
-        _mm_store_si128((__m128i*)(ptr + 16), v2);
-        _mm_store_si128((__m128i*)(ptr + 24), v3);
-    }
-    else
-    {
-        _mm_storeu_si128((__m128i*)(ptr), v0);
-        _mm_storeu_si128((__m128i*)(ptr + 8), v1);
-        _mm_storeu_si128((__m128i*)(ptr + 16), v2);
-        _mm_storeu_si128((__m128i*)(ptr + 24), v3);
-    }
-}
-
-inline void v_store_interleave( unsigned* ptr, const v_uint32x4& a, const v_uint32x4& b,
-                                hal::StoreMode mode = hal::STORE_UNALIGNED)
-{
-    __m128i v0 = _mm_unpacklo_epi32(a.val, b.val);
-    __m128i v1 = _mm_unpackhi_epi32(a.val, b.val);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm_stream_si128((__m128i*)(ptr), v0);
-        _mm_stream_si128((__m128i*)(ptr + 4), v1);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm_store_si128((__m128i*)(ptr), v0);
-        _mm_store_si128((__m128i*)(ptr + 4), v1);
-    }
-    else
-    {
-        _mm_storeu_si128((__m128i*)(ptr), v0);
-        _mm_storeu_si128((__m128i*)(ptr + 4), v1);
-    }
-}
-
-inline void v_store_interleave( unsigned* ptr, const v_uint32x4& a, const v_uint32x4& b,
-                                const v_uint32x4& c, hal::StoreMode mode = hal::STORE_UNALIGNED)
-{
-    v_uint32x4 z = v_setzero_u32(), u0, u1, u2, u3;
-    v_transpose4x4(a, b, c, z, u0, u1, u2, u3);
-
-    __m128i v0 = _mm_or_si128(u0.val, _mm_slli_si128(u1.val, 12));
-    __m128i v1 = _mm_or_si128(_mm_srli_si128(u1.val, 4), _mm_slli_si128(u2.val, 8));
-    __m128i v2 = _mm_or_si128(_mm_srli_si128(u2.val, 8), _mm_slli_si128(u3.val, 4));
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm_stream_si128((__m128i*)(ptr), v0);
-        _mm_stream_si128((__m128i*)(ptr + 4), v1);
-        _mm_stream_si128((__m128i*)(ptr + 8), v2);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm_store_si128((__m128i*)(ptr), v0);
-        _mm_store_si128((__m128i*)(ptr + 4), v1);
-        _mm_store_si128((__m128i*)(ptr + 8), v2);
-    }
-    else
-    {
-        _mm_storeu_si128((__m128i*)(ptr), v0);
-        _mm_storeu_si128((__m128i*)(ptr + 4), v1);
-        _mm_storeu_si128((__m128i*)(ptr + 8), v2);
-    }
-}
-
-inline void v_store_interleave(unsigned* ptr, const v_uint32x4& a, const v_uint32x4& b,
-                               const v_uint32x4& c, const v_uint32x4& d,
-                               hal::StoreMode mode = hal::STORE_UNALIGNED)
-{
-    v_uint32x4 v0, v1, v2, v3;
-    v_transpose4x4(a, b, c, d, v0, v1, v2, v3);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm_stream_si128((__m128i*)(ptr), v0.val);
-        _mm_stream_si128((__m128i*)(ptr + 4), v1.val);
-        _mm_stream_si128((__m128i*)(ptr + 8), v2.val);
-        _mm_stream_si128((__m128i*)(ptr + 12), v3.val);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm_store_si128((__m128i*)(ptr), v0.val);
-        _mm_store_si128((__m128i*)(ptr + 4), v1.val);
-        _mm_store_si128((__m128i*)(ptr + 8), v2.val);
-        _mm_store_si128((__m128i*)(ptr + 12), v3.val);
-    }
-    else
-    {
-        _mm_storeu_si128((__m128i*)(ptr), v0.val);
-        _mm_storeu_si128((__m128i*)(ptr + 4), v1.val);
-        _mm_storeu_si128((__m128i*)(ptr + 8), v2.val);
-        _mm_storeu_si128((__m128i*)(ptr + 12), v3.val);
-    }
-}
-
-// 2-channel, float only
-inline void v_store_interleave(float* ptr, const v_float32x4& a, const v_float32x4& b,
-                               hal::StoreMode mode = hal::STORE_UNALIGNED)
-{
-    __m128 v0 = _mm_unpacklo_ps(a.val, b.val); // a0 b0 a1 b1
-    __m128 v1 = _mm_unpackhi_ps(a.val, b.val); // a2 b2 a3 b3
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm_stream_ps(ptr, v0);
-        _mm_stream_ps(ptr + 4, v1);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm_store_ps(ptr, v0);
-        _mm_store_ps(ptr + 4, v1);
-    }
-    else
-    {
-        _mm_storeu_ps(ptr, v0);
-        _mm_storeu_ps(ptr + 4, v1);
-    }
-}
-
-inline void v_store_interleave(float* ptr, const v_float32x4& a, const v_float32x4& b,
-                               const v_float32x4& c, hal::StoreMode mode = hal::STORE_UNALIGNED)
-{
-    __m128 u0 = _mm_shuffle_ps(a.val, b.val, _MM_SHUFFLE(0, 0, 0, 0));
-    __m128 u1 = _mm_shuffle_ps(c.val, a.val, _MM_SHUFFLE(1, 1, 0, 0));
-    __m128 v0 = _mm_shuffle_ps(u0, u1, _MM_SHUFFLE(2, 0, 2, 0));
-    __m128 u2 = _mm_shuffle_ps(b.val, c.val, _MM_SHUFFLE(1, 1, 1, 1));
-    __m128 u3 = _mm_shuffle_ps(a.val, b.val, _MM_SHUFFLE(2, 2, 2, 2));
-    __m128 v1 = _mm_shuffle_ps(u2, u3, _MM_SHUFFLE(2, 0, 2, 0));
-    __m128 u4 = _mm_shuffle_ps(c.val, a.val, _MM_SHUFFLE(3, 3, 2, 2));
-    __m128 u5 = _mm_shuffle_ps(b.val, c.val, _MM_SHUFFLE(3, 3, 3, 3));
-    __m128 v2 = _mm_shuffle_ps(u4, u5, _MM_SHUFFLE(2, 0, 2, 0));
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm_stream_ps(ptr, v0);
-        _mm_stream_ps(ptr + 4, v1);
-        _mm_stream_ps(ptr + 8, v2);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm_store_ps(ptr, v0);
-        _mm_store_ps(ptr + 4, v1);
-        _mm_store_ps(ptr + 8, v2);
-    }
-    else
-    {
-        _mm_storeu_ps(ptr, v0);
-        _mm_storeu_ps(ptr + 4, v1);
-        _mm_storeu_ps(ptr + 8, v2);
-    }
-}
-
-inline void v_store_interleave(float* ptr, const v_float32x4& a, const v_float32x4& b,
-                               const v_float32x4& c, const v_float32x4& d,
-                               hal::StoreMode mode = hal::STORE_UNALIGNED)
-{
-    __m128 u0 = _mm_unpacklo_ps(a.val, c.val);
-    __m128 u1 = _mm_unpacklo_ps(b.val, d.val);
-    __m128 u2 = _mm_unpackhi_ps(a.val, c.val);
-    __m128 u3 = _mm_unpackhi_ps(b.val, d.val);
-    __m128 v0 = _mm_unpacklo_ps(u0, u1);
-    __m128 v2 = _mm_unpacklo_ps(u2, u3);
-    __m128 v1 = _mm_unpackhi_ps(u0, u1);
-    __m128 v3 = _mm_unpackhi_ps(u2, u3);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm_stream_ps(ptr, v0);
-        _mm_stream_ps(ptr + 4, v1);
-        _mm_stream_ps(ptr + 8, v2);
-        _mm_stream_ps(ptr + 12, v3);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm_store_ps(ptr, v0);
-        _mm_store_ps(ptr + 4, v1);
-        _mm_store_ps(ptr + 8, v2);
-        _mm_store_ps(ptr + 12, v3);
-    }
-    else
-    {
-        _mm_storeu_ps(ptr, v0);
-        _mm_storeu_ps(ptr + 4, v1);
-        _mm_storeu_ps(ptr + 8, v2);
-        _mm_storeu_ps(ptr + 12, v3);
-    }
-}
-
-inline void v_store_interleave(uint64 *ptr, const v_uint64x2& a, const v_uint64x2& b,
-                               hal::StoreMode mode = hal::STORE_UNALIGNED)
-{
-    __m128i v0 = _mm_unpacklo_epi64(a.val, b.val);
-    __m128i v1 = _mm_unpackhi_epi64(a.val, b.val);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm_stream_si128((__m128i*)(ptr), v0);
-        _mm_stream_si128((__m128i*)(ptr + 2), v1);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm_store_si128((__m128i*)(ptr), v0);
-        _mm_store_si128((__m128i*)(ptr + 2), v1);
-    }
-    else
-    {
-        _mm_storeu_si128((__m128i*)(ptr), v0);
-        _mm_storeu_si128((__m128i*)(ptr + 2), v1);
-    }
-}
-
-inline void v_store_interleave(uint64 *ptr, const v_uint64x2& a, const v_uint64x2& b,
-                               const v_uint64x2& c, hal::StoreMode mode = hal::STORE_UNALIGNED)
-{
-    __m128i v0 = _mm_unpacklo_epi64(a.val, b.val);
-    __m128i v1 = _mm_unpacklo_epi64(c.val, _mm_unpackhi_epi64(a.val, a.val));
-    __m128i v2 = _mm_unpackhi_epi64(b.val, c.val);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm_stream_si128((__m128i*)(ptr), v0);
-        _mm_stream_si128((__m128i*)(ptr + 2), v1);
-        _mm_stream_si128((__m128i*)(ptr + 4), v2);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm_store_si128((__m128i*)(ptr), v0);
-        _mm_store_si128((__m128i*)(ptr + 2), v1);
-        _mm_store_si128((__m128i*)(ptr + 4), v2);
-    }
-    else
-    {
-        _mm_storeu_si128((__m128i*)(ptr), v0);
-        _mm_storeu_si128((__m128i*)(ptr + 2), v1);
-        _mm_storeu_si128((__m128i*)(ptr + 4), v2);
-    }
-}
-
-inline void v_store_interleave(uint64 *ptr, const v_uint64x2& a, const v_uint64x2& b,
-                               const v_uint64x2& c, const v_uint64x2& d,
-                               hal::StoreMode mode = hal::STORE_UNALIGNED)
-{
-    __m128i v0 = _mm_unpacklo_epi64(a.val, b.val);
-    __m128i v1 = _mm_unpacklo_epi64(c.val, d.val);
-    __m128i v2 = _mm_unpackhi_epi64(a.val, b.val);
-    __m128i v3 = _mm_unpackhi_epi64(c.val, d.val);
-
-    if( mode == hal::STORE_ALIGNED_NOCACHE )
-    {
-        _mm_stream_si128((__m128i*)(ptr), v0);
-        _mm_stream_si128((__m128i*)(ptr + 2), v1);
-        _mm_stream_si128((__m128i*)(ptr + 4), v2);
-        _mm_stream_si128((__m128i*)(ptr + 6), v3);
-    }
-    else if( mode == hal::STORE_ALIGNED )
-    {
-        _mm_store_si128((__m128i*)(ptr), v0);
-        _mm_store_si128((__m128i*)(ptr + 2), v1);
-        _mm_store_si128((__m128i*)(ptr + 4), v2);
-        _mm_store_si128((__m128i*)(ptr + 6), v3);
-    }
-    else
-    {
-        _mm_storeu_si128((__m128i*)(ptr), v0);
-        _mm_storeu_si128((__m128i*)(ptr + 2), v1);
-        _mm_storeu_si128((__m128i*)(ptr + 4), v2);
-        _mm_storeu_si128((__m128i*)(ptr + 6), v3);
-    }
-}
-
-#define OPENCV_HAL_IMPL_SSE_LOADSTORE_INTERLEAVE(_Tpvec0, _Tp0, suffix0, _Tpvec1, _Tp1, suffix1) \
-inline void v_load_deinterleave( const _Tp0* ptr, _Tpvec0& a0, _Tpvec0& b0 ) \
-{ \
-    _Tpvec1 a1, b1; \
-    v_load_deinterleave((const _Tp1*)ptr, a1, b1); \
-    a0 = v_reinterpret_as_##suffix0(a1); \
-    b0 = v_reinterpret_as_##suffix0(b1); \
-} \
-inline void v_load_deinterleave( const _Tp0* ptr, _Tpvec0& a0, _Tpvec0& b0, _Tpvec0& c0 ) \
-{ \
-    _Tpvec1 a1, b1, c1; \
-    v_load_deinterleave((const _Tp1*)ptr, a1, b1, c1); \
-    a0 = v_reinterpret_as_##suffix0(a1); \
-    b0 = v_reinterpret_as_##suffix0(b1); \
-    c0 = v_reinterpret_as_##suffix0(c1); \
-} \
-inline void v_load_deinterleave( const _Tp0* ptr, _Tpvec0& a0, _Tpvec0& b0, _Tpvec0& c0, _Tpvec0& d0 ) \
-{ \
-    _Tpvec1 a1, b1, c1, d1; \
-    v_load_deinterleave((const _Tp1*)ptr, a1, b1, c1, d1); \
-    a0 = v_reinterpret_as_##suffix0(a1); \
-    b0 = v_reinterpret_as_##suffix0(b1); \
-    c0 = v_reinterpret_as_##suffix0(c1); \
-    d0 = v_reinterpret_as_##suffix0(d1); \
-} \
-inline void v_store_interleave( _Tp0* ptr, const _Tpvec0& a0, const _Tpvec0& b0, \
-                                hal::StoreMode mode = hal::STORE_UNALIGNED ) \
-{ \
-    _Tpvec1 a1 = v_reinterpret_as_##suffix1(a0); \
-    _Tpvec1 b1 = v_reinterpret_as_##suffix1(b0); \
-    v_store_interleave((_Tp1*)ptr, a1, b1, mode);      \
-} \
-inline void v_store_interleave( _Tp0* ptr, const _Tpvec0& a0, const _Tpvec0& b0, \
-                                const _Tpvec0& c0, hal::StoreMode mode = hal::STORE_UNALIGNED ) \
-{ \
-    _Tpvec1 a1 = v_reinterpret_as_##suffix1(a0); \
-    _Tpvec1 b1 = v_reinterpret_as_##suffix1(b0); \
-    _Tpvec1 c1 = v_reinterpret_as_##suffix1(c0); \
-    v_store_interleave((_Tp1*)ptr, a1, b1, c1, mode);  \
-} \
-inline void v_store_interleave( _Tp0* ptr, const _Tpvec0& a0, const _Tpvec0& b0, \
-                                const _Tpvec0& c0, const _Tpvec0& d0, \
-                                hal::StoreMode mode = hal::STORE_UNALIGNED ) \
-{ \
-    _Tpvec1 a1 = v_reinterpret_as_##suffix1(a0); \
-    _Tpvec1 b1 = v_reinterpret_as_##suffix1(b0); \
-    _Tpvec1 c1 = v_reinterpret_as_##suffix1(c0); \
-    _Tpvec1 d1 = v_reinterpret_as_##suffix1(d0); \
-    v_store_interleave((_Tp1*)ptr, a1, b1, c1, d1, mode); \
-}
-
-OPENCV_HAL_IMPL_SSE_LOADSTORE_INTERLEAVE(v_int8x16, schar, s8, v_uint8x16, uchar, u8)
-OPENCV_HAL_IMPL_SSE_LOADSTORE_INTERLEAVE(v_int16x8, short, s16, v_uint16x8, ushort, u16)
-OPENCV_HAL_IMPL_SSE_LOADSTORE_INTERLEAVE(v_int32x4, int, s32, v_uint32x4, unsigned, u32)
-OPENCV_HAL_IMPL_SSE_LOADSTORE_INTERLEAVE(v_int64x2, int64, s64, v_uint64x2, uint64, u64)
-OPENCV_HAL_IMPL_SSE_LOADSTORE_INTERLEAVE(v_float64x2, double, f64, v_uint64x2, uint64, u64)
-
-inline v_float32x4 v_cvt_f32(const v_int32x4& a)
-{
-    return v_float32x4(_mm_cvtepi32_ps(a.val));
-}
-
-inline v_float32x4 v_cvt_f32(const v_float64x2& a)
-{
-    return v_float32x4(_mm_cvtpd_ps(a.val));
-}
-
-inline v_float32x4 v_cvt_f32(const v_float64x2& a, const v_float64x2& b)
-{
-    return v_float32x4(_mm_movelh_ps(_mm_cvtpd_ps(a.val), _mm_cvtpd_ps(b.val)));
-}
-
-inline v_float64x2 v_cvt_f64(const v_int32x4& a)
-{
-    return v_float64x2(_mm_cvtepi32_pd(a.val));
-}
-
-inline v_float64x2 v_cvt_f64_high(const v_int32x4& a)
-{
-    return v_float64x2(_mm_cvtepi32_pd(_mm_srli_si128(a.val,8)));
-}
-
-inline v_float64x2 v_cvt_f64(const v_float32x4& a)
-{
-    return v_float64x2(_mm_cvtps_pd(a.val));
-}
-
-inline v_float64x2 v_cvt_f64_high(const v_float32x4& a)
-{
-    return v_float64x2(_mm_cvtps_pd(_mm_movehl_ps(a.val, a.val)));
-}
-
-// from (Mysticial and wim) https://stackoverflow.com/q/41144668
-inline v_float64x2 v_cvt_f64(const v_int64x2& v)
-{
-    // constants encoded as floating-point
-    __m128i magic_i_hi32 = _mm_set1_epi64x(0x4530000080000000); // 2^84 + 2^63
-    __m128i magic_i_all  = _mm_set1_epi64x(0x4530000080100000); // 2^84 + 2^63 + 2^52
-    __m128d magic_d_all  = _mm_castsi128_pd(magic_i_all);
-    // Blend the 32 lowest significant bits of v with magic_int_lo
-#if CV_SSE4_1
-    __m128i magic_i_lo   = _mm_set1_epi64x(0x4330000000000000); // 2^52
-    __m128i v_lo         = _mm_blend_epi16(v.val, magic_i_lo, 0xcc);
-#else
-    __m128i magic_i_lo   = _mm_set1_epi32(0x43300000); // 2^52
-    __m128i v_lo         = _mm_unpacklo_epi32(_mm_shuffle_epi32(v.val, _MM_SHUFFLE(0, 0, 2, 0)), magic_i_lo);
-#endif
-    // Extract the 32 most significant bits of v
-    __m128i v_hi         = _mm_srli_epi64(v.val, 32);
-    // Flip the msb of v_hi and blend with 0x45300000
-            v_hi         = _mm_xor_si128(v_hi, magic_i_hi32);
-    // Compute in double precision
-    __m128d v_hi_dbl     = _mm_sub_pd(_mm_castsi128_pd(v_hi), magic_d_all);
-    // (v_hi - magic_d_all) + v_lo  Do not assume associativity of floating point addition
-    __m128d result       = _mm_add_pd(v_hi_dbl, _mm_castsi128_pd(v_lo));
-    return v_float64x2(result);
-}
-
-////////////// Lookup table access ////////////////////
-
-inline v_int8x16 v_lut(const schar* tab, const int* idx)
-{
-#if defined(_MSC_VER)
-    return v_int8x16(_mm_setr_epi8(tab[idx[0]], tab[idx[1]], tab[idx[ 2]], tab[idx[ 3]], tab[idx[ 4]], tab[idx[ 5]], tab[idx[ 6]], tab[idx[ 7]],
-                                   tab[idx[8]], tab[idx[9]], tab[idx[10]], tab[idx[11]], tab[idx[12]], tab[idx[13]], tab[idx[14]], tab[idx[15]]));
-#else
-    return v_int8x16(_mm_setr_epi64(
-                        _mm_setr_pi8(tab[idx[0]], tab[idx[1]], tab[idx[ 2]], tab[idx[ 3]], tab[idx[ 4]], tab[idx[ 5]], tab[idx[ 6]], tab[idx[ 7]]),
-                        _mm_setr_pi8(tab[idx[8]], tab[idx[9]], tab[idx[10]], tab[idx[11]], tab[idx[12]], tab[idx[13]], tab[idx[14]], tab[idx[15]])
-                    ));
-#endif
-}
-inline v_int8x16 v_lut_pairs(const schar* tab, const int* idx)
-{
-#if defined(_MSC_VER)
-    return v_int8x16(_mm_setr_epi16(*(const short*)(tab + idx[0]), *(const short*)(tab + idx[1]), *(const short*)(tab + idx[2]), *(const short*)(tab + idx[3]),
-                                    *(const short*)(tab + idx[4]), *(const short*)(tab + idx[5]), *(const short*)(tab + idx[6]), *(const short*)(tab + idx[7])));
-#else
-    return v_int8x16(_mm_setr_epi64(
-                        _mm_setr_pi16(*(const short*)(tab + idx[0]), *(const short*)(tab + idx[1]), *(const short*)(tab + idx[2]), *(const short*)(tab + idx[3])),
-                        _mm_setr_pi16(*(const short*)(tab + idx[4]), *(const short*)(tab + idx[5]), *(const short*)(tab + idx[6]), *(const short*)(tab + idx[7]))
-                    ));
-#endif
-}
-inline v_int8x16 v_lut_quads(const schar* tab, const int* idx)
-{
-#if defined(_MSC_VER)
-    return v_int8x16(_mm_setr_epi32(*(const int*)(tab + idx[0]), *(const int*)(tab + idx[1]),
-                                    *(const int*)(tab + idx[2]), *(const int*)(tab + idx[3])));
-#else
-    return v_int8x16(_mm_setr_epi64(
-                        _mm_setr_pi32(*(const int*)(tab + idx[0]), *(const int*)(tab + idx[1])),
-                        _mm_setr_pi32(*(const int*)(tab + idx[2]), *(const int*)(tab + idx[3]))
-                    ));
-#endif
-}
-inline v_uint8x16 v_lut(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut((const schar *)tab, idx)); }
-inline v_uint8x16 v_lut_pairs(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut_pairs((const schar *)tab, idx)); }
-inline v_uint8x16 v_lut_quads(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut_quads((const schar *)tab, idx)); }
-
-inline v_int16x8 v_lut(const short* tab, const int* idx)
-{
-#if defined(_MSC_VER)
-    return v_int16x8(_mm_setr_epi16(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]],
-                                    tab[idx[4]], tab[idx[5]], tab[idx[6]], tab[idx[7]]));
-#else
-    return v_int16x8(_mm_setr_epi64(
-                        _mm_setr_pi16(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]]),
-                        _mm_setr_pi16(tab[idx[4]], tab[idx[5]], tab[idx[6]], tab[idx[7]])
-                    ));
-#endif
-}
-inline v_int16x8 v_lut_pairs(const short* tab, const int* idx)
-{
-#if defined(_MSC_VER)
-    return v_int16x8(_mm_setr_epi32(*(const int*)(tab + idx[0]), *(const int*)(tab + idx[1]),
-                                    *(const int*)(tab + idx[2]), *(const int*)(tab + idx[3])));
-#else
-    return v_int16x8(_mm_setr_epi64(
-                        _mm_setr_pi32(*(const int*)(tab + idx[0]), *(const int*)(tab + idx[1])),
-                        _mm_setr_pi32(*(const int*)(tab + idx[2]), *(const int*)(tab + idx[3]))
-                    ));
-#endif
-}
-inline v_int16x8 v_lut_quads(const short* tab, const int* idx)
-{
-    return v_int16x8(_mm_set_epi64x(*(const int64_t*)(tab + idx[1]), *(const int64_t*)(tab + idx[0])));
-}
-inline v_uint16x8 v_lut(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut((const short *)tab, idx)); }
-inline v_uint16x8 v_lut_pairs(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut_pairs((const short *)tab, idx)); }
-inline v_uint16x8 v_lut_quads(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut_quads((const short *)tab, idx)); }
-
-inline v_int32x4 v_lut(const int* tab, const int* idx)
-{
-#if defined(_MSC_VER)
-    return v_int32x4(_mm_setr_epi32(tab[idx[0]], tab[idx[1]],
-                                    tab[idx[2]], tab[idx[3]]));
-#else
-    return v_int32x4(_mm_setr_epi64(
-                        _mm_setr_pi32(tab[idx[0]], tab[idx[1]]),
-                        _mm_setr_pi32(tab[idx[2]], tab[idx[3]])
-                    ));
-#endif
-}
-inline v_int32x4 v_lut_pairs(const int* tab, const int* idx)
-{
-    return v_int32x4(_mm_set_epi64x(*(const int64_t*)(tab + idx[1]), *(const int64_t*)(tab + idx[0])));
-}
-inline v_int32x4 v_lut_quads(const int* tab, const int* idx)
-{
-    return v_int32x4(_mm_loadu_si128((const __m128i*)(tab + idx[0])));
-}
-inline v_uint32x4 v_lut(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut((const int *)tab, idx)); }
-inline v_uint32x4 v_lut_pairs(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut_pairs((const int *)tab, idx)); }
-inline v_uint32x4 v_lut_quads(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut_quads((const int *)tab, idx)); }
-
-inline v_int64x2 v_lut(const int64_t* tab, const int* idx)
-{
-    return v_int64x2(_mm_set_epi64x(tab[idx[1]], tab[idx[0]]));
-}
-inline v_int64x2 v_lut_pairs(const int64_t* tab, const int* idx)
-{
-    return v_int64x2(_mm_loadu_si128((const __m128i*)(tab + idx[0])));
-}
-inline v_uint64x2 v_lut(const uint64_t* tab, const int* idx) { return v_reinterpret_as_u64(v_lut((const int64_t *)tab, idx)); }
-inline v_uint64x2 v_lut_pairs(const uint64_t* tab, const int* idx) { return v_reinterpret_as_u64(v_lut_pairs((const int64_t *)tab, idx)); }
-
-inline v_float32x4 v_lut(const float* tab, const int* idx)
-{
-    return v_float32x4(_mm_setr_ps(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]]));
-}
-inline v_float32x4 v_lut_pairs(const float* tab, const int* idx) { return v_reinterpret_as_f32(v_lut_pairs((const int *)tab, idx)); }
-inline v_float32x4 v_lut_quads(const float* tab, const int* idx) { return v_reinterpret_as_f32(v_lut_quads((const int *)tab, idx)); }
-
-inline v_float64x2 v_lut(const double* tab, const int* idx)
-{
-    return v_float64x2(_mm_setr_pd(tab[idx[0]], tab[idx[1]]));
-}
-inline v_float64x2 v_lut_pairs(const double* tab, const int* idx) { return v_float64x2(_mm_castsi128_pd(_mm_loadu_si128((const __m128i*)(tab + idx[0])))); }
-
-inline v_int32x4 v_lut(const int* tab, const v_int32x4& idxvec)
-{
-    int CV_DECL_ALIGNED(32) idx[4];
-    v_store_aligned(idx, idxvec);
-    return v_int32x4(_mm_setr_epi32(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]]));
-}
-
-inline v_uint32x4 v_lut(const unsigned* tab, const v_int32x4& idxvec)
-{
-    return v_reinterpret_as_u32(v_lut((const int *)tab, idxvec));
-}
-
-inline v_float32x4 v_lut(const float* tab, const v_int32x4& idxvec)
-{
-    int CV_DECL_ALIGNED(32) idx[4];
-    v_store_aligned(idx, idxvec);
-    return v_float32x4(_mm_setr_ps(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]]));
-}
-
-inline v_float64x2 v_lut(const double* tab, const v_int32x4& idxvec)
-{
-    int idx[2];
-    v_store_low(idx, idxvec);
-    return v_float64x2(_mm_setr_pd(tab[idx[0]], tab[idx[1]]));
-}
-
-// loads pairs from the table and deinterleaves them, e.g. returns:
-//   x = (tab[idxvec[0], tab[idxvec[1]], tab[idxvec[2]], tab[idxvec[3]]),
-//   y = (tab[idxvec[0]+1], tab[idxvec[1]+1], tab[idxvec[2]+1], tab[idxvec[3]+1])
-// note that the indices are float's indices, not the float-pair indices.
-// in theory, this function can be used to implement bilinear interpolation,
-// when idxvec are the offsets within the image.
-inline void v_lut_deinterleave(const float* tab, const v_int32x4& idxvec, v_float32x4& x, v_float32x4& y)
-{
-    int CV_DECL_ALIGNED(32) idx[4];
-    v_store_aligned(idx, idxvec);
-    __m128 z = _mm_setzero_ps();
-    __m128 xy01 = _mm_loadl_pi(z, (__m64*)(tab + idx[0]));
-    __m128 xy23 = _mm_loadl_pi(z, (__m64*)(tab + idx[2]));
-    xy01 = _mm_loadh_pi(xy01, (__m64*)(tab + idx[1]));
-    xy23 = _mm_loadh_pi(xy23, (__m64*)(tab + idx[3]));
-    __m128 xxyy02 = _mm_unpacklo_ps(xy01, xy23);
-    __m128 xxyy13 = _mm_unpackhi_ps(xy01, xy23);
-    x = v_float32x4(_mm_unpacklo_ps(xxyy02, xxyy13));
-    y = v_float32x4(_mm_unpackhi_ps(xxyy02, xxyy13));
-}
-
-inline void v_lut_deinterleave(const double* tab, const v_int32x4& idxvec, v_float64x2& x, v_float64x2& y)
-{
-    int idx[2];
-    v_store_low(idx, idxvec);
-    __m128d xy0 = _mm_loadu_pd(tab + idx[0]);
-    __m128d xy1 = _mm_loadu_pd(tab + idx[1]);
-    x = v_float64x2(_mm_unpacklo_pd(xy0, xy1));
-    y = v_float64x2(_mm_unpackhi_pd(xy0, xy1));
-}
-
-inline v_int8x16 v_interleave_pairs(const v_int8x16& vec)
-{
-#if CV_SSSE3
-    return v_int8x16(_mm_shuffle_epi8(vec.val, _mm_set_epi64x(0x0f0d0e0c0b090a08, 0x0705060403010200)));
-#else
-    __m128i a = _mm_shufflelo_epi16(vec.val, _MM_SHUFFLE(3, 1, 2, 0));
-    a = _mm_shufflehi_epi16(a, _MM_SHUFFLE(3, 1, 2, 0));
-    a = _mm_shuffle_epi32(a, _MM_SHUFFLE(3, 1, 2, 0));
-    return v_int8x16(_mm_unpacklo_epi8(a, _mm_unpackhi_epi64(a, a)));
-#endif
-}
-inline v_uint8x16 v_interleave_pairs(const v_uint8x16& vec) { return v_reinterpret_as_u8(v_interleave_pairs(v_reinterpret_as_s8(vec))); }
-inline v_int8x16 v_interleave_quads(const v_int8x16& vec)
-{
-#if CV_SSSE3
-    return v_int8x16(_mm_shuffle_epi8(vec.val, _mm_set_epi64x(0x0f0b0e0a0d090c08, 0x0703060205010400)));
-#else
-    __m128i a = _mm_shuffle_epi32(vec.val, _MM_SHUFFLE(3, 1, 2, 0));
-    return v_int8x16(_mm_unpacklo_epi8(a, _mm_unpackhi_epi64(a, a)));
-#endif
-}
-inline v_uint8x16 v_interleave_quads(const v_uint8x16& vec) { return v_reinterpret_as_u8(v_interleave_quads(v_reinterpret_as_s8(vec))); }
-
-inline v_int16x8 v_interleave_pairs(const v_int16x8& vec)
-{
-#if CV_SSSE3
-    return v_int16x8(_mm_shuffle_epi8(vec.val, _mm_set_epi64x(0x0f0e0b0a0d0c0908, 0x0706030205040100)));
-#else
-    __m128i a = _mm_shufflelo_epi16(vec.val, _MM_SHUFFLE(3, 1, 2, 0));
-    return v_int16x8(_mm_shufflehi_epi16(a, _MM_SHUFFLE(3, 1, 2, 0)));
-#endif
-}
-inline v_uint16x8 v_interleave_pairs(const v_uint16x8& vec) { return v_reinterpret_as_u16(v_interleave_pairs(v_reinterpret_as_s16(vec))); }
-inline v_int16x8 v_interleave_quads(const v_int16x8& vec)
-{
-#if CV_SSSE3
-    return v_int16x8(_mm_shuffle_epi8(vec.val, _mm_set_epi64x(0x0f0e07060d0c0504, 0x0b0a030209080100)));
-#else
-    return v_int16x8(_mm_unpacklo_epi16(vec.val, _mm_unpackhi_epi64(vec.val, vec.val)));
-#endif
-}
-inline v_uint16x8 v_interleave_quads(const v_uint16x8& vec) { return v_reinterpret_as_u16(v_interleave_quads(v_reinterpret_as_s16(vec))); }
-
-inline v_int32x4 v_interleave_pairs(const v_int32x4& vec)
-{
-    return v_int32x4(_mm_shuffle_epi32(vec.val, _MM_SHUFFLE(3, 1, 2, 0)));
-}
-inline v_uint32x4 v_interleave_pairs(const v_uint32x4& vec) { return v_reinterpret_as_u32(v_interleave_pairs(v_reinterpret_as_s32(vec))); }
-inline v_float32x4 v_interleave_pairs(const v_float32x4& vec) { return v_reinterpret_as_f32(v_interleave_pairs(v_reinterpret_as_s32(vec))); }
-
-inline v_int8x16 v_pack_triplets(const v_int8x16& vec)
-{
-#if CV_SSSE3
-    return v_int8x16(_mm_shuffle_epi8(vec.val, _mm_set_epi64x(0xffffff0f0e0d0c0a, 0x0908060504020100)));
-#else
-    __m128i mask = _mm_set1_epi64x(0x00000000FFFFFFFF);
-    __m128i a = _mm_srli_si128(_mm_or_si128(_mm_andnot_si128(mask, vec.val), _mm_and_si128(mask, _mm_sll_epi32(vec.val, _mm_set_epi64x(0, 8)))), 1);
-    return v_int8x16(_mm_srli_si128(_mm_shufflelo_epi16(a, _MM_SHUFFLE(2, 1, 0, 3)), 2));
-#endif
-}
-inline v_uint8x16 v_pack_triplets(const v_uint8x16& vec) { return v_reinterpret_as_u8(v_pack_triplets(v_reinterpret_as_s8(vec))); }
-
-inline v_int16x8 v_pack_triplets(const v_int16x8& vec)
-{
-#if CV_SSSE3
-    return v_int16x8(_mm_shuffle_epi8(vec.val, _mm_set_epi64x(0xffff0f0e0d0c0b0a, 0x0908050403020100)));
-#else
-    return v_int16x8(_mm_srli_si128(_mm_shufflelo_epi16(vec.val, _MM_SHUFFLE(2, 1, 0, 3)), 2));
-#endif
-}
-inline v_uint16x8 v_pack_triplets(const v_uint16x8& vec) { return v_reinterpret_as_u16(v_pack_triplets(v_reinterpret_as_s16(vec))); }
-
-inline v_int32x4 v_pack_triplets(const v_int32x4& vec) { return vec; }
-inline v_uint32x4 v_pack_triplets(const v_uint32x4& vec) { return vec; }
-inline v_float32x4 v_pack_triplets(const v_float32x4& vec) { return vec; }
-
-template<int i>
-inline uchar v_extract_n(const v_uint8x16& v)
-{
-#if CV_SSE4_1
-    return (uchar)_mm_extract_epi8(v.val, i);
-#else
-    return v_rotate_right<i>(v).get0();
-#endif
-}
-
-template<int i>
-inline schar v_extract_n(const v_int8x16& v)
-{
-    return (schar)v_extract_n<i>(v_reinterpret_as_u8(v));
-}
-
-template<int i>
-inline ushort v_extract_n(const v_uint16x8& v)
-{
-    return (ushort)_mm_extract_epi16(v.val, i);
-}
-
-template<int i>
-inline short v_extract_n(const v_int16x8& v)
-{
-    return (short)v_extract_n<i>(v_reinterpret_as_u16(v));
-}
-
-template<int i>
-inline uint v_extract_n(const v_uint32x4& v)
-{
-#if CV_SSE4_1
-    return (uint)_mm_extract_epi32(v.val, i);
-#else
-    return v_rotate_right<i>(v).get0();
-#endif
-}
-
-template<int i>
-inline int v_extract_n(const v_int32x4& v)
-{
-    return (int)v_extract_n<i>(v_reinterpret_as_u32(v));
-}
-
-template<int i>
-inline uint64 v_extract_n(const v_uint64x2& v)
-{
-#ifdef CV__SIMD_NATIVE_mm_extract_epi64
-    return (uint64)_v128_extract_epi64<i>(v.val);
-#else
-    return v_rotate_right<i>(v).get0();
-#endif
-}
-
-template<int i>
-inline int64 v_extract_n(const v_int64x2& v)
-{
-    return (int64)v_extract_n<i>(v_reinterpret_as_u64(v));
-}
-
-template<int i>
-inline float v_extract_n(const v_float32x4& v)
-{
-    union { uint iv; float fv; } d;
-    d.iv = v_extract_n<i>(v_reinterpret_as_u32(v));
-    return d.fv;
-}
-
-template<int i>
-inline double v_extract_n(const v_float64x2& v)
-{
-    union { uint64 iv; double dv; } d;
-    d.iv = v_extract_n<i>(v_reinterpret_as_u64(v));
-    return d.dv;
-}
-
-template<int i>
-inline v_int32x4 v_broadcast_element(const v_int32x4& v)
-{
-    return v_int32x4(_mm_shuffle_epi32(v.val, _MM_SHUFFLE(i,i,i,i)));
-}
-
-template<int i>
-inline v_uint32x4 v_broadcast_element(const v_uint32x4& v)
-{
-    return v_uint32x4(_mm_shuffle_epi32(v.val, _MM_SHUFFLE(i,i,i,i)));
-}
-
-template<int i>
-inline v_float32x4 v_broadcast_element(const v_float32x4& v)
-{
-    return v_float32x4(_mm_shuffle_ps(v.val, v.val, _MM_SHUFFLE((char)i,(char)i,(char)i,(char)i)));
-}
-
-////////////// FP16 support ///////////////////////////
-
-inline v_float32x4 v_load_expand(const float16_t* ptr)
-{
-#if CV_FP16
-    return v_float32x4(_mm_cvtph_ps(_mm_loadu_si128((const __m128i*)ptr)));
-#else
-    const __m128i z = _mm_setzero_si128(), delta = _mm_set1_epi32(0x38000000);
-    const __m128i signmask = _mm_set1_epi32(0x80000000), maxexp = _mm_set1_epi32(0x7c000000);
-    const __m128 deltaf = _mm_castsi128_ps(_mm_set1_epi32(0x38800000));
-    __m128i bits = _mm_unpacklo_epi16(z, _mm_loadl_epi64((const __m128i*)ptr)); // h << 16
-    __m128i e = _mm_and_si128(bits, maxexp), sign = _mm_and_si128(bits, signmask);
-    __m128i t = _mm_add_epi32(_mm_srli_epi32(_mm_xor_si128(bits, sign), 3), delta); // ((h & 0x7fff) << 13) + delta
-    __m128i zt = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_add_epi32(t, _mm_set1_epi32(1 << 23))), deltaf));
-
-    t = _mm_add_epi32(t, _mm_and_si128(delta, _mm_cmpeq_epi32(maxexp, e)));
-    __m128i zmask = _mm_cmpeq_epi32(e, z);
-    __m128i ft = v_select_si128(zmask, zt, t);
-    return v_float32x4(_mm_castsi128_ps(_mm_or_si128(ft, sign)));
-#endif
-}
-
-inline void v_pack_store(float16_t* ptr, const v_float32x4& v)
-{
-#if CV_FP16
-    __m128i fp16_value = _mm_cvtps_ph(v.val, 0);
-    _mm_storel_epi64((__m128i*)ptr, fp16_value);
-#else
-    const __m128i signmask = _mm_set1_epi32(0x80000000);
-    const __m128i rval = _mm_set1_epi32(0x3f000000);
-
-    __m128i t = _mm_castps_si128(v.val);
-    __m128i sign = _mm_srai_epi32(_mm_and_si128(t, signmask), 16);
-    t = _mm_andnot_si128(signmask, t);
-
-    __m128i finitemask = _mm_cmpgt_epi32(_mm_set1_epi32(0x47800000), t);
-    __m128i isnan = _mm_cmpgt_epi32(t, _mm_set1_epi32(0x7f800000));
-    __m128i naninf = v_select_si128(isnan, _mm_set1_epi32(0x7e00), _mm_set1_epi32(0x7c00));
-    __m128i tinymask = _mm_cmpgt_epi32(_mm_set1_epi32(0x38800000), t);
-    __m128i tt = _mm_castps_si128(_mm_add_ps(_mm_castsi128_ps(t), _mm_castsi128_ps(rval)));
-    tt = _mm_sub_epi32(tt, rval);
-    __m128i odd = _mm_and_si128(_mm_srli_epi32(t, 13), _mm_set1_epi32(1));
-    __m128i nt = _mm_add_epi32(t, _mm_set1_epi32(0xc8000fff));
-    nt = _mm_srli_epi32(_mm_add_epi32(nt, odd), 13);
-    t = v_select_si128(tinymask, tt, nt);
-    t = v_select_si128(finitemask, t, naninf);
-    t = _mm_or_si128(t, sign);
-    t = _mm_packs_epi32(t, t);
-    _mm_storel_epi64((__m128i*)ptr, t);
-#endif
-}
-
-inline void v_cleanup() {}
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END
-
-//! @endcond
-
-}
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_sse_em.hpp b/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_sse_em.hpp
deleted file mode 100644
index 6fb08816..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_sse_em.hpp
+++ /dev/null
@@ -1,180 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html
-
-#ifndef OPENCV_HAL_INTRIN_SSE_EM_HPP
-#define OPENCV_HAL_INTRIN_SSE_EM_HPP
-
-namespace cv
-{
-
-//! @cond IGNORED
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
-
-#define OPENCV_HAL_SSE_WRAP_1(fun, tp) \
-    inline tp _v128_##fun(const tp& a) \
-    { return _mm_##fun(a); }
-
-#define OPENCV_HAL_SSE_WRAP_2(fun, tp) \
-    inline tp _v128_##fun(const tp& a, const tp& b) \
-    { return _mm_##fun(a, b); }
-
-#define OPENCV_HAL_SSE_WRAP_3(fun, tp) \
-    inline tp _v128_##fun(const tp& a, const tp& b, const tp& c) \
-    { return _mm_##fun(a, b, c); }
-
-///////////////////////////// XOP /////////////////////////////
-
-// [todo] define CV_XOP
-#if 1 // CV_XOP
-inline __m128i _v128_comgt_epu32(const __m128i& a, const __m128i& b)
-{
-    const __m128i delta = _mm_set1_epi32((int)0x80000000);
-    return _mm_cmpgt_epi32(_mm_xor_si128(a, delta), _mm_xor_si128(b, delta));
-}
-// wrapping XOP
-#else
-OPENCV_HAL_SSE_WRAP_2(_v128_comgt_epu32, __m128i)
-#endif // !CV_XOP
-
-///////////////////////////// SSE4.1 /////////////////////////////
-
-#if !CV_SSE4_1
-
-/** Swizzle **/
-inline __m128i _v128_blendv_epi8(const __m128i& a, const __m128i& b, const __m128i& mask)
-{ return _mm_xor_si128(a, _mm_and_si128(_mm_xor_si128(b, a), mask)); }
-
-/** Convert **/
-// 8 >> 16
-inline __m128i _v128_cvtepu8_epi16(const __m128i& a)
-{
-    const __m128i z = _mm_setzero_si128();
-    return _mm_unpacklo_epi8(a, z);
-}
-inline __m128i _v128_cvtepi8_epi16(const __m128i& a)
-{ return _mm_srai_epi16(_mm_unpacklo_epi8(a, a), 8); }
-// 8 >> 32
-inline __m128i _v128_cvtepu8_epi32(const __m128i& a)
-{
-    const __m128i z = _mm_setzero_si128();
-    return _mm_unpacklo_epi16(_mm_unpacklo_epi8(a, z), z);
-}
-inline __m128i _v128_cvtepi8_epi32(const __m128i& a)
-{
-    __m128i r = _mm_unpacklo_epi8(a, a);
-    r = _mm_unpacklo_epi8(r, r);
-    return _mm_srai_epi32(r, 24);
-}
-// 16 >> 32
-inline __m128i _v128_cvtepu16_epi32(const __m128i& a)
-{
-    const __m128i z = _mm_setzero_si128();
-    return _mm_unpacklo_epi16(a, z);
-}
-inline __m128i _v128_cvtepi16_epi32(const __m128i& a)
-{ return _mm_srai_epi32(_mm_unpacklo_epi16(a, a), 16); }
-// 32 >> 64
-inline __m128i _v128_cvtepu32_epi64(const __m128i& a)
-{
-    const __m128i z = _mm_setzero_si128();
-    return _mm_unpacklo_epi32(a, z);
-}
-inline __m128i _v128_cvtepi32_epi64(const __m128i& a)
-{ return _mm_unpacklo_epi32(a, _mm_srai_epi32(a, 31)); }
-
-/** Arithmetic **/
-inline __m128i _v128_mullo_epi32(const __m128i& a, const __m128i& b)
-{
-    __m128i c0 = _mm_mul_epu32(a, b);
-    __m128i c1 = _mm_mul_epu32(_mm_srli_epi64(a, 32), _mm_srli_epi64(b, 32));
-    __m128i d0 = _mm_unpacklo_epi32(c0, c1);
-    __m128i d1 = _mm_unpackhi_epi32(c0, c1);
-    return _mm_unpacklo_epi64(d0, d1);
-}
-
-/** Math **/
-inline __m128i _v128_min_epu32(const __m128i& a, const __m128i& b)
-{ return _v128_blendv_epi8(a, b, _v128_comgt_epu32(a, b)); }
-
-// wrapping SSE4.1
-#else
-OPENCV_HAL_SSE_WRAP_1(cvtepu8_epi16, __m128i)
-OPENCV_HAL_SSE_WRAP_1(cvtepi8_epi16, __m128i)
-OPENCV_HAL_SSE_WRAP_1(cvtepu8_epi32, __m128i)
-OPENCV_HAL_SSE_WRAP_1(cvtepi8_epi32, __m128i)
-OPENCV_HAL_SSE_WRAP_1(cvtepu16_epi32, __m128i)
-OPENCV_HAL_SSE_WRAP_1(cvtepi16_epi32, __m128i)
-OPENCV_HAL_SSE_WRAP_1(cvtepu32_epi64, __m128i)
-OPENCV_HAL_SSE_WRAP_1(cvtepi32_epi64, __m128i)
-OPENCV_HAL_SSE_WRAP_2(min_epu32, __m128i)
-OPENCV_HAL_SSE_WRAP_2(mullo_epi32, __m128i)
-OPENCV_HAL_SSE_WRAP_3(blendv_epi8, __m128i)
-#endif // !CV_SSE4_1
-
-///////////////////////////// Revolutionary /////////////////////////////
-
-/** Convert **/
-// 16 << 8
-inline __m128i _v128_cvtepu8_epi16_high(const __m128i& a)
-{
-    const __m128i z = _mm_setzero_si128();
-    return _mm_unpackhi_epi8(a, z);
-}
-inline __m128i _v128_cvtepi8_epi16_high(const __m128i& a)
-{ return _mm_srai_epi16(_mm_unpackhi_epi8(a, a), 8); }
-// 32 << 16
-inline __m128i _v128_cvtepu16_epi32_high(const __m128i& a)
-{
-    const __m128i z = _mm_setzero_si128();
-    return _mm_unpackhi_epi16(a, z);
-}
-inline __m128i _v128_cvtepi16_epi32_high(const __m128i& a)
-{ return _mm_srai_epi32(_mm_unpackhi_epi16(a, a), 16); }
-// 64 << 32
-inline __m128i _v128_cvtepu32_epi64_high(const __m128i& a)
-{
-    const __m128i z = _mm_setzero_si128();
-    return _mm_unpackhi_epi32(a, z);
-}
-inline __m128i _v128_cvtepi32_epi64_high(const __m128i& a)
-{ return _mm_unpackhi_epi32(a, _mm_srai_epi32(a, 31)); }
-
-/** Miscellaneous **/
-inline __m128i _v128_packs_epu32(const __m128i& a, const __m128i& b)
-{
-    const __m128i m = _mm_set1_epi32(65535);
-    __m128i am = _v128_min_epu32(a, m);
-    __m128i bm = _v128_min_epu32(b, m);
-#if CV_SSE4_1
-    return _mm_packus_epi32(am, bm);
-#else
-    const __m128i d = _mm_set1_epi32(32768), nd = _mm_set1_epi16(-32768);
-    am = _mm_sub_epi32(am, d);
-    bm = _mm_sub_epi32(bm, d);
-    am = _mm_packs_epi32(am, bm);
-    return _mm_sub_epi16(am, nd);
-#endif
-}
-
-template<int i>
-inline int64 _v128_extract_epi64(const __m128i& a)
-{
-#if defined(CV__SIMD_HAVE_mm_extract_epi64) || (CV_SSE4_1 && (defined(__x86_64__)/*GCC*/ || defined(_M_X64)/*MSVC*/))
-#define CV__SIMD_NATIVE_mm_extract_epi64 1
-    return _mm_extract_epi64(a, i);
-#else
-    CV_DECL_ALIGNED(16) int64 tmp[2];
-    _mm_store_si128((__m128i*)tmp, a);
-    return tmp[i];
-#endif
-}
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END
-
-//! @endcond
-
-} // cv::
-
-#endif // OPENCV_HAL_INTRIN_SSE_EM_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_vsx.hpp b/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_vsx.hpp
deleted file mode 100644
index b198643c..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_vsx.hpp
+++ /dev/null
@@ -1,1608 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html
-
-#ifndef OPENCV_HAL_VSX_HPP
-#define OPENCV_HAL_VSX_HPP
-
-#include <algorithm>
-#include "opencv2/core/utility.hpp"
-
-#define CV_SIMD128 1
-#define CV_SIMD128_64F 1
-
-namespace cv
-{
-
-//! @cond IGNORED
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
-
-///////// Types ////////////
-
-struct v_uint8x16
-{
-    typedef uchar lane_type;
-    enum { nlanes = 16 };
-    vec_uchar16 val;
-
-    explicit v_uint8x16(const vec_uchar16& v) : val(v)
-    {}
-    v_uint8x16()
-    {}
-    v_uint8x16(vec_bchar16 v) : val(vec_uchar16_c(v))
-    {}
-    v_uint8x16(uchar v0, uchar v1, uchar v2, uchar v3, uchar v4, uchar v5, uchar v6, uchar v7,
-               uchar v8, uchar v9, uchar v10, uchar v11, uchar v12, uchar v13, uchar v14, uchar v15)
-        : val(vec_uchar16_set(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15))
-    {}
-
-    static inline v_uint8x16 zero() { return v_uint8x16(vec_uchar16_z); }
-
-    uchar get0() const
-    { return vec_extract(val, 0); }
-};
-
-struct v_int8x16
-{
-    typedef schar lane_type;
-    enum { nlanes = 16 };
-    vec_char16 val;
-
-    explicit v_int8x16(const vec_char16& v) : val(v)
-    {}
-    v_int8x16()
-    {}
-    v_int8x16(vec_bchar16 v) : val(vec_char16_c(v))
-    {}
-    v_int8x16(schar v0, schar v1, schar v2, schar v3, schar v4, schar v5, schar v6, schar v7,
-              schar v8, schar v9, schar v10, schar v11, schar v12, schar v13, schar v14, schar v15)
-        : val(vec_char16_set(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15))
-    {}
-
-    static inline v_int8x16 zero() { return v_int8x16(vec_char16_z); }
-
-    schar get0() const
-    { return vec_extract(val, 0); }
-};
-
-struct v_uint16x8
-{
-    typedef ushort lane_type;
-    enum { nlanes = 8 };
-    vec_ushort8 val;
-
-    explicit v_uint16x8(const vec_ushort8& v) : val(v)
-    {}
-    v_uint16x8()
-    {}
-    v_uint16x8(vec_bshort8 v) : val(vec_ushort8_c(v))
-    {}
-    v_uint16x8(ushort v0, ushort v1, ushort v2, ushort v3, ushort v4, ushort v5, ushort v6, ushort v7)
-        : val(vec_ushort8_set(v0, v1, v2, v3, v4, v5, v6, v7))
-    {}
-
-    static inline v_uint16x8 zero() { return v_uint16x8(vec_ushort8_z); }
-
-    ushort get0() const
-    { return vec_extract(val, 0); }
-};
-
-struct v_int16x8
-{
-    typedef short lane_type;
-    enum { nlanes = 8 };
-    vec_short8 val;
-
-    explicit v_int16x8(const vec_short8& v) : val(v)
-    {}
-    v_int16x8()
-    {}
-    v_int16x8(vec_bshort8 v) : val(vec_short8_c(v))
-    {}
-    v_int16x8(short v0, short v1, short v2, short v3, short v4, short v5, short v6, short v7)
-        : val(vec_short8_set(v0, v1, v2, v3, v4, v5, v6, v7))
-    {}
-
-    static inline v_int16x8 zero() { return v_int16x8(vec_short8_z); }
-
-    short get0() const
-    { return vec_extract(val, 0); }
-};
-
-struct v_uint32x4
-{
-    typedef unsigned lane_type;
-    enum { nlanes = 4 };
-    vec_uint4 val;
-
-    explicit v_uint32x4(const vec_uint4& v) : val(v)
-    {}
-    v_uint32x4()
-    {}
-    v_uint32x4(vec_bint4 v) : val(vec_uint4_c(v))
-    {}
-    v_uint32x4(unsigned v0, unsigned v1, unsigned v2, unsigned v3) : val(vec_uint4_set(v0, v1, v2, v3))
-    {}
-
-    static inline v_uint32x4 zero() { return v_uint32x4(vec_uint4_z); }
-
-    uint get0() const
-    { return vec_extract(val, 0); }
-};
-
-struct v_int32x4
-{
-    typedef int lane_type;
-    enum { nlanes = 4 };
-    vec_int4 val;
-
-    explicit v_int32x4(const vec_int4& v) : val(v)
-    {}
-    v_int32x4()
-    {}
-    v_int32x4(vec_bint4 v) : val(vec_int4_c(v))
-    {}
-    v_int32x4(int v0, int v1, int v2, int v3) : val(vec_int4_set(v0, v1, v2, v3))
-    {}
-
-    static inline v_int32x4 zero() { return v_int32x4(vec_int4_z); }
-
-    int get0() const
-    { return vec_extract(val, 0); }
-};
-
-struct v_float32x4
-{
-    typedef float lane_type;
-    enum { nlanes = 4 };
-    vec_float4 val;
-
-    explicit v_float32x4(const vec_float4& v) : val(v)
-    {}
-    v_float32x4()
-    {}
-    v_float32x4(vec_bint4 v) : val(vec_float4_c(v))
-    {}
-    v_float32x4(float v0, float v1, float v2, float v3) : val(vec_float4_set(v0, v1, v2, v3))
-    {}
-
-    static inline v_float32x4 zero() { return v_float32x4(vec_float4_z); }
-
-    float get0() const
-    { return vec_extract(val, 0); }
-};
-
-struct v_uint64x2
-{
-    typedef uint64 lane_type;
-    enum { nlanes = 2 };
-    vec_udword2 val;
-
-    explicit v_uint64x2(const vec_udword2& v) : val(v)
-    {}
-    v_uint64x2()
-    {}
-    v_uint64x2(vec_bdword2 v) : val(vec_udword2_c(v))
-    {}
-    v_uint64x2(uint64 v0, uint64 v1) : val(vec_udword2_set(v0, v1))
-    {}
-
-    static inline v_uint64x2 zero() { return v_uint64x2(vec_udword2_z); }
-
-    uint64 get0() const
-    { return vec_extract(val, 0); }
-};
-
-struct v_int64x2
-{
-    typedef int64 lane_type;
-    enum { nlanes = 2 };
-    vec_dword2 val;
-
-    explicit v_int64x2(const vec_dword2& v) : val(v)
-    {}
-    v_int64x2()
-    {}
-    v_int64x2(vec_bdword2 v) : val(vec_dword2_c(v))
-    {}
-    v_int64x2(int64 v0, int64 v1) : val(vec_dword2_set(v0, v1))
-    {}
-
-    static inline v_int64x2 zero() { return v_int64x2(vec_dword2_z); }
-
-    int64 get0() const
-    { return vec_extract(val, 0); }
-};
-
-struct v_float64x2
-{
-    typedef double lane_type;
-    enum { nlanes = 2 };
-    vec_double2 val;
-
-    explicit v_float64x2(const vec_double2& v) : val(v)
-    {}
-    v_float64x2()
-    {}
-    v_float64x2(vec_bdword2 v) : val(vec_double2_c(v))
-    {}
-    v_float64x2(double v0, double v1) : val(vec_double2_set(v0, v1))
-    {}
-
-    static inline v_float64x2 zero() { return v_float64x2(vec_double2_z); }
-
-    double get0() const
-    { return vec_extract(val, 0); }
-};
-
-#define OPENCV_HAL_IMPL_VSX_EXTRACT_N(_Tpvec, _Tp) \
-template<int i> inline _Tp v_extract_n(VSX_UNUSED(_Tpvec v)) { return vec_extract(v.val, i); }
-
-OPENCV_HAL_IMPL_VSX_EXTRACT_N(v_uint8x16, uchar)
-OPENCV_HAL_IMPL_VSX_EXTRACT_N(v_int8x16, schar)
-OPENCV_HAL_IMPL_VSX_EXTRACT_N(v_uint16x8, ushort)
-OPENCV_HAL_IMPL_VSX_EXTRACT_N(v_int16x8, short)
-OPENCV_HAL_IMPL_VSX_EXTRACT_N(v_uint32x4, uint)
-OPENCV_HAL_IMPL_VSX_EXTRACT_N(v_int32x4, int)
-OPENCV_HAL_IMPL_VSX_EXTRACT_N(v_uint64x2, uint64)
-OPENCV_HAL_IMPL_VSX_EXTRACT_N(v_int64x2, int64)
-OPENCV_HAL_IMPL_VSX_EXTRACT_N(v_float32x4, float)
-OPENCV_HAL_IMPL_VSX_EXTRACT_N(v_float64x2, double)
-
-//////////////// Load and store operations ///////////////
-
-/*
- * clang-5 aborted during parse "vec_xxx_c" only if it's
- * inside a function template which is defined by preprocessor macro.
- *
- * if vec_xxx_c defined as C++ cast, clang-5 will pass it
-*/
-#define OPENCV_HAL_IMPL_VSX_INITVEC(_Tpvec, _Tp, suffix, cast)                        \
-inline _Tpvec v_setzero_##suffix() { return _Tpvec(vec_splats((_Tp)0)); }             \
-inline _Tpvec v_setall_##suffix(_Tp v) { return _Tpvec(vec_splats((_Tp)v));}          \
-template<typename _Tpvec0> inline _Tpvec v_reinterpret_as_##suffix(const _Tpvec0 &a)  \
-{ return _Tpvec((cast)a.val); }
-
-OPENCV_HAL_IMPL_VSX_INITVEC(v_uint8x16, uchar, u8, vec_uchar16)
-OPENCV_HAL_IMPL_VSX_INITVEC(v_int8x16, schar, s8, vec_char16)
-OPENCV_HAL_IMPL_VSX_INITVEC(v_uint16x8, ushort, u16, vec_ushort8)
-OPENCV_HAL_IMPL_VSX_INITVEC(v_int16x8, short, s16, vec_short8)
-OPENCV_HAL_IMPL_VSX_INITVEC(v_uint32x4, uint, u32, vec_uint4)
-OPENCV_HAL_IMPL_VSX_INITVEC(v_int32x4, int, s32, vec_int4)
-OPENCV_HAL_IMPL_VSX_INITVEC(v_uint64x2, uint64, u64, vec_udword2)
-OPENCV_HAL_IMPL_VSX_INITVEC(v_int64x2, int64, s64, vec_dword2)
-OPENCV_HAL_IMPL_VSX_INITVEC(v_float32x4, float, f32, vec_float4)
-OPENCV_HAL_IMPL_VSX_INITVEC(v_float64x2, double, f64, vec_double2)
-
-#define OPENCV_HAL_IMPL_VSX_LOADSTORE_C(_Tpvec, _Tp, ld, ld_a, st, st_a)    \
-inline _Tpvec v_load(const _Tp* ptr)                                        \
-{ return _Tpvec(ld(0, ptr)); }                                              \
-inline _Tpvec v_load_aligned(VSX_UNUSED(const _Tp* ptr))                    \
-{ return _Tpvec(ld_a(0, ptr)); }                                            \
-inline _Tpvec v_load_low(const _Tp* ptr)                                    \
-{ return _Tpvec(vec_ld_l8(ptr)); }                                          \
-inline _Tpvec v_load_halves(const _Tp* ptr0, const _Tp* ptr1)               \
-{ return _Tpvec(vec_mergesqh(vec_ld_l8(ptr0), vec_ld_l8(ptr1))); }          \
-inline void v_store(_Tp* ptr, const _Tpvec& a)                              \
-{ st(a.val, 0, ptr); }                                                      \
-inline void v_store_aligned(VSX_UNUSED(_Tp* ptr), const _Tpvec& a)          \
-{ st_a(a.val, 0, ptr); }                                                    \
-inline void v_store_aligned_nocache(VSX_UNUSED(_Tp* ptr), const _Tpvec& a)  \
-{ st_a(a.val, 0, ptr); }                                                    \
-inline void v_store(_Tp* ptr, const _Tpvec& a, hal::StoreMode mode)         \
-{ if(mode == hal::STORE_UNALIGNED) st(a.val, 0, ptr); else st_a(a.val, 0, ptr); } \
-inline void v_store_low(_Tp* ptr, const _Tpvec& a)                          \
-{ vec_st_l8(a.val, ptr); }                                                  \
-inline void v_store_high(_Tp* ptr, const _Tpvec& a)                         \
-{ vec_st_h8(a.val, ptr); }
-
-// working around gcc bug for aligned ld/st
-// if runtime check for vec_ld/st fail we failback to unaligned ld/st
-// https://github.com/opencv/opencv/issues/13211
-#ifdef CV_COMPILER_VSX_BROKEN_ALIGNED
-    #define OPENCV_HAL_IMPL_VSX_LOADSTORE(_Tpvec, _Tp) \
-    OPENCV_HAL_IMPL_VSX_LOADSTORE_C(_Tpvec, _Tp, vsx_ld, vsx_ld, vsx_st, vsx_st)
-#else
-    #define OPENCV_HAL_IMPL_VSX_LOADSTORE(_Tpvec, _Tp) \
-    OPENCV_HAL_IMPL_VSX_LOADSTORE_C(_Tpvec, _Tp, vsx_ld, vec_ld, vsx_st, vec_st)
-#endif
-
-OPENCV_HAL_IMPL_VSX_LOADSTORE(v_uint8x16,  uchar)
-OPENCV_HAL_IMPL_VSX_LOADSTORE(v_int8x16,   schar)
-OPENCV_HAL_IMPL_VSX_LOADSTORE(v_uint16x8,  ushort)
-OPENCV_HAL_IMPL_VSX_LOADSTORE(v_int16x8,   short)
-OPENCV_HAL_IMPL_VSX_LOADSTORE(v_uint32x4,  uint)
-OPENCV_HAL_IMPL_VSX_LOADSTORE(v_int32x4,   int)
-OPENCV_HAL_IMPL_VSX_LOADSTORE(v_float32x4, float)
-
-OPENCV_HAL_IMPL_VSX_LOADSTORE_C(v_float64x2, double, vsx_ld,  vsx_ld,  vsx_st,  vsx_st)
-OPENCV_HAL_IMPL_VSX_LOADSTORE_C(v_uint64x2,  uint64, vsx_ld2, vsx_ld2, vsx_st2, vsx_st2)
-OPENCV_HAL_IMPL_VSX_LOADSTORE_C(v_int64x2,    int64, vsx_ld2, vsx_ld2, vsx_st2, vsx_st2)
-
-//////////////// Value reordering ///////////////
-
-/* de&interleave */
-#define OPENCV_HAL_IMPL_VSX_INTERLEAVE(_Tp, _Tpvec)                          \
-inline void v_load_deinterleave(const _Tp* ptr, _Tpvec& a, _Tpvec& b)        \
-{ vec_ld_deinterleave(ptr, a.val, b.val);}                                   \
-inline void v_load_deinterleave(const _Tp* ptr, _Tpvec& a,                   \
-                                _Tpvec& b, _Tpvec& c)                        \
-{ vec_ld_deinterleave(ptr, a.val, b.val, c.val); }                           \
-inline void v_load_deinterleave(const _Tp* ptr, _Tpvec& a, _Tpvec& b,        \
-                                                _Tpvec& c, _Tpvec& d)        \
-{ vec_ld_deinterleave(ptr, a.val, b.val, c.val, d.val); }                    \
-inline void v_store_interleave(_Tp* ptr, const _Tpvec& a, const _Tpvec& b,   \
-                               hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ vec_st_interleave(a.val, b.val, ptr); }                                    \
-inline void v_store_interleave(_Tp* ptr, const _Tpvec& a,                    \
-                               const _Tpvec& b, const _Tpvec& c,             \
-                               hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ vec_st_interleave(a.val, b.val, c.val, ptr); }                             \
-inline void v_store_interleave(_Tp* ptr, const _Tpvec& a, const _Tpvec& b,   \
-                                         const _Tpvec& c, const _Tpvec& d,   \
-                               hal::StoreMode /*mode*/=hal::STORE_UNALIGNED) \
-{ vec_st_interleave(a.val, b.val, c.val, d.val, ptr); }
-
-OPENCV_HAL_IMPL_VSX_INTERLEAVE(uchar, v_uint8x16)
-OPENCV_HAL_IMPL_VSX_INTERLEAVE(schar, v_int8x16)
-OPENCV_HAL_IMPL_VSX_INTERLEAVE(ushort, v_uint16x8)
-OPENCV_HAL_IMPL_VSX_INTERLEAVE(short, v_int16x8)
-OPENCV_HAL_IMPL_VSX_INTERLEAVE(uint, v_uint32x4)
-OPENCV_HAL_IMPL_VSX_INTERLEAVE(int, v_int32x4)
-OPENCV_HAL_IMPL_VSX_INTERLEAVE(float, v_float32x4)
-OPENCV_HAL_IMPL_VSX_INTERLEAVE(double, v_float64x2)
-OPENCV_HAL_IMPL_VSX_INTERLEAVE(int64, v_int64x2)
-OPENCV_HAL_IMPL_VSX_INTERLEAVE(uint64, v_uint64x2)
-
-/* Expand */
-#define OPENCV_HAL_IMPL_VSX_EXPAND(_Tpvec, _Tpwvec, _Tp, fl, fh)  \
-inline void v_expand(const _Tpvec& a, _Tpwvec& b0, _Tpwvec& b1)   \
-{                                                                 \
-    b0.val = fh(a.val);                                           \
-    b1.val = fl(a.val);                                           \
-}                                                                 \
-inline _Tpwvec v_expand_low(const _Tpvec& a)                      \
-{ return _Tpwvec(fh(a.val)); }                                    \
-inline _Tpwvec v_expand_high(const _Tpvec& a)                     \
-{ return _Tpwvec(fl(a.val)); }                                    \
-inline _Tpwvec v_load_expand(const _Tp* ptr)                      \
-{ return _Tpwvec(fh(vec_ld_l8(ptr))); }
-
-OPENCV_HAL_IMPL_VSX_EXPAND(v_uint8x16, v_uint16x8, uchar, vec_unpacklu, vec_unpackhu)
-OPENCV_HAL_IMPL_VSX_EXPAND(v_int8x16, v_int16x8, schar, vec_unpackl, vec_unpackh)
-OPENCV_HAL_IMPL_VSX_EXPAND(v_uint16x8, v_uint32x4, ushort, vec_unpacklu, vec_unpackhu)
-OPENCV_HAL_IMPL_VSX_EXPAND(v_int16x8, v_int32x4, short, vec_unpackl, vec_unpackh)
-OPENCV_HAL_IMPL_VSX_EXPAND(v_uint32x4, v_uint64x2, uint, vec_unpacklu, vec_unpackhu)
-OPENCV_HAL_IMPL_VSX_EXPAND(v_int32x4, v_int64x2, int, vec_unpackl, vec_unpackh)
-
-/* Load and zero expand a 4 byte value into the second dword, first is don't care. */
-#if !defined(CV_COMPILER_VSX_BROKEN_ASM)
-    #define _LXSIWZX(out, ptr, T) __asm__ ("lxsiwzx %x0, 0, %1\r\n" : "=wa"(out) : "r" (ptr) : "memory");
-#else
-    /* This is compiler-agnostic, but will introduce an unneeded splat on the critical path. */
-    #define _LXSIWZX(out, ptr, T) out = (T)vec_udword2_sp(*(uint32_t*)(ptr));
-#endif
-
-inline v_uint32x4 v_load_expand_q(const uchar* ptr)
-{
-    // Zero-extend the extra 24B instead of unpacking. Usually faster in small kernel
-    // Likewise note, value is zero extended and upper 4 bytes are zero'ed.
-    vec_uchar16 pmu = {8, 12, 12, 12, 9, 12, 12, 12, 10, 12, 12, 12, 11, 12, 12, 12};
-    vec_uchar16 out;
-
-    _LXSIWZX(out, ptr, vec_uchar16);
-    out = vec_perm(out, out, pmu);
-    return v_uint32x4((vec_uint4)out);
-}
-
-inline v_int32x4 v_load_expand_q(const schar* ptr)
-{
-    vec_char16 out;
-    vec_short8 outs;
-    vec_int4 outw;
-
-    _LXSIWZX(out, ptr, vec_char16);
-    outs = vec_unpackl(out);
-    outw = vec_unpackh(outs);
-    return v_int32x4(outw);
-}
-
-/* pack */
-#define OPENCV_HAL_IMPL_VSX_PACK(_Tpvec, _Tp, _Tpwvec, _Tpvn, _Tpdel, sfnc, pkfnc, addfnc, pack)    \
-inline _Tpvec v_##pack(const _Tpwvec& a, const _Tpwvec& b)                                          \
-{                                                                                                   \
-    return _Tpvec(pkfnc(a.val, b.val));                                                             \
-}                                                                                                   \
-inline void v_##pack##_store(_Tp* ptr, const _Tpwvec& a)                                            \
-{                                                                                                   \
-    vec_st_l8(pkfnc(a.val, a.val), ptr);                                                            \
-}                                                                                                   \
-template<int n>                                                                                     \
-inline _Tpvec v_rshr_##pack(const _Tpwvec& a, const _Tpwvec& b)                                     \
-{                                                                                                   \
-    const __vector _Tpvn vn = vec_splats((_Tpvn)n);                                                 \
-    const __vector _Tpdel delta = vec_splats((_Tpdel)((_Tpdel)1 << (n-1)));                         \
-    return _Tpvec(pkfnc(sfnc(addfnc(a.val, delta), vn), sfnc(addfnc(b.val, delta), vn)));           \
-}                                                                                                   \
-template<int n>                                                                                     \
-inline void v_rshr_##pack##_store(_Tp* ptr, const _Tpwvec& a)                                       \
-{                                                                                                   \
-    const __vector _Tpvn vn = vec_splats((_Tpvn)n);                                                 \
-    const __vector _Tpdel delta = vec_splats((_Tpdel)((_Tpdel)1 << (n-1)));                         \
-    vec_st_l8(pkfnc(sfnc(addfnc(a.val, delta), vn), delta), ptr);                                   \
-}
-
-OPENCV_HAL_IMPL_VSX_PACK(v_uint8x16, uchar, v_uint16x8, unsigned short, unsigned short,
-                         vec_sr, vec_packs, vec_adds, pack)
-OPENCV_HAL_IMPL_VSX_PACK(v_int8x16, schar, v_int16x8, unsigned short, short,
-                         vec_sra, vec_packs, vec_adds, pack)
-
-OPENCV_HAL_IMPL_VSX_PACK(v_uint16x8, ushort, v_uint32x4, unsigned int, unsigned int,
-                         vec_sr, vec_packs, vec_add, pack)
-OPENCV_HAL_IMPL_VSX_PACK(v_int16x8, short, v_int32x4, unsigned int, int,
-                         vec_sra, vec_packs, vec_add, pack)
-
-OPENCV_HAL_IMPL_VSX_PACK(v_uint32x4, uint, v_uint64x2, unsigned long long, unsigned long long,
-                         vec_sr, vec_pack, vec_add, pack)
-OPENCV_HAL_IMPL_VSX_PACK(v_int32x4, int, v_int64x2, unsigned long long, long long,
-                         vec_sra, vec_pack, vec_add, pack)
-
-OPENCV_HAL_IMPL_VSX_PACK(v_uint8x16, uchar, v_int16x8, unsigned short, short,
-                         vec_sra, vec_packsu, vec_adds, pack_u)
-OPENCV_HAL_IMPL_VSX_PACK(v_uint16x8, ushort, v_int32x4, unsigned int, int,
-                         vec_sra, vec_packsu, vec_add, pack_u)
-// Following variant is not implemented on other platforms:
-//OPENCV_HAL_IMPL_VSX_PACK(v_uint32x4, uint, v_int64x2, unsigned long long, long long,
-//                         vec_sra, vec_packsu, vec_add, pack_u)
-
-// pack boolean
-inline v_uint8x16 v_pack_b(const v_uint16x8& a, const v_uint16x8& b)
-{
-    vec_uchar16 ab = vec_pack(a.val, b.val);
-    return v_uint8x16(ab);
-}
-
-inline v_uint8x16 v_pack_b(const v_uint32x4& a, const v_uint32x4& b,
-                           const v_uint32x4& c, const v_uint32x4& d)
-{
-    vec_ushort8 ab = vec_pack(a.val, b.val);
-    vec_ushort8 cd = vec_pack(c.val, d.val);
-    return v_uint8x16(vec_pack(ab, cd));
-}
-
-inline v_uint8x16 v_pack_b(const v_uint64x2& a, const v_uint64x2& b, const v_uint64x2& c,
-                           const v_uint64x2& d, const v_uint64x2& e, const v_uint64x2& f,
-                           const v_uint64x2& g, const v_uint64x2& h)
-{
-    vec_uint4 ab = vec_pack(a.val, b.val);
-    vec_uint4 cd = vec_pack(c.val, d.val);
-    vec_uint4 ef = vec_pack(e.val, f.val);
-    vec_uint4 gh = vec_pack(g.val, h.val);
-
-    vec_ushort8 abcd = vec_pack(ab, cd);
-    vec_ushort8 efgh = vec_pack(ef, gh);
-    return v_uint8x16(vec_pack(abcd, efgh));
-}
-
-/* Recombine */
-template <typename _Tpvec>
-inline void v_zip(const _Tpvec& a0, const _Tpvec& a1, _Tpvec& b0, _Tpvec& b1)
-{
-    b0.val = vec_mergeh(a0.val, a1.val);
-    b1.val = vec_mergel(a0.val, a1.val);
-}
-
-template <typename _Tpvec>
-inline _Tpvec v_combine_high(const _Tpvec& a, const _Tpvec& b)
-{ return _Tpvec(vec_mergesql(a.val, b.val)); }
-
-template <typename _Tpvec>
-inline _Tpvec v_combine_low(const _Tpvec& a, const _Tpvec& b)
-{ return _Tpvec(vec_mergesqh(a.val, b.val)); }
-
-template <typename _Tpvec>
-inline void v_recombine(const _Tpvec& a, const _Tpvec& b, _Tpvec& c, _Tpvec& d)
-{
-    c.val = vec_mergesqh(a.val, b.val);
-    d.val = vec_mergesql(a.val, b.val);
-}
-
-////////// Arithmetic, bitwise and comparison operations /////////
-
-/* Element-wise binary and unary operations */
-/** Arithmetics **/
-#define OPENCV_HAL_IMPL_VSX_BIN_OP(bin_op, _Tpvec, intrin)       \
-inline _Tpvec operator bin_op (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(intrin(a.val, b.val)); }                         \
-inline _Tpvec& operator bin_op##= (_Tpvec& a, const _Tpvec& b)   \
-{ a.val = intrin(a.val, b.val); return a; }
-
-OPENCV_HAL_IMPL_VSX_BIN_OP(+, v_uint8x16, vec_adds)
-OPENCV_HAL_IMPL_VSX_BIN_OP(-, v_uint8x16, vec_subs)
-OPENCV_HAL_IMPL_VSX_BIN_OP(+, v_int8x16,  vec_adds)
-OPENCV_HAL_IMPL_VSX_BIN_OP(-, v_int8x16, vec_subs)
-OPENCV_HAL_IMPL_VSX_BIN_OP(+, v_uint16x8, vec_adds)
-OPENCV_HAL_IMPL_VSX_BIN_OP(-, v_uint16x8, vec_subs)
-OPENCV_HAL_IMPL_VSX_BIN_OP(+, v_int16x8, vec_adds)
-OPENCV_HAL_IMPL_VSX_BIN_OP(-, v_int16x8, vec_subs)
-OPENCV_HAL_IMPL_VSX_BIN_OP(+, v_uint32x4, vec_add)
-OPENCV_HAL_IMPL_VSX_BIN_OP(-, v_uint32x4, vec_sub)
-OPENCV_HAL_IMPL_VSX_BIN_OP(*, v_uint32x4, vec_mul)
-OPENCV_HAL_IMPL_VSX_BIN_OP(+, v_int32x4, vec_add)
-OPENCV_HAL_IMPL_VSX_BIN_OP(-, v_int32x4, vec_sub)
-OPENCV_HAL_IMPL_VSX_BIN_OP(*, v_int32x4, vec_mul)
-OPENCV_HAL_IMPL_VSX_BIN_OP(+, v_float32x4, vec_add)
-OPENCV_HAL_IMPL_VSX_BIN_OP(-, v_float32x4, vec_sub)
-OPENCV_HAL_IMPL_VSX_BIN_OP(*, v_float32x4, vec_mul)
-OPENCV_HAL_IMPL_VSX_BIN_OP(/, v_float32x4, vec_div)
-OPENCV_HAL_IMPL_VSX_BIN_OP(+, v_float64x2, vec_add)
-OPENCV_HAL_IMPL_VSX_BIN_OP(-, v_float64x2, vec_sub)
-OPENCV_HAL_IMPL_VSX_BIN_OP(*, v_float64x2, vec_mul)
-OPENCV_HAL_IMPL_VSX_BIN_OP(/, v_float64x2, vec_div)
-OPENCV_HAL_IMPL_VSX_BIN_OP(+, v_uint64x2, vec_add)
-OPENCV_HAL_IMPL_VSX_BIN_OP(-, v_uint64x2, vec_sub)
-OPENCV_HAL_IMPL_VSX_BIN_OP(+, v_int64x2, vec_add)
-OPENCV_HAL_IMPL_VSX_BIN_OP(-, v_int64x2, vec_sub)
-
-// saturating multiply
-#define OPENCV_HAL_IMPL_VSX_MUL_SAT(_Tpvec, _Tpwvec)             \
-    inline _Tpvec operator * (const _Tpvec& a, const _Tpvec& b)  \
-    {                                                            \
-        _Tpwvec c, d;                                            \
-        v_mul_expand(a, b, c, d);                                \
-        return v_pack(c, d);                                     \
-    }                                                            \
-    inline _Tpvec& operator *= (_Tpvec& a, const _Tpvec& b)      \
-    { a = a * b; return a; }
-
-OPENCV_HAL_IMPL_VSX_MUL_SAT(v_int8x16,  v_int16x8)
-OPENCV_HAL_IMPL_VSX_MUL_SAT(v_uint8x16, v_uint16x8)
-OPENCV_HAL_IMPL_VSX_MUL_SAT(v_int16x8,  v_int32x4)
-OPENCV_HAL_IMPL_VSX_MUL_SAT(v_uint16x8, v_uint32x4)
-
-template<typename Tvec, typename Twvec>
-inline void v_mul_expand(const Tvec& a, const Tvec& b, Twvec& c, Twvec& d)
-{
-    Twvec p0 = Twvec(vec_mule(a.val, b.val));
-    Twvec p1 = Twvec(vec_mulo(a.val, b.val));
-    v_zip(p0, p1, c, d);
-}
-
-inline v_int16x8 v_mul_hi(const v_int16x8& a, const v_int16x8& b)
-{
-    vec_int4 p0 = vec_mule(a.val, b.val);
-    vec_int4 p1 = vec_mulo(a.val, b.val);
-    static const vec_uchar16 perm = {2, 3, 18, 19, 6, 7, 22, 23, 10, 11, 26, 27, 14, 15, 30, 31};
-    return v_int16x8(vec_perm(vec_short8_c(p0), vec_short8_c(p1), perm));
-}
-inline v_uint16x8 v_mul_hi(const v_uint16x8& a, const v_uint16x8& b)
-{
-    vec_uint4 p0 = vec_mule(a.val, b.val);
-    vec_uint4 p1 = vec_mulo(a.val, b.val);
-    static const vec_uchar16 perm = {2, 3, 18, 19, 6, 7, 22, 23, 10, 11, 26, 27, 14, 15, 30, 31};
-    return v_uint16x8(vec_perm(vec_ushort8_c(p0), vec_ushort8_c(p1), perm));
-}
-
-/** Non-saturating arithmetics **/
-#define OPENCV_HAL_IMPL_VSX_BIN_FUNC(func, intrin)    \
-template<typename _Tpvec>                             \
-inline _Tpvec func(const _Tpvec& a, const _Tpvec& b)  \
-{ return _Tpvec(intrin(a.val, b.val)); }
-
-OPENCV_HAL_IMPL_VSX_BIN_FUNC(v_add_wrap, vec_add)
-OPENCV_HAL_IMPL_VSX_BIN_FUNC(v_sub_wrap, vec_sub)
-OPENCV_HAL_IMPL_VSX_BIN_FUNC(v_mul_wrap, vec_mul)
-
-/** Bitwise shifts **/
-#define OPENCV_HAL_IMPL_VSX_SHIFT_OP(_Tpvec, shr, splfunc)   \
-inline _Tpvec operator << (const _Tpvec& a, int imm)         \
-{ return _Tpvec(vec_sl(a.val, splfunc(imm))); }              \
-inline _Tpvec operator >> (const _Tpvec& a, int imm)         \
-{ return _Tpvec(shr(a.val, splfunc(imm))); }                 \
-template<int imm> inline _Tpvec v_shl(const _Tpvec& a)       \
-{ return _Tpvec(vec_sl(a.val, splfunc(imm))); }              \
-template<int imm> inline _Tpvec v_shr(const _Tpvec& a)       \
-{ return _Tpvec(shr(a.val, splfunc(imm))); }
-
-OPENCV_HAL_IMPL_VSX_SHIFT_OP(v_uint8x16, vec_sr, vec_uchar16_sp)
-OPENCV_HAL_IMPL_VSX_SHIFT_OP(v_uint16x8, vec_sr, vec_ushort8_sp)
-OPENCV_HAL_IMPL_VSX_SHIFT_OP(v_uint32x4, vec_sr, vec_uint4_sp)
-OPENCV_HAL_IMPL_VSX_SHIFT_OP(v_uint64x2, vec_sr, vec_udword2_sp)
-// algebraic right shift
-OPENCV_HAL_IMPL_VSX_SHIFT_OP(v_int8x16, vec_sra, vec_uchar16_sp)
-OPENCV_HAL_IMPL_VSX_SHIFT_OP(v_int16x8, vec_sra, vec_ushort8_sp)
-OPENCV_HAL_IMPL_VSX_SHIFT_OP(v_int32x4, vec_sra, vec_uint4_sp)
-OPENCV_HAL_IMPL_VSX_SHIFT_OP(v_int64x2, vec_sra, vec_udword2_sp)
-
-/** Bitwise logic **/
-#define OPENCV_HAL_IMPL_VSX_LOGIC_OP(_Tpvec)    \
-OPENCV_HAL_IMPL_VSX_BIN_OP(&, _Tpvec, vec_and)  \
-OPENCV_HAL_IMPL_VSX_BIN_OP(|, _Tpvec, vec_or)   \
-OPENCV_HAL_IMPL_VSX_BIN_OP(^, _Tpvec, vec_xor)  \
-inline _Tpvec operator ~ (const _Tpvec& a)      \
-{ return _Tpvec(vec_not(a.val)); }
-
-OPENCV_HAL_IMPL_VSX_LOGIC_OP(v_uint8x16)
-OPENCV_HAL_IMPL_VSX_LOGIC_OP(v_int8x16)
-OPENCV_HAL_IMPL_VSX_LOGIC_OP(v_uint16x8)
-OPENCV_HAL_IMPL_VSX_LOGIC_OP(v_int16x8)
-OPENCV_HAL_IMPL_VSX_LOGIC_OP(v_uint32x4)
-OPENCV_HAL_IMPL_VSX_LOGIC_OP(v_int32x4)
-OPENCV_HAL_IMPL_VSX_LOGIC_OP(v_uint64x2)
-OPENCV_HAL_IMPL_VSX_LOGIC_OP(v_int64x2)
-OPENCV_HAL_IMPL_VSX_LOGIC_OP(v_float32x4)
-OPENCV_HAL_IMPL_VSX_LOGIC_OP(v_float64x2)
-
-/** Bitwise select **/
-#define OPENCV_HAL_IMPL_VSX_SELECT(_Tpvec, cast)                             \
-inline _Tpvec v_select(const _Tpvec& mask, const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(vec_sel(b.val, a.val, cast(mask.val))); }
-
-OPENCV_HAL_IMPL_VSX_SELECT(v_uint8x16, vec_bchar16_c)
-OPENCV_HAL_IMPL_VSX_SELECT(v_int8x16, vec_bchar16_c)
-OPENCV_HAL_IMPL_VSX_SELECT(v_uint16x8, vec_bshort8_c)
-OPENCV_HAL_IMPL_VSX_SELECT(v_int16x8, vec_bshort8_c)
-OPENCV_HAL_IMPL_VSX_SELECT(v_uint32x4, vec_bint4_c)
-OPENCV_HAL_IMPL_VSX_SELECT(v_int32x4, vec_bint4_c)
-OPENCV_HAL_IMPL_VSX_SELECT(v_float32x4, vec_bint4_c)
-OPENCV_HAL_IMPL_VSX_SELECT(v_float64x2, vec_bdword2_c)
-
-/** Comparison **/
-#define OPENCV_HAL_IMPL_VSX_INT_CMP_OP(_Tpvec)                 \
-inline _Tpvec operator == (const _Tpvec& a, const _Tpvec& b)   \
-{ return _Tpvec(vec_cmpeq(a.val, b.val)); }                    \
-inline _Tpvec operator != (const _Tpvec& a, const _Tpvec& b)   \
-{ return _Tpvec(vec_cmpne(a.val, b.val)); }                    \
-inline _Tpvec operator < (const _Tpvec& a, const _Tpvec& b)    \
-{ return _Tpvec(vec_cmplt(a.val, b.val)); }                    \
-inline _Tpvec operator > (const _Tpvec& a, const _Tpvec& b)    \
-{ return _Tpvec(vec_cmpgt(a.val, b.val)); }                    \
-inline _Tpvec operator <= (const _Tpvec& a, const _Tpvec& b)   \
-{ return _Tpvec(vec_cmple(a.val, b.val)); }                    \
-inline _Tpvec operator >= (const _Tpvec& a, const _Tpvec& b)   \
-{ return _Tpvec(vec_cmpge(a.val, b.val)); }
-
-OPENCV_HAL_IMPL_VSX_INT_CMP_OP(v_uint8x16)
-OPENCV_HAL_IMPL_VSX_INT_CMP_OP(v_int8x16)
-OPENCV_HAL_IMPL_VSX_INT_CMP_OP(v_uint16x8)
-OPENCV_HAL_IMPL_VSX_INT_CMP_OP(v_int16x8)
-OPENCV_HAL_IMPL_VSX_INT_CMP_OP(v_uint32x4)
-OPENCV_HAL_IMPL_VSX_INT_CMP_OP(v_int32x4)
-OPENCV_HAL_IMPL_VSX_INT_CMP_OP(v_float32x4)
-OPENCV_HAL_IMPL_VSX_INT_CMP_OP(v_float64x2)
-OPENCV_HAL_IMPL_VSX_INT_CMP_OP(v_uint64x2)
-OPENCV_HAL_IMPL_VSX_INT_CMP_OP(v_int64x2)
-
-inline v_float32x4 v_not_nan(const v_float32x4& a)
-{ return v_float32x4(vec_cmpeq(a.val, a.val)); }
-inline v_float64x2 v_not_nan(const v_float64x2& a)
-{ return v_float64x2(vec_cmpeq(a.val, a.val)); }
-
-/** min/max **/
-OPENCV_HAL_IMPL_VSX_BIN_FUNC(v_min, vec_min)
-OPENCV_HAL_IMPL_VSX_BIN_FUNC(v_max, vec_max)
-
-/** Rotate **/
-#define OPENCV_IMPL_VSX_ROTATE(_Tpvec, suffix, shf, cast)                       \
-template<int imm>                                                               \
-inline _Tpvec v_rotate_##suffix(const _Tpvec& a)                                \
-{                                                                               \
-    const int wd = imm * sizeof(typename _Tpvec::lane_type);                    \
-    if (wd > 15)                                                                \
-        return _Tpvec::zero();                                                  \
-    return _Tpvec((cast)shf(vec_uchar16_c(a.val), vec_uchar16_sp(wd << 3)));    \
-}
-
-#define OPENCV_IMPL_VSX_ROTATE_LR(_Tpvec, cast)     \
-OPENCV_IMPL_VSX_ROTATE(_Tpvec, left, vec_slo, cast) \
-OPENCV_IMPL_VSX_ROTATE(_Tpvec, right, vec_sro, cast)
-
-OPENCV_IMPL_VSX_ROTATE_LR(v_uint8x16, vec_uchar16)
-OPENCV_IMPL_VSX_ROTATE_LR(v_int8x16,  vec_char16)
-OPENCV_IMPL_VSX_ROTATE_LR(v_uint16x8, vec_ushort8)
-OPENCV_IMPL_VSX_ROTATE_LR(v_int16x8,  vec_short8)
-OPENCV_IMPL_VSX_ROTATE_LR(v_uint32x4, vec_uint4)
-OPENCV_IMPL_VSX_ROTATE_LR(v_int32x4,  vec_int4)
-OPENCV_IMPL_VSX_ROTATE_LR(v_float32x4, vec_float4)
-OPENCV_IMPL_VSX_ROTATE_LR(v_uint64x2, vec_udword2)
-OPENCV_IMPL_VSX_ROTATE_LR(v_int64x2,  vec_dword2)
-OPENCV_IMPL_VSX_ROTATE_LR(v_float64x2, vec_double2)
-
-template<int imm, typename _Tpvec>
-inline _Tpvec v_rotate_right(const _Tpvec& a, const _Tpvec& b)
-{
-    enum { CV_SHIFT = 16 - imm * (sizeof(typename _Tpvec::lane_type)) };
-    if (CV_SHIFT == 16)
-        return a;
-#ifdef __IBMCPP__
-    return _Tpvec(vec_sld(b.val, a.val, CV_SHIFT & 15));
-#else
-    return _Tpvec(vec_sld(b.val, a.val, CV_SHIFT));
-#endif
-}
-
-template<int imm, typename _Tpvec>
-inline _Tpvec v_rotate_left(const _Tpvec& a, const _Tpvec& b)
-{
-    enum { CV_SHIFT = imm * (sizeof(typename _Tpvec::lane_type)) };
-    if (CV_SHIFT == 16)
-        return b;
-    return _Tpvec(vec_sld(a.val, b.val, CV_SHIFT));
-}
-
-#define OPENCV_IMPL_VSX_ROTATE_64_2RG(_Tpvec, suffix, rg1, rg2)   \
-template<int imm>                                                 \
-inline _Tpvec v_rotate_##suffix(const _Tpvec& a, const _Tpvec& b) \
-{                                                                 \
-    if (imm == 1)                                                 \
-        return _Tpvec(vec_permi(rg1.val, rg2.val, 2));            \
-    return imm ? b : a;                                           \
-}
-
-#define OPENCV_IMPL_VSX_ROTATE_64_2RG_LR(_Tpvec)    \
-OPENCV_IMPL_VSX_ROTATE_64_2RG(_Tpvec, left,  b, a)  \
-OPENCV_IMPL_VSX_ROTATE_64_2RG(_Tpvec, right, a, b)
-
-OPENCV_IMPL_VSX_ROTATE_64_2RG_LR(v_float64x2)
-OPENCV_IMPL_VSX_ROTATE_64_2RG_LR(v_uint64x2)
-OPENCV_IMPL_VSX_ROTATE_64_2RG_LR(v_int64x2)
-
-/* Reverse */
-inline v_uint8x16 v_reverse(const v_uint8x16 &a)
-{
-    static const vec_uchar16 perm = {15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0};
-    vec_uchar16 vec = (vec_uchar16)a.val;
-    return v_uint8x16(vec_perm(vec, vec, perm));
-}
-
-inline v_int8x16 v_reverse(const v_int8x16 &a)
-{ return v_reinterpret_as_s8(v_reverse(v_reinterpret_as_u8(a))); }
-
-inline v_uint16x8 v_reverse(const v_uint16x8 &a)
-{
-    static const vec_uchar16 perm = {14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1};
-    vec_uchar16 vec = (vec_uchar16)a.val;
-    return v_reinterpret_as_u16(v_uint8x16(vec_perm(vec, vec, perm)));
-}
-
-inline v_int16x8 v_reverse(const v_int16x8 &a)
-{ return v_reinterpret_as_s16(v_reverse(v_reinterpret_as_u16(a))); }
-
-inline v_uint32x4 v_reverse(const v_uint32x4 &a)
-{
-    static const vec_uchar16 perm = {12, 13, 14, 15, 8, 9, 10, 11, 4, 5, 6, 7, 0, 1, 2, 3};
-    vec_uchar16 vec = (vec_uchar16)a.val;
-    return v_reinterpret_as_u32(v_uint8x16(vec_perm(vec, vec, perm)));
-}
-
-inline v_int32x4 v_reverse(const v_int32x4 &a)
-{ return v_reinterpret_as_s32(v_reverse(v_reinterpret_as_u32(a))); }
-
-inline v_float32x4 v_reverse(const v_float32x4 &a)
-{ return v_reinterpret_as_f32(v_reverse(v_reinterpret_as_u32(a))); }
-
-inline v_uint64x2 v_reverse(const v_uint64x2 &a)
-{
-    static const vec_uchar16 perm = {8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 4, 5, 6, 7};
-    vec_uchar16 vec = (vec_uchar16)a.val;
-    return v_reinterpret_as_u64(v_uint8x16(vec_perm(vec, vec, perm)));
-}
-
-inline v_int64x2 v_reverse(const v_int64x2 &a)
-{ return v_reinterpret_as_s64(v_reverse(v_reinterpret_as_u64(a))); }
-
-inline v_float64x2 v_reverse(const v_float64x2 &a)
-{ return v_reinterpret_as_f64(v_reverse(v_reinterpret_as_u64(a))); }
-
-/* Extract */
-template<int s, typename _Tpvec>
-inline _Tpvec v_extract(const _Tpvec& a, const _Tpvec& b)
-{ return v_rotate_right<s>(a, b); }
-
-////////// Reduce and mask /////////
-
-/** Reduce **/
-inline uint v_reduce_sum(const v_uint8x16& a)
-{
-    const vec_uint4 zero4 = vec_uint4_z;
-    vec_uint4 sum4 = vec_sum4s(a.val, zero4);
-    return (uint)vec_extract(vec_sums(vec_int4_c(sum4), vec_int4_c(zero4)), 3);
-}
-inline int v_reduce_sum(const v_int8x16& a)
-{
-    const vec_int4 zero4 = vec_int4_z;
-    vec_int4 sum4 = vec_sum4s(a.val, zero4);
-    return (int)vec_extract(vec_sums(sum4, zero4), 3);
-}
-inline int v_reduce_sum(const v_int16x8& a)
-{
-    const vec_int4 zero = vec_int4_z;
-    return saturate_cast<int>(vec_extract(vec_sums(vec_sum4s(a.val, zero), zero), 3));
-}
-inline uint v_reduce_sum(const v_uint16x8& a)
-{
-    const vec_int4 v4 = vec_int4_c(vec_unpackhu(vec_adds(a.val, vec_sld(a.val, a.val, 8))));
-    return saturate_cast<uint>(vec_extract(vec_sums(v4, vec_int4_z), 3));
-}
-
-#define OPENCV_HAL_IMPL_VSX_REDUCE_OP_4(_Tpvec, _Tpvec2, scalartype, suffix, func) \
-inline scalartype v_reduce_##suffix(const _Tpvec& a)                               \
-{                                                                                  \
-    const _Tpvec2 rs = func(a.val, vec_sld(a.val, a.val, 8));                      \
-    return vec_extract(func(rs, vec_sld(rs, rs, 4)), 0);                           \
-}
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_4(v_uint32x4, vec_uint4, uint, sum, vec_add)
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_4(v_uint32x4, vec_uint4, uint, max, vec_max)
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_4(v_uint32x4, vec_uint4, uint, min, vec_min)
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_4(v_int32x4, vec_int4, int, sum, vec_add)
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_4(v_int32x4, vec_int4, int, max, vec_max)
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_4(v_int32x4, vec_int4, int, min, vec_min)
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_4(v_float32x4, vec_float4, float, sum, vec_add)
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_4(v_float32x4, vec_float4, float, max, vec_max)
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_4(v_float32x4, vec_float4, float, min, vec_min)
-
-inline uint64 v_reduce_sum(const v_uint64x2& a)
-{
-    return vec_extract(vec_add(a.val, vec_permi(a.val, a.val, 3)), 0);
-}
-inline int64 v_reduce_sum(const v_int64x2& a)
-{
-    return vec_extract(vec_add(a.val, vec_permi(a.val, a.val, 3)), 0);
-}
-inline double v_reduce_sum(const v_float64x2& a)
-{
-    return vec_extract(vec_add(a.val, vec_permi(a.val, a.val, 3)), 0);
-}
-
-#define OPENCV_HAL_IMPL_VSX_REDUCE_OP_8(_Tpvec, _Tpvec2, scalartype, suffix, func) \
-inline scalartype v_reduce_##suffix(const _Tpvec& a)                               \
-{                                                                                  \
-    _Tpvec2 rs = func(a.val, vec_sld(a.val, a.val, 8));                            \
-    rs = func(rs, vec_sld(rs, rs, 4));                                             \
-    return vec_extract(func(rs, vec_sld(rs, rs, 2)), 0);                           \
-}
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_8(v_uint16x8, vec_ushort8, ushort, max, vec_max)
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_8(v_uint16x8, vec_ushort8, ushort, min, vec_min)
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_8(v_int16x8, vec_short8, short, max, vec_max)
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_8(v_int16x8, vec_short8, short, min, vec_min)
-
-#define OPENCV_HAL_IMPL_VSX_REDUCE_OP_16(_Tpvec, _Tpvec2, scalartype, suffix, func) \
-inline scalartype v_reduce_##suffix(const _Tpvec& a)                               \
-{                                                                                  \
-    _Tpvec2 rs = func(a.val, vec_sld(a.val, a.val, 8));                            \
-    rs = func(rs, vec_sld(rs, rs, 4));                                             \
-    rs = func(rs, vec_sld(rs, rs, 2));                                             \
-    return vec_extract(func(rs, vec_sld(rs, rs, 1)), 0);                           \
-}
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_16(v_uint8x16, vec_uchar16, uchar, max, vec_max)
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_16(v_uint8x16, vec_uchar16, uchar, min, vec_min)
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_16(v_int8x16, vec_char16, schar, max, vec_max)
-OPENCV_HAL_IMPL_VSX_REDUCE_OP_16(v_int8x16, vec_char16, schar, min, vec_min)
-
-inline v_float32x4 v_reduce_sum4(const v_float32x4& a, const v_float32x4& b,
-                                 const v_float32x4& c, const v_float32x4& d)
-{
-    vec_float4 ac = vec_add(vec_mergel(a.val, c.val), vec_mergeh(a.val, c.val));
-    ac = vec_add(ac, vec_sld(ac, ac, 8));
-
-    vec_float4 bd = vec_add(vec_mergel(b.val, d.val), vec_mergeh(b.val, d.val));
-    bd = vec_add(bd, vec_sld(bd, bd, 8));
-    return v_float32x4(vec_mergeh(ac, bd));
-}
-
-inline unsigned v_reduce_sad(const v_uint8x16& a, const v_uint8x16& b)
-{
-    const vec_uint4 zero4 = vec_uint4_z;
-    vec_uint4 sum4 = vec_sum4s(vec_absd(a.val, b.val), zero4);
-    return (unsigned)vec_extract(vec_sums(vec_int4_c(sum4), vec_int4_c(zero4)), 3);
-}
-inline unsigned v_reduce_sad(const v_int8x16& a, const v_int8x16& b)
-{
-    const vec_int4 zero4 = vec_int4_z;
-    vec_char16 ad = vec_abss(vec_subs(a.val, b.val));
-    vec_int4 sum4 = vec_sum4s(ad, zero4);
-    return (unsigned)vec_extract(vec_sums(sum4, zero4), 3);
-}
-inline unsigned v_reduce_sad(const v_uint16x8& a, const v_uint16x8& b)
-{
-    vec_ushort8 ad = vec_absd(a.val, b.val);
-    VSX_UNUSED(vec_int4) sum = vec_sums(vec_int4_c(vec_unpackhu(ad)) + vec_int4_c(vec_unpacklu(ad)), vec_int4_z);
-    return (unsigned)vec_extract(sum, 3);
-}
-inline unsigned v_reduce_sad(const v_int16x8& a, const v_int16x8& b)
-{
-    const vec_int4 zero4 = vec_int4_z;
-    vec_short8 ad = vec_abss(vec_subs(a.val, b.val));
-    vec_int4 sum4 = vec_sum4s(ad, zero4);
-    return (unsigned)vec_extract(vec_sums(sum4, zero4), 3);
-}
-inline unsigned v_reduce_sad(const v_uint32x4& a, const v_uint32x4& b)
-{
-    const vec_uint4 ad = vec_absd(a.val, b.val);
-    const vec_uint4 rd = vec_add(ad, vec_sld(ad, ad, 8));
-    return vec_extract(vec_add(rd, vec_sld(rd, rd, 4)), 0);
-}
-inline unsigned v_reduce_sad(const v_int32x4& a, const v_int32x4& b)
-{
-    vec_int4 ad = vec_abss(vec_sub(a.val, b.val));
-    return (unsigned)vec_extract(vec_sums(ad, vec_int4_z), 3);
-}
-inline float v_reduce_sad(const v_float32x4& a, const v_float32x4& b)
-{
-    const vec_float4 ad = vec_abs(vec_sub(a.val, b.val));
-    const vec_float4 rd = vec_add(ad, vec_sld(ad, ad, 8));
-    return vec_extract(vec_add(rd, vec_sld(rd, rd, 4)), 0);
-}
-
-/** Popcount **/
-inline v_uint8x16 v_popcount(const v_uint8x16& a)
-{ return v_uint8x16(vec_popcntu(a.val)); }
-inline v_uint8x16 v_popcount(const v_int8x16& a)
-{ return v_uint8x16(vec_popcntu(a.val)); }
-inline v_uint16x8 v_popcount(const v_uint16x8& a)
-{ return v_uint16x8(vec_popcntu(a.val)); }
-inline v_uint16x8 v_popcount(const v_int16x8& a)
-{ return v_uint16x8(vec_popcntu(a.val)); }
-inline v_uint32x4 v_popcount(const v_uint32x4& a)
-{ return v_uint32x4(vec_popcntu(a.val)); }
-inline v_uint32x4 v_popcount(const v_int32x4& a)
-{ return v_uint32x4(vec_popcntu(a.val)); }
-inline v_uint64x2 v_popcount(const v_uint64x2& a)
-{ return v_uint64x2(vec_popcntu(a.val)); }
-inline v_uint64x2 v_popcount(const v_int64x2& a)
-{ return v_uint64x2(vec_popcntu(a.val)); }
-
-/** Mask **/
-inline int v_signmask(const v_uint8x16& a)
-{
-    static const vec_uchar16 qperm = {120, 112, 104, 96, 88, 80, 72, 64, 56, 48, 40, 32, 24, 16, 8, 0};
-    return vec_extract((vec_int4)vec_vbpermq(v_reinterpret_as_u8(a).val, qperm), 2);
-}
-inline int v_signmask(const v_int8x16& a)
-{ return v_signmask(v_reinterpret_as_u8(a)); }
-
-inline int v_signmask(const v_int16x8& a)
-{
-    static const vec_uchar16 qperm = {112, 96, 80, 64, 48, 32, 16, 0, 128, 128, 128, 128, 128, 128, 128, 128};
-    return vec_extract((vec_int4)vec_vbpermq(v_reinterpret_as_u8(a).val, qperm), 2);
-}
-inline int v_signmask(const v_uint16x8& a)
-{ return v_signmask(v_reinterpret_as_s16(a)); }
-
-inline int v_signmask(const v_int32x4& a)
-{
-    static const vec_uchar16 qperm = {96, 64, 32, 0, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128};
-    return vec_extract((vec_int4)vec_vbpermq(v_reinterpret_as_u8(a).val, qperm), 2);
-}
-inline int v_signmask(const v_uint32x4& a)
-{ return v_signmask(v_reinterpret_as_s32(a)); }
-inline int v_signmask(const v_float32x4& a)
-{ return v_signmask(v_reinterpret_as_s32(a)); }
-
-inline int v_signmask(const v_int64x2& a)
-{
-    VSX_UNUSED(const vec_dword2) sv = vec_sr(a.val, vec_udword2_sp(63));
-    return (int)vec_extract(sv, 0) | (int)vec_extract(sv, 1) << 1;
-}
-inline int v_signmask(const v_uint64x2& a)
-{ return v_signmask(v_reinterpret_as_s64(a)); }
-inline int v_signmask(const v_float64x2& a)
-{ return v_signmask(v_reinterpret_as_s64(a)); }
-
-inline int v_scan_forward(const v_int8x16& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_uint8x16& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_int16x8& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_uint16x8& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_int32x4& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_uint32x4& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_float32x4& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_int64x2& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_uint64x2& a) { return trailingZeros32(v_signmask(a)); }
-inline int v_scan_forward(const v_float64x2& a) { return trailingZeros32(v_signmask(a)); }
-
-template<typename _Tpvec>
-inline bool v_check_all(const _Tpvec& a)
-{ return vec_all_lt(a.val, _Tpvec::zero().val); }
-inline bool v_check_all(const v_uint8x16& a)
-{ return v_check_all(v_reinterpret_as_s8(a)); }
-inline bool v_check_all(const v_uint16x8& a)
-{ return v_check_all(v_reinterpret_as_s16(a)); }
-inline bool v_check_all(const v_uint32x4& a)
-{ return v_check_all(v_reinterpret_as_s32(a)); }
-inline bool v_check_all(const v_uint64x2& a)
-{ return v_check_all(v_reinterpret_as_s64(a)); }
-inline bool v_check_all(const v_float32x4& a)
-{ return v_check_all(v_reinterpret_as_s32(a)); }
-inline bool v_check_all(const v_float64x2& a)
-{ return v_check_all(v_reinterpret_as_s64(a)); }
-
-template<typename _Tpvec>
-inline bool v_check_any(const _Tpvec& a)
-{ return vec_any_lt(a.val, _Tpvec::zero().val); }
-inline bool v_check_any(const v_uint8x16& a)
-{ return v_check_any(v_reinterpret_as_s8(a)); }
-inline bool v_check_any(const v_uint16x8& a)
-{ return v_check_any(v_reinterpret_as_s16(a)); }
-inline bool v_check_any(const v_uint32x4& a)
-{ return v_check_any(v_reinterpret_as_s32(a)); }
-inline bool v_check_any(const v_uint64x2& a)
-{ return v_check_any(v_reinterpret_as_s64(a)); }
-inline bool v_check_any(const v_float32x4& a)
-{ return v_check_any(v_reinterpret_as_s32(a)); }
-inline bool v_check_any(const v_float64x2& a)
-{ return v_check_any(v_reinterpret_as_s64(a)); }
-
-////////// Other math /////////
-
-/** Some frequent operations **/
-inline v_float32x4 v_sqrt(const v_float32x4& x)
-{ return v_float32x4(vec_sqrt(x.val)); }
-inline v_float64x2 v_sqrt(const v_float64x2& x)
-{ return v_float64x2(vec_sqrt(x.val)); }
-
-inline v_float32x4 v_invsqrt(const v_float32x4& x)
-{ return v_float32x4(vec_rsqrt(x.val)); }
-inline v_float64x2 v_invsqrt(const v_float64x2& x)
-{ return v_float64x2(vec_rsqrt(x.val)); }
-
-#define OPENCV_HAL_IMPL_VSX_MULADD(_Tpvec)                                  \
-inline _Tpvec v_magnitude(const _Tpvec& a, const _Tpvec& b)                 \
-{ return _Tpvec(vec_sqrt(vec_madd(a.val, a.val, vec_mul(b.val, b.val)))); } \
-inline _Tpvec v_sqr_magnitude(const _Tpvec& a, const _Tpvec& b)             \
-{ return _Tpvec(vec_madd(a.val, a.val, vec_mul(b.val, b.val))); }           \
-inline _Tpvec v_fma(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c)      \
-{ return _Tpvec(vec_madd(a.val, b.val, c.val)); }                           \
-inline _Tpvec v_muladd(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c)   \
-{ return _Tpvec(vec_madd(a.val, b.val, c.val)); }
-
-OPENCV_HAL_IMPL_VSX_MULADD(v_float32x4)
-OPENCV_HAL_IMPL_VSX_MULADD(v_float64x2)
-
-inline v_int32x4 v_muladd(const v_int32x4& a, const v_int32x4& b, const v_int32x4& c)
-{ return a * b + c; }
-
-// TODO: exp, log, sin, cos
-
-/** Absolute values **/
-inline v_uint8x16 v_abs(const v_int8x16& x)
-{ return v_uint8x16(vec_uchar16_c(vec_abs(x.val))); }
-
-inline v_uint16x8 v_abs(const v_int16x8& x)
-{ return v_uint16x8(vec_ushort8_c(vec_abs(x.val))); }
-
-inline v_uint32x4 v_abs(const v_int32x4& x)
-{ return v_uint32x4(vec_uint4_c(vec_abs(x.val))); }
-
-inline v_float32x4 v_abs(const v_float32x4& x)
-{ return v_float32x4(vec_abs(x.val)); }
-
-inline v_float64x2 v_abs(const v_float64x2& x)
-{ return v_float64x2(vec_abs(x.val)); }
-
-/** Absolute difference **/
-// unsigned
-OPENCV_HAL_IMPL_VSX_BIN_FUNC(v_absdiff, vec_absd)
-
-inline v_uint8x16 v_absdiff(const v_int8x16& a, const v_int8x16& b)
-{ return v_reinterpret_as_u8(v_sub_wrap(v_max(a, b), v_min(a, b))); }
-inline v_uint16x8 v_absdiff(const v_int16x8& a, const v_int16x8& b)
-{ return v_reinterpret_as_u16(v_sub_wrap(v_max(a, b), v_min(a, b))); }
-inline v_uint32x4 v_absdiff(const v_int32x4& a, const v_int32x4& b)
-{ return v_reinterpret_as_u32(v_max(a, b) - v_min(a, b)); }
-
-inline v_float32x4 v_absdiff(const v_float32x4& a, const v_float32x4& b)
-{ return v_abs(a - b); }
-inline v_float64x2 v_absdiff(const v_float64x2& a, const v_float64x2& b)
-{ return v_abs(a - b); }
-
-/** Absolute difference for signed integers **/
-inline v_int8x16 v_absdiffs(const v_int8x16& a, const v_int8x16& b)
-{ return v_int8x16(vec_abss(vec_subs(a.val, b.val))); }
-inline v_int16x8 v_absdiffs(const v_int16x8& a, const v_int16x8& b)
-{ return v_int16x8(vec_abss(vec_subs(a.val, b.val))); }
-
-////////// Conversions /////////
-
-/** Rounding **/
-inline v_int32x4 v_round(const v_float32x4& a)
-{ return v_int32x4(vec_cts(vec_rint(a.val))); }
-
-inline v_int32x4 v_round(const v_float64x2& a)
-{ return v_int32x4(vec_mergesqo(vec_ctso(vec_rint(a.val)), vec_int4_z)); }
-
-inline v_int32x4 v_round(const v_float64x2& a, const v_float64x2& b)
-{ return v_int32x4(vec_mergesqo(vec_ctso(vec_rint(a.val)), vec_ctso(vec_rint(b.val)))); }
-
-inline v_int32x4 v_floor(const v_float32x4& a)
-{ return v_int32x4(vec_cts(vec_floor(a.val))); }
-
-inline v_int32x4 v_floor(const v_float64x2& a)
-{ return v_int32x4(vec_mergesqo(vec_ctso(vec_floor(a.val)), vec_int4_z)); }
-
-inline v_int32x4 v_ceil(const v_float32x4& a)
-{ return v_int32x4(vec_cts(vec_ceil(a.val))); }
-
-inline v_int32x4 v_ceil(const v_float64x2& a)
-{ return v_int32x4(vec_mergesqo(vec_ctso(vec_ceil(a.val)), vec_int4_z)); }
-
-inline v_int32x4 v_trunc(const v_float32x4& a)
-{ return v_int32x4(vec_cts(a.val)); }
-
-inline v_int32x4 v_trunc(const v_float64x2& a)
-{ return v_int32x4(vec_mergesqo(vec_ctso(a.val), vec_int4_z)); }
-
-/** To float **/
-inline v_float32x4 v_cvt_f32(const v_int32x4& a)
-{ return v_float32x4(vec_ctf(a.val)); }
-
-inline v_float32x4 v_cvt_f32(const v_float64x2& a)
-{ return v_float32x4(vec_mergesqo(vec_cvfo(a.val), vec_float4_z)); }
-
-inline v_float32x4 v_cvt_f32(const v_float64x2& a, const v_float64x2& b)
-{ return v_float32x4(vec_mergesqo(vec_cvfo(a.val), vec_cvfo(b.val))); }
-
-inline v_float64x2 v_cvt_f64(const v_int32x4& a)
-{ return v_float64x2(vec_ctdo(vec_mergeh(a.val, a.val))); }
-
-inline v_float64x2 v_cvt_f64_high(const v_int32x4& a)
-{ return v_float64x2(vec_ctdo(vec_mergel(a.val, a.val))); }
-
-inline v_float64x2 v_cvt_f64(const v_float32x4& a)
-{ return v_float64x2(vec_cvfo(vec_mergeh(a.val, a.val))); }
-
-inline v_float64x2 v_cvt_f64_high(const v_float32x4& a)
-{ return v_float64x2(vec_cvfo(vec_mergel(a.val, a.val))); }
-
-inline v_float64x2 v_cvt_f64(const v_int64x2& a)
-{ return v_float64x2(vec_ctd(a.val)); }
-
-////////////// Lookup table access ////////////////////
-
-inline v_int8x16 v_lut(const schar* tab, const int* idx)
-{
-    return v_int8x16(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]], tab[idx[4]], tab[idx[5]], tab[idx[6]], tab[idx[7]],
-                     tab[idx[8]], tab[idx[9]], tab[idx[10]], tab[idx[11]], tab[idx[12]], tab[idx[13]], tab[idx[14]], tab[idx[15]]);
-}
-inline v_int8x16 v_lut_pairs(const schar* tab, const int* idx)
-{
-    return v_reinterpret_as_s8(v_int16x8(*(const short*)(tab+idx[0]), *(const short*)(tab+idx[1]), *(const short*)(tab+idx[2]), *(const short*)(tab+idx[3]),
-                                       *(const short*)(tab+idx[4]), *(const short*)(tab+idx[5]), *(const short*)(tab+idx[6]), *(const short*)(tab+idx[7])));
-}
-inline v_int8x16 v_lut_quads(const schar* tab, const int* idx)
-{
-    return v_reinterpret_as_s8(v_int32x4(*(const int*)(tab+idx[0]), *(const int*)(tab+idx[1]), *(const int*)(tab+idx[2]), *(const int*)(tab+idx[3])));
-}
-inline v_uint8x16 v_lut(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut((const schar*)tab, idx)); }
-inline v_uint8x16 v_lut_pairs(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut_pairs((const schar*)tab, idx)); }
-inline v_uint8x16 v_lut_quads(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut_quads((const schar*)tab, idx)); }
-
-inline v_int16x8 v_lut(const short* tab, const int* idx)
-{
-    return v_int16x8(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]], tab[idx[4]], tab[idx[5]], tab[idx[6]], tab[idx[7]]);
-}
-inline v_int16x8 v_lut_pairs(const short* tab, const int* idx)
-{
-    return v_reinterpret_as_s16(v_int32x4(*(const int*)(tab + idx[0]), *(const int*)(tab + idx[1]), *(const int*)(tab + idx[2]), *(const int*)(tab + idx[3])));
-}
-inline v_int16x8 v_lut_quads(const short* tab, const int* idx)
-{
-    return v_reinterpret_as_s16(v_int64x2(*(const int64*)(tab + idx[0]), *(const int64*)(tab + idx[1])));
-}
-inline v_uint16x8 v_lut(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut((const short*)tab, idx)); }
-inline v_uint16x8 v_lut_pairs(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut_pairs((const short*)tab, idx)); }
-inline v_uint16x8 v_lut_quads(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut_quads((const short*)tab, idx)); }
-
-inline v_int32x4 v_lut(const int* tab, const int* idx)
-{
-    return v_int32x4(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]]);
-}
-inline v_int32x4 v_lut_pairs(const int* tab, const int* idx)
-{
-    return v_reinterpret_as_s32(v_int64x2(*(const int64*)(tab + idx[0]), *(const int64*)(tab + idx[1])));
-}
-inline v_int32x4 v_lut_quads(const int* tab, const int* idx)
-{
-    return v_int32x4(vsx_ld(0, tab + idx[0]));
-}
-inline v_uint32x4 v_lut(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut((const int*)tab, idx)); }
-inline v_uint32x4 v_lut_pairs(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut_pairs((const int*)tab, idx)); }
-inline v_uint32x4 v_lut_quads(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut_quads((const int*)tab, idx)); }
-
-inline v_int64x2 v_lut(const int64_t* tab, const int* idx)
-{
-    return v_int64x2(tab[idx[0]], tab[idx[1]]);
-}
-inline v_int64x2 v_lut_pairs(const int64_t* tab, const int* idx)
-{
-    return v_int64x2(vsx_ld2(0, tab + idx[0]));
-}
-inline v_uint64x2 v_lut(const uint64_t* tab, const int* idx) { return v_reinterpret_as_u64(v_lut((const int64_t *)tab, idx)); }
-inline v_uint64x2 v_lut_pairs(const uint64_t* tab, const int* idx) { return v_reinterpret_as_u64(v_lut_pairs((const int64_t *)tab, idx)); }
-
-inline v_float32x4 v_lut(const float* tab, const int* idx)
-{
-    return v_float32x4(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]]);
-}
-inline v_float32x4 v_lut_pairs(const float* tab, const int* idx) { return v_reinterpret_as_f32(v_lut_pairs((const int*)tab, idx)); }
-inline v_float32x4 v_lut_quads(const float* tab, const int* idx) { return v_load(tab + *idx); }
-
-inline v_float64x2 v_lut(const double* tab, const int* idx)
-{
-    return v_float64x2(tab[idx[0]], tab[idx[1]]);
-}
-inline v_float64x2 v_lut_pairs(const double* tab, const int* idx) { return v_load(tab + *idx); }
-
-inline v_int32x4 v_lut(const int* tab, const v_int32x4& idxvec)
-{
-    const int idx[4] = {
-        vec_extract(idxvec.val, 0),
-        vec_extract(idxvec.val, 1),
-        vec_extract(idxvec.val, 2),
-        vec_extract(idxvec.val, 3)
-    };
-    return v_int32x4(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]]);
-}
-
-inline v_uint32x4 v_lut(const unsigned* tab, const v_int32x4& idxvec)
-{
-    const int idx[4] = {
-        vec_extract(idxvec.val, 0),
-        vec_extract(idxvec.val, 1),
-        vec_extract(idxvec.val, 2),
-        vec_extract(idxvec.val, 3)
-    };
-    return v_uint32x4(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]]);
-}
-
-inline v_float32x4 v_lut(const float* tab, const v_int32x4& idxvec)
-{
-    const int idx[4] = {
-        vec_extract(idxvec.val, 0),
-        vec_extract(idxvec.val, 1),
-        vec_extract(idxvec.val, 2),
-        vec_extract(idxvec.val, 3)
-    };
-    return v_float32x4(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]]);
-}
-
-inline v_float64x2 v_lut(const double* tab, const v_int32x4& idxvec)
-{
-    const int idx[2] = {
-        vec_extract(idxvec.val, 0),
-        vec_extract(idxvec.val, 1)
-    };
-    return v_float64x2(tab[idx[0]], tab[idx[1]]);
-}
-
-inline void v_lut_deinterleave(const float* tab, const v_int32x4& idxvec, v_float32x4& x, v_float32x4& y)
-{
-    vec_float4 xy0 = vec_ld_l8(tab + vec_extract(idxvec.val, 0));
-    vec_float4 xy1 = vec_ld_l8(tab + vec_extract(idxvec.val, 1));
-    vec_float4 xy2 = vec_ld_l8(tab + vec_extract(idxvec.val, 2));
-    vec_float4 xy3 = vec_ld_l8(tab + vec_extract(idxvec.val, 3));
-    vec_float4 xy02 = vec_mergeh(xy0, xy2); // x0, x2, y0, y2
-    vec_float4 xy13 = vec_mergeh(xy1, xy3); // x1, x3, y1, y3
-    x.val = vec_mergeh(xy02, xy13);
-    y.val = vec_mergel(xy02, xy13);
-}
-inline void v_lut_deinterleave(const double* tab, const v_int32x4& idxvec, v_float64x2& x, v_float64x2& y)
-{
-    vec_double2 xy0 = vsx_ld(vec_extract(idxvec.val, 0), tab);
-    vec_double2 xy1 = vsx_ld(vec_extract(idxvec.val, 1), tab);
-    x.val = vec_mergeh(xy0, xy1);
-    y.val = vec_mergel(xy0, xy1);
-}
-
-inline v_int8x16 v_interleave_pairs(const v_int8x16& vec)
-{
-    static const vec_uchar16 perm = {0, 2, 1, 3, 4, 6, 5, 7, 8, 10, 9, 11, 12, 14, 13, 15};
-    return v_int8x16(vec_perm(vec.val, vec.val, perm));
-}
-inline v_uint8x16 v_interleave_pairs(const v_uint8x16& vec)
-{ return v_reinterpret_as_u8(v_interleave_pairs(v_reinterpret_as_s8(vec))); }
-
-inline v_int8x16 v_interleave_quads(const v_int8x16& vec)
-{
-    static const vec_uchar16 perm = {0, 4, 1, 5, 2, 6, 3, 7, 8, 12, 9, 13, 10, 14, 11, 15};
-    return v_int8x16(vec_perm(vec.val, vec.val, perm));
-}
-inline v_uint8x16 v_interleave_quads(const v_uint8x16& vec)
-{ return v_reinterpret_as_u8(v_interleave_quads(v_reinterpret_as_s8(vec))); }
-
-inline v_int16x8 v_interleave_pairs(const v_int16x8& vec)
-{
-    static const vec_uchar16 perm = {0,1, 4,5, 2,3, 6,7, 8,9, 12,13, 10,11, 14,15};
-    return v_int16x8(vec_perm(vec.val, vec.val, perm));
-}
-inline v_uint16x8 v_interleave_pairs(const v_uint16x8& vec)
-{ return v_reinterpret_as_u16(v_interleave_pairs(v_reinterpret_as_s16(vec))); }
-
-inline v_int16x8 v_interleave_quads(const v_int16x8& vec)
-{
-    static const vec_uchar16 perm = {0,1, 8,9, 2,3, 10,11, 4,5, 12,13, 6,7, 14,15};
-    return v_int16x8(vec_perm(vec.val, vec.val, perm));
-}
-inline v_uint16x8 v_interleave_quads(const v_uint16x8& vec)
-{ return v_reinterpret_as_u16(v_interleave_quads(v_reinterpret_as_s16(vec))); }
-
-inline v_int32x4 v_interleave_pairs(const v_int32x4& vec)
-{
-    static const vec_uchar16 perm = {0,1,2,3, 8,9,10,11, 4,5,6,7, 12,13,14,15};
-    return v_int32x4(vec_perm(vec.val, vec.val, perm));
-}
-inline v_uint32x4 v_interleave_pairs(const v_uint32x4& vec)
-{ return v_reinterpret_as_u32(v_interleave_pairs(v_reinterpret_as_s32(vec))); }
-inline v_float32x4 v_interleave_pairs(const v_float32x4& vec)
-{ return v_reinterpret_as_f32(v_interleave_pairs(v_reinterpret_as_s32(vec))); }
-
-inline v_int8x16 v_pack_triplets(const v_int8x16& vec)
-{
-    static const vec_uchar16 perm = {0, 1, 2, 4, 5, 6, 8, 9, 10, 12, 13, 14, 15, 15, 15, 15};
-    return v_int8x16(vec_perm(vec.val, vec.val, perm));
-}
-inline v_uint8x16 v_pack_triplets(const v_uint8x16& vec)
-{ return v_reinterpret_as_u8(v_pack_triplets(v_reinterpret_as_s8(vec))); }
-
-inline v_int16x8 v_pack_triplets(const v_int16x8& vec)
-{
-    static const vec_uchar16 perm = {0,1, 2,3, 4,5, 8,9, 10,11, 12,13, 14,15, 14,15};
-    return v_int16x8(vec_perm(vec.val, vec.val, perm));
-}
-inline v_uint16x8 v_pack_triplets(const v_uint16x8& vec)
-{ return v_reinterpret_as_u16(v_pack_triplets(v_reinterpret_as_s16(vec))); }
-
-inline v_int32x4 v_pack_triplets(const v_int32x4& vec)
-{ return vec; }
-inline v_uint32x4 v_pack_triplets(const v_uint32x4& vec)
-{ return vec; }
-inline v_float32x4 v_pack_triplets(const v_float32x4& vec)
-{ return vec; }
-
-/////// FP16 support ////////
-
-inline v_float32x4 v_load_expand(const float16_t* ptr)
-{
-    vec_ushort8 vf16 = vec_ld_l8((const ushort*)ptr);
-#if CV_VSX3 && defined(vec_extract_fp_from_shorth)
-    return v_float32x4(vec_extract_fp_from_shorth(vf16));
-#elif CV_VSX3 && !defined(CV_COMPILER_VSX_BROKEN_ASM)
-    vec_float4 vf32;
-    __asm__ __volatile__ ("xvcvhpsp %x0,%x1" : "=wa" (vf32) : "wa" (vec_mergeh(vf16, vf16)));
-    return v_float32x4(vf32);
-#else
-    const vec_int4 z = vec_int4_z, delta = vec_int4_sp(0x38000000);
-    const vec_int4 signmask = vec_int4_sp(0x80000000);
-    const vec_int4 maxexp = vec_int4_sp(0x7c000000);
-    const vec_float4 deltaf = vec_float4_c(vec_int4_sp(0x38800000));
-
-    vec_int4 bits = vec_int4_c(vec_mergeh(vec_short8_c(z), vec_short8_c(vf16)));
-    vec_int4 e = vec_and(bits, maxexp), sign = vec_and(bits, signmask);
-    vec_int4 t = vec_add(vec_sr(vec_xor(bits, sign), vec_uint4_sp(3)), delta); // ((h & 0x7fff) << 13) + delta
-    vec_int4 zt = vec_int4_c(vec_sub(vec_float4_c(vec_add(t, vec_int4_sp(1 << 23))), deltaf));
-
-    t = vec_add(t, vec_and(delta, vec_cmpeq(maxexp, e)));
-    vec_bint4 zmask = vec_cmpeq(e, z);
-    vec_int4 ft = vec_sel(t, zt, zmask);
-    return v_float32x4(vec_float4_c(vec_or(ft, sign)));
-#endif
-}
-
-inline void v_pack_store(float16_t* ptr, const v_float32x4& v)
-{
-// fixme: Is there any builtin op or intrinsic that cover "xvcvsphp"?
-#if CV_VSX3 && !defined(CV_COMPILER_VSX_BROKEN_ASM)
-    vec_ushort8 vf16;
-    __asm__ __volatile__ ("xvcvsphp %x0,%x1" : "=wa" (vf16) : "wa" (v.val));
-    vec_st_l8(vec_mergesqe(vf16, vf16), ptr);
-#else
-    const vec_int4 signmask = vec_int4_sp(0x80000000);
-    const vec_int4 rval = vec_int4_sp(0x3f000000);
-
-    vec_int4 t = vec_int4_c(v.val);
-    vec_int4 sign = vec_sra(vec_and(t, signmask), vec_uint4_sp(16));
-    t = vec_and(vec_nor(signmask, signmask), t);
-
-    vec_bint4 finitemask = vec_cmpgt(vec_int4_sp(0x47800000), t);
-    vec_bint4 isnan = vec_cmpgt(t, vec_int4_sp(0x7f800000));
-    vec_int4 naninf = vec_sel(vec_int4_sp(0x7c00), vec_int4_sp(0x7e00), isnan);
-    vec_bint4 tinymask = vec_cmpgt(vec_int4_sp(0x38800000), t);
-    vec_int4 tt = vec_int4_c(vec_add(vec_float4_c(t), vec_float4_c(rval)));
-    tt = vec_sub(tt, rval);
-    vec_int4 odd = vec_and(vec_sr(t, vec_uint4_sp(13)), vec_int4_sp(1));
-    vec_int4 nt = vec_add(t, vec_int4_sp(0xc8000fff));
-    nt = vec_sr(vec_add(nt, odd), vec_uint4_sp(13));
-    t = vec_sel(nt, tt, tinymask);
-    t = vec_sel(naninf, t, finitemask);
-    t = vec_or(t, sign);
-    vec_st_l8(vec_packs(t, t), ptr);
-#endif
-}
-
-inline void v_cleanup() {}
-
-
-/** Reinterpret **/
-/** its up there with load and store operations **/
-
-////////// Matrix operations /////////
-
-//////// Dot Product ////////
-// 16 >> 32
-inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b)
-{ return v_int32x4(vec_msum(a.val, b.val, vec_int4_z)); }
-inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b, const v_int32x4& c)
-{ return v_int32x4(vec_msum(a.val, b.val, c.val)); }
-
-// 32 >> 64
-inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b)
-{
-    vec_dword2 even = vec_mule(a.val, b.val);
-    vec_dword2 odd = vec_mulo(a.val, b.val);
-    return v_int64x2(vec_add(even, odd));
-}
-inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b, const v_int64x2& c)
-{ return v_dotprod(a, b) + c; }
-
-// 8 >> 32
-inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b, const v_uint32x4& c)
-{ return v_uint32x4(vec_msum(a.val, b.val, c.val)); }
-inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b)
-{ return v_uint32x4(vec_msum(a.val, b.val, vec_uint4_z)); }
-
-inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b)
-{
-    const vec_ushort8 eight = vec_ushort8_sp(8);
-    vec_short8 a0 = vec_sra((vec_short8)vec_sld(a.val, a.val, 1), eight); // even
-    vec_short8 a1 = vec_sra((vec_short8)a.val, eight); // odd
-    vec_short8 b0 = vec_sra((vec_short8)vec_sld(b.val, b.val, 1), eight);
-    vec_short8 b1 = vec_sra((vec_short8)b.val, eight);
-    return v_int32x4(vec_msum(a0, b0, vec_msum(a1, b1, vec_int4_z)));
-}
-
-inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b, const v_int32x4& c)
-{
-    const vec_ushort8 eight = vec_ushort8_sp(8);
-    vec_short8 a0 = vec_sra((vec_short8)vec_sld(a.val, a.val, 1), eight); // even
-    vec_short8 a1 = vec_sra((vec_short8)a.val, eight); // odd
-    vec_short8 b0 = vec_sra((vec_short8)vec_sld(b.val, b.val, 1), eight);
-    vec_short8 b1 = vec_sra((vec_short8)b.val, eight);
-    return v_int32x4(vec_msum(a0, b0, vec_msum(a1, b1, c.val)));
-}
-
-// 16 >> 64
-inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b)
-{
-    const vec_uint4 zero = vec_uint4_z;
-    vec_uint4 even = vec_mule(a.val, b.val);
-    vec_uint4 odd  = vec_mulo(a.val, b.val);
-    vec_udword2 e0 = (vec_udword2)vec_mergee(even, zero);
-    vec_udword2 e1 = (vec_udword2)vec_mergeo(even, zero);
-    vec_udword2 o0 = (vec_udword2)vec_mergee(odd, zero);
-    vec_udword2 o1 = (vec_udword2)vec_mergeo(odd, zero);
-    vec_udword2 s0 = vec_add(e0, o0);
-    vec_udword2 s1 = vec_add(e1, o1);
-    return v_uint64x2(vec_add(s0, s1));
-}
-inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b, const v_uint64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-inline v_int64x2 v_dotprod_expand(const v_int16x8& a, const v_int16x8& b)
-{
-    v_int32x4 prod = v_dotprod(a, b);
-    v_int64x2 c, d;
-    v_expand(prod, c, d);
-    return v_int64x2(vec_add(vec_mergeh(c.val, d.val), vec_mergel(c.val, d.val)));
-}
-inline v_int64x2 v_dotprod_expand(const v_int16x8& a, const v_int16x8& b, const v_int64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-// 32 >> 64f
-inline v_float64x2 v_dotprod_expand(const v_int32x4& a, const v_int32x4& b)
-{ return v_cvt_f64(v_dotprod(a, b)); }
-inline v_float64x2 v_dotprod_expand(const v_int32x4& a, const v_int32x4& b, const v_float64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-//////// Fast Dot Product ////////
-
-// 16 >> 32
-inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b)
-{ return v_dotprod(a, b); }
-inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b, const v_int32x4& c)
-{ return v_int32x4(vec_msum(a.val, b.val, vec_int4_z)) + c; }
-// 32 >> 64
-inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b)
-{ return v_dotprod(a, b); }
-inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b, const v_int64x2& c)
-{ return v_dotprod(a, b, c); }
-
-// 8 >> 32
-inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b)
-{ return v_dotprod_expand(a, b); }
-inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b, const v_uint32x4& c)
-{ return v_uint32x4(vec_msum(a.val, b.val, vec_uint4_z)) + c; }
-
-inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b)
-{
-    vec_short8 a0 = vec_unpackh(a.val);
-    vec_short8 a1 = vec_unpackl(a.val);
-    vec_short8 b0 = vec_unpackh(b.val);
-    vec_short8 b1 = vec_unpackl(b.val);
-    return v_int32x4(vec_msum(a0, b0, vec_msum(a1, b1, vec_int4_z)));
-}
-inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b, const v_int32x4& c)
-{ return v_dotprod_expand_fast(a, b) + c; }
-
-// 16 >> 64
-inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b)
-{ return v_dotprod_expand(a, b); }
-inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b, const v_uint64x2& c)
-{ return v_dotprod_expand(a, b, c); }
-
-inline v_int64x2 v_dotprod_expand_fast(const v_int16x8& a, const v_int16x8& b)
-{
-    v_int32x4 prod = v_dotprod(a, b);
-    v_int64x2 c, d;
-    v_expand(prod, c, d);
-    return c + d;
-}
-inline v_int64x2 v_dotprod_expand_fast(const v_int16x8& a, const v_int16x8& b, const v_int64x2& c)
-{ return v_dotprod_expand_fast(a, b) + c; }
-
-// 32 >> 64f
-inline v_float64x2 v_dotprod_expand_fast(const v_int32x4& a, const v_int32x4& b)
-{ return v_dotprod_expand(a, b); }
-inline v_float64x2 v_dotprod_expand_fast(const v_int32x4& a, const v_int32x4& b, const v_float64x2& c)
-{ return v_dotprod_expand(a, b, c); }
-
-inline v_float32x4 v_matmul(const v_float32x4& v, const v_float32x4& m0,
-                            const v_float32x4& m1, const v_float32x4& m2,
-                            const v_float32x4& m3)
-{
-    const vec_float4 v0 = vec_splat(v.val, 0);
-    const vec_float4 v1 = vec_splat(v.val, 1);
-    const vec_float4 v2 = vec_splat(v.val, 2);
-    VSX_UNUSED(const vec_float4) v3 = vec_splat(v.val, 3);
-    return v_float32x4(vec_madd(v0, m0.val, vec_madd(v1, m1.val, vec_madd(v2, m2.val, vec_mul(v3, m3.val)))));
-}
-
-inline v_float32x4 v_matmuladd(const v_float32x4& v, const v_float32x4& m0,
-                               const v_float32x4& m1, const v_float32x4& m2,
-                               const v_float32x4& a)
-{
-    const vec_float4 v0 = vec_splat(v.val, 0);
-    const vec_float4 v1 = vec_splat(v.val, 1);
-    const vec_float4 v2 = vec_splat(v.val, 2);
-    return v_float32x4(vec_madd(v0, m0.val, vec_madd(v1, m1.val, vec_madd(v2, m2.val, a.val))));
-}
-
-#define OPENCV_HAL_IMPL_VSX_TRANSPOSE4x4(_Tpvec, _Tpvec2)                        \
-inline void v_transpose4x4(const _Tpvec& a0, const _Tpvec& a1,                   \
-                           const _Tpvec& a2, const _Tpvec& a3,                   \
-                           _Tpvec& b0, _Tpvec& b1, _Tpvec& b2, _Tpvec& b3)       \
-{                                                                                \
-    _Tpvec2 a02 = vec_mergeh(a0.val, a2.val);                                    \
-    _Tpvec2 a13 = vec_mergeh(a1.val, a3.val);                                    \
-    b0.val = vec_mergeh(a02, a13);                                               \
-    b1.val = vec_mergel(a02, a13);                                               \
-    a02 = vec_mergel(a0.val, a2.val);                                            \
-    a13 = vec_mergel(a1.val, a3.val);                                            \
-    b2.val  = vec_mergeh(a02, a13);                                              \
-    b3.val  = vec_mergel(a02, a13);                                              \
-}
-OPENCV_HAL_IMPL_VSX_TRANSPOSE4x4(v_uint32x4, vec_uint4)
-OPENCV_HAL_IMPL_VSX_TRANSPOSE4x4(v_int32x4, vec_int4)
-OPENCV_HAL_IMPL_VSX_TRANSPOSE4x4(v_float32x4, vec_float4)
-
-template<int i, typename Tvec>
-inline Tvec v_broadcast_element(const Tvec& v)
-{ return Tvec(vec_splat(v.val, i)); }
-
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END
-
-//! @endcond
-
-}
-
-#endif // OPENCV_HAL_VSX_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_wasm.hpp b/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_wasm.hpp
deleted file mode 100644
index b4178af8..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/intrin_wasm.hpp
+++ /dev/null
@@ -1,2782 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_HAL_INTRIN_WASM_HPP
-#define OPENCV_HAL_INTRIN_WASM_HPP
-
-#include <limits>
-#include <cstring>
-#include <algorithm>
-#include "opencv2/core/saturate.hpp"
-
-#define CV_SIMD128 1
-#define CV_SIMD128_64F 0 // Now all implementation of f64 use fallback, so disable it.
-#define CV_SIMD128_FP16 0
-
-namespace cv
-{
-
-//! @cond IGNORED
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
-
-#if (__EMSCRIPTEN_major__ * 1000000 + __EMSCRIPTEN_minor__ * 1000 + __EMSCRIPTEN_tiny__) < (1038046)
-// handle renames: https://github.com/emscripten-core/emscripten/pull/9440 (https://github.com/emscripten-core/emscripten/commit/755d5b46cb84d0aa120c10981b11d05646c29673)
-#define wasm_i32x4_trunc_saturate_f32x4 wasm_trunc_saturate_i32x4_f32x4
-#define wasm_u32x4_trunc_saturate_f32x4 wasm_trunc_saturate_u32x4_f32x4
-#define wasm_i64x2_trunc_saturate_f64x2 wasm_trunc_saturate_i64x2_f64x2
-#define wasm_u64x2_trunc_saturate_f64x2 wasm_trunc_saturate_u64x2_f64x2
-#define wasm_f32x4_convert_i32x4 wasm_convert_f32x4_i32x4
-#define wasm_f32x4_convert_u32x4 wasm_convert_f32x4_u32x4
-#define wasm_f64x2_convert_i64x2 wasm_convert_f64x2_i64x2
-#define wasm_f64x2_convert_u64x2 wasm_convert_f64x2_u64x2
-#endif // COMPATIBILITY: <1.38.46
-
-///////// Types ///////////
-
-struct v_uint8x16
-{
-    typedef uchar lane_type;
-    typedef v128_t vector_type;
-    enum { nlanes = 16 };
-
-    v_uint8x16() {}
-    explicit v_uint8x16(v128_t v) : val(v) {}
-    v_uint8x16(uchar v0, uchar v1, uchar v2, uchar v3, uchar v4, uchar v5, uchar v6, uchar v7,
-            uchar v8, uchar v9, uchar v10, uchar v11, uchar v12, uchar v13, uchar v14, uchar v15)
-    {
-        uchar v[] = {v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15};
-        val = wasm_v128_load(v);
-    }
-
-    uchar get0() const
-    {
-        return (uchar)wasm_i8x16_extract_lane(val, 0);
-    }
-
-    v128_t val;
-};
-
-struct v_int8x16
-{
-    typedef schar lane_type;
-    typedef v128_t vector_type;
-    enum { nlanes = 16 };
-
-    v_int8x16() {}
-    explicit v_int8x16(v128_t v) : val(v) {}
-    v_int8x16(schar v0, schar v1, schar v2, schar v3, schar v4, schar v5, schar v6, schar v7,
-            schar v8, schar v9, schar v10, schar v11, schar v12, schar v13, schar v14, schar v15)
-    {
-        schar v[] = {v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15};
-        val = wasm_v128_load(v);
-    }
-
-    schar get0() const
-    {
-        return wasm_i8x16_extract_lane(val, 0);
-    }
-
-    v128_t val;
-};
-
-struct v_uint16x8
-{
-    typedef ushort lane_type;
-    typedef v128_t vector_type;
-    enum { nlanes = 8 };
-
-    v_uint16x8() {}
-    explicit v_uint16x8(v128_t v) : val(v) {}
-    v_uint16x8(ushort v0, ushort v1, ushort v2, ushort v3, ushort v4, ushort v5, ushort v6, ushort v7)
-    {
-        ushort v[] = {v0, v1, v2, v3, v4, v5, v6, v7};
-        val = wasm_v128_load(v);
-    }
-
-    ushort get0() const
-    {
-        return (ushort)wasm_i16x8_extract_lane(val, 0);    // wasm_u16x8_extract_lane() unimplemented yet
-    }
-
-    v128_t val;
-};
-
-struct v_int16x8
-{
-    typedef short lane_type;
-    typedef v128_t vector_type;
-    enum { nlanes = 8 };
-
-    v_int16x8() {}
-    explicit v_int16x8(v128_t v) : val(v) {}
-    v_int16x8(short v0, short v1, short v2, short v3, short v4, short v5, short v6, short v7)
-    {
-        short v[] = {v0, v1, v2, v3, v4, v5, v6, v7};
-        val = wasm_v128_load(v);
-    }
-
-    short get0() const
-    {
-        return wasm_i16x8_extract_lane(val, 0);
-    }
-
-    v128_t val;
-};
-
-struct v_uint32x4
-{
-    typedef unsigned lane_type;
-    typedef v128_t vector_type;
-    enum { nlanes = 4 };
-
-    v_uint32x4() {}
-    explicit v_uint32x4(v128_t v) : val(v) {}
-    v_uint32x4(unsigned v0, unsigned v1, unsigned v2, unsigned v3)
-    {
-        unsigned v[] = {v0, v1, v2, v3};
-        val = wasm_v128_load(v);
-    }
-
-    unsigned get0() const
-    {
-        return (unsigned)wasm_i32x4_extract_lane(val, 0);
-    }
-
-    v128_t val;
-};
-
-struct v_int32x4
-{
-    typedef int lane_type;
-    typedef v128_t vector_type;
-    enum { nlanes = 4 };
-
-    v_int32x4() {}
-    explicit v_int32x4(v128_t v) : val(v) {}
-    v_int32x4(int v0, int v1, int v2, int v3)
-    {
-        int v[] = {v0, v1, v2, v3};
-        val = wasm_v128_load(v);
-    }
-
-    int get0() const
-    {
-        return wasm_i32x4_extract_lane(val, 0);
-    }
-
-    v128_t val;
-};
-
-struct v_float32x4
-{
-    typedef float lane_type;
-    typedef v128_t vector_type;
-    enum { nlanes = 4 };
-
-    v_float32x4() {}
-    explicit v_float32x4(v128_t v) : val(v) {}
-    v_float32x4(float v0, float v1, float v2, float v3)
-    {
-        float v[] = {v0, v1, v2, v3};
-        val = wasm_v128_load(v);
-    }
-
-    float get0() const
-    {
-        return wasm_f32x4_extract_lane(val, 0);
-    }
-
-    v128_t val;
-};
-
-struct v_uint64x2
-{
-    typedef uint64 lane_type;
-    typedef v128_t vector_type;
-    enum { nlanes = 2 };
-
-    v_uint64x2() {}
-    explicit v_uint64x2(v128_t v) : val(v) {}
-    v_uint64x2(uint64 v0, uint64 v1)
-    {
-        uint64 v[] = {v0, v1};
-        val = wasm_v128_load(v);
-    }
-
-    uint64 get0() const
-    {
-        return (uint64)wasm_i64x2_extract_lane(val, 0);
-    }
-
-    v128_t val;
-};
-
-struct v_int64x2
-{
-    typedef int64 lane_type;
-    typedef v128_t vector_type;
-    enum { nlanes = 2 };
-
-    v_int64x2() {}
-    explicit v_int64x2(v128_t v) : val(v) {}
-    v_int64x2(int64 v0, int64 v1)
-    {
-        int64 v[] = {v0, v1};
-        val = wasm_v128_load(v);
-    }
-
-    int64 get0() const
-    {
-        return wasm_i64x2_extract_lane(val, 0);
-    }
-
-    v128_t val;
-};
-
-struct v_float64x2
-{
-    typedef double lane_type;
-    typedef v128_t vector_type;
-    enum { nlanes = 2 };
-
-    v_float64x2() {}
-    explicit v_float64x2(v128_t v) : val(v) {}
-    v_float64x2(double v0, double v1)
-    {
-        double v[] = {v0, v1};
-        val = wasm_v128_load(v);
-    }
-
-    double get0() const
-    {
-        return wasm_f64x2_extract_lane(val, 0);
-    }
-
-    v128_t val;
-};
-
-namespace
-{
-#define OPENCV_HAL_IMPL_REINTERPRET_INT(ft, tt) \
-inline tt reinterpret_int(ft x) { union { ft l; tt i; } v; v.l = x; return v.i; }
-OPENCV_HAL_IMPL_REINTERPRET_INT(uchar, schar)
-OPENCV_HAL_IMPL_REINTERPRET_INT(schar, schar)
-OPENCV_HAL_IMPL_REINTERPRET_INT(ushort, short)
-OPENCV_HAL_IMPL_REINTERPRET_INT(short, short)
-OPENCV_HAL_IMPL_REINTERPRET_INT(unsigned, int)
-OPENCV_HAL_IMPL_REINTERPRET_INT(int, int)
-OPENCV_HAL_IMPL_REINTERPRET_INT(float, int)
-OPENCV_HAL_IMPL_REINTERPRET_INT(uint64, int64)
-OPENCV_HAL_IMPL_REINTERPRET_INT(int64, int64)
-OPENCV_HAL_IMPL_REINTERPRET_INT(double, int64)
-
-static const unsigned char popCountTable[] =
-{
-    0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7,
-    4, 5, 5, 6, 5, 6, 6, 7, 5, 6, 6, 7, 6, 7, 7, 8,
-};
-}  // namespace
-
-static v128_t wasm_unpacklo_i8x16(v128_t a, v128_t b) {
-    return wasm_v8x16_shuffle(a, b, 0,16,1,17,2,18,3,19,4,20,5,21,6,22,7,23);
-}
-
-static v128_t wasm_unpacklo_i16x8(v128_t a, v128_t b) {
-    return wasm_v8x16_shuffle(a, b, 0,1,16,17,2,3,18,19,4,5,20,21,6,7,22,23);
-}
-
-static v128_t wasm_unpacklo_i32x4(v128_t a, v128_t b) {
-    return wasm_v8x16_shuffle(a, b, 0,1,2,3,16,17,18,19,4,5,6,7,20,21,22,23);
-}
-
-static v128_t wasm_unpacklo_i64x2(v128_t a, v128_t b) {
-    return wasm_v8x16_shuffle(a, b, 0,1,2,3,4,5,6,7,16,17,18,19,20,21,22,23);
-}
-
-static v128_t wasm_unpackhi_i8x16(v128_t a, v128_t b) {
-    return wasm_v8x16_shuffle(a, b, 8,24,9,25,10,26,11,27,12,28,13,29,14,30,15,31);
-}
-
-static v128_t wasm_unpackhi_i16x8(v128_t a, v128_t b) {
-    return wasm_v8x16_shuffle(a, b, 8,9,24,25,10,11,26,27,12,13,28,29,14,15,30,31);
-}
-
-static v128_t wasm_unpackhi_i32x4(v128_t a, v128_t b) {
-    return wasm_v8x16_shuffle(a, b, 8,9,10,11,24,25,26,27,12,13,14,15,28,29,30,31);
-}
-
-static v128_t wasm_unpackhi_i64x2(v128_t a, v128_t b) {
-    return wasm_v8x16_shuffle(a, b, 8,9,10,11,12,13,14,15,24,25,26,27,28,29,30,31);
-}
-
-/** Convert **/
-// 8 >> 16
-inline v128_t v128_cvtu8x16_i16x8(const v128_t& a)
-{
-    const v128_t z = wasm_i8x16_splat(0);
-    return wasm_unpacklo_i8x16(a, z);
-}
-inline v128_t v128_cvti8x16_i16x8(const v128_t& a)
-{ return wasm_i16x8_shr(wasm_unpacklo_i8x16(a, a), 8); }
-// 8 >> 32
-inline v128_t v128_cvtu8x16_i32x4(const v128_t& a)
-{
-    const v128_t z = wasm_i8x16_splat(0);
-    return wasm_unpacklo_i16x8(wasm_unpacklo_i8x16(a, z), z);
-}
-inline v128_t v128_cvti8x16_i32x4(const v128_t& a)
-{
-    v128_t r = wasm_unpacklo_i8x16(a, a);
-    r = wasm_unpacklo_i8x16(r, r);
-    return wasm_i32x4_shr(r, 24);
-}
-// 16 >> 32
-inline v128_t v128_cvtu16x8_i32x4(const v128_t& a)
-{
-    const v128_t z = wasm_i8x16_splat(0);
-    return wasm_unpacklo_i16x8(a, z);
-}
-inline v128_t v128_cvti16x8_i32x4(const v128_t& a)
-{ return wasm_i32x4_shr(wasm_unpacklo_i16x8(a, a), 16); }
-// 32 >> 64
-inline v128_t v128_cvtu32x4_i64x2(const v128_t& a)
-{
-    const v128_t z = wasm_i8x16_splat(0);
-    return wasm_unpacklo_i32x4(a, z);
-}
-inline v128_t v128_cvti32x4_i64x2(const v128_t& a)
-{ return wasm_unpacklo_i32x4(a, wasm_i32x4_shr(a, 31)); }
-
-// 16 << 8
-inline v128_t v128_cvtu8x16_i16x8_high(const v128_t& a)
-{
-    const v128_t z = wasm_i8x16_splat(0);
-    return wasm_unpackhi_i8x16(a, z);
-}
-inline v128_t v128_cvti8x16_i16x8_high(const v128_t& a)
-{ return wasm_i16x8_shr(wasm_unpackhi_i8x16(a, a), 8); }
-// 32 << 16
-inline v128_t v128_cvtu16x8_i32x4_high(const v128_t& a)
-{
-    const v128_t z = wasm_i8x16_splat(0);
-    return wasm_unpackhi_i16x8(a, z);
-}
-inline v128_t v128_cvti16x8_i32x4_high(const v128_t& a)
-{ return wasm_i32x4_shr(wasm_unpackhi_i16x8(a, a), 16); }
-// 64 << 32
-inline v128_t v128_cvtu32x4_i64x2_high(const v128_t& a)
-{
-    const v128_t z = wasm_i8x16_splat(0);
-    return wasm_unpackhi_i32x4(a, z);
-}
-inline v128_t v128_cvti32x4_i64x2_high(const v128_t& a)
-{ return wasm_unpackhi_i32x4(a, wasm_i32x4_shr(a, 31)); }
-
-#define OPENCV_HAL_IMPL_WASM_INITVEC(_Tpvec, _Tp, suffix, zsuffix, _Tps) \
-inline _Tpvec v_setzero_##suffix() { return _Tpvec(wasm_##zsuffix##_splat((_Tps)0)); } \
-inline _Tpvec v_setall_##suffix(_Tp v) { return _Tpvec(wasm_##zsuffix##_splat((_Tps)v)); } \
-template<typename _Tpvec0> inline _Tpvec v_reinterpret_as_##suffix(const _Tpvec0& a) \
-{ return _Tpvec(a.val); }
-
-OPENCV_HAL_IMPL_WASM_INITVEC(v_uint8x16, uchar, u8, i8x16, schar)
-OPENCV_HAL_IMPL_WASM_INITVEC(v_int8x16, schar, s8, i8x16, schar)
-OPENCV_HAL_IMPL_WASM_INITVEC(v_uint16x8, ushort, u16, i16x8, short)
-OPENCV_HAL_IMPL_WASM_INITVEC(v_int16x8, short, s16, i16x8, short)
-OPENCV_HAL_IMPL_WASM_INITVEC(v_uint32x4, unsigned, u32, i32x4, int)
-OPENCV_HAL_IMPL_WASM_INITVEC(v_int32x4, int, s32, i32x4, int)
-OPENCV_HAL_IMPL_WASM_INITVEC(v_float32x4, float, f32, f32x4, float)
-OPENCV_HAL_IMPL_WASM_INITVEC(v_uint64x2, uint64, u64, i64x2, int64)
-OPENCV_HAL_IMPL_WASM_INITVEC(v_int64x2, int64, s64, i64x2, int64)
-OPENCV_HAL_IMPL_WASM_INITVEC(v_float64x2, double, f64, f64x2, double)
-
-//////////////// PACK ///////////////
-inline v_uint8x16 v_pack(const v_uint16x8& a, const v_uint16x8& b)
-{
-    v128_t maxval = wasm_i16x8_splat(255);
-    v128_t a1 = wasm_v128_bitselect(maxval, a.val, wasm_u16x8_gt(a.val, maxval));
-    v128_t b1 = wasm_v128_bitselect(maxval, b.val, wasm_u16x8_gt(b.val, maxval));
-    return v_uint8x16(wasm_v8x16_shuffle(a1, b1, 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30));
-}
-inline v_int8x16 v_pack(const v_int16x8& a, const v_int16x8& b)
-{
-    v128_t maxval = wasm_i16x8_splat(127);
-    v128_t minval = wasm_i16x8_splat(-128);
-    v128_t a1 = wasm_v128_bitselect(maxval, a.val, wasm_i16x8_gt(a.val, maxval));
-    v128_t b1 = wasm_v128_bitselect(maxval, b.val, wasm_i16x8_gt(b.val, maxval));
-    v128_t a2 = wasm_v128_bitselect(minval, a1, wasm_i16x8_lt(a1, minval));
-    v128_t b2 = wasm_v128_bitselect(minval, b1, wasm_i16x8_lt(b1, minval));
-    return v_int8x16(wasm_v8x16_shuffle(a2, b2, 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30));
-}
-inline v_uint16x8 v_pack(const v_uint32x4& a, const v_uint32x4& b)
-{
-    v128_t maxval = wasm_i32x4_splat(65535);
-    v128_t a1 = wasm_v128_bitselect(maxval, a.val, wasm_u32x4_gt(a.val, maxval));
-    v128_t b1 = wasm_v128_bitselect(maxval, b.val, wasm_u32x4_gt(b.val, maxval));
-    return v_uint16x8(wasm_v8x16_shuffle(a1, b1, 0,1,4,5,8,9,12,13,16,17,20,21,24,25,28,29));
-}
-inline v_int16x8 v_pack(const v_int32x4& a, const v_int32x4& b)
-{
-    v128_t maxval = wasm_i32x4_splat(32767);
-    v128_t minval = wasm_i32x4_splat(-32768);
-    v128_t a1 = wasm_v128_bitselect(maxval, a.val, wasm_i32x4_gt(a.val, maxval));
-    v128_t b1 = wasm_v128_bitselect(maxval, b.val, wasm_i32x4_gt(b.val, maxval));
-    v128_t a2 = wasm_v128_bitselect(minval, a1, wasm_i32x4_lt(a1, minval));
-    v128_t b2 = wasm_v128_bitselect(minval, b1, wasm_i32x4_lt(b1, minval));
-    return v_int16x8(wasm_v8x16_shuffle(a2, b2, 0,1,4,5,8,9,12,13,16,17,20,21,24,25,28,29));
-}
-inline v_uint32x4 v_pack(const v_uint64x2& a, const v_uint64x2& b)
-{
-    return v_uint32x4(wasm_v8x16_shuffle(a.val, b.val, 0,1,2,3,8,9,10,11,16,17,18,19,24,25,26,27));
-}
-inline v_int32x4 v_pack(const v_int64x2& a, const v_int64x2& b)
-{
-    return v_int32x4(wasm_v8x16_shuffle(a.val, b.val, 0,1,2,3,8,9,10,11,16,17,18,19,24,25,26,27));
-}
-inline v_uint8x16 v_pack_u(const v_int16x8& a, const v_int16x8& b)
-{
-    v128_t maxval = wasm_i16x8_splat(255);
-    v128_t minval = wasm_i16x8_splat(0);
-    v128_t a1 = wasm_v128_bitselect(maxval, a.val, wasm_i16x8_gt(a.val, maxval));
-    v128_t b1 = wasm_v128_bitselect(maxval, b.val, wasm_i16x8_gt(b.val, maxval));
-    v128_t a2 = wasm_v128_bitselect(minval, a1, wasm_i16x8_lt(a1, minval));
-    v128_t b2 = wasm_v128_bitselect(minval, b1, wasm_i16x8_lt(b1, minval));
-    return v_uint8x16(wasm_v8x16_shuffle(a2, b2, 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30));
-}
-inline v_uint16x8 v_pack_u(const v_int32x4& a, const v_int32x4& b)
-{
-    v128_t maxval = wasm_i32x4_splat(65535);
-    v128_t minval = wasm_i32x4_splat(0);
-    v128_t a1 = wasm_v128_bitselect(maxval, a.val, wasm_i32x4_gt(a.val, maxval));
-    v128_t b1 = wasm_v128_bitselect(maxval, b.val, wasm_i32x4_gt(b.val, maxval));
-    v128_t a2 = wasm_v128_bitselect(minval, a1, wasm_i32x4_lt(a1, minval));
-    v128_t b2 = wasm_v128_bitselect(minval, b1, wasm_i32x4_lt(b1, minval));
-    return v_uint16x8(wasm_v8x16_shuffle(a2, b2, 0,1,4,5,8,9,12,13,16,17,20,21,24,25,28,29));
-}
-
-template<int n>
-inline v_uint8x16 v_rshr_pack(const v_uint16x8& a, const v_uint16x8& b)
-{
-    v128_t delta = wasm_i16x8_splat(((short)1 << (n-1)));
-    v128_t a1 = wasm_u16x8_shr(wasm_i16x8_add(a.val, delta), n);
-    v128_t b1 = wasm_u16x8_shr(wasm_i16x8_add(b.val, delta), n);
-    v128_t maxval = wasm_i16x8_splat(255);
-    v128_t a2 = wasm_v128_bitselect(maxval, a1, wasm_u16x8_gt(a1, maxval));
-    v128_t b2 = wasm_v128_bitselect(maxval, b1, wasm_u16x8_gt(b1, maxval));
-    return v_uint8x16(wasm_v8x16_shuffle(a2, b2, 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30));
-}
-template<int n>
-inline v_int8x16 v_rshr_pack(const v_int16x8& a, const v_int16x8& b)
-{
-    v128_t delta = wasm_i16x8_splat(((short)1 << (n-1)));
-    v128_t a1 = wasm_i16x8_shr(wasm_i16x8_add(a.val, delta), n);
-    v128_t b1 = wasm_i16x8_shr(wasm_i16x8_add(b.val, delta), n);
-    v128_t maxval = wasm_i16x8_splat(127);
-    v128_t minval = wasm_i16x8_splat(-128);
-    v128_t a2 = wasm_v128_bitselect(maxval, a1, wasm_i16x8_gt(a1, maxval));
-    v128_t b2 = wasm_v128_bitselect(maxval, b1, wasm_i16x8_gt(b1, maxval));
-    v128_t a3 = wasm_v128_bitselect(minval, a2, wasm_i16x8_lt(a1, minval));
-    v128_t b3 = wasm_v128_bitselect(minval, b2, wasm_i16x8_lt(b1, minval));
-    return v_int8x16(wasm_v8x16_shuffle(a3, b3, 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30));
-}
-template<int n>
-inline v_uint16x8 v_rshr_pack(const v_uint32x4& a, const v_uint32x4& b)
-{
-    v128_t delta = wasm_i32x4_splat(((int)1 << (n-1)));
-    v128_t a1 = wasm_u32x4_shr(wasm_i32x4_add(a.val, delta), n);
-    v128_t b1 = wasm_u32x4_shr(wasm_i32x4_add(b.val, delta), n);
-    v128_t maxval = wasm_i32x4_splat(65535);
-    v128_t a2 = wasm_v128_bitselect(maxval, a1, wasm_u32x4_gt(a1, maxval));
-    v128_t b2 = wasm_v128_bitselect(maxval, b1, wasm_u32x4_gt(b1, maxval));
-    return v_uint16x8(wasm_v8x16_shuffle(a2, b2, 0,1,4,5,8,9,12,13,16,17,20,21,24,25,28,29));
-}
-template<int n>
-inline v_int16x8 v_rshr_pack(const v_int32x4& a, const v_int32x4& b)
-{
-    v128_t delta = wasm_i32x4_splat(((int)1 << (n-1)));
-    v128_t a1 = wasm_i32x4_shr(wasm_i32x4_add(a.val, delta), n);
-    v128_t b1 = wasm_i32x4_shr(wasm_i32x4_add(b.val, delta), n);
-    v128_t maxval = wasm_i32x4_splat(32767);
-    v128_t minval = wasm_i16x8_splat(-32768);
-    v128_t a2 = wasm_v128_bitselect(maxval, a1, wasm_i32x4_gt(a1, maxval));
-    v128_t b2 = wasm_v128_bitselect(maxval, b1, wasm_i32x4_gt(b1, maxval));
-    v128_t a3 = wasm_v128_bitselect(minval, a2, wasm_i32x4_lt(a1, minval));
-    v128_t b3 = wasm_v128_bitselect(minval, b2, wasm_i32x4_lt(b1, minval));
-    return v_int16x8(wasm_v8x16_shuffle(a3, b3, 0,1,4,5,8,9,12,13,16,17,20,21,24,25,28,29));
-}
-template<int n>
-inline v_uint32x4 v_rshr_pack(const v_uint64x2& a, const v_uint64x2& b)
-{
-    v128_t delta = wasm_i64x2_splat(((int64)1 << (n-1)));
-    v128_t a1 = wasm_u64x2_shr(wasm_i64x2_add(a.val, delta), n);
-    v128_t b1 = wasm_u64x2_shr(wasm_i64x2_add(b.val, delta), n);
-    return v_uint32x4(wasm_v8x16_shuffle(a1, b1, 0,1,2,3,8,9,10,11,16,17,18,19,24,25,26,27));
-}
-template<int n>
-inline v_int32x4 v_rshr_pack(const v_int64x2& a, const v_int64x2& b)
-{
-    v128_t delta = wasm_i64x2_splat(((int64)1 << (n-1)));
-    v128_t a1 = wasm_i64x2_shr(wasm_i64x2_add(a.val, delta), n);
-    v128_t b1 = wasm_i64x2_shr(wasm_i64x2_add(b.val, delta), n);
-    return v_int32x4(wasm_v8x16_shuffle(a1, b1, 0,1,2,3,8,9,10,11,16,17,18,19,24,25,26,27));
-}
-template<int n>
-inline v_uint8x16 v_rshr_pack_u(const v_int16x8& a, const v_int16x8& b)
-{
-    v128_t delta = wasm_i16x8_splat(((short)1 << (n-1)));
-    v128_t a1 = wasm_i16x8_shr(wasm_i16x8_add(a.val, delta), n);
-    v128_t b1 = wasm_i16x8_shr(wasm_i16x8_add(b.val, delta), n);
-    v128_t maxval = wasm_i16x8_splat(255);
-    v128_t minval = wasm_i16x8_splat(0);
-    v128_t a2 = wasm_v128_bitselect(maxval, a1, wasm_i16x8_gt(a1, maxval));
-    v128_t b2 = wasm_v128_bitselect(maxval, b1, wasm_i16x8_gt(b1, maxval));
-    v128_t a3 = wasm_v128_bitselect(minval, a2, wasm_i16x8_lt(a1, minval));
-    v128_t b3 = wasm_v128_bitselect(minval, b2, wasm_i16x8_lt(b1, minval));
-    return v_uint8x16(wasm_v8x16_shuffle(a3, b3, 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30));
-}
-template<int n>
-inline v_uint16x8 v_rshr_pack_u(const v_int32x4& a, const v_int32x4& b)
-{
-    v128_t delta = wasm_i32x4_splat(((int)1 << (n-1)));
-    v128_t a1 = wasm_i32x4_shr(wasm_i32x4_add(a.val, delta), n);
-    v128_t b1 = wasm_i32x4_shr(wasm_i32x4_add(b.val, delta), n);
-    v128_t maxval = wasm_i32x4_splat(65535);
-    v128_t minval = wasm_i16x8_splat(0);
-    v128_t a2 = wasm_v128_bitselect(maxval, a1, wasm_i32x4_gt(a1, maxval));
-    v128_t b2 = wasm_v128_bitselect(maxval, b1, wasm_i32x4_gt(b1, maxval));
-    v128_t a3 = wasm_v128_bitselect(minval, a2, wasm_i32x4_lt(a1, minval));
-    v128_t b3 = wasm_v128_bitselect(minval, b2, wasm_i32x4_lt(b1, minval));
-    return v_uint16x8(wasm_v8x16_shuffle(a3, b3, 0,1,4,5,8,9,12,13,16,17,20,21,24,25,28,29));
-}
-
-inline void v_pack_store(uchar* ptr, const v_uint16x8& a)
-{
-    v128_t maxval = wasm_i16x8_splat(255);
-    v128_t a1 = wasm_v128_bitselect(maxval, a.val, wasm_u16x8_gt(a.val, maxval));
-    v128_t r = wasm_v8x16_shuffle(a1, a1, 0,2,4,6,8,10,12,14,0,2,4,6,8,10,12,14);
-    uchar t_ptr[16];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<8; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-inline void v_pack_store(schar* ptr, const v_int16x8& a)
-{
-    v128_t maxval = wasm_i16x8_splat(127);
-    v128_t minval = wasm_i16x8_splat(-128);
-    v128_t a1 = wasm_v128_bitselect(maxval, a.val, wasm_i16x8_gt(a.val, maxval));
-    v128_t a2 = wasm_v128_bitselect(minval, a1, wasm_i16x8_lt(a1, minval));
-    v128_t r = wasm_v8x16_shuffle(a2, a2, 0,2,4,6,8,10,12,14,0,2,4,6,8,10,12,14);
-    schar t_ptr[16];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<8; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-inline void v_pack_store(ushort* ptr, const v_uint32x4& a)
-{
-    v128_t maxval = wasm_i32x4_splat(65535);
-    v128_t a1 = wasm_v128_bitselect(maxval, a.val, wasm_u32x4_gt(a.val, maxval));
-    v128_t r = wasm_v8x16_shuffle(a1, a1, 0,1,4,5,8,9,12,13,0,1,4,5,8,9,12,13);
-    ushort t_ptr[8];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<4; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-inline void v_pack_store(short* ptr, const v_int32x4& a)
-{
-    v128_t maxval = wasm_i32x4_splat(32767);
-    v128_t minval = wasm_i32x4_splat(-32768);
-    v128_t a1 = wasm_v128_bitselect(maxval, a.val, wasm_i32x4_gt(a.val, maxval));
-    v128_t a2 = wasm_v128_bitselect(minval, a1, wasm_i32x4_lt(a1, minval));
-    v128_t r = wasm_v8x16_shuffle(a2, a2, 0,1,4,5,8,9,12,13,0,1,4,5,8,9,12,13);
-    short t_ptr[8];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<4; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-inline void v_pack_store(unsigned* ptr, const v_uint64x2& a)
-{
-    v128_t r = wasm_v8x16_shuffle(a.val, a.val, 0,1,2,3,8,9,10,11,0,1,2,3,8,9,10,11);
-    unsigned t_ptr[4];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<2; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-inline void v_pack_store(int* ptr, const v_int64x2& a)
-{
-    v128_t r = wasm_v8x16_shuffle(a.val, a.val, 0,1,2,3,8,9,10,11,0,1,2,3,8,9,10,11);
-    int t_ptr[4];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<2; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-inline void v_pack_u_store(uchar* ptr, const v_int16x8& a)
-{
-    v128_t maxval = wasm_i16x8_splat(255);
-    v128_t minval = wasm_i16x8_splat(0);
-    v128_t a1 = wasm_v128_bitselect(maxval, a.val, wasm_i16x8_gt(a.val, maxval));
-    v128_t a2 = wasm_v128_bitselect(minval, a1, wasm_i16x8_lt(a1, minval));
-    v128_t r = wasm_v8x16_shuffle(a2, a2, 0,2,4,6,8,10,12,14,0,2,4,6,8,10,12,14);
-    uchar t_ptr[16];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<8; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-inline void v_pack_u_store(ushort* ptr, const v_int32x4& a)
-{
-    v128_t maxval = wasm_i32x4_splat(65535);
-    v128_t minval = wasm_i32x4_splat(0);
-    v128_t a1 = wasm_v128_bitselect(maxval, a.val, wasm_i32x4_gt(a.val, maxval));
-    v128_t a2 = wasm_v128_bitselect(minval, a1, wasm_i32x4_lt(a1, minval));
-    v128_t r = wasm_v8x16_shuffle(a2, a2, 0,1,4,5,8,9,12,13,0,1,4,5,8,9,12,13);
-    ushort t_ptr[8];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<4; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-
-template<int n>
-inline void v_rshr_pack_store(uchar* ptr, const v_uint16x8& a)
-{
-    v128_t delta = wasm_i16x8_splat((short)(1 << (n-1)));
-    v128_t a1 = wasm_u16x8_shr(wasm_i16x8_add(a.val, delta), n);
-    v128_t maxval = wasm_i16x8_splat(255);
-    v128_t a2 = wasm_v128_bitselect(maxval, a1, wasm_u16x8_gt(a1, maxval));
-    v128_t r = wasm_v8x16_shuffle(a2, a2, 0,2,4,6,8,10,12,14,0,2,4,6,8,10,12,14);
-    uchar t_ptr[16];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<8; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-template<int n>
-inline void v_rshr_pack_store(schar* ptr, const v_int16x8& a)
-{
-    v128_t delta = wasm_i16x8_splat(((short)1 << (n-1)));
-    v128_t a1 = wasm_i16x8_shr(wasm_i16x8_add(a.val, delta), n);
-    v128_t maxval = wasm_i16x8_splat(127);
-    v128_t minval = wasm_i16x8_splat(-128);
-    v128_t a2 = wasm_v128_bitselect(maxval, a1, wasm_i16x8_gt(a1, maxval));
-    v128_t a3 = wasm_v128_bitselect(minval, a2, wasm_i16x8_lt(a1, minval));
-    v128_t r = wasm_v8x16_shuffle(a3, a3, 0,2,4,6,8,10,12,14,0,2,4,6,8,10,12,14);
-    schar t_ptr[16];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<8; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-template<int n>
-inline void v_rshr_pack_store(ushort* ptr, const v_uint32x4& a)
-{
-    v128_t delta = wasm_i32x4_splat(((int)1 << (n-1)));
-    v128_t a1 = wasm_u32x4_shr(wasm_i32x4_add(a.val, delta), n);
-    v128_t maxval = wasm_i32x4_splat(65535);
-    v128_t a2 = wasm_v128_bitselect(maxval, a1, wasm_u32x4_gt(a1, maxval));
-    v128_t r = wasm_v8x16_shuffle(a2, a2, 0,1,4,5,8,9,12,13,0,1,4,5,8,9,12,13);
-    ushort t_ptr[8];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<4; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-template<int n>
-inline void v_rshr_pack_store(short* ptr, const v_int32x4& a)
-{
-    v128_t delta = wasm_i32x4_splat(((int)1 << (n-1)));
-    v128_t a1 = wasm_i32x4_shr(wasm_i32x4_add(a.val, delta), n);
-    v128_t maxval = wasm_i32x4_splat(32767);
-    v128_t minval = wasm_i32x4_splat(-32768);
-    v128_t a2 = wasm_v128_bitselect(maxval, a1, wasm_i32x4_gt(a1, maxval));
-    v128_t a3 = wasm_v128_bitselect(minval, a2, wasm_i32x4_lt(a1, minval));
-    v128_t r = wasm_v8x16_shuffle(a3, a3, 0,1,4,5,8,9,12,13,0,1,4,5,8,9,12,13);
-    short t_ptr[8];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<4; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-template<int n>
-inline void v_rshr_pack_store(unsigned* ptr, const v_uint64x2& a)
-{
-    v128_t delta = wasm_i64x2_splat(((int64)1 << (n-1)));
-    v128_t a1 = wasm_u64x2_shr(wasm_i64x2_add(a.val, delta), n);
-    v128_t r = wasm_v8x16_shuffle(a1, a1, 0,1,2,3,8,9,10,11,0,1,2,3,8,9,10,11);
-    unsigned t_ptr[4];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<2; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-template<int n>
-inline void v_rshr_pack_store(int* ptr, const v_int64x2& a)
-{
-    v128_t delta = wasm_i64x2_splat(((int64)1 << (n-1)));
-    v128_t a1 = wasm_i64x2_shr(wasm_i64x2_add(a.val, delta), n);
-    v128_t r = wasm_v8x16_shuffle(a1, a1, 0,1,2,3,8,9,10,11,0,1,2,3,8,9,10,11);
-    int t_ptr[4];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<2; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-template<int n>
-inline void v_rshr_pack_u_store(uchar* ptr, const v_int16x8& a)
-{
-    v128_t delta = wasm_i16x8_splat(((short)1 << (n-1)));
-    v128_t a1 = wasm_i16x8_shr(wasm_i16x8_add(a.val, delta), n);
-    v128_t maxval = wasm_i16x8_splat(255);
-    v128_t minval = wasm_i16x8_splat(0);
-    v128_t a2 = wasm_v128_bitselect(maxval, a1, wasm_i16x8_gt(a1, maxval));
-    v128_t a3 = wasm_v128_bitselect(minval, a2, wasm_i16x8_lt(a1, minval));
-    v128_t r = wasm_v8x16_shuffle(a3, a3, 0,2,4,6,8,10,12,14,0,2,4,6,8,10,12,14);
-    uchar t_ptr[16];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<8; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-template<int n>
-inline void v_rshr_pack_u_store(ushort* ptr, const v_int32x4& a)
-{
-    v128_t delta = wasm_i32x4_splat(((int)1 << (n-1)));
-    v128_t a1 = wasm_i32x4_shr(wasm_i32x4_add(a.val, delta), n);
-    v128_t maxval = wasm_i32x4_splat(65535);
-    v128_t minval = wasm_i32x4_splat(0);
-    v128_t a2 = wasm_v128_bitselect(maxval, a1, wasm_i32x4_gt(a1, maxval));
-    v128_t a3 = wasm_v128_bitselect(minval, a2, wasm_i32x4_lt(a1, minval));
-    v128_t r = wasm_v8x16_shuffle(a3, a3, 0,1,4,5,8,9,12,13,0,1,4,5,8,9,12,13);
-    ushort t_ptr[8];
-    wasm_v128_store(t_ptr, r);
-    for (int i=0; i<4; ++i) {
-        ptr[i] = t_ptr[i];
-    }
-}
-
-inline v_uint8x16 v_pack_b(const v_uint16x8& a, const v_uint16x8& b)
-{
-    v128_t maxval = wasm_i16x8_splat(255);
-    v128_t a1 = wasm_v128_bitselect(maxval, a.val, wasm_u16x8_gt(a.val, maxval));
-    v128_t b1 = wasm_v128_bitselect(maxval, b.val, wasm_u16x8_gt(b.val, maxval));
-    return v_uint8x16(wasm_v8x16_shuffle(a1, b1, 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30));
-}
-
-inline v_uint8x16 v_pack_b(const v_uint32x4& a, const v_uint32x4& b,
-                           const v_uint32x4& c, const v_uint32x4& d)
-{
-    v128_t maxval = wasm_i32x4_splat(255);
-    v128_t a1 = wasm_v128_bitselect(maxval, a.val, wasm_u32x4_gt(a.val, maxval));
-    v128_t b1 = wasm_v128_bitselect(maxval, b.val, wasm_u32x4_gt(b.val, maxval));
-    v128_t c1 = wasm_v128_bitselect(maxval, c.val, wasm_u32x4_gt(c.val, maxval));
-    v128_t d1 = wasm_v128_bitselect(maxval, d.val, wasm_u32x4_gt(d.val, maxval));
-    v128_t ab = wasm_v8x16_shuffle(a1, b1, 0,4,8,12,16,20,24,28,0,4,8,12,16,20,24,28);
-    v128_t cd = wasm_v8x16_shuffle(c1, d1, 0,4,8,12,16,20,24,28,0,4,8,12,16,20,24,28);
-    return v_uint8x16(wasm_v8x16_shuffle(ab, cd, 0,1,2,3,4,5,6,7,16,17,18,19,20,21,22,23));
-}
-
-inline v_uint8x16 v_pack_b(const v_uint64x2& a, const v_uint64x2& b, const v_uint64x2& c,
-                           const v_uint64x2& d, const v_uint64x2& e, const v_uint64x2& f,
-                           const v_uint64x2& g, const v_uint64x2& h)
-{
-    v128_t maxval = wasm_i32x4_splat(255);
-    v128_t a1 = wasm_v128_bitselect(maxval, a.val, ((__u64x2)(a.val) > (__u64x2)maxval));
-    v128_t b1 = wasm_v128_bitselect(maxval, b.val, ((__u64x2)(b.val) > (__u64x2)maxval));
-    v128_t c1 = wasm_v128_bitselect(maxval, c.val, ((__u64x2)(c.val) > (__u64x2)maxval));
-    v128_t d1 = wasm_v128_bitselect(maxval, d.val, ((__u64x2)(d.val) > (__u64x2)maxval));
-    v128_t e1 = wasm_v128_bitselect(maxval, e.val, ((__u64x2)(e.val) > (__u64x2)maxval));
-    v128_t f1 = wasm_v128_bitselect(maxval, f.val, ((__u64x2)(f.val) > (__u64x2)maxval));
-    v128_t g1 = wasm_v128_bitselect(maxval, g.val, ((__u64x2)(g.val) > (__u64x2)maxval));
-    v128_t h1 = wasm_v128_bitselect(maxval, h.val, ((__u64x2)(h.val) > (__u64x2)maxval));
-    v128_t ab = wasm_v8x16_shuffle(a1, b1, 0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24);
-    v128_t cd = wasm_v8x16_shuffle(c1, d1, 0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24);
-    v128_t ef = wasm_v8x16_shuffle(e1, f1, 0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24);
-    v128_t gh = wasm_v8x16_shuffle(g1, h1, 0,8,16,24,0,8,16,24,0,8,16,24,0,8,16,24);
-    v128_t abcd = wasm_v8x16_shuffle(ab, cd, 0,1,2,3,16,17,18,19,0,1,2,3,16,17,18,19);
-    v128_t efgh = wasm_v8x16_shuffle(ef, gh, 0,1,2,3,16,17,18,19,0,1,2,3,16,17,18,19);
-    return v_uint8x16(wasm_v8x16_shuffle(abcd, efgh, 0,1,2,3,4,5,6,7,16,17,18,19,20,21,22,23));
-}
-
-inline v_float32x4 v_matmul(const v_float32x4& v, const v_float32x4& m0,
-                            const v_float32x4& m1, const v_float32x4& m2,
-                            const v_float32x4& m3)
-{
-    v128_t v0 = wasm_f32x4_splat(wasm_f32x4_extract_lane(v.val, 0));
-    v128_t v1 = wasm_f32x4_splat(wasm_f32x4_extract_lane(v.val, 1));
-    v128_t v2 = wasm_f32x4_splat(wasm_f32x4_extract_lane(v.val, 2));
-    v128_t v3 = wasm_f32x4_splat(wasm_f32x4_extract_lane(v.val, 3));
-    v0 = wasm_f32x4_mul(v0, m0.val);
-    v1 = wasm_f32x4_mul(v1, m1.val);
-    v2 = wasm_f32x4_mul(v2, m2.val);
-    v3 = wasm_f32x4_mul(v3, m3.val);
-
-    return v_float32x4(wasm_f32x4_add(wasm_f32x4_add(v0, v1), wasm_f32x4_add(v2, v3)));
-}
-
-inline v_float32x4 v_matmuladd(const v_float32x4& v, const v_float32x4& m0,
-                               const v_float32x4& m1, const v_float32x4& m2,
-                               const v_float32x4& a)
-{
-    v128_t v0 = wasm_f32x4_splat(wasm_f32x4_extract_lane(v.val, 0));
-    v128_t v1 = wasm_f32x4_splat(wasm_f32x4_extract_lane(v.val, 1));
-    v128_t v2 = wasm_f32x4_splat(wasm_f32x4_extract_lane(v.val, 2));
-    v0 = wasm_f32x4_mul(v0, m0.val);
-    v1 = wasm_f32x4_mul(v1, m1.val);
-    v2 = wasm_f32x4_mul(v2, m2.val);
-
-    return v_float32x4(wasm_f32x4_add(wasm_f32x4_add(v0, v1), wasm_f32x4_add(v2, a.val)));
-}
-
-#define OPENCV_HAL_IMPL_WASM_BIN_OP(bin_op, _Tpvec, intrin) \
-inline _Tpvec operator bin_op (const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(intrin(a.val, b.val)); \
-} \
-inline _Tpvec& operator bin_op##= (_Tpvec& a, const _Tpvec& b) \
-{ \
-    a.val = intrin(a.val, b.val); \
-    return a; \
-}
-
-OPENCV_HAL_IMPL_WASM_BIN_OP(+, v_uint8x16, wasm_u8x16_add_saturate)
-OPENCV_HAL_IMPL_WASM_BIN_OP(-, v_uint8x16, wasm_u8x16_sub_saturate)
-OPENCV_HAL_IMPL_WASM_BIN_OP(+, v_int8x16, wasm_i8x16_add_saturate)
-OPENCV_HAL_IMPL_WASM_BIN_OP(-, v_int8x16, wasm_i8x16_sub_saturate)
-OPENCV_HAL_IMPL_WASM_BIN_OP(+, v_uint16x8, wasm_u16x8_add_saturate)
-OPENCV_HAL_IMPL_WASM_BIN_OP(-, v_uint16x8, wasm_u16x8_sub_saturate)
-OPENCV_HAL_IMPL_WASM_BIN_OP(+, v_int16x8, wasm_i16x8_add_saturate)
-OPENCV_HAL_IMPL_WASM_BIN_OP(-, v_int16x8, wasm_i16x8_sub_saturate)
-OPENCV_HAL_IMPL_WASM_BIN_OP(+, v_uint32x4, wasm_i32x4_add)
-OPENCV_HAL_IMPL_WASM_BIN_OP(-, v_uint32x4, wasm_i32x4_sub)
-OPENCV_HAL_IMPL_WASM_BIN_OP(*, v_uint32x4, wasm_i32x4_mul)
-OPENCV_HAL_IMPL_WASM_BIN_OP(+, v_int32x4, wasm_i32x4_add)
-OPENCV_HAL_IMPL_WASM_BIN_OP(-, v_int32x4, wasm_i32x4_sub)
-OPENCV_HAL_IMPL_WASM_BIN_OP(*, v_int32x4, wasm_i32x4_mul)
-OPENCV_HAL_IMPL_WASM_BIN_OP(+, v_float32x4, wasm_f32x4_add)
-OPENCV_HAL_IMPL_WASM_BIN_OP(-, v_float32x4, wasm_f32x4_sub)
-OPENCV_HAL_IMPL_WASM_BIN_OP(*, v_float32x4, wasm_f32x4_mul)
-OPENCV_HAL_IMPL_WASM_BIN_OP(/, v_float32x4, wasm_f32x4_div)
-OPENCV_HAL_IMPL_WASM_BIN_OP(+, v_uint64x2, wasm_i64x2_add)
-OPENCV_HAL_IMPL_WASM_BIN_OP(-, v_uint64x2, wasm_i64x2_sub)
-OPENCV_HAL_IMPL_WASM_BIN_OP(+, v_int64x2, wasm_i64x2_add)
-OPENCV_HAL_IMPL_WASM_BIN_OP(-, v_int64x2, wasm_i64x2_sub)
-OPENCV_HAL_IMPL_WASM_BIN_OP(+, v_float64x2, wasm_f64x2_add)
-OPENCV_HAL_IMPL_WASM_BIN_OP(-, v_float64x2, wasm_f64x2_sub)
-OPENCV_HAL_IMPL_WASM_BIN_OP(*, v_float64x2, wasm_f64x2_mul)
-OPENCV_HAL_IMPL_WASM_BIN_OP(/, v_float64x2, wasm_f64x2_div)
-
-// saturating multiply 8-bit, 16-bit
-#define OPENCV_HAL_IMPL_WASM_MUL_SAT(_Tpvec, _Tpwvec)        \
-inline _Tpvec operator * (const _Tpvec& a, const _Tpvec& b)  \
-{                                                            \
-    _Tpwvec c, d;                                            \
-    v_mul_expand(a, b, c, d);                                \
-    return v_pack(c, d);                                     \
-}                                                            \
-inline _Tpvec& operator *= (_Tpvec& a, const _Tpvec& b)      \
-{ a = a * b; return a; }
-
-OPENCV_HAL_IMPL_WASM_MUL_SAT(v_uint8x16, v_uint16x8)
-OPENCV_HAL_IMPL_WASM_MUL_SAT(v_int8x16,  v_int16x8)
-OPENCV_HAL_IMPL_WASM_MUL_SAT(v_uint16x8, v_uint32x4)
-OPENCV_HAL_IMPL_WASM_MUL_SAT(v_int16x8,  v_int32x4)
-
-//  Multiply and expand
-inline void v_mul_expand(const v_uint8x16& a, const v_uint8x16& b,
-                         v_uint16x8& c, v_uint16x8& d)
-{
-    v_uint16x8 a0, a1, b0, b1;
-    v_expand(a, a0, a1);
-    v_expand(b, b0, b1);
-    c = v_mul_wrap(a0, b0);
-    d = v_mul_wrap(a1, b1);
-}
-
-inline void v_mul_expand(const v_int8x16& a, const v_int8x16& b,
-                         v_int16x8& c, v_int16x8& d)
-{
-    v_int16x8 a0, a1, b0, b1;
-    v_expand(a, a0, a1);
-    v_expand(b, b0, b1);
-    c = v_mul_wrap(a0, b0);
-    d = v_mul_wrap(a1, b1);
-}
-
-inline void v_mul_expand(const v_int16x8& a, const v_int16x8& b,
-                         v_int32x4& c, v_int32x4& d)
-{
-    v_int32x4 a0, a1, b0, b1;
-    v_expand(a, a0, a1);
-    v_expand(b, b0, b1);
-    c.val = wasm_i32x4_mul(a0.val, b0.val);
-    d.val = wasm_i32x4_mul(a1.val, b1.val);
-}
-
-inline void v_mul_expand(const v_uint16x8& a, const v_uint16x8& b,
-                         v_uint32x4& c, v_uint32x4& d)
-{
-    v_uint32x4 a0, a1, b0, b1;
-    v_expand(a, a0, a1);
-    v_expand(b, b0, b1);
-    c.val = wasm_i32x4_mul(a0.val, b0.val);
-    d.val = wasm_i32x4_mul(a1.val, b1.val);
-}
-
-inline void v_mul_expand(const v_uint32x4& a, const v_uint32x4& b,
-                         v_uint64x2& c, v_uint64x2& d)
-{
-    v_uint64x2 a0, a1, b0, b1;
-    v_expand(a, a0, a1);
-    v_expand(b, b0, b1);
-    c.val = ((__u64x2)(a0.val) * (__u64x2)(b0.val));
-    d.val = ((__u64x2)(a1.val) * (__u64x2)(b1.val));
-}
-
-inline v_int16x8 v_mul_hi(const v_int16x8& a, const v_int16x8& b)
-{
-    v_int32x4 a0, a1, b0, b1;
-    v_expand(a, a0, a1);
-    v_expand(b, b0, b1);
-    v128_t c = wasm_i32x4_mul(a0.val, b0.val);
-    v128_t d = wasm_i32x4_mul(a1.val, b1.val);
-    return v_int16x8(wasm_v8x16_shuffle(c, d, 2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31));
-}
-inline v_uint16x8 v_mul_hi(const v_uint16x8& a, const v_uint16x8& b)
-{
-    v_uint32x4 a0, a1, b0, b1;
-    v_expand(a, a0, a1);
-    v_expand(b, b0, b1);
-    v128_t c = wasm_i32x4_mul(a0.val, b0.val);
-    v128_t d = wasm_i32x4_mul(a1.val, b1.val);
-    return v_uint16x8(wasm_v8x16_shuffle(c, d, 2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31));
-}
-
-//////// Dot Product ////////
-
-inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b)
-{
-    v128_t a0 = wasm_i32x4_shr(wasm_i32x4_shl(a.val, 16), 16);
-    v128_t a1 = wasm_i32x4_shr(a.val, 16);
-    v128_t b0 = wasm_i32x4_shr(wasm_i32x4_shl(b.val, 16), 16);
-    v128_t b1 = wasm_i32x4_shr(b.val, 16);
-    v128_t c = wasm_i32x4_mul(a0, b0);
-    v128_t d = wasm_i32x4_mul(a1, b1);
-    return v_int32x4(wasm_i32x4_add(c, d));
-}
-
-inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b, const v_int32x4& c)
-{ return v_dotprod(a, b) + c; }
-
-inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b)
-{
-    v128_t a0 = wasm_i64x2_shr(wasm_i64x2_shl(a.val, 32), 32);
-    v128_t a1 = wasm_i64x2_shr(a.val, 32);
-    v128_t b0 = wasm_i64x2_shr(wasm_i64x2_shl(b.val, 32), 32);
-    v128_t b1 = wasm_i64x2_shr(b.val, 32);
-    v128_t c = (v128_t)((__i64x2)a0 * (__i64x2)b0);
-    v128_t d = (v128_t)((__i64x2)a1 * (__i64x2)b1);
-    return v_int64x2(wasm_i64x2_add(c, d));
-}
-inline v_int64x2 v_dotprod(const v_int32x4& a, const v_int32x4& b, const v_int64x2& c)
-{
-    return v_dotprod(a, b) + c;
-}
-
-// 8 >> 32
-inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b)
-{
-    v128_t a0 = wasm_u16x8_shr(wasm_i16x8_shl(a.val, 8), 8);
-    v128_t a1 = wasm_u16x8_shr(a.val, 8);
-    v128_t b0 = wasm_u16x8_shr(wasm_i16x8_shl(b.val, 8), 8);
-    v128_t b1 = wasm_u16x8_shr(b.val, 8);
-    return v_uint32x4((
-        v_dotprod(v_int16x8(a0), v_int16x8(b0)) +
-        v_dotprod(v_int16x8(a1), v_int16x8(b1))).val
-    );
-}
-inline v_uint32x4 v_dotprod_expand(const v_uint8x16& a, const v_uint8x16& b, const v_uint32x4& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b)
-{
-    v128_t a0 = wasm_i16x8_shr(wasm_i16x8_shl(a.val, 8), 8);
-    v128_t a1 = wasm_i16x8_shr(a.val, 8);
-    v128_t b0 = wasm_i16x8_shr(wasm_i16x8_shl(b.val, 8), 8);
-    v128_t b1 = wasm_i16x8_shr(b.val, 8);
-    return v_int32x4(
-        v_dotprod(v_int16x8(a0), v_int16x8(b0)) +
-        v_dotprod(v_int16x8(a1), v_int16x8(b1))
-    );
-}
-inline v_int32x4 v_dotprod_expand(const v_int8x16& a, const v_int8x16& b, const v_int32x4& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-// 16 >> 64
-inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b)
-{
-    v128_t a0 = wasm_u32x4_shr(wasm_i32x4_shl(a.val, 16), 16);
-    v128_t a1 = wasm_u32x4_shr(a.val, 16);
-    v128_t b0 = wasm_u32x4_shr(wasm_i32x4_shl(b.val, 16), 16);
-    v128_t b1 = wasm_u32x4_shr(b.val, 16);
-    return v_uint64x2((
-        v_dotprod(v_int32x4(a0), v_int32x4(b0)) +
-        v_dotprod(v_int32x4(a1), v_int32x4(b1))).val
-    );
-}
-inline v_uint64x2 v_dotprod_expand(const v_uint16x8& a, const v_uint16x8& b, const v_uint64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-inline v_int64x2 v_dotprod_expand(const v_int16x8& a, const v_int16x8& b)
-{
-    v128_t a0 = wasm_i32x4_shr(wasm_i32x4_shl(a.val, 16), 16);
-    v128_t a1 = wasm_i32x4_shr(a.val, 16);
-    v128_t b0 = wasm_i32x4_shr(wasm_i32x4_shl(b.val, 16), 16);
-    v128_t b1 = wasm_i32x4_shr(b.val, 16);
-    return v_int64x2((
-        v_dotprod(v_int32x4(a0), v_int32x4(b0)) +
-        v_dotprod(v_int32x4(a1), v_int32x4(b1)))
-    );
-}
-
-inline v_int64x2 v_dotprod_expand(const v_int16x8& a, const v_int16x8& b, const v_int64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-// 32 >> 64f
-inline v_float64x2 v_dotprod_expand(const v_int32x4& a, const v_int32x4& b)
-{ return v_cvt_f64(v_dotprod(a, b)); }
-inline v_float64x2 v_dotprod_expand(const v_int32x4& a, const v_int32x4& b, const v_float64x2& c)
-{ return v_dotprod_expand(a, b) + c; }
-
-//////// Fast Dot Product ////////
-
-// 16 >> 32
-inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b)
-{ return v_dotprod(a, b); }
-inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b, const v_int32x4& c)
-{ return v_dotprod(a, b, c); }
-
-// 32 >> 64
-inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b)
-{ return v_dotprod(a, b); }
-inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b, const v_int64x2& c)
-{ return v_dotprod(a, b, c); }
-
-// 8 >> 32
-inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b)
-{ return v_dotprod_expand(a, b); }
-inline v_uint32x4 v_dotprod_expand_fast(const v_uint8x16& a, const v_uint8x16& b, const v_uint32x4& c)
-{ return v_dotprod_expand(a, b, c); }
-inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b)
-{ return v_dotprod_expand(a, b); }
-inline v_int32x4 v_dotprod_expand_fast(const v_int8x16& a, const v_int8x16& b, const v_int32x4& c)
-{ return v_dotprod_expand(a, b, c); }
-
-// 16 >> 64
-inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b)
-{ return v_dotprod_expand(a, b); }
-inline v_uint64x2 v_dotprod_expand_fast(const v_uint16x8& a, const v_uint16x8& b, const v_uint64x2& c)
-{ return v_dotprod_expand(a, b, c); }
-inline v_int64x2 v_dotprod_expand_fast(const v_int16x8& a, const v_int16x8& b)
-{ return v_dotprod_expand(a, b); }
-inline v_int64x2 v_dotprod_expand_fast(const v_int16x8& a, const v_int16x8& b, const v_int64x2& c)
-{ return v_dotprod_expand(a, b, c); }
-
-// 32 >> 64f
-inline v_float64x2 v_dotprod_expand_fast(const v_int32x4& a, const v_int32x4& b)
-{ return v_dotprod_expand(a, b); }
-inline v_float64x2 v_dotprod_expand_fast(const v_int32x4& a, const v_int32x4& b, const v_float64x2& c)
-{ return v_dotprod_expand(a, b, c); }
-
-#define OPENCV_HAL_IMPL_WASM_LOGIC_OP(_Tpvec) \
-OPENCV_HAL_IMPL_WASM_BIN_OP(&, _Tpvec, wasm_v128_and) \
-OPENCV_HAL_IMPL_WASM_BIN_OP(|, _Tpvec, wasm_v128_or) \
-OPENCV_HAL_IMPL_WASM_BIN_OP(^, _Tpvec, wasm_v128_xor) \
-inline _Tpvec operator ~ (const _Tpvec& a) \
-{ \
-    return _Tpvec(wasm_v128_not(a.val)); \
-}
-
-OPENCV_HAL_IMPL_WASM_LOGIC_OP(v_uint8x16)
-OPENCV_HAL_IMPL_WASM_LOGIC_OP(v_int8x16)
-OPENCV_HAL_IMPL_WASM_LOGIC_OP(v_uint16x8)
-OPENCV_HAL_IMPL_WASM_LOGIC_OP(v_int16x8)
-OPENCV_HAL_IMPL_WASM_LOGIC_OP(v_uint32x4)
-OPENCV_HAL_IMPL_WASM_LOGIC_OP(v_int32x4)
-OPENCV_HAL_IMPL_WASM_LOGIC_OP(v_uint64x2)
-OPENCV_HAL_IMPL_WASM_LOGIC_OP(v_int64x2)
-OPENCV_HAL_IMPL_WASM_LOGIC_OP(v_float32x4)
-OPENCV_HAL_IMPL_WASM_LOGIC_OP(v_float64x2)
-
-inline v_float32x4 v_sqrt(const v_float32x4& x)
-{
-    return v_float32x4(wasm_f32x4_sqrt(x.val));
-}
-
-inline v_float32x4 v_invsqrt(const v_float32x4& x)
-{
-    const v128_t _1_0 = wasm_f32x4_splat(1.0);
-    return v_float32x4(wasm_f32x4_div(_1_0, wasm_f32x4_sqrt(x.val)));
-}
-
-inline v_float64x2 v_sqrt(const v_float64x2& x)
-{
-    return v_float64x2(wasm_f64x2_sqrt(x.val));
-}
-
-inline v_float64x2 v_invsqrt(const v_float64x2& x)
-{
-    const v128_t _1_0 = wasm_f64x2_splat(1.0);
-    return v_float64x2(wasm_f64x2_div(_1_0, wasm_f64x2_sqrt(x.val)));
-}
-
-#define OPENCV_HAL_IMPL_WASM_ABS_INT_FUNC(_Tpuvec, _Tpsvec, suffix, zsuffix, shiftWidth) \
-inline _Tpuvec v_abs(const _Tpsvec& x) \
-{ \
-    v128_t s = wasm_##suffix##_shr(x.val, shiftWidth); \
-    v128_t f = wasm_##zsuffix##_shr(x.val, shiftWidth); \
-    return _Tpuvec(wasm_##zsuffix##_add(wasm_v128_xor(x.val, f), s)); \
-}
-
-OPENCV_HAL_IMPL_WASM_ABS_INT_FUNC(v_uint8x16, v_int8x16, u8x16, i8x16, 7)
-OPENCV_HAL_IMPL_WASM_ABS_INT_FUNC(v_uint16x8, v_int16x8, u16x8, i16x8, 15)
-OPENCV_HAL_IMPL_WASM_ABS_INT_FUNC(v_uint32x4, v_int32x4, u32x4, i32x4, 31)
-
-inline v_float32x4 v_abs(const v_float32x4& x)
-{ return v_float32x4(wasm_f32x4_abs(x.val)); }
-inline v_float64x2 v_abs(const v_float64x2& x)
-{
-    return v_float64x2(wasm_f64x2_abs(x.val));
-}
-
-// TODO: exp, log, sin, cos
-
-#define OPENCV_HAL_IMPL_WASM_BIN_FUNC(_Tpvec, func, intrin) \
-inline _Tpvec func(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(intrin(a.val, b.val)); \
-}
-
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_float32x4, v_min, wasm_f32x4_min)
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_float32x4, v_max, wasm_f32x4_max)
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_float64x2, v_min, wasm_f64x2_min)
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_float64x2, v_max, wasm_f64x2_max)
-
-#define OPENCV_HAL_IMPL_WASM_MINMAX_S_INIT_FUNC(_Tpvec, suffix) \
-inline _Tpvec v_min(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(wasm_v128_bitselect(b.val, a.val, wasm_##suffix##_gt(a.val, b.val))); \
-} \
-inline _Tpvec v_max(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(wasm_v128_bitselect(a.val, b.val, wasm_##suffix##_gt(a.val, b.val))); \
-}
-
-OPENCV_HAL_IMPL_WASM_MINMAX_S_INIT_FUNC(v_int8x16, i8x16)
-OPENCV_HAL_IMPL_WASM_MINMAX_S_INIT_FUNC(v_int16x8, i16x8)
-OPENCV_HAL_IMPL_WASM_MINMAX_S_INIT_FUNC(v_int32x4, i32x4)
-
-#define OPENCV_HAL_IMPL_WASM_MINMAX_U_INIT_FUNC(_Tpvec, suffix, deltaNum) \
-inline _Tpvec v_min(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    v128_t delta = wasm_##suffix##_splat(deltaNum); \
-    v128_t mask = wasm_##suffix##_gt(wasm_v128_xor(a.val, delta), wasm_v128_xor(b.val, delta)); \
-    return _Tpvec(wasm_v128_bitselect(b.val, a.val, mask)); \
-} \
-inline _Tpvec v_max(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    v128_t delta = wasm_##suffix##_splat(deltaNum); \
-    v128_t mask = wasm_##suffix##_gt(wasm_v128_xor(a.val, delta), wasm_v128_xor(b.val, delta)); \
-    return _Tpvec(wasm_v128_bitselect(a.val, b.val, mask)); \
-}
-
-OPENCV_HAL_IMPL_WASM_MINMAX_U_INIT_FUNC(v_uint8x16, i8x16, (schar)0x80)
-OPENCV_HAL_IMPL_WASM_MINMAX_U_INIT_FUNC(v_uint16x8, i16x8, (short)0x8000)
-OPENCV_HAL_IMPL_WASM_MINMAX_U_INIT_FUNC(v_uint32x4, i32x4, (int)0x80000000)
-
-#define OPENCV_HAL_IMPL_WASM_INIT_CMP_OP(_Tpvec, suffix, esuffix) \
-inline _Tpvec operator == (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(wasm_##esuffix##_eq(a.val, b.val)); } \
-inline _Tpvec operator != (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(wasm_##esuffix##_ne(a.val, b.val)); } \
-inline _Tpvec operator < (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(wasm_##suffix##_lt(a.val, b.val)); } \
-inline _Tpvec operator > (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(wasm_##suffix##_gt(a.val, b.val)); } \
-inline _Tpvec operator <= (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(wasm_##suffix##_le(a.val, b.val)); } \
-inline _Tpvec operator >= (const _Tpvec& a, const _Tpvec& b) \
-{ return _Tpvec(wasm_##suffix##_ge(a.val, b.val)); }
-
-OPENCV_HAL_IMPL_WASM_INIT_CMP_OP(v_uint8x16, u8x16, i8x16)
-OPENCV_HAL_IMPL_WASM_INIT_CMP_OP(v_int8x16, i8x16, i8x16)
-OPENCV_HAL_IMPL_WASM_INIT_CMP_OP(v_uint16x8, u16x8, i16x8)
-OPENCV_HAL_IMPL_WASM_INIT_CMP_OP(v_int16x8, i16x8, i16x8)
-OPENCV_HAL_IMPL_WASM_INIT_CMP_OP(v_uint32x4, u32x4, i32x4)
-OPENCV_HAL_IMPL_WASM_INIT_CMP_OP(v_int32x4, i32x4, i32x4)
-OPENCV_HAL_IMPL_WASM_INIT_CMP_OP(v_float32x4, f32x4, f32x4)
-OPENCV_HAL_IMPL_WASM_INIT_CMP_OP(v_float64x2, f64x2, f64x2)
-
-#define OPENCV_HAL_IMPL_WASM_64BIT_CMP_OP(_Tpvec, cast) \
-inline _Tpvec operator == (const _Tpvec& a, const _Tpvec& b) \
-{ return cast(v_reinterpret_as_f64(a) == v_reinterpret_as_f64(b)); } \
-inline _Tpvec operator != (const _Tpvec& a, const _Tpvec& b) \
-{ return cast(v_reinterpret_as_f64(a) != v_reinterpret_as_f64(b)); }
-
-OPENCV_HAL_IMPL_WASM_64BIT_CMP_OP(v_uint64x2, v_reinterpret_as_u64)
-OPENCV_HAL_IMPL_WASM_64BIT_CMP_OP(v_int64x2, v_reinterpret_as_s64)
-
-inline v_float32x4 v_not_nan(const v_float32x4& a)
-{
-    v128_t z = wasm_i32x4_splat(0x7fffffff);
-    v128_t t = wasm_i32x4_splat(0x7f800000);
-    return v_float32x4(wasm_u32x4_lt(wasm_v128_and(a.val, z), t));
-}
-inline v_float64x2 v_not_nan(const v_float64x2& a)
-{
-    v128_t z = wasm_i64x2_splat(0x7fffffffffffffff);
-    v128_t t = wasm_i64x2_splat(0x7ff0000000000000);
-    return v_float64x2((__u64x2)(wasm_v128_and(a.val, z)) < (__u64x2)t);
-}
-
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_uint8x16, v_add_wrap, wasm_i8x16_add)
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_int8x16, v_add_wrap, wasm_i8x16_add)
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_uint16x8, v_add_wrap, wasm_i16x8_add)
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_int16x8, v_add_wrap, wasm_i16x8_add)
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_uint8x16, v_sub_wrap, wasm_i8x16_sub)
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_int8x16, v_sub_wrap, wasm_i8x16_sub)
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_uint16x8, v_sub_wrap, wasm_i16x8_sub)
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_int16x8, v_sub_wrap, wasm_i16x8_sub)
-#if (__EMSCRIPTEN_major__ * 1000000 + __EMSCRIPTEN_minor__ * 1000 + __EMSCRIPTEN_tiny__) >= (1039012)
-// details: https://github.com/opencv/opencv/issues/18097 ( https://github.com/emscripten-core/emscripten/issues/12018 )
-// 1.39.12: https://github.com/emscripten-core/emscripten/commit/cd801d0f110facfd694212a3c8b2ed2ffcd630e2
-inline v_uint8x16 v_mul_wrap(const v_uint8x16& a, const v_uint8x16& b)
-{
-    uchar a_[16], b_[16];
-    wasm_v128_store(a_, a.val);
-    wasm_v128_store(b_, b.val);
-    for (int i = 0; i < 16; i++)
-        a_[i] = (uchar)(a_[i] * b_[i]);
-    return v_uint8x16(wasm_v128_load(a_));
-}
-inline v_int8x16 v_mul_wrap(const v_int8x16& a, const v_int8x16& b)
-{
-    schar a_[16], b_[16];
-    wasm_v128_store(a_, a.val);
-    wasm_v128_store(b_, b.val);
-    for (int i = 0; i < 16; i++)
-        a_[i] = (schar)(a_[i] * b_[i]);
-    return v_int8x16(wasm_v128_load(a_));
-}
-#else
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_uint8x16, v_mul_wrap, wasm_i8x16_mul)
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_int8x16, v_mul_wrap, wasm_i8x16_mul)
-#endif
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_uint16x8, v_mul_wrap, wasm_i16x8_mul)
-OPENCV_HAL_IMPL_WASM_BIN_FUNC(v_int16x8, v_mul_wrap, wasm_i16x8_mul)
-
-
-/** Absolute difference **/
-
-inline v_uint8x16 v_absdiff(const v_uint8x16& a, const v_uint8x16& b)
-{ return v_add_wrap(a - b,  b - a); }
-inline v_uint16x8 v_absdiff(const v_uint16x8& a, const v_uint16x8& b)
-{ return v_add_wrap(a - b,  b - a); }
-inline v_uint32x4 v_absdiff(const v_uint32x4& a, const v_uint32x4& b)
-{ return v_max(a, b) - v_min(a, b); }
-
-inline v_uint8x16 v_absdiff(const v_int8x16& a, const v_int8x16& b)
-{
-    v_int8x16 d = v_sub_wrap(a, b);
-    v_int8x16 m = a < b;
-    return v_reinterpret_as_u8(v_sub_wrap(d ^ m, m));
-}
-inline v_uint16x8 v_absdiff(const v_int16x8& a, const v_int16x8& b)
-{
-    return v_reinterpret_as_u16(v_sub_wrap(v_max(a, b), v_min(a, b)));
-}
-inline v_uint32x4 v_absdiff(const v_int32x4& a, const v_int32x4& b)
-{
-    v_int32x4 d = a - b;
-    v_int32x4 m = a < b;
-    return v_reinterpret_as_u32((d ^ m) - m);
-}
-
-/** Saturating absolute difference **/
-inline v_int8x16 v_absdiffs(const v_int8x16& a, const v_int8x16& b)
-{
-    v_int8x16 d = a - b;
-    v_int8x16 m = a < b;
-    return (d ^ m) - m;
- }
-inline v_int16x8 v_absdiffs(const v_int16x8& a, const v_int16x8& b)
-{ return v_max(a, b) - v_min(a, b); }
-
-
-inline v_int32x4 v_fma(const v_int32x4& a, const v_int32x4& b, const v_int32x4& c)
-{
-    return a * b + c;
-}
-
-inline v_int32x4 v_muladd(const v_int32x4& a, const v_int32x4& b, const v_int32x4& c)
-{
-    return v_fma(a, b, c);
-}
-
-inline v_float32x4 v_fma(const v_float32x4& a, const v_float32x4& b, const v_float32x4& c)
-{
-    return a * b + c;
-}
-
-inline v_float64x2 v_fma(const v_float64x2& a, const v_float64x2& b, const v_float64x2& c)
-{
-    return a * b + c;
-}
-
-inline v_float32x4 v_absdiff(const v_float32x4& a, const v_float32x4& b)
-{
-    v128_t absmask_vec = wasm_i32x4_splat(0x7fffffff);
-    return v_float32x4(wasm_v128_and(wasm_f32x4_sub(a.val, b.val), absmask_vec));
-}
-inline v_float64x2 v_absdiff(const v_float64x2& a, const v_float64x2& b)
-{
-    v128_t absmask_vec = wasm_u64x2_shr(wasm_i32x4_splat(-1), 1);
-    return v_float64x2(wasm_v128_and(wasm_f64x2_sub(a.val, b.val), absmask_vec));
-}
-
-#define OPENCV_HAL_IMPL_WASM_MISC_FLT_OP(_Tpvec, suffix) \
-inline _Tpvec v_magnitude(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    v128_t a_Square = wasm_##suffix##_mul(a.val, a.val); \
-    v128_t b_Square = wasm_##suffix##_mul(b.val, b.val); \
-    return _Tpvec(wasm_##suffix##_sqrt(wasm_##suffix##_add(a_Square, b_Square))); \
-} \
-inline _Tpvec v_sqr_magnitude(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    v128_t a_Square = wasm_##suffix##_mul(a.val, a.val); \
-    v128_t b_Square = wasm_##suffix##_mul(b.val, b.val); \
-    return _Tpvec(wasm_##suffix##_add(a_Square, b_Square)); \
-} \
-inline _Tpvec v_muladd(const _Tpvec& a, const _Tpvec& b, const _Tpvec& c) \
-{ \
-    return _Tpvec(wasm_##suffix##_add(wasm_##suffix##_mul(a.val, b.val), c.val)); \
-}
-
-OPENCV_HAL_IMPL_WASM_MISC_FLT_OP(v_float32x4, f32x4)
-OPENCV_HAL_IMPL_WASM_MISC_FLT_OP(v_float64x2, f64x2)
-
-#define OPENCV_HAL_IMPL_WASM_SHIFT_OP(_Tpuvec, _Tpsvec, suffix, ssuffix) \
-inline _Tpuvec operator << (const _Tpuvec& a, int imm) \
-{ \
-    return _Tpuvec(wasm_##suffix##_shl(a.val, imm)); \
-} \
-inline _Tpsvec operator << (const _Tpsvec& a, int imm) \
-{ \
-    return _Tpsvec(wasm_##suffix##_shl(a.val, imm)); \
-} \
-inline _Tpuvec operator >> (const _Tpuvec& a, int imm) \
-{ \
-    return _Tpuvec(wasm_##ssuffix##_shr(a.val, imm)); \
-} \
-inline _Tpsvec operator >> (const _Tpsvec& a, int imm) \
-{ \
-    return _Tpsvec(wasm_##suffix##_shr(a.val, imm)); \
-} \
-template<int imm> \
-inline _Tpuvec v_shl(const _Tpuvec& a) \
-{ \
-    return _Tpuvec(wasm_##suffix##_shl(a.val, imm)); \
-} \
-template<int imm> \
-inline _Tpsvec v_shl(const _Tpsvec& a) \
-{ \
-    return _Tpsvec(wasm_##suffix##_shl(a.val, imm)); \
-} \
-template<int imm> \
-inline _Tpuvec v_shr(const _Tpuvec& a) \
-{ \
-    return _Tpuvec(wasm_##ssuffix##_shr(a.val, imm)); \
-} \
-template<int imm> \
-inline _Tpsvec v_shr(const _Tpsvec& a) \
-{ \
-    return _Tpsvec(wasm_##suffix##_shr(a.val, imm)); \
-}
-
-OPENCV_HAL_IMPL_WASM_SHIFT_OP(v_uint8x16, v_int8x16, i8x16, u8x16)
-OPENCV_HAL_IMPL_WASM_SHIFT_OP(v_uint16x8, v_int16x8, i16x8, u16x8)
-OPENCV_HAL_IMPL_WASM_SHIFT_OP(v_uint32x4, v_int32x4, i32x4, u32x4)
-OPENCV_HAL_IMPL_WASM_SHIFT_OP(v_uint64x2, v_int64x2, i64x2, u64x2)
-
-namespace hal_wasm_internal
-{
-    template <int imm,
-        bool is_invalid = ((imm < 0) || (imm > 16)),
-        bool is_first = (imm == 0),
-        bool is_second = (imm == 16),
-        bool is_other = (((imm > 0) && (imm < 16)))>
-    class v_wasm_palignr_u8_class;
-
-    template <int imm>
-    class v_wasm_palignr_u8_class<imm, true, false, false, false>;
-
-    template <int imm>
-    class v_wasm_palignr_u8_class<imm, false, true, false, false>
-    {
-    public:
-        inline v128_t operator()(const v128_t& a, const v128_t&) const
-        {
-            return a;
-        }
-    };
-
-    template <int imm>
-    class v_wasm_palignr_u8_class<imm, false, false, true, false>
-    {
-    public:
-        inline v128_t operator()(const v128_t&, const v128_t& b) const
-        {
-            return b;
-        }
-    };
-
-    template <int imm>
-    class v_wasm_palignr_u8_class<imm, false, false, false, true>
-    {
-    public:
-        inline v128_t operator()(const v128_t& a, const v128_t& b) const
-        {
-            enum { imm2 = (sizeof(v128_t) - imm) };
-            return wasm_v8x16_shuffle(a, b,
-                                      imm, imm+1, imm+2, imm+3,
-                                      imm+4, imm+5, imm+6, imm+7,
-                                      imm+8, imm+9, imm+10, imm+11,
-                                      imm+12, imm+13, imm+14, imm+15);
-        }
-    };
-
-    template <int imm>
-    inline v128_t v_wasm_palignr_u8(const v128_t& a, const v128_t& b)
-    {
-        CV_StaticAssert((imm >= 0) && (imm <= 16), "Invalid imm for v_wasm_palignr_u8.");
-        return v_wasm_palignr_u8_class<imm>()(a, b);
-    }
-}
-
-template<int imm, typename _Tpvec>
-inline _Tpvec v_rotate_right(const _Tpvec &a)
-{
-    using namespace hal_wasm_internal;
-    enum { imm2 = (imm * sizeof(typename _Tpvec::lane_type)) };
-    v128_t z = wasm_i8x16_splat(0);
-    return _Tpvec(v_wasm_palignr_u8<imm2>(a.val, z));
-}
-
-template<int imm, typename _Tpvec>
-inline _Tpvec v_rotate_left(const _Tpvec &a)
-{
-    using namespace hal_wasm_internal;
-    enum { imm2 = ((_Tpvec::nlanes - imm) * sizeof(typename _Tpvec::lane_type)) };
-    v128_t z = wasm_i8x16_splat(0);
-    return _Tpvec(v_wasm_palignr_u8<imm2>(z, a.val));
-}
-
-template<int imm, typename _Tpvec>
-inline _Tpvec v_rotate_right(const _Tpvec &a, const _Tpvec &b)
-{
-    using namespace hal_wasm_internal;
-    enum { imm2 = (imm * sizeof(typename _Tpvec::lane_type)) };
-    return _Tpvec(v_wasm_palignr_u8<imm2>(a.val, b.val));
-}
-
-template<int imm, typename _Tpvec>
-inline _Tpvec v_rotate_left(const _Tpvec &a, const _Tpvec &b)
-{
-    using namespace hal_wasm_internal;
-    enum { imm2 = ((_Tpvec::nlanes - imm) * sizeof(typename _Tpvec::lane_type)) };
-    return _Tpvec(v_wasm_palignr_u8<imm2>(b.val, a.val));
-}
-
-#define OPENCV_HAL_IMPL_WASM_LOADSTORE_INT_OP(_Tpvec, _Tp) \
-inline _Tpvec v_load(const _Tp* ptr) \
-{ return _Tpvec(wasm_v128_load(ptr)); } \
-inline _Tpvec v_load_aligned(const _Tp* ptr) \
-{ return _Tpvec(wasm_v128_load(ptr)); } \
-inline _Tpvec v_load_low(const _Tp* ptr) \
-{ \
-    _Tp tmp[_Tpvec::nlanes] = {0}; \
-    for (int i=0; i<_Tpvec::nlanes/2; ++i) { \
-        tmp[i] = ptr[i]; \
-    } \
-    return _Tpvec(wasm_v128_load(tmp)); \
-} \
-inline _Tpvec v_load_halves(const _Tp* ptr0, const _Tp* ptr1) \
-{ \
-    _Tp tmp[_Tpvec::nlanes]; \
-    for (int i=0; i<_Tpvec::nlanes/2; ++i) { \
-        tmp[i] = ptr0[i]; \
-        tmp[i+_Tpvec::nlanes/2] = ptr1[i]; \
-    } \
-    return _Tpvec(wasm_v128_load(tmp)); \
-} \
-inline void v_store(_Tp* ptr, const _Tpvec& a) \
-{ wasm_v128_store(ptr, a.val); } \
-inline void v_store_aligned(_Tp* ptr, const _Tpvec& a) \
-{ wasm_v128_store(ptr, a.val); } \
-inline void v_store_aligned_nocache(_Tp* ptr, const _Tpvec& a) \
-{ wasm_v128_store(ptr, a.val); } \
-inline void v_store(_Tp* ptr, const _Tpvec& a, hal::StoreMode /*mode*/) \
-{ \
-    wasm_v128_store(ptr, a.val); \
-} \
-inline void v_store_low(_Tp* ptr, const _Tpvec& a) \
-{ \
-    _Tpvec::lane_type a_[_Tpvec::nlanes]; \
-    wasm_v128_store(a_, a.val); \
-    for (int i = 0; i < (_Tpvec::nlanes / 2); i++) \
-        ptr[i] = a_[i]; \
-} \
-inline void v_store_high(_Tp* ptr, const _Tpvec& a) \
-{ \
-    _Tpvec::lane_type a_[_Tpvec::nlanes]; \
-    wasm_v128_store(a_, a.val); \
-    for (int i = 0; i < (_Tpvec::nlanes / 2); i++) \
-        ptr[i] = a_[i + (_Tpvec::nlanes / 2)]; \
-}
-
-OPENCV_HAL_IMPL_WASM_LOADSTORE_INT_OP(v_uint8x16, uchar)
-OPENCV_HAL_IMPL_WASM_LOADSTORE_INT_OP(v_int8x16, schar)
-OPENCV_HAL_IMPL_WASM_LOADSTORE_INT_OP(v_uint16x8, ushort)
-OPENCV_HAL_IMPL_WASM_LOADSTORE_INT_OP(v_int16x8, short)
-OPENCV_HAL_IMPL_WASM_LOADSTORE_INT_OP(v_uint32x4, unsigned)
-OPENCV_HAL_IMPL_WASM_LOADSTORE_INT_OP(v_int32x4, int)
-OPENCV_HAL_IMPL_WASM_LOADSTORE_INT_OP(v_uint64x2, uint64)
-OPENCV_HAL_IMPL_WASM_LOADSTORE_INT_OP(v_int64x2, int64)
-OPENCV_HAL_IMPL_WASM_LOADSTORE_INT_OP(v_float32x4, float)
-OPENCV_HAL_IMPL_WASM_LOADSTORE_INT_OP(v_float64x2, double)
-
-
-/** Reverse **/
-inline v_uint8x16 v_reverse(const v_uint8x16 &a)
-{ return v_uint8x16(wasm_v8x16_shuffle(a.val, a.val, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0)); }
-
-inline v_int8x16 v_reverse(const v_int8x16 &a)
-{ return v_reinterpret_as_s8(v_reverse(v_reinterpret_as_u8(a))); }
-
-inline v_uint16x8 v_reverse(const v_uint16x8 &a)
-{ return v_uint16x8(wasm_v8x16_shuffle(a.val, a.val, 14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1)); }
-
-inline v_int16x8 v_reverse(const v_int16x8 &a)
-{ return v_reinterpret_as_s16(v_reverse(v_reinterpret_as_u16(a))); }
-
-inline v_uint32x4 v_reverse(const v_uint32x4 &a)
-{ return v_uint32x4(wasm_v8x16_shuffle(a.val, a.val, 12, 13, 14, 15, 8, 9, 10, 11, 4, 5, 6, 7, 0, 1, 2, 3)); }
-
-inline v_int32x4 v_reverse(const v_int32x4 &a)
-{ return v_reinterpret_as_s32(v_reverse(v_reinterpret_as_u32(a))); }
-
-inline v_float32x4 v_reverse(const v_float32x4 &a)
-{ return v_reinterpret_as_f32(v_reverse(v_reinterpret_as_u32(a))); }
-
-inline v_uint64x2 v_reverse(const v_uint64x2 &a)
-{ return v_uint64x2(wasm_v8x16_shuffle(a.val, a.val, 8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 4, 5, 6, 7)); }
-
-inline v_int64x2 v_reverse(const v_int64x2 &a)
-{ return v_reinterpret_as_s64(v_reverse(v_reinterpret_as_u64(a))); }
-
-inline v_float64x2 v_reverse(const v_float64x2 &a)
-{ return v_reinterpret_as_f64(v_reverse(v_reinterpret_as_u64(a))); }
-
-
-#define OPENCV_HAL_IMPL_WASM_REDUCE_OP_4_SUM(_Tpvec, scalartype, regtype, suffix, esuffix) \
-inline scalartype v_reduce_sum(const _Tpvec& a) \
-{ \
-    regtype val = a.val; \
-    val = wasm_##suffix##_add(val, wasm_v8x16_shuffle(val, val, 8,9,10,11,12,13,14,15,0,1,2,3,4,5,6,7)); \
-    val = wasm_##suffix##_add(val, wasm_v8x16_shuffle(val, val, 4,5,6,7,8,9,10,11,12,13,14,15,0,1,2,3)); \
-    return (scalartype)wasm_##esuffix##_extract_lane(val, 0); \
-}
-
-OPENCV_HAL_IMPL_WASM_REDUCE_OP_4_SUM(v_uint32x4, unsigned, v128_t, i32x4, i32x4)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP_4_SUM(v_int32x4, int, v128_t, i32x4, i32x4)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP_4_SUM(v_float32x4, float, v128_t, f32x4, f32x4)
-
-// To do: Optimize v_reduce_sum with wasm intrin.
-//        Now use fallback implementation as there is no widening op in wasm intrin.
-
-#define OPENCV_HAL_IMPL_FALLBACK_REDUCE_OP_SUM(_Tpvec, scalartype) \
-inline scalartype v_reduce_sum(const _Tpvec& a) \
-{ \
-    _Tpvec::lane_type a_[_Tpvec::nlanes]; \
-    wasm_v128_store(a_, a.val); \
-    scalartype c = a_[0]; \
-    for (int i = 1; i < _Tpvec::nlanes; i++) \
-        c += a_[i]; \
-    return c; \
-}
-
-OPENCV_HAL_IMPL_FALLBACK_REDUCE_OP_SUM(v_uint8x16, unsigned)
-OPENCV_HAL_IMPL_FALLBACK_REDUCE_OP_SUM(v_int8x16, int)
-OPENCV_HAL_IMPL_FALLBACK_REDUCE_OP_SUM(v_uint16x8, unsigned)
-OPENCV_HAL_IMPL_FALLBACK_REDUCE_OP_SUM(v_int16x8, int)
-
-
-#define OPENCV_HAL_IMPL_WASM_REDUCE_OP_2_SUM(_Tpvec, scalartype, regtype, suffix, esuffix) \
-inline scalartype v_reduce_sum(const _Tpvec& a) \
-{ \
-    regtype val = a.val; \
-    val = wasm_##suffix##_add(val, wasm_v8x16_shuffle(val, val, 8,9,10,11,12,13,14,15,0,1,2,3,4,5,6,7)); \
-    return (scalartype)wasm_##esuffix##_extract_lane(val, 0); \
-}
-OPENCV_HAL_IMPL_WASM_REDUCE_OP_2_SUM(v_uint64x2, uint64, v128_t, i64x2, i64x2)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP_2_SUM(v_int64x2, int64,  v128_t, i64x2, i64x2)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP_2_SUM(v_float64x2, double,  v128_t, f64x2,f64x2)
-
-inline v_float32x4 v_reduce_sum4(const v_float32x4& a, const v_float32x4& b,
-                                 const v_float32x4& c, const v_float32x4& d)
-{
-    v128_t ac = wasm_f32x4_add(wasm_unpacklo_i32x4(a.val, c.val), wasm_unpackhi_i32x4(a.val, c.val));
-    v128_t bd = wasm_f32x4_add(wasm_unpacklo_i32x4(b.val, d.val), wasm_unpackhi_i32x4(b.val, d.val));
-    return v_float32x4(wasm_f32x4_add(wasm_unpacklo_i32x4(ac, bd), wasm_unpackhi_i32x4(ac, bd)));
-}
-
-#define OPENCV_HAL_IMPL_WASM_REDUCE_OP(_Tpvec, scalartype, func, scalar_func) \
-inline scalartype v_reduce_##func(const _Tpvec& a) \
-{ \
-    scalartype buf[_Tpvec::nlanes]; \
-    v_store(buf, a); \
-    scalartype tmp = buf[0]; \
-    for (int i=1; i<_Tpvec::nlanes; ++i) { \
-        tmp = scalar_func(tmp, buf[i]); \
-    } \
-    return tmp; \
-}
-
-OPENCV_HAL_IMPL_WASM_REDUCE_OP(v_uint8x16, uchar, max, std::max)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP(v_uint8x16, uchar, min, std::min)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP(v_int8x16, schar, max, std::max)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP(v_int8x16, schar, min, std::min)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP(v_uint16x8, ushort, max, std::max)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP(v_uint16x8, ushort, min, std::min)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP(v_int16x8, short, max, std::max)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP(v_int16x8, short, min, std::min)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP(v_uint32x4, unsigned, max, std::max)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP(v_uint32x4, unsigned, min, std::min)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP(v_int32x4, int, max, std::max)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP(v_int32x4, int, min, std::min)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP(v_float32x4, float, max, std::max)
-OPENCV_HAL_IMPL_WASM_REDUCE_OP(v_float32x4, float, min, std::min)
-
-inline unsigned v_reduce_sad(const v_uint8x16& a, const v_uint8x16& b)
-{
-    v_uint16x8 l16, h16;
-    v_uint32x4 l16_l32, l16_h32, h16_l32, h16_h32;
-    v_expand(v_absdiff(a, b), l16, h16);
-    v_expand(l16, l16_l32, l16_h32);
-    v_expand(h16, h16_l32, h16_h32);
-    return v_reduce_sum(l16_l32+l16_h32+h16_l32+h16_h32);
-}
-inline unsigned v_reduce_sad(const v_int8x16& a, const v_int8x16& b)
-{
-    v_uint16x8 l16, h16;
-    v_uint32x4 l16_l32, l16_h32, h16_l32, h16_h32;
-    v_expand(v_absdiff(a, b), l16, h16);
-    v_expand(l16, l16_l32, l16_h32);
-    v_expand(h16, h16_l32, h16_h32);
-    return v_reduce_sum(l16_l32+l16_h32+h16_l32+h16_h32);
-}
-inline unsigned v_reduce_sad(const v_uint16x8& a, const v_uint16x8& b)
-{
-    v_uint32x4 l, h;
-    v_expand(v_absdiff(a, b), l, h);
-    return v_reduce_sum(l + h);
-}
-inline unsigned v_reduce_sad(const v_int16x8& a, const v_int16x8& b)
-{
-    v_uint32x4 l, h;
-    v_expand(v_absdiff(a, b), l, h);
-    return v_reduce_sum(l + h);
-}
-inline unsigned v_reduce_sad(const v_uint32x4& a, const v_uint32x4& b)
-{
-    return v_reduce_sum(v_absdiff(a, b));
-}
-inline unsigned v_reduce_sad(const v_int32x4& a, const v_int32x4& b)
-{
-    return v_reduce_sum(v_absdiff(a, b));
-}
-inline float v_reduce_sad(const v_float32x4& a, const v_float32x4& b)
-{
-    return v_reduce_sum(v_absdiff(a, b));
-}
-
-inline v_uint8x16 v_popcount(const v_uint8x16& a)
-{
-    v128_t m1 = wasm_i32x4_splat(0x55555555);
-    v128_t m2 = wasm_i32x4_splat(0x33333333);
-    v128_t m4 = wasm_i32x4_splat(0x0f0f0f0f);
-    v128_t p = a.val;
-    p = wasm_i32x4_add(wasm_v128_and(wasm_u32x4_shr(p, 1), m1), wasm_v128_and(p, m1));
-    p = wasm_i32x4_add(wasm_v128_and(wasm_u32x4_shr(p, 2), m2), wasm_v128_and(p, m2));
-    p = wasm_i32x4_add(wasm_v128_and(wasm_u32x4_shr(p, 4), m4), wasm_v128_and(p, m4));
-    return v_uint8x16(p);
-}
-inline v_uint16x8 v_popcount(const v_uint16x8& a)
-{
-    v_uint8x16 p = v_popcount(v_reinterpret_as_u8(a));
-    p += v_rotate_right<1>(p);
-    return v_reinterpret_as_u16(p) & v_setall_u16(0x00ff);
-}
-inline v_uint32x4 v_popcount(const v_uint32x4& a)
-{
-    v_uint8x16 p = v_popcount(v_reinterpret_as_u8(a));
-    p += v_rotate_right<1>(p);
-    p += v_rotate_right<2>(p);
-    return v_reinterpret_as_u32(p) & v_setall_u32(0x000000ff);
-}
-inline v_uint64x2 v_popcount(const v_uint64x2& a)
-{
-    uint64 a_[2], b_[2] = { 0 };
-    wasm_v128_store(a_, a.val);
-    for (int i = 0; i < 16; i++)
-        b_[i / 8] += popCountTable[((uint8_t*)a_)[i]];
-    return v_uint64x2(wasm_v128_load(b_));
-}
-inline v_uint8x16 v_popcount(const v_int8x16& a)
-{ return v_popcount(v_reinterpret_as_u8(a)); }
-inline v_uint16x8 v_popcount(const v_int16x8& a)
-{ return v_popcount(v_reinterpret_as_u16(a)); }
-inline v_uint32x4 v_popcount(const v_int32x4& a)
-{ return v_popcount(v_reinterpret_as_u32(a)); }
-inline v_uint64x2 v_popcount(const v_int64x2& a)
-{ return v_popcount(v_reinterpret_as_u64(a)); }
-
-#define OPENCV_HAL_IMPL_WASM_CHECK_SIGNS(_Tpvec, suffix, scalarType) \
-inline int v_signmask(const _Tpvec& a) \
-{ \
-    _Tpvec::lane_type a_[_Tpvec::nlanes]; \
-    wasm_v128_store(a_, a.val); \
-    int mask = 0; \
-    for (int i = 0; i < _Tpvec::nlanes; i++) \
-        mask |= (reinterpret_int(a_[i]) < 0) << i; \
-    return mask; \
-} \
-inline bool v_check_all(const _Tpvec& a) \
-{ return wasm_i8x16_all_true(wasm_##suffix##_lt(a.val, wasm_##suffix##_splat(0))); } \
-inline bool v_check_any(const _Tpvec& a) \
-{ return wasm_i8x16_any_true(wasm_##suffix##_lt(a.val, wasm_##suffix##_splat(0)));; }
-
-OPENCV_HAL_IMPL_WASM_CHECK_SIGNS(v_uint8x16, i8x16, schar)
-OPENCV_HAL_IMPL_WASM_CHECK_SIGNS(v_int8x16, i8x16, schar)
-OPENCV_HAL_IMPL_WASM_CHECK_SIGNS(v_uint16x8, i16x8, short)
-OPENCV_HAL_IMPL_WASM_CHECK_SIGNS(v_int16x8, i16x8, short)
-OPENCV_HAL_IMPL_WASM_CHECK_SIGNS(v_uint32x4, i32x4, int)
-OPENCV_HAL_IMPL_WASM_CHECK_SIGNS(v_int32x4, i32x4, int)
-OPENCV_HAL_IMPL_WASM_CHECK_SIGNS(v_float32x4, i32x4, float)
-OPENCV_HAL_IMPL_WASM_CHECK_SIGNS(v_float64x2, f64x2, double)
-
-#define OPENCV_HAL_IMPL_WASM_CHECK_ALL_ANY(_Tpvec, suffix, esuffix) \
-inline bool v_check_all(const _Tpvec& a) \
-{ \
-    v128_t masked = v_reinterpret_as_##esuffix(a).val; \
-    masked = wasm_i32x4_replace_lane(masked, 0, 0xffffffff); \
-    masked = wasm_i32x4_replace_lane(masked, 2, 0xffffffff); \
-    return wasm_i8x16_all_true(wasm_##suffix##_lt(masked, wasm_##suffix##_splat(0))); \
-} \
-inline bool v_check_any(const _Tpvec& a) \
-{ \
-    v128_t masked = v_reinterpret_as_##esuffix(a).val; \
-    masked = wasm_i32x4_replace_lane(masked, 0, 0x0); \
-    masked = wasm_i32x4_replace_lane(masked, 2, 0x0); \
-    return wasm_i8x16_any_true(wasm_##suffix##_lt(masked, wasm_##suffix##_splat(0))); \
-} \
-
-OPENCV_HAL_IMPL_WASM_CHECK_ALL_ANY(v_int64x2, i32x4, s32)
-OPENCV_HAL_IMPL_WASM_CHECK_ALL_ANY(v_uint64x2, i32x4, u32)
-
-
-inline int v_scan_forward(const v_int8x16& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))); }
-inline int v_scan_forward(const v_uint8x16& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))); }
-inline int v_scan_forward(const v_int16x8& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 2; }
-inline int v_scan_forward(const v_uint16x8& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 2; }
-inline int v_scan_forward(const v_int32x4& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 4; }
-inline int v_scan_forward(const v_uint32x4& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 4; }
-inline int v_scan_forward(const v_float32x4& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 4; }
-inline int v_scan_forward(const v_int64x2& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 8; }
-inline int v_scan_forward(const v_uint64x2& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 8; }
-inline int v_scan_forward(const v_float64x2& a) { return trailingZeros32(v_signmask(v_reinterpret_as_s8(a))) / 8; }
-
-#define OPENCV_HAL_IMPL_WASM_SELECT(_Tpvec) \
-inline _Tpvec v_select(const _Tpvec& mask, const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(wasm_v128_bitselect(a.val, b.val, mask.val)); \
-}
-
-OPENCV_HAL_IMPL_WASM_SELECT(v_uint8x16)
-OPENCV_HAL_IMPL_WASM_SELECT(v_int8x16)
-OPENCV_HAL_IMPL_WASM_SELECT(v_uint16x8)
-OPENCV_HAL_IMPL_WASM_SELECT(v_int16x8)
-OPENCV_HAL_IMPL_WASM_SELECT(v_uint32x4)
-OPENCV_HAL_IMPL_WASM_SELECT(v_int32x4)
-OPENCV_HAL_IMPL_WASM_SELECT(v_uint64x2)
-OPENCV_HAL_IMPL_WASM_SELECT(v_int64x2)
-OPENCV_HAL_IMPL_WASM_SELECT(v_float32x4)
-OPENCV_HAL_IMPL_WASM_SELECT(v_float64x2)
-
-#define OPENCV_HAL_IMPL_WASM_EXPAND(_Tpvec, _Tpwvec, _Tp, intrin)    \
-inline void v_expand(const _Tpvec& a, _Tpwvec& b0, _Tpwvec& b1)      \
-{                                                                    \
-    b0.val = intrin(a.val);                                          \
-    b1.val = __CV_CAT(intrin, _high)(a.val);                         \
-}                                                                    \
-inline _Tpwvec v_expand_low(const _Tpvec& a)                         \
-{ return _Tpwvec(intrin(a.val)); }                                   \
-inline _Tpwvec v_expand_high(const _Tpvec& a)                        \
-{ return _Tpwvec(__CV_CAT(intrin, _high)(a.val)); }                  \
-inline _Tpwvec v_load_expand(const _Tp* ptr)                         \
-{                                                                    \
-    v128_t a = wasm_v128_load(ptr);                                  \
-    return _Tpwvec(intrin(a));                                       \
-}
-
-OPENCV_HAL_IMPL_WASM_EXPAND(v_uint8x16, v_uint16x8, uchar, v128_cvtu8x16_i16x8)
-OPENCV_HAL_IMPL_WASM_EXPAND(v_int8x16,  v_int16x8,  schar, v128_cvti8x16_i16x8)
-OPENCV_HAL_IMPL_WASM_EXPAND(v_uint16x8, v_uint32x4, ushort, v128_cvtu16x8_i32x4)
-OPENCV_HAL_IMPL_WASM_EXPAND(v_int16x8,  v_int32x4,  short, v128_cvti16x8_i32x4)
-OPENCV_HAL_IMPL_WASM_EXPAND(v_uint32x4, v_uint64x2, unsigned, v128_cvtu32x4_i64x2)
-OPENCV_HAL_IMPL_WASM_EXPAND(v_int32x4,  v_int64x2,  int, v128_cvti32x4_i64x2)
-
-#define OPENCV_HAL_IMPL_WASM_EXPAND_Q(_Tpvec, _Tp, intrin)  \
-inline _Tpvec v_load_expand_q(const _Tp* ptr)               \
-{                                                           \
-    v128_t a = wasm_v128_load(ptr);                         \
-    return _Tpvec(intrin(a));                               \
-}
-
-OPENCV_HAL_IMPL_WASM_EXPAND_Q(v_uint32x4, uchar, v128_cvtu8x16_i32x4)
-OPENCV_HAL_IMPL_WASM_EXPAND_Q(v_int32x4, schar, v128_cvti8x16_i32x4)
-
-#define OPENCV_HAL_IMPL_WASM_UNPACKS(_Tpvec, suffix) \
-inline void v_zip(const _Tpvec& a0, const _Tpvec& a1, _Tpvec& b0, _Tpvec& b1) \
-{ \
-    b0.val = wasm_unpacklo_##suffix(a0.val, a1.val); \
-    b1.val = wasm_unpackhi_##suffix(a0.val, a1.val); \
-} \
-inline _Tpvec v_combine_low(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(wasm_unpacklo_i64x2(a.val, b.val)); \
-} \
-inline _Tpvec v_combine_high(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _Tpvec(wasm_unpackhi_i64x2(a.val, b.val)); \
-} \
-inline void v_recombine(const _Tpvec& a, const _Tpvec& b, _Tpvec& c, _Tpvec& d) \
-{ \
-    c.val = wasm_unpacklo_i64x2(a.val, b.val); \
-    d.val = wasm_unpackhi_i64x2(a.val, b.val); \
-}
-
-OPENCV_HAL_IMPL_WASM_UNPACKS(v_uint8x16, i8x16)
-OPENCV_HAL_IMPL_WASM_UNPACKS(v_int8x16, i8x16)
-OPENCV_HAL_IMPL_WASM_UNPACKS(v_uint16x8, i16x8)
-OPENCV_HAL_IMPL_WASM_UNPACKS(v_int16x8, i16x8)
-OPENCV_HAL_IMPL_WASM_UNPACKS(v_uint32x4, i32x4)
-OPENCV_HAL_IMPL_WASM_UNPACKS(v_int32x4, i32x4)
-OPENCV_HAL_IMPL_WASM_UNPACKS(v_float32x4, i32x4)
-OPENCV_HAL_IMPL_WASM_UNPACKS(v_float64x2, i64x2)
-
-template<int s, typename _Tpvec>
-inline _Tpvec v_extract(const _Tpvec& a, const _Tpvec& b)
-{
-    return v_rotate_right<s>(a, b);
-}
-
-inline v_int32x4 v_round(const v_float32x4& a)
-{
-    v128_t h = wasm_f32x4_splat(0.5);
-    return v_int32x4(wasm_i32x4_trunc_saturate_f32x4(wasm_f32x4_add(a.val, h)));
-}
-
-inline v_int32x4 v_floor(const v_float32x4& a)
-{
-    v128_t a1 = wasm_i32x4_trunc_saturate_f32x4(a.val);
-    v128_t mask = wasm_f32x4_lt(a.val, wasm_f32x4_convert_i32x4(a1));
-    return v_int32x4(wasm_i32x4_add(a1, mask));
-}
-
-inline v_int32x4 v_ceil(const v_float32x4& a)
-{
-    v128_t a1 = wasm_i32x4_trunc_saturate_f32x4(a.val);
-    v128_t mask = wasm_f32x4_gt(a.val, wasm_f32x4_convert_i32x4(a1));
-    return v_int32x4(wasm_i32x4_sub(a1, mask));
-}
-
-inline v_int32x4 v_trunc(const v_float32x4& a)
-{ return v_int32x4(wasm_i32x4_trunc_saturate_f32x4(a.val)); }
-
-#define OPENCV_HAL_IMPL_WASM_MATH_FUNC(func, cfunc) \
-inline v_int32x4 func(const v_float64x2& a) \
-{ \
-    double a_[2]; \
-    wasm_v128_store(a_, a.val); \
-    int c_[4]; \
-    c_[0] = cfunc(a_[0]); \
-    c_[1] = cfunc(a_[1]); \
-    c_[2] = 0; \
-    c_[3] = 0; \
-    return v_int32x4(wasm_v128_load(c_)); \
-}
-
-OPENCV_HAL_IMPL_WASM_MATH_FUNC(v_round, cvRound)
-OPENCV_HAL_IMPL_WASM_MATH_FUNC(v_floor, cvFloor)
-OPENCV_HAL_IMPL_WASM_MATH_FUNC(v_ceil, cvCeil)
-OPENCV_HAL_IMPL_WASM_MATH_FUNC(v_trunc, int)
-
-inline v_int32x4 v_round(const v_float64x2& a, const v_float64x2& b)
-{
-    double a_[2], b_[2];
-    wasm_v128_store(a_, a.val);
-    wasm_v128_store(b_, b.val);
-    int c_[4];
-    c_[0] = cvRound(a_[0]);
-    c_[1] = cvRound(a_[1]);
-    c_[2] = cvRound(b_[0]);
-    c_[3] = cvRound(b_[1]);
-    return v_int32x4(wasm_v128_load(c_));
-}
-
-#define OPENCV_HAL_IMPL_WASM_TRANSPOSE4x4(_Tpvec, suffix) \
-inline void v_transpose4x4(const _Tpvec& a0, const _Tpvec& a1, \
-                           const _Tpvec& a2, const _Tpvec& a3, \
-                           _Tpvec& b0, _Tpvec& b1, \
-                           _Tpvec& b2, _Tpvec& b3) \
-{ \
-    v128_t t0 = wasm_unpacklo_##suffix(a0.val, a1.val); \
-    v128_t t1 = wasm_unpacklo_##suffix(a2.val, a3.val); \
-    v128_t t2 = wasm_unpackhi_##suffix(a0.val, a1.val); \
-    v128_t t3 = wasm_unpackhi_##suffix(a2.val, a3.val); \
-\
-    b0.val = wasm_unpacklo_i64x2(t0, t1); \
-    b1.val = wasm_unpackhi_i64x2(t0, t1); \
-    b2.val = wasm_unpacklo_i64x2(t2, t3); \
-    b3.val = wasm_unpackhi_i64x2(t2, t3); \
-}
-
-OPENCV_HAL_IMPL_WASM_TRANSPOSE4x4(v_uint32x4, i32x4)
-OPENCV_HAL_IMPL_WASM_TRANSPOSE4x4(v_int32x4, i32x4)
-OPENCV_HAL_IMPL_WASM_TRANSPOSE4x4(v_float32x4, i32x4)
-
-// load deinterleave
-inline void v_load_deinterleave(const uchar* ptr, v_uint8x16& a, v_uint8x16& b)
-{
-    v128_t t00 = wasm_v128_load(ptr);
-    v128_t t01 = wasm_v128_load(ptr + 16);
-
-    a.val = wasm_v8x16_shuffle(t00, t01, 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30);
-    b.val = wasm_v8x16_shuffle(t00, t01, 1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31);
-}
-
-inline void v_load_deinterleave(const uchar* ptr, v_uint8x16& a, v_uint8x16& b, v_uint8x16& c)
-{
-    v128_t t00 = wasm_v128_load(ptr);
-    v128_t t01 = wasm_v128_load(ptr + 16);
-    v128_t t02 = wasm_v128_load(ptr + 32);
-
-    v128_t t10 = wasm_v8x16_shuffle(t00, t01, 0,3,6,9,12,15,18,21,24,27,30,1,2,4,5,7);
-    v128_t t11 = wasm_v8x16_shuffle(t00, t01, 1,4,7,10,13,16,19,22,25,28,31,0,2,3,5,6);
-    v128_t t12 = wasm_v8x16_shuffle(t00, t01, 2,5,8,11,14,17,20,23,26,29,0,1,3,4,6,7);
-
-    a.val = wasm_v8x16_shuffle(t10, t02, 0,1,2,3,4,5,6,7,8,9,10,17,20,23,26,29);
-    b.val = wasm_v8x16_shuffle(t11, t02, 0,1,2,3,4,5,6,7,8,9,10,18,21,24,27,30);
-    c.val = wasm_v8x16_shuffle(t12, t02, 0,1,2,3,4,5,6,7,8,9,16,19,22,25,28,31);
-}
-
-inline void v_load_deinterleave(const uchar* ptr, v_uint8x16& a, v_uint8x16& b, v_uint8x16& c, v_uint8x16& d)
-{
-    v128_t u0 = wasm_v128_load(ptr); // a0 b0 c0 d0 a1 b1 c1 d1 ...
-    v128_t u1 = wasm_v128_load(ptr + 16); // a4 b4 c4 d4 ...
-    v128_t u2 = wasm_v128_load(ptr + 32); // a8 b8 c8 d8 ...
-    v128_t u3 = wasm_v128_load(ptr + 48); // a12 b12 c12 d12 ...
-
-    v128_t v0 = wasm_v8x16_shuffle(u0, u1, 0,4,8,12,16,20,24,28,1,5,9,13,17,21,25,29);
-    v128_t v1 = wasm_v8x16_shuffle(u2, u3, 0,4,8,12,16,20,24,28,1,5,9,13,17,21,25,29);
-    v128_t v2 = wasm_v8x16_shuffle(u0, u1, 2,6,10,14,18,22,26,30,3,7,11,15,19,23,27,31);
-    v128_t v3 = wasm_v8x16_shuffle(u2, u3, 2,6,10,14,18,22,26,30,3,7,11,15,19,23,27,31);
-
-    a.val = wasm_v8x16_shuffle(v0, v1, 0,1,2,3,4,5,6,7,16,17,18,19,20,21,22,23);
-    b.val = wasm_v8x16_shuffle(v0, v1, 8,9,10,11,12,13,14,15,24,25,26,27,28,29,30,31);
-    c.val = wasm_v8x16_shuffle(v2, v3, 0,1,2,3,4,5,6,7,16,17,18,19,20,21,22,23);
-    d.val = wasm_v8x16_shuffle(v2, v3, 8,9,10,11,12,13,14,15,24,25,26,27,28,29,30,31);
-}
-
-inline void v_load_deinterleave(const ushort* ptr, v_uint16x8& a, v_uint16x8& b)
-{
-    v128_t v0 = wasm_v128_load(ptr);     // a0 b0 a1 b1 a2 b2 a3 b3
-    v128_t v1 = wasm_v128_load(ptr + 8); // a4 b4 a5 b5 a6 b6 a7 b7
-
-    a.val = wasm_v8x16_shuffle(v0, v1, 0,1,4,5,8,9,12,13,16,17,20,21,24,25,28,29); // a0 a1 a2 a3 a4 a5 a6 a7
-    b.val = wasm_v8x16_shuffle(v0, v1, 2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31); // b0 b1 ab b3 b4 b5 b6 b7
-}
-
-inline void v_load_deinterleave(const ushort* ptr, v_uint16x8& a, v_uint16x8& b, v_uint16x8& c)
-{
-    v128_t t00 = wasm_v128_load(ptr);        // a0 b0 c0 a1 b1 c1 a2 b2
-    v128_t t01 = wasm_v128_load(ptr + 8);    // c2 a3 b3 c3 a4 b4 c4 a5
-    v128_t t02 = wasm_v128_load(ptr + 16);  // b5 c5 a6 b6 c6 a7 b7 c7
-
-    v128_t t10 = wasm_v8x16_shuffle(t00, t01, 0,1,6,7,12,13,18,19,24,25,30,31,2,3,4,5);
-    v128_t t11 = wasm_v8x16_shuffle(t00, t01, 2,3,8,9,14,15,20,21,26,27,0,1,4,5,6,7);
-    v128_t t12 = wasm_v8x16_shuffle(t00, t01, 4,5,10,11,16,17,22,23,28,29,0,1,2,3,6,7);
-
-    a.val = wasm_v8x16_shuffle(t10, t02, 0,1,2,3,4,5,6,7,8,9,10,11,20,21,26,27);
-    b.val = wasm_v8x16_shuffle(t11, t02, 0,1,2,3,4,5,6,7,8,9,16,17,22,23,28,29);
-    c.val = wasm_v8x16_shuffle(t12, t02, 0,1,2,3,4,5,6,7,8,9,18,19,24,25,30,31);
-}
-
-inline void v_load_deinterleave(const ushort* ptr, v_uint16x8& a, v_uint16x8& b, v_uint16x8& c, v_uint16x8& d)
-{
-    v128_t u0 = wasm_v128_load(ptr); // a0 b0 c0 d0 a1 b1 c1 d1
-    v128_t u1 = wasm_v128_load(ptr + 8); // a2 b2 c2 d2 ...
-    v128_t u2 = wasm_v128_load(ptr + 16); // a4 b4 c4 d4 ...
-    v128_t u3 = wasm_v128_load(ptr + 24); // a6 b6 c6 d6 ...
-
-    v128_t v0 = wasm_v8x16_shuffle(u0, u1, 0,1,8,9,16,17,24,25,2,3,10,11,18,19,26,27); // a0 a1 a2 a3 b0 b1 b2 b3
-    v128_t v1 = wasm_v8x16_shuffle(u2, u3, 0,1,8,9,16,17,24,25,2,3,10,11,18,19,26,27); // a4 a5 a6 a7 b4 b5 b6 b7
-    v128_t v2 = wasm_v8x16_shuffle(u0, u1, 4,5,12,13,20,21,28,29,6,7,14,15,22,23,30,31); // c0 c1 c2 c3 d0 d1 d2 d3
-    v128_t v3 = wasm_v8x16_shuffle(u2, u3, 4,5,12,13,20,21,28,29,6,7,14,15,22,23,30,31); // c4 c5 c6 c7 d4 d5 d6 d7
-
-    a.val = wasm_v8x16_shuffle(v0, v1, 0,1,2,3,4,5,6,7,16,17,18,19,20,21,22,23);
-    b.val = wasm_v8x16_shuffle(v0, v1, 8,9,10,11,12,13,14,15,24,25,26,27,28,29,30,31);
-    c.val = wasm_v8x16_shuffle(v2, v3, 0,1,2,3,4,5,6,7,16,17,18,19,20,21,22,23);
-    d.val = wasm_v8x16_shuffle(v2, v3, 8,9,10,11,12,13,14,15,24,25,26,27,28,29,30,31);
-}
-
-inline void v_load_deinterleave(const unsigned* ptr, v_uint32x4& a, v_uint32x4& b)
-{
-    v128_t v0 = wasm_v128_load(ptr);     // a0 b0 a1 b1
-    v128_t v1 = wasm_v128_load(ptr + 4); // a2 b2 a3 b3
-
-    a.val = wasm_v8x16_shuffle(v0, v1, 0,1,2,3,8,9,10,11,16,17,18,19,24,25,26,27); // a0 a1 a2 a3
-    b.val = wasm_v8x16_shuffle(v0, v1, 4,5,6,7,12,13,14,15,20,21,22,23,28,29,30,31); // b0 b1 b2 b3
-}
-
-inline void v_load_deinterleave(const unsigned* ptr, v_uint32x4& a, v_uint32x4& b, v_uint32x4& c)
-{
-    v128_t t00 = wasm_v128_load(ptr);        // a0 b0 c0 a1
-    v128_t t01 = wasm_v128_load(ptr + 4);     // b2 c2 a3 b3
-    v128_t t02 = wasm_v128_load(ptr + 8);    // c3 a4 b4 c4
-
-    v128_t t10 = wasm_v8x16_shuffle(t00, t01, 0,1,2,3,12,13,14,15,24,25,26,27,4,5,6,7);
-    v128_t t11 = wasm_v8x16_shuffle(t00, t01, 4,5,6,7,16,17,18,19,28,29,30,31,0,1,2,3);
-    v128_t t12 = wasm_v8x16_shuffle(t00, t01, 8,9,10,11,20,21,22,23,0,1,2,3,4,5,6,7);
-
-    a.val = wasm_v8x16_shuffle(t10, t02, 0,1,2,3,4,5,6,7,8,9,10,11,20,21,22,23);
-    b.val = wasm_v8x16_shuffle(t11, t02, 0,1,2,3,4,5,6,7,8,9,10,11,24,25,26,27);
-    c.val = wasm_v8x16_shuffle(t12, t02, 0,1,2,3,4,5,6,7,16,17,18,19,28,29,30,31);
-}
-
-inline void v_load_deinterleave(const unsigned* ptr, v_uint32x4& a, v_uint32x4& b, v_uint32x4& c, v_uint32x4& d)
-{
-    v_uint32x4 s0(wasm_v128_load(ptr));      // a0 b0 c0 d0
-    v_uint32x4 s1(wasm_v128_load(ptr + 4));  // a1 b1 c1 d1
-    v_uint32x4 s2(wasm_v128_load(ptr + 8));  // a2 b2 c2 d2
-    v_uint32x4 s3(wasm_v128_load(ptr + 12)); // a3 b3 c3 d3
-
-    v_transpose4x4(s0, s1, s2, s3, a, b, c, d);
-}
-
-inline void v_load_deinterleave(const float* ptr, v_float32x4& a, v_float32x4& b)
-{
-    v128_t v0 = wasm_v128_load(ptr);       // a0 b0 a1 b1
-    v128_t v1 = wasm_v128_load((ptr + 4)); // a2 b2 a3 b3
-
-    a.val = wasm_v8x16_shuffle(v0, v1, 0,1,2,3,8,9,10,11,16,17,18,19,24,25,26,27); // a0 a1 a2 a3
-    b.val = wasm_v8x16_shuffle(v0, v1, 4,5,6,7,12,13,14,15,20,21,22,23,28,29,30,31); // b0 b1 b2 b3
-}
-
-inline void v_load_deinterleave(const float* ptr, v_float32x4& a, v_float32x4& b, v_float32x4& c)
-{
-    v128_t t00 = wasm_v128_load(ptr);        // a0 b0 c0 a1
-    v128_t t01 = wasm_v128_load(ptr + 4);     // b2 c2 a3 b3
-    v128_t t02 = wasm_v128_load(ptr + 8);    // c3 a4 b4 c4
-
-    v128_t t10 = wasm_v8x16_shuffle(t00, t01, 0,1,2,3,12,13,14,15,24,25,26,27,4,5,6,7);
-    v128_t t11 = wasm_v8x16_shuffle(t00, t01, 4,5,6,7,16,17,18,19,28,29,30,31,0,1,2,3);
-    v128_t t12 = wasm_v8x16_shuffle(t00, t01, 8,9,10,11,20,21,22,23,0,1,2,3,4,5,6,7);
-
-    a.val = wasm_v8x16_shuffle(t10, t02, 0,1,2,3,4,5,6,7,8,9,10,11,20,21,22,23);
-    b.val = wasm_v8x16_shuffle(t11, t02, 0,1,2,3,4,5,6,7,8,9,10,11,24,25,26,27);
-    c.val = wasm_v8x16_shuffle(t12, t02, 0,1,2,3,4,5,6,7,16,17,18,19,28,29,30,31);
-}
-
-inline void v_load_deinterleave(const float* ptr, v_float32x4& a, v_float32x4& b, v_float32x4& c, v_float32x4& d)
-{
-    v_float32x4 s0(wasm_v128_load(ptr));      // a0 b0 c0 d0
-    v_float32x4 s1(wasm_v128_load(ptr + 4));  // a1 b1 c1 d1
-    v_float32x4 s2(wasm_v128_load(ptr + 8));  // a2 b2 c2 d2
-    v_float32x4 s3(wasm_v128_load(ptr + 12)); // a3 b3 c3 d3
-
-    v_transpose4x4(s0, s1, s2, s3, a, b, c, d);
-}
-
-inline void v_load_deinterleave(const uint64 *ptr, v_uint64x2& a, v_uint64x2& b)
-{
-    v128_t t0 = wasm_v128_load(ptr);      // a0 b0
-    v128_t t1 = wasm_v128_load(ptr + 2);  // a1 b1
-
-    a.val = wasm_unpacklo_i64x2(t0, t1);
-    b.val = wasm_unpackhi_i64x2(t0, t1);
-}
-
-inline void v_load_deinterleave(const uint64 *ptr, v_uint64x2& a, v_uint64x2& b, v_uint64x2& c)
-{
-    v128_t t0 = wasm_v128_load(ptr);     // a0, b0
-    v128_t t1 = wasm_v128_load(ptr + 2); // c0, a1
-    v128_t t2 = wasm_v128_load(ptr + 4); // b1, c1
-
-    a.val = wasm_v8x16_shuffle(t0, t1, 0,1,2,3,4,5,6,7,24,25,26,27,28,29,30,31);
-    b.val = wasm_v8x16_shuffle(t0, t2, 8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23);
-    c.val = wasm_v8x16_shuffle(t1, t2, 0,1,2,3,4,5,6,7,24,25,26,27,28,29,30,31);
-}
-
-inline void v_load_deinterleave(const uint64 *ptr, v_uint64x2& a,
-                                v_uint64x2& b, v_uint64x2& c, v_uint64x2& d)
-{
-    v128_t t0 = wasm_v128_load(ptr);     // a0 b0
-    v128_t t1 = wasm_v128_load(ptr + 2); // c0 d0
-    v128_t t2 = wasm_v128_load(ptr + 4); // a1 b1
-    v128_t t3 = wasm_v128_load(ptr + 6); // c1 d1
-
-    a.val = wasm_unpacklo_i64x2(t0, t2);
-    b.val = wasm_unpackhi_i64x2(t0, t2);
-    c.val = wasm_unpacklo_i64x2(t1, t3);
-    d.val = wasm_unpackhi_i64x2(t1, t3);
-}
-
-// store interleave
-
-inline void v_store_interleave( uchar* ptr, const v_uint8x16& a, const v_uint8x16& b,
-                                hal::StoreMode /*mode*/ = hal::STORE_UNALIGNED)
-{
-    v128_t v0 = wasm_unpacklo_i8x16(a.val, b.val);
-    v128_t v1 = wasm_unpackhi_i8x16(a.val, b.val);
-
-    wasm_v128_store(ptr, v0);
-    wasm_v128_store(ptr + 16, v1);
-}
-
-inline void v_store_interleave( uchar* ptr, const v_uint8x16& a, const v_uint8x16& b,
-                                const v_uint8x16& c, hal::StoreMode /*mode*/ = hal::STORE_UNALIGNED)
-{
-    v128_t t00 = wasm_v8x16_shuffle(a.val, b.val, 0,16,0,1,17,0,2,18,0,3,19,0,4,20,0,5);
-    v128_t t01 = wasm_v8x16_shuffle(a.val, b.val, 21,0,6,22,0,7,23,0,8,24,0,9,25,0,10,26);
-    v128_t t02 = wasm_v8x16_shuffle(a.val, b.val, 0,11,27,0,12,28,0,13,29,0,14,30,0,15,31,0);
-
-    v128_t t10 = wasm_v8x16_shuffle(t00, c.val, 0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15);
-    v128_t t11 = wasm_v8x16_shuffle(t01, c.val, 0,21,2,3,22,5,6,23,8,9,24,11,12,25,14,15);
-    v128_t t12 = wasm_v8x16_shuffle(t02, c.val, 26,1,2,27,4,5,28,7,8,29,10,11,30,13,14,31);
-
-    wasm_v128_store(ptr, t10);
-    wasm_v128_store(ptr + 16, t11);
-    wasm_v128_store(ptr + 32, t12);
-}
-
-inline void v_store_interleave( uchar* ptr, const v_uint8x16& a, const v_uint8x16& b,
-                                const v_uint8x16& c, const v_uint8x16& d,
-                                hal::StoreMode /*mode*/ = hal::STORE_UNALIGNED)
-{
-    // a0 a1 a2 a3 ....
-    // b0 b1 b2 b3 ....
-    // c0 c1 c2 c3 ....
-    // d0 d1 d2 d3 ....
-    v128_t u0 = wasm_unpacklo_i8x16(a.val, c.val); // a0 c0 a1 c1 ...
-    v128_t u1 = wasm_unpackhi_i8x16(a.val, c.val); // a8 c8 a9 c9 ...
-    v128_t u2 = wasm_unpacklo_i8x16(b.val, d.val); // b0 d0 b1 d1 ...
-    v128_t u3 = wasm_unpackhi_i8x16(b.val, d.val); // b8 d8 b9 d9 ...
-
-    v128_t v0 = wasm_unpacklo_i8x16(u0, u2); // a0 b0 c0 d0 ...
-    v128_t v1 = wasm_unpackhi_i8x16(u0, u2); // a4 b4 c4 d4 ...
-    v128_t v2 = wasm_unpacklo_i8x16(u1, u3); // a8 b8 c8 d8 ...
-    v128_t v3 = wasm_unpackhi_i8x16(u1, u3); // a12 b12 c12 d12 ...
-
-    wasm_v128_store(ptr, v0);
-    wasm_v128_store(ptr + 16, v1);
-    wasm_v128_store(ptr + 32, v2);
-    wasm_v128_store(ptr + 48, v3);
-}
-
-inline void v_store_interleave( ushort* ptr, const v_uint16x8& a, const v_uint16x8& b,
-                                hal::StoreMode /*mode*/ = hal::STORE_UNALIGNED)
-{
-    v128_t v0 = wasm_unpacklo_i16x8(a.val, b.val);
-    v128_t v1 = wasm_unpackhi_i16x8(a.val, b.val);
-
-    wasm_v128_store(ptr, v0);
-    wasm_v128_store(ptr + 8, v1);
-}
-
-inline void v_store_interleave( ushort* ptr, const v_uint16x8& a,
-                                const v_uint16x8& b, const v_uint16x8& c,
-                                hal::StoreMode /*mode*/ = hal::STORE_UNALIGNED)
-{
-    v128_t t00 = wasm_v8x16_shuffle(a.val, b.val, 0,1,16,17,0,0,2,3,18,19,0,0,4,5,20,21);
-    v128_t t01 = wasm_v8x16_shuffle(a.val, b.val, 0,0,6,7,22,23,0,0,8,9,24,25,0,0,10,11);
-    v128_t t02 = wasm_v8x16_shuffle(a.val, b.val, 26,27,0,0,12,13,28,29,0,0,14,15,30,31,0,0);
-
-    v128_t t10 = wasm_v8x16_shuffle(t00, c.val, 0,1,2,3,16,17,6,7,8,9,18,19,12,13,14,15);
-    v128_t t11 = wasm_v8x16_shuffle(t01, c.val, 20,21,2,3,4,5,22,23,8,9,10,11,24,25,14,15);
-    v128_t t12 = wasm_v8x16_shuffle(t02, c.val, 0,1,26,27,4,5,6,7,28,29,10,11,12,13,30,31);
-
-    wasm_v128_store(ptr, t10);
-    wasm_v128_store(ptr + 8, t11);
-    wasm_v128_store(ptr + 16, t12);
-}
-
-inline void v_store_interleave( ushort* ptr, const v_uint16x8& a, const v_uint16x8& b,
-                                const v_uint16x8& c, const v_uint16x8& d,
-                                hal::StoreMode /*mode*/ = hal::STORE_UNALIGNED)
-{
-    // a0 a1 a2 a3 ....
-    // b0 b1 b2 b3 ....
-    // c0 c1 c2 c3 ....
-    // d0 d1 d2 d3 ....
-    v128_t u0 = wasm_unpacklo_i16x8(a.val, c.val); // a0 c0 a1 c1 ...
-    v128_t u1 = wasm_unpackhi_i16x8(a.val, c.val); // a4 c4 a5 c5 ...
-    v128_t u2 = wasm_unpacklo_i16x8(b.val, d.val); // b0 d0 b1 d1 ...
-    v128_t u3 = wasm_unpackhi_i16x8(b.val, d.val); // b4 d4 b5 d5 ...
-
-    v128_t v0 = wasm_unpacklo_i16x8(u0, u2); // a0 b0 c0 d0 ...
-    v128_t v1 = wasm_unpackhi_i16x8(u0, u2); // a2 b2 c2 d2 ...
-    v128_t v2 = wasm_unpacklo_i16x8(u1, u3); // a4 b4 c4 d4 ...
-    v128_t v3 = wasm_unpackhi_i16x8(u1, u3); // a6 b6 c6 d6 ...
-
-    wasm_v128_store(ptr, v0);
-    wasm_v128_store(ptr + 8, v1);
-    wasm_v128_store(ptr + 16, v2);
-    wasm_v128_store(ptr + 24, v3);
-}
-
-inline void v_store_interleave( unsigned* ptr, const v_uint32x4& a, const v_uint32x4& b,
-                                hal::StoreMode /*mode*/ = hal::STORE_UNALIGNED)
-{
-    v128_t v0 = wasm_unpacklo_i32x4(a.val, b.val);
-    v128_t v1 = wasm_unpackhi_i32x4(a.val, b.val);
-
-    wasm_v128_store(ptr, v0);
-    wasm_v128_store(ptr + 4, v1);
-}
-
-inline void v_store_interleave( unsigned* ptr, const v_uint32x4& a, const v_uint32x4& b,
-                                const v_uint32x4& c, hal::StoreMode /*mode*/ = hal::STORE_UNALIGNED)
-{
-    v128_t t00 = wasm_v8x16_shuffle(a.val, b.val, 0,1,2,3,16,17,18,19,0,0,0,0,4,5,6,7);
-    v128_t t01 = wasm_v8x16_shuffle(a.val, b.val, 20,21,22,23,0,0,0,0,8,9,10,11,24,25,26,27);
-    v128_t t02 = wasm_v8x16_shuffle(a.val, b.val, 0,0,0,0,12,13,14,15,28,29,30,31,0,0,0,0);
-
-    v128_t t10 = wasm_v8x16_shuffle(t00, c.val, 0,1,2,3,4,5,6,7,16,17,18,19,12,13,14,15);
-    v128_t t11 = wasm_v8x16_shuffle(t01, c.val, 0,1,2,3,20,21,22,23,8,9,10,11,12,13,14,15);
-    v128_t t12 = wasm_v8x16_shuffle(t02, c.val, 24,25,26,27,4,5,6,7,8,9,10,11,28,29,30,31);
-
-    wasm_v128_store(ptr, t10);
-    wasm_v128_store(ptr + 4, t11);
-    wasm_v128_store(ptr + 8, t12);
-}
-
-inline void v_store_interleave(unsigned* ptr, const v_uint32x4& a, const v_uint32x4& b,
-                               const v_uint32x4& c, const v_uint32x4& d,
-                               hal::StoreMode /*mode*/ = hal::STORE_UNALIGNED)
-{
-    v_uint32x4 v0, v1, v2, v3;
-    v_transpose4x4(a, b, c, d, v0, v1, v2, v3);
-
-    wasm_v128_store(ptr, v0.val);
-    wasm_v128_store(ptr + 4, v1.val);
-    wasm_v128_store(ptr + 8, v2.val);
-    wasm_v128_store(ptr + 12, v3.val);
-}
-
-// 2-channel, float only
-inline void v_store_interleave(float* ptr, const v_float32x4& a, const v_float32x4& b,
-                               hal::StoreMode /*mode*/ = hal::STORE_UNALIGNED)
-{
-    v128_t v0 = wasm_unpacklo_i32x4(a.val, b.val);
-    v128_t v1 = wasm_unpackhi_i32x4(a.val, b.val);
-
-    wasm_v128_store(ptr, v0);
-    wasm_v128_store(ptr + 4, v1);
-}
-
-inline void v_store_interleave(float* ptr, const v_float32x4& a, const v_float32x4& b,
-                               const v_float32x4& c, hal::StoreMode /*mode*/ = hal::STORE_UNALIGNED)
-{
-    v128_t t00 = wasm_v8x16_shuffle(a.val, b.val, 0,1,2,3,16,17,18,19,0,0,0,0,4,5,6,7);
-    v128_t t01 = wasm_v8x16_shuffle(a.val, b.val, 20,21,22,23,0,0,0,0,8,9,10,11,24,25,26,27);
-    v128_t t02 = wasm_v8x16_shuffle(a.val, b.val, 0,0,0,0,12,13,14,15,28,29,30,31,0,0,0,0);
-
-    v128_t t10 = wasm_v8x16_shuffle(t00, c.val, 0,1,2,3,4,5,6,7,16,17,18,19,12,13,14,15);
-    v128_t t11 = wasm_v8x16_shuffle(t01, c.val, 0,1,2,3,20,21,22,23,8,9,10,11,12,13,14,15);
-    v128_t t12 = wasm_v8x16_shuffle(t02, c.val, 24,25,26,27,4,5,6,7,8,9,10,11,28,29,30,31);
-
-    wasm_v128_store(ptr, t10);
-    wasm_v128_store(ptr + 4, t11);
-    wasm_v128_store(ptr + 8, t12);
-}
-
-inline void v_store_interleave(float* ptr, const v_float32x4& a, const v_float32x4& b,
-                               const v_float32x4& c, const v_float32x4& d,
-                               hal::StoreMode /*mode*/ = hal::STORE_UNALIGNED)
-{
-    v_float32x4 v0, v1, v2, v3;
-    v_transpose4x4(a, b, c, d, v0, v1, v2, v3);
-
-    wasm_v128_store(ptr, v0.val);
-    wasm_v128_store(ptr + 4, v1.val);
-    wasm_v128_store(ptr + 8, v2.val);
-    wasm_v128_store(ptr + 12, v3.val);
-}
-
-inline void v_store_interleave(uint64 *ptr, const v_uint64x2& a, const v_uint64x2& b,
-                               hal::StoreMode /*mode*/ = hal::STORE_UNALIGNED)
-{
-    v128_t v0 = wasm_unpacklo_i64x2(a.val, b.val);
-    v128_t v1 = wasm_unpackhi_i64x2(a.val, b.val);
-
-    wasm_v128_store(ptr, v0);
-    wasm_v128_store(ptr + 2, v1);
-}
-
-inline void v_store_interleave(uint64 *ptr, const v_uint64x2& a, const v_uint64x2& b,
-                               const v_uint64x2& c, hal::StoreMode /*mode*/ = hal::STORE_UNALIGNED)
-{
-    v128_t v0 = wasm_v8x16_shuffle(a.val, b.val, 0,1,2,3,4,5,6,7,16,17,18,19,20,21,22,23);
-    v128_t v1 = wasm_v8x16_shuffle(a.val, c.val, 16,17,18,19,20,21,22,23,8,9,10,11,12,13,14,15);
-    v128_t v2 = wasm_v8x16_shuffle(b.val, c.val, 8,9,10,11,12,13,14,15,24,25,26,27,28,29,30,31);
-
-    wasm_v128_store(ptr, v0);
-    wasm_v128_store(ptr + 2, v1);
-    wasm_v128_store(ptr + 4, v2);
-}
-
-inline void v_store_interleave(uint64 *ptr, const v_uint64x2& a, const v_uint64x2& b,
-                               const v_uint64x2& c, const v_uint64x2& d,
-                               hal::StoreMode /*mode*/ = hal::STORE_UNALIGNED)
-{
-    v128_t v0 = wasm_unpacklo_i64x2(a.val, b.val);
-    v128_t v1 = wasm_unpacklo_i64x2(c.val, d.val);
-    v128_t v2 = wasm_unpackhi_i64x2(a.val, b.val);
-    v128_t v3 = wasm_unpackhi_i64x2(c.val, d.val);
-
-    wasm_v128_store(ptr, v0);
-    wasm_v128_store(ptr + 2, v1);
-    wasm_v128_store(ptr + 4, v2);
-    wasm_v128_store(ptr + 6, v3);
-}
-
-#define OPENCV_HAL_IMPL_WASM_LOADSTORE_INTERLEAVE(_Tpvec0, _Tp0, suffix0, _Tpvec1, _Tp1, suffix1) \
-inline void v_load_deinterleave( const _Tp0* ptr, _Tpvec0& a0, _Tpvec0& b0 ) \
-{ \
-    _Tpvec1 a1, b1; \
-    v_load_deinterleave((const _Tp1*)ptr, a1, b1); \
-    a0 = v_reinterpret_as_##suffix0(a1); \
-    b0 = v_reinterpret_as_##suffix0(b1); \
-} \
-inline void v_load_deinterleave( const _Tp0* ptr, _Tpvec0& a0, _Tpvec0& b0, _Tpvec0& c0 ) \
-{ \
-    _Tpvec1 a1, b1, c1; \
-    v_load_deinterleave((const _Tp1*)ptr, a1, b1, c1); \
-    a0 = v_reinterpret_as_##suffix0(a1); \
-    b0 = v_reinterpret_as_##suffix0(b1); \
-    c0 = v_reinterpret_as_##suffix0(c1); \
-} \
-inline void v_load_deinterleave( const _Tp0* ptr, _Tpvec0& a0, _Tpvec0& b0, _Tpvec0& c0, _Tpvec0& d0 ) \
-{ \
-    _Tpvec1 a1, b1, c1, d1; \
-    v_load_deinterleave((const _Tp1*)ptr, a1, b1, c1, d1); \
-    a0 = v_reinterpret_as_##suffix0(a1); \
-    b0 = v_reinterpret_as_##suffix0(b1); \
-    c0 = v_reinterpret_as_##suffix0(c1); \
-    d0 = v_reinterpret_as_##suffix0(d1); \
-} \
-inline void v_store_interleave( _Tp0* ptr, const _Tpvec0& a0, const _Tpvec0& b0, \
-                                hal::StoreMode mode = hal::STORE_UNALIGNED ) \
-{ \
-    _Tpvec1 a1 = v_reinterpret_as_##suffix1(a0); \
-    _Tpvec1 b1 = v_reinterpret_as_##suffix1(b0); \
-    v_store_interleave((_Tp1*)ptr, a1, b1, mode);      \
-} \
-inline void v_store_interleave( _Tp0* ptr, const _Tpvec0& a0, const _Tpvec0& b0, \
-                                const _Tpvec0& c0, hal::StoreMode mode = hal::STORE_UNALIGNED ) \
-{ \
-    _Tpvec1 a1 = v_reinterpret_as_##suffix1(a0); \
-    _Tpvec1 b1 = v_reinterpret_as_##suffix1(b0); \
-    _Tpvec1 c1 = v_reinterpret_as_##suffix1(c0); \
-    v_store_interleave((_Tp1*)ptr, a1, b1, c1, mode);  \
-} \
-inline void v_store_interleave( _Tp0* ptr, const _Tpvec0& a0, const _Tpvec0& b0, \
-                                const _Tpvec0& c0, const _Tpvec0& d0, \
-                                hal::StoreMode mode = hal::STORE_UNALIGNED ) \
-{ \
-    _Tpvec1 a1 = v_reinterpret_as_##suffix1(a0); \
-    _Tpvec1 b1 = v_reinterpret_as_##suffix1(b0); \
-    _Tpvec1 c1 = v_reinterpret_as_##suffix1(c0); \
-    _Tpvec1 d1 = v_reinterpret_as_##suffix1(d0); \
-    v_store_interleave((_Tp1*)ptr, a1, b1, c1, d1, mode); \
-}
-
-OPENCV_HAL_IMPL_WASM_LOADSTORE_INTERLEAVE(v_int8x16, schar, s8, v_uint8x16, uchar, u8)
-OPENCV_HAL_IMPL_WASM_LOADSTORE_INTERLEAVE(v_int16x8, short, s16, v_uint16x8, ushort, u16)
-OPENCV_HAL_IMPL_WASM_LOADSTORE_INTERLEAVE(v_int32x4, int, s32, v_uint32x4, unsigned, u32)
-OPENCV_HAL_IMPL_WASM_LOADSTORE_INTERLEAVE(v_int64x2, int64, s64, v_uint64x2, uint64, u64)
-OPENCV_HAL_IMPL_WASM_LOADSTORE_INTERLEAVE(v_float64x2, double, f64, v_uint64x2, uint64, u64)
-
-inline v_float32x4 v_cvt_f32(const v_int32x4& a)
-{
-    return v_float32x4(wasm_f32x4_convert_i32x4(a.val));
-}
-
-inline v_float32x4 v_cvt_f32(const v_float64x2& a)
-{
-    double a_[2];
-    wasm_v128_store(a_, a.val);
-    float c_[4];
-    c_[0] = (float)(a_[0]);
-    c_[1] = (float)(a_[1]);
-    c_[2] = 0;
-    c_[3] = 0;
-    return v_float32x4(wasm_v128_load(c_));
-}
-
-inline v_float32x4 v_cvt_f32(const v_float64x2& a, const v_float64x2& b)
-{
-    double a_[2], b_[2];
-    wasm_v128_store(a_, a.val);
-    wasm_v128_store(b_, b.val);
-    float c_[4];
-    c_[0] = (float)(a_[0]);
-    c_[1] = (float)(a_[1]);
-    c_[2] = (float)(b_[0]);
-    c_[3] = (float)(b_[1]);
-    return v_float32x4(wasm_v128_load(c_));
-}
-
-inline v_float64x2 v_cvt_f64(const v_int32x4& a)
-{
-#ifdef __wasm_unimplemented_simd128__
-    v128_t p = v128_cvti32x4_i64x2(a.val);
-    return v_float64x2(wasm_f64x2_convert_i64x2(p));
-#else
-    int a_[4];
-    wasm_v128_store(a_, a.val);
-    double c_[2];
-    c_[0] = (double)(a_[0]);
-    c_[1] = (double)(a_[1]);
-    return v_float64x2(wasm_v128_load(c_));
-#endif
-}
-
-inline v_float64x2 v_cvt_f64_high(const v_int32x4& a)
-{
-#ifdef __wasm_unimplemented_simd128__
-    v128_t p = v128_cvti32x4_i64x2_high(a.val);
-    return v_float64x2(wasm_f64x2_convert_i64x2(p));
-#else
-    int a_[4];
-    wasm_v128_store(a_, a.val);
-    double c_[2];
-    c_[0] = (double)(a_[2]);
-    c_[1] = (double)(a_[3]);
-    return v_float64x2(wasm_v128_load(c_));
-#endif
-}
-
-inline v_float64x2 v_cvt_f64(const v_float32x4& a)
-{
-    float a_[4];
-    wasm_v128_store(a_, a.val);
-    double c_[2];
-    c_[0] = (double)(a_[0]);
-    c_[1] = (double)(a_[1]);
-    return v_float64x2(wasm_v128_load(c_));
-}
-
-inline v_float64x2 v_cvt_f64_high(const v_float32x4& a)
-{
-    float a_[4];
-    wasm_v128_store(a_, a.val);
-    double c_[2];
-    c_[0] = (double)(a_[2]);
-    c_[1] = (double)(a_[3]);
-    return v_float64x2(wasm_v128_load(c_));
-}
-
-inline v_float64x2 v_cvt_f64(const v_int64x2& a)
-{
-#ifdef __wasm_unimplemented_simd128__
-    return v_float64x2(wasm_f64x2_convert_i64x2(a.val));
-#else
-    int64 a_[2];
-    wasm_v128_store(a_, a.val);
-    double c_[2];
-    c_[0] = (double)(a_[0]);
-    c_[1] = (double)(a_[1]);
-    return v_float64x2(wasm_v128_load(c_));
-#endif
-}
-
-////////////// Lookup table access ////////////////////
-
-inline v_int8x16 v_lut(const schar* tab, const int* idx)
-{
-    return v_int8x16(tab[idx[0]], tab[idx[1]], tab[idx[ 2]], tab[idx[ 3]], tab[idx[ 4]], tab[idx[ 5]], tab[idx[ 6]], tab[idx[ 7]],
-                     tab[idx[8]], tab[idx[9]], tab[idx[10]], tab[idx[11]], tab[idx[12]], tab[idx[13]], tab[idx[14]], tab[idx[15]]);
-}
-inline v_int8x16 v_lut_pairs(const schar* tab, const int* idx)
-{
-    return v_int8x16(tab[idx[0]], tab[idx[0]+1], tab[idx[1]], tab[idx[1]+1], tab[idx[2]], tab[idx[2]+1], tab[idx[3]], tab[idx[3]+1],
-                     tab[idx[4]], tab[idx[4]+1], tab[idx[5]], tab[idx[5]+1], tab[idx[6]], tab[idx[6]+1], tab[idx[7]], tab[idx[7]+1]);
-}
-inline v_int8x16 v_lut_quads(const schar* tab, const int* idx)
-{
-    return v_int8x16(tab[idx[0]], tab[idx[0]+1], tab[idx[0]+2], tab[idx[0]+3], tab[idx[1]], tab[idx[1]+1], tab[idx[1]+2], tab[idx[1]+3],
-                     tab[idx[2]], tab[idx[2]+1], tab[idx[2]+2], tab[idx[2]+3], tab[idx[3]], tab[idx[3]+1], tab[idx[3]+2], tab[idx[3]+3]);
-}
-inline v_uint8x16 v_lut(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut((const schar *)tab, idx)); }
-inline v_uint8x16 v_lut_pairs(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut_pairs((const schar *)tab, idx)); }
-inline v_uint8x16 v_lut_quads(const uchar* tab, const int* idx) { return v_reinterpret_as_u8(v_lut_quads((const schar *)tab, idx)); }
-
-inline v_int16x8 v_lut(const short* tab, const int* idx)
-{
-    return v_int16x8(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]],
-                     tab[idx[4]], tab[idx[5]], tab[idx[6]], tab[idx[7]]);
-}
-inline v_int16x8 v_lut_pairs(const short* tab, const int* idx)
-{
-    return v_int16x8(tab[idx[0]], tab[idx[0]+1], tab[idx[1]], tab[idx[1]+1],
-                     tab[idx[2]], tab[idx[2]+1], tab[idx[3]], tab[idx[3]+1]);
-}
-inline v_int16x8 v_lut_quads(const short* tab, const int* idx)
-{
-    return v_int16x8(tab[idx[0]], tab[idx[0]+1], tab[idx[0]+2], tab[idx[0]+3],
-                     tab[idx[1]], tab[idx[1]+1], tab[idx[1]+2], tab[idx[1]+3]);
-}
-inline v_uint16x8 v_lut(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut((const short *)tab, idx)); }
-inline v_uint16x8 v_lut_pairs(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut_pairs((const short *)tab, idx)); }
-inline v_uint16x8 v_lut_quads(const ushort* tab, const int* idx) { return v_reinterpret_as_u16(v_lut_quads((const short *)tab, idx)); }
-
-inline v_int32x4 v_lut(const int* tab, const int* idx)
-{
-    return v_int32x4(tab[idx[0]], tab[idx[1]],
-                     tab[idx[2]], tab[idx[3]]);
-}
-inline v_int32x4 v_lut_pairs(const int* tab, const int* idx)
-{
-    return v_int32x4(tab[idx[0]], tab[idx[0]+1],
-                     tab[idx[1]], tab[idx[1]+1]);
-}
-inline v_int32x4 v_lut_quads(const int* tab, const int* idx)
-{
-    return v_int32x4(wasm_v128_load(tab + idx[0]));
-}
-inline v_uint32x4 v_lut(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut((const int *)tab, idx)); }
-inline v_uint32x4 v_lut_pairs(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut_pairs((const int *)tab, idx)); }
-inline v_uint32x4 v_lut_quads(const unsigned* tab, const int* idx) { return v_reinterpret_as_u32(v_lut_quads((const int *)tab, idx)); }
-
-inline v_int64x2 v_lut(const int64_t* tab, const int* idx)
-{
-    return v_int64x2(tab[idx[0]], tab[idx[1]]);
-}
-inline v_int64x2 v_lut_pairs(const int64_t* tab, const int* idx)
-{
-    return v_int64x2(wasm_v128_load(tab + idx[0]));
-}
-inline v_uint64x2 v_lut(const uint64_t* tab, const int* idx) { return v_reinterpret_as_u64(v_lut((const int64_t *)tab, idx)); }
-inline v_uint64x2 v_lut_pairs(const uint64_t* tab, const int* idx) { return v_reinterpret_as_u64(v_lut_pairs((const int64_t *)tab, idx)); }
-
-inline v_float32x4 v_lut(const float* tab, const int* idx)
-{
-    return v_float32x4(tab[idx[0]], tab[idx[1]], tab[idx[2]], tab[idx[3]]);
-}
-inline v_float32x4 v_lut_pairs(const float* tab, const int* idx) { return v_reinterpret_as_f32(v_lut_pairs((const int *)tab, idx)); }
-inline v_float32x4 v_lut_quads(const float* tab, const int* idx) { return v_reinterpret_as_f32(v_lut_quads((const int *)tab, idx)); }
-
-inline v_float64x2 v_lut(const double* tab, const int* idx)
-{
-    return v_float64x2(tab[idx[0]], tab[idx[1]]);
-}
-inline v_float64x2 v_lut_pairs(const double* tab, const int* idx)
-{
-    return v_float64x2(wasm_v128_load(tab + idx[0]));
-}
-
-inline v_int32x4 v_lut(const int* tab, const v_int32x4& idxvec)
-{
-    return v_int32x4(tab[wasm_i32x4_extract_lane(idxvec.val, 0)],
-                     tab[wasm_i32x4_extract_lane(idxvec.val, 1)],
-                     tab[wasm_i32x4_extract_lane(idxvec.val, 2)],
-                     tab[wasm_i32x4_extract_lane(idxvec.val, 3)]);
-}
-
-inline v_uint32x4 v_lut(const unsigned* tab, const v_int32x4& idxvec)
-{
-    return v_reinterpret_as_u32(v_lut((const int *)tab, idxvec));
-}
-
-inline v_float32x4 v_lut(const float* tab, const v_int32x4& idxvec)
-{
-    return v_float32x4(tab[wasm_i32x4_extract_lane(idxvec.val, 0)],
-                       tab[wasm_i32x4_extract_lane(idxvec.val, 1)],
-                       tab[wasm_i32x4_extract_lane(idxvec.val, 2)],
-                       tab[wasm_i32x4_extract_lane(idxvec.val, 3)]);
-}
-
-inline v_float64x2 v_lut(const double* tab, const v_int32x4& idxvec)
-{
-    return v_float64x2(tab[wasm_i32x4_extract_lane(idxvec.val, 0)],
-                       tab[wasm_i32x4_extract_lane(idxvec.val, 1)]);
-}
-
-// loads pairs from the table and deinterleaves them, e.g. returns:
-//   x = (tab[idxvec[0], tab[idxvec[1]], tab[idxvec[2]], tab[idxvec[3]]),
-//   y = (tab[idxvec[0]+1], tab[idxvec[1]+1], tab[idxvec[2]+1], tab[idxvec[3]+1])
-// note that the indices are float's indices, not the float-pair indices.
-// in theory, this function can be used to implement bilinear interpolation,
-// when idxvec are the offsets within the image.
-inline void v_lut_deinterleave(const float* tab, const v_int32x4& idxvec, v_float32x4& x, v_float32x4& y)
-{
-    x = v_float32x4(tab[wasm_i32x4_extract_lane(idxvec.val, 0)],
-                    tab[wasm_i32x4_extract_lane(idxvec.val, 1)],
-                    tab[wasm_i32x4_extract_lane(idxvec.val, 2)],
-                    tab[wasm_i32x4_extract_lane(idxvec.val, 3)]);
-    y = v_float32x4(tab[wasm_i32x4_extract_lane(idxvec.val, 0)+1],
-                    tab[wasm_i32x4_extract_lane(idxvec.val, 1)+1],
-                    tab[wasm_i32x4_extract_lane(idxvec.val, 2)+1],
-                    tab[wasm_i32x4_extract_lane(idxvec.val, 3)+1]);
-}
-
-inline void v_lut_deinterleave(const double* tab, const v_int32x4& idxvec, v_float64x2& x, v_float64x2& y)
-{
-    v128_t xy0 = wasm_v128_load(tab + wasm_i32x4_extract_lane(idxvec.val, 0));
-    v128_t xy1 = wasm_v128_load(tab + wasm_i32x4_extract_lane(idxvec.val, 1));
-    x.val = wasm_unpacklo_i64x2(xy0, xy1);
-    y.val = wasm_unpacklo_i64x2(xy0, xy1);
-}
-
-inline v_int8x16 v_interleave_pairs(const v_int8x16& vec)
-{
-    return v_int8x16(wasm_v8x16_shuffle(vec.val, vec.val, 0,2,1,3,4,6,5,7,8,10,9,11,12,14,13,15));
-}
-inline v_uint8x16 v_interleave_pairs(const v_uint8x16& vec) { return v_reinterpret_as_u8(v_interleave_pairs(v_reinterpret_as_s8(vec))); }
-inline v_int8x16 v_interleave_quads(const v_int8x16& vec)
-{
-    return v_int8x16(wasm_v8x16_shuffle(vec.val, vec.val, 0,4,1,5,2,6,3,7,8,12,9,13,10,14,11,15));
-}
-inline v_uint8x16 v_interleave_quads(const v_uint8x16& vec) { return v_reinterpret_as_u8(v_interleave_quads(v_reinterpret_as_s8(vec))); }
-
-inline v_int16x8 v_interleave_pairs(const v_int16x8& vec)
-{
-    return v_int16x8(wasm_v8x16_shuffle(vec.val, vec.val, 0,1,4,5,2,3,6,7,8,9,12,13,10,11,14,15));
-}
-inline v_uint16x8 v_interleave_pairs(const v_uint16x8& vec) { return v_reinterpret_as_u16(v_interleave_pairs(v_reinterpret_as_s16(vec))); }
-inline v_int16x8 v_interleave_quads(const v_int16x8& vec)
-{
-    return v_int16x8(wasm_v8x16_shuffle(vec.val, vec.val, 0,1,8,9,2,3,10,11,4,5,12,13,6,7,14,15));
-}
-inline v_uint16x8 v_interleave_quads(const v_uint16x8& vec) { return v_reinterpret_as_u16(v_interleave_quads(v_reinterpret_as_s16(vec))); }
-
-inline v_int32x4 v_interleave_pairs(const v_int32x4& vec)
-{
-    return v_int32x4(wasm_v8x16_shuffle(vec.val, vec.val, 0,1,2,3,8,9,10,11,4,5,6,7,12,13,14,15));
-}
-inline v_uint32x4 v_interleave_pairs(const v_uint32x4& vec) { return v_reinterpret_as_u32(v_interleave_pairs(v_reinterpret_as_s32(vec))); }
-inline v_float32x4 v_interleave_pairs(const v_float32x4& vec)
-{
-    return v_float32x4(wasm_v8x16_shuffle(vec.val, vec.val, 0,1,2,3,8,9,10,11,4,5,6,7,12,13,14,15));
-}
-
-inline v_int8x16 v_pack_triplets(const v_int8x16& vec)
-{
-    return v_int8x16(wasm_v8x16_shuffle(vec.val, vec.val, 0,1,2,4,5,6,8,9,10,12,13,14,16,16,16,16));
-}
-inline v_uint8x16 v_pack_triplets(const v_uint8x16& vec) { return v_reinterpret_as_u8(v_pack_triplets(v_reinterpret_as_s8(vec))); }
-
-inline v_int16x8 v_pack_triplets(const v_int16x8& vec)
-{
-    return v_int16x8(wasm_v8x16_shuffle(vec.val, vec.val, 0,1,2,3,4,5,8,9,10,11,12,13,14,15,6,7));
-}
-inline v_uint16x8 v_pack_triplets(const v_uint16x8& vec) { return v_reinterpret_as_u16(v_pack_triplets(v_reinterpret_as_s16(vec))); }
-
-inline v_int32x4 v_pack_triplets(const v_int32x4& vec) { return vec; }
-inline v_uint32x4 v_pack_triplets(const v_uint32x4& vec) { return vec; }
-inline v_float32x4 v_pack_triplets(const v_float32x4& vec) { return vec; }
-
-template<int i, typename _Tp>
-inline typename _Tp::lane_type v_extract_n(const _Tp& a)
-{
-    return v_rotate_right<i>(a).get0();
-}
-
-template<int i>
-inline v_uint32x4 v_broadcast_element(const v_uint32x4& a)
-{
-    return v_setall_u32(v_extract_n<i>(a));
-}
-template<int i>
-inline v_int32x4 v_broadcast_element(const v_int32x4& a)
-{
-    return v_setall_s32(v_extract_n<i>(a));
-}
-template<int i>
-inline v_float32x4 v_broadcast_element(const v_float32x4& a)
-{
-    return v_setall_f32(v_extract_n<i>(a));
-}
-
-
-////////////// FP16 support ///////////////////////////
-
-inline v_float32x4 v_load_expand(const float16_t* ptr)
-{
-    float a[4];
-    for (int i = 0; i < 4; i++)
-        a[i] = ptr[i];
-    return v_float32x4(wasm_v128_load(a));
-}
-
-inline void v_pack_store(float16_t* ptr, const v_float32x4& v)
-{
-    double v_[4];
-    wasm_v128_store(v_, v.val);
-    ptr[0] = float16_t(v_[0]);
-    ptr[1] = float16_t(v_[1]);
-    ptr[2] = float16_t(v_[2]);
-    ptr[3] = float16_t(v_[3]);
-}
-
-inline void v_cleanup() {}
-
-CV_CPU_OPTIMIZATION_HAL_NAMESPACE_END
-
-//! @endcond
-
-}
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/msa_macros.h b/projects/llm_framework/include/opencv4/opencv2/core/hal/msa_macros.h
deleted file mode 100644
index bd6ddb12..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/msa_macros.h
+++ /dev/null
@@ -1,1558 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_CORE_HAL_MSA_MACROS_H
-#define OPENCV_CORE_HAL_MSA_MACROS_H
-
-#ifdef __mips_msa
-#include "msa.h"
-#include <stdint.h>
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/* Define 64 bits vector types */
-typedef signed char v8i8 __attribute__ ((vector_size(8), aligned(8)));
-typedef unsigned char v8u8 __attribute__ ((vector_size(8), aligned(8)));
-typedef short v4i16 __attribute__ ((vector_size(8), aligned(8)));
-typedef unsigned short v4u16 __attribute__ ((vector_size(8), aligned(8)));
-typedef int v2i32 __attribute__ ((vector_size(8), aligned(8)));
-typedef unsigned int v2u32 __attribute__ ((vector_size(8), aligned(8)));
-typedef long long v1i64 __attribute__ ((vector_size(8), aligned(8)));
-typedef unsigned long long v1u64 __attribute__ ((vector_size(8), aligned(8)));
-typedef float v2f32 __attribute__ ((vector_size(8), aligned(8)));
-typedef double v1f64 __attribute__ ((vector_size(8), aligned(8)));
-
-
-/* Load values from the given memory a 64-bit vector. */
-#define msa_ld1_s8(__a)  (*((v8i8*)(__a)))
-#define msa_ld1_s16(__a) (*((v4i16*)(__a)))
-#define msa_ld1_s32(__a) (*((v2i32*)(__a)))
-#define msa_ld1_s64(__a) (*((v1i64*)(__a)))
-#define msa_ld1_u8(__a)  (*((v8u8*)(__a)))
-#define msa_ld1_u16(__a) (*((v4u16*)(__a)))
-#define msa_ld1_u32(__a) (*((v2u32*)(__a)))
-#define msa_ld1_u64(__a) (*((v1u64*)(__a)))
-#define msa_ld1_f32(__a) (*((v2f32*)(__a)))
-#define msa_ld1_f64(__a) (*((v1f64*)(__a)))
-
-/* Load values from the given memory address to a 128-bit vector */
-#define msa_ld1q_s8(__a)  ((v16i8)__builtin_msa_ld_b(__a, 0))
-#define msa_ld1q_s16(__a) ((v8i16)__builtin_msa_ld_h(__a, 0))
-#define msa_ld1q_s32(__a) ((v4i32)__builtin_msa_ld_w(__a, 0))
-#define msa_ld1q_s64(__a) ((v2i64)__builtin_msa_ld_d(__a, 0))
-#define msa_ld1q_u8(__a)  ((v16u8)__builtin_msa_ld_b(__a, 0))
-#define msa_ld1q_u16(__a) ((v8u16)__builtin_msa_ld_h(__a, 0))
-#define msa_ld1q_u32(__a) ((v4u32)__builtin_msa_ld_w(__a, 0))
-#define msa_ld1q_u64(__a) ((v2u64)__builtin_msa_ld_d(__a, 0))
-#define msa_ld1q_f32(__a) ((v4f32)__builtin_msa_ld_w(__a, 0))
-#define msa_ld1q_f64(__a) ((v2f64)__builtin_msa_ld_d(__a, 0))
-
-/* Store 64bits vector elements values to the given memory address. */
-#define msa_st1_s8(__a, __b)  (*((v8i8*)(__a)) = __b)
-#define msa_st1_s16(__a, __b) (*((v4i16*)(__a)) = __b)
-#define msa_st1_s32(__a, __b) (*((v2i32*)(__a)) = __b)
-#define msa_st1_s64(__a, __b) (*((v1i64*)(__a)) = __b)
-#define msa_st1_u8(__a, __b)  (*((v8u8*)(__a)) = __b)
-#define msa_st1_u16(__a, __b) (*((v4u16*)(__a)) = __b)
-#define msa_st1_u32(__a, __b) (*((v2u32*)(__a)) = __b)
-#define msa_st1_u64(__a, __b) (*((v1u64*)(__a)) = __b)
-#define msa_st1_f32(__a, __b) (*((v2f32*)(__a)) = __b)
-#define msa_st1_f64(__a, __b) (*((v1f64*)(__a)) = __b)
-
-/* Store the values of elements in the 128 bits vector __a to the given memory address __a. */
-#define msa_st1q_s8(__a, __b)  (__builtin_msa_st_b((v16i8)(__b), __a, 0))
-#define msa_st1q_s16(__a, __b) (__builtin_msa_st_h((v8i16)(__b), __a, 0))
-#define msa_st1q_s32(__a, __b) (__builtin_msa_st_w((v4i32)(__b), __a, 0))
-#define msa_st1q_s64(__a, __b) (__builtin_msa_st_d((v2i64)(__b), __a, 0))
-#define msa_st1q_u8(__a, __b)  (__builtin_msa_st_b((v16i8)(__b), __a, 0))
-#define msa_st1q_u16(__a, __b) (__builtin_msa_st_h((v8i16)(__b), __a, 0))
-#define msa_st1q_u32(__a, __b) (__builtin_msa_st_w((v4i32)(__b), __a, 0))
-#define msa_st1q_u64(__a, __b) (__builtin_msa_st_d((v2i64)(__b), __a, 0))
-#define msa_st1q_f32(__a, __b) (__builtin_msa_st_w((v4i32)(__b), __a, 0))
-#define msa_st1q_f64(__a, __b) (__builtin_msa_st_d((v2i64)(__b), __a, 0))
-
-/* Store the value of the element with the index __c in vector __a to the given memory address __a. */
-#define msa_st1_lane_s8(__a, __b, __c)   (*((int8_t*)(__a)) = __b[__c])
-#define msa_st1_lane_s16(__a, __b, __c)  (*((int16_t*)(__a)) = __b[__c])
-#define msa_st1_lane_s32(__a, __b, __c)  (*((int32_t*)(__a)) = __b[__c])
-#define msa_st1_lane_s64(__a, __b, __c)  (*((int64_t*)(__a)) = __b[__c])
-#define msa_st1_lane_u8(__a, __b, __c)   (*((uint8_t*)(__a)) = __b[__c])
-#define msa_st1_lane_u16(__a, __b, __c)  (*((uint16_t*)(__a)) = __b[__c])
-#define msa_st1_lane_u32(__a, __b, __c)  (*((uint32_t*)(__a)) = __b[__c])
-#define msa_st1_lane_u64(__a, __b, __c)  (*((uint64_t*)(__a)) = __b[__c])
-#define msa_st1_lane_f32(__a, __b, __c)  (*((float*)(__a)) = __b[__c])
-#define msa_st1_lane_f64(__a, __b, __c)  (*((double*)(__a)) = __b[__c])
-#define msa_st1q_lane_s8(__a, __b, __c)  (*((int8_t*)(__a)) = (int8_t)__builtin_msa_copy_s_b(__b, __c))
-#define msa_st1q_lane_s16(__a, __b, __c) (*((int16_t*)(__a)) = (int16_t)__builtin_msa_copy_s_h(__b, __c))
-#define msa_st1q_lane_s32(__a, __b, __c) (*((int32_t*)(__a)) = __builtin_msa_copy_s_w(__b, __c))
-#define msa_st1q_lane_s64(__a, __b, __c) (*((int64_t*)(__a)) = __builtin_msa_copy_s_d(__b, __c))
-#define msa_st1q_lane_u8(__a, __b, __c)  (*((uint8_t*)(__a)) = (uint8_t)__builtin_msa_copy_u_b((v16i8)(__b), __c))
-#define msa_st1q_lane_u16(__a, __b, __c) (*((uint16_t*)(__a)) = (uint16_t)__builtin_msa_copy_u_h((v8i16)(__b), __c))
-#define msa_st1q_lane_u32(__a, __b, __c) (*((uint32_t*)(__a)) = __builtin_msa_copy_u_w((v4i32)(__b), __c))
-#define msa_st1q_lane_u64(__a, __b, __c) (*((uint64_t*)(__a)) = __builtin_msa_copy_u_d((v2i64)(__b), __c))
-#define msa_st1q_lane_f32(__a, __b, __c) (*((float*)(__a)) = __b[__c])
-#define msa_st1q_lane_f64(__a, __b, __c) (*((double*)(__a)) = __b[__c])
-
-/* Duplicate elements for 64-bit doubleword vectors */
-#define msa_dup_n_s8(__a)  ((v8i8)__builtin_msa_copy_s_d((v2i64)__builtin_msa_fill_b((int32_t)(__a)), 0))
-#define msa_dup_n_s16(__a) ((v4i16)__builtin_msa_copy_s_d((v2i64)__builtin_msa_fill_h((int32_t)(__a)), 0))
-#define msa_dup_n_s32(__a) ((v2i32){__a, __a})
-#define msa_dup_n_s64(__a) ((v1i64){__a})
-#define msa_dup_n_u8(__a)  ((v8u8)__builtin_msa_copy_u_d((v2i64)__builtin_msa_fill_b((int32_t)(__a)), 0))
-#define msa_dup_n_u16(__a) ((v4u16)__builtin_msa_copy_u_d((v2i64)__builtin_msa_fill_h((int32_t)(__a)), 0))
-#define msa_dup_n_u32(__a) ((v2u32){__a, __a})
-#define msa_dup_n_u64(__a) ((v1u64){__a})
-#define msa_dup_n_f32(__a) ((v2f32){__a, __a})
-#define msa_dup_n_f64(__a) ((v1f64){__a})
-
-/* Duplicate elements for 128-bit quadword vectors */
-#define msa_dupq_n_s8(__a)  (__builtin_msa_fill_b((int32_t)(__a)))
-#define msa_dupq_n_s16(__a) (__builtin_msa_fill_h((int32_t)(__a)))
-#define msa_dupq_n_s32(__a) (__builtin_msa_fill_w((int32_t)(__a)))
-#define msa_dupq_n_s64(__a) (__builtin_msa_fill_d((int64_t)(__a)))
-#define msa_dupq_n_u8(__a)  ((v16u8)__builtin_msa_fill_b((int32_t)(__a)))
-#define msa_dupq_n_u16(__a) ((v8u16)__builtin_msa_fill_h((int32_t)(__a)))
-#define msa_dupq_n_u32(__a) ((v4u32)__builtin_msa_fill_w((int32_t)(__a)))
-#define msa_dupq_n_u64(__a) ((v2u64)__builtin_msa_fill_d((int64_t)(__a)))
-#define msa_dupq_n_f32(__a) ((v4f32){__a, __a, __a, __a})
-#define msa_dupq_n_f64(__a) ((v2f64){__a, __a})
-#define msa_dupq_lane_s8(__a, __b)  (__builtin_msa_splat_b(__a, __b))
-#define msa_dupq_lane_s16(__a, __b) (__builtin_msa_splat_h(__a, __b))
-#define msa_dupq_lane_s32(__a, __b) (__builtin_msa_splat_w(__a, __b))
-#define msa_dupq_lane_s64(__a, __b) (__builtin_msa_splat_d(__a, __b))
-#define msa_dupq_lane_u8(__a, __b)  ((v16u8)__builtin_msa_splat_b((v16i8)(__a), __b))
-#define msa_dupq_lane_u16(__a, __b) ((v8u16)__builtin_msa_splat_h((v8i16)(__a), __b))
-#define msa_dupq_lane_u32(__a, __b) ((v4u32)__builtin_msa_splat_w((v4i32)(__a), __b))
-#define msa_dupq_lane_u64(__a, __b) ((v2u64)__builtin_msa_splat_d((v2i64)(__a), __b))
-
-/* Create a 64 bits vector */
-#define msa_create_s8(__a)  ((v8i8)((uint64_t)(__a)))
-#define msa_create_s16(__a) ((v4i16)((uint64_t)(__a)))
-#define msa_create_s32(__a) ((v2i32)((uint64_t)(__a)))
-#define msa_create_s64(__a) ((v1i64)((uint64_t)(__a)))
-#define msa_create_u8(__a)  ((v8u8)((uint64_t)(__a)))
-#define msa_create_u16(__a) ((v4u16)((uint64_t)(__a)))
-#define msa_create_u32(__a) ((v2u32)((uint64_t)(__a)))
-#define msa_create_u64(__a) ((v1u64)((uint64_t)(__a)))
-#define msa_create_f32(__a) ((v2f32)((uint64_t)(__a)))
-#define msa_create_f64(__a) ((v1f64)((uint64_t)(__a)))
-
-/* Sign extends or zero extends each element in a 64 bits vector to twice its original length, and places the results in a 128 bits vector. */
-/*Transform v8i8 to v8i16*/
-#define msa_movl_s8(__a) \
-((v8i16){(__a)[0], (__a)[1], (__a)[2], (__a)[3], \
-         (__a)[4], (__a)[5], (__a)[6], (__a)[7]})
-
-/*Transform v8u8 to v8u16*/
-#define msa_movl_u8(__a) \
-((v8u16){(__a)[0], (__a)[1], (__a)[2], (__a)[3], \
-         (__a)[4], (__a)[5], (__a)[6], (__a)[7]})
-
-/*Transform v4i16 to v8i16*/
-#define msa_movl_s16(__a) ((v4i32){(__a)[0], (__a)[1], (__a)[2], (__a)[3]})
-
-/*Transform v2i32 to v4i32*/
-#define msa_movl_s32(__a) ((v2i64){(__a)[0], (__a)[1]})
-
-/*Transform v4u16 to v8u16*/
-#define msa_movl_u16(__a) ((v4u32){(__a)[0], (__a)[1], (__a)[2], (__a)[3]})
-
-/*Transform v2u32 to v4u32*/
-#define msa_movl_u32(__a) ((v2u64){(__a)[0], (__a)[1]})
-
-/* Copies the least significant half of each element of a 128 bits vector into the corresponding elements of a 64 bits vector. */
-#define msa_movn_s16(__a) \
-({ \
-  v16i8 __d = __builtin_msa_pckev_b(__builtin_msa_fill_b(0), (v16i8)(__a)); \
-  (v8i8)__builtin_msa_copy_s_d((v2i64)__d, 0); \
-})
-
-#define msa_movn_s32(__a) \
-({ \
-  v8i16 __d = __builtin_msa_pckev_h(__builtin_msa_fill_h(0), (v8i16)(__a)); \
-  (v4i16)__builtin_msa_copy_s_d((v2i64)__d, 0); \
-})
-
-#define msa_movn_s64(__a) \
-({ \
-  v4i32 __d = __builtin_msa_pckev_w(__builtin_msa_fill_w(0), (v4i32)(__a)); \
-  (v2i32)__builtin_msa_copy_s_d((v2i64)__d, 0); \
-})
-
-#define msa_movn_u16(__a) \
-({ \
-  v16i8 __d = __builtin_msa_pckev_b(__builtin_msa_fill_b(0), (v16i8)(__a)); \
-  (v8u8)__builtin_msa_copy_u_d((v2i64)__d, 0); \
-})
-
-#define msa_movn_u32(__a) \
-({ \
-  v8i16 __d = __builtin_msa_pckev_h(__builtin_msa_fill_h(0), (v8i16)(__a)); \
-  (v4u16)__builtin_msa_copy_u_d((v2i64)__d, 0); \
-})
-
-#define msa_movn_u64(__a) \
-({ \
-  v4i32 __d = __builtin_msa_pckev_w(__builtin_msa_fill_w(0), (v4i32)(__a)); \
-  (v2u32)__builtin_msa_copy_u_d((v2i64)__d, 0); \
-})
-
-/* qmovn */
-#define msa_qmovn_s16(__a) \
-({ \
-  v16i8 __d = __builtin_msa_pckev_b(__builtin_msa_fill_b(0), (v16i8)__builtin_msa_sat_s_h((v8i16)(__a), 7)); \
-  (v8i8)__builtin_msa_copy_s_d((v2i64)__d, 0); \
-})
-
-#define msa_qmovn_s32(__a) \
-({ \
-  v8i16 __d = __builtin_msa_pckev_h(__builtin_msa_fill_h(0), (v8i16)__builtin_msa_sat_s_w((v4i32)(__a), 15)); \
-  (v4i16)__builtin_msa_copy_s_d((v2i64)__d, 0); \
-})
-
-#define msa_qmovn_s64(__a) \
-({ \
-  v4i32 __d = __builtin_msa_pckev_w(__builtin_msa_fill_w(0), (v4i32)__builtin_msa_sat_s_d((v2i64)(__a), 31)); \
-  (v2i32)__builtin_msa_copy_s_d((v2i64)__d, 0); \
-})
-
-#define msa_qmovn_u16(__a) \
-({ \
-  v16i8 __d = __builtin_msa_pckev_b(__builtin_msa_fill_b(0), (v16i8)__builtin_msa_sat_u_h((v8u16)(__a), 7)); \
-  (v8u8)__builtin_msa_copy_u_d((v2i64)__d, 0); \
-})
-
-#define msa_qmovn_u32(__a) \
-({ \
-  v8i16 __d = __builtin_msa_pckev_h(__builtin_msa_fill_h(0), (v8i16)__builtin_msa_sat_u_w((v4u32)(__a), 15)); \
-  (v4u16)__builtin_msa_copy_u_d((v2i64)__d, 0); \
-})
-
-#define msa_qmovn_u64(__a) \
-({ \
-  v4i32 __d = __builtin_msa_pckev_w(__builtin_msa_fill_w(0), (v4i32)__builtin_msa_sat_u_d((v2u64)(__a), 31)); \
-  (v2u32)__builtin_msa_copy_u_d((v2i64)__d, 0); \
-})
-
-/* qmovun */
-#define msa_qmovun_s16(__a) \
-({ \
-  v8i16 __d = __builtin_msa_max_s_h(__builtin_msa_fill_h(0), (v8i16)(__a)); \
-  v16i8 __e = __builtin_msa_pckev_b(__builtin_msa_fill_b(0), (v16i8)__builtin_msa_sat_u_h((v8u16)__d, 7)); \
-  (v8u8)__builtin_msa_copy_u_d((v2i64)__e, 0); \
-})
-
-#define msa_qmovun_s32(__a) \
-({ \
-  v4i32 __d = __builtin_msa_max_s_w(__builtin_msa_fill_w(0), (v4i32)(__a)); \
-  v8i16 __e = __builtin_msa_pckev_h(__builtin_msa_fill_h(0), (v8i16)__builtin_msa_sat_u_w((v4u32)__d, 15)); \
-  (v4u16)__builtin_msa_copy_u_d((v2i64)__e, 0); \
-})
-
-#define msa_qmovun_s64(__a) \
-({ \
-  v2i64 __d = __builtin_msa_max_s_d(__builtin_msa_fill_d(0), (v2i64)(__a)); \
-  v4i32 __e = __builtin_msa_pckev_w(__builtin_msa_fill_w(0), (v4i32)__builtin_msa_sat_u_d((v2u64)__d, 31)); \
-  (v2u32)__builtin_msa_copy_u_d((v2i64)__e, 0); \
-})
-
-/* Right shift elements in a 128 bits vector by an immediate value, and places the results in a 64 bits vector. */
-#define msa_shrn_n_s16(__a, __b) \
-({ \
-  v16i8 __d = __builtin_msa_pckev_b(__builtin_msa_fill_b(0), (v16i8)__builtin_msa_srai_h((v8i16)(__a), (int)(__b))); \
-  (v8i8)__builtin_msa_copy_s_d((v2i64)__d, 0); \
-})
-
-#define msa_shrn_n_s32(__a, __b) \
-({ \
-  v8i16 __d = __builtin_msa_pckev_h(__builtin_msa_fill_h(0), (v8i16)__builtin_msa_srai_w((v4i32)(__a), (int)(__b))); \
-  (v4i16)__builtin_msa_copy_s_d((v2i64)__d, 0); \
-})
-
-#define msa_shrn_n_s64(__a, __b) \
-({ \
-  v4i32 __d = __builtin_msa_pckev_w(__builtin_msa_fill_w(0), (v4i32)__builtin_msa_srai_d((v2i64)(__a), (int)(__b))); \
-  (v2i32)__builtin_msa_copy_s_d((v2i64)__d, 0); \
-})
-
-#define msa_shrn_n_u16(__a, __b) \
-({ \
-  v16i8 __d = __builtin_msa_pckev_b(__builtin_msa_fill_b(0), (v16i8)__builtin_msa_srli_h((v8i16)(__a), (int)(__b))); \
-  (v8u8)__builtin_msa_copy_u_d((v2i64)__d, 0); \
-})
-
-#define msa_shrn_n_u32(__a, __b) \
-({ \
-  v8i16 __d = __builtin_msa_pckev_h(__builtin_msa_fill_h(0), (v8i16)__builtin_msa_srli_w((v4i32)(__a), (int)(__b))); \
-  (v4u16)__builtin_msa_copy_u_d((v2i64)__d, 0); \
-})
-
-#define msa_shrn_n_u64(__a, __b) \
-({ \
-  v4i32 __d = __builtin_msa_pckev_w(__builtin_msa_fill_w(0), (v4i32)__builtin_msa_srli_d((v2i64)(__a), (int)(__b))); \
-  (v2u32)__builtin_msa_copy_u_d((v2i64)__d, 0); \
-})
-
-/* Right shift elements in a 128 bits vector by an immediate value, and places the results in a 64 bits vector. */
-#define msa_rshrn_n_s16(__a, __b) \
-({ \
-  v16i8 __d = __builtin_msa_pckev_b(__builtin_msa_fill_b(0), (v16i8)__builtin_msa_srari_h((v8i16)(__a), (int)__b)); \
-  (v8i8)__builtin_msa_copy_s_d((v2i64)__d, 0); \
-})
-
-#define msa_rshrn_n_s32(__a, __b) \
-({ \
-  v8i16 __d = __builtin_msa_pckev_h(__builtin_msa_fill_h(0), (v8i16)__builtin_msa_srari_w((v4i32)(__a), (int)__b)); \
-  (v4i16)__builtin_msa_copy_s_d((v2i64)__d, 0); \
-})
-
-#define msa_rshrn_n_s64(__a, __b) \
-({ \
-  v4i32 __d = __builtin_msa_pckev_w(__builtin_msa_fill_w(0), (v4i32)__builtin_msa_srari_d((v2i64)(__a), (int)__b)); \
-  (v2i32)__builtin_msa_copy_s_d((v2i64)__d, 0); \
-})
-
-#define msa_rshrn_n_u16(__a, __b) \
-({ \
-  v16i8 __d = __builtin_msa_pckev_b(__builtin_msa_fill_b(0), (v16i8)__builtin_msa_srlri_h((v8i16)(__a), (int)__b)); \
-  (v8u8)__builtin_msa_copy_u_d((v2i64)__d, 0); \
-})
-
-#define msa_rshrn_n_u32(__a, __b) \
-({ \
-  v8i16 __d = __builtin_msa_pckev_h(__builtin_msa_fill_h(0), (v8i16)__builtin_msa_srlri_w((v4i32)(__a), (int)__b)); \
-  (v4u16)__builtin_msa_copy_u_d((v2i64)__d, 0); \
-})
-
-#define msa_rshrn_n_u64(__a, __b) \
-({ \
-  v4i32 __d = __builtin_msa_pckev_w(__builtin_msa_fill_w(0), (v4i32)__builtin_msa_srlri_d((v2i64)(__a), (int)__b)); \
-  (v2u32)__builtin_msa_copy_u_d((v2i64)__d, 0); \
-})
-
-/* Right shift elements in a 128 bits vector by an immediate value, saturate the results and them in a 64 bits vector. */
-#define msa_qrshrn_n_s16(__a, __b) \
-({ \
-  v8i16 __d = __builtin_msa_sat_s_h(__builtin_msa_srari_h((v8i16)(__a), (int)(__b)), 7); \
-  v16i8 __e = __builtin_msa_pckev_b(__builtin_msa_fill_b(0), (v16i8)__d); \
-  (v8i8)__builtin_msa_copy_s_d((v2i64)__e, 0); \
-})
-
-#define msa_qrshrn_n_s32(__a, __b) \
-({ \
-  v4i32 __d = __builtin_msa_sat_s_w(__builtin_msa_srari_w((v4i32)(__a), (int)(__b)), 15); \
-  v8i16 __e = __builtin_msa_pckev_h(__builtin_msa_fill_h(0), (v8i16)__d); \
-  (v4i16)__builtin_msa_copy_s_d((v2i64)__e, 0); \
-})
-
-#define msa_qrshrn_n_s64(__a, __b) \
-({ \
-  v2i64 __d = __builtin_msa_sat_s_d(__builtin_msa_srari_d((v2i64)(__a), (int)(__b)), 31); \
-  v4i32 __e = __builtin_msa_pckev_w(__builtin_msa_fill_w(0), (v4i32)__d); \
-  (v2i32)__builtin_msa_copy_s_d((v2i64)__e, 0); \
-})
-
-#define msa_qrshrn_n_u16(__a, __b) \
-({ \
-  v8u16 __d = __builtin_msa_sat_u_h((v8u16)__builtin_msa_srlri_h((v8i16)(__a), (int)(__b)), 7); \
-  v16i8 __e = __builtin_msa_pckev_b(__builtin_msa_fill_b(0), (v16i8)__d); \
-  (v8u8)__builtin_msa_copy_u_d((v2i64)__e, 0); \
-})
-
-#define msa_qrshrn_n_u32(__a, __b) \
-({ \
-  v4u32 __d = __builtin_msa_sat_u_w((v4u32)__builtin_msa_srlri_w((v4i32)(__a), (int)(__b)), 15); \
-  v8i16 __e = __builtin_msa_pckev_h(__builtin_msa_fill_h(0), (v8i16)__d); \
-  (v4u16)__builtin_msa_copy_u_d((v2i64)__e, 0); \
-})
-
-#define msa_qrshrn_n_u64(__a, __b) \
-({ \
-  v2u64 __d = __builtin_msa_sat_u_d((v2u64)__builtin_msa_srlri_d((v2i64)(__a), (int)(__b)), 31); \
-  v4i32 __e = __builtin_msa_pckev_w(__builtin_msa_fill_w(0), (v4i32)__d); \
-  (v2u32)__builtin_msa_copy_u_d((v2i64)__e, 0); \
-})
-
-/* Right shift elements in a 128 bits vector by an immediate value, saturate the results and them in a 64 bits vector.
-   Input is signed and output is unsigned. */
-#define msa_qrshrun_n_s16(__a, __b) \
-({ \
-  v8i16 __d = __builtin_msa_srlri_h(__builtin_msa_max_s_h(__builtin_msa_fill_h(0), (v8i16)(__a)), (int)(__b)); \
-  v16i8 __e = __builtin_msa_pckev_b(__builtin_msa_fill_b(0), (v16i8)__builtin_msa_sat_u_h((v8u16)__d, 7)); \
-  (v8u8)__builtin_msa_copy_u_d((v2i64)__e, 0); \
-})
-
-#define msa_qrshrun_n_s32(__a, __b) \
-({ \
-  v4i32 __d = __builtin_msa_srlri_w(__builtin_msa_max_s_w(__builtin_msa_fill_w(0), (v4i32)(__a)), (int)(__b)); \
-  v8i16 __e = __builtin_msa_pckev_h(__builtin_msa_fill_h(0), (v8i16)__builtin_msa_sat_u_w((v4u32)__d, 15)); \
-  (v4u16)__builtin_msa_copy_u_d((v2i64)__e, 0); \
-})
-
-#define msa_qrshrun_n_s64(__a, __b) \
-({ \
-  v2i64 __d = __builtin_msa_srlri_d(__builtin_msa_max_s_d(__builtin_msa_fill_d(0), (v2i64)(__a)), (int)(__b)); \
-  v4i32 __e = __builtin_msa_pckev_w(__builtin_msa_fill_w(0), (v4i32)__builtin_msa_sat_u_d((v2u64)__d, 31)); \
-  (v2u32)__builtin_msa_copy_u_d((v2i64)__e, 0); \
-})
-
-/* pack */
-#define msa_pack_s16(__a, __b) (__builtin_msa_pckev_b((v16i8)(__b), (v16i8)(__a)))
-#define msa_pack_s32(__a, __b) (__builtin_msa_pckev_h((v8i16)(__b), (v8i16)(__a)))
-#define msa_pack_s64(__a, __b) (__builtin_msa_pckev_w((v4i32)(__b), (v4i32)(__a)))
-#define msa_pack_u16(__a, __b) ((v16u8)__builtin_msa_pckev_b((v16i8)(__b), (v16i8)(__a)))
-#define msa_pack_u32(__a, __b) ((v8u16)__builtin_msa_pckev_h((v8i16)(__b), (v8i16)(__a)))
-#define msa_pack_u64(__a, __b) ((v4u32)__builtin_msa_pckev_w((v4i32)(__b), (v4i32)(__a)))
-
-/* qpack */
-#define msa_qpack_s16(__a, __b) \
-(__builtin_msa_pckev_b((v16i8)__builtin_msa_sat_s_h((v8i16)(__b), 7), (v16i8)__builtin_msa_sat_s_h((v8i16)(__a), 7)))
-#define msa_qpack_s32(__a, __b) \
-(__builtin_msa_pckev_h((v8i16)__builtin_msa_sat_s_w((v4i32)(__b), 15), (v8i16)__builtin_msa_sat_s_w((v4i32)(__a), 15)))
-#define msa_qpack_s64(__a, __b) \
-(__builtin_msa_pckev_w((v4i32)__builtin_msa_sat_s_d((v2i64)(__b), 31), (v4i32)__builtin_msa_sat_s_d((v2i64)(__a), 31)))
-#define msa_qpack_u16(__a, __b) \
-((v16u8)__builtin_msa_pckev_b((v16i8)__builtin_msa_sat_u_h((v8u16)(__b), 7), (v16i8)__builtin_msa_sat_u_h((v8u16)(__a), 7)))
-#define msa_qpack_u32(__a, __b) \
-((v8u16)__builtin_msa_pckev_h((v8i16)__builtin_msa_sat_u_w((v4u32)(__b), 15), (v8i16)__builtin_msa_sat_u_w((v4u32)(__a), 15)))
-#define msa_qpack_u64(__a, __b) \
-((v4u32)__builtin_msa_pckev_w((v4i32)__builtin_msa_sat_u_d((v2u64)(__b), 31), (v4i32)__builtin_msa_sat_u_d((v2u64)(__a), 31)))
-
-/* qpacku */
-#define msa_qpacku_s16(__a, __b) \
-((v16u8)__builtin_msa_pckev_b((v16i8)__builtin_msa_sat_u_h((v8u16)(__builtin_msa_max_s_h(__builtin_msa_fill_h(0), (v8i16)(__b))), 7), \
-                              (v16i8)__builtin_msa_sat_u_h((v8u16)(__builtin_msa_max_s_h(__builtin_msa_fill_h(0), (v8i16)(__a))), 7)))
-#define msa_qpacku_s32(__a, __b) \
-((v8u16)__builtin_msa_pckev_h((v8i16)__builtin_msa_sat_u_w((v4u32)(__builtin_msa_max_s_w(__builtin_msa_fill_w(0), (v4i32)(__b))), 15), \
-                              (v8i16)__builtin_msa_sat_u_w((v4u32)(__builtin_msa_max_s_w(__builtin_msa_fill_w(0), (v4i32)(__a))), 15)))
-#define msa_qpacku_s64(__a, __b) \
-((v4u32)__builtin_msa_pckev_w((v4i32)__builtin_msa_sat_u_d((v2u64)(__builtin_msa_max_s_d(__builtin_msa_fill_d(0), (v2i64)(__b))), 31), \
-                              (v4i32)__builtin_msa_sat_u_d((v2u64)(__builtin_msa_max_s_d(__builtin_msa_fill_d(0), (v2i64)(__a))), 31)))
-
-/* packr */
-#define msa_packr_s16(__a, __b, __c) \
-(__builtin_msa_pckev_b((v16i8)__builtin_msa_srai_h((v8i16)(__b), (int)(__c)), (v16i8)__builtin_msa_srai_h((v8i16)(__a), (int)(__c))))
-#define msa_packr_s32(__a, __b, __c) \
-(__builtin_msa_pckev_h((v8i16)__builtin_msa_srai_w((v4i32)(__b), (int)(__c)), (v8i16)__builtin_msa_srai_w((v4i32)(__a), (int)(__c))))
-#define msa_packr_s64(__a, __b, __c) \
-(__builtin_msa_pckev_w((v4i32)__builtin_msa_srai_d((v2i64)(__b), (int)(__c)), (v4i32)__builtin_msa_srai_d((v2i64)(__a), (int)(__c))))
-#define msa_packr_u16(__a, __b, __c) \
-((v16u8)__builtin_msa_pckev_b((v16i8)__builtin_msa_srli_h((v8i16)(__b), (int)(__c)), (v16i8)__builtin_msa_srli_h((v8i16)(__a), (int)(__c))))
-#define msa_packr_u32(__a, __b, __c) \
-((v8u16)__builtin_msa_pckev_h((v8i16)__builtin_msa_srli_w((v4i32)(__b), (int)(__c)), (v8i16)__builtin_msa_srli_w((v4i32)(__a), (int)(__c))))
-#define msa_packr_u64(__a, __b, __c) \
-((v4u32)__builtin_msa_pckev_w((v4i32)__builtin_msa_srli_d((v2i64)(__b), (int)(__c)), (v4i32)__builtin_msa_srli_d((v2i64)(__a), (int)(__c))))
-
-/* rpackr */
-#define msa_rpackr_s16(__a, __b, __c) \
-(__builtin_msa_pckev_b((v16i8)__builtin_msa_srari_h((v8i16)(__b), (int)(__c)), (v16i8)__builtin_msa_srari_h((v8i16)(__a), (int)(__c))))
-#define msa_rpackr_s32(__a, __b, __c) \
-(__builtin_msa_pckev_h((v8i16)__builtin_msa_srari_w((v4i32)(__b), (int)(__c)), (v8i16)__builtin_msa_srari_w((v4i32)(__a), (int)(__c))))
-#define msa_rpackr_s64(__a, __b, __c) \
-(__builtin_msa_pckev_w((v4i32)__builtin_msa_srari_d((v2i64)(__b), (int)(__c)), (v4i32)__builtin_msa_srari_d((v2i64)(__a), (int)(__c))))
-#define msa_rpackr_u16(__a, __b, __c) \
-((v16u8)__builtin_msa_pckev_b((v16i8)__builtin_msa_srlri_h((v8i16)(__b), (int)(__c)), (v16i8)__builtin_msa_srlri_h((v8i16)(__a), (int)(__c))))
-#define msa_rpackr_u32(__a, __b, __c) \
-((v8u16)__builtin_msa_pckev_h((v8i16)__builtin_msa_srlri_w((v4i32)(__b), (int)(__c)), (v8i16)__builtin_msa_srlri_w((v4i32)(__a), (int)(__c))))
-#define msa_rpackr_u64(__a, __b, __c) \
-((v4u32)__builtin_msa_pckev_w((v4i32)__builtin_msa_srlri_d((v2i64)(__b), (int)(__c)), (v4i32)__builtin_msa_srlri_d((v2i64)(__a), (int)(__c))))
-
-/* qrpackr */
-#define msa_qrpackr_s16(__a, __b, __c) \
-(__builtin_msa_pckev_b((v16i8)__builtin_msa_sat_s_h(__builtin_msa_srari_h((v8i16)(__b), (int)(__c)), 7), \
-                       (v16i8)__builtin_msa_sat_s_h(__builtin_msa_srari_h((v8i16)(__a), (int)(__c)), 7)))
-#define msa_qrpackr_s32(__a, __b, __c) \
-(__builtin_msa_pckev_h((v8i16)__builtin_msa_sat_s_w(__builtin_msa_srari_w((v4i32)(__b), (int)(__c)), 15), \
-                       (v8i16)__builtin_msa_sat_s_w(__builtin_msa_srari_w((v4i32)(__a), (int)(__c)), 15)))
-#define msa_qrpackr_s64(__a, __b, __c) \
-(__builtin_msa_pckev_w((v4i32)__builtin_msa_sat_s_d(__builtin_msa_srari_d((v2i64)(__b), (int)(__c)), 31), \
-                       (v4i32)__builtin_msa_sat_s_d(__builtin_msa_srari_d((v2i64)(__a), (int)(__c)), 31)))
-#define msa_qrpackr_u16(__a, __b, __c) \
-((v16u8)__builtin_msa_pckev_b((v16i8)__builtin_msa_sat_u_h((v8u16)__builtin_msa_srlri_h((v8i16)(__b), (int)(__c)), 7), \
-                              (v16i8)__builtin_msa_sat_u_h((v8u16)__builtin_msa_srlri_h((v8i16)(__a), (int)(__c)), 7)))
-#define msa_qrpackr_u32(__a, __b, __c) \
-((v8u16)__builtin_msa_pckev_h((v8i16)__builtin_msa_sat_u_w((v4u32)__builtin_msa_srlri_w((v4i32)(__b), (int)(__c)), 15), \
-                              (v8i16)__builtin_msa_sat_u_w((v4u32)__builtin_msa_srlri_w((v4i32)(__a), (int)(__c)), 15)))
-#define msa_qrpackr_u64(__a, __b, __c) \
-((v4u32)__builtin_msa_pckev_w((v4i32)__builtin_msa_sat_u_d((v2u64)__builtin_msa_srlri_d((v2i64)(__b), (int)(__c)), 31), \
-                              (v4i32)__builtin_msa_sat_u_d((v2u64)__builtin_msa_srlri_d((v2i64)(__a), (int)(__c)), 31)))
-
-/* qrpackru */
-#define msa_qrpackru_s16(__a, __b, __c) \
-({ \
-  v8i16 __d = __builtin_msa_srlri_h(__builtin_msa_max_s_h(__builtin_msa_fill_h(0), (v8i16)(__a)), (int)(__c)); \
-  v8i16 __e = __builtin_msa_srlri_h(__builtin_msa_max_s_h(__builtin_msa_fill_h(0), (v8i16)(__b)), (int)(__c)); \
-  (v16u8)__builtin_msa_pckev_b((v16i8)__builtin_msa_sat_u_h((v8u16)__e, 7), (v16i8)__builtin_msa_sat_u_h((v8u16)__d, 7)); \
-})
-
-#define msa_qrpackru_s32(__a, __b, __c) \
-({ \
-  v4i32 __d = __builtin_msa_srlri_w(__builtin_msa_max_s_w(__builtin_msa_fill_w(0), (v4i32)(__a)), (int)(__c)); \
-  v4i32 __e = __builtin_msa_srlri_w(__builtin_msa_max_s_w(__builtin_msa_fill_w(0), (v4i32)(__b)), (int)(__c)); \
-  (v8u16)__builtin_msa_pckev_h((v8i16)__builtin_msa_sat_u_w((v4u32)__e, 15), (v8i16)__builtin_msa_sat_u_w((v4u32)__d, 15)); \
-})
-
-#define msa_qrpackru_s64(__a, __b, __c) \
-({ \
-  v2i64 __d = __builtin_msa_srlri_d(__builtin_msa_max_s_d(__builtin_msa_fill_d(0), (v2i64)(__a)), (int)(__c)); \
-  v2i64 __e = __builtin_msa_srlri_d(__builtin_msa_max_s_d(__builtin_msa_fill_d(0), (v2i64)(__b)), (int)(__c)); \
-  (v4u32)__builtin_msa_pckev_w((v4i32)__builtin_msa_sat_u_d((v2u64)__e, 31), (v4i32)__builtin_msa_sat_u_d((v2u64)__d, 31)); \
-})
-
-/* Minimum values between corresponding elements in the two vectors are written to the returned vector. */
-#define msa_minq_s8(__a, __b)  (__builtin_msa_min_s_b(__a, __b))
-#define msa_minq_s16(__a, __b) (__builtin_msa_min_s_h(__a, __b))
-#define msa_minq_s32(__a, __b) (__builtin_msa_min_s_w(__a, __b))
-#define msa_minq_s64(__a, __b) (__builtin_msa_min_s_d(__a, __b))
-#define msa_minq_u8(__a, __b)  ((v16u8)__builtin_msa_min_u_b(__a, __b))
-#define msa_minq_u16(__a, __b) ((v8u16)__builtin_msa_min_u_h(__a, __b))
-#define msa_minq_u32(__a, __b) ((v4u32)__builtin_msa_min_u_w(__a, __b))
-#define msa_minq_u64(__a, __b) ((v2u64)__builtin_msa_min_u_d(__a, __b))
-#define msa_minq_f32(__a, __b) (__builtin_msa_fmin_w(__a, __b))
-#define msa_minq_f64(__a, __b) (__builtin_msa_fmin_d(__a, __b))
-
-/* Maximum values between corresponding elements in the two vectors are written to the returned vector. */
-#define msa_maxq_s8(__a, __b)  (__builtin_msa_max_s_b(__a, __b))
-#define msa_maxq_s16(__a, __b) (__builtin_msa_max_s_h(__a, __b))
-#define msa_maxq_s32(__a, __b) (__builtin_msa_max_s_w(__a, __b))
-#define msa_maxq_s64(__a, __b) (__builtin_msa_max_s_d(__a, __b))
-#define msa_maxq_u8(__a, __b)  ((v16u8)__builtin_msa_max_u_b(__a, __b))
-#define msa_maxq_u16(__a, __b) ((v8u16)__builtin_msa_max_u_h(__a, __b))
-#define msa_maxq_u32(__a, __b) ((v4u32)__builtin_msa_max_u_w(__a, __b))
-#define msa_maxq_u64(__a, __b) ((v2u64)__builtin_msa_max_u_d(__a, __b))
-#define msa_maxq_f32(__a, __b) (__builtin_msa_fmax_w(__a, __b))
-#define msa_maxq_f64(__a, __b) (__builtin_msa_fmax_d(__a, __b))
-
-/* Vector type reinterpretion */
-#define MSA_TPV_REINTERPRET(_Tpv, Vec) ((_Tpv)(Vec))
-
-/* Add the odd elements in vector __a with the even elements in vector __b to double width elements in the returned vector. */
-/* v8i16 msa_hadd_s16 ((v16i8)__a, (v16i8)__b) */
-#define msa_hadd_s16(__a, __b) (__builtin_msa_hadd_s_h((v16i8)(__a), (v16i8)(__b)))
-/* v4i32 msa_hadd_s32 ((v8i16)__a, (v8i16)__b) */
-#define msa_hadd_s32(__a, __b) (__builtin_msa_hadd_s_w((v8i16)(__a), (v8i16)(__b)))
-/* v2i64 msa_hadd_s64 ((v4i32)__a, (v4i32)__b) */
-#define msa_hadd_s64(__a, __b) (__builtin_msa_hadd_s_d((v4i32)(__a), (v4i32)(__b)))
-
-/* Copy even elements in __a to the left half and even elements in __b to the right half and return the result vector. */
-#define msa_pckev_s8(__a, __b)  (__builtin_msa_pckev_b((v16i8)(__a), (v16i8)(__b)))
-#define msa_pckev_s16(__a, __b) (__builtin_msa_pckev_h((v8i16)(__a), (v8i16)(__b)))
-#define msa_pckev_s32(__a, __b) (__builtin_msa_pckev_w((v4i32)(__a), (v4i32)(__b)))
-#define msa_pckev_s64(__a, __b) (__builtin_msa_pckev_d((v2i64)(__a), (v2i64)(__b)))
-
-/* Copy even elements in __a to the left half and even elements in __b to the right half and return the result vector. */
-#define msa_pckod_s8(__a, __b)  (__builtin_msa_pckod_b((v16i8)(__a), (v16i8)(__b)))
-#define msa_pckod_s16(__a, __b) (__builtin_msa_pckod_h((v8i16)(__a), (v8i16)(__b)))
-#define msa_pckod_s32(__a, __b) (__builtin_msa_pckod_w((v4i32)(__a), (v4i32)(__b)))
-#define msa_pckod_s64(__a, __b) (__builtin_msa_pckod_d((v2i64)(__a), (v2i64)(__b)))
-
-#ifdef _MIPSEB
-#define LANE_IMM0_1(x)  (0b1 - ((x) & 0b1))
-#define LANE_IMM0_3(x)  (0b11 - ((x) & 0b11))
-#define LANE_IMM0_7(x)  (0b111 - ((x) & 0b111))
-#define LANE_IMM0_15(x) (0b1111 - ((x) & 0b1111))
-#else
-#define LANE_IMM0_1(x)  ((x) & 0b1)
-#define LANE_IMM0_3(x)  ((x) & 0b11)
-#define LANE_IMM0_7(x)  ((x) & 0b111)
-#define LANE_IMM0_15(x) ((x) & 0b1111)
-#endif
-
-#define msa_get_lane_u8(__a, __b)        ((uint8_t)(__a)[LANE_IMM0_7(__b)])
-#define msa_get_lane_s8(__a, __b)        ((int8_t)(__a)[LANE_IMM0_7(__b)])
-#define msa_get_lane_u16(__a, __b)       ((uint16_t)(__a)[LANE_IMM0_3(__b)])
-#define msa_get_lane_s16(__a, __b)       ((int16_t)(__a)[LANE_IMM0_3(__b)])
-#define msa_get_lane_u32(__a, __b)       ((uint32_t)(__a)[LANE_IMM0_1(__b)])
-#define msa_get_lane_s32(__a, __b)       ((int32_t)(__a)[LANE_IMM0_1(__b)])
-#define msa_get_lane_f32(__a, __b)       ((float)(__a)[LANE_IMM0_3(__b)])
-#define msa_get_lane_s64(__a, __b)       ((int64_t)(__a)[LANE_IMM0_1(__b)])
-#define msa_get_lane_u64(__a, __b)       ((uint64_t)(__a)[LANE_IMM0_1(__b)])
-#define msa_get_lane_f64(__a, __b)       ((double)(__a)[LANE_IMM0_1(__b)])
-#define msa_getq_lane_u8(__a, imm0_15)   ((uint8_t)__builtin_msa_copy_u_b((v16i8)(__a), imm0_15))
-#define msa_getq_lane_s8(__a, imm0_15)   ((int8_t)__builtin_msa_copy_s_b(__a, imm0_15))
-#define msa_getq_lane_u16(__a, imm0_7)   ((uint16_t)__builtin_msa_copy_u_h((v8i16)(__a), imm0_7))
-#define msa_getq_lane_s16(__a, imm0_7)   ((int16_t)__builtin_msa_copy_s_h(__a, imm0_7))
-#define msa_getq_lane_u32(__a, imm0_3)   __builtin_msa_copy_u_w((v4i32)(__a), imm0_3)
-#define msa_getq_lane_s32                __builtin_msa_copy_s_w
-#define msa_getq_lane_f32(__a, __b)      ((float)(__a)[LANE_IMM0_3(__b)])
-#define msa_getq_lane_f64(__a, __b)      ((double)(__a)[LANE_IMM0_1(__b)])
-#if (__mips == 64)
-#define msa_getq_lane_u64(__a, imm0_1)   __builtin_msa_copy_u_d((v2i64)(__a), imm0_1)
-#define msa_getq_lane_s64                __builtin_msa_copy_s_d
-#else
-#define msa_getq_lane_u64(__a, imm0_1)   ((uint64_t)(__a)[LANE_IMM0_1(imm0_1)])
-#define msa_getq_lane_s64(__a, imm0_1)   ((int64_t)(__a)[LANE_IMM0_1(imm0_1)])
-#endif
-
-/* combine */
-#if (__mips == 64)
-#define __COMBINE_64_64(__TYPE, a, b)    ((__TYPE)((v2u64){((v1u64)(a))[0], ((v1u64)(b))[0]}))
-#else
-#define __COMBINE_64_64(__TYPE, a, b)    ((__TYPE)((v4u32){((v2u32)(a))[0], ((v2u32)(a))[1],  \
-                                                           ((v2u32)(b))[0], ((v2u32)(b))[1]}))
-#endif
-
-/* v16i8 msa_combine_s8 (v8i8 __a, v8i8 __b) */
-#define msa_combine_s8(__a, __b)  __COMBINE_64_64(v16i8, __a, __b)
-
-/* v8i16 msa_combine_s16(v4i16 __a, v4i16 __b) */
-#define msa_combine_s16(__a, __b)  __COMBINE_64_64(v8i16, __a, __b)
-
-/* v4i32 msa_combine_s32(v2i32 __a, v2i32 __b) */
-#define msa_combine_s32(__a, __b)  __COMBINE_64_64(v4i32, __a, __b)
-
-/* v2i64 msa_combine_s64(v1i64 __a, v1i64 __b) */
-#define msa_combine_s64(__a, __b)  __COMBINE_64_64(v2i64, __a, __b)
-
-/* v4f32 msa_combine_f32(v2f32 __a, v2f32 __b) */
-#define msa_combine_f32(__a, __b)  __COMBINE_64_64(v4f32, __a, __b)
-
-/* v16u8 msa_combine_u8(v8u8 __a, v8u8 __b) */
-#define msa_combine_u8(__a, __b)  __COMBINE_64_64(v16u8, __a, __b)
-
-/* v8u16 msa_combine_u16(v4u16 __a, v4u16 __b) */
-#define msa_combine_u16(__a, __b)  __COMBINE_64_64(v8u16, __a, __b)
-
-/* v4u32 msa_combine_u32(v2u32 __a, v2u32 __b) */
-#define msa_combine_u32(__a, __b)  __COMBINE_64_64(v4u32, __a, __b)
-
-/* v2u64 msa_combine_u64(v1u64 __a, v1u64 __b) */
-#define msa_combine_u64(__a, __b)  __COMBINE_64_64(v2u64, __a, __b)
-
-/* v2f64 msa_combine_f64(v1f64 __a, v1f64 __b) */
-#define msa_combine_f64(__a, __b)  __COMBINE_64_64(v2f64, __a, __b)
-
-/* get_low, get_high */
-#if (__mips == 64)
-#define __GET_LOW(__TYPE, a)   ((__TYPE)((v1u64)(__builtin_msa_copy_u_d((v2i64)(a), 0))))
-#define __GET_HIGH(__TYPE, a)  ((__TYPE)((v1u64)(__builtin_msa_copy_u_d((v2i64)(a), 1))))
-#else
-#define __GET_LOW(__TYPE, a)   ((__TYPE)(((v2u64)(a))[0]))
-#define __GET_HIGH(__TYPE, a)  ((__TYPE)(((v2u64)(a))[1]))
-#endif
-
-/* v8i8 msa_get_low_s8(v16i8 __a) */
-#define msa_get_low_s8(__a)  __GET_LOW(v8i8, __a)
-
-/* v4i16 msa_get_low_s16(v8i16 __a) */
-#define msa_get_low_s16(__a)  __GET_LOW(v4i16, __a)
-
-/* v2i32 msa_get_low_s32(v4i32 __a) */
-#define msa_get_low_s32(__a)  __GET_LOW(v2i32, __a)
-
-/* v1i64 msa_get_low_s64(v2i64 __a) */
-#define msa_get_low_s64(__a)  __GET_LOW(v1i64, __a)
-
-/* v8u8 msa_get_low_u8(v16u8 __a) */
-#define msa_get_low_u8(__a)  __GET_LOW(v8u8, __a)
-
-/* v4u16 msa_get_low_u16(v8u16 __a) */
-#define msa_get_low_u16(__a)  __GET_LOW(v4u16, __a)
-
-/* v2u32 msa_get_low_u32(v4u32 __a) */
-#define msa_get_low_u32(__a)  __GET_LOW(v2u32, __a)
-
-/* v1u64 msa_get_low_u64(v2u64 __a) */
-#define msa_get_low_u64(__a)  __GET_LOW(v1u64, __a)
-
-/* v2f32 msa_get_low_f32(v4f32 __a) */
-#define msa_get_low_f32(__a)  __GET_LOW(v2f32, __a)
-
-/* v1f64 msa_get_low_f64(v2f64 __a) */
-#define msa_get_low_f64(__a)  __GET_LOW(v1f64, __a)
-
-/* v8i8 msa_get_high_s8(v16i8 __a) */
-#define msa_get_high_s8(__a)  __GET_HIGH(v8i8, __a)
-
-/* v4i16 msa_get_high_s16(v8i16 __a) */
-#define msa_get_high_s16(__a)  __GET_HIGH(v4i16, __a)
-
-/* v2i32 msa_get_high_s32(v4i32 __a) */
-#define msa_get_high_s32(__a)  __GET_HIGH(v2i32, __a)
-
-/* v1i64 msa_get_high_s64(v2i64 __a) */
-#define msa_get_high_s64(__a)  __GET_HIGH(v1i64, __a)
-
-/* v8u8 msa_get_high_u8(v16u8 __a) */
-#define msa_get_high_u8(__a)  __GET_HIGH(v8u8, __a)
-
-/* v4u16 msa_get_high_u16(v8u16 __a) */
-#define msa_get_high_u16(__a)  __GET_HIGH(v4u16, __a)
-
-/* v2u32 msa_get_high_u32(v4u32 __a) */
-#define msa_get_high_u32(__a)  __GET_HIGH(v2u32, __a)
-
-/* v1u64 msa_get_high_u64(v2u64 __a) */
-#define msa_get_high_u64(__a)  __GET_HIGH(v1u64, __a)
-
-/* v2f32 msa_get_high_f32(v4f32 __a) */
-#define msa_get_high_f32(__a)  __GET_HIGH(v2f32, __a)
-
-/* v1f64 msa_get_high_f64(v2f64 __a) */
-#define msa_get_high_f64(__a)  __GET_HIGH(v1f64, __a)
-
-/* ri = ai * b[lane] */
-/* v4f32 msa_mulq_lane_f32(v4f32 __a, v4f32 __b, const int __lane) */
-#define msa_mulq_lane_f32(__a, __b, __lane)  ((__a) * msa_getq_lane_f32(__b, __lane))
-
-/* ri = ai + bi * c[lane] */
-/* v4f32 msa_mlaq_lane_f32(v4f32 __a, v4f32 __b, v4f32 __c, const int __lane) */
-#define msa_mlaq_lane_f32(__a, __b, __c, __lane)  ((__a) + ((__b) * msa_getq_lane_f32(__c, __lane)))
-
-/* uint16_t msa_sum_u16(v8u16 __a)*/
-#define msa_sum_u16(__a)                         \
-({                                               \
-  v4u32 _b;                                      \
-  v2u64 _c;                                      \
-  _b = __builtin_msa_hadd_u_w(__a, __a);         \
-  _c = __builtin_msa_hadd_u_d(_b, _b);           \
-  (uint16_t)(_c[0] + _c[1]);                     \
-})
-
-/* int16_t msa_sum_s16(v8i16 __a) */
-#define msa_sum_s16(__a)                        \
-({                                              \
-  v4i32 _b;                                     \
-  v2i64 _c;                                     \
-  _b = __builtin_msa_hadd_s_w(__a, __a);        \
-  _c = __builtin_msa_hadd_s_d(_b, _b);          \
-  (int16_t)(_c[0] + _c[1]);                     \
-})
-
-
-/* uint32_t msa_sum_u32(v4u32 __a)*/
-#define msa_sum_u32(__a)                       \
-({                                             \
-  v2u64 _b;                                    \
-  _b = __builtin_msa_hadd_u_d(__a, __a);       \
-  (uint32_t)(_b[0] + _b[1]);                   \
-})
-
-/* int32_t  msa_sum_s32(v4i32 __a)*/
-#define msa_sum_s32(__a)                       \
-({                                             \
-  v2i64 _b;                                    \
-  _b = __builtin_msa_hadd_s_d(__a, __a);       \
-  (int32_t)(_b[0] + _b[1]);                    \
-})
-
-/* uint8_t msa_sum_u8(v16u8 __a)*/
-#define msa_sum_u8(__a)                        \
-({                                             \
-  v8u16 _b16;                                    \
-  v4u32 _c32;                                    \
-  _b16 = __builtin_msa_hadd_u_h(__a, __a);       \
-  _c32 = __builtin_msa_hadd_u_w(_b16, _b16);         \
-  (uint8_t)msa_sum_u32(_c32);                    \
-})
-
-/* int8_t msa_sum_s8(v16s8 __a)*/
-#define msa_sum_s8(__a)                        \
-({                                             \
-  v8i16 _b16;                                    \
-  v4i32 _c32;                                    \
-  _b16 = __builtin_msa_hadd_s_h(__a, __a);       \
-  _c32 = __builtin_msa_hadd_s_w(_b16, _b16);         \
-  (int8_t)msa_sum_s32(_c32);                     \
-})
-
-/* float msa_sum_f32(v4f32 __a)*/
-#define msa_sum_f32(__a)  ((__a)[0] + (__a)[1] + (__a)[2] + (__a)[3])
-
-/* v8u16 msa_paddlq_u8(v16u8 __a) */
-#define msa_paddlq_u8(__a)  (__builtin_msa_hadd_u_h(__a, __a))
-
-/* v8i16 msa_paddlq_s8(v16i8 __a) */
-#define msa_paddlq_s8(__a)  (__builtin_msa_hadd_s_h(__a, __a))
-
-/* v4u32 msa_paddlq_u16 (v8u16 __a)*/
-#define msa_paddlq_u16(__a)  (__builtin_msa_hadd_u_w(__a, __a))
-
-/* v4i32 msa_paddlq_s16 (v8i16 __a)*/
-#define msa_paddlq_s16(__a)  (__builtin_msa_hadd_s_w(__a, __a))
-
-/* v2u64 msa_paddlq_u32(v4u32 __a) */
-#define msa_paddlq_u32(__a)  (__builtin_msa_hadd_u_d(__a, __a))
-
-/* v2i64 msa_paddlq_s32(v4i32 __a) */
-#define msa_paddlq_s32(__a)  (__builtin_msa_hadd_s_d(__a, __a))
-
-#define V8U8_2_V8U16(x)   {(uint16_t)x[0], (uint16_t)x[1], (uint16_t)x[2], (uint16_t)x[3], \
-                           (uint16_t)x[4], (uint16_t)x[5], (uint16_t)x[6], (uint16_t)x[7]}
-#define V8U8_2_V8I16(x)   {(int16_t)x[0], (int16_t)x[1], (int16_t)x[2], (int16_t)x[3], \
-                           (int16_t)x[4], (int16_t)x[5], (int16_t)x[6], (int16_t)x[7]}
-#define V8I8_2_V8I16(x)   {(int16_t)x[0], (int16_t)x[1], (int16_t)x[2], (int16_t)x[3], \
-                           (int16_t)x[4], (int16_t)x[5], (int16_t)x[6], (int16_t)x[7]}
-#define V4U16_2_V4U32(x)  {(uint32_t)x[0], (uint32_t)x[1], (uint32_t)x[2], (uint32_t)x[3]}
-#define V4U16_2_V4I32(x)  {(int32_t)x[0], (int32_t)x[1], (int32_t)x[2], (int32_t)x[3]}
-#define V4I16_2_V4I32(x)  {(int32_t)x[0], (int32_t)x[1], (int32_t)x[2], (int32_t)x[3]}
-#define V2U32_2_V2U64(x)  {(uint64_t)x[0], (uint64_t)x[1]}
-#define V2U32_2_V2I64(x)  {(int64_t)x[0], (int64_t)x[1]}
-
-/* v8u16 msa_mull_u8(v8u8 __a, v8u8 __b) */
-#define msa_mull_u8(__a, __b)  ((v8u16)__builtin_msa_mulv_h((v8i16)V8U8_2_V8I16(__a), (v8i16)V8U8_2_V8I16(__b)))
-
-/* v8i16 msa_mull_s8(v8i8 __a, v8i8 __b)*/
-#define msa_mull_s8(__a, __b)  (__builtin_msa_mulv_h((v8i16)V8I8_2_V8I16(__a), (v8i16)V8I8_2_V8I16(__b)))
-
-/* v4u32 msa_mull_u16(v4u16 __a, v4u16 __b) */
-#define msa_mull_u16(__a, __b)  ((v4u32)__builtin_msa_mulv_w((v4i32)V4U16_2_V4I32(__a), (v4i32)V4U16_2_V4I32(__b)))
-
-/* v4i32 msa_mull_s16(v4i16 __a, v4i16 __b) */
-#define msa_mull_s16(__a, __b)  (__builtin_msa_mulv_w((v4i32)V4I16_2_V4I32(__a), (v4i32)V4I16_2_V4I32(__b)))
-
-/* v2u64 msa_mull_u32(v2u32 __a, v2u32 __b) */
-#define msa_mull_u32(__a, __b)  ((v2u64)__builtin_msa_mulv_d((v2i64)V2U32_2_V2I64(__a), (v2i64)V2U32_2_V2I64(__b)))
-
-/* bitwise and: __builtin_msa_and_v */
-#define msa_andq_u8(__a, __b)  ((v16u8)__builtin_msa_and_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_andq_s8(__a, __b)  ((v16i8)__builtin_msa_and_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_andq_u16(__a, __b) ((v8u16)__builtin_msa_and_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_andq_s16(__a, __b) ((v8i16)__builtin_msa_and_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_andq_u32(__a, __b) ((v4u32)__builtin_msa_and_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_andq_s32(__a, __b) ((v4i32)__builtin_msa_and_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_andq_u64(__a, __b) ((v2u64)__builtin_msa_and_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_andq_s64(__a, __b) ((v2i64)__builtin_msa_and_v((v16u8)(__a), (v16u8)(__b)))
-
-/* bitwise or: __builtin_msa_or_v */
-#define msa_orrq_u8(__a, __b)  ((v16u8)__builtin_msa_or_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_orrq_s8(__a, __b)  ((v16i8)__builtin_msa_or_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_orrq_u16(__a, __b) ((v8u16)__builtin_msa_or_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_orrq_s16(__a, __b) ((v8i16)__builtin_msa_or_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_orrq_u32(__a, __b) ((v4u32)__builtin_msa_or_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_orrq_s32(__a, __b) ((v4i32)__builtin_msa_or_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_orrq_u64(__a, __b) ((v2u64)__builtin_msa_or_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_orrq_s64(__a, __b) ((v2i64)__builtin_msa_or_v((v16u8)(__a), (v16u8)(__b)))
-
-/* bitwise xor: __builtin_msa_xor_v */
-#define msa_eorq_u8(__a, __b)  ((v16u8)__builtin_msa_xor_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_eorq_s8(__a, __b)  ((v16i8)__builtin_msa_xor_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_eorq_u16(__a, __b) ((v8u16)__builtin_msa_xor_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_eorq_s16(__a, __b) ((v8i16)__builtin_msa_xor_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_eorq_u32(__a, __b) ((v4u32)__builtin_msa_xor_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_eorq_s32(__a, __b) ((v4i32)__builtin_msa_xor_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_eorq_u64(__a, __b) ((v2u64)__builtin_msa_xor_v((v16u8)(__a), (v16u8)(__b)))
-#define msa_eorq_s64(__a, __b) ((v2i64)__builtin_msa_xor_v((v16u8)(__a), (v16u8)(__b)))
-
-/* bitwise not: v16u8 __builtin_msa_xori_b (v16u8, 0xff) */
-#define msa_mvnq_u8(__a)  ((v16u8)__builtin_msa_xori_b((v16u8)(__a), 0xFF))
-#define msa_mvnq_s8(__a)  ((v16i8)__builtin_msa_xori_b((v16u8)(__a), 0xFF))
-#define msa_mvnq_u16(__a) ((v8u16)__builtin_msa_xori_b((v16u8)(__a), 0xFF))
-#define msa_mvnq_s16(__a) ((v8i16)__builtin_msa_xori_b((v16u8)(__a), 0xFF))
-#define msa_mvnq_u32(__a) ((v4u32)__builtin_msa_xori_b((v16u8)(__a), 0xFF))
-#define msa_mvnq_s32(__a) ((v4i32)__builtin_msa_xori_b((v16u8)(__a), 0xFF))
-#define msa_mvnq_u64(__a) ((v2u64)__builtin_msa_xori_b((v16u8)(__a), 0xFF))
-#define msa_mvnq_s64(__a) ((v2i64)__builtin_msa_xori_b((v16u8)(__a), 0xFF))
-
-/* compare equal: ceq -> ri = ai == bi ? 1...1:0...0 */
-#define msa_ceqq_u8(__a, __b)  ((v16u8)__builtin_msa_ceq_b((v16i8)(__a), (v16i8)(__b)))
-#define msa_ceqq_s8(__a, __b)  ((v16u8)__builtin_msa_ceq_b((v16i8)(__a), (v16i8)(__b)))
-#define msa_ceqq_u16(__a, __b) ((v8u16)__builtin_msa_ceq_h((v8i16)(__a), (v8i16)(__b)))
-#define msa_ceqq_s16(__a, __b) ((v8u16)__builtin_msa_ceq_h((v8i16)(__a), (v8i16)(__b)))
-#define msa_ceqq_u32(__a, __b) ((v4u32)__builtin_msa_ceq_w((v4i32)(__a), (v4i32)(__b)))
-#define msa_ceqq_s32(__a, __b) ((v4u32)__builtin_msa_ceq_w((v4i32)(__a), (v4i32)(__b)))
-#define msa_ceqq_f32(__a, __b) ((v4u32)__builtin_msa_fceq_w((v4f32)(__a), (v4f32)(__b)))
-#define msa_ceqq_u64(__a, __b) ((v2u64)__builtin_msa_ceq_d((v2i64)(__a), (v2i64)(__b)))
-#define msa_ceqq_s64(__a, __b) ((v2u64)__builtin_msa_ceq_d((v2i64)(__a), (v2i64)(__b)))
-#define msa_ceqq_f64(__a, __b) ((v2u64)__builtin_msa_fceq_d((v2f64)(__a), (v2f64)(__b)))
-
-/* Compare less-than: clt -> ri = ai < bi ? 1...1:0...0 */
-#define msa_cltq_u8(__a, __b)  ((v16u8)__builtin_msa_clt_u_b((v16u8)(__a), (v16u8)(__b)))
-#define msa_cltq_s8(__a, __b)  ((v16u8)__builtin_msa_clt_s_b((v16i8)(__a), (v16i8)(__b)))
-#define msa_cltq_u16(__a, __b) ((v8u16)__builtin_msa_clt_u_h((v8u16)(__a), (v8u16)(__b)))
-#define msa_cltq_s16(__a, __b) ((v8u16)__builtin_msa_clt_s_h((v8i16)(__a), (v8i16)(__b)))
-#define msa_cltq_u32(__a, __b) ((v4u32)__builtin_msa_clt_u_w((v4u32)(__a), (v4u32)(__b)))
-#define msa_cltq_s32(__a, __b) ((v4u32)__builtin_msa_clt_s_w((v4i32)(__a), (v4i32)(__b)))
-#define msa_cltq_f32(__a, __b) ((v4u32)__builtin_msa_fclt_w((v4f32)(__a), (v4f32)(__b)))
-#define msa_cltq_u64(__a, __b) ((v2u64)__builtin_msa_clt_u_d((v2u64)(__a), (v2u64)(__b)))
-#define msa_cltq_s64(__a, __b) ((v2u64)__builtin_msa_clt_s_d((v2i64)(__a), (v2i64)(__b)))
-#define msa_cltq_f64(__a, __b) ((v2u64)__builtin_msa_fclt_d((v2f64)(__a), (v2f64)(__b)))
-
-/* compare greater-than: cgt -> ri = ai > bi ? 1...1:0...0 */
-#define msa_cgtq_u8(__a, __b)  ((v16u8)__builtin_msa_clt_u_b((v16u8)(__b), (v16u8)(__a)))
-#define msa_cgtq_s8(__a, __b)  ((v16u8)__builtin_msa_clt_s_b((v16i8)(__b), (v16i8)(__a)))
-#define msa_cgtq_u16(__a, __b) ((v8u16)__builtin_msa_clt_u_h((v8u16)(__b), (v8u16)(__a)))
-#define msa_cgtq_s16(__a, __b) ((v8u16)__builtin_msa_clt_s_h((v8i16)(__b), (v8i16)(__a)))
-#define msa_cgtq_u32(__a, __b) ((v4u32)__builtin_msa_clt_u_w((v4u32)(__b), (v4u32)(__a)))
-#define msa_cgtq_s32(__a, __b) ((v4u32)__builtin_msa_clt_s_w((v4i32)(__b), (v4i32)(__a)))
-#define msa_cgtq_f32(__a, __b) ((v4u32)__builtin_msa_fclt_w((v4f32)(__b), (v4f32)(__a)))
-#define msa_cgtq_u64(__a, __b) ((v2u64)__builtin_msa_clt_u_d((v2u64)(__b), (v2u64)(__a)))
-#define msa_cgtq_s64(__a, __b) ((v2u64)__builtin_msa_clt_s_d((v2i64)(__b), (v2i64)(__a)))
-#define msa_cgtq_f64(__a, __b) ((v2u64)__builtin_msa_fclt_d((v2f64)(__b), (v2f64)(__a)))
-
-/* compare less-equal: cle -> ri = ai <= bi ? 1...1:0...0 */
-#define msa_cleq_u8(__a, __b)  ((v16u8)__builtin_msa_cle_u_b((v16u8)(__a), (v16u8)(__b)))
-#define msa_cleq_s8(__a, __b)  ((v16u8)__builtin_msa_cle_s_b((v16i8)(__a), (v16i8)(__b)))
-#define msa_cleq_u16(__a, __b) ((v8u16)__builtin_msa_cle_u_h((v8u16)(__a), (v8u16)(__b)))
-#define msa_cleq_s16(__a, __b) ((v8u16)__builtin_msa_cle_s_h((v8i16)(__a), (v8i16)(__b)))
-#define msa_cleq_u32(__a, __b) ((v4u32)__builtin_msa_cle_u_w((v4u32)(__a), (v4u32)(__b)))
-#define msa_cleq_s32(__a, __b) ((v4u32)__builtin_msa_cle_s_w((v4i32)(__a), (v4i32)(__b)))
-#define msa_cleq_f32(__a, __b) ((v4u32)__builtin_msa_fcle_w((v4f32)(__a), (v4f32)(__b)))
-#define msa_cleq_u64(__a, __b) ((v2u64)__builtin_msa_cle_u_d((v2u64)(__a), (v2u64)(__b)))
-#define msa_cleq_s64(__a, __b) ((v2u64)__builtin_msa_cle_s_d((v2i64)(__a), (v2i64)(__b)))
-#define msa_cleq_f64(__a, __b) ((v2u64)__builtin_msa_fcle_d((v2f64)(__a), (v2f64)(__b)))
-
-/* compare greater-equal: cge -> ri = ai >= bi ? 1...1:0...0 */
-#define msa_cgeq_u8(__a, __b)  ((v16u8)__builtin_msa_cle_u_b((v16u8)(__b), (v16u8)(__a)))
-#define msa_cgeq_s8(__a, __b)  ((v16u8)__builtin_msa_cle_s_b((v16i8)(__b), (v16i8)(__a)))
-#define msa_cgeq_u16(__a, __b) ((v8u16)__builtin_msa_cle_u_h((v8u16)(__b), (v8u16)(__a)))
-#define msa_cgeq_s16(__a, __b) ((v8u16)__builtin_msa_cle_s_h((v8i16)(__b), (v8i16)(__a)))
-#define msa_cgeq_u32(__a, __b) ((v4u32)__builtin_msa_cle_u_w((v4u32)(__b), (v4u32)(__a)))
-#define msa_cgeq_s32(__a, __b) ((v4u32)__builtin_msa_cle_s_w((v4i32)(__b), (v4i32)(__a)))
-#define msa_cgeq_f32(__a, __b) ((v4u32)__builtin_msa_fcle_w((v4f32)(__b), (v4f32)(__a)))
-#define msa_cgeq_u64(__a, __b) ((v2u64)__builtin_msa_cle_u_d((v2u64)(__b), (v2u64)(__a)))
-#define msa_cgeq_s64(__a, __b) ((v2u64)__builtin_msa_cle_s_d((v2i64)(__b), (v2i64)(__a)))
-#define msa_cgeq_f64(__a, __b) ((v2u64)__builtin_msa_fcle_d((v2f64)(__b), (v2f64)(__a)))
-
-/* Shift Left Logical: shl -> ri = ai << bi; */
-#define msa_shlq_u8(__a, __b)  ((v16u8)__builtin_msa_sll_b((v16i8)(__a), (v16i8)(__b)))
-#define msa_shlq_s8(__a, __b)  ((v16i8)__builtin_msa_sll_b((v16i8)(__a), (v16i8)(__b)))
-#define msa_shlq_u16(__a, __b) ((v8u16)__builtin_msa_sll_h((v8i16)(__a), (v8i16)(__b)))
-#define msa_shlq_s16(__a, __b) ((v8i16)__builtin_msa_sll_h((v8i16)(__a), (v8i16)(__b)))
-#define msa_shlq_u32(__a, __b) ((v4u32)__builtin_msa_sll_w((v4i32)(__a), (v4i32)(__b)))
-#define msa_shlq_s32(__a, __b) ((v4i32)__builtin_msa_sll_w((v4i32)(__a), (v4i32)(__b)))
-#define msa_shlq_u64(__a, __b) ((v2u64)__builtin_msa_sll_d((v2i64)(__a), (v2i64)(__b)))
-#define msa_shlq_s64(__a, __b) ((v2i64)__builtin_msa_sll_d((v2i64)(__a), (v2i64)(__b)))
-
-/* Immediate Shift Left Logical: shl -> ri = ai << imm; */
-#define msa_shlq_n_u8(__a, __imm)  ((v16u8)__builtin_msa_slli_b((v16i8)(__a), __imm))
-#define msa_shlq_n_s8(__a, __imm)  ((v16i8)__builtin_msa_slli_b((v16i8)(__a), __imm))
-#define msa_shlq_n_u16(__a, __imm) ((v8u16)__builtin_msa_slli_h((v8i16)(__a), __imm))
-#define msa_shlq_n_s16(__a, __imm) ((v8i16)__builtin_msa_slli_h((v8i16)(__a), __imm))
-#define msa_shlq_n_u32(__a, __imm) ((v4u32)__builtin_msa_slli_w((v4i32)(__a), __imm))
-#define msa_shlq_n_s32(__a, __imm) ((v4i32)__builtin_msa_slli_w((v4i32)(__a), __imm))
-#define msa_shlq_n_u64(__a, __imm) ((v2u64)__builtin_msa_slli_d((v2i64)(__a), __imm))
-#define msa_shlq_n_s64(__a, __imm) ((v2i64)__builtin_msa_slli_d((v2i64)(__a), __imm))
-
-/* shift right: shrq -> ri = ai >> bi; */
-#define msa_shrq_u8(__a, __b)  ((v16u8)__builtin_msa_srl_b((v16i8)(__a), (v16i8)(__b)))
-#define msa_shrq_s8(__a, __b)  ((v16i8)__builtin_msa_sra_b((v16i8)(__a), (v16i8)(__b)))
-#define msa_shrq_u16(__a, __b) ((v8u16)__builtin_msa_srl_h((v8i16)(__a), (v8i16)(__b)))
-#define msa_shrq_s16(__a, __b) ((v8i16)__builtin_msa_sra_h((v8i16)(__a), (v8i16)(__b)))
-#define msa_shrq_u32(__a, __b) ((v4u32)__builtin_msa_srl_w((v4i32)(__a), (v4i32)(__b)))
-#define msa_shrq_s32(__a, __b) ((v4i32)__builtin_msa_sra_w((v4i32)(__a), (v4i32)(__b)))
-#define msa_shrq_u64(__a, __b) ((v2u64)__builtin_msa_srl_d((v2i64)(__a), (v2i64)(__b)))
-#define msa_shrq_s64(__a, __b) ((v2i64)__builtin_msa_sra_d((v2i64)(__a), (v2i64)(__b)))
-
-/* Immediate Shift Right: shr -> ri = ai >> imm; */
-#define msa_shrq_n_u8(__a, __imm)  ((v16u8)__builtin_msa_srli_b((v16i8)(__a), __imm))
-#define msa_shrq_n_s8(__a, __imm)  ((v16i8)__builtin_msa_srai_b((v16i8)(__a), __imm))
-#define msa_shrq_n_u16(__a, __imm) ((v8u16)__builtin_msa_srli_h((v8i16)(__a), __imm))
-#define msa_shrq_n_s16(__a, __imm) ((v8i16)__builtin_msa_srai_h((v8i16)(__a), __imm))
-#define msa_shrq_n_u32(__a, __imm) ((v4u32)__builtin_msa_srli_w((v4i32)(__a), __imm))
-#define msa_shrq_n_s32(__a, __imm) ((v4i32)__builtin_msa_srai_w((v4i32)(__a), __imm))
-#define msa_shrq_n_u64(__a, __imm) ((v2u64)__builtin_msa_srli_d((v2i64)(__a), __imm))
-#define msa_shrq_n_s64(__a, __imm) ((v2i64)__builtin_msa_srai_d((v2i64)(__a), __imm))
-
-/* Immediate Shift Right Rounded: shr -> ri = ai >> (rounded)imm; */
-#define msa_rshrq_n_u8(__a, __imm)  ((v16u8)__builtin_msa_srlri_b((v16i8)(__a), __imm))
-#define msa_rshrq_n_s8(__a, __imm)  ((v16i8)__builtin_msa_srari_b((v16i8)(__a), __imm))
-#define msa_rshrq_n_u16(__a, __imm) ((v8u16)__builtin_msa_srlri_h((v8i16)(__a), __imm))
-#define msa_rshrq_n_s16(__a, __imm) ((v8i16)__builtin_msa_srari_h((v8i16)(__a), __imm))
-#define msa_rshrq_n_u32(__a, __imm) ((v4u32)__builtin_msa_srlri_w((v4i32)(__a), __imm))
-#define msa_rshrq_n_s32(__a, __imm) ((v4i32)__builtin_msa_srari_w((v4i32)(__a), __imm))
-#define msa_rshrq_n_u64(__a, __imm) ((v2u64)__builtin_msa_srlri_d((v2i64)(__a), __imm))
-#define msa_rshrq_n_s64(__a, __imm) ((v2i64)__builtin_msa_srari_d((v2i64)(__a), __imm))
-
-/* Vector saturating rounding shift left, qrshl -> ri = ai << bi; */
-#define msa_qrshrq_s32(a, b)  ((v4i32)__msa_srar_w((v4i32)(a), (v4i32)(b)))
-
-/* Rename the msa builtin func to unify the name style for intrin_msa.hpp */
-#define msa_qaddq_u8          __builtin_msa_adds_u_b
-#define msa_qaddq_s8          __builtin_msa_adds_s_b
-#define msa_qaddq_u16         __builtin_msa_adds_u_h
-#define msa_qaddq_s16         __builtin_msa_adds_s_h
-#define msa_qaddq_u32         __builtin_msa_adds_u_w
-#define msa_qaddq_s32         __builtin_msa_adds_s_w
-#define msa_qaddq_u64         __builtin_msa_adds_u_d
-#define msa_qaddq_s64         __builtin_msa_adds_s_d
-#define msa_addq_u8(a, b)     ((v16u8)__builtin_msa_addv_b((v16i8)(a), (v16i8)(b)))
-#define msa_addq_s8           __builtin_msa_addv_b
-#define msa_addq_u16(a, b)    ((v8u16)__builtin_msa_addv_h((v8i16)(a), (v8i16)(b)))
-#define msa_addq_s16          __builtin_msa_addv_h
-#define msa_addq_u32(a, b)    ((v4u32)__builtin_msa_addv_w((v4i32)(a), (v4i32)(b)))
-#define msa_addq_s32          __builtin_msa_addv_w
-#define msa_addq_f32          __builtin_msa_fadd_w
-#define msa_addq_u64(a, b)    ((v2u64)__builtin_msa_addv_d((v2i64)(a), (v2i64)(b)))
-#define msa_addq_s64          __builtin_msa_addv_d
-#define msa_addq_f64          __builtin_msa_fadd_d
-#define msa_qsubq_u8          __builtin_msa_subs_u_b
-#define msa_qsubq_s8          __builtin_msa_subs_s_b
-#define msa_qsubq_u16         __builtin_msa_subs_u_h
-#define msa_qsubq_s16         __builtin_msa_subs_s_h
-#define msa_subq_u8(a, b)     ((v16u8)__builtin_msa_subv_b((v16i8)(a), (v16i8)(b)))
-#define msa_subq_s8           __builtin_msa_subv_b
-#define msa_subq_u16(a, b)    ((v8u16)__builtin_msa_subv_h((v8i16)(a), (v8i16)(b)))
-#define msa_subq_s16          __builtin_msa_subv_h
-#define msa_subq_u32(a, b)    ((v4u32)__builtin_msa_subv_w((v4i32)(a), (v4i32)(b)))
-#define msa_subq_s32          __builtin_msa_subv_w
-#define msa_subq_f32          __builtin_msa_fsub_w
-#define msa_subq_u64(a, b)    ((v2u64)__builtin_msa_subv_d((v2i64)(a), (v2i64)(b)))
-#define msa_subq_s64          __builtin_msa_subv_d
-#define msa_subq_f64          __builtin_msa_fsub_d
-#define msa_mulq_u8(a, b)     ((v16u8)__builtin_msa_mulv_b((v16i8)(a), (v16i8)(b)))
-#define msa_mulq_s8(a, b)     ((v16i8)__builtin_msa_mulv_b((v16i8)(a), (v16i8)(b)))
-#define msa_mulq_u16(a, b)    ((v8u16)__builtin_msa_mulv_h((v8i16)(a), (v8i16)(b)))
-#define msa_mulq_s16(a, b)    ((v8i16)__builtin_msa_mulv_h((v8i16)(a), (v8i16)(b)))
-#define msa_mulq_u32(a, b)    ((v4u32)__builtin_msa_mulv_w((v4i32)(a), (v4i32)(b)))
-#define msa_mulq_s32(a, b)    ((v4i32)__builtin_msa_mulv_w((v4i32)(a), (v4i32)(b)))
-#define msa_mulq_u64(a, b)    ((v2u64)__builtin_msa_mulv_d((v2i64)(a), (v2i64)(b)))
-#define msa_mulq_s64(a, b)    ((v2i64)__builtin_msa_mulv_d((v2i64)(a), (v2i64)(b)))
-#define msa_mulq_f32          __builtin_msa_fmul_w
-#define msa_mulq_f64          __builtin_msa_fmul_d
-#define msa_divq_f32          __builtin_msa_fdiv_w
-#define msa_divq_f64          __builtin_msa_fdiv_d
-#define msa_dotp_s_h          __builtin_msa_dotp_s_h
-#define msa_dotp_s_w          __builtin_msa_dotp_s_w
-#define msa_dotp_s_d          __builtin_msa_dotp_s_d
-#define msa_dotp_u_h          __builtin_msa_dotp_u_h
-#define msa_dotp_u_w          __builtin_msa_dotp_u_w
-#define msa_dotp_u_d          __builtin_msa_dotp_u_d
-#define msa_dpadd_s_h         __builtin_msa_dpadd_s_h
-#define msa_dpadd_s_w         __builtin_msa_dpadd_s_w
-#define msa_dpadd_s_d         __builtin_msa_dpadd_s_d
-#define msa_dpadd_u_h         __builtin_msa_dpadd_u_h
-#define msa_dpadd_u_w         __builtin_msa_dpadd_u_w
-#define msa_dpadd_u_d         __builtin_msa_dpadd_u_d
-
-#define ILVRL_B2(RTYPE, in0, in1, low, hi) do {       \
-      low = (RTYPE)__builtin_msa_ilvr_b((v16i8)(in0), (v16i8)(in1));  \
-      hi  = (RTYPE)__builtin_msa_ilvl_b((v16i8)(in0), (v16i8)(in1));  \
-    } while (0)
-#define ILVRL_B2_UB(...) ILVRL_B2(v16u8, __VA_ARGS__)
-#define ILVRL_B2_SB(...) ILVRL_B2(v16i8, __VA_ARGS__)
-#define ILVRL_B2_UH(...) ILVRL_B2(v8u16, __VA_ARGS__)
-#define ILVRL_B2_SH(...) ILVRL_B2(v8i16, __VA_ARGS__)
-#define ILVRL_B2_SW(...) ILVRL_B2(v4i32, __VA_ARGS__)
-
-#define ILVRL_H2(RTYPE, in0, in1, low, hi) do {       \
-      low = (RTYPE)__builtin_msa_ilvr_h((v8i16)(in0), (v8i16)(in1));  \
-      hi  = (RTYPE)__builtin_msa_ilvl_h((v8i16)(in0), (v8i16)(in1));  \
-    } while (0)
-#define ILVRL_H2_UB(...) ILVRL_H2(v16u8, __VA_ARGS__)
-#define ILVRL_H2_SB(...) ILVRL_H2(v16i8, __VA_ARGS__)
-#define ILVRL_H2_UH(...) ILVRL_H2(v8u16, __VA_ARGS__)
-#define ILVRL_H2_SH(...) ILVRL_H2(v8i16, __VA_ARGS__)
-#define ILVRL_H2_SW(...) ILVRL_H2(v4i32, __VA_ARGS__)
-#define ILVRL_H2_UW(...) ILVRL_H2(v4u32, __VA_ARGS__)
-
-#define ILVRL_W2(RTYPE, in0, in1, low, hi) do {       \
-      low = (RTYPE)__builtin_msa_ilvr_w((v4i32)(in0), (v4i32)(in1));  \
-      hi  = (RTYPE)__builtin_msa_ilvl_w((v4i32)(in0), (v4i32)(in1));  \
-    } while (0)
-#define ILVRL_W2_UB(...) ILVRL_W2(v16u8, __VA_ARGS__)
-#define ILVRL_W2_SH(...) ILVRL_W2(v8i16, __VA_ARGS__)
-#define ILVRL_W2_SW(...) ILVRL_W2(v4i32, __VA_ARGS__)
-#define ILVRL_W2_UW(...) ILVRL_W2(v4u32, __VA_ARGS__)
-
-/* absq, qabsq (r = |a|;) */
-#define msa_absq_s8(a)        __builtin_msa_add_a_b(a, __builtin_msa_fill_b(0))
-#define msa_absq_s16(a)       __builtin_msa_add_a_h(a, __builtin_msa_fill_h(0))
-#define msa_absq_s32(a)       __builtin_msa_add_a_w(a, __builtin_msa_fill_w(0))
-#define msa_absq_s64(a)       __builtin_msa_add_a_d(a, __builtin_msa_fill_d(0))
-#define msa_absq_f32(a)       ((v4f32)__builtin_msa_bclri_w((v4u32)(a), 31))
-#define msa_absq_f64(a)       ((v2f64)__builtin_msa_bclri_d((v2u64)(a), 63))
-#define msa_qabsq_s8(a)       __builtin_msa_adds_a_b(a, __builtin_msa_fill_b(0))
-#define msa_qabsq_s16(a)      __builtin_msa_adds_a_h(a, __builtin_msa_fill_h(0))
-#define msa_qabsq_s32(a)      __builtin_msa_adds_a_w(a, __builtin_msa_fill_w(0))
-#define msa_qabsq_s64(a)      __builtin_msa_adds_a_d(a, __builtin_msa_fill_d(0))
-
-/* abdq, qabdq (r = |a - b|;) */
-#define msa_abdq_u8           __builtin_msa_asub_u_b
-#define msa_abdq_s8           __builtin_msa_asub_s_b
-#define msa_abdq_u16          __builtin_msa_asub_u_h
-#define msa_abdq_s16          __builtin_msa_asub_s_h
-#define msa_abdq_u32          __builtin_msa_asub_u_w
-#define msa_abdq_s32          __builtin_msa_asub_s_w
-#define msa_abdq_u64          __builtin_msa_asub_u_d
-#define msa_abdq_s64          __builtin_msa_asub_s_d
-#define msa_abdq_f32(a, b)    msa_absq_f32(__builtin_msa_fsub_w(a, b))
-#define msa_abdq_f64(a, b)    msa_absq_f64(__builtin_msa_fsub_d(a, b))
-#define msa_qabdq_s8(a, b)    msa_qabsq_s8(__builtin_msa_subs_s_b(a, b))
-#define msa_qabdq_s16(a, b)   msa_qabsq_s16(__builtin_msa_subs_s_h(a, b))
-#define msa_qabdq_s32(a, b)   msa_qabsq_s32(__builtin_msa_subs_s_w(a, b))
-#define msa_qabdq_s64(a, b)   msa_qabsq_s64(__builtin_msa_subs_s_d(a, b))
-
-/* sqrtq, rsqrtq */
-#define msa_sqrtq_f32         __builtin_msa_fsqrt_w
-#define msa_sqrtq_f64         __builtin_msa_fsqrt_d
-#define msa_rsqrtq_f32        __builtin_msa_frsqrt_w
-#define msa_rsqrtq_f64        __builtin_msa_frsqrt_d
-
-
-/* mlaq: r = a + b * c; */
-__extension__ extern __inline v4i32
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
-msa_mlaq_s32(v4i32 __a, v4i32 __b, v4i32 __c)
-{
-  __asm__ volatile("maddv.w %w[__a], %w[__b], %w[__c]\n"
-               // Outputs
-               : [__a] "+f"(__a)
-               // Inputs
-               : [__b] "f"(__b), [__c] "f"(__c));
-  return __a;
-}
-
-__extension__ extern __inline v2i64
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
-msa_mlaq_s64(v2i64 __a, v2i64 __b, v2i64 __c)
-{
-  __asm__ volatile("maddv.d %w[__a], %w[__b], %w[__c]\n"
-               // Outputs
-               : [__a] "+f"(__a)
-               // Inputs
-               : [__b] "f"(__b), [__c] "f"(__c));
-  return __a;
-}
-
-__extension__ extern __inline v4f32
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
-msa_mlaq_f32(v4f32 __a, v4f32 __b, v4f32 __c)
-{
-  __asm__ volatile("fmadd.w %w[__a], %w[__b], %w[__c]\n"
-               // Outputs
-               : [__a] "+f"(__a)
-               // Inputs
-               : [__b] "f"(__b), [__c] "f"(__c));
-  return __a;
-}
-
-__extension__ extern __inline v2f64
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
-msa_mlaq_f64(v2f64 __a, v2f64 __b, v2f64 __c)
-{
-  __asm__ volatile("fmadd.d %w[__a], %w[__b], %w[__c]\n"
-               // Outputs
-               : [__a] "+f"(__a)
-               // Inputs
-               : [__b] "f"(__b), [__c] "f"(__c));
-  return __a;
-}
-
-/* cntq */
-#define msa_cntq_s8           __builtin_msa_pcnt_b
-#define msa_cntq_s16          __builtin_msa_pcnt_h
-#define msa_cntq_s32          __builtin_msa_pcnt_w
-#define msa_cntq_s64          __builtin_msa_pcnt_d
-
-/* bslq (a: mask; r = b(if a == 0); r = c(if a == 1);) */
-#define msa_bslq_u8           __builtin_msa_bsel_v
-
-/* ilvrq, ilvlq (For EL only, ilvrq: b0, a0, b1, a1; ilvlq: b2, a2, b3, a3;) */
-#define msa_ilvrq_s8          __builtin_msa_ilvr_b
-#define msa_ilvrq_s16         __builtin_msa_ilvr_h
-#define msa_ilvrq_s32         __builtin_msa_ilvr_w
-#define msa_ilvrq_s64         __builtin_msa_ilvr_d
-#define msa_ilvlq_s8          __builtin_msa_ilvl_b
-#define msa_ilvlq_s16         __builtin_msa_ilvl_h
-#define msa_ilvlq_s32         __builtin_msa_ilvl_w
-#define msa_ilvlq_s64         __builtin_msa_ilvl_d
-
-/* ilvevq, ilvodq (ilvevq: b0, a0, b2, a2; ilvodq: b1, a1, b3, a3; ) */
-#define msa_ilvevq_s8         __builtin_msa_ilvev_b
-#define msa_ilvevq_s16        __builtin_msa_ilvev_h
-#define msa_ilvevq_s32        __builtin_msa_ilvev_w
-#define msa_ilvevq_s64        __builtin_msa_ilvev_d
-#define msa_ilvodq_s8         __builtin_msa_ilvod_b
-#define msa_ilvodq_s16        __builtin_msa_ilvod_h
-#define msa_ilvodq_s32        __builtin_msa_ilvod_w
-#define msa_ilvodq_s64        __builtin_msa_ilvod_d
-
-/* extq (r = (a || b); a concatenation b and get elements from index c) */
-#ifdef _MIPSEB
-#define msa_extq_s8(a, b, c)  \
-(__builtin_msa_vshf_b(__builtin_msa_subv_b((v16i8)((v2i64){0x1716151413121110, 0x1F1E1D1C1B1A1918}), __builtin_msa_fill_b(c)), a, b))
-#define msa_extq_s16(a, b, c) \
-(__builtin_msa_vshf_h(__builtin_msa_subv_h((v8i16)((v2i64){0x000B000A00090008, 0x000F000E000D000C}), __builtin_msa_fill_h(c)), a, b))
-#define msa_extq_s32(a, b, c) \
-(__builtin_msa_vshf_w(__builtin_msa_subv_w((v4i32)((v2i64){0x0000000500000004, 0x0000000700000006}), __builtin_msa_fill_w(c)), a, b))
-#define msa_extq_s64(a, b, c) \
-(__builtin_msa_vshf_d(__builtin_msa_subv_d((v2i64){0x0000000000000002, 0x0000000000000003}, __builtin_msa_fill_d(c)), a, b))
-#else
-#define msa_extq_s8(a, b, c)  \
-(__builtin_msa_vshf_b(__builtin_msa_addv_b((v16i8)((v2i64){0x0706050403020100, 0x0F0E0D0C0B0A0908}), __builtin_msa_fill_b(c)), b, a))
-#define msa_extq_s16(a, b, c) \
-(__builtin_msa_vshf_h(__builtin_msa_addv_h((v8i16)((v2i64){0x0003000200010000, 0x0007000600050004}), __builtin_msa_fill_h(c)), b, a))
-#define msa_extq_s32(a, b, c) \
-(__builtin_msa_vshf_w(__builtin_msa_addv_w((v4i32)((v2i64){0x0000000100000000, 0x0000000300000002}), __builtin_msa_fill_w(c)), b, a))
-#define msa_extq_s64(a, b, c) \
-(__builtin_msa_vshf_d(__builtin_msa_addv_d((v2i64){0x0000000000000000, 0x0000000000000001}, __builtin_msa_fill_d(c)), b, a))
-#endif /* _MIPSEB */
-
-/* cvttruncq, cvttintq, cvtrintq */
-#define msa_cvttruncq_u32_f32 __builtin_msa_ftrunc_u_w
-#define msa_cvttruncq_s32_f32 __builtin_msa_ftrunc_s_w
-#define msa_cvttruncq_u64_f64 __builtin_msa_ftrunc_u_d
-#define msa_cvttruncq_s64_f64 __builtin_msa_ftrunc_s_d
-#define msa_cvttintq_u32_f32  __builtin_msa_ftint_u_w
-#define msa_cvttintq_s32_f32  __builtin_msa_ftint_s_w
-#define msa_cvttintq_u64_f64  __builtin_msa_ftint_u_d
-#define msa_cvttintq_s64_f64  __builtin_msa_ftint_s_d
-#define msa_cvtrintq_f32      __builtin_msa_frint_w
-#define msa_cvtrintq_f64      __builtin_msa_frint_d
-
-/* cvtfintq, cvtfq */
-#define msa_cvtfintq_f32_u32  __builtin_msa_ffint_u_w
-#define msa_cvtfintq_f32_s32  __builtin_msa_ffint_s_w
-#define msa_cvtfintq_f64_u64  __builtin_msa_ffint_u_d
-#define msa_cvtfintq_f64_s64  __builtin_msa_ffint_s_d
-#define msa_cvtfq_f32_f64     __builtin_msa_fexdo_w
-#define msa_cvtflq_f64_f32    __builtin_msa_fexupr_d
-#define msa_cvtfhq_f64_f32    __builtin_msa_fexupl_d
-
-#define msa_addl_u8(a, b)     ((v8u16)__builtin_msa_addv_h((v8i16)V8U8_2_V8I16(a), (v8i16)V8U8_2_V8I16(b)))
-#define msa_addl_s8(a, b)     (__builtin_msa_addv_h((v8i16)V8I8_2_V8I16(a), (v8i16)V8I8_2_V8I16(b)))
-#define msa_addl_u16(a, b)    ((v4u32)__builtin_msa_addv_w((v4i32)V4U16_2_V4I32(a), (v4i32)V4U16_2_V4I32(b)))
-#define msa_addl_s16(a, b)    (__builtin_msa_addv_w((v4i32)V4I16_2_V4I32(a), (v4i32)V4I16_2_V4I32(b)))
-#define msa_subl_s16(a, b)    (__builtin_msa_subv_w((v4i32)V4I16_2_V4I32(a), (v4i32)V4I16_2_V4I32(b)))
-#define msa_recpeq_f32        __builtin_msa_frcp_w
-#define msa_recpsq_f32(a, b)  (__builtin_msa_fsub_w(msa_dupq_n_f32(2.0f), __builtin_msa_fmul_w(a, b)))
-
-#define MSA_INTERLEAVED_IMPL_LOAD2_STORE2(_Tp, _Tpv, _Tpvs, suffix, df, nlanes) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_ld2q_##suffix(const _Tp* ptr, _Tpv* a, _Tpv* b) \
-{ \
-  _Tpv v0 = msa_ld1q_##suffix(ptr); \
-  _Tpv v1 = msa_ld1q_##suffix(ptr + nlanes); \
-  *a = (_Tpv)__builtin_msa_pckev_##df((_Tpvs)v1, (_Tpvs)v0); \
-  *b = (_Tpv)__builtin_msa_pckod_##df((_Tpvs)v1, (_Tpvs)v0); \
-} \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_st2q_##suffix(_Tp* ptr, const _Tpv a, const _Tpv b) \
-{ \
-  msa_st1q_##suffix(ptr, (_Tpv)__builtin_msa_ilvr_##df((_Tpvs)b, (_Tpvs)a)); \
-  msa_st1q_##suffix(ptr + nlanes, (_Tpv)__builtin_msa_ilvl_##df((_Tpvs)b, (_Tpvs)a)); \
-}
-
-MSA_INTERLEAVED_IMPL_LOAD2_STORE2(uint8_t, v16u8, v16i8, u8, b, 16)
-MSA_INTERLEAVED_IMPL_LOAD2_STORE2(int8_t, v16i8, v16i8, s8, b, 16)
-MSA_INTERLEAVED_IMPL_LOAD2_STORE2(uint16_t, v8u16, v8i16, u16, h, 8)
-MSA_INTERLEAVED_IMPL_LOAD2_STORE2(int16_t, v8i16, v8i16, s16, h, 8)
-MSA_INTERLEAVED_IMPL_LOAD2_STORE2(uint32_t, v4u32, v4i32, u32, w, 4)
-MSA_INTERLEAVED_IMPL_LOAD2_STORE2(int32_t, v4i32, v4i32, s32, w, 4)
-MSA_INTERLEAVED_IMPL_LOAD2_STORE2(float, v4f32, v4i32, f32, w, 4)
-MSA_INTERLEAVED_IMPL_LOAD2_STORE2(uint64_t, v2u64, v2i64, u64, d, 2)
-MSA_INTERLEAVED_IMPL_LOAD2_STORE2(int64_t, v2i64, v2i64, s64, d, 2)
-MSA_INTERLEAVED_IMPL_LOAD2_STORE2(double, v2f64, v2i64, f64, d, 2)
-
-#ifdef _MIPSEB
-#define MSA_INTERLEAVED_IMPL_LOAD3_8(_Tp, _Tpv, _Tpvs, suffix) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_ld3q_##suffix(const _Tp* ptr, _Tpv* a, _Tpv* b, _Tpv* c) \
-{ \
-  _Tpv v0 = msa_ld1q_##suffix(ptr); \
-  _Tpv v1 = msa_ld1q_##suffix(ptr + 16); \
-  _Tpv v2 = msa_ld1q_##suffix(ptr + 32); \
-  _Tpvs v3 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x0704011F1F1F1F1F, 0x1F1C191613100D0A}), (_Tpvs)v0, (_Tpvs)v1); \
-  *a = (_Tpv)__builtin_msa_vshf_b((_Tpvs)((v2i64){0x1716150E0B080502, 0x1F1E1D1C1B1A1918}), v3, (_Tpvs)v2); \
-  v3 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x0603001F1F1F1F1F, 0x1E1B1815120F0C09}), (_Tpvs)v0, (_Tpvs)v1); \
-  *b = (_Tpv)__builtin_msa_vshf_b((_Tpvs)((v2i64){0x1716150D0A070401, 0x1F1E1D1C1B1A1918}), v3, (_Tpvs)v2); \
-  v3 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x05021F1F1F1F1F1F, 0x1D1A1714110E0B08}), (_Tpvs)v0, (_Tpvs)v1); \
-  *c = (_Tpv)__builtin_msa_vshf_b((_Tpvs)((v2i64){0x17160F0C09060300, 0x1F1E1D1C1B1A1918}), v3, (_Tpvs)v2); \
-}
-#else
-#define MSA_INTERLEAVED_IMPL_LOAD3_8(_Tp, _Tpv, _Tpvs, suffix) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_ld3q_##suffix(const _Tp* ptr, _Tpv* a, _Tpv* b, _Tpv* c) \
-{ \
-  _Tpv v0 = msa_ld1q_##suffix(ptr); \
-  _Tpv v1 = msa_ld1q_##suffix(ptr + 16); \
-  _Tpv v2 = msa_ld1q_##suffix(ptr + 32); \
-  _Tpvs v3 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x15120F0C09060300, 0x00000000001E1B18}), (_Tpvs)v1, (_Tpvs)v0); \
-  *a = (_Tpv)__builtin_msa_vshf_b((_Tpvs)((v2i64){0x0706050403020100, 0x1D1A1714110A0908}), (_Tpvs)v2, v3); \
-  v3 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x1613100D0A070401, 0x00000000001F1C19}), (_Tpvs)v1, (_Tpvs)v0); \
-  *b = (_Tpv)__builtin_msa_vshf_b((_Tpvs)((v2i64){0x0706050403020100, 0x1E1B1815120A0908}), (_Tpvs)v2, v3); \
-  v3 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x1714110E0B080502, 0x0000000000001D1A}), (_Tpvs)v1, (_Tpvs)v0); \
-  *c = (_Tpv)__builtin_msa_vshf_b((_Tpvs)((v2i64){0x0706050403020100, 0x1F1C191613100908}), (_Tpvs)v2, v3); \
-}
-#endif
-
-MSA_INTERLEAVED_IMPL_LOAD3_8(uint8_t, v16u8, v16i8, u8)
-MSA_INTERLEAVED_IMPL_LOAD3_8(int8_t, v16i8, v16i8, s8)
-
-#ifdef _MIPSEB
-#define MSA_INTERLEAVED_IMPL_LOAD3_16(_Tp, _Tpv, _Tpvs, suffix) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_ld3q_##suffix(const _Tp* ptr, _Tpv* a, _Tpv* b, _Tpv* c) \
-{ \
-  _Tpv v0 = msa_ld1q_##suffix(ptr); \
-  _Tpv v1 = msa_ld1q_##suffix(ptr + 8); \
-  _Tpv v2 = msa_ld1q_##suffix(ptr + 16); \
-  _Tpvs v3 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x00030000000F000F, 0x000F000C00090006}), (_Tpvs)v1, (_Tpvs)v0); \
-  *a = (_Tpv)__builtin_msa_vshf_h((_Tpvs)((v2i64){0x000B000A00050002, 0x000F000E000D000C}), (_Tpvs)v2, v3); \
-  v3 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x0002000F000F000F, 0x000E000B00080005}), (_Tpvs)v1, (_Tpvs)v0); \
-  *b = (_Tpv)__builtin_msa_vshf_h((_Tpvs)((v2i64){0x000B000700040001, 0x000F000E000D000C}), (_Tpvs)v2, v3); \
-  v3 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x0001000F000F000F, 0x000D000A00070004}), (_Tpvs)v1, (_Tpvs)v0); \
-  *c = (_Tpv)__builtin_msa_vshf_h((_Tpvs)((v2i64){0x000B000600030000, 0x000F000E000D000C}), (_Tpvs)v2, v3); \
-}
-#else
-#define MSA_INTERLEAVED_IMPL_LOAD3_16(_Tp, _Tpv, _Tpvs, suffix) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_ld3q_##suffix(const _Tp* ptr, _Tpv* a, _Tpv* b, _Tpv* c) \
-{ \
-  _Tpv v0 = msa_ld1q_##suffix(ptr); \
-  _Tpv v1 = msa_ld1q_##suffix(ptr + 8); \
-  _Tpv v2 = msa_ld1q_##suffix(ptr + 16); \
-  _Tpvs v3 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x0009000600030000, 0x00000000000F000C}), (_Tpvs)v1, (_Tpvs)v0); \
-  *a = (_Tpv)__builtin_msa_vshf_h((_Tpvs)((v2i64){0x0003000200010000, 0x000D000A00050004}), (_Tpvs)v2, v3); \
-  v3 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x000A000700040001, 0x000000000000000D}), (_Tpvs)v1, (_Tpvs)v0); \
-  *b = (_Tpv)__builtin_msa_vshf_h((_Tpvs)((v2i64){0x0003000200010000, 0x000E000B00080004}), (_Tpvs)v2, v3); \
-  v3 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x000B000800050002, 0x000000000000000E}), (_Tpvs)v1, (_Tpvs)v0); \
-  *c = (_Tpv)__builtin_msa_vshf_h((_Tpvs)((v2i64){0x0003000200010000, 0x000F000C00090004}), (_Tpvs)v2, v3); \
-}
-#endif
-
-MSA_INTERLEAVED_IMPL_LOAD3_16(uint16_t, v8u16, v8i16, u16)
-MSA_INTERLEAVED_IMPL_LOAD3_16(int16_t, v8i16, v8i16, s16)
-
-#define MSA_INTERLEAVED_IMPL_LOAD3_32(_Tp, _Tpv, _Tpvs, suffix) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_ld3q_##suffix(const _Tp* ptr, _Tpv* a, _Tpv* b, _Tpv* c) \
-{ \
-  _Tpv v00 = msa_ld1q_##suffix(ptr); \
-  _Tpv v01 = msa_ld1q_##suffix(ptr + 4); \
-  _Tpv v02 = msa_ld1q_##suffix(ptr + 8); \
-  _Tpvs v10 = __builtin_msa_ilvr_w((_Tpvs)__builtin_msa_ilvl_d((v2i64)v01, (v2i64)v01), (_Tpvs)v00); \
-  _Tpvs v11 = __builtin_msa_ilvr_w((_Tpvs)v02, (_Tpvs)__builtin_msa_ilvl_d((v2i64)v00, (v2i64)v00)); \
-  _Tpvs v12 = __builtin_msa_ilvr_w((_Tpvs)__builtin_msa_ilvl_d((v2i64)v02, (v2i64)v02), (_Tpvs)v01); \
-  *a = (_Tpv)__builtin_msa_ilvr_w((_Tpvs)__builtin_msa_ilvl_d((v2i64)v11, (v2i64)v11), v10); \
-  *b = (_Tpv)__builtin_msa_ilvr_w(v12, (_Tpvs)__builtin_msa_ilvl_d((v2i64)v10, (v2i64)v10)); \
-  *c = (_Tpv)__builtin_msa_ilvr_w((_Tpvs)__builtin_msa_ilvl_d((v2i64)v12, (v2i64)v12), v11); \
-}
-
-MSA_INTERLEAVED_IMPL_LOAD3_32(uint32_t, v4u32, v4i32, u32)
-MSA_INTERLEAVED_IMPL_LOAD3_32(int32_t, v4i32, v4i32, s32)
-MSA_INTERLEAVED_IMPL_LOAD3_32(float, v4f32, v4i32, f32)
-
-#define MSA_INTERLEAVED_IMPL_LOAD3_64(_Tp, _Tpv, suffix) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_ld3q_##suffix(const _Tp* ptr, _Tpv* a, _Tpv* b, _Tpv* c) \
-{ \
-  *((_Tp*)a) = *ptr;           *((_Tp*)b) = *(ptr + 1);     *((_Tp*)c) = *(ptr + 2);     \
-  *((_Tp*)a + 1) = *(ptr + 3); *((_Tp*)b + 1) = *(ptr + 4); *((_Tp*)c + 1) = *(ptr + 5); \
-}
-
-MSA_INTERLEAVED_IMPL_LOAD3_64(uint64_t, v2u64, u64)
-MSA_INTERLEAVED_IMPL_LOAD3_64(int64_t, v2i64, s64)
-MSA_INTERLEAVED_IMPL_LOAD3_64(double, v2f64, f64)
-
-#ifdef _MIPSEB
-#define MSA_INTERLEAVED_IMPL_STORE3_8(_Tp, _Tpv, _Tpvs, suffix) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_st3q_##suffix(_Tp* ptr, const _Tpv a, const _Tpv b, const _Tpv c) \
-{ \
-  _Tpvs v0 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x0F0E0D0C0B1F1F1F, 0x1F1E1D1C1B1A1F1F}), (_Tpvs)b, (_Tpvs)a); \
-  _Tpvs v1 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x0D1C140C1B130B1A, 0x1F170F1E160E1D15}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr, (_Tpv)v1); \
-  v0 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x0A09080706051F1F, 0x19181716151F1F1F}), (_Tpvs)b, (_Tpvs)a); \
-  v1 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x1D14071C13061B12, 0x170A1F16091E1508}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr + 16, (_Tpv)v1); \
-  v0 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x04030201001F1F1F, 0x14131211101F1F1F}), (_Tpvs)b, (_Tpvs)a); \
-  v1 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x15021C14011B1300, 0x051F17041E16031D}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr + 32, (_Tpv)v1); \
-}
-#else
-#define MSA_INTERLEAVED_IMPL_STORE3_8(_Tp, _Tpv, _Tpvs, suffix) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_st3q_##suffix(_Tp* ptr, const _Tpv a, const _Tpv b, const _Tpv c) \
-{ \
-  _Tpvs v0 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x0000050403020100, 0x0000001413121110}), (_Tpvs)b, (_Tpvs)a); \
-  _Tpvs v1 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x0A02110901100800, 0x05140C04130B0312}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr, (_Tpv)v1); \
-  v0 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x0000000A09080706, 0x00001A1918171615}), (_Tpvs)b, (_Tpvs)a); \
-  v1 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x170A011609001508, 0x0D04190C03180B02}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr + 16, (_Tpv)v1); \
-  v0 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x0000000F0E0D0C0B, 0x0000001F1E1D1C1B}), (_Tpvs)b, (_Tpvs)a); \
-  v1 = __builtin_msa_vshf_b((_Tpvs)((v2i64){0x021C09011B08001A, 0x1F0C041E0B031D0A}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr + 32, (_Tpv)v1); \
-}
-#endif
-
-MSA_INTERLEAVED_IMPL_STORE3_8(uint8_t, v16u8, v16i8, u8)
-MSA_INTERLEAVED_IMPL_STORE3_8(int8_t, v16i8, v16i8, s8)
-
-#ifdef _MIPSEB
-#define MSA_INTERLEAVED_IMPL_STORE3_16(_Tp, _Tpv, _Tpvs, suffix) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_st3q_##suffix(_Tp* ptr, const _Tpv a, const _Tpv b, const _Tpv c) \
-{ \
-  _Tpvs v0 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x000700060005000F, 0x000F000E000D000F}), (_Tpvs)b, (_Tpvs)a); \
-  _Tpvs v1 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x000A0006000D0009, 0x000F000B0007000E}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr, (_Tpv)v1); \
-  v0 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x00040003000F000F, 0x000C000B000A000F}), (_Tpvs)b, (_Tpvs)a); \
-  v1 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x000E000A0003000D, 0x0005000F000B0004}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr + 8, (_Tpv)v1); \
-  v0 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x000200010000000F, 0x00090008000F000F}), (_Tpvs)b, (_Tpvs)a); \
-  v1 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x0001000E00090000, 0x000B0002000F000A}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr + 16, (_Tpv)v1); \
-}
-#else
-#define MSA_INTERLEAVED_IMPL_STORE3_16(_Tp, _Tpv, _Tpvs, suffix) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_st3q_##suffix(_Tp* ptr, const _Tpv a, const _Tpv b, const _Tpv c) \
-{ \
-  _Tpvs v0 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x0000000200010000, 0x0000000A00090008}), (_Tpvs)b, (_Tpvs)a); \
-  _Tpvs v1 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x0001000800040000, 0x0006000200090005}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr, (_Tpv)v1); \
-  v0 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x0000000500040003, 0x00000000000C000B}), (_Tpvs)b, (_Tpvs)a); \
-  v1 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x000B00040000000A, 0x0002000C00050001}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr + 8, (_Tpv)v1); \
-  v0 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x0000000000070006, 0x0000000F000E000D}), (_Tpvs)b, (_Tpvs)a); \
-  v1 = __builtin_msa_vshf_h((_Tpvs)((v2i64){0x00050000000D0004, 0x000F00060001000E}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr + 16, (_Tpv)v1); \
-}
-#endif
-
-MSA_INTERLEAVED_IMPL_STORE3_16(uint16_t, v8u16, v8i16, u16)
-MSA_INTERLEAVED_IMPL_STORE3_16(int16_t, v8i16, v8i16, s16)
-
-#ifdef _MIPSEB
-#define MSA_INTERLEAVED_IMPL_STORE3_32(_Tp, _Tpv, _Tpvs, suffix) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_st3q_##suffix(_Tp* ptr, const _Tpv a, const _Tpv b, const _Tpv c) \
-{ \
-  _Tpvs v0 = __builtin_msa_vshf_w((_Tpvs)((v2i64){0x0000000300000007, 0x0000000700000006}), (_Tpvs)b, (_Tpvs)a); \
-  _Tpvs v1 = __builtin_msa_vshf_w((_Tpvs)((v2i64){0x0000000300000006, 0x0000000700000005}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr, (_Tpv)v1); \
-  v0 = __builtin_msa_vshf_w((_Tpvs)((v2i64){0x0000000200000001, 0x0000000500000007}), (_Tpvs)b, (_Tpvs)a); \
-  v1 = __builtin_msa_vshf_w((_Tpvs)((v2i64){0x0000000700000004, 0x0000000500000002}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr + 4, (_Tpv)v1); \
-  v0 = __builtin_msa_vshf_w((_Tpvs)((v2i64){0x0000000000000007, 0x0000000400000007}), (_Tpvs)b, (_Tpvs)a); \
-  v1 = __builtin_msa_vshf_w((_Tpvs)((v2i64){0x0000000500000000, 0x0000000100000007}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr + 8, (_Tpv)v1); \
-}
-#else
-#define MSA_INTERLEAVED_IMPL_STORE3_32(_Tp, _Tpv, _Tpvs, suffix) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_st3q_##suffix(_Tp* ptr, const _Tpv a, const _Tpv b, const _Tpv c) \
-{ \
-  _Tpvs v0 = __builtin_msa_vshf_w((_Tpvs)((v2i64){0x0000000100000000, 0x0000000000000004}), (_Tpvs)b, (_Tpvs)a); \
-  _Tpvs v1 = __builtin_msa_vshf_w((_Tpvs)((v2i64){0x0000000200000000, 0x0000000100000004}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr, (_Tpv)v1); \
-  v0 = __builtin_msa_vshf_w((_Tpvs)((v2i64){0x0000000000000002, 0x0000000600000005}), (_Tpvs)b, (_Tpvs)a); \
-  v1 = __builtin_msa_vshf_w((_Tpvs)((v2i64){0x0000000500000002, 0x0000000300000000}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr + 4, (_Tpv)v1); \
-  v0 = __builtin_msa_vshf_w((_Tpvs)((v2i64){0x0000000000000003, 0x0000000000000007}), (_Tpvs)b, (_Tpvs)a); \
-  v1 = __builtin_msa_vshf_w((_Tpvs)((v2i64){0x0000000000000006, 0x0000000700000002}), (_Tpvs)c, (_Tpvs)v0); \
-  msa_st1q_##suffix(ptr + 8, (_Tpv)v1); \
-}
-#endif
-
-MSA_INTERLEAVED_IMPL_STORE3_32(uint32_t, v4u32, v4i32, u32)
-MSA_INTERLEAVED_IMPL_STORE3_32(int32_t, v4i32, v4i32, s32)
-MSA_INTERLEAVED_IMPL_STORE3_32(float, v4f32, v4i32, f32)
-
-#define MSA_INTERLEAVED_IMPL_STORE3_64(_Tp, _Tpv, suffix) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_st3q_##suffix(_Tp* ptr, const _Tpv a, const _Tpv b, const _Tpv c) \
-{ \
-  *ptr = a[0];       *(ptr + 1) = b[0]; *(ptr + 2) = c[0]; \
-  *(ptr + 3) = a[1]; *(ptr + 4) = b[1]; *(ptr + 5) = c[1]; \
-}
-
-MSA_INTERLEAVED_IMPL_STORE3_64(uint64_t, v2u64, u64)
-MSA_INTERLEAVED_IMPL_STORE3_64(int64_t, v2i64, s64)
-MSA_INTERLEAVED_IMPL_STORE3_64(double, v2f64, f64)
-
-#define MSA_INTERLEAVED_IMPL_LOAD4_STORE4(_Tp, _Tpv, _Tpvs, suffix, df, nlanes) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_ld4q_##suffix(const _Tp* ptr, _Tpv* a, _Tpv* b, _Tpv* c, _Tpv* d) \
-{ \
-  _Tpv v0 = msa_ld1q_##suffix(ptr); \
-  _Tpv v1 = msa_ld1q_##suffix(ptr + nlanes); \
-  _Tpv v2 = msa_ld1q_##suffix(ptr + nlanes * 2); \
-  _Tpv v3 = msa_ld1q_##suffix(ptr + nlanes * 3); \
-  _Tpvs t0 = __builtin_msa_pckev_##df((_Tpvs)v1, (_Tpvs)v0); \
-  _Tpvs t1 = __builtin_msa_pckev_##df((_Tpvs)v3, (_Tpvs)v2); \
-  _Tpvs t2 = __builtin_msa_pckod_##df((_Tpvs)v1, (_Tpvs)v0); \
-  _Tpvs t3 = __builtin_msa_pckod_##df((_Tpvs)v3, (_Tpvs)v2); \
-  *a = (_Tpv)__builtin_msa_pckev_##df(t1, t0); \
-  *b = (_Tpv)__builtin_msa_pckev_##df(t3, t2); \
-  *c = (_Tpv)__builtin_msa_pckod_##df(t1, t0); \
-  *d = (_Tpv)__builtin_msa_pckod_##df(t3, t2); \
-} \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_st4q_##suffix(_Tp* ptr, const _Tpv a, const _Tpv b, const _Tpv c, const _Tpv d) \
-{ \
-  _Tpvs v0 = __builtin_msa_ilvr_##df((_Tpvs)c, (_Tpvs)a); \
-  _Tpvs v1 = __builtin_msa_ilvr_##df((_Tpvs)d, (_Tpvs)b); \
-  _Tpvs v2 = __builtin_msa_ilvl_##df((_Tpvs)c, (_Tpvs)a); \
-  _Tpvs v3 = __builtin_msa_ilvl_##df((_Tpvs)d, (_Tpvs)b); \
-  msa_st1q_##suffix(ptr, (_Tpv)__builtin_msa_ilvr_##df(v1, v0)); \
-  msa_st1q_##suffix(ptr + nlanes, (_Tpv)__builtin_msa_ilvl_##df(v1, v0)); \
-  msa_st1q_##suffix(ptr + 2 * nlanes, (_Tpv)__builtin_msa_ilvr_##df(v3, v2)); \
-  msa_st1q_##suffix(ptr + 3 * nlanes, (_Tpv)__builtin_msa_ilvl_##df(v3, v2)); \
-}
-
-MSA_INTERLEAVED_IMPL_LOAD4_STORE4(uint8_t, v16u8, v16i8, u8, b, 16)
-MSA_INTERLEAVED_IMPL_LOAD4_STORE4(int8_t, v16i8, v16i8, s8, b, 16)
-MSA_INTERLEAVED_IMPL_LOAD4_STORE4(uint16_t, v8u16, v8i16, u16, h, 8)
-MSA_INTERLEAVED_IMPL_LOAD4_STORE4(int16_t, v8i16, v8i16, s16, h, 8)
-MSA_INTERLEAVED_IMPL_LOAD4_STORE4(uint32_t, v4u32, v4i32, u32, w, 4)
-MSA_INTERLEAVED_IMPL_LOAD4_STORE4(int32_t, v4i32, v4i32, s32, w, 4)
-MSA_INTERLEAVED_IMPL_LOAD4_STORE4(float, v4f32, v4i32, f32, w, 4)
-
-#define MSA_INTERLEAVED_IMPL_LOAD4_STORE4_64(_Tp, _Tpv, _Tpvs, suffix) \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_ld4q_##suffix(const _Tp* ptr, _Tpv* a, _Tpv* b, _Tpv* c, _Tpv* d) \
-{ \
-  _Tpv v0 = msa_ld1q_##suffix(ptr); \
-  _Tpv v1 = msa_ld1q_##suffix(ptr + 2); \
-  _Tpv v2 = msa_ld1q_##suffix(ptr + 4); \
-  _Tpv v3 = msa_ld1q_##suffix(ptr + 6); \
-  *a = (_Tpv)__builtin_msa_ilvr_d((_Tpvs)v2, (_Tpvs)v0); \
-  *b = (_Tpv)__builtin_msa_ilvl_d((_Tpvs)v2, (_Tpvs)v0); \
-  *c = (_Tpv)__builtin_msa_ilvr_d((_Tpvs)v3, (_Tpvs)v1); \
-  *d = (_Tpv)__builtin_msa_ilvl_d((_Tpvs)v3, (_Tpvs)v1); \
-} \
-__extension__ extern __inline void \
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) \
-msa_st4q_##suffix(_Tp* ptr, const _Tpv a, const _Tpv b, const _Tpv c, const _Tpv d) \
-{ \
-  msa_st1q_##suffix(ptr, (_Tpv)__builtin_msa_ilvr_d((_Tpvs)b, (_Tpvs)a)); \
-  msa_st1q_##suffix(ptr + 2, (_Tpv)__builtin_msa_ilvr_d((_Tpvs)d, (_Tpvs)c)); \
-  msa_st1q_##suffix(ptr + 4, (_Tpv)__builtin_msa_ilvl_d((_Tpvs)b, (_Tpvs)a)); \
-  msa_st1q_##suffix(ptr + 6, (_Tpv)__builtin_msa_ilvl_d((_Tpvs)d, (_Tpvs)c)); \
-}
-
-MSA_INTERLEAVED_IMPL_LOAD4_STORE4_64(uint64_t, v2u64, v2i64, u64)
-MSA_INTERLEAVED_IMPL_LOAD4_STORE4_64(int64_t, v2i64, v2i64, s64)
-MSA_INTERLEAVED_IMPL_LOAD4_STORE4_64(double, v2f64, v2i64, f64)
-
-__extension__ extern __inline v8i16
-__attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
-msa_qdmulhq_n_s16(v8i16 a, int16_t b)
-{
-  v8i16 a_lo, a_hi;
-  ILVRL_H2_SH(a, msa_dupq_n_s16(0), a_lo, a_hi);
-  return msa_packr_s32(msa_shlq_n_s32(msa_mulq_s32(msa_paddlq_s16(a_lo), msa_dupq_n_s32(b)), 1),
-                       msa_shlq_n_s32(msa_mulq_s32(msa_paddlq_s16(a_hi), msa_dupq_n_s32(b)), 1), 16);
-}
-
-#ifdef __cplusplus
-} // extern "C"
-#endif
-
-#endif /*__mips_msa*/
-#endif /* OPENCV_CORE_MSA_MACROS_H */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/hal/simd_utils.impl.hpp b/projects/llm_framework/include/opencv4/opencv2/core/hal/simd_utils.impl.hpp
deleted file mode 100644
index fff8f942..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/hal/simd_utils.impl.hpp
+++ /dev/null
@@ -1,146 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html
-
-// This header is not standalone. Don't include directly, use "intrin.hpp" instead.
-#ifdef OPENCV_HAL_INTRIN_HPP  // defined in intrin.hpp
-
-
-#if CV_SIMD128 || CV_SIMD128_CPP
-
-template<typename _T> struct Type2Vec128_Traits;
-#define CV_INTRIN_DEF_TYPE2VEC128_TRAITS(type_, vec_type_) \
-    template<> struct Type2Vec128_Traits<type_> \
-    { \
-        typedef vec_type_ vec_type; \
-    }
-
-CV_INTRIN_DEF_TYPE2VEC128_TRAITS(uchar, v_uint8x16);
-CV_INTRIN_DEF_TYPE2VEC128_TRAITS(schar, v_int8x16);
-CV_INTRIN_DEF_TYPE2VEC128_TRAITS(ushort, v_uint16x8);
-CV_INTRIN_DEF_TYPE2VEC128_TRAITS(short, v_int16x8);
-CV_INTRIN_DEF_TYPE2VEC128_TRAITS(unsigned, v_uint32x4);
-CV_INTRIN_DEF_TYPE2VEC128_TRAITS(int, v_int32x4);
-CV_INTRIN_DEF_TYPE2VEC128_TRAITS(float, v_float32x4);
-CV_INTRIN_DEF_TYPE2VEC128_TRAITS(uint64, v_uint64x2);
-CV_INTRIN_DEF_TYPE2VEC128_TRAITS(int64, v_int64x2);
-#if CV_SIMD128_64F
-CV_INTRIN_DEF_TYPE2VEC128_TRAITS(double, v_float64x2);
-#endif
-
-template<typename _T> static inline
-typename Type2Vec128_Traits<_T>::vec_type v_setall(const _T& a);
-
-template<> inline Type2Vec128_Traits< uchar>::vec_type v_setall< uchar>(const  uchar& a) { return v_setall_u8(a); }
-template<> inline Type2Vec128_Traits< schar>::vec_type v_setall< schar>(const  schar& a) { return v_setall_s8(a); }
-template<> inline Type2Vec128_Traits<ushort>::vec_type v_setall<ushort>(const ushort& a) { return v_setall_u16(a); }
-template<> inline Type2Vec128_Traits< short>::vec_type v_setall< short>(const  short& a) { return v_setall_s16(a); }
-template<> inline Type2Vec128_Traits<  uint>::vec_type v_setall<  uint>(const   uint& a) { return v_setall_u32(a); }
-template<> inline Type2Vec128_Traits<   int>::vec_type v_setall<   int>(const    int& a) { return v_setall_s32(a); }
-template<> inline Type2Vec128_Traits<uint64>::vec_type v_setall<uint64>(const uint64& a) { return v_setall_u64(a); }
-template<> inline Type2Vec128_Traits< int64>::vec_type v_setall< int64>(const  int64& a) { return v_setall_s64(a); }
-template<> inline Type2Vec128_Traits< float>::vec_type v_setall< float>(const  float& a) { return v_setall_f32(a); }
-#if CV_SIMD128_64F
-template<> inline Type2Vec128_Traits<double>::vec_type v_setall<double>(const double& a) { return v_setall_f64(a); }
-#endif
-
-#endif  // SIMD128
-
-
-#if CV_SIMD256
-
-template<typename _T> struct Type2Vec256_Traits;
-#define CV_INTRIN_DEF_TYPE2VEC256_TRAITS(type_, vec_type_) \
-    template<> struct Type2Vec256_Traits<type_> \
-    { \
-        typedef vec_type_ vec_type; \
-    }
-
-CV_INTRIN_DEF_TYPE2VEC256_TRAITS(uchar, v_uint8x32);
-CV_INTRIN_DEF_TYPE2VEC256_TRAITS(schar, v_int8x32);
-CV_INTRIN_DEF_TYPE2VEC256_TRAITS(ushort, v_uint16x16);
-CV_INTRIN_DEF_TYPE2VEC256_TRAITS(short, v_int16x16);
-CV_INTRIN_DEF_TYPE2VEC256_TRAITS(unsigned, v_uint32x8);
-CV_INTRIN_DEF_TYPE2VEC256_TRAITS(int, v_int32x8);
-CV_INTRIN_DEF_TYPE2VEC256_TRAITS(float, v_float32x8);
-CV_INTRIN_DEF_TYPE2VEC256_TRAITS(uint64, v_uint64x4);
-CV_INTRIN_DEF_TYPE2VEC256_TRAITS(int64, v_int64x4);
-#if CV_SIMD256_64F
-CV_INTRIN_DEF_TYPE2VEC256_TRAITS(double, v_float64x4);
-#endif
-
-template<typename _T> static inline
-typename Type2Vec256_Traits<_T>::vec_type v256_setall(const _T& a);
-
-template<> inline Type2Vec256_Traits< uchar>::vec_type v256_setall< uchar>(const  uchar& a) { return v256_setall_u8(a); }
-template<> inline Type2Vec256_Traits< schar>::vec_type v256_setall< schar>(const  schar& a) { return v256_setall_s8(a); }
-template<> inline Type2Vec256_Traits<ushort>::vec_type v256_setall<ushort>(const ushort& a) { return v256_setall_u16(a); }
-template<> inline Type2Vec256_Traits< short>::vec_type v256_setall< short>(const  short& a) { return v256_setall_s16(a); }
-template<> inline Type2Vec256_Traits<  uint>::vec_type v256_setall<  uint>(const   uint& a) { return v256_setall_u32(a); }
-template<> inline Type2Vec256_Traits<   int>::vec_type v256_setall<   int>(const    int& a) { return v256_setall_s32(a); }
-template<> inline Type2Vec256_Traits<uint64>::vec_type v256_setall<uint64>(const uint64& a) { return v256_setall_u64(a); }
-template<> inline Type2Vec256_Traits< int64>::vec_type v256_setall< int64>(const  int64& a) { return v256_setall_s64(a); }
-template<> inline Type2Vec256_Traits< float>::vec_type v256_setall< float>(const  float& a) { return v256_setall_f32(a); }
-#if CV_SIMD256_64F
-template<> inline Type2Vec256_Traits<double>::vec_type v256_setall<double>(const double& a) { return v256_setall_f64(a); }
-#endif
-
-#endif  // SIMD256
-
-
-#if CV_SIMD512
-
-template<typename _T> struct Type2Vec512_Traits;
-#define CV_INTRIN_DEF_TYPE2VEC512_TRAITS(type_, vec_type_) \
-    template<> struct Type2Vec512_Traits<type_> \
-    { \
-        typedef vec_type_ vec_type; \
-    }
-
-CV_INTRIN_DEF_TYPE2VEC512_TRAITS(uchar, v_uint8x64);
-CV_INTRIN_DEF_TYPE2VEC512_TRAITS(schar, v_int8x64);
-CV_INTRIN_DEF_TYPE2VEC512_TRAITS(ushort, v_uint16x32);
-CV_INTRIN_DEF_TYPE2VEC512_TRAITS(short, v_int16x32);
-CV_INTRIN_DEF_TYPE2VEC512_TRAITS(unsigned, v_uint32x16);
-CV_INTRIN_DEF_TYPE2VEC512_TRAITS(int, v_int32x16);
-CV_INTRIN_DEF_TYPE2VEC512_TRAITS(float, v_float32x16);
-CV_INTRIN_DEF_TYPE2VEC512_TRAITS(uint64, v_uint64x8);
-CV_INTRIN_DEF_TYPE2VEC512_TRAITS(int64, v_int64x8);
-#if CV_SIMD512_64F
-CV_INTRIN_DEF_TYPE2VEC512_TRAITS(double, v_float64x8);
-#endif
-
-template<typename _T> static inline
-typename Type2Vec512_Traits<_T>::vec_type v512_setall(const _T& a);
-
-template<> inline Type2Vec512_Traits< uchar>::vec_type v512_setall< uchar>(const  uchar& a) { return v512_setall_u8(a); }
-template<> inline Type2Vec512_Traits< schar>::vec_type v512_setall< schar>(const  schar& a) { return v512_setall_s8(a); }
-template<> inline Type2Vec512_Traits<ushort>::vec_type v512_setall<ushort>(const ushort& a) { return v512_setall_u16(a); }
-template<> inline Type2Vec512_Traits< short>::vec_type v512_setall< short>(const  short& a) { return v512_setall_s16(a); }
-template<> inline Type2Vec512_Traits<  uint>::vec_type v512_setall<  uint>(const   uint& a) { return v512_setall_u32(a); }
-template<> inline Type2Vec512_Traits<   int>::vec_type v512_setall<   int>(const    int& a) { return v512_setall_s32(a); }
-template<> inline Type2Vec512_Traits<uint64>::vec_type v512_setall<uint64>(const uint64& a) { return v512_setall_u64(a); }
-template<> inline Type2Vec512_Traits< int64>::vec_type v512_setall< int64>(const  int64& a) { return v512_setall_s64(a); }
-template<> inline Type2Vec512_Traits< float>::vec_type v512_setall< float>(const  float& a) { return v512_setall_f32(a); }
-#if CV_SIMD512_64F
-template<> inline Type2Vec512_Traits<double>::vec_type v512_setall<double>(const double& a) { return v512_setall_f64(a); }
-#endif
-
-#endif  // SIMD512
-
-
-#if CV_SIMD_WIDTH == 16
-template<typename _T> static inline
-typename Type2Vec128_Traits<_T>::vec_type vx_setall(const _T& a) { return v_setall(a); }
-#elif CV_SIMD_WIDTH == 32
-template<typename _T> static inline
-typename Type2Vec256_Traits<_T>::vec_type vx_setall(const _T& a) { return v256_setall(a); }
-#elif CV_SIMD_WIDTH == 64
-template<typename _T> static inline
-typename Type2Vec512_Traits<_T>::vec_type vx_setall(const _T& a) { return v512_setall(a); }
-#else
-#error "Build configuration error, unsupported CV_SIMD_WIDTH"
-#endif
-
-
-#endif  // OPENCV_HAL_INTRIN_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/mat.hpp b/projects/llm_framework/include/opencv4/opencv2/core/mat.hpp
deleted file mode 100644
index 1ef0eb5a..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/mat.hpp
+++ /dev/null
@@ -1,3766 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_MAT_HPP
-#define OPENCV_CORE_MAT_HPP
-
-#ifndef __cplusplus
-#  error mat.hpp header must be compiled as C++
-#endif
-
-#include "opencv2/core/matx.hpp"
-#include "opencv2/core/types.hpp"
-
-#include "opencv2/core/bufferpool.hpp"
-
-#include <type_traits>
-
-namespace cv
-{
-
-//! @addtogroup core_basic
-//! @{
-
-enum AccessFlag { ACCESS_READ=1<<24, ACCESS_WRITE=1<<25,
-    ACCESS_RW=3<<24, ACCESS_MASK=ACCESS_RW, ACCESS_FAST=1<<26 };
-CV_ENUM_FLAGS(AccessFlag)
-__CV_ENUM_FLAGS_BITWISE_AND(AccessFlag, int, AccessFlag)
-
-CV__DEBUG_NS_BEGIN
-
-class CV_EXPORTS _OutputArray;
-
-//////////////////////// Input/Output Array Arguments /////////////////////////////////
-
-/** @brief This is the proxy class for passing read-only input arrays into OpenCV functions.
-
-It is defined as:
-@code
-    typedef const _InputArray& InputArray;
-@endcode
-where _InputArray is a class that can be constructed from `Mat`, `Mat_<T>`, `Matx<T, m, n>`,
-`std::vector<T>`, `std::vector<std::vector<T> >`, `std::vector<Mat>`, `std::vector<Mat_<T> >`,
-`UMat`, `std::vector<UMat>` or `double`. It can also be constructed from a matrix expression.
-
-Since this is mostly implementation-level class, and its interface may change in future versions, we
-do not describe it in details. There are a few key things, though, that should be kept in mind:
-
--   When you see in the reference manual or in OpenCV source code a function that takes
-    InputArray, it means that you can actually pass `Mat`, `Matx`, `vector<T>` etc. (see above the
-    complete list).
--   Optional input arguments: If some of the input arrays may be empty, pass cv::noArray() (or
-    simply cv::Mat() as you probably did before).
--   The class is designed solely for passing parameters. That is, normally you *should not*
-    declare class members, local and global variables of this type.
--   If you want to design your own function or a class method that can operate of arrays of
-    multiple types, you can use InputArray (or OutputArray) for the respective parameters. Inside
-    a function you should use _InputArray::getMat() method to construct a matrix header for the
-    array (without copying data). _InputArray::kind() can be used to distinguish Mat from
-    `vector<>` etc., but normally it is not needed.
-
-Here is how you can use a function that takes InputArray :
-@code
-    std::vector<Point2f> vec;
-    // points or a circle
-    for( int i = 0; i < 30; i++ )
-        vec.push_back(Point2f((float)(100 + 30*cos(i*CV_PI*2/5)),
-                              (float)(100 - 30*sin(i*CV_PI*2/5))));
-    cv::transform(vec, vec, cv::Matx23f(0.707, -0.707, 10, 0.707, 0.707, 20));
-@endcode
-That is, we form an STL vector containing points, and apply in-place affine transformation to the
-vector using the 2x3 matrix created inline as `Matx<float, 2, 3>` instance.
-
-Here is how such a function can be implemented (for simplicity, we implement a very specific case of
-it, according to the assertion statement inside) :
-@code
-    void myAffineTransform(InputArray _src, OutputArray _dst, InputArray _m)
-    {
-        // get Mat headers for input arrays. This is O(1) operation,
-        // unless _src and/or _m are matrix expressions.
-        Mat src = _src.getMat(), m = _m.getMat();
-        CV_Assert( src.type() == CV_32FC2 && m.type() == CV_32F && m.size() == Size(3, 2) );
-
-        // [re]create the output array so that it has the proper size and type.
-        // In case of Mat it calls Mat::create, in case of STL vector it calls vector::resize.
-        _dst.create(src.size(), src.type());
-        Mat dst = _dst.getMat();
-
-        for( int i = 0; i < src.rows; i++ )
-            for( int j = 0; j < src.cols; j++ )
-            {
-                Point2f pt = src.at<Point2f>(i, j);
-                dst.at<Point2f>(i, j) = Point2f(m.at<float>(0, 0)*pt.x +
-                                                m.at<float>(0, 1)*pt.y +
-                                                m.at<float>(0, 2),
-                                                m.at<float>(1, 0)*pt.x +
-                                                m.at<float>(1, 1)*pt.y +
-                                                m.at<float>(1, 2));
-            }
-    }
-@endcode
-There is another related type, InputArrayOfArrays, which is currently defined as a synonym for
-InputArray:
-@code
-    typedef InputArray InputArrayOfArrays;
-@endcode
-It denotes function arguments that are either vectors of vectors or vectors of matrices. A separate
-synonym is needed to generate Python/Java etc. wrappers properly. At the function implementation
-level their use is similar, but _InputArray::getMat(idx) should be used to get header for the
-idx-th component of the outer vector and _InputArray::size().area() should be used to find the
-number of components (vectors/matrices) of the outer vector.
-
-In general, type support is limited to cv::Mat types. Other types are forbidden.
-But in some cases we need to support passing of custom non-general Mat types, like arrays of cv::KeyPoint, cv::DMatch, etc.
-This data is not intended to be interpreted as an image data, or processed somehow like regular cv::Mat.
-To pass such custom type use rawIn() / rawOut() / rawInOut() wrappers.
-Custom type is wrapped as Mat-compatible `CV_8UC<N>` values (N = sizeof(T), N <= CV_CN_MAX).
- */
-class CV_EXPORTS _InputArray
-{
-public:
-    enum KindFlag {
-        KIND_SHIFT = 16,
-        FIXED_TYPE = 0x8000 << KIND_SHIFT,
-        FIXED_SIZE = 0x4000 << KIND_SHIFT,
-        KIND_MASK = 31 << KIND_SHIFT,
-
-        NONE              = 0 << KIND_SHIFT,
-        MAT               = 1 << KIND_SHIFT,
-        MATX              = 2 << KIND_SHIFT,
-        STD_VECTOR        = 3 << KIND_SHIFT,
-        STD_VECTOR_VECTOR = 4 << KIND_SHIFT,
-        STD_VECTOR_MAT    = 5 << KIND_SHIFT,
-#if OPENCV_ABI_COMPATIBILITY < 500
-        EXPR              = 6 << KIND_SHIFT,  //!< removed: https://github.com/opencv/opencv/pull/17046
-#endif
-        OPENGL_BUFFER     = 7 << KIND_SHIFT,
-        CUDA_HOST_MEM     = 8 << KIND_SHIFT,
-        CUDA_GPU_MAT      = 9 << KIND_SHIFT,
-        UMAT              =10 << KIND_SHIFT,
-        STD_VECTOR_UMAT   =11 << KIND_SHIFT,
-        STD_BOOL_VECTOR   =12 << KIND_SHIFT,
-        STD_VECTOR_CUDA_GPU_MAT = 13 << KIND_SHIFT,
-#if OPENCV_ABI_COMPATIBILITY < 500
-        STD_ARRAY         =14 << KIND_SHIFT,  //!< removed: https://github.com/opencv/opencv/issues/18897
-#endif
-        STD_ARRAY_MAT     =15 << KIND_SHIFT
-    };
-
-    _InputArray();
-    _InputArray(int _flags, void* _obj);
-    _InputArray(const Mat& m);
-    _InputArray(const MatExpr& expr);
-    _InputArray(const std::vector<Mat>& vec);
-    template<typename _Tp> _InputArray(const Mat_<_Tp>& m);
-    template<typename _Tp> _InputArray(const std::vector<_Tp>& vec);
-    _InputArray(const std::vector<bool>& vec);
-    template<typename _Tp> _InputArray(const std::vector<std::vector<_Tp> >& vec);
-    _InputArray(const std::vector<std::vector<bool> >&) = delete;  // not supported
-    template<typename _Tp> _InputArray(const std::vector<Mat_<_Tp> >& vec);
-    template<typename _Tp> _InputArray(const _Tp* vec, int n);
-    template<typename _Tp, int m, int n> _InputArray(const Matx<_Tp, m, n>& matx);
-    _InputArray(const double& val);
-    _InputArray(const cuda::GpuMat& d_mat);
-    _InputArray(const std::vector<cuda::GpuMat>& d_mat_array);
-    _InputArray(const ogl::Buffer& buf);
-    _InputArray(const cuda::HostMem& cuda_mem);
-    template<typename _Tp> _InputArray(const cudev::GpuMat_<_Tp>& m);
-    _InputArray(const UMat& um);
-    _InputArray(const std::vector<UMat>& umv);
-
-    template<typename _Tp, std::size_t _Nm> _InputArray(const std::array<_Tp, _Nm>& arr);
-    template<std::size_t _Nm> _InputArray(const std::array<Mat, _Nm>& arr);
-
-    template<typename _Tp> static _InputArray rawIn(const std::vector<_Tp>& vec);
-    template<typename _Tp, std::size_t _Nm> static _InputArray rawIn(const std::array<_Tp, _Nm>& arr);
-
-    Mat getMat(int idx=-1) const;
-    Mat getMat_(int idx=-1) const;
-    UMat getUMat(int idx=-1) const;
-    void getMatVector(std::vector<Mat>& mv) const;
-    void getUMatVector(std::vector<UMat>& umv) const;
-    void getGpuMatVector(std::vector<cuda::GpuMat>& gpumv) const;
-    cuda::GpuMat getGpuMat() const;
-    ogl::Buffer getOGlBuffer() const;
-
-    int getFlags() const;
-    void* getObj() const;
-    Size getSz() const;
-
-    _InputArray::KindFlag kind() const;
-    int dims(int i=-1) const;
-    int cols(int i=-1) const;
-    int rows(int i=-1) const;
-    Size size(int i=-1) const;
-    int sizend(int* sz, int i=-1) const;
-    bool sameSize(const _InputArray& arr) const;
-    size_t total(int i=-1) const;
-    int type(int i=-1) const;
-    int depth(int i=-1) const;
-    int channels(int i=-1) const;
-    bool isContinuous(int i=-1) const;
-    bool isSubmatrix(int i=-1) const;
-    bool empty() const;
-    void copyTo(const _OutputArray& arr) const;
-    void copyTo(const _OutputArray& arr, const _InputArray & mask) const;
-    size_t offset(int i=-1) const;
-    size_t step(int i=-1) const;
-    bool isMat() const;
-    bool isUMat() const;
-    bool isMatVector() const;
-    bool isUMatVector() const;
-    bool isMatx() const;
-    bool isVector() const;
-    bool isGpuMat() const;
-    bool isGpuMatVector() const;
-    ~_InputArray();
-
-protected:
-    int flags;
-    void* obj;
-    Size sz;
-
-    void init(int _flags, const void* _obj);
-    void init(int _flags, const void* _obj, Size _sz);
-};
-CV_ENUM_FLAGS(_InputArray::KindFlag)
-__CV_ENUM_FLAGS_BITWISE_AND(_InputArray::KindFlag, int, _InputArray::KindFlag)
-
-/** @brief This type is very similar to InputArray except that it is used for input/output and output function
-parameters.
-
-Just like with InputArray, OpenCV users should not care about OutputArray, they just pass `Mat`,
-`vector<T>` etc. to the functions. The same limitation as for `InputArray`: *Do not explicitly
-create OutputArray instances* applies here too.
-
-If you want to make your function polymorphic (i.e. accept different arrays as output parameters),
-it is also not very difficult. Take the sample above as the reference. Note that
-_OutputArray::create() needs to be called before _OutputArray::getMat(). This way you guarantee
-that the output array is properly allocated.
-
-Optional output parameters. If you do not need certain output array to be computed and returned to
-you, pass cv::noArray(), just like you would in the case of optional input array. At the
-implementation level, use _OutputArray::needed() to check if certain output array needs to be
-computed or not.
-
-There are several synonyms for OutputArray that are used to assist automatic Python/Java/... wrapper
-generators:
-@code
-    typedef OutputArray OutputArrayOfArrays;
-    typedef OutputArray InputOutputArray;
-    typedef OutputArray InputOutputArrayOfArrays;
-@endcode
- */
-class CV_EXPORTS _OutputArray : public _InputArray
-{
-public:
-    enum DepthMask
-    {
-        DEPTH_MASK_8U = 1 << CV_8U,
-        DEPTH_MASK_8S = 1 << CV_8S,
-        DEPTH_MASK_16U = 1 << CV_16U,
-        DEPTH_MASK_16S = 1 << CV_16S,
-        DEPTH_MASK_32S = 1 << CV_32S,
-        DEPTH_MASK_32F = 1 << CV_32F,
-        DEPTH_MASK_64F = 1 << CV_64F,
-        DEPTH_MASK_16F = 1 << CV_16F,
-        DEPTH_MASK_ALL = (DEPTH_MASK_64F<<1)-1,
-        DEPTH_MASK_ALL_BUT_8S = DEPTH_MASK_ALL & ~DEPTH_MASK_8S,
-        DEPTH_MASK_ALL_16F = (DEPTH_MASK_16F<<1)-1,
-        DEPTH_MASK_FLT = DEPTH_MASK_32F + DEPTH_MASK_64F
-    };
-
-    _OutputArray();
-    _OutputArray(int _flags, void* _obj);
-    _OutputArray(Mat& m);
-    _OutputArray(std::vector<Mat>& vec);
-    _OutputArray(cuda::GpuMat& d_mat);
-    _OutputArray(std::vector<cuda::GpuMat>& d_mat);
-    _OutputArray(ogl::Buffer& buf);
-    _OutputArray(cuda::HostMem& cuda_mem);
-    template<typename _Tp> _OutputArray(cudev::GpuMat_<_Tp>& m);
-    template<typename _Tp> _OutputArray(std::vector<_Tp>& vec);
-    _OutputArray(std::vector<bool>& vec) = delete;  // not supported
-    template<typename _Tp> _OutputArray(std::vector<std::vector<_Tp> >& vec);
-    _OutputArray(std::vector<std::vector<bool> >&) = delete;  // not supported
-    template<typename _Tp> _OutputArray(std::vector<Mat_<_Tp> >& vec);
-    template<typename _Tp> _OutputArray(Mat_<_Tp>& m);
-    template<typename _Tp> _OutputArray(_Tp* vec, int n);
-    template<typename _Tp, int m, int n> _OutputArray(Matx<_Tp, m, n>& matx);
-    _OutputArray(UMat& m);
-    _OutputArray(std::vector<UMat>& vec);
-
-    _OutputArray(const Mat& m);
-    _OutputArray(const std::vector<Mat>& vec);
-    _OutputArray(const cuda::GpuMat& d_mat);
-    _OutputArray(const std::vector<cuda::GpuMat>& d_mat);
-    _OutputArray(const ogl::Buffer& buf);
-    _OutputArray(const cuda::HostMem& cuda_mem);
-    template<typename _Tp> _OutputArray(const cudev::GpuMat_<_Tp>& m);
-    template<typename _Tp> _OutputArray(const std::vector<_Tp>& vec);
-    template<typename _Tp> _OutputArray(const std::vector<std::vector<_Tp> >& vec);
-    template<typename _Tp> _OutputArray(const std::vector<Mat_<_Tp> >& vec);
-    template<typename _Tp> _OutputArray(const Mat_<_Tp>& m);
-    template<typename _Tp> _OutputArray(const _Tp* vec, int n);
-    template<typename _Tp, int m, int n> _OutputArray(const Matx<_Tp, m, n>& matx);
-    _OutputArray(const UMat& m);
-    _OutputArray(const std::vector<UMat>& vec);
-
-    template<typename _Tp, std::size_t _Nm> _OutputArray(std::array<_Tp, _Nm>& arr);
-    template<typename _Tp, std::size_t _Nm> _OutputArray(const std::array<_Tp, _Nm>& arr);
-    template<std::size_t _Nm> _OutputArray(std::array<Mat, _Nm>& arr);
-    template<std::size_t _Nm> _OutputArray(const std::array<Mat, _Nm>& arr);
-
-    template<typename _Tp> static _OutputArray rawOut(std::vector<_Tp>& vec);
-    template<typename _Tp, std::size_t _Nm> static _OutputArray rawOut(std::array<_Tp, _Nm>& arr);
-
-    bool fixedSize() const;
-    bool fixedType() const;
-    bool needed() const;
-    Mat& getMatRef(int i=-1) const;
-    UMat& getUMatRef(int i=-1) const;
-    cuda::GpuMat& getGpuMatRef() const;
-    std::vector<cuda::GpuMat>& getGpuMatVecRef() const;
-    ogl::Buffer& getOGlBufferRef() const;
-    cuda::HostMem& getHostMemRef() const;
-    void create(Size sz, int type, int i=-1, bool allowTransposed=false, _OutputArray::DepthMask fixedDepthMask=static_cast<_OutputArray::DepthMask>(0)) const;
-    void create(int rows, int cols, int type, int i=-1, bool allowTransposed=false, _OutputArray::DepthMask fixedDepthMask=static_cast<_OutputArray::DepthMask>(0)) const;
-    void create(int dims, const int* size, int type, int i=-1, bool allowTransposed=false, _OutputArray::DepthMask fixedDepthMask=static_cast<_OutputArray::DepthMask>(0)) const;
-    void createSameSize(const _InputArray& arr, int mtype) const;
-    void release() const;
-    void clear() const;
-    void setTo(const _InputArray& value, const _InputArray & mask = _InputArray()) const;
-
-    void assign(const UMat& u) const;
-    void assign(const Mat& m) const;
-
-    void assign(const std::vector<UMat>& v) const;
-    void assign(const std::vector<Mat>& v) const;
-
-    void move(UMat& u) const;
-    void move(Mat& m) const;
-};
-
-
-class CV_EXPORTS _InputOutputArray : public _OutputArray
-{
-public:
-    _InputOutputArray();
-    _InputOutputArray(int _flags, void* _obj);
-    _InputOutputArray(Mat& m);
-    _InputOutputArray(std::vector<Mat>& vec);
-    _InputOutputArray(cuda::GpuMat& d_mat);
-    _InputOutputArray(ogl::Buffer& buf);
-    _InputOutputArray(cuda::HostMem& cuda_mem);
-    template<typename _Tp> _InputOutputArray(cudev::GpuMat_<_Tp>& m);
-    template<typename _Tp> _InputOutputArray(std::vector<_Tp>& vec);
-    _InputOutputArray(std::vector<bool>& vec) = delete;  // not supported
-    template<typename _Tp> _InputOutputArray(std::vector<std::vector<_Tp> >& vec);
-    template<typename _Tp> _InputOutputArray(std::vector<Mat_<_Tp> >& vec);
-    template<typename _Tp> _InputOutputArray(Mat_<_Tp>& m);
-    template<typename _Tp> _InputOutputArray(_Tp* vec, int n);
-    template<typename _Tp, int m, int n> _InputOutputArray(Matx<_Tp, m, n>& matx);
-    _InputOutputArray(UMat& m);
-    _InputOutputArray(std::vector<UMat>& vec);
-
-    _InputOutputArray(const Mat& m);
-    _InputOutputArray(const std::vector<Mat>& vec);
-    _InputOutputArray(const cuda::GpuMat& d_mat);
-    _InputOutputArray(const std::vector<cuda::GpuMat>& d_mat);
-    _InputOutputArray(const ogl::Buffer& buf);
-    _InputOutputArray(const cuda::HostMem& cuda_mem);
-    template<typename _Tp> _InputOutputArray(const cudev::GpuMat_<_Tp>& m);
-    template<typename _Tp> _InputOutputArray(const std::vector<_Tp>& vec);
-    template<typename _Tp> _InputOutputArray(const std::vector<std::vector<_Tp> >& vec);
-    template<typename _Tp> _InputOutputArray(const std::vector<Mat_<_Tp> >& vec);
-    template<typename _Tp> _InputOutputArray(const Mat_<_Tp>& m);
-    template<typename _Tp> _InputOutputArray(const _Tp* vec, int n);
-    template<typename _Tp, int m, int n> _InputOutputArray(const Matx<_Tp, m, n>& matx);
-    _InputOutputArray(const UMat& m);
-    _InputOutputArray(const std::vector<UMat>& vec);
-
-    template<typename _Tp, std::size_t _Nm> _InputOutputArray(std::array<_Tp, _Nm>& arr);
-    template<typename _Tp, std::size_t _Nm> _InputOutputArray(const std::array<_Tp, _Nm>& arr);
-    template<std::size_t _Nm> _InputOutputArray(std::array<Mat, _Nm>& arr);
-    template<std::size_t _Nm> _InputOutputArray(const std::array<Mat, _Nm>& arr);
-
-    template<typename _Tp> static _InputOutputArray rawInOut(std::vector<_Tp>& vec);
-    template<typename _Tp, std::size_t _Nm> _InputOutputArray rawInOut(std::array<_Tp, _Nm>& arr);
-
-};
-
-/** Helper to wrap custom types. @see InputArray */
-template<typename _Tp> static inline _InputArray rawIn(_Tp& v);
-/** Helper to wrap custom types. @see InputArray */
-template<typename _Tp> static inline _OutputArray rawOut(_Tp& v);
-/** Helper to wrap custom types. @see InputArray */
-template<typename _Tp> static inline _InputOutputArray rawInOut(_Tp& v);
-
-CV__DEBUG_NS_END
-
-typedef const _InputArray& InputArray;
-typedef InputArray InputArrayOfArrays;
-typedef const _OutputArray& OutputArray;
-typedef OutputArray OutputArrayOfArrays;
-typedef const _InputOutputArray& InputOutputArray;
-typedef InputOutputArray InputOutputArrayOfArrays;
-
-CV_EXPORTS InputOutputArray noArray();
-
-/////////////////////////////////// MatAllocator //////////////////////////////////////
-
-//! Usage flags for allocator
-enum UMatUsageFlags
-{
-    USAGE_DEFAULT = 0,
-
-    // buffer allocation policy is platform and usage specific
-    USAGE_ALLOCATE_HOST_MEMORY = 1 << 0,
-    USAGE_ALLOCATE_DEVICE_MEMORY = 1 << 1,
-    USAGE_ALLOCATE_SHARED_MEMORY = 1 << 2, // It is not equal to: USAGE_ALLOCATE_HOST_MEMORY | USAGE_ALLOCATE_DEVICE_MEMORY
-
-    __UMAT_USAGE_FLAGS_32BIT = 0x7fffffff // Binary compatibility hint
-};
-
-struct CV_EXPORTS UMatData;
-
-/** @brief  Custom array allocator
-*/
-class CV_EXPORTS MatAllocator
-{
-public:
-    MatAllocator() {}
-    virtual ~MatAllocator() {}
-
-    // let's comment it off for now to detect and fix all the uses of allocator
-    //virtual void allocate(int dims, const int* sizes, int type, int*& refcount,
-    //                      uchar*& datastart, uchar*& data, size_t* step) = 0;
-    //virtual void deallocate(int* refcount, uchar* datastart, uchar* data) = 0;
-    virtual UMatData* allocate(int dims, const int* sizes, int type,
-                               void* data, size_t* step, AccessFlag flags, UMatUsageFlags usageFlags) const = 0;
-    virtual bool allocate(UMatData* data, AccessFlag accessflags, UMatUsageFlags usageFlags) const = 0;
-    virtual void deallocate(UMatData* data) const = 0;
-    virtual void map(UMatData* data, AccessFlag accessflags) const;
-    virtual void unmap(UMatData* data) const;
-    virtual void download(UMatData* data, void* dst, int dims, const size_t sz[],
-                          const size_t srcofs[], const size_t srcstep[],
-                          const size_t dststep[]) const;
-    virtual void upload(UMatData* data, const void* src, int dims, const size_t sz[],
-                        const size_t dstofs[], const size_t dststep[],
-                        const size_t srcstep[]) const;
-    virtual void copy(UMatData* srcdata, UMatData* dstdata, int dims, const size_t sz[],
-                      const size_t srcofs[], const size_t srcstep[],
-                      const size_t dstofs[], const size_t dststep[], bool sync) const;
-
-    // default implementation returns DummyBufferPoolController
-    virtual BufferPoolController* getBufferPoolController(const char* id = NULL) const;
-};
-
-
-//////////////////////////////// MatCommaInitializer //////////////////////////////////
-
-/** @brief  Comma-separated Matrix Initializer
-
- The class instances are usually not created explicitly.
- Instead, they are created on "matrix << firstValue" operator.
-
- The sample below initializes 2x2 rotation matrix:
-
- \code
- double angle = 30, a = cos(angle*CV_PI/180), b = sin(angle*CV_PI/180);
- Mat R = (Mat_<double>(2,2) << a, -b, b, a);
- \endcode
-*/
-template<typename _Tp> class MatCommaInitializer_
-{
-public:
-    //! the constructor, created by "matrix << firstValue" operator, where matrix is cv::Mat
-    MatCommaInitializer_(Mat_<_Tp>* _m);
-    //! the operator that takes the next value and put it to the matrix
-    template<typename T2> MatCommaInitializer_<_Tp>& operator , (T2 v);
-    //! another form of conversion operator
-    operator Mat_<_Tp>() const;
-protected:
-    MatIterator_<_Tp> it;
-};
-
-
-/////////////////////////////////////// Mat ///////////////////////////////////////////
-
-// note that umatdata might be allocated together
-// with the matrix data, not as a separate object.
-// therefore, it does not have constructor or destructor;
-// it should be explicitly initialized using init().
-struct CV_EXPORTS UMatData
-{
-    enum MemoryFlag { COPY_ON_MAP=1, HOST_COPY_OBSOLETE=2,
-        DEVICE_COPY_OBSOLETE=4, TEMP_UMAT=8, TEMP_COPIED_UMAT=24,
-        USER_ALLOCATED=32, DEVICE_MEM_MAPPED=64,
-        ASYNC_CLEANUP=128
-    };
-    UMatData(const MatAllocator* allocator);
-    ~UMatData();
-
-    // provide atomic access to the structure
-    void lock();
-    void unlock();
-
-    bool hostCopyObsolete() const;
-    bool deviceCopyObsolete() const;
-    bool deviceMemMapped() const;
-    bool copyOnMap() const;
-    bool tempUMat() const;
-    bool tempCopiedUMat() const;
-    void markHostCopyObsolete(bool flag);
-    void markDeviceCopyObsolete(bool flag);
-    void markDeviceMemMapped(bool flag);
-
-    const MatAllocator* prevAllocator;
-    const MatAllocator* currAllocator;
-    int urefcount;
-    int refcount;
-    uchar* data;
-    uchar* origdata;
-    size_t size;
-
-    UMatData::MemoryFlag flags;
-    void* handle;
-    void* userdata;
-    int allocatorFlags_;
-    int mapcount;
-    UMatData* originalUMatData;
-    std::shared_ptr<void> allocatorContext;
-};
-CV_ENUM_FLAGS(UMatData::MemoryFlag)
-
-
-struct CV_EXPORTS MatSize
-{
-    explicit MatSize(int* _p) CV_NOEXCEPT;
-    int dims() const CV_NOEXCEPT;
-    Size operator()() const;
-    const int& operator[](int i) const;
-    int& operator[](int i);
-    operator const int*() const CV_NOEXCEPT;  // TODO OpenCV 4.0: drop this
-    bool operator == (const MatSize& sz) const CV_NOEXCEPT;
-    bool operator != (const MatSize& sz) const CV_NOEXCEPT;
-
-    int* p;
-};
-
-struct CV_EXPORTS MatStep
-{
-    MatStep() CV_NOEXCEPT;
-    explicit MatStep(size_t s) CV_NOEXCEPT;
-    const size_t& operator[](int i) const CV_NOEXCEPT;
-    size_t& operator[](int i) CV_NOEXCEPT;
-    operator size_t() const;
-    MatStep& operator = (size_t s);
-
-    size_t* p;
-    size_t buf[2];
-protected:
-    MatStep& operator = (const MatStep&);
-};
-
-/** @example samples/cpp/cout_mat.cpp
-An example demonstrating the serial out capabilities of cv::Mat
-*/
-
- /** @brief n-dimensional dense array class \anchor CVMat_Details
-
-The class Mat represents an n-dimensional dense numerical single-channel or multi-channel array. It
-can be used to store real or complex-valued vectors and matrices, grayscale or color images, voxel
-volumes, vector fields, point clouds, tensors, histograms (though, very high-dimensional histograms
-may be better stored in a SparseMat ). The data layout of the array `M` is defined by the array
-`M.step[]`, so that the address of element \f$(i_0,...,i_{M.dims-1})\f$, where \f$0\leq i_k<M.size[k]\f$, is
-computed as:
-\f[addr(M_{i_0,...,i_{M.dims-1}}) = M.data + M.step[0]*i_0 + M.step[1]*i_1 + ... + M.step[M.dims-1]*i_{M.dims-1}\f]
-In case of a 2-dimensional array, the above formula is reduced to:
-\f[addr(M_{i,j}) = M.data + M.step[0]*i + M.step[1]*j\f]
-Note that `M.step[i] >= M.step[i+1]` (in fact, `M.step[i] >= M.step[i+1]*M.size[i+1]` ). This means
-that 2-dimensional matrices are stored row-by-row, 3-dimensional matrices are stored plane-by-plane,
-and so on. M.step[M.dims-1] is minimal and always equal to the element size M.elemSize() .
-
-So, the data layout in Mat is compatible with the majority of dense array types from the standard
-toolkits and SDKs, such as Numpy (ndarray), Win32 (independent device bitmaps), and others,
-that is, with any array that uses *steps* (or *strides*) to compute the position of a pixel.
-Due to this compatibility, it is possible to make a Mat header for user-allocated data and process
-it in-place using OpenCV functions.
-
-There are many different ways to create a Mat object. The most popular options are listed below:
-
-- Use the create(nrows, ncols, type) method or the similar Mat(nrows, ncols, type[, fillValue])
-constructor. A new array of the specified size and type is allocated. type has the same meaning as
-in the cvCreateMat method. For example, CV_8UC1 means a 8-bit single-channel array, CV_32FC2
-means a 2-channel (complex) floating-point array, and so on.
-@code
-    // make a 7x7 complex matrix filled with 1+3j.
-    Mat M(7,7,CV_32FC2,Scalar(1,3));
-    // and now turn M to a 100x60 15-channel 8-bit matrix.
-    // The old content will be deallocated
-    M.create(100,60,CV_8UC(15));
-@endcode
-As noted in the introduction to this chapter, create() allocates only a new array when the shape
-or type of the current array are different from the specified ones.
-
-- Create a multi-dimensional array:
-@code
-    // create a 100x100x100 8-bit array
-    int sz[] = {100, 100, 100};
-    Mat bigCube(3, sz, CV_8U, Scalar::all(0));
-@endcode
-It passes the number of dimensions =1 to the Mat constructor but the created array will be
-2-dimensional with the number of columns set to 1. So, Mat::dims is always \>= 2 (can also be 0
-when the array is empty).
-
-- Use a copy constructor or assignment operator where there can be an array or expression on the
-right side (see below). As noted in the introduction, the array assignment is an O(1) operation
-because it only copies the header and increases the reference counter. The Mat::clone() method can
-be used to get a full (deep) copy of the array when you need it.
-
-- Construct a header for a part of another array. It can be a single row, single column, several
-rows, several columns, rectangular region in the array (called a *minor* in algebra) or a
-diagonal. Such operations are also O(1) because the new header references the same data. You can
-actually modify a part of the array using this feature, for example:
-@code
-    // add the 5-th row, multiplied by 3 to the 3rd row
-    M.row(3) = M.row(3) + M.row(5)*3;
-    // now copy the 7-th column to the 1-st column
-    // M.col(1) = M.col(7); // this will not work
-    Mat M1 = M.col(1);
-    M.col(7).copyTo(M1);
-    // create a new 320x240 image
-    Mat img(Size(320,240),CV_8UC3);
-    // select a ROI
-    Mat roi(img, Rect(10,10,100,100));
-    // fill the ROI with (0,255,0) (which is green in RGB space);
-    // the original 320x240 image will be modified
-    roi = Scalar(0,255,0);
-@endcode
-Due to the additional datastart and dataend members, it is possible to compute a relative
-sub-array position in the main *container* array using locateROI():
-@code
-    Mat A = Mat::eye(10, 10, CV_32S);
-    // extracts A columns, 1 (inclusive) to 3 (exclusive).
-    Mat B = A(Range::all(), Range(1, 3));
-    // extracts B rows, 5 (inclusive) to 9 (exclusive).
-    // that is, C \~ A(Range(5, 9), Range(1, 3))
-    Mat C = B(Range(5, 9), Range::all());
-    Size size; Point ofs;
-    C.locateROI(size, ofs);
-    // size will be (width=10,height=10) and the ofs will be (x=1, y=5)
-@endcode
-As in case of whole matrices, if you need a deep copy, use the `clone()` method of the extracted
-sub-matrices.
-
-- Make a header for user-allocated data. It can be useful to do the following:
-    -# Process "foreign" data using OpenCV (for example, when you implement a DirectShow\* filter or
-    a processing module for gstreamer, and so on). For example:
-    @code
-        Mat process_video_frame(const unsigned char* pixels,
-                                int width, int height, int step)
-        {
-            // wrap input buffer
-            Mat img(height, width, CV_8UC3, (unsigned char*)pixels, step);
-
-            Mat result;
-            GaussianBlur(img, result, Size(7, 7), 1.5, 1.5);
-
-            return result;
-        }
-    @endcode
-    -# Quickly initialize small matrices and/or get a super-fast element access.
-    @code
-        double m[3][3] = {{a, b, c}, {d, e, f}, {g, h, i}};
-        Mat M = Mat(3, 3, CV_64F, m).inv();
-    @endcode
-    .
-
-- Use MATLAB-style array initializers, zeros(), ones(), eye(), for example:
-@code
-    // create a double-precision identity matrix and add it to M.
-    M += Mat::eye(M.rows, M.cols, CV_64F);
-@endcode
-
-- Use a comma-separated initializer:
-@code
-    // create a 3x3 double-precision identity matrix
-    Mat M = (Mat_<double>(3,3) << 1, 0, 0, 0, 1, 0, 0, 0, 1);
-@endcode
-With this approach, you first call a constructor of the Mat class with the proper parameters, and
-then you just put `<< operator` followed by comma-separated values that can be constants,
-variables, expressions, and so on. Also, note the extra parentheses required to avoid compilation
-errors.
-
-Once the array is created, it is automatically managed via a reference-counting mechanism. If the
-array header is built on top of user-allocated data, you should handle the data by yourself. The
-array data is deallocated when no one points to it. If you want to release the data pointed by a
-array header before the array destructor is called, use Mat::release().
-
-The next important thing to learn about the array class is element access. This manual already
-described how to compute an address of each array element. Normally, you are not required to use the
-formula directly in the code. If you know the array element type (which can be retrieved using the
-method Mat::type() ), you can access the element \f$M_{ij}\f$ of a 2-dimensional array as:
-@code
-    M.at<double>(i,j) += 1.f;
-@endcode
-assuming that `M` is a double-precision floating-point array. There are several variants of the method
-at for a different number of dimensions.
-
-If you need to process a whole row of a 2D array, the most efficient way is to get the pointer to
-the row first, and then just use the plain C operator [] :
-@code
-    // compute sum of positive matrix elements
-    // (assuming that M is a double-precision matrix)
-    double sum=0;
-    for(int i = 0; i < M.rows; i++)
-    {
-        const double* Mi = M.ptr<double>(i);
-        for(int j = 0; j < M.cols; j++)
-            sum += std::max(Mi[j], 0.);
-    }
-@endcode
-Some operations, like the one above, do not actually depend on the array shape. They just process
-elements of an array one by one (or elements from multiple arrays that have the same coordinates,
-for example, array addition). Such operations are called *element-wise*. It makes sense to check
-whether all the input/output arrays are continuous, namely, have no gaps at the end of each row. If
-yes, process them as a long single row:
-@code
-    // compute the sum of positive matrix elements, optimized variant
-    double sum=0;
-    int cols = M.cols, rows = M.rows;
-    if(M.isContinuous())
-    {
-        cols *= rows;
-        rows = 1;
-    }
-    for(int i = 0; i < rows; i++)
-    {
-        const double* Mi = M.ptr<double>(i);
-        for(int j = 0; j < cols; j++)
-            sum += std::max(Mi[j], 0.);
-    }
-@endcode
-In case of the continuous matrix, the outer loop body is executed just once. So, the overhead is
-smaller, which is especially noticeable in case of small matrices.
-
-Finally, there are STL-style iterators that are smart enough to skip gaps between successive rows:
-@code
-    // compute sum of positive matrix elements, iterator-based variant
-    double sum=0;
-    MatConstIterator_<double> it = M.begin<double>(), it_end = M.end<double>();
-    for(; it != it_end; ++it)
-        sum += std::max(*it, 0.);
-@endcode
-The matrix iterators are random-access iterators, so they can be passed to any STL algorithm,
-including std::sort().
-
-@note Matrix Expressions and arithmetic see MatExpr
-*/
-class CV_EXPORTS Mat
-{
-public:
-    /**
-    These are various constructors that form a matrix. As noted in the AutomaticAllocation, often
-    the default constructor is enough, and the proper matrix will be allocated by an OpenCV function.
-    The constructed matrix can further be assigned to another matrix or matrix expression or can be
-    allocated with Mat::create . In the former case, the old content is de-referenced.
-     */
-    Mat() CV_NOEXCEPT;
-
-    /** @overload
-    @param rows Number of rows in a 2D array.
-    @param cols Number of columns in a 2D array.
-    @param type Array type. Use CV_8UC1, ..., CV_64FC4 to create 1-4 channel matrices, or
-    CV_8UC(n), ..., CV_64FC(n) to create multi-channel (up to CV_CN_MAX channels) matrices.
-    */
-    Mat(int rows, int cols, int type);
-
-    /** @overload
-    @param size 2D array size: Size(cols, rows) . In the Size() constructor, the number of rows and the
-    number of columns go in the reverse order.
-    @param type Array type. Use CV_8UC1, ..., CV_64FC4 to create 1-4 channel matrices, or
-    CV_8UC(n), ..., CV_64FC(n) to create multi-channel (up to CV_CN_MAX channels) matrices.
-      */
-    Mat(Size size, int type);
-
-    /** @overload
-    @param rows Number of rows in a 2D array.
-    @param cols Number of columns in a 2D array.
-    @param type Array type. Use CV_8UC1, ..., CV_64FC4 to create 1-4 channel matrices, or
-    CV_8UC(n), ..., CV_64FC(n) to create multi-channel (up to CV_CN_MAX channels) matrices.
-    @param s An optional value to initialize each matrix element with. To set all the matrix elements to
-    the particular value after the construction, use the assignment operator
-    Mat::operator=(const Scalar& value) .
-    */
-    Mat(int rows, int cols, int type, const Scalar& s);
-
-    /** @overload
-    @param size 2D array size: Size(cols, rows) . In the Size() constructor, the number of rows and the
-    number of columns go in the reverse order.
-    @param type Array type. Use CV_8UC1, ..., CV_64FC4 to create 1-4 channel matrices, or
-    CV_8UC(n), ..., CV_64FC(n) to create multi-channel (up to CV_CN_MAX channels) matrices.
-    @param s An optional value to initialize each matrix element with. To set all the matrix elements to
-    the particular value after the construction, use the assignment operator
-    Mat::operator=(const Scalar& value) .
-      */
-    Mat(Size size, int type, const Scalar& s);
-
-    /** @overload
-    @param ndims Array dimensionality.
-    @param sizes Array of integers specifying an n-dimensional array shape.
-    @param type Array type. Use CV_8UC1, ..., CV_64FC4 to create 1-4 channel matrices, or
-    CV_8UC(n), ..., CV_64FC(n) to create multi-channel (up to CV_CN_MAX channels) matrices.
-    */
-    Mat(int ndims, const int* sizes, int type);
-
-    /** @overload
-    @param sizes Array of integers specifying an n-dimensional array shape.
-    @param type Array type. Use CV_8UC1, ..., CV_64FC4 to create 1-4 channel matrices, or
-    CV_8UC(n), ..., CV_64FC(n) to create multi-channel (up to CV_CN_MAX channels) matrices.
-    */
-    Mat(const std::vector<int>& sizes, int type);
-
-    /** @overload
-    @param ndims Array dimensionality.
-    @param sizes Array of integers specifying an n-dimensional array shape.
-    @param type Array type. Use CV_8UC1, ..., CV_64FC4 to create 1-4 channel matrices, or
-    CV_8UC(n), ..., CV_64FC(n) to create multi-channel (up to CV_CN_MAX channels) matrices.
-    @param s An optional value to initialize each matrix element with. To set all the matrix elements to
-    the particular value after the construction, use the assignment operator
-    Mat::operator=(const Scalar& value) .
-    */
-    Mat(int ndims, const int* sizes, int type, const Scalar& s);
-
-    /** @overload
-    @param sizes Array of integers specifying an n-dimensional array shape.
-    @param type Array type. Use CV_8UC1, ..., CV_64FC4 to create 1-4 channel matrices, or
-    CV_8UC(n), ..., CV_64FC(n) to create multi-channel (up to CV_CN_MAX channels) matrices.
-    @param s An optional value to initialize each matrix element with. To set all the matrix elements to
-    the particular value after the construction, use the assignment operator
-    Mat::operator=(const Scalar& value) .
-    */
-    Mat(const std::vector<int>& sizes, int type, const Scalar& s);
-
-
-    /** @overload
-    @param m Array that (as a whole or partly) is assigned to the constructed matrix. No data is copied
-    by these constructors. Instead, the header pointing to m data or its sub-array is constructed and
-    associated with it. The reference counter, if any, is incremented. So, when you modify the matrix
-    formed using such a constructor, you also modify the corresponding elements of m . If you want to
-    have an independent copy of the sub-array, use Mat::clone() .
-    */
-    Mat(const Mat& m);
-
-    /** @overload
-    @param rows Number of rows in a 2D array.
-    @param cols Number of columns in a 2D array.
-    @param type Array type. Use CV_8UC1, ..., CV_64FC4 to create 1-4 channel matrices, or
-    CV_8UC(n), ..., CV_64FC(n) to create multi-channel (up to CV_CN_MAX channels) matrices.
-    @param data Pointer to the user data. Matrix constructors that take data and step parameters do not
-    allocate matrix data. Instead, they just initialize the matrix header that points to the specified
-    data, which means that no data is copied. This operation is very efficient and can be used to
-    process external data using OpenCV functions. The external data is not automatically deallocated, so
-    you should take care of it.
-    @param step Number of bytes each matrix row occupies. The value should include the padding bytes at
-    the end of each row, if any. If the parameter is missing (set to AUTO_STEP ), no padding is assumed
-    and the actual step is calculated as cols*elemSize(). See Mat::elemSize.
-    */
-    Mat(int rows, int cols, int type, void* data, size_t step=AUTO_STEP);
-
-    /** @overload
-    @param size 2D array size: Size(cols, rows) . In the Size() constructor, the number of rows and the
-    number of columns go in the reverse order.
-    @param type Array type. Use CV_8UC1, ..., CV_64FC4 to create 1-4 channel matrices, or
-    CV_8UC(n), ..., CV_64FC(n) to create multi-channel (up to CV_CN_MAX channels) matrices.
-    @param data Pointer to the user data. Matrix constructors that take data and step parameters do not
-    allocate matrix data. Instead, they just initialize the matrix header that points to the specified
-    data, which means that no data is copied. This operation is very efficient and can be used to
-    process external data using OpenCV functions. The external data is not automatically deallocated, so
-    you should take care of it.
-    @param step Number of bytes each matrix row occupies. The value should include the padding bytes at
-    the end of each row, if any. If the parameter is missing (set to AUTO_STEP ), no padding is assumed
-    and the actual step is calculated as cols*elemSize(). See Mat::elemSize.
-    */
-    Mat(Size size, int type, void* data, size_t step=AUTO_STEP);
-
-    /** @overload
-    @param ndims Array dimensionality.
-    @param sizes Array of integers specifying an n-dimensional array shape.
-    @param type Array type. Use CV_8UC1, ..., CV_64FC4 to create 1-4 channel matrices, or
-    CV_8UC(n), ..., CV_64FC(n) to create multi-channel (up to CV_CN_MAX channels) matrices.
-    @param data Pointer to the user data. Matrix constructors that take data and step parameters do not
-    allocate matrix data. Instead, they just initialize the matrix header that points to the specified
-    data, which means that no data is copied. This operation is very efficient and can be used to
-    process external data using OpenCV functions. The external data is not automatically deallocated, so
-    you should take care of it.
-    @param steps Array of ndims-1 steps in case of a multi-dimensional array (the last step is always
-    set to the element size). If not specified, the matrix is assumed to be continuous.
-    */
-    Mat(int ndims, const int* sizes, int type, void* data, const size_t* steps=0);
-
-    /** @overload
-    @param sizes Array of integers specifying an n-dimensional array shape.
-    @param type Array type. Use CV_8UC1, ..., CV_64FC4 to create 1-4 channel matrices, or
-    CV_8UC(n), ..., CV_64FC(n) to create multi-channel (up to CV_CN_MAX channels) matrices.
-    @param data Pointer to the user data. Matrix constructors that take data and step parameters do not
-    allocate matrix data. Instead, they just initialize the matrix header that points to the specified
-    data, which means that no data is copied. This operation is very efficient and can be used to
-    process external data using OpenCV functions. The external data is not automatically deallocated, so
-    you should take care of it.
-    @param steps Array of ndims-1 steps in case of a multi-dimensional array (the last step is always
-    set to the element size). If not specified, the matrix is assumed to be continuous.
-    */
-    Mat(const std::vector<int>& sizes, int type, void* data, const size_t* steps=0);
-
-    /** @overload
-    @param m Array that (as a whole or partly) is assigned to the constructed matrix. No data is copied
-    by these constructors. Instead, the header pointing to m data or its sub-array is constructed and
-    associated with it. The reference counter, if any, is incremented. So, when you modify the matrix
-    formed using such a constructor, you also modify the corresponding elements of m . If you want to
-    have an independent copy of the sub-array, use Mat::clone() .
-    @param rowRange Range of the m rows to take. As usual, the range start is inclusive and the range
-    end is exclusive. Use Range::all() to take all the rows.
-    @param colRange Range of the m columns to take. Use Range::all() to take all the columns.
-    */
-    Mat(const Mat& m, const Range& rowRange, const Range& colRange=Range::all());
-
-    /** @overload
-    @param m Array that (as a whole or partly) is assigned to the constructed matrix. No data is copied
-    by these constructors. Instead, the header pointing to m data or its sub-array is constructed and
-    associated with it. The reference counter, if any, is incremented. So, when you modify the matrix
-    formed using such a constructor, you also modify the corresponding elements of m . If you want to
-    have an independent copy of the sub-array, use Mat::clone() .
-    @param roi Region of interest.
-    */
-    Mat(const Mat& m, const Rect& roi);
-
-    /** @overload
-    @param m Array that (as a whole or partly) is assigned to the constructed matrix. No data is copied
-    by these constructors. Instead, the header pointing to m data or its sub-array is constructed and
-    associated with it. The reference counter, if any, is incremented. So, when you modify the matrix
-    formed using such a constructor, you also modify the corresponding elements of m . If you want to
-    have an independent copy of the sub-array, use Mat::clone() .
-    @param ranges Array of selected ranges of m along each dimensionality.
-    */
-    Mat(const Mat& m, const Range* ranges);
-
-    /** @overload
-    @param m Array that (as a whole or partly) is assigned to the constructed matrix. No data is copied
-    by these constructors. Instead, the header pointing to m data or its sub-array is constructed and
-    associated with it. The reference counter, if any, is incremented. So, when you modify the matrix
-    formed using such a constructor, you also modify the corresponding elements of m . If you want to
-    have an independent copy of the sub-array, use Mat::clone() .
-    @param ranges Array of selected ranges of m along each dimensionality.
-    */
-    Mat(const Mat& m, const std::vector<Range>& ranges);
-
-    /** @overload
-    @param vec STL vector whose elements form the matrix. The matrix has a single column and the number
-    of rows equal to the number of vector elements. Type of the matrix matches the type of vector
-    elements. The constructor can handle arbitrary types, for which there is a properly declared
-    DataType . This means that the vector elements must be primitive numbers or uni-type numerical
-    tuples of numbers. Mixed-type structures are not supported. The corresponding constructor is
-    explicit. Since STL vectors are not automatically converted to Mat instances, you should write
-    Mat(vec) explicitly. Unless you copy the data into the matrix ( copyData=true ), no new elements
-    will be added to the vector because it can potentially yield vector data reallocation, and, thus,
-    the matrix data pointer will be invalid.
-    @param copyData Flag to specify whether the underlying data of the STL vector should be copied
-    to (true) or shared with (false) the newly constructed matrix. When the data is copied, the
-    allocated buffer is managed using Mat reference counting mechanism. While the data is shared,
-    the reference counter is NULL, and you should not deallocate the data until the matrix is not
-    destructed.
-    */
-    template<typename _Tp> explicit Mat(const std::vector<_Tp>& vec, bool copyData=false);
-
-    /** @overload
-    */
-    template<typename _Tp, typename = typename std::enable_if<std::is_arithmetic<_Tp>::value>::type>
-    explicit Mat(const std::initializer_list<_Tp> list);
-
-    /** @overload
-    */
-    template<typename _Tp> explicit Mat(const std::initializer_list<int> sizes, const std::initializer_list<_Tp> list);
-
-    /** @overload
-    */
-    template<typename _Tp, size_t _Nm> explicit Mat(const std::array<_Tp, _Nm>& arr, bool copyData=false);
-
-    /** @overload
-    */
-    template<typename _Tp, int n> explicit Mat(const Vec<_Tp, n>& vec, bool copyData=true);
-
-    /** @overload
-    */
-    template<typename _Tp, int m, int n> explicit Mat(const Matx<_Tp, m, n>& mtx, bool copyData=true);
-
-    /** @overload
-    */
-    template<typename _Tp> explicit Mat(const Point_<_Tp>& pt, bool copyData=true);
-
-    /** @overload
-    */
-    template<typename _Tp> explicit Mat(const Point3_<_Tp>& pt, bool copyData=true);
-
-    /** @overload
-    */
-    template<typename _Tp> explicit Mat(const MatCommaInitializer_<_Tp>& commaInitializer);
-
-    //! download data from GpuMat
-    explicit Mat(const cuda::GpuMat& m);
-
-    //! destructor - calls release()
-    ~Mat();
-
-    /** @brief assignment operators
-
-    These are available assignment operators. Since they all are very different, make sure to read the
-    operator parameters description.
-    @param m Assigned, right-hand-side matrix. Matrix assignment is an O(1) operation. This means that
-    no data is copied but the data is shared and the reference counter, if any, is incremented. Before
-    assigning new data, the old data is de-referenced via Mat::release .
-     */
-    Mat& operator = (const Mat& m);
-
-    /** @overload
-    @param expr Assigned matrix expression object. As opposite to the first form of the assignment
-    operation, the second form can reuse already allocated matrix if it has the right size and type to
-    fit the matrix expression result. It is automatically handled by the real function that the matrix
-    expressions is expanded to. For example, C=A+B is expanded to add(A, B, C), and add takes care of
-    automatic C reallocation.
-    */
-    Mat& operator = (const MatExpr& expr);
-
-    //! retrieve UMat from Mat
-    UMat getUMat(AccessFlag accessFlags, UMatUsageFlags usageFlags = USAGE_DEFAULT) const;
-
-    /** @brief Creates a matrix header for the specified matrix row.
-
-    The method makes a new header for the specified matrix row and returns it. This is an O(1)
-    operation, regardless of the matrix size. The underlying data of the new matrix is shared with the
-    original matrix. Here is the example of one of the classical basic matrix processing operations,
-    axpy, used by LU and many other algorithms:
-    @code
-        inline void matrix_axpy(Mat& A, int i, int j, double alpha)
-        {
-            A.row(i) += A.row(j)*alpha;
-        }
-    @endcode
-    @note In the current implementation, the following code does not work as expected:
-    @code
-        Mat A;
-        ...
-        A.row(i) = A.row(j); // will not work
-    @endcode
-    This happens because A.row(i) forms a temporary header that is further assigned to another header.
-    Remember that each of these operations is O(1), that is, no data is copied. Thus, the above
-    assignment is not true if you may have expected the j-th row to be copied to the i-th row. To
-    achieve that, you should either turn this simple assignment into an expression or use the
-    Mat::copyTo method:
-    @code
-        Mat A;
-        ...
-        // works, but looks a bit obscure.
-        A.row(i) = A.row(j) + 0;
-        // this is a bit longer, but the recommended method.
-        A.row(j).copyTo(A.row(i));
-    @endcode
-    @param y A 0-based row index.
-     */
-    Mat row(int y) const;
-
-    /** @brief Creates a matrix header for the specified matrix column.
-
-    The method makes a new header for the specified matrix column and returns it. This is an O(1)
-    operation, regardless of the matrix size. The underlying data of the new matrix is shared with the
-    original matrix. See also the Mat::row description.
-    @param x A 0-based column index.
-     */
-    Mat col(int x) const;
-
-    /** @brief Creates a matrix header for the specified row span.
-
-    The method makes a new header for the specified row span of the matrix. Similarly to Mat::row and
-    Mat::col , this is an O(1) operation.
-    @param startrow An inclusive 0-based start index of the row span.
-    @param endrow An exclusive 0-based ending index of the row span.
-     */
-    Mat rowRange(int startrow, int endrow) const;
-
-    /** @overload
-    @param r Range structure containing both the start and the end indices.
-    */
-    Mat rowRange(const Range& r) const;
-
-    /** @brief Creates a matrix header for the specified column span.
-
-    The method makes a new header for the specified column span of the matrix. Similarly to Mat::row and
-    Mat::col , this is an O(1) operation.
-    @param startcol An inclusive 0-based start index of the column span.
-    @param endcol An exclusive 0-based ending index of the column span.
-     */
-    Mat colRange(int startcol, int endcol) const;
-
-    /** @overload
-    @param r Range structure containing both the start and the end indices.
-    */
-    Mat colRange(const Range& r) const;
-
-    /** @brief Extracts a diagonal from a matrix
-
-    The method makes a new header for the specified matrix diagonal. The new matrix is represented as a
-    single-column matrix. Similarly to Mat::row and Mat::col, this is an O(1) operation.
-    @param d index of the diagonal, with the following values:
-    - `d=0` is the main diagonal.
-    - `d<0` is a diagonal from the lower half. For example, d=-1 means the diagonal is set
-      immediately below the main one.
-    - `d>0` is a diagonal from the upper half. For example, d=1 means the diagonal is set
-      immediately above the main one.
-    For example:
-    @code
-        Mat m = (Mat_<int>(3,3) <<
-                    1,2,3,
-                    4,5,6,
-                    7,8,9);
-        Mat d0 = m.diag(0);
-        Mat d1 = m.diag(1);
-        Mat d_1 = m.diag(-1);
-    @endcode
-    The resulting matrices are
-    @code
-     d0 =
-       [1;
-        5;
-        9]
-     d1 =
-       [2;
-        6]
-     d_1 =
-       [4;
-        8]
-    @endcode
-     */
-    Mat diag(int d=0) const;
-
-    /** @brief creates a diagonal matrix
-
-    The method creates a square diagonal matrix from specified main diagonal.
-    @param d One-dimensional matrix that represents the main diagonal.
-     */
-    CV_NODISCARD_STD static Mat diag(const Mat& d);
-
-    /** @brief Creates a full copy of the array and the underlying data.
-
-    The method creates a full copy of the array. The original step[] is not taken into account. So, the
-    array copy is a continuous array occupying total()*elemSize() bytes.
-     */
-    CV_NODISCARD_STD Mat clone() const;
-
-    /** @brief Copies the matrix to another one.
-
-    The method copies the matrix data to another matrix. Before copying the data, the method invokes :
-    @code
-        m.create(this->size(), this->type());
-    @endcode
-    so that the destination matrix is reallocated if needed. While m.copyTo(m); works flawlessly, the
-    function does not handle the case of a partial overlap between the source and the destination
-    matrices.
-
-    When the operation mask is specified, if the Mat::create call shown above reallocates the matrix,
-    the newly allocated matrix is initialized with all zeros before copying the data.
-    @param m Destination matrix. If it does not have a proper size or type before the operation, it is
-    reallocated.
-     */
-    void copyTo( OutputArray m ) const;
-
-    /** @overload
-    @param m Destination matrix. If it does not have a proper size or type before the operation, it is
-    reallocated.
-    @param mask Operation mask of the same size as \*this. Its non-zero elements indicate which matrix
-    elements need to be copied. The mask has to be of type CV_8U and can have 1 or multiple channels.
-    */
-    void copyTo( OutputArray m, InputArray mask ) const;
-
-    /** @brief Converts an array to another data type with optional scaling.
-
-    The method converts source pixel values to the target data type. saturate_cast\<\> is applied at
-    the end to avoid possible overflows:
-
-    \f[m(x,y) = saturate \_ cast<rType>( \alpha (*this)(x,y) +  \beta )\f]
-    @param m output matrix; if it does not have a proper size or type before the operation, it is
-    reallocated.
-    @param rtype desired output matrix type or, rather, the depth since the number of channels are the
-    same as the input has; if rtype is negative, the output matrix will have the same type as the input.
-    @param alpha optional scale factor.
-    @param beta optional delta added to the scaled values.
-     */
-    void convertTo( OutputArray m, int rtype, double alpha=1, double beta=0 ) const;
-
-    /** @brief Provides a functional form of convertTo.
-
-    This is an internally used method called by the @ref MatrixExpressions engine.
-    @param m Destination array.
-    @param type Desired destination array depth (or -1 if it should be the same as the source type).
-     */
-    void assignTo( Mat& m, int type=-1 ) const;
-
-    /** @brief Sets all or some of the array elements to the specified value.
-    @param s Assigned scalar converted to the actual array type.
-    */
-    Mat& operator = (const Scalar& s);
-
-    /** @brief Sets all or some of the array elements to the specified value.
-
-    This is an advanced variant of the Mat::operator=(const Scalar& s) operator.
-    @param value Assigned scalar converted to the actual array type.
-    @param mask Operation mask of the same size as \*this. Its non-zero elements indicate which matrix
-    elements need to be copied. The mask has to be of type CV_8U and can have 1 or multiple channels
-     */
-    Mat& setTo(InputArray value, InputArray mask=noArray());
-
-    /** @brief Changes the shape and/or the number of channels of a 2D matrix without copying the data.
-
-    The method makes a new matrix header for \*this elements. The new matrix may have a different size
-    and/or different number of channels. Any combination is possible if:
-    -   No extra elements are included into the new matrix and no elements are excluded. Consequently,
-        the product rows\*cols\*channels() must stay the same after the transformation.
-    -   No data is copied. That is, this is an O(1) operation. Consequently, if you change the number of
-        rows, or the operation changes the indices of elements row in some other way, the matrix must be
-        continuous. See Mat::isContinuous .
-
-    For example, if there is a set of 3D points stored as an STL vector, and you want to represent the
-    points as a 3xN matrix, do the following:
-    @code
-        std::vector<Point3f> vec;
-        ...
-        Mat pointMat = Mat(vec). // convert vector to Mat, O(1) operation
-                          reshape(1). // make Nx3 1-channel matrix out of Nx1 3-channel.
-                                      // Also, an O(1) operation
-                             t(); // finally, transpose the Nx3 matrix.
-                                  // This involves copying all the elements
-    @endcode
-    @param cn New number of channels. If the parameter is 0, the number of channels remains the same.
-    @param rows New number of rows. If the parameter is 0, the number of rows remains the same.
-     */
-    Mat reshape(int cn, int rows=0) const;
-
-    /** @overload */
-    Mat reshape(int cn, int newndims, const int* newsz) const;
-
-    /** @overload */
-    Mat reshape(int cn, const std::vector<int>& newshape) const;
-
-    /** @brief Transposes a matrix.
-
-    The method performs matrix transposition by means of matrix expressions. It does not perform the
-    actual transposition but returns a temporary matrix transposition object that can be further used as
-    a part of more complex matrix expressions or can be assigned to a matrix:
-    @code
-        Mat A1 = A + Mat::eye(A.size(), A.type())*lambda;
-        Mat C = A1.t()*A1; // compute (A + lambda*I)^t * (A + lamda*I)
-    @endcode
-     */
-    MatExpr t() const;
-
-    /** @brief Inverses a matrix.
-
-    The method performs a matrix inversion by means of matrix expressions. This means that a temporary
-    matrix inversion object is returned by the method and can be used further as a part of more complex
-    matrix expressions or can be assigned to a matrix.
-    @param method Matrix inversion method. One of cv::DecompTypes
-     */
-    MatExpr inv(int method=DECOMP_LU) const;
-
-    /** @brief Performs an element-wise multiplication or division of the two matrices.
-
-    The method returns a temporary object encoding per-element array multiplication, with optional
-    scale. Note that this is not a matrix multiplication that corresponds to a simpler "\*" operator.
-
-    Example:
-    @code
-        Mat C = A.mul(5/B); // equivalent to divide(A, B, C, 5)
-    @endcode
-    @param m Another array of the same type and the same size as \*this, or a matrix expression.
-    @param scale Optional scale factor.
-     */
-    MatExpr mul(InputArray m, double scale=1) const;
-
-    /** @brief Computes a cross-product of two 3-element vectors.
-
-    The method computes a cross-product of two 3-element vectors. The vectors must be 3-element
-    floating-point vectors of the same shape and size. The result is another 3-element vector of the
-    same shape and type as operands.
-    @param m Another cross-product operand.
-     */
-    Mat cross(InputArray m) const;
-
-    /** @brief Computes a dot-product of two vectors.
-
-    The method computes a dot-product of two matrices. If the matrices are not single-column or
-    single-row vectors, the top-to-bottom left-to-right scan ordering is used to treat them as 1D
-    vectors. The vectors must have the same size and type. If the matrices have more than one channel,
-    the dot products from all the channels are summed together.
-    @param m another dot-product operand.
-     */
-    double dot(InputArray m) const;
-
-    /** @brief Returns a zero array of the specified size and type.
-
-    The method returns a Matlab-style zero array initializer. It can be used to quickly form a constant
-    array as a function parameter, part of a matrix expression, or as a matrix initializer:
-    @code
-        Mat A;
-        A = Mat::zeros(3, 3, CV_32F);
-    @endcode
-    In the example above, a new matrix is allocated only if A is not a 3x3 floating-point matrix.
-    Otherwise, the existing matrix A is filled with zeros.
-    @param rows Number of rows.
-    @param cols Number of columns.
-    @param type Created matrix type.
-     */
-    CV_NODISCARD_STD static MatExpr zeros(int rows, int cols, int type);
-
-    /** @overload
-    @param size Alternative to the matrix size specification Size(cols, rows) .
-    @param type Created matrix type.
-    */
-    CV_NODISCARD_STD static MatExpr zeros(Size size, int type);
-
-    /** @overload
-    @param ndims Array dimensionality.
-    @param sz Array of integers specifying the array shape.
-    @param type Created matrix type.
-    */
-    CV_NODISCARD_STD static MatExpr zeros(int ndims, const int* sz, int type);
-
-    /** @brief Returns an array of all 1's of the specified size and type.
-
-    The method returns a Matlab-style 1's array initializer, similarly to Mat::zeros. Note that using
-    this method you can initialize an array with an arbitrary value, using the following Matlab idiom:
-    @code
-        Mat A = Mat::ones(100, 100, CV_8U)*3; // make 100x100 matrix filled with 3.
-    @endcode
-    The above operation does not form a 100x100 matrix of 1's and then multiply it by 3. Instead, it
-    just remembers the scale factor (3 in this case) and use it when actually invoking the matrix
-    initializer.
-    @note In case of multi-channels type, only the first channel will be initialized with 1's, the
-    others will be set to 0's.
-    @param rows Number of rows.
-    @param cols Number of columns.
-    @param type Created matrix type.
-     */
-    CV_NODISCARD_STD static MatExpr ones(int rows, int cols, int type);
-
-    /** @overload
-    @param size Alternative to the matrix size specification Size(cols, rows) .
-    @param type Created matrix type.
-    */
-    CV_NODISCARD_STD static MatExpr ones(Size size, int type);
-
-    /** @overload
-    @param ndims Array dimensionality.
-    @param sz Array of integers specifying the array shape.
-    @param type Created matrix type.
-    */
-    CV_NODISCARD_STD static MatExpr ones(int ndims, const int* sz, int type);
-
-    /** @brief Returns an identity matrix of the specified size and type.
-
-    The method returns a Matlab-style identity matrix initializer, similarly to Mat::zeros. Similarly to
-    Mat::ones, you can use a scale operation to create a scaled identity matrix efficiently:
-    @code
-        // make a 4x4 diagonal matrix with 0.1's on the diagonal.
-        Mat A = Mat::eye(4, 4, CV_32F)*0.1;
-    @endcode
-    @note In case of multi-channels type, identity matrix will be initialized only for the first channel,
-    the others will be set to 0's
-    @param rows Number of rows.
-    @param cols Number of columns.
-    @param type Created matrix type.
-     */
-    CV_NODISCARD_STD static MatExpr eye(int rows, int cols, int type);
-
-    /** @overload
-    @param size Alternative matrix size specification as Size(cols, rows) .
-    @param type Created matrix type.
-    */
-    CV_NODISCARD_STD static MatExpr eye(Size size, int type);
-
-    /** @brief Allocates new array data if needed.
-
-    This is one of the key Mat methods. Most new-style OpenCV functions and methods that produce arrays
-    call this method for each output array. The method uses the following algorithm:
-
-    -# If the current array shape and the type match the new ones, return immediately. Otherwise,
-       de-reference the previous data by calling Mat::release.
-    -# Initialize the new header.
-    -# Allocate the new data of total()\*elemSize() bytes.
-    -# Allocate the new, associated with the data, reference counter and set it to 1.
-
-    Such a scheme makes the memory management robust and efficient at the same time and helps avoid
-    extra typing for you. This means that usually there is no need to explicitly allocate output arrays.
-    That is, instead of writing:
-    @code
-        Mat color;
-        ...
-        Mat gray(color.rows, color.cols, color.depth());
-        cvtColor(color, gray, COLOR_BGR2GRAY);
-    @endcode
-    you can simply write:
-    @code
-        Mat color;
-        ...
-        Mat gray;
-        cvtColor(color, gray, COLOR_BGR2GRAY);
-    @endcode
-    because cvtColor, as well as the most of OpenCV functions, calls Mat::create() for the output array
-    internally.
-    @param rows New number of rows.
-    @param cols New number of columns.
-    @param type New matrix type.
-     */
-    void create(int rows, int cols, int type);
-
-    /** @overload
-    @param size Alternative new matrix size specification: Size(cols, rows)
-    @param type New matrix type.
-    */
-    void create(Size size, int type);
-
-    /** @overload
-    @param ndims New array dimensionality.
-    @param sizes Array of integers specifying a new array shape.
-    @param type New matrix type.
-    */
-    void create(int ndims, const int* sizes, int type);
-
-    /** @overload
-    @param sizes Array of integers specifying a new array shape.
-    @param type New matrix type.
-    */
-    void create(const std::vector<int>& sizes, int type);
-
-    /** @brief Increments the reference counter.
-
-    The method increments the reference counter associated with the matrix data. If the matrix header
-    points to an external data set (see Mat::Mat ), the reference counter is NULL, and the method has no
-    effect in this case. Normally, to avoid memory leaks, the method should not be called explicitly. It
-    is called implicitly by the matrix assignment operator. The reference counter increment is an atomic
-    operation on the platforms that support it. Thus, it is safe to operate on the same matrices
-    asynchronously in different threads.
-     */
-    void addref();
-
-    /** @brief Decrements the reference counter and deallocates the matrix if needed.
-
-    The method decrements the reference counter associated with the matrix data. When the reference
-    counter reaches 0, the matrix data is deallocated and the data and the reference counter pointers
-    are set to NULL's. If the matrix header points to an external data set (see Mat::Mat ), the
-    reference counter is NULL, and the method has no effect in this case.
-
-    This method can be called manually to force the matrix data deallocation. But since this method is
-    automatically called in the destructor, or by any other method that changes the data pointer, it is
-    usually not needed. The reference counter decrement and check for 0 is an atomic operation on the
-    platforms that support it. Thus, it is safe to operate on the same matrices asynchronously in
-    different threads.
-     */
-    void release();
-
-    //! internal use function, consider to use 'release' method instead; deallocates the matrix data
-    void deallocate();
-    //! internal use function; properly re-allocates _size, _step arrays
-    void copySize(const Mat& m);
-
-    /** @brief Reserves space for the certain number of rows.
-
-    The method reserves space for sz rows. If the matrix already has enough space to store sz rows,
-    nothing happens. If the matrix is reallocated, the first Mat::rows rows are preserved. The method
-    emulates the corresponding method of the STL vector class.
-    @param sz Number of rows.
-     */
-    void reserve(size_t sz);
-
-    /** @brief Reserves space for the certain number of bytes.
-
-    The method reserves space for sz bytes. If the matrix already has enough space to store sz bytes,
-    nothing happens. If matrix has to be reallocated its previous content could be lost.
-    @param sz Number of bytes.
-    */
-    void reserveBuffer(size_t sz);
-
-    /** @brief Changes the number of matrix rows.
-
-    The methods change the number of matrix rows. If the matrix is reallocated, the first
-    min(Mat::rows, sz) rows are preserved. The methods emulate the corresponding methods of the STL
-    vector class.
-    @param sz New number of rows.
-     */
-    void resize(size_t sz);
-
-    /** @overload
-    @param sz New number of rows.
-    @param s Value assigned to the newly added elements.
-     */
-    void resize(size_t sz, const Scalar& s);
-
-    //! internal function
-    void push_back_(const void* elem);
-
-    /** @brief Adds elements to the bottom of the matrix.
-
-    The methods add one or more elements to the bottom of the matrix. They emulate the corresponding
-    method of the STL vector class. When elem is Mat , its type and the number of columns must be the
-    same as in the container matrix.
-    @param elem Added element(s).
-     */
-    template<typename _Tp> void push_back(const _Tp& elem);
-
-    /** @overload
-    @param elem Added element(s).
-    */
-    template<typename _Tp> void push_back(const Mat_<_Tp>& elem);
-
-    /** @overload
-    @param elem Added element(s).
-    */
-    template<typename _Tp> void push_back(const std::vector<_Tp>& elem);
-
-    /** @overload
-    @param m Added line(s).
-    */
-    void push_back(const Mat& m);
-
-    /** @brief Removes elements from the bottom of the matrix.
-
-    The method removes one or more rows from the bottom of the matrix.
-    @param nelems Number of removed rows. If it is greater than the total number of rows, an exception
-    is thrown.
-     */
-    void pop_back(size_t nelems=1);
-
-    /** @brief Locates the matrix header within a parent matrix.
-
-    After you extracted a submatrix from a matrix using Mat::row, Mat::col, Mat::rowRange,
-    Mat::colRange, and others, the resultant submatrix points just to the part of the original big
-    matrix. However, each submatrix contains information (represented by datastart and dataend
-    fields) that helps reconstruct the original matrix size and the position of the extracted
-    submatrix within the original matrix. The method locateROI does exactly that.
-    @param wholeSize Output parameter that contains the size of the whole matrix containing *this*
-    as a part.
-    @param ofs Output parameter that contains an offset of *this* inside the whole matrix.
-     */
-    void locateROI( Size& wholeSize, Point& ofs ) const;
-
-    /** @brief Adjusts a submatrix size and position within the parent matrix.
-
-    The method is complimentary to Mat::locateROI . The typical use of these functions is to determine
-    the submatrix position within the parent matrix and then shift the position somehow. Typically, it
-    can be required for filtering operations when pixels outside of the ROI should be taken into
-    account. When all the method parameters are positive, the ROI needs to grow in all directions by the
-    specified amount, for example:
-    @code
-        A.adjustROI(2, 2, 2, 2);
-    @endcode
-    In this example, the matrix size is increased by 4 elements in each direction. The matrix is shifted
-    by 2 elements to the left and 2 elements up, which brings in all the necessary pixels for the
-    filtering with the 5x5 kernel.
-
-    adjustROI forces the adjusted ROI to be inside of the parent matrix that is boundaries of the
-    adjusted ROI are constrained by boundaries of the parent matrix. For example, if the submatrix A is
-    located in the first row of a parent matrix and you called A.adjustROI(2, 2, 2, 2) then A will not
-    be increased in the upward direction.
-
-    The function is used internally by the OpenCV filtering functions, like filter2D , morphological
-    operations, and so on.
-    @param dtop Shift of the top submatrix boundary upwards.
-    @param dbottom Shift of the bottom submatrix boundary downwards.
-    @param dleft Shift of the left submatrix boundary to the left.
-    @param dright Shift of the right submatrix boundary to the right.
-    @sa copyMakeBorder
-     */
-    Mat& adjustROI( int dtop, int dbottom, int dleft, int dright );
-
-    /** @brief Extracts a rectangular submatrix.
-
-    The operators make a new header for the specified sub-array of \*this . They are the most
-    generalized forms of Mat::row, Mat::col, Mat::rowRange, and Mat::colRange . For example,
-    `A(Range(0, 10), Range::all())` is equivalent to `A.rowRange(0, 10)`. Similarly to all of the above,
-    the operators are O(1) operations, that is, no matrix data is copied.
-    @param rowRange Start and end row of the extracted submatrix. The upper boundary is not included. To
-    select all the rows, use Range::all().
-    @param colRange Start and end column of the extracted submatrix. The upper boundary is not included.
-    To select all the columns, use Range::all().
-     */
-    Mat operator()( Range rowRange, Range colRange ) const;
-
-    /** @overload
-    @param roi Extracted submatrix specified as a rectangle.
-    */
-    Mat operator()( const Rect& roi ) const;
-
-    /** @overload
-    @param ranges Array of selected ranges along each array dimension.
-    */
-    Mat operator()( const Range* ranges ) const;
-
-    /** @overload
-    @param ranges Array of selected ranges along each array dimension.
-    */
-    Mat operator()(const std::vector<Range>& ranges) const;
-
-    template<typename _Tp> operator std::vector<_Tp>() const;
-    template<typename _Tp, int n> operator Vec<_Tp, n>() const;
-    template<typename _Tp, int m, int n> operator Matx<_Tp, m, n>() const;
-
-    template<typename _Tp, std::size_t _Nm> operator std::array<_Tp, _Nm>() const;
-
-    /** @brief Reports whether the matrix is continuous or not.
-
-    The method returns true if the matrix elements are stored continuously without gaps at the end of
-    each row. Otherwise, it returns false. Obviously, 1x1 or 1xN matrices are always continuous.
-    Matrices created with Mat::create are always continuous. But if you extract a part of the matrix
-    using Mat::col, Mat::diag, and so on, or constructed a matrix header for externally allocated data,
-    such matrices may no longer have this property.
-
-    The continuity flag is stored as a bit in the Mat::flags field and is computed automatically when
-    you construct a matrix header. Thus, the continuity check is a very fast operation, though
-    theoretically it could be done as follows:
-    @code
-        // alternative implementation of Mat::isContinuous()
-        bool myCheckMatContinuity(const Mat& m)
-        {
-            //return (m.flags & Mat::CONTINUOUS_FLAG) != 0;
-            return m.rows == 1 || m.step == m.cols*m.elemSize();
-        }
-    @endcode
-    The method is used in quite a few of OpenCV functions. The point is that element-wise operations
-    (such as arithmetic and logical operations, math functions, alpha blending, color space
-    transformations, and others) do not depend on the image geometry. Thus, if all the input and output
-    arrays are continuous, the functions can process them as very long single-row vectors. The example
-    below illustrates how an alpha-blending function can be implemented:
-    @code
-        template<typename T>
-        void alphaBlendRGBA(const Mat& src1, const Mat& src2, Mat& dst)
-        {
-            const float alpha_scale = (float)std::numeric_limits<T>::max(),
-                        inv_scale = 1.f/alpha_scale;
-
-            CV_Assert( src1.type() == src2.type() &&
-                       src1.type() == CV_MAKETYPE(traits::Depth<T>::value, 4) &&
-                       src1.size() == src2.size());
-            Size size = src1.size();
-            dst.create(size, src1.type());
-
-            // here is the idiom: check the arrays for continuity and,
-            // if this is the case,
-            // treat the arrays as 1D vectors
-            if( src1.isContinuous() && src2.isContinuous() && dst.isContinuous() )
-            {
-                size.width *= size.height;
-                size.height = 1;
-            }
-            size.width *= 4;
-
-            for( int i = 0; i < size.height; i++ )
-            {
-                // when the arrays are continuous,
-                // the outer loop is executed only once
-                const T* ptr1 = src1.ptr<T>(i);
-                const T* ptr2 = src2.ptr<T>(i);
-                T* dptr = dst.ptr<T>(i);
-
-                for( int j = 0; j < size.width; j += 4 )
-                {
-                    float alpha = ptr1[j+3]*inv_scale, beta = ptr2[j+3]*inv_scale;
-                    dptr[j] = saturate_cast<T>(ptr1[j]*alpha + ptr2[j]*beta);
-                    dptr[j+1] = saturate_cast<T>(ptr1[j+1]*alpha + ptr2[j+1]*beta);
-                    dptr[j+2] = saturate_cast<T>(ptr1[j+2]*alpha + ptr2[j+2]*beta);
-                    dptr[j+3] = saturate_cast<T>((1 - (1-alpha)*(1-beta))*alpha_scale);
-                }
-            }
-        }
-    @endcode
-    This approach, while being very simple, can boost the performance of a simple element-operation by
-    10-20 percents, especially if the image is rather small and the operation is quite simple.
-
-    Another OpenCV idiom in this function, a call of Mat::create for the destination array, that
-    allocates the destination array unless it already has the proper size and type. And while the newly
-    allocated arrays are always continuous, you still need to check the destination array because
-    Mat::create does not always allocate a new matrix.
-     */
-    bool isContinuous() const;
-
-    //! returns true if the matrix is a submatrix of another matrix
-    bool isSubmatrix() const;
-
-    /** @brief Returns the matrix element size in bytes.
-
-    The method returns the matrix element size in bytes. For example, if the matrix type is CV_16SC3 ,
-    the method returns 3\*sizeof(short) or 6.
-     */
-    size_t elemSize() const;
-
-    /** @brief Returns the size of each matrix element channel in bytes.
-
-    The method returns the matrix element channel size in bytes, that is, it ignores the number of
-    channels. For example, if the matrix type is CV_16SC3 , the method returns sizeof(short) or 2.
-     */
-    size_t elemSize1() const;
-
-    /** @brief Returns the type of a matrix element.
-
-    The method returns a matrix element type. This is an identifier compatible with the CvMat type
-    system, like CV_16SC3 or 16-bit signed 3-channel array, and so on.
-     */
-    int type() const;
-
-    /** @brief Returns the depth of a matrix element.
-
-    The method returns the identifier of the matrix element depth (the type of each individual channel).
-    For example, for a 16-bit signed element array, the method returns CV_16S . A complete list of
-    matrix types contains the following values:
-    -   CV_8U - 8-bit unsigned integers ( 0..255 )
-    -   CV_8S - 8-bit signed integers ( -128..127 )
-    -   CV_16U - 16-bit unsigned integers ( 0..65535 )
-    -   CV_16S - 16-bit signed integers ( -32768..32767 )
-    -   CV_32S - 32-bit signed integers ( -2147483648..2147483647 )
-    -   CV_32F - 32-bit floating-point numbers ( -FLT_MAX..FLT_MAX, INF, NAN )
-    -   CV_64F - 64-bit floating-point numbers ( -DBL_MAX..DBL_MAX, INF, NAN )
-     */
-    int depth() const;
-
-    /** @brief Returns the number of matrix channels.
-
-    The method returns the number of matrix channels.
-     */
-    int channels() const;
-
-    /** @brief Returns a normalized step.
-
-    The method returns a matrix step divided by Mat::elemSize1() . It can be useful to quickly access an
-    arbitrary matrix element.
-     */
-    size_t step1(int i=0) const;
-
-    /** @brief Returns true if the array has no elements.
-
-    The method returns true if Mat::total() is 0 or if Mat::data is NULL. Because of pop_back() and
-    resize() methods `M.total() == 0` does not imply that `M.data == NULL`.
-     */
-    bool empty() const;
-
-    /** @brief Returns the total number of array elements.
-
-    The method returns the number of array elements (a number of pixels if the array represents an
-    image).
-     */
-    size_t total() const;
-
-    /** @brief Returns the total number of array elements.
-
-     The method returns the number of elements within a certain sub-array slice with startDim <= dim < endDim
-     */
-    size_t total(int startDim, int endDim=INT_MAX) const;
-
-    /**
-     * @param elemChannels Number of channels or number of columns the matrix should have.
-     *                     For a 2-D matrix, when the matrix has only 1 column, then it should have
-     *                     elemChannels channels; When the matrix has only 1 channel,
-     *                     then it should have elemChannels columns.
-     *                     For a 3-D matrix, it should have only one channel. Furthermore,
-     *                     if the number of planes is not one, then the number of rows
-     *                     within every plane has to be 1; if the number of rows within
-     *                     every plane is not 1, then the number of planes has to be 1.
-     * @param depth The depth the matrix should have. Set it to -1 when any depth is fine.
-     * @param requireContinuous Set it to true to require the matrix to be continuous
-     * @return -1 if the requirement is not satisfied.
-     *         Otherwise, it returns the number of elements in the matrix. Note
-     *         that an element may have multiple channels.
-     *
-     * The following code demonstrates its usage for a 2-d matrix:
-     * @snippet snippets/core_mat_checkVector.cpp example-2d
-     *
-     * The following code demonstrates its usage for a 3-d matrix:
-     * @snippet snippets/core_mat_checkVector.cpp example-3d
-     */
-    int checkVector(int elemChannels, int depth=-1, bool requireContinuous=true) const;
-
-    /** @brief Returns a pointer to the specified matrix row.
-
-    The methods return `uchar*` or typed pointer to the specified matrix row. See the sample in
-    Mat::isContinuous to know how to use these methods.
-    @param i0 A 0-based row index.
-     */
-    uchar* ptr(int i0=0);
-    /** @overload */
-    const uchar* ptr(int i0=0) const;
-
-    /** @overload
-    @param row Index along the dimension 0
-    @param col Index along the dimension 1
-    */
-    uchar* ptr(int row, int col);
-    /** @overload
-    @param row Index along the dimension 0
-    @param col Index along the dimension 1
-    */
-    const uchar* ptr(int row, int col) const;
-
-    /** @overload */
-    uchar* ptr(int i0, int i1, int i2);
-    /** @overload */
-    const uchar* ptr(int i0, int i1, int i2) const;
-
-    /** @overload */
-    uchar* ptr(const int* idx);
-    /** @overload */
-    const uchar* ptr(const int* idx) const;
-    /** @overload */
-    template<int n> uchar* ptr(const Vec<int, n>& idx);
-    /** @overload */
-    template<int n> const uchar* ptr(const Vec<int, n>& idx) const;
-
-    /** @overload */
-    template<typename _Tp> _Tp* ptr(int i0=0);
-    /** @overload */
-    template<typename _Tp> const _Tp* ptr(int i0=0) const;
-    /** @overload
-    @param row Index along the dimension 0
-    @param col Index along the dimension 1
-    */
-    template<typename _Tp> _Tp* ptr(int row, int col);
-    /** @overload
-    @param row Index along the dimension 0
-    @param col Index along the dimension 1
-    */
-    template<typename _Tp> const _Tp* ptr(int row, int col) const;
-    /** @overload */
-    template<typename _Tp> _Tp* ptr(int i0, int i1, int i2);
-    /** @overload */
-    template<typename _Tp> const _Tp* ptr(int i0, int i1, int i2) const;
-    /** @overload */
-    template<typename _Tp> _Tp* ptr(const int* idx);
-    /** @overload */
-    template<typename _Tp> const _Tp* ptr(const int* idx) const;
-    /** @overload */
-    template<typename _Tp, int n> _Tp* ptr(const Vec<int, n>& idx);
-    /** @overload */
-    template<typename _Tp, int n> const _Tp* ptr(const Vec<int, n>& idx) const;
-
-    /** @brief Returns a reference to the specified array element.
-
-    The template methods return a reference to the specified array element. For the sake of higher
-    performance, the index range checks are only performed in the Debug configuration.
-
-    Note that the variants with a single index (i) can be used to access elements of single-row or
-    single-column 2-dimensional arrays. That is, if, for example, A is a 1 x N floating-point matrix and
-    B is an M x 1 integer matrix, you can simply write `A.at<float>(k+4)` and `B.at<int>(2*i+1)`
-    instead of `A.at<float>(0,k+4)` and `B.at<int>(2*i+1,0)`, respectively.
-
-    The example below initializes a Hilbert matrix:
-    @code
-        Mat H(100, 100, CV_64F);
-        for(int i = 0; i < H.rows; i++)
-            for(int j = 0; j < H.cols; j++)
-                H.at<double>(i,j)=1./(i+j+1);
-    @endcode
-
-    Keep in mind that the size identifier used in the at operator cannot be chosen at random. It depends
-    on the image from which you are trying to retrieve the data. The table below gives a better insight in this:
-     - If matrix is of type `CV_8U` then use `Mat.at<uchar>(y,x)`.
-     - If matrix is of type `CV_8S` then use `Mat.at<schar>(y,x)`.
-     - If matrix is of type `CV_16U` then use `Mat.at<ushort>(y,x)`.
-     - If matrix is of type `CV_16S` then use `Mat.at<short>(y,x)`.
-     - If matrix is of type `CV_32S`  then use `Mat.at<int>(y,x)`.
-     - If matrix is of type `CV_32F`  then use `Mat.at<float>(y,x)`.
-     - If matrix is of type `CV_64F` then use `Mat.at<double>(y,x)`.
-
-    @param i0 Index along the dimension 0
-     */
-    template<typename _Tp> _Tp& at(int i0=0);
-    /** @overload
-    @param i0 Index along the dimension 0
-    */
-    template<typename _Tp> const _Tp& at(int i0=0) const;
-    /** @overload
-    @param row Index along the dimension 0
-    @param col Index along the dimension 1
-    */
-    template<typename _Tp> _Tp& at(int row, int col);
-    /** @overload
-    @param row Index along the dimension 0
-    @param col Index along the dimension 1
-    */
-    template<typename _Tp> const _Tp& at(int row, int col) const;
-
-    /** @overload
-    @param i0 Index along the dimension 0
-    @param i1 Index along the dimension 1
-    @param i2 Index along the dimension 2
-    */
-    template<typename _Tp> _Tp& at(int i0, int i1, int i2);
-    /** @overload
-    @param i0 Index along the dimension 0
-    @param i1 Index along the dimension 1
-    @param i2 Index along the dimension 2
-    */
-    template<typename _Tp> const _Tp& at(int i0, int i1, int i2) const;
-
-    /** @overload
-    @param idx Array of Mat::dims indices.
-    */
-    template<typename _Tp> _Tp& at(const int* idx);
-    /** @overload
-    @param idx Array of Mat::dims indices.
-    */
-    template<typename _Tp> const _Tp& at(const int* idx) const;
-
-    /** @overload */
-    template<typename _Tp, int n> _Tp& at(const Vec<int, n>& idx);
-    /** @overload */
-    template<typename _Tp, int n> const _Tp& at(const Vec<int, n>& idx) const;
-
-    /** @overload
-    special versions for 2D arrays (especially convenient for referencing image pixels)
-    @param pt Element position specified as Point(j,i) .
-    */
-    template<typename _Tp> _Tp& at(Point pt);
-    /** @overload
-    special versions for 2D arrays (especially convenient for referencing image pixels)
-    @param pt Element position specified as Point(j,i) .
-    */
-    template<typename _Tp> const _Tp& at(Point pt) const;
-
-    /** @brief Returns the matrix iterator and sets it to the first matrix element.
-
-    The methods return the matrix read-only or read-write iterators. The use of matrix iterators is very
-    similar to the use of bi-directional STL iterators. In the example below, the alpha blending
-    function is rewritten using the matrix iterators:
-    @code
-        template<typename T>
-        void alphaBlendRGBA(const Mat& src1, const Mat& src2, Mat& dst)
-        {
-            typedef Vec<T, 4> VT;
-
-            const float alpha_scale = (float)std::numeric_limits<T>::max(),
-                        inv_scale = 1.f/alpha_scale;
-
-            CV_Assert( src1.type() == src2.type() &&
-                       src1.type() == traits::Type<VT>::value &&
-                       src1.size() == src2.size());
-            Size size = src1.size();
-            dst.create(size, src1.type());
-
-            MatConstIterator_<VT> it1 = src1.begin<VT>(), it1_end = src1.end<VT>();
-            MatConstIterator_<VT> it2 = src2.begin<VT>();
-            MatIterator_<VT> dst_it = dst.begin<VT>();
-
-            for( ; it1 != it1_end; ++it1, ++it2, ++dst_it )
-            {
-                VT pix1 = *it1, pix2 = *it2;
-                float alpha = pix1[3]*inv_scale, beta = pix2[3]*inv_scale;
-                *dst_it = VT(saturate_cast<T>(pix1[0]*alpha + pix2[0]*beta),
-                             saturate_cast<T>(pix1[1]*alpha + pix2[1]*beta),
-                             saturate_cast<T>(pix1[2]*alpha + pix2[2]*beta),
-                             saturate_cast<T>((1 - (1-alpha)*(1-beta))*alpha_scale));
-            }
-        }
-    @endcode
-     */
-    template<typename _Tp> MatIterator_<_Tp> begin();
-    template<typename _Tp> MatConstIterator_<_Tp> begin() const;
-
-    /** @brief Same as begin() but for inverse traversal
-     */
-    template<typename _Tp> std::reverse_iterator<MatIterator_<_Tp>> rbegin();
-    template<typename _Tp> std::reverse_iterator<MatConstIterator_<_Tp>> rbegin() const;
-
-    /** @brief Returns the matrix iterator and sets it to the after-last matrix element.
-
-    The methods return the matrix read-only or read-write iterators, set to the point following the last
-    matrix element.
-     */
-    template<typename _Tp> MatIterator_<_Tp> end();
-    template<typename _Tp> MatConstIterator_<_Tp> end() const;
-
-    /** @brief Same as end() but for inverse traversal
-     */
-    template<typename _Tp> std::reverse_iterator< MatIterator_<_Tp>> rend();
-    template<typename _Tp> std::reverse_iterator< MatConstIterator_<_Tp>> rend() const;
-
-
-    /** @brief Runs the given functor over all matrix elements in parallel.
-
-    The operation passed as argument has to be a function pointer, a function object or a lambda(C++11).
-
-    Example 1. All of the operations below put 0xFF the first channel of all matrix elements:
-    @code
-        Mat image(1920, 1080, CV_8UC3);
-        typedef cv::Point3_<uint8_t> Pixel;
-
-        // first. raw pointer access.
-        for (int r = 0; r < image.rows; ++r) {
-            Pixel* ptr = image.ptr<Pixel>(r, 0);
-            const Pixel* ptr_end = ptr + image.cols;
-            for (; ptr != ptr_end; ++ptr) {
-                ptr->x = 255;
-            }
-        }
-
-        // Using MatIterator. (Simple but there are a Iterator's overhead)
-        for (Pixel &p : cv::Mat_<Pixel>(image)) {
-            p.x = 255;
-        }
-
-        // Parallel execution with function object.
-        struct Operator {
-            void operator ()(Pixel &pixel, const int * position) {
-                pixel.x = 255;
-            }
-        };
-        image.forEach<Pixel>(Operator());
-
-        // Parallel execution using C++11 lambda.
-        image.forEach<Pixel>([](Pixel &p, const int * position) -> void {
-            p.x = 255;
-        });
-    @endcode
-    Example 2. Using the pixel's position:
-    @code
-        // Creating 3D matrix (255 x 255 x 255) typed uint8_t
-        // and initialize all elements by the value which equals elements position.
-        // i.e. pixels (x,y,z) = (1,2,3) is (b,g,r) = (1,2,3).
-
-        int sizes[] = { 255, 255, 255 };
-        typedef cv::Point3_<uint8_t> Pixel;
-
-        Mat_<Pixel> image = Mat::zeros(3, sizes, CV_8UC3);
-
-        image.forEach<Pixel>([&](Pixel& pixel, const int position[]) -> void {
-            pixel.x = position[0];
-            pixel.y = position[1];
-            pixel.z = position[2];
-        });
-    @endcode
-     */
-    template<typename _Tp, typename Functor> void forEach(const Functor& operation);
-    /** @overload */
-    template<typename _Tp, typename Functor> void forEach(const Functor& operation) const;
-
-    Mat(Mat&& m);
-    Mat& operator = (Mat&& m);
-
-    enum { MAGIC_VAL  = 0x42FF0000, AUTO_STEP = 0, CONTINUOUS_FLAG = CV_MAT_CONT_FLAG, SUBMATRIX_FLAG = CV_SUBMAT_FLAG };
-    enum { MAGIC_MASK = 0xFFFF0000, TYPE_MASK = 0x00000FFF, DEPTH_MASK = 7 };
-
-    /*! includes several bit-fields:
-         - the magic signature
-         - continuity flag
-         - depth
-         - number of channels
-     */
-    int flags;
-    //! the matrix dimensionality, >= 2
-    int dims;
-    //! the number of rows and columns or (-1, -1) when the matrix has more than 2 dimensions
-    int rows, cols;
-    //! pointer to the data
-    uchar* data;
-
-    //! helper fields used in locateROI and adjustROI
-    const uchar* datastart;
-    const uchar* dataend;
-    const uchar* datalimit;
-
-    //! custom allocator
-    MatAllocator* allocator;
-    //! and the standard allocator
-    static MatAllocator* getStdAllocator();
-    static MatAllocator* getDefaultAllocator();
-    static void setDefaultAllocator(MatAllocator* allocator);
-
-    //! internal use method: updates the continuity flag
-    void updateContinuityFlag();
-
-    //! interaction with UMat
-    UMatData* u;
-
-    MatSize size;
-    MatStep step;
-
-protected:
-    template<typename _Tp, typename Functor> void forEach_impl(const Functor& operation);
-};
-
-
-///////////////////////////////// Mat_<_Tp> ////////////////////////////////////
-
-/** @brief Template matrix class derived from Mat
-
-@code{.cpp}
-    template<typename _Tp> class Mat_ : public Mat
-    {
-    public:
-        // ... some specific methods
-        //         and
-        // no new extra fields
-    };
-@endcode
-The class `Mat_<_Tp>` is a *thin* template wrapper on top of the Mat class. It does not have any
-extra data fields. Nor this class nor Mat has any virtual methods. Thus, references or pointers to
-these two classes can be freely but carefully converted one to another. For example:
-@code{.cpp}
-    // create a 100x100 8-bit matrix
-    Mat M(100,100,CV_8U);
-    // this will be compiled fine. no any data conversion will be done.
-    Mat_<float>& M1 = (Mat_<float>&)M;
-    // the program is likely to crash at the statement below
-    M1(99,99) = 1.f;
-@endcode
-While Mat is sufficient in most cases, Mat_ can be more convenient if you use a lot of element
-access operations and if you know matrix type at the compilation time. Note that
-`Mat::at(int y,int x)` and `Mat_::operator()(int y,int x)` do absolutely the same
-and run at the same speed, but the latter is certainly shorter:
-@code{.cpp}
-    Mat_<double> M(20,20);
-    for(int i = 0; i < M.rows; i++)
-        for(int j = 0; j < M.cols; j++)
-            M(i,j) = 1./(i+j+1);
-    Mat E, V;
-    eigen(M,E,V);
-    cout << E.at<double>(0,0)/E.at<double>(M.rows-1,0);
-@endcode
-To use Mat_ for multi-channel images/matrices, pass Vec as a Mat_ parameter:
-@code{.cpp}
-    // allocate a 320x240 color image and fill it with green (in RGB space)
-    Mat_<Vec3b> img(240, 320, Vec3b(0,255,0));
-    // now draw a diagonal white line
-    for(int i = 0; i < 100; i++)
-        img(i,i)=Vec3b(255,255,255);
-    // and now scramble the 2nd (red) channel of each pixel
-    for(int i = 0; i < img.rows; i++)
-        for(int j = 0; j < img.cols; j++)
-            img(i,j)[2] ^= (uchar)(i ^ j);
-@endcode
-Mat_ is fully compatible with C++11 range-based for loop. For example such loop
-can be used to safely apply look-up table:
-@code{.cpp}
-void applyTable(Mat_<uchar>& I, const uchar* const table)
-{
-    for(auto& pixel : I)
-    {
-        pixel = table[pixel];
-    }
-}
-@endcode
- */
-template<typename _Tp> class Mat_ : public Mat
-{
-public:
-    typedef _Tp value_type;
-    typedef typename DataType<_Tp>::channel_type channel_type;
-    typedef MatIterator_<_Tp> iterator;
-    typedef MatConstIterator_<_Tp> const_iterator;
-
-    //! default constructor
-    Mat_() CV_NOEXCEPT;
-    //! equivalent to Mat(_rows, _cols, DataType<_Tp>::type)
-    Mat_(int _rows, int _cols);
-    //! constructor that sets each matrix element to specified value
-    Mat_(int _rows, int _cols, const _Tp& value);
-    //! equivalent to Mat(_size, DataType<_Tp>::type)
-    explicit Mat_(Size _size);
-    //! constructor that sets each matrix element to specified value
-    Mat_(Size _size, const _Tp& value);
-    //! n-dim array constructor
-    Mat_(int _ndims, const int* _sizes);
-    //! n-dim array constructor that sets each matrix element to specified value
-    Mat_(int _ndims, const int* _sizes, const _Tp& value);
-    //! copy/conversion constructor. If m is of different type, it's converted
-    Mat_(const Mat& m);
-    //! copy constructor
-    Mat_(const Mat_& m);
-    //! constructs a matrix on top of user-allocated data. step is in bytes(!!!), regardless of the type
-    Mat_(int _rows, int _cols, _Tp* _data, size_t _step=AUTO_STEP);
-    //! constructs n-dim matrix on top of user-allocated data. steps are in bytes(!!!), regardless of the type
-    Mat_(int _ndims, const int* _sizes, _Tp* _data, const size_t* _steps=0);
-    //! selects a submatrix
-    Mat_(const Mat_& m, const Range& rowRange, const Range& colRange=Range::all());
-    //! selects a submatrix
-    Mat_(const Mat_& m, const Rect& roi);
-    //! selects a submatrix, n-dim version
-    Mat_(const Mat_& m, const Range* ranges);
-    //! selects a submatrix, n-dim version
-    Mat_(const Mat_& m, const std::vector<Range>& ranges);
-    //! from a matrix expression
-    explicit Mat_(const MatExpr& e);
-    //! makes a matrix out of Vec, std::vector, Point_ or Point3_. The matrix will have a single column
-    explicit Mat_(const std::vector<_Tp>& vec, bool copyData=false);
-    template<int n> explicit Mat_(const Vec<typename DataType<_Tp>::channel_type, n>& vec, bool copyData=true);
-    template<int m, int n> explicit Mat_(const Matx<typename DataType<_Tp>::channel_type, m, n>& mtx, bool copyData=true);
-    explicit Mat_(const Point_<typename DataType<_Tp>::channel_type>& pt, bool copyData=true);
-    explicit Mat_(const Point3_<typename DataType<_Tp>::channel_type>& pt, bool copyData=true);
-    explicit Mat_(const MatCommaInitializer_<_Tp>& commaInitializer);
-
-    Mat_(std::initializer_list<_Tp> values);
-    explicit Mat_(const std::initializer_list<int> sizes, const std::initializer_list<_Tp> values);
-
-    template <std::size_t _Nm> explicit Mat_(const std::array<_Tp, _Nm>& arr, bool copyData=false);
-
-    Mat_& operator = (const Mat& m);
-    Mat_& operator = (const Mat_& m);
-    //! set all the elements to s.
-    Mat_& operator = (const _Tp& s);
-    //! assign a matrix expression
-    Mat_& operator = (const MatExpr& e);
-
-    //! iterators; they are smart enough to skip gaps in the end of rows
-    iterator begin();
-    iterator end();
-    const_iterator begin() const;
-    const_iterator end() const;
-
-    //reverse iterators
-    std::reverse_iterator<iterator> rbegin();
-    std::reverse_iterator<iterator> rend();
-    std::reverse_iterator<const_iterator> rbegin() const;
-    std::reverse_iterator<const_iterator> rend() const;
-
-    //! template methods for for operation over all matrix elements.
-    // the operations take care of skipping gaps in the end of rows (if any)
-    template<typename Functor> void forEach(const Functor& operation);
-    template<typename Functor> void forEach(const Functor& operation) const;
-
-    //! equivalent to Mat::create(_rows, _cols, DataType<_Tp>::type)
-    void create(int _rows, int _cols);
-    //! equivalent to Mat::create(_size, DataType<_Tp>::type)
-    void create(Size _size);
-    //! equivalent to Mat::create(_ndims, _sizes, DatType<_Tp>::type)
-    void create(int _ndims, const int* _sizes);
-    //! equivalent to Mat::release()
-    void release();
-    //! cross-product
-    Mat_ cross(const Mat_& m) const;
-    //! data type conversion
-    template<typename T2> operator Mat_<T2>() const;
-    //! overridden forms of Mat::row() etc.
-    Mat_ row(int y) const;
-    Mat_ col(int x) const;
-    Mat_ diag(int d=0) const;
-    CV_NODISCARD_STD Mat_ clone() const;
-
-    //! overridden forms of Mat::elemSize() etc.
-    size_t elemSize() const;
-    size_t elemSize1() const;
-    int type() const;
-    int depth() const;
-    int channels() const;
-    size_t step1(int i=0) const;
-    //! returns step()/sizeof(_Tp)
-    size_t stepT(int i=0) const;
-
-    //! overridden forms of Mat::zeros() etc. Data type is omitted, of course
-    CV_NODISCARD_STD static MatExpr zeros(int rows, int cols);
-    CV_NODISCARD_STD static MatExpr zeros(Size size);
-    CV_NODISCARD_STD static MatExpr zeros(int _ndims, const int* _sizes);
-    CV_NODISCARD_STD static MatExpr ones(int rows, int cols);
-    CV_NODISCARD_STD static MatExpr ones(Size size);
-    CV_NODISCARD_STD static MatExpr ones(int _ndims, const int* _sizes);
-    CV_NODISCARD_STD static MatExpr eye(int rows, int cols);
-    CV_NODISCARD_STD static MatExpr eye(Size size);
-
-    //! some more overridden methods
-    Mat_& adjustROI( int dtop, int dbottom, int dleft, int dright );
-    Mat_ operator()( const Range& rowRange, const Range& colRange ) const;
-    Mat_ operator()( const Rect& roi ) const;
-    Mat_ operator()( const Range* ranges ) const;
-    Mat_ operator()(const std::vector<Range>& ranges) const;
-
-    //! more convenient forms of row and element access operators
-    _Tp* operator [](int y);
-    const _Tp* operator [](int y) const;
-
-    //! returns reference to the specified element
-    _Tp& operator ()(const int* idx);
-    //! returns read-only reference to the specified element
-    const _Tp& operator ()(const int* idx) const;
-
-    //! returns reference to the specified element
-    template<int n> _Tp& operator ()(const Vec<int, n>& idx);
-    //! returns read-only reference to the specified element
-    template<int n> const _Tp& operator ()(const Vec<int, n>& idx) const;
-
-    //! returns reference to the specified element (1D case)
-    _Tp& operator ()(int idx0);
-    //! returns read-only reference to the specified element (1D case)
-    const _Tp& operator ()(int idx0) const;
-    //! returns reference to the specified element (2D case)
-    _Tp& operator ()(int row, int col);
-    //! returns read-only reference to the specified element (2D case)
-    const _Tp& operator ()(int row, int col) const;
-    //! returns reference to the specified element (3D case)
-    _Tp& operator ()(int idx0, int idx1, int idx2);
-    //! returns read-only reference to the specified element (3D case)
-    const _Tp& operator ()(int idx0, int idx1, int idx2) const;
-
-    _Tp& operator ()(Point pt);
-    const _Tp& operator ()(Point pt) const;
-
-    //! conversion to vector.
-    operator std::vector<_Tp>() const;
-
-    //! conversion to array.
-    template<std::size_t _Nm> operator std::array<_Tp, _Nm>() const;
-
-    //! conversion to Vec
-    template<int n> operator Vec<typename DataType<_Tp>::channel_type, n>() const;
-    //! conversion to Matx
-    template<int m, int n> operator Matx<typename DataType<_Tp>::channel_type, m, n>() const;
-
-    Mat_(Mat_&& m);
-    Mat_& operator = (Mat_&& m);
-
-    Mat_(Mat&& m);
-    Mat_& operator = (Mat&& m);
-
-    Mat_(MatExpr&& e);
-};
-
-typedef Mat_<uchar> Mat1b;
-typedef Mat_<Vec2b> Mat2b;
-typedef Mat_<Vec3b> Mat3b;
-typedef Mat_<Vec4b> Mat4b;
-
-typedef Mat_<short> Mat1s;
-typedef Mat_<Vec2s> Mat2s;
-typedef Mat_<Vec3s> Mat3s;
-typedef Mat_<Vec4s> Mat4s;
-
-typedef Mat_<ushort> Mat1w;
-typedef Mat_<Vec2w> Mat2w;
-typedef Mat_<Vec3w> Mat3w;
-typedef Mat_<Vec4w> Mat4w;
-
-typedef Mat_<int>   Mat1i;
-typedef Mat_<Vec2i> Mat2i;
-typedef Mat_<Vec3i> Mat3i;
-typedef Mat_<Vec4i> Mat4i;
-
-typedef Mat_<float> Mat1f;
-typedef Mat_<Vec2f> Mat2f;
-typedef Mat_<Vec3f> Mat3f;
-typedef Mat_<Vec4f> Mat4f;
-
-typedef Mat_<double> Mat1d;
-typedef Mat_<Vec2d> Mat2d;
-typedef Mat_<Vec3d> Mat3d;
-typedef Mat_<Vec4d> Mat4d;
-
-/** @todo document */
-class CV_EXPORTS UMat
-{
-public:
-    //! default constructor
-    UMat(UMatUsageFlags usageFlags = USAGE_DEFAULT) CV_NOEXCEPT;
-    //! constructs 2D matrix of the specified size and type
-    // (_type is CV_8UC1, CV_64FC3, CV_32SC(12) etc.)
-    UMat(int rows, int cols, int type, UMatUsageFlags usageFlags = USAGE_DEFAULT);
-    UMat(Size size, int type, UMatUsageFlags usageFlags = USAGE_DEFAULT);
-    //! constructs 2D matrix and fills it with the specified value _s.
-    UMat(int rows, int cols, int type, const Scalar& s, UMatUsageFlags usageFlags = USAGE_DEFAULT);
-    UMat(Size size, int type, const Scalar& s, UMatUsageFlags usageFlags = USAGE_DEFAULT);
-
-    //! constructs n-dimensional matrix
-    UMat(int ndims, const int* sizes, int type, UMatUsageFlags usageFlags = USAGE_DEFAULT);
-    UMat(int ndims, const int* sizes, int type, const Scalar& s, UMatUsageFlags usageFlags = USAGE_DEFAULT);
-
-    //! copy constructor
-    UMat(const UMat& m);
-
-    //! creates a matrix header for a part of the bigger matrix
-    UMat(const UMat& m, const Range& rowRange, const Range& colRange=Range::all());
-    UMat(const UMat& m, const Rect& roi);
-    UMat(const UMat& m, const Range* ranges);
-    UMat(const UMat& m, const std::vector<Range>& ranges);
-
-    // FIXIT copyData=false is not implemented, drop this in favor of cv::Mat (OpenCV 5.0)
-    //! builds matrix from std::vector with or without copying the data
-    template<typename _Tp> explicit UMat(const std::vector<_Tp>& vec, bool copyData=false);
-
-    //! destructor - calls release()
-    ~UMat();
-    //! assignment operators
-    UMat& operator = (const UMat& m);
-
-    Mat getMat(AccessFlag flags) const;
-
-    //! returns a new matrix header for the specified row
-    UMat row(int y) const;
-    //! returns a new matrix header for the specified column
-    UMat col(int x) const;
-    //! ... for the specified row span
-    UMat rowRange(int startrow, int endrow) const;
-    UMat rowRange(const Range& r) const;
-    //! ... for the specified column span
-    UMat colRange(int startcol, int endcol) const;
-    UMat colRange(const Range& r) const;
-    //! ... for the specified diagonal
-    //! (d=0 - the main diagonal,
-    //!  >0 - a diagonal from the upper half,
-    //!  <0 - a diagonal from the lower half)
-    UMat diag(int d=0) const;
-    //! constructs a square diagonal matrix which main diagonal is vector "d"
-    CV_NODISCARD_STD static UMat diag(const UMat& d, UMatUsageFlags usageFlags /*= USAGE_DEFAULT*/);
-    CV_NODISCARD_STD static UMat diag(const UMat& d) { return diag(d, USAGE_DEFAULT); }  // OpenCV 5.0: remove abi compatibility overload
-
-    //! returns deep copy of the matrix, i.e. the data is copied
-    CV_NODISCARD_STD UMat clone() const;
-    //! copies the matrix content to "m".
-    // It calls m.create(this->size(), this->type()).
-    void copyTo( OutputArray m ) const;
-    //! copies those matrix elements to "m" that are marked with non-zero mask elements.
-    void copyTo( OutputArray m, InputArray mask ) const;
-    //! converts matrix to another datatype with optional scaling. See cvConvertScale.
-    void convertTo( OutputArray m, int rtype, double alpha=1, double beta=0 ) const;
-
-    void assignTo( UMat& m, int type=-1 ) const;
-
-    //! sets every matrix element to s
-    UMat& operator = (const Scalar& s);
-    //! sets some of the matrix elements to s, according to the mask
-    UMat& setTo(InputArray value, InputArray mask=noArray());
-    //! creates alternative matrix header for the same data, with different
-    // number of channels and/or different number of rows. see cvReshape.
-    UMat reshape(int cn, int rows=0) const;
-    UMat reshape(int cn, int newndims, const int* newsz) const;
-
-    //! matrix transposition by means of matrix expressions
-    UMat t() const;
-    //! matrix inversion by means of matrix expressions
-    UMat inv(int method=DECOMP_LU) const;
-    //! per-element matrix multiplication by means of matrix expressions
-    UMat mul(InputArray m, double scale=1) const;
-
-    //! computes dot-product
-    double dot(InputArray m) const;
-
-    //! Matlab-style matrix initialization
-    CV_NODISCARD_STD static UMat zeros(int rows, int cols, int type, UMatUsageFlags usageFlags /*= USAGE_DEFAULT*/);
-    CV_NODISCARD_STD static UMat zeros(Size size, int type, UMatUsageFlags usageFlags /*= USAGE_DEFAULT*/);
-    CV_NODISCARD_STD static UMat zeros(int ndims, const int* sz, int type, UMatUsageFlags usageFlags /*= USAGE_DEFAULT*/);
-    CV_NODISCARD_STD static UMat zeros(int rows, int cols, int type) { return zeros(rows, cols, type, USAGE_DEFAULT); }  // OpenCV 5.0: remove abi compatibility overload
-    CV_NODISCARD_STD static UMat zeros(Size size, int type) { return zeros(size, type, USAGE_DEFAULT); }  // OpenCV 5.0: remove abi compatibility overload
-    CV_NODISCARD_STD static UMat zeros(int ndims, const int* sz, int type) { return zeros(ndims, sz, type, USAGE_DEFAULT); }  // OpenCV 5.0: remove abi compatibility overload
-    CV_NODISCARD_STD static UMat ones(int rows, int cols, int type, UMatUsageFlags usageFlags /*= USAGE_DEFAULT*/);
-    CV_NODISCARD_STD static UMat ones(Size size, int type, UMatUsageFlags usageFlags /*= USAGE_DEFAULT*/);
-    CV_NODISCARD_STD static UMat ones(int ndims, const int* sz, int type, UMatUsageFlags usageFlags /*= USAGE_DEFAULT*/);
-    CV_NODISCARD_STD static UMat ones(int rows, int cols, int type) { return ones(rows, cols, type, USAGE_DEFAULT); }  // OpenCV 5.0: remove abi compatibility overload
-    CV_NODISCARD_STD static UMat ones(Size size, int type) { return ones(size, type, USAGE_DEFAULT); }  // OpenCV 5.0: remove abi compatibility overload
-    CV_NODISCARD_STD static UMat ones(int ndims, const int* sz, int type) { return ones(ndims, sz, type, USAGE_DEFAULT); }  // OpenCV 5.0: remove abi compatibility overload
-    CV_NODISCARD_STD static UMat eye(int rows, int cols, int type, UMatUsageFlags usageFlags /*= USAGE_DEFAULT*/);
-    CV_NODISCARD_STD static UMat eye(Size size, int type, UMatUsageFlags usageFlags /*= USAGE_DEFAULT*/);
-    CV_NODISCARD_STD static UMat eye(int rows, int cols, int type) { return eye(rows, cols, type, USAGE_DEFAULT); }  // OpenCV 5.0: remove abi compatibility overload
-    CV_NODISCARD_STD static UMat eye(Size size, int type) { return eye(size, type, USAGE_DEFAULT); }  // OpenCV 5.0: remove abi compatibility overload
-
-    //! allocates new matrix data unless the matrix already has specified size and type.
-    // previous data is unreferenced if needed.
-    void create(int rows, int cols, int type, UMatUsageFlags usageFlags = USAGE_DEFAULT);
-    void create(Size size, int type, UMatUsageFlags usageFlags = USAGE_DEFAULT);
-    void create(int ndims, const int* sizes, int type, UMatUsageFlags usageFlags = USAGE_DEFAULT);
-    void create(const std::vector<int>& sizes, int type, UMatUsageFlags usageFlags = USAGE_DEFAULT);
-
-    //! increases the reference counter; use with care to avoid memleaks
-    void addref();
-    //! decreases reference counter;
-    // deallocates the data when reference counter reaches 0.
-    void release();
-
-    //! deallocates the matrix data
-    void deallocate();
-    //! internal use function; properly re-allocates _size, _step arrays
-    void copySize(const UMat& m);
-
-    //! locates matrix header within a parent matrix. See below
-    void locateROI( Size& wholeSize, Point& ofs ) const;
-    //! moves/resizes the current matrix ROI inside the parent matrix.
-    UMat& adjustROI( int dtop, int dbottom, int dleft, int dright );
-    //! extracts a rectangular sub-matrix
-    // (this is a generalized form of row, rowRange etc.)
-    UMat operator()( Range rowRange, Range colRange ) const;
-    UMat operator()( const Rect& roi ) const;
-    UMat operator()( const Range* ranges ) const;
-    UMat operator()(const std::vector<Range>& ranges) const;
-
-    //! returns true iff the matrix data is continuous
-    // (i.e. when there are no gaps between successive rows).
-    // similar to CV_IS_MAT_CONT(cvmat->type)
-    bool isContinuous() const;
-
-    //! returns true if the matrix is a submatrix of another matrix
-    bool isSubmatrix() const;
-
-    //! returns element size in bytes,
-    // similar to CV_ELEM_SIZE(cvmat->type)
-    size_t elemSize() const;
-    //! returns the size of element channel in bytes.
-    size_t elemSize1() const;
-    //! returns element type, similar to CV_MAT_TYPE(cvmat->type)
-    int type() const;
-    //! returns element type, similar to CV_MAT_DEPTH(cvmat->type)
-    int depth() const;
-    //! returns element type, similar to CV_MAT_CN(cvmat->type)
-    int channels() const;
-    //! returns step/elemSize1()
-    size_t step1(int i=0) const;
-    //! returns true if matrix data is NULL
-    bool empty() const;
-    //! returns the total number of matrix elements
-    size_t total() const;
-
-    //! returns N if the matrix is 1-channel (N x ptdim) or ptdim-channel (1 x N) or (N x 1); negative number otherwise
-    int checkVector(int elemChannels, int depth=-1, bool requireContinuous=true) const;
-
-    UMat(UMat&& m);
-    UMat& operator = (UMat&& m);
-
-    /*! Returns the OpenCL buffer handle on which UMat operates on.
-        The UMat instance should be kept alive during the use of the handle to prevent the buffer to be
-        returned to the OpenCV buffer pool.
-     */
-    void* handle(AccessFlag accessFlags) const;
-    void ndoffset(size_t* ofs) const;
-
-    enum { MAGIC_VAL  = 0x42FF0000, AUTO_STEP = 0, CONTINUOUS_FLAG = CV_MAT_CONT_FLAG, SUBMATRIX_FLAG = CV_SUBMAT_FLAG };
-    enum { MAGIC_MASK = 0xFFFF0000, TYPE_MASK = 0x00000FFF, DEPTH_MASK = 7 };
-
-    /*! includes several bit-fields:
-         - the magic signature
-         - continuity flag
-         - depth
-         - number of channels
-     */
-    int flags;
-
-    //! the matrix dimensionality, >= 2
-    int dims;
-
-    //! number of rows in the matrix; -1 when the matrix has more than 2 dimensions
-    int rows;
-
-    //! number of columns in the matrix; -1 when the matrix has more than 2 dimensions
-    int cols;
-
-    //! custom allocator
-    MatAllocator* allocator;
-
-    //! usage flags for allocator; recommend do not set directly, instead set during construct/create/getUMat
-    UMatUsageFlags usageFlags;
-
-    //! and the standard allocator
-    static MatAllocator* getStdAllocator();
-
-    //! internal use method: updates the continuity flag
-    void updateContinuityFlag();
-
-    //! black-box container of UMat data
-    UMatData* u;
-
-    //! offset of the submatrix (or 0)
-    size_t offset;
-
-    //! dimensional size of the matrix; accessible in various formats
-    MatSize size;
-
-    //! number of bytes each matrix element/row/plane/dimension occupies
-    MatStep step;
-
-protected:
-};
-
-
-/////////////////////////// multi-dimensional sparse matrix //////////////////////////
-
-/** @brief The class SparseMat represents multi-dimensional sparse numerical arrays.
-
-Such a sparse array can store elements of any type that Mat can store. *Sparse* means that only
-non-zero elements are stored (though, as a result of operations on a sparse matrix, some of its
-stored elements can actually become 0. It is up to you to detect such elements and delete them
-using SparseMat::erase ). The non-zero elements are stored in a hash table that grows when it is
-filled so that the search time is O(1) in average (regardless of whether element is there or not).
-Elements can be accessed using the following methods:
--   Query operations (SparseMat::ptr and the higher-level SparseMat::ref, SparseMat::value and
-    SparseMat::find), for example:
-    @code
-        const int dims = 5;
-        int size[5] = {10, 10, 10, 10, 10};
-        SparseMat sparse_mat(dims, size, CV_32F);
-        for(int i = 0; i < 1000; i++)
-        {
-            int idx[dims];
-            for(int k = 0; k < dims; k++)
-                idx[k] = rand() % size[k];
-            sparse_mat.ref<float>(idx) += 1.f;
-        }
-        cout << "nnz = " << sparse_mat.nzcount() << endl;
-    @endcode
--   Sparse matrix iterators. They are similar to MatIterator but different from NAryMatIterator.
-    That is, the iteration loop is familiar to STL users:
-    @code
-        // prints elements of a sparse floating-point matrix
-        // and the sum of elements.
-        SparseMatConstIterator_<float>
-            it = sparse_mat.begin<float>(),
-            it_end = sparse_mat.end<float>();
-        double s = 0;
-        int dims = sparse_mat.dims();
-        for(; it != it_end; ++it)
-        {
-            // print element indices and the element value
-            const SparseMat::Node* n = it.node();
-            printf("(");
-            for(int i = 0; i < dims; i++)
-                printf("%d%s", n->idx[i], i < dims-1 ? ", " : ")");
-            printf(": %g\n", it.value<float>());
-            s += *it;
-        }
-        printf("Element sum is %g\n", s);
-    @endcode
-    If you run this loop, you will notice that elements are not enumerated in a logical order
-    (lexicographical, and so on). They come in the same order as they are stored in the hash table
-    (semi-randomly). You may collect pointers to the nodes and sort them to get the proper ordering.
-    Note, however, that pointers to the nodes may become invalid when you add more elements to the
-    matrix. This may happen due to possible buffer reallocation.
--   Combination of the above 2 methods when you need to process 2 or more sparse matrices
-    simultaneously. For example, this is how you can compute unnormalized cross-correlation of the 2
-    floating-point sparse matrices:
-    @code
-        double cross_corr(const SparseMat& a, const SparseMat& b)
-        {
-            const SparseMat *_a = &a, *_b = &b;
-            // if b contains less elements than a,
-            // it is faster to iterate through b
-            if(_a->nzcount() > _b->nzcount())
-                std::swap(_a, _b);
-            SparseMatConstIterator_<float> it = _a->begin<float>(),
-                                           it_end = _a->end<float>();
-            double ccorr = 0;
-            for(; it != it_end; ++it)
-            {
-                // take the next element from the first matrix
-                float avalue = *it;
-                const Node* anode = it.node();
-                // and try to find an element with the same index in the second matrix.
-                // since the hash value depends only on the element index,
-                // reuse the hash value stored in the node
-                float bvalue = _b->value<float>(anode->idx,&anode->hashval);
-                ccorr += avalue*bvalue;
-            }
-            return ccorr;
-        }
-    @endcode
- */
-class CV_EXPORTS SparseMat
-{
-public:
-    typedef SparseMatIterator iterator;
-    typedef SparseMatConstIterator const_iterator;
-
-    enum { MAGIC_VAL=0x42FD0000, MAX_DIM=32, HASH_SCALE=0x5bd1e995, HASH_BIT=0x80000000 };
-
-    //! the sparse matrix header
-    struct CV_EXPORTS Hdr
-    {
-        Hdr(int _dims, const int* _sizes, int _type);
-        void clear();
-        int refcount;
-        int dims;
-        int valueOffset;
-        size_t nodeSize;
-        size_t nodeCount;
-        size_t freeList;
-        std::vector<uchar> pool;
-        std::vector<size_t> hashtab;
-        int size[MAX_DIM];
-    };
-
-    //! sparse matrix node - element of a hash table
-    struct CV_EXPORTS Node
-    {
-        //! hash value
-        size_t hashval;
-        //! index of the next node in the same hash table entry
-        size_t next;
-        //! index of the matrix element
-        int idx[MAX_DIM];
-    };
-
-    /** @brief Various SparseMat constructors.
-     */
-    SparseMat();
-
-    /** @overload
-    @param dims Array dimensionality.
-    @param _sizes Sparce matrix size on all dementions.
-    @param _type Sparse matrix data type.
-    */
-    SparseMat(int dims, const int* _sizes, int _type);
-
-    /** @overload
-    @param m Source matrix for copy constructor. If m is dense matrix (ocvMat) then it will be converted
-    to sparse representation.
-    */
-    SparseMat(const SparseMat& m);
-
-    /** @overload
-    @param m Source matrix for copy constructor. If m is dense matrix (ocvMat) then it will be converted
-    to sparse representation.
-    */
-    explicit SparseMat(const Mat& m);
-
-    //! the destructor
-    ~SparseMat();
-
-    //! assignment operator. This is O(1) operation, i.e. no data is copied
-    SparseMat& operator = (const SparseMat& m);
-    //! equivalent to the corresponding constructor
-    SparseMat& operator = (const Mat& m);
-
-    //! creates full copy of the matrix
-    CV_NODISCARD_STD SparseMat clone() const;
-
-    //! copies all the data to the destination matrix. All the previous content of m is erased
-    void copyTo( SparseMat& m ) const;
-    //! converts sparse matrix to dense matrix.
-    void copyTo( Mat& m ) const;
-    //! multiplies all the matrix elements by the specified scale factor alpha and converts the results to the specified data type
-    void convertTo( SparseMat& m, int rtype, double alpha=1 ) const;
-    //! converts sparse matrix to dense n-dim matrix with optional type conversion and scaling.
-    /*!
-        @param [out] m - output matrix; if it does not have a proper size or type before the operation,
-            it is reallocated
-        @param [in] rtype - desired output matrix type or, rather, the depth since the number of channels
-            are the same as the input has; if rtype is negative, the output matrix will have the
-            same type as the input.
-        @param [in] alpha - optional scale factor
-        @param [in] beta - optional delta added to the scaled values
-    */
-    void convertTo( Mat& m, int rtype, double alpha=1, double beta=0 ) const;
-
-    // not used now
-    void assignTo( SparseMat& m, int type=-1 ) const;
-
-    //! reallocates sparse matrix.
-    /*!
-        If the matrix already had the proper size and type,
-        it is simply cleared with clear(), otherwise,
-        the old matrix is released (using release()) and the new one is allocated.
-    */
-    void create(int dims, const int* _sizes, int _type);
-    //! sets all the sparse matrix elements to 0, which means clearing the hash table.
-    void clear();
-    //! manually increments the reference counter to the header.
-    void addref();
-    // decrements the header reference counter. When the counter reaches 0, the header and all the underlying data are deallocated.
-    void release();
-
-    //! converts sparse matrix to the old-style representation; all the elements are copied.
-    //operator CvSparseMat*() const;
-    //! returns the size of each element in bytes (not including the overhead - the space occupied by SparseMat::Node elements)
-    size_t elemSize() const;
-    //! returns elemSize()/channels()
-    size_t elemSize1() const;
-
-    //! returns type of sparse matrix elements
-    int type() const;
-    //! returns the depth of sparse matrix elements
-    int depth() const;
-    //! returns the number of channels
-    int channels() const;
-
-    //! returns the array of sizes, or NULL if the matrix is not allocated
-    const int* size() const;
-    //! returns the size of i-th matrix dimension (or 0)
-    int size(int i) const;
-    //! returns the matrix dimensionality
-    int dims() const;
-    //! returns the number of non-zero elements (=the number of hash table nodes)
-    size_t nzcount() const;
-
-    //! computes the element hash value (1D case)
-    size_t hash(int i0) const;
-    //! computes the element hash value (2D case)
-    size_t hash(int i0, int i1) const;
-    //! computes the element hash value (3D case)
-    size_t hash(int i0, int i1, int i2) const;
-    //! computes the element hash value (nD case)
-    size_t hash(const int* idx) const;
-
-    //!@{
-    /*!
-     specialized variants for 1D, 2D, 3D cases and the generic_type one for n-D case.
-     return pointer to the matrix element.
-      - if the element is there (it's non-zero), the pointer to it is returned
-      - if it's not there and createMissing=false, NULL pointer is returned
-      - if it's not there and createMissing=true, then the new element
-        is created and initialized with 0. Pointer to it is returned
-      - if the optional hashval pointer is not NULL, the element hash value is
-        not computed, but *hashval is taken instead.
-    */
-    //! returns pointer to the specified element (1D case)
-    uchar* ptr(int i0, bool createMissing, size_t* hashval=0);
-    //! returns pointer to the specified element (2D case)
-    uchar* ptr(int i0, int i1, bool createMissing, size_t* hashval=0);
-    //! returns pointer to the specified element (3D case)
-    uchar* ptr(int i0, int i1, int i2, bool createMissing, size_t* hashval=0);
-    //! returns pointer to the specified element (nD case)
-    uchar* ptr(const int* idx, bool createMissing, size_t* hashval=0);
-    //!@}
-
-    //!@{
-    /*!
-     return read-write reference to the specified sparse matrix element.
-
-     `ref<_Tp>(i0,...[,hashval])` is equivalent to `*(_Tp*)ptr(i0,...,true[,hashval])`.
-     The methods always return a valid reference.
-     If the element did not exist, it is created and initialized with 0.
-    */
-    //! returns reference to the specified element (1D case)
-    template<typename _Tp> _Tp& ref(int i0, size_t* hashval=0);
-    //! returns reference to the specified element (2D case)
-    template<typename _Tp> _Tp& ref(int i0, int i1, size_t* hashval=0);
-    //! returns reference to the specified element (3D case)
-    template<typename _Tp> _Tp& ref(int i0, int i1, int i2, size_t* hashval=0);
-    //! returns reference to the specified element (nD case)
-    template<typename _Tp> _Tp& ref(const int* idx, size_t* hashval=0);
-    //!@}
-
-    //!@{
-    /*!
-     return value of the specified sparse matrix element.
-
-     `value<_Tp>(i0,...[,hashval])` is equivalent to
-     @code
-     { const _Tp* p = find<_Tp>(i0,...[,hashval]); return p ? *p : _Tp(); }
-     @endcode
-
-     That is, if the element did not exist, the methods return 0.
-     */
-    //! returns value of the specified element (1D case)
-    template<typename _Tp> _Tp value(int i0, size_t* hashval=0) const;
-    //! returns value of the specified element (2D case)
-    template<typename _Tp> _Tp value(int i0, int i1, size_t* hashval=0) const;
-    //! returns value of the specified element (3D case)
-    template<typename _Tp> _Tp value(int i0, int i1, int i2, size_t* hashval=0) const;
-    //! returns value of the specified element (nD case)
-    template<typename _Tp> _Tp value(const int* idx, size_t* hashval=0) const;
-    //!@}
-
-    //!@{
-    /*!
-     Return pointer to the specified sparse matrix element if it exists
-
-     `find<_Tp>(i0,...[,hashval])` is equivalent to `(_const Tp*)ptr(i0,...false[,hashval])`.
-
-     If the specified element does not exist, the methods return NULL.
-    */
-    //! returns pointer to the specified element (1D case)
-    template<typename _Tp> const _Tp* find(int i0, size_t* hashval=0) const;
-    //! returns pointer to the specified element (2D case)
-    template<typename _Tp> const _Tp* find(int i0, int i1, size_t* hashval=0) const;
-    //! returns pointer to the specified element (3D case)
-    template<typename _Tp> const _Tp* find(int i0, int i1, int i2, size_t* hashval=0) const;
-    //! returns pointer to the specified element (nD case)
-    template<typename _Tp> const _Tp* find(const int* idx, size_t* hashval=0) const;
-    //!@}
-
-    //! erases the specified element (2D case)
-    void erase(int i0, int i1, size_t* hashval=0);
-    //! erases the specified element (3D case)
-    void erase(int i0, int i1, int i2, size_t* hashval=0);
-    //! erases the specified element (nD case)
-    void erase(const int* idx, size_t* hashval=0);
-
-    //!@{
-    /*!
-       return the sparse matrix iterator pointing to the first sparse matrix element
-    */
-    //! returns the sparse matrix iterator at the matrix beginning
-    SparseMatIterator begin();
-    //! returns the sparse matrix iterator at the matrix beginning
-    template<typename _Tp> SparseMatIterator_<_Tp> begin();
-    //! returns the read-only sparse matrix iterator at the matrix beginning
-    SparseMatConstIterator begin() const;
-    //! returns the read-only sparse matrix iterator at the matrix beginning
-    template<typename _Tp> SparseMatConstIterator_<_Tp> begin() const;
-    //!@}
-    /*!
-       return the sparse matrix iterator pointing to the element following the last sparse matrix element
-    */
-    //! returns the sparse matrix iterator at the matrix end
-    SparseMatIterator end();
-    //! returns the read-only sparse matrix iterator at the matrix end
-    SparseMatConstIterator end() const;
-    //! returns the typed sparse matrix iterator at the matrix end
-    template<typename _Tp> SparseMatIterator_<_Tp> end();
-    //! returns the typed read-only sparse matrix iterator at the matrix end
-    template<typename _Tp> SparseMatConstIterator_<_Tp> end() const;
-
-    //! returns the value stored in the sparse martix node
-    template<typename _Tp> _Tp& value(Node* n);
-    //! returns the value stored in the sparse martix node
-    template<typename _Tp> const _Tp& value(const Node* n) const;
-
-    ////////////// some internal-use methods ///////////////
-    Node* node(size_t nidx);
-    const Node* node(size_t nidx) const;
-
-    uchar* newNode(const int* idx, size_t hashval);
-    void removeNode(size_t hidx, size_t nidx, size_t previdx);
-    void resizeHashTab(size_t newsize);
-
-    int flags;
-    Hdr* hdr;
-};
-
-
-
-///////////////////////////////// SparseMat_<_Tp> ////////////////////////////////////
-
-/** @brief Template sparse n-dimensional array class derived from SparseMat
-
-SparseMat_ is a thin wrapper on top of SparseMat created in the same way as Mat_ . It simplifies
-notation of some operations:
-@code
-    int sz[] = {10, 20, 30};
-    SparseMat_<double> M(3, sz);
-    ...
-    M.ref(1, 2, 3) = M(4, 5, 6) + M(7, 8, 9);
-@endcode
- */
-template<typename _Tp> class SparseMat_ : public SparseMat
-{
-public:
-    typedef SparseMatIterator_<_Tp> iterator;
-    typedef SparseMatConstIterator_<_Tp> const_iterator;
-
-    //! the default constructor
-    SparseMat_();
-    //! the full constructor equivalent to SparseMat(dims, _sizes, DataType<_Tp>::type)
-    SparseMat_(int dims, const int* _sizes);
-    //! the copy constructor. If DataType<_Tp>.type != m.type(), the m elements are converted
-    SparseMat_(const SparseMat& m);
-    //! the copy constructor. This is O(1) operation - no data is copied
-    SparseMat_(const SparseMat_& m);
-    //! converts dense matrix to the sparse form
-    SparseMat_(const Mat& m);
-    //! converts the old-style sparse matrix to the C++ class. All the elements are copied
-    //SparseMat_(const CvSparseMat* m);
-    //! the assignment operator. If DataType<_Tp>.type != m.type(), the m elements are converted
-    SparseMat_& operator = (const SparseMat& m);
-    //! the assignment operator. This is O(1) operation - no data is copied
-    SparseMat_& operator = (const SparseMat_& m);
-    //! converts dense matrix to the sparse form
-    SparseMat_& operator = (const Mat& m);
-
-    //! makes full copy of the matrix. All the elements are duplicated
-    CV_NODISCARD_STD SparseMat_ clone() const;
-    //! equivalent to cv::SparseMat::create(dims, _sizes, DataType<_Tp>::type)
-    void create(int dims, const int* _sizes);
-    //! converts sparse matrix to the old-style CvSparseMat. All the elements are copied
-    //operator CvSparseMat*() const;
-
-    //! returns type of the matrix elements
-    int type() const;
-    //! returns depth of the matrix elements
-    int depth() const;
-    //! returns the number of channels in each matrix element
-    int channels() const;
-
-    //! equivalent to SparseMat::ref<_Tp>(i0, hashval)
-    _Tp& ref(int i0, size_t* hashval=0);
-    //! equivalent to SparseMat::ref<_Tp>(i0, i1, hashval)
-    _Tp& ref(int i0, int i1, size_t* hashval=0);
-    //! equivalent to SparseMat::ref<_Tp>(i0, i1, i2, hashval)
-    _Tp& ref(int i0, int i1, int i2, size_t* hashval=0);
-    //! equivalent to SparseMat::ref<_Tp>(idx, hashval)
-    _Tp& ref(const int* idx, size_t* hashval=0);
-
-    //! equivalent to SparseMat::value<_Tp>(i0, hashval)
-    _Tp operator()(int i0, size_t* hashval=0) const;
-    //! equivalent to SparseMat::value<_Tp>(i0, i1, hashval)
-    _Tp operator()(int i0, int i1, size_t* hashval=0) const;
-    //! equivalent to SparseMat::value<_Tp>(i0, i1, i2, hashval)
-    _Tp operator()(int i0, int i1, int i2, size_t* hashval=0) const;
-    //! equivalent to SparseMat::value<_Tp>(idx, hashval)
-    _Tp operator()(const int* idx, size_t* hashval=0) const;
-
-    //! returns sparse matrix iterator pointing to the first sparse matrix element
-    SparseMatIterator_<_Tp> begin();
-    //! returns read-only sparse matrix iterator pointing to the first sparse matrix element
-    SparseMatConstIterator_<_Tp> begin() const;
-    //! returns sparse matrix iterator pointing to the element following the last sparse matrix element
-    SparseMatIterator_<_Tp> end();
-    //! returns read-only sparse matrix iterator pointing to the element following the last sparse matrix element
-    SparseMatConstIterator_<_Tp> end() const;
-};
-
-
-
-////////////////////////////////// MatConstIterator //////////////////////////////////
-
-class CV_EXPORTS MatConstIterator
-{
-public:
-    typedef uchar* value_type;
-    typedef ptrdiff_t difference_type;
-    typedef const uchar** pointer;
-    typedef uchar* reference;
-
-    typedef std::random_access_iterator_tag iterator_category;
-
-    //! default constructor
-    MatConstIterator();
-    //! constructor that sets the iterator to the beginning of the matrix
-    MatConstIterator(const Mat* _m);
-    //! constructor that sets the iterator to the specified element of the matrix
-    MatConstIterator(const Mat* _m, int _row, int _col=0);
-    //! constructor that sets the iterator to the specified element of the matrix
-    MatConstIterator(const Mat* _m, Point _pt);
-    //! constructor that sets the iterator to the specified element of the matrix
-    MatConstIterator(const Mat* _m, const int* _idx);
-    //! copy constructor
-    MatConstIterator(const MatConstIterator& it);
-
-    //! copy operator
-    MatConstIterator& operator = (const MatConstIterator& it);
-    //! returns the current matrix element
-    const uchar* operator *() const;
-    //! returns the i-th matrix element, relative to the current
-    const uchar* operator [](ptrdiff_t i) const;
-
-    //! shifts the iterator forward by the specified number of elements
-    MatConstIterator& operator += (ptrdiff_t ofs);
-    //! shifts the iterator backward by the specified number of elements
-    MatConstIterator& operator -= (ptrdiff_t ofs);
-    //! decrements the iterator
-    MatConstIterator& operator --();
-    //! decrements the iterator
-    MatConstIterator operator --(int);
-    //! increments the iterator
-    MatConstIterator& operator ++();
-    //! increments the iterator
-    MatConstIterator operator ++(int);
-    //! returns the current iterator position
-    Point pos() const;
-    //! returns the current iterator position
-    void pos(int* _idx) const;
-
-    ptrdiff_t lpos() const;
-    void seek(ptrdiff_t ofs, bool relative = false);
-    void seek(const int* _idx, bool relative = false);
-
-    const Mat* m;
-    size_t elemSize;
-    const uchar* ptr;
-    const uchar* sliceStart;
-    const uchar* sliceEnd;
-};
-
-
-
-////////////////////////////////// MatConstIterator_ /////////////////////////////////
-
-/** @brief Matrix read-only iterator
- */
-template<typename _Tp>
-class MatConstIterator_ : public MatConstIterator
-{
-public:
-    typedef _Tp value_type;
-    typedef ptrdiff_t difference_type;
-    typedef const _Tp* pointer;
-    typedef const _Tp& reference;
-
-    typedef std::random_access_iterator_tag iterator_category;
-
-    //! default constructor
-    MatConstIterator_();
-    //! constructor that sets the iterator to the beginning of the matrix
-    MatConstIterator_(const Mat_<_Tp>* _m);
-    //! constructor that sets the iterator to the specified element of the matrix
-    MatConstIterator_(const Mat_<_Tp>* _m, int _row, int _col=0);
-    //! constructor that sets the iterator to the specified element of the matrix
-    MatConstIterator_(const Mat_<_Tp>* _m, Point _pt);
-    //! constructor that sets the iterator to the specified element of the matrix
-    MatConstIterator_(const Mat_<_Tp>* _m, const int* _idx);
-    //! copy constructor
-    MatConstIterator_(const MatConstIterator_& it);
-
-    //! copy operator
-    MatConstIterator_& operator = (const MatConstIterator_& it);
-    //! returns the current matrix element
-    const _Tp& operator *() const;
-    //! returns the i-th matrix element, relative to the current
-    const _Tp& operator [](ptrdiff_t i) const;
-
-    //! shifts the iterator forward by the specified number of elements
-    MatConstIterator_& operator += (ptrdiff_t ofs);
-    //! shifts the iterator backward by the specified number of elements
-    MatConstIterator_& operator -= (ptrdiff_t ofs);
-    //! decrements the iterator
-    MatConstIterator_& operator --();
-    //! decrements the iterator
-    MatConstIterator_ operator --(int);
-    //! increments the iterator
-    MatConstIterator_& operator ++();
-    //! increments the iterator
-    MatConstIterator_ operator ++(int);
-    //! returns the current iterator position
-    Point pos() const;
-};
-
-
-
-//////////////////////////////////// MatIterator_ ////////////////////////////////////
-
-/** @brief Matrix read-write iterator
-*/
-template<typename _Tp>
-class MatIterator_ : public MatConstIterator_<_Tp>
-{
-public:
-    typedef _Tp* pointer;
-    typedef _Tp& reference;
-
-    typedef std::random_access_iterator_tag iterator_category;
-
-    //! the default constructor
-    MatIterator_();
-    //! constructor that sets the iterator to the beginning of the matrix
-    MatIterator_(Mat_<_Tp>* _m);
-    //! constructor that sets the iterator to the specified element of the matrix
-    MatIterator_(Mat_<_Tp>* _m, int _row, int _col=0);
-    //! constructor that sets the iterator to the specified element of the matrix
-    MatIterator_(Mat_<_Tp>* _m, Point _pt);
-    //! constructor that sets the iterator to the specified element of the matrix
-    MatIterator_(Mat_<_Tp>* _m, const int* _idx);
-    //! copy constructor
-    MatIterator_(const MatIterator_& it);
-    //! copy operator
-    MatIterator_& operator = (const MatIterator_<_Tp>& it );
-
-    //! returns the current matrix element
-    _Tp& operator *() const;
-    //! returns the i-th matrix element, relative to the current
-    _Tp& operator [](ptrdiff_t i) const;
-
-    //! shifts the iterator forward by the specified number of elements
-    MatIterator_& operator += (ptrdiff_t ofs);
-    //! shifts the iterator backward by the specified number of elements
-    MatIterator_& operator -= (ptrdiff_t ofs);
-    //! decrements the iterator
-    MatIterator_& operator --();
-    //! decrements the iterator
-    MatIterator_ operator --(int);
-    //! increments the iterator
-    MatIterator_& operator ++();
-    //! increments the iterator
-    MatIterator_ operator ++(int);
-};
-
-
-
-/////////////////////////////// SparseMatConstIterator ///////////////////////////////
-
-/**  @brief Read-Only Sparse Matrix Iterator.
-
- Here is how to use the iterator to compute the sum of floating-point sparse matrix elements:
-
- \code
- SparseMatConstIterator it = m.begin(), it_end = m.end();
- double s = 0;
- CV_Assert( m.type() == CV_32F );
- for( ; it != it_end; ++it )
-    s += it.value<float>();
- \endcode
-*/
-class CV_EXPORTS SparseMatConstIterator
-{
-public:
-    //! the default constructor
-    SparseMatConstIterator();
-    //! the full constructor setting the iterator to the first sparse matrix element
-    SparseMatConstIterator(const SparseMat* _m);
-    //! the copy constructor
-    SparseMatConstIterator(const SparseMatConstIterator& it);
-
-    //! the assignment operator
-    SparseMatConstIterator& operator = (const SparseMatConstIterator& it);
-
-    //! template method returning the current matrix element
-    template<typename _Tp> const _Tp& value() const;
-    //! returns the current node of the sparse matrix. it.node->idx is the current element index
-    const SparseMat::Node* node() const;
-
-    //! moves iterator to the previous element
-    SparseMatConstIterator& operator --();
-    //! moves iterator to the previous element
-    SparseMatConstIterator operator --(int);
-    //! moves iterator to the next element
-    SparseMatConstIterator& operator ++();
-    //! moves iterator to the next element
-    SparseMatConstIterator operator ++(int);
-
-    //! moves iterator to the element after the last element
-    void seekEnd();
-
-    const SparseMat* m;
-    size_t hashidx;
-    uchar* ptr;
-};
-
-
-
-////////////////////////////////// SparseMatIterator /////////////////////////////////
-
-/** @brief  Read-write Sparse Matrix Iterator
-
- The class is similar to cv::SparseMatConstIterator,
- but can be used for in-place modification of the matrix elements.
-*/
-class CV_EXPORTS SparseMatIterator : public SparseMatConstIterator
-{
-public:
-    //! the default constructor
-    SparseMatIterator();
-    //! the full constructor setting the iterator to the first sparse matrix element
-    SparseMatIterator(SparseMat* _m);
-    //! the full constructor setting the iterator to the specified sparse matrix element
-    SparseMatIterator(SparseMat* _m, const int* idx);
-    //! the copy constructor
-    SparseMatIterator(const SparseMatIterator& it);
-
-    //! the assignment operator
-    SparseMatIterator& operator = (const SparseMatIterator& it);
-    //! returns read-write reference to the current sparse matrix element
-    template<typename _Tp> _Tp& value() const;
-    //! returns pointer to the current sparse matrix node. it.node->idx is the index of the current element (do not modify it!)
-    SparseMat::Node* node() const;
-
-    //! moves iterator to the next element
-    SparseMatIterator& operator ++();
-    //! moves iterator to the next element
-    SparseMatIterator operator ++(int);
-};
-
-
-
-/////////////////////////////// SparseMatConstIterator_ //////////////////////////////
-
-/** @brief  Template Read-Only Sparse Matrix Iterator Class.
-
- This is the derived from SparseMatConstIterator class that
- introduces more convenient operator *() for accessing the current element.
-*/
-template<typename _Tp> class SparseMatConstIterator_ : public SparseMatConstIterator
-{
-public:
-
-    typedef std::forward_iterator_tag iterator_category;
-
-    //! the default constructor
-    SparseMatConstIterator_();
-    //! the full constructor setting the iterator to the first sparse matrix element
-    SparseMatConstIterator_(const SparseMat_<_Tp>* _m);
-    SparseMatConstIterator_(const SparseMat* _m);
-    //! the copy constructor
-    SparseMatConstIterator_(const SparseMatConstIterator_& it);
-
-    //! the assignment operator
-    SparseMatConstIterator_& operator = (const SparseMatConstIterator_& it);
-    //! the element access operator
-    const _Tp& operator *() const;
-
-    //! moves iterator to the next element
-    SparseMatConstIterator_& operator ++();
-    //! moves iterator to the next element
-    SparseMatConstIterator_ operator ++(int);
-};
-
-
-
-///////////////////////////////// SparseMatIterator_ /////////////////////////////////
-
-/** @brief  Template Read-Write Sparse Matrix Iterator Class.
-
- This is the derived from cv::SparseMatConstIterator_ class that
- introduces more convenient operator *() for accessing the current element.
-*/
-template<typename _Tp> class SparseMatIterator_ : public SparseMatConstIterator_<_Tp>
-{
-public:
-
-    typedef std::forward_iterator_tag iterator_category;
-
-    //! the default constructor
-    SparseMatIterator_();
-    //! the full constructor setting the iterator to the first sparse matrix element
-    SparseMatIterator_(SparseMat_<_Tp>* _m);
-    SparseMatIterator_(SparseMat* _m);
-    //! the copy constructor
-    SparseMatIterator_(const SparseMatIterator_& it);
-
-    //! the assignment operator
-    SparseMatIterator_& operator = (const SparseMatIterator_& it);
-    //! returns the reference to the current element
-    _Tp& operator *() const;
-
-    //! moves the iterator to the next element
-    SparseMatIterator_& operator ++();
-    //! moves the iterator to the next element
-    SparseMatIterator_ operator ++(int);
-};
-
-
-
-/////////////////////////////////// NAryMatIterator //////////////////////////////////
-
-/** @brief n-ary multi-dimensional array iterator.
-
-Use the class to implement unary, binary, and, generally, n-ary element-wise operations on
-multi-dimensional arrays. Some of the arguments of an n-ary function may be continuous arrays, some
-may be not. It is possible to use conventional MatIterator 's for each array but incrementing all of
-the iterators after each small operations may be a big overhead. In this case consider using
-NAryMatIterator to iterate through several matrices simultaneously as long as they have the same
-geometry (dimensionality and all the dimension sizes are the same). On each iteration `it.planes[0]`,
-`it.planes[1]`,... will be the slices of the corresponding matrices.
-
-The example below illustrates how you can compute a normalized and threshold 3D color histogram:
-@code
-    void computeNormalizedColorHist(const Mat& image, Mat& hist, int N, double minProb)
-    {
-        const int histSize[] = {N, N, N};
-
-        // make sure that the histogram has a proper size and type
-        hist.create(3, histSize, CV_32F);
-
-        // and clear it
-        hist = Scalar(0);
-
-        // the loop below assumes that the image
-        // is a 8-bit 3-channel. check it.
-        CV_Assert(image.type() == CV_8UC3);
-        MatConstIterator_<Vec3b> it = image.begin<Vec3b>(),
-                                 it_end = image.end<Vec3b>();
-        for( ; it != it_end; ++it )
-        {
-            const Vec3b& pix = *it;
-            hist.at<float>(pix[0]*N/256, pix[1]*N/256, pix[2]*N/256) += 1.f;
-        }
-
-        minProb *= image.rows*image.cols;
-
-        // initialize iterator (the style is different from STL).
-        // after initialization the iterator will contain
-        // the number of slices or planes the iterator will go through.
-        // it simultaneously increments iterators for several matrices
-        // supplied as a null terminated list of pointers
-        const Mat* arrays[] = {&hist, 0};
-        Mat planes[1];
-        NAryMatIterator itNAry(arrays, planes, 1);
-        double s = 0;
-        // iterate through the matrix. on each iteration
-        // itNAry.planes[i] (of type Mat) will be set to the current plane
-        // of the i-th n-dim matrix passed to the iterator constructor.
-        for(int p = 0; p < itNAry.nplanes; p++, ++itNAry)
-        {
-            threshold(itNAry.planes[0], itNAry.planes[0], minProb, 0, THRESH_TOZERO);
-            s += sum(itNAry.planes[0])[0];
-        }
-
-        s = 1./s;
-        itNAry = NAryMatIterator(arrays, planes, 1);
-        for(int p = 0; p < itNAry.nplanes; p++, ++itNAry)
-            itNAry.planes[0] *= s;
-    }
-@endcode
- */
-class CV_EXPORTS NAryMatIterator
-{
-public:
-    //! the default constructor
-    NAryMatIterator();
-    //! the full constructor taking arbitrary number of n-dim matrices
-    NAryMatIterator(const Mat** arrays, uchar** ptrs, int narrays=-1);
-    //! the full constructor taking arbitrary number of n-dim matrices
-    NAryMatIterator(const Mat** arrays, Mat* planes, int narrays=-1);
-    //! the separate iterator initialization method
-    void init(const Mat** arrays, Mat* planes, uchar** ptrs, int narrays=-1);
-
-    //! proceeds to the next plane of every iterated matrix
-    NAryMatIterator& operator ++();
-    //! proceeds to the next plane of every iterated matrix (postfix increment operator)
-    NAryMatIterator operator ++(int);
-
-    //! the iterated arrays
-    const Mat** arrays;
-    //! the current planes
-    Mat* planes;
-    //! data pointers
-    uchar** ptrs;
-    //! the number of arrays
-    int narrays;
-    //! the number of hyper-planes that the iterator steps through
-    size_t nplanes;
-    //! the size of each segment (in elements)
-    size_t size;
-protected:
-    int iterdepth;
-    size_t idx;
-};
-
-
-
-///////////////////////////////// Matrix Expressions /////////////////////////////////
-
-class CV_EXPORTS MatOp
-{
-public:
-    MatOp();
-    virtual ~MatOp();
-
-    virtual bool elementWise(const MatExpr& expr) const;
-    virtual void assign(const MatExpr& expr, Mat& m, int type=-1) const = 0;
-    virtual void roi(const MatExpr& expr, const Range& rowRange,
-                     const Range& colRange, MatExpr& res) const;
-    virtual void diag(const MatExpr& expr, int d, MatExpr& res) const;
-    virtual void augAssignAdd(const MatExpr& expr, Mat& m) const;
-    virtual void augAssignSubtract(const MatExpr& expr, Mat& m) const;
-    virtual void augAssignMultiply(const MatExpr& expr, Mat& m) const;
-    virtual void augAssignDivide(const MatExpr& expr, Mat& m) const;
-    virtual void augAssignAnd(const MatExpr& expr, Mat& m) const;
-    virtual void augAssignOr(const MatExpr& expr, Mat& m) const;
-    virtual void augAssignXor(const MatExpr& expr, Mat& m) const;
-
-    virtual void add(const MatExpr& expr1, const MatExpr& expr2, MatExpr& res) const;
-    virtual void add(const MatExpr& expr1, const Scalar& s, MatExpr& res) const;
-
-    virtual void subtract(const MatExpr& expr1, const MatExpr& expr2, MatExpr& res) const;
-    virtual void subtract(const Scalar& s, const MatExpr& expr, MatExpr& res) const;
-
-    virtual void multiply(const MatExpr& expr1, const MatExpr& expr2, MatExpr& res, double scale=1) const;
-    virtual void multiply(const MatExpr& expr1, double s, MatExpr& res) const;
-
-    virtual void divide(const MatExpr& expr1, const MatExpr& expr2, MatExpr& res, double scale=1) const;
-    virtual void divide(double s, const MatExpr& expr, MatExpr& res) const;
-
-    virtual void abs(const MatExpr& expr, MatExpr& res) const;
-
-    virtual void transpose(const MatExpr& expr, MatExpr& res) const;
-    virtual void matmul(const MatExpr& expr1, const MatExpr& expr2, MatExpr& res) const;
-    virtual void invert(const MatExpr& expr, int method, MatExpr& res) const;
-
-    virtual Size size(const MatExpr& expr) const;
-    virtual int type(const MatExpr& expr) const;
-};
-
-/** @brief Matrix expression representation
-@anchor MatrixExpressions
-This is a list of implemented matrix operations that can be combined in arbitrary complex
-expressions (here A, B stand for matrices ( Mat ), s for a scalar ( Scalar ), alpha for a
-real-valued scalar ( double )):
--   Addition, subtraction, negation: `A+B`, `A-B`, `A+s`, `A-s`, `s+A`, `s-A`, `-A`
--   Scaling: `A*alpha`
--   Per-element multiplication and division: `A.mul(B)`, `A/B`, `alpha/A`
--   Matrix multiplication: `A*B`
--   Transposition: `A.t()` (means A<sup>T</sup>)
--   Matrix inversion and pseudo-inversion, solving linear systems and least-squares problems:
-    `A.inv([method]) (~ A<sup>-1</sup>)`,   `A.inv([method])*B (~ X: AX=B)`
--   Comparison: `A cmpop B`, `A cmpop alpha`, `alpha cmpop A`, where *cmpop* is one of
-  `>`, `>=`, `==`, `!=`, `<=`, `<`. The result of comparison is an 8-bit single channel mask whose
-    elements are set to 255 (if the particular element or pair of elements satisfy the condition) or
-    0.
--   Bitwise logical operations: `A logicop B`, `A logicop s`, `s logicop A`, `~A`, where *logicop* is one of
-  `&`, `|`, `^`.
--   Element-wise minimum and maximum: `min(A, B)`, `min(A, alpha)`, `max(A, B)`, `max(A, alpha)`
--   Element-wise absolute value: `abs(A)`
--   Cross-product, dot-product: `A.cross(B)`, `A.dot(B)`
--   Any function of matrix or matrices and scalars that returns a matrix or a scalar, such as norm,
-    mean, sum, countNonZero, trace, determinant, repeat, and others.
--   Matrix initializers ( Mat::eye(), Mat::zeros(), Mat::ones() ), matrix comma-separated
-    initializers, matrix constructors and operators that extract sub-matrices (see Mat description).
--   Mat_<destination_type>() constructors to cast the result to the proper type.
-@note Comma-separated initializers and probably some other operations may require additional
-explicit Mat() or Mat_<T>() constructor calls to resolve a possible ambiguity.
-
-Here are examples of matrix expressions:
-@code
-    // compute pseudo-inverse of A, equivalent to A.inv(DECOMP_SVD)
-    SVD svd(A);
-    Mat pinvA = svd.vt.t()*Mat::diag(1./svd.w)*svd.u.t();
-
-    // compute the new vector of parameters in the Levenberg-Marquardt algorithm
-    x -= (A.t()*A + lambda*Mat::eye(A.cols,A.cols,A.type())).inv(DECOMP_CHOLESKY)*(A.t()*err);
-
-    // sharpen image using "unsharp mask" algorithm
-    Mat blurred; double sigma = 1, threshold = 5, amount = 1;
-    GaussianBlur(img, blurred, Size(), sigma, sigma);
-    Mat lowContrastMask = abs(img - blurred) < threshold;
-    Mat sharpened = img*(1+amount) + blurred*(-amount);
-    img.copyTo(sharpened, lowContrastMask);
-@endcode
-*/
-class CV_EXPORTS MatExpr
-{
-public:
-    MatExpr();
-    explicit MatExpr(const Mat& m);
-
-    MatExpr(const MatOp* _op, int _flags, const Mat& _a = Mat(), const Mat& _b = Mat(),
-            const Mat& _c = Mat(), double _alpha = 1, double _beta = 1, const Scalar& _s = Scalar());
-
-    operator Mat() const;
-    template<typename _Tp> operator Mat_<_Tp>() const;
-
-    Size size() const;
-    int type() const;
-
-    MatExpr row(int y) const;
-    MatExpr col(int x) const;
-    MatExpr diag(int d = 0) const;
-    MatExpr operator()( const Range& rowRange, const Range& colRange ) const;
-    MatExpr operator()( const Rect& roi ) const;
-
-    MatExpr t() const;
-    MatExpr inv(int method = DECOMP_LU) const;
-    MatExpr mul(const MatExpr& e, double scale=1) const;
-    MatExpr mul(const Mat& m, double scale=1) const;
-
-    Mat cross(const Mat& m) const;
-    double dot(const Mat& m) const;
-
-    void swap(MatExpr& b);
-
-    const MatOp* op;
-    int flags;
-
-    Mat a, b, c;
-    double alpha, beta;
-    Scalar s;
-};
-
-//! @} core_basic
-
-//! @relates cv::MatExpr
-//! @{
-CV_EXPORTS MatExpr operator + (const Mat& a, const Mat& b);
-CV_EXPORTS MatExpr operator + (const Mat& a, const Scalar& s);
-CV_EXPORTS MatExpr operator + (const Scalar& s, const Mat& a);
-CV_EXPORTS MatExpr operator + (const MatExpr& e, const Mat& m);
-CV_EXPORTS MatExpr operator + (const Mat& m, const MatExpr& e);
-CV_EXPORTS MatExpr operator + (const MatExpr& e, const Scalar& s);
-CV_EXPORTS MatExpr operator + (const Scalar& s, const MatExpr& e);
-CV_EXPORTS MatExpr operator + (const MatExpr& e1, const MatExpr& e2);
-template<typename _Tp, int m, int n> static inline
-MatExpr operator + (const Mat& a, const Matx<_Tp, m, n>& b) { return a + Mat(b); }
-template<typename _Tp, int m, int n> static inline
-MatExpr operator + (const Matx<_Tp, m, n>& a, const Mat& b) { return Mat(a) + b; }
-
-CV_EXPORTS MatExpr operator - (const Mat& a, const Mat& b);
-CV_EXPORTS MatExpr operator - (const Mat& a, const Scalar& s);
-CV_EXPORTS MatExpr operator - (const Scalar& s, const Mat& a);
-CV_EXPORTS MatExpr operator - (const MatExpr& e, const Mat& m);
-CV_EXPORTS MatExpr operator - (const Mat& m, const MatExpr& e);
-CV_EXPORTS MatExpr operator - (const MatExpr& e, const Scalar& s);
-CV_EXPORTS MatExpr operator - (const Scalar& s, const MatExpr& e);
-CV_EXPORTS MatExpr operator - (const MatExpr& e1, const MatExpr& e2);
-template<typename _Tp, int m, int n> static inline
-MatExpr operator - (const Mat& a, const Matx<_Tp, m, n>& b) { return a - Mat(b); }
-template<typename _Tp, int m, int n> static inline
-MatExpr operator - (const Matx<_Tp, m, n>& a, const Mat& b) { return Mat(a) - b; }
-
-CV_EXPORTS MatExpr operator - (const Mat& m);
-CV_EXPORTS MatExpr operator - (const MatExpr& e);
-
-CV_EXPORTS MatExpr operator * (const Mat& a, const Mat& b);
-CV_EXPORTS MatExpr operator * (const Mat& a, double s);
-CV_EXPORTS MatExpr operator * (double s, const Mat& a);
-CV_EXPORTS MatExpr operator * (const MatExpr& e, const Mat& m);
-CV_EXPORTS MatExpr operator * (const Mat& m, const MatExpr& e);
-CV_EXPORTS MatExpr operator * (const MatExpr& e, double s);
-CV_EXPORTS MatExpr operator * (double s, const MatExpr& e);
-CV_EXPORTS MatExpr operator * (const MatExpr& e1, const MatExpr& e2);
-template<typename _Tp, int m, int n> static inline
-MatExpr operator * (const Mat& a, const Matx<_Tp, m, n>& b) { return a * Mat(b); }
-template<typename _Tp, int m, int n> static inline
-MatExpr operator * (const Matx<_Tp, m, n>& a, const Mat& b) { return Mat(a) * b; }
-
-CV_EXPORTS MatExpr operator / (const Mat& a, const Mat& b);
-CV_EXPORTS MatExpr operator / (const Mat& a, double s);
-CV_EXPORTS MatExpr operator / (double s, const Mat& a);
-CV_EXPORTS MatExpr operator / (const MatExpr& e, const Mat& m);
-CV_EXPORTS MatExpr operator / (const Mat& m, const MatExpr& e);
-CV_EXPORTS MatExpr operator / (const MatExpr& e, double s);
-CV_EXPORTS MatExpr operator / (double s, const MatExpr& e);
-CV_EXPORTS MatExpr operator / (const MatExpr& e1, const MatExpr& e2);
-template<typename _Tp, int m, int n> static inline
-MatExpr operator / (const Mat& a, const Matx<_Tp, m, n>& b) { return a / Mat(b); }
-template<typename _Tp, int m, int n> static inline
-MatExpr operator / (const Matx<_Tp, m, n>& a, const Mat& b) { return Mat(a) / b; }
-
-CV_EXPORTS MatExpr operator < (const Mat& a, const Mat& b);
-CV_EXPORTS MatExpr operator < (const Mat& a, double s);
-CV_EXPORTS MatExpr operator < (double s, const Mat& a);
-template<typename _Tp, int m, int n> static inline
-MatExpr operator < (const Mat& a, const Matx<_Tp, m, n>& b) { return a < Mat(b); }
-template<typename _Tp, int m, int n> static inline
-MatExpr operator < (const Matx<_Tp, m, n>& a, const Mat& b) { return Mat(a) < b; }
-
-CV_EXPORTS MatExpr operator <= (const Mat& a, const Mat& b);
-CV_EXPORTS MatExpr operator <= (const Mat& a, double s);
-CV_EXPORTS MatExpr operator <= (double s, const Mat& a);
-template<typename _Tp, int m, int n> static inline
-MatExpr operator <= (const Mat& a, const Matx<_Tp, m, n>& b) { return a <= Mat(b); }
-template<typename _Tp, int m, int n> static inline
-MatExpr operator <= (const Matx<_Tp, m, n>& a, const Mat& b) { return Mat(a) <= b; }
-
-CV_EXPORTS MatExpr operator == (const Mat& a, const Mat& b);
-CV_EXPORTS MatExpr operator == (const Mat& a, double s);
-CV_EXPORTS MatExpr operator == (double s, const Mat& a);
-template<typename _Tp, int m, int n> static inline
-MatExpr operator == (const Mat& a, const Matx<_Tp, m, n>& b) { return a == Mat(b); }
-template<typename _Tp, int m, int n> static inline
-MatExpr operator == (const Matx<_Tp, m, n>& a, const Mat& b) { return Mat(a) == b; }
-
-CV_EXPORTS MatExpr operator != (const Mat& a, const Mat& b);
-CV_EXPORTS MatExpr operator != (const Mat& a, double s);
-CV_EXPORTS MatExpr operator != (double s, const Mat& a);
-template<typename _Tp, int m, int n> static inline
-MatExpr operator != (const Mat& a, const Matx<_Tp, m, n>& b) { return a != Mat(b); }
-template<typename _Tp, int m, int n> static inline
-MatExpr operator != (const Matx<_Tp, m, n>& a, const Mat& b) { return Mat(a) != b; }
-
-CV_EXPORTS MatExpr operator >= (const Mat& a, const Mat& b);
-CV_EXPORTS MatExpr operator >= (const Mat& a, double s);
-CV_EXPORTS MatExpr operator >= (double s, const Mat& a);
-template<typename _Tp, int m, int n> static inline
-MatExpr operator >= (const Mat& a, const Matx<_Tp, m, n>& b) { return a >= Mat(b); }
-template<typename _Tp, int m, int n> static inline
-MatExpr operator >= (const Matx<_Tp, m, n>& a, const Mat& b) { return Mat(a) >= b; }
-
-CV_EXPORTS MatExpr operator > (const Mat& a, const Mat& b);
-CV_EXPORTS MatExpr operator > (const Mat& a, double s);
-CV_EXPORTS MatExpr operator > (double s, const Mat& a);
-template<typename _Tp, int m, int n> static inline
-MatExpr operator > (const Mat& a, const Matx<_Tp, m, n>& b) { return a > Mat(b); }
-template<typename _Tp, int m, int n> static inline
-MatExpr operator > (const Matx<_Tp, m, n>& a, const Mat& b) { return Mat(a) > b; }
-
-CV_EXPORTS MatExpr operator & (const Mat& a, const Mat& b);
-CV_EXPORTS MatExpr operator & (const Mat& a, const Scalar& s);
-CV_EXPORTS MatExpr operator & (const Scalar& s, const Mat& a);
-template<typename _Tp, int m, int n> static inline
-MatExpr operator & (const Mat& a, const Matx<_Tp, m, n>& b) { return a & Mat(b); }
-template<typename _Tp, int m, int n> static inline
-MatExpr operator & (const Matx<_Tp, m, n>& a, const Mat& b) { return Mat(a) & b; }
-
-CV_EXPORTS MatExpr operator | (const Mat& a, const Mat& b);
-CV_EXPORTS MatExpr operator | (const Mat& a, const Scalar& s);
-CV_EXPORTS MatExpr operator | (const Scalar& s, const Mat& a);
-template<typename _Tp, int m, int n> static inline
-MatExpr operator | (const Mat& a, const Matx<_Tp, m, n>& b) { return a | Mat(b); }
-template<typename _Tp, int m, int n> static inline
-MatExpr operator | (const Matx<_Tp, m, n>& a, const Mat& b) { return Mat(a) | b; }
-
-CV_EXPORTS MatExpr operator ^ (const Mat& a, const Mat& b);
-CV_EXPORTS MatExpr operator ^ (const Mat& a, const Scalar& s);
-CV_EXPORTS MatExpr operator ^ (const Scalar& s, const Mat& a);
-template<typename _Tp, int m, int n> static inline
-MatExpr operator ^ (const Mat& a, const Matx<_Tp, m, n>& b) { return a ^ Mat(b); }
-template<typename _Tp, int m, int n> static inline
-MatExpr operator ^ (const Matx<_Tp, m, n>& a, const Mat& b) { return Mat(a) ^ b; }
-
-CV_EXPORTS MatExpr operator ~(const Mat& m);
-
-CV_EXPORTS MatExpr min(const Mat& a, const Mat& b);
-CV_EXPORTS MatExpr min(const Mat& a, double s);
-CV_EXPORTS MatExpr min(double s, const Mat& a);
-template<typename _Tp, int m, int n> static inline
-MatExpr min (const Mat& a, const Matx<_Tp, m, n>& b) { return min(a, Mat(b)); }
-template<typename _Tp, int m, int n> static inline
-MatExpr min (const Matx<_Tp, m, n>& a, const Mat& b) { return min(Mat(a), b); }
-
-CV_EXPORTS MatExpr max(const Mat& a, const Mat& b);
-CV_EXPORTS MatExpr max(const Mat& a, double s);
-CV_EXPORTS MatExpr max(double s, const Mat& a);
-template<typename _Tp, int m, int n> static inline
-MatExpr max (const Mat& a, const Matx<_Tp, m, n>& b) { return max(a, Mat(b)); }
-template<typename _Tp, int m, int n> static inline
-MatExpr max (const Matx<_Tp, m, n>& a, const Mat& b) { return max(Mat(a), b); }
-
-/** @brief Calculates an absolute value of each matrix element.
-
-abs is a meta-function that is expanded to one of absdiff or convertScaleAbs forms:
-- C = abs(A-B) is equivalent to `absdiff(A, B, C)`
-- C = abs(A) is equivalent to `absdiff(A, Scalar::all(0), C)`
-- C = `Mat_<Vec<uchar,n> >(abs(A*alpha + beta))` is equivalent to `convertScaleAbs(A, C, alpha,
-beta)`
-
-The output matrix has the same size and the same type as the input one except for the last case,
-where C is depth=CV_8U .
-@param m matrix.
-@sa @ref MatrixExpressions, absdiff, convertScaleAbs
- */
-CV_EXPORTS MatExpr abs(const Mat& m);
-/** @overload
-@param e matrix expression.
-*/
-CV_EXPORTS MatExpr abs(const MatExpr& e);
-//! @} relates cv::MatExpr
-
-} // cv
-
-#include "opencv2/core/mat.inl.hpp"
-
-#endif // OPENCV_CORE_MAT_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/mat.inl.hpp b/projects/llm_framework/include/opencv4/opencv2/core/mat.inl.hpp
deleted file mode 100644
index 886b82c6..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/mat.inl.hpp
+++ /dev/null
@@ -1,3422 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Copyright (C) 2015, Itseez Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_MATRIX_OPERATIONS_HPP
-#define OPENCV_CORE_MATRIX_OPERATIONS_HPP
-
-#ifndef __cplusplus
-#  error mat.inl.hpp header must be compiled as C++
-#endif
-
-#ifdef _MSC_VER
-#pragma warning( push )
-#pragma warning( disable: 4127 )
-#endif
-
-#if defined(CV_SKIP_DISABLE_CLANG_ENUM_WARNINGS)
-  // nothing
-#elif defined(CV_FORCE_DISABLE_CLANG_ENUM_WARNINGS)
-  #define CV_DISABLE_CLANG_ENUM_WARNINGS
-#elif defined(__clang__) && defined(__has_warning)
-  #if __has_warning("-Wdeprecated-enum-enum-conversion") && __has_warning("-Wdeprecated-anon-enum-enum-conversion")
-    #define CV_DISABLE_CLANG_ENUM_WARNINGS
-  #endif
-#endif
-#ifdef CV_DISABLE_CLANG_ENUM_WARNINGS
-#pragma clang diagnostic push
-#pragma clang diagnostic ignored "-Wdeprecated-enum-enum-conversion"
-#pragma clang diagnostic ignored "-Wdeprecated-anon-enum-enum-conversion"
-#endif
-
-namespace cv
-{
-CV__DEBUG_NS_BEGIN
-
-
-//! @cond IGNORED
-
-////////////////////////// Custom (raw) type wrapper //////////////////////////
-
-template<typename _Tp> static inline
-int rawType()
-{
-    CV_StaticAssert(sizeof(_Tp) <= CV_CN_MAX, "sizeof(_Tp) is too large");
-    const int elemSize = sizeof(_Tp);
-    return (int)CV_MAKETYPE(CV_8U, elemSize);
-}
-
-//////////////////////// Input/Output Arrays ////////////////////////
-
-inline void _InputArray::init(int _flags, const void* _obj)
-{ flags = _flags; obj = (void*)_obj; }
-
-inline void _InputArray::init(int _flags, const void* _obj, Size _sz)
-{ flags = _flags; obj = (void*)_obj; sz = _sz; }
-
-inline void* _InputArray::getObj() const { return obj; }
-inline int _InputArray::getFlags() const { return flags; }
-inline Size _InputArray::getSz() const { return sz; }
-
-inline _InputArray::_InputArray() { init(0 + NONE, 0); }
-inline _InputArray::_InputArray(int _flags, void* _obj) { init(_flags, _obj); }
-inline _InputArray::_InputArray(const Mat& m) { init(MAT+ACCESS_READ, &m); }
-inline _InputArray::_InputArray(const std::vector<Mat>& vec) { init(STD_VECTOR_MAT+ACCESS_READ, &vec); }
-inline _InputArray::_InputArray(const UMat& m) { init(UMAT+ACCESS_READ, &m); }
-inline _InputArray::_InputArray(const std::vector<UMat>& vec) { init(STD_VECTOR_UMAT+ACCESS_READ, &vec); }
-
-template<typename _Tp> inline
-_InputArray::_InputArray(const std::vector<_Tp>& vec)
-{ init(FIXED_TYPE + STD_VECTOR + traits::Type<_Tp>::value + ACCESS_READ, &vec); }
-
-template<typename _Tp, std::size_t _Nm> inline
-_InputArray::_InputArray(const std::array<_Tp, _Nm>& arr)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_READ, arr.data(), Size(1, _Nm)); }
-
-template<std::size_t _Nm> inline
-_InputArray::_InputArray(const std::array<Mat, _Nm>& arr)
-{ init(STD_ARRAY_MAT + ACCESS_READ, arr.data(), Size(1, _Nm)); }
-
-inline
-_InputArray::_InputArray(const std::vector<bool>& vec)
-{ init(FIXED_TYPE + STD_BOOL_VECTOR + traits::Type<bool>::value + ACCESS_READ, &vec); }
-
-template<typename _Tp> inline
-_InputArray::_InputArray(const std::vector<std::vector<_Tp> >& vec)
-{ init(FIXED_TYPE + STD_VECTOR_VECTOR + traits::Type<_Tp>::value + ACCESS_READ, &vec); }
-
-template<typename _Tp> inline
-_InputArray::_InputArray(const std::vector<Mat_<_Tp> >& vec)
-{ init(FIXED_TYPE + STD_VECTOR_MAT + traits::Type<_Tp>::value + ACCESS_READ, &vec); }
-
-template<typename _Tp, int m, int n> inline
-_InputArray::_InputArray(const Matx<_Tp, m, n>& mtx)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_READ, &mtx, Size(n, m)); }
-
-template<typename _Tp> inline
-_InputArray::_InputArray(const _Tp* vec, int n)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_READ, vec, Size(n, 1)); }
-
-template<typename _Tp> inline
-_InputArray::_InputArray(const Mat_<_Tp>& m)
-{ init(FIXED_TYPE + MAT + traits::Type<_Tp>::value + ACCESS_READ, &m); }
-
-inline _InputArray::_InputArray(const double& val)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + CV_64F + ACCESS_READ, &val, Size(1,1)); }
-
-inline _InputArray::_InputArray(const cuda::GpuMat& d_mat)
-{ init(CUDA_GPU_MAT + ACCESS_READ, &d_mat); }
-
-inline _InputArray::_InputArray(const std::vector<cuda::GpuMat>& d_mat)
-{	init(STD_VECTOR_CUDA_GPU_MAT + ACCESS_READ, &d_mat);}
-
-inline _InputArray::_InputArray(const ogl::Buffer& buf)
-{ init(OPENGL_BUFFER + ACCESS_READ, &buf); }
-
-inline _InputArray::_InputArray(const cuda::HostMem& cuda_mem)
-{ init(CUDA_HOST_MEM + ACCESS_READ, &cuda_mem); }
-
-template<typename _Tp> inline
-_InputArray _InputArray::rawIn(const std::vector<_Tp>& vec)
-{
-    _InputArray v;
-    v.flags = _InputArray::FIXED_TYPE + _InputArray::STD_VECTOR + rawType<_Tp>() + ACCESS_READ;
-    v.obj = (void*)&vec;
-    return v;
-}
-
-template<typename _Tp, std::size_t _Nm> inline
-_InputArray _InputArray::rawIn(const std::array<_Tp, _Nm>& arr)
-{
-    _InputArray v;
-    v.flags = FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_READ;
-    v.obj = (void*)arr.data();
-    v.sz = Size(1, _Nm);
-    return v;
-}
-
-inline _InputArray::~_InputArray() {}
-
-inline Mat _InputArray::getMat(int i) const
-{
-    if( kind() == MAT && i < 0 )
-        return *(const Mat*)obj;
-    return getMat_(i);
-}
-
-inline bool _InputArray::isMat() const { return kind() == _InputArray::MAT; }
-inline bool _InputArray::isUMat() const  { return kind() == _InputArray::UMAT; }
-inline bool _InputArray::isMatVector() const { return kind() == _InputArray::STD_VECTOR_MAT; }
-inline bool _InputArray::isUMatVector() const  { return kind() == _InputArray::STD_VECTOR_UMAT; }
-inline bool _InputArray::isMatx() const { return kind() == _InputArray::MATX; }
-inline bool _InputArray::isVector() const { return kind() == _InputArray::STD_VECTOR ||
-                                                   kind() == _InputArray::STD_BOOL_VECTOR ||
-                                                   (kind() == _InputArray::MATX && (sz.width <= 1 || sz.height <= 1)); }
-inline bool _InputArray::isGpuMat() const { return kind() == _InputArray::CUDA_GPU_MAT; }
-inline bool _InputArray::isGpuMatVector() const { return kind() == _InputArray::STD_VECTOR_CUDA_GPU_MAT; }
-
-////////////////////////////////////////////////////////////////////////////////////////
-
-inline _OutputArray::_OutputArray() { init(NONE + ACCESS_WRITE, 0); }
-inline _OutputArray::_OutputArray(int _flags, void* _obj) { init(_flags + ACCESS_WRITE, _obj); }
-inline _OutputArray::_OutputArray(Mat& m) { init(MAT+ACCESS_WRITE, &m); }
-inline _OutputArray::_OutputArray(std::vector<Mat>& vec) { init(STD_VECTOR_MAT + ACCESS_WRITE, &vec); }
-inline _OutputArray::_OutputArray(UMat& m) { init(UMAT + ACCESS_WRITE, &m); }
-inline _OutputArray::_OutputArray(std::vector<UMat>& vec) { init(STD_VECTOR_UMAT + ACCESS_WRITE, &vec); }
-
-template<typename _Tp> inline
-_OutputArray::_OutputArray(std::vector<_Tp>& vec)
-{ init(FIXED_TYPE + STD_VECTOR + traits::Type<_Tp>::value + ACCESS_WRITE, &vec); }
-
-template<typename _Tp, std::size_t _Nm> inline
-_OutputArray::_OutputArray(std::array<_Tp, _Nm>& arr)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_WRITE, arr.data(), Size(1, _Nm)); }
-
-template<std::size_t _Nm> inline
-_OutputArray::_OutputArray(std::array<Mat, _Nm>& arr)
-{ init(STD_ARRAY_MAT + ACCESS_WRITE, arr.data(), Size(1, _Nm)); }
-
-template<typename _Tp> inline
-_OutputArray::_OutputArray(std::vector<std::vector<_Tp> >& vec)
-{ init(FIXED_TYPE + STD_VECTOR_VECTOR + traits::Type<_Tp>::value + ACCESS_WRITE, &vec); }
-
-template<typename _Tp> inline
-_OutputArray::_OutputArray(std::vector<Mat_<_Tp> >& vec)
-{ init(FIXED_TYPE + STD_VECTOR_MAT + traits::Type<_Tp>::value + ACCESS_WRITE, &vec); }
-
-template<typename _Tp> inline
-_OutputArray::_OutputArray(Mat_<_Tp>& m)
-{ init(FIXED_TYPE + MAT + traits::Type<_Tp>::value + ACCESS_WRITE, &m); }
-
-template<typename _Tp, int m, int n> inline
-_OutputArray::_OutputArray(Matx<_Tp, m, n>& mtx)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_WRITE, &mtx, Size(n, m)); }
-
-template<typename _Tp> inline
-_OutputArray::_OutputArray(_Tp* vec, int n)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_WRITE, vec, Size(n, 1)); }
-
-template<typename _Tp> inline
-_OutputArray::_OutputArray(const std::vector<_Tp>& vec)
-{ init(FIXED_TYPE + FIXED_SIZE + STD_VECTOR + traits::Type<_Tp>::value + ACCESS_WRITE, &vec); }
-
-template<typename _Tp, std::size_t _Nm> inline
-_OutputArray::_OutputArray(const std::array<_Tp, _Nm>& arr)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_WRITE, arr.data(), Size(1, _Nm)); }
-
-template<std::size_t _Nm> inline
-_OutputArray::_OutputArray(const std::array<Mat, _Nm>& arr)
-{ init(FIXED_SIZE + STD_ARRAY_MAT + ACCESS_WRITE, arr.data(), Size(1, _Nm)); }
-
-template<typename _Tp> inline
-_OutputArray::_OutputArray(const std::vector<std::vector<_Tp> >& vec)
-{ init(FIXED_TYPE + FIXED_SIZE + STD_VECTOR_VECTOR + traits::Type<_Tp>::value + ACCESS_WRITE, &vec); }
-
-template<typename _Tp> inline
-_OutputArray::_OutputArray(const std::vector<Mat_<_Tp> >& vec)
-{ init(FIXED_TYPE + FIXED_SIZE + STD_VECTOR_MAT + traits::Type<_Tp>::value + ACCESS_WRITE, &vec); }
-
-template<typename _Tp> inline
-_OutputArray::_OutputArray(const Mat_<_Tp>& m)
-{ init(FIXED_TYPE + FIXED_SIZE + MAT + traits::Type<_Tp>::value + ACCESS_WRITE, &m); }
-
-template<typename _Tp, int m, int n> inline
-_OutputArray::_OutputArray(const Matx<_Tp, m, n>& mtx)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_WRITE, &mtx, Size(n, m)); }
-
-template<typename _Tp> inline
-_OutputArray::_OutputArray(const _Tp* vec, int n)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_WRITE, vec, Size(n, 1)); }
-
-inline _OutputArray::_OutputArray(cuda::GpuMat& d_mat)
-{ init(CUDA_GPU_MAT + ACCESS_WRITE, &d_mat); }
-
-inline _OutputArray::_OutputArray(std::vector<cuda::GpuMat>& d_mat)
-{	init(STD_VECTOR_CUDA_GPU_MAT + ACCESS_WRITE, &d_mat);}
-
-inline _OutputArray::_OutputArray(ogl::Buffer& buf)
-{ init(OPENGL_BUFFER + ACCESS_WRITE, &buf); }
-
-inline _OutputArray::_OutputArray(cuda::HostMem& cuda_mem)
-{ init(CUDA_HOST_MEM + ACCESS_WRITE, &cuda_mem); }
-
-inline _OutputArray::_OutputArray(const Mat& m)
-{ init(FIXED_TYPE + FIXED_SIZE + MAT + ACCESS_WRITE, &m); }
-
-inline _OutputArray::_OutputArray(const std::vector<Mat>& vec)
-{ init(FIXED_SIZE + STD_VECTOR_MAT + ACCESS_WRITE, &vec); }
-
-inline _OutputArray::_OutputArray(const UMat& m)
-{ init(FIXED_TYPE + FIXED_SIZE + UMAT + ACCESS_WRITE, &m); }
-
-inline _OutputArray::_OutputArray(const std::vector<UMat>& vec)
-{ init(FIXED_SIZE + STD_VECTOR_UMAT + ACCESS_WRITE, &vec); }
-
-inline _OutputArray::_OutputArray(const cuda::GpuMat& d_mat)
-{ init(FIXED_TYPE + FIXED_SIZE + CUDA_GPU_MAT + ACCESS_WRITE, &d_mat); }
-
-
-inline _OutputArray::_OutputArray(const ogl::Buffer& buf)
-{ init(FIXED_TYPE + FIXED_SIZE + OPENGL_BUFFER + ACCESS_WRITE, &buf); }
-
-inline _OutputArray::_OutputArray(const cuda::HostMem& cuda_mem)
-{ init(FIXED_TYPE + FIXED_SIZE + CUDA_HOST_MEM + ACCESS_WRITE, &cuda_mem); }
-
-template<typename _Tp> inline
-_OutputArray _OutputArray::rawOut(std::vector<_Tp>& vec)
-{
-    _OutputArray v;
-    v.flags = _InputArray::FIXED_TYPE + _InputArray::STD_VECTOR + rawType<_Tp>() + ACCESS_WRITE;
-    v.obj = (void*)&vec;
-    return v;
-}
-
-template<typename _Tp, std::size_t _Nm> inline
-_OutputArray _OutputArray::rawOut(std::array<_Tp, _Nm>& arr)
-{
-    _OutputArray v;
-    v.flags = FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_WRITE;
-    v.obj = (void*)arr.data();
-    v.sz = Size(1, _Nm);
-    return v;
-}
-
-///////////////////////////////////////////////////////////////////////////////////////////
-
-inline _InputOutputArray::_InputOutputArray() { init(0+ACCESS_RW, 0); }
-inline _InputOutputArray::_InputOutputArray(int _flags, void* _obj) { init(_flags+ACCESS_RW, _obj); }
-inline _InputOutputArray::_InputOutputArray(Mat& m) { init(MAT+ACCESS_RW, &m); }
-inline _InputOutputArray::_InputOutputArray(std::vector<Mat>& vec) { init(STD_VECTOR_MAT+ACCESS_RW, &vec); }
-inline _InputOutputArray::_InputOutputArray(UMat& m) { init(UMAT+ACCESS_RW, &m); }
-inline _InputOutputArray::_InputOutputArray(std::vector<UMat>& vec) { init(STD_VECTOR_UMAT+ACCESS_RW, &vec); }
-
-template<typename _Tp> inline
-_InputOutputArray::_InputOutputArray(std::vector<_Tp>& vec)
-{ init(FIXED_TYPE + STD_VECTOR + traits::Type<_Tp>::value + ACCESS_RW, &vec); }
-
-template<typename _Tp, std::size_t _Nm> inline
-_InputOutputArray::_InputOutputArray(std::array<_Tp, _Nm>& arr)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_RW, arr.data(), Size(1, _Nm)); }
-
-template<std::size_t _Nm> inline
-_InputOutputArray::_InputOutputArray(std::array<Mat, _Nm>& arr)
-{ init(STD_ARRAY_MAT + ACCESS_RW, arr.data(), Size(1, _Nm)); }
-
-template<typename _Tp> inline
-_InputOutputArray::_InputOutputArray(std::vector<std::vector<_Tp> >& vec)
-{ init(FIXED_TYPE + STD_VECTOR_VECTOR + traits::Type<_Tp>::value + ACCESS_RW, &vec); }
-
-template<typename _Tp> inline
-_InputOutputArray::_InputOutputArray(std::vector<Mat_<_Tp> >& vec)
-{ init(FIXED_TYPE + STD_VECTOR_MAT + traits::Type<_Tp>::value + ACCESS_RW, &vec); }
-
-template<typename _Tp> inline
-_InputOutputArray::_InputOutputArray(Mat_<_Tp>& m)
-{ init(FIXED_TYPE + MAT + traits::Type<_Tp>::value + ACCESS_RW, &m); }
-
-template<typename _Tp, int m, int n> inline
-_InputOutputArray::_InputOutputArray(Matx<_Tp, m, n>& mtx)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_RW, &mtx, Size(n, m)); }
-
-template<typename _Tp> inline
-_InputOutputArray::_InputOutputArray(_Tp* vec, int n)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_RW, vec, Size(n, 1)); }
-
-template<typename _Tp> inline
-_InputOutputArray::_InputOutputArray(const std::vector<_Tp>& vec)
-{ init(FIXED_TYPE + FIXED_SIZE + STD_VECTOR + traits::Type<_Tp>::value + ACCESS_RW, &vec); }
-
-template<typename _Tp, std::size_t _Nm> inline
-_InputOutputArray::_InputOutputArray(const std::array<_Tp, _Nm>& arr)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_RW, arr.data(), Size(1, _Nm)); }
-
-template<std::size_t _Nm> inline
-_InputOutputArray::_InputOutputArray(const std::array<Mat, _Nm>& arr)
-{ init(FIXED_SIZE + STD_ARRAY_MAT + ACCESS_RW, arr.data(), Size(1, _Nm)); }
-
-template<typename _Tp> inline
-_InputOutputArray::_InputOutputArray(const std::vector<std::vector<_Tp> >& vec)
-{ init(FIXED_TYPE + FIXED_SIZE + STD_VECTOR_VECTOR + traits::Type<_Tp>::value + ACCESS_RW, &vec); }
-
-template<typename _Tp> inline
-_InputOutputArray::_InputOutputArray(const std::vector<Mat_<_Tp> >& vec)
-{ init(FIXED_TYPE + FIXED_SIZE + STD_VECTOR_MAT + traits::Type<_Tp>::value + ACCESS_RW, &vec); }
-
-template<typename _Tp> inline
-_InputOutputArray::_InputOutputArray(const Mat_<_Tp>& m)
-{ init(FIXED_TYPE + FIXED_SIZE + MAT + traits::Type<_Tp>::value + ACCESS_RW, &m); }
-
-template<typename _Tp, int m, int n> inline
-_InputOutputArray::_InputOutputArray(const Matx<_Tp, m, n>& mtx)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_RW, &mtx, Size(n, m)); }
-
-template<typename _Tp> inline
-_InputOutputArray::_InputOutputArray(const _Tp* vec, int n)
-{ init(FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_RW, vec, Size(n, 1)); }
-
-inline _InputOutputArray::_InputOutputArray(cuda::GpuMat& d_mat)
-{ init(CUDA_GPU_MAT + ACCESS_RW, &d_mat); }
-
-inline _InputOutputArray::_InputOutputArray(ogl::Buffer& buf)
-{ init(OPENGL_BUFFER + ACCESS_RW, &buf); }
-
-inline _InputOutputArray::_InputOutputArray(cuda::HostMem& cuda_mem)
-{ init(CUDA_HOST_MEM + ACCESS_RW, &cuda_mem); }
-
-inline _InputOutputArray::_InputOutputArray(const Mat& m)
-{ init(FIXED_TYPE + FIXED_SIZE + MAT + ACCESS_RW, &m); }
-
-inline _InputOutputArray::_InputOutputArray(const std::vector<Mat>& vec)
-{ init(FIXED_SIZE + STD_VECTOR_MAT + ACCESS_RW, &vec); }
-
-inline _InputOutputArray::_InputOutputArray(const UMat& m)
-{ init(FIXED_TYPE + FIXED_SIZE + UMAT + ACCESS_RW, &m); }
-
-inline _InputOutputArray::_InputOutputArray(const std::vector<UMat>& vec)
-{ init(FIXED_SIZE + STD_VECTOR_UMAT + ACCESS_RW, &vec); }
-
-inline _InputOutputArray::_InputOutputArray(const cuda::GpuMat& d_mat)
-{ init(FIXED_TYPE + FIXED_SIZE + CUDA_GPU_MAT + ACCESS_RW, &d_mat); }
-
-inline _InputOutputArray::_InputOutputArray(const std::vector<cuda::GpuMat>& d_mat)
-{ init(FIXED_TYPE + FIXED_SIZE + STD_VECTOR_CUDA_GPU_MAT + ACCESS_RW, &d_mat);}
-
-template<> inline _InputOutputArray::_InputOutputArray(std::vector<cuda::GpuMat>& d_mat)
-{ init(FIXED_TYPE + FIXED_SIZE + STD_VECTOR_CUDA_GPU_MAT + ACCESS_RW, &d_mat);}
-
-inline _InputOutputArray::_InputOutputArray(const ogl::Buffer& buf)
-{ init(FIXED_TYPE + FIXED_SIZE + OPENGL_BUFFER + ACCESS_RW, &buf); }
-
-inline _InputOutputArray::_InputOutputArray(const cuda::HostMem& cuda_mem)
-{ init(FIXED_TYPE + FIXED_SIZE + CUDA_HOST_MEM + ACCESS_RW, &cuda_mem); }
-
-template<typename _Tp> inline
-_InputOutputArray _InputOutputArray::rawInOut(std::vector<_Tp>& vec)
-{
-    _InputOutputArray v;
-    v.flags = _InputArray::FIXED_TYPE + _InputArray::STD_VECTOR + rawType<_Tp>() + ACCESS_RW;
-    v.obj = (void*)&vec;
-    return v;
-}
-
-template<typename _Tp, std::size_t _Nm> inline
-_InputOutputArray _InputOutputArray::rawInOut(std::array<_Tp, _Nm>& arr)
-{
-    _InputOutputArray v;
-    v.flags = FIXED_TYPE + FIXED_SIZE + MATX + traits::Type<_Tp>::value + ACCESS_RW;
-    v.obj = (void*)arr.data();
-    v.sz = Size(1, _Nm);
-    return v;
-}
-
-
-template<typename _Tp> static inline _InputArray rawIn(_Tp& v) { return _InputArray::rawIn(v); }
-template<typename _Tp> static inline _OutputArray rawOut(_Tp& v) { return _OutputArray::rawOut(v); }
-template<typename _Tp> static inline _InputOutputArray rawInOut(_Tp& v) { return _InputOutputArray::rawInOut(v); }
-
-CV__DEBUG_NS_END
-
-//////////////////////////////////////////// Mat //////////////////////////////////////////
-
-template<typename _Tp> inline
-Mat::Mat(const std::vector<_Tp>& vec, bool copyData)
-    : flags(MAGIC_VAL + traits::Type<_Tp>::value + CV_MAT_CONT_FLAG), dims(2), rows((int)vec.size()),
-      cols(1), data(0), datastart(0), dataend(0), datalimit(0), allocator(0), u(0), size(&rows), step(0)
-{
-    if(vec.empty())
-        return;
-    if( !copyData )
-    {
-        step[0] = step[1] = sizeof(_Tp);
-        datastart = data = (uchar*)&vec[0];
-        datalimit = dataend = datastart + rows * step[0];
-    }
-    else
-        Mat((int)vec.size(), 1, traits::Type<_Tp>::value, (uchar*)&vec[0]).copyTo(*this);
-}
-
-template<typename _Tp, typename> inline
-Mat::Mat(const std::initializer_list<_Tp> list)
-    : Mat()
-{
-    CV_Assert(list.size() != 0);
-    Mat((int)list.size(), 1, traits::Type<_Tp>::value, (uchar*)list.begin()).copyTo(*this);
-}
-
-template<typename _Tp> inline
-Mat::Mat(const std::initializer_list<int> sizes, const std::initializer_list<_Tp> list)
-    : Mat()
-{
-    size_t size_total = 1;
-    for(auto s : sizes)
-        size_total *= s;
-    CV_Assert(list.size() != 0);
-    CV_Assert(size_total == list.size());
-    Mat((int)sizes.size(), (int*)sizes.begin(), traits::Type<_Tp>::value, (uchar*)list.begin()).copyTo(*this);
-}
-
-template<typename _Tp, std::size_t _Nm> inline
-Mat::Mat(const std::array<_Tp, _Nm>& arr, bool copyData)
-    : flags(MAGIC_VAL + traits::Type<_Tp>::value + CV_MAT_CONT_FLAG), dims(2), rows((int)arr.size()),
-      cols(1), data(0), datastart(0), dataend(0), datalimit(0), allocator(0), u(0), size(&rows), step(0)
-{
-    if(arr.empty())
-        return;
-    if( !copyData )
-    {
-        step[0] = step[1] = sizeof(_Tp);
-        datastart = data = (uchar*)arr.data();
-        datalimit = dataend = datastart + rows * step[0];
-    }
-    else
-        Mat((int)arr.size(), 1, traits::Type<_Tp>::value, (uchar*)arr.data()).copyTo(*this);
-}
-
-template<typename _Tp, int n> inline
-Mat::Mat(const Vec<_Tp, n>& vec, bool copyData)
-    : flags(MAGIC_VAL + traits::Type<_Tp>::value + CV_MAT_CONT_FLAG), dims(2), rows(n), cols(1), data(0),
-      datastart(0), dataend(0), datalimit(0), allocator(0), u(0), size(&rows), step(0)
-{
-    if( !copyData )
-    {
-        step[0] = step[1] = sizeof(_Tp);
-        datastart = data = (uchar*)vec.val;
-        datalimit = dataend = datastart + rows * step[0];
-    }
-    else
-        Mat(n, 1, traits::Type<_Tp>::value, (void*)vec.val).copyTo(*this);
-}
-
-
-template<typename _Tp, int m, int n> inline
-Mat::Mat(const Matx<_Tp,m,n>& M, bool copyData)
-    : flags(MAGIC_VAL + traits::Type<_Tp>::value + CV_MAT_CONT_FLAG), dims(2), rows(m), cols(n), data(0),
-      datastart(0), dataend(0), datalimit(0), allocator(0), u(0), size(&rows), step(0)
-{
-    if( !copyData )
-    {
-        step[0] = cols * sizeof(_Tp);
-        step[1] = sizeof(_Tp);
-        datastart = data = (uchar*)M.val;
-        datalimit = dataend = datastart + rows * step[0];
-    }
-    else
-        Mat(m, n, traits::Type<_Tp>::value, (uchar*)M.val).copyTo(*this);
-}
-
-template<typename _Tp> inline
-Mat::Mat(const Point_<_Tp>& pt, bool copyData)
-    : flags(MAGIC_VAL + traits::Type<_Tp>::value + CV_MAT_CONT_FLAG), dims(2), rows(2), cols(1), data(0),
-      datastart(0), dataend(0), datalimit(0), allocator(0), u(0), size(&rows), step(0)
-{
-    if( !copyData )
-    {
-        step[0] = step[1] = sizeof(_Tp);
-        datastart = data = (uchar*)&pt.x;
-        datalimit = dataend = datastart + rows * step[0];
-    }
-    else
-    {
-        create(2, 1, traits::Type<_Tp>::value);
-        ((_Tp*)data)[0] = pt.x;
-        ((_Tp*)data)[1] = pt.y;
-    }
-}
-
-template<typename _Tp> inline
-Mat::Mat(const Point3_<_Tp>& pt, bool copyData)
-    : flags(MAGIC_VAL + traits::Type<_Tp>::value + CV_MAT_CONT_FLAG), dims(2), rows(3), cols(1), data(0),
-      datastart(0), dataend(0), datalimit(0), allocator(0), u(0), size(&rows), step(0)
-{
-    if( !copyData )
-    {
-        step[0] = step[1] = sizeof(_Tp);
-        datastart = data = (uchar*)&pt.x;
-        datalimit = dataend = datastart + rows * step[0];
-    }
-    else
-    {
-        create(3, 1, traits::Type<_Tp>::value);
-        ((_Tp*)data)[0] = pt.x;
-        ((_Tp*)data)[1] = pt.y;
-        ((_Tp*)data)[2] = pt.z;
-    }
-}
-
-template<typename _Tp> inline
-Mat::Mat(const MatCommaInitializer_<_Tp>& commaInitializer)
-    : flags(MAGIC_VAL + traits::Type<_Tp>::value + CV_MAT_CONT_FLAG), dims(0), rows(0), cols(0), data(0),
-      datastart(0), dataend(0), allocator(0), u(0), size(&rows)
-{
-    *this = commaInitializer.operator Mat_<_Tp>();
-}
-
-inline
-Mat Mat::row(int y) const
-{
-    return Mat(*this, Range(y, y + 1), Range::all());
-}
-
-inline
-Mat Mat::col(int x) const
-{
-    return Mat(*this, Range::all(), Range(x, x + 1));
-}
-
-inline
-Mat Mat::rowRange(int startrow, int endrow) const
-{
-    return Mat(*this, Range(startrow, endrow), Range::all());
-}
-
-inline
-Mat Mat::rowRange(const Range& r) const
-{
-    return Mat(*this, r, Range::all());
-}
-
-inline
-Mat Mat::colRange(int startcol, int endcol) const
-{
-    return Mat(*this, Range::all(), Range(startcol, endcol));
-}
-
-inline
-Mat Mat::colRange(const Range& r) const
-{
-    return Mat(*this, Range::all(), r);
-}
-
-inline
-Mat Mat::operator()( Range _rowRange, Range _colRange ) const
-{
-    return Mat(*this, _rowRange, _colRange);
-}
-
-inline
-Mat Mat::operator()( const Rect& roi ) const
-{
-    return Mat(*this, roi);
-}
-
-inline
-Mat Mat::operator()(const Range* ranges) const
-{
-    return Mat(*this, ranges);
-}
-
-inline
-Mat Mat::operator()(const std::vector<Range>& ranges) const
-{
-    return Mat(*this, ranges);
-}
-
-inline
-bool Mat::isContinuous() const
-{
-    return (flags & CONTINUOUS_FLAG) != 0;
-}
-
-inline
-bool Mat::isSubmatrix() const
-{
-    return (flags & SUBMATRIX_FLAG) != 0;
-}
-
-inline
-size_t Mat::elemSize() const
-{
-    size_t res = dims > 0 ? step.p[dims - 1] : 0;
-    CV_DbgAssert(res != 0);
-    return res;
-}
-
-inline
-size_t Mat::elemSize1() const
-{
-    return CV_ELEM_SIZE1(flags);
-}
-
-inline
-int Mat::type() const
-{
-    return CV_MAT_TYPE(flags);
-}
-
-inline
-int Mat::depth() const
-{
-    return CV_MAT_DEPTH(flags);
-}
-
-inline
-int Mat::channels() const
-{
-    return CV_MAT_CN(flags);
-}
-
-inline
-uchar* Mat::ptr(int y)
-{
-    CV_DbgAssert( y == 0 || (data && dims >= 1 && (unsigned)y < (unsigned)size.p[0]) );
-    return data + step.p[0] * y;
-}
-
-inline
-const uchar* Mat::ptr(int y) const
-{
-    CV_DbgAssert( y == 0 || (data && dims >= 1 && (unsigned)y < (unsigned)size.p[0]) );
-    return data + step.p[0] * y;
-}
-
-template<typename _Tp> inline
-_Tp* Mat::ptr(int y)
-{
-    CV_DbgAssert( y == 0 || (data && dims >= 1 && (unsigned)y < (unsigned)size.p[0]) );
-    return (_Tp*)(data + step.p[0] * y);
-}
-
-template<typename _Tp> inline
-const _Tp* Mat::ptr(int y) const
-{
-    CV_DbgAssert( y == 0 || (data && dims >= 1 && (unsigned)y < (unsigned)size.p[0]) );
-    return (const _Tp*)(data + step.p[0] * y);
-}
-
-inline
-uchar* Mat::ptr(int i0, int i1)
-{
-    CV_DbgAssert(dims >= 2);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)i0 < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)i1 < (unsigned)size.p[1]);
-    return data + i0 * step.p[0] + i1 * step.p[1];
-}
-
-inline
-const uchar* Mat::ptr(int i0, int i1) const
-{
-    CV_DbgAssert(dims >= 2);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)i0 < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)i1 < (unsigned)size.p[1]);
-    return data + i0 * step.p[0] + i1 * step.p[1];
-}
-
-template<typename _Tp> inline
-_Tp* Mat::ptr(int i0, int i1)
-{
-    CV_DbgAssert(dims >= 2);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)i0 < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)i1 < (unsigned)size.p[1]);
-    return (_Tp*)(data + i0 * step.p[0] + i1 * step.p[1]);
-}
-
-template<typename _Tp> inline
-const _Tp* Mat::ptr(int i0, int i1) const
-{
-    CV_DbgAssert(dims >= 2);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)i0 < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)i1 < (unsigned)size.p[1]);
-    return (const _Tp*)(data + i0 * step.p[0] + i1 * step.p[1]);
-}
-
-inline
-uchar* Mat::ptr(int i0, int i1, int i2)
-{
-    CV_DbgAssert(dims >= 3);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)i0 < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)i1 < (unsigned)size.p[1]);
-    CV_DbgAssert((unsigned)i2 < (unsigned)size.p[2]);
-    return data + i0 * step.p[0] + i1 * step.p[1] + i2 * step.p[2];
-}
-
-inline
-const uchar* Mat::ptr(int i0, int i1, int i2) const
-{
-    CV_DbgAssert(dims >= 3);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)i0 < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)i1 < (unsigned)size.p[1]);
-    CV_DbgAssert((unsigned)i2 < (unsigned)size.p[2]);
-    return data + i0 * step.p[0] + i1 * step.p[1] + i2 * step.p[2];
-}
-
-template<typename _Tp> inline
-_Tp* Mat::ptr(int i0, int i1, int i2)
-{
-    CV_DbgAssert(dims >= 3);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)i0 < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)i1 < (unsigned)size.p[1]);
-    CV_DbgAssert((unsigned)i2 < (unsigned)size.p[2]);
-    return (_Tp*)(data + i0 * step.p[0] + i1 * step.p[1] + i2 * step.p[2]);
-}
-
-template<typename _Tp> inline
-const _Tp* Mat::ptr(int i0, int i1, int i2) const
-{
-    CV_DbgAssert(dims >= 3);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)i0 < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)i1 < (unsigned)size.p[1]);
-    CV_DbgAssert((unsigned)i2 < (unsigned)size.p[2]);
-    return (const _Tp*)(data + i0 * step.p[0] + i1 * step.p[1] + i2 * step.p[2]);
-}
-
-inline
-uchar* Mat::ptr(const int* idx)
-{
-    int i, d = dims;
-    uchar* p = data;
-    CV_DbgAssert( d >= 1 && p );
-    for( i = 0; i < d; i++ )
-    {
-        CV_DbgAssert( (unsigned)idx[i] < (unsigned)size.p[i] );
-        p += idx[i] * step.p[i];
-    }
-    return p;
-}
-
-inline
-const uchar* Mat::ptr(const int* idx) const
-{
-    int i, d = dims;
-    uchar* p = data;
-    CV_DbgAssert( d >= 1 && p );
-    for( i = 0; i < d; i++ )
-    {
-        CV_DbgAssert( (unsigned)idx[i] < (unsigned)size.p[i] );
-        p += idx[i] * step.p[i];
-    }
-    return p;
-}
-
-template<typename _Tp> inline
-_Tp* Mat::ptr(const int* idx)
-{
-    int i, d = dims;
-    uchar* p = data;
-    CV_DbgAssert( d >= 1 && p );
-    for( i = 0; i < d; i++ )
-    {
-        CV_DbgAssert( (unsigned)idx[i] < (unsigned)size.p[i] );
-        p += idx[i] * step.p[i];
-    }
-    return (_Tp*)p;
-}
-
-template<typename _Tp> inline
-const _Tp* Mat::ptr(const int* idx) const
-{
-    int i, d = dims;
-    uchar* p = data;
-    CV_DbgAssert( d >= 1 && p );
-    for( i = 0; i < d; i++ )
-    {
-        CV_DbgAssert( (unsigned)idx[i] < (unsigned)size.p[i] );
-        p += idx[i] * step.p[i];
-    }
-    return (const _Tp*)p;
-}
-
-template<int n> inline
-uchar* Mat::ptr(const Vec<int, n>& idx)
-{
-    return Mat::ptr(idx.val);
-}
-
-template<int n> inline
-const uchar* Mat::ptr(const Vec<int, n>& idx) const
-{
-    return Mat::ptr(idx.val);
-}
-
-template<typename _Tp, int n> inline
-_Tp* Mat::ptr(const Vec<int, n>& idx)
-{
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    return Mat::ptr<_Tp>(idx.val);
-}
-
-template<typename _Tp, int n> inline
-const _Tp* Mat::ptr(const Vec<int, n>& idx) const
-{
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    return Mat::ptr<_Tp>(idx.val);
-}
-
-
-template<typename _Tp> inline
-_Tp& Mat::at(int i0, int i1)
-{
-    CV_DbgAssert(dims <= 2);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)i0 < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)(i1 * DataType<_Tp>::channels) < (unsigned)(size.p[1] * channels()));
-    CV_DbgAssert(CV_ELEM_SIZE1(traits::Depth<_Tp>::value) == elemSize1());
-    return ((_Tp*)(data + step.p[0] * i0))[i1];
-}
-
-template<typename _Tp> inline
-const _Tp& Mat::at(int i0, int i1) const
-{
-    CV_DbgAssert(dims <= 2);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)i0 < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)(i1 * DataType<_Tp>::channels) < (unsigned)(size.p[1] * channels()));
-    CV_DbgAssert(CV_ELEM_SIZE1(traits::Depth<_Tp>::value) == elemSize1());
-    return ((const _Tp*)(data + step.p[0] * i0))[i1];
-}
-
-template<typename _Tp> inline
-_Tp& Mat::at(Point pt)
-{
-    CV_DbgAssert(dims <= 2);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)pt.y < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)(pt.x * DataType<_Tp>::channels) < (unsigned)(size.p[1] * channels()));
-    CV_DbgAssert(CV_ELEM_SIZE1(traits::Depth<_Tp>::value) == elemSize1());
-    return ((_Tp*)(data + step.p[0] * pt.y))[pt.x];
-}
-
-template<typename _Tp> inline
-const _Tp& Mat::at(Point pt) const
-{
-    CV_DbgAssert(dims <= 2);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)pt.y < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)(pt.x * DataType<_Tp>::channels) < (unsigned)(size.p[1] * channels()));
-    CV_DbgAssert(CV_ELEM_SIZE1(traits::Depth<_Tp>::value) == elemSize1());
-    return ((const _Tp*)(data + step.p[0] * pt.y))[pt.x];
-}
-
-template<typename _Tp> inline
-_Tp& Mat::at(int i0)
-{
-    CV_DbgAssert(dims <= 2);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)i0 < (unsigned)(size.p[0] * size.p[1]));
-    CV_DbgAssert(elemSize() == sizeof(_Tp));
-    if( isContinuous() || size.p[0] == 1 )
-        return ((_Tp*)data)[i0];
-    if( size.p[1] == 1 )
-        return *(_Tp*)(data + step.p[0] * i0);
-    int i = i0 / cols, j = i0 - i * cols;
-    return ((_Tp*)(data + step.p[0] * i))[j];
-}
-
-template<typename _Tp> inline
-const _Tp& Mat::at(int i0) const
-{
-    CV_DbgAssert(dims <= 2);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)i0 < (unsigned)(size.p[0] * size.p[1]));
-    CV_DbgAssert(elemSize() == sizeof(_Tp));
-    if( isContinuous() || size.p[0] == 1 )
-        return ((const _Tp*)data)[i0];
-    if( size.p[1] == 1 )
-        return *(const _Tp*)(data + step.p[0] * i0);
-    int i = i0 / cols, j = i0 - i * cols;
-    return ((const _Tp*)(data + step.p[0] * i))[j];
-}
-
-template<typename _Tp> inline
-_Tp& Mat::at(int i0, int i1, int i2)
-{
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    return *(_Tp*)ptr(i0, i1, i2);
-}
-
-template<typename _Tp> inline
-const _Tp& Mat::at(int i0, int i1, int i2) const
-{
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    return *(const _Tp*)ptr(i0, i1, i2);
-}
-
-template<typename _Tp> inline
-_Tp& Mat::at(const int* idx)
-{
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    return *(_Tp*)ptr(idx);
-}
-
-template<typename _Tp> inline
-const _Tp& Mat::at(const int* idx) const
-{
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    return *(const _Tp*)ptr(idx);
-}
-
-template<typename _Tp, int n> inline
-_Tp& Mat::at(const Vec<int, n>& idx)
-{
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    return *(_Tp*)ptr(idx.val);
-}
-
-template<typename _Tp, int n> inline
-const _Tp& Mat::at(const Vec<int, n>& idx) const
-{
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    return *(const _Tp*)ptr(idx.val);
-}
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp> Mat::begin() const
-{
-    if (empty())
-        return MatConstIterator_<_Tp>();
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    return MatConstIterator_<_Tp>((const Mat_<_Tp>*)this);
-}
-
-template<typename _Tp> inline
-std::reverse_iterator<MatConstIterator_<_Tp>> Mat::rbegin() const
-{
-    if (empty())
-        return std::reverse_iterator<MatConstIterator_<_Tp>>();
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    MatConstIterator_<_Tp> it((const Mat_<_Tp>*)this);
-    it += total();
-    return std::reverse_iterator<MatConstIterator_<_Tp>> (it);
-}
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp> Mat::end() const
-{
-    if (empty())
-        return MatConstIterator_<_Tp>();
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    MatConstIterator_<_Tp> it((const Mat_<_Tp>*)this);
-    it += total();
-    return it;
-}
-
-template<typename _Tp> inline
-std::reverse_iterator<MatConstIterator_<_Tp>> Mat::rend() const
-{
-    if (empty())
-        return std::reverse_iterator<MatConstIterator_<_Tp>>();
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    return std::reverse_iterator<MatConstIterator_<_Tp>>((const Mat_<_Tp>*)this);
-}
-
-template<typename _Tp> inline
-MatIterator_<_Tp> Mat::begin()
-{
-    if (empty())
-        return MatIterator_<_Tp>();
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    return MatIterator_<_Tp>((Mat_<_Tp>*)this);
-}
-
-template<typename _Tp> inline
-std::reverse_iterator<MatIterator_<_Tp>> Mat::rbegin()
-{
-    if (empty())
-        return std::reverse_iterator<MatIterator_<_Tp>>();
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    MatIterator_<_Tp> it((Mat_<_Tp>*)this);
-    it += total();
-    return std::reverse_iterator<MatIterator_<_Tp>>(it);
-}
-
-template<typename _Tp> inline
-MatIterator_<_Tp> Mat::end()
-{
-    if (empty())
-        return MatIterator_<_Tp>();
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    MatIterator_<_Tp> it((Mat_<_Tp>*)this);
-    it += total();
-    return it;
-}
-
-template<typename _Tp> inline
-std::reverse_iterator<MatIterator_<_Tp>> Mat::rend()
-{
-    if (empty())
-        return std::reverse_iterator<MatIterator_<_Tp>>();
-    CV_DbgAssert( elemSize() == sizeof(_Tp) );
-    return std::reverse_iterator<MatIterator_<_Tp>>(MatIterator_<_Tp>((Mat_<_Tp>*)this));
-}
-
-template<typename _Tp, typename Functor> inline
-void Mat::forEach(const Functor& operation) {
-    this->forEach_impl<_Tp>(operation);
-}
-
-template<typename _Tp, typename Functor> inline
-void Mat::forEach(const Functor& operation) const {
-    // call as not const
-    (const_cast<Mat*>(this))->forEach<_Tp>(operation);
-}
-
-template<typename _Tp> inline
-Mat::operator std::vector<_Tp>() const
-{
-    std::vector<_Tp> v;
-    copyTo(v);
-    return v;
-}
-
-template<typename _Tp, std::size_t _Nm> inline
-Mat::operator std::array<_Tp, _Nm>() const
-{
-    std::array<_Tp, _Nm> v;
-    copyTo(v);
-    return v;
-}
-
-template<typename _Tp, int n> inline
-Mat::operator Vec<_Tp, n>() const
-{
-    CV_Assert( data && dims <= 2 && (rows == 1 || cols == 1) &&
-               rows + cols - 1 == n && channels() == 1 );
-
-    if( isContinuous() && type() == traits::Type<_Tp>::value )
-        return Vec<_Tp, n>((_Tp*)data);
-    Vec<_Tp, n> v;
-    Mat tmp(rows, cols, traits::Type<_Tp>::value, v.val);
-    convertTo(tmp, tmp.type());
-    return v;
-}
-
-template<typename _Tp, int m, int n> inline
-Mat::operator Matx<_Tp, m, n>() const
-{
-    CV_Assert( data && dims <= 2 && rows == m && cols == n && channels() == 1 );
-
-    if( isContinuous() && type() == traits::Type<_Tp>::value )
-        return Matx<_Tp, m, n>((_Tp*)data);
-    Matx<_Tp, m, n> mtx;
-    Mat tmp(rows, cols, traits::Type<_Tp>::value, mtx.val);
-    convertTo(tmp, tmp.type());
-    return mtx;
-}
-
-template<typename _Tp> inline
-void Mat::push_back(const _Tp& elem)
-{
-    if( !data )
-    {
-        *this = Mat(1, 1, traits::Type<_Tp>::value, (void*)&elem).clone();
-        return;
-    }
-    CV_Assert(traits::Type<_Tp>::value == type() && cols == 1
-              /* && dims == 2 (cols == 1 implies dims == 2) */);
-    const uchar* tmp = dataend + step[0];
-    if( !isSubmatrix() && isContinuous() && tmp <= datalimit )
-    {
-        *(_Tp*)(data + (size.p[0]++) * step.p[0]) = elem;
-        dataend = tmp;
-    }
-    else
-        push_back_(&elem);
-}
-
-template<typename _Tp> inline
-void Mat::push_back(const Mat_<_Tp>& m)
-{
-    push_back((const Mat&)m);
-}
-
-template<> inline
-void Mat::push_back(const MatExpr& expr)
-{
-    push_back(static_cast<Mat>(expr));
-}
-
-
-template<typename _Tp> inline
-void Mat::push_back(const std::vector<_Tp>& v)
-{
-    push_back(Mat(v));
-}
-
-
-///////////////////////////// MatSize ////////////////////////////
-
-inline
-MatSize::MatSize(int* _p) CV_NOEXCEPT
-    : p(_p) {}
-
-inline
-int MatSize::dims() const CV_NOEXCEPT
-{
-    return (p - 1)[0];
-}
-
-inline
-Size MatSize::operator()() const
-{
-    CV_DbgAssert(dims() <= 2);
-    return Size(p[1], p[0]);
-}
-
-inline
-const int& MatSize::operator[](int i) const
-{
-    CV_DbgAssert(i < dims());
-#ifdef __OPENCV_BUILD
-    CV_DbgAssert(i >= 0);
-#endif
-    return p[i];
-}
-
-inline
-int& MatSize::operator[](int i)
-{
-    CV_DbgAssert(i < dims());
-#ifdef __OPENCV_BUILD
-    CV_DbgAssert(i >= 0);
-#endif
-    return p[i];
-}
-
-inline
-MatSize::operator const int*() const CV_NOEXCEPT
-{
-    return p;
-}
-
-inline
-bool MatSize::operator != (const MatSize& sz) const CV_NOEXCEPT
-{
-    return !(*this == sz);
-}
-
-
-
-///////////////////////////// MatStep ////////////////////////////
-
-inline
-MatStep::MatStep() CV_NOEXCEPT
-{
-    p = buf; p[0] = p[1] = 0;
-}
-
-inline
-MatStep::MatStep(size_t s) CV_NOEXCEPT
-{
-    p = buf; p[0] = s; p[1] = 0;
-}
-
-inline
-const size_t& MatStep::operator[](int i) const CV_NOEXCEPT
-{
-    return p[i];
-}
-
-inline
-size_t& MatStep::operator[](int i) CV_NOEXCEPT
-{
-    return p[i];
-}
-
-inline MatStep::operator size_t() const
-{
-    CV_DbgAssert( p == buf );
-    return buf[0];
-}
-
-inline MatStep& MatStep::operator = (size_t s)
-{
-    CV_DbgAssert( p == buf );
-    buf[0] = s;
-    return *this;
-}
-
-
-
-////////////////////////////// Mat_<_Tp> ////////////////////////////
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_() CV_NOEXCEPT
-    : Mat()
-{
-    flags = (flags & ~CV_MAT_TYPE_MASK) + traits::Type<_Tp>::value;
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(int _rows, int _cols)
-    : Mat(_rows, _cols, traits::Type<_Tp>::value)
-{
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(int _rows, int _cols, const _Tp& value)
-    : Mat(_rows, _cols, traits::Type<_Tp>::value)
-{
-    *this = value;
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(Size _sz)
-    : Mat(_sz.height, _sz.width, traits::Type<_Tp>::value)
-{}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(Size _sz, const _Tp& value)
-    : Mat(_sz.height, _sz.width, traits::Type<_Tp>::value)
-{
-    *this = value;
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(int _dims, const int* _sz)
-    : Mat(_dims, _sz, traits::Type<_Tp>::value)
-{}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(int _dims, const int* _sz, const _Tp& _s)
-    : Mat(_dims, _sz, traits::Type<_Tp>::value, Scalar(_s))
-{}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(int _dims, const int* _sz, _Tp* _data, const size_t* _steps)
-    : Mat(_dims, _sz, traits::Type<_Tp>::value, _data, _steps)
-{}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(const Mat_<_Tp>& m, const Range* ranges)
-    : Mat(m, ranges)
-{}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(const Mat_<_Tp>& m, const std::vector<Range>& ranges)
-    : Mat(m, ranges)
-{}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(const Mat& m)
-    : Mat()
-{
-    flags = (flags & ~CV_MAT_TYPE_MASK) + traits::Type<_Tp>::value;
-    *this = m;
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(const Mat_& m)
-    : Mat(m)
-{}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(int _rows, int _cols, _Tp* _data, size_t steps)
-    : Mat(_rows, _cols, traits::Type<_Tp>::value, _data, steps)
-{}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(const Mat_& m, const Range& _rowRange, const Range& _colRange)
-    : Mat(m, _rowRange, _colRange)
-{}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(const Mat_& m, const Rect& roi)
-    : Mat(m, roi)
-{}
-
-template<typename _Tp> template<int n> inline
-Mat_<_Tp>::Mat_(const Vec<typename DataType<_Tp>::channel_type, n>& vec, bool copyData)
-    : Mat(n / DataType<_Tp>::channels, 1, traits::Type<_Tp>::value, (void*)&vec)
-{
-    CV_Assert(n%DataType<_Tp>::channels == 0);
-    if( copyData )
-        *this = clone();
-}
-
-template<typename _Tp> template<int m, int n> inline
-Mat_<_Tp>::Mat_(const Matx<typename DataType<_Tp>::channel_type, m, n>& M, bool copyData)
-    : Mat(m, n / DataType<_Tp>::channels, traits::Type<_Tp>::value, (void*)&M)
-{
-    CV_Assert(n % DataType<_Tp>::channels == 0);
-    if( copyData )
-        *this = clone();
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(const Point_<typename DataType<_Tp>::channel_type>& pt, bool copyData)
-    : Mat(2 / DataType<_Tp>::channels, 1, traits::Type<_Tp>::value, (void*)&pt)
-{
-    CV_Assert(2 % DataType<_Tp>::channels == 0);
-    if( copyData )
-        *this = clone();
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(const Point3_<typename DataType<_Tp>::channel_type>& pt, bool copyData)
-    : Mat(3 / DataType<_Tp>::channels, 1, traits::Type<_Tp>::value, (void*)&pt)
-{
-    CV_Assert(3 % DataType<_Tp>::channels == 0);
-    if( copyData )
-        *this = clone();
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(const MatCommaInitializer_<_Tp>& commaInitializer)
-    : Mat(commaInitializer)
-{}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(const std::vector<_Tp>& vec, bool copyData)
-    : Mat(vec, copyData)
-{}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(std::initializer_list<_Tp> list)
-    : Mat(list)
-{}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(const std::initializer_list<int> sizes, std::initializer_list<_Tp> list)
-    : Mat(sizes, list)
-{}
-
-template<typename _Tp> template<std::size_t _Nm> inline
-Mat_<_Tp>::Mat_(const std::array<_Tp, _Nm>& arr, bool copyData)
-    : Mat(arr, copyData)
-{}
-
-template<typename _Tp> inline
-Mat_<_Tp>& Mat_<_Tp>::operator = (const Mat& m)
-{
-    if (m.empty())
-    {
-        release();
-        return *this;
-    }
-    if( traits::Type<_Tp>::value == m.type() )
-    {
-        Mat::operator = (m);
-        return *this;
-    }
-    if( traits::Depth<_Tp>::value == m.depth() )
-    {
-        return (*this = m.reshape(DataType<_Tp>::channels, m.dims, 0));
-    }
-    CV_Assert(DataType<_Tp>::channels == m.channels() || m.empty());
-    m.convertTo(*this, type());
-    return *this;
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>& Mat_<_Tp>::operator = (const Mat_& m)
-{
-    Mat::operator=(m);
-    return *this;
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>& Mat_<_Tp>::operator = (const _Tp& s)
-{
-    typedef typename DataType<_Tp>::vec_type VT;
-    Mat::operator=(Scalar((const VT&)s));
-    return *this;
-}
-
-template<typename _Tp> inline
-void Mat_<_Tp>::create(int _rows, int _cols)
-{
-    Mat::create(_rows, _cols, traits::Type<_Tp>::value);
-}
-
-template<typename _Tp> inline
-void Mat_<_Tp>::create(Size _sz)
-{
-    Mat::create(_sz, traits::Type<_Tp>::value);
-}
-
-template<typename _Tp> inline
-void Mat_<_Tp>::create(int _dims, const int* _sz)
-{
-    Mat::create(_dims, _sz, traits::Type<_Tp>::value);
-}
-
-template<typename _Tp> inline
-void Mat_<_Tp>::release()
-{
-    Mat::release();
-    flags = (flags & ~CV_MAT_TYPE_MASK) + traits::Type<_Tp>::value;
-}
-
-template<typename _Tp> inline
-Mat_<_Tp> Mat_<_Tp>::cross(const Mat_& m) const
-{
-    return Mat_<_Tp>(Mat::cross(m));
-}
-
-template<typename _Tp> template<typename T2> inline
-Mat_<_Tp>::operator Mat_<T2>() const
-{
-    return Mat_<T2>(static_cast<const Mat&>(*this));
-}
-
-template<typename _Tp> inline
-Mat_<_Tp> Mat_<_Tp>::row(int y) const
-{
-    return Mat_(*this, Range(y, y+1), Range::all());
-}
-
-template<typename _Tp> inline
-Mat_<_Tp> Mat_<_Tp>::col(int x) const
-{
-    return Mat_(*this, Range::all(), Range(x, x+1));
-}
-
-template<typename _Tp> inline
-Mat_<_Tp> Mat_<_Tp>::diag(int d) const
-{
-    return Mat_(Mat::diag(d));
-}
-
-template<typename _Tp> inline
-Mat_<_Tp> Mat_<_Tp>::clone() const
-{
-    return Mat_(Mat::clone());
-}
-
-template<typename _Tp> inline
-size_t Mat_<_Tp>::elemSize() const
-{
-    CV_DbgAssert( Mat::elemSize() == sizeof(_Tp) );
-    return sizeof(_Tp);
-}
-
-template<typename _Tp> inline
-size_t Mat_<_Tp>::elemSize1() const
-{
-    CV_DbgAssert( Mat::elemSize1() == sizeof(_Tp) / DataType<_Tp>::channels );
-    return sizeof(_Tp) / DataType<_Tp>::channels;
-}
-
-template<typename _Tp> inline
-int Mat_<_Tp>::type() const
-{
-    CV_DbgAssert( Mat::type() == traits::Type<_Tp>::value );
-    return traits::Type<_Tp>::value;
-}
-
-template<typename _Tp> inline
-int Mat_<_Tp>::depth() const
-{
-    CV_DbgAssert( Mat::depth() == traits::Depth<_Tp>::value );
-    return traits::Depth<_Tp>::value;
-}
-
-template<typename _Tp> inline
-int Mat_<_Tp>::channels() const
-{
-    CV_DbgAssert( Mat::channels() == DataType<_Tp>::channels );
-    return DataType<_Tp>::channels;
-}
-
-template<typename _Tp> inline
-size_t Mat_<_Tp>::stepT(int i) const
-{
-    return step.p[i] / elemSize();
-}
-
-template<typename _Tp> inline
-size_t Mat_<_Tp>::step1(int i) const
-{
-    return step.p[i] / elemSize1();
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>& Mat_<_Tp>::adjustROI( int dtop, int dbottom, int dleft, int dright )
-{
-    return (Mat_<_Tp>&)(Mat::adjustROI(dtop, dbottom, dleft, dright));
-}
-
-template<typename _Tp> inline
-Mat_<_Tp> Mat_<_Tp>::operator()( const Range& _rowRange, const Range& _colRange ) const
-{
-    return Mat_<_Tp>(*this, _rowRange, _colRange);
-}
-
-template<typename _Tp> inline
-Mat_<_Tp> Mat_<_Tp>::operator()( const Rect& roi ) const
-{
-    return Mat_<_Tp>(*this, roi);
-}
-
-template<typename _Tp> inline
-Mat_<_Tp> Mat_<_Tp>::operator()( const Range* ranges ) const
-{
-    return Mat_<_Tp>(*this, ranges);
-}
-
-template<typename _Tp> inline
-Mat_<_Tp> Mat_<_Tp>::operator()(const std::vector<Range>& ranges) const
-{
-    return Mat_<_Tp>(*this, ranges);
-}
-
-template<typename _Tp> inline
-_Tp* Mat_<_Tp>::operator [](int y)
-{
-    CV_DbgAssert( 0 <= y && y < size.p[0] );
-    return (_Tp*)(data + y*step.p[0]);
-}
-
-template<typename _Tp> inline
-const _Tp* Mat_<_Tp>::operator [](int y) const
-{
-    CV_DbgAssert( 0 <= y && y < size.p[0] );
-    return (const _Tp*)(data + y*step.p[0]);
-}
-
-template<typename _Tp> inline
-_Tp& Mat_<_Tp>::operator ()(int i0, int i1)
-{
-    CV_DbgAssert(dims <= 2);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)i0 < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)i1 < (unsigned)size.p[1]);
-    CV_DbgAssert(type() == traits::Type<_Tp>::value);
-    return ((_Tp*)(data + step.p[0] * i0))[i1];
-}
-
-template<typename _Tp> inline
-const _Tp& Mat_<_Tp>::operator ()(int i0, int i1) const
-{
-    CV_DbgAssert(dims <= 2);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)i0 < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)i1 < (unsigned)size.p[1]);
-    CV_DbgAssert(type() == traits::Type<_Tp>::value);
-    return ((const _Tp*)(data + step.p[0] * i0))[i1];
-}
-
-template<typename _Tp> inline
-_Tp& Mat_<_Tp>::operator ()(Point pt)
-{
-    CV_DbgAssert(dims <= 2);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)pt.y < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)pt.x < (unsigned)size.p[1]);
-    CV_DbgAssert(type() == traits::Type<_Tp>::value);
-    return ((_Tp*)(data + step.p[0] * pt.y))[pt.x];
-}
-
-template<typename _Tp> inline
-const _Tp& Mat_<_Tp>::operator ()(Point pt) const
-{
-    CV_DbgAssert(dims <= 2);
-    CV_DbgAssert(data);
-    CV_DbgAssert((unsigned)pt.y < (unsigned)size.p[0]);
-    CV_DbgAssert((unsigned)pt.x < (unsigned)size.p[1]);
-    CV_DbgAssert(type() == traits::Type<_Tp>::value);
-    return ((const _Tp*)(data + step.p[0] * pt.y))[pt.x];
-}
-
-template<typename _Tp> inline
-_Tp& Mat_<_Tp>::operator ()(const int* idx)
-{
-    return Mat::at<_Tp>(idx);
-}
-
-template<typename _Tp> inline
-const _Tp& Mat_<_Tp>::operator ()(const int* idx) const
-{
-    return Mat::at<_Tp>(idx);
-}
-
-template<typename _Tp> template<int n> inline
-_Tp& Mat_<_Tp>::operator ()(const Vec<int, n>& idx)
-{
-    return Mat::at<_Tp>(idx);
-}
-
-template<typename _Tp> template<int n> inline
-const _Tp& Mat_<_Tp>::operator ()(const Vec<int, n>& idx) const
-{
-    return Mat::at<_Tp>(idx);
-}
-
-template<typename _Tp> inline
-_Tp& Mat_<_Tp>::operator ()(int i0)
-{
-    return this->at<_Tp>(i0);
-}
-
-template<typename _Tp> inline
-const _Tp& Mat_<_Tp>::operator ()(int i0) const
-{
-    return this->at<_Tp>(i0);
-}
-
-template<typename _Tp> inline
-_Tp& Mat_<_Tp>::operator ()(int i0, int i1, int i2)
-{
-    return this->at<_Tp>(i0, i1, i2);
-}
-
-template<typename _Tp> inline
-const _Tp& Mat_<_Tp>::operator ()(int i0, int i1, int i2) const
-{
-    return this->at<_Tp>(i0, i1, i2);
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>::operator std::vector<_Tp>() const
-{
-    std::vector<_Tp> v;
-    copyTo(v);
-    return v;
-}
-
-template<typename _Tp> template<std::size_t _Nm> inline
-Mat_<_Tp>::operator std::array<_Tp, _Nm>() const
-{
-    std::array<_Tp, _Nm> a;
-    copyTo(a);
-    return a;
-}
-
-template<typename _Tp> template<int n> inline
-Mat_<_Tp>::operator Vec<typename DataType<_Tp>::channel_type, n>() const
-{
-    CV_Assert(n % DataType<_Tp>::channels == 0);
-
-#if defined _MSC_VER
-    const Mat* pMat = (const Mat*)this; // workaround for MSVS <= 2012 compiler bugs (but GCC 4.6 dislikes this workaround)
-    return pMat->operator Vec<typename DataType<_Tp>::channel_type, n>();
-#else
-    return this->Mat::operator Vec<typename DataType<_Tp>::channel_type, n>();
-#endif
-}
-
-template<typename _Tp> template<int m, int n> inline
-Mat_<_Tp>::operator Matx<typename DataType<_Tp>::channel_type, m, n>() const
-{
-    CV_Assert(n % DataType<_Tp>::channels == 0);
-
-#if defined _MSC_VER
-    const Mat* pMat = (const Mat*)this; // workaround for MSVS <= 2012 compiler bugs (but GCC 4.6 dislikes this workaround)
-    Matx<typename DataType<_Tp>::channel_type, m, n> res = pMat->operator Matx<typename DataType<_Tp>::channel_type, m, n>();
-    return res;
-#else
-    Matx<typename DataType<_Tp>::channel_type, m, n> res = this->Mat::operator Matx<typename DataType<_Tp>::channel_type, m, n>();
-    return res;
-#endif
-}
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp> Mat_<_Tp>::begin() const
-{
-    return Mat::begin<_Tp>();
-}
-
-template<typename _Tp> inline
-std::reverse_iterator<MatConstIterator_<_Tp>> Mat_<_Tp>::rbegin() const
-{
-    return Mat::rbegin<_Tp>();
-}
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp> Mat_<_Tp>::end() const
-{
-    return Mat::end<_Tp>();
-}
-
-template<typename _Tp> inline
-std::reverse_iterator<MatConstIterator_<_Tp>> Mat_<_Tp>::rend() const
-{
-    return Mat::rend<_Tp>();
-}
-
-template<typename _Tp> inline
-MatIterator_<_Tp> Mat_<_Tp>::begin()
-{
-    return Mat::begin<_Tp>();
-}
-
-template<typename _Tp> inline
-std::reverse_iterator<MatIterator_<_Tp>> Mat_<_Tp>::rbegin()
-{
-    return Mat::rbegin<_Tp>();
-}
-
-template<typename _Tp> inline
-MatIterator_<_Tp> Mat_<_Tp>::end()
-{
-    return Mat::end<_Tp>();
-}
-
-template<typename _Tp> inline
-std::reverse_iterator<MatIterator_<_Tp>> Mat_<_Tp>::rend()
-{
-    return Mat::rend<_Tp>();
-}
-
-template<typename _Tp> template<typename Functor> inline
-void Mat_<_Tp>::forEach(const Functor& operation) {
-    Mat::forEach<_Tp, Functor>(operation);
-}
-
-template<typename _Tp> template<typename Functor> inline
-void Mat_<_Tp>::forEach(const Functor& operation) const {
-    Mat::forEach<_Tp, Functor>(operation);
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(Mat_&& m)
-    : Mat(std::move(m))
-{
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>& Mat_<_Tp>::operator = (Mat_&& m)
-{
-    Mat::operator = (std::move(m));
-    return *this;
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(Mat&& m)
-    : Mat()
-{
-    flags = (flags & ~CV_MAT_TYPE_MASK) + traits::Type<_Tp>::value;
-    *this = std::move(m);
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>& Mat_<_Tp>::operator = (Mat&& m)
-{
-    if (m.empty())
-    {
-        release();
-        return *this;
-    }
-    if( traits::Type<_Tp>::value == m.type() )
-    {
-        Mat::operator = ((Mat&&)m);
-        return *this;
-    }
-    if( traits::Depth<_Tp>::value == m.depth() )
-    {
-        Mat::operator = ((Mat&&)m.reshape(DataType<_Tp>::channels, m.dims, 0));
-        return *this;
-    }
-    CV_DbgAssert(DataType<_Tp>::channels == m.channels());
-    m.convertTo(*this, type());
-    return *this;
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(MatExpr&& e)
-    : Mat()
-{
-    flags = (flags & ~CV_MAT_TYPE_MASK) + traits::Type<_Tp>::value;
-    *this = Mat(e);
-}
-
-
-///////////////////////////// SparseMat /////////////////////////////
-
-inline
-SparseMat SparseMat::clone() const
-{
-    SparseMat temp;
-    this->copyTo(temp);
-    return temp;
-}
-
-inline
-size_t SparseMat::elemSize() const
-{
-    return CV_ELEM_SIZE(flags);
-}
-
-inline
-size_t SparseMat::elemSize1() const
-{
-    return CV_ELEM_SIZE1(flags);
-}
-
-inline
-int SparseMat::type() const
-{
-    return CV_MAT_TYPE(flags);
-}
-
-inline
-int SparseMat::depth() const
-{
-    return CV_MAT_DEPTH(flags);
-}
-
-inline
-int SparseMat::channels() const
-{
-    return CV_MAT_CN(flags);
-}
-
-inline
-const int* SparseMat::size() const
-{
-    return hdr ? hdr->size : 0;
-}
-
-inline
-int SparseMat::size(int i) const
-{
-    if( hdr )
-    {
-        CV_DbgAssert((unsigned)i < (unsigned)hdr->dims);
-        return hdr->size[i];
-    }
-    return 0;
-}
-
-inline
-int SparseMat::dims() const
-{
-    return hdr ? hdr->dims : 0;
-}
-
-inline
-size_t SparseMat::nzcount() const
-{
-    return hdr ? hdr->nodeCount : 0;
-}
-
-template<typename _Tp> inline
-_Tp& SparseMat::ref(int i0, size_t* hashval)
-{
-    return *(_Tp*)((SparseMat*)this)->ptr(i0, true, hashval);
-}
-
-template<typename _Tp> inline
-_Tp& SparseMat::ref(int i0, int i1, size_t* hashval)
-{
-    return *(_Tp*)((SparseMat*)this)->ptr(i0, i1, true, hashval);
-}
-
-template<typename _Tp> inline
-_Tp& SparseMat::ref(int i0, int i1, int i2, size_t* hashval)
-{
-    return *(_Tp*)((SparseMat*)this)->ptr(i0, i1, i2, true, hashval);
-}
-
-template<typename _Tp> inline
-_Tp& SparseMat::ref(const int* idx, size_t* hashval)
-{
-    return *(_Tp*)((SparseMat*)this)->ptr(idx, true, hashval);
-}
-
-template<typename _Tp> inline
-_Tp SparseMat::value(int i0, size_t* hashval) const
-{
-    const _Tp* p = (const _Tp*)((SparseMat*)this)->ptr(i0, false, hashval);
-    return p ? *p : _Tp();
-}
-
-template<typename _Tp> inline
-_Tp SparseMat::value(int i0, int i1, size_t* hashval) const
-{
-    const _Tp* p = (const _Tp*)((SparseMat*)this)->ptr(i0, i1, false, hashval);
-    return p ? *p : _Tp();
-}
-
-template<typename _Tp> inline
-_Tp SparseMat::value(int i0, int i1, int i2, size_t* hashval) const
-{
-    const _Tp* p = (const _Tp*)((SparseMat*)this)->ptr(i0, i1, i2, false, hashval);
-    return p ? *p : _Tp();
-}
-
-template<typename _Tp> inline
-_Tp SparseMat::value(const int* idx, size_t* hashval) const
-{
-    const _Tp* p = (const _Tp*)((SparseMat*)this)->ptr(idx, false, hashval);
-    return p ? *p : _Tp();
-}
-
-template<typename _Tp> inline
-const _Tp* SparseMat::find(int i0, size_t* hashval) const
-{
-    return (const _Tp*)((SparseMat*)this)->ptr(i0, false, hashval);
-}
-
-template<typename _Tp> inline
-const _Tp* SparseMat::find(int i0, int i1, size_t* hashval) const
-{
-    return (const _Tp*)((SparseMat*)this)->ptr(i0, i1, false, hashval);
-}
-
-template<typename _Tp> inline
-const _Tp* SparseMat::find(int i0, int i1, int i2, size_t* hashval) const
-{
-    return (const _Tp*)((SparseMat*)this)->ptr(i0, i1, i2, false, hashval);
-}
-
-template<typename _Tp> inline
-const _Tp* SparseMat::find(const int* idx, size_t* hashval) const
-{
-    return (const _Tp*)((SparseMat*)this)->ptr(idx, false, hashval);
-}
-
-template<typename _Tp> inline
-_Tp& SparseMat::value(Node* n)
-{
-    return *(_Tp*)((uchar*)n + hdr->valueOffset);
-}
-
-template<typename _Tp> inline
-const _Tp& SparseMat::value(const Node* n) const
-{
-    return *(const _Tp*)((const uchar*)n + hdr->valueOffset);
-}
-
-inline
-SparseMat::Node* SparseMat::node(size_t nidx)
-{
-    return (Node*)(void*)&hdr->pool[nidx];
-}
-
-inline
-const SparseMat::Node* SparseMat::node(size_t nidx) const
-{
-    return (const Node*)(const void*)&hdr->pool[nidx];
-}
-
-inline
-SparseMatIterator SparseMat::begin()
-{
-    return SparseMatIterator(this);
-}
-
-inline
-SparseMatConstIterator SparseMat::begin() const
-{
-    return SparseMatConstIterator(this);
-}
-
-inline
-SparseMatIterator SparseMat::end()
-{
-    SparseMatIterator it(this);
-    it.seekEnd();
-    return it;
-}
-
-inline
-SparseMatConstIterator SparseMat::end() const
-{
-    SparseMatConstIterator it(this);
-    it.seekEnd();
-    return it;
-}
-
-template<typename _Tp> inline
-SparseMatIterator_<_Tp> SparseMat::begin()
-{
-    return SparseMatIterator_<_Tp>(this);
-}
-
-template<typename _Tp> inline
-SparseMatConstIterator_<_Tp> SparseMat::begin() const
-{
-    return SparseMatConstIterator_<_Tp>(this);
-}
-
-template<typename _Tp> inline
-SparseMatIterator_<_Tp> SparseMat::end()
-{
-    SparseMatIterator_<_Tp> it(this);
-    it.seekEnd();
-    return it;
-}
-
-template<typename _Tp> inline
-SparseMatConstIterator_<_Tp> SparseMat::end() const
-{
-    SparseMatConstIterator_<_Tp> it(this);
-    it.seekEnd();
-    return it;
-}
-
-
-
-///////////////////////////// SparseMat_ ////////////////////////////
-
-template<typename _Tp> inline
-SparseMat_<_Tp>::SparseMat_()
-{
-    flags = MAGIC_VAL + traits::Type<_Tp>::value;
-}
-
-template<typename _Tp> inline
-SparseMat_<_Tp>::SparseMat_(int _dims, const int* _sizes)
-    : SparseMat(_dims, _sizes, traits::Type<_Tp>::value)
-{}
-
-template<typename _Tp> inline
-SparseMat_<_Tp>::SparseMat_(const SparseMat& m)
-{
-    if( m.type() == traits::Type<_Tp>::value )
-        *this = (const SparseMat_<_Tp>&)m;
-    else
-        m.convertTo(*this, traits::Type<_Tp>::value);
-}
-
-template<typename _Tp> inline
-SparseMat_<_Tp>::SparseMat_(const SparseMat_<_Tp>& m)
-{
-    this->flags = m.flags;
-    this->hdr = m.hdr;
-    if( this->hdr )
-        CV_XADD(&this->hdr->refcount, 1);
-}
-
-template<typename _Tp> inline
-SparseMat_<_Tp>::SparseMat_(const Mat& m)
-{
-    SparseMat sm(m);
-    *this = sm;
-}
-
-template<typename _Tp> inline
-SparseMat_<_Tp>& SparseMat_<_Tp>::operator = (const SparseMat_<_Tp>& m)
-{
-    if( this != &m )
-    {
-        if( m.hdr ) CV_XADD(&m.hdr->refcount, 1);
-        release();
-        flags = m.flags;
-        hdr = m.hdr;
-    }
-    return *this;
-}
-
-template<typename _Tp> inline
-SparseMat_<_Tp>& SparseMat_<_Tp>::operator = (const SparseMat& m)
-{
-    if( m.type() == traits::Type<_Tp>::value )
-        return (*this = (const SparseMat_<_Tp>&)m);
-    m.convertTo(*this, traits::Type<_Tp>::value);
-    return *this;
-}
-
-template<typename _Tp> inline
-SparseMat_<_Tp>& SparseMat_<_Tp>::operator = (const Mat& m)
-{
-    return (*this = SparseMat(m));
-}
-
-template<typename _Tp> inline
-SparseMat_<_Tp> SparseMat_<_Tp>::clone() const
-{
-    SparseMat_<_Tp> m;
-    this->copyTo(m);
-    return m;
-}
-
-template<typename _Tp> inline
-void SparseMat_<_Tp>::create(int _dims, const int* _sizes)
-{
-    SparseMat::create(_dims, _sizes, traits::Type<_Tp>::value);
-}
-
-template<typename _Tp> inline
-int SparseMat_<_Tp>::type() const
-{
-    return traits::Type<_Tp>::value;
-}
-
-template<typename _Tp> inline
-int SparseMat_<_Tp>::depth() const
-{
-    return traits::Depth<_Tp>::value;
-}
-
-template<typename _Tp> inline
-int SparseMat_<_Tp>::channels() const
-{
-    return DataType<_Tp>::channels;
-}
-
-template<typename _Tp> inline
-_Tp& SparseMat_<_Tp>::ref(int i0, size_t* hashval)
-{
-    return SparseMat::ref<_Tp>(i0, hashval);
-}
-
-template<typename _Tp> inline
-_Tp SparseMat_<_Tp>::operator()(int i0, size_t* hashval) const
-{
-    return SparseMat::value<_Tp>(i0, hashval);
-}
-
-template<typename _Tp> inline
-_Tp& SparseMat_<_Tp>::ref(int i0, int i1, size_t* hashval)
-{
-    return SparseMat::ref<_Tp>(i0, i1, hashval);
-}
-
-template<typename _Tp> inline
-_Tp SparseMat_<_Tp>::operator()(int i0, int i1, size_t* hashval) const
-{
-    return SparseMat::value<_Tp>(i0, i1, hashval);
-}
-
-template<typename _Tp> inline
-_Tp& SparseMat_<_Tp>::ref(int i0, int i1, int i2, size_t* hashval)
-{
-    return SparseMat::ref<_Tp>(i0, i1, i2, hashval);
-}
-
-template<typename _Tp> inline
-_Tp SparseMat_<_Tp>::operator()(int i0, int i1, int i2, size_t* hashval) const
-{
-    return SparseMat::value<_Tp>(i0, i1, i2, hashval);
-}
-
-template<typename _Tp> inline
-_Tp& SparseMat_<_Tp>::ref(const int* idx, size_t* hashval)
-{
-    return SparseMat::ref<_Tp>(idx, hashval);
-}
-
-template<typename _Tp> inline
-_Tp SparseMat_<_Tp>::operator()(const int* idx, size_t* hashval) const
-{
-    return SparseMat::value<_Tp>(idx, hashval);
-}
-
-template<typename _Tp> inline
-SparseMatIterator_<_Tp> SparseMat_<_Tp>::begin()
-{
-    return SparseMatIterator_<_Tp>(this);
-}
-
-template<typename _Tp> inline
-SparseMatConstIterator_<_Tp> SparseMat_<_Tp>::begin() const
-{
-    return SparseMatConstIterator_<_Tp>(this);
-}
-
-template<typename _Tp> inline
-SparseMatIterator_<_Tp> SparseMat_<_Tp>::end()
-{
-    SparseMatIterator_<_Tp> it(this);
-    it.seekEnd();
-    return it;
-}
-
-template<typename _Tp> inline
-SparseMatConstIterator_<_Tp> SparseMat_<_Tp>::end() const
-{
-    SparseMatConstIterator_<_Tp> it(this);
-    it.seekEnd();
-    return it;
-}
-
-
-
-////////////////////////// MatConstIterator /////////////////////////
-
-inline
-MatConstIterator::MatConstIterator()
-    : m(0), elemSize(0), ptr(0), sliceStart(0), sliceEnd(0)
-{}
-
-inline
-MatConstIterator::MatConstIterator(const Mat* _m)
-    : m(_m), elemSize(_m->elemSize()), ptr(0), sliceStart(0), sliceEnd(0)
-{
-    if( m && m->isContinuous() )
-    {
-        CV_Assert(!m->empty());
-        sliceStart = m->ptr();
-        sliceEnd = sliceStart + m->total()*elemSize;
-    }
-    seek((const int*)0);
-}
-
-inline
-MatConstIterator::MatConstIterator(const Mat* _m, int _row, int _col)
-    : m(_m), elemSize(_m->elemSize()), ptr(0), sliceStart(0), sliceEnd(0)
-{
-    CV_Assert(m && m->dims <= 2);
-    if( m->isContinuous() )
-    {
-        CV_Assert(!m->empty());
-        sliceStart = m->ptr();
-        sliceEnd = sliceStart + m->total()*elemSize;
-    }
-    int idx[] = {_row, _col};
-    seek(idx);
-}
-
-inline
-MatConstIterator::MatConstIterator(const Mat* _m, Point _pt)
-    : m(_m), elemSize(_m->elemSize()), ptr(0), sliceStart(0), sliceEnd(0)
-{
-    CV_Assert(m && m->dims <= 2);
-    if( m->isContinuous() )
-    {
-        CV_Assert(!m->empty());
-        sliceStart = m->ptr();
-        sliceEnd = sliceStart + m->total()*elemSize;
-    }
-    int idx[] = {_pt.y, _pt.x};
-    seek(idx);
-}
-
-inline
-MatConstIterator::MatConstIterator(const MatConstIterator& it)
-    : m(it.m), elemSize(it.elemSize), ptr(it.ptr), sliceStart(it.sliceStart), sliceEnd(it.sliceEnd)
-{}
-
-inline
-MatConstIterator& MatConstIterator::operator = (const MatConstIterator& it )
-{
-    m = it.m; elemSize = it.elemSize; ptr = it.ptr;
-    sliceStart = it.sliceStart; sliceEnd = it.sliceEnd;
-    return *this;
-}
-
-inline
-const uchar* MatConstIterator::operator *() const
-{
-    return ptr;
-}
-
-inline MatConstIterator& MatConstIterator::operator += (ptrdiff_t ofs)
-{
-    if( !m || ofs == 0 )
-        return *this;
-    ptrdiff_t ofsb = ofs*elemSize;
-    ptr += ofsb;
-    if( ptr < sliceStart || sliceEnd <= ptr )
-    {
-        ptr -= ofsb;
-        seek(ofs, true);
-    }
-    return *this;
-}
-
-inline
-MatConstIterator& MatConstIterator::operator -= (ptrdiff_t ofs)
-{
-    return (*this += -ofs);
-}
-
-inline
-MatConstIterator& MatConstIterator::operator --()
-{
-    if( m && (ptr -= elemSize) < sliceStart )
-    {
-        ptr += elemSize;
-        seek(-1, true);
-    }
-    return *this;
-}
-
-inline
-MatConstIterator MatConstIterator::operator --(int)
-{
-    MatConstIterator b = *this;
-    *this += -1;
-    return b;
-}
-
-inline
-MatConstIterator& MatConstIterator::operator ++()
-{
-    if( m && (ptr += elemSize) >= sliceEnd )
-    {
-        ptr -= elemSize;
-        seek(1, true);
-    }
-    return *this;
-}
-
-inline MatConstIterator MatConstIterator::operator ++(int)
-{
-    MatConstIterator b = *this;
-    *this += 1;
-    return b;
-}
-
-
-static inline
-bool operator == (const MatConstIterator& a, const MatConstIterator& b)
-{
-    return a.m == b.m && a.ptr == b.ptr;
-}
-
-static inline
-bool operator != (const MatConstIterator& a, const MatConstIterator& b)
-{
-    return !(a == b);
-}
-
-static inline
-bool operator < (const MatConstIterator& a, const MatConstIterator& b)
-{
-    return a.ptr < b.ptr;
-}
-
-static inline
-bool operator > (const MatConstIterator& a, const MatConstIterator& b)
-{
-    return a.ptr > b.ptr;
-}
-
-static inline
-bool operator <= (const MatConstIterator& a, const MatConstIterator& b)
-{
-    return a.ptr <= b.ptr;
-}
-
-static inline
-bool operator >= (const MatConstIterator& a, const MatConstIterator& b)
-{
-    return a.ptr >= b.ptr;
-}
-
-static inline
-ptrdiff_t operator - (const MatConstIterator& b, const MatConstIterator& a)
-{
-    if( a.m != b.m )
-        return ((size_t)(-1) >> 1);
-    if( a.sliceEnd == b.sliceEnd )
-        return (b.ptr - a.ptr)/static_cast<ptrdiff_t>(b.elemSize);
-
-    return b.lpos() - a.lpos();
-}
-
-static inline
-MatConstIterator operator + (const MatConstIterator& a, ptrdiff_t ofs)
-{
-    MatConstIterator b = a;
-    return b += ofs;
-}
-
-static inline
-MatConstIterator operator + (ptrdiff_t ofs, const MatConstIterator& a)
-{
-    MatConstIterator b = a;
-    return b += ofs;
-}
-
-static inline
-MatConstIterator operator - (const MatConstIterator& a, ptrdiff_t ofs)
-{
-    MatConstIterator b = a;
-    return b += -ofs;
-}
-
-
-inline
-const uchar* MatConstIterator::operator [](ptrdiff_t i) const
-{
-    return *(*this + i);
-}
-
-
-
-///////////////////////// MatConstIterator_ /////////////////////////
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp>::MatConstIterator_()
-{}
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp>::MatConstIterator_(const Mat_<_Tp>* _m)
-    : MatConstIterator(_m)
-{}
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp>::MatConstIterator_(const Mat_<_Tp>* _m, int _row, int _col)
-    : MatConstIterator(_m, _row, _col)
-{}
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp>::MatConstIterator_(const Mat_<_Tp>* _m, Point _pt)
-    : MatConstIterator(_m, _pt)
-{}
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp>::MatConstIterator_(const MatConstIterator_& it)
-    : MatConstIterator(it)
-{}
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp>& MatConstIterator_<_Tp>::operator = (const MatConstIterator_& it )
-{
-    MatConstIterator::operator = (it);
-    return *this;
-}
-
-template<typename _Tp> inline
-const _Tp& MatConstIterator_<_Tp>::operator *() const
-{
-    return *(_Tp*)(this->ptr);
-}
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp>& MatConstIterator_<_Tp>::operator += (ptrdiff_t ofs)
-{
-    MatConstIterator::operator += (ofs);
-    return *this;
-}
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp>& MatConstIterator_<_Tp>::operator -= (ptrdiff_t ofs)
-{
-    return (*this += -ofs);
-}
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp>& MatConstIterator_<_Tp>::operator --()
-{
-    MatConstIterator::operator --();
-    return *this;
-}
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp> MatConstIterator_<_Tp>::operator --(int)
-{
-    MatConstIterator_ b = *this;
-    MatConstIterator::operator --();
-    return b;
-}
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp>& MatConstIterator_<_Tp>::operator ++()
-{
-    MatConstIterator::operator ++();
-    return *this;
-}
-
-template<typename _Tp> inline
-MatConstIterator_<_Tp> MatConstIterator_<_Tp>::operator ++(int)
-{
-    MatConstIterator_ b = *this;
-    MatConstIterator::operator ++();
-    return b;
-}
-
-
-template<typename _Tp> inline
-Point MatConstIterator_<_Tp>::pos() const
-{
-    if( !m )
-        return Point();
-    CV_DbgAssert( m->dims <= 2 );
-    if( m->isContinuous() )
-    {
-        ptrdiff_t ofs = (const _Tp*)ptr - (const _Tp*)m->data;
-        int y = (int)(ofs / m->cols);
-        int x = (int)(ofs - (ptrdiff_t)y * m->cols);
-        return Point(x, y);
-    }
-    else
-    {
-        ptrdiff_t ofs = (uchar*)ptr - m->data;
-        int y = (int)(ofs / m->step);
-        int x = (int)((ofs - y * m->step)/sizeof(_Tp));
-        return Point(x, y);
-    }
-}
-
-
-template<typename _Tp> static inline
-bool operator == (const MatConstIterator_<_Tp>& a, const MatConstIterator_<_Tp>& b)
-{
-    return a.m == b.m && a.ptr == b.ptr;
-}
-
-template<typename _Tp> static inline
-bool operator != (const MatConstIterator_<_Tp>& a, const MatConstIterator_<_Tp>& b)
-{
-    return a.m != b.m || a.ptr != b.ptr;
-}
-
-template<typename _Tp> static inline
-MatConstIterator_<_Tp> operator + (const MatConstIterator_<_Tp>& a, ptrdiff_t ofs)
-{
-    MatConstIterator t = (const MatConstIterator&)a + ofs;
-    return (MatConstIterator_<_Tp>&)t;
-}
-
-template<typename _Tp> static inline
-MatConstIterator_<_Tp> operator + (ptrdiff_t ofs, const MatConstIterator_<_Tp>& a)
-{
-    MatConstIterator t = (const MatConstIterator&)a + ofs;
-    return (MatConstIterator_<_Tp>&)t;
-}
-
-template<typename _Tp> static inline
-MatConstIterator_<_Tp> operator - (const MatConstIterator_<_Tp>& a, ptrdiff_t ofs)
-{
-    MatConstIterator t = (const MatConstIterator&)a - ofs;
-    return (MatConstIterator_<_Tp>&)t;
-}
-
-template<typename _Tp> inline
-const _Tp& MatConstIterator_<_Tp>::operator [](ptrdiff_t i) const
-{
-    return *(_Tp*)MatConstIterator::operator [](i);
-}
-
-
-
-//////////////////////////// MatIterator_ ///////////////////////////
-
-template<typename _Tp> inline
-MatIterator_<_Tp>::MatIterator_()
-    : MatConstIterator_<_Tp>()
-{}
-
-template<typename _Tp> inline
-MatIterator_<_Tp>::MatIterator_(Mat_<_Tp>* _m)
-    : MatConstIterator_<_Tp>(_m)
-{}
-
-template<typename _Tp> inline
-MatIterator_<_Tp>::MatIterator_(Mat_<_Tp>* _m, int _row, int _col)
-    : MatConstIterator_<_Tp>(_m, _row, _col)
-{}
-
-template<typename _Tp> inline
-MatIterator_<_Tp>::MatIterator_(Mat_<_Tp>* _m, Point _pt)
-    : MatConstIterator_<_Tp>(_m, _pt)
-{}
-
-template<typename _Tp> inline
-MatIterator_<_Tp>::MatIterator_(Mat_<_Tp>* _m, const int* _idx)
-    : MatConstIterator_<_Tp>(_m, _idx)
-{}
-
-template<typename _Tp> inline
-MatIterator_<_Tp>::MatIterator_(const MatIterator_& it)
-    : MatConstIterator_<_Tp>(it)
-{}
-
-template<typename _Tp> inline
-MatIterator_<_Tp>& MatIterator_<_Tp>::operator = (const MatIterator_<_Tp>& it )
-{
-    MatConstIterator::operator = (it);
-    return *this;
-}
-
-template<typename _Tp> inline
-_Tp& MatIterator_<_Tp>::operator *() const
-{
-    return *(_Tp*)(this->ptr);
-}
-
-template<typename _Tp> inline
-MatIterator_<_Tp>& MatIterator_<_Tp>::operator += (ptrdiff_t ofs)
-{
-    MatConstIterator::operator += (ofs);
-    return *this;
-}
-
-template<typename _Tp> inline
-MatIterator_<_Tp>& MatIterator_<_Tp>::operator -= (ptrdiff_t ofs)
-{
-    MatConstIterator::operator += (-ofs);
-    return *this;
-}
-
-template<typename _Tp> inline
-MatIterator_<_Tp>& MatIterator_<_Tp>::operator --()
-{
-    MatConstIterator::operator --();
-    return *this;
-}
-
-template<typename _Tp> inline
-MatIterator_<_Tp> MatIterator_<_Tp>::operator --(int)
-{
-    MatIterator_ b = *this;
-    MatConstIterator::operator --();
-    return b;
-}
-
-template<typename _Tp> inline
-MatIterator_<_Tp>& MatIterator_<_Tp>::operator ++()
-{
-    MatConstIterator::operator ++();
-    return *this;
-}
-
-template<typename _Tp> inline
-MatIterator_<_Tp> MatIterator_<_Tp>::operator ++(int)
-{
-    MatIterator_ b = *this;
-    MatConstIterator::operator ++();
-    return b;
-}
-
-template<typename _Tp> inline
-_Tp& MatIterator_<_Tp>::operator [](ptrdiff_t i) const
-{
-    return *(*this + i);
-}
-
-
-template<typename _Tp> static inline
-bool operator == (const MatIterator_<_Tp>& a, const MatIterator_<_Tp>& b)
-{
-    return a.m == b.m && a.ptr == b.ptr;
-}
-
-template<typename _Tp> static inline
-bool operator != (const MatIterator_<_Tp>& a, const MatIterator_<_Tp>& b)
-{
-    return a.m != b.m || a.ptr != b.ptr;
-}
-
-template<typename _Tp> static inline
-MatIterator_<_Tp> operator + (const MatIterator_<_Tp>& a, ptrdiff_t ofs)
-{
-    MatConstIterator t = (const MatConstIterator&)a + ofs;
-    return (MatIterator_<_Tp>&)t;
-}
-
-template<typename _Tp> static inline
-MatIterator_<_Tp> operator + (ptrdiff_t ofs, const MatIterator_<_Tp>& a)
-{
-    MatConstIterator t = (const MatConstIterator&)a + ofs;
-    return (MatIterator_<_Tp>&)t;
-}
-
-template<typename _Tp> static inline
-MatIterator_<_Tp> operator - (const MatIterator_<_Tp>& a, ptrdiff_t ofs)
-{
-    MatConstIterator t = (const MatConstIterator&)a - ofs;
-    return (MatIterator_<_Tp>&)t;
-}
-
-
-
-/////////////////////// SparseMatConstIterator //////////////////////
-
-inline
-SparseMatConstIterator::SparseMatConstIterator()
-    : m(0), hashidx(0), ptr(0)
-{}
-
-inline
-SparseMatConstIterator::SparseMatConstIterator(const SparseMatConstIterator& it)
-    : m(it.m), hashidx(it.hashidx), ptr(it.ptr)
-{}
-
-inline SparseMatConstIterator& SparseMatConstIterator::operator = (const SparseMatConstIterator& it)
-{
-    if( this != &it )
-    {
-        m = it.m;
-        hashidx = it.hashidx;
-        ptr = it.ptr;
-    }
-    return *this;
-}
-
-template<typename _Tp> inline
-const _Tp& SparseMatConstIterator::value() const
-{
-    return *(const _Tp*)ptr;
-}
-
-inline
-const SparseMat::Node* SparseMatConstIterator::node() const
-{
-    return (ptr && m && m->hdr) ? (const SparseMat::Node*)(const void*)(ptr - m->hdr->valueOffset) : 0;
-}
-
-inline
-SparseMatConstIterator SparseMatConstIterator::operator ++(int)
-{
-    SparseMatConstIterator it = *this;
-    ++*this;
-    return it;
-}
-
-inline
-void SparseMatConstIterator::seekEnd()
-{
-    if( m && m->hdr )
-    {
-        hashidx = m->hdr->hashtab.size();
-        ptr = 0;
-    }
-}
-
-
-static inline
-bool operator == (const SparseMatConstIterator& it1, const SparseMatConstIterator& it2)
-{
-    return it1.m == it2.m && it1.ptr == it2.ptr;
-}
-
-static inline
-bool operator != (const SparseMatConstIterator& it1, const SparseMatConstIterator& it2)
-{
-    return !(it1 == it2);
-}
-
-
-
-///////////////////////// SparseMatIterator /////////////////////////
-
-inline
-SparseMatIterator::SparseMatIterator()
-{}
-
-inline
-SparseMatIterator::SparseMatIterator(SparseMat* _m)
-    : SparseMatConstIterator(_m)
-{}
-
-inline
-SparseMatIterator::SparseMatIterator(const SparseMatIterator& it)
-    : SparseMatConstIterator(it)
-{}
-
-inline
-SparseMatIterator& SparseMatIterator::operator = (const SparseMatIterator& it)
-{
-    (SparseMatConstIterator&)*this = it;
-    return *this;
-}
-
-template<typename _Tp> inline
-_Tp& SparseMatIterator::value() const
-{
-    return *(_Tp*)ptr;
-}
-
-inline
-SparseMat::Node* SparseMatIterator::node() const
-{
-    return (SparseMat::Node*)SparseMatConstIterator::node();
-}
-
-inline
-SparseMatIterator& SparseMatIterator::operator ++()
-{
-    SparseMatConstIterator::operator ++();
-    return *this;
-}
-
-inline
-SparseMatIterator SparseMatIterator::operator ++(int)
-{
-    SparseMatIterator it = *this;
-    ++*this;
-    return it;
-}
-
-
-
-////////////////////// SparseMatConstIterator_ //////////////////////
-
-template<typename _Tp> inline
-SparseMatConstIterator_<_Tp>::SparseMatConstIterator_()
-{}
-
-template<typename _Tp> inline
-SparseMatConstIterator_<_Tp>::SparseMatConstIterator_(const SparseMat_<_Tp>* _m)
-    : SparseMatConstIterator(_m)
-{}
-
-template<typename _Tp> inline
-SparseMatConstIterator_<_Tp>::SparseMatConstIterator_(const SparseMat* _m)
-    : SparseMatConstIterator(_m)
-{
-    CV_Assert( _m->type() == traits::Type<_Tp>::value );
-}
-
-template<typename _Tp> inline
-SparseMatConstIterator_<_Tp>::SparseMatConstIterator_(const SparseMatConstIterator_<_Tp>& it)
-    : SparseMatConstIterator(it)
-{}
-
-template<typename _Tp> inline
-SparseMatConstIterator_<_Tp>& SparseMatConstIterator_<_Tp>::operator = (const SparseMatConstIterator_<_Tp>& it)
-{
-    return reinterpret_cast<SparseMatConstIterator_<_Tp>&>
-         (*reinterpret_cast<SparseMatConstIterator*>(this) =
-           reinterpret_cast<const SparseMatConstIterator&>(it));
-}
-
-template<typename _Tp> inline
-const _Tp& SparseMatConstIterator_<_Tp>::operator *() const
-{
-    return *(const _Tp*)this->ptr;
-}
-
-template<typename _Tp> inline
-SparseMatConstIterator_<_Tp>& SparseMatConstIterator_<_Tp>::operator ++()
-{
-    SparseMatConstIterator::operator ++();
-    return *this;
-}
-
-template<typename _Tp> inline
-SparseMatConstIterator_<_Tp> SparseMatConstIterator_<_Tp>::operator ++(int)
-{
-    SparseMatConstIterator_<_Tp> it = *this;
-    SparseMatConstIterator::operator ++();
-    return it;
-}
-
-
-
-///////////////////////// SparseMatIterator_ ////////////////////////
-
-template<typename _Tp> inline
-SparseMatIterator_<_Tp>::SparseMatIterator_()
-{}
-
-template<typename _Tp> inline
-SparseMatIterator_<_Tp>::SparseMatIterator_(SparseMat_<_Tp>* _m)
-    : SparseMatConstIterator_<_Tp>(_m)
-{}
-
-template<typename _Tp> inline
-SparseMatIterator_<_Tp>::SparseMatIterator_(SparseMat* _m)
-    : SparseMatConstIterator_<_Tp>(_m)
-{}
-
-template<typename _Tp> inline
-SparseMatIterator_<_Tp>::SparseMatIterator_(const SparseMatIterator_<_Tp>& it)
-    : SparseMatConstIterator_<_Tp>(it)
-{}
-
-template<typename _Tp> inline
-SparseMatIterator_<_Tp>& SparseMatIterator_<_Tp>::operator = (const SparseMatIterator_<_Tp>& it)
-{
-    return reinterpret_cast<SparseMatIterator_<_Tp>&>
-         (*reinterpret_cast<SparseMatConstIterator*>(this) =
-           reinterpret_cast<const SparseMatConstIterator&>(it));
-}
-
-template<typename _Tp> inline
-_Tp& SparseMatIterator_<_Tp>::operator *() const
-{
-    return *(_Tp*)this->ptr;
-}
-
-template<typename _Tp> inline
-SparseMatIterator_<_Tp>& SparseMatIterator_<_Tp>::operator ++()
-{
-    SparseMatConstIterator::operator ++();
-    return *this;
-}
-
-template<typename _Tp> inline
-SparseMatIterator_<_Tp> SparseMatIterator_<_Tp>::operator ++(int)
-{
-    SparseMatIterator_<_Tp> it = *this;
-    SparseMatConstIterator::operator ++();
-    return it;
-}
-
-
-
-//////////////////////// MatCommaInitializer_ ///////////////////////
-
-template<typename _Tp> inline
-MatCommaInitializer_<_Tp>::MatCommaInitializer_(Mat_<_Tp>* _m)
-    : it(_m)
-{}
-
-template<typename _Tp> template<typename T2> inline
-MatCommaInitializer_<_Tp>& MatCommaInitializer_<_Tp>::operator , (T2 v)
-{
-    CV_DbgAssert( this->it < ((const Mat_<_Tp>*)this->it.m)->end() );
-    *this->it = _Tp(v);
-    ++this->it;
-    return *this;
-}
-
-template<typename _Tp> inline
-MatCommaInitializer_<_Tp>::operator Mat_<_Tp>() const
-{
-    CV_DbgAssert( this->it == ((const Mat_<_Tp>*)this->it.m)->end() );
-    return Mat_<_Tp>(*this->it.m);
-}
-
-
-template<typename _Tp, typename T2> static inline
-MatCommaInitializer_<_Tp> operator << (const Mat_<_Tp>& m, T2 val)
-{
-    MatCommaInitializer_<_Tp> commaInitializer((Mat_<_Tp>*)&m);
-    return (commaInitializer, val);
-}
-
-
-
-///////////////////////// Matrix Expressions ////////////////////////
-
-inline
-Mat& Mat::operator = (const MatExpr& e)
-{
-    e.op->assign(e, *this);
-    return *this;
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>::Mat_(const MatExpr& e)
-{
-    e.op->assign(e, *this, traits::Type<_Tp>::value);
-}
-
-template<typename _Tp> inline
-Mat_<_Tp>& Mat_<_Tp>::operator = (const MatExpr& e)
-{
-    e.op->assign(e, *this, traits::Type<_Tp>::value);
-    return *this;
-}
-
-template<typename _Tp> inline
-MatExpr Mat_<_Tp>::zeros(int rows, int cols)
-{
-    return Mat::zeros(rows, cols, traits::Type<_Tp>::value);
-}
-
-template<typename _Tp> inline
-MatExpr Mat_<_Tp>::zeros(Size sz)
-{
-    return Mat::zeros(sz, traits::Type<_Tp>::value);
-}
-
-template<typename _Tp> inline
-MatExpr Mat_<_Tp>::ones(int rows, int cols)
-{
-    return Mat::ones(rows, cols, traits::Type<_Tp>::value);
-}
-
-template<typename _Tp> inline
-MatExpr Mat_<_Tp>::ones(Size sz)
-{
-    return Mat::ones(sz, traits::Type<_Tp>::value);
-}
-
-template<typename _Tp> inline
-MatExpr Mat_<_Tp>::eye(int rows, int cols)
-{
-    return Mat::eye(rows, cols, traits::Type<_Tp>::value);
-}
-
-template<typename _Tp> inline
-MatExpr Mat_<_Tp>::eye(Size sz)
-{
-    return Mat::eye(sz, traits::Type<_Tp>::value);
-}
-
-inline
-MatExpr::MatExpr()
-    : op(0), flags(0), a(Mat()), b(Mat()), c(Mat()), alpha(0), beta(0), s()
-{}
-
-inline
-MatExpr::MatExpr(const MatOp* _op, int _flags, const Mat& _a, const Mat& _b,
-                 const Mat& _c, double _alpha, double _beta, const Scalar& _s)
-    : op(_op), flags(_flags), a(_a), b(_b), c(_c), alpha(_alpha), beta(_beta), s(_s)
-{}
-
-inline
-MatExpr::operator Mat() const
-{
-    Mat m;
-    op->assign(*this, m);
-    return m;
-}
-
-template<typename _Tp> inline
-MatExpr::operator Mat_<_Tp>() const
-{
-    Mat_<_Tp> m;
-    op->assign(*this, m, traits::Type<_Tp>::value);
-    return m;
-}
-
-
-template<typename _Tp> static inline
-MatExpr min(const Mat_<_Tp>& a, const Mat_<_Tp>& b)
-{
-    return cv::min((const Mat&)a, (const Mat&)b);
-}
-
-template<typename _Tp> static inline
-MatExpr min(const Mat_<_Tp>& a, double s)
-{
-    return cv::min((const Mat&)a, s);
-}
-
-template<typename _Tp> static inline
-MatExpr min(double s, const Mat_<_Tp>& a)
-{
-    return cv::min((const Mat&)a, s);
-}
-
-template<typename _Tp> static inline
-MatExpr max(const Mat_<_Tp>& a, const Mat_<_Tp>& b)
-{
-    return cv::max((const Mat&)a, (const Mat&)b);
-}
-
-template<typename _Tp> static inline
-MatExpr max(const Mat_<_Tp>& a, double s)
-{
-    return cv::max((const Mat&)a, s);
-}
-
-template<typename _Tp> static inline
-MatExpr max(double s, const Mat_<_Tp>& a)
-{
-    return cv::max((const Mat&)a, s);
-}
-
-template<typename _Tp> static inline
-MatExpr abs(const Mat_<_Tp>& m)
-{
-    return cv::abs((const Mat&)m);
-}
-
-
-static inline
-Mat& operator += (Mat& a, const MatExpr& b)
-{
-    b.op->augAssignAdd(b, a);
-    return a;
-}
-
-static inline
-const Mat& operator += (const Mat& a, const MatExpr& b)
-{
-    b.op->augAssignAdd(b, (Mat&)a);
-    return a;
-}
-
-template<typename _Tp> static inline
-Mat_<_Tp>& operator += (Mat_<_Tp>& a, const MatExpr& b)
-{
-    b.op->augAssignAdd(b, a);
-    return a;
-}
-
-template<typename _Tp> static inline
-const Mat_<_Tp>& operator += (const Mat_<_Tp>& a, const MatExpr& b)
-{
-    b.op->augAssignAdd(b, (Mat&)a);
-    return a;
-}
-
-static inline
-Mat& operator -= (Mat& a, const MatExpr& b)
-{
-    b.op->augAssignSubtract(b, a);
-    return a;
-}
-
-static inline
-const Mat& operator -= (const Mat& a, const MatExpr& b)
-{
-    b.op->augAssignSubtract(b, (Mat&)a);
-    return a;
-}
-
-template<typename _Tp> static inline
-Mat_<_Tp>& operator -= (Mat_<_Tp>& a, const MatExpr& b)
-{
-    b.op->augAssignSubtract(b, a);
-    return a;
-}
-
-template<typename _Tp> static inline
-const Mat_<_Tp>& operator -= (const Mat_<_Tp>& a, const MatExpr& b)
-{
-    b.op->augAssignSubtract(b, (Mat&)a);
-    return a;
-}
-
-static inline
-Mat& operator *= (Mat& a, const MatExpr& b)
-{
-    b.op->augAssignMultiply(b, a);
-    return a;
-}
-
-static inline
-const Mat& operator *= (const Mat& a, const MatExpr& b)
-{
-    b.op->augAssignMultiply(b, (Mat&)a);
-    return a;
-}
-
-template<typename _Tp> static inline
-Mat_<_Tp>& operator *= (Mat_<_Tp>& a, const MatExpr& b)
-{
-    b.op->augAssignMultiply(b, a);
-    return a;
-}
-
-template<typename _Tp> static inline
-const Mat_<_Tp>& operator *= (const Mat_<_Tp>& a, const MatExpr& b)
-{
-    b.op->augAssignMultiply(b, (Mat&)a);
-    return a;
-}
-
-static inline
-Mat& operator /= (Mat& a, const MatExpr& b)
-{
-    b.op->augAssignDivide(b, a);
-    return a;
-}
-
-static inline
-const Mat& operator /= (const Mat& a, const MatExpr& b)
-{
-    b.op->augAssignDivide(b, (Mat&)a);
-    return a;
-}
-
-template<typename _Tp> static inline
-Mat_<_Tp>& operator /= (Mat_<_Tp>& a, const MatExpr& b)
-{
-    b.op->augAssignDivide(b, a);
-    return a;
-}
-
-template<typename _Tp> static inline
-const Mat_<_Tp>& operator /= (const Mat_<_Tp>& a, const MatExpr& b)
-{
-    b.op->augAssignDivide(b, (Mat&)a);
-    return a;
-}
-
-
-//////////////////////////////// UMat ////////////////////////////////
-
-template<typename _Tp> inline
-UMat::UMat(const std::vector<_Tp>& vec, bool copyData)
-: flags(MAGIC_VAL + traits::Type<_Tp>::value + CV_MAT_CONT_FLAG), dims(2), rows((int)vec.size()),
-cols(1), allocator(0), usageFlags(USAGE_DEFAULT), u(0), offset(0), size(&rows)
-{
-    if(vec.empty())
-        return;
-    if( !copyData )
-    {
-        // !!!TODO!!!
-        CV_Error(Error::StsNotImplemented, "");
-    }
-    else
-        Mat((int)vec.size(), 1, traits::Type<_Tp>::value, (uchar*)&vec[0]).copyTo(*this);
-}
-
-inline
-UMat UMat::row(int y) const
-{
-    return UMat(*this, Range(y, y + 1), Range::all());
-}
-
-inline
-UMat UMat::col(int x) const
-{
-    return UMat(*this, Range::all(), Range(x, x + 1));
-}
-
-inline
-UMat UMat::rowRange(int startrow, int endrow) const
-{
-    return UMat(*this, Range(startrow, endrow), Range::all());
-}
-
-inline
-UMat UMat::rowRange(const Range& r) const
-{
-    return UMat(*this, r, Range::all());
-}
-
-inline
-UMat UMat::colRange(int startcol, int endcol) const
-{
-    return UMat(*this, Range::all(), Range(startcol, endcol));
-}
-
-inline
-UMat UMat::colRange(const Range& r) const
-{
-    return UMat(*this, Range::all(), r);
-}
-
-inline
-UMat UMat::operator()( Range _rowRange, Range _colRange ) const
-{
-    return UMat(*this, _rowRange, _colRange);
-}
-
-inline
-UMat UMat::operator()( const Rect& roi ) const
-{
-    return UMat(*this, roi);
-}
-
-inline
-UMat UMat::operator()(const Range* ranges) const
-{
-    return UMat(*this, ranges);
-}
-
-inline
-UMat UMat::operator()(const std::vector<Range>& ranges) const
-{
-    return UMat(*this, ranges);
-}
-
-inline
-bool UMat::isContinuous() const
-{
-    return (flags & CONTINUOUS_FLAG) != 0;
-}
-
-inline
-bool UMat::isSubmatrix() const
-{
-    return (flags & SUBMATRIX_FLAG) != 0;
-}
-
-inline
-size_t UMat::elemSize() const
-{
-    size_t res = dims > 0 ? step.p[dims - 1] : 0;
-    CV_DbgAssert(res != 0);
-    return res;
-}
-
-inline
-size_t UMat::elemSize1() const
-{
-    return CV_ELEM_SIZE1(flags);
-}
-
-inline
-int UMat::type() const
-{
-    return CV_MAT_TYPE(flags);
-}
-
-inline
-int UMat::depth() const
-{
-    return CV_MAT_DEPTH(flags);
-}
-
-inline
-int UMat::channels() const
-{
-    return CV_MAT_CN(flags);
-}
-
-inline
-size_t UMat::step1(int i) const
-{
-    return step.p[i] / elemSize1();
-}
-
-
-inline bool UMatData::hostCopyObsolete() const { return (flags & HOST_COPY_OBSOLETE) != 0; }
-inline bool UMatData::deviceCopyObsolete() const { return (flags & DEVICE_COPY_OBSOLETE) != 0; }
-inline bool UMatData::deviceMemMapped() const { return (flags & DEVICE_MEM_MAPPED) != 0; }
-inline bool UMatData::copyOnMap() const { return (flags & COPY_ON_MAP) != 0; }
-inline bool UMatData::tempUMat() const { return (flags & TEMP_UMAT) != 0; }
-inline bool UMatData::tempCopiedUMat() const { return (flags & TEMP_COPIED_UMAT) == TEMP_COPIED_UMAT; }
-
-inline void UMatData::markDeviceMemMapped(bool flag)
-{
-  if(flag)
-    flags |= DEVICE_MEM_MAPPED;
-  else
-    flags &= ~DEVICE_MEM_MAPPED;
-}
-
-inline void UMatData::markHostCopyObsolete(bool flag)
-{
-    if(flag)
-        flags |= HOST_COPY_OBSOLETE;
-    else
-        flags &= ~HOST_COPY_OBSOLETE;
-}
-inline void UMatData::markDeviceCopyObsolete(bool flag)
-{
-    if(flag)
-        flags |= DEVICE_COPY_OBSOLETE;
-    else
-        flags &= ~DEVICE_COPY_OBSOLETE;
-}
-
-//! @endcond
-
-static inline
-void swap(MatExpr& a, MatExpr& b) { a.swap(b); }
-
-} //cv
-
-#ifdef _MSC_VER
-#pragma warning( pop )
-#endif
-
-#ifdef CV_DISABLE_CLANG_ENUM_WARNINGS
-#undef CV_DISABLE_CLANG_ENUM_WARNINGS
-#pragma clang diagnostic pop
-#endif
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/matx.hpp b/projects/llm_framework/include/opencv4/opencv2/core/matx.hpp
deleted file mode 100644
index 3c92e3a2..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/matx.hpp
+++ /dev/null
@@ -1,1508 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_MATX_HPP
-#define OPENCV_CORE_MATX_HPP
-
-#ifndef __cplusplus
-#  error matx.hpp header must be compiled as C++
-#endif
-
-#include "opencv2/core/cvdef.h"
-#include "opencv2/core/base.hpp"
-#include "opencv2/core/traits.hpp"
-#include "opencv2/core/saturate.hpp"
-
-#include <initializer_list>
-
-namespace cv
-{
-
-//! @addtogroup core_basic
-//! @{
-
-////////////////////////////// Small Matrix ///////////////////////////
-
-//! @cond IGNORED
-// FIXIT Remove this (especially CV_EXPORTS modifier)
-struct CV_EXPORTS Matx_AddOp { Matx_AddOp() {} Matx_AddOp(const Matx_AddOp&) {} };
-struct CV_EXPORTS Matx_SubOp { Matx_SubOp() {} Matx_SubOp(const Matx_SubOp&) {} };
-struct CV_EXPORTS Matx_ScaleOp { Matx_ScaleOp() {} Matx_ScaleOp(const Matx_ScaleOp&) {} };
-struct CV_EXPORTS Matx_MulOp { Matx_MulOp() {} Matx_MulOp(const Matx_MulOp&) {} };
-struct CV_EXPORTS Matx_DivOp { Matx_DivOp() {} Matx_DivOp(const Matx_DivOp&) {} };
-struct CV_EXPORTS Matx_MatMulOp { Matx_MatMulOp() {} Matx_MatMulOp(const Matx_MatMulOp&) {} };
-struct CV_EXPORTS Matx_TOp { Matx_TOp() {} Matx_TOp(const Matx_TOp&) {} };
-//! @endcond
-
-/** @brief Template class for small matrices whose type and size are known at compilation time
-
-If you need a more flexible type, use Mat . The elements of the matrix M are accessible using the
-M(i,j) notation. Most of the common matrix operations (see also @ref MatrixExpressions ) are
-available. To do an operation on Matx that is not implemented, you can easily convert the matrix to
-Mat and backwards:
-@code{.cpp}
-    Matx33f m(1, 2, 3,
-              4, 5, 6,
-              7, 8, 9);
-    cout << sum(Mat(m*m.t())) << endl;
-@endcode
-Except of the plain constructor which takes a list of elements, Matx can be initialized from a C-array:
-@code{.cpp}
-    float values[] = { 1, 2, 3};
-    Matx31f m(values);
-@endcode
-In case if C++11 features are available, std::initializer_list can be also used to initialize Matx:
-@code{.cpp}
-    Matx31f m = { 1, 2, 3};
-@endcode
- */
-template<typename _Tp, int m, int n> class Matx
-{
-public:
-    enum {
-           rows     = m,
-           cols     = n,
-           channels = rows*cols,
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-           depth    = traits::Type<_Tp>::value,
-           type     = CV_MAKETYPE(depth, channels),
-#endif
-           shortdim = (m < n ? m : n)
-         };
-
-    typedef _Tp                           value_type;
-    typedef Matx<_Tp, m, n>               mat_type;
-    typedef Matx<_Tp, shortdim, 1> diag_type;
-
-    //! default constructor
-    Matx();
-
-    explicit Matx(_Tp v0); //!< 1x1 matrix
-    Matx(_Tp v0, _Tp v1); //!< 1x2 or 2x1 matrix
-    Matx(_Tp v0, _Tp v1, _Tp v2); //!< 1x3 or 3x1 matrix
-    Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3); //!< 1x4, 2x2 or 4x1 matrix
-    Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4); //!< 1x5 or 5x1 matrix
-    Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5); //!< 1x6, 2x3, 3x2 or 6x1 matrix
-    Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6); //!< 1x7 or 7x1 matrix
-    Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7); //!< 1x8, 2x4, 4x2 or 8x1 matrix
-    Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7, _Tp v8); //!< 1x9, 3x3 or 9x1 matrix
-    Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7, _Tp v8, _Tp v9); //!< 1x10, 2x5 or 5x2 or 10x1 matrix
-    Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3,
-         _Tp v4, _Tp v5, _Tp v6, _Tp v7,
-         _Tp v8, _Tp v9, _Tp v10, _Tp v11); //!< 1x12, 2x6, 3x4, 4x3, 6x2 or 12x1 matrix
-    Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3,
-         _Tp v4, _Tp v5, _Tp v6, _Tp v7,
-         _Tp v8, _Tp v9, _Tp v10, _Tp v11,
-         _Tp v12, _Tp v13); //!< 1x14, 2x7, 7x2 or 14x1 matrix
-    Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3,
-         _Tp v4, _Tp v5, _Tp v6, _Tp v7,
-         _Tp v8, _Tp v9, _Tp v10, _Tp v11,
-         _Tp v12, _Tp v13, _Tp v14, _Tp v15); //!< 1x16, 4x4 or 16x1 matrix
-    explicit Matx(const _Tp* vals); //!< initialize from a plain array
-
-    Matx(std::initializer_list<_Tp>); //!< initialize from an initializer list
-
-    CV_NODISCARD_STD static Matx all(_Tp alpha);
-    CV_NODISCARD_STD static Matx zeros();
-    CV_NODISCARD_STD static Matx ones();
-    CV_NODISCARD_STD static Matx eye();
-    CV_NODISCARD_STD static Matx diag(const diag_type& d);
-    /** @brief Generates uniformly distributed random numbers
-    @param a Range boundary.
-    @param b The other range boundary (boundaries don't have to be ordered, the lower boundary is inclusive,
-    the upper one is exclusive).
-     */
-    CV_NODISCARD_STD static Matx randu(_Tp a, _Tp b);
-    /** @brief Generates normally distributed random numbers
-    @param a Mean value.
-    @param b Standard deviation.
-     */
-    CV_NODISCARD_STD static Matx randn(_Tp a, _Tp b);
-
-    //! dot product computed with the default precision
-    _Tp dot(const Matx<_Tp, m, n>& v) const;
-
-    //! dot product computed in double-precision arithmetics
-    double ddot(const Matx<_Tp, m, n>& v) const;
-
-    //! conversion to another data type
-    template<typename T2> operator Matx<T2, m, n>() const;
-
-    //! change the matrix shape
-    template<int m1, int n1> Matx<_Tp, m1, n1> reshape() const;
-
-    //! extract part of the matrix
-    template<int m1, int n1> Matx<_Tp, m1, n1> get_minor(int base_row, int base_col) const;
-
-    //! extract the matrix row
-    Matx<_Tp, 1, n> row(int i) const;
-
-    //! extract the matrix column
-    Matx<_Tp, m, 1> col(int i) const;
-
-    //! extract the matrix diagonal
-    diag_type diag() const;
-
-    //! transpose the matrix
-    Matx<_Tp, n, m> t() const;
-
-    //! invert the matrix
-    Matx<_Tp, n, m> inv(int method=DECOMP_LU, bool *p_is_ok = NULL) const;
-
-    //! solve linear system
-    template<int l> Matx<_Tp, n, l> solve(const Matx<_Tp, m, l>& rhs, int flags=DECOMP_LU) const;
-    Vec<_Tp, n> solve(const Vec<_Tp, m>& rhs, int method) const;
-
-    //! multiply two matrices element-wise
-    Matx<_Tp, m, n> mul(const Matx<_Tp, m, n>& a) const;
-
-    //! divide two matrices element-wise
-    Matx<_Tp, m, n> div(const Matx<_Tp, m, n>& a) const;
-
-    //! element access
-    const _Tp& operator ()(int row, int col) const;
-    _Tp& operator ()(int row, int col);
-
-    //! 1D element access
-    const _Tp& operator ()(int i) const;
-    _Tp& operator ()(int i);
-
-    Matx(const Matx<_Tp, m, n>& a, const Matx<_Tp, m, n>& b, Matx_AddOp);
-    Matx(const Matx<_Tp, m, n>& a, const Matx<_Tp, m, n>& b, Matx_SubOp);
-    template<typename _T2> Matx(const Matx<_Tp, m, n>& a, _T2 alpha, Matx_ScaleOp);
-    Matx(const Matx<_Tp, m, n>& a, const Matx<_Tp, m, n>& b, Matx_MulOp);
-    Matx(const Matx<_Tp, m, n>& a, const Matx<_Tp, m, n>& b, Matx_DivOp);
-    template<int l> Matx(const Matx<_Tp, m, l>& a, const Matx<_Tp, l, n>& b, Matx_MatMulOp);
-    Matx(const Matx<_Tp, n, m>& a, Matx_TOp);
-
-    _Tp val[m*n]; //< matrix elements
-};
-
-typedef Matx<float, 1, 2> Matx12f;
-typedef Matx<double, 1, 2> Matx12d;
-typedef Matx<float, 1, 3> Matx13f;
-typedef Matx<double, 1, 3> Matx13d;
-typedef Matx<float, 1, 4> Matx14f;
-typedef Matx<double, 1, 4> Matx14d;
-typedef Matx<float, 1, 6> Matx16f;
-typedef Matx<double, 1, 6> Matx16d;
-
-typedef Matx<float, 2, 1> Matx21f;
-typedef Matx<double, 2, 1> Matx21d;
-typedef Matx<float, 3, 1> Matx31f;
-typedef Matx<double, 3, 1> Matx31d;
-typedef Matx<float, 4, 1> Matx41f;
-typedef Matx<double, 4, 1> Matx41d;
-typedef Matx<float, 6, 1> Matx61f;
-typedef Matx<double, 6, 1> Matx61d;
-
-typedef Matx<float, 2, 2> Matx22f;
-typedef Matx<double, 2, 2> Matx22d;
-typedef Matx<float, 2, 3> Matx23f;
-typedef Matx<double, 2, 3> Matx23d;
-typedef Matx<float, 3, 2> Matx32f;
-typedef Matx<double, 3, 2> Matx32d;
-
-typedef Matx<float, 3, 3> Matx33f;
-typedef Matx<double, 3, 3> Matx33d;
-
-typedef Matx<float, 3, 4> Matx34f;
-typedef Matx<double, 3, 4> Matx34d;
-typedef Matx<float, 4, 3> Matx43f;
-typedef Matx<double, 4, 3> Matx43d;
-
-typedef Matx<float, 4, 4> Matx44f;
-typedef Matx<double, 4, 4> Matx44d;
-typedef Matx<float, 6, 6> Matx66f;
-typedef Matx<double, 6, 6> Matx66d;
-
-/*!
-  traits
-*/
-template<typename _Tp, int m, int n> class DataType< Matx<_Tp, m, n> >
-{
-public:
-    typedef Matx<_Tp, m, n>                               value_type;
-    typedef Matx<typename DataType<_Tp>::work_type, m, n> work_type;
-    typedef _Tp                                           channel_type;
-    typedef value_type                                    vec_type;
-
-    enum { generic_type = 0,
-           channels     = m * n,
-           fmt          = traits::SafeFmt<channel_type>::fmt + ((channels - 1) << 8)
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-           ,depth        = DataType<channel_type>::depth
-           ,type         = CV_MAKETYPE(depth, channels)
-#endif
-         };
-};
-
-namespace traits {
-template<typename _Tp, int m, int n>
-struct Depth< Matx<_Tp, m, n> > { enum { value = Depth<_Tp>::value }; };
-template<typename _Tp, int m, int n>
-struct Type< Matx<_Tp, m, n> > { enum { value = CV_MAKETYPE(Depth<_Tp>::value, n*m) }; };
-} // namespace
-
-
-/** @brief  Comma-separated Matrix Initializer
-*/
-template<typename _Tp, int m, int n> class MatxCommaInitializer
-{
-public:
-    MatxCommaInitializer(Matx<_Tp, m, n>* _mtx);
-    template<typename T2> MatxCommaInitializer<_Tp, m, n>& operator , (T2 val);
-    Matx<_Tp, m, n> operator *() const;
-
-    Matx<_Tp, m, n>* dst;
-    int idx;
-};
-
-/*
- Utility methods
-*/
-template<typename _Tp, int m> static double determinant(const Matx<_Tp, m, m>& a);
-template<typename _Tp, int m, int n> static double trace(const Matx<_Tp, m, n>& a);
-template<typename _Tp, int m, int n> static double norm(const Matx<_Tp, m, n>& M);
-template<typename _Tp, int m, int n> static double norm(const Matx<_Tp, m, n>& M, int normType);
-
-
-
-/////////////////////// Vec (used as element of multi-channel images /////////////////////
-
-/** @brief Template class for short numerical vectors, a partial case of Matx
-
-This template class represents short numerical vectors (of 1, 2, 3, 4 ... elements) on which you
-can perform basic arithmetical operations, access individual elements using [] operator etc. The
-vectors are allocated on stack, as opposite to std::valarray, std::vector, cv::Mat etc., which
-elements are dynamically allocated in the heap.
-
-The template takes 2 parameters:
-@tparam _Tp element type
-@tparam cn the number of elements
-
-In addition to the universal notation like Vec<float, 3>, you can use shorter aliases
-for the most popular specialized variants of Vec, e.g. Vec3f ~ Vec<float, 3>.
-
-It is possible to convert Vec\<T,2\> to/from Point_, Vec\<T,3\> to/from Point3_ , and Vec\<T,4\>
-to CvScalar or Scalar_. Use operator[] to access the elements of Vec.
-
-All the expected vector operations are also implemented:
--   v1 = v2 + v3
--   v1 = v2 - v3
--   v1 = v2 \* scale
--   v1 = scale \* v2
--   v1 = -v2
--   v1 += v2 and other augmenting operations
--   v1 == v2, v1 != v2
--   norm(v1) (euclidean norm)
-The Vec class is commonly used to describe pixel types of multi-channel arrays. See Mat for details.
-*/
-template<typename _Tp, int cn> class Vec : public Matx<_Tp, cn, 1>
-{
-public:
-    typedef _Tp value_type;
-    enum {
-           channels = cn,
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-           depth    = Matx<_Tp, cn, 1>::depth,
-           type     = CV_MAKETYPE(depth, channels),
-#endif
-           _dummy_enum_finalizer = 0
-         };
-
-    //! default constructor
-    Vec();
-
-    Vec(_Tp v0); //!< 1-element vector constructor
-    Vec(_Tp v0, _Tp v1); //!< 2-element vector constructor
-    Vec(_Tp v0, _Tp v1, _Tp v2); //!< 3-element vector constructor
-    Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3); //!< 4-element vector constructor
-    Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4); //!< 5-element vector constructor
-    Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5); //!< 6-element vector constructor
-    Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6); //!< 7-element vector constructor
-    Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7); //!< 8-element vector constructor
-    Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7, _Tp v8); //!< 9-element vector constructor
-    Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7, _Tp v8, _Tp v9); //!< 10-element vector constructor
-    Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7, _Tp v8, _Tp v9, _Tp v10, _Tp v11, _Tp v12, _Tp v13); //!< 14-element vector constructor
-    explicit Vec(const _Tp* values);
-
-    Vec(std::initializer_list<_Tp>);
-
-    Vec(const Vec<_Tp, cn>& v);
-
-    static Vec all(_Tp alpha);
-
-    //! per-element multiplication
-    Vec mul(const Vec<_Tp, cn>& v) const;
-
-    //! conjugation (makes sense for complex numbers and quaternions)
-    Vec conj() const;
-
-    /*!
-      cross product of the two 3D vectors.
-
-      For other dimensionalities the exception is raised
-    */
-    Vec cross(const Vec& v) const;
-    //! conversion to another data type
-    template<typename T2> operator Vec<T2, cn>() const;
-
-    /*! element access */
-    const _Tp& operator [](int i) const;
-    _Tp& operator[](int i);
-    const _Tp& operator ()(int i) const;
-    _Tp& operator ()(int i);
-
-#ifdef CV_CXX11
-    Vec<_Tp, cn>& operator=(const Vec<_Tp, cn>& rhs) = default;
-#endif
-
-    Vec(const Matx<_Tp, cn, 1>& a, const Matx<_Tp, cn, 1>& b, Matx_AddOp);
-    Vec(const Matx<_Tp, cn, 1>& a, const Matx<_Tp, cn, 1>& b, Matx_SubOp);
-    template<typename _T2> Vec(const Matx<_Tp, cn, 1>& a, _T2 alpha, Matx_ScaleOp);
-};
-
-/** @name Shorter aliases for the most popular specializations of Vec<T,n>
-  @{
-*/
-typedef Vec<uchar, 2> Vec2b;
-typedef Vec<uchar, 3> Vec3b;
-typedef Vec<uchar, 4> Vec4b;
-
-typedef Vec<short, 2> Vec2s;
-typedef Vec<short, 3> Vec3s;
-typedef Vec<short, 4> Vec4s;
-
-typedef Vec<ushort, 2> Vec2w;
-typedef Vec<ushort, 3> Vec3w;
-typedef Vec<ushort, 4> Vec4w;
-
-typedef Vec<int, 2> Vec2i;
-typedef Vec<int, 3> Vec3i;
-typedef Vec<int, 4> Vec4i;
-typedef Vec<int, 6> Vec6i;
-typedef Vec<int, 8> Vec8i;
-
-typedef Vec<float, 2> Vec2f;
-typedef Vec<float, 3> Vec3f;
-typedef Vec<float, 4> Vec4f;
-typedef Vec<float, 6> Vec6f;
-
-typedef Vec<double, 2> Vec2d;
-typedef Vec<double, 3> Vec3d;
-typedef Vec<double, 4> Vec4d;
-typedef Vec<double, 6> Vec6d;
-/** @} */
-
-/*!
-  traits
-*/
-template<typename _Tp, int cn> class DataType< Vec<_Tp, cn> >
-{
-public:
-    typedef Vec<_Tp, cn>                               value_type;
-    typedef Vec<typename DataType<_Tp>::work_type, cn> work_type;
-    typedef _Tp                                        channel_type;
-    typedef value_type                                 vec_type;
-
-    enum { generic_type = 0,
-           channels     = cn,
-           fmt          = DataType<channel_type>::fmt + ((channels - 1) << 8),
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-           depth        = DataType<channel_type>::depth,
-           type         = CV_MAKETYPE(depth, channels),
-#endif
-           _dummy_enum_finalizer = 0
-         };
-};
-
-namespace traits {
-template<typename _Tp, int cn>
-struct Depth< Vec<_Tp, cn> > { enum { value = Depth<_Tp>::value }; };
-template<typename _Tp, int cn>
-struct Type< Vec<_Tp, cn> > { enum { value = CV_MAKETYPE(Depth<_Tp>::value, cn) }; };
-} // namespace
-
-
-/** @brief  Comma-separated Vec Initializer
-*/
-template<typename _Tp, int m> class VecCommaInitializer : public MatxCommaInitializer<_Tp, m, 1>
-{
-public:
-    VecCommaInitializer(Vec<_Tp, m>* _vec);
-    template<typename T2> VecCommaInitializer<_Tp, m>& operator , (T2 val);
-    Vec<_Tp, m> operator *() const;
-};
-
-template<typename _Tp, int cn> static Vec<_Tp, cn> normalize(const Vec<_Tp, cn>& v);
-
-//! @} core_basic
-
-//! @cond IGNORED
-
-///////////////////////////////////// helper classes /////////////////////////////////////
-namespace internal
-{
-
-template<typename _Tp, int m> struct Matx_DetOp
-{
-    double operator ()(const Matx<_Tp, m, m>& a) const
-    {
-        Matx<_Tp, m, m> temp = a;
-        double p = LU(temp.val, m*sizeof(_Tp), m, 0, 0, 0);
-        if( p == 0 )
-            return p;
-        for( int i = 0; i < m; i++ )
-            p *= temp(i, i);
-        return p;
-    }
-};
-
-template<typename _Tp> struct Matx_DetOp<_Tp, 1>
-{
-    double operator ()(const Matx<_Tp, 1, 1>& a) const
-    {
-        return a(0,0);
-    }
-};
-
-template<typename _Tp> struct Matx_DetOp<_Tp, 2>
-{
-    double operator ()(const Matx<_Tp, 2, 2>& a) const
-    {
-        return a(0,0)*a(1,1) - a(0,1)*a(1,0);
-    }
-};
-
-template<typename _Tp> struct Matx_DetOp<_Tp, 3>
-{
-    double operator ()(const Matx<_Tp, 3, 3>& a) const
-    {
-        return a(0,0)*(a(1,1)*a(2,2) - a(2,1)*a(1,2)) -
-            a(0,1)*(a(1,0)*a(2,2) - a(2,0)*a(1,2)) +
-            a(0,2)*(a(1,0)*a(2,1) - a(2,0)*a(1,1));
-    }
-};
-
-template<typename _Tp> Vec<_Tp, 2> inline conjugate(const Vec<_Tp, 2>& v)
-{
-    return Vec<_Tp, 2>(v[0], -v[1]);
-}
-
-template<typename _Tp> Vec<_Tp, 4> inline conjugate(const Vec<_Tp, 4>& v)
-{
-    return Vec<_Tp, 4>(v[0], -v[1], -v[2], -v[3]);
-}
-
-} // internal
-
-
-
-////////////////////////////////// Matx Implementation ///////////////////////////////////
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n>::Matx()
-{
-    for(int i = 0; i < channels; i++) val[i] = _Tp(0);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n>::Matx(_Tp v0)
-{
-    val[0] = v0;
-    for(int i = 1; i < channels; i++) val[i] = _Tp(0);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n>::Matx(_Tp v0, _Tp v1)
-{
-    CV_StaticAssert(channels >= 2, "Matx should have at least 2 elements.");
-    val[0] = v0; val[1] = v1;
-    for(int i = 2; i < channels; i++) val[i] = _Tp(0);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n>::Matx(_Tp v0, _Tp v1, _Tp v2)
-{
-    CV_StaticAssert(channels >= 3, "Matx should have at least 3 elements.");
-    val[0] = v0; val[1] = v1; val[2] = v2;
-    for(int i = 3; i < channels; i++) val[i] = _Tp(0);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n>::Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3)
-{
-    CV_StaticAssert(channels >= 4, "Matx should have at least 4 elements.");
-    val[0] = v0; val[1] = v1; val[2] = v2; val[3] = v3;
-    for(int i = 4; i < channels; i++) val[i] = _Tp(0);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n>::Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4)
-{
-    CV_StaticAssert(channels >= 5, "Matx should have at least 5 elements.");
-    val[0] = v0; val[1] = v1; val[2] = v2; val[3] = v3; val[4] = v4;
-    for(int i = 5; i < channels; i++) val[i] = _Tp(0);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n>::Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5)
-{
-    CV_StaticAssert(channels >= 6, "Matx should have at least 6 elements.");
-    val[0] = v0; val[1] = v1; val[2] = v2; val[3] = v3;
-    val[4] = v4; val[5] = v5;
-    for(int i = 6; i < channels; i++) val[i] = _Tp(0);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n>::Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6)
-{
-    CV_StaticAssert(channels >= 7, "Matx should have at least 7 elements.");
-    val[0] = v0; val[1] = v1; val[2] = v2; val[3] = v3;
-    val[4] = v4; val[5] = v5; val[6] = v6;
-    for(int i = 7; i < channels; i++) val[i] = _Tp(0);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n>::Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7)
-{
-    CV_StaticAssert(channels >= 8, "Matx should have at least 8 elements.");
-    val[0] = v0; val[1] = v1; val[2] = v2; val[3] = v3;
-    val[4] = v4; val[5] = v5; val[6] = v6; val[7] = v7;
-    for(int i = 8; i < channels; i++) val[i] = _Tp(0);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n>::Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7, _Tp v8)
-{
-    CV_StaticAssert(channels >= 9, "Matx should have at least 9 elements.");
-    val[0] = v0; val[1] = v1; val[2] = v2; val[3] = v3;
-    val[4] = v4; val[5] = v5; val[6] = v6; val[7] = v7;
-    val[8] = v8;
-    for(int i = 9; i < channels; i++) val[i] = _Tp(0);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n>::Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7, _Tp v8, _Tp v9)
-{
-    CV_StaticAssert(channels >= 10, "Matx should have at least 10 elements.");
-    val[0] = v0; val[1] = v1; val[2] = v2; val[3] = v3;
-    val[4] = v4; val[5] = v5; val[6] = v6; val[7] = v7;
-    val[8] = v8; val[9] = v9;
-    for(int i = 10; i < channels; i++) val[i] = _Tp(0);
-}
-
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp,m,n>::Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7, _Tp v8, _Tp v9, _Tp v10, _Tp v11)
-{
-    CV_StaticAssert(channels >= 12, "Matx should have at least 12 elements.");
-    val[0] = v0; val[1] = v1; val[2] = v2; val[3] = v3;
-    val[4] = v4; val[5] = v5; val[6] = v6; val[7] = v7;
-    val[8] = v8; val[9] = v9; val[10] = v10; val[11] = v11;
-    for(int i = 12; i < channels; i++) val[i] = _Tp(0);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp,m,n>::Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7, _Tp v8, _Tp v9, _Tp v10, _Tp v11, _Tp v12, _Tp v13)
-{
-    CV_StaticAssert(channels >= 14, "Matx should have at least 14 elements.");
-    val[0] = v0; val[1] = v1; val[2] = v2; val[3] = v3;
-    val[4] = v4; val[5] = v5; val[6] = v6; val[7] = v7;
-    val[8] = v8; val[9] = v9; val[10] = v10; val[11] = v11;
-    val[12] = v12; val[13] = v13;
-    for (int i = 14; i < channels; i++) val[i] = _Tp(0);
-}
-
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp,m,n>::Matx(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7, _Tp v8, _Tp v9, _Tp v10, _Tp v11, _Tp v12, _Tp v13, _Tp v14, _Tp v15)
-{
-    CV_StaticAssert(channels >= 16, "Matx should have at least 16 elements.");
-    val[0] = v0; val[1] = v1; val[2] = v2; val[3] = v3;
-    val[4] = v4; val[5] = v5; val[6] = v6; val[7] = v7;
-    val[8] = v8; val[9] = v9; val[10] = v10; val[11] = v11;
-    val[12] = v12; val[13] = v13; val[14] = v14; val[15] = v15;
-    for(int i = 16; i < channels; i++) val[i] = _Tp(0);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n>::Matx(const _Tp* values)
-{
-    for( int i = 0; i < channels; i++ ) val[i] = values[i];
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n>::Matx(std::initializer_list<_Tp> list)
-{
-    CV_DbgAssert(list.size() == channels);
-    int i = 0;
-    for(const auto& elem : list)
-    {
-        val[i++] = elem;
-    }
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n> Matx<_Tp, m, n>::all(_Tp alpha)
-{
-    Matx<_Tp, m, n> M;
-    for( int i = 0; i < m*n; i++ ) M.val[i] = alpha;
-    return M;
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp,m,n> Matx<_Tp,m,n>::zeros()
-{
-    return all(0);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp,m,n> Matx<_Tp,m,n>::ones()
-{
-    return all(1);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp,m,n> Matx<_Tp,m,n>::eye()
-{
-    Matx<_Tp,m,n> M;
-    for(int i = 0; i < shortdim; i++)
-        M(i,i) = 1;
-    return M;
-}
-
-template<typename _Tp, int m, int n> inline
-_Tp Matx<_Tp, m, n>::dot(const Matx<_Tp, m, n>& M) const
-{
-    _Tp s = 0;
-    for( int i = 0; i < channels; i++ ) s += val[i]*M.val[i];
-    return s;
-}
-
-template<typename _Tp, int m, int n> inline
-double Matx<_Tp, m, n>::ddot(const Matx<_Tp, m, n>& M) const
-{
-    double s = 0;
-    for( int i = 0; i < channels; i++ ) s += (double)val[i]*M.val[i];
-    return s;
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp,m,n> Matx<_Tp,m,n>::diag(const typename Matx<_Tp,m,n>::diag_type& d)
-{
-    Matx<_Tp,m,n> M;
-    for(int i = 0; i < shortdim; i++)
-        M(i,i) = d(i, 0);
-    return M;
-}
-
-template<typename _Tp, int m, int n> template<typename T2>
-inline Matx<_Tp, m, n>::operator Matx<T2, m, n>() const
-{
-    Matx<T2, m, n> M;
-    for( int i = 0; i < m*n; i++ ) M.val[i] = saturate_cast<T2>(val[i]);
-    return M;
-}
-
-template<typename _Tp, int m, int n> template<int m1, int n1> inline
-Matx<_Tp, m1, n1> Matx<_Tp, m, n>::reshape() const
-{
-    CV_StaticAssert(m1*n1 == m*n, "Input and destnarion matrices must have the same number of elements");
-    return (const Matx<_Tp, m1, n1>&)*this;
-}
-
-template<typename _Tp, int m, int n>
-template<int m1, int n1> inline
-Matx<_Tp, m1, n1> Matx<_Tp, m, n>::get_minor(int base_row, int base_col) const
-{
-    CV_DbgAssert(0 <= base_row && base_row+m1 <= m && 0 <= base_col && base_col+n1 <= n);
-    Matx<_Tp, m1, n1> s;
-    for( int di = 0; di < m1; di++ )
-        for( int dj = 0; dj < n1; dj++ )
-            s(di, dj) = (*this)(base_row+di, base_col+dj);
-    return s;
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, 1, n> Matx<_Tp, m, n>::row(int i) const
-{
-    CV_DbgAssert((unsigned)i < (unsigned)m);
-    return Matx<_Tp, 1, n>(&val[i*n]);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, 1> Matx<_Tp, m, n>::col(int j) const
-{
-    CV_DbgAssert((unsigned)j < (unsigned)n);
-    Matx<_Tp, m, 1> v;
-    for( int i = 0; i < m; i++ )
-        v.val[i] = val[i*n + j];
-    return v;
-}
-
-template<typename _Tp, int m, int n> inline
-typename Matx<_Tp, m, n>::diag_type Matx<_Tp, m, n>::diag() const
-{
-    diag_type d;
-    for( int i = 0; i < shortdim; i++ )
-        d.val[i] = val[i*n + i];
-    return d;
-}
-
-template<typename _Tp, int m, int n> inline
-const _Tp& Matx<_Tp, m, n>::operator()(int row_idx, int col_idx) const
-{
-    CV_DbgAssert( (unsigned)row_idx < (unsigned)m && (unsigned)col_idx < (unsigned)n );
-    return this->val[row_idx*n + col_idx];
-}
-
-template<typename _Tp, int m, int n> inline
-_Tp& Matx<_Tp, m, n>::operator ()(int row_idx, int col_idx)
-{
-    CV_DbgAssert( (unsigned)row_idx < (unsigned)m && (unsigned)col_idx < (unsigned)n );
-    return val[row_idx*n + col_idx];
-}
-
-template<typename _Tp, int m, int n> inline
-const _Tp& Matx<_Tp, m, n>::operator ()(int i) const
-{
-    CV_StaticAssert(m == 1 || n == 1, "Single index indexation requires matrix to be a column or a row");
-    CV_DbgAssert( (unsigned)i < (unsigned)(m+n-1) );
-    return val[i];
-}
-
-template<typename _Tp, int m, int n> inline
-_Tp& Matx<_Tp, m, n>::operator ()(int i)
-{
-    CV_StaticAssert(m == 1 || n == 1, "Single index indexation requires matrix to be a column or a row");
-    CV_DbgAssert( (unsigned)i < (unsigned)(m+n-1) );
-    return val[i];
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp,m,n>::Matx(const Matx<_Tp, m, n>& a, const Matx<_Tp, m, n>& b, Matx_AddOp)
-{
-    for( int i = 0; i < channels; i++ )
-        val[i] = saturate_cast<_Tp>(a.val[i] + b.val[i]);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp,m,n>::Matx(const Matx<_Tp, m, n>& a, const Matx<_Tp, m, n>& b, Matx_SubOp)
-{
-    for( int i = 0; i < channels; i++ )
-        val[i] = saturate_cast<_Tp>(a.val[i] - b.val[i]);
-}
-
-template<typename _Tp, int m, int n> template<typename _T2> inline
-Matx<_Tp,m,n>::Matx(const Matx<_Tp, m, n>& a, _T2 alpha, Matx_ScaleOp)
-{
-    for( int i = 0; i < channels; i++ )
-        val[i] = saturate_cast<_Tp>(a.val[i] * alpha);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp,m,n>::Matx(const Matx<_Tp, m, n>& a, const Matx<_Tp, m, n>& b, Matx_MulOp)
-{
-    for( int i = 0; i < channels; i++ )
-        val[i] = saturate_cast<_Tp>(a.val[i] * b.val[i]);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp,m,n>::Matx(const Matx<_Tp, m, n>& a, const Matx<_Tp, m, n>& b, Matx_DivOp)
-{
-    for( int i = 0; i < channels; i++ )
-        val[i] = saturate_cast<_Tp>(a.val[i] / b.val[i]);
-}
-
-template<typename _Tp, int m, int n> template<int l> inline
-Matx<_Tp,m,n>::Matx(const Matx<_Tp, m, l>& a, const Matx<_Tp, l, n>& b, Matx_MatMulOp)
-{
-    for( int i = 0; i < m; i++ )
-        for( int j = 0; j < n; j++ )
-        {
-            _Tp s = 0;
-            for( int k = 0; k < l; k++ )
-                s += a(i, k) * b(k, j);
-            val[i*n + j] = s;
-        }
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp,m,n>::Matx(const Matx<_Tp, n, m>& a, Matx_TOp)
-{
-    for( int i = 0; i < m; i++ )
-        for( int j = 0; j < n; j++ )
-            val[i*n + j] = a(j, i);
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n> Matx<_Tp, m, n>::mul(const Matx<_Tp, m, n>& a) const
-{
-    return Matx<_Tp, m, n>(*this, a, Matx_MulOp());
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n> Matx<_Tp, m, n>::div(const Matx<_Tp, m, n>& a) const
-{
-    return Matx<_Tp, m, n>(*this, a, Matx_DivOp());
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, n, m> Matx<_Tp, m, n>::t() const
-{
-    return Matx<_Tp, n, m>(*this, Matx_TOp());
-}
-
-template<typename _Tp, int m, int n> inline
-Vec<_Tp, n> Matx<_Tp, m, n>::solve(const Vec<_Tp, m>& rhs, int method) const
-{
-    Matx<_Tp, n, 1> x = solve((const Matx<_Tp, m, 1>&)(rhs), method);
-    return (Vec<_Tp, n>&)(x);
-}
-
-template<typename _Tp, int m> static inline
-double determinant(const Matx<_Tp, m, m>& a)
-{
-    return cv::internal::Matx_DetOp<_Tp, m>()(a);
-}
-
-template<typename _Tp, int m, int n> static inline
-double trace(const Matx<_Tp, m, n>& a)
-{
-    _Tp s = 0;
-    for( int i = 0; i < std::min(m, n); i++ )
-        s += a(i,i);
-    return s;
-}
-
-template<typename _Tp, int m, int n> static inline
-double norm(const Matx<_Tp, m, n>& M)
-{
-    return std::sqrt(normL2Sqr<_Tp, double>(M.val, m*n));
-}
-
-template<typename _Tp, int m, int n> static inline
-double norm(const Matx<_Tp, m, n>& M, int normType)
-{
-    switch(normType) {
-    case NORM_INF:
-        return (double)normInf<_Tp, typename DataType<_Tp>::work_type>(M.val, m*n);
-    case NORM_L1:
-        return (double)normL1<_Tp, typename DataType<_Tp>::work_type>(M.val, m*n);
-    case NORM_L2SQR:
-        return (double)normL2Sqr<_Tp, typename DataType<_Tp>::work_type>(M.val, m*n);
-    default:
-    case NORM_L2:
-        return std::sqrt((double)normL2Sqr<_Tp, typename DataType<_Tp>::work_type>(M.val, m*n));
-    }
-}
-
-
-
-//////////////////////////////// matx comma initializer //////////////////////////////////
-
-template<typename _Tp, typename _T2, int m, int n> static inline
-MatxCommaInitializer<_Tp, m, n> operator << (const Matx<_Tp, m, n>& mtx, _T2 val)
-{
-    MatxCommaInitializer<_Tp, m, n> commaInitializer((Matx<_Tp, m, n>*)&mtx);
-    return (commaInitializer, val);
-}
-
-template<typename _Tp, int m, int n> inline
-MatxCommaInitializer<_Tp, m, n>::MatxCommaInitializer(Matx<_Tp, m, n>* _mtx)
-    : dst(_mtx), idx(0)
-{}
-
-template<typename _Tp, int m, int n> template<typename _T2> inline
-MatxCommaInitializer<_Tp, m, n>& MatxCommaInitializer<_Tp, m, n>::operator , (_T2 value)
-{
-    CV_DbgAssert( idx < m*n );
-    dst->val[idx++] = saturate_cast<_Tp>(value);
-    return *this;
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, m, n> MatxCommaInitializer<_Tp, m, n>::operator *() const
-{
-    CV_DbgAssert( idx == n*m );
-    return *dst;
-}
-
-
-
-/////////////////////////////////// Vec Implementation ///////////////////////////////////
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec() {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(_Tp v0)
-    : Matx<_Tp, cn, 1>(v0) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(_Tp v0, _Tp v1)
-    : Matx<_Tp, cn, 1>(v0, v1) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(_Tp v0, _Tp v1, _Tp v2)
-    : Matx<_Tp, cn, 1>(v0, v1, v2) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3)
-    : Matx<_Tp, cn, 1>(v0, v1, v2, v3) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4)
-    : Matx<_Tp, cn, 1>(v0, v1, v2, v3, v4) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5)
-    : Matx<_Tp, cn, 1>(v0, v1, v2, v3, v4, v5) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6)
-    : Matx<_Tp, cn, 1>(v0, v1, v2, v3, v4, v5, v6) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7)
-    : Matx<_Tp, cn, 1>(v0, v1, v2, v3, v4, v5, v6, v7) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7, _Tp v8)
-    : Matx<_Tp, cn, 1>(v0, v1, v2, v3, v4, v5, v6, v7, v8) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7, _Tp v8, _Tp v9)
-    : Matx<_Tp, cn, 1>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(_Tp v0, _Tp v1, _Tp v2, _Tp v3, _Tp v4, _Tp v5, _Tp v6, _Tp v7, _Tp v8, _Tp v9, _Tp v10, _Tp v11, _Tp v12, _Tp v13)
-    : Matx<_Tp, cn, 1>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(const _Tp* values)
-    : Matx<_Tp, cn, 1>(values) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(std::initializer_list<_Tp> list)
-    : Matx<_Tp, cn, 1>(list) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(const Vec<_Tp, cn>& m)
-    : Matx<_Tp, cn, 1>(m.val) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(const Matx<_Tp, cn, 1>& a, const Matx<_Tp, cn, 1>& b, Matx_AddOp op)
-    : Matx<_Tp, cn, 1>(a, b, op) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn>::Vec(const Matx<_Tp, cn, 1>& a, const Matx<_Tp, cn, 1>& b, Matx_SubOp op)
-    : Matx<_Tp, cn, 1>(a, b, op) {}
-
-template<typename _Tp, int cn> template<typename _T2> inline
-Vec<_Tp, cn>::Vec(const Matx<_Tp, cn, 1>& a, _T2 alpha, Matx_ScaleOp op)
-    : Matx<_Tp, cn, 1>(a, alpha, op) {}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn> Vec<_Tp, cn>::all(_Tp alpha)
-{
-    Vec v;
-    for( int i = 0; i < cn; i++ ) v.val[i] = alpha;
-    return v;
-}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn> Vec<_Tp, cn>::mul(const Vec<_Tp, cn>& v) const
-{
-    Vec<_Tp, cn> w;
-    for( int i = 0; i < cn; i++ ) w.val[i] = saturate_cast<_Tp>(this->val[i]*v.val[i]);
-    return w;
-}
-
-template<> inline
-Vec<float, 2> Vec<float, 2>::conj() const
-{
-    return cv::internal::conjugate(*this);
-}
-
-template<> inline
-Vec<double, 2> Vec<double, 2>::conj() const
-{
-    return cv::internal::conjugate(*this);
-}
-
-template<> inline
-Vec<float, 4> Vec<float, 4>::conj() const
-{
-    return cv::internal::conjugate(*this);
-}
-
-template<> inline
-Vec<double, 4> Vec<double, 4>::conj() const
-{
-    return cv::internal::conjugate(*this);
-}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn> Vec<_Tp, cn>::cross(const Vec<_Tp, cn>&) const
-{
-    CV_StaticAssert(cn == 3, "for arbitrary-size vector there is no cross-product defined");
-    return Vec<_Tp, cn>();
-}
-
-template<> inline
-Vec<float, 3> Vec<float, 3>::cross(const Vec<float, 3>& v) const
-{
-    return Vec<float,3>(this->val[1]*v.val[2] - this->val[2]*v.val[1],
-                     this->val[2]*v.val[0] - this->val[0]*v.val[2],
-                     this->val[0]*v.val[1] - this->val[1]*v.val[0]);
-}
-
-template<> inline
-Vec<double, 3> Vec<double, 3>::cross(const Vec<double, 3>& v) const
-{
-    return Vec<double,3>(this->val[1]*v.val[2] - this->val[2]*v.val[1],
-                     this->val[2]*v.val[0] - this->val[0]*v.val[2],
-                     this->val[0]*v.val[1] - this->val[1]*v.val[0]);
-}
-
-template<typename _Tp, int cn> template<typename T2> inline
-Vec<_Tp, cn>::operator Vec<T2, cn>() const
-{
-    Vec<T2, cn> v;
-    for( int i = 0; i < cn; i++ ) v.val[i] = saturate_cast<T2>(this->val[i]);
-    return v;
-}
-
-template<typename _Tp, int cn> inline
-const _Tp& Vec<_Tp, cn>::operator [](int i) const
-{
-    CV_DbgAssert( (unsigned)i < (unsigned)cn );
-    return this->val[i];
-}
-
-template<typename _Tp, int cn> inline
-_Tp& Vec<_Tp, cn>::operator [](int i)
-{
-    CV_DbgAssert( (unsigned)i < (unsigned)cn );
-    return this->val[i];
-}
-
-template<typename _Tp, int cn> inline
-const _Tp& Vec<_Tp, cn>::operator ()(int i) const
-{
-    CV_DbgAssert( (unsigned)i < (unsigned)cn );
-    return this->val[i];
-}
-
-template<typename _Tp, int cn> inline
-_Tp& Vec<_Tp, cn>::operator ()(int i)
-{
-    CV_DbgAssert( (unsigned)i < (unsigned)cn );
-    return this->val[i];
-}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn> normalize(const Vec<_Tp, cn>& v)
-{
-    double nv = norm(v);
-    return v * (nv ? 1./nv : 0.);
-}
-
-
-
-//////////////////////////////// vec comma initializer //////////////////////////////////
-
-
-template<typename _Tp, typename _T2, int cn> static inline
-VecCommaInitializer<_Tp, cn> operator << (const Vec<_Tp, cn>& vec, _T2 val)
-{
-    VecCommaInitializer<_Tp, cn> commaInitializer((Vec<_Tp, cn>*)&vec);
-    return (commaInitializer, val);
-}
-
-template<typename _Tp, int cn> inline
-VecCommaInitializer<_Tp, cn>::VecCommaInitializer(Vec<_Tp, cn>* _vec)
-    : MatxCommaInitializer<_Tp, cn, 1>(_vec)
-{}
-
-template<typename _Tp, int cn> template<typename _T2> inline
-VecCommaInitializer<_Tp, cn>& VecCommaInitializer<_Tp, cn>::operator , (_T2 value)
-{
-    CV_DbgAssert( this->idx < cn );
-    this->dst->val[this->idx++] = saturate_cast<_Tp>(value);
-    return *this;
-}
-
-template<typename _Tp, int cn> inline
-Vec<_Tp, cn> VecCommaInitializer<_Tp, cn>::operator *() const
-{
-    CV_DbgAssert( this->idx == cn );
-    return *this->dst;
-}
-
-//! @endcond
-
-///////////////////////////// Matx out-of-class operators ////////////////////////////////
-
-//! @relates cv::Matx
-//! @{
-
-template<typename _Tp1, typename _Tp2, int m, int n> static inline
-Matx<_Tp1, m, n>& operator += (Matx<_Tp1, m, n>& a, const Matx<_Tp2, m, n>& b)
-{
-    for( int i = 0; i < m*n; i++ )
-        a.val[i] = saturate_cast<_Tp1>(a.val[i] + b.val[i]);
-    return a;
-}
-
-template<typename _Tp1, typename _Tp2, int m, int n> static inline
-Matx<_Tp1, m, n>& operator -= (Matx<_Tp1, m, n>& a, const Matx<_Tp2, m, n>& b)
-{
-    for( int i = 0; i < m*n; i++ )
-        a.val[i] = saturate_cast<_Tp1>(a.val[i] - b.val[i]);
-    return a;
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n> operator + (const Matx<_Tp, m, n>& a, const Matx<_Tp, m, n>& b)
-{
-    return Matx<_Tp, m, n>(a, b, Matx_AddOp());
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n> operator - (const Matx<_Tp, m, n>& a, const Matx<_Tp, m, n>& b)
-{
-    return Matx<_Tp, m, n>(a, b, Matx_SubOp());
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n>& operator *= (Matx<_Tp, m, n>& a, int alpha)
-{
-    for( int i = 0; i < m*n; i++ )
-        a.val[i] = saturate_cast<_Tp>(a.val[i] * alpha);
-    return a;
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n>& operator *= (Matx<_Tp, m, n>& a, float alpha)
-{
-    for( int i = 0; i < m*n; i++ )
-        a.val[i] = saturate_cast<_Tp>(a.val[i] * alpha);
-    return a;
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n>& operator *= (Matx<_Tp, m, n>& a, double alpha)
-{
-    for( int i = 0; i < m*n; i++ )
-        a.val[i] = saturate_cast<_Tp>(a.val[i] * alpha);
-    return a;
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n> operator * (const Matx<_Tp, m, n>& a, int alpha)
-{
-    return Matx<_Tp, m, n>(a, alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n> operator * (const Matx<_Tp, m, n>& a, float alpha)
-{
-    return Matx<_Tp, m, n>(a, alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n> operator * (const Matx<_Tp, m, n>& a, double alpha)
-{
-    return Matx<_Tp, m, n>(a, alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n> operator * (int alpha, const Matx<_Tp, m, n>& a)
-{
-    return Matx<_Tp, m, n>(a, alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n> operator * (float alpha, const Matx<_Tp, m, n>& a)
-{
-    return Matx<_Tp, m, n>(a, alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n> operator * (double alpha, const Matx<_Tp, m, n>& a)
-{
-    return Matx<_Tp, m, n>(a, alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n>& operator /= (Matx<_Tp, m, n>& a, float alpha)
-{
-    for( int i = 0; i < m*n; i++ )
-        a.val[i] = a.val[i] / alpha;
-    return a;
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n>& operator /= (Matx<_Tp, m, n>& a, double alpha)
-{
-    for( int i = 0; i < m*n; i++ )
-        a.val[i] = a.val[i] / alpha;
-    return a;
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n> operator / (const Matx<_Tp, m, n>& a, float alpha)
-{
-    return Matx<_Tp, m, n>(a, 1.f/alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n> operator / (const Matx<_Tp, m, n>& a, double alpha)
-{
-    return Matx<_Tp, m, n>(a, 1./alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int m, int n> static inline
-Matx<_Tp, m, n> operator - (const Matx<_Tp, m, n>& a)
-{
-    return Matx<_Tp, m, n>(a, -1, Matx_ScaleOp());
-}
-
-template<typename _Tp, int m, int n, int l> static inline
-Matx<_Tp, m, n> operator * (const Matx<_Tp, m, l>& a, const Matx<_Tp, l, n>& b)
-{
-    return Matx<_Tp, m, n>(a, b, Matx_MatMulOp());
-}
-
-template<typename _Tp, int m, int n> static inline
-Vec<_Tp, m> operator * (const Matx<_Tp, m, n>& a, const Vec<_Tp, n>& b)
-{
-    Matx<_Tp, m, 1> c(a, b, Matx_MatMulOp());
-    return (const Vec<_Tp, m>&)(c);
-}
-
-template<typename _Tp, int m, int n> static inline
-bool operator == (const Matx<_Tp, m, n>& a, const Matx<_Tp, m, n>& b)
-{
-    for( int i = 0; i < m*n; i++ )
-        if( a.val[i] != b.val[i] ) return false;
-    return true;
-}
-
-template<typename _Tp, int m, int n> static inline
-bool operator != (const Matx<_Tp, m, n>& a, const Matx<_Tp, m, n>& b)
-{
-    return !(a == b);
-}
-
-//! @}
-
-////////////////////////////// Vec out-of-class operators ////////////////////////////////
-
-//! @relates cv::Vec
-//! @{
-
-template<typename _Tp1, typename _Tp2, int cn> static inline
-Vec<_Tp1, cn>& operator += (Vec<_Tp1, cn>& a, const Vec<_Tp2, cn>& b)
-{
-    for( int i = 0; i < cn; i++ )
-        a.val[i] = saturate_cast<_Tp1>(a.val[i] + b.val[i]);
-    return a;
-}
-
-template<typename _Tp1, typename _Tp2, int cn> static inline
-Vec<_Tp1, cn>& operator -= (Vec<_Tp1, cn>& a, const Vec<_Tp2, cn>& b)
-{
-    for( int i = 0; i < cn; i++ )
-        a.val[i] = saturate_cast<_Tp1>(a.val[i] - b.val[i]);
-    return a;
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn> operator + (const Vec<_Tp, cn>& a, const Vec<_Tp, cn>& b)
-{
-    return Vec<_Tp, cn>(a, b, Matx_AddOp());
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn> operator - (const Vec<_Tp, cn>& a, const Vec<_Tp, cn>& b)
-{
-    return Vec<_Tp, cn>(a, b, Matx_SubOp());
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn>& operator *= (Vec<_Tp, cn>& a, int alpha)
-{
-    for( int i = 0; i < cn; i++ )
-        a[i] = saturate_cast<_Tp>(a[i]*alpha);
-    return a;
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn>& operator *= (Vec<_Tp, cn>& a, float alpha)
-{
-    for( int i = 0; i < cn; i++ )
-        a[i] = saturate_cast<_Tp>(a[i]*alpha);
-    return a;
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn>& operator *= (Vec<_Tp, cn>& a, double alpha)
-{
-    for( int i = 0; i < cn; i++ )
-        a[i] = saturate_cast<_Tp>(a[i]*alpha);
-    return a;
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn>& operator /= (Vec<_Tp, cn>& a, int alpha)
-{
-    double ialpha = 1./alpha;
-    for( int i = 0; i < cn; i++ )
-        a[i] = saturate_cast<_Tp>(a[i]*ialpha);
-    return a;
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn>& operator /= (Vec<_Tp, cn>& a, float alpha)
-{
-    float ialpha = 1.f/alpha;
-    for( int i = 0; i < cn; i++ )
-        a[i] = saturate_cast<_Tp>(a[i]*ialpha);
-    return a;
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn>& operator /= (Vec<_Tp, cn>& a, double alpha)
-{
-    double ialpha = 1./alpha;
-    for( int i = 0; i < cn; i++ )
-        a[i] = saturate_cast<_Tp>(a[i]*ialpha);
-    return a;
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn> operator * (const Vec<_Tp, cn>& a, int alpha)
-{
-    return Vec<_Tp, cn>(a, alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn> operator * (int alpha, const Vec<_Tp, cn>& a)
-{
-    return Vec<_Tp, cn>(a, alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn> operator * (const Vec<_Tp, cn>& a, float alpha)
-{
-    return Vec<_Tp, cn>(a, alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn> operator * (float alpha, const Vec<_Tp, cn>& a)
-{
-    return Vec<_Tp, cn>(a, alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn> operator * (const Vec<_Tp, cn>& a, double alpha)
-{
-    return Vec<_Tp, cn>(a, alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn> operator * (double alpha, const Vec<_Tp, cn>& a)
-{
-    return Vec<_Tp, cn>(a, alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn> operator / (const Vec<_Tp, cn>& a, int alpha)
-{
-    return Vec<_Tp, cn>(a, 1./alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn> operator / (const Vec<_Tp, cn>& a, float alpha)
-{
-    return Vec<_Tp, cn>(a, 1.f/alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn> operator / (const Vec<_Tp, cn>& a, double alpha)
-{
-    return Vec<_Tp, cn>(a, 1./alpha, Matx_ScaleOp());
-}
-
-template<typename _Tp, int cn> static inline
-Vec<_Tp, cn> operator - (const Vec<_Tp, cn>& a)
-{
-    Vec<_Tp,cn> t;
-    for( int i = 0; i < cn; i++ ) t.val[i] = saturate_cast<_Tp>(-a.val[i]);
-    return t;
-}
-
-template<typename _Tp> inline Vec<_Tp, 4> operator * (const Vec<_Tp, 4>& v1, const Vec<_Tp, 4>& v2)
-{
-    return Vec<_Tp, 4>(saturate_cast<_Tp>(v1[0]*v2[0] - v1[1]*v2[1] - v1[2]*v2[2] - v1[3]*v2[3]),
-                       saturate_cast<_Tp>(v1[0]*v2[1] + v1[1]*v2[0] + v1[2]*v2[3] - v1[3]*v2[2]),
-                       saturate_cast<_Tp>(v1[0]*v2[2] - v1[1]*v2[3] + v1[2]*v2[0] + v1[3]*v2[1]),
-                       saturate_cast<_Tp>(v1[0]*v2[3] + v1[1]*v2[2] - v1[2]*v2[1] + v1[3]*v2[0]));
-}
-
-template<typename _Tp> inline Vec<_Tp, 4>& operator *= (Vec<_Tp, 4>& v1, const Vec<_Tp, 4>& v2)
-{
-    v1 = v1 * v2;
-    return v1;
-}
-
-//! @}
-
-} // cv
-
-#endif // OPENCV_CORE_MATX_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/neon_utils.hpp b/projects/llm_framework/include/opencv4/opencv2/core/neon_utils.hpp
deleted file mode 100644
index 573ba99e..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/neon_utils.hpp
+++ /dev/null
@@ -1,128 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2015, Itseez Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_HAL_NEON_UTILS_HPP
-#define OPENCV_HAL_NEON_UTILS_HPP
-
-#include "opencv2/core/cvdef.h"
-
-//! @addtogroup core_utils_neon
-//! @{
-
-#if CV_NEON
-
-inline int32x2_t cv_vrnd_s32_f32(float32x2_t v)
-{
-    static int32x2_t v_sign = vdup_n_s32(1 << 31),
-        v_05 = vreinterpret_s32_f32(vdup_n_f32(0.5f));
-
-    int32x2_t v_addition = vorr_s32(v_05, vand_s32(v_sign, vreinterpret_s32_f32(v)));
-    return vcvt_s32_f32(vadd_f32(v, vreinterpret_f32_s32(v_addition)));
-}
-
-inline int32x4_t cv_vrndq_s32_f32(float32x4_t v)
-{
-    static int32x4_t v_sign = vdupq_n_s32(1 << 31),
-        v_05 = vreinterpretq_s32_f32(vdupq_n_f32(0.5f));
-
-    int32x4_t v_addition = vorrq_s32(v_05, vandq_s32(v_sign, vreinterpretq_s32_f32(v)));
-    return vcvtq_s32_f32(vaddq_f32(v, vreinterpretq_f32_s32(v_addition)));
-}
-
-inline uint32x2_t cv_vrnd_u32_f32(float32x2_t v)
-{
-    static float32x2_t v_05 = vdup_n_f32(0.5f);
-    return vcvt_u32_f32(vadd_f32(v, v_05));
-}
-
-inline uint32x4_t cv_vrndq_u32_f32(float32x4_t v)
-{
-    static float32x4_t v_05 = vdupq_n_f32(0.5f);
-    return vcvtq_u32_f32(vaddq_f32(v, v_05));
-}
-
-inline float32x4_t cv_vrecpq_f32(float32x4_t val)
-{
-    float32x4_t reciprocal = vrecpeq_f32(val);
-    reciprocal = vmulq_f32(vrecpsq_f32(val, reciprocal), reciprocal);
-    reciprocal = vmulq_f32(vrecpsq_f32(val, reciprocal), reciprocal);
-    return reciprocal;
-}
-
-inline float32x2_t cv_vrecp_f32(float32x2_t val)
-{
-    float32x2_t reciprocal = vrecpe_f32(val);
-    reciprocal = vmul_f32(vrecps_f32(val, reciprocal), reciprocal);
-    reciprocal = vmul_f32(vrecps_f32(val, reciprocal), reciprocal);
-    return reciprocal;
-}
-
-inline float32x4_t cv_vrsqrtq_f32(float32x4_t val)
-{
-    float32x4_t e = vrsqrteq_f32(val);
-    e = vmulq_f32(vrsqrtsq_f32(vmulq_f32(e, e), val), e);
-    e = vmulq_f32(vrsqrtsq_f32(vmulq_f32(e, e), val), e);
-    return e;
-}
-
-inline float32x2_t cv_vrsqrt_f32(float32x2_t val)
-{
-    float32x2_t e = vrsqrte_f32(val);
-    e = vmul_f32(vrsqrts_f32(vmul_f32(e, e), val), e);
-    e = vmul_f32(vrsqrts_f32(vmul_f32(e, e), val), e);
-    return e;
-}
-
-inline float32x4_t cv_vsqrtq_f32(float32x4_t val)
-{
-    return cv_vrecpq_f32(cv_vrsqrtq_f32(val));
-}
-
-inline float32x2_t cv_vsqrt_f32(float32x2_t val)
-{
-    return cv_vrecp_f32(cv_vrsqrt_f32(val));
-}
-
-#endif
-
-//! @}
-
-#endif // OPENCV_HAL_NEON_UTILS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/ocl.hpp b/projects/llm_framework/include/opencv4/opencv2/core/ocl.hpp
deleted file mode 100644
index 4503fa00..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/ocl.hpp
+++ /dev/null
@@ -1,917 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the OpenCV Foundation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_OPENCL_HPP
-#define OPENCV_OPENCL_HPP
-
-#include "opencv2/core.hpp"
-#include <typeinfo>
-#include <typeindex>
-
-namespace cv { namespace ocl {
-
-//! @addtogroup core_opencl
-//! @{
-
-CV_EXPORTS_W bool haveOpenCL();
-CV_EXPORTS_W bool useOpenCL();
-CV_EXPORTS_W bool haveAmdBlas();
-CV_EXPORTS_W bool haveAmdFft();
-CV_EXPORTS_W void setUseOpenCL(bool flag);
-CV_EXPORTS_W void finish();
-
-CV_EXPORTS bool haveSVM();
-
-class CV_EXPORTS Context;
-class CV_EXPORTS_W_SIMPLE Device;
-class CV_EXPORTS Kernel;
-class CV_EXPORTS Program;
-class CV_EXPORTS ProgramSource;
-class CV_EXPORTS Queue;
-class CV_EXPORTS PlatformInfo;
-class CV_EXPORTS Image2D;
-
-class CV_EXPORTS_W_SIMPLE Device
-{
-public:
-    CV_WRAP Device() CV_NOEXCEPT;
-    explicit Device(void* d);
-    Device(const Device& d);
-    Device& operator = (const Device& d);
-    Device(Device&& d) CV_NOEXCEPT;
-    Device& operator = (Device&& d) CV_NOEXCEPT;
-    CV_WRAP ~Device();
-
-    void set(void* d);
-
-    enum
-    {
-        TYPE_DEFAULT     = (1 << 0),
-        TYPE_CPU         = (1 << 1),
-        TYPE_GPU         = (1 << 2),
-        TYPE_ACCELERATOR = (1 << 3),
-        TYPE_DGPU        = TYPE_GPU + (1 << 16),
-        TYPE_IGPU        = TYPE_GPU + (1 << 17),
-        TYPE_ALL         = 0xFFFFFFFF
-    };
-
-    CV_WRAP String name() const;
-    CV_WRAP String extensions() const;
-    CV_WRAP bool isExtensionSupported(const String& extensionName) const;
-    CV_WRAP String version() const;
-    CV_WRAP String vendorName() const;
-    CV_WRAP String OpenCL_C_Version() const;
-    CV_WRAP String OpenCLVersion() const;
-    CV_WRAP int deviceVersionMajor() const;
-    CV_WRAP int deviceVersionMinor() const;
-    CV_WRAP String driverVersion() const;
-    void* ptr() const;
-
-    CV_WRAP int type() const;
-
-    CV_WRAP int addressBits() const;
-    CV_WRAP bool available() const;
-    CV_WRAP bool compilerAvailable() const;
-    CV_WRAP bool linkerAvailable() const;
-
-    enum
-    {
-        FP_DENORM=(1 << 0),
-        FP_INF_NAN=(1 << 1),
-        FP_ROUND_TO_NEAREST=(1 << 2),
-        FP_ROUND_TO_ZERO=(1 << 3),
-        FP_ROUND_TO_INF=(1 << 4),
-        FP_FMA=(1 << 5),
-        FP_SOFT_FLOAT=(1 << 6),
-        FP_CORRECTLY_ROUNDED_DIVIDE_SQRT=(1 << 7)
-    };
-    CV_WRAP int doubleFPConfig() const;
-    CV_WRAP int singleFPConfig() const;
-    CV_WRAP int halfFPConfig() const;
-
-    CV_WRAP bool endianLittle() const;
-    CV_WRAP bool errorCorrectionSupport() const;
-
-    enum
-    {
-        EXEC_KERNEL=(1 << 0),
-        EXEC_NATIVE_KERNEL=(1 << 1)
-    };
-    CV_WRAP int executionCapabilities() const;
-
-    CV_WRAP size_t globalMemCacheSize() const;
-
-    enum
-    {
-        NO_CACHE=0,
-        READ_ONLY_CACHE=1,
-        READ_WRITE_CACHE=2
-    };
-    CV_WRAP int globalMemCacheType() const;
-    CV_WRAP int globalMemCacheLineSize() const;
-    CV_WRAP size_t globalMemSize() const;
-
-    CV_WRAP size_t localMemSize() const;
-    enum
-    {
-        NO_LOCAL_MEM=0,
-        LOCAL_IS_LOCAL=1,
-        LOCAL_IS_GLOBAL=2
-    };
-    CV_WRAP int localMemType() const;
-    CV_WRAP bool hostUnifiedMemory() const;
-
-    CV_WRAP bool imageSupport() const;
-
-    CV_WRAP bool imageFromBufferSupport() const;
-    uint imagePitchAlignment() const;
-    uint imageBaseAddressAlignment() const;
-
-    /// deprecated, use isExtensionSupported() method (probably with "cl_khr_subgroups" value)
-    CV_WRAP bool intelSubgroupsSupport() const;
-
-    CV_WRAP size_t image2DMaxWidth() const;
-    CV_WRAP size_t image2DMaxHeight() const;
-
-    CV_WRAP size_t image3DMaxWidth() const;
-    CV_WRAP size_t image3DMaxHeight() const;
-    CV_WRAP size_t image3DMaxDepth() const;
-
-    CV_WRAP size_t imageMaxBufferSize() const;
-    CV_WRAP size_t imageMaxArraySize() const;
-
-    enum
-    {
-        UNKNOWN_VENDOR=0,
-        VENDOR_AMD=1,
-        VENDOR_INTEL=2,
-        VENDOR_NVIDIA=3
-    };
-    CV_WRAP int vendorID() const;
-    // FIXIT
-    // dev.isAMD() doesn't work for OpenCL CPU devices from AMD OpenCL platform.
-    // This method should use platform name instead of vendor name.
-    // After fix restore code in arithm.cpp: ocl_compare()
-    CV_WRAP inline bool isAMD() const { return vendorID() == VENDOR_AMD; }
-    CV_WRAP inline bool isIntel() const { return vendorID() == VENDOR_INTEL; }
-    CV_WRAP inline bool isNVidia() const { return vendorID() == VENDOR_NVIDIA; }
-
-    CV_WRAP int maxClockFrequency() const;
-    CV_WRAP int maxComputeUnits() const;
-    CV_WRAP int maxConstantArgs() const;
-    CV_WRAP size_t maxConstantBufferSize() const;
-
-    CV_WRAP size_t maxMemAllocSize() const;
-    CV_WRAP size_t maxParameterSize() const;
-
-    CV_WRAP int maxReadImageArgs() const;
-    CV_WRAP int maxWriteImageArgs() const;
-    CV_WRAP int maxSamplers() const;
-
-    CV_WRAP size_t maxWorkGroupSize() const;
-    CV_WRAP int maxWorkItemDims() const;
-    void maxWorkItemSizes(size_t*) const;
-
-    CV_WRAP int memBaseAddrAlign() const;
-
-    CV_WRAP int nativeVectorWidthChar() const;
-    CV_WRAP int nativeVectorWidthShort() const;
-    CV_WRAP int nativeVectorWidthInt() const;
-    CV_WRAP int nativeVectorWidthLong() const;
-    CV_WRAP int nativeVectorWidthFloat() const;
-    CV_WRAP int nativeVectorWidthDouble() const;
-    CV_WRAP int nativeVectorWidthHalf() const;
-
-    CV_WRAP int preferredVectorWidthChar() const;
-    CV_WRAP int preferredVectorWidthShort() const;
-    CV_WRAP int preferredVectorWidthInt() const;
-    CV_WRAP int preferredVectorWidthLong() const;
-    CV_WRAP int preferredVectorWidthFloat() const;
-    CV_WRAP int preferredVectorWidthDouble() const;
-    CV_WRAP int preferredVectorWidthHalf() const;
-
-    CV_WRAP size_t printfBufferSize() const;
-    CV_WRAP size_t profilingTimerResolution() const;
-
-    CV_WRAP static const Device& getDefault();
-
-    /**
-     * @param d OpenCL handle (cl_device_id). clRetainDevice() is called on success.
-     *
-     * @note Ownership of the passed device is passed to OpenCV on success.
-     * The caller should additionally call `clRetainDevice` on it if it intends
-     * to continue using the device.
-      */
-    static Device fromHandle(void* d);
-
-    struct Impl;
-    inline Impl* getImpl() const { return (Impl*)p; }
-    inline bool empty() const { return !p; }
-protected:
-    Impl* p;
-};
-
-
-class CV_EXPORTS Context
-{
-public:
-    Context() CV_NOEXCEPT;
-    explicit Context(int dtype);  //!< @deprecated
-    ~Context();
-    Context(const Context& c);
-    Context& operator= (const Context& c);
-    Context(Context&& c) CV_NOEXCEPT;
-    Context& operator = (Context&& c) CV_NOEXCEPT;
-
-    /** @deprecated */
-    bool create();
-    /** @deprecated */
-    bool create(int dtype);
-
-    size_t ndevices() const;
-    Device& device(size_t idx) const;
-    Program getProg(const ProgramSource& prog,
-                    const String& buildopt, String& errmsg);
-    void unloadProg(Program& prog);
-
-
-    /** Get thread-local OpenCL context (initialize if necessary) */
-#if 0  // OpenCV 5.0
-    static Context& getDefault();
-#else
-    static Context& getDefault(bool initialize = true);
-#endif
-
-    /** @returns cl_context value */
-    void* ptr() const;
-
-    /**
-     * @brief Get OpenCL context property specified on context creation
-     * @param propertyId Property id (CL_CONTEXT_* as defined in cl_context_properties type)
-     * @returns Property value if property was specified on clCreateContext, or NULL if context created without the property
-     */
-    void* getOpenCLContextProperty(int propertyId) const;
-
-    bool useSVM() const;
-    void setUseSVM(bool enabled);
-
-    /**
-     * @param context OpenCL handle (cl_context). clRetainContext() is called on success
-     */
-    static Context fromHandle(void* context);
-    static Context fromDevice(const ocl::Device& device);
-    static Context create(const std::string& configuration);
-
-    void release();
-
-    class CV_EXPORTS UserContext {
-    public:
-        virtual ~UserContext();
-    };
-    template <typename T>
-    inline void setUserContext(const std::shared_ptr<T>& userContext) {
-        setUserContext(typeid(T), userContext);
-    }
-    template <typename T>
-    inline std::shared_ptr<T> getUserContext() {
-        return std::dynamic_pointer_cast<T>(getUserContext(typeid(T)));
-    }
-    void setUserContext(std::type_index typeId, const std::shared_ptr<UserContext>& userContext);
-    std::shared_ptr<UserContext> getUserContext(std::type_index typeId);
-
-    struct Impl;
-    inline Impl* getImpl() const { return (Impl*)p; }
-    inline bool empty() const { return !p; }
-// TODO OpenCV 5.0
-//protected:
-    Impl* p;
-};
-
-/** @deprecated */
-class CV_EXPORTS Platform
-{
-public:
-    Platform() CV_NOEXCEPT;
-    ~Platform();
-    Platform(const Platform& p);
-    Platform& operator = (const Platform& p);
-    Platform(Platform&& p) CV_NOEXCEPT;
-    Platform& operator = (Platform&& p) CV_NOEXCEPT;
-
-    void* ptr() const;
-
-    /** @deprecated */
-    static Platform& getDefault();
-
-    struct Impl;
-    inline Impl* getImpl() const { return (Impl*)p; }
-    inline bool empty() const { return !p; }
-protected:
-    Impl* p;
-};
-
-/** @brief Attaches OpenCL context to OpenCV
-@note
-  OpenCV will check if available OpenCL platform has platformName name, then assign context to
-  OpenCV and call `clRetainContext` function. The deviceID device will be used as target device and
-  new command queue will be created.
-@param platformName name of OpenCL platform to attach, this string is used to check if platform is available to OpenCV at runtime
-@param platformID ID of platform attached context was created for
-@param context OpenCL context to be attached to OpenCV
-@param deviceID ID of device, must be created from attached context
-*/
-CV_EXPORTS void attachContext(const String& platformName, void* platformID, void* context, void* deviceID);
-
-/** @brief Convert OpenCL buffer to UMat
-@note
-  OpenCL buffer (cl_mem_buffer) should contain 2D image data, compatible with OpenCV. Memory
-  content is not copied from `clBuffer` to UMat. Instead, buffer handle assigned to UMat and
-  `clRetainMemObject` is called.
-@param cl_mem_buffer source clBuffer handle
-@param step num of bytes in single row
-@param rows number of rows
-@param cols number of cols
-@param type OpenCV type of image
-@param dst destination UMat
-*/
-CV_EXPORTS void convertFromBuffer(void* cl_mem_buffer, size_t step, int rows, int cols, int type, UMat& dst);
-
-/** @brief Convert OpenCL image2d_t to UMat
-@note
-  OpenCL `image2d_t` (cl_mem_image), should be compatible with OpenCV UMat formats. Memory content
-  is copied from image to UMat with `clEnqueueCopyImageToBuffer` function.
-@param cl_mem_image source image2d_t handle
-@param dst destination UMat
-*/
-CV_EXPORTS void convertFromImage(void* cl_mem_image, UMat& dst);
-
-// TODO Move to internal header
-/// @deprecated
-void initializeContextFromHandle(Context& ctx, void* platform, void* context, void* device);
-
-class CV_EXPORTS Queue
-{
-public:
-    Queue() CV_NOEXCEPT;
-    explicit Queue(const Context& c, const Device& d=Device());
-    ~Queue();
-    Queue(const Queue& q);
-    Queue& operator = (const Queue& q);
-    Queue(Queue&& q) CV_NOEXCEPT;
-    Queue& operator = (Queue&& q) CV_NOEXCEPT;
-
-    bool create(const Context& c=Context(), const Device& d=Device());
-    void finish();
-    void* ptr() const;
-    static Queue& getDefault();
-
-    /// @brief Returns OpenCL command queue with enable profiling mode support
-    const Queue& getProfilingQueue() const;
-
-    struct Impl; friend struct Impl;
-    inline Impl* getImpl() const { return p; }
-    inline bool empty() const { return !p; }
-protected:
-    Impl* p;
-};
-
-
-class CV_EXPORTS KernelArg
-{
-public:
-    enum { LOCAL=1, READ_ONLY=2, WRITE_ONLY=4, READ_WRITE=6, CONSTANT=8, PTR_ONLY = 16, NO_SIZE=256 };
-    KernelArg(int _flags, UMat* _m, int wscale=1, int iwscale=1, const void* _obj=0, size_t _sz=0);
-    KernelArg() CV_NOEXCEPT;
-
-    static KernelArg Local(size_t localMemSize)
-    { return KernelArg(LOCAL, 0, 1, 1, 0, localMemSize); }
-    static KernelArg PtrWriteOnly(const UMat& m)
-    { return KernelArg(PTR_ONLY+WRITE_ONLY, (UMat*)&m); }
-    static KernelArg PtrReadOnly(const UMat& m)
-    { return KernelArg(PTR_ONLY+READ_ONLY, (UMat*)&m); }
-    static KernelArg PtrReadWrite(const UMat& m)
-    { return KernelArg(PTR_ONLY+READ_WRITE, (UMat*)&m); }
-    static KernelArg ReadWrite(const UMat& m, int wscale=1, int iwscale=1)
-    { return KernelArg(READ_WRITE, (UMat*)&m, wscale, iwscale); }
-    static KernelArg ReadWriteNoSize(const UMat& m, int wscale=1, int iwscale=1)
-    { return KernelArg(READ_WRITE+NO_SIZE, (UMat*)&m, wscale, iwscale); }
-    static KernelArg ReadOnly(const UMat& m, int wscale=1, int iwscale=1)
-    { return KernelArg(READ_ONLY, (UMat*)&m, wscale, iwscale); }
-    static KernelArg WriteOnly(const UMat& m, int wscale=1, int iwscale=1)
-    { return KernelArg(WRITE_ONLY, (UMat*)&m, wscale, iwscale); }
-    static KernelArg ReadOnlyNoSize(const UMat& m, int wscale=1, int iwscale=1)
-    { return KernelArg(READ_ONLY+NO_SIZE, (UMat*)&m, wscale, iwscale); }
-    static KernelArg WriteOnlyNoSize(const UMat& m, int wscale=1, int iwscale=1)
-    { return KernelArg(WRITE_ONLY+NO_SIZE, (UMat*)&m, wscale, iwscale); }
-    static KernelArg Constant(const Mat& m);
-    template<typename _Tp> static KernelArg Constant(const _Tp* arr, size_t n)
-    { return KernelArg(CONSTANT, 0, 1, 1, (void*)arr, n); }
-
-    int flags;
-    UMat* m;
-    const void* obj;
-    size_t sz;
-    int wscale, iwscale;
-};
-
-
-class CV_EXPORTS Kernel
-{
-public:
-    Kernel() CV_NOEXCEPT;
-    Kernel(const char* kname, const Program& prog);
-    Kernel(const char* kname, const ProgramSource& prog,
-           const String& buildopts = String(), String* errmsg=0);
-    ~Kernel();
-    Kernel(const Kernel& k);
-    Kernel& operator = (const Kernel& k);
-    Kernel(Kernel&& k) CV_NOEXCEPT;
-    Kernel& operator = (Kernel&& k) CV_NOEXCEPT;
-
-    bool empty() const;
-    bool create(const char* kname, const Program& prog);
-    bool create(const char* kname, const ProgramSource& prog,
-                const String& buildopts, String* errmsg=0);
-
-    int set(int i, const void* value, size_t sz);
-    int set(int i, const Image2D& image2D);
-    int set(int i, const UMat& m);
-    int set(int i, const KernelArg& arg);
-    template<typename _Tp> int set(int i, const _Tp& value)
-    { return set(i, &value, sizeof(value)); }
-
-
-protected:
-    template<typename _Tp0> inline
-    int set_args_(int i, const _Tp0& a0) { return set(i, a0); }
-    template<typename _Tp0, typename... _Tps> inline
-    int set_args_(int i, const _Tp0& a0, const _Tps&... rest_args) { i = set(i, a0); return set_args_(i, rest_args...); }
-public:
-    /** @brief Setup OpenCL Kernel arguments.
-    Avoid direct using of set(i, ...) methods.
-    @code
-    bool ok = kernel
-        .args(
-            srcUMat, dstUMat,
-            (float)some_float_param
-        ).run(ndims, globalSize, localSize);
-    if (!ok) return false;
-    @endcode
-    */
-    template<typename... _Tps> inline
-    Kernel& args(const _Tps&... kernel_args) { set_args_(0, kernel_args...); return *this; }
-
-    /** @brief Run the OpenCL kernel (globalsize value may be adjusted)
-
-    @param dims the work problem dimensions. It is the length of globalsize and localsize. It can be either 1, 2 or 3.
-    @param globalsize work items for each dimension. It is not the final globalsize passed to
-      OpenCL. Each dimension will be adjusted to the nearest integer divisible by the corresponding
-      value in localsize. If localsize is NULL, it will still be adjusted depending on dims. The
-      adjusted values are greater than or equal to the original values.
-    @param localsize work-group size for each dimension.
-    @param sync specify whether to wait for OpenCL computation to finish before return.
-    @param q command queue
-
-    @note Use run_() if your kernel code doesn't support adjusted globalsize.
-    */
-    bool run(int dims, size_t globalsize[],
-             size_t localsize[], bool sync, const Queue& q=Queue());
-
-    /** @brief Run the OpenCL kernel
-     *
-     * @param dims the work problem dimensions. It is the length of globalsize and localsize. It can be either 1, 2 or 3.
-     * @param globalsize work items for each dimension. This value is passed to OpenCL without changes.
-     * @param localsize work-group size for each dimension.
-     * @param sync specify whether to wait for OpenCL computation to finish before return.
-     * @param q command queue
-     */
-    bool run_(int dims, size_t globalsize[], size_t localsize[], bool sync, const Queue& q=Queue());
-
-    bool runTask(bool sync, const Queue& q=Queue());
-
-    /** @brief Similar to synchronized run_() call with returning of kernel execution time
-     *
-     * Separate OpenCL command queue may be used (with CL_QUEUE_PROFILING_ENABLE)
-     * @return Execution time in nanoseconds or negative number on error
-     */
-    int64 runProfiling(int dims, size_t globalsize[], size_t localsize[], const Queue& q=Queue());
-
-    size_t workGroupSize() const;
-    size_t preferedWorkGroupSizeMultiple() const;
-    bool compileWorkGroupSize(size_t wsz[]) const;
-    size_t localMemSize() const;
-
-    void* ptr() const;
-    struct Impl;
-
-protected:
-    Impl* p;
-};
-
-class CV_EXPORTS Program
-{
-public:
-    Program() CV_NOEXCEPT;
-    Program(const ProgramSource& src,
-            const String& buildflags, String& errmsg);
-    Program(const Program& prog);
-    Program& operator = (const Program& prog);
-    Program(Program&& prog) CV_NOEXCEPT;
-    Program& operator = (Program&& prog) CV_NOEXCEPT;
-    ~Program();
-
-    bool create(const ProgramSource& src,
-                const String& buildflags, String& errmsg);
-
-    void* ptr() const;
-
-    /**
-     * @brief Query device-specific program binary.
-     *
-     * Returns RAW OpenCL executable binary without additional attachments.
-     *
-     * @sa ProgramSource::fromBinary
-     *
-     * @param[out] binary output buffer
-     */
-    void getBinary(std::vector<char>& binary) const;
-
-    struct Impl; friend struct Impl;
-    inline Impl* getImpl() const { return (Impl*)p; }
-    inline bool empty() const { return !p; }
-protected:
-    Impl* p;
-public:
-#ifndef OPENCV_REMOVE_DEPRECATED_API
-    // TODO Remove this
-    CV_DEPRECATED bool read(const String& buf, const String& buildflags); // removed, use ProgramSource instead
-    CV_DEPRECATED bool write(String& buf) const; // removed, use getBinary() method instead (RAW OpenCL binary)
-    CV_DEPRECATED const ProgramSource& source() const; // implementation removed
-    CV_DEPRECATED String getPrefix() const; // deprecated, implementation replaced
-    CV_DEPRECATED static String getPrefix(const String& buildflags); // deprecated, implementation replaced
-#endif
-};
-
-
-class CV_EXPORTS ProgramSource
-{
-public:
-    typedef uint64 hash_t; // deprecated
-
-    ProgramSource() CV_NOEXCEPT;
-    explicit ProgramSource(const String& module, const String& name, const String& codeStr, const String& codeHash);
-    explicit ProgramSource(const String& prog); // deprecated
-    explicit ProgramSource(const char* prog); // deprecated
-    ~ProgramSource();
-    ProgramSource(const ProgramSource& prog);
-    ProgramSource& operator = (const ProgramSource& prog);
-    ProgramSource(ProgramSource&& prog) CV_NOEXCEPT;
-    ProgramSource& operator = (ProgramSource&& prog) CV_NOEXCEPT;
-
-    const String& source() const; // deprecated
-    hash_t hash() const; // deprecated
-
-
-    /** @brief Describe OpenCL program binary.
-     * Do not call clCreateProgramWithBinary() and/or clBuildProgram().
-     *
-     * Caller should guarantee binary buffer lifetime greater than ProgramSource object (and any of its copies).
-     *
-     * This kind of binary is not portable between platforms in general - it is specific to OpenCL vendor / device / driver version.
-     *
-     * @param module name of program owner module
-     * @param name unique name of program (module+name is used as key for OpenCL program caching)
-     * @param binary buffer address. See buffer lifetime requirement in description.
-     * @param size buffer size
-     * @param buildOptions additional program-related build options passed to clBuildProgram()
-     * @return created ProgramSource object
-     */
-    static ProgramSource fromBinary(const String& module, const String& name,
-            const unsigned char* binary, const size_t size,
-            const cv::String& buildOptions = cv::String());
-
-    /** @brief Describe OpenCL program in SPIR format.
-     * Do not call clCreateProgramWithBinary() and/or clBuildProgram().
-     *
-     * Supports SPIR 1.2 by default (pass '-spir-std=X.Y' in buildOptions to override this behavior)
-     *
-     * Caller should guarantee binary buffer lifetime greater than ProgramSource object (and any of its copies).
-     *
-     * Programs in this format are portable between OpenCL implementations with 'khr_spir' extension:
-     * https://www.khronos.org/registry/OpenCL/sdk/2.0/docs/man/xhtml/cl_khr_spir.html
-     * (but they are not portable between different platforms: 32-bit / 64-bit)
-     *
-     * Note: these programs can't support vendor specific extensions, like 'cl_intel_subgroups'.
-     *
-     * @param module name of program owner module
-     * @param name unique name of program (module+name is used as key for OpenCL program caching)
-     * @param binary buffer address. See buffer lifetime requirement in description.
-     * @param size buffer size
-     * @param buildOptions additional program-related build options passed to clBuildProgram()
-     *        (these options are added automatically: '-x spir' and '-spir-std=1.2')
-     * @return created ProgramSource object.
-     */
-    static ProgramSource fromSPIR(const String& module, const String& name,
-            const unsigned char* binary, const size_t size,
-            const cv::String& buildOptions = cv::String());
-
-    //OpenCL 2.1+ only
-    //static Program fromSPIRV(const String& module, const String& name,
-    //        const unsigned char* binary, const size_t size,
-    //        const cv::String& buildOptions = cv::String());
-
-    struct Impl; friend struct Impl;
-    inline Impl* getImpl() const { return (Impl*)p; }
-    inline bool empty() const { return !p; }
-protected:
-    Impl* p;
-};
-
-class CV_EXPORTS PlatformInfo
-{
-public:
-    PlatformInfo() CV_NOEXCEPT;
-    /**
-     * @param id pointer cl_platform_id (cl_platform_id*)
-     */
-    explicit PlatformInfo(void* id);
-    ~PlatformInfo();
-
-    PlatformInfo(const PlatformInfo& i);
-    PlatformInfo& operator =(const PlatformInfo& i);
-    PlatformInfo(PlatformInfo&& i) CV_NOEXCEPT;
-    PlatformInfo& operator = (PlatformInfo&& i) CV_NOEXCEPT;
-
-    String name() const;
-    String vendor() const;
-
-    /// See CL_PLATFORM_VERSION
-    String version() const;
-    int versionMajor() const;
-    int versionMinor() const;
-
-    int deviceNumber() const;
-    void getDevice(Device& device, int d) const;
-
-    struct Impl;
-    bool empty() const { return !p; }
-protected:
-    Impl* p;
-};
-
-CV_EXPORTS const char* convertTypeStr(int sdepth, int ddepth, int cn, char* buf);
-CV_EXPORTS const char* typeToStr(int t);
-CV_EXPORTS const char* memopTypeToStr(int t);
-CV_EXPORTS const char* vecopTypeToStr(int t);
-CV_EXPORTS const char* getOpenCLErrorString(int errorCode);
-CV_EXPORTS String kernelToStr(InputArray _kernel, int ddepth = -1, const char * name = NULL);
-CV_EXPORTS void getPlatfomsInfo(std::vector<PlatformInfo>& platform_info);
-
-
-enum OclVectorStrategy
-{
-    // all matrices have its own vector width
-    OCL_VECTOR_OWN = 0,
-    // all matrices have maximal vector width among all matrices
-    // (useful for cases when matrices have different data types)
-    OCL_VECTOR_MAX = 1,
-
-    // default strategy
-    OCL_VECTOR_DEFAULT = OCL_VECTOR_OWN
-};
-
-CV_EXPORTS int predictOptimalVectorWidth(InputArray src1, InputArray src2 = noArray(), InputArray src3 = noArray(),
-                                         InputArray src4 = noArray(), InputArray src5 = noArray(), InputArray src6 = noArray(),
-                                         InputArray src7 = noArray(), InputArray src8 = noArray(), InputArray src9 = noArray(),
-                                         OclVectorStrategy strat = OCL_VECTOR_DEFAULT);
-
-CV_EXPORTS int checkOptimalVectorWidth(const int *vectorWidths,
-                                       InputArray src1, InputArray src2 = noArray(), InputArray src3 = noArray(),
-                                       InputArray src4 = noArray(), InputArray src5 = noArray(), InputArray src6 = noArray(),
-                                       InputArray src7 = noArray(), InputArray src8 = noArray(), InputArray src9 = noArray(),
-                                       OclVectorStrategy strat = OCL_VECTOR_DEFAULT);
-
-// with OCL_VECTOR_MAX strategy
-CV_EXPORTS int predictOptimalVectorWidthMax(InputArray src1, InputArray src2 = noArray(), InputArray src3 = noArray(),
-                                            InputArray src4 = noArray(), InputArray src5 = noArray(), InputArray src6 = noArray(),
-                                            InputArray src7 = noArray(), InputArray src8 = noArray(), InputArray src9 = noArray());
-
-CV_EXPORTS void buildOptionsAddMatrixDescription(String& buildOptions, const String& name, InputArray _m);
-
-class CV_EXPORTS Image2D
-{
-public:
-    Image2D() CV_NOEXCEPT;
-
-    /**
-    @param src UMat object from which to get image properties and data
-    @param norm flag to enable the use of normalized channel data types
-    @param alias flag indicating that the image should alias the src UMat. If true, changes to the
-        image or src will be reflected in both objects.
-    */
-    explicit Image2D(const UMat &src, bool norm = false, bool alias = false);
-    Image2D(const Image2D & i);
-    ~Image2D();
-
-    Image2D & operator = (const Image2D & i);
-    Image2D(Image2D &&) CV_NOEXCEPT;
-    Image2D &operator=(Image2D &&) CV_NOEXCEPT;
-
-    /** Indicates if creating an aliased image should succeed.
-    Depends on the underlying platform and the dimensions of the UMat.
-    */
-    static bool canCreateAlias(const UMat &u);
-
-    /** Indicates if the image format is supported.
-    */
-    static bool isFormatSupported(int depth, int cn, bool norm);
-
-    void* ptr() const;
-protected:
-    struct Impl;
-    Impl* p;
-};
-
-class CV_EXPORTS Timer
-{
-public:
-    Timer(const Queue& q);
-    ~Timer();
-    void start();
-    void stop();
-
-    uint64 durationNS() const; //< duration in nanoseconds
-
-protected:
-    struct Impl;
-    Impl* const p;
-
-private:
-    Timer(const Timer&); // disabled
-    Timer& operator=(const Timer&); // disabled
-};
-
-CV_EXPORTS MatAllocator* getOpenCLAllocator();
-
-
-class CV_EXPORTS_W OpenCLExecutionContext
-{
-public:
-    OpenCLExecutionContext() = default;
-    ~OpenCLExecutionContext() = default;
-
-    OpenCLExecutionContext(const OpenCLExecutionContext&) = default;
-    OpenCLExecutionContext(OpenCLExecutionContext&&) = default;
-
-    OpenCLExecutionContext& operator=(const OpenCLExecutionContext&) = default;
-    OpenCLExecutionContext& operator=(OpenCLExecutionContext&&) = default;
-
-    /** Get associated ocl::Context */
-    Context& getContext() const;
-    /** Get the single default associated ocl::Device */
-    Device& getDevice() const;
-    /** Get the single ocl::Queue that is associated with the ocl::Context and
-     *  the single default ocl::Device
-     */
-    Queue& getQueue() const;
-
-    bool useOpenCL() const;
-    void setUseOpenCL(bool flag);
-
-    /** Get OpenCL execution context of current thread.
-     *
-     * Initialize OpenCL execution context if it is empty
-     * - create new
-     * - reuse context of the main thread (threadID = 0)
-     */
-    static OpenCLExecutionContext& getCurrent();
-
-    /** Get OpenCL execution context of current thread (can be empty) */
-    static OpenCLExecutionContext& getCurrentRef();
-
-    /** Bind this OpenCL execution context to current thread.
-     *
-     * Context can't be empty.
-     *
-     * @note clFinish is not called for queue of previous execution context
-     */
-    void bind() const;
-
-    /** Creates new execution context with same OpenCV context and device
-     *
-     * @param q OpenCL queue
-     */
-    OpenCLExecutionContext cloneWithNewQueue(const ocl::Queue& q) const;
-    /** @overload */
-    OpenCLExecutionContext cloneWithNewQueue() const;
-
-    /** @brief Creates OpenCL execution context
-     * OpenCV will check if available OpenCL platform has platformName name,
-     * then assign context to OpenCV.
-     * The deviceID device will be used as target device and a new command queue will be created.
-     *
-     * @note On success, ownership of one reference of the context and device is taken.
-     * The caller should additionally call `clRetainContext` and/or `clRetainDevice`
-     * to increase the reference count if it wishes to continue using them.
-     *
-     * @param platformName name of OpenCL platform to attach, this string is used to check if platform is available to OpenCV at runtime
-     * @param platformID ID of platform attached context was created for (cl_platform_id)
-     * @param context OpenCL context to be attached to OpenCV (cl_context)
-     * @param deviceID OpenCL device (cl_device_id)
-     */
-    static OpenCLExecutionContext create(const std::string& platformName, void* platformID, void* context, void* deviceID);
-
-    /** @brief Creates OpenCL execution context
-     *
-     * @param context non-empty OpenCL context
-     * @param device non-empty OpenCL device (must be a part of context)
-     * @param queue non-empty OpenCL queue for provided context and device
-     */
-    static OpenCLExecutionContext create(const Context& context, const Device& device, const ocl::Queue& queue);
-    /** @overload */
-    static OpenCLExecutionContext create(const Context& context, const Device& device);
-
-    struct Impl;
-    inline bool empty() const { return !p; }
-    void release();
-protected:
-    std::shared_ptr<Impl> p;
-};
-
-class OpenCLExecutionContextScope
-{
-    OpenCLExecutionContext ctx_;
-public:
-    inline OpenCLExecutionContextScope(const OpenCLExecutionContext& ctx)
-    {
-        CV_Assert(!ctx.empty());
-        ctx_ = OpenCLExecutionContext::getCurrentRef();
-        ctx.bind();
-    }
-
-    inline ~OpenCLExecutionContextScope()
-    {
-        if (!ctx_.empty())
-        {
-            ctx_.bind();
-        }
-    }
-};
-
-#ifdef __OPENCV_BUILD
-namespace internal {
-
-CV_EXPORTS bool isOpenCLForced();
-#define OCL_FORCE_CHECK(condition) (cv::ocl::internal::isOpenCLForced() || (condition))
-
-CV_EXPORTS bool isPerformanceCheckBypassed();
-#define OCL_PERFORMANCE_CHECK(condition) (cv::ocl::internal::isPerformanceCheckBypassed() || (condition))
-
-CV_EXPORTS bool isCLBuffer(UMat& u);
-
-} // namespace internal
-#endif
-
-//! @}
-
-}}
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/ocl_genbase.hpp b/projects/llm_framework/include/opencv4/opencv2/core/ocl_genbase.hpp
deleted file mode 100644
index 5334cf1f..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/ocl_genbase.hpp
+++ /dev/null
@@ -1,69 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the OpenCV Foundation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_OPENCL_GENBASE_HPP
-#define OPENCV_OPENCL_GENBASE_HPP
-
-//! @cond IGNORED
-
-namespace cv {
-namespace ocl {
-
-class ProgramSource;
-
-namespace internal {
-
-struct CV_EXPORTS ProgramEntry
-{
-    const char* module;
-    const char* name;
-    const char* programCode;
-    const char* programHash;
-    ProgramSource* pProgramSource;
-
-    operator ProgramSource& () const;
-};
-
-} } } // namespace
-
-//! @endcond
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/ocl_defs.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/ocl_defs.hpp
deleted file mode 100644
index 14df750f..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/ocl_defs.hpp
+++ /dev/null
@@ -1,82 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-// Copyright (C) 2014, Advanced Micro Devices, Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-
-#ifndef OPENCV_CORE_OPENCL_DEFS_HPP
-#define OPENCV_CORE_OPENCL_DEFS_HPP
-
-#include "opencv2/core/utility.hpp"
-#include "cvconfig.h"
-
-namespace cv { namespace ocl {
-#ifdef HAVE_OPENCL
-/// Call is similar to useOpenCL() but doesn't try to load OpenCL runtime or create OpenCL context
-CV_EXPORTS bool isOpenCLActivated();
-#else
-static inline bool isOpenCLActivated() { return false; }
-#endif
-}} // namespace
-
-
-//#define CV_OPENCL_RUN_ASSERT
-
-#ifdef HAVE_OPENCL
-
-#ifdef CV_OPENCL_RUN_VERBOSE
-#define CV_OCL_RUN_(condition, func, ...)                                   \
-    {                                                                       \
-        if (cv::ocl::isOpenCLActivated() && (condition) && func)            \
-        {                                                                   \
-            printf("%s: OpenCL implementation is running\n", CV_Func);      \
-            fflush(stdout);                                                 \
-            CV_IMPL_ADD(CV_IMPL_OCL);                                       \
-            return __VA_ARGS__;                                             \
-        }                                                                   \
-        else                                                                \
-        {                                                                   \
-            printf("%s: Plain implementation is running\n", CV_Func);       \
-            fflush(stdout);                                                 \
-        }                                                                   \
-    }
-#elif defined CV_OPENCL_RUN_ASSERT
-#define CV_OCL_RUN_(condition, func, ...)                                   \
-    {                                                                       \
-        if (cv::ocl::isOpenCLActivated() && (condition))                    \
-        {                                                                   \
-            if(func)                                                        \
-            {                                                               \
-                CV_IMPL_ADD(CV_IMPL_OCL);                                   \
-            }                                                               \
-            else                                                            \
-            {                                                               \
-                CV_Error(cv::Error::StsAssert, #func);                      \
-            }                                                               \
-            return __VA_ARGS__;                                             \
-        }                                                                   \
-    }
-#else
-#define CV_OCL_RUN_(condition, func, ...)                                   \
-try \
-{ \
-    if (cv::ocl::isOpenCLActivated() && (condition) && func)                \
-    {                                                                       \
-        CV_IMPL_ADD(CV_IMPL_OCL);                                           \
-        return __VA_ARGS__;                                                 \
-    } \
-} \
-catch (const cv::Exception& e) \
-{ \
-    CV_UNUSED(e); /* TODO: Add some logging here */ \
-}
-#endif
-
-#else
-#define CV_OCL_RUN_(condition, func, ...)
-#endif
-
-#define CV_OCL_RUN(condition, func) CV_OCL_RUN_(condition, func)
-
-#endif // OPENCV_CORE_OPENCL_DEFS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/opencl_info.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/opencl_info.hpp
deleted file mode 100644
index 3ead76e5..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/opencl_info.hpp
+++ /dev/null
@@ -1,212 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#include <iostream>
-
-#include <opencv2/core.hpp>
-#include <opencv2/core/ocl.hpp>
-
-#ifndef DUMP_CONFIG_PROPERTY
-#define DUMP_CONFIG_PROPERTY(...)
-#endif
-
-#ifndef DUMP_MESSAGE_STDOUT
-#define DUMP_MESSAGE_STDOUT(...) do { std::cout << __VA_ARGS__ << std::endl; } while (false)
-#endif
-
-namespace cv {
-
-namespace {
-static std::string bytesToStringRepr(size_t value)
-{
-    size_t b = value % 1024;
-    value /= 1024;
-
-    size_t kb = value % 1024;
-    value /= 1024;
-
-    size_t mb = value % 1024;
-    value /= 1024;
-
-    size_t gb = value;
-
-    std::ostringstream stream;
-
-    if (gb > 0)
-        stream << gb << " GB ";
-    if (mb > 0)
-        stream << mb << " MB ";
-    if (kb > 0)
-        stream << kb << " KB ";
-    if (b > 0)
-        stream << b << " B";
-
-    std::string s = stream.str();
-    if (s[s.size() - 1] == ' ')
-        s = s.substr(0, s.size() - 1);
-    return s;
-}
-
-static String getDeviceTypeString(const cv::ocl::Device& device)
-{
-    if (device.type() == cv::ocl::Device::TYPE_CPU) {
-        return "CPU";
-    }
-
-    if (device.type() == cv::ocl::Device::TYPE_GPU) {
-        if (device.hostUnifiedMemory()) {
-            return "iGPU";
-        } else {
-            return "dGPU";
-        }
-    }
-
-    return "unknown";
-}
-} // namespace
-
-static void dumpOpenCLInformation()
-{
-    using namespace cv::ocl;
-
-    try
-    {
-        if (!haveOpenCL() || !useOpenCL())
-        {
-            DUMP_MESSAGE_STDOUT("OpenCL is disabled");
-            DUMP_CONFIG_PROPERTY("cv_ocl", "disabled");
-            return;
-        }
-
-        std::vector<PlatformInfo> platforms;
-        cv::ocl::getPlatfomsInfo(platforms);
-        if (platforms.empty())
-        {
-            DUMP_MESSAGE_STDOUT("OpenCL is not available");
-            DUMP_CONFIG_PROPERTY("cv_ocl", "not available");
-            return;
-        }
-
-        DUMP_MESSAGE_STDOUT("OpenCL Platforms: ");
-        for (size_t i = 0; i < platforms.size(); i++)
-        {
-            const PlatformInfo* platform = &platforms[i];
-            DUMP_MESSAGE_STDOUT("    " << platform->name());
-            Device current_device;
-            for (int j = 0; j < platform->deviceNumber(); j++)
-            {
-                platform->getDevice(current_device, j);
-                String deviceTypeStr = getDeviceTypeString(current_device);
-                DUMP_MESSAGE_STDOUT( "        " << deviceTypeStr << ": " << current_device.name() << " (" << current_device.version() << ")");
-                DUMP_CONFIG_PROPERTY( cv::format("cv_ocl_platform_%d_device_%d", (int)i, j ),
-                    cv::format("(Platform=%s)(Type=%s)(Name=%s)(Version=%s)",
-                    platform->name().c_str(), deviceTypeStr.c_str(), current_device.name().c_str(), current_device.version().c_str()) );
-            }
-        }
-        const Device& device = Device::getDefault();
-        if (!device.available())
-            CV_Error(Error::OpenCLInitError, "OpenCL device is not available");
-
-        DUMP_MESSAGE_STDOUT("Current OpenCL device: ");
-
-        String deviceTypeStr = getDeviceTypeString(device);
-        DUMP_MESSAGE_STDOUT("    Type = " << deviceTypeStr);
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_deviceType", deviceTypeStr);
-
-        DUMP_MESSAGE_STDOUT("    Name = " << device.name());
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_deviceName", device.name());
-
-        DUMP_MESSAGE_STDOUT("    Version = " << device.version());
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_deviceVersion", device.version());
-
-        DUMP_MESSAGE_STDOUT("    Driver version = " << device.driverVersion());
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_driverVersion", device.driverVersion());
-
-        DUMP_MESSAGE_STDOUT("    Address bits = " << device.addressBits());
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_addressBits", device.addressBits());
-
-        DUMP_MESSAGE_STDOUT("    Compute units = " << device.maxComputeUnits());
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_maxComputeUnits", device.maxComputeUnits());
-
-        DUMP_MESSAGE_STDOUT("    Max work group size = " << device.maxWorkGroupSize());
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_maxWorkGroupSize", device.maxWorkGroupSize());
-
-        std::string localMemorySizeStr = bytesToStringRepr(device.localMemSize());
-        DUMP_MESSAGE_STDOUT("    Local memory size = " << localMemorySizeStr);
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_localMemSize", device.localMemSize());
-
-        std::string maxMemAllocSizeStr = bytesToStringRepr(device.maxMemAllocSize());
-        DUMP_MESSAGE_STDOUT("    Max memory allocation size = " << maxMemAllocSizeStr);
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_maxMemAllocSize", device.maxMemAllocSize());
-
-        const char* doubleSupportStr = device.doubleFPConfig() > 0 ? "Yes" : "No";
-        DUMP_MESSAGE_STDOUT("    Double support = " << doubleSupportStr);
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_haveDoubleSupport", device.doubleFPConfig() > 0);
-
-        const char* halfSupportStr = device.halfFPConfig() > 0 ? "Yes" : "No";
-        DUMP_MESSAGE_STDOUT("    Half support = " << halfSupportStr);
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_haveHalfSupport", device.halfFPConfig() > 0);
-
-        const char* isUnifiedMemoryStr = device.hostUnifiedMemory() ? "Yes" : "No";
-        DUMP_MESSAGE_STDOUT("    Host unified memory = " << isUnifiedMemoryStr);
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_hostUnifiedMemory", device.hostUnifiedMemory());
-
-        DUMP_MESSAGE_STDOUT("    Device extensions:");
-        String extensionsStr = device.extensions();
-        size_t pos = 0;
-        while (pos < extensionsStr.size())
-        {
-            size_t pos2 = extensionsStr.find(' ', pos);
-            if (pos2 == String::npos)
-                pos2 = extensionsStr.size();
-            if (pos2 > pos)
-            {
-                String extensionName = extensionsStr.substr(pos, pos2 - pos);
-                DUMP_MESSAGE_STDOUT("        " << extensionName);
-            }
-            pos = pos2 + 1;
-        }
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_extensions", extensionsStr);
-
-        const char* haveAmdBlasStr = haveAmdBlas() ? "Yes" : "No";
-        DUMP_MESSAGE_STDOUT("    Has AMD Blas = " << haveAmdBlasStr);
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_AmdBlas", haveAmdBlas());
-
-        const char* haveAmdFftStr = haveAmdFft() ? "Yes" : "No";
-        DUMP_MESSAGE_STDOUT("    Has AMD Fft = " << haveAmdFftStr);
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_AmdFft", haveAmdFft());
-
-
-        DUMP_MESSAGE_STDOUT("    Preferred vector width char = " << device.preferredVectorWidthChar());
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_preferredVectorWidthChar", device.preferredVectorWidthChar());
-
-        DUMP_MESSAGE_STDOUT("    Preferred vector width short = " << device.preferredVectorWidthShort());
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_preferredVectorWidthShort", device.preferredVectorWidthShort());
-
-        DUMP_MESSAGE_STDOUT("    Preferred vector width int = " << device.preferredVectorWidthInt());
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_preferredVectorWidthInt", device.preferredVectorWidthInt());
-
-        DUMP_MESSAGE_STDOUT("    Preferred vector width long = " << device.preferredVectorWidthLong());
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_preferredVectorWidthLong", device.preferredVectorWidthLong());
-
-        DUMP_MESSAGE_STDOUT("    Preferred vector width float = " << device.preferredVectorWidthFloat());
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_preferredVectorWidthFloat", device.preferredVectorWidthFloat());
-
-        DUMP_MESSAGE_STDOUT("    Preferred vector width double = " << device.preferredVectorWidthDouble());
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_preferredVectorWidthDouble", device.preferredVectorWidthDouble());
-
-        DUMP_MESSAGE_STDOUT("    Preferred vector width half = " << device.preferredVectorWidthHalf());
-        DUMP_CONFIG_PROPERTY("cv_ocl_current_preferredVectorWidthHalf", device.preferredVectorWidthHalf());
-    }
-    catch (...)
-    {
-        DUMP_MESSAGE_STDOUT("Exception. Can't dump OpenCL info");
-        DUMP_MESSAGE_STDOUT("OpenCL device not available");
-        DUMP_CONFIG_PROPERTY("cv_ocl", "not available");
-    }
-}
-#undef DUMP_MESSAGE_STDOUT
-#undef DUMP_CONFIG_PROPERTY
-
-} // namespace
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/opencl_svm.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/opencl_svm.hpp
deleted file mode 100644
index 7453082a..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/opencl_svm.hpp
+++ /dev/null
@@ -1,81 +0,0 @@
-/* See LICENSE file in the root OpenCV directory */
-
-#ifndef OPENCV_CORE_OPENCL_SVM_HPP
-#define OPENCV_CORE_OPENCL_SVM_HPP
-
-//
-// Internal usage only (binary compatibility is not guaranteed)
-//
-#ifndef __OPENCV_BUILD
-#error Internal header file
-#endif
-
-#if defined(HAVE_OPENCL) && defined(HAVE_OPENCL_SVM)
-#include "runtime/opencl_core.hpp"
-#include "runtime/opencl_svm_20.hpp"
-#include "runtime/opencl_svm_hsa_extension.hpp"
-
-namespace cv { namespace ocl { namespace svm {
-
-struct SVMCapabilities
-{
-    enum Value
-    {
-        SVM_COARSE_GRAIN_BUFFER = (1 << 0),
-        SVM_FINE_GRAIN_BUFFER = (1 << 1),
-        SVM_FINE_GRAIN_SYSTEM = (1 << 2),
-        SVM_ATOMICS = (1 << 3),
-    };
-    int value_;
-
-    SVMCapabilities(int capabilities = 0) : value_(capabilities) { }
-    operator int() const { return value_; }
-
-    inline bool isNoSVMSupport() const { return value_ == 0; }
-    inline bool isSupportCoarseGrainBuffer() const { return (value_ & SVM_COARSE_GRAIN_BUFFER) != 0; }
-    inline bool isSupportFineGrainBuffer() const { return (value_ & SVM_FINE_GRAIN_BUFFER) != 0; }
-    inline bool isSupportFineGrainSystem() const { return (value_ & SVM_FINE_GRAIN_SYSTEM) != 0; }
-    inline bool isSupportAtomics() const { return (value_ & SVM_ATOMICS) != 0; }
-};
-
-CV_EXPORTS const SVMCapabilities getSVMCapabilitites(const ocl::Context& context);
-
-struct SVMFunctions
-{
-    clSVMAllocAMD_fn fn_clSVMAlloc;
-    clSVMFreeAMD_fn fn_clSVMFree;
-    clSetKernelArgSVMPointerAMD_fn fn_clSetKernelArgSVMPointer;
-    //clSetKernelExecInfoAMD_fn fn_clSetKernelExecInfo;
-    //clEnqueueSVMFreeAMD_fn fn_clEnqueueSVMFree;
-    clEnqueueSVMMemcpyAMD_fn fn_clEnqueueSVMMemcpy;
-    clEnqueueSVMMemFillAMD_fn fn_clEnqueueSVMMemFill;
-    clEnqueueSVMMapAMD_fn fn_clEnqueueSVMMap;
-    clEnqueueSVMUnmapAMD_fn fn_clEnqueueSVMUnmap;
-
-    inline SVMFunctions()
-        : fn_clSVMAlloc(NULL), fn_clSVMFree(NULL),
-          fn_clSetKernelArgSVMPointer(NULL), /*fn_clSetKernelExecInfo(NULL),*/
-          /*fn_clEnqueueSVMFree(NULL),*/ fn_clEnqueueSVMMemcpy(NULL), fn_clEnqueueSVMMemFill(NULL),
-          fn_clEnqueueSVMMap(NULL), fn_clEnqueueSVMUnmap(NULL)
-    {
-        // nothing
-    }
-
-    inline bool isValid() const
-    {
-        return fn_clSVMAlloc != NULL && fn_clSVMFree && fn_clSetKernelArgSVMPointer &&
-                /*fn_clSetKernelExecInfo && fn_clEnqueueSVMFree &&*/ fn_clEnqueueSVMMemcpy &&
-                fn_clEnqueueSVMMemFill && fn_clEnqueueSVMMap && fn_clEnqueueSVMUnmap;
-    }
-};
-
-// We should guarantee that SVMFunctions lifetime is not less than context's lifetime
-CV_EXPORTS const SVMFunctions* getSVMFunctions(const ocl::Context& context);
-
-CV_EXPORTS bool useSVM(UMatUsageFlags usageFlags);
-
-}}} //namespace cv::ocl::svm
-#endif
-
-#endif // OPENCV_CORE_OPENCL_SVM_HPP
-/* End of file. */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_clblas.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_clblas.hpp
deleted file mode 100644
index 2749927b..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_clblas.hpp
+++ /dev/null
@@ -1,602 +0,0 @@
-//
-// AUTOGENERATED, DO NOT EDIT
-//
-#ifndef OPENCV_CORE_OCL_RUNTIME_CLAMDBLAS_HPP
-#error "Invalid usage"
-#endif
-
-// generated by parser_clblas.py
-#define clblasCaxpy clblasCaxpy_
-#define clblasCcopy clblasCcopy_
-#define clblasCdotc clblasCdotc_
-#define clblasCdotu clblasCdotu_
-#define clblasCgbmv clblasCgbmv_
-#define clblasCgemm clblasCgemm_
-#define clblasCgemv clblasCgemv_
-#define clblasCgerc clblasCgerc_
-#define clblasCgeru clblasCgeru_
-#define clblasChbmv clblasChbmv_
-#define clblasChemm clblasChemm_
-#define clblasChemv clblasChemv_
-#define clblasCher clblasCher_
-#define clblasCher2 clblasCher2_
-#define clblasCher2k clblasCher2k_
-#define clblasCherk clblasCherk_
-#define clblasChpmv clblasChpmv_
-#define clblasChpr clblasChpr_
-#define clblasChpr2 clblasChpr2_
-#define clblasCrotg clblasCrotg_
-#define clblasCscal clblasCscal_
-#define clblasCsrot clblasCsrot_
-#define clblasCsscal clblasCsscal_
-#define clblasCswap clblasCswap_
-#define clblasCsymm clblasCsymm_
-#define clblasCsyr2k clblasCsyr2k_
-#define clblasCsyrk clblasCsyrk_
-#define clblasCtbmv clblasCtbmv_
-#define clblasCtbsv clblasCtbsv_
-#define clblasCtpmv clblasCtpmv_
-#define clblasCtpsv clblasCtpsv_
-#define clblasCtrmm clblasCtrmm_
-#define clblasCtrmv clblasCtrmv_
-#define clblasCtrsm clblasCtrsm_
-#define clblasCtrsv clblasCtrsv_
-#define clblasDasum clblasDasum_
-#define clblasDaxpy clblasDaxpy_
-#define clblasDcopy clblasDcopy_
-#define clblasDdot clblasDdot_
-#define clblasDgbmv clblasDgbmv_
-#define clblasDgemm clblasDgemm_
-#define clblasDgemv clblasDgemv_
-#define clblasDger clblasDger_
-#define clblasDnrm2 clblasDnrm2_
-#define clblasDrot clblasDrot_
-#define clblasDrotg clblasDrotg_
-#define clblasDrotm clblasDrotm_
-#define clblasDrotmg clblasDrotmg_
-#define clblasDsbmv clblasDsbmv_
-#define clblasDscal clblasDscal_
-#define clblasDspmv clblasDspmv_
-#define clblasDspr clblasDspr_
-#define clblasDspr2 clblasDspr2_
-#define clblasDswap clblasDswap_
-#define clblasDsymm clblasDsymm_
-#define clblasDsymv clblasDsymv_
-#define clblasDsyr clblasDsyr_
-#define clblasDsyr2 clblasDsyr2_
-#define clblasDsyr2k clblasDsyr2k_
-#define clblasDsyrk clblasDsyrk_
-#define clblasDtbmv clblasDtbmv_
-#define clblasDtbsv clblasDtbsv_
-#define clblasDtpmv clblasDtpmv_
-#define clblasDtpsv clblasDtpsv_
-#define clblasDtrmm clblasDtrmm_
-#define clblasDtrmv clblasDtrmv_
-#define clblasDtrsm clblasDtrsm_
-#define clblasDtrsv clblasDtrsv_
-#define clblasDzasum clblasDzasum_
-#define clblasDznrm2 clblasDznrm2_
-#define clblasGetVersion clblasGetVersion_
-#define clblasSasum clblasSasum_
-#define clblasSaxpy clblasSaxpy_
-#define clblasScasum clblasScasum_
-#define clblasScnrm2 clblasScnrm2_
-#define clblasScopy clblasScopy_
-#define clblasSdot clblasSdot_
-#define clblasSetup clblasSetup_
-#define clblasSgbmv clblasSgbmv_
-#define clblasSgemm clblasSgemm_
-#define clblasSgemv clblasSgemv_
-#define clblasSger clblasSger_
-#define clblasSnrm2 clblasSnrm2_
-#define clblasSrot clblasSrot_
-#define clblasSrotg clblasSrotg_
-#define clblasSrotm clblasSrotm_
-#define clblasSrotmg clblasSrotmg_
-#define clblasSsbmv clblasSsbmv_
-#define clblasSscal clblasSscal_
-#define clblasSspmv clblasSspmv_
-#define clblasSspr clblasSspr_
-#define clblasSspr2 clblasSspr2_
-#define clblasSswap clblasSswap_
-#define clblasSsymm clblasSsymm_
-#define clblasSsymv clblasSsymv_
-#define clblasSsyr clblasSsyr_
-#define clblasSsyr2 clblasSsyr2_
-#define clblasSsyr2k clblasSsyr2k_
-#define clblasSsyrk clblasSsyrk_
-#define clblasStbmv clblasStbmv_
-#define clblasStbsv clblasStbsv_
-#define clblasStpmv clblasStpmv_
-#define clblasStpsv clblasStpsv_
-#define clblasStrmm clblasStrmm_
-#define clblasStrmv clblasStrmv_
-#define clblasStrsm clblasStrsm_
-#define clblasStrsv clblasStrsv_
-#define clblasTeardown clblasTeardown_
-#define clblasZaxpy clblasZaxpy_
-#define clblasZcopy clblasZcopy_
-#define clblasZdotc clblasZdotc_
-#define clblasZdotu clblasZdotu_
-#define clblasZdrot clblasZdrot_
-#define clblasZdscal clblasZdscal_
-#define clblasZgbmv clblasZgbmv_
-#define clblasZgemm clblasZgemm_
-#define clblasZgemv clblasZgemv_
-#define clblasZgerc clblasZgerc_
-#define clblasZgeru clblasZgeru_
-#define clblasZhbmv clblasZhbmv_
-#define clblasZhemm clblasZhemm_
-#define clblasZhemv clblasZhemv_
-#define clblasZher clblasZher_
-#define clblasZher2 clblasZher2_
-#define clblasZher2k clblasZher2k_
-#define clblasZherk clblasZherk_
-#define clblasZhpmv clblasZhpmv_
-#define clblasZhpr clblasZhpr_
-#define clblasZhpr2 clblasZhpr2_
-#define clblasZrotg clblasZrotg_
-#define clblasZscal clblasZscal_
-#define clblasZswap clblasZswap_
-#define clblasZsymm clblasZsymm_
-#define clblasZsyr2k clblasZsyr2k_
-#define clblasZsyrk clblasZsyrk_
-#define clblasZtbmv clblasZtbmv_
-#define clblasZtbsv clblasZtbsv_
-#define clblasZtpmv clblasZtpmv_
-#define clblasZtpsv clblasZtpsv_
-#define clblasZtrmm clblasZtrmm_
-#define clblasZtrmv clblasZtrmv_
-#define clblasZtrsm clblasZtrsm_
-#define clblasZtrsv clblasZtrsv_
-#define clblasiCamax clblasiCamax_
-#define clblasiDamax clblasiDamax_
-#define clblasiSamax clblasiSamax_
-#define clblasiZamax clblasiZamax_
-
-#include <clBLAS.h>
-
-// generated by parser_clblas.py
-#undef clblasCaxpy
-//#define clblasCaxpy clblasCaxpy_pfn
-#undef clblasCcopy
-//#define clblasCcopy clblasCcopy_pfn
-#undef clblasCdotc
-//#define clblasCdotc clblasCdotc_pfn
-#undef clblasCdotu
-//#define clblasCdotu clblasCdotu_pfn
-#undef clblasCgbmv
-//#define clblasCgbmv clblasCgbmv_pfn
-#undef clblasCgemm
-#define clblasCgemm clblasCgemm_pfn
-#undef clblasCgemv
-//#define clblasCgemv clblasCgemv_pfn
-#undef clblasCgerc
-//#define clblasCgerc clblasCgerc_pfn
-#undef clblasCgeru
-//#define clblasCgeru clblasCgeru_pfn
-#undef clblasChbmv
-//#define clblasChbmv clblasChbmv_pfn
-#undef clblasChemm
-//#define clblasChemm clblasChemm_pfn
-#undef clblasChemv
-//#define clblasChemv clblasChemv_pfn
-#undef clblasCher
-//#define clblasCher clblasCher_pfn
-#undef clblasCher2
-//#define clblasCher2 clblasCher2_pfn
-#undef clblasCher2k
-//#define clblasCher2k clblasCher2k_pfn
-#undef clblasCherk
-//#define clblasCherk clblasCherk_pfn
-#undef clblasChpmv
-//#define clblasChpmv clblasChpmv_pfn
-#undef clblasChpr
-//#define clblasChpr clblasChpr_pfn
-#undef clblasChpr2
-//#define clblasChpr2 clblasChpr2_pfn
-#undef clblasCrotg
-//#define clblasCrotg clblasCrotg_pfn
-#undef clblasCscal
-//#define clblasCscal clblasCscal_pfn
-#undef clblasCsrot
-//#define clblasCsrot clblasCsrot_pfn
-#undef clblasCsscal
-//#define clblasCsscal clblasCsscal_pfn
-#undef clblasCswap
-//#define clblasCswap clblasCswap_pfn
-#undef clblasCsymm
-//#define clblasCsymm clblasCsymm_pfn
-#undef clblasCsyr2k
-//#define clblasCsyr2k clblasCsyr2k_pfn
-#undef clblasCsyrk
-//#define clblasCsyrk clblasCsyrk_pfn
-#undef clblasCtbmv
-//#define clblasCtbmv clblasCtbmv_pfn
-#undef clblasCtbsv
-//#define clblasCtbsv clblasCtbsv_pfn
-#undef clblasCtpmv
-//#define clblasCtpmv clblasCtpmv_pfn
-#undef clblasCtpsv
-//#define clblasCtpsv clblasCtpsv_pfn
-#undef clblasCtrmm
-//#define clblasCtrmm clblasCtrmm_pfn
-#undef clblasCtrmv
-//#define clblasCtrmv clblasCtrmv_pfn
-#undef clblasCtrsm
-//#define clblasCtrsm clblasCtrsm_pfn
-#undef clblasCtrsv
-//#define clblasCtrsv clblasCtrsv_pfn
-#undef clblasDasum
-//#define clblasDasum clblasDasum_pfn
-#undef clblasDaxpy
-//#define clblasDaxpy clblasDaxpy_pfn
-#undef clblasDcopy
-//#define clblasDcopy clblasDcopy_pfn
-#undef clblasDdot
-//#define clblasDdot clblasDdot_pfn
-#undef clblasDgbmv
-//#define clblasDgbmv clblasDgbmv_pfn
-#undef clblasDgemm
-#define clblasDgemm clblasDgemm_pfn
-#undef clblasDgemv
-//#define clblasDgemv clblasDgemv_pfn
-#undef clblasDger
-//#define clblasDger clblasDger_pfn
-#undef clblasDnrm2
-//#define clblasDnrm2 clblasDnrm2_pfn
-#undef clblasDrot
-//#define clblasDrot clblasDrot_pfn
-#undef clblasDrotg
-//#define clblasDrotg clblasDrotg_pfn
-#undef clblasDrotm
-//#define clblasDrotm clblasDrotm_pfn
-#undef clblasDrotmg
-//#define clblasDrotmg clblasDrotmg_pfn
-#undef clblasDsbmv
-//#define clblasDsbmv clblasDsbmv_pfn
-#undef clblasDscal
-//#define clblasDscal clblasDscal_pfn
-#undef clblasDspmv
-//#define clblasDspmv clblasDspmv_pfn
-#undef clblasDspr
-//#define clblasDspr clblasDspr_pfn
-#undef clblasDspr2
-//#define clblasDspr2 clblasDspr2_pfn
-#undef clblasDswap
-//#define clblasDswap clblasDswap_pfn
-#undef clblasDsymm
-//#define clblasDsymm clblasDsymm_pfn
-#undef clblasDsymv
-//#define clblasDsymv clblasDsymv_pfn
-#undef clblasDsyr
-//#define clblasDsyr clblasDsyr_pfn
-#undef clblasDsyr2
-//#define clblasDsyr2 clblasDsyr2_pfn
-#undef clblasDsyr2k
-//#define clblasDsyr2k clblasDsyr2k_pfn
-#undef clblasDsyrk
-//#define clblasDsyrk clblasDsyrk_pfn
-#undef clblasDtbmv
-//#define clblasDtbmv clblasDtbmv_pfn
-#undef clblasDtbsv
-//#define clblasDtbsv clblasDtbsv_pfn
-#undef clblasDtpmv
-//#define clblasDtpmv clblasDtpmv_pfn
-#undef clblasDtpsv
-//#define clblasDtpsv clblasDtpsv_pfn
-#undef clblasDtrmm
-//#define clblasDtrmm clblasDtrmm_pfn
-#undef clblasDtrmv
-//#define clblasDtrmv clblasDtrmv_pfn
-#undef clblasDtrsm
-//#define clblasDtrsm clblasDtrsm_pfn
-#undef clblasDtrsv
-//#define clblasDtrsv clblasDtrsv_pfn
-#undef clblasDzasum
-//#define clblasDzasum clblasDzasum_pfn
-#undef clblasDznrm2
-//#define clblasDznrm2 clblasDznrm2_pfn
-#undef clblasGetVersion
-//#define clblasGetVersion clblasGetVersion_pfn
-#undef clblasSasum
-//#define clblasSasum clblasSasum_pfn
-#undef clblasSaxpy
-//#define clblasSaxpy clblasSaxpy_pfn
-#undef clblasScasum
-//#define clblasScasum clblasScasum_pfn
-#undef clblasScnrm2
-//#define clblasScnrm2 clblasScnrm2_pfn
-#undef clblasScopy
-//#define clblasScopy clblasScopy_pfn
-#undef clblasSdot
-//#define clblasSdot clblasSdot_pfn
-#undef clblasSetup
-#define clblasSetup clblasSetup_pfn
-#undef clblasSgbmv
-//#define clblasSgbmv clblasSgbmv_pfn
-#undef clblasSgemm
-#define clblasSgemm clblasSgemm_pfn
-#undef clblasSgemv
-//#define clblasSgemv clblasSgemv_pfn
-#undef clblasSger
-//#define clblasSger clblasSger_pfn
-#undef clblasSnrm2
-//#define clblasSnrm2 clblasSnrm2_pfn
-#undef clblasSrot
-//#define clblasSrot clblasSrot_pfn
-#undef clblasSrotg
-//#define clblasSrotg clblasSrotg_pfn
-#undef clblasSrotm
-//#define clblasSrotm clblasSrotm_pfn
-#undef clblasSrotmg
-//#define clblasSrotmg clblasSrotmg_pfn
-#undef clblasSsbmv
-//#define clblasSsbmv clblasSsbmv_pfn
-#undef clblasSscal
-//#define clblasSscal clblasSscal_pfn
-#undef clblasSspmv
-//#define clblasSspmv clblasSspmv_pfn
-#undef clblasSspr
-//#define clblasSspr clblasSspr_pfn
-#undef clblasSspr2
-//#define clblasSspr2 clblasSspr2_pfn
-#undef clblasSswap
-//#define clblasSswap clblasSswap_pfn
-#undef clblasSsymm
-//#define clblasSsymm clblasSsymm_pfn
-#undef clblasSsymv
-//#define clblasSsymv clblasSsymv_pfn
-#undef clblasSsyr
-//#define clblasSsyr clblasSsyr_pfn
-#undef clblasSsyr2
-//#define clblasSsyr2 clblasSsyr2_pfn
-#undef clblasSsyr2k
-//#define clblasSsyr2k clblasSsyr2k_pfn
-#undef clblasSsyrk
-//#define clblasSsyrk clblasSsyrk_pfn
-#undef clblasStbmv
-//#define clblasStbmv clblasStbmv_pfn
-#undef clblasStbsv
-//#define clblasStbsv clblasStbsv_pfn
-#undef clblasStpmv
-//#define clblasStpmv clblasStpmv_pfn
-#undef clblasStpsv
-//#define clblasStpsv clblasStpsv_pfn
-#undef clblasStrmm
-//#define clblasStrmm clblasStrmm_pfn
-#undef clblasStrmv
-//#define clblasStrmv clblasStrmv_pfn
-#undef clblasStrsm
-//#define clblasStrsm clblasStrsm_pfn
-#undef clblasStrsv
-//#define clblasStrsv clblasStrsv_pfn
-#undef clblasTeardown
-#define clblasTeardown clblasTeardown_pfn
-#undef clblasZaxpy
-//#define clblasZaxpy clblasZaxpy_pfn
-#undef clblasZcopy
-//#define clblasZcopy clblasZcopy_pfn
-#undef clblasZdotc
-//#define clblasZdotc clblasZdotc_pfn
-#undef clblasZdotu
-//#define clblasZdotu clblasZdotu_pfn
-#undef clblasZdrot
-//#define clblasZdrot clblasZdrot_pfn
-#undef clblasZdscal
-//#define clblasZdscal clblasZdscal_pfn
-#undef clblasZgbmv
-//#define clblasZgbmv clblasZgbmv_pfn
-#undef clblasZgemm
-#define clblasZgemm clblasZgemm_pfn
-#undef clblasZgemv
-//#define clblasZgemv clblasZgemv_pfn
-#undef clblasZgerc
-//#define clblasZgerc clblasZgerc_pfn
-#undef clblasZgeru
-//#define clblasZgeru clblasZgeru_pfn
-#undef clblasZhbmv
-//#define clblasZhbmv clblasZhbmv_pfn
-#undef clblasZhemm
-//#define clblasZhemm clblasZhemm_pfn
-#undef clblasZhemv
-//#define clblasZhemv clblasZhemv_pfn
-#undef clblasZher
-//#define clblasZher clblasZher_pfn
-#undef clblasZher2
-//#define clblasZher2 clblasZher2_pfn
-#undef clblasZher2k
-//#define clblasZher2k clblasZher2k_pfn
-#undef clblasZherk
-//#define clblasZherk clblasZherk_pfn
-#undef clblasZhpmv
-//#define clblasZhpmv clblasZhpmv_pfn
-#undef clblasZhpr
-//#define clblasZhpr clblasZhpr_pfn
-#undef clblasZhpr2
-//#define clblasZhpr2 clblasZhpr2_pfn
-#undef clblasZrotg
-//#define clblasZrotg clblasZrotg_pfn
-#undef clblasZscal
-//#define clblasZscal clblasZscal_pfn
-#undef clblasZswap
-//#define clblasZswap clblasZswap_pfn
-#undef clblasZsymm
-//#define clblasZsymm clblasZsymm_pfn
-#undef clblasZsyr2k
-//#define clblasZsyr2k clblasZsyr2k_pfn
-#undef clblasZsyrk
-//#define clblasZsyrk clblasZsyrk_pfn
-#undef clblasZtbmv
-//#define clblasZtbmv clblasZtbmv_pfn
-#undef clblasZtbsv
-//#define clblasZtbsv clblasZtbsv_pfn
-#undef clblasZtpmv
-//#define clblasZtpmv clblasZtpmv_pfn
-#undef clblasZtpsv
-//#define clblasZtpsv clblasZtpsv_pfn
-#undef clblasZtrmm
-//#define clblasZtrmm clblasZtrmm_pfn
-#undef clblasZtrmv
-//#define clblasZtrmv clblasZtrmv_pfn
-#undef clblasZtrsm
-//#define clblasZtrsm clblasZtrsm_pfn
-#undef clblasZtrsv
-//#define clblasZtrsv clblasZtrsv_pfn
-#undef clblasiCamax
-//#define clblasiCamax clblasiCamax_pfn
-#undef clblasiDamax
-//#define clblasiDamax clblasiDamax_pfn
-#undef clblasiSamax
-//#define clblasiSamax clblasiSamax_pfn
-#undef clblasiZamax
-//#define clblasiZamax clblasiZamax_pfn
-
-// generated by parser_clblas.py
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCaxpy)(size_t N, cl_float2 alpha, const cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCcopy)(size_t N, const cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCdotc)(size_t N, cl_mem dotProduct, size_t offDP, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCdotu)(size_t N, cl_mem dotProduct, size_t offDP, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCgbmv)(clblasOrder order, clblasTranspose trans, size_t M, size_t N, size_t KL, size_t KU, cl_float2 alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem X, size_t offx, int incx, cl_float2 beta, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-extern CL_RUNTIME_EXPORT clblasStatus (*clblasCgemm)(clblasOrder order, clblasTranspose transA, clblasTranspose transB, size_t M, size_t N, size_t K, FloatComplex alpha, const cl_mem A, size_t offA, size_t lda, const cl_mem B, size_t offB, size_t ldb, FloatComplex beta, cl_mem C, size_t offC, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCgemv)(clblasOrder order, clblasTranspose transA, size_t M, size_t N, FloatComplex alpha, const cl_mem A, size_t offA, size_t lda, const cl_mem x, size_t offx, int incx, FloatComplex beta, cl_mem y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCgerc)(clblasOrder order, size_t M, size_t N, cl_float2 alpha, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem A, size_t offa, size_t lda, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCgeru)(clblasOrder order, size_t M, size_t N, cl_float2 alpha, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem A, size_t offa, size_t lda, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasChbmv)(clblasOrder order, clblasUplo uplo, size_t N, size_t K, cl_float2 alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem X, size_t offx, int incx, cl_float2 beta, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasChemm)(clblasOrder order, clblasSide side, clblasUplo uplo, size_t M, size_t N, cl_float2 alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem B, size_t offb, size_t ldb, cl_float2 beta, cl_mem C, size_t offc, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasChemv)(clblasOrder order, clblasUplo uplo, size_t N, FloatComplex alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem X, size_t offx, int incx, FloatComplex beta, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCher)(clblasOrder order, clblasUplo uplo, size_t N, cl_float alpha, const cl_mem X, size_t offx, int incx, cl_mem A, size_t offa, size_t lda, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCher2)(clblasOrder order, clblasUplo uplo, size_t N, cl_float2 alpha, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem A, size_t offa, size_t lda, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCher2k)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, size_t N, size_t K, FloatComplex alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem B, size_t offb, size_t ldb, cl_float beta, cl_mem C, size_t offc, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCherk)(clblasOrder order, clblasUplo uplo, clblasTranspose transA, size_t N, size_t K, float alpha, const cl_mem A, size_t offa, size_t lda, float beta, cl_mem C, size_t offc, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasChpmv)(clblasOrder order, clblasUplo uplo, size_t N, cl_float2 alpha, const cl_mem AP, size_t offa, const cl_mem X, size_t offx, int incx, cl_float2 beta, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasChpr)(clblasOrder order, clblasUplo uplo, size_t N, cl_float alpha, const cl_mem X, size_t offx, int incx, cl_mem AP, size_t offa, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasChpr2)(clblasOrder order, clblasUplo uplo, size_t N, cl_float2 alpha, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem AP, size_t offa, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCrotg)(cl_mem CA, size_t offCA, cl_mem CB, size_t offCB, cl_mem C, size_t offC, cl_mem S, size_t offS, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCscal)(size_t N, cl_float2 alpha, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCsrot)(size_t N, cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_float C, cl_float S, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCsscal)(size_t N, cl_float alpha, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCswap)(size_t N, cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCsymm)(clblasOrder order, clblasSide side, clblasUplo uplo, size_t M, size_t N, cl_float2 alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem B, size_t offb, size_t ldb, cl_float2 beta, cl_mem C, size_t offc, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCsyr2k)(clblasOrder order, clblasUplo uplo, clblasTranspose transAB, size_t N, size_t K, FloatComplex alpha, const cl_mem A, size_t offA, size_t lda, const cl_mem B, size_t offB, size_t ldb, FloatComplex beta, cl_mem C, size_t offC, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCsyrk)(clblasOrder order, clblasUplo uplo, clblasTranspose transA, size_t N, size_t K, FloatComplex alpha, const cl_mem A, size_t offA, size_t lda, FloatComplex beta, cl_mem C, size_t offC, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCtbmv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, size_t K, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCtbsv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, size_t K, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCtpmv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem AP, size_t offa, cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCtpsv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem A, size_t offa, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCtrmm)(clblasOrder order, clblasSide side, clblasUplo uplo, clblasTranspose transA, clblasDiag diag, size_t M, size_t N, FloatComplex alpha, const cl_mem A, size_t offA, size_t lda, cl_mem B, size_t offB, size_t ldb, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCtrmv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCtrsm)(clblasOrder order, clblasSide side, clblasUplo uplo, clblasTranspose transA, clblasDiag diag, size_t M, size_t N, FloatComplex alpha, const cl_mem A, size_t offA, size_t lda, cl_mem B, size_t offB, size_t ldb, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasCtrsv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDasum)(size_t N, cl_mem asum, size_t offAsum, const cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDaxpy)(size_t N, cl_double alpha, const cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDcopy)(size_t N, const cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDdot)(size_t N, cl_mem dotProduct, size_t offDP, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDgbmv)(clblasOrder order, clblasTranspose trans, size_t M, size_t N, size_t KL, size_t KU, cl_double alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem X, size_t offx, int incx, cl_double beta, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-extern CL_RUNTIME_EXPORT clblasStatus (*clblasDgemm)(clblasOrder order, clblasTranspose transA, clblasTranspose transB, size_t M, size_t N, size_t K, cl_double alpha, const cl_mem A, size_t offA, size_t lda, const cl_mem B, size_t offB, size_t ldb, cl_double beta, cl_mem C, size_t offC, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDgemv)(clblasOrder order, clblasTranspose transA, size_t M, size_t N, cl_double alpha, const cl_mem A, size_t offA, size_t lda, const cl_mem x, size_t offx, int incx, cl_double beta, cl_mem y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDger)(clblasOrder order, size_t M, size_t N, cl_double alpha, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem A, size_t offa, size_t lda, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDnrm2)(size_t N, cl_mem NRM2, size_t offNRM2, const cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDrot)(size_t N, cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_double C, cl_double S, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDrotg)(cl_mem DA, size_t offDA, cl_mem DB, size_t offDB, cl_mem C, size_t offC, cl_mem S, size_t offS, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDrotm)(size_t N, cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, const cl_mem DPARAM, size_t offDparam, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDrotmg)(cl_mem DD1, size_t offDD1, cl_mem DD2, size_t offDD2, cl_mem DX1, size_t offDX1, const cl_mem DY1, size_t offDY1, cl_mem DPARAM, size_t offDparam, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDsbmv)(clblasOrder order, clblasUplo uplo, size_t N, size_t K, cl_double alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem X, size_t offx, int incx, cl_double beta, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDscal)(size_t N, cl_double alpha, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDspmv)(clblasOrder order, clblasUplo uplo, size_t N, cl_double alpha, const cl_mem AP, size_t offa, const cl_mem X, size_t offx, int incx, cl_double beta, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDspr)(clblasOrder order, clblasUplo uplo, size_t N, cl_double alpha, const cl_mem X, size_t offx, int incx, cl_mem AP, size_t offa, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDspr2)(clblasOrder order, clblasUplo uplo, size_t N, cl_double alpha, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem AP, size_t offa, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDswap)(size_t N, cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDsymm)(clblasOrder order, clblasSide side, clblasUplo uplo, size_t M, size_t N, cl_double alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem B, size_t offb, size_t ldb, cl_double beta, cl_mem C, size_t offc, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDsymv)(clblasOrder order, clblasUplo uplo, size_t N, cl_double alpha, const cl_mem A, size_t offA, size_t lda, const cl_mem x, size_t offx, int incx, cl_double beta, cl_mem y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDsyr)(clblasOrder order, clblasUplo uplo, size_t N, cl_double alpha, const cl_mem X, size_t offx, int incx, cl_mem A, size_t offa, size_t lda, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDsyr2)(clblasOrder order, clblasUplo uplo, size_t N, cl_double alpha, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem A, size_t offa, size_t lda, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDsyr2k)(clblasOrder order, clblasUplo uplo, clblasTranspose transAB, size_t N, size_t K, cl_double alpha, const cl_mem A, size_t offA, size_t lda, const cl_mem B, size_t offB, size_t ldb, cl_double beta, cl_mem C, size_t offC, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDsyrk)(clblasOrder order, clblasUplo uplo, clblasTranspose transA, size_t N, size_t K, cl_double alpha, const cl_mem A, size_t offA, size_t lda, cl_double beta, cl_mem C, size_t offC, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDtbmv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, size_t K, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDtbsv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, size_t K, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDtpmv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem AP, size_t offa, cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDtpsv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem A, size_t offa, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDtrmm)(clblasOrder order, clblasSide side, clblasUplo uplo, clblasTranspose transA, clblasDiag diag, size_t M, size_t N, cl_double alpha, const cl_mem A, size_t offA, size_t lda, cl_mem B, size_t offB, size_t ldb, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDtrmv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDtrsm)(clblasOrder order, clblasSide side, clblasUplo uplo, clblasTranspose transA, clblasDiag diag, size_t M, size_t N, cl_double alpha, const cl_mem A, size_t offA, size_t lda, cl_mem B, size_t offB, size_t ldb, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDtrsv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDzasum)(size_t N, cl_mem asum, size_t offAsum, const cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasDznrm2)(size_t N, cl_mem NRM2, size_t offNRM2, const cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasGetVersion)(cl_uint* major, cl_uint* minor, cl_uint* patch);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSasum)(size_t N, cl_mem asum, size_t offAsum, const cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSaxpy)(size_t N, cl_float alpha, const cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasScasum)(size_t N, cl_mem asum, size_t offAsum, const cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasScnrm2)(size_t N, cl_mem NRM2, size_t offNRM2, const cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasScopy)(size_t N, const cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSdot)(size_t N, cl_mem dotProduct, size_t offDP, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-extern CL_RUNTIME_EXPORT clblasStatus (*clblasSetup)();
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSgbmv)(clblasOrder order, clblasTranspose trans, size_t M, size_t N, size_t KL, size_t KU, cl_float alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem X, size_t offx, int incx, cl_float beta, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-extern CL_RUNTIME_EXPORT clblasStatus (*clblasSgemm)(clblasOrder order, clblasTranspose transA, clblasTranspose transB, size_t M, size_t N, size_t K, cl_float alpha, const cl_mem A, size_t offA, size_t lda, const cl_mem B, size_t offB, size_t ldb, cl_float beta, cl_mem C, size_t offC, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSgemv)(clblasOrder order, clblasTranspose transA, size_t M, size_t N, cl_float alpha, const cl_mem A, size_t offA, size_t lda, const cl_mem x, size_t offx, int incx, cl_float beta, cl_mem y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSger)(clblasOrder order, size_t M, size_t N, cl_float alpha, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem A, size_t offa, size_t lda, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSnrm2)(size_t N, cl_mem NRM2, size_t offNRM2, const cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSrot)(size_t N, cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_float C, cl_float S, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSrotg)(cl_mem SA, size_t offSA, cl_mem SB, size_t offSB, cl_mem C, size_t offC, cl_mem S, size_t offS, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSrotm)(size_t N, cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, const cl_mem SPARAM, size_t offSparam, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSrotmg)(cl_mem SD1, size_t offSD1, cl_mem SD2, size_t offSD2, cl_mem SX1, size_t offSX1, const cl_mem SY1, size_t offSY1, cl_mem SPARAM, size_t offSparam, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSsbmv)(clblasOrder order, clblasUplo uplo, size_t N, size_t K, cl_float alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem X, size_t offx, int incx, cl_float beta, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSscal)(size_t N, cl_float alpha, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSspmv)(clblasOrder order, clblasUplo uplo, size_t N, cl_float alpha, const cl_mem AP, size_t offa, const cl_mem X, size_t offx, int incx, cl_float beta, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSspr)(clblasOrder order, clblasUplo uplo, size_t N, cl_float alpha, const cl_mem X, size_t offx, int incx, cl_mem AP, size_t offa, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSspr2)(clblasOrder order, clblasUplo uplo, size_t N, cl_float alpha, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem AP, size_t offa, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSswap)(size_t N, cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSsymm)(clblasOrder order, clblasSide side, clblasUplo uplo, size_t M, size_t N, cl_float alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem B, size_t offb, size_t ldb, cl_float beta, cl_mem C, size_t offc, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSsymv)(clblasOrder order, clblasUplo uplo, size_t N, cl_float alpha, const cl_mem A, size_t offA, size_t lda, const cl_mem x, size_t offx, int incx, cl_float beta, cl_mem y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSsyr)(clblasOrder order, clblasUplo uplo, size_t N, cl_float alpha, const cl_mem X, size_t offx, int incx, cl_mem A, size_t offa, size_t lda, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSsyr2)(clblasOrder order, clblasUplo uplo, size_t N, cl_float alpha, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem A, size_t offa, size_t lda, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSsyr2k)(clblasOrder order, clblasUplo uplo, clblasTranspose transAB, size_t N, size_t K, cl_float alpha, const cl_mem A, size_t offA, size_t lda, const cl_mem B, size_t offB, size_t ldb, cl_float beta, cl_mem C, size_t offC, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasSsyrk)(clblasOrder order, clblasUplo uplo, clblasTranspose transA, size_t N, size_t K, cl_float alpha, const cl_mem A, size_t offA, size_t lda, cl_float beta, cl_mem C, size_t offC, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasStbmv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, size_t K, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasStbsv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, size_t K, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasStpmv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem AP, size_t offa, cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasStpsv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem A, size_t offa, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasStrmm)(clblasOrder order, clblasSide side, clblasUplo uplo, clblasTranspose transA, clblasDiag diag, size_t M, size_t N, cl_float alpha, const cl_mem A, size_t offA, size_t lda, cl_mem B, size_t offB, size_t ldb, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasStrmv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasStrsm)(clblasOrder order, clblasSide side, clblasUplo uplo, clblasTranspose transA, clblasDiag diag, size_t M, size_t N, cl_float alpha, const cl_mem A, size_t offA, size_t lda, cl_mem B, size_t offB, size_t ldb, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasStrsv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-extern CL_RUNTIME_EXPORT void (*clblasTeardown)();
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZaxpy)(size_t N, cl_double2 alpha, const cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZcopy)(size_t N, const cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZdotc)(size_t N, cl_mem dotProduct, size_t offDP, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZdotu)(size_t N, cl_mem dotProduct, size_t offDP, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZdrot)(size_t N, cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_double C, cl_double S, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZdscal)(size_t N, cl_double alpha, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZgbmv)(clblasOrder order, clblasTranspose trans, size_t M, size_t N, size_t KL, size_t KU, cl_double2 alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem X, size_t offx, int incx, cl_double2 beta, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-extern CL_RUNTIME_EXPORT clblasStatus (*clblasZgemm)(clblasOrder order, clblasTranspose transA, clblasTranspose transB, size_t M, size_t N, size_t K, DoubleComplex alpha, const cl_mem A, size_t offA, size_t lda, const cl_mem B, size_t offB, size_t ldb, DoubleComplex beta, cl_mem C, size_t offC, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZgemv)(clblasOrder order, clblasTranspose transA, size_t M, size_t N, DoubleComplex alpha, const cl_mem A, size_t offA, size_t lda, const cl_mem x, size_t offx, int incx, DoubleComplex beta, cl_mem y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZgerc)(clblasOrder order, size_t M, size_t N, cl_double2 alpha, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem A, size_t offa, size_t lda, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZgeru)(clblasOrder order, size_t M, size_t N, cl_double2 alpha, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem A, size_t offa, size_t lda, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZhbmv)(clblasOrder order, clblasUplo uplo, size_t N, size_t K, cl_double2 alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem X, size_t offx, int incx, cl_double2 beta, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZhemm)(clblasOrder order, clblasSide side, clblasUplo uplo, size_t M, size_t N, cl_double2 alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem B, size_t offb, size_t ldb, cl_double2 beta, cl_mem C, size_t offc, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZhemv)(clblasOrder order, clblasUplo uplo, size_t N, DoubleComplex alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem X, size_t offx, int incx, DoubleComplex beta, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZher)(clblasOrder order, clblasUplo uplo, size_t N, cl_double alpha, const cl_mem X, size_t offx, int incx, cl_mem A, size_t offa, size_t lda, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZher2)(clblasOrder order, clblasUplo uplo, size_t N, cl_double2 alpha, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem A, size_t offa, size_t lda, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZher2k)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, size_t N, size_t K, DoubleComplex alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem B, size_t offb, size_t ldb, cl_double beta, cl_mem C, size_t offc, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZherk)(clblasOrder order, clblasUplo uplo, clblasTranspose transA, size_t N, size_t K, double alpha, const cl_mem A, size_t offa, size_t lda, double beta, cl_mem C, size_t offc, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZhpmv)(clblasOrder order, clblasUplo uplo, size_t N, cl_double2 alpha, const cl_mem AP, size_t offa, const cl_mem X, size_t offx, int incx, cl_double2 beta, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZhpr)(clblasOrder order, clblasUplo uplo, size_t N, cl_double alpha, const cl_mem X, size_t offx, int incx, cl_mem AP, size_t offa, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZhpr2)(clblasOrder order, clblasUplo uplo, size_t N, cl_double2 alpha, const cl_mem X, size_t offx, int incx, const cl_mem Y, size_t offy, int incy, cl_mem AP, size_t offa, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZrotg)(cl_mem CA, size_t offCA, cl_mem CB, size_t offCB, cl_mem C, size_t offC, cl_mem S, size_t offS, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZscal)(size_t N, cl_double2 alpha, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZswap)(size_t N, cl_mem X, size_t offx, int incx, cl_mem Y, size_t offy, int incy, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZsymm)(clblasOrder order, clblasSide side, clblasUplo uplo, size_t M, size_t N, cl_double2 alpha, const cl_mem A, size_t offa, size_t lda, const cl_mem B, size_t offb, size_t ldb, cl_double2 beta, cl_mem C, size_t offc, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZsyr2k)(clblasOrder order, clblasUplo uplo, clblasTranspose transAB, size_t N, size_t K, DoubleComplex alpha, const cl_mem A, size_t offA, size_t lda, const cl_mem B, size_t offB, size_t ldb, DoubleComplex beta, cl_mem C, size_t offC, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZsyrk)(clblasOrder order, clblasUplo uplo, clblasTranspose transA, size_t N, size_t K, DoubleComplex alpha, const cl_mem A, size_t offA, size_t lda, DoubleComplex beta, cl_mem C, size_t offC, size_t ldc, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZtbmv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, size_t K, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZtbsv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, size_t K, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZtpmv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem AP, size_t offa, cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZtpsv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem A, size_t offa, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZtrmm)(clblasOrder order, clblasSide side, clblasUplo uplo, clblasTranspose transA, clblasDiag diag, size_t M, size_t N, DoubleComplex alpha, const cl_mem A, size_t offA, size_t lda, cl_mem B, size_t offB, size_t ldb, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZtrmv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZtrsm)(clblasOrder order, clblasSide side, clblasUplo uplo, clblasTranspose transA, clblasDiag diag, size_t M, size_t N, DoubleComplex alpha, const cl_mem A, size_t offA, size_t lda, cl_mem B, size_t offB, size_t ldb, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasZtrsv)(clblasOrder order, clblasUplo uplo, clblasTranspose trans, clblasDiag diag, size_t N, const cl_mem A, size_t offa, size_t lda, cl_mem X, size_t offx, int incx, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasiCamax)(size_t N, cl_mem iMax, size_t offiMax, const cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasiDamax)(size_t N, cl_mem iMax, size_t offiMax, const cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasiSamax)(size_t N, cl_mem iMax, size_t offiMax, const cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
-//extern CL_RUNTIME_EXPORT clblasStatus (*clblasiZamax)(size_t N, cl_mem iMax, size_t offiMax, const cl_mem X, size_t offx, int incx, cl_mem scratchBuff, cl_uint numCommandQueues, cl_command_queue* commandQueues, cl_uint numEventsInWaitList, const cl_event* eventWaitList, cl_event* events);
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_clfft.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_clfft.hpp
deleted file mode 100644
index dff3b406..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_clfft.hpp
+++ /dev/null
@@ -1,146 +0,0 @@
-//
-// AUTOGENERATED, DO NOT EDIT
-//
-#ifndef OPENCV_CORE_OCL_RUNTIME_CLAMDFFT_HPP
-#error "Invalid usage"
-#endif
-
-// generated by parser_clfft.py
-#define clfftBakePlan clfftBakePlan_
-#define clfftCopyPlan clfftCopyPlan_
-#define clfftCreateDefaultPlan clfftCreateDefaultPlan_
-#define clfftDestroyPlan clfftDestroyPlan_
-#define clfftEnqueueTransform clfftEnqueueTransform_
-#define clfftGetLayout clfftGetLayout_
-#define clfftGetPlanBatchSize clfftGetPlanBatchSize_
-#define clfftGetPlanContext clfftGetPlanContext_
-#define clfftGetPlanDim clfftGetPlanDim_
-#define clfftGetPlanDistance clfftGetPlanDistance_
-#define clfftGetPlanInStride clfftGetPlanInStride_
-#define clfftGetPlanLength clfftGetPlanLength_
-#define clfftGetPlanOutStride clfftGetPlanOutStride_
-#define clfftGetPlanPrecision clfftGetPlanPrecision_
-#define clfftGetPlanScale clfftGetPlanScale_
-#define clfftGetPlanTransposeResult clfftGetPlanTransposeResult_
-#define clfftGetResultLocation clfftGetResultLocation_
-#define clfftGetTmpBufSize clfftGetTmpBufSize_
-#define clfftGetVersion clfftGetVersion_
-#define clfftSetLayout clfftSetLayout_
-#define clfftSetPlanBatchSize clfftSetPlanBatchSize_
-#define clfftSetPlanCallback clfftSetPlanCallback_
-#define clfftSetPlanDim clfftSetPlanDim_
-#define clfftSetPlanDistance clfftSetPlanDistance_
-#define clfftSetPlanInStride clfftSetPlanInStride_
-#define clfftSetPlanLength clfftSetPlanLength_
-#define clfftSetPlanOutStride clfftSetPlanOutStride_
-#define clfftSetPlanPrecision clfftSetPlanPrecision_
-#define clfftSetPlanScale clfftSetPlanScale_
-#define clfftSetPlanTransposeResult clfftSetPlanTransposeResult_
-#define clfftSetResultLocation clfftSetResultLocation_
-#define clfftSetup clfftSetup_
-#define clfftTeardown clfftTeardown_
-
-#include <clFFT.h>
-
-// generated by parser_clfft.py
-#undef clfftBakePlan
-#define clfftBakePlan clfftBakePlan_pfn
-#undef clfftCopyPlan
-//#define clfftCopyPlan clfftCopyPlan_pfn
-#undef clfftCreateDefaultPlan
-#define clfftCreateDefaultPlan clfftCreateDefaultPlan_pfn
-#undef clfftDestroyPlan
-#define clfftDestroyPlan clfftDestroyPlan_pfn
-#undef clfftEnqueueTransform
-#define clfftEnqueueTransform clfftEnqueueTransform_pfn
-#undef clfftGetLayout
-//#define clfftGetLayout clfftGetLayout_pfn
-#undef clfftGetPlanBatchSize
-//#define clfftGetPlanBatchSize clfftGetPlanBatchSize_pfn
-#undef clfftGetPlanContext
-//#define clfftGetPlanContext clfftGetPlanContext_pfn
-#undef clfftGetPlanDim
-//#define clfftGetPlanDim clfftGetPlanDim_pfn
-#undef clfftGetPlanDistance
-//#define clfftGetPlanDistance clfftGetPlanDistance_pfn
-#undef clfftGetPlanInStride
-//#define clfftGetPlanInStride clfftGetPlanInStride_pfn
-#undef clfftGetPlanLength
-//#define clfftGetPlanLength clfftGetPlanLength_pfn
-#undef clfftGetPlanOutStride
-//#define clfftGetPlanOutStride clfftGetPlanOutStride_pfn
-#undef clfftGetPlanPrecision
-//#define clfftGetPlanPrecision clfftGetPlanPrecision_pfn
-#undef clfftGetPlanScale
-//#define clfftGetPlanScale clfftGetPlanScale_pfn
-#undef clfftGetPlanTransposeResult
-//#define clfftGetPlanTransposeResult clfftGetPlanTransposeResult_pfn
-#undef clfftGetResultLocation
-//#define clfftGetResultLocation clfftGetResultLocation_pfn
-#undef clfftGetTmpBufSize
-#define clfftGetTmpBufSize clfftGetTmpBufSize_pfn
-#undef clfftGetVersion
-#define clfftGetVersion clfftGetVersion_pfn
-#undef clfftSetLayout
-#define clfftSetLayout clfftSetLayout_pfn
-#undef clfftSetPlanBatchSize
-#define clfftSetPlanBatchSize clfftSetPlanBatchSize_pfn
-#undef clfftSetPlanCallback
-//#define clfftSetPlanCallback clfftSetPlanCallback_pfn
-#undef clfftSetPlanDim
-//#define clfftSetPlanDim clfftSetPlanDim_pfn
-#undef clfftSetPlanDistance
-#define clfftSetPlanDistance clfftSetPlanDistance_pfn
-#undef clfftSetPlanInStride
-#define clfftSetPlanInStride clfftSetPlanInStride_pfn
-#undef clfftSetPlanLength
-//#define clfftSetPlanLength clfftSetPlanLength_pfn
-#undef clfftSetPlanOutStride
-#define clfftSetPlanOutStride clfftSetPlanOutStride_pfn
-#undef clfftSetPlanPrecision
-#define clfftSetPlanPrecision clfftSetPlanPrecision_pfn
-#undef clfftSetPlanScale
-#define clfftSetPlanScale clfftSetPlanScale_pfn
-#undef clfftSetPlanTransposeResult
-//#define clfftSetPlanTransposeResult clfftSetPlanTransposeResult_pfn
-#undef clfftSetResultLocation
-#define clfftSetResultLocation clfftSetResultLocation_pfn
-#undef clfftSetup
-#define clfftSetup clfftSetup_pfn
-#undef clfftTeardown
-#define clfftTeardown clfftTeardown_pfn
-
-// generated by parser_clfft.py
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftBakePlan)(clfftPlanHandle plHandle, cl_uint numQueues, cl_command_queue* commQueueFFT, void (CL_CALLBACK* pfn_notify) (clfftPlanHandle plHandle, void* user_data), void* user_data);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftCopyPlan)(clfftPlanHandle* out_plHandle, cl_context new_context, clfftPlanHandle in_plHandle);
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftCreateDefaultPlan)(clfftPlanHandle* plHandle, cl_context context, const clfftDim dim, const size_t* clLengths);
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftDestroyPlan)(clfftPlanHandle* plHandle);
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftEnqueueTransform)(clfftPlanHandle plHandle, clfftDirection dir, cl_uint numQueuesAndEvents, cl_command_queue* commQueues, cl_uint numWaitEvents, const cl_event* waitEvents, cl_event* outEvents, cl_mem* inputBuffers, cl_mem* outputBuffers, cl_mem tmpBuffer);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftGetLayout)(const clfftPlanHandle plHandle, clfftLayout* iLayout, clfftLayout* oLayout);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftGetPlanBatchSize)(const clfftPlanHandle plHandle, size_t* batchSize);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftGetPlanContext)(const clfftPlanHandle plHandle, cl_context* context);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftGetPlanDim)(const clfftPlanHandle plHandle, clfftDim* dim, cl_uint* size);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftGetPlanDistance)(const clfftPlanHandle plHandle, size_t* iDist, size_t* oDist);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftGetPlanInStride)(const clfftPlanHandle plHandle, const clfftDim dim, size_t* clStrides);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftGetPlanLength)(const clfftPlanHandle plHandle, const clfftDim dim, size_t* clLengths);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftGetPlanOutStride)(const clfftPlanHandle plHandle, const clfftDim dim, size_t* clStrides);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftGetPlanPrecision)(const clfftPlanHandle plHandle, clfftPrecision* precision);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftGetPlanScale)(const clfftPlanHandle plHandle, clfftDirection dir, cl_float* scale);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftGetPlanTransposeResult)(const clfftPlanHandle plHandle, clfftResultTransposed* transposed);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftGetResultLocation)(const clfftPlanHandle plHandle, clfftResultLocation* placeness);
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftGetTmpBufSize)(const clfftPlanHandle plHandle, size_t* buffersize);
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftGetVersion)(cl_uint* major, cl_uint* minor, cl_uint* patch);
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftSetLayout)(clfftPlanHandle plHandle, clfftLayout iLayout, clfftLayout oLayout);
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftSetPlanBatchSize)(clfftPlanHandle plHandle, size_t batchSize);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftSetPlanCallback)(clfftPlanHandle plHandle, const char* funcName, const char* funcString, int localMemSize, clfftCallbackType callbackType, cl_mem* userdata, int numUserdataBuffers);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftSetPlanDim)(clfftPlanHandle plHandle, const clfftDim dim);
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftSetPlanDistance)(clfftPlanHandle plHandle, size_t iDist, size_t oDist);
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftSetPlanInStride)(clfftPlanHandle plHandle, const clfftDim dim, size_t* clStrides);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftSetPlanLength)(clfftPlanHandle plHandle, const clfftDim dim, const size_t* clLengths);
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftSetPlanOutStride)(clfftPlanHandle plHandle, const clfftDim dim, size_t* clStrides);
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftSetPlanPrecision)(clfftPlanHandle plHandle, clfftPrecision precision);
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftSetPlanScale)(clfftPlanHandle plHandle, clfftDirection dir, cl_float scale);
-//extern CL_RUNTIME_EXPORT clfftStatus (*clfftSetPlanTransposeResult)(clfftPlanHandle plHandle, clfftResultTransposed transposed);
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftSetResultLocation)(clfftPlanHandle plHandle, clfftResultLocation placeness);
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftSetup)(const clfftSetupData* setupData);
-extern CL_RUNTIME_EXPORT clfftStatus (*clfftTeardown)();
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_core.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_core.hpp
deleted file mode 100644
index 28618a1f..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_core.hpp
+++ /dev/null
@@ -1,371 +0,0 @@
-//
-// AUTOGENERATED, DO NOT EDIT
-//
-#ifndef OPENCV_CORE_OCL_RUNTIME_OPENCL_CORE_HPP
-#error "Invalid usage"
-#endif
-
-// generated by parser_cl.py
-#define clBuildProgram clBuildProgram_
-#define clCompileProgram clCompileProgram_
-#define clCreateBuffer clCreateBuffer_
-#define clCreateCommandQueue clCreateCommandQueue_
-#define clCreateContext clCreateContext_
-#define clCreateContextFromType clCreateContextFromType_
-#define clCreateImage clCreateImage_
-#define clCreateImage2D clCreateImage2D_
-#define clCreateImage3D clCreateImage3D_
-#define clCreateKernel clCreateKernel_
-#define clCreateKernelsInProgram clCreateKernelsInProgram_
-#define clCreateProgramWithBinary clCreateProgramWithBinary_
-#define clCreateProgramWithBuiltInKernels clCreateProgramWithBuiltInKernels_
-#define clCreateProgramWithSource clCreateProgramWithSource_
-#define clCreateSampler clCreateSampler_
-#define clCreateSubBuffer clCreateSubBuffer_
-#define clCreateSubDevices clCreateSubDevices_
-#define clCreateUserEvent clCreateUserEvent_
-#define clEnqueueBarrier clEnqueueBarrier_
-#define clEnqueueBarrierWithWaitList clEnqueueBarrierWithWaitList_
-#define clEnqueueCopyBuffer clEnqueueCopyBuffer_
-#define clEnqueueCopyBufferRect clEnqueueCopyBufferRect_
-#define clEnqueueCopyBufferToImage clEnqueueCopyBufferToImage_
-#define clEnqueueCopyImage clEnqueueCopyImage_
-#define clEnqueueCopyImageToBuffer clEnqueueCopyImageToBuffer_
-#define clEnqueueFillBuffer clEnqueueFillBuffer_
-#define clEnqueueFillImage clEnqueueFillImage_
-#define clEnqueueMapBuffer clEnqueueMapBuffer_
-#define clEnqueueMapImage clEnqueueMapImage_
-#define clEnqueueMarker clEnqueueMarker_
-#define clEnqueueMarkerWithWaitList clEnqueueMarkerWithWaitList_
-#define clEnqueueMigrateMemObjects clEnqueueMigrateMemObjects_
-#define clEnqueueNDRangeKernel clEnqueueNDRangeKernel_
-#define clEnqueueNativeKernel clEnqueueNativeKernel_
-#define clEnqueueReadBuffer clEnqueueReadBuffer_
-#define clEnqueueReadBufferRect clEnqueueReadBufferRect_
-#define clEnqueueReadImage clEnqueueReadImage_
-#define clEnqueueTask clEnqueueTask_
-#define clEnqueueUnmapMemObject clEnqueueUnmapMemObject_
-#define clEnqueueWaitForEvents clEnqueueWaitForEvents_
-#define clEnqueueWriteBuffer clEnqueueWriteBuffer_
-#define clEnqueueWriteBufferRect clEnqueueWriteBufferRect_
-#define clEnqueueWriteImage clEnqueueWriteImage_
-#define clFinish clFinish_
-#define clFlush clFlush_
-#define clGetCommandQueueInfo clGetCommandQueueInfo_
-#define clGetContextInfo clGetContextInfo_
-#define clGetDeviceIDs clGetDeviceIDs_
-#define clGetDeviceInfo clGetDeviceInfo_
-#define clGetEventInfo clGetEventInfo_
-#define clGetEventProfilingInfo clGetEventProfilingInfo_
-#define clGetExtensionFunctionAddress clGetExtensionFunctionAddress_
-#define clGetExtensionFunctionAddressForPlatform clGetExtensionFunctionAddressForPlatform_
-#define clGetImageInfo clGetImageInfo_
-#define clGetKernelArgInfo clGetKernelArgInfo_
-#define clGetKernelInfo clGetKernelInfo_
-#define clGetKernelWorkGroupInfo clGetKernelWorkGroupInfo_
-#define clGetMemObjectInfo clGetMemObjectInfo_
-#define clGetPlatformIDs clGetPlatformIDs_
-#define clGetPlatformInfo clGetPlatformInfo_
-#define clGetProgramBuildInfo clGetProgramBuildInfo_
-#define clGetProgramInfo clGetProgramInfo_
-#define clGetSamplerInfo clGetSamplerInfo_
-#define clGetSupportedImageFormats clGetSupportedImageFormats_
-#define clLinkProgram clLinkProgram_
-#define clReleaseCommandQueue clReleaseCommandQueue_
-#define clReleaseContext clReleaseContext_
-#define clReleaseDevice clReleaseDevice_
-#define clReleaseEvent clReleaseEvent_
-#define clReleaseKernel clReleaseKernel_
-#define clReleaseMemObject clReleaseMemObject_
-#define clReleaseProgram clReleaseProgram_
-#define clReleaseSampler clReleaseSampler_
-#define clRetainCommandQueue clRetainCommandQueue_
-#define clRetainContext clRetainContext_
-#define clRetainDevice clRetainDevice_
-#define clRetainEvent clRetainEvent_
-#define clRetainKernel clRetainKernel_
-#define clRetainMemObject clRetainMemObject_
-#define clRetainProgram clRetainProgram_
-#define clRetainSampler clRetainSampler_
-#define clSetEventCallback clSetEventCallback_
-#define clSetKernelArg clSetKernelArg_
-#define clSetMemObjectDestructorCallback clSetMemObjectDestructorCallback_
-#define clSetUserEventStatus clSetUserEventStatus_
-#define clUnloadCompiler clUnloadCompiler_
-#define clUnloadPlatformCompiler clUnloadPlatformCompiler_
-#define clWaitForEvents clWaitForEvents_
-
-#if defined __APPLE__
-#define CL_SILENCE_DEPRECATION
-#include <OpenCL/cl.h>
-#else
-#include <CL/cl.h>
-#endif
-
-// generated by parser_cl.py
-#undef clBuildProgram
-#define clBuildProgram clBuildProgram_pfn
-#undef clCompileProgram
-#define clCompileProgram clCompileProgram_pfn
-#undef clCreateBuffer
-#define clCreateBuffer clCreateBuffer_pfn
-#undef clCreateCommandQueue
-#define clCreateCommandQueue clCreateCommandQueue_pfn
-#undef clCreateContext
-#define clCreateContext clCreateContext_pfn
-#undef clCreateContextFromType
-#define clCreateContextFromType clCreateContextFromType_pfn
-#undef clCreateImage
-#define clCreateImage clCreateImage_pfn
-#undef clCreateImage2D
-#define clCreateImage2D clCreateImage2D_pfn
-#undef clCreateImage3D
-#define clCreateImage3D clCreateImage3D_pfn
-#undef clCreateKernel
-#define clCreateKernel clCreateKernel_pfn
-#undef clCreateKernelsInProgram
-#define clCreateKernelsInProgram clCreateKernelsInProgram_pfn
-#undef clCreateProgramWithBinary
-#define clCreateProgramWithBinary clCreateProgramWithBinary_pfn
-#undef clCreateProgramWithBuiltInKernels
-#define clCreateProgramWithBuiltInKernels clCreateProgramWithBuiltInKernels_pfn
-#undef clCreateProgramWithSource
-#define clCreateProgramWithSource clCreateProgramWithSource_pfn
-#undef clCreateSampler
-#define clCreateSampler clCreateSampler_pfn
-#undef clCreateSubBuffer
-#define clCreateSubBuffer clCreateSubBuffer_pfn
-#undef clCreateSubDevices
-#define clCreateSubDevices clCreateSubDevices_pfn
-#undef clCreateUserEvent
-#define clCreateUserEvent clCreateUserEvent_pfn
-#undef clEnqueueBarrier
-#define clEnqueueBarrier clEnqueueBarrier_pfn
-#undef clEnqueueBarrierWithWaitList
-#define clEnqueueBarrierWithWaitList clEnqueueBarrierWithWaitList_pfn
-#undef clEnqueueCopyBuffer
-#define clEnqueueCopyBuffer clEnqueueCopyBuffer_pfn
-#undef clEnqueueCopyBufferRect
-#define clEnqueueCopyBufferRect clEnqueueCopyBufferRect_pfn
-#undef clEnqueueCopyBufferToImage
-#define clEnqueueCopyBufferToImage clEnqueueCopyBufferToImage_pfn
-#undef clEnqueueCopyImage
-#define clEnqueueCopyImage clEnqueueCopyImage_pfn
-#undef clEnqueueCopyImageToBuffer
-#define clEnqueueCopyImageToBuffer clEnqueueCopyImageToBuffer_pfn
-#undef clEnqueueFillBuffer
-#define clEnqueueFillBuffer clEnqueueFillBuffer_pfn
-#undef clEnqueueFillImage
-#define clEnqueueFillImage clEnqueueFillImage_pfn
-#undef clEnqueueMapBuffer
-#define clEnqueueMapBuffer clEnqueueMapBuffer_pfn
-#undef clEnqueueMapImage
-#define clEnqueueMapImage clEnqueueMapImage_pfn
-#undef clEnqueueMarker
-#define clEnqueueMarker clEnqueueMarker_pfn
-#undef clEnqueueMarkerWithWaitList
-#define clEnqueueMarkerWithWaitList clEnqueueMarkerWithWaitList_pfn
-#undef clEnqueueMigrateMemObjects
-#define clEnqueueMigrateMemObjects clEnqueueMigrateMemObjects_pfn
-#undef clEnqueueNDRangeKernel
-#define clEnqueueNDRangeKernel clEnqueueNDRangeKernel_pfn
-#undef clEnqueueNativeKernel
-#define clEnqueueNativeKernel clEnqueueNativeKernel_pfn
-#undef clEnqueueReadBuffer
-#define clEnqueueReadBuffer clEnqueueReadBuffer_pfn
-#undef clEnqueueReadBufferRect
-#define clEnqueueReadBufferRect clEnqueueReadBufferRect_pfn
-#undef clEnqueueReadImage
-#define clEnqueueReadImage clEnqueueReadImage_pfn
-#undef clEnqueueTask
-#define clEnqueueTask clEnqueueTask_pfn
-#undef clEnqueueUnmapMemObject
-#define clEnqueueUnmapMemObject clEnqueueUnmapMemObject_pfn
-#undef clEnqueueWaitForEvents
-#define clEnqueueWaitForEvents clEnqueueWaitForEvents_pfn
-#undef clEnqueueWriteBuffer
-#define clEnqueueWriteBuffer clEnqueueWriteBuffer_pfn
-#undef clEnqueueWriteBufferRect
-#define clEnqueueWriteBufferRect clEnqueueWriteBufferRect_pfn
-#undef clEnqueueWriteImage
-#define clEnqueueWriteImage clEnqueueWriteImage_pfn
-#undef clFinish
-#define clFinish clFinish_pfn
-#undef clFlush
-#define clFlush clFlush_pfn
-#undef clGetCommandQueueInfo
-#define clGetCommandQueueInfo clGetCommandQueueInfo_pfn
-#undef clGetContextInfo
-#define clGetContextInfo clGetContextInfo_pfn
-#undef clGetDeviceIDs
-#define clGetDeviceIDs clGetDeviceIDs_pfn
-#undef clGetDeviceInfo
-#define clGetDeviceInfo clGetDeviceInfo_pfn
-#undef clGetEventInfo
-#define clGetEventInfo clGetEventInfo_pfn
-#undef clGetEventProfilingInfo
-#define clGetEventProfilingInfo clGetEventProfilingInfo_pfn
-#undef clGetExtensionFunctionAddress
-#define clGetExtensionFunctionAddress clGetExtensionFunctionAddress_pfn
-#undef clGetExtensionFunctionAddressForPlatform
-#define clGetExtensionFunctionAddressForPlatform clGetExtensionFunctionAddressForPlatform_pfn
-#undef clGetImageInfo
-#define clGetImageInfo clGetImageInfo_pfn
-#undef clGetKernelArgInfo
-#define clGetKernelArgInfo clGetKernelArgInfo_pfn
-#undef clGetKernelInfo
-#define clGetKernelInfo clGetKernelInfo_pfn
-#undef clGetKernelWorkGroupInfo
-#define clGetKernelWorkGroupInfo clGetKernelWorkGroupInfo_pfn
-#undef clGetMemObjectInfo
-#define clGetMemObjectInfo clGetMemObjectInfo_pfn
-#undef clGetPlatformIDs
-#define clGetPlatformIDs clGetPlatformIDs_pfn
-#undef clGetPlatformInfo
-#define clGetPlatformInfo clGetPlatformInfo_pfn
-#undef clGetProgramBuildInfo
-#define clGetProgramBuildInfo clGetProgramBuildInfo_pfn
-#undef clGetProgramInfo
-#define clGetProgramInfo clGetProgramInfo_pfn
-#undef clGetSamplerInfo
-#define clGetSamplerInfo clGetSamplerInfo_pfn
-#undef clGetSupportedImageFormats
-#define clGetSupportedImageFormats clGetSupportedImageFormats_pfn
-#undef clLinkProgram
-#define clLinkProgram clLinkProgram_pfn
-#undef clReleaseCommandQueue
-#define clReleaseCommandQueue clReleaseCommandQueue_pfn
-#undef clReleaseContext
-#define clReleaseContext clReleaseContext_pfn
-#undef clReleaseDevice
-#define clReleaseDevice clReleaseDevice_pfn
-#undef clReleaseEvent
-#define clReleaseEvent clReleaseEvent_pfn
-#undef clReleaseKernel
-#define clReleaseKernel clReleaseKernel_pfn
-#undef clReleaseMemObject
-#define clReleaseMemObject clReleaseMemObject_pfn
-#undef clReleaseProgram
-#define clReleaseProgram clReleaseProgram_pfn
-#undef clReleaseSampler
-#define clReleaseSampler clReleaseSampler_pfn
-#undef clRetainCommandQueue
-#define clRetainCommandQueue clRetainCommandQueue_pfn
-#undef clRetainContext
-#define clRetainContext clRetainContext_pfn
-#undef clRetainDevice
-#define clRetainDevice clRetainDevice_pfn
-#undef clRetainEvent
-#define clRetainEvent clRetainEvent_pfn
-#undef clRetainKernel
-#define clRetainKernel clRetainKernel_pfn
-#undef clRetainMemObject
-#define clRetainMemObject clRetainMemObject_pfn
-#undef clRetainProgram
-#define clRetainProgram clRetainProgram_pfn
-#undef clRetainSampler
-#define clRetainSampler clRetainSampler_pfn
-#undef clSetEventCallback
-#define clSetEventCallback clSetEventCallback_pfn
-#undef clSetKernelArg
-#define clSetKernelArg clSetKernelArg_pfn
-#undef clSetMemObjectDestructorCallback
-#define clSetMemObjectDestructorCallback clSetMemObjectDestructorCallback_pfn
-#undef clSetUserEventStatus
-#define clSetUserEventStatus clSetUserEventStatus_pfn
-#undef clUnloadCompiler
-#define clUnloadCompiler clUnloadCompiler_pfn
-#undef clUnloadPlatformCompiler
-#define clUnloadPlatformCompiler clUnloadPlatformCompiler_pfn
-#undef clWaitForEvents
-#define clWaitForEvents clWaitForEvents_pfn
-
-// generated by parser_cl.py
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clBuildProgram)(cl_program, cl_uint, const cl_device_id*, const char*, void (CL_CALLBACK*) (cl_program, void*), void*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clCompileProgram)(cl_program, cl_uint, const cl_device_id*, const char*, cl_uint, const cl_program*, const char**, void (CL_CALLBACK*) (cl_program, void*), void*);
-extern CL_RUNTIME_EXPORT cl_mem (CL_API_CALL*clCreateBuffer)(cl_context, cl_mem_flags, size_t, void*, cl_int*);
-extern CL_RUNTIME_EXPORT cl_command_queue (CL_API_CALL*clCreateCommandQueue)(cl_context, cl_device_id, cl_command_queue_properties, cl_int*);
-extern CL_RUNTIME_EXPORT cl_context (CL_API_CALL*clCreateContext)(const cl_context_properties*, cl_uint, const cl_device_id*, void (CL_CALLBACK*) (const char*, const void*, size_t, void*), void*, cl_int*);
-extern CL_RUNTIME_EXPORT cl_context (CL_API_CALL*clCreateContextFromType)(const cl_context_properties*, cl_device_type, void (CL_CALLBACK*) (const char*, const void*, size_t, void*), void*, cl_int*);
-extern CL_RUNTIME_EXPORT cl_mem (CL_API_CALL*clCreateImage)(cl_context, cl_mem_flags, const cl_image_format*, const cl_image_desc*, void*, cl_int*);
-extern CL_RUNTIME_EXPORT cl_mem (CL_API_CALL*clCreateImage2D)(cl_context, cl_mem_flags, const cl_image_format*, size_t, size_t, size_t, void*, cl_int*);
-extern CL_RUNTIME_EXPORT cl_mem (CL_API_CALL*clCreateImage3D)(cl_context, cl_mem_flags, const cl_image_format*, size_t, size_t, size_t, size_t, size_t, void*, cl_int*);
-extern CL_RUNTIME_EXPORT cl_kernel (CL_API_CALL*clCreateKernel)(cl_program, const char*, cl_int*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clCreateKernelsInProgram)(cl_program, cl_uint, cl_kernel*, cl_uint*);
-extern CL_RUNTIME_EXPORT cl_program (CL_API_CALL*clCreateProgramWithBinary)(cl_context, cl_uint, const cl_device_id*, const size_t*, const unsigned char**, cl_int*, cl_int*);
-extern CL_RUNTIME_EXPORT cl_program (CL_API_CALL*clCreateProgramWithBuiltInKernels)(cl_context, cl_uint, const cl_device_id*, const char*, cl_int*);
-extern CL_RUNTIME_EXPORT cl_program (CL_API_CALL*clCreateProgramWithSource)(cl_context, cl_uint, const char**, const size_t*, cl_int*);
-extern CL_RUNTIME_EXPORT cl_sampler (CL_API_CALL*clCreateSampler)(cl_context, cl_bool, cl_addressing_mode, cl_filter_mode, cl_int*);
-extern CL_RUNTIME_EXPORT cl_mem (CL_API_CALL*clCreateSubBuffer)(cl_mem, cl_mem_flags, cl_buffer_create_type, const void*, cl_int*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clCreateSubDevices)(cl_device_id, const cl_device_partition_property*, cl_uint, cl_device_id*, cl_uint*);
-extern CL_RUNTIME_EXPORT cl_event (CL_API_CALL*clCreateUserEvent)(cl_context, cl_int*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueBarrier)(cl_command_queue);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueBarrierWithWaitList)(cl_command_queue, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueCopyBuffer)(cl_command_queue, cl_mem, cl_mem, size_t, size_t, size_t, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueCopyBufferRect)(cl_command_queue, cl_mem, cl_mem, const size_t*, const size_t*, const size_t*, size_t, size_t, size_t, size_t, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueCopyBufferToImage)(cl_command_queue, cl_mem, cl_mem, size_t, const size_t*, const size_t*, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueCopyImage)(cl_command_queue, cl_mem, cl_mem, const size_t*, const size_t*, const size_t*, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueCopyImageToBuffer)(cl_command_queue, cl_mem, cl_mem, const size_t*, const size_t*, size_t, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueFillBuffer)(cl_command_queue, cl_mem, const void*, size_t, size_t, size_t, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueFillImage)(cl_command_queue, cl_mem, const void*, const size_t*, const size_t*, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT void* (CL_API_CALL*clEnqueueMapBuffer)(cl_command_queue, cl_mem, cl_bool, cl_map_flags, size_t, size_t, cl_uint, const cl_event*, cl_event*, cl_int*);
-extern CL_RUNTIME_EXPORT void* (CL_API_CALL*clEnqueueMapImage)(cl_command_queue, cl_mem, cl_bool, cl_map_flags, const size_t*, const size_t*, size_t*, size_t*, cl_uint, const cl_event*, cl_event*, cl_int*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueMarker)(cl_command_queue, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueMarkerWithWaitList)(cl_command_queue, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueMigrateMemObjects)(cl_command_queue, cl_uint, const cl_mem*, cl_mem_migration_flags, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueNDRangeKernel)(cl_command_queue, cl_kernel, cl_uint, const size_t*, const size_t*, const size_t*, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueNativeKernel)(cl_command_queue, void (CL_CALLBACK*) (void*), void*, size_t, cl_uint, const cl_mem*, const void**, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueReadBuffer)(cl_command_queue, cl_mem, cl_bool, size_t, size_t, void*, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueReadBufferRect)(cl_command_queue, cl_mem, cl_bool, const size_t*, const size_t*, const size_t*, size_t, size_t, size_t, size_t, void*, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueReadImage)(cl_command_queue, cl_mem, cl_bool, const size_t*, const size_t*, size_t, size_t, void*, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueTask)(cl_command_queue, cl_kernel, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueUnmapMemObject)(cl_command_queue, cl_mem, void*, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueWaitForEvents)(cl_command_queue, cl_uint, const cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueWriteBuffer)(cl_command_queue, cl_mem, cl_bool, size_t, size_t, const void*, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueWriteBufferRect)(cl_command_queue, cl_mem, cl_bool, const size_t*, const size_t*, const size_t*, size_t, size_t, size_t, size_t, const void*, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueWriteImage)(cl_command_queue, cl_mem, cl_bool, const size_t*, const size_t*, size_t, size_t, const void*, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clFinish)(cl_command_queue);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clFlush)(cl_command_queue);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetCommandQueueInfo)(cl_command_queue, cl_command_queue_info, size_t, void*, size_t*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetContextInfo)(cl_context, cl_context_info, size_t, void*, size_t*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetDeviceIDs)(cl_platform_id, cl_device_type, cl_uint, cl_device_id*, cl_uint*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetDeviceInfo)(cl_device_id, cl_device_info, size_t, void*, size_t*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetEventInfo)(cl_event, cl_event_info, size_t, void*, size_t*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetEventProfilingInfo)(cl_event, cl_profiling_info, size_t, void*, size_t*);
-extern CL_RUNTIME_EXPORT void* (CL_API_CALL*clGetExtensionFunctionAddress)(const char*);
-extern CL_RUNTIME_EXPORT void* (CL_API_CALL*clGetExtensionFunctionAddressForPlatform)(cl_platform_id, const char*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetImageInfo)(cl_mem, cl_image_info, size_t, void*, size_t*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetKernelArgInfo)(cl_kernel, cl_uint, cl_kernel_arg_info, size_t, void*, size_t*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetKernelInfo)(cl_kernel, cl_kernel_info, size_t, void*, size_t*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetKernelWorkGroupInfo)(cl_kernel, cl_device_id, cl_kernel_work_group_info, size_t, void*, size_t*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetMemObjectInfo)(cl_mem, cl_mem_info, size_t, void*, size_t*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetPlatformIDs)(cl_uint, cl_platform_id*, cl_uint*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetPlatformInfo)(cl_platform_id, cl_platform_info, size_t, void*, size_t*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetProgramBuildInfo)(cl_program, cl_device_id, cl_program_build_info, size_t, void*, size_t*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetProgramInfo)(cl_program, cl_program_info, size_t, void*, size_t*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetSamplerInfo)(cl_sampler, cl_sampler_info, size_t, void*, size_t*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetSupportedImageFormats)(cl_context, cl_mem_flags, cl_mem_object_type, cl_uint, cl_image_format*, cl_uint*);
-extern CL_RUNTIME_EXPORT cl_program (CL_API_CALL*clLinkProgram)(cl_context, cl_uint, const cl_device_id*, const char*, cl_uint, const cl_program*, void (CL_CALLBACK*) (cl_program, void*), void*, cl_int*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clReleaseCommandQueue)(cl_command_queue);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clReleaseContext)(cl_context);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clReleaseDevice)(cl_device_id);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clReleaseEvent)(cl_event);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clReleaseKernel)(cl_kernel);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clReleaseMemObject)(cl_mem);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clReleaseProgram)(cl_program);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clReleaseSampler)(cl_sampler);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clRetainCommandQueue)(cl_command_queue);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clRetainContext)(cl_context);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clRetainDevice)(cl_device_id);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clRetainEvent)(cl_event);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clRetainKernel)(cl_kernel);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clRetainMemObject)(cl_mem);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clRetainProgram)(cl_program);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clRetainSampler)(cl_sampler);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clSetEventCallback)(cl_event, cl_int, void (CL_CALLBACK*) (cl_event, cl_int, void*), void*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clSetKernelArg)(cl_kernel, cl_uint, size_t, const void*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clSetMemObjectDestructorCallback)(cl_mem, void (CL_CALLBACK*) (cl_mem, void*), void*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clSetUserEventStatus)(cl_event, cl_int);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clUnloadCompiler)();
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clUnloadPlatformCompiler)(cl_platform_id);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clWaitForEvents)(cl_uint, const cl_event*);
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_core_wrappers.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_core_wrappers.hpp
deleted file mode 100644
index 216b22b8..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_core_wrappers.hpp
+++ /dev/null
@@ -1,272 +0,0 @@
-//
-// AUTOGENERATED, DO NOT EDIT
-//
-#ifndef OPENCV_CORE_OCL_RUNTIME_OPENCL_WRAPPERS_HPP
-#error "Invalid usage"
-#endif
-
-// generated by parser_cl.py
-#undef clBuildProgram
-#define clBuildProgram clBuildProgram_fn
-inline cl_int clBuildProgram(cl_program p0, cl_uint p1, const cl_device_id* p2, const char* p3, void (CL_CALLBACK*p4) (cl_program, void*), void* p5) { return clBuildProgram_pfn(p0, p1, p2, p3, p4, p5); }
-#undef clCompileProgram
-#define clCompileProgram clCompileProgram_fn
-inline cl_int clCompileProgram(cl_program p0, cl_uint p1, const cl_device_id* p2, const char* p3, cl_uint p4, const cl_program* p5, const char** p6, void (CL_CALLBACK*p7) (cl_program, void*), void* p8) { return clCompileProgram_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8); }
-#undef clCreateBuffer
-#define clCreateBuffer clCreateBuffer_fn
-inline cl_mem clCreateBuffer(cl_context p0, cl_mem_flags p1, size_t p2, void* p3, cl_int* p4) { return clCreateBuffer_pfn(p0, p1, p2, p3, p4); }
-#undef clCreateCommandQueue
-#define clCreateCommandQueue clCreateCommandQueue_fn
-inline cl_command_queue clCreateCommandQueue(cl_context p0, cl_device_id p1, cl_command_queue_properties p2, cl_int* p3) { return clCreateCommandQueue_pfn(p0, p1, p2, p3); }
-#undef clCreateContext
-#define clCreateContext clCreateContext_fn
-inline cl_context clCreateContext(const cl_context_properties* p0, cl_uint p1, const cl_device_id* p2, void (CL_CALLBACK*p3) (const char*, const void*, size_t, void*), void* p4, cl_int* p5) { return clCreateContext_pfn(p0, p1, p2, p3, p4, p5); }
-#undef clCreateContextFromType
-#define clCreateContextFromType clCreateContextFromType_fn
-inline cl_context clCreateContextFromType(const cl_context_properties* p0, cl_device_type p1, void (CL_CALLBACK*p2) (const char*, const void*, size_t, void*), void* p3, cl_int* p4) { return clCreateContextFromType_pfn(p0, p1, p2, p3, p4); }
-#undef clCreateImage
-#define clCreateImage clCreateImage_fn
-inline cl_mem clCreateImage(cl_context p0, cl_mem_flags p1, const cl_image_format* p2, const cl_image_desc* p3, void* p4, cl_int* p5) { return clCreateImage_pfn(p0, p1, p2, p3, p4, p5); }
-#undef clCreateImage2D
-#define clCreateImage2D clCreateImage2D_fn
-inline cl_mem clCreateImage2D(cl_context p0, cl_mem_flags p1, const cl_image_format* p2, size_t p3, size_t p4, size_t p5, void* p6, cl_int* p7) { return clCreateImage2D_pfn(p0, p1, p2, p3, p4, p5, p6, p7); }
-#undef clCreateImage3D
-#define clCreateImage3D clCreateImage3D_fn
-inline cl_mem clCreateImage3D(cl_context p0, cl_mem_flags p1, const cl_image_format* p2, size_t p3, size_t p4, size_t p5, size_t p6, size_t p7, void* p8, cl_int* p9) { return clCreateImage3D_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8, p9); }
-#undef clCreateKernel
-#define clCreateKernel clCreateKernel_fn
-inline cl_kernel clCreateKernel(cl_program p0, const char* p1, cl_int* p2) { return clCreateKernel_pfn(p0, p1, p2); }
-#undef clCreateKernelsInProgram
-#define clCreateKernelsInProgram clCreateKernelsInProgram_fn
-inline cl_int clCreateKernelsInProgram(cl_program p0, cl_uint p1, cl_kernel* p2, cl_uint* p3) { return clCreateKernelsInProgram_pfn(p0, p1, p2, p3); }
-#undef clCreateProgramWithBinary
-#define clCreateProgramWithBinary clCreateProgramWithBinary_fn
-inline cl_program clCreateProgramWithBinary(cl_context p0, cl_uint p1, const cl_device_id* p2, const size_t* p3, const unsigned char** p4, cl_int* p5, cl_int* p6) { return clCreateProgramWithBinary_pfn(p0, p1, p2, p3, p4, p5, p6); }
-#undef clCreateProgramWithBuiltInKernels
-#define clCreateProgramWithBuiltInKernels clCreateProgramWithBuiltInKernels_fn
-inline cl_program clCreateProgramWithBuiltInKernels(cl_context p0, cl_uint p1, const cl_device_id* p2, const char* p3, cl_int* p4) { return clCreateProgramWithBuiltInKernels_pfn(p0, p1, p2, p3, p4); }
-#undef clCreateProgramWithSource
-#define clCreateProgramWithSource clCreateProgramWithSource_fn
-inline cl_program clCreateProgramWithSource(cl_context p0, cl_uint p1, const char** p2, const size_t* p3, cl_int* p4) { return clCreateProgramWithSource_pfn(p0, p1, p2, p3, p4); }
-#undef clCreateSampler
-#define clCreateSampler clCreateSampler_fn
-inline cl_sampler clCreateSampler(cl_context p0, cl_bool p1, cl_addressing_mode p2, cl_filter_mode p3, cl_int* p4) { return clCreateSampler_pfn(p0, p1, p2, p3, p4); }
-#undef clCreateSubBuffer
-#define clCreateSubBuffer clCreateSubBuffer_fn
-inline cl_mem clCreateSubBuffer(cl_mem p0, cl_mem_flags p1, cl_buffer_create_type p2, const void* p3, cl_int* p4) { return clCreateSubBuffer_pfn(p0, p1, p2, p3, p4); }
-#undef clCreateSubDevices
-#define clCreateSubDevices clCreateSubDevices_fn
-inline cl_int clCreateSubDevices(cl_device_id p0, const cl_device_partition_property* p1, cl_uint p2, cl_device_id* p3, cl_uint* p4) { return clCreateSubDevices_pfn(p0, p1, p2, p3, p4); }
-#undef clCreateUserEvent
-#define clCreateUserEvent clCreateUserEvent_fn
-inline cl_event clCreateUserEvent(cl_context p0, cl_int* p1) { return clCreateUserEvent_pfn(p0, p1); }
-#undef clEnqueueBarrier
-#define clEnqueueBarrier clEnqueueBarrier_fn
-inline cl_int clEnqueueBarrier(cl_command_queue p0) { return clEnqueueBarrier_pfn(p0); }
-#undef clEnqueueBarrierWithWaitList
-#define clEnqueueBarrierWithWaitList clEnqueueBarrierWithWaitList_fn
-inline cl_int clEnqueueBarrierWithWaitList(cl_command_queue p0, cl_uint p1, const cl_event* p2, cl_event* p3) { return clEnqueueBarrierWithWaitList_pfn(p0, p1, p2, p3); }
-#undef clEnqueueCopyBuffer
-#define clEnqueueCopyBuffer clEnqueueCopyBuffer_fn
-inline cl_int clEnqueueCopyBuffer(cl_command_queue p0, cl_mem p1, cl_mem p2, size_t p3, size_t p4, size_t p5, cl_uint p6, const cl_event* p7, cl_event* p8) { return clEnqueueCopyBuffer_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8); }
-#undef clEnqueueCopyBufferRect
-#define clEnqueueCopyBufferRect clEnqueueCopyBufferRect_fn
-inline cl_int clEnqueueCopyBufferRect(cl_command_queue p0, cl_mem p1, cl_mem p2, const size_t* p3, const size_t* p4, const size_t* p5, size_t p6, size_t p7, size_t p8, size_t p9, cl_uint p10, const cl_event* p11, cl_event* p12) { return clEnqueueCopyBufferRect_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, p10, p11, p12); }
-#undef clEnqueueCopyBufferToImage
-#define clEnqueueCopyBufferToImage clEnqueueCopyBufferToImage_fn
-inline cl_int clEnqueueCopyBufferToImage(cl_command_queue p0, cl_mem p1, cl_mem p2, size_t p3, const size_t* p4, const size_t* p5, cl_uint p6, const cl_event* p7, cl_event* p8) { return clEnqueueCopyBufferToImage_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8); }
-#undef clEnqueueCopyImage
-#define clEnqueueCopyImage clEnqueueCopyImage_fn
-inline cl_int clEnqueueCopyImage(cl_command_queue p0, cl_mem p1, cl_mem p2, const size_t* p3, const size_t* p4, const size_t* p5, cl_uint p6, const cl_event* p7, cl_event* p8) { return clEnqueueCopyImage_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8); }
-#undef clEnqueueCopyImageToBuffer
-#define clEnqueueCopyImageToBuffer clEnqueueCopyImageToBuffer_fn
-inline cl_int clEnqueueCopyImageToBuffer(cl_command_queue p0, cl_mem p1, cl_mem p2, const size_t* p3, const size_t* p4, size_t p5, cl_uint p6, const cl_event* p7, cl_event* p8) { return clEnqueueCopyImageToBuffer_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8); }
-#undef clEnqueueFillBuffer
-#define clEnqueueFillBuffer clEnqueueFillBuffer_fn
-inline cl_int clEnqueueFillBuffer(cl_command_queue p0, cl_mem p1, const void* p2, size_t p3, size_t p4, size_t p5, cl_uint p6, const cl_event* p7, cl_event* p8) { return clEnqueueFillBuffer_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8); }
-#undef clEnqueueFillImage
-#define clEnqueueFillImage clEnqueueFillImage_fn
-inline cl_int clEnqueueFillImage(cl_command_queue p0, cl_mem p1, const void* p2, const size_t* p3, const size_t* p4, cl_uint p5, const cl_event* p6, cl_event* p7) { return clEnqueueFillImage_pfn(p0, p1, p2, p3, p4, p5, p6, p7); }
-#undef clEnqueueMapBuffer
-#define clEnqueueMapBuffer clEnqueueMapBuffer_fn
-inline void* clEnqueueMapBuffer(cl_command_queue p0, cl_mem p1, cl_bool p2, cl_map_flags p3, size_t p4, size_t p5, cl_uint p6, const cl_event* p7, cl_event* p8, cl_int* p9) { return clEnqueueMapBuffer_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8, p9); }
-#undef clEnqueueMapImage
-#define clEnqueueMapImage clEnqueueMapImage_fn
-inline void* clEnqueueMapImage(cl_command_queue p0, cl_mem p1, cl_bool p2, cl_map_flags p3, const size_t* p4, const size_t* p5, size_t* p6, size_t* p7, cl_uint p8, const cl_event* p9, cl_event* p10, cl_int* p11) { return clEnqueueMapImage_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, p10, p11); }
-#undef clEnqueueMarker
-#define clEnqueueMarker clEnqueueMarker_fn
-inline cl_int clEnqueueMarker(cl_command_queue p0, cl_event* p1) { return clEnqueueMarker_pfn(p0, p1); }
-#undef clEnqueueMarkerWithWaitList
-#define clEnqueueMarkerWithWaitList clEnqueueMarkerWithWaitList_fn
-inline cl_int clEnqueueMarkerWithWaitList(cl_command_queue p0, cl_uint p1, const cl_event* p2, cl_event* p3) { return clEnqueueMarkerWithWaitList_pfn(p0, p1, p2, p3); }
-#undef clEnqueueMigrateMemObjects
-#define clEnqueueMigrateMemObjects clEnqueueMigrateMemObjects_fn
-inline cl_int clEnqueueMigrateMemObjects(cl_command_queue p0, cl_uint p1, const cl_mem* p2, cl_mem_migration_flags p3, cl_uint p4, const cl_event* p5, cl_event* p6) { return clEnqueueMigrateMemObjects_pfn(p0, p1, p2, p3, p4, p5, p6); }
-#undef clEnqueueNDRangeKernel
-#define clEnqueueNDRangeKernel clEnqueueNDRangeKernel_fn
-inline cl_int clEnqueueNDRangeKernel(cl_command_queue p0, cl_kernel p1, cl_uint p2, const size_t* p3, const size_t* p4, const size_t* p5, cl_uint p6, const cl_event* p7, cl_event* p8) { return clEnqueueNDRangeKernel_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8); }
-#undef clEnqueueNativeKernel
-#define clEnqueueNativeKernel clEnqueueNativeKernel_fn
-inline cl_int clEnqueueNativeKernel(cl_command_queue p0, void (CL_CALLBACK*p1) (void*), void* p2, size_t p3, cl_uint p4, const cl_mem* p5, const void** p6, cl_uint p7, const cl_event* p8, cl_event* p9) { return clEnqueueNativeKernel_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8, p9); }
-#undef clEnqueueReadBuffer
-#define clEnqueueReadBuffer clEnqueueReadBuffer_fn
-inline cl_int clEnqueueReadBuffer(cl_command_queue p0, cl_mem p1, cl_bool p2, size_t p3, size_t p4, void* p5, cl_uint p6, const cl_event* p7, cl_event* p8) { return clEnqueueReadBuffer_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8); }
-#undef clEnqueueReadBufferRect
-#define clEnqueueReadBufferRect clEnqueueReadBufferRect_fn
-inline cl_int clEnqueueReadBufferRect(cl_command_queue p0, cl_mem p1, cl_bool p2, const size_t* p3, const size_t* p4, const size_t* p5, size_t p6, size_t p7, size_t p8, size_t p9, void* p10, cl_uint p11, const cl_event* p12, cl_event* p13) { return clEnqueueReadBufferRect_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, p10, p11, p12, p13); }
-#undef clEnqueueReadImage
-#define clEnqueueReadImage clEnqueueReadImage_fn
-inline cl_int clEnqueueReadImage(cl_command_queue p0, cl_mem p1, cl_bool p2, const size_t* p3, const size_t* p4, size_t p5, size_t p6, void* p7, cl_uint p8, const cl_event* p9, cl_event* p10) { return clEnqueueReadImage_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, p10); }
-#undef clEnqueueTask
-#define clEnqueueTask clEnqueueTask_fn
-inline cl_int clEnqueueTask(cl_command_queue p0, cl_kernel p1, cl_uint p2, const cl_event* p3, cl_event* p4) { return clEnqueueTask_pfn(p0, p1, p2, p3, p4); }
-#undef clEnqueueUnmapMemObject
-#define clEnqueueUnmapMemObject clEnqueueUnmapMemObject_fn
-inline cl_int clEnqueueUnmapMemObject(cl_command_queue p0, cl_mem p1, void* p2, cl_uint p3, const cl_event* p4, cl_event* p5) { return clEnqueueUnmapMemObject_pfn(p0, p1, p2, p3, p4, p5); }
-#undef clEnqueueWaitForEvents
-#define clEnqueueWaitForEvents clEnqueueWaitForEvents_fn
-inline cl_int clEnqueueWaitForEvents(cl_command_queue p0, cl_uint p1, const cl_event* p2) { return clEnqueueWaitForEvents_pfn(p0, p1, p2); }
-#undef clEnqueueWriteBuffer
-#define clEnqueueWriteBuffer clEnqueueWriteBuffer_fn
-inline cl_int clEnqueueWriteBuffer(cl_command_queue p0, cl_mem p1, cl_bool p2, size_t p3, size_t p4, const void* p5, cl_uint p6, const cl_event* p7, cl_event* p8) { return clEnqueueWriteBuffer_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8); }
-#undef clEnqueueWriteBufferRect
-#define clEnqueueWriteBufferRect clEnqueueWriteBufferRect_fn
-inline cl_int clEnqueueWriteBufferRect(cl_command_queue p0, cl_mem p1, cl_bool p2, const size_t* p3, const size_t* p4, const size_t* p5, size_t p6, size_t p7, size_t p8, size_t p9, const void* p10, cl_uint p11, const cl_event* p12, cl_event* p13) { return clEnqueueWriteBufferRect_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, p10, p11, p12, p13); }
-#undef clEnqueueWriteImage
-#define clEnqueueWriteImage clEnqueueWriteImage_fn
-inline cl_int clEnqueueWriteImage(cl_command_queue p0, cl_mem p1, cl_bool p2, const size_t* p3, const size_t* p4, size_t p5, size_t p6, const void* p7, cl_uint p8, const cl_event* p9, cl_event* p10) { return clEnqueueWriteImage_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, p10); }
-#undef clFinish
-#define clFinish clFinish_fn
-inline cl_int clFinish(cl_command_queue p0) { return clFinish_pfn(p0); }
-#undef clFlush
-#define clFlush clFlush_fn
-inline cl_int clFlush(cl_command_queue p0) { return clFlush_pfn(p0); }
-#undef clGetCommandQueueInfo
-#define clGetCommandQueueInfo clGetCommandQueueInfo_fn
-inline cl_int clGetCommandQueueInfo(cl_command_queue p0, cl_command_queue_info p1, size_t p2, void* p3, size_t* p4) { return clGetCommandQueueInfo_pfn(p0, p1, p2, p3, p4); }
-#undef clGetContextInfo
-#define clGetContextInfo clGetContextInfo_fn
-inline cl_int clGetContextInfo(cl_context p0, cl_context_info p1, size_t p2, void* p3, size_t* p4) { return clGetContextInfo_pfn(p0, p1, p2, p3, p4); }
-#undef clGetDeviceIDs
-#define clGetDeviceIDs clGetDeviceIDs_fn
-inline cl_int clGetDeviceIDs(cl_platform_id p0, cl_device_type p1, cl_uint p2, cl_device_id* p3, cl_uint* p4) { return clGetDeviceIDs_pfn(p0, p1, p2, p3, p4); }
-#undef clGetDeviceInfo
-#define clGetDeviceInfo clGetDeviceInfo_fn
-inline cl_int clGetDeviceInfo(cl_device_id p0, cl_device_info p1, size_t p2, void* p3, size_t* p4) { return clGetDeviceInfo_pfn(p0, p1, p2, p3, p4); }
-#undef clGetEventInfo
-#define clGetEventInfo clGetEventInfo_fn
-inline cl_int clGetEventInfo(cl_event p0, cl_event_info p1, size_t p2, void* p3, size_t* p4) { return clGetEventInfo_pfn(p0, p1, p2, p3, p4); }
-#undef clGetEventProfilingInfo
-#define clGetEventProfilingInfo clGetEventProfilingInfo_fn
-inline cl_int clGetEventProfilingInfo(cl_event p0, cl_profiling_info p1, size_t p2, void* p3, size_t* p4) { return clGetEventProfilingInfo_pfn(p0, p1, p2, p3, p4); }
-#undef clGetExtensionFunctionAddress
-#define clGetExtensionFunctionAddress clGetExtensionFunctionAddress_fn
-inline void* clGetExtensionFunctionAddress(const char* p0) { return clGetExtensionFunctionAddress_pfn(p0); }
-#undef clGetExtensionFunctionAddressForPlatform
-#define clGetExtensionFunctionAddressForPlatform clGetExtensionFunctionAddressForPlatform_fn
-inline void* clGetExtensionFunctionAddressForPlatform(cl_platform_id p0, const char* p1) { return clGetExtensionFunctionAddressForPlatform_pfn(p0, p1); }
-#undef clGetImageInfo
-#define clGetImageInfo clGetImageInfo_fn
-inline cl_int clGetImageInfo(cl_mem p0, cl_image_info p1, size_t p2, void* p3, size_t* p4) { return clGetImageInfo_pfn(p0, p1, p2, p3, p4); }
-#undef clGetKernelArgInfo
-#define clGetKernelArgInfo clGetKernelArgInfo_fn
-inline cl_int clGetKernelArgInfo(cl_kernel p0, cl_uint p1, cl_kernel_arg_info p2, size_t p3, void* p4, size_t* p5) { return clGetKernelArgInfo_pfn(p0, p1, p2, p3, p4, p5); }
-#undef clGetKernelInfo
-#define clGetKernelInfo clGetKernelInfo_fn
-inline cl_int clGetKernelInfo(cl_kernel p0, cl_kernel_info p1, size_t p2, void* p3, size_t* p4) { return clGetKernelInfo_pfn(p0, p1, p2, p3, p4); }
-#undef clGetKernelWorkGroupInfo
-#define clGetKernelWorkGroupInfo clGetKernelWorkGroupInfo_fn
-inline cl_int clGetKernelWorkGroupInfo(cl_kernel p0, cl_device_id p1, cl_kernel_work_group_info p2, size_t p3, void* p4, size_t* p5) { return clGetKernelWorkGroupInfo_pfn(p0, p1, p2, p3, p4, p5); }
-#undef clGetMemObjectInfo
-#define clGetMemObjectInfo clGetMemObjectInfo_fn
-inline cl_int clGetMemObjectInfo(cl_mem p0, cl_mem_info p1, size_t p2, void* p3, size_t* p4) { return clGetMemObjectInfo_pfn(p0, p1, p2, p3, p4); }
-#undef clGetPlatformIDs
-#define clGetPlatformIDs clGetPlatformIDs_fn
-inline cl_int clGetPlatformIDs(cl_uint p0, cl_platform_id* p1, cl_uint* p2) { return clGetPlatformIDs_pfn(p0, p1, p2); }
-#undef clGetPlatformInfo
-#define clGetPlatformInfo clGetPlatformInfo_fn
-inline cl_int clGetPlatformInfo(cl_platform_id p0, cl_platform_info p1, size_t p2, void* p3, size_t* p4) { return clGetPlatformInfo_pfn(p0, p1, p2, p3, p4); }
-#undef clGetProgramBuildInfo
-#define clGetProgramBuildInfo clGetProgramBuildInfo_fn
-inline cl_int clGetProgramBuildInfo(cl_program p0, cl_device_id p1, cl_program_build_info p2, size_t p3, void* p4, size_t* p5) { return clGetProgramBuildInfo_pfn(p0, p1, p2, p3, p4, p5); }
-#undef clGetProgramInfo
-#define clGetProgramInfo clGetProgramInfo_fn
-inline cl_int clGetProgramInfo(cl_program p0, cl_program_info p1, size_t p2, void* p3, size_t* p4) { return clGetProgramInfo_pfn(p0, p1, p2, p3, p4); }
-#undef clGetSamplerInfo
-#define clGetSamplerInfo clGetSamplerInfo_fn
-inline cl_int clGetSamplerInfo(cl_sampler p0, cl_sampler_info p1, size_t p2, void* p3, size_t* p4) { return clGetSamplerInfo_pfn(p0, p1, p2, p3, p4); }
-#undef clGetSupportedImageFormats
-#define clGetSupportedImageFormats clGetSupportedImageFormats_fn
-inline cl_int clGetSupportedImageFormats(cl_context p0, cl_mem_flags p1, cl_mem_object_type p2, cl_uint p3, cl_image_format* p4, cl_uint* p5) { return clGetSupportedImageFormats_pfn(p0, p1, p2, p3, p4, p5); }
-#undef clLinkProgram
-#define clLinkProgram clLinkProgram_fn
-inline cl_program clLinkProgram(cl_context p0, cl_uint p1, const cl_device_id* p2, const char* p3, cl_uint p4, const cl_program* p5, void (CL_CALLBACK*p6) (cl_program, void*), void* p7, cl_int* p8) { return clLinkProgram_pfn(p0, p1, p2, p3, p4, p5, p6, p7, p8); }
-#undef clReleaseCommandQueue
-#define clReleaseCommandQueue clReleaseCommandQueue_fn
-inline cl_int clReleaseCommandQueue(cl_command_queue p0) { return clReleaseCommandQueue_pfn(p0); }
-#undef clReleaseContext
-#define clReleaseContext clReleaseContext_fn
-inline cl_int clReleaseContext(cl_context p0) { return clReleaseContext_pfn(p0); }
-#undef clReleaseDevice
-#define clReleaseDevice clReleaseDevice_fn
-inline cl_int clReleaseDevice(cl_device_id p0) { return clReleaseDevice_pfn(p0); }
-#undef clReleaseEvent
-#define clReleaseEvent clReleaseEvent_fn
-inline cl_int clReleaseEvent(cl_event p0) { return clReleaseEvent_pfn(p0); }
-#undef clReleaseKernel
-#define clReleaseKernel clReleaseKernel_fn
-inline cl_int clReleaseKernel(cl_kernel p0) { return clReleaseKernel_pfn(p0); }
-#undef clReleaseMemObject
-#define clReleaseMemObject clReleaseMemObject_fn
-inline cl_int clReleaseMemObject(cl_mem p0) { return clReleaseMemObject_pfn(p0); }
-#undef clReleaseProgram
-#define clReleaseProgram clReleaseProgram_fn
-inline cl_int clReleaseProgram(cl_program p0) { return clReleaseProgram_pfn(p0); }
-#undef clReleaseSampler
-#define clReleaseSampler clReleaseSampler_fn
-inline cl_int clReleaseSampler(cl_sampler p0) { return clReleaseSampler_pfn(p0); }
-#undef clRetainCommandQueue
-#define clRetainCommandQueue clRetainCommandQueue_fn
-inline cl_int clRetainCommandQueue(cl_command_queue p0) { return clRetainCommandQueue_pfn(p0); }
-#undef clRetainContext
-#define clRetainContext clRetainContext_fn
-inline cl_int clRetainContext(cl_context p0) { return clRetainContext_pfn(p0); }
-#undef clRetainDevice
-#define clRetainDevice clRetainDevice_fn
-inline cl_int clRetainDevice(cl_device_id p0) { return clRetainDevice_pfn(p0); }
-#undef clRetainEvent
-#define clRetainEvent clRetainEvent_fn
-inline cl_int clRetainEvent(cl_event p0) { return clRetainEvent_pfn(p0); }
-#undef clRetainKernel
-#define clRetainKernel clRetainKernel_fn
-inline cl_int clRetainKernel(cl_kernel p0) { return clRetainKernel_pfn(p0); }
-#undef clRetainMemObject
-#define clRetainMemObject clRetainMemObject_fn
-inline cl_int clRetainMemObject(cl_mem p0) { return clRetainMemObject_pfn(p0); }
-#undef clRetainProgram
-#define clRetainProgram clRetainProgram_fn
-inline cl_int clRetainProgram(cl_program p0) { return clRetainProgram_pfn(p0); }
-#undef clRetainSampler
-#define clRetainSampler clRetainSampler_fn
-inline cl_int clRetainSampler(cl_sampler p0) { return clRetainSampler_pfn(p0); }
-#undef clSetEventCallback
-#define clSetEventCallback clSetEventCallback_fn
-inline cl_int clSetEventCallback(cl_event p0, cl_int p1, void (CL_CALLBACK*p2) (cl_event, cl_int, void*), void* p3) { return clSetEventCallback_pfn(p0, p1, p2, p3); }
-#undef clSetKernelArg
-#define clSetKernelArg clSetKernelArg_fn
-inline cl_int clSetKernelArg(cl_kernel p0, cl_uint p1, size_t p2, const void* p3) { return clSetKernelArg_pfn(p0, p1, p2, p3); }
-#undef clSetMemObjectDestructorCallback
-#define clSetMemObjectDestructorCallback clSetMemObjectDestructorCallback_fn
-inline cl_int clSetMemObjectDestructorCallback(cl_mem p0, void (CL_CALLBACK*p1) (cl_mem, void*), void* p2) { return clSetMemObjectDestructorCallback_pfn(p0, p1, p2); }
-#undef clSetUserEventStatus
-#define clSetUserEventStatus clSetUserEventStatus_fn
-inline cl_int clSetUserEventStatus(cl_event p0, cl_int p1) { return clSetUserEventStatus_pfn(p0, p1); }
-#undef clUnloadCompiler
-#define clUnloadCompiler clUnloadCompiler_fn
-inline cl_int clUnloadCompiler() { return clUnloadCompiler_pfn(); }
-#undef clUnloadPlatformCompiler
-#define clUnloadPlatformCompiler clUnloadPlatformCompiler_fn
-inline cl_int clUnloadPlatformCompiler(cl_platform_id p0) { return clUnloadPlatformCompiler_pfn(p0); }
-#undef clWaitForEvents
-#define clWaitForEvents clWaitForEvents_fn
-inline cl_int clWaitForEvents(cl_uint p0, const cl_event* p1) { return clWaitForEvents_pfn(p0, p1); }
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_gl.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_gl.hpp
deleted file mode 100644
index 0b12aed6..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_gl.hpp
+++ /dev/null
@@ -1,62 +0,0 @@
-//
-// AUTOGENERATED, DO NOT EDIT
-//
-#ifndef OPENCV_CORE_OCL_RUNTIME_OPENCL_GL_HPP
-#error "Invalid usage"
-#endif
-
-// generated by parser_cl.py
-#define clCreateFromGLBuffer clCreateFromGLBuffer_
-#define clCreateFromGLRenderbuffer clCreateFromGLRenderbuffer_
-#define clCreateFromGLTexture clCreateFromGLTexture_
-#define clCreateFromGLTexture2D clCreateFromGLTexture2D_
-#define clCreateFromGLTexture3D clCreateFromGLTexture3D_
-#define clEnqueueAcquireGLObjects clEnqueueAcquireGLObjects_
-#define clEnqueueReleaseGLObjects clEnqueueReleaseGLObjects_
-#define clGetGLContextInfoKHR clGetGLContextInfoKHR_
-#define clGetGLObjectInfo clGetGLObjectInfo_
-#define clGetGLTextureInfo clGetGLTextureInfo_
-
-#if defined __APPLE__
-#include <OpenCL/cl_gl.h>
-#else
-#include <CL/cl_gl.h>
-#endif
-
-// generated by parser_cl.py
-#undef clCreateFromGLBuffer
-#define clCreateFromGLBuffer clCreateFromGLBuffer_pfn
-#undef clCreateFromGLRenderbuffer
-#define clCreateFromGLRenderbuffer clCreateFromGLRenderbuffer_pfn
-#undef clCreateFromGLTexture
-#define clCreateFromGLTexture clCreateFromGLTexture_pfn
-#undef clCreateFromGLTexture2D
-#define clCreateFromGLTexture2D clCreateFromGLTexture2D_pfn
-#undef clCreateFromGLTexture3D
-#define clCreateFromGLTexture3D clCreateFromGLTexture3D_pfn
-#undef clEnqueueAcquireGLObjects
-#define clEnqueueAcquireGLObjects clEnqueueAcquireGLObjects_pfn
-#undef clEnqueueReleaseGLObjects
-#define clEnqueueReleaseGLObjects clEnqueueReleaseGLObjects_pfn
-#undef clGetGLContextInfoKHR
-#define clGetGLContextInfoKHR clGetGLContextInfoKHR_pfn
-#undef clGetGLObjectInfo
-#define clGetGLObjectInfo clGetGLObjectInfo_pfn
-#undef clGetGLTextureInfo
-#define clGetGLTextureInfo clGetGLTextureInfo_pfn
-
-#ifdef cl_khr_gl_sharing
-
-// generated by parser_cl.py
-extern CL_RUNTIME_EXPORT cl_mem (CL_API_CALL*clCreateFromGLBuffer)(cl_context, cl_mem_flags, cl_GLuint, int*);
-extern CL_RUNTIME_EXPORT cl_mem (CL_API_CALL*clCreateFromGLRenderbuffer)(cl_context, cl_mem_flags, cl_GLuint, cl_int*);
-extern CL_RUNTIME_EXPORT cl_mem (CL_API_CALL*clCreateFromGLTexture)(cl_context, cl_mem_flags, cl_GLenum, cl_GLint, cl_GLuint, cl_int*);
-extern CL_RUNTIME_EXPORT cl_mem (CL_API_CALL*clCreateFromGLTexture2D)(cl_context, cl_mem_flags, cl_GLenum, cl_GLint, cl_GLuint, cl_int*);
-extern CL_RUNTIME_EXPORT cl_mem (CL_API_CALL*clCreateFromGLTexture3D)(cl_context, cl_mem_flags, cl_GLenum, cl_GLint, cl_GLuint, cl_int*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueAcquireGLObjects)(cl_command_queue, cl_uint, const cl_mem*, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clEnqueueReleaseGLObjects)(cl_command_queue, cl_uint, const cl_mem*, cl_uint, const cl_event*, cl_event*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetGLContextInfoKHR)(const cl_context_properties*, cl_gl_context_info, size_t, void*, size_t*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetGLObjectInfo)(cl_mem, cl_gl_object_type*, cl_GLuint*);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL*clGetGLTextureInfo)(cl_mem, cl_gl_texture_info, size_t, void*, size_t*);
-
-#endif // cl_khr_gl_sharing
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_gl_wrappers.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_gl_wrappers.hpp
deleted file mode 100644
index 12f342b2..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/autogenerated/opencl_gl_wrappers.hpp
+++ /dev/null
@@ -1,42 +0,0 @@
-//
-// AUTOGENERATED, DO NOT EDIT
-//
-#ifndef OPENCV_CORE_OCL_RUNTIME_OPENCL_GL_WRAPPERS_HPP
-#error "Invalid usage"
-#endif
-
-#ifdef cl_khr_gl_sharing
-
-// generated by parser_cl.py
-#undef clCreateFromGLBuffer
-#define clCreateFromGLBuffer clCreateFromGLBuffer_fn
-inline cl_mem clCreateFromGLBuffer(cl_context p0, cl_mem_flags p1, cl_GLuint p2, int* p3) { return clCreateFromGLBuffer_pfn(p0, p1, p2, p3); }
-#undef clCreateFromGLRenderbuffer
-#define clCreateFromGLRenderbuffer clCreateFromGLRenderbuffer_fn
-inline cl_mem clCreateFromGLRenderbuffer(cl_context p0, cl_mem_flags p1, cl_GLuint p2, cl_int* p3) { return clCreateFromGLRenderbuffer_pfn(p0, p1, p2, p3); }
-#undef clCreateFromGLTexture
-#define clCreateFromGLTexture clCreateFromGLTexture_fn
-inline cl_mem clCreateFromGLTexture(cl_context p0, cl_mem_flags p1, cl_GLenum p2, cl_GLint p3, cl_GLuint p4, cl_int* p5) { return clCreateFromGLTexture_pfn(p0, p1, p2, p3, p4, p5); }
-#undef clCreateFromGLTexture2D
-#define clCreateFromGLTexture2D clCreateFromGLTexture2D_fn
-inline cl_mem clCreateFromGLTexture2D(cl_context p0, cl_mem_flags p1, cl_GLenum p2, cl_GLint p3, cl_GLuint p4, cl_int* p5) { return clCreateFromGLTexture2D_pfn(p0, p1, p2, p3, p4, p5); }
-#undef clCreateFromGLTexture3D
-#define clCreateFromGLTexture3D clCreateFromGLTexture3D_fn
-inline cl_mem clCreateFromGLTexture3D(cl_context p0, cl_mem_flags p1, cl_GLenum p2, cl_GLint p3, cl_GLuint p4, cl_int* p5) { return clCreateFromGLTexture3D_pfn(p0, p1, p2, p3, p4, p5); }
-#undef clEnqueueAcquireGLObjects
-#define clEnqueueAcquireGLObjects clEnqueueAcquireGLObjects_fn
-inline cl_int clEnqueueAcquireGLObjects(cl_command_queue p0, cl_uint p1, const cl_mem* p2, cl_uint p3, const cl_event* p4, cl_event* p5) { return clEnqueueAcquireGLObjects_pfn(p0, p1, p2, p3, p4, p5); }
-#undef clEnqueueReleaseGLObjects
-#define clEnqueueReleaseGLObjects clEnqueueReleaseGLObjects_fn
-inline cl_int clEnqueueReleaseGLObjects(cl_command_queue p0, cl_uint p1, const cl_mem* p2, cl_uint p3, const cl_event* p4, cl_event* p5) { return clEnqueueReleaseGLObjects_pfn(p0, p1, p2, p3, p4, p5); }
-#undef clGetGLContextInfoKHR
-#define clGetGLContextInfoKHR clGetGLContextInfoKHR_fn
-inline cl_int clGetGLContextInfoKHR(const cl_context_properties* p0, cl_gl_context_info p1, size_t p2, void* p3, size_t* p4) { return clGetGLContextInfoKHR_pfn(p0, p1, p2, p3, p4); }
-#undef clGetGLObjectInfo
-#define clGetGLObjectInfo clGetGLObjectInfo_fn
-inline cl_int clGetGLObjectInfo(cl_mem p0, cl_gl_object_type* p1, cl_GLuint* p2) { return clGetGLObjectInfo_pfn(p0, p1, p2); }
-#undef clGetGLTextureInfo
-#define clGetGLTextureInfo clGetGLTextureInfo_fn
-inline cl_int clGetGLTextureInfo(cl_mem p0, cl_gl_texture_info p1, size_t p2, void* p3, size_t* p4) { return clGetGLTextureInfo_pfn(p0, p1, p2, p3, p4); }
-
-#endif // cl_khr_gl_sharing
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_clblas.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_clblas.hpp
deleted file mode 100644
index ccddf8f7..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_clblas.hpp
+++ /dev/null
@@ -1,53 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2010-2013, Advanced Micro Devices, Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the OpenCV Foundation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_OCL_RUNTIME_CLAMDBLAS_HPP
-#define OPENCV_CORE_OCL_RUNTIME_CLAMDBLAS_HPP
-
-#ifdef HAVE_CLAMDBLAS
-
-#include "opencl_core.hpp"
-
-#include "autogenerated/opencl_clblas.hpp"
-
-#endif // HAVE_CLAMDBLAS
-
-#endif // OPENCV_CORE_OCL_RUNTIME_CLAMDBLAS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_clfft.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_clfft.hpp
deleted file mode 100644
index 7f4af5e6..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_clfft.hpp
+++ /dev/null
@@ -1,53 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2010-2013, Advanced Micro Devices, Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the OpenCV Foundation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_OCL_RUNTIME_CLAMDFFT_HPP
-#define OPENCV_CORE_OCL_RUNTIME_CLAMDFFT_HPP
-
-#ifdef HAVE_CLAMDFFT
-
-#include "opencl_core.hpp"
-
-#include "autogenerated/opencl_clfft.hpp"
-
-#endif // HAVE_CLAMDFFT
-
-#endif // OPENCV_CORE_OCL_RUNTIME_CLAMDFFT_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_core.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_core.hpp
deleted file mode 100644
index 0404b317..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_core.hpp
+++ /dev/null
@@ -1,84 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2010-2013, Advanced Micro Devices, Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the OpenCV Foundation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_OCL_RUNTIME_OPENCL_CORE_HPP
-#define OPENCV_CORE_OCL_RUNTIME_OPENCL_CORE_HPP
-
-#ifdef HAVE_OPENCL
-
-#ifndef CL_RUNTIME_EXPORT
-#if (defined(BUILD_SHARED_LIBS) || defined(OPENCV_CORE_SHARED)) && (defined _WIN32 || defined WINCE) && \
-    !(defined(__OPENCV_BUILD) && defined(OPENCV_MODULE_IS_PART_OF_WORLD))
-#define CL_RUNTIME_EXPORT __declspec(dllimport)
-#else
-#define CL_RUNTIME_EXPORT
-#endif
-#endif
-
-#ifdef HAVE_OPENCL_SVM
-#define clSVMAlloc clSVMAlloc_
-#define clSVMFree clSVMFree_
-#define clSetKernelArgSVMPointer clSetKernelArgSVMPointer_
-#define clSetKernelExecInfo clSetKernelExecInfo_
-#define clEnqueueSVMFree clEnqueueSVMFree_
-#define clEnqueueSVMMemcpy clEnqueueSVMMemcpy_
-#define clEnqueueSVMMemFill clEnqueueSVMMemFill_
-#define clEnqueueSVMMap clEnqueueSVMMap_
-#define clEnqueueSVMUnmap clEnqueueSVMUnmap_
-#endif
-
-#include "autogenerated/opencl_core.hpp"
-
-#ifndef CL_DEVICE_DOUBLE_FP_CONFIG
-#define CL_DEVICE_DOUBLE_FP_CONFIG 0x1032
-#endif
-
-#ifndef CL_DEVICE_HALF_FP_CONFIG
-#define CL_DEVICE_HALF_FP_CONFIG 0x1033
-#endif
-
-#ifndef CL_VERSION_1_2
-#define CV_REQUIRE_OPENCL_1_2_ERROR CV_Error(cv::Error::OpenCLApiCallError, "OpenCV compiled without OpenCL v1.2 support, so we can't use functionality from OpenCL v1.2")
-#endif
-
-#endif // HAVE_OPENCL
-
-#endif // OPENCV_CORE_OCL_RUNTIME_OPENCL_CORE_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_core_wrappers.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_core_wrappers.hpp
deleted file mode 100644
index 38fcae99..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_core_wrappers.hpp
+++ /dev/null
@@ -1,47 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2010-2013, Advanced Micro Devices, Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the OpenCV Foundation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_OCL_RUNTIME_OPENCL_WRAPPERS_HPP
-#define OPENCV_CORE_OCL_RUNTIME_OPENCL_WRAPPERS_HPP
-
-#include "autogenerated/opencl_core_wrappers.hpp"
-
-#endif // OPENCV_CORE_OCL_RUNTIME_OPENCL_WRAPPERS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_gl.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_gl.hpp
deleted file mode 100644
index 659c7d80..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_gl.hpp
+++ /dev/null
@@ -1,53 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2010-2013, Advanced Micro Devices, Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the OpenCV Foundation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_OCL_RUNTIME_OPENCL_GL_HPP
-#define OPENCV_CORE_OCL_RUNTIME_OPENCL_GL_HPP
-
-#if defined HAVE_OPENCL && defined HAVE_OPENGL
-
-#include "opencl_core.hpp"
-
-#include "autogenerated/opencl_gl.hpp"
-
-#endif // defined HAVE_OPENCL && defined HAVE_OPENGL
-
-#endif // OPENCV_CORE_OCL_RUNTIME_OPENCL_GL_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_gl_wrappers.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_gl_wrappers.hpp
deleted file mode 100644
index 9700004c..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_gl_wrappers.hpp
+++ /dev/null
@@ -1,47 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2010-2013, Advanced Micro Devices, Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the OpenCV Foundation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_OCL_RUNTIME_OPENCL_GL_WRAPPERS_HPP
-#define OPENCV_CORE_OCL_RUNTIME_OPENCL_GL_WRAPPERS_HPP
-
-#include "autogenerated/opencl_gl_wrappers.hpp"
-
-#endif // OPENCV_CORE_OCL_RUNTIME_OPENCL_GL_WRAPPERS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_svm_20.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_svm_20.hpp
deleted file mode 100644
index 9636b19b..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_svm_20.hpp
+++ /dev/null
@@ -1,48 +0,0 @@
-/* See LICENSE file in the root OpenCV directory */
-
-#ifndef OPENCV_CORE_OCL_RUNTIME_OPENCL_SVM_2_0_HPP
-#define OPENCV_CORE_OCL_RUNTIME_OPENCL_SVM_2_0_HPP
-
-#if defined(HAVE_OPENCL_SVM)
-#include "opencl_core.hpp"
-
-#include "opencl_svm_definitions.hpp"
-
-#undef clSVMAlloc
-#define clSVMAlloc clSVMAlloc_pfn
-#undef clSVMFree
-#define clSVMFree clSVMFree_pfn
-#undef clSetKernelArgSVMPointer
-#define clSetKernelArgSVMPointer clSetKernelArgSVMPointer_pfn
-#undef clSetKernelExecInfo
-//#define clSetKernelExecInfo clSetKernelExecInfo_pfn
-#undef clEnqueueSVMFree
-//#define clEnqueueSVMFree clEnqueueSVMFree_pfn
-#undef clEnqueueSVMMemcpy
-#define clEnqueueSVMMemcpy clEnqueueSVMMemcpy_pfn
-#undef clEnqueueSVMMemFill
-#define clEnqueueSVMMemFill clEnqueueSVMMemFill_pfn
-#undef clEnqueueSVMMap
-#define clEnqueueSVMMap clEnqueueSVMMap_pfn
-#undef clEnqueueSVMUnmap
-#define clEnqueueSVMUnmap clEnqueueSVMUnmap_pfn
-
-extern CL_RUNTIME_EXPORT void* (CL_API_CALL *clSVMAlloc)(cl_context context, cl_svm_mem_flags flags, size_t size, unsigned int alignment);
-extern CL_RUNTIME_EXPORT void (CL_API_CALL *clSVMFree)(cl_context context, void* svm_pointer);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL *clSetKernelArgSVMPointer)(cl_kernel kernel, cl_uint arg_index, const void* arg_value);
-//extern CL_RUNTIME_EXPORT void* (CL_API_CALL *clSetKernelExecInfo)(cl_kernel kernel, cl_kernel_exec_info param_name, size_t param_value_size, const void* param_value);
-//extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL *clEnqueueSVMFree)(cl_command_queue command_queue, cl_uint num_svm_pointers, void* svm_pointers[],
-//        void (CL_CALLBACK *pfn_free_func)(cl_command_queue queue, cl_uint num_svm_pointers, void* svm_pointers[], void* user_data), void* user_data,
-//        cl_uint num_events_in_wait_list, const cl_event* event_wait_list, cl_event* event);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL *clEnqueueSVMMemcpy)(cl_command_queue command_queue, cl_bool blocking_copy, void* dst_ptr, const void* src_ptr, size_t size,
-        cl_uint num_events_in_wait_list, const cl_event* event_wait_list, cl_event* event);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL *clEnqueueSVMMemFill)(cl_command_queue command_queue, void* svm_ptr, const void* pattern, size_t pattern_size, size_t size,
-        cl_uint num_events_in_wait_list, const cl_event* event_wait_list, cl_event* event);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL *clEnqueueSVMMap)(cl_command_queue command_queue, cl_bool blocking_map, cl_map_flags map_flags, void* svm_ptr, size_t size,
-        cl_uint num_events_in_wait_list, const cl_event* event_wait_list, cl_event* event);
-extern CL_RUNTIME_EXPORT cl_int (CL_API_CALL *clEnqueueSVMUnmap)(cl_command_queue command_queue, void* svm_ptr,
-        cl_uint num_events_in_wait_list, const cl_event* event_wait_list, cl_event* event);
-
-#endif // HAVE_OPENCL_SVM
-
-#endif // OPENCV_CORE_OCL_RUNTIME_OPENCL_SVM_2_0_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_svm_definitions.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_svm_definitions.hpp
deleted file mode 100644
index 97c927b4..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_svm_definitions.hpp
+++ /dev/null
@@ -1,42 +0,0 @@
-/* See LICENSE file in the root OpenCV directory */
-
-#ifndef OPENCV_CORE_OCL_RUNTIME_OPENCL_SVM_DEFINITIONS_HPP
-#define OPENCV_CORE_OCL_RUNTIME_OPENCL_SVM_DEFINITIONS_HPP
-
-#if defined(HAVE_OPENCL_SVM)
-#if defined(CL_VERSION_2_0)
-
-// OpenCL 2.0 contains SVM definitions
-
-#else
-
-typedef cl_bitfield cl_device_svm_capabilities;
-typedef cl_bitfield cl_svm_mem_flags;
-typedef cl_uint     cl_kernel_exec_info;
-
-//
-// TODO Add real values after OpenCL 2.0 release
-//
-
-#ifndef CL_DEVICE_SVM_CAPABILITIES
-#define CL_DEVICE_SVM_CAPABILITIES 0x1053
-
-#define CL_DEVICE_SVM_COARSE_GRAIN_BUFFER             (1 << 0)
-#define CL_DEVICE_SVM_FINE_GRAIN_BUFFER               (1 << 1)
-#define CL_DEVICE_SVM_FINE_GRAIN_SYSTEM               (1 << 2)
-#define CL_DEVICE_SVM_ATOMICS                         (1 << 3)
-#endif
-
-#ifndef CL_MEM_SVM_FINE_GRAIN_BUFFER
-#define CL_MEM_SVM_FINE_GRAIN_BUFFER (1 << 10)
-#endif
-
-#ifndef CL_MEM_SVM_ATOMICS
-#define CL_MEM_SVM_ATOMICS (1 << 11)
-#endif
-
-
-#endif // CL_VERSION_2_0
-#endif // HAVE_OPENCL_SVM
-
-#endif // OPENCV_CORE_OCL_RUNTIME_OPENCL_SVM_DEFINITIONS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_svm_hsa_extension.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_svm_hsa_extension.hpp
deleted file mode 100644
index 497bc3de..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opencl/runtime/opencl_svm_hsa_extension.hpp
+++ /dev/null
@@ -1,166 +0,0 @@
-/* See LICENSE file in the root OpenCV directory */
-
-#ifndef OPENCV_CORE_OCL_RUNTIME_OPENCL_SVM_HSA_EXTENSION_HPP
-#define OPENCV_CORE_OCL_RUNTIME_OPENCL_SVM_HSA_EXTENSION_HPP
-
-#if defined(HAVE_OPENCL_SVM)
-#include "opencl_core.hpp"
-
-#ifndef CL_DEVICE_SVM_CAPABILITIES_AMD
-//
-//  Part of the file is an extract from the cl_ext.h file from AMD APP SDK package.
-//  Below is the original copyright.
-//
-/*******************************************************************************
- * Copyright (c) 2008-2013 The Khronos Group Inc.
- *
- * Permission is hereby granted, free of charge, to any person obtaining a
- * copy of this software and/or associated documentation files (the
- * "Materials"), to deal in the Materials without restriction, including
- * without limitation the rights to use, copy, modify, merge, publish,
- * distribute, sublicense, and/or sell copies of the Materials, and to
- * permit persons to whom the Materials are furnished to do so, subject to
- * the following conditions:
- *
- * The above copyright notice and this permission notice shall be included
- * in all copies or substantial portions of the Materials.
- *
- * THE MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
- * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
- * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
- * IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
- * CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
- * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
- * MATERIALS OR THE USE OR OTHER DEALINGS IN THE MATERIALS.
- ******************************************************************************/
-
-/*******************************************
- * Shared Virtual Memory (SVM) extension
- *******************************************/
-typedef cl_bitfield                      cl_device_svm_capabilities_amd;
-typedef cl_bitfield                      cl_svm_mem_flags_amd;
-typedef cl_uint                          cl_kernel_exec_info_amd;
-
-/* cl_device_info */
-#define CL_DEVICE_SVM_CAPABILITIES_AMD                     0x1053
-#define CL_DEVICE_PREFERRED_PLATFORM_ATOMIC_ALIGNMENT_AMD  0x1054
-
-/* cl_device_svm_capabilities_amd */
-#define CL_DEVICE_SVM_COARSE_GRAIN_BUFFER_AMD             (1 << 0)
-#define CL_DEVICE_SVM_FINE_GRAIN_BUFFER_AMD               (1 << 1)
-#define CL_DEVICE_SVM_FINE_GRAIN_SYSTEM_AMD               (1 << 2)
-#define CL_DEVICE_SVM_ATOMICS_AMD                         (1 << 3)
-
-/* cl_svm_mem_flags_amd */
-#define CL_MEM_SVM_FINE_GRAIN_BUFFER_AMD                  (1 << 10)
-#define CL_MEM_SVM_ATOMICS_AMD                            (1 << 11)
-
-/* cl_mem_info */
-#define CL_MEM_USES_SVM_POINTER_AMD                       0x1109
-
-/* cl_kernel_exec_info_amd */
-#define CL_KERNEL_EXEC_INFO_SVM_PTRS_AMD                  0x11B6
-#define CL_KERNEL_EXEC_INFO_SVM_FINE_GRAIN_SYSTEM_AMD     0x11B7
-
-/* cl_command_type */
-#define CL_COMMAND_SVM_FREE_AMD                           0x1209
-#define CL_COMMAND_SVM_MEMCPY_AMD                         0x120A
-#define CL_COMMAND_SVM_MEMFILL_AMD                        0x120B
-#define CL_COMMAND_SVM_MAP_AMD                            0x120C
-#define CL_COMMAND_SVM_UNMAP_AMD                          0x120D
-
-typedef CL_API_ENTRY void*
-(CL_API_CALL * clSVMAllocAMD_fn)(
-    cl_context            /* context */,
-    cl_svm_mem_flags_amd  /* flags */,
-    size_t                /* size */,
-    unsigned int          /* alignment */
-) CL_EXT_SUFFIX__VERSION_1_2;
-
-typedef CL_API_ENTRY void
-(CL_API_CALL * clSVMFreeAMD_fn)(
-    cl_context  /* context */,
-    void*       /* svm_pointer */
-) CL_EXT_SUFFIX__VERSION_1_2;
-
-typedef CL_API_ENTRY cl_int
-(CL_API_CALL * clEnqueueSVMFreeAMD_fn)(
-    cl_command_queue /* command_queue */,
-    cl_uint          /* num_svm_pointers */,
-    void**           /* svm_pointers */,
-    void (CL_CALLBACK *)( /*pfn_free_func*/
-        cl_command_queue /* queue */,
-        cl_uint          /* num_svm_pointers */,
-        void**           /* svm_pointers */,
-        void*            /* user_data */),
-    void*             /* user_data */,
-    cl_uint           /* num_events_in_wait_list */,
-    const cl_event*   /* event_wait_list */,
-    cl_event*         /* event */
-) CL_EXT_SUFFIX__VERSION_1_2;
-
-typedef CL_API_ENTRY cl_int
-(CL_API_CALL * clEnqueueSVMMemcpyAMD_fn)(
-    cl_command_queue /* command_queue */,
-    cl_bool          /* blocking_copy */,
-    void*            /* dst_ptr */,
-    const void*      /* src_ptr */,
-    size_t           /* size */,
-    cl_uint          /* num_events_in_wait_list */,
-    const cl_event*  /* event_wait_list */,
-    cl_event*        /* event */
-) CL_EXT_SUFFIX__VERSION_1_2;
-
-typedef CL_API_ENTRY cl_int
-(CL_API_CALL * clEnqueueSVMMemFillAMD_fn)(
-    cl_command_queue /* command_queue */,
-    void*            /* svm_ptr */,
-    const void*      /* pattern */,
-    size_t           /* pattern_size */,
-    size_t           /* size */,
-    cl_uint          /* num_events_in_wait_list */,
-    const cl_event*  /* event_wait_list */,
-    cl_event*        /* event */
-) CL_EXT_SUFFIX__VERSION_1_2;
-
-typedef CL_API_ENTRY cl_int
-(CL_API_CALL * clEnqueueSVMMapAMD_fn)(
-    cl_command_queue /* command_queue */,
-    cl_bool          /* blocking_map */,
-    cl_map_flags     /* map_flags */,
-    void*            /* svm_ptr */,
-    size_t           /* size */,
-    cl_uint          /* num_events_in_wait_list */,
-    const cl_event*  /* event_wait_list */,
-    cl_event*        /* event */
-) CL_EXT_SUFFIX__VERSION_1_2;
-
-typedef CL_API_ENTRY cl_int
-(CL_API_CALL * clEnqueueSVMUnmapAMD_fn)(
-    cl_command_queue /* command_queue */,
-    void*            /* svm_ptr */,
-    cl_uint          /* num_events_in_wait_list */,
-    const cl_event*  /* event_wait_list */,
-    cl_event*        /* event */
-) CL_EXT_SUFFIX__VERSION_1_2;
-
-typedef CL_API_ENTRY cl_int
-(CL_API_CALL * clSetKernelArgSVMPointerAMD_fn)(
-    cl_kernel     /* kernel */,
-    cl_uint       /* arg_index */,
-    const void *  /* arg_value */
-) CL_EXT_SUFFIX__VERSION_1_2;
-
-typedef CL_API_ENTRY cl_int
-(CL_API_CALL * clSetKernelExecInfoAMD_fn)(
-     cl_kernel                /* kernel */,
-     cl_kernel_exec_info_amd  /* param_name */,
-     size_t                   /* param_value_size */,
-     const void *             /* param_value */
-) CL_EXT_SUFFIX__VERSION_1_2;
-
-#endif
-
-#endif // HAVE_OPENCL_SVM
-
-#endif // OPENCV_CORE_OCL_RUNTIME_OPENCL_SVM_HSA_EXTENSION_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/opengl.hpp b/projects/llm_framework/include/opencv4/opencv2/core/opengl.hpp
deleted file mode 100644
index a311ce25..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/opengl.hpp
+++ /dev/null
@@ -1,725 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_OPENGL_HPP
-#define OPENCV_CORE_OPENGL_HPP
-
-#ifndef __cplusplus
-#  error opengl.hpp header must be compiled as C++
-#endif
-
-#include "opencv2/core.hpp"
-#include "ocl.hpp"
-
-namespace cv { namespace ogl {
-
-/** @addtogroup core_opengl
-This section describes OpenGL interoperability.
-
-To enable OpenGL support, configure OpenCV using CMake with WITH_OPENGL=ON . Currently OpenGL is
-supported only with WIN32, GTK and Qt backends on Windows and Linux (MacOS and Android are not
-supported). For GTK backend gtkglext-1.0 library is required.
-
-To use OpenGL functionality you should first create OpenGL context (window or frame buffer). You can
-do this with namedWindow function or with other OpenGL toolkit (GLUT, for example).
-*/
-//! @{
-
-/////////////////// OpenGL Objects ///////////////////
-
-/** @brief Smart pointer for OpenGL buffer object with reference counting.
-
-Buffer Objects are OpenGL objects that store an array of unformatted memory allocated by the OpenGL
-context. These can be used to store vertex data, pixel data retrieved from images or the
-framebuffer, and a variety of other things.
-
-ogl::Buffer has interface similar with Mat interface and represents 2D array memory.
-
-ogl::Buffer supports memory transfers between host and device and also can be mapped to CUDA memory.
- */
-class CV_EXPORTS Buffer
-{
-public:
-    /** @brief The target defines how you intend to use the buffer object.
-    */
-    enum Target
-    {
-        ARRAY_BUFFER         = 0x8892, //!< The buffer will be used as a source for vertex data
-        ELEMENT_ARRAY_BUFFER = 0x8893, //!< The buffer will be used for indices (in glDrawElements, for example)
-        PIXEL_PACK_BUFFER    = 0x88EB, //!< The buffer will be used for reading from OpenGL textures
-        PIXEL_UNPACK_BUFFER  = 0x88EC  //!< The buffer will be used for writing to OpenGL textures
-    };
-
-    enum Access
-    {
-        READ_ONLY  = 0x88B8,
-        WRITE_ONLY = 0x88B9,
-        READ_WRITE = 0x88BA
-    };
-
-    /** @brief The constructors.
-
-    Creates empty ogl::Buffer object, creates ogl::Buffer object from existed buffer ( abufId
-    parameter), allocates memory for ogl::Buffer object or copies from host/device memory.
-     */
-    Buffer();
-
-    /** @overload
-    @param arows Number of rows in a 2D array.
-    @param acols Number of columns in a 2D array.
-    @param atype Array type ( CV_8UC1, ..., CV_64FC4 ). See Mat for details.
-    @param abufId Buffer object name.
-    @param autoRelease Auto release mode (if true, release will be called in object's destructor).
-    */
-    Buffer(int arows, int acols, int atype, unsigned int abufId, bool autoRelease = false);
-
-    /** @overload
-    @param asize 2D array size.
-    @param atype Array type ( CV_8UC1, ..., CV_64FC4 ). See Mat for details.
-    @param abufId Buffer object name.
-    @param autoRelease Auto release mode (if true, release will be called in object's destructor).
-    */
-    Buffer(Size asize, int atype, unsigned int abufId, bool autoRelease = false);
-
-    /** @overload
-    @param arows Number of rows in a 2D array.
-    @param acols Number of columns in a 2D array.
-    @param atype Array type ( CV_8UC1, ..., CV_64FC4 ). See Mat for details.
-    @param target Buffer usage. See cv::ogl::Buffer::Target .
-    @param autoRelease Auto release mode (if true, release will be called in object's destructor).
-    */
-    Buffer(int arows, int acols, int atype, Target target = ARRAY_BUFFER, bool autoRelease = false);
-
-    /** @overload
-    @param asize 2D array size.
-    @param atype Array type ( CV_8UC1, ..., CV_64FC4 ). See Mat for details.
-    @param target Buffer usage. See cv::ogl::Buffer::Target .
-    @param autoRelease Auto release mode (if true, release will be called in object's destructor).
-    */
-    Buffer(Size asize, int atype, Target target = ARRAY_BUFFER, bool autoRelease = false);
-
-    /** @overload
-    @param arr Input array (host or device memory, it can be Mat , cuda::GpuMat or std::vector ).
-    @param target Buffer usage. See cv::ogl::Buffer::Target .
-    @param autoRelease Auto release mode (if true, release will be called in object's destructor).
-    */
-    explicit Buffer(InputArray arr, Target target = ARRAY_BUFFER, bool autoRelease = false);
-
-    /** @brief Allocates memory for ogl::Buffer object.
-
-    @param arows Number of rows in a 2D array.
-    @param acols Number of columns in a 2D array.
-    @param atype Array type ( CV_8UC1, ..., CV_64FC4 ). See Mat for details.
-    @param target Buffer usage. See cv::ogl::Buffer::Target .
-    @param autoRelease Auto release mode (if true, release will be called in object's destructor).
-     */
-    void create(int arows, int acols, int atype, Target target = ARRAY_BUFFER, bool autoRelease = false);
-
-    /** @overload
-    @param asize 2D array size.
-    @param atype Array type ( CV_8UC1, ..., CV_64FC4 ). See Mat for details.
-    @param target Buffer usage. See cv::ogl::Buffer::Target .
-    @param autoRelease Auto release mode (if true, release will be called in object's destructor).
-    */
-    void create(Size asize, int atype, Target target = ARRAY_BUFFER, bool autoRelease = false);
-
-    /** @brief Decrements the reference counter and destroys the buffer object if needed.
-
-    The function will call setAutoRelease(true) .
-     */
-    void release();
-
-    /** @brief Sets auto release mode.
-
-    The lifetime of the OpenGL object is tied to the lifetime of the context. If OpenGL context was
-    bound to a window it could be released at any time (user can close a window). If object's destructor
-    is called after destruction of the context it will cause an error. Thus ogl::Buffer doesn't destroy
-    OpenGL object in destructor by default (all OpenGL resources will be released with OpenGL context).
-    This function can force ogl::Buffer destructor to destroy OpenGL object.
-    @param flag Auto release mode (if true, release will be called in object's destructor).
-     */
-    void setAutoRelease(bool flag);
-
-    /** @brief Copies from host/device memory to OpenGL buffer.
-    @param arr Input array (host or device memory, it can be Mat , cuda::GpuMat or std::vector ).
-    @param target Buffer usage. See cv::ogl::Buffer::Target .
-    @param autoRelease Auto release mode (if true, release will be called in object's destructor).
-     */
-    void copyFrom(InputArray arr, Target target = ARRAY_BUFFER, bool autoRelease = false);
-
-    /** @overload */
-    void copyFrom(InputArray arr, cuda::Stream& stream, Target target = ARRAY_BUFFER, bool autoRelease = false);
-
-    /** @brief Copies from OpenGL buffer to host/device memory or another OpenGL buffer object.
-
-    @param arr Destination array (host or device memory, can be Mat , cuda::GpuMat , std::vector or
-    ogl::Buffer ).
-     */
-    void copyTo(OutputArray arr) const;
-
-    /** @overload */
-    void copyTo(OutputArray arr, cuda::Stream& stream) const;
-
-    /** @brief Creates a full copy of the buffer object and the underlying data.
-
-    @param target Buffer usage for destination buffer.
-    @param autoRelease Auto release mode for destination buffer.
-     */
-    Buffer clone(Target target = ARRAY_BUFFER, bool autoRelease = false) const;
-
-    /** @brief Binds OpenGL buffer to the specified buffer binding point.
-
-    @param target Binding point. See cv::ogl::Buffer::Target .
-     */
-    void bind(Target target) const;
-
-    /** @brief Unbind any buffers from the specified binding point.
-
-    @param target Binding point. See cv::ogl::Buffer::Target .
-     */
-    static void unbind(Target target);
-
-    /** @brief Maps OpenGL buffer to host memory.
-
-    mapHost maps to the client's address space the entire data store of the buffer object. The data can
-    then be directly read and/or written relative to the returned pointer, depending on the specified
-    access policy.
-
-    A mapped data store must be unmapped with ogl::Buffer::unmapHost before its buffer object is used.
-
-    This operation can lead to memory transfers between host and device.
-
-    Only one buffer object can be mapped at a time.
-    @param access Access policy, indicating whether it will be possible to read from, write to, or both
-    read from and write to the buffer object's mapped data store. The symbolic constant must be
-    ogl::Buffer::READ_ONLY , ogl::Buffer::WRITE_ONLY or ogl::Buffer::READ_WRITE .
-     */
-    Mat mapHost(Access access);
-
-    /** @brief Unmaps OpenGL buffer.
-    */
-    void unmapHost();
-
-    //! map to device memory (blocking)
-    cuda::GpuMat mapDevice();
-    void unmapDevice();
-
-    /** @brief Maps OpenGL buffer to CUDA device memory.
-
-    This operation doesn't copy data. Several buffer objects can be mapped to CUDA memory at a time.
-
-    A mapped data store must be unmapped with ogl::Buffer::unmapDevice before its buffer object is used.
-     */
-    cuda::GpuMat mapDevice(cuda::Stream& stream);
-
-    /** @brief Unmaps OpenGL buffer.
-    */
-    void unmapDevice(cuda::Stream& stream);
-
-    int rows() const;
-    int cols() const;
-    Size size() const;
-    bool empty() const;
-
-    int type() const;
-    int depth() const;
-    int channels() const;
-    int elemSize() const;
-    int elemSize1() const;
-
-    //! get OpenGL opject id
-    unsigned int bufId() const;
-
-    class Impl;
-
-private:
-    Ptr<Impl> impl_;
-    int rows_;
-    int cols_;
-    int type_;
-};
-
-/** @brief Smart pointer for OpenGL 2D texture memory with reference counting.
- */
-class CV_EXPORTS Texture2D
-{
-public:
-    /** @brief An Image Format describes the way that the images in Textures store their data.
-    */
-    enum Format
-    {
-        NONE            = 0,
-        DEPTH_COMPONENT = 0x1902, //!< Depth
-        RGB             = 0x1907, //!< Red, Green, Blue
-        RGBA            = 0x1908  //!< Red, Green, Blue, Alpha
-    };
-
-    /** @brief The constructors.
-
-    Creates empty ogl::Texture2D object, allocates memory for ogl::Texture2D object or copies from
-    host/device memory.
-     */
-    Texture2D();
-
-    /** @overload */
-    Texture2D(int arows, int acols, Format aformat, unsigned int atexId, bool autoRelease = false);
-
-    /** @overload */
-    Texture2D(Size asize, Format aformat, unsigned int atexId, bool autoRelease = false);
-
-    /** @overload
-    @param arows Number of rows.
-    @param acols Number of columns.
-    @param aformat Image format. See cv::ogl::Texture2D::Format .
-    @param autoRelease Auto release mode (if true, release will be called in object's destructor).
-    */
-    Texture2D(int arows, int acols, Format aformat, bool autoRelease = false);
-
-    /** @overload
-    @param asize 2D array size.
-    @param aformat Image format. See cv::ogl::Texture2D::Format .
-    @param autoRelease Auto release mode (if true, release will be called in object's destructor).
-    */
-    Texture2D(Size asize, Format aformat, bool autoRelease = false);
-
-    /** @overload
-    @param arr Input array (host or device memory, it can be Mat , cuda::GpuMat or ogl::Buffer ).
-    @param autoRelease Auto release mode (if true, release will be called in object's destructor).
-    */
-    explicit Texture2D(InputArray arr, bool autoRelease = false);
-
-    /** @brief Allocates memory for ogl::Texture2D object.
-
-    @param arows Number of rows.
-    @param acols Number of columns.
-    @param aformat Image format. See cv::ogl::Texture2D::Format .
-    @param autoRelease Auto release mode (if true, release will be called in object's destructor).
-     */
-    void create(int arows, int acols, Format aformat, bool autoRelease = false);
-    /** @overload
-    @param asize 2D array size.
-    @param aformat Image format. See cv::ogl::Texture2D::Format .
-    @param autoRelease Auto release mode (if true, release will be called in object's destructor).
-    */
-    void create(Size asize, Format aformat, bool autoRelease = false);
-
-    /** @brief Decrements the reference counter and destroys the texture object if needed.
-
-    The function will call setAutoRelease(true) .
-     */
-    void release();
-
-    /** @brief Sets auto release mode.
-
-    @param flag Auto release mode (if true, release will be called in object's destructor).
-
-    The lifetime of the OpenGL object is tied to the lifetime of the context. If OpenGL context was
-    bound to a window it could be released at any time (user can close a window). If object's destructor
-    is called after destruction of the context it will cause an error. Thus ogl::Texture2D doesn't
-    destroy OpenGL object in destructor by default (all OpenGL resources will be released with OpenGL
-    context). This function can force ogl::Texture2D destructor to destroy OpenGL object.
-     */
-    void setAutoRelease(bool flag);
-
-    /** @brief Copies from host/device memory to OpenGL texture.
-
-    @param arr Input array (host or device memory, it can be Mat , cuda::GpuMat or ogl::Buffer ).
-    @param autoRelease Auto release mode (if true, release will be called in object's destructor).
-     */
-    void copyFrom(InputArray arr, bool autoRelease = false);
-
-    /** @brief Copies from OpenGL texture to host/device memory or another OpenGL texture object.
-
-    @param arr Destination array (host or device memory, can be Mat , cuda::GpuMat , ogl::Buffer or
-    ogl::Texture2D ).
-    @param ddepth Destination depth.
-    @param autoRelease Auto release mode for destination buffer (if arr is OpenGL buffer or texture).
-     */
-    void copyTo(OutputArray arr, int ddepth = CV_32F, bool autoRelease = false) const;
-
-    /** @brief Binds texture to current active texture unit for GL_TEXTURE_2D target.
-    */
-    void bind() const;
-
-    int rows() const;
-    int cols() const;
-    Size size() const;
-    bool empty() const;
-
-    Format format() const;
-
-    //! get OpenGL opject id
-    unsigned int texId() const;
-
-    class Impl;
-
-private:
-    Ptr<Impl> impl_;
-    int rows_;
-    int cols_;
-    Format format_;
-};
-
-/** @brief Wrapper for OpenGL Client-Side Vertex arrays.
-
-ogl::Arrays stores vertex data in ogl::Buffer objects.
- */
-class CV_EXPORTS Arrays
-{
-public:
-    /** @brief Default constructor
-     */
-    Arrays();
-
-    /** @brief Sets an array of vertex coordinates.
-    @param vertex array with vertex coordinates, can be both host and device memory.
-    */
-    void setVertexArray(InputArray vertex);
-
-    /** @brief Resets vertex coordinates.
-    */
-    void resetVertexArray();
-
-    /** @brief Sets an array of vertex colors.
-    @param color array with vertex colors, can be both host and device memory.
-     */
-    void setColorArray(InputArray color);
-
-    /** @brief Resets vertex colors.
-    */
-    void resetColorArray();
-
-    /** @brief Sets an array of vertex normals.
-    @param normal array with vertex normals, can be both host and device memory.
-     */
-    void setNormalArray(InputArray normal);
-
-    /** @brief Resets vertex normals.
-    */
-    void resetNormalArray();
-
-    /** @brief Sets an array of vertex texture coordinates.
-    @param texCoord array with vertex texture coordinates, can be both host and device memory.
-     */
-    void setTexCoordArray(InputArray texCoord);
-
-    /** @brief Resets vertex texture coordinates.
-    */
-    void resetTexCoordArray();
-
-    /** @brief Releases all inner buffers.
-    */
-    void release();
-
-    /** @brief Sets auto release mode all inner buffers.
-    @param flag Auto release mode.
-     */
-    void setAutoRelease(bool flag);
-
-    /** @brief Binds all vertex arrays.
-    */
-    void bind() const;
-
-    /** @brief Returns the vertex count.
-    */
-    int size() const;
-    bool empty() const;
-
-private:
-    int size_;
-    Buffer vertex_;
-    Buffer color_;
-    Buffer normal_;
-    Buffer texCoord_;
-};
-
-/////////////////// Render Functions ///////////////////
-
-//! render mode
-enum RenderModes {
-    POINTS         = 0x0000,
-    LINES          = 0x0001,
-    LINE_LOOP      = 0x0002,
-    LINE_STRIP     = 0x0003,
-    TRIANGLES      = 0x0004,
-    TRIANGLE_STRIP = 0x0005,
-    TRIANGLE_FAN   = 0x0006,
-    QUADS          = 0x0007,
-    QUAD_STRIP     = 0x0008,
-    POLYGON        = 0x0009
-};
-
-/** @brief Render OpenGL texture or primitives.
-@param tex Texture to draw.
-@param wndRect Region of window, where to draw a texture (normalized coordinates).
-@param texRect Region of texture to draw (normalized coordinates).
- */
-CV_EXPORTS void render(const Texture2D& tex,
-    Rect_<double> wndRect = Rect_<double>(0.0, 0.0, 1.0, 1.0),
-    Rect_<double> texRect = Rect_<double>(0.0, 0.0, 1.0, 1.0));
-
-/** @overload
-@param arr Array of privitives vertices.
-@param mode Render mode. One of cv::ogl::RenderModes
-@param color Color for all vertices. Will be used if arr doesn't contain color array.
-*/
-CV_EXPORTS void render(const Arrays& arr, int mode = POINTS, Scalar color = Scalar::all(255));
-
-/** @overload
-@param arr Array of privitives vertices.
-@param indices Array of vertices indices (host or device memory).
-@param mode Render mode. One of cv::ogl::RenderModes
-@param color Color for all vertices. Will be used if arr doesn't contain color array.
-*/
-CV_EXPORTS void render(const Arrays& arr, InputArray indices, int mode = POINTS, Scalar color = Scalar::all(255));
-
-/////////////////// CL-GL Interoperability Functions ///////////////////
-
-namespace ocl {
-using namespace cv::ocl;
-
-// TODO static functions in the Context class
-/** @brief Creates OpenCL context from GL.
-@return Returns reference to OpenCL Context
- */
-CV_EXPORTS Context& initializeContextFromGL();
-
-} // namespace cv::ogl::ocl
-
-/** @brief Converts InputArray to Texture2D object.
-@param src     - source InputArray.
-@param texture - destination Texture2D object.
- */
-CV_EXPORTS void convertToGLTexture2D(InputArray src, Texture2D& texture);
-
-/** @brief Converts Texture2D object to OutputArray.
-@param texture - source Texture2D object.
-@param dst     - destination OutputArray.
- */
-CV_EXPORTS void convertFromGLTexture2D(const Texture2D& texture, OutputArray dst);
-
-/** @brief Maps Buffer object to process on CL side (convert to UMat).
-
-Function creates CL buffer from GL one, and then constructs UMat that can be used
-to process buffer data with OpenCV functions. Note that in current implementation
-UMat constructed this way doesn't own corresponding GL buffer object, so it is
-the user responsibility to close down CL/GL buffers relationships by explicitly
-calling unmapGLBuffer() function.
-@param buffer      - source Buffer object.
-@param accessFlags - data access flags (ACCESS_READ|ACCESS_WRITE).
-@return Returns UMat object
- */
-CV_EXPORTS UMat mapGLBuffer(const Buffer& buffer, AccessFlag accessFlags = ACCESS_READ | ACCESS_WRITE);
-
-/** @brief Unmaps Buffer object (releases UMat, previously mapped from Buffer).
-
-Function must be called explicitly by the user for each UMat previously constructed
-by the call to mapGLBuffer() function.
-@param u           - source UMat, created by mapGLBuffer().
- */
-CV_EXPORTS void unmapGLBuffer(UMat& u);
-
-//! @}
-}} // namespace cv::ogl
-
-namespace cv { namespace cuda {
-
-/** @brief Sets a CUDA device and initializes it for the current thread with OpenGL interoperability.
-
-This function should be explicitly called after OpenGL context creation and before any CUDA calls.
-@param device System index of a CUDA device starting with 0.
-@ingroup core_opengl
- */
-CV_EXPORTS void setGlDevice(int device = 0);
-
-}}
-
-//! @cond IGNORED
-
-////////////////////////////////////////////////////////////////////////
-////////////////////////////////////////////////////////////////////////
-////////////////////////////////////////////////////////////////////////
-
-inline
-cv::ogl::Buffer::Buffer(int arows, int acols, int atype, Target target, bool autoRelease) : rows_(0), cols_(0), type_(0)
-{
-    create(arows, acols, atype, target, autoRelease);
-}
-
-inline
-cv::ogl::Buffer::Buffer(Size asize, int atype, Target target, bool autoRelease) : rows_(0), cols_(0), type_(0)
-{
-    create(asize, atype, target, autoRelease);
-}
-
-inline
-void cv::ogl::Buffer::create(Size asize, int atype, Target target, bool autoRelease)
-{
-    create(asize.height, asize.width, atype, target, autoRelease);
-}
-
-inline
-int cv::ogl::Buffer::rows() const
-{
-    return rows_;
-}
-
-inline
-int cv::ogl::Buffer::cols() const
-{
-    return cols_;
-}
-
-inline
-cv::Size cv::ogl::Buffer::size() const
-{
-    return Size(cols_, rows_);
-}
-
-inline
-bool cv::ogl::Buffer::empty() const
-{
-    return rows_ == 0 || cols_ == 0;
-}
-
-inline
-int cv::ogl::Buffer::type() const
-{
-    return type_;
-}
-
-inline
-int cv::ogl::Buffer::depth() const
-{
-    return CV_MAT_DEPTH(type_);
-}
-
-inline
-int cv::ogl::Buffer::channels() const
-{
-    return CV_MAT_CN(type_);
-}
-
-inline
-int cv::ogl::Buffer::elemSize() const
-{
-    return CV_ELEM_SIZE(type_);
-}
-
-inline
-int cv::ogl::Buffer::elemSize1() const
-{
-    return CV_ELEM_SIZE1(type_);
-}
-
-///////
-
-inline
-cv::ogl::Texture2D::Texture2D(int arows, int acols, Format aformat, bool autoRelease) : rows_(0), cols_(0), format_(NONE)
-{
-    create(arows, acols, aformat, autoRelease);
-}
-
-inline
-cv::ogl::Texture2D::Texture2D(Size asize, Format aformat, bool autoRelease) : rows_(0), cols_(0), format_(NONE)
-{
-    create(asize, aformat, autoRelease);
-}
-
-inline
-void cv::ogl::Texture2D::create(Size asize, Format aformat, bool autoRelease)
-{
-    create(asize.height, asize.width, aformat, autoRelease);
-}
-
-inline
-int cv::ogl::Texture2D::rows() const
-{
-    return rows_;
-}
-
-inline
-int cv::ogl::Texture2D::cols() const
-{
-    return cols_;
-}
-
-inline
-cv::Size cv::ogl::Texture2D::size() const
-{
-    return Size(cols_, rows_);
-}
-
-inline
-bool cv::ogl::Texture2D::empty() const
-{
-    return rows_ == 0 || cols_ == 0;
-}
-
-inline
-cv::ogl::Texture2D::Format cv::ogl::Texture2D::format() const
-{
-    return format_;
-}
-
-///////
-
-inline
-cv::ogl::Arrays::Arrays() : size_(0)
-{
-}
-
-inline
-int cv::ogl::Arrays::size() const
-{
-    return size_;
-}
-
-inline
-bool cv::ogl::Arrays::empty() const
-{
-    return size_ == 0;
-}
-
-//! @endcond
-
-#endif /* OPENCV_CORE_OPENGL_HPP */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/operations.hpp b/projects/llm_framework/include/opencv4/opencv2/core/operations.hpp
deleted file mode 100644
index bde28c49..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/operations.hpp
+++ /dev/null
@@ -1,594 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Copyright (C) 2015, Itseez Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_OPERATIONS_HPP
-#define OPENCV_CORE_OPERATIONS_HPP
-
-#ifndef __cplusplus
-#  error operations.hpp header must be compiled as C++
-#endif
-
-#include <cstdio>
-
-#if defined(__GNUC__) || defined(__clang__) // at least GCC 3.1+, clang 3.5+
-#  if defined(__MINGW_PRINTF_FORMAT)  // https://sourceforge.net/p/mingw-w64/wiki2/gnu%20printf/.
-#    define CV_FORMAT_PRINTF(string_idx, first_to_check) __attribute__ ((format (__MINGW_PRINTF_FORMAT, string_idx, first_to_check)))
-#  else
-#    define CV_FORMAT_PRINTF(string_idx, first_to_check) __attribute__ ((format (printf, string_idx, first_to_check)))
-#  endif
-#else
-#  define CV_FORMAT_PRINTF(A, B)
-#endif
-
-//! @cond IGNORED
-
-namespace cv
-{
-
-////////////////////////////// Matx methods depending on core API /////////////////////////////
-
-namespace internal
-{
-
-template<typename _Tp, int m, int n> struct Matx_FastInvOp
-{
-    bool operator()(const Matx<_Tp, m, n>& a, Matx<_Tp, n, m>& b, int method) const
-    {
-        return invert(a, b, method) != 0;
-    }
-};
-
-template<typename _Tp, int m> struct Matx_FastInvOp<_Tp, m, m>
-{
-    bool operator()(const Matx<_Tp, m, m>& a, Matx<_Tp, m, m>& b, int method) const
-    {
-        if (method == DECOMP_LU || method == DECOMP_CHOLESKY)
-        {
-            Matx<_Tp, m, m> temp = a;
-
-            // assume that b is all 0's on input => make it a unity matrix
-            for (int i = 0; i < m; i++)
-                b(i, i) = (_Tp)1;
-
-            if (method == DECOMP_CHOLESKY)
-                return Cholesky(temp.val, m*sizeof(_Tp), m, b.val, m*sizeof(_Tp), m);
-
-            return LU(temp.val, m*sizeof(_Tp), m, b.val, m*sizeof(_Tp), m) != 0;
-        }
-        else
-        {
-            return invert(a, b, method) != 0;
-        }
-    }
-};
-
-template<typename _Tp> struct Matx_FastInvOp<_Tp, 2, 2>
-{
-    bool operator()(const Matx<_Tp, 2, 2>& a, Matx<_Tp, 2, 2>& b, int /*method*/) const
-    {
-        _Tp d = (_Tp)determinant(a);
-        if (d == 0)
-            return false;
-        d = 1/d;
-        b(1,1) = a(0,0)*d;
-        b(0,0) = a(1,1)*d;
-        b(0,1) = -a(0,1)*d;
-        b(1,0) = -a(1,0)*d;
-        return true;
-    }
-};
-
-template<typename _Tp> struct Matx_FastInvOp<_Tp, 3, 3>
-{
-    bool operator()(const Matx<_Tp, 3, 3>& a, Matx<_Tp, 3, 3>& b, int /*method*/) const
-    {
-        _Tp d = (_Tp)determinant(a);
-        if (d == 0)
-            return false;
-        d = 1/d;
-        b(0,0) = (a(1,1) * a(2,2) - a(1,2) * a(2,1)) * d;
-        b(0,1) = (a(0,2) * a(2,1) - a(0,1) * a(2,2)) * d;
-        b(0,2) = (a(0,1) * a(1,2) - a(0,2) * a(1,1)) * d;
-
-        b(1,0) = (a(1,2) * a(2,0) - a(1,0) * a(2,2)) * d;
-        b(1,1) = (a(0,0) * a(2,2) - a(0,2) * a(2,0)) * d;
-        b(1,2) = (a(0,2) * a(1,0) - a(0,0) * a(1,2)) * d;
-
-        b(2,0) = (a(1,0) * a(2,1) - a(1,1) * a(2,0)) * d;
-        b(2,1) = (a(0,1) * a(2,0) - a(0,0) * a(2,1)) * d;
-        b(2,2) = (a(0,0) * a(1,1) - a(0,1) * a(1,0)) * d;
-        return true;
-    }
-};
-
-
-template<typename _Tp, int m, int l, int n> struct Matx_FastSolveOp
-{
-    bool operator()(const Matx<_Tp, m, l>& a, const Matx<_Tp, m, n>& b,
-                    Matx<_Tp, l, n>& x, int method) const
-    {
-        return cv::solve(a, b, x, method);
-    }
-};
-
-template<typename _Tp, int m, int n> struct Matx_FastSolveOp<_Tp, m, m, n>
-{
-    bool operator()(const Matx<_Tp, m, m>& a, const Matx<_Tp, m, n>& b,
-                    Matx<_Tp, m, n>& x, int method) const
-    {
-        if (method == DECOMP_LU || method == DECOMP_CHOLESKY)
-        {
-            Matx<_Tp, m, m> temp = a;
-            x = b;
-            if( method == DECOMP_CHOLESKY )
-                return Cholesky(temp.val, m*sizeof(_Tp), m, x.val, n*sizeof(_Tp), n);
-
-            return LU(temp.val, m*sizeof(_Tp), m, x.val, n*sizeof(_Tp), n) != 0;
-        }
-        else
-        {
-            return cv::solve(a, b, x, method);
-        }
-    }
-};
-
-template<typename _Tp> struct Matx_FastSolveOp<_Tp, 2, 2, 1>
-{
-    bool operator()(const Matx<_Tp, 2, 2>& a, const Matx<_Tp, 2, 1>& b,
-                    Matx<_Tp, 2, 1>& x, int) const
-    {
-        _Tp d = (_Tp)determinant(a);
-        if (d == 0)
-            return false;
-        d = 1/d;
-        x(0) = (b(0)*a(1,1) - b(1)*a(0,1))*d;
-        x(1) = (b(1)*a(0,0) - b(0)*a(1,0))*d;
-        return true;
-    }
-};
-
-template<typename _Tp> struct Matx_FastSolveOp<_Tp, 3, 3, 1>
-{
-    bool operator()(const Matx<_Tp, 3, 3>& a, const Matx<_Tp, 3, 1>& b,
-                    Matx<_Tp, 3, 1>& x, int) const
-    {
-        _Tp d = (_Tp)determinant(a);
-        if (d == 0)
-            return false;
-        d = 1/d;
-        x(0) = d*(b(0)*(a(1,1)*a(2,2) - a(1,2)*a(2,1)) -
-                a(0,1)*(b(1)*a(2,2) - a(1,2)*b(2)) +
-                a(0,2)*(b(1)*a(2,1) - a(1,1)*b(2)));
-
-        x(1) = d*(a(0,0)*(b(1)*a(2,2) - a(1,2)*b(2)) -
-                b(0)*(a(1,0)*a(2,2) - a(1,2)*a(2,0)) +
-                a(0,2)*(a(1,0)*b(2) - b(1)*a(2,0)));
-
-        x(2) = d*(a(0,0)*(a(1,1)*b(2) - b(1)*a(2,1)) -
-                a(0,1)*(a(1,0)*b(2) - b(1)*a(2,0)) +
-                b(0)*(a(1,0)*a(2,1) - a(1,1)*a(2,0)));
-        return true;
-    }
-};
-
-} // internal
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp,m,n> Matx<_Tp,m,n>::randu(_Tp a, _Tp b)
-{
-    Matx<_Tp,m,n> M;
-    cv::randu(M, Scalar(a), Scalar(b));
-    return M;
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp,m,n> Matx<_Tp,m,n>::randn(_Tp a, _Tp b)
-{
-    Matx<_Tp,m,n> M;
-    cv::randn(M, Scalar(a), Scalar(b));
-    return M;
-}
-
-template<typename _Tp, int m, int n> inline
-Matx<_Tp, n, m> Matx<_Tp, m, n>::inv(int method, bool *p_is_ok /*= NULL*/) const
-{
-    Matx<_Tp, n, m> b;
-    bool ok = cv::internal::Matx_FastInvOp<_Tp, m, n>()(*this, b, method);
-    if (p_is_ok) *p_is_ok = ok;
-    return ok ? b : Matx<_Tp, n, m>::zeros();
-}
-
-template<typename _Tp, int m, int n> template<int l> inline
-Matx<_Tp, n, l> Matx<_Tp, m, n>::solve(const Matx<_Tp, m, l>& rhs, int method) const
-{
-    Matx<_Tp, n, l> x;
-    bool ok = cv::internal::Matx_FastSolveOp<_Tp, m, n, l>()(*this, rhs, x, method);
-    return ok ? x : Matx<_Tp, n, l>::zeros();
-}
-
-
-
-////////////////////////// Augmenting algebraic & logical operations //////////////////////////
-
-#define CV_MAT_AUG_OPERATOR1(op, cvop, A, B) \
-    static inline A& operator op (A& a, const B& b) { cvop; return a; }
-
-#define CV_MAT_AUG_OPERATOR(op, cvop, A, B)   \
-    CV_MAT_AUG_OPERATOR1(op, cvop, A, B)      \
-    CV_MAT_AUG_OPERATOR1(op, cvop, const A, B)
-
-#define CV_MAT_AUG_OPERATOR_T(op, cvop, A, B)                   \
-    template<typename _Tp> CV_MAT_AUG_OPERATOR1(op, cvop, A, B) \
-    template<typename _Tp> CV_MAT_AUG_OPERATOR1(op, cvop, const A, B)
-
-#define CV_MAT_AUG_OPERATOR_TN(op, cvop, A)                                \
-    template<typename _Tp, int m, int n> static inline A& operator op (A& a, const Matx<_Tp,m,n>& b) { cvop; return a; } \
-    template<typename _Tp, int m, int n> static inline const A& operator op (const A& a, const Matx<_Tp,m,n>& b) { cvop; return a; }
-
-CV_MAT_AUG_OPERATOR  (+=, cv::add(a, b, (const Mat&)a), Mat, Mat)
-CV_MAT_AUG_OPERATOR  (+=, cv::add(a, b, (const Mat&)a), Mat, Scalar)
-CV_MAT_AUG_OPERATOR_T(+=, cv::add(a, b, (const Mat&)a), Mat_<_Tp>, Mat)
-CV_MAT_AUG_OPERATOR_T(+=, cv::add(a, b, (const Mat&)a), Mat_<_Tp>, Scalar)
-CV_MAT_AUG_OPERATOR_T(+=, cv::add(a, b, (const Mat&)a), Mat_<_Tp>, Mat_<_Tp>)
-CV_MAT_AUG_OPERATOR_TN(+=, cv::add(a, Mat(b), (const Mat&)a), Mat)
-CV_MAT_AUG_OPERATOR_TN(+=, cv::add(a, Mat(b), (const Mat&)a), Mat_<_Tp>)
-
-CV_MAT_AUG_OPERATOR  (-=, cv::subtract(a, b, (const Mat&)a), Mat, Mat)
-CV_MAT_AUG_OPERATOR  (-=, cv::subtract(a, b, (const Mat&)a), Mat, Scalar)
-CV_MAT_AUG_OPERATOR_T(-=, cv::subtract(a, b, (const Mat&)a), Mat_<_Tp>, Mat)
-CV_MAT_AUG_OPERATOR_T(-=, cv::subtract(a, b, (const Mat&)a), Mat_<_Tp>, Scalar)
-CV_MAT_AUG_OPERATOR_T(-=, cv::subtract(a, b, (const Mat&)a), Mat_<_Tp>, Mat_<_Tp>)
-CV_MAT_AUG_OPERATOR_TN(-=, cv::subtract(a, Mat(b), (const Mat&)a), Mat)
-CV_MAT_AUG_OPERATOR_TN(-=, cv::subtract(a, Mat(b), (const Mat&)a), Mat_<_Tp>)
-
-CV_MAT_AUG_OPERATOR  (*=, cv::gemm(a, b, 1, Mat(), 0, a, 0), Mat, Mat)
-CV_MAT_AUG_OPERATOR_T(*=, cv::gemm(a, b, 1, Mat(), 0, a, 0), Mat_<_Tp>, Mat)
-CV_MAT_AUG_OPERATOR_T(*=, cv::gemm(a, b, 1, Mat(), 0, a, 0), Mat_<_Tp>, Mat_<_Tp>)
-CV_MAT_AUG_OPERATOR  (*=, a.convertTo(a, -1, b), Mat, double)
-CV_MAT_AUG_OPERATOR_T(*=, a.convertTo(a, -1, b), Mat_<_Tp>, double)
-CV_MAT_AUG_OPERATOR_TN(*=, cv::gemm(a, Mat(b), 1, Mat(), 0, a, 0), Mat)
-CV_MAT_AUG_OPERATOR_TN(*=, cv::gemm(a, Mat(b), 1, Mat(), 0, a, 0), Mat_<_Tp>)
-
-CV_MAT_AUG_OPERATOR  (/=, cv::divide(a, b, (const Mat&)a), Mat, Mat)
-CV_MAT_AUG_OPERATOR_T(/=, cv::divide(a, b, (const Mat&)a), Mat_<_Tp>, Mat)
-CV_MAT_AUG_OPERATOR_T(/=, cv::divide(a, b, (const Mat&)a), Mat_<_Tp>, Mat_<_Tp>)
-CV_MAT_AUG_OPERATOR  (/=, a.convertTo((Mat&)a, -1, 1./b), Mat, double)
-CV_MAT_AUG_OPERATOR_T(/=, a.convertTo((Mat&)a, -1, 1./b), Mat_<_Tp>, double)
-CV_MAT_AUG_OPERATOR_TN(/=, cv::divide(a, Mat(b), (const Mat&)a), Mat)
-CV_MAT_AUG_OPERATOR_TN(/=, cv::divide(a, Mat(b), (const Mat&)a), Mat_<_Tp>)
-
-CV_MAT_AUG_OPERATOR  (&=, cv::bitwise_and(a, b, (const Mat&)a), Mat, Mat)
-CV_MAT_AUG_OPERATOR  (&=, cv::bitwise_and(a, b, (const Mat&)a), Mat, Scalar)
-CV_MAT_AUG_OPERATOR_T(&=, cv::bitwise_and(a, b, (const Mat&)a), Mat_<_Tp>, Mat)
-CV_MAT_AUG_OPERATOR_T(&=, cv::bitwise_and(a, b, (const Mat&)a), Mat_<_Tp>, Scalar)
-CV_MAT_AUG_OPERATOR_T(&=, cv::bitwise_and(a, b, (const Mat&)a), Mat_<_Tp>, Mat_<_Tp>)
-CV_MAT_AUG_OPERATOR_TN(&=, cv::bitwise_and(a, Mat(b), (const Mat&)a), Mat)
-CV_MAT_AUG_OPERATOR_TN(&=, cv::bitwise_and(a, Mat(b), (const Mat&)a), Mat_<_Tp>)
-
-CV_MAT_AUG_OPERATOR  (|=, cv::bitwise_or(a, b, (const Mat&)a), Mat, Mat)
-CV_MAT_AUG_OPERATOR  (|=, cv::bitwise_or(a, b, (const Mat&)a), Mat, Scalar)
-CV_MAT_AUG_OPERATOR_T(|=, cv::bitwise_or(a, b, (const Mat&)a), Mat_<_Tp>, Mat)
-CV_MAT_AUG_OPERATOR_T(|=, cv::bitwise_or(a, b, (const Mat&)a), Mat_<_Tp>, Scalar)
-CV_MAT_AUG_OPERATOR_T(|=, cv::bitwise_or(a, b, (const Mat&)a), Mat_<_Tp>, Mat_<_Tp>)
-CV_MAT_AUG_OPERATOR_TN(|=, cv::bitwise_or(a, Mat(b), (const Mat&)a), Mat)
-CV_MAT_AUG_OPERATOR_TN(|=, cv::bitwise_or(a, Mat(b), (const Mat&)a), Mat_<_Tp>)
-
-CV_MAT_AUG_OPERATOR  (^=, cv::bitwise_xor(a, b, (const Mat&)a), Mat, Mat)
-CV_MAT_AUG_OPERATOR  (^=, cv::bitwise_xor(a, b, (const Mat&)a), Mat, Scalar)
-CV_MAT_AUG_OPERATOR_T(^=, cv::bitwise_xor(a, b, (const Mat&)a), Mat_<_Tp>, Mat)
-CV_MAT_AUG_OPERATOR_T(^=, cv::bitwise_xor(a, b, (const Mat&)a), Mat_<_Tp>, Scalar)
-CV_MAT_AUG_OPERATOR_T(^=, cv::bitwise_xor(a, b, (const Mat&)a), Mat_<_Tp>, Mat_<_Tp>)
-CV_MAT_AUG_OPERATOR_TN(^=, cv::bitwise_xor(a, Mat(b), (const Mat&)a), Mat)
-CV_MAT_AUG_OPERATOR_TN(^=, cv::bitwise_xor(a, Mat(b), (const Mat&)a), Mat_<_Tp>)
-
-#undef CV_MAT_AUG_OPERATOR_TN
-#undef CV_MAT_AUG_OPERATOR_T
-#undef CV_MAT_AUG_OPERATOR
-#undef CV_MAT_AUG_OPERATOR1
-
-
-
-///////////////////////////////////////////// SVD /////////////////////////////////////////////
-
-inline SVD::SVD() {}
-inline SVD::SVD( InputArray m, int flags ) { operator ()(m, flags); }
-inline void SVD::solveZ( InputArray m, OutputArray _dst )
-{
-    Mat mtx = m.getMat();
-    SVD svd(mtx, (mtx.rows >= mtx.cols ? 0 : SVD::FULL_UV));
-    _dst.create(svd.vt.cols, 1, svd.vt.type());
-    Mat dst = _dst.getMat();
-    svd.vt.row(svd.vt.rows-1).reshape(1,svd.vt.cols).copyTo(dst);
-}
-
-template<typename _Tp, int m, int n, int nm> inline void
-    SVD::compute( const Matx<_Tp, m, n>& a, Matx<_Tp, nm, 1>& w, Matx<_Tp, m, nm>& u, Matx<_Tp, n, nm>& vt )
-{
-    CV_StaticAssert( nm == MIN(m, n), "Invalid size of output vector.");
-    Mat _a(a, false), _u(u, false), _w(w, false), _vt(vt, false);
-    SVD::compute(_a, _w, _u, _vt);
-    CV_Assert(_w.data == (uchar*)&w.val[0] && _u.data == (uchar*)&u.val[0] && _vt.data == (uchar*)&vt.val[0]);
-}
-
-template<typename _Tp, int m, int n, int nm> inline void
-SVD::compute( const Matx<_Tp, m, n>& a, Matx<_Tp, nm, 1>& w )
-{
-    CV_StaticAssert( nm == MIN(m, n), "Invalid size of output vector.");
-    Mat _a(a, false), _w(w, false);
-    SVD::compute(_a, _w);
-    CV_Assert(_w.data == (uchar*)&w.val[0]);
-}
-
-template<typename _Tp, int m, int n, int nm, int nb> inline void
-SVD::backSubst( const Matx<_Tp, nm, 1>& w, const Matx<_Tp, m, nm>& u,
-                const Matx<_Tp, n, nm>& vt, const Matx<_Tp, m, nb>& rhs,
-                Matx<_Tp, n, nb>& dst )
-{
-    CV_StaticAssert( nm == MIN(m, n), "Invalid size of output vector.");
-    Mat _u(u, false), _w(w, false), _vt(vt, false), _rhs(rhs, false), _dst(dst, false);
-    SVD::backSubst(_w, _u, _vt, _rhs, _dst);
-    CV_Assert(_dst.data == (uchar*)&dst.val[0]);
-}
-
-
-
-/////////////////////////////////// Multiply-with-Carry RNG ///////////////////////////////////
-
-inline RNG::RNG()              { state = 0xffffffff; }
-inline RNG::RNG(uint64 _state) { state = _state ? _state : 0xffffffff; }
-
-inline RNG::operator uchar()    { return (uchar)next(); }
-inline RNG::operator schar()    { return (schar)next(); }
-inline RNG::operator ushort()   { return (ushort)next(); }
-inline RNG::operator short()    { return (short)next(); }
-inline RNG::operator int()      { return (int)next(); }
-inline RNG::operator unsigned() { return next(); }
-inline RNG::operator float()    { return next()*2.3283064365386962890625e-10f; }
-inline RNG::operator double()   { unsigned t = next(); return (((uint64)t << 32) | next()) * 5.4210108624275221700372640043497e-20; }
-
-inline unsigned RNG::operator ()(unsigned N) { return (unsigned)uniform(0,N); }
-inline unsigned RNG::operator ()()           { return next(); }
-
-inline int    RNG::uniform(int a, int b)       { return a == b ? a : (int)(next() % (b - a) + a); }
-inline float  RNG::uniform(float a, float b)   { return ((float)*this)*(b - a) + a; }
-inline double RNG::uniform(double a, double b) { return ((double)*this)*(b - a) + a; }
-
-inline bool RNG::operator ==(const RNG& other) const { return state == other.state; }
-
-inline unsigned RNG::next()
-{
-    state = (uint64)(unsigned)state* /*CV_RNG_COEFF*/ 4164903690U + (unsigned)(state >> 32);
-    return (unsigned)state;
-}
-
-//! returns the next uniformly-distributed random number of the specified type
-template<typename _Tp> static inline _Tp randu()
-{
-  return (_Tp)theRNG();
-}
-
-///////////////////////////////// Formatted string generation /////////////////////////////////
-
-/** @brief Returns a text string formatted using the printf-like expression.
-
-The function acts like sprintf but forms and returns an STL string. It can be used to form an error
-message in the Exception constructor.
-@param fmt printf-compatible formatting specifiers.
-
-**Note**:
-|Type|Specifier|
-|-|-|
-|`const char*`|`%s`|
-|`char`|`%c`|
-|`float` / `double`|`%f`,`%g`|
-|`int`, `long`, `long long`|`%d`, `%ld`, ``%lld`|
-|`unsigned`, `unsigned long`, `unsigned long long`|`%u`, `%lu`, `%llu`|
-|`uint64` -> `uintmax_t`, `int64` -> `intmax_t`|`%ju`, `%jd`|
-|`size_t`|`%zu`|
- */
-CV_EXPORTS String format( const char* fmt, ... ) CV_FORMAT_PRINTF(1, 2);
-
-///////////////////////////////// Formatted output of cv::Mat /////////////////////////////////
-
-static inline
-Ptr<Formatted> format(InputArray mtx, Formatter::FormatType fmt)
-{
-    return Formatter::get(fmt)->format(mtx.getMat());
-}
-
-static inline
-int print(Ptr<Formatted> fmtd, FILE* stream = stdout)
-{
-    int written = 0;
-    fmtd->reset();
-    for(const char* str = fmtd->next(); str; str = fmtd->next())
-        written += fputs(str, stream);
-
-    return written;
-}
-
-static inline
-int print(const Mat& mtx, FILE* stream = stdout)
-{
-    return print(Formatter::get()->format(mtx), stream);
-}
-
-static inline
-int print(const UMat& mtx, FILE* stream = stdout)
-{
-    return print(Formatter::get()->format(mtx.getMat(ACCESS_READ)), stream);
-}
-
-template<typename _Tp> static inline
-int print(const std::vector<Point_<_Tp> >& vec, FILE* stream = stdout)
-{
-    return print(Formatter::get()->format(Mat(vec)), stream);
-}
-
-template<typename _Tp> static inline
-int print(const std::vector<Point3_<_Tp> >& vec, FILE* stream = stdout)
-{
-    return print(Formatter::get()->format(Mat(vec)), stream);
-}
-
-template<typename _Tp, int m, int n> static inline
-int print(const Matx<_Tp, m, n>& matx, FILE* stream = stdout)
-{
-    return print(Formatter::get()->format(cv::Mat(matx)), stream);
-}
-
-//! @endcond
-
-/****************************************************************************************\
-*                                  Auxiliary algorithms                                  *
-\****************************************************************************************/
-
-/** @brief Splits an element set into equivalency classes.
-
-The generic function partition implements an \f$O(N^2)\f$ algorithm for splitting a set of \f$N\f$ elements
-into one or more equivalency classes, as described in
-<http://en.wikipedia.org/wiki/Disjoint-set_data_structure> . The function returns the number of
-equivalency classes.
-@param _vec Set of elements stored as a vector.
-@param labels Output vector of labels. It contains as many elements as vec. Each label labels[i] is
-a 0-based cluster index of `vec[i]`.
-@param predicate Equivalence predicate (pointer to a boolean function of two arguments or an
-instance of the class that has the method bool operator()(const _Tp& a, const _Tp& b) ). The
-predicate returns true when the elements are certainly in the same class, and returns false if they
-may or may not be in the same class.
-@ingroup core_cluster
-*/
-template<typename _Tp, class _EqPredicate> int
-partition( const std::vector<_Tp>& _vec, std::vector<int>& labels,
-          _EqPredicate predicate=_EqPredicate())
-{
-    int i, j, N = (int)_vec.size();
-    const _Tp* vec = &_vec[0];
-
-    const int PARENT=0;
-    const int RANK=1;
-
-    std::vector<int> _nodes(N*2);
-    int (*nodes)[2] = (int(*)[2])&_nodes[0];
-
-    // The first O(N) pass: create N single-vertex trees
-    for(i = 0; i < N; i++)
-    {
-        nodes[i][PARENT]=-1;
-        nodes[i][RANK] = 0;
-    }
-
-    // The main O(N^2) pass: merge connected components
-    for( i = 0; i < N; i++ )
-    {
-        int root = i;
-
-        // find root
-        while( nodes[root][PARENT] >= 0 )
-            root = nodes[root][PARENT];
-
-        for( j = 0; j < N; j++ )
-        {
-            if( i == j || !predicate(vec[i], vec[j]))
-                continue;
-            int root2 = j;
-
-            while( nodes[root2][PARENT] >= 0 )
-                root2 = nodes[root2][PARENT];
-
-            if( root2 != root )
-            {
-                // unite both trees
-                int rank = nodes[root][RANK], rank2 = nodes[root2][RANK];
-                if( rank > rank2 )
-                    nodes[root2][PARENT] = root;
-                else
-                {
-                    nodes[root][PARENT] = root2;
-                    nodes[root2][RANK] += rank == rank2;
-                    root = root2;
-                }
-                CV_Assert( nodes[root][PARENT] < 0 );
-
-                int k = j, parent;
-
-                // compress the path from node2 to root
-                while( (parent = nodes[k][PARENT]) >= 0 )
-                {
-                    nodes[k][PARENT] = root;
-                    k = parent;
-                }
-
-                // compress the path from node to root
-                k = i;
-                while( (parent = nodes[k][PARENT]) >= 0 )
-                {
-                    nodes[k][PARENT] = root;
-                    k = parent;
-                }
-            }
-        }
-    }
-
-    // Final O(N) pass: enumerate classes
-    labels.resize(N);
-    int nclasses = 0;
-
-    for( i = 0; i < N; i++ )
-    {
-        int root = i;
-        while( nodes[root][PARENT] >= 0 )
-            root = nodes[root][PARENT];
-        // re-use the rank as the class label
-        if( nodes[root][RANK] >= 0 )
-            nodes[root][RANK] = ~nclasses++;
-        labels[i] = ~nodes[root][RANK];
-    }
-
-    return nclasses;
-}
-
-} // cv
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/optim.hpp b/projects/llm_framework/include/opencv4/opencv2/core/optim.hpp
deleted file mode 100644
index f61a2b94..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/optim.hpp
+++ /dev/null
@@ -1,302 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the OpenCV Foundation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_OPTIM_HPP
-#define OPENCV_OPTIM_HPP
-
-#include "opencv2/core.hpp"
-
-namespace cv
-{
-
-/** @addtogroup core_optim
-The algorithms in this section minimize or maximize function value within specified constraints or
-without any constraints.
-@{
-*/
-
-/** @brief Basic interface for all solvers
- */
-class CV_EXPORTS MinProblemSolver : public Algorithm
-{
-public:
-    /** @brief Represents function being optimized
-     */
-    class CV_EXPORTS Function
-    {
-    public:
-        virtual ~Function() {}
-        virtual int getDims() const = 0;
-        virtual double getGradientEps() const;
-        virtual double calc(const double* x) const = 0;
-        virtual void getGradient(const double* x,double* grad);
-    };
-
-    /** @brief Getter for the optimized function.
-
-    The optimized function is represented by Function interface, which requires derivatives to
-    implement the calc(double*) and getDim() methods to evaluate the function.
-
-    @return Smart-pointer to an object that implements Function interface - it represents the
-    function that is being optimized. It can be empty, if no function was given so far.
-     */
-    virtual Ptr<Function> getFunction() const = 0;
-
-    /** @brief Setter for the optimized function.
-
-    *It should be called at least once before the call to* minimize(), as default value is not usable.
-
-    @param f The new function to optimize.
-     */
-    virtual void setFunction(const Ptr<Function>& f) = 0;
-
-    /** @brief Getter for the previously set terminal criteria for this algorithm.
-
-    @return Deep copy of the terminal criteria used at the moment.
-     */
-    virtual TermCriteria getTermCriteria() const = 0;
-
-    /** @brief Set terminal criteria for solver.
-
-    This method *is not necessary* to be called before the first call to minimize(), as the default
-    value is sensible.
-
-    Algorithm stops when the number of function evaluations done exceeds termcrit.maxCount, when
-    the function values at the vertices of simplex are within termcrit.epsilon range or simplex
-    becomes so small that it can enclosed in a box with termcrit.epsilon sides, whatever comes
-    first.
-    @param termcrit Terminal criteria to be used, represented as cv::TermCriteria structure.
-     */
-    virtual void setTermCriteria(const TermCriteria& termcrit) = 0;
-
-    /** @brief actually runs the algorithm and performs the minimization.
-
-    The sole input parameter determines the centroid of the starting simplex (roughly, it tells
-    where to start), all the others (terminal criteria, initial step, function to be minimized) are
-    supposed to be set via the setters before the call to this method or the default values (not
-    always sensible) will be used.
-
-    @param x The initial point, that will become a centroid of an initial simplex. After the algorithm
-    will terminate, it will be set to the point where the algorithm stops, the point of possible
-    minimum.
-    @return The value of a function at the point found.
-     */
-    virtual double minimize(InputOutputArray x) = 0;
-};
-
-/** @brief This class is used to perform the non-linear non-constrained minimization of a function,
-
-defined on an `n`-dimensional Euclidean space, using the **Nelder-Mead method**, also known as
-**downhill simplex method**. The basic idea about the method can be obtained from
-<http://en.wikipedia.org/wiki/Nelder-Mead_method>.
-
-It should be noted, that this method, although deterministic, is rather a heuristic and therefore
-may converge to a local minima, not necessary a global one. It is iterative optimization technique,
-which at each step uses an information about the values of a function evaluated only at `n+1`
-points, arranged as a *simplex* in `n`-dimensional space (hence the second name of the method). At
-each step new point is chosen to evaluate function at, obtained value is compared with previous
-ones and based on this information simplex changes it's shape , slowly moving to the local minimum.
-Thus this method is using *only* function values to make decision, on contrary to, say, Nonlinear
-Conjugate Gradient method (which is also implemented in optim).
-
-Algorithm stops when the number of function evaluations done exceeds termcrit.maxCount, when the
-function values at the vertices of simplex are within termcrit.epsilon range or simplex becomes so
-small that it can enclosed in a box with termcrit.epsilon sides, whatever comes first, for some
-defined by user positive integer termcrit.maxCount and positive non-integer termcrit.epsilon.
-
-@note DownhillSolver is a derivative of the abstract interface
-cv::MinProblemSolver, which in turn is derived from the Algorithm interface and is used to
-encapsulate the functionality, common to all non-linear optimization algorithms in the optim
-module.
-
-@note term criteria should meet following condition:
-@code
-    termcrit.type == (TermCriteria::MAX_ITER + TermCriteria::EPS) && termcrit.epsilon > 0 && termcrit.maxCount > 0
-@endcode
- */
-class CV_EXPORTS DownhillSolver : public MinProblemSolver
-{
-public:
-    /** @brief Returns the initial step that will be used in downhill simplex algorithm.
-
-    @param step Initial step that will be used in algorithm. Note, that although corresponding setter
-    accepts column-vectors as well as row-vectors, this method will return a row-vector.
-    @see DownhillSolver::setInitStep
-     */
-    virtual void getInitStep(OutputArray step) const=0;
-
-    /** @brief Sets the initial step that will be used in downhill simplex algorithm.
-
-    Step, together with initial point (given in DownhillSolver::minimize) are two `n`-dimensional
-    vectors that are used to determine the shape of initial simplex. Roughly said, initial point
-    determines the position of a simplex (it will become simplex's centroid), while step determines the
-    spread (size in each dimension) of a simplex. To be more precise, if \f$s,x_0\in\mathbb{R}^n\f$ are
-    the initial step and initial point respectively, the vertices of a simplex will be:
-    \f$v_0:=x_0-\frac{1}{2} s\f$ and \f$v_i:=x_0+s_i\f$ for \f$i=1,2,\dots,n\f$ where \f$s_i\f$ denotes
-    projections of the initial step of *n*-th coordinate (the result of projection is treated to be
-    vector given by \f$s_i:=e_i\cdot\left<e_i\cdot s\right>\f$, where \f$e_i\f$ form canonical basis)
-
-    @param step Initial step that will be used in algorithm. Roughly said, it determines the spread
-    (size in each dimension) of an initial simplex.
-     */
-    virtual void setInitStep(InputArray step)=0;
-
-    /** @brief This function returns the reference to the ready-to-use DownhillSolver object.
-
-    All the parameters are optional, so this procedure can be called even without parameters at
-    all. In this case, the default values will be used. As default value for terminal criteria are
-    the only sensible ones, MinProblemSolver::setFunction() and DownhillSolver::setInitStep()
-    should be called upon the obtained object, if the respective parameters were not given to
-    create(). Otherwise, the two ways (give parameters to createDownhillSolver() or miss them out
-    and call the MinProblemSolver::setFunction() and DownhillSolver::setInitStep()) are absolutely
-    equivalent (and will drop the same errors in the same way, should invalid input be detected).
-    @param f Pointer to the function that will be minimized, similarly to the one you submit via
-    MinProblemSolver::setFunction.
-    @param initStep Initial step, that will be used to construct the initial simplex, similarly to the one
-    you submit via MinProblemSolver::setInitStep.
-    @param termcrit Terminal criteria to the algorithm, similarly to the one you submit via
-    MinProblemSolver::setTermCriteria.
-     */
-    static Ptr<DownhillSolver> create(const Ptr<MinProblemSolver::Function>& f=Ptr<MinProblemSolver::Function>(),
-                                      InputArray initStep=Mat_<double>(1,1,0.0),
-                                      TermCriteria termcrit=TermCriteria(TermCriteria::MAX_ITER+TermCriteria::EPS,5000,0.000001));
-};
-
-/** @brief This class is used to perform the non-linear non-constrained minimization of a function
-with known gradient,
-
-defined on an *n*-dimensional Euclidean space, using the **Nonlinear Conjugate Gradient method**.
-The implementation was done based on the beautifully clear explanatory article [An Introduction to
-the Conjugate Gradient Method Without the Agonizing
-Pain](http://www.cs.cmu.edu/~quake-papers/painless-conjugate-gradient.pdf) by Jonathan Richard
-Shewchuk. The method can be seen as an adaptation of a standard Conjugate Gradient method (see, for
-example <http://en.wikipedia.org/wiki/Conjugate_gradient_method>) for numerically solving the
-systems of linear equations.
-
-It should be noted, that this method, although deterministic, is rather a heuristic method and
-therefore may converge to a local minima, not necessary a global one. What is even more disastrous,
-most of its behaviour is ruled by gradient, therefore it essentially cannot distinguish between
-local minima and maxima. Therefore, if it starts sufficiently near to the local maximum, it may
-converge to it. Another obvious restriction is that it should be possible to compute the gradient of
-a function at any point, thus it is preferable to have analytic expression for gradient and
-computational burden should be born by the user.
-
-The latter responsibility is accomplished via the getGradient method of a
-MinProblemSolver::Function interface (which represents function being optimized). This method takes
-point a point in *n*-dimensional space (first argument represents the array of coordinates of that
-point) and compute its gradient (it should be stored in the second argument as an array).
-
-@note class ConjGradSolver thus does not add any new methods to the basic MinProblemSolver interface.
-
-@note term criteria should meet following condition:
-@code
-    termcrit.type == (TermCriteria::MAX_ITER + TermCriteria::EPS) && termcrit.epsilon > 0 && termcrit.maxCount > 0
-    // or
-    termcrit.type == TermCriteria::MAX_ITER) && termcrit.maxCount > 0
-@endcode
- */
-class CV_EXPORTS ConjGradSolver : public MinProblemSolver
-{
-public:
-    /** @brief This function returns the reference to the ready-to-use ConjGradSolver object.
-
-    All the parameters are optional, so this procedure can be called even without parameters at
-    all. In this case, the default values will be used. As default value for terminal criteria are
-    the only sensible ones, MinProblemSolver::setFunction() should be called upon the obtained
-    object, if the function was not given to create(). Otherwise, the two ways (submit it to
-    create() or miss it out and call the MinProblemSolver::setFunction()) are absolutely equivalent
-    (and will drop the same errors in the same way, should invalid input be detected).
-    @param f Pointer to the function that will be minimized, similarly to the one you submit via
-    MinProblemSolver::setFunction.
-    @param termcrit Terminal criteria to the algorithm, similarly to the one you submit via
-    MinProblemSolver::setTermCriteria.
-    */
-    static Ptr<ConjGradSolver> create(const Ptr<MinProblemSolver::Function>& f=Ptr<ConjGradSolver::Function>(),
-                                      TermCriteria termcrit=TermCriteria(TermCriteria::MAX_ITER+TermCriteria::EPS,5000,0.000001));
-};
-
-//! return codes for cv::solveLP() function
-enum SolveLPResult
-{
-    SOLVELP_UNBOUNDED    = -2, //!< problem is unbounded (target function can achieve arbitrary high values)
-    SOLVELP_UNFEASIBLE    = -1, //!< problem is unfeasible (there are no points that satisfy all the constraints imposed)
-    SOLVELP_SINGLE    = 0, //!< there is only one maximum for target function
-    SOLVELP_MULTI    = 1 //!< there are multiple maxima for target function - the arbitrary one is returned
-};
-
-/** @brief Solve given (non-integer) linear programming problem using the Simplex Algorithm (Simplex Method).
-
-What we mean here by "linear programming problem" (or LP problem, for short) can be formulated as:
-
-\f[\mbox{Maximize } c\cdot x\\
- \mbox{Subject to:}\\
- Ax\leq b\\
- x\geq 0\f]
-
-Where \f$c\f$ is fixed `1`-by-`n` row-vector, \f$A\f$ is fixed `m`-by-`n` matrix, \f$b\f$ is fixed `m`-by-`1`
-column vector and \f$x\f$ is an arbitrary `n`-by-`1` column vector, which satisfies the constraints.
-
-Simplex algorithm is one of many algorithms that are designed to handle this sort of problems
-efficiently. Although it is not optimal in theoretical sense (there exist algorithms that can solve
-any problem written as above in polynomial time, while simplex method degenerates to exponential
-time for some special cases), it is well-studied, easy to implement and is shown to work well for
-real-life purposes.
-
-The particular implementation is taken almost verbatim from **Introduction to Algorithms, third
-edition** by T. H. Cormen, C. E. Leiserson, R. L. Rivest and Clifford Stein. In particular, the
-Bland's rule <http://en.wikipedia.org/wiki/Bland%27s_rule> is used to prevent cycling.
-
-@param Func This row-vector corresponds to \f$c\f$ in the LP problem formulation (see above). It should
-contain 32- or 64-bit floating point numbers. As a convenience, column-vector may be also submitted,
-in the latter case it is understood to correspond to \f$c^T\f$.
-@param Constr `m`-by-`n+1` matrix, whose rightmost column corresponds to \f$b\f$ in formulation above
-and the remaining to \f$A\f$. It should contain 32- or 64-bit floating point numbers.
-@param z The solution will be returned here as a column-vector - it corresponds to \f$c\f$ in the
-formulation above. It will contain 64-bit floating point numbers.
-@return One of cv::SolveLPResult
- */
-CV_EXPORTS_W int solveLP(InputArray Func, InputArray Constr, OutputArray z);
-
-//! @}
-
-}// cv
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/ovx.hpp b/projects/llm_framework/include/opencv4/opencv2/core/ovx.hpp
deleted file mode 100644
index 8bb7d549..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/ovx.hpp
+++ /dev/null
@@ -1,28 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-// Copyright (C) 2016, Intel Corporation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-
-// OpenVX related definitions and declarations
-
-#pragma once
-#ifndef OPENCV_OVX_HPP
-#define OPENCV_OVX_HPP
-
-#include "cvdef.h"
-
-namespace cv
-{
-/// Check if use of OpenVX is possible
-CV_EXPORTS_W bool haveOpenVX();
-
-/// Check if use of OpenVX is enabled
-CV_EXPORTS_W bool useOpenVX();
-
-/// Enable/disable use of OpenVX
-CV_EXPORTS_W void setUseOpenVX(bool flag);
-} // namespace cv
-
-#endif // OPENCV_OVX_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/parallel/backend/parallel_for.openmp.hpp b/projects/llm_framework/include/opencv4/opencv2/core/parallel/backend/parallel_for.openmp.hpp
deleted file mode 100644
index b172cac3..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/parallel/backend/parallel_for.openmp.hpp
+++ /dev/null
@@ -1,72 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_CORE_PARALLEL_FOR_OPENMP_HPP
-#define OPENCV_CORE_PARALLEL_FOR_OPENMP_HPP
-
-#include "opencv2/core/parallel/parallel_backend.hpp"
-
-#if !defined(_OPENMP) && !defined(OPENCV_SKIP_OPENMP_PRESENSE_CHECK)
-#error "This file must be compiled with enabled OpenMP"
-#endif
-
-#include <omp.h>
-
-namespace cv { namespace parallel { namespace openmp {
-
-/** OpenMP parallel_for API implementation
- *
- * @sa setParallelForBackend
- * @ingroup core_parallel_backend
- */
-class ParallelForBackend : public ParallelForAPI
-{
-protected:
-    int numThreads;
-    int numThreadsMax;
-public:
-    ParallelForBackend()
-    {
-        numThreads = 0;
-        numThreadsMax = omp_get_max_threads();
-    }
-
-    virtual ~ParallelForBackend() {}
-
-    virtual void parallel_for(int tasks, FN_parallel_for_body_cb_t body_callback, void* callback_data) CV_OVERRIDE
-    {
-#pragma omp parallel for schedule(dynamic) num_threads(numThreads > 0 ? numThreads : numThreadsMax)
-        for (int i = 0; i < tasks; ++i)
-            body_callback(i, i + 1, callback_data);
-    }
-
-    virtual int getThreadNum() const CV_OVERRIDE
-    {
-        return omp_get_thread_num();
-    }
-
-    virtual int getNumThreads() const CV_OVERRIDE
-    {
-        return numThreads > 0
-               ? numThreads
-               : numThreadsMax;
-    }
-
-    virtual int setNumThreads(int nThreads) CV_OVERRIDE
-    {
-        int oldNumThreads = numThreads;
-        numThreads = nThreads;
-        // nothing needed as numThreads is used in #pragma omp parallel for directly
-        return oldNumThreads;
-    }
-
-    const char* getName() const CV_OVERRIDE
-    {
-        return "openmp";
-    }
-};
-
-}}}  // namespace
-
-#endif  // OPENCV_CORE_PARALLEL_FOR_OPENMP_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/parallel/backend/parallel_for.tbb.hpp b/projects/llm_framework/include/opencv4/opencv2/core/parallel/backend/parallel_for.tbb.hpp
deleted file mode 100644
index 04b0c4c6..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/parallel/backend/parallel_for.tbb.hpp
+++ /dev/null
@@ -1,153 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_CORE_PARALLEL_FOR_TBB_HPP
-#define OPENCV_CORE_PARALLEL_FOR_TBB_HPP
-
-#include "opencv2/core/parallel/parallel_backend.hpp"
-#include <opencv2/core/utils/logger.hpp>
-
-#ifndef TBB_SUPPRESS_DEPRECATED_MESSAGES  // supress warning
-#define TBB_SUPPRESS_DEPRECATED_MESSAGES 1
-#endif
-#include "tbb/tbb.h"
-#if !defined(TBB_INTERFACE_VERSION)
-#error "Unknows/unsupported TBB version"
-#endif
-
-#if TBB_INTERFACE_VERSION >= 8000
-#include "tbb/task_arena.h"
-#endif
-
-namespace cv { namespace parallel { namespace tbb {
-
-using namespace ::tbb;
-
-#if TBB_INTERFACE_VERSION >= 8000
-static tbb::task_arena& getArena()
-{
-    static tbb::task_arena tbbArena(tbb::task_arena::automatic);
-    return tbbArena;
-}
-#else
-static tbb::task_scheduler_init& getScheduler()
-{
-    static tbb::task_scheduler_init tbbScheduler(tbb::task_scheduler_init::deferred);
-    return tbbScheduler;
-}
-#endif
-
-/** TBB parallel_for API implementation
- *
- * @sa setParallelForBackend
- * @ingroup core_parallel_backend
- */
-class ParallelForBackend : public ParallelForAPI
-{
-protected:
-    int numThreads;
-    int numThreadsMax;
-public:
-    ParallelForBackend()
-    {
-        CV_LOG_INFO(NULL, "Initializing TBB parallel backend: TBB_INTERFACE_VERSION=" << TBB_INTERFACE_VERSION);
-        numThreads = 0;
-#if TBB_INTERFACE_VERSION >= 8000
-        (void)getArena();
-#else
-        (void)getScheduler();
-#endif
-    }
-
-    virtual ~ParallelForBackend() {}
-
-    class CallbackProxy
-    {
-        const FN_parallel_for_body_cb_t& callback;
-        void* const callback_data;
-        const int tasks;
-    public:
-        inline CallbackProxy(int tasks_, FN_parallel_for_body_cb_t& callback_, void* callback_data_)
-            : callback(callback_), callback_data(callback_data_), tasks(tasks_)
-        {
-            // nothing
-        }
-
-        void operator()(const tbb::blocked_range<int>& range) const
-        {
-            this->callback(range.begin(), range.end(), callback_data);
-        }
-
-        void operator()() const
-        {
-            tbb::parallel_for(tbb::blocked_range<int>(0, tasks), *this);
-        }
-    };
-
-    virtual void parallel_for(int tasks, FN_parallel_for_body_cb_t body_callback, void* callback_data) CV_OVERRIDE
-    {
-        CallbackProxy task(tasks, body_callback, callback_data);
-#if TBB_INTERFACE_VERSION >= 8000
-        getArena().execute(task);
-#else
-        task();
-#endif
-    }
-
-    virtual int getThreadNum() const CV_OVERRIDE
-    {
-#if TBB_INTERFACE_VERSION >= 9100
-        return tbb::this_task_arena::current_thread_index();
-#elif TBB_INTERFACE_VERSION >= 8000
-        return tbb::task_arena::current_thread_index();
-#else
-        return 0;
-#endif
-    }
-
-    virtual int getNumThreads() const CV_OVERRIDE
-    {
-#if TBB_INTERFACE_VERSION >= 9100
-    return getArena().max_concurrency();
-#elif TBB_INTERFACE_VERSION >= 8000
-    return numThreads > 0
-        ? numThreads
-        : tbb::task_scheduler_init::default_num_threads();
-#else
-    return getScheduler().is_active()
-           ? numThreads
-           : tbb::task_scheduler_init::default_num_threads();
-#endif
-    }
-
-    virtual int setNumThreads(int nThreads) CV_OVERRIDE
-    {
-        int oldNumThreads = numThreads;
-        numThreads = nThreads;
-
-#if TBB_INTERFACE_VERSION >= 8000
-        auto& tbbArena = getArena();
-        if (tbbArena.is_active())
-            tbbArena.terminate();
-        if (numThreads > 0)
-            tbbArena.initialize(numThreads);
-#else
-        auto& tbbScheduler = getScheduler();
-        if (tbbScheduler.is_active())
-            tbbScheduler.terminate();
-        if (numThreads > 0)
-            tbbScheduler.initialize(numThreads);
-#endif
-        return oldNumThreads;
-    }
-
-    const char* getName() const CV_OVERRIDE
-    {
-        return "tbb";
-    }
-};
-
-}}}  // namespace
-
-#endif  // OPENCV_CORE_PARALLEL_FOR_TBB_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/parallel/parallel_backend.hpp b/projects/llm_framework/include/opencv4/opencv2/core/parallel/parallel_backend.hpp
deleted file mode 100644
index c3e8333c..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/parallel/parallel_backend.hpp
+++ /dev/null
@@ -1,90 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_CORE_PARALLEL_BACKEND_HPP
-#define OPENCV_CORE_PARALLEL_BACKEND_HPP
-
-#include "opencv2/core/cvdef.h"
-#include <memory>
-
-namespace cv { namespace parallel {
-#ifndef CV_API_CALL
-#define CV_API_CALL
-#endif
-
-/** @addtogroup core_parallel_backend
- * @{
- * API below is provided to resolve problem of CPU resource over-subscription by multiple thread pools from different multi-threading frameworks.
- * This is common problem for cases when OpenCV compiled threading framework is different from the Users Applications framework.
- *
- * Applications can replace OpenCV `parallel_for()` backend with own implementation (to reuse Application's thread pool).
- *
- *
- * ### Backend API usage examples
- *
- * #### Intel TBB
- *
- * - include header with simple implementation of TBB backend:
- *   @snippet parallel_backend/example-tbb.cpp tbb_include
- * - execute backend replacement code:
- *   @snippet parallel_backend/example-tbb.cpp tbb_backend
- * - configuration of compiler/linker options is responsibility of Application's scripts
- *
- * #### OpenMP
- *
- * - include header with simple implementation of OpenMP backend:
- *   @snippet parallel_backend/example-openmp.cpp openmp_include
- * - execute backend replacement code:
- *   @snippet parallel_backend/example-openmp.cpp openmp_backend
- * - Configuration of compiler/linker options is responsibility of Application's scripts
- *
- *
- * ### Plugins support
- *
- * Runtime configuration options:
- * - change backend priority: `OPENCV_PARALLEL_PRIORITY_<backend>=9999`
- * - disable backend: `OPENCV_PARALLEL_PRIORITY_<backend>=0`
- * - specify list of backends with high priority (>100000): `OPENCV_PARALLEL_PRIORITY_LIST=TBB,OPENMP`. Unknown backends are registered as new plugins.
- *
- */
-
-/** Interface for parallel_for backends implementations
- *
- * @sa setParallelForBackend
- */
-class CV_EXPORTS ParallelForAPI
-{
-public:
-    virtual ~ParallelForAPI();
-
-    typedef void (CV_API_CALL *FN_parallel_for_body_cb_t)(int start, int end, void* data);
-
-    virtual void parallel_for(int tasks, FN_parallel_for_body_cb_t body_callback, void* callback_data) = 0;
-
-    virtual int getThreadNum() const = 0;
-
-    virtual int getNumThreads() const = 0;
-
-    virtual int setNumThreads(int nThreads) = 0;
-
-    virtual const char* getName() const = 0;
-};
-
-/** @brief Replace OpenCV parallel_for backend
- *
- * Application can replace OpenCV `parallel_for()` backend with own implementation.
- *
- * @note This call is not thread-safe. Consider calling this function from the `main()` before any other OpenCV processing functions (and without any other created threads).
- */
-CV_EXPORTS void setParallelForBackend(const std::shared_ptr<ParallelForAPI>& api, bool propagateNumThreads = true);
-
-/** @brief Change OpenCV parallel_for backend
- *
- * @note This call is not thread-safe. Consider calling this function from the `main()` before any other OpenCV processing functions (and without any other created threads).
- */
-CV_EXPORTS_W bool setParallelForBackend(const std::string& backendName, bool propagateNumThreads = true);
-
-//! @}
-}}  // namespace
-#endif  // OPENCV_CORE_PARALLEL_BACKEND_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/persistence.hpp b/projects/llm_framework/include/opencv4/opencv2/core/persistence.hpp
deleted file mode 100644
index 276f6403..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/persistence.hpp
+++ /dev/null
@@ -1,1350 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_PERSISTENCE_HPP
-#define OPENCV_CORE_PERSISTENCE_HPP
-
-#ifndef CV_DOXYGEN
-/// Define to support persistence legacy formats
-#define CV__LEGACY_PERSISTENCE
-#endif
-
-#ifndef __cplusplus
-#  error persistence.hpp header must be compiled as C++
-#endif
-
-//! @addtogroup core_c
-//! @{
-
-/** @brief "black box" representation of the file storage associated with a file on disk.
-
-Several functions that are described below take CvFileStorage\* as inputs and allow the user to
-save or to load hierarchical collections that consist of scalar values, standard CXCore objects
-(such as matrices, sequences, graphs), and user-defined objects.
-
-OpenCV can read and write data in XML (<http://www.w3c.org/XML>), YAML (<http://www.yaml.org>) or
-JSON (<http://www.json.org/>) formats. Below is an example of 3x3 floating-point identity matrix A,
-stored in XML and YAML files
-using CXCore functions:
-XML:
-@code{.xml}
-    <?xml version="1.0">
-    <opencv_storage>
-    <A type_id="opencv-matrix">
-      <rows>3</rows>
-      <cols>3</cols>
-      <dt>f</dt>
-      <data>1. 0. 0. 0. 1. 0. 0. 0. 1.</data>
-    </A>
-    </opencv_storage>
-@endcode
-YAML:
-@code{.yaml}
-    %YAML:1.0
-    A: !!opencv-matrix
-      rows: 3
-      cols: 3
-      dt: f
-      data: [ 1., 0., 0., 0., 1., 0., 0., 0., 1.]
-@endcode
-As it can be seen from the examples, XML uses nested tags to represent hierarchy, while YAML uses
-indentation for that purpose (similar to the Python programming language).
-
-The same functions can read and write data in both formats; the particular format is determined by
-the extension of the opened file, ".xml" for XML files, ".yml" or ".yaml" for YAML and ".json" for
-JSON.
- */
-
-//! @} core_c
-
-#include "opencv2/core/types.hpp"
-#include "opencv2/core/mat.hpp"
-
-namespace cv {
-
-/** @addtogroup core_xml
-
-XML/YAML/JSON file storages.     {#xml_storage}
-=======================
-Writing to a file storage.
---------------------------
-You can store and then restore various OpenCV data structures to/from XML (<http://www.w3c.org/XML>),
-YAML (<http://www.yaml.org>) or JSON (<http://www.json.org/>) formats. Also, it is possible to store
-and load arbitrarily complex data structures, which include OpenCV data structures, as well as
-primitive data types (integer and floating-point numbers and text strings) as their elements.
-
-Use the following procedure to write something to XML, YAML or JSON:
--# Create new FileStorage and open it for writing. It can be done with a single call to
-FileStorage::FileStorage constructor that takes a filename, or you can use the default constructor
-and then call FileStorage::open. Format of the file (XML, YAML or JSON) is determined from the filename
-extension (".xml", ".yml"/".yaml" and ".json", respectively)
--# Write all the data you want using the streaming operator `<<`, just like in the case of STL
-streams.
--# Close the file using FileStorage::release. FileStorage destructor also closes the file.
-
-Here is an example:
-@code
-    #include "opencv2/core.hpp"
-    #include <time.h>
-
-    using namespace cv;
-
-    int main(int, char** argv)
-    {
-        FileStorage fs("test.yml", FileStorage::WRITE);
-
-        fs << "frameCount" << 5;
-        time_t rawtime; time(&rawtime);
-        fs << "calibrationDate" << asctime(localtime(&rawtime));
-        Mat cameraMatrix = (Mat_<double>(3,3) << 1000, 0, 320, 0, 1000, 240, 0, 0, 1);
-        Mat distCoeffs = (Mat_<double>(5,1) << 0.1, 0.01, -0.001, 0, 0);
-        fs << "cameraMatrix" << cameraMatrix << "distCoeffs" << distCoeffs;
-        fs << "features" << "[";
-        for( int i = 0; i < 3; i++ )
-        {
-            int x = rand() % 640;
-            int y = rand() % 480;
-            uchar lbp = rand() % 256;
-
-            fs << "{:" << "x" << x << "y" << y << "lbp" << "[:";
-            for( int j = 0; j < 8; j++ )
-                fs << ((lbp >> j) & 1);
-            fs << "]" << "}";
-        }
-        fs << "]";
-        fs.release();
-        return 0;
-    }
-@endcode
-The sample above stores to YML an integer, a text string (calibration date), 2 matrices, and a custom
-structure "feature", which includes feature coordinates and LBP (local binary pattern) value. Here
-is output of the sample:
-@code{.yaml}
-%YAML:1.0
-frameCount: 5
-calibrationDate: "Fri Jun 17 14:09:29 2011\n"
-cameraMatrix: !!opencv-matrix
-   rows: 3
-   cols: 3
-   dt: d
-   data: [ 1000., 0., 320., 0., 1000., 240., 0., 0., 1. ]
-distCoeffs: !!opencv-matrix
-   rows: 5
-   cols: 1
-   dt: d
-   data: [ 1.0000000000000001e-01, 1.0000000000000000e-02,
-       -1.0000000000000000e-03, 0., 0. ]
-features:
-   - { x:167, y:49, lbp:[ 1, 0, 0, 1, 1, 0, 1, 1 ] }
-   - { x:298, y:130, lbp:[ 0, 0, 0, 1, 0, 0, 1, 1 ] }
-   - { x:344, y:158, lbp:[ 1, 1, 0, 0, 0, 0, 1, 0 ] }
-@endcode
-
-As an exercise, you can replace ".yml" with ".xml" or ".json" in the sample above and see, how the
-corresponding XML file will look like.
-
-Several things can be noted by looking at the sample code and the output:
-
--   The produced YAML (and XML/JSON) consists of heterogeneous collections that can be nested. There are
-    2 types of collections: named collections (mappings) and unnamed collections (sequences). In mappings
-    each element has a name and is accessed by name. This is similar to structures and std::map in
-    C/C++ and dictionaries in Python. In sequences elements do not have names, they are accessed by
-    indices. This is similar to arrays and std::vector in C/C++ and lists, tuples in Python.
-    "Heterogeneous" means that elements of each single collection can have different types.
-
-    Top-level collection in YAML/XML/JSON is a mapping. Each matrix is stored as a mapping, and the matrix
-    elements are stored as a sequence. Then, there is a sequence of features, where each feature is
-    represented a mapping, and lbp value in a nested sequence.
-
--   When you write to a mapping (a structure), you write element name followed by its value. When you
-    write to a sequence, you simply write the elements one by one. OpenCV data structures (such as
-    cv::Mat) are written in absolutely the same way as simple C data structures - using `<<`
-    operator.
-
--   To write a mapping, you first write the special string `{` to the storage, then write the
-    elements as pairs (`fs << <element_name> << <element_value>`) and then write the closing
-    `}`.
-
--   To write a sequence, you first write the special string `[`, then write the elements, then
-    write the closing `]`.
-
--   In YAML/JSON (but not XML), mappings and sequences can be written in a compact Python-like inline
-    form. In the sample above matrix elements, as well as each feature, including its lbp value, is
-    stored in such inline form. To store a mapping/sequence in a compact form, put `:` after the
-    opening character, e.g. use `{:` instead of `{` and `[:` instead of `[`. When the
-    data is written to XML, those extra `:` are ignored.
-
-Reading data from a file storage.
----------------------------------
-To read the previously written XML, YAML or JSON file, do the following:
--#  Open the file storage using FileStorage::FileStorage constructor or FileStorage::open method.
-    In the current implementation the whole file is parsed and the whole representation of file
-    storage is built in memory as a hierarchy of file nodes (see FileNode)
-
--#  Read the data you are interested in. Use FileStorage::operator [], FileNode::operator []
-    and/or FileNodeIterator.
-
--#  Close the storage using FileStorage::release.
-
-Here is how to read the file created by the code sample above:
-@code
-    FileStorage fs2("test.yml", FileStorage::READ);
-
-    // first method: use (type) operator on FileNode.
-    int frameCount = (int)fs2["frameCount"];
-
-    String date;
-    // second method: use FileNode::operator >>
-    fs2["calibrationDate"] >> date;
-
-    Mat cameraMatrix2, distCoeffs2;
-    fs2["cameraMatrix"] >> cameraMatrix2;
-    fs2["distCoeffs"] >> distCoeffs2;
-
-    cout << "frameCount: " << frameCount << endl
-         << "calibration date: " << date << endl
-         << "camera matrix: " << cameraMatrix2 << endl
-         << "distortion coeffs: " << distCoeffs2 << endl;
-
-    FileNode features = fs2["features"];
-    FileNodeIterator it = features.begin(), it_end = features.end();
-    int idx = 0;
-    std::vector<uchar> lbpval;
-
-    // iterate through a sequence using FileNodeIterator
-    for( ; it != it_end; ++it, idx++ )
-    {
-        cout << "feature #" << idx << ": ";
-        cout << "x=" << (int)(*it)["x"] << ", y=" << (int)(*it)["y"] << ", lbp: (";
-        // you can also easily read numerical arrays using FileNode >> std::vector operator.
-        (*it)["lbp"] >> lbpval;
-        for( int i = 0; i < (int)lbpval.size(); i++ )
-            cout << " " << (int)lbpval[i];
-        cout << ")" << endl;
-    }
-    fs2.release();
-@endcode
-
-Format specification    {#format_spec}
---------------------
-`([count]{u|c|w|s|i|f|d})`... where the characters correspond to fundamental C++ types:
--   `u` 8-bit unsigned number
--   `c` 8-bit signed number
--   `w` 16-bit unsigned number
--   `s` 16-bit signed number
--   `i` 32-bit signed number
--   `f` single precision floating-point number
--   `d` double precision floating-point number
--   `r` pointer, 32 lower bits of which are written as a signed integer. The type can be used to
-    store structures with links between the elements.
-
-`count` is the optional counter of values of a given type. For example, `2if` means that each array
-element is a structure of 2 integers, followed by a single-precision floating-point number. The
-equivalent notations of the above specification are `iif`, `2i1f` and so forth. Other examples: `u`
-means that the array consists of bytes, and `2d` means the array consists of pairs of doubles.
-
-@see @ref samples/cpp/filestorage.cpp
-*/
-
-//! @{
-
-/** @example samples/cpp/filestorage.cpp
-A complete example using the FileStorage interface
-*/
-
-////////////////////////// XML & YAML I/O //////////////////////////
-
-class CV_EXPORTS FileNode;
-class CV_EXPORTS FileNodeIterator;
-
-/** @brief XML/YAML/JSON file storage class that encapsulates all the information necessary for writing or
-reading data to/from a file.
- */
-class CV_EXPORTS_W FileStorage
-{
-public:
-    //! file storage mode
-    enum Mode
-    {
-        READ        = 0, //!< value, open the file for reading
-        WRITE       = 1, //!< value, open the file for writing
-        APPEND      = 2, //!< value, open the file for appending
-        MEMORY      = 4, //!< flag, read data from source or write data to the internal buffer (which is
-        //!< returned by FileStorage::release)
-        FORMAT_MASK = (7<<3), //!< mask for format flags
-        FORMAT_AUTO = 0,      //!< flag, auto format
-        FORMAT_XML  = (1<<3), //!< flag, XML format
-        FORMAT_YAML = (2<<3), //!< flag, YAML format
-        FORMAT_JSON = (3<<3), //!< flag, JSON format
-
-        BASE64      = 64,     //!< flag, write rawdata in Base64 by default. (consider using WRITE_BASE64)
-        WRITE_BASE64 = BASE64 | WRITE, //!< flag, enable both WRITE and BASE64
-    };
-    enum State
-    {
-        UNDEFINED      = 0,
-        VALUE_EXPECTED = 1,
-        NAME_EXPECTED  = 2,
-        INSIDE_MAP     = 4
-    };
-
-    /** @brief The constructors.
-
-     The full constructor opens the file. Alternatively you can use the default constructor and then
-     call FileStorage::open.
-     */
-    CV_WRAP FileStorage();
-
-    /** @overload
-     @copydoc open()
-     */
-    CV_WRAP FileStorage(const String& filename, int flags, const String& encoding=String());
-
-    //! the destructor. calls release()
-    virtual ~FileStorage();
-
-    /** @brief Opens a file.
-
-     See description of parameters in FileStorage::FileStorage. The method calls FileStorage::release
-     before opening the file.
-     @param filename Name of the file to open or the text string to read the data from.
-     Extension of the file (.xml, .yml/.yaml or .json) determines its format (XML, YAML or JSON
-     respectively). Also you can append .gz to work with compressed files, for example myHugeMatrix.xml.gz. If both
-     FileStorage::WRITE and FileStorage::MEMORY flags are specified, source is used just to specify
-     the output file format (e.g. mydata.xml, .yml etc.). A file name can also contain parameters.
-     You can use this format, "*?base64" (e.g. "file.json?base64" (case sensitive)), as an alternative to
-     FileStorage::BASE64 flag.
-     @param flags Mode of operation. One of FileStorage::Mode
-     @param encoding Encoding of the file. Note that UTF-16 XML encoding is not supported currently and
-     you should use 8-bit encoding instead of it.
-     */
-    CV_WRAP virtual bool open(const String& filename, int flags, const String& encoding=String());
-
-    /** @brief Checks whether the file is opened.
-
-     @returns true if the object is associated with the current file and false otherwise. It is a
-     good practice to call this method after you tried to open a file.
-     */
-    CV_WRAP virtual bool isOpened() const;
-
-    /** @brief Closes the file and releases all the memory buffers.
-
-     Call this method after all I/O operations with the storage are finished.
-     */
-    CV_WRAP virtual void release();
-
-    /** @brief Closes the file and releases all the memory buffers.
-
-     Call this method after all I/O operations with the storage are finished. If the storage was
-     opened for writing data and FileStorage::WRITE was specified
-     */
-    CV_WRAP virtual String releaseAndGetString();
-
-    /** @brief Returns the first element of the top-level mapping.
-     @returns The first element of the top-level mapping.
-     */
-    CV_WRAP FileNode getFirstTopLevelNode() const;
-
-    /** @brief Returns the top-level mapping
-     @param streamidx Zero-based index of the stream. In most cases there is only one stream in the file.
-     However, YAML supports multiple streams and so there can be several.
-     @returns The top-level mapping.
-     */
-    CV_WRAP FileNode root(int streamidx=0) const;
-
-    /** @brief Returns the specified element of the top-level mapping.
-     @param nodename Name of the file node.
-     @returns Node with the given name.
-     */
-    FileNode operator[](const String& nodename) const;
-
-    /** @overload */
-    CV_WRAP_AS(getNode) FileNode operator[](const char* nodename) const;
-
-    /**
-     * @brief Simplified writing API to use with bindings.
-     * @param name Name of the written object. When writing to sequences (a.k.a. "arrays"), pass an empty string.
-     * @param val Value of the written object.
-     */
-    CV_WRAP void write(const String& name, int val);
-    /// @overload
-    CV_WRAP void write(const String& name, double val);
-    /// @overload
-    CV_WRAP void write(const String& name, const String& val);
-    /// @overload
-    CV_WRAP void write(const String& name, const Mat& val);
-    /// @overload
-    CV_WRAP void write(const String& name, const std::vector<String>& val);
-
-    /** @brief Writes multiple numbers.
-
-     Writes one or more numbers of the specified format to the currently written structure. Usually it is
-     more convenient to use operator `<<` instead of this method.
-     @param fmt Specification of each array element, see @ref format_spec "format specification"
-     @param vec Pointer to the written array.
-     @param len Number of the uchar elements to write.
-     */
-    void writeRaw( const String& fmt, const void* vec, size_t len );
-
-    /** @brief Writes a comment.
-
-     The function writes a comment into file storage. The comments are skipped when the storage is read.
-     @param comment The written comment, single-line or multi-line
-     @param append If true, the function tries to put the comment at the end of current line.
-     Else if the comment is multi-line, or if it does not fit at the end of the current
-     line, the comment starts a new line.
-     */
-    CV_WRAP void writeComment(const String& comment, bool append = false);
-
-    /** @brief Starts to write a nested structure (sequence or a mapping).
-    @param name name of the structure. When writing to sequences (a.k.a. "arrays"), pass an empty string.
-    @param flags type of the structure (FileNode::MAP or FileNode::SEQ (both with optional FileNode::FLOW)).
-    @param typeName optional name of the type you store. The effect of setting this depends on the storage format.
-    I.e. if the format has a specification for storing type information, this parameter is used.
-    */
-    CV_WRAP void startWriteStruct(const String& name, int flags, const String& typeName=String());
-
-    /** @brief Finishes writing nested structure (should pair startWriteStruct())
-    */
-    CV_WRAP void endWriteStruct();
-
-    /** @brief Returns the normalized object name for the specified name of a file.
-    @param filename Name of a file
-    @returns The normalized object name.
-     */
-    static String getDefaultObjectName(const String& filename);
-
-    /** @brief Returns the current format.
-     * @returns The current format, see FileStorage::Mode
-     */
-    CV_WRAP int getFormat() const;
-
-    int state;
-    std::string elname;
-
-    class Impl;
-    Ptr<Impl> p;
-};
-
-/** @brief File Storage Node class.
-
-The node is used to store each and every element of the file storage opened for reading. When
-XML/YAML file is read, it is first parsed and stored in the memory as a hierarchical collection of
-nodes. Each node can be a "leaf" that is contain a single number or a string, or be a collection of
-other nodes. There can be named collections (mappings) where each element has a name and it is
-accessed by a name, and ordered collections (sequences) where elements do not have names but rather
-accessed by index. Type of the file node can be determined using FileNode::type method.
-
-Note that file nodes are only used for navigating file storages opened for reading. When a file
-storage is opened for writing, no data is stored in memory after it is written.
- */
-class CV_EXPORTS_W_SIMPLE FileNode
-{
-public:
-    //! type of the file storage node
-    enum
-    {
-        NONE      = 0, //!< empty node
-        INT       = 1, //!< an integer
-        REAL      = 2, //!< floating-point number
-        FLOAT     = REAL, //!< synonym or REAL
-        STR       = 3, //!< text string in UTF-8 encoding
-        STRING    = STR, //!< synonym for STR
-        SEQ       = 4, //!< sequence
-        MAP       = 5, //!< mapping
-        TYPE_MASK = 7,
-
-        FLOW      = 8,  //!< compact representation of a sequence or mapping. Used only by YAML writer
-        UNIFORM   = 8,  //!< if set, means that all the collection elements are numbers of the same type (real's or int's).
-        //!< UNIFORM is used only when reading FileStorage; FLOW is used only when writing. So they share the same bit
-        EMPTY     = 16, //!< empty structure (sequence or mapping)
-        NAMED     = 32  //!< the node has a name (i.e. it is element of a mapping).
-    };
-    /** @brief The constructors.
-
-     These constructors are used to create a default file node, construct it from obsolete structures or
-     from the another file node.
-     */
-    CV_WRAP FileNode();
-
-    /** @overload
-     @param fs Pointer to the file storage structure.
-     @param blockIdx Index of the memory block where the file node is stored
-     @param ofs Offset in bytes from the beginning of the serialized storage
-
-     @deprecated
-     */
-    FileNode(const FileStorage* fs, size_t blockIdx, size_t ofs);
-
-    /** @overload
-     @param node File node to be used as initialization for the created file node.
-     */
-    FileNode(const FileNode& node);
-
-    FileNode& operator=(const FileNode& node);
-
-    /** @brief Returns element of a mapping node or a sequence node.
-     @param nodename Name of an element in the mapping node.
-     @returns Returns the element with the given identifier.
-     */
-    FileNode operator[](const String& nodename) const;
-
-    /** @overload
-     @param nodename Name of an element in the mapping node.
-     */
-    CV_WRAP_AS(getNode) FileNode operator[](const char* nodename) const;
-
-    /** @overload
-     @param i Index of an element in the sequence node.
-     */
-    CV_WRAP_AS(at) FileNode operator[](int i) const;
-
-    /** @brief Returns keys of a mapping node.
-     @returns Keys of a mapping node.
-     */
-    CV_WRAP std::vector<String> keys() const;
-
-    /** @brief Returns type of the node.
-     @returns Type of the node. See FileNode::Type
-     */
-    CV_WRAP int type() const;
-
-    //! returns true if the node is empty
-    CV_WRAP bool empty() const;
-    //! returns true if the node is a "none" object
-    CV_WRAP bool isNone() const;
-    //! returns true if the node is a sequence
-    CV_WRAP bool isSeq() const;
-    //! returns true if the node is a mapping
-    CV_WRAP bool isMap() const;
-    //! returns true if the node is an integer
-    CV_WRAP bool isInt() const;
-    //! returns true if the node is a floating-point number
-    CV_WRAP bool isReal() const;
-    //! returns true if the node is a text string
-    CV_WRAP bool isString() const;
-    //! returns true if the node has a name
-    CV_WRAP bool isNamed() const;
-    //! returns the node name or an empty string if the node is nameless
-    CV_WRAP std::string name() const;
-    //! returns the number of elements in the node, if it is a sequence or mapping, or 1 otherwise.
-    CV_WRAP size_t size() const;
-    //! returns raw size of the FileNode in bytes
-    CV_WRAP size_t rawSize() const;
-    //! returns the node content as an integer. If the node stores floating-point number, it is rounded.
-    operator int() const;
-    //! returns the node content as float
-    operator float() const;
-    //! returns the node content as double
-    operator double() const;
-    //! returns the node content as text string
-    inline operator std::string() const { return this->string(); }
-
-    static bool isMap(int flags);
-    static bool isSeq(int flags);
-    static bool isCollection(int flags);
-    static bool isEmptyCollection(int flags);
-    static bool isFlow(int flags);
-
-    uchar* ptr();
-    const uchar* ptr() const;
-
-    //! returns iterator pointing to the first node element
-    FileNodeIterator begin() const;
-    //! returns iterator pointing to the element following the last node element
-    FileNodeIterator end() const;
-
-    /** @brief Reads node elements to the buffer with the specified format.
-
-    Usually it is more convenient to use operator `>>` instead of this method.
-    @param fmt Specification of each array element. See @ref format_spec "format specification"
-    @param vec Pointer to the destination array.
-    @param len Number of bytes to read (buffer size limit). If it is greater than number of
-               remaining elements then all of them will be read.
-     */
-    void readRaw( const String& fmt, void* vec, size_t len ) const;
-
-    /** Internal method used when reading FileStorage.
-     Sets the type (int, real or string) and value of the previously created node.
-     */
-    void setValue( int type, const void* value, int len=-1 );
-
-    //! Simplified reading API to use with bindings.
-    CV_WRAP double real() const;
-    //! Simplified reading API to use with bindings.
-    CV_WRAP std::string string() const;
-    //! Simplified reading API to use with bindings.
-    CV_WRAP Mat mat() const;
-
-    //protected:
-    FileNode(FileStorage::Impl* fs, size_t blockIdx, size_t ofs);
-
-    FileStorage::Impl* fs;
-    size_t blockIdx;
-    size_t ofs;
-};
-
-
-/** @brief used to iterate through sequences and mappings.
-
- A standard STL notation, with node.begin(), node.end() denoting the beginning and the end of a
- sequence, stored in node. See the data reading sample in the beginning of the section.
- */
-class CV_EXPORTS FileNodeIterator
-{
-public:
-    /** @brief The constructors.
-
-     These constructors are used to create a default iterator, set it to specific element in a file node
-     or construct it from another iterator.
-     */
-    FileNodeIterator();
-
-    /** @overload
-     @param node File node - the collection to iterate over;
-        it can be a scalar (equivalent to 1-element collection) or "none" (equivalent to empty collection).
-     @param seekEnd - true if iterator needs to be set after the last element of the node;
-        that is:
-            * node.begin() => FileNodeIterator(node, false)
-            * node.end() => FileNodeIterator(node, true)
-     */
-    FileNodeIterator(const FileNode& node, bool seekEnd);
-
-    /** @overload
-     @param it Iterator to be used as initialization for the created iterator.
-     */
-    FileNodeIterator(const FileNodeIterator& it);
-
-    FileNodeIterator& operator=(const FileNodeIterator& it);
-
-    //! returns the currently observed element
-    FileNode operator *() const;
-
-    //! moves iterator to the next node
-    FileNodeIterator& operator ++ ();
-    //! moves iterator to the next node
-    FileNodeIterator operator ++ (int);
-    //! moves iterator forward by the specified offset (possibly negative)
-    FileNodeIterator& operator += (int ofs);
-
-    /** @brief Reads node elements to the buffer with the specified format.
-
-    Usually it is more convenient to use operator `>>` instead of this method.
-    @param fmt Specification of each array element. See @ref format_spec "format specification"
-    @param vec Pointer to the destination array.
-    @param len Number of bytes to read (buffer size limit). If it is greater than number of
-               remaining elements then all of them will be read.
-     */
-    FileNodeIterator& readRaw( const String& fmt, void* vec,
-                               size_t len=(size_t)INT_MAX );
-
-    //! returns the number of remaining (not read yet) elements
-    size_t remaining() const;
-
-    bool equalTo(const FileNodeIterator& it) const;
-
-protected:
-    FileStorage::Impl* fs;
-    size_t blockIdx;
-    size_t ofs;
-    size_t blockSize;
-    size_t nodeNElems;
-    size_t idx;
-};
-
-//! @} core_xml
-
-/////////////////// XML & YAML I/O implementation //////////////////
-
-//! @relates cv::FileStorage
-//! @{
-
-CV_EXPORTS void write( FileStorage& fs, const String& name, int value );
-CV_EXPORTS void write( FileStorage& fs, const String& name, float value );
-CV_EXPORTS void write( FileStorage& fs, const String& name, double value );
-CV_EXPORTS void write( FileStorage& fs, const String& name, const String& value );
-CV_EXPORTS void write( FileStorage& fs, const String& name, const Mat& value );
-CV_EXPORTS void write( FileStorage& fs, const String& name, const SparseMat& value );
-#ifdef CV__LEGACY_PERSISTENCE
-CV_EXPORTS void write( FileStorage& fs, const String& name, const std::vector<KeyPoint>& value);
-CV_EXPORTS void write( FileStorage& fs, const String& name, const std::vector<DMatch>& value);
-#endif
-
-CV_EXPORTS void writeScalar( FileStorage& fs, int value );
-CV_EXPORTS void writeScalar( FileStorage& fs, float value );
-CV_EXPORTS void writeScalar( FileStorage& fs, double value );
-CV_EXPORTS void writeScalar( FileStorage& fs, const String& value );
-
-//! @}
-
-//! @relates cv::FileNode
-//! @{
-
-CV_EXPORTS void read(const FileNode& node, int& value, int default_value);
-CV_EXPORTS void read(const FileNode& node, float& value, float default_value);
-CV_EXPORTS void read(const FileNode& node, double& value, double default_value);
-CV_EXPORTS void read(const FileNode& node, std::string& value, const std::string& default_value);
-CV_EXPORTS void read(const FileNode& node, Mat& mat, const Mat& default_mat = Mat() );
-CV_EXPORTS void read(const FileNode& node, SparseMat& mat, const SparseMat& default_mat = SparseMat() );
-#ifdef CV__LEGACY_PERSISTENCE
-CV_EXPORTS void read(const FileNode& node, std::vector<KeyPoint>& keypoints);
-CV_EXPORTS void read(const FileNode& node, std::vector<DMatch>& matches);
-#endif
-CV_EXPORTS void read(const FileNode& node, KeyPoint& value, const KeyPoint& default_value);
-CV_EXPORTS void read(const FileNode& node, DMatch& value, const DMatch& default_value);
-
-template<typename _Tp> static inline void read(const FileNode& node, Point_<_Tp>& value, const Point_<_Tp>& default_value)
-{
-    std::vector<_Tp> temp; FileNodeIterator it = node.begin(); it >> temp;
-    value = temp.size() != 2 ? default_value : Point_<_Tp>(saturate_cast<_Tp>(temp[0]), saturate_cast<_Tp>(temp[1]));
-}
-
-template<typename _Tp> static inline void read(const FileNode& node, Point3_<_Tp>& value, const Point3_<_Tp>& default_value)
-{
-    std::vector<_Tp> temp; FileNodeIterator it = node.begin(); it >> temp;
-    value = temp.size() != 3 ? default_value : Point3_<_Tp>(saturate_cast<_Tp>(temp[0]), saturate_cast<_Tp>(temp[1]),
-                                                            saturate_cast<_Tp>(temp[2]));
-}
-
-template<typename _Tp> static inline void read(const FileNode& node, Size_<_Tp>& value, const Size_<_Tp>& default_value)
-{
-    std::vector<_Tp> temp; FileNodeIterator it = node.begin(); it >> temp;
-    value = temp.size() != 2 ? default_value : Size_<_Tp>(saturate_cast<_Tp>(temp[0]), saturate_cast<_Tp>(temp[1]));
-}
-
-template<typename _Tp> static inline void read(const FileNode& node, Complex<_Tp>& value, const Complex<_Tp>& default_value)
-{
-    std::vector<_Tp> temp; FileNodeIterator it = node.begin(); it >> temp;
-    value = temp.size() != 2 ? default_value : Complex<_Tp>(saturate_cast<_Tp>(temp[0]), saturate_cast<_Tp>(temp[1]));
-}
-
-template<typename _Tp> static inline void read(const FileNode& node, Rect_<_Tp>& value, const Rect_<_Tp>& default_value)
-{
-    std::vector<_Tp> temp; FileNodeIterator it = node.begin(); it >> temp;
-    value = temp.size() != 4 ? default_value : Rect_<_Tp>(saturate_cast<_Tp>(temp[0]), saturate_cast<_Tp>(temp[1]),
-                                                          saturate_cast<_Tp>(temp[2]), saturate_cast<_Tp>(temp[3]));
-}
-
-template<typename _Tp, int cn> static inline void read(const FileNode& node, Vec<_Tp, cn>& value, const Vec<_Tp, cn>& default_value)
-{
-    std::vector<_Tp> temp; FileNodeIterator it = node.begin(); it >> temp;
-    value = temp.size() != cn ? default_value : Vec<_Tp, cn>(&temp[0]);
-}
-
-template<typename _Tp, int m, int n> static inline void read(const FileNode& node, Matx<_Tp, m, n>& value, const Matx<_Tp, m, n>& default_matx = Matx<_Tp, m, n>())
-{
-    Mat temp;
-    read(node, temp); // read as a Mat class
-
-    if (temp.empty())
-        value = default_matx;
-    else
-        value = Matx<_Tp, m, n>(temp);
-}
-
-template<typename _Tp> static inline void read(const FileNode& node, Scalar_<_Tp>& value, const Scalar_<_Tp>& default_value)
-{
-    std::vector<_Tp> temp; FileNodeIterator it = node.begin(); it >> temp;
-    value = temp.size() != 4 ? default_value : Scalar_<_Tp>(saturate_cast<_Tp>(temp[0]), saturate_cast<_Tp>(temp[1]),
-                                                            saturate_cast<_Tp>(temp[2]), saturate_cast<_Tp>(temp[3]));
-}
-
-static inline void read(const FileNode& node, Range& value, const Range& default_value)
-{
-    Point2i temp(value.start, value.end); const Point2i default_temp = Point2i(default_value.start, default_value.end);
-    read(node, temp, default_temp);
-    value.start = temp.x; value.end = temp.y;
-}
-
-//! @}
-
-/** @brief Writes string to a file storage.
-@relates cv::FileStorage
- */
-CV_EXPORTS FileStorage& operator << (FileStorage& fs, const String& str);
-
-//! @cond IGNORED
-
-namespace internal
-{
-    class CV_EXPORTS WriteStructContext
-    {
-    public:
-        WriteStructContext(FileStorage& _fs, const String& name, int flags, const String& typeName = String());
-        ~WriteStructContext();
-    private:
-        FileStorage* fs;
-    };
-
-    template<typename _Tp, int numflag> class VecWriterProxy
-    {
-    public:
-        VecWriterProxy( FileStorage* _fs ) : fs(_fs) {}
-        void operator()(const std::vector<_Tp>& vec) const
-        {
-            size_t count = vec.size();
-            for (size_t i = 0; i < count; i++)
-                write(*fs, vec[i]);
-        }
-    private:
-        FileStorage* fs;
-    };
-
-    template<typename _Tp> class VecWriterProxy<_Tp, 1>
-    {
-    public:
-        VecWriterProxy( FileStorage* _fs ) : fs(_fs) {}
-        void operator()(const std::vector<_Tp>& vec) const
-        {
-            int _fmt = traits::SafeFmt<_Tp>::fmt;
-            char fmt[] = { (char)((_fmt >> 8) + '1'), (char)_fmt, '\0' };
-            fs->writeRaw(fmt, !vec.empty() ? (uchar*)&vec[0] : 0, vec.size() * sizeof(_Tp));
-        }
-    private:
-        FileStorage* fs;
-    };
-
-    template<typename _Tp, int numflag> class VecReaderProxy
-    {
-    public:
-        VecReaderProxy( FileNodeIterator* _it ) : it(_it) {}
-        void operator()(std::vector<_Tp>& vec, size_t count) const
-        {
-            count = std::min(count, it->remaining());
-            vec.resize(count);
-            for (size_t i = 0; i < count; i++, ++(*it))
-                read(**it, vec[i], _Tp());
-        }
-    private:
-        FileNodeIterator* it;
-    };
-
-    template<typename _Tp> class VecReaderProxy<_Tp, 1>
-    {
-    public:
-        VecReaderProxy( FileNodeIterator* _it ) : it(_it) {}
-        void operator()(std::vector<_Tp>& vec, size_t count) const
-        {
-            size_t remaining = it->remaining();
-            size_t cn = DataType<_Tp>::channels;
-            int _fmt = traits::SafeFmt<_Tp>::fmt;
-            CV_Assert((_fmt >> 8) < 9);
-            char fmt[] = { (char)((_fmt >> 8)+'1'), (char)_fmt, '\0' };
-            CV_Assert((remaining % cn) == 0);
-            size_t remaining1 = remaining / cn;
-            count = count > remaining1 ? remaining1 : count;
-            vec.resize(count);
-            it->readRaw(fmt, !vec.empty() ? (uchar*)&vec[0] : 0, count*sizeof(_Tp));
-        }
-    private:
-        FileNodeIterator* it;
-    };
-
-} // internal
-
-//! @endcond
-
-//! @relates cv::FileStorage
-//! @{
-
-template<typename _Tp> static inline
-void write(FileStorage& fs, const _Tp& value)
-{
-    write(fs, String(), value);
-}
-
-template<> inline
-void write( FileStorage& fs, const int& value )
-{
-    writeScalar(fs, value);
-}
-
-template<> inline
-void write( FileStorage& fs, const float& value )
-{
-    writeScalar(fs, value);
-}
-
-template<> inline
-void write( FileStorage& fs, const double& value )
-{
-    writeScalar(fs, value);
-}
-
-template<> inline
-void write( FileStorage& fs, const String& value )
-{
-    writeScalar(fs, value);
-}
-
-template<typename _Tp> static inline
-void write(FileStorage& fs, const Point_<_Tp>& pt )
-{
-    write(fs, pt.x);
-    write(fs, pt.y);
-}
-
-template<typename _Tp> static inline
-void write(FileStorage& fs, const Point3_<_Tp>& pt )
-{
-    write(fs, pt.x);
-    write(fs, pt.y);
-    write(fs, pt.z);
-}
-
-template<typename _Tp> static inline
-void write(FileStorage& fs, const Size_<_Tp>& sz )
-{
-    write(fs, sz.width);
-    write(fs, sz.height);
-}
-
-template<typename _Tp> static inline
-void write(FileStorage& fs, const Complex<_Tp>& c )
-{
-    write(fs, c.re);
-    write(fs, c.im);
-}
-
-template<typename _Tp> static inline
-void write(FileStorage& fs, const Rect_<_Tp>& r )
-{
-    write(fs, r.x);
-    write(fs, r.y);
-    write(fs, r.width);
-    write(fs, r.height);
-}
-
-template<typename _Tp, int cn> static inline
-void write(FileStorage& fs, const Vec<_Tp, cn>& v )
-{
-    for(int i = 0; i < cn; i++)
-        write(fs, v.val[i]);
-}
-
-template<typename _Tp, int m, int n> static inline
-void write(FileStorage& fs, const Matx<_Tp, m, n>& x )
-{
-    write(fs, Mat(x)); // write as a Mat class
-}
-
-template<typename _Tp> static inline
-void write(FileStorage& fs, const Scalar_<_Tp>& s )
-{
-    write(fs, s.val[0]);
-    write(fs, s.val[1]);
-    write(fs, s.val[2]);
-    write(fs, s.val[3]);
-}
-
-static inline
-void write(FileStorage& fs, const Range& r )
-{
-    write(fs, r.start);
-    write(fs, r.end);
-}
-
-template<typename _Tp> static inline
-void write( FileStorage& fs, const std::vector<_Tp>& vec )
-{
-    cv::internal::VecWriterProxy<_Tp, traits::SafeFmt<_Tp>::fmt != 0> w(&fs);
-    w(vec);
-}
-
-template<typename _Tp> static inline
-void write(FileStorage& fs, const String& name, const Point_<_Tp>& pt )
-{
-    cv::internal::WriteStructContext ws(fs, name, FileNode::SEQ+FileNode::FLOW);
-    write(fs, pt);
-}
-
-template<typename _Tp> static inline
-void write(FileStorage& fs, const String& name, const Point3_<_Tp>& pt )
-{
-    cv::internal::WriteStructContext ws(fs, name, FileNode::SEQ+FileNode::FLOW);
-    write(fs, pt);
-}
-
-template<typename _Tp> static inline
-void write(FileStorage& fs, const String& name, const Size_<_Tp>& sz )
-{
-    cv::internal::WriteStructContext ws(fs, name, FileNode::SEQ+FileNode::FLOW);
-    write(fs, sz);
-}
-
-template<typename _Tp> static inline
-void write(FileStorage& fs, const String& name, const Complex<_Tp>& c )
-{
-    cv::internal::WriteStructContext ws(fs, name, FileNode::SEQ+FileNode::FLOW);
-    write(fs, c);
-}
-
-template<typename _Tp> static inline
-void write(FileStorage& fs, const String& name, const Rect_<_Tp>& r )
-{
-    cv::internal::WriteStructContext ws(fs, name, FileNode::SEQ+FileNode::FLOW);
-    write(fs, r);
-}
-
-template<typename _Tp, int cn> static inline
-void write(FileStorage& fs, const String& name, const Vec<_Tp, cn>& v )
-{
-    cv::internal::WriteStructContext ws(fs, name, FileNode::SEQ+FileNode::FLOW);
-    write(fs, v);
-}
-
-template<typename _Tp, int m, int n> static inline
-void write(FileStorage& fs, const String& name, const Matx<_Tp, m, n>& x )
-{
-    write(fs, name, Mat(x)); // write as a Mat class
-}
-
-template<typename _Tp> static inline
-void write(FileStorage& fs, const String& name, const Scalar_<_Tp>& s )
-{
-    cv::internal::WriteStructContext ws(fs, name, FileNode::SEQ+FileNode::FLOW);
-    write(fs, s);
-}
-
-static inline
-void write(FileStorage& fs, const String& name, const Range& r )
-{
-    cv::internal::WriteStructContext ws(fs, name, FileNode::SEQ+FileNode::FLOW);
-    write(fs, r);
-}
-
-static inline
-void write(FileStorage& fs, const String& name, const KeyPoint& kpt)
-{
-    cv::internal::WriteStructContext ws(fs, name, FileNode::SEQ+FileNode::FLOW);
-    write(fs, kpt.pt.x);
-    write(fs, kpt.pt.y);
-    write(fs, kpt.size);
-    write(fs, kpt.angle);
-    write(fs, kpt.response);
-    write(fs, kpt.octave);
-    write(fs, kpt.class_id);
-}
-
-static inline
-void write(FileStorage& fs, const String& name, const DMatch& m)
-{
-    cv::internal::WriteStructContext ws(fs, name, FileNode::SEQ+FileNode::FLOW);
-    write(fs, m.queryIdx);
-    write(fs, m.trainIdx);
-    write(fs, m.imgIdx);
-    write(fs, m.distance);
-}
-
-template<typename _Tp, typename std::enable_if< std::is_enum<_Tp>::value >::type* = nullptr>
-static inline void write( FileStorage& fs, const String& name, const _Tp& val )
-{
-    write(fs, name, static_cast<int>(val));
-}
-
-template<typename _Tp> static inline
-void write( FileStorage& fs, const String& name, const std::vector<_Tp>& vec )
-{
-    cv::internal::WriteStructContext ws(fs, name, FileNode::SEQ+(traits::SafeFmt<_Tp>::fmt != 0 ? FileNode::FLOW : 0));
-    write(fs, vec);
-}
-
-template<typename _Tp> static inline
-void write( FileStorage& fs, const String& name, const std::vector< std::vector<_Tp> >& vec )
-{
-    cv::internal::WriteStructContext ws(fs, name, FileNode::SEQ);
-    for(size_t i = 0; i < vec.size(); i++)
-    {
-        cv::internal::WriteStructContext ws_(fs, name, FileNode::SEQ+(traits::SafeFmt<_Tp>::fmt != 0 ? FileNode::FLOW : 0));
-        write(fs, vec[i]);
-    }
-}
-
-#ifdef CV__LEGACY_PERSISTENCE
-// This code is not needed anymore, but it is preserved here to keep source compatibility
-// Implementation is similar to templates instantiations
-static inline void write(FileStorage& fs, const KeyPoint& kpt) { write(fs, String(), kpt); }
-static inline void write(FileStorage& fs, const DMatch& m) { write(fs, String(), m); }
-static inline void write(FileStorage& fs, const std::vector<KeyPoint>& vec)
-{
-    cv::internal::VecWriterProxy<KeyPoint, 0> w(&fs);
-    w(vec);
-}
-static inline void write(FileStorage& fs, const std::vector<DMatch>& vec)
-{
-    cv::internal::VecWriterProxy<DMatch, 0> w(&fs);
-    w(vec);
-
-}
-#endif
-
-//! @} FileStorage
-
-//! @relates cv::FileNode
-//! @{
-
-static inline
-void read(const FileNode& node, bool& value, bool default_value)
-{
-    int temp;
-    read(node, temp, (int)default_value);
-    value = temp != 0;
-}
-
-static inline
-void read(const FileNode& node, uchar& value, uchar default_value)
-{
-    int temp;
-    read(node, temp, (int)default_value);
-    value = saturate_cast<uchar>(temp);
-}
-
-static inline
-void read(const FileNode& node, schar& value, schar default_value)
-{
-    int temp;
-    read(node, temp, (int)default_value);
-    value = saturate_cast<schar>(temp);
-}
-
-static inline
-void read(const FileNode& node, ushort& value, ushort default_value)
-{
-    int temp;
-    read(node, temp, (int)default_value);
-    value = saturate_cast<ushort>(temp);
-}
-
-static inline
-void read(const FileNode& node, short& value, short default_value)
-{
-    int temp;
-    read(node, temp, (int)default_value);
-    value = saturate_cast<short>(temp);
-}
-
-template<typename _Tp> static inline
-void read( FileNodeIterator& it, std::vector<_Tp>& vec, size_t maxCount = (size_t)INT_MAX )
-{
-    cv::internal::VecReaderProxy<_Tp, traits::SafeFmt<_Tp>::fmt != 0> r(&it);
-    r(vec, maxCount);
-}
-
-template<typename _Tp, typename std::enable_if< std::is_enum<_Tp>::value >::type* = nullptr>
-static inline void read(const FileNode& node, _Tp& value, const _Tp& default_value = static_cast<_Tp>(0))
-{
-    int temp;
-    read(node, temp, static_cast<int>(default_value));
-    value = static_cast<_Tp>(temp);
-}
-
-template<typename _Tp> static inline
-void read( const FileNode& node, std::vector<_Tp>& vec, const std::vector<_Tp>& default_value = std::vector<_Tp>() )
-{
-    if(node.empty())
-        vec = default_value;
-    else
-    {
-        FileNodeIterator it = node.begin();
-        read( it, vec );
-    }
-}
-
-static inline
-void read( const FileNode& node, std::vector<KeyPoint>& vec, const std::vector<KeyPoint>& default_value )
-{
-    if(node.empty())
-        vec = default_value;
-    else
-        read(node, vec);
-}
-
-static inline
-void read( const FileNode& node, std::vector<DMatch>& vec, const std::vector<DMatch>& default_value )
-{
-    if(node.empty())
-        vec = default_value;
-    else
-        read(node, vec);
-}
-
-//! @} FileNode
-
-//! @relates cv::FileStorage
-//! @{
-
-/** @brief Writes data to a file storage.
- */
-template<typename _Tp> static inline
-FileStorage& operator << (FileStorage& fs, const _Tp& value)
-{
-    if( !fs.isOpened() )
-        return fs;
-    if( fs.state == FileStorage::NAME_EXPECTED + FileStorage::INSIDE_MAP )
-        CV_Error( Error::StsError, "No element name has been given" );
-    write( fs, fs.elname, value );
-    if( fs.state & FileStorage::INSIDE_MAP )
-        fs.state = FileStorage::NAME_EXPECTED + FileStorage::INSIDE_MAP;
-    return fs;
-}
-
-/** @brief Writes data to a file storage.
- */
-static inline
-FileStorage& operator << (FileStorage& fs, const char* str)
-{
-    return (fs << String(str));
-}
-
-/** @brief Writes data to a file storage.
- */
-static inline
-FileStorage& operator << (FileStorage& fs, char* value)
-{
-    return (fs << String(value));
-}
-
-//! @} FileStorage
-
-//! @relates cv::FileNodeIterator
-//! @{
-
-/** @brief Reads data from a file storage.
- */
-template<typename _Tp> static inline
-FileNodeIterator& operator >> (FileNodeIterator& it, _Tp& value)
-{
-    read( *it, value, _Tp());
-    return ++it;
-}
-
-/** @brief Reads data from a file storage.
- */
-template<typename _Tp> static inline
-FileNodeIterator& operator >> (FileNodeIterator& it, std::vector<_Tp>& vec)
-{
-    cv::internal::VecReaderProxy<_Tp, traits::SafeFmt<_Tp>::fmt != 0> r(&it);
-    r(vec, (size_t)INT_MAX);
-    return it;
-}
-
-//! @} FileNodeIterator
-
-//! @relates cv::FileNode
-//! @{
-
-/** @brief Reads data from a file storage.
- */
-template<typename _Tp> static inline
-void operator >> (const FileNode& n, _Tp& value)
-{
-    read( n, value, _Tp());
-}
-
-/** @brief Reads data from a file storage.
- */
-template<typename _Tp> static inline
-void operator >> (const FileNode& n, std::vector<_Tp>& vec)
-{
-    FileNodeIterator it = n.begin();
-    it >> vec;
-}
-
-/** @brief Reads KeyPoint from a file storage.
-*/
-//It needs special handling because it contains two types of fields, int & float.
-static inline
-void operator >> (const FileNode& n, KeyPoint& kpt)
-{
-    FileNodeIterator it = n.begin();
-    it >> kpt.pt.x >> kpt.pt.y >> kpt.size >> kpt.angle >> kpt.response >> kpt.octave >> kpt.class_id;
-}
-
-#ifdef CV__LEGACY_PERSISTENCE
-static inline
-void operator >> (const FileNode& n, std::vector<KeyPoint>& vec)
-{
-    read(n, vec);
-}
-static inline
-void operator >> (const FileNode& n, std::vector<DMatch>& vec)
-{
-    read(n, vec);
-}
-#endif
-
-/** @brief Reads DMatch from a file storage.
-*/
-//It needs special handling because it contains two types of fields, int & float.
-static inline
-void operator >> (const FileNode& n, DMatch& m)
-{
-    FileNodeIterator it = n.begin();
-    it >> m.queryIdx >> m.trainIdx >> m.imgIdx >> m.distance;
-}
-
-//! @} FileNode
-
-//! @relates cv::FileNodeIterator
-//! @{
-
-CV_EXPORTS bool operator == (const FileNodeIterator& it1, const FileNodeIterator& it2);
-CV_EXPORTS bool operator != (const FileNodeIterator& it1, const FileNodeIterator& it2);
-
-static inline
-ptrdiff_t operator - (const FileNodeIterator& it1, const FileNodeIterator& it2)
-{
-    return it2.remaining() - it1.remaining();
-}
-
-static inline
-bool operator < (const FileNodeIterator& it1, const FileNodeIterator& it2)
-{
-    return it1.remaining() > it2.remaining();
-}
-
-//! @} FileNodeIterator
-
-} // cv
-
-#endif // OPENCV_CORE_PERSISTENCE_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/quaternion.hpp b/projects/llm_framework/include/opencv4/opencv2/core/quaternion.hpp
deleted file mode 100644
index 8c21501e..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/quaternion.hpp
+++ /dev/null
@@ -1,1696 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2020, Huawei Technologies Co., Ltd. All rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//       http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Author: Liangqian Kong <chargerKong@126.com>
-//         Longbu Wang <riskiest@gmail.com>
-#ifndef OPENCV_CORE_QUATERNION_HPP
-#define OPENCV_CORE_QUATERNION_HPP
-
-#include <opencv2/core.hpp>
-#include <opencv2/core/utils/logger.hpp>
-#include <iostream>
-namespace cv
-{
-//! @addtogroup core
-//! @{
-
-//! Unit quaternion flag
-enum QuatAssumeType
-{
-    /**
-     * This flag is specified by default.
-     * If this flag is specified, the input quaternions are assumed to be not unit quaternions.
-     * It can guarantee the correctness of the calculations,
-     * although the calculation speed will be slower than the flag QUAT_ASSUME_UNIT.
-     */
-    QUAT_ASSUME_NOT_UNIT,
-    /**
-     * If this flag is specified, the input quaternions are assumed to be unit quaternions which
-     * will save some computations. However, if this flag is specified without unit quaternion,
-     * the program correctness of the result will not be guaranteed.
-     */
-    QUAT_ASSUME_UNIT
-};
-
-class QuatEnum
-{
-public:
-    /** @brief Enum of Euler angles type.
-     *
-     * Without considering the possibility of using two different convertions for the definition of the rotation axes ,
-     * there exists twelve possible sequences of rotation axes, divided into two groups:
-     * - Proper Euler angles (Z-X-Z, X-Y-X, Y-Z-Y, Z-Y-Z, X-Z-X, Y-X-Y)
-     * - Tait–Bryan angles (X-Y-Z, Y-Z-X, Z-X-Y, X-Z-Y, Z-Y-X, Y-X-Z).
-     *
-     * The three elemental rotations may be [extrinsic](https://en.wikipedia.org/wiki/Euler_angles#Definition_by_extrinsic_rotations)
-     * (rotations about the axes *xyz* of the original coordinate system, which is assumed to remain motionless),
-     * or [intrinsic](https://en.wikipedia.org/wiki/Euler_angles#Definition_by_intrinsic_rotations)(rotations about the axes of the rotating coordinate system *XYZ*, solidary with the moving body, which changes its orientation after each elemental rotation).
-     *
-     *
-     * Extrinsic and intrinsic rotations are relevant.
-     *
-     * The definition of the Euler angles is as following,
-     * - \f$\theta_1 \f$ represents the first rotation angle,
-     * - \f$\theta_2 \f$ represents the second rotation angle,
-     * - \f$\theta_3 \f$ represents the third rotation angle.
-     *
-     * For intrinsic rotations in the order of X-Y-Z, the rotation matrix R can be calculated by:\f[R =X(\theta_1) Y(\theta_2) Z(\theta_3) \f]
-     * For extrinsic rotations in the order of X-Y-Z, the rotation matrix R can be calculated by:\f[R =Z({\theta_3}) Y({\theta_2}) X({\theta_1})\f]
-     * where
-     * \f[X({\theta})={\begin{bmatrix}1&0&0\\0&\cos {\theta_1} &-\sin {\theta_1} \\0&\sin {\theta_1} &\cos {\theta_1} \\\end{bmatrix}},
-     * Y({\theta})={\begin{bmatrix}\cos \theta_{2}&0&\sin \theta_{2}\\0&1 &0 \\\ -sin \theta_2& 0&\cos \theta_{2} \\\end{bmatrix}},
-     * Z({\theta})={\begin{bmatrix}\cos\theta_{3} &-\sin \theta_3&0\\\sin \theta_3 &\cos \theta_3 &0\\0&0&1\\\end{bmatrix}}.
-     * \f]
-     *
-     * The function is designed according to this set of conventions:
-     * - [Right handed](https://en.wikipedia.org/wiki/Right_hand_rule) reference frames are adopted, and the [right hand rule](https://en.wikipedia.org/wiki/Right_hand_rule) is used to determine the sign of angles.
-     * - Each matrix is meant to represent an [active rotation](https://en.wikipedia.org/wiki/Active_and_passive_transformation) (the composing and composed matrices
-     * are supposed to act on the coordinates of vectors defined in the initial fixed reference frame and give as a result the coordinates of a rotated vector defined in the same reference frame).
-     * - For \f$\theta_1\f$ and \f$\theta_3\f$, the valid range is (−π, π].
-     *
-     *   For \f$\theta_2\f$, the valid range is [−π/2, π/2] or [0, π].
-     *
-     *   For Tait–Bryan angles, the valid range of \f$\theta_2\f$ is [−π/2, π/2]. When transforming a quaternion to Euler angles, the solution of Euler angles is unique in condition of \f$ \theta_2 \in (−π/2, π/2)\f$ .
-     *   If \f$\theta_2 = −π/2 \f$ or \f$ \theta_2 = π/2\f$, there are infinite solutions. The common name for this situation is gimbal lock.
-     *   For Proper Euler angles,the valid range of \f$\theta_2\f$ is in [0, π]. The solutions of Euler angles are unique in condition of  \f$ \theta_2 \in (0, π)\f$ . If \f$\theta_2 =0 \f$ or \f$\theta_2 =π \f$,
-     *   there are infinite solutions and gimbal lock will occur.
-     */
-    enum EulerAnglesType
-    {
-        INT_XYZ, ///< Intrinsic rotations with the Euler angles type X-Y-Z
-        INT_XZY, ///< Intrinsic rotations with the Euler angles type X-Z-Y
-        INT_YXZ, ///< Intrinsic rotations with the Euler angles type Y-X-Z
-        INT_YZX, ///< Intrinsic rotations with the Euler angles type Y-Z-X
-        INT_ZXY, ///< Intrinsic rotations with the Euler angles type Z-X-Y
-        INT_ZYX, ///< Intrinsic rotations with the Euler angles type Z-Y-X
-        INT_XYX, ///< Intrinsic rotations with the Euler angles type X-Y-X
-        INT_XZX, ///< Intrinsic rotations with the Euler angles type X-Z-X
-        INT_YXY, ///< Intrinsic rotations with the Euler angles type Y-X-Y
-        INT_YZY, ///< Intrinsic rotations with the Euler angles type Y-Z-Y
-        INT_ZXZ, ///< Intrinsic rotations with the Euler angles type Z-X-Z
-        INT_ZYZ, ///< Intrinsic rotations with the Euler angles type Z-Y-Z
-
-        EXT_XYZ, ///< Extrinsic rotations with the Euler angles type X-Y-Z
-        EXT_XZY, ///< Extrinsic rotations with the Euler angles type X-Z-Y
-        EXT_YXZ, ///< Extrinsic rotations with the Euler angles type Y-X-Z
-        EXT_YZX, ///< Extrinsic rotations with the Euler angles type Y-Z-X
-        EXT_ZXY, ///< Extrinsic rotations with the Euler angles type Z-X-Y
-        EXT_ZYX, ///< Extrinsic rotations with the Euler angles type Z-Y-X
-        EXT_XYX, ///< Extrinsic rotations with the Euler angles type X-Y-X
-        EXT_XZX, ///< Extrinsic rotations with the Euler angles type X-Z-X
-        EXT_YXY, ///< Extrinsic rotations with the Euler angles type Y-X-Y
-        EXT_YZY,  ///< Extrinsic rotations with the Euler angles type Y-Z-Y
-        EXT_ZXZ, ///< Extrinsic rotations with the Euler angles type Z-X-Z
-        EXT_ZYZ, ///< Extrinsic rotations with the Euler angles type Z-Y-Z
-        #ifndef CV_DOXYGEN
-            EULER_ANGLES_MAX_VALUE
-        #endif
-    };
-
-};
-
-template <typename _Tp> class Quat;
-template <typename _Tp> std::ostream& operator<<(std::ostream&, const Quat<_Tp>&);
-
-/**
- * Quaternion is a number system that extends the complex numbers. It can be expressed as a
- * rotation in three-dimensional space.
- * A quaternion is generally represented in the form:
- *      \f[q = w + x\boldsymbol{i} + y\boldsymbol{j} + z\boldsymbol{k}\f]
- *      \f[q = [w, x, y, z]\f]
- *      \f[q = [w, \boldsymbol{v}] \f]
- *      \f[q = ||q||[\cos\psi, u_x\sin\psi,u_y\sin\psi,  u_z\sin\psi].\f]
- *      \f[q = ||q||[\cos\psi, \boldsymbol{u}\sin\psi]\f]
- * where \f$\psi = \frac{\theta}{2}\f$, \f$\theta\f$ represents rotation angle,
- * \f$\boldsymbol{u} = [u_x, u_y, u_z]\f$ represents normalized rotation axis,
- * and \f$||q||\f$ represents the norm of \f$q\f$.
- *
- * A unit quaternion is usually represents rotation, which has the form:
- *      \f[q = [\cos\psi, u_x\sin\psi,u_y\sin\psi,  u_z\sin\psi].\f]
- *
- * To create a quaternion representing the rotation around the axis \f$\boldsymbol{u}\f$
- * with angle \f$\theta\f$, you can use
- * ```
- * using namespace cv;
- * double angle = CV_PI;
- * Vec3d axis = {0, 0, 1};
- * Quatd q = Quatd::createFromAngleAxis(angle, axis);
- * ```
- *
- * You can simply use four same type number to create a quaternion
- * ```
- * Quatd q(1, 2, 3, 4);
- * ```
- * Or use a Vec4d or Vec4f vector.
- * ```
- * Vec4d vec{1, 2, 3, 4};
- * Quatd q(vec);
- * ```
- *
- * ```
- * Vec4f vec{1, 2, 3, 4};
- * Quatf q(vec);
- * ```
- *
- * If you already have a 3x3 rotation matrix R, then you can use
- * ```
- * Quatd q = Quatd::createFromRotMat(R);
- * ```
- *
- * If you already have a rotation vector rvec which has the form of `angle * axis`, then you can use
- * ```
- * Quatd q = Quatd::createFromRvec(rvec);
- * ```
- *
- * To extract the rotation matrix from quaternion, see toRotMat3x3()
- *
- * To extract the Vec4d or Vec4f, see toVec()
- *
- * To extract the rotation vector, see toRotVec()
- *
- * If there are two quaternions \f$q_0, q_1\f$ are needed to interpolate, you can use nlerp(), slerp() or spline()
- * ```
- * Quatd::nlerp(q0, q1, t)
- *
- * Quatd::slerp(q0, q1, t)
- *
- * Quatd::spline(q0, q0, q1, q1, t)
- * ```
- * spline can smoothly connect rotations of  multiple quaternions
- *
- * Three ways to get an element in Quaternion
- * ```
- * Quatf q(1,2,3,4);
- * std::cout << q.w << std::endl; // w=1, x=2, y=3, z=4
- * std::cout << q[0] << std::endl; // q[0]=1, q[1]=2, q[2]=3, q[3]=4
- * std::cout << q.at(0) << std::endl;
- * ```
- */
-template <typename _Tp>
-class Quat
-{
-    static_assert(std::is_floating_point<_Tp>::value, "Quaternion only make sense with type of float or double");
-    using value_type = _Tp;
-public:
-    static constexpr _Tp CV_QUAT_EPS = (_Tp)1.e-6;
-    static constexpr _Tp CV_QUAT_CONVERT_THRESHOLD = (_Tp)1.e-6;
-
-    Quat();
-
-    /**
-     * @brief From Vec4d or Vec4f.
-     */
-    explicit Quat(const Vec<_Tp, 4> &coeff);
-
-    /**
-     * @brief from four numbers.
-     */
-    Quat(_Tp w, _Tp x, _Tp y, _Tp z);
-
-    /**
-     * @brief from an angle, axis. Axis will be normalized in this function. And
-     * it generates
-     * \f[q = [\cos\psi, u_x\sin\psi,u_y\sin\psi,  u_z\sin\psi].\f]
-     * where \f$\psi = \frac{\theta}{2}\f$, \f$\theta\f$ is the rotation angle.
-     */
-    static Quat<_Tp> createFromAngleAxis(const _Tp angle, const Vec<_Tp, 3> &axis);
-
-    /**
-     * @brief from a 3x3 rotation matrix.
-     */
-    static Quat<_Tp> createFromRotMat(InputArray R);
-
-    /**
-     * @brief from a rotation vector
-     * \f$r\f$ has the form \f$\theta \cdot \boldsymbol{u}\f$, where \f$\theta\f$
-     * represents rotation angle and \f$\boldsymbol{u}\f$ represents normalized rotation axis.
-     *
-     * Angle and axis could be easily derived as:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * \psi &= ||r||\\
-     * \boldsymbol{u} &= \frac{r}{\theta}
-     * \end{split}
-     * \end{equation}
-     * \f]
-     * Then a quaternion can be calculated by
-     *  \f[q = [\cos\psi, \boldsymbol{u}\sin\psi]\f]
-     *  where \f$\psi = \theta / 2 \f$
-     */
-    static Quat<_Tp> createFromRvec(InputArray rvec);
-
-     /**
-     * @brief
-     * from Euler angles
-     *
-     * A quaternion can be generated from Euler angles by combining the quaternion representations of the Euler rotations.
-     *
-     * For example, if we use intrinsic rotations in the order of X-Y-Z,\f$\theta_1 \f$ is rotation around the X-axis, \f$\theta_2 \f$ is rotation around the Y-axis,
-     * \f$\theta_3 \f$ is rotation around the Z-axis. The final quaternion q can be calculated by
-     *
-     * \f[ {q} = q_{X, \theta_1}  q_{Y, \theta_2} q_{Z, \theta_3}\f]
-     * where \f$ q_{X, \theta_1} \f$ is created from @ref createFromXRot,  \f$ q_{Y, \theta_2} \f$ is created from @ref createFromYRot,
-     *  \f$ q_{Z, \theta_3} \f$ is created from @ref createFromZRot.
-     * @param angles the Euler angles in a vector of length 3
-     * @param eulerAnglesType the convertion Euler angles type
-     */
-    static Quat<_Tp> createFromEulerAngles(const Vec<_Tp, 3> &angles, QuatEnum::EulerAnglesType eulerAnglesType);
-
-    /**
-     * @brief get a quaternion from a rotation about the Y-axis by \f$\theta\f$ .
-     * \f[q = \cos(\theta/2)+0 i+ sin(\theta/2) j +0k \f]
-     */
-    static Quat<_Tp> createFromYRot(const _Tp theta);
-
-    /**
-     * @brief get a quaternion from a rotation about the X-axis by \f$\theta\f$ .
-     * \f[q = \cos(\theta/2)+sin(\theta/2) i +0 j +0 k \f]
-     */
-    static Quat<_Tp> createFromXRot(const _Tp theta);
-
-    /**
-     * @brief get a quaternion from a rotation about the Z-axis by \f$\theta\f$.
-     * \f[q = \cos(\theta/2)+0 i +0 j +sin(\theta/2) k \f]
-     */
-    static Quat<_Tp> createFromZRot(const _Tp theta);
-
-    /**
-     * @brief a way to get element.
-     * @param index over a range [0, 3].
-     *
-     * A quaternion q
-     *
-     * q.at(0) is equivalent to q.w,
-     *
-     * q.at(1) is equivalent to q.x,
-     *
-     * q.at(2) is equivalent to q.y,
-     *
-     * q.at(3) is equivalent to q.z.
-     */
-    _Tp at(size_t index) const;
-
-    /**
-     * @brief return the conjugate of this quaternion.
-     * \f[q.conjugate() = (w, -x, -y, -z).\f]
-     */
-    Quat<_Tp> conjugate() const;
-
-    /**
-     *
-     * @brief return the value of exponential value.
-     * \f[\exp(q) = e^w (\cos||\boldsymbol{v}||+ \frac{v}{||\boldsymbol{v}||})\sin||\boldsymbol{v}||\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     * @param q a quaternion.
-     *
-     * For example:
-     * ```
-     * Quatd q{1,2,3,4};
-     * cout << exp(q) << endl;
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> exp(const Quat<T> &q);
-
-    /**
-     * @brief return the value of exponential value.
-     * \f[\exp(q) = e^w (\cos||\boldsymbol{v}||+ \frac{v}{||\boldsymbol{v}||}\sin||\boldsymbol{v}||)\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     *
-     * For example
-     * ```
-     * Quatd q{1,2,3,4};
-     * cout << q.exp() << endl;
-     * ```
-     */
-    Quat<_Tp> exp() const;
-
-    /**
-     * @brief return the value of logarithm function.
-     * \f[\ln(q) = \ln||q|| + \frac{\boldsymbol{v}}{||\boldsymbol{v}||}\arccos\frac{w}{||q||}.\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     * @param q a quaternion.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, q assume to be a unit quaternion and this function will save some computations.
-     *
-     * For example
-     * ```
-     * Quatd q1{1,2,3,4};
-     * cout << log(q1) << endl;
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> log(const Quat<T> &q, QuatAssumeType assumeUnit);
-
-    /**
-     * @brief return the value of logarithm function.
-     *  \f[\ln(q) = \ln||q|| + \frac{\boldsymbol{v}}{||\boldsymbol{v}||}\arccos\frac{w}{||q||}\f].
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     * @param assumeUnit if QUAT_ASSUME_UNIT, this quaternion assume to be a unit quaternion and this function will save some computations.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.log();
-     *
-     * QuatAssumeType assumeUnit = QUAT_ASSUME_UNIT;
-     * Quatd q1(1,2,3,4);
-     * q1.normalize().log(assumeUnit);
-     * ```
-     */
-    Quat<_Tp> log(QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief return the value of power function with index \f$x\f$.
-     * \f[q^x = ||q||(cos(x\theta) + \boldsymbol{u}sin(x\theta))).\f]
-     * @param q a quaternion.
-     * @param x index of exponentiation.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, quaternion q assume to be a unit quaternion and this function will save some computations.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * power(q, 2.0);
-     *
-     * QuatAssumeType assumeUnit = QUAT_ASSUME_UNIT;
-     * double angle = CV_PI;
-     * Vec3d axis{0, 0, 1};
-     * Quatd q1 = Quatd::createFromAngleAxis(angle, axis); //generate a unit quat by axis and angle
-     * power(q1, 2.0, assumeUnit);//This assumeUnit means q1 is a unit quaternion.
-     * ```
-     * @note the type of the index should be the same as the quaternion.
-     */
-    template <typename T>
-    friend Quat<T> power(const Quat<T> &q, const T x, QuatAssumeType assumeUnit);
-
-    /**
-     * @brief return the value of power function with index \f$x\f$.
-     * \f[q^x = ||q||(\cos(x\theta) + \boldsymbol{u}\sin(x\theta))).\f]
-     * @param x index of exponentiation.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, this quaternion assume to be a unit quaternion and this function will save some computations.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.power(2.0);
-     *
-     * QuatAssumeType assumeUnit = QUAT_ASSUME_UNIT;
-     * double angle = CV_PI;
-     * Vec3d axis{0, 0, 1};
-     * Quatd q1 = Quatd::createFromAngleAxis(angle, axis); //generate a unit quat by axis and angle
-     * q1.power(2.0, assumeUnit); //This assumeUnt means q1 is a unit quaternion
-     * ```
-     */
-    Quat<_Tp> power(const _Tp x, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief return \f$\sqrt{q}\f$.
-     * @param q a quaternion.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, quaternion q assume to be a unit quaternion and this function will save some computations.
-     *
-     * For example
-     * ```
-     * Quatf q(1,2,3,4);
-     * sqrt(q);
-     *
-     * QuatAssumeType assumeUnit = QUAT_ASSUME_UNIT;
-     * q = {1,0,0,0};
-     * sqrt(q, assumeUnit); //This assumeUnit means q is a unit quaternion.
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> sqrt(const Quat<T> &q, QuatAssumeType assumeUnit);
-
-    /**
-     * @brief return \f$\sqrt{q}\f$.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, this quaternion assume to be a unit quaternion and this function will save some computations.
-     *
-     * For example
-     * ```
-     * Quatf q(1,2,3,4);
-     * q.sqrt();
-     *
-     * QuatAssumeType assumeUnit = QUAT_ASSUME_UNIT;
-     * q = {1,0,0,0};
-     * q.sqrt(assumeUnit); //This assumeUnit means q is a unit quaternion
-     * ```
-     */
-    Quat<_Tp> sqrt(QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief return the value of power function with quaternion \f$q\f$.
-     * \f[p^q = e^{q\ln(p)}.\f]
-     * @param p base quaternion of power function.
-     * @param q index quaternion of power function.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, quaternion \f$p\f$ assume to be a unit quaternion and this function will save some computations.
-     *
-     * For example
-     * ```
-     * Quatd p(1,2,3,4);
-     * Quatd q(5,6,7,8);
-     * power(p, q);
-     *
-     * QuatAssumeType assumeUnit = QUAT_ASSUME_UNIT;
-     * p = p.normalize();
-     * power(p, q, assumeUnit); //This assumeUnit means p is a unit quaternion
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> power(const Quat<T> &p, const Quat<T> &q, QuatAssumeType assumeUnit);
-
-    /**
-     * @brief return the value of power function with quaternion \f$q\f$.
-     * \f[p^q = e^{q\ln(p)}.\f]
-     * @param q index quaternion of power function.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, this quaternion assume to be a unit quaternion and this function will save some computations.
-     *
-     * For example
-     * ```
-     * Quatd p(1,2,3,4);
-     * Quatd q(5,6,7,8);
-     * p.power(q);
-     *
-     * QuatAssumeType assumeUnit = QUAT_ASSUME_UNIT;
-     * p = p.normalize();
-     * p.power(q, assumeUnit); //This assumeUnit means p is a unit quaternion
-     * ```
-     */
-    Quat<_Tp> power(const Quat<_Tp> &q, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief return the crossProduct between \f$p = (a, b, c, d) = (a, \boldsymbol{u})\f$ and \f$q = (w, x, y, z) = (w, \boldsymbol{v})\f$.
-     * \f[p \times q = \frac{pq- qp}{2}\f]
-     * \f[p \times q = \boldsymbol{u} \times \boldsymbol{v}\f]
-     * \f[p \times q = (cz-dy)i + (dx-bz)j + (by-xc)k \f]
-     *
-     * For example
-     * ```
-     * Quatd q{1,2,3,4};
-     * Quatd p{5,6,7,8};
-     * crossProduct(p, q);
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> crossProduct(const Quat<T> &p, const Quat<T> &q);
-
-    /**
-     * @brief return the crossProduct between \f$p = (a, b, c, d) = (a, \boldsymbol{u})\f$ and \f$q = (w, x, y, z) = (w, \boldsymbol{v})\f$.
-     * \f[p \times q = \frac{pq- qp}{2}.\f]
-     * \f[p \times q = \boldsymbol{u} \times \boldsymbol{v}.\f]
-     * \f[p \times q = (cz-dy)i + (dx-bz)j + (by-xc)k. \f]
-     *
-     * For example
-     * ```
-     * Quatd q{1,2,3,4};
-     * Quatd p{5,6,7,8};
-     * p.crossProduct(q)
-     * ```
-     */
-    Quat<_Tp> crossProduct(const Quat<_Tp> &q) const;
-
-    /**
-     * @brief return the norm of quaternion.
-     * \f[||q|| = \sqrt{w^2 + x^2 + y^2 + z^2}.\f]
-     */
-    _Tp norm() const;
-
-    /**
-     * @brief return a normalized \f$p\f$.
-     * \f[p = \frac{q}{||q||}\f]
-     * where \f$p\f$ satisfies \f$(p.x)^2 + (p.y)^2 + (p.z)^2 + (p.w)^2 = 1.\f$
-     */
-    Quat<_Tp> normalize() const;
-
-    /**
-     * @brief return \f$q^{-1}\f$ which is an inverse of \f$q\f$
-     * which satisfies \f$q * q^{-1} = 1\f$.
-     * @param q a quaternion.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, quaternion q assume to be a unit quaternion and this function will save some computations.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * inv(q);
-     *
-     * QuatAssumeType assumeUnit = QUAT_ASSUME_UNIT;
-     * q = q.normalize();
-     * inv(q, assumeUnit);//This assumeUnit means p is a unit quaternion
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> inv(const Quat<T> &q, QuatAssumeType assumeUnit);
-
-    /**
-     * @brief return \f$q^{-1}\f$ which is an inverse of \f$q\f$
-     * satisfying \f$q * q^{-1} = 1\f$.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, quaternion q assume to be a unit quaternion and this function will save some computations.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.inv();
-     *
-     * QuatAssumeType assumeUnit = QUAT_ASSUME_UNIT;
-     * q = q.normalize();
-     * q.inv(assumeUnit);  //assumeUnit means p is a unit quaternion
-     * ```
-     */
-    Quat<_Tp> inv(QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief return sinh value of quaternion q, sinh could be calculated as:
-     * \f[\sinh(p) = \sin(w)\cos(||\boldsymbol{v}||) + \cosh(w)\frac{v}{||\boldsymbol{v}||}\sin||\boldsymbol{v}||\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     * @param q a quaternion.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * sinh(q);
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> sinh(const Quat<T> &q);
-
-    /**
-     * @brief return sinh value of this quaternion, sinh could be calculated as:
-     * \f$\sinh(p) = \sin(w)\cos(||\boldsymbol{v}||) + \cosh(w)\frac{v}{||\boldsymbol{v}||}\sin||\boldsymbol{v}||\f$
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.sinh();
-     * ```
-     */
-    Quat<_Tp> sinh() const;
-
-    /**
-     * @brief return cosh value of quaternion q, cosh could be calculated as:
-     * \f[\cosh(p) = \cosh(w) * \cos(||\boldsymbol{v}||) + \sinh(w)\frac{\boldsymbol{v}}{||\boldsymbol{v}||}\sin(||\boldsymbol{v}||)\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     * @param q a quaternion.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * cosh(q);
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> cosh(const Quat<T> &q);
-
-    /**
-     * @brief return cosh value of this quaternion, cosh could be calculated as:
-     * \f[\cosh(p) = \cosh(w) * \cos(||\boldsymbol{v}||) + \sinh(w)\frac{\boldsymbol{v}}{||\boldsymbol{v}||}sin(||\boldsymbol{v}||)\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.cosh();
-     * ```
-     */
-    Quat<_Tp> cosh() const;
-
-    /**
-     * @brief return tanh value of quaternion q, tanh could be calculated as:
-     * \f[ \tanh(q) = \frac{\sinh(q)}{\cosh(q)}.\f]
-     * @param q a quaternion.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * tanh(q);
-     * ```
-     * @sa sinh, cosh
-     */
-    template <typename T>
-    friend Quat<T> tanh(const Quat<T> &q);
-
-    /**
-     * @brief return tanh value of this quaternion, tanh could be calculated as:
-     * \f[ \tanh(q) = \frac{\sinh(q)}{\cosh(q)}.\f]
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.tanh();
-     * ```
-     * @sa sinh, cosh
-     */
-    Quat<_Tp> tanh() const;
-
-    /**
-     * @brief return tanh value of quaternion q, sin could be calculated as:
-     * \f[\sin(p) = \sin(w) * \cosh(||\boldsymbol{v}||) + \cos(w)\frac{\boldsymbol{v}}{||\boldsymbol{v}||}\sinh(||\boldsymbol{v}||)\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     * @param q a quaternion.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * sin(q);
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> sin(const Quat<T> &q);
-
-    /**
-     * @brief return sin value of this quaternion, sin could be calculated as:
-     * \f[\sin(p) = \sin(w) * \cosh(||\boldsymbol{v}||) + \cos(w)\frac{\boldsymbol{v}}{||\boldsymbol{v}||}\sinh(||\boldsymbol{v}||)\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.sin();
-     * ```
-     */
-    Quat<_Tp> sin() const;
-
-    /**
-     * @brief return sin value of quaternion q, cos could be calculated as:
-     * \f[\cos(p) = \cos(w) * \cosh(||\boldsymbol{v}||) - \sin(w)\frac{\boldsymbol{v}}{||\boldsymbol{v}||}\sinh(||\boldsymbol{v}||)\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     * @param q a quaternion.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * cos(q);
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> cos(const Quat<T> &q);
-
-    /**
-     * @brief return cos value of this quaternion, cos could be calculated as:
-     * \f[\cos(p) = \cos(w) * \cosh(||\boldsymbol{v}||) - \sin(w)\frac{\boldsymbol{v}}{||\boldsymbol{v}||}\sinh(||\boldsymbol{v}||)\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.cos();
-     * ```
-     */
-    Quat<_Tp> cos() const;
-
-    /**
-     * @brief return tan value of quaternion q, tan could be calculated as:
-     * \f[\tan(q) = \frac{\sin(q)}{\cos(q)}.\f]
-     * @param q a quaternion.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * tan(q);
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> tan(const Quat<T> &q);
-
-    /**
-     * @brief return tan value of this quaternion, tan could be calculated as:
-     * \f[\tan(q) = \frac{\sin(q)}{\cos(q)}.\f]
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.tan();
-     * ```
-     */
-    Quat<_Tp> tan() const;
-
-    /**
-     * @brief return arcsin value of quaternion q, arcsin could be calculated as:
-     * \f[\arcsin(q) = -\frac{\boldsymbol{v}}{||\boldsymbol{v}||}arcsinh(q\frac{\boldsymbol{v}}{||\boldsymbol{v}||})\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     * @param q a quaternion.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * asin(q);
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> asin(const Quat<T> &q);
-
-    /**
-     * @brief return arcsin value of this quaternion, arcsin could be calculated as:
-     * \f[\arcsin(q) = -\frac{\boldsymbol{v}}{||\boldsymbol{v}||}arcsinh(q\frac{\boldsymbol{v}}{||\boldsymbol{v}||})\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.asin();
-     * ```
-     */
-    Quat<_Tp> asin() const;
-
-    /**
-     * @brief return arccos value of quaternion q, arccos could be calculated as:
-     * \f[\arccos(q) = -\frac{\boldsymbol{v}}{||\boldsymbol{v}||}arccosh(q)\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     * @param q a quaternion.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * acos(q);
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> acos(const Quat<T> &q);
-
-    /**
-     * @brief return arccos value of this quaternion, arccos could be calculated as:
-     * \f[\arccos(q) = -\frac{\boldsymbol{v}}{||\boldsymbol{v}||}arccosh(q)\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.acos();
-     * ```
-     */
-    Quat<_Tp> acos() const;
-
-    /**
-     * @brief return arctan value of quaternion q, arctan could be calculated as:
-     * \f[\arctan(q) = -\frac{\boldsymbol{v}}{||\boldsymbol{v}||}arctanh(q\frac{\boldsymbol{v}}{||\boldsymbol{v}||})\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     * @param q a quaternion.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * atan(q);
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> atan(const Quat<T> &q);
-
-    /**
-     * @brief return arctan value of this quaternion, arctan could be calculated as:
-     * \f[\arctan(q) = -\frac{\boldsymbol{v}}{||\boldsymbol{v}||}arctanh(q\frac{\boldsymbol{v}}{||\boldsymbol{v}||})\f]
-     * where \f$\boldsymbol{v} = [x, y, z].\f$
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.atan();
-     * ```
-     */
-    Quat<_Tp> atan() const;
-
-    /**
-     * @brief return arcsinh value of quaternion q, arcsinh could be calculated as:
-     * \f[arcsinh(q) = \ln(q + \sqrt{q^2 + 1})\f].
-     * @param q a quaternion.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * asinh(q);
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> asinh(const Quat<T> &q);
-
-    /**
-     * @brief return arcsinh value of this quaternion, arcsinh could be calculated as:
-     * \f[arcsinh(q) = \ln(q + \sqrt{q^2 + 1})\f].
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.asinh();
-     * ```
-     */
-    Quat<_Tp> asinh() const;
-
-    /**
-     * @brief return arccosh value of quaternion q, arccosh could be calculated as:
-     * \f[arccosh(q) = \ln(q + \sqrt{q^2 - 1})\f].
-     * @param q a quaternion.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * acosh(q);
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> acosh(const Quat<T> &q);
-
-    /**
-     * @brief return arccosh value of this quaternion, arccosh could be calculated as:
-     * \f[arcosh(q) = \ln(q + \sqrt{q^2 - 1})\f].
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.acosh();
-     * ```
-     */
-    Quat<_Tp> acosh() const;
-
-    /**
-     * @brief return arctanh value of quaternion q, arctanh could be calculated as:
-     * \f[arctanh(q) = \frac{\ln(q + 1) - \ln(1 - q)}{2}\f].
-     * @param q a quaternion.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * atanh(q);
-     * ```
-     */
-    template <typename T>
-    friend Quat<T> atanh(const Quat<T> &q);
-
-    /**
-     * @brief return arctanh value of this quaternion, arctanh could be calculated as:
-     * \f[arcsinh(q) = \frac{\ln(q + 1) - \ln(1 - q)}{2}\f].
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.atanh();
-     * ```
-     */
-    Quat<_Tp> atanh() const;
-
-    /**
-     * @brief return true if this quaternion is a unit quaternion.
-     * @param eps tolerance scope of normalization. The eps could be defined as
-     *
-     * \f[eps = |1 - dotValue|\f] where \f[dotValue = (this.w^2 + this.x^2 + this,y^2 + this.z^2).\f]
-     * And this function will consider it is normalized when the dotValue over a range \f$[1-eps, 1+eps]\f$.
-     */
-    bool isNormal(_Tp eps=CV_QUAT_EPS) const;
-
-    /**
-     * @brief to throw an error if this quaternion is not a unit quaternion.
-     * @param eps tolerance scope of normalization.
-     * @sa isNormal
-     */
-    void assertNormal(_Tp eps=CV_QUAT_EPS) const;
-
-    /**
-     * @brief transform a quaternion to a 3x3 rotation matrix.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, this quaternion assume to be a unit quaternion and
-     * this function will save some computations. Otherwise, this function will normalize this
-     * quaternion at first then do the transformation.
-     *
-     * @note Matrix A which is to be rotated should have the form
-     * \f[\begin{bmatrix}
-     * x_0& x_1& x_2&...&x_n\\
-     * y_0& y_1& y_2&...&y_n\\
-     * z_0& z_1& z_2&...&z_n
-     * \end{bmatrix}\f]
-     * where the same subscript represents a point. The shape of A assume to be [3, n]
-     * The points matrix A can be rotated by toRotMat3x3() * A.
-     * The result has 3 rows and n columns too.
-
-     * For example
-     * ```
-     * double angle = CV_PI;
-     * Vec3d axis{0,0,1};
-     * Quatd q_unit = Quatd::createFromAngleAxis(angle, axis); //quaternion could also be get by interpolation by two or more quaternions.
-     *
-     * //assume there is two points (1,0,0) and (1,0,1) to be rotated
-     * Mat pointsA = (Mat_<double>(2, 3) << 1,0,0,1,0,1);
-     * //change the shape
-     * pointsA = pointsA.t();
-     * // rotate 180 degrees around the z axis
-     * Mat new_point = q_unit.toRotMat3x3() * pointsA;
-     * // print two points
-     * cout << new_point << endl;
-     * ```
-     */
-    Matx<_Tp, 3, 3> toRotMat3x3(QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief transform a quaternion to a 4x4 rotation matrix.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, this quaternion assume to be a unit quaternion and
-     * this function will save some computations. Otherwise, this function will normalize this
-     * quaternion at first then do the transformation.
-     *
-     * The operations is similar as toRotMat3x3
-     * except that the points matrix should have the form
-     * \f[\begin{bmatrix}
-     * x_0& x_1& x_2&...&x_n\\
-     * y_0& y_1& y_2&...&y_n\\
-     * z_0& z_1& z_2&...&z_n\\
-     * 0&0&0&...&0
-     * \end{bmatrix}\f]
-     *
-     * @sa toRotMat3x3
-     */
-
-    Matx<_Tp, 4, 4> toRotMat4x4(QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief transform the this quaternion to a Vec<T, 4>.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.toVec();
-     * ```
-     */
-    Vec<_Tp, 4> toVec() const;
-
-    /**
-     * @brief transform this quaternion to a Rotation vector.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, this quaternion assume to be a unit quaternion and
-     * this function will save some computations.
-     * Rotation vector rVec is defined as:
-     * \f[ rVec = [\theta v_x, \theta v_y, \theta v_z]\f]
-     * where \f$\theta\f$ represents rotation angle, and \f$\boldsymbol{v}\f$ represents the normalized rotation axis.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.toRotVec();
-     *
-     * QuatAssumeType assumeUnit = QUAT_ASSUME_UNIT;
-     * q.normalize().toRotVec(assumeUnit); //answer is same as q.toRotVec().
-     * ```
-     */
-    Vec<_Tp, 3> toRotVec(QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief get the angle of quaternion, it returns the rotation angle.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, this quaternion assume to be a unit quaternion and
-     * this function will save some computations.
-     * \f[\psi = 2 *arccos(\frac{w}{||q||})\f]
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.getAngle();
-     *
-     * QuatAssumeType assumeUnit = QUAT_ASSUME_UNIT;
-     * q.normalize().getAngle(assumeUnit);//same as q.getAngle().
-     * ```
-     * @note It always return the value between \f$[0, 2\pi]\f$.
-     */
-    _Tp getAngle(QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief get the axis of quaternion, it returns a vector of length 3.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, this quaternion assume to be a unit quaternion and
-     * this function will save some computations.
-     *
-     * the unit axis \f$\boldsymbol{u}\f$ is defined by
-     * \f[\begin{equation}
-     *    \begin{split}
-     *      \boldsymbol{v}
-     *      &= \boldsymbol{u} ||\boldsymbol{v}||\\
-     *      &= \boldsymbol{u}||q||sin(\frac{\theta}{2})
-     *    \end{split}
-     *    \end{equation}\f]
-     *  where \f$v=[x, y ,z]\f$ and \f$\theta\f$ represents rotation angle.
-     *
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * q.getAxis();
-     *
-     * QuatAssumeType assumeUnit = QUAT_ASSUME_UNIT;
-     * q.normalize().getAxis(assumeUnit);//same as q.getAxis()
-     * ```
-     */
-    Vec<_Tp, 3> getAxis(QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT) const;
-
-    /**
-     * @brief return the dot between quaternion \f$q\f$ and this quaternion.
-     *
-     * dot(p, q) is a good metric of how close the quaternions are.
-     * Indeed, consider the unit quaternion difference \f$p^{-1} * q\f$, its real part is dot(p, q).
-     * At the same time its real part is equal to \f$\cos(\beta/2)\f$ where \f$\beta\f$ is
-     * an angle of rotation between p and q, i.e.,
-     * Therefore, the closer dot(p, q) to 1,
-     * the smaller rotation between them.
-     * \f[p \cdot q = p.w \cdot q.w + p.x \cdot q.x + p.y \cdot q.y + p.z \cdot q.z\f]
-     * @param q the other quaternion.
-     *
-     * For example
-     * ```
-     * Quatd q(1,2,3,4);
-     * Quatd p(5,6,7,8);
-     * p.dot(q);
-     * ```
-     */
-    _Tp dot(Quat<_Tp> q) const;
-
-    /**
-     * @brief To calculate the interpolation from \f$q_0\f$ to \f$q_1\f$ by Linear Interpolation(Nlerp)
-     * For two quaternions, this interpolation curve can be displayed as:
-     * \f[Lerp(q_0, q_1, t) = (1 - t)q_0 + tq_1.\f]
-     * Obviously, the lerp will interpolate along a straight line if we think of \f$q_0\f$ and \f$q_1\f$ as a vector
-     * in a two-dimensional space. When \f$t = 0\f$, it returns \f$q_0\f$ and when \f$t= 1\f$, it returns \f$q_1\f$.
-     * \f$t\f$ should to be ranged in \f$[0, 1]\f$ normally.
-     * @param q0 a quaternion used in linear interpolation.
-     * @param q1 a quaternion used in linear interpolation.
-     * @param t percent of vector \f$\overrightarrow{q_0q_1}\f$ over a range [0, 1].
-     * @note it returns a non-unit quaternion.
-     */
-    static Quat<_Tp> lerp(const Quat<_Tp> &q0, const Quat &q1, const _Tp t);
-
-    /**
-     * @brief To calculate the interpolation from \f$q_0\f$ to \f$q_1\f$ by Normalized Linear Interpolation(Nlerp).
-     * it returns a normalized quaternion of Linear Interpolation(Lerp).
-     * \f[ Nlerp(q_0, q_1, t) = \frac{(1 - t)q_0 + tq_1}{||(1 - t)q_0 + tq_1||}.\f]
-     * The interpolation will always choose the shortest path but the constant speed is not guaranteed.
-     * @param q0 a quaternion used in normalized linear interpolation.
-     * @param q1 a quaternion used in normalized linear interpolation.
-     * @param t percent of vector \f$\overrightarrow{q_0q_1}\f$ over a range [0, 1].
-     * @param assumeUnit if QUAT_ASSUME_UNIT, all input quaternions assume to be unit quaternion. Otherwise, all inputs
-     quaternion will be normalized inside the function.
-     * @sa lerp
-     */
-    static Quat<_Tp> nlerp(const Quat<_Tp> &q0, const Quat &q1, const _Tp t, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT);
-
-    /**
-     @brief To calculate the interpolation between \f$q_0\f$ and \f$q_1\f$ by Spherical Linear
-     Interpolation(Slerp), which can be defined as:
-    \f[ Slerp(q_0, q_1, t) = \frac{\sin((1-t)\theta)}{\sin(\theta)}q_0 + \frac{\sin(t\theta)}{\sin(\theta)}q_1\f]
-    where \f$\theta\f$ can be calculated as:
-    \f[\theta=cos^{-1}(q_0\cdot q_1)\f]
-    resulting from the both of their norm is unit.
-    @param q0 a quaternion used in Slerp.
-    @param q1 a quaternion used in Slerp.
-    @param t percent of angle between \f$q_0\f$ and \f$q_1\f$ over a range [0, 1].
-    @param assumeUnit if QUAT_ASSUME_UNIT, all input quaternions assume to be unit quaternions. Otherwise, all input
-    quaternions will be normalized inside the function.
-    @param directChange if QUAT_ASSUME_UNIT, the interpolation will choose the nearest path.
-    @note If the interpolation angle is small, the error between Nlerp and Slerp is not so large. To improve efficiency and
-    avoid zero division error, we use Nlerp instead of Slerp.
-    */
-    static Quat<_Tp> slerp(const Quat<_Tp> &q0, const Quat &q1, const _Tp t, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT, bool directChange=true);
-
-    /**
-     * @brief To calculate the interpolation between \f$q_0\f$,\f$q_1\f$,\f$q_2\f$,\f$q_3\f$  by Spherical and quadrangle(Squad). This could be defined as:
-     * \f[Squad(q_i, s_i, s_{i+1}, q_{i+1}, t) = Slerp(Slerp(q_i, q_{i+1}, t), Slerp(s_i, s_{i+1}, t), 2t(1-t))\f]
-     * where
-     * \f[s_i = q_i\exp(-\frac{\log(q^*_iq_{i+1}) + \log(q^*_iq_{i-1})}{4})\f]
-     *
-     * The Squad expression is analogous to the \f$B\acute{e}zier\f$ curve, but involves spherical linear
-     * interpolation instead of simple linear interpolation. Each \f$s_i\f$ needs to be calculated by three
-     * quaternions.
-     *
-     * @param q0 the first quaternion.
-     * @param s0 the second quaternion.
-     * @param s1 the third quaternion.
-     * @param q1 thr fourth quaternion.
-     * @param t interpolation parameter of quadratic and linear interpolation over a range \f$[0, 1]\f$.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, all input quaternions assume to be unit quaternion. Otherwise, all input
-     * quaternions will be normalized inside the function.
-     * @param directChange if QUAT_ASSUME_UNIT, squad will find the nearest path to interpolate.
-     * @sa interPoint, spline
-     */
-    static Quat<_Tp> squad(const Quat<_Tp> &q0, const Quat<_Tp> &s0,
-                            const Quat<_Tp> &s1, const Quat<_Tp> &q1,
-                            const _Tp t, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT,
-                            bool directChange=true);
-
-    /**
-     * @brief This is the part calculation of squad.
-     * To calculate the intermedia quaternion \f$s_i\f$ between each three quaternion
-     * \f[s_i = q_i\exp(-\frac{\log(q^*_iq_{i+1}) + \log(q^*_iq_{i-1})}{4}).\f]
-     * @param q0 the first quaternion.
-     * @param q1 the second quaternion.
-     * @param q2 the third quaternion.
-     * @param assumeUnit if QUAT_ASSUME_UNIT, all input quaternions assume to be unit quaternion. Otherwise, all input
-     * quaternions will be normalized inside the function.
-     * @sa squad
-     */
-    static Quat<_Tp> interPoint(const Quat<_Tp> &q0, const Quat<_Tp> &q1,
-                                 const Quat<_Tp> &q2, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT);
-
-    /**
-     * @brief to calculate a quaternion which is the result of a \f$C^1\f$ continuous
-     * spline curve constructed by squad at the ratio t. Here, the interpolation values are
-     * between \f$q_1\f$ and \f$q_2\f$. \f$q_0\f$ and \f$q_2\f$ are used to ensure the \f$C^1\f$
-     * continuity. if t = 0, it returns \f$q_1\f$, if t = 1, it returns \f$q_2\f$.
-     * @param q0 the first input quaternion to ensure \f$C^1\f$ continuity.
-     * @param q1 the second input quaternion.
-     * @param q2 the third input quaternion.
-     * @param q3 the fourth input quaternion the same use of \f$q1\f$.
-     * @param t ratio over a range [0, 1].
-     * @param assumeUnit if QUAT_ASSUME_UNIT, \f$q_0, q_1, q_2, q_3\f$ assume to be unit quaternion. Otherwise, all input
-     * quaternions will be normalized inside the function.
-     *
-     * For example:
-     *
-     * If there are three double quaternions \f$v_0, v_1, v_2\f$ waiting to be interpolated.
-     *
-     * Interpolation between \f$v_0\f$ and \f$v_1\f$ with a ratio \f$t_0\f$ could be calculated as
-     * ```
-     * Quatd::spline(v0, v0, v1, v2, t0);
-     * ```
-     * Interpolation between \f$v_1\f$ and \f$v_2\f$ with a ratio \f$t_0\f$ could be calculated as
-     * ```
-     * Quatd::spline(v0, v1, v2, v2, t0);
-     * ```
-     * @sa squad, slerp
-     */
-    static Quat<_Tp> spline(const Quat<_Tp> &q0, const Quat<_Tp> &q1,
-                            const Quat<_Tp> &q2, const Quat<_Tp> &q3,
-                            const _Tp t, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT);
-
-    /**
-     * @brief Return opposite quaternion \f$-p\f$
-     * which satisfies \f$p + (-p) = 0.\f$
-     *
-     * For example
-     * ```
-     * Quatd q{1, 2, 3, 4};
-     * std::cout << -q << std::endl; // [-1, -2, -3, -4]
-     * ```
-     */
-    Quat<_Tp> operator-() const;
-
-    /**
-     * @brief return true if two quaternions p and q are nearly equal, i.e. when the absolute
-     * value of each \f$p_i\f$ and \f$q_i\f$ is less than CV_QUAT_EPS.
-     */
-    bool operator==(const Quat<_Tp>&) const;
-
-    /**
-     * @brief Addition operator of two quaternions p and q.
-     * It returns a new quaternion that each value is the sum of \f$p_i\f$ and \f$q_i\f$.
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * Quatd q{5, 6, 7, 8};
-     * std::cout << p + q << std::endl; //[6, 8, 10, 12]
-     * ```
-     */
-    Quat<_Tp> operator+(const Quat<_Tp>&) const;
-
-    /**
-     * @brief Addition assignment operator of two quaternions p and q.
-     * It adds right operand to the left operand and assign the result to left operand.
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * Quatd q{5, 6, 7, 8};
-     * p += q; // equivalent to p = p + q
-     * std::cout << p << std::endl; //[6, 8, 10, 12]
-     *
-     * ```
-     */
-    Quat<_Tp>& operator+=(const Quat<_Tp>&);
-
-    /**
-     * @brief Subtraction operator of two quaternions p and q.
-     * It returns a new quaternion that each value is the sum of \f$p_i\f$ and \f$-q_i\f$.
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * Quatd q{5, 6, 7, 8};
-     * std::cout << p - q << std::endl; //[-4, -4, -4, -4]
-     * ```
-     */
-    Quat<_Tp> operator-(const Quat<_Tp>&) const;
-
-    /**
-     * @brief Subtraction assignment operator of two quaternions p and q.
-     * It subtracts right operand from the left operand and assign the result to left operand.
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * Quatd q{5, 6, 7, 8};
-     * p -= q; // equivalent to p = p - q
-     * std::cout << p << std::endl; //[-4, -4, -4, -4]
-     *
-     * ```
-     */
-    Quat<_Tp>& operator-=(const Quat<_Tp>&);
-
-    /**
-     * @brief Multiplication assignment operator of two quaternions q and p.
-     * It multiplies right operand with the left operand and assign the result to left operand.
-     *
-     * Rule of quaternion multiplication:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p * q &= [p_0, \boldsymbol{u}]*[q_0, \boldsymbol{v}]\\
-     * &=[p_0q_0 - \boldsymbol{u}\cdot \boldsymbol{v}, p_0\boldsymbol{v} + q_0\boldsymbol{u}+ \boldsymbol{u}\times \boldsymbol{v}].
-     * \end{split}
-     * \end{equation}
-     * \f]
-     * where \f$\cdot\f$ means dot product and \f$\times \f$ means cross product.
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * Quatd q{5, 6, 7, 8};
-     * p *= q; // equivalent to p = p * q
-     * std::cout << p << std::endl; //[-60, 12, 30, 24]
-     * ```
-     */
-    Quat<_Tp>& operator*=(const Quat<_Tp>&);
-
-    /**
-     * @brief Multiplication assignment operator of a quaternions and a scalar.
-     * It multiplies right operand with the left operand and assign the result to left operand.
-     *
-     * Rule of quaternion multiplication with a scalar:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p * s &= [w, x, y, z] * s\\
-     * &=[w * s, x * s, y * s, z * s].
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * double s = 2.0;
-     * p *= s; // equivalent to p = p * s
-     * std::cout << p << std::endl; //[2.0, 4.0, 6.0, 8.0]
-     * ```
-     * @note the type of scalar should be equal to the quaternion.
-     */
-    Quat<_Tp>& operator*=(const _Tp s);
-
-    /**
-     * @brief Multiplication operator of two quaternions q and p.
-     * Multiplies values on either side of the operator.
-     *
-     * Rule of quaternion multiplication:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p * q &= [p_0, \boldsymbol{u}]*[q_0, \boldsymbol{v}]\\
-     * &=[p_0q_0 - \boldsymbol{u}\cdot \boldsymbol{v}, p_0\boldsymbol{v} + q_0\boldsymbol{u}+ \boldsymbol{u}\times \boldsymbol{v}].
-     * \end{split}
-     * \end{equation}
-     * \f]
-     * where \f$\cdot\f$ means dot product and \f$\times \f$ means cross product.
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * Quatd q{5, 6, 7, 8};
-     * std::cout << p * q << std::endl; //[-60, 12, 30, 24]
-     * ```
-     */
-    Quat<_Tp> operator*(const Quat<_Tp>&) const;
-
-    /**
-     * @brief Division operator of a quaternions and a scalar.
-     * It divides left operand with the right operand and assign the result to left operand.
-     *
-     * Rule of quaternion division with a scalar:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p / s &= [w, x, y, z] / s\\
-     * &=[w/s, x/s, y/s, z/s].
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * double s = 2.0;
-     * p /= s; // equivalent to p = p / s
-     * std::cout << p << std::endl; //[0.5, 1, 1.5, 2]
-     * ```
-     * @note the type of scalar should be equal to this quaternion.
-     */
-    Quat<_Tp> operator/(const _Tp s) const;
-
-    /**
-     * @brief Division operator of two quaternions p and q.
-     * Divides left hand operand by right hand operand.
-     *
-     * Rule of quaternion division with a scalar:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p / q &= p * q.inv()\\
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * Quatd q{5, 6, 7, 8};
-     * std::cout << p / q << std::endl; // equivalent to p * q.inv()
-     * ```
-     */
-    Quat<_Tp> operator/(const Quat<_Tp>&) const;
-
-    /**
-     * @brief Division assignment operator of a quaternions and a scalar.
-     * It divides left operand with the right operand and assign the result to left operand.
-     *
-     * Rule of quaternion division with a scalar:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p / s &= [w, x, y, z] / s\\
-     * &=[w / s, x / s, y / s, z / s].
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * double s = 2.0;;
-     * p /= s; // equivalent to p = p / s
-     * std::cout << p << std::endl; //[0.5, 1.0, 1.5, 2.0]
-     * ```
-     * @note the type of scalar should be equal to the quaternion.
-     */
-    Quat<_Tp>& operator/=(const _Tp s);
-
-    /**
-     * @brief Division assignment operator of two quaternions p and q;
-     * It divides left operand with the right operand and assign the result to left operand.
-     *
-     * Rule of quaternion division with a quaternion:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p / q&= p * q.inv()\\
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * Quatd q{5, 6, 7, 8};
-     * p /= q; // equivalent to p = p * q.inv()
-     * std::cout << p << std::endl;
-     * ```
-     */
-    Quat<_Tp>& operator/=(const Quat<_Tp>&);
-
-    _Tp& operator[](std::size_t n);
-
-    const _Tp& operator[](std::size_t n) const;
-
-    /**
-     * @brief Subtraction operator of a scalar and a quaternions.
-     * Subtracts right hand operand from left hand operand.
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * double scalar = 2.0;
-     * std::cout << scalar - p << std::endl; //[1.0, -2, -3, -4]
-     * ```
-     * @note the type of scalar should be equal to the quaternion.
-     */
-    template <typename T>
-    friend Quat<T> cv::operator-(const T s, const Quat<T>&);
-
-    /**
-     * @brief Subtraction operator of a quaternions and a scalar.
-     * Subtracts right hand operand from left hand operand.
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * double scalar = 2.0;
-     * std::cout << p - scalar << std::endl; //[-1.0, 2, 3, 4]
-     * ```
-     * @note the type of scalar should be equal to the quaternion.
-     */
-    template <typename T>
-    friend Quat<T> cv::operator-(const Quat<T>&, const T s);
-
-    /**
-     * @brief Addition operator of a quaternions and a scalar.
-     * Adds right hand operand from left hand operand.
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * double scalar = 2.0;
-     * std::cout << scalar + p << std::endl; //[3.0, 2, 3, 4]
-     * ```
-     * @note the type of scalar should be equal to the quaternion.
-     */
-    template <typename T>
-    friend Quat<T> cv::operator+(const T s, const Quat<T>&);
-
-    /**
-     * @brief Addition operator of a quaternions and a scalar.
-     * Adds right hand operand from left hand operand.
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * double scalar = 2.0;
-     * std::cout << p + scalar << std::endl; //[3.0, 2, 3, 4]
-     * ```
-     * @note the type of scalar should be equal to the quaternion.
-     */
-    template <typename T>
-    friend Quat<T> cv::operator+(const Quat<T>&, const T s);
-
-    /**
-     * @brief Multiplication operator of a scalar and a quaternions.
-     * It multiplies right operand with the left operand and assign the result to left operand.
-     *
-     * Rule of quaternion multiplication with a scalar:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p * s &= [w, x, y, z] * s\\
-     * &=[w * s, x * s, y * s, z * s].
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * double s = 2.0;
-     * std::cout << s * p << std::endl; //[2.0, 4.0, 6.0, 8.0]
-     * ```
-     * @note the type of scalar should be equal to the quaternion.
-     */
-    template <typename T>
-    friend Quat<T> cv::operator*(const T s, const Quat<T>&);
-
-    /**
-     * @brief Multiplication operator of a quaternion and a scalar.
-     * It multiplies right operand with the left operand and assign the result to left operand.
-     *
-     * Rule of quaternion multiplication with a scalar:
-     * \f[
-     * \begin{equation}
-     * \begin{split}
-     * p * s &= [w, x, y, z] * s\\
-     * &=[w * s, x * s, y * s, z * s].
-     * \end{split}
-     * \end{equation}
-     * \f]
-     *
-     * For example
-     * ```
-     * Quatd p{1, 2, 3, 4};
-     * double s = 2.0;
-     * std::cout << p * s << std::endl; //[2.0, 4.0, 6.0, 8.0]
-     * ```
-     * @note the type of scalar should be equal to the quaternion.
-     */
-    template <typename T>
-    friend Quat<T> cv::operator*(const Quat<T>&, const T s);
-
-    template <typename S>
-    friend std::ostream& cv::operator<<(std::ostream&, const Quat<S>&);
-
-    /**
-     * @brief Transform a quaternion q to Euler angles.
-     *
-     *
-     * When transforming a quaternion \f$q = w + x\boldsymbol{i} + y\boldsymbol{j} + z\boldsymbol{k}\f$ to Euler angles, rotation matrix M can be calculated by:
-     * \f[ \begin{aligned} {M} &={\begin{bmatrix}1-2(y^{2}+z^{2})&2(xy-zx)&2(xz+yw)\\2(xy+zw)&1-2(x^{2}+z^{2})&2(yz-xw)\\2(xz-yw)&2(yz+xw)&1-2(x^{2}+y^{2})\end{bmatrix}}\end{aligned}.\f]
-     * On the other hand, the rotation matrix can be obtained from Euler angles.
-     * Using intrinsic rotations with Euler angles type XYZ as an example,
-     * \f$\theta_1 \f$, \f$\theta_2 \f$, \f$\theta_3 \f$ are three angles for Euler angles, the rotation matrix R can be calculated by:\f[R =X(\theta_1)Y(\theta_2)Z(\theta_3)
-     * ={\begin{bmatrix}\cos\theta_{2}\cos\theta_{3}&-\cos\theta_{2}\sin\theta_{3}&\sin\theta_{2}\\\cos\theta_{1}\sin\theta_{3}+\cos\theta_{3}\sin\theta_{1}\sin\theta_{2}&\cos\theta_{1}\cos\theta_{3}-\sin\theta_{1}\sin\theta_{2}\sin\theta_{3}&-\cos\theta_{2}\sin\theta_{1}\\\sin\theta_{1}\sin\theta_{3}-\cos\theta_{1}\cos\theta_{3}\sin\theta_{2}&\cos\theta_{3}\sin\theta_{1}+\cos\theta_{1}\sin\theta_{2}\sin\theta_{3}&\cos\theta_{1}\cos_{2}\end{bmatrix}}\f]
-     * Rotation matrix M and R are equal. As long as \f$ s_{2} \neq 1 \f$, by comparing each element of two matrices ,the solution is\f$\begin{cases} \theta_1 = \arctan2(-m_{23},m_{33})\\\theta_2 = arcsin(m_{13}) \\\theta_3 = \arctan2(-m_{12},m_{11}) \end{cases}\f$.
-     *
-     * When \f$ s_{2}=1\f$ or \f$ s_{2}=-1\f$, the gimbal lock occurs. The function will prompt "WARNING: Gimbal Lock will occur. Euler angles is non-unique. For intrinsic rotations, we set the third angle to 0, and for external rotation, we set the first angle to 0.".
-     *
-     * When \f$ s_{2}=1\f$ ,
-     * The rotation matrix R is \f$R = {\begin{bmatrix}0&0&1\\\sin(\theta_1+\theta_3)&\cos(\theta_1+\theta_3)&0\\-\cos(\theta_1+\theta_3)&\sin(\theta_1+\theta_3)&0\end{bmatrix}}\f$.
-     *
-     * The number of solutions is infinite with the condition \f$\begin{cases} \theta_1+\theta_3 = \arctan2(m_{21},m_{22})\\ \theta_2=\pi/2 \end{cases}\ \f$.
-     *
-     * We set \f$ \theta_3 = 0\f$, the solution is \f$\begin{cases} \theta_1=\arctan2(m_{21},m_{22})\\ \theta_2=\pi/2\\ \theta_3=0 \end{cases}\f$.
-     *
-     * When \f$ s_{2}=-1\f$,
-     * The rotation matrix R is \f$X_{1}Y_{2}Z_{3}={\begin{bmatrix}0&0&-1\\-\sin(\theta_1-\theta_3)&\cos(\theta_1-\theta_3)&0\\\cos(\theta_1-\theta_3)&\sin(\theta_1-\theta_3)&0\end{bmatrix}}\f$.
-     *
-     * The number of solutions is infinite with the condition \f$\begin{cases} \theta_1+\theta_3 = \arctan2(m_{32},m_{22})\\ \theta_2=\pi/2 \end{cases}\ \f$.
-     *
-     * We set \f$ \theta_3 = 0\f$, the solution is \f$ \begin{cases}\theta_1=\arctan2(m_{32},m_{22}) \\ \theta_2=-\pi/2\\  \theta_3=0\end{cases}\f$.
-     *
-     * Since \f$ sin \theta\in [-1,1] \f$ and \f$ cos \theta \in [-1,1] \f$, the unnormalized quaternion will cause computational troubles. For this reason, this function will normalize the quaternion at first and @ref QuatAssumeType is not needed.
-     *
-     * When the gimbal lock occurs, we set \f$\theta_3 = 0\f$ for intrinsic rotations or \f$\theta_1 = 0\f$ for extrinsic rotations.
-     *
-     * As a result, for every Euler angles type, we can get solution as shown in the following table.
-     * EulerAnglesType  | Ordinary | \f$\theta_2 = π/2\f$ | \f$\theta_2 = -π/2\f$
-     * ------------- | -------------| -------------| -------------
-     * INT_XYZ|\f$ \theta_1 = \arctan2(-m_{23},m_{33})\\\theta_2 = \arcsin(m_{13}) \\\theta_3= \arctan2(-m_{12},m_{11}) \f$|\f$ \theta_1=\arctan2(m_{21},m_{22})\\ \theta_2=\pi/2\\ \theta_3=0 \f$|\f$ \theta_1=\arctan2(m_{32},m_{22})\\ \theta_2=-\pi/2\\ \theta_3=0 \f$
-     * INT_XZY|\f$ \theta_1 = \arctan2(m_{32},m_{22})\\\theta_2 = -\arcsin(m_{12}) \\\theta_3= \arctan2(m_{13},m_{11}) \f$|\f$ \theta_1=\arctan2(m_{31},m_{33})\\ \theta_2=\pi/2\\ \theta_3=0 \f$|\f$ \theta_1=\arctan2(-m_{23},m_{33})\\ \theta_2=-\pi/2\\ \theta_3=0 \f$
-     * INT_YXZ|\f$ \theta_1 = \arctan2(m_{13},m_{33})\\\theta_2 = -\arcsin(m_{23}) \\\theta_3= \arctan2(m_{21},m_{22}) \f$|\f$ \theta_1=\arctan2(m_{12},m_{11})\\ \theta_2=\pi/2\\ \theta_3=0 \f$|\f$ \theta_1=\arctan2(-m_{12},m_{11})\\ \theta_2=-\pi/2\\ \theta_3=0 \f$
-     * INT_YZX|\f$ \theta_1 = \arctan2(-m_{31},m_{11})\\\theta_2 = \arcsin(m_{21}) \\\theta_3= \arctan2(-m_{23},m_{22}) \f$|\f$ \theta_1=\arctan2(m_{13},m_{33})\\ \theta_2=\pi/2\\ \theta_3=0 \f$|\f$ \theta_1=\arctan2(m_{13},m_{12})\\ \theta_2=-\pi/2\\ \theta_3=0 \f$
-     * INT_ZXY|\f$ \theta_1 = \arctan2(-m_{12},m_{22})\\\theta_2 = \arcsin(m_{32}) \\\theta_3= \arctan2(-m_{31},m_{33}) \f$|\f$ \theta_1=\arctan2(m_{21},m_{11})\\ \theta_2=\pi/2\\ \theta_3=0 \f$|\f$ \theta_1=\arctan2(m_{21},m_{11})\\ \theta_2=-\pi/2\\ \theta_3=0 \f$
-     * INT_ZYX|\f$ \theta_1 = \arctan2(m_{21},m_{11})\\\theta_2 = \arcsin(-m_{31}) \\\theta_3= \arctan2(m_{32},m_{33}) \f$|\f$ \theta_1=\arctan2(m_{23},m_{22})\\ \theta_2=\pi/2\\ \theta_3=0 \f$|\f$ \theta_1=\arctan2(-m_{12},m_{22})\\ \theta_2=-\pi/2\\ \theta_3=0 \f$
-     * EXT_XYZ|\f$ \theta_1 = \arctan2(m_{32},m_{33})\\\theta_2 = \arcsin(-m_{31}) \\\ \theta_3 = \arctan2(m_{21},m_{11})\f$|\f$ \theta_1= 0\\ \theta_2=\pi/2\\ \theta_3=\arctan2(m_{23},m_{22}) \f$|\f$ \theta_1=0\\ \theta_2=-\pi/2\\ \theta_3=\arctan2(-m_{12},m_{22}) \f$
-     * EXT_XZY|\f$ \theta_1 = \arctan2(-m_{23},m_{22})\\\theta_2 = \arcsin(m_{21}) \\\theta_3=  \arctan2(-m_{31},m_{11})\f$|\f$ \theta_1= 0\\ \theta_2=\pi/2\\ \theta_3=\arctan2(m_{13},m_{33}) \f$|\f$ \theta_1=0\\ \theta_2=-\pi/2\\ \theta_3=\arctan2(m_{13},m_{12}) \f$
-     * EXT_YXZ|\f$ \theta_1 = \arctan2(-m_{31},m_{33}) \\\theta_2 = \arcsin(m_{32}) \\\theta_3= \arctan2(-m_{12},m_{22})\f$|\f$ \theta_1= 0\\ \theta_2=\pi/2\\ \theta_3=\arctan2(m_{21},m_{11}) \f$|\f$ \theta_1=0\\ \theta_2=-\pi/2\\ \theta_3=\arctan2(m_{21},m_{11}) \f$
-     * EXT_YZX|\f$ \theta_1 = \arctan2(m_{13},m_{11})\\\theta_2 = -\arcsin(m_{12}) \\\theta_3= \arctan2(m_{32},m_{22})\f$|\f$ \theta_1= 0\\ \theta_2=\pi/2\\ \theta_3=\arctan2(m_{31},m_{33}) \f$|\f$ \theta_1=0\\ \theta_2=-\pi/2\\ \theta_3=\arctan2(-m_{23},m_{33}) \f$
-     * EXT_ZXY|\f$ \theta_1 = \arctan2(m_{21},m_{22})\\\theta_2 = -\arcsin(m_{23}) \\\theta_3= \arctan2(m_{13},m_{33})\f$|\f$ \theta_1= 0\\ \theta_2=\pi/2\\ \theta_3=\arctan2(m_{12},m_{11}) \f$|\f$ \theta_1= 0\\ \theta_2=-\pi/2\\ \theta_3=\arctan2(-m_{12},m_{11}) \f$
-     * EXT_ZYX|\f$ \theta_1 = \arctan2(-m_{12},m_{11})\\\theta_2 = \arcsin(m_{13}) \\\theta_3= \arctan2(-m_{23},m_{33})\f$|\f$ \theta_1=0\\ \theta_2=\pi/2\\ \theta_3=\arctan2(m_{21},m_{22}) \f$|\f$ \theta_1=0\\ \theta_2=-\pi/2\\ \theta_3=\arctan2(m_{32},m_{22}) \f$
-     *
-     *  EulerAnglesType  | Ordinary | \f$\theta_2 = 0\f$ | \f$\theta_2 = π\f$
-     * ------------- | -------------| -------------| -------------
-     * INT_XYX| \f$ \theta_1 = \arctan2(m_{21},-m_{31})\\\theta_2 =\arccos(m_{11}) \\\theta_3 = \arctan2(m_{12},m_{13}) \f$| \f$ \theta_1=\arctan2(m_{32},m_{33})\\ \theta_2=0\\ \theta_3=0 \f$| \f$ \theta_1=\arctan2(m_{23},m_{22})\\ \theta_2=\pi\\ \theta_3=0 \f$
-     * INT_XZX| \f$ \theta_1 = \arctan2(m_{31},m_{21})\\\theta_2 = \arccos(m_{11}) \\\theta_3 = \arctan2(m_{13},-m_{12}) \f$| \f$ \theta_1=\arctan2(m_{32},m_{33})\\ \theta_2=0\\ \theta_3=0 \f$| \f$ \theta_1=\arctan2(-m_{32},m_{33})\\ \theta_2=\pi\\ \theta_3=0 \f$
-     * INT_YXY| \f$ \theta_1 = \arctan2(m_{12},m_{32})\\\theta_2 = \arccos(m_{22}) \\\theta_3 = \arctan2(m_{21},-m_{23}) \f$| \f$ \theta_1=\arctan2(m_{13},m_{11})\\ \theta_2=0\\ \theta_3=0 \f$| \f$ \theta_1=\arctan2(-m_{31},m_{11})\\ \theta_2=\pi\\ \theta_3=0 \f$
-     * INT_YZY| \f$ \theta_1 = \arctan2(m_{32},-m_{12})\\\theta_2 = \arccos(m_{22}) \\\theta_3 =\arctan2(m_{23},m_{21}) \f$| \f$ \theta_1=\arctan2(m_{13},m_{11})\\ \theta_2=0\\ \theta_3=0 \f$| \f$ \theta_1=\arctan2(m_{13},-m_{11})\\ \theta_2=\pi\\ \theta_3=0 \f$
-     * INT_ZXZ| \f$ \theta_1 = \arctan2(-m_{13},m_{23})\\\theta_2 = \arccos(m_{33}) \\\theta_3 =\arctan2(m_{31},m_{32}) \f$| \f$ \theta_1=\arctan2(m_{21},m_{22})\\ \theta_2=0\\ \theta_3=0 \f$| \f$ \theta_1=\arctan2(m_{21},m_{11})\\ \theta_2=\pi\\ \theta_3=0 \f$
-     * INT_ZYZ| \f$ \theta_1 = \arctan2(m_{23},m_{13})\\\theta_2 = \arccos(m_{33}) \\\theta_3 = \arctan2(m_{32},-m_{31}) \f$| \f$ \theta_1=\arctan2(m_{21},m_{11})\\ \theta_2=0\\ \theta_3=0 \f$| \f$ \theta_1=\arctan2(m_{21},m_{11})\\ \theta_2=\pi\\ \theta_3=0 \f$
-     * EXT_XYX| \f$ \theta_1 = \arctan2(m_{12},m_{13}) \\\theta_2 = \arccos(m_{11}) \\\theta_3 = \arctan2(m_{21},-m_{31})\f$| \f$ \theta_1=0\\ \theta_2=0\\ \theta_3=\arctan2(m_{32},m_{33}) \f$| \f$ \theta_1= 0\\ \theta_2=\pi\\ \theta_3= \arctan2(m_{23},m_{22}) \f$
-     * EXT_XZX| \f$ \theta_1 = \arctan2(m_{13},-m_{12})\\\theta_2 = \arccos(m_{11}) \\\theta_3 = \arctan2(m_{31},m_{21})\f$| \f$ \theta_1= 0\\ \theta_2=0\\ \theta_3=\arctan2(m_{32},m_{33}) \f$| \f$ \theta_1= 0\\ \theta_2=\pi\\ \theta_3=\arctan2(-m_{32},m_{33}) \f$
-     * EXT_YXY| \f$ \theta_1 = \arctan2(m_{21},-m_{23})\\\theta_2 = \arccos(m_{22}) \\\theta_3 = \arctan2(m_{12},m_{32}) \f$| \f$ \theta_1= 0\\ \theta_2=0\\ \theta_3=\arctan2(m_{13},m_{11}) \f$| \f$ \theta_1= 0\\ \theta_2=\pi\\ \theta_3=\arctan2(-m_{31},m_{11}) \f$
-     * EXT_YZY| \f$ \theta_1 = \arctan2(m_{23},m_{21}) \\\theta_2 = \arccos(m_{22}) \\\theta_3 = \arctan2(m_{32},-m_{12}) \f$| \f$ \theta_1= 0\\ \theta_2=0\\ \theta_3=\arctan2(m_{13},m_{11}) \f$| \f$ \theta_1=0\\ \theta_2=\pi\\ \theta_3=\arctan2(m_{13},-m_{11}) \f$
-     * EXT_ZXZ| \f$ \theta_1 = \arctan2(m_{31},m_{32}) \\\theta_2 = \arccos(m_{33}) \\\theta_3 = \arctan2(-m_{13},m_{23})\f$| \f$ \theta_1=0\\ \theta_2=0\\ \theta_3=\arctan2(m_{21},m_{22}) \f$| \f$ \theta_1= 0\\ \theta_2=\pi\\ \theta_3=\arctan2(m_{21},m_{11}) \f$
-     * EXT_ZYZ| \f$ \theta_1 = \arctan2(m_{32},-m_{31})\\\theta_2 = \arccos(m_{33}) \\\theta_3 = \arctan2(m_{23},m_{13}) \f$| \f$ \theta_1=0\\ \theta_2=0\\ \theta_3=\arctan2(m_{21},m_{11}) \f$| \f$ \theta_1= 0\\ \theta_2=\pi\\ \theta_3=\arctan2(m_{21},m_{11}) \f$
-     *
-     * @param eulerAnglesType the convertion Euler angles type
-     */
-
-    Vec<_Tp, 3> toEulerAngles(QuatEnum::EulerAnglesType eulerAnglesType);
-
-    _Tp w, x, y, z;
-
-};
-
-template <typename T>
-Quat<T> inv(const Quat<T> &q, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT);
-
-template <typename T>
-Quat<T> sinh(const Quat<T> &q);
-
-template <typename T>
-Quat<T> cosh(const Quat<T> &q);
-
-template <typename T>
-Quat<T> tanh(const Quat<T> &q);
-
-template <typename T>
-Quat<T> sin(const Quat<T> &q);
-
-template <typename T>
-Quat<T> cos(const Quat<T> &q);
-
-template <typename T>
-Quat<T> tan(const Quat<T> &q);
-
-template <typename T>
-Quat<T> asinh(const Quat<T> &q);
-
-template <typename T>
-Quat<T> acosh(const Quat<T> &q);
-
-template <typename T>
-Quat<T> atanh(const Quat<T> &q);
-
-template <typename T>
-Quat<T> asin(const Quat<T> &q);
-
-template <typename T>
-Quat<T> acos(const Quat<T> &q);
-
-template <typename T>
-Quat<T> atan(const Quat<T> &q);
-
-template <typename T>
-Quat<T> power(const Quat<T> &q, const Quat<T> &p, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT);
-
-template <typename T>
-Quat<T> exp(const Quat<T> &q);
-
-template <typename T>
-Quat<T> log(const Quat<T> &q, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT);
-
-template <typename T>
-Quat<T> power(const Quat<T>& q, const T x, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT);
-
-template <typename T>
-Quat<T> crossProduct(const Quat<T> &p, const Quat<T> &q);
-
-template <typename S>
-Quat<S> sqrt(const Quat<S> &q, QuatAssumeType assumeUnit=QUAT_ASSUME_NOT_UNIT);
-
-template <typename T>
-Quat<T> operator*(const T, const Quat<T>&);
-
-template <typename T>
-Quat<T> operator*(const Quat<T>&, const T);
-
-template <typename S>
-std::ostream& operator<<(std::ostream&, const Quat<S>&);
-
-using Quatd = Quat<double>;
-using Quatf = Quat<float>;
-
-//! @} core
-}
-
-#include "opencv2/core/quaternion.inl.hpp"
-
-#endif /* OPENCV_CORE_QUATERNION_HPP */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/quaternion.inl.hpp b/projects/llm_framework/include/opencv4/opencv2/core/quaternion.inl.hpp
deleted file mode 100644
index 29a16d9f..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/quaternion.inl.hpp
+++ /dev/null
@@ -1,1063 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2020, Huawei Technologies Co., Ltd. All rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//       http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Author: Liangqian Kong <chargerKong@126.com>
-//         Longbu Wang <riskiest@gmail.com>
-
-#ifndef OPENCV_CORE_QUATERNION_INL_HPP
-#define OPENCV_CORE_QUATERNION_INL_HPP
-
-#ifndef OPENCV_CORE_QUATERNION_HPP
-#erorr This is not a standalone header. Include quaternion.hpp instead.
-#endif
-
-//@cond IGNORE
-///////////////////////////////////////////////////////////////////////////////////////
-//Implementation
-namespace cv {
-
-template <typename T>
-Quat<T>::Quat() : w(0), x(0), y(0), z(0) {}
-
-template <typename T>
-Quat<T>::Quat(const Vec<T, 4> &coeff):w(coeff[0]), x(coeff[1]), y(coeff[2]), z(coeff[3]){}
-
-template <typename T>
-Quat<T>::Quat(const T qw, const T qx, const T qy, const T qz):w(qw), x(qx), y(qy), z(qz){}
-
-template <typename T>
-Quat<T> Quat<T>::createFromAngleAxis(const T angle, const Vec<T, 3> &axis)
-{
-    T w, x, y, z;
-    T vNorm = std::sqrt(axis.dot(axis));
-    if (vNorm < CV_QUAT_EPS)
-    {
-        CV_Error(Error::StsBadArg, "this quaternion does not represent a rotation");
-    }
-    const T angle_half = angle * T(0.5);
-    w = std::cos(angle_half);
-    const T sin_v = std::sin(angle_half);
-    const T sin_norm = sin_v / vNorm;
-    x = sin_norm * axis[0];
-    y = sin_norm * axis[1];
-    z = sin_norm * axis[2];
-    return Quat<T>(w, x, y, z);
-}
-
-template <typename T>
-Quat<T> Quat<T>::createFromRotMat(InputArray _R)
-{
-    CV_CheckTypeEQ(_R.type(), cv::traits::Type<T>::value, "");
-    if (_R.rows() != 3 || _R.cols() != 3)
-    {
-        CV_Error(Error::StsBadArg, "Cannot convert matrix to quaternion: rotation matrix should be a 3x3 matrix");
-    }
-    Matx<T, 3, 3> R;
-    _R.copyTo(R);
-
-    T S, w, x, y, z;
-    T trace = R(0, 0) + R(1, 1) + R(2, 2);
-    if (trace > 0)
-    {
-        S = std::sqrt(trace + 1) * T(2);
-        x = (R(1, 2) - R(2, 1)) / S;
-        y = (R(2, 0) - R(0, 2)) / S;
-        z = (R(0, 1) - R(1, 0)) / S;
-        w = -T(0.25) * S;
-    }
-    else if (R(0, 0) > R(1, 1) && R(0, 0) > R(2, 2))
-    {
-
-        S = std::sqrt(T(1.0) + R(0, 0) - R(1, 1) - R(2, 2)) * T(2);
-        x = -T(0.25) * S;
-        y = -(R(1, 0) + R(0, 1)) / S;
-        z = -(R(0, 2) + R(2, 0)) / S;
-        w = (R(1, 2) - R(2, 1)) / S;
-    }
-    else if (R(1, 1) > R(2, 2))
-    {
-        S = std::sqrt(T(1.0) - R(0, 0) + R(1, 1) - R(2, 2)) * T(2);
-        x = (R(0, 1) + R(1, 0)) / S;
-        y = T(0.25) * S;
-        z = (R(1, 2) + R(2, 1)) / S;
-        w = (R(0, 2) - R(2, 0)) / S;
-    }
-    else
-    {
-        S = std::sqrt(T(1.0) - R(0, 0) - R(1, 1) + R(2, 2)) * T(2);
-        x = (R(0, 2) + R(2, 0)) / S;
-        y = (R(1, 2) + R(2, 1)) / S;
-        z = T(0.25) * S;
-        w = -(R(0, 1) - R(1, 0)) / S;
-    }
-    return Quat<T> (w, x, y, z);
-}
-
-template <typename T>
-Quat<T> Quat<T>::createFromRvec(InputArray _rvec)
-{
-    if (!((_rvec.cols() == 1 && _rvec.rows() == 3) || (_rvec.cols() == 3 && _rvec.rows() == 1))) {
-        CV_Error(Error::StsBadArg, "Cannot convert rotation vector to quaternion: The length of rotation vector should be 3");
-    }
-    Vec<T, 3> rvec;
-    _rvec.copyTo(rvec);
-    T psi = std::sqrt(rvec.dot(rvec));
-    if (abs(psi) < CV_QUAT_EPS) {
-        return Quat<T> (1, 0, 0, 0);
-    }
-    Vec<T, 3> axis = rvec / psi;
-    return createFromAngleAxis(psi, axis);
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::operator-() const
-{
-    return Quat<T>(-w, -x, -y, -z);
-}
-
-
-template <typename T>
-inline bool Quat<T>::operator==(const Quat<T> &q) const
-{
-    return (abs(w - q.w) < CV_QUAT_EPS && abs(x - q.x) < CV_QUAT_EPS && abs(y - q.y) < CV_QUAT_EPS && abs(z - q.z) < CV_QUAT_EPS);
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::operator+(const Quat<T> &q1) const
-{
-    return Quat<T>(w + q1.w, x + q1.x, y + q1.y, z + q1.z);
-}
-
-template <typename T>
-inline Quat<T> operator+(const T a, const Quat<T>& q)
-{
-    return Quat<T>(q.w + a, q.x, q.y, q.z);
-}
-
-template <typename T>
-inline Quat<T> operator+(const Quat<T>& q, const T a)
-{
-    return Quat<T>(q.w + a, q.x, q.y, q.z);
-}
-
-template <typename T>
-inline Quat<T> operator-(const T a, const Quat<T>& q)
-{
-    return Quat<T>(a - q.w, -q.x, -q.y, -q.z);
-}
-
-template <typename T>
-inline Quat<T> operator-(const Quat<T>& q, const T a)
-{
-    return Quat<T>(q.w - a, q.x, q.y, q.z);
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::operator-(const Quat<T> &q1) const
-{
-    return Quat<T>(w - q1.w, x - q1.x, y - q1.y, z - q1.z);
-}
-
-template <typename T>
-inline Quat<T>& Quat<T>::operator+=(const Quat<T> &q1)
-{
-    w += q1.w;
-    x += q1.x;
-    y += q1.y;
-    z += q1.z;
-    return *this;
-}
-
-template <typename T>
-inline Quat<T>& Quat<T>::operator-=(const Quat<T> &q1)
-{
-    w -= q1.w;
-    x -= q1.x;
-    y -= q1.y;
-    z -= q1.z;
-    return *this;
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::operator*(const Quat<T> &q1) const
-{
-    Vec<T, 4> q{w, x, y, z};
-    Vec<T, 4> q2{q1.w, q1.x, q1.y, q1.z};
-    return Quat<T>(q * q2);
-}
-
-
-template <typename T>
-Quat<T> operator*(const Quat<T> &q1, const T a)
-{
-    return Quat<T>(a * q1.w, a * q1.x, a * q1.y, a * q1.z);
-}
-
-template <typename T>
-Quat<T> operator*(const T a, const Quat<T> &q1)
-{
-    return Quat<T>(a * q1.w, a * q1.x, a * q1.y, a * q1.z);
-}
-
-template <typename T>
-inline Quat<T>& Quat<T>::operator*=(const Quat<T> &q1)
-{
-    T qw, qx, qy, qz;
-    qw = w * q1.w - x * q1.x - y * q1.y - z * q1.z;
-    qx = x * q1.w + w * q1.x + y * q1.z - z * q1.y;
-    qy = y * q1.w + w * q1.y + z * q1.x - x * q1.z;
-    qz = z * q1.w + w * q1.z + x * q1.y - y * q1.x;
-    w = qw;
-    x = qx;
-    y = qy;
-    z = qz;
-    return *this;
-}
-
-template <typename T>
-inline Quat<T>& Quat<T>::operator/=(const Quat<T> &q1)
-{
-    Quat<T> q(*this * q1.inv());
-    w = q.w;
-    x = q.x;
-    y = q.y;
-    z = q.z;
-    return *this;
-}
-template <typename T>
-Quat<T>& Quat<T>::operator*=(const T q1)
-{
-    w *= q1;
-    x *= q1;
-    y *= q1;
-    z *= q1;
-    return *this;
-}
-
-template <typename T>
-inline Quat<T>& Quat<T>::operator/=(const T a)
-{
-    const T a_inv = 1.0 / a;
-    w *= a_inv;
-    x *= a_inv;
-    y *= a_inv;
-    z *= a_inv;
-    return *this;
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::operator/(const T a) const
-{
-    const T a_inv = T(1.0) / a;
-    return Quat<T>(w * a_inv, x * a_inv, y * a_inv, z * a_inv);
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::operator/(const Quat<T> &q) const
-{
-    return *this * q.inv();
-}
-
-template <typename T>
-inline const T& Quat<T>::operator[](std::size_t n) const
-{
-    switch (n) {
-        case 0:
-            return w;
-        case 1:
-            return x;
-        case 2:
-            return y;
-        case 3:
-            return z;
-        default:
-            CV_Error(Error::StsOutOfRange, "subscript exceeds the index range");
-    }
-}
-
-template <typename T>
-inline T& Quat<T>::operator[](std::size_t n)
-{
-    switch (n) {
-        case 0:
-            return w;
-        case 1:
-            return x;
-        case 2:
-            return y;
-        case 3:
-            return z;
-        default:
-            CV_Error(Error::StsOutOfRange, "subscript exceeds the index range");
-    }
-}
-
-template <typename T>
-std::ostream & operator<<(std::ostream &os, const Quat<T> &q)
-{
-    os << "Quat " << Vec<T, 4>{q.w, q.x, q.y, q.z};
-    return os;
-}
-
-template <typename T>
-inline T Quat<T>::at(size_t index) const
-{
-    return (*this)[index];
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::conjugate() const
-{
-    return Quat<T>(w, -x, -y, -z);
-}
-
-template <typename T>
-inline T Quat<T>::norm() const
-{
-    return std::sqrt(dot(*this));
-}
-
-template <typename T>
-Quat<T> exp(const Quat<T> &q)
-{
-    return q.exp();
-}
-
-template <typename T>
-Quat<T> Quat<T>::exp() const
-{
-    Vec<T, 3> v{x, y, z};
-    T normV = std::sqrt(v.dot(v));
-    T k = normV < CV_QUAT_EPS ? 1 : std::sin(normV) / normV;
-    return std::exp(w) * Quat<T>(std::cos(normV), v[0] * k, v[1] * k, v[2] * k);
-}
-
-template <typename T>
-Quat<T> log(const Quat<T> &q, QuatAssumeType assumeUnit)
-{
-    return q.log(assumeUnit);
-}
-
-template <typename T>
-Quat<T> Quat<T>::log(QuatAssumeType assumeUnit) const
-{
-    Vec<T, 3> v{x, y, z};
-    T vNorm = std::sqrt(v.dot(v));
-    if (assumeUnit)
-    {
-        T k = vNorm < CV_QUAT_EPS ? 1 : std::acos(w) / vNorm;
-        return Quat<T>(0, v[0] * k, v[1] * k, v[2] * k);
-    }
-    T qNorm = norm();
-    if (qNorm < CV_QUAT_EPS)
-    {
-        CV_Error(Error::StsBadArg, "Cannot apply this quaternion to log function: undefined");
-    }
-    T k = vNorm < CV_QUAT_EPS ? 1 : std::acos(w / qNorm) / vNorm;
-    return Quat<T>(std::log(qNorm), v[0] * k, v[1] * k, v[2] *k);
-}
-
-template <typename T>
-inline Quat<T> power(const Quat<T> &q1, const T alpha, QuatAssumeType assumeUnit)
-{
-    return q1.power(alpha, assumeUnit);
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::power(const T alpha, QuatAssumeType assumeUnit) const
-{
-    if (x * x + y * y + z * z > CV_QUAT_EPS)
-    {
-        T angle = getAngle(assumeUnit);
-        Vec<T, 3> axis = getAxis(assumeUnit);
-        if (assumeUnit)
-        {
-            return createFromAngleAxis(alpha * angle, axis);
-        }
-        return std::pow(norm(), alpha) * createFromAngleAxis(alpha * angle, axis);
-    }
-    else
-    {
-        return std::pow(norm(), alpha) * Quat<T>(w, x, y, z);
-    }
-}
-
-
-template <typename T>
-inline Quat<T> sqrt(const Quat<T> &q, QuatAssumeType assumeUnit)
-{
-    return q.sqrt(assumeUnit);
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::sqrt(QuatAssumeType assumeUnit) const
-{
-    return power(0.5, assumeUnit);
-}
-
-
-template <typename T>
-inline Quat<T> power(const Quat<T> &p, const Quat<T> &q, QuatAssumeType assumeUnit)
-{
-    return p.power(q, assumeUnit);
-}
-
-
-template <typename T>
-inline Quat<T> Quat<T>::power(const Quat<T> &q, QuatAssumeType assumeUnit) const
-{
-    return cv::exp(q * log(assumeUnit));
-}
-
-template <typename T>
-inline T Quat<T>::dot(Quat<T> q1) const
-{
-    return w * q1.w + x * q1.x + y * q1.y + z * q1.z;
-}
-
-
-template <typename T>
-inline Quat<T> crossProduct(const Quat<T> &p, const Quat<T> &q)
-{
-    return p.crossProduct(q);
-}
-
-
-template <typename T>
-inline Quat<T> Quat<T>::crossProduct(const Quat<T> &q) const
-{
-    return Quat<T> (0, y * q.z - z * q.y, z * q.x - x * q.z, x * q.y - q.x * y);
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::normalize() const
-{
-    T normVal = norm();
-    if (normVal < CV_QUAT_EPS)
-    {
-        CV_Error(Error::StsBadArg, "Cannot normalize this quaternion: the norm is too small.");
-    }
-    return Quat<T>(w / normVal, x / normVal, y / normVal, z / normVal) ;
-}
-
-template <typename T>
-inline Quat<T> inv(const Quat<T> &q, QuatAssumeType assumeUnit)
-{
-    return q.inv(assumeUnit);
-}
-
-
-template <typename T>
-inline Quat<T> Quat<T>::inv(QuatAssumeType assumeUnit) const
-{
-    if (assumeUnit)
-    {
-        return conjugate();
-    }
-    T norm2 = dot(*this);
-    if (norm2 < CV_QUAT_EPS)
-    {
-        CV_Error(Error::StsBadArg, "This quaternion do not have inverse quaternion");
-    }
-    return conjugate() / norm2;
-}
-
-template <typename T>
-inline Quat<T> sinh(const Quat<T> &q)
-{
-    return q.sinh();
-}
-
-
-template <typename T>
-inline Quat<T> Quat<T>::sinh() const
-{
-    Vec<T, 3> v{x, y ,z};
-    T vNorm = std::sqrt(v.dot(v));
-    T k = vNorm < CV_QUAT_EPS ? 1 : std::cosh(w) * std::sin(vNorm) / vNorm;
-    return Quat<T>(std::sinh(w) * std::cos(vNorm), v[0] * k, v[1] * k, v[2] * k);
-}
-
-
-template <typename T>
-inline Quat<T> cosh(const Quat<T> &q)
-{
-    return q.cosh();
-}
-
-
-template <typename T>
-inline Quat<T> Quat<T>::cosh() const
-{
-    Vec<T, 3> v{x, y ,z};
-    T vNorm = std::sqrt(v.dot(v));
-    T k = vNorm < CV_QUAT_EPS ? 1 : std::sinh(w) * std::sin(vNorm) / vNorm;
-    return Quat<T>(std::cosh(w) * std::cos(vNorm), v[0] * k, v[1] * k, v[2] * k);
-}
-
-template <typename T>
-inline Quat<T> tanh(const Quat<T> &q)
-{
-    return q.tanh();
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::tanh() const
-{
-    return sinh() * cosh().inv();
-}
-
-
-template <typename T>
-inline Quat<T> sin(const Quat<T> &q)
-{
-    return q.sin();
-}
-
-
-template <typename T>
-inline Quat<T> Quat<T>::sin() const
-{
-    Vec<T, 3> v{x, y ,z};
-    T vNorm = std::sqrt(v.dot(v));
-    T k = vNorm < CV_QUAT_EPS ? 1 : std::cos(w) * std::sinh(vNorm) / vNorm;
-    return Quat<T>(std::sin(w) * std::cosh(vNorm), v[0] * k, v[1] * k, v[2] * k);
-}
-
-template <typename T>
-inline Quat<T> cos(const Quat<T> &q)
-{
-    return q.cos();
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::cos() const
-{
-    Vec<T, 3> v{x, y ,z};
-    T vNorm = std::sqrt(v.dot(v));
-    T k = vNorm < CV_QUAT_EPS ? 1 : std::sin(w) * std::sinh(vNorm) / vNorm;
-    return Quat<T>(std::cos(w) * std::cosh(vNorm), -v[0] * k, -v[1] * k, -v[2] * k);
-}
-
-template <typename T>
-inline Quat<T> tan(const Quat<T> &q)
-{
-    return q.tan();
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::tan() const
-{
-    return sin() * cos().inv();
-}
-
-template <typename T>
-inline Quat<T> asinh(const Quat<T> &q)
-{
-    return q.asinh();
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::asinh() const
-{
-    return cv::log(*this + cv::power(*this * *this + Quat<T>(1, 0, 0, 0), 0.5));
-}
-
-template <typename T>
-inline Quat<T> acosh(const Quat<T> &q)
-{
-    return q.acosh();
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::acosh() const
-{
-    return cv::log(*this + cv::power(*this * *this - Quat<T>(1,0,0,0), 0.5));
-}
-
-template <typename T>
-inline Quat<T> atanh(const Quat<T> &q)
-{
-    return q.atanh();
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::atanh() const
-{
-    Quat<T> ident(1, 0, 0, 0);
-    Quat<T> c1 = (ident + *this).log();
-    Quat<T> c2 = (ident - *this).log();
-    return 0.5 * (c1 - c2);
-}
-
-template <typename T>
-inline Quat<T> asin(const Quat<T> &q)
-{
-    return q.asin();
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::asin() const
-{
-    Quat<T> v(0, x, y, z);
-    T vNorm = v.norm();
-    T k = vNorm < CV_QUAT_EPS ? 1 : vNorm;
-    return -v / k * (*this * v / k).asinh();
-}
-
-template <typename T>
-inline Quat<T> acos(const Quat<T> &q)
-{
-    return q.acos();
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::acos() const
-{
-    Quat<T> v(0, x, y, z);
-    T vNorm = v.norm();
-    T k = vNorm < CV_QUAT_EPS ? 1 : vNorm;
-    return -v / k * acosh();
-}
-
-template <typename T>
-inline Quat<T> atan(const Quat<T> &q)
-{
-    return q.atan();
-}
-
-template <typename T>
-inline Quat<T> Quat<T>::atan() const
-{
-    Quat<T> v(0, x, y, z);
-    T vNorm = v.norm();
-    T k = vNorm < CV_QUAT_EPS ? 1 : vNorm;
-    return -v / k * (*this * v / k).atanh();
-}
-
-template <typename T>
-inline T Quat<T>::getAngle(QuatAssumeType assumeUnit) const
-{
-    if (assumeUnit)
-    {
-        return 2 * std::acos(w);
-    }
-    if (norm() < CV_QUAT_EPS)
-    {
-        CV_Error(Error::StsBadArg, "This quaternion does not represent a rotation");
-    }
-    return 2 * std::acos(w / norm());
-}
-
-template <typename T>
-inline Vec<T, 3> Quat<T>::getAxis(QuatAssumeType assumeUnit) const
-{
-    T angle = getAngle(assumeUnit);
-    const T sin_v = std::sin(angle * 0.5);
-    if (assumeUnit)
-    {
-        return Vec<T, 3>{x, y, z} / sin_v;
-    }
-    return Vec<T, 3> {x, y, z} / (norm() * sin_v);
-}
-
-template <typename T>
-Matx<T, 4, 4> Quat<T>::toRotMat4x4(QuatAssumeType assumeUnit) const
-{
-    T a = w, b = x, c = y, d = z;
-    if (!assumeUnit)
-    {
-        Quat<T> qTemp = normalize();
-        a = qTemp.w;
-        b = qTemp.x;
-        c = qTemp.y;
-        d = qTemp.z;
-    }
-    Matx<T, 4, 4> R{
-        1 - 2 * (c * c + d * d), 2 * (b * c - a * d)    , 2 * (b * d + a * c)    , 0,
-        2 * (b * c + a * d)    , 1 - 2 * (b * b + d * d), 2 * (c * d - a * b)    , 0,
-        2 * (b * d - a * c)    , 2 * (c * d + a * b)    , 1 - 2 * (b * b + c * c), 0,
-        0                      , 0                      , 0                      , 1,
-    };
-    return R;
-}
-
-template <typename T>
-Matx<T, 3, 3> Quat<T>::toRotMat3x3(QuatAssumeType assumeUnit) const
-{
-    T a = w, b = x, c = y, d = z;
-    if (!assumeUnit)
-    {
-        Quat<T> qTemp = normalize();
-        a = qTemp.w;
-        b = qTemp.x;
-        c = qTemp.y;
-        d = qTemp.z;
-    }
-    Matx<T, 3, 3> R{
-        1 - 2 * (c * c + d * d), 2 * (b * c - a * d)    , 2 * (b * d + a * c),
-        2 * (b * c + a * d)    , 1 - 2 * (b * b + d * d), 2 * (c * d - a * b),
-        2 * (b * d - a * c)    , 2 * (c * d + a * b)    , 1 - 2 * (b * b + c * c)
-    };
-    return R;
-}
-
-template <typename T>
-Vec<T, 3> Quat<T>::toRotVec(QuatAssumeType assumeUnit) const
-{
-    T angle = getAngle(assumeUnit);
-    Vec<T, 3> axis = getAxis(assumeUnit);
-    return angle * axis;
-}
-
-template <typename T>
-Vec<T, 4> Quat<T>::toVec() const
-{
-    return Vec<T, 4>{w, x, y, z};
-}
-
-template <typename T>
-Quat<T> Quat<T>::lerp(const Quat<T> &q0, const Quat<T> &q1, const T t)
-{
-    return (1 - t) * q0 + t * q1;
-}
-
-template <typename T>
-Quat<T> Quat<T>::slerp(const Quat<T> &q0, const Quat<T> &q1, const T t, QuatAssumeType assumeUnit, bool directChange)
-{
-    Quatd v0(q0);
-    Quatd v1(q1);
-    if (!assumeUnit)
-    {
-        v0 = v0.normalize();
-        v1 = v1.normalize();
-    }
-    T cosTheta = v0.dot(v1);
-    constexpr T DOT_THRESHOLD = 0.995;
-    if (cosTheta > DOT_THRESHOLD)
-    {
-        return nlerp(v0, v1, t, QUAT_ASSUME_UNIT);
-    }
-
-    if (directChange && cosTheta < 0)
-    {
-        v0 = -v0;
-        cosTheta = -cosTheta;
-    }
-    T sinTheta = std::sqrt(1 - cosTheta * cosTheta);
-    T angle = atan2(sinTheta, cosTheta);
-    return (std::sin((1 - t) * angle) / (sinTheta) * v0 + std::sin(t * angle) / (sinTheta) * v1).normalize();
-}
-
-
-template <typename T>
-inline Quat<T> Quat<T>::nlerp(const Quat<T> &q0, const Quat<T> &q1, const T t, QuatAssumeType assumeUnit)
-{
-    Quat<T> v0(q0), v1(q1);
-    if (v1.dot(v0) < 0)
-    {
-        v0 = -v0;
-    }
-    if (assumeUnit)
-    {
-        return ((1 - t) * v0 + t * v1).normalize();
-    }
-    v0 = v0.normalize();
-    v1 = v1.normalize();
-    return ((1 - t) * v0 + t * v1).normalize();
-}
-
-
-template <typename T>
-inline bool Quat<T>::isNormal(T eps) const
-{
-
-    double normVar = norm();
-    if ((normVar > 1 - eps) && (normVar < 1 + eps))
-        return true;
-    return false;
-}
-
-template <typename T>
-inline void Quat<T>::assertNormal(T eps) const
-{
-    if (!isNormal(eps))
-        CV_Error(Error::StsBadArg, "Quaternion should be normalized");
-}
-
-
-template <typename T>
-inline Quat<T> Quat<T>::squad(const Quat<T> &q0, const Quat<T> &q1,
-                            const Quat<T> &q2, const Quat<T> &q3,
-                            const T t, QuatAssumeType assumeUnit,
-                            bool directChange)
-{
-    Quat<T> v0(q0), v1(q1), v2(q2), v3(q3);
-    if (!assumeUnit)
-    {
-        v0 = v0.normalize();
-        v1 = v1.normalize();
-        v2 = v2.normalize();
-        v3 = v3.normalize();
-    }
-
-    Quat<T> c0 = slerp(v0, v3, t, assumeUnit, directChange);
-    Quat<T> c1 = slerp(v1, v2, t, assumeUnit, directChange);
-    return slerp(c0, c1, 2 * t * (1 - t), assumeUnit, directChange);
-}
-
-template <typename T>
-Quat<T> Quat<T>::interPoint(const Quat<T> &q0, const Quat<T> &q1,
-                            const Quat<T> &q2, QuatAssumeType assumeUnit)
-{
-    Quat<T> v0(q0), v1(q1), v2(q2);
-    if (!assumeUnit)
-    {
-        v0 = v0.normalize();
-        v1 = v1.normalize();
-        v2 = v2.normalize();
-    }
-    return v1 * cv::exp(-(cv::log(v1.conjugate() * v0, assumeUnit) + (cv::log(v1.conjugate() * v2, assumeUnit))) / 4);
-}
-
-template <typename T>
-Quat<T> Quat<T>::spline(const Quat<T> &q0, const Quat<T> &q1, const Quat<T> &q2, const Quat<T> &q3, const T t, QuatAssumeType assumeUnit)
-{
-    Quatd v0(q0), v1(q1), v2(q2), v3(q3);
-    if (!assumeUnit)
-    {
-        v0 = v0.normalize();
-        v1 = v1.normalize();
-        v2 = v2.normalize();
-        v3 = v3.normalize();
-    }
-    T cosTheta;
-    std::vector<Quat<T>> vec{v0, v1, v2, v3};
-    for (size_t i = 0; i < 3; ++i)
-    {
-        cosTheta = vec[i].dot(vec[i + 1]);
-        if (cosTheta < 0)
-        {
-            vec[i + 1] = -vec[i + 1];
-        }
-    }
-    Quat<T> s1 = interPoint(vec[0], vec[1], vec[2], QUAT_ASSUME_UNIT);
-    Quat<T> s2 = interPoint(vec[1], vec[2], vec[3], QUAT_ASSUME_UNIT);
-    return squad(vec[1], s1, s2, vec[2], t, assumeUnit, QUAT_ASSUME_NOT_UNIT);
-}
-
-namespace detail {
-
-template <typename T> static
-Quat<T> createFromAxisRot(int axis, const T theta)
-{
-    if (axis == 0)
-        return Quat<T>::createFromXRot(theta);
-    if (axis == 1)
-        return Quat<T>::createFromYRot(theta);
-    if (axis == 2)
-        return Quat<T>::createFromZRot(theta);
-    CV_Assert(0);
-}
-
-inline bool isIntAngleType(QuatEnum::EulerAnglesType eulerAnglesType)
-{
-    return eulerAnglesType < QuatEnum::EXT_XYZ;
-}
-
-inline bool isTaitBryan(QuatEnum::EulerAnglesType eulerAnglesType)
-{
-    return eulerAnglesType/6 == 1 || eulerAnglesType/6 == 3;
-}
-}  // namespace detail
-
-template <typename T>
-Quat<T> Quat<T>::createFromYRot(const T theta)
-{
-    return Quat<T>{std::cos(theta * 0.5f), 0, std::sin(theta * 0.5f), 0};
-}
-
-template <typename T>
-Quat<T> Quat<T>::createFromXRot(const T theta){
-    return Quat<T>{std::cos(theta * 0.5f), std::sin(theta * 0.5f), 0, 0};
-}
-
-template <typename T>
-Quat<T> Quat<T>::createFromZRot(const T theta){
-    return Quat<T>{std::cos(theta * 0.5f), 0, 0, std::sin(theta * 0.5f)};
-}
-
-
-template <typename T>
-Quat<T> Quat<T>::createFromEulerAngles(const Vec<T, 3> &angles, QuatEnum::EulerAnglesType eulerAnglesType) {
-    CV_Assert(eulerAnglesType < QuatEnum::EulerAnglesType::EULER_ANGLES_MAX_VALUE);
-    static const int rotationAxis[24][3] = {
-        {0, 1, 2}, ///< Intrinsic rotations with the Euler angles type X-Y-Z
-        {0, 2, 1}, ///< Intrinsic rotations with the Euler angles type X-Z-Y
-        {1, 0, 2}, ///< Intrinsic rotations with the Euler angles type Y-X-Z
-        {1, 2, 0}, ///< Intrinsic rotations with the Euler angles type Y-Z-X
-        {2, 0, 1}, ///< Intrinsic rotations with the Euler angles type Z-X-Y
-        {2, 1, 0}, ///< Intrinsic rotations with the Euler angles type Z-Y-X
-        {0, 1, 0}, ///< Intrinsic rotations with the Euler angles type X-Y-X
-        {0, 2, 0}, ///< Intrinsic rotations with the Euler angles type X-Z-X
-        {1, 0, 1}, ///< Intrinsic rotations with the Euler angles type Y-X-Y
-        {1, 2, 1}, ///< Intrinsic rotations with the Euler angles type Y-Z-Y
-        {2, 0, 2}, ///< Intrinsic rotations with the Euler angles type Z-X-Z
-        {2, 1, 2}, ///< Intrinsic rotations with the Euler angles type Z-Y-Z
-        {0, 1, 2}, ///< Extrinsic rotations with the Euler angles type X-Y-Z
-        {0, 2, 1}, ///< Extrinsic rotations with the Euler angles type X-Z-Y
-        {1, 0, 2}, ///< Extrinsic rotations with the Euler angles type Y-X-Z
-        {1, 2, 0}, ///< Extrinsic rotations with the Euler angles type Y-Z-X
-        {2, 0, 1}, ///< Extrinsic rotations with the Euler angles type Z-X-Y
-        {2, 1, 0}, ///< Extrinsic rotations with the Euler angles type Z-Y-X
-        {0, 1, 0}, ///< Extrinsic rotations with the Euler angles type X-Y-X
-        {0, 2, 0}, ///< Extrinsic rotations with the Euler angles type X-Z-X
-        {1, 0, 1}, ///< Extrinsic rotations with the Euler angles type Y-X-Y
-        {1, 2, 1}, ///< Extrinsic rotations with the Euler angles type Y-Z-Y
-        {2, 0, 2}, ///< Extrinsic rotations with the Euler angles type Z-X-Z
-        {2, 1, 2}  ///< Extrinsic rotations with the Euler angles type Z-Y-Z
-    };
-    Quat<T> q1 = detail::createFromAxisRot(rotationAxis[eulerAnglesType][0], angles(0));
-    Quat<T> q2 = detail::createFromAxisRot(rotationAxis[eulerAnglesType][1], angles(1));
-    Quat<T> q3 = detail::createFromAxisRot(rotationAxis[eulerAnglesType][2], angles(2));
-    if (detail::isIntAngleType(eulerAnglesType))
-    {
-        return q1 * q2 * q3;
-    }
-    else // (!detail::isIntAngleType<T>(eulerAnglesType))
-    {
-        return q3 * q2 * q1;
-    }
-}
-
-template <typename T>
-Vec<T, 3> Quat<T>::toEulerAngles(QuatEnum::EulerAnglesType eulerAnglesType){
-    CV_Assert(eulerAnglesType < QuatEnum::EulerAnglesType::EULER_ANGLES_MAX_VALUE);
-    Matx33d R = toRotMat3x3();
-    enum {
-        C_ZERO,
-        C_PI,
-        C_PI_2,
-        N_CONSTANTS,
-        R_0_0 = N_CONSTANTS, R_0_1, R_0_2,
-        R_1_0, R_1_1, R_1_2,
-        R_2_0, R_2_1, R_2_2
-    };
-    static const T constants_[N_CONSTANTS] = {
-        0,  // C_ZERO
-        (T)CV_PI,  // C_PI
-        (T)(CV_PI * 0.5)  // C_PI_2, -C_PI_2
-    };
-    static const int rotationR_[24][12] = {
-        {+R_0_2,    +R_1_0, +R_1_1, C_PI_2,     +R_2_1, +R_1_1, -C_PI_2,    -R_1_2, +R_2_2,    +R_0_2,    -R_0_1, +R_0_0},  // INT_XYZ
-        {+R_0_1,    -R_1_2, +R_2_2, -C_PI_2,    +R_2_0, +R_2_2, C_PI_2,     +R_2_1, +R_1_1,    -R_0_1,    +R_0_2, +R_0_0},  // INT_XZY
-        {+R_1_2,    -R_0_1, +R_0_0, -C_PI_2,    +R_0_1, +R_0_0, C_PI_2,     +R_0_2, +R_2_2,    -R_1_2,    +R_1_0, +R_1_1},  // INT_YXZ
-        {+R_1_0,    +R_0_2, +R_2_2, C_PI_2,     +R_0_2, +R_0_1, -C_PI_2,    -R_2_0, +R_0_0,    +R_1_0,    -R_1_2, +R_1_1},  // INT_YZX
-        {+R_2_1,    +R_1_0, +R_0_0, C_PI_2,     +R_1_0, +R_0_0, -C_PI_2,    -R_0_1, +R_1_1,    +R_2_1,    -R_2_0, +R_2_2},  // INT_ZXY
-        {+R_2_0,    -R_0_1, +R_1_1, -C_PI_2,    +R_1_2, +R_1_1, C_PI_2,     +R_1_0, +R_0_0,    -R_2_0,    +R_2_1, +R_2_2},  // INT_ZYX
-        {+R_0_0,    +R_2_1, +R_2_2, C_ZERO,     +R_1_2, +R_1_1, C_PI,       +R_1_0, -R_2_0,    +R_0_0,    +R_0_1, +R_0_2},  // INT_XYX
-        {+R_0_0,    +R_2_1, +R_2_2, C_ZERO,     -R_2_1, +R_2_2, C_PI,       +R_2_0, +R_1_0,    +R_0_0,    +R_0_2, -R_0_1},  // INT_XZX
-        {+R_1_1,    +R_0_2, +R_0_0, C_ZERO,     -R_2_0, +R_0_0, C_PI,       +R_0_1, +R_2_1,    +R_1_1,    +R_1_0, -R_1_2},  // INT_YXY
-        {+R_1_1,    +R_0_2, +R_0_0, C_ZERO,     +R_0_2, -R_0_0, C_PI,       +R_2_1, -R_0_1,    +R_1_1,    +R_1_2, +R_1_0},  // INT_YZY
-        {+R_2_2,    +R_1_0, +R_1_1, C_ZERO,     +R_1_0, +R_0_0, C_PI,       +R_0_2, -R_1_2,    +R_2_2,    +R_2_0, +R_2_1},  // INT_ZXZ
-        {+R_2_2,    +R_1_0, +R_0_0, C_ZERO,     +R_1_0, +R_0_0, C_PI,       +R_1_2, +R_0_2,    +R_2_2,    +R_2_1, -R_2_0},  // INT_ZYZ
-
-        {+R_2_0,    -C_PI_2, -R_0_1, +R_1_1,    C_PI_2,  +R_1_2, +R_1_1,    +R_2_1, +R_2_2,    -R_2_0,    +R_1_0, +R_0_0},  // EXT_XYZ
-        {+R_1_0,    C_PI_2,  +R_0_2, +R_2_2,    -C_PI_2, +R_0_2, +R_0_1,    -R_1_2, +R_1_1,    +R_1_0,    -R_2_0, +R_0_0},  // EXT_XZY
-        {+R_2_1,    C_PI_2,  +R_1_0, +R_0_0,    -C_PI_2, +R_1_0, +R_0_0,    -R_2_0, +R_2_2,    +R_2_1,    -R_0_1, +R_1_1},  // EXT_YXZ
-        {+R_0_2,    -C_PI_2, -R_1_2, +R_2_2,    C_PI_2,  +R_2_0, +R_2_2,    +R_0_2, +R_0_0,    -R_0_1,    +R_2_1, +R_1_1},  // EXT_YZX
-        {+R_1_2,    -C_PI_2, -R_0_1, +R_0_0,    C_PI_2,  +R_0_1, +R_0_0,    +R_1_0, +R_1_1,    -R_1_2,    +R_0_2, +R_2_2},  // EXT_ZXY
-        {+R_0_2,    C_PI_2,  +R_1_0, +R_1_1,    -C_PI_2, +R_2_1, +R_1_1,    -R_0_1, +R_0_0,    +R_0_2,    -R_1_2, +R_2_2},  // EXT_ZYX
-        {+R_0_0,    C_ZERO,  +R_2_1, +R_2_2,    C_PI,    +R_1_2, +R_1_1,    +R_0_1, +R_0_2,    +R_0_0,    +R_1_0, -R_2_0},  // EXT_XYX
-        {+R_0_0,    C_ZERO,  +R_2_1, +R_2_2,    C_PI,    +R_2_1, +R_2_2,    +R_0_2, -R_0_1,    +R_0_0,    +R_2_0, +R_1_0},  // EXT_XZX
-        {+R_1_1,    C_ZERO,  +R_0_2, +R_0_0,    C_PI,    -R_2_0, +R_0_0,    +R_1_0, -R_1_2,    +R_1_1,    +R_0_1, +R_2_1},  // EXT_YXY
-        {+R_1_1,    C_ZERO,  +R_0_2, +R_0_0,    C_PI,    +R_0_2, -R_0_0,    +R_1_2, +R_1_0,    +R_1_1,    +R_2_1, -R_0_1},  // EXT_YZY
-        {+R_2_2,    C_ZERO,  +R_1_0, +R_1_1,    C_PI,    +R_1_0, +R_0_0,    +R_2_0, +R_2_1,    +R_2_2,    +R_0_2, -R_1_2},  // EXT_ZXZ
-        {+R_2_2,    C_ZERO,  +R_1_0, +R_0_0,    C_PI,    +R_1_0, +R_0_0,    +R_2_1, -R_2_0,    +R_2_2,    +R_1_2, +R_0_2},  // EXT_ZYZ
-    };
-    T rotationR[12];
-    for (int i = 0; i < 12; i++)
-    {
-        int id = rotationR_[eulerAnglesType][i];
-        unsigned idx = std::abs(id);
-        T value = 0.0f;
-        if (idx < N_CONSTANTS)
-        {
-            value = constants_[idx];
-        }
-        else
-        {
-            unsigned r_idx = idx - N_CONSTANTS;
-            CV_DbgAssert(r_idx < 9);
-            value = R.val[r_idx];
-        }
-        bool isNegative = id < 0;
-        if (isNegative)
-            value = -value;
-        rotationR[i] = value;
-    }
-    Vec<T, 3> angles;
-    if (detail::isIntAngleType(eulerAnglesType))
-    {
-        if (abs(rotationR[0] - 1) < CV_QUAT_CONVERT_THRESHOLD)
-        {
-            CV_LOG_WARNING(NULL,"Gimbal Lock occurs. Euler angles are non-unique, we set the third angle to 0");
-            angles = {std::atan2(rotationR[1], rotationR[2]), rotationR[3], 0};
-            return angles;
-        }
-        else if(abs(rotationR[0] + 1) < CV_QUAT_CONVERT_THRESHOLD)
-        {
-            CV_LOG_WARNING(NULL,"Gimbal Lock occurs. Euler angles are non-unique, we set the third angle to 0");
-            angles = {std::atan2(rotationR[4], rotationR[5]), rotationR[6], 0};
-            return angles;
-        }
-    }
-    else // (!detail::isIntAngleType<T>(eulerAnglesType))
-    {
-        if (abs(rotationR[0] - 1) < CV_QUAT_CONVERT_THRESHOLD)
-        {
-            CV_LOG_WARNING(NULL,"Gimbal Lock occurs. Euler angles are non-unique, we set the first angle to 0");
-            angles = {0, rotationR[1], std::atan2(rotationR[2], rotationR[3])};
-            return angles;
-        }
-        else if (abs(rotationR[0] + 1) < CV_QUAT_CONVERT_THRESHOLD)
-        {
-            CV_LOG_WARNING(NULL,"Gimbal Lock occurs. Euler angles are non-unique, we set the first angle to 0");
-            angles = {0, rotationR[4], std::atan2(rotationR[5], rotationR[6])};
-            return angles;
-        }
-    }
-
-    angles(0) = std::atan2(rotationR[7], rotationR[8]);
-    if (detail::isTaitBryan(eulerAnglesType))
-        angles(1) = std::acos(rotationR[9]);
-    else
-        angles(1) = std::asin(rotationR[9]);
-    angles(2) = std::atan2(rotationR[10], rotationR[11]);
-    return angles;
-}
-
-}  // namepsace
-//! @endcond
-
-#endif /*OPENCV_CORE_QUATERNION_INL_HPP*/
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/saturate.hpp b/projects/llm_framework/include/opencv4/opencv2/core/saturate.hpp
deleted file mode 100644
index 8127e3d9..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/saturate.hpp
+++ /dev/null
@@ -1,179 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Copyright (C) 2014, Itseez Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_SATURATE_HPP
-#define OPENCV_CORE_SATURATE_HPP
-
-#include "opencv2/core/cvdef.h"
-#include "opencv2/core/fast_math.hpp"
-
-namespace cv
-{
-
-//! @addtogroup core_utils
-//! @{
-
-/////////////// saturate_cast (used in image & signal processing) ///////////////////
-
-/** @brief Template function for accurate conversion from one primitive type to another.
-
- The function saturate_cast resembles the standard C++ cast operations, such as static_cast\<T\>()
- and others. It perform an efficient and accurate conversion from one primitive type to another
- (see the introduction chapter). saturate in the name means that when the input value v is out of the
- range of the target type, the result is not formed just by taking low bits of the input, but instead
- the value is clipped. For example:
- @code
- uchar a = saturate_cast<uchar>(-100); // a = 0 (UCHAR_MIN)
- short b = saturate_cast<short>(33333.33333); // b = 32767 (SHRT_MAX)
- @endcode
- Such clipping is done when the target type is unsigned char , signed char , unsigned short or
- signed short . For 32-bit integers, no clipping is done.
-
- When the parameter is a floating-point value and the target type is an integer (8-, 16- or 32-bit),
- the floating-point value is first rounded to the nearest integer and then clipped if needed (when
- the target type is 8- or 16-bit).
-
- @param v Function parameter.
- @sa add, subtract, multiply, divide, Mat::convertTo
- */
-template<typename _Tp> static inline _Tp saturate_cast(uchar v)    { return _Tp(v); }
-/** @overload */
-template<typename _Tp> static inline _Tp saturate_cast(schar v)    { return _Tp(v); }
-/** @overload */
-template<typename _Tp> static inline _Tp saturate_cast(ushort v)   { return _Tp(v); }
-/** @overload */
-template<typename _Tp> static inline _Tp saturate_cast(short v)    { return _Tp(v); }
-/** @overload */
-template<typename _Tp> static inline _Tp saturate_cast(unsigned v) { return _Tp(v); }
-/** @overload */
-template<typename _Tp> static inline _Tp saturate_cast(int v)      { return _Tp(v); }
-/** @overload */
-template<typename _Tp> static inline _Tp saturate_cast(float v)    { return _Tp(v); }
-/** @overload */
-template<typename _Tp> static inline _Tp saturate_cast(double v)   { return _Tp(v); }
-/** @overload */
-template<typename _Tp> static inline _Tp saturate_cast(int64 v)    { return _Tp(v); }
-/** @overload */
-template<typename _Tp> static inline _Tp saturate_cast(uint64 v)   { return _Tp(v); }
-
-template<> inline uchar saturate_cast<uchar>(schar v)        { return (uchar)std::max((int)v, 0); }
-template<> inline uchar saturate_cast<uchar>(ushort v)       { return (uchar)std::min((unsigned)v, (unsigned)UCHAR_MAX); }
-template<> inline uchar saturate_cast<uchar>(int v)          { return (uchar)((unsigned)v <= UCHAR_MAX ? v : v > 0 ? UCHAR_MAX : 0); }
-template<> inline uchar saturate_cast<uchar>(short v)        { return saturate_cast<uchar>((int)v); }
-template<> inline uchar saturate_cast<uchar>(unsigned v)     { return (uchar)std::min(v, (unsigned)UCHAR_MAX); }
-template<> inline uchar saturate_cast<uchar>(float v)        { int iv = cvRound(v); return saturate_cast<uchar>(iv); }
-template<> inline uchar saturate_cast<uchar>(double v)       { int iv = cvRound(v); return saturate_cast<uchar>(iv); }
-template<> inline uchar saturate_cast<uchar>(int64 v)        { return (uchar)((uint64)v <= (uint64)UCHAR_MAX ? v : v > 0 ? UCHAR_MAX : 0); }
-template<> inline uchar saturate_cast<uchar>(uint64 v)       { return (uchar)std::min(v, (uint64)UCHAR_MAX); }
-
-template<> inline schar saturate_cast<schar>(uchar v)        { return (schar)std::min((int)v, SCHAR_MAX); }
-template<> inline schar saturate_cast<schar>(ushort v)       { return (schar)std::min((unsigned)v, (unsigned)SCHAR_MAX); }
-template<> inline schar saturate_cast<schar>(int v)          { return (schar)((unsigned)(v-SCHAR_MIN) <= (unsigned)UCHAR_MAX ? v : v > 0 ? SCHAR_MAX : SCHAR_MIN); }
-template<> inline schar saturate_cast<schar>(short v)        { return saturate_cast<schar>((int)v); }
-template<> inline schar saturate_cast<schar>(unsigned v)     { return (schar)std::min(v, (unsigned)SCHAR_MAX); }
-template<> inline schar saturate_cast<schar>(float v)        { int iv = cvRound(v); return saturate_cast<schar>(iv); }
-template<> inline schar saturate_cast<schar>(double v)       { int iv = cvRound(v); return saturate_cast<schar>(iv); }
-template<> inline schar saturate_cast<schar>(int64 v)        { return (schar)((uint64)((int64)v-SCHAR_MIN) <= (uint64)UCHAR_MAX ? v : v > 0 ? SCHAR_MAX : SCHAR_MIN); }
-template<> inline schar saturate_cast<schar>(uint64 v)       { return (schar)std::min(v, (uint64)SCHAR_MAX); }
-
-template<> inline ushort saturate_cast<ushort>(schar v)      { return (ushort)std::max((int)v, 0); }
-template<> inline ushort saturate_cast<ushort>(short v)      { return (ushort)std::max((int)v, 0); }
-template<> inline ushort saturate_cast<ushort>(int v)        { return (ushort)((unsigned)v <= (unsigned)USHRT_MAX ? v : v > 0 ? USHRT_MAX : 0); }
-template<> inline ushort saturate_cast<ushort>(unsigned v)   { return (ushort)std::min(v, (unsigned)USHRT_MAX); }
-template<> inline ushort saturate_cast<ushort>(float v)      { int iv = cvRound(v); return saturate_cast<ushort>(iv); }
-template<> inline ushort saturate_cast<ushort>(double v)     { int iv = cvRound(v); return saturate_cast<ushort>(iv); }
-template<> inline ushort saturate_cast<ushort>(int64 v)      { return (ushort)((uint64)v <= (uint64)USHRT_MAX ? v : v > 0 ? USHRT_MAX : 0); }
-template<> inline ushort saturate_cast<ushort>(uint64 v)     { return (ushort)std::min(v, (uint64)USHRT_MAX); }
-
-template<> inline short saturate_cast<short>(ushort v)       { return (short)std::min((int)v, SHRT_MAX); }
-template<> inline short saturate_cast<short>(int v)          { return (short)((unsigned)(v - SHRT_MIN) <= (unsigned)USHRT_MAX ? v : v > 0 ? SHRT_MAX : SHRT_MIN); }
-template<> inline short saturate_cast<short>(unsigned v)     { return (short)std::min(v, (unsigned)SHRT_MAX); }
-template<> inline short saturate_cast<short>(float v)        { int iv = cvRound(v); return saturate_cast<short>(iv); }
-template<> inline short saturate_cast<short>(double v)       { int iv = cvRound(v); return saturate_cast<short>(iv); }
-template<> inline short saturate_cast<short>(int64 v)        { return (short)((uint64)((int64)v - SHRT_MIN) <= (uint64)USHRT_MAX ? v : v > 0 ? SHRT_MAX : SHRT_MIN); }
-template<> inline short saturate_cast<short>(uint64 v)       { return (short)std::min(v, (uint64)SHRT_MAX); }
-
-template<> inline int saturate_cast<int>(unsigned v)         { return (int)std::min(v, (unsigned)INT_MAX); }
-template<> inline int saturate_cast<int>(int64 v)            { return (int)((uint64)(v - INT_MIN) <= (uint64)UINT_MAX ? v : v > 0 ? INT_MAX : INT_MIN); }
-template<> inline int saturate_cast<int>(uint64 v)           { return (int)std::min(v, (uint64)INT_MAX); }
-template<> inline int saturate_cast<int>(float v)            { return cvRound(v); }
-template<> inline int saturate_cast<int>(double v)           { return cvRound(v); }
-
-template<> inline unsigned saturate_cast<unsigned>(schar v)  { return (unsigned)std::max(v, (schar)0); }
-template<> inline unsigned saturate_cast<unsigned>(short v)  { return (unsigned)std::max(v, (short)0); }
-template<> inline unsigned saturate_cast<unsigned>(int v)    { return (unsigned)std::max(v, (int)0); }
-template<> inline unsigned saturate_cast<unsigned>(int64 v)  { return (unsigned)((uint64)v <= (uint64)UINT_MAX ? v : v > 0 ? UINT_MAX : 0); }
-template<> inline unsigned saturate_cast<unsigned>(uint64 v) { return (unsigned)std::min(v, (uint64)UINT_MAX); }
-// we intentionally do not clip negative numbers, to make -1 become 0xffffffff etc.
-template<> inline unsigned saturate_cast<unsigned>(float v)  { return static_cast<unsigned>(cvRound(v)); }
-template<> inline unsigned saturate_cast<unsigned>(double v) { return static_cast<unsigned>(cvRound(v)); }
-
-template<> inline uint64 saturate_cast<uint64>(schar v)      { return (uint64)std::max(v, (schar)0); }
-template<> inline uint64 saturate_cast<uint64>(short v)      { return (uint64)std::max(v, (short)0); }
-template<> inline uint64 saturate_cast<uint64>(int v)        { return (uint64)std::max(v, (int)0); }
-template<> inline uint64 saturate_cast<uint64>(int64 v)      { return (uint64)std::max(v, (int64)0); }
-
-template<> inline int64 saturate_cast<int64>(uint64 v)       { return (int64)std::min(v, (uint64)LLONG_MAX); }
-
-/** @overload */
-template<typename _Tp> static inline _Tp saturate_cast(float16_t v) { return saturate_cast<_Tp>((float)v); }
-
-// in theory, we could use a LUT for 8u/8s->16f conversion,
-// but with hardware support for FP32->FP16 conversion the current approach is preferable
-template<> inline float16_t saturate_cast<float16_t>(uchar v)   { return float16_t((float)v); }
-template<> inline float16_t saturate_cast<float16_t>(schar v)   { return float16_t((float)v); }
-template<> inline float16_t saturate_cast<float16_t>(ushort v)  { return float16_t((float)v); }
-template<> inline float16_t saturate_cast<float16_t>(short v)   { return float16_t((float)v); }
-template<> inline float16_t saturate_cast<float16_t>(unsigned v){ return float16_t((float)v); }
-template<> inline float16_t saturate_cast<float16_t>(int v)     { return float16_t((float)v); }
-template<> inline float16_t saturate_cast<float16_t>(uint64 v)  { return float16_t((float)v); }
-template<> inline float16_t saturate_cast<float16_t>(int64 v)   { return float16_t((float)v); }
-template<> inline float16_t saturate_cast<float16_t>(float v)   { return float16_t(v); }
-template<> inline float16_t saturate_cast<float16_t>(double v)  { return float16_t((float)v); }
-
-//! @}
-
-} // cv
-
-#endif // OPENCV_CORE_SATURATE_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/simd_intrinsics.hpp b/projects/llm_framework/include/opencv4/opencv2/core/simd_intrinsics.hpp
deleted file mode 100644
index 2658d92c..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/simd_intrinsics.hpp
+++ /dev/null
@@ -1,87 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_CORE_SIMD_INTRINSICS_HPP
-#define OPENCV_CORE_SIMD_INTRINSICS_HPP
-
-/**
-Helper header to support SIMD intrinsics (universal intrinsics) in user code.
-Intrinsics documentation: https://docs.opencv.org/4.x/df/d91/group__core__hal__intrin.html
-
-
-Checks of target CPU instruction set based on compiler definitions don't work well enough.
-More reliable solutions require utilization of configuration systems (like CMake).
-
-So, probably you need to specify your own configuration.
-
-You can do that via CMake in this way:
-    add_definitions(/DOPENCV_SIMD_CONFIG_HEADER=opencv_simd_config_custom.hpp)
-or
-    add_definitions(/DOPENCV_SIMD_CONFIG_INCLUDE_DIR=1)
-
-Additionally you may need to add include directory to your files:
-    include_directories("${CMAKE_CURRENT_LIST_DIR}/opencv_config_${MYTARGET}")
-
-These files can be pre-generated for target configurations of your application
-or generated by CMake on the fly (use CMAKE_BINARY_DIR for that).
-
-Notes:
-- H/W capability checks are still responsibility of your application
-- runtime dispatching is not covered by this helper header
-*/
-
-#ifdef __OPENCV_BUILD
-#error "Use core/hal/intrin.hpp during OpenCV build"
-#endif
-
-#ifdef OPENCV_HAL_INTRIN_HPP
-#error "core/simd_intrinsics.hpp must be included before core/hal/intrin.hpp"
-#endif
-
-#include "opencv2/core/cvdef.h"
-
-#ifdef OPENCV_SIMD_CONFIG_HEADER
-#include CVAUX_STR(OPENCV_SIMD_CONFIG_HEADER)
-#elif defined(OPENCV_SIMD_CONFIG_INCLUDE_DIR)
-#include "opencv_simd_config.hpp"  // corresponding directory should be added via -I compiler parameter
-#else  // custom config headers
-
-#if (!defined(CV_AVX_512F) || !CV_AVX_512F) && (defined(__AVX512__) || defined(__AVX512F__))
-#  include <immintrin.h>
-#  undef CV_AVX_512F
-#  define CV_AVX_512F 1
-#  ifndef OPENCV_SIMD_DONT_ASSUME_SKX  // Skylake-X with AVX-512F/CD/BW/DQ/VL
-#    undef CV_AVX512_SKX
-#    define CV_AVX512_SKX 1
-#    undef CV_AVX_512CD
-#    define CV_AVX_512CD 1
-#    undef CV_AVX_512BW
-#    define CV_AVX_512BW 1
-#    undef CV_AVX_512DQ
-#    define CV_AVX_512DQ 1
-#    undef CV_AVX_512VL
-#    define CV_AVX_512VL 1
-#  endif
-#endif // AVX512
-
-// GCC/Clang: -mavx2
-// MSVC: /arch:AVX2
-#if defined __AVX2__
-#  include <immintrin.h>
-#  undef CV_AVX2
-#  define CV_AVX2 1
-#  if defined __F16C__
-#    undef CV_FP16
-#    define CV_FP16 1
-#  endif
-#endif
-
-#endif
-
-// SSE / NEON / VSX is handled by cv_cpu_dispatch.h compatibility block
-#include "cv_cpu_dispatch.h"
-
-#include "hal/intrin.hpp"
-
-#endif // OPENCV_CORE_SIMD_INTRINSICS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/softfloat.hpp b/projects/llm_framework/include/opencv4/opencv2/core/softfloat.hpp
deleted file mode 100644
index 485e15c4..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/softfloat.hpp
+++ /dev/null
@@ -1,514 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html
-
-// This file is based on files from package issued with the following license:
-
-/*============================================================================
-
-This C header file is part of the SoftFloat IEEE Floating-Point Arithmetic
-Package, Release 3c, by John R. Hauser.
-
-Copyright 2011, 2012, 2013, 2014, 2015, 2016, 2017 The Regents of the
-University of California.  All rights reserved.
-
-Redistribution and use in source and binary forms, with or without
-modification, are permitted provided that the following conditions are met:
-
- 1. Redistributions of source code must retain the above copyright notice,
-    this list of conditions, and the following disclaimer.
-
- 2. Redistributions in binary form must reproduce the above copyright notice,
-    this list of conditions, and the following disclaimer in the documentation
-    and/or other materials provided with the distribution.
-
- 3. Neither the name of the University nor the names of its contributors may
-    be used to endorse or promote products derived from this software without
-    specific prior written permission.
-
-THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS "AS IS", AND ANY
-EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
-WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE, ARE
-DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE FOR ANY
-DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
-(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
-LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
-ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
-SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-=============================================================================*/
-
-#pragma once
-#ifndef softfloat_h
-#define softfloat_h 1
-
-#include "cvdef.h"
-
-namespace cv
-{
-
-/** @addtogroup core_utils_softfloat
-
-  [SoftFloat](http://www.jhauser.us/arithmetic/SoftFloat.html) is a software implementation
-  of floating-point calculations according to IEEE 754 standard.
-  All calculations are done in integers, that's why they are machine-independent and bit-exact.
-  This library can be useful in accuracy-critical parts like look-up tables generation, tests, etc.
-  OpenCV contains a subset of SoftFloat partially rewritten to C++.
-
-  ### Types
-
-  There are two basic types: @ref softfloat and @ref softdouble.
-  These types are binary compatible with float and double types respectively
-  and support conversions to/from them.
-  Other types from original SoftFloat library like fp16 or fp128 were thrown away
-  as well as quiet/signaling NaN support, on-the-fly rounding mode switch
-  and exception flags (though exceptions can be implemented in the future).
-
-  ### Operations
-
-  Both types support the following:
-  - Construction from signed and unsigned 32-bit and 64 integers,
-  float/double or raw binary representation
-  - Conversions between each other, to float or double and to int
-  using @ref cvRound, @ref cvTrunc, @ref cvFloor, @ref cvCeil or a bunch of
-  saturate_cast functions
-  - Add, subtract, multiply, divide, remainder, square root, FMA with absolute precision
-  - Comparison operations
-  - Explicit sign, exponent and significand manipulation through get/set methods,
- number state indicators (isInf, isNan, isSubnormal)
-  - Type-specific constants like eps, minimum/maximum value, best pi approximation, etc.
-  - min(), max(), abs(), exp(), log() and pow() functions
-
-*/
-//! @{
-
-struct softfloat;
-struct softdouble;
-
-struct CV_EXPORTS softfloat
-{
-public:
-    /** @brief Default constructor */
-    softfloat() { v = 0; }
-    /** @brief Copy constructor */
-    softfloat( const softfloat& c) { v = c.v; }
-    /** @brief Assign constructor */
-    softfloat& operator=( const softfloat& c )
-    {
-        if(&c != this) v = c.v;
-        return *this;
-    }
-    /** @brief Construct from raw
-
-    Builds new value from raw binary representation
-    */
-    static const softfloat fromRaw( const uint32_t a ) { softfloat x; x.v = a; return x; }
-
-    /** @brief Construct from integer */
-    explicit softfloat( const uint32_t );
-    explicit softfloat( const uint64_t );
-    explicit softfloat( const int32_t );
-    explicit softfloat( const int64_t );
-
-#ifdef CV_INT32_T_IS_LONG_INT
-    // for platforms with int32_t = long int
-    explicit softfloat( const int a ) { *this = softfloat(static_cast<int32_t>(a)); }
-#endif
-
-    /** @brief Construct from float */
-    explicit softfloat( const float a ) { Cv32suf s; s.f = a; v = s.u; }
-
-    /** @brief Type casts  */
-    operator softdouble() const;
-    operator float() const { Cv32suf s; s.u = v; return s.f; }
-
-    /** @brief Basic arithmetics */
-    softfloat operator + (const softfloat&) const;
-    softfloat operator - (const softfloat&) const;
-    softfloat operator * (const softfloat&) const;
-    softfloat operator / (const softfloat&) const;
-    softfloat operator - () const { softfloat x; x.v = v ^ (1U << 31); return x; }
-
-    /** @brief Remainder operator
-
-    A quote from original SoftFloat manual:
-
-    > The IEEE Standard remainder operation computes the value
-    > a - n * b, where n is the integer closest to a / b.
-    > If a / b is exactly halfway between two integers, n is the even integer
-    > closest to a / b. The IEEE Standard’s remainder operation is always exact and so requires no rounding.
-    > Depending on the relative magnitudes of the operands, the remainder functions
-    > can take considerably longer to execute than the other SoftFloat functions.
-    > This is an inherent characteristic of the remainder operation itself and is not a flaw
-    > in the SoftFloat implementation.
-    */
-    softfloat operator % (const softfloat&) const;
-
-    softfloat& operator += (const softfloat& a) { *this = *this + a; return *this; }
-    softfloat& operator -= (const softfloat& a) { *this = *this - a; return *this; }
-    softfloat& operator *= (const softfloat& a) { *this = *this * a; return *this; }
-    softfloat& operator /= (const softfloat& a) { *this = *this / a; return *this; }
-    softfloat& operator %= (const softfloat& a) { *this = *this % a; return *this; }
-
-    /** @brief Comparison operations
-
-     - Any operation with NaN produces false
-       + The only exception is when x is NaN: x != y for any y.
-     - Positive and negative zeros are equal
-    */
-    bool operator == ( const softfloat& ) const;
-    bool operator != ( const softfloat& ) const;
-    bool operator >  ( const softfloat& ) const;
-    bool operator >= ( const softfloat& ) const;
-    bool operator <  ( const softfloat& ) const;
-    bool operator <= ( const softfloat& ) const;
-
-    /** @brief NaN state indicator */
-    inline bool isNaN() const { return (v & 0x7fffffff)  > 0x7f800000; }
-    /** @brief Inf state indicator */
-    inline bool isInf() const { return (v & 0x7fffffff) == 0x7f800000; }
-    /** @brief Subnormal number indicator */
-    inline bool isSubnormal() const { return ((v >> 23) & 0xFF) == 0; }
-
-    /** @brief Get sign bit */
-    inline bool getSign() const { return (v >> 31) != 0; }
-    /** @brief Construct a copy with new sign bit */
-    inline softfloat setSign(bool sign) const { softfloat x; x.v = (v & ((1U << 31) - 1)) | ((uint32_t)sign << 31); return x; }
-    /** @brief Get 0-based exponent */
-    inline int getExp() const { return ((v >> 23) & 0xFF) - 127; }
-    /** @brief Construct a copy with new 0-based exponent */
-    inline softfloat setExp(int e) const { softfloat x; x.v = (v & 0x807fffff) | (((e + 127) & 0xFF) << 23 ); return x; }
-
-    /** @brief Get a fraction part
-
-    Returns a number 1 <= x < 2 with the same significand
-    */
-    inline softfloat getFrac() const
-    {
-        uint_fast32_t vv = (v & 0x007fffff) | (127 << 23);
-        return softfloat::fromRaw(vv);
-    }
-    /** @brief Construct a copy with provided significand
-
-    Constructs a copy of a number with significand taken from parameter
-    */
-    inline softfloat setFrac(const softfloat& s) const
-    {
-        softfloat x;
-        x.v = (v & 0xff800000) | (s.v & 0x007fffff);
-        return x;
-    }
-
-    /** @brief Zero constant */
-    static softfloat zero() { return softfloat::fromRaw( 0 ); }
-    /** @brief Positive infinity constant */
-    static softfloat  inf() { return softfloat::fromRaw( 0xFF << 23 ); }
-    /** @brief Default NaN constant */
-    static softfloat  nan() { return softfloat::fromRaw( 0x7fffffff ); }
-    /** @brief One constant */
-    static softfloat  one() { return softfloat::fromRaw(  127 << 23 ); }
-    /** @brief Smallest normalized value */
-    static softfloat  min() { return softfloat::fromRaw( 0x01 << 23 ); }
-    /** @brief Difference between 1 and next representable value */
-    static softfloat  eps() { return softfloat::fromRaw( (127 - 23) << 23 ); }
-    /** @brief Biggest finite value */
-    static softfloat  max() { return softfloat::fromRaw( (0xFF << 23) - 1 ); }
-    /** @brief Correct pi approximation */
-    static softfloat   pi() { return softfloat::fromRaw( 0x40490fdb ); }
-
-    uint32_t v;
-};
-
-/*----------------------------------------------------------------------------
-*----------------------------------------------------------------------------*/
-
-struct CV_EXPORTS softdouble
-{
-public:
-    /** @brief Default constructor */
-    softdouble() : v(0) { }
-    /** @brief Copy constructor */
-    softdouble( const softdouble& c) { v = c.v; }
-    /** @brief Assign constructor */
-    softdouble& operator=( const softdouble& c )
-    {
-        if(&c != this) v = c.v;
-        return *this;
-    }
-    /** @brief Construct from raw
-
-    Builds new value from raw binary representation
-    */
-    static softdouble fromRaw( const uint64_t a ) { softdouble x; x.v = a; return x; }
-
-    /** @brief Construct from integer */
-    explicit softdouble( const uint32_t );
-    explicit softdouble( const uint64_t );
-    explicit softdouble( const  int32_t );
-    explicit softdouble( const  int64_t );
-
-#ifdef CV_INT32_T_IS_LONG_INT
-    // for platforms with int32_t = long int
-    explicit softdouble( const int a ) { *this = softdouble(static_cast<int32_t>(a)); }
-#endif
-
-    /** @brief Construct from double */
-    explicit softdouble( const double a ) { Cv64suf s; s.f = a; v = s.u; }
-
-    /** @brief Type casts  */
-    operator softfloat() const;
-    operator double() const { Cv64suf s; s.u = v; return s.f; }
-
-    /** @brief Basic arithmetics */
-    softdouble operator + (const softdouble&) const;
-    softdouble operator - (const softdouble&) const;
-    softdouble operator * (const softdouble&) const;
-    softdouble operator / (const softdouble&) const;
-    softdouble operator - () const { softdouble x; x.v = v ^ (1ULL << 63); return x; }
-
-    /** @brief Remainder operator
-
-    A quote from original SoftFloat manual:
-
-    > The IEEE Standard remainder operation computes the value
-    > a - n * b, where n is the integer closest to a / b.
-    > If a / b is exactly halfway between two integers, n is the even integer
-    > closest to a / b. The IEEE Standard’s remainder operation is always exact and so requires no rounding.
-    > Depending on the relative magnitudes of the operands, the remainder functions
-    > can take considerably longer to execute than the other SoftFloat functions.
-    > This is an inherent characteristic of the remainder operation itself and is not a flaw
-    > in the SoftFloat implementation.
-    */
-    softdouble operator % (const softdouble&) const;
-
-    softdouble& operator += (const softdouble& a) { *this = *this + a; return *this; }
-    softdouble& operator -= (const softdouble& a) { *this = *this - a; return *this; }
-    softdouble& operator *= (const softdouble& a) { *this = *this * a; return *this; }
-    softdouble& operator /= (const softdouble& a) { *this = *this / a; return *this; }
-    softdouble& operator %= (const softdouble& a) { *this = *this % a; return *this; }
-
-    /** @brief Comparison operations
-
-     - Any operation with NaN produces false
-       + The only exception is when x is NaN: x != y for any y.
-     - Positive and negative zeros are equal
-    */
-    bool operator == ( const softdouble& ) const;
-    bool operator != ( const softdouble& ) const;
-    bool operator >  ( const softdouble& ) const;
-    bool operator >= ( const softdouble& ) const;
-    bool operator <  ( const softdouble& ) const;
-    bool operator <= ( const softdouble& ) const;
-
-    /** @brief NaN state indicator */
-    inline bool isNaN() const { return (v & 0x7fffffffffffffff)  > 0x7ff0000000000000; }
-    /** @brief Inf state indicator */
-    inline bool isInf() const { return (v & 0x7fffffffffffffff) == 0x7ff0000000000000; }
-    /** @brief Subnormal number indicator */
-    inline bool isSubnormal() const { return ((v >> 52) & 0x7FF) == 0; }
-
-    /** @brief Get sign bit */
-    inline bool getSign() const { return (v >> 63) != 0; }
-    /** @brief Construct a copy with new sign bit */
-    softdouble setSign(bool sign) const { softdouble x; x.v = (v & ((1ULL << 63) - 1)) | ((uint_fast64_t)(sign) << 63); return x; }
-    /** @brief Get 0-based exponent */
-    inline int getExp() const { return ((v >> 52) & 0x7FF) - 1023; }
-    /** @brief Construct a copy with new 0-based exponent */
-    inline softdouble setExp(int e) const
-    {
-        softdouble x;
-        x.v = (v & 0x800FFFFFFFFFFFFF) | ((uint_fast64_t)((e + 1023) & 0x7FF) << 52);
-        return x;
-    }
-
-    /** @brief Get a fraction part
-
-    Returns a number 1 <= x < 2 with the same significand
-    */
-    inline softdouble getFrac() const
-    {
-        uint_fast64_t vv = (v & 0x000FFFFFFFFFFFFF) | ((uint_fast64_t)(1023) << 52);
-        return softdouble::fromRaw(vv);
-    }
-    /** @brief Construct a copy with provided significand
-
-    Constructs a copy of a number with significand taken from parameter
-    */
-    inline softdouble setFrac(const softdouble& s) const
-    {
-        softdouble x;
-        x.v = (v & 0xFFF0000000000000) | (s.v & 0x000FFFFFFFFFFFFF);
-        return x;
-    }
-
-    /** @brief Zero constant */
-    static softdouble zero() { return softdouble::fromRaw( 0 ); }
-    /** @brief Positive infinity constant */
-    static softdouble  inf() { return softdouble::fromRaw( (uint_fast64_t)(0x7FF) << 52 ); }
-    /** @brief Default NaN constant */
-    static softdouble  nan() { return softdouble::fromRaw( CV_BIG_INT(0x7FFFFFFFFFFFFFFF) ); }
-    /** @brief One constant */
-    static softdouble  one() { return softdouble::fromRaw( (uint_fast64_t)( 1023) << 52 ); }
-    /** @brief Smallest normalized value */
-    static softdouble  min() { return softdouble::fromRaw( (uint_fast64_t)( 0x01) << 52 ); }
-    /** @brief Difference between 1 and next representable value */
-    static softdouble  eps() { return softdouble::fromRaw( (uint_fast64_t)( 1023 - 52 ) << 52 ); }
-    /** @brief Biggest finite value */
-    static softdouble  max() { return softdouble::fromRaw( ((uint_fast64_t)(0x7FF) << 52) - 1 ); }
-    /** @brief Correct pi approximation */
-    static softdouble   pi() { return softdouble::fromRaw( CV_BIG_INT(0x400921FB54442D18) ); }
-
-    uint64_t v;
-};
-
-/*----------------------------------------------------------------------------
-*----------------------------------------------------------------------------*/
-
-/** @brief Fused Multiplication and Addition
-
-Computes (a*b)+c with single rounding
-*/
-CV_EXPORTS softfloat  mulAdd( const softfloat&  a, const softfloat&  b, const softfloat & c);
-CV_EXPORTS softdouble mulAdd( const softdouble& a, const softdouble& b, const softdouble& c);
-
-/** @brief Square root */
-CV_EXPORTS softfloat  sqrt( const softfloat&  a );
-CV_EXPORTS softdouble sqrt( const softdouble& a );
-}
-
-/*----------------------------------------------------------------------------
-| Ported from OpenCV and added for usability
-*----------------------------------------------------------------------------*/
-
-/** @brief Truncates number to integer with minimum magnitude */
-CV_EXPORTS int cvTrunc(const cv::softfloat&  a);
-CV_EXPORTS int cvTrunc(const cv::softdouble& a);
-
-/** @brief Rounds a number to nearest even integer */
-CV_EXPORTS int cvRound(const cv::softfloat&  a);
-CV_EXPORTS int cvRound(const cv::softdouble& a);
-
-/** @brief Rounds a number to nearest even long long integer */
-CV_EXPORTS int64_t cvRound64(const cv::softdouble& a);
-
-/** @brief Rounds a number down to integer */
-CV_EXPORTS int cvFloor(const cv::softfloat&  a);
-CV_EXPORTS int cvFloor(const cv::softdouble& a);
-
-/** @brief Rounds number up to integer */
-CV_EXPORTS int  cvCeil(const cv::softfloat&  a);
-CV_EXPORTS int  cvCeil(const cv::softdouble& a);
-
-namespace cv
-{
-/** @brief Saturate casts */
-template<typename _Tp> static inline _Tp saturate_cast(softfloat  a) { return _Tp(a); }
-template<typename _Tp> static inline _Tp saturate_cast(softdouble a) { return _Tp(a); }
-
-template<> inline uchar saturate_cast<uchar>(softfloat  a) { return (uchar)std::max(std::min(cvRound(a), (int)UCHAR_MAX), 0); }
-template<> inline uchar saturate_cast<uchar>(softdouble a) { return (uchar)std::max(std::min(cvRound(a), (int)UCHAR_MAX), 0); }
-
-template<> inline schar saturate_cast<schar>(softfloat  a) { return (schar)std::min(std::max(cvRound(a), (int)SCHAR_MIN), (int)SCHAR_MAX); }
-template<> inline schar saturate_cast<schar>(softdouble a) { return (schar)std::min(std::max(cvRound(a), (int)SCHAR_MIN), (int)SCHAR_MAX); }
-
-template<> inline ushort saturate_cast<ushort>(softfloat  a) { return (ushort)std::max(std::min(cvRound(a), (int)USHRT_MAX), 0); }
-template<> inline ushort saturate_cast<ushort>(softdouble a) { return (ushort)std::max(std::min(cvRound(a), (int)USHRT_MAX), 0); }
-
-template<> inline short saturate_cast<short>(softfloat  a) { return (short)std::min(std::max(cvRound(a), (int)SHRT_MIN), (int)SHRT_MAX); }
-template<> inline short saturate_cast<short>(softdouble a) { return (short)std::min(std::max(cvRound(a), (int)SHRT_MIN), (int)SHRT_MAX); }
-
-template<> inline int saturate_cast<int>(softfloat  a) { return cvRound(a); }
-template<> inline int saturate_cast<int>(softdouble a) { return cvRound(a); }
-
-template<> inline int64_t saturate_cast<int64_t>(softfloat  a) { return cvRound(a); }
-template<> inline int64_t saturate_cast<int64_t>(softdouble a) { return cvRound64(a); }
-
-/** @brief Saturate cast to unsigned integer and unsigned long long integer
-We intentionally do not clip negative numbers, to make -1 become 0xffffffff etc.
-*/
-template<> inline unsigned saturate_cast<unsigned>(softfloat  a) { return cvRound(a); }
-template<> inline unsigned saturate_cast<unsigned>(softdouble a) { return cvRound(a); }
-
-template<> inline uint64_t saturate_cast<uint64_t>(softfloat  a) { return cvRound(a); }
-template<> inline uint64_t saturate_cast<uint64_t>(softdouble a) { return cvRound64(a); }
-
-/** @brief Min and Max functions */
-inline softfloat  min(const softfloat&  a, const softfloat&  b) { return (a > b) ? b : a; }
-inline softdouble min(const softdouble& a, const softdouble& b) { return (a > b) ? b : a; }
-
-inline softfloat  max(const softfloat&  a, const softfloat&  b) { return (a > b) ? a : b; }
-inline softdouble max(const softdouble& a, const softdouble& b) { return (a > b) ? a : b; }
-
-/** @brief Absolute value */
-inline softfloat  abs( softfloat  a) { softfloat  x; x.v = a.v & ((1U   << 31) - 1); return x; }
-inline softdouble abs( softdouble a) { softdouble x; x.v = a.v & ((1ULL << 63) - 1); return x; }
-
-/** @brief Exponent
-
-Special cases:
-- exp(NaN) is NaN
-- exp(-Inf) == 0
-- exp(+Inf) == +Inf
-*/
-CV_EXPORTS softfloat  exp( const softfloat&  a);
-CV_EXPORTS softdouble exp( const softdouble& a);
-
-/** @brief Natural logarithm
-
-Special cases:
-- log(NaN), log(x < 0) are NaN
-- log(0) == -Inf
-*/
-CV_EXPORTS softfloat  log( const softfloat&  a );
-CV_EXPORTS softdouble log( const softdouble& a );
-
-/** @brief Raising to the power
-
-Special cases:
-- x**NaN is NaN for any x
-- ( |x| == 1 )**Inf is NaN
-- ( |x|  > 1 )**+Inf or ( |x| < 1 )**-Inf is +Inf
-- ( |x|  > 1 )**-Inf or ( |x| < 1 )**+Inf is 0
-- x ** 0 == 1 for any x
-- x ** 1 == 1 for any x
-- NaN ** y is NaN for any other y
-- Inf**(y < 0) == 0
-- Inf ** y is +Inf for any other y
-- (x < 0)**y is NaN for any other y if x can't be correctly rounded to integer
-- 0 ** 0 == 1
-- 0 ** (y < 0) is +Inf
-- 0 ** (y > 0) is 0
-*/
-CV_EXPORTS softfloat  pow( const softfloat&  a, const softfloat&  b);
-CV_EXPORTS softdouble pow( const softdouble& a, const softdouble& b);
-
-/** @brief Cube root
-
-Special cases:
-- cbrt(NaN) is NaN
-- cbrt(+/-Inf) is +/-Inf
-*/
-CV_EXPORTS softfloat cbrt( const softfloat& a );
-
-/** @brief Sine
-
-Special cases:
-- sin(Inf) or sin(NaN) is NaN
-- sin(x) == x when sin(x) is close to zero
-*/
-CV_EXPORTS softdouble sin( const softdouble& a );
-
-/** @brief Cosine
- *
-Special cases:
-- cos(Inf) or cos(NaN) is NaN
-- cos(x) == +/- 1 when cos(x) is close to +/- 1
-*/
-CV_EXPORTS softdouble cos( const softdouble& a );
-
-//! @} core_utils_softfloat
-
-} // cv::
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/sse_utils.hpp b/projects/llm_framework/include/opencv4/opencv2/core/sse_utils.hpp
deleted file mode 100644
index 0906583e..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/sse_utils.hpp
+++ /dev/null
@@ -1,652 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2015, Itseez Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_SSE_UTILS_HPP
-#define OPENCV_CORE_SSE_UTILS_HPP
-
-#ifndef __cplusplus
-#  error sse_utils.hpp header must be compiled as C++
-#endif
-
-#include "opencv2/core/cvdef.h"
-
-//! @addtogroup core_utils_sse
-//! @{
-
-#if CV_SSE2
-
-inline void _mm_deinterleave_epi8(__m128i & v_r0, __m128i & v_r1, __m128i & v_g0, __m128i & v_g1)
-{
-    __m128i layer1_chunk0 = _mm_unpacklo_epi8(v_r0, v_g0);
-    __m128i layer1_chunk1 = _mm_unpackhi_epi8(v_r0, v_g0);
-    __m128i layer1_chunk2 = _mm_unpacklo_epi8(v_r1, v_g1);
-    __m128i layer1_chunk3 = _mm_unpackhi_epi8(v_r1, v_g1);
-
-    __m128i layer2_chunk0 = _mm_unpacklo_epi8(layer1_chunk0, layer1_chunk2);
-    __m128i layer2_chunk1 = _mm_unpackhi_epi8(layer1_chunk0, layer1_chunk2);
-    __m128i layer2_chunk2 = _mm_unpacklo_epi8(layer1_chunk1, layer1_chunk3);
-    __m128i layer2_chunk3 = _mm_unpackhi_epi8(layer1_chunk1, layer1_chunk3);
-
-    __m128i layer3_chunk0 = _mm_unpacklo_epi8(layer2_chunk0, layer2_chunk2);
-    __m128i layer3_chunk1 = _mm_unpackhi_epi8(layer2_chunk0, layer2_chunk2);
-    __m128i layer3_chunk2 = _mm_unpacklo_epi8(layer2_chunk1, layer2_chunk3);
-    __m128i layer3_chunk3 = _mm_unpackhi_epi8(layer2_chunk1, layer2_chunk3);
-
-    __m128i layer4_chunk0 = _mm_unpacklo_epi8(layer3_chunk0, layer3_chunk2);
-    __m128i layer4_chunk1 = _mm_unpackhi_epi8(layer3_chunk0, layer3_chunk2);
-    __m128i layer4_chunk2 = _mm_unpacklo_epi8(layer3_chunk1, layer3_chunk3);
-    __m128i layer4_chunk3 = _mm_unpackhi_epi8(layer3_chunk1, layer3_chunk3);
-
-    v_r0 = _mm_unpacklo_epi8(layer4_chunk0, layer4_chunk2);
-    v_r1 = _mm_unpackhi_epi8(layer4_chunk0, layer4_chunk2);
-    v_g0 = _mm_unpacklo_epi8(layer4_chunk1, layer4_chunk3);
-    v_g1 = _mm_unpackhi_epi8(layer4_chunk1, layer4_chunk3);
-}
-
-inline void _mm_deinterleave_epi8(__m128i & v_r0, __m128i & v_r1, __m128i & v_g0,
-                                  __m128i & v_g1, __m128i & v_b0, __m128i & v_b1)
-{
-    __m128i layer1_chunk0 = _mm_unpacklo_epi8(v_r0, v_g1);
-    __m128i layer1_chunk1 = _mm_unpackhi_epi8(v_r0, v_g1);
-    __m128i layer1_chunk2 = _mm_unpacklo_epi8(v_r1, v_b0);
-    __m128i layer1_chunk3 = _mm_unpackhi_epi8(v_r1, v_b0);
-    __m128i layer1_chunk4 = _mm_unpacklo_epi8(v_g0, v_b1);
-    __m128i layer1_chunk5 = _mm_unpackhi_epi8(v_g0, v_b1);
-
-    __m128i layer2_chunk0 = _mm_unpacklo_epi8(layer1_chunk0, layer1_chunk3);
-    __m128i layer2_chunk1 = _mm_unpackhi_epi8(layer1_chunk0, layer1_chunk3);
-    __m128i layer2_chunk2 = _mm_unpacklo_epi8(layer1_chunk1, layer1_chunk4);
-    __m128i layer2_chunk3 = _mm_unpackhi_epi8(layer1_chunk1, layer1_chunk4);
-    __m128i layer2_chunk4 = _mm_unpacklo_epi8(layer1_chunk2, layer1_chunk5);
-    __m128i layer2_chunk5 = _mm_unpackhi_epi8(layer1_chunk2, layer1_chunk5);
-
-    __m128i layer3_chunk0 = _mm_unpacklo_epi8(layer2_chunk0, layer2_chunk3);
-    __m128i layer3_chunk1 = _mm_unpackhi_epi8(layer2_chunk0, layer2_chunk3);
-    __m128i layer3_chunk2 = _mm_unpacklo_epi8(layer2_chunk1, layer2_chunk4);
-    __m128i layer3_chunk3 = _mm_unpackhi_epi8(layer2_chunk1, layer2_chunk4);
-    __m128i layer3_chunk4 = _mm_unpacklo_epi8(layer2_chunk2, layer2_chunk5);
-    __m128i layer3_chunk5 = _mm_unpackhi_epi8(layer2_chunk2, layer2_chunk5);
-
-    __m128i layer4_chunk0 = _mm_unpacklo_epi8(layer3_chunk0, layer3_chunk3);
-    __m128i layer4_chunk1 = _mm_unpackhi_epi8(layer3_chunk0, layer3_chunk3);
-    __m128i layer4_chunk2 = _mm_unpacklo_epi8(layer3_chunk1, layer3_chunk4);
-    __m128i layer4_chunk3 = _mm_unpackhi_epi8(layer3_chunk1, layer3_chunk4);
-    __m128i layer4_chunk4 = _mm_unpacklo_epi8(layer3_chunk2, layer3_chunk5);
-    __m128i layer4_chunk5 = _mm_unpackhi_epi8(layer3_chunk2, layer3_chunk5);
-
-    v_r0 = _mm_unpacklo_epi8(layer4_chunk0, layer4_chunk3);
-    v_r1 = _mm_unpackhi_epi8(layer4_chunk0, layer4_chunk3);
-    v_g0 = _mm_unpacklo_epi8(layer4_chunk1, layer4_chunk4);
-    v_g1 = _mm_unpackhi_epi8(layer4_chunk1, layer4_chunk4);
-    v_b0 = _mm_unpacklo_epi8(layer4_chunk2, layer4_chunk5);
-    v_b1 = _mm_unpackhi_epi8(layer4_chunk2, layer4_chunk5);
-}
-
-inline void _mm_deinterleave_epi8(__m128i & v_r0, __m128i & v_r1, __m128i & v_g0, __m128i & v_g1,
-                                  __m128i & v_b0, __m128i & v_b1, __m128i & v_a0, __m128i & v_a1)
-{
-    __m128i layer1_chunk0 = _mm_unpacklo_epi8(v_r0, v_b0);
-    __m128i layer1_chunk1 = _mm_unpackhi_epi8(v_r0, v_b0);
-    __m128i layer1_chunk2 = _mm_unpacklo_epi8(v_r1, v_b1);
-    __m128i layer1_chunk3 = _mm_unpackhi_epi8(v_r1, v_b1);
-    __m128i layer1_chunk4 = _mm_unpacklo_epi8(v_g0, v_a0);
-    __m128i layer1_chunk5 = _mm_unpackhi_epi8(v_g0, v_a0);
-    __m128i layer1_chunk6 = _mm_unpacklo_epi8(v_g1, v_a1);
-    __m128i layer1_chunk7 = _mm_unpackhi_epi8(v_g1, v_a1);
-
-    __m128i layer2_chunk0 = _mm_unpacklo_epi8(layer1_chunk0, layer1_chunk4);
-    __m128i layer2_chunk1 = _mm_unpackhi_epi8(layer1_chunk0, layer1_chunk4);
-    __m128i layer2_chunk2 = _mm_unpacklo_epi8(layer1_chunk1, layer1_chunk5);
-    __m128i layer2_chunk3 = _mm_unpackhi_epi8(layer1_chunk1, layer1_chunk5);
-    __m128i layer2_chunk4 = _mm_unpacklo_epi8(layer1_chunk2, layer1_chunk6);
-    __m128i layer2_chunk5 = _mm_unpackhi_epi8(layer1_chunk2, layer1_chunk6);
-    __m128i layer2_chunk6 = _mm_unpacklo_epi8(layer1_chunk3, layer1_chunk7);
-    __m128i layer2_chunk7 = _mm_unpackhi_epi8(layer1_chunk3, layer1_chunk7);
-
-    __m128i layer3_chunk0 = _mm_unpacklo_epi8(layer2_chunk0, layer2_chunk4);
-    __m128i layer3_chunk1 = _mm_unpackhi_epi8(layer2_chunk0, layer2_chunk4);
-    __m128i layer3_chunk2 = _mm_unpacklo_epi8(layer2_chunk1, layer2_chunk5);
-    __m128i layer3_chunk3 = _mm_unpackhi_epi8(layer2_chunk1, layer2_chunk5);
-    __m128i layer3_chunk4 = _mm_unpacklo_epi8(layer2_chunk2, layer2_chunk6);
-    __m128i layer3_chunk5 = _mm_unpackhi_epi8(layer2_chunk2, layer2_chunk6);
-    __m128i layer3_chunk6 = _mm_unpacklo_epi8(layer2_chunk3, layer2_chunk7);
-    __m128i layer3_chunk7 = _mm_unpackhi_epi8(layer2_chunk3, layer2_chunk7);
-
-    __m128i layer4_chunk0 = _mm_unpacklo_epi8(layer3_chunk0, layer3_chunk4);
-    __m128i layer4_chunk1 = _mm_unpackhi_epi8(layer3_chunk0, layer3_chunk4);
-    __m128i layer4_chunk2 = _mm_unpacklo_epi8(layer3_chunk1, layer3_chunk5);
-    __m128i layer4_chunk3 = _mm_unpackhi_epi8(layer3_chunk1, layer3_chunk5);
-    __m128i layer4_chunk4 = _mm_unpacklo_epi8(layer3_chunk2, layer3_chunk6);
-    __m128i layer4_chunk5 = _mm_unpackhi_epi8(layer3_chunk2, layer3_chunk6);
-    __m128i layer4_chunk6 = _mm_unpacklo_epi8(layer3_chunk3, layer3_chunk7);
-    __m128i layer4_chunk7 = _mm_unpackhi_epi8(layer3_chunk3, layer3_chunk7);
-
-    v_r0 = _mm_unpacklo_epi8(layer4_chunk0, layer4_chunk4);
-    v_r1 = _mm_unpackhi_epi8(layer4_chunk0, layer4_chunk4);
-    v_g0 = _mm_unpacklo_epi8(layer4_chunk1, layer4_chunk5);
-    v_g1 = _mm_unpackhi_epi8(layer4_chunk1, layer4_chunk5);
-    v_b0 = _mm_unpacklo_epi8(layer4_chunk2, layer4_chunk6);
-    v_b1 = _mm_unpackhi_epi8(layer4_chunk2, layer4_chunk6);
-    v_a0 = _mm_unpacklo_epi8(layer4_chunk3, layer4_chunk7);
-    v_a1 = _mm_unpackhi_epi8(layer4_chunk3, layer4_chunk7);
-}
-
-inline void _mm_interleave_epi8(__m128i & v_r0, __m128i & v_r1, __m128i & v_g0, __m128i & v_g1)
-{
-    __m128i v_mask = _mm_set1_epi16(0x00ff);
-
-    __m128i layer4_chunk0 = _mm_packus_epi16(_mm_and_si128(v_r0, v_mask), _mm_and_si128(v_r1, v_mask));
-    __m128i layer4_chunk2 = _mm_packus_epi16(_mm_srli_epi16(v_r0, 8), _mm_srli_epi16(v_r1, 8));
-    __m128i layer4_chunk1 = _mm_packus_epi16(_mm_and_si128(v_g0, v_mask), _mm_and_si128(v_g1, v_mask));
-    __m128i layer4_chunk3 = _mm_packus_epi16(_mm_srli_epi16(v_g0, 8), _mm_srli_epi16(v_g1, 8));
-
-    __m128i layer3_chunk0 = _mm_packus_epi16(_mm_and_si128(layer4_chunk0, v_mask), _mm_and_si128(layer4_chunk1, v_mask));
-    __m128i layer3_chunk2 = _mm_packus_epi16(_mm_srli_epi16(layer4_chunk0, 8), _mm_srli_epi16(layer4_chunk1, 8));
-    __m128i layer3_chunk1 = _mm_packus_epi16(_mm_and_si128(layer4_chunk2, v_mask), _mm_and_si128(layer4_chunk3, v_mask));
-    __m128i layer3_chunk3 = _mm_packus_epi16(_mm_srli_epi16(layer4_chunk2, 8), _mm_srli_epi16(layer4_chunk3, 8));
-
-    __m128i layer2_chunk0 = _mm_packus_epi16(_mm_and_si128(layer3_chunk0, v_mask), _mm_and_si128(layer3_chunk1, v_mask));
-    __m128i layer2_chunk2 = _mm_packus_epi16(_mm_srli_epi16(layer3_chunk0, 8), _mm_srli_epi16(layer3_chunk1, 8));
-    __m128i layer2_chunk1 = _mm_packus_epi16(_mm_and_si128(layer3_chunk2, v_mask), _mm_and_si128(layer3_chunk3, v_mask));
-    __m128i layer2_chunk3 = _mm_packus_epi16(_mm_srli_epi16(layer3_chunk2, 8), _mm_srli_epi16(layer3_chunk3, 8));
-
-    __m128i layer1_chunk0 = _mm_packus_epi16(_mm_and_si128(layer2_chunk0, v_mask), _mm_and_si128(layer2_chunk1, v_mask));
-    __m128i layer1_chunk2 = _mm_packus_epi16(_mm_srli_epi16(layer2_chunk0, 8), _mm_srli_epi16(layer2_chunk1, 8));
-    __m128i layer1_chunk1 = _mm_packus_epi16(_mm_and_si128(layer2_chunk2, v_mask), _mm_and_si128(layer2_chunk3, v_mask));
-    __m128i layer1_chunk3 = _mm_packus_epi16(_mm_srli_epi16(layer2_chunk2, 8), _mm_srli_epi16(layer2_chunk3, 8));
-
-    v_r0 = _mm_packus_epi16(_mm_and_si128(layer1_chunk0, v_mask), _mm_and_si128(layer1_chunk1, v_mask));
-    v_g0 = _mm_packus_epi16(_mm_srli_epi16(layer1_chunk0, 8), _mm_srli_epi16(layer1_chunk1, 8));
-    v_r1 = _mm_packus_epi16(_mm_and_si128(layer1_chunk2, v_mask), _mm_and_si128(layer1_chunk3, v_mask));
-    v_g1 = _mm_packus_epi16(_mm_srli_epi16(layer1_chunk2, 8), _mm_srli_epi16(layer1_chunk3, 8));
-}
-
-inline void _mm_interleave_epi8(__m128i & v_r0, __m128i & v_r1, __m128i & v_g0,
-                                __m128i & v_g1, __m128i & v_b0, __m128i & v_b1)
-{
-    __m128i v_mask = _mm_set1_epi16(0x00ff);
-
-    __m128i layer4_chunk0 = _mm_packus_epi16(_mm_and_si128(v_r0, v_mask), _mm_and_si128(v_r1, v_mask));
-    __m128i layer4_chunk3 = _mm_packus_epi16(_mm_srli_epi16(v_r0, 8), _mm_srli_epi16(v_r1, 8));
-    __m128i layer4_chunk1 = _mm_packus_epi16(_mm_and_si128(v_g0, v_mask), _mm_and_si128(v_g1, v_mask));
-    __m128i layer4_chunk4 = _mm_packus_epi16(_mm_srli_epi16(v_g0, 8), _mm_srli_epi16(v_g1, 8));
-    __m128i layer4_chunk2 = _mm_packus_epi16(_mm_and_si128(v_b0, v_mask), _mm_and_si128(v_b1, v_mask));
-    __m128i layer4_chunk5 = _mm_packus_epi16(_mm_srli_epi16(v_b0, 8), _mm_srli_epi16(v_b1, 8));
-
-    __m128i layer3_chunk0 = _mm_packus_epi16(_mm_and_si128(layer4_chunk0, v_mask), _mm_and_si128(layer4_chunk1, v_mask));
-    __m128i layer3_chunk3 = _mm_packus_epi16(_mm_srli_epi16(layer4_chunk0, 8), _mm_srli_epi16(layer4_chunk1, 8));
-    __m128i layer3_chunk1 = _mm_packus_epi16(_mm_and_si128(layer4_chunk2, v_mask), _mm_and_si128(layer4_chunk3, v_mask));
-    __m128i layer3_chunk4 = _mm_packus_epi16(_mm_srli_epi16(layer4_chunk2, 8), _mm_srli_epi16(layer4_chunk3, 8));
-    __m128i layer3_chunk2 = _mm_packus_epi16(_mm_and_si128(layer4_chunk4, v_mask), _mm_and_si128(layer4_chunk5, v_mask));
-    __m128i layer3_chunk5 = _mm_packus_epi16(_mm_srli_epi16(layer4_chunk4, 8), _mm_srli_epi16(layer4_chunk5, 8));
-
-    __m128i layer2_chunk0 = _mm_packus_epi16(_mm_and_si128(layer3_chunk0, v_mask), _mm_and_si128(layer3_chunk1, v_mask));
-    __m128i layer2_chunk3 = _mm_packus_epi16(_mm_srli_epi16(layer3_chunk0, 8), _mm_srli_epi16(layer3_chunk1, 8));
-    __m128i layer2_chunk1 = _mm_packus_epi16(_mm_and_si128(layer3_chunk2, v_mask), _mm_and_si128(layer3_chunk3, v_mask));
-    __m128i layer2_chunk4 = _mm_packus_epi16(_mm_srli_epi16(layer3_chunk2, 8), _mm_srli_epi16(layer3_chunk3, 8));
-    __m128i layer2_chunk2 = _mm_packus_epi16(_mm_and_si128(layer3_chunk4, v_mask), _mm_and_si128(layer3_chunk5, v_mask));
-    __m128i layer2_chunk5 = _mm_packus_epi16(_mm_srli_epi16(layer3_chunk4, 8), _mm_srli_epi16(layer3_chunk5, 8));
-
-    __m128i layer1_chunk0 = _mm_packus_epi16(_mm_and_si128(layer2_chunk0, v_mask), _mm_and_si128(layer2_chunk1, v_mask));
-    __m128i layer1_chunk3 = _mm_packus_epi16(_mm_srli_epi16(layer2_chunk0, 8), _mm_srli_epi16(layer2_chunk1, 8));
-    __m128i layer1_chunk1 = _mm_packus_epi16(_mm_and_si128(layer2_chunk2, v_mask), _mm_and_si128(layer2_chunk3, v_mask));
-    __m128i layer1_chunk4 = _mm_packus_epi16(_mm_srli_epi16(layer2_chunk2, 8), _mm_srli_epi16(layer2_chunk3, 8));
-    __m128i layer1_chunk2 = _mm_packus_epi16(_mm_and_si128(layer2_chunk4, v_mask), _mm_and_si128(layer2_chunk5, v_mask));
-    __m128i layer1_chunk5 = _mm_packus_epi16(_mm_srli_epi16(layer2_chunk4, 8), _mm_srli_epi16(layer2_chunk5, 8));
-
-    v_r0 = _mm_packus_epi16(_mm_and_si128(layer1_chunk0, v_mask), _mm_and_si128(layer1_chunk1, v_mask));
-    v_g1 = _mm_packus_epi16(_mm_srli_epi16(layer1_chunk0, 8), _mm_srli_epi16(layer1_chunk1, 8));
-    v_r1 = _mm_packus_epi16(_mm_and_si128(layer1_chunk2, v_mask), _mm_and_si128(layer1_chunk3, v_mask));
-    v_b0 = _mm_packus_epi16(_mm_srli_epi16(layer1_chunk2, 8), _mm_srli_epi16(layer1_chunk3, 8));
-    v_g0 = _mm_packus_epi16(_mm_and_si128(layer1_chunk4, v_mask), _mm_and_si128(layer1_chunk5, v_mask));
-    v_b1 = _mm_packus_epi16(_mm_srli_epi16(layer1_chunk4, 8), _mm_srli_epi16(layer1_chunk5, 8));
-}
-
-inline void _mm_interleave_epi8(__m128i & v_r0, __m128i & v_r1, __m128i & v_g0, __m128i & v_g1,
-                                __m128i & v_b0, __m128i & v_b1, __m128i & v_a0, __m128i & v_a1)
-{
-    __m128i v_mask = _mm_set1_epi16(0x00ff);
-
-    __m128i layer4_chunk0 = _mm_packus_epi16(_mm_and_si128(v_r0, v_mask), _mm_and_si128(v_r1, v_mask));
-    __m128i layer4_chunk4 = _mm_packus_epi16(_mm_srli_epi16(v_r0, 8), _mm_srli_epi16(v_r1, 8));
-    __m128i layer4_chunk1 = _mm_packus_epi16(_mm_and_si128(v_g0, v_mask), _mm_and_si128(v_g1, v_mask));
-    __m128i layer4_chunk5 = _mm_packus_epi16(_mm_srli_epi16(v_g0, 8), _mm_srli_epi16(v_g1, 8));
-    __m128i layer4_chunk2 = _mm_packus_epi16(_mm_and_si128(v_b0, v_mask), _mm_and_si128(v_b1, v_mask));
-    __m128i layer4_chunk6 = _mm_packus_epi16(_mm_srli_epi16(v_b0, 8), _mm_srli_epi16(v_b1, 8));
-    __m128i layer4_chunk3 = _mm_packus_epi16(_mm_and_si128(v_a0, v_mask), _mm_and_si128(v_a1, v_mask));
-    __m128i layer4_chunk7 = _mm_packus_epi16(_mm_srli_epi16(v_a0, 8), _mm_srli_epi16(v_a1, 8));
-
-    __m128i layer3_chunk0 = _mm_packus_epi16(_mm_and_si128(layer4_chunk0, v_mask), _mm_and_si128(layer4_chunk1, v_mask));
-    __m128i layer3_chunk4 = _mm_packus_epi16(_mm_srli_epi16(layer4_chunk0, 8), _mm_srli_epi16(layer4_chunk1, 8));
-    __m128i layer3_chunk1 = _mm_packus_epi16(_mm_and_si128(layer4_chunk2, v_mask), _mm_and_si128(layer4_chunk3, v_mask));
-    __m128i layer3_chunk5 = _mm_packus_epi16(_mm_srli_epi16(layer4_chunk2, 8), _mm_srli_epi16(layer4_chunk3, 8));
-    __m128i layer3_chunk2 = _mm_packus_epi16(_mm_and_si128(layer4_chunk4, v_mask), _mm_and_si128(layer4_chunk5, v_mask));
-    __m128i layer3_chunk6 = _mm_packus_epi16(_mm_srli_epi16(layer4_chunk4, 8), _mm_srli_epi16(layer4_chunk5, 8));
-    __m128i layer3_chunk3 = _mm_packus_epi16(_mm_and_si128(layer4_chunk6, v_mask), _mm_and_si128(layer4_chunk7, v_mask));
-    __m128i layer3_chunk7 = _mm_packus_epi16(_mm_srli_epi16(layer4_chunk6, 8), _mm_srli_epi16(layer4_chunk7, 8));
-
-    __m128i layer2_chunk0 = _mm_packus_epi16(_mm_and_si128(layer3_chunk0, v_mask), _mm_and_si128(layer3_chunk1, v_mask));
-    __m128i layer2_chunk4 = _mm_packus_epi16(_mm_srli_epi16(layer3_chunk0, 8), _mm_srli_epi16(layer3_chunk1, 8));
-    __m128i layer2_chunk1 = _mm_packus_epi16(_mm_and_si128(layer3_chunk2, v_mask), _mm_and_si128(layer3_chunk3, v_mask));
-    __m128i layer2_chunk5 = _mm_packus_epi16(_mm_srli_epi16(layer3_chunk2, 8), _mm_srli_epi16(layer3_chunk3, 8));
-    __m128i layer2_chunk2 = _mm_packus_epi16(_mm_and_si128(layer3_chunk4, v_mask), _mm_and_si128(layer3_chunk5, v_mask));
-    __m128i layer2_chunk6 = _mm_packus_epi16(_mm_srli_epi16(layer3_chunk4, 8), _mm_srli_epi16(layer3_chunk5, 8));
-    __m128i layer2_chunk3 = _mm_packus_epi16(_mm_and_si128(layer3_chunk6, v_mask), _mm_and_si128(layer3_chunk7, v_mask));
-    __m128i layer2_chunk7 = _mm_packus_epi16(_mm_srli_epi16(layer3_chunk6, 8), _mm_srli_epi16(layer3_chunk7, 8));
-
-    __m128i layer1_chunk0 = _mm_packus_epi16(_mm_and_si128(layer2_chunk0, v_mask), _mm_and_si128(layer2_chunk1, v_mask));
-    __m128i layer1_chunk4 = _mm_packus_epi16(_mm_srli_epi16(layer2_chunk0, 8), _mm_srli_epi16(layer2_chunk1, 8));
-    __m128i layer1_chunk1 = _mm_packus_epi16(_mm_and_si128(layer2_chunk2, v_mask), _mm_and_si128(layer2_chunk3, v_mask));
-    __m128i layer1_chunk5 = _mm_packus_epi16(_mm_srli_epi16(layer2_chunk2, 8), _mm_srli_epi16(layer2_chunk3, 8));
-    __m128i layer1_chunk2 = _mm_packus_epi16(_mm_and_si128(layer2_chunk4, v_mask), _mm_and_si128(layer2_chunk5, v_mask));
-    __m128i layer1_chunk6 = _mm_packus_epi16(_mm_srli_epi16(layer2_chunk4, 8), _mm_srli_epi16(layer2_chunk5, 8));
-    __m128i layer1_chunk3 = _mm_packus_epi16(_mm_and_si128(layer2_chunk6, v_mask), _mm_and_si128(layer2_chunk7, v_mask));
-    __m128i layer1_chunk7 = _mm_packus_epi16(_mm_srli_epi16(layer2_chunk6, 8), _mm_srli_epi16(layer2_chunk7, 8));
-
-    v_r0 = _mm_packus_epi16(_mm_and_si128(layer1_chunk0, v_mask), _mm_and_si128(layer1_chunk1, v_mask));
-    v_b0 = _mm_packus_epi16(_mm_srli_epi16(layer1_chunk0, 8), _mm_srli_epi16(layer1_chunk1, 8));
-    v_r1 = _mm_packus_epi16(_mm_and_si128(layer1_chunk2, v_mask), _mm_and_si128(layer1_chunk3, v_mask));
-    v_b1 = _mm_packus_epi16(_mm_srli_epi16(layer1_chunk2, 8), _mm_srli_epi16(layer1_chunk3, 8));
-    v_g0 = _mm_packus_epi16(_mm_and_si128(layer1_chunk4, v_mask), _mm_and_si128(layer1_chunk5, v_mask));
-    v_a0 = _mm_packus_epi16(_mm_srli_epi16(layer1_chunk4, 8), _mm_srli_epi16(layer1_chunk5, 8));
-    v_g1 = _mm_packus_epi16(_mm_and_si128(layer1_chunk6, v_mask), _mm_and_si128(layer1_chunk7, v_mask));
-    v_a1 = _mm_packus_epi16(_mm_srli_epi16(layer1_chunk6, 8), _mm_srli_epi16(layer1_chunk7, 8));
-}
-
-inline void _mm_deinterleave_epi16(__m128i & v_r0, __m128i & v_r1, __m128i & v_g0, __m128i & v_g1)
-{
-    __m128i layer1_chunk0 = _mm_unpacklo_epi16(v_r0, v_g0);
-    __m128i layer1_chunk1 = _mm_unpackhi_epi16(v_r0, v_g0);
-    __m128i layer1_chunk2 = _mm_unpacklo_epi16(v_r1, v_g1);
-    __m128i layer1_chunk3 = _mm_unpackhi_epi16(v_r1, v_g1);
-
-    __m128i layer2_chunk0 = _mm_unpacklo_epi16(layer1_chunk0, layer1_chunk2);
-    __m128i layer2_chunk1 = _mm_unpackhi_epi16(layer1_chunk0, layer1_chunk2);
-    __m128i layer2_chunk2 = _mm_unpacklo_epi16(layer1_chunk1, layer1_chunk3);
-    __m128i layer2_chunk3 = _mm_unpackhi_epi16(layer1_chunk1, layer1_chunk3);
-
-    __m128i layer3_chunk0 = _mm_unpacklo_epi16(layer2_chunk0, layer2_chunk2);
-    __m128i layer3_chunk1 = _mm_unpackhi_epi16(layer2_chunk0, layer2_chunk2);
-    __m128i layer3_chunk2 = _mm_unpacklo_epi16(layer2_chunk1, layer2_chunk3);
-    __m128i layer3_chunk3 = _mm_unpackhi_epi16(layer2_chunk1, layer2_chunk3);
-
-    v_r0 = _mm_unpacklo_epi16(layer3_chunk0, layer3_chunk2);
-    v_r1 = _mm_unpackhi_epi16(layer3_chunk0, layer3_chunk2);
-    v_g0 = _mm_unpacklo_epi16(layer3_chunk1, layer3_chunk3);
-    v_g1 = _mm_unpackhi_epi16(layer3_chunk1, layer3_chunk3);
-}
-
-inline void _mm_deinterleave_epi16(__m128i & v_r0, __m128i & v_r1, __m128i & v_g0,
-                                   __m128i & v_g1, __m128i & v_b0, __m128i & v_b1)
-{
-    __m128i layer1_chunk0 = _mm_unpacklo_epi16(v_r0, v_g1);
-    __m128i layer1_chunk1 = _mm_unpackhi_epi16(v_r0, v_g1);
-    __m128i layer1_chunk2 = _mm_unpacklo_epi16(v_r1, v_b0);
-    __m128i layer1_chunk3 = _mm_unpackhi_epi16(v_r1, v_b0);
-    __m128i layer1_chunk4 = _mm_unpacklo_epi16(v_g0, v_b1);
-    __m128i layer1_chunk5 = _mm_unpackhi_epi16(v_g0, v_b1);
-
-    __m128i layer2_chunk0 = _mm_unpacklo_epi16(layer1_chunk0, layer1_chunk3);
-    __m128i layer2_chunk1 = _mm_unpackhi_epi16(layer1_chunk0, layer1_chunk3);
-    __m128i layer2_chunk2 = _mm_unpacklo_epi16(layer1_chunk1, layer1_chunk4);
-    __m128i layer2_chunk3 = _mm_unpackhi_epi16(layer1_chunk1, layer1_chunk4);
-    __m128i layer2_chunk4 = _mm_unpacklo_epi16(layer1_chunk2, layer1_chunk5);
-    __m128i layer2_chunk5 = _mm_unpackhi_epi16(layer1_chunk2, layer1_chunk5);
-
-    __m128i layer3_chunk0 = _mm_unpacklo_epi16(layer2_chunk0, layer2_chunk3);
-    __m128i layer3_chunk1 = _mm_unpackhi_epi16(layer2_chunk0, layer2_chunk3);
-    __m128i layer3_chunk2 = _mm_unpacklo_epi16(layer2_chunk1, layer2_chunk4);
-    __m128i layer3_chunk3 = _mm_unpackhi_epi16(layer2_chunk1, layer2_chunk4);
-    __m128i layer3_chunk4 = _mm_unpacklo_epi16(layer2_chunk2, layer2_chunk5);
-    __m128i layer3_chunk5 = _mm_unpackhi_epi16(layer2_chunk2, layer2_chunk5);
-
-    v_r0 = _mm_unpacklo_epi16(layer3_chunk0, layer3_chunk3);
-    v_r1 = _mm_unpackhi_epi16(layer3_chunk0, layer3_chunk3);
-    v_g0 = _mm_unpacklo_epi16(layer3_chunk1, layer3_chunk4);
-    v_g1 = _mm_unpackhi_epi16(layer3_chunk1, layer3_chunk4);
-    v_b0 = _mm_unpacklo_epi16(layer3_chunk2, layer3_chunk5);
-    v_b1 = _mm_unpackhi_epi16(layer3_chunk2, layer3_chunk5);
-}
-
-inline void _mm_deinterleave_epi16(__m128i & v_r0, __m128i & v_r1, __m128i & v_g0, __m128i & v_g1,
-                                   __m128i & v_b0, __m128i & v_b1, __m128i & v_a0, __m128i & v_a1)
-{
-    __m128i layer1_chunk0 = _mm_unpacklo_epi16(v_r0, v_b0);
-    __m128i layer1_chunk1 = _mm_unpackhi_epi16(v_r0, v_b0);
-    __m128i layer1_chunk2 = _mm_unpacklo_epi16(v_r1, v_b1);
-    __m128i layer1_chunk3 = _mm_unpackhi_epi16(v_r1, v_b1);
-    __m128i layer1_chunk4 = _mm_unpacklo_epi16(v_g0, v_a0);
-    __m128i layer1_chunk5 = _mm_unpackhi_epi16(v_g0, v_a0);
-    __m128i layer1_chunk6 = _mm_unpacklo_epi16(v_g1, v_a1);
-    __m128i layer1_chunk7 = _mm_unpackhi_epi16(v_g1, v_a1);
-
-    __m128i layer2_chunk0 = _mm_unpacklo_epi16(layer1_chunk0, layer1_chunk4);
-    __m128i layer2_chunk1 = _mm_unpackhi_epi16(layer1_chunk0, layer1_chunk4);
-    __m128i layer2_chunk2 = _mm_unpacklo_epi16(layer1_chunk1, layer1_chunk5);
-    __m128i layer2_chunk3 = _mm_unpackhi_epi16(layer1_chunk1, layer1_chunk5);
-    __m128i layer2_chunk4 = _mm_unpacklo_epi16(layer1_chunk2, layer1_chunk6);
-    __m128i layer2_chunk5 = _mm_unpackhi_epi16(layer1_chunk2, layer1_chunk6);
-    __m128i layer2_chunk6 = _mm_unpacklo_epi16(layer1_chunk3, layer1_chunk7);
-    __m128i layer2_chunk7 = _mm_unpackhi_epi16(layer1_chunk3, layer1_chunk7);
-
-    __m128i layer3_chunk0 = _mm_unpacklo_epi16(layer2_chunk0, layer2_chunk4);
-    __m128i layer3_chunk1 = _mm_unpackhi_epi16(layer2_chunk0, layer2_chunk4);
-    __m128i layer3_chunk2 = _mm_unpacklo_epi16(layer2_chunk1, layer2_chunk5);
-    __m128i layer3_chunk3 = _mm_unpackhi_epi16(layer2_chunk1, layer2_chunk5);
-    __m128i layer3_chunk4 = _mm_unpacklo_epi16(layer2_chunk2, layer2_chunk6);
-    __m128i layer3_chunk5 = _mm_unpackhi_epi16(layer2_chunk2, layer2_chunk6);
-    __m128i layer3_chunk6 = _mm_unpacklo_epi16(layer2_chunk3, layer2_chunk7);
-    __m128i layer3_chunk7 = _mm_unpackhi_epi16(layer2_chunk3, layer2_chunk7);
-
-    v_r0 = _mm_unpacklo_epi16(layer3_chunk0, layer3_chunk4);
-    v_r1 = _mm_unpackhi_epi16(layer3_chunk0, layer3_chunk4);
-    v_g0 = _mm_unpacklo_epi16(layer3_chunk1, layer3_chunk5);
-    v_g1 = _mm_unpackhi_epi16(layer3_chunk1, layer3_chunk5);
-    v_b0 = _mm_unpacklo_epi16(layer3_chunk2, layer3_chunk6);
-    v_b1 = _mm_unpackhi_epi16(layer3_chunk2, layer3_chunk6);
-    v_a0 = _mm_unpacklo_epi16(layer3_chunk3, layer3_chunk7);
-    v_a1 = _mm_unpackhi_epi16(layer3_chunk3, layer3_chunk7);
-}
-
-#if CV_SSE4_1
-
-inline void _mm_interleave_epi16(__m128i & v_r0, __m128i & v_r1, __m128i & v_g0, __m128i & v_g1)
-{
-    __m128i v_mask = _mm_set1_epi32(0x0000ffff);
-
-    __m128i layer3_chunk0 = _mm_packus_epi32(_mm_and_si128(v_r0, v_mask), _mm_and_si128(v_r1, v_mask));
-    __m128i layer3_chunk2 = _mm_packus_epi32(_mm_srli_epi32(v_r0, 16), _mm_srli_epi32(v_r1, 16));
-    __m128i layer3_chunk1 = _mm_packus_epi32(_mm_and_si128(v_g0, v_mask), _mm_and_si128(v_g1, v_mask));
-    __m128i layer3_chunk3 = _mm_packus_epi32(_mm_srli_epi32(v_g0, 16), _mm_srli_epi32(v_g1, 16));
-
-    __m128i layer2_chunk0 = _mm_packus_epi32(_mm_and_si128(layer3_chunk0, v_mask), _mm_and_si128(layer3_chunk1, v_mask));
-    __m128i layer2_chunk2 = _mm_packus_epi32(_mm_srli_epi32(layer3_chunk0, 16), _mm_srli_epi32(layer3_chunk1, 16));
-    __m128i layer2_chunk1 = _mm_packus_epi32(_mm_and_si128(layer3_chunk2, v_mask), _mm_and_si128(layer3_chunk3, v_mask));
-    __m128i layer2_chunk3 = _mm_packus_epi32(_mm_srli_epi32(layer3_chunk2, 16), _mm_srli_epi32(layer3_chunk3, 16));
-
-    __m128i layer1_chunk0 = _mm_packus_epi32(_mm_and_si128(layer2_chunk0, v_mask), _mm_and_si128(layer2_chunk1, v_mask));
-    __m128i layer1_chunk2 = _mm_packus_epi32(_mm_srli_epi32(layer2_chunk0, 16), _mm_srli_epi32(layer2_chunk1, 16));
-    __m128i layer1_chunk1 = _mm_packus_epi32(_mm_and_si128(layer2_chunk2, v_mask), _mm_and_si128(layer2_chunk3, v_mask));
-    __m128i layer1_chunk3 = _mm_packus_epi32(_mm_srli_epi32(layer2_chunk2, 16), _mm_srli_epi32(layer2_chunk3, 16));
-
-    v_r0 = _mm_packus_epi32(_mm_and_si128(layer1_chunk0, v_mask), _mm_and_si128(layer1_chunk1, v_mask));
-    v_g0 = _mm_packus_epi32(_mm_srli_epi32(layer1_chunk0, 16), _mm_srli_epi32(layer1_chunk1, 16));
-    v_r1 = _mm_packus_epi32(_mm_and_si128(layer1_chunk2, v_mask), _mm_and_si128(layer1_chunk3, v_mask));
-    v_g1 = _mm_packus_epi32(_mm_srli_epi32(layer1_chunk2, 16), _mm_srli_epi32(layer1_chunk3, 16));
-}
-
-inline void _mm_interleave_epi16(__m128i & v_r0, __m128i & v_r1, __m128i & v_g0,
-                                 __m128i & v_g1, __m128i & v_b0, __m128i & v_b1)
-{
-    __m128i v_mask = _mm_set1_epi32(0x0000ffff);
-
-    __m128i layer3_chunk0 = _mm_packus_epi32(_mm_and_si128(v_r0, v_mask), _mm_and_si128(v_r1, v_mask));
-    __m128i layer3_chunk3 = _mm_packus_epi32(_mm_srli_epi32(v_r0, 16), _mm_srli_epi32(v_r1, 16));
-    __m128i layer3_chunk1 = _mm_packus_epi32(_mm_and_si128(v_g0, v_mask), _mm_and_si128(v_g1, v_mask));
-    __m128i layer3_chunk4 = _mm_packus_epi32(_mm_srli_epi32(v_g0, 16), _mm_srli_epi32(v_g1, 16));
-    __m128i layer3_chunk2 = _mm_packus_epi32(_mm_and_si128(v_b0, v_mask), _mm_and_si128(v_b1, v_mask));
-    __m128i layer3_chunk5 = _mm_packus_epi32(_mm_srli_epi32(v_b0, 16), _mm_srli_epi32(v_b1, 16));
-
-    __m128i layer2_chunk0 = _mm_packus_epi32(_mm_and_si128(layer3_chunk0, v_mask), _mm_and_si128(layer3_chunk1, v_mask));
-    __m128i layer2_chunk3 = _mm_packus_epi32(_mm_srli_epi32(layer3_chunk0, 16), _mm_srli_epi32(layer3_chunk1, 16));
-    __m128i layer2_chunk1 = _mm_packus_epi32(_mm_and_si128(layer3_chunk2, v_mask), _mm_and_si128(layer3_chunk3, v_mask));
-    __m128i layer2_chunk4 = _mm_packus_epi32(_mm_srli_epi32(layer3_chunk2, 16), _mm_srli_epi32(layer3_chunk3, 16));
-    __m128i layer2_chunk2 = _mm_packus_epi32(_mm_and_si128(layer3_chunk4, v_mask), _mm_and_si128(layer3_chunk5, v_mask));
-    __m128i layer2_chunk5 = _mm_packus_epi32(_mm_srli_epi32(layer3_chunk4, 16), _mm_srli_epi32(layer3_chunk5, 16));
-
-    __m128i layer1_chunk0 = _mm_packus_epi32(_mm_and_si128(layer2_chunk0, v_mask), _mm_and_si128(layer2_chunk1, v_mask));
-    __m128i layer1_chunk3 = _mm_packus_epi32(_mm_srli_epi32(layer2_chunk0, 16), _mm_srli_epi32(layer2_chunk1, 16));
-    __m128i layer1_chunk1 = _mm_packus_epi32(_mm_and_si128(layer2_chunk2, v_mask), _mm_and_si128(layer2_chunk3, v_mask));
-    __m128i layer1_chunk4 = _mm_packus_epi32(_mm_srli_epi32(layer2_chunk2, 16), _mm_srli_epi32(layer2_chunk3, 16));
-    __m128i layer1_chunk2 = _mm_packus_epi32(_mm_and_si128(layer2_chunk4, v_mask), _mm_and_si128(layer2_chunk5, v_mask));
-    __m128i layer1_chunk5 = _mm_packus_epi32(_mm_srli_epi32(layer2_chunk4, 16), _mm_srli_epi32(layer2_chunk5, 16));
-
-    v_r0 = _mm_packus_epi32(_mm_and_si128(layer1_chunk0, v_mask), _mm_and_si128(layer1_chunk1, v_mask));
-    v_g1 = _mm_packus_epi32(_mm_srli_epi32(layer1_chunk0, 16), _mm_srli_epi32(layer1_chunk1, 16));
-    v_r1 = _mm_packus_epi32(_mm_and_si128(layer1_chunk2, v_mask), _mm_and_si128(layer1_chunk3, v_mask));
-    v_b0 = _mm_packus_epi32(_mm_srli_epi32(layer1_chunk2, 16), _mm_srli_epi32(layer1_chunk3, 16));
-    v_g0 = _mm_packus_epi32(_mm_and_si128(layer1_chunk4, v_mask), _mm_and_si128(layer1_chunk5, v_mask));
-    v_b1 = _mm_packus_epi32(_mm_srli_epi32(layer1_chunk4, 16), _mm_srli_epi32(layer1_chunk5, 16));
-}
-
-inline void _mm_interleave_epi16(__m128i & v_r0, __m128i & v_r1, __m128i & v_g0, __m128i & v_g1,
-                                 __m128i & v_b0, __m128i & v_b1, __m128i & v_a0, __m128i & v_a1)
-{
-    __m128i v_mask = _mm_set1_epi32(0x0000ffff);
-
-    __m128i layer3_chunk0 = _mm_packus_epi32(_mm_and_si128(v_r0, v_mask), _mm_and_si128(v_r1, v_mask));
-    __m128i layer3_chunk4 = _mm_packus_epi32(_mm_srli_epi32(v_r0, 16), _mm_srli_epi32(v_r1, 16));
-    __m128i layer3_chunk1 = _mm_packus_epi32(_mm_and_si128(v_g0, v_mask), _mm_and_si128(v_g1, v_mask));
-    __m128i layer3_chunk5 = _mm_packus_epi32(_mm_srli_epi32(v_g0, 16), _mm_srli_epi32(v_g1, 16));
-    __m128i layer3_chunk2 = _mm_packus_epi32(_mm_and_si128(v_b0, v_mask), _mm_and_si128(v_b1, v_mask));
-    __m128i layer3_chunk6 = _mm_packus_epi32(_mm_srli_epi32(v_b0, 16), _mm_srli_epi32(v_b1, 16));
-    __m128i layer3_chunk3 = _mm_packus_epi32(_mm_and_si128(v_a0, v_mask), _mm_and_si128(v_a1, v_mask));
-    __m128i layer3_chunk7 = _mm_packus_epi32(_mm_srli_epi32(v_a0, 16), _mm_srli_epi32(v_a1, 16));
-
-    __m128i layer2_chunk0 = _mm_packus_epi32(_mm_and_si128(layer3_chunk0, v_mask), _mm_and_si128(layer3_chunk1, v_mask));
-    __m128i layer2_chunk4 = _mm_packus_epi32(_mm_srli_epi32(layer3_chunk0, 16), _mm_srli_epi32(layer3_chunk1, 16));
-    __m128i layer2_chunk1 = _mm_packus_epi32(_mm_and_si128(layer3_chunk2, v_mask), _mm_and_si128(layer3_chunk3, v_mask));
-    __m128i layer2_chunk5 = _mm_packus_epi32(_mm_srli_epi32(layer3_chunk2, 16), _mm_srli_epi32(layer3_chunk3, 16));
-    __m128i layer2_chunk2 = _mm_packus_epi32(_mm_and_si128(layer3_chunk4, v_mask), _mm_and_si128(layer3_chunk5, v_mask));
-    __m128i layer2_chunk6 = _mm_packus_epi32(_mm_srli_epi32(layer3_chunk4, 16), _mm_srli_epi32(layer3_chunk5, 16));
-    __m128i layer2_chunk3 = _mm_packus_epi32(_mm_and_si128(layer3_chunk6, v_mask), _mm_and_si128(layer3_chunk7, v_mask));
-    __m128i layer2_chunk7 = _mm_packus_epi32(_mm_srli_epi32(layer3_chunk6, 16), _mm_srli_epi32(layer3_chunk7, 16));
-
-    __m128i layer1_chunk0 = _mm_packus_epi32(_mm_and_si128(layer2_chunk0, v_mask), _mm_and_si128(layer2_chunk1, v_mask));
-    __m128i layer1_chunk4 = _mm_packus_epi32(_mm_srli_epi32(layer2_chunk0, 16), _mm_srli_epi32(layer2_chunk1, 16));
-    __m128i layer1_chunk1 = _mm_packus_epi32(_mm_and_si128(layer2_chunk2, v_mask), _mm_and_si128(layer2_chunk3, v_mask));
-    __m128i layer1_chunk5 = _mm_packus_epi32(_mm_srli_epi32(layer2_chunk2, 16), _mm_srli_epi32(layer2_chunk3, 16));
-    __m128i layer1_chunk2 = _mm_packus_epi32(_mm_and_si128(layer2_chunk4, v_mask), _mm_and_si128(layer2_chunk5, v_mask));
-    __m128i layer1_chunk6 = _mm_packus_epi32(_mm_srli_epi32(layer2_chunk4, 16), _mm_srli_epi32(layer2_chunk5, 16));
-    __m128i layer1_chunk3 = _mm_packus_epi32(_mm_and_si128(layer2_chunk6, v_mask), _mm_and_si128(layer2_chunk7, v_mask));
-    __m128i layer1_chunk7 = _mm_packus_epi32(_mm_srli_epi32(layer2_chunk6, 16), _mm_srli_epi32(layer2_chunk7, 16));
-
-    v_r0 = _mm_packus_epi32(_mm_and_si128(layer1_chunk0, v_mask), _mm_and_si128(layer1_chunk1, v_mask));
-    v_b0 = _mm_packus_epi32(_mm_srli_epi32(layer1_chunk0, 16), _mm_srli_epi32(layer1_chunk1, 16));
-    v_r1 = _mm_packus_epi32(_mm_and_si128(layer1_chunk2, v_mask), _mm_and_si128(layer1_chunk3, v_mask));
-    v_b1 = _mm_packus_epi32(_mm_srli_epi32(layer1_chunk2, 16), _mm_srli_epi32(layer1_chunk3, 16));
-    v_g0 = _mm_packus_epi32(_mm_and_si128(layer1_chunk4, v_mask), _mm_and_si128(layer1_chunk5, v_mask));
-    v_a0 = _mm_packus_epi32(_mm_srli_epi32(layer1_chunk4, 16), _mm_srli_epi32(layer1_chunk5, 16));
-    v_g1 = _mm_packus_epi32(_mm_and_si128(layer1_chunk6, v_mask), _mm_and_si128(layer1_chunk7, v_mask));
-    v_a1 = _mm_packus_epi32(_mm_srli_epi32(layer1_chunk6, 16), _mm_srli_epi32(layer1_chunk7, 16));
-}
-
-#endif // CV_SSE4_1
-
-inline void _mm_deinterleave_ps(__m128 & v_r0, __m128 & v_r1, __m128 & v_g0, __m128 & v_g1)
-{
-    __m128 layer1_chunk0 = _mm_unpacklo_ps(v_r0, v_g0);
-    __m128 layer1_chunk1 = _mm_unpackhi_ps(v_r0, v_g0);
-    __m128 layer1_chunk2 = _mm_unpacklo_ps(v_r1, v_g1);
-    __m128 layer1_chunk3 = _mm_unpackhi_ps(v_r1, v_g1);
-
-    __m128 layer2_chunk0 = _mm_unpacklo_ps(layer1_chunk0, layer1_chunk2);
-    __m128 layer2_chunk1 = _mm_unpackhi_ps(layer1_chunk0, layer1_chunk2);
-    __m128 layer2_chunk2 = _mm_unpacklo_ps(layer1_chunk1, layer1_chunk3);
-    __m128 layer2_chunk3 = _mm_unpackhi_ps(layer1_chunk1, layer1_chunk3);
-
-    v_r0 = _mm_unpacklo_ps(layer2_chunk0, layer2_chunk2);
-    v_r1 = _mm_unpackhi_ps(layer2_chunk0, layer2_chunk2);
-    v_g0 = _mm_unpacklo_ps(layer2_chunk1, layer2_chunk3);
-    v_g1 = _mm_unpackhi_ps(layer2_chunk1, layer2_chunk3);
-}
-
-inline void _mm_deinterleave_ps(__m128 & v_r0, __m128 & v_r1, __m128 & v_g0,
-                                __m128 & v_g1, __m128 & v_b0, __m128 & v_b1)
-{
-    __m128 layer1_chunk0 = _mm_unpacklo_ps(v_r0, v_g1);
-    __m128 layer1_chunk1 = _mm_unpackhi_ps(v_r0, v_g1);
-    __m128 layer1_chunk2 = _mm_unpacklo_ps(v_r1, v_b0);
-    __m128 layer1_chunk3 = _mm_unpackhi_ps(v_r1, v_b0);
-    __m128 layer1_chunk4 = _mm_unpacklo_ps(v_g0, v_b1);
-    __m128 layer1_chunk5 = _mm_unpackhi_ps(v_g0, v_b1);
-
-    __m128 layer2_chunk0 = _mm_unpacklo_ps(layer1_chunk0, layer1_chunk3);
-    __m128 layer2_chunk1 = _mm_unpackhi_ps(layer1_chunk0, layer1_chunk3);
-    __m128 layer2_chunk2 = _mm_unpacklo_ps(layer1_chunk1, layer1_chunk4);
-    __m128 layer2_chunk3 = _mm_unpackhi_ps(layer1_chunk1, layer1_chunk4);
-    __m128 layer2_chunk4 = _mm_unpacklo_ps(layer1_chunk2, layer1_chunk5);
-    __m128 layer2_chunk5 = _mm_unpackhi_ps(layer1_chunk2, layer1_chunk5);
-
-    v_r0 = _mm_unpacklo_ps(layer2_chunk0, layer2_chunk3);
-    v_r1 = _mm_unpackhi_ps(layer2_chunk0, layer2_chunk3);
-    v_g0 = _mm_unpacklo_ps(layer2_chunk1, layer2_chunk4);
-    v_g1 = _mm_unpackhi_ps(layer2_chunk1, layer2_chunk4);
-    v_b0 = _mm_unpacklo_ps(layer2_chunk2, layer2_chunk5);
-    v_b1 = _mm_unpackhi_ps(layer2_chunk2, layer2_chunk5);
-}
-
-inline void _mm_deinterleave_ps(__m128 & v_r0, __m128 & v_r1, __m128 & v_g0, __m128 & v_g1,
-                                __m128 & v_b0, __m128 & v_b1, __m128 & v_a0, __m128 & v_a1)
-{
-    __m128 layer1_chunk0 = _mm_unpacklo_ps(v_r0, v_b0);
-    __m128 layer1_chunk1 = _mm_unpackhi_ps(v_r0, v_b0);
-    __m128 layer1_chunk2 = _mm_unpacklo_ps(v_r1, v_b1);
-    __m128 layer1_chunk3 = _mm_unpackhi_ps(v_r1, v_b1);
-    __m128 layer1_chunk4 = _mm_unpacklo_ps(v_g0, v_a0);
-    __m128 layer1_chunk5 = _mm_unpackhi_ps(v_g0, v_a0);
-    __m128 layer1_chunk6 = _mm_unpacklo_ps(v_g1, v_a1);
-    __m128 layer1_chunk7 = _mm_unpackhi_ps(v_g1, v_a1);
-
-    __m128 layer2_chunk0 = _mm_unpacklo_ps(layer1_chunk0, layer1_chunk4);
-    __m128 layer2_chunk1 = _mm_unpackhi_ps(layer1_chunk0, layer1_chunk4);
-    __m128 layer2_chunk2 = _mm_unpacklo_ps(layer1_chunk1, layer1_chunk5);
-    __m128 layer2_chunk3 = _mm_unpackhi_ps(layer1_chunk1, layer1_chunk5);
-    __m128 layer2_chunk4 = _mm_unpacklo_ps(layer1_chunk2, layer1_chunk6);
-    __m128 layer2_chunk5 = _mm_unpackhi_ps(layer1_chunk2, layer1_chunk6);
-    __m128 layer2_chunk6 = _mm_unpacklo_ps(layer1_chunk3, layer1_chunk7);
-    __m128 layer2_chunk7 = _mm_unpackhi_ps(layer1_chunk3, layer1_chunk7);
-
-    v_r0 = _mm_unpacklo_ps(layer2_chunk0, layer2_chunk4);
-    v_r1 = _mm_unpackhi_ps(layer2_chunk0, layer2_chunk4);
-    v_g0 = _mm_unpacklo_ps(layer2_chunk1, layer2_chunk5);
-    v_g1 = _mm_unpackhi_ps(layer2_chunk1, layer2_chunk5);
-    v_b0 = _mm_unpacklo_ps(layer2_chunk2, layer2_chunk6);
-    v_b1 = _mm_unpackhi_ps(layer2_chunk2, layer2_chunk6);
-    v_a0 = _mm_unpacklo_ps(layer2_chunk3, layer2_chunk7);
-    v_a1 = _mm_unpackhi_ps(layer2_chunk3, layer2_chunk7);
-}
-
-inline void _mm_interleave_ps(__m128 & v_r0, __m128 & v_r1, __m128 & v_g0, __m128 & v_g1)
-{
-    enum { mask_lo = _MM_SHUFFLE(2, 0, 2, 0), mask_hi = _MM_SHUFFLE(3, 1, 3, 1) };
-
-    __m128 layer2_chunk0 = _mm_shuffle_ps(v_r0, v_r1, mask_lo);
-    __m128 layer2_chunk2 = _mm_shuffle_ps(v_r0, v_r1, mask_hi);
-    __m128 layer2_chunk1 = _mm_shuffle_ps(v_g0, v_g1, mask_lo);
-    __m128 layer2_chunk3 = _mm_shuffle_ps(v_g0, v_g1, mask_hi);
-
-    __m128 layer1_chunk0 = _mm_shuffle_ps(layer2_chunk0, layer2_chunk1, mask_lo);
-    __m128 layer1_chunk2 = _mm_shuffle_ps(layer2_chunk0, layer2_chunk1, mask_hi);
-    __m128 layer1_chunk1 = _mm_shuffle_ps(layer2_chunk2, layer2_chunk3, mask_lo);
-    __m128 layer1_chunk3 = _mm_shuffle_ps(layer2_chunk2, layer2_chunk3, mask_hi);
-
-    v_r0 = _mm_shuffle_ps(layer1_chunk0, layer1_chunk1, mask_lo);
-    v_g0 = _mm_shuffle_ps(layer1_chunk0, layer1_chunk1, mask_hi);
-    v_r1 = _mm_shuffle_ps(layer1_chunk2, layer1_chunk3, mask_lo);
-    v_g1 = _mm_shuffle_ps(layer1_chunk2, layer1_chunk3, mask_hi);
-}
-
-inline void _mm_interleave_ps(__m128 & v_r0, __m128 & v_r1, __m128 & v_g0,
-                              __m128 & v_g1, __m128 & v_b0, __m128 & v_b1)
-{
-    enum { mask_lo = _MM_SHUFFLE(2, 0, 2, 0), mask_hi = _MM_SHUFFLE(3, 1, 3, 1) };
-
-    __m128 layer2_chunk0 = _mm_shuffle_ps(v_r0, v_r1, mask_lo);
-    __m128 layer2_chunk3 = _mm_shuffle_ps(v_r0, v_r1, mask_hi);
-    __m128 layer2_chunk1 = _mm_shuffle_ps(v_g0, v_g1, mask_lo);
-    __m128 layer2_chunk4 = _mm_shuffle_ps(v_g0, v_g1, mask_hi);
-    __m128 layer2_chunk2 = _mm_shuffle_ps(v_b0, v_b1, mask_lo);
-    __m128 layer2_chunk5 = _mm_shuffle_ps(v_b0, v_b1, mask_hi);
-
-    __m128 layer1_chunk0 = _mm_shuffle_ps(layer2_chunk0, layer2_chunk1, mask_lo);
-    __m128 layer1_chunk3 = _mm_shuffle_ps(layer2_chunk0, layer2_chunk1, mask_hi);
-    __m128 layer1_chunk1 = _mm_shuffle_ps(layer2_chunk2, layer2_chunk3, mask_lo);
-    __m128 layer1_chunk4 = _mm_shuffle_ps(layer2_chunk2, layer2_chunk3, mask_hi);
-    __m128 layer1_chunk2 = _mm_shuffle_ps(layer2_chunk4, layer2_chunk5, mask_lo);
-    __m128 layer1_chunk5 = _mm_shuffle_ps(layer2_chunk4, layer2_chunk5, mask_hi);
-
-    v_r0 = _mm_shuffle_ps(layer1_chunk0, layer1_chunk1, mask_lo);
-    v_g1 = _mm_shuffle_ps(layer1_chunk0, layer1_chunk1, mask_hi);
-    v_r1 = _mm_shuffle_ps(layer1_chunk2, layer1_chunk3, mask_lo);
-    v_b0 = _mm_shuffle_ps(layer1_chunk2, layer1_chunk3, mask_hi);
-    v_g0 = _mm_shuffle_ps(layer1_chunk4, layer1_chunk5, mask_lo);
-    v_b1 = _mm_shuffle_ps(layer1_chunk4, layer1_chunk5, mask_hi);
-}
-
-inline void _mm_interleave_ps(__m128 & v_r0, __m128 & v_r1, __m128 & v_g0, __m128 & v_g1,
-                              __m128 & v_b0, __m128 & v_b1, __m128 & v_a0, __m128 & v_a1)
-{
-    enum { mask_lo = _MM_SHUFFLE(2, 0, 2, 0), mask_hi = _MM_SHUFFLE(3, 1, 3, 1) };
-
-    __m128 layer2_chunk0 = _mm_shuffle_ps(v_r0, v_r1, mask_lo);
-    __m128 layer2_chunk4 = _mm_shuffle_ps(v_r0, v_r1, mask_hi);
-    __m128 layer2_chunk1 = _mm_shuffle_ps(v_g0, v_g1, mask_lo);
-    __m128 layer2_chunk5 = _mm_shuffle_ps(v_g0, v_g1, mask_hi);
-    __m128 layer2_chunk2 = _mm_shuffle_ps(v_b0, v_b1, mask_lo);
-    __m128 layer2_chunk6 = _mm_shuffle_ps(v_b0, v_b1, mask_hi);
-    __m128 layer2_chunk3 = _mm_shuffle_ps(v_a0, v_a1, mask_lo);
-    __m128 layer2_chunk7 = _mm_shuffle_ps(v_a0, v_a1, mask_hi);
-
-    __m128 layer1_chunk0 = _mm_shuffle_ps(layer2_chunk0, layer2_chunk1, mask_lo);
-    __m128 layer1_chunk4 = _mm_shuffle_ps(layer2_chunk0, layer2_chunk1, mask_hi);
-    __m128 layer1_chunk1 = _mm_shuffle_ps(layer2_chunk2, layer2_chunk3, mask_lo);
-    __m128 layer1_chunk5 = _mm_shuffle_ps(layer2_chunk2, layer2_chunk3, mask_hi);
-    __m128 layer1_chunk2 = _mm_shuffle_ps(layer2_chunk4, layer2_chunk5, mask_lo);
-    __m128 layer1_chunk6 = _mm_shuffle_ps(layer2_chunk4, layer2_chunk5, mask_hi);
-    __m128 layer1_chunk3 = _mm_shuffle_ps(layer2_chunk6, layer2_chunk7, mask_lo);
-    __m128 layer1_chunk7 = _mm_shuffle_ps(layer2_chunk6, layer2_chunk7, mask_hi);
-
-    v_r0 = _mm_shuffle_ps(layer1_chunk0, layer1_chunk1, mask_lo);
-    v_b0 = _mm_shuffle_ps(layer1_chunk0, layer1_chunk1, mask_hi);
-    v_r1 = _mm_shuffle_ps(layer1_chunk2, layer1_chunk3, mask_lo);
-    v_b1 = _mm_shuffle_ps(layer1_chunk2, layer1_chunk3, mask_hi);
-    v_g0 = _mm_shuffle_ps(layer1_chunk4, layer1_chunk5, mask_lo);
-    v_a0 = _mm_shuffle_ps(layer1_chunk4, layer1_chunk5, mask_hi);
-    v_g1 = _mm_shuffle_ps(layer1_chunk6, layer1_chunk7, mask_lo);
-    v_a1 = _mm_shuffle_ps(layer1_chunk6, layer1_chunk7, mask_hi);
-}
-
-#endif // CV_SSE2
-
-//! @}
-
-#endif //OPENCV_CORE_SSE_UTILS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/traits.hpp b/projects/llm_framework/include/opencv4/opencv2/core/traits.hpp
deleted file mode 100644
index 52ab083c..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/traits.hpp
+++ /dev/null
@@ -1,417 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_TRAITS_HPP
-#define OPENCV_CORE_TRAITS_HPP
-
-#include "opencv2/core/cvdef.h"
-
-namespace cv
-{
-
-//#define OPENCV_TRAITS_ENABLE_DEPRECATED
-
-//! @addtogroup core_basic
-//! @{
-
-/** @brief Template "trait" class for OpenCV primitive data types.
-
-@note Deprecated. This is replaced by "single purpose" traits: traits::Type and traits::Depth
-
-A primitive OpenCV data type is one of unsigned char, bool, signed char, unsigned short, signed
-short, int, float, double, or a tuple of values of one of these types, where all the values in the
-tuple have the same type. Any primitive type from the list can be defined by an identifier in the
-form CV_\<bit-depth\>{U|S|F}C(\<number_of_channels\>), for example: uchar \~ CV_8UC1, 3-element
-floating-point tuple \~ CV_32FC3, and so on. A universal OpenCV structure that is able to store a
-single instance of such a primitive data type is Vec. Multiple instances of such a type can be
-stored in a std::vector, Mat, Mat_, SparseMat, SparseMat_, or any other container that is able to
-store Vec instances.
-
-The DataType class is basically used to provide a description of such primitive data types without
-adding any fields or methods to the corresponding classes (and it is actually impossible to add
-anything to primitive C/C++ data types). This technique is known in C++ as class traits. It is not
-DataType itself that is used but its specialized versions, such as:
-@code
-    template<> class DataType<uchar>
-    {
-        typedef uchar value_type;
-        typedef int work_type;
-        typedef uchar channel_type;
-        enum { channel_type = CV_8U, channels = 1, fmt='u', type = CV_8U };
-    };
-    ...
-    template<typename _Tp> DataType<std::complex<_Tp> >
-    {
-        typedef std::complex<_Tp> value_type;
-        typedef std::complex<_Tp> work_type;
-        typedef _Tp channel_type;
-        // DataDepth is another helper trait class
-        enum { depth = DataDepth<_Tp>::value, channels=2,
-            fmt=(channels-1)*256+DataDepth<_Tp>::fmt,
-            type=CV_MAKETYPE(depth, channels) };
-    };
-    ...
-@endcode
-The main purpose of this class is to convert compilation-time type information to an
-OpenCV-compatible data type identifier, for example:
-@code
-    // allocates a 30x40 floating-point matrix
-    Mat A(30, 40, DataType<float>::type);
-
-    Mat B = Mat_<std::complex<double> >(3, 3);
-    // the statement below will print 6, 2 , that is depth == CV_64F, channels == 2
-    cout << B.depth() << ", " << B.channels() << endl;
-@endcode
-So, such traits are used to tell OpenCV which data type you are working with, even if such a type is
-not native to OpenCV. For example, the matrix B initialization above is compiled because OpenCV
-defines the proper specialized template class DataType\<complex\<_Tp\> \> . This mechanism is also
-useful (and used in OpenCV this way) for generic algorithms implementations.
-
-@note Default values were dropped to stop confusing developers about using of unsupported types (see #7599)
-*/
-template<typename _Tp> class DataType
-{
-public:
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-    typedef _Tp         value_type;
-    typedef value_type  work_type;
-    typedef value_type  channel_type;
-    typedef value_type  vec_type;
-    enum { generic_type = 1,
-           depth        = -1,
-           channels     = 1,
-           fmt          = 0,
-           type = CV_MAKETYPE(depth, channels)
-         };
-#endif
-};
-
-template<> class DataType<bool>
-{
-public:
-    typedef bool        value_type;
-    typedef int         work_type;
-    typedef value_type  channel_type;
-    typedef value_type  vec_type;
-    enum { generic_type = 0,
-           depth        = CV_8U,
-           channels     = 1,
-           fmt          = (int)'u',
-           type         = CV_MAKETYPE(depth, channels)
-         };
-};
-
-template<> class DataType<uchar>
-{
-public:
-    typedef uchar       value_type;
-    typedef int         work_type;
-    typedef value_type  channel_type;
-    typedef value_type  vec_type;
-    enum { generic_type = 0,
-           depth        = CV_8U,
-           channels     = 1,
-           fmt          = (int)'u',
-           type         = CV_MAKETYPE(depth, channels)
-         };
-};
-
-template<> class DataType<schar>
-{
-public:
-    typedef schar       value_type;
-    typedef int         work_type;
-    typedef value_type  channel_type;
-    typedef value_type  vec_type;
-    enum { generic_type = 0,
-           depth        = CV_8S,
-           channels     = 1,
-           fmt          = (int)'c',
-           type         = CV_MAKETYPE(depth, channels)
-         };
-};
-
-template<> class DataType<char>
-{
-public:
-    typedef schar       value_type;
-    typedef int         work_type;
-    typedef value_type  channel_type;
-    typedef value_type  vec_type;
-    enum { generic_type = 0,
-           depth        = CV_8S,
-           channels     = 1,
-           fmt          = (int)'c',
-           type         = CV_MAKETYPE(depth, channels)
-         };
-};
-
-template<> class DataType<ushort>
-{
-public:
-    typedef ushort      value_type;
-    typedef int         work_type;
-    typedef value_type  channel_type;
-    typedef value_type  vec_type;
-    enum { generic_type = 0,
-           depth        = CV_16U,
-           channels     = 1,
-           fmt          = (int)'w',
-           type         = CV_MAKETYPE(depth, channels)
-         };
-};
-
-template<> class DataType<short>
-{
-public:
-    typedef short       value_type;
-    typedef int         work_type;
-    typedef value_type  channel_type;
-    typedef value_type  vec_type;
-    enum { generic_type = 0,
-           depth        = CV_16S,
-           channels     = 1,
-           fmt          = (int)'s',
-           type         = CV_MAKETYPE(depth, channels)
-         };
-};
-
-template<> class DataType<int>
-{
-public:
-    typedef int         value_type;
-    typedef value_type  work_type;
-    typedef value_type  channel_type;
-    typedef value_type  vec_type;
-    enum { generic_type = 0,
-           depth        = CV_32S,
-           channels     = 1,
-           fmt          = (int)'i',
-           type         = CV_MAKETYPE(depth, channels)
-         };
-};
-
-template<> class DataType<float>
-{
-public:
-    typedef float       value_type;
-    typedef value_type  work_type;
-    typedef value_type  channel_type;
-    typedef value_type  vec_type;
-    enum { generic_type = 0,
-           depth        = CV_32F,
-           channels     = 1,
-           fmt          = (int)'f',
-           type         = CV_MAKETYPE(depth, channels)
-         };
-};
-
-template<> class DataType<double>
-{
-public:
-    typedef double      value_type;
-    typedef value_type  work_type;
-    typedef value_type  channel_type;
-    typedef value_type  vec_type;
-    enum { generic_type = 0,
-           depth        = CV_64F,
-           channels     = 1,
-           fmt          = (int)'d',
-           type         = CV_MAKETYPE(depth, channels)
-         };
-};
-
-template<> class DataType<float16_t>
-{
-public:
-    typedef float16_t   value_type;
-    typedef float       work_type;
-    typedef value_type  channel_type;
-    typedef value_type  vec_type;
-    enum { generic_type = 0,
-           depth        = CV_16F,
-           channels     = 1,
-           fmt          = (int)'h',
-           type         = CV_MAKETYPE(depth, channels)
-         };
-};
-
-/** @brief A helper class for cv::DataType
-
-The class is specialized for each fundamental numerical data type supported by OpenCV. It provides
-DataDepth<T>::value constant.
-*/
-template<typename _Tp> class DataDepth
-{
-public:
-    enum
-    {
-        value = DataType<_Tp>::depth,
-        fmt   = DataType<_Tp>::fmt
-    };
-};
-
-
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-
-template<int _depth> class TypeDepth
-{
-#ifdef OPENCV_TRAITS_ENABLE_LEGACY_DEFAULTS
-    enum { depth = CV_USRTYPE1 };
-    typedef void value_type;
-#endif
-};
-
-template<> class TypeDepth<CV_8U>
-{
-    enum { depth = CV_8U };
-    typedef uchar value_type;
-};
-
-template<> class TypeDepth<CV_8S>
-{
-    enum { depth = CV_8S };
-    typedef schar value_type;
-};
-
-template<> class TypeDepth<CV_16U>
-{
-    enum { depth = CV_16U };
-    typedef ushort value_type;
-};
-
-template<> class TypeDepth<CV_16S>
-{
-    enum { depth = CV_16S };
-    typedef short value_type;
-};
-
-template<> class TypeDepth<CV_32S>
-{
-    enum { depth = CV_32S };
-    typedef int value_type;
-};
-
-template<> class TypeDepth<CV_32F>
-{
-    enum { depth = CV_32F };
-    typedef float value_type;
-};
-
-template<> class TypeDepth<CV_64F>
-{
-    enum { depth = CV_64F };
-    typedef double value_type;
-};
-
-template<> class TypeDepth<CV_16F>
-{
-    enum { depth = CV_16F };
-    typedef float16_t value_type;
-};
-
-#endif
-
-//! @}
-
-namespace traits {
-
-namespace internal {
-#define CV_CREATE_MEMBER_CHECK(X) \
-template<typename T> class CheckMember_##X { \
-    struct Fallback { int X; }; \
-    struct Derived : T, Fallback { }; \
-    template<typename U, U> struct Check; \
-    typedef char CV_NO[1]; \
-    typedef char CV_YES[2]; \
-    template<typename U> static CV_NO & func(Check<int Fallback::*, &U::X> *); \
-    template<typename U> static CV_YES & func(...); \
-public: \
-    typedef CheckMember_##X type; \
-    enum { value = sizeof(func<Derived>(0)) == sizeof(CV_YES) }; \
-};
-
-CV_CREATE_MEMBER_CHECK(fmt)
-CV_CREATE_MEMBER_CHECK(type)
-
-} // namespace internal
-
-
-template<typename T>
-struct Depth
-{ enum { value = DataType<T>::depth }; };
-
-template<typename T>
-struct Type
-{ enum { value = DataType<T>::type }; };
-
-/** Similar to traits::Type<T> but has value = -1 in case of unknown type (instead of compiler error) */
-template<typename T, bool available = internal::CheckMember_type< DataType<T> >::value >
-struct SafeType {};
-
-template<typename T>
-struct SafeType<T, false>
-{ enum { value = -1 }; };
-
-template<typename T>
-struct SafeType<T, true>
-{ enum { value = Type<T>::value }; };
-
-
-template<typename T, bool available = internal::CheckMember_fmt< DataType<T> >::value >
-struct SafeFmt {};
-
-template<typename T>
-struct SafeFmt<T, false>
-{ enum { fmt = 0 }; };
-
-template<typename T>
-struct SafeFmt<T, true>
-{ enum { fmt = DataType<T>::fmt }; };
-
-
-} // namespace
-
-} // cv
-
-#endif // OPENCV_CORE_TRAITS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/types.hpp b/projects/llm_framework/include/opencv4/opencv2/core/types.hpp
deleted file mode 100644
index 28675203..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/types.hpp
+++ /dev/null
@@ -1,2439 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_TYPES_HPP
-#define OPENCV_CORE_TYPES_HPP
-
-#ifndef __cplusplus
-#  error types.hpp header must be compiled as C++
-#endif
-
-#include <climits>
-#include <cfloat>
-#include <vector>
-#include <limits>
-
-#include "opencv2/core/cvdef.h"
-#include "opencv2/core/cvstd.hpp"
-#include "opencv2/core/matx.hpp"
-
-namespace cv
-{
-
-//! @addtogroup core_basic
-//! @{
-
-//////////////////////////////// Complex //////////////////////////////
-
-/** @brief  A complex number class.
-
-  The template class is similar and compatible with std::complex, however it provides slightly
-  more convenient access to the real and imaginary parts using through the simple field access, as opposite
-  to std::complex::real() and std::complex::imag().
-*/
-template<typename _Tp> class Complex
-{
-public:
-
-    //! default constructor
-    Complex();
-    Complex( _Tp _re, _Tp _im = 0 );
-
-    //! conversion to another data type
-    template<typename T2> operator Complex<T2>() const;
-    //! conjugation
-    Complex conj() const;
-
-    _Tp re, im; //< the real and the imaginary parts
-};
-
-typedef Complex<float> Complexf;
-typedef Complex<double> Complexd;
-
-template<typename _Tp> class DataType< Complex<_Tp> >
-{
-public:
-    typedef Complex<_Tp> value_type;
-    typedef value_type   work_type;
-    typedef _Tp          channel_type;
-
-    enum { generic_type = 0,
-           channels     = 2,
-           fmt          = DataType<channel_type>::fmt + ((channels - 1) << 8)
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-           ,depth        = DataType<channel_type>::depth
-           ,type         = CV_MAKETYPE(depth, channels)
-#endif
-    };
-
-    typedef Vec<channel_type, channels> vec_type;
-};
-
-namespace traits {
-template<typename _Tp>
-struct Depth< Complex<_Tp> > { enum { value = Depth<_Tp>::value }; };
-template<typename _Tp>
-struct Type< Complex<_Tp> > { enum { value = CV_MAKETYPE(Depth<_Tp>::value, 2) }; };
-} // namespace
-
-
-//////////////////////////////// Point_ ////////////////////////////////
-
-/** @brief Template class for 2D points specified by its coordinates `x` and `y`.
-
-An instance of the class is interchangeable with C structures, CvPoint and CvPoint2D32f . There is
-also a cast operator to convert point coordinates to the specified type. The conversion from
-floating-point coordinates to integer coordinates is done by rounding. Commonly, the conversion
-uses this operation for each of the coordinates. Besides the class members listed in the
-declaration above, the following operations on points are implemented:
-@code
-    pt1 = pt2 + pt3;
-    pt1 = pt2 - pt3;
-    pt1 = pt2 * a;
-    pt1 = a * pt2;
-    pt1 = pt2 / a;
-    pt1 += pt2;
-    pt1 -= pt2;
-    pt1 *= a;
-    pt1 /= a;
-    double value = norm(pt); // L2 norm
-    pt1 == pt2;
-    pt1 != pt2;
-@endcode
-For your convenience, the following type aliases are defined:
-@code
-    typedef Point_<int> Point2i;
-    typedef Point2i Point;
-    typedef Point_<float> Point2f;
-    typedef Point_<double> Point2d;
-@endcode
-Example:
-@code
-    Point2f a(0.3f, 0.f), b(0.f, 0.4f);
-    Point pt = (a + b)*10.f;
-    cout << pt.x << ", " << pt.y << endl;
-@endcode
-*/
-template<typename _Tp> class Point_
-{
-public:
-    typedef _Tp value_type;
-
-    //! default constructor
-    Point_();
-    Point_(_Tp _x, _Tp _y);
-#if (defined(__GNUC__) && __GNUC__ < 5) && !defined(__clang__)  // GCC 4.x bug. Details: https://github.com/opencv/opencv/pull/20837
-    Point_(const Point_& pt);
-    Point_(Point_&& pt) CV_NOEXCEPT = default;
-#elif OPENCV_ABI_COMPATIBILITY < 500
-    Point_(const Point_& pt) = default;
-    Point_(Point_&& pt) CV_NOEXCEPT = default;
-#endif
-    Point_(const Size_<_Tp>& sz);
-    Point_(const Vec<_Tp, 2>& v);
-
-#if (defined(__GNUC__) && __GNUC__ < 5) && !defined(__clang__)  // GCC 4.x bug. Details: https://github.com/opencv/opencv/pull/20837
-    Point_& operator = (const Point_& pt);
-    Point_& operator = (Point_&& pt) CV_NOEXCEPT = default;
-#elif OPENCV_ABI_COMPATIBILITY < 500
-    Point_& operator = (const Point_& pt) = default;
-    Point_& operator = (Point_&& pt) CV_NOEXCEPT = default;
-#endif
-    //! conversion to another data type
-    template<typename _Tp2> operator Point_<_Tp2>() const;
-
-    //! conversion to the old-style C structures
-    operator Vec<_Tp, 2>() const;
-
-    //! dot product
-    _Tp dot(const Point_& pt) const;
-    //! dot product computed in double-precision arithmetics
-    double ddot(const Point_& pt) const;
-    //! cross-product
-    double cross(const Point_& pt) const;
-    //! checks whether the point is inside the specified rectangle
-    bool inside(const Rect_<_Tp>& r) const;
-    _Tp x; //!< x coordinate of the point
-    _Tp y; //!< y coordinate of the point
-};
-
-typedef Point_<int> Point2i;
-typedef Point_<int64> Point2l;
-typedef Point_<float> Point2f;
-typedef Point_<double> Point2d;
-typedef Point2i Point;
-
-template<typename _Tp> class DataType< Point_<_Tp> >
-{
-public:
-    typedef Point_<_Tp>                               value_type;
-    typedef Point_<typename DataType<_Tp>::work_type> work_type;
-    typedef _Tp                                       channel_type;
-
-    enum { generic_type = 0,
-           channels     = 2,
-           fmt          = traits::SafeFmt<channel_type>::fmt + ((channels - 1) << 8)
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-           ,depth        = DataType<channel_type>::depth
-           ,type         = CV_MAKETYPE(depth, channels)
-#endif
-         };
-
-    typedef Vec<channel_type, channels> vec_type;
-};
-
-namespace traits {
-template<typename _Tp>
-struct Depth< Point_<_Tp> > { enum { value = Depth<_Tp>::value }; };
-template<typename _Tp>
-struct Type< Point_<_Tp> > { enum { value = CV_MAKETYPE(Depth<_Tp>::value, 2) }; };
-} // namespace
-
-
-//////////////////////////////// Point3_ ////////////////////////////////
-
-/** @brief Template class for 3D points specified by its coordinates `x`, `y` and `z`.
-
-An instance of the class is interchangeable with the C structure CvPoint2D32f . Similarly to
-Point_ , the coordinates of 3D points can be converted to another type. The vector arithmetic and
-comparison operations are also supported.
-
-The following Point3_\<\> aliases are available:
-@code
-    typedef Point3_<int> Point3i;
-    typedef Point3_<float> Point3f;
-    typedef Point3_<double> Point3d;
-@endcode
-@see cv::Point3i, cv::Point3f and cv::Point3d
-*/
-template<typename _Tp> class Point3_
-{
-public:
-    typedef _Tp value_type;
-
-    //! default constructor
-    Point3_();
-    Point3_(_Tp _x, _Tp _y, _Tp _z);
-#if OPENCV_ABI_COMPATIBILITY < 500
-    Point3_(const Point3_& pt) = default;
-    Point3_(Point3_&& pt) CV_NOEXCEPT = default;
-#endif
-    explicit Point3_(const Point_<_Tp>& pt);
-    Point3_(const Vec<_Tp, 3>& v);
-
-#if OPENCV_ABI_COMPATIBILITY < 500
-    Point3_& operator = (const Point3_& pt) = default;
-    Point3_& operator = (Point3_&& pt) CV_NOEXCEPT = default;
-#endif
-    //! conversion to another data type
-    template<typename _Tp2> operator Point3_<_Tp2>() const;
-    //! conversion to cv::Vec<>
-    operator Vec<_Tp, 3>() const;
-
-    //! dot product
-    _Tp dot(const Point3_& pt) const;
-    //! dot product computed in double-precision arithmetics
-    double ddot(const Point3_& pt) const;
-    //! cross product of the 2 3D points
-    Point3_ cross(const Point3_& pt) const;
-    _Tp x; //!< x coordinate of the 3D point
-    _Tp y; //!< y coordinate of the 3D point
-    _Tp z; //!< z coordinate of the 3D point
-};
-
-typedef Point3_<int> Point3i;
-typedef Point3_<float> Point3f;
-typedef Point3_<double> Point3d;
-
-template<typename _Tp> class DataType< Point3_<_Tp> >
-{
-public:
-    typedef Point3_<_Tp>                               value_type;
-    typedef Point3_<typename DataType<_Tp>::work_type> work_type;
-    typedef _Tp                                        channel_type;
-
-    enum { generic_type = 0,
-           channels     = 3,
-           fmt          = traits::SafeFmt<channel_type>::fmt + ((channels - 1) << 8)
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-           ,depth        = DataType<channel_type>::depth
-           ,type         = CV_MAKETYPE(depth, channels)
-#endif
-         };
-
-    typedef Vec<channel_type, channels> vec_type;
-};
-
-namespace traits {
-template<typename _Tp>
-struct Depth< Point3_<_Tp> > { enum { value = Depth<_Tp>::value }; };
-template<typename _Tp>
-struct Type< Point3_<_Tp> > { enum { value = CV_MAKETYPE(Depth<_Tp>::value, 3) }; };
-} // namespace
-
-//////////////////////////////// Size_ ////////////////////////////////
-
-/** @brief Template class for specifying the size of an image or rectangle.
-
-The class includes two members called width and height. The structure can be converted to and from
-the old OpenCV structures CvSize and CvSize2D32f . The same set of arithmetic and comparison
-operations as for Point_ is available.
-
-OpenCV defines the following Size_\<\> aliases:
-@code
-    typedef Size_<int> Size2i;
-    typedef Size2i Size;
-    typedef Size_<float> Size2f;
-@endcode
-*/
-template<typename _Tp> class Size_
-{
-public:
-    typedef _Tp value_type;
-
-    //! default constructor
-    Size_();
-    Size_(_Tp _width, _Tp _height);
-#if OPENCV_ABI_COMPATIBILITY < 500
-    Size_(const Size_& sz) = default;
-    Size_(Size_&& sz) CV_NOEXCEPT = default;
-#endif
-    Size_(const Point_<_Tp>& pt);
-
-#if OPENCV_ABI_COMPATIBILITY < 500
-    Size_& operator = (const Size_& sz) = default;
-    Size_& operator = (Size_&& sz) CV_NOEXCEPT = default;
-#endif
-    //! the area (width*height)
-    _Tp area() const;
-    //! aspect ratio (width/height)
-    double aspectRatio() const;
-    //! true if empty
-    bool empty() const;
-
-    //! conversion of another data type.
-    template<typename _Tp2> operator Size_<_Tp2>() const;
-
-    _Tp width; //!< the width
-    _Tp height; //!< the height
-};
-
-typedef Size_<int> Size2i;
-typedef Size_<int64> Size2l;
-typedef Size_<float> Size2f;
-typedef Size_<double> Size2d;
-typedef Size2i Size;
-
-template<typename _Tp> class DataType< Size_<_Tp> >
-{
-public:
-    typedef Size_<_Tp>                               value_type;
-    typedef Size_<typename DataType<_Tp>::work_type> work_type;
-    typedef _Tp                                      channel_type;
-
-    enum { generic_type = 0,
-           channels     = 2,
-           fmt          = DataType<channel_type>::fmt + ((channels - 1) << 8)
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-           ,depth        = DataType<channel_type>::depth
-           ,type         = CV_MAKETYPE(depth, channels)
-#endif
-         };
-
-    typedef Vec<channel_type, channels> vec_type;
-};
-
-namespace traits {
-template<typename _Tp>
-struct Depth< Size_<_Tp> > { enum { value = Depth<_Tp>::value }; };
-template<typename _Tp>
-struct Type< Size_<_Tp> > { enum { value = CV_MAKETYPE(Depth<_Tp>::value, 2) }; };
-} // namespace
-
-//////////////////////////////// Rect_ ////////////////////////////////
-
-/** @brief Template class for 2D rectangles
-
-described by the following parameters:
--   Coordinates of the top-left corner. This is a default interpretation of Rect_::x and Rect_::y
-    in OpenCV. Though, in your algorithms you may count x and y from the bottom-left corner.
--   Rectangle width and height.
-
-OpenCV typically assumes that the top and left boundary of the rectangle are inclusive, while the
-right and bottom boundaries are not. For example, the method Rect_::contains returns true if
-
-\f[x  \leq pt.x < x+width,
-      y  \leq pt.y < y+height\f]
-
-Virtually every loop over an image ROI in OpenCV (where ROI is specified by Rect_\<int\> ) is
-implemented as:
-@code
-    for(int y = roi.y; y < roi.y + roi.height; y++)
-        for(int x = roi.x; x < roi.x + roi.width; x++)
-        {
-            // ...
-        }
-@endcode
-In addition to the class members, the following operations on rectangles are implemented:
--   \f$\texttt{rect} = \texttt{rect} \pm \texttt{point}\f$ (shifting a rectangle by a certain offset)
--   \f$\texttt{rect} = \texttt{rect} \pm \texttt{size}\f$ (expanding or shrinking a rectangle by a
-    certain amount)
--   rect += point, rect -= point, rect += size, rect -= size (augmenting operations)
--   rect = rect1 & rect2 (rectangle intersection)
--   rect = rect1 | rect2 (minimum area rectangle containing rect1 and rect2 )
--   rect &= rect1, rect |= rect1 (and the corresponding augmenting operations)
--   rect == rect1, rect != rect1 (rectangle comparison)
-
-This is an example how the partial ordering on rectangles can be established (rect1 \f$\subseteq\f$
-rect2):
-@code
-    template<typename _Tp> inline bool
-    operator <= (const Rect_<_Tp>& r1, const Rect_<_Tp>& r2)
-    {
-        return (r1 & r2) == r1;
-    }
-@endcode
-For your convenience, the Rect_\<\> alias is available: cv::Rect
-*/
-template<typename _Tp> class Rect_
-{
-public:
-    typedef _Tp value_type;
-
-    //! default constructor
-    Rect_();
-    Rect_(_Tp _x, _Tp _y, _Tp _width, _Tp _height);
-#if OPENCV_ABI_COMPATIBILITY < 500
-    Rect_(const Rect_& r) = default;
-    Rect_(Rect_&& r) CV_NOEXCEPT = default;
-#endif
-    Rect_(const Point_<_Tp>& org, const Size_<_Tp>& sz);
-    Rect_(const Point_<_Tp>& pt1, const Point_<_Tp>& pt2);
-
-#if OPENCV_ABI_COMPATIBILITY < 500
-    Rect_& operator = (const Rect_& r) = default;
-    Rect_& operator = (Rect_&& r) CV_NOEXCEPT = default;
-#endif
-    //! the top-left corner
-    Point_<_Tp> tl() const;
-    //! the bottom-right corner
-    Point_<_Tp> br() const;
-
-    //! size (width, height) of the rectangle
-    Size_<_Tp> size() const;
-    //! area (width*height) of the rectangle
-    _Tp area() const;
-    //! true if empty
-    bool empty() const;
-
-    //! conversion to another data type
-    template<typename _Tp2> operator Rect_<_Tp2>() const;
-
-    //! checks whether the rectangle contains the point
-    bool contains(const Point_<_Tp>& pt) const;
-
-    _Tp x; //!< x coordinate of the top-left corner
-    _Tp y; //!< y coordinate of the top-left corner
-    _Tp width; //!< width of the rectangle
-    _Tp height; //!< height of the rectangle
-};
-
-typedef Rect_<int> Rect2i;
-typedef Rect_<float> Rect2f;
-typedef Rect_<double> Rect2d;
-typedef Rect2i Rect;
-
-template<typename _Tp> class DataType< Rect_<_Tp> >
-{
-public:
-    typedef Rect_<_Tp>                               value_type;
-    typedef Rect_<typename DataType<_Tp>::work_type> work_type;
-    typedef _Tp                                      channel_type;
-
-    enum { generic_type = 0,
-           channels     = 4,
-           fmt          = traits::SafeFmt<channel_type>::fmt + ((channels - 1) << 8)
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-           ,depth        = DataType<channel_type>::depth
-           ,type         = CV_MAKETYPE(depth, channels)
-#endif
-         };
-
-    typedef Vec<channel_type, channels> vec_type;
-};
-
-namespace traits {
-template<typename _Tp>
-struct Depth< Rect_<_Tp> > { enum { value = Depth<_Tp>::value }; };
-template<typename _Tp>
-struct Type< Rect_<_Tp> > { enum { value = CV_MAKETYPE(Depth<_Tp>::value, 4) }; };
-} // namespace
-
-///////////////////////////// RotatedRect /////////////////////////////
-
-/** @brief The class represents rotated (i.e. not up-right) rectangles on a plane.
-
-Each rectangle is specified by the center point (mass center), length of each side (represented by
-#Size2f structure) and the rotation angle in degrees.
-
-The sample below demonstrates how to use RotatedRect:
-@snippet snippets/core_various.cpp RotatedRect_demo
-![image](pics/rotatedrect.png)
-
-@sa CamShift, fitEllipse, minAreaRect, CvBox2D
-*/
-class CV_EXPORTS RotatedRect
-{
-public:
-    //! default constructor
-    RotatedRect();
-    /** full constructor
-    @param center The rectangle mass center.
-    @param size Width and height of the rectangle.
-    @param angle The rotation angle in a clockwise direction. When the angle is 0, 90, 180, 270 etc.,
-    the rectangle becomes an up-right rectangle.
-    */
-    RotatedRect(const Point2f& center, const Size2f& size, float angle);
-    /**
-    Any 3 end points of the RotatedRect. They must be given in order (either clockwise or
-    anticlockwise).
-     */
-    RotatedRect(const Point2f& point1, const Point2f& point2, const Point2f& point3);
-
-    /** returns 4 vertices of the rectangle
-    @param pts The points array for storing rectangle vertices. The order is bottomLeft, topLeft, topRight, bottomRight.
-    */
-    void points(Point2f pts[]) const;
-    //! returns the minimal up-right integer rectangle containing the rotated rectangle
-    Rect boundingRect() const;
-    //! returns the minimal (exact) floating point rectangle containing the rotated rectangle, not intended for use with images
-    Rect_<float> boundingRect2f() const;
-    //! returns the rectangle mass center
-    Point2f center;
-    //! returns width and height of the rectangle
-    Size2f size;
-    //! returns the rotation angle. When the angle is 0, 90, 180, 270 etc., the rectangle becomes an up-right rectangle.
-    float angle;
-};
-
-template<> class DataType< RotatedRect >
-{
-public:
-    typedef RotatedRect  value_type;
-    typedef value_type   work_type;
-    typedef float        channel_type;
-
-    enum { generic_type = 0,
-           channels     = (int)sizeof(value_type)/sizeof(channel_type), // 5
-           fmt          = traits::SafeFmt<channel_type>::fmt + ((channels - 1) << 8)
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-           ,depth        = DataType<channel_type>::depth
-           ,type         = CV_MAKETYPE(depth, channels)
-#endif
-         };
-
-    typedef Vec<channel_type, channels> vec_type;
-};
-
-namespace traits {
-template<>
-struct Depth< RotatedRect > { enum { value = Depth<float>::value }; };
-template<>
-struct Type< RotatedRect > { enum { value = CV_MAKETYPE(Depth<float>::value, (int)sizeof(RotatedRect)/sizeof(float)) }; };
-} // namespace
-
-
-//////////////////////////////// Range /////////////////////////////////
-
-/** @brief Template class specifying a continuous subsequence (slice) of a sequence.
-
-The class is used to specify a row or a column span in a matrix ( Mat ) and for many other purposes.
-Range(a,b) is basically the same as a:b in Matlab or a..b in Python. As in Python, start is an
-inclusive left boundary of the range and end is an exclusive right boundary of the range. Such a
-half-opened interval is usually denoted as \f$[start,end)\f$ .
-
-The static method Range::all() returns a special variable that means "the whole sequence" or "the
-whole range", just like " : " in Matlab or " ... " in Python. All the methods and functions in
-OpenCV that take Range support this special Range::all() value. But, of course, in case of your own
-custom processing, you will probably have to check and handle it explicitly:
-@code
-    void my_function(..., const Range& r, ....)
-    {
-        if(r == Range::all()) {
-            // process all the data
-        }
-        else {
-            // process [r.start, r.end)
-        }
-    }
-@endcode
-*/
-class CV_EXPORTS Range
-{
-public:
-    Range();
-    Range(int _start, int _end);
-    int size() const;
-    bool empty() const;
-    static Range all();
-
-    int start, end;
-};
-
-template<> class DataType<Range>
-{
-public:
-    typedef Range      value_type;
-    typedef value_type work_type;
-    typedef int        channel_type;
-
-    enum { generic_type = 0,
-           channels     = 2,
-           fmt          = traits::SafeFmt<channel_type>::fmt + ((channels - 1) << 8)
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-           ,depth        = DataType<channel_type>::depth
-           ,type         = CV_MAKETYPE(depth, channels)
-#endif
-         };
-
-    typedef Vec<channel_type, channels> vec_type;
-};
-
-namespace traits {
-template<>
-struct Depth< Range > { enum { value = Depth<int>::value }; };
-template<>
-struct Type< Range > { enum { value = CV_MAKETYPE(Depth<int>::value, 2) }; };
-} // namespace
-
-
-//////////////////////////////// Scalar_ ///////////////////////////////
-
-/** @brief Template class for a 4-element vector derived from Vec.
-
-Being derived from Vec\<_Tp, 4\> , Scalar\_ and Scalar can be used just as typical 4-element
-vectors. In addition, they can be converted to/from CvScalar . The type Scalar is widely used in
-OpenCV to pass pixel values.
-*/
-template<typename _Tp> class Scalar_ : public Vec<_Tp, 4>
-{
-public:
-    //! default constructor
-    Scalar_();
-    Scalar_(_Tp v0, _Tp v1, _Tp v2=0, _Tp v3=0);
-    Scalar_(_Tp v0);
-
-    Scalar_(const Scalar_& s);
-    Scalar_(Scalar_&& s) CV_NOEXCEPT;
-
-    Scalar_& operator=(const Scalar_& s);
-    Scalar_& operator=(Scalar_&& s) CV_NOEXCEPT;
-
-    template<typename _Tp2, int cn>
-    Scalar_(const Vec<_Tp2, cn>& v);
-
-    //! returns a scalar with all elements set to v0
-    static Scalar_<_Tp> all(_Tp v0);
-
-    //! conversion to another data type
-    template<typename T2> operator Scalar_<T2>() const;
-
-    //! per-element product
-    Scalar_<_Tp> mul(const Scalar_<_Tp>& a, double scale=1 ) const;
-
-    //! returns (v0, -v1, -v2, -v3)
-    Scalar_<_Tp> conj() const;
-
-    //! returns true iff v1 == v2 == v3 == 0
-    bool isReal() const;
-};
-
-typedef Scalar_<double> Scalar;
-
-template<typename _Tp> class DataType< Scalar_<_Tp> >
-{
-public:
-    typedef Scalar_<_Tp>                               value_type;
-    typedef Scalar_<typename DataType<_Tp>::work_type> work_type;
-    typedef _Tp                                        channel_type;
-
-    enum { generic_type = 0,
-           channels     = 4,
-           fmt          = traits::SafeFmt<channel_type>::fmt + ((channels - 1) << 8)
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-           ,depth        = DataType<channel_type>::depth
-           ,type         = CV_MAKETYPE(depth, channels)
-#endif
-         };
-
-    typedef Vec<channel_type, channels> vec_type;
-};
-
-namespace traits {
-template<typename _Tp>
-struct Depth< Scalar_<_Tp> > { enum { value = Depth<_Tp>::value }; };
-template<typename _Tp>
-struct Type< Scalar_<_Tp> > { enum { value = CV_MAKETYPE(Depth<_Tp>::value, 4) }; };
-} // namespace
-
-
-/////////////////////////////// KeyPoint ////////////////////////////////
-
-/** @brief Data structure for salient point detectors.
-
-The class instance stores a keypoint, i.e. a point feature found by one of many available keypoint
-detectors, such as Harris corner detector, #FAST, %StarDetector, %SURF, %SIFT etc.
-
-The keypoint is characterized by the 2D position, scale (proportional to the diameter of the
-neighborhood that needs to be taken into account), orientation and some other parameters. The
-keypoint neighborhood is then analyzed by another algorithm that builds a descriptor (usually
-represented as a feature vector). The keypoints representing the same object in different images
-can then be matched using %KDTree or another method.
-*/
-class CV_EXPORTS_W_SIMPLE KeyPoint
-{
-public:
-    //! the default constructor
-    CV_WRAP KeyPoint();
-    /**
-    @param pt x & y coordinates of the keypoint
-    @param size keypoint diameter
-    @param angle keypoint orientation
-    @param response keypoint detector response on the keypoint (that is, strength of the keypoint)
-    @param octave pyramid octave in which the keypoint has been detected
-    @param class_id object id
-     */
-    KeyPoint(Point2f pt, float size, float angle=-1, float response=0, int octave=0, int class_id=-1);
-    /**
-    @param x x-coordinate of the keypoint
-    @param y y-coordinate of the keypoint
-    @param size keypoint diameter
-    @param angle keypoint orientation
-    @param response keypoint detector response on the keypoint (that is, strength of the keypoint)
-    @param octave pyramid octave in which the keypoint has been detected
-    @param class_id object id
-     */
-    CV_WRAP KeyPoint(float x, float y, float size, float angle=-1, float response=0, int octave=0, int class_id=-1);
-
-    size_t hash() const;
-
-    /**
-    This method converts vector of keypoints to vector of points or the reverse, where each keypoint is
-    assigned the same size and the same orientation.
-
-    @param keypoints Keypoints obtained from any feature detection algorithm like SIFT/SURF/ORB
-    @param points2f Array of (x,y) coordinates of each keypoint
-    @param keypointIndexes Array of indexes of keypoints to be converted to points. (Acts like a mask to
-    convert only specified keypoints)
-    */
-    CV_WRAP static void convert(const std::vector<KeyPoint>& keypoints,
-                                CV_OUT std::vector<Point2f>& points2f,
-                                const std::vector<int>& keypointIndexes=std::vector<int>());
-    /** @overload
-    @param points2f Array of (x,y) coordinates of each keypoint
-    @param keypoints Keypoints obtained from any feature detection algorithm like SIFT/SURF/ORB
-    @param size keypoint diameter
-    @param response keypoint detector response on the keypoint (that is, strength of the keypoint)
-    @param octave pyramid octave in which the keypoint has been detected
-    @param class_id object id
-    */
-    CV_WRAP static void convert(const std::vector<Point2f>& points2f,
-                                CV_OUT std::vector<KeyPoint>& keypoints,
-                                float size=1, float response=1, int octave=0, int class_id=-1);
-
-    /**
-    This method computes overlap for pair of keypoints. Overlap is the ratio between area of keypoint
-    regions' intersection and area of keypoint regions' union (considering keypoint region as circle).
-    If they don't overlap, we get zero. If they coincide at same location with same size, we get 1.
-    @param kp1 First keypoint
-    @param kp2 Second keypoint
-    */
-    CV_WRAP static float overlap(const KeyPoint& kp1, const KeyPoint& kp2);
-
-    CV_PROP_RW Point2f pt; //!< coordinates of the keypoints
-    CV_PROP_RW float size; //!< diameter of the meaningful keypoint neighborhood
-    CV_PROP_RW float angle; //!< computed orientation of the keypoint (-1 if not applicable);
-                            //!< it's in [0,360) degrees and measured relative to
-                            //!< image coordinate system, ie in clockwise.
-    CV_PROP_RW float response; //!< the response by which the most strong keypoints have been selected. Can be used for the further sorting or subsampling
-    CV_PROP_RW int octave; //!< octave (pyramid layer) from which the keypoint has been extracted
-    CV_PROP_RW int class_id; //!< object class (if the keypoints need to be clustered by an object they belong to)
-};
-
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-template<> class DataType<KeyPoint>
-{
-public:
-    typedef KeyPoint      value_type;
-    typedef float         work_type;
-    typedef float         channel_type;
-
-    enum { generic_type = 0,
-           depth        = DataType<channel_type>::depth,
-           channels     = (int)(sizeof(value_type)/sizeof(channel_type)), // 7
-           fmt          = DataType<channel_type>::fmt + ((channels - 1) << 8),
-           type         = CV_MAKETYPE(depth, channels)
-         };
-
-    typedef Vec<channel_type, channels> vec_type;
-};
-#endif
-
-
-//////////////////////////////// DMatch /////////////////////////////////
-
-/** @brief Class for matching keypoint descriptors
-
-query descriptor index, train descriptor index, train image index, and distance between
-descriptors.
-*/
-class CV_EXPORTS_W_SIMPLE DMatch
-{
-public:
-    CV_WRAP DMatch();
-    CV_WRAP DMatch(int _queryIdx, int _trainIdx, float _distance);
-    CV_WRAP DMatch(int _queryIdx, int _trainIdx, int _imgIdx, float _distance);
-
-    CV_PROP_RW int queryIdx; //!< query descriptor index
-    CV_PROP_RW int trainIdx; //!< train descriptor index
-    CV_PROP_RW int imgIdx;   //!< train image index
-
-    CV_PROP_RW float distance;
-
-    // less is better
-    bool operator<(const DMatch &m) const;
-};
-
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-template<> class DataType<DMatch>
-{
-public:
-    typedef DMatch      value_type;
-    typedef int         work_type;
-    typedef int         channel_type;
-
-    enum { generic_type = 0,
-           depth        = DataType<channel_type>::depth,
-           channels     = (int)(sizeof(value_type)/sizeof(channel_type)), // 4
-           fmt          = DataType<channel_type>::fmt + ((channels - 1) << 8),
-           type         = CV_MAKETYPE(depth, channels)
-         };
-
-    typedef Vec<channel_type, channels> vec_type;
-};
-#endif
-
-
-///////////////////////////// TermCriteria //////////////////////////////
-
-/** @brief The class defining termination criteria for iterative algorithms.
-
-You can initialize it by default constructor and then override any parameters, or the structure may
-be fully initialized using the advanced variant of the constructor.
-*/
-class CV_EXPORTS TermCriteria
-{
-public:
-    /**
-      Criteria type, can be one of: COUNT, EPS or COUNT + EPS
-    */
-    enum Type
-    {
-        COUNT=1, //!< the maximum number of iterations or elements to compute
-        MAX_ITER=COUNT, //!< ditto
-        EPS=2 //!< the desired accuracy or change in parameters at which the iterative algorithm stops
-    };
-
-    //! default constructor
-    TermCriteria();
-    /**
-    @param type The type of termination criteria, one of TermCriteria::Type
-    @param maxCount The maximum number of iterations or elements to compute.
-    @param epsilon The desired accuracy or change in parameters at which the iterative algorithm stops.
-    */
-    TermCriteria(int type, int maxCount, double epsilon);
-
-    inline bool isValid() const
-    {
-        const bool isCount = (type & COUNT) && maxCount > 0;
-        const bool isEps = (type & EPS) && !cvIsNaN(epsilon);
-        return isCount || isEps;
-    }
-
-    int type; //!< the type of termination criteria: COUNT, EPS or COUNT + EPS
-    int maxCount; //!< the maximum number of iterations/elements
-    double epsilon; //!< the desired accuracy
-};
-
-
-//! @} core_basic
-
-///////////////////////// raster image moments //////////////////////////
-
-//! @addtogroup imgproc_shape
-//! @{
-
-/** @brief struct returned by cv::moments
-
-The spatial moments \f$\texttt{Moments::m}_{ji}\f$ are computed as:
-
-\f[\texttt{m} _{ji}= \sum _{x,y}  \left ( \texttt{array} (x,y)  \cdot x^j  \cdot y^i \right )\f]
-
-The central moments \f$\texttt{Moments::mu}_{ji}\f$ are computed as:
-
-\f[\texttt{mu} _{ji}= \sum _{x,y}  \left ( \texttt{array} (x,y)  \cdot (x -  \bar{x} )^j  \cdot (y -  \bar{y} )^i \right )\f]
-
-where \f$(\bar{x}, \bar{y})\f$ is the mass center:
-
-\f[\bar{x} = \frac{\texttt{m}_{10}}{\texttt{m}_{00}} , \; \bar{y} = \frac{\texttt{m}_{01}}{\texttt{m}_{00}}\f]
-
-The normalized central moments \f$\texttt{Moments::nu}_{ij}\f$ are computed as:
-
-\f[\texttt{nu} _{ji}= \frac{\texttt{mu}_{ji}}{\texttt{m}_{00}^{(i+j)/2+1}} .\f]
-
-@note
-\f$\texttt{mu}_{00}=\texttt{m}_{00}\f$, \f$\texttt{nu}_{00}=1\f$
-\f$\texttt{nu}_{10}=\texttt{mu}_{10}=\texttt{mu}_{01}=\texttt{mu}_{10}=0\f$ , hence the values are not
-stored.
-
-The moments of a contour are defined in the same way but computed using the Green's formula (see
-<http://en.wikipedia.org/wiki/Green_theorem>). So, due to a limited raster resolution, the moments
-computed for a contour are slightly different from the moments computed for the same rasterized
-contour.
-
-@note
-Since the contour moments are computed using Green formula, you may get seemingly odd results for
-contours with self-intersections, e.g. a zero area (m00) for butterfly-shaped contours.
- */
-class CV_EXPORTS_W_MAP Moments
-{
-public:
-    //! the default constructor
-    Moments();
-    //! the full constructor
-    Moments(double m00, double m10, double m01, double m20, double m11,
-            double m02, double m30, double m21, double m12, double m03 );
-    ////! the conversion from CvMoments
-    //Moments( const CvMoments& moments );
-    ////! the conversion to CvMoments
-    //operator CvMoments() const;
-
-    //! @name spatial moments
-    //! @{
-    CV_PROP_RW double  m00, m10, m01, m20, m11, m02, m30, m21, m12, m03;
-    //! @}
-
-    //! @name central moments
-    //! @{
-    CV_PROP_RW double  mu20, mu11, mu02, mu30, mu21, mu12, mu03;
-    //! @}
-
-    //! @name central normalized moments
-    //! @{
-    CV_PROP_RW double  nu20, nu11, nu02, nu30, nu21, nu12, nu03;
-    //! @}
-};
-
-template<> class DataType<Moments>
-{
-public:
-    typedef Moments     value_type;
-    typedef double      work_type;
-    typedef double      channel_type;
-
-    enum { generic_type = 0,
-           channels     = (int)(sizeof(value_type)/sizeof(channel_type)), // 24
-           fmt          = DataType<channel_type>::fmt + ((channels - 1) << 8)
-#ifdef OPENCV_TRAITS_ENABLE_DEPRECATED
-           ,depth        = DataType<channel_type>::depth
-           ,type         = CV_MAKETYPE(depth, channels)
-#endif
-         };
-
-    typedef Vec<channel_type, channels> vec_type;
-};
-
-namespace traits {
-template<>
-struct Depth< Moments > { enum { value = Depth<double>::value }; };
-template<>
-struct Type< Moments > { enum { value = CV_MAKETYPE(Depth<double>::value, (int)(sizeof(Moments)/sizeof(double))) }; };
-} // namespace
-
-//! @} imgproc_shape
-
-//! @cond IGNORED
-
-/////////////////////////////////////////////////////////////////////////
-///////////////////////////// Implementation ////////////////////////////
-/////////////////////////////////////////////////////////////////////////
-
-//////////////////////////////// Complex ////////////////////////////////
-
-template<typename _Tp> inline
-Complex<_Tp>::Complex()
-    : re(0), im(0) {}
-
-template<typename _Tp> inline
-Complex<_Tp>::Complex( _Tp _re, _Tp _im )
-    : re(_re), im(_im) {}
-
-template<typename _Tp> template<typename T2> inline
-Complex<_Tp>::operator Complex<T2>() const
-{
-    return Complex<T2>(saturate_cast<T2>(re), saturate_cast<T2>(im));
-}
-
-template<typename _Tp> inline
-Complex<_Tp> Complex<_Tp>::conj() const
-{
-    return Complex<_Tp>(re, -im);
-}
-
-
-template<typename _Tp> static inline
-bool operator == (const Complex<_Tp>& a, const Complex<_Tp>& b)
-{
-    return a.re == b.re && a.im == b.im;
-}
-
-template<typename _Tp> static inline
-bool operator != (const Complex<_Tp>& a, const Complex<_Tp>& b)
-{
-    return a.re != b.re || a.im != b.im;
-}
-
-template<typename _Tp> static inline
-Complex<_Tp> operator + (const Complex<_Tp>& a, const Complex<_Tp>& b)
-{
-    return Complex<_Tp>( a.re + b.re, a.im + b.im );
-}
-
-template<typename _Tp> static inline
-Complex<_Tp>& operator += (Complex<_Tp>& a, const Complex<_Tp>& b)
-{
-    a.re += b.re; a.im += b.im;
-    return a;
-}
-
-template<typename _Tp> static inline
-Complex<_Tp> operator - (const Complex<_Tp>& a, const Complex<_Tp>& b)
-{
-    return Complex<_Tp>( a.re - b.re, a.im - b.im );
-}
-
-template<typename _Tp> static inline
-Complex<_Tp>& operator -= (Complex<_Tp>& a, const Complex<_Tp>& b)
-{
-    a.re -= b.re; a.im -= b.im;
-    return a;
-}
-
-template<typename _Tp> static inline
-Complex<_Tp> operator - (const Complex<_Tp>& a)
-{
-    return Complex<_Tp>(-a.re, -a.im);
-}
-
-template<typename _Tp> static inline
-Complex<_Tp> operator * (const Complex<_Tp>& a, const Complex<_Tp>& b)
-{
-    return Complex<_Tp>( a.re*b.re - a.im*b.im, a.re*b.im + a.im*b.re );
-}
-
-template<typename _Tp> static inline
-Complex<_Tp> operator * (const Complex<_Tp>& a, _Tp b)
-{
-    return Complex<_Tp>( a.re*b, a.im*b );
-}
-
-template<typename _Tp> static inline
-Complex<_Tp> operator * (_Tp b, const Complex<_Tp>& a)
-{
-    return Complex<_Tp>( a.re*b, a.im*b );
-}
-
-template<typename _Tp> static inline
-Complex<_Tp> operator + (const Complex<_Tp>& a, _Tp b)
-{
-    return Complex<_Tp>( a.re + b, a.im );
-}
-
-template<typename _Tp> static inline
-Complex<_Tp> operator - (const Complex<_Tp>& a, _Tp b)
-{ return Complex<_Tp>( a.re - b, a.im ); }
-
-template<typename _Tp> static inline
-Complex<_Tp> operator + (_Tp b, const Complex<_Tp>& a)
-{
-    return Complex<_Tp>( a.re + b, a.im );
-}
-
-template<typename _Tp> static inline
-Complex<_Tp> operator - (_Tp b, const Complex<_Tp>& a)
-{
-    return Complex<_Tp>( b - a.re, -a.im );
-}
-
-template<typename _Tp> static inline
-Complex<_Tp>& operator += (Complex<_Tp>& a, _Tp b)
-{
-    a.re += b; return a;
-}
-
-template<typename _Tp> static inline
-Complex<_Tp>& operator -= (Complex<_Tp>& a, _Tp b)
-{
-    a.re -= b; return a;
-}
-
-template<typename _Tp> static inline
-Complex<_Tp>& operator *= (Complex<_Tp>& a, _Tp b)
-{
-    a.re *= b; a.im *= b; return a;
-}
-
-template<typename _Tp> static inline
-double abs(const Complex<_Tp>& a)
-{
-    return std::sqrt( (double)a.re*a.re + (double)a.im*a.im);
-}
-
-template<typename _Tp> static inline
-Complex<_Tp> operator / (const Complex<_Tp>& a, const Complex<_Tp>& b)
-{
-    double t = 1./((double)b.re*b.re + (double)b.im*b.im);
-    return Complex<_Tp>( (_Tp)((a.re*b.re + a.im*b.im)*t),
-                        (_Tp)((-a.re*b.im + a.im*b.re)*t) );
-}
-
-template<typename _Tp> static inline
-Complex<_Tp>& operator /= (Complex<_Tp>& a, const Complex<_Tp>& b)
-{
-    a = a / b;
-    return a;
-}
-
-template<typename _Tp> static inline
-Complex<_Tp> operator / (const Complex<_Tp>& a, _Tp b)
-{
-    _Tp t = (_Tp)1/b;
-    return Complex<_Tp>( a.re*t, a.im*t );
-}
-
-template<typename _Tp> static inline
-Complex<_Tp> operator / (_Tp b, const Complex<_Tp>& a)
-{
-    return Complex<_Tp>(b)/a;
-}
-
-template<typename _Tp> static inline
-Complex<_Tp> operator /= (const Complex<_Tp>& a, _Tp b)
-{
-    _Tp t = (_Tp)1/b;
-    a.re *= t; a.im *= t; return a;
-}
-
-
-
-//////////////////////////////// 2D Point ///////////////////////////////
-
-template<typename _Tp> inline
-Point_<_Tp>::Point_()
-    : x(0), y(0) {}
-
-template<typename _Tp> inline
-Point_<_Tp>::Point_(_Tp _x, _Tp _y)
-    : x(_x), y(_y) {}
-
-#if (defined(__GNUC__) && __GNUC__ < 5) && !defined(__clang__)  // GCC 4.x bug. Details: https://github.com/opencv/opencv/pull/20837
-template<typename _Tp> inline
-Point_<_Tp>::Point_(const Point_& pt)
-    : x(pt.x), y(pt.y) {}
-#endif
-
-template<typename _Tp> inline
-Point_<_Tp>::Point_(const Size_<_Tp>& sz)
-    : x(sz.width), y(sz.height) {}
-
-template<typename _Tp> inline
-Point_<_Tp>::Point_(const Vec<_Tp,2>& v)
-    : x(v[0]), y(v[1]) {}
-
-#if (defined(__GNUC__) && __GNUC__ < 5) && !defined(__clang__)  // GCC 4.x bug. Details: https://github.com/opencv/opencv/pull/20837
-template<typename _Tp> inline
-Point_<_Tp>& Point_<_Tp>::operator = (const Point_& pt)
-{
-    x = pt.x; y = pt.y;
-    return *this;
-}
-#endif
-
-template<typename _Tp> template<typename _Tp2> inline
-Point_<_Tp>::operator Point_<_Tp2>() const
-{
-    return Point_<_Tp2>(saturate_cast<_Tp2>(x), saturate_cast<_Tp2>(y));
-}
-
-template<typename _Tp> inline
-Point_<_Tp>::operator Vec<_Tp, 2>() const
-{
-    return Vec<_Tp, 2>(x, y);
-}
-
-template<typename _Tp> inline
-_Tp Point_<_Tp>::dot(const Point_& pt) const
-{
-    return saturate_cast<_Tp>(x*pt.x + y*pt.y);
-}
-
-template<typename _Tp> inline
-double Point_<_Tp>::ddot(const Point_& pt) const
-{
-    return (double)x*(double)(pt.x) + (double)y*(double)(pt.y);
-}
-
-template<typename _Tp> inline
-double Point_<_Tp>::cross(const Point_& pt) const
-{
-    return (double)x*pt.y - (double)y*pt.x;
-}
-
-template<typename _Tp> inline bool
-Point_<_Tp>::inside( const Rect_<_Tp>& r ) const
-{
-    return r.contains(*this);
-}
-
-
-template<typename _Tp> static inline
-Point_<_Tp>& operator += (Point_<_Tp>& a, const Point_<_Tp>& b)
-{
-    a.x += b.x;
-    a.y += b.y;
-    return a;
-}
-
-template<typename _Tp> static inline
-Point_<_Tp>& operator -= (Point_<_Tp>& a, const Point_<_Tp>& b)
-{
-    a.x -= b.x;
-    a.y -= b.y;
-    return a;
-}
-
-template<typename _Tp> static inline
-Point_<_Tp>& operator *= (Point_<_Tp>& a, int b)
-{
-    a.x = saturate_cast<_Tp>(a.x * b);
-    a.y = saturate_cast<_Tp>(a.y * b);
-    return a;
-}
-
-template<typename _Tp> static inline
-Point_<_Tp>& operator *= (Point_<_Tp>& a, float b)
-{
-    a.x = saturate_cast<_Tp>(a.x * b);
-    a.y = saturate_cast<_Tp>(a.y * b);
-    return a;
-}
-
-template<typename _Tp> static inline
-Point_<_Tp>& operator *= (Point_<_Tp>& a, double b)
-{
-    a.x = saturate_cast<_Tp>(a.x * b);
-    a.y = saturate_cast<_Tp>(a.y * b);
-    return a;
-}
-
-template<typename _Tp> static inline
-Point_<_Tp>& operator /= (Point_<_Tp>& a, int b)
-{
-    a.x = saturate_cast<_Tp>(a.x / b);
-    a.y = saturate_cast<_Tp>(a.y / b);
-    return a;
-}
-
-template<typename _Tp> static inline
-Point_<_Tp>& operator /= (Point_<_Tp>& a, float b)
-{
-    a.x = saturate_cast<_Tp>(a.x / b);
-    a.y = saturate_cast<_Tp>(a.y / b);
-    return a;
-}
-
-template<typename _Tp> static inline
-Point_<_Tp>& operator /= (Point_<_Tp>& a, double b)
-{
-    a.x = saturate_cast<_Tp>(a.x / b);
-    a.y = saturate_cast<_Tp>(a.y / b);
-    return a;
-}
-
-template<typename _Tp> static inline
-double norm(const Point_<_Tp>& pt)
-{
-    return std::sqrt((double)pt.x*pt.x + (double)pt.y*pt.y);
-}
-
-template<typename _Tp> static inline
-bool operator == (const Point_<_Tp>& a, const Point_<_Tp>& b)
-{
-    return a.x == b.x && a.y == b.y;
-}
-
-template<typename _Tp> static inline
-bool operator != (const Point_<_Tp>& a, const Point_<_Tp>& b)
-{
-    return a.x != b.x || a.y != b.y;
-}
-
-template<typename _Tp> static inline
-Point_<_Tp> operator + (const Point_<_Tp>& a, const Point_<_Tp>& b)
-{
-    return Point_<_Tp>( saturate_cast<_Tp>(a.x + b.x), saturate_cast<_Tp>(a.y + b.y) );
-}
-
-template<typename _Tp> static inline
-Point_<_Tp> operator - (const Point_<_Tp>& a, const Point_<_Tp>& b)
-{
-    return Point_<_Tp>( saturate_cast<_Tp>(a.x - b.x), saturate_cast<_Tp>(a.y - b.y) );
-}
-
-template<typename _Tp> static inline
-Point_<_Tp> operator - (const Point_<_Tp>& a)
-{
-    return Point_<_Tp>( saturate_cast<_Tp>(-a.x), saturate_cast<_Tp>(-a.y) );
-}
-
-template<typename _Tp> static inline
-Point_<_Tp> operator * (const Point_<_Tp>& a, int b)
-{
-    return Point_<_Tp>( saturate_cast<_Tp>(a.x*b), saturate_cast<_Tp>(a.y*b) );
-}
-
-template<typename _Tp> static inline
-Point_<_Tp> operator * (int a, const Point_<_Tp>& b)
-{
-    return Point_<_Tp>( saturate_cast<_Tp>(b.x*a), saturate_cast<_Tp>(b.y*a) );
-}
-
-template<typename _Tp> static inline
-Point_<_Tp> operator * (const Point_<_Tp>& a, float b)
-{
-    return Point_<_Tp>( saturate_cast<_Tp>(a.x*b), saturate_cast<_Tp>(a.y*b) );
-}
-
-template<typename _Tp> static inline
-Point_<_Tp> operator * (float a, const Point_<_Tp>& b)
-{
-    return Point_<_Tp>( saturate_cast<_Tp>(b.x*a), saturate_cast<_Tp>(b.y*a) );
-}
-
-template<typename _Tp> static inline
-Point_<_Tp> operator * (const Point_<_Tp>& a, double b)
-{
-    return Point_<_Tp>( saturate_cast<_Tp>(a.x*b), saturate_cast<_Tp>(a.y*b) );
-}
-
-template<typename _Tp> static inline
-Point_<_Tp> operator * (double a, const Point_<_Tp>& b)
-{
-    return Point_<_Tp>( saturate_cast<_Tp>(b.x*a), saturate_cast<_Tp>(b.y*a) );
-}
-
-template<typename _Tp> static inline
-Point_<_Tp> operator * (const Matx<_Tp, 2, 2>& a, const Point_<_Tp>& b)
-{
-    Matx<_Tp, 2, 1> tmp = a * Vec<_Tp,2>(b.x, b.y);
-    return Point_<_Tp>(tmp.val[0], tmp.val[1]);
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp> operator * (const Matx<_Tp, 3, 3>& a, const Point_<_Tp>& b)
-{
-    Matx<_Tp, 3, 1> tmp = a * Vec<_Tp,3>(b.x, b.y, 1);
-    return Point3_<_Tp>(tmp.val[0], tmp.val[1], tmp.val[2]);
-}
-
-template<typename _Tp> static inline
-Point_<_Tp> operator / (const Point_<_Tp>& a, int b)
-{
-    Point_<_Tp> tmp(a);
-    tmp /= b;
-    return tmp;
-}
-
-template<typename _Tp> static inline
-Point_<_Tp> operator / (const Point_<_Tp>& a, float b)
-{
-    Point_<_Tp> tmp(a);
-    tmp /= b;
-    return tmp;
-}
-
-template<typename _Tp> static inline
-Point_<_Tp> operator / (const Point_<_Tp>& a, double b)
-{
-    Point_<_Tp> tmp(a);
-    tmp /= b;
-    return tmp;
-}
-
-
-template<typename _AccTp> static inline _AccTp normL2Sqr(const Point_<int>& pt);
-template<typename _AccTp> static inline _AccTp normL2Sqr(const Point_<int64>& pt);
-template<typename _AccTp> static inline _AccTp normL2Sqr(const Point_<float>& pt);
-template<typename _AccTp> static inline _AccTp normL2Sqr(const Point_<double>& pt);
-
-template<> inline int normL2Sqr<int>(const Point_<int>& pt) { return pt.dot(pt); }
-template<> inline int64 normL2Sqr<int64>(const Point_<int64>& pt) { return pt.dot(pt); }
-template<> inline float normL2Sqr<float>(const Point_<float>& pt) { return pt.dot(pt); }
-template<> inline double normL2Sqr<double>(const Point_<int>& pt) { return pt.dot(pt); }
-
-template<> inline double normL2Sqr<double>(const Point_<float>& pt) { return pt.ddot(pt); }
-template<> inline double normL2Sqr<double>(const Point_<double>& pt) { return pt.ddot(pt); }
-
-
-
-//////////////////////////////// 3D Point ///////////////////////////////
-
-template<typename _Tp> inline
-Point3_<_Tp>::Point3_()
-    : x(0), y(0), z(0) {}
-
-template<typename _Tp> inline
-Point3_<_Tp>::Point3_(_Tp _x, _Tp _y, _Tp _z)
-    : x(_x), y(_y), z(_z) {}
-
-template<typename _Tp> inline
-Point3_<_Tp>::Point3_(const Point_<_Tp>& pt)
-    : x(pt.x), y(pt.y), z(_Tp()) {}
-
-template<typename _Tp> inline
-Point3_<_Tp>::Point3_(const Vec<_Tp, 3>& v)
-    : x(v[0]), y(v[1]), z(v[2]) {}
-
-template<typename _Tp> template<typename _Tp2> inline
-Point3_<_Tp>::operator Point3_<_Tp2>() const
-{
-    return Point3_<_Tp2>(saturate_cast<_Tp2>(x), saturate_cast<_Tp2>(y), saturate_cast<_Tp2>(z));
-}
-
-template<typename _Tp> inline
-Point3_<_Tp>::operator Vec<_Tp, 3>() const
-{
-    return Vec<_Tp, 3>(x, y, z);
-}
-
-template<typename _Tp> inline
-_Tp Point3_<_Tp>::dot(const Point3_& pt) const
-{
-    return saturate_cast<_Tp>(x*pt.x + y*pt.y + z*pt.z);
-}
-
-template<typename _Tp> inline
-double Point3_<_Tp>::ddot(const Point3_& pt) const
-{
-    return (double)x*pt.x + (double)y*pt.y + (double)z*pt.z;
-}
-
-template<typename _Tp> inline
-Point3_<_Tp> Point3_<_Tp>::cross(const Point3_<_Tp>& pt) const
-{
-    return Point3_<_Tp>(y*pt.z - z*pt.y, z*pt.x - x*pt.z, x*pt.y - y*pt.x);
-}
-
-
-template<typename _Tp> static inline
-Point3_<_Tp>& operator += (Point3_<_Tp>& a, const Point3_<_Tp>& b)
-{
-    a.x += b.x;
-    a.y += b.y;
-    a.z += b.z;
-    return a;
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp>& operator -= (Point3_<_Tp>& a, const Point3_<_Tp>& b)
-{
-    a.x -= b.x;
-    a.y -= b.y;
-    a.z -= b.z;
-    return a;
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp>& operator *= (Point3_<_Tp>& a, int b)
-{
-    a.x = saturate_cast<_Tp>(a.x * b);
-    a.y = saturate_cast<_Tp>(a.y * b);
-    a.z = saturate_cast<_Tp>(a.z * b);
-    return a;
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp>& operator *= (Point3_<_Tp>& a, float b)
-{
-    a.x = saturate_cast<_Tp>(a.x * b);
-    a.y = saturate_cast<_Tp>(a.y * b);
-    a.z = saturate_cast<_Tp>(a.z * b);
-    return a;
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp>& operator *= (Point3_<_Tp>& a, double b)
-{
-    a.x = saturate_cast<_Tp>(a.x * b);
-    a.y = saturate_cast<_Tp>(a.y * b);
-    a.z = saturate_cast<_Tp>(a.z * b);
-    return a;
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp>& operator /= (Point3_<_Tp>& a, int b)
-{
-    a.x = saturate_cast<_Tp>(a.x / b);
-    a.y = saturate_cast<_Tp>(a.y / b);
-    a.z = saturate_cast<_Tp>(a.z / b);
-    return a;
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp>& operator /= (Point3_<_Tp>& a, float b)
-{
-    a.x = saturate_cast<_Tp>(a.x / b);
-    a.y = saturate_cast<_Tp>(a.y / b);
-    a.z = saturate_cast<_Tp>(a.z / b);
-    return a;
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp>& operator /= (Point3_<_Tp>& a, double b)
-{
-    a.x = saturate_cast<_Tp>(a.x / b);
-    a.y = saturate_cast<_Tp>(a.y / b);
-    a.z = saturate_cast<_Tp>(a.z / b);
-    return a;
-}
-
-template<typename _Tp> static inline
-double norm(const Point3_<_Tp>& pt)
-{
-    return std::sqrt((double)pt.x*pt.x + (double)pt.y*pt.y + (double)pt.z*pt.z);
-}
-
-template<typename _Tp> static inline
-bool operator == (const Point3_<_Tp>& a, const Point3_<_Tp>& b)
-{
-    return a.x == b.x && a.y == b.y && a.z == b.z;
-}
-
-template<typename _Tp> static inline
-bool operator != (const Point3_<_Tp>& a, const Point3_<_Tp>& b)
-{
-    return a.x != b.x || a.y != b.y || a.z != b.z;
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp> operator + (const Point3_<_Tp>& a, const Point3_<_Tp>& b)
-{
-    return Point3_<_Tp>( saturate_cast<_Tp>(a.x + b.x), saturate_cast<_Tp>(a.y + b.y), saturate_cast<_Tp>(a.z + b.z));
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp> operator - (const Point3_<_Tp>& a, const Point3_<_Tp>& b)
-{
-    return Point3_<_Tp>( saturate_cast<_Tp>(a.x - b.x), saturate_cast<_Tp>(a.y - b.y), saturate_cast<_Tp>(a.z - b.z));
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp> operator - (const Point3_<_Tp>& a)
-{
-    return Point3_<_Tp>( saturate_cast<_Tp>(-a.x), saturate_cast<_Tp>(-a.y), saturate_cast<_Tp>(-a.z) );
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp> operator * (const Point3_<_Tp>& a, int b)
-{
-    return Point3_<_Tp>( saturate_cast<_Tp>(a.x*b), saturate_cast<_Tp>(a.y*b), saturate_cast<_Tp>(a.z*b) );
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp> operator * (int a, const Point3_<_Tp>& b)
-{
-    return Point3_<_Tp>( saturate_cast<_Tp>(b.x * a), saturate_cast<_Tp>(b.y * a), saturate_cast<_Tp>(b.z * a) );
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp> operator * (const Point3_<_Tp>& a, float b)
-{
-    return Point3_<_Tp>( saturate_cast<_Tp>(a.x * b), saturate_cast<_Tp>(a.y * b), saturate_cast<_Tp>(a.z * b) );
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp> operator * (float a, const Point3_<_Tp>& b)
-{
-    return Point3_<_Tp>( saturate_cast<_Tp>(b.x * a), saturate_cast<_Tp>(b.y * a), saturate_cast<_Tp>(b.z * a) );
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp> operator * (const Point3_<_Tp>& a, double b)
-{
-    return Point3_<_Tp>( saturate_cast<_Tp>(a.x * b), saturate_cast<_Tp>(a.y * b), saturate_cast<_Tp>(a.z * b) );
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp> operator * (double a, const Point3_<_Tp>& b)
-{
-    return Point3_<_Tp>( saturate_cast<_Tp>(b.x * a), saturate_cast<_Tp>(b.y * a), saturate_cast<_Tp>(b.z * a) );
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp> operator * (const Matx<_Tp, 3, 3>& a, const Point3_<_Tp>& b)
-{
-    Matx<_Tp, 3, 1> tmp = a * Vec<_Tp,3>(b.x, b.y, b.z);
-    return Point3_<_Tp>(tmp.val[0], tmp.val[1], tmp.val[2]);
-}
-
-template<typename _Tp> static inline
-Matx<_Tp, 4, 1> operator * (const Matx<_Tp, 4, 4>& a, const Point3_<_Tp>& b)
-{
-    return a * Matx<_Tp, 4, 1>(b.x, b.y, b.z, 1);
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp> operator / (const Point3_<_Tp>& a, int b)
-{
-    Point3_<_Tp> tmp(a);
-    tmp /= b;
-    return tmp;
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp> operator / (const Point3_<_Tp>& a, float b)
-{
-    Point3_<_Tp> tmp(a);
-    tmp /= b;
-    return tmp;
-}
-
-template<typename _Tp> static inline
-Point3_<_Tp> operator / (const Point3_<_Tp>& a, double b)
-{
-    Point3_<_Tp> tmp(a);
-    tmp /= b;
-    return tmp;
-}
-
-
-
-////////////////////////////////// Size /////////////////////////////////
-
-template<typename _Tp> inline
-Size_<_Tp>::Size_()
-    : width(0), height(0) {}
-
-template<typename _Tp> inline
-Size_<_Tp>::Size_(_Tp _width, _Tp _height)
-    : width(_width), height(_height) {}
-
-template<typename _Tp> inline
-Size_<_Tp>::Size_(const Point_<_Tp>& pt)
-    : width(pt.x), height(pt.y) {}
-
-template<typename _Tp> template<typename _Tp2> inline
-Size_<_Tp>::operator Size_<_Tp2>() const
-{
-    return Size_<_Tp2>(saturate_cast<_Tp2>(width), saturate_cast<_Tp2>(height));
-}
-
-template<typename _Tp> inline
-_Tp Size_<_Tp>::area() const
-{
-    const _Tp result = width * height;
-    CV_DbgAssert(!std::numeric_limits<_Tp>::is_integer
-        || width == 0 || result / width == height); // make sure the result fits in the return value
-    return result;
-}
-
-template<typename _Tp> inline
-double Size_<_Tp>::aspectRatio() const
-{
-    return width / static_cast<double>(height);
-}
-
-template<typename _Tp> inline
-bool Size_<_Tp>::empty() const
-{
-    return width <= 0 || height <= 0;
-}
-
-
-template<typename _Tp> static inline
-Size_<_Tp>& operator *= (Size_<_Tp>& a, _Tp b)
-{
-    a.width *= b;
-    a.height *= b;
-    return a;
-}
-
-template<typename _Tp> static inline
-Size_<_Tp> operator * (const Size_<_Tp>& a, _Tp b)
-{
-    Size_<_Tp> tmp(a);
-    tmp *= b;
-    return tmp;
-}
-
-template<typename _Tp> static inline
-Size_<_Tp>& operator /= (Size_<_Tp>& a, _Tp b)
-{
-    a.width /= b;
-    a.height /= b;
-    return a;
-}
-
-template<typename _Tp> static inline
-Size_<_Tp> operator / (const Size_<_Tp>& a, _Tp b)
-{
-    Size_<_Tp> tmp(a);
-    tmp /= b;
-    return tmp;
-}
-
-template<typename _Tp> static inline
-Size_<_Tp>& operator += (Size_<_Tp>& a, const Size_<_Tp>& b)
-{
-    a.width += b.width;
-    a.height += b.height;
-    return a;
-}
-
-template<typename _Tp> static inline
-Size_<_Tp> operator + (const Size_<_Tp>& a, const Size_<_Tp>& b)
-{
-    Size_<_Tp> tmp(a);
-    tmp += b;
-    return tmp;
-}
-
-template<typename _Tp> static inline
-Size_<_Tp>& operator -= (Size_<_Tp>& a, const Size_<_Tp>& b)
-{
-    a.width -= b.width;
-    a.height -= b.height;
-    return a;
-}
-
-template<typename _Tp> static inline
-Size_<_Tp> operator - (const Size_<_Tp>& a, const Size_<_Tp>& b)
-{
-    Size_<_Tp> tmp(a);
-    tmp -= b;
-    return tmp;
-}
-
-template<typename _Tp> static inline
-bool operator == (const Size_<_Tp>& a, const Size_<_Tp>& b)
-{
-    return a.width == b.width && a.height == b.height;
-}
-
-template<typename _Tp> static inline
-bool operator != (const Size_<_Tp>& a, const Size_<_Tp>& b)
-{
-    return !(a == b);
-}
-
-
-
-////////////////////////////////// Rect /////////////////////////////////
-
-template<typename _Tp> inline
-Rect_<_Tp>::Rect_()
-    : x(0), y(0), width(0), height(0) {}
-
-template<typename _Tp> inline
-Rect_<_Tp>::Rect_(_Tp _x, _Tp _y, _Tp _width, _Tp _height)
-    : x(_x), y(_y), width(_width), height(_height) {}
-
-template<typename _Tp> inline
-Rect_<_Tp>::Rect_(const Point_<_Tp>& org, const Size_<_Tp>& sz)
-    : x(org.x), y(org.y), width(sz.width), height(sz.height) {}
-
-template<typename _Tp> inline
-Rect_<_Tp>::Rect_(const Point_<_Tp>& pt1, const Point_<_Tp>& pt2)
-{
-    x = std::min(pt1.x, pt2.x);
-    y = std::min(pt1.y, pt2.y);
-    width = std::max(pt1.x, pt2.x) - x;
-    height = std::max(pt1.y, pt2.y) - y;
-}
-
-template<typename _Tp> inline
-Point_<_Tp> Rect_<_Tp>::tl() const
-{
-    return Point_<_Tp>(x,y);
-}
-
-template<typename _Tp> inline
-Point_<_Tp> Rect_<_Tp>::br() const
-{
-    return Point_<_Tp>(x + width, y + height);
-}
-
-template<typename _Tp> inline
-Size_<_Tp> Rect_<_Tp>::size() const
-{
-    return Size_<_Tp>(width, height);
-}
-
-template<typename _Tp> inline
-_Tp Rect_<_Tp>::area() const
-{
-    const _Tp result = width * height;
-    CV_DbgAssert(!std::numeric_limits<_Tp>::is_integer
-        || width == 0 || result / width == height); // make sure the result fits in the return value
-    return result;
-}
-
-template<typename _Tp> inline
-bool Rect_<_Tp>::empty() const
-{
-    return width <= 0 || height <= 0;
-}
-
-template<typename _Tp> template<typename _Tp2> inline
-Rect_<_Tp>::operator Rect_<_Tp2>() const
-{
-    return Rect_<_Tp2>(saturate_cast<_Tp2>(x), saturate_cast<_Tp2>(y), saturate_cast<_Tp2>(width), saturate_cast<_Tp2>(height));
-}
-
-template<typename _Tp> inline
-bool Rect_<_Tp>::contains(const Point_<_Tp>& pt) const
-{
-    return x <= pt.x && pt.x < x + width && y <= pt.y && pt.y < y + height;
-}
-
-
-template<typename _Tp> static inline
-Rect_<_Tp>& operator += ( Rect_<_Tp>& a, const Point_<_Tp>& b )
-{
-    a.x += b.x;
-    a.y += b.y;
-    return a;
-}
-
-template<typename _Tp> static inline
-Rect_<_Tp>& operator -= ( Rect_<_Tp>& a, const Point_<_Tp>& b )
-{
-    a.x -= b.x;
-    a.y -= b.y;
-    return a;
-}
-
-template<typename _Tp> static inline
-Rect_<_Tp>& operator += ( Rect_<_Tp>& a, const Size_<_Tp>& b )
-{
-    a.width += b.width;
-    a.height += b.height;
-    return a;
-}
-
-template<typename _Tp> static inline
-Rect_<_Tp>& operator -= ( Rect_<_Tp>& a, const Size_<_Tp>& b )
-{
-    const _Tp width = a.width - b.width;
-    const _Tp height = a.height - b.height;
-    CV_DbgAssert(width >= 0 && height >= 0);
-    a.width = width;
-    a.height = height;
-    return a;
-}
-
-template<typename _Tp> static inline
-Rect_<_Tp>& operator &= ( Rect_<_Tp>& a, const Rect_<_Tp>& b )
-{
-    if (a.empty() || b.empty()) {
-        a = Rect();
-        return a;
-    }
-    const Rect_<_Tp>& Rx_min = (a.x < b.x) ? a : b;
-    const Rect_<_Tp>& Rx_max = (a.x < b.x) ? b : a;
-    const Rect_<_Tp>& Ry_min = (a.y < b.y) ? a : b;
-    const Rect_<_Tp>& Ry_max = (a.y < b.y) ? b : a;
-    // Looking at the formula below, we will compute Rx_min.width - (Rx_max.x - Rx_min.x)
-    // but we want to avoid overflows. Rx_min.width >= 0 and (Rx_max.x - Rx_min.x) >= 0
-    // by definition so the difference does not overflow. The only thing that can overflow
-    // is (Rx_max.x - Rx_min.x). And it can only overflow if Rx_min.x < 0.
-    // Let us first deal with the following case.
-    if ((Rx_min.x < 0 && Rx_min.x + Rx_min.width < Rx_max.x) ||
-        (Ry_min.y < 0 && Ry_min.y + Ry_min.height < Ry_max.y)) {
-        a = Rect();
-        return a;
-    }
-    // We now know that either Rx_min.x >= 0, or
-    // Rx_min.x < 0 && Rx_min.x + Rx_min.width >= Rx_max.x and therefore
-    // Rx_min.width >= (Rx_max.x - Rx_min.x) which means (Rx_max.x - Rx_min.x)
-    // is inferior to a valid int and therefore does not overflow.
-    a.width = std::min(Rx_min.width - (Rx_max.x - Rx_min.x), Rx_max.width);
-    a.height = std::min(Ry_min.height - (Ry_max.y - Ry_min.y), Ry_max.height);
-    a.x = Rx_max.x;
-    a.y = Ry_max.y;
-    if (a.empty())
-        a = Rect();
-    return a;
-}
-
-template<typename _Tp> static inline
-Rect_<_Tp>& operator |= ( Rect_<_Tp>& a, const Rect_<_Tp>& b )
-{
-    if (a.empty()) {
-        a = b;
-    }
-    else if (!b.empty()) {
-        _Tp x1 = std::min(a.x, b.x);
-        _Tp y1 = std::min(a.y, b.y);
-        a.width = std::max(a.x + a.width, b.x + b.width) - x1;
-        a.height = std::max(a.y + a.height, b.y + b.height) - y1;
-        a.x = x1;
-        a.y = y1;
-    }
-    return a;
-}
-
-template<typename _Tp> static inline
-bool operator == (const Rect_<_Tp>& a, const Rect_<_Tp>& b)
-{
-    return a.x == b.x && a.y == b.y && a.width == b.width && a.height == b.height;
-}
-
-template<typename _Tp> static inline
-bool operator != (const Rect_<_Tp>& a, const Rect_<_Tp>& b)
-{
-    return a.x != b.x || a.y != b.y || a.width != b.width || a.height != b.height;
-}
-
-template<typename _Tp> static inline
-Rect_<_Tp> operator + (const Rect_<_Tp>& a, const Point_<_Tp>& b)
-{
-    return Rect_<_Tp>( a.x + b.x, a.y + b.y, a.width, a.height );
-}
-
-template<typename _Tp> static inline
-Rect_<_Tp> operator - (const Rect_<_Tp>& a, const Point_<_Tp>& b)
-{
-    return Rect_<_Tp>( a.x - b.x, a.y - b.y, a.width, a.height );
-}
-
-template<typename _Tp> static inline
-Rect_<_Tp> operator + (const Rect_<_Tp>& a, const Size_<_Tp>& b)
-{
-    return Rect_<_Tp>( a.x, a.y, a.width + b.width, a.height + b.height );
-}
-
-template<typename _Tp> static inline
-Rect_<_Tp> operator - (const Rect_<_Tp>& a, const Size_<_Tp>& b)
-{
-    const _Tp width = a.width - b.width;
-    const _Tp height = a.height - b.height;
-    CV_DbgAssert(width >= 0 && height >= 0);
-    return Rect_<_Tp>( a.x, a.y, width, height );
-}
-
-template<typename _Tp> static inline
-Rect_<_Tp> operator & (const Rect_<_Tp>& a, const Rect_<_Tp>& b)
-{
-    Rect_<_Tp> c = a;
-    return c &= b;
-}
-
-template<typename _Tp> static inline
-Rect_<_Tp> operator | (const Rect_<_Tp>& a, const Rect_<_Tp>& b)
-{
-    Rect_<_Tp> c = a;
-    return c |= b;
-}
-
-/**
- * @brief measure dissimilarity between two sample sets
- *
- * computes the complement of the Jaccard Index as described in <https://en.wikipedia.org/wiki/Jaccard_index>.
- * For rectangles this reduces to computing the intersection over the union.
- */
-template<typename _Tp> static inline
-double jaccardDistance(const Rect_<_Tp>& a, const Rect_<_Tp>& b) {
-    _Tp Aa = a.area();
-    _Tp Ab = b.area();
-
-    if ((Aa + Ab) <= std::numeric_limits<_Tp>::epsilon()) {
-        // jaccard_index = 1 -> distance = 0
-        return 0.0;
-    }
-
-    double Aab = (a & b).area();
-    // distance = 1 - jaccard_index
-    return 1.0 - Aab / (Aa + Ab - Aab);
-}
-
-////////////////////////////// RotatedRect //////////////////////////////
-
-inline
-RotatedRect::RotatedRect()
-    : center(), size(), angle(0) {}
-
-inline
-RotatedRect::RotatedRect(const Point2f& _center, const Size2f& _size, float _angle)
-    : center(_center), size(_size), angle(_angle) {}
-
-///////////////////////////////// Range /////////////////////////////////
-
-inline
-Range::Range()
-    : start(0), end(0) {}
-
-inline
-Range::Range(int _start, int _end)
-    : start(_start), end(_end) {}
-
-inline
-int Range::size() const
-{
-    return end - start;
-}
-
-inline
-bool Range::empty() const
-{
-    return start == end;
-}
-
-inline
-Range Range::all()
-{
-    return Range(INT_MIN, INT_MAX);
-}
-
-
-static inline
-bool operator == (const Range& r1, const Range& r2)
-{
-    return r1.start == r2.start && r1.end == r2.end;
-}
-
-static inline
-bool operator != (const Range& r1, const Range& r2)
-{
-    return !(r1 == r2);
-}
-
-static inline
-bool operator !(const Range& r)
-{
-    return r.start == r.end;
-}
-
-static inline
-Range operator & (const Range& r1, const Range& r2)
-{
-    Range r(std::max(r1.start, r2.start), std::min(r1.end, r2.end));
-    r.end = std::max(r.end, r.start);
-    return r;
-}
-
-static inline
-Range& operator &= (Range& r1, const Range& r2)
-{
-    r1 = r1 & r2;
-    return r1;
-}
-
-static inline
-Range operator + (const Range& r1, int delta)
-{
-    return Range(r1.start + delta, r1.end + delta);
-}
-
-static inline
-Range operator + (int delta, const Range& r1)
-{
-    return Range(r1.start + delta, r1.end + delta);
-}
-
-static inline
-Range operator - (const Range& r1, int delta)
-{
-    return r1 + (-delta);
-}
-
-
-
-///////////////////////////////// Scalar ////////////////////////////////
-
-template<typename _Tp> inline
-Scalar_<_Tp>::Scalar_()
-{
-    this->val[0] = this->val[1] = this->val[2] = this->val[3] = 0;
-}
-
-template<typename _Tp> inline
-Scalar_<_Tp>::Scalar_(_Tp v0, _Tp v1, _Tp v2, _Tp v3)
-{
-    this->val[0] = v0;
-    this->val[1] = v1;
-    this->val[2] = v2;
-    this->val[3] = v3;
-}
-
-template<typename _Tp> inline
-Scalar_<_Tp>::Scalar_(const Scalar_<_Tp>& s) : Vec<_Tp, 4>(s) {
-}
-
-template<typename _Tp> inline
-Scalar_<_Tp>::Scalar_(Scalar_<_Tp>&& s) CV_NOEXCEPT {
-    this->val[0] = std::move(s.val[0]);
-    this->val[1] = std::move(s.val[1]);
-    this->val[2] = std::move(s.val[2]);
-    this->val[3] = std::move(s.val[3]);
-}
-
-template<typename _Tp> inline
-Scalar_<_Tp>& Scalar_<_Tp>::operator=(const Scalar_<_Tp>& s) {
-    this->val[0] = s.val[0];
-    this->val[1] = s.val[1];
-    this->val[2] = s.val[2];
-    this->val[3] = s.val[3];
-    return *this;
-}
-
-template<typename _Tp> inline
-Scalar_<_Tp>& Scalar_<_Tp>::operator=(Scalar_<_Tp>&& s) CV_NOEXCEPT {
-    this->val[0] = std::move(s.val[0]);
-    this->val[1] = std::move(s.val[1]);
-    this->val[2] = std::move(s.val[2]);
-    this->val[3] = std::move(s.val[3]);
-    return *this;
-}
-
-template<typename _Tp> template<typename _Tp2, int cn> inline
-Scalar_<_Tp>::Scalar_(const Vec<_Tp2, cn>& v)
-{
-    int i;
-    for( i = 0; i < (cn < 4 ? cn : 4); i++ )
-        this->val[i] = cv::saturate_cast<_Tp>(v.val[i]);
-    for( ; i < 4; i++ )
-        this->val[i] = 0;
-}
-
-template<typename _Tp> inline
-Scalar_<_Tp>::Scalar_(_Tp v0)
-{
-    this->val[0] = v0;
-    this->val[1] = this->val[2] = this->val[3] = 0;
-}
-
-template<typename _Tp> inline
-Scalar_<_Tp> Scalar_<_Tp>::all(_Tp v0)
-{
-    return Scalar_<_Tp>(v0, v0, v0, v0);
-}
-
-
-template<typename _Tp> inline
-Scalar_<_Tp> Scalar_<_Tp>::mul(const Scalar_<_Tp>& a, double scale ) const
-{
-    return Scalar_<_Tp>(saturate_cast<_Tp>(this->val[0] * a.val[0] * scale),
-                        saturate_cast<_Tp>(this->val[1] * a.val[1] * scale),
-                        saturate_cast<_Tp>(this->val[2] * a.val[2] * scale),
-                        saturate_cast<_Tp>(this->val[3] * a.val[3] * scale));
-}
-
-template<typename _Tp> inline
-Scalar_<_Tp> Scalar_<_Tp>::conj() const
-{
-    return Scalar_<_Tp>(saturate_cast<_Tp>( this->val[0]),
-                        saturate_cast<_Tp>(-this->val[1]),
-                        saturate_cast<_Tp>(-this->val[2]),
-                        saturate_cast<_Tp>(-this->val[3]));
-}
-
-template<typename _Tp> inline
-bool Scalar_<_Tp>::isReal() const
-{
-    return this->val[1] == 0 && this->val[2] == 0 && this->val[3] == 0;
-}
-
-
-template<typename _Tp> template<typename T2> inline
-Scalar_<_Tp>::operator Scalar_<T2>() const
-{
-    return Scalar_<T2>(saturate_cast<T2>(this->val[0]),
-                       saturate_cast<T2>(this->val[1]),
-                       saturate_cast<T2>(this->val[2]),
-                       saturate_cast<T2>(this->val[3]));
-}
-
-
-template<typename _Tp> static inline
-Scalar_<_Tp>& operator += (Scalar_<_Tp>& a, const Scalar_<_Tp>& b)
-{
-    a.val[0] += b.val[0];
-    a.val[1] += b.val[1];
-    a.val[2] += b.val[2];
-    a.val[3] += b.val[3];
-    return a;
-}
-
-template<typename _Tp> static inline
-Scalar_<_Tp>& operator -= (Scalar_<_Tp>& a, const Scalar_<_Tp>& b)
-{
-    a.val[0] -= b.val[0];
-    a.val[1] -= b.val[1];
-    a.val[2] -= b.val[2];
-    a.val[3] -= b.val[3];
-    return a;
-}
-
-template<typename _Tp> static inline
-Scalar_<_Tp>& operator *= ( Scalar_<_Tp>& a, _Tp v )
-{
-    a.val[0] *= v;
-    a.val[1] *= v;
-    a.val[2] *= v;
-    a.val[3] *= v;
-    return a;
-}
-
-template<typename _Tp> static inline
-bool operator == ( const Scalar_<_Tp>& a, const Scalar_<_Tp>& b )
-{
-    return a.val[0] == b.val[0] && a.val[1] == b.val[1] &&
-           a.val[2] == b.val[2] && a.val[3] == b.val[3];
-}
-
-template<typename _Tp> static inline
-bool operator != ( const Scalar_<_Tp>& a, const Scalar_<_Tp>& b )
-{
-    return a.val[0] != b.val[0] || a.val[1] != b.val[1] ||
-           a.val[2] != b.val[2] || a.val[3] != b.val[3];
-}
-
-template<typename _Tp> static inline
-Scalar_<_Tp> operator + (const Scalar_<_Tp>& a, const Scalar_<_Tp>& b)
-{
-    return Scalar_<_Tp>(a.val[0] + b.val[0],
-                        a.val[1] + b.val[1],
-                        a.val[2] + b.val[2],
-                        a.val[3] + b.val[3]);
-}
-
-template<typename _Tp> static inline
-Scalar_<_Tp> operator - (const Scalar_<_Tp>& a, const Scalar_<_Tp>& b)
-{
-    return Scalar_<_Tp>(saturate_cast<_Tp>(a.val[0] - b.val[0]),
-                        saturate_cast<_Tp>(a.val[1] - b.val[1]),
-                        saturate_cast<_Tp>(a.val[2] - b.val[2]),
-                        saturate_cast<_Tp>(a.val[3] - b.val[3]));
-}
-
-template<typename _Tp> static inline
-Scalar_<_Tp> operator * (const Scalar_<_Tp>& a, _Tp alpha)
-{
-    return Scalar_<_Tp>(a.val[0] * alpha,
-                        a.val[1] * alpha,
-                        a.val[2] * alpha,
-                        a.val[3] * alpha);
-}
-
-template<typename _Tp> static inline
-Scalar_<_Tp> operator * (_Tp alpha, const Scalar_<_Tp>& a)
-{
-    return a*alpha;
-}
-
-template<typename _Tp> static inline
-Scalar_<_Tp> operator - (const Scalar_<_Tp>& a)
-{
-    return Scalar_<_Tp>(saturate_cast<_Tp>(-a.val[0]),
-                        saturate_cast<_Tp>(-a.val[1]),
-                        saturate_cast<_Tp>(-a.val[2]),
-                        saturate_cast<_Tp>(-a.val[3]));
-}
-
-
-template<typename _Tp> static inline
-Scalar_<_Tp> operator * (const Scalar_<_Tp>& a, const Scalar_<_Tp>& b)
-{
-    return Scalar_<_Tp>(saturate_cast<_Tp>(a[0]*b[0] - a[1]*b[1] - a[2]*b[2] - a[3]*b[3]),
-                        saturate_cast<_Tp>(a[0]*b[1] + a[1]*b[0] + a[2]*b[3] - a[3]*b[2]),
-                        saturate_cast<_Tp>(a[0]*b[2] - a[1]*b[3] + a[2]*b[0] + a[3]*b[1]),
-                        saturate_cast<_Tp>(a[0]*b[3] + a[1]*b[2] - a[2]*b[1] + a[3]*b[0]));
-}
-
-template<typename _Tp> static inline
-Scalar_<_Tp>& operator *= (Scalar_<_Tp>& a, const Scalar_<_Tp>& b)
-{
-    a = a * b;
-    return a;
-}
-
-template<typename _Tp> static inline
-Scalar_<_Tp> operator / (const Scalar_<_Tp>& a, _Tp alpha)
-{
-    return Scalar_<_Tp>(a.val[0] / alpha,
-                        a.val[1] / alpha,
-                        a.val[2] / alpha,
-                        a.val[3] / alpha);
-}
-
-template<typename _Tp> static inline
-Scalar_<float> operator / (const Scalar_<float>& a, float alpha)
-{
-    float s = 1 / alpha;
-    return Scalar_<float>(a.val[0] * s, a.val[1] * s, a.val[2] * s, a.val[3] * s);
-}
-
-template<typename _Tp> static inline
-Scalar_<double> operator / (const Scalar_<double>& a, double alpha)
-{
-    double s = 1 / alpha;
-    return Scalar_<double>(a.val[0] * s, a.val[1] * s, a.val[2] * s, a.val[3] * s);
-}
-
-template<typename _Tp> static inline
-Scalar_<_Tp>& operator /= (Scalar_<_Tp>& a, _Tp alpha)
-{
-    a = a / alpha;
-    return a;
-}
-
-template<typename _Tp> static inline
-Scalar_<_Tp> operator / (_Tp a, const Scalar_<_Tp>& b)
-{
-    _Tp s = a / (b[0]*b[0] + b[1]*b[1] + b[2]*b[2] + b[3]*b[3]);
-    return b.conj() * s;
-}
-
-template<typename _Tp> static inline
-Scalar_<_Tp> operator / (const Scalar_<_Tp>& a, const Scalar_<_Tp>& b)
-{
-    return a * ((_Tp)1 / b);
-}
-
-template<typename _Tp> static inline
-Scalar_<_Tp>& operator /= (Scalar_<_Tp>& a, const Scalar_<_Tp>& b)
-{
-    a = a / b;
-    return a;
-}
-
-template<typename _Tp> static inline
-Scalar operator * (const Matx<_Tp, 4, 4>& a, const Scalar& b)
-{
-    Matx<double, 4, 1> c((Matx<double, 4, 4>)a, b, Matx_MatMulOp());
-    return reinterpret_cast<const Scalar&>(c);
-}
-
-template<> inline
-Scalar operator * (const Matx<double, 4, 4>& a, const Scalar& b)
-{
-    Matx<double, 4, 1> c(a, b, Matx_MatMulOp());
-    return reinterpret_cast<const Scalar&>(c);
-}
-
-
-
-//////////////////////////////// KeyPoint ///////////////////////////////
-
-inline
-KeyPoint::KeyPoint()
-    : pt(0,0), size(0), angle(-1), response(0), octave(0), class_id(-1) {}
-
-inline
-KeyPoint::KeyPoint(Point2f _pt, float _size, float _angle, float _response, int _octave, int _class_id)
-    : pt(_pt), size(_size), angle(_angle), response(_response), octave(_octave), class_id(_class_id) {}
-
-inline
-KeyPoint::KeyPoint(float x, float y, float _size, float _angle, float _response, int _octave, int _class_id)
-    : pt(x, y), size(_size), angle(_angle), response(_response), octave(_octave), class_id(_class_id) {}
-
-
-
-///////////////////////////////// DMatch ////////////////////////////////
-
-inline
-DMatch::DMatch()
-    : queryIdx(-1), trainIdx(-1), imgIdx(-1), distance(FLT_MAX) {}
-
-inline
-DMatch::DMatch(int _queryIdx, int _trainIdx, float _distance)
-    : queryIdx(_queryIdx), trainIdx(_trainIdx), imgIdx(-1), distance(_distance) {}
-
-inline
-DMatch::DMatch(int _queryIdx, int _trainIdx, int _imgIdx, float _distance)
-    : queryIdx(_queryIdx), trainIdx(_trainIdx), imgIdx(_imgIdx), distance(_distance) {}
-
-inline
-bool DMatch::operator < (const DMatch &m) const
-{
-    return distance < m.distance;
-}
-
-
-
-////////////////////////////// TermCriteria /////////////////////////////
-
-inline
-TermCriteria::TermCriteria()
-    : type(0), maxCount(0), epsilon(0) {}
-
-inline
-TermCriteria::TermCriteria(int _type, int _maxCount, double _epsilon)
-    : type(_type), maxCount(_maxCount), epsilon(_epsilon) {}
-
-//! @endcond
-
-} // cv
-
-#endif //OPENCV_CORE_TYPES_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/types_c.h b/projects/llm_framework/include/opencv4/opencv2/core/types_c.h
deleted file mode 100644
index 32f3c8c9..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/types_c.h
+++ /dev/null
@@ -1,2126 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_TYPES_H
-#define OPENCV_CORE_TYPES_H
-
-#ifdef CV__ENABLE_C_API_CTORS  // invalid C API ctors (must be removed)
-#if defined(_WIN32) && !defined(CV__SKIP_MESSAGE_MALFORMED_C_API_CTORS)
-#error "C API ctors don't work on Win32: https://github.com/opencv/opencv/issues/15990"
-#endif
-#endif
-
-//#define CV__VALIDATE_UNUNITIALIZED_VARS 1  // C++11 & GCC only
-
-#ifdef __cplusplus
-
-#ifdef CV__VALIDATE_UNUNITIALIZED_VARS
-#pragma GCC diagnostic ignored "-Wmissing-field-initializers"
-#define CV_STRUCT_INITIALIZER {0,}
-#else
-#if defined(__GNUC__) && __GNUC__ == 4  // GCC 4.x warns on "= {}" initialization, fixed in GCC 5.0
-#pragma GCC diagnostic ignored "-Wmissing-field-initializers"
-#endif
-#define CV_STRUCT_INITIALIZER {}
-#endif
-
-#else
-#define CV_STRUCT_INITIALIZER {0}
-#endif
-
-
-#ifdef HAVE_IPL
-#  ifndef __IPL_H__
-#    if defined _WIN32
-#      include <ipl.h>
-#    else
-#      include <ipl/ipl.h>
-#    endif
-#  endif
-#elif defined __IPL_H__
-#  define HAVE_IPL
-#endif
-
-#include "opencv2/core/cvdef.h"
-
-#ifndef SKIP_INCLUDES
-#include <assert.h>
-#include <stdlib.h>
-#include <string.h>
-#include <float.h>
-#endif // SKIP_INCLUDES
-
-#if defined _WIN32
-#  define CV_CDECL __cdecl
-#  define CV_STDCALL __stdcall
-#else
-#  define CV_CDECL
-#  define CV_STDCALL
-#endif
-
-#ifndef CV_DEFAULT
-#  ifdef __cplusplus
-#    define CV_DEFAULT(val) = val
-#  else
-#    define CV_DEFAULT(val)
-#  endif
-#endif
-
-#ifndef CV_EXTERN_C_FUNCPTR
-#  ifdef __cplusplus
-#    define CV_EXTERN_C_FUNCPTR(x) extern "C" { typedef x; }
-#  else
-#    define CV_EXTERN_C_FUNCPTR(x) typedef x
-#  endif
-#endif
-
-#ifndef CVAPI
-#  define CVAPI(rettype) CV_EXTERN_C CV_EXPORTS rettype CV_CDECL
-#endif
-
-#ifndef CV_IMPL
-#  define CV_IMPL CV_EXTERN_C
-#endif
-
-#ifdef __cplusplus
-#  include "opencv2/core.hpp"
-#endif
-
-/** @addtogroup core_c
-    @{
-*/
-
-/** @brief This is the "metatype" used *only* as a function parameter.
-
-It denotes that the function accepts arrays of multiple types, such as IplImage*, CvMat* or even
-CvSeq* sometimes. The particular array type is determined at runtime by analyzing the first 4
-bytes of the header. In C++ interface the role of CvArr is played by InputArray and OutputArray.
- */
-typedef void CvArr;
-
-typedef int CVStatus;
-
-/** @see cv::Error::Code */
-enum {
- CV_StsOk=                       0,  /**< everything is ok                */
- CV_StsBackTrace=               -1,  /**< pseudo error for back trace     */
- CV_StsError=                   -2,  /**< unknown /unspecified error      */
- CV_StsInternal=                -3,  /**< internal error (bad state)      */
- CV_StsNoMem=                   -4,  /**< insufficient memory             */
- CV_StsBadArg=                  -5,  /**< function arg/param is bad       */
- CV_StsBadFunc=                 -6,  /**< unsupported function            */
- CV_StsNoConv=                  -7,  /**< iter. didn't converge           */
- CV_StsAutoTrace=               -8,  /**< tracing                         */
- CV_HeaderIsNull=               -9,  /**< image header is NULL            */
- CV_BadImageSize=              -10,  /**< image size is invalid           */
- CV_BadOffset=                 -11,  /**< offset is invalid               */
- CV_BadDataPtr=                -12,  /**/
- CV_BadStep=                   -13,  /**< image step is wrong, this may happen for a non-continuous matrix */
- CV_BadModelOrChSeq=           -14,  /**/
- CV_BadNumChannels=            -15,  /**< bad number of channels, for example, some functions accept only single channel matrices */
- CV_BadNumChannel1U=           -16,  /**/
- CV_BadDepth=                  -17,  /**< input image depth is not supported by the function */
- CV_BadAlphaChannel=           -18,  /**/
- CV_BadOrder=                  -19,  /**< number of dimensions is out of range */
- CV_BadOrigin=                 -20,  /**< incorrect input origin               */
- CV_BadAlign=                  -21,  /**< incorrect input align                */
- CV_BadCallBack=               -22,  /**/
- CV_BadTileSize=               -23,  /**/
- CV_BadCOI=                    -24,  /**< input COI is not supported           */
- CV_BadROISize=                -25,  /**< incorrect input roi                  */
- CV_MaskIsTiled=               -26,  /**/
- CV_StsNullPtr=                -27,  /**< null pointer */
- CV_StsVecLengthErr=           -28,  /**< incorrect vector length */
- CV_StsFilterStructContentErr= -29,  /**< incorrect filter structure content */
- CV_StsKernelStructContentErr= -30,  /**< incorrect transform kernel content */
- CV_StsFilterOffsetErr=        -31,  /**< incorrect filter offset value */
- CV_StsBadSize=                -201, /**< the input/output structure size is incorrect  */
- CV_StsDivByZero=              -202, /**< division by zero */
- CV_StsInplaceNotSupported=    -203, /**< in-place operation is not supported */
- CV_StsObjectNotFound=         -204, /**< request can't be completed */
- CV_StsUnmatchedFormats=       -205, /**< formats of input/output arrays differ */
- CV_StsBadFlag=                -206, /**< flag is wrong or not supported */
- CV_StsBadPoint=               -207, /**< bad CvPoint */
- CV_StsBadMask=                -208, /**< bad format of mask (neither 8uC1 nor 8sC1)*/
- CV_StsUnmatchedSizes=         -209, /**< sizes of input/output structures do not match */
- CV_StsUnsupportedFormat=      -210, /**< the data format/type is not supported by the function*/
- CV_StsOutOfRange=             -211, /**< some of parameters are out of range */
- CV_StsParseError=             -212, /**< invalid syntax/structure of the parsed file */
- CV_StsNotImplemented=         -213, /**< the requested function/feature is not implemented */
- CV_StsBadMemBlock=            -214, /**< an allocated block has been corrupted */
- CV_StsAssert=                 -215, /**< assertion failed   */
- CV_GpuNotSupported=           -216, /**< no CUDA support    */
- CV_GpuApiCallError=           -217, /**< GPU API call error */
- CV_OpenGlNotSupported=        -218, /**< no OpenGL support  */
- CV_OpenGlApiCallError=        -219, /**< OpenGL API call error */
- CV_OpenCLApiCallError=        -220, /**< OpenCL API call error */
- CV_OpenCLDoubleNotSupported=  -221,
- CV_OpenCLInitError=           -222, /**< OpenCL initialization error */
- CV_OpenCLNoAMDBlasFft=        -223
-};
-
-/****************************************************************************************\
-*                             Common macros and inline functions                         *
-\****************************************************************************************/
-
-#define CV_SWAP(a,b,t) ((t) = (a), (a) = (b), (b) = (t))
-
-/** min & max without jumps */
-#define  CV_IMIN(a, b)  ((a) ^ (((a)^(b)) & (((a) < (b)) - 1)))
-
-#define  CV_IMAX(a, b)  ((a) ^ (((a)^(b)) & (((a) > (b)) - 1)))
-
-/** absolute value without jumps */
-#ifndef __cplusplus
-#  define  CV_IABS(a)     (((a) ^ ((a) < 0 ? -1 : 0)) - ((a) < 0 ? -1 : 0))
-#else
-#  define  CV_IABS(a)     abs(a)
-#endif
-#define  CV_CMP(a,b)    (((a) > (b)) - ((a) < (b)))
-#define  CV_SIGN(a)     CV_CMP((a),0)
-
-#define cvInvSqrt(value) ((float)(1./sqrt(value)))
-#define cvSqrt(value)  ((float)sqrt(value))
-
-
-/*************** Random number generation *******************/
-
-typedef uint64 CvRNG;
-
-#define CV_RNG_COEFF 4164903690U
-
-/** @brief Initializes a random number generator state.
-
-The function initializes a random number generator and returns the state. The pointer to the state
-can be then passed to the cvRandInt, cvRandReal and cvRandArr functions. In the current
-implementation a multiply-with-carry generator is used.
-@param seed 64-bit value used to initiate a random sequence
-@sa the C++ class RNG replaced CvRNG.
- */
-CV_INLINE CvRNG cvRNG( int64 seed CV_DEFAULT(-1))
-{
-    CvRNG rng = seed ? (uint64)seed : (uint64)(int64)-1;
-    return rng;
-}
-
-/** @brief Returns a 32-bit unsigned integer and updates RNG.
-
-The function returns a uniformly-distributed random 32-bit unsigned integer and updates the RNG
-state. It is similar to the rand() function from the C runtime library, except that OpenCV functions
-always generates a 32-bit random number, regardless of the platform.
-@param rng CvRNG state initialized by cvRNG.
- */
-CV_INLINE unsigned cvRandInt( CvRNG* rng )
-{
-    uint64 temp = *rng;
-    temp = (uint64)(unsigned)temp*CV_RNG_COEFF + (temp >> 32);
-    *rng = temp;
-    return (unsigned)temp;
-}
-
-/** @brief Returns a floating-point random number and updates RNG.
-
-The function returns a uniformly-distributed random floating-point number between 0 and 1 (1 is not
-included).
-@param rng RNG state initialized by cvRNG
- */
-CV_INLINE double cvRandReal( CvRNG* rng )
-{
-    return cvRandInt(rng)*2.3283064365386962890625e-10 /* 2^-32 */;
-}
-
-/****************************************************************************************\
-*                                  Image type (IplImage)                                 *
-\****************************************************************************************/
-
-#ifndef HAVE_IPL
-
-/*
- * The following definitions (until #endif)
- * is an extract from IPL headers.
- * Copyright (c) 1995 Intel Corporation.
- */
-#define IPL_DEPTH_SIGN 0x80000000
-
-#define IPL_DEPTH_1U     1
-#define IPL_DEPTH_8U     8
-#define IPL_DEPTH_16U   16
-#define IPL_DEPTH_32F   32
-
-#define IPL_DEPTH_8S  (IPL_DEPTH_SIGN| 8)
-#define IPL_DEPTH_16S (IPL_DEPTH_SIGN|16)
-#define IPL_DEPTH_32S (IPL_DEPTH_SIGN|32)
-
-#define IPL_DATA_ORDER_PIXEL  0
-#define IPL_DATA_ORDER_PLANE  1
-
-#define IPL_ORIGIN_TL 0
-#define IPL_ORIGIN_BL 1
-
-#define IPL_ALIGN_4BYTES   4
-#define IPL_ALIGN_8BYTES   8
-#define IPL_ALIGN_16BYTES 16
-#define IPL_ALIGN_32BYTES 32
-
-#define IPL_ALIGN_DWORD   IPL_ALIGN_4BYTES
-#define IPL_ALIGN_QWORD   IPL_ALIGN_8BYTES
-
-#define IPL_BORDER_CONSTANT   0
-#define IPL_BORDER_REPLICATE  1
-#define IPL_BORDER_REFLECT    2
-#define IPL_BORDER_WRAP       3
-
-#ifdef __cplusplus
-typedef struct _IplImage IplImage;
-CV_EXPORTS _IplImage cvIplImage(const cv::Mat& m);
-#endif
-
-/** The IplImage is taken from the Intel Image Processing Library, in which the format is native. OpenCV
-only supports a subset of possible IplImage formats, as outlined in the parameter list above.
-
-In addition to the above restrictions, OpenCV handles ROIs differently. OpenCV functions require
-that the image size or ROI size of all source and destination images match exactly. On the other
-hand, the Intel Image Processing Library processes the area of intersection between the source and
-destination images (or ROIs), allowing them to vary independently.
-*/
-typedef struct
-_IplImage
-{
-    int  nSize;             /**< sizeof(IplImage) */
-    int  ID;                /**< version (=0)*/
-    int  nChannels;         /**< Most of OpenCV functions support 1,2,3 or 4 channels */
-    int  alphaChannel;      /**< Ignored by OpenCV */
-    int  depth;             /**< Pixel depth in bits: IPL_DEPTH_8U, IPL_DEPTH_8S, IPL_DEPTH_16S,
-                               IPL_DEPTH_32S, IPL_DEPTH_32F and IPL_DEPTH_64F are supported.  */
-    char colorModel[4];     /**< Ignored by OpenCV */
-    char channelSeq[4];     /**< ditto */
-    int  dataOrder;         /**< 0 - interleaved color channels, 1 - separate color channels.
-                               cvCreateImage can only create interleaved images */
-    int  origin;            /**< 0 - top-left origin,
-                               1 - bottom-left origin (Windows bitmaps style).  */
-    int  align;             /**< Alignment of image rows (4 or 8).
-                               OpenCV ignores it and uses widthStep instead.    */
-    int  width;             /**< Image width in pixels.                           */
-    int  height;            /**< Image height in pixels.                          */
-    struct _IplROI *roi;    /**< Image ROI. If NULL, the whole image is selected. */
-    struct _IplImage *maskROI;      /**< Must be NULL. */
-    void  *imageId;                 /**< "           " */
-    struct _IplTileInfo *tileInfo;  /**< "           " */
-    int  imageSize;         /**< Image data size in bytes
-                               (==image->height*image->widthStep
-                               in case of interleaved data)*/
-    char *imageData;        /**< Pointer to aligned image data.         */
-    int  widthStep;         /**< Size of aligned image row in bytes.    */
-    int  BorderMode[4];     /**< Ignored by OpenCV.                     */
-    int  BorderConst[4];    /**< Ditto.                                 */
-    char *imageDataOrigin;  /**< Pointer to very origin of image data
-                               (not necessarily aligned) -
-                               needed for correct deallocation */
-
-#if defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus)
-    _IplImage()
-    {
-        memset(this, 0, sizeof(*this));  // valid for POD structure
-        nSize = sizeof(IplImage);
-    }
-    _IplImage(const cv::Mat& m) { *this = cvIplImage(m); }
-#endif
-}
-IplImage;
-
-CV_INLINE IplImage cvIplImage()
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    IplImage self = CV_STRUCT_INITIALIZER; self.nSize = sizeof(IplImage); return self;
-#else
-    return _IplImage();
-#endif
-}
-
-typedef struct _IplTileInfo IplTileInfo;
-
-typedef struct _IplROI
-{
-    int  coi; /**< 0 - no COI (all channels are selected), 1 - 0th channel is selected ...*/
-    int  xOffset;
-    int  yOffset;
-    int  width;
-    int  height;
-}
-IplROI;
-
-typedef struct _IplConvKernel
-{
-    int  nCols;
-    int  nRows;
-    int  anchorX;
-    int  anchorY;
-    int *values;
-    int  nShiftR;
-}
-IplConvKernel;
-
-typedef struct _IplConvKernelFP
-{
-    int  nCols;
-    int  nRows;
-    int  anchorX;
-    int  anchorY;
-    float *values;
-}
-IplConvKernelFP;
-
-#define IPL_IMAGE_HEADER 1
-#define IPL_IMAGE_DATA   2
-#define IPL_IMAGE_ROI    4
-
-#endif/*HAVE_IPL*/
-
-/** extra border mode */
-#define IPL_BORDER_REFLECT_101    4
-#define IPL_BORDER_TRANSPARENT    5
-
-#define IPL_IMAGE_MAGIC_VAL  ((int)sizeof(IplImage))
-#define CV_TYPE_NAME_IMAGE "opencv-image"
-
-#define CV_IS_IMAGE_HDR(img) \
-    ((img) != NULL && ((const IplImage*)(img))->nSize == sizeof(IplImage))
-
-#define CV_IS_IMAGE(img) \
-    (CV_IS_IMAGE_HDR(img) && ((IplImage*)img)->imageData != NULL)
-
-/** for storing double-precision
-   floating point data in IplImage's */
-#define IPL_DEPTH_64F  64
-
-/** get reference to pixel at (col,row),
-   for multi-channel images (col) should be multiplied by number of channels */
-#define CV_IMAGE_ELEM( image, elemtype, row, col )       \
-    (((elemtype*)((image)->imageData + (image)->widthStep*(row)))[(col)])
-
-/****************************************************************************************\
-*                                  Matrix type (CvMat)                                   *
-\****************************************************************************************/
-
-#define CV_AUTO_STEP  0x7fffffff
-#define CV_WHOLE_ARR  cvSlice( 0, 0x3fffffff )
-
-#define CV_MAGIC_MASK       0xFFFF0000
-#define CV_MAT_MAGIC_VAL    0x42420000
-#define CV_TYPE_NAME_MAT    "opencv-matrix"
-
-#ifdef __cplusplus
-typedef struct CvMat CvMat;
-CV_INLINE CvMat cvMat(const cv::Mat& m);
-#endif
-
-/** Matrix elements are stored row by row. Element (i, j) (i - 0-based row index, j - 0-based column
-index) of a matrix can be retrieved or modified using CV_MAT_ELEM macro:
-
-    uchar pixval = CV_MAT_ELEM(grayimg, uchar, i, j)
-    CV_MAT_ELEM(cameraMatrix, float, 0, 2) = image.width*0.5f;
-
-To access multiple-channel matrices, you can use
-CV_MAT_ELEM(matrix, type, i, j\*nchannels + channel_idx).
-
-@deprecated CvMat is now obsolete; consider using Mat instead.
- */
-typedef struct CvMat
-{
-    int type;
-    int step;
-
-    /* for internal use only */
-    int* refcount;
-    int hdr_refcount;
-
-    union
-    {
-        uchar* ptr;
-        short* s;
-        int* i;
-        float* fl;
-        double* db;
-    } data;
-
-#ifdef __cplusplus
-    union
-    {
-        int rows;
-        int height;
-    };
-
-    union
-    {
-        int cols;
-        int width;
-    };
-#else
-    int rows;
-    int cols;
-#endif
-
-#if defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus)
-    CvMat() {}
-    CvMat(const cv::Mat& m) { *this = cvMat(m); }
-#endif
-}
-CvMat;
-
-
-#define CV_IS_MAT_HDR(mat) \
-    ((mat) != NULL && \
-    (((const CvMat*)(mat))->type & CV_MAGIC_MASK) == CV_MAT_MAGIC_VAL && \
-    ((const CvMat*)(mat))->cols > 0 && ((const CvMat*)(mat))->rows > 0)
-
-#define CV_IS_MAT_HDR_Z(mat) \
-    ((mat) != NULL && \
-    (((const CvMat*)(mat))->type & CV_MAGIC_MASK) == CV_MAT_MAGIC_VAL && \
-    ((const CvMat*)(mat))->cols >= 0 && ((const CvMat*)(mat))->rows >= 0)
-
-#define CV_IS_MAT(mat) \
-    (CV_IS_MAT_HDR(mat) && ((const CvMat*)(mat))->data.ptr != NULL)
-
-#define CV_IS_MASK_ARR(mat) \
-    (((mat)->type & (CV_MAT_TYPE_MASK & ~CV_8SC1)) == 0)
-
-#define CV_ARE_TYPES_EQ(mat1, mat2) \
-    ((((mat1)->type ^ (mat2)->type) & CV_MAT_TYPE_MASK) == 0)
-
-#define CV_ARE_CNS_EQ(mat1, mat2) \
-    ((((mat1)->type ^ (mat2)->type) & CV_MAT_CN_MASK) == 0)
-
-#define CV_ARE_DEPTHS_EQ(mat1, mat2) \
-    ((((mat1)->type ^ (mat2)->type) & CV_MAT_DEPTH_MASK) == 0)
-
-#define CV_ARE_SIZES_EQ(mat1, mat2) \
-    ((mat1)->rows == (mat2)->rows && (mat1)->cols == (mat2)->cols)
-
-#define CV_IS_MAT_CONST(mat)  \
-    (((mat)->rows|(mat)->cols) == 1)
-
-#define IPL2CV_DEPTH(depth) \
-    ((((CV_8U)+(CV_16U<<4)+(CV_32F<<8)+(CV_64F<<16)+(CV_8S<<20)+ \
-    (CV_16S<<24)+(CV_32S<<28)) >> ((((depth) & 0xF0) >> 2) + \
-    (((depth) & IPL_DEPTH_SIGN) ? 20 : 0))) & 15)
-
-/** Inline constructor. No data is allocated internally!!!
- * (Use together with cvCreateData, or use cvCreateMat instead to
- * get a matrix with allocated data):
- */
-CV_INLINE CvMat cvMat( int rows, int cols, int type, void* data CV_DEFAULT(NULL))
-{
-    CvMat m;
-
-    assert( (unsigned)CV_MAT_DEPTH(type) <= CV_64F );
-    type = CV_MAT_TYPE(type);
-    m.type = CV_MAT_MAGIC_VAL | CV_MAT_CONT_FLAG | type;
-    m.cols = cols;
-    m.rows = rows;
-    m.step = m.cols*CV_ELEM_SIZE(type);
-    m.data.ptr = (uchar*)data;
-    m.refcount = NULL;
-    m.hdr_refcount = 0;
-
-    return m;
-}
-
-#ifdef __cplusplus
-
-CV_INLINE CvMat cvMat(const cv::Mat& m)
-{
-    CvMat self;
-    CV_DbgAssert(m.dims <= 2);
-    self = cvMat(m.rows, m.dims == 1 ? 1 : m.cols, m.type(), m.data);
-    self.step = (int)m.step[0];
-    self.type = (self.type & ~cv::Mat::CONTINUOUS_FLAG) | (m.flags & cv::Mat::CONTINUOUS_FLAG);
-    return self;
-}
-CV_INLINE CvMat cvMat()
-{
-#if !defined(CV__ENABLE_C_API_CTORS)
-    CvMat self = CV_STRUCT_INITIALIZER; return self;
-#else
-    return CvMat();
-#endif
-}
-CV_INLINE CvMat cvMat(const CvMat& m)
-{
-#if !defined(CV__ENABLE_C_API_CTORS)
-    CvMat self = CV_STRUCT_INITIALIZER; memcpy(&self, &m, sizeof(self)); return self;
-#else
-    return CvMat(m);
-#endif
-}
-
-#endif // __cplusplus
-
-
-#define CV_MAT_ELEM_PTR_FAST( mat, row, col, pix_size )  \
-    (assert( (unsigned)(row) < (unsigned)(mat).rows &&   \
-             (unsigned)(col) < (unsigned)(mat).cols ),   \
-     (mat).data.ptr + (size_t)(mat).step*(row) + (pix_size)*(col))
-
-#define CV_MAT_ELEM_PTR( mat, row, col )                 \
-    CV_MAT_ELEM_PTR_FAST( mat, row, col, CV_ELEM_SIZE((mat).type) )
-
-#define CV_MAT_ELEM( mat, elemtype, row, col )           \
-    (*(elemtype*)CV_MAT_ELEM_PTR_FAST( mat, row, col, sizeof(elemtype)))
-
-/** @brief Returns the particular element of single-channel floating-point matrix.
-
-The function is a fast replacement for cvGetReal2D in the case of single-channel floating-point
-matrices. It is faster because it is inline, it does fewer checks for array type and array element
-type, and it checks for the row and column ranges only in debug mode.
-@param mat Input matrix
-@param row The zero-based index of row
-@param col The zero-based index of column
- */
-CV_INLINE  double  cvmGet( const CvMat* mat, int row, int col )
-{
-    int type;
-
-    type = CV_MAT_TYPE(mat->type);
-    assert( (unsigned)row < (unsigned)mat->rows &&
-            (unsigned)col < (unsigned)mat->cols );
-
-    if( type == CV_32FC1 )
-        return ((float*)(void*)(mat->data.ptr + (size_t)mat->step*row))[col];
-    else
-    {
-        assert( type == CV_64FC1 );
-        return ((double*)(void*)(mat->data.ptr + (size_t)mat->step*row))[col];
-    }
-}
-
-/** @brief Sets a specific element of a single-channel floating-point matrix.
-
-The function is a fast replacement for cvSetReal2D in the case of single-channel floating-point
-matrices. It is faster because it is inline, it does fewer checks for array type and array element
-type, and it checks for the row and column ranges only in debug mode.
-@param mat The matrix
-@param row The zero-based index of row
-@param col The zero-based index of column
-@param value The new value of the matrix element
- */
-CV_INLINE  void  cvmSet( CvMat* mat, int row, int col, double value )
-{
-    int type;
-    type = CV_MAT_TYPE(mat->type);
-    assert( (unsigned)row < (unsigned)mat->rows &&
-            (unsigned)col < (unsigned)mat->cols );
-
-    if( type == CV_32FC1 )
-        ((float*)(void*)(mat->data.ptr + (size_t)mat->step*row))[col] = (float)value;
-    else
-    {
-        assert( type == CV_64FC1 );
-        ((double*)(void*)(mat->data.ptr + (size_t)mat->step*row))[col] = value;
-    }
-}
-
-
-CV_INLINE int cvIplDepth( int type )
-{
-    int depth = CV_MAT_DEPTH(type);
-    return CV_ELEM_SIZE1(depth)*8 | (depth == CV_8S || depth == CV_16S ||
-           depth == CV_32S ? IPL_DEPTH_SIGN : 0);
-}
-
-
-/****************************************************************************************\
-*                       Multi-dimensional dense array (CvMatND)                          *
-\****************************************************************************************/
-
-#define CV_MATND_MAGIC_VAL    0x42430000
-#define CV_TYPE_NAME_MATND    "opencv-nd-matrix"
-
-#define CV_MAX_DIM            32
-
-#ifdef __cplusplus
-typedef struct CvMatND CvMatND;
-CV_EXPORTS CvMatND cvMatND(const cv::Mat& m);
-#endif
-
-/**
-  @deprecated consider using cv::Mat instead
-  */
-typedef struct
-CvMatND
-{
-    int type;
-    int dims;
-
-    int* refcount;
-    int hdr_refcount;
-
-    union
-    {
-        uchar* ptr;
-        float* fl;
-        double* db;
-        int* i;
-        short* s;
-    } data;
-
-    struct
-    {
-        int size;
-        int step;
-    }
-    dim[CV_MAX_DIM];
-
-#if defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus)
-    CvMatND() {}
-    CvMatND(const cv::Mat& m) { *this = cvMatND(m); }
-#endif
-}
-CvMatND;
-
-
-CV_INLINE CvMatND cvMatND()
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvMatND self = CV_STRUCT_INITIALIZER; return self;
-#else
-    return CvMatND();
-#endif
-}
-
-#define CV_IS_MATND_HDR(mat) \
-    ((mat) != NULL && (((const CvMatND*)(mat))->type & CV_MAGIC_MASK) == CV_MATND_MAGIC_VAL)
-
-#define CV_IS_MATND(mat) \
-    (CV_IS_MATND_HDR(mat) && ((const CvMatND*)(mat))->data.ptr != NULL)
-
-
-/****************************************************************************************\
-*                      Multi-dimensional sparse array (CvSparseMat)                      *
-\****************************************************************************************/
-
-#define CV_SPARSE_MAT_MAGIC_VAL    0x42440000
-#define CV_TYPE_NAME_SPARSE_MAT    "opencv-sparse-matrix"
-
-struct CvSet;
-
-typedef struct CvSparseMat
-{
-    int type;
-    int dims;
-    int* refcount;
-    int hdr_refcount;
-
-    struct CvSet* heap;
-    void** hashtable;
-    int hashsize;
-    int valoffset;
-    int idxoffset;
-    int size[CV_MAX_DIM];
-
-#ifdef __cplusplus
-    CV_EXPORTS void copyToSparseMat(cv::SparseMat& m) const;
-#endif
-}
-CvSparseMat;
-
-#ifdef __cplusplus
-CV_EXPORTS CvSparseMat* cvCreateSparseMat(const cv::SparseMat& m);
-#endif
-
-#define CV_IS_SPARSE_MAT_HDR(mat) \
-    ((mat) != NULL && \
-    (((const CvSparseMat*)(mat))->type & CV_MAGIC_MASK) == CV_SPARSE_MAT_MAGIC_VAL)
-
-#define CV_IS_SPARSE_MAT(mat) \
-    CV_IS_SPARSE_MAT_HDR(mat)
-
-/**************** iteration through a sparse array *****************/
-
-typedef struct CvSparseNode
-{
-    unsigned hashval;
-    struct CvSparseNode* next;
-}
-CvSparseNode;
-
-typedef struct CvSparseMatIterator
-{
-    CvSparseMat* mat;
-    CvSparseNode* node;
-    int curidx;
-}
-CvSparseMatIterator;
-
-#define CV_NODE_VAL(mat,node)   ((void*)((uchar*)(node) + (mat)->valoffset))
-#define CV_NODE_IDX(mat,node)   ((int*)((uchar*)(node) + (mat)->idxoffset))
-
-/****************************************************************************************\
-*                                         Histogram                                      *
-\****************************************************************************************/
-
-typedef int CvHistType;
-
-#define CV_HIST_MAGIC_VAL     0x42450000
-#define CV_HIST_UNIFORM_FLAG  (1 << 10)
-
-/** indicates whether bin ranges are set already or not */
-#define CV_HIST_RANGES_FLAG   (1 << 11)
-
-#define CV_HIST_ARRAY         0
-#define CV_HIST_SPARSE        1
-#define CV_HIST_TREE          CV_HIST_SPARSE
-
-/** should be used as a parameter only,
-   it turns to CV_HIST_UNIFORM_FLAG of hist->type */
-#define CV_HIST_UNIFORM       1
-
-typedef struct CvHistogram
-{
-    int     type;
-    CvArr*  bins;
-    float   thresh[CV_MAX_DIM][2];  /**< For uniform histograms.                      */
-    float** thresh2;                /**< For non-uniform histograms.                  */
-    CvMatND mat;                    /**< Embedded matrix header for array histograms. */
-}
-CvHistogram;
-
-#define CV_IS_HIST( hist ) \
-    ((hist) != NULL  && \
-     (((CvHistogram*)(hist))->type & CV_MAGIC_MASK) == CV_HIST_MAGIC_VAL && \
-     (hist)->bins != NULL)
-
-#define CV_IS_UNIFORM_HIST( hist ) \
-    (((hist)->type & CV_HIST_UNIFORM_FLAG) != 0)
-
-#define CV_IS_SPARSE_HIST( hist ) \
-    CV_IS_SPARSE_MAT((hist)->bins)
-
-#define CV_HIST_HAS_RANGES( hist ) \
-    (((hist)->type & CV_HIST_RANGES_FLAG) != 0)
-
-/****************************************************************************************\
-*                      Other supplementary data type definitions                         *
-\****************************************************************************************/
-
-/*************************************** CvRect *****************************************/
-/** @sa Rect_ */
-typedef struct CvRect
-{
-    int x;
-    int y;
-    int width;
-    int height;
-
-#ifdef CV__VALIDATE_UNUNITIALIZED_VARS
-    CvRect() __attribute__(( warning("Non-initialized variable") )) {};
-    template<typename _Tp> CvRect(const std::initializer_list<_Tp> list)
-    {
-        CV_Assert(list.size() == 0 || list.size() == 4);
-        x = y = width = height = 0;
-        if (list.size() == 4)
-        {
-            x = list.begin()[0]; y = list.begin()[1]; width = list.begin()[2]; height = list.begin()[3];
-        }
-    };
-#elif defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus)
-    CvRect(int _x = 0, int _y = 0, int w = 0, int h = 0): x(_x), y(_y), width(w), height(h) {}
-    template<typename _Tp>
-    CvRect(const cv::Rect_<_Tp>& r): x(cv::saturate_cast<int>(r.x)), y(cv::saturate_cast<int>(r.y)), width(cv::saturate_cast<int>(r.width)), height(cv::saturate_cast<int>(r.height)) {}
-#endif
-#ifdef __cplusplus
-    template<typename _Tp>
-    operator cv::Rect_<_Tp>() const { return cv::Rect_<_Tp>((_Tp)x, (_Tp)y, (_Tp)width, (_Tp)height); }
-#endif
-}
-CvRect;
-
-/** constructs CvRect structure. */
-CV_INLINE  CvRect  cvRect( int x, int y, int width, int height )
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvRect r = {x, y, width, height};
-#else
-    CvRect r(x, y , width, height);
-#endif
-    return r;
-}
-#ifdef __cplusplus
-CV_INLINE CvRect cvRect(const cv::Rect& rc) { return cvRect(rc.x, rc.y, rc.width, rc.height); }
-#endif
-
-CV_INLINE  IplROI  cvRectToROI( CvRect rect, int coi )
-{
-    IplROI roi;
-    roi.xOffset = rect.x;
-    roi.yOffset = rect.y;
-    roi.width = rect.width;
-    roi.height = rect.height;
-    roi.coi = coi;
-
-    return roi;
-}
-
-
-CV_INLINE  CvRect  cvROIToRect( IplROI roi )
-{
-    return cvRect( roi.xOffset, roi.yOffset, roi.width, roi.height );
-}
-
-/*********************************** CvTermCriteria *************************************/
-
-#define CV_TERMCRIT_ITER    1
-#define CV_TERMCRIT_NUMBER  CV_TERMCRIT_ITER
-#define CV_TERMCRIT_EPS     2
-
-/** @sa TermCriteria
- */
-typedef struct CvTermCriteria
-{
-    int    type;  /**< may be combination of
-                     CV_TERMCRIT_ITER
-                     CV_TERMCRIT_EPS */
-    int    max_iter;
-    double epsilon;
-#if defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus)
-    CvTermCriteria(int _type = 0, int _iter = 0, double _eps = 0) : type(_type), max_iter(_iter), epsilon(_eps)  {}
-    CvTermCriteria(const cv::TermCriteria& t) : type(t.type), max_iter(t.maxCount), epsilon(t.epsilon)  {}
-#endif
-#ifdef __cplusplus
-    operator cv::TermCriteria() const { return cv::TermCriteria(type, max_iter, epsilon); }
-#endif
-}
-CvTermCriteria;
-
-CV_INLINE  CvTermCriteria  cvTermCriteria( int type, int max_iter, double epsilon )
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvTermCriteria t = { type, max_iter, (float)epsilon};
-#else
-    CvTermCriteria t(type, max_iter, epsilon);
-#endif
-    return t;
-}
-#ifdef __cplusplus
-CV_INLINE CvTermCriteria cvTermCriteria(const cv::TermCriteria& t) { return cvTermCriteria(t.type, t.maxCount, t.epsilon); }
-#endif
-
-
-/******************************* CvPoint and variants ***********************************/
-
-typedef struct CvPoint
-{
-    int x;
-    int y;
-
-#ifdef CV__VALIDATE_UNUNITIALIZED_VARS
-    CvPoint() __attribute__(( warning("Non-initialized variable") )) {}
-    template<typename _Tp> CvPoint(const std::initializer_list<_Tp> list)
-    {
-        CV_Assert(list.size() == 0 || list.size() == 2);
-        x = y = 0;
-        if (list.size() == 2)
-        {
-            x = list.begin()[0]; y = list.begin()[1];
-        }
-    };
-#elif defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus)
-    CvPoint(int _x = 0, int _y = 0): x(_x), y(_y) {}
-    template<typename _Tp>
-    CvPoint(const cv::Point_<_Tp>& pt): x((int)pt.x), y((int)pt.y) {}
-#endif
-#ifdef __cplusplus
-    template<typename _Tp>
-    operator cv::Point_<_Tp>() const { return cv::Point_<_Tp>(cv::saturate_cast<_Tp>(x), cv::saturate_cast<_Tp>(y)); }
-#endif
-}
-CvPoint;
-
-/** constructs CvPoint structure. */
-CV_INLINE  CvPoint  cvPoint( int x, int y )
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvPoint p = {x, y};
-#else
-    CvPoint p(x, y);
-#endif
-    return p;
-}
-#ifdef __cplusplus
-CV_INLINE CvPoint cvPoint(const cv::Point& pt) { return cvPoint(pt.x, pt.y); }
-#endif
-
-typedef struct CvPoint2D32f
-{
-    float x;
-    float y;
-
-#ifdef CV__VALIDATE_UNUNITIALIZED_VARS
-    CvPoint2D32f() __attribute__(( warning("Non-initialized variable") )) {}
-    template<typename _Tp> CvPoint2D32f(const std::initializer_list<_Tp> list)
-    {
-        CV_Assert(list.size() == 0 || list.size() == 2);
-        x = y = 0;
-        if (list.size() == 2)
-        {
-            x = list.begin()[0]; y = list.begin()[1];
-        }
-    };
-#elif defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus)
-    CvPoint2D32f(float _x = 0, float _y = 0): x(_x), y(_y) {}
-    template<typename _Tp>
-    CvPoint2D32f(const cv::Point_<_Tp>& pt): x((float)pt.x), y((float)pt.y) {}
-#endif
-#ifdef __cplusplus
-    template<typename _Tp>
-    operator cv::Point_<_Tp>() const { return cv::Point_<_Tp>(cv::saturate_cast<_Tp>(x), cv::saturate_cast<_Tp>(y)); }
-#endif
-}
-CvPoint2D32f;
-
-/** constructs CvPoint2D32f structure. */
-CV_INLINE  CvPoint2D32f  cvPoint2D32f( double x, double y )
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvPoint2D32f p = { (float)x, (float)y };
-#else
-    CvPoint2D32f p((float)x, (float)y);
-#endif
-    return p;
-}
-
-#ifdef __cplusplus
-template<typename _Tp>
-CvPoint2D32f cvPoint2D32f(const cv::Point_<_Tp>& pt)
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvPoint2D32f p = { (float)pt.x, (float)pt.y };
-#else
-    CvPoint2D32f p((float)pt.x, (float)pt.y);
-#endif
-    return p;
-}
-#endif
-
-/** converts CvPoint to CvPoint2D32f. */
-CV_INLINE  CvPoint2D32f  cvPointTo32f( CvPoint point )
-{
-    return cvPoint2D32f( (float)point.x, (float)point.y );
-}
-
-/** converts CvPoint2D32f to CvPoint. */
-CV_INLINE  CvPoint  cvPointFrom32f( CvPoint2D32f point )
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvPoint ipt = { cvRound(point.x), cvRound(point.y) };
-#else
-    CvPoint ipt(cvRound(point.x), cvRound(point.y));
-#endif
-    return ipt;
-}
-
-
-typedef struct CvPoint3D32f
-{
-    float x;
-    float y;
-    float z;
-
-#ifdef CV__VALIDATE_UNUNITIALIZED_VARS
-    CvPoint3D32f() __attribute__(( warning("Non-initialized variable") )) {}
-    template<typename _Tp> CvPoint3D32f(const std::initializer_list<_Tp> list)
-    {
-        CV_Assert(list.size() == 0 || list.size() == 3);
-        x = y = z = 0;
-        if (list.size() == 3)
-        {
-            x = list.begin()[0]; y = list.begin()[1]; z = list.begin()[2];
-        }
-    };
-#elif defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus)
-    CvPoint3D32f(float _x = 0, float _y = 0, float _z = 0): x(_x), y(_y), z(_z) {}
-    template<typename _Tp>
-    CvPoint3D32f(const cv::Point3_<_Tp>& pt): x((float)pt.x), y((float)pt.y), z((float)pt.z) {}
-#endif
-#ifdef __cplusplus
-    template<typename _Tp>
-    operator cv::Point3_<_Tp>() const { return cv::Point3_<_Tp>(cv::saturate_cast<_Tp>(x), cv::saturate_cast<_Tp>(y), cv::saturate_cast<_Tp>(z)); }
-#endif
-}
-CvPoint3D32f;
-
-/** constructs CvPoint3D32f structure. */
-CV_INLINE  CvPoint3D32f  cvPoint3D32f( double x, double y, double z )
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvPoint3D32f p = { (float)x, (float)y, (float)z };
-#else
-    CvPoint3D32f p((float)x, (float)y, (float)z);
-#endif
-    return p;
-}
-
-#ifdef __cplusplus
-template<typename _Tp>
-CvPoint3D32f cvPoint3D32f(const cv::Point3_<_Tp>& pt)
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvPoint3D32f p  = { (float)pt.x, (float)pt.y, (float)pt.z };
-#else
-    CvPoint3D32f p((float)pt.x, (float)pt.y, (float)pt.z);
-#endif
-    return p;
-}
-#endif
-
-
-typedef struct CvPoint2D64f
-{
-    double x;
-    double y;
-#ifdef CV__VALIDATE_UNUNITIALIZED_VARS
-    CvPoint2D64f() __attribute__(( warning("Non-initialized variable") )) {}
-    template<typename _Tp> CvPoint2D64f(const std::initializer_list<_Tp> list)
-    {
-        CV_Assert(list.size() == 0 || list.size() == 2);
-        x = y = 0;
-        if (list.size() == 2)
-        {
-            x = list.begin()[0]; y = list.begin()[1];
-        }
-    };
-#endif
-}
-CvPoint2D64f;
-
-/** constructs CvPoint2D64f structure.*/
-CV_INLINE  CvPoint2D64f  cvPoint2D64f( double x, double y )
-{
-    CvPoint2D64f p = { x, y };
-    return p;
-}
-
-
-typedef struct CvPoint3D64f
-{
-    double x;
-    double y;
-    double z;
-#ifdef CV__VALIDATE_UNUNITIALIZED_VARS
-    CvPoint3D64f() __attribute__(( warning("Non-initialized variable") )) {}
-    template<typename _Tp> CvPoint3D64f(const std::initializer_list<_Tp> list)
-    {
-        CV_Assert(list.size() == 0 || list.size() == 3);
-        x = y = z = 0;
-        if (list.size() == 3)
-        {
-            x = list.begin()[0]; y = list.begin()[1]; z = list.begin()[2];
-        }
-    };
-#endif
-}
-CvPoint3D64f;
-
-/** constructs CvPoint3D64f structure. */
-CV_INLINE  CvPoint3D64f  cvPoint3D64f( double x, double y, double z )
-{
-    CvPoint3D64f p = { x, y, z };
-    return p;
-}
-
-
-/******************************** CvSize's & CvBox **************************************/
-
-typedef struct CvSize
-{
-    int width;
-    int height;
-
-#ifdef CV__VALIDATE_UNUNITIALIZED_VARS
-    CvSize() __attribute__(( warning("Non-initialized variable") )) {}
-    template<typename _Tp> CvSize(const std::initializer_list<_Tp> list)
-    {
-        CV_Assert(list.size() == 0 || list.size() == 2);
-        width = 0; height = 0;
-        if (list.size() == 2)
-        {
-            width = list.begin()[0]; height = list.begin()[1];
-        }
-    };
-#elif defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus)
-    CvSize(int w = 0, int h = 0): width(w), height(h) {}
-    template<typename _Tp>
-    CvSize(const cv::Size_<_Tp>& sz): width(cv::saturate_cast<int>(sz.width)), height(cv::saturate_cast<int>(sz.height)) {}
-#endif
-#ifdef __cplusplus
-    template<typename _Tp>
-    operator cv::Size_<_Tp>() const { return cv::Size_<_Tp>(cv::saturate_cast<_Tp>(width), cv::saturate_cast<_Tp>(height)); }
-#endif
-}
-CvSize;
-
-/** constructs CvSize structure. */
-CV_INLINE  CvSize  cvSize( int width, int height )
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvSize s = { width, height };
-#else
-    CvSize s(width, height);
-#endif
-    return s;
-}
-
-#ifdef __cplusplus
-CV_INLINE CvSize cvSize(const cv::Size& sz)
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvSize s = { sz.width, sz.height };
-#else
-    CvSize s(sz.width, sz.height);
-#endif
-    return s;
-}
-#endif
-
-typedef struct CvSize2D32f
-{
-    float width;
-    float height;
-
-#ifdef CV__VALIDATE_UNUNITIALIZED_VARS
-    CvSize2D32f() __attribute__(( warning("Non-initialized variable") )) {}
-    template<typename _Tp> CvSize2D32f(const std::initializer_list<_Tp> list)
-    {
-        CV_Assert(list.size() == 0 || list.size() == 2);
-        width = 0; height = 0;
-        if (list.size() == 2)
-        {
-            width = list.begin()[0]; height = list.begin()[1];
-        }
-    };
-#elif defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus)
-    CvSize2D32f(float w = 0, float h = 0): width(w), height(h) {}
-    template<typename _Tp>
-    CvSize2D32f(const cv::Size_<_Tp>& sz): width(cv::saturate_cast<float>(sz.width)), height(cv::saturate_cast<float>(sz.height)) {}
-#endif
-#ifdef __cplusplus
-    template<typename _Tp>
-    operator cv::Size_<_Tp>() const { return cv::Size_<_Tp>(cv::saturate_cast<_Tp>(width), cv::saturate_cast<_Tp>(height)); }
-#endif
-}
-CvSize2D32f;
-
-/** constructs CvSize2D32f structure. */
-CV_INLINE  CvSize2D32f  cvSize2D32f( double width, double height )
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvSize2D32f s = { (float)width, (float)height };
-#else
-    CvSize2D32f s((float)width, (float)height);
-#endif
-    return s;
-}
-#ifdef __cplusplus
-template<typename _Tp>
-CvSize2D32f cvSize2D32f(const cv::Size_<_Tp>& sz)
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvSize2D32f s = { (float)sz.width, (float)sz.height };
-#else
-    CvSize2D32f s((float)sz.width, (float)sz.height);
-#endif
-    return s;
-}
-#endif
-
-/** @sa RotatedRect
- */
-typedef struct CvBox2D
-{
-    CvPoint2D32f center;  /**< Center of the box.                          */
-    CvSize2D32f  size;    /**< Box width and length.                       */
-    float angle;          /**< Angle between the horizontal axis           */
-                          /**< and the first side (i.e. length) in degrees */
-
-#if defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus)
-    CvBox2D(CvPoint2D32f c = CvPoint2D32f(), CvSize2D32f s = CvSize2D32f(), float a = 0) : center(c), size(s), angle(a) {}
-    CvBox2D(const cv::RotatedRect& rr) : center(rr.center), size(rr.size), angle(rr.angle) {}
-#endif
-#ifdef __cplusplus
-    operator cv::RotatedRect() const { return cv::RotatedRect(center, size, angle); }
-#endif
-}
-CvBox2D;
-
-
-#ifdef __cplusplus
-CV_INLINE CvBox2D cvBox2D(CvPoint2D32f c = CvPoint2D32f(), CvSize2D32f s = CvSize2D32f(), float a = 0)
-{
-    CvBox2D self;
-    self.center = c;
-    self.size = s;
-    self.angle = a;
-    return self;
-}
-CV_INLINE CvBox2D cvBox2D(const cv::RotatedRect& rr)
-{
-    CvBox2D self;
-    self.center = cvPoint2D32f(rr.center);
-    self.size = cvSize2D32f(rr.size);
-    self.angle = rr.angle;
-    return self;
-}
-#endif
-
-
-/** Line iterator state: */
-typedef struct CvLineIterator
-{
-    /** Pointer to the current point: */
-    uchar* ptr;
-
-    /* Bresenham algorithm state: */
-    int  err;
-    int  plus_delta;
-    int  minus_delta;
-    int  plus_step;
-    int  minus_step;
-}
-CvLineIterator;
-
-
-
-/************************************* CvSlice ******************************************/
-#define CV_WHOLE_SEQ_END_INDEX 0x3fffffff
-#define CV_WHOLE_SEQ  cvSlice(0, CV_WHOLE_SEQ_END_INDEX)
-
-typedef struct CvSlice
-{
-    int  start_index, end_index;
-
-#ifdef CV__VALIDATE_UNUNITIALIZED_VARS
-    CvSlice() __attribute__(( warning("Non-initialized variable") )) {}
-    template<typename _Tp> CvSlice(const std::initializer_list<_Tp> list)
-    {
-        CV_Assert(list.size() == 0 || list.size() == 2);
-        start_index = end_index = 0;
-        if (list.size() == 2)
-        {
-            start_index = list.begin()[0]; end_index = list.begin()[1];
-        }
-    };
-#endif
-#if defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus) && !defined(__CUDACC__)
-    CvSlice(int start = 0, int end = 0) : start_index(start), end_index(end) {}
-    CvSlice(const cv::Range& r) { *this = (r.start != INT_MIN && r.end != INT_MAX) ? CvSlice(r.start, r.end) : CvSlice(0, CV_WHOLE_SEQ_END_INDEX); }
-    operator cv::Range() const { return (start_index == 0 && end_index == CV_WHOLE_SEQ_END_INDEX ) ? cv::Range::all() : cv::Range(start_index, end_index); }
-#endif
-}
-CvSlice;
-
-CV_INLINE  CvSlice  cvSlice( int start, int end )
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus) && !defined(__CUDACC__))
-    CvSlice slice = { start, end };
-#else
-    CvSlice slice(start, end);
-#endif
-    return slice;
-}
-
-#if defined(__cplusplus)
-CV_INLINE  CvSlice  cvSlice(const cv::Range& r)
-{
-    CvSlice slice = (r.start != INT_MIN && r.end != INT_MAX) ? cvSlice(r.start, r.end) : cvSlice(0, CV_WHOLE_SEQ_END_INDEX);
-    return slice;
-}
-#endif
-
-
-/************************************* CvScalar *****************************************/
-/** @sa Scalar_
- */
-typedef struct CvScalar
-{
-    double val[4];
-
-#ifdef CV__VALIDATE_UNUNITIALIZED_VARS
-    CvScalar() __attribute__(( warning("Non-initialized variable") )) {}
-    CvScalar(const std::initializer_list<double> list)
-    {
-        CV_Assert(list.size() == 0 || list.size() == 4);
-        val[0] = val[1] = val[2] = val[3] = 0;
-        if (list.size() == 4)
-        {
-            val[0] = list.begin()[0]; val[1] = list.begin()[1]; val[2] = list.begin()[2]; val[3] = list.begin()[3];
-        }
-    };
-#elif defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus)
-    CvScalar() {}
-    CvScalar(double d0, double d1 = 0, double d2 = 0, double d3 = 0) { val[0] = d0; val[1] = d1; val[2] = d2; val[3] = d3; }
-    template<typename _Tp>
-    CvScalar(const cv::Scalar_<_Tp>& s) { val[0] = s.val[0]; val[1] = s.val[1]; val[2] = s.val[2]; val[3] = s.val[3]; }
-    template<typename _Tp, int cn>
-    CvScalar(const cv::Vec<_Tp, cn>& v)
-    {
-        int i;
-        for( i = 0; i < (cn < 4 ? cn : 4); i++ ) val[i] = v.val[i];
-        for( ; i < 4; i++ ) val[i] = 0;
-    }
-#endif
-#ifdef __cplusplus
-    template<typename _Tp>
-    operator cv::Scalar_<_Tp>() const { return cv::Scalar_<_Tp>(cv::saturate_cast<_Tp>(val[0]), cv::saturate_cast<_Tp>(val[1]), cv::saturate_cast<_Tp>(val[2]), cv::saturate_cast<_Tp>(val[3])); }
-#endif
-}
-CvScalar;
-
-CV_INLINE  CvScalar  cvScalar( double val0, double val1 CV_DEFAULT(0),
-                               double val2 CV_DEFAULT(0), double val3 CV_DEFAULT(0))
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvScalar scalar = CV_STRUCT_INITIALIZER;
-#else
-    CvScalar scalar;
-#endif
-    scalar.val[0] = val0; scalar.val[1] = val1;
-    scalar.val[2] = val2; scalar.val[3] = val3;
-    return scalar;
-}
-
-#ifdef __cplusplus
-CV_INLINE CvScalar cvScalar()
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvScalar scalar = CV_STRUCT_INITIALIZER;
-#else
-    CvScalar scalar;
-#endif
-    scalar.val[0] = scalar.val[1] = scalar.val[2] = scalar.val[3] = 0;
-    return scalar;
-}
-CV_INLINE CvScalar cvScalar(const cv::Scalar& s)
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvScalar scalar = CV_STRUCT_INITIALIZER;
-#else
-    CvScalar scalar;
-#endif
-    scalar.val[0] = s.val[0];
-    scalar.val[1] = s.val[1];
-    scalar.val[2] = s.val[2];
-    scalar.val[3] = s.val[3];
-    return scalar;
-}
-#endif
-
-CV_INLINE  CvScalar  cvRealScalar( double val0 )
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvScalar scalar = CV_STRUCT_INITIALIZER;
-#else
-    CvScalar scalar;
-#endif
-    scalar.val[0] = val0;
-    scalar.val[1] = scalar.val[2] = scalar.val[3] = 0;
-    return scalar;
-}
-
-CV_INLINE  CvScalar  cvScalarAll( double val0123 )
-{
-#if !(defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus))
-    CvScalar scalar = CV_STRUCT_INITIALIZER;
-#else
-    CvScalar scalar;
-#endif
-    scalar.val[0] = val0123;
-    scalar.val[1] = val0123;
-    scalar.val[2] = val0123;
-    scalar.val[3] = val0123;
-    return scalar;
-}
-
-/****************************************************************************************\
-*                                   Dynamic Data structures                              *
-\****************************************************************************************/
-
-/******************************** Memory storage ****************************************/
-
-typedef struct CvMemBlock
-{
-    struct CvMemBlock*  prev;
-    struct CvMemBlock*  next;
-}
-CvMemBlock;
-
-#define CV_STORAGE_MAGIC_VAL    0x42890000
-
-typedef struct CvMemStorage
-{
-    int signature;
-    CvMemBlock* bottom;           /**< First allocated block.                   */
-    CvMemBlock* top;              /**< Current memory block - top of the stack. */
-    struct  CvMemStorage* parent; /**< We get new blocks from parent as needed. */
-    int block_size;               /**< Block size.                              */
-    int free_space;               /**< Remaining free space in current block.   */
-}
-CvMemStorage;
-
-#define CV_IS_STORAGE(storage)  \
-    ((storage) != NULL &&       \
-    (((CvMemStorage*)(storage))->signature & CV_MAGIC_MASK) == CV_STORAGE_MAGIC_VAL)
-
-
-typedef struct CvMemStoragePos
-{
-    CvMemBlock* top;
-    int free_space;
-}
-CvMemStoragePos;
-
-
-/*********************************** Sequence *******************************************/
-
-typedef struct CvSeqBlock
-{
-    struct CvSeqBlock*  prev; /**< Previous sequence block.                   */
-    struct CvSeqBlock*  next; /**< Next sequence block.                       */
-    int    start_index;       /**< Index of the first element in the block +  */
-                              /**< sequence->first->start_index.              */
-    int    count;             /**< Number of elements in the block.           */
-    schar* data;              /**< Pointer to the first element of the block. */
-}
-CvSeqBlock;
-
-
-#define CV_TREE_NODE_FIELDS(node_type)                               \
-    int       flags;             /**< Miscellaneous flags.     */      \
-    int       header_size;       /**< Size of sequence header. */      \
-    struct    node_type* h_prev; /**< Previous sequence.       */      \
-    struct    node_type* h_next; /**< Next sequence.           */      \
-    struct    node_type* v_prev; /**< 2nd previous sequence.   */      \
-    struct    node_type* v_next  /**< 2nd next sequence.       */
-
-/**
-   Read/Write sequence.
-   Elements can be dynamically inserted to or deleted from the sequence.
-*/
-#define CV_SEQUENCE_FIELDS()                                              \
-    CV_TREE_NODE_FIELDS(CvSeq);                                           \
-    int       total;          /**< Total number of elements.            */  \
-    int       elem_size;      /**< Size of sequence element in bytes.   */  \
-    schar*    block_max;      /**< Maximal bound of the last block.     */  \
-    schar*    ptr;            /**< Current write pointer.               */  \
-    int       delta_elems;    /**< Grow seq this many at a time.        */  \
-    CvMemStorage* storage;    /**< Where the seq is stored.             */  \
-    CvSeqBlock* free_blocks;  /**< Free blocks list.                    */  \
-    CvSeqBlock* first;        /**< Pointer to the first sequence block. */
-
-typedef struct CvSeq
-{
-    CV_SEQUENCE_FIELDS()
-}
-CvSeq;
-
-#define CV_TYPE_NAME_SEQ             "opencv-sequence"
-#define CV_TYPE_NAME_SEQ_TREE        "opencv-sequence-tree"
-
-/*************************************** Set ********************************************/
-/** @brief Set
-  Order is not preserved. There can be gaps between sequence elements.
-  After the element has been inserted it stays in the same place all the time.
-  The MSB(most-significant or sign bit) of the first field (flags) is 0 iff the element exists.
-*/
-#define CV_SET_ELEM_FIELDS(elem_type)   \
-    int  flags;                         \
-    struct elem_type* next_free;
-
-typedef struct CvSetElem
-{
-    CV_SET_ELEM_FIELDS(CvSetElem)
-}
-CvSetElem;
-
-#define CV_SET_FIELDS()      \
-    CV_SEQUENCE_FIELDS()     \
-    CvSetElem* free_elems;   \
-    int active_count;
-
-typedef struct CvSet
-{
-    CV_SET_FIELDS()
-}
-CvSet;
-
-
-#define CV_SET_ELEM_IDX_MASK   ((1 << 26) - 1)
-#define CV_SET_ELEM_FREE_FLAG  (1 << (sizeof(int)*8-1))
-
-/** Checks whether the element pointed by ptr belongs to a set or not */
-#define CV_IS_SET_ELEM( ptr )  (((CvSetElem*)(ptr))->flags >= 0)
-
-/************************************* Graph ********************************************/
-
-/** @name Graph
-
-We represent a graph as a set of vertices. Vertices contain their adjacency lists (more exactly,
-pointers to first incoming or outcoming edge (or 0 if isolated vertex)). Edges are stored in
-another set. There is a singly-linked list of incoming/outcoming edges for each vertex.
-
-Each edge consists of:
-
-- Two pointers to the starting and ending vertices (vtx[0] and vtx[1] respectively).
-
-    A graph may be oriented or not. In the latter case, edges between vertex i to vertex j are not
-distinguished during search operations.
-
-- Two pointers to next edges for the starting and ending vertices, where next[0] points to the
-next edge in the vtx[0] adjacency list and next[1] points to the next edge in the vtx[1]
-adjacency list.
-
-@see CvGraphEdge, CvGraphVtx, CvGraphVtx2D, CvGraph
-@{
-*/
-#define CV_GRAPH_EDGE_FIELDS()      \
-    int flags;                      \
-    float weight;                   \
-    struct CvGraphEdge* next[2];    \
-    struct CvGraphVtx* vtx[2];
-
-
-#define CV_GRAPH_VERTEX_FIELDS()    \
-    int flags;                      \
-    struct CvGraphEdge* first;
-
-
-typedef struct CvGraphEdge
-{
-    CV_GRAPH_EDGE_FIELDS()
-}
-CvGraphEdge;
-
-typedef struct CvGraphVtx
-{
-    CV_GRAPH_VERTEX_FIELDS()
-}
-CvGraphVtx;
-
-typedef struct CvGraphVtx2D
-{
-    CV_GRAPH_VERTEX_FIELDS()
-    CvPoint2D32f* ptr;
-}
-CvGraphVtx2D;
-
-/**
-   Graph is "derived" from the set (this is set a of vertices)
-   and includes another set (edges)
-*/
-#define  CV_GRAPH_FIELDS()   \
-    CV_SET_FIELDS()          \
-    CvSet* edges;
-
-typedef struct CvGraph
-{
-    CV_GRAPH_FIELDS()
-}
-CvGraph;
-
-#define CV_TYPE_NAME_GRAPH "opencv-graph"
-
-/** @} */
-
-/*********************************** Chain/Contour *************************************/
-
-typedef struct CvChain
-{
-    CV_SEQUENCE_FIELDS()
-    CvPoint  origin;
-}
-CvChain;
-
-#define CV_CONTOUR_FIELDS()  \
-    CV_SEQUENCE_FIELDS()     \
-    CvRect rect;             \
-    int color;               \
-    int reserved[3];
-
-typedef struct CvContour
-{
-    CV_CONTOUR_FIELDS()
-}
-CvContour;
-
-typedef CvContour CvPoint2DSeq;
-
-/****************************************************************************************\
-*                                    Sequence types                                      *
-\****************************************************************************************/
-
-#define CV_SEQ_MAGIC_VAL             0x42990000
-
-#define CV_IS_SEQ(seq) \
-    ((seq) != NULL && (((CvSeq*)(seq))->flags & CV_MAGIC_MASK) == CV_SEQ_MAGIC_VAL)
-
-#define CV_SET_MAGIC_VAL             0x42980000
-#define CV_IS_SET(set) \
-    ((set) != NULL && (((CvSeq*)(set))->flags & CV_MAGIC_MASK) == CV_SET_MAGIC_VAL)
-
-#define CV_SEQ_ELTYPE_BITS           12
-#define CV_SEQ_ELTYPE_MASK           ((1 << CV_SEQ_ELTYPE_BITS) - 1)
-
-#define CV_SEQ_ELTYPE_POINT          CV_32SC2  /**< (x,y) */
-#define CV_SEQ_ELTYPE_CODE           CV_8UC1   /**< freeman code: 0..7 */
-#define CV_SEQ_ELTYPE_GENERIC        0
-#define CV_SEQ_ELTYPE_PTR            CV_MAKE_TYPE(CV_8U, 8 /*sizeof(void*)*/)
-#define CV_SEQ_ELTYPE_PPOINT         CV_SEQ_ELTYPE_PTR  /**< &(x,y) */
-#define CV_SEQ_ELTYPE_INDEX          CV_32SC1  /**< #(x,y) */
-#define CV_SEQ_ELTYPE_GRAPH_EDGE     0  /**< &next_o, &next_d, &vtx_o, &vtx_d */
-#define CV_SEQ_ELTYPE_GRAPH_VERTEX   0  /**< first_edge, &(x,y) */
-#define CV_SEQ_ELTYPE_TRIAN_ATR      0  /**< vertex of the binary tree   */
-#define CV_SEQ_ELTYPE_CONNECTED_COMP 0  /**< connected component  */
-#define CV_SEQ_ELTYPE_POINT3D        CV_32FC3  /**< (x,y,z)  */
-
-#define CV_SEQ_KIND_BITS        2
-#define CV_SEQ_KIND_MASK        (((1 << CV_SEQ_KIND_BITS) - 1)<<CV_SEQ_ELTYPE_BITS)
-
-/** types of sequences */
-#define CV_SEQ_KIND_GENERIC     (0 << CV_SEQ_ELTYPE_BITS)
-#define CV_SEQ_KIND_CURVE       (1 << CV_SEQ_ELTYPE_BITS)
-#define CV_SEQ_KIND_BIN_TREE    (2 << CV_SEQ_ELTYPE_BITS)
-
-/** types of sparse sequences (sets) */
-#define CV_SEQ_KIND_GRAPH       (1 << CV_SEQ_ELTYPE_BITS)
-#define CV_SEQ_KIND_SUBDIV2D    (2 << CV_SEQ_ELTYPE_BITS)
-
-#define CV_SEQ_FLAG_SHIFT       (CV_SEQ_KIND_BITS + CV_SEQ_ELTYPE_BITS)
-
-/** flags for curves */
-#define CV_SEQ_FLAG_CLOSED     (1 << CV_SEQ_FLAG_SHIFT)
-#define CV_SEQ_FLAG_SIMPLE     (0 << CV_SEQ_FLAG_SHIFT)
-#define CV_SEQ_FLAG_CONVEX     (0 << CV_SEQ_FLAG_SHIFT)
-#define CV_SEQ_FLAG_HOLE       (2 << CV_SEQ_FLAG_SHIFT)
-
-/** flags for graphs */
-#define CV_GRAPH_FLAG_ORIENTED (1 << CV_SEQ_FLAG_SHIFT)
-
-#define CV_GRAPH               CV_SEQ_KIND_GRAPH
-#define CV_ORIENTED_GRAPH      (CV_SEQ_KIND_GRAPH|CV_GRAPH_FLAG_ORIENTED)
-
-/** point sets */
-#define CV_SEQ_POINT_SET       (CV_SEQ_KIND_GENERIC| CV_SEQ_ELTYPE_POINT)
-#define CV_SEQ_POINT3D_SET     (CV_SEQ_KIND_GENERIC| CV_SEQ_ELTYPE_POINT3D)
-#define CV_SEQ_POLYLINE        (CV_SEQ_KIND_CURVE  | CV_SEQ_ELTYPE_POINT)
-#define CV_SEQ_POLYGON         (CV_SEQ_FLAG_CLOSED | CV_SEQ_POLYLINE )
-#define CV_SEQ_CONTOUR         CV_SEQ_POLYGON
-#define CV_SEQ_SIMPLE_POLYGON  (CV_SEQ_FLAG_SIMPLE | CV_SEQ_POLYGON  )
-
-/** chain-coded curves */
-#define CV_SEQ_CHAIN           (CV_SEQ_KIND_CURVE  | CV_SEQ_ELTYPE_CODE)
-#define CV_SEQ_CHAIN_CONTOUR   (CV_SEQ_FLAG_CLOSED | CV_SEQ_CHAIN)
-
-/** binary tree for the contour */
-#define CV_SEQ_POLYGON_TREE    (CV_SEQ_KIND_BIN_TREE  | CV_SEQ_ELTYPE_TRIAN_ATR)
-
-/** sequence of the connected components */
-#define CV_SEQ_CONNECTED_COMP  (CV_SEQ_KIND_GENERIC  | CV_SEQ_ELTYPE_CONNECTED_COMP)
-
-/** sequence of the integer numbers */
-#define CV_SEQ_INDEX           (CV_SEQ_KIND_GENERIC  | CV_SEQ_ELTYPE_INDEX)
-
-#define CV_SEQ_ELTYPE( seq )   ((seq)->flags & CV_SEQ_ELTYPE_MASK)
-#define CV_SEQ_KIND( seq )     ((seq)->flags & CV_SEQ_KIND_MASK )
-
-/** flag checking */
-#define CV_IS_SEQ_INDEX( seq )      ((CV_SEQ_ELTYPE(seq) == CV_SEQ_ELTYPE_INDEX) && \
-                                     (CV_SEQ_KIND(seq) == CV_SEQ_KIND_GENERIC))
-
-#define CV_IS_SEQ_CURVE( seq )      (CV_SEQ_KIND(seq) == CV_SEQ_KIND_CURVE)
-#define CV_IS_SEQ_CLOSED( seq )     (((seq)->flags & CV_SEQ_FLAG_CLOSED) != 0)
-#define CV_IS_SEQ_CONVEX( seq )     0
-#define CV_IS_SEQ_HOLE( seq )       (((seq)->flags & CV_SEQ_FLAG_HOLE) != 0)
-#define CV_IS_SEQ_SIMPLE( seq )     1
-
-/** type checking macros */
-#define CV_IS_SEQ_POINT_SET( seq ) \
-    ((CV_SEQ_ELTYPE(seq) == CV_32SC2 || CV_SEQ_ELTYPE(seq) == CV_32FC2))
-
-#define CV_IS_SEQ_POINT_SUBSET( seq ) \
-    (CV_IS_SEQ_INDEX( seq ) || CV_SEQ_ELTYPE(seq) == CV_SEQ_ELTYPE_PPOINT)
-
-#define CV_IS_SEQ_POLYLINE( seq )   \
-    (CV_SEQ_KIND(seq) == CV_SEQ_KIND_CURVE && CV_IS_SEQ_POINT_SET(seq))
-
-#define CV_IS_SEQ_POLYGON( seq )   \
-    (CV_IS_SEQ_POLYLINE(seq) && CV_IS_SEQ_CLOSED(seq))
-
-#define CV_IS_SEQ_CHAIN( seq )   \
-    (CV_SEQ_KIND(seq) == CV_SEQ_KIND_CURVE && (seq)->elem_size == 1)
-
-#define CV_IS_SEQ_CONTOUR( seq )   \
-    (CV_IS_SEQ_CLOSED(seq) && (CV_IS_SEQ_POLYLINE(seq) || CV_IS_SEQ_CHAIN(seq)))
-
-#define CV_IS_SEQ_CHAIN_CONTOUR( seq ) \
-    (CV_IS_SEQ_CHAIN( seq ) && CV_IS_SEQ_CLOSED( seq ))
-
-#define CV_IS_SEQ_POLYGON_TREE( seq ) \
-    (CV_SEQ_ELTYPE (seq) ==  CV_SEQ_ELTYPE_TRIAN_ATR &&    \
-    CV_SEQ_KIND( seq ) ==  CV_SEQ_KIND_BIN_TREE )
-
-#define CV_IS_GRAPH( seq )    \
-    (CV_IS_SET(seq) && CV_SEQ_KIND((CvSet*)(seq)) == CV_SEQ_KIND_GRAPH)
-
-#define CV_IS_GRAPH_ORIENTED( seq )   \
-    (((seq)->flags & CV_GRAPH_FLAG_ORIENTED) != 0)
-
-#define CV_IS_SUBDIV2D( seq )  \
-    (CV_IS_SET(seq) && CV_SEQ_KIND((CvSet*)(seq)) == CV_SEQ_KIND_SUBDIV2D)
-
-/****************************************************************************************/
-/*                            Sequence writer & reader                                  */
-/****************************************************************************************/
-
-#define CV_SEQ_WRITER_FIELDS()                                     \
-    int          header_size;                                      \
-    CvSeq*       seq;        /**< the sequence written */            \
-    CvSeqBlock*  block;      /**< current block */                   \
-    schar*       ptr;        /**< pointer to free space */           \
-    schar*       block_min;  /**< pointer to the beginning of block*/\
-    schar*       block_max;  /**< pointer to the end of block */
-
-typedef struct CvSeqWriter
-{
-    CV_SEQ_WRITER_FIELDS()
-}
-CvSeqWriter;
-
-
-#define CV_SEQ_READER_FIELDS()                                      \
-    int          header_size;                                       \
-    CvSeq*       seq;        /**< sequence, beign read */             \
-    CvSeqBlock*  block;      /**< current block */                    \
-    schar*       ptr;        /**< pointer to element be read next */  \
-    schar*       block_min;  /**< pointer to the beginning of block */\
-    schar*       block_max;  /**< pointer to the end of block */      \
-    int          delta_index;/**< = seq->first->start_index   */      \
-    schar*       prev_elem;  /**< pointer to previous element */
-
-typedef struct CvSeqReader
-{
-    CV_SEQ_READER_FIELDS()
-}
-CvSeqReader;
-
-/****************************************************************************************/
-/*                                Operations on sequences                               */
-/****************************************************************************************/
-
-#define  CV_SEQ_ELEM( seq, elem_type, index )                    \
-/** assert gives some guarantee that <seq> parameter is valid */  \
-(   assert(sizeof((seq)->first[0]) == sizeof(CvSeqBlock) &&      \
-    (seq)->elem_size == sizeof(elem_type)),                      \
-    (elem_type*)((seq)->first && (unsigned)index <               \
-    (unsigned)((seq)->first->count) ?                            \
-    (seq)->first->data + (index) * sizeof(elem_type) :           \
-    cvGetSeqElem( (CvSeq*)(seq), (index) )))
-#define CV_GET_SEQ_ELEM( elem_type, seq, index ) CV_SEQ_ELEM( (seq), elem_type, (index) )
-
-/** Add element to sequence: */
-#define CV_WRITE_SEQ_ELEM_VAR( elem_ptr, writer )     \
-{                                                     \
-    if( (writer).ptr >= (writer).block_max )          \
-    {                                                 \
-        cvCreateSeqBlock( &writer);                   \
-    }                                                 \
-    memcpy((writer).ptr, elem_ptr, (writer).seq->elem_size);\
-    (writer).ptr += (writer).seq->elem_size;          \
-}
-
-#define CV_WRITE_SEQ_ELEM( elem, writer )             \
-{                                                     \
-    assert( (writer).seq->elem_size == sizeof(elem)); \
-    if( (writer).ptr >= (writer).block_max )          \
-    {                                                 \
-        cvCreateSeqBlock( &writer);                   \
-    }                                                 \
-    assert( (writer).ptr <= (writer).block_max - sizeof(elem));\
-    memcpy((writer).ptr, &(elem), sizeof(elem));      \
-    (writer).ptr += sizeof(elem);                     \
-}
-
-
-/** Move reader position forward: */
-#define CV_NEXT_SEQ_ELEM( elem_size, reader )                 \
-{                                                             \
-    if( ((reader).ptr += (elem_size)) >= (reader).block_max ) \
-    {                                                         \
-        cvChangeSeqBlock( &(reader), 1 );                     \
-    }                                                         \
-}
-
-
-/** Move reader position backward: */
-#define CV_PREV_SEQ_ELEM( elem_size, reader )                \
-{                                                            \
-    if( ((reader).ptr -= (elem_size)) < (reader).block_min ) \
-    {                                                        \
-        cvChangeSeqBlock( &(reader), -1 );                   \
-    }                                                        \
-}
-
-/** Read element and move read position forward: */
-#define CV_READ_SEQ_ELEM( elem, reader )                       \
-{                                                              \
-    assert( (reader).seq->elem_size == sizeof(elem));          \
-    memcpy( &(elem), (reader).ptr, sizeof((elem)));            \
-    CV_NEXT_SEQ_ELEM( sizeof(elem), reader )                   \
-}
-
-/** Read element and move read position backward: */
-#define CV_REV_READ_SEQ_ELEM( elem, reader )                     \
-{                                                                \
-    assert( (reader).seq->elem_size == sizeof(elem));            \
-    memcpy(&(elem), (reader).ptr, sizeof((elem)));               \
-    CV_PREV_SEQ_ELEM( sizeof(elem), reader )                     \
-}
-
-
-#define CV_READ_CHAIN_POINT( _pt, reader )                              \
-{                                                                       \
-    (_pt) = (reader).pt;                                                \
-    if( (reader).ptr )                                                  \
-    {                                                                   \
-        CV_READ_SEQ_ELEM( (reader).code, (reader));                     \
-        assert( ((reader).code & ~7) == 0 );                            \
-        (reader).pt.x += (reader).deltas[(int)(reader).code][0];        \
-        (reader).pt.y += (reader).deltas[(int)(reader).code][1];        \
-    }                                                                   \
-}
-
-#define CV_CURRENT_POINT( reader )  (*((CvPoint*)((reader).ptr)))
-#define CV_PREV_POINT( reader )     (*((CvPoint*)((reader).prev_elem)))
-
-#define CV_READ_EDGE( pt1, pt2, reader )               \
-{                                                      \
-    assert( sizeof(pt1) == sizeof(CvPoint) &&          \
-            sizeof(pt2) == sizeof(CvPoint) &&          \
-            reader.seq->elem_size == sizeof(CvPoint)); \
-    (pt1) = CV_PREV_POINT( reader );                   \
-    (pt2) = CV_CURRENT_POINT( reader );                \
-    (reader).prev_elem = (reader).ptr;                 \
-    CV_NEXT_SEQ_ELEM( sizeof(CvPoint), (reader));      \
-}
-
-/************ Graph macros ************/
-
-/** Return next graph edge for given vertex: */
-#define  CV_NEXT_GRAPH_EDGE( edge, vertex )                              \
-     (assert((edge)->vtx[0] == (vertex) || (edge)->vtx[1] == (vertex)),  \
-      (edge)->next[(edge)->vtx[1] == (vertex)])
-
-
-
-/****************************************************************************************\
-*             Data structures for persistence (a.k.a serialization) functionality        *
-\****************************************************************************************/
-
-#if 0
-
-/** "black box" file storage */
-typedef struct CvFileStorage CvFileStorage;
-
-/** Storage flags: */
-#define CV_STORAGE_READ          0
-#define CV_STORAGE_WRITE         1
-#define CV_STORAGE_WRITE_TEXT    CV_STORAGE_WRITE
-#define CV_STORAGE_WRITE_BINARY  CV_STORAGE_WRITE
-#define CV_STORAGE_APPEND        2
-#define CV_STORAGE_MEMORY        4
-#define CV_STORAGE_FORMAT_MASK   (7<<3)
-#define CV_STORAGE_FORMAT_AUTO   0
-#define CV_STORAGE_FORMAT_XML    8
-#define CV_STORAGE_FORMAT_YAML  16
-#define CV_STORAGE_FORMAT_JSON  24
-#define CV_STORAGE_BASE64       64
-#define CV_STORAGE_WRITE_BASE64  (CV_STORAGE_BASE64 | CV_STORAGE_WRITE)
-
-/** @brief List of attributes. :
-
-In the current implementation, attributes are used to pass extra parameters when writing user
-objects (see cvWrite). XML attributes inside tags are not supported, aside from the object type
-specification (type_id attribute).
-@see cvAttrList, cvAttrValue
- */
-typedef struct CvAttrList
-{
-    const char** attr;         /**< NULL-terminated array of (attribute_name,attribute_value) pairs. */
-    struct CvAttrList* next;   /**< Pointer to next chunk of the attributes list.                    */
-}
-CvAttrList;
-
-/** initializes CvAttrList structure */
-CV_INLINE CvAttrList cvAttrList( const char** attr CV_DEFAULT(NULL),
-                                 CvAttrList* next CV_DEFAULT(NULL) )
-{
-    CvAttrList l;
-    l.attr = attr;
-    l.next = next;
-
-    return l;
-}
-
-struct CvTypeInfo;
-
-#define CV_NODE_NONE        0
-#define CV_NODE_INT         1
-#define CV_NODE_INTEGER     CV_NODE_INT
-#define CV_NODE_REAL        2
-#define CV_NODE_FLOAT       CV_NODE_REAL
-#define CV_NODE_STR         3
-#define CV_NODE_STRING      CV_NODE_STR
-#define CV_NODE_REF         4 /**< not used */
-#define CV_NODE_SEQ         5
-#define CV_NODE_MAP         6
-#define CV_NODE_TYPE_MASK   7
-
-#define CV_NODE_TYPE(flags)  ((flags) & CV_NODE_TYPE_MASK)
-
-/** file node flags */
-#define CV_NODE_FLOW        8 /**<Used only for writing structures in YAML format. */
-#define CV_NODE_USER        16
-#define CV_NODE_EMPTY       32
-#define CV_NODE_NAMED       64
-
-#define CV_NODE_IS_INT(flags)        (CV_NODE_TYPE(flags) == CV_NODE_INT)
-#define CV_NODE_IS_REAL(flags)       (CV_NODE_TYPE(flags) == CV_NODE_REAL)
-#define CV_NODE_IS_STRING(flags)     (CV_NODE_TYPE(flags) == CV_NODE_STRING)
-#define CV_NODE_IS_SEQ(flags)        (CV_NODE_TYPE(flags) == CV_NODE_SEQ)
-#define CV_NODE_IS_MAP(flags)        (CV_NODE_TYPE(flags) == CV_NODE_MAP)
-#define CV_NODE_IS_COLLECTION(flags) (CV_NODE_TYPE(flags) >= CV_NODE_SEQ)
-#define CV_NODE_IS_FLOW(flags)       (((flags) & CV_NODE_FLOW) != 0)
-#define CV_NODE_IS_EMPTY(flags)      (((flags) & CV_NODE_EMPTY) != 0)
-#define CV_NODE_IS_USER(flags)       (((flags) & CV_NODE_USER) != 0)
-#define CV_NODE_HAS_NAME(flags)      (((flags) & CV_NODE_NAMED) != 0)
-
-#define CV_NODE_SEQ_SIMPLE 256
-#define CV_NODE_SEQ_IS_SIMPLE(seq) (((seq)->flags & CV_NODE_SEQ_SIMPLE) != 0)
-
-typedef struct CvString
-{
-    int len;
-    char* ptr;
-}
-CvString;
-
-/** All the keys (names) of elements in the read file storage
-   are stored in the hash to speed up the lookup operations: */
-typedef struct CvStringHashNode
-{
-    unsigned hashval;
-    CvString str;
-    struct CvStringHashNode* next;
-}
-CvStringHashNode;
-
-typedef struct CvGenericHash CvFileNodeHash;
-
-/** Basic element of the file storage - scalar or collection: */
-typedef struct CvFileNode
-{
-    int tag;
-    struct CvTypeInfo* info; /**< type information
-            (only for user-defined object, for others it is 0) */
-    union
-    {
-        double f; /**< scalar floating-point number */
-        int i;    /**< scalar integer number */
-        CvString str; /**< text string */
-        CvSeq* seq; /**< sequence (ordered collection of file nodes) */
-        CvFileNodeHash* map; /**< map (collection of named file nodes) */
-    } data;
-}
-CvFileNode;
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-typedef int (CV_CDECL *CvIsInstanceFunc)( const void* struct_ptr );
-typedef void (CV_CDECL *CvReleaseFunc)( void** struct_dblptr );
-typedef void* (CV_CDECL *CvReadFunc)( CvFileStorage* storage, CvFileNode* node );
-typedef void (CV_CDECL *CvWriteFunc)( CvFileStorage* storage, const char* name,
-                                      const void* struct_ptr, CvAttrList attributes );
-typedef void* (CV_CDECL *CvCloneFunc)( const void* struct_ptr );
-#ifdef __cplusplus
-}
-#endif
-
-/** @brief Type information
-
-The structure contains information about one of the standard or user-defined types. Instances of the
-type may or may not contain a pointer to the corresponding CvTypeInfo structure. In any case, there
-is a way to find the type info structure for a given object using the cvTypeOf function.
-Alternatively, type info can be found by type name using cvFindType, which is used when an object
-is read from file storage. The user can register a new type with cvRegisterType that adds the type
-information structure into the beginning of the type list. Thus, it is possible to create
-specialized types from generic standard types and override the basic methods.
- */
-typedef struct CvTypeInfo
-{
-    int flags; /**< not used */
-    int header_size; /**< sizeof(CvTypeInfo) */
-    struct CvTypeInfo* prev; /**< previous registered type in the list */
-    struct CvTypeInfo* next; /**< next registered type in the list */
-    const char* type_name; /**< type name, written to file storage */
-    CvIsInstanceFunc is_instance; /**< checks if the passed object belongs to the type */
-    CvReleaseFunc release; /**< releases object (memory etc.) */
-    CvReadFunc read; /**< reads object from file storage */
-    CvWriteFunc write; /**< writes object to file storage */
-    CvCloneFunc clone; /**< creates a copy of the object */
-}
-CvTypeInfo;
-#endif
-
-/** @} */
-
-#endif /*OPENCV_CORE_TYPES_H*/
-
-/* End of file. */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/utility.hpp b/projects/llm_framework/include/opencv4/opencv2/core/utility.hpp
deleted file mode 100644
index 108c0d93..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/utility.hpp
+++ /dev/null
@@ -1,1229 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Copyright (C) 2015, Itseez Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_CORE_UTILITY_H
-#define OPENCV_CORE_UTILITY_H
-
-#ifndef __cplusplus
-#  error utility.hpp header must be compiled as C++
-#endif
-
-#if defined(check)
-#  warning Detected Apple 'check' macro definition, it can cause build conflicts. Please, include this header before any Apple headers.
-#endif
-
-#include "opencv2/core.hpp"
-#include <ostream>
-
-#include <functional>
-
-#if !defined(_M_CEE)
-#include <mutex>  // std::mutex, std::lock_guard
-#endif
-
-namespace cv
-{
-
-//! @addtogroup core_utils
-//! @{
-
-/** @brief  Automatically Allocated Buffer Class
-
- The class is used for temporary buffers in functions and methods.
- If a temporary buffer is usually small (a few K's of memory),
- but its size depends on the parameters, it makes sense to create a small
- fixed-size array on stack and use it if it's large enough. If the required buffer size
- is larger than the fixed size, another buffer of sufficient size is allocated dynamically
- and released after the processing. Therefore, in typical cases, when the buffer size is small,
- there is no overhead associated with malloc()/free().
- At the same time, there is no limit on the size of processed data.
-
- This is what AutoBuffer does. The template takes 2 parameters - type of the buffer elements and
- the number of stack-allocated elements. Here is how the class is used:
-
- \code
- void my_func(const cv::Mat& m)
- {
-    cv::AutoBuffer<float> buf(1000); // create automatic buffer containing 1000 floats
-
-    buf.allocate(m.rows); // if m.rows <= 1000, the pre-allocated buffer is used,
-                          // otherwise the buffer of "m.rows" floats will be allocated
-                          // dynamically and deallocated in cv::AutoBuffer destructor
-    ...
- }
- \endcode
-*/
-#ifdef OPENCV_ENABLE_MEMORY_SANITIZER
-template<typename _Tp, size_t fixed_size = 0> class AutoBuffer
-#else
-template<typename _Tp, size_t fixed_size = 1024/sizeof(_Tp)+8> class AutoBuffer
-#endif
-{
-public:
-    typedef _Tp value_type;
-
-    //! the default constructor
-    AutoBuffer();
-    //! constructor taking the real buffer size
-    explicit AutoBuffer(size_t _size);
-
-    //! the copy constructor
-    AutoBuffer(const AutoBuffer<_Tp, fixed_size>& buf);
-    //! the assignment operator
-    AutoBuffer<_Tp, fixed_size>& operator = (const AutoBuffer<_Tp, fixed_size>& buf);
-
-    //! destructor. calls deallocate()
-    ~AutoBuffer();
-
-    //! allocates the new buffer of size _size. if the _size is small enough, stack-allocated buffer is used
-    void allocate(size_t _size);
-    //! deallocates the buffer if it was dynamically allocated
-    void deallocate();
-    //! resizes the buffer and preserves the content
-    void resize(size_t _size);
-    //! returns the current buffer size
-    size_t size() const;
-    //! returns pointer to the real buffer, stack-allocated or heap-allocated
-    inline _Tp* data() { return ptr; }
-    //! returns read-only pointer to the real buffer, stack-allocated or heap-allocated
-    inline const _Tp* data() const { return ptr; }
-
-#if !defined(OPENCV_DISABLE_DEPRECATED_COMPATIBILITY) // use to .data() calls instead
-    //! returns pointer to the real buffer, stack-allocated or heap-allocated
-    operator _Tp* () { return ptr; }
-    //! returns read-only pointer to the real buffer, stack-allocated or heap-allocated
-    operator const _Tp* () const { return ptr; }
-#else
-    //! returns a reference to the element at specified location. No bounds checking is performed in Release builds.
-    inline _Tp& operator[] (size_t i) { CV_DbgCheckLT(i, sz, "out of range"); return ptr[i]; }
-    //! returns a reference to the element at specified location. No bounds checking is performed in Release builds.
-    inline const _Tp& operator[] (size_t i) const { CV_DbgCheckLT(i, sz, "out of range"); return ptr[i]; }
-#endif
-
-protected:
-    //! pointer to the real buffer, can point to buf if the buffer is small enough
-    _Tp* ptr;
-    //! size of the real buffer
-    size_t sz;
-    //! pre-allocated buffer. At least 1 element to confirm C++ standard requirements
-    _Tp buf[(fixed_size > 0) ? fixed_size : 1];
-};
-
-/**  @brief Sets/resets the break-on-error mode.
-
-When the break-on-error mode is set, the default error handler issues a hardware exception, which
-can make debugging more convenient.
-
-\return the previous state
- */
-CV_EXPORTS bool setBreakOnError(bool flag);
-
-extern "C" typedef int (*ErrorCallback)( int status, const char* func_name,
-                                       const char* err_msg, const char* file_name,
-                                       int line, void* userdata );
-
-
-/** @brief Sets the new error handler and the optional user data.
-
-  The function sets the new error handler, called from cv::error().
-
-  \param errCallback the new error handler. If NULL, the default error handler is used.
-  \param userdata the optional user data pointer, passed to the callback.
-  \param prevUserdata the optional output parameter where the previous user data pointer is stored
-
-  \return the previous error handler
-*/
-CV_EXPORTS ErrorCallback redirectError( ErrorCallback errCallback, void* userdata=0, void** prevUserdata=0);
-
-CV_EXPORTS String tempfile( const char* suffix = 0);
-CV_EXPORTS void glob(String pattern, std::vector<String>& result, bool recursive = false);
-
-/** @brief OpenCV will try to set the number of threads for the next parallel region.
-
-If threads == 0, OpenCV will disable threading optimizations and run all it's functions
-sequentially. Passing threads \< 0 will reset threads number to system default. This function must
-be called outside of parallel region.
-
-OpenCV will try to run its functions with specified threads number, but some behaviour differs from
-framework:
--   `TBB` - User-defined parallel constructions will run with the same threads number, if
-    another is not specified. If later on user creates his own scheduler, OpenCV will use it.
--   `OpenMP` - No special defined behaviour.
--   `Concurrency` - If threads == 1, OpenCV will disable threading optimizations and run its
-    functions sequentially.
--   `GCD` - Supports only values \<= 0.
--   `C=` - No special defined behaviour.
-@param nthreads Number of threads used by OpenCV.
-@sa getNumThreads, getThreadNum
- */
-CV_EXPORTS_W void setNumThreads(int nthreads);
-
-/** @brief Returns the number of threads used by OpenCV for parallel regions.
-
-Always returns 1 if OpenCV is built without threading support.
-
-The exact meaning of return value depends on the threading framework used by OpenCV library:
-- `TBB` - The number of threads, that OpenCV will try to use for parallel regions. If there is
-  any tbb::thread_scheduler_init in user code conflicting with OpenCV, then function returns
-  default number of threads used by TBB library.
-- `OpenMP` - An upper bound on the number of threads that could be used to form a new team.
-- `Concurrency` - The number of threads, that OpenCV will try to use for parallel regions.
-- `GCD` - Unsupported; returns the GCD thread pool limit (512) for compatibility.
-- `C=` - The number of threads, that OpenCV will try to use for parallel regions, if before
-  called setNumThreads with threads \> 0, otherwise returns the number of logical CPUs,
-  available for the process.
-@sa setNumThreads, getThreadNum
- */
-CV_EXPORTS_W int getNumThreads();
-
-/** @brief Returns the index of the currently executed thread within the current parallel region. Always
-returns 0 if called outside of parallel region.
-
-@deprecated Current implementation doesn't corresponding to this documentation.
-
-The exact meaning of the return value depends on the threading framework used by OpenCV library:
-- `TBB` - Unsupported with current 4.1 TBB release. Maybe will be supported in future.
-- `OpenMP` - The thread number, within the current team, of the calling thread.
-- `Concurrency` - An ID for the virtual processor that the current context is executing on (0
-  for master thread and unique number for others, but not necessary 1,2,3,...).
-- `GCD` - System calling thread's ID. Never returns 0 inside parallel region.
-- `C=` - The index of the current parallel task.
-@sa setNumThreads, getNumThreads
- */
-CV_EXPORTS_W int getThreadNum();
-
-/** @brief Returns full configuration time cmake output.
-
-Returned value is raw cmake output including version control system revision, compiler version,
-compiler flags, enabled modules and third party libraries, etc. Output format depends on target
-architecture.
- */
-CV_EXPORTS_W const String& getBuildInformation();
-
-/** @brief Returns library version string
-
-For example "3.4.1-dev".
-
-@sa getMajorVersion, getMinorVersion, getRevisionVersion
-*/
-CV_EXPORTS_W String getVersionString();
-
-/** @brief Returns major library version */
-CV_EXPORTS_W int getVersionMajor();
-
-/** @brief Returns minor library version */
-CV_EXPORTS_W int getVersionMinor();
-
-/** @brief Returns revision field of the library version */
-CV_EXPORTS_W int getVersionRevision();
-
-/** @brief Returns the number of ticks.
-
-The function returns the number of ticks after the certain event (for example, when the machine was
-turned on). It can be used to initialize RNG or to measure a function execution time by reading the
-tick count before and after the function call.
-@sa getTickFrequency, TickMeter
- */
-CV_EXPORTS_W int64 getTickCount();
-
-/** @brief Returns the number of ticks per second.
-
-The function returns the number of ticks per second. That is, the following code computes the
-execution time in seconds:
-@code
-    double t = (double)getTickCount();
-    // do something ...
-    t = ((double)getTickCount() - t)/getTickFrequency();
-@endcode
-@sa getTickCount, TickMeter
- */
-CV_EXPORTS_W double getTickFrequency();
-
-/** @brief a Class to measure passing time.
-
-The class computes passing time by counting the number of ticks per second. That is, the following code computes the
-execution time in seconds:
-@snippet snippets/core_various.cpp TickMeter_total
-
-It is also possible to compute the average time over multiple runs:
-@snippet snippets/core_various.cpp TickMeter_average
-
-@sa getTickCount, getTickFrequency
-*/
-class CV_EXPORTS_W TickMeter
-{
-public:
-    //! the default constructor
-    CV_WRAP TickMeter()
-    {
-        reset();
-    }
-
-    //! starts counting ticks.
-    CV_WRAP void start()
-    {
-        startTime = cv::getTickCount();
-    }
-
-    //! stops counting ticks.
-    CV_WRAP void stop()
-    {
-        int64 time = cv::getTickCount();
-        if (startTime == 0)
-            return;
-        ++counter;
-        sumTime += (time - startTime);
-        startTime = 0;
-    }
-
-    //! returns counted ticks.
-    CV_WRAP int64 getTimeTicks() const
-    {
-        return sumTime;
-    }
-
-    //! returns passed time in microseconds.
-    CV_WRAP double getTimeMicro() const
-    {
-        return getTimeMilli()*1e3;
-    }
-
-    //! returns passed time in milliseconds.
-    CV_WRAP double getTimeMilli() const
-    {
-        return getTimeSec()*1e3;
-    }
-
-    //! returns passed time in seconds.
-    CV_WRAP double getTimeSec()   const
-    {
-        return (double)getTimeTicks() / getTickFrequency();
-    }
-
-    //! returns internal counter value.
-    CV_WRAP int64 getCounter() const
-    {
-        return counter;
-    }
-
-    //! returns average FPS (frames per second) value.
-    CV_WRAP double getFPS() const
-    {
-        const double sec = getTimeSec();
-        if (sec < DBL_EPSILON)
-            return 0.;
-        return counter / sec;
-    }
-
-    //! returns average time in seconds
-    CV_WRAP double getAvgTimeSec() const
-    {
-        if (counter <= 0)
-            return 0.;
-        return getTimeSec() / counter;
-    }
-
-    //! returns average time in milliseconds
-    CV_WRAP double getAvgTimeMilli() const
-    {
-        return getAvgTimeSec() * 1e3;
-    }
-
-    //! resets internal values.
-    CV_WRAP void reset()
-    {
-        startTime = 0;
-        sumTime = 0;
-        counter = 0;
-    }
-
-private:
-    int64 counter;
-    int64 sumTime;
-    int64 startTime;
-};
-
-/** @brief output operator
-@code
-TickMeter tm;
-tm.start();
-// do something ...
-tm.stop();
-std::cout << tm;
-@endcode
-*/
-
-static inline
-std::ostream& operator << (std::ostream& out, const TickMeter& tm)
-{
-    return out << tm.getTimeSec() << "sec";
-}
-
-/** @brief Returns the number of CPU ticks.
-
-The function returns the current number of CPU ticks on some architectures (such as x86, x64,
-PowerPC). On other platforms the function is equivalent to getTickCount. It can also be used for
-very accurate time measurements, as well as for RNG initialization. Note that in case of multi-CPU
-systems a thread, from which getCPUTickCount is called, can be suspended and resumed at another CPU
-with its own counter. So, theoretically (and practically) the subsequent calls to the function do
-not necessary return the monotonously increasing values. Also, since a modern CPU varies the CPU
-frequency depending on the load, the number of CPU clocks spent in some code cannot be directly
-converted to time units. Therefore, getTickCount is generally a preferable solution for measuring
-execution time.
- */
-CV_EXPORTS_W int64 getCPUTickCount();
-
-/** @brief Returns true if the specified feature is supported by the host hardware.
-
-The function returns true if the host hardware supports the specified feature. When user calls
-setUseOptimized(false), the subsequent calls to checkHardwareSupport() will return false until
-setUseOptimized(true) is called. This way user can dynamically switch on and off the optimized code
-in OpenCV.
-@param feature The feature of interest, one of cv::CpuFeatures
- */
-CV_EXPORTS_W bool checkHardwareSupport(int feature);
-
-/** @brief Returns feature name by ID
-
-Returns empty string if feature is not defined
-*/
-CV_EXPORTS_W String getHardwareFeatureName(int feature);
-
-/** @brief Returns list of CPU features enabled during compilation.
-
-Returned value is a string containing space separated list of CPU features with following markers:
-
-- no markers - baseline features
-- prefix `*` - features enabled in dispatcher
-- suffix `?` - features enabled but not available in HW
-
-Example: `SSE SSE2 SSE3 *SSE4.1 *SSE4.2 *FP16 *AVX *AVX2 *AVX512-SKX?`
-*/
-CV_EXPORTS_W std::string getCPUFeaturesLine();
-
-/** @brief Returns the number of logical CPUs available for the process.
- */
-CV_EXPORTS_W int getNumberOfCPUs();
-
-
-/** @brief Aligns a pointer to the specified number of bytes.
-
-The function returns the aligned pointer of the same type as the input pointer:
-\f[\texttt{(_Tp*)(((size_t)ptr + n-1) & -n)}\f]
-@param ptr Aligned pointer.
-@param n Alignment size that must be a power of two.
- */
-template<typename _Tp> static inline _Tp* alignPtr(_Tp* ptr, int n=(int)sizeof(_Tp))
-{
-    CV_DbgAssert((n & (n - 1)) == 0); // n is a power of 2
-    return (_Tp*)(((size_t)ptr + n-1) & -n);
-}
-
-/** @brief Aligns a buffer size to the specified number of bytes.
-
-The function returns the minimum number that is greater than or equal to sz and is divisible by n :
-\f[\texttt{(sz + n-1) & -n}\f]
-@param sz Buffer size to align.
-@param n Alignment size that must be a power of two.
- */
-static inline size_t alignSize(size_t sz, int n)
-{
-    CV_DbgAssert((n & (n - 1)) == 0); // n is a power of 2
-    return (sz + n-1) & -n;
-}
-
-/** @brief Integer division with result round up.
-
-Use this function instead of `ceil((float)a / b)` expressions.
-
-@sa alignSize
-*/
-static inline int divUp(int a, unsigned int b)
-{
-    CV_DbgAssert(a >= 0);
-    return (a + b - 1) / b;
-}
-/** @overload */
-static inline size_t divUp(size_t a, unsigned int b)
-{
-    return (a + b - 1) / b;
-}
-
-/** @brief Round first value up to the nearest multiple of second value.
-
-Use this function instead of `ceil((float)a / b) * b` expressions.
-
-@sa divUp
-*/
-static inline int roundUp(int a, unsigned int b)
-{
-    CV_DbgAssert(a >= 0);
-    return a + b - 1 - (a + b -1) % b;
-}
-/** @overload */
-static inline size_t roundUp(size_t a, unsigned int b)
-{
-    return a + b - 1 - (a + b - 1) % b;
-}
-
-/** @brief Alignment check of passed values
-
-Usage: `isAligned<sizeof(int)>(...)`
-
-@note Alignment(N) must be a power of 2 (2**k, 2^k)
-*/
-template<int N, typename T> static inline
-bool isAligned(const T& data)
-{
-    CV_StaticAssert((N & (N - 1)) == 0, "");  // power of 2
-    return (((size_t)data) & (N - 1)) == 0;
-}
-/** @overload */
-template<int N> static inline
-bool isAligned(const void* p1)
-{
-    return isAligned<N>((size_t)p1);
-}
-/** @overload */
-template<int N> static inline
-bool isAligned(const void* p1, const void* p2)
-{
-    return isAligned<N>(((size_t)p1)|((size_t)p2));
-}
-/** @overload */
-template<int N> static inline
-bool isAligned(const void* p1, const void* p2, const void* p3)
-{
-    return isAligned<N>(((size_t)p1)|((size_t)p2)|((size_t)p3));
-}
-/** @overload */
-template<int N> static inline
-bool isAligned(const void* p1, const void* p2, const void* p3, const void* p4)
-{
-    return isAligned<N>(((size_t)p1)|((size_t)p2)|((size_t)p3)|((size_t)p4));
-}
-
-/** @brief Enables or disables the optimized code.
-
-The function can be used to dynamically turn on and off optimized dispatched code (code that uses SSE4.2, AVX/AVX2,
-and other instructions on the platforms that support it). It sets a global flag that is further
-checked by OpenCV functions. Since the flag is not checked in the inner OpenCV loops, it is only
-safe to call the function on the very top level in your application where you can be sure that no
-other OpenCV function is currently executed.
-
-By default, the optimized code is enabled unless you disable it in CMake. The current status can be
-retrieved using useOptimized.
-@param onoff The boolean flag specifying whether the optimized code should be used (onoff=true)
-or not (onoff=false).
- */
-CV_EXPORTS_W void setUseOptimized(bool onoff);
-
-/** @brief Returns the status of optimized code usage.
-
-The function returns true if the optimized code is enabled. Otherwise, it returns false.
- */
-CV_EXPORTS_W bool useOptimized();
-
-static inline size_t getElemSize(int type) { return (size_t)CV_ELEM_SIZE(type); }
-
-/////////////////////////////// Parallel Primitives //////////////////////////////////
-
-/** @brief Base class for parallel data processors
-
-@ingroup core_parallel
-*/
-class CV_EXPORTS ParallelLoopBody
-{
-public:
-    virtual ~ParallelLoopBody();
-    virtual void operator() (const Range& range) const = 0;
-};
-
-/** @brief Parallel data processor
-
-@ingroup core_parallel
-*/
-CV_EXPORTS void parallel_for_(const Range& range, const ParallelLoopBody& body, double nstripes=-1.);
-
-//! @ingroup core_parallel
-class ParallelLoopBodyLambdaWrapper : public ParallelLoopBody
-{
-private:
-    std::function<void(const Range&)> m_functor;
-public:
-    inline
-    ParallelLoopBodyLambdaWrapper(std::function<void(const Range&)> functor)
-        : m_functor(functor)
-    {
-        // nothing
-    }
-
-    virtual void operator() (const cv::Range& range) const CV_OVERRIDE
-    {
-        m_functor(range);
-    }
-};
-
-//! @ingroup core_parallel
-static inline
-void parallel_for_(const Range& range, std::function<void(const Range&)> functor, double nstripes=-1.)
-{
-    parallel_for_(range, ParallelLoopBodyLambdaWrapper(functor), nstripes);
-}
-
-
-/////////////////////////////// forEach method of cv::Mat ////////////////////////////
-template<typename _Tp, typename Functor> inline
-void Mat::forEach_impl(const Functor& operation) {
-    if (false) {
-        operation(*reinterpret_cast<_Tp*>(0), reinterpret_cast<int*>(0));
-        // If your compiler fails in this line.
-        // Please check that your functor signature is
-        //     (_Tp&, const int*)   <- multi-dimensional
-        //  or (_Tp&, void*)        <- in case you don't need current idx.
-    }
-
-    CV_Assert(!empty());
-    CV_Assert(this->total() / this->size[this->dims - 1] <= INT_MAX);
-    const int LINES = static_cast<int>(this->total() / this->size[this->dims - 1]);
-
-    class PixelOperationWrapper :public ParallelLoopBody
-    {
-    public:
-        PixelOperationWrapper(Mat_<_Tp>* const frame, const Functor& _operation)
-            : mat(frame), op(_operation) {}
-        virtual ~PixelOperationWrapper(){}
-        // ! Overloaded virtual operator
-        // convert range call to row call.
-        virtual void operator()(const Range &range) const CV_OVERRIDE
-        {
-            const int DIMS = mat->dims;
-            const int COLS = mat->size[DIMS - 1];
-            if (DIMS <= 2) {
-                for (int row = range.start; row < range.end; ++row) {
-                    this->rowCall2(row, COLS);
-                }
-            } else {
-                std::vector<int> idx(DIMS); /// idx is modified in this->rowCall
-                idx[DIMS - 2] = range.start - 1;
-
-                for (int line_num = range.start; line_num < range.end; ++line_num) {
-                    idx[DIMS - 2]++;
-                    for (int i = DIMS - 2; i >= 0; --i) {
-                        if (idx[i] >= mat->size[i]) {
-                            idx[i - 1] += idx[i] / mat->size[i];
-                            idx[i] %= mat->size[i];
-                            continue; // carry-over;
-                        }
-                        else {
-                            break;
-                        }
-                    }
-                    this->rowCall(&idx[0], COLS, DIMS);
-                }
-            }
-        }
-    private:
-        Mat_<_Tp>* const mat;
-        const Functor op;
-        // ! Call operator for each elements in this row.
-        inline void rowCall(int* const idx, const int COLS, const int DIMS) const {
-            int &col = idx[DIMS - 1];
-            col = 0;
-            _Tp* pixel = &(mat->template at<_Tp>(idx));
-
-            while (col < COLS) {
-                op(*pixel, const_cast<const int*>(idx));
-                pixel++; col++;
-            }
-            col = 0;
-        }
-        // ! Call operator for each elements in this row. 2d mat special version.
-        inline void rowCall2(const int row, const int COLS) const {
-            union Index{
-                int body[2];
-                operator const int*() const {
-                    return reinterpret_cast<const int*>(this);
-                }
-                int& operator[](const int i) {
-                    return body[i];
-                }
-            } idx = {{row, 0}};
-            // Special union is needed to avoid
-            // "error: array subscript is above array bounds [-Werror=array-bounds]"
-            // when call the functor `op` such that access idx[3].
-
-            _Tp* pixel = &(mat->template at<_Tp>(idx));
-            const _Tp* const pixel_end = pixel + COLS;
-            while(pixel < pixel_end) {
-                op(*pixel++, static_cast<const int*>(idx));
-                idx[1]++;
-            }
-        }
-        PixelOperationWrapper& operator=(const PixelOperationWrapper &) {
-            CV_Assert(false);
-            // We can not remove this implementation because Visual Studio warning C4822.
-            return *this;
-        }
-    };
-
-    parallel_for_(cv::Range(0, LINES), PixelOperationWrapper(reinterpret_cast<Mat_<_Tp>*>(this), operation));
-}
-
-/////////////////////////// Synchronization Primitives ///////////////////////////////
-
-#if !defined(_M_CEE)
-#ifndef OPENCV_DISABLE_THREAD_SUPPORT
-typedef std::recursive_mutex Mutex;
-typedef std::lock_guard<cv::Mutex> AutoLock;
-#else // OPENCV_DISABLE_THREAD_SUPPORT
-// Custom (failing) implementation of `std::recursive_mutex`.
-struct Mutex {
-    void lock(){
-        CV_Error(cv::Error::StsNotImplemented,
-                 "cv::Mutex is disabled by OPENCV_DISABLE_THREAD_SUPPORT=ON");
-    }
-    void unlock(){
-        CV_Error(cv::Error::StsNotImplemented,
-                 "cv::Mutex is disabled by OPENCV_DISABLE_THREAD_SUPPORT=ON");
-    }
-};
-// Stub for cv::AutoLock when threads are disabled.
-struct AutoLock {
-    AutoLock(Mutex &) { }
-};
-#endif // OPENCV_DISABLE_THREAD_SUPPORT
-#endif // !defined(_M_CEE)
-
-
-/** @brief Designed for command line parsing
-
-The sample below demonstrates how to use CommandLineParser:
-@code
-    CommandLineParser parser(argc, argv, keys);
-    parser.about("Application name v1.0.0");
-
-    if (parser.has("help"))
-    {
-        parser.printMessage();
-        return 0;
-    }
-
-    int N = parser.get<int>("N");
-    double fps = parser.get<double>("fps");
-    String path = parser.get<String>("path");
-
-    use_time_stamp = parser.has("timestamp");
-
-    String img1 = parser.get<String>(0);
-    String img2 = parser.get<String>(1);
-
-    int repeat = parser.get<int>(2);
-
-    if (!parser.check())
-    {
-        parser.printErrors();
-        return 0;
-    }
-@endcode
-
-### Keys syntax
-
-The keys parameter is a string containing several blocks, each one is enclosed in curly braces and
-describes one argument. Each argument contains three parts separated by the `|` symbol:
-
--# argument names is a space-separated list of option synonyms (to mark argument as positional, prefix it with the `@` symbol)
--# default value will be used if the argument was not provided (can be empty)
--# help message (can be empty)
-
-For example:
-
-@code{.cpp}
-    const String keys =
-        "{help h usage ? |      | print this message   }"
-        "{@image1        |      | image1 for compare   }"
-        "{@image2        |<none>| image2 for compare   }"
-        "{@repeat        |1     | number               }"
-        "{path           |.     | path to file         }"
-        "{fps            | -1.0 | fps for output video }"
-        "{N count        |100   | count of objects     }"
-        "{ts timestamp   |      | use time stamp       }"
-        ;
-}
-@endcode
-
-Note that there are no default values for `help` and `timestamp` so we can check their presence using the `has()` method.
-Arguments with default values are considered to be always present. Use the `get()` method in these cases to check their
-actual value instead.
-
-String keys like `get<String>("@image1")` return the empty string `""` by default - even with an empty default value.
-Use the special `<none>` default value to enforce that the returned string must not be empty. (like in `get<String>("@image2")`)
-
-### Usage
-
-For the described keys:
-
-@code{.sh}
-    # Good call (3 positional parameters: image1, image2 and repeat; N is 200, ts is true)
-    $ ./app -N=200 1.png 2.jpg 19 -ts
-
-    # Bad call
-    $ ./app -fps=aaa
-    ERRORS:
-    Parameter 'fps': can not convert: [aaa] to [double]
-@endcode
- */
-class CV_EXPORTS CommandLineParser
-{
-public:
-
-    /** @brief Constructor
-
-    Initializes command line parser object
-
-    @param argc number of command line arguments (from main())
-    @param argv array of command line arguments (from main())
-    @param keys string describing acceptable command line parameters (see class description for syntax)
-    */
-    CommandLineParser(int argc, const char* const argv[], const String& keys);
-
-    /** @brief Copy constructor */
-    CommandLineParser(const CommandLineParser& parser);
-
-    /** @brief Assignment operator */
-    CommandLineParser& operator = (const CommandLineParser& parser);
-
-    /** @brief Destructor */
-    ~CommandLineParser();
-
-    /** @brief Returns application path
-
-    This method returns the path to the executable from the command line (`argv[0]`).
-
-    For example, if the application has been started with such a command:
-    @code{.sh}
-    $ ./bin/my-executable
-    @endcode
-    this method will return `./bin`.
-    */
-    String getPathToApplication() const;
-
-    /** @brief Access arguments by name
-
-    Returns argument converted to selected type. If the argument is not known or can not be
-    converted to selected type, the error flag is set (can be checked with @ref check).
-
-    For example, define:
-    @code{.cpp}
-    String keys = "{N count||}";
-    @endcode
-
-    Call:
-    @code{.sh}
-    $ ./my-app -N=20
-    # or
-    $ ./my-app --count=20
-    @endcode
-
-    Access:
-    @code{.cpp}
-    int N = parser.get<int>("N");
-    @endcode
-
-    @param name name of the argument
-    @param space_delete remove spaces from the left and right of the string
-    @tparam T the argument will be converted to this type if possible
-
-    @note You can access positional arguments by their `@`-prefixed name:
-    @code{.cpp}
-    parser.get<String>("@image");
-    @endcode
-     */
-    template <typename T>
-    T get(const String& name, bool space_delete = true) const
-    {
-        T val = T();
-        getByName(name, space_delete, ParamType<T>::type, (void*)&val);
-        return val;
-    }
-
-    /** @brief Access positional arguments by index
-
-    Returns argument converted to selected type. Indexes are counted from zero.
-
-    For example, define:
-    @code{.cpp}
-    String keys = "{@arg1||}{@arg2||}"
-    @endcode
-
-    Call:
-    @code{.sh}
-    ./my-app abc qwe
-    @endcode
-
-    Access arguments:
-    @code{.cpp}
-    String val_1 = parser.get<String>(0); // returns "abc", arg1
-    String val_2 = parser.get<String>(1); // returns "qwe", arg2
-    @endcode
-
-    @param index index of the argument
-    @param space_delete remove spaces from the left and right of the string
-    @tparam T the argument will be converted to this type if possible
-     */
-    template <typename T>
-    T get(int index, bool space_delete = true) const
-    {
-        T val = T();
-        getByIndex(index, space_delete, ParamType<T>::type, (void*)&val);
-        return val;
-    }
-
-    /** @brief Check if field was provided in the command line
-
-    @param name argument name to check
-    */
-    bool has(const String& name) const;
-
-    /** @brief Check for parsing errors
-
-    Returns false if error occurred while accessing the parameters (bad conversion, missing arguments,
-    etc.). Call @ref printErrors to print error messages list.
-     */
-    bool check() const;
-
-    /** @brief Set the about message
-
-    The about message will be shown when @ref printMessage is called, right before arguments table.
-     */
-    void about(const String& message);
-
-    /** @brief Print help message
-
-    This method will print standard help message containing the about message and arguments description.
-
-    @sa about
-    */
-    void printMessage() const;
-
-    /** @brief Print list of errors occurred
-
-    @sa check
-    */
-    void printErrors() const;
-
-protected:
-    void getByName(const String& name, bool space_delete, Param type, void* dst) const;
-    void getByIndex(int index, bool space_delete, Param type, void* dst) const;
-
-    struct Impl;
-    Impl* impl;
-};
-
-//! @} core_utils
-
-//! @cond IGNORED
-
-/////////////////////////////// AutoBuffer implementation ////////////////////////////////////////
-
-template<typename _Tp, size_t fixed_size> inline
-AutoBuffer<_Tp, fixed_size>::AutoBuffer()
-{
-    ptr = buf;
-    sz = fixed_size;
-}
-
-template<typename _Tp, size_t fixed_size> inline
-AutoBuffer<_Tp, fixed_size>::AutoBuffer(size_t _size)
-{
-    ptr = buf;
-    sz = fixed_size;
-    allocate(_size);
-}
-
-template<typename _Tp, size_t fixed_size> inline
-AutoBuffer<_Tp, fixed_size>::AutoBuffer(const AutoBuffer<_Tp, fixed_size>& abuf )
-{
-    ptr = buf;
-    sz = fixed_size;
-    allocate(abuf.size());
-    for( size_t i = 0; i < sz; i++ )
-        ptr[i] = abuf.ptr[i];
-}
-
-template<typename _Tp, size_t fixed_size> inline AutoBuffer<_Tp, fixed_size>&
-AutoBuffer<_Tp, fixed_size>::operator = (const AutoBuffer<_Tp, fixed_size>& abuf)
-{
-    if( this != &abuf )
-    {
-        deallocate();
-        allocate(abuf.size());
-        for( size_t i = 0; i < sz; i++ )
-            ptr[i] = abuf.ptr[i];
-    }
-    return *this;
-}
-
-template<typename _Tp, size_t fixed_size> inline
-AutoBuffer<_Tp, fixed_size>::~AutoBuffer()
-{ deallocate(); }
-
-template<typename _Tp, size_t fixed_size> inline void
-AutoBuffer<_Tp, fixed_size>::allocate(size_t _size)
-{
-    if(_size <= sz)
-    {
-        sz = _size;
-        return;
-    }
-    deallocate();
-    sz = _size;
-    if(_size > fixed_size)
-    {
-        ptr = new _Tp[_size];
-    }
-}
-
-template<typename _Tp, size_t fixed_size> inline void
-AutoBuffer<_Tp, fixed_size>::deallocate()
-{
-    if( ptr != buf )
-    {
-        delete[] ptr;
-        ptr = buf;
-        sz = fixed_size;
-    }
-}
-
-template<typename _Tp, size_t fixed_size> inline void
-AutoBuffer<_Tp, fixed_size>::resize(size_t _size)
-{
-    if(_size <= sz)
-    {
-        sz = _size;
-        return;
-    }
-    size_t i, prevsize = sz, minsize = MIN(prevsize, _size);
-    _Tp* prevptr = ptr;
-
-    ptr = _size > fixed_size ? new _Tp[_size] : buf;
-    sz = _size;
-
-    if( ptr != prevptr )
-        for( i = 0; i < minsize; i++ )
-            ptr[i] = prevptr[i];
-    for( i = prevsize; i < _size; i++ )
-        ptr[i] = _Tp();
-
-    if( prevptr != buf )
-        delete[] prevptr;
-}
-
-template<typename _Tp, size_t fixed_size> inline size_t
-AutoBuffer<_Tp, fixed_size>::size() const
-{ return sz; }
-
-//! @endcond
-
-
-// Basic Node class for tree building
-template<class OBJECT>
-class CV_EXPORTS Node
-{
-public:
-    Node()
-    {
-        m_pParent  = 0;
-    }
-    Node(OBJECT& payload) : m_payload(payload)
-    {
-        m_pParent  = 0;
-    }
-    ~Node()
-    {
-        removeChilds();
-        if (m_pParent)
-        {
-            int idx = m_pParent->findChild(this);
-            if (idx >= 0)
-                m_pParent->m_childs.erase(m_pParent->m_childs.begin() + idx);
-        }
-    }
-
-    Node<OBJECT>* findChild(OBJECT& payload) const
-    {
-        for(size_t i = 0; i < this->m_childs.size(); i++)
-        {
-            if(this->m_childs[i]->m_payload == payload)
-                return this->m_childs[i];
-        }
-        return NULL;
-    }
-
-    int findChild(Node<OBJECT> *pNode) const
-    {
-        for (size_t i = 0; i < this->m_childs.size(); i++)
-        {
-            if(this->m_childs[i] == pNode)
-                return (int)i;
-        }
-        return -1;
-    }
-
-    void addChild(Node<OBJECT> *pNode)
-    {
-        if(!pNode)
-            return;
-
-        CV_Assert(pNode->m_pParent == 0);
-        pNode->m_pParent = this;
-        this->m_childs.push_back(pNode);
-    }
-
-    void removeChilds()
-    {
-        for(size_t i = 0; i < m_childs.size(); i++)
-        {
-            m_childs[i]->m_pParent = 0; // avoid excessive parent vector trimming
-            delete m_childs[i];
-        }
-        m_childs.clear();
-    }
-
-    int getDepth()
-    {
-        int   count   = 0;
-        Node *pParent = m_pParent;
-        while(pParent) count++, pParent = pParent->m_pParent;
-        return count;
-    }
-
-public:
-    OBJECT                     m_payload;
-    Node<OBJECT>*              m_pParent;
-    std::vector<Node<OBJECT>*> m_childs;
-};
-
-
-namespace samples {
-
-//! @addtogroup core_utils_samples
-// This section describes utility functions for OpenCV samples.
-//
-// @note Implementation of these utilities is not thread-safe.
-//
-//! @{
-
-/** @brief Try to find requested data file
-
-Search directories:
-
-1. Directories passed via `addSamplesDataSearchPath()`
-2. OPENCV_SAMPLES_DATA_PATH_HINT environment variable
-3. OPENCV_SAMPLES_DATA_PATH environment variable
-   If parameter value is not empty and nothing is found then stop searching.
-4. Detects build/install path based on:
-   a. current working directory (CWD)
-   b. and/or binary module location (opencv_core/opencv_world, doesn't work with static linkage)
-5. Scan `<source>/{,data,samples/data}` directories if build directory is detected or the current directory is in source tree.
-6. Scan `<install>/share/OpenCV` directory if install directory is detected.
-
-@see cv::utils::findDataFile
-
-@param relative_path Relative path to data file
-@param required Specify "file not found" handling.
-       If true, function prints information message and raises cv::Exception.
-       If false, function returns empty result
-@param silentMode Disables messages
-@return Returns path (absolute or relative to the current directory) or empty string if file is not found
-*/
-CV_EXPORTS_W cv::String findFile(const cv::String& relative_path, bool required = true, bool silentMode = false);
-
-CV_EXPORTS_W cv::String findFileOrKeep(const cv::String& relative_path, bool silentMode = false);
-
-inline cv::String findFileOrKeep(const cv::String& relative_path, bool silentMode)
-{
-    cv::String res = findFile(relative_path, false, silentMode);
-    if (res.empty())
-        return relative_path;
-    return res;
-}
-
-/** @brief Override search data path by adding new search location
-
-Use this only to override default behavior
-Passed paths are used in LIFO order.
-
-@param path Path to used samples data
-*/
-CV_EXPORTS_W void addSamplesDataSearchPath(const cv::String& path);
-
-/** @brief Append samples search data sub directory
-
-General usage is to add OpenCV modules name (`<opencv_contrib>/modules/<name>/samples/data` -> `<name>/samples/data` + `modules/<name>/samples/data`).
-Passed subdirectories are used in LIFO order.
-
-@param subdir samples data sub directory
-*/
-CV_EXPORTS_W void addSamplesDataSearchSubDirectory(const cv::String& subdir);
-
-//! @}
-} // namespace samples
-
-namespace utils {
-
-CV_EXPORTS int getThreadID();
-
-} // namespace
-
-} //namespace cv
-
-#ifdef CV_COLLECT_IMPL_DATA
-#include "opencv2/core/utils/instrumentation.hpp"
-#else
-/// Collect implementation data on OpenCV function call. Requires ENABLE_IMPL_COLLECTION build option.
-#define CV_IMPL_ADD(impl)
-#endif
-
-#endif //OPENCV_CORE_UTILITY_H
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/utils/allocator_stats.hpp b/projects/llm_framework/include/opencv4/opencv2/core/utils/allocator_stats.hpp
deleted file mode 100644
index 79e93382..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/utils/allocator_stats.hpp
+++ /dev/null
@@ -1,29 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_CORE_ALLOCATOR_STATS_HPP
-#define OPENCV_CORE_ALLOCATOR_STATS_HPP
-
-#include "../cvdef.h"
-
-namespace cv { namespace utils {
-
-class AllocatorStatisticsInterface
-{
-protected:
-    AllocatorStatisticsInterface() {}
-    virtual ~AllocatorStatisticsInterface() {}
-public:
-    virtual uint64_t getCurrentUsage() const = 0;
-    virtual uint64_t getTotalUsage() const = 0;
-    virtual uint64_t getNumberOfAllocations() const = 0;
-    virtual uint64_t getPeakUsage() const = 0;
-
-    /** set peak usage = current usage */
-    virtual void resetPeakUsage() = 0;
-};
-
-}} // namespace
-
-#endif // OPENCV_CORE_ALLOCATOR_STATS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/utils/allocator_stats.impl.hpp b/projects/llm_framework/include/opencv4/opencv2/core/utils/allocator_stats.impl.hpp
deleted file mode 100644
index eb5ecde1..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/utils/allocator_stats.impl.hpp
+++ /dev/null
@@ -1,158 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_CORE_ALLOCATOR_STATS_IMPL_HPP
-#define OPENCV_CORE_ALLOCATOR_STATS_IMPL_HPP
-
-#include "./allocator_stats.hpp"
-
-//#define OPENCV_DISABLE_ALLOCATOR_STATS
-
-#ifdef CV_CXX11
-
-#include <atomic>
-
-#ifndef OPENCV_ALLOCATOR_STATS_COUNTER_TYPE
-#if defined(__GNUC__) && (\
-        (defined(__SIZEOF_POINTER__) && __SIZEOF_POINTER__ == 4) || \
-        (defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4) && !defined(__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8)) \
-    )
-#define OPENCV_ALLOCATOR_STATS_COUNTER_TYPE int
-#endif
-#endif
-
-#ifndef OPENCV_ALLOCATOR_STATS_COUNTER_TYPE
-#define OPENCV_ALLOCATOR_STATS_COUNTER_TYPE long long
-#endif
-
-#else  // CV_CXX11
-
-#ifndef OPENCV_ALLOCATOR_STATS_COUNTER_TYPE
-#define OPENCV_ALLOCATOR_STATS_COUNTER_TYPE int  // CV_XADD supports int only
-#endif
-
-#endif  // CV_CXX11
-
-namespace cv { namespace utils {
-
-#ifdef CV__ALLOCATOR_STATS_LOG
-namespace {
-#endif
-
-class AllocatorStatistics : public AllocatorStatisticsInterface
-{
-#ifdef OPENCV_DISABLE_ALLOCATOR_STATS
-
-public:
-    AllocatorStatistics() {}
-    ~AllocatorStatistics() CV_OVERRIDE {}
-
-    uint64_t getCurrentUsage() const CV_OVERRIDE { return 0; }
-    uint64_t getTotalUsage() const CV_OVERRIDE { return 0; }
-    uint64_t getNumberOfAllocations() const CV_OVERRIDE { return 0; }
-    uint64_t getPeakUsage() const CV_OVERRIDE { return 0; }
-
-    /** set peak usage = current usage */
-    void resetPeakUsage() CV_OVERRIDE {};
-
-    void onAllocate(size_t /*sz*/) {}
-    void onFree(size_t /*sz*/) {}
-
-#elif defined(CV_CXX11)
-
-protected:
-    typedef OPENCV_ALLOCATOR_STATS_COUNTER_TYPE counter_t;
-    std::atomic<counter_t> curr, total, total_allocs, peak;
-public:
-    AllocatorStatistics() {}
-    ~AllocatorStatistics() CV_OVERRIDE {}
-
-    uint64_t getCurrentUsage() const CV_OVERRIDE { return (uint64_t)curr.load(); }
-    uint64_t getTotalUsage() const CV_OVERRIDE { return (uint64_t)total.load(); }
-    uint64_t getNumberOfAllocations() const CV_OVERRIDE { return (uint64_t)total_allocs.load(); }
-    uint64_t getPeakUsage() const CV_OVERRIDE { return (uint64_t)peak.load(); }
-
-    /** set peak usage = current usage */
-    void resetPeakUsage() CV_OVERRIDE { peak.store(curr.load()); }
-
-    // Controller interface
-    void onAllocate(size_t sz)
-    {
-#ifdef CV__ALLOCATOR_STATS_LOG
-        CV__ALLOCATOR_STATS_LOG(cv::format("allocate: %lld (curr=%lld)", (long long int)sz, (long long int)curr.load()));
-#endif
-
-        counter_t new_curr = curr.fetch_add((counter_t)sz) + (counter_t)sz;
-
-        // peak = std::max((uint64_t)peak, new_curr);
-        auto prev_peak = peak.load();
-        while (prev_peak < new_curr)
-        {
-            if (peak.compare_exchange_weak(prev_peak, new_curr))
-                break;
-        }
-        // end of peak = max(...)
-
-        total += (counter_t)sz;
-        total_allocs++;
-    }
-    void onFree(size_t sz)
-    {
-#ifdef CV__ALLOCATOR_STATS_LOG
-        CV__ALLOCATOR_STATS_LOG(cv::format("free: %lld (curr=%lld)", (long long int)sz, (long long int)curr.load()));
-#endif
-        curr -= (counter_t)sz;
-    }
-
-#else  // non C++11
-
-protected:
-    typedef OPENCV_ALLOCATOR_STATS_COUNTER_TYPE counter_t;
-    volatile counter_t curr, total, total_allocs, peak;  // overflow is possible, CV_XADD operates with 'int' only
-public:
-    AllocatorStatistics()
-        : curr(0), total(0), total_allocs(0), peak(0)
-    {}
-    ~AllocatorStatistics() CV_OVERRIDE {}
-
-    uint64_t getCurrentUsage() const CV_OVERRIDE { return (uint64_t)curr; }
-    uint64_t getTotalUsage() const CV_OVERRIDE { return (uint64_t)total; }
-    uint64_t getNumberOfAllocations() const CV_OVERRIDE { return (uint64_t)total_allocs; }
-    uint64_t getPeakUsage() const CV_OVERRIDE { return (uint64_t)peak; }
-
-    void resetPeakUsage() CV_OVERRIDE { peak = curr; }
-
-    // Controller interface
-    void onAllocate(size_t sz)
-    {
-#ifdef CV__ALLOCATOR_STATS_LOG
-        CV__ALLOCATOR_STATS_LOG(cv::format("allocate: %lld (curr=%lld)", (long long int)sz, (long long int)curr));
-#endif
-
-        counter_t new_curr = (counter_t)CV_XADD(&curr, (counter_t)sz) + (counter_t)sz;
-
-        peak = std::max((counter_t)peak, new_curr);  // non-thread safe
-
-        //CV_XADD(&total, (uint64_t)sz);  // overflow with int, non-reliable...
-        total += sz;
-
-        CV_XADD(&total_allocs, (counter_t)1);
-    }
-    void onFree(size_t sz)
-    {
-#ifdef CV__ALLOCATOR_STATS_LOG
-        CV__ALLOCATOR_STATS_LOG(cv::format("free: %lld (curr=%lld)", (long long int)sz, (long long int)curr));
-#endif
-        CV_XADD(&curr, (counter_t)-sz);
-    }
-#endif
-};
-
-#ifdef CV__ALLOCATOR_STATS_LOG
-} // namespace
-#endif
-
-}} // namespace
-
-#endif // OPENCV_CORE_ALLOCATOR_STATS_IMPL_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/utils/filesystem.hpp b/projects/llm_framework/include/opencv4/opencv2/core/utils/filesystem.hpp
deleted file mode 100644
index a98d2202..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/utils/filesystem.hpp
+++ /dev/null
@@ -1,82 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_UTILS_FILESYSTEM_HPP
-#define OPENCV_UTILS_FILESYSTEM_HPP
-
-namespace cv { namespace utils { namespace fs {
-
-
-CV_EXPORTS bool exists(const cv::String& path);
-CV_EXPORTS bool isDirectory(const cv::String& path);
-
-CV_EXPORTS void remove_all(const cv::String& path);
-
-
-CV_EXPORTS cv::String getcwd();
-
-/** @brief Converts path p to a canonical absolute path
- * Symlinks are processed if there is support for them on running platform.
- *
- * @param path input path. Target file/directory should exist.
- */
-CV_EXPORTS cv::String canonical(const cv::String& path);
-
-/** Join path components */
-CV_EXPORTS cv::String join(const cv::String& base, const cv::String& path);
-
-/** Get parent directory */
-CV_EXPORTS cv::String getParent(const cv::String &path);
-CV_EXPORTS std::wstring getParent(const std::wstring& path);
-
-/**
- * Generate a list of all files that match the globbing pattern.
- *
- * Result entries are prefixed by base directory path.
- *
- * @param directory base directory
- * @param pattern filter pattern (based on '*'/'?' symbols). Use empty string to disable filtering and return all results
- * @param[out] result result of globing.
- * @param recursive scan nested directories too
- * @param includeDirectories include directories into results list
- */
-CV_EXPORTS void glob(const cv::String& directory, const cv::String& pattern,
-        CV_OUT std::vector<cv::String>& result,
-        bool recursive = false, bool includeDirectories = false);
-
-/**
- * Generate a list of all files that match the globbing pattern.
- *
- * @param directory base directory
- * @param pattern filter pattern (based on '*'/'?' symbols). Use empty string to disable filtering and return all results
- * @param[out] result globbing result with relative paths from base directory
- * @param recursive scan nested directories too
- * @param includeDirectories include directories into results list
- */
-CV_EXPORTS void glob_relative(const cv::String& directory, const cv::String& pattern,
-        CV_OUT std::vector<cv::String>& result,
-        bool recursive = false, bool includeDirectories = false);
-
-
-CV_EXPORTS bool createDirectory(const cv::String& path);
-CV_EXPORTS bool createDirectories(const cv::String& path);
-
-#ifdef __OPENCV_BUILD
-// TODO
-//CV_EXPORTS cv::String getTempDirectory();
-
-/**
- * @brief Returns directory to store OpenCV cache files
- * Create sub-directory in common OpenCV cache directory if it doesn't exist.
- * @param sub_directory_name name of sub-directory. NULL or "" value asks to return root cache directory.
- * @param configuration_name optional name of configuration parameter name which overrides default behavior.
- * @return Path to cache directory. Returns empty string if cache directories support is not available. Returns "disabled" if cache disabled by user.
- */
-CV_EXPORTS cv::String getCacheDirectory(const char* sub_directory_name, const char* configuration_name = NULL);
-
-#endif
-
-}}} // namespace
-
-#endif // OPENCV_UTILS_FILESYSTEM_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/utils/instrumentation.hpp b/projects/llm_framework/include/opencv4/opencv2/core/utils/instrumentation.hpp
deleted file mode 100644
index 36398670..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/utils/instrumentation.hpp
+++ /dev/null
@@ -1,125 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_UTILS_INSTR_HPP
-#define OPENCV_UTILS_INSTR_HPP
-
-#include <opencv2/core/utility.hpp>
-#include <opencv2/core/utils/tls.hpp>
-
-namespace cv {
-
-//! @addtogroup core_utils
-//! @{
-
-#ifdef CV_COLLECT_IMPL_DATA
-CV_EXPORTS void setImpl(int flags); // set implementation flags and reset storage arrays
-CV_EXPORTS void addImpl(int flag, const char* func = 0); // add implementation and function name to storage arrays
-// Get stored implementation flags and functions names arrays
-// Each implementation entry correspond to function name entry, so you can find which implementation was executed in which function
-CV_EXPORTS int getImpl(std::vector<int> &impl, std::vector<String> &funName);
-
-CV_EXPORTS bool useCollection(); // return implementation collection state
-CV_EXPORTS void setUseCollection(bool flag); // set implementation collection state
-
-#define CV_IMPL_PLAIN  0x01 // native CPU OpenCV implementation
-#define CV_IMPL_OCL    0x02 // OpenCL implementation
-#define CV_IMPL_IPP    0x04 // IPP implementation
-#define CV_IMPL_MT     0x10 // multithreaded implementation
-
-#undef CV_IMPL_ADD
-#define CV_IMPL_ADD(impl)                                                   \
-    if(cv::useCollection())                                                 \
-    {                                                                       \
-        cv::addImpl(impl, CV_Func);                                         \
-    }
-#endif
-
-// Instrumentation external interface
-namespace instr
-{
-
-#if !defined OPENCV_ABI_CHECK
-
-enum TYPE
-{
-    TYPE_GENERAL = 0,   // OpenCV API function, e.g. exported function
-    TYPE_MARKER,        // Information marker
-    TYPE_WRAPPER,       // Wrapper function for implementation
-    TYPE_FUN,           // Simple function call
-};
-
-enum IMPL
-{
-    IMPL_PLAIN = 0,
-    IMPL_IPP,
-    IMPL_OPENCL,
-};
-
-struct NodeDataTls
-{
-    NodeDataTls()
-    {
-        m_ticksTotal = 0;
-    }
-    uint64      m_ticksTotal;
-};
-
-class CV_EXPORTS NodeData
-{
-public:
-    NodeData(const char* funName = 0, const char* fileName = NULL, int lineNum = 0, void* retAddress = NULL, bool alwaysExpand = false, cv::instr::TYPE instrType = TYPE_GENERAL, cv::instr::IMPL implType = IMPL_PLAIN);
-    NodeData(NodeData &ref);
-    ~NodeData();
-    NodeData& operator=(const NodeData&);
-
-    cv::String          m_funName;
-    cv::instr::TYPE     m_instrType;
-    cv::instr::IMPL     m_implType;
-    const char*         m_fileName;
-    int                 m_lineNum;
-    void*               m_retAddress;
-    bool                m_alwaysExpand;
-    bool                m_funError;
-
-    volatile int         m_counter;
-    volatile uint64      m_ticksTotal;
-    TLSDataAccumulator<NodeDataTls> m_tls;
-    int                  m_threads;
-
-    // No synchronization
-    double getTotalMs()   const { return ((double)m_ticksTotal / cv::getTickFrequency()) * 1000; }
-    double getMeanMs()    const { return (((double)m_ticksTotal/m_counter) / cv::getTickFrequency()) * 1000; }
-};
-bool operator==(const NodeData& lhs, const NodeData& rhs);
-
-typedef Node<NodeData> InstrNode;
-
-CV_EXPORTS InstrNode* getTrace();
-
-#endif // !defined OPENCV_ABI_CHECK
-
-
-CV_EXPORTS bool       useInstrumentation();
-CV_EXPORTS void       setUseInstrumentation(bool flag);
-CV_EXPORTS void       resetTrace();
-
-enum FLAGS
-{
-    FLAGS_NONE              = 0,
-    FLAGS_MAPPING           = 0x01,
-    FLAGS_EXPAND_SAME_NAMES = 0x02,
-};
-
-CV_EXPORTS void       setFlags(FLAGS modeFlags);
-static inline void    setFlags(int modeFlags) { setFlags((FLAGS)modeFlags); }
-CV_EXPORTS FLAGS      getFlags();
-
-} // namespace instr
-
-//! @}
-
-} // namespace
-
-#endif // OPENCV_UTILS_TLS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/utils/logger.defines.hpp b/projects/llm_framework/include/opencv4/opencv2/core/utils/logger.defines.hpp
deleted file mode 100644
index 7d73f02b..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/utils/logger.defines.hpp
+++ /dev/null
@@ -1,42 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_LOGGER_DEFINES_HPP
-#define OPENCV_LOGGER_DEFINES_HPP
-
-//! @addtogroup core_logging
-//! @{
-
-// Supported logging levels and their semantic
-#define CV_LOG_LEVEL_SILENT 0          //!< for using in setLogLevel() call
-#define CV_LOG_LEVEL_FATAL 1           //!< Fatal (critical) error (unrecoverable internal error)
-#define CV_LOG_LEVEL_ERROR 2           //!< Error message
-#define CV_LOG_LEVEL_WARN 3            //!< Warning message
-#define CV_LOG_LEVEL_INFO 4            //!< Info message
-#define CV_LOG_LEVEL_DEBUG 5           //!< Debug message. Disabled in the "Release" build.
-#define CV_LOG_LEVEL_VERBOSE 6         //!< Verbose (trace) messages. Requires verbosity level. Disabled in the "Release" build.
-
-namespace cv {
-namespace utils {
-namespace logging {
-
-//! Supported logging levels and their semantic
-enum LogLevel {
-    LOG_LEVEL_SILENT = 0,              //!< for using in setLogVevel() call
-    LOG_LEVEL_FATAL = 1,               //!< Fatal (critical) error (unrecoverable internal error)
-    LOG_LEVEL_ERROR = 2,               //!< Error message
-    LOG_LEVEL_WARNING = 3,             //!< Warning message
-    LOG_LEVEL_INFO = 4,                //!< Info message
-    LOG_LEVEL_DEBUG = 5,               //!< Debug message. Disabled in the "Release" build.
-    LOG_LEVEL_VERBOSE = 6,             //!< Verbose (trace) messages. Requires verbosity level. Disabled in the "Release" build.
-#ifndef CV_DOXYGEN
-    ENUM_LOG_LEVEL_FORCE_INT = INT_MAX
-#endif
-};
-
-}}} // namespace
-
-//! @}
-
-#endif // OPENCV_LOGGER_DEFINES_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/utils/logger.hpp b/projects/llm_framework/include/opencv4/opencv2/core/utils/logger.hpp
deleted file mode 100644
index accb860a..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/utils/logger.hpp
+++ /dev/null
@@ -1,218 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_LOGGER_HPP
-#define OPENCV_LOGGER_HPP
-
-#include <iostream>
-#include <sstream>
-#include <limits.h> // INT_MAX
-
-#include "logger.defines.hpp"
-#include "logtag.hpp"
-
-namespace cv {
-namespace utils {
-namespace logging {
-
-//! @addtogroup core_logging
-//! @{
-
-/** Set global logging level
-@return previous logging level
-*/
-CV_EXPORTS LogLevel setLogLevel(LogLevel logLevel);
-/** Get global logging level */
-CV_EXPORTS LogLevel getLogLevel();
-
-CV_EXPORTS void registerLogTag(cv::utils::logging::LogTag* plogtag);
-
-CV_EXPORTS void setLogTagLevel(const char* tag, cv::utils::logging::LogLevel level);
-
-CV_EXPORTS cv::utils::logging::LogLevel getLogTagLevel(const char* tag);
-
-namespace internal {
-
-/** Get global log tag */
-CV_EXPORTS cv::utils::logging::LogTag* getGlobalLogTag();
-
-/** Write log message */
-CV_EXPORTS void writeLogMessage(LogLevel logLevel, const char* message);
-
-/** Write log message */
-CV_EXPORTS void writeLogMessageEx(LogLevel logLevel, const char* tag, const char* file, int line, const char* func, const char* message);
-
-} // namespace
-
-struct LogTagAuto
-    : public LogTag
-{
-    inline LogTagAuto(const char* _name, LogLevel _level)
-        : LogTag(_name, _level)
-    {
-        registerLogTag(this);
-    }
-};
-
-/**
- * \def CV_LOG_STRIP_LEVEL
- *
- * Define CV_LOG_STRIP_LEVEL=CV_LOG_LEVEL_[DEBUG|INFO|WARN|ERROR|FATAL|SILENT] to compile out anything at that and before that logging level
- */
-#ifndef CV_LOG_STRIP_LEVEL
-# if defined NDEBUG
-#   define CV_LOG_STRIP_LEVEL CV_LOG_LEVEL_DEBUG
-# else
-#   define CV_LOG_STRIP_LEVEL CV_LOG_LEVEL_VERBOSE
-# endif
-#endif
-
-#define CV_LOGTAG_PTR_CAST(expr) static_cast<const cv::utils::logging::LogTag*>(expr)
-
-// CV_LOGTAG_EXPAND_NAME is intended to be re-defined (undef and then define again)
-// to allows logging users to use a shorter name argument when calling
-// CV_LOG_WITH_TAG or its related macros such as CV_LOG_INFO.
-//
-// This macro is intended to modify the tag argument as a string (token), via
-// preprocessor token pasting or metaprogramming techniques. A typical usage
-// is to apply a prefix, such as
-// ...... #define CV_LOGTAG_EXPAND_NAME(tag) cv_logtag_##tag
-//
-// It is permitted to re-define to a hard-coded expression, ignoring the tag.
-// This would work identically like the CV_LOGTAG_FALLBACK macro.
-//
-// Important: When the logging macro is called with tag being NULL, a user-defined
-// CV_LOGTAG_EXPAND_NAME may expand it into cv_logtag_0, cv_logtag_NULL, or
-// cv_logtag_nullptr. Use with care. Also be mindful of C++ symbol redefinitions.
-//
-// If there is significant amount of logging code with tag being NULL, it is
-// recommended to use (re-define) CV_LOGTAG_FALLBACK to inject locally a default
-// tag at the beginning of a compilation unit, to minimize lines of code changes.
-//
-#define CV_LOGTAG_EXPAND_NAME(tag) tag
-
-// CV_LOGTAG_FALLBACK is intended to be re-defined (undef and then define again)
-// by any other compilation units to provide a log tag when the logging statement
-// does not specify one. The macro needs to expand into a C++ expression that can
-// be static_cast into (cv::utils::logging::LogTag*). Null (nullptr) is permitted.
-#define CV_LOGTAG_FALLBACK nullptr
-
-// CV_LOGTAG_GLOBAL is the tag used when a log tag is not specified in the logging
-// statement nor the compilation unit. The macro needs to expand into a C++
-// expression that can be static_cast into (cv::utils::logging::LogTag*). Must be
-// non-null. Do not re-define.
-#define CV_LOGTAG_GLOBAL cv::utils::logging::internal::getGlobalLogTag()
-
-#define CV_LOG_WITH_TAG(tag, msgLevel, extra_check0, extra_check1, ...) \
-    for(;;) { \
-        extra_check0; \
-        const auto cv_temp_msglevel = (cv::utils::logging::LogLevel)(msgLevel); \
-        if (cv_temp_msglevel >= (CV_LOG_STRIP_LEVEL)) break; \
-        auto cv_temp_logtagptr = CV_LOGTAG_PTR_CAST(CV_LOGTAG_EXPAND_NAME(tag)); \
-        if (!cv_temp_logtagptr) cv_temp_logtagptr = CV_LOGTAG_PTR_CAST(CV_LOGTAG_FALLBACK); \
-        if (!cv_temp_logtagptr) cv_temp_logtagptr = CV_LOGTAG_PTR_CAST(CV_LOGTAG_GLOBAL); \
-        if (cv_temp_logtagptr && (cv_temp_msglevel > cv_temp_logtagptr->level)) break; \
-        extra_check1; \
-        std::stringstream cv_temp_logstream; \
-        cv_temp_logstream << __VA_ARGS__; \
-        cv::utils::logging::internal::writeLogMessageEx( \
-            cv_temp_msglevel, \
-            (cv_temp_logtagptr ? cv_temp_logtagptr->name : nullptr), \
-            __FILE__, \
-            __LINE__, \
-            CV_Func, \
-            cv_temp_logstream.str().c_str()); \
-        break; \
-    }
-
-#define CV_LOG_FATAL(tag, ...)   CV_LOG_WITH_TAG(tag, cv::utils::logging::LOG_LEVEL_FATAL, , , __VA_ARGS__)
-#define CV_LOG_ERROR(tag, ...)   CV_LOG_WITH_TAG(tag, cv::utils::logging::LOG_LEVEL_ERROR, , , __VA_ARGS__)
-#define CV_LOG_WARNING(tag, ...) CV_LOG_WITH_TAG(tag, cv::utils::logging::LOG_LEVEL_WARNING, , , __VA_ARGS__)
-#define CV_LOG_INFO(tag, ...) CV_LOG_WITH_TAG(tag, cv::utils::logging::LOG_LEVEL_INFO, , , __VA_ARGS__)
-#define CV_LOG_DEBUG(tag, ...) CV_LOG_WITH_TAG(tag, cv::utils::logging::LOG_LEVEL_DEBUG, , , __VA_ARGS__)
-#define CV_LOG_VERBOSE(tag, v, ...) CV_LOG_WITH_TAG(tag, (cv::utils::logging::LOG_LEVEL_VERBOSE + (int)(v)), , , __VA_ARGS__)
-
-#if CV_LOG_STRIP_LEVEL <= CV_LOG_LEVEL_INFO
-#undef CV_LOG_INFO
-#define CV_LOG_INFO(tag, ...)
-#endif
-
-#if CV_LOG_STRIP_LEVEL <= CV_LOG_LEVEL_DEBUG
-#undef CV_LOG_DEBUG
-#define CV_LOG_DEBUG(tag, ...)
-#endif
-
-#if CV_LOG_STRIP_LEVEL <= CV_LOG_LEVEL_VERBOSE
-#undef CV_LOG_VERBOSE
-#define CV_LOG_VERBOSE(tag, v, ...)
-#endif
-
-//! @cond IGNORED
-#define CV__LOG_ONCE_CHECK_PRE \
-    static bool _cv_log_once_ ## __LINE__ = false; \
-    if (_cv_log_once_ ## __LINE__) break;
-
-#define CV__LOG_ONCE_CHECK_POST \
-    _cv_log_once_ ## __LINE__ = true;
-
-#define CV__LOG_IF_CHECK(logging_cond) \
-    if (!(logging_cond)) break;
-
-//! @endcond
-
-
-// CV_LOG_ONCE_XXX macros
-
-#define CV_LOG_ONCE_ERROR(tag, ...) CV_LOG_WITH_TAG(tag, cv::utils::logging::LOG_LEVEL_ERROR, CV__LOG_ONCE_CHECK_PRE, CV__LOG_ONCE_CHECK_POST, __VA_ARGS__)
-#define CV_LOG_ONCE_WARNING(tag, ...) CV_LOG_WITH_TAG(tag, cv::utils::logging::LOG_LEVEL_WARNING, CV__LOG_ONCE_CHECK_PRE, CV__LOG_ONCE_CHECK_POST, __VA_ARGS__)
-#define CV_LOG_ONCE_INFO(tag, ...) CV_LOG_WITH_TAG(tag, cv::utils::logging::LOG_LEVEL_INFO, CV__LOG_ONCE_CHECK_PRE, CV__LOG_ONCE_CHECK_POST, __VA_ARGS__)
-#define CV_LOG_ONCE_DEBUG(tag, ...) CV_LOG_WITH_TAG(tag, cv::utils::logging::LOG_LEVEL_DEBUG, CV__LOG_ONCE_CHECK_PRE, CV__LOG_ONCE_CHECK_POST, __VA_ARGS__)
-#define CV_LOG_ONCE_VERBOSE(tag, v, ...) CV_LOG_WITH_TAG(tag, (cv::utils::logging::LOG_LEVEL_VERBOSE + (int)(v)), CV__LOG_ONCE_CHECK_PRE, CV__LOG_ONCE_CHECK_POST, __VA_ARGS__)
-
-#if CV_LOG_STRIP_LEVEL <= CV_LOG_LEVEL_INFO
-#undef CV_LOG_ONCE_INFO
-#define CV_LOG_ONCE_INFO(tag, ...)
-#endif
-
-#if CV_LOG_STRIP_LEVEL <= CV_LOG_LEVEL_DEBUG
-#undef CV_LOG_ONCE_DEBUG
-#define CV_LOG_ONCE_DEBUG(tag, ...)
-#endif
-
-#if CV_LOG_STRIP_LEVEL <= CV_LOG_LEVEL_VERBOSE
-#undef CV_LOG_ONCE_VERBOSE
-#define CV_LOG_ONCE_VERBOSE(tag, v, ...)
-#endif
-
-
-// CV_LOG_IF_XXX macros
-
-#define CV_LOG_IF_FATAL(tag, logging_cond, ...) CV_LOG_WITH_TAG(tag, cv::utils::logging::LOG_LEVEL_FATAL, , CV__LOG_IF_CHECK(logging_cond), __VA_ARGS__)
-#define CV_LOG_IF_ERROR(tag, logging_cond, ...) CV_LOG_WITH_TAG(tag, cv::utils::logging::LOG_LEVEL_ERROR, , CV__LOG_IF_CHECK(logging_cond), __VA_ARGS__)
-#define CV_LOG_IF_WARNING(tag, logging_cond, ...) CV_LOG_WITH_TAG(tag, cv::utils::logging::LOG_LEVEL_WARNING, , CV__LOG_IF_CHECK(logging_cond), __VA_ARGS__)
-#define CV_LOG_IF_INFO(tag, logging_cond, ...) CV_LOG_WITH_TAG(tag, cv::utils::logging::LOG_LEVEL_INFO, , CV__LOG_IF_CHECK(logging_cond), __VA_ARGS__)
-#define CV_LOG_IF_DEBUG(tag, logging_cond, ...) CV_LOG_WITH_TAG(tag, cv::utils::logging::LOG_LEVEL_DEBUG, , CV__LOG_IF_CHECK(logging_cond), __VA_ARGS__)
-#define CV_LOG_IF_VERBOSE(tag, v, logging_cond, ...) CV_LOG_WITH_TAG(tag, (cv::utils::logging::LOG_LEVEL_VERBOSE + (int)(v)), , CV__LOG_IF_CHECK(logging_cond), __VA_ARGS__)
-
-#if CV_LOG_STRIP_LEVEL <= CV_LOG_LEVEL_INFO
-#undef CV_LOG_IF_INFO
-#define CV_LOG_IF_INFO(tag, logging_cond, ...)
-#endif
-
-#if CV_LOG_STRIP_LEVEL <= CV_LOG_LEVEL_DEBUG
-#undef CV_LOG_IF_DEBUG
-#define CV_LOG_IF_DEBUG(tag, logging_cond, ...)
-#endif
-
-#if CV_LOG_STRIP_LEVEL <= CV_LOG_LEVEL_VERBOSE
-#undef CV_LOG_IF_VERBOSE
-#define CV_LOG_IF_VERBOSE(tag, v, logging_cond, ...)
-#endif
-
-
-//! @}
-
-}}} // namespace
-
-#endif // OPENCV_LOGGER_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/utils/logtag.hpp b/projects/llm_framework/include/opencv4/opencv2/core/utils/logtag.hpp
deleted file mode 100644
index 40897207..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/utils/logtag.hpp
+++ /dev/null
@@ -1,28 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_CORE_LOGTAG_HPP
-#define OPENCV_CORE_LOGTAG_HPP
-
-#include "opencv2/core/cvstd.hpp"
-#include "logger.defines.hpp"
-
-namespace cv {
-namespace utils {
-namespace logging {
-
-struct LogTag
-{
-    const char* name;
-    LogLevel level;
-
-    inline LogTag(const char* _name, LogLevel _level)
-        : name(_name)
-        , level(_level)
-    {}
-};
-
-}}}
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/utils/tls.hpp b/projects/llm_framework/include/opencv4/opencv2/core/utils/tls.hpp
deleted file mode 100644
index 124caebc..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/utils/tls.hpp
+++ /dev/null
@@ -1,235 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_UTILS_TLS_HPP
-#define OPENCV_UTILS_TLS_HPP
-
-#ifndef OPENCV_CORE_UTILITY_H
-#error "tls.hpp must be included after opencv2/core/utility.hpp or opencv2/core.hpp"
-#endif
-
-namespace cv {
-
-//! @addtogroup core_utils
-//! @{
-
-namespace details { class TlsStorage; }
-
-/** TLS container base implementation
- *
- * Don't use directly.
- *
- * @sa TLSData, TLSDataAccumulator templates
- */
-class CV_EXPORTS TLSDataContainer
-{
-protected:
-    TLSDataContainer();
-    virtual ~TLSDataContainer();
-
-    /// @deprecated use detachData() instead
-    void  gatherData(std::vector<void*> &data) const;
-    /// get TLS data and detach all data from threads (similar to cleanup() call)
-    void  detachData(std::vector<void*>& data);
-
-    void* getData() const;
-    void  release();
-
-protected:
-    virtual void* createDataInstance() const = 0;
-    virtual void  deleteDataInstance(void* pData) const = 0;
-
-private:
-    int key_;
-
-    friend class cv::details::TlsStorage;  // core/src/system.cpp
-
-public:
-    void cleanup(); //!< Release created TLS data container objects. It is similar to release() call, but it keeps TLS container valid.
-
-private:
-    // Disable copy/assign (noncopyable pattern)
-    TLSDataContainer(TLSDataContainer &) = delete;
-    TLSDataContainer& operator =(const TLSDataContainer &) = delete;
-};
-
-
-/** @brief Simple TLS data class
- *
- * @sa TLSDataAccumulator
- */
-template <typename T>
-class TLSData : protected TLSDataContainer
-{
-public:
-    inline TLSData() {}
-    inline ~TLSData() { release(); }
-
-    inline T* get() const   { return (T*)getData(); }  //!< Get data associated with key
-    inline T& getRef() const { T* ptr = (T*)getData(); CV_DbgAssert(ptr); return *ptr; }  //!< Get data associated with key
-
-    /// Release associated thread data
-    inline void cleanup()
-    {
-        TLSDataContainer::cleanup();
-    }
-
-protected:
-    /// Wrapper to allocate data by template
-    virtual void* createDataInstance() const CV_OVERRIDE { return new T; }
-    /// Wrapper to release data by template
-    virtual void  deleteDataInstance(void* pData) const CV_OVERRIDE { delete (T*)pData; }
-};
-
-
-/// TLS data accumulator with gathering methods
-template <typename T>
-class TLSDataAccumulator : public TLSData<T>
-{
-    mutable cv::Mutex mutex;
-    mutable std::vector<T*> dataFromTerminatedThreads;
-    std::vector<T*> detachedData;
-    bool cleanupMode;
-public:
-    TLSDataAccumulator() : cleanupMode(false) {}
-    ~TLSDataAccumulator()
-    {
-        release();
-    }
-
-    /** @brief Get data from all threads
-     * @deprecated replaced by detachData()
-     *
-     * Lifetime of vector data is valid until next detachData()/cleanup()/release() calls
-     *
-     * @param[out] data result buffer (should be empty)
-     */
-    void gather(std::vector<T*> &data) const
-    {
-        CV_Assert(cleanupMode == false);  // state is not valid
-        CV_Assert(data.empty());
-        {
-            std::vector<void*> &dataVoid = reinterpret_cast<std::vector<void*>&>(data);
-            TLSDataContainer::gatherData(dataVoid);
-        }
-        {
-            AutoLock lock(mutex);
-            data.reserve(data.size() + dataFromTerminatedThreads.size());
-            for (typename std::vector<T*>::const_iterator i = dataFromTerminatedThreads.begin(); i != dataFromTerminatedThreads.end(); ++i)
-            {
-                data.push_back((T*)*i);
-            }
-        }
-    }
-
-    /** @brief Get and detach data from all threads
-     *
-     * Call cleanupDetachedData() when returned vector is not needed anymore.
-     *
-     * @return Vector with associated data. Content is preserved (including lifetime of attached data pointers) until next detachData()/cleanupDetachedData()/cleanup()/release() calls
-     */
-    std::vector<T*>& detachData()
-    {
-        CV_Assert(cleanupMode == false);  // state is not valid
-        std::vector<void*> dataVoid;
-        {
-            TLSDataContainer::detachData(dataVoid);
-        }
-        {
-            AutoLock lock(mutex);
-            detachedData.reserve(dataVoid.size() + dataFromTerminatedThreads.size());
-            for (typename std::vector<T*>::const_iterator i = dataFromTerminatedThreads.begin(); i != dataFromTerminatedThreads.end(); ++i)
-            {
-                detachedData.push_back((T*)*i);
-            }
-            dataFromTerminatedThreads.clear();
-            for (typename std::vector<void*>::const_iterator i = dataVoid.begin(); i != dataVoid.end(); ++i)
-            {
-                detachedData.push_back((T*)(void*)*i);
-            }
-        }
-        dataVoid.clear();
-        return detachedData;
-    }
-
-    /// Release associated thread data returned by detachData() call
-    void cleanupDetachedData()
-    {
-        AutoLock lock(mutex);
-        cleanupMode = true;
-        _cleanupDetachedData();
-        cleanupMode = false;
-    }
-
-    /// Release associated thread data
-    void cleanup()
-    {
-        cleanupMode = true;
-        TLSDataContainer::cleanup();
-
-        AutoLock lock(mutex);
-        _cleanupDetachedData();
-        _cleanupTerminatedData();
-        cleanupMode = false;
-    }
-
-    /// Release associated thread data and free TLS key
-    void release()
-    {
-        cleanupMode = true;
-        TLSDataContainer::release();
-        {
-            AutoLock lock(mutex);
-            _cleanupDetachedData();
-            _cleanupTerminatedData();
-        }
-    }
-
-protected:
-    // synchronized
-    void _cleanupDetachedData()
-    {
-        for (typename std::vector<T*>::iterator i = detachedData.begin(); i != detachedData.end(); ++i)
-        {
-            deleteDataInstance((T*)*i);
-        }
-        detachedData.clear();
-    }
-
-    // synchronized
-    void _cleanupTerminatedData()
-    {
-        for (typename std::vector<T*>::iterator i = dataFromTerminatedThreads.begin(); i != dataFromTerminatedThreads.end(); ++i)
-        {
-            deleteDataInstance((T*)*i);
-        }
-        dataFromTerminatedThreads.clear();
-    }
-
-protected:
-    virtual void* createDataInstance() const CV_OVERRIDE
-    {
-        // Note: we can collect all allocated data here, but this would require raced mutex locks
-        return new T;
-    }
-    virtual void  deleteDataInstance(void* pData) const CV_OVERRIDE
-    {
-        if (cleanupMode)
-        {
-            delete (T*)pData;
-        }
-        else
-        {
-            AutoLock lock(mutex);
-            dataFromTerminatedThreads.push_back((T*)pData);
-        }
-    }
-};
-
-
-//! @}
-
-} // namespace
-
-#endif // OPENCV_UTILS_TLS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/utils/trace.hpp b/projects/llm_framework/include/opencv4/opencv2/core/utils/trace.hpp
deleted file mode 100644
index ef5d35b4..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/utils/trace.hpp
+++ /dev/null
@@ -1,252 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_TRACE_HPP
-#define OPENCV_TRACE_HPP
-
-#include <opencv2/core/cvdef.h>
-
-namespace cv {
-namespace utils {
-namespace trace {
-
-//! @addtogroup core_logging
-//! @{
-
-//! Macro to trace function
-#define CV_TRACE_FUNCTION()
-
-#define CV_TRACE_FUNCTION_SKIP_NESTED()
-
-//! Trace code scope.
-//! @note Dynamic names are not supported in this macro (on stack or heap). Use string literals here only, like "initialize".
-#define CV_TRACE_REGION(name_as_static_string_literal)
-//! mark completed of the current opened region and create new one
-//! @note Dynamic names are not supported in this macro (on stack or heap). Use string literals here only, like "step1".
-#define CV_TRACE_REGION_NEXT(name_as_static_string_literal)
-
-//! Macro to trace argument value
-#define CV_TRACE_ARG(arg_id)
-
-//! Macro to trace argument value (expanded version)
-#define CV_TRACE_ARG_VALUE(arg_id, arg_name, value)
-
-//! @cond IGNORED
-#define CV_TRACE_NS cv::utils::trace
-
-#if !defined(OPENCV_DISABLE_TRACE) && defined(__EMSCRIPTEN__)
-#define OPENCV_DISABLE_TRACE 1
-#endif
-
-namespace details {
-
-#ifndef __OPENCV_TRACE
-# if defined __OPENCV_BUILD && !defined __OPENCV_TESTS && !defined __OPENCV_APPS
-#   define __OPENCV_TRACE 1
-# else
-#   define __OPENCV_TRACE 0
-# endif
-#endif
-
-#ifndef CV_TRACE_FILENAME
-# define CV_TRACE_FILENAME __FILE__
-#endif
-
-#ifndef CV__TRACE_FUNCTION
-# if defined _MSC_VER
-#   define CV__TRACE_FUNCTION __FUNCSIG__
-# elif defined __GNUC__
-#   define CV__TRACE_FUNCTION __PRETTY_FUNCTION__
-# else
-#   define CV__TRACE_FUNCTION "<unknown>"
-# endif
-#endif
-
-//! Thread-local instance (usually allocated on stack)
-class CV_EXPORTS Region
-{
-public:
-    struct LocationExtraData;
-    struct LocationStaticStorage
-    {
-        LocationExtraData** ppExtra;   //< implementation specific data
-        const char* name;              //< region name (function name or other custom name)
-        const char* filename;          //< source code filename
-        int line;                      //< source code line
-        int flags;                     //< flags (implementation code path: Plain, IPP, OpenCL)
-    };
-
-    Region(const LocationStaticStorage& location);
-    inline ~Region()
-    {
-        if (implFlags != 0)
-            destroy();
-        CV_DbgAssert(implFlags == 0);
-        CV_DbgAssert(pImpl == NULL);
-    }
-
-    class Impl;
-    Impl* pImpl; // NULL if current region is not active
-    int implFlags; // see RegionFlag, 0 if region is ignored
-
-    bool isActive() const { return pImpl != NULL; }
-
-    void destroy();
-private:
-    Region(const Region&); // disabled
-    Region& operator= (const Region&); // disabled
-};
-
-//! Specify region flags
-enum RegionLocationFlag {
-    REGION_FLAG_FUNCTION = (1 << 0),             //< region is function (=1) / nested named region (=0)
-    REGION_FLAG_APP_CODE = (1 << 1),             //< region is Application code (=1) / OpenCV library code (=0)
-    REGION_FLAG_SKIP_NESTED = (1 << 2),          //< avoid processing of nested regions
-
-    REGION_FLAG_IMPL_IPP = (1 << 16),            //< region is part of IPP code path
-    REGION_FLAG_IMPL_OPENCL = (2 << 16),         //< region is part of OpenCL code path
-    REGION_FLAG_IMPL_OPENVX = (3 << 16),         //< region is part of OpenVX code path
-
-    REGION_FLAG_IMPL_MASK = (15 << 16),
-
-    REGION_FLAG_REGION_FORCE = (1 << 30),
-    REGION_FLAG_REGION_NEXT = (1 << 31),         //< close previous region (see #CV_TRACE_REGION_NEXT macro)
-
-    ENUM_REGION_FLAG_FORCE_INT = INT_MAX
-};
-
-struct CV_EXPORTS TraceArg {
-public:
-    struct ExtraData;
-    ExtraData** ppExtra;
-    const char* name;
-    int flags;
-};
-/** @brief Add meta information to current region (function)
- * See CV_TRACE_ARG macro
- * @param arg argument information structure (global static cache)
- * @param value argument value (can by dynamic string literal in case of string, static allocation is not required)
- */
-CV_EXPORTS void traceArg(const TraceArg& arg, const char* value);
-//! @overload
-CV_EXPORTS void traceArg(const TraceArg& arg, int value);
-//! @overload
-CV_EXPORTS void traceArg(const TraceArg& arg, int64 value);
-//! @overload
-CV_EXPORTS void traceArg(const TraceArg& arg, double value);
-
-#define CV__TRACE_LOCATION_VARNAME(loc_id) CVAUX_CONCAT(CVAUX_CONCAT(__cv_trace_location_, loc_id), __LINE__)
-#define CV__TRACE_LOCATION_EXTRA_VARNAME(loc_id) CVAUX_CONCAT(CVAUX_CONCAT(__cv_trace_location_extra_, loc_id) , __LINE__)
-
-#define CV__TRACE_DEFINE_LOCATION_(loc_id, name, flags) \
-    static CV_TRACE_NS::details::Region::LocationExtraData* CV__TRACE_LOCATION_EXTRA_VARNAME(loc_id) = 0; \
-    static const CV_TRACE_NS::details::Region::LocationStaticStorage \
-        CV__TRACE_LOCATION_VARNAME(loc_id) = { &(CV__TRACE_LOCATION_EXTRA_VARNAME(loc_id)), name, CV_TRACE_FILENAME, __LINE__, flags};
-
-#define CV__TRACE_DEFINE_LOCATION_FN(name, flags) CV__TRACE_DEFINE_LOCATION_(fn, name, ((flags) | CV_TRACE_NS::details::REGION_FLAG_FUNCTION))
-
-
-#define CV__TRACE_OPENCV_FUNCTION() \
-    CV__TRACE_DEFINE_LOCATION_FN(CV__TRACE_FUNCTION, 0); \
-    const CV_TRACE_NS::details::Region __region_fn(CV__TRACE_LOCATION_VARNAME(fn));
-
-#define CV__TRACE_OPENCV_FUNCTION_NAME(name) \
-    CV__TRACE_DEFINE_LOCATION_FN(name, 0); \
-    const CV_TRACE_NS::details::Region __region_fn(CV__TRACE_LOCATION_VARNAME(fn));
-
-#define CV__TRACE_APP_FUNCTION() \
-    CV__TRACE_DEFINE_LOCATION_FN(CV__TRACE_FUNCTION, CV_TRACE_NS::details::REGION_FLAG_APP_CODE); \
-    const CV_TRACE_NS::details::Region __region_fn(CV__TRACE_LOCATION_VARNAME(fn));
-
-#define CV__TRACE_APP_FUNCTION_NAME(name) \
-    CV__TRACE_DEFINE_LOCATION_FN(name, CV_TRACE_NS::details::REGION_FLAG_APP_CODE); \
-    const CV_TRACE_NS::details::Region __region_fn(CV__TRACE_LOCATION_VARNAME(fn));
-
-
-#define CV__TRACE_OPENCV_FUNCTION_SKIP_NESTED() \
-    CV__TRACE_DEFINE_LOCATION_FN(CV__TRACE_FUNCTION, CV_TRACE_NS::details::REGION_FLAG_SKIP_NESTED); \
-    const CV_TRACE_NS::details::Region __region_fn(CV__TRACE_LOCATION_VARNAME(fn));
-
-#define CV__TRACE_OPENCV_FUNCTION_NAME_SKIP_NESTED(name) \
-    CV__TRACE_DEFINE_LOCATION_FN(name, CV_TRACE_NS::details::REGION_FLAG_SKIP_NESTED); \
-    const CV_TRACE_NS::details::Region __region_fn(CV__TRACE_LOCATION_VARNAME(fn));
-
-#define CV__TRACE_APP_FUNCTION_SKIP_NESTED() \
-    CV__TRACE_DEFINE_LOCATION_FN(CV__TRACE_FUNCTION, CV_TRACE_NS::details::REGION_FLAG_SKIP_NESTED | CV_TRACE_NS::details::REGION_FLAG_APP_CODE); \
-    const CV_TRACE_NS::details::Region __region_fn(CV__TRACE_LOCATION_VARNAME(fn));
-
-
-#define CV__TRACE_REGION_(name_as_static_string_literal, flags) \
-    CV__TRACE_DEFINE_LOCATION_(region, name_as_static_string_literal, flags); \
-    CV_TRACE_NS::details::Region CVAUX_CONCAT(__region_, __LINE__)(CV__TRACE_LOCATION_VARNAME(region));
-
-#define CV__TRACE_REGION(name_as_static_string_literal) CV__TRACE_REGION_(name_as_static_string_literal, 0)
-#define CV__TRACE_REGION_NEXT(name_as_static_string_literal) CV__TRACE_REGION_(name_as_static_string_literal, CV_TRACE_NS::details::REGION_FLAG_REGION_NEXT)
-
-#define CV__TRACE_ARG_VARNAME(arg_id) CVAUX_CONCAT(__cv_trace_arg_ ## arg_id, __LINE__)
-#define CV__TRACE_ARG_EXTRA_VARNAME(arg_id) CVAUX_CONCAT(__cv_trace_arg_extra_ ## arg_id, __LINE__)
-
-#define CV__TRACE_DEFINE_ARG_(arg_id, name, flags) \
-    static CV_TRACE_NS::details::TraceArg::ExtraData* CV__TRACE_ARG_EXTRA_VARNAME(arg_id) = 0; \
-    static const CV_TRACE_NS::details::TraceArg \
-        CV__TRACE_ARG_VARNAME(arg_id) = { &(CV__TRACE_ARG_EXTRA_VARNAME(arg_id)), name, flags };
-
-#define CV__TRACE_ARG_VALUE(arg_id, arg_name, value) \
-        CV__TRACE_DEFINE_ARG_(arg_id, arg_name, 0); \
-        CV_TRACE_NS::details::traceArg((CV__TRACE_ARG_VARNAME(arg_id)), value);
-
-#define CV__TRACE_ARG(arg_id) CV_TRACE_ARG_VALUE(arg_id, #arg_id, (arg_id))
-
-} // namespace
-
-#ifndef OPENCV_DISABLE_TRACE
-#undef CV_TRACE_FUNCTION
-#undef CV_TRACE_FUNCTION_SKIP_NESTED
-#if __OPENCV_TRACE
-#define CV_TRACE_FUNCTION CV__TRACE_OPENCV_FUNCTION
-#define CV_TRACE_FUNCTION_SKIP_NESTED CV__TRACE_OPENCV_FUNCTION_SKIP_NESTED
-#else
-#define CV_TRACE_FUNCTION CV__TRACE_APP_FUNCTION
-#define CV_TRACE_FUNCTION_SKIP_NESTED CV__TRACE_APP_FUNCTION_SKIP_NESTED
-#endif
-
-#undef CV_TRACE_REGION
-#define CV_TRACE_REGION CV__TRACE_REGION
-
-#undef CV_TRACE_REGION_NEXT
-#define CV_TRACE_REGION_NEXT CV__TRACE_REGION_NEXT
-
-#undef CV_TRACE_ARG_VALUE
-#define CV_TRACE_ARG_VALUE(arg_id, arg_name, value) \
-        if (__region_fn.isActive()) \
-        { \
-            CV__TRACE_ARG_VALUE(arg_id, arg_name, value); \
-        }
-
-#undef CV_TRACE_ARG
-#define CV_TRACE_ARG CV__TRACE_ARG
-
-#endif // OPENCV_DISABLE_TRACE
-
-#ifdef OPENCV_TRACE_VERBOSE
-#define CV_TRACE_FUNCTION_VERBOSE CV_TRACE_FUNCTION
-#define CV_TRACE_REGION_VERBOSE CV_TRACE_REGION
-#define CV_TRACE_REGION_NEXT_VERBOSE CV_TRACE_REGION_NEXT
-#define CV_TRACE_ARG_VALUE_VERBOSE CV_TRACE_ARG_VALUE
-#define CV_TRACE_ARG_VERBOSE CV_TRACE_ARG
-#else
-#define CV_TRACE_FUNCTION_VERBOSE(...)
-#define CV_TRACE_REGION_VERBOSE(...)
-#define CV_TRACE_REGION_NEXT_VERBOSE(...)
-#define CV_TRACE_ARG_VALUE_VERBOSE(...)
-#define CV_TRACE_ARG_VERBOSE(...)
-#endif
-
-//! @endcond
-
-//! @}
-
-}}} // namespace
-
-#endif // OPENCV_TRACE_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/va_intel.hpp b/projects/llm_framework/include/opencv4/opencv2/core/va_intel.hpp
deleted file mode 100644
index b37ce751..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/va_intel.hpp
+++ /dev/null
@@ -1,75 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-// Copyright (C) 2015, Itseez, Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-
-#ifndef OPENCV_CORE_VA_INTEL_HPP
-#define OPENCV_CORE_VA_INTEL_HPP
-
-#ifndef __cplusplus
-#  error va_intel.hpp header must be compiled as C++
-#endif
-
-#include "opencv2/core.hpp"
-#include "ocl.hpp"
-
-#if defined(HAVE_VA)
-# include "va/va.h"
-#else  // HAVE_VA
-# if !defined(_VA_H_)
-    typedef void* VADisplay;
-    typedef unsigned int VASurfaceID;
-# endif // !_VA_H_
-#endif // HAVE_VA
-
-namespace cv { namespace va_intel {
-
-/** @addtogroup core_va_intel
-This section describes Intel VA-API/OpenCL (CL-VA) interoperability.
-
-To enable basic VA interoperability build OpenCV with libva library integration enabled: `-DWITH_VA=ON` (corresponding dev package should be installed).
-
-To enable advanced CL-VA interoperability support on Intel HW, enable option: `-DWITH_VA_INTEL=ON` (OpenCL integration should be enabled which is the default setting). Special runtime environment should be set up in order to use this feature: correct combination of [libva](https://github.com/intel/libva), [OpenCL runtime](https://github.com/intel/compute-runtime) and [media driver](https://github.com/intel/media-driver) should be installed.
-
-Check usage example for details: samples/va_intel/va_intel_interop.cpp
-*/
-//! @{
-
-/////////////////// CL-VA Interoperability Functions ///////////////////
-
-namespace ocl {
-using namespace cv::ocl;
-
-// TODO static functions in the Context class
-/** @brief Creates OpenCL context from VA.
-@param display    - VADisplay for which CL interop should be established.
-@param tryInterop - try to set up for interoperability, if true; set up for use slow copy if false.
-@return Returns reference to OpenCL Context
- */
-CV_EXPORTS Context& initializeContextFromVA(VADisplay display, bool tryInterop = true);
-
-} // namespace cv::va_intel::ocl
-
-/** @brief Converts InputArray to VASurfaceID object.
-@param display - VADisplay object.
-@param src     - source InputArray.
-@param surface - destination VASurfaceID object.
-@param size    - size of image represented by VASurfaceID object.
- */
-CV_EXPORTS void convertToVASurface(VADisplay display, InputArray src, VASurfaceID surface, Size size);
-
-/** @brief Converts VASurfaceID object to OutputArray.
-@param display - VADisplay object.
-@param surface - source VASurfaceID object.
-@param size    - size of image represented by VASurfaceID object.
-@param dst     - destination OutputArray.
- */
-CV_EXPORTS void convertFromVASurface(VADisplay display, VASurfaceID surface, Size size, OutputArray dst);
-
-//! @}
-
-}} // namespace cv::va_intel
-
-#endif /* OPENCV_CORE_VA_INTEL_HPP */
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/version.hpp b/projects/llm_framework/include/opencv4/opencv2/core/version.hpp
deleted file mode 100644
index 2fbd4995..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/version.hpp
+++ /dev/null
@@ -1,26 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_VERSION_HPP
-#define OPENCV_VERSION_HPP
-
-#define CV_VERSION_MAJOR    4
-#define CV_VERSION_MINOR    5
-#define CV_VERSION_REVISION 5
-#define CV_VERSION_STATUS   ""
-
-#define CVAUX_STR_EXP(__A)  #__A
-#define CVAUX_STR(__A)      CVAUX_STR_EXP(__A)
-
-#define CVAUX_STRW_EXP(__A)  L ## #__A
-#define CVAUX_STRW(__A)      CVAUX_STRW_EXP(__A)
-
-#define CV_VERSION          CVAUX_STR(CV_VERSION_MAJOR) "." CVAUX_STR(CV_VERSION_MINOR) "." CVAUX_STR(CV_VERSION_REVISION) CV_VERSION_STATUS
-
-/* old  style version constants*/
-#define CV_MAJOR_VERSION    CV_VERSION_MAJOR
-#define CV_MINOR_VERSION    CV_VERSION_MINOR
-#define CV_SUBMINOR_VERSION CV_VERSION_REVISION
-
-#endif // OPENCV_VERSION_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/core/vsx_utils.hpp b/projects/llm_framework/include/opencv4/opencv2/core/vsx_utils.hpp
deleted file mode 100644
index 68863ffb..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/core/vsx_utils.hpp
+++ /dev/null
@@ -1,1042 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html
-
-#ifndef OPENCV_HAL_VSX_UTILS_HPP
-#define OPENCV_HAL_VSX_UTILS_HPP
-
-#include "opencv2/core/cvdef.h"
-
-#ifndef SKIP_INCLUDES
-#   include <assert.h>
-#endif
-
-//! @addtogroup core_utils_vsx
-//! @{
-#if CV_VSX
-
-#define __VSX_S16__(c, v) (c){v, v, v, v, v, v, v, v, v, v, v, v, v, v, v, v}
-#define __VSX_S8__(c, v)  (c){v, v, v, v, v, v, v, v}
-#define __VSX_S4__(c, v)  (c){v, v, v, v}
-#define __VSX_S2__(c, v)  (c){v, v}
-
-typedef __vector unsigned char vec_uchar16;
-#define vec_uchar16_set(...) (vec_uchar16){__VA_ARGS__}
-#define vec_uchar16_sp(c)    (__VSX_S16__(vec_uchar16, (unsigned char)c))
-#define vec_uchar16_c(v)     ((vec_uchar16)(v))
-#define vec_uchar16_z        vec_uchar16_sp(0)
-
-typedef __vector signed char vec_char16;
-#define vec_char16_set(...) (vec_char16){__VA_ARGS__}
-#define vec_char16_sp(c)    (__VSX_S16__(vec_char16, (signed char)c))
-#define vec_char16_c(v)     ((vec_char16)(v))
-#define vec_char16_z        vec_char16_sp(0)
-
-typedef __vector unsigned short vec_ushort8;
-#define vec_ushort8_set(...) (vec_ushort8){__VA_ARGS__}
-#define vec_ushort8_sp(c)    (__VSX_S8__(vec_ushort8, (unsigned short)c))
-#define vec_ushort8_c(v)     ((vec_ushort8)(v))
-#define vec_ushort8_z        vec_ushort8_sp(0)
-
-typedef __vector signed short vec_short8;
-#define vec_short8_set(...) (vec_short8){__VA_ARGS__}
-#define vec_short8_sp(c)    (__VSX_S8__(vec_short8, (signed short)c))
-#define vec_short8_c(v)     ((vec_short8)(v))
-#define vec_short8_z        vec_short8_sp(0)
-
-typedef __vector unsigned int vec_uint4;
-#define vec_uint4_set(...) (vec_uint4){__VA_ARGS__}
-#define vec_uint4_sp(c)    (__VSX_S4__(vec_uint4, (unsigned int)c))
-#define vec_uint4_c(v)     ((vec_uint4)(v))
-#define vec_uint4_z        vec_uint4_sp(0)
-
-typedef __vector signed int vec_int4;
-#define vec_int4_set(...)  (vec_int4){__VA_ARGS__}
-#define vec_int4_sp(c)     (__VSX_S4__(vec_int4, (signed int)c))
-#define vec_int4_c(v)      ((vec_int4)(v))
-#define vec_int4_z         vec_int4_sp(0)
-
-typedef __vector float vec_float4;
-#define vec_float4_set(...)  (vec_float4){__VA_ARGS__}
-#define vec_float4_sp(c)     (__VSX_S4__(vec_float4, c))
-#define vec_float4_c(v)      ((vec_float4)(v))
-#define vec_float4_z         vec_float4_sp(0)
-
-typedef __vector unsigned long long vec_udword2;
-#define vec_udword2_set(...) (vec_udword2){__VA_ARGS__}
-#define vec_udword2_sp(c)    (__VSX_S2__(vec_udword2, (unsigned long long)c))
-#define vec_udword2_c(v)     ((vec_udword2)(v))
-#define vec_udword2_z        vec_udword2_sp(0)
-
-typedef __vector signed long long vec_dword2;
-#define vec_dword2_set(...) (vec_dword2){__VA_ARGS__}
-#define vec_dword2_sp(c)    (__VSX_S2__(vec_dword2, (signed long long)c))
-#define vec_dword2_c(v)     ((vec_dword2)(v))
-#define vec_dword2_z        vec_dword2_sp(0)
-
-typedef  __vector double vec_double2;
-#define vec_double2_set(...) (vec_double2){__VA_ARGS__}
-#define vec_double2_c(v)     ((vec_double2)(v))
-#define vec_double2_sp(c)    (__VSX_S2__(vec_double2, c))
-#define vec_double2_z        vec_double2_sp(0)
-
-#define vec_bchar16           __vector __bool char
-#define vec_bchar16_set(...) (vec_bchar16){__VA_ARGS__}
-#define vec_bchar16_c(v)     ((vec_bchar16)(v))
-
-#define vec_bshort8           __vector __bool short
-#define vec_bshort8_set(...) (vec_bshort8){__VA_ARGS__}
-#define vec_bshort8_c(v)     ((vec_bshort8)(v))
-
-#define vec_bint4             __vector __bool int
-#define vec_bint4_set(...)   (vec_bint4){__VA_ARGS__}
-#define vec_bint4_c(v)       ((vec_bint4)(v))
-
-#define vec_bdword2            __vector __bool long long
-#define vec_bdword2_set(...)  (vec_bdword2){__VA_ARGS__}
-#define vec_bdword2_c(v)      ((vec_bdword2)(v))
-
-#define VSX_FINLINE(tp) extern inline tp __attribute__((always_inline))
-
-#define VSX_REDIRECT_1RG(rt, rg, fnm, fn2)   \
-VSX_FINLINE(rt) fnm(const rg& a) { return fn2(a); }
-
-#define VSX_REDIRECT_2RG(rt, rg, fnm, fn2)   \
-VSX_FINLINE(rt) fnm(const rg& a, const rg& b) { return fn2(a, b); }
-
-/*
- * GCC VSX compatibility
-**/
-#if defined(__GNUG__) && !defined(__clang__)
-
-// inline asm helper
-#define VSX_IMPL_1RG(rt, rg, opc, fnm) \
-VSX_FINLINE(rt) fnm(const rg& a)       \
-{ rt rs; __asm__ __volatile__(#opc" %x0,%x1" : "=wa" (rs) : "wa" (a)); return rs; }
-
-#define VSX_IMPL_1VRG(rt, rg, opc, fnm) \
-VSX_FINLINE(rt) fnm(const rg& a)        \
-{ rt rs; __asm__ __volatile__(#opc" %0,%1" : "=v" (rs) : "v" (a)); return rs; }
-
-#define VSX_IMPL_2VRG_F(rt, rg, fopc, fnm)     \
-VSX_FINLINE(rt) fnm(const rg& a, const rg& b)  \
-{ rt rs; __asm__ __volatile__(fopc : "=v" (rs) : "v" (a), "v" (b)); return rs; }
-
-#define VSX_IMPL_2VRG(rt, rg, opc, fnm) VSX_IMPL_2VRG_F(rt, rg, #opc" %0,%1,%2", fnm)
-
-#if __GNUG__ < 8
-
-    // Support for int4 -> dword2 expanding multiply was added in GCC 8.
-    #ifdef vec_mule
-        #undef vec_mule
-    #endif
-    #ifdef vec_mulo
-        #undef vec_mulo
-    #endif
-
-    VSX_REDIRECT_2RG(vec_ushort8,  vec_uchar16,  vec_mule, __builtin_vec_mule)
-    VSX_REDIRECT_2RG(vec_short8,  vec_char16,  vec_mule, __builtin_vec_mule)
-    VSX_REDIRECT_2RG(vec_int4,  vec_short8,  vec_mule, __builtin_vec_mule)
-    VSX_REDIRECT_2RG(vec_uint4,  vec_ushort8,  vec_mule, __builtin_vec_mule)
-    VSX_REDIRECT_2RG(vec_ushort8,  vec_uchar16,  vec_mulo, __builtin_vec_mulo)
-    VSX_REDIRECT_2RG(vec_short8,  vec_char16,  vec_mulo, __builtin_vec_mulo)
-    VSX_REDIRECT_2RG(vec_int4,  vec_short8,  vec_mulo, __builtin_vec_mulo)
-    VSX_REDIRECT_2RG(vec_uint4,  vec_ushort8,  vec_mulo, __builtin_vec_mulo)
-
-    // dword2 support arrived in ISA 2.07 and GCC 8+
-    VSX_IMPL_2VRG(vec_dword2,  vec_int4,  vmulosw, vec_mule)
-    VSX_IMPL_2VRG(vec_udword2, vec_uint4, vmulouw, vec_mule)
-    VSX_IMPL_2VRG(vec_dword2,  vec_int4,  vmulesw, vec_mulo)
-    VSX_IMPL_2VRG(vec_udword2, vec_uint4, vmuleuw, vec_mulo)
-
-#endif
-
-#if __GNUG__ < 7
-// up to GCC 6 vec_mul only supports precisions and llong
-#   ifdef vec_mul
-#       undef vec_mul
-#   endif
-/*
- * there's no a direct instruction for supporting 8-bit, 16-bit multiplication in ISA 2.07,
- * XLC Implement it by using instruction "multiply even", "multiply odd" and "permute"
-**/
-#   define VSX_IMPL_MULH(Tvec, cperm)                                        \
-    VSX_FINLINE(Tvec) vec_mul(const Tvec& a, const Tvec& b)                  \
-    {                                                                        \
-        static const vec_uchar16 ev_od = {cperm};                            \
-        return vec_perm((Tvec)vec_mule(a, b), (Tvec)vec_mulo(a, b), ev_od);  \
-    }
-    #define VSX_IMPL_MULH_P16 0, 16, 2, 18, 4, 20, 6, 22, 8, 24, 10, 26, 12, 28, 14, 30
-    VSX_IMPL_MULH(vec_char16,  VSX_IMPL_MULH_P16)
-    VSX_IMPL_MULH(vec_uchar16, VSX_IMPL_MULH_P16)
-    #define VSX_IMPL_MULH_P8 0, 1, 16, 17, 4, 5, 20, 21, 8, 9, 24, 25, 12, 13, 28, 29
-    VSX_IMPL_MULH(vec_short8,  VSX_IMPL_MULH_P8)
-    VSX_IMPL_MULH(vec_ushort8, VSX_IMPL_MULH_P8)
-    // vmuluwm can be used for unsigned or signed integers, that's what they said
-    VSX_IMPL_2VRG(vec_int4,  vec_int4,  vmuluwm, vec_mul)
-    VSX_IMPL_2VRG(vec_uint4, vec_uint4, vmuluwm, vec_mul)
-    // redirect to GCC builtin vec_mul, since it already supports precisions and llong
-    VSX_REDIRECT_2RG(vec_float4,  vec_float4,  vec_mul, __builtin_vec_mul)
-    VSX_REDIRECT_2RG(vec_double2, vec_double2, vec_mul, __builtin_vec_mul)
-    VSX_REDIRECT_2RG(vec_dword2,  vec_dword2,  vec_mul, __builtin_vec_mul)
-    VSX_REDIRECT_2RG(vec_udword2, vec_udword2, vec_mul, __builtin_vec_mul)
-#endif // __GNUG__ < 7
-
-#if __GNUG__ < 6
-/*
- * Instruction "compare greater than or equal" in ISA 2.07 only supports single
- * and double precision.
- * In XLC and new versions of GCC implement integers by using instruction "greater than" and NOR.
-**/
-#   ifdef vec_cmpge
-#       undef vec_cmpge
-#   endif
-#   ifdef vec_cmple
-#       undef vec_cmple
-#   endif
-#   define vec_cmple(a, b) vec_cmpge(b, a)
-#   define VSX_IMPL_CMPGE(rt, rg, opc, fnm) \
-    VSX_IMPL_2VRG_F(rt, rg, #opc" %0,%2,%1\n\t xxlnor %x0,%x0,%x0", fnm)
-
-    VSX_IMPL_CMPGE(vec_bchar16, vec_char16,  vcmpgtsb, vec_cmpge)
-    VSX_IMPL_CMPGE(vec_bchar16, vec_uchar16, vcmpgtub, vec_cmpge)
-    VSX_IMPL_CMPGE(vec_bshort8, vec_short8,  vcmpgtsh, vec_cmpge)
-    VSX_IMPL_CMPGE(vec_bshort8, vec_ushort8, vcmpgtuh, vec_cmpge)
-    VSX_IMPL_CMPGE(vec_bint4,   vec_int4,    vcmpgtsw, vec_cmpge)
-    VSX_IMPL_CMPGE(vec_bint4,   vec_uint4,   vcmpgtuw, vec_cmpge)
-    VSX_IMPL_CMPGE(vec_bdword2, vec_dword2,  vcmpgtsd, vec_cmpge)
-    VSX_IMPL_CMPGE(vec_bdword2, vec_udword2, vcmpgtud, vec_cmpge)
-
-// redirect to GCC builtin cmpge, since it already supports precisions
-    VSX_REDIRECT_2RG(vec_bint4,   vec_float4,  vec_cmpge, __builtin_vec_cmpge)
-    VSX_REDIRECT_2RG(vec_bdword2, vec_double2, vec_cmpge, __builtin_vec_cmpge)
-
-// up to gcc5 vec_nor doesn't support bool long long
-#   undef vec_nor
-    template<typename T>
-    VSX_REDIRECT_2RG(T, T, vec_nor, __builtin_vec_nor)
-
-    VSX_FINLINE(vec_bdword2) vec_nor(const vec_bdword2& a, const vec_bdword2& b)
-    { return vec_bdword2_c(__builtin_vec_nor(vec_dword2_c(a), vec_dword2_c(b))); }
-
-// vec_packs doesn't support double words in gcc4 and old versions of gcc5
-#   undef vec_packs
-    VSX_REDIRECT_2RG(vec_char16,  vec_short8,  vec_packs, __builtin_vec_packs)
-    VSX_REDIRECT_2RG(vec_uchar16, vec_ushort8, vec_packs, __builtin_vec_packs)
-    VSX_REDIRECT_2RG(vec_short8,  vec_int4,    vec_packs, __builtin_vec_packs)
-    VSX_REDIRECT_2RG(vec_ushort8, vec_uint4,   vec_packs, __builtin_vec_packs)
-
-    VSX_IMPL_2VRG_F(vec_int4,  vec_dword2,  "vpksdss %0,%2,%1", vec_packs)
-    VSX_IMPL_2VRG_F(vec_uint4, vec_udword2, "vpkudus %0,%2,%1", vec_packs)
-#endif // __GNUG__ < 6
-
-#if __GNUG__ < 5
-// vec_xxpermdi in gcc4 missing little-endian supports just like clang
-#   define vec_permi(a, b, c) vec_xxpermdi(b, a, (3 ^ (((c) & 1) << 1 | (c) >> 1)))
-// same as vec_xxpermdi
-#   undef vec_vbpermq
-    VSX_IMPL_2VRG(vec_udword2, vec_uchar16, vbpermq, vec_vbpermq)
-    VSX_IMPL_2VRG(vec_dword2,  vec_char16, vbpermq, vec_vbpermq)
-#else
-#   define vec_permi vec_xxpermdi
-#endif // __GNUG__ < 5
-
-// shift left double by word immediate
-#ifndef vec_sldw
-#   define vec_sldw __builtin_vsx_xxsldwi
-#endif
-
-// vector population count
-VSX_IMPL_1VRG(vec_uchar16, vec_uchar16, vpopcntb, vec_popcntu)
-VSX_IMPL_1VRG(vec_uchar16, vec_char16,  vpopcntb, vec_popcntu)
-VSX_IMPL_1VRG(vec_ushort8, vec_ushort8, vpopcnth, vec_popcntu)
-VSX_IMPL_1VRG(vec_ushort8, vec_short8,  vpopcnth, vec_popcntu)
-VSX_IMPL_1VRG(vec_uint4,   vec_uint4,   vpopcntw, vec_popcntu)
-VSX_IMPL_1VRG(vec_uint4,   vec_int4,    vpopcntw, vec_popcntu)
-VSX_IMPL_1VRG(vec_udword2, vec_udword2, vpopcntd, vec_popcntu)
-VSX_IMPL_1VRG(vec_udword2, vec_dword2,  vpopcntd, vec_popcntu)
-
-// converts between single and double-precision
-VSX_REDIRECT_1RG(vec_float4,  vec_double2, vec_cvfo, __builtin_vsx_xvcvdpsp)
-VSX_REDIRECT_1RG(vec_double2, vec_float4,  vec_cvfo, __builtin_vsx_xvcvspdp)
-
-// converts word and doubleword to double-precision
-#undef vec_ctd
-VSX_IMPL_1RG(vec_double2, vec_int4,    xvcvsxwdp, vec_ctdo)
-VSX_IMPL_1RG(vec_double2, vec_uint4,   xvcvuxwdp, vec_ctdo)
-VSX_IMPL_1RG(vec_double2, vec_dword2,  xvcvsxddp, vec_ctd)
-VSX_IMPL_1RG(vec_double2, vec_udword2, xvcvuxddp, vec_ctd)
-
-// converts word and doubleword to single-precision
-#undef vec_ctf
-VSX_IMPL_1RG(vec_float4, vec_int4,    xvcvsxwsp, vec_ctf)
-VSX_IMPL_1RG(vec_float4, vec_uint4,   xvcvuxwsp, vec_ctf)
-VSX_IMPL_1RG(vec_float4, vec_dword2,  xvcvsxdsp, vec_ctfo)
-VSX_IMPL_1RG(vec_float4, vec_udword2, xvcvuxdsp, vec_ctfo)
-
-// converts single and double precision to signed word
-#undef vec_cts
-VSX_IMPL_1RG(vec_int4,  vec_double2, xvcvdpsxws, vec_ctso)
-VSX_IMPL_1RG(vec_int4,  vec_float4,  xvcvspsxws, vec_cts)
-
-// converts single and double precision to unsigned word
-#undef vec_ctu
-VSX_IMPL_1RG(vec_uint4, vec_double2, xvcvdpuxws, vec_ctuo)
-VSX_IMPL_1RG(vec_uint4, vec_float4,  xvcvspuxws, vec_ctu)
-
-// converts single and double precision to signed doubleword
-#undef vec_ctsl
-VSX_IMPL_1RG(vec_dword2, vec_double2, xvcvdpsxds, vec_ctsl)
-VSX_IMPL_1RG(vec_dword2, vec_float4,  xvcvspsxds, vec_ctslo)
-
-// converts single and double precision to unsigned doubleword
-#undef vec_ctul
-VSX_IMPL_1RG(vec_udword2, vec_double2, xvcvdpuxds, vec_ctul)
-VSX_IMPL_1RG(vec_udword2, vec_float4,  xvcvspuxds, vec_ctulo)
-
-// just in case if GCC doesn't define it
-#ifndef vec_xl
-#   define vec_xl vec_vsx_ld
-#   define vec_xst vec_vsx_st
-#endif
-
-#endif // GCC VSX compatibility
-
-/*
- * CLANG VSX compatibility
-**/
-#if defined(__clang__) && !defined(__IBMCPP__)
-
-/*
- * CLANG doesn't support %x<n> in the inline asm template which fixes register number
- * when using any of the register constraints wa, wd, wf
- *
- * For more explanation checkout PowerPC and IBM RS6000 in https://gcc.gnu.org/onlinedocs/gcc/Machine-Constraints.html
- * Also there's already an open bug https://bugs.llvm.org/show_bug.cgi?id=31837
- *
- * So we're not able to use inline asm and only use built-in functions that CLANG supports
- * and use __builtin_convertvector if clang missing any of vector conversions built-in functions
- *
- * todo: clang asm template bug is fixed, need to reconsider the current workarounds.
-*/
-
-// convert vector helper
-#define VSX_IMPL_CONVERT(rt, rg, fnm) \
-VSX_FINLINE(rt) fnm(const rg& a) { return __builtin_convertvector(a, rt); }
-
-#if __clang_major__ < 5
-// implement vec_permi in a dirty way
-#   define VSX_IMPL_CLANG_4_PERMI(Tvec)                                                 \
-    VSX_FINLINE(Tvec) vec_permi(const Tvec& a, const Tvec& b, unsigned const char c)    \
-    {                                                                                   \
-        switch (c)                                                                      \
-        {                                                                               \
-        case 0:                                                                         \
-            return vec_mergeh(a, b);                                                    \
-        case 1:                                                                         \
-            return vec_mergel(vec_mergeh(a, a), b);                                     \
-        case 2:                                                                         \
-            return vec_mergeh(vec_mergel(a, a), b);                                     \
-        default:                                                                        \
-            return vec_mergel(a, b);                                                    \
-        }                                                                               \
-    }
-    VSX_IMPL_CLANG_4_PERMI(vec_udword2)
-    VSX_IMPL_CLANG_4_PERMI(vec_dword2)
-    VSX_IMPL_CLANG_4_PERMI(vec_double2)
-
-// vec_xxsldwi is missing in clang 4
-#   define vec_xxsldwi(a, b, c) vec_sld(a, b, (c) * 4)
-#else
-// vec_xxpermdi is missing little-endian supports in clang 4 just like gcc4
-#   define vec_permi(a, b, c) vec_xxpermdi(b, a, (3 ^ (((c) & 1) << 1 | (c) >> 1)))
-#endif // __clang_major__ < 5
-
-// shift left double by word immediate
-#ifndef vec_sldw
-#   define vec_sldw vec_xxsldwi
-#endif
-
-// Implement vec_rsqrt since clang only supports vec_rsqrte
-#ifndef vec_rsqrt
-    VSX_FINLINE(vec_float4) vec_rsqrt(const vec_float4& a)
-    { return vec_div(vec_float4_sp(1), vec_sqrt(a)); }
-
-    VSX_FINLINE(vec_double2) vec_rsqrt(const vec_double2& a)
-    { return vec_div(vec_double2_sp(1), vec_sqrt(a)); }
-#endif
-
-// vec_promote missing support for doubleword
-VSX_FINLINE(vec_dword2) vec_promote(long long a, int b)
-{
-    vec_dword2 ret = vec_dword2_z;
-    ret[b & 1] = a;
-    return ret;
-}
-
-VSX_FINLINE(vec_udword2) vec_promote(unsigned long long a, int b)
-{
-    vec_udword2 ret = vec_udword2_z;
-    ret[b & 1] = a;
-    return ret;
-}
-
-// vec_popcnt should return unsigned but clang has different thought just like gcc in vec_vpopcnt
-#define VSX_IMPL_POPCNTU(Tvec, Tvec2, ucast)   \
-VSX_FINLINE(Tvec) vec_popcntu(const Tvec2& a)  \
-{ return ucast(vec_popcnt(a)); }
-VSX_IMPL_POPCNTU(vec_uchar16, vec_char16, vec_uchar16_c);
-VSX_IMPL_POPCNTU(vec_ushort8, vec_short8, vec_ushort8_c);
-VSX_IMPL_POPCNTU(vec_uint4,   vec_int4,   vec_uint4_c);
-VSX_IMPL_POPCNTU(vec_udword2, vec_dword2, vec_udword2_c);
-// redirect unsigned types
-VSX_REDIRECT_1RG(vec_uchar16, vec_uchar16, vec_popcntu, vec_popcnt)
-VSX_REDIRECT_1RG(vec_ushort8, vec_ushort8, vec_popcntu, vec_popcnt)
-VSX_REDIRECT_1RG(vec_uint4,   vec_uint4,   vec_popcntu, vec_popcnt)
-VSX_REDIRECT_1RG(vec_udword2, vec_udword2, vec_popcntu, vec_popcnt)
-
-// converts between single and double precision
-VSX_REDIRECT_1RG(vec_float4,  vec_double2, vec_cvfo, __builtin_vsx_xvcvdpsp)
-VSX_REDIRECT_1RG(vec_double2, vec_float4,  vec_cvfo, __builtin_vsx_xvcvspdp)
-
-// converts word and doubleword to double-precision
-#ifdef vec_ctd
-#   undef vec_ctd
-#endif
-VSX_REDIRECT_1RG(vec_double2, vec_int4,  vec_ctdo, __builtin_vsx_xvcvsxwdp)
-VSX_REDIRECT_1RG(vec_double2, vec_uint4, vec_ctdo, __builtin_vsx_xvcvuxwdp)
-
-VSX_IMPL_CONVERT(vec_double2, vec_dword2,  vec_ctd)
-VSX_IMPL_CONVERT(vec_double2, vec_udword2, vec_ctd)
-
-// converts word and doubleword to single-precision
-#if __clang_major__ > 4
-#   undef vec_ctf
-#endif
-VSX_IMPL_CONVERT(vec_float4, vec_int4,    vec_ctf)
-VSX_IMPL_CONVERT(vec_float4, vec_uint4,   vec_ctf)
-VSX_REDIRECT_1RG(vec_float4, vec_dword2,  vec_ctfo, __builtin_vsx_xvcvsxdsp)
-VSX_REDIRECT_1RG(vec_float4, vec_udword2, vec_ctfo, __builtin_vsx_xvcvuxdsp)
-
-// converts single and double precision to signed word
-#if __clang_major__ > 4
-#   undef vec_cts
-#endif
-VSX_REDIRECT_1RG(vec_int4,  vec_double2, vec_ctso, __builtin_vsx_xvcvdpsxws)
-VSX_IMPL_CONVERT(vec_int4,  vec_float4,  vec_cts)
-
-// converts single and double precision to unsigned word
-#if __clang_major__ > 4
-#   undef vec_ctu
-#endif
-VSX_REDIRECT_1RG(vec_uint4, vec_double2, vec_ctuo, __builtin_vsx_xvcvdpuxws)
-VSX_IMPL_CONVERT(vec_uint4, vec_float4,  vec_ctu)
-
-// converts single and double precision to signed doubleword
-#ifdef vec_ctsl
-#   undef vec_ctsl
-#endif
-VSX_IMPL_CONVERT(vec_dword2, vec_double2, vec_ctsl)
-// __builtin_convertvector unable to convert, xvcvspsxds is missing on it
-VSX_FINLINE(vec_dword2) vec_ctslo(const vec_float4& a)
-{ return vec_ctsl(vec_cvfo(a)); }
-
-// converts single and double precision to unsigned doubleword
-#ifdef vec_ctul
-#   undef vec_ctul
-#endif
-VSX_IMPL_CONVERT(vec_udword2, vec_double2, vec_ctul)
-// __builtin_convertvector unable to convert, xvcvspuxds is missing on it
-VSX_FINLINE(vec_udword2) vec_ctulo(const vec_float4& a)
-{ return vec_ctul(vec_cvfo(a)); }
-
-#endif // CLANG VSX compatibility
-
-/*
- * Common GCC, CLANG compatibility
-**/
-#if defined(__GNUG__) && !defined(__IBMCPP__)
-
-#ifdef vec_cvf
-#   undef vec_cvf
-#endif
-
-#define VSX_IMPL_CONV_EVEN_4_2(rt, rg, fnm, fn2) \
-VSX_FINLINE(rt) fnm(const rg& a)                 \
-{ return fn2(vec_sldw(a, a, 1)); }
-
-VSX_IMPL_CONV_EVEN_4_2(vec_double2, vec_float4, vec_cvf,  vec_cvfo)
-VSX_IMPL_CONV_EVEN_4_2(vec_double2, vec_int4,   vec_ctd,  vec_ctdo)
-VSX_IMPL_CONV_EVEN_4_2(vec_double2, vec_uint4,  vec_ctd,  vec_ctdo)
-
-VSX_IMPL_CONV_EVEN_4_2(vec_dword2,  vec_float4, vec_ctsl, vec_ctslo)
-VSX_IMPL_CONV_EVEN_4_2(vec_udword2, vec_float4, vec_ctul, vec_ctulo)
-
-#define VSX_IMPL_CONV_EVEN_2_4(rt, rg, fnm, fn2) \
-VSX_FINLINE(rt) fnm(const rg& a)                 \
-{                                                \
-    rt v4 = fn2(a);                              \
-    return vec_sldw(v4, v4, 3);                  \
-}
-
-VSX_IMPL_CONV_EVEN_2_4(vec_float4, vec_double2, vec_cvf, vec_cvfo)
-VSX_IMPL_CONV_EVEN_2_4(vec_float4, vec_dword2,  vec_ctf, vec_ctfo)
-VSX_IMPL_CONV_EVEN_2_4(vec_float4, vec_udword2, vec_ctf, vec_ctfo)
-
-VSX_IMPL_CONV_EVEN_2_4(vec_int4,   vec_double2, vec_cts, vec_ctso)
-VSX_IMPL_CONV_EVEN_2_4(vec_uint4,  vec_double2, vec_ctu, vec_ctuo)
-
-// Only for Eigen!
-/*
- * changing behavior of conversion intrinsics for gcc has effect on Eigen
- * so we redefine old behavior again only on gcc, clang
-*/
-#if !defined(__clang__) || __clang_major__ > 4
-    // ignoring second arg since Eigen only truncates toward zero
-#   define VSX_IMPL_CONV_2VARIANT(rt, rg, fnm, fn2)     \
-    VSX_FINLINE(rt) fnm(const rg& a, int only_truncate) \
-    {                                                   \
-        assert(only_truncate == 0);                     \
-        CV_UNUSED(only_truncate);                       \
-        return fn2(a);                                  \
-    }
-    VSX_IMPL_CONV_2VARIANT(vec_int4,   vec_float4,  vec_cts, vec_cts)
-    VSX_IMPL_CONV_2VARIANT(vec_uint4,  vec_float4,  vec_ctu, vec_ctu)
-    VSX_IMPL_CONV_2VARIANT(vec_float4, vec_int4,    vec_ctf, vec_ctf)
-    VSX_IMPL_CONV_2VARIANT(vec_float4, vec_uint4,   vec_ctf, vec_ctf)
-    // define vec_cts for converting double precision to signed doubleword
-    // which isn't compatible with xlc but its okay since Eigen only uses it for gcc
-    VSX_IMPL_CONV_2VARIANT(vec_dword2, vec_double2, vec_cts, vec_ctsl)
-#endif // Eigen
-
-#endif // Common GCC, CLANG compatibility
-
-/*
- * XLC VSX compatibility
-**/
-#if defined(__IBMCPP__)
-
-// vector population count
-#define vec_popcntu vec_popcnt
-
-// overload and redirect with setting second arg to zero
-// since we only support conversions without the second arg
-#define VSX_IMPL_OVERLOAD_Z2(rt, rg, fnm) \
-VSX_FINLINE(rt) fnm(const rg& a) { return fnm(a, 0); }
-
-VSX_IMPL_OVERLOAD_Z2(vec_double2, vec_int4,    vec_ctd)
-VSX_IMPL_OVERLOAD_Z2(vec_double2, vec_uint4,   vec_ctd)
-VSX_IMPL_OVERLOAD_Z2(vec_double2, vec_dword2,  vec_ctd)
-VSX_IMPL_OVERLOAD_Z2(vec_double2, vec_udword2, vec_ctd)
-
-VSX_IMPL_OVERLOAD_Z2(vec_float4,  vec_int4,    vec_ctf)
-VSX_IMPL_OVERLOAD_Z2(vec_float4,  vec_uint4,   vec_ctf)
-VSX_IMPL_OVERLOAD_Z2(vec_float4,  vec_dword2,  vec_ctf)
-VSX_IMPL_OVERLOAD_Z2(vec_float4,  vec_udword2, vec_ctf)
-
-VSX_IMPL_OVERLOAD_Z2(vec_int4,    vec_double2, vec_cts)
-VSX_IMPL_OVERLOAD_Z2(vec_int4,    vec_float4,  vec_cts)
-
-VSX_IMPL_OVERLOAD_Z2(vec_uint4,   vec_double2, vec_ctu)
-VSX_IMPL_OVERLOAD_Z2(vec_uint4,   vec_float4,  vec_ctu)
-
-VSX_IMPL_OVERLOAD_Z2(vec_dword2,  vec_double2, vec_ctsl)
-VSX_IMPL_OVERLOAD_Z2(vec_dword2,  vec_float4,  vec_ctsl)
-
-VSX_IMPL_OVERLOAD_Z2(vec_udword2, vec_double2, vec_ctul)
-VSX_IMPL_OVERLOAD_Z2(vec_udword2, vec_float4,  vec_ctul)
-
-// fixme: implement conversions of odd-numbered elements in a dirty way
-// since xlc doesn't support VSX registers operand in inline asm.
-#define VSX_IMPL_CONV_ODD_4_2(rt, rg, fnm, fn2) \
-VSX_FINLINE(rt) fnm(const rg& a) { return fn2(vec_sldw(a, a, 3)); }
-
-VSX_IMPL_CONV_ODD_4_2(vec_double2, vec_float4, vec_cvfo,  vec_cvf)
-VSX_IMPL_CONV_ODD_4_2(vec_double2, vec_int4,   vec_ctdo,  vec_ctd)
-VSX_IMPL_CONV_ODD_4_2(vec_double2, vec_uint4,  vec_ctdo,  vec_ctd)
-
-VSX_IMPL_CONV_ODD_4_2(vec_dword2,  vec_float4, vec_ctslo, vec_ctsl)
-VSX_IMPL_CONV_ODD_4_2(vec_udword2, vec_float4, vec_ctulo, vec_ctul)
-
-#define VSX_IMPL_CONV_ODD_2_4(rt, rg, fnm, fn2)  \
-VSX_FINLINE(rt) fnm(const rg& a)                 \
-{                                                \
-    rt v4 = fn2(a);                              \
-    return vec_sldw(v4, v4, 1);                  \
-}
-
-VSX_IMPL_CONV_ODD_2_4(vec_float4, vec_double2, vec_cvfo, vec_cvf)
-VSX_IMPL_CONV_ODD_2_4(vec_float4, vec_dword2,  vec_ctfo, vec_ctf)
-VSX_IMPL_CONV_ODD_2_4(vec_float4, vec_udword2, vec_ctfo, vec_ctf)
-
-VSX_IMPL_CONV_ODD_2_4(vec_int4,   vec_double2, vec_ctso, vec_cts)
-VSX_IMPL_CONV_ODD_2_4(vec_uint4,  vec_double2, vec_ctuo, vec_ctu)
-
-#endif // XLC VSX compatibility
-
-// ignore GCC warning that caused by -Wunused-but-set-variable in rare cases
-#if defined(__GNUG__) && !defined(__clang__)
-#   define VSX_UNUSED(Tvec) Tvec __attribute__((__unused__))
-#else // CLANG, XLC
-#   define VSX_UNUSED(Tvec) Tvec
-#endif
-
-// gcc can find his way in casting log int and XLC, CLANG ambiguous
-#if defined(__clang__) || defined(__IBMCPP__)
-    VSX_FINLINE(vec_udword2) vec_splats(uint64 v)
-    { return vec_splats((unsigned long long) v); }
-
-    VSX_FINLINE(vec_dword2) vec_splats(int64 v)
-    { return vec_splats((long long) v); }
-
-    VSX_FINLINE(vec_udword2) vec_promote(uint64 a, int b)
-    { return vec_promote((unsigned long long) a, b); }
-
-    VSX_FINLINE(vec_dword2) vec_promote(int64 a, int b)
-    { return vec_promote((long long) a, b); }
-#endif
-
-/*
- * implement vsx_ld(offset, pointer), vsx_st(vector, offset, pointer)
- * load and set using offset depend on the pointer type
- *
- * implement vsx_ldf(offset, pointer), vsx_stf(vector, offset, pointer)
- * load and set using offset depend on fixed bytes size
- *
- * Note: In clang vec_xl and vec_xst fails to load unaligned addresses
- * so we are using vec_vsx_ld, vec_vsx_st instead
-*/
-
-#if defined(__clang__) && !defined(__IBMCPP__)
-#   define vsx_ldf  vec_vsx_ld
-#   define vsx_stf  vec_vsx_st
-#else // GCC , XLC
-#   define vsx_ldf  vec_xl
-#   define vsx_stf  vec_xst
-#endif
-
-#define VSX_OFFSET(o, p) ((o) * sizeof(*(p)))
-#define vsx_ld(o, p) vsx_ldf(VSX_OFFSET(o, p), p)
-#define vsx_st(v, o, p) vsx_stf(v, VSX_OFFSET(o, p), p)
-
-/*
- * implement vsx_ld2(offset, pointer), vsx_st2(vector, offset, pointer) to load and store double words
- * In GCC vec_xl and vec_xst it maps to vec_vsx_ld, vec_vsx_st which doesn't support long long
- * and in CLANG we are using vec_vsx_ld, vec_vsx_st because vec_xl, vec_xst fails to load unaligned addresses
- *
- * In XLC vec_xl and vec_xst fail to cast int64(long int) to long long
-*/
-#if (defined(__GNUG__) || defined(__clang__)) && !defined(__IBMCPP__)
-    VSX_FINLINE(vec_udword2) vsx_ld2(long o, const uint64* p)
-    { return vec_udword2_c(vsx_ldf(VSX_OFFSET(o, p), (unsigned int*)p)); }
-
-    VSX_FINLINE(vec_dword2) vsx_ld2(long o, const int64* p)
-    { return vec_dword2_c(vsx_ldf(VSX_OFFSET(o, p), (int*)p)); }
-
-    VSX_FINLINE(void) vsx_st2(const vec_udword2& vec, long o, uint64* p)
-    { vsx_stf(vec_uint4_c(vec), VSX_OFFSET(o, p), (unsigned int*)p); }
-
-    VSX_FINLINE(void) vsx_st2(const vec_dword2& vec, long o, int64* p)
-    { vsx_stf(vec_int4_c(vec), VSX_OFFSET(o, p), (int*)p); }
-#else // XLC
-    VSX_FINLINE(vec_udword2) vsx_ld2(long o, const uint64* p)
-    { return vsx_ldf(VSX_OFFSET(o, p), (unsigned long long*)p); }
-
-    VSX_FINLINE(vec_dword2) vsx_ld2(long o, const int64* p)
-    { return vsx_ldf(VSX_OFFSET(o, p), (long long*)p); }
-
-    VSX_FINLINE(void) vsx_st2(const vec_udword2& vec, long o, uint64* p)
-    { vsx_stf(vec, VSX_OFFSET(o, p), (unsigned long long*)p); }
-
-    VSX_FINLINE(void) vsx_st2(const vec_dword2& vec, long o, int64* p)
-    { vsx_stf(vec, VSX_OFFSET(o, p), (long long*)p); }
-#endif
-
-// Store lower 8 byte
-#define vec_st_l8(v, p) *((uint64*)(p)) = vec_extract(vec_udword2_c(v), 0)
-
-// Store higher 8 byte
-#define vec_st_h8(v, p) *((uint64*)(p)) = vec_extract(vec_udword2_c(v), 1)
-
-// Load 64-bits of integer data to lower part
-#define VSX_IMPL_LOAD_L8(Tvec, Tp)                  \
-VSX_FINLINE(Tvec) vec_ld_l8(const Tp *p)            \
-{ return ((Tvec)vec_promote(*((uint64*)p), 0)); }
-
-VSX_IMPL_LOAD_L8(vec_uchar16, uchar)
-VSX_IMPL_LOAD_L8(vec_char16,  schar)
-VSX_IMPL_LOAD_L8(vec_ushort8, ushort)
-VSX_IMPL_LOAD_L8(vec_short8,  short)
-VSX_IMPL_LOAD_L8(vec_uint4,   uint)
-VSX_IMPL_LOAD_L8(vec_int4,    int)
-VSX_IMPL_LOAD_L8(vec_float4,  float)
-VSX_IMPL_LOAD_L8(vec_udword2, uint64)
-VSX_IMPL_LOAD_L8(vec_dword2,  int64)
-VSX_IMPL_LOAD_L8(vec_double2, double)
-
-// logical not
-#define vec_not(a) vec_nor(a, a)
-
-// power9 yaya
-// not equal
-#ifndef vec_cmpne
-#   define vec_cmpne(a, b) vec_not(vec_cmpeq(a, b))
-#endif
-
-// absolute difference
-#ifndef vec_absd
-#   define vec_absd(a, b) vec_sub(vec_max(a, b), vec_min(a, b))
-#endif
-
-/*
- * Implement vec_unpacklu and vec_unpackhu
- * since vec_unpackl, vec_unpackh only support signed integers
-**/
-#define VSX_IMPL_UNPACKU(rt, rg, zero)      \
-VSX_FINLINE(rt) vec_unpacklu(const rg& a)   \
-{ return (rt)(vec_mergel(a, zero)); }       \
-VSX_FINLINE(rt) vec_unpackhu(const rg& a)   \
-{ return (rt)(vec_mergeh(a, zero));  }
-
-VSX_IMPL_UNPACKU(vec_ushort8, vec_uchar16, vec_uchar16_z)
-VSX_IMPL_UNPACKU(vec_uint4,   vec_ushort8, vec_ushort8_z)
-VSX_IMPL_UNPACKU(vec_udword2, vec_uint4,   vec_uint4_z)
-
-/*
- * Implement vec_mergesqe and vec_mergesqo
- * Merges the sequence values of even and odd elements of two vectors
-*/
-#define VSX_IMPL_PERM(rt, fnm, ...)            \
-VSX_FINLINE(rt) fnm(const rt& a, const rt& b)  \
-{ static const vec_uchar16 perm = {__VA_ARGS__}; return vec_perm(a, b, perm); }
-
-// 16
-#define perm16_mergesqe 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
-#define perm16_mergesqo 1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31
-VSX_IMPL_PERM(vec_uchar16, vec_mergesqe, perm16_mergesqe)
-VSX_IMPL_PERM(vec_uchar16, vec_mergesqo, perm16_mergesqo)
-VSX_IMPL_PERM(vec_char16,  vec_mergesqe, perm16_mergesqe)
-VSX_IMPL_PERM(vec_char16,  vec_mergesqo, perm16_mergesqo)
-// 8
-#define perm8_mergesqe 0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29
-#define perm8_mergesqo 2, 3, 6, 7, 10, 11, 14, 15, 18, 19, 22, 23, 26, 27, 30, 31
-VSX_IMPL_PERM(vec_ushort8, vec_mergesqe, perm8_mergesqe)
-VSX_IMPL_PERM(vec_ushort8, vec_mergesqo, perm8_mergesqo)
-VSX_IMPL_PERM(vec_short8,  vec_mergesqe, perm8_mergesqe)
-VSX_IMPL_PERM(vec_short8,  vec_mergesqo, perm8_mergesqo)
-// 4
-#define perm4_mergesqe 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27
-#define perm4_mergesqo 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
-VSX_IMPL_PERM(vec_uint4,  vec_mergesqe, perm4_mergesqe)
-VSX_IMPL_PERM(vec_uint4,  vec_mergesqo, perm4_mergesqo)
-VSX_IMPL_PERM(vec_int4,   vec_mergesqe, perm4_mergesqe)
-VSX_IMPL_PERM(vec_int4,   vec_mergesqo, perm4_mergesqo)
-VSX_IMPL_PERM(vec_float4, vec_mergesqe, perm4_mergesqe)
-VSX_IMPL_PERM(vec_float4, vec_mergesqo, perm4_mergesqo)
-// 2
-VSX_REDIRECT_2RG(vec_double2, vec_double2, vec_mergesqe, vec_mergeh)
-VSX_REDIRECT_2RG(vec_double2, vec_double2, vec_mergesqo, vec_mergel)
-VSX_REDIRECT_2RG(vec_dword2,  vec_dword2,  vec_mergesqe, vec_mergeh)
-VSX_REDIRECT_2RG(vec_dword2,  vec_dword2,  vec_mergesqo, vec_mergel)
-VSX_REDIRECT_2RG(vec_udword2, vec_udword2, vec_mergesqe, vec_mergeh)
-VSX_REDIRECT_2RG(vec_udword2, vec_udword2, vec_mergesqo, vec_mergel)
-
-/*
- * Implement vec_mergesqh and vec_mergesql
- * Merges the sequence most and least significant halves of two vectors
-*/
-#define VSX_IMPL_MERGESQHL(Tvec)                                    \
-VSX_FINLINE(Tvec) vec_mergesqh(const Tvec& a, const Tvec& b)        \
-{ return (Tvec)vec_mergeh(vec_udword2_c(a), vec_udword2_c(b)); }    \
-VSX_FINLINE(Tvec) vec_mergesql(const Tvec& a, const Tvec& b)        \
-{ return (Tvec)vec_mergel(vec_udword2_c(a), vec_udword2_c(b)); }
-VSX_IMPL_MERGESQHL(vec_uchar16)
-VSX_IMPL_MERGESQHL(vec_char16)
-VSX_IMPL_MERGESQHL(vec_ushort8)
-VSX_IMPL_MERGESQHL(vec_short8)
-VSX_IMPL_MERGESQHL(vec_uint4)
-VSX_IMPL_MERGESQHL(vec_int4)
-VSX_IMPL_MERGESQHL(vec_float4)
-VSX_REDIRECT_2RG(vec_udword2, vec_udword2, vec_mergesqh, vec_mergeh)
-VSX_REDIRECT_2RG(vec_udword2, vec_udword2, vec_mergesql, vec_mergel)
-VSX_REDIRECT_2RG(vec_dword2,  vec_dword2,  vec_mergesqh, vec_mergeh)
-VSX_REDIRECT_2RG(vec_dword2,  vec_dword2,  vec_mergesql, vec_mergel)
-VSX_REDIRECT_2RG(vec_double2, vec_double2, vec_mergesqh, vec_mergeh)
-VSX_REDIRECT_2RG(vec_double2, vec_double2, vec_mergesql, vec_mergel)
-
-
-// 2 and 4 channels interleave for all types except 2 lanes
-#define VSX_IMPL_ST_INTERLEAVE(Tp, Tvec)                                    \
-VSX_FINLINE(void) vec_st_interleave(const Tvec& a, const Tvec& b, Tp* ptr)  \
-{                                                                           \
-    vsx_stf(vec_mergeh(a, b), 0, ptr);                                      \
-    vsx_stf(vec_mergel(a, b), 16, ptr);                                     \
-}                                                                           \
-VSX_FINLINE(void) vec_st_interleave(const Tvec& a, const Tvec& b,           \
-                                     const Tvec& c, const Tvec& d, Tp* ptr) \
-{                                                                           \
-    Tvec ac = vec_mergeh(a, c);                                             \
-    Tvec bd = vec_mergeh(b, d);                                             \
-    vsx_stf(vec_mergeh(ac, bd), 0, ptr);                                    \
-    vsx_stf(vec_mergel(ac, bd), 16, ptr);                                   \
-    ac = vec_mergel(a, c);                                                  \
-    bd = vec_mergel(b, d);                                                  \
-    vsx_stf(vec_mergeh(ac, bd), 32, ptr);                                   \
-    vsx_stf(vec_mergel(ac, bd), 48, ptr);                                   \
-}
-VSX_IMPL_ST_INTERLEAVE(uchar,  vec_uchar16)
-VSX_IMPL_ST_INTERLEAVE(schar,  vec_char16)
-VSX_IMPL_ST_INTERLEAVE(ushort, vec_ushort8)
-VSX_IMPL_ST_INTERLEAVE(short,  vec_short8)
-VSX_IMPL_ST_INTERLEAVE(uint,   vec_uint4)
-VSX_IMPL_ST_INTERLEAVE(int,    vec_int4)
-VSX_IMPL_ST_INTERLEAVE(float,  vec_float4)
-
-// 2 and 4 channels deinterleave for 16 lanes
-#define VSX_IMPL_ST_DINTERLEAVE_8(Tp, Tvec)                                 \
-VSX_FINLINE(void) vec_ld_deinterleave(const Tp* ptr, Tvec& a, Tvec& b)      \
-{                                                                           \
-    Tvec v0 = vsx_ld(0, ptr);                                               \
-    Tvec v1 = vsx_ld(16, ptr);                                              \
-    a = vec_mergesqe(v0, v1);                                               \
-    b = vec_mergesqo(v0, v1);                                               \
-}                                                                           \
-VSX_FINLINE(void) vec_ld_deinterleave(const Tp* ptr, Tvec& a, Tvec& b,      \
-                                       Tvec& c, Tvec& d)                    \
-{                                                                           \
-    Tvec v0 = vsx_ld(0, ptr);                                               \
-    Tvec v1 = vsx_ld(16, ptr);                                              \
-    Tvec v2 = vsx_ld(32, ptr);                                              \
-    Tvec v3 = vsx_ld(48, ptr);                                              \
-    Tvec m0 = vec_mergesqe(v0, v1);                                         \
-    Tvec m1 = vec_mergesqe(v2, v3);                                         \
-    a = vec_mergesqe(m0, m1);                                               \
-    c = vec_mergesqo(m0, m1);                                               \
-    m0 = vec_mergesqo(v0, v1);                                              \
-    m1 = vec_mergesqo(v2, v3);                                              \
-    b = vec_mergesqe(m0, m1);                                               \
-    d = vec_mergesqo(m0, m1);                                               \
-}
-VSX_IMPL_ST_DINTERLEAVE_8(uchar, vec_uchar16)
-VSX_IMPL_ST_DINTERLEAVE_8(schar, vec_char16)
-
-// 2 and 4 channels deinterleave for 8 lanes
-#define VSX_IMPL_ST_DINTERLEAVE_16(Tp, Tvec)                                \
-VSX_FINLINE(void) vec_ld_deinterleave(const Tp* ptr, Tvec& a, Tvec& b)      \
-{                                                                           \
-    Tvec v0 = vsx_ld(0, ptr);                                               \
-    Tvec v1 = vsx_ld(8, ptr);                                               \
-    a = vec_mergesqe(v0, v1);                                               \
-    b = vec_mergesqo(v0, v1);                                               \
-}                                                                           \
-VSX_FINLINE(void) vec_ld_deinterleave(const Tp* ptr, Tvec& a, Tvec& b,      \
-                                       Tvec& c, Tvec& d)                    \
-{                                                                           \
-    Tvec v0 = vsx_ld(0, ptr);                                               \
-    Tvec v1 = vsx_ld(8, ptr);                                               \
-    Tvec m0 = vec_mergeh(v0, v1);                                           \
-    Tvec m1 = vec_mergel(v0, v1);                                           \
-    Tvec ab0 = vec_mergeh(m0, m1);                                          \
-    Tvec cd0 = vec_mergel(m0, m1);                                          \
-    v0 = vsx_ld(16, ptr);                                                   \
-    v1 = vsx_ld(24, ptr);                                                   \
-    m0 = vec_mergeh(v0, v1);                                                \
-    m1 = vec_mergel(v0, v1);                                                \
-    Tvec ab1 = vec_mergeh(m0, m1);                                          \
-    Tvec cd1 = vec_mergel(m0, m1);                                          \
-    a = vec_mergesqh(ab0, ab1);                                             \
-    b = vec_mergesql(ab0, ab1);                                             \
-    c = vec_mergesqh(cd0, cd1);                                             \
-    d = vec_mergesql(cd0, cd1);                                             \
-}
-VSX_IMPL_ST_DINTERLEAVE_16(ushort, vec_ushort8)
-VSX_IMPL_ST_DINTERLEAVE_16(short,  vec_short8)
-
-// 2 and 4 channels deinterleave for 4 lanes
-#define VSX_IMPL_ST_DINTERLEAVE_32(Tp, Tvec)                                \
-VSX_FINLINE(void) vec_ld_deinterleave(const Tp* ptr, Tvec& a, Tvec& b)      \
-{                                                                           \
-    a = vsx_ld(0, ptr);                                                     \
-    b = vsx_ld(4, ptr);                                                     \
-    Tvec m0 = vec_mergeh(a, b);                                             \
-    Tvec m1 = vec_mergel(a, b);                                             \
-    a = vec_mergeh(m0, m1);                                                 \
-    b = vec_mergel(m0, m1);                                                 \
-}                                                                           \
-VSX_FINLINE(void) vec_ld_deinterleave(const Tp* ptr, Tvec& a, Tvec& b,      \
-                                       Tvec& c, Tvec& d)                    \
-{                                                                           \
-    Tvec v0 = vsx_ld(0, ptr);                                               \
-    Tvec v1 = vsx_ld(4, ptr);                                               \
-    Tvec v2 = vsx_ld(8, ptr);                                               \
-    Tvec v3 = vsx_ld(12, ptr);                                              \
-    Tvec m0 = vec_mergeh(v0, v2);                                           \
-    Tvec m1 = vec_mergeh(v1, v3);                                           \
-    a = vec_mergeh(m0, m1);                                                 \
-    b = vec_mergel(m0, m1);                                                 \
-    m0 = vec_mergel(v0, v2);                                                \
-    m1 = vec_mergel(v1, v3);                                                \
-    c = vec_mergeh(m0, m1);                                                 \
-    d = vec_mergel(m0, m1);                                                 \
-}
-VSX_IMPL_ST_DINTERLEAVE_32(uint,  vec_uint4)
-VSX_IMPL_ST_DINTERLEAVE_32(int,   vec_int4)
-VSX_IMPL_ST_DINTERLEAVE_32(float, vec_float4)
-
-// 2 and 4 channels interleave and deinterleave for 2 lanes
-#define VSX_IMPL_ST_D_INTERLEAVE_64(Tp, Tvec, ld_func, st_func)             \
-VSX_FINLINE(void) vec_st_interleave(const Tvec& a, const Tvec& b, Tp* ptr)  \
-{                                                                           \
-    st_func(vec_mergeh(a, b), 0, ptr);                                      \
-    st_func(vec_mergel(a, b), 2, ptr);                                      \
-}                                                                           \
-VSX_FINLINE(void) vec_st_interleave(const Tvec& a, const Tvec& b,           \
-                                     const Tvec& c, const Tvec& d, Tp* ptr) \
-{                                                                           \
-    st_func(vec_mergeh(a, b), 0, ptr);                                      \
-    st_func(vec_mergeh(c, d), 2, ptr);                                      \
-    st_func(vec_mergel(a, b), 4, ptr);                                      \
-    st_func(vec_mergel(c, d), 6, ptr);                                      \
-}                                                                           \
-VSX_FINLINE(void) vec_ld_deinterleave(const Tp* ptr, Tvec& a, Tvec& b)      \
-{                                                                           \
-    Tvec m0 = ld_func(0, ptr);                                              \
-    Tvec m1 = ld_func(2, ptr);                                              \
-    a = vec_mergeh(m0, m1);                                                 \
-    b = vec_mergel(m0, m1);                                                 \
-}                                                                           \
-VSX_FINLINE(void) vec_ld_deinterleave(const Tp* ptr, Tvec& a, Tvec& b,      \
-                                       Tvec& c, Tvec& d)                    \
-{                                                                           \
-    Tvec v0 = ld_func(0, ptr);                                              \
-    Tvec v1 = ld_func(2, ptr);                                              \
-    Tvec v2 = ld_func(4, ptr);                                              \
-    Tvec v3 = ld_func(6, ptr);                                              \
-    a = vec_mergeh(v0, v2);                                                 \
-    b = vec_mergel(v0, v2);                                                 \
-    c = vec_mergeh(v1, v3);                                                 \
-    d = vec_mergel(v1, v3);                                                 \
-}
-VSX_IMPL_ST_D_INTERLEAVE_64(int64,  vec_dword2,  vsx_ld2, vsx_st2)
-VSX_IMPL_ST_D_INTERLEAVE_64(uint64, vec_udword2, vsx_ld2, vsx_st2)
-VSX_IMPL_ST_D_INTERLEAVE_64(double, vec_double2, vsx_ld,  vsx_st)
-
-/* 3 channels */
-#define VSX_IMPL_ST_INTERLEAVE_3CH_16(Tp, Tvec)                                                   \
-VSX_FINLINE(void) vec_st_interleave(const Tvec& a, const Tvec& b,                                 \
-                                     const Tvec& c, Tp* ptr)                                      \
-{                                                                                                 \
-    static const vec_uchar16 a12 = {0, 16, 0, 1, 17, 0, 2, 18, 0, 3, 19, 0, 4, 20, 0, 5};         \
-    static const vec_uchar16 a123 = {0, 1, 16, 3, 4, 17, 6, 7, 18, 9, 10, 19, 12, 13, 20, 15};    \
-    vsx_st(vec_perm(vec_perm(a, b, a12), c, a123), 0, ptr);                                       \
-    static const vec_uchar16 b12 = {21, 0, 6, 22, 0, 7, 23, 0, 8, 24, 0, 9, 25, 0, 10, 26};       \
-    static const vec_uchar16 b123 = {0, 21, 2, 3, 22, 5, 6, 23, 8, 9, 24, 11, 12, 25, 14, 15};    \
-    vsx_st(vec_perm(vec_perm(a, b, b12), c, b123), 16, ptr);                                      \
-    static const vec_uchar16 c12 = {0, 11, 27, 0, 12, 28, 0, 13, 29, 0, 14, 30, 0, 15, 31, 0};    \
-    static const vec_uchar16 c123 = {26, 1, 2, 27, 4, 5, 28, 7, 8, 29, 10, 11, 30, 13, 14, 31};   \
-    vsx_st(vec_perm(vec_perm(a, b, c12), c, c123), 32, ptr);                                      \
-}                                                                                                 \
-VSX_FINLINE(void) vec_ld_deinterleave(const Tp* ptr, Tvec& a, Tvec& b, Tvec& c)                   \
-{                                                                                                 \
-    Tvec v1 = vsx_ld(0, ptr);                                                                     \
-    Tvec v2 = vsx_ld(16, ptr);                                                                    \
-    Tvec v3 = vsx_ld(32, ptr);                                                                    \
-    static const vec_uchar16 a12_perm = {0, 3, 6, 9, 12, 15, 18, 21, 24, 27, 30, 0, 0, 0, 0, 0};  \
-    static const vec_uchar16 a123_perm = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 17, 20, 23, 26, 29};  \
-    a = vec_perm(vec_perm(v1, v2, a12_perm), v3, a123_perm);                                      \
-    static const vec_uchar16 b12_perm = {1, 4, 7, 10, 13, 16, 19, 22, 25, 28, 31, 0, 0, 0, 0, 0}; \
-    static const vec_uchar16 b123_perm = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 18, 21, 24, 27, 30};  \
-    b = vec_perm(vec_perm(v1, v2, b12_perm), v3, b123_perm);                                      \
-    static const vec_uchar16 c12_perm = {2, 5, 8, 11, 14, 17, 20, 23, 26, 29, 0, 0, 0, 0, 0, 0};  \
-    static const vec_uchar16 c123_perm = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 16, 19, 22, 25, 28, 31};  \
-    c = vec_perm(vec_perm(v1, v2, c12_perm), v3, c123_perm);                                      \
-}
-VSX_IMPL_ST_INTERLEAVE_3CH_16(uchar, vec_uchar16)
-VSX_IMPL_ST_INTERLEAVE_3CH_16(schar, vec_char16)
-
-#define VSX_IMPL_ST_INTERLEAVE_3CH_8(Tp, Tvec)                                                    \
-VSX_FINLINE(void) vec_st_interleave(const Tvec& a, const Tvec& b,                                 \
-                                     const Tvec& c, Tp* ptr)                                      \
-{                                                                                                 \
-    static const vec_uchar16 a12 = {0, 1, 16, 17, 0, 0, 2, 3, 18, 19, 0, 0, 4, 5, 20, 21};        \
-    static const vec_uchar16 a123 = {0, 1, 2, 3, 16, 17, 6, 7, 8, 9, 18, 19, 12, 13, 14, 15};     \
-    vsx_st(vec_perm(vec_perm(a, b, a12), c, a123), 0, ptr);                                       \
-    static const vec_uchar16 b12 = {0, 0, 6, 7, 22, 23, 0, 0, 8, 9, 24, 25, 0, 0, 10, 11};        \
-    static const vec_uchar16 b123 = {20, 21, 2, 3, 4, 5, 22, 23, 8, 9, 10, 11, 24, 25, 14, 15};   \
-    vsx_st(vec_perm(vec_perm(a, b, b12), c, b123), 8, ptr);                                       \
-    static const vec_uchar16 c12 = {26, 27, 0, 0, 12, 13, 28, 29, 0, 0, 14, 15, 30, 31, 0, 0};    \
-    static const vec_uchar16 c123 = {0, 1, 26, 27, 4, 5, 6, 7, 28, 29, 10, 11, 12, 13, 30, 31};   \
-    vsx_st(vec_perm(vec_perm(a, b, c12), c, c123), 16, ptr);                                      \
-}                                                                                                 \
-VSX_FINLINE(void) vec_ld_deinterleave(const Tp* ptr, Tvec& a, Tvec& b, Tvec& c)                   \
-{                                                                                                 \
-    Tvec v1 = vsx_ld(0, ptr);                                                                     \
-    Tvec v2 = vsx_ld(8, ptr);                                                                     \
-    Tvec v3 = vsx_ld(16, ptr);                                                                    \
-    static const vec_uchar16 a12_perm = {0, 1, 6, 7, 12, 13, 18, 19, 24, 25, 30, 31, 0, 0, 0, 0}; \
-    static const vec_uchar16 a123_perm = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 20, 21, 26, 27};  \
-    a = vec_perm(vec_perm(v1, v2, a12_perm), v3, a123_perm);                                      \
-    static const vec_uchar16 b12_perm = {2, 3, 8, 9, 14, 15, 20, 21, 26, 27, 0, 0, 0, 0, 0, 0};   \
-    static const vec_uchar16 b123_perm = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 16, 17, 22, 23, 28, 29};  \
-    b = vec_perm(vec_perm(v1, v2, b12_perm), v3, b123_perm);                                      \
-    static const vec_uchar16 c12_perm = {4, 5, 10, 11, 16, 17, 22, 23, 28, 29, 0, 0, 0, 0, 0, 0}; \
-    static const vec_uchar16 c123_perm = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 18, 19, 24, 25, 30, 31};  \
-    c = vec_perm(vec_perm(v1, v2, c12_perm), v3, c123_perm);                                      \
-}
-VSX_IMPL_ST_INTERLEAVE_3CH_8(ushort, vec_ushort8)
-VSX_IMPL_ST_INTERLEAVE_3CH_8(short,  vec_short8)
-
-#define VSX_IMPL_ST_INTERLEAVE_3CH_4(Tp, Tvec)                                                     \
-VSX_FINLINE(void) vec_st_interleave(const Tvec& a, const Tvec& b,                                  \
-                                     const Tvec& c, Tp* ptr)                                       \
-{                                                                                                  \
-    Tvec hbc = vec_mergeh(b, c);                                                                   \
-    static const vec_uchar16 ahbc = {0, 1, 2, 3, 16, 17, 18, 19, 20, 21, 22, 23, 4, 5, 6, 7};      \
-    vsx_st(vec_perm(a, hbc, ahbc), 0, ptr);                                                        \
-    Tvec lab = vec_mergel(a, b);                                                                   \
-    vsx_st(vec_sld(lab, hbc, 8), 4, ptr);                                                          \
-    static const vec_uchar16 clab = {8, 9, 10, 11, 24, 25, 26, 27, 28, 29, 30, 31, 12, 13, 14, 15};\
-    vsx_st(vec_perm(c, lab, clab), 8, ptr);                                                        \
-}                                                                                                  \
-VSX_FINLINE(void) vec_ld_deinterleave(const Tp* ptr, Tvec& a, Tvec& b, Tvec& c)                    \
-{                                                                                                  \
-    Tvec v1 = vsx_ld(0, ptr);                                                                      \
-    Tvec v2 = vsx_ld(4, ptr);                                                                      \
-    Tvec v3 = vsx_ld(8, ptr);                                                                      \
-    static const vec_uchar16 flp = {0, 1, 2, 3, 12, 13, 14, 15, 16, 17, 18, 19, 28, 29, 30, 31};   \
-    a = vec_perm(v1, vec_sld(v3, v2, 8), flp);                                                     \
-    static const vec_uchar16 flp2 = {28, 29, 30, 31, 0, 1, 2, 3, 12, 13, 14, 15, 16, 17, 18, 19};  \
-    b = vec_perm(v2, vec_sld(v1, v3, 8), flp2);                                                    \
-    c = vec_perm(vec_sld(v2, v1, 8), v3, flp);                                                     \
-}
-VSX_IMPL_ST_INTERLEAVE_3CH_4(uint,  vec_uint4)
-VSX_IMPL_ST_INTERLEAVE_3CH_4(int,   vec_int4)
-VSX_IMPL_ST_INTERLEAVE_3CH_4(float, vec_float4)
-
-#define VSX_IMPL_ST_INTERLEAVE_3CH_2(Tp, Tvec, ld_func, st_func)     \
-VSX_FINLINE(void) vec_st_interleave(const Tvec& a, const Tvec& b,    \
-                                     const Tvec& c, Tp* ptr)         \
-{                                                                    \
-    st_func(vec_mergeh(a, b), 0, ptr);                               \
-    st_func(vec_permi(c, a, 1), 2, ptr);                             \
-    st_func(vec_mergel(b, c), 4, ptr);                               \
-}                                                                    \
-VSX_FINLINE(void) vec_ld_deinterleave(const Tp* ptr, Tvec& a,        \
-                                       Tvec& b, Tvec& c)             \
-{                                                                    \
-    Tvec v1 = ld_func(0, ptr);                                       \
-    Tvec v2 = ld_func(2, ptr);                                       \
-    Tvec v3 = ld_func(4, ptr);                                       \
-    a = vec_permi(v1, v2, 1);                                        \
-    b = vec_permi(v1, v3, 2);                                        \
-    c = vec_permi(v2, v3, 1);                                        \
-}
-VSX_IMPL_ST_INTERLEAVE_3CH_2(int64,  vec_dword2,  vsx_ld2, vsx_st2)
-VSX_IMPL_ST_INTERLEAVE_3CH_2(uint64, vec_udword2, vsx_ld2, vsx_st2)
-VSX_IMPL_ST_INTERLEAVE_3CH_2(double, vec_double2, vsx_ld,  vsx_st)
-
-#endif // CV_VSX
-
-//! @}
-
-#endif // OPENCV_HAL_VSX_UTILS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/cvconfig.h b/projects/llm_framework/include/opencv4/opencv2/cvconfig.h
deleted file mode 100644
index 47abe048..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/cvconfig.h
+++ /dev/null
@@ -1,149 +0,0 @@
-#ifndef OPENCV_CVCONFIG_H_INCLUDED
-#define OPENCV_CVCONFIG_H_INCLUDED
-
-/* OpenCV compiled as static or dynamic libs */
-/* #undef BUILD_SHARED_LIBS */
-
-/* OpenCV intrinsics optimized code */
-#define CV_ENABLE_INTRINSICS
-
-/* OpenCV additional optimized code */
-/* #undef CV_DISABLE_OPTIMIZATION */
-
-/* Compile for 'real' NVIDIA GPU architectures */
-#define CUDA_ARCH_BIN ""
-
-/* NVIDIA GPU features are used */
-#define CUDA_ARCH_FEATURES ""
-
-/* Compile for 'virtual' NVIDIA PTX architectures */
-#define CUDA_ARCH_PTX ""
-
-/* AMD's Basic Linear Algebra Subprograms Library*/
-/* #undef HAVE_CLAMDBLAS */
-
-/* AMD's OpenCL Fast Fourier Transform Library*/
-/* #undef HAVE_CLAMDFFT */
-
-/* Clp support */
-/* #undef HAVE_CLP */
-
-/* NVIDIA CUDA Runtime API*/
-/* #undef HAVE_CUDA */
-
-/* NVIDIA CUDA Basic Linear Algebra Subprograms (BLAS) API*/
-/* #undef HAVE_CUBLAS */
-
-/* NVIDIA CUDA Deep Neural Network (cuDNN) API*/
-/* #undef HAVE_CUDNN */
-
-/* NVIDIA CUDA Fast Fourier Transform (FFT) API*/
-/* #undef HAVE_CUFFT */
-
-/* DirectX */
-/* #undef HAVE_DIRECTX */
-/* #undef HAVE_DIRECTX_NV12 */
-/* #undef HAVE_D3D11 */
-/* #undef HAVE_D3D10 */
-/* #undef HAVE_D3D9 */
-
-/* Eigen Matrix & Linear Algebra Library */
-/* #undef HAVE_EIGEN */
-
-/* Geospatial Data Abstraction Library */
-/* #undef HAVE_GDAL */
-
-/* Halide support */
-/* #undef HAVE_HALIDE */
-
-/* Vulkan support */
-/* #undef HAVE_VULKAN */
-
-/* Define to 1 if you have the <inttypes.h> header file. */
-#define HAVE_INTTYPES_H 1
-
-/* Intel Integrated Performance Primitives */
-/* #undef HAVE_IPP */
-/* #undef HAVE_IPP_ICV */
-/* #undef HAVE_IPP_IW */
-/* #undef HAVE_IPP_IW_LL */
-
-/* JPEG-2000 codec */
-#define HAVE_OPENJPEG
-/* #undef HAVE_JASPER */
-
-/* IJG JPEG codec */
-#define HAVE_JPEG
-
-/* libpng/png.h needs to be included */
-/* #undef HAVE_LIBPNG_PNG_H */
-
-/* GDCM DICOM codec */
-/* #undef HAVE_GDCM */
-
-/* NVIDIA Video Decoding API*/
-/* #undef HAVE_NVCUVID */
-/* #undef HAVE_NVCUVID_HEADER */
-/* #undef HAVE_DYNLINK_NVCUVID_HEADER */
-
-/* NVIDIA Video Encoding API*/
-/* #undef HAVE_NVCUVENC */
-
-/* OpenCL Support */
-/* #undef HAVE_OPENCL */
-/* #undef HAVE_OPENCL_STATIC */
-/* #undef HAVE_OPENCL_SVM */
-
-/* NVIDIA OpenCL D3D Extensions support */
-/* #undef HAVE_OPENCL_D3D11_NV */
-
-/* OpenEXR codec */
-/* #undef HAVE_OPENEXR */
-
-/* OpenGL support*/
-/* #undef HAVE_OPENGL */
-
-/* PNG codec */
-#define HAVE_PNG
-
-/* Posix threads (pthreads) */
-#define HAVE_PTHREAD
-
-/* parallel_for with pthreads */
-#define HAVE_PTHREADS_PF
-
-/* Intel Threading Building Blocks */
-/* #undef HAVE_TBB */
-
-/* Ste||ar Group High Performance ParallelX */
-/* #undef HAVE_HPX */
-
-/* TIFF codec */
-#define HAVE_TIFF
-
-/* Define if your processor stores words with the most significant byte
-   first (like Motorola and SPARC, unlike Intel and VAX). */
-/* #undef WORDS_BIGENDIAN */
-
-/* VA library (libva) */
-/* #undef HAVE_VA */
-
-/* Intel VA-API/OpenCL */
-/* #undef HAVE_VA_INTEL */
-
-/* Lapack */
-/* #undef HAVE_LAPACK */
-
-/* Library was compiled with functions instrumentation */
-/* #undef ENABLE_INSTRUMENTATION */
-
-/* OpenVX */
-/* #undef HAVE_OPENVX */
-
-/* OpenCV trace utilities */
-#define OPENCV_TRACE
-
-/* Library QR-code decoding */
-/* #undef HAVE_QUIRC */
-
-#endif // OPENCV_CVCONFIG_H_INCLUDED
diff --git a/projects/llm_framework/include/opencv4/opencv2/highgui.hpp b/projects/llm_framework/include/opencv4/opencv2/highgui.hpp
deleted file mode 100644
index 7ae23957..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/highgui.hpp
+++ /dev/null
@@ -1,868 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_HIGHGUI_HPP
-#define OPENCV_HIGHGUI_HPP
-
-#include "opencv2/core.hpp"
-#ifdef HAVE_OPENCV_IMGCODECS
-#include "opencv2/imgcodecs.hpp"
-#endif
-#ifdef HAVE_OPENCV_VIDEOIO
-#include "opencv2/videoio.hpp"
-#endif
-
-/**
-@defgroup highgui High-level GUI
-
-While OpenCV was designed for use in full-scale applications and can be used within functionally
-rich UI frameworks (such as Qt\*, WinForms\*, or Cocoa\*) or without any UI at all, sometimes there
-it is required to try functionality quickly and visualize the results. This is what the HighGUI
-module has been designed for.
-
-It provides easy interface to:
-
--   Create and manipulate windows that can display images and "remember" their content (no need to
-    handle repaint events from OS).
--   Add trackbars to the windows, handle simple mouse events as well as keyboard commands.
-
-@{
-    @defgroup highgui_window_flags Flags related creating and manipulating HighGUI windows and mouse events
-    @defgroup highgui_opengl OpenGL support
-    @defgroup highgui_qt Qt New Functions
-
-    ![image](pics/qtgui.png)
-
-    This figure explains new functionality implemented with Qt\* GUI. The new GUI provides a statusbar,
-    a toolbar, and a control panel. The control panel can have trackbars and buttonbars attached to it.
-    If you cannot see the control panel, press Ctrl+P or right-click any Qt window and select **Display
-    properties window**.
-
-    -   To attach a trackbar, the window name parameter must be NULL.
-
-    -   To attach a buttonbar, a button must be created. If the last bar attached to the control panel
-        is a buttonbar, the new button is added to the right of the last button. If the last bar
-        attached to the control panel is a trackbar, or the control panel is empty, a new buttonbar is
-        created. Then, a new button is attached to it.
-
-    See below the example used to generate the figure:
-    @code
-        int main(int argc, char *argv[])
-        {
-
-            int value = 50;
-            int value2 = 0;
-
-
-            namedWindow("main1",WINDOW_NORMAL);
-            namedWindow("main2",WINDOW_AUTOSIZE | WINDOW_GUI_NORMAL);
-            createTrackbar( "track1", "main1", &value, 255,  NULL);
-
-            String nameb1 = "button1";
-            String nameb2 = "button2";
-
-            createButton(nameb1,callbackButton,&nameb1,QT_CHECKBOX,1);
-            createButton(nameb2,callbackButton,NULL,QT_CHECKBOX,0);
-            createTrackbar( "track2", NULL, &value2, 255, NULL);
-            createButton("button5",callbackButton1,NULL,QT_RADIOBOX,0);
-            createButton("button6",callbackButton2,NULL,QT_RADIOBOX,1);
-
-            setMouseCallback( "main2",on_mouse,NULL );
-
-            Mat img1 = imread("files/flower.jpg");
-            VideoCapture video;
-            video.open("files/hockey.avi");
-
-            Mat img2,img3;
-
-            while( waitKey(33) != 27 )
-            {
-                img1.convertTo(img2,-1,1,value);
-                video >> img3;
-
-                imshow("main1",img2);
-                imshow("main2",img3);
-            }
-
-            destroyAllWindows();
-
-            return 0;
-        }
-    @endcode
-
-
-    @defgroup highgui_winrt WinRT support
-
-    This figure explains new functionality implemented with WinRT GUI. The new GUI provides an Image control,
-    and a slider panel. Slider panel holds trackbars attached to it.
-
-    Sliders are attached below the image control. Every new slider is added below the previous one.
-
-    See below the example used to generate the figure:
-    @code
-        void sample_app::MainPage::ShowWindow()
-        {
-            static cv::String windowName("sample");
-            cv::winrt_initContainer(this->cvContainer);
-            cv::namedWindow(windowName); // not required
-
-            cv::Mat image = cv::imread("Assets/sample.jpg");
-            cv::Mat converted = cv::Mat(image.rows, image.cols, CV_8UC4);
-            cv::cvtColor(image, converted, COLOR_BGR2BGRA);
-            cv::imshow(windowName, converted); // this will create window if it hasn't been created before
-
-            int state = 42;
-            cv::TrackbarCallback callback = [](int pos, void* userdata)
-            {
-                if (pos == 0) {
-                    cv::destroyWindow(windowName);
-                }
-            };
-            cv::TrackbarCallback callbackTwin = [](int pos, void* userdata)
-            {
-                if (pos >= 70) {
-                    cv::destroyAllWindows();
-                }
-            };
-            cv::createTrackbar("Sample trackbar", windowName, &state, 100, callback);
-            cv::createTrackbar("Twin brother", windowName, &state, 100, callbackTwin);
-        }
-    @endcode
-
-    @defgroup highgui_c C API
-@}
-*/
-
-///////////////////////// graphical user interface //////////////////////////
-namespace cv
-{
-
-//! @addtogroup highgui
-//! @{
-
-//! @addtogroup highgui_window_flags
-//! @{
-
-//! Flags for cv::namedWindow
-enum WindowFlags {
-       WINDOW_NORMAL     = 0x00000000, //!< the user can resize the window (no constraint) / also use to switch a fullscreen window to a normal size.
-       WINDOW_AUTOSIZE   = 0x00000001, //!< the user cannot resize the window, the size is constrainted by the image displayed.
-       WINDOW_OPENGL     = 0x00001000, //!< window with opengl support.
-
-       WINDOW_FULLSCREEN = 1,          //!< change the window to fullscreen.
-       WINDOW_FREERATIO  = 0x00000100, //!< the image expends as much as it can (no ratio constraint).
-       WINDOW_KEEPRATIO  = 0x00000000, //!< the ratio of the image is respected.
-       WINDOW_GUI_EXPANDED=0x00000000, //!< status bar and tool bar
-       WINDOW_GUI_NORMAL = 0x00000010, //!< old fashious way
-    };
-
-//! Flags for cv::setWindowProperty / cv::getWindowProperty
-enum WindowPropertyFlags {
-       WND_PROP_FULLSCREEN   = 0, //!< fullscreen property    (can be WINDOW_NORMAL or WINDOW_FULLSCREEN).
-       WND_PROP_AUTOSIZE     = 1, //!< autosize property      (can be WINDOW_NORMAL or WINDOW_AUTOSIZE).
-       WND_PROP_ASPECT_RATIO = 2, //!< window's aspect ration (can be set to WINDOW_FREERATIO or WINDOW_KEEPRATIO).
-       WND_PROP_OPENGL       = 3, //!< opengl support.
-       WND_PROP_VISIBLE      = 4, //!< checks whether the window exists and is visible
-       WND_PROP_TOPMOST      = 5, //!< property to toggle normal window being topmost or not
-       WND_PROP_VSYNC        = 6  //!< enable or disable VSYNC (in OpenGL mode)
-     };
-
-//! Mouse Events see cv::MouseCallback
-enum MouseEventTypes {
-       EVENT_MOUSEMOVE      = 0, //!< indicates that the mouse pointer has moved over the window.
-       EVENT_LBUTTONDOWN    = 1, //!< indicates that the left mouse button is pressed.
-       EVENT_RBUTTONDOWN    = 2, //!< indicates that the right mouse button is pressed.
-       EVENT_MBUTTONDOWN    = 3, //!< indicates that the middle mouse button is pressed.
-       EVENT_LBUTTONUP      = 4, //!< indicates that left mouse button is released.
-       EVENT_RBUTTONUP      = 5, //!< indicates that right mouse button is released.
-       EVENT_MBUTTONUP      = 6, //!< indicates that middle mouse button is released.
-       EVENT_LBUTTONDBLCLK  = 7, //!< indicates that left mouse button is double clicked.
-       EVENT_RBUTTONDBLCLK  = 8, //!< indicates that right mouse button is double clicked.
-       EVENT_MBUTTONDBLCLK  = 9, //!< indicates that middle mouse button is double clicked.
-       EVENT_MOUSEWHEEL     = 10,//!< positive and negative values mean forward and backward scrolling, respectively.
-       EVENT_MOUSEHWHEEL    = 11 //!< positive and negative values mean right and left scrolling, respectively.
-     };
-
-//! Mouse Event Flags see cv::MouseCallback
-enum MouseEventFlags {
-       EVENT_FLAG_LBUTTON   = 1, //!< indicates that the left mouse button is down.
-       EVENT_FLAG_RBUTTON   = 2, //!< indicates that the right mouse button is down.
-       EVENT_FLAG_MBUTTON   = 4, //!< indicates that the middle mouse button is down.
-       EVENT_FLAG_CTRLKEY   = 8, //!< indicates that CTRL Key is pressed.
-       EVENT_FLAG_SHIFTKEY  = 16,//!< indicates that SHIFT Key is pressed.
-       EVENT_FLAG_ALTKEY    = 32 //!< indicates that ALT Key is pressed.
-     };
-
-//! @} highgui_window_flags
-
-//! @addtogroup highgui_qt
-//! @{
-
-//! Qt font weight
-enum QtFontWeights {
-        QT_FONT_LIGHT           = 25, //!< Weight of 25
-        QT_FONT_NORMAL          = 50, //!< Weight of 50
-        QT_FONT_DEMIBOLD        = 63, //!< Weight of 63
-        QT_FONT_BOLD            = 75, //!< Weight of 75
-        QT_FONT_BLACK           = 87  //!< Weight of 87
-     };
-
-//! Qt font style
-enum QtFontStyles {
-        QT_STYLE_NORMAL         = 0, //!< Normal font.
-        QT_STYLE_ITALIC         = 1, //!< Italic font.
-        QT_STYLE_OBLIQUE        = 2  //!< Oblique font.
-     };
-
-//! Qt "button" type
-enum QtButtonTypes {
-       QT_PUSH_BUTTON   = 0,    //!< Push button.
-       QT_CHECKBOX      = 1,    //!< Checkbox button.
-       QT_RADIOBOX      = 2,    //!< Radiobox button.
-       QT_NEW_BUTTONBAR = 1024  //!< Button should create a new buttonbar
-     };
-
-//! @} highgui_qt
-
-/** @brief Callback function for mouse events. see cv::setMouseCallback
-@param event one of the cv::MouseEventTypes constants.
-@param x The x-coordinate of the mouse event.
-@param y The y-coordinate of the mouse event.
-@param flags one of the cv::MouseEventFlags constants.
-@param userdata The optional parameter.
- */
-typedef void (*MouseCallback)(int event, int x, int y, int flags, void* userdata);
-
-/** @brief Callback function for Trackbar see cv::createTrackbar
-@param pos current position of the specified trackbar.
-@param userdata The optional parameter.
- */
-typedef void (*TrackbarCallback)(int pos, void* userdata);
-
-/** @brief Callback function defined to be called every frame. See cv::setOpenGlDrawCallback
-@param userdata The optional parameter.
- */
-typedef void (*OpenGlDrawCallback)(void* userdata);
-
-/** @brief Callback function for a button created by cv::createButton
-@param state current state of the button. It could be -1 for a push button, 0 or 1 for a check/radio box button.
-@param userdata The optional parameter.
- */
-typedef void (*ButtonCallback)(int state, void* userdata);
-
-/** @brief Creates a window.
-
-The function namedWindow creates a window that can be used as a placeholder for images and
-trackbars. Created windows are referred to by their names.
-
-If a window with the same name already exists, the function does nothing.
-
-You can call cv::destroyWindow or cv::destroyAllWindows to close the window and de-allocate any associated
-memory usage. For a simple program, you do not really have to call these functions because all the
-resources and windows of the application are closed automatically by the operating system upon exit.
-
-@note
-
-Qt backend supports additional flags:
- -   **WINDOW_NORMAL or WINDOW_AUTOSIZE:** WINDOW_NORMAL enables you to resize the
-     window, whereas WINDOW_AUTOSIZE adjusts automatically the window size to fit the
-     displayed image (see imshow ), and you cannot change the window size manually.
- -   **WINDOW_FREERATIO or WINDOW_KEEPRATIO:** WINDOW_FREERATIO adjusts the image
-     with no respect to its ratio, whereas WINDOW_KEEPRATIO keeps the image ratio.
- -   **WINDOW_GUI_NORMAL or WINDOW_GUI_EXPANDED:** WINDOW_GUI_NORMAL is the old way to draw the window
-     without statusbar and toolbar, whereas WINDOW_GUI_EXPANDED is a new enhanced GUI.
-By default, flags == WINDOW_AUTOSIZE | WINDOW_KEEPRATIO | WINDOW_GUI_EXPANDED
-
-@param winname Name of the window in the window caption that may be used as a window identifier.
-@param flags Flags of the window. The supported flags are: (cv::WindowFlags)
- */
-CV_EXPORTS_W void namedWindow(const String& winname, int flags = WINDOW_AUTOSIZE);
-
-/** @brief Destroys the specified window.
-
-The function destroyWindow destroys the window with the given name.
-
-@param winname Name of the window to be destroyed.
- */
-CV_EXPORTS_W void destroyWindow(const String& winname);
-
-/** @brief Destroys all of the HighGUI windows.
-
-The function destroyAllWindows destroys all of the opened HighGUI windows.
- */
-CV_EXPORTS_W void destroyAllWindows();
-
-CV_EXPORTS_W int startWindowThread();
-
-/** @brief Similar to #waitKey, but returns full key code.
-
-@note
-
-Key code is implementation specific and depends on used backend: QT/GTK/Win32/etc
-
-*/
-CV_EXPORTS_W int waitKeyEx(int delay = 0);
-
-/** @brief Waits for a pressed key.
-
-The function waitKey waits for a key event infinitely (when \f$\texttt{delay}\leq 0\f$ ) or for delay
-milliseconds, when it is positive. Since the OS has a minimum time between switching threads, the
-function will not wait exactly delay ms, it will wait at least delay ms, depending on what else is
-running on your computer at that time. It returns the code of the pressed key or -1 if no key was
-pressed before the specified time had elapsed. To check for a key press but not wait for it, use
-#pollKey.
-
-@note The functions #waitKey and #pollKey are the only methods in HighGUI that can fetch and handle
-GUI events, so one of them needs to be called periodically for normal event processing unless
-HighGUI is used within an environment that takes care of event processing.
-
-@note The function only works if there is at least one HighGUI window created and the window is
-active. If there are several HighGUI windows, any of them can be active.
-
-@param delay Delay in milliseconds. 0 is the special value that means "forever".
- */
-CV_EXPORTS_W int waitKey(int delay = 0);
-
-/** @brief Polls for a pressed key.
-
-The function pollKey polls for a key event without waiting. It returns the code of the pressed key
-or -1 if no key was pressed since the last invocation. To wait until a key was pressed, use #waitKey.
-
-@note The functions #waitKey and #pollKey are the only methods in HighGUI that can fetch and handle
-GUI events, so one of them needs to be called periodically for normal event processing unless
-HighGUI is used within an environment that takes care of event processing.
-
-@note The function only works if there is at least one HighGUI window created and the window is
-active. If there are several HighGUI windows, any of them can be active.
- */
-CV_EXPORTS_W int pollKey();
-
-/** @brief Displays an image in the specified window.
-
-The function imshow displays an image in the specified window. If the window was created with the
-cv::WINDOW_AUTOSIZE flag, the image is shown with its original size, however it is still limited by the screen resolution.
-Otherwise, the image is scaled to fit the window. The function may scale the image, depending on its depth:
-
--   If the image is 8-bit unsigned, it is displayed as is.
--   If the image is 16-bit unsigned, the pixels are divided by 256. That is, the
-    value range [0,255\*256] is mapped to [0,255].
--   If the image is 32-bit or 64-bit floating-point, the pixel values are multiplied by 255. That is, the
-    value range [0,1] is mapped to [0,255].
--   32-bit integer images are not processed anymore due to ambiguouty of required transform.
-    Convert to 8-bit unsigned matrix using a custom preprocessing specific to image's context.
-
-If window was created with OpenGL support, cv::imshow also support ogl::Buffer , ogl::Texture2D and
-cuda::GpuMat as input.
-
-If the window was not created before this function, it is assumed creating a window with cv::WINDOW_AUTOSIZE.
-
-If you need to show an image that is bigger than the screen resolution, you will need to call namedWindow("", WINDOW_NORMAL) before the imshow.
-
-@note This function should be followed by a call to cv::waitKey or cv::pollKey to perform GUI
-housekeeping tasks that are necessary to actually show the given image and make the window respond
-to mouse and keyboard events. Otherwise, it won't display the image and the window might lock up.
-For example, **waitKey(0)** will display the window infinitely until any keypress (it is suitable
-for image display). **waitKey(25)** will display a frame and wait approximately 25 ms for a key
-press (suitable for displaying a video frame-by-frame). To remove the window, use cv::destroyWindow.
-
-@note
-
-[__Windows Backend Only__] Pressing Ctrl+C will copy the image to the clipboard.
-
-[__Windows Backend Only__] Pressing Ctrl+S will show a dialog to save the image.
-
-@param winname Name of the window.
-@param mat Image to be shown.
- */
-CV_EXPORTS_W void imshow(const String& winname, InputArray mat);
-
-/** @brief Resizes the window to the specified size
-
-@note
-
--   The specified window size is for the image area. Toolbars are not counted.
--   Only windows created without cv::WINDOW_AUTOSIZE flag can be resized.
-
-@param winname Window name.
-@param width The new window width.
-@param height The new window height.
- */
-CV_EXPORTS_W void resizeWindow(const String& winname, int width, int height);
-
-/** @overload
-@param winname Window name.
-@param size The new window size.
-*/
-CV_EXPORTS_W void resizeWindow(const String& winname, const cv::Size& size);
-
-/** @brief Moves the window to the specified position
-
-@param winname Name of the window.
-@param x The new x-coordinate of the window.
-@param y The new y-coordinate of the window.
- */
-CV_EXPORTS_W void moveWindow(const String& winname, int x, int y);
-
-/** @brief Changes parameters of a window dynamically.
-
-The function setWindowProperty enables changing properties of a window.
-
-@param winname Name of the window.
-@param prop_id Window property to edit. The supported operation flags are: (cv::WindowPropertyFlags)
-@param prop_value New value of the window property. The supported flags are: (cv::WindowFlags)
- */
-CV_EXPORTS_W void setWindowProperty(const String& winname, int prop_id, double prop_value);
-
-/** @brief Updates window title
-@param winname Name of the window.
-@param title New title.
-*/
-CV_EXPORTS_W void setWindowTitle(const String& winname, const String& title);
-
-/** @brief Provides parameters of a window.
-
-The function getWindowProperty returns properties of a window.
-
-@param winname Name of the window.
-@param prop_id Window property to retrieve. The following operation flags are available: (cv::WindowPropertyFlags)
-
-@sa setWindowProperty
- */
-CV_EXPORTS_W double getWindowProperty(const String& winname, int prop_id);
-
-/** @brief Provides rectangle of image in the window.
-
-The function getWindowImageRect returns the client screen coordinates, width and height of the image rendering area.
-
-@param winname Name of the window.
-
-@sa resizeWindow moveWindow
- */
-CV_EXPORTS_W Rect getWindowImageRect(const String& winname);
-
-/** @example samples/cpp/create_mask.cpp
-This program demonstrates using mouse events and how to make and use a mask image (black and white) .
-*/
-/** @brief Sets mouse handler for the specified window
-
-@param winname Name of the window.
-@param onMouse Callback function for mouse events. See OpenCV samples on how to specify and use the callback.
-@param userdata The optional parameter passed to the callback.
- */
-CV_EXPORTS void setMouseCallback(const String& winname, MouseCallback onMouse, void* userdata = 0);
-
-/** @brief Gets the mouse-wheel motion delta, when handling mouse-wheel events cv::EVENT_MOUSEWHEEL and
-cv::EVENT_MOUSEHWHEEL.
-
-For regular mice with a scroll-wheel, delta will be a multiple of 120. The value 120 corresponds to
-a one notch rotation of the wheel or the threshold for action to be taken and one such action should
-occur for each delta. Some high-precision mice with higher-resolution freely-rotating wheels may
-generate smaller values.
-
-For cv::EVENT_MOUSEWHEEL positive and negative values mean forward and backward scrolling,
-respectively. For cv::EVENT_MOUSEHWHEEL, where available, positive and negative values mean right and
-left scrolling, respectively.
-
-@note
-
-Mouse-wheel events are currently supported only on Windows.
-
-@param flags The mouse callback flags parameter.
- */
-CV_EXPORTS int getMouseWheelDelta(int flags);
-
-/** @brief Allows users to select a ROI on the given image.
-
-The function creates a window and allows users to select a ROI using the mouse.
-Controls: use `space` or `enter` to finish selection, use key `c` to cancel selection (function will return the zero cv::Rect).
-
-@param windowName name of the window where selection process will be shown.
-@param img image to select a ROI.
-@param showCrosshair if true crosshair of selection rectangle will be shown.
-@param fromCenter if true center of selection will match initial mouse position. In opposite case a corner of
-selection rectangle will correspont to the initial mouse position.
-@return selected ROI or empty rect if selection canceled.
-
-@note The function sets it's own mouse callback for specified window using cv::setMouseCallback(windowName, ...).
-After finish of work an empty callback will be set for the used window.
- */
-CV_EXPORTS_W Rect selectROI(const String& windowName, InputArray img, bool showCrosshair = true, bool fromCenter = false);
-
-/** @overload
- */
-CV_EXPORTS_W Rect selectROI(InputArray img, bool showCrosshair = true, bool fromCenter = false);
-
-/** @brief Allows users to select multiple ROIs on the given image.
-
-The function creates a window and allows users to select multiple ROIs using the mouse.
-Controls: use `space` or `enter` to finish current selection and start a new one,
-use `esc` to terminate multiple ROI selection process.
-
-@param windowName name of the window where selection process will be shown.
-@param img image to select a ROI.
-@param boundingBoxes selected ROIs.
-@param showCrosshair if true crosshair of selection rectangle will be shown.
-@param fromCenter if true center of selection will match initial mouse position. In opposite case a corner of
-selection rectangle will correspont to the initial mouse position.
-
-@note The function sets it's own mouse callback for specified window using cv::setMouseCallback(windowName, ...).
-After finish of work an empty callback will be set for the used window.
- */
-CV_EXPORTS_W void selectROIs(const String& windowName, InputArray img,
-                             CV_OUT std::vector<Rect>& boundingBoxes, bool showCrosshair = true, bool fromCenter = false);
-
-/** @brief Creates a trackbar and attaches it to the specified window.
-
-The function createTrackbar creates a trackbar (a slider or range control) with the specified name
-and range, assigns a variable value to be a position synchronized with the trackbar and specifies
-the callback function onChange to be called on the trackbar position change. The created trackbar is
-displayed in the specified window winname.
-
-@note
-
-[__Qt Backend Only__] winname can be empty if the trackbar should be attached to the
-control panel.
-
-Clicking the label of each trackbar enables editing the trackbar values manually.
-
-@param trackbarname Name of the created trackbar.
-@param winname Name of the window that will be used as a parent of the created trackbar.
-@param value Optional pointer to an integer variable whose value reflects the position of the
-slider. Upon creation, the slider position is defined by this variable.
-@param count Maximal position of the slider. The minimal position is always 0.
-@param onChange Pointer to the function to be called every time the slider changes position. This
-function should be prototyped as void Foo(int,void\*); , where the first parameter is the trackbar
-position and the second parameter is the user data (see the next parameter). If the callback is
-the NULL pointer, no callbacks are called, but only value is updated.
-@param userdata User data that is passed as is to the callback. It can be used to handle trackbar
-events without using global variables.
- */
-CV_EXPORTS int createTrackbar(const String& trackbarname, const String& winname,
-                              int* value, int count,
-                              TrackbarCallback onChange = 0,
-                              void* userdata = 0);
-
-/** @brief Returns the trackbar position.
-
-The function returns the current position of the specified trackbar.
-
-@note
-
-[__Qt Backend Only__] winname can be empty if the trackbar is attached to the control
-panel.
-
-@param trackbarname Name of the trackbar.
-@param winname Name of the window that is the parent of the trackbar.
- */
-CV_EXPORTS_W int getTrackbarPos(const String& trackbarname, const String& winname);
-
-/** @brief Sets the trackbar position.
-
-The function sets the position of the specified trackbar in the specified window.
-
-@note
-
-[__Qt Backend Only__] winname can be empty if the trackbar is attached to the control
-panel.
-
-@param trackbarname Name of the trackbar.
-@param winname Name of the window that is the parent of trackbar.
-@param pos New position.
- */
-CV_EXPORTS_W void setTrackbarPos(const String& trackbarname, const String& winname, int pos);
-
-/** @brief Sets the trackbar maximum position.
-
-The function sets the maximum position of the specified trackbar in the specified window.
-
-@note
-
-[__Qt Backend Only__] winname can be empty if the trackbar is attached to the control
-panel.
-
-@param trackbarname Name of the trackbar.
-@param winname Name of the window that is the parent of trackbar.
-@param maxval New maximum position.
- */
-CV_EXPORTS_W void setTrackbarMax(const String& trackbarname, const String& winname, int maxval);
-
-/** @brief Sets the trackbar minimum position.
-
-The function sets the minimum position of the specified trackbar in the specified window.
-
-@note
-
-[__Qt Backend Only__] winname can be empty if the trackbar is attached to the control
-panel.
-
-@param trackbarname Name of the trackbar.
-@param winname Name of the window that is the parent of trackbar.
-@param minval New minimum position.
- */
-CV_EXPORTS_W void setTrackbarMin(const String& trackbarname, const String& winname, int minval);
-
-//! @addtogroup highgui_opengl OpenGL support
-//! @{
-
-/** @brief Displays OpenGL 2D texture in the specified window.
-
-@param winname Name of the window.
-@param tex OpenGL 2D texture data.
- */
-CV_EXPORTS void imshow(const String& winname, const ogl::Texture2D& tex);
-
-/** @brief Sets a callback function to be called to draw on top of displayed image.
-
-The function setOpenGlDrawCallback can be used to draw 3D data on the window. See the example of
-callback function below:
-@code
-    void on_opengl(void* param)
-    {
-        glLoadIdentity();
-
-        glTranslated(0.0, 0.0, -1.0);
-
-        glRotatef( 55, 1, 0, 0 );
-        glRotatef( 45, 0, 1, 0 );
-        glRotatef( 0, 0, 0, 1 );
-
-        static const int coords[6][4][3] = {
-            { { +1, -1, -1 }, { -1, -1, -1 }, { -1, +1, -1 }, { +1, +1, -1 } },
-            { { +1, +1, -1 }, { -1, +1, -1 }, { -1, +1, +1 }, { +1, +1, +1 } },
-            { { +1, -1, +1 }, { +1, -1, -1 }, { +1, +1, -1 }, { +1, +1, +1 } },
-            { { -1, -1, -1 }, { -1, -1, +1 }, { -1, +1, +1 }, { -1, +1, -1 } },
-            { { +1, -1, +1 }, { -1, -1, +1 }, { -1, -1, -1 }, { +1, -1, -1 } },
-            { { -1, -1, +1 }, { +1, -1, +1 }, { +1, +1, +1 }, { -1, +1, +1 } }
-        };
-
-        for (int i = 0; i < 6; ++i) {
-                    glColor3ub( i*20, 100+i*10, i*42 );
-                    glBegin(GL_QUADS);
-                    for (int j = 0; j < 4; ++j) {
-                            glVertex3d(0.2 * coords[i][j][0], 0.2 * coords[i][j][1], 0.2 * coords[i][j][2]);
-                    }
-                    glEnd();
-        }
-    }
-@endcode
-
-@param winname Name of the window.
-@param onOpenGlDraw Pointer to the function to be called every frame. This function should be
-prototyped as void Foo(void\*) .
-@param userdata Pointer passed to the callback function.(__Optional__)
- */
-CV_EXPORTS void setOpenGlDrawCallback(const String& winname, OpenGlDrawCallback onOpenGlDraw, void* userdata = 0);
-
-/** @brief Sets the specified window as current OpenGL context.
-
-@param winname Name of the window.
- */
-CV_EXPORTS void setOpenGlContext(const String& winname);
-
-/** @brief Force window to redraw its context and call draw callback ( See cv::setOpenGlDrawCallback ).
-
-@param winname Name of the window.
- */
-CV_EXPORTS void updateWindow(const String& winname);
-
-//! @} highgui_opengl
-
-//! @addtogroup highgui_qt
-//! @{
-
-/** @brief QtFont available only for Qt. See cv::fontQt
- */
-struct QtFont
-{
-    const char* nameFont;  //!< Name of the font
-    Scalar      color;     //!< Color of the font. Scalar(blue_component, green_component, red_component[, alpha_component])
-    int         font_face; //!< See cv::QtFontStyles
-    const int*  ascii;     //!< font data and metrics
-    const int*  greek;
-    const int*  cyrillic;
-    float       hscale, vscale;
-    float       shear;     //!< slope coefficient: 0 - normal, >0 - italic
-    int         thickness; //!< See cv::QtFontWeights
-    float       dx;        //!< horizontal interval between letters
-    int         line_type; //!< PointSize
-};
-
-/** @brief Creates the font to draw a text on an image.
-
-The function fontQt creates a cv::QtFont object. This cv::QtFont is not compatible with putText .
-
-A basic usage of this function is the following: :
-@code
-    QtFont font = fontQt("Times");
-    addText( img1, "Hello World !", Point(50,50), font);
-@endcode
-
-@param nameFont Name of the font. The name should match the name of a system font (such as
-*Times*). If the font is not found, a default one is used.
-@param pointSize Size of the font. If not specified, equal zero or negative, the point size of the
-font is set to a system-dependent default value. Generally, this is 12 points.
-@param color Color of the font in BGRA where A = 255 is fully transparent. Use the macro CV_RGB
-for simplicity.
-@param weight Font weight. Available operation flags are : cv::QtFontWeights You can also specify a positive integer for better control.
-@param style Font style. Available operation flags are : cv::QtFontStyles
-@param spacing Spacing between characters. It can be negative or positive.
- */
-CV_EXPORTS QtFont fontQt(const String& nameFont, int pointSize = -1,
-                         Scalar color = Scalar::all(0), int weight = QT_FONT_NORMAL,
-                         int style = QT_STYLE_NORMAL, int spacing = 0);
-
-/** @brief Draws a text on the image.
-
-The function addText draws *text* on the image *img* using a specific font *font* (see example cv::fontQt
-)
-
-@param img 8-bit 3-channel image where the text should be drawn.
-@param text Text to write on an image.
-@param org Point(x,y) where the text should start on an image.
-@param font Font to use to draw a text.
- */
-CV_EXPORTS void addText( const Mat& img, const String& text, Point org, const QtFont& font);
-
-/** @brief Draws a text on the image.
-
-@param img 8-bit 3-channel image where the text should be drawn.
-@param text Text to write on an image.
-@param org Point(x,y) where the text should start on an image.
-@param nameFont Name of the font. The name should match the name of a system font (such as
-*Times*). If the font is not found, a default one is used.
-@param pointSize Size of the font. If not specified, equal zero or negative, the point size of the
-font is set to a system-dependent default value. Generally, this is 12 points.
-@param color Color of the font in BGRA where A = 255 is fully transparent.
-@param weight Font weight. Available operation flags are : cv::QtFontWeights You can also specify a positive integer for better control.
-@param style Font style. Available operation flags are : cv::QtFontStyles
-@param spacing Spacing between characters. It can be negative or positive.
- */
-CV_EXPORTS_W void addText(const Mat& img, const String& text, Point org, const String& nameFont, int pointSize = -1, Scalar color = Scalar::all(0),
-        int weight = QT_FONT_NORMAL, int style = QT_STYLE_NORMAL, int spacing = 0);
-
-/** @brief Displays a text on a window image as an overlay for a specified duration.
-
-The function displayOverlay displays useful information/tips on top of the window for a certain
-amount of time *delayms*. The function does not modify the image, displayed in the window, that is,
-after the specified delay the original content of the window is restored.
-
-@param winname Name of the window.
-@param text Overlay text to write on a window image.
-@param delayms The period (in milliseconds), during which the overlay text is displayed. If this
-function is called before the previous overlay text timed out, the timer is restarted and the text
-is updated. If this value is zero, the text never disappears.
- */
-CV_EXPORTS_W void displayOverlay(const String& winname, const String& text, int delayms = 0);
-
-/** @brief Displays a text on the window statusbar during the specified period of time.
-
-The function displayStatusBar displays useful information/tips on top of the window for a certain
-amount of time *delayms* . This information is displayed on the window statusbar (the window must be
-created with the CV_GUI_EXPANDED flags).
-
-@param winname Name of the window.
-@param text Text to write on the window statusbar.
-@param delayms Duration (in milliseconds) to display the text. If this function is called before
-the previous text timed out, the timer is restarted and the text is updated. If this value is
-zero, the text never disappears.
- */
-CV_EXPORTS_W void displayStatusBar(const String& winname, const String& text, int delayms = 0);
-
-/** @brief Saves parameters of the specified window.
-
-The function saveWindowParameters saves size, location, flags, trackbars value, zoom and panning
-location of the window windowName.
-
-@param windowName Name of the window.
- */
-CV_EXPORTS void saveWindowParameters(const String& windowName);
-
-/** @brief Loads parameters of the specified window.
-
-The function loadWindowParameters loads size, location, flags, trackbars value, zoom and panning
-location of the window windowName.
-
-@param windowName Name of the window.
- */
-CV_EXPORTS void loadWindowParameters(const String& windowName);
-
-CV_EXPORTS  int startLoop(int (*pt2Func)(int argc, char *argv[]), int argc, char* argv[]);
-
-CV_EXPORTS  void stopLoop();
-
-/** @brief Attaches a button to the control panel.
-
-The function createButton attaches a button to the control panel. Each button is added to a
-buttonbar to the right of the last button. A new buttonbar is created if nothing was attached to the
-control panel before, or if the last element attached to the control panel was a trackbar or if the
-QT_NEW_BUTTONBAR flag is added to the type.
-
-See below various examples of the cv::createButton function call: :
-@code
-    createButton("",callbackButton);//create a push button "button 0", that will call callbackButton.
-    createButton("button2",callbackButton,NULL,QT_CHECKBOX,0);
-    createButton("button3",callbackButton,&value);
-    createButton("button5",callbackButton1,NULL,QT_RADIOBOX);
-    createButton("button6",callbackButton2,NULL,QT_PUSH_BUTTON,1);
-    createButton("button6",callbackButton2,NULL,QT_PUSH_BUTTON|QT_NEW_BUTTONBAR);// create a push button in a new row
-@endcode
-
-@param  bar_name Name of the button.
-@param on_change Pointer to the function to be called every time the button changes its state.
-This function should be prototyped as void Foo(int state,\*void); . *state* is the current state
-of the button. It could be -1 for a push button, 0 or 1 for a check/radio box button.
-@param userdata Pointer passed to the callback function.
-@param type Optional type of the button. Available types are: (cv::QtButtonTypes)
-@param initial_button_state Default state of the button. Use for checkbox and radiobox. Its
-value could be 0 or 1. (__Optional__)
-*/
-CV_EXPORTS int createButton( const String& bar_name, ButtonCallback on_change,
-                             void* userdata = 0, int type = QT_PUSH_BUTTON,
-                             bool initial_button_state = false);
-
-//! @} highgui_qt
-
-//! @} highgui
-
-} // cv
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/highgui/highgui.hpp b/projects/llm_framework/include/opencv4/opencv2/highgui/highgui.hpp
deleted file mode 100644
index 160c9cf4..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/highgui/highgui.hpp
+++ /dev/null
@@ -1,48 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifdef __OPENCV_BUILD
-#error this is a compatibility header which should not be used inside the OpenCV library
-#endif
-
-#include "opencv2/highgui.hpp"
diff --git a/projects/llm_framework/include/opencv4/opencv2/highgui/highgui_c.h b/projects/llm_framework/include/opencv4/opencv2/highgui/highgui_c.h
deleted file mode 100644
index e508e149..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/highgui/highgui_c.h
+++ /dev/null
@@ -1,251 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                        Intel License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000, Intel Corporation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of Intel Corporation may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_HIGHGUI_H
-#define OPENCV_HIGHGUI_H
-
-#include "opencv2/core/core_c.h"
-#include "opencv2/imgproc/imgproc_c.h"
-
-#ifdef __cplusplus
-extern "C" {
-#endif /* __cplusplus */
-
-/** @addtogroup highgui_c
-  @{
-  */
-
-/****************************************************************************************\
-*                                  Basic GUI functions                                   *
-\****************************************************************************************/
-//YV
-//-----------New for Qt
-/* For font */
-enum {  CV_FONT_LIGHT           = 25,//QFont::Light,
-        CV_FONT_NORMAL          = 50,//QFont::Normal,
-        CV_FONT_DEMIBOLD        = 63,//QFont::DemiBold,
-        CV_FONT_BOLD            = 75,//QFont::Bold,
-        CV_FONT_BLACK           = 87 //QFont::Black
-};
-
-enum {  CV_STYLE_NORMAL         = 0,//QFont::StyleNormal,
-        CV_STYLE_ITALIC         = 1,//QFont::StyleItalic,
-        CV_STYLE_OBLIQUE        = 2 //QFont::StyleOblique
-};
-/* ---------*/
-
-//for color cvScalar(blue_component, green_component, red_component[, alpha_component])
-//and alpha= 0 <-> 0xFF (not transparent <-> transparent)
-CVAPI(CvFont) cvFontQt(const char* nameFont, int pointSize CV_DEFAULT(-1), CvScalar color CV_DEFAULT(cvScalarAll(0)), int weight CV_DEFAULT(CV_FONT_NORMAL),  int style CV_DEFAULT(CV_STYLE_NORMAL), int spacing CV_DEFAULT(0));
-
-CVAPI(void) cvAddText(const CvArr* img, const char* text, CvPoint org, CvFont *arg2);
-
-CVAPI(void) cvDisplayOverlay(const char* name, const char* text, int delayms CV_DEFAULT(0));
-CVAPI(void) cvDisplayStatusBar(const char* name, const char* text, int delayms CV_DEFAULT(0));
-
-CVAPI(void) cvSaveWindowParameters(const char* name);
-CVAPI(void) cvLoadWindowParameters(const char* name);
-CVAPI(int) cvStartLoop(int (*pt2Func)(int argc, char *argv[]), int argc, char* argv[]);
-CVAPI(void) cvStopLoop( void );
-
-typedef void (CV_CDECL *CvButtonCallback)(int state, void* userdata);
-enum {CV_PUSH_BUTTON = 0, CV_CHECKBOX = 1, CV_RADIOBOX = 2};
-CVAPI(int) cvCreateButton( const char* button_name CV_DEFAULT(NULL),CvButtonCallback on_change CV_DEFAULT(NULL), void* userdata CV_DEFAULT(NULL) , int button_type CV_DEFAULT(CV_PUSH_BUTTON), int initial_button_state CV_DEFAULT(0));
-//----------------------
-
-
-/* this function is used to set some external parameters in case of X Window */
-CVAPI(int) cvInitSystem( int argc, char** argv );
-
-CVAPI(int) cvStartWindowThread( void );
-
-// ---------  YV ---------
-enum
-{
-    //These 3 flags are used by cvSet/GetWindowProperty
-    CV_WND_PROP_FULLSCREEN = 0, //to change/get window's fullscreen property
-    CV_WND_PROP_AUTOSIZE   = 1, //to change/get window's autosize property
-    CV_WND_PROP_ASPECTRATIO= 2, //to change/get window's aspectratio property
-    CV_WND_PROP_OPENGL     = 3, //to change/get window's opengl support
-    CV_WND_PROP_VISIBLE    = 4,
-
-    //These 2 flags are used by cvNamedWindow and cvSet/GetWindowProperty
-    CV_WINDOW_NORMAL       = 0x00000000, //the user can resize the window (no constraint)  / also use to switch a fullscreen window to a normal size
-    CV_WINDOW_AUTOSIZE     = 0x00000001, //the user cannot resize the window, the size is constrainted by the image displayed
-    CV_WINDOW_OPENGL       = 0x00001000, //window with opengl support
-
-    //Those flags are only for Qt
-    CV_GUI_EXPANDED         = 0x00000000, //status bar and tool bar
-    CV_GUI_NORMAL           = 0x00000010, //old fashious way
-
-    //These 3 flags are used by cvNamedWindow and cvSet/GetWindowProperty
-    CV_WINDOW_FULLSCREEN   = 1,//change the window to fullscreen
-    CV_WINDOW_FREERATIO    = 0x00000100,//the image expends as much as it can (no ratio constraint)
-    CV_WINDOW_KEEPRATIO    = 0x00000000//the ration image is respected.
-};
-
-/* create window */
-CVAPI(int) cvNamedWindow( const char* name, int flags CV_DEFAULT(CV_WINDOW_AUTOSIZE) );
-
-/* Set and Get Property of the window */
-CVAPI(void) cvSetWindowProperty(const char* name, int prop_id, double prop_value);
-CVAPI(double) cvGetWindowProperty(const char* name, int prop_id);
-
-/* display image within window (highgui windows remember their content) */
-CVAPI(void) cvShowImage( const char* name, const CvArr* image );
-
-/* resize/move window */
-CVAPI(void) cvResizeWindow( const char* name, int width, int height );
-CVAPI(void) cvMoveWindow( const char* name, int x, int y );
-
-
-/* destroy window and all the trackers associated with it */
-CVAPI(void) cvDestroyWindow( const char* name );
-
-CVAPI(void) cvDestroyAllWindows(void);
-
-/* get native window handle (HWND in case of Win32 and Widget in case of X Window) */
-CVAPI(void*) cvGetWindowHandle( const char* name );
-
-/* get name of highgui window given its native handle */
-CVAPI(const char*) cvGetWindowName( void* window_handle );
-
-
-typedef void (CV_CDECL *CvTrackbarCallback)(int pos);
-
-/* create trackbar and display it on top of given window, set callback */
-CVAPI(int) cvCreateTrackbar( const char* trackbar_name, const char* window_name,
-                             int* value, int count, CvTrackbarCallback on_change CV_DEFAULT(NULL));
-
-typedef void (CV_CDECL *CvTrackbarCallback2)(int pos, void* userdata);
-
-CVAPI(int) cvCreateTrackbar2( const char* trackbar_name, const char* window_name,
-                              int* value, int count, CvTrackbarCallback2 on_change,
-                              void* userdata CV_DEFAULT(0));
-
-/* retrieve or set trackbar position */
-CVAPI(int) cvGetTrackbarPos( const char* trackbar_name, const char* window_name );
-CVAPI(void) cvSetTrackbarPos( const char* trackbar_name, const char* window_name, int pos );
-CVAPI(void) cvSetTrackbarMax(const char* trackbar_name, const char* window_name, int maxval);
-CVAPI(void) cvSetTrackbarMin(const char* trackbar_name, const char* window_name, int minval);
-
-enum
-{
-    CV_EVENT_MOUSEMOVE      =0,
-    CV_EVENT_LBUTTONDOWN    =1,
-    CV_EVENT_RBUTTONDOWN    =2,
-    CV_EVENT_MBUTTONDOWN    =3,
-    CV_EVENT_LBUTTONUP      =4,
-    CV_EVENT_RBUTTONUP      =5,
-    CV_EVENT_MBUTTONUP      =6,
-    CV_EVENT_LBUTTONDBLCLK  =7,
-    CV_EVENT_RBUTTONDBLCLK  =8,
-    CV_EVENT_MBUTTONDBLCLK  =9,
-    CV_EVENT_MOUSEWHEEL     =10,
-    CV_EVENT_MOUSEHWHEEL    =11
-};
-
-enum
-{
-    CV_EVENT_FLAG_LBUTTON   =1,
-    CV_EVENT_FLAG_RBUTTON   =2,
-    CV_EVENT_FLAG_MBUTTON   =4,
-    CV_EVENT_FLAG_CTRLKEY   =8,
-    CV_EVENT_FLAG_SHIFTKEY  =16,
-    CV_EVENT_FLAG_ALTKEY    =32
-};
-
-
-#define CV_GET_WHEEL_DELTA(flags) ((short)((flags >> 16) & 0xffff)) // upper 16 bits
-
-typedef void (CV_CDECL *CvMouseCallback )(int event, int x, int y, int flags, void* param);
-
-/* assign callback for mouse events */
-CVAPI(void) cvSetMouseCallback( const char* window_name, CvMouseCallback on_mouse,
-                                void* param CV_DEFAULT(NULL));
-
-/* wait for key event infinitely (delay<=0) or for "delay" milliseconds */
-CVAPI(int) cvWaitKey(int delay CV_DEFAULT(0));
-
-// OpenGL support
-
-typedef void (CV_CDECL *CvOpenGlDrawCallback)(void* userdata);
-CVAPI(void) cvSetOpenGlDrawCallback(const char* window_name, CvOpenGlDrawCallback callback, void* userdata CV_DEFAULT(NULL));
-
-CVAPI(void) cvSetOpenGlContext(const char* window_name);
-CVAPI(void) cvUpdateWindow(const char* window_name);
-
-
-/****************************************************************************************\
-
-*                              Obsolete functions/synonyms                               *
-\****************************************************************************************/
-
-#define cvAddSearchPath(path)
-#define cvvInitSystem cvInitSystem
-#define cvvNamedWindow cvNamedWindow
-#define cvvShowImage cvShowImage
-#define cvvResizeWindow cvResizeWindow
-#define cvvDestroyWindow cvDestroyWindow
-#define cvvCreateTrackbar cvCreateTrackbar
-#define cvvAddSearchPath cvAddSearchPath
-#define cvvWaitKey(name) cvWaitKey(0)
-#define cvvWaitKeyEx(name,delay) cvWaitKey(delay)
-#define HG_AUTOSIZE CV_WINDOW_AUTOSIZE
-#define set_preprocess_func cvSetPreprocessFuncWin32
-#define set_postprocess_func cvSetPostprocessFuncWin32
-
-#if defined _WIN32
-
-CVAPI(void) cvSetPreprocessFuncWin32_(const void* callback);
-CVAPI(void) cvSetPostprocessFuncWin32_(const void* callback);
-#define cvSetPreprocessFuncWin32(callback) cvSetPreprocessFuncWin32_((const void*)(callback))
-#define cvSetPostprocessFuncWin32(callback) cvSetPostprocessFuncWin32_((const void*)(callback))
-
-#endif
-
-/** @} highgui_c */
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/imgcodecs.hpp b/projects/llm_framework/include/opencv4/opencv2/imgcodecs.hpp
deleted file mode 100644
index d7ff9a17..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/imgcodecs.hpp
+++ /dev/null
@@ -1,330 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_IMGCODECS_HPP
-#define OPENCV_IMGCODECS_HPP
-
-#include "opencv2/core.hpp"
-
-/**
-  @defgroup imgcodecs Image file reading and writing
-  @{
-    @defgroup imgcodecs_c C API
-    @defgroup imgcodecs_flags Flags used for image file reading and writing
-    @defgroup imgcodecs_ios iOS glue
-    @defgroup imgcodecs_macosx MacOS(OSX) glue
-  @}
-*/
-
-//////////////////////////////// image codec ////////////////////////////////
-namespace cv
-{
-
-//! @addtogroup imgcodecs
-//! @{
-
-//! @addtogroup imgcodecs_flags
-//! @{
-
-//! Imread flags
-enum ImreadModes {
-       IMREAD_UNCHANGED            = -1, //!< If set, return the loaded image as is (with alpha channel, otherwise it gets cropped). Ignore EXIF orientation.
-       IMREAD_GRAYSCALE            = 0,  //!< If set, always convert image to the single channel grayscale image (codec internal conversion).
-       IMREAD_COLOR                = 1,  //!< If set, always convert image to the 3 channel BGR color image.
-       IMREAD_ANYDEPTH             = 2,  //!< If set, return 16-bit/32-bit image when the input has the corresponding depth, otherwise convert it to 8-bit.
-       IMREAD_ANYCOLOR             = 4,  //!< If set, the image is read in any possible color format.
-       IMREAD_LOAD_GDAL            = 8,  //!< If set, use the gdal driver for loading the image.
-       IMREAD_REDUCED_GRAYSCALE_2  = 16, //!< If set, always convert image to the single channel grayscale image and the image size reduced 1/2.
-       IMREAD_REDUCED_COLOR_2      = 17, //!< If set, always convert image to the 3 channel BGR color image and the image size reduced 1/2.
-       IMREAD_REDUCED_GRAYSCALE_4  = 32, //!< If set, always convert image to the single channel grayscale image and the image size reduced 1/4.
-       IMREAD_REDUCED_COLOR_4      = 33, //!< If set, always convert image to the 3 channel BGR color image and the image size reduced 1/4.
-       IMREAD_REDUCED_GRAYSCALE_8  = 64, //!< If set, always convert image to the single channel grayscale image and the image size reduced 1/8.
-       IMREAD_REDUCED_COLOR_8      = 65, //!< If set, always convert image to the 3 channel BGR color image and the image size reduced 1/8.
-       IMREAD_IGNORE_ORIENTATION   = 128 //!< If set, do not rotate the image according to EXIF's orientation flag.
-     };
-
-//! Imwrite flags
-enum ImwriteFlags {
-       IMWRITE_JPEG_QUALITY        = 1,  //!< For JPEG, it can be a quality from 0 to 100 (the higher is the better). Default value is 95.
-       IMWRITE_JPEG_PROGRESSIVE    = 2,  //!< Enable JPEG features, 0 or 1, default is False.
-       IMWRITE_JPEG_OPTIMIZE       = 3,  //!< Enable JPEG features, 0 or 1, default is False.
-       IMWRITE_JPEG_RST_INTERVAL   = 4,  //!< JPEG restart interval, 0 - 65535, default is 0 - no restart.
-       IMWRITE_JPEG_LUMA_QUALITY   = 5,  //!< Separate luma quality level, 0 - 100, default is 0 - don't use.
-       IMWRITE_JPEG_CHROMA_QUALITY = 6,  //!< Separate chroma quality level, 0 - 100, default is 0 - don't use.
-       IMWRITE_PNG_COMPRESSION     = 16, //!< For PNG, it can be the compression level from 0 to 9. A higher value means a smaller size and longer compression time. If specified, strategy is changed to IMWRITE_PNG_STRATEGY_DEFAULT (Z_DEFAULT_STRATEGY). Default value is 1 (best speed setting).
-       IMWRITE_PNG_STRATEGY        = 17, //!< One of cv::ImwritePNGFlags, default is IMWRITE_PNG_STRATEGY_RLE.
-       IMWRITE_PNG_BILEVEL         = 18, //!< Binary level PNG, 0 or 1, default is 0.
-       IMWRITE_PXM_BINARY          = 32, //!< For PPM, PGM, or PBM, it can be a binary format flag, 0 or 1. Default value is 1.
-       IMWRITE_EXR_TYPE            = (3 << 4) + 0, /* 48 */ //!< override EXR storage type (FLOAT (FP32) is default)
-       IMWRITE_EXR_COMPRESSION     = (3 << 4) + 1, /* 49 */ //!< override EXR compression type (ZIP_COMPRESSION = 3 is default)
-       IMWRITE_WEBP_QUALITY        = 64, //!< For WEBP, it can be a quality from 1 to 100 (the higher is the better). By default (without any parameter) and for quality above 100 the lossless compression is used.
-       IMWRITE_PAM_TUPLETYPE       = 128,//!< For PAM, sets the TUPLETYPE field to the corresponding string value that is defined for the format
-       IMWRITE_TIFF_RESUNIT = 256,//!< For TIFF, use to specify which DPI resolution unit to set; see libtiff documentation for valid values
-       IMWRITE_TIFF_XDPI = 257,//!< For TIFF, use to specify the X direction DPI
-       IMWRITE_TIFF_YDPI = 258, //!< For TIFF, use to specify the Y direction DPI
-       IMWRITE_TIFF_COMPRESSION = 259, //!< For TIFF, use to specify the image compression scheme. See libtiff for integer constants corresponding to compression formats. Note, for images whose depth is CV_32F, only libtiff's SGILOG compression scheme is used. For other supported depths, the compression scheme can be specified by this flag; LZW compression is the default.
-       IMWRITE_JPEG2000_COMPRESSION_X1000 = 272 //!< For JPEG2000, use to specify the target compression rate (multiplied by 1000). The value can be from 0 to 1000. Default is 1000.
-     };
-
-enum ImwriteEXRTypeFlags {
-       /*IMWRITE_EXR_TYPE_UNIT = 0, //!< not supported */
-       IMWRITE_EXR_TYPE_HALF = 1,   //!< store as HALF (FP16)
-       IMWRITE_EXR_TYPE_FLOAT = 2   //!< store as FP32 (default)
-     };
-
-enum ImwriteEXRCompressionFlags {
-       IMWRITE_EXR_COMPRESSION_NO    = 0, //!< no compression
-       IMWRITE_EXR_COMPRESSION_RLE   = 1, //!< run length encoding
-       IMWRITE_EXR_COMPRESSION_ZIPS  = 2, //!< zlib compression, one scan line at a time
-       IMWRITE_EXR_COMPRESSION_ZIP   = 3, //!< zlib compression, in blocks of 16 scan lines
-       IMWRITE_EXR_COMPRESSION_PIZ   = 4, //!< piz-based wavelet compression
-       IMWRITE_EXR_COMPRESSION_PXR24 = 5, //!< lossy 24-bit float compression
-       IMWRITE_EXR_COMPRESSION_B44   = 6, //!< lossy 4-by-4 pixel block compression, fixed compression rate
-       IMWRITE_EXR_COMPRESSION_B44A  = 7, //!< lossy 4-by-4 pixel block compression, flat fields are compressed more
-       IMWRITE_EXR_COMPRESSION_DWAA  = 8, //!< lossy DCT based compression, in blocks of 32 scanlines. More efficient for partial buffer access. Supported since OpenEXR 2.2.0.
-       IMWRITE_EXR_COMPRESSION_DWAB  = 9, //!< lossy DCT based compression, in blocks of 256 scanlines. More efficient space wise and faster to decode full frames than DWAA_COMPRESSION. Supported since OpenEXR 2.2.0.
-     };
-
-//! Imwrite PNG specific flags used to tune the compression algorithm.
-/** These flags will be modify the way of PNG image compression and will be passed to the underlying zlib processing stage.
-
--   The effect of IMWRITE_PNG_STRATEGY_FILTERED is to force more Huffman coding and less string matching; it is somewhat intermediate between IMWRITE_PNG_STRATEGY_DEFAULT and IMWRITE_PNG_STRATEGY_HUFFMAN_ONLY.
--   IMWRITE_PNG_STRATEGY_RLE is designed to be almost as fast as IMWRITE_PNG_STRATEGY_HUFFMAN_ONLY, but give better compression for PNG image data.
--   The strategy parameter only affects the compression ratio but not the correctness of the compressed output even if it is not set appropriately.
--   IMWRITE_PNG_STRATEGY_FIXED prevents the use of dynamic Huffman codes, allowing for a simpler decoder for special applications.
-*/
-enum ImwritePNGFlags {
-       IMWRITE_PNG_STRATEGY_DEFAULT      = 0, //!< Use this value for normal data.
-       IMWRITE_PNG_STRATEGY_FILTERED     = 1, //!< Use this value for data produced by a filter (or predictor).Filtered data consists mostly of small values with a somewhat random distribution. In this case, the compression algorithm is tuned to compress them better.
-       IMWRITE_PNG_STRATEGY_HUFFMAN_ONLY = 2, //!< Use this value to force Huffman encoding only (no string match).
-       IMWRITE_PNG_STRATEGY_RLE          = 3, //!< Use this value to limit match distances to one (run-length encoding).
-       IMWRITE_PNG_STRATEGY_FIXED        = 4  //!< Using this value prevents the use of dynamic Huffman codes, allowing for a simpler decoder for special applications.
-     };
-
-//! Imwrite PAM specific tupletype flags used to define the 'TUPETYPE' field of a PAM file.
-enum ImwritePAMFlags {
-       IMWRITE_PAM_FORMAT_NULL = 0,
-       IMWRITE_PAM_FORMAT_BLACKANDWHITE = 1,
-       IMWRITE_PAM_FORMAT_GRAYSCALE = 2,
-       IMWRITE_PAM_FORMAT_GRAYSCALE_ALPHA = 3,
-       IMWRITE_PAM_FORMAT_RGB = 4,
-       IMWRITE_PAM_FORMAT_RGB_ALPHA = 5,
-     };
-
-//! @} imgcodecs_flags
-
-/** @brief Loads an image from a file.
-
-@anchor imread
-
-The function imread loads an image from the specified file and returns it. If the image cannot be
-read (because of missing file, improper permissions, unsupported or invalid format), the function
-returns an empty matrix ( Mat::data==NULL ).
-
-Currently, the following file formats are supported:
-
--   Windows bitmaps - \*.bmp, \*.dib (always supported)
--   JPEG files - \*.jpeg, \*.jpg, \*.jpe (see the *Note* section)
--   JPEG 2000 files - \*.jp2 (see the *Note* section)
--   Portable Network Graphics - \*.png (see the *Note* section)
--   WebP - \*.webp (see the *Note* section)
--   Portable image format - \*.pbm, \*.pgm, \*.ppm \*.pxm, \*.pnm (always supported)
--   PFM files - \*.pfm (see the *Note* section)
--   Sun rasters - \*.sr, \*.ras (always supported)
--   TIFF files - \*.tiff, \*.tif (see the *Note* section)
--   OpenEXR Image files - \*.exr (see the *Note* section)
--   Radiance HDR - \*.hdr, \*.pic (always supported)
--   Raster and Vector geospatial data supported by GDAL (see the *Note* section)
-
-@note
--   The function determines the type of an image by the content, not by the file extension.
--   In the case of color images, the decoded images will have the channels stored in **B G R** order.
--   When using IMREAD_GRAYSCALE, the codec's internal grayscale conversion will be used, if available.
-    Results may differ to the output of cvtColor()
--   On Microsoft Windows\* OS and MacOSX\*, the codecs shipped with an OpenCV image (libjpeg,
-    libpng, libtiff, and libjasper) are used by default. So, OpenCV can always read JPEGs, PNGs,
-    and TIFFs. On MacOSX, there is also an option to use native MacOSX image readers. But beware
-    that currently these native image loaders give images with different pixel values because of
-    the color management embedded into MacOSX.
--   On Linux\*, BSD flavors and other Unix-like open-source operating systems, OpenCV looks for
-    codecs supplied with an OS image. Install the relevant packages (do not forget the development
-    files, for example, "libjpeg-dev", in Debian\* and Ubuntu\*) to get the codec support or turn
-    on the OPENCV_BUILD_3RDPARTY_LIBS flag in CMake.
--   In the case you set *WITH_GDAL* flag to true in CMake and @ref IMREAD_LOAD_GDAL to load the image,
-    then the [GDAL](http://www.gdal.org) driver will be used in order to decode the image, supporting
-    the following formats: [Raster](http://www.gdal.org/formats_list.html),
-    [Vector](http://www.gdal.org/ogr_formats.html).
--   If EXIF information is embedded in the image file, the EXIF orientation will be taken into account
-    and thus the image will be rotated accordingly except if the flags @ref IMREAD_IGNORE_ORIENTATION
-    or @ref IMREAD_UNCHANGED are passed.
--   Use the IMREAD_UNCHANGED flag to keep the floating point values from PFM image.
--   By default number of pixels must be less than 2^30. Limit can be set using system
-    variable OPENCV_IO_MAX_IMAGE_PIXELS
-
-@param filename Name of file to be loaded.
-@param flags Flag that can take values of cv::ImreadModes
-*/
-CV_EXPORTS_W Mat imread( const String& filename, int flags = IMREAD_COLOR );
-
-/** @brief Loads a multi-page image from a file.
-
-The function imreadmulti loads a multi-page image from the specified file into a vector of Mat objects.
-@param filename Name of file to be loaded.
-@param flags Flag that can take values of cv::ImreadModes, default with cv::IMREAD_ANYCOLOR.
-@param mats A vector of Mat objects holding each page, if more than one.
-@sa cv::imread
-*/
-CV_EXPORTS_W bool imreadmulti(const String& filename, CV_OUT std::vector<Mat>& mats, int flags = IMREAD_ANYCOLOR);
-
-/** @brief Loads a of images of a multi-page image from a file.
-
-The function imreadmulti loads a specified range from a multi-page image from the specified file into a vector of Mat objects.
-@param filename Name of file to be loaded.
-@param start Start index of the image to load
-@param count Count number of images to load
-@param flags Flag that can take values of cv::ImreadModes, default with cv::IMREAD_ANYCOLOR.
-@param mats A vector of Mat objects holding each page, if more than one.
-@sa cv::imread
-*/
-CV_EXPORTS_W bool imreadmulti(const String& filename, CV_OUT std::vector<Mat>& mats, int start, int count, int flags = IMREAD_ANYCOLOR);
-
-/** @brief Returns the number of images inside the give file
-
-The function imcount will return the number of pages in a multi-page image, or 1 for single-page images
-@param filename Name of file to be loaded.
-@param flags Flag that can take values of cv::ImreadModes, default with cv::IMREAD_ANYCOLOR.
-*/
-CV_EXPORTS_W size_t imcount(const String& filename, int flags = IMREAD_ANYCOLOR);
-
-/** @brief Saves an image to a specified file.
-
-The function imwrite saves the image to the specified file. The image format is chosen based on the
-filename extension (see cv::imread for the list of extensions). In general, only 8-bit
-single-channel or 3-channel (with 'BGR' channel order) images
-can be saved using this function, with these exceptions:
-
-- 16-bit unsigned (CV_16U) images can be saved in the case of PNG, JPEG 2000, and TIFF formats
-- 32-bit float (CV_32F) images can be saved in PFM, TIFF, OpenEXR, and Radiance HDR formats;
-  3-channel (CV_32FC3) TIFF images will be saved using the LogLuv high dynamic range encoding
-  (4 bytes per pixel)
-- PNG images with an alpha channel can be saved using this function. To do this, create
-8-bit (or 16-bit) 4-channel image BGRA, where the alpha channel goes last. Fully transparent pixels
-should have alpha set to 0, fully opaque pixels should have alpha set to 255/65535 (see the code sample below).
-- Multiple images (vector of Mat) can be saved in TIFF format (see the code sample below).
-
-If the image format is not supported, the image will be converted to 8-bit unsigned (CV_8U) and saved that way.
-
-If the format, depth or channel order is different, use
-Mat::convertTo and cv::cvtColor to convert it before saving. Or, use the universal FileStorage I/O
-functions to save the image to XML or YAML format.
-
-The sample below shows how to create a BGRA image, how to set custom compression parameters and save it to a PNG file.
-It also demonstrates how to save multiple images in a TIFF file:
-@include snippets/imgcodecs_imwrite.cpp
-@param filename Name of the file.
-@param img (Mat or vector of Mat) Image or Images to be saved.
-@param params Format-specific parameters encoded as pairs (paramId_1, paramValue_1, paramId_2, paramValue_2, ... .) see cv::ImwriteFlags
-*/
-CV_EXPORTS_W bool imwrite( const String& filename, InputArray img,
-              const std::vector<int>& params = std::vector<int>());
-
-/// @overload multi-image overload for bindings
-CV_WRAP static inline
-bool imwritemulti(const String& filename, InputArrayOfArrays img,
-                  const std::vector<int>& params = std::vector<int>())
-{
-    return imwrite(filename, img, params);
-}
-
-/** @brief Reads an image from a buffer in memory.
-
-The function imdecode reads an image from the specified buffer in the memory. If the buffer is too short or
-contains invalid data, the function returns an empty matrix ( Mat::data==NULL ).
-
-See cv::imread for the list of supported formats and flags description.
-
-@note In the case of color images, the decoded images will have the channels stored in **B G R** order.
-@param buf Input array or vector of bytes.
-@param flags The same flags as in cv::imread, see cv::ImreadModes.
-*/
-CV_EXPORTS_W Mat imdecode( InputArray buf, int flags );
-
-/** @overload
-@param buf
-@param flags
-@param dst The optional output placeholder for the decoded matrix. It can save the image
-reallocations when the function is called repeatedly for images of the same size.
-*/
-CV_EXPORTS Mat imdecode( InputArray buf, int flags, Mat* dst);
-
-/** @brief Encodes an image into a memory buffer.
-
-The function imencode compresses the image and stores it in the memory buffer that is resized to fit the
-result. See cv::imwrite for the list of supported formats and flags description.
-
-@param ext File extension that defines the output format.
-@param img Image to be written.
-@param buf Output buffer resized to fit the compressed image.
-@param params Format-specific parameters. See cv::imwrite and cv::ImwriteFlags.
-*/
-CV_EXPORTS_W bool imencode( const String& ext, InputArray img,
-                            CV_OUT std::vector<uchar>& buf,
-                            const std::vector<int>& params = std::vector<int>());
-
-/** @brief Returns true if the specified image can be decoded by OpenCV
-
-@param filename File name of the image
-*/
-CV_EXPORTS_W bool haveImageReader( const String& filename );
-
-/** @brief Returns true if an image with the specified filename can be encoded by OpenCV
-
- @param filename File name of the image
- */
-CV_EXPORTS_W bool haveImageWriter( const String& filename );
-
-
-//! @} imgcodecs
-
-} // cv
-
-#endif //OPENCV_IMGCODECS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/imgcodecs/imgcodecs.hpp b/projects/llm_framework/include/opencv4/opencv2/imgcodecs/imgcodecs.hpp
deleted file mode 100644
index a3cd2326..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/imgcodecs/imgcodecs.hpp
+++ /dev/null
@@ -1,48 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifdef __OPENCV_BUILD
-#error this is a compatibility header which should not be used inside the OpenCV library
-#endif
-
-#include "opencv2/imgcodecs.hpp"
diff --git a/projects/llm_framework/include/opencv4/opencv2/imgcodecs/imgcodecs_c.h b/projects/llm_framework/include/opencv4/opencv2/imgcodecs/imgcodecs_c.h
deleted file mode 100644
index c78b3f72..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/imgcodecs/imgcodecs_c.h
+++ /dev/null
@@ -1 +0,0 @@
-#error "This header with legacy C API declarations has been removed from OpenCV. Legacy constants are available from legacy/constants_c.h file."
diff --git a/projects/llm_framework/include/opencv4/opencv2/imgcodecs/ios.h b/projects/llm_framework/include/opencv4/opencv2/imgcodecs/ios.h
deleted file mode 100644
index 5f172181..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/imgcodecs/ios.h
+++ /dev/null
@@ -1,59 +0,0 @@
-
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#import <UIKit/UIKit.h>
-#import <Accelerate/Accelerate.h>
-#import <AVFoundation/AVFoundation.h>
-#import <ImageIO/ImageIO.h>
-#include "opencv2/core.hpp"
-
-//! @addtogroup imgcodecs_ios
-//! @{
-
-CV_EXPORTS CGImageRef MatToCGImage(const cv::Mat& image) CF_RETURNS_RETAINED;
-CV_EXPORTS void CGImageToMat(const CGImageRef image, cv::Mat& m, bool alphaExist = false);
-CV_EXPORTS UIImage* MatToUIImage(const cv::Mat& image);
-CV_EXPORTS void UIImageToMat(const UIImage* image,
-                             cv::Mat& m, bool alphaExist = false);
-
-//! @}
diff --git a/projects/llm_framework/include/opencv4/opencv2/imgcodecs/legacy/constants_c.h b/projects/llm_framework/include/opencv4/opencv2/imgcodecs/legacy/constants_c.h
deleted file mode 100644
index de7be4f7..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/imgcodecs/legacy/constants_c.h
+++ /dev/null
@@ -1,54 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_IMGCODECS_LEGACY_CONSTANTS_H
-#define OPENCV_IMGCODECS_LEGACY_CONSTANTS_H
-
-/* duplicate of "ImreadModes" enumeration for better compatibility with OpenCV 3.x */
-enum
-{
-/* 8bit, color or not */
-    CV_LOAD_IMAGE_UNCHANGED  =-1,
-/* 8bit, gray */
-    CV_LOAD_IMAGE_GRAYSCALE  =0,
-/* ?, color */
-    CV_LOAD_IMAGE_COLOR      =1,
-/* any depth, ? */
-    CV_LOAD_IMAGE_ANYDEPTH   =2,
-/* ?, any color */
-    CV_LOAD_IMAGE_ANYCOLOR   =4,
-/* ?, no rotate */
-    CV_LOAD_IMAGE_IGNORE_ORIENTATION  =128
-};
-
-/* duplicate of "ImwriteFlags" enumeration for better compatibility with OpenCV 3.x */
-enum
-{
-    CV_IMWRITE_JPEG_QUALITY =1,
-    CV_IMWRITE_JPEG_PROGRESSIVE =2,
-    CV_IMWRITE_JPEG_OPTIMIZE =3,
-    CV_IMWRITE_JPEG_RST_INTERVAL =4,
-    CV_IMWRITE_JPEG_LUMA_QUALITY =5,
-    CV_IMWRITE_JPEG_CHROMA_QUALITY =6,
-    CV_IMWRITE_PNG_COMPRESSION =16,
-    CV_IMWRITE_PNG_STRATEGY =17,
-    CV_IMWRITE_PNG_BILEVEL =18,
-    CV_IMWRITE_PNG_STRATEGY_DEFAULT =0,
-    CV_IMWRITE_PNG_STRATEGY_FILTERED =1,
-    CV_IMWRITE_PNG_STRATEGY_HUFFMAN_ONLY =2,
-    CV_IMWRITE_PNG_STRATEGY_RLE =3,
-    CV_IMWRITE_PNG_STRATEGY_FIXED =4,
-    CV_IMWRITE_PXM_BINARY =32,
-    CV_IMWRITE_EXR_TYPE = 48,
-    CV_IMWRITE_WEBP_QUALITY =64,
-    CV_IMWRITE_PAM_TUPLETYPE = 128,
-    CV_IMWRITE_PAM_FORMAT_NULL = 0,
-    CV_IMWRITE_PAM_FORMAT_BLACKANDWHITE = 1,
-    CV_IMWRITE_PAM_FORMAT_GRAYSCALE = 2,
-    CV_IMWRITE_PAM_FORMAT_GRAYSCALE_ALPHA = 3,
-    CV_IMWRITE_PAM_FORMAT_RGB = 4,
-    CV_IMWRITE_PAM_FORMAT_RGB_ALPHA = 5,
-};
-
-#endif // OPENCV_IMGCODECS_LEGACY_CONSTANTS_H
diff --git a/projects/llm_framework/include/opencv4/opencv2/imgcodecs/macosx.h b/projects/llm_framework/include/opencv4/opencv2/imgcodecs/macosx.h
deleted file mode 100644
index cfb07707..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/imgcodecs/macosx.h
+++ /dev/null
@@ -1,20 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#if !defined(__APPLE__) || !defined(__MACH__)
-#error This header should be used in macOS ObjC/Swift projects.
-#endif
-
-#import <AppKit/AppKit.h>
-#include "opencv2/core.hpp"
-
-//! @addtogroup imgcodecs_macosx
-//! @{
-
-CV_EXPORTS CGImageRef MatToCGImage(const cv::Mat& image) CF_RETURNS_RETAINED;
-CV_EXPORTS void CGImageToMat(const CGImageRef image, cv::Mat& m, bool alphaExist = false);
-CV_EXPORTS NSImage* MatToNSImage(const cv::Mat& image);
-CV_EXPORTS void NSImageToMat(const NSImage* image, cv::Mat& m, bool alphaExist = false);
-
-//! @}
diff --git a/projects/llm_framework/include/opencv4/opencv2/imgproc.hpp b/projects/llm_framework/include/opencv4/opencv2/imgproc.hpp
deleted file mode 100644
index eacec611..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/imgproc.hpp
+++ /dev/null
@@ -1,4994 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_IMGPROC_HPP
-#define OPENCV_IMGPROC_HPP
-
-#include "opencv2/core.hpp"
-
-/**
-  @defgroup imgproc Image Processing
-
-This module includes image-processing functions.
-
-  @{
-    @defgroup imgproc_filter Image Filtering
-
-Functions and classes described in this section are used to perform various linear or non-linear
-filtering operations on 2D images (represented as Mat's). It means that for each pixel location
-\f$(x,y)\f$ in the source image (normally, rectangular), its neighborhood is considered and used to
-compute the response. In case of a linear filter, it is a weighted sum of pixel values. In case of
-morphological operations, it is the minimum or maximum values, and so on. The computed response is
-stored in the destination image at the same location \f$(x,y)\f$. It means that the output image
-will be of the same size as the input image. Normally, the functions support multi-channel arrays,
-in which case every channel is processed independently. Therefore, the output image will also have
-the same number of channels as the input one.
-
-Another common feature of the functions and classes described in this section is that, unlike
-simple arithmetic functions, they need to extrapolate values of some non-existing pixels. For
-example, if you want to smooth an image using a Gaussian \f$3 \times 3\f$ filter, then, when
-processing the left-most pixels in each row, you need pixels to the left of them, that is, outside
-of the image. You can let these pixels be the same as the left-most image pixels ("replicated
-border" extrapolation method), or assume that all the non-existing pixels are zeros ("constant
-border" extrapolation method), and so on. OpenCV enables you to specify the extrapolation method.
-For details, see #BorderTypes
-
-@anchor filter_depths
-### Depth combinations
-Input depth (src.depth()) | Output depth (ddepth)
---------------------------|----------------------
-CV_8U                     | -1/CV_16S/CV_32F/CV_64F
-CV_16U/CV_16S             | -1/CV_32F/CV_64F
-CV_32F                    | -1/CV_32F/CV_64F
-CV_64F                    | -1/CV_64F
-
-@note when ddepth=-1, the output image will have the same depth as the source.
-
-    @defgroup imgproc_transform Geometric Image Transformations
-
-The functions in this section perform various geometrical transformations of 2D images. They do not
-change the image content but deform the pixel grid and map this deformed grid to the destination
-image. In fact, to avoid sampling artifacts, the mapping is done in the reverse order, from
-destination to the source. That is, for each pixel \f$(x, y)\f$ of the destination image, the
-functions compute coordinates of the corresponding "donor" pixel in the source image and copy the
-pixel value:
-
-\f[\texttt{dst} (x,y)= \texttt{src} (f_x(x,y), f_y(x,y))\f]
-
-In case when you specify the forward mapping \f$\left<g_x, g_y\right>: \texttt{src} \rightarrow
-\texttt{dst}\f$, the OpenCV functions first compute the corresponding inverse mapping
-\f$\left<f_x, f_y\right>: \texttt{dst} \rightarrow \texttt{src}\f$ and then use the above formula.
-
-The actual implementations of the geometrical transformations, from the most generic remap and to
-the simplest and the fastest resize, need to solve two main problems with the above formula:
-
-- Extrapolation of non-existing pixels. Similarly to the filtering functions described in the
-previous section, for some \f$(x,y)\f$, either one of \f$f_x(x,y)\f$, or \f$f_y(x,y)\f$, or both
-of them may fall outside of the image. In this case, an extrapolation method needs to be used.
-OpenCV provides the same selection of extrapolation methods as in the filtering functions. In
-addition, it provides the method #BORDER_TRANSPARENT. This means that the corresponding pixels in
-the destination image will not be modified at all.
-
-- Interpolation of pixel values. Usually \f$f_x(x,y)\f$ and \f$f_y(x,y)\f$ are floating-point
-numbers. This means that \f$\left<f_x, f_y\right>\f$ can be either an affine or perspective
-transformation, or radial lens distortion correction, and so on. So, a pixel value at fractional
-coordinates needs to be retrieved. In the simplest case, the coordinates can be just rounded to the
-nearest integer coordinates and the corresponding pixel can be used. This is called a
-nearest-neighbor interpolation. However, a better result can be achieved by using more
-sophisticated [interpolation methods](http://en.wikipedia.org/wiki/Multivariate_interpolation) ,
-where a polynomial function is fit into some neighborhood of the computed pixel \f$(f_x(x,y),
-f_y(x,y))\f$, and then the value of the polynomial at \f$(f_x(x,y), f_y(x,y))\f$ is taken as the
-interpolated pixel value. In OpenCV, you can choose between several interpolation methods. See
-resize for details.
-
-@note The geometrical transformations do not work with `CV_8S` or `CV_32S` images.
-
-    @defgroup imgproc_misc Miscellaneous Image Transformations
-    @defgroup imgproc_draw Drawing Functions
-
-Drawing functions work with matrices/images of arbitrary depth. The boundaries of the shapes can be
-rendered with antialiasing (implemented only for 8-bit images for now). All the functions include
-the parameter color that uses an RGB value (that may be constructed with the Scalar constructor )
-for color images and brightness for grayscale images. For color images, the channel ordering is
-normally *Blue, Green, Red*. This is what imshow, imread, and imwrite expect. So, if you form a
-color using the Scalar constructor, it should look like:
-
-\f[\texttt{Scalar} (blue \_ component, green \_ component, red \_ component[, alpha \_ component])\f]
-
-If you are using your own image rendering and I/O functions, you can use any channel ordering. The
-drawing functions process each channel independently and do not depend on the channel order or even
-on the used color space. The whole image can be converted from BGR to RGB or to a different color
-space using cvtColor .
-
-If a drawn figure is partially or completely outside the image, the drawing functions clip it. Also,
-many drawing functions can handle pixel coordinates specified with sub-pixel accuracy. This means
-that the coordinates can be passed as fixed-point numbers encoded as integers. The number of
-fractional bits is specified by the shift parameter and the real point coordinates are calculated as
-\f$\texttt{Point}(x,y)\rightarrow\texttt{Point2f}(x*2^{-shift},y*2^{-shift})\f$ . This feature is
-especially effective when rendering antialiased shapes.
-
-@note The functions do not support alpha-transparency when the target image is 4-channel. In this
-case, the color[3] is simply copied to the repainted pixels. Thus, if you want to paint
-semi-transparent shapes, you can paint them in a separate buffer and then blend it with the main
-image.
-
-    @defgroup imgproc_color_conversions Color Space Conversions
-    @defgroup imgproc_colormap ColorMaps in OpenCV
-
-The human perception isn't built for observing fine changes in grayscale images. Human eyes are more
-sensitive to observing changes between colors, so you often need to recolor your grayscale images to
-get a clue about them. OpenCV now comes with various colormaps to enhance the visualization in your
-computer vision application.
-
-In OpenCV you only need applyColorMap to apply a colormap on a given image. The following sample
-code reads the path to an image from command line, applies a Jet colormap on it and shows the
-result:
-
-@include snippets/imgproc_applyColorMap.cpp
-
-@see #ColormapTypes
-
-    @defgroup imgproc_subdiv2d Planar Subdivision
-
-The Subdiv2D class described in this section is used to perform various planar subdivision on
-a set of 2D points (represented as vector of Point2f). OpenCV subdivides a plane into triangles
-using the Delaunay's algorithm, which corresponds to the dual graph of the Voronoi diagram.
-In the figure below, the Delaunay's triangulation is marked with black lines and the Voronoi
-diagram with red lines.
-
-![Delaunay triangulation (black) and Voronoi (red)](pics/delaunay_voronoi.png)
-
-The subdivisions can be used for the 3D piece-wise transformation of a plane, morphing, fast
-location of points on the plane, building special graphs (such as NNG,RNG), and so forth.
-
-    @defgroup imgproc_hist Histograms
-    @defgroup imgproc_shape Structural Analysis and Shape Descriptors
-    @defgroup imgproc_motion Motion Analysis and Object Tracking
-    @defgroup imgproc_feature Feature Detection
-    @defgroup imgproc_object Object Detection
-    @defgroup imgproc_segmentation Image Segmentation
-    @defgroup imgproc_c C API
-    @defgroup imgproc_hal Hardware Acceleration Layer
-    @{
-        @defgroup imgproc_hal_functions Functions
-        @defgroup imgproc_hal_interface Interface
-    @}
-  @}
-*/
-
-namespace cv
-{
-
-/** @addtogroup imgproc
-@{
-*/
-
-//! @addtogroup imgproc_filter
-//! @{
-
-enum SpecialFilter {
-    FILTER_SCHARR = -1
-};
-
-//! type of morphological operation
-enum MorphTypes{
-    MORPH_ERODE    = 0, //!< see #erode
-    MORPH_DILATE   = 1, //!< see #dilate
-    MORPH_OPEN     = 2, //!< an opening operation
-                        //!< \f[\texttt{dst} = \mathrm{open} ( \texttt{src} , \texttt{element} )= \mathrm{dilate} ( \mathrm{erode} ( \texttt{src} , \texttt{element} ))\f]
-    MORPH_CLOSE    = 3, //!< a closing operation
-                        //!< \f[\texttt{dst} = \mathrm{close} ( \texttt{src} , \texttt{element} )= \mathrm{erode} ( \mathrm{dilate} ( \texttt{src} , \texttt{element} ))\f]
-    MORPH_GRADIENT = 4, //!< a morphological gradient
-                        //!< \f[\texttt{dst} = \mathrm{morph\_grad} ( \texttt{src} , \texttt{element} )= \mathrm{dilate} ( \texttt{src} , \texttt{element} )- \mathrm{erode} ( \texttt{src} , \texttt{element} )\f]
-    MORPH_TOPHAT   = 5, //!< "top hat"
-                        //!< \f[\texttt{dst} = \mathrm{tophat} ( \texttt{src} , \texttt{element} )= \texttt{src} - \mathrm{open} ( \texttt{src} , \texttt{element} )\f]
-    MORPH_BLACKHAT = 6, //!< "black hat"
-                        //!< \f[\texttt{dst} = \mathrm{blackhat} ( \texttt{src} , \texttt{element} )= \mathrm{close} ( \texttt{src} , \texttt{element} )- \texttt{src}\f]
-    MORPH_HITMISS  = 7  //!< "hit or miss"
-                        //!<   .- Only supported for CV_8UC1 binary images. A tutorial can be found in the documentation
-};
-
-//! shape of the structuring element
-enum MorphShapes {
-    MORPH_RECT    = 0, //!< a rectangular structuring element:  \f[E_{ij}=1\f]
-    MORPH_CROSS   = 1, //!< a cross-shaped structuring element:
-                       //!< \f[E_{ij} = \begin{cases} 1 & \texttt{if } {i=\texttt{anchor.y } {or } {j=\texttt{anchor.x}}} \\0 & \texttt{otherwise} \end{cases}\f]
-    MORPH_ELLIPSE = 2 //!< an elliptic structuring element, that is, a filled ellipse inscribed
-                      //!< into the rectangle Rect(0, 0, esize.width, 0.esize.height)
-};
-
-//! @} imgproc_filter
-
-//! @addtogroup imgproc_transform
-//! @{
-
-//! interpolation algorithm
-enum InterpolationFlags{
-    /** nearest neighbor interpolation */
-    INTER_NEAREST        = 0,
-    /** bilinear interpolation */
-    INTER_LINEAR         = 1,
-    /** bicubic interpolation */
-    INTER_CUBIC          = 2,
-    /** resampling using pixel area relation. It may be a preferred method for image decimation, as
-    it gives moire'-free results. But when the image is zoomed, it is similar to the INTER_NEAREST
-    method. */
-    INTER_AREA           = 3,
-    /** Lanczos interpolation over 8x8 neighborhood */
-    INTER_LANCZOS4       = 4,
-    /** Bit exact bilinear interpolation */
-    INTER_LINEAR_EXACT = 5,
-    /** Bit exact nearest neighbor interpolation. This will produce same results as
-    the nearest neighbor method in PIL, scikit-image or Matlab. */
-    INTER_NEAREST_EXACT  = 6,
-    /** mask for interpolation codes */
-    INTER_MAX            = 7,
-    /** flag, fills all of the destination image pixels. If some of them correspond to outliers in the
-    source image, they are set to zero */
-    WARP_FILL_OUTLIERS   = 8,
-    /** flag, inverse transformation
-
-    For example, #linearPolar or #logPolar transforms:
-    - flag is __not__ set: \f$dst( \rho , \phi ) = src(x,y)\f$
-    - flag is set: \f$dst(x,y) = src( \rho , \phi )\f$
-    */
-    WARP_INVERSE_MAP     = 16
-};
-
-/** \brief Specify the polar mapping mode
-@sa warpPolar
-*/
-enum WarpPolarMode
-{
-    WARP_POLAR_LINEAR = 0, ///< Remaps an image to/from polar space.
-    WARP_POLAR_LOG = 256   ///< Remaps an image to/from semilog-polar space.
-};
-
-enum InterpolationMasks {
-       INTER_BITS      = 5,
-       INTER_BITS2     = INTER_BITS * 2,
-       INTER_TAB_SIZE  = 1 << INTER_BITS,
-       INTER_TAB_SIZE2 = INTER_TAB_SIZE * INTER_TAB_SIZE
-     };
-
-//! @} imgproc_transform
-
-//! @addtogroup imgproc_misc
-//! @{
-
-//! Distance types for Distance Transform and M-estimators
-//! @see distanceTransform, fitLine
-enum DistanceTypes {
-    DIST_USER    = -1,  //!< User defined distance
-    DIST_L1      = 1,   //!< distance = |x1-x2| + |y1-y2|
-    DIST_L2      = 2,   //!< the simple euclidean distance
-    DIST_C       = 3,   //!< distance = max(|x1-x2|,|y1-y2|)
-    DIST_L12     = 4,   //!< L1-L2 metric: distance = 2(sqrt(1+x*x/2) - 1))
-    DIST_FAIR    = 5,   //!< distance = c^2(|x|/c-log(1+|x|/c)), c = 1.3998
-    DIST_WELSCH  = 6,   //!< distance = c^2/2(1-exp(-(x/c)^2)), c = 2.9846
-    DIST_HUBER   = 7    //!< distance = |x|<c ? x^2/2 : c(|x|-c/2), c=1.345
-};
-
-//! Mask size for distance transform
-enum DistanceTransformMasks {
-    DIST_MASK_3       = 3, //!< mask=3
-    DIST_MASK_5       = 5, //!< mask=5
-    DIST_MASK_PRECISE = 0  //!<
-};
-
-//! type of the threshold operation
-//! ![threshold types](pics/threshold.png)
-enum ThresholdTypes {
-    THRESH_BINARY     = 0, //!< \f[\texttt{dst} (x,y) =  \fork{\texttt{maxval}}{if \(\texttt{src}(x,y) > \texttt{thresh}\)}{0}{otherwise}\f]
-    THRESH_BINARY_INV = 1, //!< \f[\texttt{dst} (x,y) =  \fork{0}{if \(\texttt{src}(x,y) > \texttt{thresh}\)}{\texttt{maxval}}{otherwise}\f]
-    THRESH_TRUNC      = 2, //!< \f[\texttt{dst} (x,y) =  \fork{\texttt{threshold}}{if \(\texttt{src}(x,y) > \texttt{thresh}\)}{\texttt{src}(x,y)}{otherwise}\f]
-    THRESH_TOZERO     = 3, //!< \f[\texttt{dst} (x,y) =  \fork{\texttt{src}(x,y)}{if \(\texttt{src}(x,y) > \texttt{thresh}\)}{0}{otherwise}\f]
-    THRESH_TOZERO_INV = 4, //!< \f[\texttt{dst} (x,y) =  \fork{0}{if \(\texttt{src}(x,y) > \texttt{thresh}\)}{\texttt{src}(x,y)}{otherwise}\f]
-    THRESH_MASK       = 7,
-    THRESH_OTSU       = 8, //!< flag, use Otsu algorithm to choose the optimal threshold value
-    THRESH_TRIANGLE   = 16 //!< flag, use Triangle algorithm to choose the optimal threshold value
-};
-
-//! adaptive threshold algorithm
-//! @see adaptiveThreshold
-enum AdaptiveThresholdTypes {
-    /** the threshold value \f$T(x,y)\f$ is a mean of the \f$\texttt{blockSize} \times
-    \texttt{blockSize}\f$ neighborhood of \f$(x, y)\f$ minus C */
-    ADAPTIVE_THRESH_MEAN_C     = 0,
-    /** the threshold value \f$T(x, y)\f$ is a weighted sum (cross-correlation with a Gaussian
-    window) of the \f$\texttt{blockSize} \times \texttt{blockSize}\f$ neighborhood of \f$(x, y)\f$
-    minus C . The default sigma (standard deviation) is used for the specified blockSize . See
-    #getGaussianKernel*/
-    ADAPTIVE_THRESH_GAUSSIAN_C = 1
-};
-
-//! class of the pixel in GrabCut algorithm
-enum GrabCutClasses {
-    GC_BGD    = 0,  //!< an obvious background pixels
-    GC_FGD    = 1,  //!< an obvious foreground (object) pixel
-    GC_PR_BGD = 2,  //!< a possible background pixel
-    GC_PR_FGD = 3   //!< a possible foreground pixel
-};
-
-//! GrabCut algorithm flags
-enum GrabCutModes {
-    /** The function initializes the state and the mask using the provided rectangle. After that it
-    runs iterCount iterations of the algorithm. */
-    GC_INIT_WITH_RECT  = 0,
-    /** The function initializes the state using the provided mask. Note that GC_INIT_WITH_RECT
-    and GC_INIT_WITH_MASK can be combined. Then, all the pixels outside of the ROI are
-    automatically initialized with GC_BGD .*/
-    GC_INIT_WITH_MASK  = 1,
-    /** The value means that the algorithm should just resume. */
-    GC_EVAL            = 2,
-    /** The value means that the algorithm should just run the grabCut algorithm (a single iteration) with the fixed model */
-    GC_EVAL_FREEZE_MODEL = 3
-};
-
-//! distanceTransform algorithm flags
-enum DistanceTransformLabelTypes {
-    /** each connected component of zeros in src (as well as all the non-zero pixels closest to the
-    connected component) will be assigned the same label */
-    DIST_LABEL_CCOMP = 0,
-    /** each zero pixel (and all the non-zero pixels closest to it) gets its own label. */
-    DIST_LABEL_PIXEL = 1
-};
-
-//! floodfill algorithm flags
-enum FloodFillFlags {
-    /** If set, the difference between the current pixel and seed pixel is considered. Otherwise,
-    the difference between neighbor pixels is considered (that is, the range is floating). */
-    FLOODFILL_FIXED_RANGE = 1 << 16,
-    /** If set, the function does not change the image ( newVal is ignored), and only fills the
-    mask with the value specified in bits 8-16 of flags as described above. This option only make
-    sense in function variants that have the mask parameter. */
-    FLOODFILL_MASK_ONLY   = 1 << 17
-};
-
-//! @} imgproc_misc
-
-//! @addtogroup imgproc_shape
-//! @{
-
-//! connected components statistics
-enum ConnectedComponentsTypes {
-    CC_STAT_LEFT   = 0, //!< The leftmost (x) coordinate which is the inclusive start of the bounding
-                        //!< box in the horizontal direction.
-    CC_STAT_TOP    = 1, //!< The topmost (y) coordinate which is the inclusive start of the bounding
-                        //!< box in the vertical direction.
-    CC_STAT_WIDTH  = 2, //!< The horizontal size of the bounding box
-    CC_STAT_HEIGHT = 3, //!< The vertical size of the bounding box
-    CC_STAT_AREA   = 4, //!< The total area (in pixels) of the connected component
-#ifndef CV_DOXYGEN
-    CC_STAT_MAX    = 5 //!< Max enumeration value. Used internally only for memory allocation
-#endif
-};
-
-//! connected components algorithm
-enum ConnectedComponentsAlgorithmsTypes {
-    CCL_DEFAULT   = -1, //!< Spaghetti @cite Bolelli2019 algorithm for 8-way connectivity, Spaghetti4C @cite Bolelli2021 algorithm for 4-way connectivity.
-    CCL_WU        = 0,  //!< SAUF @cite Wu2009 algorithm for 8-way connectivity, SAUF algorithm for 4-way connectivity. The parallel implementation described in @cite Bolelli2017 is available for SAUF.
-    CCL_GRANA     = 1,  //!< BBDT @cite Grana2010 algorithm for 8-way connectivity, SAUF algorithm for 4-way connectivity. The parallel implementation described in @cite Bolelli2017 is available for both BBDT and SAUF.
-    CCL_BOLELLI   = 2,  //!< Spaghetti @cite Bolelli2019 algorithm for 8-way connectivity, Spaghetti4C @cite Bolelli2021 algorithm for 4-way connectivity. The parallel implementation described in @cite Bolelli2017 is available for both Spaghetti and Spaghetti4C.
-    CCL_SAUF      = 3,  //!< Same as CCL_WU. It is preferable to use the flag with the name of the algorithm (CCL_SAUF) rather than the one with the name of the first author (CCL_WU).
-    CCL_BBDT      = 4,  //!< Same as CCL_GRANA. It is preferable to use the flag with the name of the algorithm (CCL_BBDT) rather than the one with the name of the first author (CCL_GRANA).
-    CCL_SPAGHETTI = 5,  //!< Same as CCL_BOLELLI. It is preferable to use the flag with the name of the algorithm (CCL_SPAGHETTI) rather than the one with the name of the first author (CCL_BOLELLI).
-};
-
-//! mode of the contour retrieval algorithm
-enum RetrievalModes {
-    /** retrieves only the extreme outer contours. It sets `hierarchy[i][2]=hierarchy[i][3]=-1` for
-    all the contours. */
-    RETR_EXTERNAL  = 0,
-    /** retrieves all of the contours without establishing any hierarchical relationships. */
-    RETR_LIST      = 1,
-    /** retrieves all of the contours and organizes them into a two-level hierarchy. At the top
-    level, there are external boundaries of the components. At the second level, there are
-    boundaries of the holes. If there is another contour inside a hole of a connected component, it
-    is still put at the top level. */
-    RETR_CCOMP     = 2,
-    /** retrieves all of the contours and reconstructs a full hierarchy of nested contours.*/
-    RETR_TREE      = 3,
-    RETR_FLOODFILL = 4 //!<
-};
-
-//! the contour approximation algorithm
-enum ContourApproximationModes {
-    /** stores absolutely all the contour points. That is, any 2 subsequent points (x1,y1) and
-    (x2,y2) of the contour will be either horizontal, vertical or diagonal neighbors, that is,
-    max(abs(x1-x2),abs(y2-y1))==1. */
-    CHAIN_APPROX_NONE      = 1,
-    /** compresses horizontal, vertical, and diagonal segments and leaves only their end points.
-    For example, an up-right rectangular contour is encoded with 4 points. */
-    CHAIN_APPROX_SIMPLE    = 2,
-    /** applies one of the flavors of the Teh-Chin chain approximation algorithm @cite TehChin89 */
-    CHAIN_APPROX_TC89_L1   = 3,
-    /** applies one of the flavors of the Teh-Chin chain approximation algorithm @cite TehChin89 */
-    CHAIN_APPROX_TC89_KCOS = 4
-};
-
-/** @brief Shape matching methods
-
-\f$A\f$ denotes object1,\f$B\f$ denotes object2
-
-\f$\begin{array}{l} m^A_i =  \mathrm{sign} (h^A_i)  \cdot \log{h^A_i} \\ m^B_i =  \mathrm{sign} (h^B_i)  \cdot \log{h^B_i} \end{array}\f$
-
-and \f$h^A_i, h^B_i\f$ are the Hu moments of \f$A\f$ and \f$B\f$ , respectively.
-*/
-enum ShapeMatchModes {
-    CONTOURS_MATCH_I1  =1, //!< \f[I_1(A,B) =  \sum _{i=1...7}  \left |  \frac{1}{m^A_i} -  \frac{1}{m^B_i} \right |\f]
-    CONTOURS_MATCH_I2  =2, //!< \f[I_2(A,B) =  \sum _{i=1...7}  \left | m^A_i - m^B_i  \right |\f]
-    CONTOURS_MATCH_I3  =3  //!< \f[I_3(A,B) =  \max _{i=1...7}  \frac{ \left| m^A_i - m^B_i \right| }{ \left| m^A_i \right| }\f]
-};
-
-//! @} imgproc_shape
-
-//! @addtogroup imgproc_feature
-//! @{
-
-//! Variants of a Hough transform
-enum HoughModes {
-
-    /** classical or standard Hough transform. Every line is represented by two floating-point
-    numbers \f$(\rho, \theta)\f$ , where \f$\rho\f$ is a distance between (0,0) point and the line,
-    and \f$\theta\f$ is the angle between x-axis and the normal to the line. Thus, the matrix must
-    be (the created sequence will be) of CV_32FC2 type */
-    HOUGH_STANDARD      = 0,
-    /** probabilistic Hough transform (more efficient in case if the picture contains a few long
-    linear segments). It returns line segments rather than the whole line. Each segment is
-    represented by starting and ending points, and the matrix must be (the created sequence will
-    be) of the CV_32SC4 type. */
-    HOUGH_PROBABILISTIC = 1,
-    /** multi-scale variant of the classical Hough transform. The lines are encoded the same way as
-    HOUGH_STANDARD. */
-    HOUGH_MULTI_SCALE   = 2,
-    HOUGH_GRADIENT      = 3, //!< basically *21HT*, described in @cite Yuen90
-    HOUGH_GRADIENT_ALT  = 4, //!< variation of HOUGH_GRADIENT to get better accuracy
-};
-
-//! Variants of Line Segment %Detector
-enum LineSegmentDetectorModes {
-    LSD_REFINE_NONE = 0, //!< No refinement applied
-    LSD_REFINE_STD  = 1, //!< Standard refinement is applied. E.g. breaking arches into smaller straighter line approximations.
-    LSD_REFINE_ADV  = 2  //!< Advanced refinement. Number of false alarms is calculated, lines are
-                         //!< refined through increase of precision, decrement in size, etc.
-};
-
-//! @} imgproc_feature
-
-/** Histogram comparison methods
-  @ingroup imgproc_hist
-*/
-enum HistCompMethods {
-    /** Correlation
-    \f[d(H_1,H_2) =  \frac{\sum_I (H_1(I) - \bar{H_1}) (H_2(I) - \bar{H_2})}{\sqrt{\sum_I(H_1(I) - \bar{H_1})^2 \sum_I(H_2(I) - \bar{H_2})^2}}\f]
-    where
-    \f[\bar{H_k} =  \frac{1}{N} \sum _J H_k(J)\f]
-    and \f$N\f$ is a total number of histogram bins. */
-    HISTCMP_CORREL        = 0,
-    /** Chi-Square
-    \f[d(H_1,H_2) =  \sum _I  \frac{\left(H_1(I)-H_2(I)\right)^2}{H_1(I)}\f] */
-    HISTCMP_CHISQR        = 1,
-    /** Intersection
-    \f[d(H_1,H_2) =  \sum _I  \min (H_1(I), H_2(I))\f] */
-    HISTCMP_INTERSECT     = 2,
-    /** Bhattacharyya distance
-    (In fact, OpenCV computes Hellinger distance, which is related to Bhattacharyya coefficient.)
-    \f[d(H_1,H_2) =  \sqrt{1 - \frac{1}{\sqrt{\bar{H_1} \bar{H_2} N^2}} \sum_I \sqrt{H_1(I) \cdot H_2(I)}}\f] */
-    HISTCMP_BHATTACHARYYA = 3,
-    HISTCMP_HELLINGER     = HISTCMP_BHATTACHARYYA, //!< Synonym for HISTCMP_BHATTACHARYYA
-    /** Alternative Chi-Square
-    \f[d(H_1,H_2) =  2 * \sum _I  \frac{\left(H_1(I)-H_2(I)\right)^2}{H_1(I)+H_2(I)}\f]
-    This alternative formula is regularly used for texture comparison. See e.g. @cite Puzicha1997 */
-    HISTCMP_CHISQR_ALT    = 4,
-    /** Kullback-Leibler divergence
-    \f[d(H_1,H_2) = \sum _I H_1(I) \log \left(\frac{H_1(I)}{H_2(I)}\right)\f] */
-    HISTCMP_KL_DIV        = 5
-};
-
-/** the color conversion codes
-@see @ref imgproc_color_conversions
-@ingroup imgproc_color_conversions
- */
-enum ColorConversionCodes {
-    COLOR_BGR2BGRA     = 0, //!< add alpha channel to RGB or BGR image
-    COLOR_RGB2RGBA     = COLOR_BGR2BGRA,
-
-    COLOR_BGRA2BGR     = 1, //!< remove alpha channel from RGB or BGR image
-    COLOR_RGBA2RGB     = COLOR_BGRA2BGR,
-
-    COLOR_BGR2RGBA     = 2, //!< convert between RGB and BGR color spaces (with or without alpha channel)
-    COLOR_RGB2BGRA     = COLOR_BGR2RGBA,
-
-    COLOR_RGBA2BGR     = 3,
-    COLOR_BGRA2RGB     = COLOR_RGBA2BGR,
-
-    COLOR_BGR2RGB      = 4,
-    COLOR_RGB2BGR      = COLOR_BGR2RGB,
-
-    COLOR_BGRA2RGBA    = 5,
-    COLOR_RGBA2BGRA    = COLOR_BGRA2RGBA,
-
-    COLOR_BGR2GRAY     = 6, //!< convert between RGB/BGR and grayscale, @ref color_convert_rgb_gray "color conversions"
-    COLOR_RGB2GRAY     = 7,
-    COLOR_GRAY2BGR     = 8,
-    COLOR_GRAY2RGB     = COLOR_GRAY2BGR,
-    COLOR_GRAY2BGRA    = 9,
-    COLOR_GRAY2RGBA    = COLOR_GRAY2BGRA,
-    COLOR_BGRA2GRAY    = 10,
-    COLOR_RGBA2GRAY    = 11,
-
-    COLOR_BGR2BGR565   = 12, //!< convert between RGB/BGR and BGR565 (16-bit images)
-    COLOR_RGB2BGR565   = 13,
-    COLOR_BGR5652BGR   = 14,
-    COLOR_BGR5652RGB   = 15,
-    COLOR_BGRA2BGR565  = 16,
-    COLOR_RGBA2BGR565  = 17,
-    COLOR_BGR5652BGRA  = 18,
-    COLOR_BGR5652RGBA  = 19,
-
-    COLOR_GRAY2BGR565  = 20, //!< convert between grayscale to BGR565 (16-bit images)
-    COLOR_BGR5652GRAY  = 21,
-
-    COLOR_BGR2BGR555   = 22,  //!< convert between RGB/BGR and BGR555 (16-bit images)
-    COLOR_RGB2BGR555   = 23,
-    COLOR_BGR5552BGR   = 24,
-    COLOR_BGR5552RGB   = 25,
-    COLOR_BGRA2BGR555  = 26,
-    COLOR_RGBA2BGR555  = 27,
-    COLOR_BGR5552BGRA  = 28,
-    COLOR_BGR5552RGBA  = 29,
-
-    COLOR_GRAY2BGR555  = 30, //!< convert between grayscale and BGR555 (16-bit images)
-    COLOR_BGR5552GRAY  = 31,
-
-    COLOR_BGR2XYZ      = 32, //!< convert RGB/BGR to CIE XYZ, @ref color_convert_rgb_xyz "color conversions"
-    COLOR_RGB2XYZ      = 33,
-    COLOR_XYZ2BGR      = 34,
-    COLOR_XYZ2RGB      = 35,
-
-    COLOR_BGR2YCrCb    = 36, //!< convert RGB/BGR to luma-chroma (aka YCC), @ref color_convert_rgb_ycrcb "color conversions"
-    COLOR_RGB2YCrCb    = 37,
-    COLOR_YCrCb2BGR    = 38,
-    COLOR_YCrCb2RGB    = 39,
-
-    COLOR_BGR2HSV      = 40, //!< convert RGB/BGR to HSV (hue saturation value) with H range 0..180 if 8 bit image, @ref color_convert_rgb_hsv "color conversions"
-    COLOR_RGB2HSV      = 41,
-
-    COLOR_BGR2Lab      = 44, //!< convert RGB/BGR to CIE Lab, @ref color_convert_rgb_lab "color conversions"
-    COLOR_RGB2Lab      = 45,
-
-    COLOR_BGR2Luv      = 50, //!< convert RGB/BGR to CIE Luv, @ref color_convert_rgb_luv "color conversions"
-    COLOR_RGB2Luv      = 51,
-    COLOR_BGR2HLS      = 52, //!< convert RGB/BGR to HLS (hue lightness saturation) with H range 0..180 if 8 bit image, @ref color_convert_rgb_hls "color conversions"
-    COLOR_RGB2HLS      = 53,
-
-    COLOR_HSV2BGR      = 54, //!< backward conversions HSV to RGB/BGR with H range 0..180 if 8 bit image
-    COLOR_HSV2RGB      = 55,
-
-    COLOR_Lab2BGR      = 56,
-    COLOR_Lab2RGB      = 57,
-    COLOR_Luv2BGR      = 58,
-    COLOR_Luv2RGB      = 59,
-    COLOR_HLS2BGR      = 60, //!< backward conversions HLS to RGB/BGR with H range 0..180 if 8 bit image
-    COLOR_HLS2RGB      = 61,
-
-    COLOR_BGR2HSV_FULL = 66, //!< convert RGB/BGR to HSV (hue saturation value) with H range 0..255 if 8 bit image, @ref color_convert_rgb_hsv "color conversions"
-    COLOR_RGB2HSV_FULL = 67,
-    COLOR_BGR2HLS_FULL = 68, //!< convert RGB/BGR to HLS (hue lightness saturation) with H range 0..255 if 8 bit image, @ref color_convert_rgb_hls "color conversions"
-    COLOR_RGB2HLS_FULL = 69,
-
-    COLOR_HSV2BGR_FULL = 70, //!< backward conversions HSV to RGB/BGR with H range 0..255 if 8 bit image
-    COLOR_HSV2RGB_FULL = 71,
-    COLOR_HLS2BGR_FULL = 72, //!< backward conversions HLS to RGB/BGR with H range 0..255 if 8 bit image
-    COLOR_HLS2RGB_FULL = 73,
-
-    COLOR_LBGR2Lab     = 74,
-    COLOR_LRGB2Lab     = 75,
-    COLOR_LBGR2Luv     = 76,
-    COLOR_LRGB2Luv     = 77,
-
-    COLOR_Lab2LBGR     = 78,
-    COLOR_Lab2LRGB     = 79,
-    COLOR_Luv2LBGR     = 80,
-    COLOR_Luv2LRGB     = 81,
-
-    COLOR_BGR2YUV      = 82, //!< convert between RGB/BGR and YUV
-    COLOR_RGB2YUV      = 83,
-    COLOR_YUV2BGR      = 84,
-    COLOR_YUV2RGB      = 85,
-
-    //! YUV 4:2:0 family to RGB
-    COLOR_YUV2RGB_NV12  = 90,
-    COLOR_YUV2BGR_NV12  = 91,
-    COLOR_YUV2RGB_NV21  = 92,
-    COLOR_YUV2BGR_NV21  = 93,
-    COLOR_YUV420sp2RGB  = COLOR_YUV2RGB_NV21,
-    COLOR_YUV420sp2BGR  = COLOR_YUV2BGR_NV21,
-
-    COLOR_YUV2RGBA_NV12 = 94,
-    COLOR_YUV2BGRA_NV12 = 95,
-    COLOR_YUV2RGBA_NV21 = 96,
-    COLOR_YUV2BGRA_NV21 = 97,
-    COLOR_YUV420sp2RGBA = COLOR_YUV2RGBA_NV21,
-    COLOR_YUV420sp2BGRA = COLOR_YUV2BGRA_NV21,
-
-    COLOR_YUV2RGB_YV12  = 98,
-    COLOR_YUV2BGR_YV12  = 99,
-    COLOR_YUV2RGB_IYUV  = 100,
-    COLOR_YUV2BGR_IYUV  = 101,
-    COLOR_YUV2RGB_I420  = COLOR_YUV2RGB_IYUV,
-    COLOR_YUV2BGR_I420  = COLOR_YUV2BGR_IYUV,
-    COLOR_YUV420p2RGB   = COLOR_YUV2RGB_YV12,
-    COLOR_YUV420p2BGR   = COLOR_YUV2BGR_YV12,
-
-    COLOR_YUV2RGBA_YV12 = 102,
-    COLOR_YUV2BGRA_YV12 = 103,
-    COLOR_YUV2RGBA_IYUV = 104,
-    COLOR_YUV2BGRA_IYUV = 105,
-    COLOR_YUV2RGBA_I420 = COLOR_YUV2RGBA_IYUV,
-    COLOR_YUV2BGRA_I420 = COLOR_YUV2BGRA_IYUV,
-    COLOR_YUV420p2RGBA  = COLOR_YUV2RGBA_YV12,
-    COLOR_YUV420p2BGRA  = COLOR_YUV2BGRA_YV12,
-
-    COLOR_YUV2GRAY_420  = 106,
-    COLOR_YUV2GRAY_NV21 = COLOR_YUV2GRAY_420,
-    COLOR_YUV2GRAY_NV12 = COLOR_YUV2GRAY_420,
-    COLOR_YUV2GRAY_YV12 = COLOR_YUV2GRAY_420,
-    COLOR_YUV2GRAY_IYUV = COLOR_YUV2GRAY_420,
-    COLOR_YUV2GRAY_I420 = COLOR_YUV2GRAY_420,
-    COLOR_YUV420sp2GRAY = COLOR_YUV2GRAY_420,
-    COLOR_YUV420p2GRAY  = COLOR_YUV2GRAY_420,
-
-    //! YUV 4:2:2 family to RGB
-    COLOR_YUV2RGB_UYVY = 107,
-    COLOR_YUV2BGR_UYVY = 108,
-    //COLOR_YUV2RGB_VYUY = 109,
-    //COLOR_YUV2BGR_VYUY = 110,
-    COLOR_YUV2RGB_Y422 = COLOR_YUV2RGB_UYVY,
-    COLOR_YUV2BGR_Y422 = COLOR_YUV2BGR_UYVY,
-    COLOR_YUV2RGB_UYNV = COLOR_YUV2RGB_UYVY,
-    COLOR_YUV2BGR_UYNV = COLOR_YUV2BGR_UYVY,
-
-    COLOR_YUV2RGBA_UYVY = 111,
-    COLOR_YUV2BGRA_UYVY = 112,
-    //COLOR_YUV2RGBA_VYUY = 113,
-    //COLOR_YUV2BGRA_VYUY = 114,
-    COLOR_YUV2RGBA_Y422 = COLOR_YUV2RGBA_UYVY,
-    COLOR_YUV2BGRA_Y422 = COLOR_YUV2BGRA_UYVY,
-    COLOR_YUV2RGBA_UYNV = COLOR_YUV2RGBA_UYVY,
-    COLOR_YUV2BGRA_UYNV = COLOR_YUV2BGRA_UYVY,
-
-    COLOR_YUV2RGB_YUY2 = 115,
-    COLOR_YUV2BGR_YUY2 = 116,
-    COLOR_YUV2RGB_YVYU = 117,
-    COLOR_YUV2BGR_YVYU = 118,
-    COLOR_YUV2RGB_YUYV = COLOR_YUV2RGB_YUY2,
-    COLOR_YUV2BGR_YUYV = COLOR_YUV2BGR_YUY2,
-    COLOR_YUV2RGB_YUNV = COLOR_YUV2RGB_YUY2,
-    COLOR_YUV2BGR_YUNV = COLOR_YUV2BGR_YUY2,
-
-    COLOR_YUV2RGBA_YUY2 = 119,
-    COLOR_YUV2BGRA_YUY2 = 120,
-    COLOR_YUV2RGBA_YVYU = 121,
-    COLOR_YUV2BGRA_YVYU = 122,
-    COLOR_YUV2RGBA_YUYV = COLOR_YUV2RGBA_YUY2,
-    COLOR_YUV2BGRA_YUYV = COLOR_YUV2BGRA_YUY2,
-    COLOR_YUV2RGBA_YUNV = COLOR_YUV2RGBA_YUY2,
-    COLOR_YUV2BGRA_YUNV = COLOR_YUV2BGRA_YUY2,
-
-    COLOR_YUV2GRAY_UYVY = 123,
-    COLOR_YUV2GRAY_YUY2 = 124,
-    //CV_YUV2GRAY_VYUY    = CV_YUV2GRAY_UYVY,
-    COLOR_YUV2GRAY_Y422 = COLOR_YUV2GRAY_UYVY,
-    COLOR_YUV2GRAY_UYNV = COLOR_YUV2GRAY_UYVY,
-    COLOR_YUV2GRAY_YVYU = COLOR_YUV2GRAY_YUY2,
-    COLOR_YUV2GRAY_YUYV = COLOR_YUV2GRAY_YUY2,
-    COLOR_YUV2GRAY_YUNV = COLOR_YUV2GRAY_YUY2,
-
-    //! alpha premultiplication
-    COLOR_RGBA2mRGBA    = 125,
-    COLOR_mRGBA2RGBA    = 126,
-
-    //! RGB to YUV 4:2:0 family
-    COLOR_RGB2YUV_I420  = 127,
-    COLOR_BGR2YUV_I420  = 128,
-    COLOR_RGB2YUV_IYUV  = COLOR_RGB2YUV_I420,
-    COLOR_BGR2YUV_IYUV  = COLOR_BGR2YUV_I420,
-
-    COLOR_RGBA2YUV_I420 = 129,
-    COLOR_BGRA2YUV_I420 = 130,
-    COLOR_RGBA2YUV_IYUV = COLOR_RGBA2YUV_I420,
-    COLOR_BGRA2YUV_IYUV = COLOR_BGRA2YUV_I420,
-    COLOR_RGB2YUV_YV12  = 131,
-    COLOR_BGR2YUV_YV12  = 132,
-    COLOR_RGBA2YUV_YV12 = 133,
-    COLOR_BGRA2YUV_YV12 = 134,
-
-    //! Demosaicing, see @ref color_convert_bayer "color conversions" for additional information
-    COLOR_BayerBG2BGR = 46, //!< equivalent to RGGB Bayer pattern
-    COLOR_BayerGB2BGR = 47, //!< equivalent to GRBG Bayer pattern
-    COLOR_BayerRG2BGR = 48, //!< equivalent to BGGR Bayer pattern
-    COLOR_BayerGR2BGR = 49, //!< equivalent to GBRG Bayer pattern
-
-    COLOR_BayerRGGB2BGR = COLOR_BayerBG2BGR,
-    COLOR_BayerGRBG2BGR = COLOR_BayerGB2BGR,
-    COLOR_BayerBGGR2BGR = COLOR_BayerRG2BGR,
-    COLOR_BayerGBRG2BGR = COLOR_BayerGR2BGR,
-
-    COLOR_BayerRGGB2RGB = COLOR_BayerBGGR2BGR,
-    COLOR_BayerGRBG2RGB = COLOR_BayerGBRG2BGR,
-    COLOR_BayerBGGR2RGB = COLOR_BayerRGGB2BGR,
-    COLOR_BayerGBRG2RGB = COLOR_BayerGRBG2BGR,
-
-    COLOR_BayerBG2RGB = COLOR_BayerRG2BGR, //!< equivalent to RGGB Bayer pattern
-    COLOR_BayerGB2RGB = COLOR_BayerGR2BGR, //!< equivalent to GRBG Bayer pattern
-    COLOR_BayerRG2RGB = COLOR_BayerBG2BGR, //!< equivalent to BGGR Bayer pattern
-    COLOR_BayerGR2RGB = COLOR_BayerGB2BGR, //!< equivalent to GBRG Bayer pattern
-
-    COLOR_BayerBG2GRAY = 86, //!< equivalent to RGGB Bayer pattern
-    COLOR_BayerGB2GRAY = 87, //!< equivalent to GRBG Bayer pattern
-    COLOR_BayerRG2GRAY = 88, //!< equivalent to BGGR Bayer pattern
-    COLOR_BayerGR2GRAY = 89, //!< equivalent to GBRG Bayer pattern
-
-    COLOR_BayerRGGB2GRAY = COLOR_BayerBG2GRAY,
-    COLOR_BayerGRBG2GRAY = COLOR_BayerGB2GRAY,
-    COLOR_BayerBGGR2GRAY = COLOR_BayerRG2GRAY,
-    COLOR_BayerGBRG2GRAY = COLOR_BayerGR2GRAY,
-
-    //! Demosaicing using Variable Number of Gradients
-    COLOR_BayerBG2BGR_VNG = 62, //!< equivalent to RGGB Bayer pattern
-    COLOR_BayerGB2BGR_VNG = 63, //!< equivalent to GRBG Bayer pattern
-    COLOR_BayerRG2BGR_VNG = 64, //!< equivalent to BGGR Bayer pattern
-    COLOR_BayerGR2BGR_VNG = 65, //!< equivalent to GBRG Bayer pattern
-
-    COLOR_BayerRGGB2BGR_VNG = COLOR_BayerBG2BGR_VNG,
-    COLOR_BayerGRBG2BGR_VNG = COLOR_BayerGB2BGR_VNG,
-    COLOR_BayerBGGR2BGR_VNG = COLOR_BayerRG2BGR_VNG,
-    COLOR_BayerGBRG2BGR_VNG = COLOR_BayerGR2BGR_VNG,
-
-    COLOR_BayerRGGB2RGB_VNG = COLOR_BayerBGGR2BGR_VNG,
-    COLOR_BayerGRBG2RGB_VNG = COLOR_BayerGBRG2BGR_VNG,
-    COLOR_BayerBGGR2RGB_VNG = COLOR_BayerRGGB2BGR_VNG,
-    COLOR_BayerGBRG2RGB_VNG = COLOR_BayerGRBG2BGR_VNG,
-
-    COLOR_BayerBG2RGB_VNG = COLOR_BayerRG2BGR_VNG, //!< equivalent to RGGB Bayer pattern
-    COLOR_BayerGB2RGB_VNG = COLOR_BayerGR2BGR_VNG, //!< equivalent to GRBG Bayer pattern
-    COLOR_BayerRG2RGB_VNG = COLOR_BayerBG2BGR_VNG, //!< equivalent to BGGR Bayer pattern
-    COLOR_BayerGR2RGB_VNG = COLOR_BayerGB2BGR_VNG, //!< equivalent to GBRG Bayer pattern
-
-    //! Edge-Aware Demosaicing
-    COLOR_BayerBG2BGR_EA  = 135, //!< equivalent to RGGB Bayer pattern
-    COLOR_BayerGB2BGR_EA  = 136, //!< equivalent to GRBG Bayer pattern
-    COLOR_BayerRG2BGR_EA  = 137, //!< equivalent to BGGR Bayer pattern
-    COLOR_BayerGR2BGR_EA  = 138, //!< equivalent to GBRG Bayer pattern
-
-    COLOR_BayerRGGB2BGR_EA  = COLOR_BayerBG2BGR_EA,
-    COLOR_BayerGRBG2BGR_EA  = COLOR_BayerGB2BGR_EA,
-    COLOR_BayerBGGR2BGR_EA  = COLOR_BayerRG2BGR_EA,
-    COLOR_BayerGBRG2BGR_EA  = COLOR_BayerGR2BGR_EA,
-
-    COLOR_BayerRGGB2RGB_EA  = COLOR_BayerBGGR2BGR_EA,
-    COLOR_BayerGRBG2RGB_EA  = COLOR_BayerGBRG2BGR_EA,
-    COLOR_BayerBGGR2RGB_EA  = COLOR_BayerRGGB2BGR_EA,
-    COLOR_BayerGBRG2RGB_EA  = COLOR_BayerGRBG2BGR_EA,
-
-    COLOR_BayerBG2RGB_EA  = COLOR_BayerRG2BGR_EA, //!< equivalent to RGGB Bayer pattern
-    COLOR_BayerGB2RGB_EA  = COLOR_BayerGR2BGR_EA, //!< equivalent to GRBG Bayer pattern
-    COLOR_BayerRG2RGB_EA  = COLOR_BayerBG2BGR_EA, //!< equivalent to BGGR Bayer pattern
-    COLOR_BayerGR2RGB_EA  = COLOR_BayerGB2BGR_EA, //!< equivalent to GBRG Bayer pattern
-
-    //! Demosaicing with alpha channel
-    COLOR_BayerBG2BGRA = 139, //!< equivalent to RGGB Bayer pattern
-    COLOR_BayerGB2BGRA = 140, //!< equivalent to GRBG Bayer pattern
-    COLOR_BayerRG2BGRA = 141, //!< equivalent to BGGR Bayer pattern
-    COLOR_BayerGR2BGRA = 142, //!< equivalent to GBRG Bayer pattern
-
-    COLOR_BayerRGGB2BGRA = COLOR_BayerBG2BGRA,
-    COLOR_BayerGRBG2BGRA = COLOR_BayerGB2BGRA,
-    COLOR_BayerBGGR2BGRA = COLOR_BayerRG2BGRA,
-    COLOR_BayerGBRG2BGRA = COLOR_BayerGR2BGRA,
-
-    COLOR_BayerRGGB2RGBA = COLOR_BayerBGGR2BGRA,
-    COLOR_BayerGRBG2RGBA = COLOR_BayerGBRG2BGRA,
-    COLOR_BayerBGGR2RGBA = COLOR_BayerRGGB2BGRA,
-    COLOR_BayerGBRG2RGBA = COLOR_BayerGRBG2BGRA,
-
-    COLOR_BayerBG2RGBA = COLOR_BayerRG2BGRA, //!< equivalent to RGGB Bayer pattern
-    COLOR_BayerGB2RGBA = COLOR_BayerGR2BGRA, //!< equivalent to GRBG Bayer pattern
-    COLOR_BayerRG2RGBA = COLOR_BayerBG2BGRA, //!< equivalent to BGGR Bayer pattern
-    COLOR_BayerGR2RGBA = COLOR_BayerGB2BGRA, //!< equivalent to GBRG Bayer pattern
-
-    COLOR_COLORCVT_MAX  = 143
-};
-
-//! @addtogroup imgproc_shape
-//! @{
-
-//! types of intersection between rectangles
-enum RectanglesIntersectTypes {
-    INTERSECT_NONE = 0, //!< No intersection
-    INTERSECT_PARTIAL  = 1, //!< There is a partial intersection
-    INTERSECT_FULL  = 2 //!< One of the rectangle is fully enclosed in the other
-};
-
-/** types of line
-@ingroup imgproc_draw
-*/
-enum LineTypes {
-    FILLED  = -1,
-    LINE_4  = 4, //!< 4-connected line
-    LINE_8  = 8, //!< 8-connected line
-    LINE_AA = 16 //!< antialiased line
-};
-
-/** Only a subset of Hershey fonts <https://en.wikipedia.org/wiki/Hershey_fonts> are supported
-@ingroup imgproc_draw
-*/
-enum HersheyFonts {
-    FONT_HERSHEY_SIMPLEX        = 0, //!< normal size sans-serif font
-    FONT_HERSHEY_PLAIN          = 1, //!< small size sans-serif font
-    FONT_HERSHEY_DUPLEX         = 2, //!< normal size sans-serif font (more complex than FONT_HERSHEY_SIMPLEX)
-    FONT_HERSHEY_COMPLEX        = 3, //!< normal size serif font
-    FONT_HERSHEY_TRIPLEX        = 4, //!< normal size serif font (more complex than FONT_HERSHEY_COMPLEX)
-    FONT_HERSHEY_COMPLEX_SMALL  = 5, //!< smaller version of FONT_HERSHEY_COMPLEX
-    FONT_HERSHEY_SCRIPT_SIMPLEX = 6, //!< hand-writing style font
-    FONT_HERSHEY_SCRIPT_COMPLEX = 7, //!< more complex variant of FONT_HERSHEY_SCRIPT_SIMPLEX
-    FONT_ITALIC                 = 16 //!< flag for italic font
-};
-
-/** Possible set of marker types used for the cv::drawMarker function
-@ingroup imgproc_draw
-*/
-enum MarkerTypes
-{
-    MARKER_CROSS = 0,           //!< A crosshair marker shape
-    MARKER_TILTED_CROSS = 1,    //!< A 45 degree tilted crosshair marker shape
-    MARKER_STAR = 2,            //!< A star marker shape, combination of cross and tilted cross
-    MARKER_DIAMOND = 3,         //!< A diamond marker shape
-    MARKER_SQUARE = 4,          //!< A square marker shape
-    MARKER_TRIANGLE_UP = 5,     //!< An upwards pointing triangle marker shape
-    MARKER_TRIANGLE_DOWN = 6    //!< A downwards pointing triangle marker shape
-};
-
-/** @brief finds arbitrary template in the grayscale image using Generalized Hough Transform
-*/
-class CV_EXPORTS_W GeneralizedHough : public Algorithm
-{
-public:
-    //! set template to search
-    CV_WRAP virtual void setTemplate(InputArray templ, Point templCenter = Point(-1, -1)) = 0;
-    CV_WRAP virtual void setTemplate(InputArray edges, InputArray dx, InputArray dy, Point templCenter = Point(-1, -1)) = 0;
-
-    //! find template on image
-    CV_WRAP virtual void detect(InputArray image, OutputArray positions, OutputArray votes = noArray()) = 0;
-    CV_WRAP virtual void detect(InputArray edges, InputArray dx, InputArray dy, OutputArray positions, OutputArray votes = noArray()) = 0;
-
-    //! Canny low threshold.
-    CV_WRAP virtual void setCannyLowThresh(int cannyLowThresh) = 0;
-    CV_WRAP virtual int getCannyLowThresh() const = 0;
-
-    //! Canny high threshold.
-    CV_WRAP virtual void setCannyHighThresh(int cannyHighThresh) = 0;
-    CV_WRAP virtual int getCannyHighThresh() const = 0;
-
-    //! Minimum distance between the centers of the detected objects.
-    CV_WRAP virtual void setMinDist(double minDist) = 0;
-    CV_WRAP virtual double getMinDist() const = 0;
-
-    //! Inverse ratio of the accumulator resolution to the image resolution.
-    CV_WRAP virtual void setDp(double dp) = 0;
-    CV_WRAP virtual double getDp() const = 0;
-
-    //! Maximal size of inner buffers.
-    CV_WRAP virtual void setMaxBufferSize(int maxBufferSize) = 0;
-    CV_WRAP virtual int getMaxBufferSize() const = 0;
-};
-
-/** @brief finds arbitrary template in the grayscale image using Generalized Hough Transform
-
-Detects position only without translation and rotation @cite Ballard1981 .
-*/
-class CV_EXPORTS_W GeneralizedHoughBallard : public GeneralizedHough
-{
-public:
-    //! R-Table levels.
-    CV_WRAP virtual void setLevels(int levels) = 0;
-    CV_WRAP virtual int getLevels() const = 0;
-
-    //! The accumulator threshold for the template centers at the detection stage. The smaller it is, the more false positions may be detected.
-    CV_WRAP virtual void setVotesThreshold(int votesThreshold) = 0;
-    CV_WRAP virtual int getVotesThreshold() const = 0;
-};
-
-/** @brief finds arbitrary template in the grayscale image using Generalized Hough Transform
-
-Detects position, translation and rotation @cite Guil1999 .
-*/
-class CV_EXPORTS_W GeneralizedHoughGuil : public GeneralizedHough
-{
-public:
-    //! Angle difference in degrees between two points in feature.
-    CV_WRAP virtual void setXi(double xi) = 0;
-    CV_WRAP virtual double getXi() const = 0;
-
-    //! Feature table levels.
-    CV_WRAP virtual void setLevels(int levels) = 0;
-    CV_WRAP virtual int getLevels() const = 0;
-
-    //! Maximal difference between angles that treated as equal.
-    CV_WRAP virtual void setAngleEpsilon(double angleEpsilon) = 0;
-    CV_WRAP virtual double getAngleEpsilon() const = 0;
-
-    //! Minimal rotation angle to detect in degrees.
-    CV_WRAP virtual void setMinAngle(double minAngle) = 0;
-    CV_WRAP virtual double getMinAngle() const = 0;
-
-    //! Maximal rotation angle to detect in degrees.
-    CV_WRAP virtual void setMaxAngle(double maxAngle) = 0;
-    CV_WRAP virtual double getMaxAngle() const = 0;
-
-    //! Angle step in degrees.
-    CV_WRAP virtual void setAngleStep(double angleStep) = 0;
-    CV_WRAP virtual double getAngleStep() const = 0;
-
-    //! Angle votes threshold.
-    CV_WRAP virtual void setAngleThresh(int angleThresh) = 0;
-    CV_WRAP virtual int getAngleThresh() const = 0;
-
-    //! Minimal scale to detect.
-    CV_WRAP virtual void setMinScale(double minScale) = 0;
-    CV_WRAP virtual double getMinScale() const = 0;
-
-    //! Maximal scale to detect.
-    CV_WRAP virtual void setMaxScale(double maxScale) = 0;
-    CV_WRAP virtual double getMaxScale() const = 0;
-
-    //! Scale step.
-    CV_WRAP virtual void setScaleStep(double scaleStep) = 0;
-    CV_WRAP virtual double getScaleStep() const = 0;
-
-    //! Scale votes threshold.
-    CV_WRAP virtual void setScaleThresh(int scaleThresh) = 0;
-    CV_WRAP virtual int getScaleThresh() const = 0;
-
-    //! Position votes threshold.
-    CV_WRAP virtual void setPosThresh(int posThresh) = 0;
-    CV_WRAP virtual int getPosThresh() const = 0;
-};
-
-//! @} imgproc_shape
-
-//! @addtogroup imgproc_hist
-//! @{
-
-/** @brief Base class for Contrast Limited Adaptive Histogram Equalization.
-*/
-class CV_EXPORTS_W CLAHE : public Algorithm
-{
-public:
-    /** @brief Equalizes the histogram of a grayscale image using Contrast Limited Adaptive Histogram Equalization.
-
-    @param src Source image of type CV_8UC1 or CV_16UC1.
-    @param dst Destination image.
-     */
-    CV_WRAP virtual void apply(InputArray src, OutputArray dst) = 0;
-
-    /** @brief Sets threshold for contrast limiting.
-
-    @param clipLimit threshold value.
-    */
-    CV_WRAP virtual void setClipLimit(double clipLimit) = 0;
-
-    //! Returns threshold value for contrast limiting.
-    CV_WRAP virtual double getClipLimit() const = 0;
-
-    /** @brief Sets size of grid for histogram equalization. Input image will be divided into
-    equally sized rectangular tiles.
-
-    @param tileGridSize defines the number of tiles in row and column.
-    */
-    CV_WRAP virtual void setTilesGridSize(Size tileGridSize) = 0;
-
-    //!@brief Returns Size defines the number of tiles in row and column.
-    CV_WRAP virtual Size getTilesGridSize() const = 0;
-
-    CV_WRAP virtual void collectGarbage() = 0;
-};
-
-//! @} imgproc_hist
-
-//! @addtogroup imgproc_subdiv2d
-//! @{
-
-class CV_EXPORTS_W Subdiv2D
-{
-public:
-    /** Subdiv2D point location cases */
-    enum { PTLOC_ERROR        = -2, //!< Point location error
-           PTLOC_OUTSIDE_RECT = -1, //!< Point outside the subdivision bounding rect
-           PTLOC_INSIDE       = 0, //!< Point inside some facet
-           PTLOC_VERTEX       = 1, //!< Point coincides with one of the subdivision vertices
-           PTLOC_ON_EDGE      = 2  //!< Point on some edge
-         };
-
-    /** Subdiv2D edge type navigation (see: getEdge()) */
-    enum { NEXT_AROUND_ORG   = 0x00,
-           NEXT_AROUND_DST   = 0x22,
-           PREV_AROUND_ORG   = 0x11,
-           PREV_AROUND_DST   = 0x33,
-           NEXT_AROUND_LEFT  = 0x13,
-           NEXT_AROUND_RIGHT = 0x31,
-           PREV_AROUND_LEFT  = 0x20,
-           PREV_AROUND_RIGHT = 0x02
-         };
-
-    /** creates an empty Subdiv2D object.
-    To create a new empty Delaunay subdivision you need to use the #initDelaunay function.
-     */
-    CV_WRAP Subdiv2D();
-
-    /** @overload
-
-    @param rect Rectangle that includes all of the 2D points that are to be added to the subdivision.
-
-    The function creates an empty Delaunay subdivision where 2D points can be added using the function
-    insert() . All of the points to be added must be within the specified rectangle, otherwise a runtime
-    error is raised.
-     */
-    CV_WRAP Subdiv2D(Rect rect);
-
-    /** @brief Creates a new empty Delaunay subdivision
-
-    @param rect Rectangle that includes all of the 2D points that are to be added to the subdivision.
-
-     */
-    CV_WRAP void initDelaunay(Rect rect);
-
-    /** @brief Insert a single point into a Delaunay triangulation.
-
-    @param pt Point to insert.
-
-    The function inserts a single point into a subdivision and modifies the subdivision topology
-    appropriately. If a point with the same coordinates exists already, no new point is added.
-    @returns the ID of the point.
-
-    @note If the point is outside of the triangulation specified rect a runtime error is raised.
-     */
-    CV_WRAP int insert(Point2f pt);
-
-    /** @brief Insert multiple points into a Delaunay triangulation.
-
-    @param ptvec Points to insert.
-
-    The function inserts a vector of points into a subdivision and modifies the subdivision topology
-    appropriately.
-     */
-    CV_WRAP void insert(const std::vector<Point2f>& ptvec);
-
-    /** @brief Returns the location of a point within a Delaunay triangulation.
-
-    @param pt Point to locate.
-    @param edge Output edge that the point belongs to or is located to the right of it.
-    @param vertex Optional output vertex the input point coincides with.
-
-    The function locates the input point within the subdivision and gives one of the triangle edges
-    or vertices.
-
-    @returns an integer which specify one of the following five cases for point location:
-    -  The point falls into some facet. The function returns #PTLOC_INSIDE and edge will contain one of
-       edges of the facet.
-    -  The point falls onto the edge. The function returns #PTLOC_ON_EDGE and edge will contain this edge.
-    -  The point coincides with one of the subdivision vertices. The function returns #PTLOC_VERTEX and
-       vertex will contain a pointer to the vertex.
-    -  The point is outside the subdivision reference rectangle. The function returns #PTLOC_OUTSIDE_RECT
-       and no pointers are filled.
-    -  One of input arguments is invalid. A runtime error is raised or, if silent or "parent" error
-       processing mode is selected, #PTLOC_ERROR is returned.
-     */
-    CV_WRAP int locate(Point2f pt, CV_OUT int& edge, CV_OUT int& vertex);
-
-    /** @brief Finds the subdivision vertex closest to the given point.
-
-    @param pt Input point.
-    @param nearestPt Output subdivision vertex point.
-
-    The function is another function that locates the input point within the subdivision. It finds the
-    subdivision vertex that is the closest to the input point. It is not necessarily one of vertices
-    of the facet containing the input point, though the facet (located using locate() ) is used as a
-    starting point.
-
-    @returns vertex ID.
-     */
-    CV_WRAP int findNearest(Point2f pt, CV_OUT Point2f* nearestPt = 0);
-
-    /** @brief Returns a list of all edges.
-
-    @param edgeList Output vector.
-
-    The function gives each edge as a 4 numbers vector, where each two are one of the edge
-    vertices. i.e. org_x = v[0], org_y = v[1], dst_x = v[2], dst_y = v[3].
-     */
-    CV_WRAP void getEdgeList(CV_OUT std::vector<Vec4f>& edgeList) const;
-
-    /** @brief Returns a list of the leading edge ID connected to each triangle.
-
-    @param leadingEdgeList Output vector.
-
-    The function gives one edge ID for each triangle.
-     */
-    CV_WRAP void getLeadingEdgeList(CV_OUT std::vector<int>& leadingEdgeList) const;
-
-    /** @brief Returns a list of all triangles.
-
-    @param triangleList Output vector.
-
-    The function gives each triangle as a 6 numbers vector, where each two are one of the triangle
-    vertices. i.e. p1_x = v[0], p1_y = v[1], p2_x = v[2], p2_y = v[3], p3_x = v[4], p3_y = v[5].
-     */
-    CV_WRAP void getTriangleList(CV_OUT std::vector<Vec6f>& triangleList) const;
-
-    /** @brief Returns a list of all Voronoi facets.
-
-    @param idx Vector of vertices IDs to consider. For all vertices you can pass empty vector.
-    @param facetList Output vector of the Voronoi facets.
-    @param facetCenters Output vector of the Voronoi facets center points.
-
-     */
-    CV_WRAP void getVoronoiFacetList(const std::vector<int>& idx, CV_OUT std::vector<std::vector<Point2f> >& facetList,
-                                     CV_OUT std::vector<Point2f>& facetCenters);
-
-    /** @brief Returns vertex location from vertex ID.
-
-    @param vertex vertex ID.
-    @param firstEdge Optional. The first edge ID which is connected to the vertex.
-    @returns vertex (x,y)
-
-     */
-    CV_WRAP Point2f getVertex(int vertex, CV_OUT int* firstEdge = 0) const;
-
-    /** @brief Returns one of the edges related to the given edge.
-
-    @param edge Subdivision edge ID.
-    @param nextEdgeType Parameter specifying which of the related edges to return.
-    The following values are possible:
-    -   NEXT_AROUND_ORG next around the edge origin ( eOnext on the picture below if e is the input edge)
-    -   NEXT_AROUND_DST next around the edge vertex ( eDnext )
-    -   PREV_AROUND_ORG previous around the edge origin (reversed eRnext )
-    -   PREV_AROUND_DST previous around the edge destination (reversed eLnext )
-    -   NEXT_AROUND_LEFT next around the left facet ( eLnext )
-    -   NEXT_AROUND_RIGHT next around the right facet ( eRnext )
-    -   PREV_AROUND_LEFT previous around the left facet (reversed eOnext )
-    -   PREV_AROUND_RIGHT previous around the right facet (reversed eDnext )
-
-    ![sample output](pics/quadedge.png)
-
-    @returns edge ID related to the input edge.
-     */
-    CV_WRAP int getEdge( int edge, int nextEdgeType ) const;
-
-    /** @brief Returns next edge around the edge origin.
-
-    @param edge Subdivision edge ID.
-
-    @returns an integer which is next edge ID around the edge origin: eOnext on the
-    picture above if e is the input edge).
-     */
-    CV_WRAP int nextEdge(int edge) const;
-
-    /** @brief Returns another edge of the same quad-edge.
-
-    @param edge Subdivision edge ID.
-    @param rotate Parameter specifying which of the edges of the same quad-edge as the input
-    one to return. The following values are possible:
-    -   0 - the input edge ( e on the picture below if e is the input edge)
-    -   1 - the rotated edge ( eRot )
-    -   2 - the reversed edge (reversed e (in green))
-    -   3 - the reversed rotated edge (reversed eRot (in green))
-
-    @returns one of the edges ID of the same quad-edge as the input edge.
-     */
-    CV_WRAP int rotateEdge(int edge, int rotate) const;
-    CV_WRAP int symEdge(int edge) const;
-
-    /** @brief Returns the edge origin.
-
-    @param edge Subdivision edge ID.
-    @param orgpt Output vertex location.
-
-    @returns vertex ID.
-     */
-    CV_WRAP int edgeOrg(int edge, CV_OUT Point2f* orgpt = 0) const;
-
-    /** @brief Returns the edge destination.
-
-    @param edge Subdivision edge ID.
-    @param dstpt Output vertex location.
-
-    @returns vertex ID.
-     */
-    CV_WRAP int edgeDst(int edge, CV_OUT Point2f* dstpt = 0) const;
-
-protected:
-    int newEdge();
-    void deleteEdge(int edge);
-    int newPoint(Point2f pt, bool isvirtual, int firstEdge = 0);
-    void deletePoint(int vtx);
-    void setEdgePoints( int edge, int orgPt, int dstPt );
-    void splice( int edgeA, int edgeB );
-    int connectEdges( int edgeA, int edgeB );
-    void swapEdges( int edge );
-    int isRightOf(Point2f pt, int edge) const;
-    void calcVoronoi();
-    void clearVoronoi();
-    void checkSubdiv() const;
-
-    struct CV_EXPORTS Vertex
-    {
-        Vertex();
-        Vertex(Point2f pt, bool isvirtual, int firstEdge=0);
-        bool isvirtual() const;
-        bool isfree() const;
-
-        int firstEdge;
-        int type;
-        Point2f pt;
-    };
-
-    struct CV_EXPORTS QuadEdge
-    {
-        QuadEdge();
-        QuadEdge(int edgeidx);
-        bool isfree() const;
-
-        int next[4];
-        int pt[4];
-    };
-
-    //! All of the vertices
-    std::vector<Vertex> vtx;
-    //! All of the edges
-    std::vector<QuadEdge> qedges;
-    int freeQEdge;
-    int freePoint;
-    bool validGeometry;
-
-    int recentEdge;
-    //! Top left corner of the bounding rect
-    Point2f topLeft;
-    //! Bottom right corner of the bounding rect
-    Point2f bottomRight;
-};
-
-//! @} imgproc_subdiv2d
-
-//! @addtogroup imgproc_feature
-//! @{
-
-/** @example samples/cpp/lsd_lines.cpp
-An example using the LineSegmentDetector
-\image html building_lsd.png "Sample output image" width=434 height=300
-*/
-
-/** @brief Line segment detector class
-
-following the algorithm described at @cite Rafael12 .
-
-@note Implementation has been removed from OpenCV version 3.4.6 to 3.4.15 and version 4.1.0 to 4.5.3 due original code license conflict.
-restored again after [Computation of a NFA](https://github.com/rafael-grompone-von-gioi/binomial_nfa) code published under the MIT license.
-*/
-class CV_EXPORTS_W LineSegmentDetector : public Algorithm
-{
-public:
-
-    /** @brief Finds lines in the input image.
-
-    This is the output of the default parameters of the algorithm on the above shown image.
-
-    ![image](pics/building_lsd.png)
-
-    @param image A grayscale (CV_8UC1) input image. If only a roi needs to be selected, use:
-    `lsd_ptr-\>detect(image(roi), lines, ...); lines += Scalar(roi.x, roi.y, roi.x, roi.y);`
-    @param lines A vector of Vec4f elements specifying the beginning and ending point of a line. Where
-    Vec4f is (x1, y1, x2, y2), point 1 is the start, point 2 - end. Returned lines are strictly
-    oriented depending on the gradient.
-    @param width Vector of widths of the regions, where the lines are found. E.g. Width of line.
-    @param prec Vector of precisions with which the lines are found.
-    @param nfa Vector containing number of false alarms in the line region, with precision of 10%. The
-    bigger the value, logarithmically better the detection.
-    - -1 corresponds to 10 mean false alarms
-    - 0 corresponds to 1 mean false alarm
-    - 1 corresponds to 0.1 mean false alarms
-    This vector will be calculated only when the objects type is #LSD_REFINE_ADV.
-    */
-    CV_WRAP virtual void detect(InputArray image, OutputArray lines,
-                        OutputArray width = noArray(), OutputArray prec = noArray(),
-                        OutputArray nfa = noArray()) = 0;
-
-    /** @brief Draws the line segments on a given image.
-    @param image The image, where the lines will be drawn. Should be bigger or equal to the image,
-    where the lines were found.
-    @param lines A vector of the lines that needed to be drawn.
-     */
-    CV_WRAP virtual void drawSegments(InputOutputArray image, InputArray lines) = 0;
-
-    /** @brief Draws two groups of lines in blue and red, counting the non overlapping (mismatching) pixels.
-
-    @param size The size of the image, where lines1 and lines2 were found.
-    @param lines1 The first group of lines that needs to be drawn. It is visualized in blue color.
-    @param lines2 The second group of lines. They visualized in red color.
-    @param image Optional image, where the lines will be drawn. The image should be color(3-channel)
-    in order for lines1 and lines2 to be drawn in the above mentioned colors.
-     */
-    CV_WRAP virtual int compareSegments(const Size& size, InputArray lines1, InputArray lines2, InputOutputArray image = noArray()) = 0;
-
-    virtual ~LineSegmentDetector() { }
-};
-
-/** @brief Creates a smart pointer to a LineSegmentDetector object and initializes it.
-
-The LineSegmentDetector algorithm is defined using the standard values. Only advanced users may want
-to edit those, as to tailor it for their own application.
-
-@param refine The way found lines will be refined, see #LineSegmentDetectorModes
-@param scale The scale of the image that will be used to find the lines. Range (0..1].
-@param sigma_scale Sigma for Gaussian filter. It is computed as sigma = sigma_scale/scale.
-@param quant Bound to the quantization error on the gradient norm.
-@param ang_th Gradient angle tolerance in degrees.
-@param log_eps Detection threshold: -log10(NFA) \> log_eps. Used only when advance refinement is chosen.
-@param density_th Minimal density of aligned region points in the enclosing rectangle.
-@param n_bins Number of bins in pseudo-ordering of gradient modulus.
- */
-CV_EXPORTS_W Ptr<LineSegmentDetector> createLineSegmentDetector(
-    int refine = LSD_REFINE_STD, double scale = 0.8,
-    double sigma_scale = 0.6, double quant = 2.0, double ang_th = 22.5,
-    double log_eps = 0, double density_th = 0.7, int n_bins = 1024);
-
-//! @} imgproc_feature
-
-//! @addtogroup imgproc_filter
-//! @{
-
-/** @brief Returns Gaussian filter coefficients.
-
-The function computes and returns the \f$\texttt{ksize} \times 1\f$ matrix of Gaussian filter
-coefficients:
-
-\f[G_i= \alpha *e^{-(i-( \texttt{ksize} -1)/2)^2/(2* \texttt{sigma}^2)},\f]
-
-where \f$i=0..\texttt{ksize}-1\f$ and \f$\alpha\f$ is the scale factor chosen so that \f$\sum_i G_i=1\f$.
-
-Two of such generated kernels can be passed to sepFilter2D. Those functions automatically recognize
-smoothing kernels (a symmetrical kernel with sum of weights equal to 1) and handle them accordingly.
-You may also use the higher-level GaussianBlur.
-@param ksize Aperture size. It should be odd ( \f$\texttt{ksize} \mod 2 = 1\f$ ) and positive.
-@param sigma Gaussian standard deviation. If it is non-positive, it is computed from ksize as
-`sigma = 0.3*((ksize-1)*0.5 - 1) + 0.8`.
-@param ktype Type of filter coefficients. It can be CV_32F or CV_64F .
-@sa  sepFilter2D, getDerivKernels, getStructuringElement, GaussianBlur
- */
-CV_EXPORTS_W Mat getGaussianKernel( int ksize, double sigma, int ktype = CV_64F );
-
-/** @brief Returns filter coefficients for computing spatial image derivatives.
-
-The function computes and returns the filter coefficients for spatial image derivatives. When
-`ksize=FILTER_SCHARR`, the Scharr \f$3 \times 3\f$ kernels are generated (see #Scharr). Otherwise, Sobel
-kernels are generated (see #Sobel). The filters are normally passed to #sepFilter2D or to
-
-@param kx Output matrix of row filter coefficients. It has the type ktype .
-@param ky Output matrix of column filter coefficients. It has the type ktype .
-@param dx Derivative order in respect of x.
-@param dy Derivative order in respect of y.
-@param ksize Aperture size. It can be FILTER_SCHARR, 1, 3, 5, or 7.
-@param normalize Flag indicating whether to normalize (scale down) the filter coefficients or not.
-Theoretically, the coefficients should have the denominator \f$=2^{ksize*2-dx-dy-2}\f$. If you are
-going to filter floating-point images, you are likely to use the normalized kernels. But if you
-compute derivatives of an 8-bit image, store the results in a 16-bit image, and wish to preserve
-all the fractional bits, you may want to set normalize=false .
-@param ktype Type of filter coefficients. It can be CV_32f or CV_64F .
- */
-CV_EXPORTS_W void getDerivKernels( OutputArray kx, OutputArray ky,
-                                   int dx, int dy, int ksize,
-                                   bool normalize = false, int ktype = CV_32F );
-
-/** @brief Returns Gabor filter coefficients.
-
-For more details about gabor filter equations and parameters, see: [Gabor
-Filter](http://en.wikipedia.org/wiki/Gabor_filter).
-
-@param ksize Size of the filter returned.
-@param sigma Standard deviation of the gaussian envelope.
-@param theta Orientation of the normal to the parallel stripes of a Gabor function.
-@param lambd Wavelength of the sinusoidal factor.
-@param gamma Spatial aspect ratio.
-@param psi Phase offset.
-@param ktype Type of filter coefficients. It can be CV_32F or CV_64F .
- */
-CV_EXPORTS_W Mat getGaborKernel( Size ksize, double sigma, double theta, double lambd,
-                                 double gamma, double psi = CV_PI*0.5, int ktype = CV_64F );
-
-//! returns "magic" border value for erosion and dilation. It is automatically transformed to Scalar::all(-DBL_MAX) for dilation.
-static inline Scalar morphologyDefaultBorderValue() { return Scalar::all(DBL_MAX); }
-
-/** @brief Returns a structuring element of the specified size and shape for morphological operations.
-
-The function constructs and returns the structuring element that can be further passed to #erode,
-#dilate or #morphologyEx. But you can also construct an arbitrary binary mask yourself and use it as
-the structuring element.
-
-@param shape Element shape that could be one of #MorphShapes
-@param ksize Size of the structuring element.
-@param anchor Anchor position within the element. The default value \f$(-1, -1)\f$ means that the
-anchor is at the center. Note that only the shape of a cross-shaped element depends on the anchor
-position. In other cases the anchor just regulates how much the result of the morphological
-operation is shifted.
- */
-CV_EXPORTS_W Mat getStructuringElement(int shape, Size ksize, Point anchor = Point(-1,-1));
-
-/** @example samples/cpp/tutorial_code/ImgProc/Smoothing/Smoothing.cpp
-Sample code for simple filters
-![Sample screenshot](Smoothing_Tutorial_Result_Median_Filter.jpg)
-Check @ref tutorial_gausian_median_blur_bilateral_filter "the corresponding tutorial" for more details
- */
-
-/** @brief Blurs an image using the median filter.
-
-The function smoothes an image using the median filter with the \f$\texttt{ksize} \times
-\texttt{ksize}\f$ aperture. Each channel of a multi-channel image is processed independently.
-In-place operation is supported.
-
-@note The median filter uses #BORDER_REPLICATE internally to cope with border pixels, see #BorderTypes
-
-@param src input 1-, 3-, or 4-channel image; when ksize is 3 or 5, the image depth should be
-CV_8U, CV_16U, or CV_32F, for larger aperture sizes, it can only be CV_8U.
-@param dst destination array of the same size and type as src.
-@param ksize aperture linear size; it must be odd and greater than 1, for example: 3, 5, 7 ...
-@sa  bilateralFilter, blur, boxFilter, GaussianBlur
- */
-CV_EXPORTS_W void medianBlur( InputArray src, OutputArray dst, int ksize );
-
-/** @brief Blurs an image using a Gaussian filter.
-
-The function convolves the source image with the specified Gaussian kernel. In-place filtering is
-supported.
-
-@param src input image; the image can have any number of channels, which are processed
-independently, but the depth should be CV_8U, CV_16U, CV_16S, CV_32F or CV_64F.
-@param dst output image of the same size and type as src.
-@param ksize Gaussian kernel size. ksize.width and ksize.height can differ but they both must be
-positive and odd. Or, they can be zero's and then they are computed from sigma.
-@param sigmaX Gaussian kernel standard deviation in X direction.
-@param sigmaY Gaussian kernel standard deviation in Y direction; if sigmaY is zero, it is set to be
-equal to sigmaX, if both sigmas are zeros, they are computed from ksize.width and ksize.height,
-respectively (see #getGaussianKernel for details); to fully control the result regardless of
-possible future modifications of all this semantics, it is recommended to specify all of ksize,
-sigmaX, and sigmaY.
-@param borderType pixel extrapolation method, see #BorderTypes. #BORDER_WRAP is not supported.
-
-@sa  sepFilter2D, filter2D, blur, boxFilter, bilateralFilter, medianBlur
- */
-CV_EXPORTS_W void GaussianBlur( InputArray src, OutputArray dst, Size ksize,
-                                double sigmaX, double sigmaY = 0,
-                                int borderType = BORDER_DEFAULT );
-
-/** @brief Applies the bilateral filter to an image.
-
-The function applies bilateral filtering to the input image, as described in
-http://www.dai.ed.ac.uk/CVonline/LOCAL_COPIES/MANDUCHI1/Bilateral_Filtering.html
-bilateralFilter can reduce unwanted noise very well while keeping edges fairly sharp. However, it is
-very slow compared to most filters.
-
-_Sigma values_: For simplicity, you can set the 2 sigma values to be the same. If they are small (\<
-10), the filter will not have much effect, whereas if they are large (\> 150), they will have a very
-strong effect, making the image look "cartoonish".
-
-_Filter size_: Large filters (d \> 5) are very slow, so it is recommended to use d=5 for real-time
-applications, and perhaps d=9 for offline applications that need heavy noise filtering.
-
-This filter does not work inplace.
-@param src Source 8-bit or floating-point, 1-channel or 3-channel image.
-@param dst Destination image of the same size and type as src .
-@param d Diameter of each pixel neighborhood that is used during filtering. If it is non-positive,
-it is computed from sigmaSpace.
-@param sigmaColor Filter sigma in the color space. A larger value of the parameter means that
-farther colors within the pixel neighborhood (see sigmaSpace) will be mixed together, resulting
-in larger areas of semi-equal color.
-@param sigmaSpace Filter sigma in the coordinate space. A larger value of the parameter means that
-farther pixels will influence each other as long as their colors are close enough (see sigmaColor
-). When d\>0, it specifies the neighborhood size regardless of sigmaSpace. Otherwise, d is
-proportional to sigmaSpace.
-@param borderType border mode used to extrapolate pixels outside of the image, see #BorderTypes
- */
-CV_EXPORTS_W void bilateralFilter( InputArray src, OutputArray dst, int d,
-                                   double sigmaColor, double sigmaSpace,
-                                   int borderType = BORDER_DEFAULT );
-
-/** @brief Blurs an image using the box filter.
-
-The function smooths an image using the kernel:
-
-\f[\texttt{K} =  \alpha \begin{bmatrix} 1 & 1 & 1 &  \cdots & 1 & 1  \\ 1 & 1 & 1 &  \cdots & 1 & 1  \\ \hdotsfor{6} \\ 1 & 1 & 1 &  \cdots & 1 & 1 \end{bmatrix}\f]
-
-where
-
-\f[\alpha = \begin{cases} \frac{1}{\texttt{ksize.width*ksize.height}} & \texttt{when } \texttt{normalize=true}  \\1 & \texttt{otherwise}\end{cases}\f]
-
-Unnormalized box filter is useful for computing various integral characteristics over each pixel
-neighborhood, such as covariance matrices of image derivatives (used in dense optical flow
-algorithms, and so on). If you need to compute pixel sums over variable-size windows, use #integral.
-
-@param src input image.
-@param dst output image of the same size and type as src.
-@param ddepth the output image depth (-1 to use src.depth()).
-@param ksize blurring kernel size.
-@param anchor anchor point; default value Point(-1,-1) means that the anchor is at the kernel
-center.
-@param normalize flag, specifying whether the kernel is normalized by its area or not.
-@param borderType border mode used to extrapolate pixels outside of the image, see #BorderTypes. #BORDER_WRAP is not supported.
-@sa  blur, bilateralFilter, GaussianBlur, medianBlur, integral
- */
-CV_EXPORTS_W void boxFilter( InputArray src, OutputArray dst, int ddepth,
-                             Size ksize, Point anchor = Point(-1,-1),
-                             bool normalize = true,
-                             int borderType = BORDER_DEFAULT );
-
-/** @brief Calculates the normalized sum of squares of the pixel values overlapping the filter.
-
-For every pixel \f$ (x, y) \f$ in the source image, the function calculates the sum of squares of those neighboring
-pixel values which overlap the filter placed over the pixel \f$ (x, y) \f$.
-
-The unnormalized square box filter can be useful in computing local image statistics such as the the local
-variance and standard deviation around the neighborhood of a pixel.
-
-@param src input image
-@param dst output image of the same size and type as src
-@param ddepth the output image depth (-1 to use src.depth())
-@param ksize kernel size
-@param anchor kernel anchor point. The default value of Point(-1, -1) denotes that the anchor is at the kernel
-center.
-@param normalize flag, specifying whether the kernel is to be normalized by it's area or not.
-@param borderType border mode used to extrapolate pixels outside of the image, see #BorderTypes. #BORDER_WRAP is not supported.
-@sa boxFilter
-*/
-CV_EXPORTS_W void sqrBoxFilter( InputArray src, OutputArray dst, int ddepth,
-                                Size ksize, Point anchor = Point(-1, -1),
-                                bool normalize = true,
-                                int borderType = BORDER_DEFAULT );
-
-/** @brief Blurs an image using the normalized box filter.
-
-The function smooths an image using the kernel:
-
-\f[\texttt{K} =  \frac{1}{\texttt{ksize.width*ksize.height}} \begin{bmatrix} 1 & 1 & 1 &  \cdots & 1 & 1  \\ 1 & 1 & 1 &  \cdots & 1 & 1  \\ \hdotsfor{6} \\ 1 & 1 & 1 &  \cdots & 1 & 1  \\ \end{bmatrix}\f]
-
-The call `blur(src, dst, ksize, anchor, borderType)` is equivalent to `boxFilter(src, dst, src.type(), ksize,
-anchor, true, borderType)`.
-
-@param src input image; it can have any number of channels, which are processed independently, but
-the depth should be CV_8U, CV_16U, CV_16S, CV_32F or CV_64F.
-@param dst output image of the same size and type as src.
-@param ksize blurring kernel size.
-@param anchor anchor point; default value Point(-1,-1) means that the anchor is at the kernel
-center.
-@param borderType border mode used to extrapolate pixels outside of the image, see #BorderTypes. #BORDER_WRAP is not supported.
-@sa  boxFilter, bilateralFilter, GaussianBlur, medianBlur
- */
-CV_EXPORTS_W void blur( InputArray src, OutputArray dst,
-                        Size ksize, Point anchor = Point(-1,-1),
-                        int borderType = BORDER_DEFAULT );
-
-/** @brief Convolves an image with the kernel.
-
-The function applies an arbitrary linear filter to an image. In-place operation is supported. When
-the aperture is partially outside the image, the function interpolates outlier pixel values
-according to the specified border mode.
-
-The function does actually compute correlation, not the convolution:
-
-\f[\texttt{dst} (x,y) =  \sum _{ \substack{0\leq x' < \texttt{kernel.cols}\\{0\leq y' < \texttt{kernel.rows}}}}  \texttt{kernel} (x',y')* \texttt{src} (x+x'- \texttt{anchor.x} ,y+y'- \texttt{anchor.y} )\f]
-
-That is, the kernel is not mirrored around the anchor point. If you need a real convolution, flip
-the kernel using #flip and set the new anchor to `(kernel.cols - anchor.x - 1, kernel.rows -
-anchor.y - 1)`.
-
-The function uses the DFT-based algorithm in case of sufficiently large kernels (~`11 x 11` or
-larger) and the direct algorithm for small kernels.
-
-@param src input image.
-@param dst output image of the same size and the same number of channels as src.
-@param ddepth desired depth of the destination image, see @ref filter_depths "combinations"
-@param kernel convolution kernel (or rather a correlation kernel), a single-channel floating point
-matrix; if you want to apply different kernels to different channels, split the image into
-separate color planes using split and process them individually.
-@param anchor anchor of the kernel that indicates the relative position of a filtered point within
-the kernel; the anchor should lie within the kernel; default value (-1,-1) means that the anchor
-is at the kernel center.
-@param delta optional value added to the filtered pixels before storing them in dst.
-@param borderType pixel extrapolation method, see #BorderTypes. #BORDER_WRAP is not supported.
-@sa  sepFilter2D, dft, matchTemplate
- */
-CV_EXPORTS_W void filter2D( InputArray src, OutputArray dst, int ddepth,
-                            InputArray kernel, Point anchor = Point(-1,-1),
-                            double delta = 0, int borderType = BORDER_DEFAULT );
-
-/** @brief Applies a separable linear filter to an image.
-
-The function applies a separable linear filter to the image. That is, first, every row of src is
-filtered with the 1D kernel kernelX. Then, every column of the result is filtered with the 1D
-kernel kernelY. The final result shifted by delta is stored in dst .
-
-@param src Source image.
-@param dst Destination image of the same size and the same number of channels as src .
-@param ddepth Destination image depth, see @ref filter_depths "combinations"
-@param kernelX Coefficients for filtering each row.
-@param kernelY Coefficients for filtering each column.
-@param anchor Anchor position within the kernel. The default value \f$(-1,-1)\f$ means that the anchor
-is at the kernel center.
-@param delta Value added to the filtered results before storing them.
-@param borderType Pixel extrapolation method, see #BorderTypes. #BORDER_WRAP is not supported.
-@sa  filter2D, Sobel, GaussianBlur, boxFilter, blur
- */
-CV_EXPORTS_W void sepFilter2D( InputArray src, OutputArray dst, int ddepth,
-                               InputArray kernelX, InputArray kernelY,
-                               Point anchor = Point(-1,-1),
-                               double delta = 0, int borderType = BORDER_DEFAULT );
-
-/** @example samples/cpp/tutorial_code/ImgTrans/Sobel_Demo.cpp
-Sample code using Sobel and/or Scharr OpenCV functions to make a simple Edge Detector
-![Sample screenshot](Sobel_Derivatives_Tutorial_Result.jpg)
-Check @ref tutorial_sobel_derivatives "the corresponding tutorial" for more details
-*/
-
-/** @brief Calculates the first, second, third, or mixed image derivatives using an extended Sobel operator.
-
-In all cases except one, the \f$\texttt{ksize} \times \texttt{ksize}\f$ separable kernel is used to
-calculate the derivative. When \f$\texttt{ksize = 1}\f$, the \f$3 \times 1\f$ or \f$1 \times 3\f$
-kernel is used (that is, no Gaussian smoothing is done). `ksize = 1` can only be used for the first
-or the second x- or y- derivatives.
-
-There is also the special value `ksize = #FILTER_SCHARR (-1)` that corresponds to the \f$3\times3\f$ Scharr
-filter that may give more accurate results than the \f$3\times3\f$ Sobel. The Scharr aperture is
-
-\f[\vecthreethree{-3}{0}{3}{-10}{0}{10}{-3}{0}{3}\f]
-
-for the x-derivative, or transposed for the y-derivative.
-
-The function calculates an image derivative by convolving the image with the appropriate kernel:
-
-\f[\texttt{dst} =  \frac{\partial^{xorder+yorder} \texttt{src}}{\partial x^{xorder} \partial y^{yorder}}\f]
-
-The Sobel operators combine Gaussian smoothing and differentiation, so the result is more or less
-resistant to the noise. Most often, the function is called with ( xorder = 1, yorder = 0, ksize = 3)
-or ( xorder = 0, yorder = 1, ksize = 3) to calculate the first x- or y- image derivative. The first
-case corresponds to a kernel of:
-
-\f[\vecthreethree{-1}{0}{1}{-2}{0}{2}{-1}{0}{1}\f]
-
-The second case corresponds to a kernel of:
-
-\f[\vecthreethree{-1}{-2}{-1}{0}{0}{0}{1}{2}{1}\f]
-
-@param src input image.
-@param dst output image of the same size and the same number of channels as src .
-@param ddepth output image depth, see @ref filter_depths "combinations"; in the case of
-    8-bit input images it will result in truncated derivatives.
-@param dx order of the derivative x.
-@param dy order of the derivative y.
-@param ksize size of the extended Sobel kernel; it must be 1, 3, 5, or 7.
-@param scale optional scale factor for the computed derivative values; by default, no scaling is
-applied (see #getDerivKernels for details).
-@param delta optional delta value that is added to the results prior to storing them in dst.
-@param borderType pixel extrapolation method, see #BorderTypes. #BORDER_WRAP is not supported.
-@sa  Scharr, Laplacian, sepFilter2D, filter2D, GaussianBlur, cartToPolar
- */
-CV_EXPORTS_W void Sobel( InputArray src, OutputArray dst, int ddepth,
-                         int dx, int dy, int ksize = 3,
-                         double scale = 1, double delta = 0,
-                         int borderType = BORDER_DEFAULT );
-
-/** @brief Calculates the first order image derivative in both x and y using a Sobel operator
-
-Equivalent to calling:
-
-@code
-Sobel( src, dx, CV_16SC1, 1, 0, 3 );
-Sobel( src, dy, CV_16SC1, 0, 1, 3 );
-@endcode
-
-@param src input image.
-@param dx output image with first-order derivative in x.
-@param dy output image with first-order derivative in y.
-@param ksize size of Sobel kernel. It must be 3.
-@param borderType pixel extrapolation method, see #BorderTypes.
-                  Only #BORDER_DEFAULT=#BORDER_REFLECT_101 and #BORDER_REPLICATE are supported.
-
-@sa Sobel
- */
-
-CV_EXPORTS_W void spatialGradient( InputArray src, OutputArray dx,
-                                   OutputArray dy, int ksize = 3,
-                                   int borderType = BORDER_DEFAULT );
-
-/** @brief Calculates the first x- or y- image derivative using Scharr operator.
-
-The function computes the first x- or y- spatial image derivative using the Scharr operator. The
-call
-
-\f[\texttt{Scharr(src, dst, ddepth, dx, dy, scale, delta, borderType)}\f]
-
-is equivalent to
-
-\f[\texttt{Sobel(src, dst, ddepth, dx, dy, FILTER_SCHARR, scale, delta, borderType)} .\f]
-
-@param src input image.
-@param dst output image of the same size and the same number of channels as src.
-@param ddepth output image depth, see @ref filter_depths "combinations"
-@param dx order of the derivative x.
-@param dy order of the derivative y.
-@param scale optional scale factor for the computed derivative values; by default, no scaling is
-applied (see #getDerivKernels for details).
-@param delta optional delta value that is added to the results prior to storing them in dst.
-@param borderType pixel extrapolation method, see #BorderTypes. #BORDER_WRAP is not supported.
-@sa  cartToPolar
- */
-CV_EXPORTS_W void Scharr( InputArray src, OutputArray dst, int ddepth,
-                          int dx, int dy, double scale = 1, double delta = 0,
-                          int borderType = BORDER_DEFAULT );
-
-/** @example samples/cpp/laplace.cpp
-An example using Laplace transformations for edge detection
-*/
-
-/** @brief Calculates the Laplacian of an image.
-
-The function calculates the Laplacian of the source image by adding up the second x and y
-derivatives calculated using the Sobel operator:
-
-\f[\texttt{dst} =  \Delta \texttt{src} =  \frac{\partial^2 \texttt{src}}{\partial x^2} +  \frac{\partial^2 \texttt{src}}{\partial y^2}\f]
-
-This is done when `ksize > 1`. When `ksize == 1`, the Laplacian is computed by filtering the image
-with the following \f$3 \times 3\f$ aperture:
-
-\f[\vecthreethree {0}{1}{0}{1}{-4}{1}{0}{1}{0}\f]
-
-@param src Source image.
-@param dst Destination image of the same size and the same number of channels as src .
-@param ddepth Desired depth of the destination image.
-@param ksize Aperture size used to compute the second-derivative filters. See #getDerivKernels for
-details. The size must be positive and odd.
-@param scale Optional scale factor for the computed Laplacian values. By default, no scaling is
-applied. See #getDerivKernels for details.
-@param delta Optional delta value that is added to the results prior to storing them in dst .
-@param borderType Pixel extrapolation method, see #BorderTypes. #BORDER_WRAP is not supported.
-@sa  Sobel, Scharr
- */
-CV_EXPORTS_W void Laplacian( InputArray src, OutputArray dst, int ddepth,
-                             int ksize = 1, double scale = 1, double delta = 0,
-                             int borderType = BORDER_DEFAULT );
-
-//! @} imgproc_filter
-
-//! @addtogroup imgproc_feature
-//! @{
-
-/** @example samples/cpp/edge.cpp
-This program demonstrates usage of the Canny edge detector
-
-Check @ref tutorial_canny_detector "the corresponding tutorial" for more details
-*/
-
-/** @brief Finds edges in an image using the Canny algorithm @cite Canny86 .
-
-The function finds edges in the input image and marks them in the output map edges using the
-Canny algorithm. The smallest value between threshold1 and threshold2 is used for edge linking. The
-largest value is used to find initial segments of strong edges. See
-<http://en.wikipedia.org/wiki/Canny_edge_detector>
-
-@param image 8-bit input image.
-@param edges output edge map; single channels 8-bit image, which has the same size as image .
-@param threshold1 first threshold for the hysteresis procedure.
-@param threshold2 second threshold for the hysteresis procedure.
-@param apertureSize aperture size for the Sobel operator.
-@param L2gradient a flag, indicating whether a more accurate \f$L_2\f$ norm
-\f$=\sqrt{(dI/dx)^2 + (dI/dy)^2}\f$ should be used to calculate the image gradient magnitude (
-L2gradient=true ), or whether the default \f$L_1\f$ norm \f$=|dI/dx|+|dI/dy|\f$ is enough (
-L2gradient=false ).
- */
-CV_EXPORTS_W void Canny( InputArray image, OutputArray edges,
-                         double threshold1, double threshold2,
-                         int apertureSize = 3, bool L2gradient = false );
-
-/** \overload
-
-Finds edges in an image using the Canny algorithm with custom image gradient.
-
-@param dx 16-bit x derivative of input image (CV_16SC1 or CV_16SC3).
-@param dy 16-bit y derivative of input image (same type as dx).
-@param edges output edge map; single channels 8-bit image, which has the same size as image .
-@param threshold1 first threshold for the hysteresis procedure.
-@param threshold2 second threshold for the hysteresis procedure.
-@param L2gradient a flag, indicating whether a more accurate \f$L_2\f$ norm
-\f$=\sqrt{(dI/dx)^2 + (dI/dy)^2}\f$ should be used to calculate the image gradient magnitude (
-L2gradient=true ), or whether the default \f$L_1\f$ norm \f$=|dI/dx|+|dI/dy|\f$ is enough (
-L2gradient=false ).
- */
-CV_EXPORTS_W void Canny( InputArray dx, InputArray dy,
-                         OutputArray edges,
-                         double threshold1, double threshold2,
-                         bool L2gradient = false );
-
-/** @brief Calculates the minimal eigenvalue of gradient matrices for corner detection.
-
-The function is similar to cornerEigenValsAndVecs but it calculates and stores only the minimal
-eigenvalue of the covariance matrix of derivatives, that is, \f$\min(\lambda_1, \lambda_2)\f$ in terms
-of the formulae in the cornerEigenValsAndVecs description.
-
-@param src Input single-channel 8-bit or floating-point image.
-@param dst Image to store the minimal eigenvalues. It has the type CV_32FC1 and the same size as
-src .
-@param blockSize Neighborhood size (see the details on #cornerEigenValsAndVecs ).
-@param ksize Aperture parameter for the Sobel operator.
-@param borderType Pixel extrapolation method. See #BorderTypes. #BORDER_WRAP is not supported.
- */
-CV_EXPORTS_W void cornerMinEigenVal( InputArray src, OutputArray dst,
-                                     int blockSize, int ksize = 3,
-                                     int borderType = BORDER_DEFAULT );
-
-/** @brief Harris corner detector.
-
-The function runs the Harris corner detector on the image. Similarly to cornerMinEigenVal and
-cornerEigenValsAndVecs , for each pixel \f$(x, y)\f$ it calculates a \f$2\times2\f$ gradient covariance
-matrix \f$M^{(x,y)}\f$ over a \f$\texttt{blockSize} \times \texttt{blockSize}\f$ neighborhood. Then, it
-computes the following characteristic:
-
-\f[\texttt{dst} (x,y) =  \mathrm{det} M^{(x,y)} - k  \cdot \left ( \mathrm{tr} M^{(x,y)} \right )^2\f]
-
-Corners in the image can be found as the local maxima of this response map.
-
-@param src Input single-channel 8-bit or floating-point image.
-@param dst Image to store the Harris detector responses. It has the type CV_32FC1 and the same
-size as src .
-@param blockSize Neighborhood size (see the details on #cornerEigenValsAndVecs ).
-@param ksize Aperture parameter for the Sobel operator.
-@param k Harris detector free parameter. See the formula above.
-@param borderType Pixel extrapolation method. See #BorderTypes. #BORDER_WRAP is not supported.
- */
-CV_EXPORTS_W void cornerHarris( InputArray src, OutputArray dst, int blockSize,
-                                int ksize, double k,
-                                int borderType = BORDER_DEFAULT );
-
-/** @brief Calculates eigenvalues and eigenvectors of image blocks for corner detection.
-
-For every pixel \f$p\f$ , the function cornerEigenValsAndVecs considers a blockSize \f$\times\f$ blockSize
-neighborhood \f$S(p)\f$ . It calculates the covariation matrix of derivatives over the neighborhood as:
-
-\f[M =  \begin{bmatrix} \sum _{S(p)}(dI/dx)^2 &  \sum _{S(p)}dI/dx dI/dy  \\ \sum _{S(p)}dI/dx dI/dy &  \sum _{S(p)}(dI/dy)^2 \end{bmatrix}\f]
-
-where the derivatives are computed using the Sobel operator.
-
-After that, it finds eigenvectors and eigenvalues of \f$M\f$ and stores them in the destination image as
-\f$(\lambda_1, \lambda_2, x_1, y_1, x_2, y_2)\f$ where
-
--   \f$\lambda_1, \lambda_2\f$ are the non-sorted eigenvalues of \f$M\f$
--   \f$x_1, y_1\f$ are the eigenvectors corresponding to \f$\lambda_1\f$
--   \f$x_2, y_2\f$ are the eigenvectors corresponding to \f$\lambda_2\f$
-
-The output of the function can be used for robust edge or corner detection.
-
-@param src Input single-channel 8-bit or floating-point image.
-@param dst Image to store the results. It has the same size as src and the type CV_32FC(6) .
-@param blockSize Neighborhood size (see details below).
-@param ksize Aperture parameter for the Sobel operator.
-@param borderType Pixel extrapolation method. See #BorderTypes. #BORDER_WRAP is not supported.
-
-@sa  cornerMinEigenVal, cornerHarris, preCornerDetect
- */
-CV_EXPORTS_W void cornerEigenValsAndVecs( InputArray src, OutputArray dst,
-                                          int blockSize, int ksize,
-                                          int borderType = BORDER_DEFAULT );
-
-/** @brief Calculates a feature map for corner detection.
-
-The function calculates the complex spatial derivative-based function of the source image
-
-\f[\texttt{dst} = (D_x  \texttt{src} )^2  \cdot D_{yy}  \texttt{src} + (D_y  \texttt{src} )^2  \cdot D_{xx}  \texttt{src} - 2 D_x  \texttt{src} \cdot D_y  \texttt{src} \cdot D_{xy}  \texttt{src}\f]
-
-where \f$D_x\f$,\f$D_y\f$ are the first image derivatives, \f$D_{xx}\f$,\f$D_{yy}\f$ are the second image
-derivatives, and \f$D_{xy}\f$ is the mixed derivative.
-
-The corners can be found as local maximums of the functions, as shown below:
-@code
-    Mat corners, dilated_corners;
-    preCornerDetect(image, corners, 3);
-    // dilation with 3x3 rectangular structuring element
-    dilate(corners, dilated_corners, Mat(), 1);
-    Mat corner_mask = corners == dilated_corners;
-@endcode
-
-@param src Source single-channel 8-bit of floating-point image.
-@param dst Output image that has the type CV_32F and the same size as src .
-@param ksize %Aperture size of the Sobel .
-@param borderType Pixel extrapolation method. See #BorderTypes. #BORDER_WRAP is not supported.
- */
-CV_EXPORTS_W void preCornerDetect( InputArray src, OutputArray dst, int ksize,
-                                   int borderType = BORDER_DEFAULT );
-
-/** @brief Refines the corner locations.
-
-The function iterates to find the sub-pixel accurate location of corners or radial saddle
-points as described in @cite forstner1987fast, and as shown on the figure below.
-
-![image](pics/cornersubpix.png)
-
-Sub-pixel accurate corner locator is based on the observation that every vector from the center \f$q\f$
-to a point \f$p\f$ located within a neighborhood of \f$q\f$ is orthogonal to the image gradient at \f$p\f$
-subject to image and measurement noise. Consider the expression:
-
-\f[\epsilon _i = {DI_{p_i}}^T  \cdot (q - p_i)\f]
-
-where \f${DI_{p_i}}\f$ is an image gradient at one of the points \f$p_i\f$ in a neighborhood of \f$q\f$ . The
-value of \f$q\f$ is to be found so that \f$\epsilon_i\f$ is minimized. A system of equations may be set up
-with \f$\epsilon_i\f$ set to zero:
-
-\f[\sum _i(DI_{p_i}  \cdot {DI_{p_i}}^T) \cdot q -  \sum _i(DI_{p_i}  \cdot {DI_{p_i}}^T  \cdot p_i)\f]
-
-where the gradients are summed within a neighborhood ("search window") of \f$q\f$ . Calling the first
-gradient term \f$G\f$ and the second gradient term \f$b\f$ gives:
-
-\f[q = G^{-1}  \cdot b\f]
-
-The algorithm sets the center of the neighborhood window at this new center \f$q\f$ and then iterates
-until the center stays within a set threshold.
-
-@param image Input single-channel, 8-bit or float image.
-@param corners Initial coordinates of the input corners and refined coordinates provided for
-output.
-@param winSize Half of the side length of the search window. For example, if winSize=Size(5,5) ,
-then a \f$(5*2+1) \times (5*2+1) = 11 \times 11\f$ search window is used.
-@param zeroZone Half of the size of the dead region in the middle of the search zone over which
-the summation in the formula below is not done. It is used sometimes to avoid possible
-singularities of the autocorrelation matrix. The value of (-1,-1) indicates that there is no such
-a size.
-@param criteria Criteria for termination of the iterative process of corner refinement. That is,
-the process of corner position refinement stops either after criteria.maxCount iterations or when
-the corner position moves by less than criteria.epsilon on some iteration.
- */
-CV_EXPORTS_W void cornerSubPix( InputArray image, InputOutputArray corners,
-                                Size winSize, Size zeroZone,
-                                TermCriteria criteria );
-
-/** @brief Determines strong corners on an image.
-
-The function finds the most prominent corners in the image or in the specified image region, as
-described in @cite Shi94
-
--   Function calculates the corner quality measure at every source image pixel using the
-    #cornerMinEigenVal or #cornerHarris .
--   Function performs a non-maximum suppression (the local maximums in *3 x 3* neighborhood are
-    retained).
--   The corners with the minimal eigenvalue less than
-    \f$\texttt{qualityLevel} \cdot \max_{x,y} qualityMeasureMap(x,y)\f$ are rejected.
--   The remaining corners are sorted by the quality measure in the descending order.
--   Function throws away each corner for which there is a stronger corner at a distance less than
-    maxDistance.
-
-The function can be used to initialize a point-based tracker of an object.
-
-@note If the function is called with different values A and B of the parameter qualityLevel , and
-A \> B, the vector of returned corners with qualityLevel=A will be the prefix of the output vector
-with qualityLevel=B .
-
-@param image Input 8-bit or floating-point 32-bit, single-channel image.
-@param corners Output vector of detected corners.
-@param maxCorners Maximum number of corners to return. If there are more corners than are found,
-the strongest of them is returned. `maxCorners <= 0` implies that no limit on the maximum is set
-and all detected corners are returned.
-@param qualityLevel Parameter characterizing the minimal accepted quality of image corners. The
-parameter value is multiplied by the best corner quality measure, which is the minimal eigenvalue
-(see #cornerMinEigenVal ) or the Harris function response (see #cornerHarris ). The corners with the
-quality measure less than the product are rejected. For example, if the best corner has the
-quality measure = 1500, and the qualityLevel=0.01 , then all the corners with the quality measure
-less than 15 are rejected.
-@param minDistance Minimum possible Euclidean distance between the returned corners.
-@param mask Optional region of interest. If the image is not empty (it needs to have the type
-CV_8UC1 and the same size as image ), it specifies the region in which the corners are detected.
-@param blockSize Size of an average block for computing a derivative covariation matrix over each
-pixel neighborhood. See cornerEigenValsAndVecs .
-@param useHarrisDetector Parameter indicating whether to use a Harris detector (see #cornerHarris)
-or #cornerMinEigenVal.
-@param k Free parameter of the Harris detector.
-
-@sa  cornerMinEigenVal, cornerHarris, calcOpticalFlowPyrLK, estimateRigidTransform,
- */
-
-CV_EXPORTS_W void goodFeaturesToTrack( InputArray image, OutputArray corners,
-                                     int maxCorners, double qualityLevel, double minDistance,
-                                     InputArray mask = noArray(), int blockSize = 3,
-                                     bool useHarrisDetector = false, double k = 0.04 );
-
-CV_EXPORTS_W void goodFeaturesToTrack( InputArray image, OutputArray corners,
-                                     int maxCorners, double qualityLevel, double minDistance,
-                                     InputArray mask, int blockSize,
-                                     int gradientSize, bool useHarrisDetector = false,
-                                     double k = 0.04 );
-
-/** @brief Same as above, but returns also quality measure of the detected corners.
-
-@param image Input 8-bit or floating-point 32-bit, single-channel image.
-@param corners Output vector of detected corners.
-@param maxCorners Maximum number of corners to return. If there are more corners than are found,
-the strongest of them is returned. `maxCorners <= 0` implies that no limit on the maximum is set
-and all detected corners are returned.
-@param qualityLevel Parameter characterizing the minimal accepted quality of image corners. The
-parameter value is multiplied by the best corner quality measure, which is the minimal eigenvalue
-(see #cornerMinEigenVal ) or the Harris function response (see #cornerHarris ). The corners with the
-quality measure less than the product are rejected. For example, if the best corner has the
-quality measure = 1500, and the qualityLevel=0.01 , then all the corners with the quality measure
-less than 15 are rejected.
-@param minDistance Minimum possible Euclidean distance between the returned corners.
-@param mask Region of interest. If the image is not empty (it needs to have the type
-CV_8UC1 and the same size as image ), it specifies the region in which the corners are detected.
-@param cornersQuality Output vector of quality measure of the detected corners.
-@param blockSize Size of an average block for computing a derivative covariation matrix over each
-pixel neighborhood. See cornerEigenValsAndVecs .
-@param gradientSize Aperture parameter for the Sobel operator used for derivatives computation.
-See cornerEigenValsAndVecs .
-@param useHarrisDetector Parameter indicating whether to use a Harris detector (see #cornerHarris)
-or #cornerMinEigenVal.
-@param k Free parameter of the Harris detector.
- */
-CV_EXPORTS CV_WRAP_AS(goodFeaturesToTrackWithQuality) void goodFeaturesToTrack(
-        InputArray image, OutputArray corners,
-        int maxCorners, double qualityLevel, double minDistance,
-        InputArray mask, OutputArray cornersQuality, int blockSize = 3,
-        int gradientSize = 3, bool useHarrisDetector = false, double k = 0.04);
-
-/** @example samples/cpp/tutorial_code/ImgTrans/houghlines.cpp
-An example using the Hough line detector
-![Sample input image](Hough_Lines_Tutorial_Original_Image.jpg) ![Output image](Hough_Lines_Tutorial_Result.jpg)
-*/
-
-/** @brief Finds lines in a binary image using the standard Hough transform.
-
-The function implements the standard or standard multi-scale Hough transform algorithm for line
-detection. See <http://homepages.inf.ed.ac.uk/rbf/HIPR2/hough.htm> for a good explanation of Hough
-transform.
-
-@param image 8-bit, single-channel binary source image. The image may be modified by the function.
-@param lines Output vector of lines. Each line is represented by a 2 or 3 element vector
-\f$(\rho, \theta)\f$ or \f$(\rho, \theta, \textrm{votes})\f$ . \f$\rho\f$ is the distance from the coordinate origin \f$(0,0)\f$ (top-left corner of
-the image). \f$\theta\f$ is the line rotation angle in radians (
-\f$0 \sim \textrm{vertical line}, \pi/2 \sim \textrm{horizontal line}\f$ ).
-\f$\textrm{votes}\f$ is the value of accumulator.
-@param rho Distance resolution of the accumulator in pixels.
-@param theta Angle resolution of the accumulator in radians.
-@param threshold Accumulator threshold parameter. Only those lines are returned that get enough
-votes ( \f$>\texttt{threshold}\f$ ).
-@param srn For the multi-scale Hough transform, it is a divisor for the distance resolution rho .
-The coarse accumulator distance resolution is rho and the accurate accumulator resolution is
-rho/srn . If both srn=0 and stn=0 , the classical Hough transform is used. Otherwise, both these
-parameters should be positive.
-@param stn For the multi-scale Hough transform, it is a divisor for the distance resolution theta.
-@param min_theta For standard and multi-scale Hough transform, minimum angle to check for lines.
-Must fall between 0 and max_theta.
-@param max_theta For standard and multi-scale Hough transform, maximum angle to check for lines.
-Must fall between min_theta and CV_PI.
- */
-CV_EXPORTS_W void HoughLines( InputArray image, OutputArray lines,
-                              double rho, double theta, int threshold,
-                              double srn = 0, double stn = 0,
-                              double min_theta = 0, double max_theta = CV_PI );
-
-/** @brief Finds line segments in a binary image using the probabilistic Hough transform.
-
-The function implements the probabilistic Hough transform algorithm for line detection, described
-in @cite Matas00
-
-See the line detection example below:
-@include snippets/imgproc_HoughLinesP.cpp
-This is a sample picture the function parameters have been tuned for:
-
-![image](pics/building.jpg)
-
-And this is the output of the above program in case of the probabilistic Hough transform:
-
-![image](pics/houghp.png)
-
-@param image 8-bit, single-channel binary source image. The image may be modified by the function.
-@param lines Output vector of lines. Each line is represented by a 4-element vector
-\f$(x_1, y_1, x_2, y_2)\f$ , where \f$(x_1,y_1)\f$ and \f$(x_2, y_2)\f$ are the ending points of each detected
-line segment.
-@param rho Distance resolution of the accumulator in pixels.
-@param theta Angle resolution of the accumulator in radians.
-@param threshold Accumulator threshold parameter. Only those lines are returned that get enough
-votes ( \f$>\texttt{threshold}\f$ ).
-@param minLineLength Minimum line length. Line segments shorter than that are rejected.
-@param maxLineGap Maximum allowed gap between points on the same line to link them.
-
-@sa LineSegmentDetector
- */
-CV_EXPORTS_W void HoughLinesP( InputArray image, OutputArray lines,
-                               double rho, double theta, int threshold,
-                               double minLineLength = 0, double maxLineGap = 0 );
-
-/** @brief Finds lines in a set of points using the standard Hough transform.
-
-The function finds lines in a set of points using a modification of the Hough transform.
-@include snippets/imgproc_HoughLinesPointSet.cpp
-@param point Input vector of points. Each vector must be encoded as a Point vector \f$(x,y)\f$. Type must be CV_32FC2 or CV_32SC2.
-@param lines Output vector of found lines. Each vector is encoded as a vector<Vec3d> \f$(votes, rho, theta)\f$.
-The larger the value of 'votes', the higher the reliability of the Hough line.
-@param lines_max Max count of Hough lines.
-@param threshold Accumulator threshold parameter. Only those lines are returned that get enough
-votes ( \f$>\texttt{threshold}\f$ ).
-@param min_rho Minimum value for \f$\rho\f$ for the accumulator (Note: \f$\rho\f$ can be negative. The absolute value \f$|\rho|\f$ is the distance of a line to the origin.).
-@param max_rho Maximum value for \f$\rho\f$ for the accumulator.
-@param rho_step Distance resolution of the accumulator.
-@param min_theta Minimum angle value of the accumulator in radians.
-@param max_theta Maximum angle value of the accumulator in radians.
-@param theta_step Angle resolution of the accumulator in radians.
- */
-CV_EXPORTS_W void HoughLinesPointSet( InputArray point, OutputArray lines, int lines_max, int threshold,
-                                      double min_rho, double max_rho, double rho_step,
-                                      double min_theta, double max_theta, double theta_step );
-
-/** @example samples/cpp/tutorial_code/ImgTrans/houghcircles.cpp
-An example using the Hough circle detector
-*/
-
-/** @brief Finds circles in a grayscale image using the Hough transform.
-
-The function finds circles in a grayscale image using a modification of the Hough transform.
-
-Example: :
-@include snippets/imgproc_HoughLinesCircles.cpp
-
-@note Usually the function detects the centers of circles well. However, it may fail to find correct
-radii. You can assist to the function by specifying the radius range ( minRadius and maxRadius ) if
-you know it. Or, in the case of #HOUGH_GRADIENT method you may set maxRadius to a negative number
-to return centers only without radius search, and find the correct radius using an additional procedure.
-
-It also helps to smooth image a bit unless it's already soft. For example,
-GaussianBlur() with 7x7 kernel and 1.5x1.5 sigma or similar blurring may help.
-
-@param image 8-bit, single-channel, grayscale input image.
-@param circles Output vector of found circles. Each vector is encoded as  3 or 4 element
-floating-point vector \f$(x, y, radius)\f$ or \f$(x, y, radius, votes)\f$ .
-@param method Detection method, see #HoughModes. The available methods are #HOUGH_GRADIENT and #HOUGH_GRADIENT_ALT.
-@param dp Inverse ratio of the accumulator resolution to the image resolution. For example, if
-dp=1 , the accumulator has the same resolution as the input image. If dp=2 , the accumulator has
-half as big width and height. For #HOUGH_GRADIENT_ALT the recommended value is dp=1.5,
-unless some small very circles need to be detected.
-@param minDist Minimum distance between the centers of the detected circles. If the parameter is
-too small, multiple neighbor circles may be falsely detected in addition to a true one. If it is
-too large, some circles may be missed.
-@param param1 First method-specific parameter. In case of #HOUGH_GRADIENT and #HOUGH_GRADIENT_ALT,
-it is the higher threshold of the two passed to the Canny edge detector (the lower one is twice smaller).
-Note that #HOUGH_GRADIENT_ALT uses #Scharr algorithm to compute image derivatives, so the threshold value
-shough normally be higher, such as 300 or normally exposed and contrasty images.
-@param param2 Second method-specific parameter. In case of #HOUGH_GRADIENT, it is the
-accumulator threshold for the circle centers at the detection stage. The smaller it is, the more
-false circles may be detected. Circles, corresponding to the larger accumulator values, will be
-returned first. In the case of #HOUGH_GRADIENT_ALT algorithm, this is the circle "perfectness" measure.
-The closer it to 1, the better shaped circles algorithm selects. In most cases 0.9 should be fine.
-If you want get better detection of small circles, you may decrease it to 0.85, 0.8 or even less.
-But then also try to limit the search range [minRadius, maxRadius] to avoid many false circles.
-@param minRadius Minimum circle radius.
-@param maxRadius Maximum circle radius. If <= 0, uses the maximum image dimension. If < 0, #HOUGH_GRADIENT returns
-centers without finding the radius. #HOUGH_GRADIENT_ALT always computes circle radiuses.
-
-@sa fitEllipse, minEnclosingCircle
- */
-CV_EXPORTS_W void HoughCircles( InputArray image, OutputArray circles,
-                               int method, double dp, double minDist,
-                               double param1 = 100, double param2 = 100,
-                               int minRadius = 0, int maxRadius = 0 );
-
-//! @} imgproc_feature
-
-//! @addtogroup imgproc_filter
-//! @{
-
-/** @example samples/cpp/tutorial_code/ImgProc/Morphology_2.cpp
-Advanced morphology Transformations sample code
-![Sample screenshot](Morphology_2_Tutorial_Result.jpg)
-Check @ref tutorial_opening_closing_hats "the corresponding tutorial" for more details
-*/
-
-/** @brief Erodes an image by using a specific structuring element.
-
-The function erodes the source image using the specified structuring element that determines the
-shape of a pixel neighborhood over which the minimum is taken:
-
-\f[\texttt{dst} (x,y) =  \min _{(x',y'):  \, \texttt{element} (x',y') \ne0 } \texttt{src} (x+x',y+y')\f]
-
-The function supports the in-place mode. Erosion can be applied several ( iterations ) times. In
-case of multi-channel images, each channel is processed independently.
-
-@param src input image; the number of channels can be arbitrary, but the depth should be one of
-CV_8U, CV_16U, CV_16S, CV_32F or CV_64F.
-@param dst output image of the same size and type as src.
-@param kernel structuring element used for erosion; if `element=Mat()`, a `3 x 3` rectangular
-structuring element is used. Kernel can be created using #getStructuringElement.
-@param anchor position of the anchor within the element; default value (-1, -1) means that the
-anchor is at the element center.
-@param iterations number of times erosion is applied.
-@param borderType pixel extrapolation method, see #BorderTypes. #BORDER_WRAP is not supported.
-@param borderValue border value in case of a constant border
-@sa  dilate, morphologyEx, getStructuringElement
- */
-CV_EXPORTS_W void erode( InputArray src, OutputArray dst, InputArray kernel,
-                         Point anchor = Point(-1,-1), int iterations = 1,
-                         int borderType = BORDER_CONSTANT,
-                         const Scalar& borderValue = morphologyDefaultBorderValue() );
-
-/** @example samples/cpp/tutorial_code/ImgProc/Morphology_1.cpp
-Erosion and Dilation sample code
-![Sample Screenshot-Erosion](Morphology_1_Tutorial_Erosion_Result.jpg)![Sample Screenshot-Dilation](Morphology_1_Tutorial_Dilation_Result.jpg)
-Check @ref tutorial_erosion_dilatation "the corresponding tutorial" for more details
-*/
-
-/** @brief Dilates an image by using a specific structuring element.
-
-The function dilates the source image using the specified structuring element that determines the
-shape of a pixel neighborhood over which the maximum is taken:
-\f[\texttt{dst} (x,y) =  \max _{(x',y'):  \, \texttt{element} (x',y') \ne0 } \texttt{src} (x+x',y+y')\f]
-
-The function supports the in-place mode. Dilation can be applied several ( iterations ) times. In
-case of multi-channel images, each channel is processed independently.
-
-@param src input image; the number of channels can be arbitrary, but the depth should be one of
-CV_8U, CV_16U, CV_16S, CV_32F or CV_64F.
-@param dst output image of the same size and type as src.
-@param kernel structuring element used for dilation; if elemenat=Mat(), a 3 x 3 rectangular
-structuring element is used. Kernel can be created using #getStructuringElement
-@param anchor position of the anchor within the element; default value (-1, -1) means that the
-anchor is at the element center.
-@param iterations number of times dilation is applied.
-@param borderType pixel extrapolation method, see #BorderTypes. #BORDER_WRAP is not suported.
-@param borderValue border value in case of a constant border
-@sa  erode, morphologyEx, getStructuringElement
- */
-CV_EXPORTS_W void dilate( InputArray src, OutputArray dst, InputArray kernel,
-                          Point anchor = Point(-1,-1), int iterations = 1,
-                          int borderType = BORDER_CONSTANT,
-                          const Scalar& borderValue = morphologyDefaultBorderValue() );
-
-/** @brief Performs advanced morphological transformations.
-
-The function cv::morphologyEx can perform advanced morphological transformations using an erosion and dilation as
-basic operations.
-
-Any of the operations can be done in-place. In case of multi-channel images, each channel is
-processed independently.
-
-@param src Source image. The number of channels can be arbitrary. The depth should be one of
-CV_8U, CV_16U, CV_16S, CV_32F or CV_64F.
-@param dst Destination image of the same size and type as source image.
-@param op Type of a morphological operation, see #MorphTypes
-@param kernel Structuring element. It can be created using #getStructuringElement.
-@param anchor Anchor position with the kernel. Negative values mean that the anchor is at the
-kernel center.
-@param iterations Number of times erosion and dilation are applied.
-@param borderType Pixel extrapolation method, see #BorderTypes. #BORDER_WRAP is not supported.
-@param borderValue Border value in case of a constant border. The default value has a special
-meaning.
-@sa  dilate, erode, getStructuringElement
-@note The number of iterations is the number of times erosion or dilatation operation will be applied.
-For instance, an opening operation (#MORPH_OPEN) with two iterations is equivalent to apply
-successively: erode -> erode -> dilate -> dilate (and not erode -> dilate -> erode -> dilate).
- */
-CV_EXPORTS_W void morphologyEx( InputArray src, OutputArray dst,
-                                int op, InputArray kernel,
-                                Point anchor = Point(-1,-1), int iterations = 1,
-                                int borderType = BORDER_CONSTANT,
-                                const Scalar& borderValue = morphologyDefaultBorderValue() );
-
-//! @} imgproc_filter
-
-//! @addtogroup imgproc_transform
-//! @{
-
-/** @brief Resizes an image.
-
-The function resize resizes the image src down to or up to the specified size. Note that the
-initial dst type or size are not taken into account. Instead, the size and type are derived from
-the `src`,`dsize`,`fx`, and `fy`. If you want to resize src so that it fits the pre-created dst,
-you may call the function as follows:
-@code
-    // explicitly specify dsize=dst.size(); fx and fy will be computed from that.
-    resize(src, dst, dst.size(), 0, 0, interpolation);
-@endcode
-If you want to decimate the image by factor of 2 in each direction, you can call the function this
-way:
-@code
-    // specify fx and fy and let the function compute the destination image size.
-    resize(src, dst, Size(), 0.5, 0.5, interpolation);
-@endcode
-To shrink an image, it will generally look best with #INTER_AREA interpolation, whereas to
-enlarge an image, it will generally look best with c#INTER_CUBIC (slow) or #INTER_LINEAR
-(faster but still looks OK).
-
-@param src input image.
-@param dst output image; it has the size dsize (when it is non-zero) or the size computed from
-src.size(), fx, and fy; the type of dst is the same as of src.
-@param dsize output image size; if it equals zero (`None` in Python), it is computed as:
- \f[\texttt{dsize = Size(round(fx*src.cols), round(fy*src.rows))}\f]
- Either dsize or both fx and fy must be non-zero.
-@param fx scale factor along the horizontal axis; when it equals 0, it is computed as
-\f[\texttt{(double)dsize.width/src.cols}\f]
-@param fy scale factor along the vertical axis; when it equals 0, it is computed as
-\f[\texttt{(double)dsize.height/src.rows}\f]
-@param interpolation interpolation method, see #InterpolationFlags
-
-@sa  warpAffine, warpPerspective, remap
- */
-CV_EXPORTS_W void resize( InputArray src, OutputArray dst,
-                          Size dsize, double fx = 0, double fy = 0,
-                          int interpolation = INTER_LINEAR );
-
-/** @brief Applies an affine transformation to an image.
-
-The function warpAffine transforms the source image using the specified matrix:
-
-\f[\texttt{dst} (x,y) =  \texttt{src} ( \texttt{M} _{11} x +  \texttt{M} _{12} y +  \texttt{M} _{13}, \texttt{M} _{21} x +  \texttt{M} _{22} y +  \texttt{M} _{23})\f]
-
-when the flag #WARP_INVERSE_MAP is set. Otherwise, the transformation is first inverted
-with #invertAffineTransform and then put in the formula above instead of M. The function cannot
-operate in-place.
-
-@param src input image.
-@param dst output image that has the size dsize and the same type as src .
-@param M \f$2\times 3\f$ transformation matrix.
-@param dsize size of the output image.
-@param flags combination of interpolation methods (see #InterpolationFlags) and the optional
-flag #WARP_INVERSE_MAP that means that M is the inverse transformation (
-\f$\texttt{dst}\rightarrow\texttt{src}\f$ ).
-@param borderMode pixel extrapolation method (see #BorderTypes); when
-borderMode=#BORDER_TRANSPARENT, it means that the pixels in the destination image corresponding to
-the "outliers" in the source image are not modified by the function.
-@param borderValue value used in case of a constant border; by default, it is 0.
-
-@sa  warpPerspective, resize, remap, getRectSubPix, transform
- */
-CV_EXPORTS_W void warpAffine( InputArray src, OutputArray dst,
-                              InputArray M, Size dsize,
-                              int flags = INTER_LINEAR,
-                              int borderMode = BORDER_CONSTANT,
-                              const Scalar& borderValue = Scalar());
-
-/** @example samples/cpp/warpPerspective_demo.cpp
-An example program shows using cv::getPerspectiveTransform and cv::warpPerspective for image warping
-*/
-
-/** @brief Applies a perspective transformation to an image.
-
-The function warpPerspective transforms the source image using the specified matrix:
-
-\f[\texttt{dst} (x,y) =  \texttt{src} \left ( \frac{M_{11} x + M_{12} y + M_{13}}{M_{31} x + M_{32} y + M_{33}} ,
-     \frac{M_{21} x + M_{22} y + M_{23}}{M_{31} x + M_{32} y + M_{33}} \right )\f]
-
-when the flag #WARP_INVERSE_MAP is set. Otherwise, the transformation is first inverted with invert
-and then put in the formula above instead of M. The function cannot operate in-place.
-
-@param src input image.
-@param dst output image that has the size dsize and the same type as src .
-@param M \f$3\times 3\f$ transformation matrix.
-@param dsize size of the output image.
-@param flags combination of interpolation methods (#INTER_LINEAR or #INTER_NEAREST) and the
-optional flag #WARP_INVERSE_MAP, that sets M as the inverse transformation (
-\f$\texttt{dst}\rightarrow\texttt{src}\f$ ).
-@param borderMode pixel extrapolation method (#BORDER_CONSTANT or #BORDER_REPLICATE).
-@param borderValue value used in case of a constant border; by default, it equals 0.
-
-@sa  warpAffine, resize, remap, getRectSubPix, perspectiveTransform
- */
-CV_EXPORTS_W void warpPerspective( InputArray src, OutputArray dst,
-                                   InputArray M, Size dsize,
-                                   int flags = INTER_LINEAR,
-                                   int borderMode = BORDER_CONSTANT,
-                                   const Scalar& borderValue = Scalar());
-
-/** @brief Applies a generic geometrical transformation to an image.
-
-The function remap transforms the source image using the specified map:
-
-\f[\texttt{dst} (x,y) =  \texttt{src} (map_x(x,y),map_y(x,y))\f]
-
-where values of pixels with non-integer coordinates are computed using one of available
-interpolation methods. \f$map_x\f$ and \f$map_y\f$ can be encoded as separate floating-point maps
-in \f$map_1\f$ and \f$map_2\f$ respectively, or interleaved floating-point maps of \f$(x,y)\f$ in
-\f$map_1\f$, or fixed-point maps created by using convertMaps. The reason you might want to
-convert from floating to fixed-point representations of a map is that they can yield much faster
-(\~2x) remapping operations. In the converted case, \f$map_1\f$ contains pairs (cvFloor(x),
-cvFloor(y)) and \f$map_2\f$ contains indices in a table of interpolation coefficients.
-
-This function cannot operate in-place.
-
-@param src Source image.
-@param dst Destination image. It has the same size as map1 and the same type as src .
-@param map1 The first map of either (x,y) points or just x values having the type CV_16SC2 ,
-CV_32FC1, or CV_32FC2. See convertMaps for details on converting a floating point
-representation to fixed-point for speed.
-@param map2 The second map of y values having the type CV_16UC1, CV_32FC1, or none (empty map
-if map1 is (x,y) points), respectively.
-@param interpolation Interpolation method (see #InterpolationFlags). The methods #INTER_AREA
-and #INTER_LINEAR_EXACT are not supported by this function.
-@param borderMode Pixel extrapolation method (see #BorderTypes). When
-borderMode=#BORDER_TRANSPARENT, it means that the pixels in the destination image that
-corresponds to the "outliers" in the source image are not modified by the function.
-@param borderValue Value used in case of a constant border. By default, it is 0.
-@note
-Due to current implementation limitations the size of an input and output images should be less than 32767x32767.
- */
-CV_EXPORTS_W void remap( InputArray src, OutputArray dst,
-                         InputArray map1, InputArray map2,
-                         int interpolation, int borderMode = BORDER_CONSTANT,
-                         const Scalar& borderValue = Scalar());
-
-/** @brief Converts image transformation maps from one representation to another.
-
-The function converts a pair of maps for remap from one representation to another. The following
-options ( (map1.type(), map2.type()) \f$\rightarrow\f$ (dstmap1.type(), dstmap2.type()) ) are
-supported:
-
-- \f$\texttt{(CV_32FC1, CV_32FC1)} \rightarrow \texttt{(CV_16SC2, CV_16UC1)}\f$. This is the
-most frequently used conversion operation, in which the original floating-point maps (see remap )
-are converted to a more compact and much faster fixed-point representation. The first output array
-contains the rounded coordinates and the second array (created only when nninterpolation=false )
-contains indices in the interpolation tables.
-
-- \f$\texttt{(CV_32FC2)} \rightarrow \texttt{(CV_16SC2, CV_16UC1)}\f$. The same as above but
-the original maps are stored in one 2-channel matrix.
-
-- Reverse conversion. Obviously, the reconstructed floating-point maps will not be exactly the same
-as the originals.
-
-@param map1 The first input map of type CV_16SC2, CV_32FC1, or CV_32FC2 .
-@param map2 The second input map of type CV_16UC1, CV_32FC1, or none (empty matrix),
-respectively.
-@param dstmap1 The first output map that has the type dstmap1type and the same size as src .
-@param dstmap2 The second output map.
-@param dstmap1type Type of the first output map that should be CV_16SC2, CV_32FC1, or
-CV_32FC2 .
-@param nninterpolation Flag indicating whether the fixed-point maps are used for the
-nearest-neighbor or for a more complex interpolation.
-
-@sa  remap, undistort, initUndistortRectifyMap
- */
-CV_EXPORTS_W void convertMaps( InputArray map1, InputArray map2,
-                               OutputArray dstmap1, OutputArray dstmap2,
-                               int dstmap1type, bool nninterpolation = false );
-
-/** @brief Calculates an affine matrix of 2D rotation.
-
-The function calculates the following matrix:
-
-\f[\begin{bmatrix} \alpha &  \beta & (1- \alpha )  \cdot \texttt{center.x} -  \beta \cdot \texttt{center.y} \\ - \beta &  \alpha &  \beta \cdot \texttt{center.x} + (1- \alpha )  \cdot \texttt{center.y} \end{bmatrix}\f]
-
-where
-
-\f[\begin{array}{l} \alpha =  \texttt{scale} \cdot \cos \texttt{angle} , \\ \beta =  \texttt{scale} \cdot \sin \texttt{angle} \end{array}\f]
-
-The transformation maps the rotation center to itself. If this is not the target, adjust the shift.
-
-@param center Center of the rotation in the source image.
-@param angle Rotation angle in degrees. Positive values mean counter-clockwise rotation (the
-coordinate origin is assumed to be the top-left corner).
-@param scale Isotropic scale factor.
-
-@sa  getAffineTransform, warpAffine, transform
- */
-CV_EXPORTS_W Mat getRotationMatrix2D(Point2f center, double angle, double scale);
-
-/** @sa getRotationMatrix2D */
-CV_EXPORTS Matx23d getRotationMatrix2D_(Point2f center, double angle, double scale);
-
-inline
-Mat getRotationMatrix2D(Point2f center, double angle, double scale)
-{
-    return Mat(getRotationMatrix2D_(center, angle, scale), true);
-}
-
-/** @brief Calculates an affine transform from three pairs of the corresponding points.
-
-The function calculates the \f$2 \times 3\f$ matrix of an affine transform so that:
-
-\f[\begin{bmatrix} x'_i \\ y'_i \end{bmatrix} = \texttt{map_matrix} \cdot \begin{bmatrix} x_i \\ y_i \\ 1 \end{bmatrix}\f]
-
-where
-
-\f[dst(i)=(x'_i,y'_i), src(i)=(x_i, y_i), i=0,1,2\f]
-
-@param src Coordinates of triangle vertices in the source image.
-@param dst Coordinates of the corresponding triangle vertices in the destination image.
-
-@sa  warpAffine, transform
- */
-CV_EXPORTS Mat getAffineTransform( const Point2f src[], const Point2f dst[] );
-
-/** @brief Inverts an affine transformation.
-
-The function computes an inverse affine transformation represented by \f$2 \times 3\f$ matrix M:
-
-\f[\begin{bmatrix} a_{11} & a_{12} & b_1  \\ a_{21} & a_{22} & b_2 \end{bmatrix}\f]
-
-The result is also a \f$2 \times 3\f$ matrix of the same type as M.
-
-@param M Original affine transformation.
-@param iM Output reverse affine transformation.
- */
-CV_EXPORTS_W void invertAffineTransform( InputArray M, OutputArray iM );
-
-/** @brief Calculates a perspective transform from four pairs of the corresponding points.
-
-The function calculates the \f$3 \times 3\f$ matrix of a perspective transform so that:
-
-\f[\begin{bmatrix} t_i x'_i \\ t_i y'_i \\ t_i \end{bmatrix} = \texttt{map_matrix} \cdot \begin{bmatrix} x_i \\ y_i \\ 1 \end{bmatrix}\f]
-
-where
-
-\f[dst(i)=(x'_i,y'_i), src(i)=(x_i, y_i), i=0,1,2,3\f]
-
-@param src Coordinates of quadrangle vertices in the source image.
-@param dst Coordinates of the corresponding quadrangle vertices in the destination image.
-@param solveMethod method passed to cv::solve (#DecompTypes)
-
-@sa  findHomography, warpPerspective, perspectiveTransform
- */
-CV_EXPORTS_W Mat getPerspectiveTransform(InputArray src, InputArray dst, int solveMethod = DECOMP_LU);
-
-/** @overload */
-CV_EXPORTS Mat getPerspectiveTransform(const Point2f src[], const Point2f dst[], int solveMethod = DECOMP_LU);
-
-
-CV_EXPORTS_W Mat getAffineTransform( InputArray src, InputArray dst );
-
-/** @brief Retrieves a pixel rectangle from an image with sub-pixel accuracy.
-
-The function getRectSubPix extracts pixels from src:
-
-\f[patch(x, y) = src(x +  \texttt{center.x} - ( \texttt{dst.cols} -1)*0.5, y +  \texttt{center.y} - ( \texttt{dst.rows} -1)*0.5)\f]
-
-where the values of the pixels at non-integer coordinates are retrieved using bilinear
-interpolation. Every channel of multi-channel images is processed independently. Also
-the image should be a single channel or three channel image. While the center of the
-rectangle must be inside the image, parts of the rectangle may be outside.
-
-@param image Source image.
-@param patchSize Size of the extracted patch.
-@param center Floating point coordinates of the center of the extracted rectangle within the
-source image. The center must be inside the image.
-@param patch Extracted patch that has the size patchSize and the same number of channels as src .
-@param patchType Depth of the extracted pixels. By default, they have the same depth as src .
-
-@sa  warpAffine, warpPerspective
- */
-CV_EXPORTS_W void getRectSubPix( InputArray image, Size patchSize,
-                                 Point2f center, OutputArray patch, int patchType = -1 );
-
-/** @example samples/cpp/polar_transforms.cpp
-An example using the cv::linearPolar and cv::logPolar operations
-*/
-
-/** @brief Remaps an image to semilog-polar coordinates space.
-
-@deprecated This function produces same result as cv::warpPolar(src, dst, src.size(), center, maxRadius, flags+WARP_POLAR_LOG);
-
-@internal
-Transform the source image using the following transformation (See @ref polar_remaps_reference_image "Polar remaps reference image d)"):
-\f[\begin{array}{l}
-  dst( \rho , \phi ) = src(x,y) \\
-  dst.size() \leftarrow src.size()
-\end{array}\f]
-
-where
-\f[\begin{array}{l}
-  I = (dx,dy) = (x - center.x,y - center.y) \\
-  \rho = M \cdot log_e(\texttt{magnitude} (I)) ,\\
-  \phi = Kangle \cdot \texttt{angle} (I) \\
-\end{array}\f]
-
-and
-\f[\begin{array}{l}
-  M = src.cols / log_e(maxRadius) \\
-  Kangle = src.rows / 2\Pi \\
-\end{array}\f]
-
-The function emulates the human "foveal" vision and can be used for fast scale and
-rotation-invariant template matching, for object tracking and so forth.
-@param src Source image
-@param dst Destination image. It will have same size and type as src.
-@param center The transformation center; where the output precision is maximal
-@param M Magnitude scale parameter. It determines the radius of the bounding circle to transform too.
-@param flags A combination of interpolation methods, see #InterpolationFlags
-
-@note
--   The function can not operate in-place.
--   To calculate magnitude and angle in degrees #cartToPolar is used internally thus angles are measured from 0 to 360 with accuracy about 0.3 degrees.
-
-@sa cv::linearPolar
-@endinternal
-*/
-CV_EXPORTS_W void logPolar( InputArray src, OutputArray dst,
-                            Point2f center, double M, int flags );
-
-/** @brief Remaps an image to polar coordinates space.
-
-@deprecated This function produces same result as cv::warpPolar(src, dst, src.size(), center, maxRadius, flags)
-
-@internal
-Transform the source image using the following transformation (See @ref polar_remaps_reference_image "Polar remaps reference image c)"):
-\f[\begin{array}{l}
-  dst( \rho , \phi ) = src(x,y) \\
-  dst.size() \leftarrow src.size()
-\end{array}\f]
-
-where
-\f[\begin{array}{l}
-  I = (dx,dy) = (x - center.x,y - center.y) \\
-  \rho = Kmag \cdot \texttt{magnitude} (I) ,\\
-  \phi = angle \cdot \texttt{angle} (I)
-\end{array}\f]
-
-and
-\f[\begin{array}{l}
-  Kx = src.cols / maxRadius \\
-  Ky = src.rows / 2\Pi
-\end{array}\f]
-
-
-@param src Source image
-@param dst Destination image. It will have same size and type as src.
-@param center The transformation center;
-@param maxRadius The radius of the bounding circle to transform. It determines the inverse magnitude scale parameter too.
-@param flags A combination of interpolation methods, see #InterpolationFlags
-
-@note
--   The function can not operate in-place.
--   To calculate magnitude and angle in degrees #cartToPolar is used internally thus angles are measured from 0 to 360 with accuracy about 0.3 degrees.
-
-@sa cv::logPolar
-@endinternal
-*/
-CV_EXPORTS_W void linearPolar( InputArray src, OutputArray dst,
-                               Point2f center, double maxRadius, int flags );
-
-
-/** \brief Remaps an image to polar or semilog-polar coordinates space
-
-@anchor polar_remaps_reference_image
-![Polar remaps reference](pics/polar_remap_doc.png)
-
-Transform the source image using the following transformation:
-\f[
-dst(\rho , \phi ) = src(x,y)
-\f]
-
-where
-\f[
-\begin{array}{l}
-\vec{I} = (x - center.x, \;y - center.y) \\
-\phi = Kangle \cdot \texttt{angle} (\vec{I}) \\
-\rho = \left\{\begin{matrix}
-Klin \cdot \texttt{magnitude} (\vec{I}) & default \\
-Klog \cdot log_e(\texttt{magnitude} (\vec{I})) & if \; semilog \\
-\end{matrix}\right.
-\end{array}
-\f]
-
-and
-\f[
-\begin{array}{l}
-Kangle = dsize.height / 2\Pi \\
-Klin = dsize.width / maxRadius \\
-Klog = dsize.width / log_e(maxRadius) \\
-\end{array}
-\f]
-
-
-\par Linear vs semilog mapping
-
-Polar mapping can be linear or semi-log. Add one of #WarpPolarMode to `flags` to specify the polar mapping mode.
-
-Linear is the default mode.
-
-The semilog mapping emulates the human "foveal" vision that permit very high acuity on the line of sight (central vision)
-in contrast to peripheral vision where acuity is minor.
-
-\par Option on `dsize`:
-
-- if both values in `dsize <=0 ` (default),
-the destination image will have (almost) same area of source bounding circle:
-\f[\begin{array}{l}
-dsize.area  \leftarrow (maxRadius^2 \cdot \Pi) \\
-dsize.width = \texttt{cvRound}(maxRadius) \\
-dsize.height = \texttt{cvRound}(maxRadius \cdot \Pi) \\
-\end{array}\f]
-
-
-- if only `dsize.height <= 0`,
-the destination image area will be proportional to the bounding circle area but scaled by `Kx * Kx`:
-\f[\begin{array}{l}
-dsize.height = \texttt{cvRound}(dsize.width \cdot \Pi) \\
-\end{array}
-\f]
-
-- if both values in `dsize > 0 `,
-the destination image will have the given size therefore the area of the bounding circle will be scaled to `dsize`.
-
-
-\par Reverse mapping
-
-You can get reverse mapping adding #WARP_INVERSE_MAP to `flags`
-\snippet polar_transforms.cpp InverseMap
-
-In addiction, to calculate the original coordinate from a polar mapped coordinate \f$(rho, phi)->(x, y)\f$:
-\snippet polar_transforms.cpp InverseCoordinate
-
-@param src Source image.
-@param dst Destination image. It will have same type as src.
-@param dsize The destination image size (see description for valid options).
-@param center The transformation center.
-@param maxRadius The radius of the bounding circle to transform. It determines the inverse magnitude scale parameter too.
-@param flags A combination of interpolation methods, #InterpolationFlags + #WarpPolarMode.
-            - Add #WARP_POLAR_LINEAR to select linear polar mapping (default)
-            - Add #WARP_POLAR_LOG to select semilog polar mapping
-            - Add #WARP_INVERSE_MAP for reverse mapping.
-@note
--  The function can not operate in-place.
--  To calculate magnitude and angle in degrees #cartToPolar is used internally thus angles are measured from 0 to 360 with accuracy about 0.3 degrees.
--  This function uses #remap. Due to current implementation limitations the size of an input and output images should be less than 32767x32767.
-
-@sa cv::remap
-*/
-CV_EXPORTS_W void warpPolar(InputArray src, OutputArray dst, Size dsize,
-                            Point2f center, double maxRadius, int flags);
-
-
-//! @} imgproc_transform
-
-//! @addtogroup imgproc_misc
-//! @{
-
-/** @brief Calculates the integral of an image.
-
-The function calculates one or more integral images for the source image as follows:
-
-\f[\texttt{sum} (X,Y) =  \sum _{x<X,y<Y}  \texttt{image} (x,y)\f]
-
-\f[\texttt{sqsum} (X,Y) =  \sum _{x<X,y<Y}  \texttt{image} (x,y)^2\f]
-
-\f[\texttt{tilted} (X,Y) =  \sum _{y<Y,abs(x-X+1) \leq Y-y-1}  \texttt{image} (x,y)\f]
-
-Using these integral images, you can calculate sum, mean, and standard deviation over a specific
-up-right or rotated rectangular region of the image in a constant time, for example:
-
-\f[\sum _{x_1 \leq x < x_2,  \, y_1  \leq y < y_2}  \texttt{image} (x,y) =  \texttt{sum} (x_2,y_2)- \texttt{sum} (x_1,y_2)- \texttt{sum} (x_2,y_1)+ \texttt{sum} (x_1,y_1)\f]
-
-It makes possible to do a fast blurring or fast block correlation with a variable window size, for
-example. In case of multi-channel images, sums for each channel are accumulated independently.
-
-As a practical example, the next figure shows the calculation of the integral of a straight
-rectangle Rect(3,3,3,2) and of a tilted rectangle Rect(5,1,2,3) . The selected pixels in the
-original image are shown, as well as the relative pixels in the integral images sum and tilted .
-
-![integral calculation example](pics/integral.png)
-
-@param src input image as \f$W \times H\f$, 8-bit or floating-point (32f or 64f).
-@param sum integral image as \f$(W+1)\times (H+1)\f$ , 32-bit integer or floating-point (32f or 64f).
-@param sqsum integral image for squared pixel values; it is \f$(W+1)\times (H+1)\f$, double-precision
-floating-point (64f) array.
-@param tilted integral for the image rotated by 45 degrees; it is \f$(W+1)\times (H+1)\f$ array with
-the same data type as sum.
-@param sdepth desired depth of the integral and the tilted integral images, CV_32S, CV_32F, or
-CV_64F.
-@param sqdepth desired depth of the integral image of squared pixel values, CV_32F or CV_64F.
- */
-CV_EXPORTS_AS(integral3) void integral( InputArray src, OutputArray sum,
-                                        OutputArray sqsum, OutputArray tilted,
-                                        int sdepth = -1, int sqdepth = -1 );
-
-/** @overload */
-CV_EXPORTS_W void integral( InputArray src, OutputArray sum, int sdepth = -1 );
-
-/** @overload */
-CV_EXPORTS_AS(integral2) void integral( InputArray src, OutputArray sum,
-                                        OutputArray sqsum, int sdepth = -1, int sqdepth = -1 );
-
-//! @} imgproc_misc
-
-//! @addtogroup imgproc_motion
-//! @{
-
-/** @brief Adds an image to the accumulator image.
-
-The function adds src or some of its elements to dst :
-
-\f[\texttt{dst} (x,y)  \leftarrow \texttt{dst} (x,y) +  \texttt{src} (x,y)  \quad \text{if} \quad \texttt{mask} (x,y)  \ne 0\f]
-
-The function supports multi-channel images. Each channel is processed independently.
-
-The function cv::accumulate can be used, for example, to collect statistics of a scene background
-viewed by a still camera and for the further foreground-background segmentation.
-
-@param src Input image of type CV_8UC(n), CV_16UC(n), CV_32FC(n) or CV_64FC(n), where n is a positive integer.
-@param dst %Accumulator image with the same number of channels as input image, and a depth of CV_32F or CV_64F.
-@param mask Optional operation mask.
-
-@sa  accumulateSquare, accumulateProduct, accumulateWeighted
- */
-CV_EXPORTS_W void accumulate( InputArray src, InputOutputArray dst,
-                              InputArray mask = noArray() );
-
-/** @brief Adds the square of a source image to the accumulator image.
-
-The function adds the input image src or its selected region, raised to a power of 2, to the
-accumulator dst :
-
-\f[\texttt{dst} (x,y)  \leftarrow \texttt{dst} (x,y) +  \texttt{src} (x,y)^2  \quad \text{if} \quad \texttt{mask} (x,y)  \ne 0\f]
-
-The function supports multi-channel images. Each channel is processed independently.
-
-@param src Input image as 1- or 3-channel, 8-bit or 32-bit floating point.
-@param dst %Accumulator image with the same number of channels as input image, 32-bit or 64-bit
-floating-point.
-@param mask Optional operation mask.
-
-@sa  accumulateSquare, accumulateProduct, accumulateWeighted
- */
-CV_EXPORTS_W void accumulateSquare( InputArray src, InputOutputArray dst,
-                                    InputArray mask = noArray() );
-
-/** @brief Adds the per-element product of two input images to the accumulator image.
-
-The function adds the product of two images or their selected regions to the accumulator dst :
-
-\f[\texttt{dst} (x,y)  \leftarrow \texttt{dst} (x,y) +  \texttt{src1} (x,y)  \cdot \texttt{src2} (x,y)  \quad \text{if} \quad \texttt{mask} (x,y)  \ne 0\f]
-
-The function supports multi-channel images. Each channel is processed independently.
-
-@param src1 First input image, 1- or 3-channel, 8-bit or 32-bit floating point.
-@param src2 Second input image of the same type and the same size as src1 .
-@param dst %Accumulator image with the same number of channels as input images, 32-bit or 64-bit
-floating-point.
-@param mask Optional operation mask.
-
-@sa  accumulate, accumulateSquare, accumulateWeighted
- */
-CV_EXPORTS_W void accumulateProduct( InputArray src1, InputArray src2,
-                                     InputOutputArray dst, InputArray mask=noArray() );
-
-/** @brief Updates a running average.
-
-The function calculates the weighted sum of the input image src and the accumulator dst so that dst
-becomes a running average of a frame sequence:
-
-\f[\texttt{dst} (x,y)  \leftarrow (1- \texttt{alpha} )  \cdot \texttt{dst} (x,y) +  \texttt{alpha} \cdot \texttt{src} (x,y)  \quad \text{if} \quad \texttt{mask} (x,y)  \ne 0\f]
-
-That is, alpha regulates the update speed (how fast the accumulator "forgets" about earlier images).
-The function supports multi-channel images. Each channel is processed independently.
-
-@param src Input image as 1- or 3-channel, 8-bit or 32-bit floating point.
-@param dst %Accumulator image with the same number of channels as input image, 32-bit or 64-bit
-floating-point.
-@param alpha Weight of the input image.
-@param mask Optional operation mask.
-
-@sa  accumulate, accumulateSquare, accumulateProduct
- */
-CV_EXPORTS_W void accumulateWeighted( InputArray src, InputOutputArray dst,
-                                      double alpha, InputArray mask = noArray() );
-
-/** @brief The function is used to detect translational shifts that occur between two images.
-
-The operation takes advantage of the Fourier shift theorem for detecting the translational shift in
-the frequency domain. It can be used for fast image registration as well as motion estimation. For
-more information please see <http://en.wikipedia.org/wiki/Phase_correlation>
-
-Calculates the cross-power spectrum of two supplied source arrays. The arrays are padded if needed
-with getOptimalDFTSize.
-
-The function performs the following equations:
-- First it applies a Hanning window (see <http://en.wikipedia.org/wiki/Hann_function>) to each
-image to remove possible edge effects. This window is cached until the array size changes to speed
-up processing time.
-- Next it computes the forward DFTs of each source array:
-\f[\mathbf{G}_a = \mathcal{F}\{src_1\}, \; \mathbf{G}_b = \mathcal{F}\{src_2\}\f]
-where \f$\mathcal{F}\f$ is the forward DFT.
-- It then computes the cross-power spectrum of each frequency domain array:
-\f[R = \frac{ \mathbf{G}_a \mathbf{G}_b^*}{|\mathbf{G}_a \mathbf{G}_b^*|}\f]
-- Next the cross-correlation is converted back into the time domain via the inverse DFT:
-\f[r = \mathcal{F}^{-1}\{R\}\f]
-- Finally, it computes the peak location and computes a 5x5 weighted centroid around the peak to
-achieve sub-pixel accuracy.
-\f[(\Delta x, \Delta y) = \texttt{weightedCentroid} \{\arg \max_{(x, y)}\{r\}\}\f]
-- If non-zero, the response parameter is computed as the sum of the elements of r within the 5x5
-centroid around the peak location. It is normalized to a maximum of 1 (meaning there is a single
-peak) and will be smaller when there are multiple peaks.
-
-@param src1 Source floating point array (CV_32FC1 or CV_64FC1)
-@param src2 Source floating point array (CV_32FC1 or CV_64FC1)
-@param window Floating point array with windowing coefficients to reduce edge effects (optional).
-@param response Signal power within the 5x5 centroid around the peak, between 0 and 1 (optional).
-@returns detected phase shift (sub-pixel) between the two arrays.
-
-@sa dft, getOptimalDFTSize, idft, mulSpectrums createHanningWindow
- */
-CV_EXPORTS_W Point2d phaseCorrelate(InputArray src1, InputArray src2,
-                                    InputArray window = noArray(), CV_OUT double* response = 0);
-
-/** @brief This function computes a Hanning window coefficients in two dimensions.
-
-See (http://en.wikipedia.org/wiki/Hann_function) and (http://en.wikipedia.org/wiki/Window_function)
-for more information.
-
-An example is shown below:
-@code
-    // create hanning window of size 100x100 and type CV_32F
-    Mat hann;
-    createHanningWindow(hann, Size(100, 100), CV_32F);
-@endcode
-@param dst Destination array to place Hann coefficients in
-@param winSize The window size specifications (both width and height must be > 1)
-@param type Created array type
- */
-CV_EXPORTS_W void createHanningWindow(OutputArray dst, Size winSize, int type);
-
-/** @brief Performs the per-element division of the first Fourier spectrum by the second Fourier spectrum.
-
-The function cv::divSpectrums performs the per-element division of the first array by the second array.
-The arrays are CCS-packed or complex matrices that are results of a real or complex Fourier transform.
-
-@param a first input array.
-@param b second input array of the same size and type as src1 .
-@param c output array of the same size and type as src1 .
-@param flags operation flags; currently, the only supported flag is cv::DFT_ROWS, which indicates that
-each row of src1 and src2 is an independent 1D Fourier spectrum. If you do not want to use this flag, then simply add a `0` as value.
-@param conjB optional flag that conjugates the second input array before the multiplication (true)
-or not (false).
-*/
-CV_EXPORTS_W void divSpectrums(InputArray a, InputArray b, OutputArray c,
-                               int flags, bool conjB = false);
-
-//! @} imgproc_motion
-
-//! @addtogroup imgproc_misc
-//! @{
-
-/** @brief Applies a fixed-level threshold to each array element.
-
-The function applies fixed-level thresholding to a multiple-channel array. The function is typically
-used to get a bi-level (binary) image out of a grayscale image ( #compare could be also used for
-this purpose) or for removing a noise, that is, filtering out pixels with too small or too large
-values. There are several types of thresholding supported by the function. They are determined by
-type parameter.
-
-Also, the special values #THRESH_OTSU or #THRESH_TRIANGLE may be combined with one of the
-above values. In these cases, the function determines the optimal threshold value using the Otsu's
-or Triangle algorithm and uses it instead of the specified thresh.
-
-@note Currently, the Otsu's and Triangle methods are implemented only for 8-bit single-channel images.
-
-@param src input array (multiple-channel, 8-bit or 32-bit floating point).
-@param dst output array of the same size  and type and the same number of channels as src.
-@param thresh threshold value.
-@param maxval maximum value to use with the #THRESH_BINARY and #THRESH_BINARY_INV thresholding
-types.
-@param type thresholding type (see #ThresholdTypes).
-@return the computed threshold value if Otsu's or Triangle methods used.
-
-@sa  adaptiveThreshold, findContours, compare, min, max
- */
-CV_EXPORTS_W double threshold( InputArray src, OutputArray dst,
-                               double thresh, double maxval, int type );
-
-
-/** @brief Applies an adaptive threshold to an array.
-
-The function transforms a grayscale image to a binary image according to the formulae:
--   **THRESH_BINARY**
-    \f[dst(x,y) =  \fork{\texttt{maxValue}}{if \(src(x,y) > T(x,y)\)}{0}{otherwise}\f]
--   **THRESH_BINARY_INV**
-    \f[dst(x,y) =  \fork{0}{if \(src(x,y) > T(x,y)\)}{\texttt{maxValue}}{otherwise}\f]
-where \f$T(x,y)\f$ is a threshold calculated individually for each pixel (see adaptiveMethod parameter).
-
-The function can process the image in-place.
-
-@param src Source 8-bit single-channel image.
-@param dst Destination image of the same size and the same type as src.
-@param maxValue Non-zero value assigned to the pixels for which the condition is satisfied
-@param adaptiveMethod Adaptive thresholding algorithm to use, see #AdaptiveThresholdTypes.
-The #BORDER_REPLICATE | #BORDER_ISOLATED is used to process boundaries.
-@param thresholdType Thresholding type that must be either #THRESH_BINARY or #THRESH_BINARY_INV,
-see #ThresholdTypes.
-@param blockSize Size of a pixel neighborhood that is used to calculate a threshold value for the
-pixel: 3, 5, 7, and so on.
-@param C Constant subtracted from the mean or weighted mean (see the details below). Normally, it
-is positive but may be zero or negative as well.
-
-@sa  threshold, blur, GaussianBlur
- */
-CV_EXPORTS_W void adaptiveThreshold( InputArray src, OutputArray dst,
-                                     double maxValue, int adaptiveMethod,
-                                     int thresholdType, int blockSize, double C );
-
-//! @} imgproc_misc
-
-//! @addtogroup imgproc_filter
-//! @{
-
-/** @example samples/cpp/tutorial_code/ImgProc/Pyramids/Pyramids.cpp
-An example using pyrDown and pyrUp functions
-*/
-
-/** @brief Blurs an image and downsamples it.
-
-By default, size of the output image is computed as `Size((src.cols+1)/2, (src.rows+1)/2)`, but in
-any case, the following conditions should be satisfied:
-
-\f[\begin{array}{l} | \texttt{dstsize.width} *2-src.cols| \leq 2 \\ | \texttt{dstsize.height} *2-src.rows| \leq 2 \end{array}\f]
-
-The function performs the downsampling step of the Gaussian pyramid construction. First, it
-convolves the source image with the kernel:
-
-\f[\frac{1}{256} \begin{bmatrix} 1 & 4 & 6 & 4 & 1  \\ 4 & 16 & 24 & 16 & 4  \\ 6 & 24 & 36 & 24 & 6  \\ 4 & 16 & 24 & 16 & 4  \\ 1 & 4 & 6 & 4 & 1 \end{bmatrix}\f]
-
-Then, it downsamples the image by rejecting even rows and columns.
-
-@param src input image.
-@param dst output image; it has the specified size and the same type as src.
-@param dstsize size of the output image.
-@param borderType Pixel extrapolation method, see #BorderTypes (#BORDER_CONSTANT isn't supported)
- */
-CV_EXPORTS_W void pyrDown( InputArray src, OutputArray dst,
-                           const Size& dstsize = Size(), int borderType = BORDER_DEFAULT );
-
-/** @brief Upsamples an image and then blurs it.
-
-By default, size of the output image is computed as `Size(src.cols\*2, (src.rows\*2)`, but in any
-case, the following conditions should be satisfied:
-
-\f[\begin{array}{l} | \texttt{dstsize.width} -src.cols*2| \leq  ( \texttt{dstsize.width}   \mod  2)  \\ | \texttt{dstsize.height} -src.rows*2| \leq  ( \texttt{dstsize.height}   \mod  2) \end{array}\f]
-
-The function performs the upsampling step of the Gaussian pyramid construction, though it can
-actually be used to construct the Laplacian pyramid. First, it upsamples the source image by
-injecting even zero rows and columns and then convolves the result with the same kernel as in
-pyrDown multiplied by 4.
-
-@param src input image.
-@param dst output image. It has the specified size and the same type as src .
-@param dstsize size of the output image.
-@param borderType Pixel extrapolation method, see #BorderTypes (only #BORDER_DEFAULT is supported)
- */
-CV_EXPORTS_W void pyrUp( InputArray src, OutputArray dst,
-                         const Size& dstsize = Size(), int borderType = BORDER_DEFAULT );
-
-/** @brief Constructs the Gaussian pyramid for an image.
-
-The function constructs a vector of images and builds the Gaussian pyramid by recursively applying
-pyrDown to the previously built pyramid layers, starting from `dst[0]==src`.
-
-@param src Source image. Check pyrDown for the list of supported types.
-@param dst Destination vector of maxlevel+1 images of the same type as src. dst[0] will be the
-same as src. dst[1] is the next pyramid layer, a smoothed and down-sized src, and so on.
-@param maxlevel 0-based index of the last (the smallest) pyramid layer. It must be non-negative.
-@param borderType Pixel extrapolation method, see #BorderTypes (#BORDER_CONSTANT isn't supported)
- */
-CV_EXPORTS void buildPyramid( InputArray src, OutputArrayOfArrays dst,
-                              int maxlevel, int borderType = BORDER_DEFAULT );
-
-//! @} imgproc_filter
-
-//! @addtogroup imgproc_hist
-//! @{
-
-/** @example samples/cpp/demhist.cpp
-An example for creating histograms of an image
-*/
-
-/** @brief Calculates a histogram of a set of arrays.
-
-The function cv::calcHist calculates the histogram of one or more arrays. The elements of a tuple used
-to increment a histogram bin are taken from the corresponding input arrays at the same location. The
-sample below shows how to compute a 2D Hue-Saturation histogram for a color image. :
-@include snippets/imgproc_calcHist.cpp
-
-@param images Source arrays. They all should have the same depth, CV_8U, CV_16U or CV_32F , and the same
-size. Each of them can have an arbitrary number of channels.
-@param nimages Number of source images.
-@param channels List of the dims channels used to compute the histogram. The first array channels
-are numerated from 0 to images[0].channels()-1 , the second array channels are counted from
-images[0].channels() to images[0].channels() + images[1].channels()-1, and so on.
-@param mask Optional mask. If the matrix is not empty, it must be an 8-bit array of the same size
-as images[i] . The non-zero mask elements mark the array elements counted in the histogram.
-@param hist Output histogram, which is a dense or sparse dims -dimensional array.
-@param dims Histogram dimensionality that must be positive and not greater than CV_MAX_DIMS
-(equal to 32 in the current OpenCV version).
-@param histSize Array of histogram sizes in each dimension.
-@param ranges Array of the dims arrays of the histogram bin boundaries in each dimension. When the
-histogram is uniform ( uniform =true), then for each dimension i it is enough to specify the lower
-(inclusive) boundary \f$L_0\f$ of the 0-th histogram bin and the upper (exclusive) boundary
-\f$U_{\texttt{histSize}[i]-1}\f$ for the last histogram bin histSize[i]-1 . That is, in case of a
-uniform histogram each of ranges[i] is an array of 2 elements. When the histogram is not uniform (
-uniform=false ), then each of ranges[i] contains histSize[i]+1 elements:
-\f$L_0, U_0=L_1, U_1=L_2, ..., U_{\texttt{histSize[i]}-2}=L_{\texttt{histSize[i]}-1}, U_{\texttt{histSize[i]}-1}\f$
-. The array elements, that are not between \f$L_0\f$ and \f$U_{\texttt{histSize[i]}-1}\f$ , are not
-counted in the histogram.
-@param uniform Flag indicating whether the histogram is uniform or not (see above).
-@param accumulate Accumulation flag. If it is set, the histogram is not cleared in the beginning
-when it is allocated. This feature enables you to compute a single histogram from several sets of
-arrays, or to update the histogram in time.
-*/
-CV_EXPORTS void calcHist( const Mat* images, int nimages,
-                          const int* channels, InputArray mask,
-                          OutputArray hist, int dims, const int* histSize,
-                          const float** ranges, bool uniform = true, bool accumulate = false );
-
-/** @overload
-
-this variant uses %SparseMat for output
-*/
-CV_EXPORTS void calcHist( const Mat* images, int nimages,
-                          const int* channels, InputArray mask,
-                          SparseMat& hist, int dims,
-                          const int* histSize, const float** ranges,
-                          bool uniform = true, bool accumulate = false );
-
-/** @overload */
-CV_EXPORTS_W void calcHist( InputArrayOfArrays images,
-                            const std::vector<int>& channels,
-                            InputArray mask, OutputArray hist,
-                            const std::vector<int>& histSize,
-                            const std::vector<float>& ranges,
-                            bool accumulate = false );
-
-/** @brief Calculates the back projection of a histogram.
-
-The function cv::calcBackProject calculates the back project of the histogram. That is, similarly to
-#calcHist , at each location (x, y) the function collects the values from the selected channels
-in the input images and finds the corresponding histogram bin. But instead of incrementing it, the
-function reads the bin value, scales it by scale , and stores in backProject(x,y) . In terms of
-statistics, the function computes probability of each element value in respect with the empirical
-probability distribution represented by the histogram. See how, for example, you can find and track
-a bright-colored object in a scene:
-
-- Before tracking, show the object to the camera so that it covers almost the whole frame.
-Calculate a hue histogram. The histogram may have strong maximums, corresponding to the dominant
-colors in the object.
-
-- When tracking, calculate a back projection of a hue plane of each input video frame using that
-pre-computed histogram. Threshold the back projection to suppress weak colors. It may also make
-sense to suppress pixels with non-sufficient color saturation and too dark or too bright pixels.
-
-- Find connected components in the resulting picture and choose, for example, the largest
-component.
-
-This is an approximate algorithm of the CamShift color object tracker.
-
-@param images Source arrays. They all should have the same depth, CV_8U, CV_16U or CV_32F , and the same
-size. Each of them can have an arbitrary number of channels.
-@param nimages Number of source images.
-@param channels The list of channels used to compute the back projection. The number of channels
-must match the histogram dimensionality. The first array channels are numerated from 0 to
-images[0].channels()-1 , the second array channels are counted from images[0].channels() to
-images[0].channels() + images[1].channels()-1, and so on.
-@param hist Input histogram that can be dense or sparse.
-@param backProject Destination back projection array that is a single-channel array of the same
-size and depth as images[0] .
-@param ranges Array of arrays of the histogram bin boundaries in each dimension. See #calcHist .
-@param scale Optional scale factor for the output back projection.
-@param uniform Flag indicating whether the histogram is uniform or not (see above).
-
-@sa calcHist, compareHist
- */
-CV_EXPORTS void calcBackProject( const Mat* images, int nimages,
-                                 const int* channels, InputArray hist,
-                                 OutputArray backProject, const float** ranges,
-                                 double scale = 1, bool uniform = true );
-
-/** @overload */
-CV_EXPORTS void calcBackProject( const Mat* images, int nimages,
-                                 const int* channels, const SparseMat& hist,
-                                 OutputArray backProject, const float** ranges,
-                                 double scale = 1, bool uniform = true );
-
-/** @overload */
-CV_EXPORTS_W void calcBackProject( InputArrayOfArrays images, const std::vector<int>& channels,
-                                   InputArray hist, OutputArray dst,
-                                   const std::vector<float>& ranges,
-                                   double scale );
-
-/** @brief Compares two histograms.
-
-The function cv::compareHist compares two dense or two sparse histograms using the specified method.
-
-The function returns \f$d(H_1, H_2)\f$ .
-
-While the function works well with 1-, 2-, 3-dimensional dense histograms, it may not be suitable
-for high-dimensional sparse histograms. In such histograms, because of aliasing and sampling
-problems, the coordinates of non-zero histogram bins can slightly shift. To compare such histograms
-or more general sparse configurations of weighted points, consider using the #EMD function.
-
-@param H1 First compared histogram.
-@param H2 Second compared histogram of the same size as H1 .
-@param method Comparison method, see #HistCompMethods
- */
-CV_EXPORTS_W double compareHist( InputArray H1, InputArray H2, int method );
-
-/** @overload */
-CV_EXPORTS double compareHist( const SparseMat& H1, const SparseMat& H2, int method );
-
-/** @brief Equalizes the histogram of a grayscale image.
-
-The function equalizes the histogram of the input image using the following algorithm:
-
-- Calculate the histogram \f$H\f$ for src .
-- Normalize the histogram so that the sum of histogram bins is 255.
-- Compute the integral of the histogram:
-\f[H'_i =  \sum _{0  \le j < i} H(j)\f]
-- Transform the image using \f$H'\f$ as a look-up table: \f$\texttt{dst}(x,y) = H'(\texttt{src}(x,y))\f$
-
-The algorithm normalizes the brightness and increases the contrast of the image.
-
-@param src Source 8-bit single channel image.
-@param dst Destination image of the same size and type as src .
- */
-CV_EXPORTS_W void equalizeHist( InputArray src, OutputArray dst );
-
-/** @brief Creates a smart pointer to a cv::CLAHE class and initializes it.
-
-@param clipLimit Threshold for contrast limiting.
-@param tileGridSize Size of grid for histogram equalization. Input image will be divided into
-equally sized rectangular tiles. tileGridSize defines the number of tiles in row and column.
- */
-CV_EXPORTS_W Ptr<CLAHE> createCLAHE(double clipLimit = 40.0, Size tileGridSize = Size(8, 8));
-
-/** @brief Computes the "minimal work" distance between two weighted point configurations.
-
-The function computes the earth mover distance and/or a lower boundary of the distance between the
-two weighted point configurations. One of the applications described in @cite RubnerSept98,
-@cite Rubner2000 is multi-dimensional histogram comparison for image retrieval. EMD is a transportation
-problem that is solved using some modification of a simplex algorithm, thus the complexity is
-exponential in the worst case, though, on average it is much faster. In the case of a real metric
-the lower boundary can be calculated even faster (using linear-time algorithm) and it can be used
-to determine roughly whether the two signatures are far enough so that they cannot relate to the
-same object.
-
-@param signature1 First signature, a \f$\texttt{size1}\times \texttt{dims}+1\f$ floating-point matrix.
-Each row stores the point weight followed by the point coordinates. The matrix is allowed to have
-a single column (weights only) if the user-defined cost matrix is used. The weights must be
-non-negative and have at least one non-zero value.
-@param signature2 Second signature of the same format as signature1 , though the number of rows
-may be different. The total weights may be different. In this case an extra "dummy" point is added
-to either signature1 or signature2. The weights must be non-negative and have at least one non-zero
-value.
-@param distType Used metric. See #DistanceTypes.
-@param cost User-defined \f$\texttt{size1}\times \texttt{size2}\f$ cost matrix. Also, if a cost matrix
-is used, lower boundary lowerBound cannot be calculated because it needs a metric function.
-@param lowerBound Optional input/output parameter: lower boundary of a distance between the two
-signatures that is a distance between mass centers. The lower boundary may not be calculated if
-the user-defined cost matrix is used, the total weights of point configurations are not equal, or
-if the signatures consist of weights only (the signature matrices have a single column). You
-**must** initialize \*lowerBound . If the calculated distance between mass centers is greater or
-equal to \*lowerBound (it means that the signatures are far enough), the function does not
-calculate EMD. In any case \*lowerBound is set to the calculated distance between mass centers on
-return. Thus, if you want to calculate both distance between mass centers and EMD, \*lowerBound
-should be set to 0.
-@param flow Resultant \f$\texttt{size1} \times \texttt{size2}\f$ flow matrix: \f$\texttt{flow}_{i,j}\f$ is
-a flow from \f$i\f$ -th point of signature1 to \f$j\f$ -th point of signature2 .
- */
-CV_EXPORTS float EMD( InputArray signature1, InputArray signature2,
-                      int distType, InputArray cost=noArray(),
-                      float* lowerBound = 0, OutputArray flow = noArray() );
-
-CV_EXPORTS_AS(EMD) float wrapperEMD( InputArray signature1, InputArray signature2,
-                      int distType, InputArray cost=noArray(),
-                      CV_IN_OUT Ptr<float> lowerBound = Ptr<float>(), OutputArray flow = noArray() );
-
-//! @} imgproc_hist
-
-//! @addtogroup imgproc_segmentation
-//! @{
-
-/** @example samples/cpp/watershed.cpp
-An example using the watershed algorithm
-*/
-
-/** @brief Performs a marker-based image segmentation using the watershed algorithm.
-
-The function implements one of the variants of watershed, non-parametric marker-based segmentation
-algorithm, described in @cite Meyer92 .
-
-Before passing the image to the function, you have to roughly outline the desired regions in the
-image markers with positive (\>0) indices. So, every region is represented as one or more connected
-components with the pixel values 1, 2, 3, and so on. Such markers can be retrieved from a binary
-mask using #findContours and #drawContours (see the watershed.cpp demo). The markers are "seeds" of
-the future image regions. All the other pixels in markers , whose relation to the outlined regions
-is not known and should be defined by the algorithm, should be set to 0's. In the function output,
-each pixel in markers is set to a value of the "seed" components or to -1 at boundaries between the
-regions.
-
-@note Any two neighbor connected components are not necessarily separated by a watershed boundary
-(-1's pixels); for example, they can touch each other in the initial marker image passed to the
-function.
-
-@param image Input 8-bit 3-channel image.
-@param markers Input/output 32-bit single-channel image (map) of markers. It should have the same
-size as image .
-
-@sa findContours
- */
-CV_EXPORTS_W void watershed( InputArray image, InputOutputArray markers );
-
-//! @} imgproc_segmentation
-
-//! @addtogroup imgproc_filter
-//! @{
-
-/** @brief Performs initial step of meanshift segmentation of an image.
-
-The function implements the filtering stage of meanshift segmentation, that is, the output of the
-function is the filtered "posterized" image with color gradients and fine-grain texture flattened.
-At every pixel (X,Y) of the input image (or down-sized input image, see below) the function executes
-meanshift iterations, that is, the pixel (X,Y) neighborhood in the joint space-color hyperspace is
-considered:
-
-\f[(x,y): X- \texttt{sp} \le x  \le X+ \texttt{sp} , Y- \texttt{sp} \le y  \le Y+ \texttt{sp} , ||(R,G,B)-(r,g,b)||   \le \texttt{sr}\f]
-
-where (R,G,B) and (r,g,b) are the vectors of color components at (X,Y) and (x,y), respectively
-(though, the algorithm does not depend on the color space used, so any 3-component color space can
-be used instead). Over the neighborhood the average spatial value (X',Y') and average color vector
-(R',G',B') are found and they act as the neighborhood center on the next iteration:
-
-\f[(X,Y)~(X',Y'), (R,G,B)~(R',G',B').\f]
-
-After the iterations over, the color components of the initial pixel (that is, the pixel from where
-the iterations started) are set to the final value (average color at the last iteration):
-
-\f[I(X,Y) <- (R*,G*,B*)\f]
-
-When maxLevel \> 0, the gaussian pyramid of maxLevel+1 levels is built, and the above procedure is
-run on the smallest layer first. After that, the results are propagated to the larger layer and the
-iterations are run again only on those pixels where the layer colors differ by more than sr from the
-lower-resolution layer of the pyramid. That makes boundaries of color regions sharper. Note that the
-results will be actually different from the ones obtained by running the meanshift procedure on the
-whole original image (i.e. when maxLevel==0).
-
-@param src The source 8-bit, 3-channel image.
-@param dst The destination image of the same format and the same size as the source.
-@param sp The spatial window radius.
-@param sr The color window radius.
-@param maxLevel Maximum level of the pyramid for the segmentation.
-@param termcrit Termination criteria: when to stop meanshift iterations.
- */
-CV_EXPORTS_W void pyrMeanShiftFiltering( InputArray src, OutputArray dst,
-                                         double sp, double sr, int maxLevel = 1,
-                                         TermCriteria termcrit=TermCriteria(TermCriteria::MAX_ITER+TermCriteria::EPS,5,1) );
-
-//! @}
-
-//! @addtogroup imgproc_segmentation
-//! @{
-
-/** @example samples/cpp/grabcut.cpp
-An example using the GrabCut algorithm
-![Sample Screenshot](grabcut_output1.jpg)
-*/
-
-/** @brief Runs the GrabCut algorithm.
-
-The function implements the [GrabCut image segmentation algorithm](http://en.wikipedia.org/wiki/GrabCut).
-
-@param img Input 8-bit 3-channel image.
-@param mask Input/output 8-bit single-channel mask. The mask is initialized by the function when
-mode is set to #GC_INIT_WITH_RECT. Its elements may have one of the #GrabCutClasses.
-@param rect ROI containing a segmented object. The pixels outside of the ROI are marked as
-"obvious background". The parameter is only used when mode==#GC_INIT_WITH_RECT .
-@param bgdModel Temporary array for the background model. Do not modify it while you are
-processing the same image.
-@param fgdModel Temporary arrays for the foreground model. Do not modify it while you are
-processing the same image.
-@param iterCount Number of iterations the algorithm should make before returning the result. Note
-that the result can be refined with further calls with mode==#GC_INIT_WITH_MASK or
-mode==GC_EVAL .
-@param mode Operation mode that could be one of the #GrabCutModes
- */
-CV_EXPORTS_W void grabCut( InputArray img, InputOutputArray mask, Rect rect,
-                           InputOutputArray bgdModel, InputOutputArray fgdModel,
-                           int iterCount, int mode = GC_EVAL );
-
-//! @} imgproc_segmentation
-
-//! @addtogroup imgproc_misc
-//! @{
-
-/** @example samples/cpp/distrans.cpp
-An example on using the distance transform
-*/
-
-/** @brief Calculates the distance to the closest zero pixel for each pixel of the source image.
-
-The function cv::distanceTransform calculates the approximate or precise distance from every binary
-image pixel to the nearest zero pixel. For zero image pixels, the distance will obviously be zero.
-
-When maskSize == #DIST_MASK_PRECISE and distanceType == #DIST_L2 , the function runs the
-algorithm described in @cite Felzenszwalb04 . This algorithm is parallelized with the TBB library.
-
-In other cases, the algorithm @cite Borgefors86 is used. This means that for a pixel the function
-finds the shortest path to the nearest zero pixel consisting of basic shifts: horizontal, vertical,
-diagonal, or knight's move (the latest is available for a \f$5\times 5\f$ mask). The overall
-distance is calculated as a sum of these basic distances. Since the distance function should be
-symmetric, all of the horizontal and vertical shifts must have the same cost (denoted as a ), all
-the diagonal shifts must have the same cost (denoted as `b`), and all knight's moves must have the
-same cost (denoted as `c`). For the #DIST_C and #DIST_L1 types, the distance is calculated
-precisely, whereas for #DIST_L2 (Euclidean distance) the distance can be calculated only with a
-relative error (a \f$5\times 5\f$ mask gives more accurate results). For `a`,`b`, and `c`, OpenCV
-uses the values suggested in the original paper:
-- DIST_L1: `a = 1, b = 2`
-- DIST_L2:
-    - `3 x 3`: `a=0.955, b=1.3693`
-    - `5 x 5`: `a=1, b=1.4, c=2.1969`
-- DIST_C: `a = 1, b = 1`
-
-Typically, for a fast, coarse distance estimation #DIST_L2, a \f$3\times 3\f$ mask is used. For a
-more accurate distance estimation #DIST_L2, a \f$5\times 5\f$ mask or the precise algorithm is used.
-Note that both the precise and the approximate algorithms are linear on the number of pixels.
-
-This variant of the function does not only compute the minimum distance for each pixel \f$(x, y)\f$
-but also identifies the nearest connected component consisting of zero pixels
-(labelType==#DIST_LABEL_CCOMP) or the nearest zero pixel (labelType==#DIST_LABEL_PIXEL). Index of the
-component/pixel is stored in `labels(x, y)`. When labelType==#DIST_LABEL_CCOMP, the function
-automatically finds connected components of zero pixels in the input image and marks them with
-distinct labels. When labelType==#DIST_LABEL_PIXEL, the function scans through the input image and
-marks all the zero pixels with distinct labels.
-
-In this mode, the complexity is still linear. That is, the function provides a very fast way to
-compute the Voronoi diagram for a binary image. Currently, the second variant can use only the
-approximate distance transform algorithm, i.e. maskSize=#DIST_MASK_PRECISE is not supported
-yet.
-
-@param src 8-bit, single-channel (binary) source image.
-@param dst Output image with calculated distances. It is a 8-bit or 32-bit floating-point,
-single-channel image of the same size as src.
-@param labels Output 2D array of labels (the discrete Voronoi diagram). It has the type
-CV_32SC1 and the same size as src.
-@param distanceType Type of distance, see #DistanceTypes
-@param maskSize Size of the distance transform mask, see #DistanceTransformMasks.
-#DIST_MASK_PRECISE is not supported by this variant. In case of the #DIST_L1 or #DIST_C distance type,
-the parameter is forced to 3 because a \f$3\times 3\f$ mask gives the same result as \f$5\times
-5\f$ or any larger aperture.
-@param labelType Type of the label array to build, see #DistanceTransformLabelTypes.
- */
-CV_EXPORTS_AS(distanceTransformWithLabels) void distanceTransform( InputArray src, OutputArray dst,
-                                     OutputArray labels, int distanceType, int maskSize,
-                                     int labelType = DIST_LABEL_CCOMP );
-
-/** @overload
-@param src 8-bit, single-channel (binary) source image.
-@param dst Output image with calculated distances. It is a 8-bit or 32-bit floating-point,
-single-channel image of the same size as src .
-@param distanceType Type of distance, see #DistanceTypes
-@param maskSize Size of the distance transform mask, see #DistanceTransformMasks. In case of the
-#DIST_L1 or #DIST_C distance type, the parameter is forced to 3 because a \f$3\times 3\f$ mask gives
-the same result as \f$5\times 5\f$ or any larger aperture.
-@param dstType Type of output image. It can be CV_8U or CV_32F. Type CV_8U can be used only for
-the first variant of the function and distanceType == #DIST_L1.
-*/
-CV_EXPORTS_W void distanceTransform( InputArray src, OutputArray dst,
-                                     int distanceType, int maskSize, int dstType=CV_32F);
-
-/** @brief Fills a connected component with the given color.
-
-The function cv::floodFill fills a connected component starting from the seed point with the specified
-color. The connectivity is determined by the color/brightness closeness of the neighbor pixels. The
-pixel at \f$(x,y)\f$ is considered to belong to the repainted domain if:
-
-- in case of a grayscale image and floating range
-\f[\texttt{src} (x',y')- \texttt{loDiff} \leq \texttt{src} (x,y)  \leq \texttt{src} (x',y')+ \texttt{upDiff}\f]
-
-
-- in case of a grayscale image and fixed range
-\f[\texttt{src} ( \texttt{seedPoint} .x, \texttt{seedPoint} .y)- \texttt{loDiff} \leq \texttt{src} (x,y)  \leq \texttt{src} ( \texttt{seedPoint} .x, \texttt{seedPoint} .y)+ \texttt{upDiff}\f]
-
-
-- in case of a color image and floating range
-\f[\texttt{src} (x',y')_r- \texttt{loDiff} _r \leq \texttt{src} (x,y)_r \leq \texttt{src} (x',y')_r+ \texttt{upDiff} _r,\f]
-\f[\texttt{src} (x',y')_g- \texttt{loDiff} _g \leq \texttt{src} (x,y)_g \leq \texttt{src} (x',y')_g+ \texttt{upDiff} _g\f]
-and
-\f[\texttt{src} (x',y')_b- \texttt{loDiff} _b \leq \texttt{src} (x,y)_b \leq \texttt{src} (x',y')_b+ \texttt{upDiff} _b\f]
-
-
-- in case of a color image and fixed range
-\f[\texttt{src} ( \texttt{seedPoint} .x, \texttt{seedPoint} .y)_r- \texttt{loDiff} _r \leq \texttt{src} (x,y)_r \leq \texttt{src} ( \texttt{seedPoint} .x, \texttt{seedPoint} .y)_r+ \texttt{upDiff} _r,\f]
-\f[\texttt{src} ( \texttt{seedPoint} .x, \texttt{seedPoint} .y)_g- \texttt{loDiff} _g \leq \texttt{src} (x,y)_g \leq \texttt{src} ( \texttt{seedPoint} .x, \texttt{seedPoint} .y)_g+ \texttt{upDiff} _g\f]
-and
-\f[\texttt{src} ( \texttt{seedPoint} .x, \texttt{seedPoint} .y)_b- \texttt{loDiff} _b \leq \texttt{src} (x,y)_b \leq \texttt{src} ( \texttt{seedPoint} .x, \texttt{seedPoint} .y)_b+ \texttt{upDiff} _b\f]
-
-
-where \f$src(x',y')\f$ is the value of one of pixel neighbors that is already known to belong to the
-component. That is, to be added to the connected component, a color/brightness of the pixel should
-be close enough to:
-- Color/brightness of one of its neighbors that already belong to the connected component in case
-of a floating range.
-- Color/brightness of the seed point in case of a fixed range.
-
-Use these functions to either mark a connected component with the specified color in-place, or build
-a mask and then extract the contour, or copy the region to another image, and so on.
-
-@param image Input/output 1- or 3-channel, 8-bit, or floating-point image. It is modified by the
-function unless the #FLOODFILL_MASK_ONLY flag is set in the second variant of the function. See
-the details below.
-@param mask Operation mask that should be a single-channel 8-bit image, 2 pixels wider and 2 pixels
-taller than image. Since this is both an input and output parameter, you must take responsibility
-of initializing it. Flood-filling cannot go across non-zero pixels in the input mask. For example,
-an edge detector output can be used as a mask to stop filling at edges. On output, pixels in the
-mask corresponding to filled pixels in the image are set to 1 or to the a value specified in flags
-as described below. Additionally, the function fills the border of the mask with ones to simplify
-internal processing. It is therefore possible to use the same mask in multiple calls to the function
-to make sure the filled areas do not overlap.
-@param seedPoint Starting point.
-@param newVal New value of the repainted domain pixels.
-@param loDiff Maximal lower brightness/color difference between the currently observed pixel and
-one of its neighbors belonging to the component, or a seed pixel being added to the component.
-@param upDiff Maximal upper brightness/color difference between the currently observed pixel and
-one of its neighbors belonging to the component, or a seed pixel being added to the component.
-@param rect Optional output parameter set by the function to the minimum bounding rectangle of the
-repainted domain.
-@param flags Operation flags. The first 8 bits contain a connectivity value. The default value of
-4 means that only the four nearest neighbor pixels (those that share an edge) are considered. A
-connectivity value of 8 means that the eight nearest neighbor pixels (those that share a corner)
-will be considered. The next 8 bits (8-16) contain a value between 1 and 255 with which to fill
-the mask (the default value is 1). For example, 4 | ( 255 \<\< 8 ) will consider 4 nearest
-neighbours and fill the mask with a value of 255. The following additional options occupy higher
-bits and therefore may be further combined with the connectivity and mask fill values using
-bit-wise or (|), see #FloodFillFlags.
-
-@note Since the mask is larger than the filled image, a pixel \f$(x, y)\f$ in image corresponds to the
-pixel \f$(x+1, y+1)\f$ in the mask .
-
-@sa findContours
- */
-CV_EXPORTS_W int floodFill( InputOutputArray image, InputOutputArray mask,
-                            Point seedPoint, Scalar newVal, CV_OUT Rect* rect=0,
-                            Scalar loDiff = Scalar(), Scalar upDiff = Scalar(),
-                            int flags = 4 );
-
-/** @example samples/cpp/ffilldemo.cpp
-An example using the FloodFill technique
-*/
-
-/** @overload
-
-variant without `mask` parameter
-*/
-CV_EXPORTS int floodFill( InputOutputArray image,
-                          Point seedPoint, Scalar newVal, CV_OUT Rect* rect = 0,
-                          Scalar loDiff = Scalar(), Scalar upDiff = Scalar(),
-                          int flags = 4 );
-
-//! Performs linear blending of two images:
-//! \f[ \texttt{dst}(i,j) = \texttt{weights1}(i,j)*\texttt{src1}(i,j) + \texttt{weights2}(i,j)*\texttt{src2}(i,j) \f]
-//! @param src1 It has a type of CV_8UC(n) or CV_32FC(n), where n is a positive integer.
-//! @param src2 It has the same type and size as src1.
-//! @param weights1 It has a type of CV_32FC1 and the same size with src1.
-//! @param weights2 It has a type of CV_32FC1 and the same size with src1.
-//! @param dst It is created if it does not have the same size and type with src1.
-CV_EXPORTS_W void blendLinear(InputArray src1, InputArray src2, InputArray weights1, InputArray weights2, OutputArray dst);
-
-//! @} imgproc_misc
-
-//! @addtogroup imgproc_color_conversions
-//! @{
-
-/** @brief Converts an image from one color space to another.
-
-The function converts an input image from one color space to another. In case of a transformation
-to-from RGB color space, the order of the channels should be specified explicitly (RGB or BGR). Note
-that the default color format in OpenCV is often referred to as RGB but it is actually BGR (the
-bytes are reversed). So the first byte in a standard (24-bit) color image will be an 8-bit Blue
-component, the second byte will be Green, and the third byte will be Red. The fourth, fifth, and
-sixth bytes would then be the second pixel (Blue, then Green, then Red), and so on.
-
-The conventional ranges for R, G, and B channel values are:
--   0 to 255 for CV_8U images
--   0 to 65535 for CV_16U images
--   0 to 1 for CV_32F images
-
-In case of linear transformations, the range does not matter. But in case of a non-linear
-transformation, an input RGB image should be normalized to the proper value range to get the correct
-results, for example, for RGB \f$\rightarrow\f$ L\*u\*v\* transformation. For example, if you have a
-32-bit floating-point image directly converted from an 8-bit image without any scaling, then it will
-have the 0..255 value range instead of 0..1 assumed by the function. So, before calling #cvtColor ,
-you need first to scale the image down:
-@code
-    img *= 1./255;
-    cvtColor(img, img, COLOR_BGR2Luv);
-@endcode
-If you use #cvtColor with 8-bit images, the conversion will have some information lost. For many
-applications, this will not be noticeable but it is recommended to use 32-bit images in applications
-that need the full range of colors or that convert an image before an operation and then convert
-back.
-
-If conversion adds the alpha channel, its value will set to the maximum of corresponding channel
-range: 255 for CV_8U, 65535 for CV_16U, 1 for CV_32F.
-
-@param src input image: 8-bit unsigned, 16-bit unsigned ( CV_16UC... ), or single-precision
-floating-point.
-@param dst output image of the same size and depth as src.
-@param code color space conversion code (see #ColorConversionCodes).
-@param dstCn number of channels in the destination image; if the parameter is 0, the number of the
-channels is derived automatically from src and code.
-
-@see @ref imgproc_color_conversions
- */
-CV_EXPORTS_W void cvtColor( InputArray src, OutputArray dst, int code, int dstCn = 0 );
-
-/** @brief Converts an image from one color space to another where the source image is
-stored in two planes.
-
-This function only supports YUV420 to RGB conversion as of now.
-
-@param src1: 8-bit image (#CV_8U) of the Y plane.
-@param src2: image containing interleaved U/V plane.
-@param dst: output image.
-@param code: Specifies the type of conversion. It can take any of the following values:
-- #COLOR_YUV2BGR_NV12
-- #COLOR_YUV2RGB_NV12
-- #COLOR_YUV2BGRA_NV12
-- #COLOR_YUV2RGBA_NV12
-- #COLOR_YUV2BGR_NV21
-- #COLOR_YUV2RGB_NV21
-- #COLOR_YUV2BGRA_NV21
-- #COLOR_YUV2RGBA_NV21
-*/
-CV_EXPORTS_W void cvtColorTwoPlane( InputArray src1, InputArray src2, OutputArray dst, int code );
-
-/** @brief main function for all demosaicing processes
-
-@param src input image: 8-bit unsigned or 16-bit unsigned.
-@param dst output image of the same size and depth as src.
-@param code Color space conversion code (see the description below).
-@param dstCn number of channels in the destination image; if the parameter is 0, the number of the
-channels is derived automatically from src and code.
-
-The function can do the following transformations:
-
--   Demosaicing using bilinear interpolation
-
-    #COLOR_BayerBG2BGR , #COLOR_BayerGB2BGR , #COLOR_BayerRG2BGR , #COLOR_BayerGR2BGR
-
-    #COLOR_BayerBG2GRAY , #COLOR_BayerGB2GRAY , #COLOR_BayerRG2GRAY , #COLOR_BayerGR2GRAY
-
--   Demosaicing using Variable Number of Gradients.
-
-    #COLOR_BayerBG2BGR_VNG , #COLOR_BayerGB2BGR_VNG , #COLOR_BayerRG2BGR_VNG , #COLOR_BayerGR2BGR_VNG
-
--   Edge-Aware Demosaicing.
-
-    #COLOR_BayerBG2BGR_EA , #COLOR_BayerGB2BGR_EA , #COLOR_BayerRG2BGR_EA , #COLOR_BayerGR2BGR_EA
-
--   Demosaicing with alpha channel
-
-    #COLOR_BayerBG2BGRA , #COLOR_BayerGB2BGRA , #COLOR_BayerRG2BGRA , #COLOR_BayerGR2BGRA
-
-@sa cvtColor
-*/
-CV_EXPORTS_W void demosaicing(InputArray src, OutputArray dst, int code, int dstCn = 0);
-
-//! @} imgproc_color_conversions
-
-//! @addtogroup imgproc_shape
-//! @{
-
-/** @brief Calculates all of the moments up to the third order of a polygon or rasterized shape.
-
-The function computes moments, up to the 3rd order, of a vector shape or a rasterized shape. The
-results are returned in the structure cv::Moments.
-
-@param array Raster image (single-channel, 8-bit or floating-point 2D array) or an array (
-\f$1 \times N\f$ or \f$N \times 1\f$ ) of 2D points (Point or Point2f ).
-@param binaryImage If it is true, all non-zero image pixels are treated as 1's. The parameter is
-used for images only.
-@returns moments.
-
-@note Only applicable to contour moments calculations from Python bindings: Note that the numpy
-type for the input array should be either np.int32 or np.float32.
-
-@sa  contourArea, arcLength
- */
-CV_EXPORTS_W Moments moments( InputArray array, bool binaryImage = false );
-
-/** @brief Calculates seven Hu invariants.
-
-The function calculates seven Hu invariants (introduced in @cite Hu62; see also
-<http://en.wikipedia.org/wiki/Image_moment>) defined as:
-
-\f[\begin{array}{l} hu[0]= \eta _{20}+ \eta _{02} \\ hu[1]=( \eta _{20}- \eta _{02})^{2}+4 \eta _{11}^{2} \\ hu[2]=( \eta _{30}-3 \eta _{12})^{2}+ (3 \eta _{21}- \eta _{03})^{2} \\ hu[3]=( \eta _{30}+ \eta _{12})^{2}+ ( \eta _{21}+ \eta _{03})^{2} \\ hu[4]=( \eta _{30}-3 \eta _{12})( \eta _{30}+ \eta _{12})[( \eta _{30}+ \eta _{12})^{2}-3( \eta _{21}+ \eta _{03})^{2}]+(3 \eta _{21}- \eta _{03})( \eta _{21}+ \eta _{03})[3( \eta _{30}+ \eta _{12})^{2}-( \eta _{21}+ \eta _{03})^{2}] \\ hu[5]=( \eta _{20}- \eta _{02})[( \eta _{30}+ \eta _{12})^{2}- ( \eta _{21}+ \eta _{03})^{2}]+4 \eta _{11}( \eta _{30}+ \eta _{12})( \eta _{21}+ \eta _{03}) \\ hu[6]=(3 \eta _{21}- \eta _{03})( \eta _{21}+ \eta _{03})[3( \eta _{30}+ \eta _{12})^{2}-( \eta _{21}+ \eta _{03})^{2}]-( \eta _{30}-3 \eta _{12})( \eta _{21}+ \eta _{03})[3( \eta _{30}+ \eta _{12})^{2}-( \eta _{21}+ \eta _{03})^{2}] \\ \end{array}\f]
-
-where \f$\eta_{ji}\f$ stands for \f$\texttt{Moments::nu}_{ji}\f$ .
-
-These values are proved to be invariants to the image scale, rotation, and reflection except the
-seventh one, whose sign is changed by reflection. This invariance is proved with the assumption of
-infinite image resolution. In case of raster images, the computed Hu invariants for the original and
-transformed images are a bit different.
-
-@param moments Input moments computed with moments .
-@param hu Output Hu invariants.
-
-@sa matchShapes
- */
-CV_EXPORTS void HuMoments( const Moments& moments, double hu[7] );
-
-/** @overload */
-CV_EXPORTS_W void HuMoments( const Moments& m, OutputArray hu );
-
-//! @} imgproc_shape
-
-//! @addtogroup imgproc_object
-//! @{
-
-//! type of the template matching operation
-enum TemplateMatchModes {
-    TM_SQDIFF        = 0, /*!< \f[R(x,y)= \sum _{x',y'} (T(x',y')-I(x+x',y+y'))^2\f]
-                               with mask:
-                               \f[R(x,y)= \sum _{x',y'} \left( (T(x',y')-I(x+x',y+y')) \cdot
-                                  M(x',y') \right)^2\f] */
-    TM_SQDIFF_NORMED = 1, /*!< \f[R(x,y)= \frac{\sum_{x',y'} (T(x',y')-I(x+x',y+y'))^2}{\sqrt{\sum_{
-                                  x',y'}T(x',y')^2 \cdot \sum_{x',y'} I(x+x',y+y')^2}}\f]
-                               with mask:
-                               \f[R(x,y)= \frac{\sum _{x',y'} \left( (T(x',y')-I(x+x',y+y')) \cdot
-                                  M(x',y') \right)^2}{\sqrt{\sum_{x',y'} \left( T(x',y') \cdot
-                                  M(x',y') \right)^2 \cdot \sum_{x',y'} \left( I(x+x',y+y') \cdot
-                                  M(x',y') \right)^2}}\f] */
-    TM_CCORR         = 2, /*!< \f[R(x,y)= \sum _{x',y'} (T(x',y') \cdot I(x+x',y+y'))\f]
-                               with mask:
-                               \f[R(x,y)= \sum _{x',y'} (T(x',y') \cdot I(x+x',y+y') \cdot M(x',y')
-                                  ^2)\f] */
-    TM_CCORR_NORMED  = 3, /*!< \f[R(x,y)= \frac{\sum_{x',y'} (T(x',y') \cdot I(x+x',y+y'))}{\sqrt{
-                                  \sum_{x',y'}T(x',y')^2 \cdot \sum_{x',y'} I(x+x',y+y')^2}}\f]
-                               with mask:
-                               \f[R(x,y)= \frac{\sum_{x',y'} (T(x',y') \cdot I(x+x',y+y') \cdot
-                                  M(x',y')^2)}{\sqrt{\sum_{x',y'} \left( T(x',y') \cdot M(x',y')
-                                  \right)^2 \cdot \sum_{x',y'} \left( I(x+x',y+y') \cdot M(x',y')
-                                  \right)^2}}\f] */
-    TM_CCOEFF        = 4, /*!< \f[R(x,y)= \sum _{x',y'} (T'(x',y') \cdot I'(x+x',y+y'))\f]
-                               where
-                               \f[\begin{array}{l} T'(x',y')=T(x',y') - 1/(w \cdot h) \cdot \sum _{
-                                  x'',y''} T(x'',y'') \\ I'(x+x',y+y')=I(x+x',y+y') - 1/(w \cdot h)
-                                  \cdot \sum _{x'',y''} I(x+x'',y+y'') \end{array}\f]
-                               with mask:
-                               \f[\begin{array}{l} T'(x',y')=M(x',y') \cdot \left( T(x',y') -
-                                  \frac{1}{\sum _{x'',y''} M(x'',y'')} \cdot \sum _{x'',y''}
-                                  (T(x'',y'') \cdot M(x'',y'')) \right) \\ I'(x+x',y+y')=M(x',y')
-                                  \cdot \left( I(x+x',y+y') - \frac{1}{\sum _{x'',y''} M(x'',y'')}
-                                  \cdot \sum _{x'',y''} (I(x+x'',y+y'') \cdot M(x'',y'')) \right)
-                                  \end{array} \f] */
-    TM_CCOEFF_NORMED = 5  /*!< \f[R(x,y)= \frac{ \sum_{x',y'} (T'(x',y') \cdot I'(x+x',y+y')) }{
-                                  \sqrt{\sum_{x',y'}T'(x',y')^2 \cdot \sum_{x',y'} I'(x+x',y+y')^2}
-                                  }\f] */
-};
-
-/** @example samples/cpp/tutorial_code/Histograms_Matching/MatchTemplate_Demo.cpp
-An example using Template Matching algorithm
-*/
-
-/** @brief Compares a template against overlapped image regions.
-
-The function slides through image , compares the overlapped patches of size \f$w \times h\f$ against
-templ using the specified method and stores the comparison results in result . #TemplateMatchModes
-describes the formulae for the available comparison methods ( \f$I\f$ denotes image, \f$T\f$
-template, \f$R\f$ result, \f$M\f$ the optional mask ). The summation is done over template and/or
-the image patch: \f$x' = 0...w-1, y' = 0...h-1\f$
-
-After the function finishes the comparison, the best matches can be found as global minimums (when
-#TM_SQDIFF was used) or maximums (when #TM_CCORR or #TM_CCOEFF was used) using the
-#minMaxLoc function. In case of a color image, template summation in the numerator and each sum in
-the denominator is done over all of the channels and separate mean values are used for each channel.
-That is, the function can take a color template and a color image. The result will still be a
-single-channel image, which is easier to analyze.
-
-@param image Image where the search is running. It must be 8-bit or 32-bit floating-point.
-@param templ Searched template. It must be not greater than the source image and have the same
-data type.
-@param result Map of comparison results. It must be single-channel 32-bit floating-point. If image
-is \f$W \times H\f$ and templ is \f$w \times h\f$ , then result is \f$(W-w+1) \times (H-h+1)\f$ .
-@param method Parameter specifying the comparison method, see #TemplateMatchModes
-@param mask Optional mask. It must have the same size as templ. It must either have the same number
-            of channels as template or only one channel, which is then used for all template and
-            image channels. If the data type is #CV_8U, the mask is interpreted as a binary mask,
-            meaning only elements where mask is nonzero are used and are kept unchanged independent
-            of the actual mask value (weight equals 1). For data tpye #CV_32F, the mask values are
-            used as weights. The exact formulas are documented in #TemplateMatchModes.
- */
-CV_EXPORTS_W void matchTemplate( InputArray image, InputArray templ,
-                                 OutputArray result, int method, InputArray mask = noArray() );
-
-//! @}
-
-//! @addtogroup imgproc_shape
-//! @{
-
-/** @example samples/cpp/connected_components.cpp
-This program demonstrates connected components and use of the trackbar
-*/
-
-/** @brief computes the connected components labeled image of boolean image
-
-image with 4 or 8 way connectivity - returns N, the total number of labels [0, N-1] where 0
-represents the background label. ltype specifies the output label image type, an important
-consideration based on the total number of labels or alternatively the total number of pixels in
-the source image. ccltype specifies the connected components labeling algorithm to use, currently
-Bolelli (Spaghetti) @cite Bolelli2019, Grana (BBDT) @cite Grana2010 and Wu's (SAUF) @cite Wu2009 algorithms
-are supported, see the #ConnectedComponentsAlgorithmsTypes for details. Note that SAUF algorithm forces
-a row major ordering of labels while Spaghetti and BBDT do not.
-This function uses parallel version of the algorithms if at least one allowed
-parallel framework is enabled and if the rows of the image are at least twice the number returned by #getNumberOfCPUs.
-
-@param image the 8-bit single-channel image to be labeled
-@param labels destination labeled image
-@param connectivity 8 or 4 for 8-way or 4-way connectivity respectively
-@param ltype output image label type. Currently CV_32S and CV_16U are supported.
-@param ccltype connected components algorithm type (see the #ConnectedComponentsAlgorithmsTypes).
-*/
-CV_EXPORTS_AS(connectedComponentsWithAlgorithm) int connectedComponents(InputArray image, OutputArray labels,
-                                                                        int connectivity, int ltype, int ccltype);
-
-
-/** @overload
-
-@param image the 8-bit single-channel image to be labeled
-@param labels destination labeled image
-@param connectivity 8 or 4 for 8-way or 4-way connectivity respectively
-@param ltype output image label type. Currently CV_32S and CV_16U are supported.
-*/
-CV_EXPORTS_W int connectedComponents(InputArray image, OutputArray labels,
-                                     int connectivity = 8, int ltype = CV_32S);
-
-
-/** @brief computes the connected components labeled image of boolean image and also produces a statistics output for each label
-
-image with 4 or 8 way connectivity - returns N, the total number of labels [0, N-1] where 0
-represents the background label. ltype specifies the output label image type, an important
-consideration based on the total number of labels or alternatively the total number of pixels in
-the source image. ccltype specifies the connected components labeling algorithm to use, currently
-Bolelli (Spaghetti) @cite Bolelli2019, Grana (BBDT) @cite Grana2010 and Wu's (SAUF) @cite Wu2009 algorithms
-are supported, see the #ConnectedComponentsAlgorithmsTypes for details. Note that SAUF algorithm forces
-a row major ordering of labels while Spaghetti and BBDT do not.
-This function uses parallel version of the algorithms (statistics included) if at least one allowed
-parallel framework is enabled and if the rows of the image are at least twice the number returned by #getNumberOfCPUs.
-
-@param image the 8-bit single-channel image to be labeled
-@param labels destination labeled image
-@param stats statistics output for each label, including the background label.
-Statistics are accessed via stats(label, COLUMN) where COLUMN is one of
-#ConnectedComponentsTypes, selecting the statistic. The data type is CV_32S.
-@param centroids centroid output for each label, including the background label. Centroids are
-accessed via centroids(label, 0) for x and centroids(label, 1) for y. The data type CV_64F.
-@param connectivity 8 or 4 for 8-way or 4-way connectivity respectively
-@param ltype output image label type. Currently CV_32S and CV_16U are supported.
-@param ccltype connected components algorithm type (see #ConnectedComponentsAlgorithmsTypes).
-*/
-CV_EXPORTS_AS(connectedComponentsWithStatsWithAlgorithm) int connectedComponentsWithStats(InputArray image, OutputArray labels,
-                                                                                          OutputArray stats, OutputArray centroids,
-                                                                                          int connectivity, int ltype, int ccltype);
-
-/** @overload
-@param image the 8-bit single-channel image to be labeled
-@param labels destination labeled image
-@param stats statistics output for each label, including the background label.
-Statistics are accessed via stats(label, COLUMN) where COLUMN is one of
-#ConnectedComponentsTypes, selecting the statistic. The data type is CV_32S.
-@param centroids centroid output for each label, including the background label. Centroids are
-accessed via centroids(label, 0) for x and centroids(label, 1) for y. The data type CV_64F.
-@param connectivity 8 or 4 for 8-way or 4-way connectivity respectively
-@param ltype output image label type. Currently CV_32S and CV_16U are supported.
-*/
-CV_EXPORTS_W int connectedComponentsWithStats(InputArray image, OutputArray labels,
-                                              OutputArray stats, OutputArray centroids,
-                                              int connectivity = 8, int ltype = CV_32S);
-
-
-/** @brief Finds contours in a binary image.
-
-The function retrieves contours from the binary image using the algorithm @cite Suzuki85 . The contours
-are a useful tool for shape analysis and object detection and recognition. See squares.cpp in the
-OpenCV sample directory.
-@note Since opencv 3.2 source image is not modified by this function.
-
-@param image Source, an 8-bit single-channel image. Non-zero pixels are treated as 1's. Zero
-pixels remain 0's, so the image is treated as binary . You can use #compare, #inRange, #threshold ,
-#adaptiveThreshold, #Canny, and others to create a binary image out of a grayscale or color one.
-If mode equals to #RETR_CCOMP or #RETR_FLOODFILL, the input can also be a 32-bit integer image of labels (CV_32SC1).
-@param contours Detected contours. Each contour is stored as a vector of points (e.g.
-std::vector<std::vector<cv::Point> >).
-@param hierarchy Optional output vector (e.g. std::vector<cv::Vec4i>), containing information about the image topology. It has
-as many elements as the number of contours. For each i-th contour contours[i], the elements
-hierarchy[i][0] , hierarchy[i][1] , hierarchy[i][2] , and hierarchy[i][3] are set to 0-based indices
-in contours of the next and previous contours at the same hierarchical level, the first child
-contour and the parent contour, respectively. If for the contour i there are no next, previous,
-parent, or nested contours, the corresponding elements of hierarchy[i] will be negative.
-@note In Python, hierarchy is nested inside a top level array. Use hierarchy[0][i] to access hierarchical elements of i-th contour.
-@param mode Contour retrieval mode, see #RetrievalModes
-@param method Contour approximation method, see #ContourApproximationModes
-@param offset Optional offset by which every contour point is shifted. This is useful if the
-contours are extracted from the image ROI and then they should be analyzed in the whole image
-context.
- */
-CV_EXPORTS_W void findContours( InputArray image, OutputArrayOfArrays contours,
-                              OutputArray hierarchy, int mode,
-                              int method, Point offset = Point());
-
-/** @overload */
-CV_EXPORTS void findContours( InputArray image, OutputArrayOfArrays contours,
-                              int mode, int method, Point offset = Point());
-
-/** @example samples/cpp/squares.cpp
-A program using pyramid scaling, Canny, contours and contour simplification to find
-squares in a list of images (pic1-6.png). Returns sequence of squares detected on the image.
-*/
-
-/** @example samples/tapi/squares.cpp
-A program using pyramid scaling, Canny, contours and contour simplification to find
-squares in the input image.
-*/
-
-/** @brief Approximates a polygonal curve(s) with the specified precision.
-
-The function cv::approxPolyDP approximates a curve or a polygon with another curve/polygon with less
-vertices so that the distance between them is less or equal to the specified precision. It uses the
-Douglas-Peucker algorithm <http://en.wikipedia.org/wiki/Ramer-Douglas-Peucker_algorithm>
-
-@param curve Input vector of a 2D point stored in std::vector or Mat
-@param approxCurve Result of the approximation. The type should match the type of the input curve.
-@param epsilon Parameter specifying the approximation accuracy. This is the maximum distance
-between the original curve and its approximation.
-@param closed If true, the approximated curve is closed (its first and last vertices are
-connected). Otherwise, it is not closed.
- */
-CV_EXPORTS_W void approxPolyDP( InputArray curve,
-                                OutputArray approxCurve,
-                                double epsilon, bool closed );
-
-/** @brief Calculates a contour perimeter or a curve length.
-
-The function computes a curve length or a closed contour perimeter.
-
-@param curve Input vector of 2D points, stored in std::vector or Mat.
-@param closed Flag indicating whether the curve is closed or not.
- */
-CV_EXPORTS_W double arcLength( InputArray curve, bool closed );
-
-/** @brief Calculates the up-right bounding rectangle of a point set or non-zero pixels of gray-scale image.
-
-The function calculates and returns the minimal up-right bounding rectangle for the specified point set or
-non-zero pixels of gray-scale image.
-
-@param array Input gray-scale image or 2D point set, stored in std::vector or Mat.
- */
-CV_EXPORTS_W Rect boundingRect( InputArray array );
-
-/** @brief Calculates a contour area.
-
-The function computes a contour area. Similarly to moments , the area is computed using the Green
-formula. Thus, the returned area and the number of non-zero pixels, if you draw the contour using
-#drawContours or #fillPoly , can be different. Also, the function will most certainly give a wrong
-results for contours with self-intersections.
-
-Example:
-@code
-    vector<Point> contour;
-    contour.push_back(Point2f(0, 0));
-    contour.push_back(Point2f(10, 0));
-    contour.push_back(Point2f(10, 10));
-    contour.push_back(Point2f(5, 4));
-
-    double area0 = contourArea(contour);
-    vector<Point> approx;
-    approxPolyDP(contour, approx, 5, true);
-    double area1 = contourArea(approx);
-
-    cout << "area0 =" << area0 << endl <<
-            "area1 =" << area1 << endl <<
-            "approx poly vertices" << approx.size() << endl;
-@endcode
-@param contour Input vector of 2D points (contour vertices), stored in std::vector or Mat.
-@param oriented Oriented area flag. If it is true, the function returns a signed area value,
-depending on the contour orientation (clockwise or counter-clockwise). Using this feature you can
-determine orientation of a contour by taking the sign of an area. By default, the parameter is
-false, which means that the absolute value is returned.
- */
-CV_EXPORTS_W double contourArea( InputArray contour, bool oriented = false );
-
-/** @brief Finds a rotated rectangle of the minimum area enclosing the input 2D point set.
-
-The function calculates and returns the minimum-area bounding rectangle (possibly rotated) for a
-specified point set. Developer should keep in mind that the returned RotatedRect can contain negative
-indices when data is close to the containing Mat element boundary.
-
-@param points Input vector of 2D points, stored in std::vector\<\> or Mat
- */
-CV_EXPORTS_W RotatedRect minAreaRect( InputArray points );
-
-/** @brief Finds the four vertices of a rotated rect. Useful to draw the rotated rectangle.
-
-The function finds the four vertices of a rotated rectangle. This function is useful to draw the
-rectangle. In C++, instead of using this function, you can directly use RotatedRect::points method. Please
-visit the @ref tutorial_bounding_rotated_ellipses "tutorial on Creating Bounding rotated boxes and ellipses for contours" for more information.
-
-@param box The input rotated rectangle. It may be the output of
-@param points The output array of four vertices of rectangles.
- */
-CV_EXPORTS_W void boxPoints(RotatedRect box, OutputArray points);
-
-/** @brief Finds a circle of the minimum area enclosing a 2D point set.
-
-The function finds the minimal enclosing circle of a 2D point set using an iterative algorithm.
-
-@param points Input vector of 2D points, stored in std::vector\<\> or Mat
-@param center Output center of the circle.
-@param radius Output radius of the circle.
- */
-CV_EXPORTS_W void minEnclosingCircle( InputArray points,
-                                      CV_OUT Point2f& center, CV_OUT float& radius );
-
-/** @example samples/cpp/minarea.cpp
-*/
-
-/** @brief Finds a triangle of minimum area enclosing a 2D point set and returns its area.
-
-The function finds a triangle of minimum area enclosing the given set of 2D points and returns its
-area. The output for a given 2D point set is shown in the image below. 2D points are depicted in
-*red* and the enclosing triangle in *yellow*.
-
-![Sample output of the minimum enclosing triangle function](pics/minenclosingtriangle.png)
-
-The implementation of the algorithm is based on O'Rourke's @cite ORourke86 and Klee and Laskowski's
-@cite KleeLaskowski85 papers. O'Rourke provides a \f$\theta(n)\f$ algorithm for finding the minimal
-enclosing triangle of a 2D convex polygon with n vertices. Since the #minEnclosingTriangle function
-takes a 2D point set as input an additional preprocessing step of computing the convex hull of the
-2D point set is required. The complexity of the #convexHull function is \f$O(n log(n))\f$ which is higher
-than \f$\theta(n)\f$. Thus the overall complexity of the function is \f$O(n log(n))\f$.
-
-@param points Input vector of 2D points with depth CV_32S or CV_32F, stored in std::vector\<\> or Mat
-@param triangle Output vector of three 2D points defining the vertices of the triangle. The depth
-of the OutputArray must be CV_32F.
- */
-CV_EXPORTS_W double minEnclosingTriangle( InputArray points, CV_OUT OutputArray triangle );
-
-/** @brief Compares two shapes.
-
-The function compares two shapes. All three implemented methods use the Hu invariants (see #HuMoments)
-
-@param contour1 First contour or grayscale image.
-@param contour2 Second contour or grayscale image.
-@param method Comparison method, see #ShapeMatchModes
-@param parameter Method-specific parameter (not supported now).
- */
-CV_EXPORTS_W double matchShapes( InputArray contour1, InputArray contour2,
-                                 int method, double parameter );
-
-/** @example samples/cpp/convexhull.cpp
-An example using the convexHull functionality
-*/
-
-/** @brief Finds the convex hull of a point set.
-
-The function cv::convexHull finds the convex hull of a 2D point set using the Sklansky's algorithm @cite Sklansky82
-that has *O(N logN)* complexity in the current implementation.
-
-@param points Input 2D point set, stored in std::vector or Mat.
-@param hull Output convex hull. It is either an integer vector of indices or vector of points. In
-the first case, the hull elements are 0-based indices of the convex hull points in the original
-array (since the set of convex hull points is a subset of the original point set). In the second
-case, hull elements are the convex hull points themselves.
-@param clockwise Orientation flag. If it is true, the output convex hull is oriented clockwise.
-Otherwise, it is oriented counter-clockwise. The assumed coordinate system has its X axis pointing
-to the right, and its Y axis pointing upwards.
-@param returnPoints Operation flag. In case of a matrix, when the flag is true, the function
-returns convex hull points. Otherwise, it returns indices of the convex hull points. When the
-output array is std::vector, the flag is ignored, and the output depends on the type of the
-vector: std::vector\<int\> implies returnPoints=false, std::vector\<Point\> implies
-returnPoints=true.
-
-@note `points` and `hull` should be different arrays, inplace processing isn't supported.
-
-Check @ref tutorial_hull "the corresponding tutorial" for more details.
-
-useful links:
-
-https://www.learnopencv.com/convex-hull-using-opencv-in-python-and-c/
- */
-CV_EXPORTS_W void convexHull( InputArray points, OutputArray hull,
-                              bool clockwise = false, bool returnPoints = true );
-
-/** @brief Finds the convexity defects of a contour.
-
-The figure below displays convexity defects of a hand contour:
-
-![image](pics/defects.png)
-
-@param contour Input contour.
-@param convexhull Convex hull obtained using convexHull that should contain indices of the contour
-points that make the hull.
-@param convexityDefects The output vector of convexity defects. In C++ and the new Python/Java
-interface each convexity defect is represented as 4-element integer vector (a.k.a. #Vec4i):
-(start_index, end_index, farthest_pt_index, fixpt_depth), where indices are 0-based indices
-in the original contour of the convexity defect beginning, end and the farthest point, and
-fixpt_depth is fixed-point approximation (with 8 fractional bits) of the distance between the
-farthest contour point and the hull. That is, to get the floating-point value of the depth will be
-fixpt_depth/256.0.
- */
-CV_EXPORTS_W void convexityDefects( InputArray contour, InputArray convexhull, OutputArray convexityDefects );
-
-/** @brief Tests a contour convexity.
-
-The function tests whether the input contour is convex or not. The contour must be simple, that is,
-without self-intersections. Otherwise, the function output is undefined.
-
-@param contour Input vector of 2D points, stored in std::vector\<\> or Mat
- */
-CV_EXPORTS_W bool isContourConvex( InputArray contour );
-
-/** @example samples/cpp/intersectExample.cpp
-Examples of how intersectConvexConvex works
-*/
-
-/** @brief Finds intersection of two convex polygons
-
-@param p1 First polygon
-@param p2 Second polygon
-@param p12 Output polygon describing the intersecting area
-@param handleNested When true, an intersection is found if one of the polygons is fully enclosed in the other.
-When false, no intersection is found. If the polygons share a side or the vertex of one polygon lies on an edge
-of the other, they are not considered nested and an intersection will be found regardless of the value of handleNested.
-
-@returns Absolute value of area of intersecting polygon
-
-@note intersectConvexConvex doesn't confirm that both polygons are convex and will return invalid results if they aren't.
- */
-CV_EXPORTS_W float intersectConvexConvex( InputArray p1, InputArray p2,
-                                          OutputArray p12, bool handleNested = true );
-
-/** @example samples/cpp/fitellipse.cpp
-An example using the fitEllipse technique
-*/
-
-/** @brief Fits an ellipse around a set of 2D points.
-
-The function calculates the ellipse that fits (in a least-squares sense) a set of 2D points best of
-all. It returns the rotated rectangle in which the ellipse is inscribed. The first algorithm described by @cite Fitzgibbon95
-is used. Developer should keep in mind that it is possible that the returned
-ellipse/rotatedRect data contains negative indices, due to the data points being close to the
-border of the containing Mat element.
-
-@param points Input 2D point set, stored in std::vector\<\> or Mat
- */
-CV_EXPORTS_W RotatedRect fitEllipse( InputArray points );
-
-/** @brief Fits an ellipse around a set of 2D points.
-
- The function calculates the ellipse that fits a set of 2D points.
- It returns the rotated rectangle in which the ellipse is inscribed.
- The Approximate Mean Square (AMS) proposed by @cite Taubin1991 is used.
-
- For an ellipse, this basis set is \f$ \chi= \left(x^2, x y, y^2, x, y, 1\right) \f$,
- which is a set of six free coefficients \f$ A^T=\left\{A_{\text{xx}},A_{\text{xy}},A_{\text{yy}},A_x,A_y,A_0\right\} \f$.
- However, to specify an ellipse, all that is needed is five numbers; the major and minor axes lengths \f$ (a,b) \f$,
- the position \f$ (x_0,y_0) \f$, and the orientation \f$ \theta \f$. This is because the basis set includes lines,
- quadratics, parabolic and hyperbolic functions as well as elliptical functions as possible fits.
- If the fit is found to be a parabolic or hyperbolic function then the standard #fitEllipse method is used.
- The AMS method restricts the fit to parabolic, hyperbolic and elliptical curves
- by imposing the condition that \f$ A^T ( D_x^T D_x  +   D_y^T D_y) A = 1 \f$ where
- the matrices \f$ Dx \f$ and \f$ Dy \f$ are the partial derivatives of the design matrix \f$ D \f$ with
- respect to x and y. The matrices are formed row by row applying the following to
- each of the points in the set:
- \f{align*}{
- D(i,:)&=\left\{x_i^2, x_i y_i, y_i^2, x_i, y_i, 1\right\} &
- D_x(i,:)&=\left\{2 x_i,y_i,0,1,0,0\right\} &
- D_y(i,:)&=\left\{0,x_i,2 y_i,0,1,0\right\}
- \f}
- The AMS method minimizes the cost function
- \f{equation*}{
- \epsilon ^2=\frac{ A^T D^T D A }{ A^T (D_x^T D_x +  D_y^T D_y) A^T }
- \f}
-
- The minimum cost is found by solving the generalized eigenvalue problem.
-
- \f{equation*}{
- D^T D A = \lambda  \left( D_x^T D_x +  D_y^T D_y\right) A
- \f}
-
- @param points Input 2D point set, stored in std::vector\<\> or Mat
- */
-CV_EXPORTS_W RotatedRect fitEllipseAMS( InputArray points );
-
-
-/** @brief Fits an ellipse around a set of 2D points.
-
- The function calculates the ellipse that fits a set of 2D points.
- It returns the rotated rectangle in which the ellipse is inscribed.
- The Direct least square (Direct) method by @cite Fitzgibbon1999 is used.
-
- For an ellipse, this basis set is \f$ \chi= \left(x^2, x y, y^2, x, y, 1\right) \f$,
- which is a set of six free coefficients \f$ A^T=\left\{A_{\text{xx}},A_{\text{xy}},A_{\text{yy}},A_x,A_y,A_0\right\} \f$.
- However, to specify an ellipse, all that is needed is five numbers; the major and minor axes lengths \f$ (a,b) \f$,
- the position \f$ (x_0,y_0) \f$, and the orientation \f$ \theta \f$. This is because the basis set includes lines,
- quadratics, parabolic and hyperbolic functions as well as elliptical functions as possible fits.
- The Direct method confines the fit to ellipses by ensuring that \f$ 4 A_{xx} A_{yy}- A_{xy}^2 > 0 \f$.
- The condition imposed is that \f$ 4 A_{xx} A_{yy}- A_{xy}^2=1 \f$ which satisfies the inequality
- and as the coefficients can be arbitrarily scaled is not overly restrictive.
-
- \f{equation*}{
- \epsilon ^2= A^T D^T D A \quad \text{with} \quad A^T C A =1 \quad \text{and} \quad C=\left(\begin{matrix}
- 0 & 0  & 2  & 0  & 0  &  0  \\
- 0 & -1  & 0  & 0  & 0  &  0 \\
- 2 & 0  & 0  & 0  & 0  &  0 \\
- 0 & 0  & 0  & 0  & 0  &  0 \\
- 0 & 0  & 0  & 0  & 0  &  0 \\
- 0 & 0  & 0  & 0  & 0  &  0
- \end{matrix} \right)
- \f}
-
- The minimum cost is found by solving the generalized eigenvalue problem.
-
- \f{equation*}{
- D^T D A = \lambda  \left( C\right) A
- \f}
-
- The system produces only one positive eigenvalue \f$ \lambda\f$ which is chosen as the solution
- with its eigenvector \f$\mathbf{u}\f$. These are used to find the coefficients
-
- \f{equation*}{
- A = \sqrt{\frac{1}{\mathbf{u}^T C \mathbf{u}}}  \mathbf{u}
- \f}
- The scaling factor guarantees that  \f$A^T C A =1\f$.
-
- @param points Input 2D point set, stored in std::vector\<\> or Mat
- */
-CV_EXPORTS_W RotatedRect fitEllipseDirect( InputArray points );
-
-/** @brief Fits a line to a 2D or 3D point set.
-
-The function fitLine fits a line to a 2D or 3D point set by minimizing \f$\sum_i \rho(r_i)\f$ where
-\f$r_i\f$ is a distance between the \f$i^{th}\f$ point, the line and \f$\rho(r)\f$ is a distance function, one
-of the following:
--  DIST_L2
-\f[\rho (r) = r^2/2  \quad \text{(the simplest and the fastest least-squares method)}\f]
-- DIST_L1
-\f[\rho (r) = r\f]
-- DIST_L12
-\f[\rho (r) = 2  \cdot ( \sqrt{1 + \frac{r^2}{2}} - 1)\f]
-- DIST_FAIR
-\f[\rho \left (r \right ) = C^2  \cdot \left (  \frac{r}{C} -  \log{\left(1 + \frac{r}{C}\right)} \right )  \quad \text{where} \quad C=1.3998\f]
-- DIST_WELSCH
-\f[\rho \left (r \right ) =  \frac{C^2}{2} \cdot \left ( 1 -  \exp{\left(-\left(\frac{r}{C}\right)^2\right)} \right )  \quad \text{where} \quad C=2.9846\f]
-- DIST_HUBER
-\f[\rho (r) =  \fork{r^2/2}{if \(r < C\)}{C \cdot (r-C/2)}{otherwise} \quad \text{where} \quad C=1.345\f]
-
-The algorithm is based on the M-estimator ( <http://en.wikipedia.org/wiki/M-estimator> ) technique
-that iteratively fits the line using the weighted least-squares algorithm. After each iteration the
-weights \f$w_i\f$ are adjusted to be inversely proportional to \f$\rho(r_i)\f$ .
-
-@param points Input vector of 2D or 3D points, stored in std::vector\<\> or Mat.
-@param line Output line parameters. In case of 2D fitting, it should be a vector of 4 elements
-(like Vec4f) - (vx, vy, x0, y0), where (vx, vy) is a normalized vector collinear to the line and
-(x0, y0) is a point on the line. In case of 3D fitting, it should be a vector of 6 elements (like
-Vec6f) - (vx, vy, vz, x0, y0, z0), where (vx, vy, vz) is a normalized vector collinear to the line
-and (x0, y0, z0) is a point on the line.
-@param distType Distance used by the M-estimator, see #DistanceTypes
-@param param Numerical parameter ( C ) for some types of distances. If it is 0, an optimal value
-is chosen.
-@param reps Sufficient accuracy for the radius (distance between the coordinate origin and the line).
-@param aeps Sufficient accuracy for the angle. 0.01 would be a good default value for reps and aeps.
- */
-CV_EXPORTS_W void fitLine( InputArray points, OutputArray line, int distType,
-                           double param, double reps, double aeps );
-
-/** @brief Performs a point-in-contour test.
-
-The function determines whether the point is inside a contour, outside, or lies on an edge (or
-coincides with a vertex). It returns positive (inside), negative (outside), or zero (on an edge)
-value, correspondingly. When measureDist=false , the return value is +1, -1, and 0, respectively.
-Otherwise, the return value is a signed distance between the point and the nearest contour edge.
-
-See below a sample output of the function where each image pixel is tested against the contour:
-
-![sample output](pics/pointpolygon.png)
-
-@param contour Input contour.
-@param pt Point tested against the contour.
-@param measureDist If true, the function estimates the signed distance from the point to the
-nearest contour edge. Otherwise, the function only checks if the point is inside a contour or not.
- */
-CV_EXPORTS_W double pointPolygonTest( InputArray contour, Point2f pt, bool measureDist );
-
-/** @brief Finds out if there is any intersection between two rotated rectangles.
-
-If there is then the vertices of the intersecting region are returned as well.
-
-Below are some examples of intersection configurations. The hatched pattern indicates the
-intersecting region and the red vertices are returned by the function.
-
-![intersection examples](pics/intersection.png)
-
-@param rect1 First rectangle
-@param rect2 Second rectangle
-@param intersectingRegion The output array of the vertices of the intersecting region. It returns
-at most 8 vertices. Stored as std::vector\<cv::Point2f\> or cv::Mat as Mx1 of type CV_32FC2.
-@returns One of #RectanglesIntersectTypes
- */
-CV_EXPORTS_W int rotatedRectangleIntersection( const RotatedRect& rect1, const RotatedRect& rect2, OutputArray intersectingRegion  );
-
-/** @brief Creates a smart pointer to a cv::GeneralizedHoughBallard class and initializes it.
-*/
-CV_EXPORTS_W Ptr<GeneralizedHoughBallard> createGeneralizedHoughBallard();
-
-/** @brief Creates a smart pointer to a cv::GeneralizedHoughGuil class and initializes it.
-*/
-CV_EXPORTS_W Ptr<GeneralizedHoughGuil> createGeneralizedHoughGuil();
-
-//! @} imgproc_shape
-
-//! @addtogroup imgproc_colormap
-//! @{
-
-//! GNU Octave/MATLAB equivalent colormaps
-enum ColormapTypes
-{
-    COLORMAP_AUTUMN = 0, //!< ![autumn](pics/colormaps/colorscale_autumn.jpg)
-    COLORMAP_BONE = 1, //!< ![bone](pics/colormaps/colorscale_bone.jpg)
-    COLORMAP_JET = 2, //!< ![jet](pics/colormaps/colorscale_jet.jpg)
-    COLORMAP_WINTER = 3, //!< ![winter](pics/colormaps/colorscale_winter.jpg)
-    COLORMAP_RAINBOW = 4, //!< ![rainbow](pics/colormaps/colorscale_rainbow.jpg)
-    COLORMAP_OCEAN = 5, //!< ![ocean](pics/colormaps/colorscale_ocean.jpg)
-    COLORMAP_SUMMER = 6, //!< ![summer](pics/colormaps/colorscale_summer.jpg)
-    COLORMAP_SPRING = 7, //!< ![spring](pics/colormaps/colorscale_spring.jpg)
-    COLORMAP_COOL = 8, //!< ![cool](pics/colormaps/colorscale_cool.jpg)
-    COLORMAP_HSV = 9, //!< ![HSV](pics/colormaps/colorscale_hsv.jpg)
-    COLORMAP_PINK = 10, //!< ![pink](pics/colormaps/colorscale_pink.jpg)
-    COLORMAP_HOT = 11, //!< ![hot](pics/colormaps/colorscale_hot.jpg)
-    COLORMAP_PARULA = 12, //!< ![parula](pics/colormaps/colorscale_parula.jpg)
-    COLORMAP_MAGMA = 13, //!< ![magma](pics/colormaps/colorscale_magma.jpg)
-    COLORMAP_INFERNO = 14, //!< ![inferno](pics/colormaps/colorscale_inferno.jpg)
-    COLORMAP_PLASMA = 15, //!< ![plasma](pics/colormaps/colorscale_plasma.jpg)
-    COLORMAP_VIRIDIS = 16, //!< ![viridis](pics/colormaps/colorscale_viridis.jpg)
-    COLORMAP_CIVIDIS = 17, //!< ![cividis](pics/colormaps/colorscale_cividis.jpg)
-    COLORMAP_TWILIGHT = 18, //!< ![twilight](pics/colormaps/colorscale_twilight.jpg)
-    COLORMAP_TWILIGHT_SHIFTED = 19, //!< ![twilight shifted](pics/colormaps/colorscale_twilight_shifted.jpg)
-    COLORMAP_TURBO = 20, //!< ![turbo](pics/colormaps/colorscale_turbo.jpg)
-    COLORMAP_DEEPGREEN = 21  //!< ![deepgreen](pics/colormaps/colorscale_deepgreen.jpg)
-};
-
-/** @example samples/cpp/falsecolor.cpp
-An example using applyColorMap function
-*/
-
-/** @brief Applies a GNU Octave/MATLAB equivalent colormap on a given image.
-
-@param src The source image, grayscale or colored of type CV_8UC1 or CV_8UC3.
-@param dst The result is the colormapped source image. Note: Mat::create is called on dst.
-@param colormap The colormap to apply, see #ColormapTypes
-*/
-CV_EXPORTS_W void applyColorMap(InputArray src, OutputArray dst, int colormap);
-
-/** @brief Applies a user colormap on a given image.
-
-@param src The source image, grayscale or colored of type CV_8UC1 or CV_8UC3.
-@param dst The result is the colormapped source image. Note: Mat::create is called on dst.
-@param userColor The colormap to apply of type CV_8UC1 or CV_8UC3 and size 256
-*/
-CV_EXPORTS_W void applyColorMap(InputArray src, OutputArray dst, InputArray userColor);
-
-//! @} imgproc_colormap
-
-//! @addtogroup imgproc_draw
-//! @{
-
-
-/** OpenCV color channel order is BGR[A] */
-#define CV_RGB(r, g, b)  cv::Scalar((b), (g), (r), 0)
-
-/** @brief Draws a line segment connecting two points.
-
-The function line draws the line segment between pt1 and pt2 points in the image. The line is
-clipped by the image boundaries. For non-antialiased lines with integer coordinates, the 8-connected
-or 4-connected Bresenham algorithm is used. Thick lines are drawn with rounding endings. Antialiased
-lines are drawn using Gaussian filtering.
-
-@param img Image.
-@param pt1 First point of the line segment.
-@param pt2 Second point of the line segment.
-@param color Line color.
-@param thickness Line thickness.
-@param lineType Type of the line. See #LineTypes.
-@param shift Number of fractional bits in the point coordinates.
- */
-CV_EXPORTS_W void line(InputOutputArray img, Point pt1, Point pt2, const Scalar& color,
-                     int thickness = 1, int lineType = LINE_8, int shift = 0);
-
-/** @brief Draws an arrow segment pointing from the first point to the second one.
-
-The function cv::arrowedLine draws an arrow between pt1 and pt2 points in the image. See also #line.
-
-@param img Image.
-@param pt1 The point the arrow starts from.
-@param pt2 The point the arrow points to.
-@param color Line color.
-@param thickness Line thickness.
-@param line_type Type of the line. See #LineTypes
-@param shift Number of fractional bits in the point coordinates.
-@param tipLength The length of the arrow tip in relation to the arrow length
- */
-CV_EXPORTS_W void arrowedLine(InputOutputArray img, Point pt1, Point pt2, const Scalar& color,
-                     int thickness=1, int line_type=8, int shift=0, double tipLength=0.1);
-
-/** @brief Draws a simple, thick, or filled up-right rectangle.
-
-The function cv::rectangle draws a rectangle outline or a filled rectangle whose two opposite corners
-are pt1 and pt2.
-
-@param img Image.
-@param pt1 Vertex of the rectangle.
-@param pt2 Vertex of the rectangle opposite to pt1 .
-@param color Rectangle color or brightness (grayscale image).
-@param thickness Thickness of lines that make up the rectangle. Negative values, like #FILLED,
-mean that the function has to draw a filled rectangle.
-@param lineType Type of the line. See #LineTypes
-@param shift Number of fractional bits in the point coordinates.
- */
-CV_EXPORTS_W void rectangle(InputOutputArray img, Point pt1, Point pt2,
-                          const Scalar& color, int thickness = 1,
-                          int lineType = LINE_8, int shift = 0);
-
-/** @overload
-
-use `rec` parameter as alternative specification of the drawn rectangle: `r.tl() and
-r.br()-Point(1,1)` are opposite corners
-*/
-CV_EXPORTS_W void rectangle(InputOutputArray img, Rect rec,
-                          const Scalar& color, int thickness = 1,
-                          int lineType = LINE_8, int shift = 0);
-
-/** @example samples/cpp/tutorial_code/ImgProc/basic_drawing/Drawing_2.cpp
-An example using drawing functions
-*/
-
-/** @brief Draws a circle.
-
-The function cv::circle draws a simple or filled circle with a given center and radius.
-@param img Image where the circle is drawn.
-@param center Center of the circle.
-@param radius Radius of the circle.
-@param color Circle color.
-@param thickness Thickness of the circle outline, if positive. Negative values, like #FILLED,
-mean that a filled circle is to be drawn.
-@param lineType Type of the circle boundary. See #LineTypes
-@param shift Number of fractional bits in the coordinates of the center and in the radius value.
- */
-CV_EXPORTS_W void circle(InputOutputArray img, Point center, int radius,
-                       const Scalar& color, int thickness = 1,
-                       int lineType = LINE_8, int shift = 0);
-
-/** @brief Draws a simple or thick elliptic arc or fills an ellipse sector.
-
-The function cv::ellipse with more parameters draws an ellipse outline, a filled ellipse, an elliptic
-arc, or a filled ellipse sector. The drawing code uses general parametric form.
-A piecewise-linear curve is used to approximate the elliptic arc
-boundary. If you need more control of the ellipse rendering, you can retrieve the curve using
-#ellipse2Poly and then render it with #polylines or fill it with #fillPoly. If you use the first
-variant of the function and want to draw the whole ellipse, not an arc, pass `startAngle=0` and
-`endAngle=360`. If `startAngle` is greater than `endAngle`, they are swapped. The figure below explains
-the meaning of the parameters to draw the blue arc.
-
-![Parameters of Elliptic Arc](pics/ellipse.svg)
-
-@param img Image.
-@param center Center of the ellipse.
-@param axes Half of the size of the ellipse main axes.
-@param angle Ellipse rotation angle in degrees.
-@param startAngle Starting angle of the elliptic arc in degrees.
-@param endAngle Ending angle of the elliptic arc in degrees.
-@param color Ellipse color.
-@param thickness Thickness of the ellipse arc outline, if positive. Otherwise, this indicates that
-a filled ellipse sector is to be drawn.
-@param lineType Type of the ellipse boundary. See #LineTypes
-@param shift Number of fractional bits in the coordinates of the center and values of axes.
- */
-CV_EXPORTS_W void ellipse(InputOutputArray img, Point center, Size axes,
-                        double angle, double startAngle, double endAngle,
-                        const Scalar& color, int thickness = 1,
-                        int lineType = LINE_8, int shift = 0);
-
-/** @overload
-@param img Image.
-@param box Alternative ellipse representation via RotatedRect. This means that the function draws
-an ellipse inscribed in the rotated rectangle.
-@param color Ellipse color.
-@param thickness Thickness of the ellipse arc outline, if positive. Otherwise, this indicates that
-a filled ellipse sector is to be drawn.
-@param lineType Type of the ellipse boundary. See #LineTypes
-*/
-CV_EXPORTS_W void ellipse(InputOutputArray img, const RotatedRect& box, const Scalar& color,
-                        int thickness = 1, int lineType = LINE_8);
-
-/* ----------------------------------------------------------------------------------------- */
-/* ADDING A SET OF PREDEFINED MARKERS WHICH COULD BE USED TO HIGHLIGHT POSITIONS IN AN IMAGE */
-/* ----------------------------------------------------------------------------------------- */
-
-/** @brief Draws a marker on a predefined position in an image.
-
-The function cv::drawMarker draws a marker on a given position in the image. For the moment several
-marker types are supported, see #MarkerTypes for more information.
-
-@param img Image.
-@param position The point where the crosshair is positioned.
-@param color Line color.
-@param markerType The specific type of marker you want to use, see #MarkerTypes
-@param thickness Line thickness.
-@param line_type Type of the line, See #LineTypes
-@param markerSize The length of the marker axis [default = 20 pixels]
- */
-CV_EXPORTS_W void drawMarker(InputOutputArray img, Point position, const Scalar& color,
-                             int markerType = MARKER_CROSS, int markerSize=20, int thickness=1,
-                             int line_type=8);
-
-/* ----------------------------------------------------------------------------------------- */
-/* END OF MARKER SECTION */
-/* ----------------------------------------------------------------------------------------- */
-
-/** @brief Fills a convex polygon.
-
-The function cv::fillConvexPoly draws a filled convex polygon. This function is much faster than the
-function #fillPoly . It can fill not only convex polygons but any monotonic polygon without
-self-intersections, that is, a polygon whose contour intersects every horizontal line (scan line)
-twice at the most (though, its top-most and/or the bottom edge could be horizontal).
-
-@param img Image.
-@param points Polygon vertices.
-@param color Polygon color.
-@param lineType Type of the polygon boundaries. See #LineTypes
-@param shift Number of fractional bits in the vertex coordinates.
- */
-CV_EXPORTS_W void fillConvexPoly(InputOutputArray img, InputArray points,
-                                 const Scalar& color, int lineType = LINE_8,
-                                 int shift = 0);
-
-/** @overload */
-CV_EXPORTS void fillConvexPoly(InputOutputArray img, const Point* pts, int npts,
-                               const Scalar& color, int lineType = LINE_8,
-                               int shift = 0);
-
-/** @example samples/cpp/tutorial_code/ImgProc/basic_drawing/Drawing_1.cpp
-An example using drawing functions
-Check @ref tutorial_random_generator_and_text "the corresponding tutorial" for more details
-*/
-
-/** @brief Fills the area bounded by one or more polygons.
-
-The function cv::fillPoly fills an area bounded by several polygonal contours. The function can fill
-complex areas, for example, areas with holes, contours with self-intersections (some of their
-parts), and so forth.
-
-@param img Image.
-@param pts Array of polygons where each polygon is represented as an array of points.
-@param color Polygon color.
-@param lineType Type of the polygon boundaries. See #LineTypes
-@param shift Number of fractional bits in the vertex coordinates.
-@param offset Optional offset of all points of the contours.
- */
-CV_EXPORTS_W void fillPoly(InputOutputArray img, InputArrayOfArrays pts,
-                           const Scalar& color, int lineType = LINE_8, int shift = 0,
-                           Point offset = Point() );
-
-/** @overload */
-CV_EXPORTS void fillPoly(InputOutputArray img, const Point** pts,
-                         const int* npts, int ncontours,
-                         const Scalar& color, int lineType = LINE_8, int shift = 0,
-                         Point offset = Point() );
-
-/** @brief Draws several polygonal curves.
-
-@param img Image.
-@param pts Array of polygonal curves.
-@param isClosed Flag indicating whether the drawn polylines are closed or not. If they are closed,
-the function draws a line from the last vertex of each curve to its first vertex.
-@param color Polyline color.
-@param thickness Thickness of the polyline edges.
-@param lineType Type of the line segments. See #LineTypes
-@param shift Number of fractional bits in the vertex coordinates.
-
-The function cv::polylines draws one or more polygonal curves.
- */
-CV_EXPORTS_W void polylines(InputOutputArray img, InputArrayOfArrays pts,
-                            bool isClosed, const Scalar& color,
-                            int thickness = 1, int lineType = LINE_8, int shift = 0 );
-
-/** @overload */
-CV_EXPORTS void polylines(InputOutputArray img, const Point* const* pts, const int* npts,
-                          int ncontours, bool isClosed, const Scalar& color,
-                          int thickness = 1, int lineType = LINE_8, int shift = 0 );
-
-/** @example samples/cpp/contours2.cpp
-An example program illustrates the use of cv::findContours and cv::drawContours
-\image html WindowsQtContoursOutput.png "Screenshot of the program"
-*/
-
-/** @example samples/cpp/segment_objects.cpp
-An example using drawContours to clean up a background segmentation result
-*/
-
-/** @brief Draws contours outlines or filled contours.
-
-The function draws contour outlines in the image if \f$\texttt{thickness} \ge 0\f$ or fills the area
-bounded by the contours if \f$\texttt{thickness}<0\f$ . The example below shows how to retrieve
-connected components from the binary image and label them: :
-@include snippets/imgproc_drawContours.cpp
-
-@param image Destination image.
-@param contours All the input contours. Each contour is stored as a point vector.
-@param contourIdx Parameter indicating a contour to draw. If it is negative, all the contours are drawn.
-@param color Color of the contours.
-@param thickness Thickness of lines the contours are drawn with. If it is negative (for example,
-thickness=#FILLED ), the contour interiors are drawn.
-@param lineType Line connectivity. See #LineTypes
-@param hierarchy Optional information about hierarchy. It is only needed if you want to draw only
-some of the contours (see maxLevel ).
-@param maxLevel Maximal level for drawn contours. If it is 0, only the specified contour is drawn.
-If it is 1, the function draws the contour(s) and all the nested contours. If it is 2, the function
-draws the contours, all the nested contours, all the nested-to-nested contours, and so on. This
-parameter is only taken into account when there is hierarchy available.
-@param offset Optional contour shift parameter. Shift all the drawn contours by the specified
-\f$\texttt{offset}=(dx,dy)\f$ .
-@note When thickness=#FILLED, the function is designed to handle connected components with holes correctly
-even when no hierarchy data is provided. This is done by analyzing all the outlines together
-using even-odd rule. This may give incorrect results if you have a joint collection of separately retrieved
-contours. In order to solve this problem, you need to call #drawContours separately for each sub-group
-of contours, or iterate over the collection using contourIdx parameter.
- */
-CV_EXPORTS_W void drawContours( InputOutputArray image, InputArrayOfArrays contours,
-                              int contourIdx, const Scalar& color,
-                              int thickness = 1, int lineType = LINE_8,
-                              InputArray hierarchy = noArray(),
-                              int maxLevel = INT_MAX, Point offset = Point() );
-
-/** @brief Clips the line against the image rectangle.
-
-The function cv::clipLine calculates a part of the line segment that is entirely within the specified
-rectangle. It returns false if the line segment is completely outside the rectangle. Otherwise,
-it returns true .
-@param imgSize Image size. The image rectangle is Rect(0, 0, imgSize.width, imgSize.height) .
-@param pt1 First line point.
-@param pt2 Second line point.
- */
-CV_EXPORTS bool clipLine(Size imgSize, CV_IN_OUT Point& pt1, CV_IN_OUT Point& pt2);
-
-/** @overload
-@param imgSize Image size. The image rectangle is Rect(0, 0, imgSize.width, imgSize.height) .
-@param pt1 First line point.
-@param pt2 Second line point.
-*/
-CV_EXPORTS bool clipLine(Size2l imgSize, CV_IN_OUT Point2l& pt1, CV_IN_OUT Point2l& pt2);
-
-/** @overload
-@param imgRect Image rectangle.
-@param pt1 First line point.
-@param pt2 Second line point.
-*/
-CV_EXPORTS_W bool clipLine(Rect imgRect, CV_OUT CV_IN_OUT Point& pt1, CV_OUT CV_IN_OUT Point& pt2);
-
-/** @brief Approximates an elliptic arc with a polyline.
-
-The function ellipse2Poly computes the vertices of a polyline that approximates the specified
-elliptic arc. It is used by #ellipse. If `arcStart` is greater than `arcEnd`, they are swapped.
-
-@param center Center of the arc.
-@param axes Half of the size of the ellipse main axes. See #ellipse for details.
-@param angle Rotation angle of the ellipse in degrees. See #ellipse for details.
-@param arcStart Starting angle of the elliptic arc in degrees.
-@param arcEnd Ending angle of the elliptic arc in degrees.
-@param delta Angle between the subsequent polyline vertices. It defines the approximation
-accuracy.
-@param pts Output vector of polyline vertices.
- */
-CV_EXPORTS_W void ellipse2Poly( Point center, Size axes, int angle,
-                                int arcStart, int arcEnd, int delta,
-                                CV_OUT std::vector<Point>& pts );
-
-/** @overload
-@param center Center of the arc.
-@param axes Half of the size of the ellipse main axes. See #ellipse for details.
-@param angle Rotation angle of the ellipse in degrees. See #ellipse for details.
-@param arcStart Starting angle of the elliptic arc in degrees.
-@param arcEnd Ending angle of the elliptic arc in degrees.
-@param delta Angle between the subsequent polyline vertices. It defines the approximation accuracy.
-@param pts Output vector of polyline vertices.
-*/
-CV_EXPORTS void ellipse2Poly(Point2d center, Size2d axes, int angle,
-                             int arcStart, int arcEnd, int delta,
-                             CV_OUT std::vector<Point2d>& pts);
-
-/** @brief Draws a text string.
-
-The function cv::putText renders the specified text string in the image. Symbols that cannot be rendered
-using the specified font are replaced by question marks. See #getTextSize for a text rendering code
-example.
-
-@param img Image.
-@param text Text string to be drawn.
-@param org Bottom-left corner of the text string in the image.
-@param fontFace Font type, see #HersheyFonts.
-@param fontScale Font scale factor that is multiplied by the font-specific base size.
-@param color Text color.
-@param thickness Thickness of the lines used to draw a text.
-@param lineType Line type. See #LineTypes
-@param bottomLeftOrigin When true, the image data origin is at the bottom-left corner. Otherwise,
-it is at the top-left corner.
- */
-CV_EXPORTS_W void putText( InputOutputArray img, const String& text, Point org,
-                         int fontFace, double fontScale, Scalar color,
-                         int thickness = 1, int lineType = LINE_8,
-                         bool bottomLeftOrigin = false );
-
-/** @brief Calculates the width and height of a text string.
-
-The function cv::getTextSize calculates and returns the size of a box that contains the specified text.
-That is, the following code renders some text, the tight box surrounding it, and the baseline: :
-@code
-    String text = "Funny text inside the box";
-    int fontFace = FONT_HERSHEY_SCRIPT_SIMPLEX;
-    double fontScale = 2;
-    int thickness = 3;
-
-    Mat img(600, 800, CV_8UC3, Scalar::all(0));
-
-    int baseline=0;
-    Size textSize = getTextSize(text, fontFace,
-                                fontScale, thickness, &baseline);
-    baseline += thickness;
-
-    // center the text
-    Point textOrg((img.cols - textSize.width)/2,
-                  (img.rows + textSize.height)/2);
-
-    // draw the box
-    rectangle(img, textOrg + Point(0, baseline),
-              textOrg + Point(textSize.width, -textSize.height),
-              Scalar(0,0,255));
-    // ... and the baseline first
-    line(img, textOrg + Point(0, thickness),
-         textOrg + Point(textSize.width, thickness),
-         Scalar(0, 0, 255));
-
-    // then put the text itself
-    putText(img, text, textOrg, fontFace, fontScale,
-            Scalar::all(255), thickness, 8);
-@endcode
-
-@param text Input text string.
-@param fontFace Font to use, see #HersheyFonts.
-@param fontScale Font scale factor that is multiplied by the font-specific base size.
-@param thickness Thickness of lines used to render the text. See #putText for details.
-@param[out] baseLine y-coordinate of the baseline relative to the bottom-most text
-point.
-@return The size of a box that contains the specified text.
-
-@see putText
- */
-CV_EXPORTS_W Size getTextSize(const String& text, int fontFace,
-                            double fontScale, int thickness,
-                            CV_OUT int* baseLine);
-
-
-/** @brief Calculates the font-specific size to use to achieve a given height in pixels.
-
-@param fontFace Font to use, see cv::HersheyFonts.
-@param pixelHeight Pixel height to compute the fontScale for
-@param thickness Thickness of lines used to render the text.See putText for details.
-@return The fontSize to use for cv::putText
-
-@see cv::putText
-*/
-CV_EXPORTS_W double getFontScaleFromHeight(const int fontFace,
-                                           const int pixelHeight,
-                                           const int thickness = 1);
-
-/** @brief Line iterator
-
-The class is used to iterate over all the pixels on the raster line
-segment connecting two specified points.
-
-The class LineIterator is used to get each pixel of a raster line. It
-can be treated as versatile implementation of the Bresenham algorithm
-where you can stop at each pixel and do some extra processing, for
-example, grab pixel values along the line or draw a line with an effect
-(for example, with XOR operation).
-
-The number of pixels along the line is stored in LineIterator::count.
-The method LineIterator::pos returns the current position in the image:
-
-@code{.cpp}
-// grabs pixels along the line (pt1, pt2)
-// from 8-bit 3-channel image to the buffer
-LineIterator it(img, pt1, pt2, 8);
-LineIterator it2 = it;
-vector<Vec3b> buf(it.count);
-
-for(int i = 0; i < it.count; i++, ++it)
-    buf[i] = *(const Vec3b*)*it;
-
-// alternative way of iterating through the line
-for(int i = 0; i < it2.count; i++, ++it2)
-{
-    Vec3b val = img.at<Vec3b>(it2.pos());
-    CV_Assert(buf[i] == val);
-}
-@endcode
-*/
-class CV_EXPORTS LineIterator
-{
-public:
-    /** @brief initializes the iterator
-
-    creates iterators for the line connecting pt1 and pt2
-    the line will be clipped on the image boundaries
-    the line is 8-connected or 4-connected
-    If leftToRight=true, then the iteration is always done
-    from the left-most point to the right most,
-    not to depend on the ordering of pt1 and pt2 parameters;
-    */
-    LineIterator( const Mat& img, Point pt1, Point pt2,
-                  int connectivity = 8, bool leftToRight = false )
-    {
-        init(&img, Rect(0, 0, img.cols, img.rows), pt1, pt2, connectivity, leftToRight);
-        ptmode = false;
-    }
-    LineIterator( Point pt1, Point pt2,
-                  int connectivity = 8, bool leftToRight = false )
-    {
-        init(0, Rect(std::min(pt1.x, pt2.x),
-                     std::min(pt1.y, pt2.y),
-                     std::max(pt1.x, pt2.x) - std::min(pt1.x, pt2.x) + 1,
-                     std::max(pt1.y, pt2.y) - std::min(pt1.y, pt2.y) + 1),
-             pt1, pt2, connectivity, leftToRight);
-        ptmode = true;
-    }
-    LineIterator( Size boundingAreaSize, Point pt1, Point pt2,
-                  int connectivity = 8, bool leftToRight = false )
-    {
-        init(0, Rect(0, 0, boundingAreaSize.width, boundingAreaSize.height),
-             pt1, pt2, connectivity, leftToRight);
-        ptmode = true;
-    }
-    LineIterator( Rect boundingAreaRect, Point pt1, Point pt2,
-                  int connectivity = 8, bool leftToRight = false )
-    {
-        init(0, boundingAreaRect, pt1, pt2, connectivity, leftToRight);
-        ptmode = true;
-    }
-    void init(const Mat* img, Rect boundingAreaRect, Point pt1, Point pt2, int connectivity, bool leftToRight);
-
-    /** @brief returns pointer to the current pixel
-    */
-    uchar* operator *();
-    /** @brief prefix increment operator (++it). shifts iterator to the next pixel
-    */
-    LineIterator& operator ++();
-    /** @brief postfix increment operator (it++). shifts iterator to the next pixel
-    */
-    LineIterator operator ++(int);
-    /** @brief returns coordinates of the current pixel
-    */
-    Point pos() const;
-
-    uchar* ptr;
-    const uchar* ptr0;
-    int step, elemSize;
-    int err, count;
-    int minusDelta, plusDelta;
-    int minusStep, plusStep;
-    int minusShift, plusShift;
-    Point p;
-    bool ptmode;
-};
-
-//! @cond IGNORED
-
-// === LineIterator implementation ===
-
-inline
-uchar* LineIterator::operator *()
-{
-    return ptmode ? 0 : ptr;
-}
-
-inline
-LineIterator& LineIterator::operator ++()
-{
-    int mask = err < 0 ? -1 : 0;
-    err += minusDelta + (plusDelta & mask);
-    if(!ptmode)
-    {
-        ptr += minusStep + (plusStep & mask);
-    }
-    else
-    {
-        p.x += minusShift + (plusShift & mask);
-        p.y += minusStep + (plusStep & mask);
-    }
-    return *this;
-}
-
-inline
-LineIterator LineIterator::operator ++(int)
-{
-    LineIterator it = *this;
-    ++(*this);
-    return it;
-}
-
-inline
-Point LineIterator::pos() const
-{
-    if(!ptmode)
-    {
-        size_t offset = (size_t)(ptr - ptr0);
-        int y = (int)(offset/step);
-        int x = (int)((offset - (size_t)y*step)/elemSize);
-        return Point(x, y);
-    }
-    return p;
-}
-
-//! @endcond
-
-//! @} imgproc_draw
-
-//! @} imgproc
-
-} // cv
-
-
-#include "./imgproc/segmentation.hpp"
-
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/imgproc/bindings.hpp b/projects/llm_framework/include/opencv4/opencv2/imgproc/bindings.hpp
deleted file mode 100644
index c69527a7..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/imgproc/bindings.hpp
+++ /dev/null
@@ -1,34 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_IMGPROC_BINDINGS_HPP
-#define OPENCV_IMGPROC_BINDINGS_HPP
-
-// This file contains special overloads for OpenCV bindings
-// No need to use these functions in C++ code.
-
-namespace cv {
-
-/** @brief Finds lines in a binary image using the standard Hough transform and get accumulator.
- *
- * @note This function is for bindings use only. Use original function in C++ code
- *
- * @sa HoughLines
- */
-CV_WRAP static inline
-void HoughLinesWithAccumulator(
-        InputArray image, OutputArray lines,
-        double rho, double theta, int threshold,
-        double srn = 0, double stn = 0,
-        double min_theta = 0, double max_theta = CV_PI
-)
-{
-    std::vector<Vec3f> lines_acc;
-    HoughLines(image, lines_acc, rho, theta, threshold, srn, stn, min_theta, max_theta);
-    Mat(lines_acc).copyTo(lines);
-}
-
-}  // namespace
-
-#endif  // OPENCV_IMGPROC_BINDINGS_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/imgproc/detail/gcgraph.hpp b/projects/llm_framework/include/opencv4/opencv2/imgproc/detail/gcgraph.hpp
deleted file mode 100644
index f17c6e7a..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/imgproc/detail/gcgraph.hpp
+++ /dev/null
@@ -1,395 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                        Intel License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000, Intel Corporation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of Intel Corporation may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_IMGPROC_DETAIL_GCGRAPH_HPP
-#define OPENCV_IMGPROC_DETAIL_GCGRAPH_HPP
-
-//! @cond IGNORED
-
-namespace cv { namespace detail {
-template <class TWeight> class GCGraph
-{
-public:
-    GCGraph();
-    GCGraph( unsigned int vtxCount, unsigned int edgeCount );
-    ~GCGraph();
-    void create( unsigned int vtxCount, unsigned int edgeCount );
-    int addVtx();
-    void addEdges( int i, int j, TWeight w, TWeight revw );
-    void addTermWeights( int i, TWeight sourceW, TWeight sinkW );
-    TWeight maxFlow();
-    bool inSourceSegment( int i );
-private:
-    class Vtx
-    {
-    public:
-        Vtx *next; // initialized and used in maxFlow() only
-        int parent;
-        int first;
-        int ts;
-        int dist;
-        TWeight weight;
-        uchar t;
-    };
-    class Edge
-    {
-    public:
-        int dst;
-        int next;
-        TWeight weight;
-    };
-
-    std::vector<Vtx> vtcs;
-    std::vector<Edge> edges;
-    TWeight flow;
-};
-
-template <class TWeight>
-GCGraph<TWeight>::GCGraph()
-{
-    flow = 0;
-}
-template <class TWeight>
-GCGraph<TWeight>::GCGraph( unsigned int vtxCount, unsigned int edgeCount )
-{
-    create( vtxCount, edgeCount );
-}
-template <class TWeight>
-GCGraph<TWeight>::~GCGraph()
-{
-}
-template <class TWeight>
-void GCGraph<TWeight>::create( unsigned int vtxCount, unsigned int edgeCount )
-{
-    vtcs.reserve( vtxCount );
-    edges.reserve( edgeCount + 2 );
-    flow = 0;
-}
-
-template <class TWeight>
-int GCGraph<TWeight>::addVtx()
-{
-    Vtx v;
-    memset( &v, 0, sizeof(Vtx));
-    vtcs.push_back(v);
-    return (int)vtcs.size() - 1;
-}
-
-template <class TWeight>
-void GCGraph<TWeight>::addEdges( int i, int j, TWeight w, TWeight revw )
-{
-    CV_Assert( i>=0 && i<(int)vtcs.size() );
-    CV_Assert( j>=0 && j<(int)vtcs.size() );
-    CV_Assert( w>=0 && revw>=0 );
-    CV_Assert( i != j );
-
-    if( !edges.size() )
-        edges.resize( 2 );
-
-    Edge fromI, toI;
-    fromI.dst = j;
-    fromI.next = vtcs[i].first;
-    fromI.weight = w;
-    vtcs[i].first = (int)edges.size();
-    edges.push_back( fromI );
-
-    toI.dst = i;
-    toI.next = vtcs[j].first;
-    toI.weight = revw;
-    vtcs[j].first = (int)edges.size();
-    edges.push_back( toI );
-}
-
-template <class TWeight>
-void GCGraph<TWeight>::addTermWeights( int i, TWeight sourceW, TWeight sinkW )
-{
-    CV_Assert( i>=0 && i<(int)vtcs.size() );
-
-    TWeight dw = vtcs[i].weight;
-    if( dw > 0 )
-        sourceW += dw;
-    else
-        sinkW -= dw;
-    flow += (sourceW < sinkW) ? sourceW : sinkW;
-    vtcs[i].weight = sourceW - sinkW;
-}
-
-template <class TWeight>
-TWeight GCGraph<TWeight>::maxFlow()
-{
-    CV_Assert(!vtcs.empty());
-    CV_Assert(!edges.empty());
-    const int TERMINAL = -1, ORPHAN = -2;
-    Vtx stub, *nilNode = &stub, *first = nilNode, *last = nilNode;
-    int curr_ts = 0;
-    stub.next = nilNode;
-    Vtx *vtxPtr = &vtcs[0];
-    Edge *edgePtr = &edges[0];
-
-    std::vector<Vtx*> orphans;
-
-    // initialize the active queue and the graph vertices
-    for( int i = 0; i < (int)vtcs.size(); i++ )
-    {
-        Vtx* v = vtxPtr + i;
-        v->ts = 0;
-        if( v->weight != 0 )
-        {
-            last = last->next = v;
-            v->dist = 1;
-            v->parent = TERMINAL;
-            v->t = v->weight < 0;
-        }
-        else
-            v->parent = 0;
-    }
-    first = first->next;
-    last->next = nilNode;
-    nilNode->next = 0;
-
-    // run the search-path -> augment-graph -> restore-trees loop
-    for(;;)
-    {
-        Vtx* v, *u;
-        int e0 = -1, ei = 0, ej = 0;
-        TWeight minWeight, weight;
-        uchar vt;
-
-        // grow S & T search trees, find an edge connecting them
-        while( first != nilNode )
-        {
-            v = first;
-            if( v->parent )
-            {
-                vt = v->t;
-                for( ei = v->first; ei != 0; ei = edgePtr[ei].next )
-                {
-                    if( edgePtr[ei^vt].weight == 0 )
-                        continue;
-                    u = vtxPtr+edgePtr[ei].dst;
-                    if( !u->parent )
-                    {
-                        u->t = vt;
-                        u->parent = ei ^ 1;
-                        u->ts = v->ts;
-                        u->dist = v->dist + 1;
-                        if( !u->next )
-                        {
-                            u->next = nilNode;
-                            last = last->next = u;
-                        }
-                        continue;
-                    }
-
-                    if( u->t != vt )
-                    {
-                        e0 = ei ^ vt;
-                        break;
-                    }
-
-                    if( u->dist > v->dist+1 && u->ts <= v->ts )
-                    {
-                        // reassign the parent
-                        u->parent = ei ^ 1;
-                        u->ts = v->ts;
-                        u->dist = v->dist + 1;
-                    }
-                }
-                if( e0 > 0 )
-                    break;
-            }
-            // exclude the vertex from the active list
-            first = first->next;
-            v->next = 0;
-        }
-
-        if( e0 <= 0 )
-            break;
-
-        // find the minimum edge weight along the path
-        minWeight = edgePtr[e0].weight;
-        CV_Assert( minWeight > 0 );
-        // k = 1: source tree, k = 0: destination tree
-        for( int k = 1; k >= 0; k-- )
-        {
-            for( v = vtxPtr+edgePtr[e0^k].dst;; v = vtxPtr+edgePtr[ei].dst )
-            {
-                if( (ei = v->parent) < 0 )
-                    break;
-                weight = edgePtr[ei^k].weight;
-                minWeight = MIN(minWeight, weight);
-                CV_Assert( minWeight > 0 );
-            }
-            weight = fabs(v->weight);
-            minWeight = MIN(minWeight, weight);
-            CV_Assert( minWeight > 0 );
-        }
-
-        // modify weights of the edges along the path and collect orphans
-        edgePtr[e0].weight -= minWeight;
-        edgePtr[e0^1].weight += minWeight;
-        flow += minWeight;
-
-        // k = 1: source tree, k = 0: destination tree
-        for( int k = 1; k >= 0; k-- )
-        {
-            for( v = vtxPtr+edgePtr[e0^k].dst;; v = vtxPtr+edgePtr[ei].dst )
-            {
-                if( (ei = v->parent) < 0 )
-                    break;
-                edgePtr[ei^(k^1)].weight += minWeight;
-                if( (edgePtr[ei^k].weight -= minWeight) == 0 )
-                {
-                    orphans.push_back(v);
-                    v->parent = ORPHAN;
-                }
-            }
-
-            v->weight = v->weight + minWeight*(1-k*2);
-            if( v->weight == 0 )
-            {
-               orphans.push_back(v);
-               v->parent = ORPHAN;
-            }
-        }
-
-        // restore the search trees by finding new parents for the orphans
-        curr_ts++;
-        while( !orphans.empty() )
-        {
-            Vtx* v2 = orphans.back();
-            orphans.pop_back();
-
-            int d, minDist = INT_MAX;
-            e0 = 0;
-            vt = v2->t;
-
-            for( ei = v2->first; ei != 0; ei = edgePtr[ei].next )
-            {
-                if( edgePtr[ei^(vt^1)].weight == 0 )
-                    continue;
-                u = vtxPtr+edgePtr[ei].dst;
-                if( u->t != vt || u->parent == 0 )
-                    continue;
-                // compute the distance to the tree root
-                for( d = 0;; )
-                {
-                    if( u->ts == curr_ts )
-                    {
-                        d += u->dist;
-                        break;
-                    }
-                    ej = u->parent;
-                    d++;
-                    if( ej < 0 )
-                    {
-                        if( ej == ORPHAN )
-                            d = INT_MAX-1;
-                        else
-                        {
-                            u->ts = curr_ts;
-                            u->dist = 1;
-                        }
-                        break;
-                    }
-                    u = vtxPtr+edgePtr[ej].dst;
-                }
-
-                // update the distance
-                if( ++d < INT_MAX )
-                {
-                    if( d < minDist )
-                    {
-                        minDist = d;
-                        e0 = ei;
-                    }
-                    for( u = vtxPtr+edgePtr[ei].dst; u->ts != curr_ts; u = vtxPtr+edgePtr[u->parent].dst )
-                    {
-                        u->ts = curr_ts;
-                        u->dist = --d;
-                    }
-                }
-            }
-
-            if( (v2->parent = e0) > 0 )
-            {
-                v2->ts = curr_ts;
-                v2->dist = minDist;
-                continue;
-            }
-
-            /* no parent is found */
-            v2->ts = 0;
-            for( ei = v2->first; ei != 0; ei = edgePtr[ei].next )
-            {
-                u = vtxPtr+edgePtr[ei].dst;
-                ej = u->parent;
-                if( u->t != vt || !ej )
-                    continue;
-                if( edgePtr[ei^(vt^1)].weight && !u->next )
-                {
-                    u->next = nilNode;
-                    last = last->next = u;
-                }
-                if( ej > 0 && vtxPtr+edgePtr[ej].dst == v2 )
-                {
-                    orphans.push_back(u);
-                    u->parent = ORPHAN;
-                }
-            }
-        }
-    }
-    return flow;
-}
-
-template <class TWeight>
-bool GCGraph<TWeight>::inSourceSegment( int i )
-{
-    CV_Assert( i>=0 && i<(int)vtcs.size() );
-    return vtcs[i].t == 0;
-}
-
-}} // namespace detail, cv
-
-
-//! @endcond
-
-#endif  // OPENCV_IMGPROC_DETAIL_GCGRAPH_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/imgproc/hal/hal.hpp b/projects/llm_framework/include/opencv4/opencv2/imgproc/hal/hal.hpp
deleted file mode 100644
index f129012b..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/imgproc/hal/hal.hpp
+++ /dev/null
@@ -1,246 +0,0 @@
-#ifndef CV_IMGPROC_HAL_HPP
-#define CV_IMGPROC_HAL_HPP
-
-#include "opencv2/core/cvdef.h"
-#include "opencv2/core/cvstd.hpp"
-#include "opencv2/core/hal/interface.h"
-
-namespace cv { namespace hal {
-
-//! @addtogroup imgproc_hal_functions
-//! @{
-
-//---------------------------
-//! @cond IGNORED
-
-struct CV_EXPORTS Filter2D
-{
-    CV_DEPRECATED static Ptr<hal::Filter2D> create(uchar * , size_t , int ,
-                                     int , int ,
-                                     int , int ,
-                                     int , int ,
-                                     int , double ,
-                                     int , int ,
-                                     bool , bool );
-    virtual void apply(uchar * , size_t ,
-                       uchar * , size_t ,
-                       int , int ,
-                       int , int ,
-                       int , int ) = 0;
-    virtual ~Filter2D() {}
-};
-
-struct CV_EXPORTS SepFilter2D
-{
-    CV_DEPRECATED static Ptr<hal::SepFilter2D> create(int , int , int ,
-                                        uchar * , int ,
-                                        uchar * , int ,
-                                        int , int ,
-                                        double , int );
-    virtual void apply(uchar * , size_t ,
-                       uchar * , size_t ,
-                       int , int ,
-                       int , int ,
-                       int , int ) = 0;
-    virtual ~SepFilter2D() {}
-};
-
-
-struct CV_EXPORTS Morph
-{
-    CV_DEPRECATED static Ptr<hal::Morph> create(int , int , int , int , int ,
-                                    int , uchar * , size_t ,
-                                    int , int ,
-                                    int , int ,
-                                    int , const double *,
-                                    int , bool , bool );
-    virtual void apply(uchar * , size_t , uchar * , size_t , int , int ,
-                       int , int , int , int ,
-                       int , int , int , int ) = 0;
-    virtual ~Morph() {}
-};
-
-//! @endcond
-//---------------------------
-
-CV_EXPORTS void filter2D(int stype, int dtype, int kernel_type,
-                         uchar * src_data, size_t src_step,
-                         uchar * dst_data, size_t dst_step,
-                         int width, int height,
-                         int full_width, int full_height,
-                         int offset_x, int offset_y,
-                         uchar * kernel_data, size_t kernel_step,
-                         int kernel_width, int kernel_height,
-                         int anchor_x, int anchor_y,
-                         double delta, int borderType,
-                         bool isSubmatrix);
-
-CV_EXPORTS void sepFilter2D(int stype, int dtype, int ktype,
-                            uchar * src_data, size_t src_step,
-                            uchar * dst_data, size_t dst_step,
-                            int width, int height,
-                            int full_width, int full_height,
-                            int offset_x, int offset_y,
-                            uchar * kernelx_data, int kernelx_len,
-                            uchar * kernely_data, int kernely_len,
-                            int anchor_x, int anchor_y,
-                            double delta, int borderType);
-
-CV_EXPORTS void morph(int op, int src_type, int dst_type,
-                      uchar * src_data, size_t src_step,
-                      uchar * dst_data, size_t dst_step,
-                      int width, int height,
-                      int roi_width, int roi_height, int roi_x, int roi_y,
-                      int roi_width2, int roi_height2, int roi_x2, int roi_y2,
-                      int kernel_type, uchar * kernel_data, size_t kernel_step,
-                      int kernel_width, int kernel_height, int anchor_x, int anchor_y,
-                      int borderType, const double borderValue[4],
-                      int iterations, bool isSubmatrix);
-
-
-CV_EXPORTS void resize(int src_type,
-                       const uchar * src_data, size_t src_step, int src_width, int src_height,
-                       uchar * dst_data, size_t dst_step, int dst_width, int dst_height,
-                       double inv_scale_x, double inv_scale_y, int interpolation);
-
-CV_EXPORTS void warpAffine(int src_type,
-                           const uchar * src_data, size_t src_step, int src_width, int src_height,
-                           uchar * dst_data, size_t dst_step, int dst_width, int dst_height,
-                           const double M[6], int interpolation, int borderType, const double borderValue[4]);
-
-CV_EXPORTS void warpPerspective(int src_type,
-                               const uchar * src_data, size_t src_step, int src_width, int src_height,
-                               uchar * dst_data, size_t dst_step, int dst_width, int dst_height,
-                               const double M[9], int interpolation, int borderType, const double borderValue[4]);
-
-CV_EXPORTS void cvtBGRtoBGR(const uchar * src_data, size_t src_step,
-                            uchar * dst_data, size_t dst_step,
-                            int width, int height,
-                            int depth, int scn, int dcn, bool swapBlue);
-
-CV_EXPORTS void cvtBGRtoBGR5x5(const uchar * src_data, size_t src_step,
-                               uchar * dst_data, size_t dst_step,
-                               int width, int height,
-                               int scn, bool swapBlue, int greenBits);
-
-CV_EXPORTS void cvtBGR5x5toBGR(const uchar * src_data, size_t src_step,
-                               uchar * dst_data, size_t dst_step,
-                               int width, int height,
-                               int dcn, bool swapBlue, int greenBits);
-
-CV_EXPORTS void cvtBGRtoGray(const uchar * src_data, size_t src_step,
-                             uchar * dst_data, size_t dst_step,
-                             int width, int height,
-                             int depth, int scn, bool swapBlue);
-
-CV_EXPORTS void cvtGraytoBGR(const uchar * src_data, size_t src_step,
-                             uchar * dst_data, size_t dst_step,
-                             int width, int height,
-                             int depth, int dcn);
-
-CV_EXPORTS void cvtBGR5x5toGray(const uchar * src_data, size_t src_step,
-                                uchar * dst_data, size_t dst_step,
-                                int width, int height,
-                                int greenBits);
-
-CV_EXPORTS void cvtGraytoBGR5x5(const uchar * src_data, size_t src_step,
-                                uchar * dst_data, size_t dst_step,
-                                int width, int height,
-                                int greenBits);
-CV_EXPORTS void cvtBGRtoYUV(const uchar * src_data, size_t src_step,
-                            uchar * dst_data, size_t dst_step,
-                            int width, int height,
-                            int depth, int scn, bool swapBlue, bool isCbCr);
-
-CV_EXPORTS void cvtYUVtoBGR(const uchar * src_data, size_t src_step,
-                            uchar * dst_data, size_t dst_step,
-                            int width, int height,
-                            int depth, int dcn, bool swapBlue, bool isCbCr);
-
-CV_EXPORTS void cvtBGRtoXYZ(const uchar * src_data, size_t src_step,
-                            uchar * dst_data, size_t dst_step,
-                            int width, int height,
-                            int depth, int scn, bool swapBlue);
-
-CV_EXPORTS void cvtXYZtoBGR(const uchar * src_data, size_t src_step,
-                            uchar * dst_data, size_t dst_step,
-                            int width, int height,
-                            int depth, int dcn, bool swapBlue);
-
-CV_EXPORTS void cvtBGRtoHSV(const uchar * src_data, size_t src_step,
-                            uchar * dst_data, size_t dst_step,
-                            int width, int height,
-                            int depth, int scn, bool swapBlue, bool isFullRange, bool isHSV);
-
-CV_EXPORTS void cvtHSVtoBGR(const uchar * src_data, size_t src_step,
-                            uchar * dst_data, size_t dst_step,
-                            int width, int height,
-                            int depth, int dcn, bool swapBlue, bool isFullRange, bool isHSV);
-
-CV_EXPORTS void cvtBGRtoLab(const uchar * src_data, size_t src_step,
-                            uchar * dst_data, size_t dst_step,
-                            int width, int height,
-                            int depth, int scn, bool swapBlue, bool isLab, bool srgb);
-
-CV_EXPORTS void cvtLabtoBGR(const uchar * src_data, size_t src_step,
-                            uchar * dst_data, size_t dst_step,
-                            int width, int height,
-                            int depth, int dcn, bool swapBlue, bool isLab, bool srgb);
-
-CV_EXPORTS void cvtTwoPlaneYUVtoBGR(const uchar * src_data, size_t src_step,
-                                    uchar * dst_data, size_t dst_step,
-                                    int dst_width, int dst_height,
-                                    int dcn, bool swapBlue, int uIdx);
-
-//! Separate Y and UV planes
-CV_EXPORTS void cvtTwoPlaneYUVtoBGR(const uchar * y_data, const uchar * uv_data, size_t src_step,
-                                    uchar * dst_data, size_t dst_step,
-                                    int dst_width, int dst_height,
-                                    int dcn, bool swapBlue, int uIdx);
-
-CV_EXPORTS void cvtTwoPlaneYUVtoBGR(const uchar * y_data, size_t y_step, const uchar * uv_data, size_t uv_step,
-                                    uchar * dst_data, size_t dst_step,
-                                    int dst_width, int dst_height,
-                                    int dcn, bool swapBlue, int uIdx);
-
-CV_EXPORTS void cvtThreePlaneYUVtoBGR(const uchar * src_data, size_t src_step,
-                                      uchar * dst_data, size_t dst_step,
-                                      int dst_width, int dst_height,
-                                      int dcn, bool swapBlue, int uIdx);
-
-CV_EXPORTS void cvtBGRtoThreePlaneYUV(const uchar * src_data, size_t src_step,
-                                      uchar * dst_data, size_t dst_step,
-                                      int width, int height,
-                                      int scn, bool swapBlue, int uIdx);
-
-//! Separate Y and UV planes
-CV_EXPORTS void cvtBGRtoTwoPlaneYUV(const uchar * src_data, size_t src_step,
-                                    uchar * y_data, uchar * uv_data, size_t dst_step,
-                                    int width, int height,
-                                    int scn, bool swapBlue, int uIdx);
-
-CV_EXPORTS void cvtOnePlaneYUVtoBGR(const uchar * src_data, size_t src_step,
-                                    uchar * dst_data, size_t dst_step,
-                                    int width, int height,
-                                    int dcn, bool swapBlue, int uIdx, int ycn);
-
-CV_EXPORTS void cvtRGBAtoMultipliedRGBA(const uchar * src_data, size_t src_step,
-                                        uchar * dst_data, size_t dst_step,
-                                        int width, int height);
-
-CV_EXPORTS void cvtMultipliedRGBAtoRGBA(const uchar * src_data, size_t src_step,
-                                        uchar * dst_data, size_t dst_step,
-                                        int width, int height);
-
-CV_EXPORTS void integral(int depth, int sdepth, int sqdepth,
-                         const uchar* src, size_t srcstep,
-                         uchar* sum, size_t sumstep,
-                         uchar* sqsum, size_t sqsumstep,
-                         uchar* tilted, size_t tstep,
-                         int width, int height, int cn);
-
-//! @}
-
-}}
-
-#endif // CV_IMGPROC_HAL_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/imgproc/hal/interface.h b/projects/llm_framework/include/opencv4/opencv2/imgproc/hal/interface.h
deleted file mode 100644
index f8dbcfe7..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/imgproc/hal/interface.h
+++ /dev/null
@@ -1,46 +0,0 @@
-#ifndef OPENCV_IMGPROC_HAL_INTERFACE_H
-#define OPENCV_IMGPROC_HAL_INTERFACE_H
-
-//! @addtogroup imgproc_hal_interface
-//! @{
-
-//! @name Interpolation modes
-//! @sa cv::InterpolationFlags
-//! @{
-#define CV_HAL_INTER_NEAREST 0
-#define CV_HAL_INTER_LINEAR 1
-#define CV_HAL_INTER_CUBIC 2
-#define CV_HAL_INTER_AREA 3
-#define CV_HAL_INTER_LANCZOS4 4
-//! @}
-
-//! @name Morphology operations
-//! @sa cv::MorphTypes
-//! @{
-#define CV_HAL_MORPH_ERODE 0
-#define CV_HAL_MORPH_DILATE 1
-//! @}
-
-//! @name Threshold types
-//! @sa cv::ThresholdTypes
-//! @{
-#define CV_HAL_THRESH_BINARY      0
-#define CV_HAL_THRESH_BINARY_INV  1
-#define CV_HAL_THRESH_TRUNC       2
-#define CV_HAL_THRESH_TOZERO      3
-#define CV_HAL_THRESH_TOZERO_INV  4
-#define CV_HAL_THRESH_MASK        7
-#define CV_HAL_THRESH_OTSU        8
-#define CV_HAL_THRESH_TRIANGLE    16
-//! @}
-
-//! @name Adaptive threshold algorithm
-//! @sa cv::AdaptiveThresholdTypes
-//! @{
-#define CV_HAL_ADAPTIVE_THRESH_MEAN_C     0
-#define CV_HAL_ADAPTIVE_THRESH_GAUSSIAN_C 1
-//! @}
-
-//! @}
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/imgproc/imgproc.hpp b/projects/llm_framework/include/opencv4/opencv2/imgproc/imgproc.hpp
deleted file mode 100644
index 4175bd0b..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/imgproc/imgproc.hpp
+++ /dev/null
@@ -1,48 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifdef __OPENCV_BUILD
-#error this is a compatibility header which should not be used inside the OpenCV library
-#endif
-
-#include "opencv2/imgproc.hpp"
diff --git a/projects/llm_framework/include/opencv4/opencv2/imgproc/imgproc_c.h b/projects/llm_framework/include/opencv4/opencv2/imgproc/imgproc_c.h
deleted file mode 100644
index 86dc119f..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/imgproc/imgproc_c.h
+++ /dev/null
@@ -1,1177 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_IMGPROC_IMGPROC_C_H
-#define OPENCV_IMGPROC_IMGPROC_C_H
-
-#include "opencv2/imgproc/types_c.h"
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/** @addtogroup imgproc_c
-@{
-*/
-
-/*********************** Background statistics accumulation *****************************/
-
-/** @brief Adds image to accumulator
-@see cv::accumulate
-*/
-CVAPI(void)  cvAcc( const CvArr* image, CvArr* sum,
-                   const CvArr* mask CV_DEFAULT(NULL) );
-
-/** @brief Adds squared image to accumulator
-@see cv::accumulateSquare
-*/
-CVAPI(void)  cvSquareAcc( const CvArr* image, CvArr* sqsum,
-                         const CvArr* mask CV_DEFAULT(NULL) );
-
-/** @brief Adds a product of two images to accumulator
-@see cv::accumulateProduct
-*/
-CVAPI(void)  cvMultiplyAcc( const CvArr* image1, const CvArr* image2, CvArr* acc,
-                           const CvArr* mask CV_DEFAULT(NULL) );
-
-/** @brief Adds image to accumulator with weights: acc = acc*(1-alpha) + image*alpha
-@see cv::accumulateWeighted
-*/
-CVAPI(void)  cvRunningAvg( const CvArr* image, CvArr* acc, double alpha,
-                          const CvArr* mask CV_DEFAULT(NULL) );
-
-/****************************************************************************************\
-*                                    Image Processing                                    *
-\****************************************************************************************/
-
-/** Copies source 2D array inside of the larger destination array and
-   makes a border of the specified type (IPL_BORDER_*) around the copied area. */
-CVAPI(void) cvCopyMakeBorder( const CvArr* src, CvArr* dst, CvPoint offset,
-                              int bordertype, CvScalar value CV_DEFAULT(cvScalarAll(0)));
-
-/** @brief Smooths the image in one of several ways.
-
-@param src The source image
-@param dst The destination image
-@param smoothtype Type of the smoothing, see SmoothMethod_c
-@param size1 The first parameter of the smoothing operation, the aperture width. Must be a
-positive odd number (1, 3, 5, ...)
-@param size2 The second parameter of the smoothing operation, the aperture height. Ignored by
-CV_MEDIAN and CV_BILATERAL methods. In the case of simple scaled/non-scaled and Gaussian blur if
-size2 is zero, it is set to size1. Otherwise it must be a positive odd number.
-@param sigma1 In the case of a Gaussian parameter this parameter may specify Gaussian \f$\sigma\f$
-(standard deviation). If it is zero, it is calculated from the kernel size:
-\f[\sigma  = 0.3 (n/2 - 1) + 0.8  \quad   \text{where}   \quad  n= \begin{array}{l l} \mbox{\texttt{size1} for horizontal kernel} \\ \mbox{\texttt{size2} for vertical kernel} \end{array}\f]
-Using standard sigma for small kernels ( \f$3\times 3\f$ to \f$7\times 7\f$ ) gives better speed. If
-sigma1 is not zero, while size1 and size2 are zeros, the kernel size is calculated from the
-sigma (to provide accurate enough operation).
-@param sigma2 additional parameter for bilateral filtering
-
-@see cv::GaussianBlur, cv::blur, cv::medianBlur, cv::bilateralFilter.
- */
-CVAPI(void) cvSmooth( const CvArr* src, CvArr* dst,
-                      int smoothtype CV_DEFAULT(CV_GAUSSIAN),
-                      int size1 CV_DEFAULT(3),
-                      int size2 CV_DEFAULT(0),
-                      double sigma1 CV_DEFAULT(0),
-                      double sigma2 CV_DEFAULT(0));
-
-/** @brief Convolves an image with the kernel.
-
-@param src input image.
-@param dst output image of the same size and the same number of channels as src.
-@param kernel convolution kernel (or rather a correlation kernel), a single-channel floating point
-matrix; if you want to apply different kernels to different channels, split the image into
-separate color planes using split and process them individually.
-@param anchor anchor of the kernel that indicates the relative position of a filtered point within
-the kernel; the anchor should lie within the kernel; default value (-1,-1) means that the anchor
-is at the kernel center.
-
-@see cv::filter2D
- */
-CVAPI(void) cvFilter2D( const CvArr* src, CvArr* dst, const CvMat* kernel,
-                        CvPoint anchor CV_DEFAULT(cvPoint(-1,-1)));
-
-/** @brief Finds integral image: SUM(X,Y) = sum(x<X,y<Y)I(x,y)
-@see cv::integral
-*/
-CVAPI(void) cvIntegral( const CvArr* image, CvArr* sum,
-                       CvArr* sqsum CV_DEFAULT(NULL),
-                       CvArr* tilted_sum CV_DEFAULT(NULL));
-
-/** @brief Smoothes the input image with gaussian kernel and then down-samples it.
-
-   dst_width = floor(src_width/2)[+1],
-   dst_height = floor(src_height/2)[+1]
-   @see cv::pyrDown
-*/
-CVAPI(void)  cvPyrDown( const CvArr* src, CvArr* dst,
-                        int filter CV_DEFAULT(CV_GAUSSIAN_5x5) );
-
-/** @brief Up-samples image and smoothes the result with gaussian kernel.
-
-   dst_width = src_width*2,
-   dst_height = src_height*2
-   @see cv::pyrUp
-*/
-CVAPI(void)  cvPyrUp( const CvArr* src, CvArr* dst,
-                      int filter CV_DEFAULT(CV_GAUSSIAN_5x5) );
-
-/** @brief Builds pyramid for an image
-@see buildPyramid
-*/
-CVAPI(CvMat**) cvCreatePyramid( const CvArr* img, int extra_layers, double rate,
-                                const CvSize* layer_sizes CV_DEFAULT(0),
-                                CvArr* bufarr CV_DEFAULT(0),
-                                int calc CV_DEFAULT(1),
-                                int filter CV_DEFAULT(CV_GAUSSIAN_5x5) );
-
-/** @brief Releases pyramid */
-CVAPI(void)  cvReleasePyramid( CvMat*** pyramid, int extra_layers );
-
-
-/** @brief Filters image using meanshift algorithm
-@see cv::pyrMeanShiftFiltering
-*/
-CVAPI(void) cvPyrMeanShiftFiltering( const CvArr* src, CvArr* dst,
-    double sp, double sr, int max_level CV_DEFAULT(1),
-    CvTermCriteria termcrit CV_DEFAULT(cvTermCriteria(CV_TERMCRIT_ITER+CV_TERMCRIT_EPS,5,1)));
-
-/** @brief Segments image using seed "markers"
-@see cv::watershed
-*/
-CVAPI(void) cvWatershed( const CvArr* image, CvArr* markers );
-
-/** @brief Calculates an image derivative using generalized Sobel
-
-   (aperture_size = 1,3,5,7) or Scharr (aperture_size = -1) operator.
-   Scharr can be used only for the first dx or dy derivative
-@see cv::Sobel
-*/
-CVAPI(void) cvSobel( const CvArr* src, CvArr* dst,
-                    int xorder, int yorder,
-                    int aperture_size CV_DEFAULT(3));
-
-/** @brief Calculates the image Laplacian: (d2/dx + d2/dy)I
-@see cv::Laplacian
-*/
-CVAPI(void) cvLaplace( const CvArr* src, CvArr* dst,
-                      int aperture_size CV_DEFAULT(3) );
-
-/** @brief Converts input array pixels from one color space to another
-@see cv::cvtColor
-*/
-CVAPI(void)  cvCvtColor( const CvArr* src, CvArr* dst, int code );
-
-
-/** @brief Resizes image (input array is resized to fit the destination array)
-@see cv::resize
-*/
-CVAPI(void)  cvResize( const CvArr* src, CvArr* dst,
-                       int interpolation CV_DEFAULT( CV_INTER_LINEAR ));
-
-/** @brief Warps image with affine transform
-@note ::cvGetQuadrangleSubPix is similar to ::cvWarpAffine, but the outliers are extrapolated using
-replication border mode.
-@see cv::warpAffine
-*/
-CVAPI(void)  cvWarpAffine( const CvArr* src, CvArr* dst, const CvMat* map_matrix,
-                           int flags CV_DEFAULT(CV_INTER_LINEAR+CV_WARP_FILL_OUTLIERS),
-                           CvScalar fillval CV_DEFAULT(cvScalarAll(0)) );
-
-/** @brief Computes affine transform matrix for mapping src[i] to dst[i] (i=0,1,2)
-@see cv::getAffineTransform
-*/
-CVAPI(CvMat*) cvGetAffineTransform( const CvPoint2D32f * src,
-                                    const CvPoint2D32f * dst,
-                                    CvMat * map_matrix );
-
-/** @brief Computes rotation_matrix matrix
-@see cv::getRotationMatrix2D
-*/
-CVAPI(CvMat*)  cv2DRotationMatrix( CvPoint2D32f center, double angle,
-                                   double scale, CvMat* map_matrix );
-
-/** @brief Warps image with perspective (projective) transform
-@see cv::warpPerspective
-*/
-CVAPI(void)  cvWarpPerspective( const CvArr* src, CvArr* dst, const CvMat* map_matrix,
-                                int flags CV_DEFAULT(CV_INTER_LINEAR+CV_WARP_FILL_OUTLIERS),
-                                CvScalar fillval CV_DEFAULT(cvScalarAll(0)) );
-
-/** @brief Computes perspective transform matrix for mapping src[i] to dst[i] (i=0,1,2,3)
-@see cv::getPerspectiveTransform
-*/
-CVAPI(CvMat*) cvGetPerspectiveTransform( const CvPoint2D32f* src,
-                                         const CvPoint2D32f* dst,
-                                         CvMat* map_matrix );
-
-/** @brief Performs generic geometric transformation using the specified coordinate maps
-@see cv::remap
-*/
-CVAPI(void)  cvRemap( const CvArr* src, CvArr* dst,
-                      const CvArr* mapx, const CvArr* mapy,
-                      int flags CV_DEFAULT(CV_INTER_LINEAR+CV_WARP_FILL_OUTLIERS),
-                      CvScalar fillval CV_DEFAULT(cvScalarAll(0)) );
-
-/** @brief Converts mapx & mapy from floating-point to integer formats for cvRemap
-@see cv::convertMaps
-*/
-CVAPI(void)  cvConvertMaps( const CvArr* mapx, const CvArr* mapy,
-                            CvArr* mapxy, CvArr* mapalpha );
-
-/** @brief Performs forward or inverse log-polar image transform
-@see cv::warpPolar
-*/
-CVAPI(void)  cvLogPolar( const CvArr* src, CvArr* dst,
-                         CvPoint2D32f center, double M,
-                         int flags CV_DEFAULT(CV_INTER_LINEAR+CV_WARP_FILL_OUTLIERS));
-
-/** Performs forward or inverse linear-polar image transform
-@see cv::warpPolar
-*/
-CVAPI(void)  cvLinearPolar( const CvArr* src, CvArr* dst,
-                         CvPoint2D32f center, double maxRadius,
-                         int flags CV_DEFAULT(CV_INTER_LINEAR+CV_WARP_FILL_OUTLIERS));
-
-/** @brief Returns a structuring element of the specified size and shape for morphological operations.
-
-@note the created structuring element IplConvKernel\* element must be released in the end using
-`cvReleaseStructuringElement(&element)`.
-
-@param cols Width of the structuring element
-@param rows Height of the structuring element
-@param anchor_x x-coordinate of the anchor
-@param anchor_y y-coordinate of the anchor
-@param shape element shape that could be one of the cv::MorphShapes_c
-@param values integer array of cols*rows elements that specifies the custom shape of the
-structuring element, when shape=CV_SHAPE_CUSTOM.
-
-@see cv::getStructuringElement
- */
- CVAPI(IplConvKernel*)  cvCreateStructuringElementEx(
-            int cols, int  rows, int  anchor_x, int  anchor_y,
-            int shape, int* values CV_DEFAULT(NULL) );
-
-/** @brief releases structuring element
-@see cvCreateStructuringElementEx
-*/
-CVAPI(void)  cvReleaseStructuringElement( IplConvKernel** element );
-
-/** @brief erodes input image (applies minimum filter) one or more times.
-   If element pointer is NULL, 3x3 rectangular element is used
-@see cv::erode
-*/
-CVAPI(void)  cvErode( const CvArr* src, CvArr* dst,
-                      IplConvKernel* element CV_DEFAULT(NULL),
-                      int iterations CV_DEFAULT(1) );
-
-/** @brief dilates input image (applies maximum filter) one or more times.
-
-   If element pointer is NULL, 3x3 rectangular element is used
-@see cv::dilate
-*/
-CVAPI(void)  cvDilate( const CvArr* src, CvArr* dst,
-                       IplConvKernel* element CV_DEFAULT(NULL),
-                       int iterations CV_DEFAULT(1) );
-
-/** @brief Performs complex morphological transformation
-@see cv::morphologyEx
-*/
-CVAPI(void)  cvMorphologyEx( const CvArr* src, CvArr* dst,
-                             CvArr* temp, IplConvKernel* element,
-                             int operation, int iterations CV_DEFAULT(1) );
-
-/** @brief Calculates all spatial and central moments up to the 3rd order
-@see cv::moments
-*/
-CVAPI(void) cvMoments( const CvArr* arr, CvMoments* moments, int binary CV_DEFAULT(0));
-
-/** @brief Retrieve spatial moments */
-CVAPI(double)  cvGetSpatialMoment( CvMoments* moments, int x_order, int y_order );
-/** @brief Retrieve central moments */
-CVAPI(double)  cvGetCentralMoment( CvMoments* moments, int x_order, int y_order );
-/** @brief Retrieve normalized central moments */
-CVAPI(double)  cvGetNormalizedCentralMoment( CvMoments* moments,
-                                             int x_order, int y_order );
-
-/** @brief Calculates 7 Hu's invariants from precalculated spatial and central moments
-@see cv::HuMoments
-*/
-CVAPI(void) cvGetHuMoments( CvMoments*  moments, CvHuMoments*  hu_moments );
-
-/*********************************** data sampling **************************************/
-
-/** @brief Fetches pixels that belong to the specified line segment and stores them to the buffer.
-
-   Returns the number of retrieved points.
-@see cv::LineSegmentDetector
-*/
-CVAPI(int)  cvSampleLine( const CvArr* image, CvPoint pt1, CvPoint pt2, void* buffer,
-                          int connectivity CV_DEFAULT(8));
-
-/** @brief Retrieves the rectangular image region with specified center from the input array.
-
- dst(x,y) <- src(x + center.x - dst_width/2, y + center.y - dst_height/2).
- Values of pixels with fractional coordinates are retrieved using bilinear interpolation
-@see cv::getRectSubPix
-*/
-CVAPI(void)  cvGetRectSubPix( const CvArr* src, CvArr* dst, CvPoint2D32f center );
-
-
-/** @brief Retrieves quadrangle from the input array.
-
-    matrixarr = ( a11  a12 | b1 )   dst(x,y) <- src(A[x y]' + b)
-                ( a21  a22 | b2 )   (bilinear interpolation is used to retrieve pixels
-                                     with fractional coordinates)
-@see cvWarpAffine
-*/
-CVAPI(void)  cvGetQuadrangleSubPix( const CvArr* src, CvArr* dst,
-                                    const CvMat* map_matrix );
-
-/** @brief Measures similarity between template and overlapped windows in the source image
-   and fills the resultant image with the measurements
-@see cv::matchTemplate
-*/
-CVAPI(void)  cvMatchTemplate( const CvArr* image, const CvArr* templ,
-                              CvArr* result, int method );
-
-/** @brief Computes earth mover distance between
-   two weighted point sets (called signatures)
-@see cv::EMD
-*/
-CVAPI(float)  cvCalcEMD2( const CvArr* signature1,
-                          const CvArr* signature2,
-                          int distance_type,
-                          CvDistanceFunction distance_func CV_DEFAULT(NULL),
-                          const CvArr* cost_matrix CV_DEFAULT(NULL),
-                          CvArr* flow CV_DEFAULT(NULL),
-                          float* lower_bound CV_DEFAULT(NULL),
-                          void* userdata CV_DEFAULT(NULL));
-
-/****************************************************************************************\
-*                              Contours retrieving                                       *
-\****************************************************************************************/
-
-/** @brief Retrieves outer and optionally inner boundaries of white (non-zero) connected
-   components in the black (zero) background
-@see cv::findContours, cvStartFindContours, cvFindNextContour, cvSubstituteContour, cvEndFindContours
-*/
-CVAPI(int)  cvFindContours( CvArr* image, CvMemStorage* storage, CvSeq** first_contour,
-                            int header_size CV_DEFAULT(sizeof(CvContour)),
-                            int mode CV_DEFAULT(CV_RETR_LIST),
-                            int method CV_DEFAULT(CV_CHAIN_APPROX_SIMPLE),
-                            CvPoint offset CV_DEFAULT(cvPoint(0,0)));
-
-/** @brief Initializes contour retrieving process.
-
-   Calls cvStartFindContours.
-   Calls cvFindNextContour until null pointer is returned
-   or some other condition becomes true.
-   Calls cvEndFindContours at the end.
-@see cvFindContours
-*/
-CVAPI(CvContourScanner)  cvStartFindContours( CvArr* image, CvMemStorage* storage,
-                            int header_size CV_DEFAULT(sizeof(CvContour)),
-                            int mode CV_DEFAULT(CV_RETR_LIST),
-                            int method CV_DEFAULT(CV_CHAIN_APPROX_SIMPLE),
-                            CvPoint offset CV_DEFAULT(cvPoint(0,0)));
-
-/** @brief Retrieves next contour
-@see cvFindContours
-*/
-CVAPI(CvSeq*)  cvFindNextContour( CvContourScanner scanner );
-
-
-/** @brief Substitutes the last retrieved contour with the new one
-
-   (if the substitutor is null, the last retrieved contour is removed from the tree)
-@see cvFindContours
-*/
-CVAPI(void)   cvSubstituteContour( CvContourScanner scanner, CvSeq* new_contour );
-
-
-/** @brief Releases contour scanner and returns pointer to the first outer contour
-@see cvFindContours
-*/
-CVAPI(CvSeq*)  cvEndFindContours( CvContourScanner* scanner );
-
-/** @brief Approximates Freeman chain(s) with a polygonal curve.
-
-This is a standalone contour approximation routine, not represented in the new interface. When
-cvFindContours retrieves contours as Freeman chains, it calls the function to get approximated
-contours, represented as polygons.
-
-@param src_seq Pointer to the approximated Freeman chain that can refer to other chains.
-@param storage Storage location for the resulting polylines.
-@param method Approximation method (see the description of the function :ocvFindContours ).
-@param parameter Method parameter (not used now).
-@param minimal_perimeter Approximates only those contours whose perimeters are not less than
-minimal_perimeter . Other chains are removed from the resulting structure.
-@param recursive Recursion flag. If it is non-zero, the function approximates all chains that can
-be obtained from chain by using the h_next or v_next links. Otherwise, the single input chain is
-approximated.
-@see cvStartReadChainPoints, cvReadChainPoint
- */
-CVAPI(CvSeq*) cvApproxChains( CvSeq* src_seq, CvMemStorage* storage,
-                            int method CV_DEFAULT(CV_CHAIN_APPROX_SIMPLE),
-                            double parameter CV_DEFAULT(0),
-                            int  minimal_perimeter CV_DEFAULT(0),
-                            int  recursive CV_DEFAULT(0));
-
-/** @brief Initializes Freeman chain reader.
-
-   The reader is used to iteratively get coordinates of all the chain points.
-   If the Freeman codes should be read as is, a simple sequence reader should be used
-@see cvApproxChains
-*/
-CVAPI(void) cvStartReadChainPoints( CvChain* chain, CvChainPtReader* reader );
-
-/** @brief Retrieves the next chain point
-@see cvApproxChains
-*/
-CVAPI(CvPoint) cvReadChainPoint( CvChainPtReader* reader );
-
-
-/****************************************************************************************\
-*                            Contour Processing and Shape Analysis                       *
-\****************************************************************************************/
-
-/** @brief Approximates a single polygonal curve (contour) or
-   a tree of polygonal curves (contours)
-@see cv::approxPolyDP
-*/
-CVAPI(CvSeq*)  cvApproxPoly( const void* src_seq,
-                             int header_size, CvMemStorage* storage,
-                             int method, double eps,
-                             int recursive CV_DEFAULT(0));
-
-/** @brief Calculates perimeter of a contour or length of a part of contour
-@see cv::arcLength
-*/
-CVAPI(double)  cvArcLength( const void* curve,
-                            CvSlice slice CV_DEFAULT(CV_WHOLE_SEQ),
-                            int is_closed CV_DEFAULT(-1));
-
-/** same as cvArcLength for closed contour
-*/
-CV_INLINE double cvContourPerimeter( const void* contour )
-{
-    return cvArcLength( contour, CV_WHOLE_SEQ, 1 );
-}
-
-
-/** @brief Calculates contour bounding rectangle (update=1) or
-   just retrieves pre-calculated rectangle (update=0)
-@see cv::boundingRect
-*/
-CVAPI(CvRect)  cvBoundingRect( CvArr* points, int update CV_DEFAULT(0) );
-
-/** @brief Calculates area of a contour or contour segment
-@see cv::contourArea
-*/
-CVAPI(double)  cvContourArea( const CvArr* contour,
-                              CvSlice slice CV_DEFAULT(CV_WHOLE_SEQ),
-                              int oriented CV_DEFAULT(0));
-
-/** @brief Finds minimum area rotated rectangle bounding a set of points
-@see cv::minAreaRect
-*/
-CVAPI(CvBox2D)  cvMinAreaRect2( const CvArr* points,
-                                CvMemStorage* storage CV_DEFAULT(NULL));
-
-/** @brief Finds minimum enclosing circle for a set of points
-@see cv::minEnclosingCircle
-*/
-CVAPI(int)  cvMinEnclosingCircle( const CvArr* points,
-                                  CvPoint2D32f* center, float* radius );
-
-/** @brief Compares two contours by matching their moments
-@see cv::matchShapes
-*/
-CVAPI(double)  cvMatchShapes( const void* object1, const void* object2,
-                              int method, double parameter CV_DEFAULT(0));
-
-/** @brief Calculates exact convex hull of 2d point set
-@see cv::convexHull
-*/
-CVAPI(CvSeq*) cvConvexHull2( const CvArr* input,
-                             void* hull_storage CV_DEFAULT(NULL),
-                             int orientation CV_DEFAULT(CV_CLOCKWISE),
-                             int return_points CV_DEFAULT(0));
-
-/** @brief Checks whether the contour is convex or not (returns 1 if convex, 0 if not)
-@see cv::isContourConvex
-*/
-CVAPI(int)  cvCheckContourConvexity( const CvArr* contour );
-
-
-/** @brief Finds convexity defects for the contour
-@see cv::convexityDefects
-*/
-CVAPI(CvSeq*)  cvConvexityDefects( const CvArr* contour, const CvArr* convexhull,
-                                   CvMemStorage* storage CV_DEFAULT(NULL));
-
-/** @brief Fits ellipse into a set of 2d points
-@see cv::fitEllipse
-*/
-CVAPI(CvBox2D) cvFitEllipse2( const CvArr* points );
-
-/** @brief Finds minimum rectangle containing two given rectangles */
-CVAPI(CvRect)  cvMaxRect( const CvRect* rect1, const CvRect* rect2 );
-
-/** @brief Finds coordinates of the box vertices */
-CVAPI(void) cvBoxPoints( CvBox2D box, CvPoint2D32f pt[4] );
-
-/** @brief Initializes sequence header for a matrix (column or row vector) of points
-
-   a wrapper for cvMakeSeqHeaderForArray (it does not initialize bounding rectangle!!!) */
-CVAPI(CvSeq*) cvPointSeqFromMat( int seq_kind, const CvArr* mat,
-                                 CvContour* contour_header,
-                                 CvSeqBlock* block );
-
-/** @brief Checks whether the point is inside polygon, outside, on an edge (at a vertex).
-
-   Returns positive, negative or zero value, correspondingly.
-   Optionally, measures a signed distance between
-   the point and the nearest polygon edge (measure_dist=1)
-@see cv::pointPolygonTest
-*/
-CVAPI(double) cvPointPolygonTest( const CvArr* contour,
-                                  CvPoint2D32f pt, int measure_dist );
-
-/****************************************************************************************\
-*                                  Histogram functions                                   *
-\****************************************************************************************/
-
-/** @brief Creates a histogram.
-
-The function creates a histogram of the specified size and returns a pointer to the created
-histogram. If the array ranges is 0, the histogram bin ranges must be specified later via the
-function cvSetHistBinRanges. Though cvCalcHist and cvCalcBackProject may process 8-bit images
-without setting bin ranges, they assume they are equally spaced in 0 to 255 bins.
-
-@param dims Number of histogram dimensions.
-@param sizes Array of the histogram dimension sizes.
-@param type Histogram representation format. CV_HIST_ARRAY means that the histogram data is
-represented as a multi-dimensional dense array CvMatND. CV_HIST_SPARSE means that histogram data
-is represented as a multi-dimensional sparse array CvSparseMat.
-@param ranges Array of ranges for the histogram bins. Its meaning depends on the uniform parameter
-value. The ranges are used when the histogram is calculated or backprojected to determine which
-histogram bin corresponds to which value/tuple of values from the input image(s).
-@param uniform Uniformity flag. If not zero, the histogram has evenly spaced bins and for every
-\f$0<=i<cDims\f$ ranges[i] is an array of two numbers: lower and upper boundaries for the i-th
-histogram dimension. The whole range [lower,upper] is then split into dims[i] equal parts to
-determine the i-th input tuple value ranges for every histogram bin. And if uniform=0 , then the
-i-th element of the ranges array contains dims[i]+1 elements: \f$\texttt{lower}_0,
-\texttt{upper}_0, \texttt{lower}_1, \texttt{upper}_1 = \texttt{lower}_2,
-...
-\texttt{upper}_{dims[i]-1}\f$ where \f$\texttt{lower}_j\f$ and \f$\texttt{upper}_j\f$ are lower
-and upper boundaries of the i-th input tuple value for the j-th bin, respectively. In either
-case, the input values that are beyond the specified range for a histogram bin are not counted
-by cvCalcHist and filled with 0 by cvCalcBackProject.
- */
-CVAPI(CvHistogram*)  cvCreateHist( int dims, int* sizes, int type,
-                                   float** ranges CV_DEFAULT(NULL),
-                                   int uniform CV_DEFAULT(1));
-
-/** @brief Sets the bounds of the histogram bins.
-
-This is a standalone function for setting bin ranges in the histogram. For a more detailed
-description of the parameters ranges and uniform, see the :ocvCalcHist function that can initialize
-the ranges as well. Ranges for the histogram bins must be set before the histogram is calculated or
-the backproject of the histogram is calculated.
-
-@param hist Histogram.
-@param ranges Array of bin ranges arrays. See :ocvCreateHist for details.
-@param uniform Uniformity flag. See :ocvCreateHist for details.
- */
-CVAPI(void)  cvSetHistBinRanges( CvHistogram* hist, float** ranges,
-                                int uniform CV_DEFAULT(1));
-
-/** @brief Makes a histogram out of an array.
-
-The function initializes the histogram, whose header and bins are allocated by the user.
-cvReleaseHist does not need to be called afterwards. Only dense histograms can be initialized this
-way. The function returns hist.
-
-@param dims Number of the histogram dimensions.
-@param sizes Array of the histogram dimension sizes.
-@param hist Histogram header initialized by the function.
-@param data Array used to store histogram bins.
-@param ranges Histogram bin ranges. See cvCreateHist for details.
-@param uniform Uniformity flag. See cvCreateHist for details.
- */
-CVAPI(CvHistogram*)  cvMakeHistHeaderForArray(
-                            int  dims, int* sizes, CvHistogram* hist,
-                            float* data, float** ranges CV_DEFAULT(NULL),
-                            int uniform CV_DEFAULT(1));
-
-/** @brief Releases the histogram.
-
-The function releases the histogram (header and the data). The pointer to the histogram is cleared
-by the function. If \*hist pointer is already NULL, the function does nothing.
-
-@param hist Double pointer to the released histogram.
- */
-CVAPI(void)  cvReleaseHist( CvHistogram** hist );
-
-/** @brief Clears the histogram.
-
-The function sets all of the histogram bins to 0 in case of a dense histogram and removes all
-histogram bins in case of a sparse array.
-
-@param hist Histogram.
- */
-CVAPI(void)  cvClearHist( CvHistogram* hist );
-
-/** @brief Finds the minimum and maximum histogram bins.
-
-The function finds the minimum and maximum histogram bins and their positions. All of output
-arguments are optional. Among several extremas with the same value the ones with the minimum index
-(in the lexicographical order) are returned. In case of several maximums or minimums, the earliest
-in the lexicographical order (extrema locations) is returned.
-
-@param hist Histogram.
-@param min_value Pointer to the minimum value of the histogram.
-@param max_value Pointer to the maximum value of the histogram.
-@param min_idx Pointer to the array of coordinates for the minimum.
-@param max_idx Pointer to the array of coordinates for the maximum.
- */
-CVAPI(void)  cvGetMinMaxHistValue( const CvHistogram* hist,
-                                   float* min_value, float* max_value,
-                                   int* min_idx CV_DEFAULT(NULL),
-                                   int* max_idx CV_DEFAULT(NULL));
-
-
-/** @brief Normalizes the histogram.
-
-The function normalizes the histogram bins by scaling them so that the sum of the bins becomes equal
-to factor.
-
-@param hist Pointer to the histogram.
-@param factor Normalization factor.
- */
-CVAPI(void)  cvNormalizeHist( CvHistogram* hist, double factor );
-
-
-/** @brief Thresholds the histogram.
-
-The function clears histogram bins that are below the specified threshold.
-
-@param hist Pointer to the histogram.
-@param threshold Threshold level.
- */
-CVAPI(void)  cvThreshHist( CvHistogram* hist, double threshold );
-
-
-/** Compares two histogram */
-CVAPI(double)  cvCompareHist( const CvHistogram* hist1,
-                              const CvHistogram* hist2,
-                              int method);
-
-/** @brief Copies a histogram.
-
-The function makes a copy of the histogram. If the second histogram pointer \*dst is NULL, a new
-histogram of the same size as src is created. Otherwise, both histograms must have equal types and
-sizes. Then the function copies the bin values of the source histogram to the destination histogram
-and sets the same bin value ranges as in src.
-
-@param src Source histogram.
-@param dst Pointer to the destination histogram.
- */
-CVAPI(void)  cvCopyHist( const CvHistogram* src, CvHistogram** dst );
-
-
-/** @brief Calculates bayesian probabilistic histograms
-   (each or src and dst is an array of _number_ histograms */
-CVAPI(void)  cvCalcBayesianProb( CvHistogram** src, int number,
-                                CvHistogram** dst);
-
-/** @brief Calculates array histogram
-@see cv::calcHist
-*/
-CVAPI(void)  cvCalcArrHist( CvArr** arr, CvHistogram* hist,
-                            int accumulate CV_DEFAULT(0),
-                            const CvArr* mask CV_DEFAULT(NULL) );
-
-/** @overload */
-CV_INLINE  void  cvCalcHist( IplImage** image, CvHistogram* hist,
-                             int accumulate CV_DEFAULT(0),
-                             const CvArr* mask CV_DEFAULT(NULL) )
-{
-    cvCalcArrHist( (CvArr**)image, hist, accumulate, mask );
-}
-
-/** @brief Calculates back project
-@see cvCalcBackProject, cv::calcBackProject
-*/
-CVAPI(void)  cvCalcArrBackProject( CvArr** image, CvArr* dst,
-                                   const CvHistogram* hist );
-
-#define  cvCalcBackProject(image, dst, hist) cvCalcArrBackProject((CvArr**)image, dst, hist)
-
-
-/** @brief Locates a template within an image by using a histogram comparison.
-
-The function calculates the back projection by comparing histograms of the source image patches with
-the given histogram. The function is similar to matchTemplate, but instead of comparing the raster
-patch with all its possible positions within the search window, the function CalcBackProjectPatch
-compares histograms. See the algorithm diagram below:
-
-![image](pics/backprojectpatch.png)
-
-@param image Source images (though, you may pass CvMat\*\* as well).
-@param dst Destination image.
-@param range
-@param hist Histogram.
-@param method Comparison method passed to cvCompareHist (see the function description).
-@param factor Normalization factor for histograms that affects the normalization scale of the
-destination image. Pass 1 if not sure.
-
-@see cvCalcBackProjectPatch
- */
-CVAPI(void)  cvCalcArrBackProjectPatch( CvArr** image, CvArr* dst, CvSize range,
-                                        CvHistogram* hist, int method,
-                                        double factor );
-
-#define  cvCalcBackProjectPatch( image, dst, range, hist, method, factor ) \
-     cvCalcArrBackProjectPatch( (CvArr**)image, dst, range, hist, method, factor )
-
-
-/** @brief Divides one histogram by another.
-
-The function calculates the object probability density from two histograms as:
-
-\f[\texttt{disthist} (I)= \forkthree{0}{if \(\texttt{hist1}(I)=0\)}{\texttt{scale}}{if \(\texttt{hist1}(I) \ne 0\) and \(\texttt{hist2}(I) > \texttt{hist1}(I)\)}{\frac{\texttt{hist2}(I) \cdot \texttt{scale}}{\texttt{hist1}(I)}}{if \(\texttt{hist1}(I) \ne 0\) and \(\texttt{hist2}(I) \le \texttt{hist1}(I)\)}\f]
-
-@param hist1 First histogram (the divisor).
-@param hist2 Second histogram.
-@param dst_hist Destination histogram.
-@param scale Scale factor for the destination histogram.
- */
-CVAPI(void)  cvCalcProbDensity( const CvHistogram* hist1, const CvHistogram* hist2,
-                                CvHistogram* dst_hist, double scale CV_DEFAULT(255) );
-
-/** @brief equalizes histogram of 8-bit single-channel image
-@see cv::equalizeHist
-*/
-CVAPI(void)  cvEqualizeHist( const CvArr* src, CvArr* dst );
-
-
-/** @brief Applies distance transform to binary image
-@see cv::distanceTransform
-*/
-CVAPI(void)  cvDistTransform( const CvArr* src, CvArr* dst,
-                              int distance_type CV_DEFAULT(CV_DIST_L2),
-                              int mask_size CV_DEFAULT(3),
-                              const float* mask CV_DEFAULT(NULL),
-                              CvArr* labels CV_DEFAULT(NULL),
-                              int labelType CV_DEFAULT(CV_DIST_LABEL_CCOMP));
-
-
-/** @brief Applies fixed-level threshold to grayscale image.
-
-   This is a basic operation applied before retrieving contours
-@see cv::threshold
-*/
-CVAPI(double)  cvThreshold( const CvArr*  src, CvArr*  dst,
-                            double  threshold, double  max_value,
-                            int threshold_type );
-
-/** @brief Applies adaptive threshold to grayscale image.
-
-   The two parameters for methods CV_ADAPTIVE_THRESH_MEAN_C and
-   CV_ADAPTIVE_THRESH_GAUSSIAN_C are:
-   neighborhood size (3, 5, 7 etc.),
-   and a constant subtracted from mean (...,-3,-2,-1,0,1,2,3,...)
-@see cv::adaptiveThreshold
-*/
-CVAPI(void)  cvAdaptiveThreshold( const CvArr* src, CvArr* dst, double max_value,
-                                  int adaptive_method CV_DEFAULT(CV_ADAPTIVE_THRESH_MEAN_C),
-                                  int threshold_type CV_DEFAULT(CV_THRESH_BINARY),
-                                  int block_size CV_DEFAULT(3),
-                                  double param1 CV_DEFAULT(5));
-
-/** @brief Fills the connected component until the color difference gets large enough
-@see cv::floodFill
-*/
-CVAPI(void)  cvFloodFill( CvArr* image, CvPoint seed_point,
-                          CvScalar new_val, CvScalar lo_diff CV_DEFAULT(cvScalarAll(0)),
-                          CvScalar up_diff CV_DEFAULT(cvScalarAll(0)),
-                          CvConnectedComp* comp CV_DEFAULT(NULL),
-                          int flags CV_DEFAULT(4),
-                          CvArr* mask CV_DEFAULT(NULL));
-
-/****************************************************************************************\
-*                                  Feature detection                                     *
-\****************************************************************************************/
-
-/** @brief Runs canny edge detector
-@see cv::Canny
-*/
-CVAPI(void)  cvCanny( const CvArr* image, CvArr* edges, double threshold1,
-                      double threshold2, int  aperture_size CV_DEFAULT(3) );
-
-/** @brief Calculates constraint image for corner detection
-
-   Dx^2 * Dyy + Dxx * Dy^2 - 2 * Dx * Dy * Dxy.
-   Applying threshold to the result gives coordinates of corners
-@see cv::preCornerDetect
-*/
-CVAPI(void) cvPreCornerDetect( const CvArr* image, CvArr* corners,
-                               int aperture_size CV_DEFAULT(3) );
-
-/** @brief Calculates eigen values and vectors of 2x2
-   gradient covariation matrix at every image pixel
-@see cv::cornerEigenValsAndVecs
-*/
-CVAPI(void)  cvCornerEigenValsAndVecs( const CvArr* image, CvArr* eigenvv,
-                                       int block_size, int aperture_size CV_DEFAULT(3) );
-
-/** @brief Calculates minimal eigenvalue for 2x2 gradient covariation matrix at
-   every image pixel
-@see cv::cornerMinEigenVal
-*/
-CVAPI(void)  cvCornerMinEigenVal( const CvArr* image, CvArr* eigenval,
-                                  int block_size, int aperture_size CV_DEFAULT(3) );
-
-/** @brief Harris corner detector:
-
-   Calculates det(M) - k*(trace(M)^2), where M is 2x2 gradient covariation matrix for each pixel
-@see cv::cornerHarris
-*/
-CVAPI(void)  cvCornerHarris( const CvArr* image, CvArr* harris_response,
-                             int block_size, int aperture_size CV_DEFAULT(3),
-                             double k CV_DEFAULT(0.04) );
-
-/** @brief Adjust corner position using some sort of gradient search
-@see cv::cornerSubPix
-*/
-CVAPI(void)  cvFindCornerSubPix( const CvArr* image, CvPoint2D32f* corners,
-                                 int count, CvSize win, CvSize zero_zone,
-                                 CvTermCriteria  criteria );
-
-/** @brief Finds a sparse set of points within the selected region
-   that seem to be easy to track
-@see cv::goodFeaturesToTrack
-*/
-CVAPI(void)  cvGoodFeaturesToTrack( const CvArr* image, CvArr* eig_image,
-                                    CvArr* temp_image, CvPoint2D32f* corners,
-                                    int* corner_count, double  quality_level,
-                                    double  min_distance,
-                                    const CvArr* mask CV_DEFAULT(NULL),
-                                    int block_size CV_DEFAULT(3),
-                                    int use_harris CV_DEFAULT(0),
-                                    double k CV_DEFAULT(0.04) );
-
-/** @brief Finds lines on binary image using one of several methods.
-
-   line_storage is either memory storage or 1 x _max number of lines_ CvMat, its
-   number of columns is changed by the function.
-   method is one of CV_HOUGH_*;
-   rho, theta and threshold are used for each of those methods;
-   param1 ~ line length, param2 ~ line gap - for probabilistic,
-   param1 ~ srn, param2 ~ stn - for multi-scale
-@see cv::HoughLines
-*/
-CVAPI(CvSeq*)  cvHoughLines2( CvArr* image, void* line_storage, int method,
-                              double rho, double theta, int threshold,
-                              double param1 CV_DEFAULT(0), double param2 CV_DEFAULT(0),
-                              double min_theta CV_DEFAULT(0), double max_theta CV_DEFAULT(CV_PI));
-
-/** @brief Finds circles in the image
-@see cv::HoughCircles
-*/
-CVAPI(CvSeq*) cvHoughCircles( CvArr* image, void* circle_storage,
-                              int method, double dp, double min_dist,
-                              double param1 CV_DEFAULT(100),
-                              double param2 CV_DEFAULT(100),
-                              int min_radius CV_DEFAULT(0),
-                              int max_radius CV_DEFAULT(0));
-
-/** @brief Fits a line into set of 2d or 3d points in a robust way (M-estimator technique)
-@see cv::fitLine
-*/
-CVAPI(void)  cvFitLine( const CvArr* points, int dist_type, double param,
-                        double reps, double aeps, float* line );
-
-/****************************************************************************************\
-*                                     Drawing                                            *
-\****************************************************************************************/
-
-/****************************************************************************************\
-*       Drawing functions work with images/matrices of arbitrary type.                   *
-*       For color images the channel order is BGR[A]                                     *
-*       Antialiasing is supported only for 8-bit image now.                              *
-*       All the functions include parameter color that means rgb value (that may be      *
-*       constructed with CV_RGB macro) for color images and brightness                   *
-*       for grayscale images.                                                            *
-*       If a drawn figure is partially or completely outside of the image, it is clipped.*
-\****************************************************************************************/
-
-#define CV_FILLED -1
-
-#define CV_AA 16
-
-/** @brief Draws 4-connected, 8-connected or antialiased line segment connecting two points
-@see cv::line
-*/
-CVAPI(void)  cvLine( CvArr* img, CvPoint pt1, CvPoint pt2,
-                     CvScalar color, int thickness CV_DEFAULT(1),
-                     int line_type CV_DEFAULT(8), int shift CV_DEFAULT(0) );
-
-/** @brief Draws a rectangle given two opposite corners of the rectangle (pt1 & pt2)
-
-   if thickness<0 (e.g. thickness == CV_FILLED), the filled box is drawn
-@see cv::rectangle
-*/
-CVAPI(void)  cvRectangle( CvArr* img, CvPoint pt1, CvPoint pt2,
-                          CvScalar color, int thickness CV_DEFAULT(1),
-                          int line_type CV_DEFAULT(8),
-                          int shift CV_DEFAULT(0));
-
-/** @brief Draws a rectangle specified by a CvRect structure
-@see cv::rectangle
-*/
-CVAPI(void)  cvRectangleR( CvArr* img, CvRect r,
-                           CvScalar color, int thickness CV_DEFAULT(1),
-                           int line_type CV_DEFAULT(8),
-                           int shift CV_DEFAULT(0));
-
-
-/** @brief Draws a circle with specified center and radius.
-
-   Thickness works in the same way as with cvRectangle
-@see cv::circle
-*/
-CVAPI(void)  cvCircle( CvArr* img, CvPoint center, int radius,
-                       CvScalar color, int thickness CV_DEFAULT(1),
-                       int line_type CV_DEFAULT(8), int shift CV_DEFAULT(0));
-
-/** @brief Draws ellipse outline, filled ellipse, elliptic arc or filled elliptic sector
-
-   depending on _thickness_, _start_angle_ and _end_angle_ parameters. The resultant figure
-   is rotated by _angle_. All the angles are in degrees
-@see cv::ellipse
-*/
-CVAPI(void)  cvEllipse( CvArr* img, CvPoint center, CvSize axes,
-                        double angle, double start_angle, double end_angle,
-                        CvScalar color, int thickness CV_DEFAULT(1),
-                        int line_type CV_DEFAULT(8), int shift CV_DEFAULT(0));
-
-CV_INLINE  void  cvEllipseBox( CvArr* img, CvBox2D box, CvScalar color,
-                               int thickness CV_DEFAULT(1),
-                               int line_type CV_DEFAULT(8), int shift CV_DEFAULT(0) )
-{
-    CvSize axes = cvSize(
-        cvRound(box.size.width*0.5),
-        cvRound(box.size.height*0.5)
-    );
-
-    cvEllipse( img, cvPointFrom32f( box.center ), axes, box.angle,
-               0, 360, color, thickness, line_type, shift );
-}
-
-/** @brief Fills convex or monotonous polygon.
-@see cv::fillConvexPoly
-*/
-CVAPI(void)  cvFillConvexPoly( CvArr* img, const CvPoint* pts, int npts, CvScalar color,
-                               int line_type CV_DEFAULT(8), int shift CV_DEFAULT(0));
-
-/** @brief Fills an area bounded by one or more arbitrary polygons
-@see cv::fillPoly
-*/
-CVAPI(void)  cvFillPoly( CvArr* img, CvPoint** pts, const int* npts,
-                         int contours, CvScalar color,
-                         int line_type CV_DEFAULT(8), int shift CV_DEFAULT(0) );
-
-/** @brief Draws one or more polygonal curves
-@see cv::polylines
-*/
-CVAPI(void)  cvPolyLine( CvArr* img, CvPoint** pts, const int* npts, int contours,
-                         int is_closed, CvScalar color, int thickness CV_DEFAULT(1),
-                         int line_type CV_DEFAULT(8), int shift CV_DEFAULT(0) );
-
-#define cvDrawRect cvRectangle
-#define cvDrawLine cvLine
-#define cvDrawCircle cvCircle
-#define cvDrawEllipse cvEllipse
-#define cvDrawPolyLine cvPolyLine
-
-/** @brief Clips the line segment connecting *pt1 and *pt2
-   by the rectangular window
-
-   (0<=x<img_size.width, 0<=y<img_size.height).
-@see cv::clipLine
-*/
-CVAPI(int) cvClipLine( CvSize img_size, CvPoint* pt1, CvPoint* pt2 );
-
-/** @brief Initializes line iterator.
-
-Initially, line_iterator->ptr will point to pt1 (or pt2, see left_to_right description) location in
-the image. Returns the number of pixels on the line between the ending points.
-@see cv::LineIterator
-*/
-CVAPI(int)  cvInitLineIterator( const CvArr* image, CvPoint pt1, CvPoint pt2,
-                                CvLineIterator* line_iterator,
-                                int connectivity CV_DEFAULT(8),
-                                int left_to_right CV_DEFAULT(0));
-
-#define CV_NEXT_LINE_POINT( line_iterator )                     \
-{                                                               \
-    int _line_iterator_mask = (line_iterator).err < 0 ? -1 : 0; \
-    (line_iterator).err += (line_iterator).minus_delta +        \
-        ((line_iterator).plus_delta & _line_iterator_mask);     \
-    (line_iterator).ptr += (line_iterator).minus_step +         \
-        ((line_iterator).plus_step & _line_iterator_mask);      \
-}
-
-
-#define CV_FONT_HERSHEY_SIMPLEX         0
-#define CV_FONT_HERSHEY_PLAIN           1
-#define CV_FONT_HERSHEY_DUPLEX          2
-#define CV_FONT_HERSHEY_COMPLEX         3
-#define CV_FONT_HERSHEY_TRIPLEX         4
-#define CV_FONT_HERSHEY_COMPLEX_SMALL   5
-#define CV_FONT_HERSHEY_SCRIPT_SIMPLEX  6
-#define CV_FONT_HERSHEY_SCRIPT_COMPLEX  7
-
-#define CV_FONT_ITALIC                 16
-
-#define CV_FONT_VECTOR0    CV_FONT_HERSHEY_SIMPLEX
-
-
-/** Font structure */
-typedef struct CvFont
-{
-  const char* nameFont;   //Qt:nameFont
-  CvScalar color;       //Qt:ColorFont -> cvScalar(blue_component, green_component, red_component[, alpha_component])
-    int         font_face;    //Qt: bool italic         /** =CV_FONT_* */
-    const int*  ascii;      //!< font data and metrics
-    const int*  greek;
-    const int*  cyrillic;
-    float       hscale, vscale;
-    float       shear;      //!< slope coefficient: 0 - normal, >0 - italic
-    int         thickness;    //!< Qt: weight               /** letters thickness */
-    float       dx;       //!< horizontal interval between letters
-    int         line_type;    //!< Qt: PointSize
-}
-CvFont;
-
-/** @brief Initializes font structure (OpenCV 1.x API).
-
-The function initializes the font structure that can be passed to text rendering functions.
-
-@param font Pointer to the font structure initialized by the function
-@param font_face Font name identifier. See cv::HersheyFonts and corresponding old CV_* identifiers.
-@param hscale Horizontal scale. If equal to 1.0f , the characters have the original width
-depending on the font type. If equal to 0.5f , the characters are of half the original width.
-@param vscale Vertical scale. If equal to 1.0f , the characters have the original height depending
-on the font type. If equal to 0.5f , the characters are of half the original height.
-@param shear Approximate tangent of the character slope relative to the vertical line. A zero
-value means a non-italic font, 1.0f means about a 45 degree slope, etc.
-@param thickness Thickness of the text strokes
-@param line_type Type of the strokes, see line description
-
-@sa cvPutText
- */
-CVAPI(void)  cvInitFont( CvFont* font, int font_face,
-                         double hscale, double vscale,
-                         double shear CV_DEFAULT(0),
-                         int thickness CV_DEFAULT(1),
-                         int line_type CV_DEFAULT(8));
-
-CV_INLINE CvFont cvFont( double scale, int thickness CV_DEFAULT(1) )
-{
-    CvFont font;
-    cvInitFont( &font, CV_FONT_HERSHEY_PLAIN, scale, scale, 0, thickness, CV_AA );
-    return font;
-}
-
-/** @brief Renders text stroke with specified font and color at specified location.
-   CvFont should be initialized with cvInitFont
-@see cvInitFont, cvGetTextSize, cvFont, cv::putText
-*/
-CVAPI(void)  cvPutText( CvArr* img, const char* text, CvPoint org,
-                        const CvFont* font, CvScalar color );
-
-/** @brief Calculates bounding box of text stroke (useful for alignment)
-@see cv::getTextSize
-*/
-CVAPI(void)  cvGetTextSize( const char* text_string, const CvFont* font,
-                            CvSize* text_size, int* baseline );
-
-/** @brief Unpacks color value
-
-if arrtype is CV_8UC?, _color_ is treated as packed color value, otherwise the first channels
-(depending on arrtype) of destination scalar are set to the same value = _color_
-*/
-CVAPI(CvScalar)  cvColorToScalar( double packed_color, int arrtype );
-
-/** @brief Returns the polygon points which make up the given ellipse.
-
-The ellipse is define by the box of size 'axes' rotated 'angle' around the 'center'. A partial
-sweep of the ellipse arc can be done by specifying arc_start and arc_end to be something other than
-0 and 360, respectively. The input array 'pts' must be large enough to hold the result. The total
-number of points stored into 'pts' is returned by this function.
-@see cv::ellipse2Poly
-*/
-CVAPI(int) cvEllipse2Poly( CvPoint center, CvSize axes,
-                 int angle, int arc_start, int arc_end, CvPoint * pts, int delta );
-
-/** @brief Draws contour outlines or filled interiors on the image
-@see cv::drawContours
-*/
-CVAPI(void)  cvDrawContours( CvArr *img, CvSeq* contour,
-                             CvScalar external_color, CvScalar hole_color,
-                             int max_level, int thickness CV_DEFAULT(1),
-                             int line_type CV_DEFAULT(8),
-                             CvPoint offset CV_DEFAULT(cvPoint(0,0)));
-
-/** @} */
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/imgproc/segmentation.hpp b/projects/llm_framework/include/opencv4/opencv2/imgproc/segmentation.hpp
deleted file mode 100644
index 26882f44..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/imgproc/segmentation.hpp
+++ /dev/null
@@ -1,141 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_IMGPROC_SEGMENTATION_HPP
-#define OPENCV_IMGPROC_SEGMENTATION_HPP
-
-#include "opencv2/imgproc.hpp"
-
-namespace cv {
-
-namespace segmentation {
-
-//! @addtogroup imgproc_segmentation
-//! @{
-
-
-/** @brief Intelligent Scissors image segmentation
- *
- * This class is used to find the path (contour) between two points
- * which can be used for image segmentation.
- *
- * Usage example:
- * @snippet snippets/imgproc_segmentation.cpp usage_example_intelligent_scissors
- *
- * Reference: <a href="http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.138.3811&rep=rep1&type=pdf">"Intelligent Scissors for Image Composition"</a>
- * algorithm designed by Eric N. Mortensen and William A. Barrett, Brigham Young University
- * @cite Mortensen95intelligentscissors
- */
-class CV_EXPORTS_W_SIMPLE IntelligentScissorsMB
-{
-public:
-    CV_WRAP
-    IntelligentScissorsMB();
-
-    /** @brief Specify weights of feature functions
-     *
-     * Consider keeping weights normalized (sum of weights equals to 1.0)
-     * Discrete dynamic programming (DP) goal is minimization of costs between pixels.
-     *
-     * @param weight_non_edge Specify cost of non-edge pixels (default: 0.43f)
-     * @param weight_gradient_direction Specify cost of gradient direction function (default: 0.43f)
-     * @param weight_gradient_magnitude Specify cost of gradient magnitude function (default: 0.14f)
-     */
-    CV_WRAP
-    IntelligentScissorsMB& setWeights(float weight_non_edge, float weight_gradient_direction, float weight_gradient_magnitude);
-
-    /** @brief Specify gradient magnitude max value threshold
-     *
-     * Zero limit value is used to disable gradient magnitude thresholding (default behavior, as described in original article).
-     * Otherwize pixels with `gradient magnitude >= threshold` have zero cost.
-     *
-     * @note Thresholding should be used for images with irregular regions (to avoid stuck on parameters from high-contract areas, like embedded logos).
-     *
-     * @param gradient_magnitude_threshold_max Specify gradient magnitude max value threshold (default: 0, disabled)
-     */
-    CV_WRAP
-    IntelligentScissorsMB& setGradientMagnitudeMaxLimit(float gradient_magnitude_threshold_max = 0.0f);
-
-    /** @brief Switch to "Laplacian Zero-Crossing" edge feature extractor and specify its parameters
-     *
-     * This feature extractor is used by default according to article.
-     *
-     * Implementation has additional filtering for regions with low-amplitude noise.
-     * This filtering is enabled through parameter of minimal gradient amplitude (use some small value 4, 8, 16).
-     *
-     * @note Current implementation of this feature extractor is based on processing of grayscale images (color image is converted to grayscale image first).
-     *
-     * @note Canny edge detector is a bit slower, but provides better results (especially on color images): use setEdgeFeatureCannyParameters().
-     *
-     * @param gradient_magnitude_min_value Minimal gradient magnitude value for edge pixels (default: 0, check is disabled)
-     */
-    CV_WRAP
-    IntelligentScissorsMB& setEdgeFeatureZeroCrossingParameters(float gradient_magnitude_min_value = 0.0f);
-
-    /** @brief Switch edge feature extractor to use Canny edge detector
-     *
-     * @note "Laplacian Zero-Crossing" feature extractor is used by default (following to original article)
-     *
-     * @sa Canny
-     */
-    CV_WRAP
-    IntelligentScissorsMB& setEdgeFeatureCannyParameters(
-            double threshold1, double threshold2,
-            int apertureSize = 3, bool L2gradient = false
-    );
-
-    /** @brief Specify input image and extract image features
-     *
-     * @param image input image. Type is #CV_8UC1 / #CV_8UC3
-     */
-    CV_WRAP
-    IntelligentScissorsMB& applyImage(InputArray image);
-
-    /** @brief Specify custom features of imput image
-     *
-     * Customized advanced variant of applyImage() call.
-     *
-     * @param non_edge Specify cost of non-edge pixels. Type is CV_8UC1. Expected values are `{0, 1}`.
-     * @param gradient_direction Specify gradient direction feature. Type is CV_32FC2. Values are expected to be normalized: `x^2 + y^2 == 1`
-     * @param gradient_magnitude Specify cost of gradient magnitude function: Type is CV_32FC1. Values should be in range `[0, 1]`.
-     * @param image **Optional parameter**. Must be specified if subset of features is specified (non-specified features are calculated internally)
-     */
-    CV_WRAP
-    IntelligentScissorsMB& applyImageFeatures(
-            InputArray non_edge, InputArray gradient_direction, InputArray gradient_magnitude,
-            InputArray image = noArray()
-    );
-
-    /** @brief Prepares a map of optimal paths for the given source point on the image
-     *
-     * @note applyImage() / applyImageFeatures() must be called before this call
-     *
-     * @param sourcePt The source point used to find the paths
-     */
-    CV_WRAP void buildMap(const Point& sourcePt);
-
-    /** @brief Extracts optimal contour for the given target point on the image
-     *
-     * @note buildMap() must be called before this call
-     *
-     * @param targetPt The target point
-     * @param[out] contour The list of pixels which contains optimal path between the source and the target points of the image. Type is CV_32SC2 (compatible with `std::vector<Point>`)
-     * @param backward Flag to indicate reverse order of retrived pixels (use "true" value to fetch points from the target to the source point)
-     */
-    CV_WRAP void getContour(const Point& targetPt, OutputArray contour, bool backward = false) const;
-
-#ifndef CV_DOXYGEN
-    struct Impl;
-    inline Impl* getImpl() const { return impl.get(); }
-protected:
-    std::shared_ptr<Impl> impl;
-#endif
-};
-
-//! @}
-
-}  // namespace segmentation
-}  // namespace cv
-
-#endif // OPENCV_IMGPROC_SEGMENTATION_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/imgproc/types_c.h b/projects/llm_framework/include/opencv4/opencv2/imgproc/types_c.h
deleted file mode 100644
index d3e55f57..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/imgproc/types_c.h
+++ /dev/null
@@ -1,659 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_IMGPROC_TYPES_C_H
-#define OPENCV_IMGPROC_TYPES_C_H
-
-#include "opencv2/core/core_c.h"
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/** @addtogroup imgproc_c
-  @{
-*/
-
-/** Connected component structure */
-typedef struct CvConnectedComp
-{
-    double area;    /**<area of the connected component  */
-    CvScalar value; /**<average color of the connected component */
-    CvRect rect;    /**<ROI of the component  */
-    CvSeq* contour; /**<optional component boundary
-                      (the contour might have child contours corresponding to the holes)*/
-}
-CvConnectedComp;
-
-/** Image smooth methods */
-enum SmoothMethod_c
-{
-    /** linear convolution with \f$\texttt{size1}\times\texttt{size2}\f$ box kernel (all 1's). If
-    you want to smooth different pixels with different-size box kernels, you can use the integral
-    image that is computed using integral */
-    CV_BLUR_NO_SCALE =0,
-    /** linear convolution with \f$\texttt{size1}\times\texttt{size2}\f$ box kernel (all
-    1's) with subsequent scaling by \f$1/(\texttt{size1}\cdot\texttt{size2})\f$ */
-    CV_BLUR  =1,
-    /** linear convolution with a \f$\texttt{size1}\times\texttt{size2}\f$ Gaussian kernel */
-    CV_GAUSSIAN  =2,
-    /** median filter with a \f$\texttt{size1}\times\texttt{size1}\f$ square aperture */
-    CV_MEDIAN =3,
-    /** bilateral filter with a \f$\texttt{size1}\times\texttt{size1}\f$ square aperture, color
-    sigma= sigma1 and spatial sigma= sigma2. If size1=0, the aperture square side is set to
-    cvRound(sigma2\*1.5)\*2+1. See cv::bilateralFilter */
-    CV_BILATERAL =4
-};
-
-/** Filters used in pyramid decomposition */
-enum
-{
-    CV_GAUSSIAN_5x5 = 7
-};
-
-/** Special filters */
-enum
-{
-    CV_SCHARR =-1,
-    CV_MAX_SOBEL_KSIZE =7
-};
-
-/** Constants for color conversion */
-enum
-{
-    CV_BGR2BGRA    =0,
-    CV_RGB2RGBA    =CV_BGR2BGRA,
-
-    CV_BGRA2BGR    =1,
-    CV_RGBA2RGB    =CV_BGRA2BGR,
-
-    CV_BGR2RGBA    =2,
-    CV_RGB2BGRA    =CV_BGR2RGBA,
-
-    CV_RGBA2BGR    =3,
-    CV_BGRA2RGB    =CV_RGBA2BGR,
-
-    CV_BGR2RGB     =4,
-    CV_RGB2BGR     =CV_BGR2RGB,
-
-    CV_BGRA2RGBA   =5,
-    CV_RGBA2BGRA   =CV_BGRA2RGBA,
-
-    CV_BGR2GRAY    =6,
-    CV_RGB2GRAY    =7,
-    CV_GRAY2BGR    =8,
-    CV_GRAY2RGB    =CV_GRAY2BGR,
-    CV_GRAY2BGRA   =9,
-    CV_GRAY2RGBA   =CV_GRAY2BGRA,
-    CV_BGRA2GRAY   =10,
-    CV_RGBA2GRAY   =11,
-
-    CV_BGR2BGR565  =12,
-    CV_RGB2BGR565  =13,
-    CV_BGR5652BGR  =14,
-    CV_BGR5652RGB  =15,
-    CV_BGRA2BGR565 =16,
-    CV_RGBA2BGR565 =17,
-    CV_BGR5652BGRA =18,
-    CV_BGR5652RGBA =19,
-
-    CV_GRAY2BGR565 =20,
-    CV_BGR5652GRAY =21,
-
-    CV_BGR2BGR555  =22,
-    CV_RGB2BGR555  =23,
-    CV_BGR5552BGR  =24,
-    CV_BGR5552RGB  =25,
-    CV_BGRA2BGR555 =26,
-    CV_RGBA2BGR555 =27,
-    CV_BGR5552BGRA =28,
-    CV_BGR5552RGBA =29,
-
-    CV_GRAY2BGR555 =30,
-    CV_BGR5552GRAY =31,
-
-    CV_BGR2XYZ     =32,
-    CV_RGB2XYZ     =33,
-    CV_XYZ2BGR     =34,
-    CV_XYZ2RGB     =35,
-
-    CV_BGR2YCrCb   =36,
-    CV_RGB2YCrCb   =37,
-    CV_YCrCb2BGR   =38,
-    CV_YCrCb2RGB   =39,
-
-    CV_BGR2HSV     =40,
-    CV_RGB2HSV     =41,
-
-    CV_BGR2Lab     =44,
-    CV_RGB2Lab     =45,
-
-    CV_BayerBG2BGR =46,
-    CV_BayerGB2BGR =47,
-    CV_BayerRG2BGR =48,
-    CV_BayerGR2BGR =49,
-
-    CV_BayerBG2RGB =CV_BayerRG2BGR,
-    CV_BayerGB2RGB =CV_BayerGR2BGR,
-    CV_BayerRG2RGB =CV_BayerBG2BGR,
-    CV_BayerGR2RGB =CV_BayerGB2BGR,
-
-    CV_BGR2Luv     =50,
-    CV_RGB2Luv     =51,
-    CV_BGR2HLS     =52,
-    CV_RGB2HLS     =53,
-
-    CV_HSV2BGR     =54,
-    CV_HSV2RGB     =55,
-
-    CV_Lab2BGR     =56,
-    CV_Lab2RGB     =57,
-    CV_Luv2BGR     =58,
-    CV_Luv2RGB     =59,
-    CV_HLS2BGR     =60,
-    CV_HLS2RGB     =61,
-
-    CV_BayerBG2BGR_VNG =62,
-    CV_BayerGB2BGR_VNG =63,
-    CV_BayerRG2BGR_VNG =64,
-    CV_BayerGR2BGR_VNG =65,
-
-    CV_BayerBG2RGB_VNG =CV_BayerRG2BGR_VNG,
-    CV_BayerGB2RGB_VNG =CV_BayerGR2BGR_VNG,
-    CV_BayerRG2RGB_VNG =CV_BayerBG2BGR_VNG,
-    CV_BayerGR2RGB_VNG =CV_BayerGB2BGR_VNG,
-
-    CV_BGR2HSV_FULL = 66,
-    CV_RGB2HSV_FULL = 67,
-    CV_BGR2HLS_FULL = 68,
-    CV_RGB2HLS_FULL = 69,
-
-    CV_HSV2BGR_FULL = 70,
-    CV_HSV2RGB_FULL = 71,
-    CV_HLS2BGR_FULL = 72,
-    CV_HLS2RGB_FULL = 73,
-
-    CV_LBGR2Lab     = 74,
-    CV_LRGB2Lab     = 75,
-    CV_LBGR2Luv     = 76,
-    CV_LRGB2Luv     = 77,
-
-    CV_Lab2LBGR     = 78,
-    CV_Lab2LRGB     = 79,
-    CV_Luv2LBGR     = 80,
-    CV_Luv2LRGB     = 81,
-
-    CV_BGR2YUV      = 82,
-    CV_RGB2YUV      = 83,
-    CV_YUV2BGR      = 84,
-    CV_YUV2RGB      = 85,
-
-    CV_BayerBG2GRAY = 86,
-    CV_BayerGB2GRAY = 87,
-    CV_BayerRG2GRAY = 88,
-    CV_BayerGR2GRAY = 89,
-
-    //YUV 4:2:0 formats family
-    CV_YUV2RGB_NV12 = 90,
-    CV_YUV2BGR_NV12 = 91,
-    CV_YUV2RGB_NV21 = 92,
-    CV_YUV2BGR_NV21 = 93,
-    CV_YUV420sp2RGB = CV_YUV2RGB_NV21,
-    CV_YUV420sp2BGR = CV_YUV2BGR_NV21,
-
-    CV_YUV2RGBA_NV12 = 94,
-    CV_YUV2BGRA_NV12 = 95,
-    CV_YUV2RGBA_NV21 = 96,
-    CV_YUV2BGRA_NV21 = 97,
-    CV_YUV420sp2RGBA = CV_YUV2RGBA_NV21,
-    CV_YUV420sp2BGRA = CV_YUV2BGRA_NV21,
-
-    CV_YUV2RGB_YV12 = 98,
-    CV_YUV2BGR_YV12 = 99,
-    CV_YUV2RGB_IYUV = 100,
-    CV_YUV2BGR_IYUV = 101,
-    CV_YUV2RGB_I420 = CV_YUV2RGB_IYUV,
-    CV_YUV2BGR_I420 = CV_YUV2BGR_IYUV,
-    CV_YUV420p2RGB = CV_YUV2RGB_YV12,
-    CV_YUV420p2BGR = CV_YUV2BGR_YV12,
-
-    CV_YUV2RGBA_YV12 = 102,
-    CV_YUV2BGRA_YV12 = 103,
-    CV_YUV2RGBA_IYUV = 104,
-    CV_YUV2BGRA_IYUV = 105,
-    CV_YUV2RGBA_I420 = CV_YUV2RGBA_IYUV,
-    CV_YUV2BGRA_I420 = CV_YUV2BGRA_IYUV,
-    CV_YUV420p2RGBA = CV_YUV2RGBA_YV12,
-    CV_YUV420p2BGRA = CV_YUV2BGRA_YV12,
-
-    CV_YUV2GRAY_420 = 106,
-    CV_YUV2GRAY_NV21 = CV_YUV2GRAY_420,
-    CV_YUV2GRAY_NV12 = CV_YUV2GRAY_420,
-    CV_YUV2GRAY_YV12 = CV_YUV2GRAY_420,
-    CV_YUV2GRAY_IYUV = CV_YUV2GRAY_420,
-    CV_YUV2GRAY_I420 = CV_YUV2GRAY_420,
-    CV_YUV420sp2GRAY = CV_YUV2GRAY_420,
-    CV_YUV420p2GRAY = CV_YUV2GRAY_420,
-
-    //YUV 4:2:2 formats family
-    CV_YUV2RGB_UYVY = 107,
-    CV_YUV2BGR_UYVY = 108,
-    //CV_YUV2RGB_VYUY = 109,
-    //CV_YUV2BGR_VYUY = 110,
-    CV_YUV2RGB_Y422 = CV_YUV2RGB_UYVY,
-    CV_YUV2BGR_Y422 = CV_YUV2BGR_UYVY,
-    CV_YUV2RGB_UYNV = CV_YUV2RGB_UYVY,
-    CV_YUV2BGR_UYNV = CV_YUV2BGR_UYVY,
-
-    CV_YUV2RGBA_UYVY = 111,
-    CV_YUV2BGRA_UYVY = 112,
-    //CV_YUV2RGBA_VYUY = 113,
-    //CV_YUV2BGRA_VYUY = 114,
-    CV_YUV2RGBA_Y422 = CV_YUV2RGBA_UYVY,
-    CV_YUV2BGRA_Y422 = CV_YUV2BGRA_UYVY,
-    CV_YUV2RGBA_UYNV = CV_YUV2RGBA_UYVY,
-    CV_YUV2BGRA_UYNV = CV_YUV2BGRA_UYVY,
-
-    CV_YUV2RGB_YUY2 = 115,
-    CV_YUV2BGR_YUY2 = 116,
-    CV_YUV2RGB_YVYU = 117,
-    CV_YUV2BGR_YVYU = 118,
-    CV_YUV2RGB_YUYV = CV_YUV2RGB_YUY2,
-    CV_YUV2BGR_YUYV = CV_YUV2BGR_YUY2,
-    CV_YUV2RGB_YUNV = CV_YUV2RGB_YUY2,
-    CV_YUV2BGR_YUNV = CV_YUV2BGR_YUY2,
-
-    CV_YUV2RGBA_YUY2 = 119,
-    CV_YUV2BGRA_YUY2 = 120,
-    CV_YUV2RGBA_YVYU = 121,
-    CV_YUV2BGRA_YVYU = 122,
-    CV_YUV2RGBA_YUYV = CV_YUV2RGBA_YUY2,
-    CV_YUV2BGRA_YUYV = CV_YUV2BGRA_YUY2,
-    CV_YUV2RGBA_YUNV = CV_YUV2RGBA_YUY2,
-    CV_YUV2BGRA_YUNV = CV_YUV2BGRA_YUY2,
-
-    CV_YUV2GRAY_UYVY = 123,
-    CV_YUV2GRAY_YUY2 = 124,
-    //CV_YUV2GRAY_VYUY = CV_YUV2GRAY_UYVY,
-    CV_YUV2GRAY_Y422 = CV_YUV2GRAY_UYVY,
-    CV_YUV2GRAY_UYNV = CV_YUV2GRAY_UYVY,
-    CV_YUV2GRAY_YVYU = CV_YUV2GRAY_YUY2,
-    CV_YUV2GRAY_YUYV = CV_YUV2GRAY_YUY2,
-    CV_YUV2GRAY_YUNV = CV_YUV2GRAY_YUY2,
-
-    // alpha premultiplication
-    CV_RGBA2mRGBA = 125,
-    CV_mRGBA2RGBA = 126,
-
-    CV_RGB2YUV_I420 = 127,
-    CV_BGR2YUV_I420 = 128,
-    CV_RGB2YUV_IYUV = CV_RGB2YUV_I420,
-    CV_BGR2YUV_IYUV = CV_BGR2YUV_I420,
-
-    CV_RGBA2YUV_I420 = 129,
-    CV_BGRA2YUV_I420 = 130,
-    CV_RGBA2YUV_IYUV = CV_RGBA2YUV_I420,
-    CV_BGRA2YUV_IYUV = CV_BGRA2YUV_I420,
-    CV_RGB2YUV_YV12  = 131,
-    CV_BGR2YUV_YV12  = 132,
-    CV_RGBA2YUV_YV12 = 133,
-    CV_BGRA2YUV_YV12 = 134,
-
-    // Edge-Aware Demosaicing
-    CV_BayerBG2BGR_EA = 135,
-    CV_BayerGB2BGR_EA = 136,
-    CV_BayerRG2BGR_EA = 137,
-    CV_BayerGR2BGR_EA = 138,
-
-    CV_BayerBG2RGB_EA = CV_BayerRG2BGR_EA,
-    CV_BayerGB2RGB_EA = CV_BayerGR2BGR_EA,
-    CV_BayerRG2RGB_EA = CV_BayerBG2BGR_EA,
-    CV_BayerGR2RGB_EA = CV_BayerGB2BGR_EA,
-
-    CV_BayerBG2BGRA =139,
-    CV_BayerGB2BGRA =140,
-    CV_BayerRG2BGRA =141,
-    CV_BayerGR2BGRA =142,
-
-    CV_BayerBG2RGBA =CV_BayerRG2BGRA,
-    CV_BayerGB2RGBA =CV_BayerGR2BGRA,
-    CV_BayerRG2RGBA =CV_BayerBG2BGRA,
-    CV_BayerGR2RGBA =CV_BayerGB2BGRA,
-
-    CV_COLORCVT_MAX  = 143
-};
-
-
-/** Sub-pixel interpolation methods */
-enum
-{
-    CV_INTER_NN        =0,
-    CV_INTER_LINEAR    =1,
-    CV_INTER_CUBIC     =2,
-    CV_INTER_AREA      =3,
-    CV_INTER_LANCZOS4  =4
-};
-
-/** ... and other image warping flags */
-enum
-{
-    CV_WARP_FILL_OUTLIERS =8,
-    CV_WARP_INVERSE_MAP  =16
-};
-
-/** Shapes of a structuring element for morphological operations
-@see cv::MorphShapes, cv::getStructuringElement
-*/
-enum MorphShapes_c
-{
-    CV_SHAPE_RECT      =0,
-    CV_SHAPE_CROSS     =1,
-    CV_SHAPE_ELLIPSE   =2,
-    CV_SHAPE_CUSTOM    =100 //!< custom structuring element
-};
-
-/** Morphological operations */
-enum
-{
-    CV_MOP_ERODE        =0,
-    CV_MOP_DILATE       =1,
-    CV_MOP_OPEN         =2,
-    CV_MOP_CLOSE        =3,
-    CV_MOP_GRADIENT     =4,
-    CV_MOP_TOPHAT       =5,
-    CV_MOP_BLACKHAT     =6
-};
-
-/** Spatial and central moments */
-typedef struct CvMoments
-{
-    double  m00, m10, m01, m20, m11, m02, m30, m21, m12, m03; /**< spatial moments */
-    double  mu20, mu11, mu02, mu30, mu21, mu12, mu03; /**< central moments */
-    double  inv_sqrt_m00; /**< m00 != 0 ? 1/sqrt(m00) : 0 */
-
-#if defined(CV__ENABLE_C_API_CTORS) && defined(__cplusplus)
-    CvMoments(){}
-    CvMoments(const cv::Moments& m)
-    {
-        m00 = m.m00; m10 = m.m10; m01 = m.m01;
-        m20 = m.m20; m11 = m.m11; m02 = m.m02;
-        m30 = m.m30; m21 = m.m21; m12 = m.m12; m03 = m.m03;
-        mu20 = m.mu20; mu11 = m.mu11; mu02 = m.mu02;
-        mu30 = m.mu30; mu21 = m.mu21; mu12 = m.mu12; mu03 = m.mu03;
-        double am00 = std::abs(m.m00);
-        inv_sqrt_m00 = am00 > DBL_EPSILON ? 1./std::sqrt(am00) : 0;
-    }
-    operator cv::Moments() const
-    {
-        return cv::Moments(m00, m10, m01, m20, m11, m02, m30, m21, m12, m03);
-    }
-#endif
-}
-CvMoments;
-
-#ifdef __cplusplus
-} // extern "C"
-
-CV_INLINE CvMoments cvMoments()
-{
-#if !defined(CV__ENABLE_C_API_CTORS)
-    CvMoments self = CV_STRUCT_INITIALIZER; return self;
-#else
-    return CvMoments();
-#endif
-}
-
-CV_INLINE CvMoments cvMoments(const cv::Moments& m)
-{
-#if !defined(CV__ENABLE_C_API_CTORS)
-    double am00 = std::abs(m.m00);
-    CvMoments self = {
-        m.m00, m.m10, m.m01, m.m20, m.m11, m.m02, m.m30, m.m21, m.m12, m.m03,
-        m.mu20, m.mu11, m.mu02, m.mu30, m.mu21, m.mu12, m.mu03,
-        am00 > DBL_EPSILON ? 1./std::sqrt(am00) : 0
-    };
-    return self;
-#else
-    return CvMoments(m);
-#endif
-}
-
-extern "C" {
-#endif // __cplusplus
-
-/** Hu invariants */
-typedef struct CvHuMoments
-{
-    double hu1, hu2, hu3, hu4, hu5, hu6, hu7; /**< Hu invariants */
-}
-CvHuMoments;
-
-/** Template matching methods */
-enum
-{
-    CV_TM_SQDIFF        =0,
-    CV_TM_SQDIFF_NORMED =1,
-    CV_TM_CCORR         =2,
-    CV_TM_CCORR_NORMED  =3,
-    CV_TM_CCOEFF        =4,
-    CV_TM_CCOEFF_NORMED =5
-};
-
-typedef float (CV_CDECL * CvDistanceFunction)( const float* a, const float* b, void* user_param );
-
-/** Contour retrieval modes */
-enum
-{
-    CV_RETR_EXTERNAL=0,
-    CV_RETR_LIST=1,
-    CV_RETR_CCOMP=2,
-    CV_RETR_TREE=3,
-    CV_RETR_FLOODFILL=4
-};
-
-/** Contour approximation methods */
-enum
-{
-    CV_CHAIN_CODE=0,
-    CV_CHAIN_APPROX_NONE=1,
-    CV_CHAIN_APPROX_SIMPLE=2,
-    CV_CHAIN_APPROX_TC89_L1=3,
-    CV_CHAIN_APPROX_TC89_KCOS=4,
-    CV_LINK_RUNS=5
-};
-
-/*
-Internal structure that is used for sequential retrieving contours from the image.
-It supports both hierarchical and plane variants of Suzuki algorithm.
-*/
-typedef struct _CvContourScanner* CvContourScanner;
-
-/** Freeman chain reader state */
-typedef struct CvChainPtReader
-{
-    CV_SEQ_READER_FIELDS()
-    char      code;
-    CvPoint   pt;
-    schar     deltas[8][2];
-}
-CvChainPtReader;
-
-/** initializes 8-element array for fast access to 3x3 neighborhood of a pixel */
-#define  CV_INIT_3X3_DELTAS( deltas, step, nch )            \
-    ((deltas)[0] =  (nch),  (deltas)[1] = -(step) + (nch),  \
-     (deltas)[2] = -(step), (deltas)[3] = -(step) - (nch),  \
-     (deltas)[4] = -(nch),  (deltas)[5] =  (step) - (nch),  \
-     (deltas)[6] =  (step), (deltas)[7] =  (step) + (nch))
-
-
-/** Contour approximation algorithms */
-enum
-{
-    CV_POLY_APPROX_DP = 0
-};
-
-/** Shape matching methods */
-enum
-{
-    CV_CONTOURS_MATCH_I1  =1, //!< \f[I_1(A,B) =  \sum _{i=1...7}  \left |  \frac{1}{m^A_i} -  \frac{1}{m^B_i} \right |\f]
-    CV_CONTOURS_MATCH_I2  =2, //!< \f[I_2(A,B) =  \sum _{i=1...7}  \left | m^A_i - m^B_i  \right |\f]
-    CV_CONTOURS_MATCH_I3  =3  //!< \f[I_3(A,B) =  \max _{i=1...7}  \frac{ \left| m^A_i - m^B_i \right| }{ \left| m^A_i \right| }\f]
-};
-
-/** Shape orientation */
-enum
-{
-    CV_CLOCKWISE         =1,
-    CV_COUNTER_CLOCKWISE =2
-};
-
-
-/** Convexity defect */
-typedef struct CvConvexityDefect
-{
-    CvPoint* start; /**< point of the contour where the defect begins */
-    CvPoint* end; /**< point of the contour where the defect ends */
-    CvPoint* depth_point; /**< the farthest from the convex hull point within the defect */
-    float depth; /**< distance between the farthest point and the convex hull */
-} CvConvexityDefect;
-
-
-/** Histogram comparison methods */
-enum
-{
-    CV_COMP_CORREL        =0,
-    CV_COMP_CHISQR        =1,
-    CV_COMP_INTERSECT     =2,
-    CV_COMP_BHATTACHARYYA =3,
-    CV_COMP_HELLINGER     =CV_COMP_BHATTACHARYYA,
-    CV_COMP_CHISQR_ALT    =4,
-    CV_COMP_KL_DIV        =5
-};
-
-/** Mask size for distance transform */
-enum
-{
-    CV_DIST_MASK_3   =3,
-    CV_DIST_MASK_5   =5,
-    CV_DIST_MASK_PRECISE =0
-};
-
-/** Content of output label array: connected components or pixels */
-enum
-{
-  CV_DIST_LABEL_CCOMP = 0,
-  CV_DIST_LABEL_PIXEL = 1
-};
-
-/** Distance types for Distance Transform and M-estimators */
-enum
-{
-    CV_DIST_USER    =-1,  /**< User defined distance */
-    CV_DIST_L1      =1,   /**< distance = |x1-x2| + |y1-y2| */
-    CV_DIST_L2      =2,   /**< the simple euclidean distance */
-    CV_DIST_C       =3,   /**< distance = max(|x1-x2|,|y1-y2|) */
-    CV_DIST_L12     =4,   /**< L1-L2 metric: distance = 2(sqrt(1+x*x/2) - 1)) */
-    CV_DIST_FAIR    =5,   /**< distance = c^2(|x|/c-log(1+|x|/c)), c = 1.3998 */
-    CV_DIST_WELSCH  =6,   /**< distance = c^2/2(1-exp(-(x/c)^2)), c = 2.9846 */
-    CV_DIST_HUBER   =7    /**< distance = |x|<c ? x^2/2 : c(|x|-c/2), c=1.345 */
-};
-
-
-/** Threshold types */
-enum
-{
-    CV_THRESH_BINARY      =0,  /**< value = value > threshold ? max_value : 0       */
-    CV_THRESH_BINARY_INV  =1,  /**< value = value > threshold ? 0 : max_value       */
-    CV_THRESH_TRUNC       =2,  /**< value = value > threshold ? threshold : value   */
-    CV_THRESH_TOZERO      =3,  /**< value = value > threshold ? value : 0           */
-    CV_THRESH_TOZERO_INV  =4,  /**< value = value > threshold ? 0 : value           */
-    CV_THRESH_MASK        =7,
-    CV_THRESH_OTSU        =8, /**< use Otsu algorithm to choose the optimal threshold value;
-                                 combine the flag with one of the above CV_THRESH_* values */
-    CV_THRESH_TRIANGLE    =16  /**< use Triangle algorithm to choose the optimal threshold value;
-                                 combine the flag with one of the above CV_THRESH_* values, but not
-                                 with CV_THRESH_OTSU */
-};
-
-/** Adaptive threshold methods */
-enum
-{
-    CV_ADAPTIVE_THRESH_MEAN_C  =0,
-    CV_ADAPTIVE_THRESH_GAUSSIAN_C  =1
-};
-
-/** FloodFill flags */
-enum
-{
-    CV_FLOODFILL_FIXED_RANGE =(1 << 16),
-    CV_FLOODFILL_MASK_ONLY   =(1 << 17)
-};
-
-
-/** Canny edge detector flags */
-enum
-{
-    CV_CANNY_L2_GRADIENT  =(1 << 31)
-};
-
-/** Variants of a Hough transform */
-enum
-{
-    CV_HOUGH_STANDARD =0,
-    CV_HOUGH_PROBABILISTIC =1,
-    CV_HOUGH_MULTI_SCALE =2,
-    CV_HOUGH_GRADIENT =3
-};
-
-
-/* Fast search data structures  */
-struct CvFeatureTree;
-struct CvLSH;
-struct CvLSHOperations;
-
-/** @} */
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/opencv.hpp b/projects/llm_framework/include/opencv4/opencv2/opencv.hpp
deleted file mode 100644
index d17b94a4..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/opencv.hpp
+++ /dev/null
@@ -1,95 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009-2010, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_ALL_HPP
-#define OPENCV_ALL_HPP
-
-// File that defines what modules where included during the build of OpenCV
-// These are purely the defines of the correct HAVE_OPENCV_modulename values
-#include "opencv2/opencv_modules.hpp"
-
-// Then the list of defines is checked to include the correct headers
-// Core library is always included --> without no OpenCV functionality available
-#include "opencv2/core.hpp"
-
-// Then the optional modules are checked
-#ifdef HAVE_OPENCV_CALIB3D
-#include "opencv2/calib3d.hpp"
-#endif
-#ifdef HAVE_OPENCV_FEATURES2D
-#include "opencv2/features2d.hpp"
-#endif
-#ifdef HAVE_OPENCV_DNN
-#include "opencv2/dnn.hpp"
-#endif
-#ifdef HAVE_OPENCV_FLANN
-#include "opencv2/flann.hpp"
-#endif
-#ifdef HAVE_OPENCV_HIGHGUI
-#include "opencv2/highgui.hpp"
-#endif
-#ifdef HAVE_OPENCV_IMGCODECS
-#include "opencv2/imgcodecs.hpp"
-#endif
-#ifdef HAVE_OPENCV_IMGPROC
-#include "opencv2/imgproc.hpp"
-#endif
-#ifdef HAVE_OPENCV_ML
-#include "opencv2/ml.hpp"
-#endif
-#ifdef HAVE_OPENCV_OBJDETECT
-#include "opencv2/objdetect.hpp"
-#endif
-#ifdef HAVE_OPENCV_PHOTO
-#include "opencv2/photo.hpp"
-#endif
-#ifdef HAVE_OPENCV_STITCHING
-#include "opencv2/stitching.hpp"
-#endif
-#ifdef HAVE_OPENCV_VIDEO
-#include "opencv2/video.hpp"
-#endif
-#ifdef HAVE_OPENCV_VIDEOIO
-#include "opencv2/videoio.hpp"
-#endif
-
-#endif
diff --git a/projects/llm_framework/include/opencv4/opencv2/opencv_modules.hpp b/projects/llm_framework/include/opencv4/opencv2/opencv_modules.hpp
deleted file mode 100644
index 61e037b5..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/opencv_modules.hpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*
- *      ** File generated automatically, do not modify **
- *
- * This file defines the list of modules available in current build configuration
- *
- *
-*/
-
-// This definition means that OpenCV is built with enabled non-free code.
-// For example, patented algorithms for non-profit/non-commercial use only.
-/* #undef OPENCV_ENABLE_NONFREE */
-
-#define HAVE_OPENCV_CORE
-#define HAVE_OPENCV_HIGHGUI
-#define HAVE_OPENCV_IMGCODECS
-#define HAVE_OPENCV_IMGPROC
-#define HAVE_OPENCV_VIDEOIO
-
-
diff --git a/projects/llm_framework/include/opencv4/opencv2/videoio.hpp b/projects/llm_framework/include/opencv4/opencv2/videoio.hpp
deleted file mode 100644
index 93ea8cdd..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/videoio.hpp
+++ /dev/null
@@ -1,1129 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_VIDEOIO_HPP
-#define OPENCV_VIDEOIO_HPP
-
-#include "opencv2/core.hpp"
-
-/**
-  @defgroup videoio Video I/O
-
-  @brief Read and write video or images sequence with OpenCV
-
-  ### See also:
-  - @ref videoio_overview
-  - Tutorials: @ref tutorial_table_of_content_app
-  @{
-    @defgroup videoio_flags_base Flags for video I/O
-    @defgroup videoio_flags_others Additional flags for video I/O API backends
-    @defgroup videoio_hwaccel Hardware-accelerated video decoding and encoding
-    @defgroup videoio_c C API for video I/O
-    @defgroup videoio_ios iOS glue for video I/O
-    @defgroup videoio_winrt WinRT glue for video I/O
-    @defgroup videoio_registry Query I/O API backends registry
-  @}
-*/
-
-////////////////////////////////// video io /////////////////////////////////
-
-typedef struct CvCapture CvCapture;
-typedef struct CvVideoWriter CvVideoWriter;
-
-namespace cv
-{
-
-//! @addtogroup videoio
-//! @{
-
-//! @addtogroup videoio_flags_base
-//! @{
-
-
-/** @brief cv::VideoCapture API backends identifier.
-
-Select preferred API for a capture object.
-To be used in the VideoCapture::VideoCapture() constructor or VideoCapture::open()
-
-@note Backends are available only if they have been built with your OpenCV binaries.
-See @ref videoio_overview for more information.
-*/
-enum VideoCaptureAPIs {
-       CAP_ANY          = 0,            //!< Auto detect == 0
-       CAP_VFW          = 200,          //!< Video For Windows (obsolete, removed)
-       CAP_V4L          = 200,          //!< V4L/V4L2 capturing support
-       CAP_V4L2         = CAP_V4L,      //!< Same as CAP_V4L
-       CAP_FIREWIRE     = 300,          //!< IEEE 1394 drivers
-       CAP_FIREWARE     = CAP_FIREWIRE, //!< Same value as CAP_FIREWIRE
-       CAP_IEEE1394     = CAP_FIREWIRE, //!< Same value as CAP_FIREWIRE
-       CAP_DC1394       = CAP_FIREWIRE, //!< Same value as CAP_FIREWIRE
-       CAP_CMU1394      = CAP_FIREWIRE, //!< Same value as CAP_FIREWIRE
-       CAP_QT           = 500,          //!< QuickTime (obsolete, removed)
-       CAP_UNICAP       = 600,          //!< Unicap drivers (obsolete, removed)
-       CAP_DSHOW        = 700,          //!< DirectShow (via videoInput)
-       CAP_PVAPI        = 800,          //!< PvAPI, Prosilica GigE SDK
-       CAP_OPENNI       = 900,          //!< OpenNI (for Kinect)
-       CAP_OPENNI_ASUS  = 910,          //!< OpenNI (for Asus Xtion)
-       CAP_ANDROID      = 1000,         //!< Android - not used
-       CAP_XIAPI        = 1100,         //!< XIMEA Camera API
-       CAP_AVFOUNDATION = 1200,         //!< AVFoundation framework for iOS (OS X Lion will have the same API)
-       CAP_GIGANETIX    = 1300,         //!< Smartek Giganetix GigEVisionSDK
-       CAP_MSMF         = 1400,         //!< Microsoft Media Foundation (via videoInput)
-       CAP_WINRT        = 1410,         //!< Microsoft Windows Runtime using Media Foundation
-       CAP_INTELPERC    = 1500,         //!< RealSense (former Intel Perceptual Computing SDK)
-       CAP_REALSENSE    = 1500,         //!< Synonym for CAP_INTELPERC
-       CAP_OPENNI2      = 1600,         //!< OpenNI2 (for Kinect)
-       CAP_OPENNI2_ASUS = 1610,         //!< OpenNI2 (for Asus Xtion and Occipital Structure sensors)
-       CAP_OPENNI2_ASTRA= 1620,         //!< OpenNI2 (for Orbbec Astra)
-       CAP_GPHOTO2      = 1700,         //!< gPhoto2 connection
-       CAP_GSTREAMER    = 1800,         //!< GStreamer
-       CAP_FFMPEG       = 1900,         //!< Open and record video file or stream using the FFMPEG library
-       CAP_IMAGES       = 2000,         //!< OpenCV Image Sequence (e.g. img_%02d.jpg)
-       CAP_ARAVIS       = 2100,         //!< Aravis SDK
-       CAP_OPENCV_MJPEG = 2200,         //!< Built-in OpenCV MotionJPEG codec
-       CAP_INTEL_MFX    = 2300,         //!< Intel MediaSDK
-       CAP_XINE         = 2400,         //!< XINE engine (Linux)
-       CAP_UEYE         = 2500,         //!< uEye Camera API
-     };
-
-/** @brief cv::VideoCapture generic properties identifier.
-
- Reading / writing properties involves many layers. Some unexpected result might happens along this chain.
- Effective behaviour depends from device hardware, driver and API Backend.
- @sa videoio_flags_others, VideoCapture::get(), VideoCapture::set()
-*/
-enum VideoCaptureProperties {
-       CAP_PROP_POS_MSEC       =0, //!< Current position of the video file in milliseconds.
-       CAP_PROP_POS_FRAMES     =1, //!< 0-based index of the frame to be decoded/captured next.
-       CAP_PROP_POS_AVI_RATIO  =2, //!< Relative position of the video file: 0=start of the film, 1=end of the film.
-       CAP_PROP_FRAME_WIDTH    =3, //!< Width of the frames in the video stream.
-       CAP_PROP_FRAME_HEIGHT   =4, //!< Height of the frames in the video stream.
-       CAP_PROP_FPS            =5, //!< Frame rate.
-       CAP_PROP_FOURCC         =6, //!< 4-character code of codec. see VideoWriter::fourcc .
-       CAP_PROP_FRAME_COUNT    =7, //!< Number of frames in the video file.
-       CAP_PROP_FORMAT         =8, //!< Format of the %Mat objects (see Mat::type()) returned by VideoCapture::retrieve().
-                                   //!< Set value -1 to fetch undecoded RAW video streams (as Mat 8UC1).
-       CAP_PROP_MODE           =9, //!< Backend-specific value indicating the current capture mode.
-       CAP_PROP_BRIGHTNESS    =10, //!< Brightness of the image (only for those cameras that support).
-       CAP_PROP_CONTRAST      =11, //!< Contrast of the image (only for cameras).
-       CAP_PROP_SATURATION    =12, //!< Saturation of the image (only for cameras).
-       CAP_PROP_HUE           =13, //!< Hue of the image (only for cameras).
-       CAP_PROP_GAIN          =14, //!< Gain of the image (only for those cameras that support).
-       CAP_PROP_EXPOSURE      =15, //!< Exposure (only for those cameras that support).
-       CAP_PROP_CONVERT_RGB   =16, //!< Boolean flags indicating whether images should be converted to RGB. <br/>
-                                   //!< *GStreamer note*: The flag is ignored in case if custom pipeline is used. It's user responsibility to interpret pipeline output.
-       CAP_PROP_WHITE_BALANCE_BLUE_U =17, //!< Currently unsupported.
-       CAP_PROP_RECTIFICATION =18, //!< Rectification flag for stereo cameras (note: only supported by DC1394 v 2.x backend currently).
-       CAP_PROP_MONOCHROME    =19,
-       CAP_PROP_SHARPNESS     =20,
-       CAP_PROP_AUTO_EXPOSURE =21, //!< DC1394: exposure control done by camera, user can adjust reference level using this feature.
-       CAP_PROP_GAMMA         =22,
-       CAP_PROP_TEMPERATURE   =23,
-       CAP_PROP_TRIGGER       =24,
-       CAP_PROP_TRIGGER_DELAY =25,
-       CAP_PROP_WHITE_BALANCE_RED_V =26,
-       CAP_PROP_ZOOM          =27,
-       CAP_PROP_FOCUS         =28,
-       CAP_PROP_GUID          =29,
-       CAP_PROP_ISO_SPEED     =30,
-       CAP_PROP_BACKLIGHT     =32,
-       CAP_PROP_PAN           =33,
-       CAP_PROP_TILT          =34,
-       CAP_PROP_ROLL          =35,
-       CAP_PROP_IRIS          =36,
-       CAP_PROP_SETTINGS      =37, //!< Pop up video/camera filter dialog (note: only supported by DSHOW backend currently. The property value is ignored)
-       CAP_PROP_BUFFERSIZE    =38,
-       CAP_PROP_AUTOFOCUS     =39,
-       CAP_PROP_SAR_NUM       =40, //!< Sample aspect ratio: num/den (num)
-       CAP_PROP_SAR_DEN       =41, //!< Sample aspect ratio: num/den (den)
-       CAP_PROP_BACKEND       =42, //!< Current backend (enum VideoCaptureAPIs). Read-only property
-       CAP_PROP_CHANNEL       =43, //!< Video input or Channel Number (only for those cameras that support)
-       CAP_PROP_AUTO_WB       =44, //!< enable/ disable auto white-balance
-       CAP_PROP_WB_TEMPERATURE=45, //!< white-balance color temperature
-       CAP_PROP_CODEC_PIXEL_FORMAT =46,    //!< (read-only) codec's pixel format. 4-character code - see VideoWriter::fourcc . Subset of [AV_PIX_FMT_*](https://github.com/FFmpeg/FFmpeg/blob/master/libavcodec/raw.c) or -1 if unknown
-       CAP_PROP_BITRATE       =47, //!< (read-only) Video bitrate in kbits/s
-       CAP_PROP_ORIENTATION_META=48, //!< (read-only) Frame rotation defined by stream meta (applicable for FFmpeg back-end only)
-       CAP_PROP_ORIENTATION_AUTO=49, //!< if true - rotates output frames of CvCapture considering video file's metadata  (applicable for FFmpeg back-end only) (https://github.com/opencv/opencv/issues/15499)
-       CAP_PROP_HW_ACCELERATION=50, //!< (**open-only**) Hardware acceleration type (see #VideoAccelerationType). Setting supported only via `params` parameter in cv::VideoCapture constructor / .open() method. Default value is backend-specific.
-       CAP_PROP_HW_DEVICE      =51, //!< (**open-only**) Hardware device index (select GPU if multiple available). Device enumeration is acceleration type specific.
-       CAP_PROP_HW_ACCELERATION_USE_OPENCL=52, //!< (**open-only**) If non-zero, create new OpenCL context and bind it to current thread. The OpenCL context created with Video Acceleration context attached it (if not attached yet) for optimized GPU data copy between HW accelerated decoder and cv::UMat.
-       CAP_PROP_OPEN_TIMEOUT_MSEC=53, //!< (**open-only**) timeout in milliseconds for opening a video capture (applicable for FFmpeg back-end only)
-       CAP_PROP_READ_TIMEOUT_MSEC=54, //!< (**open-only**) timeout in milliseconds for reading from a video capture (applicable for FFmpeg back-end only)
-       CAP_PROP_STREAM_OPEN_TIME_USEC =55, //<! (read-only) time in microseconds since Jan 1 1970 when stream was opened. Applicable for FFmpeg backend only. Useful for RTSP and other live streams
-       CAP_PROP_VIDEO_TOTAL_CHANNELS = 56, //!< (read-only) Number of video channels
-       CAP_PROP_VIDEO_STREAM = 57, //!< (**open-only**) Specify video stream, 0-based index. Use -1 to disable video stream from file or IP cameras. Default value is 0.
-       CAP_PROP_AUDIO_STREAM = 58, //!< (**open-only**) Specify stream in multi-language media files, -1 - disable audio processing or microphone. Default value is -1.
-       CAP_PROP_AUDIO_POS = 59, //!< (read-only) Audio position is measured in samples. Accurate audio sample timestamp of previous grabbed fragment. See CAP_PROP_AUDIO_SAMPLES_PER_SECOND and CAP_PROP_AUDIO_SHIFT_NSEC.
-       CAP_PROP_AUDIO_SHIFT_NSEC = 60, //!< (read only) Contains the time difference between the start of the audio stream and the video stream in nanoseconds. Positive value means that audio is started after the first video frame. Negative value means that audio is started before the first video frame.
-       CAP_PROP_AUDIO_DATA_DEPTH = 61, //!< (open, read) Alternative definition to bits-per-sample, but with clear handling of 32F / 32S
-       CAP_PROP_AUDIO_SAMPLES_PER_SECOND = 62, //!< (open, read) determined from file/codec input. If not specified, then selected audio sample rate is 44100
-       CAP_PROP_AUDIO_BASE_INDEX = 63, //!< (read-only) Index of the first audio channel for .retrieve() calls. That audio channel number continues enumeration after video channels.
-       CAP_PROP_AUDIO_TOTAL_CHANNELS = 64, //!< (read-only) Number of audio channels in the selected audio stream (mono, stereo, etc)
-       CAP_PROP_AUDIO_TOTAL_STREAMS = 65, //!< (read-only) Number of audio streams.
-       CAP_PROP_AUDIO_SYNCHRONIZE = 66, //!< (open, read) Enables audio synchronization.
-       CAP_PROP_LRF_HAS_KEY_FRAME = 67, //!< FFmpeg back-end only - Indicates whether the Last Raw Frame (LRF), output from VideoCapture::read() when VideoCapture is initialized with VideoCapture::open(CAP_FFMPEG, {CAP_PROP_FORMAT, -1}) or VideoCapture::set(CAP_PROP_FORMAT,-1) is called before the first call to VideoCapture::read(), contains encoded data for a key frame.
-       CAP_PROP_CODEC_EXTRADATA_INDEX = 68, //!< Positive index indicates that returning extra data is supported by the video back end.  This can be retrieved as cap.retrieve(data, <returned index>).  E.g. When reading from a h264 encoded RTSP stream, the FFmpeg backend could return the SPS and/or PPS if available (if sent in reply to a DESCRIBE request), from calls to cap.retrieve(data, <returned index>).
-#ifndef CV_DOXYGEN
-       CV__CAP_PROP_LATEST
-#endif
-     };
-
-/** @brief cv::VideoWriter generic properties identifier.
- @sa VideoWriter::get(), VideoWriter::set()
-*/
-enum VideoWriterProperties {
-  VIDEOWRITER_PROP_QUALITY = 1,    //!< Current quality (0..100%) of the encoded videostream. Can be adjusted dynamically in some codecs.
-  VIDEOWRITER_PROP_FRAMEBYTES = 2, //!< (Read-only): Size of just encoded video frame. Note that the encoding order may be different from representation order.
-  VIDEOWRITER_PROP_NSTRIPES = 3,   //!< Number of stripes for parallel encoding. -1 for auto detection.
-  VIDEOWRITER_PROP_IS_COLOR = 4,   //!< If it is not zero, the encoder will expect and encode color frames, otherwise it
-                                   //!< will work with grayscale frames.
-  VIDEOWRITER_PROP_DEPTH = 5,      //!< Defaults to CV_8U.
-  VIDEOWRITER_PROP_HW_ACCELERATION = 6, //!< (**open-only**) Hardware acceleration type (see #VideoAccelerationType). Setting supported only via `params` parameter in VideoWriter constructor / .open() method. Default value is backend-specific.
-  VIDEOWRITER_PROP_HW_DEVICE       = 7, //!< (**open-only**) Hardware device index (select GPU if multiple available). Device enumeration is acceleration type specific.
-  VIDEOWRITER_PROP_HW_ACCELERATION_USE_OPENCL= 8, //!< (**open-only**) If non-zero, create new OpenCL context and bind it to current thread. The OpenCL context created with Video Acceleration context attached it (if not attached yet) for optimized GPU data copy between cv::UMat and HW accelerated encoder.
-#ifndef CV_DOXYGEN
-  CV__VIDEOWRITER_PROP_LATEST
-#endif
-};
-
-//! @} videoio_flags_base
-
-//! @addtogroup videoio_flags_others
-//! @{
-
-/** @name Hardware acceleration support
-    @{
-*/
-
-/** @brief Video Acceleration type
- *
- * Used as value in #CAP_PROP_HW_ACCELERATION and #VIDEOWRITER_PROP_HW_ACCELERATION
- *
- * @note In case of FFmpeg backend, it translated to enum AVHWDeviceType (https://github.com/FFmpeg/FFmpeg/blob/master/libavutil/hwcontext.h)
- */
-enum VideoAccelerationType
-{
-    VIDEO_ACCELERATION_NONE     =  0,  //!< Do not require any specific H/W acceleration, prefer software processing.
-                                       //!< Reading of this value means that special H/W accelerated handling is not added or not detected by OpenCV.
-
-    VIDEO_ACCELERATION_ANY      =  1,  //!< Prefer to use H/W acceleration. If no one supported, then fallback to software processing.
-                                       //!< @note H/W acceleration may require special configuration of used environment.
-                                       //!< @note Results in encoding scenario may differ between software and hardware accelerated encoders.
-
-    VIDEO_ACCELERATION_D3D11    =  2,  //!< DirectX 11
-    VIDEO_ACCELERATION_VAAPI    =  3,  //!< VAAPI
-    VIDEO_ACCELERATION_MFX      =  4,  //!< libmfx (Intel MediaSDK/oneVPL)
-};
-
-//! @} Hardware acceleration support
-
-/** @name IEEE 1394 drivers
-    @{
-*/
-
-/** @brief Modes of the IEEE 1394 controlling registers
-(can be: auto, manual, auto single push, absolute Latter allowed with any other mode)
-every feature can have only one mode turned on at a time
-*/
-enum { CAP_PROP_DC1394_OFF                = -4, //!< turn the feature off (not controlled manually nor automatically).
-       CAP_PROP_DC1394_MODE_MANUAL        = -3, //!< set automatically when a value of the feature is set by the user.
-       CAP_PROP_DC1394_MODE_AUTO          = -2,
-       CAP_PROP_DC1394_MODE_ONE_PUSH_AUTO = -1,
-       CAP_PROP_DC1394_MAX                = 31
-     };
-
-//! @} IEEE 1394 drivers
-
-/** @name OpenNI (for Kinect)
-    @{
-*/
-
-//! OpenNI map generators
-enum { CAP_OPENNI_DEPTH_GENERATOR = 1 << 31,
-       CAP_OPENNI_IMAGE_GENERATOR = 1 << 30,
-       CAP_OPENNI_IR_GENERATOR    = 1 << 29,
-       CAP_OPENNI_GENERATORS_MASK = CAP_OPENNI_DEPTH_GENERATOR + CAP_OPENNI_IMAGE_GENERATOR + CAP_OPENNI_IR_GENERATOR
-     };
-
-//! Properties of cameras available through OpenNI backend
-enum { CAP_PROP_OPENNI_OUTPUT_MODE       = 100,
-       CAP_PROP_OPENNI_FRAME_MAX_DEPTH   = 101, //!< In mm
-       CAP_PROP_OPENNI_BASELINE          = 102, //!< In mm
-       CAP_PROP_OPENNI_FOCAL_LENGTH      = 103, //!< In pixels
-       CAP_PROP_OPENNI_REGISTRATION      = 104, //!< Flag that synchronizes the remapping depth map to image map
-                                                //!< by changing depth generator's view point (if the flag is "on") or
-                                                //!< sets this view point to its normal one (if the flag is "off").
-       CAP_PROP_OPENNI_REGISTRATION_ON   = CAP_PROP_OPENNI_REGISTRATION,
-       CAP_PROP_OPENNI_APPROX_FRAME_SYNC = 105,
-       CAP_PROP_OPENNI_MAX_BUFFER_SIZE   = 106,
-       CAP_PROP_OPENNI_CIRCLE_BUFFER     = 107,
-       CAP_PROP_OPENNI_MAX_TIME_DURATION = 108,
-       CAP_PROP_OPENNI_GENERATOR_PRESENT = 109,
-       CAP_PROP_OPENNI2_SYNC             = 110,
-       CAP_PROP_OPENNI2_MIRROR           = 111
-     };
-
-//! OpenNI shortcuts
-enum { CAP_OPENNI_IMAGE_GENERATOR_PRESENT         = CAP_OPENNI_IMAGE_GENERATOR + CAP_PROP_OPENNI_GENERATOR_PRESENT,
-       CAP_OPENNI_IMAGE_GENERATOR_OUTPUT_MODE     = CAP_OPENNI_IMAGE_GENERATOR + CAP_PROP_OPENNI_OUTPUT_MODE,
-       CAP_OPENNI_DEPTH_GENERATOR_PRESENT         = CAP_OPENNI_DEPTH_GENERATOR + CAP_PROP_OPENNI_GENERATOR_PRESENT,
-       CAP_OPENNI_DEPTH_GENERATOR_BASELINE        = CAP_OPENNI_DEPTH_GENERATOR + CAP_PROP_OPENNI_BASELINE,
-       CAP_OPENNI_DEPTH_GENERATOR_FOCAL_LENGTH    = CAP_OPENNI_DEPTH_GENERATOR + CAP_PROP_OPENNI_FOCAL_LENGTH,
-       CAP_OPENNI_DEPTH_GENERATOR_REGISTRATION    = CAP_OPENNI_DEPTH_GENERATOR + CAP_PROP_OPENNI_REGISTRATION,
-       CAP_OPENNI_DEPTH_GENERATOR_REGISTRATION_ON = CAP_OPENNI_DEPTH_GENERATOR_REGISTRATION,
-       CAP_OPENNI_IR_GENERATOR_PRESENT            = CAP_OPENNI_IR_GENERATOR + CAP_PROP_OPENNI_GENERATOR_PRESENT,
-     };
-
-//! OpenNI data given from depth generator
-enum { CAP_OPENNI_DEPTH_MAP         = 0, //!< Depth values in mm (CV_16UC1)
-       CAP_OPENNI_POINT_CLOUD_MAP   = 1, //!< XYZ in meters (CV_32FC3)
-       CAP_OPENNI_DISPARITY_MAP     = 2, //!< Disparity in pixels (CV_8UC1)
-       CAP_OPENNI_DISPARITY_MAP_32F = 3, //!< Disparity in pixels (CV_32FC1)
-       CAP_OPENNI_VALID_DEPTH_MASK  = 4, //!< CV_8UC1
-
-       CAP_OPENNI_BGR_IMAGE         = 5, //!< Data given from RGB image generator
-       CAP_OPENNI_GRAY_IMAGE        = 6, //!< Data given from RGB image generator
-
-       CAP_OPENNI_IR_IMAGE          = 7  //!< Data given from IR image generator
-     };
-
-//! Supported output modes of OpenNI image generator
-enum { CAP_OPENNI_VGA_30HZ  = 0,
-       CAP_OPENNI_SXGA_15HZ = 1,
-       CAP_OPENNI_SXGA_30HZ = 2,
-       CAP_OPENNI_QVGA_30HZ = 3,
-       CAP_OPENNI_QVGA_60HZ = 4
-     };
-
-//! @} OpenNI
-
-/** @name GStreamer
-    @{
-*/
-
-enum { CAP_PROP_GSTREAMER_QUEUE_LENGTH = 200 //!< Default is 1
-     };
-
-//! @} GStreamer
-
-/** @name PvAPI, Prosilica GigE SDK
-    @{
-*/
-
-//! PVAPI
-enum { CAP_PROP_PVAPI_MULTICASTIP           = 300, //!< IP for enable multicast master mode. 0 for disable multicast.
-       CAP_PROP_PVAPI_FRAMESTARTTRIGGERMODE = 301, //!< FrameStartTriggerMode: Determines how a frame is initiated.
-       CAP_PROP_PVAPI_DECIMATIONHORIZONTAL  = 302, //!< Horizontal sub-sampling of the image.
-       CAP_PROP_PVAPI_DECIMATIONVERTICAL    = 303, //!< Vertical sub-sampling of the image.
-       CAP_PROP_PVAPI_BINNINGX              = 304, //!< Horizontal binning factor.
-       CAP_PROP_PVAPI_BINNINGY              = 305, //!< Vertical binning factor.
-       CAP_PROP_PVAPI_PIXELFORMAT           = 306  //!< Pixel format.
-     };
-
-//! PVAPI: FrameStartTriggerMode
-enum { CAP_PVAPI_FSTRIGMODE_FREERUN     = 0,    //!< Freerun
-       CAP_PVAPI_FSTRIGMODE_SYNCIN1     = 1,    //!< SyncIn1
-       CAP_PVAPI_FSTRIGMODE_SYNCIN2     = 2,    //!< SyncIn2
-       CAP_PVAPI_FSTRIGMODE_FIXEDRATE   = 3,    //!< FixedRate
-       CAP_PVAPI_FSTRIGMODE_SOFTWARE    = 4     //!< Software
-     };
-
-//! PVAPI: DecimationHorizontal, DecimationVertical
-enum { CAP_PVAPI_DECIMATION_OFF       = 1,    //!< Off
-       CAP_PVAPI_DECIMATION_2OUTOF4   = 2,    //!< 2 out of 4 decimation
-       CAP_PVAPI_DECIMATION_2OUTOF8   = 4,    //!< 2 out of 8 decimation
-       CAP_PVAPI_DECIMATION_2OUTOF16  = 8     //!< 2 out of 16 decimation
-     };
-
-//! PVAPI: PixelFormat
-enum { CAP_PVAPI_PIXELFORMAT_MONO8    = 1,    //!< Mono8
-       CAP_PVAPI_PIXELFORMAT_MONO16   = 2,    //!< Mono16
-       CAP_PVAPI_PIXELFORMAT_BAYER8   = 3,    //!< Bayer8
-       CAP_PVAPI_PIXELFORMAT_BAYER16  = 4,    //!< Bayer16
-       CAP_PVAPI_PIXELFORMAT_RGB24    = 5,    //!< Rgb24
-       CAP_PVAPI_PIXELFORMAT_BGR24    = 6,    //!< Bgr24
-       CAP_PVAPI_PIXELFORMAT_RGBA32   = 7,    //!< Rgba32
-       CAP_PVAPI_PIXELFORMAT_BGRA32   = 8,    //!< Bgra32
-     };
-
-//! @} PvAPI
-
-/** @name XIMEA Camera API
-    @{
-*/
-
-//! Properties of cameras available through XIMEA SDK backend
-enum { CAP_PROP_XI_DOWNSAMPLING                                 = 400, //!< Change image resolution by binning or skipping.
-       CAP_PROP_XI_DATA_FORMAT                                  = 401, //!< Output data format.
-       CAP_PROP_XI_OFFSET_X                                     = 402, //!< Horizontal offset from the origin to the area of interest (in pixels).
-       CAP_PROP_XI_OFFSET_Y                                     = 403, //!< Vertical offset from the origin to the area of interest (in pixels).
-       CAP_PROP_XI_TRG_SOURCE                                   = 404, //!< Defines source of trigger.
-       CAP_PROP_XI_TRG_SOFTWARE                                 = 405, //!< Generates an internal trigger. PRM_TRG_SOURCE must be set to TRG_SOFTWARE.
-       CAP_PROP_XI_GPI_SELECTOR                                 = 406, //!< Selects general purpose input.
-       CAP_PROP_XI_GPI_MODE                                     = 407, //!< Set general purpose input mode.
-       CAP_PROP_XI_GPI_LEVEL                                    = 408, //!< Get general purpose level.
-       CAP_PROP_XI_GPO_SELECTOR                                 = 409, //!< Selects general purpose output.
-       CAP_PROP_XI_GPO_MODE                                     = 410, //!< Set general purpose output mode.
-       CAP_PROP_XI_LED_SELECTOR                                 = 411, //!< Selects camera signalling LED.
-       CAP_PROP_XI_LED_MODE                                     = 412, //!< Define camera signalling LED functionality.
-       CAP_PROP_XI_MANUAL_WB                                    = 413, //!< Calculates White Balance(must be called during acquisition).
-       CAP_PROP_XI_AUTO_WB                                      = 414, //!< Automatic white balance.
-       CAP_PROP_XI_AEAG                                         = 415, //!< Automatic exposure/gain.
-       CAP_PROP_XI_EXP_PRIORITY                                 = 416, //!< Exposure priority (0.5 - exposure 50%, gain 50%).
-       CAP_PROP_XI_AE_MAX_LIMIT                                 = 417, //!< Maximum limit of exposure in AEAG procedure.
-       CAP_PROP_XI_AG_MAX_LIMIT                                 = 418, //!< Maximum limit of gain in AEAG procedure.
-       CAP_PROP_XI_AEAG_LEVEL                                   = 419, //!< Average intensity of output signal AEAG should achieve(in %).
-       CAP_PROP_XI_TIMEOUT                                      = 420, //!< Image capture timeout in milliseconds.
-       CAP_PROP_XI_EXPOSURE                                     = 421, //!< Exposure time in microseconds.
-       CAP_PROP_XI_EXPOSURE_BURST_COUNT                         = 422, //!< Sets the number of times of exposure in one frame.
-       CAP_PROP_XI_GAIN_SELECTOR                                = 423, //!< Gain selector for parameter Gain allows to select different type of gains.
-       CAP_PROP_XI_GAIN                                         = 424, //!< Gain in dB.
-       CAP_PROP_XI_DOWNSAMPLING_TYPE                            = 426, //!< Change image downsampling type.
-       CAP_PROP_XI_BINNING_SELECTOR                             = 427, //!< Binning engine selector.
-       CAP_PROP_XI_BINNING_VERTICAL                             = 428, //!< Vertical Binning - number of vertical photo-sensitive cells to combine together.
-       CAP_PROP_XI_BINNING_HORIZONTAL                           = 429, //!< Horizontal Binning - number of horizontal photo-sensitive cells to combine together.
-       CAP_PROP_XI_BINNING_PATTERN                              = 430, //!< Binning pattern type.
-       CAP_PROP_XI_DECIMATION_SELECTOR                          = 431, //!< Decimation engine selector.
-       CAP_PROP_XI_DECIMATION_VERTICAL                          = 432, //!< Vertical Decimation - vertical sub-sampling of the image - reduces the vertical resolution of the image by the specified vertical decimation factor.
-       CAP_PROP_XI_DECIMATION_HORIZONTAL                        = 433, //!< Horizontal Decimation - horizontal sub-sampling of the image - reduces the horizontal resolution of the image by the specified vertical decimation factor.
-       CAP_PROP_XI_DECIMATION_PATTERN                           = 434, //!< Decimation pattern type.
-       CAP_PROP_XI_TEST_PATTERN_GENERATOR_SELECTOR              = 587, //!< Selects which test pattern generator is controlled by the TestPattern feature.
-       CAP_PROP_XI_TEST_PATTERN                                 = 588, //!< Selects which test pattern type is generated by the selected generator.
-       CAP_PROP_XI_IMAGE_DATA_FORMAT                            = 435, //!< Output data format.
-       CAP_PROP_XI_SHUTTER_TYPE                                 = 436, //!< Change sensor shutter type(CMOS sensor).
-       CAP_PROP_XI_SENSOR_TAPS                                  = 437, //!< Number of taps.
-       CAP_PROP_XI_AEAG_ROI_OFFSET_X                            = 439, //!< Automatic exposure/gain ROI offset X.
-       CAP_PROP_XI_AEAG_ROI_OFFSET_Y                            = 440, //!< Automatic exposure/gain ROI offset Y.
-       CAP_PROP_XI_AEAG_ROI_WIDTH                               = 441, //!< Automatic exposure/gain ROI Width.
-       CAP_PROP_XI_AEAG_ROI_HEIGHT                              = 442, //!< Automatic exposure/gain ROI Height.
-       CAP_PROP_XI_BPC                                          = 445, //!< Correction of bad pixels.
-       CAP_PROP_XI_WB_KR                                        = 448, //!< White balance red coefficient.
-       CAP_PROP_XI_WB_KG                                        = 449, //!< White balance green coefficient.
-       CAP_PROP_XI_WB_KB                                        = 450, //!< White balance blue coefficient.
-       CAP_PROP_XI_WIDTH                                        = 451, //!< Width of the Image provided by the device (in pixels).
-       CAP_PROP_XI_HEIGHT                                       = 452, //!< Height of the Image provided by the device (in pixels).
-       CAP_PROP_XI_REGION_SELECTOR                              = 589, //!< Selects Region in Multiple ROI which parameters are set by width, height, ... ,region mode.
-       CAP_PROP_XI_REGION_MODE                                  = 595, //!< Activates/deactivates Region selected by Region Selector.
-       CAP_PROP_XI_LIMIT_BANDWIDTH                              = 459, //!< Set/get bandwidth(datarate)(in Megabits).
-       CAP_PROP_XI_SENSOR_DATA_BIT_DEPTH                        = 460, //!< Sensor output data bit depth.
-       CAP_PROP_XI_OUTPUT_DATA_BIT_DEPTH                        = 461, //!< Device output data bit depth.
-       CAP_PROP_XI_IMAGE_DATA_BIT_DEPTH                         = 462, //!< bitdepth of data returned by function xiGetImage.
-       CAP_PROP_XI_OUTPUT_DATA_PACKING                          = 463, //!< Device output data packing (or grouping) enabled. Packing could be enabled if output_data_bit_depth > 8 and packing capability is available.
-       CAP_PROP_XI_OUTPUT_DATA_PACKING_TYPE                     = 464, //!< Data packing type. Some cameras supports only specific packing type.
-       CAP_PROP_XI_IS_COOLED                                    = 465, //!< Returns 1 for cameras that support cooling.
-       CAP_PROP_XI_COOLING                                      = 466, //!< Start camera cooling.
-       CAP_PROP_XI_TARGET_TEMP                                  = 467, //!< Set sensor target temperature for cooling.
-       CAP_PROP_XI_CHIP_TEMP                                    = 468, //!< Camera sensor temperature.
-       CAP_PROP_XI_HOUS_TEMP                                    = 469, //!< Camera housing temperature.
-       CAP_PROP_XI_HOUS_BACK_SIDE_TEMP                          = 590, //!< Camera housing back side temperature.
-       CAP_PROP_XI_SENSOR_BOARD_TEMP                            = 596, //!< Camera sensor board temperature.
-       CAP_PROP_XI_CMS                                          = 470, //!< Mode of color management system.
-       CAP_PROP_XI_APPLY_CMS                                    = 471, //!< Enable applying of CMS profiles to xiGetImage (see XI_PRM_INPUT_CMS_PROFILE, XI_PRM_OUTPUT_CMS_PROFILE).
-       CAP_PROP_XI_IMAGE_IS_COLOR                               = 474, //!< Returns 1 for color cameras.
-       CAP_PROP_XI_COLOR_FILTER_ARRAY                           = 475, //!< Returns color filter array type of RAW data.
-       CAP_PROP_XI_GAMMAY                                       = 476, //!< Luminosity gamma.
-       CAP_PROP_XI_GAMMAC                                       = 477, //!< Chromaticity gamma.
-       CAP_PROP_XI_SHARPNESS                                    = 478, //!< Sharpness Strength.
-       CAP_PROP_XI_CC_MATRIX_00                                 = 479, //!< Color Correction Matrix element [0][0].
-       CAP_PROP_XI_CC_MATRIX_01                                 = 480, //!< Color Correction Matrix element [0][1].
-       CAP_PROP_XI_CC_MATRIX_02                                 = 481, //!< Color Correction Matrix element [0][2].
-       CAP_PROP_XI_CC_MATRIX_03                                 = 482, //!< Color Correction Matrix element [0][3].
-       CAP_PROP_XI_CC_MATRIX_10                                 = 483, //!< Color Correction Matrix element [1][0].
-       CAP_PROP_XI_CC_MATRIX_11                                 = 484, //!< Color Correction Matrix element [1][1].
-       CAP_PROP_XI_CC_MATRIX_12                                 = 485, //!< Color Correction Matrix element [1][2].
-       CAP_PROP_XI_CC_MATRIX_13                                 = 486, //!< Color Correction Matrix element [1][3].
-       CAP_PROP_XI_CC_MATRIX_20                                 = 487, //!< Color Correction Matrix element [2][0].
-       CAP_PROP_XI_CC_MATRIX_21                                 = 488, //!< Color Correction Matrix element [2][1].
-       CAP_PROP_XI_CC_MATRIX_22                                 = 489, //!< Color Correction Matrix element [2][2].
-       CAP_PROP_XI_CC_MATRIX_23                                 = 490, //!< Color Correction Matrix element [2][3].
-       CAP_PROP_XI_CC_MATRIX_30                                 = 491, //!< Color Correction Matrix element [3][0].
-       CAP_PROP_XI_CC_MATRIX_31                                 = 492, //!< Color Correction Matrix element [3][1].
-       CAP_PROP_XI_CC_MATRIX_32                                 = 493, //!< Color Correction Matrix element [3][2].
-       CAP_PROP_XI_CC_MATRIX_33                                 = 494, //!< Color Correction Matrix element [3][3].
-       CAP_PROP_XI_DEFAULT_CC_MATRIX                            = 495, //!< Set default Color Correction Matrix.
-       CAP_PROP_XI_TRG_SELECTOR                                 = 498, //!< Selects the type of trigger.
-       CAP_PROP_XI_ACQ_FRAME_BURST_COUNT                        = 499, //!< Sets number of frames acquired by burst. This burst is used only if trigger is set to FrameBurstStart.
-       CAP_PROP_XI_DEBOUNCE_EN                                  = 507, //!< Enable/Disable debounce to selected GPI.
-       CAP_PROP_XI_DEBOUNCE_T0                                  = 508, //!< Debounce time (x * 10us).
-       CAP_PROP_XI_DEBOUNCE_T1                                  = 509, //!< Debounce time (x * 10us).
-       CAP_PROP_XI_DEBOUNCE_POL                                 = 510, //!< Debounce polarity (pol = 1 t0 - falling edge, t1 - rising edge).
-       CAP_PROP_XI_LENS_MODE                                    = 511, //!< Status of lens control interface. This shall be set to XI_ON before any Lens operations.
-       CAP_PROP_XI_LENS_APERTURE_VALUE                          = 512, //!< Current lens aperture value in stops. Examples: 2.8, 4, 5.6, 8, 11.
-       CAP_PROP_XI_LENS_FOCUS_MOVEMENT_VALUE                    = 513, //!< Lens current focus movement value to be used by XI_PRM_LENS_FOCUS_MOVE in motor steps.
-       CAP_PROP_XI_LENS_FOCUS_MOVE                              = 514, //!< Moves lens focus motor by steps set in XI_PRM_LENS_FOCUS_MOVEMENT_VALUE.
-       CAP_PROP_XI_LENS_FOCUS_DISTANCE                          = 515, //!< Lens focus distance in cm.
-       CAP_PROP_XI_LENS_FOCAL_LENGTH                            = 516, //!< Lens focal distance in mm.
-       CAP_PROP_XI_LENS_FEATURE_SELECTOR                        = 517, //!< Selects the current feature which is accessible by XI_PRM_LENS_FEATURE.
-       CAP_PROP_XI_LENS_FEATURE                                 = 518, //!< Allows access to lens feature value currently selected by XI_PRM_LENS_FEATURE_SELECTOR.
-       CAP_PROP_XI_DEVICE_MODEL_ID                              = 521, //!< Returns device model id.
-       CAP_PROP_XI_DEVICE_SN                                    = 522, //!< Returns device serial number.
-       CAP_PROP_XI_IMAGE_DATA_FORMAT_RGB32_ALPHA                = 529, //!< The alpha channel of RGB32 output image format.
-       CAP_PROP_XI_IMAGE_PAYLOAD_SIZE                           = 530, //!< Buffer size in bytes sufficient for output image returned by xiGetImage.
-       CAP_PROP_XI_TRANSPORT_PIXEL_FORMAT                       = 531, //!< Current format of pixels on transport layer.
-       CAP_PROP_XI_SENSOR_CLOCK_FREQ_HZ                         = 532, //!< Sensor clock frequency in Hz.
-       CAP_PROP_XI_SENSOR_CLOCK_FREQ_INDEX                      = 533, //!< Sensor clock frequency index. Sensor with selected frequencies have possibility to set the frequency only by this index.
-       CAP_PROP_XI_SENSOR_OUTPUT_CHANNEL_COUNT                  = 534, //!< Number of output channels from sensor used for data transfer.
-       CAP_PROP_XI_FRAMERATE                                    = 535, //!< Define framerate in Hz.
-       CAP_PROP_XI_COUNTER_SELECTOR                             = 536, //!< Select counter.
-       CAP_PROP_XI_COUNTER_VALUE                                = 537, //!< Counter status.
-       CAP_PROP_XI_ACQ_TIMING_MODE                              = 538, //!< Type of sensor frames timing.
-       CAP_PROP_XI_AVAILABLE_BANDWIDTH                          = 539, //!< Calculate and returns available interface bandwidth(int Megabits).
-       CAP_PROP_XI_BUFFER_POLICY                                = 540, //!< Data move policy.
-       CAP_PROP_XI_LUT_EN                                       = 541, //!< Activates LUT.
-       CAP_PROP_XI_LUT_INDEX                                    = 542, //!< Control the index (offset) of the coefficient to access in the LUT.
-       CAP_PROP_XI_LUT_VALUE                                    = 543, //!< Value at entry LUTIndex of the LUT.
-       CAP_PROP_XI_TRG_DELAY                                    = 544, //!< Specifies the delay in microseconds (us) to apply after the trigger reception before activating it.
-       CAP_PROP_XI_TS_RST_MODE                                  = 545, //!< Defines how time stamp reset engine will be armed.
-       CAP_PROP_XI_TS_RST_SOURCE                                = 546, //!< Defines which source will be used for timestamp reset. Writing this parameter will trigger settings of engine (arming).
-       CAP_PROP_XI_IS_DEVICE_EXIST                              = 547, //!< Returns 1 if camera connected and works properly.
-       CAP_PROP_XI_ACQ_BUFFER_SIZE                              = 548, //!< Acquisition buffer size in buffer_size_unit. Default bytes.
-       CAP_PROP_XI_ACQ_BUFFER_SIZE_UNIT                         = 549, //!< Acquisition buffer size unit in bytes. Default 1. E.g. Value 1024 means that buffer_size is in KiBytes.
-       CAP_PROP_XI_ACQ_TRANSPORT_BUFFER_SIZE                    = 550, //!< Acquisition transport buffer size in bytes.
-       CAP_PROP_XI_BUFFERS_QUEUE_SIZE                           = 551, //!< Queue of field/frame buffers.
-       CAP_PROP_XI_ACQ_TRANSPORT_BUFFER_COMMIT                  = 552, //!< Number of buffers to commit to low level.
-       CAP_PROP_XI_RECENT_FRAME                                 = 553, //!< GetImage returns most recent frame.
-       CAP_PROP_XI_DEVICE_RESET                                 = 554, //!< Resets the camera to default state.
-       CAP_PROP_XI_COLUMN_FPN_CORRECTION                        = 555, //!< Correction of column FPN.
-       CAP_PROP_XI_ROW_FPN_CORRECTION                           = 591, //!< Correction of row FPN.
-       CAP_PROP_XI_SENSOR_MODE                                  = 558, //!< Current sensor mode. Allows to select sensor mode by one integer. Setting of this parameter affects: image dimensions and downsampling.
-       CAP_PROP_XI_HDR                                          = 559, //!< Enable High Dynamic Range feature.
-       CAP_PROP_XI_HDR_KNEEPOINT_COUNT                          = 560, //!< The number of kneepoints in the PWLR.
-       CAP_PROP_XI_HDR_T1                                       = 561, //!< Position of first kneepoint(in % of XI_PRM_EXPOSURE).
-       CAP_PROP_XI_HDR_T2                                       = 562, //!< Position of second kneepoint (in % of XI_PRM_EXPOSURE).
-       CAP_PROP_XI_KNEEPOINT1                                   = 563, //!< Value of first kneepoint (% of sensor saturation).
-       CAP_PROP_XI_KNEEPOINT2                                   = 564, //!< Value of second kneepoint (% of sensor saturation).
-       CAP_PROP_XI_IMAGE_BLACK_LEVEL                            = 565, //!< Last image black level counts. Can be used for Offline processing to recall it.
-       CAP_PROP_XI_HW_REVISION                                  = 571, //!< Returns hardware revision number.
-       CAP_PROP_XI_DEBUG_LEVEL                                  = 572, //!< Set debug level.
-       CAP_PROP_XI_AUTO_BANDWIDTH_CALCULATION                   = 573, //!< Automatic bandwidth calculation.
-       CAP_PROP_XI_FFS_FILE_ID                                  = 594, //!< File number.
-       CAP_PROP_XI_FFS_FILE_SIZE                                = 580, //!< Size of file.
-       CAP_PROP_XI_FREE_FFS_SIZE                                = 581, //!< Size of free camera FFS.
-       CAP_PROP_XI_USED_FFS_SIZE                                = 582, //!< Size of used camera FFS.
-       CAP_PROP_XI_FFS_ACCESS_KEY                               = 583, //!< Setting of key enables file operations on some cameras.
-       CAP_PROP_XI_SENSOR_FEATURE_SELECTOR                      = 585, //!< Selects the current feature which is accessible by XI_PRM_SENSOR_FEATURE_VALUE.
-       CAP_PROP_XI_SENSOR_FEATURE_VALUE                         = 586, //!< Allows access to sensor feature value currently selected by XI_PRM_SENSOR_FEATURE_SELECTOR.
-     };
-
-//! @} XIMEA
-
-
-/** @name ARAVIS Camera API
-    @{
-*/
-
-//! Properties of cameras available through ARAVIS backend
-enum { CAP_PROP_ARAVIS_AUTOTRIGGER                              = 600 //!< Automatically trigger frame capture if camera is configured with software trigger
-};
-
-//! @} ARAVIS
-
-/** @name AVFoundation framework for iOS
-    @{
-*/
-
-//! Properties of cameras available through AVFOUNDATION backend
-enum { CAP_PROP_IOS_DEVICE_FOCUS        = 9001,
-       CAP_PROP_IOS_DEVICE_EXPOSURE     = 9002,
-       CAP_PROP_IOS_DEVICE_FLASH        = 9003,
-       CAP_PROP_IOS_DEVICE_WHITEBALANCE = 9004,
-       CAP_PROP_IOS_DEVICE_TORCH        = 9005
-     };
-
-//! @} AVFoundation framework for iOS
-
-
-/** @name Smartek Giganetix GigEVisionSDK
-    @{
-*/
-
-//! Properties of cameras available through Smartek Giganetix Ethernet Vision backend
-/* --- Vladimir Litvinenko (litvinenko.vladimir@gmail.com) --- */
-enum { CAP_PROP_GIGA_FRAME_OFFSET_X   = 10001,
-       CAP_PROP_GIGA_FRAME_OFFSET_Y   = 10002,
-       CAP_PROP_GIGA_FRAME_WIDTH_MAX  = 10003,
-       CAP_PROP_GIGA_FRAME_HEIGH_MAX  = 10004,
-       CAP_PROP_GIGA_FRAME_SENS_WIDTH = 10005,
-       CAP_PROP_GIGA_FRAME_SENS_HEIGH = 10006
-     };
-
-//! @} Smartek
-
-/** @name Intel Perceptual Computing SDK
-    @{
-*/
-enum { CAP_PROP_INTELPERC_PROFILE_COUNT               = 11001,
-       CAP_PROP_INTELPERC_PROFILE_IDX                 = 11002,
-       CAP_PROP_INTELPERC_DEPTH_LOW_CONFIDENCE_VALUE  = 11003,
-       CAP_PROP_INTELPERC_DEPTH_SATURATION_VALUE      = 11004,
-       CAP_PROP_INTELPERC_DEPTH_CONFIDENCE_THRESHOLD  = 11005,
-       CAP_PROP_INTELPERC_DEPTH_FOCAL_LENGTH_HORZ     = 11006,
-       CAP_PROP_INTELPERC_DEPTH_FOCAL_LENGTH_VERT     = 11007
-     };
-
-//! Intel Perceptual Streams
-enum { CAP_INTELPERC_DEPTH_GENERATOR = 1 << 29,
-       CAP_INTELPERC_IMAGE_GENERATOR = 1 << 28,
-       CAP_INTELPERC_IR_GENERATOR    = 1 << 27,
-       CAP_INTELPERC_GENERATORS_MASK = CAP_INTELPERC_DEPTH_GENERATOR + CAP_INTELPERC_IMAGE_GENERATOR + CAP_INTELPERC_IR_GENERATOR
-     };
-
-enum { CAP_INTELPERC_DEPTH_MAP              = 0, //!< Each pixel is a 16-bit integer. The value indicates the distance from an object to the camera's XY plane or the Cartesian depth.
-       CAP_INTELPERC_UVDEPTH_MAP            = 1, //!< Each pixel contains two 32-bit floating point values in the range of 0-1, representing the mapping of depth coordinates to the color coordinates.
-       CAP_INTELPERC_IR_MAP                 = 2, //!< Each pixel is a 16-bit integer. The value indicates the intensity of the reflected laser beam.
-       CAP_INTELPERC_IMAGE                  = 3
-     };
-
-//! @} Intel Perceptual
-
-/** @name gPhoto2 connection
-    @{
-*/
-
-/** @brief gPhoto2 properties
-
-If `propertyId` is less than 0 then work on widget with that __additive inversed__ camera setting ID
-Get IDs by using CAP_PROP_GPHOTO2_WIDGET_ENUMERATE.
-@see CvCaptureCAM_GPHOTO2 for more info
-*/
-enum { CAP_PROP_GPHOTO2_PREVIEW           = 17001, //!< Capture only preview from liveview mode.
-       CAP_PROP_GPHOTO2_WIDGET_ENUMERATE  = 17002, //!< Readonly, returns (const char *).
-       CAP_PROP_GPHOTO2_RELOAD_CONFIG     = 17003, //!< Trigger, only by set. Reload camera settings.
-       CAP_PROP_GPHOTO2_RELOAD_ON_CHANGE  = 17004, //!< Reload all settings on set.
-       CAP_PROP_GPHOTO2_COLLECT_MSGS      = 17005, //!< Collect messages with details.
-       CAP_PROP_GPHOTO2_FLUSH_MSGS        = 17006, //!< Readonly, returns (const char *).
-       CAP_PROP_SPEED                     = 17007, //!< Exposure speed. Can be readonly, depends on camera program.
-       CAP_PROP_APERTURE                  = 17008, //!< Aperture. Can be readonly, depends on camera program.
-       CAP_PROP_EXPOSUREPROGRAM           = 17009, //!< Camera exposure program.
-       CAP_PROP_VIEWFINDER                = 17010  //!< Enter liveview mode.
-     };
-
-//! @} gPhoto2
-
-
-/** @name Images backend
-    @{
-*/
-
-/** @brief Images backend properties
-
-*/
-enum { CAP_PROP_IMAGES_BASE = 18000,
-       CAP_PROP_IMAGES_LAST = 19000 // excluding
-     };
-
-//! @} Images
-
-//! @} videoio_flags_others
-
-
-class IVideoCapture;
-//! @cond IGNORED
-namespace internal { class VideoCapturePrivateAccessor; }
-//! @endcond IGNORED
-
-/** @brief Class for video capturing from video files, image sequences or cameras.
-
-The class provides C++ API for capturing video from cameras or for reading video files and image sequences.
-
-Here is how the class can be used:
-@include samples/cpp/videocapture_basic.cpp
-
-@note In @ref videoio_c "C API" the black-box structure `CvCapture` is used instead of %VideoCapture.
-@note
--   (C++) A basic sample on using the %VideoCapture interface can be found at
-    `OPENCV_SOURCE_CODE/samples/cpp/videocapture_starter.cpp`
--   (Python) A basic sample on using the %VideoCapture interface can be found at
-    `OPENCV_SOURCE_CODE/samples/python/video.py`
--   (Python) A multi threaded video processing sample can be found at
-    `OPENCV_SOURCE_CODE/samples/python/video_threaded.py`
--   (Python) %VideoCapture sample showcasing some features of the Video4Linux2 backend
-    `OPENCV_SOURCE_CODE/samples/python/video_v4l2.py`
- */
-class CV_EXPORTS_W VideoCapture
-{
-public:
-    /** @brief Default constructor
-    @note In @ref videoio_c "C API", when you finished working with video, release CvCapture structure with
-    cvReleaseCapture(), or use Ptr\<CvCapture\> that calls cvReleaseCapture() automatically in the
-    destructor.
-     */
-    CV_WRAP VideoCapture();
-
-    /** @overload
-    @brief  Opens a video file or a capturing device or an IP video stream for video capturing with API Preference
-
-    @param filename it can be:
-    - name of video file (eg. `video.avi`)
-    - or image sequence (eg. `img_%02d.jpg`, which will read samples like `img_00.jpg, img_01.jpg, img_02.jpg, ...`)
-    - or URL of video stream (eg. `protocol://host:port/script_name?script_params|auth`)
-    - or GStreamer pipeline string in gst-launch tool format in case if GStreamer is used as backend
-      Note that each video stream or IP camera feed has its own URL scheme. Please refer to the
-      documentation of source stream to know the right URL.
-    @param apiPreference preferred Capture API backends to use. Can be used to enforce a specific reader
-    implementation if multiple are available: e.g. cv::CAP_FFMPEG or cv::CAP_IMAGES or cv::CAP_DSHOW.
-
-    @sa cv::VideoCaptureAPIs
-    */
-    CV_WRAP explicit VideoCapture(const String& filename, int apiPreference = CAP_ANY);
-
-    /** @overload
-    @brief Opens a video file or a capturing device or an IP video stream for video capturing with API Preference and parameters
-
-    The `params` parameter allows to specify extra parameters encoded as pairs `(paramId_1, paramValue_1, paramId_2, paramValue_2, ...)`.
-    See cv::VideoCaptureProperties
-    */
-    CV_WRAP explicit VideoCapture(const String& filename, int apiPreference, const std::vector<int>& params);
-
-    /** @overload
-    @brief  Opens a camera for video capturing
-
-    @param index id of the video capturing device to open. To open default camera using default backend just pass 0.
-    (to backward compatibility usage of camera_id + domain_offset (CAP_*) is valid when apiPreference is CAP_ANY)
-    @param apiPreference preferred Capture API backends to use. Can be used to enforce a specific reader
-    implementation if multiple are available: e.g. cv::CAP_DSHOW or cv::CAP_MSMF or cv::CAP_V4L.
-
-    @sa cv::VideoCaptureAPIs
-    */
-    CV_WRAP explicit VideoCapture(int index, int apiPreference = CAP_ANY);
-
-    /** @overload
-    @brief Opens a camera for video capturing with API Preference and parameters
-
-    The `params` parameter allows to specify extra parameters encoded as pairs `(paramId_1, paramValue_1, paramId_2, paramValue_2, ...)`.
-    See cv::VideoCaptureProperties
-    */
-    CV_WRAP explicit VideoCapture(int index, int apiPreference, const std::vector<int>& params);
-
-    /** @brief Default destructor
-
-    The method first calls VideoCapture::release to close the already opened file or camera.
-    */
-    virtual ~VideoCapture();
-
-    /** @brief  Opens a video file or a capturing device or an IP video stream for video capturing.
-
-    @overload
-
-    Parameters are same as the constructor VideoCapture(const String& filename, int apiPreference = CAP_ANY)
-    @return `true` if the file has been successfully opened
-
-    The method first calls VideoCapture::release to close the already opened file or camera.
-     */
-    CV_WRAP virtual bool open(const String& filename, int apiPreference = CAP_ANY);
-
-    /** @brief  Opens a camera for video capturing
-
-    @overload
-
-    The `params` parameter allows to specify extra parameters encoded as pairs `(paramId_1, paramValue_1, paramId_2, paramValue_2, ...)`.
-    See cv::VideoCaptureProperties
-
-    @return `true` if the file has been successfully opened
-
-    The method first calls VideoCapture::release to close the already opened file or camera.
-     */
-    CV_WRAP virtual bool open(const String& filename, int apiPreference, const std::vector<int>& params);
-
-    /** @brief  Opens a camera for video capturing
-
-    @overload
-
-    Parameters are same as the constructor VideoCapture(int index, int apiPreference = CAP_ANY)
-    @return `true` if the camera has been successfully opened.
-
-    The method first calls VideoCapture::release to close the already opened file or camera.
-    */
-    CV_WRAP virtual bool open(int index, int apiPreference = CAP_ANY);
-
-    /** @brief Returns true if video capturing has been initialized already.
-
-    @overload
-
-    The `params` parameter allows to specify extra parameters encoded as pairs `(paramId_1, paramValue_1, paramId_2, paramValue_2, ...)`.
-    See cv::VideoCaptureProperties
-
-    @return `true` if the camera has been successfully opened.
-
-    The method first calls VideoCapture::release to close the already opened file or camera.
-    */
-    CV_WRAP virtual bool open(int index, int apiPreference, const std::vector<int>& params);
-
-    /** @brief Returns true if video capturing has been initialized already.
-
-    If the previous call to VideoCapture constructor or VideoCapture::open() succeeded, the method returns
-    true.
-     */
-    CV_WRAP virtual bool isOpened() const;
-
-    /** @brief Closes video file or capturing device.
-
-    The method is automatically called by subsequent VideoCapture::open and by VideoCapture
-    destructor.
-
-    The C function also deallocates memory and clears \*capture pointer.
-     */
-    CV_WRAP virtual void release();
-
-    /** @brief Grabs the next frame from video file or capturing device.
-
-    @return `true` (non-zero) in the case of success.
-
-    The method/function grabs the next frame from video file or camera and returns true (non-zero) in
-    the case of success.
-
-    The primary use of the function is in multi-camera environments, especially when the cameras do not
-    have hardware synchronization. That is, you call VideoCapture::grab() for each camera and after that
-    call the slower method VideoCapture::retrieve() to decode and get frame from each camera. This way
-    the overhead on demosaicing or motion jpeg decompression etc. is eliminated and the retrieved frames
-    from different cameras will be closer in time.
-
-    Also, when a connected camera is multi-head (for example, a stereo camera or a Kinect device), the
-    correct way of retrieving data from it is to call VideoCapture::grab() first and then call
-    VideoCapture::retrieve() one or more times with different values of the channel parameter.
-
-    @ref tutorial_kinect_openni
-     */
-    CV_WRAP virtual bool grab();
-
-    /** @brief Decodes and returns the grabbed video frame.
-
-    @param [out] image the video frame is returned here. If no frames has been grabbed the image will be empty.
-    @param flag it could be a frame index or a driver specific flag
-    @return `false` if no frames has been grabbed
-
-    The method decodes and returns the just grabbed frame. If no frames has been grabbed
-    (camera has been disconnected, or there are no more frames in video file), the method returns false
-    and the function returns an empty image (with %cv::Mat, test it with Mat::empty()).
-
-    @sa read()
-
-    @note In @ref videoio_c "C API", functions cvRetrieveFrame() and cv.RetrieveFrame() return image stored inside the video
-    capturing structure. It is not allowed to modify or release the image! You can copy the frame using
-    cvCloneImage and then do whatever you want with the copy.
-     */
-    CV_WRAP virtual bool retrieve(OutputArray image, int flag = 0);
-
-    /** @brief Stream operator to read the next video frame.
-    @sa read()
-    */
-    virtual VideoCapture& operator >> (CV_OUT Mat& image);
-
-    /** @overload
-    @sa read()
-    */
-    virtual VideoCapture& operator >> (CV_OUT UMat& image);
-
-    /** @brief Grabs, decodes and returns the next video frame.
-
-    @param [out] image the video frame is returned here. If no frames has been grabbed the image will be empty.
-    @return `false` if no frames has been grabbed
-
-    The method/function combines VideoCapture::grab() and VideoCapture::retrieve() in one call. This is the
-    most convenient method for reading video files or capturing data from decode and returns the just
-    grabbed frame. If no frames has been grabbed (camera has been disconnected, or there are no more
-    frames in video file), the method returns false and the function returns empty image (with %cv::Mat, test it with Mat::empty()).
-
-    @note In @ref videoio_c "C API", functions cvRetrieveFrame() and cv.RetrieveFrame() return image stored inside the video
-    capturing structure. It is not allowed to modify or release the image! You can copy the frame using
-    cvCloneImage and then do whatever you want with the copy.
-     */
-    CV_WRAP virtual bool read(OutputArray image);
-
-    /** @brief Sets a property in the VideoCapture.
-
-    @param propId Property identifier from cv::VideoCaptureProperties (eg. cv::CAP_PROP_POS_MSEC, cv::CAP_PROP_POS_FRAMES, ...)
-    or one from @ref videoio_flags_others
-    @param value Value of the property.
-    @return `true` if the property is supported by backend used by the VideoCapture instance.
-    @note Even if it returns `true` this doesn't ensure that the property
-    value has been accepted by the capture device. See note in VideoCapture::get()
-     */
-    CV_WRAP virtual bool set(int propId, double value);
-
-    /** @brief Returns the specified VideoCapture property
-
-    @param propId Property identifier from cv::VideoCaptureProperties (eg. cv::CAP_PROP_POS_MSEC, cv::CAP_PROP_POS_FRAMES, ...)
-    or one from @ref videoio_flags_others
-    @return Value for the specified property. Value 0 is returned when querying a property that is
-    not supported by the backend used by the VideoCapture instance.
-
-    @note Reading / writing properties involves many layers. Some unexpected result might happens
-    along this chain.
-    @code{.txt}
-    VideoCapture -> API Backend -> Operating System -> Device Driver -> Device Hardware
-    @endcode
-    The returned value might be different from what really used by the device or it could be encoded
-    using device dependent rules (eg. steps or percentage). Effective behaviour depends from device
-    driver and API Backend
-
-    */
-    CV_WRAP virtual double get(int propId) const;
-
-    /** @brief Returns used backend API name
-
-     @note Stream should be opened.
-     */
-    CV_WRAP String getBackendName() const;
-
-    /** Switches exceptions mode
-     *
-     * methods raise exceptions if not successful instead of returning an error code
-     */
-    CV_WRAP void setExceptionMode(bool enable) { throwOnFail = enable; }
-
-    /// query if exception mode is active
-    CV_WRAP bool getExceptionMode() { return throwOnFail; }
-
-
-    /** @brief Wait for ready frames from VideoCapture.
-
-    @param streams input video streams
-    @param readyIndex stream indexes with grabbed frames (ready to use .retrieve() to fetch actual frame)
-    @param timeoutNs number of nanoseconds (0 - infinite)
-    @return `true` if streamReady is not empty
-
-    @throws Exception %Exception on stream errors (check .isOpened() to filter out malformed streams) or VideoCapture type is not supported
-
-    The primary use of the function is in multi-camera environments.
-    The method fills the ready state vector, grabs video frame, if camera is ready.
-
-    After this call use VideoCapture::retrieve() to decode and fetch frame data.
-    */
-    static /*CV_WRAP*/
-    bool waitAny(
-            const std::vector<VideoCapture>& streams,
-            CV_OUT std::vector<int>& readyIndex,
-            int64 timeoutNs = 0);
-
-protected:
-    Ptr<CvCapture> cap;
-    Ptr<IVideoCapture> icap;
-    bool throwOnFail;
-
-    friend class internal::VideoCapturePrivateAccessor;
-};
-
-class IVideoWriter;
-
-/** @example samples/cpp/tutorial_code/videoio/video-write/video-write.cpp
-Check @ref tutorial_video_write "the corresponding tutorial" for more details
-*/
-
-/** @example samples/cpp/videowriter_basic.cpp
-An example using VideoCapture and VideoWriter class
-*/
-
-/** @brief Video writer class.
-
-The class provides C++ API for writing video files or image sequences.
-*/
-class CV_EXPORTS_W VideoWriter
-{
-public:
-    /** @brief Default constructors
-
-    The constructors/functions initialize video writers.
-    -   On Linux FFMPEG is used to write videos;
-    -   On Windows FFMPEG or MSWF or DSHOW is used;
-    -   On MacOSX AVFoundation is used.
-     */
-    CV_WRAP VideoWriter();
-
-    /** @overload
-    @param filename Name of the output video file.
-    @param fourcc 4-character code of codec used to compress the frames. For example,
-    VideoWriter::fourcc('P','I','M','1') is a MPEG-1 codec, VideoWriter::fourcc('M','J','P','G') is a
-    motion-jpeg codec etc. List of codes can be obtained at [Video Codecs by
-    FOURCC](http://www.fourcc.org/codecs.php) page. FFMPEG backend with MP4 container natively uses
-    other values as fourcc code: see [ObjectType](http://mp4ra.org/#/codecs),
-    so you may receive a warning message from OpenCV about fourcc code conversion.
-    @param fps Framerate of the created video stream.
-    @param frameSize Size of the video frames.
-    @param isColor If it is not zero, the encoder will expect and encode color frames, otherwise it
-    will work with grayscale frames.
-
-    @b Tips:
-    - With some backends `fourcc=-1` pops up the codec selection dialog from the system.
-    - To save image sequence use a proper filename (eg. `img_%02d.jpg`) and `fourcc=0`
-      OR `fps=0`. Use uncompressed image format (eg. `img_%02d.BMP`) to save raw frames.
-    - Most codecs are lossy. If you want lossless video file you need to use a lossless codecs
-      (eg. FFMPEG FFV1, Huffman HFYU, Lagarith LAGS, etc...)
-    - If FFMPEG is enabled, using `codec=0; fps=0;` you can create an uncompressed (raw) video file.
-    */
-    CV_WRAP VideoWriter(const String& filename, int fourcc, double fps,
-                Size frameSize, bool isColor = true);
-
-    /** @overload
-    The `apiPreference` parameter allows to specify API backends to use. Can be used to enforce a specific reader implementation
-    if multiple are available: e.g. cv::CAP_FFMPEG or cv::CAP_GSTREAMER.
-     */
-    CV_WRAP VideoWriter(const String& filename, int apiPreference, int fourcc, double fps,
-                Size frameSize, bool isColor = true);
-
-    /** @overload
-     * The `params` parameter allows to specify extra encoder parameters encoded as pairs (paramId_1, paramValue_1, paramId_2, paramValue_2, ... .)
-     * see cv::VideoWriterProperties
-     */
-    CV_WRAP VideoWriter(const String& filename, int fourcc, double fps, const Size& frameSize,
-                        const std::vector<int>& params);
-
-    /** @overload
-     */
-    CV_WRAP VideoWriter(const String& filename, int apiPreference, int fourcc, double fps,
-                        const Size& frameSize, const std::vector<int>& params);
-
-    /** @brief Default destructor
-
-    The method first calls VideoWriter::release to close the already opened file.
-    */
-    virtual ~VideoWriter();
-
-    /** @brief Initializes or reinitializes video writer.
-
-    The method opens video writer. Parameters are the same as in the constructor
-    VideoWriter::VideoWriter.
-    @return `true` if video writer has been successfully initialized
-
-    The method first calls VideoWriter::release to close the already opened file.
-     */
-    CV_WRAP virtual bool open(const String& filename, int fourcc, double fps,
-                      Size frameSize, bool isColor = true);
-
-    /** @overload
-     */
-    CV_WRAP bool open(const String& filename, int apiPreference, int fourcc, double fps,
-                      Size frameSize, bool isColor = true);
-
-    /** @overload
-     */
-    CV_WRAP bool open(const String& filename, int fourcc, double fps, const Size& frameSize,
-                      const std::vector<int>& params);
-
-    /** @overload
-     */
-    CV_WRAP bool open(const String& filename, int apiPreference, int fourcc, double fps,
-                      const Size& frameSize, const std::vector<int>& params);
-
-    /** @brief Returns true if video writer has been successfully initialized.
-    */
-    CV_WRAP virtual bool isOpened() const;
-
-    /** @brief Closes the video writer.
-
-    The method is automatically called by subsequent VideoWriter::open and by the VideoWriter
-    destructor.
-     */
-    CV_WRAP virtual void release();
-
-    /** @brief Stream operator to write the next video frame.
-    @sa write
-    */
-    virtual VideoWriter& operator << (const Mat& image);
-
-    /** @overload
-    @sa write
-    */
-    virtual VideoWriter& operator << (const UMat& image);
-
-    /** @brief Writes the next video frame
-
-    @param image The written frame. In general, color images are expected in BGR format.
-
-    The function/method writes the specified image to video file. It must have the same size as has
-    been specified when opening the video writer.
-     */
-    CV_WRAP virtual void write(InputArray image);
-
-    /** @brief Sets a property in the VideoWriter.
-
-     @param propId Property identifier from cv::VideoWriterProperties (eg. cv::VIDEOWRITER_PROP_QUALITY)
-     or one of @ref videoio_flags_others
-
-     @param value Value of the property.
-     @return  `true` if the property is supported by the backend used by the VideoWriter instance.
-     */
-    CV_WRAP virtual bool set(int propId, double value);
-
-    /** @brief Returns the specified VideoWriter property
-
-     @param propId Property identifier from cv::VideoWriterProperties (eg. cv::VIDEOWRITER_PROP_QUALITY)
-     or one of @ref videoio_flags_others
-
-     @return Value for the specified property. Value 0 is returned when querying a property that is
-     not supported by the backend used by the VideoWriter instance.
-     */
-    CV_WRAP virtual double get(int propId) const;
-
-    /** @brief Concatenates 4 chars to a fourcc code
-
-    @return a fourcc code
-
-    This static method constructs the fourcc code of the codec to be used in the constructor
-    VideoWriter::VideoWriter or VideoWriter::open.
-     */
-    CV_WRAP static int fourcc(char c1, char c2, char c3, char c4);
-
-    /** @brief Returns used backend API name
-
-     @note Stream should be opened.
-     */
-    CV_WRAP String getBackendName() const;
-
-protected:
-    Ptr<CvVideoWriter> writer;
-    Ptr<IVideoWriter> iwriter;
-
-    static Ptr<IVideoWriter> create(const String& filename, int fourcc, double fps,
-                                    Size frameSize, bool isColor = true);
-};
-
-//! @cond IGNORED
-template<> struct DefaultDeleter<CvCapture>{ CV_EXPORTS void operator ()(CvCapture* obj) const; };
-template<> struct DefaultDeleter<CvVideoWriter>{ CV_EXPORTS void operator ()(CvVideoWriter* obj) const; };
-//! @endcond IGNORED
-
-//! @} videoio
-
-} // cv
-
-#endif //OPENCV_VIDEOIO_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/videoio/cap_ios.h b/projects/llm_framework/include/opencv4/opencv2/videoio/cap_ios.h
deleted file mode 100644
index 207ad46c..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/videoio/cap_ios.h
+++ /dev/null
@@ -1,150 +0,0 @@
-/*  For iOS video I/O
- *  by Eduard Feicho on 29/07/12
- *  Copyright 2012. All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *
- * 1. Redistributions of source code must retain the above copyright notice,
- *    this list of conditions and the following disclaimer.
- * 2. Redistributions in binary form must reproduce the above copyright notice,
- *    this list of conditions and the following disclaimer in the documentation
- *    and/or other materials provided with the distribution.
- * 3. The name of the author may not be used to endorse or promote products
- *    derived from this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE AUTHOR "AS IS" AND ANY EXPRESS OR IMPLIED
- * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
- * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO
- * EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
- * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
- * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
- * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
- * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
- * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
- * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- */
-
-#import <UIKit/UIKit.h>
-#import <Accelerate/Accelerate.h>
-#import <AVFoundation/AVFoundation.h>
-#import <ImageIO/ImageIO.h>
-#include "opencv2/core.hpp"
-
-//! @addtogroup videoio_ios
-//! @{
-
-/////////////////////////////////////// CvAbstractCamera /////////////////////////////////////
-
-@class CvAbstractCamera;
-
-CV_EXPORTS @interface CvAbstractCamera : NSObject
-{
-    UIDeviceOrientation currentDeviceOrientation;
-
-    BOOL cameraAvailable;
-}
-
-@property (nonatomic, strong) AVCaptureSession* captureSession;
-@property (nonatomic, strong) AVCaptureConnection* videoCaptureConnection;
-
-@property (nonatomic, readonly) BOOL running;
-@property (nonatomic, readonly) BOOL captureSessionLoaded;
-
-@property (nonatomic, assign) int defaultFPS;
-@property (nonatomic, readonly) AVCaptureVideoPreviewLayer *captureVideoPreviewLayer;
-@property (nonatomic, assign) AVCaptureDevicePosition defaultAVCaptureDevicePosition;
-@property (nonatomic, assign) AVCaptureVideoOrientation defaultAVCaptureVideoOrientation;
-@property (nonatomic, assign) BOOL useAVCaptureVideoPreviewLayer;
-@property (nonatomic, strong) NSString *const defaultAVCaptureSessionPreset;
-
-@property (nonatomic, assign) int imageWidth;
-@property (nonatomic, assign) int imageHeight;
-
-@property (nonatomic, strong) UIView* parentView;
-
-- CV_UNUSED(start);
-- CV_UNUSED(stop);
-- CV_UNUSED(switchCameras);
-
-- (id)initWithParentView:(UIView*)parent;
-
-- CV_UNUSED(createCaptureOutput);
-- CV_UNUSED(createVideoPreviewLayer);
-- CV_UNUSED(updateOrientation);
-
-- CV_UNUSED(lockFocus);
-- CV_UNUSED(unlockFocus);
-- CV_UNUSED(lockExposure);
-- CV_UNUSED(unlockExposure);
-- CV_UNUSED(lockBalance);
-- CV_UNUSED(unlockBalance);
-
-@end
-
-///////////////////////////////// CvVideoCamera ///////////////////////////////////////////
-
-@class CvVideoCamera;
-
-CV_EXPORTS @protocol CvVideoCameraDelegate <NSObject>
-
-#ifdef __cplusplus
-// delegate method for processing image frames
-- (void)processImage:(cv::Mat&)image;
-#endif
-
-@end
-
-CV_EXPORTS @interface CvVideoCamera : CvAbstractCamera<AVCaptureVideoDataOutputSampleBufferDelegate>
-{
-    AVCaptureVideoDataOutput *videoDataOutput;
-
-    dispatch_queue_t videoDataOutputQueue;
-    CALayer *customPreviewLayer;
-
-    CMTime lastSampleTime;
-
-}
-
-@property (nonatomic, weak) id<CvVideoCameraDelegate> delegate;
-@property (nonatomic, assign) BOOL grayscaleMode;
-
-@property (nonatomic, assign) BOOL recordVideo;
-@property (nonatomic, assign) BOOL rotateVideo;
-@property (nonatomic, strong) AVAssetWriterInput* recordAssetWriterInput;
-@property (nonatomic, strong) AVAssetWriterInputPixelBufferAdaptor* recordPixelBufferAdaptor;
-@property (nonatomic, strong) AVAssetWriter* recordAssetWriter;
-
-- (void)adjustLayoutToInterfaceOrientation:(UIInterfaceOrientation)interfaceOrientation;
-- CV_UNUSED(layoutPreviewLayer);
-- CV_UNUSED(saveVideo);
-- (NSURL *)videoFileURL;
-- (NSString *)videoFileString;
-
-
-@end
-
-///////////////////////////////// CvPhotoCamera ///////////////////////////////////////////
-
-@class CvPhotoCamera;
-
-CV_EXPORTS @protocol CvPhotoCameraDelegate <NSObject>
-
-- (void)photoCamera:(CvPhotoCamera*)photoCamera capturedImage:(UIImage *)image;
-- (void)photoCameraCancel:(CvPhotoCamera*)photoCamera;
-
-@end
-
-CV_EXPORTS @interface CvPhotoCamera : CvAbstractCamera
-{
-    AVCaptureStillImageOutput *stillImageOutput;
-}
-
-@property (nonatomic, weak) id<CvPhotoCameraDelegate> delegate;
-
-- CV_UNUSED(takePicture);
-
-@end
-
-//! @} videoio_ios
diff --git a/projects/llm_framework/include/opencv4/opencv2/videoio/legacy/constants_c.h b/projects/llm_framework/include/opencv4/opencv2/videoio/legacy/constants_c.h
deleted file mode 100644
index 91f85f87..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/videoio/legacy/constants_c.h
+++ /dev/null
@@ -1,434 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_VIDEOIO_LEGACY_CONSTANTS_H
-#define OPENCV_VIDEOIO_LEGACY_CONSTANTS_H
-
-enum
-{
-    CV_CAP_ANY      =0,     // autodetect
-
-    CV_CAP_MIL      =100,   // MIL proprietary drivers
-
-    CV_CAP_VFW      =200,   // platform native
-    CV_CAP_V4L      =200,
-    CV_CAP_V4L2     =200,
-
-    CV_CAP_FIREWARE =300,   // IEEE 1394 drivers
-    CV_CAP_FIREWIRE =300,
-    CV_CAP_IEEE1394 =300,
-    CV_CAP_DC1394   =300,
-    CV_CAP_CMU1394  =300,
-
-    CV_CAP_STEREO   =400,   // TYZX proprietary drivers
-    CV_CAP_TYZX     =400,
-    CV_TYZX_LEFT    =400,
-    CV_TYZX_RIGHT   =401,
-    CV_TYZX_COLOR   =402,
-    CV_TYZX_Z       =403,
-
-    CV_CAP_QT       =500,   // QuickTime
-
-    CV_CAP_UNICAP   =600,   // Unicap drivers
-
-    CV_CAP_DSHOW    =700,   // DirectShow (via videoInput)
-    CV_CAP_MSMF     =1400,  // Microsoft Media Foundation (via videoInput)
-
-    CV_CAP_PVAPI    =800,   // PvAPI, Prosilica GigE SDK
-
-    CV_CAP_OPENNI   =900,   // OpenNI (for Kinect)
-    CV_CAP_OPENNI_ASUS =910,   // OpenNI (for Asus Xtion)
-
-    CV_CAP_ANDROID  =1000,  // Android - not used
-    CV_CAP_ANDROID_BACK =CV_CAP_ANDROID+99, // Android back camera - not used
-    CV_CAP_ANDROID_FRONT =CV_CAP_ANDROID+98, // Android front camera - not used
-
-    CV_CAP_XIAPI    =1100,   // XIMEA Camera API
-
-    CV_CAP_AVFOUNDATION = 1200,  // AVFoundation framework for iOS (OS X Lion will have the same API)
-
-    CV_CAP_GIGANETIX = 1300,  // Smartek Giganetix GigEVisionSDK
-
-    CV_CAP_INTELPERC = 1500, // Intel Perceptual Computing
-
-    CV_CAP_OPENNI2 = 1600,   // OpenNI2 (for Kinect)
-    CV_CAP_GPHOTO2 = 1700,
-    CV_CAP_GSTREAMER = 1800, // GStreamer
-    CV_CAP_FFMPEG = 1900,    // FFMPEG
-    CV_CAP_IMAGES = 2000,    // OpenCV Image Sequence (e.g. img_%02d.jpg)
-
-    CV_CAP_ARAVIS = 2100     // Aravis GigE SDK
-};
-
-enum
-{
-    // modes of the controlling registers (can be: auto, manual, auto single push, absolute Latter allowed with any other mode)
-    // every feature can have only one mode turned on at a time
-    CV_CAP_PROP_DC1394_OFF         = -4,  //turn the feature off (not controlled manually nor automatically)
-    CV_CAP_PROP_DC1394_MODE_MANUAL = -3, //set automatically when a value of the feature is set by the user
-    CV_CAP_PROP_DC1394_MODE_AUTO = -2,
-    CV_CAP_PROP_DC1394_MODE_ONE_PUSH_AUTO = -1,
-    CV_CAP_PROP_POS_MSEC       =0,
-    CV_CAP_PROP_POS_FRAMES     =1,
-    CV_CAP_PROP_POS_AVI_RATIO  =2,
-    CV_CAP_PROP_FRAME_WIDTH    =3,
-    CV_CAP_PROP_FRAME_HEIGHT   =4,
-    CV_CAP_PROP_FPS            =5,
-    CV_CAP_PROP_FOURCC         =6,
-    CV_CAP_PROP_FRAME_COUNT    =7,
-    CV_CAP_PROP_FORMAT         =8,
-    CV_CAP_PROP_MODE           =9,
-    CV_CAP_PROP_BRIGHTNESS    =10,
-    CV_CAP_PROP_CONTRAST      =11,
-    CV_CAP_PROP_SATURATION    =12,
-    CV_CAP_PROP_HUE           =13,
-    CV_CAP_PROP_GAIN          =14,
-    CV_CAP_PROP_EXPOSURE      =15,
-    CV_CAP_PROP_CONVERT_RGB   =16,
-    CV_CAP_PROP_WHITE_BALANCE_BLUE_U =17,
-    CV_CAP_PROP_RECTIFICATION =18,
-    CV_CAP_PROP_MONOCHROME    =19,
-    CV_CAP_PROP_SHARPNESS     =20,
-    CV_CAP_PROP_AUTO_EXPOSURE =21, // exposure control done by camera,
-                                   // user can adjust reference level
-                                   // using this feature
-    CV_CAP_PROP_GAMMA         =22,
-    CV_CAP_PROP_TEMPERATURE   =23,
-    CV_CAP_PROP_TRIGGER       =24,
-    CV_CAP_PROP_TRIGGER_DELAY =25,
-    CV_CAP_PROP_WHITE_BALANCE_RED_V =26,
-    CV_CAP_PROP_ZOOM          =27,
-    CV_CAP_PROP_FOCUS         =28,
-    CV_CAP_PROP_GUID          =29,
-    CV_CAP_PROP_ISO_SPEED     =30,
-    CV_CAP_PROP_MAX_DC1394    =31,
-    CV_CAP_PROP_BACKLIGHT     =32,
-    CV_CAP_PROP_PAN           =33,
-    CV_CAP_PROP_TILT          =34,
-    CV_CAP_PROP_ROLL          =35,
-    CV_CAP_PROP_IRIS          =36,
-    CV_CAP_PROP_SETTINGS      =37,
-    CV_CAP_PROP_BUFFERSIZE    =38,
-    CV_CAP_PROP_AUTOFOCUS     =39,
-    CV_CAP_PROP_SAR_NUM       =40,
-    CV_CAP_PROP_SAR_DEN       =41,
-
-    CV_CAP_PROP_AUTOGRAB      =1024, // property for videoio class CvCapture_Android only
-    CV_CAP_PROP_SUPPORTED_PREVIEW_SIZES_STRING=1025, // readonly, tricky property, returns cpnst char* indeed
-    CV_CAP_PROP_PREVIEW_FORMAT=1026, // readonly, tricky property, returns cpnst char* indeed
-
-    // OpenNI map generators
-    CV_CAP_OPENNI_DEPTH_GENERATOR = 1 << 31,
-    CV_CAP_OPENNI_IMAGE_GENERATOR = 1 << 30,
-    CV_CAP_OPENNI_IR_GENERATOR    = 1 << 29,
-    CV_CAP_OPENNI_GENERATORS_MASK = CV_CAP_OPENNI_DEPTH_GENERATOR + CV_CAP_OPENNI_IMAGE_GENERATOR + CV_CAP_OPENNI_IR_GENERATOR,
-
-    // Properties of cameras available through OpenNI interfaces
-    CV_CAP_PROP_OPENNI_OUTPUT_MODE     = 100,
-    CV_CAP_PROP_OPENNI_FRAME_MAX_DEPTH = 101, // in mm
-    CV_CAP_PROP_OPENNI_BASELINE        = 102, // in mm
-    CV_CAP_PROP_OPENNI_FOCAL_LENGTH    = 103, // in pixels
-    CV_CAP_PROP_OPENNI_REGISTRATION    = 104, // flag
-    CV_CAP_PROP_OPENNI_REGISTRATION_ON = CV_CAP_PROP_OPENNI_REGISTRATION, // flag that synchronizes the remapping depth map to image map
-                                                                          // by changing depth generator's view point (if the flag is "on") or
-                                                                          // sets this view point to its normal one (if the flag is "off").
-    CV_CAP_PROP_OPENNI_APPROX_FRAME_SYNC = 105,
-    CV_CAP_PROP_OPENNI_MAX_BUFFER_SIZE   = 106,
-    CV_CAP_PROP_OPENNI_CIRCLE_BUFFER     = 107,
-    CV_CAP_PROP_OPENNI_MAX_TIME_DURATION = 108,
-
-    CV_CAP_PROP_OPENNI_GENERATOR_PRESENT = 109,
-    CV_CAP_PROP_OPENNI2_SYNC = 110,
-    CV_CAP_PROP_OPENNI2_MIRROR = 111,
-
-    CV_CAP_OPENNI_IMAGE_GENERATOR_PRESENT         = CV_CAP_OPENNI_IMAGE_GENERATOR + CV_CAP_PROP_OPENNI_GENERATOR_PRESENT,
-    CV_CAP_OPENNI_IMAGE_GENERATOR_OUTPUT_MODE     = CV_CAP_OPENNI_IMAGE_GENERATOR + CV_CAP_PROP_OPENNI_OUTPUT_MODE,
-    CV_CAP_OPENNI_DEPTH_GENERATOR_PRESENT         = CV_CAP_OPENNI_DEPTH_GENERATOR + CV_CAP_PROP_OPENNI_GENERATOR_PRESENT,
-    CV_CAP_OPENNI_DEPTH_GENERATOR_BASELINE        = CV_CAP_OPENNI_DEPTH_GENERATOR + CV_CAP_PROP_OPENNI_BASELINE,
-    CV_CAP_OPENNI_DEPTH_GENERATOR_FOCAL_LENGTH    = CV_CAP_OPENNI_DEPTH_GENERATOR + CV_CAP_PROP_OPENNI_FOCAL_LENGTH,
-    CV_CAP_OPENNI_DEPTH_GENERATOR_REGISTRATION    = CV_CAP_OPENNI_DEPTH_GENERATOR + CV_CAP_PROP_OPENNI_REGISTRATION,
-    CV_CAP_OPENNI_DEPTH_GENERATOR_REGISTRATION_ON = CV_CAP_OPENNI_DEPTH_GENERATOR_REGISTRATION,
-    CV_CAP_OPENNI_IR_GENERATOR_PRESENT            = CV_CAP_OPENNI_IR_GENERATOR + CV_CAP_PROP_OPENNI_GENERATOR_PRESENT,
-
-    // Properties of cameras available through GStreamer interface
-    CV_CAP_GSTREAMER_QUEUE_LENGTH           = 200, // default is 1
-
-    // PVAPI
-    CV_CAP_PROP_PVAPI_MULTICASTIP           = 300, // ip for anable multicast master mode. 0 for disable multicast
-    CV_CAP_PROP_PVAPI_FRAMESTARTTRIGGERMODE = 301, // FrameStartTriggerMode: Determines how a frame is initiated
-    CV_CAP_PROP_PVAPI_DECIMATIONHORIZONTAL  = 302, // Horizontal sub-sampling of the image
-    CV_CAP_PROP_PVAPI_DECIMATIONVERTICAL    = 303, // Vertical sub-sampling of the image
-    CV_CAP_PROP_PVAPI_BINNINGX              = 304, // Horizontal binning factor
-    CV_CAP_PROP_PVAPI_BINNINGY              = 305, // Vertical binning factor
-    CV_CAP_PROP_PVAPI_PIXELFORMAT           = 306, // Pixel format
-
-    // Properties of cameras available through XIMEA SDK interface
-    CV_CAP_PROP_XI_DOWNSAMPLING                                 = 400, // Change image resolution by binning or skipping.
-    CV_CAP_PROP_XI_DATA_FORMAT                                  = 401, // Output data format.
-    CV_CAP_PROP_XI_OFFSET_X                                     = 402, // Horizontal offset from the origin to the area of interest (in pixels).
-    CV_CAP_PROP_XI_OFFSET_Y                                     = 403, // Vertical offset from the origin to the area of interest (in pixels).
-    CV_CAP_PROP_XI_TRG_SOURCE                                   = 404, // Defines source of trigger.
-    CV_CAP_PROP_XI_TRG_SOFTWARE                                 = 405, // Generates an internal trigger. PRM_TRG_SOURCE must be set to TRG_SOFTWARE.
-    CV_CAP_PROP_XI_GPI_SELECTOR                                 = 406, // Selects general purpose input
-    CV_CAP_PROP_XI_GPI_MODE                                     = 407, // Set general purpose input mode
-    CV_CAP_PROP_XI_GPI_LEVEL                                    = 408, // Get general purpose level
-    CV_CAP_PROP_XI_GPO_SELECTOR                                 = 409, // Selects general purpose output
-    CV_CAP_PROP_XI_GPO_MODE                                     = 410, // Set general purpose output mode
-    CV_CAP_PROP_XI_LED_SELECTOR                                 = 411, // Selects camera signalling LED
-    CV_CAP_PROP_XI_LED_MODE                                     = 412, // Define camera signalling LED functionality
-    CV_CAP_PROP_XI_MANUAL_WB                                    = 413, // Calculates White Balance(must be called during acquisition)
-    CV_CAP_PROP_XI_AUTO_WB                                      = 414, // Automatic white balance
-    CV_CAP_PROP_XI_AEAG                                         = 415, // Automatic exposure/gain
-    CV_CAP_PROP_XI_EXP_PRIORITY                                 = 416, // Exposure priority (0.5 - exposure 50%, gain 50%).
-    CV_CAP_PROP_XI_AE_MAX_LIMIT                                 = 417, // Maximum limit of exposure in AEAG procedure
-    CV_CAP_PROP_XI_AG_MAX_LIMIT                                 = 418,  // Maximum limit of gain in AEAG procedure
-    CV_CAP_PROP_XI_AEAG_LEVEL                                   = 419, // Average intensity of output signal AEAG should achieve(in %)
-    CV_CAP_PROP_XI_TIMEOUT                                      = 420, // Image capture timeout in milliseconds
-    CV_CAP_PROP_XI_EXPOSURE                                     = 421, // Exposure time in microseconds
-    CV_CAP_PROP_XI_EXPOSURE_BURST_COUNT                         = 422, // Sets the number of times of exposure in one frame.
-    CV_CAP_PROP_XI_GAIN_SELECTOR                                = 423, // Gain selector for parameter Gain allows to select different type of gains.
-    CV_CAP_PROP_XI_GAIN                                         = 424, // Gain in dB
-    CV_CAP_PROP_XI_DOWNSAMPLING_TYPE                            = 426, // Change image downsampling type.
-    CV_CAP_PROP_XI_BINNING_SELECTOR                             = 427, // Binning engine selector.
-    CV_CAP_PROP_XI_BINNING_VERTICAL                             = 428, // Vertical Binning - number of vertical photo-sensitive cells to combine together.
-    CV_CAP_PROP_XI_BINNING_HORIZONTAL                           = 429, // Horizontal Binning - number of horizontal photo-sensitive cells to combine together.
-    CV_CAP_PROP_XI_BINNING_PATTERN                              = 430, // Binning pattern type.
-    CV_CAP_PROP_XI_DECIMATION_SELECTOR                          = 431, // Decimation engine selector.
-    CV_CAP_PROP_XI_DECIMATION_VERTICAL                          = 432, // Vertical Decimation - vertical sub-sampling of the image - reduces the vertical resolution of the image by the specified vertical decimation factor.
-    CV_CAP_PROP_XI_DECIMATION_HORIZONTAL                        = 433, // Horizontal Decimation - horizontal sub-sampling of the image - reduces the horizontal resolution of the image by the specified vertical decimation factor.
-    CV_CAP_PROP_XI_DECIMATION_PATTERN                           = 434, // Decimation pattern type.
-    CV_CAP_PROP_XI_TEST_PATTERN_GENERATOR_SELECTOR              = 587, // Selects which test pattern generator is controlled by the TestPattern feature.
-    CV_CAP_PROP_XI_TEST_PATTERN                                 = 588, // Selects which test pattern type is generated by the selected generator.
-    CV_CAP_PROP_XI_IMAGE_DATA_FORMAT                            = 435, // Output data format.
-    CV_CAP_PROP_XI_SHUTTER_TYPE                                 = 436, // Change sensor shutter type(CMOS sensor).
-    CV_CAP_PROP_XI_SENSOR_TAPS                                  = 437, // Number of taps
-    CV_CAP_PROP_XI_AEAG_ROI_OFFSET_X                            = 439, // Automatic exposure/gain ROI offset X
-    CV_CAP_PROP_XI_AEAG_ROI_OFFSET_Y                            = 440, // Automatic exposure/gain ROI offset Y
-    CV_CAP_PROP_XI_AEAG_ROI_WIDTH                               = 441, // Automatic exposure/gain ROI Width
-    CV_CAP_PROP_XI_AEAG_ROI_HEIGHT                              = 442, // Automatic exposure/gain ROI Height
-    CV_CAP_PROP_XI_BPC                                          = 445, // Correction of bad pixels
-    CV_CAP_PROP_XI_WB_KR                                        = 448, // White balance red coefficient
-    CV_CAP_PROP_XI_WB_KG                                        = 449, // White balance green coefficient
-    CV_CAP_PROP_XI_WB_KB                                        = 450, // White balance blue coefficient
-    CV_CAP_PROP_XI_WIDTH                                        = 451, // Width of the Image provided by the device (in pixels).
-    CV_CAP_PROP_XI_HEIGHT                                       = 452, // Height of the Image provided by the device (in pixels).
-    CV_CAP_PROP_XI_REGION_SELECTOR                              = 589, // Selects Region in Multiple ROI which parameters are set by width, height, ... ,region mode
-    CV_CAP_PROP_XI_REGION_MODE                                  = 595, // Activates/deactivates Region selected by Region Selector
-    CV_CAP_PROP_XI_LIMIT_BANDWIDTH                              = 459, // Set/get bandwidth(datarate)(in Megabits)
-    CV_CAP_PROP_XI_SENSOR_DATA_BIT_DEPTH                        = 460, // Sensor output data bit depth.
-    CV_CAP_PROP_XI_OUTPUT_DATA_BIT_DEPTH                        = 461, // Device output data bit depth.
-    CV_CAP_PROP_XI_IMAGE_DATA_BIT_DEPTH                         = 462, // bitdepth of data returned by function xiGetImage
-    CV_CAP_PROP_XI_OUTPUT_DATA_PACKING                          = 463, // Device output data packing (or grouping) enabled. Packing could be enabled if output_data_bit_depth > 8 and packing capability is available.
-    CV_CAP_PROP_XI_OUTPUT_DATA_PACKING_TYPE                     = 464, // Data packing type. Some cameras supports only specific packing type.
-    CV_CAP_PROP_XI_IS_COOLED                                    = 465, // Returns 1 for cameras that support cooling.
-    CV_CAP_PROP_XI_COOLING                                      = 466, // Start camera cooling.
-    CV_CAP_PROP_XI_TARGET_TEMP                                  = 467, // Set sensor target temperature for cooling.
-    CV_CAP_PROP_XI_CHIP_TEMP                                    = 468, // Camera sensor temperature
-    CV_CAP_PROP_XI_HOUS_TEMP                                    = 469, // Camera housing temperature
-    CV_CAP_PROP_XI_HOUS_BACK_SIDE_TEMP                          = 590, // Camera housing back side temperature
-    CV_CAP_PROP_XI_SENSOR_BOARD_TEMP                            = 596, // Camera sensor board temperature
-    CV_CAP_PROP_XI_CMS                                          = 470, // Mode of color management system.
-    CV_CAP_PROP_XI_APPLY_CMS                                    = 471, // Enable applying of CMS profiles to xiGetImage (see XI_PRM_INPUT_CMS_PROFILE, XI_PRM_OUTPUT_CMS_PROFILE).
-    CV_CAP_PROP_XI_IMAGE_IS_COLOR                               = 474, // Returns 1 for color cameras.
-    CV_CAP_PROP_XI_COLOR_FILTER_ARRAY                           = 475, // Returns color filter array type of RAW data.
-    CV_CAP_PROP_XI_GAMMAY                                       = 476, // Luminosity gamma
-    CV_CAP_PROP_XI_GAMMAC                                       = 477, // Chromaticity gamma
-    CV_CAP_PROP_XI_SHARPNESS                                    = 478, // Sharpness Strength
-    CV_CAP_PROP_XI_CC_MATRIX_00                                 = 479, // Color Correction Matrix element [0][0]
-    CV_CAP_PROP_XI_CC_MATRIX_01                                 = 480, // Color Correction Matrix element [0][1]
-    CV_CAP_PROP_XI_CC_MATRIX_02                                 = 481, // Color Correction Matrix element [0][2]
-    CV_CAP_PROP_XI_CC_MATRIX_03                                 = 482, // Color Correction Matrix element [0][3]
-    CV_CAP_PROP_XI_CC_MATRIX_10                                 = 483, // Color Correction Matrix element [1][0]
-    CV_CAP_PROP_XI_CC_MATRIX_11                                 = 484, // Color Correction Matrix element [1][1]
-    CV_CAP_PROP_XI_CC_MATRIX_12                                 = 485, // Color Correction Matrix element [1][2]
-    CV_CAP_PROP_XI_CC_MATRIX_13                                 = 486, // Color Correction Matrix element [1][3]
-    CV_CAP_PROP_XI_CC_MATRIX_20                                 = 487, // Color Correction Matrix element [2][0]
-    CV_CAP_PROP_XI_CC_MATRIX_21                                 = 488, // Color Correction Matrix element [2][1]
-    CV_CAP_PROP_XI_CC_MATRIX_22                                 = 489, // Color Correction Matrix element [2][2]
-    CV_CAP_PROP_XI_CC_MATRIX_23                                 = 490, // Color Correction Matrix element [2][3]
-    CV_CAP_PROP_XI_CC_MATRIX_30                                 = 491, // Color Correction Matrix element [3][0]
-    CV_CAP_PROP_XI_CC_MATRIX_31                                 = 492, // Color Correction Matrix element [3][1]
-    CV_CAP_PROP_XI_CC_MATRIX_32                                 = 493, // Color Correction Matrix element [3][2]
-    CV_CAP_PROP_XI_CC_MATRIX_33                                 = 494, // Color Correction Matrix element [3][3]
-    CV_CAP_PROP_XI_DEFAULT_CC_MATRIX                            = 495, // Set default Color Correction Matrix
-    CV_CAP_PROP_XI_TRG_SELECTOR                                 = 498, // Selects the type of trigger.
-    CV_CAP_PROP_XI_ACQ_FRAME_BURST_COUNT                        = 499, // Sets number of frames acquired by burst. This burst is used only if trigger is set to FrameBurstStart
-    CV_CAP_PROP_XI_DEBOUNCE_EN                                  = 507, // Enable/Disable debounce to selected GPI
-    CV_CAP_PROP_XI_DEBOUNCE_T0                                  = 508, // Debounce time (x * 10us)
-    CV_CAP_PROP_XI_DEBOUNCE_T1                                  = 509, // Debounce time (x * 10us)
-    CV_CAP_PROP_XI_DEBOUNCE_POL                                 = 510, // Debounce polarity (pol = 1 t0 - falling edge, t1 - rising edge)
-    CV_CAP_PROP_XI_LENS_MODE                                    = 511, // Status of lens control interface. This shall be set to XI_ON before any Lens operations.
-    CV_CAP_PROP_XI_LENS_APERTURE_VALUE                          = 512, // Current lens aperture value in stops. Examples: 2.8, 4, 5.6, 8, 11
-    CV_CAP_PROP_XI_LENS_FOCUS_MOVEMENT_VALUE                    = 513, // Lens current focus movement value to be used by XI_PRM_LENS_FOCUS_MOVE in motor steps.
-    CV_CAP_PROP_XI_LENS_FOCUS_MOVE                              = 514, // Moves lens focus motor by steps set in XI_PRM_LENS_FOCUS_MOVEMENT_VALUE.
-    CV_CAP_PROP_XI_LENS_FOCUS_DISTANCE                          = 515, // Lens focus distance in cm.
-    CV_CAP_PROP_XI_LENS_FOCAL_LENGTH                            = 516, // Lens focal distance in mm.
-    CV_CAP_PROP_XI_LENS_FEATURE_SELECTOR                        = 517, // Selects the current feature which is accessible by XI_PRM_LENS_FEATURE.
-    CV_CAP_PROP_XI_LENS_FEATURE                                 = 518, // Allows access to lens feature value currently selected by XI_PRM_LENS_FEATURE_SELECTOR.
-    CV_CAP_PROP_XI_DEVICE_MODEL_ID                              = 521, // Return device model id
-    CV_CAP_PROP_XI_DEVICE_SN                                    = 522, // Return device serial number
-    CV_CAP_PROP_XI_IMAGE_DATA_FORMAT_RGB32_ALPHA                = 529, // The alpha channel of RGB32 output image format.
-    CV_CAP_PROP_XI_IMAGE_PAYLOAD_SIZE                           = 530, // Buffer size in bytes sufficient for output image returned by xiGetImage
-    CV_CAP_PROP_XI_TRANSPORT_PIXEL_FORMAT                       = 531, // Current format of pixels on transport layer.
-    CV_CAP_PROP_XI_SENSOR_CLOCK_FREQ_HZ                         = 532, // Sensor clock frequency in Hz.
-    CV_CAP_PROP_XI_SENSOR_CLOCK_FREQ_INDEX                      = 533, // Sensor clock frequency index. Sensor with selected frequencies have possibility to set the frequency only by this index.
-    CV_CAP_PROP_XI_SENSOR_OUTPUT_CHANNEL_COUNT                  = 534, // Number of output channels from sensor used for data transfer.
-    CV_CAP_PROP_XI_FRAMERATE                                    = 535, // Define framerate in Hz
-    CV_CAP_PROP_XI_COUNTER_SELECTOR                             = 536, // Select counter
-    CV_CAP_PROP_XI_COUNTER_VALUE                                = 537, // Counter status
-    CV_CAP_PROP_XI_ACQ_TIMING_MODE                              = 538, // Type of sensor frames timing.
-    CV_CAP_PROP_XI_AVAILABLE_BANDWIDTH                          = 539, // Calculate and return available interface bandwidth(int Megabits)
-    CV_CAP_PROP_XI_BUFFER_POLICY                                = 540, // Data move policy
-    CV_CAP_PROP_XI_LUT_EN                                       = 541, // Activates LUT.
-    CV_CAP_PROP_XI_LUT_INDEX                                    = 542, // Control the index (offset) of the coefficient to access in the LUT.
-    CV_CAP_PROP_XI_LUT_VALUE                                    = 543, // Value at entry LUTIndex of the LUT
-    CV_CAP_PROP_XI_TRG_DELAY                                    = 544, // Specifies the delay in microseconds (us) to apply after the trigger reception before activating it.
-    CV_CAP_PROP_XI_TS_RST_MODE                                  = 545, // Defines how time stamp reset engine will be armed
-    CV_CAP_PROP_XI_TS_RST_SOURCE                                = 546, // Defines which source will be used for timestamp reset. Writing this parameter will trigger settings of engine (arming)
-    CV_CAP_PROP_XI_IS_DEVICE_EXIST                              = 547, // Returns 1 if camera connected and works properly.
-    CV_CAP_PROP_XI_ACQ_BUFFER_SIZE                              = 548, // Acquisition buffer size in buffer_size_unit. Default bytes.
-    CV_CAP_PROP_XI_ACQ_BUFFER_SIZE_UNIT                         = 549, // Acquisition buffer size unit in bytes. Default 1. E.g. Value 1024 means that buffer_size is in KiBytes
-    CV_CAP_PROP_XI_ACQ_TRANSPORT_BUFFER_SIZE                    = 550, // Acquisition transport buffer size in bytes
-    CV_CAP_PROP_XI_BUFFERS_QUEUE_SIZE                           = 551, // Queue of field/frame buffers
-    CV_CAP_PROP_XI_ACQ_TRANSPORT_BUFFER_COMMIT                  = 552, // Number of buffers to commit to low level
-    CV_CAP_PROP_XI_RECENT_FRAME                                 = 553, // GetImage returns most recent frame
-    CV_CAP_PROP_XI_DEVICE_RESET                                 = 554, // Resets the camera to default state.
-    CV_CAP_PROP_XI_COLUMN_FPN_CORRECTION                        = 555, // Correction of column FPN
-    CV_CAP_PROP_XI_ROW_FPN_CORRECTION                           = 591, // Correction of row FPN
-    CV_CAP_PROP_XI_SENSOR_MODE                                  = 558, // Current sensor mode. Allows to select sensor mode by one integer. Setting of this parameter affects: image dimensions and downsampling.
-    CV_CAP_PROP_XI_HDR                                          = 559, // Enable High Dynamic Range feature.
-    CV_CAP_PROP_XI_HDR_KNEEPOINT_COUNT                          = 560, // The number of kneepoints in the PWLR.
-    CV_CAP_PROP_XI_HDR_T1                                       = 561, // position of first kneepoint(in % of XI_PRM_EXPOSURE)
-    CV_CAP_PROP_XI_HDR_T2                                       = 562, // position of second kneepoint (in % of XI_PRM_EXPOSURE)
-    CV_CAP_PROP_XI_KNEEPOINT1                                   = 563, // value of first kneepoint (% of sensor saturation)
-    CV_CAP_PROP_XI_KNEEPOINT2                                   = 564, // value of second kneepoint (% of sensor saturation)
-    CV_CAP_PROP_XI_IMAGE_BLACK_LEVEL                            = 565, // Last image black level counts. Can be used for Offline processing to recall it.
-    CV_CAP_PROP_XI_HW_REVISION                                  = 571, // Returns hardware revision number.
-    CV_CAP_PROP_XI_DEBUG_LEVEL                                  = 572, // Set debug level
-    CV_CAP_PROP_XI_AUTO_BANDWIDTH_CALCULATION                   = 573, // Automatic bandwidth calculation,
-    CV_CAP_PROP_XI_FFS_FILE_ID                                  = 594, // File number.
-    CV_CAP_PROP_XI_FFS_FILE_SIZE                                = 580, // Size of file.
-    CV_CAP_PROP_XI_FREE_FFS_SIZE                                = 581, // Size of free camera FFS.
-    CV_CAP_PROP_XI_USED_FFS_SIZE                                = 582, // Size of used camera FFS.
-    CV_CAP_PROP_XI_FFS_ACCESS_KEY                               = 583, // Setting of key enables file operations on some cameras.
-    CV_CAP_PROP_XI_SENSOR_FEATURE_SELECTOR                      = 585, // Selects the current feature which is accessible by XI_PRM_SENSOR_FEATURE_VALUE.
-    CV_CAP_PROP_XI_SENSOR_FEATURE_VALUE                         = 586, // Allows access to sensor feature value currently selected by XI_PRM_SENSOR_FEATURE_SELECTOR.
-
-
-    // Properties for Android cameras
-    CV_CAP_PROP_ANDROID_FLASH_MODE = 8001,
-    CV_CAP_PROP_ANDROID_FOCUS_MODE = 8002,
-    CV_CAP_PROP_ANDROID_WHITE_BALANCE = 8003,
-    CV_CAP_PROP_ANDROID_ANTIBANDING = 8004,
-    CV_CAP_PROP_ANDROID_FOCAL_LENGTH = 8005,
-    CV_CAP_PROP_ANDROID_FOCUS_DISTANCE_NEAR = 8006,
-    CV_CAP_PROP_ANDROID_FOCUS_DISTANCE_OPTIMAL = 8007,
-    CV_CAP_PROP_ANDROID_FOCUS_DISTANCE_FAR = 8008,
-    CV_CAP_PROP_ANDROID_EXPOSE_LOCK = 8009,
-    CV_CAP_PROP_ANDROID_WHITEBALANCE_LOCK = 8010,
-
-    // Properties of cameras available through AVFOUNDATION interface
-    CV_CAP_PROP_IOS_DEVICE_FOCUS = 9001,
-    CV_CAP_PROP_IOS_DEVICE_EXPOSURE = 9002,
-    CV_CAP_PROP_IOS_DEVICE_FLASH = 9003,
-    CV_CAP_PROP_IOS_DEVICE_WHITEBALANCE = 9004,
-    CV_CAP_PROP_IOS_DEVICE_TORCH = 9005,
-
-    // Properties of cameras available through Smartek Giganetix Ethernet Vision interface
-    /* --- Vladimir Litvinenko (litvinenko.vladimir@gmail.com) --- */
-    CV_CAP_PROP_GIGA_FRAME_OFFSET_X = 10001,
-    CV_CAP_PROP_GIGA_FRAME_OFFSET_Y = 10002,
-    CV_CAP_PROP_GIGA_FRAME_WIDTH_MAX = 10003,
-    CV_CAP_PROP_GIGA_FRAME_HEIGH_MAX = 10004,
-    CV_CAP_PROP_GIGA_FRAME_SENS_WIDTH = 10005,
-    CV_CAP_PROP_GIGA_FRAME_SENS_HEIGH = 10006,
-
-    CV_CAP_PROP_INTELPERC_PROFILE_COUNT               = 11001,
-    CV_CAP_PROP_INTELPERC_PROFILE_IDX                 = 11002,
-    CV_CAP_PROP_INTELPERC_DEPTH_LOW_CONFIDENCE_VALUE  = 11003,
-    CV_CAP_PROP_INTELPERC_DEPTH_SATURATION_VALUE      = 11004,
-    CV_CAP_PROP_INTELPERC_DEPTH_CONFIDENCE_THRESHOLD  = 11005,
-    CV_CAP_PROP_INTELPERC_DEPTH_FOCAL_LENGTH_HORZ     = 11006,
-    CV_CAP_PROP_INTELPERC_DEPTH_FOCAL_LENGTH_VERT     = 11007,
-
-    // Intel PerC streams
-    CV_CAP_INTELPERC_DEPTH_GENERATOR = 1 << 29,
-    CV_CAP_INTELPERC_IMAGE_GENERATOR = 1 << 28,
-    CV_CAP_INTELPERC_GENERATORS_MASK = CV_CAP_INTELPERC_DEPTH_GENERATOR + CV_CAP_INTELPERC_IMAGE_GENERATOR
-};
-
-enum
-{
-    // Data given from depth generator.
-    CV_CAP_OPENNI_DEPTH_MAP                 = 0, // Depth values in mm (CV_16UC1)
-    CV_CAP_OPENNI_POINT_CLOUD_MAP           = 1, // XYZ in meters (CV_32FC3)
-    CV_CAP_OPENNI_DISPARITY_MAP             = 2, // Disparity in pixels (CV_8UC1)
-    CV_CAP_OPENNI_DISPARITY_MAP_32F         = 3, // Disparity in pixels (CV_32FC1)
-    CV_CAP_OPENNI_VALID_DEPTH_MASK          = 4, // CV_8UC1
-
-    // Data given from RGB image generator.
-    CV_CAP_OPENNI_BGR_IMAGE                 = 5,
-    CV_CAP_OPENNI_GRAY_IMAGE                = 6,
-
-    // Data given from IR image generator.
-    CV_CAP_OPENNI_IR_IMAGE                  = 7
-};
-
-// Supported output modes of OpenNI image generator
-enum
-{
-    CV_CAP_OPENNI_VGA_30HZ     = 0,
-    CV_CAP_OPENNI_SXGA_15HZ    = 1,
-    CV_CAP_OPENNI_SXGA_30HZ    = 2,
-    CV_CAP_OPENNI_QVGA_30HZ    = 3,
-    CV_CAP_OPENNI_QVGA_60HZ    = 4
-};
-
-enum
-{
-    CV_CAP_INTELPERC_DEPTH_MAP              = 0, // Each pixel is a 16-bit integer. The value indicates the distance from an object to the camera's XY plane or the Cartesian depth.
-    CV_CAP_INTELPERC_UVDEPTH_MAP            = 1, // Each pixel contains two 32-bit floating point values in the range of 0-1, representing the mapping of depth coordinates to the color coordinates.
-    CV_CAP_INTELPERC_IR_MAP                 = 2, // Each pixel is a 16-bit integer. The value indicates the intensity of the reflected laser beam.
-    CV_CAP_INTELPERC_IMAGE                  = 3
-};
-
-// gPhoto2 properties, if propertyId is less than 0 then work on widget with that __additive inversed__ camera setting ID
-// Get IDs by using CAP_PROP_GPHOTO2_WIDGET_ENUMERATE.
-// @see CvCaptureCAM_GPHOTO2 for more info
-enum
-{
-    CV_CAP_PROP_GPHOTO2_PREVIEW           = 17001, // Capture only preview from liveview mode.
-    CV_CAP_PROP_GPHOTO2_WIDGET_ENUMERATE  = 17002, // Readonly, returns (const char *).
-    CV_CAP_PROP_GPHOTO2_RELOAD_CONFIG     = 17003, // Trigger, only by set. Reload camera settings.
-    CV_CAP_PROP_GPHOTO2_RELOAD_ON_CHANGE  = 17004, // Reload all settings on set.
-    CV_CAP_PROP_GPHOTO2_COLLECT_MSGS      = 17005, // Collect messages with details.
-    CV_CAP_PROP_GPHOTO2_FLUSH_MSGS        = 17006, // Readonly, returns (const char *).
-    CV_CAP_PROP_SPEED                     = 17007, // Exposure speed. Can be readonly, depends on camera program.
-    CV_CAP_PROP_APERTURE                  = 17008, // Aperture. Can be readonly, depends on camera program.
-    CV_CAP_PROP_EXPOSUREPROGRAM           = 17009, // Camera exposure program.
-    CV_CAP_PROP_VIEWFINDER                = 17010  // Enter liveview mode.
-};
-
-//! Macro to construct the fourcc code of the codec. Same as CV_FOURCC()
-#define CV_FOURCC_MACRO(c1, c2, c3, c4) (((c1) & 255) + (((c2) & 255) << 8) + (((c3) & 255) << 16) + (((c4) & 255) << 24))
-
-/** @brief Constructs the fourcc code of the codec function
-
-Simply call it with 4 chars fourcc code like `CV_FOURCC('I', 'Y', 'U', 'V')`
-
-List of codes can be obtained at [Video Codecs by FOURCC](http://www.fourcc.org/codecs.php) page.
-FFMPEG backend with MP4 container natively uses other values as fourcc code:
-see [ObjectType](http://mp4ra.org/#/codecs).
-*/
-CV_INLINE int CV_FOURCC(char c1, char c2, char c3, char c4)
-{
-    return CV_FOURCC_MACRO(c1, c2, c3, c4);
-}
-
-//! (Windows only) Open Codec Selection Dialog
-#define CV_FOURCC_PROMPT -1
-//! (Linux only) Use default codec for specified filename
-#define CV_FOURCC_DEFAULT CV_FOURCC('I', 'Y', 'U', 'V')
-
-#endif // OPENCV_VIDEOIO_LEGACY_CONSTANTS_H
diff --git a/projects/llm_framework/include/opencv4/opencv2/videoio/registry.hpp b/projects/llm_framework/include/opencv4/opencv2/videoio/registry.hpp
deleted file mode 100644
index cf72247b..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/videoio/registry.hpp
+++ /dev/null
@@ -1,72 +0,0 @@
-// This file is part of OpenCV project.
-// It is subject to the license terms in the LICENSE file found in the top-level directory
-// of this distribution and at http://opencv.org/license.html.
-
-#ifndef OPENCV_VIDEOIO_REGISTRY_HPP
-#define OPENCV_VIDEOIO_REGISTRY_HPP
-
-#include <opencv2/videoio.hpp>
-
-namespace cv { namespace videoio_registry {
-/** @addtogroup videoio_registry
-This section contains API description how to query/configure available Video I/O backends.
-
-Runtime configuration options:
-- enable debug mode: `OPENCV_VIDEOIO_DEBUG=1`
-- change backend priority: `OPENCV_VIDEOIO_PRIORITY_<backend>=9999`
-- disable backend: `OPENCV_VIDEOIO_PRIORITY_<backend>=0`
-- specify list of backends with high priority (>100000): `OPENCV_VIDEOIO_PRIORITY_LIST=FFMPEG,GSTREAMER`
-
-@{
- */
-
-
-/** @brief Returns backend API name or "UnknownVideoAPI(xxx)"
-@param api backend ID (#VideoCaptureAPIs)
-*/
-CV_EXPORTS_W cv::String getBackendName(VideoCaptureAPIs api);
-
-/** @brief Returns list of all available backends */
-CV_EXPORTS_W std::vector<VideoCaptureAPIs> getBackends();
-
-/** @brief Returns list of available backends which works via `cv::VideoCapture(int index)` */
-CV_EXPORTS_W std::vector<VideoCaptureAPIs> getCameraBackends();
-
-/** @brief Returns list of available backends which works via `cv::VideoCapture(filename)` */
-CV_EXPORTS_W std::vector<VideoCaptureAPIs> getStreamBackends();
-
-/** @brief Returns list of available backends which works via `cv::VideoWriter()` */
-CV_EXPORTS_W std::vector<VideoCaptureAPIs> getWriterBackends();
-
-/** @brief Returns true if backend is available */
-CV_EXPORTS_W bool hasBackend(VideoCaptureAPIs api);
-
-/** @brief Returns true if backend is built in (false if backend is used as plugin) */
-CV_EXPORTS_W bool isBackendBuiltIn(VideoCaptureAPIs api);
-
-/** @brief Returns description and ABI/API version of videoio plugin's camera interface */
-CV_EXPORTS_W std::string getCameraBackendPluginVersion(
-    VideoCaptureAPIs api,
-    CV_OUT int& version_ABI,
-    CV_OUT int& version_API
-);
-
-/** @brief Returns description and ABI/API version of videoio plugin's stream capture interface */
-CV_EXPORTS_W std::string getStreamBackendPluginVersion(
-    VideoCaptureAPIs api,
-    CV_OUT int& version_ABI,
-    CV_OUT int& version_API
-);
-
-/** @brief Returns description and ABI/API version of videoio plugin's writer interface */
-CV_EXPORTS_W std::string getWriterBackendPluginVersion(
-    VideoCaptureAPIs api,
-    CV_OUT int& version_ABI,
-    CV_OUT int& version_API
-);
-
-
-//! @}
-}} // namespace
-
-#endif // OPENCV_VIDEOIO_REGISTRY_HPP
diff --git a/projects/llm_framework/include/opencv4/opencv2/videoio/videoio.hpp b/projects/llm_framework/include/opencv4/opencv2/videoio/videoio.hpp
deleted file mode 100644
index ec84cf7a..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/videoio/videoio.hpp
+++ /dev/null
@@ -1,48 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                          License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifdef __OPENCV_BUILD
-#error this is a compatibility header which should not be used inside the OpenCV library
-#endif
-
-#include "opencv2/videoio.hpp"
diff --git a/projects/llm_framework/include/opencv4/opencv2/videoio/videoio_c.h b/projects/llm_framework/include/opencv4/opencv2/videoio/videoio_c.h
deleted file mode 100644
index cf1a6d04..00000000
--- a/projects/llm_framework/include/opencv4/opencv2/videoio/videoio_c.h
+++ /dev/null
@@ -1,153 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                        Intel License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000, Intel Corporation, all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of Intel Corporation may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#ifndef OPENCV_VIDEOIO_H
-#define OPENCV_VIDEOIO_H
-
-#include "opencv2/core/core_c.h"
-
-#include "opencv2/videoio/legacy/constants_c.h"
-
-#ifdef __cplusplus
-extern "C" {
-#endif /* __cplusplus */
-
-/**
-  @addtogroup videoio_c
-  @{
-*/
-
-/****************************************************************************************\
-*                         Working with Video Files and Cameras                           *
-\****************************************************************************************/
-
-/** @brief "black box" capture structure
-
-In C++ use cv::VideoCapture
-*/
-typedef struct CvCapture CvCapture;
-
-/** @brief start capturing frames from video file
-*/
-CVAPI(CvCapture*) cvCreateFileCapture( const char* filename );
-
-/** @brief start capturing frames from video file. allows specifying a preferred API to use
-*/
-CVAPI(CvCapture*) cvCreateFileCaptureWithPreference( const char* filename , int apiPreference);
-
-/** @brief start capturing frames from camera: index = camera_index + domain_offset (CV_CAP_*)
-*/
-CVAPI(CvCapture*) cvCreateCameraCapture( int index );
-
-/** @brief grab a frame, return 1 on success, 0 on fail.
-
-  this function is thought to be fast
-*/
-CVAPI(int) cvGrabFrame( CvCapture* capture );
-
-/** @brief get the frame grabbed with cvGrabFrame(..)
-
-  This function may apply some frame processing like
-  frame decompression, flipping etc.
-  @warning !!!DO NOT RELEASE or MODIFY the retrieved frame!!!
-*/
-CVAPI(IplImage*) cvRetrieveFrame( CvCapture* capture, int streamIdx CV_DEFAULT(0) );
-
-/** @brief Just a combination of cvGrabFrame and cvRetrieveFrame
-
-  @warning !!!DO NOT RELEASE or MODIFY the retrieved frame!!!
-*/
-CVAPI(IplImage*) cvQueryFrame( CvCapture* capture );
-
-/** @brief stop capturing/reading and free resources
-*/
-CVAPI(void) cvReleaseCapture( CvCapture** capture );
-
-/** @brief retrieve capture properties
-*/
-CVAPI(double) cvGetCaptureProperty( CvCapture* capture, int property_id );
-/** @brief set capture properties
-*/
-CVAPI(int)    cvSetCaptureProperty( CvCapture* capture, int property_id, double value );
-
-/** @brief Return the type of the capturer (eg, ::CV_CAP_VFW, ::CV_CAP_UNICAP)
-
-It is unknown if created with ::CV_CAP_ANY
-*/
-CVAPI(int)    cvGetCaptureDomain( CvCapture* capture);
-
-/** @brief "black box" video file writer structure
-
-In C++ use cv::VideoWriter
-*/
-typedef struct CvVideoWriter CvVideoWriter;
-
-/** @brief initialize video file writer
-*/
-CVAPI(CvVideoWriter*) cvCreateVideoWriter( const char* filename, int fourcc,
-                                           double fps, CvSize frame_size,
-                                           int is_color CV_DEFAULT(1));
-
-/** @brief write frame to video file
-*/
-CVAPI(int) cvWriteFrame( CvVideoWriter* writer, const IplImage* image );
-
-/** @brief close video file writer
-*/
-CVAPI(void) cvReleaseVideoWriter( CvVideoWriter** writer );
-
-// ***************************************************************************************
-//! @name Obsolete functions/synonyms
-//! @{
-#define cvCaptureFromCAM cvCreateCameraCapture //!< @deprecated use cvCreateCameraCapture() instead
-#define cvCaptureFromFile cvCreateFileCapture  //!< @deprecated use cvCreateFileCapture() instead
-#define cvCaptureFromAVI cvCaptureFromFile     //!< @deprecated use cvCreateFileCapture() instead
-#define cvCreateAVIWriter cvCreateVideoWriter  //!< @deprecated use cvCreateVideoWriter() instead
-#define cvWriteToAVI cvWriteFrame              //!< @deprecated use cvWriteFrame() instead
-//!  @} Obsolete...
-
-//! @} videoio_c
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif //OPENCV_VIDEOIO_H
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/any.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/any.h
deleted file mode 100644
index e8f2cacf..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/any.h
+++ /dev/null
@@ -1,150 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_ANY_H__
-#define GOOGLE_PROTOBUF_ANY_H__
-
-#include <string>
-
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/arenastring.h>
-#include <google/protobuf/message_lite.h>
-
-#include <google/protobuf/port_def.inc>
-
-namespace google {
-namespace protobuf {
-
-class FieldDescriptor;
-class Message;
-
-namespace internal {
-
-extern const char kAnyFullTypeName[];          // "google.protobuf.Any".
-extern const char kTypeGoogleApisComPrefix[];  // "type.googleapis.com/".
-extern const char kTypeGoogleProdComPrefix[];  // "type.googleprod.com/".
-
-std::string GetTypeUrl(StringPiece message_name,
-                       StringPiece type_url_prefix);
-
-// Helper class used to implement google::protobuf::Any.
-class PROTOBUF_EXPORT AnyMetadata {
-  typedef ArenaStringPtr UrlType;
-  typedef ArenaStringPtr ValueType;
- public:
-  // AnyMetadata does not take ownership of "type_url" and "value".
-  constexpr AnyMetadata(UrlType* type_url, ValueType* value)
-      : type_url_(type_url), value_(value) {}
-
-  // Packs a message using the default type URL prefix: "type.googleapis.com".
-  // The resulted type URL will be "type.googleapis.com/<message_full_name>".
-  template <typename T>
-  void PackFrom(const T& message) {
-    InternalPackFrom(message, kTypeGoogleApisComPrefix, T::FullMessageName());
-  }
-
-  void PackFrom(const Message& message);
-
-  // Packs a message using the given type URL prefix. The type URL will be
-  // constructed by concatenating the message type's full name to the prefix
-  // with an optional "/" separator if the prefix doesn't already end with "/".
-  // For example, both PackFrom(message, "type.googleapis.com") and
-  // PackFrom(message, "type.googleapis.com/") yield the same result type
-  // URL: "type.googleapis.com/<message_full_name>".
-  template <typename T>
-  void PackFrom(const T& message, StringPiece type_url_prefix) {
-    InternalPackFrom(message, type_url_prefix, T::FullMessageName());
-  }
-
-  void PackFrom(const Message& message, StringPiece type_url_prefix);
-
-  // Unpacks the payload into the given message. Returns false if the message's
-  // type doesn't match the type specified in the type URL (i.e., the full
-  // name after the last "/" of the type URL doesn't match the message's actual
-  // full name) or parsing the payload has failed.
-  template <typename T>
-  bool UnpackTo(T* message) const {
-    return InternalUnpackTo(T::FullMessageName(), message);
-  }
-
-  bool UnpackTo(Message* message) const;
-
-  // Checks whether the type specified in the type URL matches the given type.
-  // A type is considered matching if its full name matches the full name after
-  // the last "/" in the type URL.
-  template <typename T>
-  bool Is() const {
-    return InternalIs(T::FullMessageName());
-  }
-
- private:
-  void InternalPackFrom(const MessageLite& message,
-                        StringPiece type_url_prefix,
-                        StringPiece type_name);
-  bool InternalUnpackTo(StringPiece type_name,
-                        MessageLite* message) const;
-  bool InternalIs(StringPiece type_name) const;
-
-  UrlType* type_url_;
-  ValueType* value_;
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(AnyMetadata);
-};
-
-// Get the proto type name from Any::type_url value. For example, passing
-// "type.googleapis.com/rpc.QueryOrigin" will return "rpc.QueryOrigin" in
-// *full_type_name. Returns false if the type_url does not have a "/"
-// in the type url separating the full type name.
-//
-// NOTE: this function is available publicly as:
-//   google::protobuf::Any()  // static method on the generated message type.
-bool ParseAnyTypeUrl(StringPiece type_url, std::string* full_type_name);
-
-// Get the proto type name and prefix from Any::type_url value. For example,
-// passing "type.googleapis.com/rpc.QueryOrigin" will return
-// "type.googleapis.com/" in *url_prefix and "rpc.QueryOrigin" in
-// *full_type_name. Returns false if the type_url does not have a "/" in the
-// type url separating the full type name.
-bool ParseAnyTypeUrl(StringPiece type_url, std::string* url_prefix,
-                     std::string* full_type_name);
-
-// See if message is of type google.protobuf.Any, if so, return the descriptors
-// for "type_url" and "value" fields.
-bool GetAnyFieldDescriptors(const Message& message,
-                            const FieldDescriptor** type_url_field,
-                            const FieldDescriptor** value_field);
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_ANY_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/arena.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/arena.h
deleted file mode 100644
index f28bebfd..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/arena.h
+++ /dev/null
@@ -1,694 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// This file defines an Arena allocator for better allocation performance.
-
-#ifndef GOOGLE_PROTOBUF_ARENA_H__
-#define GOOGLE_PROTOBUF_ARENA_H__
-
-
-#include <limits>
-#include <type_traits>
-#include <utility>
-#ifdef max
-#undef max  // Visual Studio defines this macro
-#endif
-#if defined(_MSC_VER) && !defined(_LIBCPP_STD_VER) && !_HAS_EXCEPTIONS
-// Work around bugs in MSVC <typeinfo> header when _HAS_EXCEPTIONS=0.
-#include <exception>
-#include <typeinfo>
-namespace std {
-using type_info = ::type_info;
-}
-#else
-#include <typeinfo>
-#endif
-
-#include <type_traits>
-#include <google/protobuf/arena_impl.h>
-#include <google/protobuf/port.h>
-
-#include <google/protobuf/port_def.inc>
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-namespace google {
-namespace protobuf {
-
-struct ArenaOptions;  // defined below
-
-}  // namespace protobuf
-}  // namespace google
-
-namespace google {
-namespace protobuf {
-
-class Arena;    // defined below
-class Message;  // defined in message.h
-class MessageLite;
-template <typename Key, typename T>
-class Map;
-
-namespace arena_metrics {
-
-void EnableArenaMetrics(ArenaOptions* options);
-
-}  // namespace arena_metrics
-
-namespace internal {
-
-struct ArenaStringPtr;  // defined in arenastring.h
-class LazyField;        // defined in lazy_field.h
-class EpsCopyInputStream;  // defined in parse_context.h
-
-template <typename Type>
-class GenericTypeHandler;  // defined in repeated_field.h
-
-// Templated cleanup methods.
-template <typename T>
-void arena_destruct_object(void* object) {
-  reinterpret_cast<T*>(object)->~T();
-}
-template <typename T>
-void arena_delete_object(void* object) {
-  delete reinterpret_cast<T*>(object);
-}
-}  // namespace internal
-
-// ArenaOptions provides optional additional parameters to arena construction
-// that control its block-allocation behavior.
-struct ArenaOptions {
-  // This defines the size of the first block requested from the system malloc.
-  // Subsequent block sizes will increase in a geometric series up to a maximum.
-  size_t start_block_size;
-
-  // This defines the maximum block size requested from system malloc (unless an
-  // individual arena allocation request occurs with a size larger than this
-  // maximum). Requested block sizes increase up to this value, then remain
-  // here.
-  size_t max_block_size;
-
-  // An initial block of memory for the arena to use, or NULL for none. If
-  // provided, the block must live at least as long as the arena itself. The
-  // creator of the Arena retains ownership of the block after the Arena is
-  // destroyed.
-  char* initial_block;
-
-  // The size of the initial block, if provided.
-  size_t initial_block_size;
-
-  // A function pointer to an alloc method that returns memory blocks of size
-  // requested. By default, it contains a ptr to the malloc function.
-  //
-  // NOTE: block_alloc and dealloc functions are expected to behave like
-  // malloc and free, including Asan poisoning.
-  void* (*block_alloc)(size_t);
-  // A function pointer to a dealloc method that takes ownership of the blocks
-  // from the arena. By default, it contains a ptr to a wrapper function that
-  // calls free.
-  void (*block_dealloc)(void*, size_t);
-
-  ArenaOptions()
-      : start_block_size(kDefaultStartBlockSize),
-        max_block_size(kDefaultMaxBlockSize),
-        initial_block(NULL),
-        initial_block_size(0),
-        block_alloc(kDefaultBlockAlloc),
-        block_dealloc(&internal::ArenaFree),
-        make_metrics_collector(nullptr) {}
-
-  PROTOBUF_EXPORT static void* (*const kDefaultBlockAlloc)(size_t);
-
- private:
-  // If make_metrics_collector is not nullptr, it will be called at Arena init
-  // time. It may return a pointer to a collector instance that will be notified
-  // of interesting events related to the arena.
-  internal::ArenaMetricsCollector* (*make_metrics_collector)();
-
-  // Constants define default starting block size and max block size for
-  // arena allocator behavior -- see descriptions above.
-  static const size_t kDefaultStartBlockSize =
-      internal::ArenaImpl::kDefaultStartBlockSize;
-  static const size_t kDefaultMaxBlockSize =
-      internal::ArenaImpl::kDefaultMaxBlockSize;
-
-  friend void arena_metrics::EnableArenaMetrics(ArenaOptions*);
-
-  friend class Arena;
-  friend class ArenaOptionsTestFriend;
-  friend class internal::ArenaImpl;
-};
-
-// Support for non-RTTI environments. (The metrics hooks API uses type
-// information.)
-#if PROTOBUF_RTTI
-#define RTTI_TYPE_ID(type) (&typeid(type))
-#else
-#define RTTI_TYPE_ID(type) (NULL)
-#endif
-
-// Arena allocator. Arena allocation replaces ordinary (heap-based) allocation
-// with new/delete, and improves performance by aggregating allocations into
-// larger blocks and freeing allocations all at once. Protocol messages are
-// allocated on an arena by using Arena::CreateMessage<T>(Arena*), below, and
-// are automatically freed when the arena is destroyed.
-//
-// This is a thread-safe implementation: multiple threads may allocate from the
-// arena concurrently. Destruction is not thread-safe and the destructing
-// thread must synchronize with users of the arena first.
-//
-// An arena provides two allocation interfaces: CreateMessage<T>, which works
-// for arena-enabled proto2 message types as well as other types that satisfy
-// the appropriate protocol (described below), and Create<T>, which works for
-// any arbitrary type T. CreateMessage<T> is better when the type T supports it,
-// because this interface (i) passes the arena pointer to the created object so
-// that its sub-objects and internal allocations can use the arena too, and (ii)
-// elides the object's destructor call when possible. Create<T> does not place
-// any special requirements on the type T, and will invoke the object's
-// destructor when the arena is destroyed.
-//
-// The arena message allocation protocol, required by
-// CreateMessage<T>(Arena* arena, Args&&... args), is as follows:
-//
-// - The type T must have (at least) two constructors: a constructor callable
-//   with `args` (without `arena`), called when a T is allocated on the heap;
-//   and a constructor callable with `Arena* arena, Args&&... args`, called when
-//   a T is allocated on an arena. If the second constructor is called with a
-//   NULL arena pointer, it must be equivalent to invoking the first
-//   (`args`-only) constructor.
-//
-// - The type T must have a particular type trait: a nested type
-//   |InternalArenaConstructable_|. This is usually a typedef to |void|. If no
-//   such type trait exists, then the instantiation CreateMessage<T> will fail
-//   to compile.
-//
-// - The type T *may* have the type trait |DestructorSkippable_|. If this type
-//   trait is present in the type, then its destructor will not be called if and
-//   only if it was passed a non-NULL arena pointer. If this type trait is not
-//   present on the type, then its destructor is always called when the
-//   containing arena is destroyed.
-//
-// This protocol is implemented by all arena-enabled proto2 message classes as
-// well as protobuf container types like RepeatedPtrField and Map. The protocol
-// is internal to protobuf and is not guaranteed to be stable. Non-proto types
-// should not rely on this protocol.
-class PROTOBUF_EXPORT PROTOBUF_ALIGNAS(8) Arena final {
- public:
-  // Default constructor with sensible default options, tuned for average
-  // use-cases.
-  inline Arena() : impl_() {}
-
-  // Construct an arena with default options, except for the supplied
-  // initial block. It is more efficient to use this constructor
-  // instead of passing ArenaOptions if the only configuration needed
-  // by the caller is supplying an initial block.
-  inline Arena(char* initial_block, size_t initial_block_size)
-      : impl_(initial_block, initial_block_size) {}
-
-  // Arena constructor taking custom options. See ArenaOptions above for
-  // descriptions of the options available.
-  explicit Arena(const ArenaOptions& options) : impl_(options) {}
-
-  // Block overhead.  Use this as a guide for how much to over-allocate the
-  // initial block if you want an allocation of size N to fit inside it.
-  //
-  // WARNING: if you allocate multiple objects, it is difficult to guarantee
-  // that a series of allocations will fit in the initial block, especially if
-  // Arena changes its alignment guarantees in the future!
-  static const size_t kBlockOverhead = internal::ArenaImpl::kBlockHeaderSize +
-                                       internal::ArenaImpl::kSerialArenaSize;
-
-  inline ~Arena() {}
-
-  // TODO(protobuf-team): Fix callers to use constructor and delete this method.
-  void Init(const ArenaOptions&) {}
-
-  // API to create proto2 message objects on the arena. If the arena passed in
-  // is NULL, then a heap allocated object is returned. Type T must be a message
-  // defined in a .proto file with cc_enable_arenas set to true, otherwise a
-  // compilation error will occur.
-  //
-  // RepeatedField and RepeatedPtrField may also be instantiated directly on an
-  // arena with this method.
-  //
-  // This function also accepts any type T that satisfies the arena message
-  // allocation protocol, documented above.
-  template <typename T, typename... Args>
-  PROTOBUF_ALWAYS_INLINE static T* CreateMessage(Arena* arena, Args&&... args) {
-    static_assert(
-        InternalHelper<T>::is_arena_constructable::value,
-        "CreateMessage can only construct types that are ArenaConstructable");
-    // We must delegate to CreateMaybeMessage() and NOT CreateMessageInternal()
-    // because protobuf generated classes specialize CreateMaybeMessage() and we
-    // need to use that specialization for code size reasons.
-    return Arena::CreateMaybeMessage<T>(arena, std::forward<Args>(args)...);
-  }
-
-  // API to create any objects on the arena. Note that only the object will
-  // be created on the arena; the underlying ptrs (in case of a proto2 message)
-  // will be still heap allocated. Proto messages should usually be allocated
-  // with CreateMessage<T>() instead.
-  //
-  // Note that even if T satisfies the arena message construction protocol
-  // (InternalArenaConstructable_ trait and optional DestructorSkippable_
-  // trait), as described above, this function does not follow the protocol;
-  // instead, it treats T as a black-box type, just as if it did not have these
-  // traits. Specifically, T's constructor arguments will always be only those
-  // passed to Create<T>() -- no additional arena pointer is implicitly added.
-  // Furthermore, the destructor will always be called at arena destruction time
-  // (unless the destructor is trivial). Hence, from T's point of view, it is as
-  // if the object were allocated on the heap (except that the underlying memory
-  // is obtained from the arena).
-  template <typename T, typename... Args>
-  PROTOBUF_ALWAYS_INLINE static T* Create(Arena* arena, Args&&... args) {
-    return CreateNoMessage<T>(arena, is_arena_constructable<T>(),
-                              std::forward<Args>(args)...);
-  }
-
-  // Create an array of object type T on the arena *without* invoking the
-  // constructor of T. If `arena` is null, then the return value should be freed
-  // with `delete[] x;` (or `::operator delete[](x);`).
-  // To ensure safe uses, this function checks at compile time
-  // (when compiled as C++11) that T is trivially default-constructible and
-  // trivially destructible.
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE static T* CreateArray(Arena* arena,
-                                               size_t num_elements) {
-    static_assert(std::is_pod<T>::value,
-                  "CreateArray requires a trivially constructible type");
-    static_assert(std::is_trivially_destructible<T>::value,
-                  "CreateArray requires a trivially destructible type");
-    GOOGLE_CHECK_LE(num_elements, std::numeric_limits<size_t>::max() / sizeof(T))
-        << "Requested size is too large to fit into size_t.";
-    if (arena == NULL) {
-      return static_cast<T*>(::operator new[](num_elements * sizeof(T)));
-    } else {
-      return arena->CreateInternalRawArray<T>(num_elements);
-    }
-  }
-
-  // Returns the total space allocated by the arena, which is the sum of the
-  // sizes of the underlying blocks. This method is relatively fast; a counter
-  // is kept as blocks are allocated.
-  uint64 SpaceAllocated() const { return impl_.SpaceAllocated(); }
-  // Returns the total space used by the arena. Similar to SpaceAllocated but
-  // does not include free space and block overhead. The total space returned
-  // may not include space used by other threads executing concurrently with
-  // the call to this method.
-  uint64 SpaceUsed() const { return impl_.SpaceUsed(); }
-
-  // Frees all storage allocated by this arena after calling destructors
-  // registered with OwnDestructor() and freeing objects registered with Own().
-  // Any objects allocated on this arena are unusable after this call. It also
-  // returns the total space used by the arena which is the sums of the sizes
-  // of the allocated blocks. This method is not thread-safe.
-  uint64 Reset() { return impl_.Reset(); }
-
-  // Adds |object| to a list of heap-allocated objects to be freed with |delete|
-  // when the arena is destroyed or reset.
-  template <typename T>
-  PROTOBUF_NOINLINE void Own(T* object) {
-    OwnInternal(object, std::is_convertible<T*, Message*>());
-  }
-
-  // Adds |object| to a list of objects whose destructors will be manually
-  // called when the arena is destroyed or reset. This differs from Own() in
-  // that it does not free the underlying memory with |delete|; hence, it is
-  // normally only used for objects that are placement-newed into
-  // arena-allocated memory.
-  template <typename T>
-  PROTOBUF_NOINLINE void OwnDestructor(T* object) {
-    if (object != NULL) {
-      impl_.AddCleanup(object, &internal::arena_destruct_object<T>);
-    }
-  }
-
-  // Adds a custom member function on an object to the list of destructors that
-  // will be manually called when the arena is destroyed or reset. This differs
-  // from OwnDestructor() in that any member function may be specified, not only
-  // the class destructor.
-  PROTOBUF_NOINLINE void OwnCustomDestructor(void* object,
-                                             void (*destruct)(void*)) {
-    impl_.AddCleanup(object, destruct);
-  }
-
-  // Retrieves the arena associated with |value| if |value| is an arena-capable
-  // message, or NULL otherwise. If possible, the call resolves at compile time.
-  // Note that we can often devirtualize calls to `value->GetArena()` so usually
-  // calling this method is unnecessary.
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE static Arena* GetArena(const T* value) {
-    return GetArenaInternal(value);
-  }
-
-  template <typename T>
-  class InternalHelper {
-    template <typename U>
-    static char DestructorSkippable(const typename U::DestructorSkippable_*);
-    template <typename U>
-    static double DestructorSkippable(...);
-
-    typedef std::integral_constant<
-        bool, sizeof(DestructorSkippable<T>(static_cast<const T*>(0))) ==
-                      sizeof(char) ||
-                  std::is_trivially_destructible<T>::value>
-        is_destructor_skippable;
-
-    template <typename U>
-    static char ArenaConstructable(
-        const typename U::InternalArenaConstructable_*);
-    template <typename U>
-    static double ArenaConstructable(...);
-
-    typedef std::integral_constant<bool, sizeof(ArenaConstructable<T>(
-                                             static_cast<const T*>(0))) ==
-                                             sizeof(char)>
-        is_arena_constructable;
-
-    template <typename U,
-              typename std::enable_if<
-                  std::is_same<Arena*, decltype(std::declval<const U>()
-                                                    .GetArena())>::value,
-                  int>::type = 0>
-    static char HasGetArena(decltype(&U::GetArena));
-    template <typename U>
-    static double HasGetArena(...);
-
-    typedef std::integral_constant<bool, sizeof(HasGetArena<T>(nullptr)) ==
-                                             sizeof(char)>
-        has_get_arena;
-
-    template <typename... Args>
-    static T* Construct(void* ptr, Args&&... args) {
-      return new (ptr) T(std::forward<Args>(args)...);
-    }
-
-    static Arena* GetArena(const T* p) { return p->GetArena(); }
-
-    friend class Arena;
-  };
-
-  // Helper typetraits that indicates support for arenas in a type T at compile
-  // time. This is public only to allow construction of higher-level templated
-  // utilities.
-  //
-  // is_arena_constructable<T>::value is true if the message type T has arena
-  // support enabled, and false otherwise.
-  //
-  // is_destructor_skippable<T>::value is true if the message type T has told
-  // the arena that it is safe to skip the destructor, and false otherwise.
-  //
-  // This is inside Arena because only Arena has the friend relationships
-  // necessary to see the underlying generated code traits.
-  template <typename T>
-  struct is_arena_constructable : InternalHelper<T>::is_arena_constructable {};
-  template <typename T>
-  struct is_destructor_skippable : InternalHelper<T>::is_destructor_skippable {
-  };
-
- private:
-  template <typename T>
-  struct has_get_arena : InternalHelper<T>::has_get_arena {};
-
-  template <typename T, typename... Args>
-  PROTOBUF_ALWAYS_INLINE static T* CreateMessageInternal(Arena* arena,
-                                                         Args&&... args) {
-    static_assert(
-        InternalHelper<T>::is_arena_constructable::value,
-        "CreateMessage can only construct types that are ArenaConstructable");
-    if (arena == NULL) {
-      return new T(nullptr, std::forward<Args>(args)...);
-    } else {
-      return arena->DoCreateMessage<T>(std::forward<Args>(args)...);
-    }
-  }
-
-  // This specialization for no arguments is necessary, because its behavior is
-  // slightly different.  When the arena pointer is nullptr, it calls T()
-  // instead of T(nullptr).
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE static T* CreateMessageInternal(Arena* arena) {
-    static_assert(
-        InternalHelper<T>::is_arena_constructable::value,
-        "CreateMessage can only construct types that are ArenaConstructable");
-    if (arena == NULL) {
-      return new T();
-    } else {
-      return arena->DoCreateMessage<T>();
-    }
-  }
-
-  template <typename T, typename... Args>
-  PROTOBUF_ALWAYS_INLINE static T* CreateInternal(Arena* arena,
-                                                  Args&&... args) {
-    if (arena == NULL) {
-      return new T(std::forward<Args>(args)...);
-    } else {
-      return arena->DoCreate<T>(std::is_trivially_destructible<T>::value,
-                                std::forward<Args>(args)...);
-    }
-  }
-
-  inline void AllocHook(const std::type_info* allocated_type, size_t n) const {
-    impl_.RecordAlloc(allocated_type, n);
-  }
-
-  // Allocate and also optionally call collector with the allocated type info
-  // when allocation recording is enabled.
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE void* AllocateInternal(bool skip_explicit_ownership) {
-    const size_t n = internal::AlignUpTo8(sizeof(T));
-    // Monitor allocation if needed.
-    impl_.RecordAlloc(RTTI_TYPE_ID(T), n);
-    if (skip_explicit_ownership) {
-      return AllocateAlignedTo<alignof(T)>(sizeof(T));
-    } else {
-      if (alignof(T) <= 8) {
-        return impl_.AllocateAlignedAndAddCleanup(
-            n, &internal::arena_destruct_object<T>);
-      } else {
-        auto ptr =
-            reinterpret_cast<uintptr_t>(impl_.AllocateAlignedAndAddCleanup(
-                sizeof(T) + alignof(T) - 8,
-                &internal::arena_destruct_object<T>));
-        return reinterpret_cast<void*>((ptr + alignof(T) - 8) &
-                                       (~alignof(T) + 1));
-      }
-    }
-  }
-
-  // CreateMessage<T> requires that T supports arenas, but this private method
-  // works whether or not T supports arenas. These are not exposed to user code
-  // as it can cause confusing API usages, and end up having double free in
-  // user code. These are used only internally from LazyField and Repeated
-  // fields, since they are designed to work in all mode combinations.
-  template <typename Msg, typename... Args>
-  PROTOBUF_ALWAYS_INLINE static Msg* DoCreateMaybeMessage(Arena* arena,
-                                                          std::true_type,
-                                                          Args&&... args) {
-    return CreateMessageInternal<Msg>(arena, std::forward<Args>(args)...);
-  }
-
-  template <typename T, typename... Args>
-  PROTOBUF_ALWAYS_INLINE static T* DoCreateMaybeMessage(Arena* arena,
-                                                        std::false_type,
-                                                        Args&&... args) {
-    return CreateInternal<T>(arena, std::forward<Args>(args)...);
-  }
-
-  template <typename T, typename... Args>
-  PROTOBUF_ALWAYS_INLINE static T* CreateMaybeMessage(Arena* arena,
-                                                      Args&&... args) {
-    return DoCreateMaybeMessage<T>(arena, is_arena_constructable<T>(),
-                                   std::forward<Args>(args)...);
-  }
-
-  template <typename T, typename... Args>
-  PROTOBUF_ALWAYS_INLINE static T* CreateNoMessage(Arena* arena, std::true_type,
-                                                   Args&&... args) {
-    // User is constructing with Create() despite the fact that T supports arena
-    // construction.  In this case we have to delegate to CreateInternal(), and
-    // we can't use any CreateMaybeMessage() specialization that may be defined.
-    return CreateInternal<T>(arena, std::forward<Args>(args)...);
-  }
-
-  template <typename T, typename... Args>
-  PROTOBUF_ALWAYS_INLINE static T* CreateNoMessage(Arena* arena,
-                                                   std::false_type,
-                                                   Args&&... args) {
-    // User is constructing with Create() and the type does not support arena
-    // construction.  In this case we can delegate to CreateMaybeMessage() and
-    // use any specialization that may be available for that.
-    return CreateMaybeMessage<T>(arena, std::forward<Args>(args)...);
-  }
-
-  // Just allocate the required size for the given type assuming the
-  // type has a trivial constructor.
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE T* CreateInternalRawArray(size_t num_elements) {
-    GOOGLE_CHECK_LE(num_elements, std::numeric_limits<size_t>::max() / sizeof(T))
-        << "Requested size is too large to fit into size_t.";
-    // We count on compiler to realize that if sizeof(T) is a multiple of
-    // 8 AlignUpTo can be elided.
-    const size_t n = internal::AlignUpTo8(sizeof(T) * num_elements);
-    // Monitor allocation if needed.
-    impl_.RecordAlloc(RTTI_TYPE_ID(T), n);
-    return static_cast<T*>(AllocateAlignedTo<alignof(T)>(n));
-  }
-
-  template <typename T, typename... Args>
-  PROTOBUF_ALWAYS_INLINE T* DoCreate(bool skip_explicit_ownership,
-                                     Args&&... args) {
-    return new (AllocateInternal<T>(skip_explicit_ownership))
-        T(std::forward<Args>(args)...);
-  }
-  template <typename T, typename... Args>
-  PROTOBUF_ALWAYS_INLINE T* DoCreateMessage(Args&&... args) {
-    return InternalHelper<T>::Construct(
-        AllocateInternal<T>(InternalHelper<T>::is_destructor_skippable::value),
-        this, std::forward<Args>(args)...);
-  }
-
-  // CreateInArenaStorage is used to implement map field. Without it,
-  // Map need to call generated message's protected arena constructor,
-  // which needs to declare Map as friend of generated message.
-  template <typename T, typename... Args>
-  static void CreateInArenaStorage(T* ptr, Arena* arena, Args&&... args) {
-    CreateInArenaStorageInternal(ptr, arena,
-                                 typename is_arena_constructable<T>::type(),
-                                 std::forward<Args>(args)...);
-    RegisterDestructorInternal(
-        ptr, arena,
-        typename InternalHelper<T>::is_destructor_skippable::type());
-  }
-
-  template <typename T, typename... Args>
-  static void CreateInArenaStorageInternal(T* ptr, Arena* arena,
-                                           std::true_type, Args&&... args) {
-    InternalHelper<T>::Construct(ptr, arena, std::forward<Args>(args)...);
-  }
-  template <typename T, typename... Args>
-  static void CreateInArenaStorageInternal(T* ptr, Arena* /* arena */,
-                                           std::false_type, Args&&... args) {
-    new (ptr) T(std::forward<Args>(args)...);
-  }
-
-  template <typename T>
-  static void RegisterDestructorInternal(T* /* ptr */, Arena* /* arena */,
-                                         std::true_type) {}
-  template <typename T>
-  static void RegisterDestructorInternal(T* ptr, Arena* arena,
-                                         std::false_type) {
-    arena->OwnDestructor(ptr);
-  }
-
-  // These implement Own(), which registers an object for deletion (destructor
-  // call and operator delete()). The second parameter has type 'true_type' if T
-  // is a subtype of Message and 'false_type' otherwise. Collapsing
-  // all template instantiations to one for generic Message reduces code size,
-  // using the virtual destructor instead.
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE void OwnInternal(T* object, std::true_type) {
-    if (object != NULL) {
-      impl_.AddCleanup(object, &internal::arena_delete_object<Message>);
-    }
-  }
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE void OwnInternal(T* object, std::false_type) {
-    if (object != NULL) {
-      impl_.AddCleanup(object, &internal::arena_delete_object<T>);
-    }
-  }
-
-  // Implementation for GetArena(). Only message objects with
-  // InternalArenaConstructable_ tags can be associated with an arena, and such
-  // objects must implement a GetArena() method.
-  template <typename T, typename std::enable_if<
-                            is_arena_constructable<T>::value, int>::type = 0>
-  PROTOBUF_ALWAYS_INLINE static Arena* GetArenaInternal(const T* value) {
-    return InternalHelper<T>::GetArena(value);
-  }
-  template <typename T,
-            typename std::enable_if<!is_arena_constructable<T>::value &&
-                                        has_get_arena<T>::value,
-                                    int>::type = 0>
-  PROTOBUF_ALWAYS_INLINE static Arena* GetArenaInternal(const T* value) {
-    return value->GetArena();
-  }
-  template <typename T,
-            typename std::enable_if<!is_arena_constructable<T>::value &&
-                                        !has_get_arena<T>::value,
-                                    int>::type = 0>
-  PROTOBUF_ALWAYS_INLINE static Arena* GetArenaInternal(const T* value) {
-    (void)value;
-    return nullptr;
-  }
-
-  // For friends of arena.
-  void* AllocateAligned(size_t n) {
-    return AllocateAlignedNoHook(internal::AlignUpTo8(n));
-  }
-  template<size_t Align>
-  void* AllocateAlignedTo(size_t n) {
-    static_assert(Align > 0, "Alignment must be greater than 0");
-    static_assert((Align & (Align - 1)) == 0, "Alignment must be power of two");
-    if (Align <= 8) return AllocateAligned(n);
-    // TODO(b/151247138): if the pointer would have been aligned already,
-    // this is wasting space. We should pass the alignment down.
-    uintptr_t ptr = reinterpret_cast<uintptr_t>(AllocateAligned(n + Align - 8));
-    ptr = (ptr + Align - 1) & (~Align + 1);
-    return reinterpret_cast<void*>(ptr);
-  }
-
-  void* AllocateAlignedNoHook(size_t n);
-
-  internal::ArenaImpl impl_;
-
-  template <typename Type>
-  friend class internal::GenericTypeHandler;
-  friend struct internal::ArenaStringPtr;  // For AllocateAligned.
-  friend class internal::LazyField;        // For CreateMaybeMessage.
-  friend class internal::EpsCopyInputStream;  // For parser performance
-  friend class MessageLite;
-  template <typename Key, typename T>
-  friend class Map;
-};
-
-// Defined above for supporting environments without RTTI.
-#undef RTTI_TYPE_ID
-
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_ARENA_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/arena_impl.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/arena_impl.h
deleted file mode 100644
index 13772686..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/arena_impl.h
+++ /dev/null
@@ -1,491 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// This file defines an Arena allocator for better allocation performance.
-
-#ifndef GOOGLE_PROTOBUF_ARENA_IMPL_H__
-#define GOOGLE_PROTOBUF_ARENA_IMPL_H__
-
-#include <atomic>
-#include <limits>
-
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/stubs/logging.h>
-
-#ifdef ADDRESS_SANITIZER
-#include <sanitizer/asan_interface.h>
-#endif  // ADDRESS_SANITIZER
-
-#include <google/protobuf/port_def.inc>
-
-
-namespace google {
-namespace protobuf {
-
-struct ArenaOptions;
-
-namespace internal {
-
-inline size_t AlignUpTo8(size_t n) {
-  // Align n to next multiple of 8 (from Hacker's Delight, Chapter 3.)
-  return (n + 7) & static_cast<size_t>(-8);
-}
-
-using LifecycleIdAtomic = uint64_t;
-
-void PROTOBUF_EXPORT ArenaFree(void* object, size_t size);
-
-// MetricsCollector collects stats for a particular arena.
-class PROTOBUF_EXPORT ArenaMetricsCollector {
- public:
-  virtual ~ArenaMetricsCollector();
-
-  // Invoked when the arena is about to be destroyed. This method will
-  // typically finalize any metric collection and delete the collector.
-  // space_allocated is the space used by the arena.
-  virtual void OnDestroy(uint64 space_allocated) = 0;
-
-  // OnReset() is called when the associated arena is reset.
-  // space_allocated is the space used by the arena just before the reset.
-  virtual void OnReset(uint64 space_allocated) = 0;
-
-  // Does OnAlloc() need to be called?  If false, metric collection overhead
-  // will be reduced since we will not do extra work per allocation.
-  virtual bool RecordAllocs() = 0;
-
-  // OnAlloc is called when an allocation happens.
-  // type_info is promised to be static - its lifetime extends to
-  // match program's lifetime (It is given by typeid operator).
-  // Note: typeid(void) will be passed as allocated_type every time we
-  // intentionally want to avoid monitoring an allocation. (i.e. internal
-  // allocations for managing the arena)
-  virtual void OnAlloc(const std::type_info* allocated_type,
-                       uint64 alloc_size) = 0;
-};
-
-class ArenaImpl;
-
-// A thread-unsafe Arena that can only be used within its owning thread.
-class PROTOBUF_EXPORT SerialArena {
- public:
-  // Blocks are variable length malloc-ed objects.  The following structure
-  // describes the common header for all blocks.
-  class PROTOBUF_EXPORT Block {
-   public:
-    Block(size_t size, Block* next, bool special, bool user_owned)
-        : next_and_bits_(reinterpret_cast<uintptr_t>(next) | (special ? 1 : 0) |
-                         (user_owned ? 2 : 0)),
-          pos_(kBlockHeaderSize),
-          size_(size) {
-      GOOGLE_DCHECK_EQ(reinterpret_cast<uintptr_t>(next) & 3, 0u);
-    }
-
-    char* Pointer(size_t n) {
-      GOOGLE_DCHECK(n <= size_);
-      return reinterpret_cast<char*>(this) + n;
-    }
-
-    // One of the blocks may be special. This is either a user-supplied
-    // initial block, or a block we created at startup to hold Options info.
-    // A special block is not deleted by Reset.
-    bool special() const { return (next_and_bits_ & 1) != 0; }
-
-    // Whether or not this current block is owned by the user.
-    // Only special blocks can be user_owned.
-    bool user_owned() const { return (next_and_bits_ & 2) != 0; }
-
-    Block* next() const {
-      const uintptr_t bottom_bits = 3;
-      return reinterpret_cast<Block*>(next_and_bits_ & ~bottom_bits);
-    }
-
-    void clear_next() {
-      next_and_bits_ &= 3;  // Set next to nullptr, preserve bottom bits.
-    }
-
-    size_t pos() const { return pos_; }
-    size_t size() const { return size_; }
-    void set_pos(size_t pos) { pos_ = pos; }
-
-   private:
-    // Holds pointer to next block for this thread + special/user_owned bits.
-    uintptr_t next_and_bits_;
-
-    size_t pos_;
-    size_t size_;
-    // data follows
-  };
-
-  // The allocate/free methods here are a little strange, since SerialArena is
-  // allocated inside a Block which it also manages.  This is to avoid doing
-  // an extra allocation for the SerialArena itself.
-
-  // Creates a new SerialArena inside Block* and returns it.
-  static SerialArena* New(Block* b, void* owner, ArenaImpl* arena);
-
-  void CleanupList();
-  uint64 SpaceUsed() const;
-
-  bool HasSpace(size_t n) { return n <= static_cast<size_t>(limit_ - ptr_); }
-
-  void* AllocateAligned(size_t n) {
-    GOOGLE_DCHECK_EQ(internal::AlignUpTo8(n), n);  // Must be already aligned.
-    GOOGLE_DCHECK_GE(limit_, ptr_);
-    if (PROTOBUF_PREDICT_FALSE(!HasSpace(n))) {
-      return AllocateAlignedFallback(n);
-    }
-    void* ret = ptr_;
-    ptr_ += n;
-#ifdef ADDRESS_SANITIZER
-    ASAN_UNPOISON_MEMORY_REGION(ret, n);
-#endif  // ADDRESS_SANITIZER
-    return ret;
-  }
-
-  // Allocate space if the current region provides enough space.
-  bool MaybeAllocateAligned(size_t n, void** out) {
-    GOOGLE_DCHECK_EQ(internal::AlignUpTo8(n), n);  // Must be already aligned.
-    GOOGLE_DCHECK_GE(limit_, ptr_);
-    if (PROTOBUF_PREDICT_FALSE(!HasSpace(n))) return false;
-    void* ret = ptr_;
-    ptr_ += n;
-#ifdef ADDRESS_SANITIZER
-    ASAN_UNPOISON_MEMORY_REGION(ret, n);
-#endif  // ADDRESS_SANITIZER
-    *out = ret;
-    return true;
-  }
-
-  void AddCleanup(void* elem, void (*cleanup)(void*)) {
-    if (PROTOBUF_PREDICT_FALSE(cleanup_ptr_ == cleanup_limit_)) {
-      AddCleanupFallback(elem, cleanup);
-      return;
-    }
-    cleanup_ptr_->elem = elem;
-    cleanup_ptr_->cleanup = cleanup;
-    cleanup_ptr_++;
-  }
-
-  void* AllocateAlignedAndAddCleanup(size_t n, void (*cleanup)(void*)) {
-    void* ret = AllocateAligned(n);
-    AddCleanup(ret, cleanup);
-    return ret;
-  }
-
-  Block* head() const { return head_; }
-  void* owner() const { return owner_; }
-  SerialArena* next() const { return next_; }
-  void set_next(SerialArena* next) { next_ = next; }
-  static Block* NewBlock(Block* last_block, size_t min_bytes, ArenaImpl* arena);
-
- private:
-  // Node contains the ptr of the object to be cleaned up and the associated
-  // cleanup function ptr.
-  struct CleanupNode {
-    void* elem;              // Pointer to the object to be cleaned up.
-    void (*cleanup)(void*);  // Function pointer to the destructor or deleter.
-  };
-
-  // Cleanup uses a chunked linked list, to reduce pointer chasing.
-  struct CleanupChunk {
-    static size_t SizeOf(size_t i) {
-      return sizeof(CleanupChunk) + (sizeof(CleanupNode) * (i - 1));
-    }
-    size_t size;           // Total elements in the list.
-    CleanupChunk* next;    // Next node in the list.
-    CleanupNode nodes[1];  // True length is |size|.
-  };
-
-  ArenaImpl* arena_;       // Containing arena.
-  void* owner_;            // &ThreadCache of this thread;
-  Block* head_;            // Head of linked list of blocks.
-  CleanupChunk* cleanup_;  // Head of cleanup list.
-  SerialArena* next_;      // Next SerialArena in this linked list.
-
-  // Next pointer to allocate from.  Always 8-byte aligned.  Points inside
-  // head_ (and head_->pos will always be non-canonical).  We keep these
-  // here to reduce indirection.
-  char* ptr_;
-  char* limit_;
-
-  // Next CleanupList members to append to.  These point inside cleanup_.
-  CleanupNode* cleanup_ptr_;
-  CleanupNode* cleanup_limit_;
-
-  void* AllocateAlignedFallback(size_t n);
-  void AddCleanupFallback(void* elem, void (*cleanup)(void*));
-  void CleanupListFallback();
-
- public:
-  static constexpr size_t kBlockHeaderSize =
-      (sizeof(Block) + 7) & static_cast<size_t>(-8);
-};
-
-// This class provides the core Arena memory allocation library. Different
-// implementations only need to implement the public interface below.
-// Arena is not a template type as that would only be useful if all protos
-// in turn would be templates, which will/cannot happen. However separating
-// the memory allocation part from the cruft of the API users expect we can
-// use #ifdef the select the best implementation based on hardware / OS.
-class PROTOBUF_EXPORT ArenaImpl {
- public:
-  static const size_t kDefaultStartBlockSize = 256;
-  static const size_t kDefaultMaxBlockSize = 8192;
-
-  ArenaImpl() { Init(false); }
-
-  ArenaImpl(char* mem, size_t size) {
-    GOOGLE_DCHECK_EQ(reinterpret_cast<uintptr_t>(mem) & 7, 0u);
-    Init(false);
-
-    // Ignore initial block if it is too small.
-    if (mem != nullptr && size >= kBlockHeaderSize + kSerialArenaSize) {
-      SetInitialBlock(new (mem) SerialArena::Block(size, nullptr, true, true));
-    }
-  }
-
-  explicit ArenaImpl(const ArenaOptions& options);
-
-  // Destructor deletes all owned heap allocated objects, and destructs objects
-  // that have non-trivial destructors, except for proto2 message objects whose
-  // destructors can be skipped. Also, frees all blocks except the initial block
-  // if it was passed in.
-  ~ArenaImpl();
-
-  uint64 Reset();
-
-  uint64 SpaceAllocated() const;
-  uint64 SpaceUsed() const;
-
-  void* AllocateAligned(size_t n) {
-    SerialArena* arena;
-    if (PROTOBUF_PREDICT_TRUE(GetSerialArenaFast(&arena))) {
-      return arena->AllocateAligned(n);
-    } else {
-      return AllocateAlignedFallback(n);
-    }
-  }
-
-  // This function allocates n bytes if the common happy case is true and
-  // returns true. Otherwise does nothing and returns false. This strange
-  // semantics is necessary to allow callers to program functions that only
-  // have fallback function calls in tail position. This substantially improves
-  // code for the happy path.
-  PROTOBUF_ALWAYS_INLINE bool MaybeAllocateAligned(size_t n, void** out) {
-    SerialArena* a;
-    if (PROTOBUF_PREDICT_TRUE(GetSerialArenaFromThreadCache(&a))) {
-      return a->MaybeAllocateAligned(n, out);
-    }
-    return false;
-  }
-
-  void* AllocateAlignedAndAddCleanup(size_t n, void (*cleanup)(void*));
-
-  // Add object pointer and cleanup function pointer to the list.
-  void AddCleanup(void* elem, void (*cleanup)(void*));
-
-  inline void RecordAlloc(const std::type_info* allocated_type,
-                          size_t n) const {
-    if (PROTOBUF_PREDICT_FALSE(record_allocs())) {
-      options_->metrics_collector->OnAlloc(allocated_type, n);
-    }
-  }
-
-  std::pair<void*, size_t> NewBuffer(size_t last_size, size_t min_bytes);
-
- private:
-  // Pointer to a linked list of SerialArena.
-  std::atomic<SerialArena*> threads_;
-  std::atomic<SerialArena*> hint_;       // Fast thread-local block access
-  std::atomic<size_t> space_allocated_;  // Total size of all allocated blocks.
-
-  // Unique for each arena. Changes on Reset().
-  // Least-significant-bit is 1 iff allocations should be recorded.
-  uint64 lifecycle_id_;
-
-  struct Options {
-    size_t start_block_size;
-    size_t max_block_size;
-    void* (*block_alloc)(size_t);
-    void (*block_dealloc)(void*, size_t);
-    ArenaMetricsCollector* metrics_collector;
-  };
-
-  Options* options_ = nullptr;
-
-  void* AllocateAlignedFallback(size_t n);
-  void* AllocateAlignedAndAddCleanupFallback(size_t n, void (*cleanup)(void*));
-  void AddCleanupFallback(void* elem, void (*cleanup)(void*));
-
-  void Init(bool record_allocs);
-  void SetInitialBlock(
-      SerialArena::Block* block);  // Can be called right after Init()
-
-  // Return true iff allocations should be recorded in a metrics collector.
-  inline bool record_allocs() const { return lifecycle_id_ & 1; }
-
-  // Invoke fn(b) for every Block* b.
-  template <typename Functor>
-  void PerBlock(Functor fn) {
-    // By omitting an Acquire barrier we ensure that any user code that doesn't
-    // properly synchronize Reset() or the destructor will throw a TSAN warning.
-    SerialArena* serial = threads_.load(std::memory_order_relaxed);
-    while (serial) {
-      // fn() may delete blocks and arenas, so fetch next pointers before fn();
-      SerialArena* cur = serial;
-      serial = serial->next();
-      for (auto* block = cur->head(); block != nullptr;) {
-        auto* b = block;
-        block = b->next();
-        fn(b);
-      }
-    }
-  }
-
-  // Delete or Destruct all objects owned by the arena.
-  void CleanupList();
-
-  inline void CacheSerialArena(SerialArena* serial) {
-    thread_cache().last_serial_arena = serial;
-    thread_cache().last_lifecycle_id_seen = lifecycle_id_;
-    // TODO(haberman): evaluate whether we would gain efficiency by getting rid
-    // of hint_.  It's the only write we do to ArenaImpl in the allocation path,
-    // which will dirty the cache line.
-
-    hint_.store(serial, std::memory_order_release);
-  }
-
-  PROTOBUF_ALWAYS_INLINE bool GetSerialArenaFast(SerialArena** arena) {
-    if (GetSerialArenaFromThreadCache(arena)) return true;
-
-    // Check whether we own the last accessed SerialArena on this arena.  This
-    // fast path optimizes the case where a single thread uses multiple arenas.
-    ThreadCache* tc = &thread_cache();
-    SerialArena* serial = hint_.load(std::memory_order_acquire);
-    if (PROTOBUF_PREDICT_TRUE(serial != NULL && serial->owner() == tc)) {
-      *arena = serial;
-      return true;
-    }
-    return false;
-  }
-
-  PROTOBUF_ALWAYS_INLINE bool GetSerialArenaFromThreadCache(
-      SerialArena** arena) {
-    // If this thread already owns a block in this arena then try to use that.
-    // This fast path optimizes the case where multiple threads allocate from
-    // the same arena.
-    ThreadCache* tc = &thread_cache();
-    if (PROTOBUF_PREDICT_TRUE(tc->last_lifecycle_id_seen == lifecycle_id_)) {
-      *arena = tc->last_serial_arena;
-      return true;
-    }
-    return false;
-  }
-  SerialArena* GetSerialArenaFallback(void* me);
-
-#ifdef _MSC_VER
-#pragma warning(disable : 4324)
-#endif
-  struct alignas(64) ThreadCache {
-#if defined(GOOGLE_PROTOBUF_NO_THREADLOCAL)
-    // If we are using the ThreadLocalStorage class to store the ThreadCache,
-    // then the ThreadCache's default constructor has to be responsible for
-    // initializing it.
-    ThreadCache()
-        : next_lifecycle_id(0),
-          last_lifecycle_id_seen(-1),
-          last_serial_arena(NULL) {}
-#endif
-
-    // Number of per-thread lifecycle IDs to reserve. Must be power of two.
-    // To reduce contention on a global atomic, each thread reserves a batch of
-    // IDs.  The following number is calculated based on a stress test with
-    // ~6500 threads all frequently allocating a new arena.
-    static constexpr size_t kPerThreadIds = 256;
-    // Next lifecycle ID available to this thread. We need to reserve a new
-    // batch, if `next_lifecycle_id & (kPerThreadIds - 1) == 0`.
-    uint64 next_lifecycle_id;
-    // The ThreadCache is considered valid as long as this matches the
-    // lifecycle_id of the arena being used.
-    uint64 last_lifecycle_id_seen;
-    SerialArena* last_serial_arena;
-  };
-
-  // Lifecycle_id can be highly contended variable in a situation of lots of
-  // arena creation. Make sure that other global variables are not sharing the
-  // cacheline.
-#ifdef _MSC_VER
-#pragma warning(disable : 4324)
-#endif
-  struct alignas(64) CacheAlignedLifecycleIdGenerator {
-    std::atomic<LifecycleIdAtomic> id;
-  };
-  static CacheAlignedLifecycleIdGenerator lifecycle_id_generator_;
-#if defined(GOOGLE_PROTOBUF_NO_THREADLOCAL)
-  // Android ndk does not support __thread keyword so we use a custom thread
-  // local storage class we implemented.
-  // iOS also does not support the __thread keyword.
-  static ThreadCache& thread_cache();
-#elif defined(PROTOBUF_USE_DLLS)
-  // Thread local variables cannot be exposed through DLL interface but we can
-  // wrap them in static functions.
-  static ThreadCache& thread_cache();
-#else
-  static PROTOBUF_THREAD_LOCAL ThreadCache thread_cache_;
-  static ThreadCache& thread_cache() { return thread_cache_; }
-#endif
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ArenaImpl);
-  // All protos have pointers back to the arena hence Arena must have
-  // pointer stability.
-  ArenaImpl(ArenaImpl&&) = delete;
-  ArenaImpl& operator=(ArenaImpl&&) = delete;
-
- public:
-  // kBlockHeaderSize is sizeof(Block), aligned up to the nearest multiple of 8
-  // to protect the invariant that pos is always at a multiple of 8.
-  static constexpr size_t kBlockHeaderSize = SerialArena::kBlockHeaderSize;
-  static constexpr size_t kSerialArenaSize =
-      (sizeof(SerialArena) + 7) & static_cast<size_t>(-8);
-  static constexpr size_t kOptionsSize =
-      (sizeof(Options) + 7) & static_cast<size_t>(-8);
-  static_assert(kBlockHeaderSize % 8 == 0,
-                "kBlockHeaderSize must be a multiple of 8.");
-  static_assert(kSerialArenaSize % 8 == 0,
-                "kSerialArenaSize must be a multiple of 8.");
-};
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_ARENA_IMPL_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/arenastring.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/arenastring.h
deleted file mode 100644
index 60307b27..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/arenastring.h
+++ /dev/null
@@ -1,375 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_ARENASTRING_H__
-#define GOOGLE_PROTOBUF_ARENASTRING_H__
-
-#include <string>
-#include <type_traits>
-#include <utility>
-
-#include <google/protobuf/stubs/logging.h>
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/arena.h>
-#include <google/protobuf/port.h>
-
-#include <google/protobuf/port_def.inc>
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-// Lazy string instance to support string fields with non-empty default.
-// These are initialized on the first call to .get().
-class PROTOBUF_EXPORT LazyString {
- public:
-  // We explicitly make LazyString an aggregate so that MSVC can do constant
-  // initialization on it without marking it `constexpr`.
-  // We do not want to use `constexpr` because it makes it harder to have extern
-  // storage for it and causes library bloat.
-  struct InitValue {
-    const char* ptr;
-    size_t size;
-  };
-  // We keep a union of the initialization value and the std::string to save on
-  // space. We don't need the string array after Init() is done.
-  union {
-    mutable InitValue init_value_;
-    alignas(std::string) mutable char string_buf_[sizeof(std::string)];
-  };
-  mutable std::atomic<const std::string*> inited_;
-
-  const std::string& get() const {
-    // This check generates less code than a call-once invocation.
-    auto* res = inited_.load(std::memory_order_acquire);
-    if (PROTOBUF_PREDICT_FALSE(res == nullptr)) return Init();
-    return *res;
-  }
-
- private:
-  // Initialize the string in `string_buf_`, update `inited_` and return it.
-  // We return it here to avoid having to read it again in the inlined code.
-  const std::string& Init() const;
-};
-
-template <typename T>
-class TaggedPtr {
- public:
-  TaggedPtr() = default;
-  explicit constexpr TaggedPtr(const std::string* ptr)
-      : ptr_(const_cast<std::string*>(ptr)) {}
-
-  void SetTagged(T* p) {
-    Set(p);
-    ptr_ = reinterpret_cast<void*>(as_int() | 1);
-  }
-  void Set(T* p) { ptr_ = p; }
-  T* Get() const { return reinterpret_cast<T*>(as_int() & -2); }
-  bool IsTagged() const { return as_int() & 1; }
-
-  // Returned value is only safe to dereference if IsTagged() == false.
-  // It is safe to compare.
-  T* UnsafeGet() const { return static_cast<T*>(ptr_); }
-
-  bool IsNull() { return ptr_ == nullptr; }
-
- private:
-  uintptr_t as_int() const { return reinterpret_cast<uintptr_t>(ptr_); }
-  void* ptr_;
-};
-
-static_assert(std::is_trivial<TaggedPtr<std::string>>::value,
-              "TaggedPtr must be trivial");
-
-// This class encapsulates a pointer to a std::string with or without a donated
-// buffer, tagged by bottom bit. It is a high-level wrapper that almost directly
-// corresponds to the interface required by string fields in generated
-// code. It replaces the old std::string* pointer in such cases.
-//
-// The object has different but similar code paths for when the default value is
-// the empty string and when it is a non-empty string.
-// The empty string is handled different throughout the library and there is a
-// single global instance of it we can share.
-//
-// For fields with an empty string default value, there are three distinct
-// states:
-//
-// - Pointer set to 'String' tag (LSB is 0), equal to
-//   &GetEmptyStringAlreadyInited(): field is set to its default value. Points
-//   to a true std::string*, but we do not own that std::string* (it's a
-//   globally shared instance).
-//
-// - Pointer set to 'String' tag (LSB is 0), but not equal to the global empty
-//   string: field points to a true std::string* instance that we own. This
-//   instance is either on the heap or on the arena (i.e. registered on
-//   free()/destructor-call list) as appropriate.
-//
-// - Pointer set to 'DonatedString' tag (LSB is 1): points to a std::string
-//   instance with a buffer on the arena (arena != NULL, always, in this case).
-//
-// For fields with a non-empty string default value, there are three distinct
-// states:
-//
-// - Pointer set to 'String' tag (LSB is 0), equal to `nullptr`:
-//   Field is in "default" mode and does not point to any actual instance.
-//   Methods that might need to create an instance of the object will pass a
-//   `const LazyString&` for it.
-//
-// - Pointer set to 'String' tag (LSB is 0), but not equal to `nullptr`:
-//   field points to a true std::string* instance that we own. This instance is
-//   either on the heap or on the arena (i.e. registered on
-//   free()/destructor-call list) as appropriate.
-//
-// - Pointer set to 'DonatedString' tag (LSB is 1): points to a std::string
-//   instance with a buffer on the arena (arena != NULL, always, in this case).
-//
-// Generated code and reflection code both ensure that ptr_ is never null for
-// fields with an empty default.
-// Because ArenaStringPtr is used in oneof unions, its constructor is a NOP and
-// so the field is always manually initialized via method calls.
-//
-// Side-note: why pass information about the default on every API call? Because
-// we don't want to hold it in a member variable, or else this would go into
-// every proto message instance. This would be a huge waste of space, since the
-// default instance pointer is typically a global (static class field). We want
-// the generated code to be as efficient as possible, and if we take
-// the default value information as a parameter that's in practice taken from a
-// static class field, and compare ptr_ to the default value, we end up with a
-// single "cmp %reg, GLOBAL" in the resulting machine code. (Note that this also
-// requires the String tag to be 0 so we can avoid the mask before comparing.)
-struct PROTOBUF_EXPORT ArenaStringPtr {
-  ArenaStringPtr() = default;
-  explicit constexpr ArenaStringPtr(const std::string* default_value)
-      : tagged_ptr_(default_value) {}
-
-  // Some methods below are overloaded on a `default_value` and on tags.
-  // The tagged overloads help reduce code size in the callers in generated
-  // code, while the `default_value` overloads are useful from reflection.
-  // By-value empty struct arguments are elided in the ABI.
-  struct EmptyDefault {};
-  struct NonEmptyDefault {};
-
-  void Set(const std::string* default_value, ConstStringParam value,
-           ::google::protobuf::Arena* arena);
-  void Set(const std::string* default_value, std::string&& value,
-           ::google::protobuf::Arena* arena);
-  void Set(EmptyDefault, ConstStringParam value, ::google::protobuf::Arena* arena);
-  void Set(EmptyDefault, std::string&& value, ::google::protobuf::Arena* arena);
-  void Set(NonEmptyDefault, ConstStringParam value, ::google::protobuf::Arena* arena);
-  void Set(NonEmptyDefault, std::string&& value, ::google::protobuf::Arena* arena);
-
-  // Basic accessors.
-  const std::string& Get() const PROTOBUF_ALWAYS_INLINE {
-    // Unconditionally mask away the tag.
-    return *tagged_ptr_.Get();
-  }
-  const std::string* GetPointer() const PROTOBUF_ALWAYS_INLINE {
-    // Unconditionally mask away the tag.
-    return tagged_ptr_.Get();
-  }
-
-  // For fields with an empty default value.
-  std::string* Mutable(EmptyDefault, ::google::protobuf::Arena* arena);
-  // For fields with a non-empty default value.
-  std::string* Mutable(const LazyString& default_value, ::google::protobuf::Arena* arena);
-
-  // Release returns a std::string* instance that is heap-allocated and is not
-  // Own()'d by any arena. If the field is not set, this returns NULL. The
-  // caller retains ownership. Clears this field back to NULL state. Used to
-  // implement release_<field>() methods on generated classes.
-  std::string* Release(const std::string* default_value,
-                       ::google::protobuf::Arena* arena);
-  std::string* ReleaseNonDefault(const std::string* default_value,
-                                 ::google::protobuf::Arena* arena);
-
-  // Takes a std::string that is heap-allocated, and takes ownership. The
-  // std::string's destructor is registered with the arena. Used to implement
-  // set_allocated_<field> in generated classes.
-  void SetAllocated(const std::string* default_value, std::string* value,
-                    ::google::protobuf::Arena* arena);
-
-  // Swaps internal pointers. Arena-safety semantics: this is guarded by the
-  // logic in Swap()/UnsafeArenaSwap() at the message level, so this method is
-  // 'unsafe' if called directly.
-  inline void Swap(ArenaStringPtr* other, const std::string* default_value,
-                   Arena* arena) PROTOBUF_ALWAYS_INLINE;
-
-  // Frees storage (if not on an arena).
-  void Destroy(const std::string* default_value, ::google::protobuf::Arena* arena);
-  void Destroy(EmptyDefault, ::google::protobuf::Arena* arena);
-  void Destroy(NonEmptyDefault, ::google::protobuf::Arena* arena);
-
-  // Clears content, but keeps allocated std::string, to avoid the overhead of
-  // heap operations. After this returns, the content (as seen by the user) will
-  // always be the empty std::string. Assumes that |default_value| is an empty
-  // std::string.
-  void ClearToEmpty();
-
-  // Clears content, assuming that the current value is not the empty
-  // string default.
-  void ClearNonDefaultToEmpty();
-
-  // Clears content, but keeps allocated std::string if arena != NULL, to avoid
-  // the overhead of heap operations. After this returns, the content (as seen
-  // by the user) will always be equal to |default_value|.
-  void ClearToDefault(const LazyString& default_value, ::google::protobuf::Arena* arena);
-
-  // Called from generated code / reflection runtime only. Resets value to point
-  // to a default string pointer, with the semantics that this
-  // ArenaStringPtr does not own the pointed-to memory. Disregards initial value
-  // of ptr_ (so this is the *ONLY* safe method to call after construction or
-  // when reinitializing after becoming the active field in a oneof union).
-  inline void UnsafeSetDefault(const std::string* default_value);
-
-  // Returns a mutable pointer, but doesn't initialize the string to the
-  // default value.
-  std::string* MutableNoArenaNoDefault(const std::string* default_value);
-
-  // Get a mutable pointer with unspecified contents.
-  // Similar to `MutableNoArenaNoDefault`, but also handles the arena case.
-  // If the value was donated, the contents are discarded.
-  std::string* MutableNoCopy(const std::string* default_value,
-                             ::google::protobuf::Arena* arena);
-
-  // Destroy the string. Assumes `arena == nullptr`.
-  void DestroyNoArena(const std::string* default_value);
-
-  // Internal setter used only at parse time to directly set a donated string
-  // value.
-  void UnsafeSetTaggedPointer(TaggedPtr<std::string> value) {
-    tagged_ptr_ = value;
-  }
-  // Generated code only! An optimization, in certain cases the generated
-  // code is certain we can obtain a std::string with no default checks and
-  // tag tests.
-  std::string* UnsafeMutablePointer() PROTOBUF_RETURNS_NONNULL;
-
-  inline bool IsDefault(const std::string* default_value) const {
-    // Relies on the fact that kPtrTagString == 0, so if IsString(), ptr_ is the
-    // actual std::string pointer (and if !IsString(), ptr_ will never be equal
-    // to any aligned |default_value| pointer). The key is that we want to avoid
-    // masking in the fastpath const-pointer Get() case for non-arena code.
-    return tagged_ptr_.UnsafeGet() == default_value;
-  }
-
- private:
-  TaggedPtr<std::string> tagged_ptr_;
-
-  bool IsDonatedString() const { return false; }
-
-  // Slow paths.
-
-  // MutableSlow requires that !IsString() || IsDefault
-  // Variadic to support 0 args for EmptyDefault and 1 arg for LazyString.
-  template <typename... Lazy>
-  std::string* MutableSlow(::google::protobuf::Arena* arena, const Lazy&... lazy_default);
-
-};
-
-inline void ArenaStringPtr::UnsafeSetDefault(const std::string* value) {
-  tagged_ptr_.Set(const_cast<std::string*>(value));
-}
-
-inline void ArenaStringPtr::Swap(ArenaStringPtr* other,
-                                 const std::string* default_value,
-                                 Arena* arena) {
-#ifndef NDEBUG
-  // For debug builds, we swap the contents of the string, rather than the
-  // std::string instances themselves.  This invalidates previously taken const
-  // references that are (per our documentation) invalidated by calling Swap()
-  // on the message.
-  //
-  // If both strings are the default_value, swapping is uninteresting.
-  // Otherwise, we use ArenaStringPtr::Mutable() to access the std::string, to
-  // ensure that we do not try to mutate default_value itself.
-  if (IsDefault(default_value) && other->IsDefault(default_value)) {
-    return;
-  }
-
-  if (default_value == nullptr) {
-    // If we have non-empty default, then `default_value` is null and we can't
-    // call Mutable the same way. Just do the regular swap.
-    std::swap(tagged_ptr_, other->tagged_ptr_);
-  } else {
-    std::string* this_ptr = Mutable(EmptyDefault{}, arena);
-    std::string* other_ptr = other->Mutable(EmptyDefault{}, arena);
-
-    this_ptr->swap(*other_ptr);
-  }
-#else
-  std::swap(tagged_ptr_, other->tagged_ptr_);
-#endif
-}
-
-inline void ArenaStringPtr::ClearNonDefaultToEmpty() {
-  // Unconditionally mask away the tag.
-  tagged_ptr_.Get()->clear();
-}
-
-inline std::string* ArenaStringPtr::MutableNoArenaNoDefault(
-    const std::string* default_value) {
-  // VERY IMPORTANT for performance and code size: this will reduce to a member
-  // variable load, a pointer check (against |default_value|, in practice a
-  // static global) and a branch to the slowpath (which calls operator new and
-  // the ctor). DO NOT add any tagged-pointer operations here.
-  if (IsDefault(default_value)) {
-    std::string* new_string = new std::string();
-    tagged_ptr_.Set(new_string);
-    return new_string;
-  } else {
-    return UnsafeMutablePointer();
-  }
-}
-
-inline void ArenaStringPtr::DestroyNoArena(const std::string* default_value) {
-  if (!IsDefault(default_value)) {
-    delete UnsafeMutablePointer();
-  }
-}
-
-inline std::string* ArenaStringPtr::UnsafeMutablePointer() {
-  GOOGLE_DCHECK(!tagged_ptr_.IsTagged());
-  GOOGLE_DCHECK(tagged_ptr_.UnsafeGet() != nullptr);
-  return tagged_ptr_.UnsafeGet();
-}
-
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_ARENASTRING_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/descriptor.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/descriptor.h
deleted file mode 100644
index 5bfecf50..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/descriptor.h
+++ /dev/null
@@ -1,2324 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// Author: kenton@google.com (Kenton Varda)
-//  Based on original Protocol Buffers design by
-//  Sanjay Ghemawat, Jeff Dean, and others.
-//
-// This file contains classes which describe a type of protocol message.
-// You can use a message's descriptor to learn at runtime what fields
-// it contains and what the types of those fields are.  The Message
-// interface also allows you to dynamically access and modify individual
-// fields by passing the FieldDescriptor of the field you are interested
-// in.
-//
-// Most users will not care about descriptors, because they will write
-// code specific to certain protocol types and will simply use the classes
-// generated by the protocol compiler directly.  Advanced users who want
-// to operate on arbitrary types (not known at compile time) may want to
-// read descriptors in order to learn about the contents of a message.
-// A very small number of users will want to construct their own
-// Descriptors, either because they are implementing Message manually or
-// because they are writing something like the protocol compiler.
-//
-// For an example of how you might use descriptors, see the code example
-// at the top of message.h.
-
-#ifndef GOOGLE_PROTOBUF_DESCRIPTOR_H__
-#define GOOGLE_PROTOBUF_DESCRIPTOR_H__
-
-#include <atomic>
-#include <map>
-#include <memory>
-#include <set>
-#include <string>
-#include <vector>
-
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/stubs/mutex.h>
-#include <google/protobuf/stubs/once.h>
-#include <google/protobuf/port.h>
-#include <google/protobuf/port_def.inc>
-
-// TYPE_BOOL is defined in the MacOS's ConditionalMacros.h.
-#ifdef TYPE_BOOL
-#undef TYPE_BOOL
-#endif  // TYPE_BOOL
-
-#ifdef SWIG
-#define PROTOBUF_EXPORT
-#endif
-
-
-namespace google {
-namespace protobuf {
-
-// Defined in this file.
-class Descriptor;
-class FieldDescriptor;
-class OneofDescriptor;
-class EnumDescriptor;
-class EnumValueDescriptor;
-class ServiceDescriptor;
-class MethodDescriptor;
-class FileDescriptor;
-class DescriptorDatabase;
-class DescriptorPool;
-
-// Defined in descriptor.proto
-class DescriptorProto;
-class DescriptorProto_ExtensionRange;
-class FieldDescriptorProto;
-class OneofDescriptorProto;
-class EnumDescriptorProto;
-class EnumValueDescriptorProto;
-class ServiceDescriptorProto;
-class MethodDescriptorProto;
-class FileDescriptorProto;
-class MessageOptions;
-class FieldOptions;
-class OneofOptions;
-class EnumOptions;
-class EnumValueOptions;
-class ExtensionRangeOptions;
-class ServiceOptions;
-class MethodOptions;
-class FileOptions;
-class UninterpretedOption;
-class SourceCodeInfo;
-
-// Defined in message.h
-class Message;
-class Reflection;
-
-// Defined in descriptor.cc
-class DescriptorBuilder;
-class FileDescriptorTables;
-struct Symbol;
-
-// Defined in unknown_field_set.h.
-class UnknownField;
-
-// Defined in command_line_interface.cc
-namespace compiler {
-class CommandLineInterface;
-namespace cpp {
-// Defined in helpers.h
-class Formatter;
-}  // namespace cpp
-}  // namespace compiler
-
-namespace descriptor_unittest {
-class DescriptorTest;
-}  // namespace descriptor_unittest
-
-// Defined in printer.h
-namespace io {
-class Printer;
-}  // namespace io
-
-// NB, all indices are zero-based.
-struct SourceLocation {
-  int start_line;
-  int end_line;
-  int start_column;
-  int end_column;
-
-  // Doc comments found at the source location.
-  // See the comments in SourceCodeInfo.Location (descriptor.proto) for details.
-  std::string leading_comments;
-  std::string trailing_comments;
-  std::vector<std::string> leading_detached_comments;
-};
-
-// Options when generating machine-parsable output from a descriptor with
-// DebugString().
-struct DebugStringOptions {
-  // include original user comments as recorded in SourceLocation entries. N.B.
-  // that this must be |false| by default: several other pieces of code (for
-  // example, the C++ code generation for fields in the proto compiler) rely on
-  // DebugString() output being unobstructed by user comments.
-  bool include_comments;
-  // If true, elide the braced body in the debug string.
-  bool elide_group_body;
-  bool elide_oneof_body;
-
-  DebugStringOptions()
-      : include_comments(false),
-        elide_group_body(false),
-        elide_oneof_body(false) {
-  }
-};
-
-// A class to handle the simplest cases of a lazily linked descriptor
-// for a message type that isn't built at the time of cross linking,
-// which is needed when a pool has lazily_build_dependencies_ set.
-// Must be instantiated as mutable in a descriptor.
-namespace internal {
-class PROTOBUF_EXPORT LazyDescriptor {
- public:
-  // Init function to be called at init time of a descriptor containing
-  // a LazyDescriptor.
-  void Init() {
-    descriptor_ = nullptr;
-    name_ = nullptr;
-    once_ = nullptr;
-    file_ = nullptr;
-  }
-
-  // Sets the value of the descriptor if it is known during the descriptor
-  // building process. Not thread safe, should only be called during the
-  // descriptor build process. Should not be called after SetLazy has been
-  // called.
-  void Set(const Descriptor* descriptor);
-
-  // Sets the information needed to lazily cross link the descriptor at a later
-  // time, SetLazy is not thread safe, should be called only once at descriptor
-  // build time if the symbol wasn't found and building of the file containing
-  // that type is delayed because lazily_build_dependencies_ is set on the pool.
-  // Should not be called after Set() has been called.
-  void SetLazy(StringPiece name, const FileDescriptor* file);
-
-  // Returns the current value of the descriptor, thread-safe. If SetLazy(...)
-  // has been called, will do a one-time cross link of the type specified,
-  // building the descriptor file that contains the type if necessary.
-  inline const Descriptor* Get() {
-    Once();
-    return descriptor_;
-  }
-
- private:
-  static void OnceStatic(LazyDescriptor* lazy);
-  void OnceInternal();
-  void Once();
-
-  const Descriptor* descriptor_;
-  const std::string* name_;
-  internal::once_flag* once_;
-  const FileDescriptor* file_;
-};
-}  // namespace internal
-
-// Describes a type of protocol message, or a particular group within a
-// message.  To obtain the Descriptor for a given message object, call
-// Message::GetDescriptor().  Generated message classes also have a
-// static method called descriptor() which returns the type's descriptor.
-// Use DescriptorPool to construct your own descriptors.
-class PROTOBUF_EXPORT Descriptor {
- public:
-  typedef DescriptorProto Proto;
-
-  // The name of the message type, not including its scope.
-  const std::string& name() const;
-
-  // The fully-qualified name of the message type, scope delimited by
-  // periods.  For example, message type "Foo" which is declared in package
-  // "bar" has full name "bar.Foo".  If a type "Baz" is nested within
-  // Foo, Baz's full_name is "bar.Foo.Baz".  To get only the part that
-  // comes after the last '.', use name().
-  const std::string& full_name() const;
-
-  // Index of this descriptor within the file or containing type's message
-  // type array.
-  int index() const;
-
-  // The .proto file in which this message type was defined.  Never nullptr.
-  const FileDescriptor* file() const;
-
-  // If this Descriptor describes a nested type, this returns the type
-  // in which it is nested.  Otherwise, returns nullptr.
-  const Descriptor* containing_type() const;
-
-  // Get options for this message type.  These are specified in the .proto file
-  // by placing lines like "option foo = 1234;" in the message definition.
-  // Allowed options are defined by MessageOptions in descriptor.proto, and any
-  // available extensions of that message.
-  const MessageOptions& options() const;
-
-  // Write the contents of this Descriptor into the given DescriptorProto.
-  // The target DescriptorProto must be clear before calling this; if it
-  // isn't, the result may be garbage.
-  void CopyTo(DescriptorProto* proto) const;
-
-  // Write the contents of this descriptor in a human-readable form. Output
-  // will be suitable for re-parsing.
-  std::string DebugString() const;
-
-  // Similar to DebugString(), but additionally takes options (e.g.,
-  // include original user comments in output).
-  std::string DebugStringWithOptions(const DebugStringOptions& options) const;
-
-  // Returns true if this is a placeholder for an unknown type. This will
-  // only be the case if this descriptor comes from a DescriptorPool
-  // with AllowUnknownDependencies() set.
-  bool is_placeholder() const;
-
-  enum WellKnownType {
-    WELLKNOWNTYPE_UNSPECIFIED,  // Not a well-known type.
-
-    // Wrapper types.
-    WELLKNOWNTYPE_DOUBLEVALUE,  // google.protobuf.DoubleValue
-    WELLKNOWNTYPE_FLOATVALUE,   // google.protobuf.FloatValue
-    WELLKNOWNTYPE_INT64VALUE,   // google.protobuf.Int64Value
-    WELLKNOWNTYPE_UINT64VALUE,  // google.protobuf.UInt64Value
-    WELLKNOWNTYPE_INT32VALUE,   // google.protobuf.Int32Value
-    WELLKNOWNTYPE_UINT32VALUE,  // google.protobuf.UInt32Value
-    WELLKNOWNTYPE_STRINGVALUE,  // google.protobuf.StringValue
-    WELLKNOWNTYPE_BYTESVALUE,   // google.protobuf.BytesValue
-    WELLKNOWNTYPE_BOOLVALUE,    // google.protobuf.BoolValue
-
-    // Other well known types.
-    WELLKNOWNTYPE_ANY,        // google.protobuf.Any
-    WELLKNOWNTYPE_FIELDMASK,  // google.protobuf.FieldMask
-    WELLKNOWNTYPE_DURATION,   // google.protobuf.Duration
-    WELLKNOWNTYPE_TIMESTAMP,  // google.protobuf.Timestamp
-    WELLKNOWNTYPE_VALUE,      // google.protobuf.Value
-    WELLKNOWNTYPE_LISTVALUE,  // google.protobuf.ListValue
-    WELLKNOWNTYPE_STRUCT,     // google.protobuf.Struct
-
-    // New well-known types may be added in the future.
-    // Please make sure any switch() statements have a 'default' case.
-    __WELLKNOWNTYPE__DO_NOT_USE__ADD_DEFAULT_INSTEAD__,
-  };
-
-  WellKnownType well_known_type() const;
-
-  // Field stuff -----------------------------------------------------
-
-  // The number of fields in this message type.
-  int field_count() const;
-  // Gets a field by index, where 0 <= index < field_count().
-  // These are returned in the order they were defined in the .proto file.
-  const FieldDescriptor* field(int index) const;
-
-  // Looks up a field by declared tag number.  Returns nullptr if no such field
-  // exists.
-  const FieldDescriptor* FindFieldByNumber(int number) const;
-  // Looks up a field by name.  Returns nullptr if no such field exists.
-  const FieldDescriptor* FindFieldByName(ConstStringParam name) const;
-
-  // Looks up a field by lowercased name (as returned by lowercase_name()).
-  // This lookup may be ambiguous if multiple field names differ only by case,
-  // in which case the field returned is chosen arbitrarily from the matches.
-  const FieldDescriptor* FindFieldByLowercaseName(
-      ConstStringParam lowercase_name) const;
-
-  // Looks up a field by camel-case name (as returned by camelcase_name()).
-  // This lookup may be ambiguous if multiple field names differ in a way that
-  // leads them to have identical camel-case names, in which case the field
-  // returned is chosen arbitrarily from the matches.
-  const FieldDescriptor* FindFieldByCamelcaseName(
-      ConstStringParam camelcase_name) const;
-
-  // The number of oneofs in this message type.
-  int oneof_decl_count() const;
-  // The number of oneofs in this message type, excluding synthetic oneofs.
-  // Real oneofs always come first, so iterating up to real_oneof_decl_cout()
-  // will yield all real oneofs.
-  int real_oneof_decl_count() const;
-  // Get a oneof by index, where 0 <= index < oneof_decl_count().
-  // These are returned in the order they were defined in the .proto file.
-  const OneofDescriptor* oneof_decl(int index) const;
-
-  // Looks up a oneof by name.  Returns nullptr if no such oneof exists.
-  const OneofDescriptor* FindOneofByName(ConstStringParam name) const;
-
-  // Nested type stuff -----------------------------------------------
-
-  // The number of nested types in this message type.
-  int nested_type_count() const;
-  // Gets a nested type by index, where 0 <= index < nested_type_count().
-  // These are returned in the order they were defined in the .proto file.
-  const Descriptor* nested_type(int index) const;
-
-  // Looks up a nested type by name.  Returns nullptr if no such nested type
-  // exists.
-  const Descriptor* FindNestedTypeByName(ConstStringParam name) const;
-
-  // Enum stuff ------------------------------------------------------
-
-  // The number of enum types in this message type.
-  int enum_type_count() const;
-  // Gets an enum type by index, where 0 <= index < enum_type_count().
-  // These are returned in the order they were defined in the .proto file.
-  const EnumDescriptor* enum_type(int index) const;
-
-  // Looks up an enum type by name.  Returns nullptr if no such enum type
-  // exists.
-  const EnumDescriptor* FindEnumTypeByName(ConstStringParam name) const;
-
-  // Looks up an enum value by name, among all enum types in this message.
-  // Returns nullptr if no such value exists.
-  const EnumValueDescriptor* FindEnumValueByName(ConstStringParam name) const;
-
-  // Extensions ------------------------------------------------------
-
-  // A range of field numbers which are designated for third-party
-  // extensions.
-  struct ExtensionRange {
-    typedef DescriptorProto_ExtensionRange Proto;
-
-    typedef ExtensionRangeOptions OptionsType;
-
-    // See Descriptor::CopyTo().
-    void CopyTo(DescriptorProto_ExtensionRange* proto) const;
-
-    int start;  // inclusive
-    int end;    // exclusive
-
-    const ExtensionRangeOptions* options_;
-  };
-
-  // The number of extension ranges in this message type.
-  int extension_range_count() const;
-  // Gets an extension range by index, where 0 <= index <
-  // extension_range_count(). These are returned in the order they were defined
-  // in the .proto file.
-  const ExtensionRange* extension_range(int index) const;
-
-  // Returns true if the number is in one of the extension ranges.
-  bool IsExtensionNumber(int number) const;
-
-  // Returns nullptr if no extension range contains the given number.
-  const ExtensionRange* FindExtensionRangeContainingNumber(int number) const;
-
-  // The number of extensions defined nested within this message type's scope.
-  // See doc:
-  // https://developers.google.com/protocol-buffers/docs/proto#nested-extensions
-  //
-  // Note that the extensions may be extending *other* messages.
-  //
-  // For example:
-  // message M1 {
-  //   extensions 1 to max;
-  // }
-  //
-  // message M2 {
-  //   extend M1 {
-  //     optional int32 foo = 1;
-  //   }
-  // }
-  //
-  // In this case,
-  // DescriptorPool::generated_pool()
-  //     ->FindMessageTypeByName("M2")
-  //     ->extension(0)
-  // will return "foo", even though "foo" is an extension of M1.
-  // To find all known extensions of a given message, instead use
-  // DescriptorPool::FindAllExtensions.
-  int extension_count() const;
-  // Get an extension by index, where 0 <= index < extension_count().
-  // These are returned in the order they were defined in the .proto file.
-  const FieldDescriptor* extension(int index) const;
-
-  // Looks up a named extension (which extends some *other* message type)
-  // defined within this message type's scope.
-  const FieldDescriptor* FindExtensionByName(ConstStringParam name) const;
-
-  // Similar to FindFieldByLowercaseName(), but finds extensions defined within
-  // this message type's scope.
-  const FieldDescriptor* FindExtensionByLowercaseName(
-      ConstStringParam name) const;
-
-  // Similar to FindFieldByCamelcaseName(), but finds extensions defined within
-  // this message type's scope.
-  const FieldDescriptor* FindExtensionByCamelcaseName(
-      ConstStringParam name) const;
-
-  // Reserved fields -------------------------------------------------
-
-  // A range of reserved field numbers.
-  struct ReservedRange {
-    int start;  // inclusive
-    int end;    // exclusive
-  };
-
-  // The number of reserved ranges in this message type.
-  int reserved_range_count() const;
-  // Gets an reserved range by index, where 0 <= index <
-  // reserved_range_count(). These are returned in the order they were defined
-  // in the .proto file.
-  const ReservedRange* reserved_range(int index) const;
-
-  // Returns true if the number is in one of the reserved ranges.
-  bool IsReservedNumber(int number) const;
-
-  // Returns nullptr if no reserved range contains the given number.
-  const ReservedRange* FindReservedRangeContainingNumber(int number) const;
-
-  // The number of reserved field names in this message type.
-  int reserved_name_count() const;
-
-  // Gets a reserved name by index, where 0 <= index < reserved_name_count().
-  const std::string& reserved_name(int index) const;
-
-  // Returns true if the field name is reserved.
-  bool IsReservedName(ConstStringParam name) const;
-
-  // Source Location ---------------------------------------------------
-
-  // Updates |*out_location| to the source location of the complete
-  // extent of this message declaration.  Returns false and leaves
-  // |*out_location| unchanged iff location information was not available.
-  bool GetSourceLocation(SourceLocation* out_location) const;
-
-  // Maps --------------------------------------------------------------
-
-  // Returns the FieldDescriptor for the "key" field. If this isn't a map entry
-  // field, returns nullptr.
-  const FieldDescriptor* map_key() const;
-
-  // Returns the FieldDescriptor for the "value" field. If this isn't a map
-  // entry field, returns nullptr.
-  const FieldDescriptor* map_value() const;
-
- private:
-  typedef MessageOptions OptionsType;
-
-  // Allows tests to test CopyTo(proto, true).
-  friend class descriptor_unittest::DescriptorTest;
-
-  // Allows access to GetLocationPath for annotations.
-  friend class io::Printer;
-  friend class compiler::cpp::Formatter;
-
-  // Fill the json_name field of FieldDescriptorProto.
-  void CopyJsonNameTo(DescriptorProto* proto) const;
-
-  // Internal version of DebugString; controls the level of indenting for
-  // correct depth. Takes |options| to control debug-string options, and
-  // |include_opening_clause| to indicate whether the "message ... " part of the
-  // clause has already been generated (this varies depending on context).
-  void DebugString(int depth, std::string* contents,
-                   const DebugStringOptions& options,
-                   bool include_opening_clause) const;
-
-  // Walks up the descriptor tree to generate the source location path
-  // to this descriptor from the file root.
-  void GetLocationPath(std::vector<int>* output) const;
-
-  const std::string* name_;
-  const std::string* full_name_;
-  const FileDescriptor* file_;
-  const Descriptor* containing_type_;
-  const MessageOptions* options_;
-
-  // These arrays are separated from their sizes to minimize padding on 64-bit.
-  FieldDescriptor* fields_;
-  OneofDescriptor* oneof_decls_;
-  Descriptor* nested_types_;
-  EnumDescriptor* enum_types_;
-  ExtensionRange* extension_ranges_;
-  FieldDescriptor* extensions_;
-  ReservedRange* reserved_ranges_;
-  const std::string** reserved_names_;
-
-  int field_count_;
-  int oneof_decl_count_;
-  int real_oneof_decl_count_;
-  int nested_type_count_;
-  int enum_type_count_;
-  int extension_range_count_;
-  int extension_count_;
-  int reserved_range_count_;
-  int reserved_name_count_;
-
-  // True if this is a placeholder for an unknown type.
-  bool is_placeholder_;
-  // True if this is a placeholder and the type name wasn't fully-qualified.
-  bool is_unqualified_placeholder_;
-  // Well known type.  Stored as char to conserve space.
-  char well_known_type_;
-
-  // IMPORTANT:  If you add a new field, make sure to search for all instances
-  // of Allocate<Descriptor>() and AllocateArray<Descriptor>() in descriptor.cc
-  // and update them to initialize the field.
-
-  // Must be constructed using DescriptorPool.
-  Descriptor() {}
-  friend class DescriptorBuilder;
-  friend class DescriptorPool;
-  friend class EnumDescriptor;
-  friend class FieldDescriptor;
-  friend class OneofDescriptor;
-  friend class MethodDescriptor;
-  friend class FileDescriptor;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(Descriptor);
-};
-
-
-// Describes a single field of a message.  To get the descriptor for a given
-// field, first get the Descriptor for the message in which it is defined,
-// then call Descriptor::FindFieldByName().  To get a FieldDescriptor for
-// an extension, do one of the following:
-// - Get the Descriptor or FileDescriptor for its containing scope, then
-//   call Descriptor::FindExtensionByName() or
-//   FileDescriptor::FindExtensionByName().
-// - Given a DescriptorPool, call DescriptorPool::FindExtensionByNumber() or
-//   DescriptorPool::FindExtensionByPrintableName().
-// Use DescriptorPool to construct your own descriptors.
-class PROTOBUF_EXPORT FieldDescriptor {
- public:
-  typedef FieldDescriptorProto Proto;
-
-  // Identifies a field type.  0 is reserved for errors.  The order is weird
-  // for historical reasons.  Types 12 and up are new in proto2.
-  enum Type {
-    TYPE_DOUBLE = 1,    // double, exactly eight bytes on the wire.
-    TYPE_FLOAT = 2,     // float, exactly four bytes on the wire.
-    TYPE_INT64 = 3,     // int64, varint on the wire.  Negative numbers
-                        // take 10 bytes.  Use TYPE_SINT64 if negative
-                        // values are likely.
-    TYPE_UINT64 = 4,    // uint64, varint on the wire.
-    TYPE_INT32 = 5,     // int32, varint on the wire.  Negative numbers
-                        // take 10 bytes.  Use TYPE_SINT32 if negative
-                        // values are likely.
-    TYPE_FIXED64 = 6,   // uint64, exactly eight bytes on the wire.
-    TYPE_FIXED32 = 7,   // uint32, exactly four bytes on the wire.
-    TYPE_BOOL = 8,      // bool, varint on the wire.
-    TYPE_STRING = 9,    // UTF-8 text.
-    TYPE_GROUP = 10,    // Tag-delimited message.  Deprecated.
-    TYPE_MESSAGE = 11,  // Length-delimited message.
-
-    TYPE_BYTES = 12,     // Arbitrary byte array.
-    TYPE_UINT32 = 13,    // uint32, varint on the wire
-    TYPE_ENUM = 14,      // Enum, varint on the wire
-    TYPE_SFIXED32 = 15,  // int32, exactly four bytes on the wire
-    TYPE_SFIXED64 = 16,  // int64, exactly eight bytes on the wire
-    TYPE_SINT32 = 17,    // int32, ZigZag-encoded varint on the wire
-    TYPE_SINT64 = 18,    // int64, ZigZag-encoded varint on the wire
-
-    MAX_TYPE = 18,  // Constant useful for defining lookup tables
-                    // indexed by Type.
-  };
-
-  // Specifies the C++ data type used to represent the field.  There is a
-  // fixed mapping from Type to CppType where each Type maps to exactly one
-  // CppType.  0 is reserved for errors.
-  enum CppType {
-    CPPTYPE_INT32 = 1,     // TYPE_INT32, TYPE_SINT32, TYPE_SFIXED32
-    CPPTYPE_INT64 = 2,     // TYPE_INT64, TYPE_SINT64, TYPE_SFIXED64
-    CPPTYPE_UINT32 = 3,    // TYPE_UINT32, TYPE_FIXED32
-    CPPTYPE_UINT64 = 4,    // TYPE_UINT64, TYPE_FIXED64
-    CPPTYPE_DOUBLE = 5,    // TYPE_DOUBLE
-    CPPTYPE_FLOAT = 6,     // TYPE_FLOAT
-    CPPTYPE_BOOL = 7,      // TYPE_BOOL
-    CPPTYPE_ENUM = 8,      // TYPE_ENUM
-    CPPTYPE_STRING = 9,    // TYPE_STRING, TYPE_BYTES
-    CPPTYPE_MESSAGE = 10,  // TYPE_MESSAGE, TYPE_GROUP
-
-    MAX_CPPTYPE = 10,  // Constant useful for defining lookup tables
-                       // indexed by CppType.
-  };
-
-  // Identifies whether the field is optional, required, or repeated.  0 is
-  // reserved for errors.
-  enum Label {
-    LABEL_OPTIONAL = 1,  // optional
-    LABEL_REQUIRED = 2,  // required
-    LABEL_REPEATED = 3,  // repeated
-
-    MAX_LABEL = 3,  // Constant useful for defining lookup tables
-                    // indexed by Label.
-  };
-
-  // Valid field numbers are positive integers up to kMaxNumber.
-  static const int kMaxNumber = (1 << 29) - 1;
-
-  // First field number reserved for the protocol buffer library implementation.
-  // Users may not declare fields that use reserved numbers.
-  static const int kFirstReservedNumber = 19000;
-  // Last field number reserved for the protocol buffer library implementation.
-  // Users may not declare fields that use reserved numbers.
-  static const int kLastReservedNumber = 19999;
-
-  const std::string& name() const;  // Name of this field within the message.
-  const std::string& full_name() const;  // Fully-qualified name of the field.
-  const std::string& json_name() const;  // JSON name of this field.
-  const FileDescriptor* file() const;  // File in which this field was defined.
-  bool is_extension() const;           // Is this an extension field?
-  int number() const;                  // Declared tag number.
-
-  // Same as name() except converted to lower-case.  This (and especially the
-  // FindFieldByLowercaseName() method) can be useful when parsing formats
-  // which prefer to use lowercase naming style.  (Although, technically
-  // field names should be lowercased anyway according to the protobuf style
-  // guide, so this only makes a difference when dealing with old .proto files
-  // which do not follow the guide.)
-  const std::string& lowercase_name() const;
-
-  // Same as name() except converted to camel-case.  In this conversion, any
-  // time an underscore appears in the name, it is removed and the next
-  // letter is capitalized.  Furthermore, the first letter of the name is
-  // lower-cased.  Examples:
-  //   FooBar -> fooBar
-  //   foo_bar -> fooBar
-  //   fooBar -> fooBar
-  // This (and especially the FindFieldByCamelcaseName() method) can be useful
-  // when parsing formats which prefer to use camel-case naming style.
-  const std::string& camelcase_name() const;
-
-  Type type() const;                  // Declared type of this field.
-  const char* type_name() const;      // Name of the declared type.
-  CppType cpp_type() const;           // C++ type of this field.
-  const char* cpp_type_name() const;  // Name of the C++ type.
-  Label label() const;                // optional/required/repeated
-
-  bool is_required() const;  // shorthand for label() == LABEL_REQUIRED
-  bool is_optional() const;  // shorthand for label() == LABEL_OPTIONAL
-  bool is_repeated() const;  // shorthand for label() == LABEL_REPEATED
-  bool is_packable() const;  // shorthand for is_repeated() &&
-                             //               IsTypePackable(type())
-  bool is_packed() const;    // shorthand for is_packable() &&
-                             //               options().packed()
-  bool is_map() const;       // shorthand for type() == TYPE_MESSAGE &&
-                             // message_type()->options().map_entry()
-
-  // Returns true if this field was syntactically written with "optional" in the
-  // .proto file. Excludes singular proto3 fields that do not have a label.
-  bool has_optional_keyword() const;
-
-  // Returns true if this field tracks presence, ie. does the field
-  // distinguish between "unset" and "present with default value."
-  // This includes required, optional, and oneof fields. It excludes maps,
-  // repeated fields, and singular proto3 fields without "optional".
-  //
-  // For fields where has_presence() == true, the return value of
-  // Reflection::HasField() is semantically meaningful.
-  bool has_presence() const;
-
-  // Index of this field within the message's field array, or the file or
-  // extension scope's extensions array.
-  int index() const;
-
-  // Does this field have an explicitly-declared default value?
-  bool has_default_value() const;
-
-  // Whether the user has specified the json_name field option in the .proto
-  // file.
-  bool has_json_name() const;
-
-  // Get the field default value if cpp_type() == CPPTYPE_INT32.  If no
-  // explicit default was defined, the default is 0.
-  int32 default_value_int32() const;
-  // Get the field default value if cpp_type() == CPPTYPE_INT64.  If no
-  // explicit default was defined, the default is 0.
-  int64 default_value_int64() const;
-  // Get the field default value if cpp_type() == CPPTYPE_UINT32.  If no
-  // explicit default was defined, the default is 0.
-  uint32 default_value_uint32() const;
-  // Get the field default value if cpp_type() == CPPTYPE_UINT64.  If no
-  // explicit default was defined, the default is 0.
-  uint64 default_value_uint64() const;
-  // Get the field default value if cpp_type() == CPPTYPE_FLOAT.  If no
-  // explicit default was defined, the default is 0.0.
-  float default_value_float() const;
-  // Get the field default value if cpp_type() == CPPTYPE_DOUBLE.  If no
-  // explicit default was defined, the default is 0.0.
-  double default_value_double() const;
-  // Get the field default value if cpp_type() == CPPTYPE_BOOL.  If no
-  // explicit default was defined, the default is false.
-  bool default_value_bool() const;
-  // Get the field default value if cpp_type() == CPPTYPE_ENUM.  If no
-  // explicit default was defined, the default is the first value defined
-  // in the enum type (all enum types are required to have at least one value).
-  // This never returns nullptr.
-  const EnumValueDescriptor* default_value_enum() const;
-  // Get the field default value if cpp_type() == CPPTYPE_STRING.  If no
-  // explicit default was defined, the default is the empty string.
-  const std::string& default_value_string() const;
-
-  // The Descriptor for the message of which this is a field.  For extensions,
-  // this is the extended type.  Never nullptr.
-  const Descriptor* containing_type() const;
-
-  // If the field is a member of a oneof, this is the one, otherwise this is
-  // nullptr.
-  const OneofDescriptor* containing_oneof() const;
-
-  // If the field is a member of a non-synthetic oneof, returns the descriptor
-  // for the oneof, otherwise returns nullptr.
-  const OneofDescriptor* real_containing_oneof() const;
-
-  // If the field is a member of a oneof, returns the index in that oneof.
-  int index_in_oneof() const;
-
-  // An extension may be declared within the scope of another message.  If this
-  // field is an extension (is_extension() is true), then extension_scope()
-  // returns that message, or nullptr if the extension was declared at global
-  // scope.  If this is not an extension, extension_scope() is undefined (may
-  // assert-fail).
-  const Descriptor* extension_scope() const;
-
-  // If type is TYPE_MESSAGE or TYPE_GROUP, returns a descriptor for the
-  // message or the group type.  Otherwise, returns null.
-  const Descriptor* message_type() const;
-  // If type is TYPE_ENUM, returns a descriptor for the enum.  Otherwise,
-  // returns null.
-  const EnumDescriptor* enum_type() const;
-
-  // Get the FieldOptions for this field.  This includes things listed in
-  // square brackets after the field definition.  E.g., the field:
-  //   optional string text = 1 [ctype=CORD];
-  // has the "ctype" option set.  Allowed options are defined by FieldOptions in
-  // descriptor.proto, and any available extensions of that message.
-  const FieldOptions& options() const;
-
-  // See Descriptor::CopyTo().
-  void CopyTo(FieldDescriptorProto* proto) const;
-
-  // See Descriptor::DebugString().
-  std::string DebugString() const;
-
-  // See Descriptor::DebugStringWithOptions().
-  std::string DebugStringWithOptions(const DebugStringOptions& options) const;
-
-  // Helper method to get the CppType for a particular Type.
-  static CppType TypeToCppType(Type type);
-
-  // Helper method to get the name of a Type.
-  static const char* TypeName(Type type);
-
-  // Helper method to get the name of a CppType.
-  static const char* CppTypeName(CppType cpp_type);
-
-  // Return true iff [packed = true] is valid for fields of this type.
-  static inline bool IsTypePackable(Type field_type);
-
-  // Returns full_name() except if the field is a MessageSet extension,
-  // in which case it returns the full_name() of the containing message type
-  // for backwards compatibility with proto1.
-  //
-  // A MessageSet extension is defined as an optional message extension
-  // whose containing type has the message_set_wire_format option set.
-  // This should be true of extensions of google.protobuf.bridge.MessageSet;
-  // by convention, such extensions are named "message_set_extension".
-  //
-  // The opposite operation (looking up an extension's FieldDescriptor given
-  // its printable name) can be accomplished with
-  //     message->file()->pool()->FindExtensionByPrintableName(message, name)
-  // where the extension extends "message".
-  const std::string& PrintableNameForExtension() const;
-
-  // Source Location ---------------------------------------------------
-
-  // Updates |*out_location| to the source location of the complete
-  // extent of this field declaration.  Returns false and leaves
-  // |*out_location| unchanged iff location information was not available.
-  bool GetSourceLocation(SourceLocation* out_location) const;
-
- private:
-  typedef FieldOptions OptionsType;
-
-  // Allows access to GetLocationPath for annotations.
-  friend class io::Printer;
-  friend class compiler::cpp::Formatter;
-  friend class Reflection;
-
-  // Fill the json_name field of FieldDescriptorProto.
-  void CopyJsonNameTo(FieldDescriptorProto* proto) const;
-
-  // See Descriptor::DebugString().
-  void DebugString(int depth, std::string* contents,
-                   const DebugStringOptions& options) const;
-
-  // formats the default value appropriately and returns it as a string.
-  // Must have a default value to call this. If quote_string_type is true, then
-  // types of CPPTYPE_STRING whill be surrounded by quotes and CEscaped.
-  std::string DefaultValueAsString(bool quote_string_type) const;
-
-  // Helper function that returns the field type name for DebugString.
-  std::string FieldTypeNameDebugString() const;
-
-  // Walks up the descriptor tree to generate the source location path
-  // to this descriptor from the file root.
-  void GetLocationPath(std::vector<int>* output) const;
-
-  // Returns true if this is a map message type.
-  bool is_map_message_type() const;
-
-  const std::string* name_;
-  const std::string* full_name_;
-  const std::string* lowercase_name_;
-  const std::string* camelcase_name_;
-  // If has_json_name_ is true, it's the value specified by the user.
-  // Otherwise, it has the same value as camelcase_name_.
-  const std::string* json_name_;
-  const FileDescriptor* file_;
-  internal::once_flag* type_once_;
-  static void TypeOnceInit(const FieldDescriptor* to_init);
-  void InternalTypeOnceInit() const;
-  mutable Type type_;
-  Label label_;
-  bool has_default_value_;
-  bool proto3_optional_;
-  // Whether the user has specified the json_name field option in the .proto
-  // file.
-  bool has_json_name_;
-  bool is_extension_;
-  int number_;
-  int index_in_oneof_;
-  const Descriptor* containing_type_;
-  const OneofDescriptor* containing_oneof_;
-  const Descriptor* extension_scope_;
-  mutable const Descriptor* message_type_;
-  mutable const EnumDescriptor* enum_type_;
-  const FieldOptions* options_;
-  const std::string* type_name_;
-  const std::string* default_value_enum_name_;
-  // IMPORTANT:  If you add a new field, make sure to search for all instances
-  // of Allocate<FieldDescriptor>() and AllocateArray<FieldDescriptor>() in
-  // descriptor.cc and update them to initialize the field.
-
-  union {
-    int32 default_value_int32_;
-    int64 default_value_int64_;
-    uint32 default_value_uint32_;
-    uint64 default_value_uint64_;
-    float default_value_float_;
-    double default_value_double_;
-    bool default_value_bool_;
-
-    mutable const EnumValueDescriptor* default_value_enum_;
-    const std::string* default_value_string_;
-    mutable std::atomic<const Message*> default_generated_instance_;
-  };
-
-  static const CppType kTypeToCppTypeMap[MAX_TYPE + 1];
-
-  static const char* const kTypeToName[MAX_TYPE + 1];
-
-  static const char* const kCppTypeToName[MAX_CPPTYPE + 1];
-
-  static const char* const kLabelToName[MAX_LABEL + 1];
-
-  // Must be constructed using DescriptorPool.
-  FieldDescriptor() {}
-  friend class DescriptorBuilder;
-  friend class FileDescriptor;
-  friend class Descriptor;
-  friend class OneofDescriptor;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(FieldDescriptor);
-};
-
-
-// Describes a oneof defined in a message type.
-class PROTOBUF_EXPORT OneofDescriptor {
- public:
-  typedef OneofDescriptorProto Proto;
-
-  const std::string& name() const;       // Name of this oneof.
-  const std::string& full_name() const;  // Fully-qualified name of the oneof.
-
-  // Index of this oneof within the message's oneof array.
-  int index() const;
-
-  // Returns whether this oneof was inserted by the compiler to wrap a proto3
-  // optional field. If this returns true, code generators should *not* emit it.
-  bool is_synthetic() const;
-
-  // The .proto file in which this oneof was defined.  Never nullptr.
-  const FileDescriptor* file() const;
-  // The Descriptor for the message containing this oneof.
-  const Descriptor* containing_type() const;
-
-  // The number of (non-extension) fields which are members of this oneof.
-  int field_count() const;
-  // Get a member of this oneof, in the order in which they were declared in the
-  // .proto file.  Does not include extensions.
-  const FieldDescriptor* field(int index) const;
-
-  const OneofOptions& options() const;
-
-  // See Descriptor::CopyTo().
-  void CopyTo(OneofDescriptorProto* proto) const;
-
-  // See Descriptor::DebugString().
-  std::string DebugString() const;
-
-  // See Descriptor::DebugStringWithOptions().
-  std::string DebugStringWithOptions(const DebugStringOptions& options) const;
-
-  // Source Location ---------------------------------------------------
-
-  // Updates |*out_location| to the source location of the complete
-  // extent of this oneof declaration.  Returns false and leaves
-  // |*out_location| unchanged iff location information was not available.
-  bool GetSourceLocation(SourceLocation* out_location) const;
-
- private:
-  typedef OneofOptions OptionsType;
-
-  // Allows access to GetLocationPath for annotations.
-  friend class io::Printer;
-  friend class compiler::cpp::Formatter;
-
-  // See Descriptor::DebugString().
-  void DebugString(int depth, std::string* contents,
-                   const DebugStringOptions& options) const;
-
-  // Walks up the descriptor tree to generate the source location path
-  // to this descriptor from the file root.
-  void GetLocationPath(std::vector<int>* output) const;
-
-  const std::string* name_;
-  const std::string* full_name_;
-  const Descriptor* containing_type_;
-  int field_count_;
-  const FieldDescriptor** fields_;
-  const OneofOptions* options_;
-
-  // IMPORTANT:  If you add a new field, make sure to search for all instances
-  // of Allocate<OneofDescriptor>() and AllocateArray<OneofDescriptor>()
-  // in descriptor.cc and update them to initialize the field.
-
-  // Must be constructed using DescriptorPool.
-  OneofDescriptor() {}
-  friend class DescriptorBuilder;
-  friend class Descriptor;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(OneofDescriptor);
-};
-
-// Describes an enum type defined in a .proto file.  To get the EnumDescriptor
-// for a generated enum type, call TypeName_descriptor().  Use DescriptorPool
-// to construct your own descriptors.
-class PROTOBUF_EXPORT EnumDescriptor {
- public:
-  typedef EnumDescriptorProto Proto;
-
-  // The name of this enum type in the containing scope.
-  const std::string& name() const;
-
-  // The fully-qualified name of the enum type, scope delimited by periods.
-  const std::string& full_name() const;
-
-  // Index of this enum within the file or containing message's enum array.
-  int index() const;
-
-  // The .proto file in which this enum type was defined.  Never nullptr.
-  const FileDescriptor* file() const;
-
-  // The number of values for this EnumDescriptor.  Guaranteed to be greater
-  // than zero.
-  int value_count() const;
-  // Gets a value by index, where 0 <= index < value_count().
-  // These are returned in the order they were defined in the .proto file.
-  const EnumValueDescriptor* value(int index) const;
-
-  // Looks up a value by name.  Returns nullptr if no such value exists.
-  const EnumValueDescriptor* FindValueByName(ConstStringParam name) const;
-  // Looks up a value by number.  Returns nullptr if no such value exists.  If
-  // multiple values have this number, the first one defined is returned.
-  const EnumValueDescriptor* FindValueByNumber(int number) const;
-
-  // If this enum type is nested in a message type, this is that message type.
-  // Otherwise, nullptr.
-  const Descriptor* containing_type() const;
-
-  // Get options for this enum type.  These are specified in the .proto file by
-  // placing lines like "option foo = 1234;" in the enum definition.  Allowed
-  // options are defined by EnumOptions in descriptor.proto, and any available
-  // extensions of that message.
-  const EnumOptions& options() const;
-
-  // See Descriptor::CopyTo().
-  void CopyTo(EnumDescriptorProto* proto) const;
-
-  // See Descriptor::DebugString().
-  std::string DebugString() const;
-
-  // See Descriptor::DebugStringWithOptions().
-  std::string DebugStringWithOptions(const DebugStringOptions& options) const;
-
-  // Returns true if this is a placeholder for an unknown enum. This will
-  // only be the case if this descriptor comes from a DescriptorPool
-  // with AllowUnknownDependencies() set.
-  bool is_placeholder() const;
-
-  // Reserved fields -------------------------------------------------
-
-  // A range of reserved field numbers.
-  struct ReservedRange {
-    int start;  // inclusive
-    int end;    // inclusive
-  };
-
-  // The number of reserved ranges in this message type.
-  int reserved_range_count() const;
-  // Gets an reserved range by index, where 0 <= index <
-  // reserved_range_count(). These are returned in the order they were defined
-  // in the .proto file.
-  const EnumDescriptor::ReservedRange* reserved_range(int index) const;
-
-  // Returns true if the number is in one of the reserved ranges.
-  bool IsReservedNumber(int number) const;
-
-  // Returns nullptr if no reserved range contains the given number.
-  const EnumDescriptor::ReservedRange* FindReservedRangeContainingNumber(
-      int number) const;
-
-  // The number of reserved field names in this message type.
-  int reserved_name_count() const;
-
-  // Gets a reserved name by index, where 0 <= index < reserved_name_count().
-  const std::string& reserved_name(int index) const;
-
-  // Returns true if the field name is reserved.
-  bool IsReservedName(ConstStringParam name) const;
-
-  // Source Location ---------------------------------------------------
-
-  // Updates |*out_location| to the source location of the complete
-  // extent of this enum declaration.  Returns false and leaves
-  // |*out_location| unchanged iff location information was not available.
-  bool GetSourceLocation(SourceLocation* out_location) const;
-
- private:
-  typedef EnumOptions OptionsType;
-
-  // Allows access to GetLocationPath for annotations.
-  friend class io::Printer;
-  friend class compiler::cpp::Formatter;
-
-  // Looks up a value by number.  If the value does not exist, dynamically
-  // creates a new EnumValueDescriptor for that value, assuming that it was
-  // unknown. If a new descriptor is created, this is done in a thread-safe way,
-  // and future calls will return the same value descriptor pointer.
-  //
-  // This is private but is used by Reflection (which is friended below) to
-  // return a valid EnumValueDescriptor from GetEnum() when this feature is
-  // enabled.
-  const EnumValueDescriptor* FindValueByNumberCreatingIfUnknown(
-      int number) const;
-
-  // See Descriptor::DebugString().
-  void DebugString(int depth, std::string* contents,
-                   const DebugStringOptions& options) const;
-
-  // Walks up the descriptor tree to generate the source location path
-  // to this descriptor from the file root.
-  void GetLocationPath(std::vector<int>* output) const;
-
-  const std::string* name_;
-  const std::string* full_name_;
-  const FileDescriptor* file_;
-  const Descriptor* containing_type_;
-  const EnumOptions* options_;
-
-  // True if this is a placeholder for an unknown type.
-  bool is_placeholder_;
-  // True if this is a placeholder and the type name wasn't fully-qualified.
-  bool is_unqualified_placeholder_;
-
-  int value_count_;
-  EnumValueDescriptor* values_;
-
-  int reserved_range_count_;
-  int reserved_name_count_;
-  EnumDescriptor::ReservedRange* reserved_ranges_;
-  const std::string** reserved_names_;
-
-  // IMPORTANT:  If you add a new field, make sure to search for all instances
-  // of Allocate<EnumDescriptor>() and AllocateArray<EnumDescriptor>() in
-  // descriptor.cc and update them to initialize the field.
-
-  // Must be constructed using DescriptorPool.
-  EnumDescriptor() {}
-  friend class DescriptorBuilder;
-  friend class Descriptor;
-  friend class FieldDescriptor;
-  friend class EnumValueDescriptor;
-  friend class FileDescriptor;
-  friend class DescriptorPool;
-  friend class Reflection;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(EnumDescriptor);
-};
-
-// Describes an individual enum constant of a particular type.  To get the
-// EnumValueDescriptor for a given enum value, first get the EnumDescriptor
-// for its type, then use EnumDescriptor::FindValueByName() or
-// EnumDescriptor::FindValueByNumber().  Use DescriptorPool to construct
-// your own descriptors.
-class PROTOBUF_EXPORT EnumValueDescriptor {
- public:
-  typedef EnumValueDescriptorProto Proto;
-
-  const std::string& name() const;  // Name of this enum constant.
-  int index() const;                // Index within the enums's Descriptor.
-  int number() const;               // Numeric value of this enum constant.
-
-  // The full_name of an enum value is a sibling symbol of the enum type.
-  // e.g. the full name of FieldDescriptorProto::TYPE_INT32 is actually
-  // "google.protobuf.FieldDescriptorProto.TYPE_INT32", NOT
-  // "google.protobuf.FieldDescriptorProto.Type.TYPE_INT32".  This is to conform
-  // with C++ scoping rules for enums.
-  const std::string& full_name() const;
-
-  // The .proto file in which this value was defined.  Never nullptr.
-  const FileDescriptor* file() const;
-  // The type of this value.  Never nullptr.
-  const EnumDescriptor* type() const;
-
-  // Get options for this enum value.  These are specified in the .proto file by
-  // adding text like "[foo = 1234]" after an enum value definition.  Allowed
-  // options are defined by EnumValueOptions in descriptor.proto, and any
-  // available extensions of that message.
-  const EnumValueOptions& options() const;
-
-  // See Descriptor::CopyTo().
-  void CopyTo(EnumValueDescriptorProto* proto) const;
-
-  // See Descriptor::DebugString().
-  std::string DebugString() const;
-
-  // See Descriptor::DebugStringWithOptions().
-  std::string DebugStringWithOptions(const DebugStringOptions& options) const;
-
-  // Source Location ---------------------------------------------------
-
-  // Updates |*out_location| to the source location of the complete
-  // extent of this enum value declaration.  Returns false and leaves
-  // |*out_location| unchanged iff location information was not available.
-  bool GetSourceLocation(SourceLocation* out_location) const;
-
- private:
-  typedef EnumValueOptions OptionsType;
-
-  // Allows access to GetLocationPath for annotations.
-  friend class io::Printer;
-  friend class compiler::cpp::Formatter;
-
-  // See Descriptor::DebugString().
-  void DebugString(int depth, std::string* contents,
-                   const DebugStringOptions& options) const;
-
-  // Walks up the descriptor tree to generate the source location path
-  // to this descriptor from the file root.
-  void GetLocationPath(std::vector<int>* output) const;
-
-  const std::string* name_;
-  const std::string* full_name_;
-  int number_;
-  const EnumDescriptor* type_;
-  const EnumValueOptions* options_;
-  // IMPORTANT:  If you add a new field, make sure to search for all instances
-  // of Allocate<EnumValueDescriptor>() and AllocateArray<EnumValueDescriptor>()
-  // in descriptor.cc and update them to initialize the field.
-
-  // Must be constructed using DescriptorPool.
-  EnumValueDescriptor() {}
-  friend class DescriptorBuilder;
-  friend class EnumDescriptor;
-  friend class DescriptorPool;
-  friend class FileDescriptorTables;
-  friend class Reflection;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(EnumValueDescriptor);
-};
-
-// Describes an RPC service. Use DescriptorPool to construct your own
-// descriptors.
-class PROTOBUF_EXPORT ServiceDescriptor {
- public:
-  typedef ServiceDescriptorProto Proto;
-
-  // The name of the service, not including its containing scope.
-  const std::string& name() const;
-  // The fully-qualified name of the service, scope delimited by periods.
-  const std::string& full_name() const;
-  // Index of this service within the file's services array.
-  int index() const;
-
-  // The .proto file in which this service was defined.  Never nullptr.
-  const FileDescriptor* file() const;
-
-  // Get options for this service type.  These are specified in the .proto file
-  // by placing lines like "option foo = 1234;" in the service definition.
-  // Allowed options are defined by ServiceOptions in descriptor.proto, and any
-  // available extensions of that message.
-  const ServiceOptions& options() const;
-
-  // The number of methods this service defines.
-  int method_count() const;
-  // Gets a MethodDescriptor by index, where 0 <= index < method_count().
-  // These are returned in the order they were defined in the .proto file.
-  const MethodDescriptor* method(int index) const;
-
-  // Look up a MethodDescriptor by name.
-  const MethodDescriptor* FindMethodByName(ConstStringParam name) const;
-  // See Descriptor::CopyTo().
-  void CopyTo(ServiceDescriptorProto* proto) const;
-
-  // See Descriptor::DebugString().
-  std::string DebugString() const;
-
-  // See Descriptor::DebugStringWithOptions().
-  std::string DebugStringWithOptions(const DebugStringOptions& options) const;
-
-  // Source Location ---------------------------------------------------
-
-  // Updates |*out_location| to the source location of the complete
-  // extent of this service declaration.  Returns false and leaves
-  // |*out_location| unchanged iff location information was not available.
-  bool GetSourceLocation(SourceLocation* out_location) const;
-
- private:
-  typedef ServiceOptions OptionsType;
-
-  // Allows access to GetLocationPath for annotations.
-  friend class io::Printer;
-  friend class compiler::cpp::Formatter;
-
-  // See Descriptor::DebugString().
-  void DebugString(std::string* contents,
-                   const DebugStringOptions& options) const;
-
-  // Walks up the descriptor tree to generate the source location path
-  // to this descriptor from the file root.
-  void GetLocationPath(std::vector<int>* output) const;
-
-  const std::string* name_;
-  const std::string* full_name_;
-  const FileDescriptor* file_;
-  const ServiceOptions* options_;
-  MethodDescriptor* methods_;
-  int method_count_;
-  // IMPORTANT:  If you add a new field, make sure to search for all instances
-  // of Allocate<ServiceDescriptor>() and AllocateArray<ServiceDescriptor>() in
-  // descriptor.cc and update them to initialize the field.
-
-  // Must be constructed using DescriptorPool.
-  ServiceDescriptor() {}
-  friend class DescriptorBuilder;
-  friend class FileDescriptor;
-  friend class MethodDescriptor;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ServiceDescriptor);
-};
-
-
-// Describes an individual service method.  To obtain a MethodDescriptor given
-// a service, first get its ServiceDescriptor, then call
-// ServiceDescriptor::FindMethodByName().  Use DescriptorPool to construct your
-// own descriptors.
-class PROTOBUF_EXPORT MethodDescriptor {
- public:
-  typedef MethodDescriptorProto Proto;
-
-  // Name of this method, not including containing scope.
-  const std::string& name() const;
-  // The fully-qualified name of the method, scope delimited by periods.
-  const std::string& full_name() const;
-  // Index within the service's Descriptor.
-  int index() const;
-
-  // The .proto file in which this method was defined.  Never nullptr.
-  const FileDescriptor* file() const;
-  // Gets the service to which this method belongs.  Never nullptr.
-  const ServiceDescriptor* service() const;
-
-  // Gets the type of protocol message which this method accepts as input.
-  const Descriptor* input_type() const;
-  // Gets the type of protocol message which this message produces as output.
-  const Descriptor* output_type() const;
-
-  // Gets whether the client streams multiple requests.
-  bool client_streaming() const;
-  // Gets whether the server streams multiple responses.
-  bool server_streaming() const;
-
-  // Get options for this method.  These are specified in the .proto file by
-  // placing lines like "option foo = 1234;" in curly-braces after a method
-  // declaration.  Allowed options are defined by MethodOptions in
-  // descriptor.proto, and any available extensions of that message.
-  const MethodOptions& options() const;
-
-  // See Descriptor::CopyTo().
-  void CopyTo(MethodDescriptorProto* proto) const;
-
-  // See Descriptor::DebugString().
-  std::string DebugString() const;
-
-  // See Descriptor::DebugStringWithOptions().
-  std::string DebugStringWithOptions(const DebugStringOptions& options) const;
-
-  // Source Location ---------------------------------------------------
-
-  // Updates |*out_location| to the source location of the complete
-  // extent of this method declaration.  Returns false and leaves
-  // |*out_location| unchanged iff location information was not available.
-  bool GetSourceLocation(SourceLocation* out_location) const;
-
- private:
-  typedef MethodOptions OptionsType;
-
-  // Allows access to GetLocationPath for annotations.
-  friend class io::Printer;
-  friend class compiler::cpp::Formatter;
-
-  // See Descriptor::DebugString().
-  void DebugString(int depth, std::string* contents,
-                   const DebugStringOptions& options) const;
-
-  // Walks up the descriptor tree to generate the source location path
-  // to this descriptor from the file root.
-  void GetLocationPath(std::vector<int>* output) const;
-
-  const std::string* name_;
-  const std::string* full_name_;
-  const ServiceDescriptor* service_;
-  mutable internal::LazyDescriptor input_type_;
-  mutable internal::LazyDescriptor output_type_;
-  const MethodOptions* options_;
-  bool client_streaming_;
-  bool server_streaming_;
-  // IMPORTANT:  If you add a new field, make sure to search for all instances
-  // of Allocate<MethodDescriptor>() and AllocateArray<MethodDescriptor>() in
-  // descriptor.cc and update them to initialize the field.
-
-  // Must be constructed using DescriptorPool.
-  MethodDescriptor() {}
-  friend class DescriptorBuilder;
-  friend class ServiceDescriptor;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(MethodDescriptor);
-};
-
-
-// Describes a whole .proto file.  To get the FileDescriptor for a compiled-in
-// file, get the descriptor for something defined in that file and call
-// descriptor->file().  Use DescriptorPool to construct your own descriptors.
-class PROTOBUF_EXPORT FileDescriptor {
- public:
-  typedef FileDescriptorProto Proto;
-
-  // The filename, relative to the source tree.
-  // e.g. "foo/bar/baz.proto"
-  const std::string& name() const;
-
-  // The package, e.g. "google.protobuf.compiler".
-  const std::string& package() const;
-
-  // The DescriptorPool in which this FileDescriptor and all its contents were
-  // allocated.  Never nullptr.
-  const DescriptorPool* pool() const;
-
-  // The number of files imported by this one.
-  int dependency_count() const;
-  // Gets an imported file by index, where 0 <= index < dependency_count().
-  // These are returned in the order they were defined in the .proto file.
-  const FileDescriptor* dependency(int index) const;
-
-  // The number of files public imported by this one.
-  // The public dependency list is a subset of the dependency list.
-  int public_dependency_count() const;
-  // Gets a public imported file by index, where 0 <= index <
-  // public_dependency_count().
-  // These are returned in the order they were defined in the .proto file.
-  const FileDescriptor* public_dependency(int index) const;
-
-  // The number of files that are imported for weak fields.
-  // The weak dependency list is a subset of the dependency list.
-  int weak_dependency_count() const;
-  // Gets a weak imported file by index, where 0 <= index <
-  // weak_dependency_count().
-  // These are returned in the order they were defined in the .proto file.
-  const FileDescriptor* weak_dependency(int index) const;
-
-  // Number of top-level message types defined in this file.  (This does not
-  // include nested types.)
-  int message_type_count() const;
-  // Gets a top-level message type, where 0 <= index < message_type_count().
-  // These are returned in the order they were defined in the .proto file.
-  const Descriptor* message_type(int index) const;
-
-  // Number of top-level enum types defined in this file.  (This does not
-  // include nested types.)
-  int enum_type_count() const;
-  // Gets a top-level enum type, where 0 <= index < enum_type_count().
-  // These are returned in the order they were defined in the .proto file.
-  const EnumDescriptor* enum_type(int index) const;
-
-  // Number of services defined in this file.
-  int service_count() const;
-  // Gets a service, where 0 <= index < service_count().
-  // These are returned in the order they were defined in the .proto file.
-  const ServiceDescriptor* service(int index) const;
-
-  // Number of extensions defined at file scope.  (This does not include
-  // extensions nested within message types.)
-  int extension_count() const;
-  // Gets an extension's descriptor, where 0 <= index < extension_count().
-  // These are returned in the order they were defined in the .proto file.
-  const FieldDescriptor* extension(int index) const;
-
-  // Get options for this file.  These are specified in the .proto file by
-  // placing lines like "option foo = 1234;" at the top level, outside of any
-  // other definitions.  Allowed options are defined by FileOptions in
-  // descriptor.proto, and any available extensions of that message.
-  const FileOptions& options() const;
-
-  // Syntax of this file.
-  enum Syntax {
-    SYNTAX_UNKNOWN = 0,
-    SYNTAX_PROTO2 = 2,
-    SYNTAX_PROTO3 = 3,
-  };
-  Syntax syntax() const;
-  static const char* SyntaxName(Syntax syntax);
-
-  // Find a top-level message type by name.  Returns nullptr if not found.
-  const Descriptor* FindMessageTypeByName(ConstStringParam name) const;
-  // Find a top-level enum type by name.  Returns nullptr if not found.
-  const EnumDescriptor* FindEnumTypeByName(ConstStringParam name) const;
-  // Find an enum value defined in any top-level enum by name.  Returns nullptr
-  // if not found.
-  const EnumValueDescriptor* FindEnumValueByName(ConstStringParam name) const;
-  // Find a service definition by name.  Returns nullptr if not found.
-  const ServiceDescriptor* FindServiceByName(ConstStringParam name) const;
-  // Find a top-level extension definition by name.  Returns nullptr if not
-  // found.
-  const FieldDescriptor* FindExtensionByName(ConstStringParam name) const;
-  // Similar to FindExtensionByName(), but searches by lowercased-name.  See
-  // Descriptor::FindFieldByLowercaseName().
-  const FieldDescriptor* FindExtensionByLowercaseName(
-      ConstStringParam name) const;
-  // Similar to FindExtensionByName(), but searches by camelcased-name.  See
-  // Descriptor::FindFieldByCamelcaseName().
-  const FieldDescriptor* FindExtensionByCamelcaseName(
-      ConstStringParam name) const;
-
-  // See Descriptor::CopyTo().
-  // Notes:
-  // - This method does NOT copy source code information since it is relatively
-  //   large and rarely needed.  See CopySourceCodeInfoTo() below.
-  void CopyTo(FileDescriptorProto* proto) const;
-  // Write the source code information of this FileDescriptor into the given
-  // FileDescriptorProto.  See CopyTo() above.
-  void CopySourceCodeInfoTo(FileDescriptorProto* proto) const;
-  // Fill the json_name field of FieldDescriptorProto for all fields. Can only
-  // be called after CopyTo().
-  void CopyJsonNameTo(FileDescriptorProto* proto) const;
-
-  // See Descriptor::DebugString().
-  std::string DebugString() const;
-
-  // See Descriptor::DebugStringWithOptions().
-  std::string DebugStringWithOptions(const DebugStringOptions& options) const;
-
-  // Returns true if this is a placeholder for an unknown file. This will
-  // only be the case if this descriptor comes from a DescriptorPool
-  // with AllowUnknownDependencies() set.
-  bool is_placeholder() const;
-
-  // Updates |*out_location| to the source location of the complete extent of
-  // this file declaration (namely, the empty path).
-  bool GetSourceLocation(SourceLocation* out_location) const;
-
-  // Updates |*out_location| to the source location of the complete
-  // extent of the declaration or declaration-part denoted by |path|.
-  // Returns false and leaves |*out_location| unchanged iff location
-  // information was not available.  (See SourceCodeInfo for
-  // description of path encoding.)
-  bool GetSourceLocation(const std::vector<int>& path,
-                         SourceLocation* out_location) const;
-
- private:
-  typedef FileOptions OptionsType;
-
-  const std::string* name_;
-  const std::string* package_;
-  const DescriptorPool* pool_;
-  internal::once_flag* dependencies_once_;
-  static void DependenciesOnceInit(const FileDescriptor* to_init);
-  void InternalDependenciesOnceInit() const;
-
-  // These are arranged to minimize padding on 64-bit.
-  int dependency_count_;
-  int public_dependency_count_;
-  int weak_dependency_count_;
-  int message_type_count_;
-  int enum_type_count_;
-  int service_count_;
-  int extension_count_;
-  Syntax syntax_;
-  bool is_placeholder_;
-
-  // Indicates the FileDescriptor is completed building. Used to verify
-  // that type accessor functions that can possibly build a dependent file
-  // aren't called during the process of building the file.
-  bool finished_building_;
-
-  mutable const FileDescriptor** dependencies_;
-  const std::string** dependencies_names_;
-  int* public_dependencies_;
-  int* weak_dependencies_;
-  Descriptor* message_types_;
-  EnumDescriptor* enum_types_;
-  ServiceDescriptor* services_;
-  FieldDescriptor* extensions_;
-  const FileOptions* options_;
-
-  const FileDescriptorTables* tables_;
-  const SourceCodeInfo* source_code_info_;
-
-  // IMPORTANT:  If you add a new field, make sure to search for all instances
-  // of Allocate<FileDescriptor>() and AllocateArray<FileDescriptor>() in
-  // descriptor.cc and update them to initialize the field.
-
-  FileDescriptor() {}
-  friend class DescriptorBuilder;
-  friend class DescriptorPool;
-  friend class Descriptor;
-  friend class FieldDescriptor;
-  friend class internal::LazyDescriptor;
-  friend class OneofDescriptor;
-  friend class EnumDescriptor;
-  friend class EnumValueDescriptor;
-  friend class MethodDescriptor;
-  friend class ServiceDescriptor;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(FileDescriptor);
-};
-
-
-// ===================================================================
-
-// Used to construct descriptors.
-//
-// Normally you won't want to build your own descriptors.  Message classes
-// constructed by the protocol compiler will provide them for you.  However,
-// if you are implementing Message on your own, or if you are writing a
-// program which can operate on totally arbitrary types and needs to load
-// them from some sort of database, you might need to.
-//
-// Since Descriptors are composed of a whole lot of cross-linked bits of
-// data that would be a pain to put together manually, the
-// DescriptorPool class is provided to make the process easier.  It can
-// take a FileDescriptorProto (defined in descriptor.proto), validate it,
-// and convert it to a set of nicely cross-linked Descriptors.
-//
-// DescriptorPool also helps with memory management.  Descriptors are
-// composed of many objects containing static data and pointers to each
-// other.  In all likelihood, when it comes time to delete this data,
-// you'll want to delete it all at once.  In fact, it is not uncommon to
-// have a whole pool of descriptors all cross-linked with each other which
-// you wish to delete all at once.  This class represents such a pool, and
-// handles the memory management for you.
-//
-// You can also search for descriptors within a DescriptorPool by name, and
-// extensions by number.
-class PROTOBUF_EXPORT DescriptorPool {
- public:
-  // Create a normal, empty DescriptorPool.
-  DescriptorPool();
-
-  // Constructs a DescriptorPool that, when it can't find something among the
-  // descriptors already in the pool, looks for it in the given
-  // DescriptorDatabase.
-  // Notes:
-  // - If a DescriptorPool is constructed this way, its BuildFile*() methods
-  //   must not be called (they will assert-fail).  The only way to populate
-  //   the pool with descriptors is to call the Find*By*() methods.
-  // - The Find*By*() methods may block the calling thread if the
-  //   DescriptorDatabase blocks.  This in turn means that parsing messages
-  //   may block if they need to look up extensions.
-  // - The Find*By*() methods will use mutexes for thread-safety, thus making
-  //   them slower even when they don't have to fall back to the database.
-  //   In fact, even the Find*By*() methods of descriptor objects owned by
-  //   this pool will be slower, since they will have to obtain locks too.
-  // - An ErrorCollector may optionally be given to collect validation errors
-  //   in files loaded from the database.  If not given, errors will be printed
-  //   to GOOGLE_LOG(ERROR).  Remember that files are built on-demand, so this
-  //   ErrorCollector may be called from any thread that calls one of the
-  //   Find*By*() methods.
-  // - The DescriptorDatabase must not be mutated during the lifetime of
-  //   the DescriptorPool. Even if the client takes care to avoid data races,
-  //   changes to the content of the DescriptorDatabase may not be reflected
-  //   in subsequent lookups in the DescriptorPool.
-  class ErrorCollector;
-  explicit DescriptorPool(DescriptorDatabase* fallback_database,
-                          ErrorCollector* error_collector = nullptr);
-
-  ~DescriptorPool();
-
-  // Get a pointer to the generated pool.  Generated protocol message classes
-  // which are compiled into the binary will allocate their descriptors in
-  // this pool.  Do not add your own descriptors to this pool.
-  static const DescriptorPool* generated_pool();
-
-
-  // Find a FileDescriptor in the pool by file name.  Returns nullptr if not
-  // found.
-  const FileDescriptor* FindFileByName(ConstStringParam name) const;
-
-  // Find the FileDescriptor in the pool which defines the given symbol.
-  // If any of the Find*ByName() methods below would succeed, then this is
-  // equivalent to calling that method and calling the result's file() method.
-  // Otherwise this returns nullptr.
-  const FileDescriptor* FindFileContainingSymbol(
-      ConstStringParam symbol_name) const;
-
-  // Looking up descriptors ------------------------------------------
-  // These find descriptors by fully-qualified name.  These will find both
-  // top-level descriptors and nested descriptors.  They return nullptr if not
-  // found.
-
-  const Descriptor* FindMessageTypeByName(ConstStringParam name) const;
-  const FieldDescriptor* FindFieldByName(ConstStringParam name) const;
-  const FieldDescriptor* FindExtensionByName(ConstStringParam name) const;
-  const OneofDescriptor* FindOneofByName(ConstStringParam name) const;
-  const EnumDescriptor* FindEnumTypeByName(ConstStringParam name) const;
-  const EnumValueDescriptor* FindEnumValueByName(ConstStringParam name) const;
-  const ServiceDescriptor* FindServiceByName(ConstStringParam name) const;
-  const MethodDescriptor* FindMethodByName(ConstStringParam name) const;
-
-  // Finds an extension of the given type by number.  The extendee must be
-  // a member of this DescriptorPool or one of its underlays.
-  const FieldDescriptor* FindExtensionByNumber(const Descriptor* extendee,
-                                               int number) const;
-
-  // Finds an extension of the given type by its printable name.
-  // See comments above PrintableNameForExtension() for the definition of
-  // "printable name".  The extendee must be a member of this DescriptorPool
-  // or one of its underlays.  Returns nullptr if there is no known message
-  // extension with the given printable name.
-  const FieldDescriptor* FindExtensionByPrintableName(
-      const Descriptor* extendee, ConstStringParam printable_name) const;
-
-  // Finds extensions of extendee. The extensions will be appended to
-  // out in an undefined order. Only extensions defined directly in
-  // this DescriptorPool or one of its underlays are guaranteed to be
-  // found: extensions defined in the fallback database might not be found
-  // depending on the database implementation.
-  void FindAllExtensions(const Descriptor* extendee,
-                         std::vector<const FieldDescriptor*>* out) const;
-
-  // Building descriptors --------------------------------------------
-
-  // When converting a FileDescriptorProto to a FileDescriptor, various
-  // errors might be detected in the input.  The caller may handle these
-  // programmatically by implementing an ErrorCollector.
-  class PROTOBUF_EXPORT ErrorCollector {
-   public:
-    inline ErrorCollector() {}
-    virtual ~ErrorCollector();
-
-    // These constants specify what exact part of the construct is broken.
-    // This is useful e.g. for mapping the error back to an exact location
-    // in a .proto file.
-    enum ErrorLocation {
-      NAME,           // the symbol name, or the package name for files
-      NUMBER,         // field or extension range number
-      TYPE,           // field type
-      EXTENDEE,       // field extendee
-      DEFAULT_VALUE,  // field default value
-      INPUT_TYPE,     // method input type
-      OUTPUT_TYPE,    // method output type
-      OPTION_NAME,    // name in assignment
-      OPTION_VALUE,   // value in option assignment
-      IMPORT,         // import error
-      OTHER           // some other problem
-    };
-
-    // Reports an error in the FileDescriptorProto. Use this function if the
-    // problem occurred should interrupt building the FileDescriptorProto.
-    virtual void AddError(
-        const std::string& filename,  // File name in which the error occurred.
-        const std::string& element_name,  // Full name of the erroneous element.
-        const Message* descriptor,  // Descriptor of the erroneous element.
-        ErrorLocation location,     // One of the location constants, above.
-        const std::string& message  // Human-readable error message.
-        ) = 0;
-
-    // Reports a warning in the FileDescriptorProto. Use this function if the
-    // problem occurred should NOT interrupt building the FileDescriptorProto.
-    virtual void AddWarning(
-        const std::string& /*filename*/,      // File name in which the error
-                                              // occurred.
-        const std::string& /*element_name*/,  // Full name of the erroneous
-                                              // element.
-        const Message* /*descriptor*/,  // Descriptor of the erroneous element.
-        ErrorLocation /*location*/,     // One of the location constants, above.
-        const std::string& /*message*/  // Human-readable error message.
-    ) {}
-
-   private:
-    GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ErrorCollector);
-  };
-
-  // Convert the FileDescriptorProto to real descriptors and place them in
-  // this DescriptorPool.  All dependencies of the file must already be in
-  // the pool.  Returns the resulting FileDescriptor, or nullptr if there were
-  // problems with the input (e.g. the message was invalid, or dependencies
-  // were missing).  Details about the errors are written to GOOGLE_LOG(ERROR).
-  const FileDescriptor* BuildFile(const FileDescriptorProto& proto);
-
-  // Same as BuildFile() except errors are sent to the given ErrorCollector.
-  const FileDescriptor* BuildFileCollectingErrors(
-      const FileDescriptorProto& proto, ErrorCollector* error_collector);
-
-  // By default, it is an error if a FileDescriptorProto contains references
-  // to types or other files that are not found in the DescriptorPool (or its
-  // backing DescriptorDatabase, if any).  If you call
-  // AllowUnknownDependencies(), however, then unknown types and files
-  // will be replaced by placeholder descriptors (which can be identified by
-  // the is_placeholder() method).  This can allow you to
-  // perform some useful operations with a .proto file even if you do not
-  // have access to other .proto files on which it depends.  However, some
-  // heuristics must be used to fill in the gaps in information, and these
-  // can lead to descriptors which are inaccurate.  For example, the
-  // DescriptorPool may be forced to guess whether an unknown type is a message
-  // or an enum, as well as what package it resides in.  Furthermore,
-  // placeholder types will not be discoverable via FindMessageTypeByName()
-  // and similar methods, which could confuse some descriptor-based algorithms.
-  // Generally, the results of this option should be handled with extreme care.
-  void AllowUnknownDependencies() { allow_unknown_ = true; }
-
-  // By default, weak imports are allowed to be missing, in which case we will
-  // use a placeholder for the dependency and convert the field to be an Empty
-  // message field. If you call EnforceWeakDependencies(true), however, the
-  // DescriptorPool will report a import not found error.
-  void EnforceWeakDependencies(bool enforce) { enforce_weak_ = enforce; }
-
-  // Internal stuff --------------------------------------------------
-  // These methods MUST NOT be called from outside the proto2 library.
-  // These methods may contain hidden pitfalls and may be removed in a
-  // future library version.
-
-  // Create a DescriptorPool which is overlaid on top of some other pool.
-  // If you search for a descriptor in the overlay and it is not found, the
-  // underlay will be searched as a backup.  If the underlay has its own
-  // underlay, that will be searched next, and so on.  This also means that
-  // files built in the overlay will be cross-linked with the underlay's
-  // descriptors if necessary.  The underlay remains property of the caller;
-  // it must remain valid for the lifetime of the newly-constructed pool.
-  //
-  // Example:  Say you want to parse a .proto file at runtime in order to use
-  // its type with a DynamicMessage.  Say this .proto file has dependencies,
-  // but you know that all the dependencies will be things that are already
-  // compiled into the binary.  For ease of use, you'd like to load the types
-  // right out of generated_pool() rather than have to parse redundant copies
-  // of all these .protos and runtime.  But, you don't want to add the parsed
-  // types directly into generated_pool(): this is not allowed, and would be
-  // bad design anyway.  So, instead, you could use generated_pool() as an
-  // underlay for a new DescriptorPool in which you add only the new file.
-  //
-  // WARNING:  Use of underlays can lead to many subtle gotchas.  Instead,
-  //   try to formulate what you want to do in terms of DescriptorDatabases.
-  explicit DescriptorPool(const DescriptorPool* underlay);
-
-  // Called by generated classes at init time to add their descriptors to
-  // generated_pool.  Do NOT call this in your own code!  filename must be a
-  // permanent string (e.g. a string literal).
-  static void InternalAddGeneratedFile(const void* encoded_file_descriptor,
-                                       int size);
-
-  // Disallow [enforce_utf8 = false] in .proto files.
-  void DisallowEnforceUtf8() { disallow_enforce_utf8_ = true; }
-
-
-  // For internal use only:  Gets a non-const pointer to the generated pool.
-  // This is called at static-initialization time only, so thread-safety is
-  // not a concern.  If both an underlay and a fallback database are present,
-  // the underlay takes precedence.
-  static DescriptorPool* internal_generated_pool();
-
-  // For internal use only:  Gets a non-const pointer to the generated
-  // descriptor database.
-  // Only used for testing.
-  static DescriptorDatabase* internal_generated_database();
-
-  // For internal use only:  Changes the behavior of BuildFile() such that it
-  // allows the file to make reference to message types declared in other files
-  // which it did not officially declare as dependencies.
-  void InternalDontEnforceDependencies();
-
-  // For internal use only: Enables lazy building of dependencies of a file.
-  // Delay the building of dependencies of a file descriptor until absolutely
-  // necessary, like when message_type() is called on a field that is defined
-  // in that dependency's file. This will cause functional issues if a proto
-  // or one of its dependencies has errors. Should only be enabled for the
-  // generated_pool_ (because no descriptor build errors are guaranteed by
-  // the compilation generation process), testing, or if a lack of descriptor
-  // build errors can be guaranteed for a pool.
-  void InternalSetLazilyBuildDependencies() {
-    lazily_build_dependencies_ = true;
-    // This needs to be set when lazily building dependencies, as it breaks
-    // dependency checking.
-    InternalDontEnforceDependencies();
-  }
-
-  // For internal use only.
-  void internal_set_underlay(const DescriptorPool* underlay) {
-    underlay_ = underlay;
-  }
-
-  // For internal (unit test) use only:  Returns true if a FileDescriptor has
-  // been constructed for the given file, false otherwise.  Useful for testing
-  // lazy descriptor initialization behavior.
-  bool InternalIsFileLoaded(ConstStringParam filename) const;
-
-  // Add a file to unused_import_track_files_. DescriptorBuilder will log
-  // warnings or errors for those files if there is any unused import.
-  void AddUnusedImportTrackFile(ConstStringParam file_name,
-                                bool is_error = false);
-  void ClearUnusedImportTrackFiles();
-
- private:
-  friend class Descriptor;
-  friend class internal::LazyDescriptor;
-  friend class FieldDescriptor;
-  friend class EnumDescriptor;
-  friend class ServiceDescriptor;
-  friend class MethodDescriptor;
-  friend class FileDescriptor;
-  friend class StreamDescriptor;
-  friend class DescriptorBuilder;
-  friend class FileDescriptorTables;
-
-  // Return true if the given name is a sub-symbol of any non-package
-  // descriptor that already exists in the descriptor pool.  (The full
-  // definition of such types is already known.)
-  bool IsSubSymbolOfBuiltType(StringPiece name) const;
-
-  // Tries to find something in the fallback database and link in the
-  // corresponding proto file.  Returns true if successful, in which case
-  // the caller should search for the thing again.  These are declared
-  // const because they are called by (semantically) const methods.
-  bool TryFindFileInFallbackDatabase(StringPiece name) const;
-  bool TryFindSymbolInFallbackDatabase(StringPiece name) const;
-  bool TryFindExtensionInFallbackDatabase(const Descriptor* containing_type,
-                                          int field_number) const;
-
-  // This internal find extension method only check with its table and underlay
-  // descriptor_pool's table. It does not check with fallback DB and no
-  // additional proto file will be build in this method.
-  const FieldDescriptor* InternalFindExtensionByNumberNoLock(
-      const Descriptor* extendee, int number) const;
-
-  // Like BuildFile() but called internally when the file has been loaded from
-  // fallback_database_.  Declared const because it is called by (semantically)
-  // const methods.
-  const FileDescriptor* BuildFileFromDatabase(
-      const FileDescriptorProto& proto) const;
-
-  // Helper for when lazily_build_dependencies_ is set, can look up a symbol
-  // after the file's descriptor is built, and can build the file where that
-  // symbol is defined if necessary. Will create a placeholder if the type
-  // doesn't exist in the fallback database, or the file doesn't build
-  // successfully.
-  Symbol CrossLinkOnDemandHelper(StringPiece name,
-                                 bool expecting_enum) const;
-
-  // Create a placeholder FileDescriptor of the specified name
-  FileDescriptor* NewPlaceholderFile(StringPiece name) const;
-  FileDescriptor* NewPlaceholderFileWithMutexHeld(StringPiece name) const;
-
-  enum PlaceholderType {
-    PLACEHOLDER_MESSAGE,
-    PLACEHOLDER_ENUM,
-    PLACEHOLDER_EXTENDABLE_MESSAGE
-  };
-  // Create a placeholder Descriptor of the specified name
-  Symbol NewPlaceholder(StringPiece name,
-                        PlaceholderType placeholder_type) const;
-  Symbol NewPlaceholderWithMutexHeld(StringPiece name,
-                                     PlaceholderType placeholder_type) const;
-
-  // If fallback_database_ is nullptr, this is nullptr.  Otherwise, this is a
-  // mutex which must be locked while accessing tables_.
-  internal::WrappedMutex* mutex_;
-
-  // See constructor.
-  DescriptorDatabase* fallback_database_;
-  ErrorCollector* default_error_collector_;
-  const DescriptorPool* underlay_;
-
-  // This class contains a lot of hash maps with complicated types that
-  // we'd like to keep out of the header.
-  class Tables;
-  std::unique_ptr<Tables> tables_;
-
-  bool enforce_dependencies_;
-  bool lazily_build_dependencies_;
-  bool allow_unknown_;
-  bool enforce_weak_;
-  bool disallow_enforce_utf8_;
-
-  // Set of files to track for unused imports. The bool value when true means
-  // unused imports are treated as errors (and as warnings when false).
-  std::map<std::string, bool> unused_import_track_files_;
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(DescriptorPool);
-};
-
-
-// inline methods ====================================================
-
-// These macros makes this repetitive code more readable.
-#define PROTOBUF_DEFINE_ACCESSOR(CLASS, FIELD, TYPE) \
-  inline TYPE CLASS::FIELD() const { return FIELD##_; }
-
-// Strings fields are stored as pointers but returned as const references.
-#define PROTOBUF_DEFINE_STRING_ACCESSOR(CLASS, FIELD) \
-  inline const std::string& CLASS::FIELD() const { return *FIELD##_; }
-
-// Arrays take an index parameter, obviously.
-#define PROTOBUF_DEFINE_ARRAY_ACCESSOR(CLASS, FIELD, TYPE) \
-  inline TYPE CLASS::FIELD(int index) const { return FIELD##s_ + index; }
-
-#define PROTOBUF_DEFINE_OPTIONS_ACCESSOR(CLASS, TYPE) \
-  inline const TYPE& CLASS::options() const { return *options_; }
-
-PROTOBUF_DEFINE_STRING_ACCESSOR(Descriptor, name)
-PROTOBUF_DEFINE_STRING_ACCESSOR(Descriptor, full_name)
-PROTOBUF_DEFINE_ACCESSOR(Descriptor, file, const FileDescriptor*)
-PROTOBUF_DEFINE_ACCESSOR(Descriptor, containing_type, const Descriptor*)
-
-PROTOBUF_DEFINE_ACCESSOR(Descriptor, field_count, int)
-PROTOBUF_DEFINE_ACCESSOR(Descriptor, oneof_decl_count, int)
-PROTOBUF_DEFINE_ACCESSOR(Descriptor, real_oneof_decl_count, int)
-PROTOBUF_DEFINE_ACCESSOR(Descriptor, nested_type_count, int)
-PROTOBUF_DEFINE_ACCESSOR(Descriptor, enum_type_count, int)
-
-PROTOBUF_DEFINE_ARRAY_ACCESSOR(Descriptor, field, const FieldDescriptor*)
-PROTOBUF_DEFINE_ARRAY_ACCESSOR(Descriptor, oneof_decl, const OneofDescriptor*)
-PROTOBUF_DEFINE_ARRAY_ACCESSOR(Descriptor, nested_type, const Descriptor*)
-PROTOBUF_DEFINE_ARRAY_ACCESSOR(Descriptor, enum_type, const EnumDescriptor*)
-
-PROTOBUF_DEFINE_ACCESSOR(Descriptor, extension_range_count, int)
-PROTOBUF_DEFINE_ACCESSOR(Descriptor, extension_count, int)
-PROTOBUF_DEFINE_ARRAY_ACCESSOR(Descriptor, extension_range,
-                               const Descriptor::ExtensionRange*)
-PROTOBUF_DEFINE_ARRAY_ACCESSOR(Descriptor, extension, const FieldDescriptor*)
-
-PROTOBUF_DEFINE_ACCESSOR(Descriptor, reserved_range_count, int)
-PROTOBUF_DEFINE_ARRAY_ACCESSOR(Descriptor, reserved_range,
-                               const Descriptor::ReservedRange*)
-PROTOBUF_DEFINE_ACCESSOR(Descriptor, reserved_name_count, int)
-
-PROTOBUF_DEFINE_OPTIONS_ACCESSOR(Descriptor, MessageOptions)
-PROTOBUF_DEFINE_ACCESSOR(Descriptor, is_placeholder, bool)
-
-PROTOBUF_DEFINE_STRING_ACCESSOR(FieldDescriptor, name)
-PROTOBUF_DEFINE_STRING_ACCESSOR(FieldDescriptor, full_name)
-PROTOBUF_DEFINE_STRING_ACCESSOR(FieldDescriptor, json_name)
-PROTOBUF_DEFINE_STRING_ACCESSOR(FieldDescriptor, lowercase_name)
-PROTOBUF_DEFINE_STRING_ACCESSOR(FieldDescriptor, camelcase_name)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, file, const FileDescriptor*)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, number, int)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, is_extension, bool)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, label, FieldDescriptor::Label)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, containing_type, const Descriptor*)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, containing_oneof,
-                         const OneofDescriptor*)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, index_in_oneof, int)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, extension_scope, const Descriptor*)
-PROTOBUF_DEFINE_OPTIONS_ACCESSOR(FieldDescriptor, FieldOptions)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, has_default_value, bool)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, has_json_name, bool)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, default_value_int32, int32)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, default_value_int64, int64)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, default_value_uint32, uint32)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, default_value_uint64, uint64)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, default_value_float, float)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, default_value_double, double)
-PROTOBUF_DEFINE_ACCESSOR(FieldDescriptor, default_value_bool, bool)
-PROTOBUF_DEFINE_STRING_ACCESSOR(FieldDescriptor, default_value_string)
-
-PROTOBUF_DEFINE_STRING_ACCESSOR(OneofDescriptor, name)
-PROTOBUF_DEFINE_STRING_ACCESSOR(OneofDescriptor, full_name)
-PROTOBUF_DEFINE_ACCESSOR(OneofDescriptor, containing_type, const Descriptor*)
-PROTOBUF_DEFINE_ACCESSOR(OneofDescriptor, field_count, int)
-PROTOBUF_DEFINE_OPTIONS_ACCESSOR(OneofDescriptor, OneofOptions)
-
-PROTOBUF_DEFINE_STRING_ACCESSOR(EnumDescriptor, name)
-PROTOBUF_DEFINE_STRING_ACCESSOR(EnumDescriptor, full_name)
-PROTOBUF_DEFINE_ACCESSOR(EnumDescriptor, file, const FileDescriptor*)
-PROTOBUF_DEFINE_ACCESSOR(EnumDescriptor, containing_type, const Descriptor*)
-PROTOBUF_DEFINE_ACCESSOR(EnumDescriptor, value_count, int)
-PROTOBUF_DEFINE_ARRAY_ACCESSOR(EnumDescriptor, value,
-                               const EnumValueDescriptor*)
-PROTOBUF_DEFINE_OPTIONS_ACCESSOR(EnumDescriptor, EnumOptions)
-PROTOBUF_DEFINE_ACCESSOR(EnumDescriptor, is_placeholder, bool)
-PROTOBUF_DEFINE_ACCESSOR(EnumDescriptor, reserved_range_count, int)
-PROTOBUF_DEFINE_ARRAY_ACCESSOR(EnumDescriptor, reserved_range,
-                               const EnumDescriptor::ReservedRange*)
-PROTOBUF_DEFINE_ACCESSOR(EnumDescriptor, reserved_name_count, int)
-
-PROTOBUF_DEFINE_STRING_ACCESSOR(EnumValueDescriptor, name)
-PROTOBUF_DEFINE_STRING_ACCESSOR(EnumValueDescriptor, full_name)
-PROTOBUF_DEFINE_ACCESSOR(EnumValueDescriptor, number, int)
-PROTOBUF_DEFINE_ACCESSOR(EnumValueDescriptor, type, const EnumDescriptor*)
-PROTOBUF_DEFINE_OPTIONS_ACCESSOR(EnumValueDescriptor, EnumValueOptions)
-
-PROTOBUF_DEFINE_STRING_ACCESSOR(ServiceDescriptor, name)
-PROTOBUF_DEFINE_STRING_ACCESSOR(ServiceDescriptor, full_name)
-PROTOBUF_DEFINE_ACCESSOR(ServiceDescriptor, file, const FileDescriptor*)
-PROTOBUF_DEFINE_ACCESSOR(ServiceDescriptor, method_count, int)
-PROTOBUF_DEFINE_ARRAY_ACCESSOR(ServiceDescriptor, method,
-                               const MethodDescriptor*)
-PROTOBUF_DEFINE_OPTIONS_ACCESSOR(ServiceDescriptor, ServiceOptions)
-
-PROTOBUF_DEFINE_STRING_ACCESSOR(MethodDescriptor, name)
-PROTOBUF_DEFINE_STRING_ACCESSOR(MethodDescriptor, full_name)
-PROTOBUF_DEFINE_ACCESSOR(MethodDescriptor, service, const ServiceDescriptor*)
-PROTOBUF_DEFINE_OPTIONS_ACCESSOR(MethodDescriptor, MethodOptions)
-PROTOBUF_DEFINE_ACCESSOR(MethodDescriptor, client_streaming, bool)
-PROTOBUF_DEFINE_ACCESSOR(MethodDescriptor, server_streaming, bool)
-
-PROTOBUF_DEFINE_STRING_ACCESSOR(FileDescriptor, name)
-PROTOBUF_DEFINE_STRING_ACCESSOR(FileDescriptor, package)
-PROTOBUF_DEFINE_ACCESSOR(FileDescriptor, pool, const DescriptorPool*)
-PROTOBUF_DEFINE_ACCESSOR(FileDescriptor, dependency_count, int)
-PROTOBUF_DEFINE_ACCESSOR(FileDescriptor, public_dependency_count, int)
-PROTOBUF_DEFINE_ACCESSOR(FileDescriptor, weak_dependency_count, int)
-PROTOBUF_DEFINE_ACCESSOR(FileDescriptor, message_type_count, int)
-PROTOBUF_DEFINE_ACCESSOR(FileDescriptor, enum_type_count, int)
-PROTOBUF_DEFINE_ACCESSOR(FileDescriptor, service_count, int)
-PROTOBUF_DEFINE_ACCESSOR(FileDescriptor, extension_count, int)
-PROTOBUF_DEFINE_OPTIONS_ACCESSOR(FileDescriptor, FileOptions)
-PROTOBUF_DEFINE_ACCESSOR(FileDescriptor, is_placeholder, bool)
-
-PROTOBUF_DEFINE_ARRAY_ACCESSOR(FileDescriptor, message_type, const Descriptor*)
-PROTOBUF_DEFINE_ARRAY_ACCESSOR(FileDescriptor, enum_type, const EnumDescriptor*)
-PROTOBUF_DEFINE_ARRAY_ACCESSOR(FileDescriptor, service,
-                               const ServiceDescriptor*)
-PROTOBUF_DEFINE_ARRAY_ACCESSOR(FileDescriptor, extension,
-                               const FieldDescriptor*)
-
-#undef PROTOBUF_DEFINE_ACCESSOR
-#undef PROTOBUF_DEFINE_STRING_ACCESSOR
-#undef PROTOBUF_DEFINE_ARRAY_ACCESSOR
-
-// A few accessors differ from the macros...
-
-inline Descriptor::WellKnownType Descriptor::well_known_type() const {
-  return static_cast<Descriptor::WellKnownType>(well_known_type_);
-}
-
-inline bool Descriptor::IsExtensionNumber(int number) const {
-  return FindExtensionRangeContainingNumber(number) != nullptr;
-}
-
-inline bool Descriptor::IsReservedNumber(int number) const {
-  return FindReservedRangeContainingNumber(number) != nullptr;
-}
-
-inline bool Descriptor::IsReservedName(ConstStringParam name) const {
-  for (int i = 0; i < reserved_name_count(); i++) {
-    if (name == static_cast<ConstStringParam>(reserved_name(i))) {
-      return true;
-    }
-  }
-  return false;
-}
-
-// Can't use PROTOBUF_DEFINE_ARRAY_ACCESSOR because reserved_names_ is actually
-// an array of pointers rather than the usual array of objects.
-inline const std::string& Descriptor::reserved_name(int index) const {
-  return *reserved_names_[index];
-}
-
-inline bool EnumDescriptor::IsReservedNumber(int number) const {
-  return FindReservedRangeContainingNumber(number) != nullptr;
-}
-
-inline bool EnumDescriptor::IsReservedName(ConstStringParam name) const {
-  for (int i = 0; i < reserved_name_count(); i++) {
-    if (name == static_cast<ConstStringParam>(reserved_name(i))) {
-      return true;
-    }
-  }
-  return false;
-}
-
-// Can't use PROTOBUF_DEFINE_ARRAY_ACCESSOR because reserved_names_ is actually
-// an array of pointers rather than the usual array of objects.
-inline const std::string& EnumDescriptor::reserved_name(int index) const {
-  return *reserved_names_[index];
-}
-
-inline FieldDescriptor::Type FieldDescriptor::type() const {
-  if (type_once_) {
-    internal::call_once(*type_once_, &FieldDescriptor::TypeOnceInit, this);
-  }
-  return type_;
-}
-
-inline bool FieldDescriptor::is_required() const {
-  return label() == LABEL_REQUIRED;
-}
-
-inline bool FieldDescriptor::is_optional() const {
-  return label() == LABEL_OPTIONAL;
-}
-
-inline bool FieldDescriptor::is_repeated() const {
-  return label() == LABEL_REPEATED;
-}
-
-inline bool FieldDescriptor::is_packable() const {
-  return is_repeated() && IsTypePackable(type());
-}
-
-inline bool FieldDescriptor::is_map() const {
-  return type() == TYPE_MESSAGE && is_map_message_type();
-}
-
-inline bool FieldDescriptor::has_optional_keyword() const {
-  return proto3_optional_ ||
-         (file()->syntax() == FileDescriptor::SYNTAX_PROTO2 && is_optional() &&
-          !containing_oneof());
-}
-
-inline const OneofDescriptor* FieldDescriptor::real_containing_oneof() const {
-  return containing_oneof_ && !containing_oneof_->is_synthetic()
-             ? containing_oneof_
-             : nullptr;
-}
-
-inline bool FieldDescriptor::has_presence() const {
-  if (is_repeated()) return false;
-  return cpp_type() == CPPTYPE_MESSAGE || containing_oneof() ||
-         file()->syntax() == FileDescriptor::SYNTAX_PROTO2;
-}
-
-// To save space, index() is computed by looking at the descriptor's position
-// in the parent's array of children.
-inline int FieldDescriptor::index() const {
-  if (!is_extension_) {
-    return static_cast<int>(this - containing_type()->fields_);
-  } else if (extension_scope_ != nullptr) {
-    return static_cast<int>(this - extension_scope_->extensions_);
-  } else {
-    return static_cast<int>(this - file_->extensions_);
-  }
-}
-
-inline int Descriptor::index() const {
-  if (containing_type_ == nullptr) {
-    return static_cast<int>(this - file_->message_types_);
-  } else {
-    return static_cast<int>(this - containing_type_->nested_types_);
-  }
-}
-
-inline const FileDescriptor* OneofDescriptor::file() const {
-  return containing_type()->file();
-}
-
-inline int OneofDescriptor::index() const {
-  return static_cast<int>(this - containing_type_->oneof_decls_);
-}
-
-inline bool OneofDescriptor::is_synthetic() const {
-  return field_count() == 1 && field(0)->proto3_optional_;
-}
-
-inline int EnumDescriptor::index() const {
-  if (containing_type_ == nullptr) {
-    return static_cast<int>(this - file_->enum_types_);
-  } else {
-    return static_cast<int>(this - containing_type_->enum_types_);
-  }
-}
-
-inline const FileDescriptor* EnumValueDescriptor::file() const {
-  return type()->file();
-}
-
-inline int EnumValueDescriptor::index() const {
-  return static_cast<int>(this - type_->values_);
-}
-
-inline int ServiceDescriptor::index() const {
-  return static_cast<int>(this - file_->services_);
-}
-
-inline const FileDescriptor* MethodDescriptor::file() const {
-  return service()->file();
-}
-
-inline int MethodDescriptor::index() const {
-  return static_cast<int>(this - service_->methods_);
-}
-
-inline const char* FieldDescriptor::type_name() const {
-  return kTypeToName[type()];
-}
-
-inline FieldDescriptor::CppType FieldDescriptor::cpp_type() const {
-  return kTypeToCppTypeMap[type()];
-}
-
-inline const char* FieldDescriptor::cpp_type_name() const {
-  return kCppTypeToName[kTypeToCppTypeMap[type()]];
-}
-
-inline FieldDescriptor::CppType FieldDescriptor::TypeToCppType(Type type) {
-  return kTypeToCppTypeMap[type];
-}
-
-inline const char* FieldDescriptor::TypeName(Type type) {
-  return kTypeToName[type];
-}
-
-inline const char* FieldDescriptor::CppTypeName(CppType cpp_type) {
-  return kCppTypeToName[cpp_type];
-}
-
-inline bool FieldDescriptor::IsTypePackable(Type field_type) {
-  return (field_type != FieldDescriptor::TYPE_STRING &&
-          field_type != FieldDescriptor::TYPE_GROUP &&
-          field_type != FieldDescriptor::TYPE_MESSAGE &&
-          field_type != FieldDescriptor::TYPE_BYTES);
-}
-
-inline const FileDescriptor* FileDescriptor::public_dependency(
-    int index) const {
-  return dependency(public_dependencies_[index]);
-}
-
-inline const FileDescriptor* FileDescriptor::weak_dependency(int index) const {
-  return dependency(weak_dependencies_[index]);
-}
-
-inline FileDescriptor::Syntax FileDescriptor::syntax() const { return syntax_; }
-
-// Can't use PROTOBUF_DEFINE_ARRAY_ACCESSOR because fields_ is actually an array
-// of pointers rather than the usual array of objects.
-inline const FieldDescriptor* OneofDescriptor::field(int index) const {
-  return fields_[index];
-}
-
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_DESCRIPTOR_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/extension_set.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/extension_set.h
deleted file mode 100644
index b3a6e3a0..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/extension_set.h
+++ /dev/null
@@ -1,1591 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// Author: kenton@google.com (Kenton Varda)
-//  Based on original Protocol Buffers design by
-//  Sanjay Ghemawat, Jeff Dean, and others.
-//
-// This header is logically internal, but is made public because it is used
-// from protocol-compiler-generated code, which may reside in other components.
-
-#ifndef GOOGLE_PROTOBUF_EXTENSION_SET_H__
-#define GOOGLE_PROTOBUF_EXTENSION_SET_H__
-
-#include <algorithm>
-#include <cassert>
-#include <map>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/stubs/logging.h>
-#include <google/protobuf/parse_context.h>
-#include <google/protobuf/io/coded_stream.h>
-#include <google/protobuf/port.h>
-#include <google/protobuf/repeated_field.h>
-#include <google/protobuf/wire_format_lite.h>
-
-#include <google/protobuf/port_def.inc>
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-namespace google {
-namespace protobuf {
-class Arena;
-class Descriptor;       // descriptor.h
-class FieldDescriptor;  // descriptor.h
-class DescriptorPool;   // descriptor.h
-class MessageLite;      // message_lite.h
-class Message;          // message.h
-class MessageFactory;   // message.h
-class UnknownFieldSet;  // unknown_field_set.h
-namespace internal {
-class FieldSkipper;  // wire_format_lite.h
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-class InternalMetadata;
-
-// Used to store values of type WireFormatLite::FieldType without having to
-// #include wire_format_lite.h.  Also, ensures that we use only one byte to
-// store these values, which is important to keep the layout of
-// ExtensionSet::Extension small.
-typedef uint8 FieldType;
-
-// A function which, given an integer value, returns true if the number
-// matches one of the defined values for the corresponding enum type.  This
-// is used with RegisterEnumExtension, below.
-typedef bool EnumValidityFunc(int number);
-
-// Version of the above which takes an argument.  This is needed to deal with
-// extensions that are not compiled in.
-typedef bool EnumValidityFuncWithArg(const void* arg, int number);
-
-// Information about a registered extension.
-struct ExtensionInfo {
-  inline ExtensionInfo() {}
-  inline ExtensionInfo(FieldType type_param, bool isrepeated, bool ispacked)
-      : type(type_param),
-        is_repeated(isrepeated),
-        is_packed(ispacked),
-        descriptor(NULL) {}
-
-  FieldType type;
-  bool is_repeated;
-  bool is_packed;
-
-  struct EnumValidityCheck {
-    EnumValidityFuncWithArg* func;
-    const void* arg;
-  };
-
-  struct MessageInfo {
-    const MessageLite* prototype;
-  };
-
-  union {
-    EnumValidityCheck enum_validity_check;
-    MessageInfo message_info;
-  };
-
-  // The descriptor for this extension, if one exists and is known.  May be
-  // NULL.  Must not be NULL if the descriptor for the extension does not
-  // live in the same pool as the descriptor for the containing type.
-  const FieldDescriptor* descriptor;
-};
-
-// Abstract interface for an object which looks up extension definitions.  Used
-// when parsing.
-class PROTOBUF_EXPORT ExtensionFinder {
- public:
-  virtual ~ExtensionFinder();
-
-  // Find the extension with the given containing type and number.
-  virtual bool Find(int number, ExtensionInfo* output) = 0;
-};
-
-// Implementation of ExtensionFinder which finds extensions defined in .proto
-// files which have been compiled into the binary.
-class PROTOBUF_EXPORT GeneratedExtensionFinder : public ExtensionFinder {
- public:
-  GeneratedExtensionFinder(const MessageLite* containing_type)
-      : containing_type_(containing_type) {}
-  ~GeneratedExtensionFinder() override {}
-
-  // Returns true and fills in *output if found, otherwise returns false.
-  bool Find(int number, ExtensionInfo* output) override;
-
- private:
-  const MessageLite* containing_type_;
-};
-
-// A FieldSkipper used for parsing MessageSet.
-class MessageSetFieldSkipper;
-
-// Note:  extension_set_heavy.cc defines DescriptorPoolExtensionFinder for
-// finding extensions from a DescriptorPool.
-
-// This is an internal helper class intended for use within the protocol buffer
-// library and generated classes.  Clients should not use it directly.  Instead,
-// use the generated accessors such as GetExtension() of the class being
-// extended.
-//
-// This class manages extensions for a protocol message object.  The
-// message's HasExtension(), GetExtension(), MutableExtension(), and
-// ClearExtension() methods are just thin wrappers around the embedded
-// ExtensionSet.  When parsing, if a tag number is encountered which is
-// inside one of the message type's extension ranges, the tag is passed
-// off to the ExtensionSet for parsing.  Etc.
-class PROTOBUF_EXPORT ExtensionSet {
- public:
-  constexpr ExtensionSet();
-  explicit ExtensionSet(Arena* arena);
-  ~ExtensionSet();
-
-  // These are called at startup by protocol-compiler-generated code to
-  // register known extensions.  The registrations are used by ParseField()
-  // to look up extensions for parsed field numbers.  Note that dynamic parsing
-  // does not use ParseField(); only protocol-compiler-generated parsing
-  // methods do.
-  static void RegisterExtension(const MessageLite* containing_type, int number,
-                                FieldType type, bool is_repeated,
-                                bool is_packed);
-  static void RegisterEnumExtension(const MessageLite* containing_type,
-                                    int number, FieldType type,
-                                    bool is_repeated, bool is_packed,
-                                    EnumValidityFunc* is_valid);
-  static void RegisterMessageExtension(const MessageLite* containing_type,
-                                       int number, FieldType type,
-                                       bool is_repeated, bool is_packed,
-                                       const MessageLite* prototype);
-
-  // =================================================================
-
-  // Add all fields which are currently present to the given vector.  This
-  // is useful to implement Reflection::ListFields().
-  void AppendToList(const Descriptor* containing_type,
-                    const DescriptorPool* pool,
-                    std::vector<const FieldDescriptor*>* output) const;
-
-  // =================================================================
-  // Accessors
-  //
-  // Generated message classes include type-safe templated wrappers around
-  // these methods.  Generally you should use those rather than call these
-  // directly, unless you are doing low-level memory management.
-  //
-  // When calling any of these accessors, the extension number requested
-  // MUST exist in the DescriptorPool provided to the constructor.  Otherwise,
-  // the method will fail an assert.  Normally, though, you would not call
-  // these directly; you would either call the generated accessors of your
-  // message class (e.g. GetExtension()) or you would call the accessors
-  // of the reflection interface.  In both cases, it is impossible to
-  // trigger this assert failure:  the generated accessors only accept
-  // linked-in extension types as parameters, while the Reflection interface
-  // requires you to provide the FieldDescriptor describing the extension.
-  //
-  // When calling any of these accessors, a protocol-compiler-generated
-  // implementation of the extension corresponding to the number MUST
-  // be linked in, and the FieldDescriptor used to refer to it MUST be
-  // the one generated by that linked-in code.  Otherwise, the method will
-  // die on an assert failure.  The message objects returned by the message
-  // accessors are guaranteed to be of the correct linked-in type.
-  //
-  // These methods pretty much match Reflection except that:
-  // - They're not virtual.
-  // - They identify fields by number rather than FieldDescriptors.
-  // - They identify enum values using integers rather than descriptors.
-  // - Strings provide Mutable() in addition to Set() accessors.
-
-  bool Has(int number) const;
-  int ExtensionSize(int number) const;  // Size of a repeated extension.
-  int NumExtensions() const;            // The number of extensions
-  FieldType ExtensionType(int number) const;
-  void ClearExtension(int number);
-
-  // singular fields -------------------------------------------------
-
-  int32 GetInt32(int number, int32 default_value) const;
-  int64 GetInt64(int number, int64 default_value) const;
-  uint32 GetUInt32(int number, uint32 default_value) const;
-  uint64 GetUInt64(int number, uint64 default_value) const;
-  float GetFloat(int number, float default_value) const;
-  double GetDouble(int number, double default_value) const;
-  bool GetBool(int number, bool default_value) const;
-  int GetEnum(int number, int default_value) const;
-  const std::string& GetString(int number,
-                               const std::string& default_value) const;
-  const MessageLite& GetMessage(int number,
-                                const MessageLite& default_value) const;
-  const MessageLite& GetMessage(int number, const Descriptor* message_type,
-                                MessageFactory* factory) const;
-
-  // |descriptor| may be NULL so long as it is known that the descriptor for
-  // the extension lives in the same pool as the descriptor for the containing
-  // type.
-#define desc const FieldDescriptor* descriptor  // avoid line wrapping
-  void SetInt32(int number, FieldType type, int32 value, desc);
-  void SetInt64(int number, FieldType type, int64 value, desc);
-  void SetUInt32(int number, FieldType type, uint32 value, desc);
-  void SetUInt64(int number, FieldType type, uint64 value, desc);
-  void SetFloat(int number, FieldType type, float value, desc);
-  void SetDouble(int number, FieldType type, double value, desc);
-  void SetBool(int number, FieldType type, bool value, desc);
-  void SetEnum(int number, FieldType type, int value, desc);
-  void SetString(int number, FieldType type, std::string value, desc);
-  std::string* MutableString(int number, FieldType type, desc);
-  MessageLite* MutableMessage(int number, FieldType type,
-                              const MessageLite& prototype, desc);
-  MessageLite* MutableMessage(const FieldDescriptor* descriptor,
-                              MessageFactory* factory);
-  // Adds the given message to the ExtensionSet, taking ownership of the
-  // message object. Existing message with the same number will be deleted.
-  // If "message" is NULL, this is equivalent to "ClearExtension(number)".
-  void SetAllocatedMessage(int number, FieldType type,
-                           const FieldDescriptor* descriptor,
-                           MessageLite* message);
-  void UnsafeArenaSetAllocatedMessage(int number, FieldType type,
-                                      const FieldDescriptor* descriptor,
-                                      MessageLite* message);
-  MessageLite* ReleaseMessage(int number, const MessageLite& prototype);
-  MessageLite* UnsafeArenaReleaseMessage(int number,
-                                         const MessageLite& prototype);
-
-  MessageLite* ReleaseMessage(const FieldDescriptor* descriptor,
-                              MessageFactory* factory);
-  MessageLite* UnsafeArenaReleaseMessage(const FieldDescriptor* descriptor,
-                                         MessageFactory* factory);
-#undef desc
-  Arena* GetArena() const { return arena_; }
-
-  // repeated fields -------------------------------------------------
-
-  // Fetches a RepeatedField extension by number; returns |default_value|
-  // if no such extension exists. User should not touch this directly; it is
-  // used by the GetRepeatedExtension() method.
-  const void* GetRawRepeatedField(int number, const void* default_value) const;
-  // Fetches a mutable version of a RepeatedField extension by number,
-  // instantiating one if none exists. Similar to above, user should not use
-  // this directly; it underlies MutableRepeatedExtension().
-  void* MutableRawRepeatedField(int number, FieldType field_type, bool packed,
-                                const FieldDescriptor* desc);
-
-  // This is an overload of MutableRawRepeatedField to maintain compatibility
-  // with old code using a previous API. This version of
-  // MutableRawRepeatedField() will GOOGLE_CHECK-fail on a missing extension.
-  // (E.g.: borg/clients/internal/proto1/proto2_reflection.cc.)
-  void* MutableRawRepeatedField(int number);
-
-  int32 GetRepeatedInt32(int number, int index) const;
-  int64 GetRepeatedInt64(int number, int index) const;
-  uint32 GetRepeatedUInt32(int number, int index) const;
-  uint64 GetRepeatedUInt64(int number, int index) const;
-  float GetRepeatedFloat(int number, int index) const;
-  double GetRepeatedDouble(int number, int index) const;
-  bool GetRepeatedBool(int number, int index) const;
-  int GetRepeatedEnum(int number, int index) const;
-  const std::string& GetRepeatedString(int number, int index) const;
-  const MessageLite& GetRepeatedMessage(int number, int index) const;
-
-  void SetRepeatedInt32(int number, int index, int32 value);
-  void SetRepeatedInt64(int number, int index, int64 value);
-  void SetRepeatedUInt32(int number, int index, uint32 value);
-  void SetRepeatedUInt64(int number, int index, uint64 value);
-  void SetRepeatedFloat(int number, int index, float value);
-  void SetRepeatedDouble(int number, int index, double value);
-  void SetRepeatedBool(int number, int index, bool value);
-  void SetRepeatedEnum(int number, int index, int value);
-  void SetRepeatedString(int number, int index, std::string value);
-  std::string* MutableRepeatedString(int number, int index);
-  MessageLite* MutableRepeatedMessage(int number, int index);
-
-#define desc const FieldDescriptor* descriptor  // avoid line wrapping
-  void AddInt32(int number, FieldType type, bool packed, int32 value, desc);
-  void AddInt64(int number, FieldType type, bool packed, int64 value, desc);
-  void AddUInt32(int number, FieldType type, bool packed, uint32 value, desc);
-  void AddUInt64(int number, FieldType type, bool packed, uint64 value, desc);
-  void AddFloat(int number, FieldType type, bool packed, float value, desc);
-  void AddDouble(int number, FieldType type, bool packed, double value, desc);
-  void AddBool(int number, FieldType type, bool packed, bool value, desc);
-  void AddEnum(int number, FieldType type, bool packed, int value, desc);
-  void AddString(int number, FieldType type, std::string value, desc);
-  std::string* AddString(int number, FieldType type, desc);
-  MessageLite* AddMessage(int number, FieldType type,
-                          const MessageLite& prototype, desc);
-  MessageLite* AddMessage(const FieldDescriptor* descriptor,
-                          MessageFactory* factory);
-  void AddAllocatedMessage(const FieldDescriptor* descriptor,
-                           MessageLite* new_entry);
-#undef desc
-
-  void RemoveLast(int number);
-  MessageLite* ReleaseLast(int number);
-  void SwapElements(int number, int index1, int index2);
-
-  // -----------------------------------------------------------------
-  // TODO(kenton):  Hardcore memory management accessors
-
-  // =================================================================
-  // convenience methods for implementing methods of Message
-  //
-  // These could all be implemented in terms of the other methods of this
-  // class, but providing them here helps keep the generated code size down.
-
-  void Clear();
-  void MergeFrom(const ExtensionSet& other);
-  void Swap(ExtensionSet* other);
-  void SwapExtension(ExtensionSet* other, int number);
-  bool IsInitialized() const;
-
-  // Parses a single extension from the input. The input should start out
-  // positioned immediately after the tag.
-  bool ParseField(uint32 tag, io::CodedInputStream* input,
-                  ExtensionFinder* extension_finder,
-                  FieldSkipper* field_skipper);
-
-  // Specific versions for lite or full messages (constructs the appropriate
-  // FieldSkipper automatically).  |containing_type| is the default
-  // instance for the containing message; it is used only to look up the
-  // extension by number.  See RegisterExtension(), above.  Unlike the other
-  // methods of ExtensionSet, this only works for generated message types --
-  // it looks up extensions registered using RegisterExtension().
-  bool ParseField(uint32 tag, io::CodedInputStream* input,
-                  const MessageLite* containing_type);
-  bool ParseField(uint32 tag, io::CodedInputStream* input,
-                  const Message* containing_type,
-                  UnknownFieldSet* unknown_fields);
-  bool ParseField(uint32 tag, io::CodedInputStream* input,
-                  const MessageLite* containing_type,
-                  io::CodedOutputStream* unknown_fields);
-
-  // Lite parser
-  const char* ParseField(uint64 tag, const char* ptr,
-                         const MessageLite* containing_type,
-                         internal::InternalMetadata* metadata,
-                         internal::ParseContext* ctx);
-  // Full parser
-  const char* ParseField(uint64 tag, const char* ptr,
-                         const Message* containing_type,
-                         internal::InternalMetadata* metadata,
-                         internal::ParseContext* ctx);
-  template <typename Msg>
-  const char* ParseMessageSet(const char* ptr, const Msg* containing_type,
-                              InternalMetadata* metadata,
-                              internal::ParseContext* ctx) {
-    struct MessageSetItem {
-      const char* _InternalParse(const char* ptr, ParseContext* ctx) {
-        return me->ParseMessageSetItem(ptr, containing_type, metadata, ctx);
-      }
-      ExtensionSet* me;
-      const Msg* containing_type;
-      InternalMetadata* metadata;
-    } item{this, containing_type, metadata};
-    while (!ctx->Done(&ptr)) {
-      uint32 tag;
-      ptr = ReadTag(ptr, &tag);
-      GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-      if (tag == WireFormatLite::kMessageSetItemStartTag) {
-        ptr = ctx->ParseGroup(&item, ptr, tag);
-        GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-      } else {
-        if (tag == 0 || (tag & 7) == 4) {
-          ctx->SetLastTag(tag);
-          return ptr;
-        }
-        ptr = ParseField(tag, ptr, containing_type, metadata, ctx);
-        GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-      }
-    }
-    return ptr;
-  }
-
-  // Parse an entire message in MessageSet format.  Such messages have no
-  // fields, only extensions.
-  bool ParseMessageSetLite(io::CodedInputStream* input,
-                           ExtensionFinder* extension_finder,
-                           FieldSkipper* field_skipper);
-  bool ParseMessageSet(io::CodedInputStream* input,
-                       ExtensionFinder* extension_finder,
-                       MessageSetFieldSkipper* field_skipper);
-
-  // Specific versions for lite or full messages (constructs the appropriate
-  // FieldSkipper automatically).
-  bool ParseMessageSet(io::CodedInputStream* input,
-                       const MessageLite* containing_type,
-                       std::string* unknown_fields);
-  bool ParseMessageSet(io::CodedInputStream* input,
-                       const Message* containing_type,
-                       UnknownFieldSet* unknown_fields);
-
-  // Write all extension fields with field numbers in the range
-  //   [start_field_number, end_field_number)
-  // to the output stream, using the cached sizes computed when ByteSize() was
-  // last called.  Note that the range bounds are inclusive-exclusive.
-  void SerializeWithCachedSizes(int start_field_number, int end_field_number,
-                                io::CodedOutputStream* output) const {
-    output->SetCur(_InternalSerialize(start_field_number, end_field_number,
-                                      output->Cur(), output->EpsCopy()));
-  }
-
-  // Same as SerializeWithCachedSizes, but without any bounds checking.
-  // The caller must ensure that target has sufficient capacity for the
-  // serialized extensions.
-  //
-  // Returns a pointer past the last written byte.
-  uint8* _InternalSerialize(int start_field_number, int end_field_number,
-                            uint8* target,
-                            io::EpsCopyOutputStream* stream) const;
-
-  // Like above but serializes in MessageSet format.
-  void SerializeMessageSetWithCachedSizes(io::CodedOutputStream* output) const {
-    output->SetCur(InternalSerializeMessageSetWithCachedSizesToArray(
-        output->Cur(), output->EpsCopy()));
-  }
-  uint8* InternalSerializeMessageSetWithCachedSizesToArray(
-      uint8* target, io::EpsCopyOutputStream* stream) const;
-
-  // For backward-compatibility, versions of two of the above methods that
-  // serialize deterministically iff SetDefaultSerializationDeterministic()
-  // has been called.
-  uint8* SerializeWithCachedSizesToArray(int start_field_number,
-                                         int end_field_number,
-                                         uint8* target) const;
-  uint8* SerializeMessageSetWithCachedSizesToArray(uint8* target) const;
-
-  // Returns the total serialized size of all the extensions.
-  size_t ByteSize() const;
-
-  // Like ByteSize() but uses MessageSet format.
-  size_t MessageSetByteSize() const;
-
-  // Returns (an estimate of) the total number of bytes used for storing the
-  // extensions in memory, excluding sizeof(*this).  If the ExtensionSet is
-  // for a lite message (and thus possibly contains lite messages), the results
-  // are undefined (might work, might crash, might corrupt data, might not even
-  // be linked in).  It's up to the protocol compiler to avoid calling this on
-  // such ExtensionSets (easy enough since lite messages don't implement
-  // SpaceUsed()).
-  size_t SpaceUsedExcludingSelfLong() const;
-
-  // This method just calls SpaceUsedExcludingSelfLong() but it can not be
-  // inlined because the definition of SpaceUsedExcludingSelfLong() is not
-  // included in lite runtime and when an inline method refers to it MSVC
-  // will complain about unresolved symbols when building the lite runtime
-  // as .dll.
-  int SpaceUsedExcludingSelf() const;
-
- private:
-  // Interface of a lazily parsed singular message extension.
-  class PROTOBUF_EXPORT LazyMessageExtension {
-   public:
-    LazyMessageExtension() {}
-    virtual ~LazyMessageExtension() {}
-
-    virtual LazyMessageExtension* New(Arena* arena) const = 0;
-    virtual const MessageLite& GetMessage(
-        const MessageLite& prototype) const = 0;
-    virtual MessageLite* MutableMessage(const MessageLite& prototype) = 0;
-    virtual void SetAllocatedMessage(MessageLite* message) = 0;
-    virtual void UnsafeArenaSetAllocatedMessage(MessageLite* message) = 0;
-    virtual MessageLite* ReleaseMessage(const MessageLite& prototype) = 0;
-    virtual MessageLite* UnsafeArenaReleaseMessage(
-        const MessageLite& prototype) = 0;
-
-    virtual bool IsInitialized() const = 0;
-
-    PROTOBUF_DEPRECATED_MSG("Please use ByteSizeLong() instead")
-    virtual int ByteSize() const { return internal::ToIntSize(ByteSizeLong()); }
-    virtual size_t ByteSizeLong() const = 0;
-    virtual size_t SpaceUsedLong() const = 0;
-
-    virtual void MergeFrom(const LazyMessageExtension& other) = 0;
-    virtual void Clear() = 0;
-
-    virtual bool ReadMessage(const MessageLite& prototype,
-                             io::CodedInputStream* input) = 0;
-    virtual const char* _InternalParse(const char* ptr, ParseContext* ctx) = 0;
-    virtual uint8* WriteMessageToArray(
-        int number, uint8* target, io::EpsCopyOutputStream* stream) const = 0;
-
-   private:
-    virtual void UnusedKeyMethod();  // Dummy key method to avoid weak vtable.
-
-    GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(LazyMessageExtension);
-  };
-  struct Extension {
-    // The order of these fields packs Extension into 24 bytes when using 8
-    // byte alignment. Consider this when adding or removing fields here.
-    union {
-      int32 int32_value;
-      int64 int64_value;
-      uint32 uint32_value;
-      uint64 uint64_value;
-      float float_value;
-      double double_value;
-      bool bool_value;
-      int enum_value;
-      std::string* string_value;
-      MessageLite* message_value;
-      LazyMessageExtension* lazymessage_value;
-
-      RepeatedField<int32>* repeated_int32_value;
-      RepeatedField<int64>* repeated_int64_value;
-      RepeatedField<uint32>* repeated_uint32_value;
-      RepeatedField<uint64>* repeated_uint64_value;
-      RepeatedField<float>* repeated_float_value;
-      RepeatedField<double>* repeated_double_value;
-      RepeatedField<bool>* repeated_bool_value;
-      RepeatedField<int>* repeated_enum_value;
-      RepeatedPtrField<std::string>* repeated_string_value;
-      RepeatedPtrField<MessageLite>* repeated_message_value;
-    };
-
-    FieldType type;
-    bool is_repeated;
-
-    // For singular types, indicates if the extension is "cleared".  This
-    // happens when an extension is set and then later cleared by the caller.
-    // We want to keep the Extension object around for reuse, so instead of
-    // removing it from the map, we just set is_cleared = true.  This has no
-    // meaning for repeated types; for those, the size of the RepeatedField
-    // simply becomes zero when cleared.
-    bool is_cleared : 4;
-
-    // For singular message types, indicates whether lazy parsing is enabled
-    // for this extension. This field is only valid when type == TYPE_MESSAGE
-    // and !is_repeated because we only support lazy parsing for singular
-    // message types currently. If is_lazy = true, the extension is stored in
-    // lazymessage_value. Otherwise, the extension will be message_value.
-    bool is_lazy : 4;
-
-    // For repeated types, this indicates if the [packed=true] option is set.
-    bool is_packed;
-
-    // For packed fields, the size of the packed data is recorded here when
-    // ByteSize() is called then used during serialization.
-    // TODO(kenton):  Use atomic<int> when C++ supports it.
-    mutable int cached_size;
-
-    // The descriptor for this extension, if one exists and is known.  May be
-    // NULL.  Must not be NULL if the descriptor for the extension does not
-    // live in the same pool as the descriptor for the containing type.
-    const FieldDescriptor* descriptor;
-
-    // Some helper methods for operations on a single Extension.
-    uint8* InternalSerializeFieldWithCachedSizesToArray(
-        int number, uint8* target, io::EpsCopyOutputStream* stream) const;
-    uint8* InternalSerializeMessageSetItemWithCachedSizesToArray(
-        int number, uint8* target, io::EpsCopyOutputStream* stream) const;
-    size_t ByteSize(int number) const;
-    size_t MessageSetItemByteSize(int number) const;
-    void Clear();
-    int GetSize() const;
-    void Free();
-    size_t SpaceUsedExcludingSelfLong() const;
-    bool IsInitialized() const;
-  };
-
-  // The Extension struct is small enough to be passed by value, so we use it
-  // directly as the value type in mappings rather than use pointers.  We use
-  // sorted maps rather than hash-maps because we expect most ExtensionSets will
-  // only contain a small number of extension.  Also, we want AppendToList and
-  // deterministic serialization to order fields by field number.
-
-  struct KeyValue {
-    int first;
-    Extension second;
-
-    struct FirstComparator {
-      bool operator()(const KeyValue& lhs, const KeyValue& rhs) const {
-        return lhs.first < rhs.first;
-      }
-      bool operator()(const KeyValue& lhs, int key) const {
-        return lhs.first < key;
-      }
-      bool operator()(int key, const KeyValue& rhs) const {
-        return key < rhs.first;
-      }
-    };
-  };
-
-  typedef std::map<int, Extension> LargeMap;
-
-  // Wrapper API that switches between flat-map and LargeMap.
-
-  // Finds a key (if present) in the ExtensionSet.
-  const Extension* FindOrNull(int key) const;
-  Extension* FindOrNull(int key);
-
-  // Helper-functions that only inspect the LargeMap.
-  const Extension* FindOrNullInLargeMap(int key) const;
-  Extension* FindOrNullInLargeMap(int key);
-
-  // Inserts a new (key, Extension) into the ExtensionSet (and returns true), or
-  // finds the already-existing Extension for that key (returns false).
-  // The Extension* will point to the new-or-found Extension.
-  std::pair<Extension*, bool> Insert(int key);
-
-  // Grows the flat_capacity_.
-  // If flat_capacity_ > kMaximumFlatCapacity, converts to LargeMap.
-  void GrowCapacity(size_t minimum_new_capacity);
-  static constexpr uint16 kMaximumFlatCapacity = 256;
-  bool is_large() const { return flat_capacity_ > kMaximumFlatCapacity; }
-
-  // Removes a key from the ExtensionSet.
-  void Erase(int key);
-
-  size_t Size() const {
-    return PROTOBUF_PREDICT_FALSE(is_large()) ? map_.large->size() : flat_size_;
-  }
-
-  // Similar to std::for_each.
-  // Each Iterator is decomposed into ->first and ->second fields, so
-  // that the KeyValueFunctor can be agnostic vis-a-vis KeyValue-vs-std::pair.
-  template <typename Iterator, typename KeyValueFunctor>
-  static KeyValueFunctor ForEach(Iterator begin, Iterator end,
-                                 KeyValueFunctor func) {
-    for (Iterator it = begin; it != end; ++it) func(it->first, it->second);
-    return std::move(func);
-  }
-
-  // Applies a functor to the <int, Extension&> pairs in sorted order.
-  template <typename KeyValueFunctor>
-  KeyValueFunctor ForEach(KeyValueFunctor func) {
-    if (PROTOBUF_PREDICT_FALSE(is_large())) {
-      return ForEach(map_.large->begin(), map_.large->end(), std::move(func));
-    }
-    return ForEach(flat_begin(), flat_end(), std::move(func));
-  }
-
-  // Applies a functor to the <int, const Extension&> pairs in sorted order.
-  template <typename KeyValueFunctor>
-  KeyValueFunctor ForEach(KeyValueFunctor func) const {
-    if (PROTOBUF_PREDICT_FALSE(is_large())) {
-      return ForEach(map_.large->begin(), map_.large->end(), std::move(func));
-    }
-    return ForEach(flat_begin(), flat_end(), std::move(func));
-  }
-
-  // Merges existing Extension from other_extension
-  void InternalExtensionMergeFrom(int number, const Extension& other_extension);
-
-  // Returns true and fills field_number and extension if extension is found.
-  // Note to support packed repeated field compatibility, it also fills whether
-  // the tag on wire is packed, which can be different from
-  // extension->is_packed (whether packed=true is specified).
-  bool FindExtensionInfoFromTag(uint32 tag, ExtensionFinder* extension_finder,
-                                int* field_number, ExtensionInfo* extension,
-                                bool* was_packed_on_wire);
-
-  // Returns true and fills extension if extension is found.
-  // Note to support packed repeated field compatibility, it also fills whether
-  // the tag on wire is packed, which can be different from
-  // extension->is_packed (whether packed=true is specified).
-  bool FindExtensionInfoFromFieldNumber(int wire_type, int field_number,
-                                        ExtensionFinder* extension_finder,
-                                        ExtensionInfo* extension,
-                                        bool* was_packed_on_wire);
-
-  // Parses a single extension from the input. The input should start out
-  // positioned immediately after the wire tag. This method is called in
-  // ParseField() after field number and was_packed_on_wire is extracted from
-  // the wire tag and ExtensionInfo is found by the field number.
-  bool ParseFieldWithExtensionInfo(int field_number, bool was_packed_on_wire,
-                                   const ExtensionInfo& extension,
-                                   io::CodedInputStream* input,
-                                   FieldSkipper* field_skipper);
-
-  // Like ParseField(), but this method may parse singular message extensions
-  // lazily depending on the value of FLAGS_eagerly_parse_message_sets.
-  bool ParseFieldMaybeLazily(int wire_type, int field_number,
-                             io::CodedInputStream* input,
-                             ExtensionFinder* extension_finder,
-                             MessageSetFieldSkipper* field_skipper);
-
-  // Gets the extension with the given number, creating it if it does not
-  // already exist.  Returns true if the extension did not already exist.
-  bool MaybeNewExtension(int number, const FieldDescriptor* descriptor,
-                         Extension** result);
-
-  // Gets the repeated extension for the given descriptor, creating it if
-  // it does not exist.
-  Extension* MaybeNewRepeatedExtension(const FieldDescriptor* descriptor);
-
-  // Parse a single MessageSet item -- called just after the item group start
-  // tag has been read.
-  bool ParseMessageSetItemLite(io::CodedInputStream* input,
-                               ExtensionFinder* extension_finder,
-                               FieldSkipper* field_skipper);
-  // Parse a single MessageSet item -- called just after the item group start
-  // tag has been read.
-  bool ParseMessageSetItem(io::CodedInputStream* input,
-                           ExtensionFinder* extension_finder,
-                           MessageSetFieldSkipper* field_skipper);
-
-  bool FindExtension(int wire_type, uint32 field,
-                     const MessageLite* containing_type,
-                     const internal::ParseContext* /*ctx*/,
-                     ExtensionInfo* extension, bool* was_packed_on_wire) {
-    GeneratedExtensionFinder finder(containing_type);
-    return FindExtensionInfoFromFieldNumber(wire_type, field, &finder,
-                                            extension, was_packed_on_wire);
-  }
-  inline bool FindExtension(int wire_type, uint32 field,
-                            const Message* containing_type,
-                            const internal::ParseContext* ctx,
-                            ExtensionInfo* extension, bool* was_packed_on_wire);
-  // Used for MessageSet only
-  const char* ParseFieldMaybeLazily(uint64 tag, const char* ptr,
-                                    const MessageLite* containing_type,
-                                    internal::InternalMetadata* metadata,
-                                    internal::ParseContext* ctx) {
-    // Lite MessageSet doesn't implement lazy.
-    return ParseField(tag, ptr, containing_type, metadata, ctx);
-  }
-  const char* ParseFieldMaybeLazily(uint64 tag, const char* ptr,
-                                    const Message* containing_type,
-                                    internal::InternalMetadata* metadata,
-                                    internal::ParseContext* ctx);
-  const char* ParseMessageSetItem(const char* ptr,
-                                  const MessageLite* containing_type,
-                                  internal::InternalMetadata* metadata,
-                                  internal::ParseContext* ctx);
-  const char* ParseMessageSetItem(const char* ptr,
-                                  const Message* containing_type,
-                                  internal::InternalMetadata* metadata,
-                                  internal::ParseContext* ctx);
-
-  // Implemented in extension_set_inl.h to keep code out of the header file.
-  template <typename T>
-  const char* ParseFieldWithExtensionInfo(int number, bool was_packed_on_wire,
-                                          const ExtensionInfo& info,
-                                          internal::InternalMetadata* metadata,
-                                          const char* ptr,
-                                          internal::ParseContext* ctx);
-  template <typename Msg, typename T>
-  const char* ParseMessageSetItemTmpl(const char* ptr,
-                                      const Msg* containing_type,
-                                      internal::InternalMetadata* metadata,
-                                      internal::ParseContext* ctx);
-
-  // Hack:  RepeatedPtrFieldBase declares ExtensionSet as a friend.  This
-  //   friendship should automatically extend to ExtensionSet::Extension, but
-  //   unfortunately some older compilers (e.g. GCC 3.4.4) do not implement this
-  //   correctly.  So, we must provide helpers for calling methods of that
-  //   class.
-
-  // Defined in extension_set_heavy.cc.
-  static inline size_t RepeatedMessage_SpaceUsedExcludingSelfLong(
-      RepeatedPtrFieldBase* field);
-
-  KeyValue* flat_begin() {
-    assert(!is_large());
-    return map_.flat;
-  }
-  const KeyValue* flat_begin() const {
-    assert(!is_large());
-    return map_.flat;
-  }
-  KeyValue* flat_end() {
-    assert(!is_large());
-    return map_.flat + flat_size_;
-  }
-  const KeyValue* flat_end() const {
-    assert(!is_large());
-    return map_.flat + flat_size_;
-  }
-
-  Arena* arena_;
-
-  // Manual memory-management:
-  // map_.flat is an allocated array of flat_capacity_ elements.
-  // [map_.flat, map_.flat + flat_size_) is the currently-in-use prefix.
-  uint16 flat_capacity_;
-  uint16 flat_size_;
-  union AllocatedData {
-    KeyValue* flat;
-
-    // If flat_capacity_ > kMaximumFlatCapacity, switch to LargeMap,
-    // which guarantees O(n lg n) CPU but larger constant factors.
-    LargeMap* large;
-  } map_;
-
-  static void DeleteFlatMap(const KeyValue* flat, uint16 flat_capacity);
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ExtensionSet);
-};
-
-constexpr ExtensionSet::ExtensionSet()
-    : arena_(nullptr), flat_capacity_(0), flat_size_(0), map_{nullptr} {}
-
-// These are just for convenience...
-inline void ExtensionSet::SetString(int number, FieldType type,
-                                    std::string value,
-                                    const FieldDescriptor* descriptor) {
-  MutableString(number, type, descriptor)->assign(std::move(value));
-}
-inline void ExtensionSet::SetRepeatedString(int number, int index,
-                                            std::string value) {
-  MutableRepeatedString(number, index)->assign(std::move(value));
-}
-inline void ExtensionSet::AddString(int number, FieldType type,
-                                    std::string value,
-                                    const FieldDescriptor* descriptor) {
-  AddString(number, type, descriptor)->assign(std::move(value));
-}
-// ===================================================================
-// Glue for generated extension accessors
-
-// -------------------------------------------------------------------
-// Template magic
-
-// First we have a set of classes representing "type traits" for different
-// field types.  A type traits class knows how to implement basic accessors
-// for extensions of a particular type given an ExtensionSet.  The signature
-// for a type traits class looks like this:
-//
-//   class TypeTraits {
-//    public:
-//     typedef ? ConstType;
-//     typedef ? MutableType;
-//     // TypeTraits for singular fields and repeated fields will define the
-//     // symbol "Singular" or "Repeated" respectively. These two symbols will
-//     // be used in extension accessors to distinguish between singular
-//     // extensions and repeated extensions. If the TypeTraits for the passed
-//     // in extension doesn't have the expected symbol defined, it means the
-//     // user is passing a repeated extension to a singular accessor, or the
-//     // opposite. In that case the C++ compiler will generate an error
-//     // message "no matching member function" to inform the user.
-//     typedef ? Singular
-//     typedef ? Repeated
-//
-//     static inline ConstType Get(int number, const ExtensionSet& set);
-//     static inline void Set(int number, ConstType value, ExtensionSet* set);
-//     static inline MutableType Mutable(int number, ExtensionSet* set);
-//
-//     // Variants for repeated fields.
-//     static inline ConstType Get(int number, const ExtensionSet& set,
-//                                 int index);
-//     static inline void Set(int number, int index,
-//                            ConstType value, ExtensionSet* set);
-//     static inline MutableType Mutable(int number, int index,
-//                                       ExtensionSet* set);
-//     static inline void Add(int number, ConstType value, ExtensionSet* set);
-//     static inline MutableType Add(int number, ExtensionSet* set);
-//     This is used by the ExtensionIdentifier constructor to register
-//     the extension at dynamic initialization.
-//     template <typename ExtendeeT>
-//     static void Register(int number, FieldType type, bool is_packed);
-//   };
-//
-// Not all of these methods make sense for all field types.  For example, the
-// "Mutable" methods only make sense for strings and messages, and the
-// repeated methods only make sense for repeated types.  So, each type
-// traits class implements only the set of methods from this signature that it
-// actually supports.  This will cause a compiler error if the user tries to
-// access an extension using a method that doesn't make sense for its type.
-// For example, if "foo" is an extension of type "optional int32", then if you
-// try to write code like:
-//   my_message.MutableExtension(foo)
-// you will get a compile error because PrimitiveTypeTraits<int32> does not
-// have a "Mutable()" method.
-
-// -------------------------------------------------------------------
-// PrimitiveTypeTraits
-
-// Since the ExtensionSet has different methods for each primitive type,
-// we must explicitly define the methods of the type traits class for each
-// known type.
-template <typename Type>
-class PrimitiveTypeTraits {
- public:
-  typedef Type ConstType;
-  typedef Type MutableType;
-  typedef PrimitiveTypeTraits<Type> Singular;
-
-  static inline ConstType Get(int number, const ExtensionSet& set,
-                              ConstType default_value);
-  static inline void Set(int number, FieldType field_type, ConstType value,
-                         ExtensionSet* set);
-  template <typename ExtendeeT>
-  static void Register(int number, FieldType type, bool is_packed) {
-    ExtensionSet::RegisterExtension(&ExtendeeT::default_instance(), number,
-                                    type, false, is_packed);
-  }
-};
-
-template <typename Type>
-class RepeatedPrimitiveTypeTraits {
- public:
-  typedef Type ConstType;
-  typedef Type MutableType;
-  typedef RepeatedPrimitiveTypeTraits<Type> Repeated;
-
-  typedef RepeatedField<Type> RepeatedFieldType;
-
-  static inline Type Get(int number, const ExtensionSet& set, int index);
-  static inline void Set(int number, int index, Type value, ExtensionSet* set);
-  static inline void Add(int number, FieldType field_type, bool is_packed,
-                         Type value, ExtensionSet* set);
-
-  static inline const RepeatedField<ConstType>& GetRepeated(
-      int number, const ExtensionSet& set);
-  static inline RepeatedField<Type>* MutableRepeated(int number,
-                                                     FieldType field_type,
-                                                     bool is_packed,
-                                                     ExtensionSet* set);
-
-  static const RepeatedFieldType* GetDefaultRepeatedField();
-  template <typename ExtendeeT>
-  static void Register(int number, FieldType type, bool is_packed) {
-    ExtensionSet::RegisterExtension(&ExtendeeT::default_instance(), number,
-                                    type, true, is_packed);
-  }
-};
-
-class PROTOBUF_EXPORT RepeatedPrimitiveDefaults {
- private:
-  template <typename Type>
-  friend class RepeatedPrimitiveTypeTraits;
-  static const RepeatedPrimitiveDefaults* default_instance();
-  RepeatedField<int32> default_repeated_field_int32_;
-  RepeatedField<int64> default_repeated_field_int64_;
-  RepeatedField<uint32> default_repeated_field_uint32_;
-  RepeatedField<uint64> default_repeated_field_uint64_;
-  RepeatedField<double> default_repeated_field_double_;
-  RepeatedField<float> default_repeated_field_float_;
-  RepeatedField<bool> default_repeated_field_bool_;
-};
-
-#define PROTOBUF_DEFINE_PRIMITIVE_TYPE(TYPE, METHOD)                           \
-  template <>                                                                  \
-  inline TYPE PrimitiveTypeTraits<TYPE>::Get(                                  \
-      int number, const ExtensionSet& set, TYPE default_value) {               \
-    return set.Get##METHOD(number, default_value);                             \
-  }                                                                            \
-  template <>                                                                  \
-  inline void PrimitiveTypeTraits<TYPE>::Set(int number, FieldType field_type, \
-                                             TYPE value, ExtensionSet* set) {  \
-    set->Set##METHOD(number, field_type, value, NULL);                         \
-  }                                                                            \
-                                                                               \
-  template <>                                                                  \
-  inline TYPE RepeatedPrimitiveTypeTraits<TYPE>::Get(                          \
-      int number, const ExtensionSet& set, int index) {                        \
-    return set.GetRepeated##METHOD(number, index);                             \
-  }                                                                            \
-  template <>                                                                  \
-  inline void RepeatedPrimitiveTypeTraits<TYPE>::Set(                          \
-      int number, int index, TYPE value, ExtensionSet* set) {                  \
-    set->SetRepeated##METHOD(number, index, value);                            \
-  }                                                                            \
-  template <>                                                                  \
-  inline void RepeatedPrimitiveTypeTraits<TYPE>::Add(                          \
-      int number, FieldType field_type, bool is_packed, TYPE value,            \
-      ExtensionSet* set) {                                                     \
-    set->Add##METHOD(number, field_type, is_packed, value, NULL);              \
-  }                                                                            \
-  template <>                                                                  \
-  inline const RepeatedField<TYPE>*                                            \
-  RepeatedPrimitiveTypeTraits<TYPE>::GetDefaultRepeatedField() {               \
-    return &RepeatedPrimitiveDefaults::default_instance()                      \
-                ->default_repeated_field_##TYPE##_;                            \
-  }                                                                            \
-  template <>                                                                  \
-  inline const RepeatedField<TYPE>&                                            \
-  RepeatedPrimitiveTypeTraits<TYPE>::GetRepeated(int number,                   \
-                                                 const ExtensionSet& set) {    \
-    return *reinterpret_cast<const RepeatedField<TYPE>*>(                      \
-        set.GetRawRepeatedField(number, GetDefaultRepeatedField()));           \
-  }                                                                            \
-  template <>                                                                  \
-  inline RepeatedField<TYPE>*                                                  \
-  RepeatedPrimitiveTypeTraits<TYPE>::MutableRepeated(                          \
-      int number, FieldType field_type, bool is_packed, ExtensionSet* set) {   \
-    return reinterpret_cast<RepeatedField<TYPE>*>(                             \
-        set->MutableRawRepeatedField(number, field_type, is_packed, NULL));    \
-  }
-
-PROTOBUF_DEFINE_PRIMITIVE_TYPE(int32, Int32)
-PROTOBUF_DEFINE_PRIMITIVE_TYPE(int64, Int64)
-PROTOBUF_DEFINE_PRIMITIVE_TYPE(uint32, UInt32)
-PROTOBUF_DEFINE_PRIMITIVE_TYPE(uint64, UInt64)
-PROTOBUF_DEFINE_PRIMITIVE_TYPE(float, Float)
-PROTOBUF_DEFINE_PRIMITIVE_TYPE(double, Double)
-PROTOBUF_DEFINE_PRIMITIVE_TYPE(bool, Bool)
-
-#undef PROTOBUF_DEFINE_PRIMITIVE_TYPE
-
-// -------------------------------------------------------------------
-// StringTypeTraits
-
-// Strings support both Set() and Mutable().
-class PROTOBUF_EXPORT StringTypeTraits {
- public:
-  typedef const std::string& ConstType;
-  typedef std::string* MutableType;
-  typedef StringTypeTraits Singular;
-
-  static inline const std::string& Get(int number, const ExtensionSet& set,
-                                       ConstType default_value) {
-    return set.GetString(number, default_value);
-  }
-  static inline void Set(int number, FieldType field_type,
-                         const std::string& value, ExtensionSet* set) {
-    set->SetString(number, field_type, value, NULL);
-  }
-  static inline std::string* Mutable(int number, FieldType field_type,
-                                     ExtensionSet* set) {
-    return set->MutableString(number, field_type, NULL);
-  }
-  template <typename ExtendeeT>
-  static void Register(int number, FieldType type, bool is_packed) {
-    ExtensionSet::RegisterExtension(&ExtendeeT::default_instance(), number,
-                                    type, false, is_packed);
-  }
-};
-
-class PROTOBUF_EXPORT RepeatedStringTypeTraits {
- public:
-  typedef const std::string& ConstType;
-  typedef std::string* MutableType;
-  typedef RepeatedStringTypeTraits Repeated;
-
-  typedef RepeatedPtrField<std::string> RepeatedFieldType;
-
-  static inline const std::string& Get(int number, const ExtensionSet& set,
-                                       int index) {
-    return set.GetRepeatedString(number, index);
-  }
-  static inline void Set(int number, int index, const std::string& value,
-                         ExtensionSet* set) {
-    set->SetRepeatedString(number, index, value);
-  }
-  static inline std::string* Mutable(int number, int index, ExtensionSet* set) {
-    return set->MutableRepeatedString(number, index);
-  }
-  static inline void Add(int number, FieldType field_type, bool /*is_packed*/,
-                         const std::string& value, ExtensionSet* set) {
-    set->AddString(number, field_type, value, NULL);
-  }
-  static inline std::string* Add(int number, FieldType field_type,
-                                 ExtensionSet* set) {
-    return set->AddString(number, field_type, NULL);
-  }
-  static inline const RepeatedPtrField<std::string>& GetRepeated(
-      int number, const ExtensionSet& set) {
-    return *reinterpret_cast<const RepeatedPtrField<std::string>*>(
-        set.GetRawRepeatedField(number, GetDefaultRepeatedField()));
-  }
-
-  static inline RepeatedPtrField<std::string>* MutableRepeated(
-      int number, FieldType field_type, bool is_packed, ExtensionSet* set) {
-    return reinterpret_cast<RepeatedPtrField<std::string>*>(
-        set->MutableRawRepeatedField(number, field_type, is_packed, NULL));
-  }
-
-  static const RepeatedFieldType* GetDefaultRepeatedField();
-
-  template <typename ExtendeeT>
-  static void Register(int number, FieldType type, bool is_packed) {
-    ExtensionSet::RegisterExtension(&ExtendeeT::default_instance(), number,
-                                    type, true, is_packed);
-  }
-
- private:
-  static void InitializeDefaultRepeatedFields();
-  static void DestroyDefaultRepeatedFields();
-};
-
-// -------------------------------------------------------------------
-// EnumTypeTraits
-
-// ExtensionSet represents enums using integers internally, so we have to
-// static_cast around.
-template <typename Type, bool IsValid(int)>
-class EnumTypeTraits {
- public:
-  typedef Type ConstType;
-  typedef Type MutableType;
-  typedef EnumTypeTraits<Type, IsValid> Singular;
-
-  static inline ConstType Get(int number, const ExtensionSet& set,
-                              ConstType default_value) {
-    return static_cast<Type>(set.GetEnum(number, default_value));
-  }
-  static inline void Set(int number, FieldType field_type, ConstType value,
-                         ExtensionSet* set) {
-    GOOGLE_DCHECK(IsValid(value));
-    set->SetEnum(number, field_type, value, NULL);
-  }
-  template <typename ExtendeeT>
-  static void Register(int number, FieldType type, bool is_packed) {
-    ExtensionSet::RegisterEnumExtension(&ExtendeeT::default_instance(), number,
-                                        type, false, is_packed, IsValid);
-  }
-};
-
-template <typename Type, bool IsValid(int)>
-class RepeatedEnumTypeTraits {
- public:
-  typedef Type ConstType;
-  typedef Type MutableType;
-  typedef RepeatedEnumTypeTraits<Type, IsValid> Repeated;
-
-  typedef RepeatedField<Type> RepeatedFieldType;
-
-  static inline ConstType Get(int number, const ExtensionSet& set, int index) {
-    return static_cast<Type>(set.GetRepeatedEnum(number, index));
-  }
-  static inline void Set(int number, int index, ConstType value,
-                         ExtensionSet* set) {
-    GOOGLE_DCHECK(IsValid(value));
-    set->SetRepeatedEnum(number, index, value);
-  }
-  static inline void Add(int number, FieldType field_type, bool is_packed,
-                         ConstType value, ExtensionSet* set) {
-    GOOGLE_DCHECK(IsValid(value));
-    set->AddEnum(number, field_type, is_packed, value, NULL);
-  }
-  static inline const RepeatedField<Type>& GetRepeated(
-      int number, const ExtensionSet& set) {
-    // Hack: the `Extension` struct stores a RepeatedField<int> for enums.
-    // RepeatedField<int> cannot implicitly convert to RepeatedField<EnumType>
-    // so we need to do some casting magic. See message.h for similar
-    // contortions for non-extension fields.
-    return *reinterpret_cast<const RepeatedField<Type>*>(
-        set.GetRawRepeatedField(number, GetDefaultRepeatedField()));
-  }
-
-  static inline RepeatedField<Type>* MutableRepeated(int number,
-                                                     FieldType field_type,
-                                                     bool is_packed,
-                                                     ExtensionSet* set) {
-    return reinterpret_cast<RepeatedField<Type>*>(
-        set->MutableRawRepeatedField(number, field_type, is_packed, NULL));
-  }
-
-  static const RepeatedFieldType* GetDefaultRepeatedField() {
-    // Hack: as noted above, repeated enum fields are internally stored as a
-    // RepeatedField<int>. We need to be able to instantiate global static
-    // objects to return as default (empty) repeated fields on non-existent
-    // extensions. We would not be able to know a-priori all of the enum types
-    // (values of |Type|) to instantiate all of these, so we just re-use int32's
-    // default repeated field object.
-    return reinterpret_cast<const RepeatedField<Type>*>(
-        RepeatedPrimitiveTypeTraits<int32>::GetDefaultRepeatedField());
-  }
-  template <typename ExtendeeT>
-  static void Register(int number, FieldType type, bool is_packed) {
-    ExtensionSet::RegisterEnumExtension(&ExtendeeT::default_instance(), number,
-                                        type, true, is_packed, IsValid);
-  }
-};
-
-// -------------------------------------------------------------------
-// MessageTypeTraits
-
-// ExtensionSet guarantees that when manipulating extensions with message
-// types, the implementation used will be the compiled-in class representing
-// that type.  So, we can static_cast down to the exact type we expect.
-template <typename Type>
-class MessageTypeTraits {
- public:
-  typedef const Type& ConstType;
-  typedef Type* MutableType;
-  typedef MessageTypeTraits<Type> Singular;
-
-  static inline ConstType Get(int number, const ExtensionSet& set,
-                              ConstType default_value) {
-    return static_cast<const Type&>(set.GetMessage(number, default_value));
-  }
-  static inline MutableType Mutable(int number, FieldType field_type,
-                                    ExtensionSet* set) {
-    return static_cast<Type*>(set->MutableMessage(
-        number, field_type, Type::default_instance(), NULL));
-  }
-  static inline void SetAllocated(int number, FieldType field_type,
-                                  MutableType message, ExtensionSet* set) {
-    set->SetAllocatedMessage(number, field_type, NULL, message);
-  }
-  static inline void UnsafeArenaSetAllocated(int number, FieldType field_type,
-                                             MutableType message,
-                                             ExtensionSet* set) {
-    set->UnsafeArenaSetAllocatedMessage(number, field_type, NULL, message);
-  }
-  static inline MutableType Release(int number, FieldType /* field_type */,
-                                    ExtensionSet* set) {
-    return static_cast<Type*>(
-        set->ReleaseMessage(number, Type::default_instance()));
-  }
-  static inline MutableType UnsafeArenaRelease(int number,
-                                               FieldType /* field_type */,
-                                               ExtensionSet* set) {
-    return static_cast<Type*>(
-        set->UnsafeArenaReleaseMessage(number, Type::default_instance()));
-  }
-  template <typename ExtendeeT>
-  static void Register(int number, FieldType type, bool is_packed) {
-    ExtensionSet::RegisterMessageExtension(&ExtendeeT::default_instance(),
-                                           number, type, false, is_packed,
-                                           &Type::default_instance());
-  }
-};
-
-// forward declaration
-class RepeatedMessageGenericTypeTraits;
-
-template <typename Type>
-class RepeatedMessageTypeTraits {
- public:
-  typedef const Type& ConstType;
-  typedef Type* MutableType;
-  typedef RepeatedMessageTypeTraits<Type> Repeated;
-
-  typedef RepeatedPtrField<Type> RepeatedFieldType;
-
-  static inline ConstType Get(int number, const ExtensionSet& set, int index) {
-    return static_cast<const Type&>(set.GetRepeatedMessage(number, index));
-  }
-  static inline MutableType Mutable(int number, int index, ExtensionSet* set) {
-    return static_cast<Type*>(set->MutableRepeatedMessage(number, index));
-  }
-  static inline MutableType Add(int number, FieldType field_type,
-                                ExtensionSet* set) {
-    return static_cast<Type*>(
-        set->AddMessage(number, field_type, Type::default_instance(), NULL));
-  }
-  static inline const RepeatedPtrField<Type>& GetRepeated(
-      int number, const ExtensionSet& set) {
-    // See notes above in RepeatedEnumTypeTraits::GetRepeated(): same
-    // casting hack applies here, because a RepeatedPtrField<MessageLite>
-    // cannot naturally become a RepeatedPtrType<Type> even though Type is
-    // presumably a message. google::protobuf::Message goes through similar contortions
-    // with a reinterpret_cast<>.
-    return *reinterpret_cast<const RepeatedPtrField<Type>*>(
-        set.GetRawRepeatedField(number, GetDefaultRepeatedField()));
-  }
-  static inline RepeatedPtrField<Type>* MutableRepeated(int number,
-                                                        FieldType field_type,
-                                                        bool is_packed,
-                                                        ExtensionSet* set) {
-    return reinterpret_cast<RepeatedPtrField<Type>*>(
-        set->MutableRawRepeatedField(number, field_type, is_packed, NULL));
-  }
-
-  static const RepeatedFieldType* GetDefaultRepeatedField();
-  template <typename ExtendeeT>
-  static void Register(int number, FieldType type, bool is_packed) {
-    ExtensionSet::RegisterMessageExtension(&ExtendeeT::default_instance(),
-                                           number, type, true, is_packed,
-                                           &Type::default_instance());
-  }
-};
-
-template <typename Type>
-inline const typename RepeatedMessageTypeTraits<Type>::RepeatedFieldType*
-RepeatedMessageTypeTraits<Type>::GetDefaultRepeatedField() {
-  static auto instance = OnShutdownDelete(new RepeatedFieldType);
-  return instance;
-}
-
-// -------------------------------------------------------------------
-// ExtensionIdentifier
-
-// This is the type of actual extension objects.  E.g. if you have:
-//   extends Foo with optional int32 bar = 1234;
-// then "bar" will be defined in C++ as:
-//   ExtensionIdentifier<Foo, PrimitiveTypeTraits<int32>, 5, false> bar(1234);
-//
-// Note that we could, in theory, supply the field number as a template
-// parameter, and thus make an instance of ExtensionIdentifier have no
-// actual contents.  However, if we did that, then using an extension
-// identifier would not necessarily cause the compiler to output any sort
-// of reference to any symbol defined in the extension's .pb.o file.  Some
-// linkers will actually drop object files that are not explicitly referenced,
-// but that would be bad because it would cause this extension to not be
-// registered at static initialization, and therefore using it would crash.
-
-template <typename ExtendeeType, typename TypeTraitsType, FieldType field_type,
-          bool is_packed>
-class ExtensionIdentifier {
- public:
-  typedef TypeTraitsType TypeTraits;
-  typedef ExtendeeType Extendee;
-
-  ExtensionIdentifier(int number, typename TypeTraits::ConstType default_value)
-      : number_(number), default_value_(default_value) {
-    Register(number);
-  }
-  inline int number() const { return number_; }
-  typename TypeTraits::ConstType default_value() const {
-    return default_value_;
-  }
-
-  static void Register(int number) {
-    TypeTraits::template Register<ExtendeeType>(number, field_type, is_packed);
-  }
-
- private:
-  const int number_;
-  typename TypeTraits::ConstType default_value_;
-};
-
-// -------------------------------------------------------------------
-// Generated accessors
-
-// This macro should be expanded in the context of a generated type which
-// has extensions.
-//
-// We use "_proto_TypeTraits" as a type name below because "TypeTraits"
-// causes problems if the class has a nested message or enum type with that
-// name and "_TypeTraits" is technically reserved for the C++ library since
-// it starts with an underscore followed by a capital letter.
-//
-// For similar reason, we use "_field_type" and "_is_packed" as parameter names
-// below, so that "field_type" and "is_packed" can be used as field names.
-#define GOOGLE_PROTOBUF_EXTENSION_ACCESSORS(CLASSNAME)                       \
-  /* Has, Size, Clear */                                                      \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline bool HasExtension(                                                   \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id) const { \
-    return _extensions_.Has(id.number());                                     \
-  }                                                                           \
-                                                                              \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline void ClearExtension(                                                 \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id) {       \
-    _extensions_.ClearExtension(id.number());                                 \
-  }                                                                           \
-                                                                              \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline int ExtensionSize(                                                   \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id) const { \
-    return _extensions_.ExtensionSize(id.number());                           \
-  }                                                                           \
-                                                                              \
-  /* Singular accessors */                                                    \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline typename _proto_TypeTraits::Singular::ConstType GetExtension(        \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id) const { \
-    return _proto_TypeTraits::Get(id.number(), _extensions_,                  \
-                                  id.default_value());                        \
-  }                                                                           \
-                                                                              \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline typename _proto_TypeTraits::Singular::MutableType MutableExtension(  \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id) {       \
-    return _proto_TypeTraits::Mutable(id.number(), _field_type,               \
-                                      &_extensions_);                         \
-  }                                                                           \
-                                                                              \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline void SetExtension(                                                   \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id,         \
-      typename _proto_TypeTraits::Singular::ConstType value) {                \
-    _proto_TypeTraits::Set(id.number(), _field_type, value, &_extensions_);   \
-  }                                                                           \
-                                                                              \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline void SetAllocatedExtension(                                          \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id,         \
-      typename _proto_TypeTraits::Singular::MutableType value) {              \
-    _proto_TypeTraits::SetAllocated(id.number(), _field_type, value,          \
-                                    &_extensions_);                           \
-  }                                                                           \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline void UnsafeArenaSetAllocatedExtension(                               \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id,         \
-      typename _proto_TypeTraits::Singular::MutableType value) {              \
-    _proto_TypeTraits::UnsafeArenaSetAllocated(id.number(), _field_type,      \
-                                               value, &_extensions_);         \
-  }                                                                           \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline typename _proto_TypeTraits::Singular::MutableType ReleaseExtension(  \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id) {       \
-    return _proto_TypeTraits::Release(id.number(), _field_type,               \
-                                      &_extensions_);                         \
-  }                                                                           \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline typename _proto_TypeTraits::Singular::MutableType                    \
-  UnsafeArenaReleaseExtension(                                                \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id) {       \
-    return _proto_TypeTraits::UnsafeArenaRelease(id.number(), _field_type,    \
-                                                 &_extensions_);              \
-  }                                                                           \
-                                                                              \
-  /* Repeated accessors */                                                    \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline typename _proto_TypeTraits::Repeated::ConstType GetExtension(        \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id,         \
-      int index) const {                                                      \
-    return _proto_TypeTraits::Get(id.number(), _extensions_, index);          \
-  }                                                                           \
-                                                                              \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline typename _proto_TypeTraits::Repeated::MutableType MutableExtension(  \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id,         \
-      int index) {                                                            \
-    return _proto_TypeTraits::Mutable(id.number(), index, &_extensions_);     \
-  }                                                                           \
-                                                                              \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline void SetExtension(                                                   \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id,         \
-      int index, typename _proto_TypeTraits::Repeated::ConstType value) {     \
-    _proto_TypeTraits::Set(id.number(), index, value, &_extensions_);         \
-  }                                                                           \
-                                                                              \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline typename _proto_TypeTraits::Repeated::MutableType AddExtension(      \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id) {       \
-    return _proto_TypeTraits::Add(id.number(), _field_type, &_extensions_);   \
-  }                                                                           \
-                                                                              \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline void AddExtension(                                                   \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id,         \
-      typename _proto_TypeTraits::Repeated::ConstType value) {                \
-    _proto_TypeTraits::Add(id.number(), _field_type, _is_packed, value,       \
-                           &_extensions_);                                    \
-  }                                                                           \
-                                                                              \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline const typename _proto_TypeTraits::Repeated::RepeatedFieldType&       \
-  GetRepeatedExtension(                                                       \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id) const { \
-    return _proto_TypeTraits::GetRepeated(id.number(), _extensions_);         \
-  }                                                                           \
-                                                                              \
-  template <typename _proto_TypeTraits,                                       \
-            ::PROTOBUF_NAMESPACE_ID::internal::FieldType _field_type,         \
-            bool _is_packed>                                                  \
-  inline typename _proto_TypeTraits::Repeated::RepeatedFieldType*             \
-  MutableRepeatedExtension(                                                   \
-      const ::PROTOBUF_NAMESPACE_ID::internal::ExtensionIdentifier<           \
-          CLASSNAME, _proto_TypeTraits, _field_type, _is_packed>& id) {       \
-    return _proto_TypeTraits::MutableRepeated(id.number(), _field_type,       \
-                                              _is_packed, &_extensions_);     \
-  }
-
-}  // namespace internal
-
-// Call this function to ensure that this extensions's reflection is linked into
-// the binary:
-//
-//   google::protobuf::LinkExtensionReflection(Foo::my_extension);
-//
-// This will ensure that the following lookup will succeed:
-//
-//   DescriptorPool::generated_pool()->FindExtensionByName("Foo.my_extension");
-//
-// This is often relevant for parsing extensions in text mode.
-//
-// As a side-effect, it will also guarantee that anything else from the same
-// .proto file will also be available for lookup in the generated pool.
-//
-// This function does not actually register the extension, so it does not need
-// to be called before the lookup.  However it does need to occur in a function
-// that cannot be stripped from the binary (ie. it must be reachable from main).
-//
-// Best practice is to call this function as close as possible to where the
-// reflection is actually needed.  This function is very cheap to call, so you
-// should not need to worry about its runtime overhead except in tight loops (on
-// x86-64 it compiles into two "mov" instructions).
-template <typename ExtendeeType, typename TypeTraitsType,
-          internal::FieldType field_type, bool is_packed>
-void LinkExtensionReflection(
-    const google::protobuf::internal::ExtensionIdentifier<
-        ExtendeeType, TypeTraitsType, field_type, is_packed>& extension) {
-  internal::StrongReference(extension);
-}
-
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_EXTENSION_SET_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/extension_set_inl.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/extension_set_inl.h
deleted file mode 100644
index 074784b9..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/extension_set_inl.h
+++ /dev/null
@@ -1,276 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_EXTENSION_SET_INL_H__
-#define GOOGLE_PROTOBUF_EXTENSION_SET_INL_H__
-
-#include <google/protobuf/parse_context.h>
-#include <google/protobuf/extension_set.h>
-#include <google/protobuf/metadata_lite.h>
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-template <typename T>
-const char* ExtensionSet::ParseFieldWithExtensionInfo(
-    int number, bool was_packed_on_wire, const ExtensionInfo& extension,
-    InternalMetadata* metadata, const char* ptr, internal::ParseContext* ctx) {
-  if (was_packed_on_wire) {
-    switch (extension.type) {
-#define HANDLE_TYPE(UPPERCASE, CPP_CAMELCASE)                                \
-  case WireFormatLite::TYPE_##UPPERCASE:                                     \
-    return internal::Packed##CPP_CAMELCASE##Parser(                          \
-        MutableRawRepeatedField(number, extension.type, extension.is_packed, \
-                                extension.descriptor),                       \
-        ptr, ctx);
-      HANDLE_TYPE(INT32, Int32);
-      HANDLE_TYPE(INT64, Int64);
-      HANDLE_TYPE(UINT32, UInt32);
-      HANDLE_TYPE(UINT64, UInt64);
-      HANDLE_TYPE(SINT32, SInt32);
-      HANDLE_TYPE(SINT64, SInt64);
-      HANDLE_TYPE(FIXED32, Fixed32);
-      HANDLE_TYPE(FIXED64, Fixed64);
-      HANDLE_TYPE(SFIXED32, SFixed32);
-      HANDLE_TYPE(SFIXED64, SFixed64);
-      HANDLE_TYPE(FLOAT, Float);
-      HANDLE_TYPE(DOUBLE, Double);
-      HANDLE_TYPE(BOOL, Bool);
-#undef HANDLE_TYPE
-
-      case WireFormatLite::TYPE_ENUM:
-        return internal::PackedEnumParserArg<T>(
-            MutableRawRepeatedField(number, extension.type, extension.is_packed,
-                                    extension.descriptor),
-            ptr, ctx, extension.enum_validity_check.func,
-            extension.enum_validity_check.arg, metadata, number);
-      case WireFormatLite::TYPE_STRING:
-      case WireFormatLite::TYPE_BYTES:
-      case WireFormatLite::TYPE_GROUP:
-      case WireFormatLite::TYPE_MESSAGE:
-        GOOGLE_LOG(FATAL) << "Non-primitive types can't be packed.";
-        break;
-    }
-  } else {
-    switch (extension.type) {
-#define HANDLE_VARINT_TYPE(UPPERCASE, CPP_CAMELCASE)                        \
-  case WireFormatLite::TYPE_##UPPERCASE: {                                  \
-    uint64 value;                                                           \
-    ptr = VarintParse(ptr, &value);                                         \
-    GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);                                    \
-    if (extension.is_repeated) {                                            \
-      Add##CPP_CAMELCASE(number, WireFormatLite::TYPE_##UPPERCASE,          \
-                         extension.is_packed, value, extension.descriptor); \
-    } else {                                                                \
-      Set##CPP_CAMELCASE(number, WireFormatLite::TYPE_##UPPERCASE, value,   \
-                         extension.descriptor);                             \
-    }                                                                       \
-  } break
-
-      HANDLE_VARINT_TYPE(INT32, Int32);
-      HANDLE_VARINT_TYPE(INT64, Int64);
-      HANDLE_VARINT_TYPE(UINT32, UInt32);
-      HANDLE_VARINT_TYPE(UINT64, UInt64);
-      HANDLE_VARINT_TYPE(BOOL, Bool);
-#undef HANDLE_VARINT_TYPE
-#define HANDLE_SVARINT_TYPE(UPPERCASE, CPP_CAMELCASE, SIZE)                 \
-  case WireFormatLite::TYPE_##UPPERCASE: {                                  \
-    uint64 val;                                                             \
-    ptr = VarintParse(ptr, &val);                                           \
-    GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);                                    \
-    auto value = WireFormatLite::ZigZagDecode##SIZE(val);                   \
-    if (extension.is_repeated) {                                            \
-      Add##CPP_CAMELCASE(number, WireFormatLite::TYPE_##UPPERCASE,          \
-                         extension.is_packed, value, extension.descriptor); \
-    } else {                                                                \
-      Set##CPP_CAMELCASE(number, WireFormatLite::TYPE_##UPPERCASE, value,   \
-                         extension.descriptor);                             \
-    }                                                                       \
-  } break
-
-      HANDLE_SVARINT_TYPE(SINT32, Int32, 32);
-      HANDLE_SVARINT_TYPE(SINT64, Int64, 64);
-#undef HANDLE_SVARINT_TYPE
-#define HANDLE_FIXED_TYPE(UPPERCASE, CPP_CAMELCASE, CPPTYPE)                \
-  case WireFormatLite::TYPE_##UPPERCASE: {                                  \
-    auto value = UnalignedLoad<CPPTYPE>(ptr);                               \
-    ptr += sizeof(CPPTYPE);                                                 \
-    if (extension.is_repeated) {                                            \
-      Add##CPP_CAMELCASE(number, WireFormatLite::TYPE_##UPPERCASE,          \
-                         extension.is_packed, value, extension.descriptor); \
-    } else {                                                                \
-      Set##CPP_CAMELCASE(number, WireFormatLite::TYPE_##UPPERCASE, value,   \
-                         extension.descriptor);                             \
-    }                                                                       \
-  } break
-
-      HANDLE_FIXED_TYPE(FIXED32, UInt32, uint32);
-      HANDLE_FIXED_TYPE(FIXED64, UInt64, uint64);
-      HANDLE_FIXED_TYPE(SFIXED32, Int32, int32);
-      HANDLE_FIXED_TYPE(SFIXED64, Int64, int64);
-      HANDLE_FIXED_TYPE(FLOAT, Float, float);
-      HANDLE_FIXED_TYPE(DOUBLE, Double, double);
-#undef HANDLE_FIXED_TYPE
-
-      case WireFormatLite::TYPE_ENUM: {
-        uint64 val;
-        ptr = VarintParse(ptr, &val);
-        GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-        int value = val;
-
-        if (!extension.enum_validity_check.func(
-                extension.enum_validity_check.arg, value)) {
-          WriteVarint(number, val, metadata->mutable_unknown_fields<T>());
-        } else if (extension.is_repeated) {
-          AddEnum(number, WireFormatLite::TYPE_ENUM, extension.is_packed, value,
-                  extension.descriptor);
-        } else {
-          SetEnum(number, WireFormatLite::TYPE_ENUM, value,
-                  extension.descriptor);
-        }
-        break;
-      }
-
-      case WireFormatLite::TYPE_BYTES:
-      case WireFormatLite::TYPE_STRING: {
-        std::string* value =
-            extension.is_repeated
-                ? AddString(number, WireFormatLite::TYPE_STRING,
-                            extension.descriptor)
-                : MutableString(number, WireFormatLite::TYPE_STRING,
-                                extension.descriptor);
-        int size = ReadSize(&ptr);
-        GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-        return ctx->ReadString(ptr, size, value);
-      }
-
-      case WireFormatLite::TYPE_GROUP: {
-        MessageLite* value =
-            extension.is_repeated
-                ? AddMessage(number, WireFormatLite::TYPE_GROUP,
-                             *extension.message_info.prototype,
-                             extension.descriptor)
-                : MutableMessage(number, WireFormatLite::TYPE_GROUP,
-                                 *extension.message_info.prototype,
-                                 extension.descriptor);
-        uint32 tag = (number << 3) + WireFormatLite::WIRETYPE_START_GROUP;
-        return ctx->ParseGroup(value, ptr, tag);
-      }
-
-      case WireFormatLite::TYPE_MESSAGE: {
-        MessageLite* value =
-            extension.is_repeated
-                ? AddMessage(number, WireFormatLite::TYPE_MESSAGE,
-                             *extension.message_info.prototype,
-                             extension.descriptor)
-                : MutableMessage(number, WireFormatLite::TYPE_MESSAGE,
-                                 *extension.message_info.prototype,
-                                 extension.descriptor);
-        return ctx->ParseMessage(value, ptr);
-      }
-    }
-  }
-  return ptr;
-}
-
-template <typename Msg, typename T>
-const char* ExtensionSet::ParseMessageSetItemTmpl(
-    const char* ptr, const Msg* containing_type,
-    internal::InternalMetadata* metadata, internal::ParseContext* ctx) {
-  std::string payload;
-  uint32 type_id = 0;
-  bool payload_read = false;
-  while (!ctx->Done(&ptr)) {
-    uint32 tag = static_cast<uint8>(*ptr++);
-    if (tag == WireFormatLite::kMessageSetTypeIdTag) {
-      uint64 tmp;
-      ptr = ParseBigVarint(ptr, &tmp);
-      GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-      type_id = tmp;
-      if (payload_read) {
-        ExtensionInfo extension;
-        bool was_packed_on_wire;
-        if (!FindExtension(2, type_id, containing_type, ctx, &extension,
-                           &was_packed_on_wire)) {
-          WriteLengthDelimited(type_id, payload,
-                               metadata->mutable_unknown_fields<T>());
-        } else {
-          MessageLite* value =
-              extension.is_repeated
-                  ? AddMessage(type_id, WireFormatLite::TYPE_MESSAGE,
-                               *extension.message_info.prototype,
-                               extension.descriptor)
-                  : MutableMessage(type_id, WireFormatLite::TYPE_MESSAGE,
-                                   *extension.message_info.prototype,
-                                   extension.descriptor);
-
-          const char* p;
-          // We can't use regular parse from string as we have to track
-          // proper recursion depth and descriptor pools.
-          ParseContext tmp_ctx(ctx->depth(), false, &p, payload);
-          tmp_ctx.data().pool = ctx->data().pool;
-          tmp_ctx.data().factory = ctx->data().factory;
-          GOOGLE_PROTOBUF_PARSER_ASSERT(value->_InternalParse(p, &tmp_ctx) &&
-                                         tmp_ctx.EndedAtLimit());
-        }
-        type_id = 0;
-      }
-    } else if (tag == WireFormatLite::kMessageSetMessageTag) {
-      if (type_id != 0) {
-        ptr = ParseFieldMaybeLazily(static_cast<uint64>(type_id) * 8 + 2, ptr,
-                                    containing_type, metadata, ctx);
-        GOOGLE_PROTOBUF_PARSER_ASSERT(ptr != nullptr);
-        type_id = 0;
-      } else {
-        int32 size = ReadSize(&ptr);
-        GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-        ptr = ctx->ReadString(ptr, size, &payload);
-        GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-        payload_read = true;
-      }
-    } else {
-      ptr = ReadTag(ptr - 1, &tag);
-      if (tag == 0 || (tag & 7) == 4) {
-        ctx->SetLastTag(tag);
-        return ptr;
-      }
-      ptr = ParseField(tag, ptr, containing_type, metadata, ctx);
-      GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-    }
-  }
-  return ptr;
-}
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#endif  // GOOGLE_PROTOBUF_EXTENSION_SET_INL_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/generated_enum_reflection.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/generated_enum_reflection.h
deleted file mode 100644
index 5debc0a2..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/generated_enum_reflection.h
+++ /dev/null
@@ -1,98 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// Author: jasonh@google.com (Jason Hsueh)
-//
-// This header is logically internal, but is made public because it is used
-// from protocol-compiler-generated code, which may reside in other components.
-// It provides reflection support for generated enums, and is included in
-// generated .pb.h files and should have minimal dependencies. The methods are
-// implemented in generated_message_reflection.cc.
-
-#ifndef GOOGLE_PROTOBUF_GENERATED_ENUM_REFLECTION_H__
-#define GOOGLE_PROTOBUF_GENERATED_ENUM_REFLECTION_H__
-
-#include <string>
-
-#include <google/protobuf/generated_enum_util.h>
-#include <google/protobuf/port.h>
-#include <google/protobuf/stubs/strutil.h>
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-#include <google/protobuf/port_def.inc>
-
-namespace google {
-namespace protobuf {
-class EnumDescriptor;
-}  // namespace protobuf
-}  // namespace google
-
-namespace google {
-namespace protobuf {
-
-// Returns the EnumDescriptor for enum type E, which must be a
-// proto-declared enum type.  Code generated by the protocol compiler
-// will include specializations of this template for each enum type declared.
-template <typename E>
-const EnumDescriptor* GetEnumDescriptor();
-
-namespace internal {
-
-// Helper for EnumType_Parse functions: try to parse the string 'name' as
-// an enum name of the given type, returning true and filling in value on
-// success, or returning false and leaving value unchanged on failure.
-PROTOBUF_EXPORT bool ParseNamedEnum(const EnumDescriptor* descriptor,
-                                    ConstStringParam name, int* value);
-
-template <typename EnumType>
-bool ParseNamedEnum(const EnumDescriptor* descriptor, ConstStringParam name,
-                    EnumType* value) {
-  int tmp;
-  if (!ParseNamedEnum(descriptor, name, &tmp)) return false;
-  *value = static_cast<EnumType>(tmp);
-  return true;
-}
-
-// Just a wrapper around printing the name of a value. The main point of this
-// function is not to be inlined, so that you can do this without including
-// descriptor.h.
-PROTOBUF_EXPORT const std::string& NameOfEnum(const EnumDescriptor* descriptor,
-                                              int value);
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_GENERATED_ENUM_REFLECTION_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/generated_enum_util.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/generated_enum_util.h
deleted file mode 100644
index f1002e2d..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/generated_enum_util.h
+++ /dev/null
@@ -1,83 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_GENERATED_ENUM_UTIL_H__
-#define GOOGLE_PROTOBUF_GENERATED_ENUM_UTIL_H__
-
-#include <type_traits>
-
-#include <google/protobuf/message_lite.h>
-#include <google/protobuf/stubs/strutil.h>
-
-#include <google/protobuf/port_def.inc>
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-namespace google {
-namespace protobuf {
-
-// This type trait can be used to cause templates to only match proto2 enum
-// types.
-template <typename T>
-struct is_proto_enum : ::std::false_type {};
-
-namespace internal {
-
-// The table entry format for storing enum name-to-value mapping used with lite
-// protos. This struct and the following related functions should only be used
-// by protobuf generated code.
-struct EnumEntry {
-  StringPiece name;
-  int value;
-};
-
-// Looks up a numeric enum value given the string name.
-PROTOBUF_EXPORT bool LookUpEnumValue(const EnumEntry* enums, size_t size,
-                                     StringPiece name, int* value);
-
-// Looks up an enum name given the numeric value.
-PROTOBUF_EXPORT int LookUpEnumName(const EnumEntry* enums,
-                                   const int* sorted_indices, size_t size,
-                                   int value);
-
-// Initializes the list of enum names in std::string form.
-PROTOBUF_EXPORT bool InitializeEnumStrings(
-    const EnumEntry* enums, const int* sorted_indices, size_t size,
-    internal::ExplicitlyConstructed<std::string>* enum_strings);
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_GENERATED_ENUM_UTIL_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/generated_message_table_driven.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/generated_message_table_driven.h
deleted file mode 100644
index 731d6c52..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/generated_message_table_driven.h
+++ /dev/null
@@ -1,336 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_GENERATED_MESSAGE_TABLE_DRIVEN_H__
-#define GOOGLE_PROTOBUF_GENERATED_MESSAGE_TABLE_DRIVEN_H__
-
-#include <google/protobuf/map.h>
-#include <google/protobuf/map_entry_lite.h>
-#include <google/protobuf/map_field_lite.h>
-#include <google/protobuf/message_lite.h>
-#include <google/protobuf/wire_format_lite.h>
-
-// We require C++11 and Clang to use constexpr for variables, as GCC 4.8
-// requires constexpr to be consistent between declarations of variables
-// unnecessarily (see https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58541).
-// VS 2017 Update 3 also supports this usage of constexpr.
-#if defined(__clang__) || (defined(_MSC_VER) && _MSC_VER >= 1911)
-#define PROTOBUF_CONSTEXPR_VAR constexpr
-#else  // !__clang__
-#define PROTOBUF_CONSTEXPR_VAR
-#endif  // !_clang
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-#include <google/protobuf/port_def.inc>
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-// Processing-type masks.
-static constexpr const unsigned char kOneofMask = 0x40;
-static constexpr const unsigned char kRepeatedMask = 0x20;
-// Mask for the raw type: either a WireFormatLite::FieldType or one of the
-// ProcessingTypes below, without the oneof or repeated flag.
-static constexpr const unsigned char kTypeMask = 0x1f;
-
-// Wire type masks.
-static constexpr const unsigned char kNotPackedMask = 0x10;
-static constexpr const unsigned char kInvalidMask = 0x20;
-
-enum ProcessingTypes {
-  TYPE_STRING_CORD = 19,
-  TYPE_STRING_STRING_PIECE = 20,
-  TYPE_BYTES_CORD = 21,
-  TYPE_BYTES_STRING_PIECE = 22,
-  TYPE_MAP = 23,
-};
-
-static_assert(TYPE_MAP < kRepeatedMask, "Invalid enum");
-
-struct PROTOBUF_EXPORT FieldMetadata {
-  uint32 offset;  // offset of this field in the struct
-  uint32 tag;     // field * 8 + wire_type
-  // byte offset * 8 + bit_offset;
-  // if the high bit is set then this is the byte offset of the oneof_case
-  // for this field.
-  uint32 has_offset;
-  uint32 type;      // the type of this field.
-  const void* ptr;  // auxiliary data
-
-  // From the serializer point of view each fundamental type can occur in
-  // 4 different ways. For simplicity we treat all combinations as a cartesion
-  // product although not all combinations are allowed.
-  enum FieldTypeClass {
-    kPresence,
-    kNoPresence,
-    kRepeated,
-    kPacked,
-    kOneOf,
-    kNumTypeClasses  // must be last enum
-  };
-  // C++ protobuf has 20 fundamental types, were we added Cord and StringPiece
-  // and also distinguish the same types if they have different wire format.
-  enum {
-    kCordType = 19,
-    kStringPieceType = 20,
-    kNumTypes = 20,
-    kSpecial = kNumTypes * kNumTypeClasses,
-  };
-
-  static int CalculateType(int fundamental_type, FieldTypeClass type_class);
-};
-
-// TODO(ckennelly):  Add a static assertion to ensure that these masks do not
-// conflict with wiretypes.
-
-// ParseTableField is kept small to help simplify instructions for computing
-// offsets, as we will always need this information to parse a field.
-// Additional data, needed for some types, is stored in
-// AuxiliaryParseTableField.
-struct ParseTableField {
-  uint32 offset;
-  // The presence_index ordinarily represents a has_bit index, but for fields
-  // inside a oneof it represents the index in _oneof_case_.
-  uint32 presence_index;
-  unsigned char normal_wiretype;
-  unsigned char packed_wiretype;
-
-  // processing_type is given by:
-  //   (FieldDescriptor->type() << 1) | FieldDescriptor->is_packed()
-  unsigned char processing_type;
-
-  unsigned char tag_size;
-};
-
-struct ParseTable;
-
-union AuxiliaryParseTableField {
-  typedef bool (*EnumValidator)(int);
-
-  // Enums
-  struct enum_aux {
-    EnumValidator validator;
-  };
-  enum_aux enums;
-  // Group, messages
-  struct message_aux {
-    // ExplicitlyInitialized<T> -> T requires a reinterpret_cast, which prevents
-    // the tables from being constructed as a constexpr.  We use void to avoid
-    // the cast.
-    const void* default_message_void;
-    const MessageLite* default_message() const {
-      return static_cast<const MessageLite*>(default_message_void);
-    }
-  };
-  message_aux messages;
-  // Strings
-  struct string_aux {
-    const void* default_ptr;
-    const char* field_name;
-  };
-  string_aux strings;
-
-  struct map_aux {
-    bool (*parse_map)(io::CodedInputStream*, void*);
-  };
-  map_aux maps;
-
-  AuxiliaryParseTableField() = default;
-  constexpr AuxiliaryParseTableField(AuxiliaryParseTableField::enum_aux e)
-      : enums(e) {}
-  constexpr AuxiliaryParseTableField(AuxiliaryParseTableField::message_aux m)
-      : messages(m) {}
-  constexpr AuxiliaryParseTableField(AuxiliaryParseTableField::string_aux s)
-      : strings(s) {}
-  constexpr AuxiliaryParseTableField(AuxiliaryParseTableField::map_aux m)
-      : maps(m) {}
-};
-
-struct ParseTable {
-  const ParseTableField* fields;
-  const AuxiliaryParseTableField* aux;
-  int max_field_number;
-  // TODO(ckennelly): Do something with this padding.
-
-  // TODO(ckennelly): Vet these for sign extension.
-  int64 has_bits_offset;
-  int64 oneof_case_offset;
-  int64 extension_offset;
-  int64 arena_offset;
-
-  // ExplicitlyInitialized<T> -> T requires a reinterpret_cast, which prevents
-  // the tables from being constructed as a constexpr.  We use void to avoid
-  // the cast.
-  const void* default_instance_void;
-  const MessageLite* default_instance() const {
-    return static_cast<const MessageLite*>(default_instance_void);
-  }
-
-  bool unknown_field_set;
-};
-
-static_assert(sizeof(ParseTableField) <= 16, "ParseTableField is too large");
-// The tables must be composed of POD components to ensure link-time
-// initialization.
-static_assert(std::is_pod<ParseTableField>::value, "");
-static_assert(std::is_pod<AuxiliaryParseTableField>::value, "");
-static_assert(std::is_pod<AuxiliaryParseTableField::enum_aux>::value, "");
-static_assert(std::is_pod<AuxiliaryParseTableField::message_aux>::value, "");
-static_assert(std::is_pod<AuxiliaryParseTableField::string_aux>::value, "");
-static_assert(std::is_pod<ParseTable>::value, "");
-
-// TODO(ckennelly): Consolidate these implementations into a single one, using
-// dynamic dispatch to the appropriate unknown field handler.
-bool MergePartialFromCodedStream(MessageLite* msg, const ParseTable& table,
-                                 io::CodedInputStream* input);
-bool MergePartialFromCodedStreamLite(MessageLite* msg, const ParseTable& table,
-                                     io::CodedInputStream* input);
-
-template <typename Entry>
-bool ParseMap(io::CodedInputStream* input, void* map_field) {
-  typedef typename MapEntryToMapField<Entry>::MapFieldType MapFieldType;
-  typedef Map<typename Entry::EntryKeyType, typename Entry::EntryValueType>
-      MapType;
-  typedef typename Entry::template Parser<MapFieldType, MapType> ParserType;
-
-  ParserType parser(static_cast<MapFieldType*>(map_field));
-  return WireFormatLite::ReadMessageNoVirtual(input, &parser);
-}
-
-struct SerializationTable {
-  int num_fields;
-  const FieldMetadata* field_table;
-};
-
-PROTOBUF_EXPORT void SerializeInternal(const uint8* base,
-                                       const FieldMetadata* table,
-                                       int32 num_fields,
-                                       io::CodedOutputStream* output);
-
-inline void TableSerialize(const MessageLite& msg,
-                           const SerializationTable* table,
-                           io::CodedOutputStream* output) {
-  const FieldMetadata* field_table = table->field_table;
-  int num_fields = table->num_fields - 1;
-  const uint8* base = reinterpret_cast<const uint8*>(&msg);
-  // TODO(gerbens) This skips the first test if we could use the fast
-  // array serialization path, we should make this
-  // int cached_size =
-  //    *reinterpret_cast<const int32*>(base + field_table->offset);
-  // SerializeWithCachedSize(msg, field_table + 1, num_fields, cached_size, ...)
-  // But we keep conformance with the old way for now.
-  SerializeInternal(base, field_table + 1, num_fields, output);
-}
-
-uint8* SerializeInternalToArray(const uint8* base, const FieldMetadata* table,
-                                int32 num_fields, bool is_deterministic,
-                                uint8* buffer);
-
-inline uint8* TableSerializeToArray(const MessageLite& msg,
-                                    const SerializationTable* table,
-                                    bool is_deterministic, uint8* buffer) {
-  const uint8* base = reinterpret_cast<const uint8*>(&msg);
-  const FieldMetadata* field_table = table->field_table + 1;
-  int num_fields = table->num_fields - 1;
-  return SerializeInternalToArray(base, field_table, num_fields,
-                                  is_deterministic, buffer);
-}
-
-template <typename T>
-struct CompareHelper {
-  bool operator()(const T& a, const T& b) const { return a < b; }
-};
-
-template <>
-struct CompareHelper<ArenaStringPtr> {
-  bool operator()(const ArenaStringPtr& a, const ArenaStringPtr& b) const {
-    return a.Get() < b.Get();
-  }
-};
-
-struct CompareMapKey {
-  template <typename T>
-  bool operator()(const MapEntryHelper<T>& a,
-                  const MapEntryHelper<T>& b) const {
-    return Compare(a.key_, b.key_);
-  }
-  template <typename T>
-  bool Compare(const T& a, const T& b) const {
-    return CompareHelper<T>()(a, b);
-  }
-};
-
-template <typename MapFieldType, const SerializationTable* table>
-void MapFieldSerializer(const uint8* base, uint32 offset, uint32 tag,
-                        uint32 has_offset, io::CodedOutputStream* output) {
-  typedef MapEntryHelper<typename MapFieldType::EntryTypeTrait> Entry;
-  typedef typename MapFieldType::MapType::const_iterator Iter;
-
-  const MapFieldType& map_field =
-      *reinterpret_cast<const MapFieldType*>(base + offset);
-  const SerializationTable* t =
-      table +
-      has_offset;  // has_offset is overloaded for maps to mean table offset
-  if (!output->IsSerializationDeterministic()) {
-    for (Iter it = map_field.GetMap().begin(); it != map_field.GetMap().end();
-         ++it) {
-      Entry map_entry(*it);
-      output->WriteVarint32(tag);
-      output->WriteVarint32(map_entry._cached_size_);
-      SerializeInternal(reinterpret_cast<const uint8*>(&map_entry),
-                        t->field_table, t->num_fields, output);
-    }
-  } else {
-    std::vector<Entry> v;
-    for (Iter it = map_field.GetMap().begin(); it != map_field.GetMap().end();
-         ++it) {
-      v.push_back(Entry(*it));
-    }
-    std::sort(v.begin(), v.end(), CompareMapKey());
-    for (int i = 0; i < v.size(); i++) {
-      output->WriteVarint32(tag);
-      output->WriteVarint32(v[i]._cached_size_);
-      SerializeInternal(reinterpret_cast<const uint8*>(&v[i]), t->field_table,
-                        t->num_fields, output);
-    }
-  }
-}
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_GENERATED_MESSAGE_TABLE_DRIVEN_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/generated_message_table_driven_lite.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/generated_message_table_driven_lite.h
deleted file mode 100644
index 3c65acdf..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/generated_message_table_driven_lite.h
+++ /dev/null
@@ -1,802 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_GENERATED_MESSAGE_TABLE_DRIVEN_LITE_H__
-#define GOOGLE_PROTOBUF_GENERATED_MESSAGE_TABLE_DRIVEN_LITE_H__
-
-#include <google/protobuf/io/zero_copy_stream_impl_lite.h>
-#include <google/protobuf/extension_set.h>
-#include <google/protobuf/generated_message_table_driven.h>
-#include <google/protobuf/implicit_weak_message.h>
-#include <google/protobuf/repeated_field.h>
-#include <google/protobuf/wire_format_lite.h>
-#include <type_traits>
-
-
-#include <google/protobuf/port_def.inc>
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-
-enum StringType {
-  StringType_STRING = 0,
-};
-
-// Logically a superset of StringType, consisting of all field types that
-// require special initialization.
-enum ProcessingType {
-  ProcessingType_STRING = 0,
-  ProcessingType_CORD = 1,
-  ProcessingType_STRING_PIECE = 2,
-  ProcessingType_MESSAGE = 3,
-};
-
-enum Cardinality {
-  Cardinality_SINGULAR = 0,
-  Cardinality_REPEATED = 1,
-  Cardinality_ONEOF = 3
-};
-
-template <typename Type>
-inline Type* Raw(MessageLite* msg, int64 offset) {
-  return reinterpret_cast<Type*>(reinterpret_cast<uint8*>(msg) + offset);
-}
-
-template <typename Type>
-inline const Type* Raw(const MessageLite* msg, int64 offset) {
-  return reinterpret_cast<const Type*>(reinterpret_cast<const uint8*>(msg) +
-                                       offset);
-}
-
-inline ExtensionSet* GetExtensionSet(MessageLite* msg, int64 extension_offset) {
-  if (extension_offset == -1) {
-    return NULL;
-  }
-
-  return Raw<ExtensionSet>(msg, extension_offset);
-}
-
-template <typename Type>
-inline Type* AddField(MessageLite* msg, int64 offset) {
-  static_assert(std::is_pod<Type>::value,
-                "Do not assign");
-
-  RepeatedField<Type>* repeated = Raw<RepeatedField<Type>>(msg, offset);
-  return repeated->Add();
-}
-
-template <>
-inline std::string* AddField<std::string>(MessageLite* msg, int64 offset) {
-  RepeatedPtrField<std::string>* repeated =
-      Raw<RepeatedPtrField<std::string>>(msg, offset);
-  return repeated->Add();
-}
-
-
-template <typename Type>
-inline void AddField(MessageLite* msg, int64 offset, Type value) {
-  static_assert(std::is_pod<Type>::value,
-                "Do not assign");
-  *AddField<Type>(msg, offset) = value;
-}
-
-inline void SetBit(uint32* has_bits, uint32 has_bit_index) {
-  GOOGLE_DCHECK(has_bits != nullptr);
-
-  uint32 mask = static_cast<uint32>(1u) << (has_bit_index % 32);
-  has_bits[has_bit_index / 32u] |= mask;
-}
-
-template <typename Type>
-inline Type* MutableField(MessageLite* msg, uint32* has_bits,
-                          uint32 has_bit_index, int64 offset) {
-  SetBit(has_bits, has_bit_index);
-  return Raw<Type>(msg, offset);
-}
-
-template <typename Type>
-inline void SetField(MessageLite* msg, uint32* has_bits, uint32 has_bit_index,
-                     int64 offset, Type value) {
-  static_assert(std::is_pod<Type>::value,
-                "Do not assign");
-  *MutableField<Type>(msg, has_bits, has_bit_index, offset) = value;
-}
-
-template <typename Type>
-inline void SetOneofField(MessageLite* msg, uint32* oneof_case,
-                          uint32 oneof_case_index, int64 offset,
-                          int field_number, Type value) {
-  oneof_case[oneof_case_index] = field_number;
-  *Raw<Type>(msg, offset) = value;
-}
-
-// Clears a oneof field. The field argument should correspond to the particular
-// field that is currently set in the oneof.
-inline void ClearOneofField(const ParseTableField& field, Arena* arena,
-                            MessageLite* msg) {
-  switch (field.processing_type & kTypeMask) {
-    case WireFormatLite::TYPE_MESSAGE:
-      if (arena == NULL) {
-        delete *Raw<MessageLite*>(msg, field.offset);
-      }
-      break;
-
-    case WireFormatLite::TYPE_STRING:
-    case WireFormatLite::TYPE_BYTES:
-      Raw<ArenaStringPtr>(msg, field.offset)
-          ->Destroy(ArenaStringPtr::EmptyDefault{}, arena);
-      break;
-
-    default:
-      // No cleanup needed.
-      break;
-  }
-}
-
-// Clears and reinitializes a oneof field as necessary, in preparation for
-// parsing a new value with type field_type and field number field_number.
-//
-// Note: the oneof_case argument should point directly to the _oneof_case_
-// element corresponding to this particular oneof, not to the beginning of the
-// _oneof_case_ array.
-template <ProcessingType field_type>
-inline void ResetOneofField(const ParseTable& table, int field_number,
-                            Arena* arena, MessageLite* msg, uint32* oneof_case,
-                            int64 offset, const void* default_ptr) {
-  if (*oneof_case == field_number) {
-    // The oneof is already set to the right type, so there is no need to clear
-    // it.
-    return;
-  }
-
-  if (*oneof_case != 0) {
-    ClearOneofField(table.fields[*oneof_case], arena, msg);
-  }
-  *oneof_case = field_number;
-
-  switch (field_type) {
-    case ProcessingType_STRING:
-      Raw<ArenaStringPtr>(msg, offset)
-          ->UnsafeSetDefault(static_cast<const std::string*>(default_ptr));
-      break;
-    case ProcessingType_MESSAGE:
-      MessageLite** submessage = Raw<MessageLite*>(msg, offset);
-      const MessageLite* prototype =
-          table.aux[field_number].messages.default_message();
-      *submessage = prototype->New(arena);
-      break;
-  }
-}
-
-template <typename UnknownFieldHandler, Cardinality cardinality,
-          bool is_string_type, StringType ctype>
-static inline bool HandleString(io::CodedInputStream* input, MessageLite* msg,
-                                Arena* arena, uint32* has_bits,
-                                uint32 has_bit_index, int64 offset,
-                                const void* default_ptr,
-                                const char* field_name) {
-  StringPiece utf8_string_data;
-#ifdef GOOGLE_PROTOBUF_UTF8_VALIDATION_ENABLED
-  constexpr bool kValidateUtf8 = is_string_type;
-#else
-  constexpr bool kValidateUtf8 = false;
-#endif  // GOOGLE_PROTOBUF_UTF8_VALIDATION_ENABLED
-
-  switch (ctype) {
-    case StringType_STRING: {
-      switch (cardinality) {
-        case Cardinality_SINGULAR: {
-          ArenaStringPtr* field = MutableField<ArenaStringPtr>(
-              msg, has_bits, has_bit_index, offset);
-          std::string* value = field->MutableNoCopy(
-              static_cast<const std::string*>(default_ptr), arena);
-          if (PROTOBUF_PREDICT_FALSE(
-                  !WireFormatLite::ReadString(input, value))) {
-            return false;
-          }
-          utf8_string_data = field->Get();
-        } break;
-        case Cardinality_REPEATED: {
-          std::string* value = AddField<std::string>(msg, offset);
-          if (PROTOBUF_PREDICT_FALSE(
-                  !WireFormatLite::ReadString(input, value))) {
-            return false;
-          }
-          utf8_string_data = *value;
-        } break;
-        case Cardinality_ONEOF: {
-          ArenaStringPtr* field = Raw<ArenaStringPtr>(msg, offset);
-          std::string* value = field->MutableNoCopy(
-              static_cast<const std::string*>(default_ptr), arena);
-          if (PROTOBUF_PREDICT_FALSE(
-                  !WireFormatLite::ReadString(input, value))) {
-            return false;
-          }
-          utf8_string_data = field->Get();
-        } break;
-        default:
-          PROTOBUF_ASSUME(false);
-      }
-      break;
-    }
-    default:
-      PROTOBUF_ASSUME(false);
-  }
-
-  if (kValidateUtf8) {
-    // TODO(b/118759213): fail if proto3
-    WireFormatLite::VerifyUtf8String(utf8_string_data.data(),
-                                     utf8_string_data.length(),
-                                     WireFormatLite::PARSE, field_name);
-  }
-  return true;
-}
-
-template <typename UnknownFieldHandler, Cardinality cardinality>
-inline bool HandleEnum(const ParseTable& table, io::CodedInputStream* input,
-                       MessageLite* msg, uint32* presence,
-                       uint32 presence_index, int64 offset, uint32 tag,
-                       int field_number) {
-  int value;
-  if (PROTOBUF_PREDICT_FALSE(
-          (!WireFormatLite::ReadPrimitive<int, WireFormatLite::TYPE_ENUM>(
-              input, &value)))) {
-    return false;
-  }
-
-  AuxiliaryParseTableField::EnumValidator validator =
-      table.aux[field_number].enums.validator;
-  if (validator == nullptr || validator(value)) {
-    switch (cardinality) {
-      case Cardinality_SINGULAR:
-        SetField(msg, presence, presence_index, offset, value);
-        break;
-      case Cardinality_REPEATED:
-        AddField(msg, offset, value);
-        break;
-      case Cardinality_ONEOF:
-        ClearOneofField(table.fields[presence[presence_index]], msg->GetArena(),
-                        msg);
-        SetOneofField(msg, presence, presence_index, offset, field_number,
-                      value);
-        break;
-      default:
-        PROTOBUF_ASSUME(false);
-    }
-  } else {
-    UnknownFieldHandler::Varint(msg, table, tag, value);
-  }
-
-  return true;
-}
-
-// RepeatedMessageTypeHandler allows us to operate on RepeatedPtrField fields
-// without instantiating the specific template.
-class RepeatedMessageTypeHandler {
- public:
-  typedef MessageLite Type;
-  typedef MessageLite WeakType;
-  static Arena* GetArena(Type* t) { return t->GetArena(); }
-  static void* GetMaybeArenaPointer(Type* t) {
-    return t->GetMaybeArenaPointer();
-  }
-  static inline Type* NewFromPrototype(const Type* prototype,
-                                       Arena* arena = NULL) {
-    return prototype->New(arena);
-  }
-  static void Delete(Type* t, Arena* arena = NULL) {
-    if (arena == NULL) {
-      delete t;
-    }
-  }
-};
-
-class MergePartialFromCodedStreamHelper {
- public:
-  static MessageLite* Add(RepeatedPtrFieldBase* field,
-                          const MessageLite* prototype) {
-    return field->Add<RepeatedMessageTypeHandler>(
-        const_cast<MessageLite*>(prototype));
-  }
-};
-
-template <typename UnknownFieldHandler, uint32 kMaxTag>
-bool MergePartialFromCodedStreamInlined(MessageLite* msg,
-                                        const ParseTable& table,
-                                        io::CodedInputStream* input) {
-  // We require that has_bits are present, as to avoid having to check for them
-  // for every field.
-  //
-  // TODO(ckennelly):  Make this a compile-time parameter with templates.
-  GOOGLE_DCHECK_GE(table.has_bits_offset, 0);
-  uint32* has_bits = Raw<uint32>(msg, table.has_bits_offset);
-  GOOGLE_DCHECK(has_bits != NULL);
-
-  while (true) {
-    uint32 tag = input->ReadTagWithCutoffNoLastTag(kMaxTag).first;
-    const WireFormatLite::WireType wire_type =
-        WireFormatLite::GetTagWireType(tag);
-    const int field_number = WireFormatLite::GetTagFieldNumber(tag);
-
-    if (PROTOBUF_PREDICT_FALSE(field_number > table.max_field_number)) {
-      // check for possible extensions
-      if (UnknownFieldHandler::ParseExtension(msg, table, input, tag)) {
-        // successfully parsed
-        continue;
-      }
-
-      if (PROTOBUF_PREDICT_FALSE(
-              !UnknownFieldHandler::Skip(msg, table, input, tag))) {
-        return false;
-      }
-
-      continue;
-    }
-
-    // We implicitly verify that data points to a valid field as we check the
-    // wire types.  Entries in table.fields[i] that do not correspond to valid
-    // field numbers have their normal_wiretype and packed_wiretype fields set
-    // with the kInvalidMask value.  As wire_type cannot take on that value, we
-    // will never match.
-    const ParseTableField* data = table.fields + field_number;
-
-    // TODO(ckennelly): Avoid sign extension
-    const int64 presence_index = data->presence_index;
-    const int64 offset = data->offset;
-    const unsigned char processing_type = data->processing_type;
-
-    if (data->normal_wiretype == static_cast<unsigned char>(wire_type)) {
-      switch (processing_type) {
-#define HANDLE_TYPE(TYPE, CPPTYPE)                                             \
-  case (WireFormatLite::TYPE_##TYPE): {                                        \
-    CPPTYPE value;                                                             \
-    if (PROTOBUF_PREDICT_FALSE(                                                \
-            (!WireFormatLite::ReadPrimitive<                                   \
-                CPPTYPE, WireFormatLite::TYPE_##TYPE>(input, &value)))) {      \
-      return false;                                                            \
-    }                                                                          \
-    SetField(msg, has_bits, presence_index, offset, value);                    \
-    break;                                                                     \
-  }                                                                            \
-  case (WireFormatLite::TYPE_##TYPE) | kRepeatedMask: {                        \
-    RepeatedField<CPPTYPE>* values = Raw<RepeatedField<CPPTYPE>>(msg, offset); \
-    if (PROTOBUF_PREDICT_FALSE((!WireFormatLite::ReadRepeatedPrimitive<        \
-                                CPPTYPE, WireFormatLite::TYPE_##TYPE>(         \
-            data->tag_size, tag, input, values)))) {                           \
-      return false;                                                            \
-    }                                                                          \
-    break;                                                                     \
-  }                                                                            \
-  case (WireFormatLite::TYPE_##TYPE) | kOneofMask: {                           \
-    uint32* oneof_case = Raw<uint32>(msg, table.oneof_case_offset);            \
-    CPPTYPE value;                                                             \
-    if (PROTOBUF_PREDICT_FALSE(                                                \
-            (!WireFormatLite::ReadPrimitive<                                   \
-                CPPTYPE, WireFormatLite::TYPE_##TYPE>(input, &value)))) {      \
-      return false;                                                            \
-    }                                                                          \
-    ClearOneofField(table.fields[oneof_case[presence_index]], msg->GetArena(), \
-                    msg);                                                      \
-    SetOneofField(msg, oneof_case, presence_index, offset, field_number,       \
-                  value);                                                      \
-    break;                                                                     \
-  }
-
-        HANDLE_TYPE(INT32, int32)
-        HANDLE_TYPE(INT64, int64)
-        HANDLE_TYPE(SINT32, int32)
-        HANDLE_TYPE(SINT64, int64)
-        HANDLE_TYPE(UINT32, uint32)
-        HANDLE_TYPE(UINT64, uint64)
-
-        HANDLE_TYPE(FIXED32, uint32)
-        HANDLE_TYPE(FIXED64, uint64)
-        HANDLE_TYPE(SFIXED32, int32)
-        HANDLE_TYPE(SFIXED64, int64)
-
-        HANDLE_TYPE(FLOAT, float)
-        HANDLE_TYPE(DOUBLE, double)
-
-        HANDLE_TYPE(BOOL, bool)
-#undef HANDLE_TYPE
-        case WireFormatLite::TYPE_BYTES:
-#ifndef GOOGLE_PROTOBUF_UTF8_VALIDATION_ENABLED
-        case WireFormatLite::TYPE_STRING:
-#endif  // GOOGLE_PROTOBUF_UTF8_VALIDATION_ENABLED
-        {
-          Arena* const arena = msg->GetArena();
-          const void* default_ptr = table.aux[field_number].strings.default_ptr;
-
-          if (PROTOBUF_PREDICT_FALSE(
-                  (!HandleString<UnknownFieldHandler, Cardinality_SINGULAR,
-                                 false, StringType_STRING>(
-                      input, msg, arena, has_bits, presence_index, offset,
-                      default_ptr, nullptr)))) {
-            return false;
-          }
-          break;
-        }
-        case WireFormatLite::TYPE_BYTES | kOneofMask:
-#ifndef GOOGLE_PROTOBUF_UTF8_VALIDATION_ENABLED
-        case WireFormatLite::TYPE_STRING | kOneofMask:
-#endif  // !GOOGLE_PROTOBUF_UTF8_VALIDATION_ENABLED
-        {
-          Arena* const arena = msg->GetArena();
-          uint32* oneof_case = Raw<uint32>(msg, table.oneof_case_offset);
-          const void* default_ptr = table.aux[field_number].strings.default_ptr;
-
-          ResetOneofField<ProcessingType_STRING>(
-              table, field_number, arena, msg, oneof_case + presence_index,
-              offset, default_ptr);
-
-          if (PROTOBUF_PREDICT_FALSE(
-                  (!HandleString<UnknownFieldHandler, Cardinality_ONEOF, false,
-                                 StringType_STRING>(input, msg, arena, has_bits,
-                                                    presence_index, offset,
-                                                    default_ptr, nullptr)))) {
-            return false;
-          }
-          break;
-        }
-        case (WireFormatLite::TYPE_BYTES) | kRepeatedMask:
-#ifndef GOOGLE_PROTOBUF_UTF8_VALIDATION_ENABLED
-        case (WireFormatLite::TYPE_STRING) | kRepeatedMask:
-#endif  // !GOOGLE_PROTOBUF_UTF8_VALIDATION_ENABLED
-        {
-          Arena* const arena = msg->GetArena();
-          const void* default_ptr = table.aux[field_number].strings.default_ptr;
-
-          if (PROTOBUF_PREDICT_FALSE(
-                  (!HandleString<UnknownFieldHandler, Cardinality_REPEATED,
-                                 false, StringType_STRING>(
-                      input, msg, arena, has_bits, presence_index, offset,
-                      default_ptr, nullptr)))) {
-            return false;
-          }
-          break;
-        }
-#ifdef GOOGLE_PROTOBUF_UTF8_VALIDATION_ENABLED
-        case (WireFormatLite::TYPE_STRING): {
-          Arena* const arena = msg->GetArena();
-          const void* default_ptr = table.aux[field_number].strings.default_ptr;
-          const char* field_name = table.aux[field_number].strings.field_name;
-
-          if (PROTOBUF_PREDICT_FALSE(
-                  (!HandleString<UnknownFieldHandler, Cardinality_SINGULAR,
-                                 true, StringType_STRING>(
-                      input, msg, arena, has_bits, presence_index, offset,
-                      default_ptr, field_name)))) {
-            return false;
-          }
-          break;
-        }
-        case (WireFormatLite::TYPE_STRING) | kRepeatedMask: {
-          Arena* const arena = msg->GetArena();
-          const void* default_ptr = table.aux[field_number].strings.default_ptr;
-          const char* field_name = table.aux[field_number].strings.field_name;
-
-          if (PROTOBUF_PREDICT_FALSE(
-                  (!HandleString<UnknownFieldHandler, Cardinality_REPEATED,
-                                 true, StringType_STRING>(
-                      input, msg, arena, has_bits, presence_index, offset,
-                      default_ptr, field_name)))) {
-            return false;
-          }
-          break;
-        }
-        case (WireFormatLite::TYPE_STRING) | kOneofMask: {
-          Arena* const arena = msg->GetArena();
-          uint32* oneof_case = Raw<uint32>(msg, table.oneof_case_offset);
-          const void* default_ptr = table.aux[field_number].strings.default_ptr;
-          const char* field_name = table.aux[field_number].strings.field_name;
-
-          ResetOneofField<ProcessingType_STRING>(
-              table, field_number, arena, msg, oneof_case + presence_index,
-              offset, default_ptr);
-
-          if (PROTOBUF_PREDICT_FALSE(
-                  (!HandleString<UnknownFieldHandler, Cardinality_ONEOF, true,
-                                 StringType_STRING>(
-                      input, msg, arena, has_bits, presence_index, offset,
-                      default_ptr, field_name)))) {
-            return false;
-          }
-          break;
-        }
-#endif  // GOOGLE_PROTOBUF_UTF8_VALIDATION_ENABLED
-        case WireFormatLite::TYPE_ENUM: {
-          if (PROTOBUF_PREDICT_FALSE(
-                  (!HandleEnum<UnknownFieldHandler, Cardinality_SINGULAR>(
-                      table, input, msg, has_bits, presence_index, offset, tag,
-                      field_number)))) {
-            return false;
-          }
-          break;
-        }
-        case WireFormatLite::TYPE_ENUM | kRepeatedMask: {
-          if (PROTOBUF_PREDICT_FALSE(
-                  (!HandleEnum<UnknownFieldHandler, Cardinality_REPEATED>(
-                      table, input, msg, has_bits, presence_index, offset, tag,
-                      field_number)))) {
-            return false;
-          }
-          break;
-        }
-        case WireFormatLite::TYPE_ENUM | kOneofMask: {
-          uint32* oneof_case = Raw<uint32>(msg, table.oneof_case_offset);
-          if (PROTOBUF_PREDICT_FALSE(
-                  (!HandleEnum<UnknownFieldHandler, Cardinality_ONEOF>(
-                      table, input, msg, oneof_case, presence_index, offset,
-                      tag, field_number)))) {
-            return false;
-          }
-          break;
-        }
-        case WireFormatLite::TYPE_GROUP: {
-          MessageLite** submsg_holder =
-              MutableField<MessageLite*>(msg, has_bits, presence_index, offset);
-          MessageLite* submsg = *submsg_holder;
-
-          if (submsg == NULL) {
-            Arena* const arena = msg->GetArena();
-            const MessageLite* prototype =
-                table.aux[field_number].messages.default_message();
-            submsg = prototype->New(arena);
-            *submsg_holder = submsg;
-          }
-
-          if (PROTOBUF_PREDICT_FALSE(
-                  !WireFormatLite::ReadGroup(field_number, input, submsg))) {
-            return false;
-          }
-
-          break;
-        }
-        case WireFormatLite::TYPE_GROUP | kRepeatedMask: {
-          RepeatedPtrFieldBase* field = Raw<RepeatedPtrFieldBase>(msg, offset);
-          const MessageLite* prototype =
-              table.aux[field_number].messages.default_message();
-          GOOGLE_DCHECK(prototype != NULL);
-
-          MessageLite* submsg =
-              MergePartialFromCodedStreamHelper::Add(field, prototype);
-
-          if (PROTOBUF_PREDICT_FALSE(
-                  !WireFormatLite::ReadGroup(field_number, input, submsg))) {
-            return false;
-          }
-
-          break;
-        }
-        case WireFormatLite::TYPE_MESSAGE: {
-          MessageLite** submsg_holder =
-              MutableField<MessageLite*>(msg, has_bits, presence_index, offset);
-          MessageLite* submsg = *submsg_holder;
-
-          if (submsg == NULL) {
-            Arena* const arena = msg->GetArena();
-            const MessageLite* prototype =
-                table.aux[field_number].messages.default_message();
-            if (prototype == NULL) {
-              prototype = ImplicitWeakMessage::default_instance();
-            }
-            submsg = prototype->New(arena);
-            *submsg_holder = submsg;
-          }
-
-          if (PROTOBUF_PREDICT_FALSE(
-                  !WireFormatLite::ReadMessage(input, submsg))) {
-            return false;
-          }
-
-          break;
-        }
-        // TODO(ckennelly):  Adapt ReadMessageNoVirtualNoRecursionDepth and
-        // manage input->IncrementRecursionDepth() here.
-        case WireFormatLite::TYPE_MESSAGE | kRepeatedMask: {
-          RepeatedPtrFieldBase* field = Raw<RepeatedPtrFieldBase>(msg, offset);
-          const MessageLite* prototype =
-              table.aux[field_number].messages.default_message();
-          if (prototype == NULL) {
-            prototype = ImplicitWeakMessage::default_instance();
-          }
-
-          MessageLite* submsg =
-              MergePartialFromCodedStreamHelper::Add(field, prototype);
-
-          if (PROTOBUF_PREDICT_FALSE(
-                  !WireFormatLite::ReadMessage(input, submsg))) {
-            return false;
-          }
-
-          break;
-        }
-        case WireFormatLite::TYPE_MESSAGE | kOneofMask: {
-          Arena* const arena = msg->GetArena();
-          uint32* oneof_case = Raw<uint32>(msg, table.oneof_case_offset);
-          MessageLite** submsg_holder = Raw<MessageLite*>(msg, offset);
-          ResetOneofField<ProcessingType_MESSAGE>(
-              table, field_number, arena, msg, oneof_case + presence_index,
-              offset, NULL);
-          MessageLite* submsg = *submsg_holder;
-
-          if (PROTOBUF_PREDICT_FALSE(
-                  !WireFormatLite::ReadMessage(input, submsg))) {
-            return false;
-          }
-
-          break;
-        }
-        case TYPE_MAP: {
-          if (PROTOBUF_PREDICT_FALSE(!(*table.aux[field_number].maps.parse_map)(
-                  input, Raw<void>(msg, offset)))) {
-            return false;
-          }
-          break;
-        }
-        case 0: {
-          // Done.
-          input->SetLastTag(tag);
-          return true;
-        }
-        default:
-          PROTOBUF_ASSUME(false);
-      }
-    } else if (data->packed_wiretype == static_cast<unsigned char>(wire_type)) {
-      // Non-packable fields have their packed_wiretype masked with
-      // kNotPackedMask, which is impossible to match here.
-      GOOGLE_DCHECK(processing_type & kRepeatedMask);
-      GOOGLE_DCHECK_NE(processing_type, kRepeatedMask);
-      GOOGLE_DCHECK_EQ(0, processing_type & kOneofMask);
-
-
-      // Mask out kRepeatedMask bit, allowing the jump table to be smaller.
-      switch (static_cast<WireFormatLite::FieldType>(processing_type ^
-                                                     kRepeatedMask)) {
-#define HANDLE_PACKED_TYPE(TYPE, CPPTYPE, CPPTYPE_METHOD)                      \
-  case WireFormatLite::TYPE_##TYPE: {                                          \
-    RepeatedField<CPPTYPE>* values = Raw<RepeatedField<CPPTYPE>>(msg, offset); \
-    if (PROTOBUF_PREDICT_FALSE(                                                \
-            (!WireFormatLite::ReadPackedPrimitive<                             \
-                CPPTYPE, WireFormatLite::TYPE_##TYPE>(input, values)))) {      \
-      return false;                                                            \
-    }                                                                          \
-    break;                                                                     \
-  }
-
-        HANDLE_PACKED_TYPE(INT32, int32, Int32)
-        HANDLE_PACKED_TYPE(INT64, int64, Int64)
-        HANDLE_PACKED_TYPE(SINT32, int32, Int32)
-        HANDLE_PACKED_TYPE(SINT64, int64, Int64)
-        HANDLE_PACKED_TYPE(UINT32, uint32, UInt32)
-        HANDLE_PACKED_TYPE(UINT64, uint64, UInt64)
-
-        HANDLE_PACKED_TYPE(FIXED32, uint32, UInt32)
-        HANDLE_PACKED_TYPE(FIXED64, uint64, UInt64)
-        HANDLE_PACKED_TYPE(SFIXED32, int32, Int32)
-        HANDLE_PACKED_TYPE(SFIXED64, int64, Int64)
-
-        HANDLE_PACKED_TYPE(FLOAT, float, Float)
-        HANDLE_PACKED_TYPE(DOUBLE, double, Double)
-
-        HANDLE_PACKED_TYPE(BOOL, bool, Bool)
-#undef HANDLE_PACKED_TYPE
-        case WireFormatLite::TYPE_ENUM: {
-          // To avoid unnecessarily calling MutableUnknownFields (which mutates
-          // InternalMetadata) when all inputs in the repeated series
-          // are valid, we implement our own parser rather than call
-          // WireFormat::ReadPackedEnumPreserveUnknowns.
-          uint32 length;
-          if (PROTOBUF_PREDICT_FALSE(!input->ReadVarint32(&length))) {
-            return false;
-          }
-
-          AuxiliaryParseTableField::EnumValidator validator =
-              table.aux[field_number].enums.validator;
-          RepeatedField<int>* values = Raw<RepeatedField<int>>(msg, offset);
-
-          io::CodedInputStream::Limit limit = input->PushLimit(length);
-          while (input->BytesUntilLimit() > 0) {
-            int value;
-            if (PROTOBUF_PREDICT_FALSE(
-                    (!WireFormatLite::ReadPrimitive<
-                        int, WireFormatLite::TYPE_ENUM>(input, &value)))) {
-              return false;
-            }
-
-            if (validator == nullptr || validator(value)) {
-              values->Add(value);
-            } else {
-              // TODO(ckennelly): Consider caching here.
-              UnknownFieldHandler::Varint(msg, table, tag, value);
-            }
-          }
-          input->PopLimit(limit);
-
-          break;
-        }
-        case WireFormatLite::TYPE_STRING:
-        case WireFormatLite::TYPE_GROUP:
-        case WireFormatLite::TYPE_MESSAGE:
-        case WireFormatLite::TYPE_BYTES:
-          GOOGLE_DCHECK(false);
-          return false;
-        default:
-          PROTOBUF_ASSUME(false);
-      }
-    } else {
-      if (wire_type == WireFormatLite::WIRETYPE_END_GROUP) {
-        // Must be the end of the message.
-        input->SetLastTag(tag);
-        return true;
-      }
-
-      // check for possible extensions
-      if (UnknownFieldHandler::ParseExtension(msg, table, input, tag)) {
-        // successfully parsed
-        continue;
-      }
-
-      // process unknown field.
-      if (PROTOBUF_PREDICT_FALSE(
-              !UnknownFieldHandler::Skip(msg, table, input, tag))) {
-        return false;
-      }
-    }
-  }
-}  // NOLINT(readability/fn_size)
-
-template <typename UnknownFieldHandler>
-bool MergePartialFromCodedStreamImpl(MessageLite* msg, const ParseTable& table,
-                                     io::CodedInputStream* input) {
-  // The main beneficial cutoff values are 1 and 2 byte tags.
-  // Instantiate calls with the appropriate upper tag range
-  if (table.max_field_number <= (0x7F >> 3)) {
-    return MergePartialFromCodedStreamInlined<UnknownFieldHandler, 0x7F>(
-        msg, table, input);
-  } else if (table.max_field_number <= (0x3FFF >> 3)) {
-    return MergePartialFromCodedStreamInlined<UnknownFieldHandler, 0x3FFF>(
-        msg, table, input);
-  } else {
-    return MergePartialFromCodedStreamInlined<
-        UnknownFieldHandler, std::numeric_limits<uint32>::max()>(msg, table,
-                                                                 input);
-  }
-}
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_GENERATED_MESSAGE_TABLE_DRIVEN_LITE_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/generated_message_util.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/generated_message_util.h
deleted file mode 100644
index bae0c1f7..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/generated_message_util.h
+++ /dev/null
@@ -1,270 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// Author: kenton@google.com (Kenton Varda)
-//  Based on original Protocol Buffers design by
-//  Sanjay Ghemawat, Jeff Dean, and others.
-//
-// This file contains miscellaneous helper code used by generated code --
-// including lite types -- but which should not be used directly by users.
-
-#ifndef GOOGLE_PROTOBUF_GENERATED_MESSAGE_UTIL_H__
-#define GOOGLE_PROTOBUF_GENERATED_MESSAGE_UTIL_H__
-
-#include <assert.h>
-
-#include <atomic>
-#include <climits>
-#include <string>
-#include <vector>
-
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/any.h>
-#include <google/protobuf/has_bits.h>
-#include <google/protobuf/implicit_weak_message.h>
-#include <google/protobuf/message_lite.h>
-#include <google/protobuf/stubs/once.h>  // Add direct dep on port for pb.cc
-#include <google/protobuf/port.h>
-#include <google/protobuf/repeated_field.h>
-#include <google/protobuf/wire_format_lite.h>
-#include <google/protobuf/stubs/strutil.h>
-#include <google/protobuf/stubs/casts.h>
-
-#include <google/protobuf/port_def.inc>
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-namespace google {
-namespace protobuf {
-
-class Arena;
-class Message;
-
-namespace io {
-class CodedInputStream;
-}
-
-namespace internal {
-
-template <typename To, typename From>
-inline To DownCast(From* f) {
-  return PROTOBUF_NAMESPACE_ID::internal::down_cast<To>(f);
-}
-template <typename To, typename From>
-inline To DownCast(From& f) {
-  return PROTOBUF_NAMESPACE_ID::internal::down_cast<To>(f);
-}
-
-
-// This fastpath inlines a single branch instead of having to make the
-// InitProtobufDefaults function call.
-// It also generates less inlined code than a function-scope static initializer.
-PROTOBUF_EXPORT extern std::atomic<bool> init_protobuf_defaults_state;
-PROTOBUF_EXPORT void InitProtobufDefaultsSlow();
-PROTOBUF_EXPORT inline void InitProtobufDefaults() {
-  if (PROTOBUF_PREDICT_FALSE(
-          !init_protobuf_defaults_state.load(std::memory_order_acquire))) {
-    InitProtobufDefaultsSlow();
-  }
-}
-
-// This used by proto1
-PROTOBUF_EXPORT inline const std::string& GetEmptyString() {
-  InitProtobufDefaults();
-  return GetEmptyStringAlreadyInited();
-}
-
-
-// True if IsInitialized() is true for all elements of t.  Type is expected
-// to be a RepeatedPtrField<some message type>.  It's useful to have this
-// helper here to keep the protobuf compiler from ever having to emit loops in
-// IsInitialized() methods.  We want the C++ compiler to inline this or not
-// as it sees fit.
-template <typename Msg>
-bool AllAreInitialized(const RepeatedPtrField<Msg>& t) {
-  for (int i = t.size(); --i >= 0;) {
-    if (!t.Get(i).IsInitialized()) return false;
-  }
-  return true;
-}
-
-// "Weak" variant of AllAreInitialized, used to implement implicit weak fields.
-// This version operates on MessageLite to avoid introducing a dependency on the
-// concrete message type.
-template <class T>
-bool AllAreInitializedWeak(const RepeatedPtrField<T>& t) {
-  for (int i = t.size(); --i >= 0;) {
-    if (!reinterpret_cast<const RepeatedPtrFieldBase&>(t)
-             .Get<ImplicitWeakTypeHandler<T> >(i)
-             .IsInitialized()) {
-      return false;
-    }
-  }
-  return true;
-}
-
-inline bool IsPresent(const void* base, uint32 hasbit) {
-  const uint32* has_bits_array = static_cast<const uint32*>(base);
-  return (has_bits_array[hasbit / 32] & (1u << (hasbit & 31))) != 0;
-}
-
-inline bool IsOneofPresent(const void* base, uint32 offset, uint32 tag) {
-  const uint32* oneof =
-      reinterpret_cast<const uint32*>(static_cast<const uint8*>(base) + offset);
-  return *oneof == tag >> 3;
-}
-
-typedef void (*SpecialSerializer)(const uint8* base, uint32 offset, uint32 tag,
-                                  uint32 has_offset,
-                                  io::CodedOutputStream* output);
-
-PROTOBUF_EXPORT void ExtensionSerializer(const uint8* base, uint32 offset,
-                                         uint32 tag, uint32 has_offset,
-                                         io::CodedOutputStream* output);
-PROTOBUF_EXPORT void UnknownFieldSerializerLite(const uint8* base,
-                                                uint32 offset, uint32 tag,
-                                                uint32 has_offset,
-                                                io::CodedOutputStream* output);
-
-PROTOBUF_EXPORT MessageLite* DuplicateIfNonNullInternal(MessageLite* message);
-PROTOBUF_EXPORT MessageLite* GetOwnedMessageInternal(Arena* message_arena,
-                                                     MessageLite* submessage,
-                                                     Arena* submessage_arena);
-PROTOBUF_EXPORT void GenericSwap(MessageLite* m1, MessageLite* m2);
-// We specialize GenericSwap for non-lite messages to benefit from reflection.
-PROTOBUF_EXPORT void GenericSwap(Message* m1, Message* m2);
-
-template <typename T>
-T* DuplicateIfNonNull(T* message) {
-  // The casts must be reinterpret_cast<> because T might be a forward-declared
-  // type that the compiler doesn't know is related to MessageLite.
-  return reinterpret_cast<T*>(
-      DuplicateIfNonNullInternal(reinterpret_cast<MessageLite*>(message)));
-}
-
-template <typename T>
-T* GetOwnedMessage(Arena* message_arena, T* submessage,
-                   Arena* submessage_arena) {
-  // The casts must be reinterpret_cast<> because T might be a forward-declared
-  // type that the compiler doesn't know is related to MessageLite.
-  return reinterpret_cast<T*>(GetOwnedMessageInternal(
-      message_arena, reinterpret_cast<MessageLite*>(submessage),
-      submessage_arena));
-}
-
-// Hide atomic from the public header and allow easy change to regular int
-// on platforms where the atomic might have a perf impact.
-class PROTOBUF_EXPORT CachedSize {
- public:
-  int Get() const { return size_.load(std::memory_order_relaxed); }
-  void Set(int size) { size_.store(size, std::memory_order_relaxed); }
-
- private:
-  std::atomic<int> size_{0};
-};
-
-// SCCInfo represents information of a strongly connected component of
-// mutual dependent messages.
-struct PROTOBUF_EXPORT SCCInfoBase {
-  // We use 0 for the Initialized state, because test eax,eax, jnz is smaller
-  // and is subject to macro fusion.
-  enum {
-    kInitialized = 0,  // final state
-    kRunning = 1,
-    kUninitialized = -1,  // initial state
-  };
-#if defined(_MSC_VER) && !defined(__clang__)
-  // MSVC doesn't make std::atomic constant initialized. This union trick
-  // makes it so.
-  union {
-    int visit_status_to_make_linker_init;
-    std::atomic<int> visit_status;
-  };
-#else
-  std::atomic<int> visit_status;
-#endif
-  int num_deps;
-  int num_implicit_weak_deps;
-  void (*init_func)();
-  // This is followed by an array  of num_deps
-  // const SCCInfoBase* deps[];
-};
-
-// Zero-length arrays are a language extension available in GCC and Clang but
-// not MSVC.
-#ifdef __GNUC__
-#define PROTOBUF_ARRAY_SIZE(n) (n)
-#else
-#define PROTOBUF_ARRAY_SIZE(n) ((n) ? (n) : 1)
-#endif
-
-template <int N>
-struct SCCInfo {
-  SCCInfoBase base;
-  // Semantically this is const SCCInfo<T>* which is is a templated type.
-  // The obvious inheriting from SCCInfoBase mucks with struct initialization.
-  // Attempts showed the compiler was generating dynamic initialization code.
-  // This deps array consists of base.num_deps pointers to SCCInfoBase followed
-  // by base.num_implicit_weak_deps pointers to SCCInfoBase*. We need the extra
-  // pointer indirection for implicit weak fields. We cannot use a union type
-  // here, since that would prevent the array from being linker-initialized.
-  void* deps[PROTOBUF_ARRAY_SIZE(N)];
-};
-
-#undef PROTOBUF_ARRAY_SIZE
-
-PROTOBUF_EXPORT void InitSCCImpl(SCCInfoBase* scc);
-
-inline void InitSCC(SCCInfoBase* scc) {
-  auto status = scc->visit_status.load(std::memory_order_acquire);
-  if (PROTOBUF_PREDICT_FALSE(status != SCCInfoBase::kInitialized))
-    InitSCCImpl(scc);
-}
-
-PROTOBUF_EXPORT void DestroyMessage(const void* message);
-PROTOBUF_EXPORT void DestroyString(const void* s);
-// Destroy (not delete) the message
-inline void OnShutdownDestroyMessage(const void* ptr) {
-  OnShutdownRun(DestroyMessage, ptr);
-}
-// Destroy the string (call std::string destructor)
-inline void OnShutdownDestroyString(const std::string* ptr) {
-  OnShutdownRun(DestroyString, ptr);
-}
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_GENERATED_MESSAGE_UTIL_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/has_bits.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/has_bits.h
deleted file mode 100644
index 1144b9f4..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/has_bits.h
+++ /dev/null
@@ -1,116 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_HAS_BITS_H__
-#define GOOGLE_PROTOBUF_HAS_BITS_H__
-
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/port.h>
-
-#include <google/protobuf/port_def.inc>
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-template <size_t doublewords>
-class HasBits {
- public:
-  constexpr HasBits() PROTOBUF_ALWAYS_INLINE : has_bits_{} {}
-
-  void Clear() PROTOBUF_ALWAYS_INLINE {
-    memset(has_bits_, 0, sizeof(has_bits_));
-  }
-
-  uint32& operator[](int index) PROTOBUF_ALWAYS_INLINE {
-    return has_bits_[index];
-  }
-
-  const uint32& operator[](int index) const PROTOBUF_ALWAYS_INLINE {
-    return has_bits_[index];
-  }
-
-  bool operator==(const HasBits<doublewords>& rhs) const {
-    return memcmp(has_bits_, rhs.has_bits_, sizeof(has_bits_)) == 0;
-  }
-
-  bool operator!=(const HasBits<doublewords>& rhs) const {
-    return !(*this == rhs);
-  }
-
-  void Or(const HasBits<doublewords>& rhs) {
-    for (size_t i = 0; i < doublewords; i++) has_bits_[i] |= rhs[i];
-  }
-
-  bool empty() const;
-
- private:
-  uint32 has_bits_[doublewords];
-};
-
-template <>
-inline bool HasBits<1>::empty() const {
-  return !has_bits_[0];
-}
-
-template <>
-inline bool HasBits<2>::empty() const {
-  return !(has_bits_[0] | has_bits_[1]);
-}
-
-template <>
-inline bool HasBits<3>::empty() const {
-  return !(has_bits_[0] | has_bits_[1] | has_bits_[2]);
-}
-
-template <>
-inline bool HasBits<4>::empty() const {
-  return !(has_bits_[0] | has_bits_[1] | has_bits_[2] | has_bits_[3]);
-}
-
-template <size_t doublewords>
-inline bool HasBits<doublewords>::empty() const {
-  for (size_t i = 0; i < doublewords; ++i) {
-    if (has_bits_[i]) return false;
-  }
-  return true;
-}
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_HAS_BITS_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/implicit_weak_message.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/implicit_weak_message.h
deleted file mode 100644
index bfa6a813..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/implicit_weak_message.h
+++ /dev/null
@@ -1,190 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_IMPLICIT_WEAK_MESSAGE_H__
-#define GOOGLE_PROTOBUF_IMPLICIT_WEAK_MESSAGE_H__
-
-#include <string>
-
-#include <google/protobuf/io/coded_stream.h>
-#include <google/protobuf/arena.h>
-#include <google/protobuf/message_lite.h>
-#include <google/protobuf/repeated_field.h>
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-#include <google/protobuf/port_def.inc>
-
-// This file is logically internal-only and should only be used by protobuf
-// generated code.
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-// An implementation of MessageLite that treats all data as unknown. This type
-// acts as a placeholder for an implicit weak field in the case where the true
-// message type does not get linked into the binary.
-class PROTOBUF_EXPORT ImplicitWeakMessage : public MessageLite {
- public:
-  ImplicitWeakMessage() {}
-  explicit ImplicitWeakMessage(Arena* arena) : MessageLite(arena) {}
-
-  static const ImplicitWeakMessage* default_instance();
-
-  std::string GetTypeName() const override { return ""; }
-
-  MessageLite* New() const override { return new ImplicitWeakMessage; }
-  MessageLite* New(Arena* arena) const override {
-    return Arena::CreateMessage<ImplicitWeakMessage>(arena);
-  }
-
-  void Clear() override { data_.clear(); }
-
-  bool IsInitialized() const override { return true; }
-
-  void CheckTypeAndMergeFrom(const MessageLite& other) override {
-    data_.append(static_cast<const ImplicitWeakMessage&>(other).data_);
-  }
-
-  const char* _InternalParse(const char* ptr, ParseContext* ctx) final;
-
-  size_t ByteSizeLong() const override { return data_.size(); }
-
-  uint8* _InternalSerialize(uint8* target,
-                            io::EpsCopyOutputStream* stream) const final {
-    return stream->WriteRaw(data_.data(), static_cast<int>(data_.size()),
-                            target);
-  }
-
-  int GetCachedSize() const override { return static_cast<int>(data_.size()); }
-
-  typedef void InternalArenaConstructable_;
-
- private:
-  std::string data_;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ImplicitWeakMessage);
-};
-
-// A type handler for use with implicit weak repeated message fields.
-template <typename ImplicitWeakType>
-class ImplicitWeakTypeHandler {
- public:
-  typedef MessageLite Type;
-  static constexpr bool Moveable = false;
-
-  static inline MessageLite* NewFromPrototype(const MessageLite* prototype,
-                                              Arena* arena = NULL) {
-    return prototype->New(arena);
-  }
-
-  static inline void Delete(MessageLite* value, Arena* arena) {
-    if (arena == NULL) {
-      delete value;
-    }
-  }
-  static inline Arena* GetArena(MessageLite* value) {
-    return value->GetArena();
-  }
-  static inline void* GetMaybeArenaPointer(MessageLite* value) {
-    return value->GetArena();
-  }
-  static inline void Clear(MessageLite* value) { value->Clear(); }
-  static void Merge(const MessageLite& from, MessageLite* to) {
-    to->CheckTypeAndMergeFrom(from);
-  }
-};
-
-}  // namespace internal
-
-template <typename T>
-struct WeakRepeatedPtrField {
-  using TypeHandler = internal::ImplicitWeakTypeHandler<T>;
-  constexpr WeakRepeatedPtrField() : weak() {}
-  explicit WeakRepeatedPtrField(Arena* arena) : weak(arena) {}
-  ~WeakRepeatedPtrField() { weak.template Destroy<TypeHandler>(); }
-
-  typedef internal::RepeatedPtrIterator<MessageLite> iterator;
-  typedef internal::RepeatedPtrIterator<const MessageLite> const_iterator;
-  typedef internal::RepeatedPtrOverPtrsIterator<MessageLite*, void*>
-      pointer_iterator;
-  typedef internal::RepeatedPtrOverPtrsIterator<const MessageLite* const,
-                                                const void* const>
-      const_pointer_iterator;
-
-  iterator begin() { return iterator(base().raw_data()); }
-  const_iterator begin() const { return iterator(base().raw_data()); }
-  const_iterator cbegin() const { return begin(); }
-  iterator end() { return begin() + base().size(); }
-  const_iterator end() const { return begin() + base().size(); }
-  const_iterator cend() const { return end(); }
-  pointer_iterator pointer_begin() {
-    return pointer_iterator(base().raw_mutable_data());
-  }
-  const_pointer_iterator pointer_begin() const {
-    return const_pointer_iterator(base().raw_mutable_data());
-  }
-  pointer_iterator pointer_end() {
-    return pointer_iterator(base().raw_mutable_data() + base().size());
-  }
-  const_pointer_iterator pointer_end() const {
-    return const_pointer_iterator(base().raw_mutable_data() + base().size());
-  }
-
-  MessageLite* AddWeak(const MessageLite* prototype) {
-    return base().AddWeak(prototype);
-  }
-  T* Add() { return weak.Add(); }
-  void Clear() { base().template Clear<TypeHandler>(); }
-  void MergeFrom(const WeakRepeatedPtrField& other) {
-    base().template MergeFrom<TypeHandler>(other.base());
-  }
-  void InternalSwap(WeakRepeatedPtrField* other) {
-    base().InternalSwap(&other->base());
-  }
-
-  const internal::RepeatedPtrFieldBase& base() const { return weak; }
-  internal::RepeatedPtrFieldBase& base() { return weak; }
-  // Union disables running the destructor. Which would create a strong link.
-  // Instead we explicitly destroy the underlying base through the virtual
-  // destructor.
-  union {
-    RepeatedPtrField<T> weak;
-  };
-};
-
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_IMPLICIT_WEAK_MESSAGE_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/io/coded_stream.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/io/coded_stream.h
deleted file mode 100644
index 0fff1782..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/io/coded_stream.h
+++ /dev/null
@@ -1,1714 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// Author: kenton@google.com (Kenton Varda)
-//  Based on original Protocol Buffers design by
-//  Sanjay Ghemawat, Jeff Dean, and others.
-//
-// This file contains the CodedInputStream and CodedOutputStream classes,
-// which wrap a ZeroCopyInputStream or ZeroCopyOutputStream, respectively,
-// and allow you to read or write individual pieces of data in various
-// formats.  In particular, these implement the varint encoding for
-// integers, a simple variable-length encoding in which smaller numbers
-// take fewer bytes.
-//
-// Typically these classes will only be used internally by the protocol
-// buffer library in order to encode and decode protocol buffers.  Clients
-// of the library only need to know about this class if they wish to write
-// custom message parsing or serialization procedures.
-//
-// CodedOutputStream example:
-//   // Write some data to "myfile".  First we write a 4-byte "magic number"
-//   // to identify the file type, then write a length-delimited string.  The
-//   // string is composed of a varint giving the length followed by the raw
-//   // bytes.
-//   int fd = open("myfile", O_CREAT | O_WRONLY);
-//   ZeroCopyOutputStream* raw_output = new FileOutputStream(fd);
-//   CodedOutputStream* coded_output = new CodedOutputStream(raw_output);
-//
-//   int magic_number = 1234;
-//   char text[] = "Hello world!";
-//   coded_output->WriteLittleEndian32(magic_number);
-//   coded_output->WriteVarint32(strlen(text));
-//   coded_output->WriteRaw(text, strlen(text));
-//
-//   delete coded_output;
-//   delete raw_output;
-//   close(fd);
-//
-// CodedInputStream example:
-//   // Read a file created by the above code.
-//   int fd = open("myfile", O_RDONLY);
-//   ZeroCopyInputStream* raw_input = new FileInputStream(fd);
-//   CodedInputStream* coded_input = new CodedInputStream(raw_input);
-//
-//   coded_input->ReadLittleEndian32(&magic_number);
-//   if (magic_number != 1234) {
-//     cerr << "File not in expected format." << endl;
-//     return;
-//   }
-//
-//   uint32 size;
-//   coded_input->ReadVarint32(&size);
-//
-//   char* text = new char[size + 1];
-//   coded_input->ReadRaw(buffer, size);
-//   text[size] = '\0';
-//
-//   delete coded_input;
-//   delete raw_input;
-//   close(fd);
-//
-//   cout << "Text is: " << text << endl;
-//   delete [] text;
-//
-// For those who are interested, varint encoding is defined as follows:
-//
-// The encoding operates on unsigned integers of up to 64 bits in length.
-// Each byte of the encoded value has the format:
-// * bits 0-6: Seven bits of the number being encoded.
-// * bit 7: Zero if this is the last byte in the encoding (in which
-//   case all remaining bits of the number are zero) or 1 if
-//   more bytes follow.
-// The first byte contains the least-significant 7 bits of the number, the
-// second byte (if present) contains the next-least-significant 7 bits,
-// and so on.  So, the binary number 1011000101011 would be encoded in two
-// bytes as "10101011 00101100".
-//
-// In theory, varint could be used to encode integers of any length.
-// However, for practicality we set a limit at 64 bits.  The maximum encoded
-// length of a number is thus 10 bytes.
-
-#ifndef GOOGLE_PROTOBUF_IO_CODED_STREAM_H__
-#define GOOGLE_PROTOBUF_IO_CODED_STREAM_H__
-
-
-#include <assert.h>
-
-#include <atomic>
-#include <climits>
-#include <cstddef>
-#include <cstring>
-#include <string>
-#include <type_traits>
-#include <utility>
-
-#ifdef _MSC_VER
-// Assuming windows is always little-endian.
-#if !defined(PROTOBUF_DISABLE_LITTLE_ENDIAN_OPT_FOR_TEST)
-#define PROTOBUF_LITTLE_ENDIAN 1
-#endif
-#if _MSC_VER >= 1300 && !defined(__INTEL_COMPILER)
-// If MSVC has "/RTCc" set, it will complain about truncating casts at
-// runtime.  This file contains some intentional truncating casts.
-#pragma runtime_checks("c", off)
-#endif
-#else
-#include <sys/param.h>  // __BYTE_ORDER
-#if ((defined(__LITTLE_ENDIAN__) && !defined(__BIG_ENDIAN__)) ||    \
-     (defined(__BYTE_ORDER) && __BYTE_ORDER == __LITTLE_ENDIAN)) && \
-    !defined(PROTOBUF_DISABLE_LITTLE_ENDIAN_OPT_FOR_TEST)
-#define PROTOBUF_LITTLE_ENDIAN 1
-#endif
-#endif
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/stubs/logging.h>
-#include <google/protobuf/stubs/strutil.h>
-#include <google/protobuf/port.h>
-#include <google/protobuf/stubs/port.h>
-
-
-#include <google/protobuf/port_def.inc>
-
-namespace google {
-namespace protobuf {
-
-class DescriptorPool;
-class MessageFactory;
-class ZeroCopyCodedInputStream;
-
-namespace internal {
-void MapTestForceDeterministic();
-class EpsCopyByteStream;
-}  // namespace internal
-
-namespace io {
-
-// Defined in this file.
-class CodedInputStream;
-class CodedOutputStream;
-
-// Defined in other files.
-class ZeroCopyInputStream;   // zero_copy_stream.h
-class ZeroCopyOutputStream;  // zero_copy_stream.h
-
-// Class which reads and decodes binary data which is composed of varint-
-// encoded integers and fixed-width pieces.  Wraps a ZeroCopyInputStream.
-// Most users will not need to deal with CodedInputStream.
-//
-// Most methods of CodedInputStream that return a bool return false if an
-// underlying I/O error occurs or if the data is malformed.  Once such a
-// failure occurs, the CodedInputStream is broken and is no longer useful.
-// After a failure, callers also should assume writes to "out" args may have
-// occurred, though nothing useful can be determined from those writes.
-class PROTOBUF_EXPORT CodedInputStream {
- public:
-  // Create a CodedInputStream that reads from the given ZeroCopyInputStream.
-  explicit CodedInputStream(ZeroCopyInputStream* input);
-
-  // Create a CodedInputStream that reads from the given flat array.  This is
-  // faster than using an ArrayInputStream.  PushLimit(size) is implied by
-  // this constructor.
-  explicit CodedInputStream(const uint8* buffer, int size);
-
-  // Destroy the CodedInputStream and position the underlying
-  // ZeroCopyInputStream at the first unread byte.  If an error occurred while
-  // reading (causing a method to return false), then the exact position of
-  // the input stream may be anywhere between the last value that was read
-  // successfully and the stream's byte limit.
-  ~CodedInputStream();
-
-  // Return true if this CodedInputStream reads from a flat array instead of
-  // a ZeroCopyInputStream.
-  inline bool IsFlat() const;
-
-  // Skips a number of bytes.  Returns false if an underlying read error
-  // occurs.
-  inline bool Skip(int count);
-
-  // Sets *data to point directly at the unread part of the CodedInputStream's
-  // underlying buffer, and *size to the size of that buffer, but does not
-  // advance the stream's current position.  This will always either produce
-  // a non-empty buffer or return false.  If the caller consumes any of
-  // this data, it should then call Skip() to skip over the consumed bytes.
-  // This may be useful for implementing external fast parsing routines for
-  // types of data not covered by the CodedInputStream interface.
-  bool GetDirectBufferPointer(const void** data, int* size);
-
-  // Like GetDirectBufferPointer, but this method is inlined, and does not
-  // attempt to Refresh() if the buffer is currently empty.
-  PROTOBUF_ALWAYS_INLINE
-  void GetDirectBufferPointerInline(const void** data, int* size);
-
-  // Read raw bytes, copying them into the given buffer.
-  bool ReadRaw(void* buffer, int size);
-
-  // Like ReadRaw, but reads into a string.
-  bool ReadString(std::string* buffer, int size);
-
-
-  // Read a 32-bit little-endian integer.
-  bool ReadLittleEndian32(uint32* value);
-  // Read a 64-bit little-endian integer.
-  bool ReadLittleEndian64(uint64* value);
-
-  // These methods read from an externally provided buffer. The caller is
-  // responsible for ensuring that the buffer has sufficient space.
-  // Read a 32-bit little-endian integer.
-  static const uint8* ReadLittleEndian32FromArray(const uint8* buffer,
-                                                  uint32* value);
-  // Read a 64-bit little-endian integer.
-  static const uint8* ReadLittleEndian64FromArray(const uint8* buffer,
-                                                  uint64* value);
-
-  // Read an unsigned integer with Varint encoding, truncating to 32 bits.
-  // Reading a 32-bit value is equivalent to reading a 64-bit one and casting
-  // it to uint32, but may be more efficient.
-  bool ReadVarint32(uint32* value);
-  // Read an unsigned integer with Varint encoding.
-  bool ReadVarint64(uint64* value);
-
-  // Reads a varint off the wire into an "int". This should be used for reading
-  // sizes off the wire (sizes of strings, submessages, bytes fields, etc).
-  //
-  // The value from the wire is interpreted as unsigned.  If its value exceeds
-  // the representable value of an integer on this platform, instead of
-  // truncating we return false. Truncating (as performed by ReadVarint32()
-  // above) is an acceptable approach for fields representing an integer, but
-  // when we are parsing a size from the wire, truncating the value would result
-  // in us misparsing the payload.
-  bool ReadVarintSizeAsInt(int* value);
-
-  // Read a tag.  This calls ReadVarint32() and returns the result, or returns
-  // zero (which is not a valid tag) if ReadVarint32() fails.  Also, ReadTag
-  // (but not ReadTagNoLastTag) updates the last tag value, which can be checked
-  // with LastTagWas().
-  //
-  // Always inline because this is only called in one place per parse loop
-  // but it is called for every iteration of said loop, so it should be fast.
-  // GCC doesn't want to inline this by default.
-  PROTOBUF_ALWAYS_INLINE uint32 ReadTag() {
-    return last_tag_ = ReadTagNoLastTag();
-  }
-
-  PROTOBUF_ALWAYS_INLINE uint32 ReadTagNoLastTag();
-
-  // This usually a faster alternative to ReadTag() when cutoff is a manifest
-  // constant.  It does particularly well for cutoff >= 127.  The first part
-  // of the return value is the tag that was read, though it can also be 0 in
-  // the cases where ReadTag() would return 0.  If the second part is true
-  // then the tag is known to be in [0, cutoff].  If not, the tag either is
-  // above cutoff or is 0.  (There's intentional wiggle room when tag is 0,
-  // because that can arise in several ways, and for best performance we want
-  // to avoid an extra "is tag == 0?" check here.)
-  PROTOBUF_ALWAYS_INLINE
-  std::pair<uint32, bool> ReadTagWithCutoff(uint32 cutoff) {
-    std::pair<uint32, bool> result = ReadTagWithCutoffNoLastTag(cutoff);
-    last_tag_ = result.first;
-    return result;
-  }
-
-  PROTOBUF_ALWAYS_INLINE
-  std::pair<uint32, bool> ReadTagWithCutoffNoLastTag(uint32 cutoff);
-
-  // Usually returns true if calling ReadVarint32() now would produce the given
-  // value.  Will always return false if ReadVarint32() would not return the
-  // given value.  If ExpectTag() returns true, it also advances past
-  // the varint.  For best performance, use a compile-time constant as the
-  // parameter.
-  // Always inline because this collapses to a small number of instructions
-  // when given a constant parameter, but GCC doesn't want to inline by default.
-  PROTOBUF_ALWAYS_INLINE bool ExpectTag(uint32 expected);
-
-  // Like above, except this reads from the specified buffer. The caller is
-  // responsible for ensuring that the buffer is large enough to read a varint
-  // of the expected size. For best performance, use a compile-time constant as
-  // the expected tag parameter.
-  //
-  // Returns a pointer beyond the expected tag if it was found, or NULL if it
-  // was not.
-  PROTOBUF_ALWAYS_INLINE
-  static const uint8* ExpectTagFromArray(const uint8* buffer, uint32 expected);
-
-  // Usually returns true if no more bytes can be read.  Always returns false
-  // if more bytes can be read.  If ExpectAtEnd() returns true, a subsequent
-  // call to LastTagWas() will act as if ReadTag() had been called and returned
-  // zero, and ConsumedEntireMessage() will return true.
-  bool ExpectAtEnd();
-
-  // If the last call to ReadTag() or ReadTagWithCutoff() returned the given
-  // value, returns true.  Otherwise, returns false.
-  // ReadTagNoLastTag/ReadTagWithCutoffNoLastTag do not preserve the last
-  // returned value.
-  //
-  // This is needed because parsers for some types of embedded messages
-  // (with field type TYPE_GROUP) don't actually know that they've reached the
-  // end of a message until they see an ENDGROUP tag, which was actually part
-  // of the enclosing message.  The enclosing message would like to check that
-  // tag to make sure it had the right number, so it calls LastTagWas() on
-  // return from the embedded parser to check.
-  bool LastTagWas(uint32 expected);
-  void SetLastTag(uint32 tag) { last_tag_ = tag; }
-
-  // When parsing message (but NOT a group), this method must be called
-  // immediately after MergeFromCodedStream() returns (if it returns true)
-  // to further verify that the message ended in a legitimate way.  For
-  // example, this verifies that parsing did not end on an end-group tag.
-  // It also checks for some cases where, due to optimizations,
-  // MergeFromCodedStream() can incorrectly return true.
-  bool ConsumedEntireMessage();
-  void SetConsumed() { legitimate_message_end_ = true; }
-
-  // Limits ----------------------------------------------------------
-  // Limits are used when parsing length-delimited embedded messages.
-  // After the message's length is read, PushLimit() is used to prevent
-  // the CodedInputStream from reading beyond that length.  Once the
-  // embedded message has been parsed, PopLimit() is called to undo the
-  // limit.
-
-  // Opaque type used with PushLimit() and PopLimit().  Do not modify
-  // values of this type yourself.  The only reason that this isn't a
-  // struct with private internals is for efficiency.
-  typedef int Limit;
-
-  // Places a limit on the number of bytes that the stream may read,
-  // starting from the current position.  Once the stream hits this limit,
-  // it will act like the end of the input has been reached until PopLimit()
-  // is called.
-  //
-  // As the names imply, the stream conceptually has a stack of limits.  The
-  // shortest limit on the stack is always enforced, even if it is not the
-  // top limit.
-  //
-  // The value returned by PushLimit() is opaque to the caller, and must
-  // be passed unchanged to the corresponding call to PopLimit().
-  Limit PushLimit(int byte_limit);
-
-  // Pops the last limit pushed by PushLimit().  The input must be the value
-  // returned by that call to PushLimit().
-  void PopLimit(Limit limit);
-
-  // Returns the number of bytes left until the nearest limit on the
-  // stack is hit, or -1 if no limits are in place.
-  int BytesUntilLimit() const;
-
-  // Returns current position relative to the beginning of the input stream.
-  int CurrentPosition() const;
-
-  // Total Bytes Limit -----------------------------------------------
-  // To prevent malicious users from sending excessively large messages
-  // and causing memory exhaustion, CodedInputStream imposes a hard limit on
-  // the total number of bytes it will read.
-
-  // Sets the maximum number of bytes that this CodedInputStream will read
-  // before refusing to continue.  To prevent servers from allocating enormous
-  // amounts of memory to hold parsed messages, the maximum message length
-  // should be limited to the shortest length that will not harm usability.
-  // The default limit is INT_MAX (~2GB) and apps should set shorter limits
-  // if possible. An error will always be printed to stderr if the limit is
-  // reached.
-  //
-  // Note: setting a limit less than the current read position is interpreted
-  // as a limit on the current position.
-  //
-  // This is unrelated to PushLimit()/PopLimit().
-  void SetTotalBytesLimit(int total_bytes_limit);
-
-  PROTOBUF_DEPRECATED_MSG(
-      "Please use the single parameter version of SetTotalBytesLimit(). The "
-      "second parameter is ignored.")
-  void SetTotalBytesLimit(int total_bytes_limit, int) {
-    SetTotalBytesLimit(total_bytes_limit);
-  }
-
-  // The Total Bytes Limit minus the Current Position, or -1 if the total bytes
-  // limit is INT_MAX.
-  int BytesUntilTotalBytesLimit() const;
-
-  // Recursion Limit -------------------------------------------------
-  // To prevent corrupt or malicious messages from causing stack overflows,
-  // we must keep track of the depth of recursion when parsing embedded
-  // messages and groups.  CodedInputStream keeps track of this because it
-  // is the only object that is passed down the stack during parsing.
-
-  // Sets the maximum recursion depth.  The default is 100.
-  void SetRecursionLimit(int limit);
-  int RecursionBudget() { return recursion_budget_; }
-
-  static int GetDefaultRecursionLimit() { return default_recursion_limit_; }
-
-  // Increments the current recursion depth.  Returns true if the depth is
-  // under the limit, false if it has gone over.
-  bool IncrementRecursionDepth();
-
-  // Decrements the recursion depth if possible.
-  void DecrementRecursionDepth();
-
-  // Decrements the recursion depth blindly.  This is faster than
-  // DecrementRecursionDepth().  It should be used only if all previous
-  // increments to recursion depth were successful.
-  void UnsafeDecrementRecursionDepth();
-
-  // Shorthand for make_pair(PushLimit(byte_limit), --recursion_budget_).
-  // Using this can reduce code size and complexity in some cases.  The caller
-  // is expected to check that the second part of the result is non-negative (to
-  // bail out if the depth of recursion is too high) and, if all is well, to
-  // later pass the first part of the result to PopLimit() or similar.
-  std::pair<CodedInputStream::Limit, int> IncrementRecursionDepthAndPushLimit(
-      int byte_limit);
-
-  // Shorthand for PushLimit(ReadVarint32(&length) ? length : 0).
-  Limit ReadLengthAndPushLimit();
-
-  // Helper that is equivalent to: {
-  //  bool result = ConsumedEntireMessage();
-  //  PopLimit(limit);
-  //  UnsafeDecrementRecursionDepth();
-  //  return result; }
-  // Using this can reduce code size and complexity in some cases.
-  // Do not use unless the current recursion depth is greater than zero.
-  bool DecrementRecursionDepthAndPopLimit(Limit limit);
-
-  // Helper that is equivalent to: {
-  //  bool result = ConsumedEntireMessage();
-  //  PopLimit(limit);
-  //  return result; }
-  // Using this can reduce code size and complexity in some cases.
-  bool CheckEntireMessageConsumedAndPopLimit(Limit limit);
-
-  // Extension Registry ----------------------------------------------
-  // ADVANCED USAGE:  99.9% of people can ignore this section.
-  //
-  // By default, when parsing extensions, the parser looks for extension
-  // definitions in the pool which owns the outer message's Descriptor.
-  // However, you may call SetExtensionRegistry() to provide an alternative
-  // pool instead.  This makes it possible, for example, to parse a message
-  // using a generated class, but represent some extensions using
-  // DynamicMessage.
-
-  // Set the pool used to look up extensions.  Most users do not need to call
-  // this as the correct pool will be chosen automatically.
-  //
-  // WARNING:  It is very easy to misuse this.  Carefully read the requirements
-  //   below.  Do not use this unless you are sure you need it.  Almost no one
-  //   does.
-  //
-  // Let's say you are parsing a message into message object m, and you want
-  // to take advantage of SetExtensionRegistry().  You must follow these
-  // requirements:
-  //
-  // The given DescriptorPool must contain m->GetDescriptor().  It is not
-  // sufficient for it to simply contain a descriptor that has the same name
-  // and content -- it must be the *exact object*.  In other words:
-  //   assert(pool->FindMessageTypeByName(m->GetDescriptor()->full_name()) ==
-  //          m->GetDescriptor());
-  // There are two ways to satisfy this requirement:
-  // 1) Use m->GetDescriptor()->pool() as the pool.  This is generally useless
-  //    because this is the pool that would be used anyway if you didn't call
-  //    SetExtensionRegistry() at all.
-  // 2) Use a DescriptorPool which has m->GetDescriptor()->pool() as an
-  //    "underlay".  Read the documentation for DescriptorPool for more
-  //    information about underlays.
-  //
-  // You must also provide a MessageFactory.  This factory will be used to
-  // construct Message objects representing extensions.  The factory's
-  // GetPrototype() MUST return non-NULL for any Descriptor which can be found
-  // through the provided pool.
-  //
-  // If the provided factory might return instances of protocol-compiler-
-  // generated (i.e. compiled-in) types, or if the outer message object m is
-  // a generated type, then the given factory MUST have this property:  If
-  // GetPrototype() is given a Descriptor which resides in
-  // DescriptorPool::generated_pool(), the factory MUST return the same
-  // prototype which MessageFactory::generated_factory() would return.  That
-  // is, given a descriptor for a generated type, the factory must return an
-  // instance of the generated class (NOT DynamicMessage).  However, when
-  // given a descriptor for a type that is NOT in generated_pool, the factory
-  // is free to return any implementation.
-  //
-  // The reason for this requirement is that generated sub-objects may be
-  // accessed via the standard (non-reflection) extension accessor methods,
-  // and these methods will down-cast the object to the generated class type.
-  // If the object is not actually of that type, the results would be undefined.
-  // On the other hand, if an extension is not compiled in, then there is no
-  // way the code could end up accessing it via the standard accessors -- the
-  // only way to access the extension is via reflection.  When using reflection,
-  // DynamicMessage and generated messages are indistinguishable, so it's fine
-  // if these objects are represented using DynamicMessage.
-  //
-  // Using DynamicMessageFactory on which you have called
-  // SetDelegateToGeneratedFactory(true) should be sufficient to satisfy the
-  // above requirement.
-  //
-  // If either pool or factory is NULL, both must be NULL.
-  //
-  // Note that this feature is ignored when parsing "lite" messages as they do
-  // not have descriptors.
-  void SetExtensionRegistry(const DescriptorPool* pool,
-                            MessageFactory* factory);
-
-  // Get the DescriptorPool set via SetExtensionRegistry(), or NULL if no pool
-  // has been provided.
-  const DescriptorPool* GetExtensionPool();
-
-  // Get the MessageFactory set via SetExtensionRegistry(), or NULL if no
-  // factory has been provided.
-  MessageFactory* GetExtensionFactory();
-
- private:
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(CodedInputStream);
-
-  const uint8* buffer_;
-  const uint8* buffer_end_;  // pointer to the end of the buffer.
-  ZeroCopyInputStream* input_;
-  int total_bytes_read_;  // total bytes read from input_, including
-                          // the current buffer
-
-  // If total_bytes_read_ surpasses INT_MAX, we record the extra bytes here
-  // so that we can BackUp() on destruction.
-  int overflow_bytes_;
-
-  // LastTagWas() stuff.
-  uint32 last_tag_;  // result of last ReadTag() or ReadTagWithCutoff().
-
-  // This is set true by ReadTag{Fallback/Slow}() if it is called when exactly
-  // at EOF, or by ExpectAtEnd() when it returns true.  This happens when we
-  // reach the end of a message and attempt to read another tag.
-  bool legitimate_message_end_;
-
-  // See EnableAliasing().
-  bool aliasing_enabled_;
-
-  // Limits
-  Limit current_limit_;  // if position = -1, no limit is applied
-
-  // For simplicity, if the current buffer crosses a limit (either a normal
-  // limit created by PushLimit() or the total bytes limit), buffer_size_
-  // only tracks the number of bytes before that limit.  This field
-  // contains the number of bytes after it.  Note that this implies that if
-  // buffer_size_ == 0 and buffer_size_after_limit_ > 0, we know we've
-  // hit a limit.  However, if both are zero, it doesn't necessarily mean
-  // we aren't at a limit -- the buffer may have ended exactly at the limit.
-  int buffer_size_after_limit_;
-
-  // Maximum number of bytes to read, period.  This is unrelated to
-  // current_limit_.  Set using SetTotalBytesLimit().
-  int total_bytes_limit_;
-
-  // Current recursion budget, controlled by IncrementRecursionDepth() and
-  // similar.  Starts at recursion_limit_ and goes down: if this reaches
-  // -1 we are over budget.
-  int recursion_budget_;
-  // Recursion depth limit, set by SetRecursionLimit().
-  int recursion_limit_;
-
-  // See SetExtensionRegistry().
-  const DescriptorPool* extension_pool_;
-  MessageFactory* extension_factory_;
-
-  // Private member functions.
-
-  // Fallback when Skip() goes past the end of the current buffer.
-  bool SkipFallback(int count, int original_buffer_size);
-
-  // Advance the buffer by a given number of bytes.
-  void Advance(int amount);
-
-  // Back up input_ to the current buffer position.
-  void BackUpInputToCurrentPosition();
-
-  // Recomputes the value of buffer_size_after_limit_.  Must be called after
-  // current_limit_ or total_bytes_limit_ changes.
-  void RecomputeBufferLimits();
-
-  // Writes an error message saying that we hit total_bytes_limit_.
-  void PrintTotalBytesLimitError();
-
-  // Called when the buffer runs out to request more data.  Implies an
-  // Advance(BufferSize()).
-  bool Refresh();
-
-  // When parsing varints, we optimize for the common case of small values, and
-  // then optimize for the case when the varint fits within the current buffer
-  // piece. The Fallback method is used when we can't use the one-byte
-  // optimization. The Slow method is yet another fallback when the buffer is
-  // not large enough. Making the slow path out-of-line speeds up the common
-  // case by 10-15%. The slow path is fairly uncommon: it only triggers when a
-  // message crosses multiple buffers.  Note: ReadVarint32Fallback() and
-  // ReadVarint64Fallback() are called frequently and generally not inlined, so
-  // they have been optimized to avoid "out" parameters.  The former returns -1
-  // if it fails and the uint32 it read otherwise.  The latter has a bool
-  // indicating success or failure as part of its return type.
-  int64 ReadVarint32Fallback(uint32 first_byte_or_zero);
-  int ReadVarintSizeAsIntFallback();
-  std::pair<uint64, bool> ReadVarint64Fallback();
-  bool ReadVarint32Slow(uint32* value);
-  bool ReadVarint64Slow(uint64* value);
-  int ReadVarintSizeAsIntSlow();
-  bool ReadLittleEndian32Fallback(uint32* value);
-  bool ReadLittleEndian64Fallback(uint64* value);
-
-  // Fallback/slow methods for reading tags. These do not update last_tag_,
-  // but will set legitimate_message_end_ if we are at the end of the input
-  // stream.
-  uint32 ReadTagFallback(uint32 first_byte_or_zero);
-  uint32 ReadTagSlow();
-  bool ReadStringFallback(std::string* buffer, int size);
-
-  // Return the size of the buffer.
-  int BufferSize() const;
-
-  static const int kDefaultTotalBytesLimit = INT_MAX;
-
-  static int default_recursion_limit_;  // 100 by default.
-
-  friend class google::protobuf::ZeroCopyCodedInputStream;
-  friend class google::protobuf::internal::EpsCopyByteStream;
-};
-
-// EpsCopyOutputStream wraps a ZeroCopyOutputStream and exposes a new stream,
-// which has the property you can write kSlopBytes (16 bytes) from the current
-// position without bounds checks. The cursor into the stream is managed by
-// the user of the class and is an explicit parameter in the methods. Careful
-// use of this class, ie. keep ptr a local variable, eliminates the need to
-// for the compiler to sync the ptr value between register and memory.
-class PROTOBUF_EXPORT EpsCopyOutputStream {
- public:
-  enum { kSlopBytes = 16 };
-
-  // Initialize from a stream.
-  EpsCopyOutputStream(ZeroCopyOutputStream* stream, bool deterministic,
-                      uint8** pp)
-      : end_(buffer_),
-        stream_(stream),
-        is_serialization_deterministic_(deterministic) {
-    *pp = buffer_;
-  }
-
-  // Only for array serialization. No overflow protection, end_ will be the
-  // pointed to the end of the array. When using this the total size is already
-  // known, so no need to maintain the slop region.
-  EpsCopyOutputStream(void* data, int size, bool deterministic)
-      : end_(static_cast<uint8*>(data) + size),
-        buffer_end_(nullptr),
-        stream_(nullptr),
-        is_serialization_deterministic_(deterministic) {}
-
-  // Initialize from stream but with the first buffer already given (eager).
-  EpsCopyOutputStream(void* data, int size, ZeroCopyOutputStream* stream,
-                      bool deterministic, uint8** pp)
-      : stream_(stream), is_serialization_deterministic_(deterministic) {
-    *pp = SetInitialBuffer(data, size);
-  }
-
-  // Flush everything that's written into the underlying ZeroCopyOutputStream
-  // and trims the underlying stream to the location of ptr.
-  uint8* Trim(uint8* ptr);
-
-  // After this it's guaranteed you can safely write kSlopBytes to ptr. This
-  // will never fail! The underlying stream can produce an error. Use HadError
-  // to check for errors.
-  PROTOBUF_MUST_USE_RESULT uint8* EnsureSpace(uint8* ptr) {
-    if (PROTOBUF_PREDICT_FALSE(ptr >= end_)) {
-      return EnsureSpaceFallback(ptr);
-    }
-    return ptr;
-  }
-
-  uint8* WriteRaw(const void* data, int size, uint8* ptr) {
-    if (PROTOBUF_PREDICT_FALSE(end_ - ptr < size)) {
-      return WriteRawFallback(data, size, ptr);
-    }
-    std::memcpy(ptr, data, size);
-    return ptr + size;
-  }
-  // Writes the buffer specified by data, size to the stream. Possibly by
-  // aliasing the buffer (ie. not copying the data). The caller is responsible
-  // to make sure the buffer is alive for the duration of the
-  // ZeroCopyOutputStream.
-  uint8* WriteRawMaybeAliased(const void* data, int size, uint8* ptr) {
-    if (aliasing_enabled_) {
-      return WriteAliasedRaw(data, size, ptr);
-    } else {
-      return WriteRaw(data, size, ptr);
-    }
-  }
-
-
-  uint8* WriteStringMaybeAliased(uint32 num, const std::string& s, uint8* ptr) {
-    std::ptrdiff_t size = s.size();
-    if (PROTOBUF_PREDICT_FALSE(
-            size >= 128 || end_ - ptr + 16 - TagSize(num << 3) - 1 < size)) {
-      return WriteStringMaybeAliasedOutline(num, s, ptr);
-    }
-    ptr = UnsafeVarint((num << 3) | 2, ptr);
-    *ptr++ = static_cast<uint8>(size);
-    std::memcpy(ptr, s.data(), size);
-    return ptr + size;
-  }
-  uint8* WriteBytesMaybeAliased(uint32 num, const std::string& s, uint8* ptr) {
-    return WriteStringMaybeAliased(num, s, ptr);
-  }
-
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE uint8* WriteString(uint32 num, const T& s,
-                                            uint8* ptr) {
-    std::ptrdiff_t size = s.size();
-    if (PROTOBUF_PREDICT_FALSE(
-            size >= 128 || end_ - ptr + 16 - TagSize(num << 3) - 1 < size)) {
-      return WriteStringOutline(num, s, ptr);
-    }
-    ptr = UnsafeVarint((num << 3) | 2, ptr);
-    *ptr++ = static_cast<uint8>(size);
-    std::memcpy(ptr, s.data(), size);
-    return ptr + size;
-  }
-  template <typename T>
-  uint8* WriteBytes(uint32 num, const T& s, uint8* ptr) {
-    return WriteString(num, s, ptr);
-  }
-
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE uint8* WriteInt32Packed(int num, const T& r, int size,
-                                                 uint8* ptr) {
-    return WriteVarintPacked(num, r, size, ptr, Encode64);
-  }
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE uint8* WriteUInt32Packed(int num, const T& r, int size,
-                                                  uint8* ptr) {
-    return WriteVarintPacked(num, r, size, ptr, Encode32);
-  }
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE uint8* WriteSInt32Packed(int num, const T& r, int size,
-                                                  uint8* ptr) {
-    return WriteVarintPacked(num, r, size, ptr, ZigZagEncode32);
-  }
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE uint8* WriteInt64Packed(int num, const T& r, int size,
-                                                 uint8* ptr) {
-    return WriteVarintPacked(num, r, size, ptr, Encode64);
-  }
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE uint8* WriteUInt64Packed(int num, const T& r, int size,
-                                                  uint8* ptr) {
-    return WriteVarintPacked(num, r, size, ptr, Encode64);
-  }
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE uint8* WriteSInt64Packed(int num, const T& r, int size,
-                                                  uint8* ptr) {
-    return WriteVarintPacked(num, r, size, ptr, ZigZagEncode64);
-  }
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE uint8* WriteEnumPacked(int num, const T& r, int size,
-                                                uint8* ptr) {
-    return WriteVarintPacked(num, r, size, ptr, Encode64);
-  }
-
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE uint8* WriteFixedPacked(int num, const T& r,
-                                                 uint8* ptr) {
-    ptr = EnsureSpace(ptr);
-    constexpr auto element_size = sizeof(typename T::value_type);
-    auto size = r.size() * element_size;
-    ptr = WriteLengthDelim(num, size, ptr);
-    return WriteRawLittleEndian<element_size>(r.data(), static_cast<int>(size),
-                                              ptr);
-  }
-
-  // Returns true if there was an underlying I/O error since this object was
-  // created.
-  bool HadError() const { return had_error_; }
-
-  // Instructs the EpsCopyOutputStream to allow the underlying
-  // ZeroCopyOutputStream to hold pointers to the original structure instead of
-  // copying, if it supports it (i.e. output->AllowsAliasing() is true).  If the
-  // underlying stream does not support aliasing, then enabling it has no
-  // affect.  For now, this only affects the behavior of
-  // WriteRawMaybeAliased().
-  //
-  // NOTE: It is caller's responsibility to ensure that the chunk of memory
-  // remains live until all of the data has been consumed from the stream.
-  void EnableAliasing(bool enabled);
-
-  // See documentation on CodedOutputStream::SetSerializationDeterministic.
-  void SetSerializationDeterministic(bool value) {
-    is_serialization_deterministic_ = value;
-  }
-
-  // See documentation on CodedOutputStream::IsSerializationDeterministic.
-  bool IsSerializationDeterministic() const {
-    return is_serialization_deterministic_;
-  }
-
-  // The number of bytes written to the stream at position ptr, relative to the
-  // stream's overall position.
-  int64 ByteCount(uint8* ptr) const;
-
-
- private:
-  uint8* end_;
-  uint8* buffer_end_ = buffer_;
-  uint8 buffer_[2 * kSlopBytes];
-  ZeroCopyOutputStream* stream_;
-  bool had_error_ = false;
-  bool aliasing_enabled_ = false;  // See EnableAliasing().
-  bool is_serialization_deterministic_;
-
-  uint8* EnsureSpaceFallback(uint8* ptr);
-  inline uint8* Next();
-  int Flush(uint8* ptr);
-  std::ptrdiff_t GetSize(uint8* ptr) const {
-    GOOGLE_DCHECK(ptr <= end_ + kSlopBytes);  // NOLINT
-    return end_ + kSlopBytes - ptr;
-  }
-
-  uint8* Error() {
-    had_error_ = true;
-    // We use the patch buffer to always guarantee space to write to.
-    end_ = buffer_ + kSlopBytes;
-    return buffer_;
-  }
-
-  static constexpr int TagSize(uint32 tag) {
-    return (tag < (1 << 7))
-               ? 1
-               : (tag < (1 << 14))
-                     ? 2
-                     : (tag < (1 << 21)) ? 3 : (tag < (1 << 28)) ? 4 : 5;
-  }
-
-  PROTOBUF_ALWAYS_INLINE uint8* WriteTag(uint32 num, uint32 wt, uint8* ptr) {
-    GOOGLE_DCHECK(ptr < end_);  // NOLINT
-    return UnsafeVarint((num << 3) | wt, ptr);
-  }
-
-  PROTOBUF_ALWAYS_INLINE uint8* WriteLengthDelim(int num, uint32 size,
-                                                 uint8* ptr) {
-    ptr = WriteTag(num, 2, ptr);
-    return UnsafeWriteSize(size, ptr);
-  }
-
-  uint8* WriteRawFallback(const void* data, int size, uint8* ptr);
-
-  uint8* WriteAliasedRaw(const void* data, int size, uint8* ptr);
-
-  uint8* WriteStringMaybeAliasedOutline(uint32 num, const std::string& s,
-                                        uint8* ptr);
-  uint8* WriteStringOutline(uint32 num, const std::string& s, uint8* ptr);
-
-  template <typename T, typename E>
-  PROTOBUF_ALWAYS_INLINE uint8* WriteVarintPacked(int num, const T& r, int size,
-                                                  uint8* ptr, const E& encode) {
-    ptr = EnsureSpace(ptr);
-    ptr = WriteLengthDelim(num, size, ptr);
-    auto it = r.data();
-    auto end = it + r.size();
-    do {
-      ptr = EnsureSpace(ptr);
-      ptr = UnsafeVarint(encode(*it++), ptr);
-    } while (it < end);
-    return ptr;
-  }
-
-  static uint32 Encode32(uint32 v) { return v; }
-  static uint64 Encode64(uint64 v) { return v; }
-  static uint32 ZigZagEncode32(int32 v) {
-    return (static_cast<uint32>(v) << 1) ^ static_cast<uint32>(v >> 31);
-  }
-  static uint64 ZigZagEncode64(int64 v) {
-    return (static_cast<uint64>(v) << 1) ^ static_cast<uint64>(v >> 63);
-  }
-
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE static uint8* UnsafeVarint(T value, uint8* ptr) {
-    static_assert(std::is_unsigned<T>::value,
-                  "Varint serialization must be unsigned");
-    if (value < 0x80) {
-      ptr[0] = static_cast<uint8>(value);
-      return ptr + 1;
-    }
-    ptr[0] = static_cast<uint8>(value | 0x80);
-    value >>= 7;
-    if (value < 0x80) {
-      ptr[1] = static_cast<uint8>(value);
-      return ptr + 2;
-    }
-    ptr++;
-    do {
-      *ptr = static_cast<uint8>(value | 0x80);
-      value >>= 7;
-      ++ptr;
-    } while (PROTOBUF_PREDICT_FALSE(value >= 0x80));
-    *ptr++ = static_cast<uint8>(value);
-    return ptr;
-  }
-
-  PROTOBUF_ALWAYS_INLINE static uint8* UnsafeWriteSize(uint32 value,
-                                                       uint8* ptr) {
-    while (PROTOBUF_PREDICT_FALSE(value >= 0x80)) {
-      *ptr = static_cast<uint8>(value | 0x80);
-      value >>= 7;
-      ++ptr;
-    }
-    *ptr++ = static_cast<uint8>(value);
-    return ptr;
-  }
-
-  template <int S>
-  uint8* WriteRawLittleEndian(const void* data, int size, uint8* ptr);
-#ifndef PROTOBUF_LITTLE_ENDIAN
-  uint8* WriteRawLittleEndian32(const void* data, int size, uint8* ptr);
-  uint8* WriteRawLittleEndian64(const void* data, int size, uint8* ptr);
-#endif
-
-  // These methods are for CodedOutputStream. Ideally they should be private
-  // but to match current behavior of CodedOutputStream as close as possible
-  // we allow it some functionality.
- public:
-  uint8* SetInitialBuffer(void* data, int size) {
-    auto ptr = static_cast<uint8*>(data);
-    if (size > kSlopBytes) {
-      end_ = ptr + size - kSlopBytes;
-      buffer_end_ = nullptr;
-      return ptr;
-    } else {
-      end_ = buffer_ + size;
-      buffer_end_ = ptr;
-      return buffer_;
-    }
-  }
-
- private:
-  // Needed by CodedOutputStream HadError. HadError needs to flush the patch
-  // buffers to ensure there is no error as of yet.
-  uint8* FlushAndResetBuffer(uint8*);
-
-  // The following functions mimic the old CodedOutputStream behavior as close
-  // as possible. They flush the current state to the stream, behave as
-  // the old CodedOutputStream and then return to normal operation.
-  bool Skip(int count, uint8** pp);
-  bool GetDirectBufferPointer(void** data, int* size, uint8** pp);
-  uint8* GetDirectBufferForNBytesAndAdvance(int size, uint8** pp);
-
-  friend class CodedOutputStream;
-};
-
-template <>
-inline uint8* EpsCopyOutputStream::WriteRawLittleEndian<1>(const void* data,
-                                                           int size,
-                                                           uint8* ptr) {
-  return WriteRaw(data, size, ptr);
-}
-template <>
-inline uint8* EpsCopyOutputStream::WriteRawLittleEndian<4>(const void* data,
-                                                           int size,
-                                                           uint8* ptr) {
-#ifdef PROTOBUF_LITTLE_ENDIAN
-  return WriteRaw(data, size, ptr);
-#else
-  return WriteRawLittleEndian32(data, size, ptr);
-#endif
-}
-template <>
-inline uint8* EpsCopyOutputStream::WriteRawLittleEndian<8>(const void* data,
-                                                           int size,
-                                                           uint8* ptr) {
-#ifdef PROTOBUF_LITTLE_ENDIAN
-  return WriteRaw(data, size, ptr);
-#else
-  return WriteRawLittleEndian64(data, size, ptr);
-#endif
-}
-
-// Class which encodes and writes binary data which is composed of varint-
-// encoded integers and fixed-width pieces.  Wraps a ZeroCopyOutputStream.
-// Most users will not need to deal with CodedOutputStream.
-//
-// Most methods of CodedOutputStream which return a bool return false if an
-// underlying I/O error occurs.  Once such a failure occurs, the
-// CodedOutputStream is broken and is no longer useful. The Write* methods do
-// not return the stream status, but will invalidate the stream if an error
-// occurs. The client can probe HadError() to determine the status.
-//
-// Note that every method of CodedOutputStream which writes some data has
-// a corresponding static "ToArray" version. These versions write directly
-// to the provided buffer, returning a pointer past the last written byte.
-// They require that the buffer has sufficient capacity for the encoded data.
-// This allows an optimization where we check if an output stream has enough
-// space for an entire message before we start writing and, if there is, we
-// call only the ToArray methods to avoid doing bound checks for each
-// individual value.
-// i.e., in the example above:
-//
-//   CodedOutputStream* coded_output = new CodedOutputStream(raw_output);
-//   int magic_number = 1234;
-//   char text[] = "Hello world!";
-//
-//   int coded_size = sizeof(magic_number) +
-//                    CodedOutputStream::VarintSize32(strlen(text)) +
-//                    strlen(text);
-//
-//   uint8* buffer =
-//       coded_output->GetDirectBufferForNBytesAndAdvance(coded_size);
-//   if (buffer != nullptr) {
-//     // The output stream has enough space in the buffer: write directly to
-//     // the array.
-//     buffer = CodedOutputStream::WriteLittleEndian32ToArray(magic_number,
-//                                                            buffer);
-//     buffer = CodedOutputStream::WriteVarint32ToArray(strlen(text), buffer);
-//     buffer = CodedOutputStream::WriteRawToArray(text, strlen(text), buffer);
-//   } else {
-//     // Make bound-checked writes, which will ask the underlying stream for
-//     // more space as needed.
-//     coded_output->WriteLittleEndian32(magic_number);
-//     coded_output->WriteVarint32(strlen(text));
-//     coded_output->WriteRaw(text, strlen(text));
-//   }
-//
-//   delete coded_output;
-class PROTOBUF_EXPORT CodedOutputStream {
- public:
-  // Create an CodedOutputStream that writes to the given ZeroCopyOutputStream.
-  explicit CodedOutputStream(ZeroCopyOutputStream* stream)
-      : CodedOutputStream(stream, true) {}
-  CodedOutputStream(ZeroCopyOutputStream* stream, bool do_eager_refresh);
-
-  // Destroy the CodedOutputStream and position the underlying
-  // ZeroCopyOutputStream immediately after the last byte written.
-  ~CodedOutputStream();
-
-  // Returns true if there was an underlying I/O error since this object was
-  // created. On should call Trim before this function in order to catch all
-  // errors.
-  bool HadError() {
-    cur_ = impl_.FlushAndResetBuffer(cur_);
-    GOOGLE_DCHECK(cur_);
-    return impl_.HadError();
-  }
-
-  // Trims any unused space in the underlying buffer so that its size matches
-  // the number of bytes written by this stream. The underlying buffer will
-  // automatically be trimmed when this stream is destroyed; this call is only
-  // necessary if the underlying buffer is accessed *before* the stream is
-  // destroyed.
-  void Trim() { cur_ = impl_.Trim(cur_); }
-
-  // Skips a number of bytes, leaving the bytes unmodified in the underlying
-  // buffer.  Returns false if an underlying write error occurs.  This is
-  // mainly useful with GetDirectBufferPointer().
-  // Note of caution, the skipped bytes may contain uninitialized data. The
-  // caller must make sure that the skipped bytes are properly initialized,
-  // otherwise you might leak bytes from your heap.
-  bool Skip(int count) { return impl_.Skip(count, &cur_); }
-
-  // Sets *data to point directly at the unwritten part of the
-  // CodedOutputStream's underlying buffer, and *size to the size of that
-  // buffer, but does not advance the stream's current position.  This will
-  // always either produce a non-empty buffer or return false.  If the caller
-  // writes any data to this buffer, it should then call Skip() to skip over
-  // the consumed bytes.  This may be useful for implementing external fast
-  // serialization routines for types of data not covered by the
-  // CodedOutputStream interface.
-  bool GetDirectBufferPointer(void** data, int* size) {
-    return impl_.GetDirectBufferPointer(data, size, &cur_);
-  }
-
-  // If there are at least "size" bytes available in the current buffer,
-  // returns a pointer directly into the buffer and advances over these bytes.
-  // The caller may then write directly into this buffer (e.g. using the
-  // *ToArray static methods) rather than go through CodedOutputStream.  If
-  // there are not enough bytes available, returns NULL.  The return pointer is
-  // invalidated as soon as any other non-const method of CodedOutputStream
-  // is called.
-  inline uint8* GetDirectBufferForNBytesAndAdvance(int size) {
-    return impl_.GetDirectBufferForNBytesAndAdvance(size, &cur_);
-  }
-
-  // Write raw bytes, copying them from the given buffer.
-  void WriteRaw(const void* buffer, int size) {
-    cur_ = impl_.WriteRaw(buffer, size, cur_);
-  }
-  // Like WriteRaw()  but will try to write aliased data if aliasing is
-  // turned on.
-  void WriteRawMaybeAliased(const void* data, int size);
-  // Like WriteRaw()  but writing directly to the target array.
-  // This is _not_ inlined, as the compiler often optimizes memcpy into inline
-  // copy loops. Since this gets called by every field with string or bytes
-  // type, inlining may lead to a significant amount of code bloat, with only a
-  // minor performance gain.
-  static uint8* WriteRawToArray(const void* buffer, int size, uint8* target);
-
-  // Equivalent to WriteRaw(str.data(), str.size()).
-  void WriteString(const std::string& str);
-  // Like WriteString()  but writing directly to the target array.
-  static uint8* WriteStringToArray(const std::string& str, uint8* target);
-  // Write the varint-encoded size of str followed by str.
-  static uint8* WriteStringWithSizeToArray(const std::string& str,
-                                           uint8* target);
-
-
-  // Write a 32-bit little-endian integer.
-  void WriteLittleEndian32(uint32 value) {
-    cur_ = impl_.EnsureSpace(cur_);
-    SetCur(WriteLittleEndian32ToArray(value, Cur()));
-  }
-  // Like WriteLittleEndian32()  but writing directly to the target array.
-  static uint8* WriteLittleEndian32ToArray(uint32 value, uint8* target);
-  // Write a 64-bit little-endian integer.
-  void WriteLittleEndian64(uint64 value) {
-    cur_ = impl_.EnsureSpace(cur_);
-    SetCur(WriteLittleEndian64ToArray(value, Cur()));
-  }
-  // Like WriteLittleEndian64()  but writing directly to the target array.
-  static uint8* WriteLittleEndian64ToArray(uint64 value, uint8* target);
-
-  // Write an unsigned integer with Varint encoding.  Writing a 32-bit value
-  // is equivalent to casting it to uint64 and writing it as a 64-bit value,
-  // but may be more efficient.
-  void WriteVarint32(uint32 value);
-  // Like WriteVarint32()  but writing directly to the target array.
-  static uint8* WriteVarint32ToArray(uint32 value, uint8* target);
-  // Write an unsigned integer with Varint encoding.
-  void WriteVarint64(uint64 value);
-  // Like WriteVarint64()  but writing directly to the target array.
-  static uint8* WriteVarint64ToArray(uint64 value, uint8* target);
-
-  // Equivalent to WriteVarint32() except when the value is negative,
-  // in which case it must be sign-extended to a full 10 bytes.
-  void WriteVarint32SignExtended(int32 value);
-  // Like WriteVarint32SignExtended()  but writing directly to the target array.
-  static uint8* WriteVarint32SignExtendedToArray(int32 value, uint8* target);
-
-  // This is identical to WriteVarint32(), but optimized for writing tags.
-  // In particular, if the input is a compile-time constant, this method
-  // compiles down to a couple instructions.
-  // Always inline because otherwise the aforementioned optimization can't work,
-  // but GCC by default doesn't want to inline this.
-  void WriteTag(uint32 value);
-  // Like WriteTag()  but writing directly to the target array.
-  PROTOBUF_ALWAYS_INLINE
-  static uint8* WriteTagToArray(uint32 value, uint8* target);
-
-  // Returns the number of bytes needed to encode the given value as a varint.
-  static size_t VarintSize32(uint32 value);
-  // Returns the number of bytes needed to encode the given value as a varint.
-  static size_t VarintSize64(uint64 value);
-
-  // If negative, 10 bytes.  Otherwise, same as VarintSize32().
-  static size_t VarintSize32SignExtended(int32 value);
-
-  // Compile-time equivalent of VarintSize32().
-  template <uint32 Value>
-  struct StaticVarintSize32 {
-    static const size_t value =
-        (Value < (1 << 7))
-            ? 1
-            : (Value < (1 << 14))
-                  ? 2
-                  : (Value < (1 << 21)) ? 3 : (Value < (1 << 28)) ? 4 : 5;
-  };
-
-  // Returns the total number of bytes written since this object was created.
-  int ByteCount() const {
-    return static_cast<int>(impl_.ByteCount(cur_) - start_count_);
-  }
-
-  // Instructs the CodedOutputStream to allow the underlying
-  // ZeroCopyOutputStream to hold pointers to the original structure instead of
-  // copying, if it supports it (i.e. output->AllowsAliasing() is true).  If the
-  // underlying stream does not support aliasing, then enabling it has no
-  // affect.  For now, this only affects the behavior of
-  // WriteRawMaybeAliased().
-  //
-  // NOTE: It is caller's responsibility to ensure that the chunk of memory
-  // remains live until all of the data has been consumed from the stream.
-  void EnableAliasing(bool enabled) { impl_.EnableAliasing(enabled); }
-
-  // Indicate to the serializer whether the user wants derministic
-  // serialization. The default when this is not called comes from the global
-  // default, controlled by SetDefaultSerializationDeterministic.
-  //
-  // What deterministic serialization means is entirely up to the driver of the
-  // serialization process (i.e. the caller of methods like WriteVarint32). In
-  // the case of serializing a proto buffer message using one of the methods of
-  // MessageLite, this means that for a given binary equal messages will always
-  // be serialized to the same bytes. This implies:
-  //
-  //   * Repeated serialization of a message will return the same bytes.
-  //
-  //   * Different processes running the same binary (including on different
-  //     machines) will serialize equal messages to the same bytes.
-  //
-  // Note that this is *not* canonical across languages. It is also unstable
-  // across different builds with intervening message definition changes, due to
-  // unknown fields. Users who need canonical serialization (e.g. persistent
-  // storage in a canonical form, fingerprinting) should define their own
-  // canonicalization specification and implement the serializer using
-  // reflection APIs rather than relying on this API.
-  void SetSerializationDeterministic(bool value) {
-    impl_.SetSerializationDeterministic(value);
-  }
-
-  // Return whether the user wants deterministic serialization. See above.
-  bool IsSerializationDeterministic() const {
-    return impl_.IsSerializationDeterministic();
-  }
-
-  static bool IsDefaultSerializationDeterministic() {
-    return default_serialization_deterministic_.load(
-               std::memory_order_relaxed) != 0;
-  }
-
-  template <typename Func>
-  void Serialize(const Func& func);
-
-  uint8* Cur() const { return cur_; }
-  void SetCur(uint8* ptr) { cur_ = ptr; }
-  EpsCopyOutputStream* EpsCopy() { return &impl_; }
-
- private:
-  EpsCopyOutputStream impl_;
-  uint8* cur_;
-  int64 start_count_;
-  static std::atomic<bool> default_serialization_deterministic_;
-
-  // See above.  Other projects may use "friend" to allow them to call this.
-  // After SetDefaultSerializationDeterministic() completes, all protocol
-  // buffer serializations will be deterministic by default.  Thread safe.
-  // However, the meaning of "after" is subtle here: to be safe, each thread
-  // that wants deterministic serialization by default needs to call
-  // SetDefaultSerializationDeterministic() or ensure on its own that another
-  // thread has done so.
-  friend void internal::MapTestForceDeterministic();
-  static void SetDefaultSerializationDeterministic() {
-    default_serialization_deterministic_.store(true, std::memory_order_relaxed);
-  }
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(CodedOutputStream);
-};
-
-// inline methods ====================================================
-// The vast majority of varints are only one byte.  These inline
-// methods optimize for that case.
-
-inline bool CodedInputStream::ReadVarint32(uint32* value) {
-  uint32 v = 0;
-  if (PROTOBUF_PREDICT_TRUE(buffer_ < buffer_end_)) {
-    v = *buffer_;
-    if (v < 0x80) {
-      *value = v;
-      Advance(1);
-      return true;
-    }
-  }
-  int64 result = ReadVarint32Fallback(v);
-  *value = static_cast<uint32>(result);
-  return result >= 0;
-}
-
-inline bool CodedInputStream::ReadVarint64(uint64* value) {
-  if (PROTOBUF_PREDICT_TRUE(buffer_ < buffer_end_) && *buffer_ < 0x80) {
-    *value = *buffer_;
-    Advance(1);
-    return true;
-  }
-  std::pair<uint64, bool> p = ReadVarint64Fallback();
-  *value = p.first;
-  return p.second;
-}
-
-inline bool CodedInputStream::ReadVarintSizeAsInt(int* value) {
-  if (PROTOBUF_PREDICT_TRUE(buffer_ < buffer_end_)) {
-    int v = *buffer_;
-    if (v < 0x80) {
-      *value = v;
-      Advance(1);
-      return true;
-    }
-  }
-  *value = ReadVarintSizeAsIntFallback();
-  return *value >= 0;
-}
-
-// static
-inline const uint8* CodedInputStream::ReadLittleEndian32FromArray(
-    const uint8* buffer, uint32* value) {
-#if defined(PROTOBUF_LITTLE_ENDIAN)
-  memcpy(value, buffer, sizeof(*value));
-  return buffer + sizeof(*value);
-#else
-  *value = (static_cast<uint32>(buffer[0])) |
-           (static_cast<uint32>(buffer[1]) << 8) |
-           (static_cast<uint32>(buffer[2]) << 16) |
-           (static_cast<uint32>(buffer[3]) << 24);
-  return buffer + sizeof(*value);
-#endif
-}
-// static
-inline const uint8* CodedInputStream::ReadLittleEndian64FromArray(
-    const uint8* buffer, uint64* value) {
-#if defined(PROTOBUF_LITTLE_ENDIAN)
-  memcpy(value, buffer, sizeof(*value));
-  return buffer + sizeof(*value);
-#else
-  uint32 part0 = (static_cast<uint32>(buffer[0])) |
-                 (static_cast<uint32>(buffer[1]) << 8) |
-                 (static_cast<uint32>(buffer[2]) << 16) |
-                 (static_cast<uint32>(buffer[3]) << 24);
-  uint32 part1 = (static_cast<uint32>(buffer[4])) |
-                 (static_cast<uint32>(buffer[5]) << 8) |
-                 (static_cast<uint32>(buffer[6]) << 16) |
-                 (static_cast<uint32>(buffer[7]) << 24);
-  *value = static_cast<uint64>(part0) | (static_cast<uint64>(part1) << 32);
-  return buffer + sizeof(*value);
-#endif
-}
-
-inline bool CodedInputStream::ReadLittleEndian32(uint32* value) {
-#if defined(PROTOBUF_LITTLE_ENDIAN)
-  if (PROTOBUF_PREDICT_TRUE(BufferSize() >= static_cast<int>(sizeof(*value)))) {
-    buffer_ = ReadLittleEndian32FromArray(buffer_, value);
-    return true;
-  } else {
-    return ReadLittleEndian32Fallback(value);
-  }
-#else
-  return ReadLittleEndian32Fallback(value);
-#endif
-}
-
-inline bool CodedInputStream::ReadLittleEndian64(uint64* value) {
-#if defined(PROTOBUF_LITTLE_ENDIAN)
-  if (PROTOBUF_PREDICT_TRUE(BufferSize() >= static_cast<int>(sizeof(*value)))) {
-    buffer_ = ReadLittleEndian64FromArray(buffer_, value);
-    return true;
-  } else {
-    return ReadLittleEndian64Fallback(value);
-  }
-#else
-  return ReadLittleEndian64Fallback(value);
-#endif
-}
-
-inline uint32 CodedInputStream::ReadTagNoLastTag() {
-  uint32 v = 0;
-  if (PROTOBUF_PREDICT_TRUE(buffer_ < buffer_end_)) {
-    v = *buffer_;
-    if (v < 0x80) {
-      Advance(1);
-      return v;
-    }
-  }
-  v = ReadTagFallback(v);
-  return v;
-}
-
-inline std::pair<uint32, bool> CodedInputStream::ReadTagWithCutoffNoLastTag(
-    uint32 cutoff) {
-  // In performance-sensitive code we can expect cutoff to be a compile-time
-  // constant, and things like "cutoff >= kMax1ByteVarint" to be evaluated at
-  // compile time.
-  uint32 first_byte_or_zero = 0;
-  if (PROTOBUF_PREDICT_TRUE(buffer_ < buffer_end_)) {
-    // Hot case: buffer_ non_empty, buffer_[0] in [1, 128).
-    // TODO(gpike): Is it worth rearranging this? E.g., if the number of fields
-    // is large enough then is it better to check for the two-byte case first?
-    first_byte_or_zero = buffer_[0];
-    if (static_cast<int8>(buffer_[0]) > 0) {
-      const uint32 kMax1ByteVarint = 0x7f;
-      uint32 tag = buffer_[0];
-      Advance(1);
-      return std::make_pair(tag, cutoff >= kMax1ByteVarint || tag <= cutoff);
-    }
-    // Other hot case: cutoff >= 0x80, buffer_ has at least two bytes available,
-    // and tag is two bytes.  The latter is tested by bitwise-and-not of the
-    // first byte and the second byte.
-    if (cutoff >= 0x80 && PROTOBUF_PREDICT_TRUE(buffer_ + 1 < buffer_end_) &&
-        PROTOBUF_PREDICT_TRUE((buffer_[0] & ~buffer_[1]) >= 0x80)) {
-      const uint32 kMax2ByteVarint = (0x7f << 7) + 0x7f;
-      uint32 tag = (1u << 7) * buffer_[1] + (buffer_[0] - 0x80);
-      Advance(2);
-      // It might make sense to test for tag == 0 now, but it is so rare that
-      // that we don't bother.  A varint-encoded 0 should be one byte unless
-      // the encoder lost its mind.  The second part of the return value of
-      // this function is allowed to be either true or false if the tag is 0,
-      // so we don't have to check for tag == 0.  We may need to check whether
-      // it exceeds cutoff.
-      bool at_or_below_cutoff = cutoff >= kMax2ByteVarint || tag <= cutoff;
-      return std::make_pair(tag, at_or_below_cutoff);
-    }
-  }
-  // Slow path
-  const uint32 tag = ReadTagFallback(first_byte_or_zero);
-  return std::make_pair(tag, static_cast<uint32>(tag - 1) < cutoff);
-}
-
-inline bool CodedInputStream::LastTagWas(uint32 expected) {
-  return last_tag_ == expected;
-}
-
-inline bool CodedInputStream::ConsumedEntireMessage() {
-  return legitimate_message_end_;
-}
-
-inline bool CodedInputStream::ExpectTag(uint32 expected) {
-  if (expected < (1 << 7)) {
-    if (PROTOBUF_PREDICT_TRUE(buffer_ < buffer_end_) &&
-        buffer_[0] == expected) {
-      Advance(1);
-      return true;
-    } else {
-      return false;
-    }
-  } else if (expected < (1 << 14)) {
-    if (PROTOBUF_PREDICT_TRUE(BufferSize() >= 2) &&
-        buffer_[0] == static_cast<uint8>(expected | 0x80) &&
-        buffer_[1] == static_cast<uint8>(expected >> 7)) {
-      Advance(2);
-      return true;
-    } else {
-      return false;
-    }
-  } else {
-    // Don't bother optimizing for larger values.
-    return false;
-  }
-}
-
-inline const uint8* CodedInputStream::ExpectTagFromArray(const uint8* buffer,
-                                                         uint32 expected) {
-  if (expected < (1 << 7)) {
-    if (buffer[0] == expected) {
-      return buffer + 1;
-    }
-  } else if (expected < (1 << 14)) {
-    if (buffer[0] == static_cast<uint8>(expected | 0x80) &&
-        buffer[1] == static_cast<uint8>(expected >> 7)) {
-      return buffer + 2;
-    }
-  }
-  return nullptr;
-}
-
-inline void CodedInputStream::GetDirectBufferPointerInline(const void** data,
-                                                           int* size) {
-  *data = buffer_;
-  *size = static_cast<int>(buffer_end_ - buffer_);
-}
-
-inline bool CodedInputStream::ExpectAtEnd() {
-  // If we are at a limit we know no more bytes can be read.  Otherwise, it's
-  // hard to say without calling Refresh(), and we'd rather not do that.
-
-  if (buffer_ == buffer_end_ && ((buffer_size_after_limit_ != 0) ||
-                                 (total_bytes_read_ == current_limit_))) {
-    last_tag_ = 0;                   // Pretend we called ReadTag()...
-    legitimate_message_end_ = true;  // ... and it hit EOF.
-    return true;
-  } else {
-    return false;
-  }
-}
-
-inline int CodedInputStream::CurrentPosition() const {
-  return total_bytes_read_ - (BufferSize() + buffer_size_after_limit_);
-}
-
-inline void CodedInputStream::Advance(int amount) { buffer_ += amount; }
-
-inline void CodedInputStream::SetRecursionLimit(int limit) {
-  recursion_budget_ += limit - recursion_limit_;
-  recursion_limit_ = limit;
-}
-
-inline bool CodedInputStream::IncrementRecursionDepth() {
-  --recursion_budget_;
-  return recursion_budget_ >= 0;
-}
-
-inline void CodedInputStream::DecrementRecursionDepth() {
-  if (recursion_budget_ < recursion_limit_) ++recursion_budget_;
-}
-
-inline void CodedInputStream::UnsafeDecrementRecursionDepth() {
-  assert(recursion_budget_ < recursion_limit_);
-  ++recursion_budget_;
-}
-
-inline void CodedInputStream::SetExtensionRegistry(const DescriptorPool* pool,
-                                                   MessageFactory* factory) {
-  extension_pool_ = pool;
-  extension_factory_ = factory;
-}
-
-inline const DescriptorPool* CodedInputStream::GetExtensionPool() {
-  return extension_pool_;
-}
-
-inline MessageFactory* CodedInputStream::GetExtensionFactory() {
-  return extension_factory_;
-}
-
-inline int CodedInputStream::BufferSize() const {
-  return static_cast<int>(buffer_end_ - buffer_);
-}
-
-inline CodedInputStream::CodedInputStream(ZeroCopyInputStream* input)
-    : buffer_(nullptr),
-      buffer_end_(nullptr),
-      input_(input),
-      total_bytes_read_(0),
-      overflow_bytes_(0),
-      last_tag_(0),
-      legitimate_message_end_(false),
-      aliasing_enabled_(false),
-      current_limit_(kint32max),
-      buffer_size_after_limit_(0),
-      total_bytes_limit_(kDefaultTotalBytesLimit),
-      recursion_budget_(default_recursion_limit_),
-      recursion_limit_(default_recursion_limit_),
-      extension_pool_(nullptr),
-      extension_factory_(nullptr) {
-  // Eagerly Refresh() so buffer space is immediately available.
-  Refresh();
-}
-
-inline CodedInputStream::CodedInputStream(const uint8* buffer, int size)
-    : buffer_(buffer),
-      buffer_end_(buffer + size),
-      input_(nullptr),
-      total_bytes_read_(size),
-      overflow_bytes_(0),
-      last_tag_(0),
-      legitimate_message_end_(false),
-      aliasing_enabled_(false),
-      current_limit_(size),
-      buffer_size_after_limit_(0),
-      total_bytes_limit_(kDefaultTotalBytesLimit),
-      recursion_budget_(default_recursion_limit_),
-      recursion_limit_(default_recursion_limit_),
-      extension_pool_(nullptr),
-      extension_factory_(nullptr) {
-  // Note that setting current_limit_ == size is important to prevent some
-  // code paths from trying to access input_ and segfaulting.
-}
-
-inline bool CodedInputStream::IsFlat() const { return input_ == nullptr; }
-
-inline bool CodedInputStream::Skip(int count) {
-  if (count < 0) return false;  // security: count is often user-supplied
-
-  const int original_buffer_size = BufferSize();
-
-  if (count <= original_buffer_size) {
-    // Just skipping within the current buffer.  Easy.
-    Advance(count);
-    return true;
-  }
-
-  return SkipFallback(count, original_buffer_size);
-}
-
-inline uint8* CodedOutputStream::WriteVarint32ToArray(uint32 value,
-                                                      uint8* target) {
-  return EpsCopyOutputStream::UnsafeVarint(value, target);
-}
-
-inline uint8* CodedOutputStream::WriteVarint64ToArray(uint64 value,
-                                                      uint8* target) {
-  return EpsCopyOutputStream::UnsafeVarint(value, target);
-}
-
-inline void CodedOutputStream::WriteVarint32SignExtended(int32 value) {
-  WriteVarint64(static_cast<uint64>(value));
-}
-
-inline uint8* CodedOutputStream::WriteVarint32SignExtendedToArray(
-    int32 value, uint8* target) {
-  return WriteVarint64ToArray(static_cast<uint64>(value), target);
-}
-
-inline uint8* CodedOutputStream::WriteLittleEndian32ToArray(uint32 value,
-                                                            uint8* target) {
-#if defined(PROTOBUF_LITTLE_ENDIAN)
-  memcpy(target, &value, sizeof(value));
-#else
-  target[0] = static_cast<uint8>(value);
-  target[1] = static_cast<uint8>(value >> 8);
-  target[2] = static_cast<uint8>(value >> 16);
-  target[3] = static_cast<uint8>(value >> 24);
-#endif
-  return target + sizeof(value);
-}
-
-inline uint8* CodedOutputStream::WriteLittleEndian64ToArray(uint64 value,
-                                                            uint8* target) {
-#if defined(PROTOBUF_LITTLE_ENDIAN)
-  memcpy(target, &value, sizeof(value));
-#else
-  uint32 part0 = static_cast<uint32>(value);
-  uint32 part1 = static_cast<uint32>(value >> 32);
-
-  target[0] = static_cast<uint8>(part0);
-  target[1] = static_cast<uint8>(part0 >> 8);
-  target[2] = static_cast<uint8>(part0 >> 16);
-  target[3] = static_cast<uint8>(part0 >> 24);
-  target[4] = static_cast<uint8>(part1);
-  target[5] = static_cast<uint8>(part1 >> 8);
-  target[6] = static_cast<uint8>(part1 >> 16);
-  target[7] = static_cast<uint8>(part1 >> 24);
-#endif
-  return target + sizeof(value);
-}
-
-inline void CodedOutputStream::WriteVarint32(uint32 value) {
-  cur_ = impl_.EnsureSpace(cur_);
-  SetCur(WriteVarint32ToArray(value, Cur()));
-}
-
-inline void CodedOutputStream::WriteVarint64(uint64 value) {
-  cur_ = impl_.EnsureSpace(cur_);
-  SetCur(WriteVarint64ToArray(value, Cur()));
-}
-
-inline void CodedOutputStream::WriteTag(uint32 value) { WriteVarint32(value); }
-
-inline uint8* CodedOutputStream::WriteTagToArray(uint32 value, uint8* target) {
-  return WriteVarint32ToArray(value, target);
-}
-
-inline size_t CodedOutputStream::VarintSize32(uint32 value) {
-  // This computes value == 0 ? 1 : floor(log2(value)) / 7 + 1
-  // Use an explicit multiplication to implement the divide of
-  // a number in the 1..31 range.
-  // Explicit OR 0x1 to avoid calling Bits::Log2FloorNonZero(0), which is
-  // undefined.
-  uint32 log2value = Bits::Log2FloorNonZero(value | 0x1);
-  return static_cast<size_t>((log2value * 9 + 73) / 64);
-}
-
-inline size_t CodedOutputStream::VarintSize64(uint64 value) {
-  // This computes value == 0 ? 1 : floor(log2(value)) / 7 + 1
-  // Use an explicit multiplication to implement the divide of
-  // a number in the 1..63 range.
-  // Explicit OR 0x1 to avoid calling Bits::Log2FloorNonZero(0), which is
-  // undefined.
-  uint32 log2value = Bits::Log2FloorNonZero64(value | 0x1);
-  return static_cast<size_t>((log2value * 9 + 73) / 64);
-}
-
-inline size_t CodedOutputStream::VarintSize32SignExtended(int32 value) {
-  if (value < 0) {
-    return 10;  // TODO(kenton):  Make this a symbolic constant.
-  } else {
-    return VarintSize32(static_cast<uint32>(value));
-  }
-}
-
-inline void CodedOutputStream::WriteString(const std::string& str) {
-  WriteRaw(str.data(), static_cast<int>(str.size()));
-}
-
-inline void CodedOutputStream::WriteRawMaybeAliased(const void* data,
-                                                    int size) {
-  cur_ = impl_.WriteRawMaybeAliased(data, size, cur_);
-}
-
-inline uint8* CodedOutputStream::WriteRawToArray(const void* data, int size,
-                                                 uint8* target) {
-  memcpy(target, data, size);
-  return target + size;
-}
-
-inline uint8* CodedOutputStream::WriteStringToArray(const std::string& str,
-                                                    uint8* target) {
-  return WriteRawToArray(str.data(), static_cast<int>(str.size()), target);
-}
-
-}  // namespace io
-}  // namespace protobuf
-}  // namespace google
-
-#if defined(_MSC_VER) && _MSC_VER >= 1300 && !defined(__INTEL_COMPILER)
-#pragma runtime_checks("c", restore)
-#endif  // _MSC_VER && !defined(__INTEL_COMPILER)
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_IO_CODED_STREAM_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/io/io_win32.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/io/io_win32.h
deleted file mode 100644
index 7d11dc2b..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/io/io_win32.h
+++ /dev/null
@@ -1,139 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// Author: laszlocsomor@google.com (Laszlo Csomor)
-//  Based on original Protocol Buffers design by
-//  Sanjay Ghemawat, Jeff Dean, and others.
-
-// This file contains the declarations for Windows implementations of
-// commonly used POSIX functions such as open(2) and access(2), as well
-// as macro definitions for flags of these functions.
-//
-// By including this file you'll redefine open/access/etc. to
-// ::google::protobuf::io::win32::{open/access/etc.}.
-// Make sure you don't include a header that attempts to redeclare or
-// redefine these functions, that'll lead to confusing compilation
-// errors. It's best to #include this file as the last one to ensure that.
-//
-// This file is only used on Windows, it's empty on other platforms.
-
-#ifndef GOOGLE_PROTOBUF_IO_IO_WIN32_H__
-#define GOOGLE_PROTOBUF_IO_IO_WIN32_H__
-
-#if defined(_WIN32)
-
-#include <functional>
-#include <string>
-
-#include <google/protobuf/port.h>
-#include <google/protobuf/port_def.inc>
-
-// Compilers on Windows other than MSVC (e.g. Cygwin, MinGW32) define the
-// following functions already, except for mkdir.
-namespace google {
-namespace protobuf {
-namespace io {
-namespace win32 {
-
-PROTOBUF_EXPORT FILE* fopen(const char* path, const char* mode);
-PROTOBUF_EXPORT int access(const char* path, int mode);
-PROTOBUF_EXPORT int chdir(const char* path);
-PROTOBUF_EXPORT int close(int fd);
-PROTOBUF_EXPORT int dup(int fd);
-PROTOBUF_EXPORT int dup2(int fd1, int fd2);
-PROTOBUF_EXPORT int mkdir(const char* path, int _mode);
-PROTOBUF_EXPORT int open(const char* path, int flags, int mode = 0);
-PROTOBUF_EXPORT int read(int fd, void* buffer, size_t size);
-PROTOBUF_EXPORT int setmode(int fd, int mode);
-PROTOBUF_EXPORT int stat(const char* path, struct _stat* buffer);
-PROTOBUF_EXPORT int write(int fd, const void* buffer, size_t size);
-PROTOBUF_EXPORT std::wstring testonly_utf8_to_winpath(const char* path);
-
-enum class ExpandWildcardsResult {
-  kSuccess = 0,
-  kErrorNoMatchingFile = 1,
-  kErrorInputPathConversion = 2,
-  kErrorOutputPathConversion = 3,
-};
-
-// Expand wildcards in a path pattern, feed the result to a consumer function.
-//
-// `path` must be a valid, Windows-style path. It may be absolute, or relative
-// to the current working directory, and it may contain wildcards ("*" and "?")
-// in the last path segment. This function passes all matching file names to
-// `consume`. The resulting paths may not be absolute nor normalized.
-//
-// The function returns a value from `ExpandWildcardsResult`.
-PROTOBUF_EXPORT ExpandWildcardsResult ExpandWildcards(
-    const std::string& path, std::function<void(const std::string&)> consume);
-
-namespace strings {
-
-// Convert from UTF-16 to Active-Code-Page-encoded or to UTF-8-encoded text.
-PROTOBUF_EXPORT bool wcs_to_mbs(const wchar_t* s, std::string* out,
-                                bool outUtf8);
-
-// Convert from Active-Code-Page-encoded or UTF-8-encoded text to UTF-16.
-PROTOBUF_EXPORT bool mbs_to_wcs(const char* s, std::wstring* out, bool inUtf8);
-
-// Convert from UTF-8-encoded text to UTF-16.
-PROTOBUF_EXPORT bool utf8_to_wcs(const char* input, std::wstring* out);
-
-// Convert from UTF-16-encoded text to UTF-8.
-PROTOBUF_EXPORT bool wcs_to_utf8(const wchar_t* input, std::string* out);
-
-}  // namespace strings
-
-}  // namespace win32
-}  // namespace io
-}  // namespace protobuf
-}  // namespace google
-
-#ifndef W_OK
-#define W_OK 02  // not defined by MSVC for whatever reason
-#endif
-
-#ifndef F_OK
-#define F_OK 00  // not defined by MSVC for whatever reason
-#endif
-
-#ifndef STDIN_FILENO
-#define STDIN_FILENO 0
-#endif
-
-#ifndef STDOUT_FILENO
-#define STDOUT_FILENO 1
-#endif
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // defined(_WIN32)
-
-#endif  // GOOGLE_PROTOBUF_IO_IO_WIN32_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/io/zero_copy_stream.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/io/zero_copy_stream.h
deleted file mode 100644
index d3bd6dae..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/io/zero_copy_stream.h
+++ /dev/null
@@ -1,253 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// Author: kenton@google.com (Kenton Varda)
-//  Based on original Protocol Buffers design by
-//  Sanjay Ghemawat, Jeff Dean, and others.
-//
-// This file contains the ZeroCopyInputStream and ZeroCopyOutputStream
-// interfaces, which represent abstract I/O streams to and from which
-// protocol buffers can be read and written.  For a few simple
-// implementations of these interfaces, see zero_copy_stream_impl.h.
-//
-// These interfaces are different from classic I/O streams in that they
-// try to minimize the amount of data copying that needs to be done.
-// To accomplish this, responsibility for allocating buffers is moved to
-// the stream object, rather than being the responsibility of the caller.
-// So, the stream can return a buffer which actually points directly into
-// the final data structure where the bytes are to be stored, and the caller
-// can interact directly with that buffer, eliminating an intermediate copy
-// operation.
-//
-// As an example, consider the common case in which you are reading bytes
-// from an array that is already in memory (or perhaps an mmap()ed file).
-// With classic I/O streams, you would do something like:
-//   char buffer[BUFFER_SIZE];
-//   input->Read(buffer, BUFFER_SIZE);
-//   DoSomething(buffer, BUFFER_SIZE);
-// Then, the stream basically just calls memcpy() to copy the data from
-// the array into your buffer.  With a ZeroCopyInputStream, you would do
-// this instead:
-//   const void* buffer;
-//   int size;
-//   input->Next(&buffer, &size);
-//   DoSomething(buffer, size);
-// Here, no copy is performed.  The input stream returns a pointer directly
-// into the backing array, and the caller ends up reading directly from it.
-//
-// If you want to be able to read the old-fashion way, you can create
-// a CodedInputStream or CodedOutputStream wrapping these objects and use
-// their ReadRaw()/WriteRaw() methods.  These will, of course, add a copy
-// step, but Coded*Stream will handle buffering so at least it will be
-// reasonably efficient.
-//
-// ZeroCopyInputStream example:
-//   // Read in a file and print its contents to stdout.
-//   int fd = open("myfile", O_RDONLY);
-//   ZeroCopyInputStream* input = new FileInputStream(fd);
-//
-//   const void* buffer;
-//   int size;
-//   while (input->Next(&buffer, &size)) {
-//     cout.write(buffer, size);
-//   }
-//
-//   delete input;
-//   close(fd);
-//
-// ZeroCopyOutputStream example:
-//   // Copy the contents of "infile" to "outfile", using plain read() for
-//   // "infile" but a ZeroCopyOutputStream for "outfile".
-//   int infd = open("infile", O_RDONLY);
-//   int outfd = open("outfile", O_WRONLY);
-//   ZeroCopyOutputStream* output = new FileOutputStream(outfd);
-//
-//   void* buffer;
-//   int size;
-//   while (output->Next(&buffer, &size)) {
-//     int bytes = read(infd, buffer, size);
-//     if (bytes < size) {
-//       // Reached EOF.
-//       output->BackUp(size - bytes);
-//       break;
-//     }
-//   }
-//
-//   delete output;
-//   close(infd);
-//   close(outfd);
-
-#ifndef GOOGLE_PROTOBUF_IO_ZERO_COPY_STREAM_H__
-#define GOOGLE_PROTOBUF_IO_ZERO_COPY_STREAM_H__
-
-
-#include <string>
-
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/port_def.inc>
-
-
-namespace google {
-namespace protobuf {
-namespace io {
-
-// Defined in this file.
-class ZeroCopyInputStream;
-class ZeroCopyOutputStream;
-
-// Abstract interface similar to an input stream but designed to minimize
-// copying.
-class PROTOBUF_EXPORT ZeroCopyInputStream {
- public:
-  ZeroCopyInputStream() {}
-  virtual ~ZeroCopyInputStream() {}
-
-  // Obtains a chunk of data from the stream.
-  //
-  // Preconditions:
-  // * "size" and "data" are not NULL.
-  //
-  // Postconditions:
-  // * If the returned value is false, there is no more data to return or
-  //   an error occurred.  All errors are permanent.
-  // * Otherwise, "size" points to the actual number of bytes read and "data"
-  //   points to a pointer to a buffer containing these bytes.
-  // * Ownership of this buffer remains with the stream, and the buffer
-  //   remains valid only until some other method of the stream is called
-  //   or the stream is destroyed.
-  // * It is legal for the returned buffer to have zero size, as long
-  //   as repeatedly calling Next() eventually yields a buffer with non-zero
-  //   size.
-  virtual bool Next(const void** data, int* size) = 0;
-
-  // Backs up a number of bytes, so that the next call to Next() returns
-  // data again that was already returned by the last call to Next().  This
-  // is useful when writing procedures that are only supposed to read up
-  // to a certain point in the input, then return.  If Next() returns a
-  // buffer that goes beyond what you wanted to read, you can use BackUp()
-  // to return to the point where you intended to finish.
-  //
-  // Preconditions:
-  // * The last method called must have been Next().
-  // * count must be less than or equal to the size of the last buffer
-  //   returned by Next().
-  //
-  // Postconditions:
-  // * The last "count" bytes of the last buffer returned by Next() will be
-  //   pushed back into the stream.  Subsequent calls to Next() will return
-  //   the same data again before producing new data.
-  virtual void BackUp(int count) = 0;
-
-  // Skips a number of bytes.  Returns false if the end of the stream is
-  // reached or some input error occurred.  In the end-of-stream case, the
-  // stream is advanced to the end of the stream (so ByteCount() will return
-  // the total size of the stream).
-  virtual bool Skip(int count) = 0;
-
-  // Returns the total number of bytes read since this object was created.
-  virtual int64_t ByteCount() const = 0;
-
-
- private:
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ZeroCopyInputStream);
-};
-
-// Abstract interface similar to an output stream but designed to minimize
-// copying.
-class PROTOBUF_EXPORT ZeroCopyOutputStream {
- public:
-  ZeroCopyOutputStream() {}
-  virtual ~ZeroCopyOutputStream() {}
-
-  // Obtains a buffer into which data can be written.  Any data written
-  // into this buffer will eventually (maybe instantly, maybe later on)
-  // be written to the output.
-  //
-  // Preconditions:
-  // * "size" and "data" are not NULL.
-  //
-  // Postconditions:
-  // * If the returned value is false, an error occurred.  All errors are
-  //   permanent.
-  // * Otherwise, "size" points to the actual number of bytes in the buffer
-  //   and "data" points to the buffer.
-  // * Ownership of this buffer remains with the stream, and the buffer
-  //   remains valid only until some other method of the stream is called
-  //   or the stream is destroyed.
-  // * Any data which the caller stores in this buffer will eventually be
-  //   written to the output (unless BackUp() is called).
-  // * It is legal for the returned buffer to have zero size, as long
-  //   as repeatedly calling Next() eventually yields a buffer with non-zero
-  //   size.
-  virtual bool Next(void** data, int* size) = 0;
-
-  // Backs up a number of bytes, so that the end of the last buffer returned
-  // by Next() is not actually written.  This is needed when you finish
-  // writing all the data you want to write, but the last buffer was bigger
-  // than you needed.  You don't want to write a bunch of garbage after the
-  // end of your data, so you use BackUp() to back up.
-  //
-  // Preconditions:
-  // * The last method called must have been Next().
-  // * count must be less than or equal to the size of the last buffer
-  //   returned by Next().
-  // * The caller must not have written anything to the last "count" bytes
-  //   of that buffer.
-  //
-  // Postconditions:
-  // * The last "count" bytes of the last buffer returned by Next() will be
-  //   ignored.
-  virtual void BackUp(int count) = 0;
-
-  // Returns the total number of bytes written since this object was created.
-  virtual int64_t ByteCount() const = 0;
-
-  // Write a given chunk of data to the output.  Some output streams may
-  // implement this in a way that avoids copying. Check AllowsAliasing() before
-  // calling WriteAliasedRaw(). It will GOOGLE_CHECK fail if WriteAliasedRaw() is
-  // called on a stream that does not allow aliasing.
-  //
-  // NOTE: It is caller's responsibility to ensure that the chunk of memory
-  // remains live until all of the data has been consumed from the stream.
-  virtual bool WriteAliasedRaw(const void* data, int size);
-  virtual bool AllowsAliasing() const { return false; }
-
-
- private:
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ZeroCopyOutputStream);
-};
-
-}  // namespace io
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_IO_ZERO_COPY_STREAM_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/io/zero_copy_stream_impl.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/io/zero_copy_stream_impl.h
deleted file mode 100644
index 0206e388..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/io/zero_copy_stream_impl.h
+++ /dev/null
@@ -1,327 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// Author: kenton@google.com (Kenton Varda)
-//  Based on original Protocol Buffers design by
-//  Sanjay Ghemawat, Jeff Dean, and others.
-//
-// This file contains common implementations of the interfaces defined in
-// zero_copy_stream.h which are only included in the full (non-lite)
-// protobuf library.  These implementations include Unix file descriptors
-// and C++ iostreams.  See also:  zero_copy_stream_impl_lite.h
-
-#ifndef GOOGLE_PROTOBUF_IO_ZERO_COPY_STREAM_IMPL_H__
-#define GOOGLE_PROTOBUF_IO_ZERO_COPY_STREAM_IMPL_H__
-
-
-#include <iosfwd>
-#include <string>
-
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/io/zero_copy_stream.h>
-#include <google/protobuf/io/zero_copy_stream_impl_lite.h>
-
-#include <google/protobuf/port_def.inc>
-
-namespace google {
-namespace protobuf {
-namespace io {
-
-// ===================================================================
-
-// A ZeroCopyInputStream which reads from a file descriptor.
-//
-// FileInputStream is preferred over using an ifstream with IstreamInputStream.
-// The latter will introduce an extra layer of buffering, harming performance.
-// Also, it's conceivable that FileInputStream could someday be enhanced
-// to use zero-copy file descriptors on OSs which support them.
-class PROTOBUF_EXPORT FileInputStream : public ZeroCopyInputStream {
- public:
-  // Creates a stream that reads from the given Unix file descriptor.
-  // If a block_size is given, it specifies the number of bytes that
-  // should be read and returned with each call to Next().  Otherwise,
-  // a reasonable default is used.
-  explicit FileInputStream(int file_descriptor, int block_size = -1);
-
-  // Flushes any buffers and closes the underlying file.  Returns false if
-  // an error occurs during the process; use GetErrno() to examine the error.
-  // Even if an error occurs, the file descriptor is closed when this returns.
-  bool Close();
-
-  // By default, the file descriptor is not closed when the stream is
-  // destroyed.  Call SetCloseOnDelete(true) to change that.  WARNING:
-  // This leaves no way for the caller to detect if close() fails.  If
-  // detecting close() errors is important to you, you should arrange
-  // to close the descriptor yourself.
-  void SetCloseOnDelete(bool value) { copying_input_.SetCloseOnDelete(value); }
-
-  // If an I/O error has occurred on this file descriptor, this is the
-  // errno from that error.  Otherwise, this is zero.  Once an error
-  // occurs, the stream is broken and all subsequent operations will
-  // fail.
-  int GetErrno() const { return copying_input_.GetErrno(); }
-
-  // implements ZeroCopyInputStream ----------------------------------
-  bool Next(const void** data, int* size) override;
-  void BackUp(int count) override;
-  bool Skip(int count) override;
-  int64_t ByteCount() const override;
-
- private:
-  class PROTOBUF_EXPORT CopyingFileInputStream : public CopyingInputStream {
-   public:
-    CopyingFileInputStream(int file_descriptor);
-    ~CopyingFileInputStream() override;
-
-    bool Close();
-    void SetCloseOnDelete(bool value) { close_on_delete_ = value; }
-    int GetErrno() const { return errno_; }
-
-    // implements CopyingInputStream ---------------------------------
-    int Read(void* buffer, int size) override;
-    int Skip(int count) override;
-
-   private:
-    // The file descriptor.
-    const int file_;
-    bool close_on_delete_;
-    bool is_closed_;
-
-    // The errno of the I/O error, if one has occurred.  Otherwise, zero.
-    int errno_;
-
-    // Did we try to seek once and fail?  If so, we assume this file descriptor
-    // doesn't support seeking and won't try again.
-    bool previous_seek_failed_;
-
-    GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(CopyingFileInputStream);
-  };
-
-  CopyingFileInputStream copying_input_;
-  CopyingInputStreamAdaptor impl_;
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(FileInputStream);
-};
-
-// ===================================================================
-
-// A ZeroCopyOutputStream which writes to a file descriptor.
-//
-// FileOutputStream is preferred over using an ofstream with
-// OstreamOutputStream.  The latter will introduce an extra layer of buffering,
-// harming performance.  Also, it's conceivable that FileOutputStream could
-// someday be enhanced to use zero-copy file descriptors on OSs which
-// support them.
-class PROTOBUF_EXPORT FileOutputStream : public CopyingOutputStreamAdaptor {
- public:
-  // Creates a stream that writes to the given Unix file descriptor.
-  // If a block_size is given, it specifies the size of the buffers
-  // that should be returned by Next().  Otherwise, a reasonable default
-  // is used.
-  explicit FileOutputStream(int file_descriptor, int block_size = -1);
-
-  ~FileOutputStream() override;
-
-  // Flushes any buffers and closes the underlying file.  Returns false if
-  // an error occurs during the process; use GetErrno() to examine the error.
-  // Even if an error occurs, the file descriptor is closed when this returns.
-  bool Close();
-
-  // By default, the file descriptor is not closed when the stream is
-  // destroyed.  Call SetCloseOnDelete(true) to change that.  WARNING:
-  // This leaves no way for the caller to detect if close() fails.  If
-  // detecting close() errors is important to you, you should arrange
-  // to close the descriptor yourself.
-  void SetCloseOnDelete(bool value) { copying_output_.SetCloseOnDelete(value); }
-
-  // If an I/O error has occurred on this file descriptor, this is the
-  // errno from that error.  Otherwise, this is zero.  Once an error
-  // occurs, the stream is broken and all subsequent operations will
-  // fail.
-  int GetErrno() const { return copying_output_.GetErrno(); }
-
- private:
-  class PROTOBUF_EXPORT CopyingFileOutputStream : public CopyingOutputStream {
-   public:
-    CopyingFileOutputStream(int file_descriptor);
-    ~CopyingFileOutputStream() override;
-
-    bool Close();
-    void SetCloseOnDelete(bool value) { close_on_delete_ = value; }
-    int GetErrno() const { return errno_; }
-
-    // implements CopyingOutputStream --------------------------------
-    bool Write(const void* buffer, int size) override;
-
-   private:
-    // The file descriptor.
-    const int file_;
-    bool close_on_delete_;
-    bool is_closed_;
-
-    // The errno of the I/O error, if one has occurred.  Otherwise, zero.
-    int errno_;
-
-    GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(CopyingFileOutputStream);
-  };
-
-  CopyingFileOutputStream copying_output_;
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(FileOutputStream);
-};
-
-// ===================================================================
-
-// A ZeroCopyInputStream which reads from a C++ istream.
-//
-// Note that for reading files (or anything represented by a file descriptor),
-// FileInputStream is more efficient.
-class PROTOBUF_EXPORT IstreamInputStream : public ZeroCopyInputStream {
- public:
-  // Creates a stream that reads from the given C++ istream.
-  // If a block_size is given, it specifies the number of bytes that
-  // should be read and returned with each call to Next().  Otherwise,
-  // a reasonable default is used.
-  explicit IstreamInputStream(std::istream* stream, int block_size = -1);
-
-  // implements ZeroCopyInputStream ----------------------------------
-  bool Next(const void** data, int* size) override;
-  void BackUp(int count) override;
-  bool Skip(int count) override;
-  int64_t ByteCount() const override;
-
- private:
-  class PROTOBUF_EXPORT CopyingIstreamInputStream : public CopyingInputStream {
-   public:
-    CopyingIstreamInputStream(std::istream* input);
-    ~CopyingIstreamInputStream() override;
-
-    // implements CopyingInputStream ---------------------------------
-    int Read(void* buffer, int size) override;
-    // (We use the default implementation of Skip().)
-
-   private:
-    // The stream.
-    std::istream* input_;
-
-    GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(CopyingIstreamInputStream);
-  };
-
-  CopyingIstreamInputStream copying_input_;
-  CopyingInputStreamAdaptor impl_;
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(IstreamInputStream);
-};
-
-// ===================================================================
-
-// A ZeroCopyOutputStream which writes to a C++ ostream.
-//
-// Note that for writing files (or anything represented by a file descriptor),
-// FileOutputStream is more efficient.
-class PROTOBUF_EXPORT OstreamOutputStream : public ZeroCopyOutputStream {
- public:
-  // Creates a stream that writes to the given C++ ostream.
-  // If a block_size is given, it specifies the size of the buffers
-  // that should be returned by Next().  Otherwise, a reasonable default
-  // is used.
-  explicit OstreamOutputStream(std::ostream* stream, int block_size = -1);
-  ~OstreamOutputStream() override;
-
-  // implements ZeroCopyOutputStream ---------------------------------
-  bool Next(void** data, int* size) override;
-  void BackUp(int count) override;
-  int64_t ByteCount() const override;
-
- private:
-  class PROTOBUF_EXPORT CopyingOstreamOutputStream
-      : public CopyingOutputStream {
-   public:
-    CopyingOstreamOutputStream(std::ostream* output);
-    ~CopyingOstreamOutputStream() override;
-
-    // implements CopyingOutputStream --------------------------------
-    bool Write(const void* buffer, int size) override;
-
-   private:
-    // The stream.
-    std::ostream* output_;
-
-    GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(CopyingOstreamOutputStream);
-  };
-
-  CopyingOstreamOutputStream copying_output_;
-  CopyingOutputStreamAdaptor impl_;
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(OstreamOutputStream);
-};
-
-// ===================================================================
-
-// A ZeroCopyInputStream which reads from several other streams in sequence.
-// ConcatenatingInputStream is unable to distinguish between end-of-stream
-// and read errors in the underlying streams, so it assumes any errors mean
-// end-of-stream.  So, if the underlying streams fail for any other reason,
-// ConcatenatingInputStream may do odd things.  It is suggested that you do
-// not use ConcatenatingInputStream on streams that might produce read errors
-// other than end-of-stream.
-class PROTOBUF_EXPORT ConcatenatingInputStream : public ZeroCopyInputStream {
- public:
-  // All streams passed in as well as the array itself must remain valid
-  // until the ConcatenatingInputStream is destroyed.
-  ConcatenatingInputStream(ZeroCopyInputStream* const streams[], int count);
-  ~ConcatenatingInputStream() override = default;
-
-  // implements ZeroCopyInputStream ----------------------------------
-  bool Next(const void** data, int* size) override;
-  void BackUp(int count) override;
-  bool Skip(int count) override;
-  int64_t ByteCount() const override;
-
-
- private:
-  // As streams are retired, streams_ is incremented and count_ is
-  // decremented.
-  ZeroCopyInputStream* const* streams_;
-  int stream_count_;
-  int64 bytes_retired_;  // Bytes read from previous streams.
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ConcatenatingInputStream);
-};
-
-// ===================================================================
-
-}  // namespace io
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_IO_ZERO_COPY_STREAM_IMPL_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/io/zero_copy_stream_impl_lite.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/io/zero_copy_stream_impl_lite.h
deleted file mode 100644
index cfe81d2c..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/io/zero_copy_stream_impl_lite.h
+++ /dev/null
@@ -1,408 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// Author: kenton@google.com (Kenton Varda)
-//  Based on original Protocol Buffers design by
-//  Sanjay Ghemawat, Jeff Dean, and others.
-//
-// This file contains common implementations of the interfaces defined in
-// zero_copy_stream.h which are included in the "lite" protobuf library.
-// These implementations cover I/O on raw arrays and strings, as well as
-// adaptors which make it easy to implement streams based on traditional
-// streams.  Of course, many users will probably want to write their own
-// implementations of these interfaces specific to the particular I/O
-// abstractions they prefer to use, but these should cover the most common
-// cases.
-
-#ifndef GOOGLE_PROTOBUF_IO_ZERO_COPY_STREAM_IMPL_LITE_H__
-#define GOOGLE_PROTOBUF_IO_ZERO_COPY_STREAM_IMPL_LITE_H__
-
-
-#include <iosfwd>
-#include <memory>
-#include <string>
-
-#include <google/protobuf/stubs/callback.h>
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/io/zero_copy_stream.h>
-#include <google/protobuf/stubs/stl_util.h>
-
-
-#include <google/protobuf/port_def.inc>
-
-namespace google {
-namespace protobuf {
-namespace io {
-
-// ===================================================================
-
-// A ZeroCopyInputStream backed by an in-memory array of bytes.
-class PROTOBUF_EXPORT ArrayInputStream : public ZeroCopyInputStream {
- public:
-  // Create an InputStream that returns the bytes pointed to by "data".
-  // "data" remains the property of the caller but must remain valid until
-  // the stream is destroyed.  If a block_size is given, calls to Next()
-  // will return data blocks no larger than the given size.  Otherwise, the
-  // first call to Next() returns the entire array.  block_size is mainly
-  // useful for testing; in production you would probably never want to set
-  // it.
-  ArrayInputStream(const void* data, int size, int block_size = -1);
-  ~ArrayInputStream() override = default;
-
-  // implements ZeroCopyInputStream ----------------------------------
-  bool Next(const void** data, int* size) override;
-  void BackUp(int count) override;
-  bool Skip(int count) override;
-  int64_t ByteCount() const override;
-
-
- private:
-  const uint8* const data_;  // The byte array.
-  const int size_;           // Total size of the array.
-  const int block_size_;     // How many bytes to return at a time.
-
-  int position_;
-  int last_returned_size_;  // How many bytes we returned last time Next()
-                            // was called (used for error checking only).
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ArrayInputStream);
-};
-
-// ===================================================================
-
-// A ZeroCopyOutputStream backed by an in-memory array of bytes.
-class PROTOBUF_EXPORT ArrayOutputStream : public ZeroCopyOutputStream {
- public:
-  // Create an OutputStream that writes to the bytes pointed to by "data".
-  // "data" remains the property of the caller but must remain valid until
-  // the stream is destroyed.  If a block_size is given, calls to Next()
-  // will return data blocks no larger than the given size.  Otherwise, the
-  // first call to Next() returns the entire array.  block_size is mainly
-  // useful for testing; in production you would probably never want to set
-  // it.
-  ArrayOutputStream(void* data, int size, int block_size = -1);
-  ~ArrayOutputStream() override = default;
-
-  // implements ZeroCopyOutputStream ---------------------------------
-  bool Next(void** data, int* size) override;
-  void BackUp(int count) override;
-  int64_t ByteCount() const override;
-
- private:
-  uint8* const data_;     // The byte array.
-  const int size_;        // Total size of the array.
-  const int block_size_;  // How many bytes to return at a time.
-
-  int position_;
-  int last_returned_size_;  // How many bytes we returned last time Next()
-                            // was called (used for error checking only).
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ArrayOutputStream);
-};
-
-// ===================================================================
-
-// A ZeroCopyOutputStream which appends bytes to a string.
-class PROTOBUF_EXPORT StringOutputStream : public ZeroCopyOutputStream {
- public:
-  // Create a StringOutputStream which appends bytes to the given string.
-  // The string remains property of the caller, but it is mutated in arbitrary
-  // ways and MUST NOT be accessed in any way until you're done with the
-  // stream. Either be sure there's no further usage, or (safest) destroy the
-  // stream before using the contents.
-  //
-  // Hint:  If you call target->reserve(n) before creating the stream,
-  //   the first call to Next() will return at least n bytes of buffer
-  //   space.
-  explicit StringOutputStream(std::string* target);
-  ~StringOutputStream() override = default;
-
-  // implements ZeroCopyOutputStream ---------------------------------
-  bool Next(void** data, int* size) override;
-  void BackUp(int count) override;
-  int64_t ByteCount() const override;
-
- private:
-  static constexpr size_t kMinimumSize = 16;
-
-  std::string* target_;
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(StringOutputStream);
-};
-
-// Note:  There is no StringInputStream.  Instead, just create an
-// ArrayInputStream as follows:
-//   ArrayInputStream input(str.data(), str.size());
-
-// ===================================================================
-
-// A generic traditional input stream interface.
-//
-// Lots of traditional input streams (e.g. file descriptors, C stdio
-// streams, and C++ iostreams) expose an interface where every read
-// involves copying bytes into a buffer.  If you want to take such an
-// interface and make a ZeroCopyInputStream based on it, simply implement
-// CopyingInputStream and then use CopyingInputStreamAdaptor.
-//
-// CopyingInputStream implementations should avoid buffering if possible.
-// CopyingInputStreamAdaptor does its own buffering and will read data
-// in large blocks.
-class PROTOBUF_EXPORT CopyingInputStream {
- public:
-  virtual ~CopyingInputStream() {}
-
-  // Reads up to "size" bytes into the given buffer.  Returns the number of
-  // bytes read.  Read() waits until at least one byte is available, or
-  // returns zero if no bytes will ever become available (EOF), or -1 if a
-  // permanent read error occurred.
-  virtual int Read(void* buffer, int size) = 0;
-
-  // Skips the next "count" bytes of input.  Returns the number of bytes
-  // actually skipped.  This will always be exactly equal to "count" unless
-  // EOF was reached or a permanent read error occurred.
-  //
-  // The default implementation just repeatedly calls Read() into a scratch
-  // buffer.
-  virtual int Skip(int count);
-};
-
-// A ZeroCopyInputStream which reads from a CopyingInputStream.  This is
-// useful for implementing ZeroCopyInputStreams that read from traditional
-// streams.  Note that this class is not really zero-copy.
-//
-// If you want to read from file descriptors or C++ istreams, this is
-// already implemented for you:  use FileInputStream or IstreamInputStream
-// respectively.
-class PROTOBUF_EXPORT CopyingInputStreamAdaptor : public ZeroCopyInputStream {
- public:
-  // Creates a stream that reads from the given CopyingInputStream.
-  // If a block_size is given, it specifies the number of bytes that
-  // should be read and returned with each call to Next().  Otherwise,
-  // a reasonable default is used.  The caller retains ownership of
-  // copying_stream unless SetOwnsCopyingStream(true) is called.
-  explicit CopyingInputStreamAdaptor(CopyingInputStream* copying_stream,
-                                     int block_size = -1);
-  ~CopyingInputStreamAdaptor() override;
-
-  // Call SetOwnsCopyingStream(true) to tell the CopyingInputStreamAdaptor to
-  // delete the underlying CopyingInputStream when it is destroyed.
-  void SetOwnsCopyingStream(bool value) { owns_copying_stream_ = value; }
-
-  // implements ZeroCopyInputStream ----------------------------------
-  bool Next(const void** data, int* size) override;
-  void BackUp(int count) override;
-  bool Skip(int count) override;
-  int64_t ByteCount() const override;
-
- private:
-  // Insures that buffer_ is not NULL.
-  void AllocateBufferIfNeeded();
-  // Frees the buffer and resets buffer_used_.
-  void FreeBuffer();
-
-  // The underlying copying stream.
-  CopyingInputStream* copying_stream_;
-  bool owns_copying_stream_;
-
-  // True if we have seen a permanent error from the underlying stream.
-  bool failed_;
-
-  // The current position of copying_stream_, relative to the point where
-  // we started reading.
-  int64 position_;
-
-  // Data is read into this buffer.  It may be NULL if no buffer is currently
-  // in use.  Otherwise, it points to an array of size buffer_size_.
-  std::unique_ptr<uint8[]> buffer_;
-  const int buffer_size_;
-
-  // Number of valid bytes currently in the buffer (i.e. the size last
-  // returned by Next()).  0 <= buffer_used_ <= buffer_size_.
-  int buffer_used_;
-
-  // Number of bytes in the buffer which were backed up over by a call to
-  // BackUp().  These need to be returned again.
-  // 0 <= backup_bytes_ <= buffer_used_
-  int backup_bytes_;
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(CopyingInputStreamAdaptor);
-};
-
-// ===================================================================
-
-// A generic traditional output stream interface.
-//
-// Lots of traditional output streams (e.g. file descriptors, C stdio
-// streams, and C++ iostreams) expose an interface where every write
-// involves copying bytes from a buffer.  If you want to take such an
-// interface and make a ZeroCopyOutputStream based on it, simply implement
-// CopyingOutputStream and then use CopyingOutputStreamAdaptor.
-//
-// CopyingOutputStream implementations should avoid buffering if possible.
-// CopyingOutputStreamAdaptor does its own buffering and will write data
-// in large blocks.
-class PROTOBUF_EXPORT CopyingOutputStream {
- public:
-  virtual ~CopyingOutputStream() {}
-
-  // Writes "size" bytes from the given buffer to the output.  Returns true
-  // if successful, false on a write error.
-  virtual bool Write(const void* buffer, int size) = 0;
-};
-
-// A ZeroCopyOutputStream which writes to a CopyingOutputStream.  This is
-// useful for implementing ZeroCopyOutputStreams that write to traditional
-// streams.  Note that this class is not really zero-copy.
-//
-// If you want to write to file descriptors or C++ ostreams, this is
-// already implemented for you:  use FileOutputStream or OstreamOutputStream
-// respectively.
-class PROTOBUF_EXPORT CopyingOutputStreamAdaptor : public ZeroCopyOutputStream {
- public:
-  // Creates a stream that writes to the given Unix file descriptor.
-  // If a block_size is given, it specifies the size of the buffers
-  // that should be returned by Next().  Otherwise, a reasonable default
-  // is used.
-  explicit CopyingOutputStreamAdaptor(CopyingOutputStream* copying_stream,
-                                      int block_size = -1);
-  ~CopyingOutputStreamAdaptor() override;
-
-  // Writes all pending data to the underlying stream.  Returns false if a
-  // write error occurred on the underlying stream.  (The underlying
-  // stream itself is not necessarily flushed.)
-  bool Flush();
-
-  // Call SetOwnsCopyingStream(true) to tell the CopyingOutputStreamAdaptor to
-  // delete the underlying CopyingOutputStream when it is destroyed.
-  void SetOwnsCopyingStream(bool value) { owns_copying_stream_ = value; }
-
-  // implements ZeroCopyOutputStream ---------------------------------
-  bool Next(void** data, int* size) override;
-  void BackUp(int count) override;
-  int64_t ByteCount() const override;
-  bool WriteAliasedRaw(const void* data, int size) override;
-  bool AllowsAliasing() const override { return true; }
-
- private:
-  // Write the current buffer, if it is present.
-  bool WriteBuffer();
-  // Insures that buffer_ is not NULL.
-  void AllocateBufferIfNeeded();
-  // Frees the buffer.
-  void FreeBuffer();
-
-  // The underlying copying stream.
-  CopyingOutputStream* copying_stream_;
-  bool owns_copying_stream_;
-
-  // True if we have seen a permanent error from the underlying stream.
-  bool failed_;
-
-  // The current position of copying_stream_, relative to the point where
-  // we started writing.
-  int64 position_;
-
-  // Data is written from this buffer.  It may be NULL if no buffer is
-  // currently in use.  Otherwise, it points to an array of size buffer_size_.
-  std::unique_ptr<uint8[]> buffer_;
-  const int buffer_size_;
-
-  // Number of valid bytes currently in the buffer (i.e. the size last
-  // returned by Next()).  When BackUp() is called, we just reduce this.
-  // 0 <= buffer_used_ <= buffer_size_.
-  int buffer_used_;
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(CopyingOutputStreamAdaptor);
-};
-
-// ===================================================================
-
-// A ZeroCopyInputStream which wraps some other stream and limits it to
-// a particular byte count.
-class PROTOBUF_EXPORT LimitingInputStream : public ZeroCopyInputStream {
- public:
-  LimitingInputStream(ZeroCopyInputStream* input, int64 limit);
-  ~LimitingInputStream() override;
-
-  // implements ZeroCopyInputStream ----------------------------------
-  bool Next(const void** data, int* size) override;
-  void BackUp(int count) override;
-  bool Skip(int count) override;
-  int64_t ByteCount() const override;
-
-
- private:
-  ZeroCopyInputStream* input_;
-  int64 limit_;  // Decreases as we go, becomes negative if we overshoot.
-  int64 prior_bytes_read_;  // Bytes read on underlying stream at construction
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(LimitingInputStream);
-};
-
-
-// ===================================================================
-
-// mutable_string_data() and as_string_data() are workarounds to improve
-// the performance of writing new data to an existing string.  Unfortunately
-// the methods provided by the string class are suboptimal, and using memcpy()
-// is mildly annoying because it requires its pointer args to be non-NULL even
-// if we ask it to copy 0 bytes.  Furthermore, string_as_array() has the
-// property that it always returns NULL if its arg is the empty string, exactly
-// what we want to avoid if we're using it in conjunction with memcpy()!
-// With C++11, the desired memcpy() boils down to memcpy(..., &(*s)[0], size),
-// where s is a string*.  Without C++11, &(*s)[0] is not guaranteed to be safe,
-// so we use string_as_array(), and live with the extra logic that tests whether
-// *s is empty.
-
-// Return a pointer to mutable characters underlying the given string.  The
-// return value is valid until the next time the string is resized.  We
-// trust the caller to treat the return value as an array of length s->size().
-inline char* mutable_string_data(std::string* s) {
-  // This should be simpler & faster than string_as_array() because the latter
-  // is guaranteed to return NULL when *s is empty, so it has to check for that.
-  return &(*s)[0];
-}
-
-// as_string_data(s) is equivalent to
-//  ({ char* p = mutable_string_data(s); make_pair(p, p != NULL); })
-// Sometimes it's faster: in some scenarios p cannot be NULL, and then the
-// code can avoid that check.
-inline std::pair<char*, bool> as_string_data(std::string* s) {
-  char* p = mutable_string_data(s);
-  return std::make_pair(p, true);
-}
-
-}  // namespace io
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_IO_ZERO_COPY_STREAM_IMPL_LITE_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/map.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/map.h
deleted file mode 100644
index 2453246d..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/map.h
+++ /dev/null
@@ -1,1362 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// This file defines the map container and its helpers to support protobuf maps.
-//
-// The Map and MapIterator types are provided by this header file.
-// Please avoid using other types defined here, unless they are public
-// types within Map or MapIterator, such as Map::value_type.
-
-#ifndef GOOGLE_PROTOBUF_MAP_H__
-#define GOOGLE_PROTOBUF_MAP_H__
-
-#include <functional>
-#include <initializer_list>
-#include <iterator>
-#include <limits>  // To support Visual Studio 2008
-#include <map>
-#include <string>
-#include <type_traits>
-#include <utility>
-
-#if defined(__cpp_lib_string_view)
-#include <string_view>
-#endif  // defined(__cpp_lib_string_view)
-
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/arena.h>
-#include <google/protobuf/generated_enum_util.h>
-#include <google/protobuf/map_type_handler.h>
-#include <google/protobuf/stubs/hash.h>
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-#include <google/protobuf/port_def.inc>
-
-namespace google {
-namespace protobuf {
-
-template <typename Key, typename T>
-class Map;
-
-class MapIterator;
-
-template <typename Enum>
-struct is_proto_enum;
-
-namespace internal {
-template <typename Derived, typename Key, typename T,
-          WireFormatLite::FieldType key_wire_type,
-          WireFormatLite::FieldType value_wire_type>
-class MapFieldLite;
-
-template <typename Derived, typename Key, typename T,
-          WireFormatLite::FieldType key_wire_type,
-          WireFormatLite::FieldType value_wire_type>
-class MapField;
-
-template <typename Key, typename T>
-class TypeDefinedMapFieldBase;
-
-class DynamicMapField;
-
-class GeneratedMessageReflection;
-
-// re-implement std::allocator to use arena allocator for memory allocation.
-// Used for Map implementation. Users should not use this class
-// directly.
-template <typename U>
-class MapAllocator {
- public:
-  using value_type = U;
-  using pointer = value_type*;
-  using const_pointer = const value_type*;
-  using reference = value_type&;
-  using const_reference = const value_type&;
-  using size_type = size_t;
-  using difference_type = ptrdiff_t;
-
-  constexpr MapAllocator() : arena_(nullptr) {}
-  explicit constexpr MapAllocator(Arena* arena) : arena_(arena) {}
-  template <typename X>
-  MapAllocator(const MapAllocator<X>& allocator)  // NOLINT(runtime/explicit)
-      : arena_(allocator.arena()) {}
-
-  pointer allocate(size_type n, const void* /* hint */ = nullptr) {
-    // If arena is not given, malloc needs to be called which doesn't
-    // construct element object.
-    if (arena_ == nullptr) {
-      return static_cast<pointer>(::operator new(n * sizeof(value_type)));
-    } else {
-      return reinterpret_cast<pointer>(
-          Arena::CreateArray<uint8>(arena_, n * sizeof(value_type)));
-    }
-  }
-
-  void deallocate(pointer p, size_type n) {
-    if (arena_ == nullptr) {
-#if defined(__GXX_DELETE_WITH_SIZE__) || defined(__cpp_sized_deallocation)
-      ::operator delete(p, n * sizeof(value_type));
-#else
-      (void)n;
-      ::operator delete(p);
-#endif
-    }
-  }
-
-#if !defined(GOOGLE_PROTOBUF_OS_APPLE) && !defined(GOOGLE_PROTOBUF_OS_NACL) && \
-    !defined(GOOGLE_PROTOBUF_OS_EMSCRIPTEN)
-  template <class NodeType, class... Args>
-  void construct(NodeType* p, Args&&... args) {
-    // Clang 3.6 doesn't compile static casting to void* directly. (Issue
-    // #1266) According C++ standard 5.2.9/1: "The static_cast operator shall
-    // not cast away constness". So first the maybe const pointer is casted to
-    // const void* and after the const void* is const casted.
-    new (const_cast<void*>(static_cast<const void*>(p)))
-        NodeType(std::forward<Args>(args)...);
-  }
-
-  template <class NodeType>
-  void destroy(NodeType* p) {
-    p->~NodeType();
-  }
-#else
-  void construct(pointer p, const_reference t) { new (p) value_type(t); }
-
-  void destroy(pointer p) { p->~value_type(); }
-#endif
-
-  template <typename X>
-  struct rebind {
-    using other = MapAllocator<X>;
-  };
-
-  template <typename X>
-  bool operator==(const MapAllocator<X>& other) const {
-    return arena_ == other.arena_;
-  }
-
-  template <typename X>
-  bool operator!=(const MapAllocator<X>& other) const {
-    return arena_ != other.arena_;
-  }
-
-  // To support Visual Studio 2008
-  size_type max_size() const {
-    // parentheses around (std::...:max) prevents macro warning of max()
-    return (std::numeric_limits<size_type>::max)();
-  }
-
-  // To support gcc-4.4, which does not properly
-  // support templated friend classes
-  Arena* arena() const { return arena_; }
-
- private:
-  using DestructorSkippable_ = void;
-  Arena* arena_;
-};
-
-template <typename T>
-using KeyForTree =
-    typename std::conditional<std::is_scalar<T>::value, T,
-                              std::reference_wrapper<const T>>::type;
-
-// Default case: Not transparent.
-// We use std::hash<key_type>/std::less<key_type> and all the lookup functions
-// only accept `key_type`.
-template <typename key_type>
-struct TransparentSupport {
-  using hash = std::hash<key_type>;
-  using less = std::less<key_type>;
-
-  static bool Equals(const key_type& a, const key_type& b) { return a == b; }
-
-  template <typename K>
-  using key_arg = key_type;
-};
-
-#if defined(__cpp_lib_string_view)
-// If std::string_view is available, we add transparent support for std::string
-// keys. We use std::hash<std::string_view> as it supports the input types we
-// care about. The lookup functions accept arbitrary `K`. This will include any
-// key type that is convertible to std::string_view.
-template <>
-struct TransparentSupport<std::string> {
-  static std::string_view ImplicitConvert(std::string_view str) { return str; }
-  // If the element is not convertible to std::string_view, try to convert to
-  // std::string first.
-  // The template makes this overload lose resolution when both have the same
-  // rank otherwise.
-  template <typename = void>
-  static std::string_view ImplicitConvert(const std::string& str) {
-    return str;
-  }
-
-  struct hash : private std::hash<std::string_view> {
-    using is_transparent = void;
-
-    template <typename T>
-    size_t operator()(const T& str) const {
-      return base()(ImplicitConvert(str));
-    }
-
-   private:
-    const std::hash<std::string_view>& base() const { return *this; }
-  };
-  struct less {
-    using is_transparent = void;
-
-    template <typename T, typename U>
-    bool operator()(const T& t, const U& u) const {
-      return ImplicitConvert(t) < ImplicitConvert(u);
-    }
-  };
-
-  template <typename T, typename U>
-  static bool Equals(const T& t, const U& u) {
-    return ImplicitConvert(t) == ImplicitConvert(u);
-  }
-
-  template <typename K>
-  using key_arg = K;
-};
-#endif  // defined(__cpp_lib_string_view)
-
-template <typename Key>
-using TreeForMap =
-    std::map<KeyForTree<Key>, void*, typename TransparentSupport<Key>::less,
-             MapAllocator<std::pair<const KeyForTree<Key>, void*>>>;
-
-inline bool TableEntryIsEmpty(void* const* table, size_t b) {
-  return table[b] == nullptr;
-}
-inline bool TableEntryIsNonEmptyList(void* const* table, size_t b) {
-  return table[b] != nullptr && table[b] != table[b ^ 1];
-}
-inline bool TableEntryIsTree(void* const* table, size_t b) {
-  return !TableEntryIsEmpty(table, b) && !TableEntryIsNonEmptyList(table, b);
-}
-inline bool TableEntryIsList(void* const* table, size_t b) {
-  return !TableEntryIsTree(table, b);
-}
-
-// This captures all numeric types.
-inline size_t MapValueSpaceUsedExcludingSelfLong(bool) { return 0; }
-inline size_t MapValueSpaceUsedExcludingSelfLong(const std::string& str) {
-  return StringSpaceUsedExcludingSelfLong(str);
-}
-template <typename T,
-          typename = decltype(std::declval<const T&>().SpaceUsedLong())>
-size_t MapValueSpaceUsedExcludingSelfLong(const T& message) {
-  return message.SpaceUsedLong() - sizeof(T);
-}
-
-constexpr size_t kGlobalEmptyTableSize = 1;
-PROTOBUF_EXPORT extern void* const kGlobalEmptyTable[kGlobalEmptyTableSize];
-
-// Space used for the table, trees, and nodes.
-// Does not include the indirect space used. Eg the data of a std::string.
-template <typename Key>
-PROTOBUF_NOINLINE size_t SpaceUsedInTable(void** table, size_t num_buckets,
-                                          size_t num_elements,
-                                          size_t sizeof_node) {
-  size_t size = 0;
-  // The size of the table.
-  size += sizeof(void*) * num_buckets;
-  // All the nodes.
-  size += sizeof_node * num_elements;
-  // For each tree, count the overhead of the those nodes.
-  // Two buckets at a time because we only care about trees.
-  for (size_t b = 0; b < num_buckets; b += 2) {
-    if (internal::TableEntryIsTree(table, b)) {
-      using Tree = TreeForMap<Key>;
-      Tree* tree = static_cast<Tree*>(table[b]);
-      // Estimated cost of the red-black tree nodes, 3 pointers plus a
-      // bool (plus alignment, so 4 pointers).
-      size += tree->size() *
-              (sizeof(typename Tree::value_type) + sizeof(void*) * 4);
-    }
-  }
-  return size;
-}
-
-template <typename Map,
-          typename = typename std::enable_if<
-              !std::is_scalar<typename Map::key_type>::value ||
-              !std::is_scalar<typename Map::mapped_type>::value>::type>
-size_t SpaceUsedInValues(const Map* map) {
-  size_t size = 0;
-  for (const auto& v : *map) {
-    size += internal::MapValueSpaceUsedExcludingSelfLong(v.first) +
-            internal::MapValueSpaceUsedExcludingSelfLong(v.second);
-  }
-  return size;
-}
-
-inline size_t SpaceUsedInValues(const void*) { return 0; }
-
-}  // namespace internal
-
-// This is the class for Map's internal value_type. Instead of using
-// std::pair as value_type, we use this class which provides us more control of
-// its process of construction and destruction.
-template <typename Key, typename T>
-struct MapPair {
-  using first_type = const Key;
-  using second_type = T;
-
-  MapPair(const Key& other_first, const T& other_second)
-      : first(other_first), second(other_second) {}
-  explicit MapPair(const Key& other_first) : first(other_first), second() {}
-  explicit MapPair(Key&& other_first)
-      : first(std::move(other_first)), second() {}
-  MapPair(const MapPair& other) : first(other.first), second(other.second) {}
-
-  ~MapPair() {}
-
-  // Implicitly convertible to std::pair of compatible types.
-  template <typename T1, typename T2>
-  operator std::pair<T1, T2>() const {  // NOLINT(runtime/explicit)
-    return std::pair<T1, T2>(first, second);
-  }
-
-  const Key first;
-  T second;
-
- private:
-  friend class Arena;
-  friend class Map<Key, T>;
-};
-
-// Map is an associative container type used to store protobuf map
-// fields.  Each Map instance may or may not use a different hash function, a
-// different iteration order, and so on.  E.g., please don't examine
-// implementation details to decide if the following would work:
-//  Map<int, int> m0, m1;
-//  m0[0] = m1[0] = m0[1] = m1[1] = 0;
-//  assert(m0.begin()->first == m1.begin()->first);  // Bug!
-//
-// Map's interface is similar to std::unordered_map, except that Map is not
-// designed to play well with exceptions.
-template <typename Key, typename T>
-class Map {
- public:
-  using key_type = Key;
-  using mapped_type = T;
-  using value_type = MapPair<Key, T>;
-
-  using pointer = value_type*;
-  using const_pointer = const value_type*;
-  using reference = value_type&;
-  using const_reference = const value_type&;
-
-  using size_type = size_t;
-  using hasher = typename internal::TransparentSupport<Key>::hash;
-
-  constexpr Map() : elements_(nullptr) {}
-  explicit Map(Arena* arena) : elements_(arena) {}
-
-  Map(const Map& other) : Map() { insert(other.begin(), other.end()); }
-
-  Map(Map&& other) noexcept : Map() {
-    if (other.arena() != nullptr) {
-      *this = other;
-    } else {
-      swap(other);
-    }
-  }
-
-  Map& operator=(Map&& other) noexcept {
-    if (this != &other) {
-      if (arena() != other.arena()) {
-        *this = other;
-      } else {
-        swap(other);
-      }
-    }
-    return *this;
-  }
-
-  template <class InputIt>
-  Map(const InputIt& first, const InputIt& last) : Map() {
-    insert(first, last);
-  }
-
-  ~Map() {}
-
- private:
-  using Allocator = internal::MapAllocator<void*>;
-
-  // InnerMap is a generic hash-based map.  It doesn't contain any
-  // protocol-buffer-specific logic.  It is a chaining hash map with the
-  // additional feature that some buckets can be converted to use an ordered
-  // container.  This ensures O(lg n) bounds on find, insert, and erase, while
-  // avoiding the overheads of ordered containers most of the time.
-  //
-  // The implementation doesn't need the full generality of unordered_map,
-  // and it doesn't have it.  More bells and whistles can be added as needed.
-  // Some implementation details:
-  // 1. The hash function has type hasher and the equality function
-  //    equal_to<Key>.  We inherit from hasher to save space
-  //    (empty-base-class optimization).
-  // 2. The number of buckets is a power of two.
-  // 3. Buckets are converted to trees in pairs: if we convert bucket b then
-  //    buckets b and b^1 will share a tree.  Invariant: buckets b and b^1 have
-  //    the same non-null value iff they are sharing a tree.  (An alternative
-  //    implementation strategy would be to have a tag bit per bucket.)
-  // 4. As is typical for hash_map and such, the Keys and Values are always
-  //    stored in linked list nodes.  Pointers to elements are never invalidated
-  //    until the element is deleted.
-  // 5. The trees' payload type is pointer to linked-list node.  Tree-converting
-  //    a bucket doesn't copy Key-Value pairs.
-  // 6. Once we've tree-converted a bucket, it is never converted back. However,
-  //    the items a tree contains may wind up assigned to trees or lists upon a
-  //    rehash.
-  // 7. The code requires no C++ features from C++14 or later.
-  // 8. Mutations to a map do not invalidate the map's iterators, pointers to
-  //    elements, or references to elements.
-  // 9. Except for erase(iterator), any non-const method can reorder iterators.
-  // 10. InnerMap uses KeyForTree<Key> when using the Tree representation, which
-  //    is either `Key`, if Key is a scalar, or `reference_wrapper<const Key>`
-  //    otherwise. This avoids unnecessary copies of string keys, for example.
-  class InnerMap : private hasher {
-   public:
-    explicit constexpr InnerMap(Arena* arena)
-        : hasher(),
-          num_elements_(0),
-          num_buckets_(internal::kGlobalEmptyTableSize),
-          seed_(0),
-          index_of_first_non_null_(internal::kGlobalEmptyTableSize),
-          table_(const_cast<void**>(internal::kGlobalEmptyTable)),
-          alloc_(arena) {}
-
-    ~InnerMap() {
-      if (alloc_.arena() == nullptr &&
-          num_buckets_ != internal::kGlobalEmptyTableSize) {
-        clear();
-        Dealloc<void*>(table_, num_buckets_);
-      }
-    }
-
-   private:
-    enum { kMinTableSize = 8 };
-
-    // Linked-list nodes, as one would expect for a chaining hash table.
-    struct Node {
-      value_type kv;
-      Node* next;
-    };
-
-    // Trees. The payload type is a copy of Key, so that we can query the tree
-    // with Keys that are not in any particular data structure.
-    // The value is a void* pointing to Node. We use void* instead of Node* to
-    // avoid code bloat. That way there is only one instantiation of the tree
-    // class per key type.
-    using Tree = internal::TreeForMap<Key>;
-    using TreeIterator = typename Tree::iterator;
-
-    static Node* NodeFromTreeIterator(TreeIterator it) {
-      return static_cast<Node*>(it->second);
-    }
-
-    // iterator and const_iterator are instantiations of iterator_base.
-    template <typename KeyValueType>
-    class iterator_base {
-     public:
-      using reference = KeyValueType&;
-      using pointer = KeyValueType*;
-
-      // Invariants:
-      // node_ is always correct. This is handy because the most common
-      // operations are operator* and operator-> and they only use node_.
-      // When node_ is set to a non-null value, all the other non-const fields
-      // are updated to be correct also, but those fields can become stale
-      // if the underlying map is modified.  When those fields are needed they
-      // are rechecked, and updated if necessary.
-      iterator_base() : node_(nullptr), m_(nullptr), bucket_index_(0) {}
-
-      explicit iterator_base(const InnerMap* m) : m_(m) {
-        SearchFrom(m->index_of_first_non_null_);
-      }
-
-      // Any iterator_base can convert to any other.  This is overkill, and we
-      // rely on the enclosing class to use it wisely.  The standard "iterator
-      // can convert to const_iterator" is OK but the reverse direction is not.
-      template <typename U>
-      explicit iterator_base(const iterator_base<U>& it)
-          : node_(it.node_), m_(it.m_), bucket_index_(it.bucket_index_) {}
-
-      iterator_base(Node* n, const InnerMap* m, size_type index)
-          : node_(n), m_(m), bucket_index_(index) {}
-
-      iterator_base(TreeIterator tree_it, const InnerMap* m, size_type index)
-          : node_(NodeFromTreeIterator(tree_it)), m_(m), bucket_index_(index) {
-        // Invariant: iterators that use buckets with trees have an even
-        // bucket_index_.
-        GOOGLE_DCHECK_EQ(bucket_index_ % 2, 0u);
-      }
-
-      // Advance through buckets, looking for the first that isn't empty.
-      // If nothing non-empty is found then leave node_ == nullptr.
-      void SearchFrom(size_type start_bucket) {
-        GOOGLE_DCHECK(m_->index_of_first_non_null_ == m_->num_buckets_ ||
-               m_->table_[m_->index_of_first_non_null_] != nullptr);
-        node_ = nullptr;
-        for (bucket_index_ = start_bucket; bucket_index_ < m_->num_buckets_;
-             bucket_index_++) {
-          if (m_->TableEntryIsNonEmptyList(bucket_index_)) {
-            node_ = static_cast<Node*>(m_->table_[bucket_index_]);
-            break;
-          } else if (m_->TableEntryIsTree(bucket_index_)) {
-            Tree* tree = static_cast<Tree*>(m_->table_[bucket_index_]);
-            GOOGLE_DCHECK(!tree->empty());
-            node_ = NodeFromTreeIterator(tree->begin());
-            break;
-          }
-        }
-      }
-
-      reference operator*() const { return node_->kv; }
-      pointer operator->() const { return &(operator*()); }
-
-      friend bool operator==(const iterator_base& a, const iterator_base& b) {
-        return a.node_ == b.node_;
-      }
-      friend bool operator!=(const iterator_base& a, const iterator_base& b) {
-        return a.node_ != b.node_;
-      }
-
-      iterator_base& operator++() {
-        if (node_->next == nullptr) {
-          TreeIterator tree_it;
-          const bool is_list = revalidate_if_necessary(&tree_it);
-          if (is_list) {
-            SearchFrom(bucket_index_ + 1);
-          } else {
-            GOOGLE_DCHECK_EQ(bucket_index_ & 1, 0u);
-            Tree* tree = static_cast<Tree*>(m_->table_[bucket_index_]);
-            if (++tree_it == tree->end()) {
-              SearchFrom(bucket_index_ + 2);
-            } else {
-              node_ = NodeFromTreeIterator(tree_it);
-            }
-          }
-        } else {
-          node_ = node_->next;
-        }
-        return *this;
-      }
-
-      iterator_base operator++(int /* unused */) {
-        iterator_base tmp = *this;
-        ++*this;
-        return tmp;
-      }
-
-      // Assumes node_ and m_ are correct and non-null, but other fields may be
-      // stale.  Fix them as needed.  Then return true iff node_ points to a
-      // Node in a list.  If false is returned then *it is modified to be
-      // a valid iterator for node_.
-      bool revalidate_if_necessary(TreeIterator* it) {
-        GOOGLE_DCHECK(node_ != nullptr && m_ != nullptr);
-        // Force bucket_index_ to be in range.
-        bucket_index_ &= (m_->num_buckets_ - 1);
-        // Common case: the bucket we think is relevant points to node_.
-        if (m_->table_[bucket_index_] == static_cast<void*>(node_)) return true;
-        // Less common: the bucket is a linked list with node_ somewhere in it,
-        // but not at the head.
-        if (m_->TableEntryIsNonEmptyList(bucket_index_)) {
-          Node* l = static_cast<Node*>(m_->table_[bucket_index_]);
-          while ((l = l->next) != nullptr) {
-            if (l == node_) {
-              return true;
-            }
-          }
-        }
-        // Well, bucket_index_ still might be correct, but probably
-        // not.  Revalidate just to be sure.  This case is rare enough that we
-        // don't worry about potential optimizations, such as having a custom
-        // find-like method that compares Node* instead of the key.
-        iterator_base i(m_->find(node_->kv.first, it));
-        bucket_index_ = i.bucket_index_;
-        return m_->TableEntryIsList(bucket_index_);
-      }
-
-      Node* node_;
-      const InnerMap* m_;
-      size_type bucket_index_;
-    };
-
-   public:
-    using iterator = iterator_base<value_type>;
-    using const_iterator = iterator_base<const value_type>;
-
-    Arena* arena() const { return alloc_.arena(); }
-
-    void Swap(InnerMap* other) {
-      std::swap(num_elements_, other->num_elements_);
-      std::swap(num_buckets_, other->num_buckets_);
-      std::swap(seed_, other->seed_);
-      std::swap(index_of_first_non_null_, other->index_of_first_non_null_);
-      std::swap(table_, other->table_);
-      std::swap(alloc_, other->alloc_);
-    }
-
-    iterator begin() { return iterator(this); }
-    iterator end() { return iterator(); }
-    const_iterator begin() const { return const_iterator(this); }
-    const_iterator end() const { return const_iterator(); }
-
-    void clear() {
-      for (size_type b = 0; b < num_buckets_; b++) {
-        if (TableEntryIsNonEmptyList(b)) {
-          Node* node = static_cast<Node*>(table_[b]);
-          table_[b] = nullptr;
-          do {
-            Node* next = node->next;
-            DestroyNode(node);
-            node = next;
-          } while (node != nullptr);
-        } else if (TableEntryIsTree(b)) {
-          Tree* tree = static_cast<Tree*>(table_[b]);
-          GOOGLE_DCHECK(table_[b] == table_[b + 1] && (b & 1) == 0);
-          table_[b] = table_[b + 1] = nullptr;
-          typename Tree::iterator tree_it = tree->begin();
-          do {
-            Node* node = NodeFromTreeIterator(tree_it);
-            typename Tree::iterator next = tree_it;
-            ++next;
-            tree->erase(tree_it);
-            DestroyNode(node);
-            tree_it = next;
-          } while (tree_it != tree->end());
-          DestroyTree(tree);
-          b++;
-        }
-      }
-      num_elements_ = 0;
-      index_of_first_non_null_ = num_buckets_;
-    }
-
-    const hasher& hash_function() const { return *this; }
-
-    static size_type max_size() {
-      return static_cast<size_type>(1) << (sizeof(void**) >= 8 ? 60 : 28);
-    }
-    size_type size() const { return num_elements_; }
-    bool empty() const { return size() == 0; }
-
-    template <typename K>
-    iterator find(const K& k) {
-      return iterator(FindHelper(k).first);
-    }
-
-    template <typename K>
-    const_iterator find(const K& k) const {
-      return FindHelper(k).first;
-    }
-
-    // Insert the key into the map, if not present. In that case, the value will
-    // be value initialized.
-    template <typename K>
-    std::pair<iterator, bool> insert(K&& k) {
-      std::pair<const_iterator, size_type> p = FindHelper(k);
-      // Case 1: key was already present.
-      if (p.first.node_ != nullptr)
-        return std::make_pair(iterator(p.first), false);
-      // Case 2: insert.
-      if (ResizeIfLoadIsOutOfRange(num_elements_ + 1)) {
-        p = FindHelper(k);
-      }
-      const size_type b = p.second;  // bucket number
-      Node* node;
-      // If K is not key_type, make the conversion to key_type explicit.
-      using TypeToInit = typename std::conditional<
-          std::is_same<typename std::decay<K>::type, key_type>::value, K&&,
-          key_type>::type;
-      if (alloc_.arena() == nullptr) {
-        node = new Node{value_type(static_cast<TypeToInit>(std::forward<K>(k))),
-                        nullptr};
-      } else {
-        node = Alloc<Node>(1);
-        Arena::CreateInArenaStorage(
-            const_cast<Key*>(&node->kv.first), alloc_.arena(),
-            static_cast<TypeToInit>(std::forward<K>(k)));
-        Arena::CreateInArenaStorage(&node->kv.second, alloc_.arena());
-      }
-
-      iterator result = InsertUnique(b, node);
-      ++num_elements_;
-      return std::make_pair(result, true);
-    }
-
-    template <typename K>
-    value_type& operator[](K&& k) {
-      return *insert(std::forward<K>(k)).first;
-    }
-
-    void erase(iterator it) {
-      GOOGLE_DCHECK_EQ(it.m_, this);
-      typename Tree::iterator tree_it;
-      const bool is_list = it.revalidate_if_necessary(&tree_it);
-      size_type b = it.bucket_index_;
-      Node* const item = it.node_;
-      if (is_list) {
-        GOOGLE_DCHECK(TableEntryIsNonEmptyList(b));
-        Node* head = static_cast<Node*>(table_[b]);
-        head = EraseFromLinkedList(item, head);
-        table_[b] = static_cast<void*>(head);
-      } else {
-        GOOGLE_DCHECK(TableEntryIsTree(b));
-        Tree* tree = static_cast<Tree*>(table_[b]);
-        tree->erase(tree_it);
-        if (tree->empty()) {
-          // Force b to be the minimum of b and b ^ 1.  This is important
-          // only because we want index_of_first_non_null_ to be correct.
-          b &= ~static_cast<size_type>(1);
-          DestroyTree(tree);
-          table_[b] = table_[b + 1] = nullptr;
-        }
-      }
-      DestroyNode(item);
-      --num_elements_;
-      if (PROTOBUF_PREDICT_FALSE(b == index_of_first_non_null_)) {
-        while (index_of_first_non_null_ < num_buckets_ &&
-               table_[index_of_first_non_null_] == nullptr) {
-          ++index_of_first_non_null_;
-        }
-      }
-    }
-
-    size_t SpaceUsedInternal() const {
-      return internal::SpaceUsedInTable<Key>(table_, num_buckets_,
-                                             num_elements_, sizeof(Node));
-    }
-
-   private:
-    const_iterator find(const Key& k, TreeIterator* it) const {
-      return FindHelper(k, it).first;
-    }
-    template <typename K>
-    std::pair<const_iterator, size_type> FindHelper(const K& k) const {
-      return FindHelper(k, nullptr);
-    }
-    template <typename K>
-    std::pair<const_iterator, size_type> FindHelper(const K& k,
-                                                    TreeIterator* it) const {
-      size_type b = BucketNumber(k);
-      if (TableEntryIsNonEmptyList(b)) {
-        Node* node = static_cast<Node*>(table_[b]);
-        do {
-          if (internal::TransparentSupport<Key>::Equals(node->kv.first, k)) {
-            return std::make_pair(const_iterator(node, this, b), b);
-          } else {
-            node = node->next;
-          }
-        } while (node != nullptr);
-      } else if (TableEntryIsTree(b)) {
-        GOOGLE_DCHECK_EQ(table_[b], table_[b ^ 1]);
-        b &= ~static_cast<size_t>(1);
-        Tree* tree = static_cast<Tree*>(table_[b]);
-        auto tree_it = tree->find(k);
-        if (tree_it != tree->end()) {
-          if (it != nullptr) *it = tree_it;
-          return std::make_pair(const_iterator(tree_it, this, b), b);
-        }
-      }
-      return std::make_pair(end(), b);
-    }
-
-    // Insert the given Node in bucket b.  If that would make bucket b too big,
-    // and bucket b is not a tree, create a tree for buckets b and b^1 to share.
-    // Requires count(*KeyPtrFromNodePtr(node)) == 0 and that b is the correct
-    // bucket.  num_elements_ is not modified.
-    iterator InsertUnique(size_type b, Node* node) {
-      GOOGLE_DCHECK(index_of_first_non_null_ == num_buckets_ ||
-             table_[index_of_first_non_null_] != nullptr);
-      // In practice, the code that led to this point may have already
-      // determined whether we are inserting into an empty list, a short list,
-      // or whatever.  But it's probably cheap enough to recompute that here;
-      // it's likely that we're inserting into an empty or short list.
-      iterator result;
-      GOOGLE_DCHECK(find(node->kv.first) == end());
-      if (TableEntryIsEmpty(b)) {
-        result = InsertUniqueInList(b, node);
-      } else if (TableEntryIsNonEmptyList(b)) {
-        if (PROTOBUF_PREDICT_FALSE(TableEntryIsTooLong(b))) {
-          TreeConvert(b);
-          result = InsertUniqueInTree(b, node);
-          GOOGLE_DCHECK_EQ(result.bucket_index_, b & ~static_cast<size_type>(1));
-        } else {
-          // Insert into a pre-existing list.  This case cannot modify
-          // index_of_first_non_null_, so we skip the code to update it.
-          return InsertUniqueInList(b, node);
-        }
-      } else {
-        // Insert into a pre-existing tree.  This case cannot modify
-        // index_of_first_non_null_, so we skip the code to update it.
-        return InsertUniqueInTree(b, node);
-      }
-      // parentheses around (std::min) prevents macro expansion of min(...)
-      index_of_first_non_null_ =
-          (std::min)(index_of_first_non_null_, result.bucket_index_);
-      return result;
-    }
-
-    // Returns whether we should insert after the head of the list. For
-    // non-optimized builds, we randomly decide whether to insert right at the
-    // head of the list or just after the head. This helps add a little bit of
-    // non-determinism to the map ordering.
-    bool ShouldInsertAfterHead(void* node) {
-#ifdef NDEBUG
-      (void) node;
-      return false;
-#else
-      // Doing modulo with a prime mixes the bits more.
-      return (reinterpret_cast<uintptr_t>(node) ^ seed_) % 13 > 6;
-#endif
-    }
-
-    // Helper for InsertUnique.  Handles the case where bucket b is a
-    // not-too-long linked list.
-    iterator InsertUniqueInList(size_type b, Node* node) {
-      if (table_[b] != nullptr && ShouldInsertAfterHead(node)) {
-        Node* first = static_cast<Node*>(table_[b]);
-        node->next = first->next;
-        first->next = node;
-        return iterator(node, this, b);
-      }
-
-      node->next = static_cast<Node*>(table_[b]);
-      table_[b] = static_cast<void*>(node);
-      return iterator(node, this, b);
-    }
-
-    // Helper for InsertUnique.  Handles the case where bucket b points to a
-    // Tree.
-    iterator InsertUniqueInTree(size_type b, Node* node) {
-      GOOGLE_DCHECK_EQ(table_[b], table_[b ^ 1]);
-      // Maintain the invariant that node->next is null for all Nodes in Trees.
-      node->next = nullptr;
-      return iterator(
-          static_cast<Tree*>(table_[b])->insert({node->kv.first, node}).first,
-          this, b & ~static_cast<size_t>(1));
-    }
-
-    // Returns whether it did resize.  Currently this is only used when
-    // num_elements_ increases, though it could be used in other situations.
-    // It checks for load too low as well as load too high: because any number
-    // of erases can occur between inserts, the load could be as low as 0 here.
-    // Resizing to a lower size is not always helpful, but failing to do so can
-    // destroy the expected big-O bounds for some operations. By having the
-    // policy that sometimes we resize down as well as up, clients can easily
-    // keep O(size()) = O(number of buckets) if they want that.
-    bool ResizeIfLoadIsOutOfRange(size_type new_size) {
-      const size_type kMaxMapLoadTimes16 = 12;  // controls RAM vs CPU tradeoff
-      const size_type hi_cutoff = num_buckets_ * kMaxMapLoadTimes16 / 16;
-      const size_type lo_cutoff = hi_cutoff / 4;
-      // We don't care how many elements are in trees.  If a lot are,
-      // we may resize even though there are many empty buckets.  In
-      // practice, this seems fine.
-      if (PROTOBUF_PREDICT_FALSE(new_size >= hi_cutoff)) {
-        if (num_buckets_ <= max_size() / 2) {
-          Resize(num_buckets_ * 2);
-          return true;
-        }
-      } else if (PROTOBUF_PREDICT_FALSE(new_size <= lo_cutoff &&
-                                        num_buckets_ > kMinTableSize)) {
-        size_type lg2_of_size_reduction_factor = 1;
-        // It's possible we want to shrink a lot here... size() could even be 0.
-        // So, estimate how much to shrink by making sure we don't shrink so
-        // much that we would need to grow the table after a few inserts.
-        const size_type hypothetical_size = new_size * 5 / 4 + 1;
-        while ((hypothetical_size << lg2_of_size_reduction_factor) <
-               hi_cutoff) {
-          ++lg2_of_size_reduction_factor;
-        }
-        size_type new_num_buckets = std::max<size_type>(
-            kMinTableSize, num_buckets_ >> lg2_of_size_reduction_factor);
-        if (new_num_buckets != num_buckets_) {
-          Resize(new_num_buckets);
-          return true;
-        }
-      }
-      return false;
-    }
-
-    // Resize to the given number of buckets.
-    void Resize(size_t new_num_buckets) {
-      if (num_buckets_ == internal::kGlobalEmptyTableSize) {
-        // This is the global empty array.
-        // Just overwrite with a new one. No need to transfer or free anything.
-        num_buckets_ = index_of_first_non_null_ = kMinTableSize;
-        table_ = CreateEmptyTable(num_buckets_);
-        seed_ = Seed();
-        return;
-      }
-
-      GOOGLE_DCHECK_GE(new_num_buckets, kMinTableSize);
-      void** const old_table = table_;
-      const size_type old_table_size = num_buckets_;
-      num_buckets_ = new_num_buckets;
-      table_ = CreateEmptyTable(num_buckets_);
-      const size_type start = index_of_first_non_null_;
-      index_of_first_non_null_ = num_buckets_;
-      for (size_type i = start; i < old_table_size; i++) {
-        if (internal::TableEntryIsNonEmptyList(old_table, i)) {
-          TransferList(old_table, i);
-        } else if (internal::TableEntryIsTree(old_table, i)) {
-          TransferTree(old_table, i++);
-        }
-      }
-      Dealloc<void*>(old_table, old_table_size);
-    }
-
-    void TransferList(void* const* table, size_type index) {
-      Node* node = static_cast<Node*>(table[index]);
-      do {
-        Node* next = node->next;
-        InsertUnique(BucketNumber(node->kv.first), node);
-        node = next;
-      } while (node != nullptr);
-    }
-
-    void TransferTree(void* const* table, size_type index) {
-      Tree* tree = static_cast<Tree*>(table[index]);
-      typename Tree::iterator tree_it = tree->begin();
-      do {
-        InsertUnique(BucketNumber(std::cref(tree_it->first).get()),
-                     NodeFromTreeIterator(tree_it));
-      } while (++tree_it != tree->end());
-      DestroyTree(tree);
-    }
-
-    Node* EraseFromLinkedList(Node* item, Node* head) {
-      if (head == item) {
-        return head->next;
-      } else {
-        head->next = EraseFromLinkedList(item, head->next);
-        return head;
-      }
-    }
-
-    bool TableEntryIsEmpty(size_type b) const {
-      return internal::TableEntryIsEmpty(table_, b);
-    }
-    bool TableEntryIsNonEmptyList(size_type b) const {
-      return internal::TableEntryIsNonEmptyList(table_, b);
-    }
-    bool TableEntryIsTree(size_type b) const {
-      return internal::TableEntryIsTree(table_, b);
-    }
-    bool TableEntryIsList(size_type b) const {
-      return internal::TableEntryIsList(table_, b);
-    }
-
-    void TreeConvert(size_type b) {
-      GOOGLE_DCHECK(!TableEntryIsTree(b) && !TableEntryIsTree(b ^ 1));
-      Tree* tree =
-          Arena::Create<Tree>(alloc_.arena(), typename Tree::key_compare(),
-                              typename Tree::allocator_type(alloc_));
-      size_type count = CopyListToTree(b, tree) + CopyListToTree(b ^ 1, tree);
-      GOOGLE_DCHECK_EQ(count, tree->size());
-      table_[b] = table_[b ^ 1] = static_cast<void*>(tree);
-    }
-
-    // Copy a linked list in the given bucket to a tree.
-    // Returns the number of things it copied.
-    size_type CopyListToTree(size_type b, Tree* tree) {
-      size_type count = 0;
-      Node* node = static_cast<Node*>(table_[b]);
-      while (node != nullptr) {
-        tree->insert({node->kv.first, node});
-        ++count;
-        Node* next = node->next;
-        node->next = nullptr;
-        node = next;
-      }
-      return count;
-    }
-
-    // Return whether table_[b] is a linked list that seems awfully long.
-    // Requires table_[b] to point to a non-empty linked list.
-    bool TableEntryIsTooLong(size_type b) {
-      const size_type kMaxLength = 8;
-      size_type count = 0;
-      Node* node = static_cast<Node*>(table_[b]);
-      do {
-        ++count;
-        node = node->next;
-      } while (node != nullptr);
-      // Invariant: no linked list ever is more than kMaxLength in length.
-      GOOGLE_DCHECK_LE(count, kMaxLength);
-      return count >= kMaxLength;
-    }
-
-    template <typename K>
-    size_type BucketNumber(const K& k) const {
-      // We xor the hash value against the random seed so that we effectively
-      // have a random hash function.
-      uint64 h = hash_function()(k) ^ seed_;
-
-      // We use the multiplication method to determine the bucket number from
-      // the hash value. The constant kPhi (suggested by Knuth) is roughly
-      // (sqrt(5) - 1) / 2 * 2^64.
-      constexpr uint64 kPhi = uint64{0x9e3779b97f4a7c15};
-      return ((kPhi * h) >> 32) & (num_buckets_ - 1);
-    }
-
-    // Return a power of two no less than max(kMinTableSize, n).
-    // Assumes either n < kMinTableSize or n is a power of two.
-    size_type TableSize(size_type n) {
-      return n < static_cast<size_type>(kMinTableSize)
-                 ? static_cast<size_type>(kMinTableSize)
-                 : n;
-    }
-
-    // Use alloc_ to allocate an array of n objects of type U.
-    template <typename U>
-    U* Alloc(size_type n) {
-      using alloc_type = typename Allocator::template rebind<U>::other;
-      return alloc_type(alloc_).allocate(n);
-    }
-
-    // Use alloc_ to deallocate an array of n objects of type U.
-    template <typename U>
-    void Dealloc(U* t, size_type n) {
-      using alloc_type = typename Allocator::template rebind<U>::other;
-      alloc_type(alloc_).deallocate(t, n);
-    }
-
-    void DestroyNode(Node* node) {
-      if (alloc_.arena() == nullptr) {
-        delete node;
-      }
-    }
-
-    void DestroyTree(Tree* tree) {
-      if (alloc_.arena() == nullptr) {
-        delete tree;
-      }
-    }
-
-    void** CreateEmptyTable(size_type n) {
-      GOOGLE_DCHECK(n >= kMinTableSize);
-      GOOGLE_DCHECK_EQ(n & (n - 1), 0);
-      void** result = Alloc<void*>(n);
-      memset(result, 0, n * sizeof(result[0]));
-      return result;
-    }
-
-    // Return a randomish value.
-    size_type Seed() const {
-      // We get a little bit of randomness from the address of the map. The
-      // lower bits are not very random, due to alignment, so we discard them
-      // and shift the higher bits into their place.
-      size_type s = reinterpret_cast<uintptr_t>(this) >> 12;
-#if defined(__x86_64__) && defined(__GNUC__) && \
-    !defined(GOOGLE_PROTOBUF_NO_RDTSC)
-      uint32 hi, lo;
-      asm volatile("rdtsc" : "=a"(lo), "=d"(hi));
-      s += ((static_cast<uint64>(hi) << 32) | lo);
-#endif
-      return s;
-    }
-
-    friend class Arena;
-    using InternalArenaConstructable_ = void;
-    using DestructorSkippable_ = void;
-
-    size_type num_elements_;
-    size_type num_buckets_;
-    size_type seed_;
-    size_type index_of_first_non_null_;
-    void** table_;  // an array with num_buckets_ entries
-    Allocator alloc_;
-    GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(InnerMap);
-  };  // end of class InnerMap
-
-  template <typename LookupKey>
-  using key_arg = typename internal::TransparentSupport<
-      key_type>::template key_arg<LookupKey>;
-
- public:
-  // Iterators
-  class const_iterator {
-    using InnerIt = typename InnerMap::const_iterator;
-
-   public:
-    using iterator_category = std::forward_iterator_tag;
-    using value_type = typename Map::value_type;
-    using difference_type = ptrdiff_t;
-    using pointer = const value_type*;
-    using reference = const value_type&;
-
-    const_iterator() {}
-    explicit const_iterator(const InnerIt& it) : it_(it) {}
-
-    const_reference operator*() const { return *it_; }
-    const_pointer operator->() const { return &(operator*()); }
-
-    const_iterator& operator++() {
-      ++it_;
-      return *this;
-    }
-    const_iterator operator++(int) { return const_iterator(it_++); }
-
-    friend bool operator==(const const_iterator& a, const const_iterator& b) {
-      return a.it_ == b.it_;
-    }
-    friend bool operator!=(const const_iterator& a, const const_iterator& b) {
-      return !(a == b);
-    }
-
-   private:
-    InnerIt it_;
-  };
-
-  class iterator {
-    using InnerIt = typename InnerMap::iterator;
-
-   public:
-    using iterator_category = std::forward_iterator_tag;
-    using value_type = typename Map::value_type;
-    using difference_type = ptrdiff_t;
-    using pointer = value_type*;
-    using reference = value_type&;
-
-    iterator() {}
-    explicit iterator(const InnerIt& it) : it_(it) {}
-
-    reference operator*() const { return *it_; }
-    pointer operator->() const { return &(operator*()); }
-
-    iterator& operator++() {
-      ++it_;
-      return *this;
-    }
-    iterator operator++(int) { return iterator(it_++); }
-
-    // Allow implicit conversion to const_iterator.
-    operator const_iterator() const {  // NOLINT(runtime/explicit)
-      return const_iterator(typename InnerMap::const_iterator(it_));
-    }
-
-    friend bool operator==(const iterator& a, const iterator& b) {
-      return a.it_ == b.it_;
-    }
-    friend bool operator!=(const iterator& a, const iterator& b) {
-      return !(a == b);
-    }
-
-   private:
-    friend class Map;
-
-    InnerIt it_;
-  };
-
-  iterator begin() { return iterator(elements_.begin()); }
-  iterator end() { return iterator(elements_.end()); }
-  const_iterator begin() const { return const_iterator(elements_.begin()); }
-  const_iterator end() const { return const_iterator(elements_.end()); }
-  const_iterator cbegin() const { return begin(); }
-  const_iterator cend() const { return end(); }
-
-  // Capacity
-  size_type size() const { return elements_.size(); }
-  bool empty() const { return size() == 0; }
-
-  // Element access
-  template <typename K = key_type>
-  T& operator[](const key_arg<K>& key) {
-    return elements_[key].second;
-  }
-  template <
-      typename K = key_type,
-      // Disable for integral types to reduce code bloat.
-      typename = typename std::enable_if<!std::is_integral<K>::value>::type>
-  T& operator[](key_arg<K>&& key) {
-    return elements_[std::forward<K>(key)].second;
-  }
-
-  template <typename K = key_type>
-  const T& at(const key_arg<K>& key) const {
-    const_iterator it = find(key);
-    GOOGLE_CHECK(it != end()) << "key not found: " << static_cast<Key>(key);
-    return it->second;
-  }
-
-  template <typename K = key_type>
-  T& at(const key_arg<K>& key) {
-    iterator it = find(key);
-    GOOGLE_CHECK(it != end()) << "key not found: " << static_cast<Key>(key);
-    return it->second;
-  }
-
-  // Lookup
-  template <typename K = key_type>
-  size_type count(const key_arg<K>& key) const {
-    return find(key) == end() ? 0 : 1;
-  }
-
-  template <typename K = key_type>
-  const_iterator find(const key_arg<K>& key) const {
-    return const_iterator(elements_.find(key));
-  }
-  template <typename K = key_type>
-  iterator find(const key_arg<K>& key) {
-    return iterator(elements_.find(key));
-  }
-
-  template <typename K = key_type>
-  bool contains(const key_arg<K>& key) const {
-    return find(key) != end();
-  }
-
-  template <typename K = key_type>
-  std::pair<const_iterator, const_iterator> equal_range(
-      const key_arg<K>& key) const {
-    const_iterator it = find(key);
-    if (it == end()) {
-      return std::pair<const_iterator, const_iterator>(it, it);
-    } else {
-      const_iterator begin = it++;
-      return std::pair<const_iterator, const_iterator>(begin, it);
-    }
-  }
-
-  template <typename K = key_type>
-  std::pair<iterator, iterator> equal_range(const key_arg<K>& key) {
-    iterator it = find(key);
-    if (it == end()) {
-      return std::pair<iterator, iterator>(it, it);
-    } else {
-      iterator begin = it++;
-      return std::pair<iterator, iterator>(begin, it);
-    }
-  }
-
-  // insert
-  std::pair<iterator, bool> insert(const value_type& value) {
-    std::pair<typename InnerMap::iterator, bool> p =
-        elements_.insert(value.first);
-    if (p.second) {
-      p.first->second = value.second;
-    }
-    return std::pair<iterator, bool>(iterator(p.first), p.second);
-  }
-  template <class InputIt>
-  void insert(InputIt first, InputIt last) {
-    for (InputIt it = first; it != last; ++it) {
-      iterator exist_it = find(it->first);
-      if (exist_it == end()) {
-        operator[](it->first) = it->second;
-      }
-    }
-  }
-  void insert(std::initializer_list<value_type> values) {
-    insert(values.begin(), values.end());
-  }
-
-  // Erase and clear
-  template <typename K = key_type>
-  size_type erase(const key_arg<K>& key) {
-    iterator it = find(key);
-    if (it == end()) {
-      return 0;
-    } else {
-      erase(it);
-      return 1;
-    }
-  }
-  iterator erase(iterator pos) {
-    iterator i = pos++;
-    elements_.erase(i.it_);
-    return pos;
-  }
-  void erase(iterator first, iterator last) {
-    while (first != last) {
-      first = erase(first);
-    }
-  }
-  void clear() { elements_.clear(); }
-
-  // Assign
-  Map& operator=(const Map& other) {
-    if (this != &other) {
-      clear();
-      insert(other.begin(), other.end());
-    }
-    return *this;
-  }
-
-  void swap(Map& other) {
-    if (arena() == other.arena()) {
-      elements_.Swap(&other.elements_);
-    } else {
-      // TODO(zuguang): optimize this. The temporary copy can be allocated
-      // in the same arena as the other message, and the "other = copy" can
-      // be replaced with the fast-path swap above.
-      Map copy = *this;
-      *this = other;
-      other = copy;
-    }
-  }
-
-  // Access to hasher.  Currently this returns a copy, but it may
-  // be modified to return a const reference in the future.
-  hasher hash_function() const { return elements_.hash_function(); }
-
-  size_t SpaceUsedExcludingSelfLong() const {
-    if (empty()) return 0;
-    return elements_.SpaceUsedInternal() + internal::SpaceUsedInValues(this);
-  }
-
- private:
-  Arena* arena() const { return elements_.arena(); }
-  InnerMap elements_;
-
-  friend class Arena;
-  using InternalArenaConstructable_ = void;
-  using DestructorSkippable_ = void;
-  template <typename Derived, typename K, typename V,
-            internal::WireFormatLite::FieldType key_wire_type,
-            internal::WireFormatLite::FieldType value_wire_type>
-  friend class internal::MapFieldLite;
-};
-
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_MAP_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/map_entry_lite.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/map_entry_lite.h
deleted file mode 100644
index 1caf59df..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/map_entry_lite.h
+++ /dev/null
@@ -1,653 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_MAP_ENTRY_LITE_H__
-#define GOOGLE_PROTOBUF_MAP_ENTRY_LITE_H__
-
-#include <assert.h>
-#include <string>
-
-#include <google/protobuf/stubs/casts.h>
-#include <google/protobuf/parse_context.h>
-#include <google/protobuf/io/coded_stream.h>
-#include <google/protobuf/arena.h>
-#include <google/protobuf/arenastring.h>
-#include <google/protobuf/generated_message_util.h>
-#include <google/protobuf/map.h>
-#include <google/protobuf/map_type_handler.h>
-#include <google/protobuf/port.h>
-#include <google/protobuf/wire_format_lite.h>
-
-#include <google/protobuf/port_def.inc>
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-namespace google {
-namespace protobuf {
-namespace internal {
-template <typename Derived, typename Key, typename Value,
-          WireFormatLite::FieldType kKeyFieldType,
-          WireFormatLite::FieldType kValueFieldType>
-class MapEntry;
-template <typename Derived, typename Key, typename Value,
-          WireFormatLite::FieldType kKeyFieldType,
-          WireFormatLite::FieldType kValueFieldType>
-class MapFieldLite;
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-// MoveHelper::Move is used to set *dest.  It copies *src, or moves it (in
-// the C++11 sense), or swaps it. *src is left in a sane state for
-// subsequent destruction, but shouldn't be used for anything.
-template <bool is_enum, bool is_message, bool is_stringlike, typename T>
-struct MoveHelper {  // primitives
-  static void Move(T* src, T* dest) { *dest = *src; }
-};
-
-template <bool is_message, bool is_stringlike, typename T>
-struct MoveHelper<true, is_message, is_stringlike, T> {  // enums
-  static void Move(T* src, T* dest) { *dest = *src; }
-  // T is an enum here, so allow conversions to and from int.
-  static void Move(T* src, int* dest) { *dest = static_cast<int>(*src); }
-  static void Move(int* src, T* dest) { *dest = static_cast<T>(*src); }
-};
-
-template <bool is_stringlike, typename T>
-struct MoveHelper<false, true, is_stringlike, T> {  // messages
-  static void Move(T* src, T* dest) { dest->Swap(src); }
-};
-
-template <typename T>
-struct MoveHelper<false, false, true, T> {  // strings and similar
-  static void Move(T* src, T* dest) {
-    *dest = std::move(*src);
-  }
-};
-
-// Functions for operating on a map entry.  Does not contain any representation
-// (this class is not intended to be instantiated).
-template <typename Key, typename Value, WireFormatLite::FieldType kKeyFieldType,
-          WireFormatLite::FieldType kValueFieldType>
-struct MapEntryFuncs {
-  typedef MapTypeHandler<kKeyFieldType, Key> KeyTypeHandler;
-  typedef MapTypeHandler<kValueFieldType, Value> ValueTypeHandler;
-  static const int kKeyFieldNumber = 1;
-  static const int kValueFieldNumber = 2;
-
-  static uint8* InternalSerialize(int field_number, const Key& key,
-                                  const Value& value, uint8* ptr,
-                                  io::EpsCopyOutputStream* stream) {
-    ptr = stream->EnsureSpace(ptr);
-    ptr = WireFormatLite::WriteTagToArray(
-        field_number, WireFormatLite::WIRETYPE_LENGTH_DELIMITED, ptr);
-    ptr = io::CodedOutputStream::WriteVarint32ToArray(GetCachedSize(key, value),
-                                                      ptr);
-
-    ptr = KeyTypeHandler::Write(kKeyFieldNumber, key, ptr, stream);
-    return ValueTypeHandler::Write(kValueFieldNumber, value, ptr, stream);
-  }
-
-  static size_t ByteSizeLong(const Key& key, const Value& value) {
-    // Tags for key and value will both be one byte (field numbers 1 and 2).
-    size_t inner_length =
-        2 + KeyTypeHandler::ByteSize(key) + ValueTypeHandler::ByteSize(value);
-    return inner_length + io::CodedOutputStream::VarintSize32(
-                              static_cast<uint32>(inner_length));
-  }
-
-  static int GetCachedSize(const Key& key, const Value& value) {
-    // Tags for key and value will both be one byte (field numbers 1 and 2).
-    return 2 + KeyTypeHandler::GetCachedSize(key) +
-           ValueTypeHandler::GetCachedSize(value);
-  }
-};
-
-// MapEntryImpl is used to implement parsing and serialization of map entries.
-// It uses Curious Recursive Template Pattern (CRTP) to provide the type of
-// the eventual code to the template code.
-template <typename Derived, typename Base, typename Key, typename Value,
-          WireFormatLite::FieldType kKeyFieldType,
-          WireFormatLite::FieldType kValueFieldType>
-class MapEntryImpl : public Base {
- public:
-  typedef MapEntryFuncs<Key, Value, kKeyFieldType, kValueFieldType> Funcs;
-
- protected:
-  // Provide utilities to parse/serialize key/value.  Provide utilities to
-  // manipulate internal stored type.
-  typedef MapTypeHandler<kKeyFieldType, Key> KeyTypeHandler;
-  typedef MapTypeHandler<kValueFieldType, Value> ValueTypeHandler;
-
-  // Define internal memory layout. Strings and messages are stored as
-  // pointers, while other types are stored as values.
-  typedef typename KeyTypeHandler::TypeOnMemory KeyOnMemory;
-  typedef typename ValueTypeHandler::TypeOnMemory ValueOnMemory;
-
-  // Enum type cannot be used for MapTypeHandler::Read. Define a type
-  // which will replace Enum with int.
-  typedef typename KeyTypeHandler::MapEntryAccessorType KeyMapEntryAccessorType;
-  typedef
-      typename ValueTypeHandler::MapEntryAccessorType ValueMapEntryAccessorType;
-
-  // Constants for field number.
-  static const int kKeyFieldNumber = 1;
-  static const int kValueFieldNumber = 2;
-
-  // Constants for field tag.
-  static const uint8 kKeyTag =
-      GOOGLE_PROTOBUF_WIRE_FORMAT_MAKE_TAG(kKeyFieldNumber, KeyTypeHandler::kWireType);
-  static const uint8 kValueTag = GOOGLE_PROTOBUF_WIRE_FORMAT_MAKE_TAG(
-      kValueFieldNumber, ValueTypeHandler::kWireType);
-  static const size_t kTagSize = 1;
-
- public:
-  // Work-around for a compiler bug (see repeated_field.h).
-  typedef void MapEntryHasMergeTypeTrait;
-  typedef Derived EntryType;
-  typedef Key EntryKeyType;
-  typedef Value EntryValueType;
-  static const WireFormatLite::FieldType kEntryKeyFieldType = kKeyFieldType;
-  static const WireFormatLite::FieldType kEntryValueFieldType = kValueFieldType;
-
-  constexpr MapEntryImpl()
-      : key_(KeyTypeHandler::Constinit()),
-        value_(ValueTypeHandler::Constinit()),
-        _has_bits_{} {}
-
-  explicit MapEntryImpl(Arena* arena)
-      : Base(arena),
-        key_(KeyTypeHandler::Constinit()),
-        value_(ValueTypeHandler::Constinit()),
-        _has_bits_{} {}
-
-  ~MapEntryImpl() {
-    if (Base::GetArena() != NULL) return;
-    KeyTypeHandler::DeleteNoArena(key_);
-    ValueTypeHandler::DeleteNoArena(value_);
-  }
-
-  // accessors ======================================================
-
-  virtual inline const KeyMapEntryAccessorType& key() const {
-    return KeyTypeHandler::GetExternalReference(key_);
-  }
-  virtual inline const ValueMapEntryAccessorType& value() const {
-    return ValueTypeHandler::DefaultIfNotInitialized(value_);
-  }
-  inline KeyMapEntryAccessorType* mutable_key() {
-    set_has_key();
-    return KeyTypeHandler::EnsureMutable(&key_, Base::GetArena());
-  }
-  inline ValueMapEntryAccessorType* mutable_value() {
-    set_has_value();
-    return ValueTypeHandler::EnsureMutable(&value_, Base::GetArena());
-  }
-
-  // implements MessageLite =========================================
-
-  // MapEntryImpl is for implementation only and this function isn't called
-  // anywhere. Just provide a fake implementation here for MessageLite.
-  std::string GetTypeName() const override { return ""; }
-
-  void CheckTypeAndMergeFrom(const MessageLite& other) override {
-    MergeFromInternal(*::google::protobuf::internal::DownCast<const Derived*>(&other));
-  }
-
-  const char* _InternalParse(const char* ptr, ParseContext* ctx) final {
-    while (!ctx->Done(&ptr)) {
-      uint32 tag;
-      ptr = ReadTag(ptr, &tag);
-      GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-      if (tag == kKeyTag) {
-        set_has_key();
-        KeyMapEntryAccessorType* key = mutable_key();
-        ptr = KeyTypeHandler::Read(ptr, ctx, key);
-        if (!Derived::ValidateKey(key)) return nullptr;
-      } else if (tag == kValueTag) {
-        set_has_value();
-        ValueMapEntryAccessorType* value = mutable_value();
-        ptr = ValueTypeHandler::Read(ptr, ctx, value);
-        if (!Derived::ValidateValue(value)) return nullptr;
-      } else {
-        if (tag == 0 || WireFormatLite::GetTagWireType(tag) ==
-                            WireFormatLite::WIRETYPE_END_GROUP) {
-          ctx->SetLastTag(tag);
-          return ptr;
-        }
-        ptr = UnknownFieldParse(tag, static_cast<std::string*>(nullptr), ptr,
-                                ctx);
-      }
-      GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-    }
-    return ptr;
-  }
-
-  size_t ByteSizeLong() const override {
-    size_t size = 0;
-    size += kTagSize + static_cast<size_t>(KeyTypeHandler::ByteSize(key()));
-    size += kTagSize + static_cast<size_t>(ValueTypeHandler::ByteSize(value()));
-    return size;
-  }
-
-  ::google::protobuf::uint8* _InternalSerialize(::google::protobuf::uint8* ptr,
-                              io::EpsCopyOutputStream* stream) const override {
-    ptr = KeyTypeHandler::Write(kKeyFieldNumber, key(), ptr, stream);
-    return ValueTypeHandler::Write(kValueFieldNumber, value(), ptr, stream);
-  }
-
-  // Don't override SerializeWithCachedSizesToArray.  Use MessageLite's.
-
-  int GetCachedSize() const override {
-    int size = 0;
-    size += has_key() ? static_cast<int>(kTagSize) +
-                            KeyTypeHandler::GetCachedSize(key())
-                      : 0;
-    size += has_value() ? static_cast<int>(kTagSize) +
-                              ValueTypeHandler::GetCachedSize(value())
-                        : 0;
-    return size;
-  }
-
-  bool IsInitialized() const override {
-    return ValueTypeHandler::IsInitialized(value_);
-  }
-
-  Base* New() const override {
-    Derived* entry = new Derived;
-    return entry;
-  }
-
-  Base* New(Arena* arena) const override {
-    Derived* entry = Arena::CreateMessage<Derived>(arena);
-    return entry;
-  }
-
- protected:
-  // We can't declare this function directly here as it would hide the other
-  // overload (const Message&).
-  void MergeFromInternal(const MapEntryImpl& from) {
-    if (from._has_bits_[0]) {
-      if (from.has_key()) {
-        KeyTypeHandler::EnsureMutable(&key_, Base::GetArena());
-        KeyTypeHandler::Merge(from.key(), &key_, Base::GetArena());
-        set_has_key();
-      }
-      if (from.has_value()) {
-        ValueTypeHandler::EnsureMutable(&value_, Base::GetArena());
-        ValueTypeHandler::Merge(from.value(), &value_, Base::GetArena());
-        set_has_value();
-      }
-    }
-  }
-
- public:
-  void Clear() override {
-    KeyTypeHandler::Clear(&key_, Base::GetArena());
-    ValueTypeHandler::Clear(&value_, Base::GetArena());
-    clear_has_key();
-    clear_has_value();
-  }
-
-  // Parsing using MergePartialFromCodedStream, above, is not as
-  // efficient as it could be.  This helper class provides a speedier way.
-  template <typename MapField, typename Map>
-  class Parser {
-   public:
-    explicit Parser(MapField* mf) : mf_(mf), map_(mf->MutableMap()) {}
-    ~Parser() {
-      if (entry_ != nullptr && entry_->GetArena() == nullptr) delete entry_;
-    }
-
-    // This does what the typical MergePartialFromCodedStream() is expected to
-    // do, with the additional side-effect that if successful (i.e., if true is
-    // going to be its return value) it inserts the key-value pair into map_.
-    bool MergePartialFromCodedStream(io::CodedInputStream* input) {
-      // Look for the expected thing: a key and then a value.  If it fails,
-      // invoke the enclosing class's MergePartialFromCodedStream, or return
-      // false if that would be pointless.
-      if (input->ExpectTag(kKeyTag)) {
-        if (!KeyTypeHandler::Read(input, &key_)) {
-          return false;
-        }
-        // Peek at the next byte to see if it is kValueTag.  If not, bail out.
-        const void* data;
-        int size;
-        input->GetDirectBufferPointerInline(&data, &size);
-        // We could use memcmp here, but we don't bother. The tag is one byte.
-        static_assert(kTagSize == 1, "tag size must be 1");
-        if (size > 0 && *reinterpret_cast<const char*>(data) == kValueTag) {
-          typename Map::size_type map_size = map_->size();
-          value_ptr_ = &(*map_)[key_];
-          if (PROTOBUF_PREDICT_TRUE(map_size != map_->size())) {
-            // We created a new key-value pair.  Fill in the value.
-            typedef
-                typename MapIf<ValueTypeHandler::kIsEnum, int*, Value*>::type T;
-            input->Skip(kTagSize);  // Skip kValueTag.
-            if (!ValueTypeHandler::Read(input,
-                                        reinterpret_cast<T>(value_ptr_))) {
-              map_->erase(key_);  // Failure! Undo insertion.
-              return false;
-            }
-            if (input->ExpectAtEnd()) return true;
-            return ReadBeyondKeyValuePair(input);
-          }
-        }
-      } else {
-        key_ = Key();
-      }
-
-      NewEntry();
-      *entry_->mutable_key() = key_;
-      const bool result = entry_->MergePartialFromCodedStream(input);
-      if (result) UseKeyAndValueFromEntry();
-      return result;
-    }
-
-    const char* _InternalParse(const char* ptr, ParseContext* ctx) {
-      if (PROTOBUF_PREDICT_TRUE(!ctx->Done(&ptr) && *ptr == kKeyTag)) {
-        ptr = KeyTypeHandler::Read(ptr + 1, ctx, &key_);
-        if (PROTOBUF_PREDICT_FALSE(!ptr || !Derived::ValidateKey(&key_))) {
-          return nullptr;
-        }
-        if (PROTOBUF_PREDICT_TRUE(!ctx->Done(&ptr) && *ptr == kValueTag)) {
-          typename Map::size_type map_size = map_->size();
-          value_ptr_ = &(*map_)[key_];
-          if (PROTOBUF_PREDICT_TRUE(map_size != map_->size())) {
-            using T =
-                typename MapIf<ValueTypeHandler::kIsEnum, int*, Value*>::type;
-            ptr = ValueTypeHandler::Read(ptr + 1, ctx,
-                                         reinterpret_cast<T>(value_ptr_));
-            if (PROTOBUF_PREDICT_FALSE(!ptr ||
-                                       !Derived::ValidateValue(value_ptr_))) {
-              map_->erase(key_);  // Failure! Undo insertion.
-              return nullptr;
-            }
-            if (PROTOBUF_PREDICT_TRUE(ctx->Done(&ptr))) return ptr;
-            if (!ptr) return nullptr;
-            NewEntry();
-            ValueMover::Move(value_ptr_, entry_->mutable_value());
-            map_->erase(key_);
-            goto move_key;
-          }
-        } else {
-          if (!ptr) return nullptr;
-        }
-        NewEntry();
-      move_key:
-        KeyMover::Move(&key_, entry_->mutable_key());
-      } else {
-        if (!ptr) return nullptr;
-        NewEntry();
-      }
-      ptr = entry_->_InternalParse(ptr, ctx);
-      if (ptr) UseKeyAndValueFromEntry();
-      return ptr;
-    }
-
-    template <typename UnknownType>
-    const char* ParseWithEnumValidation(const char* ptr, ParseContext* ctx,
-                                        bool (*is_valid)(int), uint32 field_num,
-                                        InternalMetadata* metadata) {
-      auto entry = NewEntry();
-      ptr = entry->_InternalParse(ptr, ctx);
-      if (!ptr) return nullptr;
-      if (is_valid(entry->value())) {
-        UseKeyAndValueFromEntry();
-      } else {
-        WriteLengthDelimited(field_num, entry->SerializeAsString(),
-                             metadata->mutable_unknown_fields<UnknownType>());
-      }
-      return ptr;
-    }
-
-    MapEntryImpl* NewEntry() { return entry_ = mf_->NewEntry(); }
-
-    const Key& key() const { return key_; }
-    const Value& value() const { return *value_ptr_; }
-
-    const Key& entry_key() const { return entry_->key(); }
-    const Value& entry_value() const { return entry_->value(); }
-
-   private:
-    void UseKeyAndValueFromEntry() {
-      // Update key_ in case we need it later (because key() is called).
-      // This is potentially inefficient, especially if the key is
-      // expensive to copy (e.g., a long string), but this is a cold
-      // path, so it's not a big deal.
-      key_ = entry_->key();
-      value_ptr_ = &(*map_)[key_];
-      ValueMover::Move(entry_->mutable_value(), value_ptr_);
-    }
-
-    // After reading a key and value successfully, and inserting that data
-    // into map_, we are not at the end of the input.  This is unusual, but
-    // allowed by the spec.
-    bool ReadBeyondKeyValuePair(io::CodedInputStream* input) PROTOBUF_COLD {
-      NewEntry();
-      ValueMover::Move(value_ptr_, entry_->mutable_value());
-      map_->erase(key_);
-      KeyMover::Move(&key_, entry_->mutable_key());
-      const bool result = entry_->MergePartialFromCodedStream(input);
-      if (result) UseKeyAndValueFromEntry();
-      return result;
-    }
-
-    typedef MoveHelper<KeyTypeHandler::kIsEnum, KeyTypeHandler::kIsMessage,
-                       KeyTypeHandler::kWireType ==
-                           WireFormatLite::WIRETYPE_LENGTH_DELIMITED,
-                       Key>
-        KeyMover;
-    typedef MoveHelper<ValueTypeHandler::kIsEnum, ValueTypeHandler::kIsMessage,
-                       ValueTypeHandler::kWireType ==
-                           WireFormatLite::WIRETYPE_LENGTH_DELIMITED,
-                       Value>
-        ValueMover;
-
-    MapField* const mf_;
-    Map* const map_;
-    Key key_;
-    Value* value_ptr_;
-    MapEntryImpl* entry_ = nullptr;
-  };
-
- protected:
-  void set_has_key() { _has_bits_[0] |= 0x00000001u; }
-  bool has_key() const { return (_has_bits_[0] & 0x00000001u) != 0; }
-  void clear_has_key() { _has_bits_[0] &= ~0x00000001u; }
-  void set_has_value() { _has_bits_[0] |= 0x00000002u; }
-  bool has_value() const { return (_has_bits_[0] & 0x00000002u) != 0; }
-  void clear_has_value() { _has_bits_[0] &= ~0x00000002u; }
-
- public:
-  inline Arena* GetArena() const { return Base::GetArena(); }
-
- public:  // Needed for constructing tables
-  KeyOnMemory key_;
-  ValueOnMemory value_;
-  uint32 _has_bits_[1];
-
- private:
-  friend class ::PROTOBUF_NAMESPACE_ID::Arena;
-  typedef void InternalArenaConstructable_;
-  typedef void DestructorSkippable_;
-  template <typename C, typename K, typename V, WireFormatLite::FieldType,
-            WireFormatLite::FieldType>
-  friend class internal::MapEntry;
-  template <typename C, typename K, typename V, WireFormatLite::FieldType,
-            WireFormatLite::FieldType>
-  friend class internal::MapFieldLite;
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(MapEntryImpl);
-};
-
-template <typename T, typename Key, typename Value,
-          WireFormatLite::FieldType kKeyFieldType,
-          WireFormatLite::FieldType kValueFieldType>
-class MapEntryLite : public MapEntryImpl<T, MessageLite, Key, Value,
-                                         kKeyFieldType, kValueFieldType> {
- public:
-  typedef MapEntryImpl<T, MessageLite, Key, Value, kKeyFieldType,
-                       kValueFieldType>
-      SuperType;
-  constexpr MapEntryLite() {}
-  explicit MapEntryLite(Arena* arena) : SuperType(arena) {}
-  ~MapEntryLite() { MessageLite::_internal_metadata_.Delete<std::string>(); }
-  void MergeFrom(const MapEntryLite& other) { MergeFromInternal(other); }
-
- private:
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(MapEntryLite);
-};
-// The completely unprincipled and unwieldy use of template parameters in
-// the map code necessitates wrappers to make the code a little bit more
-// manageable.
-template <typename Derived>
-struct DeconstructMapEntry;
-
-template <typename T, typename K, typename V, WireFormatLite::FieldType key,
-          WireFormatLite::FieldType value>
-struct DeconstructMapEntry<MapEntryLite<T, K, V, key, value> > {
-  typedef K Key;
-  typedef V Value;
-  static const WireFormatLite::FieldType kKeyFieldType = key;
-  static const WireFormatLite::FieldType kValueFieldType = value;
-};
-
-// Helpers for deterministic serialization =============================
-
-// This struct can be used with any generic sorting algorithm.  If the Key
-// type is relatively small and easy to copy then copying Keys into an
-// array of SortItems can be beneficial.  Then all the data the sorting
-// algorithm needs to touch is in that one array.
-template <typename Key, typename PtrToKeyValuePair>
-struct SortItem {
-  SortItem() {}
-  explicit SortItem(PtrToKeyValuePair p) : first(p->first), second(p) {}
-
-  Key first;
-  PtrToKeyValuePair second;
-};
-
-template <typename T>
-struct CompareByFirstField {
-  bool operator()(const T& a, const T& b) const { return a.first < b.first; }
-};
-
-template <typename T>
-struct CompareByDerefFirst {
-  bool operator()(const T& a, const T& b) const { return a->first < b->first; }
-};
-
-// Helper for table driven serialization
-
-template <WireFormatLite::FieldType FieldType>
-struct FromHelper {
-  template <typename T>
-  static const T& From(const T& x) {
-    return x;
-  }
-};
-
-template <>
-struct FromHelper<WireFormatLite::TYPE_STRING> {
-  static ArenaStringPtr From(const std::string& x) {
-    ArenaStringPtr res;
-    TaggedPtr<std::string> ptr;
-    ptr.Set(const_cast<std::string*>(&x));
-    res.UnsafeSetTaggedPointer(ptr);
-    return res;
-  }
-};
-template <>
-struct FromHelper<WireFormatLite::TYPE_BYTES> {
-  static ArenaStringPtr From(const std::string& x) {
-    ArenaStringPtr res;
-    TaggedPtr<std::string> ptr;
-    ptr.Set(const_cast<std::string*>(&x));
-    res.UnsafeSetTaggedPointer(ptr);
-    return res;
-  }
-};
-template <>
-struct FromHelper<WireFormatLite::TYPE_MESSAGE> {
-  template <typename T>
-  static T* From(const T& x) {
-    return const_cast<T*>(&x);
-  }
-};
-
-template <typename MapEntryType>
-struct MapEntryHelper;
-
-template <typename T, typename Key, typename Value,
-          WireFormatLite::FieldType kKeyFieldType,
-          WireFormatLite::FieldType kValueFieldType>
-struct MapEntryHelper<
-    MapEntryLite<T, Key, Value, kKeyFieldType, kValueFieldType> > {
-  // Provide utilities to parse/serialize key/value.  Provide utilities to
-  // manipulate internal stored type.
-  typedef MapTypeHandler<kKeyFieldType, Key> KeyTypeHandler;
-  typedef MapTypeHandler<kValueFieldType, Value> ValueTypeHandler;
-
-  // Define internal memory layout. Strings and messages are stored as
-  // pointers, while other types are stored as values.
-  typedef typename KeyTypeHandler::TypeOnMemory KeyOnMemory;
-  typedef typename ValueTypeHandler::TypeOnMemory ValueOnMemory;
-
-  explicit MapEntryHelper(const MapPair<Key, Value>& map_pair)
-      : _has_bits_(3),
-        _cached_size_(2 + KeyTypeHandler::GetCachedSize(map_pair.first) +
-                      ValueTypeHandler::GetCachedSize(map_pair.second)),
-        key_(FromHelper<kKeyFieldType>::From(map_pair.first)),
-        value_(FromHelper<kValueFieldType>::From(map_pair.second)) {}
-
-  // Purposely not following the style guide naming. These are the names
-  // the proto compiler would generate given the map entry descriptor.
-  // The proto compiler generates the offsets in this struct as if this was
-  // a regular message. This way the table driven code barely notices it's
-  // dealing with a map field.
-  uint32 _has_bits_;     // NOLINT
-  uint32 _cached_size_;  // NOLINT
-  KeyOnMemory key_;      // NOLINT
-  ValueOnMemory value_;  // NOLINT
-};
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_MAP_ENTRY_LITE_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/map_field_lite.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/map_field_lite.h
deleted file mode 100644
index 665cb0ee..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/map_field_lite.h
+++ /dev/null
@@ -1,183 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_MAP_FIELD_LITE_H__
-#define GOOGLE_PROTOBUF_MAP_FIELD_LITE_H__
-
-#include <type_traits>
-#include <google/protobuf/parse_context.h>
-#include <google/protobuf/io/coded_stream.h>
-#include <google/protobuf/map.h>
-#include <google/protobuf/map_entry_lite.h>
-#include <google/protobuf/port.h>
-#include <google/protobuf/wire_format_lite.h>
-
-#include <google/protobuf/port_def.inc>
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-// This class provides access to map field using generated api. It is used for
-// internal generated message implementation only. Users should never use this
-// directly.
-template <typename Derived, typename Key, typename T,
-          WireFormatLite::FieldType key_wire_type,
-          WireFormatLite::FieldType value_wire_type>
-class MapFieldLite {
-  // Define message type for internal repeated field.
-  typedef Derived EntryType;
-
- public:
-  typedef Map<Key, T> MapType;
-  typedef EntryType EntryTypeTrait;
-
-  constexpr MapFieldLite() {}
-
-  explicit MapFieldLite(Arena* arena) : map_(arena) {}
-
-  // Accessors
-  const Map<Key, T>& GetMap() const { return map_; }
-  Map<Key, T>* MutableMap() { return &map_; }
-
-  // Convenient methods for generated message implementation.
-  int size() const { return static_cast<int>(map_.size()); }
-  void Clear() { return map_.clear(); }
-  void MergeFrom(const MapFieldLite& other) {
-    for (typename Map<Key, T>::const_iterator it = other.map_.begin();
-         it != other.map_.end(); ++it) {
-      map_[it->first] = it->second;
-    }
-  }
-  void Swap(MapFieldLite* other) { map_.swap(other->map_); }
-
-  // Used in the implementation of parsing. Caller should take the ownership iff
-  // arena_ is NULL.
-  EntryType* NewEntry() const {
-    return Arena::CreateMessage<EntryType>(map_.arena());
-  }
-  // Used in the implementation of serializing enum value type. Caller should
-  // take the ownership iff arena_ is NULL.
-  EntryType* NewEnumEntryWrapper(const Key& key, const T t) const {
-    return EntryType::EnumWrap(key, t, map_.arena_);
-  }
-  // Used in the implementation of serializing other value types. Caller should
-  // take the ownership iff arena_ is NULL.
-  EntryType* NewEntryWrapper(const Key& key, const T& t) const {
-    return EntryType::Wrap(key, t, map_.arena_);
-  }
-
-  const char* _InternalParse(const char* ptr, ParseContext* ctx) {
-    typename Derived::template Parser<MapFieldLite, Map<Key, T>> parser(this);
-    return parser._InternalParse(ptr, ctx);
-  }
-
-  template <typename UnknownType>
-  const char* ParseWithEnumValidation(const char* ptr, ParseContext* ctx,
-                                      bool (*is_valid)(int), uint32 field_num,
-                                      InternalMetadata* metadata) {
-    typename Derived::template Parser<MapFieldLite, Map<Key, T>> parser(this);
-    return parser.template ParseWithEnumValidation<UnknownType>(
-        ptr, ctx, is_valid, field_num, metadata);
-  }
-
- private:
-  typedef void DestructorSkippable_;
-
-  Map<Key, T> map_;
-
-  friend class ::PROTOBUF_NAMESPACE_ID::Arena;
-};
-
-template <typename UnknownType, typename T>
-struct EnumParseWrapper {
-  const char* _InternalParse(const char* ptr, ParseContext* ctx) {
-    return map_field->template ParseWithEnumValidation<UnknownType>(
-        ptr, ctx, is_valid, field_num, metadata);
-  }
-  T* map_field;
-  bool (*is_valid)(int);
-  uint32 field_num;
-  InternalMetadata* metadata;
-};
-
-// Helper function because the typenames of maps are horrendous to print. This
-// leverages compiler type deduction, to keep all type data out of the
-// generated code
-template <typename UnknownType, typename T>
-EnumParseWrapper<UnknownType, T> InitEnumParseWrapper(
-    T* map_field, bool (*is_valid)(int), uint32 field_num,
-    InternalMetadata* metadata) {
-  return EnumParseWrapper<UnknownType, T>{map_field, is_valid, field_num,
-                                          metadata};
-}
-
-// True if IsInitialized() is true for value field in all elements of t. T is
-// expected to be message.  It's useful to have this helper here to keep the
-// protobuf compiler from ever having to emit loops in IsInitialized() methods.
-// We want the C++ compiler to inline this or not as it sees fit.
-template <typename Derived, typename Key, typename T,
-          WireFormatLite::FieldType key_wire_type,
-          WireFormatLite::FieldType value_wire_type>
-bool AllAreInitialized(const MapFieldLite<Derived, Key, T, key_wire_type,
-                                          value_wire_type>& field) {
-  const auto& t = field.GetMap();
-  for (typename Map<Key, T>::const_iterator it = t.begin(); it != t.end();
-       ++it) {
-    if (!it->second.IsInitialized()) return false;
-  }
-  return true;
-}
-
-template <typename MEntry>
-struct MapEntryToMapField : MapEntryToMapField<typename MEntry::SuperType> {};
-
-template <typename T, typename Key, typename Value,
-          WireFormatLite::FieldType kKeyFieldType,
-          WireFormatLite::FieldType kValueFieldType>
-struct MapEntryToMapField<
-    MapEntryLite<T, Key, Value, kKeyFieldType, kValueFieldType>> {
-  typedef MapFieldLite<
-      MapEntryLite<T, Key, Value, kKeyFieldType, kValueFieldType>, Key, Value,
-      kKeyFieldType, kValueFieldType>
-      MapFieldType;
-};
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_MAP_FIELD_LITE_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/map_type_handler.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/map_type_handler.h
deleted file mode 100644
index 8b8fd14c..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/map_type_handler.h
+++ /dev/null
@@ -1,688 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_TYPE_HANDLER_H__
-#define GOOGLE_PROTOBUF_TYPE_HANDLER_H__
-
-#include <google/protobuf/parse_context.h>
-#include <google/protobuf/io/coded_stream.h>
-#include <google/protobuf/arena.h>
-#include <google/protobuf/wire_format_lite.h>
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-// Used for compile time type selection. MapIf::type will be TrueType if Flag is
-// true and FalseType otherwise.
-template <bool Flag, typename TrueType, typename FalseType>
-struct MapIf;
-
-template <typename TrueType, typename FalseType>
-struct MapIf<true, TrueType, FalseType> {
-  typedef TrueType type;
-};
-
-template <typename TrueType, typename FalseType>
-struct MapIf<false, TrueType, FalseType> {
-  typedef FalseType type;
-};
-
-template <typename Type, bool is_arena_constructable>
-class MapArenaMessageCreator {
- public:
-  // Use arena to create message if Type is arena constructable. Otherwise,
-  // create the message on heap.
-  static inline Type* CreateMessage(Arena* arena);
-};
-template <typename Type>
-class MapArenaMessageCreator<Type, true> {
- public:
-  static inline Type* CreateMessage(Arena* arena) {
-    return Arena::CreateMessage<Type>(arena);
-  }
-};
-template <typename Type>
-class MapArenaMessageCreator<Type, false> {
- public:
-  static inline Type* CreateMessage(Arena* arena) {
-    return Arena::Create<Type>(arena);
-  }
-};
-
-// Define constants for given wire field type
-template <WireFormatLite::FieldType field_type, typename Type>
-class MapWireFieldTypeTraits {};
-
-#define TYPE_TRAITS(FieldType, CType, WireFormatType, IsMessage, IsEnum)   \
-  template <typename Type>                                                 \
-  class MapWireFieldTypeTraits<WireFormatLite::TYPE_##FieldType, Type> {   \
-   public:                                                                 \
-    static const bool kIsMessage = IsMessage;                              \
-    static const bool kIsEnum = IsEnum;                                    \
-    typedef typename MapIf<kIsMessage, Type*, CType>::type TypeOnMemory;   \
-    typedef typename MapIf<kIsEnum, int, Type>::type MapEntryAccessorType; \
-    static const WireFormatLite::WireType kWireType =                      \
-        WireFormatLite::WIRETYPE_##WireFormatType;                         \
-  };
-
-TYPE_TRAITS(MESSAGE, Type, LENGTH_DELIMITED, true, false)
-TYPE_TRAITS(STRING, ArenaStringPtr, LENGTH_DELIMITED, false, false)
-TYPE_TRAITS(BYTES, ArenaStringPtr, LENGTH_DELIMITED, false, false)
-TYPE_TRAITS(INT64, int64, VARINT, false, false)
-TYPE_TRAITS(UINT64, uint64, VARINT, false, false)
-TYPE_TRAITS(INT32, int32, VARINT, false, false)
-TYPE_TRAITS(UINT32, uint32, VARINT, false, false)
-TYPE_TRAITS(SINT64, int64, VARINT, false, false)
-TYPE_TRAITS(SINT32, int32, VARINT, false, false)
-TYPE_TRAITS(ENUM, int, VARINT, false, true)
-TYPE_TRAITS(DOUBLE, double, FIXED64, false, false)
-TYPE_TRAITS(FLOAT, float, FIXED32, false, false)
-TYPE_TRAITS(FIXED64, uint64, FIXED64, false, false)
-TYPE_TRAITS(FIXED32, uint32, FIXED32, false, false)
-TYPE_TRAITS(SFIXED64, int64, FIXED64, false, false)
-TYPE_TRAITS(SFIXED32, int32, FIXED32, false, false)
-TYPE_TRAITS(BOOL, bool, VARINT, false, false)
-
-#undef TYPE_TRAITS
-
-template <WireFormatLite::FieldType field_type, typename Type>
-class MapTypeHandler {};
-
-template <typename Type>
-class MapTypeHandler<WireFormatLite::TYPE_MESSAGE, Type> {
- public:
-  // Enum type cannot be used for MapTypeHandler::Read. Define a type which will
-  // replace Enum with int.
-  typedef typename MapWireFieldTypeTraits<WireFormatLite::TYPE_MESSAGE,
-                                          Type>::MapEntryAccessorType
-      MapEntryAccessorType;
-  // Internal stored type in MapEntryLite for given wire field type.
-  typedef typename MapWireFieldTypeTraits<WireFormatLite::TYPE_MESSAGE,
-                                          Type>::TypeOnMemory TypeOnMemory;
-  // Corresponding wire type for field type.
-  static constexpr WireFormatLite::WireType kWireType =
-      MapWireFieldTypeTraits<WireFormatLite::TYPE_MESSAGE, Type>::kWireType;
-  // Whether wire type is for message.
-  static constexpr bool kIsMessage =
-      MapWireFieldTypeTraits<WireFormatLite::TYPE_MESSAGE, Type>::kIsMessage;
-  // Whether wire type is for enum.
-  static constexpr bool kIsEnum =
-      MapWireFieldTypeTraits<WireFormatLite::TYPE_MESSAGE, Type>::kIsEnum;
-
-  // Functions used in parsing and serialization. ===================
-  static inline size_t ByteSize(const MapEntryAccessorType& value);
-  static inline int GetCachedSize(const MapEntryAccessorType& value);
-  static inline bool Read(io::CodedInputStream* input,
-                          MapEntryAccessorType* value);
-  static inline const char* Read(const char* ptr, ParseContext* ctx,
-                                 MapEntryAccessorType* value);
-
-  static inline uint8* Write(int field, const MapEntryAccessorType& value,
-                             uint8* ptr, io::EpsCopyOutputStream* stream);
-
-  // Functions to manipulate data on memory. ========================
-  static inline const Type& GetExternalReference(const Type* value);
-  static inline void DeleteNoArena(const Type* x);
-  static inline void Merge(const Type& from, Type** to, Arena* arena);
-  static inline void Clear(Type** value, Arena* arena);
-  static constexpr TypeOnMemory Constinit();
-
-  static inline Type* EnsureMutable(Type** value, Arena* arena);
-  // SpaceUsedInMapEntry: Return bytes used by value in MapEntry, excluding
-  // those already calculate in sizeof(MapField).
-  static inline size_t SpaceUsedInMapEntryLong(const Type* value);
-  // Return default instance if value is not initialized when calling const
-  // reference accessor.
-  static inline const Type& DefaultIfNotInitialized(const Type* value);
-  // Check if all required fields have values set.
-  static inline bool IsInitialized(Type* value);
-};
-
-#define MAP_HANDLER(FieldType)                                                \
-  template <typename Type>                                                    \
-  class MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type> {              \
-   public:                                                                    \
-    typedef typename MapWireFieldTypeTraits<WireFormatLite::TYPE_##FieldType, \
-                                            Type>::MapEntryAccessorType       \
-        MapEntryAccessorType;                                                 \
-    typedef typename MapWireFieldTypeTraits<WireFormatLite::TYPE_##FieldType, \
-                                            Type>::TypeOnMemory TypeOnMemory; \
-    static const WireFormatLite::WireType kWireType =                         \
-        MapWireFieldTypeTraits<WireFormatLite::TYPE_##FieldType,              \
-                               Type>::kWireType;                              \
-    static const bool kIsMessage =                                            \
-        MapWireFieldTypeTraits<WireFormatLite::TYPE_##FieldType,              \
-                               Type>::kIsMessage;                             \
-    static const bool kIsEnum =                                               \
-        MapWireFieldTypeTraits<WireFormatLite::TYPE_##FieldType,              \
-                               Type>::kIsEnum;                                \
-    static inline int ByteSize(const MapEntryAccessorType& value);            \
-    static inline int GetCachedSize(const MapEntryAccessorType& value);       \
-    static inline bool Read(io::CodedInputStream* input,                      \
-                            MapEntryAccessorType* value);                     \
-    static inline const char* Read(const char* begin, ParseContext* ctx,      \
-                                   MapEntryAccessorType* value);              \
-    static inline uint8* Write(int field, const MapEntryAccessorType& value,  \
-                               uint8* ptr, io::EpsCopyOutputStream* stream);  \
-    static inline const MapEntryAccessorType& GetExternalReference(           \
-        const TypeOnMemory& value);                                           \
-    static inline void DeleteNoArena(const TypeOnMemory& x);                  \
-    static inline void Merge(const MapEntryAccessorType& from,                \
-                             TypeOnMemory* to, Arena* arena);                 \
-    static inline void Clear(TypeOnMemory* value, Arena* arena);              \
-    static inline size_t SpaceUsedInMapEntryLong(const TypeOnMemory& value);  \
-    static inline const MapEntryAccessorType& DefaultIfNotInitialized(        \
-        const TypeOnMemory& value);                                           \
-    static inline bool IsInitialized(const TypeOnMemory& value);              \
-    static void DeleteNoArena(TypeOnMemory& value);                           \
-    static constexpr TypeOnMemory Constinit();                                \
-    static inline MapEntryAccessorType* EnsureMutable(TypeOnMemory* value,    \
-                                                      Arena* arena);          \
-  };
-MAP_HANDLER(STRING)
-MAP_HANDLER(BYTES)
-MAP_HANDLER(INT64)
-MAP_HANDLER(UINT64)
-MAP_HANDLER(INT32)
-MAP_HANDLER(UINT32)
-MAP_HANDLER(SINT64)
-MAP_HANDLER(SINT32)
-MAP_HANDLER(ENUM)
-MAP_HANDLER(DOUBLE)
-MAP_HANDLER(FLOAT)
-MAP_HANDLER(FIXED64)
-MAP_HANDLER(FIXED32)
-MAP_HANDLER(SFIXED64)
-MAP_HANDLER(SFIXED32)
-MAP_HANDLER(BOOL)
-#undef MAP_HANDLER
-
-template <typename Type>
-inline size_t MapTypeHandler<WireFormatLite::TYPE_MESSAGE, Type>::ByteSize(
-    const MapEntryAccessorType& value) {
-  return WireFormatLite::MessageSizeNoVirtual(value);
-}
-
-#define GOOGLE_PROTOBUF_BYTE_SIZE(FieldType, DeclaredType)                     \
-  template <typename Type>                                                     \
-  inline int MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::ByteSize( \
-      const MapEntryAccessorType& value) {                                     \
-    return static_cast<int>(WireFormatLite::DeclaredType##Size(value));        \
-  }
-
-GOOGLE_PROTOBUF_BYTE_SIZE(STRING, String)
-GOOGLE_PROTOBUF_BYTE_SIZE(BYTES, Bytes)
-GOOGLE_PROTOBUF_BYTE_SIZE(INT64, Int64)
-GOOGLE_PROTOBUF_BYTE_SIZE(UINT64, UInt64)
-GOOGLE_PROTOBUF_BYTE_SIZE(INT32, Int32)
-GOOGLE_PROTOBUF_BYTE_SIZE(UINT32, UInt32)
-GOOGLE_PROTOBUF_BYTE_SIZE(SINT64, SInt64)
-GOOGLE_PROTOBUF_BYTE_SIZE(SINT32, SInt32)
-GOOGLE_PROTOBUF_BYTE_SIZE(ENUM, Enum)
-
-#undef GOOGLE_PROTOBUF_BYTE_SIZE
-
-#define FIXED_BYTE_SIZE(FieldType, DeclaredType)                               \
-  template <typename Type>                                                     \
-  inline int MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::ByteSize( \
-      const MapEntryAccessorType& /* value */) {                               \
-    return WireFormatLite::k##DeclaredType##Size;                              \
-  }
-
-FIXED_BYTE_SIZE(DOUBLE, Double)
-FIXED_BYTE_SIZE(FLOAT, Float)
-FIXED_BYTE_SIZE(FIXED64, Fixed64)
-FIXED_BYTE_SIZE(FIXED32, Fixed32)
-FIXED_BYTE_SIZE(SFIXED64, SFixed64)
-FIXED_BYTE_SIZE(SFIXED32, SFixed32)
-FIXED_BYTE_SIZE(BOOL, Bool)
-
-#undef FIXED_BYTE_SIZE
-
-template <typename Type>
-inline int MapTypeHandler<WireFormatLite::TYPE_MESSAGE, Type>::GetCachedSize(
-    const MapEntryAccessorType& value) {
-  return static_cast<int>(WireFormatLite::LengthDelimitedSize(
-      static_cast<size_t>(value.GetCachedSize())));
-}
-
-#define GET_CACHED_SIZE(FieldType, DeclaredType)                         \
-  template <typename Type>                                               \
-  inline int                                                             \
-  MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::GetCachedSize( \
-      const MapEntryAccessorType& value) {                               \
-    return static_cast<int>(WireFormatLite::DeclaredType##Size(value));  \
-  }
-
-GET_CACHED_SIZE(STRING, String)
-GET_CACHED_SIZE(BYTES, Bytes)
-GET_CACHED_SIZE(INT64, Int64)
-GET_CACHED_SIZE(UINT64, UInt64)
-GET_CACHED_SIZE(INT32, Int32)
-GET_CACHED_SIZE(UINT32, UInt32)
-GET_CACHED_SIZE(SINT64, SInt64)
-GET_CACHED_SIZE(SINT32, SInt32)
-GET_CACHED_SIZE(ENUM, Enum)
-
-#undef GET_CACHED_SIZE
-
-#define GET_FIXED_CACHED_SIZE(FieldType, DeclaredType)                   \
-  template <typename Type>                                               \
-  inline int                                                             \
-  MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::GetCachedSize( \
-      const MapEntryAccessorType& /* value */) {                         \
-    return WireFormatLite::k##DeclaredType##Size;                        \
-  }
-
-GET_FIXED_CACHED_SIZE(DOUBLE, Double)
-GET_FIXED_CACHED_SIZE(FLOAT, Float)
-GET_FIXED_CACHED_SIZE(FIXED64, Fixed64)
-GET_FIXED_CACHED_SIZE(FIXED32, Fixed32)
-GET_FIXED_CACHED_SIZE(SFIXED64, SFixed64)
-GET_FIXED_CACHED_SIZE(SFIXED32, SFixed32)
-GET_FIXED_CACHED_SIZE(BOOL, Bool)
-
-#undef GET_FIXED_CACHED_SIZE
-
-template <typename Type>
-inline uint8* MapTypeHandler<WireFormatLite::TYPE_MESSAGE, Type>::Write(
-    int field, const MapEntryAccessorType& value, uint8* ptr,
-    io::EpsCopyOutputStream* stream) {
-  ptr = stream->EnsureSpace(ptr);
-  return WireFormatLite::InternalWriteMessage(field, value, ptr, stream);
-}
-
-#define WRITE_METHOD(FieldType, DeclaredType)                                  \
-  template <typename Type>                                                     \
-  inline uint8* MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::Write( \
-      int field, const MapEntryAccessorType& value, uint8* ptr,                \
-      io::EpsCopyOutputStream* stream) {                                       \
-    ptr = stream->EnsureSpace(ptr);                                            \
-    return stream->Write##DeclaredType(field, value, ptr);                     \
-  }
-
-WRITE_METHOD(STRING, String)
-WRITE_METHOD(BYTES, Bytes)
-
-#undef WRITE_METHOD
-#define WRITE_METHOD(FieldType, DeclaredType)                                  \
-  template <typename Type>                                                     \
-  inline uint8* MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::Write( \
-      int field, const MapEntryAccessorType& value, uint8* ptr,                \
-      io::EpsCopyOutputStream* stream) {                                       \
-    ptr = stream->EnsureSpace(ptr);                                            \
-    return WireFormatLite::Write##DeclaredType##ToArray(field, value, ptr);    \
-  }
-
-WRITE_METHOD(INT64, Int64)
-WRITE_METHOD(UINT64, UInt64)
-WRITE_METHOD(INT32, Int32)
-WRITE_METHOD(UINT32, UInt32)
-WRITE_METHOD(SINT64, SInt64)
-WRITE_METHOD(SINT32, SInt32)
-WRITE_METHOD(ENUM, Enum)
-WRITE_METHOD(DOUBLE, Double)
-WRITE_METHOD(FLOAT, Float)
-WRITE_METHOD(FIXED64, Fixed64)
-WRITE_METHOD(FIXED32, Fixed32)
-WRITE_METHOD(SFIXED64, SFixed64)
-WRITE_METHOD(SFIXED32, SFixed32)
-WRITE_METHOD(BOOL, Bool)
-
-#undef WRITE_METHOD
-
-template <typename Type>
-inline bool MapTypeHandler<WireFormatLite::TYPE_MESSAGE, Type>::Read(
-    io::CodedInputStream* input, MapEntryAccessorType* value) {
-  return WireFormatLite::ReadMessageNoVirtual(input, value);
-}
-
-template <typename Type>
-inline bool MapTypeHandler<WireFormatLite::TYPE_STRING, Type>::Read(
-    io::CodedInputStream* input, MapEntryAccessorType* value) {
-  return WireFormatLite::ReadString(input, value);
-}
-
-template <typename Type>
-inline bool MapTypeHandler<WireFormatLite::TYPE_BYTES, Type>::Read(
-    io::CodedInputStream* input, MapEntryAccessorType* value) {
-  return WireFormatLite::ReadBytes(input, value);
-}
-
-template <typename Type>
-const char* MapTypeHandler<WireFormatLite::TYPE_MESSAGE, Type>::Read(
-    const char* ptr, ParseContext* ctx, MapEntryAccessorType* value) {
-  return ctx->ParseMessage(value, ptr);
-}
-
-template <typename Type>
-const char* MapTypeHandler<WireFormatLite::TYPE_STRING, Type>::Read(
-    const char* ptr, ParseContext* ctx, MapEntryAccessorType* value) {
-  int size = ReadSize(&ptr);
-  GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-  return ctx->ReadString(ptr, size, value);
-}
-
-template <typename Type>
-const char* MapTypeHandler<WireFormatLite::TYPE_BYTES, Type>::Read(
-    const char* ptr, ParseContext* ctx, MapEntryAccessorType* value) {
-  int size = ReadSize(&ptr);
-  GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-  return ctx->ReadString(ptr, size, value);
-}
-
-inline const char* ReadINT64(const char* ptr, int64* value) {
-  return VarintParse(ptr, reinterpret_cast<uint64*>(value));
-}
-inline const char* ReadUINT64(const char* ptr, uint64* value) {
-  return VarintParse(ptr, value);
-}
-inline const char* ReadINT32(const char* ptr, int32* value) {
-  return VarintParse(ptr, reinterpret_cast<uint32*>(value));
-}
-inline const char* ReadUINT32(const char* ptr, uint32* value) {
-  return VarintParse(ptr, value);
-}
-inline const char* ReadSINT64(const char* ptr, int64* value) {
-  *value = ReadVarintZigZag64(&ptr);
-  return ptr;
-}
-inline const char* ReadSINT32(const char* ptr, int32* value) {
-  *value = ReadVarintZigZag32(&ptr);
-  return ptr;
-}
-template <typename E>
-inline const char* ReadENUM(const char* ptr, E* value) {
-  *value = static_cast<E>(ReadVarint32(&ptr));
-  return ptr;
-}
-inline const char* ReadBOOL(const char* ptr, bool* value) {
-  *value = static_cast<bool>(ReadVarint32(&ptr));
-  return ptr;
-}
-
-template <typename F>
-inline const char* ReadUnaligned(const char* ptr, F* value) {
-  *value = UnalignedLoad<F>(ptr);
-  return ptr + sizeof(F);
-}
-inline const char* ReadFLOAT(const char* ptr, float* value) {
-  return ReadUnaligned(ptr, value);
-}
-inline const char* ReadDOUBLE(const char* ptr, double* value) {
-  return ReadUnaligned(ptr, value);
-}
-inline const char* ReadFIXED64(const char* ptr, uint64* value) {
-  return ReadUnaligned(ptr, value);
-}
-inline const char* ReadFIXED32(const char* ptr, uint32* value) {
-  return ReadUnaligned(ptr, value);
-}
-inline const char* ReadSFIXED64(const char* ptr, int64* value) {
-  return ReadUnaligned(ptr, value);
-}
-inline const char* ReadSFIXED32(const char* ptr, int32* value) {
-  return ReadUnaligned(ptr, value);
-}
-
-#define READ_METHOD(FieldType)                                              \
-  template <typename Type>                                                  \
-  inline bool MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::Read( \
-      io::CodedInputStream* input, MapEntryAccessorType* value) {           \
-    return WireFormatLite::ReadPrimitive<TypeOnMemory,                      \
-                                         WireFormatLite::TYPE_##FieldType>( \
-        input, value);                                                      \
-  }                                                                         \
-  template <typename Type>                                                  \
-  const char* MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::Read( \
-      const char* begin, ParseContext* ctx, MapEntryAccessorType* value) {  \
-    (void)ctx;                                                              \
-    return Read##FieldType(begin, value);                                   \
-  }
-
-READ_METHOD(INT64)
-READ_METHOD(UINT64)
-READ_METHOD(INT32)
-READ_METHOD(UINT32)
-READ_METHOD(SINT64)
-READ_METHOD(SINT32)
-READ_METHOD(ENUM)
-READ_METHOD(DOUBLE)
-READ_METHOD(FLOAT)
-READ_METHOD(FIXED64)
-READ_METHOD(FIXED32)
-READ_METHOD(SFIXED64)
-READ_METHOD(SFIXED32)
-READ_METHOD(BOOL)
-
-#undef READ_METHOD
-
-// Definition for message handler
-
-template <typename Type>
-inline const Type&
-MapTypeHandler<WireFormatLite::TYPE_MESSAGE, Type>::GetExternalReference(
-    const Type* value) {
-  return *value;
-}
-
-template <typename Type>
-inline size_t MapTypeHandler<WireFormatLite::TYPE_MESSAGE,
-                             Type>::SpaceUsedInMapEntryLong(const Type* value) {
-  return value->SpaceUsedLong();
-}
-
-template <typename Type>
-inline void MapTypeHandler<WireFormatLite::TYPE_MESSAGE, Type>::Clear(
-    Type** value, Arena* /* arena */) {
-  if (*value != NULL) (*value)->Clear();
-}
-template <typename Type>
-inline void MapTypeHandler<WireFormatLite::TYPE_MESSAGE, Type>::Merge(
-    const Type& from, Type** to, Arena* /* arena */) {
-  (*to)->MergeFrom(from);
-}
-
-template <typename Type>
-void MapTypeHandler<WireFormatLite::TYPE_MESSAGE, Type>::DeleteNoArena(
-    const Type* ptr) {
-  delete ptr;
-}
-
-template <typename Type>
-constexpr auto MapTypeHandler<WireFormatLite::TYPE_MESSAGE, Type>::Constinit()
-    -> TypeOnMemory {
-  return nullptr;
-}
-
-template <typename Type>
-inline Type* MapTypeHandler<WireFormatLite::TYPE_MESSAGE, Type>::EnsureMutable(
-    Type** value, Arena* arena) {
-  if (*value == NULL) {
-    *value = MapArenaMessageCreator<
-        Type,
-        Arena::is_arena_constructable<Type>::type::value>::CreateMessage(arena);
-  }
-  return *value;
-}
-
-template <typename Type>
-inline const Type&
-MapTypeHandler<WireFormatLite::TYPE_MESSAGE, Type>::DefaultIfNotInitialized(
-    const Type* value) {
-  return value != NULL ? *value : *Type::internal_default_instance();
-}
-
-template <typename Type>
-inline bool MapTypeHandler<WireFormatLite::TYPE_MESSAGE, Type>::IsInitialized(
-    Type* value) {
-  return value ? value->IsInitialized() : false;
-}
-
-// Definition for string/bytes handler
-
-#define STRING_OR_BYTES_HANDLER_FUNCTIONS(FieldType)                          \
-  template <typename Type>                                                    \
-  inline const typename MapTypeHandler<WireFormatLite::TYPE_##FieldType,      \
-                                       Type>::MapEntryAccessorType&           \
-  MapTypeHandler<WireFormatLite::TYPE_##FieldType,                            \
-                 Type>::GetExternalReference(const TypeOnMemory& value) {     \
-    return value.Get();                                                       \
-  }                                                                           \
-  template <typename Type>                                                    \
-  inline size_t                                                               \
-  MapTypeHandler<WireFormatLite::TYPE_##FieldType,                            \
-                 Type>::SpaceUsedInMapEntryLong(const TypeOnMemory& value) {  \
-    return sizeof(value);                                                     \
-  }                                                                           \
-  template <typename Type>                                                    \
-  inline void MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::Clear(  \
-      TypeOnMemory* value, Arena* arena) {                                    \
-    value->ClearToEmpty();                                                    \
-  }                                                                           \
-  template <typename Type>                                                    \
-  inline void MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::Merge(  \
-      const MapEntryAccessorType& from, TypeOnMemory* to, Arena* arena) {     \
-    to->Set(&internal::GetEmptyStringAlreadyInited(), from, arena);           \
-  }                                                                           \
-  template <typename Type>                                                    \
-  void MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::DeleteNoArena( \
-      TypeOnMemory& value) {                                                  \
-    value.DestroyNoArena(&internal::GetEmptyStringAlreadyInited());           \
-  }                                                                           \
-  template <typename Type>                                                    \
-  constexpr auto                                                              \
-  MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::Constinit()         \
-      ->TypeOnMemory {                                                        \
-    return TypeOnMemory(&internal::GetEmptyStringAlreadyInited());            \
-  }                                                                           \
-  template <typename Type>                                                    \
-  inline typename MapTypeHandler<WireFormatLite::TYPE_##FieldType,            \
-                                 Type>::MapEntryAccessorType*                 \
-  MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::EnsureMutable(      \
-      TypeOnMemory* value, Arena* arena) {                                    \
-    return value->Mutable(ArenaStringPtr::EmptyDefault{}, arena);             \
-  }                                                                           \
-  template <typename Type>                                                    \
-  inline const typename MapTypeHandler<WireFormatLite::TYPE_##FieldType,      \
-                                       Type>::MapEntryAccessorType&           \
-  MapTypeHandler<WireFormatLite::TYPE_##FieldType,                            \
-                 Type>::DefaultIfNotInitialized(const TypeOnMemory& value) {  \
-    return value.Get();                                                       \
-  }                                                                           \
-  template <typename Type>                                                    \
-  inline bool                                                                 \
-  MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::IsInitialized(      \
-      const TypeOnMemory& /* value */) {                                      \
-    return true;                                                              \
-  }
-STRING_OR_BYTES_HANDLER_FUNCTIONS(STRING)
-STRING_OR_BYTES_HANDLER_FUNCTIONS(BYTES)
-#undef STRING_OR_BYTES_HANDLER_FUNCTIONS
-
-#define PRIMITIVE_HANDLER_FUNCTIONS(FieldType)                               \
-  template <typename Type>                                                   \
-  inline const typename MapTypeHandler<WireFormatLite::TYPE_##FieldType,     \
-                                       Type>::MapEntryAccessorType&          \
-  MapTypeHandler<WireFormatLite::TYPE_##FieldType,                           \
-                 Type>::GetExternalReference(const TypeOnMemory& value) {    \
-    return value;                                                            \
-  }                                                                          \
-  template <typename Type>                                                   \
-  inline size_t MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::     \
-      SpaceUsedInMapEntryLong(const TypeOnMemory& /* value */) {             \
-    return 0;                                                                \
-  }                                                                          \
-  template <typename Type>                                                   \
-  inline void MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::Clear( \
-      TypeOnMemory* value, Arena* /* arena */) {                             \
-    *value = 0;                                                              \
-  }                                                                          \
-  template <typename Type>                                                   \
-  inline void MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::Merge( \
-      const MapEntryAccessorType& from, TypeOnMemory* to,                    \
-      Arena* /* arena */) {                                                  \
-    *to = from;                                                              \
-  }                                                                          \
-  template <typename Type>                                                   \
-  inline void MapTypeHandler<WireFormatLite::TYPE_##FieldType,               \
-                             Type>::DeleteNoArena(TypeOnMemory& /* x */) {}  \
-  template <typename Type>                                                   \
-  constexpr auto                                                             \
-  MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::Constinit()        \
-      ->TypeOnMemory {                                                       \
-    return 0;                                                                \
-  }                                                                          \
-  template <typename Type>                                                   \
-  inline typename MapTypeHandler<WireFormatLite::TYPE_##FieldType,           \
-                                 Type>::MapEntryAccessorType*                \
-  MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::EnsureMutable(     \
-      TypeOnMemory* value, Arena* /* arena */) {                             \
-    return value;                                                            \
-  }                                                                          \
-  template <typename Type>                                                   \
-  inline const typename MapTypeHandler<WireFormatLite::TYPE_##FieldType,     \
-                                       Type>::MapEntryAccessorType&          \
-  MapTypeHandler<WireFormatLite::TYPE_##FieldType,                           \
-                 Type>::DefaultIfNotInitialized(const TypeOnMemory& value) { \
-    return value;                                                            \
-  }                                                                          \
-  template <typename Type>                                                   \
-  inline bool                                                                \
-  MapTypeHandler<WireFormatLite::TYPE_##FieldType, Type>::IsInitialized(     \
-      const TypeOnMemory& /* value */) {                                     \
-    return true;                                                             \
-  }
-PRIMITIVE_HANDLER_FUNCTIONS(INT64)
-PRIMITIVE_HANDLER_FUNCTIONS(UINT64)
-PRIMITIVE_HANDLER_FUNCTIONS(INT32)
-PRIMITIVE_HANDLER_FUNCTIONS(UINT32)
-PRIMITIVE_HANDLER_FUNCTIONS(SINT64)
-PRIMITIVE_HANDLER_FUNCTIONS(SINT32)
-PRIMITIVE_HANDLER_FUNCTIONS(ENUM)
-PRIMITIVE_HANDLER_FUNCTIONS(DOUBLE)
-PRIMITIVE_HANDLER_FUNCTIONS(FLOAT)
-PRIMITIVE_HANDLER_FUNCTIONS(FIXED64)
-PRIMITIVE_HANDLER_FUNCTIONS(FIXED32)
-PRIMITIVE_HANDLER_FUNCTIONS(SFIXED64)
-PRIMITIVE_HANDLER_FUNCTIONS(SFIXED32)
-PRIMITIVE_HANDLER_FUNCTIONS(BOOL)
-#undef PRIMITIVE_HANDLER_FUNCTIONS
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#endif  // GOOGLE_PROTOBUF_TYPE_HANDLER_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/message_lite.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/message_lite.h
deleted file mode 100644
index a76c16e5..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/message_lite.h
+++ /dev/null
@@ -1,621 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// Authors: wink@google.com (Wink Saville),
-//          kenton@google.com (Kenton Varda)
-//  Based on original Protocol Buffers design by
-//  Sanjay Ghemawat, Jeff Dean, and others.
-//
-// Defines MessageLite, the abstract interface implemented by all (lite
-// and non-lite) protocol message objects.
-
-#ifndef GOOGLE_PROTOBUF_MESSAGE_LITE_H__
-#define GOOGLE_PROTOBUF_MESSAGE_LITE_H__
-
-#include <climits>
-#include <string>
-
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/stubs/logging.h>
-#include <google/protobuf/io/coded_stream.h>
-#include <google/protobuf/arena.h>
-#include <google/protobuf/metadata_lite.h>
-#include <google/protobuf/stubs/once.h>
-#include <google/protobuf/port.h>
-#include <google/protobuf/stubs/strutil.h>
-
-
-#include <google/protobuf/port_def.inc>
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-namespace google {
-namespace protobuf {
-
-template <typename T>
-class RepeatedPtrField;
-
-namespace io {
-
-class CodedInputStream;
-class CodedOutputStream;
-class ZeroCopyInputStream;
-class ZeroCopyOutputStream;
-
-}  // namespace io
-namespace internal {
-
-// Tag type used to invoke the constinit constructor overload of some classes.
-// Such constructors are internal implementation details of the library.
-struct ConstantInitialized {
-  explicit ConstantInitialized() = default;
-};
-
-// See parse_context.h for explanation
-class ParseContext;
-
-class RepeatedPtrFieldBase;
-class WireFormatLite;
-class WeakFieldMap;
-
-// We compute sizes as size_t but cache them as int.  This function converts a
-// computed size to a cached size.  Since we don't proceed with serialization
-// if the total size was > INT_MAX, it is not important what this function
-// returns for inputs > INT_MAX.  However this case should not error or
-// GOOGLE_CHECK-fail, because the full size_t resolution is still returned from
-// ByteSizeLong() and checked against INT_MAX; we can catch the overflow
-// there.
-inline int ToCachedSize(size_t size) { return static_cast<int>(size); }
-
-// We mainly calculate sizes in terms of size_t, but some functions that
-// compute sizes return "int".  These int sizes are expected to always be
-// positive. This function is more efficient than casting an int to size_t
-// directly on 64-bit platforms because it avoids making the compiler emit a
-// sign extending instruction, which we don't want and don't want to pay for.
-inline size_t FromIntSize(int size) {
-  // Convert to unsigned before widening so sign extension is not necessary.
-  return static_cast<unsigned int>(size);
-}
-
-// For cases where a legacy function returns an integer size.  We GOOGLE_DCHECK()
-// that the conversion will fit within an integer; if this is false then we
-// are losing information.
-inline int ToIntSize(size_t size) {
-  GOOGLE_DCHECK_LE(size, static_cast<size_t>(INT_MAX));
-  return static_cast<int>(size);
-}
-
-// This type wraps a variable whose constructor and destructor are explicitly
-// called. It is particularly useful for a global variable, without its
-// constructor and destructor run on start and end of the program lifetime.
-// This circumvents the initial construction order fiasco, while keeping
-// the address of the empty string a compile time constant.
-//
-// Pay special attention to the initialization state of the object.
-// 1. The object is "uninitialized" to begin with.
-// 2. Call Construct() or DefaultConstruct() only if the object is
-//    uninitialized. After the call, the object becomes "initialized".
-// 3. Call get() and get_mutable() only if the object is initialized.
-// 4. Call Destruct() only if the object is initialized.
-//    After the call, the object becomes uninitialized.
-template <typename T>
-class ExplicitlyConstructed {
- public:
-  void DefaultConstruct() { new (&union_) T(); }
-
-  template <typename... Args>
-  void Construct(Args&&... args) {
-    new (&union_) T(std::forward<Args>(args)...);
-  }
-
-  void Destruct() { get_mutable()->~T(); }
-
-  constexpr const T& get() const { return reinterpret_cast<const T&>(union_); }
-  T* get_mutable() { return reinterpret_cast<T*>(&union_); }
-
- private:
-  // Prefer c++14 aligned_storage, but for compatibility this will do.
-  union AlignedUnion {
-    char space[sizeof(T)];
-    int64 align_to_int64;
-    void* align_to_ptr;
-  } union_;
-};
-
-PROTOBUF_DISABLE_MSVC_UNION_WARNING
-// We need a publicly accessible `value` object to allow constexpr
-// support in C++11.
-// A constexpr accessor does not work portably.
-union EmptyString {
-  constexpr EmptyString() : dummy{} {}
-  ~EmptyString() {}
-
-  // We need a dummy object for constant initialization.
-  std::false_type dummy;
-  std::string value;
-};
-PROTOBUF_ENABLE_MSVC_UNION_WARNING
-
-// Default empty string object. Don't use this directly. Instead, call
-// GetEmptyString() to get the reference.
-PROTOBUF_EXPORT extern EmptyString fixed_address_empty_string;
-
-
-PROTOBUF_EXPORT constexpr const std::string& GetEmptyStringAlreadyInited() {
-  return fixed_address_empty_string.value;
-}
-
-PROTOBUF_EXPORT size_t StringSpaceUsedExcludingSelfLong(const std::string& str);
-
-}  // namespace internal
-
-// Interface to light weight protocol messages.
-//
-// This interface is implemented by all protocol message objects.  Non-lite
-// messages additionally implement the Message interface, which is a
-// subclass of MessageLite.  Use MessageLite instead when you only need
-// the subset of features which it supports -- namely, nothing that uses
-// descriptors or reflection.  You can instruct the protocol compiler
-// to generate classes which implement only MessageLite, not the full
-// Message interface, by adding the following line to the .proto file:
-//
-//   option optimize_for = LITE_RUNTIME;
-//
-// This is particularly useful on resource-constrained systems where
-// the full protocol buffers runtime library is too big.
-//
-// Note that on non-constrained systems (e.g. servers) when you need
-// to link in lots of protocol definitions, a better way to reduce
-// total code footprint is to use optimize_for = CODE_SIZE.  This
-// will make the generated code smaller while still supporting all the
-// same features (at the expense of speed).  optimize_for = LITE_RUNTIME
-// is best when you only have a small number of message types linked
-// into your binary, in which case the size of the protocol buffers
-// runtime itself is the biggest problem.
-//
-// Users must not derive from this class. Only the protocol compiler and
-// the internal library are allowed to create subclasses.
-class PROTOBUF_EXPORT MessageLite {
- public:
-  constexpr MessageLite() = default;
-  virtual ~MessageLite() = default;
-
-  // Basic Operations ------------------------------------------------
-
-  // Get the name of this message type, e.g. "foo.bar.BazProto".
-  virtual std::string GetTypeName() const = 0;
-
-  // Construct a new instance of the same type.  Ownership is passed to the
-  // caller.
-  virtual MessageLite* New() const = 0;
-
-  // Construct a new instance on the arena. Ownership is passed to the caller
-  // if arena is a NULL. Default implementation for backwards compatibility.
-  virtual MessageLite* New(Arena* arena) const;
-
-  // Get the arena, if any, associated with this message. Virtual method
-  // required for generic operations but most arena-related operations should
-  // use the GetArena() generated-code method. Default implementation
-  // to reduce code size by avoiding the need for per-type implementations
-  // when types do not implement arena support.
-  Arena* GetArena() const { return _internal_metadata_.arena(); }
-
-  // Get a pointer that may be equal to this message's arena, or may not be.
-  // If the value returned by this method is equal to some arena pointer, then
-  // this message is on that arena; however, if this message is on some arena,
-  // this method may or may not return that arena's pointer. As a tradeoff,
-  // this method may be more efficient than GetArena(). The intent is to allow
-  // underlying representations that use e.g. tagged pointers to sometimes
-  // store the arena pointer directly, and sometimes in a more indirect way,
-  // and allow a fastpath comparison against the arena pointer when it's easy
-  // to obtain.
-  void* GetMaybeArenaPointer() const {
-    return _internal_metadata_.raw_arena_ptr();
-  }
-
-  // Clear all fields of the message and set them to their default values.
-  // Clear() avoids freeing memory, assuming that any memory allocated
-  // to hold parts of the message will be needed again to hold the next
-  // message.  If you actually want to free the memory used by a Message,
-  // you must delete it.
-  virtual void Clear() = 0;
-
-  // Quickly check if all required fields have values set.
-  virtual bool IsInitialized() const = 0;
-
-  // This is not implemented for Lite messages -- it just returns "(cannot
-  // determine missing fields for lite message)".  However, it is implemented
-  // for full messages.  See message.h.
-  virtual std::string InitializationErrorString() const;
-
-  // If |other| is the exact same class as this, calls MergeFrom(). Otherwise,
-  // results are undefined (probably crash).
-  virtual void CheckTypeAndMergeFrom(const MessageLite& other) = 0;
-
-  // These methods return a human-readable summary of the message. Note that
-  // since the MessageLite interface does not support reflection, there is very
-  // little information that these methods can provide. They are shadowed by
-  // methods of the same name on the Message interface which provide much more
-  // information. The methods here are intended primarily to facilitate code
-  // reuse for logic that needs to interoperate with both full and lite protos.
-  //
-  // The format of the returned string is subject to change, so please do not
-  // assume it will remain stable over time.
-  std::string DebugString() const;
-  std::string ShortDebugString() const { return DebugString(); }
-  // MessageLite::DebugString is already Utf8 Safe. This is to add compatibility
-  // with Message.
-  std::string Utf8DebugString() const { return DebugString(); }
-
-  // Parsing ---------------------------------------------------------
-  // Methods for parsing in protocol buffer format.  Most of these are
-  // just simple wrappers around MergeFromCodedStream().  Clear() will be
-  // called before merging the input.
-
-  // Fill the message with a protocol buffer parsed from the given input
-  // stream. Returns false on a read error or if the input is in the wrong
-  // format.  A successful return does not indicate the entire input is
-  // consumed, ensure you call ConsumedEntireMessage() to check that if
-  // applicable.
-  PROTOBUF_ATTRIBUTE_REINITIALIZES bool ParseFromCodedStream(
-      io::CodedInputStream* input);
-  // Like ParseFromCodedStream(), but accepts messages that are missing
-  // required fields.
-  PROTOBUF_ATTRIBUTE_REINITIALIZES bool ParsePartialFromCodedStream(
-      io::CodedInputStream* input);
-  // Read a protocol buffer from the given zero-copy input stream.  If
-  // successful, the entire input will be consumed.
-  PROTOBUF_ATTRIBUTE_REINITIALIZES bool ParseFromZeroCopyStream(
-      io::ZeroCopyInputStream* input);
-  // Like ParseFromZeroCopyStream(), but accepts messages that are missing
-  // required fields.
-  PROTOBUF_ATTRIBUTE_REINITIALIZES bool ParsePartialFromZeroCopyStream(
-      io::ZeroCopyInputStream* input);
-  // Parse a protocol buffer from a file descriptor.  If successful, the entire
-  // input will be consumed.
-  PROTOBUF_ATTRIBUTE_REINITIALIZES bool ParseFromFileDescriptor(
-      int file_descriptor);
-  // Like ParseFromFileDescriptor(), but accepts messages that are missing
-  // required fields.
-  PROTOBUF_ATTRIBUTE_REINITIALIZES bool ParsePartialFromFileDescriptor(
-      int file_descriptor);
-  // Parse a protocol buffer from a C++ istream.  If successful, the entire
-  // input will be consumed.
-  PROTOBUF_ATTRIBUTE_REINITIALIZES bool ParseFromIstream(std::istream* input);
-  // Like ParseFromIstream(), but accepts messages that are missing
-  // required fields.
-  PROTOBUF_ATTRIBUTE_REINITIALIZES bool ParsePartialFromIstream(
-      std::istream* input);
-  // Read a protocol buffer from the given zero-copy input stream, expecting
-  // the message to be exactly "size" bytes long.  If successful, exactly
-  // this many bytes will have been consumed from the input.
-  bool MergePartialFromBoundedZeroCopyStream(io::ZeroCopyInputStream* input,
-                                             int size);
-  // Like ParseFromBoundedZeroCopyStream(), but accepts messages that are
-  // missing required fields.
-  bool MergeFromBoundedZeroCopyStream(io::ZeroCopyInputStream* input, int size);
-  PROTOBUF_ATTRIBUTE_REINITIALIZES bool ParseFromBoundedZeroCopyStream(
-      io::ZeroCopyInputStream* input, int size);
-  // Like ParseFromBoundedZeroCopyStream(), but accepts messages that are
-  // missing required fields.
-  PROTOBUF_ATTRIBUTE_REINITIALIZES bool ParsePartialFromBoundedZeroCopyStream(
-      io::ZeroCopyInputStream* input, int size);
-  // Parses a protocol buffer contained in a string. Returns true on success.
-  // This function takes a string in the (non-human-readable) binary wire
-  // format, matching the encoding output by MessageLite::SerializeToString().
-  // If you'd like to convert a human-readable string into a protocol buffer
-  // object, see google::protobuf::TextFormat::ParseFromString().
-  PROTOBUF_ATTRIBUTE_REINITIALIZES bool ParseFromString(ConstStringParam data);
-  // Like ParseFromString(), but accepts messages that are missing
-  // required fields.
-  PROTOBUF_ATTRIBUTE_REINITIALIZES bool ParsePartialFromString(
-      ConstStringParam data);
-  // Parse a protocol buffer contained in an array of bytes.
-  PROTOBUF_ATTRIBUTE_REINITIALIZES bool ParseFromArray(const void* data,
-                                                       int size);
-  // Like ParseFromArray(), but accepts messages that are missing
-  // required fields.
-  PROTOBUF_ATTRIBUTE_REINITIALIZES bool ParsePartialFromArray(const void* data,
-                                                              int size);
-
-
-  // Reads a protocol buffer from the stream and merges it into this
-  // Message.  Singular fields read from the what is
-  // already in the Message and repeated fields are appended to those
-  // already present.
-  //
-  // It is the responsibility of the caller to call input->LastTagWas()
-  // (for groups) or input->ConsumedEntireMessage() (for non-groups) after
-  // this returns to verify that the message's end was delimited correctly.
-  //
-  // ParseFromCodedStream() is implemented as Clear() followed by
-  // MergeFromCodedStream().
-  bool MergeFromCodedStream(io::CodedInputStream* input);
-
-  // Like MergeFromCodedStream(), but succeeds even if required fields are
-  // missing in the input.
-  //
-  // MergeFromCodedStream() is just implemented as MergePartialFromCodedStream()
-  // followed by IsInitialized().
-  bool MergePartialFromCodedStream(io::CodedInputStream* input);
-
-  // Merge a protocol buffer contained in a string.
-  bool MergeFromString(ConstStringParam data);
-
-
-  // Serialization ---------------------------------------------------
-  // Methods for serializing in protocol buffer format.  Most of these
-  // are just simple wrappers around ByteSize() and SerializeWithCachedSizes().
-
-  // Write a protocol buffer of this message to the given output.  Returns
-  // false on a write error.  If the message is missing required fields,
-  // this may GOOGLE_CHECK-fail.
-  bool SerializeToCodedStream(io::CodedOutputStream* output) const;
-  // Like SerializeToCodedStream(), but allows missing required fields.
-  bool SerializePartialToCodedStream(io::CodedOutputStream* output) const;
-  // Write the message to the given zero-copy output stream.  All required
-  // fields must be set.
-  bool SerializeToZeroCopyStream(io::ZeroCopyOutputStream* output) const;
-  // Like SerializeToZeroCopyStream(), but allows missing required fields.
-  bool SerializePartialToZeroCopyStream(io::ZeroCopyOutputStream* output) const;
-  // Serialize the message and store it in the given string.  All required
-  // fields must be set.
-  bool SerializeToString(std::string* output) const;
-  // Like SerializeToString(), but allows missing required fields.
-  bool SerializePartialToString(std::string* output) const;
-  // Serialize the message and store it in the given byte array.  All required
-  // fields must be set.
-  bool SerializeToArray(void* data, int size) const;
-  // Like SerializeToArray(), but allows missing required fields.
-  bool SerializePartialToArray(void* data, int size) const;
-
-  // Make a string encoding the message. Is equivalent to calling
-  // SerializeToString() on a string and using that.  Returns the empty
-  // string if SerializeToString() would have returned an error.
-  // Note: If you intend to generate many such strings, you may
-  // reduce heap fragmentation by instead re-using the same string
-  // object with calls to SerializeToString().
-  std::string SerializeAsString() const;
-  // Like SerializeAsString(), but allows missing required fields.
-  std::string SerializePartialAsString() const;
-
-  // Serialize the message and write it to the given file descriptor.  All
-  // required fields must be set.
-  bool SerializeToFileDescriptor(int file_descriptor) const;
-  // Like SerializeToFileDescriptor(), but allows missing required fields.
-  bool SerializePartialToFileDescriptor(int file_descriptor) const;
-  // Serialize the message and write it to the given C++ ostream.  All
-  // required fields must be set.
-  bool SerializeToOstream(std::ostream* output) const;
-  // Like SerializeToOstream(), but allows missing required fields.
-  bool SerializePartialToOstream(std::ostream* output) const;
-
-  // Like SerializeToString(), but appends to the data to the string's
-  // existing contents.  All required fields must be set.
-  bool AppendToString(std::string* output) const;
-  // Like AppendToString(), but allows missing required fields.
-  bool AppendPartialToString(std::string* output) const;
-
-
-  // Computes the serialized size of the message.  This recursively calls
-  // ByteSizeLong() on all embedded messages.
-  //
-  // ByteSizeLong() is generally linear in the number of fields defined for the
-  // proto.
-  virtual size_t ByteSizeLong() const = 0;
-
-  // Legacy ByteSize() API.
-  PROTOBUF_DEPRECATED_MSG("Please use ByteSizeLong() instead")
-  int ByteSize() const { return internal::ToIntSize(ByteSizeLong()); }
-
-  // Serializes the message without recomputing the size.  The message must not
-  // have changed since the last call to ByteSize(), and the value returned by
-  // ByteSize must be non-negative.  Otherwise the results are undefined.
-  void SerializeWithCachedSizes(io::CodedOutputStream* output) const {
-    output->SetCur(_InternalSerialize(output->Cur(), output->EpsCopy()));
-  }
-
-  // Functions below here are not part of the public interface.  It isn't
-  // enforced, but they should be treated as private, and will be private
-  // at some future time.  Unfortunately the implementation of the "friend"
-  // keyword in GCC is broken at the moment, but we expect it will be fixed.
-
-  // Like SerializeWithCachedSizes, but writes directly to *target, returning
-  // a pointer to the byte immediately after the last byte written.  "target"
-  // must point at a byte array of at least ByteSize() bytes.  Whether to use
-  // deterministic serialization, e.g., maps in sorted order, is determined by
-  // CodedOutputStream::IsDefaultSerializationDeterministic().
-  uint8* SerializeWithCachedSizesToArray(uint8* target) const;
-
-  // Returns the result of the last call to ByteSize().  An embedded message's
-  // size is needed both to serialize it (because embedded messages are
-  // length-delimited) and to compute the outer message's size.  Caching
-  // the size avoids computing it multiple times.
-  //
-  // ByteSize() does not automatically use the cached size when available
-  // because this would require invalidating it every time the message was
-  // modified, which would be too hard and expensive.  (E.g. if a deeply-nested
-  // sub-message is changed, all of its parents' cached sizes would need to be
-  // invalidated, which is too much work for an otherwise inlined setter
-  // method.)
-  virtual int GetCachedSize() const = 0;
-
-  virtual const char* _InternalParse(const char* /*ptr*/,
-                                     internal::ParseContext* /*ctx*/) {
-    return nullptr;
-  }
-
- protected:
-  template <typename T>
-  static T* CreateMaybeMessage(Arena* arena) {
-    return Arena::CreateMaybeMessage<T>(arena);
-  }
-
-  inline explicit MessageLite(Arena* arena) : _internal_metadata_(arena) {}
-
-  internal::InternalMetadata _internal_metadata_;
-
- public:
-  enum ParseFlags {
-    kMerge = 0,
-    kParse = 1,
-    kMergePartial = 2,
-    kParsePartial = 3,
-    kMergeWithAliasing = 4,
-    kParseWithAliasing = 5,
-    kMergePartialWithAliasing = 6,
-    kParsePartialWithAliasing = 7
-  };
-
-  template <ParseFlags flags, typename T>
-  bool ParseFrom(const T& input);
-
-  // Fast path when conditions match (ie. non-deterministic)
-  //  uint8* _InternalSerialize(uint8* ptr) const;
-  virtual uint8* _InternalSerialize(uint8* ptr,
-                                    io::EpsCopyOutputStream* stream) const = 0;
-
-  // Identical to IsInitialized() except that it logs an error message.
-  bool IsInitializedWithErrors() const {
-    if (IsInitialized()) return true;
-    LogInitializationErrorMessage();
-    return false;
-  }
-
- private:
-  // TODO(gerbens) make this a pure abstract function
-  virtual const void* InternalGetTable() const { return NULL; }
-
-  friend class internal::WireFormatLite;
-  friend class Message;
-  friend class internal::WeakFieldMap;
-
-  void LogInitializationErrorMessage() const;
-
-  bool MergeFromImpl(io::CodedInputStream* input, ParseFlags parse_flags);
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(MessageLite);
-};
-
-namespace internal {
-
-template <bool alias>
-bool MergeFromImpl(StringPiece input, MessageLite* msg,
-                   MessageLite::ParseFlags parse_flags);
-extern template bool MergeFromImpl<false>(StringPiece input,
-                                          MessageLite* msg,
-                                          MessageLite::ParseFlags parse_flags);
-extern template bool MergeFromImpl<true>(StringPiece input,
-                                         MessageLite* msg,
-                                         MessageLite::ParseFlags parse_flags);
-
-template <bool alias>
-bool MergeFromImpl(io::ZeroCopyInputStream* input, MessageLite* msg,
-                   MessageLite::ParseFlags parse_flags);
-extern template bool MergeFromImpl<false>(io::ZeroCopyInputStream* input,
-                                          MessageLite* msg,
-                                          MessageLite::ParseFlags parse_flags);
-extern template bool MergeFromImpl<true>(io::ZeroCopyInputStream* input,
-                                         MessageLite* msg,
-                                         MessageLite::ParseFlags parse_flags);
-
-struct BoundedZCIS {
-  io::ZeroCopyInputStream* zcis;
-  int limit;
-};
-
-template <bool alias>
-bool MergeFromImpl(BoundedZCIS input, MessageLite* msg,
-                   MessageLite::ParseFlags parse_flags);
-extern template bool MergeFromImpl<false>(BoundedZCIS input, MessageLite* msg,
-                                          MessageLite::ParseFlags parse_flags);
-extern template bool MergeFromImpl<true>(BoundedZCIS input, MessageLite* msg,
-                                         MessageLite::ParseFlags parse_flags);
-
-template <typename T>
-struct SourceWrapper;
-
-template <bool alias, typename T>
-bool MergeFromImpl(const SourceWrapper<T>& input, MessageLite* msg,
-                   MessageLite::ParseFlags parse_flags) {
-  return input.template MergeInto<alias>(msg, parse_flags);
-}
-
-}  // namespace internal
-
-template <MessageLite::ParseFlags flags, typename T>
-bool MessageLite::ParseFrom(const T& input) {
-  if (flags & kParse) Clear();
-  constexpr bool alias = (flags & kMergeWithAliasing) != 0;
-  return internal::MergeFromImpl<alias>(input, this, flags);
-}
-
-// ===================================================================
-// Shutdown support.
-
-
-// Shut down the entire protocol buffers library, deleting all static-duration
-// objects allocated by the library or by generated .pb.cc files.
-//
-// There are two reasons you might want to call this:
-// * You use a draconian definition of "memory leak" in which you expect
-//   every single malloc() to have a corresponding free(), even for objects
-//   which live until program exit.
-// * You are writing a dynamically-loaded library which needs to clean up
-//   after itself when the library is unloaded.
-//
-// It is safe to call this multiple times.  However, it is not safe to use
-// any other part of the protocol buffers library after
-// ShutdownProtobufLibrary() has been called. Furthermore this call is not
-// thread safe, user needs to synchronize multiple calls.
-PROTOBUF_EXPORT void ShutdownProtobufLibrary();
-
-namespace internal {
-
-// Register a function to be called when ShutdownProtocolBuffers() is called.
-PROTOBUF_EXPORT void OnShutdown(void (*func)());
-// Run an arbitrary function on an arg
-PROTOBUF_EXPORT void OnShutdownRun(void (*f)(const void*), const void* arg);
-
-template <typename T>
-T* OnShutdownDelete(T* p) {
-  OnShutdownRun([](const void* pp) { delete static_cast<const T*>(pp); }, p);
-  return p;
-}
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_MESSAGE_LITE_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/metadata_lite.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/metadata_lite.h
deleted file mode 100644
index ec5f9238..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/metadata_lite.h
+++ /dev/null
@@ -1,248 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_METADATA_LITE_H__
-#define GOOGLE_PROTOBUF_METADATA_LITE_H__
-
-#include <string>
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/arena.h>
-#include <google/protobuf/port.h>
-
-#include <google/protobuf/port_def.inc>
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-// This is the representation for messages that support arena allocation. It
-// uses a tagged pointer to either store the Arena pointer, if there are no
-// unknown fields, or a pointer to a block of memory with both the Arena pointer
-// and the UnknownFieldSet, if there are unknown fields. This optimization
-// allows for "zero-overhead" storage of the Arena pointer, relative to the
-// above baseline implementation.
-//
-// The tagged pointer uses the LSB to disambiguate cases, and uses bit 0 == 0 to
-// indicate an arena pointer and bit 0 == 1 to indicate a UFS+Arena-container
-// pointer.
-class InternalMetadata {
- public:
-  constexpr InternalMetadata() : ptr_(nullptr) {}
-  explicit InternalMetadata(Arena* arena) : ptr_(arena) {}
-
-  template <typename T>
-  void Delete() {
-    // Note that Delete<> should be called not more than once.
-    if (have_unknown_fields() && arena() == NULL) {
-      delete PtrValue<Container<T>>();
-    }
-  }
-
-  PROTOBUF_ALWAYS_INLINE Arena* arena() const {
-    if (PROTOBUF_PREDICT_FALSE(have_unknown_fields())) {
-      return PtrValue<ContainerBase>()->arena;
-    } else {
-      return PtrValue<Arena>();
-    }
-  }
-
-  PROTOBUF_ALWAYS_INLINE bool have_unknown_fields() const {
-    return PtrTag() == kTagContainer;
-  }
-
-  PROTOBUF_ALWAYS_INLINE void* raw_arena_ptr() const { return ptr_; }
-
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE const T& unknown_fields(
-      const T& (*default_instance)()) const {
-    if (PROTOBUF_PREDICT_FALSE(have_unknown_fields())) {
-      return PtrValue<Container<T>>()->unknown_fields;
-    } else {
-      return default_instance();
-    }
-  }
-
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE T* mutable_unknown_fields() {
-    if (PROTOBUF_PREDICT_TRUE(have_unknown_fields())) {
-      return &PtrValue<Container<T>>()->unknown_fields;
-    } else {
-      return mutable_unknown_fields_slow<T>();
-    }
-  }
-
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE void Swap(InternalMetadata* other) {
-    // Semantics here are that we swap only the unknown fields, not the arena
-    // pointer. We cannot simply swap ptr_ with other->ptr_ because we need to
-    // maintain our own arena ptr. Also, our ptr_ and other's ptr_ may be in
-    // different states (direct arena pointer vs. container with UFS) so we
-    // cannot simply swap ptr_ and then restore the arena pointers. We reuse
-    // UFS's swap implementation instead.
-    if (have_unknown_fields() || other->have_unknown_fields()) {
-      DoSwap<T>(other->mutable_unknown_fields<T>());
-    }
-  }
-
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE void MergeFrom(const InternalMetadata& other) {
-    if (other.have_unknown_fields()) {
-      DoMergeFrom<T>(other.unknown_fields<T>(nullptr));
-    }
-  }
-
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE void Clear() {
-    if (have_unknown_fields()) {
-      DoClear<T>();
-    }
-  }
-
- private:
-  void* ptr_;
-
-  // Tagged pointer implementation.
-  enum {
-    // ptr_ is an Arena*.
-    kTagArena = 0,
-    // ptr_ is a Container*.
-    kTagContainer = 1,
-  };
-  static constexpr intptr_t kPtrTagMask = 1;
-  static constexpr intptr_t kPtrValueMask = ~kPtrTagMask;
-
-  // Accessors for pointer tag and pointer value.
-  PROTOBUF_ALWAYS_INLINE int PtrTag() const {
-    return reinterpret_cast<intptr_t>(ptr_) & kPtrTagMask;
-  }
-
-  template <typename U>
-  U* PtrValue() const {
-    return reinterpret_cast<U*>(reinterpret_cast<intptr_t>(ptr_) &
-                                kPtrValueMask);
-  }
-
-  // If ptr_'s tag is kTagContainer, it points to an instance of this struct.
-  struct ContainerBase {
-    Arena* arena;
-  };
-
-  template <typename T>
-  struct Container : public ContainerBase {
-    T unknown_fields;
-  };
-
-  template <typename T>
-  PROTOBUF_NOINLINE T* mutable_unknown_fields_slow() {
-    Arena* my_arena = arena();
-    Container<T>* container = Arena::Create<Container<T>>(my_arena);
-    // Two-step assignment works around a bug in clang's static analyzer:
-    // https://bugs.llvm.org/show_bug.cgi?id=34198.
-    ptr_ = container;
-    ptr_ = reinterpret_cast<void*>(reinterpret_cast<intptr_t>(ptr_) |
-                                   kTagContainer);
-    container->arena = my_arena;
-    return &(container->unknown_fields);
-  }
-
-  // Templated functions.
-
-  template <typename T>
-  void DoClear() {
-    mutable_unknown_fields<T>()->Clear();
-  }
-
-  template <typename T>
-  void DoMergeFrom(const T& other) {
-    mutable_unknown_fields<T>()->MergeFrom(other);
-  }
-
-  template <typename T>
-  void DoSwap(T* other) {
-    mutable_unknown_fields<T>()->Swap(other);
-  }
-};
-
-// String Template specializations.
-
-template <>
-inline void InternalMetadata::DoClear<std::string>() {
-  mutable_unknown_fields<std::string>()->clear();
-}
-
-template <>
-inline void InternalMetadata::DoMergeFrom<std::string>(
-    const std::string& other) {
-  mutable_unknown_fields<std::string>()->append(other);
-}
-
-template <>
-inline void InternalMetadata::DoSwap<std::string>(std::string* other) {
-  mutable_unknown_fields<std::string>()->swap(*other);
-}
-
-// This helper RAII class is needed to efficiently parse unknown fields. We
-// should only call mutable_unknown_fields if there are actual unknown fields.
-// The obvious thing to just use a stack string and swap it at the end of
-// the parse won't work, because the destructor of StringOutputStream needs to
-// be called before we can modify the string (it check-fails). Using
-// LiteUnknownFieldSetter setter(&_internal_metadata_);
-// StringOutputStream stream(setter.buffer());
-// guarantees that the string is only swapped after stream is destroyed.
-class PROTOBUF_EXPORT LiteUnknownFieldSetter {
- public:
-  explicit LiteUnknownFieldSetter(InternalMetadata* metadata)
-      : metadata_(metadata) {
-    if (metadata->have_unknown_fields()) {
-      buffer_.swap(*metadata->mutable_unknown_fields<std::string>());
-    }
-  }
-  ~LiteUnknownFieldSetter() {
-    if (!buffer_.empty())
-      metadata_->mutable_unknown_fields<std::string>()->swap(buffer_);
-  }
-  std::string* buffer() { return &buffer_; }
-
- private:
-  InternalMetadata* metadata_;
-  std::string buffer_;
-};
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_METADATA_LITE_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/parse_context.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/parse_context.h
deleted file mode 100644
index 66100858..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/parse_context.h
+++ /dev/null
@@ -1,869 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_PARSE_CONTEXT_H__
-#define GOOGLE_PROTOBUF_PARSE_CONTEXT_H__
-
-#include <cstdint>
-#include <cstring>
-#include <string>
-
-#include <google/protobuf/io/coded_stream.h>
-#include <google/protobuf/io/zero_copy_stream.h>
-#include <google/protobuf/arena.h>
-#include <google/protobuf/arenastring.h>
-#include <google/protobuf/implicit_weak_message.h>
-#include <google/protobuf/metadata_lite.h>
-#include <google/protobuf/port.h>
-#include <google/protobuf/repeated_field.h>
-#include <google/protobuf/wire_format_lite.h>
-#include <google/protobuf/stubs/strutil.h>
-
-#include <google/protobuf/port_def.inc>
-
-
-namespace google {
-namespace protobuf {
-
-class UnknownFieldSet;
-class DescriptorPool;
-class MessageFactory;
-
-namespace internal {
-
-// Template code below needs to know about the existence of these functions.
-PROTOBUF_EXPORT void WriteVarint(uint32 num, uint64 val, std::string* s);
-PROTOBUF_EXPORT void WriteLengthDelimited(uint32 num, StringPiece val,
-                                          std::string* s);
-// Inline because it is just forwarding to s->WriteVarint
-inline void WriteVarint(uint32 num, uint64 val, UnknownFieldSet* s);
-inline void WriteLengthDelimited(uint32 num, StringPiece val,
-                                 UnknownFieldSet* s);
-
-
-// The basic abstraction the parser is designed for is a slight modification
-// of the ZeroCopyInputStream (ZCIS) abstraction. A ZCIS presents a serialized
-// stream as a series of buffers that concatenate to the full stream.
-// Pictorially a ZCIS presents a stream in chunks like so
-// [---------------------------------------------------------------]
-// [---------------------] chunk 1
-//                      [----------------------------] chunk 2
-//                                          chunk 3 [--------------]
-//
-// Where the '-' represent the bytes which are vertically lined up with the
-// bytes of the stream. The proto parser requires its input to be presented
-// similarly with the extra
-// property that each chunk has kSlopBytes past its end that overlaps with the
-// first kSlopBytes of the next chunk, or if there is no next chunk at least its
-// still valid to read those bytes. Again, pictorially, we now have
-//
-// [---------------------------------------------------------------]
-// [-------------------....] chunk 1
-//                    [------------------------....] chunk 2
-//                                    chunk 3 [------------------..**]
-//                                                      chunk 4 [--****]
-// Here '-' mean the bytes of the stream or chunk and '.' means bytes past the
-// chunk that match up with the start of the next chunk. Above each chunk has
-// 4 '.' after the chunk. In the case these 'overflow' bytes represents bytes
-// past the stream, indicated by '*' above, their values are unspecified. It is
-// still legal to read them (ie. should not segfault). Reading past the
-// end should be detected by the user and indicated as an error.
-//
-// The reason for this, admittedly, unconventional invariant is to ruthlessly
-// optimize the protobuf parser. Having an overlap helps in two important ways.
-// Firstly it alleviates having to performing bounds checks if a piece of code
-// is guaranteed to not read more than kSlopBytes. Secondly, and more
-// importantly, the protobuf wireformat is such that reading a key/value pair is
-// always less than 16 bytes. This removes the need to change to next buffer in
-// the middle of reading primitive values. Hence there is no need to store and
-// load the current position.
-
-class PROTOBUF_EXPORT EpsCopyInputStream {
- public:
-  enum { kSlopBytes = 16, kMaxCordBytesToCopy = 512 };
-
-  explicit EpsCopyInputStream(bool enable_aliasing)
-      : aliasing_(enable_aliasing ? kOnPatch : kNoAliasing) {}
-
-  void BackUp(const char* ptr) {
-    GOOGLE_DCHECK(ptr <= buffer_end_ + kSlopBytes);
-    int count;
-    if (next_chunk_ == buffer_) {
-      count = static_cast<int>(buffer_end_ + kSlopBytes - ptr);
-    } else {
-      count = size_ + static_cast<int>(buffer_end_ - ptr);
-    }
-    if (count > 0) StreamBackUp(count);
-  }
-
-  // If return value is negative it's an error
-  PROTOBUF_MUST_USE_RESULT int PushLimit(const char* ptr, int limit) {
-    GOOGLE_DCHECK(limit >= 0 && limit <= INT_MAX - kSlopBytes);
-    // This add is safe due to the invariant above, because
-    // ptr - buffer_end_ <= kSlopBytes.
-    limit += static_cast<int>(ptr - buffer_end_);
-    limit_end_ = buffer_end_ + (std::min)(0, limit);
-    auto old_limit = limit_;
-    limit_ = limit;
-    return old_limit - limit;
-  }
-
-  PROTOBUF_MUST_USE_RESULT bool PopLimit(int delta) {
-    if (PROTOBUF_PREDICT_FALSE(!EndedAtLimit())) return false;
-    limit_ = limit_ + delta;
-    // TODO(gerbens) We could remove this line and hoist the code to
-    // DoneFallback. Study the perf/bin-size effects.
-    limit_end_ = buffer_end_ + (std::min)(0, limit_);
-    return true;
-  }
-
-  PROTOBUF_MUST_USE_RESULT const char* Skip(const char* ptr, int size) {
-    if (size <= buffer_end_ + kSlopBytes - ptr) {
-      return ptr + size;
-    }
-    return SkipFallback(ptr, size);
-  }
-  PROTOBUF_MUST_USE_RESULT const char* ReadString(const char* ptr, int size,
-                                                  std::string* s) {
-    if (size <= buffer_end_ + kSlopBytes - ptr) {
-      s->assign(ptr, size);
-      return ptr + size;
-    }
-    return ReadStringFallback(ptr, size, s);
-  }
-  PROTOBUF_MUST_USE_RESULT const char* AppendString(const char* ptr, int size,
-                                                    std::string* s) {
-    if (size <= buffer_end_ + kSlopBytes - ptr) {
-      s->append(ptr, size);
-      return ptr + size;
-    }
-    return AppendStringFallback(ptr, size, s);
-  }
-
-  template <typename Tag, typename T>
-  PROTOBUF_MUST_USE_RESULT const char* ReadRepeatedFixed(const char* ptr,
-                                                         Tag expected_tag,
-                                                         RepeatedField<T>* out);
-
-  template <typename T>
-  PROTOBUF_MUST_USE_RESULT const char* ReadPackedFixed(const char* ptr,
-                                                       int size,
-                                                       RepeatedField<T>* out);
-  template <typename Add>
-  PROTOBUF_MUST_USE_RESULT const char* ReadPackedVarint(const char* ptr,
-                                                        Add add);
-
-  uint32 LastTag() const { return last_tag_minus_1_ + 1; }
-  bool ConsumeEndGroup(uint32 start_tag) {
-    bool res = last_tag_minus_1_ == start_tag;
-    last_tag_minus_1_ = 0;
-    return res;
-  }
-  bool EndedAtLimit() const { return last_tag_minus_1_ == 0; }
-  bool EndedAtEndOfStream() const { return last_tag_minus_1_ == 1; }
-  void SetLastTag(uint32 tag) { last_tag_minus_1_ = tag - 1; }
-  void SetEndOfStream() { last_tag_minus_1_ = 1; }
-  bool IsExceedingLimit(const char* ptr) {
-    return ptr > limit_end_ &&
-           (next_chunk_ == nullptr || ptr - buffer_end_ > limit_);
-  }
-  int BytesUntilLimit(const char* ptr) const {
-    return limit_ + static_cast<int>(buffer_end_ - ptr);
-  }
-  // Returns true if more data is available, if false is returned one has to
-  // call Done for further checks.
-  bool DataAvailable(const char* ptr) { return ptr < limit_end_; }
-
- protected:
-  // Returns true is limit (either an explicit limit or end of stream) is
-  // reached. It aligns *ptr across buffer seams.
-  // If limit is exceeded it returns true and ptr is set to null.
-  bool DoneWithCheck(const char** ptr, int d) {
-    GOOGLE_DCHECK(*ptr);
-    if (PROTOBUF_PREDICT_TRUE(*ptr < limit_end_)) return false;
-    int overrun = *ptr - buffer_end_;
-    GOOGLE_DCHECK_LE(overrun, kSlopBytes);  // Guaranteed by parse loop.
-    if (overrun ==
-        limit_) {  //  No need to flip buffers if we ended on a limit.
-      // If we actually overrun the buffer and next_chunk_ is null. It means
-      // the stream ended and we passed the stream end.
-      if (overrun > 0 && next_chunk_ == nullptr) *ptr = nullptr;
-      return true;
-    }
-    auto res = DoneFallback(overrun, d);
-    *ptr = res.first;
-    return res.second;
-  }
-
-  const char* InitFrom(StringPiece flat) {
-    overall_limit_ = 0;
-    if (flat.size() > kSlopBytes) {
-      limit_ = kSlopBytes;
-      limit_end_ = buffer_end_ = flat.data() + flat.size() - kSlopBytes;
-      next_chunk_ = buffer_;
-      if (aliasing_ == kOnPatch) aliasing_ = kNoDelta;
-      return flat.data();
-    } else {
-      std::memcpy(buffer_, flat.data(), flat.size());
-      limit_ = 0;
-      limit_end_ = buffer_end_ = buffer_ + flat.size();
-      next_chunk_ = nullptr;
-      if (aliasing_ == kOnPatch) {
-        aliasing_ = reinterpret_cast<std::uintptr_t>(flat.data()) -
-                    reinterpret_cast<std::uintptr_t>(buffer_);
-      }
-      return buffer_;
-    }
-  }
-
-  const char* InitFrom(io::ZeroCopyInputStream* zcis);
-
-  const char* InitFrom(io::ZeroCopyInputStream* zcis, int limit) {
-    if (limit == -1) return InitFrom(zcis);
-    overall_limit_ = limit;
-    auto res = InitFrom(zcis);
-    limit_ = limit - static_cast<int>(buffer_end_ - res);
-    limit_end_ = buffer_end_ + (std::min)(0, limit_);
-    return res;
-  }
-
- private:
-  const char* limit_end_;  // buffer_end_ + min(limit_, 0)
-  const char* buffer_end_;
-  const char* next_chunk_;
-  int size_;
-  int limit_;  // relative to buffer_end_;
-  io::ZeroCopyInputStream* zcis_ = nullptr;
-  char buffer_[2 * kSlopBytes] = {};
-  enum { kNoAliasing = 0, kOnPatch = 1, kNoDelta = 2 };
-  std::uintptr_t aliasing_ = kNoAliasing;
-  // This variable is used to communicate how the parse ended, in order to
-  // completely verify the parsed data. A wire-format parse can end because of
-  // one of the following conditions:
-  // 1) A parse can end on a pushed limit.
-  // 2) A parse can end on End Of Stream (EOS).
-  // 3) A parse can end on 0 tag (only valid for toplevel message).
-  // 4) A parse can end on an end-group tag.
-  // This variable should always be set to 0, which indicates case 1. If the
-  // parse terminated due to EOS (case 2), it's set to 1. In case the parse
-  // ended due to a terminating tag (case 3 and 4) it's set to (tag - 1).
-  // This var doesn't really belong in EpsCopyInputStream and should be part of
-  // the ParseContext, but case 2 is most easily and optimally implemented in
-  // DoneFallback.
-  uint32 last_tag_minus_1_ = 0;
-  int overall_limit_ = INT_MAX;  // Overall limit independent of pushed limits.
-  // Pretty random large number that seems like a safe allocation on most
-  // systems. TODO(gerbens) do we need to set this as build flag?
-  enum { kSafeStringSize = 50000000 };
-
-  // Advances to next buffer chunk returns a pointer to the same logical place
-  // in the stream as set by overrun. Overrun indicates the position in the slop
-  // region the parse was left (0 <= overrun <= kSlopBytes). Returns true if at
-  // limit, at which point the returned pointer maybe null if there was an
-  // error. The invariant of this function is that it's guaranteed that
-  // kSlopBytes bytes can be accessed from the returned ptr. This function might
-  // advance more buffers than one in the underlying ZeroCopyInputStream.
-  std::pair<const char*, bool> DoneFallback(int overrun, int depth);
-  // Advances to the next buffer, at most one call to Next() on the underlying
-  // ZeroCopyInputStream is made. This function DOES NOT match the returned
-  // pointer to where in the slop region the parse ends, hence no overrun
-  // parameter. This is useful for string operations where you always copy
-  // to the end of the buffer (including the slop region).
-  const char* Next();
-  // overrun is the location in the slop region the stream currently is
-  // (0 <= overrun <= kSlopBytes). To prevent flipping to the next buffer of
-  // the ZeroCopyInputStream in the case the parse will end in the last
-  // kSlopBytes of the current buffer. depth is the current depth of nested
-  // groups (or negative if the use case does not need careful tracking).
-  inline const char* NextBuffer(int overrun, int depth);
-  const char* SkipFallback(const char* ptr, int size);
-  const char* AppendStringFallback(const char* ptr, int size, std::string* str);
-  const char* ReadStringFallback(const char* ptr, int size, std::string* str);
-  bool StreamNext(const void** data) {
-    bool res = zcis_->Next(data, &size_);
-    if (res) overall_limit_ -= size_;
-    return res;
-  }
-  void StreamBackUp(int count) {
-    zcis_->BackUp(count);
-    overall_limit_ += count;
-  }
-
-  template <typename A>
-  const char* AppendSize(const char* ptr, int size, const A& append) {
-    int chunk_size = buffer_end_ + kSlopBytes - ptr;
-    do {
-      GOOGLE_DCHECK(size > chunk_size);
-      if (next_chunk_ == nullptr) return nullptr;
-      append(ptr, chunk_size);
-      ptr += chunk_size;
-      size -= chunk_size;
-      // TODO(gerbens) Next calls NextBuffer which generates buffers with
-      // overlap and thus incurs cost of copying the slop regions. This is not
-      // necessary for reading strings. We should just call Next buffers.
-      if (limit_ <= kSlopBytes) return nullptr;
-      ptr = Next();
-      if (ptr == nullptr) return nullptr;  // passed the limit
-      ptr += kSlopBytes;
-      chunk_size = buffer_end_ + kSlopBytes - ptr;
-    } while (size > chunk_size);
-    append(ptr, size);
-    return ptr + size;
-  }
-
-  // AppendUntilEnd appends data until a limit (either a PushLimit or end of
-  // stream. Normal payloads are from length delimited fields which have an
-  // explicit size. Reading until limit only comes when the string takes
-  // the place of a protobuf, ie RawMessage/StringRawMessage, lazy fields and
-  // implicit weak messages. We keep these methods private and friend them.
-  template <typename A>
-  const char* AppendUntilEnd(const char* ptr, const A& append) {
-    if (ptr - buffer_end_ > limit_) return nullptr;
-    while (limit_ > kSlopBytes) {
-      int chunk_size = buffer_end_ + kSlopBytes - ptr;
-      GOOGLE_DCHECK_GE(chunk_size, 0);
-      append(ptr, chunk_size);
-      ptr = Next();
-      if (ptr == nullptr) return limit_end_;
-      ptr += kSlopBytes;
-    }
-    auto end = buffer_end_ + limit_;
-    GOOGLE_DCHECK(end >= ptr);
-    append(ptr, end - ptr);
-    return end;
-  }
-
-  PROTOBUF_MUST_USE_RESULT const char* AppendString(const char* ptr,
-                                                    std::string* str) {
-    return AppendUntilEnd(
-        ptr, [str](const char* p, ptrdiff_t s) { str->append(p, s); });
-  }
-  friend class ImplicitWeakMessage;
-};
-
-// ParseContext holds all data that is global to the entire parse. Most
-// importantly it contains the input stream, but also recursion depth and also
-// stores the end group tag, in case a parser ended on a endgroup, to verify
-// matching start/end group tags.
-class PROTOBUF_EXPORT ParseContext : public EpsCopyInputStream {
- public:
-  struct Data {
-    const DescriptorPool* pool = nullptr;
-    MessageFactory* factory = nullptr;
-  };
-
-  template <typename... T>
-  ParseContext(int depth, bool aliasing, const char** start, T&&... args)
-      : EpsCopyInputStream(aliasing), depth_(depth) {
-    *start = InitFrom(std::forward<T>(args)...);
-  }
-
-  void TrackCorrectEnding() { group_depth_ = 0; }
-
-  bool Done(const char** ptr) { return DoneWithCheck(ptr, group_depth_); }
-
-  int depth() const { return depth_; }
-
-  Data& data() { return data_; }
-  const Data& data() const { return data_; }
-
-  template <typename T>
-  PROTOBUF_MUST_USE_RESULT const char* ParseMessage(T* msg, const char* ptr);
-  // We outline when the type is generic and we go through a virtual
-  const char* ParseMessage(MessageLite* msg, const char* ptr);
-  const char* ParseMessage(Message* msg, const char* ptr);
-
-  template <typename T>
-  PROTOBUF_MUST_USE_RESULT PROTOBUF_ALWAYS_INLINE const char* ParseGroup(
-      T* msg, const char* ptr, uint32 tag) {
-    if (--depth_ < 0) return nullptr;
-    group_depth_++;
-    ptr = msg->_InternalParse(ptr, this);
-    group_depth_--;
-    depth_++;
-    if (PROTOBUF_PREDICT_FALSE(!ConsumeEndGroup(tag))) return nullptr;
-    return ptr;
-  }
-
- private:
-  // The context keeps an internal stack to keep track of the recursive
-  // part of the parse state.
-  // Current depth of the active parser, depth counts down.
-  // This is used to limit recursion depth (to prevent overflow on malicious
-  // data), but is also used to index in stack_ to store the current state.
-  int depth_;
-  // Unfortunately necessary for the fringe case of ending on 0 or end-group tag
-  // in the last kSlopBytes of a ZeroCopyInputStream chunk.
-  int group_depth_ = INT_MIN;
-  Data data_;
-};
-
-template <uint32 tag>
-bool ExpectTag(const char* ptr) {
-  if (tag < 128) {
-    return *ptr == tag;
-  } else {
-    static_assert(tag < 128 * 128, "We only expect tags for 1 or 2 bytes");
-    char buf[2] = {static_cast<char>(tag | 0x80), static_cast<char>(tag >> 7)};
-    return std::memcmp(ptr, buf, 2) == 0;
-  }
-}
-
-template <int>
-struct EndianHelper;
-
-template <>
-struct EndianHelper<1> {
-  static uint8 Load(const void* p) { return *static_cast<const uint8*>(p); }
-};
-
-template <>
-struct EndianHelper<2> {
-  static uint16 Load(const void* p) {
-    uint16 tmp;
-    std::memcpy(&tmp, p, 2);
-#ifndef PROTOBUF_LITTLE_ENDIAN
-    tmp = bswap_16(tmp);
-#endif
-    return tmp;
-  }
-};
-
-template <>
-struct EndianHelper<4> {
-  static uint32 Load(const void* p) {
-    uint32 tmp;
-    std::memcpy(&tmp, p, 4);
-#ifndef PROTOBUF_LITTLE_ENDIAN
-    tmp = bswap_32(tmp);
-#endif
-    return tmp;
-  }
-};
-
-template <>
-struct EndianHelper<8> {
-  static uint64 Load(const void* p) {
-    uint64 tmp;
-    std::memcpy(&tmp, p, 8);
-#ifndef PROTOBUF_LITTLE_ENDIAN
-    tmp = bswap_64(tmp);
-#endif
-    return tmp;
-  }
-};
-
-template <typename T>
-T UnalignedLoad(const char* p) {
-  auto tmp = EndianHelper<sizeof(T)>::Load(p);
-  T res;
-  memcpy(&res, &tmp, sizeof(T));
-  return res;
-}
-
-PROTOBUF_EXPORT
-std::pair<const char*, uint32> VarintParseSlow32(const char* p, uint32 res);
-PROTOBUF_EXPORT
-std::pair<const char*, uint64> VarintParseSlow64(const char* p, uint32 res);
-
-inline const char* VarintParseSlow(const char* p, uint32 res, uint32* out) {
-  auto tmp = VarintParseSlow32(p, res);
-  *out = tmp.second;
-  return tmp.first;
-}
-
-inline const char* VarintParseSlow(const char* p, uint32 res, uint64* out) {
-  auto tmp = VarintParseSlow64(p, res);
-  *out = tmp.second;
-  return tmp.first;
-}
-
-template <typename T>
-PROTOBUF_MUST_USE_RESULT const char* VarintParse(const char* p, T* out) {
-  auto ptr = reinterpret_cast<const uint8*>(p);
-  uint32 res = ptr[0];
-  if (!(res & 0x80)) {
-    *out = res;
-    return p + 1;
-  }
-  uint32 byte = ptr[1];
-  res += (byte - 1) << 7;
-  if (!(byte & 0x80)) {
-    *out = res;
-    return p + 2;
-  }
-  return VarintParseSlow(p, res, out);
-}
-
-// Used for tags, could read up to 5 bytes which must be available.
-// Caller must ensure its safe to call.
-
-PROTOBUF_EXPORT
-std::pair<const char*, uint32> ReadTagFallback(const char* p, uint32 res);
-
-// Same as ParseVarint but only accept 5 bytes at most.
-inline const char* ReadTag(const char* p, uint32* out, uint32 /*max_tag*/ = 0) {
-  uint32 res = static_cast<uint8>(p[0]);
-  if (res < 128) {
-    *out = res;
-    return p + 1;
-  }
-  uint32 second = static_cast<uint8>(p[1]);
-  res += (second - 1) << 7;
-  if (second < 128) {
-    *out = res;
-    return p + 2;
-  }
-  auto tmp = ReadTagFallback(p, res);
-  *out = tmp.second;
-  return tmp.first;
-}
-
-// Decode 2 consecutive bytes of a varint and returns the value, shifted left
-// by 1. It simultaneous updates *ptr to *ptr + 1 or *ptr + 2 depending if the
-// first byte's continuation bit is set.
-// If bit 15 of return value is set (equivalent to the continuation bits of both
-// bytes being set) the varint continues, otherwise the parse is done. On x86
-// movsx eax, dil
-// add edi, eax
-// adc [rsi], 1
-// add eax, eax
-// and eax, edi
-inline uint32 DecodeTwoBytes(const char** ptr) {
-  uint32 value = UnalignedLoad<uint16>(*ptr);
-  // Sign extend the low byte continuation bit
-  uint32_t x = static_cast<int8_t>(value);
-  // This add is an amazing operation, it cancels the low byte continuation bit
-  // from y transferring it to the carry. Simultaneously it also shifts the 7
-  // LSB left by one tightly against high byte varint bits. Hence value now
-  // contains the unpacked value shifted left by 1.
-  value += x;
-  // Use the carry to update the ptr appropriately.
-  *ptr += value < x ? 2 : 1;
-  return value & (x + x);  // Mask out the high byte iff no continuation
-}
-
-// More efficient varint parsing for big varints
-inline const char* ParseBigVarint(const char* p, uint64* out) {
-  auto pnew = p;
-  auto tmp = DecodeTwoBytes(&pnew);
-  uint64 res = tmp >> 1;
-  if (PROTOBUF_PREDICT_TRUE(std::int16_t(tmp) >= 0)) {
-    *out = res;
-    return pnew;
-  }
-  for (std::uint32_t i = 1; i < 5; i++) {
-    pnew = p + 2 * i;
-    tmp = DecodeTwoBytes(&pnew);
-    res += (static_cast<std::uint64_t>(tmp) - 2) << (14 * i - 1);
-    if (PROTOBUF_PREDICT_TRUE(std::int16_t(tmp) >= 0)) {
-      *out = res;
-      return pnew;
-    }
-  }
-  return nullptr;
-}
-
-PROTOBUF_EXPORT
-std::pair<const char*, int32> ReadSizeFallback(const char* p, uint32 first);
-// Used for tags, could read up to 5 bytes which must be available. Additionally
-// it makes sure the unsigned value fits a int32, otherwise returns nullptr.
-// Caller must ensure its safe to call.
-inline uint32 ReadSize(const char** pp) {
-  auto p = *pp;
-  uint32 res = static_cast<uint8>(p[0]);
-  if (res < 128) {
-    *pp = p + 1;
-    return res;
-  }
-  auto x = ReadSizeFallback(p, res);
-  *pp = x.first;
-  return x.second;
-}
-
-// Some convenience functions to simplify the generated parse loop code.
-// Returning the value and updating the buffer pointer allows for nicer
-// function composition. We rely on the compiler to inline this.
-// Also in debug compiles having local scoped variables tend to generated
-// stack frames that scale as O(num fields).
-inline uint64 ReadVarint64(const char** p) {
-  uint64 tmp;
-  *p = VarintParse(*p, &tmp);
-  return tmp;
-}
-
-inline uint32 ReadVarint32(const char** p) {
-  uint32 tmp;
-  *p = VarintParse(*p, &tmp);
-  return tmp;
-}
-
-inline int64 ReadVarintZigZag64(const char** p) {
-  uint64 tmp;
-  *p = VarintParse(*p, &tmp);
-  return WireFormatLite::ZigZagDecode64(tmp);
-}
-
-inline int32 ReadVarintZigZag32(const char** p) {
-  uint64 tmp;
-  *p = VarintParse(*p, &tmp);
-  return WireFormatLite::ZigZagDecode32(static_cast<uint32>(tmp));
-}
-
-template <typename T>
-PROTOBUF_MUST_USE_RESULT const char* ParseContext::ParseMessage(
-    T* msg, const char* ptr) {
-  int size = ReadSize(&ptr);
-  if (!ptr) return nullptr;
-  auto old = PushLimit(ptr, size);
-  if (--depth_ < 0) return nullptr;
-  ptr = msg->_InternalParse(ptr, this);
-  if (PROTOBUF_PREDICT_FALSE(ptr == nullptr)) return nullptr;
-  depth_++;
-  if (!PopLimit(old)) return nullptr;
-  return ptr;
-}
-
-template <typename Add>
-const char* ReadPackedVarintArray(const char* ptr, const char* end, Add add) {
-  while (ptr < end) {
-    uint64 varint;
-    ptr = VarintParse(ptr, &varint);
-    if (ptr == nullptr) return nullptr;
-    add(varint);
-  }
-  return ptr;
-}
-
-template <typename Add>
-const char* EpsCopyInputStream::ReadPackedVarint(const char* ptr, Add add) {
-  int size = ReadSize(&ptr);
-  if (ptr == nullptr) return nullptr;
-  int chunk_size = buffer_end_ - ptr;
-  while (size > chunk_size) {
-    ptr = ReadPackedVarintArray(ptr, buffer_end_, add);
-    if (ptr == nullptr) return nullptr;
-    int overrun = ptr - buffer_end_;
-    GOOGLE_DCHECK(overrun >= 0 && overrun <= kSlopBytes);
-    if (size - chunk_size <= kSlopBytes) {
-      // The current buffer contains all the information needed, we don't need
-      // to flip buffers. However we must parse from a buffer with enough space
-      // so we are not prone to a buffer overflow.
-      char buf[kSlopBytes + 10] = {};
-      std::memcpy(buf, buffer_end_, kSlopBytes);
-      GOOGLE_CHECK_LE(size - chunk_size, kSlopBytes);
-      auto end = buf + (size - chunk_size);
-      auto res = ReadPackedVarintArray(buf + overrun, end, add);
-      if (res == nullptr || res != end) return nullptr;
-      return buffer_end_ + (res - buf);
-    }
-    size -= overrun + chunk_size;
-    GOOGLE_DCHECK_GT(size, 0);
-    // We must flip buffers
-    if (limit_ <= kSlopBytes) return nullptr;
-    ptr = Next();
-    if (ptr == nullptr) return nullptr;
-    ptr += overrun;
-    chunk_size = buffer_end_ - ptr;
-  }
-  auto end = ptr + size;
-  ptr = ReadPackedVarintArray(ptr, end, add);
-  return end == ptr ? ptr : nullptr;
-}
-
-// Helper for verification of utf8
-PROTOBUF_EXPORT
-bool VerifyUTF8(StringPiece s, const char* field_name);
-
-inline bool VerifyUTF8(const std::string* s, const char* field_name) {
-  return VerifyUTF8(*s, field_name);
-}
-
-// All the string parsers with or without UTF checking and for all CTypes.
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* InlineGreedyStringParser(
-    std::string* s, const char* ptr, ParseContext* ctx);
-
-
-// Add any of the following lines to debug which parse function is failing.
-
-#define GOOGLE_PROTOBUF_ASSERT_RETURN(predicate, ret) \
-  if (!(predicate)) {                                  \
-    /*  ::raise(SIGINT);  */                           \
-    /*  GOOGLE_LOG(ERROR) << "Parse failure";  */             \
-    return ret;                                        \
-  }
-
-#define GOOGLE_PROTOBUF_PARSER_ASSERT(predicate) \
-  GOOGLE_PROTOBUF_ASSERT_RETURN(predicate, nullptr)
-
-template <typename T>
-PROTOBUF_MUST_USE_RESULT const char* FieldParser(uint64 tag, T& field_parser,
-                                                 const char* ptr,
-                                                 ParseContext* ctx) {
-  uint32 number = tag >> 3;
-  GOOGLE_PROTOBUF_PARSER_ASSERT(number != 0);
-  using WireType = internal::WireFormatLite::WireType;
-  switch (tag & 7) {
-    case WireType::WIRETYPE_VARINT: {
-      uint64 value;
-      ptr = VarintParse(ptr, &value);
-      GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-      field_parser.AddVarint(number, value);
-      break;
-    }
-    case WireType::WIRETYPE_FIXED64: {
-      uint64 value = UnalignedLoad<uint64>(ptr);
-      ptr += 8;
-      field_parser.AddFixed64(number, value);
-      break;
-    }
-    case WireType::WIRETYPE_LENGTH_DELIMITED: {
-      ptr = field_parser.ParseLengthDelimited(number, ptr, ctx);
-      GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-      break;
-    }
-    case WireType::WIRETYPE_START_GROUP: {
-      ptr = field_parser.ParseGroup(number, ptr, ctx);
-      GOOGLE_PROTOBUF_PARSER_ASSERT(ptr);
-      break;
-    }
-    case WireType::WIRETYPE_END_GROUP: {
-      GOOGLE_LOG(FATAL) << "Can't happen";
-      break;
-    }
-    case WireType::WIRETYPE_FIXED32: {
-      uint32 value = UnalignedLoad<uint32>(ptr);
-      ptr += 4;
-      field_parser.AddFixed32(number, value);
-      break;
-    }
-    default:
-      return nullptr;
-  }
-  return ptr;
-}
-
-template <typename T>
-PROTOBUF_MUST_USE_RESULT const char* WireFormatParser(T& field_parser,
-                                                      const char* ptr,
-                                                      ParseContext* ctx) {
-  while (!ctx->Done(&ptr)) {
-    uint32 tag;
-    ptr = ReadTag(ptr, &tag);
-    GOOGLE_PROTOBUF_PARSER_ASSERT(ptr != nullptr);
-    if (tag == 0 || (tag & 7) == 4) {
-      ctx->SetLastTag(tag);
-      return ptr;
-    }
-    ptr = FieldParser(tag, field_parser, ptr, ctx);
-    GOOGLE_PROTOBUF_PARSER_ASSERT(ptr != nullptr);
-  }
-  return ptr;
-}
-
-// The packed parsers parse repeated numeric primitives directly into  the
-// corresponding field
-
-// These are packed varints
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* PackedInt32Parser(
-    void* object, const char* ptr, ParseContext* ctx);
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* PackedUInt32Parser(
-    void* object, const char* ptr, ParseContext* ctx);
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* PackedInt64Parser(
-    void* object, const char* ptr, ParseContext* ctx);
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* PackedUInt64Parser(
-    void* object, const char* ptr, ParseContext* ctx);
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* PackedSInt32Parser(
-    void* object, const char* ptr, ParseContext* ctx);
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* PackedSInt64Parser(
-    void* object, const char* ptr, ParseContext* ctx);
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* PackedEnumParser(
-    void* object, const char* ptr, ParseContext* ctx);
-
-template <typename T>
-PROTOBUF_MUST_USE_RESULT const char* PackedEnumParser(
-    void* object, const char* ptr, ParseContext* ctx, bool (*is_valid)(int),
-    InternalMetadata* metadata, int field_num) {
-  return ctx->ReadPackedVarint(
-      ptr, [object, is_valid, metadata, field_num](uint64 val) {
-        if (is_valid(val)) {
-          static_cast<RepeatedField<int>*>(object)->Add(val);
-        } else {
-          WriteVarint(field_num, val, metadata->mutable_unknown_fields<T>());
-        }
-      });
-}
-
-template <typename T>
-PROTOBUF_MUST_USE_RESULT const char* PackedEnumParserArg(
-    void* object, const char* ptr, ParseContext* ctx,
-    bool (*is_valid)(const void*, int), const void* data,
-    InternalMetadata* metadata, int field_num) {
-  return ctx->ReadPackedVarint(
-      ptr, [object, is_valid, data, metadata, field_num](uint64 val) {
-        if (is_valid(data, val)) {
-          static_cast<RepeatedField<int>*>(object)->Add(val);
-        } else {
-          WriteVarint(field_num, val, metadata->mutable_unknown_fields<T>());
-        }
-      });
-}
-
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* PackedBoolParser(
-    void* object, const char* ptr, ParseContext* ctx);
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* PackedFixed32Parser(
-    void* object, const char* ptr, ParseContext* ctx);
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* PackedSFixed32Parser(
-    void* object, const char* ptr, ParseContext* ctx);
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* PackedFixed64Parser(
-    void* object, const char* ptr, ParseContext* ctx);
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* PackedSFixed64Parser(
-    void* object, const char* ptr, ParseContext* ctx);
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* PackedFloatParser(
-    void* object, const char* ptr, ParseContext* ctx);
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* PackedDoubleParser(
-    void* object, const char* ptr, ParseContext* ctx);
-
-// This is the only recursive parser.
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* UnknownGroupLiteParse(
-    std::string* unknown, const char* ptr, ParseContext* ctx);
-// This is a helper to for the UnknownGroupLiteParse but is actually also
-// useful in the generated code. It uses overload on std::string* vs
-// UnknownFieldSet* to make the generated code isomorphic between full and lite.
-PROTOBUF_EXPORT PROTOBUF_MUST_USE_RESULT const char* UnknownFieldParse(
-    uint32 tag, std::string* unknown, const char* ptr, ParseContext* ctx);
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_PARSE_CONTEXT_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/port.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/port.h
deleted file mode 100644
index 555fd4eb..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/port.h
+++ /dev/null
@@ -1,43 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// A common header that is included across all protobuf headers.  We do our best
-// to avoid #defining any macros here; instead we generally put macros in
-// port_def.inc and port_undef.inc so they are not visible from outside of
-// protobuf.
-
-#ifndef GOOGLE_PROTOBUF_PORT_H__
-#define GOOGLE_PROTOBUF_PORT_H__
-
-
-#include <google/protobuf/stubs/port.h>
-
-
-#endif  // GOOGLE_PROTOBUF_PORT_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/port_def.inc b/projects/llm_framework/include/protobuf-lite/google/protobuf/port_def.inc
deleted file mode 100644
index 320e888e..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/port_def.inc
+++ /dev/null
@@ -1,606 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// This file defines common macros that are used in protobuf.
-//
-// To hide these definitions from the outside world (and to prevent collisions
-// if more than one version of protobuf is #included in the same project) you
-// must follow this pattern when #including port_def.inc in a header file:
-//
-// #include "other_header.h"
-// #include "message.h"
-// // etc.
-//
-// #include "port_def.inc"  // MUST be last header included
-//
-// // Definitions for this header.
-//
-// #include "port_undef.inc"
-//
-// This is a textual header with no include guard, because we want to
-// detect/prohibit anytime it is #included twice without a corresponding
-// #undef.
-
-// These macros are private and should always be
-// ::util::RetrieveErrorSpace(*this) headers. If any of these errors fire, you
-// should either properly #include port_undef.h at the end of your header that
-// #includes port.h, or don't #include port.h twice in a .cc file.
-#ifdef PROTOBUF_NAMESPACE
-#error PROTOBUF_NAMESPACE was previously defined
-#endif
-#ifdef PROTOBUF_NAMESPACE_ID
-#error PROTOBUF_NAMESPACE_ID was previously defined
-#endif
-#ifdef PROTOBUF_ALWAYS_INLINE
-#error PROTOBUF_ALWAYS_INLINE was previously defined
-#endif
-#ifdef PROTOBUF_COLD
-#error PROTOBUF_COLD was previously defined
-#endif
-#ifdef PROTOBUF_NOINLINE
-#error PROTOBUF_NOINLINE was previously defined
-#endif
-#ifdef PROTOBUF_SECTION_VARIABLE
-#error PROTOBUF_SECTION_VARIABLE was previously defined
-#endif
-#ifdef PROTOBUF_DEPRECATED
-#error PROTOBUF_DEPRECATED was previously defined
-#endif
-#ifdef PROTOBUF_DEPRECATED_MSG
-#error PROTOBUF_DEPRECATED_MSG was previously defined
-#endif
-#ifdef PROTOBUF_FUNC_ALIGN
-#error PROTOBUF_FUNC_ALIGN was previously defined
-#endif
-#ifdef PROTOBUF_RETURNS_NONNULL
-#error PROTOBUF_RETURNS_NONNULL was previously defined
-#endif
-#ifdef PROTOBUF_ATTRIBUTE_REINITIALIZES
-#error PROTOBUF_ATTRIBUTE_REINITIALIZES was previously defined
-#endif
-#ifdef PROTOBUF_RTTI
-#error PROTOBUF_RTTI was previously defined
-#endif
-#ifdef PROTOBUF_VERSION
-#error PROTOBUF_VERSION was previously defined
-#endif
-#ifdef PROTOBUF_VERSION_SUFFIX
-#error PROTOBUF_VERSION_SUFFIX was previously defined
-#endif
-#ifdef PROTOBUF_MIN_HEADER_VERSION_FOR_PROTOC
-#error PROTOBUF_MIN_HEADER_VERSION_FOR_PROTOC was previously defined
-#endif
-#ifdef PROTOBUF_MIN_PROTOC_VERSION
-#error PROTOBUF_MIN_PROTOC_VERSION was previously defined
-#endif
-#ifdef PROTOBUF_PREDICT_TRUE
-#error PROTOBUF_PREDICT_TRUE was previously defined
-#endif
-#ifdef PROTOBUF_PREDICT_FALSE
-#error PROTOBUF_PREDICT_FALSE was previously defined
-#endif
-#ifdef PROTOBUF_FIELD_OFFSET
-#error PROTOBUF_FIELD_OFFSET was previously defined
-#endif
-#ifdef PROTOBUF_LL_FORMAT
-#error PROTOBUF_LL_FORMAT was previously defined
-#endif
-#ifdef PROTOBUF_GUARDED_BY
-#error PROTOBUF_GUARDED_BY was previously defined
-#endif
-#ifdef PROTOBUF_LONGLONG
-#error PROTOBUF_LONGLONG was previously defined
-#endif
-#ifdef PROTOBUF_ULONGLONG
-#error PROTOBUF_ULONGLONG was previously defined
-#endif
-#ifdef PROTOBUF_FALLTHROUGH_INTENDED
-#error PROTOBUF_FALLTHROUGH_INTENDED was previously defined
-#endif
-#ifdef PROTOBUF_EXPORT
-#error PROTOBUF_EXPORT was previously defined
-#endif
-#ifdef PROTOC_EXPORT
-#error PROTOC_EXPORT was previously defined
-#endif
-#ifdef PROTOBUF_MUST_USE_RESULT
-#error PROTOBUF_MUST_USE_RESULT was previously defined
-#endif
-#ifdef PROTOBUF_UNUSED
-#error PROTOBUF_UNUSED was previously defined
-#endif
-#ifdef PROTOBUF_FINAL
-#error PROTOBUF_FINAL was previously defined
-#endif
-#ifdef PROTOBUF_DISABLE_MSVC_UNION_WARNING
-#error PROTOBUF_DISABLE_MSVC_UNION_WARNING was previously defined
-#endif
-#ifdef PROTOBUF_ENABLE_MSVC_UNION_WARNING
-#error PROTOBUF_ENABLE_MSVC_UNION_WARNING was previously defined
-#endif
-#ifdef PROTOBUF_CONSTINIT
-#error PROTOBUF_CONSTINIT was previously defined
-#endif
-#ifdef PROTOBUF_MAYBE_CONSTEXPR
-#error PROTOBUF_MAYBE_CONSTEXPR was previously defined
-#endif
-#ifdef PROTOBUF_ATTRIBUTE_NO_DESTROY
-#error PROTOBUF_ATTRIBUTE_NO_DESTROY was previously defined
-#endif
-
-
-#define PROTOBUF_NAMESPACE "google::protobuf"
-#define PROTOBUF_NAMESPACE_ID google::protobuf
-#define PROTOBUF_NAMESPACE_OPEN \
-  namespace google {            \
-  namespace protobuf {
-#define PROTOBUF_NAMESPACE_CLOSE \
-  } /* namespace protobuf */     \
-  } /* namespace google */
-
-#if defined(__GNUC__) || defined(__clang__)
-#define PROTOBUF_DEPRECATED __attribute__((deprecated))
-#define PROTOBUF_DEPRECATED_ENUM __attribute__((deprecated))
-#define PROTOBUF_DEPRECATED_MSG(msg) __attribute__((deprecated(msg)))
-#elif defined(_MSC_VER)
-#define PROTOBUF_DEPRECATED __declspec(deprecated)
-#define PROTOBUF_DEPRECATED_ENUM
-#define PROTOBUF_DEPRECATED_MSG(msg) __declspec(deprecated(msg))
-#endif
-
-#define PROTOBUF_SECTION_VARIABLE(x)
-#define PROTOBUF_MUST_USE_RESULT
-
-// ----------------------------------------------------------------------------
-// Annotations:  Some parts of the code have been annotated in ways that might
-//   be useful to some compilers or tools, but are not supported universally.
-//   You can #define these annotations yourself if the default implementation
-//   is not right for you.
-
-#ifdef GOOGLE_ATTRIBUTE_ALWAYS_INLINE
-#define PROTOBUF_ALWAYS_INLINE GOOGLE_ATTRIBUTE_ALWAYS_INLINE
-#else
-#if defined(__GNUC__) && \
-    (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 1))
-// For functions we want to force inline.
-// Introduced in gcc 3.1.
-#define PROTOBUF_ALWAYS_INLINE __attribute__((always_inline))
-#else
-// Other compilers will have to figure it out for themselves.
-#define PROTOBUF_ALWAYS_INLINE
-#endif
-#endif
-
-#ifdef GOOGLE_ATTRIBUTE_NOINLINE
-#define PROTOBUF_NOINLINE GOOGLE_ATTRIBUTE_NOINLINE
-#else
-#if defined(__GNUC__) && \
-    (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 1))
-// For functions we want to force not inline.
-// Introduced in gcc 3.1.
-#define PROTOBUF_NOINLINE __attribute__((noinline))
-#elif defined(_MSC_VER) && (_MSC_VER >= 1400)
-// Seems to have been around since at least Visual Studio 2005
-#define PROTOBUF_NOINLINE __declspec(noinline)
-#else
-// Other compilers will have to figure it out for themselves.
-#define PROTOBUF_NOINLINE
-#endif
-#endif
-
-#ifdef GOOGLE_ATTRIBUTE_FUNC_ALIGN
-#define PROTOBUF_FUNC_ALIGN GOOGLE_ATTRIBUTE_FUNC_ALIGN
-#else
-#if defined(__clang__) || \
-    defined(__GNUC__) &&  \
-        (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3))
-// Function alignment attribute introduced in gcc 4.3
-#define PROTOBUF_FUNC_ALIGN(bytes) __attribute__((aligned(bytes)))
-#else
-#define PROTOBUF_FUNC_ALIGN(bytes)
-#endif
-#endif
-
-#ifdef GOOGLE_PREDICT_TRUE
-#define PROTOBUF_PREDICT_TRUE GOOGLE_PREDICT_TRUE
-#else
-#ifdef __GNUC__
-// Provided at least since GCC 3.0.
-#define PROTOBUF_PREDICT_TRUE(x) (__builtin_expect(!!(x), 1))
-#else
-#define PROTOBUF_PREDICT_TRUE(x) (x)
-#endif
-#endif
-
-#ifdef GOOGLE_PREDICT_FALSE
-#define PROTOBUF_PREDICT_FALSE GOOGLE_PREDICT_FALSE
-#else
-#ifdef __GNUC__
-// Provided at least since GCC 3.0.
-#define PROTOBUF_PREDICT_FALSE(x) (__builtin_expect(x, 0))
-#else
-#define PROTOBUF_PREDICT_FALSE(x) (x)
-#endif
-#endif
-
-#ifdef GOOGLE_PROTOBUF_ATTRIBUTE_RETURNS_NONNULL
-#define PROTOBUF_RETURNS_NONNULL GOOGLE_PROTOBUF_ATTRIBUTE_RETURNS_NONNULL
-#else
-#if defined(__has_attribute)
-#if __has_attribute(returns_nonnull)
-#define PROTOBUF_RETURNS_NONNULL __attribute__((returns_nonnull))
-#endif
-#endif
-#endif
-#ifndef PROTOBUF_RETURNS_NONNULL
-#define PROTOBUF_RETURNS_NONNULL
-#endif
-
-#if defined(__has_cpp_attribute)
-#if __has_cpp_attribute(clang::reinitializes)
-#define PROTOBUF_ATTRIBUTE_REINITIALIZES [[clang::reinitializes]]
-#endif
-#endif
-#ifndef PROTOBUF_ATTRIBUTE_REINITIALIZES
-#define PROTOBUF_ATTRIBUTE_REINITIALIZES
-#endif
-
-#define PROTOBUF_GUARDED_BY(x)
-#define PROTOBUF_COLD
-
-// Copied from ABSL.
-#if defined(__clang__) && defined(__has_warning)
-#if __has_feature(cxx_attributes) && __has_warning("-Wimplicit-fallthrough")
-#define PROTOBUF_FALLTHROUGH_INTENDED [[clang::fallthrough]]
-#endif
-#elif defined(__GNUC__) && __GNUC__ >= 7
-#define PROTOBUF_FALLTHROUGH_INTENDED [[gnu::fallthrough]]
-#endif
-
-#ifndef PROTOBUF_FALLTHROUGH_INTENDED
-#define PROTOBUF_FALLTHROUGH_INTENDED
-#endif
-
-#if defined(__has_cpp_attribute)
-#define HAS_ATTRIBUTE(attr) __has_cpp_attribute(attr)
-#else
-#define HAS_ATTRIBUTE(attr) 0
-#endif
-
-#if HAS_ATTRIBUTE(unused) || (defined(__GNUC__) && !defined(__clang__))
-#define PROTOBUF_UNUSED __attribute__((__unused__))
-#else
-#define PROTOBUF_UNUSED
-#endif
-
-#undef HAS_ATTRIBUTE
-
-#ifdef _MSC_VER
-#define PROTOBUF_LONGLONG(x) x##I64
-#define PROTOBUF_ULONGLONG(x) x##UI64
-#define PROTOBUF_LL_FORMAT "I64"  // As in printf("%I64d", ...)
-#else
-// By long long, we actually mean int64.
-#define PROTOBUF_LONGLONG(x) x##LL
-#define PROTOBUF_ULONGLONG(x) x##ULL
-// Used to format real long long integers.
-#define PROTOBUF_LL_FORMAT \
-  "ll"  // As in "%lld". Note that "q" is poor form also.
-#endif
-
-
-// Shared google3/opensource definitions. //////////////////////////////////////
-
-#define PROTOBUF_VERSION 3014000
-#define PROTOBUF_MIN_HEADER_VERSION_FOR_PROTOC 3014000
-#define PROTOBUF_MIN_PROTOC_VERSION 3014000
-#define PROTOBUF_VERSION_SUFFIX ""
-
-// The minimum library version which works with the current version of the
-// headers.
-#define GOOGLE_PROTOBUF_MIN_LIBRARY_VERSION 3014000
-
-#if defined(GOOGLE_PROTOBUF_NO_RTTI) && GOOGLE_PROTOBUF_NO_RTTI
-#define PROTOBUF_RTTI 0
-#elif defined(__has_feature)
-// https://clang.llvm.org/docs/LanguageExtensions.html#has-feature-and-has-extension
-#define PROTOBUF_RTTI __has_feature(cxx_rtti)
-#elif !defined(__cxx_rtti)
-// https://en.cppreference.com/w/User:D41D8CD98F/feature_testing_macros#C.2B.2B98
-#define PROTOBUF_RTTI 0
-#elif defined(__GNUC__) && !defined(__GXX_RTTI)
-#https:  // gcc.gnu.org/onlinedocs/cpp/Common-Predefined-Macros.html
-#define PROTOBUF_RTTI 0
-#else
-#define PROTOBUF_RTTI 1
-#endif
-
-// Returns the offset of the given field within the given aggregate type.
-// This is equivalent to the ANSI C offsetof() macro.  However, according
-// to the C++ standard, offsetof() only works on POD types, and GCC
-// enforces this requirement with a warning.  In practice, this rule is
-// unnecessarily strict; there is probably no compiler or platform on
-// which the offsets of the direct fields of a class are non-constant.
-// Fields inherited from superclasses *can* have non-constant offsets,
-// but that's not what this macro will be used for.
-#if defined(__clang__)
-// For Clang we use __builtin_offsetof() and suppress the warning,
-// to avoid Control Flow Integrity and UBSan vptr sanitizers from
-// crashing while trying to validate the invalid reinterpet_casts.
-#define PROTOBUF_FIELD_OFFSET(TYPE, FIELD)                   \
-  _Pragma("clang diagnostic push")                           \
-  _Pragma("clang diagnostic ignored \"-Winvalid-offsetof\"") \
-  __builtin_offsetof(TYPE, FIELD)                            \
-  _Pragma("clang diagnostic pop")
-#elif defined(__GNUC__) && \
-          (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8))
-#define PROTOBUF_FIELD_OFFSET(TYPE, FIELD) __builtin_offsetof(TYPE, FIELD)
-#else  // defined(__clang__)
-// Note that we calculate relative to the pointer value 16 here since if we
-// just use zero, GCC complains about dereferencing a NULL pointer.  We
-// choose 16 rather than some other number just in case the compiler would
-// be confused by an unaligned pointer.
-#define PROTOBUF_FIELD_OFFSET(TYPE, FIELD)                                \
-  static_cast< ::google::protobuf::uint32>(reinterpret_cast<const char*>(                   \
-                             &reinterpret_cast<const TYPE*>(16)->FIELD) - \
-                         reinterpret_cast<const char*>(16))
-#endif
-
-#if defined(PROTOBUF_USE_DLLS)
-#if defined(_MSC_VER)
-#ifdef LIBPROTOBUF_EXPORTS
-#define PROTOBUF_EXPORT __declspec(dllexport)
-#define PROTOBUF_EXPORT_TEMPLATE_DECLARE
-#define PROTOBUF_EXPORT_TEMPLATE_DEFINE __declspec(dllexport)
-#else
-#define PROTOBUF_EXPORT __declspec(dllimport)
-#define PROTOBUF_EXPORT_TEMPLATE_DECLARE
-#define PROTOBUF_EXPORT_TEMPLATE_DEFINE __declspec(dllimport)
-#endif
-#ifdef LIBPROTOC_EXPORTS
-#define PROTOC_EXPORT __declspec(dllexport)
-#else
-#define PROTOC_EXPORT __declspec(dllimport)
-#endif
-#else  // defined(_MSC_VER)
-#ifdef LIBPROTOBUF_EXPORTS
-#define PROTOBUF_EXPORT __attribute__((visibility("default")))
-#define PROTOBUF_EXPORT_TEMPLATE_DECLARE __attribute__((visibility("default")))
-#define PROTOBUF_EXPORT_TEMPLATE_DEFINE
-#else
-#define PROTOBUF_EXPORT
-#define PROTOBUF_EXPORT_TEMPLATE_DECLARE
-#define PROTOBUF_EXPORT_TEMPLATE_DEFINE
-#endif
-#ifdef LIBPROTOC_EXPORTS
-#define PROTOC_EXPORT __attribute__((visibility("default")))
-#else
-#define PROTOC_EXPORT
-#endif
-#endif
-#else  // defined(PROTOBUF_USE_DLLS)
-#define PROTOBUF_EXPORT
-#define PROTOC_EXPORT
-#define PROTOBUF_EXPORT_TEMPLATE_DECLARE
-#define PROTOBUF_EXPORT_TEMPLATE_DEFINE
-#endif
-
-// Windows declares several inconvenient macro names.  We #undef them and then
-// restore them in port_undef.inc.
-#ifdef _MSC_VER
-#pragma push_macro("CREATE_NEW")
-#undef CREATE_NEW
-#pragma push_macro("DOUBLE_CLICK")
-#undef DOUBLE_CLICK
-#pragma push_macro("ERROR")
-#undef ERROR
-#pragma push_macro("ERROR_BUSY")
-#undef ERROR_BUSY
-#pragma push_macro("ERROR_NOT_FOUND")
-#undef ERROR_NOT_FOUND
-#pragma push_macro("GetMessage")
-#undef GetMessage
-#pragma push_macro("IGNORE")
-#undef IGNORE
-#pragma push_macro("IN")
-#undef IN
-#pragma push_macro("INPUT_KEYBOARD")
-#undef INPUT_KEYBOARD
-#pragma push_macro("NO_ERROR")
-#undef NO_ERROR
-#pragma push_macro("OUT")
-#undef OUT
-#pragma push_macro("OPTIONAL")
-#undef OPTIONAL
-#pragma push_macro("min")
-#undef min
-#pragma push_macro("max")
-#undef max
-#pragma push_macro("NEAR")
-#undef NEAR
-#pragma push_macro("NO_DATA")
-#undef NO_DATA
-#pragma push_macro("REASON_UNKNOWN")
-#undef REASON_UNKNOWN
-#pragma push_macro("SERVICE_DISABLED")
-#undef SERVICE_DISABLED
-#pragma push_macro("SEVERITY_ERROR")
-#undef SEVERITY_ERROR
-#pragma push_macro("STRICT")
-#undef STRICT
-#pragma push_macro("timezone")
-#undef timezone
-#endif  // _MSC_VER
-
-#if defined(__clang__) || defined(__GNUC__) || defined(_MSC_VER)
-// Don't let Objective-C Macros interfere with proto identifiers with the same 
-// name.
-#pragma push_macro("DEBUG")
-#undef DEBUG
-#pragma push_macro("TRUE")
-#undef TRUE
-#pragma push_macro("FALSE")
-#undef FALSE
-#endif // defined(__clang__) || defined(__GNUC__) || defined(_MSC_VER)
-
-#if defined(__clang__)
-#pragma clang diagnostic push
-// TODO(gerbens) ideally we cleanup the code. But a cursory try shows many
-// violations. So let's ignore for now.
-#pragma clang diagnostic ignored "-Wshorten-64-to-32"
-#elif defined(__GNUC__)
-// GCC does not allow disabling diagnostics within an expression:
-// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=60875, so we disable this one
-// globally even though it's only used for PROTOBUF_FIELD_OFFSET.
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Winvalid-offsetof"
-#endif
-
-// PROTOBUF_ASSUME(pred) tells the compiler that it can assume pred is true. To
-// be safe, we also validate the assumption with a GOOGLE_DCHECK in unoptimized
-// builds. The macro does not do anything useful if the compiler does not
-// support __builtin_assume.
-#ifdef __has_builtin
-#if __has_builtin(__builtin_assume)
-#define PROTOBUF_ASSUME(pred) \
-  GOOGLE_DCHECK(pred);               \
-  __builtin_assume(pred)
-#else
-#define PROTOBUF_ASSUME(pred) GOOGLE_DCHECK(pred)
-#endif
-#else
-#define PROTOBUF_ASSUME(pred) GOOGLE_DCHECK(pred)
-#endif
-
-// Specify memory alignment for structs, classes, etc.
-// Use like:
-//   class PROTOBUF_ALIGNAS(16) MyClass { ... }
-//   PROTOBUF_ALIGNAS(16) int array[4];
-//
-// In most places you can use the C++11 keyword "alignas", which is preferred.
-//
-// But compilers have trouble mixing __attribute__((...)) syntax with
-// alignas(...) syntax.
-//
-// Doesn't work in clang or gcc:
-//   struct alignas(16) __attribute__((packed)) S { char c; };
-// Works in clang but not gcc:
-//   struct __attribute__((packed)) alignas(16) S2 { char c; };
-// Works in clang and gcc:
-//   struct alignas(16) S3 { char c; } __attribute__((packed));
-//
-// There are also some attributes that must be specified *before* a class
-// definition: visibility (used for exporting functions/classes) is one of
-// these attributes. This means that it is not possible to use alignas() with a
-// class that is marked as exported.
-#if defined(_MSC_VER)
-#define PROTOBUF_ALIGNAS(byte_alignment) __declspec(align(byte_alignment))
-#elif defined(__GNUC__)
-#define PROTOBUF_ALIGNAS(byte_alignment) \
-  __attribute__((aligned(byte_alignment)))
-#else
-#define PROTOBUF_ALIGNAS(byte_alignment) alignas(byte_alignment)
-#endif
-
-#define PROTOBUF_FINAL final
-
-#if defined(_MSC_VER)
-#define PROTOBUF_THREAD_LOCAL __declspec(thread)
-#else
-#define PROTOBUF_THREAD_LOCAL __thread
-#endif
-
-// For enabling message owned arena, one major blocker is semantic change from
-// moving to copying when there is ownership transfer (e.g., move ctor, swap,
-// set allocated, release). This change not only causes performance regression
-// but also breaks users code (e.g., dangling reference). For top-level
-// messages, since it owns the arena, we can mitigate the issue by transferring
-// ownership of arena. However, we cannot do that for nested messages. In order
-// to tell how many usages of nested messages affected by message owned arena,
-// we need to simulate the arena ownership.
-// This experiment is purely for the purpose of gathering data. All code guarded
-// by this flag is supposed to be removed after this experiment.
-// #define PROTOBUF_MESSAGE_OWNED_ARENA_EXPERIMENT
-
-#if defined(__cpp_constinit)
-#define PROTOBUF_CONSTINIT constinit
-#elif defined(__has_cpp_attribute)
-#if __has_cpp_attribute(clang::require_constant_initialization)
-#define PROTOBUF_CONSTINIT [[clang::require_constant_initialization]]
-#endif
-#endif
-#ifndef PROTOBUF_CONSTINIT
-#define PROTOBUF_CONSTINIT
-#endif
-
-// Some constructors can't be constexpr under MSVC, but given that MSVC will not
-// do constant initialization of globals anyway we can omit `constexpr` from
-// them. These constructors are marked with PROTOBUF_MAYBE_CONSTEXPR
-#if defined(_MSC_VER)
-#define PROTOBUF_MAYBE_CONSTEXPR
-#else
-#define PROTOBUF_MAYBE_CONSTEXPR constexpr
-#endif
-
-#if _MSC_VER
-#define PROTOBUF_DISABLE_MSVC_UNION_WARNING \
-  __pragma(warning(push)) \
-  __pragma(warning(disable : 4582)) \
-  __pragma(warning(disable : 4583))
-
-#define PROTOBUF_ENABLE_MSVC_UNION_WARNING \
-  __pragma(warning(pop))
-#else
-#define PROTOBUF_DISABLE_MSVC_UNION_WARNING
-#define PROTOBUF_ENABLE_MSVC_UNION_WARNING
-#endif
-
-#if defined(__cpp_constinit)
-#define PROTOBUF_CONSTINIT constinit
-#elif defined(__has_cpp_attribute)
-#if __has_cpp_attribute(clang::require_constant_initialization)
-#define PROTOBUF_CONSTINIT [[clang::require_constant_initialization]]
-#endif
-#endif
-#ifndef PROTOBUF_CONSTINIT
-#define PROTOBUF_CONSTINIT
-#endif
-
-// Some globals with an empty non-trivial destructor are annotated with
-// no_destroy for performance reasons. It reduces the cost of these globals in
-// non-opt mode and under sanitizers.
-#if defined(__has_cpp_attribute)
-#if __has_cpp_attribute(clang::no_destroy)
-#define PROTOBUF_ATTRIBUTE_NO_DESTROY [[clang::no_destroy]]
-#endif
-#endif
-#if !defined(PROTOBUF_ATTRIBUTE_NO_DESTROY)
-#define PROTOBUF_ATTRIBUTE_NO_DESTROY
-#endif
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/port_undef.inc b/projects/llm_framework/include/protobuf-lite/google/protobuf/port_undef.inc
deleted file mode 100644
index d1414285..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/port_undef.inc
+++ /dev/null
@@ -1,116 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// #undefs all macros defined in port_def.inc.  See comments in port_def.inc
-// for more info.
-
-#ifndef PROTOBUF_NAMESPACE
-#error "port_undef.inc must be included after port_def.inc"
-#endif
-#undef PROTOBUF_NAMESPACE
-#undef PROTOBUF_NAMESPACE_ID
-#undef PROTOBUF_ALWAYS_INLINE
-#undef PROTOBUF_COLD
-#undef PROTOBUF_NOINLINE
-#undef PROTOBUF_SECTION_VARIABLE
-#undef PROTOBUF_DEPRECATED
-#undef PROTOBUF_DEPRECATED_ENUM
-#undef PROTOBUF_DEPRECATED_MSG
-#undef PROTOBUF_FUNC_ALIGN
-#undef PROTOBUF_RETURNS_NONNULL
-#undef PROTOBUF_ATTRIBUTE_REINITIALIZES
-#undef PROTOBUF_RTTI
-#undef PROTOBUF_VERSION
-#undef PROTOBUF_VERSION_SUFFIX
-#undef PROTOBUF_FIELD_OFFSET
-#undef PROTOBUF_MIN_HEADER_VERSION_FOR_PROTOC
-#undef PROTOBUF_MIN_PROTOC_VERSION
-#undef PROTOBUF_PREDICT_TRUE
-#undef PROTOBUF_PREDICT_FALSE
-#undef PROTOBUF_LONGLONG
-#undef PROTOBUF_ULONGLONG
-#undef PROTOBUF_LL_FORMAT
-#undef PROTOBUF_GUARDED_BY
-#undef PROTOBUF_FALLTHROUGH_INTENDED
-#undef PROTOBUF_EXPORT
-#undef PROTOC_EXPORT
-#undef PROTOBUF_MUST_USE_RESULT
-#undef PROTOBUF_NAMESPACE_OPEN
-#undef PROTOBUF_NAMESPACE_CLOSE
-#undef PROTOBUF_UNUSED
-#undef PROTOBUF_ASSUME
-#undef PROTOBUF_EXPORT_TEMPLATE_DECLARE
-#undef PROTOBUF_EXPORT_TEMPLATE_DEFINE
-#undef PROTOBUF_ALIGNAS
-#undef PROTOBUF_FINAL
-#undef PROTOBUF_THREAD_LOCAL
-#undef PROTOBUF_MESSAGE_OWNED_ARENA_EXPERIMENT
-#undef PROTOBUF_DISABLE_MSVC_UNION_WARNING
-#undef PROTOBUF_ENABLE_MSVC_UNION_WARNING
-#undef PROTOBUF_CONSTINIT
-#undef PROTOBUF_MAYBE_CONSTEXPR
-#undef PROTOBUF_ATTRIBUTE_NO_DESTROY
-
-// Restore macro that may have been #undef'd in port_def.inc.
-#ifdef _MSC_VER
-#pragma pop_macro("CREATE_NEW")
-#pragma pop_macro("DOUBLE_CLICK")
-#pragma pop_macro("ERROR")
-#pragma pop_macro("ERROR_BUSY")
-#pragma pop_macro("ERROR_NOT_FOUND")
-#pragma pop_macro("GetMessage")
-#pragma pop_macro("IGNORE")
-#pragma pop_macro("IN")
-#pragma pop_macro("INPUT_KEYBOARD")
-#pragma pop_macro("OUT")
-#pragma pop_macro("OPTIONAL")
-#pragma pop_macro("min")
-#pragma pop_macro("max")
-#pragma pop_macro("NEAR")
-#pragma pop_macro("NO_DATA")
-#pragma pop_macro("NO_ERROR")
-#pragma pop_macro("REASON_UNKNOWN")
-#pragma pop_macro("SERVICE_DISABLED")
-#pragma pop_macro("SEVERITY_ERROR")
-#pragma pop_macro("STRICT")
-#pragma pop_macro("timezone")
-#endif
-
-#if defined(__clang__) || defined(__GNUC__) || defined(_MSC_VER)
-#pragma pop_macro("DEBUG")
-#pragma pop_macro("TRUE")
-#pragma pop_macro("FALSE")
-#endif // defined(__clang__) || defined(__GNUC__) || defined(_MSC_VER)
-
-#if defined(__clang__)
-#pragma clang diagnostic pop
-#elif defined(__GNUC__)
-#pragma GCC diagnostic pop
-#endif
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/repeated_field.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/repeated_field.h
deleted file mode 100644
index 26d58473..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/repeated_field.h
+++ /dev/null
@@ -1,2847 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// Author: kenton@google.com (Kenton Varda)
-//  Based on original Protocol Buffers design by
-//  Sanjay Ghemawat, Jeff Dean, and others.
-//
-// RepeatedField and RepeatedPtrField are used by generated protocol message
-// classes to manipulate repeated fields.  These classes are very similar to
-// STL's vector, but include a number of optimizations found to be useful
-// specifically in the case of Protocol Buffers.  RepeatedPtrField is
-// particularly different from STL vector as it manages ownership of the
-// pointers that it contains.
-//
-// Typically, clients should not need to access RepeatedField objects directly,
-// but should instead use the accessor functions generated automatically by the
-// protocol compiler.
-
-#ifndef GOOGLE_PROTOBUF_REPEATED_FIELD_H__
-#define GOOGLE_PROTOBUF_REPEATED_FIELD_H__
-
-#include <utility>
-#ifdef _MSC_VER
-// This is required for min/max on VS2013 only.
-#include <algorithm>
-#endif
-
-#include <iterator>
-#include <limits>
-#include <string>
-#include <type_traits>
-
-#include <google/protobuf/stubs/logging.h>
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/arena.h>
-#include <google/protobuf/message_lite.h>
-#include <google/protobuf/port.h>
-#include <google/protobuf/stubs/casts.h>
-#include <type_traits>
-
-
-// Must be included last.
-#include <google/protobuf/port_def.inc>
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-namespace google {
-namespace protobuf {
-
-class Message;
-class Reflection;
-
-template <typename T>
-struct WeakRepeatedPtrField;
-
-namespace internal {
-
-class MergePartialFromCodedStreamHelper;
-
-// kRepeatedFieldLowerClampLimit is the smallest size that will be allocated
-// when growing a repeated field.
-constexpr int kRepeatedFieldLowerClampLimit = 4;
-
-// kRepeatedFieldUpperClampLimit is the lowest signed integer value that
-// overflows when multiplied by 2 (which is undefined behavior). Sizes above
-// this will clamp to the maximum int value instead of following exponential
-// growth when growing a repeated field.
-constexpr int kRepeatedFieldUpperClampLimit =
-    (std::numeric_limits<int>::max() / 2) + 1;
-
-// A utility function for logging that doesn't need any template types.
-void LogIndexOutOfBounds(int index, int size);
-
-template <typename Iter>
-inline int CalculateReserve(Iter begin, Iter end, std::forward_iterator_tag) {
-  return static_cast<int>(std::distance(begin, end));
-}
-
-template <typename Iter>
-inline int CalculateReserve(Iter /*begin*/, Iter /*end*/,
-                            std::input_iterator_tag /*unused*/) {
-  return -1;
-}
-
-template <typename Iter>
-inline int CalculateReserve(Iter begin, Iter end) {
-  typedef typename std::iterator_traits<Iter>::iterator_category Category;
-  return CalculateReserve(begin, end, Category());
-}
-
-// Swaps two blocks of memory of size sizeof(T).
-template <typename T>
-inline void SwapBlock(char* p, char* q) {
-  T tmp;
-  memcpy(&tmp, p, sizeof(T));
-  memcpy(p, q, sizeof(T));
-  memcpy(q, &tmp, sizeof(T));
-}
-
-// Swaps two blocks of memory of size kSize:
-//  template <int kSize> void memswap(char* p, char* q);
-
-template <int kSize>
-inline typename std::enable_if<(kSize == 0), void>::type memswap(char*, char*) {
-}
-
-#define PROTO_MEMSWAP_DEF_SIZE(reg_type, max_size)                           \
-  template <int kSize>                                                       \
-  typename std::enable_if<(kSize >= sizeof(reg_type) && kSize < (max_size)), \
-                          void>::type                                        \
-  memswap(char* p, char* q) {                                                \
-    SwapBlock<reg_type>(p, q);                                               \
-    memswap<kSize - sizeof(reg_type)>(p + sizeof(reg_type),                  \
-                                      q + sizeof(reg_type));                 \
-  }
-
-PROTO_MEMSWAP_DEF_SIZE(uint8, 2)
-PROTO_MEMSWAP_DEF_SIZE(uint16, 4)
-PROTO_MEMSWAP_DEF_SIZE(uint32, 8)
-
-#ifdef __SIZEOF_INT128__
-PROTO_MEMSWAP_DEF_SIZE(uint64, 16)
-PROTO_MEMSWAP_DEF_SIZE(__uint128_t, (1u << 31))
-#else
-PROTO_MEMSWAP_DEF_SIZE(uint64, (1u << 31))
-#endif
-
-#undef PROTO_MEMSWAP_DEF_SIZE
-
-}  // namespace internal
-
-// RepeatedField is used to represent repeated fields of a primitive type (in
-// other words, everything except strings and nested Messages).  Most users will
-// not ever use a RepeatedField directly; they will use the get-by-index,
-// set-by-index, and add accessors that are generated for all repeated fields.
-template <typename Element>
-class RepeatedField final {
-  static_assert(
-      alignof(Arena) >= alignof(Element),
-      "We only support types that have an alignment smaller than Arena");
-
- public:
-  constexpr RepeatedField();
-  explicit RepeatedField(Arena* arena);
-  RepeatedField(const RepeatedField& other);
-  template <typename Iter>
-  RepeatedField(Iter begin, const Iter& end);
-  ~RepeatedField();
-
-  RepeatedField& operator=(const RepeatedField& other);
-
-  RepeatedField(RepeatedField&& other) noexcept;
-  RepeatedField& operator=(RepeatedField&& other) noexcept;
-
-  bool empty() const;
-  int size() const;
-
-  const Element& Get(int index) const;
-  Element* Mutable(int index);
-
-  const Element& operator[](int index) const { return Get(index); }
-  Element& operator[](int index) { return *Mutable(index); }
-
-  const Element& at(int index) const;
-  Element& at(int index);
-
-  void Set(int index, const Element& value);
-  void Add(const Element& value);
-  // Appends a new element and return a pointer to it.
-  // The new element is uninitialized if |Element| is a POD type.
-  Element* Add();
-  // Append elements in the range [begin, end) after reserving
-  // the appropriate number of elements.
-  template <typename Iter>
-  void Add(Iter begin, Iter end);
-
-  // Remove the last element in the array.
-  void RemoveLast();
-
-  // Extract elements with indices in "[start .. start+num-1]".
-  // Copy them into "elements[0 .. num-1]" if "elements" is not NULL.
-  // Caution: implementation also moves elements with indices [start+num ..].
-  // Calling this routine inside a loop can cause quadratic behavior.
-  void ExtractSubrange(int start, int num, Element* elements);
-
-  void Clear();
-  void MergeFrom(const RepeatedField& other);
-  void CopyFrom(const RepeatedField& other);
-
-  // Reserve space to expand the field to at least the given size.  If the
-  // array is grown, it will always be at least doubled in size.
-  void Reserve(int new_size);
-
-  // Resize the RepeatedField to a new, smaller size.  This is O(1).
-  void Truncate(int new_size);
-
-  void AddAlreadyReserved(const Element& value);
-  // Appends a new element and return a pointer to it.
-  // The new element is uninitialized if |Element| is a POD type.
-  // Should be called only if Capacity() > Size().
-  Element* AddAlreadyReserved();
-  Element* AddNAlreadyReserved(int elements);
-  int Capacity() const;
-
-  // Like STL resize.  Uses value to fill appended elements.
-  // Like Truncate() if new_size <= size(), otherwise this is
-  // O(new_size - size()).
-  void Resize(int new_size, const Element& value);
-
-  // Gets the underlying array.  This pointer is possibly invalidated by
-  // any add or remove operation.
-  Element* mutable_data();
-  const Element* data() const;
-
-  // Swap entire contents with "other". If they are separate arenas then, copies
-  // data between each other.
-  void Swap(RepeatedField* other);
-
-  // Swap entire contents with "other". Should be called only if the caller can
-  // guarantee that both repeated fields are on the same arena or are on the
-  // heap. Swapping between different arenas is disallowed and caught by a
-  // GOOGLE_DCHECK (see API docs for details).
-  void UnsafeArenaSwap(RepeatedField* other);
-
-  // Swap two elements.
-  void SwapElements(int index1, int index2);
-
-  // STL-like iterator support
-  typedef Element* iterator;
-  typedef const Element* const_iterator;
-  typedef Element value_type;
-  typedef value_type& reference;
-  typedef const value_type& const_reference;
-  typedef value_type* pointer;
-  typedef const value_type* const_pointer;
-  typedef int size_type;
-  typedef ptrdiff_t difference_type;
-
-  iterator begin();
-  const_iterator begin() const;
-  const_iterator cbegin() const;
-  iterator end();
-  const_iterator end() const;
-  const_iterator cend() const;
-
-  // Reverse iterator support
-  typedef std::reverse_iterator<const_iterator> const_reverse_iterator;
-  typedef std::reverse_iterator<iterator> reverse_iterator;
-  reverse_iterator rbegin() { return reverse_iterator(end()); }
-  const_reverse_iterator rbegin() const {
-    return const_reverse_iterator(end());
-  }
-  reverse_iterator rend() { return reverse_iterator(begin()); }
-  const_reverse_iterator rend() const {
-    return const_reverse_iterator(begin());
-  }
-
-  // Returns the number of bytes used by the repeated field, excluding
-  // sizeof(*this)
-  size_t SpaceUsedExcludingSelfLong() const;
-
-  int SpaceUsedExcludingSelf() const {
-    return internal::ToIntSize(SpaceUsedExcludingSelfLong());
-  }
-
-  // Removes the element referenced by position.
-  //
-  // Returns an iterator to the element immediately following the removed
-  // element.
-  //
-  // Invalidates all iterators at or after the removed element, including end().
-  iterator erase(const_iterator position);
-
-  // Removes the elements in the range [first, last).
-  //
-  // Returns an iterator to the element immediately following the removed range.
-  //
-  // Invalidates all iterators at or after the removed range, including end().
-  iterator erase(const_iterator first, const_iterator last);
-
-  // Get the Arena on which this RepeatedField stores its elements.
-  inline Arena* GetArena() const {
-    return (total_size_ == 0) ? static_cast<Arena*>(arena_or_elements_)
-                              : rep()->arena;
-  }
-
-  // For internal use only.
-  //
-  // This is public due to it being called by generated code.
-  inline void InternalSwap(RepeatedField* other);
-
- private:
-  static constexpr int kInitialSize = 0;
-  // A note on the representation here (see also comment below for
-  // RepeatedPtrFieldBase's struct Rep):
-  //
-  // We maintain the same sizeof(RepeatedField) as before we added arena support
-  // so that we do not degrade performance by bloating memory usage. Directly
-  // adding an arena_ element to RepeatedField is quite costly. By using
-  // indirection in this way, we keep the same size when the RepeatedField is
-  // empty (common case), and add only an 8-byte header to the elements array
-  // when non-empty. We make sure to place the size fields directly in the
-  // RepeatedField class to avoid costly cache misses due to the indirection.
-  int current_size_;
-  int total_size_;
-  struct Rep {
-    Arena* arena;
-    // Here we declare a huge array as a way of approximating C's "flexible
-    // array member" feature without relying on undefined behavior.
-    Element elements[(std::numeric_limits<int>::max() - 2 * sizeof(Arena*)) /
-                     sizeof(Element)];
-  };
-  static constexpr size_t kRepHeaderSize = offsetof(Rep, elements);
-
-  // If total_size_ == 0 this points to an Arena otherwise it points to the
-  // elements member of a Rep struct. Using this invariant allows the storage of
-  // the arena pointer without an extra allocation in the constructor.
-  void* arena_or_elements_;
-
-  // Return pointer to elements array.
-  // pre-condition: the array must have been allocated.
-  Element* elements() const {
-    GOOGLE_DCHECK_GT(total_size_, 0);
-    // Because of above pre-condition this cast is safe.
-    return unsafe_elements();
-  }
-
-  // Return pointer to elements array if it exists otherwise either null or
-  // a invalid pointer is returned. This only happens for empty repeated fields,
-  // where you can't dereference this pointer anyway (it's empty).
-  Element* unsafe_elements() const {
-    return static_cast<Element*>(arena_or_elements_);
-  }
-
-  // Return pointer to the Rep struct.
-  // pre-condition: the Rep must have been allocated, ie elements() is safe.
-  Rep* rep() const {
-    char* addr = reinterpret_cast<char*>(elements()) - offsetof(Rep, elements);
-    return reinterpret_cast<Rep*>(addr);
-  }
-
-  friend class Arena;
-  typedef void InternalArenaConstructable_;
-
-  // Move the contents of |from| into |to|, possibly clobbering |from| in the
-  // process.  For primitive types this is just a memcpy(), but it could be
-  // specialized for non-primitive types to, say, swap each element instead.
-  void MoveArray(Element* to, Element* from, int size);
-
-  // Copy the elements of |from| into |to|.
-  void CopyArray(Element* to, const Element* from, int size);
-
-  // Internal helper to delete all elements and deallocate the storage.
-  void InternalDeallocate(Rep* rep, int size) {
-    if (rep != NULL) {
-      Element* e = &rep->elements[0];
-      if (!std::is_trivial<Element>::value) {
-        Element* limit = &rep->elements[size];
-        for (; e < limit; e++) {
-          e->~Element();
-        }
-      }
-      if (rep->arena == NULL) {
-#if defined(__GXX_DELETE_WITH_SIZE__) || defined(__cpp_sized_deallocation)
-        const size_t bytes = size * sizeof(*e) + kRepHeaderSize;
-        ::operator delete(static_cast<void*>(rep), bytes);
-#else
-        ::operator delete(static_cast<void*>(rep));
-#endif
-      }
-    }
-  }
-
-  // This class is a performance wrapper around RepeatedField::Add(const T&)
-  // function. In general unless a RepeatedField is a local stack variable LLVM
-  // has a hard time optimizing Add. The machine code tends to be
-  // loop:
-  // mov %size, dword ptr [%repeated_field]       // load
-  // cmp %size, dword ptr [%repeated_field + 4]
-  // jae fallback
-  // mov %buffer, qword ptr [%repeated_field + 8]
-  // mov dword [%buffer + %size * 4], %value
-  // inc %size                                    // increment
-  // mov dword ptr [%repeated_field], %size       // store
-  // jmp loop
-  //
-  // This puts a load/store in each iteration of the important loop variable
-  // size. It's a pretty bad compile that happens even in simple cases, but
-  // largely the presence of the fallback path disturbs the compilers mem-to-reg
-  // analysis.
-  //
-  // This class takes ownership of a repeated field for the duration of it's
-  // lifetime. The repeated field should not be accessed during this time, ie.
-  // only access through this class is allowed. This class should always be a
-  // function local stack variable. Intended use
-  //
-  // void AddSequence(const int* begin, const int* end, RepeatedField<int>* out)
-  // {
-  //   RepeatedFieldAdder<int> adder(out);  // Take ownership of out
-  //   for (auto it = begin; it != end; ++it) {
-  //     adder.Add(*it);
-  //   }
-  // }
-  //
-  // Typically due to the fact adder is a local stack variable. The compiler
-  // will be successful in mem-to-reg transformation and the machine code will
-  // be loop: cmp %size, %capacity jae fallback mov dword ptr [%buffer + %size *
-  // 4], %val inc %size jmp loop
-  //
-  // The first version executes at 7 cycles per iteration while the second
-  // version near 1 or 2 cycles.
-  template <int = 0, bool = std::is_pod<Element>::value>
-  class FastAdderImpl {
-   public:
-    explicit FastAdderImpl(RepeatedField* rf) : repeated_field_(rf) {
-      index_ = repeated_field_->current_size_;
-      capacity_ = repeated_field_->total_size_;
-      buffer_ = repeated_field_->unsafe_elements();
-    }
-    ~FastAdderImpl() { repeated_field_->current_size_ = index_; }
-
-    void Add(Element val) {
-      if (index_ == capacity_) {
-        repeated_field_->current_size_ = index_;
-        repeated_field_->Reserve(index_ + 1);
-        capacity_ = repeated_field_->total_size_;
-        buffer_ = repeated_field_->unsafe_elements();
-      }
-      buffer_[index_++] = val;
-    }
-
-   private:
-    RepeatedField* repeated_field_;
-    int index_;
-    int capacity_;
-    Element* buffer_;
-
-    GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(FastAdderImpl);
-  };
-
-  // FastAdder is a wrapper for adding fields. The specialization above handles
-  // POD types more efficiently than RepeatedField.
-  template <int I>
-  class FastAdderImpl<I, false> {
-   public:
-    explicit FastAdderImpl(RepeatedField* rf) : repeated_field_(rf) {}
-    void Add(const Element& val) { repeated_field_->Add(val); }
-
-   private:
-    RepeatedField* repeated_field_;
-    GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(FastAdderImpl);
-  };
-
-  using FastAdder = FastAdderImpl<>;
-
-  friend class TestRepeatedFieldHelper;
-  friend class ::google::protobuf::internal::ParseContext;
-};
-
-namespace internal {
-template <typename It>
-class RepeatedPtrIterator;
-template <typename It, typename VoidPtr>
-class RepeatedPtrOverPtrsIterator;
-}  // namespace internal
-
-namespace internal {
-
-// This is a helper template to copy an array of elements efficiently when they
-// have a trivial copy constructor, and correctly otherwise. This really
-// shouldn't be necessary, but our compiler doesn't optimize std::copy very
-// effectively.
-template <typename Element,
-          bool HasTrivialCopy =
-              std::is_pod<Element>::value>
-struct ElementCopier {
-  void operator()(Element* to, const Element* from, int array_size);
-};
-
-}  // namespace internal
-
-namespace internal {
-
-// type-traits helper for RepeatedPtrFieldBase: we only want to invoke
-// arena-related "copy if on different arena" behavior if the necessary methods
-// exist on the contained type. In particular, we rely on MergeFrom() existing
-// as a general proxy for the fact that a copy will work, and we also provide a
-// specific override for std::string*.
-template <typename T>
-struct TypeImplementsMergeBehaviorProbeForMergeFrom {
-  typedef char HasMerge;
-  typedef long HasNoMerge;
-
-  // We accept either of:
-  // - void MergeFrom(const T& other)
-  // - bool MergeFrom(const T& other)
-  //
-  // We mangle these names a bit to avoid compatibility issues in 'unclean'
-  // include environments that may have, e.g., "#define test ..." (yes, this
-  // exists).
-  template <typename U, typename RetType, RetType (U::*)(const U& arg)>
-  struct CheckType;
-  template <typename U>
-  static HasMerge Check(CheckType<U, void, &U::MergeFrom>*);
-  template <typename U>
-  static HasMerge Check(CheckType<U, bool, &U::MergeFrom>*);
-  template <typename U>
-  static HasNoMerge Check(...);
-
-  // Resolves to either std::true_type or std::false_type.
-  typedef std::integral_constant<bool,
-                                 (sizeof(Check<T>(0)) == sizeof(HasMerge))>
-      type;
-};
-
-template <typename T, typename = void>
-struct TypeImplementsMergeBehavior
-    : TypeImplementsMergeBehaviorProbeForMergeFrom<T> {};
-
-
-template <>
-struct TypeImplementsMergeBehavior<std::string> {
-  typedef std::true_type type;
-};
-
-template <typename T>
-struct IsMovable
-    : std::integral_constant<bool, std::is_move_constructible<T>::value &&
-                                       std::is_move_assignable<T>::value> {};
-
-// This is the common base class for RepeatedPtrFields.  It deals only in void*
-// pointers.  Users should not use this interface directly.
-//
-// The methods of this interface correspond to the methods of RepeatedPtrField,
-// but may have a template argument called TypeHandler.  Its signature is:
-//   class TypeHandler {
-//    public:
-//     typedef MyType Type;
-//     static Type* New();
-//     static Type* NewFromPrototype(const Type* prototype,
-//                                       Arena* arena);
-//     static void Delete(Type*);
-//     static void Clear(Type*);
-//     static void Merge(const Type& from, Type* to);
-//
-//     // Only needs to be implemented if SpaceUsedExcludingSelf() is called.
-//     static int SpaceUsedLong(const Type&);
-//   };
-class PROTOBUF_EXPORT RepeatedPtrFieldBase {
- protected:
-  constexpr RepeatedPtrFieldBase();
-  explicit RepeatedPtrFieldBase(Arena* arena);
-  ~RepeatedPtrFieldBase() {
-#ifndef NDEBUG
-    // Try to trigger segfault / asan failure in non-opt builds. If arena_
-    // lifetime has ended before the destructor.
-    if (arena_) (void)arena_->SpaceAllocated();
-#endif
-  }
-
- public:
-  // Must be called from destructor.
-  template <typename TypeHandler>
-  void Destroy();
-
- protected:
-  bool empty() const;
-  int size() const;
-
-  template <typename TypeHandler>
-  const typename TypeHandler::Type& at(int index) const;
-  template <typename TypeHandler>
-  typename TypeHandler::Type& at(int index);
-
-  template <typename TypeHandler>
-  typename TypeHandler::Type* Mutable(int index);
-  template <typename TypeHandler>
-  void Delete(int index);
-  template <typename TypeHandler>
-  typename TypeHandler::Type* Add(typename TypeHandler::Type* prototype = NULL);
-
- public:
-  // The next few methods are public so that they can be called from generated
-  // code when implicit weak fields are used, but they should never be called by
-  // application code.
-
-  template <typename TypeHandler>
-  const typename TypeHandler::Type& Get(int index) const;
-
-  // Creates and adds an element using the given prototype, without introducing
-  // a link-time dependency on the concrete message type. This method is used to
-  // implement implicit weak fields. The prototype may be NULL, in which case an
-  // ImplicitWeakMessage will be used as a placeholder.
-  MessageLite* AddWeak(const MessageLite* prototype);
-
-  template <typename TypeHandler>
-  void Clear();
-
-  template <typename TypeHandler>
-  void MergeFrom(const RepeatedPtrFieldBase& other);
-
-  inline void InternalSwap(RepeatedPtrFieldBase* other);
-
- protected:
-  template <
-      typename TypeHandler,
-      typename std::enable_if<TypeHandler::Movable::value>::type* = nullptr>
-  void Add(typename TypeHandler::Type&& value);
-
-  template <typename TypeHandler>
-  void RemoveLast();
-  template <typename TypeHandler>
-  void CopyFrom(const RepeatedPtrFieldBase& other);
-
-  void CloseGap(int start, int num);
-
-  void Reserve(int new_size);
-
-  int Capacity() const;
-
-  // Used for constructing iterators.
-  void* const* raw_data() const;
-  void** raw_mutable_data() const;
-
-  template <typename TypeHandler>
-  typename TypeHandler::Type** mutable_data();
-  template <typename TypeHandler>
-  const typename TypeHandler::Type* const* data() const;
-
-  template <typename TypeHandler>
-  PROTOBUF_ALWAYS_INLINE void Swap(RepeatedPtrFieldBase* other);
-
-  void SwapElements(int index1, int index2);
-
-  template <typename TypeHandler>
-  size_t SpaceUsedExcludingSelfLong() const;
-
-  // Advanced memory management --------------------------------------
-
-  // Like Add(), but if there are no cleared objects to use, returns NULL.
-  template <typename TypeHandler>
-  typename TypeHandler::Type* AddFromCleared();
-
-  template <typename TypeHandler>
-  void AddAllocated(typename TypeHandler::Type* value) {
-    typename TypeImplementsMergeBehavior<typename TypeHandler::Type>::type t;
-    AddAllocatedInternal<TypeHandler>(value, t);
-  }
-
-  template <typename TypeHandler>
-  void UnsafeArenaAddAllocated(typename TypeHandler::Type* value);
-
-  template <typename TypeHandler>
-  typename TypeHandler::Type* ReleaseLast() {
-    typename TypeImplementsMergeBehavior<typename TypeHandler::Type>::type t;
-    return ReleaseLastInternal<TypeHandler>(t);
-  }
-
-  // Releases last element and returns it, but does not do out-of-arena copy.
-  // And just returns the raw pointer to the contained element in the arena.
-  template <typename TypeHandler>
-  typename TypeHandler::Type* UnsafeArenaReleaseLast();
-
-  int ClearedCount() const;
-  template <typename TypeHandler>
-  void AddCleared(typename TypeHandler::Type* value);
-  template <typename TypeHandler>
-  typename TypeHandler::Type* ReleaseCleared();
-
-  template <typename TypeHandler>
-  void AddAllocatedInternal(typename TypeHandler::Type* value, std::true_type);
-  template <typename TypeHandler>
-  void AddAllocatedInternal(typename TypeHandler::Type* value, std::false_type);
-
-  template <typename TypeHandler>
-  PROTOBUF_NOINLINE void AddAllocatedSlowWithCopy(
-      typename TypeHandler::Type* value, Arena* value_arena, Arena* my_arena);
-  template <typename TypeHandler>
-  PROTOBUF_NOINLINE void AddAllocatedSlowWithoutCopy(
-      typename TypeHandler::Type* value);
-
-  template <typename TypeHandler>
-  typename TypeHandler::Type* ReleaseLastInternal(std::true_type);
-  template <typename TypeHandler>
-  typename TypeHandler::Type* ReleaseLastInternal(std::false_type);
-
-  template <typename TypeHandler>
-  PROTOBUF_NOINLINE void SwapFallback(RepeatedPtrFieldBase* other);
-
-  inline Arena* GetArena() const { return arena_; }
-
- private:
-  static constexpr int kInitialSize = 0;
-  // A few notes on internal representation:
-  //
-  // We use an indirected approach, with struct Rep, to keep
-  // sizeof(RepeatedPtrFieldBase) equivalent to what it was before arena support
-  // was added, namely, 3 8-byte machine words on x86-64. An instance of Rep is
-  // allocated only when the repeated field is non-empty, and it is a
-  // dynamically-sized struct (the header is directly followed by elements[]).
-  // We place arena_ and current_size_ directly in the object to avoid cache
-  // misses due to the indirection, because these fields are checked frequently.
-  // Placing all fields directly in the RepeatedPtrFieldBase instance costs
-  // significant performance for memory-sensitive workloads.
-  Arena* arena_;
-  int current_size_;
-  int total_size_;
-  struct Rep {
-    int allocated_size;
-    // Here we declare a huge array as a way of approximating C's "flexible
-    // array member" feature without relying on undefined behavior.
-    void* elements[(std::numeric_limits<int>::max() - 2 * sizeof(int)) /
-                   sizeof(void*)];
-  };
-  static constexpr size_t kRepHeaderSize = offsetof(Rep, elements);
-  Rep* rep_;
-
-  template <typename TypeHandler>
-  static inline typename TypeHandler::Type* cast(void* element) {
-    return reinterpret_cast<typename TypeHandler::Type*>(element);
-  }
-  template <typename TypeHandler>
-  static inline const typename TypeHandler::Type* cast(const void* element) {
-    return reinterpret_cast<const typename TypeHandler::Type*>(element);
-  }
-
-  // Non-templated inner function to avoid code duplication. Takes a function
-  // pointer to the type-specific (templated) inner allocate/merge loop.
-  void MergeFromInternal(const RepeatedPtrFieldBase& other,
-                         void (RepeatedPtrFieldBase::*inner_loop)(void**,
-                                                                  void**, int,
-                                                                  int));
-
-  template <typename TypeHandler>
-  void MergeFromInnerLoop(void** our_elems, void** other_elems, int length,
-                          int already_allocated);
-
-  // Internal helper: extend array space if necessary to contain |extend_amount|
-  // more elements, and return a pointer to the element immediately following
-  // the old list of elements.  This interface factors out common behavior from
-  // Reserve() and MergeFrom() to reduce code size. |extend_amount| must be > 0.
-  void** InternalExtend(int extend_amount);
-
-  // The reflection implementation needs to call protected methods directly,
-  // reinterpreting pointers as being to Message instead of a specific Message
-  // subclass.
-  friend class ::PROTOBUF_NAMESPACE_ID::Reflection;
-
-  // ExtensionSet stores repeated message extensions as
-  // RepeatedPtrField<MessageLite>, but non-lite ExtensionSets need to implement
-  // SpaceUsedLong(), and thus need to call SpaceUsedExcludingSelfLong()
-  // reinterpreting MessageLite as Message.  ExtensionSet also needs to make use
-  // of AddFromCleared(), which is not part of the public interface.
-  friend class ExtensionSet;
-
-  // The MapFieldBase implementation needs to call protected methods directly,
-  // reinterpreting pointers as being to Message instead of a specific Message
-  // subclass.
-  friend class MapFieldBase;
-
-  // The table-driven MergePartialFromCodedStream implementation needs to
-  // operate on RepeatedPtrField<MessageLite>.
-  friend class MergePartialFromCodedStreamHelper;
-  friend class AccessorHelper;
-  template <typename T>
-  friend struct google::protobuf::WeakRepeatedPtrField;
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(RepeatedPtrFieldBase);
-};
-
-template <typename GenericType>
-class GenericTypeHandler {
- public:
-  typedef GenericType Type;
-  using Movable = IsMovable<GenericType>;
-
-  static inline GenericType* New(Arena* arena) {
-    return Arena::CreateMaybeMessage<Type>(arena);
-  }
-  static inline GenericType* New(Arena* arena, GenericType&& value) {
-    return Arena::Create<GenericType>(arena, std::move(value));
-  }
-  static inline GenericType* NewFromPrototype(const GenericType* prototype,
-                                              Arena* arena = NULL);
-  static inline void Delete(GenericType* value, Arena* arena) {
-    if (arena == NULL) {
-      delete value;
-    }
-  }
-  static inline Arena* GetArena(GenericType* value) {
-    return Arena::GetArena<Type>(value);
-  }
-  static inline void* GetMaybeArenaPointer(GenericType* value) {
-    return Arena::GetArena<Type>(value);
-  }
-
-  static inline void Clear(GenericType* value) { value->Clear(); }
-  PROTOBUF_NOINLINE
-  static void Merge(const GenericType& from, GenericType* to);
-  static inline size_t SpaceUsedLong(const GenericType& value) {
-    return value.SpaceUsedLong();
-  }
-};
-
-template <typename GenericType>
-GenericType* GenericTypeHandler<GenericType>::NewFromPrototype(
-    const GenericType* /* prototype */, Arena* arena) {
-  return New(arena);
-}
-template <typename GenericType>
-void GenericTypeHandler<GenericType>::Merge(const GenericType& from,
-                                            GenericType* to) {
-  to->MergeFrom(from);
-}
-
-// NewFromPrototype() and Merge() are not defined inline here, as we will need
-// to do a virtual function dispatch anyways to go from Message* to call
-// New/Merge.
-template <>
-MessageLite* GenericTypeHandler<MessageLite>::NewFromPrototype(
-    const MessageLite* prototype, Arena* arena);
-template <>
-inline Arena* GenericTypeHandler<MessageLite>::GetArena(MessageLite* value) {
-  return value->GetArena();
-}
-template <>
-inline void* GenericTypeHandler<MessageLite>::GetMaybeArenaPointer(
-    MessageLite* value) {
-  return value->GetMaybeArenaPointer();
-}
-template <>
-void GenericTypeHandler<MessageLite>::Merge(const MessageLite& from,
-                                            MessageLite* to);
-template <>
-inline void GenericTypeHandler<std::string>::Clear(std::string* value) {
-  value->clear();
-}
-template <>
-void GenericTypeHandler<std::string>::Merge(const std::string& from,
-                                            std::string* to);
-
-// Message specialization bodies defined in message.cc. This split is necessary
-// to allow proto2-lite (which includes this header) to be independent of
-// Message.
-template <>
-PROTOBUF_EXPORT Message* GenericTypeHandler<Message>::NewFromPrototype(
-    const Message* prototype, Arena* arena);
-template <>
-PROTOBUF_EXPORT Arena* GenericTypeHandler<Message>::GetArena(Message* value);
-template <>
-PROTOBUF_EXPORT void* GenericTypeHandler<Message>::GetMaybeArenaPointer(
-    Message* value);
-
-class StringTypeHandler {
- public:
-  typedef std::string Type;
-  using Movable = IsMovable<Type>;
-
-  static inline std::string* New(Arena* arena) {
-    return Arena::Create<std::string>(arena);
-  }
-  static inline std::string* New(Arena* arena, std::string&& value) {
-    return Arena::Create<std::string>(arena, std::move(value));
-  }
-  static inline std::string* NewFromPrototype(const std::string*,
-                                              Arena* arena) {
-    return New(arena);
-  }
-  static inline Arena* GetArena(std::string*) { return NULL; }
-  static inline void* GetMaybeArenaPointer(std::string* /* value */) {
-    return NULL;
-  }
-  static inline void Delete(std::string* value, Arena* arena) {
-    if (arena == NULL) {
-      delete value;
-    }
-  }
-  static inline void Clear(std::string* value) { value->clear(); }
-  static inline void Merge(const std::string& from, std::string* to) {
-    *to = from;
-  }
-  static size_t SpaceUsedLong(const std::string& value) {
-    return sizeof(value) + StringSpaceUsedExcludingSelfLong(value);
-  }
-};
-
-}  // namespace internal
-
-// RepeatedPtrField is like RepeatedField, but used for repeated strings or
-// Messages.
-template <typename Element>
-class RepeatedPtrField final : private internal::RepeatedPtrFieldBase {
- public:
-  constexpr RepeatedPtrField();
-  explicit RepeatedPtrField(Arena* arena);
-
-  RepeatedPtrField(const RepeatedPtrField& other);
-  template <typename Iter>
-  RepeatedPtrField(Iter begin, const Iter& end);
-  ~RepeatedPtrField();
-
-  RepeatedPtrField& operator=(const RepeatedPtrField& other);
-
-  RepeatedPtrField(RepeatedPtrField&& other) noexcept;
-  RepeatedPtrField& operator=(RepeatedPtrField&& other) noexcept;
-
-  bool empty() const;
-  int size() const;
-
-  const Element& Get(int index) const;
-  Element* Mutable(int index);
-  Element* Add();
-  void Add(Element&& value);
-
-  const Element& operator[](int index) const { return Get(index); }
-  Element& operator[](int index) { return *Mutable(index); }
-
-  const Element& at(int index) const;
-  Element& at(int index);
-
-  // Remove the last element in the array.
-  // Ownership of the element is retained by the array.
-  void RemoveLast();
-
-  // Delete elements with indices in the range [start .. start+num-1].
-  // Caution: implementation moves all elements with indices [start+num .. ].
-  // Calling this routine inside a loop can cause quadratic behavior.
-  void DeleteSubrange(int start, int num);
-
-  void Clear();
-  void MergeFrom(const RepeatedPtrField& other);
-  void CopyFrom(const RepeatedPtrField& other);
-
-  // Reserve space to expand the field to at least the given size.  This only
-  // resizes the pointer array; it doesn't allocate any objects.  If the
-  // array is grown, it will always be at least doubled in size.
-  void Reserve(int new_size);
-
-  int Capacity() const;
-
-  // Gets the underlying array.  This pointer is possibly invalidated by
-  // any add or remove operation.
-  Element** mutable_data();
-  const Element* const* data() const;
-
-  // Swap entire contents with "other". If they are on separate arenas, then
-  // copies data.
-  void Swap(RepeatedPtrField* other);
-
-  // Swap entire contents with "other". Caller should guarantee that either both
-  // fields are on the same arena or both are on the heap. Swapping between
-  // different arenas with this function is disallowed and is caught via
-  // GOOGLE_DCHECK.
-  void UnsafeArenaSwap(RepeatedPtrField* other);
-
-  // Swap two elements.
-  void SwapElements(int index1, int index2);
-
-  // STL-like iterator support
-  typedef internal::RepeatedPtrIterator<Element> iterator;
-  typedef internal::RepeatedPtrIterator<const Element> const_iterator;
-  typedef Element value_type;
-  typedef value_type& reference;
-  typedef const value_type& const_reference;
-  typedef value_type* pointer;
-  typedef const value_type* const_pointer;
-  typedef int size_type;
-  typedef ptrdiff_t difference_type;
-
-  iterator begin();
-  const_iterator begin() const;
-  const_iterator cbegin() const;
-  iterator end();
-  const_iterator end() const;
-  const_iterator cend() const;
-
-  // Reverse iterator support
-  typedef std::reverse_iterator<const_iterator> const_reverse_iterator;
-  typedef std::reverse_iterator<iterator> reverse_iterator;
-  reverse_iterator rbegin() { return reverse_iterator(end()); }
-  const_reverse_iterator rbegin() const {
-    return const_reverse_iterator(end());
-  }
-  reverse_iterator rend() { return reverse_iterator(begin()); }
-  const_reverse_iterator rend() const {
-    return const_reverse_iterator(begin());
-  }
-
-  // Custom STL-like iterator that iterates over and returns the underlying
-  // pointers to Element rather than Element itself.
-  typedef internal::RepeatedPtrOverPtrsIterator<Element*, void*>
-      pointer_iterator;
-  typedef internal::RepeatedPtrOverPtrsIterator<const Element* const,
-                                                const void* const>
-      const_pointer_iterator;
-  pointer_iterator pointer_begin();
-  const_pointer_iterator pointer_begin() const;
-  pointer_iterator pointer_end();
-  const_pointer_iterator pointer_end() const;
-
-  // Returns (an estimate of) the number of bytes used by the repeated field,
-  // excluding sizeof(*this).
-  size_t SpaceUsedExcludingSelfLong() const;
-
-  int SpaceUsedExcludingSelf() const {
-    return internal::ToIntSize(SpaceUsedExcludingSelfLong());
-  }
-
-  // Advanced memory management --------------------------------------
-  // When hardcore memory management becomes necessary -- as it sometimes
-  // does here at Google -- the following methods may be useful.
-
-  // Add an already-allocated object, passing ownership to the
-  // RepeatedPtrField.
-  //
-  // Note that some special behavior occurs with respect to arenas:
-  //
-  //   (i) if this field holds submessages, the new submessage will be copied if
-  //   the original is in an arena and this RepeatedPtrField is either in a
-  //   different arena, or on the heap.
-  //   (ii) if this field holds strings, the passed-in string *must* be
-  //   heap-allocated, not arena-allocated. There is no way to dynamically check
-  //   this at runtime, so User Beware.
-  void AddAllocated(Element* value);
-
-  // Remove the last element and return it, passing ownership to the caller.
-  // Requires:  size() > 0
-  //
-  // If this RepeatedPtrField is on an arena, an object copy is required to pass
-  // ownership back to the user (for compatible semantics). Use
-  // UnsafeArenaReleaseLast() if this behavior is undesired.
-  Element* ReleaseLast();
-
-  // Add an already-allocated object, skipping arena-ownership checks. The user
-  // must guarantee that the given object is in the same arena as this
-  // RepeatedPtrField.
-  // It is also useful in legacy code that uses temporary ownership to avoid
-  // copies. Example:
-  //   RepeatedPtrField<T> temp_field;
-  //   temp_field.AddAllocated(new T);
-  //   ... // Do something with temp_field
-  //   temp_field.ExtractSubrange(0, temp_field.size(), nullptr);
-  // If you put temp_field on the arena this fails, because the ownership
-  // transfers to the arena at the "AddAllocated" call and is not released
-  // anymore causing a double delete. UnsafeArenaAddAllocated prevents this.
-  void UnsafeArenaAddAllocated(Element* value);
-
-  // Remove the last element and return it.  Works only when operating on an
-  // arena. The returned pointer is to the original object in the arena, hence
-  // has the arena's lifetime.
-  // Requires:  current_size_ > 0
-  Element* UnsafeArenaReleaseLast();
-
-  // Extract elements with indices in the range "[start .. start+num-1]".
-  // The caller assumes ownership of the extracted elements and is responsible
-  // for deleting them when they are no longer needed.
-  // If "elements" is non-NULL, then pointers to the extracted elements
-  // are stored in "elements[0 .. num-1]" for the convenience of the caller.
-  // If "elements" is NULL, then the caller must use some other mechanism
-  // to perform any further operations (like deletion) on these elements.
-  // Caution: implementation also moves elements with indices [start+num ..].
-  // Calling this routine inside a loop can cause quadratic behavior.
-  //
-  // Memory copying behavior is identical to ReleaseLast(), described above: if
-  // this RepeatedPtrField is on an arena, an object copy is performed for each
-  // returned element, so that all returned element pointers are to
-  // heap-allocated copies. If this copy is not desired, the user should call
-  // UnsafeArenaExtractSubrange().
-  void ExtractSubrange(int start, int num, Element** elements);
-
-  // Identical to ExtractSubrange() described above, except that when this
-  // repeated field is on an arena, no object copies are performed. Instead, the
-  // raw object pointers are returned. Thus, if on an arena, the returned
-  // objects must not be freed, because they will not be heap-allocated objects.
-  void UnsafeArenaExtractSubrange(int start, int num, Element** elements);
-
-  // When elements are removed by calls to RemoveLast() or Clear(), they
-  // are not actually freed.  Instead, they are cleared and kept so that
-  // they can be reused later.  This can save lots of CPU time when
-  // repeatedly reusing a protocol message for similar purposes.
-  //
-  // Hardcore programs may choose to manipulate these cleared objects
-  // to better optimize memory management using the following routines.
-
-  // Get the number of cleared objects that are currently being kept
-  // around for reuse.
-  int ClearedCount() const;
-  // Add an element to the pool of cleared objects, passing ownership to
-  // the RepeatedPtrField.  The element must be cleared prior to calling
-  // this method.
-  //
-  // This method cannot be called when the repeated field is on an arena or when
-  // |value| is; both cases will trigger a GOOGLE_DCHECK-failure.
-  void AddCleared(Element* value);
-  // Remove a single element from the cleared pool and return it, passing
-  // ownership to the caller.  The element is guaranteed to be cleared.
-  // Requires:  ClearedCount() > 0
-  //
-  //
-  // This method cannot be called when the repeated field is on an arena; doing
-  // so will trigger a GOOGLE_DCHECK-failure.
-  Element* ReleaseCleared();
-
-  // Removes the element referenced by position.
-  //
-  // Returns an iterator to the element immediately following the removed
-  // element.
-  //
-  // Invalidates all iterators at or after the removed element, including end().
-  iterator erase(const_iterator position);
-
-  // Removes the elements in the range [first, last).
-  //
-  // Returns an iterator to the element immediately following the removed range.
-  //
-  // Invalidates all iterators at or after the removed range, including end().
-  iterator erase(const_iterator first, const_iterator last);
-
-  // Gets the arena on which this RepeatedPtrField stores its elements.
-  inline Arena* GetArena() const;
-
-  // For internal use only.
-  //
-  // This is public due to it being called by generated code.
-  void InternalSwap(RepeatedPtrField* other) {
-    internal::RepeatedPtrFieldBase::InternalSwap(other);
-  }
-
- private:
-  // Note:  RepeatedPtrField SHOULD NOT be subclassed by users.
-  class TypeHandler;
-
-  // Implementations for ExtractSubrange(). The copying behavior must be
-  // included only if the type supports the necessary operations (e.g.,
-  // MergeFrom()), so we must resolve this at compile time. ExtractSubrange()
-  // uses SFINAE to choose one of the below implementations.
-  void ExtractSubrangeInternal(int start, int num, Element** elements,
-                               std::true_type);
-  void ExtractSubrangeInternal(int start, int num, Element** elements,
-                               std::false_type);
-
-  friend class Arena;
-
-  template <typename T>
-  friend struct WeakRepeatedPtrField;
-
-  typedef void InternalArenaConstructable_;
-
-};
-
-// implementation ====================================================
-
-template <typename Element>
-constexpr RepeatedField<Element>::RepeatedField()
-    : current_size_(0), total_size_(0), arena_or_elements_(nullptr) {}
-
-template <typename Element>
-inline RepeatedField<Element>::RepeatedField(Arena* arena)
-    : current_size_(0), total_size_(0), arena_or_elements_(arena) {}
-
-template <typename Element>
-inline RepeatedField<Element>::RepeatedField(const RepeatedField& other)
-    : current_size_(0), total_size_(0), arena_or_elements_(nullptr) {
-  if (other.current_size_ != 0) {
-    Reserve(other.size());
-    AddNAlreadyReserved(other.size());
-    CopyArray(Mutable(0), &other.Get(0), other.size());
-  }
-}
-
-template <typename Element>
-template <typename Iter>
-RepeatedField<Element>::RepeatedField(Iter begin, const Iter& end)
-    : current_size_(0), total_size_(0), arena_or_elements_(nullptr) {
-  Add(begin, end);
-}
-
-template <typename Element>
-RepeatedField<Element>::~RepeatedField() {
-  if (total_size_ > 0) {
-    InternalDeallocate(rep(), total_size_);
-  }
-}
-
-template <typename Element>
-inline RepeatedField<Element>& RepeatedField<Element>::operator=(
-    const RepeatedField& other) {
-  if (this != &other) CopyFrom(other);
-  return *this;
-}
-
-template <typename Element>
-inline RepeatedField<Element>::RepeatedField(RepeatedField&& other) noexcept
-    : RepeatedField() {
-  // We don't just call Swap(&other) here because it would perform 3 copies if
-  // other is on an arena. This field can't be on an arena because arena
-  // construction always uses the Arena* accepting constructor.
-  if (other.GetArena()) {
-    CopyFrom(other);
-  } else {
-    InternalSwap(&other);
-  }
-}
-
-template <typename Element>
-inline RepeatedField<Element>& RepeatedField<Element>::operator=(
-    RepeatedField&& other) noexcept {
-  // We don't just call Swap(&other) here because it would perform 3 copies if
-  // the two fields are on different arenas.
-  if (this != &other) {
-    if (this->GetArena() != other.GetArena()) {
-      CopyFrom(other);
-    } else {
-      InternalSwap(&other);
-    }
-  }
-  return *this;
-}
-
-template <typename Element>
-inline bool RepeatedField<Element>::empty() const {
-  return current_size_ == 0;
-}
-
-template <typename Element>
-inline int RepeatedField<Element>::size() const {
-  return current_size_;
-}
-
-template <typename Element>
-inline int RepeatedField<Element>::Capacity() const {
-  return total_size_;
-}
-
-template <typename Element>
-inline void RepeatedField<Element>::AddAlreadyReserved(const Element& value) {
-  GOOGLE_DCHECK_LT(current_size_, total_size_);
-  elements()[current_size_++] = value;
-}
-
-template <typename Element>
-inline Element* RepeatedField<Element>::AddAlreadyReserved() {
-  GOOGLE_DCHECK_LT(current_size_, total_size_);
-  return &elements()[current_size_++];
-}
-
-template <typename Element>
-inline Element* RepeatedField<Element>::AddNAlreadyReserved(int n) {
-  GOOGLE_DCHECK_GE(total_size_ - current_size_, n)
-      << total_size_ << ", " << current_size_;
-  // Warning: sometimes people call this when n == 0 and total_size_ == 0. In
-  // this case the return pointer points to a zero size array (n == 0). Hence
-  // we can just use unsafe_elements(), because the user cannot dereference the
-  // pointer anyway.
-  Element* ret = unsafe_elements() + current_size_;
-  current_size_ += n;
-  return ret;
-}
-
-template <typename Element>
-inline void RepeatedField<Element>::Resize(int new_size, const Element& value) {
-  GOOGLE_DCHECK_GE(new_size, 0);
-  if (new_size > current_size_) {
-    Reserve(new_size);
-    std::fill(&elements()[current_size_], &elements()[new_size], value);
-  }
-  current_size_ = new_size;
-}
-
-template <typename Element>
-inline const Element& RepeatedField<Element>::Get(int index) const {
-  GOOGLE_DCHECK_GE(index, 0);
-  GOOGLE_DCHECK_LT(index, current_size_);
-  return elements()[index];
-}
-
-template <typename Element>
-inline const Element& RepeatedField<Element>::at(int index) const {
-  GOOGLE_CHECK_GE(index, 0);
-  GOOGLE_CHECK_LT(index, current_size_);
-  return elements()[index];
-}
-
-template <typename Element>
-inline Element& RepeatedField<Element>::at(int index) {
-  GOOGLE_CHECK_GE(index, 0);
-  GOOGLE_CHECK_LT(index, current_size_);
-  return elements()[index];
-}
-
-template <typename Element>
-inline Element* RepeatedField<Element>::Mutable(int index) {
-  GOOGLE_DCHECK_GE(index, 0);
-  GOOGLE_DCHECK_LT(index, current_size_);
-  return &elements()[index];
-}
-
-template <typename Element>
-inline void RepeatedField<Element>::Set(int index, const Element& value) {
-  GOOGLE_DCHECK_GE(index, 0);
-  GOOGLE_DCHECK_LT(index, current_size_);
-  elements()[index] = value;
-}
-
-template <typename Element>
-inline void RepeatedField<Element>::Add(const Element& value) {
-  uint32 size = current_size_;
-  if (static_cast<int>(size) == total_size_) {
-    // value could reference an element of the array. Reserving new space will
-    // invalidate the reference. So we must make a copy first.
-    auto tmp = value;
-    Reserve(total_size_ + 1);
-    elements()[size] = std::move(tmp);
-  } else {
-    elements()[size] = value;
-  }
-  current_size_ = size + 1;
-}
-
-template <typename Element>
-inline Element* RepeatedField<Element>::Add() {
-  uint32 size = current_size_;
-  if (static_cast<int>(size) == total_size_) Reserve(total_size_ + 1);
-  auto ptr = &elements()[size];
-  current_size_ = size + 1;
-  return ptr;
-}
-
-template <typename Element>
-template <typename Iter>
-inline void RepeatedField<Element>::Add(Iter begin, Iter end) {
-  int reserve = internal::CalculateReserve(begin, end);
-  if (reserve != -1) {
-    if (reserve == 0) {
-      return;
-    }
-
-    Reserve(reserve + size());
-    // TODO(ckennelly):  The compiler loses track of the buffer freshly
-    // allocated by Reserve() by the time we call elements, so it cannot
-    // guarantee that elements does not alias [begin(), end()).
-    //
-    // If restrict is available, annotating the pointer obtained from elements()
-    // causes this to lower to memcpy instead of memmove.
-    std::copy(begin, end, elements() + size());
-    current_size_ = reserve + size();
-  } else {
-    FastAdder fast_adder(this);
-    for (; begin != end; ++begin) fast_adder.Add(*begin);
-  }
-}
-
-template <typename Element>
-inline void RepeatedField<Element>::RemoveLast() {
-  GOOGLE_DCHECK_GT(current_size_, 0);
-  current_size_--;
-}
-
-template <typename Element>
-void RepeatedField<Element>::ExtractSubrange(int start, int num,
-                                             Element* elements) {
-  GOOGLE_DCHECK_GE(start, 0);
-  GOOGLE_DCHECK_GE(num, 0);
-  GOOGLE_DCHECK_LE(start + num, this->current_size_);
-
-  // Save the values of the removed elements if requested.
-  if (elements != NULL) {
-    for (int i = 0; i < num; ++i) elements[i] = this->Get(i + start);
-  }
-
-  // Slide remaining elements down to fill the gap.
-  if (num > 0) {
-    for (int i = start + num; i < this->current_size_; ++i)
-      this->Set(i - num, this->Get(i));
-    this->Truncate(this->current_size_ - num);
-  }
-}
-
-template <typename Element>
-inline void RepeatedField<Element>::Clear() {
-  current_size_ = 0;
-}
-
-template <typename Element>
-inline void RepeatedField<Element>::MergeFrom(const RepeatedField& other) {
-  GOOGLE_DCHECK_NE(&other, this);
-  if (other.current_size_ != 0) {
-    int existing_size = size();
-    Reserve(existing_size + other.size());
-    AddNAlreadyReserved(other.size());
-    CopyArray(Mutable(existing_size), &other.Get(0), other.size());
-  }
-}
-
-template <typename Element>
-inline void RepeatedField<Element>::CopyFrom(const RepeatedField& other) {
-  if (&other == this) return;
-  Clear();
-  MergeFrom(other);
-}
-
-template <typename Element>
-inline typename RepeatedField<Element>::iterator RepeatedField<Element>::erase(
-    const_iterator position) {
-  return erase(position, position + 1);
-}
-
-template <typename Element>
-inline typename RepeatedField<Element>::iterator RepeatedField<Element>::erase(
-    const_iterator first, const_iterator last) {
-  size_type first_offset = first - cbegin();
-  if (first != last) {
-    Truncate(std::copy(last, cend(), begin() + first_offset) - cbegin());
-  }
-  return begin() + first_offset;
-}
-
-template <typename Element>
-inline Element* RepeatedField<Element>::mutable_data() {
-  return unsafe_elements();
-}
-
-template <typename Element>
-inline const Element* RepeatedField<Element>::data() const {
-  return unsafe_elements();
-}
-
-template <typename Element>
-inline void RepeatedField<Element>::InternalSwap(RepeatedField* other) {
-  GOOGLE_DCHECK(this != other);
-  GOOGLE_DCHECK(GetArena() == other->GetArena());
-
-  // Swap all fields at once.
-  static_assert(std::is_standard_layout<RepeatedField<Element>>::value,
-                "offsetof() requires standard layout before c++17");
-  internal::memswap<offsetof(RepeatedField, arena_or_elements_) +
-                    sizeof(this->arena_or_elements_) -
-                    offsetof(RepeatedField, current_size_)>(
-      reinterpret_cast<char*>(this) + offsetof(RepeatedField, current_size_),
-      reinterpret_cast<char*>(other) + offsetof(RepeatedField, current_size_));
-}
-
-template <typename Element>
-void RepeatedField<Element>::Swap(RepeatedField* other) {
-  if (this == other) return;
-  if (GetArena() == other->GetArena()) {
-    InternalSwap(other);
-  } else {
-    RepeatedField<Element> temp(other->GetArena());
-    temp.MergeFrom(*this);
-    CopyFrom(*other);
-    other->UnsafeArenaSwap(&temp);
-  }
-}
-
-template <typename Element>
-void RepeatedField<Element>::UnsafeArenaSwap(RepeatedField* other) {
-  if (this == other) return;
-  InternalSwap(other);
-}
-
-template <typename Element>
-void RepeatedField<Element>::SwapElements(int index1, int index2) {
-  using std::swap;  // enable ADL with fallback
-  swap(elements()[index1], elements()[index2]);
-}
-
-template <typename Element>
-inline typename RepeatedField<Element>::iterator
-RepeatedField<Element>::begin() {
-  return unsafe_elements();
-}
-template <typename Element>
-inline typename RepeatedField<Element>::const_iterator
-RepeatedField<Element>::begin() const {
-  return unsafe_elements();
-}
-template <typename Element>
-inline typename RepeatedField<Element>::const_iterator
-RepeatedField<Element>::cbegin() const {
-  return unsafe_elements();
-}
-template <typename Element>
-inline typename RepeatedField<Element>::iterator RepeatedField<Element>::end() {
-  return unsafe_elements() + current_size_;
-}
-template <typename Element>
-inline typename RepeatedField<Element>::const_iterator
-RepeatedField<Element>::end() const {
-  return unsafe_elements() + current_size_;
-}
-template <typename Element>
-inline typename RepeatedField<Element>::const_iterator
-RepeatedField<Element>::cend() const {
-  return unsafe_elements() + current_size_;
-}
-
-template <typename Element>
-inline size_t RepeatedField<Element>::SpaceUsedExcludingSelfLong() const {
-  return total_size_ > 0 ? (total_size_ * sizeof(Element) + kRepHeaderSize) : 0;
-}
-
-namespace internal {
-// Returns the new size for a reserved field based on its 'total_size' and the
-// requested 'new_size'. The result is clamped to the closed interval:
-//   [internal::kMinRepeatedFieldAllocationSize,
-//    std::numeric_limits<int>::max()]
-// Requires:
-//     new_size > total_size &&
-//     (total_size == 0 ||
-//      total_size >= kRepeatedFieldLowerClampLimit)
-inline int CalculateReserveSize(int total_size, int new_size) {
-  if (new_size < kRepeatedFieldLowerClampLimit) {
-    // Clamp to smallest allowed size.
-    return kRepeatedFieldLowerClampLimit;
-  }
-  if (total_size < kRepeatedFieldUpperClampLimit) {
-    return std::max(total_size * 2, new_size);
-  } else {
-    // Clamp to largest allowed size.
-    GOOGLE_DCHECK_GT(new_size, kRepeatedFieldUpperClampLimit);
-    return std::numeric_limits<int>::max();
-  }
-}
-}  // namespace internal
-
-// Avoid inlining of Reserve(): new, copy, and delete[] lead to a significant
-// amount of code bloat.
-template <typename Element>
-void RepeatedField<Element>::Reserve(int new_size) {
-  if (total_size_ >= new_size) return;
-  Rep* old_rep = total_size_ > 0 ? rep() : NULL;
-  Rep* new_rep;
-  Arena* arena = GetArena();
-  new_size = internal::CalculateReserveSize(total_size_, new_size);
-  GOOGLE_DCHECK_LE(
-      static_cast<size_t>(new_size),
-      (std::numeric_limits<size_t>::max() - kRepHeaderSize) / sizeof(Element))
-      << "Requested size is too large to fit into size_t.";
-  size_t bytes =
-      kRepHeaderSize + sizeof(Element) * static_cast<size_t>(new_size);
-  if (arena == NULL) {
-    new_rep = static_cast<Rep*>(::operator new(bytes));
-  } else {
-    new_rep = reinterpret_cast<Rep*>(Arena::CreateArray<char>(arena, bytes));
-  }
-  new_rep->arena = arena;
-  int old_total_size = total_size_;
-  // Already known: new_size >= internal::kMinRepeatedFieldAllocationSize
-  // Maintain invariant:
-  //     total_size_ == 0 ||
-  //     total_size_ >= internal::kMinRepeatedFieldAllocationSize
-  total_size_ = new_size;
-  arena_or_elements_ = new_rep->elements;
-  // Invoke placement-new on newly allocated elements. We shouldn't have to do
-  // this, since Element is supposed to be POD, but a previous version of this
-  // code allocated storage with "new Element[size]" and some code uses
-  // RepeatedField with non-POD types, relying on constructor invocation. If
-  // Element has a trivial constructor (e.g., int32), gcc (tested with -O2)
-  // completely removes this loop because the loop body is empty, so this has no
-  // effect unless its side-effects are required for correctness.
-  // Note that we do this before MoveArray() below because Element's copy
-  // assignment implementation will want an initialized instance first.
-  Element* e = &elements()[0];
-  Element* limit = e + total_size_;
-  for (; e < limit; e++) {
-    new (e) Element;
-  }
-  if (current_size_ > 0) {
-    MoveArray(&elements()[0], old_rep->elements, current_size_);
-  }
-
-  // Likewise, we need to invoke destructors on the old array.
-  InternalDeallocate(old_rep, old_total_size);
-
-}
-
-template <typename Element>
-inline void RepeatedField<Element>::Truncate(int new_size) {
-  GOOGLE_DCHECK_LE(new_size, current_size_);
-  if (current_size_ > 0) {
-    current_size_ = new_size;
-  }
-}
-
-template <typename Element>
-inline void RepeatedField<Element>::MoveArray(Element* to, Element* from,
-                                              int array_size) {
-  CopyArray(to, from, array_size);
-}
-
-template <typename Element>
-inline void RepeatedField<Element>::CopyArray(Element* to, const Element* from,
-                                              int array_size) {
-  internal::ElementCopier<Element>()(to, from, array_size);
-}
-
-namespace internal {
-
-template <typename Element, bool HasTrivialCopy>
-void ElementCopier<Element, HasTrivialCopy>::operator()(Element* to,
-                                                        const Element* from,
-                                                        int array_size) {
-  std::copy(from, from + array_size, to);
-}
-
-template <typename Element>
-struct ElementCopier<Element, true> {
-  void operator()(Element* to, const Element* from, int array_size) {
-    memcpy(to, from, static_cast<size_t>(array_size) * sizeof(Element));
-  }
-};
-
-}  // namespace internal
-
-
-// -------------------------------------------------------------------
-
-namespace internal {
-
-constexpr RepeatedPtrFieldBase::RepeatedPtrFieldBase()
-    : arena_(NULL), current_size_(0), total_size_(0), rep_(NULL) {}
-
-inline RepeatedPtrFieldBase::RepeatedPtrFieldBase(Arena* arena)
-    : arena_(arena), current_size_(0), total_size_(0), rep_(NULL) {}
-
-template <typename TypeHandler>
-void RepeatedPtrFieldBase::Destroy() {
-  if (rep_ != NULL && arena_ == NULL) {
-    int n = rep_->allocated_size;
-    void* const* elements = rep_->elements;
-    for (int i = 0; i < n; i++) {
-      TypeHandler::Delete(cast<TypeHandler>(elements[i]), NULL);
-    }
-#if defined(__GXX_DELETE_WITH_SIZE__) || defined(__cpp_sized_deallocation)
-    const size_t size = total_size_ * sizeof(elements[0]) + kRepHeaderSize;
-    ::operator delete(static_cast<void*>(rep_), size);
-#else
-    ::operator delete(static_cast<void*>(rep_));
-#endif
-  }
-  rep_ = NULL;
-}
-
-template <typename TypeHandler>
-inline void RepeatedPtrFieldBase::Swap(RepeatedPtrFieldBase* other) {
-  if (other->GetArena() == GetArena()) {
-    InternalSwap(other);
-  } else {
-    SwapFallback<TypeHandler>(other);
-  }
-}
-
-template <typename TypeHandler>
-void RepeatedPtrFieldBase::SwapFallback(RepeatedPtrFieldBase* other) {
-  GOOGLE_DCHECK(other->GetArena() != GetArena());
-
-  // Copy semantics in this case. We try to improve efficiency by placing the
-  // temporary on |other|'s arena so that messages are copied twice rather than
-  // three times.
-  RepeatedPtrFieldBase temp(other->GetArena());
-  temp.MergeFrom<TypeHandler>(*this);
-  this->Clear<TypeHandler>();
-  this->MergeFrom<TypeHandler>(*other);
-  other->InternalSwap(&temp);
-  temp.Destroy<TypeHandler>();  // Frees rep_ if `other` had no arena.
-}
-
-inline bool RepeatedPtrFieldBase::empty() const { return current_size_ == 0; }
-
-inline int RepeatedPtrFieldBase::size() const { return current_size_; }
-
-template <typename TypeHandler>
-inline const typename TypeHandler::Type& RepeatedPtrFieldBase::Get(
-    int index) const {
-  GOOGLE_DCHECK_GE(index, 0);
-  GOOGLE_DCHECK_LT(index, current_size_);
-  return *cast<TypeHandler>(rep_->elements[index]);
-}
-
-template <typename TypeHandler>
-inline const typename TypeHandler::Type& RepeatedPtrFieldBase::at(
-    int index) const {
-  GOOGLE_CHECK_GE(index, 0);
-  GOOGLE_CHECK_LT(index, current_size_);
-  return *cast<TypeHandler>(rep_->elements[index]);
-}
-
-template <typename TypeHandler>
-inline typename TypeHandler::Type& RepeatedPtrFieldBase::at(int index) {
-  GOOGLE_CHECK_GE(index, 0);
-  GOOGLE_CHECK_LT(index, current_size_);
-  return *cast<TypeHandler>(rep_->elements[index]);
-}
-
-template <typename TypeHandler>
-inline typename TypeHandler::Type* RepeatedPtrFieldBase::Mutable(int index) {
-  GOOGLE_DCHECK_GE(index, 0);
-  GOOGLE_DCHECK_LT(index, current_size_);
-  return cast<TypeHandler>(rep_->elements[index]);
-}
-
-template <typename TypeHandler>
-inline void RepeatedPtrFieldBase::Delete(int index) {
-  GOOGLE_DCHECK_GE(index, 0);
-  GOOGLE_DCHECK_LT(index, current_size_);
-  TypeHandler::Delete(cast<TypeHandler>(rep_->elements[index]), arena_);
-}
-
-template <typename TypeHandler>
-inline typename TypeHandler::Type* RepeatedPtrFieldBase::Add(
-    typename TypeHandler::Type* prototype) {
-  if (rep_ != NULL && current_size_ < rep_->allocated_size) {
-    return cast<TypeHandler>(rep_->elements[current_size_++]);
-  }
-  if (!rep_ || rep_->allocated_size == total_size_) {
-    Reserve(total_size_ + 1);
-  }
-  ++rep_->allocated_size;
-  typename TypeHandler::Type* result =
-      TypeHandler::NewFromPrototype(prototype, arena_);
-  rep_->elements[current_size_++] = result;
-  return result;
-}
-
-template <typename TypeHandler,
-          typename std::enable_if<TypeHandler::Movable::value>::type*>
-inline void RepeatedPtrFieldBase::Add(typename TypeHandler::Type&& value) {
-  if (rep_ != NULL && current_size_ < rep_->allocated_size) {
-    *cast<TypeHandler>(rep_->elements[current_size_++]) = std::move(value);
-    return;
-  }
-  if (!rep_ || rep_->allocated_size == total_size_) {
-    Reserve(total_size_ + 1);
-  }
-  ++rep_->allocated_size;
-  typename TypeHandler::Type* result =
-      TypeHandler::New(arena_, std::move(value));
-  rep_->elements[current_size_++] = result;
-}
-
-template <typename TypeHandler>
-inline void RepeatedPtrFieldBase::RemoveLast() {
-  GOOGLE_DCHECK_GT(current_size_, 0);
-  TypeHandler::Clear(cast<TypeHandler>(rep_->elements[--current_size_]));
-}
-
-template <typename TypeHandler>
-void RepeatedPtrFieldBase::Clear() {
-  const int n = current_size_;
-  GOOGLE_DCHECK_GE(n, 0);
-  if (n > 0) {
-    void* const* elements = rep_->elements;
-    int i = 0;
-    do {
-      TypeHandler::Clear(cast<TypeHandler>(elements[i++]));
-    } while (i < n);
-    current_size_ = 0;
-  }
-}
-
-// To avoid unnecessary code duplication and reduce binary size, we use a
-// layered approach to implementing MergeFrom(). The toplevel method is
-// templated, so we get a small thunk per concrete message type in the binary.
-// This calls a shared implementation with most of the logic, passing a function
-// pointer to another type-specific piece of code that calls the object-allocate
-// and merge handlers.
-template <typename TypeHandler>
-inline void RepeatedPtrFieldBase::MergeFrom(const RepeatedPtrFieldBase& other) {
-  GOOGLE_DCHECK_NE(&other, this);
-  if (other.current_size_ == 0) return;
-  MergeFromInternal(other,
-                    &RepeatedPtrFieldBase::MergeFromInnerLoop<TypeHandler>);
-}
-
-inline void RepeatedPtrFieldBase::MergeFromInternal(
-    const RepeatedPtrFieldBase& other,
-    void (RepeatedPtrFieldBase::*inner_loop)(void**, void**, int, int)) {
-  // Note: wrapper has already guaranteed that other.rep_ != NULL here.
-  int other_size = other.current_size_;
-  void** other_elements = other.rep_->elements;
-  void** new_elements = InternalExtend(other_size);
-  int allocated_elems = rep_->allocated_size - current_size_;
-  (this->*inner_loop)(new_elements, other_elements, other_size,
-                      allocated_elems);
-  current_size_ += other_size;
-  if (rep_->allocated_size < current_size_) {
-    rep_->allocated_size = current_size_;
-  }
-}
-
-// Merges other_elems to our_elems.
-template <typename TypeHandler>
-void RepeatedPtrFieldBase::MergeFromInnerLoop(void** our_elems,
-                                              void** other_elems, int length,
-                                              int already_allocated) {
-  // Split into two loops, over ranges [0, allocated) and [allocated, length),
-  // to avoid a branch within the loop.
-  for (int i = 0; i < already_allocated && i < length; i++) {
-    // Already allocated: use existing element.
-    typename TypeHandler::Type* other_elem =
-        reinterpret_cast<typename TypeHandler::Type*>(other_elems[i]);
-    typename TypeHandler::Type* new_elem =
-        reinterpret_cast<typename TypeHandler::Type*>(our_elems[i]);
-    TypeHandler::Merge(*other_elem, new_elem);
-  }
-  Arena* arena = GetArena();
-  for (int i = already_allocated; i < length; i++) {
-    // Not allocated: alloc a new element first, then merge it.
-    typename TypeHandler::Type* other_elem =
-        reinterpret_cast<typename TypeHandler::Type*>(other_elems[i]);
-    typename TypeHandler::Type* new_elem =
-        TypeHandler::NewFromPrototype(other_elem, arena);
-    TypeHandler::Merge(*other_elem, new_elem);
-    our_elems[i] = new_elem;
-  }
-}
-
-template <typename TypeHandler>
-inline void RepeatedPtrFieldBase::CopyFrom(const RepeatedPtrFieldBase& other) {
-  if (&other == this) return;
-  RepeatedPtrFieldBase::Clear<TypeHandler>();
-  RepeatedPtrFieldBase::MergeFrom<TypeHandler>(other);
-}
-
-inline int RepeatedPtrFieldBase::Capacity() const { return total_size_; }
-
-inline void* const* RepeatedPtrFieldBase::raw_data() const {
-  return rep_ ? rep_->elements : NULL;
-}
-
-inline void** RepeatedPtrFieldBase::raw_mutable_data() const {
-  return rep_ ? const_cast<void**>(rep_->elements) : NULL;
-}
-
-template <typename TypeHandler>
-inline typename TypeHandler::Type** RepeatedPtrFieldBase::mutable_data() {
-  // TODO(kenton):  Breaks C++ aliasing rules.  We should probably remove this
-  //   method entirely.
-  return reinterpret_cast<typename TypeHandler::Type**>(raw_mutable_data());
-}
-
-template <typename TypeHandler>
-inline const typename TypeHandler::Type* const* RepeatedPtrFieldBase::data()
-    const {
-  // TODO(kenton):  Breaks C++ aliasing rules.  We should probably remove this
-  //   method entirely.
-  return reinterpret_cast<const typename TypeHandler::Type* const*>(raw_data());
-}
-
-inline void RepeatedPtrFieldBase::SwapElements(int index1, int index2) {
-  using std::swap;  // enable ADL with fallback
-  swap(rep_->elements[index1], rep_->elements[index2]);
-}
-
-template <typename TypeHandler>
-inline size_t RepeatedPtrFieldBase::SpaceUsedExcludingSelfLong() const {
-  size_t allocated_bytes = static_cast<size_t>(total_size_) * sizeof(void*);
-  if (rep_ != NULL) {
-    for (int i = 0; i < rep_->allocated_size; ++i) {
-      allocated_bytes +=
-          TypeHandler::SpaceUsedLong(*cast<TypeHandler>(rep_->elements[i]));
-    }
-    allocated_bytes += kRepHeaderSize;
-  }
-  return allocated_bytes;
-}
-
-template <typename TypeHandler>
-inline typename TypeHandler::Type* RepeatedPtrFieldBase::AddFromCleared() {
-  if (rep_ != NULL && current_size_ < rep_->allocated_size) {
-    return cast<TypeHandler>(rep_->elements[current_size_++]);
-  } else {
-    return NULL;
-  }
-}
-
-// AddAllocated version that implements arena-safe copying behavior.
-template <typename TypeHandler>
-void RepeatedPtrFieldBase::AddAllocatedInternal(
-    typename TypeHandler::Type* value, std::true_type) {
-  Arena* element_arena =
-      reinterpret_cast<Arena*>(TypeHandler::GetMaybeArenaPointer(value));
-  Arena* arena = GetArena();
-  if (arena == element_arena && rep_ && rep_->allocated_size < total_size_) {
-    // Fast path: underlying arena representation (tagged pointer) is equal to
-    // our arena pointer, and we can add to array without resizing it (at least
-    // one slot that is not allocated).
-    void** elems = rep_->elements;
-    if (current_size_ < rep_->allocated_size) {
-      // Make space at [current] by moving first allocated element to end of
-      // allocated list.
-      elems[rep_->allocated_size] = elems[current_size_];
-    }
-    elems[current_size_] = value;
-    current_size_ = current_size_ + 1;
-    rep_->allocated_size = rep_->allocated_size + 1;
-  } else {
-    AddAllocatedSlowWithCopy<TypeHandler>(value, TypeHandler::GetArena(value),
-                                          arena);
-  }
-}
-
-// Slowpath handles all cases, copying if necessary.
-template <typename TypeHandler>
-void RepeatedPtrFieldBase::AddAllocatedSlowWithCopy(
-    // Pass value_arena and my_arena to avoid duplicate virtual call (value) or
-    // load (mine).
-    typename TypeHandler::Type* value, Arena* value_arena, Arena* my_arena) {
-  // Ensure that either the value is in the same arena, or if not, we do the
-  // appropriate thing: Own() it (if it's on heap and we're in an arena) or copy
-  // it to our arena/heap (otherwise).
-  if (my_arena != NULL && value_arena == NULL) {
-    my_arena->Own(value);
-  } else if (my_arena != value_arena) {
-    typename TypeHandler::Type* new_value =
-        TypeHandler::NewFromPrototype(value, my_arena);
-    TypeHandler::Merge(*value, new_value);
-    TypeHandler::Delete(value, value_arena);
-    value = new_value;
-  }
-
-  UnsafeArenaAddAllocated<TypeHandler>(value);
-}
-
-// AddAllocated version that does not implement arena-safe copying behavior.
-template <typename TypeHandler>
-void RepeatedPtrFieldBase::AddAllocatedInternal(
-    typename TypeHandler::Type* value, std::false_type) {
-  if (rep_ && rep_->allocated_size < total_size_) {
-    // Fast path: underlying arena representation (tagged pointer) is equal to
-    // our arena pointer, and we can add to array without resizing it (at least
-    // one slot that is not allocated).
-    void** elems = rep_->elements;
-    if (current_size_ < rep_->allocated_size) {
-      // Make space at [current] by moving first allocated element to end of
-      // allocated list.
-      elems[rep_->allocated_size] = elems[current_size_];
-    }
-    elems[current_size_] = value;
-    current_size_ = current_size_ + 1;
-    ++rep_->allocated_size;
-  } else {
-    UnsafeArenaAddAllocated<TypeHandler>(value);
-  }
-}
-
-template <typename TypeHandler>
-void RepeatedPtrFieldBase::UnsafeArenaAddAllocated(
-    typename TypeHandler::Type* value) {
-  // Make room for the new pointer.
-  if (!rep_ || current_size_ == total_size_) {
-    // The array is completely full with no cleared objects, so grow it.
-    Reserve(total_size_ + 1);
-    ++rep_->allocated_size;
-  } else if (rep_->allocated_size == total_size_) {
-    // There is no more space in the pointer array because it contains some
-    // cleared objects awaiting reuse.  We don't want to grow the array in this
-    // case because otherwise a loop calling AddAllocated() followed by Clear()
-    // would leak memory.
-    TypeHandler::Delete(cast<TypeHandler>(rep_->elements[current_size_]),
-                        arena_);
-  } else if (current_size_ < rep_->allocated_size) {
-    // We have some cleared objects.  We don't care about their order, so we
-    // can just move the first one to the end to make space.
-    rep_->elements[rep_->allocated_size] = rep_->elements[current_size_];
-    ++rep_->allocated_size;
-  } else {
-    // There are no cleared objects.
-    ++rep_->allocated_size;
-  }
-
-  rep_->elements[current_size_++] = value;
-}
-
-// ReleaseLast() for types that implement merge/copy behavior.
-template <typename TypeHandler>
-inline typename TypeHandler::Type* RepeatedPtrFieldBase::ReleaseLastInternal(
-    std::true_type) {
-  // First, release an element.
-  typename TypeHandler::Type* result = UnsafeArenaReleaseLast<TypeHandler>();
-  // Now perform a copy if we're on an arena.
-  Arena* arena = GetArena();
-  if (arena == NULL) {
-    return result;
-  } else {
-    typename TypeHandler::Type* new_result =
-        TypeHandler::NewFromPrototype(result, NULL);
-    TypeHandler::Merge(*result, new_result);
-    return new_result;
-  }
-}
-
-// ReleaseLast() for types that *do not* implement merge/copy behavior -- this
-// is the same as UnsafeArenaReleaseLast(). Note that we GOOGLE_DCHECK-fail if we're on
-// an arena, since the user really should implement the copy operation in this
-// case.
-template <typename TypeHandler>
-inline typename TypeHandler::Type* RepeatedPtrFieldBase::ReleaseLastInternal(
-    std::false_type) {
-  GOOGLE_DCHECK(GetArena() == NULL)
-      << "ReleaseLast() called on a RepeatedPtrField that is on an arena, "
-      << "with a type that does not implement MergeFrom. This is unsafe; "
-      << "please implement MergeFrom for your type.";
-  return UnsafeArenaReleaseLast<TypeHandler>();
-}
-
-template <typename TypeHandler>
-inline typename TypeHandler::Type*
-RepeatedPtrFieldBase::UnsafeArenaReleaseLast() {
-  GOOGLE_DCHECK_GT(current_size_, 0);
-  typename TypeHandler::Type* result =
-      cast<TypeHandler>(rep_->elements[--current_size_]);
-  --rep_->allocated_size;
-  if (current_size_ < rep_->allocated_size) {
-    // There are cleared elements on the end; replace the removed element
-    // with the last allocated element.
-    rep_->elements[current_size_] = rep_->elements[rep_->allocated_size];
-  }
-  return result;
-}
-
-inline int RepeatedPtrFieldBase::ClearedCount() const {
-  return rep_ ? (rep_->allocated_size - current_size_) : 0;
-}
-
-template <typename TypeHandler>
-inline void RepeatedPtrFieldBase::AddCleared(
-    typename TypeHandler::Type* value) {
-  GOOGLE_DCHECK(GetArena() == NULL)
-      << "AddCleared() can only be used on a RepeatedPtrField not on an arena.";
-  GOOGLE_DCHECK(TypeHandler::GetArena(value) == NULL)
-      << "AddCleared() can only accept values not on an arena.";
-  if (!rep_ || rep_->allocated_size == total_size_) {
-    Reserve(total_size_ + 1);
-  }
-  rep_->elements[rep_->allocated_size++] = value;
-}
-
-template <typename TypeHandler>
-inline typename TypeHandler::Type* RepeatedPtrFieldBase::ReleaseCleared() {
-  GOOGLE_DCHECK(GetArena() == NULL)
-      << "ReleaseCleared() can only be used on a RepeatedPtrField not on "
-      << "an arena.";
-  GOOGLE_DCHECK(GetArena() == NULL);
-  GOOGLE_DCHECK(rep_ != NULL);
-  GOOGLE_DCHECK_GT(rep_->allocated_size, current_size_);
-  return cast<TypeHandler>(rep_->elements[--rep_->allocated_size]);
-}
-
-}  // namespace internal
-
-// -------------------------------------------------------------------
-
-template <typename Element>
-class RepeatedPtrField<Element>::TypeHandler
-    : public internal::GenericTypeHandler<Element> {};
-
-template <>
-class RepeatedPtrField<std::string>::TypeHandler
-    : public internal::StringTypeHandler {};
-
-template <typename Element>
-constexpr RepeatedPtrField<Element>::RepeatedPtrField()
-    : RepeatedPtrFieldBase() {}
-
-template <typename Element>
-inline RepeatedPtrField<Element>::RepeatedPtrField(Arena* arena)
-    : RepeatedPtrFieldBase(arena) {}
-
-template <typename Element>
-inline RepeatedPtrField<Element>::RepeatedPtrField(
-    const RepeatedPtrField& other)
-    : RepeatedPtrFieldBase() {
-  MergeFrom(other);
-}
-
-template <typename Element>
-template <typename Iter>
-inline RepeatedPtrField<Element>::RepeatedPtrField(Iter begin,
-                                                   const Iter& end) {
-  int reserve = internal::CalculateReserve(begin, end);
-  if (reserve != -1) {
-    Reserve(reserve);
-  }
-  for (; begin != end; ++begin) {
-    *Add() = *begin;
-  }
-}
-
-template <typename Element>
-RepeatedPtrField<Element>::~RepeatedPtrField() {
-  Destroy<TypeHandler>();
-}
-
-template <typename Element>
-inline RepeatedPtrField<Element>& RepeatedPtrField<Element>::operator=(
-    const RepeatedPtrField& other) {
-  if (this != &other) CopyFrom(other);
-  return *this;
-}
-
-template <typename Element>
-inline RepeatedPtrField<Element>::RepeatedPtrField(
-    RepeatedPtrField&& other) noexcept
-    : RepeatedPtrField() {
-  // We don't just call Swap(&other) here because it would perform 3 copies if
-  // other is on an arena. This field can't be on an arena because arena
-  // construction always uses the Arena* accepting constructor.
-  if (other.GetArena()) {
-    CopyFrom(other);
-  } else {
-    InternalSwap(&other);
-  }
-}
-
-template <typename Element>
-inline RepeatedPtrField<Element>& RepeatedPtrField<Element>::operator=(
-    RepeatedPtrField&& other) noexcept {
-  // We don't just call Swap(&other) here because it would perform 3 copies if
-  // the two fields are on different arenas.
-  if (this != &other) {
-    if (this->GetArena() != other.GetArena()) {
-      CopyFrom(other);
-    } else {
-      InternalSwap(&other);
-    }
-  }
-  return *this;
-}
-
-template <typename Element>
-inline bool RepeatedPtrField<Element>::empty() const {
-  return RepeatedPtrFieldBase::empty();
-}
-
-template <typename Element>
-inline int RepeatedPtrField<Element>::size() const {
-  return RepeatedPtrFieldBase::size();
-}
-
-template <typename Element>
-inline const Element& RepeatedPtrField<Element>::Get(int index) const {
-  return RepeatedPtrFieldBase::Get<TypeHandler>(index);
-}
-
-template <typename Element>
-inline const Element& RepeatedPtrField<Element>::at(int index) const {
-  return RepeatedPtrFieldBase::at<TypeHandler>(index);
-}
-
-template <typename Element>
-inline Element& RepeatedPtrField<Element>::at(int index) {
-  return RepeatedPtrFieldBase::at<TypeHandler>(index);
-}
-
-
-template <typename Element>
-inline Element* RepeatedPtrField<Element>::Mutable(int index) {
-  return RepeatedPtrFieldBase::Mutable<TypeHandler>(index);
-}
-
-template <typename Element>
-inline Element* RepeatedPtrField<Element>::Add() {
-  return RepeatedPtrFieldBase::Add<TypeHandler>();
-}
-
-template <typename Element>
-inline void RepeatedPtrField<Element>::Add(Element&& value) {
-  RepeatedPtrFieldBase::Add<TypeHandler>(std::move(value));
-}
-
-template <typename Element>
-inline void RepeatedPtrField<Element>::RemoveLast() {
-  RepeatedPtrFieldBase::RemoveLast<TypeHandler>();
-}
-
-template <typename Element>
-inline void RepeatedPtrField<Element>::DeleteSubrange(int start, int num) {
-  GOOGLE_DCHECK_GE(start, 0);
-  GOOGLE_DCHECK_GE(num, 0);
-  GOOGLE_DCHECK_LE(start + num, size());
-  for (int i = 0; i < num; ++i) {
-    RepeatedPtrFieldBase::Delete<TypeHandler>(start + i);
-  }
-  ExtractSubrange(start, num, NULL);
-}
-
-template <typename Element>
-inline void RepeatedPtrField<Element>::ExtractSubrange(int start, int num,
-                                                       Element** elements) {
-  typename internal::TypeImplementsMergeBehavior<
-      typename TypeHandler::Type>::type t;
-  ExtractSubrangeInternal(start, num, elements, t);
-}
-
-// ExtractSubrange() implementation for types that implement merge/copy
-// behavior.
-template <typename Element>
-inline void RepeatedPtrField<Element>::ExtractSubrangeInternal(
-    int start, int num, Element** elements, std::true_type) {
-  GOOGLE_DCHECK_GE(start, 0);
-  GOOGLE_DCHECK_GE(num, 0);
-  GOOGLE_DCHECK_LE(start + num, size());
-
-  if (num > 0) {
-    // Save the values of the removed elements if requested.
-    if (elements != NULL) {
-      if (GetArena() != NULL) {
-        // If we're on an arena, we perform a copy for each element so that the
-        // returned elements are heap-allocated.
-        for (int i = 0; i < num; ++i) {
-          Element* element =
-              RepeatedPtrFieldBase::Mutable<TypeHandler>(i + start);
-          typename TypeHandler::Type* new_value =
-              TypeHandler::NewFromPrototype(element, NULL);
-          TypeHandler::Merge(*element, new_value);
-          elements[i] = new_value;
-        }
-      } else {
-        for (int i = 0; i < num; ++i) {
-          elements[i] = RepeatedPtrFieldBase::Mutable<TypeHandler>(i + start);
-        }
-      }
-    }
-    CloseGap(start, num);
-  }
-}
-
-// ExtractSubrange() implementation for types that do not implement merge/copy
-// behavior.
-template <typename Element>
-inline void RepeatedPtrField<Element>::ExtractSubrangeInternal(
-    int start, int num, Element** elements, std::false_type) {
-  // This case is identical to UnsafeArenaExtractSubrange(). However, since
-  // ExtractSubrange() must return heap-allocated objects by contract, and we
-  // cannot fulfill this contract if we are an on arena, we must GOOGLE_DCHECK() that
-  // we are not on an arena.
-  GOOGLE_DCHECK(GetArena() == NULL)
-      << "ExtractSubrange() when arena is non-NULL is only supported when "
-      << "the Element type supplies a MergeFrom() operation to make copies.";
-  UnsafeArenaExtractSubrange(start, num, elements);
-}
-
-template <typename Element>
-inline void RepeatedPtrField<Element>::UnsafeArenaExtractSubrange(
-    int start, int num, Element** elements) {
-  GOOGLE_DCHECK_GE(start, 0);
-  GOOGLE_DCHECK_GE(num, 0);
-  GOOGLE_DCHECK_LE(start + num, size());
-
-  if (num > 0) {
-    // Save the values of the removed elements if requested.
-    if (elements != NULL) {
-      for (int i = 0; i < num; ++i) {
-        elements[i] = RepeatedPtrFieldBase::Mutable<TypeHandler>(i + start);
-      }
-    }
-    CloseGap(start, num);
-  }
-}
-
-template <typename Element>
-inline void RepeatedPtrField<Element>::Clear() {
-  RepeatedPtrFieldBase::Clear<TypeHandler>();
-}
-
-template <typename Element>
-inline void RepeatedPtrField<Element>::MergeFrom(
-    const RepeatedPtrField& other) {
-  RepeatedPtrFieldBase::MergeFrom<TypeHandler>(other);
-}
-
-template <typename Element>
-inline void RepeatedPtrField<Element>::CopyFrom(const RepeatedPtrField& other) {
-  RepeatedPtrFieldBase::CopyFrom<TypeHandler>(other);
-}
-
-template <typename Element>
-inline typename RepeatedPtrField<Element>::iterator
-RepeatedPtrField<Element>::erase(const_iterator position) {
-  return erase(position, position + 1);
-}
-
-template <typename Element>
-inline typename RepeatedPtrField<Element>::iterator
-RepeatedPtrField<Element>::erase(const_iterator first, const_iterator last) {
-  size_type pos_offset = std::distance(cbegin(), first);
-  size_type last_offset = std::distance(cbegin(), last);
-  DeleteSubrange(pos_offset, last_offset - pos_offset);
-  return begin() + pos_offset;
-}
-
-template <typename Element>
-inline Element** RepeatedPtrField<Element>::mutable_data() {
-  return RepeatedPtrFieldBase::mutable_data<TypeHandler>();
-}
-
-template <typename Element>
-inline const Element* const* RepeatedPtrField<Element>::data() const {
-  return RepeatedPtrFieldBase::data<TypeHandler>();
-}
-
-template <typename Element>
-inline void RepeatedPtrField<Element>::Swap(RepeatedPtrField* other) {
-  if (this == other) return;
-  RepeatedPtrFieldBase::Swap<TypeHandler>(other);
-}
-
-template <typename Element>
-inline void RepeatedPtrField<Element>::UnsafeArenaSwap(
-    RepeatedPtrField* other) {
-  if (this == other) return;
-  RepeatedPtrFieldBase::InternalSwap(other);
-}
-
-template <typename Element>
-inline void RepeatedPtrField<Element>::SwapElements(int index1, int index2) {
-  RepeatedPtrFieldBase::SwapElements(index1, index2);
-}
-
-template <typename Element>
-inline Arena* RepeatedPtrField<Element>::GetArena() const {
-  return RepeatedPtrFieldBase::GetArena();
-}
-
-template <typename Element>
-inline size_t RepeatedPtrField<Element>::SpaceUsedExcludingSelfLong() const {
-  return RepeatedPtrFieldBase::SpaceUsedExcludingSelfLong<TypeHandler>();
-}
-
-template <typename Element>
-inline void RepeatedPtrField<Element>::AddAllocated(Element* value) {
-  RepeatedPtrFieldBase::AddAllocated<TypeHandler>(value);
-}
-
-template <typename Element>
-inline void RepeatedPtrField<Element>::UnsafeArenaAddAllocated(Element* value) {
-  RepeatedPtrFieldBase::UnsafeArenaAddAllocated<TypeHandler>(value);
-}
-
-template <typename Element>
-inline Element* RepeatedPtrField<Element>::ReleaseLast() {
-  return RepeatedPtrFieldBase::ReleaseLast<TypeHandler>();
-}
-
-template <typename Element>
-inline Element* RepeatedPtrField<Element>::UnsafeArenaReleaseLast() {
-  return RepeatedPtrFieldBase::UnsafeArenaReleaseLast<TypeHandler>();
-}
-
-template <typename Element>
-inline int RepeatedPtrField<Element>::ClearedCount() const {
-  return RepeatedPtrFieldBase::ClearedCount();
-}
-
-template <typename Element>
-inline void RepeatedPtrField<Element>::AddCleared(Element* value) {
-  return RepeatedPtrFieldBase::AddCleared<TypeHandler>(value);
-}
-
-template <typename Element>
-inline Element* RepeatedPtrField<Element>::ReleaseCleared() {
-  return RepeatedPtrFieldBase::ReleaseCleared<TypeHandler>();
-}
-
-template <typename Element>
-inline void RepeatedPtrField<Element>::Reserve(int new_size) {
-  return RepeatedPtrFieldBase::Reserve(new_size);
-}
-
-template <typename Element>
-inline int RepeatedPtrField<Element>::Capacity() const {
-  return RepeatedPtrFieldBase::Capacity();
-}
-
-// -------------------------------------------------------------------
-
-namespace internal {
-
-// STL-like iterator implementation for RepeatedPtrField.  You should not
-// refer to this class directly; use RepeatedPtrField<T>::iterator instead.
-//
-// The iterator for RepeatedPtrField<T>, RepeatedPtrIterator<T>, is
-// very similar to iterator_ptr<T**> in util/gtl/iterator_adaptors.h,
-// but adds random-access operators and is modified to wrap a void** base
-// iterator (since RepeatedPtrField stores its array as a void* array and
-// casting void** to T** would violate C++ aliasing rules).
-//
-// This code based on net/proto/proto-array-internal.h by Jeffrey Yasskin
-// (jyasskin@google.com).
-template <typename Element>
-class RepeatedPtrIterator {
- public:
-  using iterator = RepeatedPtrIterator<Element>;
-  using iterator_category = std::random_access_iterator_tag;
-  using value_type = typename std::remove_const<Element>::type;
-  using difference_type = std::ptrdiff_t;
-  using pointer = Element*;
-  using reference = Element&;
-
-  RepeatedPtrIterator() : it_(NULL) {}
-  explicit RepeatedPtrIterator(void* const* it) : it_(it) {}
-
-  // Allow "upcasting" from RepeatedPtrIterator<T**> to
-  // RepeatedPtrIterator<const T*const*>.
-  template <typename OtherElement>
-  RepeatedPtrIterator(const RepeatedPtrIterator<OtherElement>& other)
-      : it_(other.it_) {
-    // Force a compiler error if the other type is not convertible to ours.
-    if (false) {
-      implicit_cast<Element*>(static_cast<OtherElement*>(nullptr));
-    }
-  }
-
-  // dereferenceable
-  reference operator*() const { return *reinterpret_cast<Element*>(*it_); }
-  pointer operator->() const { return &(operator*()); }
-
-  // {inc,dec}rementable
-  iterator& operator++() {
-    ++it_;
-    return *this;
-  }
-  iterator operator++(int) { return iterator(it_++); }
-  iterator& operator--() {
-    --it_;
-    return *this;
-  }
-  iterator operator--(int) { return iterator(it_--); }
-
-  // equality_comparable
-  bool operator==(const iterator& x) const { return it_ == x.it_; }
-  bool operator!=(const iterator& x) const { return it_ != x.it_; }
-
-  // less_than_comparable
-  bool operator<(const iterator& x) const { return it_ < x.it_; }
-  bool operator<=(const iterator& x) const { return it_ <= x.it_; }
-  bool operator>(const iterator& x) const { return it_ > x.it_; }
-  bool operator>=(const iterator& x) const { return it_ >= x.it_; }
-
-  // addable, subtractable
-  iterator& operator+=(difference_type d) {
-    it_ += d;
-    return *this;
-  }
-  friend iterator operator+(iterator it, const difference_type d) {
-    it += d;
-    return it;
-  }
-  friend iterator operator+(const difference_type d, iterator it) {
-    it += d;
-    return it;
-  }
-  iterator& operator-=(difference_type d) {
-    it_ -= d;
-    return *this;
-  }
-  friend iterator operator-(iterator it, difference_type d) {
-    it -= d;
-    return it;
-  }
-
-  // indexable
-  reference operator[](difference_type d) const { return *(*this + d); }
-
-  // random access iterator
-  difference_type operator-(const iterator& x) const { return it_ - x.it_; }
-
- private:
-  template <typename OtherElement>
-  friend class RepeatedPtrIterator;
-
-  // The internal iterator.
-  void* const* it_;
-};
-
-// Provide an iterator that operates on pointers to the underlying objects
-// rather than the objects themselves as RepeatedPtrIterator does.
-// Consider using this when working with stl algorithms that change
-// the array.
-// The VoidPtr template parameter holds the type-agnostic pointer value
-// referenced by the iterator.  It should either be "void *" for a mutable
-// iterator, or "const void* const" for a constant iterator.
-template <typename Element, typename VoidPtr>
-class RepeatedPtrOverPtrsIterator {
- public:
-  using iterator = RepeatedPtrOverPtrsIterator<Element, VoidPtr>;
-  using iterator_category = std::random_access_iterator_tag;
-  using value_type = typename std::remove_const<Element>::type;
-  using difference_type = std::ptrdiff_t;
-  using pointer = Element*;
-  using reference = Element&;
-
-  RepeatedPtrOverPtrsIterator() : it_(NULL) {}
-  explicit RepeatedPtrOverPtrsIterator(VoidPtr* it) : it_(it) {}
-
-  // dereferenceable
-  reference operator*() const { return *reinterpret_cast<Element*>(it_); }
-  pointer operator->() const { return &(operator*()); }
-
-  // {inc,dec}rementable
-  iterator& operator++() {
-    ++it_;
-    return *this;
-  }
-  iterator operator++(int) { return iterator(it_++); }
-  iterator& operator--() {
-    --it_;
-    return *this;
-  }
-  iterator operator--(int) { return iterator(it_--); }
-
-  // equality_comparable
-  bool operator==(const iterator& x) const { return it_ == x.it_; }
-  bool operator!=(const iterator& x) const { return it_ != x.it_; }
-
-  // less_than_comparable
-  bool operator<(const iterator& x) const { return it_ < x.it_; }
-  bool operator<=(const iterator& x) const { return it_ <= x.it_; }
-  bool operator>(const iterator& x) const { return it_ > x.it_; }
-  bool operator>=(const iterator& x) const { return it_ >= x.it_; }
-
-  // addable, subtractable
-  iterator& operator+=(difference_type d) {
-    it_ += d;
-    return *this;
-  }
-  friend iterator operator+(iterator it, difference_type d) {
-    it += d;
-    return it;
-  }
-  friend iterator operator+(difference_type d, iterator it) {
-    it += d;
-    return it;
-  }
-  iterator& operator-=(difference_type d) {
-    it_ -= d;
-    return *this;
-  }
-  friend iterator operator-(iterator it, difference_type d) {
-    it -= d;
-    return it;
-  }
-
-  // indexable
-  reference operator[](difference_type d) const { return *(*this + d); }
-
-  // random access iterator
-  difference_type operator-(const iterator& x) const { return it_ - x.it_; }
-
- private:
-  template <typename OtherElement>
-  friend class RepeatedPtrIterator;
-
-  // The internal iterator.
-  VoidPtr* it_;
-};
-
-void RepeatedPtrFieldBase::InternalSwap(RepeatedPtrFieldBase* other) {
-  GOOGLE_DCHECK(this != other);
-  GOOGLE_DCHECK(GetArena() == other->GetArena());
-
-  // Swap all fields at once.
-  static_assert(std::is_standard_layout<RepeatedPtrFieldBase>::value,
-                "offsetof() requires standard layout before c++17");
-  internal::memswap<offsetof(RepeatedPtrFieldBase, rep_) + sizeof(this->rep_) -
-                    offsetof(RepeatedPtrFieldBase, current_size_)>(
-      reinterpret_cast<char*>(this) +
-          offsetof(RepeatedPtrFieldBase, current_size_),
-      reinterpret_cast<char*>(other) +
-          offsetof(RepeatedPtrFieldBase, current_size_));
-}
-
-}  // namespace internal
-
-template <typename Element>
-inline typename RepeatedPtrField<Element>::iterator
-RepeatedPtrField<Element>::begin() {
-  return iterator(raw_data());
-}
-template <typename Element>
-inline typename RepeatedPtrField<Element>::const_iterator
-RepeatedPtrField<Element>::begin() const {
-  return iterator(raw_data());
-}
-template <typename Element>
-inline typename RepeatedPtrField<Element>::const_iterator
-RepeatedPtrField<Element>::cbegin() const {
-  return begin();
-}
-template <typename Element>
-inline typename RepeatedPtrField<Element>::iterator
-RepeatedPtrField<Element>::end() {
-  return iterator(raw_data() + size());
-}
-template <typename Element>
-inline typename RepeatedPtrField<Element>::const_iterator
-RepeatedPtrField<Element>::end() const {
-  return iterator(raw_data() + size());
-}
-template <typename Element>
-inline typename RepeatedPtrField<Element>::const_iterator
-RepeatedPtrField<Element>::cend() const {
-  return end();
-}
-
-template <typename Element>
-inline typename RepeatedPtrField<Element>::pointer_iterator
-RepeatedPtrField<Element>::pointer_begin() {
-  return pointer_iterator(raw_mutable_data());
-}
-template <typename Element>
-inline typename RepeatedPtrField<Element>::const_pointer_iterator
-RepeatedPtrField<Element>::pointer_begin() const {
-  return const_pointer_iterator(const_cast<const void* const*>(raw_data()));
-}
-template <typename Element>
-inline typename RepeatedPtrField<Element>::pointer_iterator
-RepeatedPtrField<Element>::pointer_end() {
-  return pointer_iterator(raw_mutable_data() + size());
-}
-template <typename Element>
-inline typename RepeatedPtrField<Element>::const_pointer_iterator
-RepeatedPtrField<Element>::pointer_end() const {
-  return const_pointer_iterator(
-      const_cast<const void* const*>(raw_data() + size()));
-}
-
-// Iterators and helper functions that follow the spirit of the STL
-// std::back_insert_iterator and std::back_inserter but are tailor-made
-// for RepeatedField and RepeatedPtrField. Typical usage would be:
-//
-//   std::copy(some_sequence.begin(), some_sequence.end(),
-//             RepeatedFieldBackInserter(proto.mutable_sequence()));
-//
-// Ported by johannes from util/gtl/proto-array-iterators.h
-
-namespace internal {
-// A back inserter for RepeatedField objects.
-template <typename T>
-class RepeatedFieldBackInsertIterator
-    : public std::iterator<std::output_iterator_tag, T> {
- public:
-  explicit RepeatedFieldBackInsertIterator(
-      RepeatedField<T>* const mutable_field)
-      : field_(mutable_field) {}
-  RepeatedFieldBackInsertIterator<T>& operator=(const T& value) {
-    field_->Add(value);
-    return *this;
-  }
-  RepeatedFieldBackInsertIterator<T>& operator*() { return *this; }
-  RepeatedFieldBackInsertIterator<T>& operator++() { return *this; }
-  RepeatedFieldBackInsertIterator<T>& operator++(int /* unused */) {
-    return *this;
-  }
-
- private:
-  RepeatedField<T>* field_;
-};
-
-// A back inserter for RepeatedPtrField objects.
-template <typename T>
-class RepeatedPtrFieldBackInsertIterator
-    : public std::iterator<std::output_iterator_tag, T> {
- public:
-  RepeatedPtrFieldBackInsertIterator(RepeatedPtrField<T>* const mutable_field)
-      : field_(mutable_field) {}
-  RepeatedPtrFieldBackInsertIterator<T>& operator=(const T& value) {
-    *field_->Add() = value;
-    return *this;
-  }
-  RepeatedPtrFieldBackInsertIterator<T>& operator=(
-      const T* const ptr_to_value) {
-    *field_->Add() = *ptr_to_value;
-    return *this;
-  }
-  RepeatedPtrFieldBackInsertIterator<T>& operator=(T&& value) {
-    *field_->Add() = std::move(value);
-    return *this;
-  }
-  RepeatedPtrFieldBackInsertIterator<T>& operator*() { return *this; }
-  RepeatedPtrFieldBackInsertIterator<T>& operator++() { return *this; }
-  RepeatedPtrFieldBackInsertIterator<T>& operator++(int /* unused */) {
-    return *this;
-  }
-
- private:
-  RepeatedPtrField<T>* field_;
-};
-
-// A back inserter for RepeatedPtrFields that inserts by transferring ownership
-// of a pointer.
-template <typename T>
-class AllocatedRepeatedPtrFieldBackInsertIterator
-    : public std::iterator<std::output_iterator_tag, T> {
- public:
-  explicit AllocatedRepeatedPtrFieldBackInsertIterator(
-      RepeatedPtrField<T>* const mutable_field)
-      : field_(mutable_field) {}
-  AllocatedRepeatedPtrFieldBackInsertIterator<T>& operator=(
-      T* const ptr_to_value) {
-    field_->AddAllocated(ptr_to_value);
-    return *this;
-  }
-  AllocatedRepeatedPtrFieldBackInsertIterator<T>& operator*() { return *this; }
-  AllocatedRepeatedPtrFieldBackInsertIterator<T>& operator++() { return *this; }
-  AllocatedRepeatedPtrFieldBackInsertIterator<T>& operator++(int /* unused */) {
-    return *this;
-  }
-
- private:
-  RepeatedPtrField<T>* field_;
-};
-
-// Almost identical to AllocatedRepeatedPtrFieldBackInsertIterator. This one
-// uses the UnsafeArenaAddAllocated instead.
-template <typename T>
-class UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator
-    : public std::iterator<std::output_iterator_tag, T> {
- public:
-  explicit UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator(
-      RepeatedPtrField<T>* const mutable_field)
-      : field_(mutable_field) {}
-  UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator<T>& operator=(
-      T const* const ptr_to_value) {
-    field_->UnsafeArenaAddAllocated(const_cast<T*>(ptr_to_value));
-    return *this;
-  }
-  UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator<T>& operator*() {
-    return *this;
-  }
-  UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator<T>& operator++() {
-    return *this;
-  }
-  UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator<T>& operator++(
-      int /* unused */) {
-    return *this;
-  }
-
- private:
-  RepeatedPtrField<T>* field_;
-};
-
-}  // namespace internal
-
-// Provides a back insert iterator for RepeatedField instances,
-// similar to std::back_inserter().
-template <typename T>
-internal::RepeatedFieldBackInsertIterator<T> RepeatedFieldBackInserter(
-    RepeatedField<T>* const mutable_field) {
-  return internal::RepeatedFieldBackInsertIterator<T>(mutable_field);
-}
-
-// Provides a back insert iterator for RepeatedPtrField instances,
-// similar to std::back_inserter().
-template <typename T>
-internal::RepeatedPtrFieldBackInsertIterator<T> RepeatedPtrFieldBackInserter(
-    RepeatedPtrField<T>* const mutable_field) {
-  return internal::RepeatedPtrFieldBackInsertIterator<T>(mutable_field);
-}
-
-// Special back insert iterator for RepeatedPtrField instances, just in
-// case someone wants to write generic template code that can access both
-// RepeatedFields and RepeatedPtrFields using a common name.
-template <typename T>
-internal::RepeatedPtrFieldBackInsertIterator<T> RepeatedFieldBackInserter(
-    RepeatedPtrField<T>* const mutable_field) {
-  return internal::RepeatedPtrFieldBackInsertIterator<T>(mutable_field);
-}
-
-// Provides a back insert iterator for RepeatedPtrField instances
-// similar to std::back_inserter() which transfers the ownership while
-// copying elements.
-template <typename T>
-internal::AllocatedRepeatedPtrFieldBackInsertIterator<T>
-AllocatedRepeatedPtrFieldBackInserter(
-    RepeatedPtrField<T>* const mutable_field) {
-  return internal::AllocatedRepeatedPtrFieldBackInsertIterator<T>(
-      mutable_field);
-}
-
-// Similar to AllocatedRepeatedPtrFieldBackInserter, using
-// UnsafeArenaAddAllocated instead of AddAllocated.
-// This is slightly faster if that matters. It is also useful in legacy code
-// that uses temporary ownership to avoid copies. Example:
-//   RepeatedPtrField<T> temp_field;
-//   temp_field.AddAllocated(new T);
-//   ... // Do something with temp_field
-//   temp_field.ExtractSubrange(0, temp_field.size(), nullptr);
-// If you put temp_field on the arena this fails, because the ownership
-// transfers to the arena at the "AddAllocated" call and is not released anymore
-// causing a double delete. Using UnsafeArenaAddAllocated prevents this.
-template <typename T>
-internal::UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator<T>
-UnsafeArenaAllocatedRepeatedPtrFieldBackInserter(
-    RepeatedPtrField<T>* const mutable_field) {
-  return internal::UnsafeArenaAllocatedRepeatedPtrFieldBackInsertIterator<T>(
-      mutable_field);
-}
-
-// Extern declarations of common instantiations to reduce library bloat.
-extern template class PROTOBUF_EXPORT_TEMPLATE_DECLARE RepeatedField<bool>;
-extern template class PROTOBUF_EXPORT_TEMPLATE_DECLARE RepeatedField<int32>;
-extern template class PROTOBUF_EXPORT_TEMPLATE_DECLARE RepeatedField<uint32>;
-extern template class PROTOBUF_EXPORT_TEMPLATE_DECLARE RepeatedField<int64>;
-extern template class PROTOBUF_EXPORT_TEMPLATE_DECLARE RepeatedField<uint64>;
-extern template class PROTOBUF_EXPORT_TEMPLATE_DECLARE RepeatedField<float>;
-extern template class PROTOBUF_EXPORT_TEMPLATE_DECLARE RepeatedField<double>;
-extern template class PROTOBUF_EXPORT_TEMPLATE_DECLARE
-    RepeatedPtrField<std::string>;
-
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_REPEATED_FIELD_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/bytestream.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/bytestream.h
deleted file mode 100644
index c7a48dea..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/bytestream.h
+++ /dev/null
@@ -1,351 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// This file declares the ByteSink and ByteSource abstract interfaces. These
-// interfaces represent objects that consume (ByteSink) or produce (ByteSource)
-// a sequence of bytes. Using these abstract interfaces in your APIs can help
-// make your code work with a variety of input and output types.
-//
-// This file also declares the following commonly used implementations of these
-// interfaces.
-//
-//   ByteSink:
-//      UncheckedArrayByteSink  Writes to an array, without bounds checking
-//      CheckedArrayByteSink    Writes to an array, with bounds checking
-//      GrowingArrayByteSink    Allocates and writes to a growable buffer
-//      StringByteSink          Writes to an STL string
-//      NullByteSink            Consumes a never-ending stream of bytes
-//
-//   ByteSource:
-//      ArrayByteSource         Reads from an array or string/StringPiece
-//      LimitedByteSource       Limits the number of bytes read from an
-
-#ifndef GOOGLE_PROTOBUF_STUBS_BYTESTREAM_H_
-#define GOOGLE_PROTOBUF_STUBS_BYTESTREAM_H_
-
-#include <stddef.h>
-#include <string>
-
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/stubs/stringpiece.h>
-
-#include <google/protobuf/port_def.inc>
-
-class CordByteSink;
-
-namespace google {
-namespace protobuf {
-namespace strings {
-
-// An abstract interface for an object that consumes a sequence of bytes. This
-// interface offers a way to append data as well as a Flush() function.
-//
-// Example:
-//
-//   string my_data;
-//   ...
-//   ByteSink* sink = ...
-//   sink->Append(my_data.data(), my_data.size());
-//   sink->Flush();
-//
-class PROTOBUF_EXPORT ByteSink {
- public:
-  ByteSink() {}
-  virtual ~ByteSink() {}
-
-  // Appends the "n" bytes starting at "bytes".
-  virtual void Append(const char* bytes, size_t n) = 0;
-
-  // Flushes internal buffers. The default implementation does nothing. ByteSink
-  // subclasses may use internal buffers that require calling Flush() at the end
-  // of the stream.
-  virtual void Flush();
-
- private:
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ByteSink);
-};
-
-// An abstract interface for an object that produces a fixed-size sequence of
-// bytes.
-//
-// Example:
-//
-//   ByteSource* source = ...
-//   while (source->Available() > 0) {
-//     StringPiece data = source->Peek();
-//     ... do something with "data" ...
-//     source->Skip(data.length());
-//   }
-//
-class PROTOBUF_EXPORT ByteSource {
- public:
-  ByteSource() {}
-  virtual ~ByteSource() {}
-
-  // Returns the number of bytes left to read from the source. Available()
-  // should decrease by N each time Skip(N) is called. Available() may not
-  // increase. Available() returning 0 indicates that the ByteSource is
-  // exhausted.
-  //
-  // Note: Size() may have been a more appropriate name as it's more
-  //       indicative of the fixed-size nature of a ByteSource.
-  virtual size_t Available() const = 0;
-
-  // Returns a StringPiece of the next contiguous region of the source. Does not
-  // reposition the source. The returned region is empty iff Available() == 0.
-  //
-  // The returned region is valid until the next call to Skip() or until this
-  // object is destroyed, whichever occurs first.
-  //
-  // The length of the returned StringPiece will be <= Available().
-  virtual StringPiece Peek() = 0;
-
-  // Skips the next n bytes. Invalidates any StringPiece returned by a previous
-  // call to Peek().
-  //
-  // REQUIRES: Available() >= n
-  virtual void Skip(size_t n) = 0;
-
-  // Writes the next n bytes in this ByteSource to the given ByteSink, and
-  // advances this ByteSource past the copied bytes. The default implementation
-  // of this method just copies the bytes normally, but subclasses might
-  // override CopyTo to optimize certain cases.
-  //
-  // REQUIRES: Available() >= n
-  virtual void CopyTo(ByteSink* sink, size_t n);
-
- private:
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ByteSource);
-};
-
-//
-// Some commonly used implementations of ByteSink
-//
-
-// Implementation of ByteSink that writes to an unsized byte array. No
-// bounds-checking is performed--it is the caller's responsibility to ensure
-// that the destination array is large enough.
-//
-// Example:
-//
-//   char buf[10];
-//   UncheckedArrayByteSink sink(buf);
-//   sink.Append("hi", 2);    // OK
-//   sink.Append(data, 100);  // WOOPS! Overflows buf[10].
-//
-class PROTOBUF_EXPORT UncheckedArrayByteSink : public ByteSink {
- public:
-  explicit UncheckedArrayByteSink(char* dest) : dest_(dest) {}
-  virtual void Append(const char* data, size_t n) override;
-
-  // Returns the current output pointer so that a caller can see how many bytes
-  // were produced.
-  //
-  // Note: this method is not part of the ByteSink interface.
-  char* CurrentDestination() const { return dest_; }
-
- private:
-  char* dest_;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(UncheckedArrayByteSink);
-};
-
-// Implementation of ByteSink that writes to a sized byte array. This sink will
-// not write more than "capacity" bytes to outbuf. Once "capacity" bytes are
-// appended, subsequent bytes will be ignored and Overflowed() will return true.
-// Overflowed() does not cause a runtime error (i.e., it does not CHECK fail).
-//
-// Example:
-//
-//   char buf[10];
-//   CheckedArrayByteSink sink(buf, 10);
-//   sink.Append("hi", 2);    // OK
-//   sink.Append(data, 100);  // Will only write 8 more bytes
-//
-class PROTOBUF_EXPORT CheckedArrayByteSink : public ByteSink {
- public:
-  CheckedArrayByteSink(char* outbuf, size_t capacity);
-  virtual void Append(const char* bytes, size_t n) override;
-
-  // Returns the number of bytes actually written to the sink.
-  size_t NumberOfBytesWritten() const { return size_; }
-
-  // Returns true if any bytes were discarded, i.e., if there was an
-  // attempt to write more than 'capacity' bytes.
-  bool Overflowed() const { return overflowed_; }
-
- private:
-  char* outbuf_;
-  const size_t capacity_;
-  size_t size_;
-  bool overflowed_;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(CheckedArrayByteSink);
-};
-
-// Implementation of ByteSink that allocates an internal buffer (a char array)
-// and expands it as needed to accommodate appended data (similar to a string),
-// and allows the caller to take ownership of the internal buffer via the
-// GetBuffer() method. The buffer returned from GetBuffer() must be deleted by
-// the caller with delete[]. GetBuffer() also sets the internal buffer to be
-// empty, and subsequent appends to the sink will create a new buffer. The
-// destructor will free the internal buffer if GetBuffer() was not called.
-//
-// Example:
-//
-//   GrowingArrayByteSink sink(10);
-//   sink.Append("hi", 2);
-//   sink.Append(data, n);
-//   const char* buf = sink.GetBuffer();  // Ownership transferred
-//   delete[] buf;
-//
-class PROTOBUF_EXPORT GrowingArrayByteSink : public strings::ByteSink {
- public:
-  explicit GrowingArrayByteSink(size_t estimated_size);
-  virtual ~GrowingArrayByteSink();
-  virtual void Append(const char* bytes, size_t n) override;
-
-  // Returns the allocated buffer, and sets nbytes to its size. The caller takes
-  // ownership of the buffer and must delete it with delete[].
-  char* GetBuffer(size_t* nbytes);
-
- private:
-  void Expand(size_t amount);
-  void ShrinkToFit();
-
-  size_t capacity_;
-  char* buf_;
-  size_t size_;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(GrowingArrayByteSink);
-};
-
-// Implementation of ByteSink that appends to the given string.
-// Existing contents of "dest" are not modified; new data is appended.
-//
-// Example:
-//
-//   string dest = "Hello ";
-//   StringByteSink sink(&dest);
-//   sink.Append("World", 5);
-//   assert(dest == "Hello World");
-//
-class PROTOBUF_EXPORT StringByteSink : public ByteSink {
- public:
-  explicit StringByteSink(std::string* dest) : dest_(dest) {}
-  virtual void Append(const char* data, size_t n) override;
-
- private:
-  std::string* dest_;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(StringByteSink);
-};
-
-// Implementation of ByteSink that discards all data.
-//
-// Example:
-//
-//   NullByteSink sink;
-//   sink.Append(data, data.size());  // All data ignored.
-//
-class PROTOBUF_EXPORT NullByteSink : public ByteSink {
- public:
-  NullByteSink() {}
-  void Append(const char* /*data*/, size_t /*n*/) override {}
-
- private:
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(NullByteSink);
-};
-
-//
-// Some commonly used implementations of ByteSource
-//
-
-// Implementation of ByteSource that reads from a StringPiece.
-//
-// Example:
-//
-//   string data = "Hello";
-//   ArrayByteSource source(data);
-//   assert(source.Available() == 5);
-//   assert(source.Peek() == "Hello");
-//
-class PROTOBUF_EXPORT ArrayByteSource : public ByteSource {
- public:
-  explicit ArrayByteSource(StringPiece s) : input_(s) {}
-
-  virtual size_t Available() const override;
-  virtual StringPiece Peek() override;
-  virtual void Skip(size_t n) override;
-
- private:
-  StringPiece   input_;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ArrayByteSource);
-};
-
-// Implementation of ByteSource that wraps another ByteSource, limiting the
-// number of bytes returned.
-//
-// The caller maintains ownership of the underlying source, and may not use the
-// underlying source while using the LimitByteSource object.  The underlying
-// source's pointer is advanced by n bytes every time this LimitByteSource
-// object is advanced by n.
-//
-// Example:
-//
-//   string data = "Hello World";
-//   ArrayByteSource abs(data);
-//   assert(abs.Available() == data.size());
-//
-//   LimitByteSource limit(abs, 5);
-//   assert(limit.Available() == 5);
-//   assert(limit.Peek() == "Hello");
-//
-class PROTOBUF_EXPORT LimitByteSource : public ByteSource {
- public:
-  // Returns at most "limit" bytes from "source".
-  LimitByteSource(ByteSource* source, size_t limit);
-
-  virtual size_t Available() const override;
-  virtual StringPiece Peek() override;
-  virtual void Skip(size_t n) override;
-
-  // We override CopyTo so that we can forward to the underlying source, in
-  // case it has an efficient implementation of CopyTo.
-  virtual void CopyTo(ByteSink* sink, size_t n) override;
-
- private:
-  ByteSource* source_;
-  size_t limit_;
-};
-
-}  // namespace strings
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_STUBS_BYTESTREAM_H_
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/callback.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/callback.h
deleted file mode 100644
index 43d546d1..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/callback.h
+++ /dev/null
@@ -1,583 +0,0 @@
-#ifndef GOOGLE_PROTOBUF_STUBS_CALLBACK_H_
-#define GOOGLE_PROTOBUF_STUBS_CALLBACK_H_
-
-#include <type_traits>
-
-#include <google/protobuf/stubs/macros.h>
-
-#include <google/protobuf/port_def.inc>
-
-// ===================================================================
-// emulates google3/base/callback.h
-
-namespace google {
-namespace protobuf {
-
-// Abstract interface for a callback.  When calling an RPC, you must provide
-// a Closure to call when the procedure completes.  See the Service interface
-// in service.h.
-//
-// To automatically construct a Closure which calls a particular function or
-// method with a particular set of parameters, use the NewCallback() function.
-// Example:
-//   void FooDone(const FooResponse* response) {
-//     ...
-//   }
-//
-//   void CallFoo() {
-//     ...
-//     // When done, call FooDone() and pass it a pointer to the response.
-//     Closure* callback = NewCallback(&FooDone, response);
-//     // Make the call.
-//     service->Foo(controller, request, response, callback);
-//   }
-//
-// Example that calls a method:
-//   class Handler {
-//    public:
-//     ...
-//
-//     void FooDone(const FooResponse* response) {
-//       ...
-//     }
-//
-//     void CallFoo() {
-//       ...
-//       // When done, call FooDone() and pass it a pointer to the response.
-//       Closure* callback = NewCallback(this, &Handler::FooDone, response);
-//       // Make the call.
-//       service->Foo(controller, request, response, callback);
-//     }
-//   };
-//
-// Currently NewCallback() supports binding zero, one, or two arguments.
-//
-// Callbacks created with NewCallback() automatically delete themselves when
-// executed.  They should be used when a callback is to be called exactly
-// once (usually the case with RPC callbacks).  If a callback may be called
-// a different number of times (including zero), create it with
-// NewPermanentCallback() instead.  You are then responsible for deleting the
-// callback (using the "delete" keyword as normal).
-//
-// Note that NewCallback() is a bit touchy regarding argument types.  Generally,
-// the values you provide for the parameter bindings must exactly match the
-// types accepted by the callback function.  For example:
-//   void Foo(std::string s);
-//   NewCallback(&Foo, "foo");          // WON'T WORK:  const char* != string
-//   NewCallback(&Foo, std::string("foo"));  // WORKS
-// Also note that the arguments cannot be references:
-//   void Foo(const std::string& s);
-//   std::string my_str;
-//   NewCallback(&Foo, my_str);  // WON'T WORK:  Can't use references.
-// However, correctly-typed pointers will work just fine.
-class PROTOBUF_EXPORT Closure {
- public:
-  Closure() {}
-  virtual ~Closure();
-
-  virtual void Run() = 0;
-
- private:
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(Closure);
-};
-
-template<typename R>
-class ResultCallback {
- public:
-  ResultCallback() {}
-  virtual ~ResultCallback() {}
-
-  virtual R Run() = 0;
-
- private:
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ResultCallback);
-};
-
-template <typename R, typename A1>
-class PROTOBUF_EXPORT ResultCallback1 {
- public:
-  ResultCallback1() {}
-  virtual ~ResultCallback1() {}
-
-  virtual R Run(A1) = 0;
-
- private:
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ResultCallback1);
-};
-
-template <typename R, typename A1, typename A2>
-class PROTOBUF_EXPORT ResultCallback2 {
- public:
-  ResultCallback2() {}
-  virtual ~ResultCallback2() {}
-
-  virtual R Run(A1,A2) = 0;
-
- private:
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ResultCallback2);
-};
-
-namespace internal {
-
-class PROTOBUF_EXPORT FunctionClosure0 : public Closure {
- public:
-  typedef void (*FunctionType)();
-
-  FunctionClosure0(FunctionType function, bool self_deleting)
-    : function_(function), self_deleting_(self_deleting) {}
-  ~FunctionClosure0();
-
-  void Run() override {
-    bool needs_delete = self_deleting_;  // read in case callback deletes
-    function_();
-    if (needs_delete) delete this;
-  }
-
- private:
-  FunctionType function_;
-  bool self_deleting_;
-};
-
-template <typename Class>
-class MethodClosure0 : public Closure {
- public:
-  typedef void (Class::*MethodType)();
-
-  MethodClosure0(Class* object, MethodType method, bool self_deleting)
-    : object_(object), method_(method), self_deleting_(self_deleting) {}
-  ~MethodClosure0() {}
-
-  void Run() override {
-    bool needs_delete = self_deleting_;  // read in case callback deletes
-    (object_->*method_)();
-    if (needs_delete) delete this;
-  }
-
- private:
-  Class* object_;
-  MethodType method_;
-  bool self_deleting_;
-};
-
-template <typename Arg1>
-class FunctionClosure1 : public Closure {
- public:
-  typedef void (*FunctionType)(Arg1 arg1);
-
-  FunctionClosure1(FunctionType function, bool self_deleting,
-                   Arg1 arg1)
-    : function_(function), self_deleting_(self_deleting),
-      arg1_(arg1) {}
-  ~FunctionClosure1() {}
-
-  void Run() override {
-    bool needs_delete = self_deleting_;  // read in case callback deletes
-    function_(arg1_);
-    if (needs_delete) delete this;
-  }
-
- private:
-  FunctionType function_;
-  bool self_deleting_;
-  Arg1 arg1_;
-};
-
-template <typename Class, typename Arg1>
-class MethodClosure1 : public Closure {
- public:
-  typedef void (Class::*MethodType)(Arg1 arg1);
-
-  MethodClosure1(Class* object, MethodType method, bool self_deleting,
-                 Arg1 arg1)
-    : object_(object), method_(method), self_deleting_(self_deleting),
-      arg1_(arg1) {}
-  ~MethodClosure1() {}
-
-  void Run() override {
-    bool needs_delete = self_deleting_;  // read in case callback deletes
-    (object_->*method_)(arg1_);
-    if (needs_delete) delete this;
-  }
-
- private:
-  Class* object_;
-  MethodType method_;
-  bool self_deleting_;
-  Arg1 arg1_;
-};
-
-template <typename Arg1, typename Arg2>
-class FunctionClosure2 : public Closure {
- public:
-  typedef void (*FunctionType)(Arg1 arg1, Arg2 arg2);
-
-  FunctionClosure2(FunctionType function, bool self_deleting,
-                   Arg1 arg1, Arg2 arg2)
-    : function_(function), self_deleting_(self_deleting),
-      arg1_(arg1), arg2_(arg2) {}
-  ~FunctionClosure2() {}
-
-  void Run() override {
-    bool needs_delete = self_deleting_;  // read in case callback deletes
-    function_(arg1_, arg2_);
-    if (needs_delete) delete this;
-  }
-
- private:
-  FunctionType function_;
-  bool self_deleting_;
-  Arg1 arg1_;
-  Arg2 arg2_;
-};
-
-template <typename Class, typename Arg1, typename Arg2>
-class MethodClosure2 : public Closure {
- public:
-  typedef void (Class::*MethodType)(Arg1 arg1, Arg2 arg2);
-
-  MethodClosure2(Class* object, MethodType method, bool self_deleting,
-                 Arg1 arg1, Arg2 arg2)
-    : object_(object), method_(method), self_deleting_(self_deleting),
-      arg1_(arg1), arg2_(arg2) {}
-  ~MethodClosure2() {}
-
-  void Run() override {
-    bool needs_delete = self_deleting_;  // read in case callback deletes
-    (object_->*method_)(arg1_, arg2_);
-    if (needs_delete) delete this;
-  }
-
- private:
-  Class* object_;
-  MethodType method_;
-  bool self_deleting_;
-  Arg1 arg1_;
-  Arg2 arg2_;
-};
-
-template<typename R>
-class FunctionResultCallback_0_0 : public ResultCallback<R> {
- public:
-  typedef R (*FunctionType)();
-
-  FunctionResultCallback_0_0(FunctionType function, bool self_deleting)
-      : function_(function), self_deleting_(self_deleting) {}
-  ~FunctionResultCallback_0_0() {}
-
-  R Run() override {
-    bool needs_delete = self_deleting_;  // read in case callback deletes
-    R result = function_();
-    if (needs_delete) delete this;
-    return result;
-  }
-
- private:
-  FunctionType function_;
-  bool self_deleting_;
-};
-
-template<typename R, typename P1>
-class FunctionResultCallback_1_0 : public ResultCallback<R> {
- public:
-  typedef R (*FunctionType)(P1);
-
-  FunctionResultCallback_1_0(FunctionType function, bool self_deleting,
-                             P1 p1)
-      : function_(function), self_deleting_(self_deleting), p1_(p1) {}
-  ~FunctionResultCallback_1_0() {}
-
-  R Run() override {
-    bool needs_delete = self_deleting_;  // read in case callback deletes
-    R result = function_(p1_);
-    if (needs_delete) delete this;
-    return result;
-  }
-
- private:
-  FunctionType function_;
-  bool self_deleting_;
-  P1 p1_;
-};
-
-template<typename R, typename Arg1>
-class FunctionResultCallback_0_1 : public ResultCallback1<R, Arg1> {
- public:
-  typedef R (*FunctionType)(Arg1 arg1);
-
-  FunctionResultCallback_0_1(FunctionType function, bool self_deleting)
-      : function_(function), self_deleting_(self_deleting) {}
-  ~FunctionResultCallback_0_1() {}
-
-  R Run(Arg1 a1) override {
-    bool needs_delete = self_deleting_;  // read in case callback deletes
-    R result = function_(a1);
-    if (needs_delete) delete this;
-    return result;
-  }
-
- private:
-  FunctionType function_;
-  bool self_deleting_;
-};
-
-template<typename R, typename P1, typename A1>
-class FunctionResultCallback_1_1 : public ResultCallback1<R, A1> {
- public:
-  typedef R (*FunctionType)(P1, A1);
-
-  FunctionResultCallback_1_1(FunctionType function, bool self_deleting,
-                             P1 p1)
-      : function_(function), self_deleting_(self_deleting), p1_(p1) {}
-  ~FunctionResultCallback_1_1() {}
-
-  R Run(A1 a1) override {
-    bool needs_delete = self_deleting_;  // read in case callback deletes
-    R result = function_(p1_, a1);
-    if (needs_delete) delete this;
-    return result;
-  }
-
- private:
-  FunctionType function_;
-  bool self_deleting_;
-  P1 p1_;
-};
-
-template <typename T>
-struct InternalConstRef {
-  typedef typename std::remove_reference<T>::type base_type;
-  typedef const base_type& type;
-};
-
-template<typename R, typename T>
-class MethodResultCallback_0_0 : public ResultCallback<R> {
- public:
-  typedef R (T::*MethodType)();
-  MethodResultCallback_0_0(T* object, MethodType method, bool self_deleting)
-      : object_(object),
-        method_(method),
-        self_deleting_(self_deleting) {}
-  ~MethodResultCallback_0_0() {}
-
-  R Run() {
-    bool needs_delete = self_deleting_;
-    R result = (object_->*method_)();
-    if (needs_delete) delete this;
-    return result;
-  }
-
- private:
-  T* object_;
-  MethodType method_;
-  bool self_deleting_;
-};
-
-template <typename R, typename T, typename P1, typename P2, typename P3,
-          typename P4, typename P5, typename P6, typename A1, typename A2>
-class MethodResultCallback_6_2 : public ResultCallback2<R, A1, A2> {
- public:
-  typedef R (T::*MethodType)(P1, P2, P3, P4, P5, P6, A1, A2);
-  MethodResultCallback_6_2(T* object, MethodType method, bool self_deleting,
-                           P1 p1, P2 p2, P3 p3, P4 p4, P5 p5, P6 p6)
-      : object_(object),
-        method_(method),
-        self_deleting_(self_deleting),
-        p1_(p1),
-        p2_(p2),
-        p3_(p3),
-        p4_(p4),
-        p5_(p5),
-        p6_(p6) {}
-  ~MethodResultCallback_6_2() {}
-
-  R Run(A1 a1, A2 a2) override {
-    bool needs_delete = self_deleting_;
-    R result = (object_->*method_)(p1_, p2_, p3_, p4_, p5_, p6_, a1, a2);
-    if (needs_delete) delete this;
-    return result;
-  }
-
- private:
-  T* object_;
-  MethodType method_;
-  bool self_deleting_;
-  typename std::remove_reference<P1>::type p1_;
-  typename std::remove_reference<P2>::type p2_;
-  typename std::remove_reference<P3>::type p3_;
-  typename std::remove_reference<P4>::type p4_;
-  typename std::remove_reference<P5>::type p5_;
-  typename std::remove_reference<P6>::type p6_;
-};
-
-}  // namespace internal
-
-// See Closure.
-inline Closure* NewCallback(void (*function)()) {
-  return new internal::FunctionClosure0(function, true);
-}
-
-// See Closure.
-inline Closure* NewPermanentCallback(void (*function)()) {
-  return new internal::FunctionClosure0(function, false);
-}
-
-// See Closure.
-template <typename Class>
-inline Closure* NewCallback(Class* object, void (Class::*method)()) {
-  return new internal::MethodClosure0<Class>(object, method, true);
-}
-
-// See Closure.
-template <typename Class>
-inline Closure* NewPermanentCallback(Class* object, void (Class::*method)()) {
-  return new internal::MethodClosure0<Class>(object, method, false);
-}
-
-// See Closure.
-template <typename Arg1>
-inline Closure* NewCallback(void (*function)(Arg1),
-                            Arg1 arg1) {
-  return new internal::FunctionClosure1<Arg1>(function, true, arg1);
-}
-
-// See Closure.
-template <typename Arg1>
-inline Closure* NewPermanentCallback(void (*function)(Arg1),
-                                     Arg1 arg1) {
-  return new internal::FunctionClosure1<Arg1>(function, false, arg1);
-}
-
-// See Closure.
-template <typename Class, typename Arg1>
-inline Closure* NewCallback(Class* object, void (Class::*method)(Arg1),
-                            Arg1 arg1) {
-  return new internal::MethodClosure1<Class, Arg1>(object, method, true, arg1);
-}
-
-// See Closure.
-template <typename Class, typename Arg1>
-inline Closure* NewPermanentCallback(Class* object, void (Class::*method)(Arg1),
-                                     Arg1 arg1) {
-  return new internal::MethodClosure1<Class, Arg1>(object, method, false, arg1);
-}
-
-// See Closure.
-template <typename Arg1, typename Arg2>
-inline Closure* NewCallback(void (*function)(Arg1, Arg2),
-                            Arg1 arg1, Arg2 arg2) {
-  return new internal::FunctionClosure2<Arg1, Arg2>(
-    function, true, arg1, arg2);
-}
-
-// See Closure.
-template <typename Arg1, typename Arg2>
-inline Closure* NewPermanentCallback(void (*function)(Arg1, Arg2),
-                                     Arg1 arg1, Arg2 arg2) {
-  return new internal::FunctionClosure2<Arg1, Arg2>(
-    function, false, arg1, arg2);
-}
-
-// See Closure.
-template <typename Class, typename Arg1, typename Arg2>
-inline Closure* NewCallback(Class* object, void (Class::*method)(Arg1, Arg2),
-                            Arg1 arg1, Arg2 arg2) {
-  return new internal::MethodClosure2<Class, Arg1, Arg2>(
-    object, method, true, arg1, arg2);
-}
-
-// See Closure.
-template <typename Class, typename Arg1, typename Arg2>
-inline Closure* NewPermanentCallback(
-    Class* object, void (Class::*method)(Arg1, Arg2),
-    Arg1 arg1, Arg2 arg2) {
-  return new internal::MethodClosure2<Class, Arg1, Arg2>(
-    object, method, false, arg1, arg2);
-}
-
-// See ResultCallback
-template<typename R>
-inline ResultCallback<R>* NewCallback(R (*function)()) {
-  return new internal::FunctionResultCallback_0_0<R>(function, true);
-}
-
-// See ResultCallback
-template<typename R>
-inline ResultCallback<R>* NewPermanentCallback(R (*function)()) {
-  return new internal::FunctionResultCallback_0_0<R>(function, false);
-}
-
-// See ResultCallback
-template<typename R, typename P1>
-inline ResultCallback<R>* NewCallback(R (*function)(P1), P1 p1) {
-  return new internal::FunctionResultCallback_1_0<R, P1>(
-      function, true, p1);
-}
-
-// See ResultCallback
-template<typename R, typename P1>
-inline ResultCallback<R>* NewPermanentCallback(
-    R (*function)(P1), P1 p1) {
-  return new internal::FunctionResultCallback_1_0<R, P1>(
-      function, false, p1);
-}
-
-// See ResultCallback1
-template<typename R, typename A1>
-inline ResultCallback1<R, A1>* NewCallback(R (*function)(A1)) {
-  return new internal::FunctionResultCallback_0_1<R, A1>(function, true);
-}
-
-// See ResultCallback1
-template<typename R, typename A1>
-inline ResultCallback1<R, A1>* NewPermanentCallback(R (*function)(A1)) {
-  return new internal::FunctionResultCallback_0_1<R, A1>(function, false);
-}
-
-// See ResultCallback1
-template<typename R, typename P1, typename A1>
-inline ResultCallback1<R, A1>* NewCallback(R (*function)(P1, A1), P1 p1) {
-  return new internal::FunctionResultCallback_1_1<R, P1, A1>(
-      function, true, p1);
-}
-
-// See ResultCallback1
-template<typename R, typename P1, typename A1>
-inline ResultCallback1<R, A1>* NewPermanentCallback(
-    R (*function)(P1, A1), P1 p1) {
-  return new internal::FunctionResultCallback_1_1<R, P1, A1>(
-      function, false, p1);
-}
-
-// See MethodResultCallback_0_0
-template <typename R, typename T1, typename T2>
-inline ResultCallback<R>* NewPermanentCallback(
-    T1* object, R (T2::*function)()) {
-  return new internal::MethodResultCallback_0_0<R, T1>(object, function, false);
-}
-
-// See MethodResultCallback_6_2
-template <typename R, typename T, typename P1, typename P2, typename P3,
-          typename P4, typename P5, typename P6, typename A1, typename A2>
-inline ResultCallback2<R, A1, A2>* NewPermanentCallback(
-    T* object, R (T::*function)(P1, P2, P3, P4, P5, P6, A1, A2),
-    typename internal::InternalConstRef<P1>::type p1,
-    typename internal::InternalConstRef<P2>::type p2,
-    typename internal::InternalConstRef<P3>::type p3,
-    typename internal::InternalConstRef<P4>::type p4,
-    typename internal::InternalConstRef<P5>::type p5,
-    typename internal::InternalConstRef<P6>::type p6) {
-  return new internal::MethodResultCallback_6_2<R, T, P1, P2, P3, P4, P5, P6,
-                                                A1, A2>(object, function, false,
-                                                        p1, p2, p3, p4, p5, p6);
-}
-
-// A function which does nothing.  Useful for creating no-op callbacks, e.g.:
-//   Closure* nothing = NewCallback(&DoNothing);
-void PROTOBUF_EXPORT DoNothing();
-
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_STUBS_CALLBACK_H_
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/casts.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/casts.h
deleted file mode 100644
index d8a49cec..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/casts.h
+++ /dev/null
@@ -1,139 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2014 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_CASTS_H__
-#define GOOGLE_PROTOBUF_CASTS_H__
-
-#include <google/protobuf/stubs/common.h>
-
-#include <google/protobuf/port_def.inc>
-#include <type_traits>
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-// Use implicit_cast as a safe version of static_cast or const_cast
-// for upcasting in the type hierarchy (i.e. casting a pointer to Foo
-// to a pointer to SuperclassOfFoo or casting a pointer to Foo to
-// a const pointer to Foo).
-// When you use implicit_cast, the compiler checks that the cast is safe.
-// Such explicit implicit_casts are necessary in surprisingly many
-// situations where C++ demands an exact type match instead of an
-// argument type convertible to a target type.
-//
-// The From type can be inferred, so the preferred syntax for using
-// implicit_cast is the same as for static_cast etc.:
-//
-//   implicit_cast<ToType>(expr)
-//
-// implicit_cast would have been part of the C++ standard library,
-// but the proposal was submitted too late.  It will probably make
-// its way into the language in the future.
-template<typename To, typename From>
-inline To implicit_cast(From const &f) {
-  return f;
-}
-
-// When you upcast (that is, cast a pointer from type Foo to type
-// SuperclassOfFoo), it's fine to use implicit_cast<>, since upcasts
-// always succeed.  When you downcast (that is, cast a pointer from
-// type Foo to type SubclassOfFoo), static_cast<> isn't safe, because
-// how do you know the pointer is really of type SubclassOfFoo?  It
-// could be a bare Foo, or of type DifferentSubclassOfFoo.  Thus,
-// when you downcast, you should use this macro.  In debug mode, we
-// use dynamic_cast<> to double-check the downcast is legal (we die
-// if it's not).  In normal mode, we do the efficient static_cast<>
-// instead.  Thus, it's important to test in debug mode to make sure
-// the cast is legal!
-//    This is the only place in the code we should use dynamic_cast<>.
-// In particular, you SHOULDN'T be using dynamic_cast<> in order to
-// do RTTI (eg code like this:
-//    if (dynamic_cast<Subclass1>(foo)) HandleASubclass1Object(foo);
-//    if (dynamic_cast<Subclass2>(foo)) HandleASubclass2Object(foo);
-// You should design the code some other way not to need this.
-
-template<typename To, typename From>     // use like this: down_cast<T*>(foo);
-inline To down_cast(From* f) {                   // so we only accept pointers
-  // Ensures that To is a sub-type of From *.  This test is here only
-  // for compile-time type checking, and has no overhead in an
-  // optimized build at run-time, as it will be optimized away
-  // completely.
-  if (false) {
-    implicit_cast<From*, To>(0);
-  }
-
-#if !defined(NDEBUG) && PROTOBUF_RTTI
-  assert(f == nullptr || dynamic_cast<To>(f) != nullptr);  // RTTI: debug mode only!
-#endif
-  return static_cast<To>(f);
-}
-
-template<typename To, typename From>    // use like this: down_cast<T&>(foo);
-inline To down_cast(From& f) {
-  typedef typename std::remove_reference<To>::type* ToAsPointer;
-  // Ensures that To is a sub-type of From *.  This test is here only
-  // for compile-time type checking, and has no overhead in an
-  // optimized build at run-time, as it will be optimized away
-  // completely.
-  if (false) {
-    implicit_cast<From*, ToAsPointer>(0);
-  }
-
-#if !defined(NDEBUG) && PROTOBUF_RTTI
-  // RTTI: debug mode only!
-  assert(dynamic_cast<ToAsPointer>(&f) != nullptr);
-#endif
-  return *static_cast<ToAsPointer>(&f);
-}
-
-template<typename To, typename From>
-inline To bit_cast(const From& from) {
-  GOOGLE_COMPILE_ASSERT(sizeof(From) == sizeof(To),
-                        bit_cast_with_different_sizes);
-  To dest;
-  memcpy(&dest, &from, sizeof(dest));
-  return dest;
-}
-
-}  // namespace internal
-
-// We made these internal so that they would show up as such in the docs,
-// but we don't want to stick "internal::" in front of them everywhere.
-using internal::implicit_cast;
-using internal::down_cast;
-using internal::bit_cast;
-
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_CASTS_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/common.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/common.h
deleted file mode 100644
index 2d210d72..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/common.h
+++ /dev/null
@@ -1,202 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// Author: kenton@google.com (Kenton Varda) and others
-//
-// Contains basic types and utilities used by the rest of the library.
-
-#ifndef GOOGLE_PROTOBUF_COMMON_H__
-#define GOOGLE_PROTOBUF_COMMON_H__
-
-#include <algorithm>
-#include <iostream>
-#include <map>
-#include <memory>
-#include <set>
-#include <string>
-#include <vector>
-
-#include <google/protobuf/stubs/macros.h>
-#include <google/protobuf/stubs/platform_macros.h>
-#include <google/protobuf/stubs/port.h>
-#include <google/protobuf/stubs/stringpiece.h>
-
-#ifndef PROTOBUF_USE_EXCEPTIONS
-#if defined(_MSC_VER) && defined(_CPPUNWIND)
-  #define PROTOBUF_USE_EXCEPTIONS 1
-#elif defined(__EXCEPTIONS)
-  #define PROTOBUF_USE_EXCEPTIONS 1
-#else
-  #define PROTOBUF_USE_EXCEPTIONS 0
-#endif
-#endif
-
-#if PROTOBUF_USE_EXCEPTIONS
-#include <exception>
-#endif
-#if defined(__APPLE__)
-#include <TargetConditionals.h>  // for TARGET_OS_IPHONE
-#endif
-
-#if defined(__ANDROID__) || defined(GOOGLE_PROTOBUF_OS_ANDROID) || (defined(TARGET_OS_IPHONE) && TARGET_OS_IPHONE) || defined(GOOGLE_PROTOBUF_OS_IPHONE)
-#include <pthread.h>
-#endif
-
-#include <google/protobuf/port_def.inc>
-
-namespace std {}
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-// Some of these constants are macros rather than const ints so that they can
-// be used in #if directives.
-
-// The current version, represented as a single integer to make comparison
-// easier:  major * 10^6 + minor * 10^3 + micro
-#define GOOGLE_PROTOBUF_VERSION 3014000
-
-// A suffix string for alpha, beta or rc releases. Empty for stable releases.
-#define GOOGLE_PROTOBUF_VERSION_SUFFIX ""
-
-// The minimum header version which works with the current version of
-// the library.  This constant should only be used by protoc's C++ code
-// generator.
-static const int kMinHeaderVersionForLibrary = 3014000;
-
-// The minimum protoc version which works with the current version of the
-// headers.
-#define GOOGLE_PROTOBUF_MIN_PROTOC_VERSION 3014000
-
-// The minimum header version which works with the current version of
-// protoc.  This constant should only be used in VerifyVersion().
-static const int kMinHeaderVersionForProtoc = 3014000;
-
-// Verifies that the headers and libraries are compatible.  Use the macro
-// below to call this.
-void PROTOBUF_EXPORT VerifyVersion(int headerVersion, int minLibraryVersion,
-                                   const char* filename);
-
-// Converts a numeric version number to a string.
-std::string PROTOBUF_EXPORT VersionString(int version);
-
-}  // namespace internal
-
-// Place this macro in your main() function (or somewhere before you attempt
-// to use the protobuf library) to verify that the version you link against
-// matches the headers you compiled against.  If a version mismatch is
-// detected, the process will abort.
-#define GOOGLE_PROTOBUF_VERIFY_VERSION                                    \
-  ::google::protobuf::internal::VerifyVersion(                            \
-    GOOGLE_PROTOBUF_VERSION, GOOGLE_PROTOBUF_MIN_LIBRARY_VERSION,         \
-    __FILE__)
-
-
-// ===================================================================
-// from google3/util/utf8/public/unilib.h
-
-class StringPiece;
-namespace internal {
-
-// Checks if the buffer contains structurally-valid UTF-8.  Implemented in
-// structurally_valid.cc.
-PROTOBUF_EXPORT bool IsStructurallyValidUTF8(const char* buf, int len);
-
-inline bool IsStructurallyValidUTF8(StringPiece str) {
-  return IsStructurallyValidUTF8(str.data(), static_cast<int>(str.length()));
-}
-
-// Returns initial number of bytes of structurally valid UTF-8.
-PROTOBUF_EXPORT int UTF8SpnStructurallyValid(StringPiece str);
-
-// Coerce UTF-8 byte string in src_str to be
-// a structurally-valid equal-length string by selectively
-// overwriting illegal bytes with replace_char (typically ' ' or '?').
-// replace_char must be legal printable 7-bit Ascii 0x20..0x7e.
-// src_str is read-only.
-//
-// Returns pointer to output buffer, src_str.data() if no changes were made,
-//  or idst if some bytes were changed. idst is allocated by the caller
-//  and must be at least as big as src_str
-//
-// Optimized for: all structurally valid and no byte copying is done.
-//
-PROTOBUF_EXPORT char* UTF8CoerceToStructurallyValid(StringPiece str, char* dst,
-                                                    char replace_char);
-
-}  // namespace internal
-
-// This lives in message_lite.h now, but we leave this here for any users that
-// #include common.h and not message_lite.h.
-PROTOBUF_EXPORT void ShutdownProtobufLibrary();
-
-namespace internal {
-
-// Strongly references the given variable such that the linker will be forced
-// to pull in this variable's translation unit.
-template <typename T>
-void StrongReference(const T& var) {
-  auto volatile unused = &var;
-  (void)&unused;  // Use address to avoid an extra load of "unused".
-}
-
-}  // namespace internal
-
-#if PROTOBUF_USE_EXCEPTIONS
-class FatalException : public std::exception {
- public:
-  FatalException(const char* filename, int line, const std::string& message)
-      : filename_(filename), line_(line), message_(message) {}
-  virtual ~FatalException() throw();
-
-  virtual const char* what() const throw();
-
-  const char* filename() const { return filename_; }
-  int line() const { return line_; }
-  const std::string& message() const { return message_; }
-
- private:
-  const char* filename_;
-  const int line_;
-  const std::string message_;
-};
-#endif
-
-// This is at the end of the file instead of the beginning to work around a bug
-// in some versions of MSVC.
-using std::string;
-
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_COMMON_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/hash.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/hash.h
deleted file mode 100644
index a7ec0680..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/hash.h
+++ /dev/null
@@ -1,114 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// Author: kenton@google.com (Kenton Varda)
-
-#ifndef GOOGLE_PROTOBUF_STUBS_HASH_H__
-#define GOOGLE_PROTOBUF_STUBS_HASH_H__
-
-#include <cstring>
-#include <string>
-#include <unordered_map>
-#include <unordered_set>
-
-# define GOOGLE_PROTOBUF_HASH_NAMESPACE_DECLARATION_START \
-  namespace google {                                      \
-  namespace protobuf {
-# define GOOGLE_PROTOBUF_HASH_NAMESPACE_DECLARATION_END }}
-
-namespace google {
-namespace protobuf {
-
-template <typename Key>
-struct hash : public std::hash<Key> {};
-
-template <typename Key>
-struct hash<const Key*> {
-  inline size_t operator()(const Key* key) const {
-    return reinterpret_cast<size_t>(key);
-  }
-};
-
-// Unlike the old SGI version, the TR1 "hash" does not special-case char*.  So,
-// we go ahead and provide our own implementation.
-template <>
-struct hash<const char*> {
-  inline size_t operator()(const char* str) const {
-    size_t result = 0;
-    for (; *str != '\0'; str++) {
-      result = 5 * result + static_cast<size_t>(*str);
-    }
-    return result;
-  }
-};
-
-template<>
-struct hash<bool> {
-  size_t operator()(bool x) const {
-    return static_cast<size_t>(x);
-  }
-};
-
-template <>
-struct hash<std::string> {
-  inline size_t operator()(const std::string& key) const {
-    return hash<const char*>()(key.c_str());
-  }
-
-  static const size_t bucket_size = 4;
-  static const size_t min_buckets = 8;
-  inline bool operator()(const std::string& a, const std::string& b) const {
-    return a < b;
-  }
-};
-
-template <typename First, typename Second>
-struct hash<std::pair<First, Second> > {
-  inline size_t operator()(const std::pair<First, Second>& key) const {
-    size_t first_hash = hash<First>()(key.first);
-    size_t second_hash = hash<Second>()(key.second);
-
-    // FIXME(kenton):  What is the best way to compute this hash?  I have
-    // no idea!  This seems a bit better than an XOR.
-    return first_hash * ((1 << 16) - 1) + second_hash;
-  }
-
-  static const size_t bucket_size = 4;
-  static const size_t min_buckets = 8;
-  inline bool operator()(const std::pair<First, Second>& a,
-                           const std::pair<First, Second>& b) const {
-    return a < b;
-  }
-};
-
-}  // namespace protobuf
-}  // namespace google
-
-#endif  // GOOGLE_PROTOBUF_STUBS_HASH_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/int128.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/int128.h
deleted file mode 100644
index dc70d96e..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/int128.h
+++ /dev/null
@@ -1,387 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-#ifndef GOOGLE_PROTOBUF_STUBS_INT128_H_
-#define GOOGLE_PROTOBUF_STUBS_INT128_H_
-
-#include <google/protobuf/stubs/common.h>
-
-#include <iosfwd>
-
-#include <google/protobuf/port_def.inc>
-
-namespace google {
-namespace protobuf {
-
-struct uint128_pod;
-
-// TODO(xiaofeng): Define GOOGLE_PROTOBUF_HAS_CONSTEXPR when constexpr is
-// available.
-#ifdef GOOGLE_PROTOBUF_HAS_CONSTEXPR
-# define UINT128_CONSTEXPR constexpr
-#else
-# define UINT128_CONSTEXPR
-#endif
-
-// An unsigned 128-bit integer type. Thread-compatible.
-class PROTOBUF_EXPORT uint128 {
- public:
-  UINT128_CONSTEXPR uint128();  // Sets to 0, but don't trust on this behavior.
-  UINT128_CONSTEXPR uint128(uint64 top, uint64 bottom);
-#ifndef SWIG
-  UINT128_CONSTEXPR uint128(int bottom);
-  UINT128_CONSTEXPR uint128(uint32 bottom);   // Top 96 bits = 0
-#endif
-  UINT128_CONSTEXPR uint128(uint64 bottom);   // hi_ = 0
-  UINT128_CONSTEXPR uint128(const uint128_pod &val);
-
-  // Trivial copy constructor, assignment operator and destructor.
-
-  void Initialize(uint64 top, uint64 bottom);
-
-  // Arithmetic operators.
-  uint128& operator+=(const uint128& b);
-  uint128& operator-=(const uint128& b);
-  uint128& operator*=(const uint128& b);
-  // Long division/modulo for uint128.
-  uint128& operator/=(const uint128& b);
-  uint128& operator%=(const uint128& b);
-  uint128 operator++(int);
-  uint128 operator--(int);
-  uint128& operator<<=(int);
-  uint128& operator>>=(int);
-  uint128& operator&=(const uint128& b);
-  uint128& operator|=(const uint128& b);
-  uint128& operator^=(const uint128& b);
-  uint128& operator++();
-  uint128& operator--();
-
-  friend uint64 Uint128Low64(const uint128& v);
-  friend uint64 Uint128High64(const uint128& v);
-
-  // We add "std::" to avoid including all of port.h.
-  PROTOBUF_EXPORT friend std::ostream& operator<<(std::ostream& o,
-                                                  const uint128& b);
-
- private:
-  static void DivModImpl(uint128 dividend, uint128 divisor,
-                         uint128* quotient_ret, uint128* remainder_ret);
-
-  // Little-endian memory order optimizations can benefit from
-  // having lo_ first, hi_ last.
-  // See util/endian/endian.h and Load128/Store128 for storing a uint128.
-  uint64        lo_;
-  uint64        hi_;
-
-  // Not implemented, just declared for catching automatic type conversions.
-  uint128(uint8);
-  uint128(uint16);
-  uint128(float v);
-  uint128(double v);
-};
-
-// This is a POD form of uint128 which can be used for static variables which
-// need to be operated on as uint128.
-struct uint128_pod {
-  // Note: The ordering of fields is different than 'class uint128' but the
-  // same as its 2-arg constructor.  This enables more obvious initialization
-  // of static instances, which is the primary reason for this struct in the
-  // first place.  This does not seem to defeat any optimizations wrt
-  // operations involving this struct.
-  uint64 hi;
-  uint64 lo;
-};
-
-PROTOBUF_EXPORT extern const uint128_pod kuint128max;
-
-// allow uint128 to be logged
-PROTOBUF_EXPORT extern std::ostream& operator<<(std::ostream& o,
-                                                const uint128& b);
-
-// Methods to access low and high pieces of 128-bit value.
-// Defined externally from uint128 to facilitate conversion
-// to native 128-bit types when compilers support them.
-inline uint64 Uint128Low64(const uint128& v) { return v.lo_; }
-inline uint64 Uint128High64(const uint128& v) { return v.hi_; }
-
-// TODO: perhaps it would be nice to have int128, a signed 128-bit type?
-
-// --------------------------------------------------------------------------
-//                      Implementation details follow
-// --------------------------------------------------------------------------
-inline bool operator==(const uint128& lhs, const uint128& rhs) {
-  return (Uint128Low64(lhs) == Uint128Low64(rhs) &&
-          Uint128High64(lhs) == Uint128High64(rhs));
-}
-inline bool operator!=(const uint128& lhs, const uint128& rhs) {
-  return !(lhs == rhs);
-}
-
-inline UINT128_CONSTEXPR uint128::uint128() : lo_(0), hi_(0) {}
-inline UINT128_CONSTEXPR uint128::uint128(uint64 top, uint64 bottom)
-    : lo_(bottom), hi_(top) {}
-inline UINT128_CONSTEXPR uint128::uint128(const uint128_pod& v)
-    : lo_(v.lo), hi_(v.hi) {}
-inline UINT128_CONSTEXPR uint128::uint128(uint64 bottom)
-    : lo_(bottom), hi_(0) {}
-#ifndef SWIG
-inline UINT128_CONSTEXPR uint128::uint128(uint32 bottom)
-    : lo_(bottom), hi_(0) {}
-inline UINT128_CONSTEXPR uint128::uint128(int bottom)
-    : lo_(bottom), hi_(static_cast<int64>((bottom < 0) ? -1 : 0)) {}
-#endif
-
-#undef UINT128_CONSTEXPR
-
-inline void uint128::Initialize(uint64 top, uint64 bottom) {
-  hi_ = top;
-  lo_ = bottom;
-}
-
-// Comparison operators.
-
-#define CMP128(op)                                                \
-inline bool operator op(const uint128& lhs, const uint128& rhs) { \
-  return (Uint128High64(lhs) == Uint128High64(rhs)) ?             \
-      (Uint128Low64(lhs) op Uint128Low64(rhs)) :                  \
-      (Uint128High64(lhs) op Uint128High64(rhs));                 \
-}
-
-CMP128(<)
-CMP128(>)
-CMP128(>=)
-CMP128(<=)
-
-#undef CMP128
-
-// Unary operators
-
-inline uint128 operator-(const uint128& val) {
-  const uint64 hi_flip = ~Uint128High64(val);
-  const uint64 lo_flip = ~Uint128Low64(val);
-  const uint64 lo_add = lo_flip + 1;
-  if (lo_add < lo_flip) {
-    return uint128(hi_flip + 1, lo_add);
-  }
-  return uint128(hi_flip, lo_add);
-}
-
-inline bool operator!(const uint128& val) {
-  return !Uint128High64(val) && !Uint128Low64(val);
-}
-
-// Logical operators.
-
-inline uint128 operator~(const uint128& val) {
-  return uint128(~Uint128High64(val), ~Uint128Low64(val));
-}
-
-#define LOGIC128(op)                                                 \
-inline uint128 operator op(const uint128& lhs, const uint128& rhs) { \
-  return uint128(Uint128High64(lhs) op Uint128High64(rhs),           \
-                 Uint128Low64(lhs) op Uint128Low64(rhs));            \
-}
-
-LOGIC128(|)
-LOGIC128(&)
-LOGIC128(^)
-
-#undef LOGIC128
-
-#define LOGICASSIGN128(op)                                   \
-inline uint128& uint128::operator op(const uint128& other) { \
-  hi_ op other.hi_;                                          \
-  lo_ op other.lo_;                                          \
-  return *this;                                              \
-}
-
-LOGICASSIGN128(|=)
-LOGICASSIGN128(&=)
-LOGICASSIGN128(^=)
-
-#undef LOGICASSIGN128
-
-// Shift operators.
-
-inline uint128 operator<<(const uint128& val, int amount) {
-  // uint64 shifts of >= 64 are undefined, so we will need some special-casing.
-  if (amount < 64) {
-    if (amount == 0) {
-      return val;
-    }
-    uint64 new_hi = (Uint128High64(val) << amount) |
-                    (Uint128Low64(val) >> (64 - amount));
-    uint64 new_lo = Uint128Low64(val) << amount;
-    return uint128(new_hi, new_lo);
-  } else if (amount < 128) {
-    return uint128(Uint128Low64(val) << (amount - 64), 0);
-  } else {
-    return uint128(0, 0);
-  }
-}
-
-inline uint128 operator>>(const uint128& val, int amount) {
-  // uint64 shifts of >= 64 are undefined, so we will need some special-casing.
-  if (amount < 64) {
-    if (amount == 0) {
-      return val;
-    }
-    uint64 new_hi = Uint128High64(val) >> amount;
-    uint64 new_lo = (Uint128Low64(val) >> amount) |
-                    (Uint128High64(val) << (64 - amount));
-    return uint128(new_hi, new_lo);
-  } else if (amount < 128) {
-    return uint128(0, Uint128High64(val) >> (amount - 64));
-  } else {
-    return uint128(0, 0);
-  }
-}
-
-inline uint128& uint128::operator<<=(int amount) {
-  // uint64 shifts of >= 64 are undefined, so we will need some special-casing.
-  if (amount < 64) {
-    if (amount != 0) {
-      hi_ = (hi_ << amount) | (lo_ >> (64 - amount));
-      lo_ = lo_ << amount;
-    }
-  } else if (amount < 128) {
-    hi_ = lo_ << (amount - 64);
-    lo_ = 0;
-  } else {
-    hi_ = 0;
-    lo_ = 0;
-  }
-  return *this;
-}
-
-inline uint128& uint128::operator>>=(int amount) {
-  // uint64 shifts of >= 64 are undefined, so we will need some special-casing.
-  if (amount < 64) {
-    if (amount != 0) {
-      lo_ = (lo_ >> amount) | (hi_ << (64 - amount));
-      hi_ = hi_ >> amount;
-    }
-  } else if (amount < 128) {
-    lo_ = hi_ >> (amount - 64);
-    hi_ = 0;
-  } else {
-    lo_ = 0;
-    hi_ = 0;
-  }
-  return *this;
-}
-
-inline uint128 operator+(const uint128& lhs, const uint128& rhs) {
-  return uint128(lhs) += rhs;
-}
-
-inline uint128 operator-(const uint128& lhs, const uint128& rhs) {
-  return uint128(lhs) -= rhs;
-}
-
-inline uint128 operator*(const uint128& lhs, const uint128& rhs) {
-  return uint128(lhs) *= rhs;
-}
-
-inline uint128 operator/(const uint128& lhs, const uint128& rhs) {
-  return uint128(lhs) /= rhs;
-}
-
-inline uint128 operator%(const uint128& lhs, const uint128& rhs) {
-  return uint128(lhs) %= rhs;
-}
-
-inline uint128& uint128::operator+=(const uint128& b) {
-  hi_ += b.hi_;
-  uint64 lolo = lo_ + b.lo_;
-  if (lolo < lo_)
-    ++hi_;
-  lo_ = lolo;
-  return *this;
-}
-
-inline uint128& uint128::operator-=(const uint128& b) {
-  hi_ -= b.hi_;
-  if (b.lo_ > lo_)
-    --hi_;
-  lo_ -= b.lo_;
-  return *this;
-}
-
-inline uint128& uint128::operator*=(const uint128& b) {
-  uint64 a96 = hi_ >> 32;
-  uint64 a64 = hi_ & 0xffffffffu;
-  uint64 a32 = lo_ >> 32;
-  uint64 a00 = lo_ & 0xffffffffu;
-  uint64 b96 = b.hi_ >> 32;
-  uint64 b64 = b.hi_ & 0xffffffffu;
-  uint64 b32 = b.lo_ >> 32;
-  uint64 b00 = b.lo_ & 0xffffffffu;
-  // multiply [a96 .. a00] x [b96 .. b00]
-  // terms higher than c96 disappear off the high side
-  // terms c96 and c64 are safe to ignore carry bit
-  uint64 c96 = a96 * b00 + a64 * b32 + a32 * b64 + a00 * b96;
-  uint64 c64 = a64 * b00 + a32 * b32 + a00 * b64;
-  this->hi_ = (c96 << 32) + c64;
-  this->lo_ = 0;
-  // add terms after this one at a time to capture carry
-  *this += uint128(a32 * b00) << 32;
-  *this += uint128(a00 * b32) << 32;
-  *this += a00 * b00;
-  return *this;
-}
-
-inline uint128 uint128::operator++(int) {
-  uint128 tmp(*this);
-  *this += 1;
-  return tmp;
-}
-
-inline uint128 uint128::operator--(int) {
-  uint128 tmp(*this);
-  *this -= 1;
-  return tmp;
-}
-
-inline uint128& uint128::operator++() {
-  *this += 1;
-  return *this;
-}
-
-inline uint128& uint128::operator--() {
-  *this -= 1;
-  return *this;
-}
-
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_STUBS_INT128_H_
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/logging.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/logging.h
deleted file mode 100644
index f37048d6..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/logging.h
+++ /dev/null
@@ -1,241 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_STUBS_LOGGING_H_
-#define GOOGLE_PROTOBUF_STUBS_LOGGING_H_
-
-#include <google/protobuf/stubs/macros.h>
-#include <google/protobuf/stubs/port.h>
-
-#include <google/protobuf/port_def.inc>
-
-// ===================================================================
-// emulates google3/base/logging.h
-
-namespace google {
-namespace protobuf {
-
-enum LogLevel {
-  LOGLEVEL_INFO,     // Informational.  This is never actually used by
-                     // libprotobuf.
-  LOGLEVEL_WARNING,  // Warns about issues that, although not technically a
-                     // problem now, could cause problems in the future.  For
-                     // example, a // warning will be printed when parsing a
-                     // message that is near the message size limit.
-  LOGLEVEL_ERROR,    // An error occurred which should never happen during
-                     // normal use.
-  LOGLEVEL_FATAL,    // An error occurred from which the library cannot
-                     // recover.  This usually indicates a programming error
-                     // in the code which calls the library, especially when
-                     // compiled in debug mode.
-
-#ifdef NDEBUG
-  LOGLEVEL_DFATAL = LOGLEVEL_ERROR
-#else
-  LOGLEVEL_DFATAL = LOGLEVEL_FATAL
-#endif
-};
-
-class StringPiece;
-namespace util {
-class Status;
-}
-class uint128;
-namespace internal {
-
-class LogFinisher;
-
-class PROTOBUF_EXPORT LogMessage {
- public:
-  LogMessage(LogLevel level, const char* filename, int line);
-  ~LogMessage();
-
-  LogMessage& operator<<(const std::string& value);
-  LogMessage& operator<<(const char* value);
-  LogMessage& operator<<(char value);
-  LogMessage& operator<<(int value);
-  LogMessage& operator<<(uint value);
-  LogMessage& operator<<(long value);
-  LogMessage& operator<<(unsigned long value);
-  LogMessage& operator<<(long long value);
-  LogMessage& operator<<(unsigned long long value);
-  LogMessage& operator<<(double value);
-  LogMessage& operator<<(void* value);
-  LogMessage& operator<<(const StringPiece& value);
-  LogMessage& operator<<(const util::Status& status);
-  LogMessage& operator<<(const uint128& value);
-
- private:
-  friend class LogFinisher;
-  void Finish();
-
-  LogLevel level_;
-  const char* filename_;
-  int line_;
-  std::string message_;
-};
-
-// Used to make the entire "LOG(BLAH) << etc." expression have a void return
-// type and print a newline after each message.
-class PROTOBUF_EXPORT LogFinisher {
- public:
-  void operator=(LogMessage& other);
-};
-
-template<typename T>
-bool IsOk(T status) { return status.ok(); }
-template<>
-inline bool IsOk(bool status) { return status; }
-
-}  // namespace internal
-
-// Undef everything in case we're being mixed with some other Google library
-// which already defined them itself.  Presumably all Google libraries will
-// support the same syntax for these so it should not be a big deal if they
-// end up using our definitions instead.
-#undef GOOGLE_LOG
-#undef GOOGLE_LOG_IF
-
-#undef GOOGLE_CHECK
-#undef GOOGLE_CHECK_OK
-#undef GOOGLE_CHECK_EQ
-#undef GOOGLE_CHECK_NE
-#undef GOOGLE_CHECK_LT
-#undef GOOGLE_CHECK_LE
-#undef GOOGLE_CHECK_GT
-#undef GOOGLE_CHECK_GE
-#undef GOOGLE_CHECK_NOTNULL
-
-#undef GOOGLE_DLOG
-#undef GOOGLE_DCHECK
-#undef GOOGLE_DCHECK_OK
-#undef GOOGLE_DCHECK_EQ
-#undef GOOGLE_DCHECK_NE
-#undef GOOGLE_DCHECK_LT
-#undef GOOGLE_DCHECK_LE
-#undef GOOGLE_DCHECK_GT
-#undef GOOGLE_DCHECK_GE
-
-#define GOOGLE_LOG(LEVEL)                          \
-  ::google::protobuf::internal::LogFinisher() = \
-      ::google::protobuf::internal::LogMessage( \
-          ::google::protobuf::LOGLEVEL_##LEVEL, __FILE__, __LINE__)
-#define GOOGLE_LOG_IF(LEVEL, CONDITION) \
-  !(CONDITION) ? (void)0 : GOOGLE_LOG(LEVEL)
-
-#define GOOGLE_CHECK(EXPRESSION) \
-  GOOGLE_LOG_IF(FATAL, !(EXPRESSION)) << "CHECK failed: " #EXPRESSION ": "
-#define GOOGLE_CHECK_OK(A) GOOGLE_CHECK(::google::protobuf::internal::IsOk(A))
-#define GOOGLE_CHECK_EQ(A, B) GOOGLE_CHECK((A) == (B))
-#define GOOGLE_CHECK_NE(A, B) GOOGLE_CHECK((A) != (B))
-#define GOOGLE_CHECK_LT(A, B) GOOGLE_CHECK((A) <  (B))
-#define GOOGLE_CHECK_LE(A, B) GOOGLE_CHECK((A) <= (B))
-#define GOOGLE_CHECK_GT(A, B) GOOGLE_CHECK((A) >  (B))
-#define GOOGLE_CHECK_GE(A, B) GOOGLE_CHECK((A) >= (B))
-
-namespace internal {
-template<typename T>
-T* CheckNotNull(const char* /* file */, int /* line */,
-                const char* name, T* val) {
-  if (val == nullptr) {
-    GOOGLE_LOG(FATAL) << name;
-  }
-  return val;
-}
-}  // namespace internal
-#define GOOGLE_CHECK_NOTNULL(A)               \
-  ::google::protobuf::internal::CheckNotNull( \
-      __FILE__, __LINE__, "'" #A "' must not be nullptr", (A))
-
-#ifdef NDEBUG
-
-#define GOOGLE_DLOG(LEVEL) GOOGLE_LOG_IF(LEVEL, false)
-
-#define GOOGLE_DCHECK(EXPRESSION) while(false) GOOGLE_CHECK(EXPRESSION)
-#define GOOGLE_DCHECK_OK(E) GOOGLE_DCHECK(::google::protobuf::internal::IsOk(E))
-#define GOOGLE_DCHECK_EQ(A, B) GOOGLE_DCHECK((A) == (B))
-#define GOOGLE_DCHECK_NE(A, B) GOOGLE_DCHECK((A) != (B))
-#define GOOGLE_DCHECK_LT(A, B) GOOGLE_DCHECK((A) <  (B))
-#define GOOGLE_DCHECK_LE(A, B) GOOGLE_DCHECK((A) <= (B))
-#define GOOGLE_DCHECK_GT(A, B) GOOGLE_DCHECK((A) >  (B))
-#define GOOGLE_DCHECK_GE(A, B) GOOGLE_DCHECK((A) >= (B))
-
-#else  // NDEBUG
-
-#define GOOGLE_DLOG GOOGLE_LOG
-
-#define GOOGLE_DCHECK    GOOGLE_CHECK
-#define GOOGLE_DCHECK_OK GOOGLE_CHECK_OK
-#define GOOGLE_DCHECK_EQ GOOGLE_CHECK_EQ
-#define GOOGLE_DCHECK_NE GOOGLE_CHECK_NE
-#define GOOGLE_DCHECK_LT GOOGLE_CHECK_LT
-#define GOOGLE_DCHECK_LE GOOGLE_CHECK_LE
-#define GOOGLE_DCHECK_GT GOOGLE_CHECK_GT
-#define GOOGLE_DCHECK_GE GOOGLE_CHECK_GE
-
-#endif  // !NDEBUG
-
-typedef void LogHandler(LogLevel level, const char* filename, int line,
-                        const std::string& message);
-
-// The protobuf library sometimes writes warning and error messages to
-// stderr.  These messages are primarily useful for developers, but may
-// also help end users figure out a problem.  If you would prefer that
-// these messages be sent somewhere other than stderr, call SetLogHandler()
-// to set your own handler.  This returns the old handler.  Set the handler
-// to nullptr to ignore log messages (but see also LogSilencer, below).
-//
-// Obviously, SetLogHandler is not thread-safe.  You should only call it
-// at initialization time, and probably not from library code.  If you
-// simply want to suppress log messages temporarily (e.g. because you
-// have some code that tends to trigger them frequently and you know
-// the warnings are not important to you), use the LogSilencer class
-// below.
-PROTOBUF_EXPORT LogHandler* SetLogHandler(LogHandler* new_func);
-
-// Create a LogSilencer if you want to temporarily suppress all log
-// messages.  As long as any LogSilencer objects exist, non-fatal
-// log messages will be discarded (the current LogHandler will *not*
-// be called).  Constructing a LogSilencer is thread-safe.  You may
-// accidentally suppress log messages occurring in another thread, but
-// since messages are generally for debugging purposes only, this isn't
-// a big deal.  If you want to intercept log messages, use SetLogHandler().
-class PROTOBUF_EXPORT LogSilencer {
- public:
-  LogSilencer();
-  ~LogSilencer();
-};
-
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_STUBS_LOGGING_H_
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/macros.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/macros.h
deleted file mode 100644
index c556d022..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/macros.h
+++ /dev/null
@@ -1,120 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_MACROS_H__
-#define GOOGLE_PROTOBUF_MACROS_H__
-
-#include <google/protobuf/stubs/port.h>
-
-namespace google {
-namespace protobuf {
-
-#undef GOOGLE_DISALLOW_EVIL_CONSTRUCTORS
-#define GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(TypeName)    \
-  TypeName(const TypeName&);                           \
-  void operator=(const TypeName&)
-
-#undef GOOGLE_DISALLOW_IMPLICIT_CONSTRUCTORS
-#define GOOGLE_DISALLOW_IMPLICIT_CONSTRUCTORS(TypeName) \
-  TypeName();                                           \
-  TypeName(const TypeName&);                            \
-  void operator=(const TypeName&)
-
-// ===================================================================
-// from google3/base/basictypes.h
-
-// The GOOGLE_ARRAYSIZE(arr) macro returns the # of elements in an array arr.
-// The expression is a compile-time constant, and therefore can be
-// used in defining new arrays, for example.
-//
-// GOOGLE_ARRAYSIZE catches a few type errors.  If you see a compiler error
-//
-//   "warning: division by zero in ..."
-//
-// when using GOOGLE_ARRAYSIZE, you are (wrongfully) giving it a pointer.
-// You should only use GOOGLE_ARRAYSIZE on statically allocated arrays.
-//
-// The following comments are on the implementation details, and can
-// be ignored by the users.
-//
-// ARRAYSIZE(arr) works by inspecting sizeof(arr) (the # of bytes in
-// the array) and sizeof(*(arr)) (the # of bytes in one array
-// element).  If the former is divisible by the latter, perhaps arr is
-// indeed an array, in which case the division result is the # of
-// elements in the array.  Otherwise, arr cannot possibly be an array,
-// and we generate a compiler error to prevent the code from
-// compiling.
-//
-// Since the size of bool is implementation-defined, we need to cast
-// !(sizeof(a) & sizeof(*(a))) to size_t in order to ensure the final
-// result has type size_t.
-//
-// This macro is not perfect as it wrongfully accepts certain
-// pointers, namely where the pointer size is divisible by the pointee
-// size.  Since all our code has to go through a 32-bit compiler,
-// where a pointer is 4 bytes, this means all pointers to a type whose
-// size is 3 or greater than 4 will be (righteously) rejected.
-//
-// Kudos to Jorg Brown for this simple and elegant implementation.
-
-#undef GOOGLE_ARRAYSIZE
-#define GOOGLE_ARRAYSIZE(a) \
-  ((sizeof(a) / sizeof(*(a))) / \
-   static_cast<size_t>(!(sizeof(a) % sizeof(*(a)))))
-
-// The COMPILE_ASSERT macro can be used to verify that a compile time
-// expression is true. For example, you could use it to verify the
-// size of a static array:
-//
-//   COMPILE_ASSERT(ARRAYSIZE(content_type_names) == CONTENT_NUM_TYPES,
-//                  content_type_names_incorrect_size);
-//
-// or to make sure a struct is smaller than a certain size:
-//
-//   COMPILE_ASSERT(sizeof(foo) < 128, foo_too_large);
-//
-// The second argument to the macro is the name of the variable. If
-// the expression is false, most compilers will issue a warning/error
-// containing the name of the variable.
-
-namespace internal {
-
-template <bool>
-struct CompileAssert {
-};
-
-}  // namespace internal
-
-#define GOOGLE_COMPILE_ASSERT(expr, msg) static_assert(expr, #msg)
-
-}  // namespace protobuf
-}  // namespace google
-
-#endif  // GOOGLE_PROTOBUF_MACROS_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/map_util.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/map_util.h
deleted file mode 100644
index 24e098ad..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/map_util.h
+++ /dev/null
@@ -1,769 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2014 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// from google3/util/gtl/map_util.h
-// Author: Anton Carver
-
-#ifndef GOOGLE_PROTOBUF_STUBS_MAP_UTIL_H__
-#define GOOGLE_PROTOBUF_STUBS_MAP_UTIL_H__
-
-#include <stddef.h>
-#include <iterator>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <google/protobuf/stubs/common.h>
-
-namespace google {
-namespace protobuf {
-namespace internal {
-// Local implementation of RemoveConst to avoid including base/type_traits.h.
-template <class T> struct RemoveConst { typedef T type; };
-template <class T> struct RemoveConst<const T> : RemoveConst<T> {};
-}  // namespace internal
-
-//
-// Find*()
-//
-
-// Returns a const reference to the value associated with the given key if it
-// exists. Crashes otherwise.
-//
-// This is intended as a replacement for operator[] as an rvalue (for reading)
-// when the key is guaranteed to exist.
-//
-// operator[] for lookup is discouraged for several reasons:
-//  * It has a side-effect of inserting missing keys
-//  * It is not thread-safe (even when it is not inserting, it can still
-//      choose to resize the underlying storage)
-//  * It invalidates iterators (when it chooses to resize)
-//  * It default constructs a value object even if it doesn't need to
-//
-// This version assumes the key is printable, and includes it in the fatal log
-// message.
-template <class Collection>
-const typename Collection::value_type::second_type&
-FindOrDie(const Collection& collection,
-          const typename Collection::value_type::first_type& key) {
-  typename Collection::const_iterator it = collection.find(key);
-  GOOGLE_CHECK(it != collection.end()) << "Map key not found: " << key;
-  return it->second;
-}
-
-// Same as above, but returns a non-const reference.
-template <class Collection>
-typename Collection::value_type::second_type&
-FindOrDie(Collection& collection,  // NOLINT
-          const typename Collection::value_type::first_type& key) {
-  typename Collection::iterator it = collection.find(key);
-  GOOGLE_CHECK(it != collection.end()) << "Map key not found: " << key;
-  return it->second;
-}
-
-// Same as FindOrDie above, but doesn't log the key on failure.
-template <class Collection>
-const typename Collection::value_type::second_type&
-FindOrDieNoPrint(const Collection& collection,
-                 const typename Collection::value_type::first_type& key) {
-  typename Collection::const_iterator it = collection.find(key);
-  GOOGLE_CHECK(it != collection.end()) << "Map key not found";
-  return it->second;
-}
-
-// Same as above, but returns a non-const reference.
-template <class Collection>
-typename Collection::value_type::second_type&
-FindOrDieNoPrint(Collection& collection,  // NOLINT
-                 const typename Collection::value_type::first_type& key) {
-  typename Collection::iterator it = collection.find(key);
-  GOOGLE_CHECK(it != collection.end()) << "Map key not found";
-  return it->second;
-}
-
-// Returns a const reference to the value associated with the given key if it
-// exists, otherwise returns a const reference to the provided default value.
-//
-// WARNING: If a temporary object is passed as the default "value,"
-// this function will return a reference to that temporary object,
-// which will be destroyed at the end of the statement. A common
-// example: if you have a map with string values, and you pass a char*
-// as the default "value," either use the returned value immediately
-// or store it in a string (not string&).
-// Details: http://go/findwithdefault
-template <class Collection>
-const typename Collection::value_type::second_type&
-FindWithDefault(const Collection& collection,
-                const typename Collection::value_type::first_type& key,
-                const typename Collection::value_type::second_type& value) {
-  typename Collection::const_iterator it = collection.find(key);
-  if (it == collection.end()) {
-    return value;
-  }
-  return it->second;
-}
-
-// Returns a pointer to the const value associated with the given key if it
-// exists, or nullptr otherwise.
-template <class Collection>
-const typename Collection::value_type::second_type*
-FindOrNull(const Collection& collection,
-           const typename Collection::value_type::first_type& key) {
-  typename Collection::const_iterator it = collection.find(key);
-  if (it == collection.end()) {
-    return 0;
-  }
-  return &it->second;
-}
-
-// Same as above but returns a pointer to the non-const value.
-template <class Collection>
-typename Collection::value_type::second_type*
-FindOrNull(Collection& collection,  // NOLINT
-           const typename Collection::value_type::first_type& key) {
-  typename Collection::iterator it = collection.find(key);
-  if (it == collection.end()) {
-    return 0;
-  }
-  return &it->second;
-}
-
-// Returns the pointer value associated with the given key. If none is found,
-// nullptr is returned. The function is designed to be used with a map of keys to
-// pointers.
-//
-// This function does not distinguish between a missing key and a key mapped
-// to nullptr.
-template <class Collection>
-typename Collection::value_type::second_type
-FindPtrOrNull(const Collection& collection,
-              const typename Collection::value_type::first_type& key) {
-  typename Collection::const_iterator it = collection.find(key);
-  if (it == collection.end()) {
-    return typename Collection::value_type::second_type();
-  }
-  return it->second;
-}
-
-// Same as above, except takes non-const reference to collection.
-//
-// This function is needed for containers that propagate constness to the
-// pointee, such as boost::ptr_map.
-template <class Collection>
-typename Collection::value_type::second_type
-FindPtrOrNull(Collection& collection,  // NOLINT
-              const typename Collection::value_type::first_type& key) {
-  typename Collection::iterator it = collection.find(key);
-  if (it == collection.end()) {
-    return typename Collection::value_type::second_type();
-  }
-  return it->second;
-}
-
-// Finds the pointer value associated with the given key in a map whose values
-// are linked_ptrs. Returns nullptr if key is not found.
-template <class Collection>
-typename Collection::value_type::second_type::element_type*
-FindLinkedPtrOrNull(const Collection& collection,
-                    const typename Collection::value_type::first_type& key) {
-  typename Collection::const_iterator it = collection.find(key);
-  if (it == collection.end()) {
-    return 0;
-  }
-  // Since linked_ptr::get() is a const member returning a non const,
-  // we do not need a version of this function taking a non const collection.
-  return it->second.get();
-}
-
-// Same as above, but dies if the key is not found.
-template <class Collection>
-typename Collection::value_type::second_type::element_type&
-FindLinkedPtrOrDie(const Collection& collection,
-                   const typename Collection::value_type::first_type& key) {
-  typename Collection::const_iterator it = collection.find(key);
-  GOOGLE_CHECK(it != collection.end()) <<  "key not found: " << key;
-  // Since linked_ptr::operator*() is a const member returning a non const,
-  // we do not need a version of this function taking a non const collection.
-  return *it->second;
-}
-
-// Finds the value associated with the given key and copies it to *value (if not
-// nullptr). Returns false if the key was not found, true otherwise.
-template <class Collection, class Key, class Value>
-bool FindCopy(const Collection& collection,
-              const Key& key,
-              Value* const value) {
-  typename Collection::const_iterator it = collection.find(key);
-  if (it == collection.end()) {
-    return false;
-  }
-  if (value) {
-    *value = it->second;
-  }
-  return true;
-}
-
-//
-// Contains*()
-//
-
-// Returns true if and only if the given collection contains the given key.
-template <class Collection, class Key>
-bool ContainsKey(const Collection& collection, const Key& key) {
-  return collection.find(key) != collection.end();
-}
-
-// Returns true if and only if the given collection contains the given key-value
-// pair.
-template <class Collection, class Key, class Value>
-bool ContainsKeyValuePair(const Collection& collection,
-                          const Key& key,
-                          const Value& value) {
-  typedef typename Collection::const_iterator const_iterator;
-  std::pair<const_iterator, const_iterator> range = collection.equal_range(key);
-  for (const_iterator it = range.first; it != range.second; ++it) {
-    if (it->second == value) {
-      return true;
-    }
-  }
-  return false;
-}
-
-//
-// Insert*()
-//
-
-// Inserts the given key-value pair into the collection. Returns true if and
-// only if the key from the given pair didn't previously exist. Otherwise, the
-// value in the map is replaced with the value from the given pair.
-template <class Collection>
-bool InsertOrUpdate(Collection* const collection,
-                    const typename Collection::value_type& vt) {
-  std::pair<typename Collection::iterator, bool> ret = collection->insert(vt);
-  if (!ret.second) {
-    // update
-    ret.first->second = vt.second;
-    return false;
-  }
-  return true;
-}
-
-// Same as above, except that the key and value are passed separately.
-template <class Collection>
-bool InsertOrUpdate(Collection* const collection,
-                    const typename Collection::value_type::first_type& key,
-                    const typename Collection::value_type::second_type& value) {
-  return InsertOrUpdate(
-      collection, typename Collection::value_type(key, value));
-}
-
-// Inserts/updates all the key-value pairs from the range defined by the
-// iterators "first" and "last" into the given collection.
-template <class Collection, class InputIterator>
-void InsertOrUpdateMany(Collection* const collection,
-                        InputIterator first, InputIterator last) {
-  for (; first != last; ++first) {
-    InsertOrUpdate(collection, *first);
-  }
-}
-
-// Change the value associated with a particular key in a map or hash_map
-// of the form map<Key, Value*> which owns the objects pointed to by the
-// value pointers.  If there was an existing value for the key, it is deleted.
-// True indicates an insert took place, false indicates an update + delete.
-template <class Collection>
-bool InsertAndDeleteExisting(
-    Collection* const collection,
-    const typename Collection::value_type::first_type& key,
-    const typename Collection::value_type::second_type& value) {
-  std::pair<typename Collection::iterator, bool> ret =
-      collection->insert(typename Collection::value_type(key, value));
-  if (!ret.second) {
-    delete ret.first->second;
-    ret.first->second = value;
-    return false;
-  }
-  return true;
-}
-
-// Inserts the given key and value into the given collection if and only if the
-// given key did NOT already exist in the collection. If the key previously
-// existed in the collection, the value is not changed. Returns true if the
-// key-value pair was inserted; returns false if the key was already present.
-template <class Collection>
-bool InsertIfNotPresent(Collection* const collection,
-                        const typename Collection::value_type& vt) {
-  return collection->insert(vt).second;
-}
-
-// Same as above except the key and value are passed separately.
-template <class Collection>
-bool InsertIfNotPresent(
-    Collection* const collection,
-    const typename Collection::value_type::first_type& key,
-    const typename Collection::value_type::second_type& value) {
-  return InsertIfNotPresent(
-      collection, typename Collection::value_type(key, value));
-}
-
-// Same as above except dies if the key already exists in the collection.
-template <class Collection>
-void InsertOrDie(Collection* const collection,
-                 const typename Collection::value_type& value) {
-  GOOGLE_CHECK(InsertIfNotPresent(collection, value))
-      << "duplicate value: " << value;
-}
-
-// Same as above except doesn't log the value on error.
-template <class Collection>
-void InsertOrDieNoPrint(Collection* const collection,
-                        const typename Collection::value_type& value) {
-  GOOGLE_CHECK(InsertIfNotPresent(collection, value)) << "duplicate value.";
-}
-
-// Inserts the key-value pair into the collection. Dies if key was already
-// present.
-template <class Collection>
-void InsertOrDie(Collection* const collection,
-                 const typename Collection::value_type::first_type& key,
-                 const typename Collection::value_type::second_type& data) {
-  GOOGLE_CHECK(InsertIfNotPresent(collection, key, data))
-      << "duplicate key: " << key;
-}
-
-// Same as above except doesn't log the key on error.
-template <class Collection>
-void InsertOrDieNoPrint(
-    Collection* const collection,
-    const typename Collection::value_type::first_type& key,
-    const typename Collection::value_type::second_type& data) {
-  GOOGLE_CHECK(InsertIfNotPresent(collection, key, data)) << "duplicate key.";
-}
-
-// Inserts a new key and default-initialized value. Dies if the key was already
-// present. Returns a reference to the value. Example usage:
-//
-// map<int, SomeProto> m;
-// SomeProto& proto = InsertKeyOrDie(&m, 3);
-// proto.set_field("foo");
-template <class Collection>
-typename Collection::value_type::second_type& InsertKeyOrDie(
-    Collection* const collection,
-    const typename Collection::value_type::first_type& key) {
-  typedef typename Collection::value_type value_type;
-  std::pair<typename Collection::iterator, bool> res =
-      collection->insert(value_type(key, typename value_type::second_type()));
-  GOOGLE_CHECK(res.second) << "duplicate key: " << key;
-  return res.first->second;
-}
-
-//
-// Lookup*()
-//
-
-// Looks up a given key and value pair in a collection and inserts the key-value
-// pair if it's not already present. Returns a reference to the value associated
-// with the key.
-template <class Collection>
-typename Collection::value_type::second_type&
-LookupOrInsert(Collection* const collection,
-               const typename Collection::value_type& vt) {
-  return collection->insert(vt).first->second;
-}
-
-// Same as above except the key-value are passed separately.
-template <class Collection>
-typename Collection::value_type::second_type&
-LookupOrInsert(Collection* const collection,
-               const typename Collection::value_type::first_type& key,
-               const typename Collection::value_type::second_type& value) {
-  return LookupOrInsert(
-      collection, typename Collection::value_type(key, value));
-}
-
-// Counts the number of equivalent elements in the given "sequence", and stores
-// the results in "count_map" with element as the key and count as the value.
-//
-// Example:
-//   vector<string> v = {"a", "b", "c", "a", "b"};
-//   map<string, int> m;
-//   AddTokenCounts(v, 1, &m);
-//   assert(m["a"] == 2);
-//   assert(m["b"] == 2);
-//   assert(m["c"] == 1);
-template <typename Sequence, typename Collection>
-void AddTokenCounts(
-    const Sequence& sequence,
-    const typename Collection::value_type::second_type& increment,
-    Collection* const count_map) {
-  for (typename Sequence::const_iterator it = sequence.begin();
-       it != sequence.end(); ++it) {
-    typename Collection::value_type::second_type& value =
-        LookupOrInsert(count_map, *it,
-                       typename Collection::value_type::second_type());
-    value += increment;
-  }
-}
-
-// Returns a reference to the value associated with key. If not found, a value
-// is default constructed on the heap and added to the map.
-//
-// This function is useful for containers of the form map<Key, Value*>, where
-// inserting a new key, value pair involves constructing a new heap-allocated
-// Value, and storing a pointer to that in the collection.
-template <class Collection>
-typename Collection::value_type::second_type&
-LookupOrInsertNew(Collection* const collection,
-                  const typename Collection::value_type::first_type& key) {
-  typedef typename std::iterator_traits<
-    typename Collection::value_type::second_type>::value_type Element;
-  std::pair<typename Collection::iterator, bool> ret =
-      collection->insert(typename Collection::value_type(
-          key,
-          static_cast<typename Collection::value_type::second_type>(nullptr)));
-  if (ret.second) {
-    ret.first->second = new Element();
-  }
-  return ret.first->second;
-}
-
-// Same as above but constructs the value using the single-argument constructor
-// and the given "arg".
-template <class Collection, class Arg>
-typename Collection::value_type::second_type&
-LookupOrInsertNew(Collection* const collection,
-                  const typename Collection::value_type::first_type& key,
-                  const Arg& arg) {
-  typedef typename std::iterator_traits<
-    typename Collection::value_type::second_type>::value_type Element;
-  std::pair<typename Collection::iterator, bool> ret =
-      collection->insert(typename Collection::value_type(
-          key,
-          static_cast<typename Collection::value_type::second_type>(nullptr)));
-  if (ret.second) {
-    ret.first->second = new Element(arg);
-  }
-  return ret.first->second;
-}
-
-// Lookup of linked/shared pointers is used in two scenarios:
-//
-// Use LookupOrInsertNewLinkedPtr if the container owns the elements.
-// In this case it is fine working with the raw pointer as long as it is
-// guaranteed that no other thread can delete/update an accessed element.
-// A mutex will need to lock the container operation as well as the use
-// of the returned elements. Finding an element may be performed using
-// FindLinkedPtr*().
-//
-// Use LookupOrInsertNewSharedPtr if the container does not own the elements
-// for their whole lifetime. This is typically the case when a reader allows
-// parallel updates to the container. In this case a Mutex only needs to lock
-// container operations, but all element operations must be performed on the
-// shared pointer. Finding an element must be performed using FindPtr*() and
-// cannot be done with FindLinkedPtr*() even though it compiles.
-
-// Lookup a key in a map or hash_map whose values are linked_ptrs.  If it is
-// missing, set collection[key].reset(new Value::element_type) and return that.
-// Value::element_type must be default constructable.
-template <class Collection>
-typename Collection::value_type::second_type::element_type*
-LookupOrInsertNewLinkedPtr(
-    Collection* const collection,
-    const typename Collection::value_type::first_type& key) {
-  typedef typename Collection::value_type::second_type Value;
-  std::pair<typename Collection::iterator, bool> ret =
-      collection->insert(typename Collection::value_type(key, Value()));
-  if (ret.second) {
-    ret.first->second.reset(new typename Value::element_type);
-  }
-  return ret.first->second.get();
-}
-
-// A variant of LookupOrInsertNewLinkedPtr where the value is constructed using
-// a single-parameter constructor.  Note: the constructor argument is computed
-// even if it will not be used, so only values cheap to compute should be passed
-// here.  On the other hand it does not matter how expensive the construction of
-// the actual stored value is, as that only occurs if necessary.
-template <class Collection, class Arg>
-typename Collection::value_type::second_type::element_type*
-LookupOrInsertNewLinkedPtr(
-    Collection* const collection,
-    const typename Collection::value_type::first_type& key,
-    const Arg& arg) {
-  typedef typename Collection::value_type::second_type Value;
-  std::pair<typename Collection::iterator, bool> ret =
-      collection->insert(typename Collection::value_type(key, Value()));
-  if (ret.second) {
-    ret.first->second.reset(new typename Value::element_type(arg));
-  }
-  return ret.first->second.get();
-}
-
-// Lookup a key in a map or hash_map whose values are shared_ptrs.  If it is
-// missing, set collection[key].reset(new Value::element_type). Unlike
-// LookupOrInsertNewLinkedPtr, this function returns the shared_ptr instead of
-// the raw pointer. Value::element_type must be default constructable.
-template <class Collection>
-typename Collection::value_type::second_type&
-LookupOrInsertNewSharedPtr(
-    Collection* const collection,
-    const typename Collection::value_type::first_type& key) {
-  typedef typename Collection::value_type::second_type SharedPtr;
-  typedef typename Collection::value_type::second_type::element_type Element;
-  std::pair<typename Collection::iterator, bool> ret =
-      collection->insert(typename Collection::value_type(key, SharedPtr()));
-  if (ret.second) {
-    ret.first->second.reset(new Element());
-  }
-  return ret.first->second;
-}
-
-// A variant of LookupOrInsertNewSharedPtr where the value is constructed using
-// a single-parameter constructor.  Note: the constructor argument is computed
-// even if it will not be used, so only values cheap to compute should be passed
-// here.  On the other hand it does not matter how expensive the construction of
-// the actual stored value is, as that only occurs if necessary.
-template <class Collection, class Arg>
-typename Collection::value_type::second_type&
-LookupOrInsertNewSharedPtr(
-    Collection* const collection,
-    const typename Collection::value_type::first_type& key,
-    const Arg& arg) {
-  typedef typename Collection::value_type::second_type SharedPtr;
-  typedef typename Collection::value_type::second_type::element_type Element;
-  std::pair<typename Collection::iterator, bool> ret =
-      collection->insert(typename Collection::value_type(key, SharedPtr()));
-  if (ret.second) {
-    ret.first->second.reset(new Element(arg));
-  }
-  return ret.first->second;
-}
-
-//
-// Misc Utility Functions
-//
-
-// Updates the value associated with the given key. If the key was not already
-// present, then the key-value pair are inserted and "previous" is unchanged. If
-// the key was already present, the value is updated and "*previous" will
-// contain a copy of the old value.
-//
-// InsertOrReturnExisting has complementary behavior that returns the
-// address of an already existing value, rather than updating it.
-template <class Collection>
-bool UpdateReturnCopy(Collection* const collection,
-                      const typename Collection::value_type::first_type& key,
-                      const typename Collection::value_type::second_type& value,
-                      typename Collection::value_type::second_type* previous) {
-  std::pair<typename Collection::iterator, bool> ret =
-      collection->insert(typename Collection::value_type(key, value));
-  if (!ret.second) {
-    // update
-    if (previous) {
-      *previous = ret.first->second;
-    }
-    ret.first->second = value;
-    return true;
-  }
-  return false;
-}
-
-// Same as above except that the key and value are passed as a pair.
-template <class Collection>
-bool UpdateReturnCopy(Collection* const collection,
-                      const typename Collection::value_type& vt,
-                      typename Collection::value_type::second_type* previous) {
-  std::pair<typename Collection::iterator, bool> ret = collection->insert(vt);
-  if (!ret.second) {
-    // update
-    if (previous) {
-      *previous = ret.first->second;
-    }
-    ret.first->second = vt.second;
-    return true;
-  }
-  return false;
-}
-
-// Tries to insert the given key-value pair into the collection. Returns nullptr if
-// the insert succeeds. Otherwise, returns a pointer to the existing value.
-//
-// This complements UpdateReturnCopy in that it allows to update only after
-// verifying the old value and still insert quickly without having to look up
-// twice. Unlike UpdateReturnCopy this also does not come with the issue of an
-// undefined previous* in case new data was inserted.
-template <class Collection>
-typename Collection::value_type::second_type* InsertOrReturnExisting(
-    Collection* const collection, const typename Collection::value_type& vt) {
-  std::pair<typename Collection::iterator, bool> ret = collection->insert(vt);
-  if (ret.second) {
-    return nullptr;  // Inserted, no existing previous value.
-  } else {
-    return &ret.first->second;  // Return address of already existing value.
-  }
-}
-
-// Same as above, except for explicit key and data.
-template <class Collection>
-typename Collection::value_type::second_type* InsertOrReturnExisting(
-    Collection* const collection,
-    const typename Collection::value_type::first_type& key,
-    const typename Collection::value_type::second_type& data) {
-  return InsertOrReturnExisting(collection,
-                                typename Collection::value_type(key, data));
-}
-
-// Erases the collection item identified by the given key, and returns the value
-// associated with that key. It is assumed that the value (i.e., the
-// mapped_type) is a pointer. Returns nullptr if the key was not found in the
-// collection.
-//
-// Examples:
-//   map<string, MyType*> my_map;
-//
-// One line cleanup:
-//     delete EraseKeyReturnValuePtr(&my_map, "abc");
-//
-// Use returned value:
-//     std::unique_ptr<MyType> value_ptr(
-//         EraseKeyReturnValuePtr(&my_map, "abc"));
-//     if (value_ptr.get())
-//       value_ptr->DoSomething();
-//
-template <class Collection>
-typename Collection::value_type::second_type EraseKeyReturnValuePtr(
-    Collection* const collection,
-    const typename Collection::value_type::first_type& key) {
-  typename Collection::iterator it = collection->find(key);
-  if (it == collection->end()) {
-    return nullptr;
-  }
-  typename Collection::value_type::second_type v = it->second;
-  collection->erase(it);
-  return v;
-}
-
-// Inserts all the keys from map_container into key_container, which must
-// support insert(MapContainer::key_type).
-//
-// Note: any initial contents of the key_container are not cleared.
-template <class MapContainer, class KeyContainer>
-void InsertKeysFromMap(const MapContainer& map_container,
-                       KeyContainer* key_container) {
-  GOOGLE_CHECK(key_container != nullptr);
-  for (typename MapContainer::const_iterator it = map_container.begin();
-       it != map_container.end(); ++it) {
-    key_container->insert(it->first);
-  }
-}
-
-// Appends all the keys from map_container into key_container, which must
-// support push_back(MapContainer::key_type).
-//
-// Note: any initial contents of the key_container are not cleared.
-template <class MapContainer, class KeyContainer>
-void AppendKeysFromMap(const MapContainer& map_container,
-                       KeyContainer* key_container) {
-  GOOGLE_CHECK(key_container != nullptr);
-  for (typename MapContainer::const_iterator it = map_container.begin();
-       it != map_container.end(); ++it) {
-    key_container->push_back(it->first);
-  }
-}
-
-// A more specialized overload of AppendKeysFromMap to optimize reallocations
-// for the common case in which we're appending keys to a vector and hence can
-// (and sometimes should) call reserve() first.
-//
-// (It would be possible to play SFINAE games to call reserve() for any
-// container that supports it, but this seems to get us 99% of what we need
-// without the complexity of a SFINAE-based solution.)
-template <class MapContainer, class KeyType>
-void AppendKeysFromMap(const MapContainer& map_container,
-                       std::vector<KeyType>* key_container) {
-  GOOGLE_CHECK(key_container != nullptr);
-  // We now have the opportunity to call reserve(). Calling reserve() every
-  // time is a bad idea for some use cases: libstdc++'s implementation of
-  // vector<>::reserve() resizes the vector's backing store to exactly the
-  // given size (unless it's already at least that big). Because of this,
-  // the use case that involves appending a lot of small maps (total size
-  // N) one by one to a vector would be O(N^2). But never calling reserve()
-  // loses the opportunity to improve the use case of adding from a large
-  // map to an empty vector (this improves performance by up to 33%). A
-  // number of heuristics are possible; see the discussion in
-  // cl/34081696. Here we use the simplest one.
-  if (key_container->empty()) {
-    key_container->reserve(map_container.size());
-  }
-  for (typename MapContainer::const_iterator it = map_container.begin();
-       it != map_container.end(); ++it) {
-    key_container->push_back(it->first);
-  }
-}
-
-// Inserts all the values from map_container into value_container, which must
-// support push_back(MapContainer::mapped_type).
-//
-// Note: any initial contents of the value_container are not cleared.
-template <class MapContainer, class ValueContainer>
-void AppendValuesFromMap(const MapContainer& map_container,
-                         ValueContainer* value_container) {
-  GOOGLE_CHECK(value_container != nullptr);
-  for (typename MapContainer::const_iterator it = map_container.begin();
-       it != map_container.end(); ++it) {
-    value_container->push_back(it->second);
-  }
-}
-
-// A more specialized overload of AppendValuesFromMap to optimize reallocations
-// for the common case in which we're appending values to a vector and hence
-// can (and sometimes should) call reserve() first.
-//
-// (It would be possible to play SFINAE games to call reserve() for any
-// container that supports it, but this seems to get us 99% of what we need
-// without the complexity of a SFINAE-based solution.)
-template <class MapContainer, class ValueType>
-void AppendValuesFromMap(const MapContainer& map_container,
-                         std::vector<ValueType>* value_container) {
-  GOOGLE_CHECK(value_container != nullptr);
-  // See AppendKeysFromMap for why this is done.
-  if (value_container->empty()) {
-    value_container->reserve(map_container.size());
-  }
-  for (typename MapContainer::const_iterator it = map_container.begin();
-       it != map_container.end(); ++it) {
-    value_container->push_back(it->second);
-  }
-}
-
-}  // namespace protobuf
-}  // namespace google
-
-#endif  // GOOGLE_PROTOBUF_STUBS_MAP_UTIL_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/mutex.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/mutex.h
deleted file mode 100644
index b222ff74..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/mutex.h
+++ /dev/null
@@ -1,186 +0,0 @@
-// Copyright (c) 2006, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_STUBS_MUTEX_H_
-#define GOOGLE_PROTOBUF_STUBS_MUTEX_H_
-
-#include <mutex>
-
-#ifdef GOOGLE_PROTOBUF_SUPPORT_WINDOWS_XP
-
-#include <windows.h>
-
-// GetMessage conflicts with GeneratedMessageReflection::GetMessage().
-#ifdef GetMessage
-#undef GetMessage
-#endif
-
-#endif
-
-#include <google/protobuf/stubs/macros.h>
-
-// Define thread-safety annotations for use below, if we are building with
-// Clang.
-#if defined(__clang__) && !defined(SWIG)
-#define GOOGLE_PROTOBUF_ACQUIRE(...) \
-  __attribute__((acquire_capability(__VA_ARGS__)))
-#define GOOGLE_PROTOBUF_RELEASE(...) \
-  __attribute__((release_capability(__VA_ARGS__)))
-#define GOOGLE_PROTOBUF_CAPABILITY(x) __attribute__((capability(x)))
-#else
-#define GOOGLE_PROTOBUF_ACQUIRE(...)
-#define GOOGLE_PROTOBUF_RELEASE(...)
-#define GOOGLE_PROTOBUF_CAPABILITY(x)
-#endif
-
-#include <google/protobuf/port_def.inc>
-
-// ===================================================================
-// emulates google3/base/mutex.h
-namespace google {
-namespace protobuf {
-namespace internal {
-
-#define GOOGLE_PROTOBUF_LINKER_INITIALIZED
-
-#ifdef GOOGLE_PROTOBUF_SUPPORT_WINDOWS_XP
-
-// This class is a lightweight replacement for std::mutex on Windows platforms.
-// std::mutex does not work on Windows XP SP2 with the latest VC++ libraries,
-// because it utilizes the Concurrency Runtime that is only supported on Windows
-// XP SP3 and above.
-class PROTOBUF_EXPORT CriticalSectionLock {
- public:
-  CriticalSectionLock() { InitializeCriticalSection(&critical_section_); }
-  ~CriticalSectionLock() { DeleteCriticalSection(&critical_section_); }
-  void lock() { EnterCriticalSection(&critical_section_); }
-  void unlock() { LeaveCriticalSection(&critical_section_); }
-
- private:
-  CRITICAL_SECTION critical_section_;
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(CriticalSectionLock);
-};
-
-#endif
-
-// Mutex is a natural type to wrap. As both google and other organization have
-// specialized mutexes. gRPC also provides an injection mechanism for custom
-// mutexes.
-class GOOGLE_PROTOBUF_CAPABILITY("mutex") PROTOBUF_EXPORT WrappedMutex {
- public:
-  WrappedMutex() = default;
-  void Lock() GOOGLE_PROTOBUF_ACQUIRE() { mu_.lock(); }
-  void Unlock() GOOGLE_PROTOBUF_RELEASE() { mu_.unlock(); }
-  // Crash if this Mutex is not held exclusively by this thread.
-  // May fail to crash when it should; will never crash when it should not.
-  void AssertHeld() const {}
-
- private:
-#ifndef GOOGLE_PROTOBUF_SUPPORT_WINDOWS_XP
-  std::mutex mu_;
-#else  // ifndef GOOGLE_PROTOBUF_SUPPORT_WINDOWS_XP
-  CriticalSectionLock mu_;
-#endif  // #ifndef GOOGLE_PROTOBUF_SUPPORT_WINDOWS_XP
-};
-
-using Mutex = WrappedMutex;
-
-// MutexLock(mu) acquires mu when constructed and releases it when destroyed.
-class PROTOBUF_EXPORT MutexLock {
- public:
-  explicit MutexLock(Mutex *mu) : mu_(mu) { this->mu_->Lock(); }
-  ~MutexLock() { this->mu_->Unlock(); }
- private:
-  Mutex *const mu_;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(MutexLock);
-};
-
-// TODO(kenton):  Implement these?  Hard to implement portably.
-typedef MutexLock ReaderMutexLock;
-typedef MutexLock WriterMutexLock;
-
-// MutexLockMaybe is like MutexLock, but is a no-op when mu is nullptr.
-class PROTOBUF_EXPORT MutexLockMaybe {
- public:
-  explicit MutexLockMaybe(Mutex *mu) :
-    mu_(mu) { if (this->mu_ != nullptr) { this->mu_->Lock(); } }
-  ~MutexLockMaybe() { if (this->mu_ != nullptr) { this->mu_->Unlock(); } }
- private:
-  Mutex *const mu_;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(MutexLockMaybe);
-};
-
-#if defined(GOOGLE_PROTOBUF_NO_THREADLOCAL)
-template<typename T>
-class ThreadLocalStorage {
- public:
-  ThreadLocalStorage() {
-    pthread_key_create(&key_, &ThreadLocalStorage::Delete);
-  }
-  ~ThreadLocalStorage() {
-    pthread_key_delete(key_);
-  }
-  T* Get() {
-    T* result = static_cast<T*>(pthread_getspecific(key_));
-    if (result == nullptr) {
-      result = new T();
-      pthread_setspecific(key_, result);
-    }
-    return result;
-  }
- private:
-  static void Delete(void* value) {
-    delete static_cast<T*>(value);
-  }
-  pthread_key_t key_;
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ThreadLocalStorage);
-};
-#endif
-
-}  // namespace internal
-
-// We made these internal so that they would show up as such in the docs,
-// but we don't want to stick "internal::" in front of them everywhere.
-using internal::Mutex;
-using internal::MutexLock;
-using internal::ReaderMutexLock;
-using internal::WriterMutexLock;
-using internal::MutexLockMaybe;
-
-}  // namespace protobuf
-}  // namespace google
-
-#undef GOOGLE_PROTOBUF_ACQUIRE
-#undef GOOGLE_PROTOBUF_RELEASE
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_STUBS_MUTEX_H_
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/once.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/once.h
deleted file mode 100644
index 070d36d1..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/once.h
+++ /dev/null
@@ -1,55 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_STUBS_ONCE_H__
-#define GOOGLE_PROTOBUF_STUBS_ONCE_H__
-
-#include <mutex>
-#include <utility>
-
-#include <google/protobuf/port_def.inc>
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-using once_flag = std::once_flag;
-template <typename... Args>
-void call_once(Args&&... args ) {
-  std::call_once(std::forward<Args>(args)...);
-}
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_STUBS_ONCE_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/platform_macros.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/platform_macros.h
deleted file mode 100644
index ce1b1e36..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/platform_macros.h
+++ /dev/null
@@ -1,134 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2012 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_PLATFORM_MACROS_H_
-#define GOOGLE_PROTOBUF_PLATFORM_MACROS_H_
-
-#define GOOGLE_PROTOBUF_PLATFORM_ERROR \
-#error "Host platform was not detected as supported by protobuf"
-
-// Processor architecture detection.  For more info on what's defined, see:
-//   http://msdn.microsoft.com/en-us/library/b0084kay.aspx
-//   http://www.agner.org/optimize/calling_conventions.pdf
-//   or with gcc, run: "echo | gcc -E -dM -"
-#if defined(_M_X64) || defined(__x86_64__)
-#define GOOGLE_PROTOBUF_ARCH_X64 1
-#define GOOGLE_PROTOBUF_ARCH_64_BIT 1
-#elif defined(_M_IX86) || defined(__i386__)
-#define GOOGLE_PROTOBUF_ARCH_IA32 1
-#define GOOGLE_PROTOBUF_ARCH_32_BIT 1
-#elif defined(__QNX__)
-#define GOOGLE_PROTOBUF_ARCH_ARM_QNX 1
-#define GOOGLE_PROTOBUF_ARCH_32_BIT 1
-#elif defined(_M_ARM) || defined(__ARMEL__)
-#define GOOGLE_PROTOBUF_ARCH_ARM 1
-#define GOOGLE_PROTOBUF_ARCH_32_BIT 1
-#elif defined(_M_ARM64)
-#define GOOGLE_PROTOBUF_ARCH_ARM 1
-#define GOOGLE_PROTOBUF_ARCH_64_BIT 1
-#elif defined(__aarch64__)
-#define GOOGLE_PROTOBUF_ARCH_AARCH64 1
-#define GOOGLE_PROTOBUF_ARCH_64_BIT 1
-#elif defined(__mips__)
-#if defined(__LP64__)
-#define GOOGLE_PROTOBUF_ARCH_MIPS64 1
-#define GOOGLE_PROTOBUF_ARCH_64_BIT 1
-#else
-#define GOOGLE_PROTOBUF_ARCH_MIPS 1
-#define GOOGLE_PROTOBUF_ARCH_32_BIT 1
-#endif
-#elif defined(__pnacl__)
-#define GOOGLE_PROTOBUF_ARCH_32_BIT 1
-#elif defined(sparc)
-#define GOOGLE_PROTOBUF_ARCH_SPARC 1
-#if defined(__sparc_v9__) || defined(__sparcv9) || defined(__arch64__)
-#define GOOGLE_PROTOBUF_ARCH_64_BIT 1
-#else
-#define GOOGLE_PROTOBUF_ARCH_32_BIT 1
-#endif
-#elif defined(_POWER) || defined(__powerpc64__) || defined(__PPC64__)
-#define GOOGLE_PROTOBUF_ARCH_POWER 1
-#define GOOGLE_PROTOBUF_ARCH_64_BIT 1
-#elif defined(__PPC__)
-#define GOOGLE_PROTOBUF_ARCH_PPC 1
-#define GOOGLE_PROTOBUF_ARCH_32_BIT 1
-#elif defined(__GNUC__)
-# if (((__GNUC__ == 4) && (__GNUC_MINOR__ >= 7)) || (__GNUC__ > 4))
-// We fallback to the generic Clang/GCC >= 4.7 implementation in atomicops.h
-# elif defined(__clang__)
-#  if !__has_extension(c_atomic)
-GOOGLE_PROTOBUF_PLATFORM_ERROR
-#  endif
-// We fallback to the generic Clang/GCC >= 4.7 implementation in atomicops.h
-# endif
-# if __LP64__
-#  define GOOGLE_PROTOBUF_ARCH_64_BIT 1
-# else
-#  define GOOGLE_PROTOBUF_ARCH_32_BIT 1
-# endif
-#else
-GOOGLE_PROTOBUF_PLATFORM_ERROR
-#endif
-
-#if defined(__APPLE__)
-#define GOOGLE_PROTOBUF_OS_APPLE
-#include <Availability.h>
-#include <TargetConditionals.h>
-#if TARGET_OS_IPHONE
-#define GOOGLE_PROTOBUF_OS_IPHONE
-#endif
-#elif defined(__EMSCRIPTEN__)
-#define GOOGLE_PROTOBUF_OS_EMSCRIPTEN
-#elif defined(__native_client__)
-#define GOOGLE_PROTOBUF_OS_NACL
-#elif defined(sun)
-#define GOOGLE_PROTOBUF_OS_SOLARIS
-#elif defined(_AIX)
-#define GOOGLE_PROTOBUF_OS_AIX
-#elif defined(__ANDROID__)
-#define GOOGLE_PROTOBUF_OS_ANDROID
-#endif
-
-#undef GOOGLE_PROTOBUF_PLATFORM_ERROR
-
-#if defined(GOOGLE_PROTOBUF_OS_ANDROID) || defined(GOOGLE_PROTOBUF_OS_IPHONE) || defined(__OpenBSD__)
-// Android ndk does not support the __thread keyword very well yet. Here
-// we use pthread_key_create()/pthread_getspecific()/... methods for
-// TLS support on android.
-// iOS and OpenBSD also do not support the __thread keyword.
-#define GOOGLE_PROTOBUF_NO_THREADLOCAL
-#endif
-
-#if defined(__MAC_OS_X_VERSION_MIN_REQUIRED) && __MAC_OS_X_VERSION_MIN_REQUIRED < 1070
-// __thread keyword requires at least 10.7
-#define GOOGLE_PROTOBUF_NO_THREADLOCAL
-#endif
-
-#endif  // GOOGLE_PROTOBUF_PLATFORM_MACROS_H_
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/port.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/port.h
deleted file mode 100644
index 0fcee689..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/port.h
+++ /dev/null
@@ -1,405 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-#ifndef GOOGLE_PROTOBUF_STUBS_PORT_H_
-#define GOOGLE_PROTOBUF_STUBS_PORT_H_
-
-#include <assert.h>
-#include <cstdint>
-#include <stdlib.h>
-#include <cstddef>
-#include <string>
-#include <string.h>
-
-#include <google/protobuf/stubs/platform_macros.h>
-
-#include <google/protobuf/port_def.inc>
-
-#undef PROTOBUF_LITTLE_ENDIAN
-#ifdef _WIN32
-  // Assuming windows is always little-endian.
-  // TODO(xiaofeng): The PROTOBUF_LITTLE_ENDIAN is not only used for
-  // optimization but also for correctness. We should define an
-  // different macro to test the big-endian code path in coded_stream.
-  #if !defined(PROTOBUF_DISABLE_LITTLE_ENDIAN_OPT_FOR_TEST)
-    #define PROTOBUF_LITTLE_ENDIAN 1
-  #endif
-  #if defined(_MSC_VER) && _MSC_VER >= 1300 && !defined(__INTEL_COMPILER)
-    // If MSVC has "/RTCc" set, it will complain about truncating casts at
-    // runtime.  This file contains some intentional truncating casts.
-    #pragma runtime_checks("c", off)
-  #endif
-#else
-  #include <sys/param.h>   // __BYTE_ORDER
-  #if defined(__OpenBSD__)
-    #include <endian.h>
-  #endif
-  #if ((defined(__LITTLE_ENDIAN__) && !defined(__BIG_ENDIAN__)) || \
-         (defined(__BYTE_ORDER) && __BYTE_ORDER == __LITTLE_ENDIAN) || \
-         (defined(BYTE_ORDER) && BYTE_ORDER == LITTLE_ENDIAN)) && \
-      !defined(PROTOBUF_DISABLE_LITTLE_ENDIAN_OPT_FOR_TEST)
-    #define PROTOBUF_LITTLE_ENDIAN 1
-  #endif
-#endif
-
-// These #includes are for the byte swap functions declared later on.
-#ifdef _MSC_VER
-#include <stdlib.h>  // NOLINT(build/include)
-#include <intrin.h>
-#elif defined(__APPLE__)
-#include <libkern/OSByteOrder.h>
-#elif defined(__GLIBC__) || defined(__BIONIC__) || defined(__CYGWIN__)
-#include <byteswap.h>  // IWYU pragma: export
-#endif
-
-// Legacy: some users reference these (internal-only) macros even though we
-// don't need them any more.
-#if defined(_MSC_VER) && defined(PROTOBUF_USE_DLLS)
-  #ifdef LIBPROTOBUF_EXPORTS
-    #define LIBPROTOBUF_EXPORT __declspec(dllexport)
-  #else
-    #define LIBPROTOBUF_EXPORT __declspec(dllimport)
-  #endif
-  #ifdef LIBPROTOC_EXPORTS
-    #define LIBPROTOC_EXPORT   __declspec(dllexport)
-  #else
-    #define LIBPROTOC_EXPORT   __declspec(dllimport)
-  #endif
-#else
-  #define LIBPROTOBUF_EXPORT
-  #define LIBPROTOC_EXPORT
-#endif
-
-#define PROTOBUF_RUNTIME_DEPRECATED(message) PROTOBUF_DEPRECATED_MSG(message)
-#define GOOGLE_PROTOBUF_RUNTIME_DEPRECATED(message) \
-  PROTOBUF_DEPRECATED_MSG(message)
-
-// ===================================================================
-// from google3/base/port.h
-
-#if (defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103L || \
-     (defined(_MSC_VER) && _MSC_VER >= 1900))
-// Define this to 1 if the code is compiled in C++11 mode; leave it
-// undefined otherwise.  Do NOT define it to 0 -- that causes
-// '#ifdef LANG_CXX11' to behave differently from '#if LANG_CXX11'.
-#define LANG_CXX11 1
-#else
-#error "Protobuf requires at least C++11."
-#endif
-
-namespace google {
-namespace protobuf {
-
-using ConstStringParam = const std::string &;
-
-typedef unsigned int uint;
-
-typedef int8_t int8;
-typedef int16_t int16;
-typedef int32_t int32;
-typedef int64_t int64;
-
-typedef uint8_t uint8;
-typedef uint16_t uint16;
-typedef uint32_t uint32;
-typedef uint64_t uint64;
-
-static const int32 kint32max = 0x7FFFFFFF;
-static const int32 kint32min = -kint32max - 1;
-static const int64 kint64max = PROTOBUF_LONGLONG(0x7FFFFFFFFFFFFFFF);
-static const int64 kint64min = -kint64max - 1;
-static const uint32 kuint32max = 0xFFFFFFFFu;
-static const uint64 kuint64max = PROTOBUF_ULONGLONG(0xFFFFFFFFFFFFFFFF);
-
-#if defined(ADDRESS_SANITIZER) || defined(THREAD_SANITIZER) ||\
-    defined(MEMORY_SANITIZER)
-
-#ifdef __cplusplus
-extern "C" {
-#endif  // __cplusplus
-uint16_t __sanitizer_unaligned_load16(const void *p);
-uint32_t __sanitizer_unaligned_load32(const void *p);
-uint64_t __sanitizer_unaligned_load64(const void *p);
-void __sanitizer_unaligned_store16(void *p, uint16_t v);
-void __sanitizer_unaligned_store32(void *p, uint32_t v);
-void __sanitizer_unaligned_store64(void *p, uint64_t v);
-#ifdef __cplusplus
-}  // extern "C"
-#endif  // __cplusplus
-
-inline uint16 GOOGLE_UNALIGNED_LOAD16(const void *p) {
-  return __sanitizer_unaligned_load16(p);
-}
-
-inline uint32 GOOGLE_UNALIGNED_LOAD32(const void *p) {
-  return __sanitizer_unaligned_load32(p);
-}
-
-inline uint64 GOOGLE_UNALIGNED_LOAD64(const void *p) {
-  return __sanitizer_unaligned_load64(p);
-}
-
-inline void GOOGLE_UNALIGNED_STORE16(void *p, uint16 v) {
-  __sanitizer_unaligned_store16(p, v);
-}
-
-inline void GOOGLE_UNALIGNED_STORE32(void *p, uint32 v) {
-  __sanitizer_unaligned_store32(p, v);
-}
-
-inline void GOOGLE_UNALIGNED_STORE64(void *p, uint64 v) {
-  __sanitizer_unaligned_store64(p, v);
-}
-
-#elif defined(GOOGLE_PROTOBUF_USE_UNALIGNED) && GOOGLE_PROTOBUF_USE_UNALIGNED
-
-#define GOOGLE_UNALIGNED_LOAD16(_p) (*reinterpret_cast<const uint16 *>(_p))
-#define GOOGLE_UNALIGNED_LOAD32(_p) (*reinterpret_cast<const uint32 *>(_p))
-#define GOOGLE_UNALIGNED_LOAD64(_p) (*reinterpret_cast<const uint64 *>(_p))
-
-#define GOOGLE_UNALIGNED_STORE16(_p, _val) (*reinterpret_cast<uint16 *>(_p) = (_val))
-#define GOOGLE_UNALIGNED_STORE32(_p, _val) (*reinterpret_cast<uint32 *>(_p) = (_val))
-#define GOOGLE_UNALIGNED_STORE64(_p, _val) (*reinterpret_cast<uint64 *>(_p) = (_val))
-
-#else
-inline uint16 GOOGLE_UNALIGNED_LOAD16(const void *p) {
-  uint16 t;
-  memcpy(&t, p, sizeof t);
-  return t;
-}
-
-inline uint32 GOOGLE_UNALIGNED_LOAD32(const void *p) {
-  uint32 t;
-  memcpy(&t, p, sizeof t);
-  return t;
-}
-
-inline uint64 GOOGLE_UNALIGNED_LOAD64(const void *p) {
-  uint64 t;
-  memcpy(&t, p, sizeof t);
-  return t;
-}
-
-inline void GOOGLE_UNALIGNED_STORE16(void *p, uint16 v) {
-  memcpy(p, &v, sizeof v);
-}
-
-inline void GOOGLE_UNALIGNED_STORE32(void *p, uint32 v) {
-  memcpy(p, &v, sizeof v);
-}
-
-inline void GOOGLE_UNALIGNED_STORE64(void *p, uint64 v) {
-  memcpy(p, &v, sizeof v);
-}
-#endif
-
-#if defined(GOOGLE_PROTOBUF_OS_NACL) \
-    || (defined(__ANDROID__) && defined(__clang__) \
-        && (__clang_major__ == 3 && __clang_minor__ == 8) \
-        && (__clang_patchlevel__ < 275480))
-# define GOOGLE_PROTOBUF_USE_PORTABLE_LOG2
-#endif
-
-// The following guarantees declaration of the byte swap functions.
-#ifdef _MSC_VER
-#define bswap_16(x) _byteswap_ushort(x)
-#define bswap_32(x) _byteswap_ulong(x)
-#define bswap_64(x) _byteswap_uint64(x)
-
-#elif defined(__APPLE__)
-// Mac OS X / Darwin features
-#define bswap_16(x) OSSwapInt16(x)
-#define bswap_32(x) OSSwapInt32(x)
-#define bswap_64(x) OSSwapInt64(x)
-
-#elif !defined(__GLIBC__) && !defined(__BIONIC__) && !defined(__CYGWIN__)
-
-#ifndef bswap_16
-static inline uint16 bswap_16(uint16 x) {
-  return static_cast<uint16>(((x & 0xFF) << 8) | ((x & 0xFF00) >> 8));
-}
-#define bswap_16(x) bswap_16(x)
-#endif
-
-#ifndef bswap_32
-static inline uint32 bswap_32(uint32 x) {
-  return (((x & 0xFF) << 24) |
-          ((x & 0xFF00) << 8) |
-          ((x & 0xFF0000) >> 8) |
-          ((x & 0xFF000000) >> 24));
-}
-#define bswap_32(x) bswap_32(x)
-#endif
-
-#ifndef bswap_64
-static inline uint64 bswap_64(uint64 x) {
-  return (((x & PROTOBUF_ULONGLONG(0xFF)) << 56) |
-          ((x & PROTOBUF_ULONGLONG(0xFF00)) << 40) |
-          ((x & PROTOBUF_ULONGLONG(0xFF0000)) << 24) |
-          ((x & PROTOBUF_ULONGLONG(0xFF000000)) << 8) |
-          ((x & PROTOBUF_ULONGLONG(0xFF00000000)) >> 8) |
-          ((x & PROTOBUF_ULONGLONG(0xFF0000000000)) >> 24) |
-          ((x & PROTOBUF_ULONGLONG(0xFF000000000000)) >> 40) |
-          ((x & PROTOBUF_ULONGLONG(0xFF00000000000000)) >> 56));
-}
-#define bswap_64(x) bswap_64(x)
-#endif
-
-#endif
-
-// ===================================================================
-// from google3/util/bits/bits.h
-
-class Bits {
- public:
-  static uint32 Log2FloorNonZero(uint32 n) {
-#if defined(__GNUC__)
-  return 31 ^ static_cast<uint32>(__builtin_clz(n));
-#elif defined(_MSC_VER)
-  unsigned long where;
-  _BitScanReverse(&where, n);
-  return where;
-#else
-  return Log2FloorNonZero_Portable(n);
-#endif
-  }
-
-  static uint32 Log2FloorNonZero64(uint64 n) {
-    // Older versions of clang run into an instruction-selection failure when
-    // it encounters __builtin_clzll:
-    // https://bugs.chromium.org/p/nativeclient/issues/detail?id=4395
-    // This includes arm-nacl-clang and clang in older Android NDK versions.
-    // To work around this, when we build with those we use the portable
-    // implementation instead.
-#if defined(__GNUC__) && !defined(GOOGLE_PROTOBUF_USE_PORTABLE_LOG2)
-  return 63 ^ static_cast<uint32>(__builtin_clzll(n));
-#elif defined(_MSC_VER) && defined(_M_X64)
-  unsigned long where;
-  _BitScanReverse64(&where, n);
-  return where;
-#else
-  return Log2FloorNonZero64_Portable(n);
-#endif
-  }
- private:
-  static int Log2FloorNonZero_Portable(uint32 n) {
-    if (n == 0)
-      return -1;
-    int log = 0;
-    uint32 value = n;
-    for (int i = 4; i >= 0; --i) {
-      int shift = (1 << i);
-      uint32 x = value >> shift;
-      if (x != 0) {
-        value = x;
-        log += shift;
-      }
-    }
-    assert(value == 1);
-    return log;
-  }
-
-  static int Log2FloorNonZero64_Portable(uint64 n) {
-    const uint32 topbits = static_cast<uint32>(n >> 32);
-    if (topbits == 0) {
-      // Top bits are zero, so scan in bottom bits
-      return static_cast<int>(Log2FloorNonZero(static_cast<uint32>(n)));
-    } else {
-      return 32 + static_cast<int>(Log2FloorNonZero(topbits));
-    }
-  }
-};
-
-// ===================================================================
-// from google3/util/endian/endian.h
-PROTOBUF_EXPORT uint32 ghtonl(uint32 x);
-
-class BigEndian {
- public:
-#ifdef PROTOBUF_LITTLE_ENDIAN
-
-  static uint16 FromHost16(uint16 x) { return bswap_16(x); }
-  static uint16 ToHost16(uint16 x) { return bswap_16(x); }
-
-  static uint32 FromHost32(uint32 x) { return bswap_32(x); }
-  static uint32 ToHost32(uint32 x) { return bswap_32(x); }
-
-  static uint64 FromHost64(uint64 x) { return bswap_64(x); }
-  static uint64 ToHost64(uint64 x) { return bswap_64(x); }
-
-  static bool IsLittleEndian() { return true; }
-
-#else
-
-  static uint16 FromHost16(uint16 x) { return x; }
-  static uint16 ToHost16(uint16 x) { return x; }
-
-  static uint32 FromHost32(uint32 x) { return x; }
-  static uint32 ToHost32(uint32 x) { return x; }
-
-  static uint64 FromHost64(uint64 x) { return x; }
-  static uint64 ToHost64(uint64 x) { return x; }
-
-  static bool IsLittleEndian() { return false; }
-
-#endif /* ENDIAN */
-
-  // Functions to do unaligned loads and stores in big-endian order.
-  static uint16 Load16(const void *p) {
-    return ToHost16(GOOGLE_UNALIGNED_LOAD16(p));
-  }
-
-  static void Store16(void *p, uint16 v) {
-    GOOGLE_UNALIGNED_STORE16(p, FromHost16(v));
-  }
-
-  static uint32 Load32(const void *p) {
-    return ToHost32(GOOGLE_UNALIGNED_LOAD32(p));
-  }
-
-  static void Store32(void *p, uint32 v) {
-    GOOGLE_UNALIGNED_STORE32(p, FromHost32(v));
-  }
-
-  static uint64 Load64(const void *p) {
-    return ToHost64(GOOGLE_UNALIGNED_LOAD64(p));
-  }
-
-  static void Store64(void *p, uint64 v) {
-    GOOGLE_UNALIGNED_STORE64(p, FromHost64(v));
-  }
-};
-
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_STUBS_PORT_H_
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/status.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/status.h
deleted file mode 100644
index bededad5..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/status.h
+++ /dev/null
@@ -1,125 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-#ifndef GOOGLE_PROTOBUF_STUBS_STATUS_H_
-#define GOOGLE_PROTOBUF_STUBS_STATUS_H_
-
-#include <iosfwd>
-#include <string>
-
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/stubs/stringpiece.h>
-
-#include <google/protobuf/port_def.inc>
-
-namespace google {
-namespace protobuf {
-namespace util {
-namespace error {
-// These values must match error codes defined in google/rpc/code.proto.
-enum Code {
-  OK = 0,
-  CANCELLED = 1,
-  UNKNOWN = 2,
-  INVALID_ARGUMENT = 3,
-  DEADLINE_EXCEEDED = 4,
-  NOT_FOUND = 5,
-  ALREADY_EXISTS = 6,
-  PERMISSION_DENIED = 7,
-  UNAUTHENTICATED = 16,
-  RESOURCE_EXHAUSTED = 8,
-  FAILED_PRECONDITION = 9,
-  ABORTED = 10,
-  OUT_OF_RANGE = 11,
-  UNIMPLEMENTED = 12,
-  INTERNAL = 13,
-  UNAVAILABLE = 14,
-  DATA_LOSS = 15,
-};
-}  // namespace error
-
-class PROTOBUF_EXPORT Status {
- public:
-  // Creates a "successful" status.
-  Status();
-
-  // Create a status in the canonical error space with the specified
-  // code, and error message.  If "code == 0", error_message is
-  // ignored and a Status object identical to Status::OK is
-  // constructed.
-  Status(error::Code error_code, StringPiece error_message);
-  Status(const Status&);
-  Status& operator=(const Status& x);
-  ~Status() {}
-
-  // Some pre-defined Status objects
-  static const Status OK;             // Identical to 0-arg constructor
-  static const Status CANCELLED;
-  static const Status UNKNOWN;
-
-  // Accessor
-  bool ok() const {
-    return error_code_ == error::OK;
-  }
-  int error_code() const {
-    return error_code_;
-  }
-  error::Code code() const {
-    return error_code_;
-  }
-  StringPiece error_message() const {
-    return error_message_;
-  }
-  StringPiece message() const {
-    return error_message_;
-  }
-
-  bool operator==(const Status& x) const;
-  bool operator!=(const Status& x) const {
-    return !operator==(x);
-  }
-
-  // Return a combination of the error code name and message.
-  std::string ToString() const;
-
- private:
-  error::Code error_code_;
-  std::string error_message_;
-};
-
-// Prints a human-readable representation of 'x' to 'os'.
-PROTOBUF_EXPORT std::ostream& operator<<(std::ostream& os, const Status& x);
-
-}  // namespace util
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_STUBS_STATUS_H_
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/statusor.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/statusor.h
deleted file mode 100644
index c02e89a9..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/statusor.h
+++ /dev/null
@@ -1,272 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// StatusOr<T> is the union of a Status object and a T
-// object. StatusOr models the concept of an object that is either a
-// usable value, or an error Status explaining why such a value is
-// not present. To this end, StatusOr<T> does not allow its Status
-// value to be Status::OK. Further, StatusOr<T*> does not allow the
-// contained pointer to be nullptr.
-//
-// The primary use-case for StatusOr<T> is as the return value of a
-// function which may fail.
-//
-// Example client usage for a StatusOr<T>, where T is not a pointer:
-//
-//  StatusOr<float> result = DoBigCalculationThatCouldFail();
-//  if (result.ok()) {
-//    float answer = result.ValueOrDie();
-//    printf("Big calculation yielded: %f", answer);
-//  } else {
-//    LOG(ERROR) << result.status();
-//  }
-//
-// Example client usage for a StatusOr<T*>:
-//
-//  StatusOr<Foo*> result = FooFactory::MakeNewFoo(arg);
-//  if (result.ok()) {
-//    std::unique_ptr<Foo> foo(result.ValueOrDie());
-//    foo->DoSomethingCool();
-//  } else {
-//    LOG(ERROR) << result.status();
-//  }
-//
-// Example client usage for a StatusOr<std::unique_ptr<T>>:
-//
-//  StatusOr<std::unique_ptr<Foo>> result = FooFactory::MakeNewFoo(arg);
-//  if (result.ok()) {
-//    std::unique_ptr<Foo> foo = result.ConsumeValueOrDie();
-//    foo->DoSomethingCool();
-//  } else {
-//    LOG(ERROR) << result.status();
-//  }
-//
-// Example factory implementation returning StatusOr<T*>:
-//
-//  StatusOr<Foo*> FooFactory::MakeNewFoo(int arg) {
-//    if (arg <= 0) {
-//      return ::util::Status(::util::error::INVALID_ARGUMENT,
-//                            "Arg must be positive");
-//    } else {
-//      return new Foo(arg);
-//    }
-//  }
-//
-
-#ifndef GOOGLE_PROTOBUF_STUBS_STATUSOR_H_
-#define GOOGLE_PROTOBUF_STUBS_STATUSOR_H_
-
-#include <new>
-#include <string>
-#include <utility>
-
-#include <google/protobuf/stubs/status.h>
-
-#include <google/protobuf/port_def.inc>
-
-namespace google {
-namespace protobuf {
-namespace util {
-
-template<typename T>
-class StatusOr {
-  template<typename U> friend class StatusOr;
-
- public:
-  // Construct a new StatusOr with Status::UNKNOWN status
-  StatusOr();
-
-  // Construct a new StatusOr with the given non-ok status. After calling
-  // this constructor, calls to ValueOrDie() will CHECK-fail.
-  //
-  // NOTE: Not explicit - we want to use StatusOr<T> as a return
-  // value, so it is convenient and sensible to be able to do 'return
-  // Status()' when the return type is StatusOr<T>.
-  //
-  // REQUIRES: status != Status::OK. This requirement is DCHECKed.
-  // In optimized builds, passing Status::OK here will have the effect
-  // of passing PosixErrorSpace::EINVAL as a fallback.
-  StatusOr(const Status& status);  // NOLINT
-
-  // Construct a new StatusOr with the given value. If T is a plain pointer,
-  // value must not be nullptr. After calling this constructor, calls to
-  // ValueOrDie() will succeed, and calls to status() will return OK.
-  //
-  // NOTE: Not explicit - we want to use StatusOr<T> as a return type
-  // so it is convenient and sensible to be able to do 'return T()'
-  // when when the return type is StatusOr<T>.
-  //
-  // REQUIRES: if T is a plain pointer, value != nullptr. This requirement is
-  // DCHECKed. In optimized builds, passing a null pointer here will have
-  // the effect of passing PosixErrorSpace::EINVAL as a fallback.
-  StatusOr(const T& value);  // NOLINT
-
-  // Copy constructor.
-  StatusOr(const StatusOr& other);
-
-  // Conversion copy constructor, T must be copy constructible from U
-  template<typename U>
-  StatusOr(const StatusOr<U>& other);
-
-  // Assignment operator.
-  StatusOr& operator=(const StatusOr& other);
-
-  // Conversion assignment operator, T must be assignable from U
-  template<typename U>
-  StatusOr& operator=(const StatusOr<U>& other);
-
-  // Returns a reference to our status. If this contains a T, then
-  // returns Status::OK.
-  const Status& status() const;
-
-  // Returns this->status().ok()
-  bool ok() const;
-
-  // Returns a reference to our current value, or CHECK-fails if !this->ok().
-  // If you need to initialize a T object from the stored value,
-  // ConsumeValueOrDie() may be more efficient.
-  const T& ValueOrDie() const;
-  const T& value () const;
-
- private:
-  Status status_;
-  T value_;
-};
-
-////////////////////////////////////////////////////////////////////////////////
-// Implementation details for StatusOr<T>
-
-namespace internal {
-
-class PROTOBUF_EXPORT StatusOrHelper {
- public:
-  // Move type-agnostic error handling to the .cc.
-  static void Crash(const util::Status& status);
-
-  // Customized behavior for StatusOr<T> vs. StatusOr<T*>
-  template<typename T>
-  struct Specialize;
-};
-
-template<typename T>
-struct StatusOrHelper::Specialize {
-  // For non-pointer T, a reference can never be nullptr.
-  static inline bool IsValueNull(const T& t) { return false; }
-};
-
-template<typename T>
-struct StatusOrHelper::Specialize<T*> {
-  static inline bool IsValueNull(const T* t) { return t == nullptr; }
-};
-
-}  // namespace internal
-
-template<typename T>
-inline StatusOr<T>::StatusOr()
-    : status_(util::Status::UNKNOWN) {
-}
-
-template<typename T>
-inline StatusOr<T>::StatusOr(const Status& status) {
-  if (status.ok()) {
-    status_ = Status(error::INTERNAL, "Status::OK is not a valid argument.");
-  } else {
-    status_ = status;
-  }
-}
-
-template<typename T>
-inline StatusOr<T>::StatusOr(const T& value) {
-  if (internal::StatusOrHelper::Specialize<T>::IsValueNull(value)) {
-    status_ = Status(error::INTERNAL, "nullptr is not a valid argument.");
-  } else {
-    status_ = Status::OK;
-    value_ = value;
-  }
-}
-
-template<typename T>
-inline StatusOr<T>::StatusOr(const StatusOr<T>& other)
-    : status_(other.status_), value_(other.value_) {
-}
-
-template<typename T>
-inline StatusOr<T>& StatusOr<T>::operator=(const StatusOr<T>& other) {
-  status_ = other.status_;
-  value_ = other.value_;
-  return *this;
-}
-
-template<typename T>
-template<typename U>
-inline StatusOr<T>::StatusOr(const StatusOr<U>& other)
-    : status_(other.status_), value_(other.status_.ok() ? other.value_ : T()) {
-}
-
-template<typename T>
-template<typename U>
-inline StatusOr<T>& StatusOr<T>::operator=(const StatusOr<U>& other) {
-  status_ = other.status_;
-  if (status_.ok()) value_ = other.value_;
-  return *this;
-}
-
-template<typename T>
-inline const Status& StatusOr<T>::status() const {
-  return status_;
-}
-
-template<typename T>
-inline bool StatusOr<T>::ok() const {
-  return status().ok();
-}
-
-template<typename T>
-inline const T& StatusOr<T>::ValueOrDie() const {
-  if (!status_.ok()) {
-    internal::StatusOrHelper::Crash(status_);
-  }
-  return value_;
-}
-
-template<typename T>
-inline const T& StatusOr<T>::value() const {
-  if (!status_.ok()) {
-    internal::StatusOrHelper::Crash(status_);
-  }
-  return value_;
-}
-}  // namespace util
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_STUBS_STATUSOR_H_
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/stl_util.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/stl_util.h
deleted file mode 100644
index d01f9ec9..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/stl_util.h
+++ /dev/null
@@ -1,71 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// from google3/util/gtl/stl_util.h
-
-#ifndef GOOGLE_PROTOBUF_STUBS_STL_UTIL_H__
-#define GOOGLE_PROTOBUF_STUBS_STL_UTIL_H__
-
-#include <google/protobuf/stubs/common.h>
-
-namespace google {
-namespace protobuf {
-
-// Inside Google, this function implements a horrible, disgusting hack in which
-// we reach into the string's private implementation and resize it without
-// initializing the new bytes.  In some cases doing this can significantly
-// improve performance.  However, since it's totally non-portable it has no
-// place in open source code.  Feel free to fill this function in with your
-// own disgusting hack if you want the perf boost.
-inline void STLStringResizeUninitialized(std::string* s, size_t new_size) {
-  s->resize(new_size);
-}
-
-// Return a mutable char* pointing to a string's internal buffer,
-// which may not be null-terminated. Writing through this pointer will
-// modify the string.
-//
-// string_as_array(&str)[i] is valid for 0 <= i < str.size() until the
-// next call to a string method that invalidates iterators.
-//
-// As of 2006-04, there is no standard-blessed way of getting a
-// mutable reference to a string's internal buffer. However, issue 530
-// (http://www.open-std.org/JTC1/SC22/WG21/docs/lwg-active.html#530)
-// proposes this as the method. According to Matt Austern, this should
-// already work on all current implementations.
-inline char* string_as_array(std::string* str) {
-  // DO NOT USE const_cast<char*>(str->data())! See the unittest for why.
-  return str->empty() ? nullptr : &*str->begin();
-}
-
-}  // namespace protobuf
-}  // namespace google
-
-#endif  // GOOGLE_PROTOBUF_STUBS_STL_UTIL_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/stringpiece.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/stringpiece.h
deleted file mode 100644
index fbcb20af..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/stringpiece.h
+++ /dev/null
@@ -1,489 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// A StringPiece points to part or all of a string, Cord, double-quoted string
-// literal, or other string-like object.  A StringPiece does *not* own the
-// string to which it points.  A StringPiece is not null-terminated.
-//
-// You can use StringPiece as a function or method parameter.  A StringPiece
-// parameter can receive a double-quoted string literal argument, a "const
-// char*" argument, a string argument, or a StringPiece argument with no data
-// copying.  Systematic use of StringPiece for arguments reduces data
-// copies and strlen() calls.
-//
-// Prefer passing StringPieces by value:
-//   void MyFunction(StringPiece arg);
-// If circumstances require, you may also pass by const reference:
-//   void MyFunction(const StringPiece& arg);  // not preferred
-// Both of these have the same lifetime semantics.  Passing by value
-// generates slightly smaller code.  For more discussion, see the thread
-// go/stringpiecebyvalue on c-users.
-//
-// StringPiece is also suitable for local variables if you know that
-// the lifetime of the underlying object is longer than the lifetime
-// of your StringPiece variable.
-//
-// Beware of binding a StringPiece to a temporary:
-//   StringPiece sp = obj.MethodReturningString();  // BAD: lifetime problem
-//
-// This code is okay:
-//   string str = obj.MethodReturningString();  // str owns its contents
-//   StringPiece sp(str);  // GOOD, because str outlives sp
-//
-// StringPiece is sometimes a poor choice for a return value and usually a poor
-// choice for a data member.  If you do use a StringPiece this way, it is your
-// responsibility to ensure that the object pointed to by the StringPiece
-// outlives the StringPiece.
-//
-// A StringPiece may represent just part of a string; thus the name "Piece".
-// For example, when splitting a string, vector<StringPiece> is a natural data
-// type for the output.  For another example, a Cord is a non-contiguous,
-// potentially very long string-like object.  The Cord class has an interface
-// that iteratively provides StringPiece objects that point to the
-// successive pieces of a Cord object.
-//
-// A StringPiece is not null-terminated.  If you write code that scans a
-// StringPiece, you must check its length before reading any characters.
-// Common idioms that work on null-terminated strings do not work on
-// StringPiece objects.
-//
-// There are several ways to create a null StringPiece:
-//   StringPiece()
-//   StringPiece(nullptr)
-//   StringPiece(nullptr, 0)
-// For all of the above, sp.data() == nullptr, sp.length() == 0,
-// and sp.empty() == true.  Also, if you create a StringPiece with
-// a non-null pointer then sp.data() != nullptr.  Once created,
-// sp.data() will stay either nullptr or not-nullptr, except if you call
-// sp.clear() or sp.set().
-//
-// Thus, you can use StringPiece(nullptr) to signal an out-of-band value
-// that is different from other StringPiece values.  This is similar
-// to the way that const char* p1 = nullptr; is different from
-// const char* p2 = "";.
-//
-// There are many ways to create an empty StringPiece:
-//   StringPiece()
-//   StringPiece(nullptr)
-//   StringPiece(nullptr, 0)
-//   StringPiece("")
-//   StringPiece("", 0)
-//   StringPiece("abcdef", 0)
-//   StringPiece("abcdef"+6, 0)
-// For all of the above, sp.length() will be 0 and sp.empty() will be true.
-// For some empty StringPiece values, sp.data() will be nullptr.
-// For some empty StringPiece values, sp.data() will not be nullptr.
-//
-// Be careful not to confuse: null StringPiece and empty StringPiece.
-// The set of empty StringPieces properly includes the set of null StringPieces.
-// That is, every null StringPiece is an empty StringPiece,
-// but some non-null StringPieces are empty Stringpieces too.
-//
-// All empty StringPiece values compare equal to each other.
-// Even a null StringPieces compares equal to a non-null empty StringPiece:
-//  StringPiece() == StringPiece("", 0)
-//  StringPiece(nullptr) == StringPiece("abc", 0)
-//  StringPiece(nullptr, 0) == StringPiece("abcdef"+6, 0)
-//
-// Look carefully at this example:
-//   StringPiece("") == nullptr
-// True or false?  TRUE, because StringPiece::operator== converts
-// the right-hand side from nullptr to StringPiece(nullptr),
-// and then compares two zero-length spans of characters.
-// However, we are working to make this example produce a compile error.
-//
-// Suppose you want to write:
-//   bool TestWhat?(StringPiece sp) { return sp == nullptr; }  // BAD
-// Do not do that.  Write one of these instead:
-//   bool TestNull(StringPiece sp) { return sp.data() == nullptr; }
-//   bool TestEmpty(StringPiece sp) { return sp.empty(); }
-// The intent of TestWhat? is unclear.  Did you mean TestNull or TestEmpty?
-// Right now, TestWhat? behaves likes TestEmpty.
-// We are working to make TestWhat? produce a compile error.
-// TestNull is good to test for an out-of-band signal.
-// TestEmpty is good to test for an empty StringPiece.
-//
-// Caveats (again):
-// (1) The lifetime of the pointed-to string (or piece of a string)
-//     must be longer than the lifetime of the StringPiece.
-// (2) There may or may not be a '\0' character after the end of
-//     StringPiece data.
-// (3) A null StringPiece is empty.
-//     An empty StringPiece may or may not be a null StringPiece.
-
-#ifndef GOOGLE_PROTOBUF_STUBS_STRINGPIECE_H_
-#define GOOGLE_PROTOBUF_STUBS_STRINGPIECE_H_
-
-#include <assert.h>
-#include <stddef.h>
-#include <string.h>
-#include <iosfwd>
-#include <limits>
-#include <string>
-
-#include <google/protobuf/stubs/hash.h>
-
-#include <google/protobuf/port_def.inc>
-
-namespace google {
-namespace protobuf {
-// StringPiece has *two* size types.
-// StringPiece::size_type
-//   is unsigned
-//   is 32 bits in LP32, 64 bits in LP64, 64 bits in LLP64
-//   no future changes intended
-// stringpiece_ssize_type
-//   is signed
-//   is 32 bits in LP32, 64 bits in LP64, 64 bits in LLP64
-//   future changes intended: http://go/64BitStringPiece
-//
-typedef std::string::difference_type stringpiece_ssize_type;
-
-// STRINGPIECE_CHECK_SIZE protects us from 32-bit overflows.
-// TODO(mec): delete this after stringpiece_ssize_type goes 64 bit.
-#if !defined(NDEBUG)
-#define STRINGPIECE_CHECK_SIZE 1
-#elif defined(_FORTIFY_SOURCE) && _FORTIFY_SOURCE > 0
-#define STRINGPIECE_CHECK_SIZE 1
-#else
-#define STRINGPIECE_CHECK_SIZE 0
-#endif
-
-class PROTOBUF_EXPORT StringPiece {
- private:
-  const char* ptr_;
-  stringpiece_ssize_type length_;
-
-  // Prevent overflow in debug mode or fortified mode.
-  // sizeof(stringpiece_ssize_type) may be smaller than sizeof(size_t).
-  static stringpiece_ssize_type CheckedSsizeTFromSizeT(size_t size) {
-#if STRINGPIECE_CHECK_SIZE > 0
-#ifdef max
-#undef max
-#endif
-    if (size > static_cast<size_t>(
-        std::numeric_limits<stringpiece_ssize_type>::max())) {
-      // Some people grep for this message in logs
-      // so take care if you ever change it.
-      LogFatalSizeTooBig(size, "size_t to int conversion");
-    }
-#endif
-    return static_cast<stringpiece_ssize_type>(size);
-  }
-
-  // Out-of-line error path.
-  static void LogFatalSizeTooBig(size_t size, const char* details);
-
- public:
-  // We provide non-explicit singleton constructors so users can pass
-  // in a "const char*" or a "string" wherever a "StringPiece" is
-  // expected.
-  //
-  // Style guide exception granted:
-  // http://goto/style-guide-exception-20978288
-  StringPiece() : ptr_(nullptr), length_(0) {}
-
-  StringPiece(const char* str)  // NOLINT(runtime/explicit)
-      : ptr_(str), length_(0) {
-    if (str != nullptr) {
-      length_ = CheckedSsizeTFromSizeT(strlen(str));
-    }
-  }
-
-  template <class Allocator>
-  StringPiece(  // NOLINT(runtime/explicit)
-      const std::basic_string<char, std::char_traits<char>, Allocator>& str)
-      : ptr_(str.data()), length_(0) {
-    length_ = CheckedSsizeTFromSizeT(str.size());
-  }
-
-  StringPiece(const char* offset, stringpiece_ssize_type len)
-      : ptr_(offset), length_(len) {
-    assert(len >= 0);
-  }
-
-  // Substring of another StringPiece.
-  // pos must be non-negative and <= x.length().
-  StringPiece(StringPiece x, stringpiece_ssize_type pos);
-  // Substring of another StringPiece.
-  // pos must be non-negative and <= x.length().
-  // len must be non-negative and will be pinned to at most x.length() - pos.
-  StringPiece(StringPiece x,
-              stringpiece_ssize_type pos,
-              stringpiece_ssize_type len);
-
-  // data() may return a pointer to a buffer with embedded NULs, and the
-  // returned buffer may or may not be null terminated.  Therefore it is
-  // typically a mistake to pass data() to a routine that expects a NUL
-  // terminated string.
-  const char* data() const { return ptr_; }
-  stringpiece_ssize_type size() const { return length_; }
-  stringpiece_ssize_type length() const { return length_; }
-  bool empty() const { return length_ == 0; }
-
-  void clear() {
-    ptr_ = nullptr;
-    length_ = 0;
-  }
-
-  void set(const char* data, stringpiece_ssize_type len) {
-    assert(len >= 0);
-    ptr_ = data;
-    length_ = len;
-  }
-
-  void set(const char* str) {
-    ptr_ = str;
-    if (str != nullptr)
-      length_ = CheckedSsizeTFromSizeT(strlen(str));
-    else
-      length_ = 0;
-  }
-
-  void set(const void* data, stringpiece_ssize_type len) {
-    ptr_ = reinterpret_cast<const char*>(data);
-    length_ = len;
-  }
-
-  char operator[](stringpiece_ssize_type i) const {
-    assert(0 <= i);
-    assert(i < length_);
-    return ptr_[i];
-  }
-
-  void remove_prefix(stringpiece_ssize_type n) {
-    assert(length_ >= n);
-    ptr_ += n;
-    length_ -= n;
-  }
-
-  void remove_suffix(stringpiece_ssize_type n) {
-    assert(length_ >= n);
-    length_ -= n;
-  }
-
-  // returns {-1, 0, 1}
-  int compare(StringPiece x) const {
-    const stringpiece_ssize_type min_size =
-        length_ < x.length_ ? length_ : x.length_;
-    int r = memcmp(ptr_, x.ptr_, static_cast<size_t>(min_size));
-    if (r < 0) return -1;
-    if (r > 0) return 1;
-    if (length_ < x.length_) return -1;
-    if (length_ > x.length_) return 1;
-    return 0;
-  }
-
-  std::string as_string() const { return ToString(); }
-  // We also define ToString() here, since many other string-like
-  // interfaces name the routine that converts to a C++ string
-  // "ToString", and it's confusing to have the method that does that
-  // for a StringPiece be called "as_string()".  We also leave the
-  // "as_string()" method defined here for existing code.
-  std::string ToString() const {
-    if (ptr_ == nullptr) return "";
-    return std::string(data(), static_cast<size_type>(size()));
-  }
-
-  explicit operator std::string() const { return ToString(); }
-
-  void CopyToString(std::string* target) const;
-  void AppendToString(std::string* target) const;
-
-  bool starts_with(StringPiece x) const {
-    return (length_ >= x.length_) &&
-           (memcmp(ptr_, x.ptr_, static_cast<size_t>(x.length_)) == 0);
-  }
-
-  bool ends_with(StringPiece x) const {
-    return ((length_ >= x.length_) &&
-            (memcmp(ptr_ + (length_-x.length_), x.ptr_,
-                 static_cast<size_t>(x.length_)) == 0));
-  }
-
-  // Checks whether StringPiece starts with x and if so advances the beginning
-  // of it to past the match.  It's basically a shortcut for starts_with
-  // followed by remove_prefix.
-  bool Consume(StringPiece x);
-  // Like above but for the end of the string.
-  bool ConsumeFromEnd(StringPiece x);
-
-  // standard STL container boilerplate
-  typedef char value_type;
-  typedef const char* pointer;
-  typedef const char& reference;
-  typedef const char& const_reference;
-  typedef size_t size_type;
-  typedef ptrdiff_t difference_type;
-  static const size_type npos;
-  typedef const char* const_iterator;
-  typedef const char* iterator;
-  typedef std::reverse_iterator<const_iterator> const_reverse_iterator;
-  typedef std::reverse_iterator<iterator> reverse_iterator;
-  iterator begin() const { return ptr_; }
-  iterator end() const { return ptr_ + length_; }
-  const_reverse_iterator rbegin() const {
-    return const_reverse_iterator(ptr_ + length_);
-  }
-  const_reverse_iterator rend() const {
-    return const_reverse_iterator(ptr_);
-  }
-  stringpiece_ssize_type max_size() const { return length_; }
-  stringpiece_ssize_type capacity() const { return length_; }
-
-  // cpplint.py emits a false positive [build/include_what_you_use]
-  stringpiece_ssize_type copy(char* buf, size_type n, size_type pos = 0) const;  // NOLINT
-
-  bool contains(StringPiece s) const;
-
-  stringpiece_ssize_type find(StringPiece s, size_type pos = 0) const;
-  stringpiece_ssize_type find(char c, size_type pos = 0) const;
-  stringpiece_ssize_type rfind(StringPiece s, size_type pos = npos) const;
-  stringpiece_ssize_type rfind(char c, size_type pos = npos) const;
-
-  stringpiece_ssize_type find_first_of(StringPiece s, size_type pos = 0) const;
-  stringpiece_ssize_type find_first_of(char c, size_type pos = 0) const {
-    return find(c, pos);
-  }
-  stringpiece_ssize_type find_first_not_of(StringPiece s,
-                                           size_type pos = 0) const;
-  stringpiece_ssize_type find_first_not_of(char c, size_type pos = 0) const;
-  stringpiece_ssize_type find_last_of(StringPiece s,
-                                      size_type pos = npos) const;
-  stringpiece_ssize_type find_last_of(char c, size_type pos = npos) const {
-    return rfind(c, pos);
-  }
-  stringpiece_ssize_type find_last_not_of(StringPiece s,
-                                          size_type pos = npos) const;
-  stringpiece_ssize_type find_last_not_of(char c, size_type pos = npos) const;
-
-  StringPiece substr(size_type pos, size_type n = npos) const;
-};
-
-// This large function is defined inline so that in a fairly common case where
-// one of the arguments is a literal, the compiler can elide a lot of the
-// following comparisons.
-inline bool operator==(StringPiece x, StringPiece y) {
-  stringpiece_ssize_type len = x.size();
-  if (len != y.size()) {
-    return false;
-  }
-
-  return x.data() == y.data() || len <= 0 ||
-      memcmp(x.data(), y.data(), static_cast<size_t>(len)) == 0;
-}
-
-inline bool operator!=(StringPiece x, StringPiece y) {
-  return !(x == y);
-}
-
-inline bool operator<(StringPiece x, StringPiece y) {
-  const stringpiece_ssize_type min_size =
-      x.size() < y.size() ? x.size() : y.size();
-  const int r = memcmp(x.data(), y.data(), static_cast<size_t>(min_size));
-  return (r < 0) || (r == 0 && x.size() < y.size());
-}
-
-inline bool operator>(StringPiece x, StringPiece y) {
-  return y < x;
-}
-
-inline bool operator<=(StringPiece x, StringPiece y) {
-  return !(x > y);
-}
-
-inline bool operator>=(StringPiece x, StringPiece y) {
-  return !(x < y);
-}
-
-// allow StringPiece to be logged
-extern std::ostream& operator<<(std::ostream& o, StringPiece piece);
-
-namespace internal {
-// StringPiece is not a POD and can not be used in an union (pre C++11). We
-// need a POD version of it.
-struct StringPiecePod {
-  // Create from a StringPiece.
-  static StringPiecePod CreateFromStringPiece(StringPiece str) {
-    StringPiecePod pod;
-    pod.data_ = str.data();
-    pod.size_ = str.size();
-    return pod;
-  }
-
-  // Cast to StringPiece.
-  operator StringPiece() const { return StringPiece(data_, size_); }
-
-  bool operator==(const char* value) const {
-    return StringPiece(data_, size_) == StringPiece(value);
-  }
-
-  char operator[](stringpiece_ssize_type i) const {
-    assert(0 <= i);
-    assert(i < size_);
-    return data_[i];
-  }
-
-  const char* data() const { return data_; }
-
-  stringpiece_ssize_type size() const {
-    return size_;
-  }
-
-  std::string ToString() const {
-    return std::string(data_, static_cast<size_t>(size_));
-  }
-
-  explicit operator std::string() const { return ToString(); }
-
- private:
-  const char* data_;
-  stringpiece_ssize_type size_;
-};
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-GOOGLE_PROTOBUF_HASH_NAMESPACE_DECLARATION_START
-template<> struct hash<StringPiece> {
-  size_t operator()(const StringPiece& s) const {
-    size_t result = 0;
-    for (const char *str = s.data(), *end = str + s.size(); str < end; str++) {
-      result = 5 * result + static_cast<size_t>(*str);
-    }
-    return result;
-  }
-};
-GOOGLE_PROTOBUF_HASH_NAMESPACE_DECLARATION_END
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // STRINGS_STRINGPIECE_H_
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/stringprintf.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/stringprintf.h
deleted file mode 100644
index e3858be1..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/stringprintf.h
+++ /dev/null
@@ -1,85 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2012 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// from google3/base/stringprintf.h
-//
-// Printf variants that place their output in a C++ string.
-//
-// Usage:
-//      string result = StringPrintf("%d %s\n", 10, "hello");
-//      SStringPrintf(&result, "%d %s\n", 10, "hello");
-//      StringAppendF(&result, "%d %s\n", 20, "there");
-
-#ifndef GOOGLE_PROTOBUF_STUBS_STRINGPRINTF_H
-#define GOOGLE_PROTOBUF_STUBS_STRINGPRINTF_H
-
-#include <stdarg.h>
-#include <string>
-#include <vector>
-
-#include <google/protobuf/stubs/common.h>
-
-#include <google/protobuf/port_def.inc>
-
-namespace google {
-namespace protobuf {
-
-// Return a C++ string
-PROTOBUF_EXPORT extern std::string StringPrintf(const char* format, ...);
-
-// Store result into a supplied string and return it
-PROTOBUF_EXPORT extern const std::string& SStringPrintf(std::string* dst,
-                                                        const char* format,
-                                                        ...);
-
-// Append result to a supplied string
-PROTOBUF_EXPORT extern void StringAppendF(std::string* dst, const char* format,
-                                          ...);
-
-// Lower-level routine that takes a va_list and appends to a specified
-// string.  All other routines are just convenience wrappers around it.
-PROTOBUF_EXPORT extern void StringAppendV(std::string* dst, const char* format,
-                                          va_list ap);
-
-// The max arguments supported by StringPrintfVector
-PROTOBUF_EXPORT extern const int kStringPrintfVectorMaxArgs;
-
-// You can use this version when all your arguments are strings, but
-// you don't know how many arguments you'll have at compile time.
-// StringPrintfVector will LOG(FATAL) if v.size() > kStringPrintfVectorMaxArgs
-PROTOBUF_EXPORT extern std::string StringPrintfVector(
-    const char* format, const std::vector<std::string>& v);
-
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_STUBS_STRINGPRINTF_H
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/strutil.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/strutil.h
deleted file mode 100644
index 8ce81f28..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/strutil.h
+++ /dev/null
@@ -1,953 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// from google3/strings/strutil.h
-
-#ifndef GOOGLE_PROTOBUF_STUBS_STRUTIL_H__
-#define GOOGLE_PROTOBUF_STUBS_STRUTIL_H__
-
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/stubs/stringpiece.h>
-#include <stdlib.h>
-
-#include <cstring>
-#include <google/protobuf/port_def.inc>
-#include <vector>
-
-namespace google {
-namespace protobuf {
-
-#if defined(_MSC_VER) && _MSC_VER < 1800
-#define strtoll  _strtoi64
-#define strtoull _strtoui64
-#elif defined(__DECCXX) && defined(__osf__)
-// HP C++ on Tru64 does not have strtoll, but strtol is already 64-bit.
-#define strtoll strtol
-#define strtoull strtoul
-#endif
-
-// ----------------------------------------------------------------------
-// ascii_isalnum()
-//    Check if an ASCII character is alphanumeric.  We can't use ctype's
-//    isalnum() because it is affected by locale.  This function is applied
-//    to identifiers in the protocol buffer language, not to natural-language
-//    strings, so locale should not be taken into account.
-// ascii_isdigit()
-//    Like above, but only accepts digits.
-// ascii_isspace()
-//    Check if the character is a space character.
-// ----------------------------------------------------------------------
-
-inline bool ascii_isalnum(char c) {
-  return ('a' <= c && c <= 'z') ||
-         ('A' <= c && c <= 'Z') ||
-         ('0' <= c && c <= '9');
-}
-
-inline bool ascii_isdigit(char c) {
-  return ('0' <= c && c <= '9');
-}
-
-inline bool ascii_isspace(char c) {
-  return c == ' ' || c == '\t' || c == '\n' || c == '\v' || c == '\f' ||
-      c == '\r';
-}
-
-inline bool ascii_isupper(char c) {
-  return c >= 'A' && c <= 'Z';
-}
-
-inline bool ascii_islower(char c) {
-  return c >= 'a' && c <= 'z';
-}
-
-inline char ascii_toupper(char c) {
-  return ascii_islower(c) ? c - ('a' - 'A') : c;
-}
-
-inline char ascii_tolower(char c) {
-  return ascii_isupper(c) ? c + ('a' - 'A') : c;
-}
-
-inline int hex_digit_to_int(char c) {
-  /* Assume ASCII. */
-  int x = static_cast<unsigned char>(c);
-  if (x > '9') {
-    x += 9;
-  }
-  return x & 0xf;
-}
-
-// ----------------------------------------------------------------------
-// HasPrefixString()
-//    Check if a string begins with a given prefix.
-// StripPrefixString()
-//    Given a string and a putative prefix, returns the string minus the
-//    prefix string if the prefix matches, otherwise the original
-//    string.
-// ----------------------------------------------------------------------
-inline bool HasPrefixString(StringPiece str, StringPiece prefix) {
-  return str.size() >= prefix.size() &&
-         memcmp(str.data(), prefix.data(), prefix.size()) == 0;
-}
-
-inline std::string StripPrefixString(const std::string& str,
-                                     const std::string& prefix) {
-  if (HasPrefixString(str, prefix)) {
-    return str.substr(prefix.size());
-  } else {
-    return str;
-  }
-}
-
-// ----------------------------------------------------------------------
-// HasSuffixString()
-//    Return true if str ends in suffix.
-// StripSuffixString()
-//    Given a string and a putative suffix, returns the string minus the
-//    suffix string if the suffix matches, otherwise the original
-//    string.
-// ----------------------------------------------------------------------
-inline bool HasSuffixString(StringPiece str, StringPiece suffix) {
-  return str.size() >= suffix.size() &&
-         memcmp(str.data() + str.size() - suffix.size(), suffix.data(),
-                suffix.size()) == 0;
-}
-
-inline std::string StripSuffixString(const std::string& str,
-                                     const std::string& suffix) {
-  if (HasSuffixString(str, suffix)) {
-    return str.substr(0, str.size() - suffix.size());
-  } else {
-    return str;
-  }
-}
-
-// ----------------------------------------------------------------------
-// ReplaceCharacters
-//    Replaces any occurrence of the character 'remove' (or the characters
-//    in 'remove') with the character 'replacewith'.
-//    Good for keeping html characters or protocol characters (\t) out
-//    of places where they might cause a problem.
-// StripWhitespace
-//    Removes whitespaces from both ends of the given string.
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT void ReplaceCharacters(std::string* s, const char* remove,
-                                       char replacewith);
-
-PROTOBUF_EXPORT void StripWhitespace(std::string* s);
-
-// ----------------------------------------------------------------------
-// LowerString()
-// UpperString()
-// ToUpper()
-//    Convert the characters in "s" to lowercase or uppercase.  ASCII-only:
-//    these functions intentionally ignore locale because they are applied to
-//    identifiers used in the Protocol Buffer language, not to natural-language
-//    strings.
-// ----------------------------------------------------------------------
-
-inline void LowerString(std::string* s) {
-  std::string::iterator end = s->end();
-  for (std::string::iterator i = s->begin(); i != end; ++i) {
-    // tolower() changes based on locale.  We don't want this!
-    if ('A' <= *i && *i <= 'Z') *i += 'a' - 'A';
-  }
-}
-
-inline void UpperString(std::string* s) {
-  std::string::iterator end = s->end();
-  for (std::string::iterator i = s->begin(); i != end; ++i) {
-    // toupper() changes based on locale.  We don't want this!
-    if ('a' <= *i && *i <= 'z') *i += 'A' - 'a';
-  }
-}
-
-inline void ToUpper(std::string* s) { UpperString(s); }
-
-inline std::string ToUpper(const std::string& s) {
-  std::string out = s;
-  UpperString(&out);
-  return out;
-}
-
-// ----------------------------------------------------------------------
-// StringReplace()
-//    Give me a string and two patterns "old" and "new", and I replace
-//    the first instance of "old" in the string with "new", if it
-//    exists.  RETURN a new string, regardless of whether the replacement
-//    happened or not.
-// ----------------------------------------------------------------------
-
-PROTOBUF_EXPORT std::string StringReplace(const std::string& s,
-                                          const std::string& oldsub,
-                                          const std::string& newsub,
-                                          bool replace_all);
-
-// ----------------------------------------------------------------------
-// SplitStringUsing()
-//    Split a string using a character delimiter. Append the components
-//    to 'result'.  If there are consecutive delimiters, this function skips
-//    over all of them.
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT void SplitStringUsing(StringPiece full, const char* delim,
-                                      std::vector<std::string>* res);
-
-// Split a string using one or more byte delimiters, presented
-// as a nul-terminated c string. Append the components to 'result'.
-// If there are consecutive delimiters, this function will return
-// corresponding empty strings.  If you want to drop the empty
-// strings, try SplitStringUsing().
-//
-// If "full" is the empty string, yields an empty string as the only value.
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT void SplitStringAllowEmpty(StringPiece full, const char* delim,
-                                           std::vector<std::string>* result);
-
-// ----------------------------------------------------------------------
-// Split()
-//    Split a string using a character delimiter.
-// ----------------------------------------------------------------------
-inline std::vector<std::string> Split(StringPiece full, const char* delim,
-                                      bool skip_empty = true) {
-  std::vector<std::string> result;
-  if (skip_empty) {
-    SplitStringUsing(full, delim, &result);
-  } else {
-    SplitStringAllowEmpty(full, delim, &result);
-  }
-  return result;
-}
-
-// ----------------------------------------------------------------------
-// JoinStrings()
-//    These methods concatenate a vector of strings into a C++ string, using
-//    the C-string "delim" as a separator between components. There are two
-//    flavors of the function, one flavor returns the concatenated string,
-//    another takes a pointer to the target string. In the latter case the
-//    target string is cleared and overwritten.
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT void JoinStrings(const std::vector<std::string>& components,
-                                 const char* delim, std::string* result);
-
-inline std::string JoinStrings(const std::vector<std::string>& components,
-                               const char* delim) {
-  std::string result;
-  JoinStrings(components, delim, &result);
-  return result;
-}
-
-// ----------------------------------------------------------------------
-// UnescapeCEscapeSequences()
-//    Copies "source" to "dest", rewriting C-style escape sequences
-//    -- '\n', '\r', '\\', '\ooo', etc -- to their ASCII
-//    equivalents.  "dest" must be sufficiently large to hold all
-//    the characters in the rewritten string (i.e. at least as large
-//    as strlen(source) + 1 should be safe, since the replacements
-//    are always shorter than the original escaped sequences).  It's
-//    safe for source and dest to be the same.  RETURNS the length
-//    of dest.
-//
-//    It allows hex sequences \xhh, or generally \xhhhhh with an
-//    arbitrary number of hex digits, but all of them together must
-//    specify a value of a single byte (e.g. \x0045 is equivalent
-//    to \x45, and \x1234 is erroneous).
-//
-//    It also allows escape sequences of the form \uhhhh (exactly four
-//    hex digits, upper or lower case) or \Uhhhhhhhh (exactly eight
-//    hex digits, upper or lower case) to specify a Unicode code
-//    point. The dest array will contain the UTF8-encoded version of
-//    that code-point (e.g., if source contains \u2019, then dest will
-//    contain the three bytes 0xE2, 0x80, and 0x99).
-//
-//    Errors: In the first form of the call, errors are reported with
-//    LOG(ERROR). The same is true for the second form of the call if
-//    the pointer to the string std::vector is nullptr; otherwise, error
-//    messages are stored in the std::vector. In either case, the effect on
-//    the dest array is not defined, but rest of the source will be
-//    processed.
-//    ----------------------------------------------------------------------
-
-PROTOBUF_EXPORT int UnescapeCEscapeSequences(const char* source, char* dest);
-PROTOBUF_EXPORT int UnescapeCEscapeSequences(const char* source, char* dest,
-                                             std::vector<std::string>* errors);
-
-// ----------------------------------------------------------------------
-// UnescapeCEscapeString()
-//    This does the same thing as UnescapeCEscapeSequences, but creates
-//    a new string. The caller does not need to worry about allocating
-//    a dest buffer. This should be used for non performance critical
-//    tasks such as printing debug messages. It is safe for src and dest
-//    to be the same.
-//
-//    The second call stores its errors in a supplied string vector.
-//    If the string vector pointer is nullptr, it reports the errors with LOG().
-//
-//    In the first and second calls, the length of dest is returned. In the
-//    the third call, the new string is returned.
-// ----------------------------------------------------------------------
-
-PROTOBUF_EXPORT int UnescapeCEscapeString(const std::string& src,
-                                          std::string* dest);
-PROTOBUF_EXPORT int UnescapeCEscapeString(const std::string& src,
-                                          std::string* dest,
-                                          std::vector<std::string>* errors);
-PROTOBUF_EXPORT std::string UnescapeCEscapeString(const std::string& src);
-
-// ----------------------------------------------------------------------
-// CEscape()
-//    Escapes 'src' using C-style escape sequences and returns the resulting
-//    string.
-//
-//    Escaped chars: \n, \r, \t, ", ', \, and !isprint().
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT std::string CEscape(const std::string& src);
-
-// ----------------------------------------------------------------------
-// CEscapeAndAppend()
-//    Escapes 'src' using C-style escape sequences, and appends the escaped
-//    string to 'dest'.
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT void CEscapeAndAppend(StringPiece src, std::string* dest);
-
-namespace strings {
-// Like CEscape() but does not escape bytes with the upper bit set.
-PROTOBUF_EXPORT std::string Utf8SafeCEscape(const std::string& src);
-
-// Like CEscape() but uses hex (\x) escapes instead of octals.
-PROTOBUF_EXPORT std::string CHexEscape(const std::string& src);
-}  // namespace strings
-
-// ----------------------------------------------------------------------
-// strto32()
-// strtou32()
-// strto64()
-// strtou64()
-//    Architecture-neutral plug compatible replacements for strtol() and
-//    strtoul().  Long's have different lengths on ILP-32 and LP-64
-//    platforms, so using these is safer, from the point of view of
-//    overflow behavior, than using the standard libc functions.
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT int32 strto32_adaptor(const char* nptr, char** endptr,
-                                      int base);
-PROTOBUF_EXPORT uint32 strtou32_adaptor(const char* nptr, char** endptr,
-                                        int base);
-
-inline int32 strto32(const char *nptr, char **endptr, int base) {
-  if (sizeof(int32) == sizeof(long))
-    return strtol(nptr, endptr, base);
-  else
-    return strto32_adaptor(nptr, endptr, base);
-}
-
-inline uint32 strtou32(const char *nptr, char **endptr, int base) {
-  if (sizeof(uint32) == sizeof(unsigned long))
-    return strtoul(nptr, endptr, base);
-  else
-    return strtou32_adaptor(nptr, endptr, base);
-}
-
-// For now, long long is 64-bit on all the platforms we care about, so these
-// functions can simply pass the call to strto[u]ll.
-inline int64 strto64(const char *nptr, char **endptr, int base) {
-  GOOGLE_COMPILE_ASSERT(sizeof(int64) == sizeof(long long),
-                        sizeof_int64_is_not_sizeof_long_long);
-  return strtoll(nptr, endptr, base);
-}
-
-inline uint64 strtou64(const char *nptr, char **endptr, int base) {
-  GOOGLE_COMPILE_ASSERT(sizeof(uint64) == sizeof(unsigned long long),
-                        sizeof_uint64_is_not_sizeof_long_long);
-  return strtoull(nptr, endptr, base);
-}
-
-// ----------------------------------------------------------------------
-// safe_strtob()
-// safe_strto32()
-// safe_strtou32()
-// safe_strto64()
-// safe_strtou64()
-// safe_strtof()
-// safe_strtod()
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT bool safe_strtob(StringPiece str, bool* value);
-
-PROTOBUF_EXPORT bool safe_strto32(const std::string& str, int32* value);
-PROTOBUF_EXPORT bool safe_strtou32(const std::string& str, uint32* value);
-inline bool safe_strto32(const char* str, int32* value) {
-  return safe_strto32(std::string(str), value);
-}
-inline bool safe_strto32(StringPiece str, int32* value) {
-  return safe_strto32(str.ToString(), value);
-}
-inline bool safe_strtou32(const char* str, uint32* value) {
-  return safe_strtou32(std::string(str), value);
-}
-inline bool safe_strtou32(StringPiece str, uint32* value) {
-  return safe_strtou32(str.ToString(), value);
-}
-
-PROTOBUF_EXPORT bool safe_strto64(const std::string& str, int64* value);
-PROTOBUF_EXPORT bool safe_strtou64(const std::string& str, uint64* value);
-inline bool safe_strto64(const char* str, int64* value) {
-  return safe_strto64(std::string(str), value);
-}
-inline bool safe_strto64(StringPiece str, int64* value) {
-  return safe_strto64(str.ToString(), value);
-}
-inline bool safe_strtou64(const char* str, uint64* value) {
-  return safe_strtou64(std::string(str), value);
-}
-inline bool safe_strtou64(StringPiece str, uint64* value) {
-  return safe_strtou64(str.ToString(), value);
-}
-
-PROTOBUF_EXPORT bool safe_strtof(const char* str, float* value);
-PROTOBUF_EXPORT bool safe_strtod(const char* str, double* value);
-inline bool safe_strtof(const std::string& str, float* value) {
-  return safe_strtof(str.c_str(), value);
-}
-inline bool safe_strtod(const std::string& str, double* value) {
-  return safe_strtod(str.c_str(), value);
-}
-inline bool safe_strtof(StringPiece str, float* value) {
-  return safe_strtof(str.ToString(), value);
-}
-inline bool safe_strtod(StringPiece str, double* value) {
-  return safe_strtod(str.ToString(), value);
-}
-
-// ----------------------------------------------------------------------
-// FastIntToBuffer()
-// FastHexToBuffer()
-// FastHex64ToBuffer()
-// FastHex32ToBuffer()
-// FastTimeToBuffer()
-//    These are intended for speed.  FastIntToBuffer() assumes the
-//    integer is non-negative.  FastHexToBuffer() puts output in
-//    hex rather than decimal.  FastTimeToBuffer() puts the output
-//    into RFC822 format.
-//
-//    FastHex64ToBuffer() puts a 64-bit unsigned value in hex-format,
-//    padded to exactly 16 bytes (plus one byte for '\0')
-//
-//    FastHex32ToBuffer() puts a 32-bit unsigned value in hex-format,
-//    padded to exactly 8 bytes (plus one byte for '\0')
-//
-//       All functions take the output buffer as an arg.
-//    They all return a pointer to the beginning of the output,
-//    which may not be the beginning of the input buffer.
-// ----------------------------------------------------------------------
-
-// Suggested buffer size for FastToBuffer functions.  Also works with
-// DoubleToBuffer() and FloatToBuffer().
-static const int kFastToBufferSize = 32;
-
-PROTOBUF_EXPORT char* FastInt32ToBuffer(int32 i, char* buffer);
-PROTOBUF_EXPORT char* FastInt64ToBuffer(int64 i, char* buffer);
-char* FastUInt32ToBuffer(uint32 i, char* buffer);  // inline below
-char* FastUInt64ToBuffer(uint64 i, char* buffer);  // inline below
-PROTOBUF_EXPORT char* FastHexToBuffer(int i, char* buffer);
-PROTOBUF_EXPORT char* FastHex64ToBuffer(uint64 i, char* buffer);
-PROTOBUF_EXPORT char* FastHex32ToBuffer(uint32 i, char* buffer);
-
-// at least 22 bytes long
-inline char* FastIntToBuffer(int i, char* buffer) {
-  return (sizeof(i) == 4 ?
-          FastInt32ToBuffer(i, buffer) : FastInt64ToBuffer(i, buffer));
-}
-inline char* FastUIntToBuffer(unsigned int i, char* buffer) {
-  return (sizeof(i) == 4 ?
-          FastUInt32ToBuffer(i, buffer) : FastUInt64ToBuffer(i, buffer));
-}
-inline char* FastLongToBuffer(long i, char* buffer) {
-  return (sizeof(i) == 4 ?
-          FastInt32ToBuffer(i, buffer) : FastInt64ToBuffer(i, buffer));
-}
-inline char* FastULongToBuffer(unsigned long i, char* buffer) {
-  return (sizeof(i) == 4 ?
-          FastUInt32ToBuffer(i, buffer) : FastUInt64ToBuffer(i, buffer));
-}
-
-// ----------------------------------------------------------------------
-// FastInt32ToBufferLeft()
-// FastUInt32ToBufferLeft()
-// FastInt64ToBufferLeft()
-// FastUInt64ToBufferLeft()
-//
-// Like the Fast*ToBuffer() functions above, these are intended for speed.
-// Unlike the Fast*ToBuffer() functions, however, these functions write
-// their output to the beginning of the buffer (hence the name, as the
-// output is left-aligned).  The caller is responsible for ensuring that
-// the buffer has enough space to hold the output.
-//
-// Returns a pointer to the end of the string (i.e. the null character
-// terminating the string).
-// ----------------------------------------------------------------------
-
-PROTOBUF_EXPORT char* FastInt32ToBufferLeft(int32 i, char* buffer);
-PROTOBUF_EXPORT char* FastUInt32ToBufferLeft(uint32 i, char* buffer);
-PROTOBUF_EXPORT char* FastInt64ToBufferLeft(int64 i, char* buffer);
-PROTOBUF_EXPORT char* FastUInt64ToBufferLeft(uint64 i, char* buffer);
-
-// Just define these in terms of the above.
-inline char* FastUInt32ToBuffer(uint32 i, char* buffer) {
-  FastUInt32ToBufferLeft(i, buffer);
-  return buffer;
-}
-inline char* FastUInt64ToBuffer(uint64 i, char* buffer) {
-  FastUInt64ToBufferLeft(i, buffer);
-  return buffer;
-}
-
-inline std::string SimpleBtoa(bool value) { return value ? "true" : "false"; }
-
-// ----------------------------------------------------------------------
-// SimpleItoa()
-//    Description: converts an integer to a string.
-//
-//    Return value: string
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT std::string SimpleItoa(int i);
-PROTOBUF_EXPORT std::string SimpleItoa(unsigned int i);
-PROTOBUF_EXPORT std::string SimpleItoa(long i);
-PROTOBUF_EXPORT std::string SimpleItoa(unsigned long i);
-PROTOBUF_EXPORT std::string SimpleItoa(long long i);
-PROTOBUF_EXPORT std::string SimpleItoa(unsigned long long i);
-
-// ----------------------------------------------------------------------
-// SimpleDtoa()
-// SimpleFtoa()
-// DoubleToBuffer()
-// FloatToBuffer()
-//    Description: converts a double or float to a string which, if
-//    passed to NoLocaleStrtod(), will produce the exact same original double
-//    (except in case of NaN; all NaNs are considered the same value).
-//    We try to keep the string short but it's not guaranteed to be as
-//    short as possible.
-//
-//    DoubleToBuffer() and FloatToBuffer() write the text to the given
-//    buffer and return it.  The buffer must be at least
-//    kDoubleToBufferSize bytes for doubles and kFloatToBufferSize
-//    bytes for floats.  kFastToBufferSize is also guaranteed to be large
-//    enough to hold either.
-//
-//    Return value: string
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT std::string SimpleDtoa(double value);
-PROTOBUF_EXPORT std::string SimpleFtoa(float value);
-
-PROTOBUF_EXPORT char* DoubleToBuffer(double i, char* buffer);
-PROTOBUF_EXPORT char* FloatToBuffer(float i, char* buffer);
-
-// In practice, doubles should never need more than 24 bytes and floats
-// should never need more than 14 (including null terminators), but we
-// overestimate to be safe.
-static const int kDoubleToBufferSize = 32;
-static const int kFloatToBufferSize = 24;
-
-namespace strings {
-
-enum PadSpec {
-  NO_PAD = 1,
-  ZERO_PAD_2,
-  ZERO_PAD_3,
-  ZERO_PAD_4,
-  ZERO_PAD_5,
-  ZERO_PAD_6,
-  ZERO_PAD_7,
-  ZERO_PAD_8,
-  ZERO_PAD_9,
-  ZERO_PAD_10,
-  ZERO_PAD_11,
-  ZERO_PAD_12,
-  ZERO_PAD_13,
-  ZERO_PAD_14,
-  ZERO_PAD_15,
-  ZERO_PAD_16,
-};
-
-struct Hex {
-  uint64 value;
-  enum PadSpec spec;
-  template <class Int>
-  explicit Hex(Int v, PadSpec s = NO_PAD)
-      : spec(s) {
-    // Prevent sign-extension by casting integers to
-    // their unsigned counterparts.
-#ifdef LANG_CXX11
-    static_assert(
-        sizeof(v) == 1 || sizeof(v) == 2 || sizeof(v) == 4 || sizeof(v) == 8,
-        "Unknown integer type");
-#endif
-    value = sizeof(v) == 1 ? static_cast<uint8>(v)
-          : sizeof(v) == 2 ? static_cast<uint16>(v)
-          : sizeof(v) == 4 ? static_cast<uint32>(v)
-          : static_cast<uint64>(v);
-  }
-};
-
-struct PROTOBUF_EXPORT AlphaNum {
-  const char *piece_data_;  // move these to string_ref eventually
-  size_t piece_size_;       // move these to string_ref eventually
-
-  char digits[kFastToBufferSize];
-
-  // No bool ctor -- bools convert to an integral type.
-  // A bool ctor would also convert incoming pointers (bletch).
-
-  AlphaNum(int i32)
-      : piece_data_(digits),
-        piece_size_(FastInt32ToBufferLeft(i32, digits) - &digits[0]) {}
-  AlphaNum(unsigned int u32)
-      : piece_data_(digits),
-        piece_size_(FastUInt32ToBufferLeft(u32, digits) - &digits[0]) {}
-  AlphaNum(long long i64)
-      : piece_data_(digits),
-        piece_size_(FastInt64ToBufferLeft(i64, digits) - &digits[0]) {}
-  AlphaNum(unsigned long long u64)
-      : piece_data_(digits),
-        piece_size_(FastUInt64ToBufferLeft(u64, digits) - &digits[0]) {}
-
-  // Note: on some architectures, "long" is only 32 bits, not 64, but the
-  // performance hit of using FastInt64ToBufferLeft to handle 32-bit values
-  // is quite minor.
-  AlphaNum(long i64)
-      : piece_data_(digits),
-        piece_size_(FastInt64ToBufferLeft(i64, digits) - &digits[0]) {}
-  AlphaNum(unsigned long u64)
-      : piece_data_(digits),
-        piece_size_(FastUInt64ToBufferLeft(u64, digits) - &digits[0]) {}
-
-  AlphaNum(float f)
-    : piece_data_(digits), piece_size_(strlen(FloatToBuffer(f, digits))) {}
-  AlphaNum(double f)
-    : piece_data_(digits), piece_size_(strlen(DoubleToBuffer(f, digits))) {}
-
-  AlphaNum(Hex hex);
-
-  AlphaNum(const char* c_str)
-      : piece_data_(c_str), piece_size_(strlen(c_str)) {}
-  // TODO: Add a string_ref constructor, eventually
-  // AlphaNum(const StringPiece &pc) : piece(pc) {}
-
-  AlphaNum(const std::string& str)
-      : piece_data_(str.data()), piece_size_(str.size()) {}
-
-  AlphaNum(StringPiece str)
-      : piece_data_(str.data()), piece_size_(str.size()) {}
-
-  AlphaNum(internal::StringPiecePod str)
-      : piece_data_(str.data()), piece_size_(str.size()) {}
-
-  size_t size() const { return piece_size_; }
-  const char *data() const { return piece_data_; }
-
- private:
-  // Use ":" not ':'
-  AlphaNum(char c);  // NOLINT(runtime/explicit)
-
-  // Disallow copy and assign.
-  AlphaNum(const AlphaNum&);
-  void operator=(const AlphaNum&);
-};
-
-}  // namespace strings
-
-using strings::AlphaNum;
-
-// ----------------------------------------------------------------------
-// StrCat()
-//    This merges the given strings or numbers, with no delimiter.  This
-//    is designed to be the fastest possible way to construct a string out
-//    of a mix of raw C strings, strings, bool values,
-//    and numeric values.
-//
-//    Don't use this for user-visible strings.  The localization process
-//    works poorly on strings built up out of fragments.
-//
-//    For clarity and performance, don't use StrCat when appending to a
-//    string.  In particular, avoid using any of these (anti-)patterns:
-//      str.append(StrCat(...)
-//      str += StrCat(...)
-//      str = StrCat(str, ...)
-//    where the last is the worse, with the potential to change a loop
-//    from a linear time operation with O(1) dynamic allocations into a
-//    quadratic time operation with O(n) dynamic allocations.  StrAppend
-//    is a better choice than any of the above, subject to the restriction
-//    of StrAppend(&str, a, b, c, ...) that none of the a, b, c, ... may
-//    be a reference into str.
-// ----------------------------------------------------------------------
-
-PROTOBUF_EXPORT std::string StrCat(const AlphaNum& a, const AlphaNum& b);
-PROTOBUF_EXPORT std::string StrCat(const AlphaNum& a, const AlphaNum& b,
-                                   const AlphaNum& c);
-PROTOBUF_EXPORT std::string StrCat(const AlphaNum& a, const AlphaNum& b,
-                                   const AlphaNum& c, const AlphaNum& d);
-PROTOBUF_EXPORT std::string StrCat(const AlphaNum& a, const AlphaNum& b,
-                                   const AlphaNum& c, const AlphaNum& d,
-                                   const AlphaNum& e);
-PROTOBUF_EXPORT std::string StrCat(const AlphaNum& a, const AlphaNum& b,
-                                   const AlphaNum& c, const AlphaNum& d,
-                                   const AlphaNum& e, const AlphaNum& f);
-PROTOBUF_EXPORT std::string StrCat(const AlphaNum& a, const AlphaNum& b,
-                                   const AlphaNum& c, const AlphaNum& d,
-                                   const AlphaNum& e, const AlphaNum& f,
-                                   const AlphaNum& g);
-PROTOBUF_EXPORT std::string StrCat(const AlphaNum& a, const AlphaNum& b,
-                                   const AlphaNum& c, const AlphaNum& d,
-                                   const AlphaNum& e, const AlphaNum& f,
-                                   const AlphaNum& g, const AlphaNum& h);
-PROTOBUF_EXPORT std::string StrCat(const AlphaNum& a, const AlphaNum& b,
-                                   const AlphaNum& c, const AlphaNum& d,
-                                   const AlphaNum& e, const AlphaNum& f,
-                                   const AlphaNum& g, const AlphaNum& h,
-                                   const AlphaNum& i);
-
-inline std::string StrCat(const AlphaNum& a) {
-  return std::string(a.data(), a.size());
-}
-
-// ----------------------------------------------------------------------
-// StrAppend()
-//    Same as above, but adds the output to the given string.
-//    WARNING: For speed, StrAppend does not try to check each of its input
-//    arguments to be sure that they are not a subset of the string being
-//    appended to.  That is, while this will work:
-//
-//    string s = "foo";
-//    s += s;
-//
-//    This will not (necessarily) work:
-//
-//    string s = "foo";
-//    StrAppend(&s, s);
-//
-//    Note: while StrCat supports appending up to 9 arguments, StrAppend
-//    is currently limited to 4.  That's rarely an issue except when
-//    automatically transforming StrCat to StrAppend, and can easily be
-//    worked around as consecutive calls to StrAppend are quite efficient.
-// ----------------------------------------------------------------------
-
-PROTOBUF_EXPORT void StrAppend(std::string* dest, const AlphaNum& a);
-PROTOBUF_EXPORT void StrAppend(std::string* dest, const AlphaNum& a,
-                               const AlphaNum& b);
-PROTOBUF_EXPORT void StrAppend(std::string* dest, const AlphaNum& a,
-                               const AlphaNum& b, const AlphaNum& c);
-PROTOBUF_EXPORT void StrAppend(std::string* dest, const AlphaNum& a,
-                               const AlphaNum& b, const AlphaNum& c,
-                               const AlphaNum& d);
-
-// ----------------------------------------------------------------------
-// Join()
-//    These methods concatenate a range of components into a C++ string, using
-//    the C-string "delim" as a separator between components.
-// ----------------------------------------------------------------------
-template <typename Iterator>
-void Join(Iterator start, Iterator end, const char* delim,
-          std::string* result) {
-  for (Iterator it = start; it != end; ++it) {
-    if (it != start) {
-      result->append(delim);
-    }
-    StrAppend(result, *it);
-  }
-}
-
-template <typename Range>
-std::string Join(const Range& components, const char* delim) {
-  std::string result;
-  Join(components.begin(), components.end(), delim, &result);
-  return result;
-}
-
-// ----------------------------------------------------------------------
-// ToHex()
-//    Return a lower-case hex string representation of the given integer.
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT std::string ToHex(uint64 num);
-
-// ----------------------------------------------------------------------
-// GlobalReplaceSubstring()
-//    Replaces all instances of a substring in a string.  Does nothing
-//    if 'substring' is empty.  Returns the number of replacements.
-//
-//    NOTE: The string pieces must not overlap s.
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT int GlobalReplaceSubstring(const std::string& substring,
-                                           const std::string& replacement,
-                                           std::string* s);
-
-// ----------------------------------------------------------------------
-// Base64Unescape()
-//    Converts "src" which is encoded in Base64 to its binary equivalent and
-//    writes it to "dest". If src contains invalid characters, dest is cleared
-//    and the function returns false. Returns true on success.
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT bool Base64Unescape(StringPiece src, std::string* dest);
-
-// ----------------------------------------------------------------------
-// WebSafeBase64Unescape()
-//    This is a variation of Base64Unescape which uses '-' instead of '+', and
-//    '_' instead of '/'. src is not null terminated, instead specify len. I
-//    recommend that slen<szdest, but we honor szdest anyway.
-//    RETURNS the length of dest, or -1 if src contains invalid chars.
-
-//    The variation that stores into a string clears the string first, and
-//    returns false (with dest empty) if src contains invalid chars; for
-//    this version src and dest must be different strings.
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT int WebSafeBase64Unescape(const char* src, int slen, char* dest,
-                                          int szdest);
-PROTOBUF_EXPORT bool WebSafeBase64Unescape(StringPiece src, std::string* dest);
-
-// Return the length to use for the output buffer given to the base64 escape
-// routines. Make sure to use the same value for do_padding in both.
-// This function may return incorrect results if given input_len values that
-// are extremely high, which should happen rarely.
-PROTOBUF_EXPORT int CalculateBase64EscapedLen(int input_len, bool do_padding);
-// Use this version when calling Base64Escape without a do_padding arg.
-PROTOBUF_EXPORT int CalculateBase64EscapedLen(int input_len);
-
-// ----------------------------------------------------------------------
-// Base64Escape()
-// WebSafeBase64Escape()
-//    Encode "src" to "dest" using base64 encoding.
-//    src is not null terminated, instead specify len.
-//    'dest' should have at least CalculateBase64EscapedLen() length.
-//    RETURNS the length of dest.
-//    The WebSafe variation use '-' instead of '+' and '_' instead of '/'
-//    so that we can place the out in the URL or cookies without having
-//    to escape them.  It also has an extra parameter "do_padding",
-//    which when set to false will prevent padding with "=".
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT int Base64Escape(const unsigned char* src, int slen, char* dest,
-                                 int szdest);
-PROTOBUF_EXPORT int WebSafeBase64Escape(const unsigned char* src, int slen,
-                                        char* dest, int szdest,
-                                        bool do_padding);
-// Encode src into dest with padding.
-PROTOBUF_EXPORT void Base64Escape(StringPiece src, std::string* dest);
-// Encode src into dest web-safely without padding.
-PROTOBUF_EXPORT void WebSafeBase64Escape(StringPiece src, std::string* dest);
-// Encode src into dest web-safely with padding.
-PROTOBUF_EXPORT void WebSafeBase64EscapeWithPadding(StringPiece src,
-                                                    std::string* dest);
-
-PROTOBUF_EXPORT void Base64Escape(const unsigned char* src, int szsrc,
-                                  std::string* dest, bool do_padding);
-PROTOBUF_EXPORT void WebSafeBase64Escape(const unsigned char* src, int szsrc,
-                                         std::string* dest, bool do_padding);
-
-inline bool IsValidCodePoint(uint32 code_point) {
-  return code_point < 0xD800 ||
-         (code_point >= 0xE000 && code_point <= 0x10FFFF);
-}
-
-static const int UTFmax = 4;
-// ----------------------------------------------------------------------
-// EncodeAsUTF8Char()
-//  Helper to append a Unicode code point to a string as UTF8, without bringing
-//  in any external dependencies. The output buffer must be as least 4 bytes
-//  large.
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT int EncodeAsUTF8Char(uint32 code_point, char* output);
-
-// ----------------------------------------------------------------------
-// UTF8FirstLetterNumBytes()
-//   Length of the first UTF-8 character.
-// ----------------------------------------------------------------------
-PROTOBUF_EXPORT int UTF8FirstLetterNumBytes(const char* src, int len);
-
-// From google3/third_party/absl/strings/escaping.h
-
-// ----------------------------------------------------------------------
-// CleanStringLineEndings()
-//   Clean up a multi-line string to conform to Unix line endings.
-//   Reads from src and appends to dst, so usually dst should be empty.
-//
-//   If there is no line ending at the end of a non-empty string, it can
-//   be added automatically.
-//
-//   Four different types of input are correctly handled:
-//
-//     - Unix/Linux files: line ending is LF: pass through unchanged
-//
-//     - DOS/Windows files: line ending is CRLF: convert to LF
-//
-//     - Legacy Mac files: line ending is CR: convert to LF
-//
-//     - Garbled files: random line endings: convert gracefully
-//                      lonely CR, lonely LF, CRLF: convert to LF
-//
-//   @param src The multi-line string to convert
-//   @param dst The converted string is appended to this string
-//   @param auto_end_last_line Automatically terminate the last line
-//
-//   Limitations:
-//
-//     This does not do the right thing for CRCRLF files created by
-//     broken programs that do another Unix->DOS conversion on files
-//     that are already in CRLF format.  For this, a two-pass approach
-//     brute-force would be needed that
-//
-//       (1) determines the presence of LF (first one is ok)
-//       (2) if yes, removes any CR, else convert every CR to LF
-PROTOBUF_EXPORT void CleanStringLineEndings(const std::string& src,
-                                            std::string* dst,
-                                            bool auto_end_last_line);
-
-// Same as above, but transforms the argument in place.
-PROTOBUF_EXPORT void CleanStringLineEndings(std::string* str,
-                                            bool auto_end_last_line);
-
-namespace strings {
-inline bool EndsWith(StringPiece text, StringPiece suffix) {
-  return suffix.empty() ||
-      (text.size() >= suffix.size() &&
-       memcmp(text.data() + (text.size() - suffix.size()), suffix.data(),
-              suffix.size()) == 0);
-}
-}  // namespace strings
-
-namespace internal {
-
-// A locale-independent version of the standard strtod(), which always
-// uses a dot as the decimal separator.
-double NoLocaleStrtod(const char* str, char** endptr);
-
-}  // namespace internal
-
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_STUBS_STRUTIL_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/time.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/time.h
deleted file mode 100644
index b0611768..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/stubs/time.h
+++ /dev/null
@@ -1,80 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-#ifndef GOOGLE_PROTOBUF_STUBS_TIME_H_
-#define GOOGLE_PROTOBUF_STUBS_TIME_H_
-
-#include <google/protobuf/stubs/common.h>
-
-#include <google/protobuf/port_def.inc>
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-struct DateTime {
-  int year;
-  int month;
-  int day;
-  int hour;
-  int minute;
-  int second;
-};
-
-// Converts a timestamp (seconds elapsed since 1970-01-01T00:00:00, could be
-// negative to represent time before 1970-01-01) to DateTime. Returns false
-// if the timestamp is not in the range between 0001-01-01T00:00:00 and
-// 9999-12-31T23:59:59.
-bool PROTOBUF_EXPORT SecondsToDateTime(int64 seconds, DateTime* time);
-// Converts DateTime to a timestamp (seconds since 1970-01-01T00:00:00).
-// Returns false if the DateTime is not valid or is not in the valid range.
-bool PROTOBUF_EXPORT DateTimeToSeconds(const DateTime& time, int64* seconds);
-
-void PROTOBUF_EXPORT GetCurrentTime(int64* seconds, int32* nanos);
-
-// Formats a time string in RFC3339 format.
-//
-// For example, "2015-05-20T13:29:35.120Z". For nanos, 0, 3, 6 or 9 fractional
-// digits will be used depending on how many are required to represent the exact
-// value.
-//
-// Note that "nanos" must in the range of [0, 999999999].
-std::string PROTOBUF_EXPORT FormatTime(int64 seconds, int32 nanos);
-// Parses a time string. This method accepts RFC3339 date/time string with UTC
-// offset. For example, "2015-05-20T13:29:35.120-08:00".
-bool PROTOBUF_EXPORT ParseTime(const std::string& value, int64* seconds,
-                               int32* nanos);
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_STUBS_TIME_H_
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/unknown_field_set.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/unknown_field_set.h
deleted file mode 100644
index ab3633da..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/unknown_field_set.h
+++ /dev/null
@@ -1,411 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// Author: kenton@google.com (Kenton Varda)
-//  Based on original Protocol Buffers design by
-//  Sanjay Ghemawat, Jeff Dean, and others.
-//
-// Contains classes used to keep track of unrecognized fields seen while
-// parsing a protocol message.
-
-#ifndef GOOGLE_PROTOBUF_UNKNOWN_FIELD_SET_H__
-#define GOOGLE_PROTOBUF_UNKNOWN_FIELD_SET_H__
-
-#include <assert.h>
-
-#include <string>
-#include <vector>
-
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/stubs/logging.h>
-#include <google/protobuf/parse_context.h>
-#include <google/protobuf/io/coded_stream.h>
-#include <google/protobuf/io/zero_copy_stream_impl_lite.h>
-#include <google/protobuf/message_lite.h>
-#include <google/protobuf/port.h>
-
-#include <google/protobuf/port_def.inc>
-
-#ifdef SWIG
-#error "You cannot SWIG proto headers"
-#endif
-
-namespace google {
-namespace protobuf {
-namespace internal {
-class InternalMetadata;           // metadata_lite.h
-class WireFormat;                 // wire_format.h
-class MessageSetFieldSkipperUsingCord;
-// extension_set_heavy.cc
-}  // namespace internal
-
-class Message;       // message.h
-class UnknownField;  // below
-
-// An UnknownFieldSet contains fields that were encountered while parsing a
-// message but were not defined by its type.  Keeping track of these can be
-// useful, especially in that they may be written if the message is serialized
-// again without being cleared in between.  This means that software which
-// simply receives messages and forwards them to other servers does not need
-// to be updated every time a new field is added to the message definition.
-//
-// To get the UnknownFieldSet attached to any message, call
-// Reflection::GetUnknownFields().
-//
-// This class is necessarily tied to the protocol buffer wire format, unlike
-// the Reflection interface which is independent of any serialization scheme.
-class PROTOBUF_EXPORT UnknownFieldSet {
- public:
-  UnknownFieldSet();
-  ~UnknownFieldSet();
-
-  // Remove all fields.
-  inline void Clear();
-
-  // Remove all fields and deallocate internal data objects
-  void ClearAndFreeMemory();
-
-  // Is this set empty?
-  inline bool empty() const;
-
-  // Merge the contents of some other UnknownFieldSet with this one.
-  void MergeFrom(const UnknownFieldSet& other);
-
-  // Similar to above, but this function will destroy the contents of other.
-  void MergeFromAndDestroy(UnknownFieldSet* other);
-
-  // Merge the contents an UnknownFieldSet with the UnknownFieldSet in
-  // *metadata, if there is one.  If *metadata doesn't have an UnknownFieldSet
-  // then add one to it and make it be a copy of the first arg.
-  static void MergeToInternalMetadata(const UnknownFieldSet& other,
-                                      internal::InternalMetadata* metadata);
-
-  // Swaps the contents of some other UnknownFieldSet with this one.
-  inline void Swap(UnknownFieldSet* x);
-
-  // Computes (an estimate of) the total number of bytes currently used for
-  // storing the unknown fields in memory. Does NOT include
-  // sizeof(*this) in the calculation.
-  size_t SpaceUsedExcludingSelfLong() const;
-
-  int SpaceUsedExcludingSelf() const {
-    return internal::ToIntSize(SpaceUsedExcludingSelfLong());
-  }
-
-  // Version of SpaceUsed() including sizeof(*this).
-  size_t SpaceUsedLong() const;
-
-  int SpaceUsed() const { return internal::ToIntSize(SpaceUsedLong()); }
-
-  // Returns the number of fields present in the UnknownFieldSet.
-  inline int field_count() const;
-  // Get a field in the set, where 0 <= index < field_count().  The fields
-  // appear in the order in which they were added.
-  inline const UnknownField& field(int index) const;
-  // Get a mutable pointer to a field in the set, where
-  // 0 <= index < field_count().  The fields appear in the order in which
-  // they were added.
-  inline UnknownField* mutable_field(int index);
-
-  // Adding fields ---------------------------------------------------
-
-  void AddVarint(int number, uint64 value);
-  void AddFixed32(int number, uint32 value);
-  void AddFixed64(int number, uint64 value);
-  void AddLengthDelimited(int number, const std::string& value);
-  std::string* AddLengthDelimited(int number);
-  UnknownFieldSet* AddGroup(int number);
-
-  // Adds an unknown field from another set.
-  void AddField(const UnknownField& field);
-
-  // Delete fields with indices in the range [start .. start+num-1].
-  // Caution: implementation moves all fields with indices [start+num .. ].
-  void DeleteSubrange(int start, int num);
-
-  // Delete all fields with a specific field number. The order of left fields
-  // is preserved.
-  // Caution: implementation moves all fields after the first deleted field.
-  void DeleteByNumber(int number);
-
-  // Parsing helpers -------------------------------------------------
-  // These work exactly like the similarly-named methods of Message.
-
-  bool MergeFromCodedStream(io::CodedInputStream* input);
-  bool ParseFromCodedStream(io::CodedInputStream* input);
-  bool ParseFromZeroCopyStream(io::ZeroCopyInputStream* input);
-  bool ParseFromArray(const void* data, int size);
-  inline bool ParseFromString(const std::string& data) {
-    return ParseFromArray(data.data(), static_cast<int>(data.size()));
-  }
-
-  // Merges this message's unknown field data (if any).  This works whether
-  // the message is a lite or full proto (for legacy reasons, lite and full
-  // return different types for MessageType::unknown_fields()).
-  template <typename MessageType>
-  bool MergeFromMessage(const MessageType& message);
-
-  static const UnknownFieldSet& default_instance();
-
- private:
-  // For InternalMergeFrom
-  friend class UnknownField;
-  // Merges from other UnknownFieldSet. This method assumes, that this object
-  // is newly created and has no fields.
-  void InternalMergeFrom(const UnknownFieldSet& other);
-  void ClearFallback();
-
-  template <typename MessageType,
-            typename std::enable_if<
-                std::is_base_of<Message, MessageType>::value, int>::type = 0>
-  bool InternalMergeFromMessage(const MessageType& message) {
-    MergeFrom(message.GetReflection()->GetUnknownFields(message));
-    return true;
-  }
-
-  template <typename MessageType,
-            typename std::enable_if<
-                std::is_base_of<MessageLite, MessageType>::value &&
-                    !std::is_base_of<Message, MessageType>::value,
-                int>::type = 0>
-  bool InternalMergeFromMessage(const MessageType& message) {
-    const auto& unknown_fields = message.unknown_fields();
-    io::ArrayInputStream array_stream(unknown_fields.data(),
-                                      unknown_fields.size());
-    io::CodedInputStream coded_stream(&array_stream);
-    return MergeFromCodedStream(&coded_stream);
-  }
-
-  std::vector<UnknownField> fields_;
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(UnknownFieldSet);
-};
-
-namespace internal {
-
-inline void WriteVarint(uint32 num, uint64 val, UnknownFieldSet* unknown) {
-  unknown->AddVarint(num, val);
-}
-inline void WriteLengthDelimited(uint32 num, StringPiece val,
-                                 UnknownFieldSet* unknown) {
-  unknown->AddLengthDelimited(num)->assign(val.data(), val.size());
-}
-
-PROTOBUF_EXPORT
-const char* UnknownGroupParse(UnknownFieldSet* unknown, const char* ptr,
-                              ParseContext* ctx);
-PROTOBUF_EXPORT
-const char* UnknownFieldParse(uint64 tag, UnknownFieldSet* unknown,
-                              const char* ptr, ParseContext* ctx);
-
-}  // namespace internal
-
-// Represents one field in an UnknownFieldSet.
-class PROTOBUF_EXPORT UnknownField {
- public:
-  enum Type {
-    TYPE_VARINT,
-    TYPE_FIXED32,
-    TYPE_FIXED64,
-    TYPE_LENGTH_DELIMITED,
-    TYPE_GROUP
-  };
-
-  // The field's field number, as seen on the wire.
-  inline int number() const;
-
-  // The field type.
-  inline Type type() const;
-
-  // Accessors -------------------------------------------------------
-  // Each method works only for UnknownFields of the corresponding type.
-
-  inline uint64 varint() const;
-  inline uint32 fixed32() const;
-  inline uint64 fixed64() const;
-  inline const std::string& length_delimited() const;
-  inline const UnknownFieldSet& group() const;
-
-  inline void set_varint(uint64 value);
-  inline void set_fixed32(uint32 value);
-  inline void set_fixed64(uint64 value);
-  inline void set_length_delimited(const std::string& value);
-  inline std::string* mutable_length_delimited();
-  inline UnknownFieldSet* mutable_group();
-
-  // Serialization API.
-  // These methods can take advantage of the underlying implementation and may
-  // archieve a better performance than using getters to retrieve the data and
-  // do the serialization yourself.
-  void SerializeLengthDelimitedNoTag(io::CodedOutputStream* output) const {
-    output->SetCur(InternalSerializeLengthDelimitedNoTag(output->Cur(),
-                                                         output->EpsCopy()));
-  }
-
-  inline size_t GetLengthDelimitedSize() const;
-  uint8* InternalSerializeLengthDelimitedNoTag(
-      uint8* target, io::EpsCopyOutputStream* stream) const;
-
-
-  // If this UnknownField contains a pointer, delete it.
-  void Delete();
-
-  // Make a deep copy of any pointers in this UnknownField.
-  void DeepCopy(const UnknownField& other);
-
-  // Set the wire type of this UnknownField. Should only be used when this
-  // UnknownField is being created.
-  inline void SetType(Type type);
-
-  union LengthDelimited {
-    std::string* string_value;
-  };
-
-  uint32 number_;
-  uint32 type_;
-  union {
-    uint64 varint_;
-    uint32 fixed32_;
-    uint64 fixed64_;
-    mutable union LengthDelimited length_delimited_;
-    UnknownFieldSet* group_;
-  } data_;
-};
-
-// ===================================================================
-// inline implementations
-
-inline UnknownFieldSet::UnknownFieldSet() {}
-
-inline UnknownFieldSet::~UnknownFieldSet() { Clear(); }
-
-inline void UnknownFieldSet::ClearAndFreeMemory() { Clear(); }
-
-inline void UnknownFieldSet::Clear() {
-  if (!fields_.empty()) {
-    ClearFallback();
-  }
-}
-
-inline bool UnknownFieldSet::empty() const { return fields_.empty(); }
-
-inline void UnknownFieldSet::Swap(UnknownFieldSet* x) {
-  fields_.swap(x->fields_);
-}
-
-inline int UnknownFieldSet::field_count() const {
-  return static_cast<int>(fields_.size());
-}
-inline const UnknownField& UnknownFieldSet::field(int index) const {
-  return (fields_)[static_cast<size_t>(index)];
-}
-inline UnknownField* UnknownFieldSet::mutable_field(int index) {
-  return &(fields_)[static_cast<size_t>(index)];
-}
-
-inline void UnknownFieldSet::AddLengthDelimited(int number,
-                                                const std::string& value) {
-  AddLengthDelimited(number)->assign(value);
-}
-
-
-
-
-inline int UnknownField::number() const { return static_cast<int>(number_); }
-inline UnknownField::Type UnknownField::type() const {
-  return static_cast<Type>(type_);
-}
-
-inline uint64 UnknownField::varint() const {
-  assert(type() == TYPE_VARINT);
-  return data_.varint_;
-}
-inline uint32 UnknownField::fixed32() const {
-  assert(type() == TYPE_FIXED32);
-  return data_.fixed32_;
-}
-inline uint64 UnknownField::fixed64() const {
-  assert(type() == TYPE_FIXED64);
-  return data_.fixed64_;
-}
-inline const std::string& UnknownField::length_delimited() const {
-  assert(type() == TYPE_LENGTH_DELIMITED);
-  return *data_.length_delimited_.string_value;
-}
-inline const UnknownFieldSet& UnknownField::group() const {
-  assert(type() == TYPE_GROUP);
-  return *data_.group_;
-}
-
-inline void UnknownField::set_varint(uint64 value) {
-  assert(type() == TYPE_VARINT);
-  data_.varint_ = value;
-}
-inline void UnknownField::set_fixed32(uint32 value) {
-  assert(type() == TYPE_FIXED32);
-  data_.fixed32_ = value;
-}
-inline void UnknownField::set_fixed64(uint64 value) {
-  assert(type() == TYPE_FIXED64);
-  data_.fixed64_ = value;
-}
-inline void UnknownField::set_length_delimited(const std::string& value) {
-  assert(type() == TYPE_LENGTH_DELIMITED);
-  data_.length_delimited_.string_value->assign(value);
-}
-inline std::string* UnknownField::mutable_length_delimited() {
-  assert(type() == TYPE_LENGTH_DELIMITED);
-  return data_.length_delimited_.string_value;
-}
-inline UnknownFieldSet* UnknownField::mutable_group() {
-  assert(type() == TYPE_GROUP);
-  return data_.group_;
-}
-template <typename MessageType>
-bool UnknownFieldSet::MergeFromMessage(const MessageType& message) {
-  // SFINAE will route to the right version.
-  return InternalMergeFromMessage(message);
-}
-
-
-inline size_t UnknownField::GetLengthDelimitedSize() const {
-  GOOGLE_DCHECK_EQ(TYPE_LENGTH_DELIMITED, type());
-  return data_.length_delimited_.string_value->size();
-}
-
-inline void UnknownField::SetType(Type type) {
-  type_ = type;
-}
-
-
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-#endif  // GOOGLE_PROTOBUF_UNKNOWN_FIELD_SET_H__
diff --git a/projects/llm_framework/include/protobuf-lite/google/protobuf/wire_format_lite.h b/projects/llm_framework/include/protobuf-lite/google/protobuf/wire_format_lite.h
deleted file mode 100644
index c742fe86..00000000
--- a/projects/llm_framework/include/protobuf-lite/google/protobuf/wire_format_lite.h
+++ /dev/null
@@ -1,1866 +0,0 @@
-// Protocol Buffers - Google's data interchange format
-// Copyright 2008 Google Inc.  All rights reserved.
-// https://developers.google.com/protocol-buffers/
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// Author: kenton@google.com (Kenton Varda)
-//         atenasio@google.com (Chris Atenasio) (ZigZag transform)
-//         wink@google.com (Wink Saville) (refactored from wire_format.h)
-//  Based on original Protocol Buffers design by
-//  Sanjay Ghemawat, Jeff Dean, and others.
-//
-// This header is logically internal, but is made public because it is used
-// from protocol-compiler-generated code, which may reside in other components.
-
-#ifndef GOOGLE_PROTOBUF_WIRE_FORMAT_LITE_H__
-#define GOOGLE_PROTOBUF_WIRE_FORMAT_LITE_H__
-
-#include <string>
-
-#include <google/protobuf/stubs/common.h>
-#include <google/protobuf/stubs/logging.h>
-#include <google/protobuf/io/coded_stream.h>
-#include <google/protobuf/arenastring.h>
-#include <google/protobuf/message_lite.h>
-#include <google/protobuf/port.h>
-#include <google/protobuf/repeated_field.h>
-#include <google/protobuf/stubs/casts.h>
-
-// Do UTF-8 validation on string type in Debug build only
-#ifndef NDEBUG
-#define GOOGLE_PROTOBUF_UTF8_VALIDATION_ENABLED
-#endif
-
-// Avoid conflict with iOS where <ConditionalMacros.h> #defines TYPE_BOOL.
-//
-// If some one needs the macro TYPE_BOOL in a file that includes this header,
-// it's possible to bring it back using push/pop_macro as follows.
-//
-// #pragma push_macro("TYPE_BOOL")
-// #include this header and/or all headers that need the macro to be undefined.
-// #pragma pop_macro("TYPE_BOOL")
-#undef TYPE_BOOL
-
-
-namespace google {
-namespace protobuf {
-namespace internal {
-
-#include <google/protobuf/port_def.inc>
-
-// This class is for internal use by the protocol buffer library and by
-// protocol-compiler-generated message classes.  It must not be called
-// directly by clients.
-//
-// This class contains helpers for implementing the binary protocol buffer
-// wire format without the need for reflection. Use WireFormat when using
-// reflection.
-//
-// This class is really a namespace that contains only static methods.
-class PROTOBUF_EXPORT WireFormatLite {
- public:
-  // -----------------------------------------------------------------
-  // Helper constants and functions related to the format.  These are
-  // mostly meant for internal and generated code to use.
-
-  // The wire format is composed of a sequence of tag/value pairs, each
-  // of which contains the value of one field (or one element of a repeated
-  // field).  Each tag is encoded as a varint.  The lower bits of the tag
-  // identify its wire type, which specifies the format of the data to follow.
-  // The rest of the bits contain the field number.  Each type of field (as
-  // declared by FieldDescriptor::Type, in descriptor.h) maps to one of
-  // these wire types.  Immediately following each tag is the field's value,
-  // encoded in the format specified by the wire type.  Because the tag
-  // identifies the encoding of this data, it is possible to skip
-  // unrecognized fields for forwards compatibility.
-
-  enum WireType {
-    WIRETYPE_VARINT = 0,
-    WIRETYPE_FIXED64 = 1,
-    WIRETYPE_LENGTH_DELIMITED = 2,
-    WIRETYPE_START_GROUP = 3,
-    WIRETYPE_END_GROUP = 4,
-    WIRETYPE_FIXED32 = 5,
-  };
-
-  // Lite alternative to FieldDescriptor::Type.  Must be kept in sync.
-  enum FieldType {
-    TYPE_DOUBLE = 1,
-    TYPE_FLOAT = 2,
-    TYPE_INT64 = 3,
-    TYPE_UINT64 = 4,
-    TYPE_INT32 = 5,
-    TYPE_FIXED64 = 6,
-    TYPE_FIXED32 = 7,
-    TYPE_BOOL = 8,
-    TYPE_STRING = 9,
-    TYPE_GROUP = 10,
-    TYPE_MESSAGE = 11,
-    TYPE_BYTES = 12,
-    TYPE_UINT32 = 13,
-    TYPE_ENUM = 14,
-    TYPE_SFIXED32 = 15,
-    TYPE_SFIXED64 = 16,
-    TYPE_SINT32 = 17,
-    TYPE_SINT64 = 18,
-    MAX_FIELD_TYPE = 18,
-  };
-
-  // Lite alternative to FieldDescriptor::CppType.  Must be kept in sync.
-  enum CppType {
-    CPPTYPE_INT32 = 1,
-    CPPTYPE_INT64 = 2,
-    CPPTYPE_UINT32 = 3,
-    CPPTYPE_UINT64 = 4,
-    CPPTYPE_DOUBLE = 5,
-    CPPTYPE_FLOAT = 6,
-    CPPTYPE_BOOL = 7,
-    CPPTYPE_ENUM = 8,
-    CPPTYPE_STRING = 9,
-    CPPTYPE_MESSAGE = 10,
-    MAX_CPPTYPE = 10,
-  };
-
-  // Helper method to get the CppType for a particular Type.
-  static CppType FieldTypeToCppType(FieldType type);
-
-  // Given a FieldDescriptor::Type return its WireType
-  static inline WireFormatLite::WireType WireTypeForFieldType(
-      WireFormatLite::FieldType type) {
-    return kWireTypeForFieldType[type];
-  }
-
-  // Number of bits in a tag which identify the wire type.
-  static constexpr int kTagTypeBits = 3;
-  // Mask for those bits.
-  static constexpr uint32 kTagTypeMask = (1 << kTagTypeBits) - 1;
-
-  // Helper functions for encoding and decoding tags.  (Inlined below and in
-  // _inl.h)
-  //
-  // This is different from MakeTag(field->number(), field->type()) in the
-  // case of packed repeated fields.
-  constexpr static uint32 MakeTag(int field_number, WireType type);
-  static WireType GetTagWireType(uint32 tag);
-  static int GetTagFieldNumber(uint32 tag);
-
-  // Compute the byte size of a tag.  For groups, this includes both the start
-  // and end tags.
-  static inline size_t TagSize(int field_number,
-                               WireFormatLite::FieldType type);
-
-  // Skips a field value with the given tag.  The input should start
-  // positioned immediately after the tag.  Skipped values are simply
-  // discarded, not recorded anywhere.  See WireFormat::SkipField() for a
-  // version that records to an UnknownFieldSet.
-  static bool SkipField(io::CodedInputStream* input, uint32 tag);
-
-  // Skips a field value with the given tag.  The input should start
-  // positioned immediately after the tag. Skipped values are recorded to a
-  // CodedOutputStream.
-  static bool SkipField(io::CodedInputStream* input, uint32 tag,
-                        io::CodedOutputStream* output);
-
-  // Reads and ignores a message from the input.  Skipped values are simply
-  // discarded, not recorded anywhere.  See WireFormat::SkipMessage() for a
-  // version that records to an UnknownFieldSet.
-  static bool SkipMessage(io::CodedInputStream* input);
-
-  // Reads and ignores a message from the input.  Skipped values are recorded
-  // to a CodedOutputStream.
-  static bool SkipMessage(io::CodedInputStream* input,
-                          io::CodedOutputStream* output);
-
-  // This macro does the same thing as WireFormatLite::MakeTag(), but the
-  // result is usable as a compile-time constant, which makes it usable
-  // as a switch case or a template input.  WireFormatLite::MakeTag() is more
-  // type-safe, though, so prefer it if possible.
-#define GOOGLE_PROTOBUF_WIRE_FORMAT_MAKE_TAG(FIELD_NUMBER, TYPE) \
-  static_cast<uint32>((static_cast<uint32>(FIELD_NUMBER) << 3) | (TYPE))
-
-  // These are the tags for the old MessageSet format, which was defined as:
-  //   message MessageSet {
-  //     repeated group Item = 1 {
-  //       required int32 type_id = 2;
-  //       required string message = 3;
-  //     }
-  //   }
-  static constexpr int kMessageSetItemNumber = 1;
-  static constexpr int kMessageSetTypeIdNumber = 2;
-  static constexpr int kMessageSetMessageNumber = 3;
-  static const int kMessageSetItemStartTag = GOOGLE_PROTOBUF_WIRE_FORMAT_MAKE_TAG(
-      kMessageSetItemNumber, WireFormatLite::WIRETYPE_START_GROUP);
-  static const int kMessageSetItemEndTag = GOOGLE_PROTOBUF_WIRE_FORMAT_MAKE_TAG(
-      kMessageSetItemNumber, WireFormatLite::WIRETYPE_END_GROUP);
-  static const int kMessageSetTypeIdTag = GOOGLE_PROTOBUF_WIRE_FORMAT_MAKE_TAG(
-      kMessageSetTypeIdNumber, WireFormatLite::WIRETYPE_VARINT);
-  static const int kMessageSetMessageTag = GOOGLE_PROTOBUF_WIRE_FORMAT_MAKE_TAG(
-      kMessageSetMessageNumber, WireFormatLite::WIRETYPE_LENGTH_DELIMITED);
-
-  // Byte size of all tags of a MessageSet::Item combined.
-  static const size_t kMessageSetItemTagsSize;
-
-  // Helper functions for converting between floats/doubles and IEEE-754
-  // uint32s/uint64s so that they can be written.  (Assumes your platform
-  // uses IEEE-754 floats.)
-  static uint32 EncodeFloat(float value);
-  static float DecodeFloat(uint32 value);
-  static uint64 EncodeDouble(double value);
-  static double DecodeDouble(uint64 value);
-
-  // Helper functions for mapping signed integers to unsigned integers in
-  // such a way that numbers with small magnitudes will encode to smaller
-  // varints.  If you simply static_cast a negative number to an unsigned
-  // number and varint-encode it, it will always take 10 bytes, defeating
-  // the purpose of varint.  So, for the "sint32" and "sint64" field types,
-  // we ZigZag-encode the values.
-  static uint32 ZigZagEncode32(int32 n);
-  static int32 ZigZagDecode32(uint32 n);
-  static uint64 ZigZagEncode64(int64 n);
-  static int64 ZigZagDecode64(uint64 n);
-
-  // =================================================================
-  // Methods for reading/writing individual field.
-
-  // Read fields, not including tags.  The assumption is that you already
-  // read the tag to determine what field to read.
-
-  // For primitive fields, we just use a templatized routine parameterized by
-  // the represented type and the FieldType. These are specialized with the
-  // appropriate definition for each declared type.
-  template <typename CType, enum FieldType DeclaredType>
-  PROTOBUF_ALWAYS_INLINE static bool ReadPrimitive(io::CodedInputStream* input,
-                                                   CType* value);
-
-  // Reads repeated primitive values, with optimizations for repeats.
-  // tag_size and tag should both be compile-time constants provided by the
-  // protocol compiler.
-  template <typename CType, enum FieldType DeclaredType>
-  PROTOBUF_ALWAYS_INLINE static bool ReadRepeatedPrimitive(
-      int tag_size, uint32 tag, io::CodedInputStream* input,
-      RepeatedField<CType>* value);
-
-  // Identical to ReadRepeatedPrimitive, except will not inline the
-  // implementation.
-  template <typename CType, enum FieldType DeclaredType>
-  static bool ReadRepeatedPrimitiveNoInline(int tag_size, uint32 tag,
-                                            io::CodedInputStream* input,
-                                            RepeatedField<CType>* value);
-
-  // Reads a primitive value directly from the provided buffer. It returns a
-  // pointer past the segment of data that was read.
-  //
-  // This is only implemented for the types with fixed wire size, e.g.
-  // float, double, and the (s)fixed* types.
-  template <typename CType, enum FieldType DeclaredType>
-  PROTOBUF_ALWAYS_INLINE static const uint8* ReadPrimitiveFromArray(
-      const uint8* buffer, CType* value);
-
-  // Reads a primitive packed field.
-  //
-  // This is only implemented for packable types.
-  template <typename CType, enum FieldType DeclaredType>
-  PROTOBUF_ALWAYS_INLINE static bool ReadPackedPrimitive(
-      io::CodedInputStream* input, RepeatedField<CType>* value);
-
-  // Identical to ReadPackedPrimitive, except will not inline the
-  // implementation.
-  template <typename CType, enum FieldType DeclaredType>
-  static bool ReadPackedPrimitiveNoInline(io::CodedInputStream* input,
-                                          RepeatedField<CType>* value);
-
-  // Read a packed enum field. If the is_valid function is not NULL, values for
-  // which is_valid(value) returns false are silently dropped.
-  static bool ReadPackedEnumNoInline(io::CodedInputStream* input,
-                                     bool (*is_valid)(int),
-                                     RepeatedField<int>* values);
-
-  // Read a packed enum field. If the is_valid function is not NULL, values for
-  // which is_valid(value) returns false are appended to unknown_fields_stream.
-  static bool ReadPackedEnumPreserveUnknowns(
-      io::CodedInputStream* input, int field_number, bool (*is_valid)(int),
-      io::CodedOutputStream* unknown_fields_stream, RepeatedField<int>* values);
-
-  // Read a string.  ReadString(..., std::string* value) requires an
-  // existing std::string.
-  static inline bool ReadString(io::CodedInputStream* input,
-                                std::string* value);
-  // ReadString(..., std::string** p) is internal-only, and should only be
-  // called from generated code. It starts by setting *p to "new std::string" if
-  // *p == &GetEmptyStringAlreadyInited().  It then invokes
-  // ReadString(io::CodedInputStream* input, *p).  This is useful for reducing
-  // code size.
-  static inline bool ReadString(io::CodedInputStream* input, std::string** p);
-  // Analogous to ReadString().
-  static bool ReadBytes(io::CodedInputStream* input, std::string* value);
-  static bool ReadBytes(io::CodedInputStream* input, std::string** p);
-
-  enum Operation {
-    PARSE = 0,
-    SERIALIZE = 1,
-  };
-
-  // Returns true if the data is valid UTF-8.
-  static bool VerifyUtf8String(const char* data, int size, Operation op,
-                               const char* field_name);
-
-  template <typename MessageType>
-  static inline bool ReadGroup(int field_number, io::CodedInputStream* input,
-                               MessageType* value);
-
-  template <typename MessageType>
-  static inline bool ReadMessage(io::CodedInputStream* input,
-                                 MessageType* value);
-
-  template <typename MessageType>
-  static inline bool ReadMessageNoVirtual(io::CodedInputStream* input,
-                                          MessageType* value) {
-    return ReadMessage(input, value);
-  }
-
-  // Write a tag.  The Write*() functions typically include the tag, so
-  // normally there's no need to call this unless using the Write*NoTag()
-  // variants.
-  PROTOBUF_ALWAYS_INLINE static void WriteTag(int field_number, WireType type,
-                                              io::CodedOutputStream* output);
-
-  // Write fields, without tags.
-  PROTOBUF_ALWAYS_INLINE static void WriteInt32NoTag(
-      int32 value, io::CodedOutputStream* output);
-  PROTOBUF_ALWAYS_INLINE static void WriteInt64NoTag(
-      int64 value, io::CodedOutputStream* output);
-  PROTOBUF_ALWAYS_INLINE static void WriteUInt32NoTag(
-      uint32 value, io::CodedOutputStream* output);
-  PROTOBUF_ALWAYS_INLINE static void WriteUInt64NoTag(
-      uint64 value, io::CodedOutputStream* output);
-  PROTOBUF_ALWAYS_INLINE static void WriteSInt32NoTag(
-      int32 value, io::CodedOutputStream* output);
-  PROTOBUF_ALWAYS_INLINE static void WriteSInt64NoTag(
-      int64 value, io::CodedOutputStream* output);
-  PROTOBUF_ALWAYS_INLINE static void WriteFixed32NoTag(
-      uint32 value, io::CodedOutputStream* output);
-  PROTOBUF_ALWAYS_INLINE static void WriteFixed64NoTag(
-      uint64 value, io::CodedOutputStream* output);
-  PROTOBUF_ALWAYS_INLINE static void WriteSFixed32NoTag(
-      int32 value, io::CodedOutputStream* output);
-  PROTOBUF_ALWAYS_INLINE static void WriteSFixed64NoTag(
-      int64 value, io::CodedOutputStream* output);
-  PROTOBUF_ALWAYS_INLINE static void WriteFloatNoTag(
-      float value, io::CodedOutputStream* output);
-  PROTOBUF_ALWAYS_INLINE static void WriteDoubleNoTag(
-      double value, io::CodedOutputStream* output);
-  PROTOBUF_ALWAYS_INLINE static void WriteBoolNoTag(
-      bool value, io::CodedOutputStream* output);
-  PROTOBUF_ALWAYS_INLINE static void WriteEnumNoTag(
-      int value, io::CodedOutputStream* output);
-
-  // Write array of primitive fields, without tags
-  static void WriteFloatArray(const float* a, int n,
-                              io::CodedOutputStream* output);
-  static void WriteDoubleArray(const double* a, int n,
-                               io::CodedOutputStream* output);
-  static void WriteFixed32Array(const uint32* a, int n,
-                                io::CodedOutputStream* output);
-  static void WriteFixed64Array(const uint64* a, int n,
-                                io::CodedOutputStream* output);
-  static void WriteSFixed32Array(const int32* a, int n,
-                                 io::CodedOutputStream* output);
-  static void WriteSFixed64Array(const int64* a, int n,
-                                 io::CodedOutputStream* output);
-  static void WriteBoolArray(const bool* a, int n,
-                             io::CodedOutputStream* output);
-
-  // Write fields, including tags.
-  static void WriteInt32(int field_number, int32 value,
-                         io::CodedOutputStream* output);
-  static void WriteInt64(int field_number, int64 value,
-                         io::CodedOutputStream* output);
-  static void WriteUInt32(int field_number, uint32 value,
-                          io::CodedOutputStream* output);
-  static void WriteUInt64(int field_number, uint64 value,
-                          io::CodedOutputStream* output);
-  static void WriteSInt32(int field_number, int32 value,
-                          io::CodedOutputStream* output);
-  static void WriteSInt64(int field_number, int64 value,
-                          io::CodedOutputStream* output);
-  static void WriteFixed32(int field_number, uint32 value,
-                           io::CodedOutputStream* output);
-  static void WriteFixed64(int field_number, uint64 value,
-                           io::CodedOutputStream* output);
-  static void WriteSFixed32(int field_number, int32 value,
-                            io::CodedOutputStream* output);
-  static void WriteSFixed64(int field_number, int64 value,
-                            io::CodedOutputStream* output);
-  static void WriteFloat(int field_number, float value,
-                         io::CodedOutputStream* output);
-  static void WriteDouble(int field_number, double value,
-                          io::CodedOutputStream* output);
-  static void WriteBool(int field_number, bool value,
-                        io::CodedOutputStream* output);
-  static void WriteEnum(int field_number, int value,
-                        io::CodedOutputStream* output);
-
-  static void WriteString(int field_number, const std::string& value,
-                          io::CodedOutputStream* output);
-  static void WriteBytes(int field_number, const std::string& value,
-                         io::CodedOutputStream* output);
-  static void WriteStringMaybeAliased(int field_number,
-                                      const std::string& value,
-                                      io::CodedOutputStream* output);
-  static void WriteBytesMaybeAliased(int field_number, const std::string& value,
-                                     io::CodedOutputStream* output);
-
-  static void WriteGroup(int field_number, const MessageLite& value,
-                         io::CodedOutputStream* output);
-  static void WriteMessage(int field_number, const MessageLite& value,
-                           io::CodedOutputStream* output);
-  // Like above, but these will check if the output stream has enough
-  // space to write directly to a flat array.
-  static void WriteGroupMaybeToArray(int field_number, const MessageLite& value,
-                                     io::CodedOutputStream* output);
-  static void WriteMessageMaybeToArray(int field_number,
-                                       const MessageLite& value,
-                                       io::CodedOutputStream* output);
-
-  // Like above, but de-virtualize the call to SerializeWithCachedSizes().  The
-  // pointer must point at an instance of MessageType, *not* a subclass (or
-  // the subclass must not override SerializeWithCachedSizes()).
-  template <typename MessageType>
-  static inline void WriteGroupNoVirtual(int field_number,
-                                         const MessageType& value,
-                                         io::CodedOutputStream* output);
-  template <typename MessageType>
-  static inline void WriteMessageNoVirtual(int field_number,
-                                           const MessageType& value,
-                                           io::CodedOutputStream* output);
-
-  // Like above, but use only *ToArray methods of CodedOutputStream.
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteTagToArray(int field_number,
-                                                       WireType type,
-                                                       uint8* target);
-
-  // Write fields, without tags.
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteInt32NoTagToArray(int32 value,
-                                                              uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteInt64NoTagToArray(int64 value,
-                                                              uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteUInt32NoTagToArray(uint32 value,
-                                                               uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteUInt64NoTagToArray(uint64 value,
-                                                               uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSInt32NoTagToArray(int32 value,
-                                                               uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSInt64NoTagToArray(int64 value,
-                                                               uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteFixed32NoTagToArray(uint32 value,
-                                                                uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteFixed64NoTagToArray(uint64 value,
-                                                                uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSFixed32NoTagToArray(int32 value,
-                                                                 uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSFixed64NoTagToArray(int64 value,
-                                                                 uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteFloatNoTagToArray(float value,
-                                                              uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteDoubleNoTagToArray(double value,
-                                                               uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteBoolNoTagToArray(bool value,
-                                                             uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteEnumNoTagToArray(int value,
-                                                             uint8* target);
-
-  // Write fields, without tags.  These require that value.size() > 0.
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE static uint8* WritePrimitiveNoTagToArray(
-      const RepeatedField<T>& value, uint8* (*Writer)(T, uint8*),
-      uint8* target);
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteFixedNoTagToArray(
-      const RepeatedField<T>& value, uint8* (*Writer)(T, uint8*),
-      uint8* target);
-
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteInt32NoTagToArray(
-      const RepeatedField<int32>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteInt64NoTagToArray(
-      const RepeatedField<int64>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteUInt32NoTagToArray(
-      const RepeatedField<uint32>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteUInt64NoTagToArray(
-      const RepeatedField<uint64>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSInt32NoTagToArray(
-      const RepeatedField<int32>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSInt64NoTagToArray(
-      const RepeatedField<int64>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteFixed32NoTagToArray(
-      const RepeatedField<uint32>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteFixed64NoTagToArray(
-      const RepeatedField<uint64>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSFixed32NoTagToArray(
-      const RepeatedField<int32>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSFixed64NoTagToArray(
-      const RepeatedField<int64>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteFloatNoTagToArray(
-      const RepeatedField<float>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteDoubleNoTagToArray(
-      const RepeatedField<double>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteBoolNoTagToArray(
-      const RepeatedField<bool>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteEnumNoTagToArray(
-      const RepeatedField<int>& value, uint8* output);
-
-  // Write fields, including tags.
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteInt32ToArray(int field_number,
-                                                         int32 value,
-                                                         uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteInt64ToArray(int field_number,
-                                                         int64 value,
-                                                         uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteUInt32ToArray(int field_number,
-                                                          uint32 value,
-                                                          uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteUInt64ToArray(int field_number,
-                                                          uint64 value,
-                                                          uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSInt32ToArray(int field_number,
-                                                          int32 value,
-                                                          uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSInt64ToArray(int field_number,
-                                                          int64 value,
-                                                          uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteFixed32ToArray(int field_number,
-                                                           uint32 value,
-                                                           uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteFixed64ToArray(int field_number,
-                                                           uint64 value,
-                                                           uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSFixed32ToArray(int field_number,
-                                                            int32 value,
-                                                            uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSFixed64ToArray(int field_number,
-                                                            int64 value,
-                                                            uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteFloatToArray(int field_number,
-                                                         float value,
-                                                         uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteDoubleToArray(int field_number,
-                                                          double value,
-                                                          uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteBoolToArray(int field_number,
-                                                        bool value,
-                                                        uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteEnumToArray(int field_number,
-                                                        int value,
-                                                        uint8* target);
-
-  template <typename T>
-  PROTOBUF_ALWAYS_INLINE static uint8* WritePrimitiveToArray(
-      int field_number, const RepeatedField<T>& value,
-      uint8* (*Writer)(int, T, uint8*), uint8* target);
-
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteInt32ToArray(
-      int field_number, const RepeatedField<int32>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteInt64ToArray(
-      int field_number, const RepeatedField<int64>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteUInt32ToArray(
-      int field_number, const RepeatedField<uint32>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteUInt64ToArray(
-      int field_number, const RepeatedField<uint64>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSInt32ToArray(
-      int field_number, const RepeatedField<int32>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSInt64ToArray(
-      int field_number, const RepeatedField<int64>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteFixed32ToArray(
-      int field_number, const RepeatedField<uint32>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteFixed64ToArray(
-      int field_number, const RepeatedField<uint64>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSFixed32ToArray(
-      int field_number, const RepeatedField<int32>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteSFixed64ToArray(
-      int field_number, const RepeatedField<int64>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteFloatToArray(
-      int field_number, const RepeatedField<float>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteDoubleToArray(
-      int field_number, const RepeatedField<double>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteBoolToArray(
-      int field_number, const RepeatedField<bool>& value, uint8* output);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteEnumToArray(
-      int field_number, const RepeatedField<int>& value, uint8* output);
-
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteStringToArray(
-      int field_number, const std::string& value, uint8* target);
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteBytesToArray(
-      int field_number, const std::string& value, uint8* target);
-
-  // Whether to serialize deterministically (e.g., map keys are
-  // sorted) is a property of a CodedOutputStream, and in the process
-  // of serialization, the "ToArray" variants may be invoked.  But they don't
-  // have a CodedOutputStream available, so they get an additional parameter
-  // telling them whether to serialize deterministically.
-  template <typename MessageType>
-  PROTOBUF_ALWAYS_INLINE static uint8* InternalWriteGroup(
-      int field_number, const MessageType& value, uint8* target,
-      io::EpsCopyOutputStream* stream);
-  template <typename MessageType>
-  PROTOBUF_ALWAYS_INLINE static uint8* InternalWriteMessage(
-      int field_number, const MessageType& value, uint8* target,
-      io::EpsCopyOutputStream* stream);
-
-  // Like above, but de-virtualize the call to SerializeWithCachedSizes().  The
-  // pointer must point at an instance of MessageType, *not* a subclass (or
-  // the subclass must not override SerializeWithCachedSizes()).
-  template <typename MessageType>
-  PROTOBUF_ALWAYS_INLINE static uint8* InternalWriteGroupNoVirtualToArray(
-      int field_number, const MessageType& value, uint8* target);
-  template <typename MessageType>
-  PROTOBUF_ALWAYS_INLINE static uint8* InternalWriteMessageNoVirtualToArray(
-      int field_number, const MessageType& value, uint8* target);
-
-  // For backward-compatibility, the last four methods also have versions
-  // that are non-deterministic always.
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteGroupToArray(
-      int field_number, const MessageLite& value, uint8* target) {
-    io::EpsCopyOutputStream stream(
-        target,
-        value.GetCachedSize() +
-            static_cast<int>(2 * io::CodedOutputStream::VarintSize32(
-                                     static_cast<uint32>(field_number) << 3)),
-        io::CodedOutputStream::IsDefaultSerializationDeterministic());
-    return InternalWriteGroup(field_number, value, target, &stream);
-  }
-  PROTOBUF_ALWAYS_INLINE static uint8* WriteMessageToArray(
-      int field_number, const MessageLite& value, uint8* target) {
-    int size = value.GetCachedSize();
-    io::EpsCopyOutputStream stream(
-        target,
-        size + static_cast<int>(io::CodedOutputStream::VarintSize32(
-                                    static_cast<uint32>(field_number) << 3) +
-                                io::CodedOutputStream::VarintSize32(size)),
-        io::CodedOutputStream::IsDefaultSerializationDeterministic());
-    return InternalWriteMessage(field_number, value, target, &stream);
-  }
-
-  // Compute the byte size of a field.  The XxSize() functions do NOT include
-  // the tag, so you must also call TagSize().  (This is because, for repeated
-  // fields, you should only call TagSize() once and multiply it by the element
-  // count, but you may have to call XxSize() for each individual element.)
-  static inline size_t Int32Size(int32 value);
-  static inline size_t Int64Size(int64 value);
-  static inline size_t UInt32Size(uint32 value);
-  static inline size_t UInt64Size(uint64 value);
-  static inline size_t SInt32Size(int32 value);
-  static inline size_t SInt64Size(int64 value);
-  static inline size_t EnumSize(int value);
-
-  static size_t Int32Size(const RepeatedField<int32>& value);
-  static size_t Int64Size(const RepeatedField<int64>& value);
-  static size_t UInt32Size(const RepeatedField<uint32>& value);
-  static size_t UInt64Size(const RepeatedField<uint64>& value);
-  static size_t SInt32Size(const RepeatedField<int32>& value);
-  static size_t SInt64Size(const RepeatedField<int64>& value);
-  static size_t EnumSize(const RepeatedField<int>& value);
-
-  // These types always have the same size.
-  static constexpr size_t kFixed32Size = 4;
-  static constexpr size_t kFixed64Size = 8;
-  static constexpr size_t kSFixed32Size = 4;
-  static constexpr size_t kSFixed64Size = 8;
-  static constexpr size_t kFloatSize = 4;
-  static constexpr size_t kDoubleSize = 8;
-  static constexpr size_t kBoolSize = 1;
-
-  static inline size_t StringSize(const std::string& value);
-  static inline size_t BytesSize(const std::string& value);
-
-  template <typename MessageType>
-  static inline size_t GroupSize(const MessageType& value);
-  template <typename MessageType>
-  static inline size_t MessageSize(const MessageType& value);
-
-  // Like above, but de-virtualize the call to ByteSize().  The
-  // pointer must point at an instance of MessageType, *not* a subclass (or
-  // the subclass must not override ByteSize()).
-  template <typename MessageType>
-  static inline size_t GroupSizeNoVirtual(const MessageType& value);
-  template <typename MessageType>
-  static inline size_t MessageSizeNoVirtual(const MessageType& value);
-
-  // Given the length of data, calculate the byte size of the data on the
-  // wire if we encode the data as a length delimited field.
-  static inline size_t LengthDelimitedSize(size_t length);
-
- private:
-  // A helper method for the repeated primitive reader. This method has
-  // optimizations for primitive types that have fixed size on the wire, and
-  // can be read using potentially faster paths.
-  template <typename CType, enum FieldType DeclaredType>
-  PROTOBUF_ALWAYS_INLINE static bool ReadRepeatedFixedSizePrimitive(
-      int tag_size, uint32 tag, io::CodedInputStream* input,
-      RepeatedField<CType>* value);
-
-  // Like ReadRepeatedFixedSizePrimitive but for packed primitive fields.
-  template <typename CType, enum FieldType DeclaredType>
-  PROTOBUF_ALWAYS_INLINE static bool ReadPackedFixedSizePrimitive(
-      io::CodedInputStream* input, RepeatedField<CType>* value);
-
-  static const CppType kFieldTypeToCppTypeMap[];
-  static const WireFormatLite::WireType kWireTypeForFieldType[];
-  static void WriteSubMessageMaybeToArray(int size, const MessageLite& value,
-                                          io::CodedOutputStream* output);
-
-  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(WireFormatLite);
-};
-
-// A class which deals with unknown values.  The default implementation just
-// discards them.  WireFormat defines a subclass which writes to an
-// UnknownFieldSet.  This class is used by ExtensionSet::ParseField(), since
-// ExtensionSet is part of the lite library but UnknownFieldSet is not.
-class PROTOBUF_EXPORT FieldSkipper {
- public:
-  FieldSkipper() {}
-  virtual ~FieldSkipper() {}
-
-  // Skip a field whose tag has already been consumed.
-  virtual bool SkipField(io::CodedInputStream* input, uint32 tag);
-
-  // Skip an entire message or group, up to an end-group tag (which is consumed)
-  // or end-of-stream.
-  virtual bool SkipMessage(io::CodedInputStream* input);
-
-  // Deal with an already-parsed unrecognized enum value.  The default
-  // implementation does nothing, but the UnknownFieldSet-based implementation
-  // saves it as an unknown varint.
-  virtual void SkipUnknownEnum(int field_number, int value);
-};
-
-// Subclass of FieldSkipper which saves skipped fields to a CodedOutputStream.
-
-class PROTOBUF_EXPORT CodedOutputStreamFieldSkipper : public FieldSkipper {
- public:
-  explicit CodedOutputStreamFieldSkipper(io::CodedOutputStream* unknown_fields)
-      : unknown_fields_(unknown_fields) {}
-  ~CodedOutputStreamFieldSkipper() override {}
-
-  // implements FieldSkipper -----------------------------------------
-  bool SkipField(io::CodedInputStream* input, uint32 tag) override;
-  bool SkipMessage(io::CodedInputStream* input) override;
-  void SkipUnknownEnum(int field_number, int value) override;
-
- protected:
-  io::CodedOutputStream* unknown_fields_;
-};
-
-// inline methods ====================================================
-
-inline WireFormatLite::CppType WireFormatLite::FieldTypeToCppType(
-    FieldType type) {
-  return kFieldTypeToCppTypeMap[type];
-}
-
-constexpr inline uint32 WireFormatLite::MakeTag(int field_number,
-                                                WireType type) {
-  return GOOGLE_PROTOBUF_WIRE_FORMAT_MAKE_TAG(field_number, type);
-}
-
-inline WireFormatLite::WireType WireFormatLite::GetTagWireType(uint32 tag) {
-  return static_cast<WireType>(tag & kTagTypeMask);
-}
-
-inline int WireFormatLite::GetTagFieldNumber(uint32 tag) {
-  return static_cast<int>(tag >> kTagTypeBits);
-}
-
-inline size_t WireFormatLite::TagSize(int field_number,
-                                      WireFormatLite::FieldType type) {
-  size_t result = io::CodedOutputStream::VarintSize32(
-      static_cast<uint32>(field_number << kTagTypeBits));
-  if (type == TYPE_GROUP) {
-    // Groups have both a start and an end tag.
-    return result * 2;
-  } else {
-    return result;
-  }
-}
-
-inline uint32 WireFormatLite::EncodeFloat(float value) {
-  return bit_cast<uint32>(value);
-}
-
-inline float WireFormatLite::DecodeFloat(uint32 value) {
-  return bit_cast<float>(value);
-}
-
-inline uint64 WireFormatLite::EncodeDouble(double value) {
-  return bit_cast<uint64>(value);
-}
-
-inline double WireFormatLite::DecodeDouble(uint64 value) {
-  return bit_cast<double>(value);
-}
-
-// ZigZag Transform:  Encodes signed integers so that they can be
-// effectively used with varint encoding.
-//
-// varint operates on unsigned integers, encoding smaller numbers into
-// fewer bytes.  If you try to use it on a signed integer, it will treat
-// this number as a very large unsigned integer, which means that even
-// small signed numbers like -1 will take the maximum number of bytes
-// (10) to encode.  ZigZagEncode() maps signed integers to unsigned
-// in such a way that those with a small absolute value will have smaller
-// encoded values, making them appropriate for encoding using varint.
-//
-//       int32 ->     uint32
-// -------------------------
-//           0 ->          0
-//          -1 ->          1
-//           1 ->          2
-//          -2 ->          3
-//         ... ->        ...
-//  2147483647 -> 4294967294
-// -2147483648 -> 4294967295
-//
-//        >> encode >>
-//        << decode <<
-
-inline uint32 WireFormatLite::ZigZagEncode32(int32 n) {
-  // Note:  the right-shift must be arithmetic
-  // Note:  left shift must be unsigned because of overflow
-  return (static_cast<uint32>(n) << 1) ^ static_cast<uint32>(n >> 31);
-}
-
-inline int32 WireFormatLite::ZigZagDecode32(uint32 n) {
-  // Note:  Using unsigned types prevent undefined behavior
-  return static_cast<int32>((n >> 1) ^ (~(n & 1) + 1));
-}
-
-inline uint64 WireFormatLite::ZigZagEncode64(int64 n) {
-  // Note:  the right-shift must be arithmetic
-  // Note:  left shift must be unsigned because of overflow
-  return (static_cast<uint64>(n) << 1) ^ static_cast<uint64>(n >> 63);
-}
-
-inline int64 WireFormatLite::ZigZagDecode64(uint64 n) {
-  // Note:  Using unsigned types prevent undefined behavior
-  return static_cast<int64>((n >> 1) ^ (~(n & 1) + 1));
-}
-
-// String is for UTF-8 text only, but, even so, ReadString() can simply
-// call ReadBytes().
-
-inline bool WireFormatLite::ReadString(io::CodedInputStream* input,
-                                       std::string* value) {
-  return ReadBytes(input, value);
-}
-
-inline bool WireFormatLite::ReadString(io::CodedInputStream* input,
-                                       std::string** p) {
-  return ReadBytes(input, p);
-}
-
-inline uint8* InternalSerializeUnknownMessageSetItemsToArray(
-    const std::string& unknown_fields, uint8* target,
-    io::EpsCopyOutputStream* stream) {
-  return stream->WriteRaw(unknown_fields.data(),
-                          static_cast<int>(unknown_fields.size()), target);
-}
-
-inline size_t ComputeUnknownMessageSetItemsSize(
-    const std::string& unknown_fields) {
-  return unknown_fields.size();
-}
-
-// Implementation details of ReadPrimitive.
-
-template <>
-inline bool WireFormatLite::ReadPrimitive<int32, WireFormatLite::TYPE_INT32>(
-    io::CodedInputStream* input, int32* value) {
-  uint32 temp;
-  if (!input->ReadVarint32(&temp)) return false;
-  *value = static_cast<int32>(temp);
-  return true;
-}
-template <>
-inline bool WireFormatLite::ReadPrimitive<int64, WireFormatLite::TYPE_INT64>(
-    io::CodedInputStream* input, int64* value) {
-  uint64 temp;
-  if (!input->ReadVarint64(&temp)) return false;
-  *value = static_cast<int64>(temp);
-  return true;
-}
-template <>
-inline bool WireFormatLite::ReadPrimitive<uint32, WireFormatLite::TYPE_UINT32>(
-    io::CodedInputStream* input, uint32* value) {
-  return input->ReadVarint32(value);
-}
-template <>
-inline bool WireFormatLite::ReadPrimitive<uint64, WireFormatLite::TYPE_UINT64>(
-    io::CodedInputStream* input, uint64* value) {
-  return input->ReadVarint64(value);
-}
-template <>
-inline bool WireFormatLite::ReadPrimitive<int32, WireFormatLite::TYPE_SINT32>(
-    io::CodedInputStream* input, int32* value) {
-  uint32 temp;
-  if (!input->ReadVarint32(&temp)) return false;
-  *value = ZigZagDecode32(temp);
-  return true;
-}
-template <>
-inline bool WireFormatLite::ReadPrimitive<int64, WireFormatLite::TYPE_SINT64>(
-    io::CodedInputStream* input, int64* value) {
-  uint64 temp;
-  if (!input->ReadVarint64(&temp)) return false;
-  *value = ZigZagDecode64(temp);
-  return true;
-}
-template <>
-inline bool WireFormatLite::ReadPrimitive<uint32, WireFormatLite::TYPE_FIXED32>(
-    io::CodedInputStream* input, uint32* value) {
-  return input->ReadLittleEndian32(value);
-}
-template <>
-inline bool WireFormatLite::ReadPrimitive<uint64, WireFormatLite::TYPE_FIXED64>(
-    io::CodedInputStream* input, uint64* value) {
-  return input->ReadLittleEndian64(value);
-}
-template <>
-inline bool WireFormatLite::ReadPrimitive<int32, WireFormatLite::TYPE_SFIXED32>(
-    io::CodedInputStream* input, int32* value) {
-  uint32 temp;
-  if (!input->ReadLittleEndian32(&temp)) return false;
-  *value = static_cast<int32>(temp);
-  return true;
-}
-template <>
-inline bool WireFormatLite::ReadPrimitive<int64, WireFormatLite::TYPE_SFIXED64>(
-    io::CodedInputStream* input, int64* value) {
-  uint64 temp;
-  if (!input->ReadLittleEndian64(&temp)) return false;
-  *value = static_cast<int64>(temp);
-  return true;
-}
-template <>
-inline bool WireFormatLite::ReadPrimitive<float, WireFormatLite::TYPE_FLOAT>(
-    io::CodedInputStream* input, float* value) {
-  uint32 temp;
-  if (!input->ReadLittleEndian32(&temp)) return false;
-  *value = DecodeFloat(temp);
-  return true;
-}
-template <>
-inline bool WireFormatLite::ReadPrimitive<double, WireFormatLite::TYPE_DOUBLE>(
-    io::CodedInputStream* input, double* value) {
-  uint64 temp;
-  if (!input->ReadLittleEndian64(&temp)) return false;
-  *value = DecodeDouble(temp);
-  return true;
-}
-template <>
-inline bool WireFormatLite::ReadPrimitive<bool, WireFormatLite::TYPE_BOOL>(
-    io::CodedInputStream* input, bool* value) {
-  uint64 temp;
-  if (!input->ReadVarint64(&temp)) return false;
-  *value = temp != 0;
-  return true;
-}
-template <>
-inline bool WireFormatLite::ReadPrimitive<int, WireFormatLite::TYPE_ENUM>(
-    io::CodedInputStream* input, int* value) {
-  uint32 temp;
-  if (!input->ReadVarint32(&temp)) return false;
-  *value = static_cast<int>(temp);
-  return true;
-}
-
-template <>
-inline const uint8*
-WireFormatLite::ReadPrimitiveFromArray<uint32, WireFormatLite::TYPE_FIXED32>(
-    const uint8* buffer, uint32* value) {
-  return io::CodedInputStream::ReadLittleEndian32FromArray(buffer, value);
-}
-template <>
-inline const uint8*
-WireFormatLite::ReadPrimitiveFromArray<uint64, WireFormatLite::TYPE_FIXED64>(
-    const uint8* buffer, uint64* value) {
-  return io::CodedInputStream::ReadLittleEndian64FromArray(buffer, value);
-}
-template <>
-inline const uint8*
-WireFormatLite::ReadPrimitiveFromArray<int32, WireFormatLite::TYPE_SFIXED32>(
-    const uint8* buffer, int32* value) {
-  uint32 temp;
-  buffer = io::CodedInputStream::ReadLittleEndian32FromArray(buffer, &temp);
-  *value = static_cast<int32>(temp);
-  return buffer;
-}
-template <>
-inline const uint8*
-WireFormatLite::ReadPrimitiveFromArray<int64, WireFormatLite::TYPE_SFIXED64>(
-    const uint8* buffer, int64* value) {
-  uint64 temp;
-  buffer = io::CodedInputStream::ReadLittleEndian64FromArray(buffer, &temp);
-  *value = static_cast<int64>(temp);
-  return buffer;
-}
-template <>
-inline const uint8*
-WireFormatLite::ReadPrimitiveFromArray<float, WireFormatLite::TYPE_FLOAT>(
-    const uint8* buffer, float* value) {
-  uint32 temp;
-  buffer = io::CodedInputStream::ReadLittleEndian32FromArray(buffer, &temp);
-  *value = DecodeFloat(temp);
-  return buffer;
-}
-template <>
-inline const uint8*
-WireFormatLite::ReadPrimitiveFromArray<double, WireFormatLite::TYPE_DOUBLE>(
-    const uint8* buffer, double* value) {
-  uint64 temp;
-  buffer = io::CodedInputStream::ReadLittleEndian64FromArray(buffer, &temp);
-  *value = DecodeDouble(temp);
-  return buffer;
-}
-
-template <typename CType, enum WireFormatLite::FieldType DeclaredType>
-inline bool WireFormatLite::ReadRepeatedPrimitive(
-    int,  // tag_size, unused.
-    uint32 tag, io::CodedInputStream* input, RepeatedField<CType>* values) {
-  CType value;
-  if (!ReadPrimitive<CType, DeclaredType>(input, &value)) return false;
-  values->Add(value);
-  int elements_already_reserved = values->Capacity() - values->size();
-  while (elements_already_reserved > 0 && input->ExpectTag(tag)) {
-    if (!ReadPrimitive<CType, DeclaredType>(input, &value)) return false;
-    values->AddAlreadyReserved(value);
-    elements_already_reserved--;
-  }
-  return true;
-}
-
-template <typename CType, enum WireFormatLite::FieldType DeclaredType>
-inline bool WireFormatLite::ReadRepeatedFixedSizePrimitive(
-    int tag_size, uint32 tag, io::CodedInputStream* input,
-    RepeatedField<CType>* values) {
-  GOOGLE_DCHECK_EQ(UInt32Size(tag), static_cast<size_t>(tag_size));
-  CType value;
-  if (!ReadPrimitive<CType, DeclaredType>(input, &value)) return false;
-  values->Add(value);
-
-  // For fixed size values, repeated values can be read more quickly by
-  // reading directly from a raw array.
-  //
-  // We can get a tight loop by only reading as many elements as can be
-  // added to the RepeatedField without having to do any resizing. Additionally,
-  // we only try to read as many elements as are available from the current
-  // buffer space. Doing so avoids having to perform boundary checks when
-  // reading the value: the maximum number of elements that can be read is
-  // known outside of the loop.
-  const void* void_pointer;
-  int size;
-  input->GetDirectBufferPointerInline(&void_pointer, &size);
-  if (size > 0) {
-    const uint8* buffer = reinterpret_cast<const uint8*>(void_pointer);
-    // The number of bytes each type occupies on the wire.
-    const int per_value_size = tag_size + static_cast<int>(sizeof(value));
-
-    // parentheses around (std::min) prevents macro expansion of min(...)
-    int elements_available =
-        (std::min)(values->Capacity() - values->size(), size / per_value_size);
-    int num_read = 0;
-    while (num_read < elements_available &&
-           (buffer = io::CodedInputStream::ExpectTagFromArray(buffer, tag)) !=
-               NULL) {
-      buffer = ReadPrimitiveFromArray<CType, DeclaredType>(buffer, &value);
-      values->AddAlreadyReserved(value);
-      ++num_read;
-    }
-    const int read_bytes = num_read * per_value_size;
-    if (read_bytes > 0) {
-      input->Skip(read_bytes);
-    }
-  }
-  return true;
-}
-
-// Specializations of ReadRepeatedPrimitive for the fixed size types, which use
-// the optimized code path.
-#define READ_REPEATED_FIXED_SIZE_PRIMITIVE(CPPTYPE, DECLARED_TYPE)        \
-  template <>                                                             \
-  inline bool WireFormatLite::ReadRepeatedPrimitive<                      \
-      CPPTYPE, WireFormatLite::DECLARED_TYPE>(                            \
-      int tag_size, uint32 tag, io::CodedInputStream* input,              \
-      RepeatedField<CPPTYPE>* values) {                                   \
-    return ReadRepeatedFixedSizePrimitive<CPPTYPE,                        \
-                                          WireFormatLite::DECLARED_TYPE>( \
-        tag_size, tag, input, values);                                    \
-  }
-
-READ_REPEATED_FIXED_SIZE_PRIMITIVE(uint32, TYPE_FIXED32)
-READ_REPEATED_FIXED_SIZE_PRIMITIVE(uint64, TYPE_FIXED64)
-READ_REPEATED_FIXED_SIZE_PRIMITIVE(int32, TYPE_SFIXED32)
-READ_REPEATED_FIXED_SIZE_PRIMITIVE(int64, TYPE_SFIXED64)
-READ_REPEATED_FIXED_SIZE_PRIMITIVE(float, TYPE_FLOAT)
-READ_REPEATED_FIXED_SIZE_PRIMITIVE(double, TYPE_DOUBLE)
-
-#undef READ_REPEATED_FIXED_SIZE_PRIMITIVE
-
-template <typename CType, enum WireFormatLite::FieldType DeclaredType>
-bool WireFormatLite::ReadRepeatedPrimitiveNoInline(
-    int tag_size, uint32 tag, io::CodedInputStream* input,
-    RepeatedField<CType>* value) {
-  return ReadRepeatedPrimitive<CType, DeclaredType>(tag_size, tag, input,
-                                                    value);
-}
-
-template <typename CType, enum WireFormatLite::FieldType DeclaredType>
-inline bool WireFormatLite::ReadPackedPrimitive(io::CodedInputStream* input,
-                                                RepeatedField<CType>* values) {
-  int length;
-  if (!input->ReadVarintSizeAsInt(&length)) return false;
-  io::CodedInputStream::Limit limit = input->PushLimit(length);
-  while (input->BytesUntilLimit() > 0) {
-    CType value;
-    if (!ReadPrimitive<CType, DeclaredType>(input, &value)) return false;
-    values->Add(value);
-  }
-  input->PopLimit(limit);
-  return true;
-}
-
-template <typename CType, enum WireFormatLite::FieldType DeclaredType>
-inline bool WireFormatLite::ReadPackedFixedSizePrimitive(
-    io::CodedInputStream* input, RepeatedField<CType>* values) {
-  int length;
-  if (!input->ReadVarintSizeAsInt(&length)) return false;
-  const int old_entries = values->size();
-  const int new_entries = length / static_cast<int>(sizeof(CType));
-  const int new_bytes = new_entries * static_cast<int>(sizeof(CType));
-  if (new_bytes != length) return false;
-  // We would *like* to pre-allocate the buffer to write into (for
-  // speed), but *must* avoid performing a very large allocation due
-  // to a malicious user-supplied "length" above.  So we have a fast
-  // path that pre-allocates when the "length" is less than a bound.
-  // We determine the bound by calling BytesUntilTotalBytesLimit() and
-  // BytesUntilLimit().  These return -1 to mean "no limit set".
-  // There are four cases:
-  // TotalBytesLimit  Limit
-  // -1               -1     Use slow path.
-  // -1               >= 0   Use fast path if length <= Limit.
-  // >= 0             -1     Use slow path.
-  // >= 0             >= 0   Use fast path if length <= min(both limits).
-  int64 bytes_limit = input->BytesUntilTotalBytesLimit();
-  if (bytes_limit == -1) {
-    bytes_limit = input->BytesUntilLimit();
-  } else {
-    // parentheses around (std::min) prevents macro expansion of min(...)
-    bytes_limit =
-        (std::min)(bytes_limit, static_cast<int64>(input->BytesUntilLimit()));
-  }
-  if (bytes_limit >= new_bytes) {
-    // Fast-path that pre-allocates *values to the final size.
-#if defined(PROTOBUF_LITTLE_ENDIAN)
-    values->Resize(old_entries + new_entries, 0);
-    // values->mutable_data() may change after Resize(), so do this after:
-    void* dest = reinterpret_cast<void*>(values->mutable_data() + old_entries);
-    if (!input->ReadRaw(dest, new_bytes)) {
-      values->Truncate(old_entries);
-      return false;
-    }
-#else
-    values->Reserve(old_entries + new_entries);
-    CType value;
-    for (int i = 0; i < new_entries; ++i) {
-      if (!ReadPrimitive<CType, DeclaredType>(input, &value)) return false;
-      values->AddAlreadyReserved(value);
-    }
-#endif
-  } else {
-    // This is the slow-path case where "length" may be too large to
-    // safely allocate.  We read as much as we can into *values
-    // without pre-allocating "length" bytes.
-    CType value;
-    for (int i = 0; i < new_entries; ++i) {
-      if (!ReadPrimitive<CType, DeclaredType>(input, &value)) return false;
-      values->Add(value);
-    }
-  }
-  return true;
-}
-
-// Specializations of ReadPackedPrimitive for the fixed size types, which use
-// an optimized code path.
-#define READ_REPEATED_PACKED_FIXED_SIZE_PRIMITIVE(CPPTYPE, DECLARED_TYPE)      \
-  template <>                                                                  \
-  inline bool                                                                  \
-  WireFormatLite::ReadPackedPrimitive<CPPTYPE, WireFormatLite::DECLARED_TYPE>( \
-      io::CodedInputStream * input, RepeatedField<CPPTYPE> * values) {         \
-    return ReadPackedFixedSizePrimitive<CPPTYPE,                               \
-                                        WireFormatLite::DECLARED_TYPE>(        \
-        input, values);                                                        \
-  }
-
-READ_REPEATED_PACKED_FIXED_SIZE_PRIMITIVE(uint32, TYPE_FIXED32)
-READ_REPEATED_PACKED_FIXED_SIZE_PRIMITIVE(uint64, TYPE_FIXED64)
-READ_REPEATED_PACKED_FIXED_SIZE_PRIMITIVE(int32, TYPE_SFIXED32)
-READ_REPEATED_PACKED_FIXED_SIZE_PRIMITIVE(int64, TYPE_SFIXED64)
-READ_REPEATED_PACKED_FIXED_SIZE_PRIMITIVE(float, TYPE_FLOAT)
-READ_REPEATED_PACKED_FIXED_SIZE_PRIMITIVE(double, TYPE_DOUBLE)
-
-#undef READ_REPEATED_PACKED_FIXED_SIZE_PRIMITIVE
-
-template <typename CType, enum WireFormatLite::FieldType DeclaredType>
-bool WireFormatLite::ReadPackedPrimitiveNoInline(io::CodedInputStream* input,
-                                                 RepeatedField<CType>* values) {
-  return ReadPackedPrimitive<CType, DeclaredType>(input, values);
-}
-
-
-template <typename MessageType>
-inline bool WireFormatLite::ReadGroup(int field_number,
-                                      io::CodedInputStream* input,
-                                      MessageType* value) {
-  if (!input->IncrementRecursionDepth()) return false;
-  if (!value->MergePartialFromCodedStream(input)) return false;
-  input->UnsafeDecrementRecursionDepth();
-  // Make sure the last thing read was an end tag for this group.
-  if (!input->LastTagWas(MakeTag(field_number, WIRETYPE_END_GROUP))) {
-    return false;
-  }
-  return true;
-}
-template <typename MessageType>
-inline bool WireFormatLite::ReadMessage(io::CodedInputStream* input,
-                                        MessageType* value) {
-  int length;
-  if (!input->ReadVarintSizeAsInt(&length)) return false;
-  std::pair<io::CodedInputStream::Limit, int> p =
-      input->IncrementRecursionDepthAndPushLimit(length);
-  if (p.second < 0 || !value->MergePartialFromCodedStream(input)) return false;
-  // Make sure that parsing stopped when the limit was hit, not at an endgroup
-  // tag.
-  return input->DecrementRecursionDepthAndPopLimit(p.first);
-}
-
-// ===================================================================
-
-inline void WireFormatLite::WriteTag(int field_number, WireType type,
-                                     io::CodedOutputStream* output) {
-  output->WriteTag(MakeTag(field_number, type));
-}
-
-inline void WireFormatLite::WriteInt32NoTag(int32 value,
-                                            io::CodedOutputStream* output) {
-  output->WriteVarint32SignExtended(value);
-}
-inline void WireFormatLite::WriteInt64NoTag(int64 value,
-                                            io::CodedOutputStream* output) {
-  output->WriteVarint64(static_cast<uint64>(value));
-}
-inline void WireFormatLite::WriteUInt32NoTag(uint32 value,
-                                             io::CodedOutputStream* output) {
-  output->WriteVarint32(value);
-}
-inline void WireFormatLite::WriteUInt64NoTag(uint64 value,
-                                             io::CodedOutputStream* output) {
-  output->WriteVarint64(value);
-}
-inline void WireFormatLite::WriteSInt32NoTag(int32 value,
-                                             io::CodedOutputStream* output) {
-  output->WriteVarint32(ZigZagEncode32(value));
-}
-inline void WireFormatLite::WriteSInt64NoTag(int64 value,
-                                             io::CodedOutputStream* output) {
-  output->WriteVarint64(ZigZagEncode64(value));
-}
-inline void WireFormatLite::WriteFixed32NoTag(uint32 value,
-                                              io::CodedOutputStream* output) {
-  output->WriteLittleEndian32(value);
-}
-inline void WireFormatLite::WriteFixed64NoTag(uint64 value,
-                                              io::CodedOutputStream* output) {
-  output->WriteLittleEndian64(value);
-}
-inline void WireFormatLite::WriteSFixed32NoTag(int32 value,
-                                               io::CodedOutputStream* output) {
-  output->WriteLittleEndian32(static_cast<uint32>(value));
-}
-inline void WireFormatLite::WriteSFixed64NoTag(int64 value,
-                                               io::CodedOutputStream* output) {
-  output->WriteLittleEndian64(static_cast<uint64>(value));
-}
-inline void WireFormatLite::WriteFloatNoTag(float value,
-                                            io::CodedOutputStream* output) {
-  output->WriteLittleEndian32(EncodeFloat(value));
-}
-inline void WireFormatLite::WriteDoubleNoTag(double value,
-                                             io::CodedOutputStream* output) {
-  output->WriteLittleEndian64(EncodeDouble(value));
-}
-inline void WireFormatLite::WriteBoolNoTag(bool value,
-                                           io::CodedOutputStream* output) {
-  output->WriteVarint32(value ? 1 : 0);
-}
-inline void WireFormatLite::WriteEnumNoTag(int value,
-                                           io::CodedOutputStream* output) {
-  output->WriteVarint32SignExtended(value);
-}
-
-// See comment on ReadGroupNoVirtual to understand the need for this template
-// parameter name.
-template <typename MessageType_WorkAroundCppLookupDefect>
-inline void WireFormatLite::WriteGroupNoVirtual(
-    int field_number, const MessageType_WorkAroundCppLookupDefect& value,
-    io::CodedOutputStream* output) {
-  WriteTag(field_number, WIRETYPE_START_GROUP, output);
-  value.MessageType_WorkAroundCppLookupDefect::SerializeWithCachedSizes(output);
-  WriteTag(field_number, WIRETYPE_END_GROUP, output);
-}
-template <typename MessageType_WorkAroundCppLookupDefect>
-inline void WireFormatLite::WriteMessageNoVirtual(
-    int field_number, const MessageType_WorkAroundCppLookupDefect& value,
-    io::CodedOutputStream* output) {
-  WriteTag(field_number, WIRETYPE_LENGTH_DELIMITED, output);
-  output->WriteVarint32(
-      value.MessageType_WorkAroundCppLookupDefect::GetCachedSize());
-  value.MessageType_WorkAroundCppLookupDefect::SerializeWithCachedSizes(output);
-}
-
-// ===================================================================
-
-inline uint8* WireFormatLite::WriteTagToArray(int field_number, WireType type,
-                                              uint8* target) {
-  return io::CodedOutputStream::WriteTagToArray(MakeTag(field_number, type),
-                                                target);
-}
-
-inline uint8* WireFormatLite::WriteInt32NoTagToArray(int32 value,
-                                                     uint8* target) {
-  return io::CodedOutputStream::WriteVarint32SignExtendedToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteInt64NoTagToArray(int64 value,
-                                                     uint8* target) {
-  return io::CodedOutputStream::WriteVarint64ToArray(static_cast<uint64>(value),
-                                                     target);
-}
-inline uint8* WireFormatLite::WriteUInt32NoTagToArray(uint32 value,
-                                                      uint8* target) {
-  return io::CodedOutputStream::WriteVarint32ToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteUInt64NoTagToArray(uint64 value,
-                                                      uint8* target) {
-  return io::CodedOutputStream::WriteVarint64ToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteSInt32NoTagToArray(int32 value,
-                                                      uint8* target) {
-  return io::CodedOutputStream::WriteVarint32ToArray(ZigZagEncode32(value),
-                                                     target);
-}
-inline uint8* WireFormatLite::WriteSInt64NoTagToArray(int64 value,
-                                                      uint8* target) {
-  return io::CodedOutputStream::WriteVarint64ToArray(ZigZagEncode64(value),
-                                                     target);
-}
-inline uint8* WireFormatLite::WriteFixed32NoTagToArray(uint32 value,
-                                                       uint8* target) {
-  return io::CodedOutputStream::WriteLittleEndian32ToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteFixed64NoTagToArray(uint64 value,
-                                                       uint8* target) {
-  return io::CodedOutputStream::WriteLittleEndian64ToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteSFixed32NoTagToArray(int32 value,
-                                                        uint8* target) {
-  return io::CodedOutputStream::WriteLittleEndian32ToArray(
-      static_cast<uint32>(value), target);
-}
-inline uint8* WireFormatLite::WriteSFixed64NoTagToArray(int64 value,
-                                                        uint8* target) {
-  return io::CodedOutputStream::WriteLittleEndian64ToArray(
-      static_cast<uint64>(value), target);
-}
-inline uint8* WireFormatLite::WriteFloatNoTagToArray(float value,
-                                                     uint8* target) {
-  return io::CodedOutputStream::WriteLittleEndian32ToArray(EncodeFloat(value),
-                                                           target);
-}
-inline uint8* WireFormatLite::WriteDoubleNoTagToArray(double value,
-                                                      uint8* target) {
-  return io::CodedOutputStream::WriteLittleEndian64ToArray(EncodeDouble(value),
-                                                           target);
-}
-inline uint8* WireFormatLite::WriteBoolNoTagToArray(bool value, uint8* target) {
-  return io::CodedOutputStream::WriteVarint32ToArray(value ? 1 : 0, target);
-}
-inline uint8* WireFormatLite::WriteEnumNoTagToArray(int value, uint8* target) {
-  return io::CodedOutputStream::WriteVarint32SignExtendedToArray(value, target);
-}
-
-template <typename T>
-inline uint8* WireFormatLite::WritePrimitiveNoTagToArray(
-    const RepeatedField<T>& value, uint8* (*Writer)(T, uint8*), uint8* target) {
-  const int n = value.size();
-  GOOGLE_DCHECK_GT(n, 0);
-
-  const T* ii = value.data();
-  int i = 0;
-  do {
-    target = Writer(ii[i], target);
-  } while (++i < n);
-
-  return target;
-}
-
-template <typename T>
-inline uint8* WireFormatLite::WriteFixedNoTagToArray(
-    const RepeatedField<T>& value, uint8* (*Writer)(T, uint8*), uint8* target) {
-#if defined(PROTOBUF_LITTLE_ENDIAN)
-  (void)Writer;
-
-  const int n = value.size();
-  GOOGLE_DCHECK_GT(n, 0);
-
-  const T* ii = value.data();
-  const int bytes = n * static_cast<int>(sizeof(ii[0]));
-  memcpy(target, ii, static_cast<size_t>(bytes));
-  return target + bytes;
-#else
-  return WritePrimitiveNoTagToArray(value, Writer, target);
-#endif
-}
-
-inline uint8* WireFormatLite::WriteInt32NoTagToArray(
-    const RepeatedField<int32>& value, uint8* target) {
-  return WritePrimitiveNoTagToArray(value, WriteInt32NoTagToArray, target);
-}
-inline uint8* WireFormatLite::WriteInt64NoTagToArray(
-    const RepeatedField<int64>& value, uint8* target) {
-  return WritePrimitiveNoTagToArray(value, WriteInt64NoTagToArray, target);
-}
-inline uint8* WireFormatLite::WriteUInt32NoTagToArray(
-    const RepeatedField<uint32>& value, uint8* target) {
-  return WritePrimitiveNoTagToArray(value, WriteUInt32NoTagToArray, target);
-}
-inline uint8* WireFormatLite::WriteUInt64NoTagToArray(
-    const RepeatedField<uint64>& value, uint8* target) {
-  return WritePrimitiveNoTagToArray(value, WriteUInt64NoTagToArray, target);
-}
-inline uint8* WireFormatLite::WriteSInt32NoTagToArray(
-    const RepeatedField<int32>& value, uint8* target) {
-  return WritePrimitiveNoTagToArray(value, WriteSInt32NoTagToArray, target);
-}
-inline uint8* WireFormatLite::WriteSInt64NoTagToArray(
-    const RepeatedField<int64>& value, uint8* target) {
-  return WritePrimitiveNoTagToArray(value, WriteSInt64NoTagToArray, target);
-}
-inline uint8* WireFormatLite::WriteFixed32NoTagToArray(
-    const RepeatedField<uint32>& value, uint8* target) {
-  return WriteFixedNoTagToArray(value, WriteFixed32NoTagToArray, target);
-}
-inline uint8* WireFormatLite::WriteFixed64NoTagToArray(
-    const RepeatedField<uint64>& value, uint8* target) {
-  return WriteFixedNoTagToArray(value, WriteFixed64NoTagToArray, target);
-}
-inline uint8* WireFormatLite::WriteSFixed32NoTagToArray(
-    const RepeatedField<int32>& value, uint8* target) {
-  return WriteFixedNoTagToArray(value, WriteSFixed32NoTagToArray, target);
-}
-inline uint8* WireFormatLite::WriteSFixed64NoTagToArray(
-    const RepeatedField<int64>& value, uint8* target) {
-  return WriteFixedNoTagToArray(value, WriteSFixed64NoTagToArray, target);
-}
-inline uint8* WireFormatLite::WriteFloatNoTagToArray(
-    const RepeatedField<float>& value, uint8* target) {
-  return WriteFixedNoTagToArray(value, WriteFloatNoTagToArray, target);
-}
-inline uint8* WireFormatLite::WriteDoubleNoTagToArray(
-    const RepeatedField<double>& value, uint8* target) {
-  return WriteFixedNoTagToArray(value, WriteDoubleNoTagToArray, target);
-}
-inline uint8* WireFormatLite::WriteBoolNoTagToArray(
-    const RepeatedField<bool>& value, uint8* target) {
-  return WritePrimitiveNoTagToArray(value, WriteBoolNoTagToArray, target);
-}
-inline uint8* WireFormatLite::WriteEnumNoTagToArray(
-    const RepeatedField<int>& value, uint8* target) {
-  return WritePrimitiveNoTagToArray(value, WriteEnumNoTagToArray, target);
-}
-
-inline uint8* WireFormatLite::WriteInt32ToArray(int field_number, int32 value,
-                                                uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_VARINT, target);
-  return WriteInt32NoTagToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteInt64ToArray(int field_number, int64 value,
-                                                uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_VARINT, target);
-  return WriteInt64NoTagToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteUInt32ToArray(int field_number, uint32 value,
-                                                 uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_VARINT, target);
-  return WriteUInt32NoTagToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteUInt64ToArray(int field_number, uint64 value,
-                                                 uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_VARINT, target);
-  return WriteUInt64NoTagToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteSInt32ToArray(int field_number, int32 value,
-                                                 uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_VARINT, target);
-  return WriteSInt32NoTagToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteSInt64ToArray(int field_number, int64 value,
-                                                 uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_VARINT, target);
-  return WriteSInt64NoTagToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteFixed32ToArray(int field_number,
-                                                  uint32 value, uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_FIXED32, target);
-  return WriteFixed32NoTagToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteFixed64ToArray(int field_number,
-                                                  uint64 value, uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_FIXED64, target);
-  return WriteFixed64NoTagToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteSFixed32ToArray(int field_number,
-                                                   int32 value, uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_FIXED32, target);
-  return WriteSFixed32NoTagToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteSFixed64ToArray(int field_number,
-                                                   int64 value, uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_FIXED64, target);
-  return WriteSFixed64NoTagToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteFloatToArray(int field_number, float value,
-                                                uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_FIXED32, target);
-  return WriteFloatNoTagToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteDoubleToArray(int field_number, double value,
-                                                 uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_FIXED64, target);
-  return WriteDoubleNoTagToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteBoolToArray(int field_number, bool value,
-                                               uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_VARINT, target);
-  return WriteBoolNoTagToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteEnumToArray(int field_number, int value,
-                                               uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_VARINT, target);
-  return WriteEnumNoTagToArray(value, target);
-}
-
-template <typename T>
-inline uint8* WireFormatLite::WritePrimitiveToArray(
-    int field_number, const RepeatedField<T>& value,
-    uint8* (*Writer)(int, T, uint8*), uint8* target) {
-  const int n = value.size();
-  if (n == 0) {
-    return target;
-  }
-
-  const T* ii = value.data();
-  int i = 0;
-  do {
-    target = Writer(field_number, ii[i], target);
-  } while (++i < n);
-
-  return target;
-}
-
-inline uint8* WireFormatLite::WriteInt32ToArray(
-    int field_number, const RepeatedField<int32>& value, uint8* target) {
-  return WritePrimitiveToArray(field_number, value, WriteInt32ToArray, target);
-}
-inline uint8* WireFormatLite::WriteInt64ToArray(
-    int field_number, const RepeatedField<int64>& value, uint8* target) {
-  return WritePrimitiveToArray(field_number, value, WriteInt64ToArray, target);
-}
-inline uint8* WireFormatLite::WriteUInt32ToArray(
-    int field_number, const RepeatedField<uint32>& value, uint8* target) {
-  return WritePrimitiveToArray(field_number, value, WriteUInt32ToArray, target);
-}
-inline uint8* WireFormatLite::WriteUInt64ToArray(
-    int field_number, const RepeatedField<uint64>& value, uint8* target) {
-  return WritePrimitiveToArray(field_number, value, WriteUInt64ToArray, target);
-}
-inline uint8* WireFormatLite::WriteSInt32ToArray(
-    int field_number, const RepeatedField<int32>& value, uint8* target) {
-  return WritePrimitiveToArray(field_number, value, WriteSInt32ToArray, target);
-}
-inline uint8* WireFormatLite::WriteSInt64ToArray(
-    int field_number, const RepeatedField<int64>& value, uint8* target) {
-  return WritePrimitiveToArray(field_number, value, WriteSInt64ToArray, target);
-}
-inline uint8* WireFormatLite::WriteFixed32ToArray(
-    int field_number, const RepeatedField<uint32>& value, uint8* target) {
-  return WritePrimitiveToArray(field_number, value, WriteFixed32ToArray,
-                               target);
-}
-inline uint8* WireFormatLite::WriteFixed64ToArray(
-    int field_number, const RepeatedField<uint64>& value, uint8* target) {
-  return WritePrimitiveToArray(field_number, value, WriteFixed64ToArray,
-                               target);
-}
-inline uint8* WireFormatLite::WriteSFixed32ToArray(
-    int field_number, const RepeatedField<int32>& value, uint8* target) {
-  return WritePrimitiveToArray(field_number, value, WriteSFixed32ToArray,
-                               target);
-}
-inline uint8* WireFormatLite::WriteSFixed64ToArray(
-    int field_number, const RepeatedField<int64>& value, uint8* target) {
-  return WritePrimitiveToArray(field_number, value, WriteSFixed64ToArray,
-                               target);
-}
-inline uint8* WireFormatLite::WriteFloatToArray(
-    int field_number, const RepeatedField<float>& value, uint8* target) {
-  return WritePrimitiveToArray(field_number, value, WriteFloatToArray, target);
-}
-inline uint8* WireFormatLite::WriteDoubleToArray(
-    int field_number, const RepeatedField<double>& value, uint8* target) {
-  return WritePrimitiveToArray(field_number, value, WriteDoubleToArray, target);
-}
-inline uint8* WireFormatLite::WriteBoolToArray(int field_number,
-                                               const RepeatedField<bool>& value,
-                                               uint8* target) {
-  return WritePrimitiveToArray(field_number, value, WriteBoolToArray, target);
-}
-inline uint8* WireFormatLite::WriteEnumToArray(int field_number,
-                                               const RepeatedField<int>& value,
-                                               uint8* target) {
-  return WritePrimitiveToArray(field_number, value, WriteEnumToArray, target);
-}
-inline uint8* WireFormatLite::WriteStringToArray(int field_number,
-                                                 const std::string& value,
-                                                 uint8* target) {
-  // String is for UTF-8 text only
-  // WARNING:  In wire_format.cc, both strings and bytes are handled by
-  //   WriteString() to avoid code duplication.  If the implementations become
-  //   different, you will need to update that usage.
-  target = WriteTagToArray(field_number, WIRETYPE_LENGTH_DELIMITED, target);
-  return io::CodedOutputStream::WriteStringWithSizeToArray(value, target);
-}
-inline uint8* WireFormatLite::WriteBytesToArray(int field_number,
-                                                const std::string& value,
-                                                uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_LENGTH_DELIMITED, target);
-  return io::CodedOutputStream::WriteStringWithSizeToArray(value, target);
-}
-
-
-template <typename MessageType>
-inline uint8* WireFormatLite::InternalWriteGroup(
-    int field_number, const MessageType& value, uint8* target,
-    io::EpsCopyOutputStream* stream) {
-  target = WriteTagToArray(field_number, WIRETYPE_START_GROUP, target);
-  target = value._InternalSerialize(target, stream);
-  target = stream->EnsureSpace(target);
-  return WriteTagToArray(field_number, WIRETYPE_END_GROUP, target);
-}
-template <typename MessageType>
-inline uint8* WireFormatLite::InternalWriteMessage(
-    int field_number, const MessageType& value, uint8* target,
-    io::EpsCopyOutputStream* stream) {
-  target = WriteTagToArray(field_number, WIRETYPE_LENGTH_DELIMITED, target);
-  target = io::CodedOutputStream::WriteVarint32ToArray(
-      static_cast<uint32>(value.GetCachedSize()), target);
-  return value._InternalSerialize(target, stream);
-}
-
-// See comment on ReadGroupNoVirtual to understand the need for this template
-// parameter name.
-template <typename MessageType_WorkAroundCppLookupDefect>
-inline uint8* WireFormatLite::InternalWriteGroupNoVirtualToArray(
-    int field_number, const MessageType_WorkAroundCppLookupDefect& value,
-    uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_START_GROUP, target);
-  target = value.MessageType_WorkAroundCppLookupDefect::
-               SerializeWithCachedSizesToArray(target);
-  return WriteTagToArray(field_number, WIRETYPE_END_GROUP, target);
-}
-template <typename MessageType_WorkAroundCppLookupDefect>
-inline uint8* WireFormatLite::InternalWriteMessageNoVirtualToArray(
-    int field_number, const MessageType_WorkAroundCppLookupDefect& value,
-    uint8* target) {
-  target = WriteTagToArray(field_number, WIRETYPE_LENGTH_DELIMITED, target);
-  target = io::CodedOutputStream::WriteVarint32ToArray(
-      static_cast<uint32>(
-          value.MessageType_WorkAroundCppLookupDefect::GetCachedSize()),
-      target);
-  return value
-      .MessageType_WorkAroundCppLookupDefect::SerializeWithCachedSizesToArray(
-          target);
-}
-
-// ===================================================================
-
-inline size_t WireFormatLite::Int32Size(int32 value) {
-  return io::CodedOutputStream::VarintSize32SignExtended(value);
-}
-inline size_t WireFormatLite::Int64Size(int64 value) {
-  return io::CodedOutputStream::VarintSize64(static_cast<uint64>(value));
-}
-inline size_t WireFormatLite::UInt32Size(uint32 value) {
-  return io::CodedOutputStream::VarintSize32(value);
-}
-inline size_t WireFormatLite::UInt64Size(uint64 value) {
-  return io::CodedOutputStream::VarintSize64(value);
-}
-inline size_t WireFormatLite::SInt32Size(int32 value) {
-  return io::CodedOutputStream::VarintSize32(ZigZagEncode32(value));
-}
-inline size_t WireFormatLite::SInt64Size(int64 value) {
-  return io::CodedOutputStream::VarintSize64(ZigZagEncode64(value));
-}
-inline size_t WireFormatLite::EnumSize(int value) {
-  return io::CodedOutputStream::VarintSize32SignExtended(value);
-}
-
-inline size_t WireFormatLite::StringSize(const std::string& value) {
-  return LengthDelimitedSize(value.size());
-}
-inline size_t WireFormatLite::BytesSize(const std::string& value) {
-  return LengthDelimitedSize(value.size());
-}
-
-
-template <typename MessageType>
-inline size_t WireFormatLite::GroupSize(const MessageType& value) {
-  return value.ByteSizeLong();
-}
-template <typename MessageType>
-inline size_t WireFormatLite::MessageSize(const MessageType& value) {
-  return LengthDelimitedSize(value.ByteSizeLong());
-}
-
-// See comment on ReadGroupNoVirtual to understand the need for this template
-// parameter name.
-template <typename MessageType_WorkAroundCppLookupDefect>
-inline size_t WireFormatLite::GroupSizeNoVirtual(
-    const MessageType_WorkAroundCppLookupDefect& value) {
-  return value.MessageType_WorkAroundCppLookupDefect::ByteSizeLong();
-}
-template <typename MessageType_WorkAroundCppLookupDefect>
-inline size_t WireFormatLite::MessageSizeNoVirtual(
-    const MessageType_WorkAroundCppLookupDefect& value) {
-  return LengthDelimitedSize(
-      value.MessageType_WorkAroundCppLookupDefect::ByteSizeLong());
-}
-
-inline size_t WireFormatLite::LengthDelimitedSize(size_t length) {
-  // The static_cast here prevents an error in certain compiler configurations
-  // but is not technically correct--if length is too large to fit in a uint32
-  // then it will be silently truncated. We will need to fix this if we ever
-  // decide to start supporting serialized messages greater than 2 GiB in size.
-  return length +
-         io::CodedOutputStream::VarintSize32(static_cast<uint32>(length));
-}
-
-template <typename MS>
-bool ParseMessageSetItemImpl(io::CodedInputStream* input, MS ms) {
-  // This method parses a group which should contain two fields:
-  //   required int32 type_id = 2;
-  //   required data message = 3;
-
-  uint32 last_type_id = 0;
-
-  // If we see message data before the type_id, we'll append it to this so
-  // we can parse it later.
-  std::string message_data;
-
-  while (true) {
-    const uint32 tag = input->ReadTagNoLastTag();
-    if (tag == 0) return false;
-
-    switch (tag) {
-      case WireFormatLite::kMessageSetTypeIdTag: {
-        uint32 type_id;
-        if (!input->ReadVarint32(&type_id)) return false;
-        last_type_id = type_id;
-
-        if (!message_data.empty()) {
-          // We saw some message data before the type_id.  Have to parse it
-          // now.
-          io::CodedInputStream sub_input(
-              reinterpret_cast<const uint8*>(message_data.data()),
-              static_cast<int>(message_data.size()));
-          sub_input.SetRecursionLimit(input->RecursionBudget());
-          if (!ms.ParseField(last_type_id, &sub_input)) {
-            return false;
-          }
-          message_data.clear();
-        }
-
-        break;
-      }
-
-      case WireFormatLite::kMessageSetMessageTag: {
-        if (last_type_id == 0) {
-          // We haven't seen a type_id yet.  Append this data to message_data.
-          uint32 length;
-          if (!input->ReadVarint32(&length)) return false;
-          if (static_cast<int32>(length) < 0) return false;
-          uint32 size = static_cast<uint32>(
-              length + io::CodedOutputStream::VarintSize32(length));
-          message_data.resize(size);
-          auto ptr = reinterpret_cast<uint8*>(&message_data[0]);
-          ptr = io::CodedOutputStream::WriteVarint32ToArray(length, ptr);
-          if (!input->ReadRaw(ptr, length)) return false;
-        } else {
-          // Already saw type_id, so we can parse this directly.
-          if (!ms.ParseField(last_type_id, input)) {
-            return false;
-          }
-        }
-
-        break;
-      }
-
-      case WireFormatLite::kMessageSetItemEndTag: {
-        return true;
-      }
-
-      default: {
-        if (!ms.SkipField(tag, input)) return false;
-      }
-    }
-  }
-}
-
-}  // namespace internal
-}  // namespace protobuf
-}  // namespace google
-
-#include <google/protobuf/port_undef.inc>
-
-#endif  // GOOGLE_PROTOBUF_WIRE_FORMAT_LITE_H__
diff --git a/projects/llm_framework/include/re2/re2/bitmap256.h b/projects/llm_framework/include/re2/re2/bitmap256.h
deleted file mode 100644
index 293b31d8..00000000
--- a/projects/llm_framework/include/re2/re2/bitmap256.h
+++ /dev/null
@@ -1,86 +0,0 @@
-// Copyright 2016 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_BITMAP256_H_
-#define RE2_BITMAP256_H_
-
-#ifdef _MSC_VER
-#include <intrin.h>
-#endif
-#include <stdint.h>
-#include <string.h>
-
-#include "util/logging.h"
-
-namespace re2 {
-
-class Bitmap256 {
- public:
-  Bitmap256() {
-    Clear();
-  }
-
-  // Clears all of the bits.
-  void Clear() {
-    memset(words_, 0, sizeof words_);
-  }
-
-  // Tests the bit with index c.
-  bool Test(int c) const {
-    DCHECK_GE(c, 0);
-    DCHECK_LE(c, 255);
-
-    return (words_[c / 64] & (uint64_t{1} << (c % 64))) != 0;
-  }
-
-  // Sets the bit with index c.
-  void Set(int c) {
-    DCHECK_GE(c, 0);
-    DCHECK_LE(c, 255);
-
-    words_[c / 64] |= (uint64_t{1} << (c % 64));
-  }
-
-  // Finds the next non-zero bit with index >= c.
-  // Returns -1 if no such bit exists.
-  int FindNextSetBit(int c) const;
-
- private:
-  // Finds the least significant non-zero bit in n.
-  static int FindLSBSet(uint64_t n) {
-    DCHECK_NE(n, 0);
-#if defined(__GNUC__)
-    return __builtin_ctzll(n);
-#elif defined(_MSC_VER) && defined(_M_X64)
-    unsigned long c;
-    _BitScanForward64(&c, n);
-    return static_cast<int>(c);
-#elif defined(_MSC_VER) && defined(_M_IX86)
-    unsigned long c;
-    if (static_cast<uint32_t>(n) != 0) {
-      _BitScanForward(&c, static_cast<uint32_t>(n));
-      return static_cast<int>(c);
-    } else {
-      _BitScanForward(&c, static_cast<uint32_t>(n >> 32));
-      return static_cast<int>(c) + 32;
-    }
-#else
-    int c = 63;
-    for (int shift = 1 << 5; shift != 0; shift >>= 1) {
-      uint64_t word = n << shift;
-      if (word != 0) {
-        n = word;
-        c -= shift;
-      }
-    }
-    return c;
-#endif
-  }
-
-  uint64_t words_[4];
-};
-
-}  // namespace re2
-
-#endif  // RE2_BITMAP256_H_
diff --git a/projects/llm_framework/include/re2/re2/filtered_re2.h b/projects/llm_framework/include/re2/re2/filtered_re2.h
deleted file mode 100644
index a9abd691..00000000
--- a/projects/llm_framework/include/re2/re2/filtered_re2.h
+++ /dev/null
@@ -1,115 +0,0 @@
-// Copyright 2009 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_FILTERED_RE2_H_
-#define RE2_FILTERED_RE2_H_
-
-// The class FilteredRE2 is used as a wrapper to multiple RE2 regexps.
-// It provides a prefilter mechanism that helps in cutting down the
-// number of regexps that need to be actually searched.
-//
-// By design, it does not include a string matching engine. This is to
-// allow the user of the class to use their favorite string matching
-// engine. The overall flow is: Add all the regexps using Add, then
-// Compile the FilteredRE2. Compile returns strings that need to be
-// matched. Note that the returned strings are lowercased and distinct.
-// For applying regexps to a search text, the caller does the string
-// matching using the returned strings. When doing the string match,
-// note that the caller has to do that in a case-insensitive way or
-// on a lowercased version of the search text. Then call FirstMatch
-// or AllMatches with a vector of indices of strings that were found
-// in the text to get the actual regexp matches.
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "absl/strings/string_view.h"
-#include "re2/re2.h"
-
-namespace re2 {
-
-class PrefilterTree;
-
-class FilteredRE2 {
- public:
-  FilteredRE2();
-  explicit FilteredRE2(int min_atom_len);
-  ~FilteredRE2();
-
-  // Not copyable.
-  FilteredRE2(const FilteredRE2&) = delete;
-  FilteredRE2& operator=(const FilteredRE2&) = delete;
-  // Movable.
-  FilteredRE2(FilteredRE2&& other);
-  FilteredRE2& operator=(FilteredRE2&& other);
-
-  // Uses RE2 constructor to create a RE2 object (re). Returns
-  // re->error_code(). If error_code is other than NoError, then re is
-  // deleted and not added to re2_vec_.
-  RE2::ErrorCode Add(absl::string_view pattern,
-                     const RE2::Options& options,
-                     int* id);
-
-  // Prepares the regexps added by Add for filtering.  Returns a set
-  // of strings that the caller should check for in candidate texts.
-  // The returned strings are lowercased and distinct. When doing
-  // string matching, it should be performed in a case-insensitive
-  // way or the search text should be lowercased first.  Call after
-  // all Add calls are done.
-  void Compile(std::vector<std::string>* strings_to_match);
-
-  // Returns the index of the first matching regexp.
-  // Returns -1 on no match. Can be called prior to Compile.
-  // Does not do any filtering: simply tries to Match the
-  // regexps in a loop.
-  int SlowFirstMatch(absl::string_view text) const;
-
-  // Returns the index of the first matching regexp.
-  // Returns -1 on no match. Compile has to be called before
-  // calling this.
-  int FirstMatch(absl::string_view text,
-                 const std::vector<int>& atoms) const;
-
-  // Returns the indices of all matching regexps, after first clearing
-  // matched_regexps.
-  bool AllMatches(absl::string_view text,
-                  const std::vector<int>& atoms,
-                  std::vector<int>* matching_regexps) const;
-
-  // Returns the indices of all potentially matching regexps after first
-  // clearing potential_regexps.
-  // A regexp is potentially matching if it passes the filter.
-  // If a regexp passes the filter it may still not match.
-  // A regexp that does not pass the filter is guaranteed to not match.
-  void AllPotentials(const std::vector<int>& atoms,
-                     std::vector<int>* potential_regexps) const;
-
-  // The number of regexps added.
-  int NumRegexps() const { return static_cast<int>(re2_vec_.size()); }
-
-  // Get the individual RE2 objects.
-  const RE2& GetRE2(int regexpid) const { return *re2_vec_[regexpid]; }
-
- private:
-  // Print prefilter.
-  void PrintPrefilter(int regexpid);
-
-  // Useful for testing and debugging.
-  void RegexpsGivenStrings(const std::vector<int>& matched_atoms,
-                           std::vector<int>* passed_regexps);
-
-  // All the regexps in the FilteredRE2.
-  std::vector<RE2*> re2_vec_;
-
-  // Has the FilteredRE2 been compiled using Compile()
-  bool compiled_;
-
-  // An AND-OR tree of string atoms used for filtering regexps.
-  std::unique_ptr<PrefilterTree> prefilter_tree_;
-};
-
-}  // namespace re2
-
-#endif  // RE2_FILTERED_RE2_H_
diff --git a/projects/llm_framework/include/re2/re2/pod_array.h b/projects/llm_framework/include/re2/re2/pod_array.h
deleted file mode 100644
index f234e976..00000000
--- a/projects/llm_framework/include/re2/re2/pod_array.h
+++ /dev/null
@@ -1,55 +0,0 @@
-// Copyright 2018 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_POD_ARRAY_H_
-#define RE2_POD_ARRAY_H_
-
-#include <memory>
-#include <type_traits>
-
-namespace re2 {
-
-template <typename T>
-class PODArray {
- public:
-  static_assert(std::is_trivial<T>::value && std::is_standard_layout<T>::value,
-                "T must be POD");
-
-  PODArray()
-      : ptr_() {}
-  explicit PODArray(int len)
-      : ptr_(std::allocator<T>().allocate(len), Deleter(len)) {}
-
-  T* data() const {
-    return ptr_.get();
-  }
-
-  int size() const {
-    return ptr_.get_deleter().len_;
-  }
-
-  T& operator[](int pos) const {
-    return ptr_[pos];
-  }
-
- private:
-  struct Deleter {
-    Deleter()
-        : len_(0) {}
-    explicit Deleter(int len)
-        : len_(len) {}
-
-    void operator()(T* ptr) const {
-      std::allocator<T>().deallocate(ptr, len_);
-    }
-
-    int len_;
-  };
-
-  std::unique_ptr<T[], Deleter> ptr_;
-};
-
-}  // namespace re2
-
-#endif  // RE2_POD_ARRAY_H_
diff --git a/projects/llm_framework/include/re2/re2/prefilter.h b/projects/llm_framework/include/re2/re2/prefilter.h
deleted file mode 100644
index 018691dc..00000000
--- a/projects/llm_framework/include/re2/re2/prefilter.h
+++ /dev/null
@@ -1,167 +0,0 @@
-// Copyright 2009 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_PREFILTER_H_
-#define RE2_PREFILTER_H_
-
-// Prefilter is the class used to extract string guards from regexps.
-// Rather than using Prefilter class directly, use FilteredRE2.
-// See filtered_re2.h
-
-#include <set>
-#include <string>
-#include <vector>
-
-#include "util/logging.h"
-
-namespace re2 {
-
-class RE2;
-
-class Regexp;
-
-class Prefilter {
-  // Instead of using Prefilter directly, use FilteredRE2; see filtered_re2.h
- public:
-  enum Op {
-    ALL = 0,  // Everything matches
-    NONE,  // Nothing matches
-    ATOM,  // The string atom() must match
-    AND,   // All in subs() must match
-    OR,   // One of subs() must match
-  };
-
-  explicit Prefilter(Op op);
-  ~Prefilter();
-
-  Op op() { return op_; }
-  const std::string& atom() const { return atom_; }
-  void set_unique_id(int id) { unique_id_ = id; }
-  int unique_id() const { return unique_id_; }
-
-  // The children of the Prefilter node.
-  std::vector<Prefilter*>* subs() {
-    DCHECK(op_ == AND || op_ == OR);
-    return subs_;
-  }
-
-  // Set the children vector. Prefilter takes ownership of subs and
-  // subs_ will be deleted when Prefilter is deleted.
-  void set_subs(std::vector<Prefilter*>* subs) { subs_ = subs; }
-
-  // Given a RE2, return a Prefilter. The caller takes ownership of
-  // the Prefilter and should deallocate it. Returns NULL if Prefilter
-  // cannot be formed.
-  static Prefilter* FromRE2(const RE2* re2);
-
-  // Returns a readable debug string of the prefilter.
-  std::string DebugString() const;
-
- private:
-  template <typename H>
-  friend H AbslHashValue(H h, const Prefilter& a) {
-    h = H::combine(std::move(h), a.op_);
-    if (a.op_ == ATOM) {
-      h = H::combine(std::move(h), a.atom_);
-    } else if (a.op_ == AND || a.op_ == OR) {
-      h = H::combine(std::move(h), a.subs_->size());
-      for (size_t i = 0; i < a.subs_->size(); ++i) {
-        h = H::combine(std::move(h), (*a.subs_)[i]->unique_id_);
-      }
-    }
-    return h;
-  }
-
-  friend bool operator==(const Prefilter& a, const Prefilter& b) {
-    if (&a == &b) {
-      return true;
-    }
-    if (a.op_ != b.op_) {
-      return false;
-    }
-    if (a.op_ == ATOM) {
-      if (a.atom_ != b.atom_) {
-        return false;
-      }
-    } else if (a.op_ == AND || a.op_ == OR) {
-      if (a.subs_->size() != b.subs_->size()) {
-        return false;
-      }
-      for (size_t i = 0; i < a.subs_->size(); ++i) {
-        if ((*a.subs_)[i]->unique_id_ != (*b.subs_)[i]->unique_id_) {
-          return false;
-        }
-      }
-    }
-    return true;
-  }
-
-  // A comparator used to store exact strings. We compare by length,
-  // then lexicographically. This ordering makes it easier to reduce the
-  // set of strings in SimplifyStringSet.
-  struct LengthThenLex {
-    bool operator()(const std::string& a, const std::string& b) const {
-       return (a.size() < b.size()) || (a.size() == b.size() && a < b);
-    }
-  };
-
-  class Info;
-
-  using SSet = std::set<std::string, LengthThenLex>;
-  using SSIter = SSet::iterator;
-  using ConstSSIter = SSet::const_iterator;
-
-  // Combines two prefilters together to create an AND. The passed
-  // Prefilters will be part of the returned Prefilter or deleted.
-  static Prefilter* And(Prefilter* a, Prefilter* b);
-
-  // Combines two prefilters together to create an OR. The passed
-  // Prefilters will be part of the returned Prefilter or deleted.
-  static Prefilter* Or(Prefilter* a, Prefilter* b);
-
-  // Generalized And/Or
-  static Prefilter* AndOr(Op op, Prefilter* a, Prefilter* b);
-
-  static Prefilter* FromRegexp(Regexp* a);
-
-  static Prefilter* FromString(const std::string& str);
-
-  static Prefilter* OrStrings(SSet* ss);
-
-  static Info* BuildInfo(Regexp* re);
-
-  Prefilter* Simplify();
-
-  // Removes redundant strings from the set. A string is redundant if
-  // any of the other strings appear as a substring. The empty string
-  // is a special case, which is ignored.
-  static void SimplifyStringSet(SSet* ss);
-
-  // Adds the cross-product of a and b to dst.
-  // (For each string i in a and j in b, add i+j.)
-  static void CrossProduct(const SSet& a, const SSet& b, SSet* dst);
-
-  // Kind of Prefilter.
-  Op op_;
-
-  // Sub-matches for AND or OR Prefilter.
-  std::vector<Prefilter*>* subs_;
-
-  // Actual string to match in leaf node.
-  std::string atom_;
-
-  // If different prefilters have the same string atom, or if they are
-  // structurally the same (e.g., OR of same atom strings) they are
-  // considered the same unique nodes. This is the id for each unique
-  // node. This field is populated with a unique id for every node,
-  // and -1 for duplicate nodes.
-  int unique_id_;
-
-  Prefilter(const Prefilter&) = delete;
-  Prefilter& operator=(const Prefilter&) = delete;
-};
-
-}  // namespace re2
-
-#endif  // RE2_PREFILTER_H_
diff --git a/projects/llm_framework/include/re2/re2/prefilter_tree.h b/projects/llm_framework/include/re2/re2/prefilter_tree.h
deleted file mode 100644
index 71e7a294..00000000
--- a/projects/llm_framework/include/re2/re2/prefilter_tree.h
+++ /dev/null
@@ -1,152 +0,0 @@
-// Copyright 2009 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_PREFILTER_TREE_H_
-#define RE2_PREFILTER_TREE_H_
-
-// The PrefilterTree class is used to form an AND-OR tree of strings
-// that would trigger each regexp. The 'prefilter' of each regexp is
-// added to PrefilterTree, and then PrefilterTree is used to find all
-// the unique strings across the prefilters. During search, by using
-// matches from a string matching engine, PrefilterTree deduces the
-// set of regexps that are to be triggered. The 'string matching
-// engine' itself is outside of this class, and the caller can use any
-// favorite engine. PrefilterTree provides a set of strings (called
-// atoms) that the user of this class should use to do the string
-// matching.
-
-#include <string>
-#include <vector>
-
-#include "absl/container/flat_hash_set.h"
-#include "re2/prefilter.h"
-#include "re2/sparse_array.h"
-#include "util/logging.h"
-
-namespace re2 {
-
-class PrefilterTree {
- public:
-  PrefilterTree();
-  explicit PrefilterTree(int min_atom_len);
-  ~PrefilterTree();
-
-  // Adds the prefilter for the next regexp. Note that we assume that
-  // Add called sequentially for all regexps. All Add calls
-  // must precede Compile.
-  void Add(Prefilter* prefilter);
-
-  // The Compile returns a vector of string in atom_vec.
-  // Call this after all the prefilters are added through Add.
-  // No calls to Add after Compile are allowed.
-  // The caller should use the returned set of strings to do string matching.
-  // Each time a string matches, the corresponding index then has to be
-  // and passed to RegexpsGivenStrings below.
-  void Compile(std::vector<std::string>* atom_vec);
-
-  // Given the indices of the atoms that matched, returns the indexes
-  // of regexps that should be searched.  The matched_atoms should
-  // contain all the ids of string atoms that were found to match the
-  // content. The caller can use any string match engine to perform
-  // this function. This function is thread safe.
-  void RegexpsGivenStrings(const std::vector<int>& matched_atoms,
-                           std::vector<int>* regexps) const;
-
-  // Print debug prefilter. Also prints unique ids associated with
-  // nodes of the prefilter of the regexp.
-  void PrintPrefilter(int regexpid);
-
- private:
-  using IntMap = SparseArray<int>;
-
-  struct PrefilterHash {
-    size_t operator()(const Prefilter* a) const {
-      DCHECK(a != NULL);
-      return absl::Hash<Prefilter>()(*a);
-    }
-  };
-
-  struct PrefilterEqual {
-    bool operator()(const Prefilter* a, const Prefilter* b) const {
-      DCHECK(a != NULL);
-      DCHECK(b != NULL);
-      return *a == *b;
-    }
-  };
-
-  using NodeSet =
-      absl::flat_hash_set<Prefilter*, PrefilterHash, PrefilterEqual>;
-
-  // Each unique node has a corresponding Entry that helps in
-  // passing the matching trigger information along the tree.
-  struct Entry {
-   public:
-    // How many children should match before this node triggers the
-    // parent. For an atom and an OR node, this is 1 and for an AND
-    // node, it is the number of unique children.
-    int propagate_up_at_count;
-
-    // When this node is ready to trigger the parent, what are the indices
-    // of the parent nodes to trigger. The reason there may be more than
-    // one is because of sharing. For example (abc | def) and (xyz | def)
-    // are two different nodes, but they share the atom 'def'. So when
-    // 'def' matches, it triggers two parents, corresponding to the two
-    // different OR nodes.
-    std::vector<int> parents;
-
-    // When this node is ready to trigger the parent, what are the
-    // regexps that are triggered.
-    std::vector<int> regexps;
-  };
-
-  // Returns true if the prefilter node should be kept.
-  bool KeepNode(Prefilter* node) const;
-
-  // This function assigns unique ids to various parts of the
-  // prefilter, by looking at if these nodes are already in the
-  // PrefilterTree.
-  void AssignUniqueIds(NodeSet* nodes, std::vector<std::string>* atom_vec);
-
-  // Given the matching atoms, find the regexps to be triggered.
-  void PropagateMatch(const std::vector<int>& atom_ids,
-                      IntMap* regexps) const;
-
-  // Returns the prefilter node that has the same atom/subs as this
-  // node. For the canonical node, returns node. Assumes that the
-  // children of node have already been assigned unique ids.
-  Prefilter* CanonicalNode(NodeSet* nodes, Prefilter* node);
-
-  // Recursively constructs a readable prefilter string.
-  std::string DebugNodeString(Prefilter* node) const;
-
-  // Used for debugging.
-  void PrintDebugInfo(NodeSet* nodes);
-
-  // These are all the nodes formed by Compile. Essentially, there is
-  // one node for each unique atom and each unique AND/OR node.
-  std::vector<Entry> entries_;
-
-  // indices of regexps that always pass through the filter (since we
-  // found no required literals in these regexps).
-  std::vector<int> unfiltered_;
-
-  // vector of Prefilter for all regexps.
-  std::vector<Prefilter*> prefilter_vec_;
-
-  // Atom index in returned strings to entry id mapping.
-  std::vector<int> atom_index_to_id_;
-
-  // Has the prefilter tree been compiled.
-  bool compiled_;
-
-  // Strings less than this length are not stored as atoms.
-  const int min_atom_len_;
-
-  PrefilterTree(const PrefilterTree&) = delete;
-  PrefilterTree& operator=(const PrefilterTree&) = delete;
-};
-
-}  // namespace
-
-#endif  // RE2_PREFILTER_TREE_H_
diff --git a/projects/llm_framework/include/re2/re2/prog.h b/projects/llm_framework/include/re2/re2/prog.h
deleted file mode 100644
index 41923f31..00000000
--- a/projects/llm_framework/include/re2/re2/prog.h
+++ /dev/null
@@ -1,466 +0,0 @@
-// Copyright 2007 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_PROG_H_
-#define RE2_PROG_H_
-
-// Compiled representation of regular expressions.
-// See regexp.h for the Regexp class, which represents a regular
-// expression symbolically.
-
-#include <stdint.h>
-#include <functional>
-#include <string>
-#include <vector>
-#include <type_traits>
-
-#include "absl/base/call_once.h"
-#include "absl/strings/string_view.h"
-#include "util/logging.h"
-#include "re2/pod_array.h"
-#include "re2/re2.h"
-#include "re2/sparse_array.h"
-#include "re2/sparse_set.h"
-
-namespace re2 {
-
-// Opcodes for Inst
-enum InstOp {
-  kInstAlt = 0,      // choose between out_ and out1_
-  kInstAltMatch,     // Alt: out_ is [00-FF] and back, out1_ is match; or vice versa.
-  kInstByteRange,    // next (possible case-folded) byte must be in [lo_, hi_]
-  kInstCapture,      // capturing parenthesis number cap_
-  kInstEmptyWidth,   // empty-width special (^ $ ...); bit(s) set in empty_
-  kInstMatch,        // found a match!
-  kInstNop,          // no-op; occasionally unavoidable
-  kInstFail,         // never match; occasionally unavoidable
-  kNumInst,
-};
-
-// Bit flags for empty-width specials
-enum EmptyOp {
-  kEmptyBeginLine        = 1<<0,      // ^ - beginning of line
-  kEmptyEndLine          = 1<<1,      // $ - end of line
-  kEmptyBeginText        = 1<<2,      // \A - beginning of text
-  kEmptyEndText          = 1<<3,      // \z - end of text
-  kEmptyWordBoundary     = 1<<4,      // \b - word boundary
-  kEmptyNonWordBoundary  = 1<<5,      // \B - not \b
-  kEmptyAllFlags         = (1<<6)-1,
-};
-
-class DFA;
-class Regexp;
-
-// Compiled form of regexp program.
-class Prog {
- public:
-  Prog();
-  ~Prog();
-
-  // Single instruction in regexp program.
-  class Inst {
-   public:
-    // See the assertion below for why this is so.
-    Inst() = default;
-
-    // Copyable.
-    Inst(const Inst&) = default;
-    Inst& operator=(const Inst&) = default;
-
-    // Constructors per opcode
-    void InitAlt(uint32_t out, uint32_t out1);
-    void InitByteRange(int lo, int hi, int foldcase, uint32_t out);
-    void InitCapture(int cap, uint32_t out);
-    void InitEmptyWidth(EmptyOp empty, uint32_t out);
-    void InitMatch(int id);
-    void InitNop(uint32_t out);
-    void InitFail();
-
-    // Getters
-    int id(Prog* p) { return static_cast<int>(this - p->inst_.data()); }
-    InstOp opcode() { return static_cast<InstOp>(out_opcode_&7); }
-    int last()      { return (out_opcode_>>3)&1; }
-    int out()       { return out_opcode_>>4; }
-    int out1()      { DCHECK(opcode() == kInstAlt || opcode() == kInstAltMatch); return out1_; }
-    int cap()       { DCHECK_EQ(opcode(), kInstCapture); return cap_; }
-    int lo()        { DCHECK_EQ(opcode(), kInstByteRange); return lo_; }
-    int hi()        { DCHECK_EQ(opcode(), kInstByteRange); return hi_; }
-    int foldcase()  { DCHECK_EQ(opcode(), kInstByteRange); return hint_foldcase_&1; }
-    int hint()      { DCHECK_EQ(opcode(), kInstByteRange); return hint_foldcase_>>1; }
-    int match_id()  { DCHECK_EQ(opcode(), kInstMatch); return match_id_; }
-    EmptyOp empty() { DCHECK_EQ(opcode(), kInstEmptyWidth); return empty_; }
-
-    bool greedy(Prog* p) {
-      DCHECK_EQ(opcode(), kInstAltMatch);
-      return p->inst(out())->opcode() == kInstByteRange ||
-             (p->inst(out())->opcode() == kInstNop &&
-              p->inst(p->inst(out())->out())->opcode() == kInstByteRange);
-    }
-
-    // Does this inst (an kInstByteRange) match c?
-    inline bool Matches(int c) {
-      DCHECK_EQ(opcode(), kInstByteRange);
-      if (foldcase() && 'A' <= c && c <= 'Z')
-        c += 'a' - 'A';
-      return lo_ <= c && c <= hi_;
-    }
-
-    // Returns string representation for debugging.
-    std::string Dump();
-
-    // Maximum instruction id.
-    // (Must fit in out_opcode_. PatchList/last steal another bit.)
-    static const int kMaxInst = (1<<28) - 1;
-
-   private:
-    void set_opcode(InstOp opcode) {
-      out_opcode_ = (out()<<4) | (last()<<3) | opcode;
-    }
-
-    void set_last() {
-      out_opcode_ = (out()<<4) | (1<<3) | opcode();
-    }
-
-    void set_out(int out) {
-      out_opcode_ = (out<<4) | (last()<<3) | opcode();
-    }
-
-    void set_out_opcode(int out, InstOp opcode) {
-      out_opcode_ = (out<<4) | (last()<<3) | opcode;
-    }
-
-    uint32_t out_opcode_;  // 28 bits: out, 1 bit: last, 3 (low) bits: opcode
-    union {                // additional instruction arguments:
-      uint32_t out1_;      // opcode == kInstAlt
-                           //   alternate next instruction
-
-      int32_t cap_;        // opcode == kInstCapture
-                           //   Index of capture register (holds text
-                           //   position recorded by capturing parentheses).
-                           //   For \n (the submatch for the nth parentheses),
-                           //   the left parenthesis captures into register 2*n
-                           //   and the right one captures into register 2*n+1.
-
-      int32_t match_id_;   // opcode == kInstMatch
-                           //   Match ID to identify this match (for re2::Set).
-
-      struct {             // opcode == kInstByteRange
-        uint8_t lo_;       //   byte range is lo_-hi_ inclusive
-        uint8_t hi_;       //
-        uint16_t hint_foldcase_;  // 15 bits: hint, 1 (low) bit: foldcase
-                           //   hint to execution engines: the delta to the
-                           //   next instruction (in the current list) worth
-                           //   exploring iff this instruction matched; 0
-                           //   means there are no remaining possibilities,
-                           //   which is most likely for character classes.
-                           //   foldcase: A-Z -> a-z before checking range.
-      };
-
-      EmptyOp empty_;       // opcode == kInstEmptyWidth
-                            //   empty_ is bitwise OR of kEmpty* flags above.
-    };
-
-    friend class Compiler;
-    friend struct PatchList;
-    friend class Prog;
-  };
-
-  // Inst must be trivial so that we can freely clear it with memset(3).
-  // Arrays of Inst are initialised by copying the initial elements with
-  // memmove(3) and then clearing any remaining elements with memset(3).
-  static_assert(std::is_trivial<Inst>::value, "Inst must be trivial");
-
-  // Whether to anchor the search.
-  enum Anchor {
-    kUnanchored,  // match anywhere
-    kAnchored,    // match only starting at beginning of text
-  };
-
-  // Kind of match to look for (for anchor != kFullMatch)
-  //
-  // kLongestMatch mode finds the overall longest
-  // match but still makes its submatch choices the way
-  // Perl would, not in the way prescribed by POSIX.
-  // The POSIX rules are much more expensive to implement,
-  // and no one has needed them.
-  //
-  // kFullMatch is not strictly necessary -- we could use
-  // kLongestMatch and then check the length of the match -- but
-  // the matching code can run faster if it knows to consider only
-  // full matches.
-  enum MatchKind {
-    kFirstMatch,     // like Perl, PCRE
-    kLongestMatch,   // like egrep or POSIX
-    kFullMatch,      // match only entire text; implies anchor==kAnchored
-    kManyMatch       // for SearchDFA, records set of matches
-  };
-
-  Inst *inst(int id) { return &inst_[id]; }
-  int start() { return start_; }
-  void set_start(int start) { start_ = start; }
-  int start_unanchored() { return start_unanchored_; }
-  void set_start_unanchored(int start) { start_unanchored_ = start; }
-  int size() { return size_; }
-  bool reversed() { return reversed_; }
-  void set_reversed(bool reversed) { reversed_ = reversed; }
-  int list_count() { return list_count_; }
-  int inst_count(InstOp op) { return inst_count_[op]; }
-  uint16_t* list_heads() { return list_heads_.data(); }
-  size_t bit_state_text_max_size() { return bit_state_text_max_size_; }
-  int64_t dfa_mem() { return dfa_mem_; }
-  void set_dfa_mem(int64_t dfa_mem) { dfa_mem_ = dfa_mem; }
-  bool anchor_start() { return anchor_start_; }
-  void set_anchor_start(bool b) { anchor_start_ = b; }
-  bool anchor_end() { return anchor_end_; }
-  void set_anchor_end(bool b) { anchor_end_ = b; }
-  int bytemap_range() { return bytemap_range_; }
-  const uint8_t* bytemap() { return bytemap_; }
-  bool can_prefix_accel() { return prefix_size_ != 0; }
-
-  // Accelerates to the first likely occurrence of the prefix.
-  // Returns a pointer to the first byte or NULL if not found.
-  const void* PrefixAccel(const void* data, size_t size) {
-    DCHECK(can_prefix_accel());
-    if (prefix_foldcase_) {
-      return PrefixAccel_ShiftDFA(data, size);
-    } else if (prefix_size_ != 1) {
-      return PrefixAccel_FrontAndBack(data, size);
-    } else {
-      return memchr(data, prefix_front_, size);
-    }
-  }
-
-  // Configures prefix accel using the analysis performed during compilation.
-  void ConfigurePrefixAccel(const std::string& prefix, bool prefix_foldcase);
-
-  // An implementation of prefix accel that uses prefix_dfa_ to perform
-  // case-insensitive search.
-  const void* PrefixAccel_ShiftDFA(const void* data, size_t size);
-
-  // An implementation of prefix accel that looks for prefix_front_ and
-  // prefix_back_ to return fewer false positives than memchr(3) alone.
-  const void* PrefixAccel_FrontAndBack(const void* data, size_t size);
-
-  // Returns string representation of program for debugging.
-  std::string Dump();
-  std::string DumpUnanchored();
-  std::string DumpByteMap();
-
-  // Returns the set of kEmpty flags that are in effect at
-  // position p within context.
-  static uint32_t EmptyFlags(absl::string_view context, const char* p);
-
-  // Returns whether byte c is a word character: ASCII only.
-  // Used by the implementation of \b and \B.
-  // This is not right for Unicode, but:
-  //   - it's hard to get right in a byte-at-a-time matching world
-  //     (the DFA has only one-byte lookahead).
-  //   - even if the lookahead were possible, the Progs would be huge.
-  // This crude approximation is the same one PCRE uses.
-  static bool IsWordChar(uint8_t c) {
-    return ('A' <= c && c <= 'Z') ||
-           ('a' <= c && c <= 'z') ||
-           ('0' <= c && c <= '9') ||
-           c == '_';
-  }
-
-  // Execution engines.  They all search for the regexp (run the prog)
-  // in text, which is in the larger context (used for ^ $ \b etc).
-  // Anchor and kind control the kind of search.
-  // Returns true if match found, false if not.
-  // If match found, fills match[0..nmatch-1] with submatch info.
-  // match[0] is overall match, match[1] is first set of parens, etc.
-  // If a particular submatch is not matched during the regexp match,
-  // it is set to NULL.
-  //
-  // Matching text == absl::string_view() is treated as any other empty
-  // string, but note that on return, it will not be possible to distinguish
-  // submatches that matched that empty string from submatches that didn't
-  // match anything.  Either way, match[i] == NULL.
-
-  // Search using NFA: can find submatches but kind of slow.
-  bool SearchNFA(absl::string_view text, absl::string_view context,
-                 Anchor anchor, MatchKind kind, absl::string_view* match,
-                 int nmatch);
-
-  // Search using DFA: much faster than NFA but only finds
-  // end of match and can use a lot more memory.
-  // Returns whether a match was found.
-  // If the DFA runs out of memory, sets *failed to true and returns false.
-  // If matches != NULL and kind == kManyMatch and there is a match,
-  // SearchDFA fills matches with the match IDs of the final matching state.
-  bool SearchDFA(absl::string_view text, absl::string_view context,
-                 Anchor anchor, MatchKind kind, absl::string_view* match0,
-                 bool* failed, SparseSet* matches);
-
-  // The callback issued after building each DFA state with BuildEntireDFA().
-  // If next is null, then the memory budget has been exhausted and building
-  // will halt. Otherwise, the state has been built and next points to an array
-  // of bytemap_range()+1 slots holding the next states as per the bytemap and
-  // kByteEndText. The number of the state is implied by the callback sequence:
-  // the first callback is for state 0, the second callback is for state 1, ...
-  // match indicates whether the state is a matching state.
-  using DFAStateCallback = std::function<void(const int* next, bool match)>;
-
-  // Build the entire DFA for the given match kind.
-  // Usually the DFA is built out incrementally, as needed, which
-  // avoids lots of unnecessary work.
-  // If cb is not empty, it receives one callback per state built.
-  // Returns the number of states built.
-  // FOR TESTING OR EXPERIMENTAL PURPOSES ONLY.
-  int BuildEntireDFA(MatchKind kind, const DFAStateCallback& cb);
-
-  // Compute bytemap.
-  void ComputeByteMap();
-
-  // Run peep-hole optimizer on program.
-  void Optimize();
-
-  // One-pass NFA: only correct if IsOnePass() is true,
-  // but much faster than NFA (competitive with PCRE)
-  // for those expressions.
-  bool IsOnePass();
-  bool SearchOnePass(absl::string_view text, absl::string_view context,
-                     Anchor anchor, MatchKind kind, absl::string_view* match,
-                     int nmatch);
-
-  // Bit-state backtracking.  Fast on small cases but uses memory
-  // proportional to the product of the list count and the text size.
-  bool CanBitState() { return list_heads_.data() != NULL; }
-  bool SearchBitState(absl::string_view text, absl::string_view context,
-                      Anchor anchor, MatchKind kind, absl::string_view* match,
-                      int nmatch);
-
-  static const int kMaxOnePassCapture = 5;  // $0 through $4
-
-  // Backtracking search: the gold standard against which the other
-  // implementations are checked.  FOR TESTING ONLY.
-  // It allocates a ton of memory to avoid running forever.
-  // It is also recursive, so can't use in production (will overflow stacks).
-  // The name "Unsafe" here is supposed to be a flag that
-  // you should not be using this function.
-  bool UnsafeSearchBacktrack(absl::string_view text, absl::string_view context,
-                             Anchor anchor, MatchKind kind,
-                             absl::string_view* match, int nmatch);
-
-  // Computes range for any strings matching regexp. The min and max can in
-  // some cases be arbitrarily precise, so the caller gets to specify the
-  // maximum desired length of string returned.
-  //
-  // Assuming PossibleMatchRange(&min, &max, N) returns successfully, any
-  // string s that is an anchored match for this regexp satisfies
-  //   min <= s && s <= max.
-  //
-  // Note that PossibleMatchRange() will only consider the first copy of an
-  // infinitely repeated element (i.e., any regexp element followed by a '*' or
-  // '+' operator). Regexps with "{N}" constructions are not affected, as those
-  // do not compile down to infinite repetitions.
-  //
-  // Returns true on success, false on error.
-  bool PossibleMatchRange(std::string* min, std::string* max, int maxlen);
-
-  // Outputs the program fanout into the given sparse array.
-  void Fanout(SparseArray<int>* fanout);
-
-  // Compiles a collection of regexps to Prog.  Each regexp will have
-  // its own Match instruction recording the index in the output vector.
-  static Prog* CompileSet(Regexp* re, RE2::Anchor anchor, int64_t max_mem);
-
-  // Flattens the Prog from "tree" form to "list" form. This is an in-place
-  // operation in the sense that the old instructions are lost.
-  void Flatten();
-
-  // Walks the Prog; the "successor roots" or predecessors of the reachable
-  // instructions are marked in rootmap or predmap/predvec, respectively.
-  // reachable and stk are preallocated scratch structures.
-  void MarkSuccessors(SparseArray<int>* rootmap,
-                      SparseArray<int>* predmap,
-                      std::vector<std::vector<int>>* predvec,
-                      SparseSet* reachable, std::vector<int>* stk);
-
-  // Walks the Prog from the given "root" instruction; the "dominator root"
-  // of the reachable instructions (if such exists) is marked in rootmap.
-  // reachable and stk are preallocated scratch structures.
-  void MarkDominator(int root, SparseArray<int>* rootmap,
-                     SparseArray<int>* predmap,
-                     std::vector<std::vector<int>>* predvec,
-                     SparseSet* reachable, std::vector<int>* stk);
-
-  // Walks the Prog from the given "root" instruction; the reachable
-  // instructions are emitted in "list" form and appended to flat.
-  // reachable and stk are preallocated scratch structures.
-  void EmitList(int root, SparseArray<int>* rootmap,
-                std::vector<Inst>* flat,
-                SparseSet* reachable, std::vector<int>* stk);
-
-  // Computes hints for ByteRange instructions in [begin, end).
-  void ComputeHints(std::vector<Inst>* flat, int begin, int end);
-
-  // Controls whether the DFA should bail out early if the NFA would be faster.
-  // FOR TESTING ONLY.
-  static void TESTING_ONLY_set_dfa_should_bail_when_slow(bool b);
-
- private:
-  friend class Compiler;
-
-  DFA* GetDFA(MatchKind kind);
-  void DeleteDFA(DFA* dfa);
-
-  bool anchor_start_;       // regexp has explicit start anchor
-  bool anchor_end_;         // regexp has explicit end anchor
-  bool reversed_;           // whether program runs backward over input
-  bool did_flatten_;        // has Flatten been called?
-  bool did_onepass_;        // has IsOnePass been called?
-
-  int start_;               // entry point for program
-  int start_unanchored_;    // unanchored entry point for program
-  int size_;                // number of instructions
-  int bytemap_range_;       // bytemap_[x] < bytemap_range_
-
-  bool prefix_foldcase_;    // whether prefix is case-insensitive
-  size_t prefix_size_;      // size of prefix (0 if no prefix)
-  union {
-    uint64_t* prefix_dfa_;  // "Shift DFA" for prefix
-    struct {
-      int prefix_front_;    // first byte of prefix
-      int prefix_back_;     // last byte of prefix
-    };
-  };
-
-  int list_count_;                  // count of lists (see above)
-  int inst_count_[kNumInst];        // count of instructions by opcode
-  PODArray<uint16_t> list_heads_;   // sparse array enumerating list heads
-                                    // not populated if size_ is overly large
-  size_t bit_state_text_max_size_;  // upper bound (inclusive) on text.size()
-
-  PODArray<Inst> inst_;              // pointer to instruction array
-  PODArray<uint8_t> onepass_nodes_;  // data for OnePass nodes
-
-  int64_t dfa_mem_;         // Maximum memory for DFAs.
-  DFA* dfa_first_;          // DFA cached for kFirstMatch/kManyMatch
-  DFA* dfa_longest_;        // DFA cached for kLongestMatch/kFullMatch
-
-  uint8_t bytemap_[256];    // map from input bytes to byte classes
-
-  absl::once_flag dfa_first_once_;
-  absl::once_flag dfa_longest_once_;
-
-  Prog(const Prog&) = delete;
-  Prog& operator=(const Prog&) = delete;
-};
-
-// std::string_view in MSVC has iterators that aren't just pointers and
-// that don't allow comparisons between different objects - not even if
-// those objects are views into the same string! Thus, we provide these
-// conversion functions for convenience.
-static inline const char* BeginPtr(absl::string_view s) {
-  return s.data();
-}
-static inline const char* EndPtr(absl::string_view s) {
-  return s.data() + s.size();
-}
-
-}  // namespace re2
-
-#endif  // RE2_PROG_H_
diff --git a/projects/llm_framework/include/re2/re2/re2.h b/projects/llm_framework/include/re2/re2/re2.h
deleted file mode 100644
index 7ea44e04..00000000
--- a/projects/llm_framework/include/re2/re2/re2.h
+++ /dev/null
@@ -1,1078 +0,0 @@
-// Copyright 2003-2009 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_RE2_H_
-#define RE2_RE2_H_
-
-// C++ interface to the re2 regular-expression library.
-// RE2 supports Perl-style regular expressions (with extensions like
-// \d, \w, \s, ...).
-//
-// -----------------------------------------------------------------------
-// REGEXP SYNTAX:
-//
-// This module uses the re2 library and hence supports
-// its syntax for regular expressions, which is similar to Perl's with
-// some of the more complicated things thrown away.  In particular,
-// backreferences and generalized assertions are not available, nor is \Z.
-//
-// See https://github.com/google/re2/wiki/Syntax for the syntax
-// supported by RE2, and a comparison with PCRE and PERL regexps.
-//
-// For those not familiar with Perl's regular expressions,
-// here are some examples of the most commonly used extensions:
-//
-//   "hello (\\w+) world"  -- \w matches a "word" character
-//   "version (\\d+)"      -- \d matches a digit
-//   "hello\\s+world"      -- \s matches any whitespace character
-//   "\\b(\\w+)\\b"        -- \b matches non-empty string at word boundary
-//   "(?i)hello"           -- (?i) turns on case-insensitive matching
-//   "/\\*(.*?)\\*/"       -- .*? matches . minimum no. of times possible
-//
-// The double backslashes are needed when writing C++ string literals.
-// However, they should NOT be used when writing C++11 raw string literals:
-//
-//   R"(hello (\w+) world)"  -- \w matches a "word" character
-//   R"(version (\d+))"      -- \d matches a digit
-//   R"(hello\s+world)"      -- \s matches any whitespace character
-//   R"(\b(\w+)\b)"          -- \b matches non-empty string at word boundary
-//   R"((?i)hello)"          -- (?i) turns on case-insensitive matching
-//   R"(/\*(.*?)\*/)"        -- .*? matches . minimum no. of times possible
-//
-// When using UTF-8 encoding, case-insensitive matching will perform
-// simple case folding, not full case folding.
-//
-// -----------------------------------------------------------------------
-// MATCHING INTERFACE:
-//
-// The "FullMatch" operation checks that supplied text matches a
-// supplied pattern exactly.
-//
-// Example: successful match
-//    CHECK(RE2::FullMatch("hello", "h.*o"));
-//
-// Example: unsuccessful match (requires full match):
-//    CHECK(!RE2::FullMatch("hello", "e"));
-//
-// -----------------------------------------------------------------------
-// UTF-8 AND THE MATCHING INTERFACE:
-//
-// By default, the pattern and input text are interpreted as UTF-8.
-// The RE2::Latin1 option causes them to be interpreted as Latin-1.
-//
-// Example:
-//    CHECK(RE2::FullMatch(utf8_string, RE2(utf8_pattern)));
-//    CHECK(RE2::FullMatch(latin1_string, RE2(latin1_pattern, RE2::Latin1)));
-//
-// -----------------------------------------------------------------------
-// SUBMATCH EXTRACTION:
-//
-// You can supply extra pointer arguments to extract submatches.
-// On match failure, none of the pointees will have been modified.
-// On match success, the submatches will be converted (as necessary) and
-// their values will be assigned to their pointees until all conversions
-// have succeeded or one conversion has failed.
-// On conversion failure, the pointees will be in an indeterminate state
-// because the caller has no way of knowing which conversion failed.
-// However, conversion cannot fail for types like string and string_view
-// that do not inspect the submatch contents. Hence, in the common case
-// where all of the pointees are of such types, failure is always due to
-// match failure and thus none of the pointees will have been modified.
-//
-// Example: extracts "ruby" into "s" and 1234 into "i"
-//    int i;
-//    std::string s;
-//    CHECK(RE2::FullMatch("ruby:1234", "(\\w+):(\\d+)", &s, &i));
-//
-// Example: extracts "ruby" into "s" and no value into "i"
-//    absl::optional<int> i;
-//    std::string s;
-//    CHECK(RE2::FullMatch("ruby", "(\\w+)(?::(\\d+))?", &s, &i));
-//
-// Example: fails because string cannot be stored in integer
-//    CHECK(!RE2::FullMatch("ruby", "(.*)", &i));
-//
-// Example: fails because there aren't enough sub-patterns
-//    CHECK(!RE2::FullMatch("ruby:1234", "\\w+:\\d+", &s));
-//
-// Example: does not try to extract any extra sub-patterns
-//    CHECK(RE2::FullMatch("ruby:1234", "(\\w+):(\\d+)", &s));
-//
-// Example: does not try to extract into NULL
-//    CHECK(RE2::FullMatch("ruby:1234", "(\\w+):(\\d+)", NULL, &i));
-//
-// Example: integer overflow causes failure
-//    CHECK(!RE2::FullMatch("ruby:1234567891234", "\\w+:(\\d+)", &i));
-//
-// NOTE(rsc): Asking for submatches slows successful matches quite a bit.
-// This may get a little faster in the future, but right now is slower
-// than PCRE.  On the other hand, failed matches run *very* fast (faster
-// than PCRE), as do matches without submatch extraction.
-//
-// -----------------------------------------------------------------------
-// PARTIAL MATCHES
-//
-// You can use the "PartialMatch" operation when you want the pattern
-// to match any substring of the text.
-//
-// Example: simple search for a string:
-//      CHECK(RE2::PartialMatch("hello", "ell"));
-//
-// Example: find first number in a string
-//      int number;
-//      CHECK(RE2::PartialMatch("x*100 + 20", "(\\d+)", &number));
-//      CHECK_EQ(number, 100);
-//
-// -----------------------------------------------------------------------
-// PRE-COMPILED REGULAR EXPRESSIONS
-//
-// RE2 makes it easy to use any string as a regular expression, without
-// requiring a separate compilation step.
-//
-// If speed is of the essence, you can create a pre-compiled "RE2"
-// object from the pattern and use it multiple times.  If you do so,
-// you can typically parse text faster than with sscanf.
-//
-// Example: precompile pattern for faster matching:
-//    RE2 pattern("h.*o");
-//    while (ReadLine(&str)) {
-//      if (RE2::FullMatch(str, pattern)) ...;
-//    }
-//
-// -----------------------------------------------------------------------
-// SCANNING TEXT INCREMENTALLY
-//
-// The "Consume" operation may be useful if you want to repeatedly
-// match regular expressions at the front of a string and skip over
-// them as they match.  This requires use of the string_view type,
-// which represents a sub-range of a real string.
-//
-// Example: read lines of the form "var = value" from a string.
-//      std::string contents = ...;         // Fill string somehow
-//      absl::string_view input(contents);  // Wrap a string_view around it
-//
-//      std::string var;
-//      int value;
-//      while (RE2::Consume(&input, "(\\w+) = (\\d+)\n", &var, &value)) {
-//        ...;
-//      }
-//
-// Each successful call to "Consume" will set "var/value", and also
-// advance "input" so it points past the matched text.  Note that if the
-// regular expression matches an empty string, input will advance
-// by 0 bytes.  If the regular expression being used might match
-// an empty string, the loop body must check for this case and either
-// advance the string or break out of the loop.
-//
-// The "FindAndConsume" operation is similar to "Consume" but does not
-// anchor your match at the beginning of the string.  For example, you
-// could extract all words from a string by repeatedly calling
-//     RE2::FindAndConsume(&input, "(\\w+)", &word)
-//
-// -----------------------------------------------------------------------
-// USING VARIABLE NUMBER OF ARGUMENTS
-//
-// The above operations require you to know the number of arguments
-// when you write the code.  This is not always possible or easy (for
-// example, the regular expression may be calculated at run time).
-// You can use the "N" version of the operations when the number of
-// match arguments are determined at run time.
-//
-// Example:
-//   const RE2::Arg* args[10];
-//   int n;
-//   // ... populate args with pointers to RE2::Arg values ...
-//   // ... set n to the number of RE2::Arg objects ...
-//   bool match = RE2::FullMatchN(input, pattern, args, n);
-//
-// The last statement is equivalent to
-//
-//   bool match = RE2::FullMatch(input, pattern,
-//                               *args[0], *args[1], ..., *args[n - 1]);
-//
-// -----------------------------------------------------------------------
-// PARSING HEX/OCTAL/C-RADIX NUMBERS
-//
-// By default, if you pass a pointer to a numeric value, the
-// corresponding text is interpreted as a base-10 number.  You can
-// instead wrap the pointer with a call to one of the operators Hex(),
-// Octal(), or CRadix() to interpret the text in another base.  The
-// CRadix operator interprets C-style "0" (base-8) and "0x" (base-16)
-// prefixes, but defaults to base-10.
-//
-// Example:
-//   int a, b, c, d;
-//   CHECK(RE2::FullMatch("100 40 0100 0x40", "(.*) (.*) (.*) (.*)",
-//         RE2::Octal(&a), RE2::Hex(&b), RE2::CRadix(&c), RE2::CRadix(&d));
-// will leave 64 in a, b, c, and d.
-
-#include <stddef.h>
-#include <stdint.h>
-#include <algorithm>
-#include <map>
-#include <string>
-#include <type_traits>
-#include <vector>
-
-#if defined(__APPLE__)
-#include <TargetConditionals.h>
-#endif
-
-#include "absl/base/call_once.h"
-#include "absl/strings/string_view.h"
-#include "absl/types/optional.h"
-#include "re2/stringpiece.h"
-
-namespace re2 {
-class Prog;
-class Regexp;
-}  // namespace re2
-
-namespace re2 {
-
-// Interface for regular expression matching.  Also corresponds to a
-// pre-compiled regular expression.  An "RE2" object is safe for
-// concurrent use by multiple threads.
-class RE2 {
- public:
-  // We convert user-passed pointers into special Arg objects
-  class Arg;
-  class Options;
-
-  // Defined in set.h.
-  class Set;
-
-  enum ErrorCode {
-    NoError = 0,
-
-    // Unexpected error
-    ErrorInternal,
-
-    // Parse errors
-    ErrorBadEscape,          // bad escape sequence
-    ErrorBadCharClass,       // bad character class
-    ErrorBadCharRange,       // bad character class range
-    ErrorMissingBracket,     // missing closing ]
-    ErrorMissingParen,       // missing closing )
-    ErrorUnexpectedParen,    // unexpected closing )
-    ErrorTrailingBackslash,  // trailing \ at end of regexp
-    ErrorRepeatArgument,     // repeat argument missing, e.g. "*"
-    ErrorRepeatSize,         // bad repetition argument
-    ErrorRepeatOp,           // bad repetition operator
-    ErrorBadPerlOp,          // bad perl operator
-    ErrorBadUTF8,            // invalid UTF-8 in regexp
-    ErrorBadNamedCapture,    // bad named capture group
-    ErrorPatternTooLarge     // pattern too large (compile failed)
-  };
-
-  // Predefined common options.
-  // If you need more complicated things, instantiate
-  // an Option class, possibly passing one of these to
-  // the Option constructor, change the settings, and pass that
-  // Option class to the RE2 constructor.
-  enum CannedOptions {
-    DefaultOptions = 0,
-    Latin1, // treat input as Latin-1 (default UTF-8)
-    POSIX, // POSIX syntax, leftmost-longest match
-    Quiet // do not log about regexp parse errors
-  };
-
-  // Need to have the const char* and const std::string& forms for implicit
-  // conversions when passing string literals to FullMatch and PartialMatch.
-  // Otherwise the absl::string_view form would be sufficient.
-  RE2(const char* pattern);
-  RE2(const std::string& pattern);
-  RE2(absl::string_view pattern);
-  RE2(absl::string_view pattern, const Options& options);
-  ~RE2();
-
-  // Not copyable.
-  // RE2 objects are expensive. You should probably use std::shared_ptr<RE2>
-  // instead. If you really must copy, RE2(first.pattern(), first.options())
-  // effectively does so: it produces a second object that mimics the first.
-  RE2(const RE2&) = delete;
-  RE2& operator=(const RE2&) = delete;
-  // Not movable.
-  // RE2 objects are thread-safe and logically immutable. You should probably
-  // use std::unique_ptr<RE2> instead. Otherwise, consider std::deque<RE2> if
-  // direct emplacement into a container is desired. If you really must move,
-  // be prepared to submit a design document along with your feature request.
-  RE2(RE2&&) = delete;
-  RE2& operator=(RE2&&) = delete;
-
-  // Returns whether RE2 was created properly.
-  bool ok() const { return error_code() == NoError; }
-
-  // The string specification for this RE2.  E.g.
-  //   RE2 re("ab*c?d+");
-  //   re.pattern();    // "ab*c?d+"
-  const std::string& pattern() const { return *pattern_; }
-
-  // If RE2 could not be created properly, returns an error string.
-  // Else returns the empty string.
-  const std::string& error() const { return *error_; }
-
-  // If RE2 could not be created properly, returns an error code.
-  // Else returns RE2::NoError (== 0).
-  ErrorCode error_code() const { return error_code_; }
-
-  // If RE2 could not be created properly, returns the offending
-  // portion of the regexp.
-  const std::string& error_arg() const { return *error_arg_; }
-
-  // Returns the program size, a very approximate measure of a regexp's "cost".
-  // Larger numbers are more expensive than smaller numbers.
-  int ProgramSize() const;
-  int ReverseProgramSize() const;
-
-  // If histogram is not null, outputs the program fanout
-  // as a histogram bucketed by powers of 2.
-  // Returns the number of the largest non-empty bucket.
-  int ProgramFanout(std::vector<int>* histogram) const;
-  int ReverseProgramFanout(std::vector<int>* histogram) const;
-
-  // Returns the underlying Regexp; not for general use.
-  // Returns entire_regexp_ so that callers don't need
-  // to know about prefix_ and prefix_foldcase_.
-  re2::Regexp* Regexp() const { return entire_regexp_; }
-
-  /***** The array-based matching interface ******/
-
-  // The functions here have names ending in 'N' and are used to implement
-  // the functions whose names are the prefix before the 'N'. It is sometimes
-  // useful to invoke them directly, but the syntax is awkward, so the 'N'-less
-  // versions should be preferred.
-  static bool FullMatchN(absl::string_view text, const RE2& re,
-                         const Arg* const args[], int n);
-  static bool PartialMatchN(absl::string_view text, const RE2& re,
-                            const Arg* const args[], int n);
-  static bool ConsumeN(absl::string_view* input, const RE2& re,
-                       const Arg* const args[], int n);
-  static bool FindAndConsumeN(absl::string_view* input, const RE2& re,
-                              const Arg* const args[], int n);
-
- private:
-  template <typename F, typename SP>
-  static inline bool Apply(F f, SP sp, const RE2& re) {
-    return f(sp, re, NULL, 0);
-  }
-
-  template <typename F, typename SP, typename... A>
-  static inline bool Apply(F f, SP sp, const RE2& re, const A&... a) {
-    const Arg* const args[] = {&a...};
-    const int n = sizeof...(a);
-    return f(sp, re, args, n);
-  }
-
- public:
-  // In order to allow FullMatch() et al. to be called with a varying number
-  // of arguments of varying types, we use two layers of variadic templates.
-  // The first layer constructs the temporary Arg objects. The second layer
-  // (above) constructs the array of pointers to the temporary Arg objects.
-
-  /***** The useful part: the matching interface *****/
-
-  // Matches "text" against "re".  If pointer arguments are
-  // supplied, copies matched sub-patterns into them.
-  //
-  // You can pass in a "const char*" or a "std::string" for "text".
-  // You can pass in a "const char*" or a "std::string" or a "RE2" for "re".
-  //
-  // The provided pointer arguments can be pointers to any scalar numeric
-  // type, or one of:
-  //    std::string        (matched piece is copied to string)
-  //    absl::string_view  (string_view is mutated to point to matched piece)
-  //    absl::optional<T>  (T is a supported numeric or string type as above)
-  //    T                  ("bool T::ParseFrom(const char*, size_t)" must exist)
-  //    (void*)NULL        (the corresponding matched sub-pattern is not copied)
-  //
-  // Returns true iff all of the following conditions are satisfied:
-  //   a. "text" matches "re" fully - from the beginning to the end of "text".
-  //   b. The number of matched sub-patterns is >= number of supplied pointers.
-  //   c. The "i"th argument has a suitable type for holding the
-  //      string captured as the "i"th sub-pattern.  If you pass in
-  //      NULL for the "i"th argument, or pass fewer arguments than
-  //      number of sub-patterns, the "i"th captured sub-pattern is
-  //      ignored.
-  //
-  // CAVEAT: An optional sub-pattern that does not exist in the
-  // matched string is assigned the null string.  Therefore, the
-  // following returns false because the null string - absence of
-  // a string (not even the empty string) - is not a valid number:
-  //
-  //    int number;
-  //    RE2::FullMatch("abc", "[a-z]+(\\d+)?", &number);
-  //
-  // Use absl::optional<int> instead to handle this case correctly.
-  template <typename... A>
-  static bool FullMatch(absl::string_view text, const RE2& re, A&&... a) {
-    return Apply(FullMatchN, text, re, Arg(std::forward<A>(a))...);
-  }
-
-  // Like FullMatch(), except that "re" is allowed to match a substring
-  // of "text".
-  //
-  // Returns true iff all of the following conditions are satisfied:
-  //   a. "text" matches "re" partially - for some substring of "text".
-  //   b. The number of matched sub-patterns is >= number of supplied pointers.
-  //   c. The "i"th argument has a suitable type for holding the
-  //      string captured as the "i"th sub-pattern.  If you pass in
-  //      NULL for the "i"th argument, or pass fewer arguments than
-  //      number of sub-patterns, the "i"th captured sub-pattern is
-  //      ignored.
-  template <typename... A>
-  static bool PartialMatch(absl::string_view text, const RE2& re, A&&... a) {
-    return Apply(PartialMatchN, text, re, Arg(std::forward<A>(a))...);
-  }
-
-  // Like FullMatch() and PartialMatch(), except that "re" has to match
-  // a prefix of the text, and "input" is advanced past the matched
-  // text.  Note: "input" is modified iff this routine returns true
-  // and "re" matched a non-empty substring of "input".
-  //
-  // Returns true iff all of the following conditions are satisfied:
-  //   a. "input" matches "re" partially - for some prefix of "input".
-  //   b. The number of matched sub-patterns is >= number of supplied pointers.
-  //   c. The "i"th argument has a suitable type for holding the
-  //      string captured as the "i"th sub-pattern.  If you pass in
-  //      NULL for the "i"th argument, or pass fewer arguments than
-  //      number of sub-patterns, the "i"th captured sub-pattern is
-  //      ignored.
-  template <typename... A>
-  static bool Consume(absl::string_view* input, const RE2& re, A&&... a) {
-    return Apply(ConsumeN, input, re, Arg(std::forward<A>(a))...);
-  }
-
-  // Like Consume(), but does not anchor the match at the beginning of
-  // the text.  That is, "re" need not start its match at the beginning
-  // of "input".  For example, "FindAndConsume(s, "(\\w+)", &word)" finds
-  // the next word in "s" and stores it in "word".
-  //
-  // Returns true iff all of the following conditions are satisfied:
-  //   a. "input" matches "re" partially - for some substring of "input".
-  //   b. The number of matched sub-patterns is >= number of supplied pointers.
-  //   c. The "i"th argument has a suitable type for holding the
-  //      string captured as the "i"th sub-pattern.  If you pass in
-  //      NULL for the "i"th argument, or pass fewer arguments than
-  //      number of sub-patterns, the "i"th captured sub-pattern is
-  //      ignored.
-  template <typename... A>
-  static bool FindAndConsume(absl::string_view* input, const RE2& re, A&&... a) {
-    return Apply(FindAndConsumeN, input, re, Arg(std::forward<A>(a))...);
-  }
-
-  // Replace the first match of "re" in "str" with "rewrite".
-  // Within "rewrite", backslash-escaped digits (\1 to \9) can be
-  // used to insert text matching corresponding parenthesized group
-  // from the pattern.  \0 in "rewrite" refers to the entire matching
-  // text.  E.g.,
-  //
-  //   std::string s = "yabba dabba doo";
-  //   CHECK(RE2::Replace(&s, "b+", "d"));
-  //
-  // will leave "s" containing "yada dabba doo"
-  //
-  // Returns true if the pattern matches and a replacement occurs,
-  // false otherwise.
-  static bool Replace(std::string* str,
-                      const RE2& re,
-                      absl::string_view rewrite);
-
-  // Like Replace(), except replaces successive non-overlapping occurrences
-  // of the pattern in the string with the rewrite. E.g.
-  //
-  //   std::string s = "yabba dabba doo";
-  //   CHECK(RE2::GlobalReplace(&s, "b+", "d"));
-  //
-  // will leave "s" containing "yada dada doo"
-  // Replacements are not subject to re-matching.
-  //
-  // Because GlobalReplace only replaces non-overlapping matches,
-  // replacing "ana" within "banana" makes only one replacement, not two.
-  //
-  // Returns the number of replacements made.
-  static int GlobalReplace(std::string* str,
-                           const RE2& re,
-                           absl::string_view rewrite);
-
-  // Like Replace, except that if the pattern matches, "rewrite"
-  // is copied into "out" with substitutions.  The non-matching
-  // portions of "text" are ignored.
-  //
-  // Returns true iff a match occurred and the extraction happened
-  // successfully;  if no match occurs, the string is left unaffected.
-  //
-  // REQUIRES: "text" must not alias any part of "*out".
-  static bool Extract(absl::string_view text,
-                      const RE2& re,
-                      absl::string_view rewrite,
-                      std::string* out);
-
-  // Escapes all potentially meaningful regexp characters in
-  // 'unquoted'.  The returned string, used as a regular expression,
-  // will match exactly the original string.  For example,
-  //           1.5-2.0?
-  // may become:
-  //           1\.5\-2\.0\?
-  static std::string QuoteMeta(absl::string_view unquoted);
-
-  // Computes range for any strings matching regexp. The min and max can in
-  // some cases be arbitrarily precise, so the caller gets to specify the
-  // maximum desired length of string returned.
-  //
-  // Assuming PossibleMatchRange(&min, &max, N) returns successfully, any
-  // string s that is an anchored match for this regexp satisfies
-  //   min <= s && s <= max.
-  //
-  // Note that PossibleMatchRange() will only consider the first copy of an
-  // infinitely repeated element (i.e., any regexp element followed by a '*' or
-  // '+' operator). Regexps with "{N}" constructions are not affected, as those
-  // do not compile down to infinite repetitions.
-  //
-  // Returns true on success, false on error.
-  bool PossibleMatchRange(std::string* min, std::string* max,
-                          int maxlen) const;
-
-  // Generic matching interface
-
-  // Type of match.
-  enum Anchor {
-    UNANCHORED,         // No anchoring
-    ANCHOR_START,       // Anchor at start only
-    ANCHOR_BOTH         // Anchor at start and end
-  };
-
-  // Return the number of capturing sub-patterns, or -1 if the
-  // regexp wasn't valid on construction.  The overall match ($0)
-  // does not count: if the regexp is "(a)(b)", returns 2.
-  int NumberOfCapturingGroups() const { return num_captures_; }
-
-  // Return a map from names to capturing indices.
-  // The map records the index of the leftmost group
-  // with the given name.
-  // Only valid until the re is deleted.
-  const std::map<std::string, int>& NamedCapturingGroups() const;
-
-  // Return a map from capturing indices to names.
-  // The map has no entries for unnamed groups.
-  // Only valid until the re is deleted.
-  const std::map<int, std::string>& CapturingGroupNames() const;
-
-  // General matching routine.
-  // Match against text starting at offset startpos
-  // and stopping the search at offset endpos.
-  // Returns true if match found, false if not.
-  // On a successful match, fills in submatch[] (up to nsubmatch entries)
-  // with information about submatches.
-  // I.e. matching RE2("(foo)|(bar)baz") on "barbazbla" will return true, with
-  // submatch[0] = "barbaz", submatch[1].data() = NULL, submatch[2] = "bar",
-  // submatch[3].data() = NULL, ..., up to submatch[nsubmatch-1].data() = NULL.
-  // Caveat: submatch[] may be clobbered even on match failure.
-  //
-  // Don't ask for more match information than you will use:
-  // runs much faster with nsubmatch == 1 than nsubmatch > 1, and
-  // runs even faster if nsubmatch == 0.
-  // Doesn't make sense to use nsubmatch > 1 + NumberOfCapturingGroups(),
-  // but will be handled correctly.
-  //
-  // Passing text == absl::string_view() will be handled like any other
-  // empty string, but note that on return, it will not be possible to tell
-  // whether submatch i matched the empty string or did not match:
-  // either way, submatch[i].data() == NULL.
-  bool Match(absl::string_view text,
-             size_t startpos,
-             size_t endpos,
-             Anchor re_anchor,
-             absl::string_view* submatch,
-             int nsubmatch) const;
-
-  // Check that the given rewrite string is suitable for use with this
-  // regular expression.  It checks that:
-  //   * The regular expression has enough parenthesized subexpressions
-  //     to satisfy all of the \N tokens in rewrite
-  //   * The rewrite string doesn't have any syntax errors.  E.g.,
-  //     '\' followed by anything other than a digit or '\'.
-  // A true return value guarantees that Replace() and Extract() won't
-  // fail because of a bad rewrite string.
-  bool CheckRewriteString(absl::string_view rewrite,
-                          std::string* error) const;
-
-  // Returns the maximum submatch needed for the rewrite to be done by
-  // Replace(). E.g. if rewrite == "foo \\2,\\1", returns 2.
-  static int MaxSubmatch(absl::string_view rewrite);
-
-  // Append the "rewrite" string, with backslash substitutions from "vec",
-  // to string "out".
-  // Returns true on success.  This method can fail because of a malformed
-  // rewrite string.  CheckRewriteString guarantees that the rewrite will
-  // be sucessful.
-  bool Rewrite(std::string* out,
-               absl::string_view rewrite,
-               const absl::string_view* vec,
-               int veclen) const;
-
-  // Constructor options
-  class Options {
-   public:
-    // The options are (defaults in parentheses):
-    //
-    //   utf8             (true)  text and pattern are UTF-8; otherwise Latin-1
-    //   posix_syntax     (false) restrict regexps to POSIX egrep syntax
-    //   longest_match    (false) search for longest match, not first match
-    //   log_errors       (true)  log syntax and execution errors to ERROR
-    //   max_mem          (see below)  approx. max memory footprint of RE2
-    //   literal          (false) interpret string as literal, not regexp
-    //   never_nl         (false) never match \n, even if it is in regexp
-    //   dot_nl           (false) dot matches everything including new line
-    //   never_capture    (false) parse all parens as non-capturing
-    //   case_sensitive   (true)  match is case-sensitive (regexp can override
-    //                              with (?i) unless in posix_syntax mode)
-    //
-    // The following options are only consulted when posix_syntax == true.
-    // When posix_syntax == false, these features are always enabled and
-    // cannot be turned off; to perform multi-line matching in that case,
-    // begin the regexp with (?m).
-    //   perl_classes     (false) allow Perl's \d \s \w \D \S \W
-    //   word_boundary    (false) allow Perl's \b \B (word boundary and not)
-    //   one_line         (false) ^ and $ only match beginning and end of text
-    //
-    // The max_mem option controls how much memory can be used
-    // to hold the compiled form of the regexp (the Prog) and
-    // its cached DFA graphs.  Code Search placed limits on the number
-    // of Prog instructions and DFA states: 10,000 for both.
-    // In RE2, those limits would translate to about 240 KB per Prog
-    // and perhaps 2.5 MB per DFA (DFA state sizes vary by regexp; RE2 does a
-    // better job of keeping them small than Code Search did).
-    // Each RE2 has two Progs (one forward, one reverse), and each Prog
-    // can have two DFAs (one first match, one longest match).
-    // That makes 4 DFAs:
-    //
-    //   forward, first-match    - used for UNANCHORED or ANCHOR_START searches
-    //                               if opt.longest_match() == false
-    //   forward, longest-match  - used for all ANCHOR_BOTH searches,
-    //                               and the other two kinds if
-    //                               opt.longest_match() == true
-    //   reverse, first-match    - never used
-    //   reverse, longest-match  - used as second phase for unanchored searches
-    //
-    // The RE2 memory budget is statically divided between the two
-    // Progs and then the DFAs: two thirds to the forward Prog
-    // and one third to the reverse Prog.  The forward Prog gives half
-    // of what it has left over to each of its DFAs.  The reverse Prog
-    // gives it all to its longest-match DFA.
-    //
-    // Once a DFA fills its budget, it flushes its cache and starts over.
-    // If this happens too often, RE2 falls back on the NFA implementation.
-
-    // For now, make the default budget something close to Code Search.
-    static const int kDefaultMaxMem = 8<<20;
-
-    enum Encoding {
-      EncodingUTF8 = 1,
-      EncodingLatin1
-    };
-
-    Options() :
-      max_mem_(kDefaultMaxMem),
-      encoding_(EncodingUTF8),
-      posix_syntax_(false),
-      longest_match_(false),
-      log_errors_(true),
-      literal_(false),
-      never_nl_(false),
-      dot_nl_(false),
-      never_capture_(false),
-      case_sensitive_(true),
-      perl_classes_(false),
-      word_boundary_(false),
-      one_line_(false) {
-    }
-
-    /*implicit*/ Options(CannedOptions);
-
-    int64_t max_mem() const { return max_mem_; }
-    void set_max_mem(int64_t m) { max_mem_ = m; }
-
-    Encoding encoding() const { return encoding_; }
-    void set_encoding(Encoding encoding) { encoding_ = encoding; }
-
-    bool posix_syntax() const { return posix_syntax_; }
-    void set_posix_syntax(bool b) { posix_syntax_ = b; }
-
-    bool longest_match() const { return longest_match_; }
-    void set_longest_match(bool b) { longest_match_ = b; }
-
-    bool log_errors() const { return log_errors_; }
-    void set_log_errors(bool b) { log_errors_ = b; }
-
-    bool literal() const { return literal_; }
-    void set_literal(bool b) { literal_ = b; }
-
-    bool never_nl() const { return never_nl_; }
-    void set_never_nl(bool b) { never_nl_ = b; }
-
-    bool dot_nl() const { return dot_nl_; }
-    void set_dot_nl(bool b) { dot_nl_ = b; }
-
-    bool never_capture() const { return never_capture_; }
-    void set_never_capture(bool b) { never_capture_ = b; }
-
-    bool case_sensitive() const { return case_sensitive_; }
-    void set_case_sensitive(bool b) { case_sensitive_ = b; }
-
-    bool perl_classes() const { return perl_classes_; }
-    void set_perl_classes(bool b) { perl_classes_ = b; }
-
-    bool word_boundary() const { return word_boundary_; }
-    void set_word_boundary(bool b) { word_boundary_ = b; }
-
-    bool one_line() const { return one_line_; }
-    void set_one_line(bool b) { one_line_ = b; }
-
-    void Copy(const Options& src) {
-      *this = src;
-    }
-
-    int ParseFlags() const;
-
-   private:
-    int64_t max_mem_;
-    Encoding encoding_;
-    bool posix_syntax_;
-    bool longest_match_;
-    bool log_errors_;
-    bool literal_;
-    bool never_nl_;
-    bool dot_nl_;
-    bool never_capture_;
-    bool case_sensitive_;
-    bool perl_classes_;
-    bool word_boundary_;
-    bool one_line_;
-  };
-
-  // Returns the options set in the constructor.
-  const Options& options() const { return options_; }
-
-  // Argument converters; see below.
-  template <typename T>
-  static Arg CRadix(T* ptr);
-  template <typename T>
-  static Arg Hex(T* ptr);
-  template <typename T>
-  static Arg Octal(T* ptr);
-
-  // Controls the maximum count permitted by GlobalReplace(); -1 is unlimited.
-  // FOR FUZZING ONLY.
-  static void FUZZING_ONLY_set_maximum_global_replace_count(int i);
-
- private:
-  void Init(absl::string_view pattern, const Options& options);
-
-  bool DoMatch(absl::string_view text,
-               Anchor re_anchor,
-               size_t* consumed,
-               const Arg* const args[],
-               int n) const;
-
-  re2::Prog* ReverseProg() const;
-
-  // First cache line is relatively cold fields.
-  const std::string* pattern_;    // string regular expression
-  Options options_;               // option flags
-  re2::Regexp* entire_regexp_;    // parsed regular expression
-  re2::Regexp* suffix_regexp_;    // parsed regular expression, prefix_ removed
-  const std::string* error_;      // error indicator (or points to empty string)
-  const std::string* error_arg_;  // fragment of regexp showing error (or ditto)
-
-  // Second cache line is relatively hot fields.
-  // These are ordered oddly to pack everything.
-  int num_captures_;              // number of capturing groups
-  ErrorCode error_code_ : 29;     // error code (29 bits is more than enough)
-  bool longest_match_ : 1;        // cached copy of options_.longest_match()
-  bool is_one_pass_ : 1;          // can use prog_->SearchOnePass?
-  bool prefix_foldcase_ : 1;      // prefix_ is ASCII case-insensitive
-  std::string prefix_;            // required prefix (before suffix_regexp_)
-  re2::Prog* prog_;               // compiled program for regexp
-
-  // Reverse Prog for DFA execution only
-  mutable re2::Prog* rprog_;
-  // Map from capture names to indices
-  mutable const std::map<std::string, int>* named_groups_;
-  // Map from capture indices to names
-  mutable const std::map<int, std::string>* group_names_;
-
-  mutable absl::once_flag rprog_once_;
-  mutable absl::once_flag named_groups_once_;
-  mutable absl::once_flag group_names_once_;
-};
-
-/***** Implementation details *****/
-
-namespace re2_internal {
-
-// Types for which the 3-ary Parse() function template has specializations.
-template <typename T> struct Parse3ary : public std::false_type {};
-template <> struct Parse3ary<void> : public std::true_type {};
-template <> struct Parse3ary<std::string> : public std::true_type {};
-template <> struct Parse3ary<absl::string_view> : public std::true_type {};
-template <> struct Parse3ary<char> : public std::true_type {};
-template <> struct Parse3ary<signed char> : public std::true_type {};
-template <> struct Parse3ary<unsigned char> : public std::true_type {};
-template <> struct Parse3ary<float> : public std::true_type {};
-template <> struct Parse3ary<double> : public std::true_type {};
-
-template <typename T>
-bool Parse(const char* str, size_t n, T* dest);
-
-// Types for which the 4-ary Parse() function template has specializations.
-template <typename T> struct Parse4ary : public std::false_type {};
-template <> struct Parse4ary<long> : public std::true_type {};
-template <> struct Parse4ary<unsigned long> : public std::true_type {};
-template <> struct Parse4ary<short> : public std::true_type {};
-template <> struct Parse4ary<unsigned short> : public std::true_type {};
-template <> struct Parse4ary<int> : public std::true_type {};
-template <> struct Parse4ary<unsigned int> : public std::true_type {};
-template <> struct Parse4ary<long long> : public std::true_type {};
-template <> struct Parse4ary<unsigned long long> : public std::true_type {};
-
-template <typename T>
-bool Parse(const char* str, size_t n, T* dest, int radix);
-
-// Support absl::optional<T> for all T with a stock parser.
-template <typename T> struct Parse3ary<absl::optional<T>> : public Parse3ary<T> {};
-template <typename T> struct Parse4ary<absl::optional<T>> : public Parse4ary<T> {};
-
-template <typename T>
-bool Parse(const char* str, size_t n, absl::optional<T>* dest) {
-  if (str == NULL) {
-    if (dest != NULL)
-      dest->reset();
-    return true;
-  }
-  T tmp;
-  if (Parse(str, n, &tmp)) {
-    if (dest != NULL)
-      dest->emplace(std::move(tmp));
-    return true;
-  }
-  return false;
-}
-
-template <typename T>
-bool Parse(const char* str, size_t n, absl::optional<T>* dest, int radix) {
-  if (str == NULL) {
-    if (dest != NULL)
-      dest->reset();
-    return true;
-  }
-  T tmp;
-  if (Parse(str, n, &tmp, radix)) {
-    if (dest != NULL)
-      dest->emplace(std::move(tmp));
-    return true;
-  }
-  return false;
-}
-
-}  // namespace re2_internal
-
-class RE2::Arg {
- private:
-  template <typename T>
-  using CanParse3ary = typename std::enable_if<
-      re2_internal::Parse3ary<T>::value,
-      int>::type;
-
-  template <typename T>
-  using CanParse4ary = typename std::enable_if<
-      re2_internal::Parse4ary<T>::value,
-      int>::type;
-
-#if !defined(_MSC_VER)
-  template <typename T>
-  using CanParseFrom = typename std::enable_if<
-      std::is_member_function_pointer<
-          decltype(static_cast<bool (T::*)(const char*, size_t)>(
-              &T::ParseFrom))>::value,
-      int>::type;
-#endif
-
- public:
-  Arg() : Arg(nullptr) {}
-  Arg(std::nullptr_t ptr) : arg_(ptr), parser_(DoNothing) {}
-
-  template <typename T, CanParse3ary<T> = 0>
-  Arg(T* ptr) : arg_(ptr), parser_(DoParse3ary<T>) {}
-
-  template <typename T, CanParse4ary<T> = 0>
-  Arg(T* ptr) : arg_(ptr), parser_(DoParse4ary<T>) {}
-
-#if !defined(_MSC_VER)
-  template <typename T, CanParseFrom<T> = 0>
-  Arg(T* ptr) : arg_(ptr), parser_(DoParseFrom<T>) {}
-#endif
-
-  typedef bool (*Parser)(const char* str, size_t n, void* dest);
-
-  template <typename T>
-  Arg(T* ptr, Parser parser) : arg_(ptr), parser_(parser) {}
-
-  bool Parse(const char* str, size_t n) const {
-    return (*parser_)(str, n, arg_);
-  }
-
- private:
-  static bool DoNothing(const char* /*str*/, size_t /*n*/, void* /*dest*/) {
-    return true;
-  }
-
-  template <typename T>
-  static bool DoParse3ary(const char* str, size_t n, void* dest) {
-    return re2_internal::Parse(str, n, reinterpret_cast<T*>(dest));
-  }
-
-  template <typename T>
-  static bool DoParse4ary(const char* str, size_t n, void* dest) {
-    return re2_internal::Parse(str, n, reinterpret_cast<T*>(dest), 10);
-  }
-
-#if !defined(_MSC_VER)
-  template <typename T>
-  static bool DoParseFrom(const char* str, size_t n, void* dest) {
-    if (dest == NULL) return true;
-    return reinterpret_cast<T*>(dest)->ParseFrom(str, n);
-  }
-#endif
-
-  void*         arg_;
-  Parser        parser_;
-};
-
-template <typename T>
-inline RE2::Arg RE2::CRadix(T* ptr) {
-  return RE2::Arg(ptr, [](const char* str, size_t n, void* dest) -> bool {
-    return re2_internal::Parse(str, n, reinterpret_cast<T*>(dest), 0);
-  });
-}
-
-template <typename T>
-inline RE2::Arg RE2::Hex(T* ptr) {
-  return RE2::Arg(ptr, [](const char* str, size_t n, void* dest) -> bool {
-    return re2_internal::Parse(str, n, reinterpret_cast<T*>(dest), 16);
-  });
-}
-
-template <typename T>
-inline RE2::Arg RE2::Octal(T* ptr) {
-  return RE2::Arg(ptr, [](const char* str, size_t n, void* dest) -> bool {
-    return re2_internal::Parse(str, n, reinterpret_cast<T*>(dest), 8);
-  });
-}
-
-// Silence warnings about missing initializers for members of LazyRE2.
-#if defined(__GNUC__)
-#pragma GCC diagnostic ignored "-Wmissing-field-initializers"
-#endif
-
-// Helper for writing global or static RE2s safely.
-// Write
-//     static LazyRE2 re = {".*"};
-// and then use *re instead of writing
-//     static RE2 re(".*");
-// The former is more careful about multithreaded
-// situations than the latter.
-//
-// N.B. This class never deletes the RE2 object that
-// it constructs: that's a feature, so that it can be used
-// for global and function static variables.
-class LazyRE2 {
- private:
-  struct NoArg {};
-
- public:
-  typedef RE2 element_type;  // support std::pointer_traits
-
-  // Constructor omitted to preserve braced initialization in C++98.
-
-  // Pretend to be a pointer to Type (never NULL due to on-demand creation):
-  RE2& operator*() const { return *get(); }
-  RE2* operator->() const { return get(); }
-
-  // Named accessor/initializer:
-  RE2* get() const {
-    absl::call_once(once_, &LazyRE2::Init, this);
-    return ptr_;
-  }
-
-  // All data fields must be public to support {"foo"} initialization.
-  const char* pattern_;
-  RE2::CannedOptions options_;
-  NoArg barrier_against_excess_initializers_;
-
-  mutable RE2* ptr_;
-  mutable absl::once_flag once_;
-
- private:
-  static void Init(const LazyRE2* lazy_re2) {
-    lazy_re2->ptr_ = new RE2(lazy_re2->pattern_, lazy_re2->options_);
-  }
-
-  void operator=(const LazyRE2&);  // disallowed
-};
-
-namespace hooks {
-
-// Most platforms support thread_local. Older versions of iOS don't support
-// thread_local, but for the sake of brevity, we lump together all versions
-// of Apple platforms that aren't macOS. If an iOS application really needs
-// the context pointee someday, we can get more specific then...
-//
-// As per https://github.com/google/re2/issues/325, thread_local support in
-// MinGW seems to be buggy. (FWIW, Abseil folks also avoid it.)
-#define RE2_HAVE_THREAD_LOCAL
-#if (defined(__APPLE__) && !(defined(TARGET_OS_OSX) && TARGET_OS_OSX)) || defined(__MINGW32__)
-#undef RE2_HAVE_THREAD_LOCAL
-#endif
-
-// A hook must not make any assumptions regarding the lifetime of the context
-// pointee beyond the current invocation of the hook. Pointers and references
-// obtained via the context pointee should be considered invalidated when the
-// hook returns. Hence, any data about the context pointee (e.g. its pattern)
-// would have to be copied in order for it to be kept for an indefinite time.
-//
-// A hook must not use RE2 for matching. Control flow reentering RE2::Match()
-// could result in infinite mutual recursion. To discourage that possibility,
-// RE2 will not maintain the context pointer correctly when used in that way.
-#ifdef RE2_HAVE_THREAD_LOCAL
-extern thread_local const RE2* context;
-#endif
-
-struct DFAStateCacheReset {
-  int64_t state_budget;
-  size_t state_cache_size;
-};
-
-struct DFASearchFailure {
-  // Nothing yet...
-};
-
-#define DECLARE_HOOK(type)                  \
-  using type##Callback = void(const type&); \
-  void Set##type##Hook(type##Callback* cb); \
-  type##Callback* Get##type##Hook();
-
-DECLARE_HOOK(DFAStateCacheReset)
-DECLARE_HOOK(DFASearchFailure)
-
-#undef DECLARE_HOOK
-
-}  // namespace hooks
-
-}  // namespace re2
-
-using re2::RE2;
-using re2::LazyRE2;
-
-#endif  // RE2_RE2_H_
diff --git a/projects/llm_framework/include/re2/re2/regexp.h b/projects/llm_framework/include/re2/re2/regexp.h
deleted file mode 100644
index df498947..00000000
--- a/projects/llm_framework/include/re2/re2/regexp.h
+++ /dev/null
@@ -1,664 +0,0 @@
-// Copyright 2006 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_REGEXP_H_
-#define RE2_REGEXP_H_
-
-// --- SPONSORED LINK --------------------------------------------------
-// If you want to use this library for regular expression matching,
-// you should use re2/re2.h, which provides a class RE2 that
-// mimics the PCRE interface provided by PCRE's C++ wrappers.
-// This header describes the low-level interface used to implement RE2
-// and may change in backwards-incompatible ways from time to time.
-// In contrast, RE2's interface will not.
-// ---------------------------------------------------------------------
-
-// Regular expression library: parsing, execution, and manipulation
-// of regular expressions.
-//
-// Any operation that traverses the Regexp structures should be written
-// using Regexp::Walker (see walker-inl.h), not recursively, because deeply nested
-// regular expressions such as x++++++++++++++++++++... might cause recursive
-// traversals to overflow the stack.
-//
-// It is the caller's responsibility to provide appropriate mutual exclusion
-// around manipulation of the regexps.  RE2 does this.
-//
-// PARSING
-//
-// Regexp::Parse parses regular expressions encoded in UTF-8.
-// The default syntax is POSIX extended regular expressions,
-// with the following changes:
-//
-//   1.  Backreferences (optional in POSIX EREs) are not supported.
-//         (Supporting them precludes the use of DFA-based
-//          matching engines.)
-//
-//   2.  Collating elements and collation classes are not supported.
-//         (No one has needed or wanted them.)
-//
-// The exact syntax accepted can be modified by passing flags to
-// Regexp::Parse.  In particular, many of the basic Perl additions
-// are available.  The flags are documented below (search for LikePerl).
-//
-// If parsed with the flag Regexp::Latin1, both the regular expression
-// and the input to the matching routines are assumed to be encoded in
-// Latin-1, not UTF-8.
-//
-// EXECUTION
-//
-// Once Regexp has parsed a regular expression, it provides methods
-// to search text using that regular expression.  These methods are
-// implemented via calling out to other regular expression libraries.
-// (Let's call them the sublibraries.)
-//
-// To call a sublibrary, Regexp does not simply prepare a
-// string version of the regular expression and hand it to the
-// sublibrary.  Instead, Regexp prepares, from its own parsed form, the
-// corresponding internal representation used by the sublibrary.
-// This has the drawback of needing to know the internal representation
-// used by the sublibrary, but it has two important benefits:
-//
-//   1. The syntax and meaning of regular expressions is guaranteed
-//      to be that used by Regexp's parser, not the syntax expected
-//      by the sublibrary.  Regexp might accept a restricted or
-//      expanded syntax for regular expressions as compared with
-//      the sublibrary.  As long as Regexp can translate from its
-//      internal form into the sublibrary's, clients need not know
-//      exactly which sublibrary they are using.
-//
-//   2. The sublibrary parsers are bypassed.  For whatever reason,
-//      sublibrary regular expression parsers often have security
-//      problems.  For example, plan9grep's regular expression parser
-//      has a buffer overflow in its handling of large character
-//      classes, and PCRE's parser has had buffer overflow problems
-//      in the past.  Security-team requires sandboxing of sublibrary
-//      regular expression parsers.  Avoiding the sublibrary parsers
-//      avoids the sandbox.
-//
-// The execution methods we use now are provided by the compiled form,
-// Prog, described in prog.h
-//
-// MANIPULATION
-//
-// Unlike other regular expression libraries, Regexp makes its parsed
-// form accessible to clients, so that client code can analyze the
-// parsed regular expressions.
-
-#include <stddef.h>
-#include <stdint.h>
-#include <map>
-#include <set>
-#include <string>
-
-#include "absl/strings/string_view.h"
-#include "util/logging.h"
-#include "util/utf.h"
-
-namespace re2 {
-
-// Keep in sync with string list kOpcodeNames[] in testing/dump.cc
-enum RegexpOp {
-  // Matches no strings.
-  kRegexpNoMatch = 1,
-
-  // Matches empty string.
-  kRegexpEmptyMatch,
-
-  // Matches rune_.
-  kRegexpLiteral,
-
-  // Matches runes_.
-  kRegexpLiteralString,
-
-  // Matches concatenation of sub_[0..nsub-1].
-  kRegexpConcat,
-  // Matches union of sub_[0..nsub-1].
-  kRegexpAlternate,
-
-  // Matches sub_[0] zero or more times.
-  kRegexpStar,
-  // Matches sub_[0] one or more times.
-  kRegexpPlus,
-  // Matches sub_[0] zero or one times.
-  kRegexpQuest,
-
-  // Matches sub_[0] at least min_ times, at most max_ times.
-  // max_ == -1 means no upper limit.
-  kRegexpRepeat,
-
-  // Parenthesized (capturing) subexpression.  Index is cap_.
-  // Optionally, capturing name is name_.
-  kRegexpCapture,
-
-  // Matches any character.
-  kRegexpAnyChar,
-
-  // Matches any byte [sic].
-  kRegexpAnyByte,
-
-  // Matches empty string at beginning of line.
-  kRegexpBeginLine,
-  // Matches empty string at end of line.
-  kRegexpEndLine,
-
-  // Matches word boundary "\b".
-  kRegexpWordBoundary,
-  // Matches not-a-word boundary "\B".
-  kRegexpNoWordBoundary,
-
-  // Matches empty string at beginning of text.
-  kRegexpBeginText,
-  // Matches empty string at end of text.
-  kRegexpEndText,
-
-  // Matches character class given by cc_.
-  kRegexpCharClass,
-
-  // Forces match of entire expression right now,
-  // with match ID match_id_ (used by RE2::Set).
-  kRegexpHaveMatch,
-
-  kMaxRegexpOp = kRegexpHaveMatch,
-};
-
-// Keep in sync with string list in regexp.cc
-enum RegexpStatusCode {
-  // No error
-  kRegexpSuccess = 0,
-
-  // Unexpected error
-  kRegexpInternalError,
-
-  // Parse errors
-  kRegexpBadEscape,          // bad escape sequence
-  kRegexpBadCharClass,       // bad character class
-  kRegexpBadCharRange,       // bad character class range
-  kRegexpMissingBracket,     // missing closing ]
-  kRegexpMissingParen,       // missing closing )
-  kRegexpUnexpectedParen,    // unexpected closing )
-  kRegexpTrailingBackslash,  // at end of regexp
-  kRegexpRepeatArgument,     // repeat argument missing, e.g. "*"
-  kRegexpRepeatSize,         // bad repetition argument
-  kRegexpRepeatOp,           // bad repetition operator
-  kRegexpBadPerlOp,          // bad perl operator
-  kRegexpBadUTF8,            // invalid UTF-8 in regexp
-  kRegexpBadNamedCapture,    // bad named capture
-};
-
-// Error status for certain operations.
-class RegexpStatus {
- public:
-  RegexpStatus() : code_(kRegexpSuccess), tmp_(NULL) {}
-  ~RegexpStatus() { delete tmp_; }
-
-  void set_code(RegexpStatusCode code) { code_ = code; }
-  void set_error_arg(absl::string_view error_arg) { error_arg_ = error_arg; }
-  void set_tmp(std::string* tmp) { delete tmp_; tmp_ = tmp; }
-  RegexpStatusCode code() const { return code_; }
-  absl::string_view error_arg() const { return error_arg_; }
-  bool ok() const { return code() == kRegexpSuccess; }
-
-  // Copies state from status.
-  void Copy(const RegexpStatus& status);
-
-  // Returns text equivalent of code, e.g.:
-  //   "Bad character class"
-  static std::string CodeText(RegexpStatusCode code);
-
-  // Returns text describing error, e.g.:
-  //   "Bad character class: [z-a]"
-  std::string Text() const;
-
- private:
-  RegexpStatusCode code_;        // Kind of error.
-  absl::string_view error_arg_;  // Piece of regexp containing syntax error.
-  std::string* tmp_;             // Temporary storage, possibly for error_arg_.
-
-  RegexpStatus(const RegexpStatus&) = delete;
-  RegexpStatus& operator=(const RegexpStatus&) = delete;
-};
-
-// Compiled form; see prog.h
-class Prog;
-
-struct RuneRange {
-  RuneRange() : lo(0), hi(0) { }
-  RuneRange(int l, int h) : lo(l), hi(h) { }
-  Rune lo;
-  Rune hi;
-};
-
-// Less-than on RuneRanges treats a == b if they overlap at all.
-// This lets us look in a set to find the range covering a particular Rune.
-struct RuneRangeLess {
-  bool operator()(const RuneRange& a, const RuneRange& b) const {
-    return a.hi < b.lo;
-  }
-};
-
-class CharClassBuilder;
-
-class CharClass {
- public:
-  void Delete();
-
-  typedef RuneRange* iterator;
-  iterator begin() { return ranges_; }
-  iterator end() { return ranges_ + nranges_; }
-
-  int size() { return nrunes_; }
-  bool empty() { return nrunes_ == 0; }
-  bool full() { return nrunes_ == Runemax+1; }
-  bool FoldsASCII() { return folds_ascii_; }
-
-  bool Contains(Rune r) const;
-  CharClass* Negate();
-
- private:
-  CharClass();  // not implemented
-  ~CharClass();  // not implemented
-  static CharClass* New(size_t maxranges);
-
-  friend class CharClassBuilder;
-
-  bool folds_ascii_;
-  int nrunes_;
-  RuneRange *ranges_;
-  int nranges_;
-
-  CharClass(const CharClass&) = delete;
-  CharClass& operator=(const CharClass&) = delete;
-};
-
-class Regexp {
- public:
-
-  // Flags for parsing.  Can be ORed together.
-  enum ParseFlags {
-    NoParseFlags  = 0,
-    FoldCase      = 1<<0,   // Fold case during matching (case-insensitive).
-    Literal       = 1<<1,   // Treat s as literal string instead of a regexp.
-    ClassNL       = 1<<2,   // Allow char classes like [^a-z] and \D and \s
-                            // and [[:space:]] to match newline.
-    DotNL         = 1<<3,   // Allow . to match newline.
-    MatchNL       = ClassNL | DotNL,
-    OneLine       = 1<<4,   // Treat ^ and $ as only matching at beginning and
-                            // end of text, not around embedded newlines.
-                            // (Perl's default)
-    Latin1        = 1<<5,   // Regexp and text are in Latin1, not UTF-8.
-    NonGreedy     = 1<<6,   // Repetition operators are non-greedy by default.
-    PerlClasses   = 1<<7,   // Allow Perl character classes like \d.
-    PerlB         = 1<<8,   // Allow Perl's \b and \B.
-    PerlX         = 1<<9,   // Perl extensions:
-                            //   non-capturing parens - (?: )
-                            //   non-greedy operators - *? +? ?? {}?
-                            //   flag edits - (?i) (?-i) (?i: )
-                            //     i - FoldCase
-                            //     m - !OneLine
-                            //     s - DotNL
-                            //     U - NonGreedy
-                            //   line ends: \A \z
-                            //   \Q and \E to disable/enable metacharacters
-                            //   (?P<name>expr) for named captures
-                            //   \C to match any single byte
-    UnicodeGroups = 1<<10,  // Allow \p{Han} for Unicode Han group
-                            //   and \P{Han} for its negation.
-    NeverNL       = 1<<11,  // Never match NL, even if the regexp mentions
-                            //   it explicitly.
-    NeverCapture  = 1<<12,  // Parse all parens as non-capturing.
-
-    // As close to Perl as we can get.
-    LikePerl      = ClassNL | OneLine | PerlClasses | PerlB | PerlX |
-                    UnicodeGroups,
-
-    // Internal use only.
-    WasDollar     = 1<<13,  // on kRegexpEndText: was $ in regexp text
-    AllParseFlags = (1<<14)-1,
-  };
-
-  // Get.  No set, Regexps are logically immutable once created.
-  RegexpOp op() { return static_cast<RegexpOp>(op_); }
-  int nsub() { return nsub_; }
-  bool simple() { return simple_ != 0; }
-  ParseFlags parse_flags() { return static_cast<ParseFlags>(parse_flags_); }
-  int Ref();  // For testing.
-
-  Regexp** sub() {
-    if(nsub_ <= 1)
-      return &subone_;
-    else
-      return submany_;
-  }
-
-  int min() { DCHECK_EQ(op_, kRegexpRepeat); return min_; }
-  int max() { DCHECK_EQ(op_, kRegexpRepeat); return max_; }
-  Rune rune() { DCHECK_EQ(op_, kRegexpLiteral); return rune_; }
-  CharClass* cc() { DCHECK_EQ(op_, kRegexpCharClass); return cc_; }
-  int cap() { DCHECK_EQ(op_, kRegexpCapture); return cap_; }
-  const std::string* name() { DCHECK_EQ(op_, kRegexpCapture); return name_; }
-  Rune* runes() { DCHECK_EQ(op_, kRegexpLiteralString); return runes_; }
-  int nrunes() { DCHECK_EQ(op_, kRegexpLiteralString); return nrunes_; }
-  int match_id() { DCHECK_EQ(op_, kRegexpHaveMatch); return match_id_; }
-
-  // Increments reference count, returns object as convenience.
-  Regexp* Incref();
-
-  // Decrements reference count and deletes this object if count reaches 0.
-  void Decref();
-
-  // Parses string s to produce regular expression, returned.
-  // Caller must release return value with re->Decref().
-  // On failure, sets *status (if status != NULL) and returns NULL.
-  static Regexp* Parse(absl::string_view s, ParseFlags flags,
-                       RegexpStatus* status);
-
-  // Returns a _new_ simplified version of the current regexp.
-  // Does not edit the current regexp.
-  // Caller must release return value with re->Decref().
-  // Simplified means that counted repetition has been rewritten
-  // into simpler terms and all Perl/POSIX features have been
-  // removed.  The result will capture exactly the same
-  // subexpressions the original did, unless formatted with ToString.
-  Regexp* Simplify();
-  friend class CoalesceWalker;
-  friend class SimplifyWalker;
-
-  // Parses the regexp src and then simplifies it and sets *dst to the
-  // string representation of the simplified form.  Returns true on success.
-  // Returns false and sets *status (if status != NULL) on parse error.
-  static bool SimplifyRegexp(absl::string_view src, ParseFlags flags,
-                             std::string* dst, RegexpStatus* status);
-
-  // Returns the number of capturing groups in the regexp.
-  int NumCaptures();
-  friend class NumCapturesWalker;
-
-  // Returns a map from names to capturing group indices,
-  // or NULL if the regexp contains no named capture groups.
-  // The caller is responsible for deleting the map.
-  std::map<std::string, int>* NamedCaptures();
-
-  // Returns a map from capturing group indices to capturing group
-  // names or NULL if the regexp contains no named capture groups. The
-  // caller is responsible for deleting the map.
-  std::map<int, std::string>* CaptureNames();
-
-  // Returns a string representation of the current regexp,
-  // using as few parentheses as possible.
-  std::string ToString();
-
-  // Convenience functions.  They consume the passed reference,
-  // so in many cases you should use, e.g., Plus(re->Incref(), flags).
-  // They do not consume allocated arrays like subs or runes.
-  static Regexp* Plus(Regexp* sub, ParseFlags flags);
-  static Regexp* Star(Regexp* sub, ParseFlags flags);
-  static Regexp* Quest(Regexp* sub, ParseFlags flags);
-  static Regexp* Concat(Regexp** subs, int nsubs, ParseFlags flags);
-  static Regexp* Alternate(Regexp** subs, int nsubs, ParseFlags flags);
-  static Regexp* Capture(Regexp* sub, ParseFlags flags, int cap);
-  static Regexp* Repeat(Regexp* sub, ParseFlags flags, int min, int max);
-  static Regexp* NewLiteral(Rune rune, ParseFlags flags);
-  static Regexp* NewCharClass(CharClass* cc, ParseFlags flags);
-  static Regexp* LiteralString(Rune* runes, int nrunes, ParseFlags flags);
-  static Regexp* HaveMatch(int match_id, ParseFlags flags);
-
-  // Like Alternate but does not factor out common prefixes.
-  static Regexp* AlternateNoFactor(Regexp** subs, int nsubs, ParseFlags flags);
-
-  // Debugging function.  Returns string format for regexp
-  // that makes structure clear.  Does NOT use regexp syntax.
-  std::string Dump();
-
-  // Helper traversal class, defined fully in walker-inl.h.
-  template<typename T> class Walker;
-
-  // Compile to Prog.  See prog.h
-  // Reverse prog expects to be run over text backward.
-  // Construction and execution of prog will
-  // stay within approximately max_mem bytes of memory.
-  // If max_mem <= 0, a reasonable default is used.
-  Prog* CompileToProg(int64_t max_mem);
-  Prog* CompileToReverseProg(int64_t max_mem);
-
-  // Whether to expect this library to find exactly the same answer as PCRE
-  // when running this regexp.  Most regexps do mimic PCRE exactly, but a few
-  // obscure cases behave differently.  Technically this is more a property
-  // of the Prog than the Regexp, but the computation is much easier to do
-  // on the Regexp.  See mimics_pcre.cc for the exact conditions.
-  bool MimicsPCRE();
-
-  // Benchmarking function.
-  void NullWalk();
-
-  // Whether every match of this regexp must be anchored and
-  // begin with a non-empty fixed string (perhaps after ASCII
-  // case-folding).  If so, returns the prefix and the sub-regexp that
-  // follows it.
-  // Callers should expect *prefix, *foldcase and *suffix to be "zeroed"
-  // regardless of the return value.
-  bool RequiredPrefix(std::string* prefix, bool* foldcase,
-                      Regexp** suffix);
-
-  // Whether every match of this regexp must be unanchored and
-  // begin with a non-empty fixed string (perhaps after ASCII
-  // case-folding).  If so, returns the prefix.
-  // Callers should expect *prefix and *foldcase to be "zeroed"
-  // regardless of the return value.
-  bool RequiredPrefixForAccel(std::string* prefix, bool* foldcase);
-
-  // Controls the maximum repeat count permitted by the parser.
-  // FOR FUZZING ONLY.
-  static void FUZZING_ONLY_set_maximum_repeat_count(int i);
-
- private:
-  // Constructor allocates vectors as appropriate for operator.
-  explicit Regexp(RegexpOp op, ParseFlags parse_flags);
-
-  // Use Decref() instead of delete to release Regexps.
-  // This is private to catch deletes at compile time.
-  ~Regexp();
-  void Destroy();
-  bool QuickDestroy();
-
-  // Helpers for Parse.  Listed here so they can edit Regexps.
-  class ParseState;
-
-  friend class ParseState;
-  friend bool ParseCharClass(absl::string_view* s, Regexp** out_re,
-                             RegexpStatus* status);
-
-  // Helper for testing [sic].
-  friend bool RegexpEqualTestingOnly(Regexp*, Regexp*);
-
-  // Computes whether Regexp is already simple.
-  bool ComputeSimple();
-
-  // Constructor that generates a Star, Plus or Quest,
-  // squashing the pair if sub is also a Star, Plus or Quest.
-  static Regexp* StarPlusOrQuest(RegexpOp op, Regexp* sub, ParseFlags flags);
-
-  // Constructor that generates a concatenation or alternation,
-  // enforcing the limit on the number of subexpressions for
-  // a particular Regexp.
-  static Regexp* ConcatOrAlternate(RegexpOp op, Regexp** subs, int nsubs,
-                                   ParseFlags flags, bool can_factor);
-
-  // Returns the leading string that re starts with.
-  // The returned Rune* points into a piece of re,
-  // so it must not be used after the caller calls re->Decref().
-  static Rune* LeadingString(Regexp* re, int* nrune, ParseFlags* flags);
-
-  // Removes the first n leading runes from the beginning of re.
-  // Edits re in place.
-  static void RemoveLeadingString(Regexp* re, int n);
-
-  // Returns the leading regexp in re's top-level concatenation.
-  // The returned Regexp* points at re or a sub-expression of re,
-  // so it must not be used after the caller calls re->Decref().
-  static Regexp* LeadingRegexp(Regexp* re);
-
-  // Removes LeadingRegexp(re) from re and returns the remainder.
-  // Might edit re in place.
-  static Regexp* RemoveLeadingRegexp(Regexp* re);
-
-  // Simplifies an alternation of literal strings by factoring out
-  // common prefixes.
-  static int FactorAlternation(Regexp** sub, int nsub, ParseFlags flags);
-  friend class FactorAlternationImpl;
-
-  // Is a == b?  Only efficient on regexps that have not been through
-  // Simplify yet - the expansion of a kRegexpRepeat will make this
-  // take a long time.  Do not call on such regexps, hence private.
-  static bool Equal(Regexp* a, Regexp* b);
-
-  // Allocate space for n sub-regexps.
-  void AllocSub(int n) {
-    DCHECK(n >= 0 && static_cast<uint16_t>(n) == n);
-    if (n > 1)
-      submany_ = new Regexp*[n];
-    nsub_ = static_cast<uint16_t>(n);
-  }
-
-  // Add Rune to LiteralString
-  void AddRuneToString(Rune r);
-
-  // Swaps this with that, in place.
-  void Swap(Regexp *that);
-
-  // Operator.  See description of operators above.
-  // uint8_t instead of RegexpOp to control space usage.
-  uint8_t op_;
-
-  // Is this regexp structure already simple
-  // (has it been returned by Simplify)?
-  // uint8_t instead of bool to control space usage.
-  uint8_t simple_;
-
-  // Flags saved from parsing and used during execution.
-  // (Only FoldCase is used.)
-  // uint16_t instead of ParseFlags to control space usage.
-  uint16_t parse_flags_;
-
-  // Reference count.  Exists so that SimplifyRegexp can build
-  // regexp structures that are dags rather than trees to avoid
-  // exponential blowup in space requirements.
-  // uint16_t to control space usage.
-  // The standard regexp routines will never generate a
-  // ref greater than the maximum repeat count (kMaxRepeat),
-  // but even so, Incref and Decref consult an overflow map
-  // when ref_ reaches kMaxRef.
-  uint16_t ref_;
-  static const uint16_t kMaxRef = 0xffff;
-
-  // Subexpressions.
-  // uint16_t to control space usage.
-  // Concat and Alternate handle larger numbers of subexpressions
-  // by building concatenation or alternation trees.
-  // Other routines should call Concat or Alternate instead of
-  // filling in sub() by hand.
-  uint16_t nsub_;
-  static const uint16_t kMaxNsub = 0xffff;
-  union {
-    Regexp** submany_;  // if nsub_ > 1
-    Regexp* subone_;  // if nsub_ == 1
-  };
-
-  // Extra space for parse and teardown stacks.
-  Regexp* down_;
-
-  // Arguments to operator.  See description of operators above.
-  union {
-    struct {  // Repeat
-      int max_;
-      int min_;
-    };
-    struct {  // Capture
-      int cap_;
-      std::string* name_;
-    };
-    struct {  // LiteralString
-      int nrunes_;
-      Rune* runes_;
-    };
-    struct {  // CharClass
-      // These two could be in separate union members,
-      // but it wouldn't save any space (there are other two-word structs)
-      // and keeping them separate avoids confusion during parsing.
-      CharClass* cc_;
-      CharClassBuilder* ccb_;
-    };
-    Rune rune_;  // Literal
-    int match_id_;  // HaveMatch
-    void *the_union_[2];  // as big as any other element, for memset
-  };
-
-  Regexp(const Regexp&) = delete;
-  Regexp& operator=(const Regexp&) = delete;
-};
-
-// Character class set: contains non-overlapping, non-abutting RuneRanges.
-typedef std::set<RuneRange, RuneRangeLess> RuneRangeSet;
-
-class CharClassBuilder {
- public:
-  CharClassBuilder();
-
-  typedef RuneRangeSet::iterator iterator;
-  iterator begin() { return ranges_.begin(); }
-  iterator end() { return ranges_.end(); }
-
-  int size() { return nrunes_; }
-  bool empty() { return nrunes_ == 0; }
-  bool full() { return nrunes_ == Runemax+1; }
-
-  bool Contains(Rune r);
-  bool FoldsASCII();
-  bool AddRange(Rune lo, Rune hi);  // returns whether class changed
-  CharClassBuilder* Copy();
-  void AddCharClass(CharClassBuilder* cc);
-  void Negate();
-  void RemoveAbove(Rune r);
-  CharClass* GetCharClass();
-  void AddRangeFlags(Rune lo, Rune hi, Regexp::ParseFlags parse_flags);
-
- private:
-  static const uint32_t AlphaMask = (1<<26) - 1;
-  uint32_t upper_;  // bitmap of A-Z
-  uint32_t lower_;  // bitmap of a-z
-  int nrunes_;
-  RuneRangeSet ranges_;
-
-  CharClassBuilder(const CharClassBuilder&) = delete;
-  CharClassBuilder& operator=(const CharClassBuilder&) = delete;
-};
-
-// Bitwise ops on ParseFlags produce ParseFlags.
-inline Regexp::ParseFlags operator|(Regexp::ParseFlags a,
-                                    Regexp::ParseFlags b) {
-  return static_cast<Regexp::ParseFlags>(
-      static_cast<int>(a) | static_cast<int>(b));
-}
-
-inline Regexp::ParseFlags operator^(Regexp::ParseFlags a,
-                                    Regexp::ParseFlags b) {
-  return static_cast<Regexp::ParseFlags>(
-      static_cast<int>(a) ^ static_cast<int>(b));
-}
-
-inline Regexp::ParseFlags operator&(Regexp::ParseFlags a,
-                                    Regexp::ParseFlags b) {
-  return static_cast<Regexp::ParseFlags>(
-      static_cast<int>(a) & static_cast<int>(b));
-}
-
-inline Regexp::ParseFlags operator~(Regexp::ParseFlags a) {
-  // Attempting to produce a value out of enum's range has undefined behaviour.
-  return static_cast<Regexp::ParseFlags>(
-      ~static_cast<int>(a) & static_cast<int>(Regexp::AllParseFlags));
-}
-
-}  // namespace re2
-
-#endif  // RE2_REGEXP_H_
diff --git a/projects/llm_framework/include/re2/re2/set.h b/projects/llm_framework/include/re2/re2/set.h
deleted file mode 100644
index 3fe419ba..00000000
--- a/projects/llm_framework/include/re2/re2/set.h
+++ /dev/null
@@ -1,86 +0,0 @@
-// Copyright 2010 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_SET_H_
-#define RE2_SET_H_
-
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "absl/strings/string_view.h"
-#include "re2/re2.h"
-
-namespace re2 {
-class Prog;
-class Regexp;
-}  // namespace re2
-
-namespace re2 {
-
-// An RE2::Set represents a collection of regexps that can
-// be searched for simultaneously.
-class RE2::Set {
- public:
-  enum ErrorKind {
-    kNoError = 0,
-    kNotCompiled,   // The set is not compiled.
-    kOutOfMemory,   // The DFA ran out of memory.
-    kInconsistent,  // The result is inconsistent. This should never happen.
-  };
-
-  struct ErrorInfo {
-    ErrorKind kind;
-  };
-
-  Set(const RE2::Options& options, RE2::Anchor anchor);
-  ~Set();
-
-  // Not copyable.
-  Set(const Set&) = delete;
-  Set& operator=(const Set&) = delete;
-  // Movable.
-  Set(Set&& other);
-  Set& operator=(Set&& other);
-
-  // Adds pattern to the set using the options passed to the constructor.
-  // Returns the index that will identify the regexp in the output of Match(),
-  // or -1 if the regexp cannot be parsed.
-  // Indices are assigned in sequential order starting from 0.
-  // Errors do not increment the index; if error is not NULL, *error will hold
-  // the error message from the parser.
-  int Add(absl::string_view pattern, std::string* error);
-
-  // Compiles the set in preparation for matching.
-  // Returns false if the compiler runs out of memory.
-  // Add() must not be called again after Compile().
-  // Compile() must be called before Match().
-  bool Compile();
-
-  // Returns true if text matches at least one of the regexps in the set.
-  // Fills v (if not NULL) with the indices of the matching regexps.
-  // Callers must not expect v to be sorted.
-  bool Match(absl::string_view text, std::vector<int>* v) const;
-
-  // As above, but populates error_info (if not NULL) when none of the regexps
-  // in the set matched. This can inform callers when DFA execution fails, for
-  // example, because they might wish to handle that case differently.
-  bool Match(absl::string_view text, std::vector<int>* v,
-             ErrorInfo* error_info) const;
-
- private:
-  typedef std::pair<std::string, re2::Regexp*> Elem;
-
-  RE2::Options options_;
-  RE2::Anchor anchor_;
-  std::vector<Elem> elem_;
-  bool compiled_;
-  int size_;
-  std::unique_ptr<re2::Prog> prog_;
-};
-
-}  // namespace re2
-
-#endif  // RE2_SET_H_
diff --git a/projects/llm_framework/include/re2/re2/sparse_array.h b/projects/llm_framework/include/re2/re2/sparse_array.h
deleted file mode 100644
index 09ffe086..00000000
--- a/projects/llm_framework/include/re2/re2/sparse_array.h
+++ /dev/null
@@ -1,392 +0,0 @@
-// Copyright 2006 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_SPARSE_ARRAY_H_
-#define RE2_SPARSE_ARRAY_H_
-
-// DESCRIPTION
-//
-// SparseArray<T>(m) is a map from integers in [0, m) to T values.
-// It requires (sizeof(T)+sizeof(int))*m memory, but it provides
-// fast iteration through the elements in the array and fast clearing
-// of the array.  The array has a concept of certain elements being
-// uninitialized (having no value).
-//
-// Insertion and deletion are constant time operations.
-//
-// Allocating the array is a constant time operation
-// when memory allocation is a constant time operation.
-//
-// Clearing the array is a constant time operation (unusual!).
-//
-// Iterating through the array is an O(n) operation, where n
-// is the number of items in the array (not O(m)).
-//
-// The array iterator visits entries in the order they were first
-// inserted into the array.  It is safe to add items to the array while
-// using an iterator: the iterator will visit indices added to the array
-// during the iteration, but will not re-visit indices whose values
-// change after visiting.  Thus SparseArray can be a convenient
-// implementation of a work queue.
-//
-// The SparseArray implementation is NOT thread-safe.  It is up to the
-// caller to make sure only one thread is accessing the array.  (Typically
-// these arrays are temporary values and used in situations where speed is
-// important.)
-//
-// The SparseArray interface does not present all the usual STL bells and
-// whistles.
-//
-// Implemented with reference to Briggs & Torczon, An Efficient
-// Representation for Sparse Sets, ACM Letters on Programming Languages
-// and Systems, Volume 2, Issue 1-4 (March-Dec.  1993), pp.  59-69.
-//
-// Briggs & Torczon popularized this technique, but it had been known
-// long before their paper.  They point out that Aho, Hopcroft, and
-// Ullman's 1974 Design and Analysis of Computer Algorithms and Bentley's
-// 1986 Programming Pearls both hint at the technique in exercises to the
-// reader (in Aho & Hopcroft, exercise 2.12; in Bentley, column 1
-// exercise 8).
-//
-// Briggs & Torczon describe a sparse set implementation.  I have
-// trivially generalized it to create a sparse array (actually the original
-// target of the AHU and Bentley exercises).
-
-// IMPLEMENTATION
-//
-// SparseArray is an array dense_ and an array sparse_ of identical size.
-// At any point, the number of elements in the sparse array is size_.
-//
-// The array dense_ contains the size_ elements in the sparse array (with
-// their indices),
-// in the order that the elements were first inserted.  This array is dense:
-// the size_ pairs are dense_[0] through dense_[size_-1].
-//
-// The array sparse_ maps from indices in [0,m) to indices in [0,size_).
-// For indices present in the array, dense_[sparse_[i]].index_ == i.
-// For indices not present in the array, sparse_ can contain any value at all,
-// perhaps outside the range [0, size_) but perhaps not.
-//
-// The lax requirement on sparse_ values makes clearing the array very easy:
-// set size_ to 0.  Lookups are slightly more complicated.
-// An index i has a value in the array if and only if:
-//   sparse_[i] is in [0, size_) AND
-//   dense_[sparse_[i]].index_ == i.
-// If both these properties hold, only then it is safe to refer to
-//   dense_[sparse_[i]].value_
-// as the value associated with index i.
-//
-// To insert a new entry, set sparse_[i] to size_,
-// initialize dense_[size_], and then increment size_.
-//
-// To make the sparse array as efficient as possible for non-primitive types,
-// elements may or may not be destroyed when they are deleted from the sparse
-// array through a call to resize(). They immediately become inaccessible, but
-// they are only guaranteed to be destroyed when the SparseArray destructor is
-// called.
-//
-// A moved-from SparseArray will be empty.
-
-// Doing this simplifies the logic below.
-#ifndef __has_feature
-#define __has_feature(x) 0
-#endif
-
-#include <assert.h>
-#include <stdint.h>
-#if __has_feature(memory_sanitizer)
-#include <sanitizer/msan_interface.h>
-#endif
-#include <algorithm>
-#include <memory>
-#include <utility>
-
-#include "re2/pod_array.h"
-
-namespace re2 {
-
-template<typename Value>
-class SparseArray {
- public:
-  SparseArray();
-  explicit SparseArray(int max_size);
-  ~SparseArray();
-
-  // IndexValue pairs: exposed in SparseArray::iterator.
-  class IndexValue;
-
-  typedef IndexValue* iterator;
-  typedef const IndexValue* const_iterator;
-
-  SparseArray(const SparseArray& src);
-  SparseArray(SparseArray&& src);
-
-  SparseArray& operator=(const SparseArray& src);
-  SparseArray& operator=(SparseArray&& src);
-
-  // Return the number of entries in the array.
-  int size() const {
-    return size_;
-  }
-
-  // Indicate whether the array is empty.
-  int empty() const {
-    return size_ == 0;
-  }
-
-  // Iterate over the array.
-  iterator begin() {
-    return dense_.data();
-  }
-  iterator end() {
-    return dense_.data() + size_;
-  }
-
-  const_iterator begin() const {
-    return dense_.data();
-  }
-  const_iterator end() const {
-    return dense_.data() + size_;
-  }
-
-  // Change the maximum size of the array.
-  // Invalidates all iterators.
-  void resize(int new_max_size);
-
-  // Return the maximum size of the array.
-  // Indices can be in the range [0, max_size).
-  int max_size() const {
-    if (dense_.data() != NULL)
-      return dense_.size();
-    else
-      return 0;
-  }
-
-  // Clear the array.
-  void clear() {
-    size_ = 0;
-  }
-
-  // Check whether index i is in the array.
-  bool has_index(int i) const;
-
-  // Comparison function for sorting.
-  // Can sort the sparse array so that future iterations
-  // will visit indices in increasing order using
-  // std::sort(arr.begin(), arr.end(), arr.less);
-  static bool less(const IndexValue& a, const IndexValue& b);
-
- public:
-  // Set the value at index i to v.
-  iterator set(int i, const Value& v) {
-    return SetInternal(true, i, v);
-  }
-
-  // Set the value at new index i to v.
-  // Fast but unsafe: only use if has_index(i) is false.
-  iterator set_new(int i, const Value& v) {
-    return SetInternal(false, i, v);
-  }
-
-  // Set the value at index i to v.
-  // Fast but unsafe: only use if has_index(i) is true.
-  iterator set_existing(int i, const Value& v) {
-    return SetExistingInternal(i, v);
-  }
-
-  // Get the value at index i.
-  // Fast but unsafe: only use if has_index(i) is true.
-  Value& get_existing(int i) {
-    assert(has_index(i));
-    return dense_[sparse_[i]].value_;
-  }
-  const Value& get_existing(int i) const {
-    assert(has_index(i));
-    return dense_[sparse_[i]].value_;
-  }
-
- private:
-  iterator SetInternal(bool allow_existing, int i, const Value& v) {
-    DebugCheckInvariants();
-    if (static_cast<uint32_t>(i) >= static_cast<uint32_t>(max_size())) {
-      assert(false && "illegal index");
-      // Semantically, end() would be better here, but we already know
-      // the user did something stupid, so begin() insulates them from
-      // dereferencing an invalid pointer.
-      return begin();
-    }
-    if (!allow_existing) {
-      assert(!has_index(i));
-      create_index(i);
-    } else {
-      if (!has_index(i))
-        create_index(i);
-    }
-    return SetExistingInternal(i, v);
-  }
-
-  iterator SetExistingInternal(int i, const Value& v) {
-    DebugCheckInvariants();
-    assert(has_index(i));
-    dense_[sparse_[i]].value_ = v;
-    DebugCheckInvariants();
-    return dense_.data() + sparse_[i];
-  }
-
-  // Add the index i to the array.
-  // Only use if has_index(i) is known to be false.
-  // Since it doesn't set the value associated with i,
-  // this function is private, only intended as a helper
-  // for other methods.
-  void create_index(int i);
-
-  // In debug mode, verify that some invariant properties of the class
-  // are being maintained. This is called at the end of the constructor
-  // and at the beginning and end of all public non-const member functions.
-  void DebugCheckInvariants() const;
-
-  // Initializes memory for elements [min, max).
-  void MaybeInitializeMemory(int min, int max) {
-#if __has_feature(memory_sanitizer)
-    __msan_unpoison(sparse_.data() + min, (max - min) * sizeof sparse_[0]);
-#elif defined(RE2_ON_VALGRIND)
-    for (int i = min; i < max; i++) {
-      sparse_[i] = 0xababababU;
-    }
-#endif
-  }
-
-  int size_ = 0;
-  PODArray<int> sparse_;
-  PODArray<IndexValue> dense_;
-};
-
-template<typename Value>
-SparseArray<Value>::SparseArray() = default;
-
-template<typename Value>
-SparseArray<Value>::SparseArray(const SparseArray& src)
-    : size_(src.size_),
-      sparse_(src.max_size()),
-      dense_(src.max_size()) {
-  std::copy_n(src.sparse_.data(), src.max_size(), sparse_.data());
-  std::copy_n(src.dense_.data(), src.max_size(), dense_.data());
-}
-
-template<typename Value>
-SparseArray<Value>::SparseArray(SparseArray&& src)
-    : size_(src.size_),
-      sparse_(std::move(src.sparse_)),
-      dense_(std::move(src.dense_)) {
-  src.size_ = 0;
-}
-
-template<typename Value>
-SparseArray<Value>& SparseArray<Value>::operator=(const SparseArray& src) {
-  // Construct these first for exception safety.
-  PODArray<int> a(src.max_size());
-  PODArray<IndexValue> b(src.max_size());
-
-  size_ = src.size_;
-  sparse_ = std::move(a);
-  dense_ = std::move(b);
-  std::copy_n(src.sparse_.data(), src.max_size(), sparse_.data());
-  std::copy_n(src.dense_.data(), src.max_size(), dense_.data());
-  return *this;
-}
-
-template<typename Value>
-SparseArray<Value>& SparseArray<Value>::operator=(SparseArray&& src) {
-  size_ = src.size_;
-  sparse_ = std::move(src.sparse_);
-  dense_ = std::move(src.dense_);
-  src.size_ = 0;
-  return *this;
-}
-
-// IndexValue pairs: exposed in SparseArray::iterator.
-template<typename Value>
-class SparseArray<Value>::IndexValue {
- public:
-  int index() const { return index_; }
-  Value& value() { return value_; }
-  const Value& value() const { return value_; }
-
- private:
-  friend class SparseArray;
-  int index_;
-  Value value_;
-};
-
-// Change the maximum size of the array.
-// Invalidates all iterators.
-template<typename Value>
-void SparseArray<Value>::resize(int new_max_size) {
-  DebugCheckInvariants();
-  if (new_max_size > max_size()) {
-    const int old_max_size = max_size();
-
-    // Construct these first for exception safety.
-    PODArray<int> a(new_max_size);
-    PODArray<IndexValue> b(new_max_size);
-
-    std::copy_n(sparse_.data(), old_max_size, a.data());
-    std::copy_n(dense_.data(), old_max_size, b.data());
-
-    sparse_ = std::move(a);
-    dense_ = std::move(b);
-
-    MaybeInitializeMemory(old_max_size, new_max_size);
-  }
-  if (size_ > new_max_size)
-    size_ = new_max_size;
-  DebugCheckInvariants();
-}
-
-// Check whether index i is in the array.
-template<typename Value>
-bool SparseArray<Value>::has_index(int i) const {
-  assert(i >= 0);
-  assert(i < max_size());
-  if (static_cast<uint32_t>(i) >= static_cast<uint32_t>(max_size())) {
-    return false;
-  }
-  // Unsigned comparison avoids checking sparse_[i] < 0.
-  return (uint32_t)sparse_[i] < (uint32_t)size_ &&
-         dense_[sparse_[i]].index_ == i;
-}
-
-template<typename Value>
-void SparseArray<Value>::create_index(int i) {
-  assert(!has_index(i));
-  assert(size_ < max_size());
-  sparse_[i] = size_;
-  dense_[size_].index_ = i;
-  size_++;
-}
-
-template<typename Value> SparseArray<Value>::SparseArray(int max_size) :
-    sparse_(max_size), dense_(max_size) {
-  MaybeInitializeMemory(size_, max_size);
-  DebugCheckInvariants();
-}
-
-template<typename Value> SparseArray<Value>::~SparseArray() {
-  DebugCheckInvariants();
-}
-
-template<typename Value> void SparseArray<Value>::DebugCheckInvariants() const {
-  assert(0 <= size_);
-  assert(size_ <= max_size());
-}
-
-// Comparison function for sorting.
-template<typename Value> bool SparseArray<Value>::less(const IndexValue& a,
-                                                       const IndexValue& b) {
-  return a.index_ < b.index_;
-}
-
-}  // namespace re2
-
-#endif  // RE2_SPARSE_ARRAY_H_
diff --git a/projects/llm_framework/include/re2/re2/sparse_set.h b/projects/llm_framework/include/re2/re2/sparse_set.h
deleted file mode 100644
index 06ed88d8..00000000
--- a/projects/llm_framework/include/re2/re2/sparse_set.h
+++ /dev/null
@@ -1,264 +0,0 @@
-// Copyright 2006 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_SPARSE_SET_H_
-#define RE2_SPARSE_SET_H_
-
-// DESCRIPTION
-//
-// SparseSet(m) is a set of integers in [0, m).
-// It requires sizeof(int)*m memory, but it provides
-// fast iteration through the elements in the set and fast clearing
-// of the set.
-//
-// Insertion and deletion are constant time operations.
-//
-// Allocating the set is a constant time operation
-// when memory allocation is a constant time operation.
-//
-// Clearing the set is a constant time operation (unusual!).
-//
-// Iterating through the set is an O(n) operation, where n
-// is the number of items in the set (not O(m)).
-//
-// The set iterator visits entries in the order they were first
-// inserted into the set.  It is safe to add items to the set while
-// using an iterator: the iterator will visit indices added to the set
-// during the iteration, but will not re-visit indices whose values
-// change after visiting.  Thus SparseSet can be a convenient
-// implementation of a work queue.
-//
-// The SparseSet implementation is NOT thread-safe.  It is up to the
-// caller to make sure only one thread is accessing the set.  (Typically
-// these sets are temporary values and used in situations where speed is
-// important.)
-//
-// The SparseSet interface does not present all the usual STL bells and
-// whistles.
-//
-// Implemented with reference to Briggs & Torczon, An Efficient
-// Representation for Sparse Sets, ACM Letters on Programming Languages
-// and Systems, Volume 2, Issue 1-4 (March-Dec.  1993), pp.  59-69.
-//
-// This is a specialization of sparse array; see sparse_array.h.
-
-// IMPLEMENTATION
-//
-// See sparse_array.h for implementation details.
-
-// Doing this simplifies the logic below.
-#ifndef __has_feature
-#define __has_feature(x) 0
-#endif
-
-#include <assert.h>
-#include <stdint.h>
-#if __has_feature(memory_sanitizer)
-#include <sanitizer/msan_interface.h>
-#endif
-#include <algorithm>
-#include <memory>
-#include <utility>
-
-#include "re2/pod_array.h"
-
-namespace re2 {
-
-template<typename Value>
-class SparseSetT {
- public:
-  SparseSetT();
-  explicit SparseSetT(int max_size);
-  ~SparseSetT();
-
-  typedef int* iterator;
-  typedef const int* const_iterator;
-
-  // Return the number of entries in the set.
-  int size() const {
-    return size_;
-  }
-
-  // Indicate whether the set is empty.
-  int empty() const {
-    return size_ == 0;
-  }
-
-  // Iterate over the set.
-  iterator begin() {
-    return dense_.data();
-  }
-  iterator end() {
-    return dense_.data() + size_;
-  }
-
-  const_iterator begin() const {
-    return dense_.data();
-  }
-  const_iterator end() const {
-    return dense_.data() + size_;
-  }
-
-  // Change the maximum size of the set.
-  // Invalidates all iterators.
-  void resize(int new_max_size);
-
-  // Return the maximum size of the set.
-  // Indices can be in the range [0, max_size).
-  int max_size() const {
-    if (dense_.data() != NULL)
-      return dense_.size();
-    else
-      return 0;
-  }
-
-  // Clear the set.
-  void clear() {
-    size_ = 0;
-  }
-
-  // Check whether index i is in the set.
-  bool contains(int i) const;
-
-  // Comparison function for sorting.
-  // Can sort the sparse set so that future iterations
-  // will visit indices in increasing order using
-  // std::sort(arr.begin(), arr.end(), arr.less);
-  static bool less(int a, int b);
-
- public:
-  // Insert index i into the set.
-  iterator insert(int i) {
-    return InsertInternal(true, i);
-  }
-
-  // Insert index i into the set.
-  // Fast but unsafe: only use if contains(i) is false.
-  iterator insert_new(int i) {
-    return InsertInternal(false, i);
-  }
-
- private:
-  iterator InsertInternal(bool allow_existing, int i) {
-    DebugCheckInvariants();
-    if (static_cast<uint32_t>(i) >= static_cast<uint32_t>(max_size())) {
-      assert(false && "illegal index");
-      // Semantically, end() would be better here, but we already know
-      // the user did something stupid, so begin() insulates them from
-      // dereferencing an invalid pointer.
-      return begin();
-    }
-    if (!allow_existing) {
-      assert(!contains(i));
-      create_index(i);
-    } else {
-      if (!contains(i))
-        create_index(i);
-    }
-    DebugCheckInvariants();
-    return dense_.data() + sparse_[i];
-  }
-
-  // Add the index i to the set.
-  // Only use if contains(i) is known to be false.
-  // This function is private, only intended as a helper
-  // for other methods.
-  void create_index(int i);
-
-  // In debug mode, verify that some invariant properties of the class
-  // are being maintained. This is called at the end of the constructor
-  // and at the beginning and end of all public non-const member functions.
-  void DebugCheckInvariants() const;
-
-  // Initializes memory for elements [min, max).
-  void MaybeInitializeMemory(int min, int max) {
-#if __has_feature(memory_sanitizer)
-    __msan_unpoison(sparse_.data() + min, (max - min) * sizeof sparse_[0]);
-#elif defined(RE2_ON_VALGRIND)
-    for (int i = min; i < max; i++) {
-      sparse_[i] = 0xababababU;
-    }
-#endif
-  }
-
-  int size_ = 0;
-  PODArray<int> sparse_;
-  PODArray<int> dense_;
-};
-
-template<typename Value>
-SparseSetT<Value>::SparseSetT() = default;
-
-// Change the maximum size of the set.
-// Invalidates all iterators.
-template<typename Value>
-void SparseSetT<Value>::resize(int new_max_size) {
-  DebugCheckInvariants();
-  if (new_max_size > max_size()) {
-    const int old_max_size = max_size();
-
-    // Construct these first for exception safety.
-    PODArray<int> a(new_max_size);
-    PODArray<int> b(new_max_size);
-
-    std::copy_n(sparse_.data(), old_max_size, a.data());
-    std::copy_n(dense_.data(), old_max_size, b.data());
-
-    sparse_ = std::move(a);
-    dense_ = std::move(b);
-
-    MaybeInitializeMemory(old_max_size, new_max_size);
-  }
-  if (size_ > new_max_size)
-    size_ = new_max_size;
-  DebugCheckInvariants();
-}
-
-// Check whether index i is in the set.
-template<typename Value>
-bool SparseSetT<Value>::contains(int i) const {
-  assert(i >= 0);
-  assert(i < max_size());
-  if (static_cast<uint32_t>(i) >= static_cast<uint32_t>(max_size())) {
-    return false;
-  }
-  // Unsigned comparison avoids checking sparse_[i] < 0.
-  return (uint32_t)sparse_[i] < (uint32_t)size_ &&
-         dense_[sparse_[i]] == i;
-}
-
-template<typename Value>
-void SparseSetT<Value>::create_index(int i) {
-  assert(!contains(i));
-  assert(size_ < max_size());
-  sparse_[i] = size_;
-  dense_[size_] = i;
-  size_++;
-}
-
-template<typename Value> SparseSetT<Value>::SparseSetT(int max_size) :
-    sparse_(max_size), dense_(max_size) {
-  MaybeInitializeMemory(size_, max_size);
-  DebugCheckInvariants();
-}
-
-template<typename Value> SparseSetT<Value>::~SparseSetT() {
-  DebugCheckInvariants();
-}
-
-template<typename Value> void SparseSetT<Value>::DebugCheckInvariants() const {
-  assert(0 <= size_);
-  assert(size_ <= max_size());
-}
-
-// Comparison function for sorting.
-template<typename Value> bool SparseSetT<Value>::less(int a, int b) {
-  return a < b;
-}
-
-typedef SparseSetT<void> SparseSet;
-
-}  // namespace re2
-
-#endif  // RE2_SPARSE_SET_H_
diff --git a/projects/llm_framework/include/re2/re2/stringpiece.h b/projects/llm_framework/include/re2/re2/stringpiece.h
deleted file mode 100644
index e9367bff..00000000
--- a/projects/llm_framework/include/re2/re2/stringpiece.h
+++ /dev/null
@@ -1,18 +0,0 @@
-// Copyright 2022 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_STRINGPIECE_H_
-#define RE2_STRINGPIECE_H_
-
-#include "absl/strings/string_view.h"
-
-namespace re2 {
-
-// Until RE2 requires C++17 and uses std::string_view, allow users to
-// continue to #include "re2/stringpiece.h" and use re2::StringPiece.
-using StringPiece = absl::string_view;
-
-}  // namespace re2
-
-#endif  // RE2_STRINGPIECE_H_
diff --git a/projects/llm_framework/include/re2/re2/testing/exhaustive_tester.h b/projects/llm_framework/include/re2/re2/testing/exhaustive_tester.h
deleted file mode 100644
index 906be0c8..00000000
--- a/projects/llm_framework/include/re2/re2/testing/exhaustive_tester.h
+++ /dev/null
@@ -1,104 +0,0 @@
-// Copyright 2009 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_TESTING_EXHAUSTIVE_TESTER_H_
-#define RE2_TESTING_EXHAUSTIVE_TESTER_H_
-
-#include <stdint.h>
-#include <string>
-#include <vector>
-
-#include "re2/testing/regexp_generator.h"
-#include "re2/testing/string_generator.h"
-
-namespace re2 {
-
-// Doing this simplifies the logic below.
-#ifndef __has_feature
-#define __has_feature(x) 0
-#endif
-
-#if !defined(NDEBUG)
-// We are in a debug build.
-const bool RE2_DEBUG_MODE = true;
-#elif __has_feature(address_sanitizer) || __has_feature(memory_sanitizer) || __has_feature(thread_sanitizer)
-// Not a debug build, but still under sanitizers.
-const bool RE2_DEBUG_MODE = true;
-#else
-const bool RE2_DEBUG_MODE = false;
-#endif
-
-// Exhaustive regular expression test: generate all regexps within parameters,
-// then generate all strings of a given length over a given alphabet,
-// then check that NFA, DFA, and PCRE agree about whether each regexp matches
-// each possible string, and if so, where the match is.
-//
-// Can also be used in a "random" mode that generates a given number
-// of random regexp and strings, allowing testing of larger expressions
-// and inputs.
-class ExhaustiveTester : public RegexpGenerator {
- public:
-  ExhaustiveTester(int maxatoms,
-                   int maxops,
-                   const std::vector<std::string>& alphabet,
-                   const std::vector<std::string>& ops,
-                   int maxstrlen,
-                   const std::vector<std::string>& stralphabet,
-                   const std::string& wrapper,
-                   const std::string& topwrapper)
-    : RegexpGenerator(maxatoms, maxops, alphabet, ops),
-      strgen_(maxstrlen, stralphabet),
-      wrapper_(wrapper),
-      topwrapper_(topwrapper),
-      regexps_(0), tests_(0), failures_(0),
-      randomstrings_(0), stringseed_(0), stringcount_(0)  { }
-
-  int regexps()  { return regexps_; }
-  int tests()    { return tests_; }
-  int failures() { return failures_; }
-
-  // Needed for RegexpGenerator interface.
-  void HandleRegexp(const std::string& regexp);
-
-  // Causes testing to generate random input strings.
-  void RandomStrings(int32_t seed, int32_t count) {
-    randomstrings_ = true;
-    stringseed_ = seed;
-    stringcount_ = count;
-  }
-
- private:
-  StringGenerator strgen_;
-  std::string wrapper_;      // Regexp wrapper - either empty or has one %s.
-  std::string topwrapper_;   // Regexp top-level wrapper.
-  int regexps_;   // Number of HandleRegexp calls
-  int tests_;     // Number of regexp tests.
-  int failures_;  // Number of tests failed.
-
-  bool randomstrings_;  // Whether to use random strings
-  int32_t stringseed_;  // If so, the seed.
-  int stringcount_;     // If so, how many to generate.
-
-  ExhaustiveTester(const ExhaustiveTester&) = delete;
-  ExhaustiveTester& operator=(const ExhaustiveTester&) = delete;
-};
-
-// Runs an exhaustive test on the given parameters.
-void ExhaustiveTest(int maxatoms, int maxops,
-                    const std::vector<std::string>& alphabet,
-                    const std::vector<std::string>& ops,
-                    int maxstrlen,
-                    const std::vector<std::string>& stralphabet,
-                    const std::string& wrapper,
-                    const std::string& topwrapper);
-
-// Runs an exhaustive test using the given parameters and
-// the basic egrep operators.
-void EgrepTest(int maxatoms, int maxops, const std::string& alphabet,
-               int maxstrlen, const std::string& stralphabet,
-               const std::string& wrapper);
-
-}  // namespace re2
-
-#endif  // RE2_TESTING_EXHAUSTIVE_TESTER_H_
diff --git a/projects/llm_framework/include/re2/re2/testing/regexp_generator.h b/projects/llm_framework/include/re2/re2/testing/regexp_generator.h
deleted file mode 100644
index e1be1a93..00000000
--- a/projects/llm_framework/include/re2/re2/testing/regexp_generator.h
+++ /dev/null
@@ -1,76 +0,0 @@
-// Copyright 2008 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_TESTING_REGEXP_GENERATOR_H_
-#define RE2_TESTING_REGEXP_GENERATOR_H_
-
-// Regular expression generator: generates all possible
-// regular expressions within given parameters (see below for details).
-
-#include <stdint.h>
-#include <random>
-#include <string>
-#include <vector>
-
-#include "absl/strings/string_view.h"
-
-namespace re2 {
-
-// Regular expression generator.
-//
-// Given a set of atom expressions like "a", "b", or "."
-// and operators like "%s*", generates all possible regular expressions
-// using at most maxbases base expressions and maxops operators.
-// For each such expression re, calls HandleRegexp(re).
-//
-// Callers are expected to subclass RegexpGenerator and provide HandleRegexp.
-//
-class RegexpGenerator {
- public:
-  RegexpGenerator(int maxatoms, int maxops,
-                  const std::vector<std::string>& atoms,
-                  const std::vector<std::string>& ops);
-  virtual ~RegexpGenerator() {}
-
-  // Generates all the regular expressions, calling HandleRegexp(re) for each.
-  void Generate();
-
-  // Generates n random regular expressions, calling HandleRegexp(re) for each.
-  void GenerateRandom(int32_t seed, int n);
-
-  // Handles a regular expression.  Must be provided by subclass.
-  virtual void HandleRegexp(const std::string& regexp) = 0;
-
-  // The egrep regexp operators: * + ? | and concatenation.
-  static const std::vector<std::string>& EgrepOps();
-
- private:
-  void RunPostfix(const std::vector<std::string>& post);
-  void GeneratePostfix(std::vector<std::string>* post,
-                       int nstk, int ops, int lits);
-  bool GenerateRandomPostfix(std::vector<std::string>* post,
-                             int nstk, int ops, int lits);
-
-  int maxatoms_;                    // Maximum number of atoms allowed in expr.
-  int maxops_;                      // Maximum number of ops allowed in expr.
-  std::vector<std::string> atoms_;  // Possible atoms.
-  std::vector<std::string> ops_;    // Possible ops.
-  std::minstd_rand0 rng_;           // Random number generator.
-
-  RegexpGenerator(const RegexpGenerator&) = delete;
-  RegexpGenerator& operator=(const RegexpGenerator&) = delete;
-};
-
-// Helpers for preparing arguments to RegexpGenerator constructor.
-
-// Returns one string for each character in s.
-std::vector<std::string> Explode(absl::string_view s);
-
-// Splits string everywhere sep is found, returning
-// vector of pieces.
-std::vector<std::string> Split(absl::string_view sep, absl::string_view s);
-
-}  // namespace re2
-
-#endif  // RE2_TESTING_REGEXP_GENERATOR_H_
diff --git a/projects/llm_framework/include/re2/re2/testing/string_generator.h b/projects/llm_framework/include/re2/re2/testing/string_generator.h
deleted file mode 100644
index 0d6f5fcb..00000000
--- a/projects/llm_framework/include/re2/re2/testing/string_generator.h
+++ /dev/null
@@ -1,75 +0,0 @@
-// Copyright 2008 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_TESTING_STRING_GENERATOR_H_
-#define RE2_TESTING_STRING_GENERATOR_H_
-
-// String generator: generates all possible strings of up to
-// maxlen letters using the set of letters in alpha.
-// Fetch strings using a Java-like Next()/HasNext() interface.
-
-#include <stdint.h>
-#include <random>
-#include <string>
-#include <vector>
-
-#include "absl/strings/string_view.h"
-
-namespace re2 {
-
-class StringGenerator {
- public:
-  StringGenerator(int maxlen, const std::vector<std::string>& alphabet);
-  ~StringGenerator() {}
-
-  absl::string_view Next();
-  bool HasNext() { return hasnext_; }
-
-  // Resets generator to start sequence over.
-  void Reset();
-
-  // Causes generator to emit random strings for next n calls to Next().
-  void Random(int32_t seed, int n);
-
-  // Causes generator to emit a NULL as the next call.
-  void GenerateNULL();
-
- private:
-  bool IncrementDigits();
-  bool RandomDigits();
-
-  // Global state.
-  int maxlen_;                         // Maximum length string to generate.
-  std::vector<std::string> alphabet_;  // Alphabet, one string per letter.
-
-  // Iteration state.
-  absl::string_view sp_;     // Last string_view returned by Next().
-  std::string s_;            // String data in last string_view returned by Next().
-  bool hasnext_;             // Whether Next() can be called again.
-  std::vector<int> digits_;  // Alphabet indices for next string.
-  bool generate_null_;       // Whether to generate a NULL string_view next.
-  bool random_;              // Whether generated strings are random.
-  int nrandom_;              // Number of random strings left to generate.
-  std::minstd_rand0 rng_;    // Random number generator.
-
-  StringGenerator(const StringGenerator&) = delete;
-  StringGenerator& operator=(const StringGenerator&) = delete;
-};
-
-// Generates and returns a string over binary alphabet {0,1} that contains
-// all possible binary sequences of length n as subsequences.  The obvious
-// brute force method would generate a string of length n * 2^n, but this
-// generates a string of length n-1 + 2^n called a De Bruijn cycle.
-// See Knuth, The Art of Computer Programming, Vol 2, Exercise 3.2.2 #17.
-//
-// Such a string is useful for testing a DFA.  If you have a DFA
-// where distinct last n bytes implies distinct states, then running on a
-// DeBruijn string causes the DFA to need to create a new state at every
-// position in the input, never reusing any states until it gets to the
-// end of the string.  This is the worst possible case for DFA execution.
-std::string DeBruijnString(int n);
-
-}  // namespace re2
-
-#endif  // RE2_TESTING_STRING_GENERATOR_H_
diff --git a/projects/llm_framework/include/re2/re2/testing/tester.h b/projects/llm_framework/include/re2/re2/testing/tester.h
deleted file mode 100644
index 59be5ea0..00000000
--- a/projects/llm_framework/include/re2/re2/testing/tester.h
+++ /dev/null
@@ -1,121 +0,0 @@
-// Copyright 2008 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_TESTING_TESTER_H_
-#define RE2_TESTING_TESTER_H_
-
-// Comparative tester for regular expression matching.
-// Checks all implementations against each other.
-
-#include <vector>
-
-#include "absl/strings/string_view.h"
-#include "re2/prog.h"
-#include "re2/regexp.h"
-#include "re2/re2.h"
-#include "util/pcre.h"
-
-namespace re2 {
-
-// All the supported regexp engines.
-enum Engine {
-  kEngineBacktrack = 0,    // Prog::UnsafeSearchBacktrack
-  kEngineNFA,              // Prog::SearchNFA
-  kEngineDFA,              // Prog::SearchDFA, only ask whether it matched
-  kEngineDFA1,             // Prog::SearchDFA, ask for match[0]
-  kEngineOnePass,          // Prog::SearchOnePass, if applicable
-  kEngineBitState,         // Prog::SearchBitState
-  kEngineRE2,              // RE2, all submatches
-  kEngineRE2a,             // RE2, only ask for match[0]
-  kEngineRE2b,             // RE2, only ask whether it matched
-  kEnginePCRE,             // PCRE (util/pcre.h)
-
-  kEngineMax,
-};
-
-// Make normal math on the enum preserve the type.
-// By default, C++ doesn't define ++ on enum, and e+1 has type int.
-static inline void operator++(Engine& e, int unused) {
-  e = static_cast<Engine>(e+1);
-}
-
-static inline Engine operator+(Engine e, int i) {
-  return static_cast<Engine>(static_cast<int>(e)+i);
-}
-
-// A TestInstance caches per-regexp state for a given
-// regular expression in a given configuration
-// (UTF-8 vs Latin1, longest vs first match, etc.).
-class TestInstance {
- public:
-  struct Result;
-
-  TestInstance(absl::string_view regexp, Prog::MatchKind kind,
-               Regexp::ParseFlags flags);
-  ~TestInstance();
-  Regexp::ParseFlags flags() { return flags_; }
-  bool error() { return error_; }
-
-  // Runs a single test case: search in text, which is in context,
-  // using the given anchoring.
-  bool RunCase(absl::string_view text, absl::string_view context,
-               Prog::Anchor anchor);
-
- private:
-  // Runs a single search using the named engine type.
-  void RunSearch(Engine type, absl::string_view text, absl::string_view context,
-                 Prog::Anchor anchor, Result* result);
-
-  void LogMatch(const char* prefix, Engine e, absl::string_view text,
-                absl::string_view context, Prog::Anchor anchor);
-
-  absl::string_view regexp_str_;    // regexp being tested
-  Prog::MatchKind kind_;            // kind of match
-  Regexp::ParseFlags flags_;        // flags for parsing regexp_str_
-  bool error_;                      // error during constructor?
-
-  Regexp* regexp_;                  // parsed regexp
-  int num_captures_;                // regexp_->NumCaptures() cached
-  Prog* prog_;                      // compiled program
-  Prog* rprog_;                     // compiled reverse program
-  PCRE* re_;                        // PCRE implementation
-  RE2* re2_;                        // RE2 implementation
-
-  TestInstance(const TestInstance&) = delete;
-  TestInstance& operator=(const TestInstance&) = delete;
-};
-
-// A group of TestInstances for all possible configurations.
-class Tester {
- public:
-  explicit Tester(absl::string_view regexp);
-  ~Tester();
-
-  bool error() { return error_; }
-
-  // Runs a single test case: search in text, which is in context,
-  // using the given anchoring.
-  bool TestCase(absl::string_view text, absl::string_view context,
-                Prog::Anchor anchor);
-
-  // Run TestCase(text, text, anchor) for all anchoring modes.
-  bool TestInput(absl::string_view text);
-
-  // Run TestCase(text, context, anchor) for all anchoring modes.
-  bool TestInputInContext(absl::string_view text, absl::string_view context);
-
- private:
-  bool error_;
-  std::vector<TestInstance*> v_;
-
-  Tester(const Tester&) = delete;
-  Tester& operator=(const Tester&) = delete;
-};
-
-// Run all possible tests using regexp and text.
-bool TestRegexpOnText(absl::string_view regexp, absl::string_view text);
-
-}  // namespace re2
-
-#endif  // RE2_TESTING_TESTER_H_
diff --git a/projects/llm_framework/include/re2/re2/unicode_casefold.h b/projects/llm_framework/include/re2/re2/unicode_casefold.h
deleted file mode 100644
index 4acad680..00000000
--- a/projects/llm_framework/include/re2/re2/unicode_casefold.h
+++ /dev/null
@@ -1,77 +0,0 @@
-// Copyright 2008 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_UNICODE_CASEFOLD_H_
-#define RE2_UNICODE_CASEFOLD_H_
-
-// Unicode case folding tables.
-
-// The Unicode case folding tables encode the mapping from one Unicode point
-// to the next largest Unicode point with equivalent folding.  The largest
-// point wraps back to the first.  For example, the tables map:
-//
-//     'A' -> 'a'
-//     'a' -> 'A'
-//
-//     'K' -> 'k'
-//     'k' -> 'K'  (Kelvin symbol)
-//     'K' -> 'K'
-//
-// Like everything Unicode, these tables are big.  If we represent the table
-// as a sorted list of uint32_t pairs, it has 2049 entries and is 16 kB.
-// Most table entries look like the ones around them:
-// 'A' maps to 'A'+32, 'B' maps to 'B'+32, etc.
-// Instead of listing all the pairs explicitly, we make a list of ranges
-// and deltas, so that the table entries for 'A' through 'Z' can be represented
-// as a single entry { 'A', 'Z', +32 }.
-//
-// In addition to blocks that map to each other (A-Z mapping to a-z)
-// there are blocks of pairs that individually map to each other
-// (for example, 0100<->0101, 0102<->0103, 0104<->0105, ...).
-// For those, the special delta value EvenOdd marks even/odd pairs
-// (if even, add 1; if odd, subtract 1), and OddEven marks odd/even pairs.
-//
-// In this form, the table has 274 entries, about 3kB.  If we were to split
-// the table into one for 16-bit codes and an overflow table for larger ones,
-// we could get it down to about 1.5kB, but that's not worth the complexity.
-//
-// The grouped form also allows for efficient fold range calculations
-// rather than looping one character at a time.
-
-#include <stdint.h>
-
-#include "util/utf.h"
-
-namespace re2 {
-
-enum {
-  EvenOdd = 1,
-  OddEven = -1,
-  EvenOddSkip = 1<<30,
-  OddEvenSkip,
-};
-
-struct CaseFold {
-  Rune lo;
-  Rune hi;
-  int32_t delta;
-};
-
-extern const CaseFold unicode_casefold[];
-extern const int num_unicode_casefold;
-
-extern const CaseFold unicode_tolower[];
-extern const int num_unicode_tolower;
-
-// Returns the CaseFold* in the tables that contains rune.
-// If rune is not in the tables, returns the first CaseFold* after rune.
-// If rune is larger than any value in the tables, returns NULL.
-extern const CaseFold* LookupCaseFold(const CaseFold*, int, Rune rune);
-
-// Returns the result of applying the fold f to the rune r.
-extern Rune ApplyFold(const CaseFold *f, Rune r);
-
-}  // namespace re2
-
-#endif  // RE2_UNICODE_CASEFOLD_H_
diff --git a/projects/llm_framework/include/re2/re2/unicode_groups.h b/projects/llm_framework/include/re2/re2/unicode_groups.h
deleted file mode 100644
index 6dc65321..00000000
--- a/projects/llm_framework/include/re2/re2/unicode_groups.h
+++ /dev/null
@@ -1,66 +0,0 @@
-// Copyright 2008 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_UNICODE_GROUPS_H_
-#define RE2_UNICODE_GROUPS_H_
-
-// Unicode character groups.
-
-// The codes get split into ranges of 16-bit codes
-// and ranges of 32-bit codes.  It would be simpler
-// to use only 32-bit ranges, but these tables are large
-// enough to warrant extra care.
-//
-// Using just 32-bit ranges gives 27 kB of data.
-// Adding 16-bit ranges gives 18 kB of data.
-// Adding an extra table of 16-bit singletons would reduce
-// to 16.5 kB of data but make the data harder to use;
-// we don't bother.
-
-#include <stdint.h>
-
-#include "util/utf.h"
-
-namespace re2 {
-
-struct URange16
-{
-  uint16_t lo;
-  uint16_t hi;
-};
-
-struct URange32
-{
-  Rune lo;
-  Rune hi;
-};
-
-struct UGroup
-{
-  const char *name;
-  int sign;  // +1 for [abc], -1 for [^abc]
-  const URange16 *r16;
-  int nr16;
-  const URange32 *r32;
-  int nr32;
-};
-
-// Named by property or script name (e.g., "Nd", "N", "Han").
-// Negated groups are not included.
-extern const UGroup unicode_groups[];
-extern const int num_unicode_groups;
-
-// Named by POSIX name (e.g., "[:alpha:]", "[:^lower:]").
-// Negated groups are included.
-extern const UGroup posix_groups[];
-extern const int num_posix_groups;
-
-// Named by Perl name (e.g., "\\d", "\\D").
-// Negated groups are included.
-extern const UGroup perl_groups[];
-extern const int num_perl_groups;
-
-}  // namespace re2
-
-#endif  // RE2_UNICODE_GROUPS_H_
diff --git a/projects/llm_framework/include/re2/re2/walker-inl.h b/projects/llm_framework/include/re2/re2/walker-inl.h
deleted file mode 100644
index 45763a7b..00000000
--- a/projects/llm_framework/include/re2/re2/walker-inl.h
+++ /dev/null
@@ -1,248 +0,0 @@
-// Copyright 2006 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef RE2_WALKER_INL_H_
-#define RE2_WALKER_INL_H_
-
-// Helper class for traversing Regexps without recursion.
-// Clients should declare their own subclasses that override
-// the PreVisit and PostVisit methods, which are called before
-// and after visiting the subexpressions.
-
-// Not quite the Visitor pattern, because (among other things)
-// the Visitor pattern is recursive.
-
-#include <stack>
-
-#include "absl/base/macros.h"
-#include "util/logging.h"
-#include "re2/regexp.h"
-
-namespace re2 {
-
-template<typename T> struct WalkState;
-
-template<typename T> class Regexp::Walker {
- public:
-  Walker();
-  virtual ~Walker();
-
-  // Virtual method called before visiting re's children.
-  // PreVisit passes ownership of its return value to its caller.
-  // The Arg* that PreVisit returns will be passed to PostVisit as pre_arg
-  // and passed to the child PreVisits and PostVisits as parent_arg.
-  // At the top-most Regexp, parent_arg is arg passed to walk.
-  // If PreVisit sets *stop to true, the walk does not recurse
-  // into the children.  Instead it behaves as though the return
-  // value from PreVisit is the return value from PostVisit.
-  // The default PreVisit returns parent_arg.
-  virtual T PreVisit(Regexp* re, T parent_arg, bool* stop);
-
-  // Virtual method called after visiting re's children.
-  // The pre_arg is the T that PreVisit returned.
-  // The child_args is a vector of the T that the child PostVisits returned.
-  // PostVisit takes ownership of pre_arg.
-  // PostVisit takes ownership of the Ts
-  // in *child_args, but not the vector itself.
-  // PostVisit passes ownership of its return value
-  // to its caller.
-  // The default PostVisit simply returns pre_arg.
-  virtual T PostVisit(Regexp* re, T parent_arg, T pre_arg,
-                      T* child_args, int nchild_args);
-
-  // Virtual method called to copy a T,
-  // when Walk notices that more than one child is the same re.
-  virtual T Copy(T arg);
-
-  // Virtual method called to do a "quick visit" of the re,
-  // but not its children.  Only called once the visit budget
-  // has been used up and we're trying to abort the walk
-  // as quickly as possible.  Should return a value that
-  // makes sense for the parent PostVisits still to be run.
-  // This function is (hopefully) only called by
-  // WalkExponential, but must be implemented by all clients,
-  // just in case.
-  virtual T ShortVisit(Regexp* re, T parent_arg) = 0;
-
-  // Walks over a regular expression.
-  // Top_arg is passed as parent_arg to PreVisit and PostVisit of re.
-  // Returns the T returned by PostVisit on re.
-  T Walk(Regexp* re, T top_arg);
-
-  // Like Walk, but doesn't use Copy.  This can lead to
-  // exponential runtimes on cross-linked Regexps like the
-  // ones generated by Simplify.  To help limit this,
-  // at most max_visits nodes will be visited and then
-  // the walk will be cut off early.
-  // If the walk *is* cut off early, ShortVisit(re)
-  // will be called on regexps that cannot be fully
-  // visited rather than calling PreVisit/PostVisit.
-  T WalkExponential(Regexp* re, T top_arg, int max_visits);
-
-  // Clears the stack.  Should never be necessary, since
-  // Walk always enters and exits with an empty stack.
-  // Logs DFATAL if stack is not already clear.
-  void Reset();
-
-  // Returns whether walk was cut off.
-  bool stopped_early() { return stopped_early_; }
-
- private:
-  // Walk state for the entire traversal.
-  std::stack<WalkState<T>> stack_;
-  bool stopped_early_;
-  int max_visits_;
-
-  T WalkInternal(Regexp* re, T top_arg, bool use_copy);
-
-  Walker(const Walker&) = delete;
-  Walker& operator=(const Walker&) = delete;
-};
-
-template<typename T> T Regexp::Walker<T>::PreVisit(Regexp* re,
-                                                   T parent_arg,
-                                                   bool* stop) {
-  return parent_arg;
-}
-
-template<typename T> T Regexp::Walker<T>::PostVisit(Regexp* re,
-                                                    T parent_arg,
-                                                    T pre_arg,
-                                                    T* child_args,
-                                                    int nchild_args) {
-  return pre_arg;
-}
-
-template<typename T> T Regexp::Walker<T>::Copy(T arg) {
-  return arg;
-}
-
-// State about a single level in the traversal.
-template<typename T> struct WalkState {
-  WalkState(Regexp* re, T parent)
-    : re(re),
-      n(-1),
-      parent_arg(parent),
-      child_args(NULL) { }
-
-  Regexp* re;  // The regexp
-  int n;  // The index of the next child to process; -1 means need to PreVisit
-  T parent_arg;  // Accumulated arguments.
-  T pre_arg;
-  T child_arg;  // One-element buffer for child_args.
-  T* child_args;
-};
-
-template<typename T> Regexp::Walker<T>::Walker() {
-  stopped_early_ = false;
-}
-
-template<typename T> Regexp::Walker<T>::~Walker() {
-  Reset();
-}
-
-// Clears the stack.  Should never be necessary, since
-// Walk always enters and exits with an empty stack.
-// Logs DFATAL if stack is not already clear.
-template<typename T> void Regexp::Walker<T>::Reset() {
-  if (!stack_.empty()) {
-    LOG(DFATAL) << "Stack not empty.";
-    while (!stack_.empty()) {
-      if (stack_.top().re->nsub_ > 1)
-        delete[] stack_.top().child_args;
-      stack_.pop();
-    }
-  }
-}
-
-template<typename T> T Regexp::Walker<T>::WalkInternal(Regexp* re, T top_arg,
-                                                       bool use_copy) {
-  Reset();
-
-  if (re == NULL) {
-    LOG(DFATAL) << "Walk NULL";
-    return top_arg;
-  }
-
-  stack_.push(WalkState<T>(re, top_arg));
-
-  WalkState<T>* s;
-  for (;;) {
-    T t;
-    s = &stack_.top();
-    re = s->re;
-    switch (s->n) {
-      case -1: {
-        if (--max_visits_ < 0) {
-          stopped_early_ = true;
-          t = ShortVisit(re, s->parent_arg);
-          break;
-        }
-        bool stop = false;
-        s->pre_arg = PreVisit(re, s->parent_arg, &stop);
-        if (stop) {
-          t = s->pre_arg;
-          break;
-        }
-        s->n = 0;
-        s->child_args = NULL;
-        if (re->nsub_ == 1)
-          s->child_args = &s->child_arg;
-        else if (re->nsub_ > 1)
-          s->child_args = new T[re->nsub_];
-        ABSL_FALLTHROUGH_INTENDED;
-      }
-      default: {
-        if (re->nsub_ > 0) {
-          Regexp** sub = re->sub();
-          if (s->n < re->nsub_) {
-            if (use_copy && s->n > 0 && sub[s->n - 1] == sub[s->n]) {
-              s->child_args[s->n] = Copy(s->child_args[s->n - 1]);
-              s->n++;
-            } else {
-              stack_.push(WalkState<T>(sub[s->n], s->pre_arg));
-            }
-            continue;
-          }
-        }
-
-        t = PostVisit(re, s->parent_arg, s->pre_arg, s->child_args, s->n);
-        if (re->nsub_ > 1)
-          delete[] s->child_args;
-        break;
-      }
-    }
-
-    // We've finished stack_.top().
-    // Update next guy down.
-    stack_.pop();
-    if (stack_.empty())
-      return t;
-    s = &stack_.top();
-    if (s->child_args != NULL)
-      s->child_args[s->n] = t;
-    else
-      s->child_arg = t;
-    s->n++;
-  }
-}
-
-template<typename T> T Regexp::Walker<T>::Walk(Regexp* re, T top_arg) {
-  // Without the exponential walking behavior,
-  // this budget should be more than enough for any
-  // regexp, and yet not enough to get us in trouble
-  // as far as CPU time.
-  max_visits_ = 1000000;
-  return WalkInternal(re, top_arg, true);
-}
-
-template<typename T> T Regexp::Walker<T>::WalkExponential(Regexp* re, T top_arg,
-                                                          int max_visits) {
-  max_visits_ = max_visits;
-  return WalkInternal(re, top_arg, false);
-}
-
-}  // namespace re2
-
-#endif  // RE2_WALKER_INL_H_
diff --git a/projects/llm_framework/include/re2/util/logging.h b/projects/llm_framework/include/re2/util/logging.h
deleted file mode 100644
index 946962b3..00000000
--- a/projects/llm_framework/include/re2/util/logging.h
+++ /dev/null
@@ -1,109 +0,0 @@
-// Copyright 2009 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef UTIL_LOGGING_H_
-#define UTIL_LOGGING_H_
-
-// Simplified version of Google's logging.
-
-#include <assert.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <ostream>
-#include <sstream>
-
-#include "absl/base/attributes.h"
-
-// Debug-only checking.
-#define DCHECK(condition) assert(condition)
-#define DCHECK_EQ(val1, val2) assert((val1) == (val2))
-#define DCHECK_NE(val1, val2) assert((val1) != (val2))
-#define DCHECK_LE(val1, val2) assert((val1) <= (val2))
-#define DCHECK_LT(val1, val2) assert((val1) < (val2))
-#define DCHECK_GE(val1, val2) assert((val1) >= (val2))
-#define DCHECK_GT(val1, val2) assert((val1) > (val2))
-
-// Always-on checking
-#define CHECK(x)	if(x){}else LogMessageFatal(__FILE__, __LINE__).stream() << "Check failed: " #x
-#define CHECK_LT(x, y)	CHECK((x) < (y))
-#define CHECK_GT(x, y)	CHECK((x) > (y))
-#define CHECK_LE(x, y)	CHECK((x) <= (y))
-#define CHECK_GE(x, y)	CHECK((x) >= (y))
-#define CHECK_EQ(x, y)	CHECK((x) == (y))
-#define CHECK_NE(x, y)	CHECK((x) != (y))
-
-#define LOG_INFO LogMessage(__FILE__, __LINE__)
-#define LOG_WARNING LogMessage(__FILE__, __LINE__)
-#define LOG_ERROR LogMessage(__FILE__, __LINE__)
-#define LOG_FATAL LogMessageFatal(__FILE__, __LINE__)
-#define LOG_QFATAL LOG_FATAL
-
-// It seems that one of the Windows header files defines ERROR as 0.
-#ifdef _WIN32
-#define LOG_0 LOG_INFO
-#endif
-
-#ifdef NDEBUG
-#define LOG_DFATAL LOG_ERROR
-#else
-#define LOG_DFATAL LOG_FATAL
-#endif
-
-#define LOG(severity) LOG_ ## severity.stream()
-
-#define VLOG(x) if((x)>0){}else LOG_INFO.stream()
-
-class LogMessage {
- public:
-  LogMessage(const char* file, int line)
-      : flushed_(false) {
-    stream() << file << ":" << line << ": ";
-  }
-  void Flush() {
-    stream() << "\n";
-    std::string s = str_.str();
-    size_t n = s.size();
-    if (fwrite(s.data(), 1, n, stderr) < n) {}  // shut up gcc
-    flushed_ = true;
-  }
-  ~LogMessage() {
-    if (!flushed_) {
-      Flush();
-    }
-  }
-  std::ostream& stream() { return str_; }
-
- private:
-  bool flushed_;
-  std::ostringstream str_;
-
-  LogMessage(const LogMessage&) = delete;
-  LogMessage& operator=(const LogMessage&) = delete;
-};
-
-// Silence "destructor never returns" warning for ~LogMessageFatal().
-// Since this is a header file, push and then pop to limit the scope.
-#ifdef _MSC_VER
-#pragma warning(push)
-#pragma warning(disable: 4722)
-#endif
-
-class LogMessageFatal : public LogMessage {
- public:
-  LogMessageFatal(const char* file, int line)
-      : LogMessage(file, line) {}
-  ABSL_ATTRIBUTE_NORETURN ~LogMessageFatal() {
-    Flush();
-    abort();
-  }
- private:
-  LogMessageFatal(const LogMessageFatal&) = delete;
-  LogMessageFatal& operator=(const LogMessageFatal&) = delete;
-};
-
-#ifdef _MSC_VER
-#pragma warning(pop)
-#endif
-
-#endif  // UTIL_LOGGING_H_
diff --git a/projects/llm_framework/include/re2/util/malloc_counter.h b/projects/llm_framework/include/re2/util/malloc_counter.h
deleted file mode 100644
index 81b564ff..00000000
--- a/projects/llm_framework/include/re2/util/malloc_counter.h
+++ /dev/null
@@ -1,19 +0,0 @@
-// Copyright 2009 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef UTIL_MALLOC_COUNTER_H_
-#define UTIL_MALLOC_COUNTER_H_
-
-namespace testing {
-class MallocCounter {
- public:
-  MallocCounter(int x) {}
-  static const int THIS_THREAD_ONLY = 0;
-  long long HeapGrowth() { return 0; }
-  long long PeakHeapGrowth() { return 0; }
-  void Reset() {}
-};
-}  // namespace testing
-
-#endif  // UTIL_MALLOC_COUNTER_H_
diff --git a/projects/llm_framework/include/re2/util/pcre.h b/projects/llm_framework/include/re2/util/pcre.h
deleted file mode 100644
index 846f3001..00000000
--- a/projects/llm_framework/include/re2/util/pcre.h
+++ /dev/null
@@ -1,671 +0,0 @@
-// Copyright 2003-2010 Google Inc.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef UTIL_PCRE_H_
-#define UTIL_PCRE_H_
-
-// This is a variant of PCRE's pcrecpp.h, originally written at Google.
-// The main changes are the addition of the HitLimit method and
-// compilation as PCRE in namespace re2.
-
-// C++ interface to the pcre regular-expression library.  PCRE supports
-// Perl-style regular expressions (with extensions like \d, \w, \s,
-// ...).
-//
-// -----------------------------------------------------------------------
-// REGEXP SYNTAX:
-//
-// This module uses the pcre library and hence supports its syntax
-// for regular expressions:
-//
-//      http://www.google.com/search?q=pcre
-//
-// The syntax is pretty similar to Perl's.  For those not familiar
-// with Perl's regular expressions, here are some examples of the most
-// commonly used extensions:
-//
-//   "hello (\\w+) world"  -- \w matches a "word" character
-//   "version (\\d+)"      -- \d matches a digit
-//   "hello\\s+world"      -- \s matches any whitespace character
-//   "\\b(\\w+)\\b"        -- \b matches empty string at a word boundary
-//   "(?i)hello"           -- (?i) turns on case-insensitive matching
-//   "/\\*(.*?)\\*/"       -- .*? matches . minimum no. of times possible
-//
-// -----------------------------------------------------------------------
-// MATCHING INTERFACE:
-//
-// The "FullMatch" operation checks that supplied text matches a
-// supplied pattern exactly.
-//
-// Example: successful match
-//    CHECK(PCRE::FullMatch("hello", "h.*o"));
-//
-// Example: unsuccessful match (requires full match):
-//    CHECK(!PCRE::FullMatch("hello", "e"));
-//
-// -----------------------------------------------------------------------
-// UTF-8 AND THE MATCHING INTERFACE:
-//
-// By default, pattern and text are plain text, one byte per character.
-// The UTF8 flag, passed to the constructor, causes both pattern
-// and string to be treated as UTF-8 text, still a byte stream but
-// potentially multiple bytes per character. In practice, the text
-// is likelier to be UTF-8 than the pattern, but the match returned
-// may depend on the UTF8 flag, so always use it when matching
-// UTF8 text.  E.g., "." will match one byte normally but with UTF8
-// set may match up to three bytes of a multi-byte character.
-//
-// Example:
-//    PCRE re(utf8_pattern, PCRE::UTF8);
-//    CHECK(PCRE::FullMatch(utf8_string, re));
-//
-// -----------------------------------------------------------------------
-// MATCHING WITH SUBSTRING EXTRACTION:
-//
-// You can supply extra pointer arguments to extract matched substrings.
-//
-// Example: extracts "ruby" into "s" and 1234 into "i"
-//    int i;
-//    std::string s;
-//    CHECK(PCRE::FullMatch("ruby:1234", "(\\w+):(\\d+)", &s, &i));
-//
-// Example: fails because string cannot be stored in integer
-//    CHECK(!PCRE::FullMatch("ruby", "(.*)", &i));
-//
-// Example: fails because there aren't enough sub-patterns:
-//    CHECK(!PCRE::FullMatch("ruby:1234", "\\w+:\\d+", &s));
-//
-// Example: does not try to extract any extra sub-patterns
-//    CHECK(PCRE::FullMatch("ruby:1234", "(\\w+):(\\d+)", &s));
-//
-// Example: does not try to extract into NULL
-//    CHECK(PCRE::FullMatch("ruby:1234", "(\\w+):(\\d+)", NULL, &i));
-//
-// Example: integer overflow causes failure
-//    CHECK(!PCRE::FullMatch("ruby:1234567891234", "\\w+:(\\d+)", &i));
-//
-// -----------------------------------------------------------------------
-// PARTIAL MATCHES
-//
-// You can use the "PartialMatch" operation when you want the pattern
-// to match any substring of the text.
-//
-// Example: simple search for a string:
-//      CHECK(PCRE::PartialMatch("hello", "ell"));
-//
-// Example: find first number in a string
-//      int number;
-//      CHECK(PCRE::PartialMatch("x*100 + 20", "(\\d+)", &number));
-//      CHECK_EQ(number, 100);
-//
-// -----------------------------------------------------------------------
-// PPCRE-COMPILED PCREGULAR EXPPCRESSIONS
-//
-// PCRE makes it easy to use any string as a regular expression, without
-// requiring a separate compilation step.
-//
-// If speed is of the essence, you can create a pre-compiled "PCRE"
-// object from the pattern and use it multiple times.  If you do so,
-// you can typically parse text faster than with sscanf.
-//
-// Example: precompile pattern for faster matching:
-//    PCRE pattern("h.*o");
-//    while (ReadLine(&str)) {
-//      if (PCRE::FullMatch(str, pattern)) ...;
-//    }
-//
-// -----------------------------------------------------------------------
-// SCANNING TEXT INCPCREMENTALLY
-//
-// The "Consume" operation may be useful if you want to repeatedly
-// match regular expressions at the front of a string and skip over
-// them as they match.  This requires use of the string_view type,
-// which represents a sub-range of a real string.
-//
-// Example: read lines of the form "var = value" from a string.
-//      std::string contents = ...;         // Fill string somehow
-//      absl::string_view input(contents);  // Wrap a string_view around it
-//
-//      std::string var;
-//      int value;
-//      while (PCRE::Consume(&input, "(\\w+) = (\\d+)\n", &var, &value)) {
-//        ...;
-//      }
-//
-// Each successful call to "Consume" will set "var/value", and also
-// advance "input" so it points past the matched text.  Note that if the
-// regular expression matches an empty string, input will advance
-// by 0 bytes.  If the regular expression being used might match
-// an empty string, the loop body must check for this case and either
-// advance the string or break out of the loop.
-//
-// The "FindAndConsume" operation is similar to "Consume" but does not
-// anchor your match at the beginning of the string.  For example, you
-// could extract all words from a string by repeatedly calling
-//     PCRE::FindAndConsume(&input, "(\\w+)", &word)
-//
-// -----------------------------------------------------------------------
-// PARSING HEX/OCTAL/C-RADIX NUMBERS
-//
-// By default, if you pass a pointer to a numeric value, the
-// corresponding text is interpreted as a base-10 number.  You can
-// instead wrap the pointer with a call to one of the operators Hex(),
-// Octal(), or CRadix() to interpret the text in another base.  The
-// CRadix operator interprets C-style "0" (base-8) and "0x" (base-16)
-// prefixes, but defaults to base-10.
-//
-// Example:
-//   int a, b, c, d;
-//   CHECK(PCRE::FullMatch("100 40 0100 0x40", "(.*) (.*) (.*) (.*)",
-//         Octal(&a), Hex(&b), CRadix(&c), CRadix(&d));
-// will leave 64 in a, b, c, and d.
-
-#include "absl/strings/string_view.h"
-
-#ifdef USEPCRE
-#include <pcre.h>
-namespace re2 {
-const bool UsingPCRE = true;
-}  // namespace re2
-#else
-struct pcre;  // opaque
-namespace re2 {
-const bool UsingPCRE = false;
-}  // namespace re2
-#endif
-
-// To produce a DLL, CMake can automatically export code symbols,
-// but not data symbols, so we have to annotate those manually...
-#if defined(RE2_BUILD_TESTING_DLL)
-#define RE2_TESTING_DLL __declspec(dllexport)
-#elif defined(RE2_CONSUME_TESTING_DLL)
-#define RE2_TESTING_DLL __declspec(dllimport)
-#else
-#define RE2_TESTING_DLL
-#endif
-
-namespace re2 {
-
-class PCRE_Options;
-
-// Interface for regular expression matching.  Also corresponds to a
-// pre-compiled regular expression.  An "PCRE" object is safe for
-// concurrent use by multiple threads.
-class PCRE {
- public:
-  // We convert user-passed pointers into special Arg objects
-  class Arg;
-
-  // Marks end of arg list.
-  // ONLY USE IN OPTIONAL ARG DEFAULTS.
-  // DO NOT PASS EXPLICITLY.
-  RE2_TESTING_DLL static Arg no_more_args;
-
-  // Options are same value as those in pcre.  We provide them here
-  // to avoid users needing to include pcre.h and also to isolate
-  // users from pcre should we change the underlying library.
-  // Only those needed by Google programs are exposed here to
-  // avoid collision with options employed internally by regexp.cc
-  // Note that some options have equivalents that can be specified in
-  // the regexp itself.  For example, prefixing your regexp with
-  // "(?s)" has the same effect as the PCRE_DOTALL option.
-  enum Option {
-    None = 0x0000,
-    UTF8 = 0x0800,  // == PCRE_UTF8
-    EnabledCompileOptions = UTF8,
-    EnabledExecOptions = 0x0000,  // TODO: use to replace anchor flag
-  };
-
-  // We provide implicit conversions from strings so that users can
-  // pass in a string or a "const char*" wherever an "PCRE" is expected.
-  PCRE(const char* pattern);
-  PCRE(const char* pattern, Option option);
-  PCRE(const std::string& pattern);
-  PCRE(const std::string& pattern, Option option);
-  PCRE(const char *pattern, const PCRE_Options& re_option);
-  PCRE(const std::string& pattern, const PCRE_Options& re_option);
-
-  ~PCRE();
-
-  // The string specification for this PCRE.  E.g.
-  //   PCRE re("ab*c?d+");
-  //   re.pattern();    // "ab*c?d+"
-  const std::string& pattern() const { return pattern_; }
-
-  // If PCRE could not be created properly, returns an error string.
-  // Else returns the empty string.
-  const std::string& error() const { return *error_; }
-
-  // Whether the PCRE has hit a match limit during execution.
-  // Not thread safe.  Intended only for testing.
-  // If hitting match limits is a problem,
-  // you should be using PCRE2 (re2/re2.h)
-  // instead of checking this flag.
-  bool HitLimit();
-  void ClearHitLimit();
-
-  /***** The useful part: the matching interface *****/
-
-  // Matches "text" against "pattern".  If pointer arguments are
-  // supplied, copies matched sub-patterns into them.
-  //
-  // You can pass in a "const char*" or a "std::string" for "text".
-  // You can pass in a "const char*" or a "std::string" or a "PCRE" for "pattern".
-  //
-  // The provided pointer arguments can be pointers to any scalar numeric
-  // type, or one of:
-  //    std::string        (matched piece is copied to string)
-  //    absl::string_view  (string_view is mutated to point to matched piece)
-  //    T                  ("bool T::ParseFrom(const char*, size_t)" must exist)
-  //    (void*)NULL        (the corresponding matched sub-pattern is not copied)
-  //
-  // Returns true iff all of the following conditions are satisfied:
-  //   a. "text" matches "pattern" exactly
-  //   b. The number of matched sub-patterns is >= number of supplied pointers
-  //   c. The "i"th argument has a suitable type for holding the
-  //      string captured as the "i"th sub-pattern.  If you pass in
-  //      NULL for the "i"th argument, or pass fewer arguments than
-  //      number of sub-patterns, "i"th captured sub-pattern is
-  //      ignored.
-  //
-  // CAVEAT: An optional sub-pattern that does not exist in the
-  // matched string is assigned the empty string.  Therefore, the
-  // following will return false (because the empty string is not a
-  // valid number):
-  //    int number;
-  //    PCRE::FullMatch("abc", "[a-z]+(\\d+)?", &number);
-  struct FullMatchFunctor {
-    bool operator ()(absl::string_view text, const PCRE& re,  // 3..16 args
-                     const Arg& ptr1 = no_more_args,
-                     const Arg& ptr2 = no_more_args,
-                     const Arg& ptr3 = no_more_args,
-                     const Arg& ptr4 = no_more_args,
-                     const Arg& ptr5 = no_more_args,
-                     const Arg& ptr6 = no_more_args,
-                     const Arg& ptr7 = no_more_args,
-                     const Arg& ptr8 = no_more_args,
-                     const Arg& ptr9 = no_more_args,
-                     const Arg& ptr10 = no_more_args,
-                     const Arg& ptr11 = no_more_args,
-                     const Arg& ptr12 = no_more_args,
-                     const Arg& ptr13 = no_more_args,
-                     const Arg& ptr14 = no_more_args,
-                     const Arg& ptr15 = no_more_args,
-                     const Arg& ptr16 = no_more_args) const;
-  };
-
-  RE2_TESTING_DLL static const FullMatchFunctor FullMatch;
-
-  // Exactly like FullMatch(), except that "pattern" is allowed to match
-  // a substring of "text".
-  struct PartialMatchFunctor {
-    bool operator ()(absl::string_view text, const PCRE& re,  // 3..16 args
-                     const Arg& ptr1 = no_more_args,
-                     const Arg& ptr2 = no_more_args,
-                     const Arg& ptr3 = no_more_args,
-                     const Arg& ptr4 = no_more_args,
-                     const Arg& ptr5 = no_more_args,
-                     const Arg& ptr6 = no_more_args,
-                     const Arg& ptr7 = no_more_args,
-                     const Arg& ptr8 = no_more_args,
-                     const Arg& ptr9 = no_more_args,
-                     const Arg& ptr10 = no_more_args,
-                     const Arg& ptr11 = no_more_args,
-                     const Arg& ptr12 = no_more_args,
-                     const Arg& ptr13 = no_more_args,
-                     const Arg& ptr14 = no_more_args,
-                     const Arg& ptr15 = no_more_args,
-                     const Arg& ptr16 = no_more_args) const;
-  };
-
-  RE2_TESTING_DLL static const PartialMatchFunctor PartialMatch;
-
-  // Like FullMatch() and PartialMatch(), except that pattern has to
-  // match a prefix of "text", and "input" is advanced past the matched
-  // text.  Note: "input" is modified iff this routine returns true.
-  struct ConsumeFunctor {
-    bool operator ()(absl::string_view* input, const PCRE& pattern,  // 3..16 args
-                     const Arg& ptr1 = no_more_args,
-                     const Arg& ptr2 = no_more_args,
-                     const Arg& ptr3 = no_more_args,
-                     const Arg& ptr4 = no_more_args,
-                     const Arg& ptr5 = no_more_args,
-                     const Arg& ptr6 = no_more_args,
-                     const Arg& ptr7 = no_more_args,
-                     const Arg& ptr8 = no_more_args,
-                     const Arg& ptr9 = no_more_args,
-                     const Arg& ptr10 = no_more_args,
-                     const Arg& ptr11 = no_more_args,
-                     const Arg& ptr12 = no_more_args,
-                     const Arg& ptr13 = no_more_args,
-                     const Arg& ptr14 = no_more_args,
-                     const Arg& ptr15 = no_more_args,
-                     const Arg& ptr16 = no_more_args) const;
-  };
-
-  RE2_TESTING_DLL static const ConsumeFunctor Consume;
-
-  // Like Consume(..), but does not anchor the match at the beginning of the
-  // string.  That is, "pattern" need not start its match at the beginning of
-  // "input".  For example, "FindAndConsume(s, "(\\w+)", &word)" finds the next
-  // word in "s" and stores it in "word".
-  struct FindAndConsumeFunctor {
-    bool operator ()(absl::string_view* input, const PCRE& pattern,  // 3..16 args
-                     const Arg& ptr1 = no_more_args,
-                     const Arg& ptr2 = no_more_args,
-                     const Arg& ptr3 = no_more_args,
-                     const Arg& ptr4 = no_more_args,
-                     const Arg& ptr5 = no_more_args,
-                     const Arg& ptr6 = no_more_args,
-                     const Arg& ptr7 = no_more_args,
-                     const Arg& ptr8 = no_more_args,
-                     const Arg& ptr9 = no_more_args,
-                     const Arg& ptr10 = no_more_args,
-                     const Arg& ptr11 = no_more_args,
-                     const Arg& ptr12 = no_more_args,
-                     const Arg& ptr13 = no_more_args,
-                     const Arg& ptr14 = no_more_args,
-                     const Arg& ptr15 = no_more_args,
-                     const Arg& ptr16 = no_more_args) const;
-  };
-
-  RE2_TESTING_DLL static const FindAndConsumeFunctor FindAndConsume;
-
-  // Replace the first match of "pattern" in "str" with "rewrite".
-  // Within "rewrite", backslash-escaped digits (\1 to \9) can be
-  // used to insert text matching corresponding parenthesized group
-  // from the pattern.  \0 in "rewrite" refers to the entire matching
-  // text.  E.g.,
-  //
-  //   std::string s = "yabba dabba doo";
-  //   CHECK(PCRE::Replace(&s, "b+", "d"));
-  //
-  // will leave "s" containing "yada dabba doo"
-  //
-  // Returns true if the pattern matches and a replacement occurs,
-  // false otherwise.
-  static bool Replace(std::string* str, const PCRE& pattern,
-                      absl::string_view rewrite);
-
-  // Like Replace(), except replaces all occurrences of the pattern in
-  // the string with the rewrite.  Replacements are not subject to
-  // re-matching.  E.g.,
-  //
-  //   std::string s = "yabba dabba doo";
-  //   CHECK(PCRE::GlobalReplace(&s, "b+", "d"));
-  //
-  // will leave "s" containing "yada dada doo"
-  //
-  // Returns the number of replacements made.
-  static int GlobalReplace(std::string* str, const PCRE& pattern,
-                           absl::string_view rewrite);
-
-  // Like Replace, except that if the pattern matches, "rewrite"
-  // is copied into "out" with substitutions.  The non-matching
-  // portions of "text" are ignored.
-  //
-  // Returns true iff a match occurred and the extraction happened
-  // successfully;  if no match occurs, the string is left unaffected.
-  static bool Extract(absl::string_view text, const PCRE& pattern,
-                      absl::string_view rewrite, std::string* out);
-
-  // Check that the given @p rewrite string is suitable for use with
-  // this PCRE.  It checks that:
-  //   * The PCRE has enough parenthesized subexpressions to satisfy all
-  //       of the \N tokens in @p rewrite, and
-  //   * The @p rewrite string doesn't have any syntax errors
-  //       ('\' followed by anything besides [0-9] and '\').
-  // Making this test will guarantee that "replace" and "extract"
-  // operations won't LOG(ERROR) or fail because of a bad rewrite
-  // string.
-  // @param rewrite The proposed rewrite string.
-  // @param error An error message is recorded here, iff we return false.
-  //              Otherwise, it is unchanged.
-  // @return true, iff @p rewrite is suitable for use with the PCRE.
-  bool CheckRewriteString(absl::string_view rewrite, std::string* error) const;
-
-  // Returns a copy of 'unquoted' with all potentially meaningful
-  // regexp characters backslash-escaped.  The returned string, used
-  // as a regular expression, will exactly match the original string.
-  // For example,
-  //           1.5-2.0?
-  //  becomes:
-  //           1\.5\-2\.0\?
-  static std::string QuoteMeta(absl::string_view unquoted);
-
-  /***** Generic matching interface (not so nice to use) *****/
-
-  // Type of match (TODO: Should be restructured as an Option)
-  enum Anchor {
-    UNANCHORED,         // No anchoring
-    ANCHOR_START,       // Anchor at start only
-    ANCHOR_BOTH,        // Anchor at start and end
-  };
-
-  // General matching routine.  Stores the length of the match in
-  // "*consumed" if successful.
-  bool DoMatch(absl::string_view text, Anchor anchor, size_t* consumed,
-               const Arg* const* args, int n) const;
-
-  // Return the number of capturing subpatterns, or -1 if the
-  // regexp wasn't valid on construction.
-  int NumberOfCapturingGroups() const;
-
- private:
-  void Init(const char* pattern, Option option, int match_limit,
-            int stack_limit, bool report_errors);
-
-  // Match against "text", filling in "vec" (up to "vecsize" * 2/3) with
-  // pairs of integers for the beginning and end positions of matched
-  // text.  The first pair corresponds to the entire matched text;
-  // subsequent pairs correspond, in order, to parentheses-captured
-  // matches.  Returns the number of pairs (one more than the number of
-  // the last subpattern with a match) if matching was successful
-  // and zero if the match failed.
-  // I.e. for PCRE("(foo)|(bar)|(baz)") it will return 2, 3, and 4 when matching
-  // against "foo", "bar", and "baz" respectively.
-  // When matching PCRE("(foo)|hello") against "hello", it will return 1.
-  // But the values for all subpattern are filled in into "vec".
-  int TryMatch(absl::string_view text, size_t startpos, Anchor anchor,
-               bool empty_ok, int* vec, int vecsize) const;
-
-  // Append the "rewrite" string, with backslash substitutions from "text"
-  // and "vec", to string "out".
-  bool Rewrite(std::string* out, absl::string_view rewrite,
-               absl::string_view text, int* vec, int veclen) const;
-
-  // internal implementation for DoMatch
-  bool DoMatchImpl(absl::string_view text, Anchor anchor, size_t* consumed,
-                   const Arg* const args[], int n, int* vec, int vecsize) const;
-
-  // Compile the regexp for the specified anchoring mode
-  pcre* Compile(Anchor anchor);
-
-  std::string         pattern_;
-  Option              options_;
-  pcre*               re_full_;        // For full matches
-  pcre*               re_partial_;     // For partial matches
-  const std::string*  error_;          // Error indicator (or empty string)
-  bool                report_errors_;  // Silences error logging if false
-  int                 match_limit_;    // Limit on execution resources
-  int                 stack_limit_;    // Limit on stack resources (bytes)
-  mutable int         hit_limit_;      // Hit limit during execution (bool)
-
-  PCRE(const PCRE&) = delete;
-  PCRE& operator=(const PCRE&) = delete;
-};
-
-// PCRE_Options allow you to set the PCRE::Options, plus any pcre
-// "extra" options.  The only extras are match_limit, which limits
-// the CPU time of a match, and stack_limit, which limits the
-// stack usage.  Setting a limit to <= 0 lets PCRE pick a sensible default
-// that should not cause too many problems in production code.
-// If PCRE hits a limit during a match, it may return a false negative,
-// but (hopefully) it won't crash.
-//
-// NOTE: If you are handling regular expressions specified by
-// (external or internal) users, rather than hard-coded ones,
-// you should be using PCRE2, which uses an alternate implementation
-// that avoids these issues.  See http://go/re2quick.
-class PCRE_Options {
- public:
-  // constructor
-  PCRE_Options() : option_(PCRE::None), match_limit_(0), stack_limit_(0), report_errors_(true) {}
-  // accessors
-  PCRE::Option option() const { return option_; }
-  void set_option(PCRE::Option option) {
-    option_ = option;
-  }
-  int match_limit() const { return match_limit_; }
-  void set_match_limit(int match_limit) {
-    match_limit_ = match_limit;
-  }
-  int stack_limit() const { return stack_limit_; }
-  void set_stack_limit(int stack_limit) {
-    stack_limit_ = stack_limit;
-  }
-
-  // If the regular expression is malformed, an error message will be printed
-  // iff report_errors() is true.  Default: true.
-  bool report_errors() const { return report_errors_; }
-  void set_report_errors(bool report_errors) {
-    report_errors_ = report_errors;
-  }
- private:
-  PCRE::Option option_;
-  int match_limit_;
-  int stack_limit_;
-  bool report_errors_;
-};
-
-
-/***** Implementation details *****/
-
-// Hex/Octal/Binary?
-
-// Special class for parsing into objects that define a ParseFrom() method
-template <typename T>
-class _PCRE_MatchObject {
- public:
-  static inline bool Parse(const char* str, size_t n, void* dest) {
-    if (dest == NULL) return true;
-    T* object = reinterpret_cast<T*>(dest);
-    return object->ParseFrom(str, n);
-  }
-};
-
-class PCRE::Arg {
- public:
-  // Empty constructor so we can declare arrays of PCRE::Arg
-  Arg();
-
-  // Constructor specially designed for NULL arguments
-  Arg(void*);
-
-  typedef bool (*Parser)(const char* str, size_t n, void* dest);
-
-// Type-specific parsers
-#define MAKE_PARSER(type, name)            \
-  Arg(type* p) : arg_(p), parser_(name) {} \
-  Arg(type* p, Parser parser) : arg_(p), parser_(parser) {}
-
-  MAKE_PARSER(char,               parse_char);
-  MAKE_PARSER(signed char,        parse_schar);
-  MAKE_PARSER(unsigned char,      parse_uchar);
-  MAKE_PARSER(float,              parse_float);
-  MAKE_PARSER(double,             parse_double);
-  MAKE_PARSER(std::string,        parse_string);
-  MAKE_PARSER(absl::string_view,  parse_string_view);
-
-  MAKE_PARSER(short,              parse_short);
-  MAKE_PARSER(unsigned short,     parse_ushort);
-  MAKE_PARSER(int,                parse_int);
-  MAKE_PARSER(unsigned int,       parse_uint);
-  MAKE_PARSER(long,               parse_long);
-  MAKE_PARSER(unsigned long,      parse_ulong);
-  MAKE_PARSER(long long,          parse_longlong);
-  MAKE_PARSER(unsigned long long, parse_ulonglong);
-
-#undef MAKE_PARSER
-
-  // Generic constructor
-  template <typename T> Arg(T*, Parser parser);
-  // Generic constructor template
-  template <typename T> Arg(T* p)
-    : arg_(p), parser_(_PCRE_MatchObject<T>::Parse) {
-  }
-
-  // Parse the data
-  bool Parse(const char* str, size_t n) const;
-
- private:
-  void*         arg_;
-  Parser        parser_;
-
-  static bool parse_null        (const char* str, size_t n, void* dest);
-  static bool parse_char        (const char* str, size_t n, void* dest);
-  static bool parse_schar       (const char* str, size_t n, void* dest);
-  static bool parse_uchar       (const char* str, size_t n, void* dest);
-  static bool parse_float       (const char* str, size_t n, void* dest);
-  static bool parse_double      (const char* str, size_t n, void* dest);
-  static bool parse_string      (const char* str, size_t n, void* dest);
-  static bool parse_string_view (const char* str, size_t n, void* dest);
-
-#define DECLARE_INTEGER_PARSER(name)                                       \
- private:                                                                  \
-  static bool parse_##name(const char* str, size_t n, void* dest);         \
-  static bool parse_##name##_radix(const char* str, size_t n, void* dest,  \
-                                   int radix);                             \
-                                                                           \
- public:                                                                   \
-  static bool parse_##name##_hex(const char* str, size_t n, void* dest);   \
-  static bool parse_##name##_octal(const char* str, size_t n, void* dest); \
-  static bool parse_##name##_cradix(const char* str, size_t n, void* dest)
-
-  DECLARE_INTEGER_PARSER(short);
-  DECLARE_INTEGER_PARSER(ushort);
-  DECLARE_INTEGER_PARSER(int);
-  DECLARE_INTEGER_PARSER(uint);
-  DECLARE_INTEGER_PARSER(long);
-  DECLARE_INTEGER_PARSER(ulong);
-  DECLARE_INTEGER_PARSER(longlong);
-  DECLARE_INTEGER_PARSER(ulonglong);
-
-#undef DECLARE_INTEGER_PARSER
-
-};
-
-inline PCRE::Arg::Arg() : arg_(NULL), parser_(parse_null) { }
-inline PCRE::Arg::Arg(void* p) : arg_(p), parser_(parse_null) { }
-
-inline bool PCRE::Arg::Parse(const char* str, size_t n) const {
-  return (*parser_)(str, n, arg_);
-}
-
-// This part of the parser, appropriate only for ints, deals with bases
-#define MAKE_INTEGER_PARSER(type, name)                      \
-  inline PCRE::Arg Hex(type* ptr) {                          \
-    return PCRE::Arg(ptr, PCRE::Arg::parse_##name##_hex);    \
-  }                                                          \
-  inline PCRE::Arg Octal(type* ptr) {                        \
-    return PCRE::Arg(ptr, PCRE::Arg::parse_##name##_octal);  \
-  }                                                          \
-  inline PCRE::Arg CRadix(type* ptr) {                       \
-    return PCRE::Arg(ptr, PCRE::Arg::parse_##name##_cradix); \
-  }
-
-MAKE_INTEGER_PARSER(short,              short);
-MAKE_INTEGER_PARSER(unsigned short,     ushort);
-MAKE_INTEGER_PARSER(int,                int);
-MAKE_INTEGER_PARSER(unsigned int,       uint);
-MAKE_INTEGER_PARSER(long,               long);
-MAKE_INTEGER_PARSER(unsigned long,      ulong);
-MAKE_INTEGER_PARSER(long long,          longlong);
-MAKE_INTEGER_PARSER(unsigned long long, ulonglong);
-
-#undef MAKE_INTEGER_PARSER
-
-}  // namespace re2
-
-#endif  // UTIL_PCRE_H_
diff --git a/projects/llm_framework/include/re2/util/strutil.h b/projects/llm_framework/include/re2/util/strutil.h
deleted file mode 100644
index f5d87a53..00000000
--- a/projects/llm_framework/include/re2/util/strutil.h
+++ /dev/null
@@ -1,16 +0,0 @@
-// Copyright 2016 The RE2 Authors.  All Rights Reserved.
-// Use of this source code is governed by a BSD-style
-// license that can be found in the LICENSE file.
-
-#ifndef UTIL_STRUTIL_H_
-#define UTIL_STRUTIL_H_
-
-#include <string>
-
-namespace re2 {
-
-void PrefixSuccessor(std::string* prefix);
-
-}  // namespace re2
-
-#endif  // UTIL_STRUTIL_H_
diff --git a/projects/llm_framework/include/re2/util/utf.h b/projects/llm_framework/include/re2/util/utf.h
deleted file mode 100644
index 85b42972..00000000
--- a/projects/llm_framework/include/re2/util/utf.h
+++ /dev/null
@@ -1,44 +0,0 @@
-/*
- * The authors of this software are Rob Pike and Ken Thompson.
- *              Copyright (c) 2002 by Lucent Technologies.
- * Permission to use, copy, modify, and distribute this software for any
- * purpose without fee is hereby granted, provided that this entire notice
- * is included in all copies of any software which is or includes a copy
- * or modification of this software and in all copies of the supporting
- * documentation for such software.
- * THIS SOFTWARE IS BEING PROVIDED "AS IS", WITHOUT ANY EXPRESS OR IMPLIED
- * WARRANTY.  IN PARTICULAR, NEITHER THE AUTHORS NOR LUCENT TECHNOLOGIES MAKE ANY
- * REPRESENTATION OR WARRANTY OF ANY KIND CONCERNING THE MERCHANTABILITY
- * OF THIS SOFTWARE OR ITS FITNESS FOR ANY PARTICULAR PURPOSE.
- *
- * This file and rune.cc have been converted to compile as C++ code
- * in name space re2.
- */
-
-#ifndef UTIL_UTF_H_
-#define UTIL_UTF_H_
-
-#include <stdint.h>
-
-namespace re2 {
-
-typedef signed int Rune;	/* Code-point values in Unicode 4.0 are 21 bits wide.*/
-
-enum
-{
-  UTFmax	= 4,		/* maximum bytes per rune */
-  Runesync	= 0x80,		/* cannot represent part of a UTF sequence (<) */
-  Runeself	= 0x80,		/* rune and UTF sequences are the same (<) */
-  Runeerror	= 0xFFFD,	/* decoding error in UTF */
-  Runemax	= 0x10FFFF,	/* maximum rune value */
-};
-
-int runetochar(char* s, const Rune* r);
-int chartorune(Rune* r, const char* s);
-int fullrune(const char* s, int n);
-int utflen(const char* s);
-char* utfrune(const char*, Rune);
-
-}  // namespace re2
-
-#endif  // UTIL_UTF_H_
diff --git a/projects/llm_framework/include/sentencepiece/bpe_model.h b/projects/llm_framework/include/sentencepiece/bpe_model.h
deleted file mode 100644
index 428b8a0d..00000000
--- a/projects/llm_framework/include/sentencepiece/bpe_model.h
+++ /dev/null
@@ -1,52 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef BPE_MODEL_H_
-#define BPE_MODEL_H_
-
-#include "model_interface.h"
-#include "sentencepiece_model.pb.h"
-
-namespace sentencepiece {
-namespace bpe {
-
-// Segmentation model with BPE (Byte Pair Encoding)
-// Details:
-// Neural Machine Translation of Rare Words with Subword Units
-// https://arxiv.org/abs/1508.07909
-//
-// https://en.wikipedia.org/wiki/Byte_pair_encoding
-class Model : public ModelInterface {
- public:
-  explicit Model(const ModelProto &model_proto);
-  ~Model() override;
-
-  EncodeResult Encode(absl::string_view normalized) const override {
-    return SampleEncode(normalized, 0.0);
-  }
-
-  // Sampling with BPE-dropout: https://arxiv.org/pdf/1910.13267.pdf
-  // `alpha` is dropout probability in BPE-dropout paper.
-  // Skips merge operation with `alpha` probability.
-  // When alpha <= 0.0, no sampling is performed.
-  EncodeResult SampleEncode(absl::string_view normalized,
-                            float alpha) const override;
-
-  bool IsSampleEncodeAvailable() const override { return true; }
-
-  bool IsNBestEncodeAvailable() const override { return false; }
-};
-}  // namespace bpe
-}  // namespace sentencepiece
-#endif  // BPE_MODEL_H_
diff --git a/projects/llm_framework/include/sentencepiece/bpe_model_trainer.h b/projects/llm_framework/include/sentencepiece/bpe_model_trainer.h
deleted file mode 100644
index 15ca4798..00000000
--- a/projects/llm_framework/include/sentencepiece/bpe_model_trainer.h
+++ /dev/null
@@ -1,130 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef BPE_MODEL_TRAINER_H_
-#define BPE_MODEL_TRAINER_H_
-
-#include <cstdint>
-#include <limits>
-#include <string>
-#include <vector>
-
-#include "sentencepiece_model.pb.h"
-#include "third_party/absl/container/btree_set.h"
-#include "third_party/absl/container/flat_hash_map.h"
-#include "trainer_interface.h"
-
-namespace sentencepiece {
-namespace bpe {
-
-// Trainer class for BPE model.
-class Trainer : public TrainerInterface {
- public:
-  Trainer(const TrainerSpec &trainer_spec,
-          const NormalizerSpec &normalizer_spec,
-          const NormalizerSpec &denormalizer_spec)
-      : TrainerInterface::TrainerInterface(trainer_spec, normalizer_spec,
-                                           denormalizer_spec) {}
-
-  util::Status Train() override;
-
- private:
-  // Symbol represents a character or symbol bigram.
-  struct Symbol {
-    const Symbol *left;              // left symbol in bigram
-    const Symbol *right;             // right symbol in bigram
-    string_util::UnicodeText chars;  // all flattend chracter sequence
-    bool is_unk;                     // true if this symbol is unknown.
-    uint64_t fp;                     // fingerprint of this symbol.
-    uint64_t freq;                   // frequency of this symbol.
-
-    // Position list. Use set so that we can keep the order of occurrence.
-    // See EncodePos/DecodePos.
-    absl::btree_set<uint64_t> positions;
-
-    bool IsBigram() const { return left != nullptr && right != nullptr; }
-    std::string ToString() const;
-    Symbol() : left(nullptr), right(nullptr), is_unk(false), fp(0), freq(0) {}
-  };
-
-  struct Position {
-    int sid;    // sentence id
-    int left;   // left symbol index
-    int right;  // right symbol index
-  };
-
-  // Encodes sid, left and right bigram index into uint64_t.
-  // Encoded value keeps the order of sid, left and right.
-  static uint64_t EncodePos(int sid, int l, int r) {
-    CHECK_GE(l, 0);
-    CHECK_GE(r, 0);
-    CHECK_LE(l, std::numeric_limits<uint16_t>::max());
-    CHECK_LE(r, std::numeric_limits<uint16_t>::max());
-    const uint64_t n = (static_cast<uint64_t>(sid) << 32) |
-                       (static_cast<uint64_t>(l) << 16) | r;
-    return n;
-  }
-
-  // Decodes sid, left and right bigram index from uint64_t.
-  static Position DecodePos(uint64_t n) {
-    Position p;
-    p.sid = n >> 32;
-    p.left = (n >> 16) & 0xffff;
-    p.right = n & 0xffff;
-    return p;
-  }
-
-  // Gets unary (character) symbol from the char code |c|.
-  // The return value is cached.
-  Symbol *GetCharSymbol(char32 c);
-
-  // Gets symbol pair from left/right symbols. The return value is cached.
-  Symbol *GetPairSymbol(const Symbol *left, const Symbol *right);
-
-  // Computes the frequency of |symbol| and update symbol->freq field.
-  void ComputeFreq(Symbol *symbol) const;
-
-  // Returns the valid index before symbols_[sid][index].
-  int GetNextIndex(int sid, int index) const;
-
-  // Returns the valid index after symbols_[sid][index].
-  int GetPrevIndex(int sid, int index) const;
-
-  // Makes a new bigram from [symbols_[sid][left], symbols_[sid][right]] and
-  // Adds it to symbols_cache_ and active_symbols_.
-  void AddNewPair(int sid, int left, int right);
-
-  // Resets the fequency of bigram [symbols_[sid][left] symbols_[sid][right]],
-  // if this bigram is not |best|.
-  void ResetFreq(int sid, int left, int right, const Symbol *best);
-
-  // Updates |active_symbols_| by copying the top 5% frequent symbols in
-  // symbols_cache_.
-  void UpdateActiveSymbols();
-
-  // All unique symbols. Key is a fingerprint of Symbol.
-  absl::flat_hash_map<uint64_t, Symbol *> symbols_cache_;
-
-  // Set of symbols from which we find the best symbol in each iteration.
-  absl::btree_set<Symbol *> active_symbols_;
-
-  // Stores symbols allocated in heap so that we can delete them at onece.
-  std::vector<Symbol *> allocated_;
-
-  // Sentences. symbols_[sid][index] stores a symbol in sentence_[sid][index].
-  std::vector<std::vector<Symbol *>> symbols_;
-};
-}  // namespace bpe
-}  // namespace sentencepiece
-#endif  // BPE_MODEL_TRAINER_H_
diff --git a/projects/llm_framework/include/sentencepiece/builder.h b/projects/llm_framework/include/sentencepiece/builder.h
deleted file mode 100644
index 2e844b23..00000000
--- a/projects/llm_framework/include/sentencepiece/builder.h
+++ /dev/null
@@ -1,148 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef BUILDER_H_
-#define BUILDER_H_
-
-#include <map>
-#include <string>
-#include <vector>
-
-#include "common.h"
-#include "sentencepiece_model.pb.h"
-#include "sentencepiece_processor.h"
-#include "third_party/absl/strings/string_view.h"
-
-namespace sentencepiece {
-namespace normalizer {
-
-// Builder creates a text normalization rule from user-defined string
-// to string mappings. The normalization mapping is compiled into
-// a single and compact blob index which is stored into the model proto.
-// This class also provides pre-defined rules based on Unicode NFKC.
-// https://en.wikipedia.org/wiki/Unicode_equivalence#Normalization
-class Builder {
- public:
-  Builder() = delete;
-  ~Builder() = delete;
-
-  // Basic Unicode character sequence.
-  using Chars = std::vector<char32>;
-
-  // String-to-string mapping.
-  using CharsMap = std::map<Chars, Chars>;
-
-  static util::Status CompileCharsMap(const CharsMap &chars_map,
-                                      std::string *output);
-
-  // Decompiles `blob` into `chars_map`.
-  static util::Status DecompileCharsMap(absl::string_view blob,
-                                        CharsMap *chars_map);
-
-  // Returns a pre-compiled binary index with `name`.
-  static util::Status GetPrecompiledCharsMap(absl::string_view name,
-                                             std::string *output);
-
-  // Makes a normalization mapping based on NFKC.
-  //
-  // Note that Normalizer/Builder classes do not support
-  // full NFKC normalization, since full NFKC normalization cannot
-  // be implemented with a simple longest matching string-to-string
-  // replacement. One unsupported normalization is multiple combining
-  // marks.
-  //
-  // Strings with multiple combining marks cannot correctly
-  // be normalized, because it needs to sort the combining marks
-  // with Canonical_Combining_Class (CCC).
-  // http://unicode.org/reports/tr15/#Multiple_Mark_Figure
-  //
-  // Example:
-  //  Original:    U+1E0B U+0323
-  //  Decomposed:  U+0064 U+0307 U+0323
-  //  NFKD:        U+0064 U+0323 U+0307 (Combining characters are sorted by CCC)
-  //  NFKC:        U+1E0D U+0307 (U+0064 U+0323 => U+1E0D)
-  //
-  // To support the normalization above with a longest matching, we need to
-  // enumerate all possible permutations of combining marks in advance,
-  // which is not feasible. For example, suppose the case there are three
-  // combining marks X, Y and Z, which are sorted into one canonical order
-  // Z, Y, X with NFK(D|C).  In this case, all permutations (XYZ, XZY, YXZ...)
-  // are normalized into ZYX. When we implement this normalization with
-  // a longest matching, we need to have 3! rules. XYZ=>ZYX, XZY=>ZYX..
-  // Since Unicode has more than 100 combining characters, it is not possible
-  // to expand all permutations.
-  //
-  // We will not implement the full NFKC in SentencePiece because
-  //  1) It is unusual to see decomposed Unicode characters in real text.
-  //  2) Providing a flexible, user-customizable, and self-contained
-  //     normalizer is the goal of SentencePiece.
-  //
-  // TODO(taku): Make NFC, NFD, and NFKD mapping if necessary.
-  static util::Status BuildNFKCMap(CharsMap *chars_map);
-
-  // Makes an NFKC-based mapping with NMT specific modifications around
-  // whitespaces.
-  static util::Status BuildNmtNFKCMap(CharsMap *chars_map);
-
-  // Merge Unicode case folding mapping into `chars_map`.
-  static util::Status MergeUnicodeCaseFoldMap(CharsMap *chars_map);
-
-  static util::Status MergeNmtMap(Builder::CharsMap *chars_map);
-
-  // Makes NFKC with Unicode case folding.
-  static util::Status BuildNFKC_CFMap(CharsMap *chars_map);
-
-  // Makes NMT NFKC with Unicode case folding.
-  static util::Status BuildNmtNFKC_CFMap(CharsMap *chars_map);
-
-  // Given NFKC maps, convert them to NFKD.
-  static util::Status BuildNFKDMap(CharsMap *chars_map);
-
-  // Builds NFC map.
-  static util::Status BuildNFCMap(CharsMap *chars_map);
-
-  // Builds NFD map.
-  static util::Status BuildNFDMap(CharsMap *chars_map);
-
-  // Makes NFKD with Unicode case folding.
-  static util::Status BuildNFKD_CFMap(CharsMap *chars_map);
-
-  // Makes NFKC with Unicode case folding.
-  static util::Status BuildNFC_CFMap(CharsMap *chars_map);
-
-  // Makes NFD with Unicode case folding.
-  static util::Status BuildNFD_CFMap(CharsMap *chars_map);
-
-  // Builds Chars map save in `filename`.
-  // Format:
-  // src_uchar1 src_uchar2 ... <tab> trg_uchar1 trg_uchar2...
-  // (src|trg)_ucharX must be a hex of Unicode code point.
-  static util::Status LoadCharsMap(absl::string_view filename,
-                                   CharsMap *chars_map);
-
-  // Saves Chars map to `filename` as TSV.
-  static util::Status SaveCharsMap(absl::string_view filename,
-                                   const CharsMap &chars_map);
-
-  // Removes redundant rules from `chars_map`.
-  // When char_maps have "aa" => "bb" and "a" => "b", the first
-  // rule is not necessary since the second rule can cover the first rule.
-  static util::Status RemoveRedundantMap(CharsMap *chars_map);
-
- private:
-  FRIEND_TEST(BuilderTest, RemoveRedundantMapTest);
-};
-}  // namespace normalizer
-}  // namespace sentencepiece
-#endif  // BUILDER_H_
diff --git a/projects/llm_framework/include/sentencepiece/builtin_pb/sentencepiece.pb.h b/projects/llm_framework/include/sentencepiece/builtin_pb/sentencepiece.pb.h
deleted file mode 100644
index 24ec3370..00000000
--- a/projects/llm_framework/include/sentencepiece/builtin_pb/sentencepiece.pb.h
+++ /dev/null
@@ -1,1041 +0,0 @@
-// Generated by the protocol buffer compiler.  DO NOT EDIT!
-// source: sentencepiece.proto
-
-#ifndef GOOGLE_PROTOBUF_INCLUDED_sentencepiece_2eproto
-#define GOOGLE_PROTOBUF_INCLUDED_sentencepiece_2eproto
-
-#include <limits>
-#include <string>
-
-#include <google/protobuf/port_def.inc>
-#if PROTOBUF_VERSION < 3014000
-#error This file was generated by a newer version of protoc which is
-#error incompatible with your Protocol Buffer headers. Please update
-#error your headers.
-#endif
-#if 3014000 < PROTOBUF_MIN_PROTOC_VERSION
-#error This file was generated by an older version of protoc which is
-#error incompatible with your Protocol Buffer headers. Please
-#error regenerate this file with a newer version of protoc.
-#endif
-
-#include <google/protobuf/port_undef.inc>
-#include <google/protobuf/io/coded_stream.h>
-#include <google/protobuf/arena.h>
-#include <google/protobuf/arenastring.h>
-#include <google/protobuf/generated_message_table_driven.h>
-#include <google/protobuf/generated_message_util.h>
-#include <google/protobuf/metadata_lite.h>
-#include <google/protobuf/message_lite.h>
-#include <google/protobuf/repeated_field.h>  // IWYU pragma: export
-#include <google/protobuf/extension_set.h>  // IWYU pragma: export
-// @@protoc_insertion_point(includes)
-#include <google/protobuf/port_def.inc>
-#define PROTOBUF_INTERNAL_EXPORT_sentencepiece_2eproto
-PROTOBUF_NAMESPACE_OPEN
-namespace internal {
-class AnyMetadata;
-}  // namespace internal
-PROTOBUF_NAMESPACE_CLOSE
-
-// Internal implementation detail -- do not use these members.
-struct TableStruct_sentencepiece_2eproto {
-  static const ::PROTOBUF_NAMESPACE_ID::internal::ParseTableField entries[]
-    PROTOBUF_SECTION_VARIABLE(protodesc_cold);
-  static const ::PROTOBUF_NAMESPACE_ID::internal::AuxiliaryParseTableField aux[]
-    PROTOBUF_SECTION_VARIABLE(protodesc_cold);
-  static const ::PROTOBUF_NAMESPACE_ID::internal::ParseTable schema[3]
-    PROTOBUF_SECTION_VARIABLE(protodesc_cold);
-  static const ::PROTOBUF_NAMESPACE_ID::internal::FieldMetadata field_metadata[];
-  static const ::PROTOBUF_NAMESPACE_ID::internal::SerializationTable serialization_table[];
-  static const ::PROTOBUF_NAMESPACE_ID::uint32 offsets[];
-};
-namespace sentencepiece {
-class NBestSentencePieceText;
-class NBestSentencePieceTextDefaultTypeInternal;
-extern NBestSentencePieceTextDefaultTypeInternal _NBestSentencePieceText_default_instance_;
-class SentencePieceText;
-class SentencePieceTextDefaultTypeInternal;
-extern SentencePieceTextDefaultTypeInternal _SentencePieceText_default_instance_;
-class SentencePieceText_SentencePiece;
-class SentencePieceText_SentencePieceDefaultTypeInternal;
-extern SentencePieceText_SentencePieceDefaultTypeInternal _SentencePieceText_SentencePiece_default_instance_;
-}  // namespace sentencepiece
-PROTOBUF_NAMESPACE_OPEN
-template<> ::sentencepiece::NBestSentencePieceText* Arena::CreateMaybeMessage<::sentencepiece::NBestSentencePieceText>(Arena*);
-template<> ::sentencepiece::SentencePieceText* Arena::CreateMaybeMessage<::sentencepiece::SentencePieceText>(Arena*);
-template<> ::sentencepiece::SentencePieceText_SentencePiece* Arena::CreateMaybeMessage<::sentencepiece::SentencePieceText_SentencePiece>(Arena*);
-PROTOBUF_NAMESPACE_CLOSE
-namespace sentencepiece {
-
-// ===================================================================
-
-class SentencePieceText_SentencePiece PROTOBUF_FINAL :
-    public ::PROTOBUF_NAMESPACE_ID::MessageLite /* @@protoc_insertion_point(class_definition:sentencepiece.SentencePieceText.SentencePiece) */ {
- public:
-  inline SentencePieceText_SentencePiece() : SentencePieceText_SentencePiece(nullptr) {}
-  virtual ~SentencePieceText_SentencePiece();
-
-  SentencePieceText_SentencePiece(const SentencePieceText_SentencePiece& from);
-  SentencePieceText_SentencePiece(SentencePieceText_SentencePiece&& from) noexcept
-    : SentencePieceText_SentencePiece() {
-    *this = ::std::move(from);
-  }
-
-  inline SentencePieceText_SentencePiece& operator=(const SentencePieceText_SentencePiece& from) {
-    CopyFrom(from);
-    return *this;
-  }
-  inline SentencePieceText_SentencePiece& operator=(SentencePieceText_SentencePiece&& from) noexcept {
-    if (GetArena() == from.GetArena()) {
-      if (this != &from) InternalSwap(&from);
-    } else {
-      CopyFrom(from);
-    }
-    return *this;
-  }
-
-  inline const std::string& unknown_fields() const {
-    return _internal_metadata_.unknown_fields<std::string>(::PROTOBUF_NAMESPACE_ID::internal::GetEmptyString);
-  }
-  inline std::string* mutable_unknown_fields() {
-    return _internal_metadata_.mutable_unknown_fields<std::string>();
-  }
-
-  static const SentencePieceText_SentencePiece& default_instance();
-
-  static inline const SentencePieceText_SentencePiece* internal_default_instance() {
-    return reinterpret_cast<const SentencePieceText_SentencePiece*>(
-               &_SentencePieceText_SentencePiece_default_instance_);
-  }
-  static constexpr int kIndexInFileMessages =
-    0;
-
-  friend void swap(SentencePieceText_SentencePiece& a, SentencePieceText_SentencePiece& b) {
-    a.Swap(&b);
-  }
-  inline void Swap(SentencePieceText_SentencePiece* other) {
-    if (other == this) return;
-    if (GetArena() == other->GetArena()) {
-      InternalSwap(other);
-    } else {
-      ::PROTOBUF_NAMESPACE_ID::internal::GenericSwap(this, other);
-    }
-  }
-  void UnsafeArenaSwap(SentencePieceText_SentencePiece* other) {
-    if (other == this) return;
-    GOOGLE_DCHECK(GetArena() == other->GetArena());
-    InternalSwap(other);
-  }
-
-  // implements Message ----------------------------------------------
-
-  inline SentencePieceText_SentencePiece* New() const final {
-    return CreateMaybeMessage<SentencePieceText_SentencePiece>(nullptr);
-  }
-
-  SentencePieceText_SentencePiece* New(::PROTOBUF_NAMESPACE_ID::Arena* arena) const final {
-    return CreateMaybeMessage<SentencePieceText_SentencePiece>(arena);
-  }
-  void CheckTypeAndMergeFrom(const ::PROTOBUF_NAMESPACE_ID::MessageLite& from)
-    final;
-  void CopyFrom(const SentencePieceText_SentencePiece& from);
-  void MergeFrom(const SentencePieceText_SentencePiece& from);
-  PROTOBUF_ATTRIBUTE_REINITIALIZES void Clear() final;
-  bool IsInitialized() const final;
-
-  size_t ByteSizeLong() const final;
-  const char* _InternalParse(const char* ptr, ::PROTOBUF_NAMESPACE_ID::internal::ParseContext* ctx) final;
-  ::PROTOBUF_NAMESPACE_ID::uint8* _InternalSerialize(
-      ::PROTOBUF_NAMESPACE_ID::uint8* target, ::PROTOBUF_NAMESPACE_ID::io::EpsCopyOutputStream* stream) const final;
-  void DiscardUnknownFields();
-  int GetCachedSize() const final { return _cached_size_.Get(); }
-
-  private:
-  inline void SharedCtor();
-  inline void SharedDtor();
-  void SetCachedSize(int size) const;
-  void InternalSwap(SentencePieceText_SentencePiece* other);
-  friend class ::PROTOBUF_NAMESPACE_ID::internal::AnyMetadata;
-  static ::PROTOBUF_NAMESPACE_ID::StringPiece FullMessageName() {
-    return "sentencepiece.SentencePieceText.SentencePiece";
-  }
-  protected:
-  explicit SentencePieceText_SentencePiece(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  private:
-  static void ArenaDtor(void* object);
-  inline void RegisterArenaDtor(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  public:
-
-  std::string GetTypeName() const final;
-
-  // nested types ----------------------------------------------------
-
-  // accessors -------------------------------------------------------
-
-  enum : int {
-    kPieceFieldNumber = 1,
-    kSurfaceFieldNumber = 3,
-    kIdFieldNumber = 2,
-    kBeginFieldNumber = 4,
-    kEndFieldNumber = 5,
-  };
-  // optional string piece = 1;
-  bool has_piece() const;
-  private:
-  bool _internal_has_piece() const;
-  public:
-  void clear_piece();
-  const std::string& piece() const;
-  void set_piece(const std::string& value);
-  void set_piece(std::string&& value);
-  void set_piece(const char* value);
-  void set_piece(const char* value, size_t size);
-  std::string* mutable_piece();
-  std::string* release_piece();
-  void set_allocated_piece(std::string* piece);
-  private:
-  const std::string& _internal_piece() const;
-  void _internal_set_piece(const std::string& value);
-  std::string* _internal_mutable_piece();
-  public:
-
-  // optional string surface = 3;
-  bool has_surface() const;
-  private:
-  bool _internal_has_surface() const;
-  public:
-  void clear_surface();
-  const std::string& surface() const;
-  void set_surface(const std::string& value);
-  void set_surface(std::string&& value);
-  void set_surface(const char* value);
-  void set_surface(const char* value, size_t size);
-  std::string* mutable_surface();
-  std::string* release_surface();
-  void set_allocated_surface(std::string* surface);
-  private:
-  const std::string& _internal_surface() const;
-  void _internal_set_surface(const std::string& value);
-  std::string* _internal_mutable_surface();
-  public:
-
-  // optional uint32 id = 2;
-  bool has_id() const;
-  private:
-  bool _internal_has_id() const;
-  public:
-  void clear_id();
-  ::PROTOBUF_NAMESPACE_ID::uint32 id() const;
-  void set_id(::PROTOBUF_NAMESPACE_ID::uint32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::uint32 _internal_id() const;
-  void _internal_set_id(::PROTOBUF_NAMESPACE_ID::uint32 value);
-  public:
-
-  // optional uint32 begin = 4;
-  bool has_begin() const;
-  private:
-  bool _internal_has_begin() const;
-  public:
-  void clear_begin();
-  ::PROTOBUF_NAMESPACE_ID::uint32 begin() const;
-  void set_begin(::PROTOBUF_NAMESPACE_ID::uint32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::uint32 _internal_begin() const;
-  void _internal_set_begin(::PROTOBUF_NAMESPACE_ID::uint32 value);
-  public:
-
-  // optional uint32 end = 5;
-  bool has_end() const;
-  private:
-  bool _internal_has_end() const;
-  public:
-  void clear_end();
-  ::PROTOBUF_NAMESPACE_ID::uint32 end() const;
-  void set_end(::PROTOBUF_NAMESPACE_ID::uint32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::uint32 _internal_end() const;
-  void _internal_set_end(::PROTOBUF_NAMESPACE_ID::uint32 value);
-  public:
-
-  GOOGLE_PROTOBUF_EXTENSION_ACCESSORS(SentencePieceText_SentencePiece)
-  // @@protoc_insertion_point(class_scope:sentencepiece.SentencePieceText.SentencePiece)
- private:
-  class _Internal;
-
-  ::PROTOBUF_NAMESPACE_ID::internal::ExtensionSet _extensions_;
-
-  template <typename T> friend class ::PROTOBUF_NAMESPACE_ID::Arena::InternalHelper;
-  typedef void InternalArenaConstructable_;
-  typedef void DestructorSkippable_;
-  ::PROTOBUF_NAMESPACE_ID::internal::HasBits<1> _has_bits_;
-  mutable ::PROTOBUF_NAMESPACE_ID::internal::CachedSize _cached_size_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr piece_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr surface_;
-  ::PROTOBUF_NAMESPACE_ID::uint32 id_;
-  ::PROTOBUF_NAMESPACE_ID::uint32 begin_;
-  ::PROTOBUF_NAMESPACE_ID::uint32 end_;
-  friend struct ::TableStruct_sentencepiece_2eproto;
-};
-// -------------------------------------------------------------------
-
-class SentencePieceText PROTOBUF_FINAL :
-    public ::PROTOBUF_NAMESPACE_ID::MessageLite /* @@protoc_insertion_point(class_definition:sentencepiece.SentencePieceText) */ {
- public:
-  inline SentencePieceText() : SentencePieceText(nullptr) {}
-  virtual ~SentencePieceText();
-
-  SentencePieceText(const SentencePieceText& from);
-  SentencePieceText(SentencePieceText&& from) noexcept
-    : SentencePieceText() {
-    *this = ::std::move(from);
-  }
-
-  inline SentencePieceText& operator=(const SentencePieceText& from) {
-    CopyFrom(from);
-    return *this;
-  }
-  inline SentencePieceText& operator=(SentencePieceText&& from) noexcept {
-    if (GetArena() == from.GetArena()) {
-      if (this != &from) InternalSwap(&from);
-    } else {
-      CopyFrom(from);
-    }
-    return *this;
-  }
-
-  inline const std::string& unknown_fields() const {
-    return _internal_metadata_.unknown_fields<std::string>(::PROTOBUF_NAMESPACE_ID::internal::GetEmptyString);
-  }
-  inline std::string* mutable_unknown_fields() {
-    return _internal_metadata_.mutable_unknown_fields<std::string>();
-  }
-
-  static const SentencePieceText& default_instance();
-
-  static inline const SentencePieceText* internal_default_instance() {
-    return reinterpret_cast<const SentencePieceText*>(
-               &_SentencePieceText_default_instance_);
-  }
-  static constexpr int kIndexInFileMessages =
-    1;
-
-  friend void swap(SentencePieceText& a, SentencePieceText& b) {
-    a.Swap(&b);
-  }
-  inline void Swap(SentencePieceText* other) {
-    if (other == this) return;
-    if (GetArena() == other->GetArena()) {
-      InternalSwap(other);
-    } else {
-      ::PROTOBUF_NAMESPACE_ID::internal::GenericSwap(this, other);
-    }
-  }
-  void UnsafeArenaSwap(SentencePieceText* other) {
-    if (other == this) return;
-    GOOGLE_DCHECK(GetArena() == other->GetArena());
-    InternalSwap(other);
-  }
-
-  // implements Message ----------------------------------------------
-
-  inline SentencePieceText* New() const final {
-    return CreateMaybeMessage<SentencePieceText>(nullptr);
-  }
-
-  SentencePieceText* New(::PROTOBUF_NAMESPACE_ID::Arena* arena) const final {
-    return CreateMaybeMessage<SentencePieceText>(arena);
-  }
-  void CheckTypeAndMergeFrom(const ::PROTOBUF_NAMESPACE_ID::MessageLite& from)
-    final;
-  void CopyFrom(const SentencePieceText& from);
-  void MergeFrom(const SentencePieceText& from);
-  PROTOBUF_ATTRIBUTE_REINITIALIZES void Clear() final;
-  bool IsInitialized() const final;
-
-  size_t ByteSizeLong() const final;
-  const char* _InternalParse(const char* ptr, ::PROTOBUF_NAMESPACE_ID::internal::ParseContext* ctx) final;
-  ::PROTOBUF_NAMESPACE_ID::uint8* _InternalSerialize(
-      ::PROTOBUF_NAMESPACE_ID::uint8* target, ::PROTOBUF_NAMESPACE_ID::io::EpsCopyOutputStream* stream) const final;
-  void DiscardUnknownFields();
-  int GetCachedSize() const final { return _cached_size_.Get(); }
-
-  private:
-  inline void SharedCtor();
-  inline void SharedDtor();
-  void SetCachedSize(int size) const;
-  void InternalSwap(SentencePieceText* other);
-  friend class ::PROTOBUF_NAMESPACE_ID::internal::AnyMetadata;
-  static ::PROTOBUF_NAMESPACE_ID::StringPiece FullMessageName() {
-    return "sentencepiece.SentencePieceText";
-  }
-  protected:
-  explicit SentencePieceText(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  private:
-  static void ArenaDtor(void* object);
-  inline void RegisterArenaDtor(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  public:
-
-  std::string GetTypeName() const final;
-
-  // nested types ----------------------------------------------------
-
-  typedef SentencePieceText_SentencePiece SentencePiece;
-
-  // accessors -------------------------------------------------------
-
-  enum : int {
-    kPiecesFieldNumber = 2,
-    kTextFieldNumber = 1,
-    kScoreFieldNumber = 3,
-  };
-  // repeated .sentencepiece.SentencePieceText.SentencePiece pieces = 2;
-  int pieces_size() const;
-  private:
-  int _internal_pieces_size() const;
-  public:
-  void clear_pieces();
-  ::sentencepiece::SentencePieceText_SentencePiece* mutable_pieces(int index);
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::SentencePieceText_SentencePiece >*
-      mutable_pieces();
-  private:
-  const ::sentencepiece::SentencePieceText_SentencePiece& _internal_pieces(int index) const;
-  ::sentencepiece::SentencePieceText_SentencePiece* _internal_add_pieces();
-  public:
-  const ::sentencepiece::SentencePieceText_SentencePiece& pieces(int index) const;
-  ::sentencepiece::SentencePieceText_SentencePiece* add_pieces();
-  const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::SentencePieceText_SentencePiece >&
-      pieces() const;
-
-  // optional string text = 1;
-  bool has_text() const;
-  private:
-  bool _internal_has_text() const;
-  public:
-  void clear_text();
-  const std::string& text() const;
-  void set_text(const std::string& value);
-  void set_text(std::string&& value);
-  void set_text(const char* value);
-  void set_text(const char* value, size_t size);
-  std::string* mutable_text();
-  std::string* release_text();
-  void set_allocated_text(std::string* text);
-  private:
-  const std::string& _internal_text() const;
-  void _internal_set_text(const std::string& value);
-  std::string* _internal_mutable_text();
-  public:
-
-  // optional float score = 3;
-  bool has_score() const;
-  private:
-  bool _internal_has_score() const;
-  public:
-  void clear_score();
-  float score() const;
-  void set_score(float value);
-  private:
-  float _internal_score() const;
-  void _internal_set_score(float value);
-  public:
-
-  GOOGLE_PROTOBUF_EXTENSION_ACCESSORS(SentencePieceText)
-  // @@protoc_insertion_point(class_scope:sentencepiece.SentencePieceText)
- private:
-  class _Internal;
-
-  ::PROTOBUF_NAMESPACE_ID::internal::ExtensionSet _extensions_;
-
-  template <typename T> friend class ::PROTOBUF_NAMESPACE_ID::Arena::InternalHelper;
-  typedef void InternalArenaConstructable_;
-  typedef void DestructorSkippable_;
-  ::PROTOBUF_NAMESPACE_ID::internal::HasBits<1> _has_bits_;
-  mutable ::PROTOBUF_NAMESPACE_ID::internal::CachedSize _cached_size_;
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::SentencePieceText_SentencePiece > pieces_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr text_;
-  float score_;
-  friend struct ::TableStruct_sentencepiece_2eproto;
-};
-// -------------------------------------------------------------------
-
-class NBestSentencePieceText PROTOBUF_FINAL :
-    public ::PROTOBUF_NAMESPACE_ID::MessageLite /* @@protoc_insertion_point(class_definition:sentencepiece.NBestSentencePieceText) */ {
- public:
-  inline NBestSentencePieceText() : NBestSentencePieceText(nullptr) {}
-  virtual ~NBestSentencePieceText();
-
-  NBestSentencePieceText(const NBestSentencePieceText& from);
-  NBestSentencePieceText(NBestSentencePieceText&& from) noexcept
-    : NBestSentencePieceText() {
-    *this = ::std::move(from);
-  }
-
-  inline NBestSentencePieceText& operator=(const NBestSentencePieceText& from) {
-    CopyFrom(from);
-    return *this;
-  }
-  inline NBestSentencePieceText& operator=(NBestSentencePieceText&& from) noexcept {
-    if (GetArena() == from.GetArena()) {
-      if (this != &from) InternalSwap(&from);
-    } else {
-      CopyFrom(from);
-    }
-    return *this;
-  }
-
-  inline const std::string& unknown_fields() const {
-    return _internal_metadata_.unknown_fields<std::string>(::PROTOBUF_NAMESPACE_ID::internal::GetEmptyString);
-  }
-  inline std::string* mutable_unknown_fields() {
-    return _internal_metadata_.mutable_unknown_fields<std::string>();
-  }
-
-  static const NBestSentencePieceText& default_instance();
-
-  static inline const NBestSentencePieceText* internal_default_instance() {
-    return reinterpret_cast<const NBestSentencePieceText*>(
-               &_NBestSentencePieceText_default_instance_);
-  }
-  static constexpr int kIndexInFileMessages =
-    2;
-
-  friend void swap(NBestSentencePieceText& a, NBestSentencePieceText& b) {
-    a.Swap(&b);
-  }
-  inline void Swap(NBestSentencePieceText* other) {
-    if (other == this) return;
-    if (GetArena() == other->GetArena()) {
-      InternalSwap(other);
-    } else {
-      ::PROTOBUF_NAMESPACE_ID::internal::GenericSwap(this, other);
-    }
-  }
-  void UnsafeArenaSwap(NBestSentencePieceText* other) {
-    if (other == this) return;
-    GOOGLE_DCHECK(GetArena() == other->GetArena());
-    InternalSwap(other);
-  }
-
-  // implements Message ----------------------------------------------
-
-  inline NBestSentencePieceText* New() const final {
-    return CreateMaybeMessage<NBestSentencePieceText>(nullptr);
-  }
-
-  NBestSentencePieceText* New(::PROTOBUF_NAMESPACE_ID::Arena* arena) const final {
-    return CreateMaybeMessage<NBestSentencePieceText>(arena);
-  }
-  void CheckTypeAndMergeFrom(const ::PROTOBUF_NAMESPACE_ID::MessageLite& from)
-    final;
-  void CopyFrom(const NBestSentencePieceText& from);
-  void MergeFrom(const NBestSentencePieceText& from);
-  PROTOBUF_ATTRIBUTE_REINITIALIZES void Clear() final;
-  bool IsInitialized() const final;
-
-  size_t ByteSizeLong() const final;
-  const char* _InternalParse(const char* ptr, ::PROTOBUF_NAMESPACE_ID::internal::ParseContext* ctx) final;
-  ::PROTOBUF_NAMESPACE_ID::uint8* _InternalSerialize(
-      ::PROTOBUF_NAMESPACE_ID::uint8* target, ::PROTOBUF_NAMESPACE_ID::io::EpsCopyOutputStream* stream) const final;
-  void DiscardUnknownFields();
-  int GetCachedSize() const final { return _cached_size_.Get(); }
-
-  private:
-  inline void SharedCtor();
-  inline void SharedDtor();
-  void SetCachedSize(int size) const;
-  void InternalSwap(NBestSentencePieceText* other);
-  friend class ::PROTOBUF_NAMESPACE_ID::internal::AnyMetadata;
-  static ::PROTOBUF_NAMESPACE_ID::StringPiece FullMessageName() {
-    return "sentencepiece.NBestSentencePieceText";
-  }
-  protected:
-  explicit NBestSentencePieceText(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  private:
-  static void ArenaDtor(void* object);
-  inline void RegisterArenaDtor(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  public:
-
-  std::string GetTypeName() const final;
-
-  // nested types ----------------------------------------------------
-
-  // accessors -------------------------------------------------------
-
-  enum : int {
-    kNbestsFieldNumber = 1,
-  };
-  // repeated .sentencepiece.SentencePieceText nbests = 1;
-  int nbests_size() const;
-  private:
-  int _internal_nbests_size() const;
-  public:
-  void clear_nbests();
-  ::sentencepiece::SentencePieceText* mutable_nbests(int index);
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::SentencePieceText >*
-      mutable_nbests();
-  private:
-  const ::sentencepiece::SentencePieceText& _internal_nbests(int index) const;
-  ::sentencepiece::SentencePieceText* _internal_add_nbests();
-  public:
-  const ::sentencepiece::SentencePieceText& nbests(int index) const;
-  ::sentencepiece::SentencePieceText* add_nbests();
-  const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::SentencePieceText >&
-      nbests() const;
-
-  // @@protoc_insertion_point(class_scope:sentencepiece.NBestSentencePieceText)
- private:
-  class _Internal;
-
-  template <typename T> friend class ::PROTOBUF_NAMESPACE_ID::Arena::InternalHelper;
-  typedef void InternalArenaConstructable_;
-  typedef void DestructorSkippable_;
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::SentencePieceText > nbests_;
-  mutable ::PROTOBUF_NAMESPACE_ID::internal::CachedSize _cached_size_;
-  friend struct ::TableStruct_sentencepiece_2eproto;
-};
-// ===================================================================
-
-
-// ===================================================================
-
-#ifdef __GNUC__
-  #pragma GCC diagnostic push
-  #pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#endif  // __GNUC__
-// SentencePieceText_SentencePiece
-
-// optional string piece = 1;
-inline bool SentencePieceText_SentencePiece::_internal_has_piece() const {
-  bool value = (_has_bits_[0] & 0x00000001u) != 0;
-  return value;
-}
-inline bool SentencePieceText_SentencePiece::has_piece() const {
-  return _internal_has_piece();
-}
-inline void SentencePieceText_SentencePiece::clear_piece() {
-  piece_.ClearToEmpty();
-  _has_bits_[0] &= ~0x00000001u;
-}
-inline const std::string& SentencePieceText_SentencePiece::piece() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.SentencePieceText.SentencePiece.piece)
-  return _internal_piece();
-}
-inline void SentencePieceText_SentencePiece::set_piece(const std::string& value) {
-  _internal_set_piece(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.SentencePieceText.SentencePiece.piece)
-}
-inline std::string* SentencePieceText_SentencePiece::mutable_piece() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.SentencePieceText.SentencePiece.piece)
-  return _internal_mutable_piece();
-}
-inline const std::string& SentencePieceText_SentencePiece::_internal_piece() const {
-  return piece_.Get();
-}
-inline void SentencePieceText_SentencePiece::_internal_set_piece(const std::string& value) {
-  _has_bits_[0] |= 0x00000001u;
-  piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, value, GetArena());
-}
-inline void SentencePieceText_SentencePiece::set_piece(std::string&& value) {
-  _has_bits_[0] |= 0x00000001u;
-  piece_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.SentencePieceText.SentencePiece.piece)
-}
-inline void SentencePieceText_SentencePiece::set_piece(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000001u;
-  piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.SentencePieceText.SentencePiece.piece)
-}
-inline void SentencePieceText_SentencePiece::set_piece(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000001u;
-  piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.SentencePieceText.SentencePiece.piece)
-}
-inline std::string* SentencePieceText_SentencePiece::_internal_mutable_piece() {
-  _has_bits_[0] |= 0x00000001u;
-  return piece_.Mutable(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, GetArena());
-}
-inline std::string* SentencePieceText_SentencePiece::release_piece() {
-  // @@protoc_insertion_point(field_release:sentencepiece.SentencePieceText.SentencePiece.piece)
-  if (!_internal_has_piece()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000001u;
-  return piece_.ReleaseNonDefault(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), GetArena());
-}
-inline void SentencePieceText_SentencePiece::set_allocated_piece(std::string* piece) {
-  if (piece != nullptr) {
-    _has_bits_[0] |= 0x00000001u;
-  } else {
-    _has_bits_[0] &= ~0x00000001u;
-  }
-  piece_.SetAllocated(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), piece,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.SentencePieceText.SentencePiece.piece)
-}
-
-// optional uint32 id = 2;
-inline bool SentencePieceText_SentencePiece::_internal_has_id() const {
-  bool value = (_has_bits_[0] & 0x00000004u) != 0;
-  return value;
-}
-inline bool SentencePieceText_SentencePiece::has_id() const {
-  return _internal_has_id();
-}
-inline void SentencePieceText_SentencePiece::clear_id() {
-  id_ = 0u;
-  _has_bits_[0] &= ~0x00000004u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::uint32 SentencePieceText_SentencePiece::_internal_id() const {
-  return id_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::uint32 SentencePieceText_SentencePiece::id() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.SentencePieceText.SentencePiece.id)
-  return _internal_id();
-}
-inline void SentencePieceText_SentencePiece::_internal_set_id(::PROTOBUF_NAMESPACE_ID::uint32 value) {
-  _has_bits_[0] |= 0x00000004u;
-  id_ = value;
-}
-inline void SentencePieceText_SentencePiece::set_id(::PROTOBUF_NAMESPACE_ID::uint32 value) {
-  _internal_set_id(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.SentencePieceText.SentencePiece.id)
-}
-
-// optional string surface = 3;
-inline bool SentencePieceText_SentencePiece::_internal_has_surface() const {
-  bool value = (_has_bits_[0] & 0x00000002u) != 0;
-  return value;
-}
-inline bool SentencePieceText_SentencePiece::has_surface() const {
-  return _internal_has_surface();
-}
-inline void SentencePieceText_SentencePiece::clear_surface() {
-  surface_.ClearToEmpty();
-  _has_bits_[0] &= ~0x00000002u;
-}
-inline const std::string& SentencePieceText_SentencePiece::surface() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.SentencePieceText.SentencePiece.surface)
-  return _internal_surface();
-}
-inline void SentencePieceText_SentencePiece::set_surface(const std::string& value) {
-  _internal_set_surface(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.SentencePieceText.SentencePiece.surface)
-}
-inline std::string* SentencePieceText_SentencePiece::mutable_surface() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.SentencePieceText.SentencePiece.surface)
-  return _internal_mutable_surface();
-}
-inline const std::string& SentencePieceText_SentencePiece::_internal_surface() const {
-  return surface_.Get();
-}
-inline void SentencePieceText_SentencePiece::_internal_set_surface(const std::string& value) {
-  _has_bits_[0] |= 0x00000002u;
-  surface_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, value, GetArena());
-}
-inline void SentencePieceText_SentencePiece::set_surface(std::string&& value) {
-  _has_bits_[0] |= 0x00000002u;
-  surface_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.SentencePieceText.SentencePiece.surface)
-}
-inline void SentencePieceText_SentencePiece::set_surface(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000002u;
-  surface_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.SentencePieceText.SentencePiece.surface)
-}
-inline void SentencePieceText_SentencePiece::set_surface(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000002u;
-  surface_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.SentencePieceText.SentencePiece.surface)
-}
-inline std::string* SentencePieceText_SentencePiece::_internal_mutable_surface() {
-  _has_bits_[0] |= 0x00000002u;
-  return surface_.Mutable(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, GetArena());
-}
-inline std::string* SentencePieceText_SentencePiece::release_surface() {
-  // @@protoc_insertion_point(field_release:sentencepiece.SentencePieceText.SentencePiece.surface)
-  if (!_internal_has_surface()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000002u;
-  return surface_.ReleaseNonDefault(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), GetArena());
-}
-inline void SentencePieceText_SentencePiece::set_allocated_surface(std::string* surface) {
-  if (surface != nullptr) {
-    _has_bits_[0] |= 0x00000002u;
-  } else {
-    _has_bits_[0] &= ~0x00000002u;
-  }
-  surface_.SetAllocated(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), surface,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.SentencePieceText.SentencePiece.surface)
-}
-
-// optional uint32 begin = 4;
-inline bool SentencePieceText_SentencePiece::_internal_has_begin() const {
-  bool value = (_has_bits_[0] & 0x00000008u) != 0;
-  return value;
-}
-inline bool SentencePieceText_SentencePiece::has_begin() const {
-  return _internal_has_begin();
-}
-inline void SentencePieceText_SentencePiece::clear_begin() {
-  begin_ = 0u;
-  _has_bits_[0] &= ~0x00000008u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::uint32 SentencePieceText_SentencePiece::_internal_begin() const {
-  return begin_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::uint32 SentencePieceText_SentencePiece::begin() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.SentencePieceText.SentencePiece.begin)
-  return _internal_begin();
-}
-inline void SentencePieceText_SentencePiece::_internal_set_begin(::PROTOBUF_NAMESPACE_ID::uint32 value) {
-  _has_bits_[0] |= 0x00000008u;
-  begin_ = value;
-}
-inline void SentencePieceText_SentencePiece::set_begin(::PROTOBUF_NAMESPACE_ID::uint32 value) {
-  _internal_set_begin(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.SentencePieceText.SentencePiece.begin)
-}
-
-// optional uint32 end = 5;
-inline bool SentencePieceText_SentencePiece::_internal_has_end() const {
-  bool value = (_has_bits_[0] & 0x00000010u) != 0;
-  return value;
-}
-inline bool SentencePieceText_SentencePiece::has_end() const {
-  return _internal_has_end();
-}
-inline void SentencePieceText_SentencePiece::clear_end() {
-  end_ = 0u;
-  _has_bits_[0] &= ~0x00000010u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::uint32 SentencePieceText_SentencePiece::_internal_end() const {
-  return end_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::uint32 SentencePieceText_SentencePiece::end() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.SentencePieceText.SentencePiece.end)
-  return _internal_end();
-}
-inline void SentencePieceText_SentencePiece::_internal_set_end(::PROTOBUF_NAMESPACE_ID::uint32 value) {
-  _has_bits_[0] |= 0x00000010u;
-  end_ = value;
-}
-inline void SentencePieceText_SentencePiece::set_end(::PROTOBUF_NAMESPACE_ID::uint32 value) {
-  _internal_set_end(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.SentencePieceText.SentencePiece.end)
-}
-
-// -------------------------------------------------------------------
-
-// SentencePieceText
-
-// optional string text = 1;
-inline bool SentencePieceText::_internal_has_text() const {
-  bool value = (_has_bits_[0] & 0x00000001u) != 0;
-  return value;
-}
-inline bool SentencePieceText::has_text() const {
-  return _internal_has_text();
-}
-inline void SentencePieceText::clear_text() {
-  text_.ClearToEmpty();
-  _has_bits_[0] &= ~0x00000001u;
-}
-inline const std::string& SentencePieceText::text() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.SentencePieceText.text)
-  return _internal_text();
-}
-inline void SentencePieceText::set_text(const std::string& value) {
-  _internal_set_text(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.SentencePieceText.text)
-}
-inline std::string* SentencePieceText::mutable_text() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.SentencePieceText.text)
-  return _internal_mutable_text();
-}
-inline const std::string& SentencePieceText::_internal_text() const {
-  return text_.Get();
-}
-inline void SentencePieceText::_internal_set_text(const std::string& value) {
-  _has_bits_[0] |= 0x00000001u;
-  text_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, value, GetArena());
-}
-inline void SentencePieceText::set_text(std::string&& value) {
-  _has_bits_[0] |= 0x00000001u;
-  text_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.SentencePieceText.text)
-}
-inline void SentencePieceText::set_text(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000001u;
-  text_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.SentencePieceText.text)
-}
-inline void SentencePieceText::set_text(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000001u;
-  text_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.SentencePieceText.text)
-}
-inline std::string* SentencePieceText::_internal_mutable_text() {
-  _has_bits_[0] |= 0x00000001u;
-  return text_.Mutable(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, GetArena());
-}
-inline std::string* SentencePieceText::release_text() {
-  // @@protoc_insertion_point(field_release:sentencepiece.SentencePieceText.text)
-  if (!_internal_has_text()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000001u;
-  return text_.ReleaseNonDefault(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), GetArena());
-}
-inline void SentencePieceText::set_allocated_text(std::string* text) {
-  if (text != nullptr) {
-    _has_bits_[0] |= 0x00000001u;
-  } else {
-    _has_bits_[0] &= ~0x00000001u;
-  }
-  text_.SetAllocated(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), text,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.SentencePieceText.text)
-}
-
-// repeated .sentencepiece.SentencePieceText.SentencePiece pieces = 2;
-inline int SentencePieceText::_internal_pieces_size() const {
-  return pieces_.size();
-}
-inline int SentencePieceText::pieces_size() const {
-  return _internal_pieces_size();
-}
-inline void SentencePieceText::clear_pieces() {
-  pieces_.Clear();
-}
-inline ::sentencepiece::SentencePieceText_SentencePiece* SentencePieceText::mutable_pieces(int index) {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.SentencePieceText.pieces)
-  return pieces_.Mutable(index);
-}
-inline ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::SentencePieceText_SentencePiece >*
-SentencePieceText::mutable_pieces() {
-  // @@protoc_insertion_point(field_mutable_list:sentencepiece.SentencePieceText.pieces)
-  return &pieces_;
-}
-inline const ::sentencepiece::SentencePieceText_SentencePiece& SentencePieceText::_internal_pieces(int index) const {
-  return pieces_.Get(index);
-}
-inline const ::sentencepiece::SentencePieceText_SentencePiece& SentencePieceText::pieces(int index) const {
-  // @@protoc_insertion_point(field_get:sentencepiece.SentencePieceText.pieces)
-  return _internal_pieces(index);
-}
-inline ::sentencepiece::SentencePieceText_SentencePiece* SentencePieceText::_internal_add_pieces() {
-  return pieces_.Add();
-}
-inline ::sentencepiece::SentencePieceText_SentencePiece* SentencePieceText::add_pieces() {
-  // @@protoc_insertion_point(field_add:sentencepiece.SentencePieceText.pieces)
-  return _internal_add_pieces();
-}
-inline const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::SentencePieceText_SentencePiece >&
-SentencePieceText::pieces() const {
-  // @@protoc_insertion_point(field_list:sentencepiece.SentencePieceText.pieces)
-  return pieces_;
-}
-
-// optional float score = 3;
-inline bool SentencePieceText::_internal_has_score() const {
-  bool value = (_has_bits_[0] & 0x00000002u) != 0;
-  return value;
-}
-inline bool SentencePieceText::has_score() const {
-  return _internal_has_score();
-}
-inline void SentencePieceText::clear_score() {
-  score_ = 0;
-  _has_bits_[0] &= ~0x00000002u;
-}
-inline float SentencePieceText::_internal_score() const {
-  return score_;
-}
-inline float SentencePieceText::score() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.SentencePieceText.score)
-  return _internal_score();
-}
-inline void SentencePieceText::_internal_set_score(float value) {
-  _has_bits_[0] |= 0x00000002u;
-  score_ = value;
-}
-inline void SentencePieceText::set_score(float value) {
-  _internal_set_score(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.SentencePieceText.score)
-}
-
-// -------------------------------------------------------------------
-
-// NBestSentencePieceText
-
-// repeated .sentencepiece.SentencePieceText nbests = 1;
-inline int NBestSentencePieceText::_internal_nbests_size() const {
-  return nbests_.size();
-}
-inline int NBestSentencePieceText::nbests_size() const {
-  return _internal_nbests_size();
-}
-inline void NBestSentencePieceText::clear_nbests() {
-  nbests_.Clear();
-}
-inline ::sentencepiece::SentencePieceText* NBestSentencePieceText::mutable_nbests(int index) {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.NBestSentencePieceText.nbests)
-  return nbests_.Mutable(index);
-}
-inline ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::SentencePieceText >*
-NBestSentencePieceText::mutable_nbests() {
-  // @@protoc_insertion_point(field_mutable_list:sentencepiece.NBestSentencePieceText.nbests)
-  return &nbests_;
-}
-inline const ::sentencepiece::SentencePieceText& NBestSentencePieceText::_internal_nbests(int index) const {
-  return nbests_.Get(index);
-}
-inline const ::sentencepiece::SentencePieceText& NBestSentencePieceText::nbests(int index) const {
-  // @@protoc_insertion_point(field_get:sentencepiece.NBestSentencePieceText.nbests)
-  return _internal_nbests(index);
-}
-inline ::sentencepiece::SentencePieceText* NBestSentencePieceText::_internal_add_nbests() {
-  return nbests_.Add();
-}
-inline ::sentencepiece::SentencePieceText* NBestSentencePieceText::add_nbests() {
-  // @@protoc_insertion_point(field_add:sentencepiece.NBestSentencePieceText.nbests)
-  return _internal_add_nbests();
-}
-inline const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::SentencePieceText >&
-NBestSentencePieceText::nbests() const {
-  // @@protoc_insertion_point(field_list:sentencepiece.NBestSentencePieceText.nbests)
-  return nbests_;
-}
-
-#ifdef __GNUC__
-  #pragma GCC diagnostic pop
-#endif  // __GNUC__
-// -------------------------------------------------------------------
-
-// -------------------------------------------------------------------
-
-
-// @@protoc_insertion_point(namespace_scope)
-
-}  // namespace sentencepiece
-
-// @@protoc_insertion_point(global_scope)
-
-#include <google/protobuf/port_undef.inc>
-#endif  // GOOGLE_PROTOBUF_INCLUDED_GOOGLE_PROTOBUF_INCLUDED_sentencepiece_2eproto
diff --git a/projects/llm_framework/include/sentencepiece/builtin_pb/sentencepiece_model.pb.h b/projects/llm_framework/include/sentencepiece/builtin_pb/sentencepiece_model.pb.h
deleted file mode 100644
index 3c738813..00000000
--- a/projects/llm_framework/include/sentencepiece/builtin_pb/sentencepiece_model.pb.h
+++ /dev/null
@@ -1,5030 +0,0 @@
-// Generated by the protocol buffer compiler.  DO NOT EDIT!
-// source: sentencepiece_model.proto
-
-#ifndef GOOGLE_PROTOBUF_INCLUDED_sentencepiece_5fmodel_2eproto
-#define GOOGLE_PROTOBUF_INCLUDED_sentencepiece_5fmodel_2eproto
-
-#include <limits>
-#include <string>
-
-#include <google/protobuf/port_def.inc>
-#if PROTOBUF_VERSION < 3014000
-#error This file was generated by a newer version of protoc which is
-#error incompatible with your Protocol Buffer headers. Please update
-#error your headers.
-#endif
-#if 3014000 < PROTOBUF_MIN_PROTOC_VERSION
-#error This file was generated by an older version of protoc which is
-#error incompatible with your Protocol Buffer headers. Please
-#error regenerate this file with a newer version of protoc.
-#endif
-
-#include <google/protobuf/port_undef.inc>
-#include <google/protobuf/io/coded_stream.h>
-#include <google/protobuf/arena.h>
-#include <google/protobuf/arenastring.h>
-#include <google/protobuf/generated_message_table_driven.h>
-#include <google/protobuf/generated_message_util.h>
-#include <google/protobuf/metadata_lite.h>
-#include <google/protobuf/message_lite.h>
-#include <google/protobuf/repeated_field.h>  // IWYU pragma: export
-#include <google/protobuf/extension_set.h>  // IWYU pragma: export
-#include <google/protobuf/generated_enum_util.h>
-// @@protoc_insertion_point(includes)
-#include <google/protobuf/port_def.inc>
-#define PROTOBUF_INTERNAL_EXPORT_sentencepiece_5fmodel_2eproto
-PROTOBUF_NAMESPACE_OPEN
-namespace internal {
-class AnyMetadata;
-}  // namespace internal
-PROTOBUF_NAMESPACE_CLOSE
-
-// Internal implementation detail -- do not use these members.
-struct TableStruct_sentencepiece_5fmodel_2eproto {
-  static const ::PROTOBUF_NAMESPACE_ID::internal::ParseTableField entries[]
-    PROTOBUF_SECTION_VARIABLE(protodesc_cold);
-  static const ::PROTOBUF_NAMESPACE_ID::internal::AuxiliaryParseTableField aux[]
-    PROTOBUF_SECTION_VARIABLE(protodesc_cold);
-  static const ::PROTOBUF_NAMESPACE_ID::internal::ParseTable schema[6]
-    PROTOBUF_SECTION_VARIABLE(protodesc_cold);
-  static const ::PROTOBUF_NAMESPACE_ID::internal::FieldMetadata field_metadata[];
-  static const ::PROTOBUF_NAMESPACE_ID::internal::SerializationTable serialization_table[];
-  static const ::PROTOBUF_NAMESPACE_ID::uint32 offsets[];
-};
-namespace sentencepiece {
-class ModelProto;
-class ModelProtoDefaultTypeInternal;
-extern ModelProtoDefaultTypeInternal _ModelProto_default_instance_;
-class ModelProto_SentencePiece;
-class ModelProto_SentencePieceDefaultTypeInternal;
-extern ModelProto_SentencePieceDefaultTypeInternal _ModelProto_SentencePiece_default_instance_;
-class NormalizerSpec;
-class NormalizerSpecDefaultTypeInternal;
-extern NormalizerSpecDefaultTypeInternal _NormalizerSpec_default_instance_;
-class SelfTestData;
-class SelfTestDataDefaultTypeInternal;
-extern SelfTestDataDefaultTypeInternal _SelfTestData_default_instance_;
-class SelfTestData_Sample;
-class SelfTestData_SampleDefaultTypeInternal;
-extern SelfTestData_SampleDefaultTypeInternal _SelfTestData_Sample_default_instance_;
-class TrainerSpec;
-class TrainerSpecDefaultTypeInternal;
-extern TrainerSpecDefaultTypeInternal _TrainerSpec_default_instance_;
-}  // namespace sentencepiece
-PROTOBUF_NAMESPACE_OPEN
-template<> ::sentencepiece::ModelProto* Arena::CreateMaybeMessage<::sentencepiece::ModelProto>(Arena*);
-template<> ::sentencepiece::ModelProto_SentencePiece* Arena::CreateMaybeMessage<::sentencepiece::ModelProto_SentencePiece>(Arena*);
-template<> ::sentencepiece::NormalizerSpec* Arena::CreateMaybeMessage<::sentencepiece::NormalizerSpec>(Arena*);
-template<> ::sentencepiece::SelfTestData* Arena::CreateMaybeMessage<::sentencepiece::SelfTestData>(Arena*);
-template<> ::sentencepiece::SelfTestData_Sample* Arena::CreateMaybeMessage<::sentencepiece::SelfTestData_Sample>(Arena*);
-template<> ::sentencepiece::TrainerSpec* Arena::CreateMaybeMessage<::sentencepiece::TrainerSpec>(Arena*);
-PROTOBUF_NAMESPACE_CLOSE
-namespace sentencepiece {
-
-enum TrainerSpec_ModelType : int {
-  TrainerSpec_ModelType_UNIGRAM = 1,
-  TrainerSpec_ModelType_BPE = 2,
-  TrainerSpec_ModelType_WORD = 3,
-  TrainerSpec_ModelType_CHAR = 4
-};
-bool TrainerSpec_ModelType_IsValid(int value);
-constexpr TrainerSpec_ModelType TrainerSpec_ModelType_ModelType_MIN = TrainerSpec_ModelType_UNIGRAM;
-constexpr TrainerSpec_ModelType TrainerSpec_ModelType_ModelType_MAX = TrainerSpec_ModelType_CHAR;
-constexpr int TrainerSpec_ModelType_ModelType_ARRAYSIZE = TrainerSpec_ModelType_ModelType_MAX + 1;
-
-const std::string& TrainerSpec_ModelType_Name(TrainerSpec_ModelType value);
-template<typename T>
-inline const std::string& TrainerSpec_ModelType_Name(T enum_t_value) {
-  static_assert(::std::is_same<T, TrainerSpec_ModelType>::value ||
-    ::std::is_integral<T>::value,
-    "Incorrect type passed to function TrainerSpec_ModelType_Name.");
-  return TrainerSpec_ModelType_Name(static_cast<TrainerSpec_ModelType>(enum_t_value));
-}
-bool TrainerSpec_ModelType_Parse(
-    ::PROTOBUF_NAMESPACE_ID::ConstStringParam name, TrainerSpec_ModelType* value);
-enum ModelProto_SentencePiece_Type : int {
-  ModelProto_SentencePiece_Type_NORMAL = 1,
-  ModelProto_SentencePiece_Type_UNKNOWN = 2,
-  ModelProto_SentencePiece_Type_CONTROL = 3,
-  ModelProto_SentencePiece_Type_USER_DEFINED = 4,
-  ModelProto_SentencePiece_Type_BYTE = 6,
-  ModelProto_SentencePiece_Type_UNUSED = 5
-};
-bool ModelProto_SentencePiece_Type_IsValid(int value);
-constexpr ModelProto_SentencePiece_Type ModelProto_SentencePiece_Type_Type_MIN = ModelProto_SentencePiece_Type_NORMAL;
-constexpr ModelProto_SentencePiece_Type ModelProto_SentencePiece_Type_Type_MAX = ModelProto_SentencePiece_Type_BYTE;
-constexpr int ModelProto_SentencePiece_Type_Type_ARRAYSIZE = ModelProto_SentencePiece_Type_Type_MAX + 1;
-
-const std::string& ModelProto_SentencePiece_Type_Name(ModelProto_SentencePiece_Type value);
-template<typename T>
-inline const std::string& ModelProto_SentencePiece_Type_Name(T enum_t_value) {
-  static_assert(::std::is_same<T, ModelProto_SentencePiece_Type>::value ||
-    ::std::is_integral<T>::value,
-    "Incorrect type passed to function ModelProto_SentencePiece_Type_Name.");
-  return ModelProto_SentencePiece_Type_Name(static_cast<ModelProto_SentencePiece_Type>(enum_t_value));
-}
-bool ModelProto_SentencePiece_Type_Parse(
-    ::PROTOBUF_NAMESPACE_ID::ConstStringParam name, ModelProto_SentencePiece_Type* value);
-// ===================================================================
-
-class TrainerSpec PROTOBUF_FINAL :
-    public ::PROTOBUF_NAMESPACE_ID::MessageLite /* @@protoc_insertion_point(class_definition:sentencepiece.TrainerSpec) */ {
- public:
-  inline TrainerSpec() : TrainerSpec(nullptr) {}
-  virtual ~TrainerSpec();
-
-  TrainerSpec(const TrainerSpec& from);
-  TrainerSpec(TrainerSpec&& from) noexcept
-    : TrainerSpec() {
-    *this = ::std::move(from);
-  }
-
-  inline TrainerSpec& operator=(const TrainerSpec& from) {
-    CopyFrom(from);
-    return *this;
-  }
-  inline TrainerSpec& operator=(TrainerSpec&& from) noexcept {
-    if (GetArena() == from.GetArena()) {
-      if (this != &from) InternalSwap(&from);
-    } else {
-      CopyFrom(from);
-    }
-    return *this;
-  }
-
-  inline const std::string& unknown_fields() const {
-    return _internal_metadata_.unknown_fields<std::string>(::PROTOBUF_NAMESPACE_ID::internal::GetEmptyString);
-  }
-  inline std::string* mutable_unknown_fields() {
-    return _internal_metadata_.mutable_unknown_fields<std::string>();
-  }
-
-  static const TrainerSpec& default_instance();
-
-  static inline const TrainerSpec* internal_default_instance() {
-    return reinterpret_cast<const TrainerSpec*>(
-               &_TrainerSpec_default_instance_);
-  }
-  static constexpr int kIndexInFileMessages =
-    0;
-
-  friend void swap(TrainerSpec& a, TrainerSpec& b) {
-    a.Swap(&b);
-  }
-  inline void Swap(TrainerSpec* other) {
-    if (other == this) return;
-    if (GetArena() == other->GetArena()) {
-      InternalSwap(other);
-    } else {
-      ::PROTOBUF_NAMESPACE_ID::internal::GenericSwap(this, other);
-    }
-  }
-  void UnsafeArenaSwap(TrainerSpec* other) {
-    if (other == this) return;
-    GOOGLE_DCHECK(GetArena() == other->GetArena());
-    InternalSwap(other);
-  }
-
-  // implements Message ----------------------------------------------
-
-  inline TrainerSpec* New() const final {
-    return CreateMaybeMessage<TrainerSpec>(nullptr);
-  }
-
-  TrainerSpec* New(::PROTOBUF_NAMESPACE_ID::Arena* arena) const final {
-    return CreateMaybeMessage<TrainerSpec>(arena);
-  }
-  void CheckTypeAndMergeFrom(const ::PROTOBUF_NAMESPACE_ID::MessageLite& from)
-    final;
-  void CopyFrom(const TrainerSpec& from);
-  void MergeFrom(const TrainerSpec& from);
-  PROTOBUF_ATTRIBUTE_REINITIALIZES void Clear() final;
-  bool IsInitialized() const final;
-
-  size_t ByteSizeLong() const final;
-  const char* _InternalParse(const char* ptr, ::PROTOBUF_NAMESPACE_ID::internal::ParseContext* ctx) final;
-  ::PROTOBUF_NAMESPACE_ID::uint8* _InternalSerialize(
-      ::PROTOBUF_NAMESPACE_ID::uint8* target, ::PROTOBUF_NAMESPACE_ID::io::EpsCopyOutputStream* stream) const final;
-  void DiscardUnknownFields();
-  int GetCachedSize() const final { return _cached_size_.Get(); }
-
-  private:
-  inline void SharedCtor();
-  inline void SharedDtor();
-  void SetCachedSize(int size) const;
-  void InternalSwap(TrainerSpec* other);
-  friend class ::PROTOBUF_NAMESPACE_ID::internal::AnyMetadata;
-  static ::PROTOBUF_NAMESPACE_ID::StringPiece FullMessageName() {
-    return "sentencepiece.TrainerSpec";
-  }
-  protected:
-  explicit TrainerSpec(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  private:
-  static void ArenaDtor(void* object);
-  inline void RegisterArenaDtor(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  public:
-
-  std::string GetTypeName() const final;
-
-  // nested types ----------------------------------------------------
-
-  typedef TrainerSpec_ModelType ModelType;
-  static constexpr ModelType UNIGRAM =
-    TrainerSpec_ModelType_UNIGRAM;
-  static constexpr ModelType BPE =
-    TrainerSpec_ModelType_BPE;
-  static constexpr ModelType WORD =
-    TrainerSpec_ModelType_WORD;
-  static constexpr ModelType CHAR =
-    TrainerSpec_ModelType_CHAR;
-  static inline bool ModelType_IsValid(int value) {
-    return TrainerSpec_ModelType_IsValid(value);
-  }
-  static constexpr ModelType ModelType_MIN =
-    TrainerSpec_ModelType_ModelType_MIN;
-  static constexpr ModelType ModelType_MAX =
-    TrainerSpec_ModelType_ModelType_MAX;
-  static constexpr int ModelType_ARRAYSIZE =
-    TrainerSpec_ModelType_ModelType_ARRAYSIZE;
-  template<typename T>
-  static inline const std::string& ModelType_Name(T enum_t_value) {
-    static_assert(::std::is_same<T, ModelType>::value ||
-      ::std::is_integral<T>::value,
-      "Incorrect type passed to function ModelType_Name.");
-    return TrainerSpec_ModelType_Name(enum_t_value);
-  }
-  static inline bool ModelType_Parse(::PROTOBUF_NAMESPACE_ID::ConstStringParam name,
-      ModelType* value) {
-    return TrainerSpec_ModelType_Parse(name, value);
-  }
-
-  // accessors -------------------------------------------------------
-
-  enum : int {
-    kInputFieldNumber = 1,
-    kAcceptLanguageFieldNumber = 5,
-    kControlSymbolsFieldNumber = 30,
-    kUserDefinedSymbolsFieldNumber = 31,
-    kModelPrefixFieldNumber = 2,
-    kInputFormatFieldNumber = 7,
-    kRequiredCharsFieldNumber = 36,
-    kUnkSurfaceFieldNumber = 44,
-    kUnkPieceFieldNumber = 45,
-    kBosPieceFieldNumber = 46,
-    kEosPieceFieldNumber = 47,
-    kPadPieceFieldNumber = 48,
-    kPretokenizationDelimiterFieldNumber = 53,
-    kSeedSentencepiecesFileFieldNumber = 54,
-    kSelfTestSampleSizeFieldNumber = 6,
-    kMiningSentenceSizeFieldNumber = 12,
-    kInputSentenceSizeFieldNumber = 11,
-    kTrainingSentenceSizeFieldNumber = 13,
-    kEnableDifferentialPrivacyFieldNumber = 50,
-    kTreatWhitespaceAsSuffixFieldNumber = 24,
-    kAllowWhitespaceOnlyPiecesFieldNumber = 26,
-    kSplitDigitsFieldNumber = 25,
-    kByteFallbackFieldNumber = 35,
-    kUseAllVocabFieldNumber = 34,
-    kTrainExtremelyLargeCorpusFieldNumber = 49,
-    kUnkIdFieldNumber = 40,
-    kDifferentialPrivacyNoiseLevelFieldNumber = 51,
-    kDifferentialPrivacyClippingThresholdFieldNumber = 52,
-    kModelTypeFieldNumber = 3,
-    kVocabSizeFieldNumber = 4,
-    kCharacterCoverageFieldNumber = 10,
-    kSeedSentencepieceSizeFieldNumber = 14,
-    kShrinkingFactorFieldNumber = 15,
-    kNumThreadsFieldNumber = 16,
-    kNumSubIterationsFieldNumber = 17,
-    kMaxSentenceLengthFieldNumber = 18,
-    kMaxSentencepieceLengthFieldNumber = 20,
-    kShuffleInputSentenceFieldNumber = 19,
-    kSplitByUnicodeScriptFieldNumber = 21,
-    kSplitByNumberFieldNumber = 23,
-    kSplitByWhitespaceFieldNumber = 22,
-    kVocabularyOutputPieceScoreFieldNumber = 32,
-    kHardVocabLimitFieldNumber = 33,
-    kBosIdFieldNumber = 41,
-    kEosIdFieldNumber = 42,
-    kPadIdFieldNumber = 43,
-  };
-  // repeated string input = 1;
-  int input_size() const;
-  private:
-  int _internal_input_size() const;
-  public:
-  void clear_input();
-  const std::string& input(int index) const;
-  std::string* mutable_input(int index);
-  void set_input(int index, const std::string& value);
-  void set_input(int index, std::string&& value);
-  void set_input(int index, const char* value);
-  void set_input(int index, const char* value, size_t size);
-  std::string* add_input();
-  void add_input(const std::string& value);
-  void add_input(std::string&& value);
-  void add_input(const char* value);
-  void add_input(const char* value, size_t size);
-  const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>& input() const;
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>* mutable_input();
-  private:
-  const std::string& _internal_input(int index) const;
-  std::string* _internal_add_input();
-  public:
-
-  // repeated string accept_language = 5;
-  int accept_language_size() const;
-  private:
-  int _internal_accept_language_size() const;
-  public:
-  void clear_accept_language();
-  const std::string& accept_language(int index) const;
-  std::string* mutable_accept_language(int index);
-  void set_accept_language(int index, const std::string& value);
-  void set_accept_language(int index, std::string&& value);
-  void set_accept_language(int index, const char* value);
-  void set_accept_language(int index, const char* value, size_t size);
-  std::string* add_accept_language();
-  void add_accept_language(const std::string& value);
-  void add_accept_language(std::string&& value);
-  void add_accept_language(const char* value);
-  void add_accept_language(const char* value, size_t size);
-  const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>& accept_language() const;
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>* mutable_accept_language();
-  private:
-  const std::string& _internal_accept_language(int index) const;
-  std::string* _internal_add_accept_language();
-  public:
-
-  // repeated string control_symbols = 30;
-  int control_symbols_size() const;
-  private:
-  int _internal_control_symbols_size() const;
-  public:
-  void clear_control_symbols();
-  const std::string& control_symbols(int index) const;
-  std::string* mutable_control_symbols(int index);
-  void set_control_symbols(int index, const std::string& value);
-  void set_control_symbols(int index, std::string&& value);
-  void set_control_symbols(int index, const char* value);
-  void set_control_symbols(int index, const char* value, size_t size);
-  std::string* add_control_symbols();
-  void add_control_symbols(const std::string& value);
-  void add_control_symbols(std::string&& value);
-  void add_control_symbols(const char* value);
-  void add_control_symbols(const char* value, size_t size);
-  const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>& control_symbols() const;
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>* mutable_control_symbols();
-  private:
-  const std::string& _internal_control_symbols(int index) const;
-  std::string* _internal_add_control_symbols();
-  public:
-
-  // repeated string user_defined_symbols = 31;
-  int user_defined_symbols_size() const;
-  private:
-  int _internal_user_defined_symbols_size() const;
-  public:
-  void clear_user_defined_symbols();
-  const std::string& user_defined_symbols(int index) const;
-  std::string* mutable_user_defined_symbols(int index);
-  void set_user_defined_symbols(int index, const std::string& value);
-  void set_user_defined_symbols(int index, std::string&& value);
-  void set_user_defined_symbols(int index, const char* value);
-  void set_user_defined_symbols(int index, const char* value, size_t size);
-  std::string* add_user_defined_symbols();
-  void add_user_defined_symbols(const std::string& value);
-  void add_user_defined_symbols(std::string&& value);
-  void add_user_defined_symbols(const char* value);
-  void add_user_defined_symbols(const char* value, size_t size);
-  const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>& user_defined_symbols() const;
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>* mutable_user_defined_symbols();
-  private:
-  const std::string& _internal_user_defined_symbols(int index) const;
-  std::string* _internal_add_user_defined_symbols();
-  public:
-
-  // optional string model_prefix = 2;
-  bool has_model_prefix() const;
-  private:
-  bool _internal_has_model_prefix() const;
-  public:
-  void clear_model_prefix();
-  const std::string& model_prefix() const;
-  void set_model_prefix(const std::string& value);
-  void set_model_prefix(std::string&& value);
-  void set_model_prefix(const char* value);
-  void set_model_prefix(const char* value, size_t size);
-  std::string* mutable_model_prefix();
-  std::string* release_model_prefix();
-  void set_allocated_model_prefix(std::string* model_prefix);
-  private:
-  const std::string& _internal_model_prefix() const;
-  void _internal_set_model_prefix(const std::string& value);
-  std::string* _internal_mutable_model_prefix();
-  public:
-
-  // optional string input_format = 7;
-  bool has_input_format() const;
-  private:
-  bool _internal_has_input_format() const;
-  public:
-  void clear_input_format();
-  const std::string& input_format() const;
-  void set_input_format(const std::string& value);
-  void set_input_format(std::string&& value);
-  void set_input_format(const char* value);
-  void set_input_format(const char* value, size_t size);
-  std::string* mutable_input_format();
-  std::string* release_input_format();
-  void set_allocated_input_format(std::string* input_format);
-  private:
-  const std::string& _internal_input_format() const;
-  void _internal_set_input_format(const std::string& value);
-  std::string* _internal_mutable_input_format();
-  public:
-
-  // optional string required_chars = 36;
-  bool has_required_chars() const;
-  private:
-  bool _internal_has_required_chars() const;
-  public:
-  void clear_required_chars();
-  const std::string& required_chars() const;
-  void set_required_chars(const std::string& value);
-  void set_required_chars(std::string&& value);
-  void set_required_chars(const char* value);
-  void set_required_chars(const char* value, size_t size);
-  std::string* mutable_required_chars();
-  std::string* release_required_chars();
-  void set_allocated_required_chars(std::string* required_chars);
-  private:
-  const std::string& _internal_required_chars() const;
-  void _internal_set_required_chars(const std::string& value);
-  std::string* _internal_mutable_required_chars();
-  public:
-
-  // optional string unk_surface = 44 [default = " \342\201\207 "];
-  bool has_unk_surface() const;
-  private:
-  bool _internal_has_unk_surface() const;
-  public:
-  void clear_unk_surface();
-  const std::string& unk_surface() const;
-  void set_unk_surface(const std::string& value);
-  void set_unk_surface(std::string&& value);
-  void set_unk_surface(const char* value);
-  void set_unk_surface(const char* value, size_t size);
-  std::string* mutable_unk_surface();
-  std::string* release_unk_surface();
-  void set_allocated_unk_surface(std::string* unk_surface);
-  private:
-  const std::string& _internal_unk_surface() const;
-  void _internal_set_unk_surface(const std::string& value);
-  std::string* _internal_mutable_unk_surface();
-  public:
-
-  // optional string unk_piece = 45 [default = "<unk>"];
-  bool has_unk_piece() const;
-  private:
-  bool _internal_has_unk_piece() const;
-  public:
-  void clear_unk_piece();
-  const std::string& unk_piece() const;
-  void set_unk_piece(const std::string& value);
-  void set_unk_piece(std::string&& value);
-  void set_unk_piece(const char* value);
-  void set_unk_piece(const char* value, size_t size);
-  std::string* mutable_unk_piece();
-  std::string* release_unk_piece();
-  void set_allocated_unk_piece(std::string* unk_piece);
-  private:
-  const std::string& _internal_unk_piece() const;
-  void _internal_set_unk_piece(const std::string& value);
-  std::string* _internal_mutable_unk_piece();
-  public:
-
-  // optional string bos_piece = 46 [default = "<s>"];
-  bool has_bos_piece() const;
-  private:
-  bool _internal_has_bos_piece() const;
-  public:
-  void clear_bos_piece();
-  const std::string& bos_piece() const;
-  void set_bos_piece(const std::string& value);
-  void set_bos_piece(std::string&& value);
-  void set_bos_piece(const char* value);
-  void set_bos_piece(const char* value, size_t size);
-  std::string* mutable_bos_piece();
-  std::string* release_bos_piece();
-  void set_allocated_bos_piece(std::string* bos_piece);
-  private:
-  const std::string& _internal_bos_piece() const;
-  void _internal_set_bos_piece(const std::string& value);
-  std::string* _internal_mutable_bos_piece();
-  public:
-
-  // optional string eos_piece = 47 [default = "</s>"];
-  bool has_eos_piece() const;
-  private:
-  bool _internal_has_eos_piece() const;
-  public:
-  void clear_eos_piece();
-  const std::string& eos_piece() const;
-  void set_eos_piece(const std::string& value);
-  void set_eos_piece(std::string&& value);
-  void set_eos_piece(const char* value);
-  void set_eos_piece(const char* value, size_t size);
-  std::string* mutable_eos_piece();
-  std::string* release_eos_piece();
-  void set_allocated_eos_piece(std::string* eos_piece);
-  private:
-  const std::string& _internal_eos_piece() const;
-  void _internal_set_eos_piece(const std::string& value);
-  std::string* _internal_mutable_eos_piece();
-  public:
-
-  // optional string pad_piece = 48 [default = "<pad>"];
-  bool has_pad_piece() const;
-  private:
-  bool _internal_has_pad_piece() const;
-  public:
-  void clear_pad_piece();
-  const std::string& pad_piece() const;
-  void set_pad_piece(const std::string& value);
-  void set_pad_piece(std::string&& value);
-  void set_pad_piece(const char* value);
-  void set_pad_piece(const char* value, size_t size);
-  std::string* mutable_pad_piece();
-  std::string* release_pad_piece();
-  void set_allocated_pad_piece(std::string* pad_piece);
-  private:
-  const std::string& _internal_pad_piece() const;
-  void _internal_set_pad_piece(const std::string& value);
-  std::string* _internal_mutable_pad_piece();
-  public:
-
-  // optional string pretokenization_delimiter = 53 [default = ""];
-  bool has_pretokenization_delimiter() const;
-  private:
-  bool _internal_has_pretokenization_delimiter() const;
-  public:
-  void clear_pretokenization_delimiter();
-  const std::string& pretokenization_delimiter() const;
-  void set_pretokenization_delimiter(const std::string& value);
-  void set_pretokenization_delimiter(std::string&& value);
-  void set_pretokenization_delimiter(const char* value);
-  void set_pretokenization_delimiter(const char* value, size_t size);
-  std::string* mutable_pretokenization_delimiter();
-  std::string* release_pretokenization_delimiter();
-  void set_allocated_pretokenization_delimiter(std::string* pretokenization_delimiter);
-  private:
-  const std::string& _internal_pretokenization_delimiter() const;
-  void _internal_set_pretokenization_delimiter(const std::string& value);
-  std::string* _internal_mutable_pretokenization_delimiter();
-  public:
-
-  // optional string seed_sentencepieces_file = 54 [default = ""];
-  bool has_seed_sentencepieces_file() const;
-  private:
-  bool _internal_has_seed_sentencepieces_file() const;
-  public:
-  void clear_seed_sentencepieces_file();
-  const std::string& seed_sentencepieces_file() const;
-  void set_seed_sentencepieces_file(const std::string& value);
-  void set_seed_sentencepieces_file(std::string&& value);
-  void set_seed_sentencepieces_file(const char* value);
-  void set_seed_sentencepieces_file(const char* value, size_t size);
-  std::string* mutable_seed_sentencepieces_file();
-  std::string* release_seed_sentencepieces_file();
-  void set_allocated_seed_sentencepieces_file(std::string* seed_sentencepieces_file);
-  private:
-  const std::string& _internal_seed_sentencepieces_file() const;
-  void _internal_set_seed_sentencepieces_file(const std::string& value);
-  std::string* _internal_mutable_seed_sentencepieces_file();
-  public:
-
-  // optional int32 self_test_sample_size = 6 [default = 0];
-  bool has_self_test_sample_size() const;
-  private:
-  bool _internal_has_self_test_sample_size() const;
-  public:
-  void clear_self_test_sample_size();
-  ::PROTOBUF_NAMESPACE_ID::int32 self_test_sample_size() const;
-  void set_self_test_sample_size(::PROTOBUF_NAMESPACE_ID::int32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::int32 _internal_self_test_sample_size() const;
-  void _internal_set_self_test_sample_size(::PROTOBUF_NAMESPACE_ID::int32 value);
-  public:
-
-  // optional int32 mining_sentence_size = 12 [deprecated = true];
-  PROTOBUF_DEPRECATED bool has_mining_sentence_size() const;
-  private:
-  bool _internal_has_mining_sentence_size() const;
-  public:
-  PROTOBUF_DEPRECATED void clear_mining_sentence_size();
-  PROTOBUF_DEPRECATED ::PROTOBUF_NAMESPACE_ID::int32 mining_sentence_size() const;
-  PROTOBUF_DEPRECATED void set_mining_sentence_size(::PROTOBUF_NAMESPACE_ID::int32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::int32 _internal_mining_sentence_size() const;
-  void _internal_set_mining_sentence_size(::PROTOBUF_NAMESPACE_ID::int32 value);
-  public:
-
-  // optional uint64 input_sentence_size = 11 [default = 0];
-  bool has_input_sentence_size() const;
-  private:
-  bool _internal_has_input_sentence_size() const;
-  public:
-  void clear_input_sentence_size();
-  ::PROTOBUF_NAMESPACE_ID::uint64 input_sentence_size() const;
-  void set_input_sentence_size(::PROTOBUF_NAMESPACE_ID::uint64 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::uint64 _internal_input_sentence_size() const;
-  void _internal_set_input_sentence_size(::PROTOBUF_NAMESPACE_ID::uint64 value);
-  public:
-
-  // optional int32 training_sentence_size = 13 [deprecated = true];
-  PROTOBUF_DEPRECATED bool has_training_sentence_size() const;
-  private:
-  bool _internal_has_training_sentence_size() const;
-  public:
-  PROTOBUF_DEPRECATED void clear_training_sentence_size();
-  PROTOBUF_DEPRECATED ::PROTOBUF_NAMESPACE_ID::int32 training_sentence_size() const;
-  PROTOBUF_DEPRECATED void set_training_sentence_size(::PROTOBUF_NAMESPACE_ID::int32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::int32 _internal_training_sentence_size() const;
-  void _internal_set_training_sentence_size(::PROTOBUF_NAMESPACE_ID::int32 value);
-  public:
-
-  // optional bool enable_differential_privacy = 50 [default = false];
-  bool has_enable_differential_privacy() const;
-  private:
-  bool _internal_has_enable_differential_privacy() const;
-  public:
-  void clear_enable_differential_privacy();
-  bool enable_differential_privacy() const;
-  void set_enable_differential_privacy(bool value);
-  private:
-  bool _internal_enable_differential_privacy() const;
-  void _internal_set_enable_differential_privacy(bool value);
-  public:
-
-  // optional bool treat_whitespace_as_suffix = 24 [default = false];
-  bool has_treat_whitespace_as_suffix() const;
-  private:
-  bool _internal_has_treat_whitespace_as_suffix() const;
-  public:
-  void clear_treat_whitespace_as_suffix();
-  bool treat_whitespace_as_suffix() const;
-  void set_treat_whitespace_as_suffix(bool value);
-  private:
-  bool _internal_treat_whitespace_as_suffix() const;
-  void _internal_set_treat_whitespace_as_suffix(bool value);
-  public:
-
-  // optional bool allow_whitespace_only_pieces = 26 [default = false];
-  bool has_allow_whitespace_only_pieces() const;
-  private:
-  bool _internal_has_allow_whitespace_only_pieces() const;
-  public:
-  void clear_allow_whitespace_only_pieces();
-  bool allow_whitespace_only_pieces() const;
-  void set_allow_whitespace_only_pieces(bool value);
-  private:
-  bool _internal_allow_whitespace_only_pieces() const;
-  void _internal_set_allow_whitespace_only_pieces(bool value);
-  public:
-
-  // optional bool split_digits = 25 [default = false];
-  bool has_split_digits() const;
-  private:
-  bool _internal_has_split_digits() const;
-  public:
-  void clear_split_digits();
-  bool split_digits() const;
-  void set_split_digits(bool value);
-  private:
-  bool _internal_split_digits() const;
-  void _internal_set_split_digits(bool value);
-  public:
-
-  // optional bool byte_fallback = 35 [default = false];
-  bool has_byte_fallback() const;
-  private:
-  bool _internal_has_byte_fallback() const;
-  public:
-  void clear_byte_fallback();
-  bool byte_fallback() const;
-  void set_byte_fallback(bool value);
-  private:
-  bool _internal_byte_fallback() const;
-  void _internal_set_byte_fallback(bool value);
-  public:
-
-  // optional bool use_all_vocab = 34 [default = false];
-  bool has_use_all_vocab() const;
-  private:
-  bool _internal_has_use_all_vocab() const;
-  public:
-  void clear_use_all_vocab();
-  bool use_all_vocab() const;
-  void set_use_all_vocab(bool value);
-  private:
-  bool _internal_use_all_vocab() const;
-  void _internal_set_use_all_vocab(bool value);
-  public:
-
-  // optional bool train_extremely_large_corpus = 49 [default = false];
-  bool has_train_extremely_large_corpus() const;
-  private:
-  bool _internal_has_train_extremely_large_corpus() const;
-  public:
-  void clear_train_extremely_large_corpus();
-  bool train_extremely_large_corpus() const;
-  void set_train_extremely_large_corpus(bool value);
-  private:
-  bool _internal_train_extremely_large_corpus() const;
-  void _internal_set_train_extremely_large_corpus(bool value);
-  public:
-
-  // optional int32 unk_id = 40 [default = 0];
-  bool has_unk_id() const;
-  private:
-  bool _internal_has_unk_id() const;
-  public:
-  void clear_unk_id();
-  ::PROTOBUF_NAMESPACE_ID::int32 unk_id() const;
-  void set_unk_id(::PROTOBUF_NAMESPACE_ID::int32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::int32 _internal_unk_id() const;
-  void _internal_set_unk_id(::PROTOBUF_NAMESPACE_ID::int32 value);
-  public:
-
-  // optional float differential_privacy_noise_level = 51 [default = 0];
-  bool has_differential_privacy_noise_level() const;
-  private:
-  bool _internal_has_differential_privacy_noise_level() const;
-  public:
-  void clear_differential_privacy_noise_level();
-  float differential_privacy_noise_level() const;
-  void set_differential_privacy_noise_level(float value);
-  private:
-  float _internal_differential_privacy_noise_level() const;
-  void _internal_set_differential_privacy_noise_level(float value);
-  public:
-
-  // optional uint64 differential_privacy_clipping_threshold = 52 [default = 0];
-  bool has_differential_privacy_clipping_threshold() const;
-  private:
-  bool _internal_has_differential_privacy_clipping_threshold() const;
-  public:
-  void clear_differential_privacy_clipping_threshold();
-  ::PROTOBUF_NAMESPACE_ID::uint64 differential_privacy_clipping_threshold() const;
-  void set_differential_privacy_clipping_threshold(::PROTOBUF_NAMESPACE_ID::uint64 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::uint64 _internal_differential_privacy_clipping_threshold() const;
-  void _internal_set_differential_privacy_clipping_threshold(::PROTOBUF_NAMESPACE_ID::uint64 value);
-  public:
-
-  // optional .sentencepiece.TrainerSpec.ModelType model_type = 3 [default = UNIGRAM];
-  bool has_model_type() const;
-  private:
-  bool _internal_has_model_type() const;
-  public:
-  void clear_model_type();
-  ::sentencepiece::TrainerSpec_ModelType model_type() const;
-  void set_model_type(::sentencepiece::TrainerSpec_ModelType value);
-  private:
-  ::sentencepiece::TrainerSpec_ModelType _internal_model_type() const;
-  void _internal_set_model_type(::sentencepiece::TrainerSpec_ModelType value);
-  public:
-
-  // optional int32 vocab_size = 4 [default = 8000];
-  bool has_vocab_size() const;
-  private:
-  bool _internal_has_vocab_size() const;
-  public:
-  void clear_vocab_size();
-  ::PROTOBUF_NAMESPACE_ID::int32 vocab_size() const;
-  void set_vocab_size(::PROTOBUF_NAMESPACE_ID::int32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::int32 _internal_vocab_size() const;
-  void _internal_set_vocab_size(::PROTOBUF_NAMESPACE_ID::int32 value);
-  public:
-
-  // optional float character_coverage = 10 [default = 0.9995];
-  bool has_character_coverage() const;
-  private:
-  bool _internal_has_character_coverage() const;
-  public:
-  void clear_character_coverage();
-  float character_coverage() const;
-  void set_character_coverage(float value);
-  private:
-  float _internal_character_coverage() const;
-  void _internal_set_character_coverage(float value);
-  public:
-
-  // optional int32 seed_sentencepiece_size = 14 [default = 1000000];
-  bool has_seed_sentencepiece_size() const;
-  private:
-  bool _internal_has_seed_sentencepiece_size() const;
-  public:
-  void clear_seed_sentencepiece_size();
-  ::PROTOBUF_NAMESPACE_ID::int32 seed_sentencepiece_size() const;
-  void set_seed_sentencepiece_size(::PROTOBUF_NAMESPACE_ID::int32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::int32 _internal_seed_sentencepiece_size() const;
-  void _internal_set_seed_sentencepiece_size(::PROTOBUF_NAMESPACE_ID::int32 value);
-  public:
-
-  // optional float shrinking_factor = 15 [default = 0.75];
-  bool has_shrinking_factor() const;
-  private:
-  bool _internal_has_shrinking_factor() const;
-  public:
-  void clear_shrinking_factor();
-  float shrinking_factor() const;
-  void set_shrinking_factor(float value);
-  private:
-  float _internal_shrinking_factor() const;
-  void _internal_set_shrinking_factor(float value);
-  public:
-
-  // optional int32 num_threads = 16 [default = 16];
-  bool has_num_threads() const;
-  private:
-  bool _internal_has_num_threads() const;
-  public:
-  void clear_num_threads();
-  ::PROTOBUF_NAMESPACE_ID::int32 num_threads() const;
-  void set_num_threads(::PROTOBUF_NAMESPACE_ID::int32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::int32 _internal_num_threads() const;
-  void _internal_set_num_threads(::PROTOBUF_NAMESPACE_ID::int32 value);
-  public:
-
-  // optional int32 num_sub_iterations = 17 [default = 2];
-  bool has_num_sub_iterations() const;
-  private:
-  bool _internal_has_num_sub_iterations() const;
-  public:
-  void clear_num_sub_iterations();
-  ::PROTOBUF_NAMESPACE_ID::int32 num_sub_iterations() const;
-  void set_num_sub_iterations(::PROTOBUF_NAMESPACE_ID::int32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::int32 _internal_num_sub_iterations() const;
-  void _internal_set_num_sub_iterations(::PROTOBUF_NAMESPACE_ID::int32 value);
-  public:
-
-  // optional int32 max_sentence_length = 18 [default = 4192];
-  bool has_max_sentence_length() const;
-  private:
-  bool _internal_has_max_sentence_length() const;
-  public:
-  void clear_max_sentence_length();
-  ::PROTOBUF_NAMESPACE_ID::int32 max_sentence_length() const;
-  void set_max_sentence_length(::PROTOBUF_NAMESPACE_ID::int32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::int32 _internal_max_sentence_length() const;
-  void _internal_set_max_sentence_length(::PROTOBUF_NAMESPACE_ID::int32 value);
-  public:
-
-  // optional int32 max_sentencepiece_length = 20 [default = 16];
-  bool has_max_sentencepiece_length() const;
-  private:
-  bool _internal_has_max_sentencepiece_length() const;
-  public:
-  void clear_max_sentencepiece_length();
-  ::PROTOBUF_NAMESPACE_ID::int32 max_sentencepiece_length() const;
-  void set_max_sentencepiece_length(::PROTOBUF_NAMESPACE_ID::int32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::int32 _internal_max_sentencepiece_length() const;
-  void _internal_set_max_sentencepiece_length(::PROTOBUF_NAMESPACE_ID::int32 value);
-  public:
-
-  // optional bool shuffle_input_sentence = 19 [default = true];
-  bool has_shuffle_input_sentence() const;
-  private:
-  bool _internal_has_shuffle_input_sentence() const;
-  public:
-  void clear_shuffle_input_sentence();
-  bool shuffle_input_sentence() const;
-  void set_shuffle_input_sentence(bool value);
-  private:
-  bool _internal_shuffle_input_sentence() const;
-  void _internal_set_shuffle_input_sentence(bool value);
-  public:
-
-  // optional bool split_by_unicode_script = 21 [default = true];
-  bool has_split_by_unicode_script() const;
-  private:
-  bool _internal_has_split_by_unicode_script() const;
-  public:
-  void clear_split_by_unicode_script();
-  bool split_by_unicode_script() const;
-  void set_split_by_unicode_script(bool value);
-  private:
-  bool _internal_split_by_unicode_script() const;
-  void _internal_set_split_by_unicode_script(bool value);
-  public:
-
-  // optional bool split_by_number = 23 [default = true];
-  bool has_split_by_number() const;
-  private:
-  bool _internal_has_split_by_number() const;
-  public:
-  void clear_split_by_number();
-  bool split_by_number() const;
-  void set_split_by_number(bool value);
-  private:
-  bool _internal_split_by_number() const;
-  void _internal_set_split_by_number(bool value);
-  public:
-
-  // optional bool split_by_whitespace = 22 [default = true];
-  bool has_split_by_whitespace() const;
-  private:
-  bool _internal_has_split_by_whitespace() const;
-  public:
-  void clear_split_by_whitespace();
-  bool split_by_whitespace() const;
-  void set_split_by_whitespace(bool value);
-  private:
-  bool _internal_split_by_whitespace() const;
-  void _internal_set_split_by_whitespace(bool value);
-  public:
-
-  // optional bool vocabulary_output_piece_score = 32 [default = true];
-  bool has_vocabulary_output_piece_score() const;
-  private:
-  bool _internal_has_vocabulary_output_piece_score() const;
-  public:
-  void clear_vocabulary_output_piece_score();
-  bool vocabulary_output_piece_score() const;
-  void set_vocabulary_output_piece_score(bool value);
-  private:
-  bool _internal_vocabulary_output_piece_score() const;
-  void _internal_set_vocabulary_output_piece_score(bool value);
-  public:
-
-  // optional bool hard_vocab_limit = 33 [default = true];
-  bool has_hard_vocab_limit() const;
-  private:
-  bool _internal_has_hard_vocab_limit() const;
-  public:
-  void clear_hard_vocab_limit();
-  bool hard_vocab_limit() const;
-  void set_hard_vocab_limit(bool value);
-  private:
-  bool _internal_hard_vocab_limit() const;
-  void _internal_set_hard_vocab_limit(bool value);
-  public:
-
-  // optional int32 bos_id = 41 [default = 1];
-  bool has_bos_id() const;
-  private:
-  bool _internal_has_bos_id() const;
-  public:
-  void clear_bos_id();
-  ::PROTOBUF_NAMESPACE_ID::int32 bos_id() const;
-  void set_bos_id(::PROTOBUF_NAMESPACE_ID::int32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::int32 _internal_bos_id() const;
-  void _internal_set_bos_id(::PROTOBUF_NAMESPACE_ID::int32 value);
-  public:
-
-  // optional int32 eos_id = 42 [default = 2];
-  bool has_eos_id() const;
-  private:
-  bool _internal_has_eos_id() const;
-  public:
-  void clear_eos_id();
-  ::PROTOBUF_NAMESPACE_ID::int32 eos_id() const;
-  void set_eos_id(::PROTOBUF_NAMESPACE_ID::int32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::int32 _internal_eos_id() const;
-  void _internal_set_eos_id(::PROTOBUF_NAMESPACE_ID::int32 value);
-  public:
-
-  // optional int32 pad_id = 43 [default = -1];
-  bool has_pad_id() const;
-  private:
-  bool _internal_has_pad_id() const;
-  public:
-  void clear_pad_id();
-  ::PROTOBUF_NAMESPACE_ID::int32 pad_id() const;
-  void set_pad_id(::PROTOBUF_NAMESPACE_ID::int32 value);
-  private:
-  ::PROTOBUF_NAMESPACE_ID::int32 _internal_pad_id() const;
-  void _internal_set_pad_id(::PROTOBUF_NAMESPACE_ID::int32 value);
-  public:
-
-  GOOGLE_PROTOBUF_EXTENSION_ACCESSORS(TrainerSpec)
-  // @@protoc_insertion_point(class_scope:sentencepiece.TrainerSpec)
- private:
-  class _Internal;
-
-  ::PROTOBUF_NAMESPACE_ID::internal::ExtensionSet _extensions_;
-
-  template <typename T> friend class ::PROTOBUF_NAMESPACE_ID::Arena::InternalHelper;
-  typedef void InternalArenaConstructable_;
-  typedef void DestructorSkippable_;
-  ::PROTOBUF_NAMESPACE_ID::internal::HasBits<2> _has_bits_;
-  mutable ::PROTOBUF_NAMESPACE_ID::internal::CachedSize _cached_size_;
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string> input_;
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string> accept_language_;
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string> control_symbols_;
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string> user_defined_symbols_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr model_prefix_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr input_format_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr required_chars_;
-  static const ::PROTOBUF_NAMESPACE_ID::internal::LazyString _i_give_permission_to_break_this_code_default_unk_surface_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr unk_surface_;
-  static const ::PROTOBUF_NAMESPACE_ID::internal::LazyString _i_give_permission_to_break_this_code_default_unk_piece_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr unk_piece_;
-  static const ::PROTOBUF_NAMESPACE_ID::internal::LazyString _i_give_permission_to_break_this_code_default_bos_piece_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr bos_piece_;
-  static const ::PROTOBUF_NAMESPACE_ID::internal::LazyString _i_give_permission_to_break_this_code_default_eos_piece_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr eos_piece_;
-  static const ::PROTOBUF_NAMESPACE_ID::internal::LazyString _i_give_permission_to_break_this_code_default_pad_piece_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr pad_piece_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr pretokenization_delimiter_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr seed_sentencepieces_file_;
-  ::PROTOBUF_NAMESPACE_ID::int32 self_test_sample_size_;
-  ::PROTOBUF_NAMESPACE_ID::int32 mining_sentence_size_;
-  ::PROTOBUF_NAMESPACE_ID::uint64 input_sentence_size_;
-  ::PROTOBUF_NAMESPACE_ID::int32 training_sentence_size_;
-  bool enable_differential_privacy_;
-  bool treat_whitespace_as_suffix_;
-  bool allow_whitespace_only_pieces_;
-  bool split_digits_;
-  bool byte_fallback_;
-  bool use_all_vocab_;
-  bool train_extremely_large_corpus_;
-  ::PROTOBUF_NAMESPACE_ID::int32 unk_id_;
-  float differential_privacy_noise_level_;
-  ::PROTOBUF_NAMESPACE_ID::uint64 differential_privacy_clipping_threshold_;
-  int model_type_;
-  ::PROTOBUF_NAMESPACE_ID::int32 vocab_size_;
-  float character_coverage_;
-  ::PROTOBUF_NAMESPACE_ID::int32 seed_sentencepiece_size_;
-  float shrinking_factor_;
-  ::PROTOBUF_NAMESPACE_ID::int32 num_threads_;
-  ::PROTOBUF_NAMESPACE_ID::int32 num_sub_iterations_;
-  ::PROTOBUF_NAMESPACE_ID::int32 max_sentence_length_;
-  ::PROTOBUF_NAMESPACE_ID::int32 max_sentencepiece_length_;
-  bool shuffle_input_sentence_;
-  bool split_by_unicode_script_;
-  bool split_by_number_;
-  bool split_by_whitespace_;
-  bool vocabulary_output_piece_score_;
-  bool hard_vocab_limit_;
-  ::PROTOBUF_NAMESPACE_ID::int32 bos_id_;
-  ::PROTOBUF_NAMESPACE_ID::int32 eos_id_;
-  ::PROTOBUF_NAMESPACE_ID::int32 pad_id_;
-  friend struct ::TableStruct_sentencepiece_5fmodel_2eproto;
-};
-// -------------------------------------------------------------------
-
-class NormalizerSpec PROTOBUF_FINAL :
-    public ::PROTOBUF_NAMESPACE_ID::MessageLite /* @@protoc_insertion_point(class_definition:sentencepiece.NormalizerSpec) */ {
- public:
-  inline NormalizerSpec() : NormalizerSpec(nullptr) {}
-  virtual ~NormalizerSpec();
-
-  NormalizerSpec(const NormalizerSpec& from);
-  NormalizerSpec(NormalizerSpec&& from) noexcept
-    : NormalizerSpec() {
-    *this = ::std::move(from);
-  }
-
-  inline NormalizerSpec& operator=(const NormalizerSpec& from) {
-    CopyFrom(from);
-    return *this;
-  }
-  inline NormalizerSpec& operator=(NormalizerSpec&& from) noexcept {
-    if (GetArena() == from.GetArena()) {
-      if (this != &from) InternalSwap(&from);
-    } else {
-      CopyFrom(from);
-    }
-    return *this;
-  }
-
-  inline const std::string& unknown_fields() const {
-    return _internal_metadata_.unknown_fields<std::string>(::PROTOBUF_NAMESPACE_ID::internal::GetEmptyString);
-  }
-  inline std::string* mutable_unknown_fields() {
-    return _internal_metadata_.mutable_unknown_fields<std::string>();
-  }
-
-  static const NormalizerSpec& default_instance();
-
-  static inline const NormalizerSpec* internal_default_instance() {
-    return reinterpret_cast<const NormalizerSpec*>(
-               &_NormalizerSpec_default_instance_);
-  }
-  static constexpr int kIndexInFileMessages =
-    1;
-
-  friend void swap(NormalizerSpec& a, NormalizerSpec& b) {
-    a.Swap(&b);
-  }
-  inline void Swap(NormalizerSpec* other) {
-    if (other == this) return;
-    if (GetArena() == other->GetArena()) {
-      InternalSwap(other);
-    } else {
-      ::PROTOBUF_NAMESPACE_ID::internal::GenericSwap(this, other);
-    }
-  }
-  void UnsafeArenaSwap(NormalizerSpec* other) {
-    if (other == this) return;
-    GOOGLE_DCHECK(GetArena() == other->GetArena());
-    InternalSwap(other);
-  }
-
-  // implements Message ----------------------------------------------
-
-  inline NormalizerSpec* New() const final {
-    return CreateMaybeMessage<NormalizerSpec>(nullptr);
-  }
-
-  NormalizerSpec* New(::PROTOBUF_NAMESPACE_ID::Arena* arena) const final {
-    return CreateMaybeMessage<NormalizerSpec>(arena);
-  }
-  void CheckTypeAndMergeFrom(const ::PROTOBUF_NAMESPACE_ID::MessageLite& from)
-    final;
-  void CopyFrom(const NormalizerSpec& from);
-  void MergeFrom(const NormalizerSpec& from);
-  PROTOBUF_ATTRIBUTE_REINITIALIZES void Clear() final;
-  bool IsInitialized() const final;
-
-  size_t ByteSizeLong() const final;
-  const char* _InternalParse(const char* ptr, ::PROTOBUF_NAMESPACE_ID::internal::ParseContext* ctx) final;
-  ::PROTOBUF_NAMESPACE_ID::uint8* _InternalSerialize(
-      ::PROTOBUF_NAMESPACE_ID::uint8* target, ::PROTOBUF_NAMESPACE_ID::io::EpsCopyOutputStream* stream) const final;
-  void DiscardUnknownFields();
-  int GetCachedSize() const final { return _cached_size_.Get(); }
-
-  private:
-  inline void SharedCtor();
-  inline void SharedDtor();
-  void SetCachedSize(int size) const;
-  void InternalSwap(NormalizerSpec* other);
-  friend class ::PROTOBUF_NAMESPACE_ID::internal::AnyMetadata;
-  static ::PROTOBUF_NAMESPACE_ID::StringPiece FullMessageName() {
-    return "sentencepiece.NormalizerSpec";
-  }
-  protected:
-  explicit NormalizerSpec(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  private:
-  static void ArenaDtor(void* object);
-  inline void RegisterArenaDtor(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  public:
-
-  std::string GetTypeName() const final;
-
-  // nested types ----------------------------------------------------
-
-  // accessors -------------------------------------------------------
-
-  enum : int {
-    kNameFieldNumber = 1,
-    kPrecompiledCharsmapFieldNumber = 2,
-    kNormalizationRuleTsvFieldNumber = 6,
-    kAddDummyPrefixFieldNumber = 3,
-    kRemoveExtraWhitespacesFieldNumber = 4,
-    kEscapeWhitespacesFieldNumber = 5,
-  };
-  // optional string name = 1;
-  bool has_name() const;
-  private:
-  bool _internal_has_name() const;
-  public:
-  void clear_name();
-  const std::string& name() const;
-  void set_name(const std::string& value);
-  void set_name(std::string&& value);
-  void set_name(const char* value);
-  void set_name(const char* value, size_t size);
-  std::string* mutable_name();
-  std::string* release_name();
-  void set_allocated_name(std::string* name);
-  private:
-  const std::string& _internal_name() const;
-  void _internal_set_name(const std::string& value);
-  std::string* _internal_mutable_name();
-  public:
-
-  // optional bytes precompiled_charsmap = 2;
-  bool has_precompiled_charsmap() const;
-  private:
-  bool _internal_has_precompiled_charsmap() const;
-  public:
-  void clear_precompiled_charsmap();
-  const std::string& precompiled_charsmap() const;
-  void set_precompiled_charsmap(const std::string& value);
-  void set_precompiled_charsmap(std::string&& value);
-  void set_precompiled_charsmap(const char* value);
-  void set_precompiled_charsmap(const void* value, size_t size);
-  std::string* mutable_precompiled_charsmap();
-  std::string* release_precompiled_charsmap();
-  void set_allocated_precompiled_charsmap(std::string* precompiled_charsmap);
-  private:
-  const std::string& _internal_precompiled_charsmap() const;
-  void _internal_set_precompiled_charsmap(const std::string& value);
-  std::string* _internal_mutable_precompiled_charsmap();
-  public:
-
-  // optional string normalization_rule_tsv = 6;
-  bool has_normalization_rule_tsv() const;
-  private:
-  bool _internal_has_normalization_rule_tsv() const;
-  public:
-  void clear_normalization_rule_tsv();
-  const std::string& normalization_rule_tsv() const;
-  void set_normalization_rule_tsv(const std::string& value);
-  void set_normalization_rule_tsv(std::string&& value);
-  void set_normalization_rule_tsv(const char* value);
-  void set_normalization_rule_tsv(const char* value, size_t size);
-  std::string* mutable_normalization_rule_tsv();
-  std::string* release_normalization_rule_tsv();
-  void set_allocated_normalization_rule_tsv(std::string* normalization_rule_tsv);
-  private:
-  const std::string& _internal_normalization_rule_tsv() const;
-  void _internal_set_normalization_rule_tsv(const std::string& value);
-  std::string* _internal_mutable_normalization_rule_tsv();
-  public:
-
-  // optional bool add_dummy_prefix = 3 [default = true];
-  bool has_add_dummy_prefix() const;
-  private:
-  bool _internal_has_add_dummy_prefix() const;
-  public:
-  void clear_add_dummy_prefix();
-  bool add_dummy_prefix() const;
-  void set_add_dummy_prefix(bool value);
-  private:
-  bool _internal_add_dummy_prefix() const;
-  void _internal_set_add_dummy_prefix(bool value);
-  public:
-
-  // optional bool remove_extra_whitespaces = 4 [default = true];
-  bool has_remove_extra_whitespaces() const;
-  private:
-  bool _internal_has_remove_extra_whitespaces() const;
-  public:
-  void clear_remove_extra_whitespaces();
-  bool remove_extra_whitespaces() const;
-  void set_remove_extra_whitespaces(bool value);
-  private:
-  bool _internal_remove_extra_whitespaces() const;
-  void _internal_set_remove_extra_whitespaces(bool value);
-  public:
-
-  // optional bool escape_whitespaces = 5 [default = true];
-  bool has_escape_whitespaces() const;
-  private:
-  bool _internal_has_escape_whitespaces() const;
-  public:
-  void clear_escape_whitespaces();
-  bool escape_whitespaces() const;
-  void set_escape_whitespaces(bool value);
-  private:
-  bool _internal_escape_whitespaces() const;
-  void _internal_set_escape_whitespaces(bool value);
-  public:
-
-  GOOGLE_PROTOBUF_EXTENSION_ACCESSORS(NormalizerSpec)
-  // @@protoc_insertion_point(class_scope:sentencepiece.NormalizerSpec)
- private:
-  class _Internal;
-
-  ::PROTOBUF_NAMESPACE_ID::internal::ExtensionSet _extensions_;
-
-  template <typename T> friend class ::PROTOBUF_NAMESPACE_ID::Arena::InternalHelper;
-  typedef void InternalArenaConstructable_;
-  typedef void DestructorSkippable_;
-  ::PROTOBUF_NAMESPACE_ID::internal::HasBits<1> _has_bits_;
-  mutable ::PROTOBUF_NAMESPACE_ID::internal::CachedSize _cached_size_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr name_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr precompiled_charsmap_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr normalization_rule_tsv_;
-  bool add_dummy_prefix_;
-  bool remove_extra_whitespaces_;
-  bool escape_whitespaces_;
-  friend struct ::TableStruct_sentencepiece_5fmodel_2eproto;
-};
-// -------------------------------------------------------------------
-
-class SelfTestData_Sample PROTOBUF_FINAL :
-    public ::PROTOBUF_NAMESPACE_ID::MessageLite /* @@protoc_insertion_point(class_definition:sentencepiece.SelfTestData.Sample) */ {
- public:
-  inline SelfTestData_Sample() : SelfTestData_Sample(nullptr) {}
-  virtual ~SelfTestData_Sample();
-
-  SelfTestData_Sample(const SelfTestData_Sample& from);
-  SelfTestData_Sample(SelfTestData_Sample&& from) noexcept
-    : SelfTestData_Sample() {
-    *this = ::std::move(from);
-  }
-
-  inline SelfTestData_Sample& operator=(const SelfTestData_Sample& from) {
-    CopyFrom(from);
-    return *this;
-  }
-  inline SelfTestData_Sample& operator=(SelfTestData_Sample&& from) noexcept {
-    if (GetArena() == from.GetArena()) {
-      if (this != &from) InternalSwap(&from);
-    } else {
-      CopyFrom(from);
-    }
-    return *this;
-  }
-
-  inline const std::string& unknown_fields() const {
-    return _internal_metadata_.unknown_fields<std::string>(::PROTOBUF_NAMESPACE_ID::internal::GetEmptyString);
-  }
-  inline std::string* mutable_unknown_fields() {
-    return _internal_metadata_.mutable_unknown_fields<std::string>();
-  }
-
-  static const SelfTestData_Sample& default_instance();
-
-  static inline const SelfTestData_Sample* internal_default_instance() {
-    return reinterpret_cast<const SelfTestData_Sample*>(
-               &_SelfTestData_Sample_default_instance_);
-  }
-  static constexpr int kIndexInFileMessages =
-    2;
-
-  friend void swap(SelfTestData_Sample& a, SelfTestData_Sample& b) {
-    a.Swap(&b);
-  }
-  inline void Swap(SelfTestData_Sample* other) {
-    if (other == this) return;
-    if (GetArena() == other->GetArena()) {
-      InternalSwap(other);
-    } else {
-      ::PROTOBUF_NAMESPACE_ID::internal::GenericSwap(this, other);
-    }
-  }
-  void UnsafeArenaSwap(SelfTestData_Sample* other) {
-    if (other == this) return;
-    GOOGLE_DCHECK(GetArena() == other->GetArena());
-    InternalSwap(other);
-  }
-
-  // implements Message ----------------------------------------------
-
-  inline SelfTestData_Sample* New() const final {
-    return CreateMaybeMessage<SelfTestData_Sample>(nullptr);
-  }
-
-  SelfTestData_Sample* New(::PROTOBUF_NAMESPACE_ID::Arena* arena) const final {
-    return CreateMaybeMessage<SelfTestData_Sample>(arena);
-  }
-  void CheckTypeAndMergeFrom(const ::PROTOBUF_NAMESPACE_ID::MessageLite& from)
-    final;
-  void CopyFrom(const SelfTestData_Sample& from);
-  void MergeFrom(const SelfTestData_Sample& from);
-  PROTOBUF_ATTRIBUTE_REINITIALIZES void Clear() final;
-  bool IsInitialized() const final;
-
-  size_t ByteSizeLong() const final;
-  const char* _InternalParse(const char* ptr, ::PROTOBUF_NAMESPACE_ID::internal::ParseContext* ctx) final;
-  ::PROTOBUF_NAMESPACE_ID::uint8* _InternalSerialize(
-      ::PROTOBUF_NAMESPACE_ID::uint8* target, ::PROTOBUF_NAMESPACE_ID::io::EpsCopyOutputStream* stream) const final;
-  void DiscardUnknownFields();
-  int GetCachedSize() const final { return _cached_size_.Get(); }
-
-  private:
-  inline void SharedCtor();
-  inline void SharedDtor();
-  void SetCachedSize(int size) const;
-  void InternalSwap(SelfTestData_Sample* other);
-  friend class ::PROTOBUF_NAMESPACE_ID::internal::AnyMetadata;
-  static ::PROTOBUF_NAMESPACE_ID::StringPiece FullMessageName() {
-    return "sentencepiece.SelfTestData.Sample";
-  }
-  protected:
-  explicit SelfTestData_Sample(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  private:
-  static void ArenaDtor(void* object);
-  inline void RegisterArenaDtor(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  public:
-
-  std::string GetTypeName() const final;
-
-  // nested types ----------------------------------------------------
-
-  // accessors -------------------------------------------------------
-
-  enum : int {
-    kInputFieldNumber = 1,
-    kExpectedFieldNumber = 2,
-  };
-  // optional string input = 1;
-  bool has_input() const;
-  private:
-  bool _internal_has_input() const;
-  public:
-  void clear_input();
-  const std::string& input() const;
-  void set_input(const std::string& value);
-  void set_input(std::string&& value);
-  void set_input(const char* value);
-  void set_input(const char* value, size_t size);
-  std::string* mutable_input();
-  std::string* release_input();
-  void set_allocated_input(std::string* input);
-  private:
-  const std::string& _internal_input() const;
-  void _internal_set_input(const std::string& value);
-  std::string* _internal_mutable_input();
-  public:
-
-  // optional string expected = 2;
-  bool has_expected() const;
-  private:
-  bool _internal_has_expected() const;
-  public:
-  void clear_expected();
-  const std::string& expected() const;
-  void set_expected(const std::string& value);
-  void set_expected(std::string&& value);
-  void set_expected(const char* value);
-  void set_expected(const char* value, size_t size);
-  std::string* mutable_expected();
-  std::string* release_expected();
-  void set_allocated_expected(std::string* expected);
-  private:
-  const std::string& _internal_expected() const;
-  void _internal_set_expected(const std::string& value);
-  std::string* _internal_mutable_expected();
-  public:
-
-  // @@protoc_insertion_point(class_scope:sentencepiece.SelfTestData.Sample)
- private:
-  class _Internal;
-
-  template <typename T> friend class ::PROTOBUF_NAMESPACE_ID::Arena::InternalHelper;
-  typedef void InternalArenaConstructable_;
-  typedef void DestructorSkippable_;
-  ::PROTOBUF_NAMESPACE_ID::internal::HasBits<1> _has_bits_;
-  mutable ::PROTOBUF_NAMESPACE_ID::internal::CachedSize _cached_size_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr input_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr expected_;
-  friend struct ::TableStruct_sentencepiece_5fmodel_2eproto;
-};
-// -------------------------------------------------------------------
-
-class SelfTestData PROTOBUF_FINAL :
-    public ::PROTOBUF_NAMESPACE_ID::MessageLite /* @@protoc_insertion_point(class_definition:sentencepiece.SelfTestData) */ {
- public:
-  inline SelfTestData() : SelfTestData(nullptr) {}
-  virtual ~SelfTestData();
-
-  SelfTestData(const SelfTestData& from);
-  SelfTestData(SelfTestData&& from) noexcept
-    : SelfTestData() {
-    *this = ::std::move(from);
-  }
-
-  inline SelfTestData& operator=(const SelfTestData& from) {
-    CopyFrom(from);
-    return *this;
-  }
-  inline SelfTestData& operator=(SelfTestData&& from) noexcept {
-    if (GetArena() == from.GetArena()) {
-      if (this != &from) InternalSwap(&from);
-    } else {
-      CopyFrom(from);
-    }
-    return *this;
-  }
-
-  inline const std::string& unknown_fields() const {
-    return _internal_metadata_.unknown_fields<std::string>(::PROTOBUF_NAMESPACE_ID::internal::GetEmptyString);
-  }
-  inline std::string* mutable_unknown_fields() {
-    return _internal_metadata_.mutable_unknown_fields<std::string>();
-  }
-
-  static const SelfTestData& default_instance();
-
-  static inline const SelfTestData* internal_default_instance() {
-    return reinterpret_cast<const SelfTestData*>(
-               &_SelfTestData_default_instance_);
-  }
-  static constexpr int kIndexInFileMessages =
-    3;
-
-  friend void swap(SelfTestData& a, SelfTestData& b) {
-    a.Swap(&b);
-  }
-  inline void Swap(SelfTestData* other) {
-    if (other == this) return;
-    if (GetArena() == other->GetArena()) {
-      InternalSwap(other);
-    } else {
-      ::PROTOBUF_NAMESPACE_ID::internal::GenericSwap(this, other);
-    }
-  }
-  void UnsafeArenaSwap(SelfTestData* other) {
-    if (other == this) return;
-    GOOGLE_DCHECK(GetArena() == other->GetArena());
-    InternalSwap(other);
-  }
-
-  // implements Message ----------------------------------------------
-
-  inline SelfTestData* New() const final {
-    return CreateMaybeMessage<SelfTestData>(nullptr);
-  }
-
-  SelfTestData* New(::PROTOBUF_NAMESPACE_ID::Arena* arena) const final {
-    return CreateMaybeMessage<SelfTestData>(arena);
-  }
-  void CheckTypeAndMergeFrom(const ::PROTOBUF_NAMESPACE_ID::MessageLite& from)
-    final;
-  void CopyFrom(const SelfTestData& from);
-  void MergeFrom(const SelfTestData& from);
-  PROTOBUF_ATTRIBUTE_REINITIALIZES void Clear() final;
-  bool IsInitialized() const final;
-
-  size_t ByteSizeLong() const final;
-  const char* _InternalParse(const char* ptr, ::PROTOBUF_NAMESPACE_ID::internal::ParseContext* ctx) final;
-  ::PROTOBUF_NAMESPACE_ID::uint8* _InternalSerialize(
-      ::PROTOBUF_NAMESPACE_ID::uint8* target, ::PROTOBUF_NAMESPACE_ID::io::EpsCopyOutputStream* stream) const final;
-  void DiscardUnknownFields();
-  int GetCachedSize() const final { return _cached_size_.Get(); }
-
-  private:
-  inline void SharedCtor();
-  inline void SharedDtor();
-  void SetCachedSize(int size) const;
-  void InternalSwap(SelfTestData* other);
-  friend class ::PROTOBUF_NAMESPACE_ID::internal::AnyMetadata;
-  static ::PROTOBUF_NAMESPACE_ID::StringPiece FullMessageName() {
-    return "sentencepiece.SelfTestData";
-  }
-  protected:
-  explicit SelfTestData(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  private:
-  static void ArenaDtor(void* object);
-  inline void RegisterArenaDtor(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  public:
-
-  std::string GetTypeName() const final;
-
-  // nested types ----------------------------------------------------
-
-  typedef SelfTestData_Sample Sample;
-
-  // accessors -------------------------------------------------------
-
-  enum : int {
-    kSamplesFieldNumber = 1,
-  };
-  // repeated .sentencepiece.SelfTestData.Sample samples = 1;
-  int samples_size() const;
-  private:
-  int _internal_samples_size() const;
-  public:
-  void clear_samples();
-  ::sentencepiece::SelfTestData_Sample* mutable_samples(int index);
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::SelfTestData_Sample >*
-      mutable_samples();
-  private:
-  const ::sentencepiece::SelfTestData_Sample& _internal_samples(int index) const;
-  ::sentencepiece::SelfTestData_Sample* _internal_add_samples();
-  public:
-  const ::sentencepiece::SelfTestData_Sample& samples(int index) const;
-  ::sentencepiece::SelfTestData_Sample* add_samples();
-  const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::SelfTestData_Sample >&
-      samples() const;
-
-  GOOGLE_PROTOBUF_EXTENSION_ACCESSORS(SelfTestData)
-  // @@protoc_insertion_point(class_scope:sentencepiece.SelfTestData)
- private:
-  class _Internal;
-
-  ::PROTOBUF_NAMESPACE_ID::internal::ExtensionSet _extensions_;
-
-  template <typename T> friend class ::PROTOBUF_NAMESPACE_ID::Arena::InternalHelper;
-  typedef void InternalArenaConstructable_;
-  typedef void DestructorSkippable_;
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::SelfTestData_Sample > samples_;
-  mutable ::PROTOBUF_NAMESPACE_ID::internal::CachedSize _cached_size_;
-  friend struct ::TableStruct_sentencepiece_5fmodel_2eproto;
-};
-// -------------------------------------------------------------------
-
-class ModelProto_SentencePiece PROTOBUF_FINAL :
-    public ::PROTOBUF_NAMESPACE_ID::MessageLite /* @@protoc_insertion_point(class_definition:sentencepiece.ModelProto.SentencePiece) */ {
- public:
-  inline ModelProto_SentencePiece() : ModelProto_SentencePiece(nullptr) {}
-  virtual ~ModelProto_SentencePiece();
-
-  ModelProto_SentencePiece(const ModelProto_SentencePiece& from);
-  ModelProto_SentencePiece(ModelProto_SentencePiece&& from) noexcept
-    : ModelProto_SentencePiece() {
-    *this = ::std::move(from);
-  }
-
-  inline ModelProto_SentencePiece& operator=(const ModelProto_SentencePiece& from) {
-    CopyFrom(from);
-    return *this;
-  }
-  inline ModelProto_SentencePiece& operator=(ModelProto_SentencePiece&& from) noexcept {
-    if (GetArena() == from.GetArena()) {
-      if (this != &from) InternalSwap(&from);
-    } else {
-      CopyFrom(from);
-    }
-    return *this;
-  }
-
-  inline const std::string& unknown_fields() const {
-    return _internal_metadata_.unknown_fields<std::string>(::PROTOBUF_NAMESPACE_ID::internal::GetEmptyString);
-  }
-  inline std::string* mutable_unknown_fields() {
-    return _internal_metadata_.mutable_unknown_fields<std::string>();
-  }
-
-  static const ModelProto_SentencePiece& default_instance();
-
-  static inline const ModelProto_SentencePiece* internal_default_instance() {
-    return reinterpret_cast<const ModelProto_SentencePiece*>(
-               &_ModelProto_SentencePiece_default_instance_);
-  }
-  static constexpr int kIndexInFileMessages =
-    4;
-
-  friend void swap(ModelProto_SentencePiece& a, ModelProto_SentencePiece& b) {
-    a.Swap(&b);
-  }
-  inline void Swap(ModelProto_SentencePiece* other) {
-    if (other == this) return;
-    if (GetArena() == other->GetArena()) {
-      InternalSwap(other);
-    } else {
-      ::PROTOBUF_NAMESPACE_ID::internal::GenericSwap(this, other);
-    }
-  }
-  void UnsafeArenaSwap(ModelProto_SentencePiece* other) {
-    if (other == this) return;
-    GOOGLE_DCHECK(GetArena() == other->GetArena());
-    InternalSwap(other);
-  }
-
-  // implements Message ----------------------------------------------
-
-  inline ModelProto_SentencePiece* New() const final {
-    return CreateMaybeMessage<ModelProto_SentencePiece>(nullptr);
-  }
-
-  ModelProto_SentencePiece* New(::PROTOBUF_NAMESPACE_ID::Arena* arena) const final {
-    return CreateMaybeMessage<ModelProto_SentencePiece>(arena);
-  }
-  void CheckTypeAndMergeFrom(const ::PROTOBUF_NAMESPACE_ID::MessageLite& from)
-    final;
-  void CopyFrom(const ModelProto_SentencePiece& from);
-  void MergeFrom(const ModelProto_SentencePiece& from);
-  PROTOBUF_ATTRIBUTE_REINITIALIZES void Clear() final;
-  bool IsInitialized() const final;
-
-  size_t ByteSizeLong() const final;
-  const char* _InternalParse(const char* ptr, ::PROTOBUF_NAMESPACE_ID::internal::ParseContext* ctx) final;
-  ::PROTOBUF_NAMESPACE_ID::uint8* _InternalSerialize(
-      ::PROTOBUF_NAMESPACE_ID::uint8* target, ::PROTOBUF_NAMESPACE_ID::io::EpsCopyOutputStream* stream) const final;
-  void DiscardUnknownFields();
-  int GetCachedSize() const final { return _cached_size_.Get(); }
-
-  private:
-  inline void SharedCtor();
-  inline void SharedDtor();
-  void SetCachedSize(int size) const;
-  void InternalSwap(ModelProto_SentencePiece* other);
-  friend class ::PROTOBUF_NAMESPACE_ID::internal::AnyMetadata;
-  static ::PROTOBUF_NAMESPACE_ID::StringPiece FullMessageName() {
-    return "sentencepiece.ModelProto.SentencePiece";
-  }
-  protected:
-  explicit ModelProto_SentencePiece(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  private:
-  static void ArenaDtor(void* object);
-  inline void RegisterArenaDtor(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  public:
-
-  std::string GetTypeName() const final;
-
-  // nested types ----------------------------------------------------
-
-  typedef ModelProto_SentencePiece_Type Type;
-  static constexpr Type NORMAL =
-    ModelProto_SentencePiece_Type_NORMAL;
-  static constexpr Type UNKNOWN =
-    ModelProto_SentencePiece_Type_UNKNOWN;
-  static constexpr Type CONTROL =
-    ModelProto_SentencePiece_Type_CONTROL;
-  static constexpr Type USER_DEFINED =
-    ModelProto_SentencePiece_Type_USER_DEFINED;
-  static constexpr Type BYTE =
-    ModelProto_SentencePiece_Type_BYTE;
-  static constexpr Type UNUSED =
-    ModelProto_SentencePiece_Type_UNUSED;
-  static inline bool Type_IsValid(int value) {
-    return ModelProto_SentencePiece_Type_IsValid(value);
-  }
-  static constexpr Type Type_MIN =
-    ModelProto_SentencePiece_Type_Type_MIN;
-  static constexpr Type Type_MAX =
-    ModelProto_SentencePiece_Type_Type_MAX;
-  static constexpr int Type_ARRAYSIZE =
-    ModelProto_SentencePiece_Type_Type_ARRAYSIZE;
-  template<typename T>
-  static inline const std::string& Type_Name(T enum_t_value) {
-    static_assert(::std::is_same<T, Type>::value ||
-      ::std::is_integral<T>::value,
-      "Incorrect type passed to function Type_Name.");
-    return ModelProto_SentencePiece_Type_Name(enum_t_value);
-  }
-  static inline bool Type_Parse(::PROTOBUF_NAMESPACE_ID::ConstStringParam name,
-      Type* value) {
-    return ModelProto_SentencePiece_Type_Parse(name, value);
-  }
-
-  // accessors -------------------------------------------------------
-
-  enum : int {
-    kPieceFieldNumber = 1,
-    kScoreFieldNumber = 2,
-    kTypeFieldNumber = 3,
-  };
-  // optional string piece = 1;
-  bool has_piece() const;
-  private:
-  bool _internal_has_piece() const;
-  public:
-  void clear_piece();
-  const std::string& piece() const;
-  void set_piece(const std::string& value);
-  void set_piece(std::string&& value);
-  void set_piece(const char* value);
-  void set_piece(const char* value, size_t size);
-  std::string* mutable_piece();
-  std::string* release_piece();
-  void set_allocated_piece(std::string* piece);
-  private:
-  const std::string& _internal_piece() const;
-  void _internal_set_piece(const std::string& value);
-  std::string* _internal_mutable_piece();
-  public:
-
-  // optional float score = 2;
-  bool has_score() const;
-  private:
-  bool _internal_has_score() const;
-  public:
-  void clear_score();
-  float score() const;
-  void set_score(float value);
-  private:
-  float _internal_score() const;
-  void _internal_set_score(float value);
-  public:
-
-  // optional .sentencepiece.ModelProto.SentencePiece.Type type = 3 [default = NORMAL];
-  bool has_type() const;
-  private:
-  bool _internal_has_type() const;
-  public:
-  void clear_type();
-  ::sentencepiece::ModelProto_SentencePiece_Type type() const;
-  void set_type(::sentencepiece::ModelProto_SentencePiece_Type value);
-  private:
-  ::sentencepiece::ModelProto_SentencePiece_Type _internal_type() const;
-  void _internal_set_type(::sentencepiece::ModelProto_SentencePiece_Type value);
-  public:
-
-  GOOGLE_PROTOBUF_EXTENSION_ACCESSORS(ModelProto_SentencePiece)
-  // @@protoc_insertion_point(class_scope:sentencepiece.ModelProto.SentencePiece)
- private:
-  class _Internal;
-
-  ::PROTOBUF_NAMESPACE_ID::internal::ExtensionSet _extensions_;
-
-  template <typename T> friend class ::PROTOBUF_NAMESPACE_ID::Arena::InternalHelper;
-  typedef void InternalArenaConstructable_;
-  typedef void DestructorSkippable_;
-  ::PROTOBUF_NAMESPACE_ID::internal::HasBits<1> _has_bits_;
-  mutable ::PROTOBUF_NAMESPACE_ID::internal::CachedSize _cached_size_;
-  ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr piece_;
-  float score_;
-  int type_;
-  friend struct ::TableStruct_sentencepiece_5fmodel_2eproto;
-};
-// -------------------------------------------------------------------
-
-class ModelProto PROTOBUF_FINAL :
-    public ::PROTOBUF_NAMESPACE_ID::MessageLite /* @@protoc_insertion_point(class_definition:sentencepiece.ModelProto) */ {
- public:
-  inline ModelProto() : ModelProto(nullptr) {}
-  virtual ~ModelProto();
-
-  ModelProto(const ModelProto& from);
-  ModelProto(ModelProto&& from) noexcept
-    : ModelProto() {
-    *this = ::std::move(from);
-  }
-
-  inline ModelProto& operator=(const ModelProto& from) {
-    CopyFrom(from);
-    return *this;
-  }
-  inline ModelProto& operator=(ModelProto&& from) noexcept {
-    if (GetArena() == from.GetArena()) {
-      if (this != &from) InternalSwap(&from);
-    } else {
-      CopyFrom(from);
-    }
-    return *this;
-  }
-
-  inline const std::string& unknown_fields() const {
-    return _internal_metadata_.unknown_fields<std::string>(::PROTOBUF_NAMESPACE_ID::internal::GetEmptyString);
-  }
-  inline std::string* mutable_unknown_fields() {
-    return _internal_metadata_.mutable_unknown_fields<std::string>();
-  }
-
-  static const ModelProto& default_instance();
-
-  static inline const ModelProto* internal_default_instance() {
-    return reinterpret_cast<const ModelProto*>(
-               &_ModelProto_default_instance_);
-  }
-  static constexpr int kIndexInFileMessages =
-    5;
-
-  friend void swap(ModelProto& a, ModelProto& b) {
-    a.Swap(&b);
-  }
-  inline void Swap(ModelProto* other) {
-    if (other == this) return;
-    if (GetArena() == other->GetArena()) {
-      InternalSwap(other);
-    } else {
-      ::PROTOBUF_NAMESPACE_ID::internal::GenericSwap(this, other);
-    }
-  }
-  void UnsafeArenaSwap(ModelProto* other) {
-    if (other == this) return;
-    GOOGLE_DCHECK(GetArena() == other->GetArena());
-    InternalSwap(other);
-  }
-
-  // implements Message ----------------------------------------------
-
-  inline ModelProto* New() const final {
-    return CreateMaybeMessage<ModelProto>(nullptr);
-  }
-
-  ModelProto* New(::PROTOBUF_NAMESPACE_ID::Arena* arena) const final {
-    return CreateMaybeMessage<ModelProto>(arena);
-  }
-  void CheckTypeAndMergeFrom(const ::PROTOBUF_NAMESPACE_ID::MessageLite& from)
-    final;
-  void CopyFrom(const ModelProto& from);
-  void MergeFrom(const ModelProto& from);
-  PROTOBUF_ATTRIBUTE_REINITIALIZES void Clear() final;
-  bool IsInitialized() const final;
-
-  size_t ByteSizeLong() const final;
-  const char* _InternalParse(const char* ptr, ::PROTOBUF_NAMESPACE_ID::internal::ParseContext* ctx) final;
-  ::PROTOBUF_NAMESPACE_ID::uint8* _InternalSerialize(
-      ::PROTOBUF_NAMESPACE_ID::uint8* target, ::PROTOBUF_NAMESPACE_ID::io::EpsCopyOutputStream* stream) const final;
-  void DiscardUnknownFields();
-  int GetCachedSize() const final { return _cached_size_.Get(); }
-
-  private:
-  inline void SharedCtor();
-  inline void SharedDtor();
-  void SetCachedSize(int size) const;
-  void InternalSwap(ModelProto* other);
-  friend class ::PROTOBUF_NAMESPACE_ID::internal::AnyMetadata;
-  static ::PROTOBUF_NAMESPACE_ID::StringPiece FullMessageName() {
-    return "sentencepiece.ModelProto";
-  }
-  protected:
-  explicit ModelProto(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  private:
-  static void ArenaDtor(void* object);
-  inline void RegisterArenaDtor(::PROTOBUF_NAMESPACE_ID::Arena* arena);
-  public:
-
-  std::string GetTypeName() const final;
-
-  // nested types ----------------------------------------------------
-
-  typedef ModelProto_SentencePiece SentencePiece;
-
-  // accessors -------------------------------------------------------
-
-  enum : int {
-    kPiecesFieldNumber = 1,
-    kTrainerSpecFieldNumber = 2,
-    kNormalizerSpecFieldNumber = 3,
-    kSelfTestDataFieldNumber = 4,
-    kDenormalizerSpecFieldNumber = 5,
-  };
-  // repeated .sentencepiece.ModelProto.SentencePiece pieces = 1;
-  int pieces_size() const;
-  private:
-  int _internal_pieces_size() const;
-  public:
-  void clear_pieces();
-  ::sentencepiece::ModelProto_SentencePiece* mutable_pieces(int index);
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::ModelProto_SentencePiece >*
-      mutable_pieces();
-  private:
-  const ::sentencepiece::ModelProto_SentencePiece& _internal_pieces(int index) const;
-  ::sentencepiece::ModelProto_SentencePiece* _internal_add_pieces();
-  public:
-  const ::sentencepiece::ModelProto_SentencePiece& pieces(int index) const;
-  ::sentencepiece::ModelProto_SentencePiece* add_pieces();
-  const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::ModelProto_SentencePiece >&
-      pieces() const;
-
-  // optional .sentencepiece.TrainerSpec trainer_spec = 2;
-  bool has_trainer_spec() const;
-  private:
-  bool _internal_has_trainer_spec() const;
-  public:
-  void clear_trainer_spec();
-  const ::sentencepiece::TrainerSpec& trainer_spec() const;
-  ::sentencepiece::TrainerSpec* release_trainer_spec();
-  ::sentencepiece::TrainerSpec* mutable_trainer_spec();
-  void set_allocated_trainer_spec(::sentencepiece::TrainerSpec* trainer_spec);
-  private:
-  const ::sentencepiece::TrainerSpec& _internal_trainer_spec() const;
-  ::sentencepiece::TrainerSpec* _internal_mutable_trainer_spec();
-  public:
-  void unsafe_arena_set_allocated_trainer_spec(
-      ::sentencepiece::TrainerSpec* trainer_spec);
-  ::sentencepiece::TrainerSpec* unsafe_arena_release_trainer_spec();
-
-  // optional .sentencepiece.NormalizerSpec normalizer_spec = 3;
-  bool has_normalizer_spec() const;
-  private:
-  bool _internal_has_normalizer_spec() const;
-  public:
-  void clear_normalizer_spec();
-  const ::sentencepiece::NormalizerSpec& normalizer_spec() const;
-  ::sentencepiece::NormalizerSpec* release_normalizer_spec();
-  ::sentencepiece::NormalizerSpec* mutable_normalizer_spec();
-  void set_allocated_normalizer_spec(::sentencepiece::NormalizerSpec* normalizer_spec);
-  private:
-  const ::sentencepiece::NormalizerSpec& _internal_normalizer_spec() const;
-  ::sentencepiece::NormalizerSpec* _internal_mutable_normalizer_spec();
-  public:
-  void unsafe_arena_set_allocated_normalizer_spec(
-      ::sentencepiece::NormalizerSpec* normalizer_spec);
-  ::sentencepiece::NormalizerSpec* unsafe_arena_release_normalizer_spec();
-
-  // optional .sentencepiece.SelfTestData self_test_data = 4;
-  bool has_self_test_data() const;
-  private:
-  bool _internal_has_self_test_data() const;
-  public:
-  void clear_self_test_data();
-  const ::sentencepiece::SelfTestData& self_test_data() const;
-  ::sentencepiece::SelfTestData* release_self_test_data();
-  ::sentencepiece::SelfTestData* mutable_self_test_data();
-  void set_allocated_self_test_data(::sentencepiece::SelfTestData* self_test_data);
-  private:
-  const ::sentencepiece::SelfTestData& _internal_self_test_data() const;
-  ::sentencepiece::SelfTestData* _internal_mutable_self_test_data();
-  public:
-  void unsafe_arena_set_allocated_self_test_data(
-      ::sentencepiece::SelfTestData* self_test_data);
-  ::sentencepiece::SelfTestData* unsafe_arena_release_self_test_data();
-
-  // optional .sentencepiece.NormalizerSpec denormalizer_spec = 5;
-  bool has_denormalizer_spec() const;
-  private:
-  bool _internal_has_denormalizer_spec() const;
-  public:
-  void clear_denormalizer_spec();
-  const ::sentencepiece::NormalizerSpec& denormalizer_spec() const;
-  ::sentencepiece::NormalizerSpec* release_denormalizer_spec();
-  ::sentencepiece::NormalizerSpec* mutable_denormalizer_spec();
-  void set_allocated_denormalizer_spec(::sentencepiece::NormalizerSpec* denormalizer_spec);
-  private:
-  const ::sentencepiece::NormalizerSpec& _internal_denormalizer_spec() const;
-  ::sentencepiece::NormalizerSpec* _internal_mutable_denormalizer_spec();
-  public:
-  void unsafe_arena_set_allocated_denormalizer_spec(
-      ::sentencepiece::NormalizerSpec* denormalizer_spec);
-  ::sentencepiece::NormalizerSpec* unsafe_arena_release_denormalizer_spec();
-
-  GOOGLE_PROTOBUF_EXTENSION_ACCESSORS(ModelProto)
-  // @@protoc_insertion_point(class_scope:sentencepiece.ModelProto)
- private:
-  class _Internal;
-
-  ::PROTOBUF_NAMESPACE_ID::internal::ExtensionSet _extensions_;
-
-  template <typename T> friend class ::PROTOBUF_NAMESPACE_ID::Arena::InternalHelper;
-  typedef void InternalArenaConstructable_;
-  typedef void DestructorSkippable_;
-  ::PROTOBUF_NAMESPACE_ID::internal::HasBits<1> _has_bits_;
-  mutable ::PROTOBUF_NAMESPACE_ID::internal::CachedSize _cached_size_;
-  ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::ModelProto_SentencePiece > pieces_;
-  ::sentencepiece::TrainerSpec* trainer_spec_;
-  ::sentencepiece::NormalizerSpec* normalizer_spec_;
-  ::sentencepiece::SelfTestData* self_test_data_;
-  ::sentencepiece::NormalizerSpec* denormalizer_spec_;
-  friend struct ::TableStruct_sentencepiece_5fmodel_2eproto;
-};
-// ===================================================================
-
-
-// ===================================================================
-
-#ifdef __GNUC__
-  #pragma GCC diagnostic push
-  #pragma GCC diagnostic ignored "-Wstrict-aliasing"
-#endif  // __GNUC__
-// TrainerSpec
-
-// repeated string input = 1;
-inline int TrainerSpec::_internal_input_size() const {
-  return input_.size();
-}
-inline int TrainerSpec::input_size() const {
-  return _internal_input_size();
-}
-inline void TrainerSpec::clear_input() {
-  input_.Clear();
-}
-inline std::string* TrainerSpec::add_input() {
-  // @@protoc_insertion_point(field_add_mutable:sentencepiece.TrainerSpec.input)
-  return _internal_add_input();
-}
-inline const std::string& TrainerSpec::_internal_input(int index) const {
-  return input_.Get(index);
-}
-inline const std::string& TrainerSpec::input(int index) const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.input)
-  return _internal_input(index);
-}
-inline std::string* TrainerSpec::mutable_input(int index) {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.TrainerSpec.input)
-  return input_.Mutable(index);
-}
-inline void TrainerSpec::set_input(int index, const std::string& value) {
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.input)
-  input_.Mutable(index)->assign(value);
-}
-inline void TrainerSpec::set_input(int index, std::string&& value) {
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.input)
-  input_.Mutable(index)->assign(std::move(value));
-}
-inline void TrainerSpec::set_input(int index, const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  input_.Mutable(index)->assign(value);
-  // @@protoc_insertion_point(field_set_char:sentencepiece.TrainerSpec.input)
-}
-inline void TrainerSpec::set_input(int index, const char* value, size_t size) {
-  input_.Mutable(index)->assign(
-    reinterpret_cast<const char*>(value), size);
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.TrainerSpec.input)
-}
-inline std::string* TrainerSpec::_internal_add_input() {
-  return input_.Add();
-}
-inline void TrainerSpec::add_input(const std::string& value) {
-  input_.Add()->assign(value);
-  // @@protoc_insertion_point(field_add:sentencepiece.TrainerSpec.input)
-}
-inline void TrainerSpec::add_input(std::string&& value) {
-  input_.Add(std::move(value));
-  // @@protoc_insertion_point(field_add:sentencepiece.TrainerSpec.input)
-}
-inline void TrainerSpec::add_input(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  input_.Add()->assign(value);
-  // @@protoc_insertion_point(field_add_char:sentencepiece.TrainerSpec.input)
-}
-inline void TrainerSpec::add_input(const char* value, size_t size) {
-  input_.Add()->assign(reinterpret_cast<const char*>(value), size);
-  // @@protoc_insertion_point(field_add_pointer:sentencepiece.TrainerSpec.input)
-}
-inline const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>&
-TrainerSpec::input() const {
-  // @@protoc_insertion_point(field_list:sentencepiece.TrainerSpec.input)
-  return input_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>*
-TrainerSpec::mutable_input() {
-  // @@protoc_insertion_point(field_mutable_list:sentencepiece.TrainerSpec.input)
-  return &input_;
-}
-
-// optional string input_format = 7;
-inline bool TrainerSpec::_internal_has_input_format() const {
-  bool value = (_has_bits_[0] & 0x00000002u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_input_format() const {
-  return _internal_has_input_format();
-}
-inline void TrainerSpec::clear_input_format() {
-  input_format_.ClearToEmpty();
-  _has_bits_[0] &= ~0x00000002u;
-}
-inline const std::string& TrainerSpec::input_format() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.input_format)
-  return _internal_input_format();
-}
-inline void TrainerSpec::set_input_format(const std::string& value) {
-  _internal_set_input_format(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.input_format)
-}
-inline std::string* TrainerSpec::mutable_input_format() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.TrainerSpec.input_format)
-  return _internal_mutable_input_format();
-}
-inline const std::string& TrainerSpec::_internal_input_format() const {
-  return input_format_.Get();
-}
-inline void TrainerSpec::_internal_set_input_format(const std::string& value) {
-  _has_bits_[0] |= 0x00000002u;
-  input_format_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, value, GetArena());
-}
-inline void TrainerSpec::set_input_format(std::string&& value) {
-  _has_bits_[0] |= 0x00000002u;
-  input_format_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.TrainerSpec.input_format)
-}
-inline void TrainerSpec::set_input_format(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000002u;
-  input_format_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.TrainerSpec.input_format)
-}
-inline void TrainerSpec::set_input_format(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000002u;
-  input_format_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.TrainerSpec.input_format)
-}
-inline std::string* TrainerSpec::_internal_mutable_input_format() {
-  _has_bits_[0] |= 0x00000002u;
-  return input_format_.Mutable(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, GetArena());
-}
-inline std::string* TrainerSpec::release_input_format() {
-  // @@protoc_insertion_point(field_release:sentencepiece.TrainerSpec.input_format)
-  if (!_internal_has_input_format()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000002u;
-  return input_format_.ReleaseNonDefault(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), GetArena());
-}
-inline void TrainerSpec::set_allocated_input_format(std::string* input_format) {
-  if (input_format != nullptr) {
-    _has_bits_[0] |= 0x00000002u;
-  } else {
-    _has_bits_[0] &= ~0x00000002u;
-  }
-  input_format_.SetAllocated(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), input_format,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.TrainerSpec.input_format)
-}
-
-// optional string model_prefix = 2;
-inline bool TrainerSpec::_internal_has_model_prefix() const {
-  bool value = (_has_bits_[0] & 0x00000001u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_model_prefix() const {
-  return _internal_has_model_prefix();
-}
-inline void TrainerSpec::clear_model_prefix() {
-  model_prefix_.ClearToEmpty();
-  _has_bits_[0] &= ~0x00000001u;
-}
-inline const std::string& TrainerSpec::model_prefix() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.model_prefix)
-  return _internal_model_prefix();
-}
-inline void TrainerSpec::set_model_prefix(const std::string& value) {
-  _internal_set_model_prefix(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.model_prefix)
-}
-inline std::string* TrainerSpec::mutable_model_prefix() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.TrainerSpec.model_prefix)
-  return _internal_mutable_model_prefix();
-}
-inline const std::string& TrainerSpec::_internal_model_prefix() const {
-  return model_prefix_.Get();
-}
-inline void TrainerSpec::_internal_set_model_prefix(const std::string& value) {
-  _has_bits_[0] |= 0x00000001u;
-  model_prefix_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, value, GetArena());
-}
-inline void TrainerSpec::set_model_prefix(std::string&& value) {
-  _has_bits_[0] |= 0x00000001u;
-  model_prefix_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.TrainerSpec.model_prefix)
-}
-inline void TrainerSpec::set_model_prefix(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000001u;
-  model_prefix_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.TrainerSpec.model_prefix)
-}
-inline void TrainerSpec::set_model_prefix(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000001u;
-  model_prefix_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.TrainerSpec.model_prefix)
-}
-inline std::string* TrainerSpec::_internal_mutable_model_prefix() {
-  _has_bits_[0] |= 0x00000001u;
-  return model_prefix_.Mutable(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, GetArena());
-}
-inline std::string* TrainerSpec::release_model_prefix() {
-  // @@protoc_insertion_point(field_release:sentencepiece.TrainerSpec.model_prefix)
-  if (!_internal_has_model_prefix()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000001u;
-  return model_prefix_.ReleaseNonDefault(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), GetArena());
-}
-inline void TrainerSpec::set_allocated_model_prefix(std::string* model_prefix) {
-  if (model_prefix != nullptr) {
-    _has_bits_[0] |= 0x00000001u;
-  } else {
-    _has_bits_[0] &= ~0x00000001u;
-  }
-  model_prefix_.SetAllocated(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), model_prefix,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.TrainerSpec.model_prefix)
-}
-
-// optional .sentencepiece.TrainerSpec.ModelType model_type = 3 [default = UNIGRAM];
-inline bool TrainerSpec::_internal_has_model_type() const {
-  bool value = (_has_bits_[0] & 0x01000000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_model_type() const {
-  return _internal_has_model_type();
-}
-inline void TrainerSpec::clear_model_type() {
-  model_type_ = 1;
-  _has_bits_[0] &= ~0x01000000u;
-}
-inline ::sentencepiece::TrainerSpec_ModelType TrainerSpec::_internal_model_type() const {
-  return static_cast< ::sentencepiece::TrainerSpec_ModelType >(model_type_);
-}
-inline ::sentencepiece::TrainerSpec_ModelType TrainerSpec::model_type() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.model_type)
-  return _internal_model_type();
-}
-inline void TrainerSpec::_internal_set_model_type(::sentencepiece::TrainerSpec_ModelType value) {
-  assert(::sentencepiece::TrainerSpec_ModelType_IsValid(value));
-  _has_bits_[0] |= 0x01000000u;
-  model_type_ = value;
-}
-inline void TrainerSpec::set_model_type(::sentencepiece::TrainerSpec_ModelType value) {
-  _internal_set_model_type(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.model_type)
-}
-
-// optional int32 vocab_size = 4 [default = 8000];
-inline bool TrainerSpec::_internal_has_vocab_size() const {
-  bool value = (_has_bits_[0] & 0x02000000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_vocab_size() const {
-  return _internal_has_vocab_size();
-}
-inline void TrainerSpec::clear_vocab_size() {
-  vocab_size_ = 8000;
-  _has_bits_[0] &= ~0x02000000u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::_internal_vocab_size() const {
-  return vocab_size_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::vocab_size() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.vocab_size)
-  return _internal_vocab_size();
-}
-inline void TrainerSpec::_internal_set_vocab_size(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _has_bits_[0] |= 0x02000000u;
-  vocab_size_ = value;
-}
-inline void TrainerSpec::set_vocab_size(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _internal_set_vocab_size(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.vocab_size)
-}
-
-// repeated string accept_language = 5;
-inline int TrainerSpec::_internal_accept_language_size() const {
-  return accept_language_.size();
-}
-inline int TrainerSpec::accept_language_size() const {
-  return _internal_accept_language_size();
-}
-inline void TrainerSpec::clear_accept_language() {
-  accept_language_.Clear();
-}
-inline std::string* TrainerSpec::add_accept_language() {
-  // @@protoc_insertion_point(field_add_mutable:sentencepiece.TrainerSpec.accept_language)
-  return _internal_add_accept_language();
-}
-inline const std::string& TrainerSpec::_internal_accept_language(int index) const {
-  return accept_language_.Get(index);
-}
-inline const std::string& TrainerSpec::accept_language(int index) const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.accept_language)
-  return _internal_accept_language(index);
-}
-inline std::string* TrainerSpec::mutable_accept_language(int index) {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.TrainerSpec.accept_language)
-  return accept_language_.Mutable(index);
-}
-inline void TrainerSpec::set_accept_language(int index, const std::string& value) {
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.accept_language)
-  accept_language_.Mutable(index)->assign(value);
-}
-inline void TrainerSpec::set_accept_language(int index, std::string&& value) {
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.accept_language)
-  accept_language_.Mutable(index)->assign(std::move(value));
-}
-inline void TrainerSpec::set_accept_language(int index, const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  accept_language_.Mutable(index)->assign(value);
-  // @@protoc_insertion_point(field_set_char:sentencepiece.TrainerSpec.accept_language)
-}
-inline void TrainerSpec::set_accept_language(int index, const char* value, size_t size) {
-  accept_language_.Mutable(index)->assign(
-    reinterpret_cast<const char*>(value), size);
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.TrainerSpec.accept_language)
-}
-inline std::string* TrainerSpec::_internal_add_accept_language() {
-  return accept_language_.Add();
-}
-inline void TrainerSpec::add_accept_language(const std::string& value) {
-  accept_language_.Add()->assign(value);
-  // @@protoc_insertion_point(field_add:sentencepiece.TrainerSpec.accept_language)
-}
-inline void TrainerSpec::add_accept_language(std::string&& value) {
-  accept_language_.Add(std::move(value));
-  // @@protoc_insertion_point(field_add:sentencepiece.TrainerSpec.accept_language)
-}
-inline void TrainerSpec::add_accept_language(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  accept_language_.Add()->assign(value);
-  // @@protoc_insertion_point(field_add_char:sentencepiece.TrainerSpec.accept_language)
-}
-inline void TrainerSpec::add_accept_language(const char* value, size_t size) {
-  accept_language_.Add()->assign(reinterpret_cast<const char*>(value), size);
-  // @@protoc_insertion_point(field_add_pointer:sentencepiece.TrainerSpec.accept_language)
-}
-inline const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>&
-TrainerSpec::accept_language() const {
-  // @@protoc_insertion_point(field_list:sentencepiece.TrainerSpec.accept_language)
-  return accept_language_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>*
-TrainerSpec::mutable_accept_language() {
-  // @@protoc_insertion_point(field_mutable_list:sentencepiece.TrainerSpec.accept_language)
-  return &accept_language_;
-}
-
-// optional int32 self_test_sample_size = 6 [default = 0];
-inline bool TrainerSpec::_internal_has_self_test_sample_size() const {
-  bool value = (_has_bits_[0] & 0x00000400u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_self_test_sample_size() const {
-  return _internal_has_self_test_sample_size();
-}
-inline void TrainerSpec::clear_self_test_sample_size() {
-  self_test_sample_size_ = 0;
-  _has_bits_[0] &= ~0x00000400u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::_internal_self_test_sample_size() const {
-  return self_test_sample_size_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::self_test_sample_size() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.self_test_sample_size)
-  return _internal_self_test_sample_size();
-}
-inline void TrainerSpec::_internal_set_self_test_sample_size(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _has_bits_[0] |= 0x00000400u;
-  self_test_sample_size_ = value;
-}
-inline void TrainerSpec::set_self_test_sample_size(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _internal_set_self_test_sample_size(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.self_test_sample_size)
-}
-
-// optional bool enable_differential_privacy = 50 [default = false];
-inline bool TrainerSpec::_internal_has_enable_differential_privacy() const {
-  bool value = (_has_bits_[0] & 0x00004000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_enable_differential_privacy() const {
-  return _internal_has_enable_differential_privacy();
-}
-inline void TrainerSpec::clear_enable_differential_privacy() {
-  enable_differential_privacy_ = false;
-  _has_bits_[0] &= ~0x00004000u;
-}
-inline bool TrainerSpec::_internal_enable_differential_privacy() const {
-  return enable_differential_privacy_;
-}
-inline bool TrainerSpec::enable_differential_privacy() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.enable_differential_privacy)
-  return _internal_enable_differential_privacy();
-}
-inline void TrainerSpec::_internal_set_enable_differential_privacy(bool value) {
-  _has_bits_[0] |= 0x00004000u;
-  enable_differential_privacy_ = value;
-}
-inline void TrainerSpec::set_enable_differential_privacy(bool value) {
-  _internal_set_enable_differential_privacy(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.enable_differential_privacy)
-}
-
-// optional float differential_privacy_noise_level = 51 [default = 0];
-inline bool TrainerSpec::_internal_has_differential_privacy_noise_level() const {
-  bool value = (_has_bits_[0] & 0x00400000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_differential_privacy_noise_level() const {
-  return _internal_has_differential_privacy_noise_level();
-}
-inline void TrainerSpec::clear_differential_privacy_noise_level() {
-  differential_privacy_noise_level_ = 0;
-  _has_bits_[0] &= ~0x00400000u;
-}
-inline float TrainerSpec::_internal_differential_privacy_noise_level() const {
-  return differential_privacy_noise_level_;
-}
-inline float TrainerSpec::differential_privacy_noise_level() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.differential_privacy_noise_level)
-  return _internal_differential_privacy_noise_level();
-}
-inline void TrainerSpec::_internal_set_differential_privacy_noise_level(float value) {
-  _has_bits_[0] |= 0x00400000u;
-  differential_privacy_noise_level_ = value;
-}
-inline void TrainerSpec::set_differential_privacy_noise_level(float value) {
-  _internal_set_differential_privacy_noise_level(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.differential_privacy_noise_level)
-}
-
-// optional uint64 differential_privacy_clipping_threshold = 52 [default = 0];
-inline bool TrainerSpec::_internal_has_differential_privacy_clipping_threshold() const {
-  bool value = (_has_bits_[0] & 0x00800000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_differential_privacy_clipping_threshold() const {
-  return _internal_has_differential_privacy_clipping_threshold();
-}
-inline void TrainerSpec::clear_differential_privacy_clipping_threshold() {
-  differential_privacy_clipping_threshold_ = PROTOBUF_ULONGLONG(0);
-  _has_bits_[0] &= ~0x00800000u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::uint64 TrainerSpec::_internal_differential_privacy_clipping_threshold() const {
-  return differential_privacy_clipping_threshold_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::uint64 TrainerSpec::differential_privacy_clipping_threshold() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.differential_privacy_clipping_threshold)
-  return _internal_differential_privacy_clipping_threshold();
-}
-inline void TrainerSpec::_internal_set_differential_privacy_clipping_threshold(::PROTOBUF_NAMESPACE_ID::uint64 value) {
-  _has_bits_[0] |= 0x00800000u;
-  differential_privacy_clipping_threshold_ = value;
-}
-inline void TrainerSpec::set_differential_privacy_clipping_threshold(::PROTOBUF_NAMESPACE_ID::uint64 value) {
-  _internal_set_differential_privacy_clipping_threshold(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.differential_privacy_clipping_threshold)
-}
-
-// optional float character_coverage = 10 [default = 0.9995];
-inline bool TrainerSpec::_internal_has_character_coverage() const {
-  bool value = (_has_bits_[0] & 0x04000000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_character_coverage() const {
-  return _internal_has_character_coverage();
-}
-inline void TrainerSpec::clear_character_coverage() {
-  character_coverage_ = 0.9995f;
-  _has_bits_[0] &= ~0x04000000u;
-}
-inline float TrainerSpec::_internal_character_coverage() const {
-  return character_coverage_;
-}
-inline float TrainerSpec::character_coverage() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.character_coverage)
-  return _internal_character_coverage();
-}
-inline void TrainerSpec::_internal_set_character_coverage(float value) {
-  _has_bits_[0] |= 0x04000000u;
-  character_coverage_ = value;
-}
-inline void TrainerSpec::set_character_coverage(float value) {
-  _internal_set_character_coverage(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.character_coverage)
-}
-
-// optional uint64 input_sentence_size = 11 [default = 0];
-inline bool TrainerSpec::_internal_has_input_sentence_size() const {
-  bool value = (_has_bits_[0] & 0x00001000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_input_sentence_size() const {
-  return _internal_has_input_sentence_size();
-}
-inline void TrainerSpec::clear_input_sentence_size() {
-  input_sentence_size_ = PROTOBUF_ULONGLONG(0);
-  _has_bits_[0] &= ~0x00001000u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::uint64 TrainerSpec::_internal_input_sentence_size() const {
-  return input_sentence_size_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::uint64 TrainerSpec::input_sentence_size() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.input_sentence_size)
-  return _internal_input_sentence_size();
-}
-inline void TrainerSpec::_internal_set_input_sentence_size(::PROTOBUF_NAMESPACE_ID::uint64 value) {
-  _has_bits_[0] |= 0x00001000u;
-  input_sentence_size_ = value;
-}
-inline void TrainerSpec::set_input_sentence_size(::PROTOBUF_NAMESPACE_ID::uint64 value) {
-  _internal_set_input_sentence_size(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.input_sentence_size)
-}
-
-// optional bool shuffle_input_sentence = 19 [default = true];
-inline bool TrainerSpec::_internal_has_shuffle_input_sentence() const {
-  bool value = (_has_bits_[1] & 0x00000002u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_shuffle_input_sentence() const {
-  return _internal_has_shuffle_input_sentence();
-}
-inline void TrainerSpec::clear_shuffle_input_sentence() {
-  shuffle_input_sentence_ = true;
-  _has_bits_[1] &= ~0x00000002u;
-}
-inline bool TrainerSpec::_internal_shuffle_input_sentence() const {
-  return shuffle_input_sentence_;
-}
-inline bool TrainerSpec::shuffle_input_sentence() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.shuffle_input_sentence)
-  return _internal_shuffle_input_sentence();
-}
-inline void TrainerSpec::_internal_set_shuffle_input_sentence(bool value) {
-  _has_bits_[1] |= 0x00000002u;
-  shuffle_input_sentence_ = value;
-}
-inline void TrainerSpec::set_shuffle_input_sentence(bool value) {
-  _internal_set_shuffle_input_sentence(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.shuffle_input_sentence)
-}
-
-// optional int32 mining_sentence_size = 12 [deprecated = true];
-inline bool TrainerSpec::_internal_has_mining_sentence_size() const {
-  bool value = (_has_bits_[0] & 0x00000800u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_mining_sentence_size() const {
-  return _internal_has_mining_sentence_size();
-}
-inline void TrainerSpec::clear_mining_sentence_size() {
-  mining_sentence_size_ = 0;
-  _has_bits_[0] &= ~0x00000800u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::_internal_mining_sentence_size() const {
-  return mining_sentence_size_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::mining_sentence_size() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.mining_sentence_size)
-  return _internal_mining_sentence_size();
-}
-inline void TrainerSpec::_internal_set_mining_sentence_size(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _has_bits_[0] |= 0x00000800u;
-  mining_sentence_size_ = value;
-}
-inline void TrainerSpec::set_mining_sentence_size(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _internal_set_mining_sentence_size(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.mining_sentence_size)
-}
-
-// optional int32 training_sentence_size = 13 [deprecated = true];
-inline bool TrainerSpec::_internal_has_training_sentence_size() const {
-  bool value = (_has_bits_[0] & 0x00002000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_training_sentence_size() const {
-  return _internal_has_training_sentence_size();
-}
-inline void TrainerSpec::clear_training_sentence_size() {
-  training_sentence_size_ = 0;
-  _has_bits_[0] &= ~0x00002000u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::_internal_training_sentence_size() const {
-  return training_sentence_size_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::training_sentence_size() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.training_sentence_size)
-  return _internal_training_sentence_size();
-}
-inline void TrainerSpec::_internal_set_training_sentence_size(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _has_bits_[0] |= 0x00002000u;
-  training_sentence_size_ = value;
-}
-inline void TrainerSpec::set_training_sentence_size(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _internal_set_training_sentence_size(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.training_sentence_size)
-}
-
-// optional int32 seed_sentencepiece_size = 14 [default = 1000000];
-inline bool TrainerSpec::_internal_has_seed_sentencepiece_size() const {
-  bool value = (_has_bits_[0] & 0x08000000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_seed_sentencepiece_size() const {
-  return _internal_has_seed_sentencepiece_size();
-}
-inline void TrainerSpec::clear_seed_sentencepiece_size() {
-  seed_sentencepiece_size_ = 1000000;
-  _has_bits_[0] &= ~0x08000000u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::_internal_seed_sentencepiece_size() const {
-  return seed_sentencepiece_size_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::seed_sentencepiece_size() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.seed_sentencepiece_size)
-  return _internal_seed_sentencepiece_size();
-}
-inline void TrainerSpec::_internal_set_seed_sentencepiece_size(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _has_bits_[0] |= 0x08000000u;
-  seed_sentencepiece_size_ = value;
-}
-inline void TrainerSpec::set_seed_sentencepiece_size(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _internal_set_seed_sentencepiece_size(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.seed_sentencepiece_size)
-}
-
-// optional float shrinking_factor = 15 [default = 0.75];
-inline bool TrainerSpec::_internal_has_shrinking_factor() const {
-  bool value = (_has_bits_[0] & 0x10000000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_shrinking_factor() const {
-  return _internal_has_shrinking_factor();
-}
-inline void TrainerSpec::clear_shrinking_factor() {
-  shrinking_factor_ = 0.75f;
-  _has_bits_[0] &= ~0x10000000u;
-}
-inline float TrainerSpec::_internal_shrinking_factor() const {
-  return shrinking_factor_;
-}
-inline float TrainerSpec::shrinking_factor() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.shrinking_factor)
-  return _internal_shrinking_factor();
-}
-inline void TrainerSpec::_internal_set_shrinking_factor(float value) {
-  _has_bits_[0] |= 0x10000000u;
-  shrinking_factor_ = value;
-}
-inline void TrainerSpec::set_shrinking_factor(float value) {
-  _internal_set_shrinking_factor(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.shrinking_factor)
-}
-
-// optional int32 max_sentence_length = 18 [default = 4192];
-inline bool TrainerSpec::_internal_has_max_sentence_length() const {
-  bool value = (_has_bits_[0] & 0x80000000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_max_sentence_length() const {
-  return _internal_has_max_sentence_length();
-}
-inline void TrainerSpec::clear_max_sentence_length() {
-  max_sentence_length_ = 4192;
-  _has_bits_[0] &= ~0x80000000u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::_internal_max_sentence_length() const {
-  return max_sentence_length_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::max_sentence_length() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.max_sentence_length)
-  return _internal_max_sentence_length();
-}
-inline void TrainerSpec::_internal_set_max_sentence_length(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _has_bits_[0] |= 0x80000000u;
-  max_sentence_length_ = value;
-}
-inline void TrainerSpec::set_max_sentence_length(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _internal_set_max_sentence_length(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.max_sentence_length)
-}
-
-// optional int32 num_threads = 16 [default = 16];
-inline bool TrainerSpec::_internal_has_num_threads() const {
-  bool value = (_has_bits_[0] & 0x20000000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_num_threads() const {
-  return _internal_has_num_threads();
-}
-inline void TrainerSpec::clear_num_threads() {
-  num_threads_ = 16;
-  _has_bits_[0] &= ~0x20000000u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::_internal_num_threads() const {
-  return num_threads_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::num_threads() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.num_threads)
-  return _internal_num_threads();
-}
-inline void TrainerSpec::_internal_set_num_threads(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _has_bits_[0] |= 0x20000000u;
-  num_threads_ = value;
-}
-inline void TrainerSpec::set_num_threads(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _internal_set_num_threads(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.num_threads)
-}
-
-// optional int32 num_sub_iterations = 17 [default = 2];
-inline bool TrainerSpec::_internal_has_num_sub_iterations() const {
-  bool value = (_has_bits_[0] & 0x40000000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_num_sub_iterations() const {
-  return _internal_has_num_sub_iterations();
-}
-inline void TrainerSpec::clear_num_sub_iterations() {
-  num_sub_iterations_ = 2;
-  _has_bits_[0] &= ~0x40000000u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::_internal_num_sub_iterations() const {
-  return num_sub_iterations_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::num_sub_iterations() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.num_sub_iterations)
-  return _internal_num_sub_iterations();
-}
-inline void TrainerSpec::_internal_set_num_sub_iterations(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _has_bits_[0] |= 0x40000000u;
-  num_sub_iterations_ = value;
-}
-inline void TrainerSpec::set_num_sub_iterations(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _internal_set_num_sub_iterations(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.num_sub_iterations)
-}
-
-// optional int32 max_sentencepiece_length = 20 [default = 16];
-inline bool TrainerSpec::_internal_has_max_sentencepiece_length() const {
-  bool value = (_has_bits_[1] & 0x00000001u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_max_sentencepiece_length() const {
-  return _internal_has_max_sentencepiece_length();
-}
-inline void TrainerSpec::clear_max_sentencepiece_length() {
-  max_sentencepiece_length_ = 16;
-  _has_bits_[1] &= ~0x00000001u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::_internal_max_sentencepiece_length() const {
-  return max_sentencepiece_length_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::max_sentencepiece_length() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.max_sentencepiece_length)
-  return _internal_max_sentencepiece_length();
-}
-inline void TrainerSpec::_internal_set_max_sentencepiece_length(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _has_bits_[1] |= 0x00000001u;
-  max_sentencepiece_length_ = value;
-}
-inline void TrainerSpec::set_max_sentencepiece_length(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _internal_set_max_sentencepiece_length(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.max_sentencepiece_length)
-}
-
-// optional bool split_by_unicode_script = 21 [default = true];
-inline bool TrainerSpec::_internal_has_split_by_unicode_script() const {
-  bool value = (_has_bits_[1] & 0x00000004u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_split_by_unicode_script() const {
-  return _internal_has_split_by_unicode_script();
-}
-inline void TrainerSpec::clear_split_by_unicode_script() {
-  split_by_unicode_script_ = true;
-  _has_bits_[1] &= ~0x00000004u;
-}
-inline bool TrainerSpec::_internal_split_by_unicode_script() const {
-  return split_by_unicode_script_;
-}
-inline bool TrainerSpec::split_by_unicode_script() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.split_by_unicode_script)
-  return _internal_split_by_unicode_script();
-}
-inline void TrainerSpec::_internal_set_split_by_unicode_script(bool value) {
-  _has_bits_[1] |= 0x00000004u;
-  split_by_unicode_script_ = value;
-}
-inline void TrainerSpec::set_split_by_unicode_script(bool value) {
-  _internal_set_split_by_unicode_script(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.split_by_unicode_script)
-}
-
-// optional bool split_by_number = 23 [default = true];
-inline bool TrainerSpec::_internal_has_split_by_number() const {
-  bool value = (_has_bits_[1] & 0x00000008u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_split_by_number() const {
-  return _internal_has_split_by_number();
-}
-inline void TrainerSpec::clear_split_by_number() {
-  split_by_number_ = true;
-  _has_bits_[1] &= ~0x00000008u;
-}
-inline bool TrainerSpec::_internal_split_by_number() const {
-  return split_by_number_;
-}
-inline bool TrainerSpec::split_by_number() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.split_by_number)
-  return _internal_split_by_number();
-}
-inline void TrainerSpec::_internal_set_split_by_number(bool value) {
-  _has_bits_[1] |= 0x00000008u;
-  split_by_number_ = value;
-}
-inline void TrainerSpec::set_split_by_number(bool value) {
-  _internal_set_split_by_number(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.split_by_number)
-}
-
-// optional bool split_by_whitespace = 22 [default = true];
-inline bool TrainerSpec::_internal_has_split_by_whitespace() const {
-  bool value = (_has_bits_[1] & 0x00000010u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_split_by_whitespace() const {
-  return _internal_has_split_by_whitespace();
-}
-inline void TrainerSpec::clear_split_by_whitespace() {
-  split_by_whitespace_ = true;
-  _has_bits_[1] &= ~0x00000010u;
-}
-inline bool TrainerSpec::_internal_split_by_whitespace() const {
-  return split_by_whitespace_;
-}
-inline bool TrainerSpec::split_by_whitespace() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.split_by_whitespace)
-  return _internal_split_by_whitespace();
-}
-inline void TrainerSpec::_internal_set_split_by_whitespace(bool value) {
-  _has_bits_[1] |= 0x00000010u;
-  split_by_whitespace_ = value;
-}
-inline void TrainerSpec::set_split_by_whitespace(bool value) {
-  _internal_set_split_by_whitespace(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.split_by_whitespace)
-}
-
-// optional bool treat_whitespace_as_suffix = 24 [default = false];
-inline bool TrainerSpec::_internal_has_treat_whitespace_as_suffix() const {
-  bool value = (_has_bits_[0] & 0x00008000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_treat_whitespace_as_suffix() const {
-  return _internal_has_treat_whitespace_as_suffix();
-}
-inline void TrainerSpec::clear_treat_whitespace_as_suffix() {
-  treat_whitespace_as_suffix_ = false;
-  _has_bits_[0] &= ~0x00008000u;
-}
-inline bool TrainerSpec::_internal_treat_whitespace_as_suffix() const {
-  return treat_whitespace_as_suffix_;
-}
-inline bool TrainerSpec::treat_whitespace_as_suffix() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.treat_whitespace_as_suffix)
-  return _internal_treat_whitespace_as_suffix();
-}
-inline void TrainerSpec::_internal_set_treat_whitespace_as_suffix(bool value) {
-  _has_bits_[0] |= 0x00008000u;
-  treat_whitespace_as_suffix_ = value;
-}
-inline void TrainerSpec::set_treat_whitespace_as_suffix(bool value) {
-  _internal_set_treat_whitespace_as_suffix(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.treat_whitespace_as_suffix)
-}
-
-// optional bool allow_whitespace_only_pieces = 26 [default = false];
-inline bool TrainerSpec::_internal_has_allow_whitespace_only_pieces() const {
-  bool value = (_has_bits_[0] & 0x00010000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_allow_whitespace_only_pieces() const {
-  return _internal_has_allow_whitespace_only_pieces();
-}
-inline void TrainerSpec::clear_allow_whitespace_only_pieces() {
-  allow_whitespace_only_pieces_ = false;
-  _has_bits_[0] &= ~0x00010000u;
-}
-inline bool TrainerSpec::_internal_allow_whitespace_only_pieces() const {
-  return allow_whitespace_only_pieces_;
-}
-inline bool TrainerSpec::allow_whitespace_only_pieces() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.allow_whitespace_only_pieces)
-  return _internal_allow_whitespace_only_pieces();
-}
-inline void TrainerSpec::_internal_set_allow_whitespace_only_pieces(bool value) {
-  _has_bits_[0] |= 0x00010000u;
-  allow_whitespace_only_pieces_ = value;
-}
-inline void TrainerSpec::set_allow_whitespace_only_pieces(bool value) {
-  _internal_set_allow_whitespace_only_pieces(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.allow_whitespace_only_pieces)
-}
-
-// optional bool split_digits = 25 [default = false];
-inline bool TrainerSpec::_internal_has_split_digits() const {
-  bool value = (_has_bits_[0] & 0x00020000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_split_digits() const {
-  return _internal_has_split_digits();
-}
-inline void TrainerSpec::clear_split_digits() {
-  split_digits_ = false;
-  _has_bits_[0] &= ~0x00020000u;
-}
-inline bool TrainerSpec::_internal_split_digits() const {
-  return split_digits_;
-}
-inline bool TrainerSpec::split_digits() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.split_digits)
-  return _internal_split_digits();
-}
-inline void TrainerSpec::_internal_set_split_digits(bool value) {
-  _has_bits_[0] |= 0x00020000u;
-  split_digits_ = value;
-}
-inline void TrainerSpec::set_split_digits(bool value) {
-  _internal_set_split_digits(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.split_digits)
-}
-
-// optional string pretokenization_delimiter = 53 [default = ""];
-inline bool TrainerSpec::_internal_has_pretokenization_delimiter() const {
-  bool value = (_has_bits_[0] & 0x00000100u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_pretokenization_delimiter() const {
-  return _internal_has_pretokenization_delimiter();
-}
-inline void TrainerSpec::clear_pretokenization_delimiter() {
-  pretokenization_delimiter_.ClearToEmpty();
-  _has_bits_[0] &= ~0x00000100u;
-}
-inline const std::string& TrainerSpec::pretokenization_delimiter() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.pretokenization_delimiter)
-  return _internal_pretokenization_delimiter();
-}
-inline void TrainerSpec::set_pretokenization_delimiter(const std::string& value) {
-  _internal_set_pretokenization_delimiter(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.pretokenization_delimiter)
-}
-inline std::string* TrainerSpec::mutable_pretokenization_delimiter() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.TrainerSpec.pretokenization_delimiter)
-  return _internal_mutable_pretokenization_delimiter();
-}
-inline const std::string& TrainerSpec::_internal_pretokenization_delimiter() const {
-  return pretokenization_delimiter_.Get();
-}
-inline void TrainerSpec::_internal_set_pretokenization_delimiter(const std::string& value) {
-  _has_bits_[0] |= 0x00000100u;
-  pretokenization_delimiter_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, value, GetArena());
-}
-inline void TrainerSpec::set_pretokenization_delimiter(std::string&& value) {
-  _has_bits_[0] |= 0x00000100u;
-  pretokenization_delimiter_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.TrainerSpec.pretokenization_delimiter)
-}
-inline void TrainerSpec::set_pretokenization_delimiter(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000100u;
-  pretokenization_delimiter_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.TrainerSpec.pretokenization_delimiter)
-}
-inline void TrainerSpec::set_pretokenization_delimiter(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000100u;
-  pretokenization_delimiter_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.TrainerSpec.pretokenization_delimiter)
-}
-inline std::string* TrainerSpec::_internal_mutable_pretokenization_delimiter() {
-  _has_bits_[0] |= 0x00000100u;
-  return pretokenization_delimiter_.Mutable(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, GetArena());
-}
-inline std::string* TrainerSpec::release_pretokenization_delimiter() {
-  // @@protoc_insertion_point(field_release:sentencepiece.TrainerSpec.pretokenization_delimiter)
-  if (!_internal_has_pretokenization_delimiter()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000100u;
-  return pretokenization_delimiter_.ReleaseNonDefault(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), GetArena());
-}
-inline void TrainerSpec::set_allocated_pretokenization_delimiter(std::string* pretokenization_delimiter) {
-  if (pretokenization_delimiter != nullptr) {
-    _has_bits_[0] |= 0x00000100u;
-  } else {
-    _has_bits_[0] &= ~0x00000100u;
-  }
-  pretokenization_delimiter_.SetAllocated(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), pretokenization_delimiter,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.TrainerSpec.pretokenization_delimiter)
-}
-
-// repeated string control_symbols = 30;
-inline int TrainerSpec::_internal_control_symbols_size() const {
-  return control_symbols_.size();
-}
-inline int TrainerSpec::control_symbols_size() const {
-  return _internal_control_symbols_size();
-}
-inline void TrainerSpec::clear_control_symbols() {
-  control_symbols_.Clear();
-}
-inline std::string* TrainerSpec::add_control_symbols() {
-  // @@protoc_insertion_point(field_add_mutable:sentencepiece.TrainerSpec.control_symbols)
-  return _internal_add_control_symbols();
-}
-inline const std::string& TrainerSpec::_internal_control_symbols(int index) const {
-  return control_symbols_.Get(index);
-}
-inline const std::string& TrainerSpec::control_symbols(int index) const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.control_symbols)
-  return _internal_control_symbols(index);
-}
-inline std::string* TrainerSpec::mutable_control_symbols(int index) {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.TrainerSpec.control_symbols)
-  return control_symbols_.Mutable(index);
-}
-inline void TrainerSpec::set_control_symbols(int index, const std::string& value) {
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.control_symbols)
-  control_symbols_.Mutable(index)->assign(value);
-}
-inline void TrainerSpec::set_control_symbols(int index, std::string&& value) {
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.control_symbols)
-  control_symbols_.Mutable(index)->assign(std::move(value));
-}
-inline void TrainerSpec::set_control_symbols(int index, const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  control_symbols_.Mutable(index)->assign(value);
-  // @@protoc_insertion_point(field_set_char:sentencepiece.TrainerSpec.control_symbols)
-}
-inline void TrainerSpec::set_control_symbols(int index, const char* value, size_t size) {
-  control_symbols_.Mutable(index)->assign(
-    reinterpret_cast<const char*>(value), size);
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.TrainerSpec.control_symbols)
-}
-inline std::string* TrainerSpec::_internal_add_control_symbols() {
-  return control_symbols_.Add();
-}
-inline void TrainerSpec::add_control_symbols(const std::string& value) {
-  control_symbols_.Add()->assign(value);
-  // @@protoc_insertion_point(field_add:sentencepiece.TrainerSpec.control_symbols)
-}
-inline void TrainerSpec::add_control_symbols(std::string&& value) {
-  control_symbols_.Add(std::move(value));
-  // @@protoc_insertion_point(field_add:sentencepiece.TrainerSpec.control_symbols)
-}
-inline void TrainerSpec::add_control_symbols(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  control_symbols_.Add()->assign(value);
-  // @@protoc_insertion_point(field_add_char:sentencepiece.TrainerSpec.control_symbols)
-}
-inline void TrainerSpec::add_control_symbols(const char* value, size_t size) {
-  control_symbols_.Add()->assign(reinterpret_cast<const char*>(value), size);
-  // @@protoc_insertion_point(field_add_pointer:sentencepiece.TrainerSpec.control_symbols)
-}
-inline const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>&
-TrainerSpec::control_symbols() const {
-  // @@protoc_insertion_point(field_list:sentencepiece.TrainerSpec.control_symbols)
-  return control_symbols_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>*
-TrainerSpec::mutable_control_symbols() {
-  // @@protoc_insertion_point(field_mutable_list:sentencepiece.TrainerSpec.control_symbols)
-  return &control_symbols_;
-}
-
-// repeated string user_defined_symbols = 31;
-inline int TrainerSpec::_internal_user_defined_symbols_size() const {
-  return user_defined_symbols_.size();
-}
-inline int TrainerSpec::user_defined_symbols_size() const {
-  return _internal_user_defined_symbols_size();
-}
-inline void TrainerSpec::clear_user_defined_symbols() {
-  user_defined_symbols_.Clear();
-}
-inline std::string* TrainerSpec::add_user_defined_symbols() {
-  // @@protoc_insertion_point(field_add_mutable:sentencepiece.TrainerSpec.user_defined_symbols)
-  return _internal_add_user_defined_symbols();
-}
-inline const std::string& TrainerSpec::_internal_user_defined_symbols(int index) const {
-  return user_defined_symbols_.Get(index);
-}
-inline const std::string& TrainerSpec::user_defined_symbols(int index) const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.user_defined_symbols)
-  return _internal_user_defined_symbols(index);
-}
-inline std::string* TrainerSpec::mutable_user_defined_symbols(int index) {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.TrainerSpec.user_defined_symbols)
-  return user_defined_symbols_.Mutable(index);
-}
-inline void TrainerSpec::set_user_defined_symbols(int index, const std::string& value) {
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.user_defined_symbols)
-  user_defined_symbols_.Mutable(index)->assign(value);
-}
-inline void TrainerSpec::set_user_defined_symbols(int index, std::string&& value) {
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.user_defined_symbols)
-  user_defined_symbols_.Mutable(index)->assign(std::move(value));
-}
-inline void TrainerSpec::set_user_defined_symbols(int index, const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  user_defined_symbols_.Mutable(index)->assign(value);
-  // @@protoc_insertion_point(field_set_char:sentencepiece.TrainerSpec.user_defined_symbols)
-}
-inline void TrainerSpec::set_user_defined_symbols(int index, const char* value, size_t size) {
-  user_defined_symbols_.Mutable(index)->assign(
-    reinterpret_cast<const char*>(value), size);
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.TrainerSpec.user_defined_symbols)
-}
-inline std::string* TrainerSpec::_internal_add_user_defined_symbols() {
-  return user_defined_symbols_.Add();
-}
-inline void TrainerSpec::add_user_defined_symbols(const std::string& value) {
-  user_defined_symbols_.Add()->assign(value);
-  // @@protoc_insertion_point(field_add:sentencepiece.TrainerSpec.user_defined_symbols)
-}
-inline void TrainerSpec::add_user_defined_symbols(std::string&& value) {
-  user_defined_symbols_.Add(std::move(value));
-  // @@protoc_insertion_point(field_add:sentencepiece.TrainerSpec.user_defined_symbols)
-}
-inline void TrainerSpec::add_user_defined_symbols(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  user_defined_symbols_.Add()->assign(value);
-  // @@protoc_insertion_point(field_add_char:sentencepiece.TrainerSpec.user_defined_symbols)
-}
-inline void TrainerSpec::add_user_defined_symbols(const char* value, size_t size) {
-  user_defined_symbols_.Add()->assign(reinterpret_cast<const char*>(value), size);
-  // @@protoc_insertion_point(field_add_pointer:sentencepiece.TrainerSpec.user_defined_symbols)
-}
-inline const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>&
-TrainerSpec::user_defined_symbols() const {
-  // @@protoc_insertion_point(field_list:sentencepiece.TrainerSpec.user_defined_symbols)
-  return user_defined_symbols_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField<std::string>*
-TrainerSpec::mutable_user_defined_symbols() {
-  // @@protoc_insertion_point(field_mutable_list:sentencepiece.TrainerSpec.user_defined_symbols)
-  return &user_defined_symbols_;
-}
-
-// optional string required_chars = 36;
-inline bool TrainerSpec::_internal_has_required_chars() const {
-  bool value = (_has_bits_[0] & 0x00000004u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_required_chars() const {
-  return _internal_has_required_chars();
-}
-inline void TrainerSpec::clear_required_chars() {
-  required_chars_.ClearToEmpty();
-  _has_bits_[0] &= ~0x00000004u;
-}
-inline const std::string& TrainerSpec::required_chars() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.required_chars)
-  return _internal_required_chars();
-}
-inline void TrainerSpec::set_required_chars(const std::string& value) {
-  _internal_set_required_chars(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.required_chars)
-}
-inline std::string* TrainerSpec::mutable_required_chars() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.TrainerSpec.required_chars)
-  return _internal_mutable_required_chars();
-}
-inline const std::string& TrainerSpec::_internal_required_chars() const {
-  return required_chars_.Get();
-}
-inline void TrainerSpec::_internal_set_required_chars(const std::string& value) {
-  _has_bits_[0] |= 0x00000004u;
-  required_chars_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, value, GetArena());
-}
-inline void TrainerSpec::set_required_chars(std::string&& value) {
-  _has_bits_[0] |= 0x00000004u;
-  required_chars_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.TrainerSpec.required_chars)
-}
-inline void TrainerSpec::set_required_chars(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000004u;
-  required_chars_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.TrainerSpec.required_chars)
-}
-inline void TrainerSpec::set_required_chars(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000004u;
-  required_chars_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.TrainerSpec.required_chars)
-}
-inline std::string* TrainerSpec::_internal_mutable_required_chars() {
-  _has_bits_[0] |= 0x00000004u;
-  return required_chars_.Mutable(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, GetArena());
-}
-inline std::string* TrainerSpec::release_required_chars() {
-  // @@protoc_insertion_point(field_release:sentencepiece.TrainerSpec.required_chars)
-  if (!_internal_has_required_chars()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000004u;
-  return required_chars_.ReleaseNonDefault(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), GetArena());
-}
-inline void TrainerSpec::set_allocated_required_chars(std::string* required_chars) {
-  if (required_chars != nullptr) {
-    _has_bits_[0] |= 0x00000004u;
-  } else {
-    _has_bits_[0] &= ~0x00000004u;
-  }
-  required_chars_.SetAllocated(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), required_chars,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.TrainerSpec.required_chars)
-}
-
-// optional bool byte_fallback = 35 [default = false];
-inline bool TrainerSpec::_internal_has_byte_fallback() const {
-  bool value = (_has_bits_[0] & 0x00040000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_byte_fallback() const {
-  return _internal_has_byte_fallback();
-}
-inline void TrainerSpec::clear_byte_fallback() {
-  byte_fallback_ = false;
-  _has_bits_[0] &= ~0x00040000u;
-}
-inline bool TrainerSpec::_internal_byte_fallback() const {
-  return byte_fallback_;
-}
-inline bool TrainerSpec::byte_fallback() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.byte_fallback)
-  return _internal_byte_fallback();
-}
-inline void TrainerSpec::_internal_set_byte_fallback(bool value) {
-  _has_bits_[0] |= 0x00040000u;
-  byte_fallback_ = value;
-}
-inline void TrainerSpec::set_byte_fallback(bool value) {
-  _internal_set_byte_fallback(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.byte_fallback)
-}
-
-// optional bool vocabulary_output_piece_score = 32 [default = true];
-inline bool TrainerSpec::_internal_has_vocabulary_output_piece_score() const {
-  bool value = (_has_bits_[1] & 0x00000020u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_vocabulary_output_piece_score() const {
-  return _internal_has_vocabulary_output_piece_score();
-}
-inline void TrainerSpec::clear_vocabulary_output_piece_score() {
-  vocabulary_output_piece_score_ = true;
-  _has_bits_[1] &= ~0x00000020u;
-}
-inline bool TrainerSpec::_internal_vocabulary_output_piece_score() const {
-  return vocabulary_output_piece_score_;
-}
-inline bool TrainerSpec::vocabulary_output_piece_score() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.vocabulary_output_piece_score)
-  return _internal_vocabulary_output_piece_score();
-}
-inline void TrainerSpec::_internal_set_vocabulary_output_piece_score(bool value) {
-  _has_bits_[1] |= 0x00000020u;
-  vocabulary_output_piece_score_ = value;
-}
-inline void TrainerSpec::set_vocabulary_output_piece_score(bool value) {
-  _internal_set_vocabulary_output_piece_score(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.vocabulary_output_piece_score)
-}
-
-// optional bool hard_vocab_limit = 33 [default = true];
-inline bool TrainerSpec::_internal_has_hard_vocab_limit() const {
-  bool value = (_has_bits_[1] & 0x00000040u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_hard_vocab_limit() const {
-  return _internal_has_hard_vocab_limit();
-}
-inline void TrainerSpec::clear_hard_vocab_limit() {
-  hard_vocab_limit_ = true;
-  _has_bits_[1] &= ~0x00000040u;
-}
-inline bool TrainerSpec::_internal_hard_vocab_limit() const {
-  return hard_vocab_limit_;
-}
-inline bool TrainerSpec::hard_vocab_limit() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.hard_vocab_limit)
-  return _internal_hard_vocab_limit();
-}
-inline void TrainerSpec::_internal_set_hard_vocab_limit(bool value) {
-  _has_bits_[1] |= 0x00000040u;
-  hard_vocab_limit_ = value;
-}
-inline void TrainerSpec::set_hard_vocab_limit(bool value) {
-  _internal_set_hard_vocab_limit(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.hard_vocab_limit)
-}
-
-// optional bool use_all_vocab = 34 [default = false];
-inline bool TrainerSpec::_internal_has_use_all_vocab() const {
-  bool value = (_has_bits_[0] & 0x00080000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_use_all_vocab() const {
-  return _internal_has_use_all_vocab();
-}
-inline void TrainerSpec::clear_use_all_vocab() {
-  use_all_vocab_ = false;
-  _has_bits_[0] &= ~0x00080000u;
-}
-inline bool TrainerSpec::_internal_use_all_vocab() const {
-  return use_all_vocab_;
-}
-inline bool TrainerSpec::use_all_vocab() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.use_all_vocab)
-  return _internal_use_all_vocab();
-}
-inline void TrainerSpec::_internal_set_use_all_vocab(bool value) {
-  _has_bits_[0] |= 0x00080000u;
-  use_all_vocab_ = value;
-}
-inline void TrainerSpec::set_use_all_vocab(bool value) {
-  _internal_set_use_all_vocab(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.use_all_vocab)
-}
-
-// optional int32 unk_id = 40 [default = 0];
-inline bool TrainerSpec::_internal_has_unk_id() const {
-  bool value = (_has_bits_[0] & 0x00200000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_unk_id() const {
-  return _internal_has_unk_id();
-}
-inline void TrainerSpec::clear_unk_id() {
-  unk_id_ = 0;
-  _has_bits_[0] &= ~0x00200000u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::_internal_unk_id() const {
-  return unk_id_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::unk_id() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.unk_id)
-  return _internal_unk_id();
-}
-inline void TrainerSpec::_internal_set_unk_id(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _has_bits_[0] |= 0x00200000u;
-  unk_id_ = value;
-}
-inline void TrainerSpec::set_unk_id(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _internal_set_unk_id(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.unk_id)
-}
-
-// optional int32 bos_id = 41 [default = 1];
-inline bool TrainerSpec::_internal_has_bos_id() const {
-  bool value = (_has_bits_[1] & 0x00000080u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_bos_id() const {
-  return _internal_has_bos_id();
-}
-inline void TrainerSpec::clear_bos_id() {
-  bos_id_ = 1;
-  _has_bits_[1] &= ~0x00000080u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::_internal_bos_id() const {
-  return bos_id_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::bos_id() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.bos_id)
-  return _internal_bos_id();
-}
-inline void TrainerSpec::_internal_set_bos_id(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _has_bits_[1] |= 0x00000080u;
-  bos_id_ = value;
-}
-inline void TrainerSpec::set_bos_id(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _internal_set_bos_id(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.bos_id)
-}
-
-// optional int32 eos_id = 42 [default = 2];
-inline bool TrainerSpec::_internal_has_eos_id() const {
-  bool value = (_has_bits_[1] & 0x00000100u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_eos_id() const {
-  return _internal_has_eos_id();
-}
-inline void TrainerSpec::clear_eos_id() {
-  eos_id_ = 2;
-  _has_bits_[1] &= ~0x00000100u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::_internal_eos_id() const {
-  return eos_id_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::eos_id() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.eos_id)
-  return _internal_eos_id();
-}
-inline void TrainerSpec::_internal_set_eos_id(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _has_bits_[1] |= 0x00000100u;
-  eos_id_ = value;
-}
-inline void TrainerSpec::set_eos_id(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _internal_set_eos_id(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.eos_id)
-}
-
-// optional int32 pad_id = 43 [default = -1];
-inline bool TrainerSpec::_internal_has_pad_id() const {
-  bool value = (_has_bits_[1] & 0x00000200u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_pad_id() const {
-  return _internal_has_pad_id();
-}
-inline void TrainerSpec::clear_pad_id() {
-  pad_id_ = -1;
-  _has_bits_[1] &= ~0x00000200u;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::_internal_pad_id() const {
-  return pad_id_;
-}
-inline ::PROTOBUF_NAMESPACE_ID::int32 TrainerSpec::pad_id() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.pad_id)
-  return _internal_pad_id();
-}
-inline void TrainerSpec::_internal_set_pad_id(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _has_bits_[1] |= 0x00000200u;
-  pad_id_ = value;
-}
-inline void TrainerSpec::set_pad_id(::PROTOBUF_NAMESPACE_ID::int32 value) {
-  _internal_set_pad_id(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.pad_id)
-}
-
-// optional string unk_piece = 45 [default = "<unk>"];
-inline bool TrainerSpec::_internal_has_unk_piece() const {
-  bool value = (_has_bits_[0] & 0x00000010u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_unk_piece() const {
-  return _internal_has_unk_piece();
-}
-inline void TrainerSpec::clear_unk_piece() {
-  unk_piece_.ClearToDefault(::sentencepiece::TrainerSpec::_i_give_permission_to_break_this_code_default_unk_piece_, GetArena());
-  _has_bits_[0] &= ~0x00000010u;
-}
-inline const std::string& TrainerSpec::unk_piece() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.unk_piece)
-  if (unk_piece_.IsDefault(nullptr)) return _i_give_permission_to_break_this_code_default_unk_piece_.get();
-  return _internal_unk_piece();
-}
-inline void TrainerSpec::set_unk_piece(const std::string& value) {
-  _internal_set_unk_piece(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.unk_piece)
-}
-inline std::string* TrainerSpec::mutable_unk_piece() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.TrainerSpec.unk_piece)
-  return _internal_mutable_unk_piece();
-}
-inline const std::string& TrainerSpec::_internal_unk_piece() const {
-  return unk_piece_.Get();
-}
-inline void TrainerSpec::_internal_set_unk_piece(const std::string& value) {
-  _has_bits_[0] |= 0x00000010u;
-  unk_piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, value, GetArena());
-}
-inline void TrainerSpec::set_unk_piece(std::string&& value) {
-  _has_bits_[0] |= 0x00000010u;
-  unk_piece_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.TrainerSpec.unk_piece)
-}
-inline void TrainerSpec::set_unk_piece(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000010u;
-  unk_piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.TrainerSpec.unk_piece)
-}
-inline void TrainerSpec::set_unk_piece(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000010u;
-  unk_piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.TrainerSpec.unk_piece)
-}
-inline std::string* TrainerSpec::_internal_mutable_unk_piece() {
-  _has_bits_[0] |= 0x00000010u;
-  return unk_piece_.Mutable(::sentencepiece::TrainerSpec::_i_give_permission_to_break_this_code_default_unk_piece_, GetArena());
-}
-inline std::string* TrainerSpec::release_unk_piece() {
-  // @@protoc_insertion_point(field_release:sentencepiece.TrainerSpec.unk_piece)
-  if (!_internal_has_unk_piece()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000010u;
-  return unk_piece_.ReleaseNonDefault(nullptr, GetArena());
-}
-inline void TrainerSpec::set_allocated_unk_piece(std::string* unk_piece) {
-  if (unk_piece != nullptr) {
-    _has_bits_[0] |= 0x00000010u;
-  } else {
-    _has_bits_[0] &= ~0x00000010u;
-  }
-  unk_piece_.SetAllocated(nullptr, unk_piece,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.TrainerSpec.unk_piece)
-}
-
-// optional string bos_piece = 46 [default = "<s>"];
-inline bool TrainerSpec::_internal_has_bos_piece() const {
-  bool value = (_has_bits_[0] & 0x00000020u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_bos_piece() const {
-  return _internal_has_bos_piece();
-}
-inline void TrainerSpec::clear_bos_piece() {
-  bos_piece_.ClearToDefault(::sentencepiece::TrainerSpec::_i_give_permission_to_break_this_code_default_bos_piece_, GetArena());
-  _has_bits_[0] &= ~0x00000020u;
-}
-inline const std::string& TrainerSpec::bos_piece() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.bos_piece)
-  if (bos_piece_.IsDefault(nullptr)) return _i_give_permission_to_break_this_code_default_bos_piece_.get();
-  return _internal_bos_piece();
-}
-inline void TrainerSpec::set_bos_piece(const std::string& value) {
-  _internal_set_bos_piece(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.bos_piece)
-}
-inline std::string* TrainerSpec::mutable_bos_piece() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.TrainerSpec.bos_piece)
-  return _internal_mutable_bos_piece();
-}
-inline const std::string& TrainerSpec::_internal_bos_piece() const {
-  return bos_piece_.Get();
-}
-inline void TrainerSpec::_internal_set_bos_piece(const std::string& value) {
-  _has_bits_[0] |= 0x00000020u;
-  bos_piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, value, GetArena());
-}
-inline void TrainerSpec::set_bos_piece(std::string&& value) {
-  _has_bits_[0] |= 0x00000020u;
-  bos_piece_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.TrainerSpec.bos_piece)
-}
-inline void TrainerSpec::set_bos_piece(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000020u;
-  bos_piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.TrainerSpec.bos_piece)
-}
-inline void TrainerSpec::set_bos_piece(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000020u;
-  bos_piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.TrainerSpec.bos_piece)
-}
-inline std::string* TrainerSpec::_internal_mutable_bos_piece() {
-  _has_bits_[0] |= 0x00000020u;
-  return bos_piece_.Mutable(::sentencepiece::TrainerSpec::_i_give_permission_to_break_this_code_default_bos_piece_, GetArena());
-}
-inline std::string* TrainerSpec::release_bos_piece() {
-  // @@protoc_insertion_point(field_release:sentencepiece.TrainerSpec.bos_piece)
-  if (!_internal_has_bos_piece()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000020u;
-  return bos_piece_.ReleaseNonDefault(nullptr, GetArena());
-}
-inline void TrainerSpec::set_allocated_bos_piece(std::string* bos_piece) {
-  if (bos_piece != nullptr) {
-    _has_bits_[0] |= 0x00000020u;
-  } else {
-    _has_bits_[0] &= ~0x00000020u;
-  }
-  bos_piece_.SetAllocated(nullptr, bos_piece,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.TrainerSpec.bos_piece)
-}
-
-// optional string eos_piece = 47 [default = "</s>"];
-inline bool TrainerSpec::_internal_has_eos_piece() const {
-  bool value = (_has_bits_[0] & 0x00000040u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_eos_piece() const {
-  return _internal_has_eos_piece();
-}
-inline void TrainerSpec::clear_eos_piece() {
-  eos_piece_.ClearToDefault(::sentencepiece::TrainerSpec::_i_give_permission_to_break_this_code_default_eos_piece_, GetArena());
-  _has_bits_[0] &= ~0x00000040u;
-}
-inline const std::string& TrainerSpec::eos_piece() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.eos_piece)
-  if (eos_piece_.IsDefault(nullptr)) return _i_give_permission_to_break_this_code_default_eos_piece_.get();
-  return _internal_eos_piece();
-}
-inline void TrainerSpec::set_eos_piece(const std::string& value) {
-  _internal_set_eos_piece(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.eos_piece)
-}
-inline std::string* TrainerSpec::mutable_eos_piece() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.TrainerSpec.eos_piece)
-  return _internal_mutable_eos_piece();
-}
-inline const std::string& TrainerSpec::_internal_eos_piece() const {
-  return eos_piece_.Get();
-}
-inline void TrainerSpec::_internal_set_eos_piece(const std::string& value) {
-  _has_bits_[0] |= 0x00000040u;
-  eos_piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, value, GetArena());
-}
-inline void TrainerSpec::set_eos_piece(std::string&& value) {
-  _has_bits_[0] |= 0x00000040u;
-  eos_piece_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.TrainerSpec.eos_piece)
-}
-inline void TrainerSpec::set_eos_piece(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000040u;
-  eos_piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.TrainerSpec.eos_piece)
-}
-inline void TrainerSpec::set_eos_piece(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000040u;
-  eos_piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.TrainerSpec.eos_piece)
-}
-inline std::string* TrainerSpec::_internal_mutable_eos_piece() {
-  _has_bits_[0] |= 0x00000040u;
-  return eos_piece_.Mutable(::sentencepiece::TrainerSpec::_i_give_permission_to_break_this_code_default_eos_piece_, GetArena());
-}
-inline std::string* TrainerSpec::release_eos_piece() {
-  // @@protoc_insertion_point(field_release:sentencepiece.TrainerSpec.eos_piece)
-  if (!_internal_has_eos_piece()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000040u;
-  return eos_piece_.ReleaseNonDefault(nullptr, GetArena());
-}
-inline void TrainerSpec::set_allocated_eos_piece(std::string* eos_piece) {
-  if (eos_piece != nullptr) {
-    _has_bits_[0] |= 0x00000040u;
-  } else {
-    _has_bits_[0] &= ~0x00000040u;
-  }
-  eos_piece_.SetAllocated(nullptr, eos_piece,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.TrainerSpec.eos_piece)
-}
-
-// optional string pad_piece = 48 [default = "<pad>"];
-inline bool TrainerSpec::_internal_has_pad_piece() const {
-  bool value = (_has_bits_[0] & 0x00000080u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_pad_piece() const {
-  return _internal_has_pad_piece();
-}
-inline void TrainerSpec::clear_pad_piece() {
-  pad_piece_.ClearToDefault(::sentencepiece::TrainerSpec::_i_give_permission_to_break_this_code_default_pad_piece_, GetArena());
-  _has_bits_[0] &= ~0x00000080u;
-}
-inline const std::string& TrainerSpec::pad_piece() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.pad_piece)
-  if (pad_piece_.IsDefault(nullptr)) return _i_give_permission_to_break_this_code_default_pad_piece_.get();
-  return _internal_pad_piece();
-}
-inline void TrainerSpec::set_pad_piece(const std::string& value) {
-  _internal_set_pad_piece(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.pad_piece)
-}
-inline std::string* TrainerSpec::mutable_pad_piece() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.TrainerSpec.pad_piece)
-  return _internal_mutable_pad_piece();
-}
-inline const std::string& TrainerSpec::_internal_pad_piece() const {
-  return pad_piece_.Get();
-}
-inline void TrainerSpec::_internal_set_pad_piece(const std::string& value) {
-  _has_bits_[0] |= 0x00000080u;
-  pad_piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, value, GetArena());
-}
-inline void TrainerSpec::set_pad_piece(std::string&& value) {
-  _has_bits_[0] |= 0x00000080u;
-  pad_piece_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.TrainerSpec.pad_piece)
-}
-inline void TrainerSpec::set_pad_piece(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000080u;
-  pad_piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.TrainerSpec.pad_piece)
-}
-inline void TrainerSpec::set_pad_piece(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000080u;
-  pad_piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.TrainerSpec.pad_piece)
-}
-inline std::string* TrainerSpec::_internal_mutable_pad_piece() {
-  _has_bits_[0] |= 0x00000080u;
-  return pad_piece_.Mutable(::sentencepiece::TrainerSpec::_i_give_permission_to_break_this_code_default_pad_piece_, GetArena());
-}
-inline std::string* TrainerSpec::release_pad_piece() {
-  // @@protoc_insertion_point(field_release:sentencepiece.TrainerSpec.pad_piece)
-  if (!_internal_has_pad_piece()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000080u;
-  return pad_piece_.ReleaseNonDefault(nullptr, GetArena());
-}
-inline void TrainerSpec::set_allocated_pad_piece(std::string* pad_piece) {
-  if (pad_piece != nullptr) {
-    _has_bits_[0] |= 0x00000080u;
-  } else {
-    _has_bits_[0] &= ~0x00000080u;
-  }
-  pad_piece_.SetAllocated(nullptr, pad_piece,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.TrainerSpec.pad_piece)
-}
-
-// optional string unk_surface = 44 [default = " \342\201\207 "];
-inline bool TrainerSpec::_internal_has_unk_surface() const {
-  bool value = (_has_bits_[0] & 0x00000008u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_unk_surface() const {
-  return _internal_has_unk_surface();
-}
-inline void TrainerSpec::clear_unk_surface() {
-  unk_surface_.ClearToDefault(::sentencepiece::TrainerSpec::_i_give_permission_to_break_this_code_default_unk_surface_, GetArena());
-  _has_bits_[0] &= ~0x00000008u;
-}
-inline const std::string& TrainerSpec::unk_surface() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.unk_surface)
-  if (unk_surface_.IsDefault(nullptr)) return _i_give_permission_to_break_this_code_default_unk_surface_.get();
-  return _internal_unk_surface();
-}
-inline void TrainerSpec::set_unk_surface(const std::string& value) {
-  _internal_set_unk_surface(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.unk_surface)
-}
-inline std::string* TrainerSpec::mutable_unk_surface() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.TrainerSpec.unk_surface)
-  return _internal_mutable_unk_surface();
-}
-inline const std::string& TrainerSpec::_internal_unk_surface() const {
-  return unk_surface_.Get();
-}
-inline void TrainerSpec::_internal_set_unk_surface(const std::string& value) {
-  _has_bits_[0] |= 0x00000008u;
-  unk_surface_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, value, GetArena());
-}
-inline void TrainerSpec::set_unk_surface(std::string&& value) {
-  _has_bits_[0] |= 0x00000008u;
-  unk_surface_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.TrainerSpec.unk_surface)
-}
-inline void TrainerSpec::set_unk_surface(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000008u;
-  unk_surface_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.TrainerSpec.unk_surface)
-}
-inline void TrainerSpec::set_unk_surface(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000008u;
-  unk_surface_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::NonEmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.TrainerSpec.unk_surface)
-}
-inline std::string* TrainerSpec::_internal_mutable_unk_surface() {
-  _has_bits_[0] |= 0x00000008u;
-  return unk_surface_.Mutable(::sentencepiece::TrainerSpec::_i_give_permission_to_break_this_code_default_unk_surface_, GetArena());
-}
-inline std::string* TrainerSpec::release_unk_surface() {
-  // @@protoc_insertion_point(field_release:sentencepiece.TrainerSpec.unk_surface)
-  if (!_internal_has_unk_surface()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000008u;
-  return unk_surface_.ReleaseNonDefault(nullptr, GetArena());
-}
-inline void TrainerSpec::set_allocated_unk_surface(std::string* unk_surface) {
-  if (unk_surface != nullptr) {
-    _has_bits_[0] |= 0x00000008u;
-  } else {
-    _has_bits_[0] &= ~0x00000008u;
-  }
-  unk_surface_.SetAllocated(nullptr, unk_surface,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.TrainerSpec.unk_surface)
-}
-
-// optional bool train_extremely_large_corpus = 49 [default = false];
-inline bool TrainerSpec::_internal_has_train_extremely_large_corpus() const {
-  bool value = (_has_bits_[0] & 0x00100000u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_train_extremely_large_corpus() const {
-  return _internal_has_train_extremely_large_corpus();
-}
-inline void TrainerSpec::clear_train_extremely_large_corpus() {
-  train_extremely_large_corpus_ = false;
-  _has_bits_[0] &= ~0x00100000u;
-}
-inline bool TrainerSpec::_internal_train_extremely_large_corpus() const {
-  return train_extremely_large_corpus_;
-}
-inline bool TrainerSpec::train_extremely_large_corpus() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.train_extremely_large_corpus)
-  return _internal_train_extremely_large_corpus();
-}
-inline void TrainerSpec::_internal_set_train_extremely_large_corpus(bool value) {
-  _has_bits_[0] |= 0x00100000u;
-  train_extremely_large_corpus_ = value;
-}
-inline void TrainerSpec::set_train_extremely_large_corpus(bool value) {
-  _internal_set_train_extremely_large_corpus(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.train_extremely_large_corpus)
-}
-
-// optional string seed_sentencepieces_file = 54 [default = ""];
-inline bool TrainerSpec::_internal_has_seed_sentencepieces_file() const {
-  bool value = (_has_bits_[0] & 0x00000200u) != 0;
-  return value;
-}
-inline bool TrainerSpec::has_seed_sentencepieces_file() const {
-  return _internal_has_seed_sentencepieces_file();
-}
-inline void TrainerSpec::clear_seed_sentencepieces_file() {
-  seed_sentencepieces_file_.ClearToEmpty();
-  _has_bits_[0] &= ~0x00000200u;
-}
-inline const std::string& TrainerSpec::seed_sentencepieces_file() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.TrainerSpec.seed_sentencepieces_file)
-  return _internal_seed_sentencepieces_file();
-}
-inline void TrainerSpec::set_seed_sentencepieces_file(const std::string& value) {
-  _internal_set_seed_sentencepieces_file(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.TrainerSpec.seed_sentencepieces_file)
-}
-inline std::string* TrainerSpec::mutable_seed_sentencepieces_file() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.TrainerSpec.seed_sentencepieces_file)
-  return _internal_mutable_seed_sentencepieces_file();
-}
-inline const std::string& TrainerSpec::_internal_seed_sentencepieces_file() const {
-  return seed_sentencepieces_file_.Get();
-}
-inline void TrainerSpec::_internal_set_seed_sentencepieces_file(const std::string& value) {
-  _has_bits_[0] |= 0x00000200u;
-  seed_sentencepieces_file_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, value, GetArena());
-}
-inline void TrainerSpec::set_seed_sentencepieces_file(std::string&& value) {
-  _has_bits_[0] |= 0x00000200u;
-  seed_sentencepieces_file_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.TrainerSpec.seed_sentencepieces_file)
-}
-inline void TrainerSpec::set_seed_sentencepieces_file(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000200u;
-  seed_sentencepieces_file_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.TrainerSpec.seed_sentencepieces_file)
-}
-inline void TrainerSpec::set_seed_sentencepieces_file(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000200u;
-  seed_sentencepieces_file_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.TrainerSpec.seed_sentencepieces_file)
-}
-inline std::string* TrainerSpec::_internal_mutable_seed_sentencepieces_file() {
-  _has_bits_[0] |= 0x00000200u;
-  return seed_sentencepieces_file_.Mutable(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, GetArena());
-}
-inline std::string* TrainerSpec::release_seed_sentencepieces_file() {
-  // @@protoc_insertion_point(field_release:sentencepiece.TrainerSpec.seed_sentencepieces_file)
-  if (!_internal_has_seed_sentencepieces_file()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000200u;
-  return seed_sentencepieces_file_.ReleaseNonDefault(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), GetArena());
-}
-inline void TrainerSpec::set_allocated_seed_sentencepieces_file(std::string* seed_sentencepieces_file) {
-  if (seed_sentencepieces_file != nullptr) {
-    _has_bits_[0] |= 0x00000200u;
-  } else {
-    _has_bits_[0] &= ~0x00000200u;
-  }
-  seed_sentencepieces_file_.SetAllocated(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), seed_sentencepieces_file,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.TrainerSpec.seed_sentencepieces_file)
-}
-
-// -------------------------------------------------------------------
-
-// NormalizerSpec
-
-// optional string name = 1;
-inline bool NormalizerSpec::_internal_has_name() const {
-  bool value = (_has_bits_[0] & 0x00000001u) != 0;
-  return value;
-}
-inline bool NormalizerSpec::has_name() const {
-  return _internal_has_name();
-}
-inline void NormalizerSpec::clear_name() {
-  name_.ClearToEmpty();
-  _has_bits_[0] &= ~0x00000001u;
-}
-inline const std::string& NormalizerSpec::name() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.NormalizerSpec.name)
-  return _internal_name();
-}
-inline void NormalizerSpec::set_name(const std::string& value) {
-  _internal_set_name(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.NormalizerSpec.name)
-}
-inline std::string* NormalizerSpec::mutable_name() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.NormalizerSpec.name)
-  return _internal_mutable_name();
-}
-inline const std::string& NormalizerSpec::_internal_name() const {
-  return name_.Get();
-}
-inline void NormalizerSpec::_internal_set_name(const std::string& value) {
-  _has_bits_[0] |= 0x00000001u;
-  name_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, value, GetArena());
-}
-inline void NormalizerSpec::set_name(std::string&& value) {
-  _has_bits_[0] |= 0x00000001u;
-  name_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.NormalizerSpec.name)
-}
-inline void NormalizerSpec::set_name(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000001u;
-  name_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.NormalizerSpec.name)
-}
-inline void NormalizerSpec::set_name(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000001u;
-  name_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.NormalizerSpec.name)
-}
-inline std::string* NormalizerSpec::_internal_mutable_name() {
-  _has_bits_[0] |= 0x00000001u;
-  return name_.Mutable(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, GetArena());
-}
-inline std::string* NormalizerSpec::release_name() {
-  // @@protoc_insertion_point(field_release:sentencepiece.NormalizerSpec.name)
-  if (!_internal_has_name()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000001u;
-  return name_.ReleaseNonDefault(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), GetArena());
-}
-inline void NormalizerSpec::set_allocated_name(std::string* name) {
-  if (name != nullptr) {
-    _has_bits_[0] |= 0x00000001u;
-  } else {
-    _has_bits_[0] &= ~0x00000001u;
-  }
-  name_.SetAllocated(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), name,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.NormalizerSpec.name)
-}
-
-// optional bytes precompiled_charsmap = 2;
-inline bool NormalizerSpec::_internal_has_precompiled_charsmap() const {
-  bool value = (_has_bits_[0] & 0x00000002u) != 0;
-  return value;
-}
-inline bool NormalizerSpec::has_precompiled_charsmap() const {
-  return _internal_has_precompiled_charsmap();
-}
-inline void NormalizerSpec::clear_precompiled_charsmap() {
-  precompiled_charsmap_.ClearToEmpty();
-  _has_bits_[0] &= ~0x00000002u;
-}
-inline const std::string& NormalizerSpec::precompiled_charsmap() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.NormalizerSpec.precompiled_charsmap)
-  return _internal_precompiled_charsmap();
-}
-inline void NormalizerSpec::set_precompiled_charsmap(const std::string& value) {
-  _internal_set_precompiled_charsmap(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.NormalizerSpec.precompiled_charsmap)
-}
-inline std::string* NormalizerSpec::mutable_precompiled_charsmap() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.NormalizerSpec.precompiled_charsmap)
-  return _internal_mutable_precompiled_charsmap();
-}
-inline const std::string& NormalizerSpec::_internal_precompiled_charsmap() const {
-  return precompiled_charsmap_.Get();
-}
-inline void NormalizerSpec::_internal_set_precompiled_charsmap(const std::string& value) {
-  _has_bits_[0] |= 0x00000002u;
-  precompiled_charsmap_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, value, GetArena());
-}
-inline void NormalizerSpec::set_precompiled_charsmap(std::string&& value) {
-  _has_bits_[0] |= 0x00000002u;
-  precompiled_charsmap_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.NormalizerSpec.precompiled_charsmap)
-}
-inline void NormalizerSpec::set_precompiled_charsmap(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000002u;
-  precompiled_charsmap_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.NormalizerSpec.precompiled_charsmap)
-}
-inline void NormalizerSpec::set_precompiled_charsmap(const void* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000002u;
-  precompiled_charsmap_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.NormalizerSpec.precompiled_charsmap)
-}
-inline std::string* NormalizerSpec::_internal_mutable_precompiled_charsmap() {
-  _has_bits_[0] |= 0x00000002u;
-  return precompiled_charsmap_.Mutable(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, GetArena());
-}
-inline std::string* NormalizerSpec::release_precompiled_charsmap() {
-  // @@protoc_insertion_point(field_release:sentencepiece.NormalizerSpec.precompiled_charsmap)
-  if (!_internal_has_precompiled_charsmap()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000002u;
-  return precompiled_charsmap_.ReleaseNonDefault(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), GetArena());
-}
-inline void NormalizerSpec::set_allocated_precompiled_charsmap(std::string* precompiled_charsmap) {
-  if (precompiled_charsmap != nullptr) {
-    _has_bits_[0] |= 0x00000002u;
-  } else {
-    _has_bits_[0] &= ~0x00000002u;
-  }
-  precompiled_charsmap_.SetAllocated(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), precompiled_charsmap,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.NormalizerSpec.precompiled_charsmap)
-}
-
-// optional bool add_dummy_prefix = 3 [default = true];
-inline bool NormalizerSpec::_internal_has_add_dummy_prefix() const {
-  bool value = (_has_bits_[0] & 0x00000008u) != 0;
-  return value;
-}
-inline bool NormalizerSpec::has_add_dummy_prefix() const {
-  return _internal_has_add_dummy_prefix();
-}
-inline void NormalizerSpec::clear_add_dummy_prefix() {
-  add_dummy_prefix_ = true;
-  _has_bits_[0] &= ~0x00000008u;
-}
-inline bool NormalizerSpec::_internal_add_dummy_prefix() const {
-  return add_dummy_prefix_;
-}
-inline bool NormalizerSpec::add_dummy_prefix() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.NormalizerSpec.add_dummy_prefix)
-  return _internal_add_dummy_prefix();
-}
-inline void NormalizerSpec::_internal_set_add_dummy_prefix(bool value) {
-  _has_bits_[0] |= 0x00000008u;
-  add_dummy_prefix_ = value;
-}
-inline void NormalizerSpec::set_add_dummy_prefix(bool value) {
-  _internal_set_add_dummy_prefix(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.NormalizerSpec.add_dummy_prefix)
-}
-
-// optional bool remove_extra_whitespaces = 4 [default = true];
-inline bool NormalizerSpec::_internal_has_remove_extra_whitespaces() const {
-  bool value = (_has_bits_[0] & 0x00000010u) != 0;
-  return value;
-}
-inline bool NormalizerSpec::has_remove_extra_whitespaces() const {
-  return _internal_has_remove_extra_whitespaces();
-}
-inline void NormalizerSpec::clear_remove_extra_whitespaces() {
-  remove_extra_whitespaces_ = true;
-  _has_bits_[0] &= ~0x00000010u;
-}
-inline bool NormalizerSpec::_internal_remove_extra_whitespaces() const {
-  return remove_extra_whitespaces_;
-}
-inline bool NormalizerSpec::remove_extra_whitespaces() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.NormalizerSpec.remove_extra_whitespaces)
-  return _internal_remove_extra_whitespaces();
-}
-inline void NormalizerSpec::_internal_set_remove_extra_whitespaces(bool value) {
-  _has_bits_[0] |= 0x00000010u;
-  remove_extra_whitespaces_ = value;
-}
-inline void NormalizerSpec::set_remove_extra_whitespaces(bool value) {
-  _internal_set_remove_extra_whitespaces(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.NormalizerSpec.remove_extra_whitespaces)
-}
-
-// optional bool escape_whitespaces = 5 [default = true];
-inline bool NormalizerSpec::_internal_has_escape_whitespaces() const {
-  bool value = (_has_bits_[0] & 0x00000020u) != 0;
-  return value;
-}
-inline bool NormalizerSpec::has_escape_whitespaces() const {
-  return _internal_has_escape_whitespaces();
-}
-inline void NormalizerSpec::clear_escape_whitespaces() {
-  escape_whitespaces_ = true;
-  _has_bits_[0] &= ~0x00000020u;
-}
-inline bool NormalizerSpec::_internal_escape_whitespaces() const {
-  return escape_whitespaces_;
-}
-inline bool NormalizerSpec::escape_whitespaces() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.NormalizerSpec.escape_whitespaces)
-  return _internal_escape_whitespaces();
-}
-inline void NormalizerSpec::_internal_set_escape_whitespaces(bool value) {
-  _has_bits_[0] |= 0x00000020u;
-  escape_whitespaces_ = value;
-}
-inline void NormalizerSpec::set_escape_whitespaces(bool value) {
-  _internal_set_escape_whitespaces(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.NormalizerSpec.escape_whitespaces)
-}
-
-// optional string normalization_rule_tsv = 6;
-inline bool NormalizerSpec::_internal_has_normalization_rule_tsv() const {
-  bool value = (_has_bits_[0] & 0x00000004u) != 0;
-  return value;
-}
-inline bool NormalizerSpec::has_normalization_rule_tsv() const {
-  return _internal_has_normalization_rule_tsv();
-}
-inline void NormalizerSpec::clear_normalization_rule_tsv() {
-  normalization_rule_tsv_.ClearToEmpty();
-  _has_bits_[0] &= ~0x00000004u;
-}
-inline const std::string& NormalizerSpec::normalization_rule_tsv() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.NormalizerSpec.normalization_rule_tsv)
-  return _internal_normalization_rule_tsv();
-}
-inline void NormalizerSpec::set_normalization_rule_tsv(const std::string& value) {
-  _internal_set_normalization_rule_tsv(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.NormalizerSpec.normalization_rule_tsv)
-}
-inline std::string* NormalizerSpec::mutable_normalization_rule_tsv() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.NormalizerSpec.normalization_rule_tsv)
-  return _internal_mutable_normalization_rule_tsv();
-}
-inline const std::string& NormalizerSpec::_internal_normalization_rule_tsv() const {
-  return normalization_rule_tsv_.Get();
-}
-inline void NormalizerSpec::_internal_set_normalization_rule_tsv(const std::string& value) {
-  _has_bits_[0] |= 0x00000004u;
-  normalization_rule_tsv_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, value, GetArena());
-}
-inline void NormalizerSpec::set_normalization_rule_tsv(std::string&& value) {
-  _has_bits_[0] |= 0x00000004u;
-  normalization_rule_tsv_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.NormalizerSpec.normalization_rule_tsv)
-}
-inline void NormalizerSpec::set_normalization_rule_tsv(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000004u;
-  normalization_rule_tsv_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.NormalizerSpec.normalization_rule_tsv)
-}
-inline void NormalizerSpec::set_normalization_rule_tsv(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000004u;
-  normalization_rule_tsv_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.NormalizerSpec.normalization_rule_tsv)
-}
-inline std::string* NormalizerSpec::_internal_mutable_normalization_rule_tsv() {
-  _has_bits_[0] |= 0x00000004u;
-  return normalization_rule_tsv_.Mutable(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, GetArena());
-}
-inline std::string* NormalizerSpec::release_normalization_rule_tsv() {
-  // @@protoc_insertion_point(field_release:sentencepiece.NormalizerSpec.normalization_rule_tsv)
-  if (!_internal_has_normalization_rule_tsv()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000004u;
-  return normalization_rule_tsv_.ReleaseNonDefault(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), GetArena());
-}
-inline void NormalizerSpec::set_allocated_normalization_rule_tsv(std::string* normalization_rule_tsv) {
-  if (normalization_rule_tsv != nullptr) {
-    _has_bits_[0] |= 0x00000004u;
-  } else {
-    _has_bits_[0] &= ~0x00000004u;
-  }
-  normalization_rule_tsv_.SetAllocated(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), normalization_rule_tsv,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.NormalizerSpec.normalization_rule_tsv)
-}
-
-// -------------------------------------------------------------------
-
-// SelfTestData_Sample
-
-// optional string input = 1;
-inline bool SelfTestData_Sample::_internal_has_input() const {
-  bool value = (_has_bits_[0] & 0x00000001u) != 0;
-  return value;
-}
-inline bool SelfTestData_Sample::has_input() const {
-  return _internal_has_input();
-}
-inline void SelfTestData_Sample::clear_input() {
-  input_.ClearToEmpty();
-  _has_bits_[0] &= ~0x00000001u;
-}
-inline const std::string& SelfTestData_Sample::input() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.SelfTestData.Sample.input)
-  return _internal_input();
-}
-inline void SelfTestData_Sample::set_input(const std::string& value) {
-  _internal_set_input(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.SelfTestData.Sample.input)
-}
-inline std::string* SelfTestData_Sample::mutable_input() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.SelfTestData.Sample.input)
-  return _internal_mutable_input();
-}
-inline const std::string& SelfTestData_Sample::_internal_input() const {
-  return input_.Get();
-}
-inline void SelfTestData_Sample::_internal_set_input(const std::string& value) {
-  _has_bits_[0] |= 0x00000001u;
-  input_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, value, GetArena());
-}
-inline void SelfTestData_Sample::set_input(std::string&& value) {
-  _has_bits_[0] |= 0x00000001u;
-  input_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.SelfTestData.Sample.input)
-}
-inline void SelfTestData_Sample::set_input(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000001u;
-  input_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.SelfTestData.Sample.input)
-}
-inline void SelfTestData_Sample::set_input(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000001u;
-  input_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.SelfTestData.Sample.input)
-}
-inline std::string* SelfTestData_Sample::_internal_mutable_input() {
-  _has_bits_[0] |= 0x00000001u;
-  return input_.Mutable(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, GetArena());
-}
-inline std::string* SelfTestData_Sample::release_input() {
-  // @@protoc_insertion_point(field_release:sentencepiece.SelfTestData.Sample.input)
-  if (!_internal_has_input()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000001u;
-  return input_.ReleaseNonDefault(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), GetArena());
-}
-inline void SelfTestData_Sample::set_allocated_input(std::string* input) {
-  if (input != nullptr) {
-    _has_bits_[0] |= 0x00000001u;
-  } else {
-    _has_bits_[0] &= ~0x00000001u;
-  }
-  input_.SetAllocated(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), input,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.SelfTestData.Sample.input)
-}
-
-// optional string expected = 2;
-inline bool SelfTestData_Sample::_internal_has_expected() const {
-  bool value = (_has_bits_[0] & 0x00000002u) != 0;
-  return value;
-}
-inline bool SelfTestData_Sample::has_expected() const {
-  return _internal_has_expected();
-}
-inline void SelfTestData_Sample::clear_expected() {
-  expected_.ClearToEmpty();
-  _has_bits_[0] &= ~0x00000002u;
-}
-inline const std::string& SelfTestData_Sample::expected() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.SelfTestData.Sample.expected)
-  return _internal_expected();
-}
-inline void SelfTestData_Sample::set_expected(const std::string& value) {
-  _internal_set_expected(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.SelfTestData.Sample.expected)
-}
-inline std::string* SelfTestData_Sample::mutable_expected() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.SelfTestData.Sample.expected)
-  return _internal_mutable_expected();
-}
-inline const std::string& SelfTestData_Sample::_internal_expected() const {
-  return expected_.Get();
-}
-inline void SelfTestData_Sample::_internal_set_expected(const std::string& value) {
-  _has_bits_[0] |= 0x00000002u;
-  expected_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, value, GetArena());
-}
-inline void SelfTestData_Sample::set_expected(std::string&& value) {
-  _has_bits_[0] |= 0x00000002u;
-  expected_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.SelfTestData.Sample.expected)
-}
-inline void SelfTestData_Sample::set_expected(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000002u;
-  expected_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.SelfTestData.Sample.expected)
-}
-inline void SelfTestData_Sample::set_expected(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000002u;
-  expected_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.SelfTestData.Sample.expected)
-}
-inline std::string* SelfTestData_Sample::_internal_mutable_expected() {
-  _has_bits_[0] |= 0x00000002u;
-  return expected_.Mutable(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, GetArena());
-}
-inline std::string* SelfTestData_Sample::release_expected() {
-  // @@protoc_insertion_point(field_release:sentencepiece.SelfTestData.Sample.expected)
-  if (!_internal_has_expected()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000002u;
-  return expected_.ReleaseNonDefault(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), GetArena());
-}
-inline void SelfTestData_Sample::set_allocated_expected(std::string* expected) {
-  if (expected != nullptr) {
-    _has_bits_[0] |= 0x00000002u;
-  } else {
-    _has_bits_[0] &= ~0x00000002u;
-  }
-  expected_.SetAllocated(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), expected,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.SelfTestData.Sample.expected)
-}
-
-// -------------------------------------------------------------------
-
-// SelfTestData
-
-// repeated .sentencepiece.SelfTestData.Sample samples = 1;
-inline int SelfTestData::_internal_samples_size() const {
-  return samples_.size();
-}
-inline int SelfTestData::samples_size() const {
-  return _internal_samples_size();
-}
-inline void SelfTestData::clear_samples() {
-  samples_.Clear();
-}
-inline ::sentencepiece::SelfTestData_Sample* SelfTestData::mutable_samples(int index) {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.SelfTestData.samples)
-  return samples_.Mutable(index);
-}
-inline ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::SelfTestData_Sample >*
-SelfTestData::mutable_samples() {
-  // @@protoc_insertion_point(field_mutable_list:sentencepiece.SelfTestData.samples)
-  return &samples_;
-}
-inline const ::sentencepiece::SelfTestData_Sample& SelfTestData::_internal_samples(int index) const {
-  return samples_.Get(index);
-}
-inline const ::sentencepiece::SelfTestData_Sample& SelfTestData::samples(int index) const {
-  // @@protoc_insertion_point(field_get:sentencepiece.SelfTestData.samples)
-  return _internal_samples(index);
-}
-inline ::sentencepiece::SelfTestData_Sample* SelfTestData::_internal_add_samples() {
-  return samples_.Add();
-}
-inline ::sentencepiece::SelfTestData_Sample* SelfTestData::add_samples() {
-  // @@protoc_insertion_point(field_add:sentencepiece.SelfTestData.samples)
-  return _internal_add_samples();
-}
-inline const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::SelfTestData_Sample >&
-SelfTestData::samples() const {
-  // @@protoc_insertion_point(field_list:sentencepiece.SelfTestData.samples)
-  return samples_;
-}
-
-// -------------------------------------------------------------------
-
-// ModelProto_SentencePiece
-
-// optional string piece = 1;
-inline bool ModelProto_SentencePiece::_internal_has_piece() const {
-  bool value = (_has_bits_[0] & 0x00000001u) != 0;
-  return value;
-}
-inline bool ModelProto_SentencePiece::has_piece() const {
-  return _internal_has_piece();
-}
-inline void ModelProto_SentencePiece::clear_piece() {
-  piece_.ClearToEmpty();
-  _has_bits_[0] &= ~0x00000001u;
-}
-inline const std::string& ModelProto_SentencePiece::piece() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.ModelProto.SentencePiece.piece)
-  return _internal_piece();
-}
-inline void ModelProto_SentencePiece::set_piece(const std::string& value) {
-  _internal_set_piece(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.ModelProto.SentencePiece.piece)
-}
-inline std::string* ModelProto_SentencePiece::mutable_piece() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.ModelProto.SentencePiece.piece)
-  return _internal_mutable_piece();
-}
-inline const std::string& ModelProto_SentencePiece::_internal_piece() const {
-  return piece_.Get();
-}
-inline void ModelProto_SentencePiece::_internal_set_piece(const std::string& value) {
-  _has_bits_[0] |= 0x00000001u;
-  piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, value, GetArena());
-}
-inline void ModelProto_SentencePiece::set_piece(std::string&& value) {
-  _has_bits_[0] |= 0x00000001u;
-  piece_.Set(
-    ::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::move(value), GetArena());
-  // @@protoc_insertion_point(field_set_rvalue:sentencepiece.ModelProto.SentencePiece.piece)
-}
-inline void ModelProto_SentencePiece::set_piece(const char* value) {
-  GOOGLE_DCHECK(value != nullptr);
-  _has_bits_[0] |= 0x00000001u;
-  piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(value), GetArena());
-  // @@protoc_insertion_point(field_set_char:sentencepiece.ModelProto.SentencePiece.piece)
-}
-inline void ModelProto_SentencePiece::set_piece(const char* value,
-    size_t size) {
-  _has_bits_[0] |= 0x00000001u;
-  piece_.Set(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, ::std::string(
-      reinterpret_cast<const char*>(value), size), GetArena());
-  // @@protoc_insertion_point(field_set_pointer:sentencepiece.ModelProto.SentencePiece.piece)
-}
-inline std::string* ModelProto_SentencePiece::_internal_mutable_piece() {
-  _has_bits_[0] |= 0x00000001u;
-  return piece_.Mutable(::PROTOBUF_NAMESPACE_ID::internal::ArenaStringPtr::EmptyDefault{}, GetArena());
-}
-inline std::string* ModelProto_SentencePiece::release_piece() {
-  // @@protoc_insertion_point(field_release:sentencepiece.ModelProto.SentencePiece.piece)
-  if (!_internal_has_piece()) {
-    return nullptr;
-  }
-  _has_bits_[0] &= ~0x00000001u;
-  return piece_.ReleaseNonDefault(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), GetArena());
-}
-inline void ModelProto_SentencePiece::set_allocated_piece(std::string* piece) {
-  if (piece != nullptr) {
-    _has_bits_[0] |= 0x00000001u;
-  } else {
-    _has_bits_[0] &= ~0x00000001u;
-  }
-  piece_.SetAllocated(&::PROTOBUF_NAMESPACE_ID::internal::GetEmptyStringAlreadyInited(), piece,
-      GetArena());
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.ModelProto.SentencePiece.piece)
-}
-
-// optional float score = 2;
-inline bool ModelProto_SentencePiece::_internal_has_score() const {
-  bool value = (_has_bits_[0] & 0x00000002u) != 0;
-  return value;
-}
-inline bool ModelProto_SentencePiece::has_score() const {
-  return _internal_has_score();
-}
-inline void ModelProto_SentencePiece::clear_score() {
-  score_ = 0;
-  _has_bits_[0] &= ~0x00000002u;
-}
-inline float ModelProto_SentencePiece::_internal_score() const {
-  return score_;
-}
-inline float ModelProto_SentencePiece::score() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.ModelProto.SentencePiece.score)
-  return _internal_score();
-}
-inline void ModelProto_SentencePiece::_internal_set_score(float value) {
-  _has_bits_[0] |= 0x00000002u;
-  score_ = value;
-}
-inline void ModelProto_SentencePiece::set_score(float value) {
-  _internal_set_score(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.ModelProto.SentencePiece.score)
-}
-
-// optional .sentencepiece.ModelProto.SentencePiece.Type type = 3 [default = NORMAL];
-inline bool ModelProto_SentencePiece::_internal_has_type() const {
-  bool value = (_has_bits_[0] & 0x00000004u) != 0;
-  return value;
-}
-inline bool ModelProto_SentencePiece::has_type() const {
-  return _internal_has_type();
-}
-inline void ModelProto_SentencePiece::clear_type() {
-  type_ = 1;
-  _has_bits_[0] &= ~0x00000004u;
-}
-inline ::sentencepiece::ModelProto_SentencePiece_Type ModelProto_SentencePiece::_internal_type() const {
-  return static_cast< ::sentencepiece::ModelProto_SentencePiece_Type >(type_);
-}
-inline ::sentencepiece::ModelProto_SentencePiece_Type ModelProto_SentencePiece::type() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.ModelProto.SentencePiece.type)
-  return _internal_type();
-}
-inline void ModelProto_SentencePiece::_internal_set_type(::sentencepiece::ModelProto_SentencePiece_Type value) {
-  assert(::sentencepiece::ModelProto_SentencePiece_Type_IsValid(value));
-  _has_bits_[0] |= 0x00000004u;
-  type_ = value;
-}
-inline void ModelProto_SentencePiece::set_type(::sentencepiece::ModelProto_SentencePiece_Type value) {
-  _internal_set_type(value);
-  // @@protoc_insertion_point(field_set:sentencepiece.ModelProto.SentencePiece.type)
-}
-
-// -------------------------------------------------------------------
-
-// ModelProto
-
-// repeated .sentencepiece.ModelProto.SentencePiece pieces = 1;
-inline int ModelProto::_internal_pieces_size() const {
-  return pieces_.size();
-}
-inline int ModelProto::pieces_size() const {
-  return _internal_pieces_size();
-}
-inline void ModelProto::clear_pieces() {
-  pieces_.Clear();
-}
-inline ::sentencepiece::ModelProto_SentencePiece* ModelProto::mutable_pieces(int index) {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.ModelProto.pieces)
-  return pieces_.Mutable(index);
-}
-inline ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::ModelProto_SentencePiece >*
-ModelProto::mutable_pieces() {
-  // @@protoc_insertion_point(field_mutable_list:sentencepiece.ModelProto.pieces)
-  return &pieces_;
-}
-inline const ::sentencepiece::ModelProto_SentencePiece& ModelProto::_internal_pieces(int index) const {
-  return pieces_.Get(index);
-}
-inline const ::sentencepiece::ModelProto_SentencePiece& ModelProto::pieces(int index) const {
-  // @@protoc_insertion_point(field_get:sentencepiece.ModelProto.pieces)
-  return _internal_pieces(index);
-}
-inline ::sentencepiece::ModelProto_SentencePiece* ModelProto::_internal_add_pieces() {
-  return pieces_.Add();
-}
-inline ::sentencepiece::ModelProto_SentencePiece* ModelProto::add_pieces() {
-  // @@protoc_insertion_point(field_add:sentencepiece.ModelProto.pieces)
-  return _internal_add_pieces();
-}
-inline const ::PROTOBUF_NAMESPACE_ID::RepeatedPtrField< ::sentencepiece::ModelProto_SentencePiece >&
-ModelProto::pieces() const {
-  // @@protoc_insertion_point(field_list:sentencepiece.ModelProto.pieces)
-  return pieces_;
-}
-
-// optional .sentencepiece.TrainerSpec trainer_spec = 2;
-inline bool ModelProto::_internal_has_trainer_spec() const {
-  bool value = (_has_bits_[0] & 0x00000001u) != 0;
-  PROTOBUF_ASSUME(!value || trainer_spec_ != nullptr);
-  return value;
-}
-inline bool ModelProto::has_trainer_spec() const {
-  return _internal_has_trainer_spec();
-}
-inline void ModelProto::clear_trainer_spec() {
-  if (trainer_spec_ != nullptr) trainer_spec_->Clear();
-  _has_bits_[0] &= ~0x00000001u;
-}
-inline const ::sentencepiece::TrainerSpec& ModelProto::_internal_trainer_spec() const {
-  const ::sentencepiece::TrainerSpec* p = trainer_spec_;
-  return p != nullptr ? *p : reinterpret_cast<const ::sentencepiece::TrainerSpec&>(
-      ::sentencepiece::_TrainerSpec_default_instance_);
-}
-inline const ::sentencepiece::TrainerSpec& ModelProto::trainer_spec() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.ModelProto.trainer_spec)
-  return _internal_trainer_spec();
-}
-inline void ModelProto::unsafe_arena_set_allocated_trainer_spec(
-    ::sentencepiece::TrainerSpec* trainer_spec) {
-  if (GetArena() == nullptr) {
-    delete reinterpret_cast<::PROTOBUF_NAMESPACE_ID::MessageLite*>(trainer_spec_);
-  }
-  trainer_spec_ = trainer_spec;
-  if (trainer_spec) {
-    _has_bits_[0] |= 0x00000001u;
-  } else {
-    _has_bits_[0] &= ~0x00000001u;
-  }
-  // @@protoc_insertion_point(field_unsafe_arena_set_allocated:sentencepiece.ModelProto.trainer_spec)
-}
-inline ::sentencepiece::TrainerSpec* ModelProto::release_trainer_spec() {
-  _has_bits_[0] &= ~0x00000001u;
-  ::sentencepiece::TrainerSpec* temp = trainer_spec_;
-  trainer_spec_ = nullptr;
-  if (GetArena() != nullptr) {
-    temp = ::PROTOBUF_NAMESPACE_ID::internal::DuplicateIfNonNull(temp);
-  }
-  return temp;
-}
-inline ::sentencepiece::TrainerSpec* ModelProto::unsafe_arena_release_trainer_spec() {
-  // @@protoc_insertion_point(field_release:sentencepiece.ModelProto.trainer_spec)
-  _has_bits_[0] &= ~0x00000001u;
-  ::sentencepiece::TrainerSpec* temp = trainer_spec_;
-  trainer_spec_ = nullptr;
-  return temp;
-}
-inline ::sentencepiece::TrainerSpec* ModelProto::_internal_mutable_trainer_spec() {
-  _has_bits_[0] |= 0x00000001u;
-  if (trainer_spec_ == nullptr) {
-    auto* p = CreateMaybeMessage<::sentencepiece::TrainerSpec>(GetArena());
-    trainer_spec_ = p;
-  }
-  return trainer_spec_;
-}
-inline ::sentencepiece::TrainerSpec* ModelProto::mutable_trainer_spec() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.ModelProto.trainer_spec)
-  return _internal_mutable_trainer_spec();
-}
-inline void ModelProto::set_allocated_trainer_spec(::sentencepiece::TrainerSpec* trainer_spec) {
-  ::PROTOBUF_NAMESPACE_ID::Arena* message_arena = GetArena();
-  if (message_arena == nullptr) {
-    delete trainer_spec_;
-  }
-  if (trainer_spec) {
-    ::PROTOBUF_NAMESPACE_ID::Arena* submessage_arena =
-      ::PROTOBUF_NAMESPACE_ID::Arena::GetArena(trainer_spec);
-    if (message_arena != submessage_arena) {
-      trainer_spec = ::PROTOBUF_NAMESPACE_ID::internal::GetOwnedMessage(
-          message_arena, trainer_spec, submessage_arena);
-    }
-    _has_bits_[0] |= 0x00000001u;
-  } else {
-    _has_bits_[0] &= ~0x00000001u;
-  }
-  trainer_spec_ = trainer_spec;
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.ModelProto.trainer_spec)
-}
-
-// optional .sentencepiece.NormalizerSpec normalizer_spec = 3;
-inline bool ModelProto::_internal_has_normalizer_spec() const {
-  bool value = (_has_bits_[0] & 0x00000002u) != 0;
-  PROTOBUF_ASSUME(!value || normalizer_spec_ != nullptr);
-  return value;
-}
-inline bool ModelProto::has_normalizer_spec() const {
-  return _internal_has_normalizer_spec();
-}
-inline void ModelProto::clear_normalizer_spec() {
-  if (normalizer_spec_ != nullptr) normalizer_spec_->Clear();
-  _has_bits_[0] &= ~0x00000002u;
-}
-inline const ::sentencepiece::NormalizerSpec& ModelProto::_internal_normalizer_spec() const {
-  const ::sentencepiece::NormalizerSpec* p = normalizer_spec_;
-  return p != nullptr ? *p : reinterpret_cast<const ::sentencepiece::NormalizerSpec&>(
-      ::sentencepiece::_NormalizerSpec_default_instance_);
-}
-inline const ::sentencepiece::NormalizerSpec& ModelProto::normalizer_spec() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.ModelProto.normalizer_spec)
-  return _internal_normalizer_spec();
-}
-inline void ModelProto::unsafe_arena_set_allocated_normalizer_spec(
-    ::sentencepiece::NormalizerSpec* normalizer_spec) {
-  if (GetArena() == nullptr) {
-    delete reinterpret_cast<::PROTOBUF_NAMESPACE_ID::MessageLite*>(normalizer_spec_);
-  }
-  normalizer_spec_ = normalizer_spec;
-  if (normalizer_spec) {
-    _has_bits_[0] |= 0x00000002u;
-  } else {
-    _has_bits_[0] &= ~0x00000002u;
-  }
-  // @@protoc_insertion_point(field_unsafe_arena_set_allocated:sentencepiece.ModelProto.normalizer_spec)
-}
-inline ::sentencepiece::NormalizerSpec* ModelProto::release_normalizer_spec() {
-  _has_bits_[0] &= ~0x00000002u;
-  ::sentencepiece::NormalizerSpec* temp = normalizer_spec_;
-  normalizer_spec_ = nullptr;
-  if (GetArena() != nullptr) {
-    temp = ::PROTOBUF_NAMESPACE_ID::internal::DuplicateIfNonNull(temp);
-  }
-  return temp;
-}
-inline ::sentencepiece::NormalizerSpec* ModelProto::unsafe_arena_release_normalizer_spec() {
-  // @@protoc_insertion_point(field_release:sentencepiece.ModelProto.normalizer_spec)
-  _has_bits_[0] &= ~0x00000002u;
-  ::sentencepiece::NormalizerSpec* temp = normalizer_spec_;
-  normalizer_spec_ = nullptr;
-  return temp;
-}
-inline ::sentencepiece::NormalizerSpec* ModelProto::_internal_mutable_normalizer_spec() {
-  _has_bits_[0] |= 0x00000002u;
-  if (normalizer_spec_ == nullptr) {
-    auto* p = CreateMaybeMessage<::sentencepiece::NormalizerSpec>(GetArena());
-    normalizer_spec_ = p;
-  }
-  return normalizer_spec_;
-}
-inline ::sentencepiece::NormalizerSpec* ModelProto::mutable_normalizer_spec() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.ModelProto.normalizer_spec)
-  return _internal_mutable_normalizer_spec();
-}
-inline void ModelProto::set_allocated_normalizer_spec(::sentencepiece::NormalizerSpec* normalizer_spec) {
-  ::PROTOBUF_NAMESPACE_ID::Arena* message_arena = GetArena();
-  if (message_arena == nullptr) {
-    delete normalizer_spec_;
-  }
-  if (normalizer_spec) {
-    ::PROTOBUF_NAMESPACE_ID::Arena* submessage_arena =
-      ::PROTOBUF_NAMESPACE_ID::Arena::GetArena(normalizer_spec);
-    if (message_arena != submessage_arena) {
-      normalizer_spec = ::PROTOBUF_NAMESPACE_ID::internal::GetOwnedMessage(
-          message_arena, normalizer_spec, submessage_arena);
-    }
-    _has_bits_[0] |= 0x00000002u;
-  } else {
-    _has_bits_[0] &= ~0x00000002u;
-  }
-  normalizer_spec_ = normalizer_spec;
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.ModelProto.normalizer_spec)
-}
-
-// optional .sentencepiece.SelfTestData self_test_data = 4;
-inline bool ModelProto::_internal_has_self_test_data() const {
-  bool value = (_has_bits_[0] & 0x00000004u) != 0;
-  PROTOBUF_ASSUME(!value || self_test_data_ != nullptr);
-  return value;
-}
-inline bool ModelProto::has_self_test_data() const {
-  return _internal_has_self_test_data();
-}
-inline void ModelProto::clear_self_test_data() {
-  if (self_test_data_ != nullptr) self_test_data_->Clear();
-  _has_bits_[0] &= ~0x00000004u;
-}
-inline const ::sentencepiece::SelfTestData& ModelProto::_internal_self_test_data() const {
-  const ::sentencepiece::SelfTestData* p = self_test_data_;
-  return p != nullptr ? *p : reinterpret_cast<const ::sentencepiece::SelfTestData&>(
-      ::sentencepiece::_SelfTestData_default_instance_);
-}
-inline const ::sentencepiece::SelfTestData& ModelProto::self_test_data() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.ModelProto.self_test_data)
-  return _internal_self_test_data();
-}
-inline void ModelProto::unsafe_arena_set_allocated_self_test_data(
-    ::sentencepiece::SelfTestData* self_test_data) {
-  if (GetArena() == nullptr) {
-    delete reinterpret_cast<::PROTOBUF_NAMESPACE_ID::MessageLite*>(self_test_data_);
-  }
-  self_test_data_ = self_test_data;
-  if (self_test_data) {
-    _has_bits_[0] |= 0x00000004u;
-  } else {
-    _has_bits_[0] &= ~0x00000004u;
-  }
-  // @@protoc_insertion_point(field_unsafe_arena_set_allocated:sentencepiece.ModelProto.self_test_data)
-}
-inline ::sentencepiece::SelfTestData* ModelProto::release_self_test_data() {
-  _has_bits_[0] &= ~0x00000004u;
-  ::sentencepiece::SelfTestData* temp = self_test_data_;
-  self_test_data_ = nullptr;
-  if (GetArena() != nullptr) {
-    temp = ::PROTOBUF_NAMESPACE_ID::internal::DuplicateIfNonNull(temp);
-  }
-  return temp;
-}
-inline ::sentencepiece::SelfTestData* ModelProto::unsafe_arena_release_self_test_data() {
-  // @@protoc_insertion_point(field_release:sentencepiece.ModelProto.self_test_data)
-  _has_bits_[0] &= ~0x00000004u;
-  ::sentencepiece::SelfTestData* temp = self_test_data_;
-  self_test_data_ = nullptr;
-  return temp;
-}
-inline ::sentencepiece::SelfTestData* ModelProto::_internal_mutable_self_test_data() {
-  _has_bits_[0] |= 0x00000004u;
-  if (self_test_data_ == nullptr) {
-    auto* p = CreateMaybeMessage<::sentencepiece::SelfTestData>(GetArena());
-    self_test_data_ = p;
-  }
-  return self_test_data_;
-}
-inline ::sentencepiece::SelfTestData* ModelProto::mutable_self_test_data() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.ModelProto.self_test_data)
-  return _internal_mutable_self_test_data();
-}
-inline void ModelProto::set_allocated_self_test_data(::sentencepiece::SelfTestData* self_test_data) {
-  ::PROTOBUF_NAMESPACE_ID::Arena* message_arena = GetArena();
-  if (message_arena == nullptr) {
-    delete self_test_data_;
-  }
-  if (self_test_data) {
-    ::PROTOBUF_NAMESPACE_ID::Arena* submessage_arena =
-      ::PROTOBUF_NAMESPACE_ID::Arena::GetArena(self_test_data);
-    if (message_arena != submessage_arena) {
-      self_test_data = ::PROTOBUF_NAMESPACE_ID::internal::GetOwnedMessage(
-          message_arena, self_test_data, submessage_arena);
-    }
-    _has_bits_[0] |= 0x00000004u;
-  } else {
-    _has_bits_[0] &= ~0x00000004u;
-  }
-  self_test_data_ = self_test_data;
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.ModelProto.self_test_data)
-}
-
-// optional .sentencepiece.NormalizerSpec denormalizer_spec = 5;
-inline bool ModelProto::_internal_has_denormalizer_spec() const {
-  bool value = (_has_bits_[0] & 0x00000008u) != 0;
-  PROTOBUF_ASSUME(!value || denormalizer_spec_ != nullptr);
-  return value;
-}
-inline bool ModelProto::has_denormalizer_spec() const {
-  return _internal_has_denormalizer_spec();
-}
-inline void ModelProto::clear_denormalizer_spec() {
-  if (denormalizer_spec_ != nullptr) denormalizer_spec_->Clear();
-  _has_bits_[0] &= ~0x00000008u;
-}
-inline const ::sentencepiece::NormalizerSpec& ModelProto::_internal_denormalizer_spec() const {
-  const ::sentencepiece::NormalizerSpec* p = denormalizer_spec_;
-  return p != nullptr ? *p : reinterpret_cast<const ::sentencepiece::NormalizerSpec&>(
-      ::sentencepiece::_NormalizerSpec_default_instance_);
-}
-inline const ::sentencepiece::NormalizerSpec& ModelProto::denormalizer_spec() const {
-  // @@protoc_insertion_point(field_get:sentencepiece.ModelProto.denormalizer_spec)
-  return _internal_denormalizer_spec();
-}
-inline void ModelProto::unsafe_arena_set_allocated_denormalizer_spec(
-    ::sentencepiece::NormalizerSpec* denormalizer_spec) {
-  if (GetArena() == nullptr) {
-    delete reinterpret_cast<::PROTOBUF_NAMESPACE_ID::MessageLite*>(denormalizer_spec_);
-  }
-  denormalizer_spec_ = denormalizer_spec;
-  if (denormalizer_spec) {
-    _has_bits_[0] |= 0x00000008u;
-  } else {
-    _has_bits_[0] &= ~0x00000008u;
-  }
-  // @@protoc_insertion_point(field_unsafe_arena_set_allocated:sentencepiece.ModelProto.denormalizer_spec)
-}
-inline ::sentencepiece::NormalizerSpec* ModelProto::release_denormalizer_spec() {
-  _has_bits_[0] &= ~0x00000008u;
-  ::sentencepiece::NormalizerSpec* temp = denormalizer_spec_;
-  denormalizer_spec_ = nullptr;
-  if (GetArena() != nullptr) {
-    temp = ::PROTOBUF_NAMESPACE_ID::internal::DuplicateIfNonNull(temp);
-  }
-  return temp;
-}
-inline ::sentencepiece::NormalizerSpec* ModelProto::unsafe_arena_release_denormalizer_spec() {
-  // @@protoc_insertion_point(field_release:sentencepiece.ModelProto.denormalizer_spec)
-  _has_bits_[0] &= ~0x00000008u;
-  ::sentencepiece::NormalizerSpec* temp = denormalizer_spec_;
-  denormalizer_spec_ = nullptr;
-  return temp;
-}
-inline ::sentencepiece::NormalizerSpec* ModelProto::_internal_mutable_denormalizer_spec() {
-  _has_bits_[0] |= 0x00000008u;
-  if (denormalizer_spec_ == nullptr) {
-    auto* p = CreateMaybeMessage<::sentencepiece::NormalizerSpec>(GetArena());
-    denormalizer_spec_ = p;
-  }
-  return denormalizer_spec_;
-}
-inline ::sentencepiece::NormalizerSpec* ModelProto::mutable_denormalizer_spec() {
-  // @@protoc_insertion_point(field_mutable:sentencepiece.ModelProto.denormalizer_spec)
-  return _internal_mutable_denormalizer_spec();
-}
-inline void ModelProto::set_allocated_denormalizer_spec(::sentencepiece::NormalizerSpec* denormalizer_spec) {
-  ::PROTOBUF_NAMESPACE_ID::Arena* message_arena = GetArena();
-  if (message_arena == nullptr) {
-    delete denormalizer_spec_;
-  }
-  if (denormalizer_spec) {
-    ::PROTOBUF_NAMESPACE_ID::Arena* submessage_arena =
-      ::PROTOBUF_NAMESPACE_ID::Arena::GetArena(denormalizer_spec);
-    if (message_arena != submessage_arena) {
-      denormalizer_spec = ::PROTOBUF_NAMESPACE_ID::internal::GetOwnedMessage(
-          message_arena, denormalizer_spec, submessage_arena);
-    }
-    _has_bits_[0] |= 0x00000008u;
-  } else {
-    _has_bits_[0] &= ~0x00000008u;
-  }
-  denormalizer_spec_ = denormalizer_spec;
-  // @@protoc_insertion_point(field_set_allocated:sentencepiece.ModelProto.denormalizer_spec)
-}
-
-#ifdef __GNUC__
-  #pragma GCC diagnostic pop
-#endif  // __GNUC__
-// -------------------------------------------------------------------
-
-// -------------------------------------------------------------------
-
-// -------------------------------------------------------------------
-
-// -------------------------------------------------------------------
-
-// -------------------------------------------------------------------
-
-
-// @@protoc_insertion_point(namespace_scope)
-
-}  // namespace sentencepiece
-
-PROTOBUF_NAMESPACE_OPEN
-
-template <> struct is_proto_enum< ::sentencepiece::TrainerSpec_ModelType> : ::std::true_type {};
-template <> struct is_proto_enum< ::sentencepiece::ModelProto_SentencePiece_Type> : ::std::true_type {};
-
-PROTOBUF_NAMESPACE_CLOSE
-
-// @@protoc_insertion_point(global_scope)
-
-#include <google/protobuf/port_undef.inc>
-#endif  // GOOGLE_PROTOBUF_INCLUDED_GOOGLE_PROTOBUF_INCLUDED_sentencepiece_5fmodel_2eproto
diff --git a/projects/llm_framework/include/sentencepiece/char_model.h b/projects/llm_framework/include/sentencepiece/char_model.h
deleted file mode 100644
index cd328750..00000000
--- a/projects/llm_framework/include/sentencepiece/char_model.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef CHAR_MODEL_H_
-#define CHAR_MODEL_H_
-
-#include "model_interface.h"
-#include "sentencepiece_model.pb.h"
-
-namespace sentencepiece {
-namespace character {
-
-// Tokenize text into character sequence
-class Model : public ModelInterface {
- public:
-  explicit Model(const ModelProto &model_proto);
-  ~Model() override;
-
-  EncodeResult Encode(absl::string_view normalized) const override;
-};
-}  // namespace character
-}  // namespace sentencepiece
-#endif  // CHAR_MODEL_H_
diff --git a/projects/llm_framework/include/sentencepiece/char_model_trainer.h b/projects/llm_framework/include/sentencepiece/char_model_trainer.h
deleted file mode 100644
index e563819a..00000000
--- a/projects/llm_framework/include/sentencepiece/char_model_trainer.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef CHAR_MODEL_TRAINER_H_
-#define CHAR_MODEL_TRAINER_H_
-
-#include "sentencepiece_model.pb.h"
-#include "trainer_interface.h"
-
-namespace sentencepiece {
-namespace character {
-
-// Trainer class for character model.
-class Trainer : public TrainerInterface {
- public:
-  Trainer(const TrainerSpec &trainer_spec,
-          const NormalizerSpec &normalizer_spec,
-          const NormalizerSpec &denormalizer_spec)
-      : TrainerInterface::TrainerInterface(trainer_spec, normalizer_spec,
-                                           denormalizer_spec) {}
-
-  util::Status Train() override;
-};
-}  // namespace character
-}  // namespace sentencepiece
-#endif  // CHAR_MODEL_TRAINER_H_
diff --git a/projects/llm_framework/include/sentencepiece/common.h b/projects/llm_framework/include/sentencepiece/common.h
deleted file mode 100644
index d7e61864..00000000
--- a/projects/llm_framework/include/sentencepiece/common.h
+++ /dev/null
@@ -1,177 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef COMMON_H_
-#define COMMON_H_
-
-#include <stdint.h>
-#include <stdlib.h>
-#include <string.h>
-
-#include <iostream>
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "config.h"
-#include "third_party/absl/strings/string_view.h"
-
-#if defined(_WIN32) && !defined(__CYGWIN__)
-#define OS_WIN
-#else
-#define OS_UNIX
-#endif
-
-#ifdef OS_WIN
-#ifndef NOMINMAX
-#define NOMINMAX
-#endif
-#include <windows.h>
-#endif
-
-typedef int8_t int8;
-typedef int16_t int16;
-typedef int32_t int32;
-typedef int64_t int64;
-typedef uint8_t uint8;
-typedef uint16_t uint16;
-typedef uint32_t char32;
-typedef uint32_t uint32;
-typedef uint64_t uint64;
-
-static constexpr uint32 kUnicodeError = 0xFFFD;
-
-template <typename T, size_t N>
-char (&ArraySizeHelper(T (&array)[N]))[N];
-
-#ifndef _MSC_VER
-template <typename T, size_t N>
-char (&ArraySizeHelper(const T (&array)[N]))[N];
-#endif  // !_MSC_VER
-
-#define arraysize(array) (sizeof(ArraySizeHelper(array)))
-
-#if defined(_FREEBSD)
-#include <sys/endian.h>
-#endif
-#if !defined(__APPLE__) && !defined(_WIN32) && !defined(_FREEBSD)
-#include <endian.h>
-#if BYTE_ORDER == __BIG_ENDIAN
-#define IS_BIG_ENDIAN
-#endif
-#endif
-
-namespace sentencepiece {
-namespace util {
-#ifndef OS_WIN
-inline uint32 Swap32(uint32 x) { return __builtin_bswap32(x); }
-#endif  // OS_WIN
-}  // namespace util
-
-namespace error {
-
-void Abort();
-void Exit(int code);
-void SetTestCounter(int c);
-void ResetTestMode();
-bool GetTestCounter();
-
-class Die {
- public:
-  explicit Die(bool die) : die_(die) {}
-  ~Die() {
-    std::cerr << std::endl;
-    if (die_) {
-      Abort();
-    }
-  }
-  int operator&(std::ostream &) { return 0; }
-
- private:
-  bool die_;
-};
-}  // namespace error
-
-namespace logging {
-enum LogSeverity {
-  LOG_INFO = 0,
-  LOG_WARNING = 1,
-  LOG_ERROR = 2,
-  LOG_FATAL = 3,
-  LOG_SEVERITY_SIZE = 4,
-};
-
-int GetMinLogLevel();
-void SetMinLogLevel(int v);
-
-inline const char *BaseName(const char *path) {
-#ifdef OS_WIN
-  const char *p = strrchr(path, '\\');
-#else
-  const char *p = strrchr(path, '/');
-#endif
-  if (p == nullptr) return path;
-  return p + 1;
-}
-}  // namespace logging
-}  // namespace sentencepiece
-
-#define LOG(severity)                                                        \
-  (::sentencepiece::logging::GetMinLogLevel() >                              \
-   ::sentencepiece::logging::LOG_##severity)                                 \
-      ? 0                                                                    \
-      : ::sentencepiece::error::Die(                                         \
-            ::sentencepiece::logging::LOG_##severity >=                      \
-            ::sentencepiece::logging::LOG_FATAL) &                           \
-            std::cerr << ::sentencepiece::logging::BaseName(__FILE__) << "(" \
-                      << __LINE__ << ") "                                    \
-                      << "LOG(" << #severity << ") "
-
-#define CHECK(condition)                                                      \
-  (condition) ? 0                                                             \
-              : ::sentencepiece::error::Die(true) &                           \
-                    std::cerr << ::sentencepiece::logging::BaseName(__FILE__) \
-                              << "(" << __LINE__ << ") [" << #condition       \
-                              << "] "
-
-#define CHECK_STREQ(a, b) CHECK_EQ(std::string(a), std::string(b))
-#define CHECK_EQ(a, b) CHECK((a) == (b))
-#define CHECK_NE(a, b) CHECK((a) != (b))
-#define CHECK_GE(a, b) CHECK((a) >= (b))
-#define CHECK_LE(a, b) CHECK((a) <= (b))
-#define CHECK_GT(a, b) CHECK((a) > (b))
-#define CHECK_LT(a, b) CHECK((a) < (b))
-
-#define FRIEND_TEST(a, b) friend class a##_Test_##b;
-
-#define CHECK_OK(expr)                         \
-  do {                                         \
-    const auto _status = expr;                 \
-    CHECK(_status.ok()) << _status.ToString(); \
-  } while (0)
-
-#define CHECK_NOT_OK(expr)                      \
-  do {                                          \
-    const auto _status = expr;                  \
-    CHECK(!_status.ok()) << _status.ToString(); \
-  } while (0)
-
-#define RETURN_IF_ERROR(expr)          \
-  do {                                 \
-    const auto _status = expr;         \
-    if (!_status.ok()) return _status; \
-  } while (0)
-
-#endif  // COMMON_H_
diff --git a/projects/llm_framework/include/sentencepiece/filesystem.h b/projects/llm_framework/include/sentencepiece/filesystem.h
deleted file mode 100644
index e572b4b6..00000000
--- a/projects/llm_framework/include/sentencepiece/filesystem.h
+++ /dev/null
@@ -1,59 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef FILESYSTEM_H_
-#define FILESYSTEM_H_
-
-#include <stdio.h>
-
-#include <fstream>
-#include <memory>
-#include <string>
-
-#include "common.h"
-#include "sentencepiece_processor.h"
-#include "third_party/absl/strings/string_view.h"
-
-namespace sentencepiece {
-namespace filesystem {
-class ReadableFile {
- public:
-  ReadableFile() {}
-  explicit ReadableFile(absl::string_view filename, bool is_binary = false) {}
-  virtual ~ReadableFile() {}
-
-  virtual util::Status status() const = 0;
-  virtual bool ReadLine(std::string *line) = 0;
-  virtual bool ReadAll(std::string *line) = 0;
-};
-
-class WritableFile {
- public:
-  WritableFile() {}
-  explicit WritableFile(absl::string_view filename, bool is_binary = false) {}
-  virtual ~WritableFile() {}
-
-  virtual util::Status status() const = 0;
-  virtual bool Write(absl::string_view text) = 0;
-  virtual bool WriteLine(absl::string_view text) = 0;
-};
-
-std::unique_ptr<ReadableFile> NewReadableFile(absl::string_view filename,
-                                              bool is_binary = false);
-std::unique_ptr<WritableFile> NewWritableFile(absl::string_view filename,
-                                              bool is_binary = false);
-
-}  // namespace filesystem
-}  // namespace sentencepiece
-#endif  // FILESYSTEM_H_
diff --git a/projects/llm_framework/include/sentencepiece/freelist.h b/projects/llm_framework/include/sentencepiece/freelist.h
deleted file mode 100644
index 8038048a..00000000
--- a/projects/llm_framework/include/sentencepiece/freelist.h
+++ /dev/null
@@ -1,90 +0,0 @@
-// Copyright 2018 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef FREELIST_H_
-#define FREELIST_H_
-
-#include <string.h>
-
-#include <vector>
-
-namespace sentencepiece {
-namespace model {
-
-// Simple FreeList that allocates a chunk of T at once.
-template <class T>
-class FreeList {
- public:
-  FreeList() = delete;
-  explicit FreeList(size_t chunk_size) : chunk_size_(chunk_size) {}
-  virtual ~FreeList() {
-    for (auto& chunk : freelist_) delete[] chunk;
-  }
-
-  // `Free` doesn't free the object but reuse the allocated memory chunks.
-  void Free() {
-    const int size = std::min<int>(chunk_index_ + 1, freelist_.size());
-    for (int i = 0; i < size; ++i) {
-      T* chunk = freelist_[i];
-      memset(static_cast<void*>(chunk), 0, sizeof(*chunk) * chunk_size_);
-    }
-    chunk_index_ = 0;
-    element_index_ = 0;
-  }
-
-  // Returns the number of allocated elements.
-  size_t size() const { return chunk_size_ * chunk_index_ + element_index_; }
-
-  void swap(FreeList<T>& other) {
-    std::swap(freelist_, other.freelist_);
-    std::swap(element_index_, other.element_index_);
-    std::swap(chunk_index_, other.chunk_index_);
-    std::swap(chunk_size_, other.chunk_size_);
-  }
-
-  // Returns the element as an array.
-  T* operator[](size_t index) const {
-    return freelist_[index / chunk_size_] + index % chunk_size_;
-  }
-
-  // Allocates new element.
-  T* Allocate() {
-    if (element_index_ >= chunk_size_) {
-      ++chunk_index_;
-      element_index_ = 0;
-    }
-
-    if (chunk_index_ == freelist_.size()) {
-      T* chunk = new T[chunk_size_];
-      memset(static_cast<void*>(chunk), 0, sizeof(*chunk) * chunk_size_);
-      freelist_.push_back(chunk);
-    }
-
-    T* result = freelist_[chunk_index_] + element_index_;
-    ++element_index_;
-
-    return result;
-  }
-
- private:
-  std::vector<T*> freelist_;
-
-  // The last element is stored at freelist_[chunk_index_][element_index_]
-  size_t element_index_ = 0;
-  size_t chunk_index_ = 0;
-  size_t chunk_size_ = 0;  // Do not modify except in swap()
-};
-}  // namespace model
-}  // namespace sentencepiece
-#endif  // FREELIST_H_
diff --git a/projects/llm_framework/include/sentencepiece/init.h b/projects/llm_framework/include/sentencepiece/init.h
deleted file mode 100644
index 5d1c4879..00000000
--- a/projects/llm_framework/include/sentencepiece/init.h
+++ /dev/null
@@ -1,62 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef INIT_H_
-#define INIT_H_
-
-#include "common.h"
-#include "third_party/absl/flags/flag.h"
-#include "third_party/absl/flags/parse.h"
-#include "third_party/absl/flags/usage.h"
-
-#ifdef _USE_EXTERNAL_PROTOBUF
-#include "google/protobuf/message_lite.h"
-#else
-#include "third_party/protobuf-lite/google/protobuf/message_lite.h"
-#endif
-
-ABSL_DECLARE_FLAG(int32, minloglevel);
-
-namespace sentencepiece {
-inline void ParseCommandLineFlags(const char *usage, int *argc, char ***argv,
-                                  bool remove_arg = true) {
-  absl::SetProgramUsageMessage(*argv[0]);
-  const auto unused_args = absl::ParseCommandLine(*argc, *argv);
-
-  if (remove_arg) {
-    char **argv_val = *argv;
-    *argv = argv_val = argv_val + *argc - unused_args.size();
-    std::copy(unused_args.begin(), unused_args.end(), argv_val);
-    *argc = static_cast<int>(unused_args.size());
-  }
-
-  logging::SetMinLogLevel(absl::GetFlag(FLAGS_minloglevel));
-}
-
-inline void ShutdownLibrary() {
-  google::protobuf::ShutdownProtobufLibrary();
-#ifdef HAS_ABSL_CLEANUP_FLAGS
-  absl::CleanupFlags();
-#endif
-}
-
-class ScopedResourceDestructor {
- public:
-  ScopedResourceDestructor() {}
-  ~ScopedResourceDestructor() { ShutdownLibrary(); }
-};
-
-}  // namespace sentencepiece
-
-#endif  // INIT_H_
diff --git a/projects/llm_framework/include/sentencepiece/model_factory.h b/projects/llm_framework/include/sentencepiece/model_factory.h
deleted file mode 100644
index 76abce72..00000000
--- a/projects/llm_framework/include/sentencepiece/model_factory.h
+++ /dev/null
@@ -1,31 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef MODEL_FACTORY_H_
-#define MODEL_FACTORY_H_
-
-#include <memory>
-
-#include "model_interface.h"
-#include "sentencepiece_model.pb.h"
-
-namespace sentencepiece {
-
-class ModelFactory {
- public:
-  // Creates Model instance from |model_proto|.
-  static std::unique_ptr<ModelInterface> Create(const ModelProto &model_proto);
-};
-}  // namespace sentencepiece
-#endif  // MODEL_FACTORY_H_
diff --git a/projects/llm_framework/include/sentencepiece/model_interface.h b/projects/llm_framework/include/sentencepiece/model_interface.h
deleted file mode 100644
index 06e92430..00000000
--- a/projects/llm_framework/include/sentencepiece/model_interface.h
+++ /dev/null
@@ -1,249 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef MODEL_INTERFACE_H_
-#define MODEL_INTERFACE_H_
-
-#include <memory>
-#include <set>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "common.h"
-#include "normalizer.h"
-#include "sentencepiece_model.pb.h"
-#include "sentencepiece_processor.h"
-#include "third_party/absl/container/flat_hash_map.h"
-#include "third_party/absl/strings/string_view.h"
-#include "third_party/darts_clone/darts.h"
-#include "util.h"
-
-namespace sentencepiece {
-
-// "_this_is_a_pen" => ["_this", "_is", "_a", "_pen"]
-std::vector<absl::string_view> SplitIntoWords(
-    absl::string_view text, bool treat_ws_as_suffix = false,
-    bool allow_ws_only_pieces = false);
-
-// Converts byte (0-255) to piece (e.g., 58 -> "<0x3A>").
-std::string ByteToPiece(unsigned char c);
-
-// Converts piece to byte (e.g., "<0x3A>" -> 58). Returns -1 if `piece` is not
-// a valid byte piece.
-int PieceToByte(absl::string_view piece);
-
-using EncodeResult = std::vector<std::pair<absl::string_view, int>>;
-using NBestEncodeResult = std::vector<std::pair<EncodeResult, float>>;
-
-class ModelProto;
-
-// Underlying model interface.
-// Given a normalized string, returns a sequence of sentence pieces with ids.
-class ModelInterface {
- public:
-  using PieceToIdMap = absl::flat_hash_map<absl::string_view, int>;
-  //                                           string_util::string_view_hash>;
-
-  absl::string_view unk_piece() const;
-  absl::string_view bos_piece() const;
-  absl::string_view eos_piece() const;
-  absl::string_view pad_piece() const;
-
-  // `model_proto` should not be deleted until ModelInterface is destroyed.
-  explicit ModelInterface(const ModelProto &model_proto);
-  ModelInterface() {}
-
-  virtual ~ModelInterface();
-
-  // Returns Status.
-  // Encode/Decode functions are valid only when status is OK.
-  virtual util::Status status() const { return status_; }
-
-  virtual const ModelProto &model_proto() const { return *model_proto_; }
-
-  virtual const normalizer::PrefixMatcher *prefix_matcher() const {
-    return matcher_.get();
-  }
-
-  // Given a normalized string, returns a sequence of sentence pieces with ids.
-  // The concatenation of pieces must be the same as `normalized`.
-  virtual EncodeResult Encode(absl::string_view normalized) const = 0;
-
-  // The same as above, but returns nbest result with score.
-  virtual NBestEncodeResult NBestEncode(absl::string_view normalized,
-                                        int nbest_size) const {
-    LOG(ERROR) << "Not implemented.";
-    return NBestEncodeResult();
-  }
-
-  virtual EncodeResult SampleEncode(absl::string_view normalized,
-                                    float alpha) const {
-    LOG(ERROR) << "Not implemented.";
-    return EncodeResult();
-  }
-
-  // Sample `samples` many tokenisations from the segmentation lattice
-  // If `wor` is true, the samples are taken without replacement, and the scores
-  // are the inclusion probabilities of the elements in the sample; otherwise
-  // the samples are taken with replacement and the scores are the log-probs of
-  // sample elements
-  // If `include_best` is true, the best tokenisation is always included in the
-  // sample, and the remaining elements are sampled excluding the best.
-  virtual NBestEncodeResult SampleEncodeAndScore(absl::string_view normalized,
-                                                 float alpha, int samples,
-                                                 bool wor,
-                                                 bool include_best) const {
-    LOG(ERROR) << "Not implemented.";
-    return {{EncodeResult(), 0.0}};
-  }
-
-  // Calculates the entropy of the segmentation lattice with inverse temperature
-  // `alpha`. Uses a novel dynamic program to calculate the entropy.
-  virtual float CalculateEntropy(absl::string_view normalized,
-                                 float alpha) const {
-    LOG(ERROR) << "Not implemented.";
-    return 0.0;
-  }
-
-  // Return true if SampleEncode returns a valid result.
-  virtual bool IsSampleEncodeAvailable() const { return false; }
-
-  // Return true if NBestEncode returns a valid result.
-  virtual bool IsNBestEncodeAvailable() const { return false; }
-
-  // Return true if SampleEncodeAndScore returns a valid result.
-  virtual bool IsSampleEncodeAndScoreAvailable() const { return false; }
-
-  // Return true if CalculateEntropy returns a valid result.
-  virtual bool IsCalculateEntropyAvailable() const { return false; }
-
-  // Returns the vocab id of `piece`.
-  // Returns UNK(0) if `piece` is unknown
-  virtual int PieceToId(absl::string_view piece) const;
-
-  // Returns the string representation of vocab with `id`.
-  // id must be 0 <= id < GetPieceSize().
-  virtual const std::string &IdToPiece(int id) const {
-    return model_proto_->pieces(id).piece();
-  }
-
-  // Returns the size of sentence pieces, which is the same
-  // as the size of vocabulary for NMT.
-  virtual int GetPieceSize() const {
-    if (!model_proto_) return 0;
-    return model_proto_->pieces_size();
-  }
-
-  // Returns the score of `id`.
-  // Score represents a log probability of the piece.
-  // We can roughly estimate the unigram frequency of the piece.
-  virtual float GetScore(int id) const {
-    return model_proto_->pieces(id).score();
-  }
-
-  // Returns true if `id` is unknown symbol.
-  virtual bool IsUnknown(int id) const {
-    return (model_proto_->pieces(id).type() ==
-            ModelProto::SentencePiece::UNKNOWN);
-  }
-
-  // Returns true if `id` is control symbol.
-  virtual bool IsControl(int id) const {
-    return (model_proto_->pieces(id).type() ==
-            ModelProto::SentencePiece::CONTROL);
-  }
-
-  // Returns true if `id` is unused symbol.
-  virtual bool IsUnused(int id) const {
-    return (model_proto_->pieces(id).type() ==
-            ModelProto::SentencePiece::UNUSED);
-  }
-
-  // Returns true if `id` is user defined symbol.
-  virtual bool IsUserDefined(int id) const {
-    return (model_proto_->pieces(id).type() ==
-            ModelProto::SentencePiece::USER_DEFINED);
-  }
-
-  // Returns true if `id` is byte symbol.
-  virtual bool IsByte(int id) const {
-    return (model_proto_->pieces(id).type() == ModelProto::SentencePiece::BYTE);
-  }
-
-  virtual bool ByteFallbackEnabled() const {
-    return model_proto_ && model_proto_->trainer_spec().byte_fallback();
-  }
-
-  // Verifies if the `expected` and `actual` outputs are equivalent. `expected`
-  // and `actual` are sentence pieces joined by space (` `). Normally it means
-  // that the two strings are identical. In some model, due to float rounding
-  // errors, the strings may not be identical, but they may be still equivalent
-  // provided their scores are close enough (by some espilon).
-  virtual bool VerifyOutputsEquivalent(absl::string_view expected,
-                                       absl::string_view actual) const {
-    return expected == actual;
-  }
-
- protected:
-  void InitializePieces();
-
-  // Non-virtual (inlined) implementation for faster execution.
-  inline float GetScoreInlined(int id) const {
-    return model_proto_->pieces(id).score();
-  }
-
-  inline bool IsUnknownInlined(int id) const {
-    return (model_proto_->pieces(id).type() ==
-            ModelProto::SentencePiece::UNKNOWN);
-  }
-
-  inline bool IsControlInlined(int id) const {
-    return (model_proto_->pieces(id).type() ==
-            ModelProto::SentencePiece::CONTROL);
-  }
-
-  inline bool IsUnusedInlined(int id) const {
-    return (model_proto_->pieces(id).type() ==
-            ModelProto::SentencePiece::UNUSED);
-  }
-
-  inline bool IsUserDefinedInlined(int id) const {
-    return (model_proto_->pieces(id).type() ==
-            ModelProto::SentencePiece::USER_DEFINED);
-  }
-
-  inline bool IsByteInlined(int id) const {
-    return (model_proto_->pieces(id).type() == ModelProto::SentencePiece::BYTE);
-  }
-
-  const ModelProto *model_proto_ = nullptr;
-
-  // PrefixMatcher for user defined symbols.
-  std::unique_ptr<normalizer::PrefixMatcher> matcher_;
-
-  // piece -> id map for normal pieces
-  PieceToIdMap pieces_;
-
-  // piece -> id map for control, unknown, and byte pieces
-  PieceToIdMap reserved_id_map_;
-
-  // unknown id.
-  int unk_id_ = 0;
-
-  // status.
-  util::Status status_;
-};
-}  // namespace sentencepiece
-#endif  // MODEL_INTERFACE_H_
diff --git a/projects/llm_framework/include/sentencepiece/normalization_rule.h b/projects/llm_framework/include/sentencepiece/normalization_rule.h
deleted file mode 100644
index 86272734..00000000
--- a/projects/llm_framework/include/sentencepiece/normalization_rule.h
+++ /dev/null
@@ -1,137363 +0,0 @@
-#ifndef NORMALIZATION_RULE_H_
-#define NORMALIZATION_RULE_H_
-#include <cstdio>
-namespace sentencepiece {
-namespace {
-
-struct BinaryBlob {
-  const char *name;
-  size_t size;
-  const char *data;
-};
-
-#if defined(_WIN32) && !defined(__CYGWIN__)
-constexpr unsigned long long int kNormalizationRules_blob_uint64[] = {
-    0x0000FC000002B800, 0x0003243D00031C3E, 0x0002E4CC00032C3C,
-    0x8000238200000DB8, 0x000005B80002DCCC, 0x0002E4CC80002376,
-    0x8000238600001DB8, 0x000248CC00010D91, 0x00014D8C000038CD,
-    0x800009A28000099F, 0x0000BD8800015D8F, 0x0001598A0000AD89,
-    0x0000C984800009A5, 0x0000D98680001E38, 0x000035800000A187,
-    0x0000258200003D81, 0x0000E5800000DD83, 0x80001E3000000D81,
-    0x80001E4800001D83, 0x0000FD8080001E40, 0x800009A800001581,
-    0x80000A4A00003189, 0x80000A5080001E60, 0x80001E7080001E58,
-    0x0000758180001E68, 0x0000718300006980, 0x000169A880000C99,
-    0x0000758080000BF1, 0x00005D8900006981, 0x00003D84000169A5,
-    0x00000584800009AB, 0x80001E2880000BEB, 0x800009AE000181A3,
-    0x80000C2A00000D81, 0x80000C3C80000BBB, 0x001184790011D07A,
-    0x0011407700100478, 0x0013F87500103076, 0x001358730013A474,
-    0x80000C4200131C72, 0x000DF06F0012EC70, 0x000C8C6D000C8C6E,
-    0x000C006B000CDC6C, 0x000FA069000C646A, 0x000F0467000FF868,
-    0x000E6865000F6466, 0x0009EC6300099464, 0x0008D0610009D062,
-    0x0000058280001E20, 0x80001E7880001E50, 0x0000198680001BB8,
-    0x000B48590008EC5A, 0x000B0057000B1858, 0x000AF85500054C56,
-    0x0005305300043454, 0x80000A5600050052, 0x0007004F0005DC50,
-    0x0007904D00076C4E, 0x0006084B0007D44C, 0x000140490001C84A,
-    0x0001F0470001E448, 0x0000A04500003446, 0x0000D84300009C44,
-    0x0002F0410000C842, 0x0002498180001BC0, 0x0003F58700024180,
-    0x80001BC8000218CC, 0x000258CC80001BD0, 0x80000A5C000248CD,
-    0x80000A6880000A62, 0x000130CC80000A6E, 0x00000581000134CD,
-    0x00045CCC80001BD8, 0x00006D8200006581, 0x000071810001C4CC,
-    0x80000A7480001BE0, 0x80001BE800007587, 0x80001C0080001BF8,
-    0x00005D8C80001BF0, 0x00046CCC800009B7, 0x00001D83000364CD,
-    0x00001D81800009BA, 0x80001EA000001D80, 0x000089A380001E98,
-    0x80001EA880001EB0, 0x0000098980001E90, 0x80001C1080001C08,
-    0x80000A8680000A80, 0x00005180800009C0, 0x0000658000005181,
-    0x0000C5B100007D81, 0x0000298280001E88, 0x000208CC000401A7,
-    0x80000C4880000A92, 0x80001EB880000C4E, 0x000005860001E187,
-    0x0000298180001C28, 0x000570CC80000A8C, 0x80001C2080001C18,
-    0x00016D8080001C30, 0x000544CC00019981, 0x80000C1E000220CD,
-    0x80000A980002ACCC, 0x80000A9E80001C38, 0x026520EF028A7CF0,
-    0x80000BFD80000AA4, 0x80000AB080000AAA, 0x80001C40026400EA,
-    0x80000C9080001C50, 0x0001F0CD80001C48, 0x024870E30001ECCC,
-    0x002858E1023874E2, 0x000144CC0027ECE0, 0x0007358200015CCD,
-    0x002430DB80001C58, 0x002494D980001C60, 0x800009C300240CD8,
-    0x0004A8CC002540D6, 0x0024D0D3000564CD, 0x002534D1800009C6,
-    0x0021C0CF002334D0, 0x0016C8CD0019DCCE, 0x001714CB800009C9,
-    0x80000ABC001788CA, 0x0017FCC700013CCC, 0x0014BCC5001284C6,
-    0x0014E8C3001464C4, 0x80000AC20015E4C2, 0x00065587000104CC,
-    0x0006B18700003D84, 0x000695810006B986, 0x0006BD8300066180,
-    0x800009BD00003D82, 0x0006C18F0006C18C, 0x0006E5890006758C,
-    0x80000A7A00068988, 0x80001E8080000AC8, 0x0007298C80000C9F,
-    0x0006BD9180000AD4, 0x0007058800071D8C, 0x00074981800009CC,
-    0x0000ED8100073D87, 0x0007218200075D82, 0x00073D8700074184,
-    0x0006E5A700074586, 0x0000CDA780001C70, 0x0006F1A380001EC0,
-    0x0000D9A380000BC1, 0x0007ADAD80000C54, 0x0006C1AD80000C5A,
-    0x80000ACE80001EC8, 0x80001C680007B1A8, 0x000755AE80000ADA,
-    0x0007698080001C78, 0x80000C0300077981, 0x0006B5B10007C5B0,
-    0x80000AE080001C80, 0x0003ACCC000769A7, 0x80001C88000214CD,
-    0x0007B5A7000791A3, 0x0006A18000003981, 0x0006C5820006B181,
-    0x0006E9840006D583, 0x0001258680000AE6, 0x00017D8800015187,
-    0x000278CC0001A989, 0x0001A58C00024CCD, 0x80001C9080000AF2,
-    0x000185810001A18F, 0x000005840001E191, 0x80000AEC80001C98,
-    0x80001CA800004181, 0x0000558C80001CA0, 0x000294CC00007181,
-    0x0000F5A30002E0CD, 0x80001CB00001A58C, 0x000171B180001CB8,
-    0x0000F18180001CC0, 0x80000C24000121A3, 0x000198CC0000C9B1,
-    0x000111A800024CCD, 0x800009CF80000AF8, 0x80000B0480001CC8,
-    0x80000AFE80001CD0, 0x000179B080001CE0, 0x000115A380001CD8,
-    0x0000418000009DA7, 0x000139A700007181, 0x00009D870000E5A3,
-    0x0005D4CD0000BDAD, 0x0001A4CC0005D0CC, 0x000081810000BCCD,
-    0x00039180800009D2, 0x000150CC0003A181, 0x0000FD910003A583,
-    0x0000818400002981, 0x800009D50003B187, 0x80001EE080001EE8,
-    0x0003818C80001EF8, 0x000E219B80001EF0, 0x80000CAB80001CE8,
-    0x000E5987000E4D84, 0x000071810000C586, 0x000E59830003A980,
-    0x80001CF0000C0182, 0x00008D8F0000858C, 0x0000C9890000DD84,
-    0x0000C58B000E6188, 0x80001D0080001CF8, 0x80000B130000B189,
-    0x0000B580000339A3, 0x80000CB70000BD81, 0x00008580000325A7,
-    0x000CE4CC00008D81, 0x80000CA500008183, 0x80001ED8000309AD,
-    0x0001F58280000B19, 0x00018984000301B1, 0x000051810000AD88,
-    0x000EF9A30000A980, 0x80000BC7800009B4, 0x0000D58180000C60,
-    0x80001F1080000C66, 0x00019D81000ED1A8, 0x00009D840000BD84,
-    0x0007C0CC00002D89, 0x00079CCC0007C4CD, 0x80001F080007D4CD,
-    0x0001D18080001F20, 0x80001F1800003D81, 0x80001F2800003183,
-    0x0000218180001F00, 0x80000C0F0001CD80, 0x0002B58000002181,
-    0x000210CC00016D81, 0x0000F1870003A8CD, 0x00001D8100002587,
-    0x000104CD80001D08, 0x0007A0CC000100CC, 0x80001D10000794CC,
-    0x0002E4CD80000B22, 0x80001D18000D84CC, 0x80000B2E000FA181,
-    0x80000C7280000C6C, 0x80000B2880001D28, 0x80001D4880001D30,
-    0x000DF4CC00009DA3, 0x80000B3A00023CCD, 0x0000058780001D38,
-    0x0000058780001D50, 0x000FA18880001D58, 0x000C20CC80000C84,
-    0x80001D6080000B40, 0x80001D6880000B52, 0x80000B4C80000C8A,
-    0x80001D7080001D78, 0x800009DB000F9581, 0x80000B0D000F7981,
-    0x80000CB1800009E1, 0x80000BA3800009DE, 0x80001ED0000F9587,
-    0x80001D2080000C09, 0x000FA18C80000B34, 0x80001D4080000B46,
-    0x00006D81000FA98F, 0x800009E4000FDD91, 0x000FB187000FF582,
-    0x000FE587800009E7, 0x000368CC80000B58, 0x80001D9800031CCD,
-    0x80001DA080000B5E, 0x80000B640000418C, 0x800009EA000F998C,
-    0x000F71B100002D84, 0x000031810000A5A3, 0x80000BE500000D80,
-    0x80000BD9000F3DA7, 0x80000BD30000098C, 0x0000298080000BDF,
-    0x80001F3800001581, 0x000F29A380000B6A, 0x0000F1A3000F61B1,
-    0x000F21A780000B70, 0x000F05A7000F39A6, 0x80000BCD000F09A6,
-    0x80000C7E80000C78, 0x80001F4880000BAC, 0x000F0DAD80001F40,
-    0x0001558280001F58, 0x0001058000015983, 0x0001698600014581,
-    0x0001758480001F50, 0x00019D8A80001F60, 0x0001A9880001E98B,
-    0x80001F3000019589, 0x0001D98C0001D18F, 0x0001B18080001D80,
-    0x80000B760001B181, 0x0001A5810001A591, 0x0001A5830001A580,
-    0x80001D8880001D90, 0x000DF8CD0001899B, 0x00006589000DCCCC,
-    0x000175800000FDA3, 0x80001DA800013581, 0x80001DB00000B5A3,
-    0x000194CC80001DB8, 0x0000A5A4000254CD, 0x80000B7C80001DC0,
-    0x0000E1A880001DD0, 0x80001DD880001DC8, 0x80001DE0000264CC,
-    0x00004D800000E1AD, 0x0000A9B000005981, 0x80001DE800007587,
-    0x0000C5A380001F68, 0x0001D0CC800009ED, 0x80000B820001A4CD,
-    0x0000218880001F80, 0x80000CBD00007583, 0x00039D8180001DF0,
-    0x80000B8800038980, 0x000019880003AD82, 0x0003BD8700000989,
-    0x0003D98480001F78, 0x80001F7000038988, 0x0003E5800003D187,
-    0x0003D1820003E981, 0x000395800003D183, 0x00022CCC00039981,
-    0x80000B8B0002C8CD, 0x80001DF800003D8C, 0x00003181800009D8,
-    0x80000B9780000B91, 0x80001E0000003582, 0x00001D8780001E08,
-    0x0000D181800009F0, 0x000339A300007184, 0x00005189800009F3,
-    0x000400CC0001DD81, 0x0000D5B10003CCCD, 0x80001E3480001E3C,
-    0x0000198180001E4C, 0x0000158300001180, 0x80001E44000095A3,
-    0x00006D80800009F9, 0x00003D8300006581, 0x00003D8180000A4D,
-    0x80001E6400003D80, 0x000099A380001E5C, 0x80001E6C80001E74,
-    0x0000098980000BF4, 0x0001FD8480000BEE, 0x000EDCCC000ED8CC,
-    0x000FD8CC80001E2C, 0x000E08CC000FCCCC, 0x000F28CC000E58CC,
-    0x000E08CC000F58CC, 0x000FDCCC000E10CD, 0x000FECCC000F24CD,
-    0x0001A180000FE8CD, 0x80000C2D0001A181, 0x00012D8080000BBE,
-    0x80000C3F0001D981, 0x80000C45000EACCC, 0x000E90CD00001582,
-    0x80001E54000E94CC, 0x0000F18700000986, 0x80001BBC80001E7C,
-    0x80001BC480000A59, 0x80001BD480001BCC, 0x80000A5F0002F0CC,
-    0x000031B180000A65, 0x000268CD80000A6B, 0x000124CC0005E8CC,
-    0x80000A71000104CD, 0x00007D8180001BDC, 0x000580CC00005D81,
-    0x0007218280001BE4, 0x000098CD80000A77, 0x80001BEC000088CC,
-    0x000065A380001BFC, 0x80001BF480001C04, 0x0007018680000A08,
-    0x800009F680000A0B, 0x80000A5380001EA4, 0x80000C9C00078DA5,
-    0x800009FC000009A3, 0x800009FF80001E24, 0x80000A0580001E9C,
-    0x000334CC80001EB4, 0x0007BDA800037CCD, 0x0007D98180000A0E,
-    0x80001EAC00000D89, 0x80001E940007D182, 0x0007FD8700079987,
-    0x0000658080000A7D, 0x80001C0C00004181, 0x80001C1400004183,
-    0x0007AD8C0007CD8C, 0x0000898680000A83, 0x00065D8400009587,
-    0x0000918200077DB1, 0x0006E58000065183, 0x80000A8900061181,
-    0x0007E58C0007FD8F, 0x0000918A80000A11, 0x0000A18880001E8C,
-    0x80000A950007B189, 0x0000CD80000725A3, 0x80000C4B0000E981,
-    0x0000F1A7000731A7, 0x000089800007BD91, 0x80000C5100009581,
-    0x80001EBC80001E84, 0x80001C2C80000CA2, 0x0006FD81000711AD,
-    0x000159830006F180, 0x80000A8F00013182, 0x0001B98700014984,
-    0x0001958900016186, 0x80001C1C00018988, 0x000210CC80001C24,
-    0x0001E98F0001F58C, 0x0001A99100007987, 0x0006BD8080001C34,
-    0x0002C4CD0006B181, 0x80000C21000238CC, 0x80000A9B0000D98C,
-    0x0000318100000582, 0x0000F18780001C3C, 0x80000AA100000D86,
-    0x80000AA700001184, 0x000165A380000C00, 0x0000C98180000AAD,
-    0x00016DA7000198CC, 0x80001C4480000AB3, 0x80001C540000B5A8,
-    0x000099AD80000C93, 0x80001C5C80001C4C, 0x80001E1080001C64,
-    0x000184CC0000E9B0, 0x80000A140001FCCD, 0x80000A1A80000A17,
-    0x80000AC580000ABF, 0x80000A1D80000ACB, 0x00007DA700001981,
-    0x80001EC480001C74, 0x00028CCC80000BC4, 0x00001D9100039982,
-    0x0003B18780000C57, 0x80001ECC80000C5D, 0x80001C6C80000AD1,
-    0x80000C1B80000ADD, 0x0003E18000038588, 0x80001C7C0003E181,
-    0x00039D830003918C, 0x00039D8100039D82, 0x80000C0600039D80,
-    0x80001C840003B186, 0x00032CCC0003BD84, 0x000391890002F4CD,
-    0x00007D8F00038588, 0x80001C8C80000AE3, 0x000335A30003818C,
-    0x80000AE90000A182, 0x000329A70000ADB0, 0x000031840000AD81,
-    0x0003F4CC80000AF5, 0x80001C9C00026CCD, 0x80001CAC80000AEF,
-    0x80001CA4000309AE, 0x0000D9A30000958C, 0x000309B180001CB4,
-    0x80001CBC00004581, 0x000130CC80001CC4, 0x00002581000134CD,
-    0x80000AFB80000C27, 0x000214CD0000F1A8, 0x001CF8CC0003B4CC,
-    0x0001E5800003FCCD, 0x0001C0CD0001ED81, 0x000780CC00016CCC,
-    0x0006DCCC000784CD, 0x0007B4CC0006D8CD, 0x000638CC80000A23,
-    0x000784CC000640CC, 0x0007E4CC0007FCCD, 0x80001CCC00067CCD,
-    0x0001718100018181, 0x0001D98180000B07, 0x80000B0180001CD4,
-    0x80001CE400018D87, 0x000188CD0001F98C, 0x80001CDC00018CCC,
-    0x80000A2980000A26, 0x001DCCCC00012D8C, 0x0001F1B1001D5CCD,
-    0x000129B180001EEC, 0x0000158000003183, 0x80001EE400003D81,
-    0x80001EF480001EFC, 0x00003D8080001CEC, 0x80000CAE00001581,
-    0x000199A300001589, 0x000165A7000624CC, 0x0001A5A7000119A3,
-    0x001C01A380001CF4, 0x000151AD000604CC, 0x0000458180001CFC,
-    0x0000158080001D04, 0x80000B1600001981, 0x001EE58080000CBA,
-    0x00007191001EED81, 0x80000A2C001F2983, 0x80000A2F001FE984,
-    0x0000A59B001F5D87, 0x80000CB480000B10, 0x001F7D8C80000A32,
-    0x80000CA800000D84, 0x0000598380001EDC, 0x001F318100005582,
-    0x00007987001F3980, 0x80000B1C001FAD86, 0x0000398B00007984,
-    0x0000558980000BCA, 0x0000058F00004188, 0x80000C6980000C63,
-    0x001F2D810000198C, 0x80001F1480000BA6, 0x80001F0C001FD9A3,
-    0x80001F24001F2D84, 0x80001F1C001FC5A7, 0x80001F2C001F7188,
-    0x80001F0480001ED4, 0x00003180001FD5AD, 0x00006DA300002D81,
-    0x00007982001FB5B1, 0x80000C1280000C0C, 0x001FAD8000000584,
-    0x0003ECCC001FA581, 0x001C60CC000218CD, 0x80001D0C000021A8,
-    0x0000098180001D14, 0x0001898080000B25, 0x0001F98100019581,
-    0x0001E58700019583, 0x000248CC00002591, 0x80001D1C000250CD,
-    0x80000B3100018589, 0x80000C7580000C6F, 0x0000058480001D24,
-    0x80000B2B80001D2C, 0x0001B58180001D34, 0x0000D98C00006D87,
-    0x000238CC0001A981, 0x0000498C0002E8CD, 0x0000E98200001581,
-    0x80000B370000418F, 0x80001D4C00000D87, 0x80000B3D0000F987,
-    0x80000B4980001D3C, 0x00000987000125A3, 0x80001D540000D1B1,
-    0x0000058780001D44, 0x80000C8780001D5C, 0x0000C18180000B43,
-    0x0001A4CC80001D64, 0x80000B5580001E14, 0x80000C8D80001D6C,
-    0x80000B4F00008DA7, 0x80001D7480001D7C, 0x0001BCCC0000E9A3,
-    0x80000A380000DCCD, 0x80000A3B000071A3, 0x000051A680000A3E,
-    0x0003B587000051A7, 0x0000058100012CCC, 0x0000058880001D9C,
-    0x0000CD9180001DA4, 0x80000BE880000B67, 0x000390CC000328CC,
-    0x0006219B0003ED8C, 0x80000BD680000BDC, 0x000128CC00038188,
-    0x80000BE200005986, 0x0004018300064D84, 0x0003A9810003A182,
-    0x0000A98F0003B580, 0x0001918180001F3C, 0x0000C18B0000D18C,
-    0x00001D890000ED8A, 0x000351A700067588, 0x0000BD8100034DA6,
-    0x000355A30000AD80, 0x80000B6D0000DD8C, 0x000095810001ADB0,
-    0x00037DAD00008580, 0x80000B73000110CC, 0x000134CC00008D84,
-    0x00001D8080000BD0, 0x80000C7B00002581, 0x000025A3000035A4,
-    0x000301B180000C81, 0x80001F4480001F4C, 0x0001E1AD80001F5C,
-    0x80001F6480001F54, 0x80001D8480001F34, 0x000035830001F9A8,
-    0x80001D9480000B79, 0x001EE0CD0007C581, 0x00011CCC001EE4CC,
-    0x001EB8CC000138CC, 0x0001ACCC001E94CD, 0x80001D8C001EC0CD,
-    0x0002D18080001DAC, 0x80001DB40002CD81, 0x000195800001A183,
-    0x000574CC0001AD81, 0x80001DBC0002D8CD, 0x00079D8180001DC4,
-    0x00067D8180000B7F, 0x80001DD400018189, 0x80001E1880001DCC,
-    0x0000C1A380001DDC, 0x0000D58180001DE4, 0x80001DEC0000D580,
-    0x0000098880001F6C, 0x80000A4400015CCC, 0x000244CD00002187,
-    0x000108CD0005ECCC, 0x80000B850000B4CC, 0x80001F8400035CB7,
-    0x80000CC0000135A3, 0x000488CC80001DF4, 0x80001C9480000A47,
-    0x80001E1C80001F7C, 0x0000498180001F74, 0x00039CCD00007980,
-    0x80000B8E000564CC, 0x80000B61000138CC, 0x80000A4100077188,
-    0x80000BAF0007758A, 0x80000B9480001DFC, 0x00077D8100002DA3,
-    0x80000B9A00077D80, 0x80001E0400074982, 0x0007498780001E0C,
-    0x0007D58080000000, 0x0007F1820007E581, 0x0007998400078D83,
-    0x8000001A000061B1, 0x0007B58800079187, 0x0007898A0007B989,
-    0x8000000A800007E9, 0x80000589800004EA, 0x80000E6180000002,
-    0x800003C780000064, 0x0000ADA000009587, 0x00008D8200076D81,
-    0x0000898C8000091D, 0x0007D5A3800004B7, 0x0000E9A80000DDAF,
-    0x0000D5AA800004AB, 0x0000BDB4800005BB, 0x80000C30000045B5,
-    0x0003D0CC000729A3, 0x0000ADB20005DCCD, 0x000045BC0000A5B3,
-    0x000055BE000051BD, 0x000075B880000BF7, 0x000065BA00007DB9,
-    0x0002E08680000C36, 0x00046CCC0001CD81, 0x0001FD84000468CD,
-    0x00023CCC80000C33, 0x00003581000470CD, 0x00021CCC80000BFA,
-    0x00000981000218CD, 0x80000C3900001184, 0x00007DB280000720,
-    0x8000087500000DB3, 0x800008C580000745, 0x80000DB300000580,
-    0x00056CCC80000954, 0x0000058C000568CD, 0x0000298980000C15,
-    0x00034098000039BF, 0x0003A8A6800006E9, 0x8000083A800006ED,
-    0x800007428000073B, 0x8000076B800008B5, 0x800009148000076E,
-    0x0000F5BF800006E3, 0x80000837800006E6, 0x80000C18000238CC,
-    0x800008630000058C, 0x8000088180000D02, 0x8000093D80000D3B,
-    0x80000D4D80000D41, 0x800000128000096E, 0x8000003080000016,
-    0x8000097980000068, 0x0003E0B880000006, 0x00038D8780000034,
-    0x00038D8500038D86, 0x0001058500037184, 0x0001358500009092,
-    0x80000D7A80000CF9, 0x000951830037D581, 0x0003AD8B00038D8C,
-    0x0003AD890003AD8A, 0x000C01A10003AD88, 0x000E49A3000059A0,
-    0x80000B5B000311A2, 0x80000DBD000075A4, 0x80000DC880000DC0,
-    0x0005DCCD80000DC3, 0x80000DB000050CCC, 0x800006A580000074,
-    0x00035DB780000024, 0x00035DB500035DB6, 0x000341B3000341B4,
-    0x00035DB100035DB2, 0x0003459900035DB0, 0x0003459B00034598,
-    0x00034D9D0003459A, 0x8000099C0003419C, 0x000080CC00034DB8,
-    0x0003ACCC000728CD, 0x000361B3000C98CD, 0x00037DB1000379B2,
-    0x0005B0CC80002218, 0x000764CD0005A8CD, 0x001AFCCC0000C8CC,
-    0x001A74CC80000DCB, 0x0001598080002214, 0x8000221000015981,
-    0x0001758400016D83, 0x000F01808000215C, 0x000668CC000F0181,
-    0x000F39930007E4CD, 0x000700CC80002164, 0x00079CCC000F1D94,
-    0x000F59800007B0CD, 0x000F6982000F5981, 0x00001D858000216C,
-    0x001E08CD00003D85, 0x000195810019C4CC, 0x800021540001B980,
-    0x8000216000001D85, 0x8000216800018984, 0x000FA58000007586,
-    0x000DB0CC000FB581, 0x80002170000D38CD, 0x8000221C80002158,
-    0x0000D98580002234, 0x000570CD80000DCE, 0x80001FE80005E0CC,
-    0x80001FEC80001FF0, 0x80001FF400002D85, 0x000CC8CC80002238,
-    0x0001A1B4000CA4CD, 0x8000219C80000DD1, 0x000FACCC800021A4,
-    0x00019DBA800021AC, 0x800021A080002194, 0x000181BE800021A8,
-    0x80000970800021B0, 0x000C7584000EBD87, 0x80001FA8000E7185,
-    0x80001FB0000FCD85, 0x80001FC080001FB8, 0x80001FAC80001FB4,
-    0x80001FBC80002198, 0x000FA58080001FC4, 0x000FD985000F9981,
-    0x80001FE0000CBC97, 0x000EA581000CF495, 0x00004983000E8980,
-    0x80001FE4000DA491, 0x000009940002BC9F, 0x8000202080002018,
-    0x8000202800003D93, 0x0000A58000027499, 0x0000BD820000B581,
-    0x000F4D85000194A5, 0x80002030000F0185, 0x000F1180000C4CA1,
-    0x000F1180000F1181, 0x000F1581000F6D81, 0x0000F183000F2980,
-    0x000F258500012CA9, 0x8000201C00070CB7, 0x0000A18000073CB5,
-    0x800020240000AD81, 0x000F118500004CB1, 0x0000318000053CBF,
-    0x8000202C00001D81, 0x80002034000F0185, 0x000EC580000634B9,
-    0x000EF585000ED181, 0x00018982000EE185, 0x000169808000223C,
-    0x0001F98000016581, 0x0001D1820001D581, 0x80000DD48000225C,
-    0x0002B8CC00011985, 0x800022580002E4CD, 0x80000E2B80002254,
-    0x0001999400012193, 0x0000398080002058, 0x8000206000002581,
-    0x000EAD8000010983, 0x001424CC000EB981, 0x800020680015BCCD,
-    0x0000AD8000006985, 0x800020700000B581, 0x0000D1848000205C,
-    0x0000DD8680002064, 0x0000E9888000206C, 0x0000D18100002D82,
-    0x000025800000CD80, 0x8000207400002981, 0x00005580800022A0,
-    0x0000718200005981, 0x00005D9400009993, 0x0000F58000008183,
-    0x80000DD70000ED81, 0x0001ECCC80002090, 0x000009800001D4CD,
-    0x8000209800000981, 0x0003E181800020A0, 0x000019800003B980,
-    0x8000209C00001981, 0x0000B585800020A4, 0x8000228800003D94,
-    0x0002A0CC80002284, 0x80000DDA0002ACCD, 0x000385800003D183,
-    0x800008B30003DD81, 0x0003AD9380002280, 0x80000E2E8000227C,
-    0x0000318000040194, 0x0001CCCD00003181, 0x800020CC00013CCC,
-    0x00001D81800020D0, 0x800020D400001D80, 0x0000A58000000582,
-    0x0000AD800000B581, 0x0000DD860000BD81, 0x0000DD84800022A4,
-    0x000210CC80000DDD, 0x0000F98800033CCD, 0x00002580800021DC,
-    0x000178CD00002981, 0x0006D18000010CCC, 0x000005850006D181,
-    0x000120CD800021E4, 0x0006D994000124CC, 0x0006C194800021EC,
-    0x0006FD810006FD82, 0x00002D850006FD80, 0x0001F1850001F985,
-    0x000EA8CD0006F593, 0x800021D4000EB0CD, 0x000E68CD800021E0,
-    0x000E98CC000E50CD, 0x000E38CC00019CCD, 0x000F24CD000E64CD,
-    0x000A88CC000F28CD, 0x0007258000007585, 0x800021E800071581,
-    0x00003D85000144CD, 0x0001F8CC0001C8CD, 0x000040CD0001FCCD,
-    0x0007498100018CCC, 0x0000F98500077980, 0x0000F58500074182,
-    0x0001D8CC800021F0, 0x003A51880001DCCD, 0x800022A8800021D8,
-    0x800021FC00000585, 0x00010D8000079D83, 0x0005C8CC00010181,
-    0x80002204000504CD, 0x00001985000178CD, 0x800021F4800021F8,
-    0x00001D8500007585, 0x8000213C000124CD, 0x000791800000D0CD,
-    0x0000398500079181, 0x8000214C80002144, 0x8000214080002134,
-    0x00079180000118CD, 0x000110CC00079181, 0x800020C80000DCCD,
-    0x800020F80007ED85, 0x8000210000007985, 0x8000211080002108,
-    0x800020FC80002104, 0x8000210C00001993, 0x8000211880002114,
-    0x00006D9480000E31, 0x80001F9080001F88, 0x80001FA080001F98,
-    0x000019820007A985, 0x0000198080001F8C, 0x80001F9400001981,
-    0x0000598080001F9C, 0x0007C98600005981, 0x0007CD8480002148,
-    0x0000E18080001FA4, 0x0000F5820000E181, 0x8000215000001D85,
-    0x8000213800000D85, 0x0000918080002208, 0x0000358200009181,
-    0x0000098500009183, 0x8000220C00000985, 0x0003C0CC80002200,
-    0x8000211C0003D4CD, 0x80001FC880000E34, 0x00000D8080001FD0,
-    0x80001FD800003D81, 0x0000198080001FCC, 0x80001FD400001981,
-    0x0003F0CC80001FDC, 0x0001C98000031CCD, 0x000011800001C981,
-    0x0001D98000002581, 0x800021200001D981, 0x000005850001CD83,
-    0x80000E3780002220, 0x8000222800001D85, 0x0000658180001FF8,
-    0x8000200000007D80, 0x00001D8500007193, 0x0001B59400019D93,
-    0x0000D5948000217C, 0x000568CD80002008, 0x0000058500056CCC,
-    0x8000201080002184, 0x8000218C00000D85, 0x0000C18080002174,
-    0x0000F9820000D981, 0x8000200480001FFC, 0x000009800000E985,
-    0x0000058500003581, 0x8000200C80002180, 0x8000218800001D85,
-    0x0000058580002014, 0x0000518080002190, 0x00007D8200006D81,
-    0x8000222C80002178, 0x0000058500001985, 0x0000F98080002230,
-    0x000039820000CD81, 0x800022240000D983, 0x0001A0CC00000585,
-    0x000180CD000124CD, 0x003E0194000138CC, 0x003EC98580002124,
-    0x003EF985003E1593, 0x80000E3A000124CD, 0x8000224080002244,
-    0x80002248000118CD, 0x0000098080000DE0, 0x0000698600000981,
-    0x00006D84000124CD, 0x8000204080002250, 0x0000558080002094,
-    0x8000204800007D81, 0x000114CD80002050, 0x8000204C80002044,
-    0x003C0188000128CD, 0x00012CCD80002054, 0x0000BD810000C982,
-    0x0000F1810000BD80, 0x003A318F0000C580, 0x0000C582000124CD,
-    0x0000F980000110CD, 0x0000DD800000CD81, 0x000178CD0000C581,
-    0x0000AD80000124CC, 0x0000B5820000B581, 0x00001D828000224C,
-    0x0000D580003ED583, 0x800021280000FD81, 0x000360CC000128CD,
-    0x80000EB1000378CD, 0x00000D8080002080, 0x0006A4CD0001BD81,
-    0x000150CD000678CD, 0x000628CC0006DCCC, 0x000644CD000670CD,
-    0x0007FCCD00063CCD, 0x0007C8CD000728CD, 0x000768CC0007C8CD,
-    0x0007F8CC000740CD, 0x800020880007D0CD, 0x0001D0CD80002084,
-    0x000148CD8000208C, 0x00001D80000144CD, 0x0001E0CD00001181,
-    0x000B4CCD000694CC, 0x0006A4CD000280CC, 0x8000226C0006DCCD,
-    0x003E75800006B0CD, 0x0001698080002270, 0x0001618100017D81,
-    0x003F898300017580, 0x0000D9938000212C, 0x00010CCD003FF993,
-    0x000258CC003F4185, 0x003F3994003D04CD, 0x003EC18500003994,
-    0x003FE988003FD593, 0x000138CD00066CCC, 0x003FC19400013CCC,
-    0x000138CC80000EB4, 0x0000758000013CCD, 0x00000D8600003181,
-    0x00000D8480002264, 0x8000203880002260, 0x8000203C003C5881,
-    0x8000207880002268, 0x8000207C003C9485, 0x80000EAE80000E3D,
-    0x000025800003BC89, 0x00003D8000003181, 0x0000058200002981,
-    0x0001E59080002278, 0x0001E5920001ED91, 0x003A0194003A7D93,
-    0x0001ED960001E595, 0x00000980800020A8, 0x800020B000000981,
-    0x00001980800020B8, 0x0000058200001981, 0x800020AC800020C0,
-    0x00000D80800020B4, 0x800020BC00001D81, 0x0000098100003582,
-    0x800020C400001980, 0x0003A0CC80002274, 0x003F2D8000032CCD,
-    0x00001D82003F6981, 0x000179B00000E183, 0x000109B2000179B1,
-    0x000115B480002130, 0x8000228C00011DB5, 0x8000229480000EB7,
-    0x800020D8000129B9, 0x00000980800021BC, 0x800020E000000981,
-    0x800021C4800020E8, 0x0001C58000002D85, 0x0001E9800001D181,
-    0x000005820001E981, 0x00000585800020F0, 0x800021B4800021CC,
-    0x00000980800020DC, 0x800020E400006581, 0x800021C000007D85,
-    0x003D70CC0001AD93, 0x800020EC003DC0CD, 0x0000759400000985,
-    0x00005980800021C8, 0x0000058200003581, 0x00000985800020F4,
-    0x800021D000000985, 0x00014D80800021B8, 0x0000198200015981,
-    0x8000229800016583, 0x0000058500003585, 0x800022908000229C,
-    0x80000E5580000E43, 0x80000E9F80000E1C, 0x80000E5B80000E8D,
-    0x80000E9380000E90, 0x80000E4C80000DF8, 0x80000EC980000E16,
-    0x00000586000214CC, 0x0000058880000F6E, 0x0035C8CC80000F74,
-    0x80000EC60035CCCD, 0x00014CCC80000EC0, 0x000254CC003C7CCD,
-    0x80000F7A0003F8A3, 0x80000EC3000124CD, 0x000178CD0003FCA7,
-    0x80000F6500012CCC, 0x00011CCD00021CAB, 0x00005D8080000F89,
-    0x80000F8F000254CC, 0x0000798800026CAD, 0x000204B00002ECB3,
-    0x0002A0B600006D86, 0x00013CCD000284B7, 0x000578BA0002C4B5,
-    0x000550B8000124CD, 0x000574BE80000ECF, 0x80000EAB00005988,
-    0x80000ED580000F95, 0x0036ACCD80000F9B, 0x003E6C86003F60CC,
-    0x003A18CD00008588, 0x80000ECC003A14CC, 0x00000981000128CD,
-    0x000274CC0000A586, 0x0003FCCC00034180, 0x80000FA100028CCD,
-    0x00002C9000001C93, 0x0003E096000394CC, 0x0003E4CC0003D097,
-    0x0002D49A00001495, 0x0003609800002588, 0x0003709E80000FB6,
-    0x003DF8CC80000ED2, 0x0003ACCD003DECCD, 0x000278CC0003A8CC,
-    0x000138CC80000FBC, 0x80000FC2000124CD, 0x003FD8CC00005188,
-    0x003E70CC003FDCCD, 0x003FA4CC003E74CD, 0x003F88CC003F9CCD,
-    0x0001B586003FFCCD, 0x0001A58400012980, 0x0000098880000FC8,
-    0x80000FCE00012984, 0x0002E4CC0001DD86, 0x000118CD0001E188,
-    0x80000FB000010CCC, 0x0000758800006D8B, 0x000005860002E4CC,
-    0x80000F7780000F71, 0x0000058100002D80, 0x00020CCC80000F38,
-    0x80000F32000208CD, 0x000400CC0000D588, 0x000138CC000220CD,
-    0x00003188000128CD, 0x0005D8CC80000F7D, 0x003E1CCD80000F35,
-    0x80000F68003E74CD, 0x000128CC80000F8C, 0x000234CC0000CCCD,
-    0x003FE4CD80000F92, 0x80000F9880000F4A, 0x003FFCCD00001188,
-    0x0007D18680000EF3, 0x00020CCC0007E980, 0x80000F9E00022CCD,
-    0x000029810007E584, 0x0007CD800007E186, 0x000234CC00003588,
-    0x80000F47000230CD, 0x0007ED88000254CC, 0x00007D8880000FA4,
-    0x80000FB90002D8CC, 0x80000FBF80000F4D, 0x0003C4B580000FC5,
-    0x000224CC0003C8B4, 0x80000FCB00000D88, 0x0000058800023CCC,
-    0x000224CC80000FD1, 0x0003F4830000F588, 0x80000FB30000C986,
-    0x0003C0870000C584, 0x0000058800022CCC, 0x0003FC8B80000F41,
-    0x80000F53003800CC, 0x0000E58B000224CC, 0x0003F08D0000758F,
-    0x80000F560000ED88, 0x0000058800022CCC, 0x000228CC80000F83,
-    0x0000058800037096, 0x000224CC80000F86, 0x80000FAA00000D88,
-    0x0000058800023CCC, 0x0000058780000FAD, 0x0002F49880000FD4,
-    0x000264D90002E499, 0x800010C9000360A7, 0x800010D2800010CC,
-    0x00003193000264D9, 0x000019940003008A, 0x800010CF00035488,
-    0x0000219500007194, 0x000224D900002994, 0x80001147800010D5,
-    0x8000183A8000176B, 0x000268D9800017C9, 0x00003D948000182E,
-    0x80001851000258D9, 0x000254D900000594, 0x8000182800000D94,
-    0x0003ACB0000268E0, 0x00064D880002ECA4, 0x8000187E0003C4B3,
-    0x000009BC80001890, 0x8000185480001894, 0x0006098000039881,
-    0x00060D86000284E0, 0x0002FCA4000021BC, 0x000210A80002BCA4,
-    0x0002E4E0000218A9, 0x000081B78000185B, 0x000091B5000099B6,
-    0x00031CA880001862, 0x0003E89280001869, 0x00006DBC0000B9B8,
-    0x800018700003E095, 0x000140A480001877, 0x800018820002E4A5,
-    0x800018890002B0A7, 0x0002B8E0000055BE, 0x000129BE00006997,
-    0x000230AE000338AD, 0x000238A8000318AF, 0x0002D4A60002B0A9,
-    0x800018AD00027CA7, 0x00020CE000039CB5, 0x0003BC920003E4B7,
-    0x0003DCB2000368B1, 0x000C64BC0003E0B3, 0x000D6CBE000CFCBD,
-    0x000CA8B8000268E0, 0x000CACBA800018B1, 0x00032198000C2CBB,
-    0x0003559A00034999, 0x0003699C0003699B, 0x0000919E0003699D,
-    0x0003388700008D9F, 0x8000189F80001898, 0x800018D1800018A6,
-    0x000200AF800018D8, 0x00006D9F800018B5, 0x0000619D000075B3,
-    0x000065B60000699C, 0x800018C3800018BC, 0x00005D99800018CA,
-    0x0000319B0001B197, 0x0003B0AD0000399A, 0x0001A1BE00022CAC,
-    0x0002C8870000219E, 0x800018F100007597, 0x800018ED00028CE0,
-    0x0000099600029CE0, 0x800018F500003D97, 0x0003BC870003E086,
-    0x000288AE0003A086, 0x000338AE800018DF, 0x00003DBE000264AF,
-    0x000240B10002C8E0, 0x0000D196800018E6, 0x800018FD800018F9,
-    0x000240B30000B98A, 0x0000259D0002C8E0, 0x800019050000599C,
-    0x00039CBF0000F195, 0x8000190980001901, 0x8000191D8000190D,
-    0x000019950002E0E0, 0x80001911000294E0, 0x000204B300002597,
-    0x00026CB30003A886, 0x80001915000C0182, 0x8000192980001921,
-    0x0003D48680001925, 0x8000192D00039C87, 0x0002F4B40003AC99,
-    0x0002E0E0000214B5, 0x000D3CE0000069BE, 0x0002E4B480001935,
-    0x0000699580001939, 0x8000193D000049BE, 0x00022CB700004D96,
-    0x80001931000FCD8A, 0x0000198F000D78B7, 0x000005B3000FEDB3,
-    0x0000099C80001952, 0x800019440000159D, 0x8000194B000D48E0,
-    0x8000195900000D8C, 0x000035B6000055A9, 0x800019640000C5B7,
-    0x000FBD9F00003DB5, 0x8000196B000039B3, 0x8000197980001972,
-    0x8000195D80001980, 0x8000198E80001987, 0x0000FDB8800019D0,
-    0x800019C60000F5B9, 0x800019D7800019E1, 0x800019A380001995,
-    0x800019AA0000DD83, 0x800019B8800019B1, 0x8000199C800019BF,
-    0x000380A50000F58D, 0x0002C08000020CE1, 0x00007DAE800019E8,
-    0x000005BC0000A597, 0x00008D92800019EC, 0x0000DC85000218E1,
-    0x80003B64000114E3, 0x00008D9C80003B68, 0x00040086000114EF,
-    0x80003B6C00038887, 0x00018D8180003B70, 0x80003B7880003B74,
-    0x80003B8080003B7C, 0x80003B8880003B84, 0x0003048080003B8C,
-    0x80003B9480003B90, 0x0001308400031483, 0x80003B9C80003B98,
-    0x80003BA480003BA0, 0x0001D99380003BA8, 0x80003BB080003BAC,
-    0x80003BB880003BB4, 0x80003BC080003BBC, 0x80003BC880003BC4,
-    0x0001E99D80003BCC, 0x000015810000F582, 0x80003BD000001580,
-    0x000148BE00004884, 0x021884AC000111A2, 0x000139A7021A94AD,
-    0x0007208680003BD8, 0x80003BDC0000EC87, 0x80003BE480003BE0,
-    0x80003BE8000111AC, 0x80003BF080003BEC, 0x0218D8BA80003BF4,
-    0x80003BF80000CC84, 0x0218FCBE0218B8BD, 0x000149B90213DCBF,
-    0x80003C0080003BFC, 0x021AC0B480003C04, 0x021000B6021F30B5,
-    0x0001508080003C08, 0x00776882006A0081, 0x00BF60840052B883,
-    0x00E48C86009B9485, 0x012A248800CE1C87, 0x0113588A0137C489,
-    0x0158B08C017FA88B, 0x018E3C8E01A2108D, 0x01F6C09001EAAC8F,
-    0x023CC89201D34091, 0x00059C840002B1A3, 0x000355A600035DA5,
-    0x80003C1080003C0C, 0x000371AA80003C14, 0x000355AC000371AB,
-    0x00034DAE00035DAD, 0x000261A1000345AF, 0x00068DA380003C18,
-    0x001801A500068DA2, 0x001A1DA7000621A4, 0x001E81A9001ABDA6,
-    0x001FC1AB001C01A8, 0x001291AD001E9DAA, 0x0016B9AF0012B5AC,
-    0x0017DDB10013D1AE, 0x006AE9B30016C1B0, 0x006C01B5006801B2,
-    0x0003F4BE006A0DB4, 0x80003C1C000514BE, 0x0006A4BE80003C20,
-    0x80003C2880003C24, 0x80003C3080003C2C, 0x0003FDB380003C34,
-    0x80003C38006C2CBF, 0x80003C4080003C3C, 0x000211B5000219B6,
-    0x00023DBB80003C48, 0x0005088600023DBA, 0x000209BF00058C87,
-    0x000211BD000201BE, 0x00005182000219BC, 0x0000418080003C4C,
-    0x000000D700004181, 0x80003C5480003C50, 0x80003C5800002CEF,
-    0x80003C5C00018CE1, 0x00197CBE0000F8E3, 0x80003C6480003C60,
-    0x000265AB0018B0BE, 0x000158E0000255AA, 0x0001CDAF80003C68,
-    0x000275AD000269AE, 0x00020DA3000275AC, 0x80003C7080003C6C,
-    0x80003C7480001919, 0x000201A5000209A6, 0x80003C7C80003C78,
-    0x80003C8480003C80, 0x80003C88000261B3, 0x000265B600026DB5,
-    0x80003C8C006D6485, 0x000239BA80003C90, 0x000219BC000209BB,
-    0x000205BE000219BD, 0x80003BD40001A1BF, 0x80003C9480003C44,
-    0x80003C9C80003C98, 0x80003CA480003CA0, 0x0005BDA880003CA8,
-    0x000585AA00058DA9, 0x000589AC00059DAB, 0x0005F9AE000589AD,
-    0x000591B00005F9AF, 0x0005A1B20005A1B1, 0x0005ADB40005A1B3,
-    0x000639B6000609B5, 0x0006E5B8000639B7, 0x00068DBA000685B9,
-    0x000681BC000681BB, 0x000681BE000681BD, 0x00079DA8000681BF,
-    0x0007DDAA0007C1A9, 0x0007F1AC0007D5AB, 0x0007EDAE0007FDAD,
-    0x0000BD800007E5AF, 0x0000198200000981, 0x80003CB080003CAC,
-    0x0000C88780003CB4, 0x000711B800039C86, 0x00072DBA000725B9,
-    0x0000DDBC0000C5BB, 0x0000CDBE0000C5BD, 0x00076DB00000C5BF,
-    0x000741B2000741B1, 0x00074DB400075DB3, 0x00073DB6000735B5,
-    0x000655B3000735B7, 0x80003CBC80003CB8, 0x80003CC480003CC0,
-    0x000679B500067DB6, 0x000685BB80003CC8, 0x80003CCC0006A9BA,
-    0x0006F1BF80003CD0, 0x000689BD000689BE, 0x80003CD4000695BC,
-    0x80003CD8000611A3, 0x000639A680003CDC, 0x80003CE480003CE0,
-    0x0006EDAA00063DA5, 0x80003CE80006C1AB, 0x0006CDAE80003CEC,
-    0x0006D1AC0006B5AF, 0x000085A80006CDAD, 0x000085AA00008DA9,
-    0x00008DAC00009DAB, 0x0000B5AE0000BDAD, 0x80003CF00000DDAF,
-    0x80003CF880003CF4, 0x80003D0080003CFC, 0x80003D0880003D04,
-    0x000069B880003D0C, 0x000079BA000069B9, 0x000069BC000079BB,
-    0x0003B9BE000069BD, 0x0000A9B00003B9BF, 0x0000BDB20000A9B1,
-    0x000099B40000B5B3, 0x000059B6000099B5, 0x000759BC000059B7,
-    0x000769BE000759BD, 0x000755B8000769BF, 0x000749BA000759B9,
-    0x000715B4000749BB, 0x000715B600071DB5, 0x000729B000076DB7,
-    0x00071DB2000729B1, 0x000749AC000715B3, 0x000759AE000749AD,
-    0x0006B5A8000759AF, 0x0006A5AA0006ADA9, 0x80003D10000759AB,
-    0x80003D1880003D14, 0x0000198080003D1C, 0x0000058200001981,
-    0x0002C88480003D20, 0x0003D887000389B3, 0x0003B9B6001C2C86,
-    0x80003D2C80003D28, 0x0003E9BA000381B5, 0x80003D300003E9BB,
-    0x0003D9BE80003D34, 0x0003EDBC0003D9BF, 0x0003CDA30003E5BD,
-    0x80003D3C80003D38, 0x80003D4480003D40, 0x0003C5A50003FDA6,
-    0x0003ADAB80003D48, 0x80003D4C0003ADAA, 0x00039DAF80003D50,
-    0x0003A1AD00039DAE, 0x80003D540003A9AC, 0x000718EF80003D58,
-    0x80003D6080003D5C, 0x80003D6880003D64, 0x0002208480003D6C,
-    0x80003D7480003D70, 0x000194E180003D78, 0x0003A8E380003D7C,
-    0x80003D8480003D80, 0x80003D8C80003D88, 0x0000158080003D90,
-    0x0000158200001D81, 0x001C408680003D98, 0x80003D9C001D5487,
-    0x80003DA480003DA0, 0x80003DAC80003DA8, 0x0006B0E180003DB0,
-    0x0007C8E300043CE1, 0x80003DB4000450E3, 0x80003DBC80003DB8,
-    0x80003DC480003DC0, 0x80003DCC80003DC8, 0x80003DD480003DD0,
-    0x00079CEF80003DD8, 0x80003DDC00079CEF, 0x001DC08480003DE0,
-    0x80003DE4000165B3, 0x000111B6000111B5, 0x80003DEC80003DE8,
-    0x000135BA80003DF0, 0x00011DBC000105BB, 0x0000FDBE0000F5BD,
-    0x001E95A80000F1BF, 0x001EA1AA001E9DA9, 0x001ED5AC001EA1AB,
-    0x001ED5AE001EDDAD, 0x80003D94001ECDAF, 0x80003DF880003DF4,
-    0x0000198080003DFC, 0x0000058200001981, 0x001F59B880003E00,
-    0x001F6DBA001F59B9, 0x001F7DBC001F65BB, 0x001F65BE001F61BD,
-    0x001EBDB0001F1DBF, 0x001E85B2001E8DB1, 0x001F75B4001F6DB3,
-    0x001F69B6001F79B5, 0x001E29A3001F69B7, 0x0003C48780003E04,
-    0x80003E0800002C86, 0x001E5DA5001E5DA6, 0x001E49AB80003E0C,
-    0x80003E10001E79AA, 0x001FBDAF80003E14, 0x001FB9AD001FB1AE,
-    0x80003E18001E51AC, 0x80003E1C001FA9B3, 0x001FA9B680003E20,
-    0x80003E2880003E24, 0x001F75BA001FA9B5, 0x80003E2C001F7DBB,
-    0x001F55BE80003E30, 0x001F65BC001F5DBF, 0x80003E34001F5DBD,
-    0x80003E3C80003E38, 0x80003E4480003E40, 0x80003E48001D80E1,
-    0x80003E4C001C2CE3, 0x80003E5480003E50, 0x80003E580018A4EF,
-    0x80003E6080003E5C, 0x001E2DA380003E64, 0x0000158100002D82,
-    0x80003E6800001D80, 0x001E2DA5001E2DA6, 0x001EF9AB80003E6C,
-    0x80003E70001EF1AA, 0x001ED9AF00012084, 0x001ED9AD001ED1AE,
-    0x001EDDA8001EE1AC, 0x001EE5AA001EEDA9, 0x001EE5AC001EFDAB,
-    0x001EE5AE001EEDAD, 0x0001E8BE001E0DAF, 0x0007088680003E78,
-    0x80003E7C0000B887, 0x80003E8480003E80, 0x001E75B880003E88,
-    0x001E0DBA001E05B9, 0x001E0DBC001E1DBB, 0x001E3DBE001E35BD,
-    0x001E75B00001F9BF, 0x001E75B2001E7DB1, 0x001E55B4001E4DB3,
-    0x001E55B6001E5DB5, 0x000291A8001E4DB7, 0x0002A1AA000299A9,
-    0x0002D1AC0002A1AB, 0x0002D5AE0002DDAD, 0x80003E8C0002C9AF,
-    0x80003E9480003E90, 0x0000008C80003E98, 0x80003EA080003E9C,
-    0x000359B880003EA4, 0x00036DBA000365B9, 0x000361BC000365BB,
-    0x000361BE000361BD, 0x0002B9B0000361BF, 0x000285B200028DB1,
-    0x000375B400036DB3, 0x000375B600037DB5, 0x000229B300036DB7,
-    0x80003EAC80003EA8, 0x80003D2480003EB0, 0x000259B500025DB6,
-    0x000249BB80003EB4, 0x80003EB800027DBA, 0x0003B9BF80003EBC,
-    0x0003B9BD0003B1BE, 0x80003EC0000251BC, 0x80003EC400026DA3,
-    0x000219A680003EC8, 0x80003ED080003ECC, 0x000239AA00021DA5,
-    0x80003ED400020DAB, 0x0003F5AE80003ED8, 0x000215AC0003FDAF,
-    0x000194E10003FDAD, 0x000074E380003EDC, 0x0000158180003EE0,
-    0x000254BE0001F980, 0x00042CBE00001182, 0x80003EEC80003EE8,
-    0x80003EF480003EF0, 0x000060EF80003EF8, 0x0002F5B30002B084,
-    0x0005648780003EFC, 0x80003F0000056C86, 0x0002E5B5000175B6,
-    0x000155BB80003F04, 0x80003F08000149BA, 0x0001F9BF80003F0C,
-    0x000145BD0001F1BE, 0x80003F10000145BC, 0x80003F1880003F14,
-    0x80003F2080003F1C, 0x80003F2880003F24, 0x80003F2C001A18EF,
-    0x80003F30001EDCE1, 0x00043C84001E74E3, 0x80003F3880003F34,
-    0x000275A380003F3C, 0x80003F4480003F40, 0x80003F480004CC84,
-    0x000265A50001F5A6, 0x0001D5AB80003F4C, 0x80003F500001C9AA,
-    0x000179AF80003F58, 0x0001C5AD000171AE, 0x0005A5A80001C5AC,
-    0x0005BDAA0005B5A9, 0x0005ADAC0005B5AB, 0x00063DAE000635AD,
-    0x80003EE4000635AF, 0x0007F5810007CD82, 0x80003F540007FD80,
-    0x0003F88780003F5C, 0x0006E9B800009C86, 0x000689BA0006E9B9,
-    0x00069DBC000685BB, 0x000681BE000681BD, 0x00064DB00006B5BF,
-    0x000651B2000651B1, 0x0006F5B4000651B3, 0x0006F5B60006FDB5,
-    0x0007A5A80006E1B7, 0x0007A5AA0007ADA9, 0x0007A5AC0007BDAB,
-    0x0007C9AE0007D1AD, 0x0000A4BE0007C1AF, 0x80003F6480003F60,
-    0x80003F6C80003F68, 0x80003F7480003F70, 0x000759B880003F78,
-    0x000761BA000765B9, 0x000769BC000779BB, 0x00071DBE000769BD,
-    0x000789B0000715BF, 0x000769B2000789B1, 0x000779B4000769B3,
-    0x000769B6000779B5, 0x80003F7C000769B7, 0x80003F8480003F80,
-    0x80003F8C80003F88, 0x0004F8EF80003F90, 0x000690E180003F94,
-    0x000404E380003F98, 0x80003FA080003F9C, 0x80003FA880003FA4,
-    0x00066DB380003FAC, 0x80003FB480003FB0, 0x80003FBC80003FB8,
-    0x00066DB5000611B6, 0x00060DBB80003FC0, 0x80003FC400060DBA,
-    0x0006FDBF80003FC8, 0x0006FDBD0006FDBE, 0x0006ADA30006FDBC,
-    0x0000288680003FCC, 0x80003FD00001A887, 0x0006ADA50006D1A6,
-    0x0006CDAB80003FD4, 0x80003FD80006CDAA, 0x00063DAF80003FDC,
-    0x00063DAD00063DAE, 0x00008D8000063DAC, 0x0000958200009581,
-    0x80003FE0000169B3, 0x000169B6000179B5, 0x80003FE880003FE4,
-    0x000155BA80003FEC, 0x0000F5BC0000EDBB, 0x0000F5BE0000FDBD,
-    0x0005EDA80000EDBF, 0x0002ADAA000231A9, 0x0002BDAC0002A5AB,
-    0x0002ADAE0002A5AD, 0x0000F9810002A5AF, 0x80003FF00000F580,
-    0x80003FF40000F182, 0x80003FFC80003FF8, 0x000365B880004000,
-    0x000379BA000369B9, 0x000369BC000379BB, 0x00031DBE000369BD,
-    0x0002DDB0000315BF, 0x00036DB2000365B1, 0x00037DB4000365B3,
-    0x00036DB6000365B5, 0x80004004000361B7, 0x00022DA380004008,
-    0x00023DA58000400C, 0x8000401480004010, 0x000C00BE00022DA6,
-    0x0003A9AB80003E74, 0x0003B9AD000211AA, 0x0003A9AF0003B1AC,
-    0x800040180003B1AE, 0x800040208000401C, 0x8000402880004024,
-    0x8000402C00028CEF, 0x0003F88480004030, 0x80004038000214E3,
-    0x000CD08600017CE1, 0x8000403C000CB487, 0x00026DA880004040,
-    0x000299AA000281A9, 0x0002B9AC000291AB, 0x0002C9AE0002B9AD,
-    0x800040440002C1AF, 0x8000404C80004048, 0x8000405480004050,
-    0x8000405C80004058, 0x0001C5B880004060, 0x0001DDBA0001C9B9,
-    0x000191BC0001D5BB, 0x000191BE000191BD, 0x000281B0000191BF,
-    0x000141B2000281B1, 0x000145B400015DB3, 0x000145B600014DB5,
-    0x800040640001FDB7, 0x8000406C80004068, 0x000D10840001D4EF,
-    0x8000407480004070, 0x80004078000644EF, 0x8000407C0006F0E1,
-    0x8000408000060CE3, 0x80004084000190E1, 0x00026DB6000660E3,
-    0x8000408C80004088, 0x80004090000265B5, 0x80004094000245B3,
-    0x0001D5BE80004098, 0x0001C1BC0001DDBF, 0x000211BA0001DDBD,
-    0x8000409C000211BB, 0x80004034800040A0, 0x000DE9AA000388BE,
-    0x000E39AC000DE9AB, 0x000E2DAE000E25AD, 0x800040A4000E25AF,
-    0x800040AC800040A8, 0x800040B0000DBDA3, 0x000D95A6000D9DA5,
-    0x000F9D80800040B4, 0x000FB582000FBD81, 0x800040B8000EB1B3,
-    0x000F71B6000F71B5, 0x00034086800040BC, 0x000F5DBA00033487,
-    0x000F4DBC000F55BB, 0x000FFDBE000FF5BD, 0x000E69A8000FF5BF,
-    0x000E75AA000E7DA9, 0x000FB5AC000FADAB, 0x000FB5AE000FBDAD,
-    0x800040C0000FADAF, 0x800040C8800040C4, 0x800040D0800040CC,
-    0x800040D8800040D4, 0x000F61B8800040DC, 0x000F79BA000F6DB9,
-    0x000F15BC000F75BB, 0x000F09BE000F1DBD, 0x000FD5B0000F05BF,
-    0x000FD5B2000FDDB1, 0x000F75B4000F6DB3, 0x000F6DB6000F7DB5,
-    0x000FFDA3000F65B7, 0x800040E4800040E0, 0x800040EC800040E8,
-    0x000E3DA5000E3DA6, 0x000E19AB800040F0, 0x800040F4000E11AA,
-    0x000EB9AF800040F8, 0x000EB9AD000EB1AE, 0x800040FC000E01AC,
-    0x8000410480004100, 0x00001D8080004108, 0x0000398200000981,
-    0x800041108000410C, 0x0001A0BE0001A884, 0x0000D08680004114,
-    0x800041180001E487, 0x0001A5A88000411C, 0x0001B9AA0001ADA9,
-    0x0001C1AC0001B1AB, 0x000139AE0001C9AD, 0x0000A084000139AF,
-    0x8000412480004120, 0x8000412C80004128, 0x8000413480004130,
-    0x0000CDB880004138, 0x0000D5BA0000D5B9, 0x0000FDBC0000E5BB,
-    0x00009DBE000095BD, 0x000149B0000095BF, 0x000159B2000149B1,
-    0x000149B4000159B3, 0x0000FDB6000149B5, 0x000231B30000F5B7,
-    0x800041408000413C, 0x8000414880004144, 0x000249B5000249B6,
-    0x00028DBB8000414C, 0x8000415000028DBA, 0x00028DBF80004154,
-    0x00028DBD00028DBE, 0x007D448500028DBC, 0x80004158000275A3,
-    0x00020DA68000415C, 0x8000416480004160, 0x0002C9AA00020DA5,
-    0x800041680002C9AB, 0x0002C9AE8000416C, 0x0002C9AC0002C9AF,
-    0x000400BE0002C9AD, 0x8000417480004170, 0x00001D8080004178,
-    0x00003D8200000981, 0x00052CBE8000417C, 0x8000418880004180,
-    0x800041908000418C, 0x8000419880004194, 0x0004E0868000419C,
-    0x800041A000035487, 0x0002A8EF800041A4, 0x800041AC800041A8,
-    0x000150E3800041B0, 0x000180E1800041B4, 0x800041BC800041B8,
-    0x800041C4800041C0, 0x000279A800047084, 0x0002ADAA000279A9,
-    0x0002BDAC0002A5AB, 0x0002ADAE0002A5AD, 0x0004A8840002A5AF,
-    0x800041CC800041C8, 0x800041D4800041D0, 0x800041DC800041D8,
-    0x000109B8800041E0, 0x00011DBA000109B9, 0x000139BC000115BB,
-    0x0001DDBE000139BD, 0x0002DDB00001D5BF, 0x000161B2000161B1,
-    0x000161B4000161B3, 0x000151B6000169B5, 0x800041E4000151B7,
-    0x800041E8000F6CE1, 0x00001980000E0CE3, 0x0000318200003581,
-    0x800041EC00000CEF, 0x800041F00001DCE1, 0x000000ED0000D8E3,
-    0x800041FC800041F8, 0x0004E086000140EF, 0x8000420000057887,
-    0x80004204000269B3, 0x000269B6000279B5, 0x8000420C80004208,
-    0x000255BA80004210, 0x0001F9BC0001E9BB, 0x0001E9BE0001F9BD,
-    0x800041840001E1BF, 0x80004214800041F4, 0x800042180005A9A3,
-    0x0005A9A60005B9A5, 0x800042208000421C, 0x000595AA80004224,
-    0x000639AC000629AB, 0x000629AE000639AD, 0x0006F1B3000621AF,
-    0x8000422C80004228, 0x8000423480004230, 0x0006E1B50006E9B6,
-    0x0006B9BB80004238, 0x8000423C0006B1BA, 0x000681BF80004240,
-    0x00069DBD000699BE, 0x000629A80006A1BC, 0x000639AA000629A9,
-    0x000629AC000639AB, 0x000651AE000629AD, 0x80004244000651AF,
-    0x8000424C80004248, 0x8000425480004250, 0x0001B1810001B982,
-    0x0007E5B80001B980, 0x0007E5BA0007EDB9, 0x0007E5BC0007FDBB,
-    0x0007E5BE0007EDBD, 0x000635B000075DBF, 0x000635B200063DB1,
-    0x00061DB4000609B3, 0x00060DB6000605B5, 0x0007B1A3000605B7,
-    0x8000425C80004258, 0x00019CBE80004260, 0x0007A1A50007A9A6,
-    0x0007F9AB80004264, 0x000088860007F1AA, 0x0007C1AF00017887,
-    0x0007DDAD0007D9AE, 0x800042680007E1AC, 0x8000426C00068DB3,
-    0x0006B1B680004270, 0x8000427880004274, 0x000171BA000681B5,
-    0x8000427C000179BB, 0x000135BE80004280, 0x000159BC00013DBF,
-    0x000635A8000159BD, 0x000641AA00063DA9, 0x00064DAC00065DAB,
-    0x00067DAE000675AD, 0x0001AC84000195AF, 0x8000428880004284,
-    0x800042908000428C, 0x8000429880004294, 0x00019DB88000429C,
-    0x0001A5BA0001ADB9, 0x000175BC00016DBB, 0x000175BE00017DBD,
-    0x0001F1B000016DBF, 0x0001C9B20001F9B1, 0x0001B5B40001C9B3,
-    0x0001B5B60001BDB5, 0x0005C9A30001ADB7, 0x800042A4800042A0,
-    0x800042AC800042A8, 0x0005C5A50005F5A6, 0x00023DAB800042B0,
-    0x800042B4000235AA, 0x000279AF800042B8, 0x00021DAD000271AE,
-    0x800042BC00021DAC, 0x800042C4800042C0, 0x00000981800042C8,
-    0x800042CC00001980, 0x800042D000003D82, 0x007CC886000320BE,
-    0x800042D800035C87, 0x800042E0800042DC, 0x007CCC84800042E4,
-    0x800042EC800042E8, 0x000120E3007CC4BE, 0x0001A8E1800042F0,
-    0x00023CEF800042F4, 0x800042FC800042F8, 0x8000430480004300,
-    0x8000430C80004308, 0x00038DB3005D34BF, 0x8000431480004310,
-    0x8000431C80004318, 0x000381B50003BDB6, 0x00034DBB80004320,
-    0x8000432400034DBA, 0x00034DBF80004328, 0x00034DBD00034DBE,
-    0x000241A800034DBC, 0x000245AA00024DA9, 0x000249AC000259AB,
-    0x000275AE00027DAD, 0x007FA0BE0002B9AF, 0x800043308000432C,
-    0x00000D8080004334, 0x0000AD820000A581, 0x00014DB880004338,
-    0x000155BA000155B9, 0x00017DBC000165BB, 0x00011DBE000115BD,
-    0x0002C9B0000115BF, 0x0002D9B20002C9B1, 0x0002C9B40002D9B3,
-    0x00017DB60002C9B5, 0x000658E1000175B7, 0x0006B8E300075CE1,
-    0x8000433C000760E3, 0x007CB88680004340, 0x80004348007D5487,
-    0x800043508000434C, 0x000000E980004354, 0x000050EF80004358,
-    0x8000435C0007E4EF, 0x000285A380004360, 0x000289A580004364,
-    0x8000436C80004368, 0x800043700002B5A6, 0x000245AB80004374,
-    0x000245AD000245AA, 0x000245AF000245AC, 0x007E71A8000245AE,
-    0x007E71AA007E71A9, 0x007E95AC007E71AB, 0x007E8DAE007E9DAD,
-    0x80004344007E81AF, 0x8000437C80004378, 0x8000438480004380,
-    0x8000438C80004388, 0x007E8DB880004390, 0x007E95BA007E95B9,
-    0x007EBDBC007EA5BB, 0x007F71BE007F71BD, 0x007EC5B0007F71BF,
-    0x007EC5B2007ECDB1, 0x007EC5B4007EDDB3, 0x007EBDB6007ECDB5,
-    0x007EF9B3007EB5B7, 0x8000439880004394, 0x800043A08000439C,
-    0x007E1DB5007E1DB6, 0x007E09BB800043A4, 0x800043A8007E01BA,
-    0x007E09BF800043AC, 0x007E09BD007E01BE, 0x00001582007E11BC,
-    0x00006180007EBDA3, 0x007E59A600006181, 0x800043B4800043B0,
-    0x007E45AA007E59A5, 0x00032484007E4DAB, 0x007E45AE800043B8,
-    0x007E55AC007E4DAF, 0x007EA9B3007E4DAD, 0x00014886800043BC,
-    0x800043C00002B087, 0x007EADB5007ED5B6, 0x007EF9BB800043C4,
-    0x800043C8007EF1BA, 0x000119BF800043CC, 0x007EDDBD000111BE,
-    0x007DD0E3007EDDBC, 0x007D38E1800043D0, 0x800043D8800043D4,
-    0x800043E0800043DC, 0x800043E8800043E4, 0x800043F0800043EC,
-    0x007D04EF800043F4, 0x800043FC800043F8, 0x007E21A880004400,
-    0x007E21AA007E21A9, 0x007E25AC007E21AB, 0x007E5DAE007E55AD,
-    0x80004404007E51AF, 0x8000440C80004408, 0x8000441480004410,
-    0x0003BD810003BD82, 0x0001F5B80003BD80, 0x0001F5BA0001FDB9,
-    0x000159BC000149BB, 0x000149BE000159BD, 0x007E35B0000141BF,
-    0x007E09B2007E3DB1, 0x0001D5B4007E09B3, 0x0001D5B60001DDB5,
-    0x800044180001CDB7, 0x800044208000441C, 0x00028884007DA5A3,
-    0x007DD9A6007DA1A5, 0x80004428000390BE, 0x007DFDAA8000442C,
-    0x007DD1AC007DF5AB, 0x00021DAE007DD1AD, 0x00036DB6000215AF,
-    0x0003448480004430, 0x80004434000365B5, 0x00032887000345B3,
-    0x000301BE0004FC86, 0x000331BC000301BF, 0x000321BA000331BD,
-    0x80004438000321BB, 0x0003D1B08000443C, 0x0003D1B20003D1B1,
-    0x0003F1B40003D1B3, 0x0003F1B60003F1B5, 0x0003D5B80003F1B7,
-    0x0003D5BA0003DDB9, 0x000315BC00030DBB, 0x000315BE00031DBD,
-    0x8000444000030DBF, 0x0020F08580004444, 0x80004448800042D4,
-    0x800044508000444C, 0x00020DA880004454, 0x0003F9AA0003E1A9,
-    0x0003D9AC0003F1AB, 0x0003BDAE0003D9AD, 0x000211A30003B1AF,
-    0x8000445C80004458, 0x00051CBE80004460, 0x000231A5000239A6,
-    0x000275AB80004464, 0x80004468000275AA, 0x000255AF8000446C,
-    0x000265AD000255AE, 0x80004470000265AC, 0x8000447880004474,
-    0x800044808000447C, 0x80004484000354E3, 0x8000448800004CE1,
-    0x8000448C000380EF, 0x0000118080004490, 0x0000158200001D81,
-    0x007BD4E180004498, 0x000354E38000449C, 0x00073086800044A0,
-    0x800044A400047487, 0x800044A8007E44EF, 0x800044AC007E24E1,
-    0x800044B0000154E3, 0x007FF8EF800044B4, 0x000185B3003C00BF,
-    0x800044BC800044B8, 0x800044C000078084, 0x0001A1B50001A9B6,
-    0x000135BB800044C4, 0x800044C8000135BA, 0x0000F9BF800044CC,
-    0x000125BD0000F1BE, 0x000621A8000125BC, 0x00064DAA000621A9,
-    0x000669AC00065DAB, 0x0006D9AE000669AD, 0x800044940006D9AF,
-    0x800044D4800044D0, 0x0004E9A2800044D8, 0x0006F5A00002E5A3,
-    0x000769B800041DA1, 0x000709BA000769B9, 0x000719BC000709BB,
-    0x000709BE000719BD, 0x0006B1B0000705BF, 0x0006B1B20006B1B1,
-    0x000775B40006B1B3, 0x000775B600077DB5, 0x0006C5A3000761B7,
-    0x800044DC006C658F, 0x800044E4800044E0, 0x0006E1A50006E9A6,
-    0x000675AB800044E8, 0x800044EC000675AA, 0x0007B9AF800044F0,
-    0x000665AD0007B1AE, 0x0078C59E000665AC, 0x007DA59C0079D99F,
-    0x007CD19A0078CD9D, 0x0070E998007CF59B, 0x007245960071B599,
-    0x0074C9940070C597, 0x0068459200746995, 0x0068F99000766593,
-    0x0075858A0068F991, 0x800044F40073A18B, 0x0070798E800044F8,
-    0x0071398C007D158F, 0x006DC18200712D8D, 0x800044FC00692583,
-    0x0077A18680004500, 0x00693D8400751587, 0x0005619A00698985,
-    0x000108BE00059D9B, 0x8000450C80004508, 0x0003899C000510E1,
-    0x007D4192000504E3, 0x80004510007C0993, 0x00786D96000484EF,
-    0x007979940007B997, 0x0001A1A100797995, 0x000D8DA380004514,
-    0x000CEDA5000C01A2, 0x000991A7000DC5A4, 0x001401A9000999A6,
-    0x001485AB000965A8, 0x001191AD00158DAA, 0x001C01AF001001AC,
-    0x001DF1B1001161AE, 0x0019CDB3001DE5B0, 0x001801B5001801B2,
-    0x000078860019B5B4, 0x0000ED8000011C87, 0x0000C5820000CD81,
-    0x8000451C80004518, 0x8000452480004520, 0x00057DA880004528,
-    0x000585AA000581A9, 0x0005B1AC000595AB, 0x0005A9AE0005BDAD,
-    0x8000452C0005A9AF, 0x8000453480004530, 0x8000453C80004538,
-    0x8000454480004540, 0x00026DB880004548, 0x0002C5BA0002CDB9,
-    0x0002CDBC0002DDBB, 0x0002F5BE0002FDBD, 0x0005A5B0000355BF,
-    0x0005B9B20005ADB1, 0x000599B40005B1B3, 0x00025DB6000599B5,
-    0x00050DB3000255B7, 0x800045508000454C, 0x8000455880004554,
-    0x000539B500053DB6, 0x000569BB8000455C, 0x80004560000561BA,
-    0x00055DBF80004564, 0x000565BD00055DBE, 0x00001D8000056DBC,
-    0x00002D8200002581, 0x80004424000541A3, 0x000571A6000575A5,
-    0x8000456C80004568, 0x00052DAA80004570, 0x000521AC000525AB,
-    0x000511AE000529AD, 0x00024DA8000511AF, 0x000261AA000255A9,
-    0x0002FDAC000289AB, 0x0002EDAE0002E5AD, 0x000224840002E5AF,
-    0x000D6887000DA8BE, 0x80004578000D0C86, 0x800045808000457C,
-    0x00017DB880004584, 0x0001C1BA0001C1B9, 0x0001C1BC0001C1BB,
-    0x0001F1BE0001C9BD, 0x00029DB00001F1BF, 0x00014DB2000145B1,
-    0x00015DB4000145B3, 0x00014DB6000145B5, 0x80004588000145B7,
-    0x800045908000458C, 0x8000459880004594, 0x800045A08000459C,
-    0x800045A400002CEF, 0x800045A800062CE1, 0x000C3CBE000670E3,
-    0x800045B0800045AC, 0x000D4084800045B4, 0x0002A1B3800045B8,
-    0x00023DB5800045BC, 0x800045C4800045C0, 0x000FE0BE00022DB6,
-    0x0001C9BB800045C8, 0x0001D9BD000211BA, 0x0001C9BF0001D1BC,
-    0x000D69A30001D1BE, 0x800045CC80004574, 0x800045D4800045D0,
-    0x000DF5A5000DE5A6, 0x000E01AB800045D8, 0x800045DC000DD9AA,
-    0x000E01AF800045E0, 0x000E11AD000E19AE, 0x00005D80000E19AC,
-    0x0000658200006D81, 0x800045E4000F75B3, 0x000FDDB6000FD5B5,
-    0x00036086800045E8, 0x000FF9BA00030487, 0x000FE9BC000FF9BB,
-    0x000FD1BE000FD5BD, 0x000FB1A8000FC9BF, 0x000FB1AA000FB1A9,
-    0x000F11AC000FB1AB, 0x000F11AE000F11AD, 0x800045EC000F11AF,
-    0x800045F4800045F0, 0x800045FC800045F8, 0x8000460480004600,
-    0x000F05B880004608, 0x000F1DBA000F09B9, 0x000F35BC000F15BB,
-    0x0000D5BE0000DDBD, 0x000F71B00000CDBF, 0x000F71B2000F71B1,
-    0x000F59B4000F4DB3, 0x000F41B6000F41B5, 0x000E39A3000F41B7,
-    0x800046108000460C, 0x8000461880004614, 0x000E99A5000E91A6,
-    0x000EB5AB8000461C, 0x80004620000EB5AA, 0x000E85AF80004624,
-    0x000E99AD000E9DAE, 0x80004628000EA5AC, 0x800046308000462C,
-    0x8000463880004634, 0x800046408000463C, 0x00000D8080004644,
-    0x00001D8200001581, 0x8000464C80004648, 0x0001F48480004650,
-    0x0000208680004654, 0x8000465800010487, 0x8000465C80004504,
-    0x8000466480004660, 0x0001FDA880004668, 0x00010DAA000105A9,
-    0x00011DAC000105AB, 0x00010DAE000105AD, 0x00017DB0000105AF,
-    0x0000C1B20000C1B1, 0x0000C1B40000C1B3, 0x0000F1B60000C9B5,
-    0x0000A1B80000F1B7, 0x0000A1BA0000A1B9, 0x0000A1BC0000A1BB,
-    0x0000A1BE0000A1BD, 0x000F30E30000A1BF, 0x000DC8E18000466C,
-    0x8000467480004670, 0x8000467C80004678, 0x0004008480004680,
-    0x8000468880004684, 0x000FECEF8000468C, 0x8000469480004690,
-    0x0002E5B380004698, 0x800046A08000469C, 0x800046A8800046A4,
-    0x0002F5B50002E5B6, 0x000365BB800046AC, 0x800046B00002D9BA,
-    0x00036DBF800046B4, 0x000375BD000365BE, 0x00001180000375BC,
-    0x0000298200001981, 0x800046B8000229A3, 0x000229A6000239A5,
-    0x800046C0800046BC, 0x000215AA800046C8, 0x0003B9AC0003A9AB,
-    0x0003A9AE0003B9AD, 0x0002FCEF0003A1AF, 0x00056886000448BE,
-    0x800046CC00036087, 0x800046D4800046D0, 0x800046DC800046D8,
-    0x800046E4800046E0, 0x0002C0E3800046E8, 0x000160E1800046EC,
-    0x000289A8800046F0, 0x00029DAA000295A9, 0x00028DAC000295AB,
-    0x000115AE000111AD, 0x800046F400010DAF, 0x800046FC800046F8,
-    0x8000470480004700, 0x8000470C80004708, 0x00017DB880004710,
-    0x000101BA000101B9, 0x000101BC000101BB, 0x000131BE000109BD,
-    0x000175B0000131BF, 0x000175B200017DB1, 0x000159B4000149B3,
-    0x00014DB6000159B5, 0x000EB4E1000145B7, 0x000E58E380004714,
-    0x80004718000000E6, 0x000594BE8000471C, 0x8000472480004720,
-    0x00001D8100001982, 0x8000472800001D80, 0x000EF0EF8000472C,
-    0x0005488480004730, 0x0005288780004738, 0x8000473C0004EC86,
-    0x800047400001ACE1, 0x800047440000B4E3, 0x8000474C80004748,
-    0x80004750000EF4EF, 0x8000475880004754, 0x000595A38000475C,
-    0x80004734800046C4, 0x8000476480004760, 0x000585A5000611A6,
-    0x000639AB80004768, 0x8000476C000631AA, 0x00061DAF80004770,
-    0x00061DAD00061DAE, 0x80004774000621AC, 0x80004778000691B3,
-    0x0006B1B68000477C, 0x8000478480004780, 0x00068DBA000681B5,
-    0x80004788000691BB, 0x000755BE8000478C, 0x000685BC00075DBF,
-    0x000655A300075DBD, 0x8000479480004790, 0x8000479C80004798,
-    0x000645A5000675A6, 0x000655AB800047A0, 0x00006484000649AA,
-    0x000799AF000064BE, 0x000799AD000791AE, 0x800047A4000641AC,
-    0x0006DDBA800047A8, 0x0006CDBC0006D5BB, 0x0006F1BE0006F1BD,
-    0x00006D800006F1BF, 0x0000758200007581, 0x800047AC000579B3,
-    0x0006FDB6000569B5, 0x0007A5A8800047B0, 0x0007DDAA0007D5A9,
-    0x0007CDAC0007D5AB, 0x0007FDAE0007F5AD, 0x000100860007F5AF,
-    0x800047B400018487, 0x800047BC800047B8, 0x800047C4800047C0,
-    0x000675B8800047C8, 0x000675BA00067DB9, 0x0006C1BC0006C9BB,
-    0x0006C5BE0006C1BD, 0x00078DB00006FDBF, 0x000795B2000791B1,
-    0x000655B400064DB3, 0x000655B600065DB5, 0x000645A800064DB7,
-    0x000651AA000655A9, 0x000675AC00066DAB, 0x000669AE00067DAD,
-    0x800047CC000669AF, 0x800047D4800047D0, 0x800047DC800047D8,
-    0x800047E4800047E0, 0x000189B8800047E8, 0x00019DBA000189B9,
-    0x0001B9BC000195BB, 0x00015DBE0001B9BD, 0x000619B0000155BF,
-    0x0001E5B2000619B1, 0x0001EDB40001FDB3, 0x0001D5B60001DDB5,
-    0x0005D9B30001BDB7, 0x800047F0800047EC, 0x800047F4000340BE,
-    0x0005FDB50005FDB6, 0x0005DDBB800047F8, 0x800047FC0005DDBA,
-    0x000219BF80004800, 0x000229BD000229BE, 0x80004804000231BC,
-    0x8000480800059DA3, 0x0005B9A68000480C, 0x8000481000039884,
-    0x000599AA0005B9A5, 0x001CC886000599AB, 0x00026DAE0003FC87,
-    0x000275AC00025DAF, 0x000245A800026DAD, 0x000275AA000275A9,
-    0x0003BDAC0003A5AB, 0x0003ADAE0003A5AD, 0x0000E9800003A5AF,
-    0x0000F1820000E981, 0x8000481C80004818, 0x001CF48480004820,
-    0x00037DB880004824, 0x0000C1BA0000C1B9, 0x0000CDBC0000DDBB,
-    0x0000F5BE0000FDBD, 0x0003DDB000009DBF, 0x00034DB2000345B1,
-    0x00035DB4000345B3, 0x00034DB6000345B5, 0x0003FDB3000345B7,
-    0x8000482C80004828, 0x8000483480004830, 0x000235B5000225B6,
-    0x000281BB80004838, 0x8000483C000219BA, 0x000281BF80004840,
-    0x000291BD000299BE, 0x80004844000299BC, 0x8000484800007CE1,
-    0x8000484C000308E3, 0x8000485480004850, 0x8000485C80004858,
-    0x8000486480004860, 0x8000486C80004868, 0x8000487480004870,
-    0x000750E1000388EF, 0x000658E300013CE3, 0x00000D80000658E1,
-    0x0000318200003181, 0x001DE8BE80004878, 0x800048808000487C,
-    0x8000488C80004888, 0x0006D8EF0001E0EF, 0x00033DA380004890,
-    0x001C688680004894, 0x80004898001CEC87, 0x0002F5A50002E5A6,
-    0x000241AB8000489C, 0x800048A00002D9AA, 0x000241AF800048A4,
-    0x000251AD000259AE, 0x001EADA8000259AC, 0x001EBDAA001EB5A9,
-    0x001F15AC001EB1AB, 0x001F15AE001F1DAD, 0x80004884001F0DAF,
-    0x800048AC800048A8, 0x800048B4800048B0, 0x800048BC800048B8,
-    0x001F1DB8800048C0, 0x001F2DBA001F25B9, 0x001F3DBC001F25BB,
-    0x001F2DBE001F25BD, 0x001F7DB0001F25BF, 0x001F45B2001F4DB1,
-    0x001F39B4001F2DB3, 0x001F2DB6001F25B5, 0x001E35B3001F25B7,
-    0x800048C8800048C4, 0x800048D0800048CC, 0x001F9DB5001F9DB6,
-    0x001F89BB800048D4, 0x800048D8001FB9BA, 0x001FE1BF800048DC,
-    0x001FF9BD001FFDBE, 0x800048E0001F91BC, 0x800048E4001E71A3,
-    0x001FD9A6800048E8, 0x800048F0800048EC, 0x001FFDAA001FD9A5,
-    0x0003AC84001FCDAB, 0x001FB9AE0003A0BE, 0x001FD5AC001FA5AF,
-    0x001EADA8001FBDAD, 0x001EB5AA001EB5A9, 0x001EC1AC001EC9AB,
-    0x001EF1AE001EC9AD, 0x0001ED80001EF1AF, 0x0001F1820001F181,
-    0x00009486800048F4, 0x800048F800017487, 0x0001C9B8800048FC,
-    0x0001D9BA0001C9B9, 0x0001F9BC0001D1BB, 0x000199BE0001F9BD,
-    0x000145B0000195BF, 0x000145B200014DB1, 0x000145B400015DB3,
-    0x000145B600014DB5, 0x800049000001F9B7, 0x001E35B380004904,
-    0x001E41B580004908, 0x800049108000490C, 0x80004814001E41B6,
-    0x000181BB00012484, 0x000181BD000189BA, 0x000181BF000199BC,
-    0x001E7DA3000189BE, 0x8000491880004914, 0x800049208000491C,
-    0x001E09A5001E09A6, 0x0001C9AB80004924, 0x800049280001C1AA,
-    0x0001C9AF8000492C, 0x0001C9AD0001C1AE, 0x800049300001D1AC,
-    0x8000493880004934, 0x800049408000493C, 0x80004944000258BE,
-    0x80004948001F44EF, 0x8000494C001E54E1, 0x80004950001E84E3,
-    0x8000495880004954, 0x000079808000495C, 0x0000198200007981,
-    0x800049600003CDB3, 0x0003CDB60003D9B5, 0x0004C08680004968,
-    0x0003A1BA0003EC87, 0x0003A1BC0003A1BB, 0x000391BE0003A9BD,
-    0x8000496C000391BF, 0x8000497480004970, 0x00027CEF80004978,
-    0x800049808000497C, 0x000114E380004984, 0x000180E180004988,
-    0x800049908000498C, 0x8000499880004994, 0x000341A38000499C,
-    0x800049A4800049A0, 0x800049AC800049A8, 0x000355A5000341A6,
-    0x00032DAB800049B0, 0x800049B400032DAA, 0x00031DAF800049B8,
-    0x000325AD00031DAE, 0x00023DA900032DAC, 0x00028DAB00023DA8,
-    0x000285AD00028DAA, 0x0002DDAF000299AC, 0x00022C84000285AE,
-    0x000000CA800049BC, 0x800049C4800049C0, 0x800049CC800049C8,
-    0x00014DB9800049D0, 0x000151BB00014DB8, 0x000179BD00015DBA,
-    0x00016DBF000175BC, 0x0002ADB100016DBE, 0x0002B5B30002ADB0,
-    0x00029DB50002BDB2, 0x00017DB700029DB4, 0x0000BD8200017DB6,
-    0x0000BD80000588BE, 0x800049D80000B581, 0x800049E0800049DC,
-    0x0005A084000188EF, 0x800049E4001E74E1, 0x00041086000154E3,
-    0x800049E800057487, 0x000145B3800049EC, 0x800049F4800049F0,
-    0x800049FC800049F8, 0x000155B5000141B6, 0x00014DBB80004A00,
-    0x80004A04000141BA, 0x00014DBF80004A08, 0x00014DBD00014DBE,
-    0x00065DA800014DBC, 0x00068DAA000685A9, 0x00069DAC000685AB,
-    0x000681AE000681AD, 0x800049D4000681AF, 0x80004A1080004A0C,
-    0x80004A1880004A14, 0x80004A2080004A1C, 0x00068DB880004A24,
-    0x000695BA000695B9, 0x0006BDBC0006A5BB, 0x000775BE000771BD,
-    0x0006C5B000076DBF, 0x0006D9B20006C9B1, 0x0006C9B40006D9B3,
-    0x0006BDB60006C9B5, 0x000605A30006B5B7, 0x80004A2C80004A28,
-    0x80004A3480004A30, 0x000615A5000601A6, 0x00060DAB80004A38,
-    0x80004A3C000601AA, 0x00060DAF80004A40, 0x00060DAD00060DAE,
-    0x00005D8000060DAC, 0x0000658200006D81, 0x0003C4BE000755B3,
-    0x000765B600076DB5, 0x0000E08680004A44, 0x000725BA00031487,
-    0x00072DBC00073DBB, 0x000715BE00071DBD, 0x000691A80000D9BF,
-    0x000691AA000691A9, 0x0006A9AC0006A5AB, 0x0006DDAE0006A9AD,
-    0x80004A480006D5AF, 0x80004A5080004A4C, 0x80004A5880004A54,
-    0x80004A6080004A5C, 0x000171B880004A64, 0x000171BA000171B9,
-    0x0001D5BC000171BB, 0x0001D5BE0001DDBD, 0x0006ADB00001CDBF,
-    0x0006B5B20006B5B1, 0x00069DB4000685B3, 0x000151B6000151B5,
-    0x000639A5000151B7, 0x80004A6C80004A68, 0x80004A70000631A6,
-    0x000601A380004A74, 0x000649AD80004A78, 0x00018DAF000679AC,
-    0x80004A7C000641AE, 0x000669AB80004964, 0x80004A80000671AA,
-    0x80004A8880004A84, 0x80004A9080004A8C, 0x80004A9880004A94,
-    0x80004AA080004A9C, 0x80004AA880004AA4, 0x0000198080004AAC,
-    0x0000058200001981, 0x0002808480004AB0, 0x000774EF000388BE,
-    0x000CB08600030084, 0x80004AB800035487, 0x80004AC080004ABC,
-    0x000C24BE80004AC4, 0x000648E180004AC8, 0x000708E380004ACC,
-    0x80004AD480004AD0, 0x80004ADC80004AD8, 0x80004AE0000365B3,
-    0x00036DB6000365B5, 0x80004AE880004AE4, 0x0003EDBA80004AEC,
-    0x0003E1BC0003E5BB, 0x0003E5BE0003E1BD, 0x80004AF00003CDBF,
-    0x80004AF880004AF4, 0x80004B0080004AFC, 0x80004B0880004B04,
-    0x000381A880004B0C, 0x000381AA000381A9, 0x000381AC000381AB,
-    0x000381AE000381AD, 0x000345B0000381AF, 0x000359B200034DB1,
-    0x00034DB4000359B3, 0x000319B6000365B5, 0x000329B8000319B7,
-    0x000339BA000329B9, 0x000329BC000339BB, 0x000319BE000329BD,
-    0x000229A3000319BF, 0x80004B1480004B10, 0x80004B1C80004B18,
-    0x000229A5000221A6, 0x0002A9AB80004B20, 0x80004B280002A1AA,
-    0x000281AF80004B2C, 0x0002ADAD0002A9AE, 0x00004D800002ADAC,
-    0x0000558200005581, 0x000340E380004B30, 0x0001A0E180004B34,
-    0x00039CEF80004B38, 0x000CC88680004B3C, 0x80004B40000D4087,
-    0x80004B4880004B44, 0x80004B5080004B4C, 0x80004B5880004B54,
-    0x80004B6080004B5C, 0x80004B6880004B64, 0x000CBCBE0001B4EF,
-    0x80004B6C00060CE1, 0x80004B7000061CE3, 0x80004B7880004B74,
-    0x0001E1B380004B7C, 0x80004B8480004B80, 0x80004B8C80004B88,
-    0x0001F5B50001E1B6, 0x0001A9BB80004B90, 0x000DE0840001A1BA,
-    0x000195BF80004B98, 0x0001A5BD000195BE, 0x000E19A80001ADBC,
-    0x000E29AA000E19A9, 0x000E3DAC000E29AB, 0x000E25AE000E21AD,
-    0x80004B24000E5DAF, 0x000FF981000FF982, 0x80004B94000FE180,
-    0x00001C8680004B9C, 0x000EC9B80003A087, 0x000ED9BA000EC9B9,
-    0x000EC9BC000ED9BB, 0x000F5DBE000EC9BD, 0x000E25B0000F55BF,
-    0x000E39B2000E2DB1, 0x000E11B4000E31B3, 0x000EF9B6000E11B5,
-    0x000EA1A3000EF9B7, 0x80004BA480004BA0, 0x80004BAC80004BA8,
-    0x000EB5A5000EA1A6, 0x000EE9AB80004BB0, 0x80004BB4000EE1AA,
-    0x000ED5AF80004BB8, 0x000EE5AD000ED5AE, 0x80004BBC000EEDAC,
-    0x80004BC0000F89B3, 0x000FB5B680004BC4, 0x80004BCC80004BC8,
-    0x000F99BA000F81B5, 0x80004BD0000F99BB, 0x000F79BE80004BD4,
-    0x000F79BC000F79BF, 0x000E29A8000F79BD, 0x000E45AA000E29A9,
-    0x000E4DAC000E59AB, 0x000E69AE000E71AD, 0x80004BD8000E69AF,
-    0x80004BE080004BDC, 0x80004BE880004BE4, 0x80004BF080004BEC,
-    0x000ED9B880004BF4, 0x000EF9BA000EEDB9, 0x000EE9BC000EF9BB,
-    0x000E9DBE000EE9BD, 0x000E19B0000E95BF, 0x000EE9B2000E19B1,
-    0x000EF9B4000EE9B3, 0x000EE9B6000EF9B5, 0x80004BF8000EE9B7,
-    0x000ED5A5000EE1A6, 0x000EDDA380004BFC, 0x0000158100002982,
-    0x000E2DAF00001D80, 0x000E2DAD000E2DAE, 0x000ECDAB000E2DAC,
-    0x80004AB4000ECDAA, 0x0002C88480004C00, 0x00010CBE000141B3,
-    0x000141B680004C08, 0x80004C1080004C0C, 0x00014DBA000151B5,
-    0x00004886000125BB, 0x000125BE00007C87, 0x000131BC00012DBF,
-    0x80004C14000131BD, 0x00030484000ED4EF, 0x80004C18000480BE,
-    0x80004C2080004C1C, 0x00CEA8BF80004C24, 0x80004C2C80004C28,
-    0x80004C300005E084, 0x80004C34000E28E3, 0x80004C38000E34E1,
-    0x0002CDA380004C3C, 0x0002DDA580004C40, 0x80004C4880004C44,
-    0x80004C4C0002CDA6, 0x0002A9AB80004C50, 0x0002BDAD0002C1AA,
-    0x0002A1AF0002BDAC, 0x0002B5A90002A9AE, 0x000201AB000D69A8,
-    0x000201AD000209AA, 0x000231AF000219AC, 0x000568BE000201AE,
-    0x80004C5880004C54, 0x80004C6080004C5C, 0x80004C6880004C64,
-    0x0003E5B980004C6C, 0x0003E5BB0003E5B8, 0x0003E5BD0003E5BA,
-    0x0003E5BF0003E5BC, 0x000221B10003E5BE, 0x000225B3000249B0,
-    0x000229B5000225B2, 0x000215B7000221B4, 0x0002CDA8000215B6,
-    0x0002D1AA0002D1A9, 0x000101AC00010DAB, 0x000101AE000101AD,
-    0x80004C70000101AF, 0x80004C7C80004C78, 0x000730BE80004C80,
-    0x80004C8880004C84, 0x000119B880004C8C, 0x00012DBA000125B9,
-    0x0001F9BC0001E9BB, 0x0001E9BE0001F9BD, 0x000145B00001E9BF,
-    0x000159B200014DB1, 0x000131B4000155B3, 0x000135B6000131B5,
-    0x0005A086000129B7, 0x80004C900005D487, 0x000129800000BCEF,
-    0x0000B98200012981, 0x00076C84000EC4EF, 0x80004C94000ED4E1,
-    0x80004C98000E54E3, 0x80004C9C000194E1, 0x000219B300006CE3,
-    0x80004CA480004CA0, 0x0038008480004CA8, 0x0001C9B50001C1B6,
-    0x0001C9BB80004CAC, 0x80004CB00001C1BA, 0x0001B5BF80004CB4,
-    0x0001C1BD0001B5BE, 0x0005D9A30001CDBC, 0x80004CB880004C74,
-    0x80004CC080004CBC, 0x000609A5000601A6, 0x000609AB80004CC4,
-    0x80004CC8000601AA, 0x000675AF80004CCC, 0x000601AD000675AE,
-    0x80004CD000060DAC, 0x80004CD880004CD4, 0x0000198080004CDC,
-    0x0000058200001981, 0x80004CE480004CE0, 0x0003288680004CE8,
-    0x80004CEC0003AC87, 0x80004CF480004CF0, 0x0007A5A880004CF8,
-    0x0007A5AA0007ADA9, 0x0007A9AC0007B9AB, 0x000711AE0007A9AD,
-    0x80004CFC000735AF, 0x80004D0480004D00, 0x80004D0C80004D08,
-    0x80004D1480004D10, 0x0000C9B880004D18, 0x0000D9BA0000C9B9,
-    0x0000F9BC0000D1BB, 0x000099BE0000F9BD, 0x000751B0000099BF,
-    0x00072DB200075DB1, 0x00073DB4000725B3, 0x000711B6000729B5,
-    0x00062DB3000711B7, 0x80004D2080004D1C, 0x80004D2880004D24,
-    0x000621B5000629B6, 0x000699BB80004D2C, 0x80004D30000695BA,
-    0x0006EDBF80004D34, 0x0006F1BD0006F9BE, 0x80004D380006F9BC,
-    0x80004D3C000669A3, 0x00066DA680004D40, 0x80004D4880004D44,
-    0x0006D1AA000665A5, 0x80004D4C0006DDAB, 0x0006BDAE80004D50,
-    0x0006BDAC0006A9AF, 0x00012DB00006B5AD, 0x00013DB2000135B1,
-    0x00012DB4000135B3, 0x000091B6000091B5, 0x0000B1B8000091B7,
-    0x000081BA0000B9B9, 0x0000F1BC000081BB, 0x0000F5BE0000FDBD,
-    0x000099800000E9BF, 0x0000A5820000AD81, 0x80004D5480004C04,
-    0x0001188780004D58, 0x000185A800007C86, 0x00019DAA000195A9,
-    0x00018DAC000195AB, 0x00015DAE000155AD, 0x000275B3000155AF,
-    0x80004D6080004D5C, 0x80004D6880004D64, 0x000265B50002B5B6,
-    0x000291BB80004D6C, 0x80004D70000289BA, 0x000335BF80004D74,
-    0x000281BD000289BE, 0x000658E3000289BC, 0x0005B8E180004D78,
-    0x0000009180004D7C, 0x80004D80000190E1, 0x80004D8400079CE3,
-    0x80004D8C80004D88, 0x000684EF00024CEF, 0x80004D9480004D90,
-    0x80004D9C80004D98, 0x80004DA480004DA0, 0x80004DA80002F9A3,
-    0x000239A60002E9A5, 0x80004DB080004DAC, 0x000205AA80004DB4,
-    0x000205AC00021DAB, 0x000205AE00020DAD, 0x00000D800003B9AF,
-    0x00001D8200001581, 0x80004DBC80004DB8, 0x003C54BE80004DC0,
-    0x003CA485001F9CEF, 0x80004DC8000210E1, 0x003E5086003F48E3,
-    0x80004DCC00033487, 0x0002D5A880004DD0, 0x000101AA000101A9,
-    0x000101AC000101AB, 0x000131AE000109AD, 0x80004DD4000131AF,
-    0x80004DDC80004DD8, 0x80004DE480004DE0, 0x80004DEC80004DE8,
-    0x00013DB880004DF0, 0x0001CDBA0001C5B9, 0x0001C1BC0001C5BB,
-    0x0001F9BE0001C9BD, 0x000125B00001F9BF, 0x000131B2000129B1,
-    0x000115B400010DB3, 0x00010DB600011DB5, 0x000618E1000105B7,
-    0x0007B4E380004DF4, 0x80004DFC80004DF8, 0x80004E0480004E00,
-    0x80004E0C80004E08, 0x80004E1480004E10, 0x80004E1C80004E18,
-    0x0007E0EF80004E20, 0x0000118080004E24, 0x0000118200001181,
-    0x80004E28003C4084, 0x80004E3080004E2C, 0x003FECBE000614EF,
-    0x80004E38000194E1, 0x003E508600065CE3, 0x80004E3C003C0087,
-    0x003E09A380004E40, 0x80004DC40024258F, 0x80004E3400000884,
-    0x003E09A5003E09A6, 0x003E6DAB80004E44, 0x80004E48003E61AA,
-    0x003E59AF80004E4C, 0x003E59AD003E51AE, 0x0038619E003E65AC,
-    0x0034959C0038759F, 0x0035E59A0039F59D, 0x003079980034759B,
-    0x002CA19600317199, 0x002C6D940031C597, 0x00282192002C6595,
-    0x0029F59000290993, 0x00140DB100287991, 0x001801B3001405B0,
-    0x001879B5001455B2, 0x80004E500018B5B4, 0x0001E0BE80004E54,
-    0x80004E5C80004E58, 0x003CD1A280004E60, 0x003D75A0003D65A3,
-    0x0000C9A1003CB9A1, 0x000401A380004E64, 0x00047DA50000DDA2,
-    0x0008F9A70004F5A4, 0x000C01A9000801A6, 0x000CCDAB00086DA8,
-    0x0010D1AD000CCDAA, 0x0011D5AF001001AC, 0x80004E680010D9AE,
-    0x003F15B6003F15B5, 0x80004E7080004E6C, 0x80004E7880004E74,
-    0x003F79BC003F15B3, 0x003F4DBE003F45BD, 0x80004E7C003F45BF,
-    0x003F7DBA80004E80, 0x003FBDB3003F71BB, 0x80004E8880004E84,
-    0x80004E9080004E8C, 0x003E7DB5003E7DB6, 0x003E59BB80004E94,
-    0x80004E98003E51BA, 0x003EE1BF80004E9C, 0x003EFDBD003EF9BE,
-    0x0025A182003E41BC, 0x80004EA000280983, 0x0028ED8680004EA8,
-    0x00299D84002D9187, 0x002DA98A0029CD85, 0x00070886002D658B,
-    0x00309D8E00008C87, 0x0030198C0034018F, 0x0035E19200319D8D,
-    0x80004EAC00351993, 0x0039E19680004EB0, 0x00386194003C1997,
-    0x003DE59A0039FD95, 0x80004EB4003D859B, 0x00009D8180004EB8,
-    0x003C299C00009D80, 0x003CE9A800007582, 0x003CF5AA003CFDA9,
-    0x003D35AC003D2DAB, 0x003D35AE003D3DAD, 0x80004EBC003D2DAF,
-    0x80004EC480004EC0, 0x80004ECC80004EC8, 0x000470BE80004ED0,
-    0x003DDDB880004ED4, 0x003DEDBA003DE5B9, 0x003DFDBC003DE5BB,
-    0x003DEDBE003DE5BD, 0x003D55B0003DE1BF, 0x003D55B2003D5DB1,
-    0x003DFDB4003DE9B3, 0x003DEDB6003DE5B5, 0x003C7DA3003DE5B7,
-    0x80004EDC80004ED8, 0x80004EE00005C4BE, 0x003DBDA5003DBDA6,
-    0x003D99AB80004EE4, 0x80004EE8003D91AA, 0x003D21AF80004EEC,
-    0x003D3DAD003D39AE, 0x80004EF0003D81AC, 0x003D35B380004EF4,
-    0x0002F9B580004EF8, 0x80004F0080004EFC, 0x80004F040002F1B6,
-    0x000241BB80004F08, 0x000235BD000249BA, 0x0003F5BF000259BC,
-    0x003CE9A80003FDBE, 0x003CF5AA003CFDA9, 0x000339AC000311AB,
-    0x000381AE000339AD, 0x0001C580000381AF, 0x0000198200000981,
-    0x80004F1080004F0C, 0x80004F1800057884, 0x000259B880004F1C,
-    0x00027DBA000275B9, 0x000219BC000209BB, 0x000209BE000219BD,
-    0x0003C5B0000201BF, 0x0003C5B20003CDB1, 0x000271B40003A9B3,
-    0x000269B6000279B5, 0x80004F20000269B7, 0x0004C484003E58E3,
-    0x80004F24003F64E1, 0x80004F28003E7CE1, 0x00040086000180E3,
-    0x80004F2C00048487, 0x001A14850001F4EF, 0x80004F30003EACEF,
-    0x80004F3880004F34, 0x80004F4080004F3C, 0x00079084003C31A3,
-    0x0003F5A60003FDA5, 0x80004F4880004F44, 0x00034DAA80004F4C,
-    0x00035DAC000345AB, 0x0002F9AE000331AD, 0x00058DA80002F1AF,
-    0x0005C5AA0005CDA9, 0x0005FDAC0005E5AB, 0x00063DAE000635AD,
-    0x80004F14000635AF, 0x80004F5480004F50, 0x80004F5C80004F58,
-    0x80004F6480004F60, 0x0006E9B880004F68, 0x000689BA0006E9B9,
-    0x00069DBC000689BB, 0x000681BE000681BD, 0x00064DB00006BDBF,
-    0x00065DB2000655B1, 0x0006F5B4000651B3, 0x0006E9B60006FDB5,
-    0x00054DB30006E1B7, 0x80004F7080004F6C, 0x80004F7880004F74,
-    0x000575B50006B5B6, 0x000691BB80004F7C, 0x80004F80000689BA,
-    0x000639BF80004F84, 0x000685BD000631BE, 0x80004F88000689BC,
-    0x80004EA4000509A3, 0x0006F1A680004F8C, 0x80004F9480004F90,
-    0x0006CDAA000531A5, 0x000048860006D5AB, 0x000675AE00006C87,
-    0x0006CDAC00067DAF, 0x000639A80006C1AD, 0x000689AA000639A9,
-    0x000689AC00069DAB, 0x0006A1AE0006BDAD, 0x0001B1800006A1AF,
-    0x000181820001B981, 0x80004F9C80004F98, 0x80004FA480004FA0,
-    0x00016DB880004FA8, 0x00017DBA000175B9, 0x00016DBC000175BB,
-    0x0001D1BE0001D1BD, 0x0006A1B00001D1BF, 0x0006A1B20006A1B1,
-    0x000691B40006B1B3, 0x00015DB6000691B5, 0x000601B3000155B7,
-    0x80004FB080004FAC, 0x80004FB880004FB4, 0x000601B500063DB6,
-    0x000665BB80004FBC, 0x80004FC0000665BA, 0x000655BF80004FC4,
-    0x000665BD000655BE, 0x80004FC8000675BC, 0x80004FCC000645A3,
-    0x000679A680004FD0, 0x80004FD880004FD4, 0x000621AA000645A5,
-    0x80004FDC000621AB, 0x000611AE80004FE0, 0x000631AC000611AF,
-    0x0002C1A8000621AD, 0x0002DDAA0002C9A9, 0x00033DAC0002D5AB,
-    0x00032DAE000325AD, 0x80004FE4000325AF, 0x80004FEC80004FE8,
-    0x80004FF880004FF0, 0x8000500080004FFC, 0x000381B8000270BE,
-    0x000385BA00038DB9, 0x000389BC000399BB, 0x0003B5BE0003BDBD,
-    0x00035DB000007DBF, 0x0003E1B20003E1B1, 0x0003E1B40003E1B3,
-    0x0003D1B60003E9B5, 0x0000A5800003D1B7, 0x0000A5820000AD81,
-    0x80005004000259B3, 0x0003A9B60003B9B5, 0x000C008680005008,
-    0x000395BA0003F487, 0x000339BC000329BB, 0x000329BE000339BD,
-    0x000215A3000321BF, 0x800050108000500C, 0x8000501880005014,
-    0x0003F5A50003E5A6, 0x000365AB8000501C, 0x800050200003D9AA,
-    0x00036DAF80005024, 0x000375AD000365AE, 0x80005028000375AC,
-    0x800050308000502C, 0x0003D4EF80005034, 0x8000503C80005038,
-    0x0006E0E380005040, 0x0001D8E180005044, 0x8000504C80005048,
-    0x8000505480005050, 0x8000505C80005058, 0x00024484000660E3,
-    0x00006D800004FCE1, 0x0000058200001581, 0x000D6CBE80005060,
-    0x000CC88680005068, 0x8000506C000D5C87, 0x800050700007B4EF,
-    0x8000507880005074, 0x800050808000507C, 0x8000508880005084,
-    0x800050908000508C, 0x000DA0840007E8EF, 0x8000509400061CE1,
-    0x80005098000634E3, 0x800050A08000509C, 0x0001F9B3800050A4,
-    0x800050AC800050A8, 0x800050B4800050B0, 0x000145B5000111B6,
-    0x00013DBB800050B8, 0x800050BC000135BA, 0x0000F5BF800050C0,
-    0x000115BD0000F5BE, 0x000E31A8000115BC, 0x000E31AA000E31A9,
-    0x000E8DAC000E9DAB, 0x000ED1AE000ED1AD, 0x80005064000ED1AF,
-    0x800050C8800050C4, 0x00001980800050CC, 0x0000058200001981,
-    0x000F55B8800050D0, 0x000F55BA000F5DB9, 0x000F7DBC000F69BB,
-    0x000F6DBE000F65BD, 0x000EB1B0000F65BF, 0x000E81B2000EB9B1,
-    0x000F71B4000E81B3, 0x000F71B6000F71B5, 0x000EB9A3000F71B7,
-    0x800050D4000388BE, 0x800050DC800050D8, 0x000E05A5000E51A6,
-    0x000E7DAB800050E0, 0x00000886000E75AA, 0x000FB5AF00036C87,
-    0x000E55AD000FB5AE, 0x800050E4000E55AC, 0x800050EC800050E8,
-    0x800050F0000FB5B3, 0x000F5DB6000F55B5, 0x800050F480004FF4,
-    0x000F79BA800050F8, 0x000F69BC000F79BB, 0x000F49BE000F5DBD,
-    0x000E65A8000F49BF, 0x000E71AA000E75A9, 0x000E91AC000E71AB,
-    0x000E91AE000E91AD, 0x800050FC000E91AF, 0x8000510480005100,
-    0x8000510C80005108, 0x8000511480005110, 0x000E85B880005118,
-    0x000E85BA000E8DB9, 0x000E8DBC000E9DBB, 0x000EB5BE000EBDBD,
-    0x000EF1B0000179BF, 0x000EF1B2000EF1B1, 0x000EC1B4000EC5B3,
-    0x000EC1B6000EC1B5, 0x000EF9A3000EC1B7, 0x800051208000511C,
-    0x8000512880005124, 0x000E19A5000E11A6, 0x000E35AB8000512C,
-    0x80005130000E35AA, 0x000E05AF80005134, 0x000E11AD000E05AE,
-    0x00000D80000E25AC, 0x00001D8200001581, 0x8000513C80005138,
-    0x0001948480005140, 0x00074086000194BE, 0x800051480000F487,
-    0x800051508000514C, 0x8000515880005154, 0x00028DA88000515C,
-    0x000295AA000295A9, 0x0002D5AC0002CDAB, 0x0002C9AE0002DDAD,
-    0x800051600002FDAF, 0x8000516880005164, 0x0000008C8000516C,
-    0x8000517480005170, 0x00037DB880005178, 0x0003C1BA0003C1B9,
-    0x0003C1BC0003C1BB, 0x0003F1BE0003C9BD, 0x000285B00003F1BF,
-    0x00034DB2000345B1, 0x00035DB4000345B3, 0x00034DB6000345B5,
-    0x00021DB3000345B7, 0x800051808000517C, 0x8000518880005184,
-    0x00025DB500025DB6, 0x000381BB8000518C, 0x80005190000241BA,
-    0x000381BF80005194, 0x000391BD000399BE, 0x80005198000399BC,
-    0x8000519C000259A3, 0x000219A6800051A0, 0x800051A8800051A4,
-    0x000205AA000219A5, 0x800051AC0003C5AB, 0x0003DDAE800051B0,
-    0x0003DDAC0003C5AF, 0x800051B80003D5AD, 0x000204840000CCE3,
-    0x0001C9800001BCE1, 0x0001F5820001FD81, 0x800051BC000510BE,
-    0x800051C0000428BE, 0x800051C8800051C4, 0x800051CC000010EF,
-    0x0004E086800051D0, 0x800051D40002F487, 0x000E08E3800051D8,
-    0x000FE0E1800051DC, 0x000EF8EF800051E0, 0x800051E8800051E4,
-    0x800051F0800051EC, 0x800051F8800051F4, 0x80005200800051FC,
-    0x8000520880005204, 0x800052108000520C, 0x8000521880005214,
-    0x001524850001E0EF, 0x8000521C000ED4E1, 0x00002980000E1CE3,
-    0x0000058200001D81, 0x000231B380005220, 0x0005CC8480005228,
-    0x800052308000522C, 0x000221B5000229B6, 0x0001CDBB80005234,
-    0x800052380001CDBA, 0x000175BF8000523C, 0x0001C1BD0001C9BE,
-    0x0005E9A80001C9BC, 0x0005F9AA0005E9A9, 0x0005E9AC0005F9AB,
-    0x000639AE0005E9AD, 0x800051B4000639AF, 0x0000888680005224,
-    0x8000524000030087, 0x8000524880005244, 0x0006D1B88000524C,
-    0x0006E1BA0006D9B9, 0x000691BC0006E1BB, 0x000695BE00069DBD,
-    0x000649B0000689BF, 0x00065DB2000649B1, 0x00064DB4000655B3,
-    0x0006F1B60006F1B5, 0x000571A30006F1B7, 0x8000525480005250,
-    0x8000525C80005258, 0x000561A5000569A6, 0x00068DAB80005260,
-    0x8000526400068DAA, 0x000635AF80005268, 0x000681AD000689AE,
-    0x8000526C000689AC, 0x0007E1B380005270, 0x0007E1B580005274,
-    0x8000527C80005278, 0x800051440007E5B6, 0x0007BDBB80005280,
-    0x000795BD0007A1BA, 0x000795BF0007A9BC, 0x000601A8000795BE,
-    0x000621AA000625A9, 0x000621AC000621AB, 0x000625AE00062DAD,
-    0x80005284000655AF, 0x00001D8100001D82, 0x8000528800001D80,
-    0x800052908000528C, 0x000639B8000130BE, 0x0006C9BA000639B9,
-    0x0006D9BC0006C9BB, 0x0006C9BE0006D9BD, 0x00062DB00006C9BF,
-    0x000635B2000631B1, 0x000619B4000609B3, 0x000609B6000619B5,
-    0x0006A9A3000609B7, 0x001F288600028C84, 0x8000529800014487,
-    0x0006A9A50006ADA6, 0x0006F5AB8000529C, 0x800052A00006E9AA,
-    0x0006DDAF800052A4, 0x0006DDAD0006DDAE, 0x800052A80006E1AC,
-    0x800052AC000615B3, 0x00063DB6800052B0, 0x800052B8800052B4,
-    0x0001D9BA000635B5, 0x800052BC0001D9BB, 0x00017DBE800052C0,
-    0x00017DBC000165BF, 0x0005C5A8000175BD, 0x0005D9AA0005C9A9,
-    0x0005F9AC0005D1AB, 0x000229AE0005F9AD, 0x800052C4000229AF,
-    0x800052CC800052C8, 0x0000008C800052D0, 0x800052D8800052D4,
-    0x0002EDB8800052DC, 0x00028DBA000285B9, 0x000285BC000281BB,
-    0x0002BDBE00028DBD, 0x000259B000037DBF, 0x0002EDB2000259B1,
-    0x0002FDB40002E5B3, 0x0002E5B60002E5B5, 0x000551A30002D5B7,
-    0x800052E4800052E0, 0x800052EC800052E8, 0x000571A5000579A6,
-    0x00029DAB800052F0, 0x800052F400029DAA, 0x000221AF800052F8,
-    0x000231AD000239AE, 0x00001182000239AC, 0x00006580800052FC,
-    0x8000530000000981, 0x80005308000398BE, 0x000324848000530C,
-    0x0003688780005310, 0x80005314001C8C86, 0x8000531C80005318,
-    0x8000532480005320, 0x0002A1B380005328, 0x0002D1B5001CC084,
-    0x800053308000532C, 0x800053340002D5B6, 0x0002C1BB80005338,
-    0x000111BD0002F5BA, 0x000111BF0002D9BC, 0x8000533C000119BE,
-    0x8000534480005340, 0x8000534C80005348, 0x8000535480005350,
-    0x800053580006B8EF, 0x8000535C00049CE1, 0x80005360000674E3,
-    0x8000536880005364, 0x0000F9808000536C, 0x000005820000F981,
-    0x001C58BE80005370, 0x00005CEF001F5884, 0x000190E1000640EF,
-    0x00003CE30006FCE1, 0x80005378000674E3, 0x001C98868000537C,
-    0x000269A3001CF487, 0x80005380001F0CBE, 0x8000538880005384,
-    0x000219A500021DA6, 0x000209AB8000538C, 0x8000539000023DAA,
-    0x0001D9AF80005394, 0x0001D9AD0001D1AE, 0x001D91A8000211AC,
-    0x001DA1AA001D99A9, 0x001DD1AC001DA1AB, 0x001DD5AE001DDDAD,
-    0x80005304001DC9AF, 0x8000539880005374, 0x800053A08000539C,
-    0x800053A8800053A4, 0x001E79B8800053AC, 0x001ECDBA001E79B9,
-    0x001EDDBC001EC5BB, 0x001EC5BE001EC5BD, 0x001DB9B0001EF5BF,
-    0x001D85B2001D8DB1, 0x001E55B4001E4DB3, 0x001E55B6001E5DB5,
-    0x001FCDB8001E49B7, 0x001FDDBA001FD5B9, 0x001FF1BC001FD5BB,
-    0x001FE9BE001FFDBD, 0x001FA5B0001FE9BF, 0x001FB1B2001FB1B1,
-    0x001F95B4001F8DB3, 0x001FFDB6001FF5B5, 0x001E19A8001FF5B7,
-    0x001E2DAA001E19A9, 0x001E25AC001E3DAB, 0x001E25AE001E2DAD,
-    0x800053B0001FDDAF, 0x800053B8800053B4, 0x800053C0800053BC,
-    0x800053C480005294, 0x001FE5B3800053C8, 0x800053D0800053CC,
-    0x800053D8800053D4, 0x001FEDB5001FEDB6, 0x001E35BB800053DC,
-    0x800053E0001E21BA, 0x001E71BF800053E4, 0x001E19BD001E11BE,
-    0x00006982001E25BC, 0x00005980001FA1A3, 0x001FA9A600005181,
-    0x800053EC800053E8, 0x001E65AA001FA9A5, 0x00040086001E71AB,
-    0x001E55AE0001EC87, 0x001E61AC001E35AF, 0x001E31A8001E5DAD,
-    0x001E41AA001E31A9, 0x001E41AC001E41AB, 0x001E71AE001E49AD,
-    0x800053F0001E71AF, 0x800053F8800053F4, 0x80005400800053FC,
-    0x8000540880005404, 0x000129B88000540C, 0x000139BA000129B9,
-    0x00012DBC000135BB, 0x0000D1BE0000D1BD, 0x0001C9B00000D1BF,
-    0x0001D9B20001C9B1, 0x0001C9B40001D9B3, 0x000119B60001C9B5,
-    0x001DC9B3000119B7, 0x8000541480005410, 0x8000541C80005418,
-    0x001DC9B5001DC9B6, 0x00020DBB80005420, 0x8000542400020DBA,
-    0x00020DBF80005428, 0x00020DBD00020DBE, 0x8000542C00020DBC,
-    0x80005430001D8DA3, 0x001D8DA680005434, 0x8000543C80005438,
-    0x000249AA001D8DA5, 0x80005440000249AB, 0x000249AE80005444,
-    0x000249AC000249AF, 0x00000D80000249AD, 0x0000118200001181,
-    0x0002CCEF80005448, 0x800054508000544C, 0x00022CE300029084,
-    0x0001D8E1000328BE, 0x0004108680005458, 0x8000545C0003D487,
-    0x000361B380005460, 0x8000546880005464, 0x800054708000546C,
-    0x000371B5000379B6, 0x00035DBB80005474, 0x8000547800035DBA,
-    0x0000E1BF8000547C, 0x0000F1BD0000F9BE, 0x0002A1A30000F9BC,
-    0x8000548480005480, 0x8000548C80005488, 0x0002B1A50002B9A6,
-    0x00029DAB80005490, 0x8000549400029DAA, 0x000121AF80005498,
-    0x000131AD000139AE, 0x8000549C000139AC, 0x001F64EF800054A0,
-    0x800054A8800054A4, 0x800054B0800054AC, 0x00000D80800054B4,
-    0x0000158200001181, 0x001CE0E1800054B8, 0x001F88E3800054BC,
-    0x00028084800054C0, 0x0005488700058CBE, 0x800054C800042C86,
-    0x001EA4EF800054CC, 0x001E00E1001EF4EF, 0x001E2CE3001F50E1,
-    0x800054D0001EB0E3, 0x800054D8800054D4, 0x800054E0800054DC,
-    0x00048484800054E4, 0x800054E8000171B3, 0x000115B600011DB5,
-    0x800054F0800054EC, 0x000101BA800054F4, 0x0000F5BC000101BB,
-    0x0000F5BE0000FDBD, 0x0006ADA80000EDBF, 0x0006B5AA0006BDA9,
-    0x0006D9AC0006C9AB, 0x0006C9AE0006D9AD, 0x800054F80006C1AF,
-    0x80005500800054FC, 0x8000550880005504, 0x800055108000550C,
-    0x00076DB880005514, 0x00070DBA000705B9, 0x000701BC000701BB,
-    0x000701BE000701BD, 0x000689B0000701BF, 0x00076DB2000689B1,
-    0x00077DB4000765B3, 0x000765B6000765B5, 0x80005454000755B7,
-    0x80005518000639A3, 0x00065DA6800054C4, 0x800055208000551C,
-    0x000649AA000655A5, 0x80005524000649AB, 0x0007BDAE80005528,
-    0x0007BDAC0007A5AF, 0x00006D800007B5AD, 0x0000198200000981,
-    0x800055308000552C, 0x8000553400039CBE, 0x0000408680005538,
-    0x8000553C00006087, 0x8000554480005540, 0x8000554C80005548,
-    0x000691A880005550, 0x0006B9AA000691A9, 0x0006A9AC0006B9AB,
-    0x0006D9AE0006A9AD, 0x800055540006D9AF, 0x8000555C80005558,
-    0x8000556480005560, 0x8000556C80005568, 0x000171B880005570,
-    0x000171BA000171B9, 0x0001D9BC000171BB, 0x0001C1BE0001C1BD,
-    0x0006B1B00001F5BF, 0x000689B20006B9B1, 0x000151B4000689B3,
-    0x000151B6000151B5, 0x000631B3000151B7, 0x8000557880005574,
-    0x800055808000557C, 0x000621B5000629B6, 0x00067DBB80005584,
-    0x80005588000675BA, 0x000195BF8000558C, 0x00065DBD000195BE,
-    0x8000559000065DBC, 0x80005594000675A3, 0x00066DA680005598,
-    0x800055A08000559C, 0x000631AA000665A5, 0x0001AC84000639AB,
-    0x0001D1AE0001A0BE, 0x000619AC0001D1AF, 0x0002DDA8000619AD,
-    0x0002E5AA0002EDA9, 0x0002E5AC0002FDAB, 0x0002E5AE0002EDAD,
-    0x800055A800033DAF, 0x800055B0800055AC, 0x800055B4000C64BE,
-    0x800055BC800055B8, 0x0003C9B8800055C0, 0x0003D9BA0003C9B9,
-    0x0003F9BC0003D1BB, 0x000399BE0003F9BD, 0x000345B0000395BF,
-    0x000345B200034DB1, 0x000345B400035DB3, 0x000345B600034DB5,
-    0x000355810003F9B7, 0x000265B300034980, 0x000265B500035582,
-    0x800055C8800055C4, 0x000C0086000265B6, 0x000381BB0003E487,
-    0x000381BD000389BA, 0x000381BF000399BC, 0x00022DA3000389BE,
-    0x800055D0800055CC, 0x800055D8800055D4, 0x00022DA500022DA6,
-    0x0003C9AB800055DC, 0x800055E00003C1AA, 0x0003C9AF800055E4,
-    0x0003C9AD0003C1AE, 0x0006D8E30003D1AC, 0x00069CE10007ACE1,
-    0x800055E800069CE3, 0x800055EC000D5884, 0x800055F4800055F0,
-    0x800055FC800055F8, 0x00011CEF80005600, 0x8000560880005604,
-    0x00007980000640EF, 0x0000118200001581, 0x8000560C000C0084,
-    0x800056100000ACE3, 0x800056180001A4E1, 0x000CC886000070EF,
-    0x8000561C000DA487, 0x8000562480005620, 0x0005C9BA80005628,
-    0x0005A5B80005C9BB, 0x0005F9BE0005CDB9, 0x0005CDBC0005F9BF,
-    0x0005A5B20005C5BD, 0x000611B00005ADB3, 0x0005ADB6000611B1,
-    0x0005B5B400059DB7, 0x000661AA0005ADB5, 0x00069DA8000661AB,
-    0x000661AE000695A9, 0x000671AC000661AF, 0x8000562C000671AD,
-    0x8000563480005630, 0x8000563C80005638, 0x000CB0BE80005640,
-    0x000E21A880005644, 0x000E21AA000E21A9, 0x000E25AC000E3DAB,
-    0x000E25AE000E2DAD, 0x80005614000E89AF, 0x8000564C80005648,
-    0x8000565480005650, 0x8000565C80005658, 0x000F51B880005660,
-    0x000F55BA000F5DB9, 0x000F75BC000F6DBB, 0x000F75BE000F7DBD,
-    0x000EF9B0000F69BF, 0x000EA1B2000EA1B1, 0x000EA1B4000EA1B3,
-    0x000E91B6000EA9B5, 0x000EA9B3000E91B7, 0x800055A480005664,
-    0x8000566C80005668, 0x000EADB5000EADB6, 0x000E65BB80005670,
-    0x80005674000E3DBA, 0x000E6DBF80005678, 0x000E75BD000E65BE,
-    0x00001982000E75BC, 0x00006580000EEDA3, 0x000EE9A600001981,
-    0x800056808000567C, 0x000E79AA000EE9A5, 0x80005684000E21AB,
-    0x000E21AE80005688, 0x000E31AC000E29AF, 0x000E61A8000E31AD,
-    0x000E7DAA000E75A9, 0x000E6DAC000E75AB, 0x000EFDAE000EF5AD,
-    0x00018086000EF5AF, 0x8000568C0001A487, 0x8000569480005690,
-    0x8000569C80005698, 0x000175B8800056A0, 0x000175BA00017DB9,
-    0x0001DDBC0001C9BB, 0x0001C5BE0001C5BD, 0x000E8DB00001F5BF,
-    0x000E9DB2000E95B1, 0x000155B4000E91B3, 0x000155B600015DB5,
-    0x000EF5B300014DB7, 0x800056A8800056A4, 0x800056B0800056AC,
-    0x000EE5B5000E75B6, 0x000E51BB800056B4, 0x800056B8000E49BA,
-    0x000199BF800056BC, 0x000E45BD000191BE, 0x800056C0000E49BC,
-    0x800056C4000EB1A3, 0x000E31A6800056C8, 0x800056D0800056CC,
-    0x000E0DAA000EA1A5, 0x800056D4000E15AB, 0x0001D5AE800056D8,
-    0x000E0DAC0001DDAF, 0x0002EDA8000E01AD, 0x000325AA00032DA9,
-    0x000325AC00033DAB, 0x000325AE00032DAD, 0x800056DC000399AF,
-    0x800056E4800056E0, 0x800056EC800056E8, 0x000274BE800056F0,
-    0x00038DB8800056F8, 0x000391BA000391B9, 0x0003BDBC0003A5BB,
-    0x00007DBE000075BD, 0x0003E9B0000075BF, 0x0003F9B20003E9B1,
-    0x0003D9B40003F1B3, 0x0003B9B60003D9B5, 0x0000AD800003B5B7,
-    0x0000BD820000B581, 0x800056FC0003A1B3, 0x0003A1B60003A1B5,
-    0x0002E08480005700, 0x000321BA80005704, 0x000321BC000321BB,
-    0x000311BE000329BD, 0x0003EDA3000311BF, 0x0004E88680005708,
-    0x8000570C00034C87, 0x0003EDA50003EDA6, 0x00036DAB80005710,
-    0x8000571400036DAA, 0x00035DAF80005718, 0x000365AD00035DAE,
-    0x8000571C00036DAC, 0x0023D4EF000E00E3, 0x80005720000FB4E1,
-    0x8000572880005724, 0x0003A9A18000572C, 0x0003C1A3000FFDA0,
-    0x001284E100030DA2, 0x00F3E4E3000FE0EF, 0x8000573480005730,
-    0x0004848480005738, 0x8000573C000484BE, 0x80005740000394EF,
-    0x8000574880005744, 0x8000574C0003D4E3, 0x80005750000054E1,
-    0x8000575880005754, 0x0000F9808000575C, 0x000005820000F981,
-    0x8000576880005760, 0x001B918F8000576C, 0x0007EC84000E9CEF,
-    0x80005770000ED4E1, 0x80005774000EF0E3, 0x0007188680005778,
-    0x0017E5990004F487, 0x000BE99B0017E598, 0x000AFD9D000B399A,
-    0x000F559F000BF19C, 0x001F2991000F5D9E, 0x001F4593001BF590,
-    0x0013D195001FCD92, 0x0017199700132994, 0x000005A100137596,
-    0x00FC01A3800056F4, 0x00FC01A5000055A2, 0x00F819A700FC11A4,
-    0x00F909A900FD45A6, 0x00F409AB00F821A8, 0x00F001AD00F401AA,
-    0x00F0F1AF00F515AC, 0x00EC01B100F01DAE, 0x00E801B300EC01B0,
-    0x00E84DB500EC11B2, 0x8000576400E851B4, 0x800057808000577C,
-    0x8000578880005784, 0x800057908000578C, 0x8000579880005794,
-    0x0006BDA88000579C, 0x000615AA000615A9, 0x000655AC000629AB,
-    0x000645AE000645AD, 0x0000E8BE000675AF, 0x800057A4800057A0,
-    0x800057AC800057A8, 0x800057B4800057B0, 0x000785B8800057B8,
-    0x000791BA000795B9, 0x0007B1BC000791BB, 0x0007A9BE0007BDBD,
-    0x00060DB00007A9BF, 0x0007F9B20007E5B1, 0x0007E9B40007F9B3,
-    0x0007C5B60007E9B5, 0x00079DB30007BDB7, 0x800057C0800057BC,
-    0x800057C8800057C4, 0x0007DDB50007C5B6, 0x000621BB800057CC,
-    0x800057D00007C9BA, 0x00062DBF800057D8, 0x00062DBD00062DBE,
-    0x0000698200062DBC, 0x000059800007D9A3, 0x000781A600005181,
-    0x800057E0800057DC, 0x00078DAA000799A5, 0x0000C886000665AB,
-    0x000669AE0001A487, 0x000669AC000669AF, 0x000D758F000669AD,
-    0x0011A98D0011A18E, 0x0015198B00116D8C, 0x800057E40015B58A,
-    0x00157D87800057E8, 0x0019898500190986, 0x001DA18300196584,
-    0x800057EC001DBD82, 0x800057F4800057F0, 0x800057FC800057F8,
-    0x0005F19B0005259C, 0x8000580000055D9A, 0x0009399780005804,
-    0x0009699500096996, 0x000DD593000CC194, 0x80005808000DD592,
-    0x000591A88000580C, 0x0005A1AA000591A9, 0x0005A5AC0005BDAB,
-    0x000279AE000249AD, 0x00FCEC84000279AF, 0x8000581480005810,
-    0x8000581C80005818, 0x8000582480005820, 0x0002EDB880005828,
-    0x0002FDBA0002F5B9, 0x0002EDBC0002F5BB, 0x000351BE000351BD,
-    0x000209B0000351BF, 0x000219B2000209B1, 0x000209B4000219B3,
-    0x0002DDB6000209B5, 0x000529B30002D5B7, 0x8000582C00FD4884,
-    0x8000583480005830, 0x00052DB50002C9B6, 0x0002CDBB80005838,
-    0x8000583C0002CDBA, 0x0002CDBF80005840, 0x0002CDBD0002CDBE,
-    0x00006D820002CDBC, 0x0000558000056DA3, 0x00028DA600006581,
-    0x8000584800FC1CBE, 0x000289AA000569A5, 0x00FCC886000289AB,
-    0x000289AE00030487, 0x000289AC000289AF, 0x00022DA8000289AD,
-    0x000251AA000251A9, 0x00028DAC000295AB, 0x0002A5AE0002ADAD,
-    0x8000584C0002DDAF, 0x8000585480005850, 0x8000585C80005858,
-    0x8000586480005860, 0x00016DB880005868, 0x00017DBA000175B9,
-    0x00016DBC000175BB, 0x0001D5BE0001DDBD, 0x0002A5B00001CDBF,
-    0x0002B1B20002ADB1, 0x000289B40002B1B3, 0x00015DB6000289B5,
-    0x000750E1000155B7, 0x0006D0E38000586C, 0x8000587480005870,
-    0x8000587C80005878, 0x8000588480005880, 0x8000588C80005888,
-    0x8000589480005890, 0x00060CEF80005898, 0x8000589C00FC74BE,
-    0x00FCE886800058A0, 0x800058A400FC8487, 0x800058A8000664E3,
-    0x0000158200067CE1, 0x000031800001BCEF, 0x800058AC00001D81,
-    0x800058B8800058B0, 0x00FD59B3800058BC, 0x800057D480005844,
-    0x800058B4800058C0, 0x00FD85B500FDADB6, 0x00FDE9BB800058C4,
-    0x800058C800FDE1BA, 0x00FDCDBF800058CC, 0x00FDE5BD00FDCDBE,
-    0x800058D000FDE5BC, 0x800058D400FD1DA3, 0x00FDE9A6800058D8,
-    0x800058E0800058DC, 0x00FDA5AA00FDC1A5, 0x800058E400FDADAB,
-    0x00FD89AE800058E8, 0x00FDA1AC00FD89AF, 0x00FE59A800FDA1AD,
-    0x00FFADAA00FFA5A9, 0x00FFBDAC00FFA5AB, 0x00FFADAE00FFA5AD,
-    0x800058EC00FFA5AF, 0x800058F4800058F0, 0x800058FC800058F8,
-    0x8000590480005900, 0x00FF65B880005908, 0x00FF65BA00FF6DB9,
-    0x00FF65BC00FF7DBB, 0x00FF1DBE00FF15BD, 0x00FFDDB000FF0DBF,
-    0x00FF65B200FF61B1, 0x00FF65B400FF7DB3, 0x00FF65B600FF6DB5,
-    0x00FE25B300FF5DB7, 0x800059108000590C, 0x8000591880005914,
-    0x00FE35B500FE25B6, 0x00FEA5BB8000591C, 0x8000592000FE19BA,
-    0x00FEA9BF0001E084, 0x00FEA9BD00FEA1BE, 0x0000698200FEB1BC,
-    0x0000598000FE61A3, 0x00FE61A600005181, 0x8000592400019CBE,
-    0x00FE5DAA00FE71A5, 0x0000C88600FEE1AB, 0x00FEE5AE0001FC87,
-    0x00FEF5AC00FEEDAF, 0x00FE79A800FEEDAD, 0x000185AA00FE79A9,
-    0x0001B9AC000191AB, 0x0001D9AE0001B9AD, 0x0000A0840001D5AF,
-    0x8000592C80005928, 0x8000593480005930, 0x8000593C80005938,
-    0x00016DB880005940, 0x00017DBA000175B9, 0x00016DBC000175BB,
-    0x0000D1BE0000D1BD, 0x0001ADB00000D1BF, 0x0001B1B20001B1B1,
-    0x00019DB4000185B3, 0x00015DB6000155B5, 0x00020DB3000155B7,
-    0x8000594880005944, 0x800059508000594C, 0x000231B5000239B6,
-    0x000275BB80005954, 0x80005958000269BA, 0x000291BF8000595C,
-    0x000251BD000259BE, 0x80005960000265BC, 0x80005964000249A3,
-    0x00027DA680005968, 0x800059708000596C, 0x00022DAA000275A5,
-    0x80005974000231AB, 0x00021DAE80005978, 0x000221AC0002D5AF,
-    0x8000597C000215AD, 0x8000598480005980, 0x00001D8080005988,
-    0x0000398200000981, 0x800059908000598C, 0x80005998000400BE,
-    0x800059A08000599C, 0x800059A8800059A4, 0x0004E086800059AC,
-    0x000544BE00030487, 0x800059B000044084, 0x800059B4000164E3,
-    0x800059B80001ACE1, 0x800059BC0002B8EF, 0x800059C4800059C0,
-    0x800059CC800059C8, 0x0002A5A8800059D0, 0x0002A5AA0002ADA9,
-    0x0002A5AC0002BDAB, 0x0002A5AE0002ADAD, 0x0004A084000119AF,
-    0x800059D8800059D4, 0x800059E0800059DC, 0x800059E8800059E4,
-    0x00010DB80007B0BE, 0x000111BA000111B9, 0x00013DBC000125BB,
-    0x0001DDBE0001D5BD, 0x000169B00001D5BF, 0x000179B2000169B1,
-    0x000159B4000171B3, 0x000139B6000159B5, 0x00FF70E1000135B7,
-    0x00FF1CE300FFB4E3, 0x800059EC00FCFCE1, 0x800059F000FE44E1,
-    0x800059F40000A0E3, 0x800059FC800059F8, 0x0006FCBE000000EF,
-    0x00FEFCEF00FF98EF, 0x0000658080005A00, 0x0000758200006D81,
-    0x80005A08000261B3, 0x000261B6000261B5, 0x0004C08680005A0C,
-    0x0001E1BA0004A487, 0x0001E1BC0001E1BB, 0x0001D1BE0001E9BD,
-    0x800059940001D1BF, 0x000605AB80005A04, 0x00060DAD000605AA,
-    0x000635AF000605AC, 0x80005A10000635AE, 0x000585A380005A14,
-    0x000585A580005A18, 0x80005A2080005A1C, 0x0006F1B3000585A6,
-    0x80005A2880005A24, 0x80005A3080005A2C, 0x0006E1B50006E9B6,
-    0x000689BB80005A34, 0x80005A38000685BA, 0x000689BF80005A3C,
-    0x000689BD000681BE, 0x000629A8000691BC, 0x000639AA000629A9,
-    0x000629AC000639AB, 0x000649AE000629AD, 0x80005A40000645AF,
-    0x80005A4880005A44, 0x80005A5080005A4C, 0x80005A5880005A54,
-    0x0007E9B880005A5C, 0x0007F9BA0007E9B9, 0x0007E9BC0007F9BB,
-    0x000741BE000745BD, 0x000601B000077DBF, 0x000605B200060DB1,
-    0x000605B400061DB3, 0x000605B600060DB5, 0x0007B1A30007D9B7,
-    0x80005A6480005A60, 0x80005A6C80005A68, 0x0007A1A50007A9A6,
-    0x0007C9AB80005A70, 0x80005A740007C5AA, 0x0007C9AF80005A78,
-    0x0007C9AD0007C1AE, 0x00004D800007D1AC, 0x0000558200005581,
-    0x80005A7C00016DB3, 0x00016DB6000179B5, 0x0000C08680005A80,
-    0x000121BA00010487, 0x000121BC000121BB, 0x000111BE000129BD,
-    0x000651A8000111BF, 0x0001A9AA000659A9, 0x0001BDAC0001A9AB,
-    0x0001A1AE0001A1AD, 0x0001AC840001D9AF, 0x80005A8880005A84,
-    0x80005A9080005A8C, 0x80005A9880005A94, 0x000169B880005A9C,
-    0x000179BA000169B9, 0x000169BC000179BB, 0x0000DDBE000169BD,
-    0x0001A9B00000D5BF, 0x0001B9B20001A9B1, 0x000199B40001B1B3,
-    0x000159B6000199B5, 0x000221A3000159B7, 0x80005AA480005AA0,
-    0x80005AAC80005AA8, 0x000235A5000221A6, 0x00026DAB80005AB0,
-    0x80005AB400026DAA, 0x00025DAF80005AB8, 0x000265AD00025DAE,
-    0x80005ABC00026DAC, 0x80005AC480005AC0, 0x80005ACC80005AC8,
-    0x80005AD480005AD0, 0x00003D8080005AD8, 0x0000198200000981,
-    0x80005AE080005ADC, 0x0003448780005AE8, 0x80005AEC000CDC86,
-    0x80005AF480005AF0, 0x80005AF8000CC084, 0x80005AFC0007DCEF,
-    0x80005B0480005B00, 0x80005B08000114E3, 0x80005B0C0001D4E1,
-    0x80005B1480005B10, 0x00039DB380005B18, 0x80005B2080005B1C,
-    0x80005B2880005B24, 0x000371B5000379B6, 0x00035DBB80005B2C,
-    0x80005B3000035DBA, 0x0000E1BF80005B34, 0x0000F1BD0000F9BE,
-    0x00022DA80000F9BC, 0x000255AA000255A9, 0x0002B9AC0002A9AB,
-    0x0002A9AE0002B9AD, 0x000D68BE0002A9AF, 0x80005B3C80005B38,
-    0x80005B4480005B40, 0x0000AD810000AD82, 0x00016DB80000AD80,
-    0x00010DBA000105B9, 0x00011DBC000105BB, 0x00010DBE000105BD,
-    0x0002D9B0000105BF, 0x00016DB20002D9B1, 0x00017DB4000165B3,
-    0x000165B6000165B5, 0x000678E1000155B7, 0x0006B8E300075CE1,
-    0x80005B48000750E3, 0x80005B5080005B4C, 0x80005B5C80005B58,
-    0x80005B6480005B60, 0x000CF8BE80005B68, 0x00067CEF80005B6C,
-    0x00029DA300078CEF, 0x000C688680005B70, 0x80005B74000D0487,
-    0x000271A5000279A6, 0x00025DAB80005B78, 0x80005B7C00025DAA,
-    0x0001E1AF80005B80, 0x0001F1AD0001F9AE, 0x000E75A80001F9AC,
-    0x000E8DAA000E85A9, 0x000E81AC000E85AB, 0x000E85AE000E8DAD,
-    0x80005B54000EBDAF, 0x80005B8880005B84, 0x80005B8C80005AE4,
-    0x80005B9480005B90, 0x000E91B880005B98, 0x000EA1BA000E99B9,
-    0x000F75BC000EA1BB, 0x000F75BE000F7DBD, 0x000EC9B0000F6DBF,
-    0x000EDDB2000ED5B1, 0x000ECDB4000ED5B3, 0x000EB1B6000EB1B5,
-    0x000E05B3000EB1B7, 0x80005BA080005B9C, 0x80005BA880005BA4,
-    0x000E09B5000E0DB6, 0x000E11BB80005BAC, 0x80005BB0000E0DBA,
-    0x000E75BF80005BB4, 0x000E01BD000E09BE, 0x00002982000E09BC,
-    0x00001D80000E41A3, 0x000E49A600001581, 0x80005BBC80005BB8,
-    0x000E49AA000E4DA5, 0x80005BC0000E55AB, 0x000E4DAE0003A084,
-    0x000E4DAC000E31AF, 0x000E9DA8000E45AD, 0x000EADAA000EA5A9,
-    0x000EBDAC000EA5AB, 0x000EDDAE000EA1AD, 0x80005BC4000ECDAF,
-    0x0001888680005BC8, 0x80005BCC0001AC87, 0x80005BD480005BD0,
-    0x00017DB880005BD8, 0x0001C1BA0001C1B9, 0x0001C1BC0001C1BB,
-    0x0001F1BE0001C9BD, 0x000EBDB00001F1BF, 0x000E85B2000E8DB1,
-    0x000159B4000149B3, 0x00014DB6000145B5, 0x000E25B3000145B7,
-    0x80005BE080005BDC, 0x80005BE880005BE4, 0x000E35B5000E21B6,
-    0x000E49BB80005BEC, 0x80005BF0000E79BA, 0x000181BF80005BF4,
-    0x00019DBD00019DBE, 0x80005BF8000E51BC, 0x80005BFC000E61A3,
-    0x000E65A680005C00, 0x80005C0880005C04, 0x000E3DAA000E71A5,
-    0x80005C0C000E0DAB, 0x0001D9AE80005C10, 0x000E15AC0001C5AF,
-    0x000325A80001D9AD, 0x00033DAA000335A9, 0x00032DAC000335AB,
-    0x000391AE000391AD, 0x80005C14000391AF, 0x80005C1C80005C18,
-    0x80005C2480005C20, 0x80005C2C80005C28, 0x000399B880005C30,
-    0x0003A5BA0003ADB9, 0x000075BC00006DBB, 0x000075BE00007DBD,
-    0x0003F1B000006DBF, 0x0003C1B20003F9B1, 0x0003B1B40003C1B3,
-    0x0003B5B60003BDB5, 0x0000B5800003A9B7, 0x0000B5820000BD81,
-    0x80005C3800030CBE, 0x80005C4080005C3C, 0x0002AC840002E8EF,
-    0x80005C44000194E1, 0x80005C4800021CE3, 0x0003D88780005C4C,
-    0x0003F5B300041C86, 0x001FA48580005C50, 0x80005C5880005C54,
-    0x0003E5B5000351B6, 0x00037DBB80005C5C, 0x80005C60000375BA,
-    0x000339BF80005C64, 0x000355BD000335BE, 0x80005C68000355BC,
-    0x80005C7080005C6C, 0x80005C74000239A3, 0x00029DA6000229A5,
-    0x80005C7C80005C78, 0x0002B9AA80005C80, 0x000299AC0002B1AB,
-    0x0002F9AE000299AD, 0x80005C840002F5AF, 0x0002448480005C88,
-    0x80005C8C000540BE, 0x80005C9480005C90, 0x00003D8080005C98,
-    0x0000198200000981, 0x80005CA080005C9C, 0x0005F48780005CA8,
-    0x80005CAC0004DC86, 0x80005CB480005CB0, 0x000E48EF80005CB8,
-    0x80005CC080005CBC, 0x000F3CE3000ED4EF, 0x000FCCE1000E74E1,
-    0x80005CC4000150E3, 0x80005CCC80005CC8, 0x000159B380005CD0,
-    0x80005CD880005CD4, 0x80005CE080005CDC, 0x000169B5000161B6,
-    0x00013DBB80005CE4, 0x80005CE8000135BA, 0x0000F5BF80005CEC,
-    0x000115BD0000F5BE, 0x000691A8000115BC, 0x000691AA000691A9,
-    0x0006A1AC0006A5AB, 0x0006DDAE0006A1AD, 0x80005CA40006D1AF,
-    0x80005CF480005CF0, 0x80005C3480005CF8, 0x80005D0080005CFC,
-    0x00075DB880005D04, 0x00076DBA000765B9, 0x000715BC00070DBB,
-    0x00070DBE000705BD, 0x0006B1B0000705BF, 0x000681B20006B9B1,
-    0x000771B4000681B3, 0x000771B6000771B5, 0x000619A300076DB7,
-    0x0000158100002D82, 0x80005D0800001D80, 0x000629A5000621A6,
-    0x00067DAB80005D0C, 0x80005D10000675AA, 0x0007B5AF0002A484,
-    0x000655AD0007B5AE, 0x0003CCBE000655AC, 0x80005D14000741B3,
-    0x000745B680005D18, 0x80005D2080005D1C, 0x000755BA000751B5,
-    0x0000488600075DBB, 0x000739BE00030C87, 0x000745BC000739BF,
-    0x000675A8000739BD, 0x00068DAA000685A9, 0x00069DAC000685AB,
-    0x000681AE000681AD, 0x80005D240006ADAF, 0x80005D2C80005D28,
-    0x80005D3480005D30, 0x80005D3C80005D38, 0x0006A9B880005D40,
-    0x00016DBA0006A9B9, 0x00017DBC000165BB, 0x000161BE000161BD,
-    0x0006D5B0000161BF, 0x0006D5B20006DDB1, 0x0006B9B40006A9B3,
-    0x0006A9B60006B9B5, 0x000605A30006A1B7, 0x80005D4880005D44,
-    0x80005D5080005D4C, 0x000615A5000601A6, 0x000619AB80005D54,
-    0x80005D58000611AA, 0x00067DAF80005D5C, 0x00067DAD00067DAE,
-    0x80005D60000601AC, 0x80005D6880005D64, 0x80005D7080005D6C,
-    0x80005D7880005D74, 0x0000398080005D7C, 0x0000058200003981,
-    0x0001B8BE80005D80, 0x80005D880001B884, 0x0002CDA880005D8C,
-    0x0002D1AA0002D1A9, 0x000335AC00032DAB, 0x000335AE00033DAD,
-    0x80005D9000032DAF, 0x80005D9880005D94, 0x80005DA080005D9C,
-    0x001C34BE80005DA4, 0x0003E9B880005DA8, 0x000389BA0003E9B9,
-    0x000399BC000389BB, 0x000389BE000399BD, 0x000355B0000389BF,
-    0x000355B200035DB1, 0x0003F9B40003E9B3, 0x0003E9B60003F9B5,
-    0x000259B30003E1B7, 0x0003948780005DAC, 0x80005DB0001C4C86,
-    0x0003BDB50003BDB6, 0x000399BB80005DB4, 0x80005DB8000391BA,
-    0x00033DBF80005DBC, 0x00033DBD00033DBE, 0x80005DC0000381BC,
-    0x80005DC400021DA3, 0x0003F9A680005DC8, 0x80005DD080005DCC,
-    0x0003D5AA0003F9A5, 0x80005DD40003DDAB, 0x000379AE80005DD8,
-    0x0003C5AC000379AF, 0x80005DDC000379AD, 0x0002048480005DE0,
-    0x80005DE4001C00BE, 0x00003D8100003D82, 0x80005DE800003D80,
-    0x80005DF080005DEC, 0x80005DFC80005DF8, 0x80005E0480005E00,
-    0x80005E0C80005E08, 0x80005E100000A0E3, 0x001CF0860001BCE1,
-    0x80005E14001D0487, 0x80005E1C80005E18, 0x80005E2480005E20,
-    0x80005E2C80005E28, 0x80005E3000007CEF, 0x80005E3880005E34,
-    0x80005E4080005E3C, 0x80005E44000638EF, 0x80005E4C80005E48,
-    0x001DA8840001A4EF, 0x80005E500006C0E1, 0x80005E54000110E3,
-    0x80005E58000710E1, 0x0001E1B30006E4E3, 0x80005E6080005E5C,
-    0x80005E6C80005E68, 0x0001F1B50001F9B6, 0x000165BB80005E70,
-    0x001F48860001D9BA, 0x00015DBF001CD087, 0x000165BD00015DBE,
-    0x001E3DB000016DBC, 0x001E99B2001E81B1, 0x001EB9B4001E91B3,
-    0x001ED9B6001EB9B5, 0x001EEDB8001ED5B7, 0x001EF1BA001EF1B9,
-    0x001EDDBC001EC5BB, 0x001F3DBE001F35BD, 0x80005DF4001F2DBF,
-    0x80005E7480005D84, 0x001FE98080005E78, 0x001F9182001F9181,
-    0x001D99A880005E64, 0x001E49AA001D99A9, 0x001E59AC001E49AB,
-    0x001E49AE001E59AD, 0x001EA1A3001E49AF, 0x80005E8080005E7C,
-    0x80005E8880005E84, 0x001EB1A5001EB9A6, 0x001E25AB80005E8C,
-    0x80005E90001E99AA, 0x001E1DAF80005E94, 0x001E25AD001E1DAE,
-    0x80005E98001E2DAC, 0x80005E9C001FF1B3, 0x001F91B680005EA0,
-    0x80005EA880005EA4, 0x001FBDBA001FE1B5, 0x80005EAC001F85BB,
-    0x001F9DBE80005EB0, 0x001F9DBC001F85BF, 0x001E2DA8001F95BD,
-    0x001E3DAA001E35A9, 0x001E2DAC001E35AB, 0x001E51AE001E55AD,
-    0x80005EB4001E45AF, 0x80005EBC80005EB8, 0x80005EC480005EC0,
-    0x0000B1810000B982, 0x001EC1B80000B980, 0x001EC1BA001EC1B9,
-    0x001EC5BC001EDDBB, 0x001EF1BE001ECDBD, 0x001E3DB0001EEDBF,
-    0x001E0DB2001E05B1, 0x001E1DB4001E05B3, 0x001E01B6001E05B5,
-    0x001EB5A3001E01B7, 0x80005ECC80005EC8, 0x0002988480005ED0,
-    0x001EA5A5001ED5A6, 0x001EC1AB000190BE, 0x00000886001EF9AA,
-    0x001EC1AF00017C87, 0x001ED1AD001ED9AE, 0x80005ED8001ED9AC,
-    0x80005EDC001E19B3, 0x000155B680005EE0, 0x80005EE880005EE4,
-    0x000171BA000191B5, 0x80005EEC000179BB, 0x000131BE80005EF0,
-    0x000151BC00013DBF, 0x001DCDA8000151BD, 0x001DD5AA001DD5A9,
-    0x00025DAC000245AB, 0x000291AE000291AD, 0x80005EF4000291AF,
-    0x80005EFC80005EF8, 0x80005F0480005F00, 0x80005F0C80005F08,
-    0x000299B880005F10, 0x0002A5BA0002ADB9, 0x000375BC00036DBB,
-    0x000375BE00037DBD, 0x0002F1B000036DBF, 0x0002C1B20002F9B1,
-    0x0002B1B40002C1B3, 0x0002B5B60002BDB5, 0x001D5DA30002A9B7,
-    0x80005F1880005F14, 0x80005F2080005F1C, 0x0002D5A5000211A6,
-    0x00023DAB80005F24, 0x80005F28000235AA, 0x000279AF80005F2C,
-    0x000215AD000275AE, 0x00044CBE000215AC, 0x80005F3480005F30,
-    0x00001D8080005F38, 0x0000398200000981, 0x80005F4080005F3C,
-    0x0004C8860003A884, 0x80005F480004A887, 0x80005F5080005F4C,
-    0x0002C1A880005F54, 0x0002C5AA0002CDA9, 0x0002C9AC0002D9AB,
-    0x0002F5AE0002FDAD, 0x80005F5800013DAF, 0x80005F6080005F5C,
-    0x80005F6880005F64, 0x80005F7080005F6C, 0x0001C9B80004F4BE,
-    0x0001D9BA0001C9B9, 0x0001F9BC0001D1BB, 0x00019DBE0001F9BD,
-    0x000145B0000195BF, 0x000145B200014DB1, 0x000145B400015DB3,
-    0x000145B600014DB5, 0x00024DB30001F9B7, 0x80005F7880005F74,
-    0x80005F8080005F7C, 0x000241B500027DB6, 0x000189BB80005F84,
-    0x80005F88000181BA, 0x000189BF80005F8C, 0x000189BD000181BE,
-    0x001F54E1000191BC, 0x001ED0E3001E8CE1, 0x000050E3001E7CE3,
-    0x001E28E180005F90, 0x000084EF80005F94, 0x80005F9C80005F98,
-    0x80005FA480005FA0, 0x001F38EF80005FA8, 0x0005CDA3001E08EF,
-    0x80005FAC80005F44, 0x0003DC8480005FB0, 0x0005C1A50005FDA6,
-    0x000609AB80005FB4, 0x00030886000601AA, 0x000609AF00036487,
-    0x000609AD000601AE, 0x00004981000611AC, 0x80005ED40007B180,
-    0x80005FB800004182, 0x80005FC080005FBC, 0x80005FC880005FC4,
-    0x80005FD080005FCC, 0x80005FD880005FD4, 0x80005FE080005FDC,
-    0x000671A880005FE4, 0x000671AA000671A9, 0x0007B1AC000671AB,
-    0x0007B1AE0007B1AD, 0x80005FE80007B1AF, 0x80005FF080005FEC,
-    0x80005FF880005FF4, 0x8000600080005FFC, 0x000769B880006004,
-    0x00071DBA000769B9, 0x000735BC000711BB, 0x000735BE00073DBD,
-    0x0007D1B000072DBF, 0x0007D1B20007D1B1, 0x000779B4000769B3,
-    0x000769B6000779B5, 0x0007F9B3000761B7, 0x8000600C80006008,
-    0x8000601480006010, 0x000639B5000631B6, 0x000615BB80006018,
-    0x8000601C000615BA, 0x0006BDBF80006020, 0x0006BDBD0006BDBE,
-    0x00006D82000605BC, 0x000055800007BDA3, 0x000675A600006581,
-    0x000190BE00019884, 0x000651AA00067DA5, 0x000C0085000651AB,
-    0x0006F9AE000F60BE, 0x000641AC0006F9AF, 0x0001D1A80006F9AD,
-    0x0001D1AA0001D1A9, 0x0001F1AC0001D1AB, 0x0001F1AE0001F1AD,
-    0x000F20860001F1AF, 0x800060280000E487, 0x800060308000602C,
-    0x8000603880006034, 0x000171B88000603C, 0x000171BA000171B9,
-    0x0000D5BC000171BB, 0x0000D5BE0000DDBD, 0x000191B00000CDBF,
-    0x000191B2000191B1, 0x000151B4000191B3, 0x000151B6000151B5,
-    0x000259B3000151B7, 0x8000604480006040, 0x8000604C80006048,
-    0x000279B5000271B6, 0x000255BB80006050, 0x80006054000255BA,
-    0x000295BF80006058, 0x000245BD000295BE, 0x8000605C000245BC,
-    0x8000606000021DA3, 0x000235A680006064, 0x8000606C80006068,
-    0x000211AA00023DA5, 0x80006070000211AB, 0x0002D1AE80006074,
-    0x000201AC0002D1AF, 0x00EA8582000201AD, 0x8000607800EADD83,
-    0x0016818600030C84, 0x001691840017CD87, 0x0012A18A00169985,
-    0x8000607C0012A18B, 0x001E918E80006080, 0x0013918C001F958F,
-    0x001AF192001E918D, 0x80006084001AE593, 0x00069196002458EF,
-    0x001A119400068997, 0x0002A99A00069995, 0x800060880002E19B,
-    0x800060908000608C, 0x0002E59C000188E1, 0x001B319B003728E3,
-    0x001F2999001E259A, 0x0007519F001F2198, 0x001BD19D0007519E,
-    0x00173193001B3D9C, 0x0017219100172192, 0x0012199700EB7590,
-    0x0013399500133196, 0x0001A98000166594, 0x0001B9820001A981,
-    0x00EB558F0002FDB3, 0x0002E5B60002EDB5, 0x000CC08680006098,
-    0x000165BA000DE887, 0x00016DBC00017DBB, 0x000155BE00015DBD,
-    0x000770E100013DBF, 0x000188E38000609C, 0x0000D4E3800060A0,
-    0x0001A8E1000D9C84, 0x000000EF800060A4, 0x800060AC800060A8,
-    0x800060B4800060B0, 0x00067CEF800060B8, 0x800060C0800060BC,
-    0x000CC4BE000730E3, 0x800060C400047CE1, 0x800060CC800060C8,
-    0x0003FDA3800060D0, 0x0003A1A10003E1A2, 0x800060D40007B1A0,
-    0x800060D80007CCEF, 0x800060E0800060DC, 0x80006024800060E4,
-    0x800060E880006094, 0x800060F0800060EC, 0x000D89A8800060F4,
-    0x000D99AA000D89A9, 0x000D89AC000D99AB, 0x000E3DAE000D89AD,
-    0x000E4DB0000E35AF, 0x000E55B2000E55B1, 0x000E7DB4000E65B3,
-    0x000E11B6000E15B5, 0x000E35B8000E11B7, 0x000E01BA000E3DB9,
-    0x000EF9BC000E01BB, 0x000ED5BE000EDDBD, 0x000D31A3000EB5BF,
-    0x0000158100002D82, 0x800060F800001D80, 0x000D21A5000D29A6,
-    0x000EB1AB800060FC, 0x80006100000EA9AA, 0x000EF1AF80006104,
-    0x000E91AD000E99AE, 0x80006108000EA1AC, 0x00006886000F49B3,
-    0x000F79B600031C87, 0x800061108000610C, 0x000F55BA000F79B5,
-    0x80006114000F59BB, 0x000F25BE80006118, 0x000F39BC000F2DBF,
-    0x000E5DA8000F2DBD, 0x000E85AA000E81A9, 0x000EB5AC000E91AB,
-    0x000EB5AE000EBDAD, 0x8000611C000EADAF, 0x8000612480006120,
-    0x8000612C80006128, 0x8000613480006130, 0x000EB9B880006138,
-    0x000149BA000EB9B9, 0x000141BC000149BB, 0x000141BE000141BD,
-    0x000ED5B0000141BF, 0x000EA9B2000ED9B1, 0x000EA5B4000EBDB3,
-    0x000EA5B6000EADB5, 0x000E0DA3000E8DB7, 0x800061408000613C,
-    0x8000614880006144, 0x000E3DA5000E3DA6, 0x000E1DAB8000614C,
-    0x80006150000E11AA, 0x000E69AF80006154, 0x000E69AD000E61AE,
-    0x80006158000E7DAC, 0x800061608000615C, 0x00000D8080006164,
-    0x00003D8200003581, 0x8000616C80006168, 0x0001A48480006170,
-    0x0007D0860001A4BE, 0x800061780000F887, 0x0002C1A88000617C,
-    0x0002D1AA0002D1A9, 0x0002FDAC0002E5AB, 0x00033DAE000335AD,
-    0x80006180000335AF, 0x8000618880006184, 0x800061908000618C,
-    0x8000619880006194, 0x0003D1B88000619C, 0x0003E1BA0003D9B9,
-    0x000391BC0003E1BB, 0x000391BE000391BD, 0x00034DB0000391BF,
-    0x00035DB2000355B1, 0x00034DB4000355B3, 0x0003F1B60003F1B5,
-    0x000259B30003F1B7, 0x800061A4800061A0, 0x800061A8000400BE,
-    0x000275B50003B5B6, 0x000391BB800061AC, 0x800061B0000389BA,
-    0x000335BF800061B4, 0x000381BD000389BE, 0x800061B8000389BC,
-    0x800061BC00021DA3, 0x0003F1A6800061C0, 0x800061C8800061C4,
-    0x0003CDAA000231A5, 0x800061CC0003D5AB, 0x0003CDAE800061D0,
-    0x0003CDAC000371AF, 0x800061D40003C5AD, 0x0000158100002182,
-    0x000004EF00001580, 0x800061D800029484, 0x0000D8E3800061DC,
-    0x0001F8E1800061E0, 0x800061EC800061E8, 0x800061F4800061F0,
-    0x0004E086800061F8, 0x800061FC00051887, 0x8000620000016DB3,
-    0x000175B600017DB5, 0x8000620880006204, 0x000151BA8000620C,
-    0x0001F5BC000151BB, 0x0001E9BE0001FDBD, 0x000E04E30001E1BF,
-    0x000EF0E1000F18E3, 0x80006210000C34E1, 0x8000621880006214,
-    0x0037ADB28000621C, 0x0032B9B0003719B3, 0x000E80EF003401B1,
-    0x002801B4000C20EF, 0x003C01AA002BBDB5, 0x003BCDA8003FC1AB,
-    0x0033D5AE003BA1A9, 0x003FC9AC0033DDAF, 0x0003EDA2003001AD,
-    0x0004A0BE000349A3, 0x000781A60003E5A1, 0x000401A4003801A7,
-    0x000621A30007F9A5, 0x80006220800061E4, 0x8000622880006224,
-    0x000631A5000639A6, 0x00061DAB8000622C, 0x8000623000061DAA,
-    0x0006ADAF80006234, 0x0006B1AD0006A5AE, 0x800062380006B9AC,
-    0x8000623C000430EF, 0x8000624480006240, 0x00003D8100003D82,
-    0x8000624800003D80, 0x800062508000624C, 0x8000625400032084,
-    0x80006258000570E3, 0x000579A8000794E1, 0x0006A9AA000579A9,
-    0x0006B9AC0006A9AB, 0x0006A9AE0006B9AD, 0x000000860006A9AF,
-    0x8000625C00008487, 0x8000626480006260, 0x8000626C80006268,
-    0x000601B880006270, 0x000605BA00060DB9, 0x000609BC000619BB,
-    0x000635BE00063DBD, 0x0006D9B00007DDBF, 0x00066DB20006D9B1,
-    0x000661B4000665B3, 0x000651B6000669B5, 0x00062DA8000651B7,
-    0x00063DAA000635A9, 0x00062DAC000635AB, 0x000649AE00065DAD,
-    0x80006274000641AF, 0x8000627C80006278, 0x8000628480006280,
-    0x8000628C80006288, 0x000151B880006290, 0x000161BA000159B9,
-    0x000111BC000161BB, 0x000115BE00011DBD, 0x000601B000010DBF,
-    0x0006F9B2000601B1, 0x0006E9B40006F9B3, 0x000171B60006E9B5,
-    0x0006B5B3000171B7, 0x8000629880006294, 0x800062A08000629C,
-    0x0006A5B50006DDB6, 0x0006C5BB800062A4, 0x800062A80006C5BA,
-    0x00062DBF800062AC, 0x00062DBD00062DBE, 0x800062B000062DBC,
-    0x800062B40006F1A3, 0x000699A6800062B8, 0x800062C0800062BC,
-    0x000681AA0006E1A5, 0x0001B4BE000681AB, 0x000669AE0001A884,
-    0x000669AC000669AF, 0x000279A8000669AD, 0x0002A5AA0002ADA9,
-    0x0002A5AC0002BDAB, 0x0002A5AE0002ADAD, 0x0003ED8000033DAF,
-    0x0003F1820003F181, 0x800062CC800062C8, 0x0003AC87800062D0,
-    0x0003C9B800039C86, 0x0003D9BA0003C9B9, 0x0003F9BC0003D1BB,
-    0x000399BE0003F9BD, 0x000345B0000395BF, 0x000345B200034DB1,
-    0x000345B400035DB3, 0x000345B600034DB5, 0x800062D40003F9B7,
-    0x000225B3800062D8, 0x000225B5003C4084, 0x800062E0800062DC,
-    0x800062E4000225B6, 0x000381BB800062E8, 0x000381BD000389BA,
-    0x000381BF000399BC, 0x800062EC000389BE, 0x00027DA500027DA6,
-    0x00027DA3800062F0, 0x800062F8800062F4, 0x0003D9AF800062FC,
-    0x0003D9AD0003D1AE, 0x0003D9AB0003C1AC, 0x800063000003D1AA,
-    0x00012DA880006304, 0x00013DAA000135A9, 0x000195AC00018DAB,
-    0x000181AE000181AD, 0x800063080001BDAF, 0x800063108000630C,
-    0x8000631880006314, 0x800063208000631C, 0x000061B880006324,
-    0x000061BA000061B9, 0x000061BC000061BB, 0x000061BE000061BD,
-    0x0001C5B0000061BF, 0x0001A5B20001A5B1, 0x0001A9B40001B9B3,
-    0x000181B60001A9B5, 0x0006D8E1000181B7, 0x0001F4E380006328,
-    0x00001D818000632C, 0x8000633000001180, 0x003DACBE00000582,
-    0x003C088680006338, 0x8000633C003DFC87, 0x0006A8EF80006340,
-    0x8000634880006344, 0x800063508000634C, 0x800063540000DCEF,
-    0x8000635C80006358, 0x8000636480006360, 0x8000636C80006368,
-    0x8000637480006370, 0x8000617400014CE1, 0x003E35B3000000E3,
-    0x8000637880006334, 0x800063808000637C, 0x003E9DB5003E81B6,
-    0x003EE1BB80006384, 0x80006388003E81BA, 0x003EEDBF8000638C,
-    0x003EEDBD003EEDBE, 0x80006390003EF1BC, 0x003D10EF80006394,
-    0x00001D8080006398, 0x0000398200000981, 0x800063A08000639C,
-    0x000364BE00026C84, 0x003C54E1800063A4, 0x003CE4E3800063A8,
-    0x003EB5A3800063AC, 0x00034887800063B0, 0x800063B400002C86,
-    0x003E1DA5003E01A6, 0x003E61AB800063B8, 0x800063BC003E01AA,
-    0x003E6DAF800063C0, 0x003E6DAD003E6DAE, 0x800063C4003E71AC,
-    0x800063C8003F71B3, 0x003F15B6800063CC, 0x800063D4800063D0,
-    0x003F35BA003F1DB5, 0x800063D8003F3DBB, 0x003F15BE800063DC,
-    0x003F25BC003F1DBF, 0x003EA5A8003F1DBD, 0x003EBDAA003EB5A9,
-    0x003ED1AC003EB1AB, 0x003ED1AE003ED1AD, 0x800063E0003ECDAF,
-    0x800063E8800063E4, 0x800063F0800063EC, 0x800063F8800063F4,
-    0x000149B8800063FC, 0x000159BA000149B9, 0x000179BC000151BB,
-    0x000119BE000179BD, 0x003EB5B0000119BF, 0x003EB5B2003EBDB1,
-    0x003E95B4003E8DB3, 0x003E95B6003E9DB5, 0x003E35A3003E8DB7,
-    0x8000640480006400, 0x8000640C80006408, 0x003E59A5003E51A6,
-    0x003E79AB80006410, 0x0002A884003E71AA, 0x003E59AF0001ACBE,
-    0x003E59AD003E51AE, 0x00004D80003E61AC, 0x0000558200005581,
-    0x80006418000149B3, 0x000171B6000149B5, 0x000040868000641C,
-    0x000121BA00011487, 0x000125BC000121BB, 0x000111BE00012DBD,
-    0x000275A8000111BF, 0x00028DAA000279A9, 0x00028DAC00029DAB,
-    0x0002B5AE0002BDAD, 0x800064200002DDAF, 0x8000642880006424,
-    0x800064308000642C, 0x8000643880006434, 0x000369B88000643C,
-    0x000379BA000369B9, 0x000369BC000379BB, 0x0003DDBE000369BD,
-    0x0002A9B00003D5BF, 0x0002BDB20002A9B1, 0x000299B40002B5B3,
-    0x000359B6000299B5, 0x80006440000359B7, 0x8000644880006444,
-    0x00045C84000209A3, 0x000231A6000209A5, 0x800064508000644C,
-    0x000261AA80006454, 0x000265AC000261AB, 0x000251AE00026DAD,
-    0x80006458000251AF, 0x800064608000645C, 0x800064640004C0BE,
-    0x8000646C80006468, 0x8000647480006470, 0x8000647C80006478,
-    0x0000158080006480, 0x0000158200001D81, 0x0002D1A880006488,
-    0x000129AA0002D9A9, 0x000139AC000129AB, 0x000129AE000139AD,
-    0x8000648C000129AF, 0x8000649480006490, 0x8000649C80006498,
-    0x800064A4800064A0, 0x000181B8800064A8, 0x000181BA000181B9,
-    0x000185BC00019DBB, 0x000185BE00018DBD, 0x000159B00001BDBF,
-    0x0001E9B2000159B1, 0x0001F9B40001E9B3, 0x0001D9B60001E9B5,
-    0x003E30E10001D9B7, 0x0001FCE3800064AC, 0x800064B4800064B0,
-    0x00043886800064B8, 0x0005A88400026487, 0x800064C0800064BC,
-    0x800064C8800064C4, 0x000128EF800064CC, 0x0013FC85800064D0,
-    0x800064D8800064D4, 0x000084EF800064DC, 0x800064E4800064E0,
-    0x000058E3800064E8, 0x00017CE1800064EC, 0x00001D82800064F0,
-    0x00002580800062C4, 0x000621B300001981, 0x000304BE80006484,
-    0x800064F4800064F8, 0x000631B5000639B6, 0x00061DBB800064FC,
-    0x0000888600061DBA, 0x0006ADBF00030C87, 0x0006BDBD0006BDBE,
-    0x800065000006BDBC, 0x800065040007BCE1, 0x800065080004CCE3,
-    0x800065108000650C, 0x8000651880006514, 0x800065208000651C,
-    0x8000652880006524, 0x800065308000652C, 0x0006A1A3000448EF,
-    0x8000653880006534, 0x800065408000653C, 0x0006B1A50006B9A6,
-    0x00069DAB80006544, 0x8000654800069DAA, 0x00062DAF8000654C,
-    0x00063DAD00063DAE, 0x8000655000063DAC, 0x800065540007E1B3,
-    0x0007E1B680006558, 0x800065608000655C, 0x0007DDBA0007F1B5,
-    0x0003E084000761BB, 0x000765BE0000E4BE, 0x000775BC000751BF,
-    0x0007FDA800076DBD, 0x00062DAA000625A9, 0x00063DAC000625AB,
-    0x00062DAE000625AD, 0x0001ED80000625AF, 0x0001F1820001F181,
-    0x0000908680006568, 0x8000656C0001F087, 0x0006F1B880006570,
-    0x0006F5BA0006F1B9, 0x000699BC000689BB, 0x00068DBE000699BD,
-    0x00065DB0000685BF, 0x0006E5B20006E1B1, 0x0006E5B40006FDB3,
-    0x0006D1B60006D1B5, 0x0006A5A30006D1B7, 0x8000657880006574,
-    0x800065808000657C, 0x0006B5A50006A5A6, 0x000625AB80006584,
-    0x80006588000699AA, 0x000615AF8000658C, 0x000629AD000621AE,
-    0x80006590000631AC, 0x80006594000185B3, 0x000185B680006598,
-    0x800065A08000659C, 0x000181BA00018DB5, 0x800065A4000181BB,
-    0x000181BE800065A8, 0x000191BC000181BF, 0x00020DA8000191BD,
-    0x000249AA000219A9, 0x000241AC000245AB, 0x000241AE000241AD,
-    0x800065AC000241AF, 0x800065B4800065B0, 0x00033884800065B8,
-    0x800065C0800065BC, 0x0002C9B8800065C4, 0x0002D9BA0002C9B9,
-    0x0002F9BC0002D1BB, 0x000299BE0002F9BD, 0x000201B0000295BF,
-    0x000201B2000201B1, 0x000201B4000201B3, 0x000201B6000201B5,
-    0x800065C80002F9B7, 0x800065CC0002C5A3, 0x0002C5A6800065D0,
-    0x800065DC800065D8, 0x0002C1AA0002CDA5, 0x000C88860002C1AB,
-    0x0002C1AE00031887, 0x0002D1AC0002C1AF, 0x000049810002D1AD,
-    0x000C44BE00004580, 0x800065E000004182, 0x800065E8800065E4,
-    0x800065F0800065EC, 0x800065F8800065F4, 0x000D3C84800065FC,
-    0x000D30BE80006600, 0x00022DA880006604, 0x000235AA00023DA9,
-    0x000299AC000289AB, 0x000289AE000299AD, 0x000FEC84000281AF,
-    0x000FE0BE80006608, 0x800066108000660C, 0x8000661880006614,
-    0x0002ADB88000661C, 0x00016DBA000165B9, 0x00017DBC000165BB,
-    0x000161BE000161BD, 0x0002C9B0000161BF, 0x0002A9B20002C9B1,
-    0x0002BDB40002A5B3, 0x0002A1B60002A1B5, 0x000600E1000295B7,
-    0x000638E380006620, 0x8000662880006624, 0x8000662C0001A4E1,
-    0x000FA084000164E3, 0x8000663480006630, 0x00007D810001B8EF,
-    0x000148EF00007D80, 0x8000641400006982, 0x800066380002F9B3,
-    0x00025DB68000663C, 0x8000664880006640, 0x000249BA000255B5,
-    0x000CC886000249BB, 0x00023DBE000D5087, 0x000239BC000221BF,
-    0x000D3DA3000235BD, 0x80006644800065D4, 0x800066508000664C,
-    0x000D91A5000D99A6, 0x000D8DAB80006654, 0x80006658000D8DAA,
-    0x000DE5AF8000665C, 0x000DF1AD000DF9AE, 0x80006660000DFDAC,
-    0x8000666880006664, 0x000C70E38000666C, 0x000F70E180006670,
-    0x000C50EF80006674, 0x8000667C80006678, 0x8000668480006680,
-    0x8000668C80006688, 0x000E15A880006690, 0x000E59AA000E1DA9,
-    0x000E49AC000E59AB, 0x000E79AE000E49AD, 0x80006694000E79AF,
-    0x8000669C80006698, 0x800066A4800066A0, 0x0001B1810001B982,
-    0x000FF9B80001B980, 0x000F4DBA000FF9B9, 0x000F41BC000F41BB,
-    0x000F7DBE000F4DBD, 0x000E09B0000F75BF, 0x000E1DB2000E09B1,
-    0x000FD5B4000E11B3, 0x000FC9B6000FD9B5, 0x000FD1B3000FC9B7,
-    0x800066AC800066A8, 0x800066B4800066B0, 0x000FC1B5000FC9B6,
-    0x000FCDBB800066B8, 0x00008886000FCDBA, 0x000E1DBF0001AC87,
-    0x000FC5BD000E11BE, 0x800066BC000FC9BC, 0x800066C0000F95A3,
-    0x000F8DA6800066C4, 0x800066CC800066C8, 0x000F89AA000F85A5,
-    0x800066D0000F89AB, 0x000E55AE800066D4, 0x000F8DAC000E59AF,
-    0x000E59A8000F81AD, 0x000199AA000189A9, 0x0001B9AC000191AB,
-    0x0001D9AE0001B9AD, 0x0000FCBE0001D9AF, 0x800066DC800066D8,
-    0x800066E4800066E0, 0x800066EC800066E8, 0x0001B5B8800066F0,
-    0x0001B5BA0001BDB9, 0x000175BC00016DBB, 0x000175BE00017DBD,
-    0x0001A9B000016DBF, 0x0001BDB20001A9B1, 0x000195B40001B1B3,
-    0x000195B600019DB5, 0x000211B300018DB7, 0x800066F8800066F4,
-    0x80006700800066FC, 0x000231B5000239B6, 0x00026DBB80006704,
-    0x8000670800026DBA, 0x000259BF8000670C, 0x000265BD000251BE,
-    0x80006710000269BC, 0x80006714000255A3, 0x00027DA680006718,
-    0x800067208000671C, 0x000229AA000275A5, 0x80006728000229AB,
-    0x000215AE8000672C, 0x00022DAC00021DAF, 0x0000CD80000221AD,
-    0x0000D1820000D181, 0x8000673000040084, 0x800067340001A4E1,
-    0x00048086000F2CE3, 0x800067380004A487, 0x8000673C0001F4EF,
-    0x8000674480006740, 0x80006748000000AA, 0x800067508000674C,
-    0x80006754000385B3, 0x000381B6000381B5, 0x8000675C80006758,
-    0x00034DBA80006760, 0x00035DBC000345BB, 0x00034DBE000345BD,
-    0x00021DA8000345BF, 0x000241AA000245A9, 0x000245AC000241AB,
-    0x000271AE00024DAD, 0x0004ACBE000271AF, 0x8000676880006764,
-    0x0004B8848000676C, 0x8000677480006770, 0x000149B880006778,
-    0x000159BA000149B9, 0x000179BC000151BB, 0x000119BE000179BD,
-    0x0002C5B0000115BF, 0x0002C5B20002CDB1, 0x0002C5B40002DDB3,
-    0x0002C5B60002CDB5, 0x000F30E1000179B7, 0x000F68E3000EB4E3,
-    0x8000677C000E7CE1, 0x8000678480006780, 0x8000678C80006788,
-    0x8000679880006790, 0x800067A08000679C, 0x000FB8EF000E64EF,
-    0x800067A8800067A4, 0x00029DA500029DA6, 0x000299A3000740BE,
-    0x800067B0800067AC, 0x000259AF80006564, 0x000259AD000251AE,
-    0x000259AB000241AC, 0x0004D886000251AA, 0x00079D8000055087,
-    0x0007AD820007A581, 0x0005C8EF80006794, 0x800067B8800067B4,
-    0x0004BCE3800067BC, 0x000490E1800067C0, 0x800067C8800067C4,
-    0x800067D0800067CC, 0x0006D1B3800067D4, 0x800067DC800067D8,
-    0x800067E4800067E0, 0x0006F1B50006E5B6, 0x0006A5BB800067E8,
-    0x800067EC0006A5BA, 0x000691BF800067F0, 0x0006A9BD0006A9BE,
-    0x0006E9A80006B5BC, 0x000619AA0006E9A9, 0x000639AC000615AB,
-    0x00062DAE000625AD, 0x800067F4000625AF, 0x00001D8100001D82,
-    0x800067F800001D80, 0x80006800800067FC, 0x0007E9B880006804,
-    0x0007FDBA0007E9B9, 0x0007EDBC0007F5BB, 0x00074DBE000745BD,
-    0x000621B0000745BF, 0x000621B2000621B1, 0x00062DB400063DB3,
-    0x00061DB6000615B5, 0x000791A30007D9B7, 0x0001288680006808,
-    0x8000680C00014C87, 0x0007B1A50007A5A6, 0x0007E5AB80006810,
-    0x800068140007E5AA, 0x0007D1AF80006818, 0x0007E9AD0007E9AE,
-    0x8000681C0007F5AC, 0x800068200006F5B3, 0x000685B680006824,
-    0x8000682C80006828, 0x000179BA00069DB5, 0x80006830000179BB,
-    0x000141BE80006834, 0x00015DBC000141BF, 0x000629A8000151BD,
-    0x000639AA000629A9, 0x000651AC000631AB, 0x000641AE000641AD,
-    0x80006838000189AF, 0x800068408000683C, 0x8000684880006844,
-    0x800068508000684C, 0x000185B880006854, 0x000185BA00018DB9,
-    0x00018DBC00019DBB, 0x0001B5BE0001BDBD, 0x0001F9B000017DBF,
-    0x0001CDB20001F9B1, 0x0001C1B40001C1B3, 0x0001C1B60001C1B5,
-    0x0005B1A30001C1B7, 0x8000685C80006858, 0x8000686480006860,
-    0x0005D9A50005C1A6, 0x00023DAB80006868, 0x8000686C00023DAA,
-    0x000205AF80006870, 0x000215AD000205AE, 0x00005D80000219AC,
-    0x0000658200006D81, 0x00039CBE000351B3, 0x000319B6000379B5,
-    0x0002E08480006878, 0x00033DBA8000687C, 0x000319BC000335BB,
-    0x0003D9BE000319BD, 0x000385A80003D9BF, 0x000395AA000395A9,
-    0x0003BDAC0003A5AB, 0x0003D1AE0003D5AD, 0x800068800003D1AF,
-    0x8000688880006884, 0x00032087001C60BE, 0x8000688C001C9486,
-    0x000371B880006890, 0x000371BA000371B9, 0x0000D5BC000371BB,
-    0x0000D5BE0000DDBD, 0x0003B5B00000CDBF, 0x000381B20003BDB1,
-    0x000351B4000381B3, 0x000351B6000351B5, 0x0003A8EF000351B7,
-    0x8000689880006894, 0x00021C848000689C, 0x800068A4800068A0,
-    0x001C6C84800068A8, 0x800068B0800068AC, 0x000340E3800068B4,
-    0x000028E1800068B8, 0x00025DA3800068BC, 0x800068C4800068C0,
-    0x800068CC800068C8, 0x000275A5000215A6, 0x000239AB800068D0,
-    0x800068D4000231AA, 0x0002D5AF800068D8, 0x000215AD0002D5AE,
-    0x0006A0E3000215AC, 0x000668E1800068DC, 0x0000FD80001CC4BE,
-    0x0000058200001181, 0x800068E8800068E0, 0x800068F0800068EC,
-    0x000684EF80006724, 0x800068F8800068F4, 0x000161B3800068FC,
-    0x001C688680006900, 0x80006904001D6C87, 0x000171B500016DB6,
-    0x0000EDBB80006908, 0x8000690C0000EDBA, 0x0000D5BF80006910,
-    0x0000E1BD0000E9BE, 0x800068E40000E9BC, 0x80006914001C50EF,
-    0x8000691C80006918, 0x8000692480006920, 0x8000692C80006928,
-    0x8000693480006930, 0x8000693C80006938, 0x80006940001D34E3,
-    0x001EE1A3001C90E1, 0x8000694880006944, 0x800069508000694C,
-    0x001EF1A5001EEDA6, 0x001F6DAB00001C85, 0x000160BF001F6DAA,
-    0x001F55AF80006954, 0x001F61AD001F69AE, 0x001EDDA8001F69AC,
-    0x001E29AA001EE1A9, 0x001E39AC001E29AB, 0x001E25AE001E2DAD,
-    0x0000E980001F99AF, 0x0000F9820000E981, 0x0004008680006958,
-    0x8000695C00001487, 0x001FF5B880006960, 0x001F8DBA001FF9B9,
-    0x001F9DBC001F85BB, 0x001F85BE001F85BD, 0x001FE9B0001FB5BF,
-    0x001FFDB2001FE9B1, 0x001FEDB4001FF5B3, 0x001FD1B6001FD1B5,
-    0x001EA1A8001FD1B7, 0x001EA1AA001EA1A9, 0x001EA9AC001EB9AB,
-    0x001EDDAE001EA9AD, 0x80006964001ECDAF, 0x8000696C80006968,
-    0x8000697480006970, 0x8000697C80006978, 0x00015DB880006980,
-    0x00016DBA000165B9, 0x00017DBC000165BB, 0x000161BE000165BD,
-    0x001EBDB0000119BF, 0x001E85B2001E8DB1, 0x00017DB4000165B3,
-    0x00016DB6000165B5, 0x001E29B3000165B7, 0x8000698880006984,
-    0x800069908000698C, 0x001E21B5001E29B6, 0x001E49BB80006994,
-    0x80006998001E79BA, 0x0001A1BF8000699C, 0x0001B1BD0001B9BE,
-    0x800069A0001E51BC, 0x800069A4001E6DA3, 0x001E6DA6800069A8,
-    0x800069B0800069AC, 0x001E3DAA001E65A5, 0x800069B4001E0DAB,
-    0x0001FDAE800069B8, 0x001E15AC0001E5AF, 0x002A098F0001F5AD,
-    0x800069BC00004CE1, 0x800069C0000A58E3, 0x0004D0BE0004D884,
-    0x0005E8BE800069C8, 0x00038887800069CC, 0x800069D000040C86,
-    0x800069D8800069D4, 0x001DED9F007B4CEF, 0x001A419D001B559E,
-    0x00198D9B001A499C, 0x00165D990016459A, 0x00120D9700159D98,
-    0x00119D9500126196, 0x002E7593002E1D94, 0x002AD591002D8D92,
-    0x007629B0002A7190, 0x007AA9B2007801B1, 0x007C01B4007AB1B3,
-    0x800069DC007E91B5, 0x00034D80800069E0, 0x0000B9820000AD81,
-    0x0002B1A3800069E4, 0x001E05A100010DA2, 0x800069E8001EB1A0,
-    0x0006CDA20002A9A1, 0x000801A40006C5A3, 0x000AD1A6000AD1A5,
-    0x000EE1A8000BB9A7, 0x000FF9AA000EE1A9, 0x0072D1AC007001AB,
-    0x007401AE00722DAD, 0x002B09820076DDAF, 0x800069EC002B0983,
-    0x002F1586800069F0, 0x002BB584002F6587, 0x0010D58A002CE585,
-    0x800069F40013218B, 0x0017358E800069F8, 0x0013558C0017258F,
-    0x00189D920014FD8D, 0x00048484001B4193, 0x001CFD9600070884,
-    0x001B5194001F5997, 0x001FDD9A001A4D95, 0x800069FC0000DD9B,
-    0x80006A0480006A00, 0x0003799C80006A08, 0x80006A1080006A0C,
-    0x80006A14001E54E1, 0x80006A18001EBCE3, 0x80006A2080006A1C,
-    0x00000D8080006A24, 0x00001D8200001581, 0x80006A2C80006A28,
-    0x8000687480006A30, 0x00056C87001EC8EF, 0x80006A380004E486,
-    0x000330EF80006A3C, 0x80006A4480006A40, 0x0003E4E380006A48,
-    0x000140E180006A4C, 0x80006A5480006A50, 0x80006A5C80006A58,
-    0x0004BDB380006A60, 0x80006A34800069C4, 0x80006A6880006A64,
-    0x00053DB500053DB6, 0x000519BB80006A6C, 0x80006A70000511BA,
-    0x0005BDBF80006A74, 0x0005BDBD0005BDBE, 0x80006A78000501BC,
-    0x80006A7C0004F9A3, 0x000579A680006A80, 0x80006A8880006A84,
-    0x000555AA000579A5, 0x80006A8C00055DAB, 0x0005F9AE80006A90,
-    0x000545AC0005F9AF, 0x00069DA80005F9AD, 0x0006B9AA0006A1A9,
-    0x0006A9AC0006B9AB, 0x000605AE00060DAD, 0x80006A94000631AF,
-    0x0000118100001982, 0x0000ECBE00001980, 0x80006A9C80006A98,
-    0x0007B5B880006AA0, 0x000749BA0007B9B9, 0x000741BC000745BB,
-    0x000741BE000741BD, 0x000621B0000741BF, 0x0007E5B20007EDB1,
-    0x0007DDB40007C5B3, 0x000795B600079DB5, 0x000731B300078DB7,
-    0x0000A88680006AA8, 0x80006AAC00018887, 0x000721B500078DB6,
-    0x000629BB80006AB0, 0x80006AB40007E5BA, 0x000659BF80006AB8,
-    0x000611BD000619BE, 0x80006ABC000631BC, 0x80006AC0000775A3,
-    0x0007C9A680006AC4, 0x80006ACC80006AC8, 0x0007A1AA000765A5,
-    0x80006AD000066DAB, 0x00065DAE80006AD4, 0x000675AC00061DAF,
-    0x000649A8000655AD, 0x00065DAA000649A9, 0x00064DAC000655AB,
-    0x000675AE000671AD, 0x0000AC8400066DAF, 0x80006ADC80006AD8,
-    0x80006AE480006AE0, 0x80006AEC80006AE8, 0x000181B880006AF0,
-    0x000185BA00018DB9, 0x000189BC000199BB, 0x0001B5BE0001BDBD,
-    0x000615B000017DBF, 0x000615B200061DB1, 0x0001E1B40001E1B3,
-    0x0001D1B60001E9B5, 0x000725AC0001D1B7, 0x000725AE00072DAD,
-    0x0006BDA800073DAF, 0x0006A5AA0006ADA9, 0x000CACBE00073DAB,
-    0x80006AF880006AF4, 0x80006B0080006AFC, 0x80006B0880006B04,
-    0x00078DBC80006B0C, 0x000791BE000791BD, 0x0007F9B80007A5BF,
-    0x000799BA0007F9B9, 0x0007E9B4000795BB, 0x0007F9B60007E9B5,
-    0x000745B00007F1B7, 0x000745B200074DB1, 0x0000BD810007F9B3,
-    0x00055DB30000BD80, 0x00054DB500001982, 0x80006B1880006B10,
-    0x80006B1C000579B6, 0x000551BB80006B20, 0x0002BDBD000559BA,
-    0x0002A5BF000549BC, 0x80006B240002ADBE, 0x000515A5000521A6,
-    0x000505A3000C4C84, 0x000CF88680006B28, 0x0002FDAF000C9C87,
-    0x0002E5AD0002F5AE, 0x000509AB000511AC, 0x80006B2C000501AA,
-    0x000205A880006B30, 0x000215AA000215A9, 0x00023DAC000225AB,
-    0x000251AE000255AD, 0x000CACBE000251AF, 0x80006B3880006B34,
-    0x80006B4080006B3C, 0x80006B4880006B44, 0x0002F5B880006B4C,
-    0x0002F5BA0002FDB9, 0x000179BC000169BB, 0x000169BE000179BD,
-    0x000235B0000169BF, 0x000201B200023DB1, 0x0002D5B4000201B3,
-    0x0002D5B60002DDB5, 0x0007B4E10002CDB7, 0x0006D0E30006D8E3,
-    0x80006B50000650E1, 0x80006B5880006B54, 0x80006B6080006B5C,
-    0x80006B6880006B64, 0x000000EC80006B6C, 0x000760EF0001FCEF,
-    0x0002DDB380006B70, 0x000CE88680006B74, 0x80006B78000C8C87,
-    0x0002F5B5000295B6, 0x0002B9BB80006B7C, 0x80006B800002B1BA,
-    0x000259BF80006B88, 0x000295BD000251BE, 0x000D1DA3000295BC,
-    0x80006B8480006B14, 0x80006B9080006B8C, 0x000D35A5000D55A6,
-    0x000D79AB80006B94, 0x80006B98000D71AA, 0x000D99AF80006B9C,
-    0x000D55AD000D91AE, 0x000F9180000D55AC, 0x000FA182000F9981,
-    0x80006BA0000EE9B3, 0x000EA1B6000E81B5, 0x80006BA880006BA4,
-    0x000F41BA80006BAC, 0x000F41BC000F41BB, 0x000F41BE000F41BD,
-    0x000E19A8000F41BF, 0x000E2DAA000E25A9, 0x000E45AC000E4DAB,
-    0x000E65AE000E6DAD, 0x80006BB0000FB9AF, 0x80006BB880006BB4,
-    0x80006BC080006BBC, 0x80006BC880006BC4, 0x000F59B880006BCC,
-    0x000F69BA000F59B9, 0x000F79BC000F69BB, 0x000F65BE000F79BD,
-    0x000FC9B0000F19BF, 0x000FD9B2000FC9B1, 0x000FC9B4000FD9B3,
-    0x000F69B6000FC9B5, 0x000FA5A3000F69B7, 0x80006BD480006BD0,
-    0x80006BDC80006BD8, 0x000FCDA5000FEDA6, 0x000E0DAB80006BE0,
-    0x80006BE4000E0DAA, 0x000E0DAF80006BE8, 0x000E0DAD000E0DAE,
-    0x00006D80000E0DAC, 0x0000198200000981, 0x80006BF080006BEC,
-    0x00019484000E44E1, 0x0001A8BE000EECE3, 0x80006BF880006BF4,
-    0x0000D086000DB4EF, 0x80006BFC0001FC87, 0x000E6DA880006C00,
-    0x000199AA000189A9, 0x0001B9AC000191AB, 0x0001C9AE0001B9AD,
-    0x0000A0840001C1AF, 0x80006C0880006C04, 0x80006C1080006C0C,
-    0x80006C1880006C14, 0x000179B880006C1C, 0x0000CDBA000179B9,
-    0x0000DDBC0000C5BB, 0x0000C5BE0000C5BD, 0x000181B00000F5BF,
-    0x000149B2000181B1, 0x000159B4000149B3, 0x000149B6000159B5,
-    0x000211B3000149B7, 0x80006C2480006C20, 0x80006C2C80006C28,
-    0x000231B5000239B6, 0x000245BB80006C30, 0x80006C34000245BA,
-    0x00029DBF80006C38, 0x00029DBD00029DBE, 0x0075FC8500029DBC,
-    0x80006C3C000255A3, 0x00027DA680006C40, 0x80006C4880006C44,
-    0x000201AA000275A5, 0x80006C4C000201AB, 0x0002D9AE80006C50,
-    0x0002D9AC0002D9AF, 0x000019820002D9AD, 0x0000198080006C54,
-    0x80006C5800001981, 0x80006C6080006C5C, 0x000400BE80006C68,
-    0x80006C7080006C6C, 0x80006C7880006C74, 0x80006C8080006C7C,
-    0x0004E08680006C84, 0x80006C8800035487, 0x80006AA480006C8C,
-    0x000280EF80006C90, 0x000188E180006C94, 0x000158E380006C98,
-    0x80006CA080006C9C, 0x80006CA880006CA4, 0x000269A880006CAC,
-    0x000275AA00027DA9, 0x0002B9AC0002A9AB, 0x0002A9AE0002B9AD,
-    0x0005AC840002A9AF, 0x80006CB480006CB0, 0x80006CBC80006CB8,
-    0x80006CC480006CC0, 0x00016DB880006CC8, 0x000105BA00010DB9,
-    0x000109BC000119BB, 0x000135BE00013DBD, 0x0002D9B00001D9BF,
-    0x00016DB20002D9B1, 0x00017DB4000165B3, 0x000165B6000165B5,
-    0x000F40E1000155B7, 0x000F54E30000CCE3, 0x0000AD80000E5CE1,
-    0x0000398200000981, 0x80006CD080006CCC, 0x80006CD8000428BE,
-    0x00073CBE80006CDC, 0x000E60EF00002CEF, 0x0004E08680006CE0,
-    0x80006CE400057487, 0x80006CE80002B1B3, 0x000275B600027DB5,
-    0x80006CF080006CEC, 0x000251BA80006CF4, 0x0001F5BC000251BB,
-    0x0001F5BE0001FDBD, 0x80006C640001EDBF, 0x0005B1AA80006CD4,
-    0x000615AC0005B1AB, 0x000615AE00061DAD, 0x80006CF800060DAF,
-    0x80006D0080006CFC, 0x80006D04000551A3, 0x000595A600059DA5,
-    0x0006E9B380006D08, 0x80006D1080006D0C, 0x80006D1880006D14,
-    0x0006E9B50006E1B6, 0x0006ADBB80006D1C, 0x80006D200006ADBA,
-    0x00069DBF80006D24, 0x0006A5BD000691BE, 0x00061DA80006A9BC,
-    0x000621AA000621A9, 0x000621AC000621AB, 0x000621AE000621AD,
-    0x80006D28000621AF, 0x80006D3080006D2C, 0x00000D8080006D34,
-    0x0001B1820001B181, 0x000639B880006D38, 0x0007EDBA000639B9,
-    0x0007FDBC0007E5BB, 0x0007EDBE0007E5BD, 0x000629B00007E5BF,
-    0x00063DB2000629B1, 0x000615B4000631B3, 0x000609B6000619B5,
-    0x0007A9A3000609B7, 0x80006D4080006D3C, 0x80006D4400028484,
-    0x0007A9A50007A1A6, 0x0007EDAB80006D48, 0x000088860007EDAA,
-    0x0007DDAF00017887, 0x0007E5AD0007D1AE, 0x80006D4C0007E9AC,
-    0x80006D500006F9B3, 0x00068DB680006D54, 0x80006D5C80006D58,
-    0x000141BA00068DB5, 0x80006D60000141BB, 0x00014DBE80006D64,
-    0x000151BC000131BF, 0x000629A800014DBD, 0x000635AA000629A9,
-    0x000641AC000645AB, 0x000679AE000649AD, 0x0001AC84000679AF,
-    0x80006D6C80006D68, 0x80006D7480006D70, 0x80006D7C80006D78,
-    0x000191B880006D80, 0x0001A1BA000199B9, 0x000171BC0001A1BB,
-    0x000171BE000171BD, 0x0001C1B0000171BF, 0x0001C1B20001C1B1,
-    0x0001CDB40001DDB3, 0x0001B1B60001B1B5, 0x0005BDA30001B1B7,
-    0x80006D8880006D84, 0x80006D9080006D8C, 0x0005C9A50005C9A6,
-    0x000205AB80006D94, 0x80006D98000205AA, 0x000275AF80006D9C,
-    0x000209AD000209AE, 0x80006DA0000215AC, 0x80006DA880006DA4,
-    0x00001D8180006DAC, 0x80006DB000001D80, 0x000324BE00000582,
-    0x007CC88680006DB8, 0x80006DBC00035C87, 0x80006DC480006DC0,
-    0x80006DCC80006DC8, 0x80006DD480006DD0, 0x80006DD8007CC084,
-    0x80006DDC0001A4E1, 0x007CE8BE00012CE3, 0x80006DE480006DE0,
-    0x000000AE000234EF, 0x80006DEC80006DE8, 0x000389B380006DF0,
-    0x80006DF880006DF4, 0x80006E0080006DFC, 0x00038DB500038DB6,
-    0x000349BB80006E04, 0x80006E08000341BA, 0x000349BF80006E0C,
-    0x000349BD000341BE, 0x00021DA8000351BC, 0x000241AA000245A9,
-    0x000245AC000241AB, 0x000271AE00024DAD, 0x007FA0BE000271AF,
-    0x80006E1480006E10, 0x00001D8080006E18, 0x0000A98200000981,
-    0x000149B880006E1C, 0x000159BA000149B9, 0x000179BC000151BB,
-    0x000119BE000179BD, 0x0002C5B0000119BF, 0x0002C5B20002CDB1,
-    0x0002C5B40002DDB3, 0x0002C5B60002CDB5, 0x000658E1000179B7,
-    0x000094E300075CE1, 0x80006E200006BCE3, 0x007CB88680006E28,
-    0x80006E2C007D5487, 0x80006E3480006E30, 0x000000EE80006E38,
-    0x0006F0EF80006E3C, 0x80006E400006E0EF, 0x000281A380006E44,
-    0x000285A580006E48, 0x80006E5080006E4C, 0x80006E54000285A6,
-    0x000241AB80006E58, 0x000241AD000249AA, 0x000241AF000259AC,
-    0x007E65A8000249AE, 0x007E7DAA007E75A9, 0x007E6DAC007E75AB,
-    0x007E9DAE007E95AD, 0x80006E24007E95AF, 0x80006E6080006E5C,
-    0x80006E6880006E64, 0x80006E7080006E6C, 0x007E89B880006E74,
-    0x007E99BA007E89B9, 0x007EB9BC007E91BB, 0x007F7DBE007EB9BD,
-    0x007EF1B0007F75BF, 0x007ECDB2007EFDB1, 0x007EC5B4007EC5B3,
-    0x007EC5B6007ECDB5, 0x007EF5B3007EB9B7, 0x80006E7C80006E78,
-    0x80006E8480006E80, 0x007EE5B5007E15B6, 0x007E39BB80006E88,
-    0x80006E8C007E35BA, 0x007E09BF80006E90, 0x007E11BD007E01BE,
-    0x00006D82007E19BC, 0x00005580007EB1A3, 0x007E51A600006581,
-    0x80006E9400031884, 0x007E71AA007EA1A5, 0x00008886007E7DAB,
-    0x007E45AE00035C87, 0x007E5DAC007E4DAF, 0x007EA5B3007E55AD,
-    0x80006E9C80006E98, 0x80006EA480006EA0, 0x007EB5B5007EA1B6,
-    0x007EF5BB80006EA8, 0x80006EAC007EE9BA, 0x000115BF80006EB0,
-    0x007ED9BD007ED1BE, 0x80006EB4007EE5BC, 0x80006EBC80006EB8,
-    0x007D08EF80006EC0, 0x80006EC880006EC4, 0x007DFCE380006ECC,
-    0x007D18E180006ED0, 0x80006ED880006ED4, 0x80006EE080006EDC,
-    0x007E15A880006EE4, 0x007E2DAA007E25A9, 0x007E3DAC007E25AB,
-    0x007E59AE007E21AD, 0x80006EE8007E59AF, 0x80006EF080006EEC,
-    0x00000D8080006EF4, 0x0003BD820003B581, 0x0001F1B880006EF8,
-    0x0001F1BA0001F1B9, 0x000155BC0001F1BB, 0x000155BE00015DBD,
-    0x007E2DB000014DBF, 0x007E31B2007E31B1, 0x007E15B4007E0DB3,
-    0x0001D1B60001D1B5, 0x80006EFC0001D1B7, 0x0002908480006F00,
-    0x00039CBE007DA9A3, 0x007DADA6007DB9A5, 0x80006F0C80006F08,
-    0x007DE5AA80006F10, 0x007DE9AC007DF9AB, 0x007DDDAE007DD5AD,
-    0x000369B6000219AF, 0x0003488480006F14, 0x80006F18000351B5,
-    0x0004F886000359B3, 0x000339BE00033487, 0x000325BC000339BF,
-    0x000335BA000339BD, 0x80006F1C00033DBB, 0x0003F1B880006F20,
-    0x0003F1BA0003F1B9, 0x000331BC0003F1BB, 0x000331BE000331BD,
-    0x0003E5B0000331BF, 0x0003E5B20003EDB1, 0x0003E9B40003F9B3,
-    0x0003DDB60003E9B5, 0x0003C5A80003D1B7, 0x0003C5AA0003CDA9,
-    0x0003CDAC0003DDAB, 0x0003F5AE0003FDAD, 0x80006F2400039DAF,
-    0x80006F2C80006F28, 0x0004FCBE80006DB4, 0x80006F3480006F30,
-    0x00020DA380006F38, 0x80006F4080006F3C, 0x80006F4880006F44,
-    0x000205A500023DA6, 0x000269AB80006F4C, 0x80006F50000261AA,
-    0x00026DAF80006F54, 0x00026DAD00026DAE, 0x80006F58000271AC,
-    0x80006F6080006F5C, 0x80006F6880006F64, 0x80006F6C00034CE3,
-    0x000011820001ACE1, 0x0000398000038CEF, 0x80006F7000003581,
-    0x0004588680006F78, 0x80006F7C00054087, 0x80006F8480006F80,
-    0x80006F8C80006F88, 0x80006F9480006F90, 0x80006F98007E44EF,
-    0x80006F9C007E24E1, 0x000000EC000168E3, 0x80006FA480006FA0,
-    0x80006FAC80006FA8, 0x80006FB480006FB0, 0x80006FB80001E9B3,
-    0x0001A1B6000185B5, 0x80006FC080006FBC, 0x00010DBA80006FC4,
-    0x000105BC000105BB, 0x000105BE00010DBD, 0x00061DA80000F9BF,
-    0x00062DAA000625A9, 0x000641AC000649AB, 0x000665AE00066DAD,
-    0x80006F740006FDAF, 0x80006FCC80006FC8, 0x80006FD480006FD0,
-    0x0000B1810000B982, 0x000751B80000B980, 0x000761BA000759B9,
-    0x000711BC000761BB, 0x000711BE000711BD, 0x000685B0000711BF,
-    0x00069DB200068DB1, 0x00068DB4000695B3, 0x000771B6000771B5,
-    0x0006A5A3000771B7, 0x80006FDC80006FD8, 0x0002988480006FE0,
-    0x0006C9A50006EDA6, 0x000649AB000310BE, 0x80006FE4000641AA,
-    0x0007B5AF80006FE8, 0x000641AD000649AE, 0x80006FEC000649AC,
-    0x00006886000741B3, 0x000741B600036C87, 0x80006FF480006FF0,
-    0x000741BA000751B5, 0x80006FF800074DBB, 0x00074DBE80006FFC,
-    0x00074DBC000735BF, 0x000685A3000745BD, 0x8000700480007000,
-    0x8000700C80007008, 0x000695A5000685A6, 0x000689AB80007010,
-    0x80007014000685AA, 0x0006F1AF80007018, 0x000681AD000689AE,
-    0x8000701C000689AC, 0x8000702000054CEF, 0x8000702880007024,
-    0x800070308000702C, 0x8000703880007034, 0x800070408000703C,
-    0x8000704880007044, 0x8000704C00052CE3, 0x0006F5A800053CE1,
-    0x00060DAA000605A9, 0x00061DAC000605AB, 0x00060DAE000605AD,
-    0x0000D180000605AF, 0x000005820000DD81, 0x0001ECBE80007050,
-    0x800070580001F884, 0x000611B88000705C, 0x000629BA000619B9,
-    0x0001F5BC000629BB, 0x0001F5BE0001FDBD, 0x000641B00001EDBF,
-    0x00065DB200064DB1, 0x00064DB4000655B3, 0x00063DB6000635B5,
-    0x0002ADA8000635B7, 0x0002DDAA0002C5A9, 0x0002F9AC0002D5AB,
-    0x00031DAE0002F9AD, 0x000C8086000315AF, 0x8000706000033C87,
-    0x8000706880007064, 0x800070708000706C, 0x000311B8000378BE,
-    0x000311BA000311B9, 0x000331BC000311BB, 0x000331BE000331BD,
-    0x00036DB0000331BF, 0x000375B2000375B1, 0x00035DB4000345B3,
-    0x000331B6000335B5, 0x000255B3000331B7, 0x8000707880007074,
-    0x800070808000707C, 0x000271B5000279B6, 0x0003B1BB80007084,
-    0x800070880003A9BA, 0x0003F1BF8000708C, 0x000391BD000399BE,
-    0x800070900003A1BC, 0x000221A5000229A6, 0x000205A380007094,
-    0x8000709C80007098, 0x0003A1AF800070A0, 0x0003C1AD0003C9AE,
-    0x0003E1AB0003F1AC, 0x80006F040003F9AA, 0x800070A8800070A4,
-    0x800070B0800070AC, 0x00002582800070B4, 0x0000318000005CEF,
-    0x0002AC8400002D81, 0x800070B8000080E3, 0x800070BC0001FCE1,
-    0x000D1886800070C0, 0x000169B3000D1087, 0x000C44BE800070C8,
-    0x800070D0800070CC, 0x000169B5000161B6, 0x0001EDBB800070D4,
-    0x800070D80001EDBA, 0x0001DDBF800070DC, 0x0001E5BD0001DDBE,
-    0x800070E00001EDBC, 0x800070E8800070E4, 0x800070EC000C0084,
-    0x800070F4800070F0, 0x000BEC850001F8EF, 0x800070F80006C4E1,
-    0x800070FC00012CE3, 0x8000710480007100, 0x8000710C80007108,
-    0x0001E1A380007110, 0x0001E1A580007114, 0x8000711C80007118,
-    0x800071200001E9A6, 0x000165AB80007124, 0x00016DAD000165AA,
-    0x000155AF000165AC, 0x000DB9A8000155AE, 0x000E21AA000DB9A9,
-    0x000E21AC000E21AB, 0x000E21AE000E21AD, 0x800070C4000E99AF,
-    0x00001D8100001D82, 0x8000712800001D80, 0x800071308000712C,
-    0x000E95B880007134, 0x000E95BA000E9DB9, 0x000F51BC000EB1BB,
-    0x000F51BE000F51BD, 0x000EE9B0000F51BF, 0x000EF9B2000EE9B1,
-    0x000ED9B4000EF1B3, 0x000EB9B6000ED9B5, 0x000FC5A8000EB1B7,
-    0x000FD9AA000FC9A9, 0x000FF9AC000FD1AB, 0x000F19AE000FF9AD,
-    0x00000086000F19AF, 0x8000713800036487, 0x800071408000713C,
-    0x8000714880007144, 0x000F7DB88000714C, 0x000F0DBA000F05B9,
-    0x000F05BC000F01BB, 0x000F39BE000F0DBD, 0x000F69B0000F39BF,
-    0x000F7DB2000F75B1, 0x000F59B4000F75B3, 0x000F4DB6000F45B5,
-    0x000D38E1000F45B7, 0x000C38E380007150, 0x8000715880007154,
-    0x800071608000715C, 0x8000716880007164, 0x800071708000716C,
-    0x8000717880007174, 0x000DFCEF8000717C, 0x000E11B380007180,
-    0x8000718880007184, 0x800071908000718C, 0x000E31B5000E39B6,
-    0x000EF1BB80007194, 0x80007198000EEDBA, 0x000EC1BF8000719C,
-    0x000EE1BD000EDDBE, 0x000ED1A3000EE9BC, 0x00028884800071A0,
-    0x800071A8000180BE, 0x000EF1A5000EF9A6, 0x000E31AB800071AC,
-    0x00000886000E2DAA, 0x000E01AF00014C87, 0x000E21AD000E1DAE,
-    0x0000A980000E29AC, 0x00004D820000A981, 0x800071B000017DB3,
-    0x000165B600016DB5, 0x800071B8800071B4, 0x0000F9BA800071BC,
-    0x0000E9BC0000F9BB, 0x0000D9BE0000E9BD, 0x0002A9A80000D9BF,
-    0x0002B9AA0002A9A9, 0x0002A9AC0002B9AB, 0x00031DAE0002A9AD,
-    0x0004ACBE000311AF, 0x800071C4800071C0, 0x800071C80005E0BE,
-    0x800071D0800071CC, 0x000371B8800071D4, 0x000375BA000371B9,
-    0x000319BC000309BB, 0x000309BE000319BD, 0x000371B0000301BF,
-    0x000371B2000371B1, 0x000351B4000371B3, 0x000351B6000351B5,
-    0x000231A3000351B7, 0x800071D880007054, 0x800071E0800071DC,
-    0x000221A5000229A6, 0x0003B5AB800071E4, 0x800071E80003B5AA,
-    0x000395AF800071EC, 0x0003A5AD000395AE, 0x800071F00003A5AC,
-    0x800071F8800071F4, 0x80007200800071FC, 0x8000720880007204,
-    0x000039808000720C, 0x0000058200003981, 0x8000721880007210,
-    0x8000721C0003FC84, 0x000291A880007220, 0x0002D9AA000299A9,
-    0x0002F1AC0002D1AB, 0x00013DAE0002F1AD, 0x00050086000135AF,
-    0x8000722400056487, 0x8000722C80007228, 0x8000723480007230,
-    0x0001E9B880007238, 0x000189BA0001E9B9, 0x000199BC000189BB,
-    0x000189BE000199BD, 0x000151B0000189BF, 0x000151B2000151B1,
-    0x0001F9B4000151B3, 0x0001E9B60001F9B5, 0x0000F4EF0001E1B7,
-    0x800072408000723C, 0x8000724400078CBF, 0x0005188480007248,
-    0x8000724C000E2CEF, 0x80007250000E84E1, 0x000028E3000E30E3,
-    0x000100E180007254, 0x0020FC8580007258, 0x8000725C000215B3,
-    0x000235B680007260, 0x8000726880007264, 0x0001D1BA00023DB5,
-    0x8000726C0001D1BB, 0x000179BE80007270, 0x0001C1BC000179BF,
-    0x0005D1A30001C1BD, 0x8000727480007214, 0x8000727C80007278,
-    0x0005F9A50005F1A6, 0x000615AB80007280, 0x80007284000615AA,
-    0x0006BDAF80007288, 0x000605AD0006BDAE, 0x00004D80000605AC,
-    0x0000558200005581, 0x8000728C0006B1B3, 0x000769B6000779B5,
-    0x0001C08680007290, 0x000755BA0003E887, 0x0007F9BC0007E9BB,
-    0x0007EDBE0007E5BD, 0x000671A80007DDBF, 0x000675AA00067DA9,
-    0x0007BDAC0007A5AB, 0x0007ADAE0007A5AD, 0x800072940007A5AF,
-    0x8000729C80007298, 0x800072A4800072A0, 0x800072AC800072A8,
-    0x00077DB8800072B0, 0x000705BA00070DB9, 0x000709BC000719BB,
-    0x000735BE00073DBD, 0x0007DDB00000D5BF, 0x000761B2000761B1,
-    0x000765B400077DB3, 0x000751B6000755B5, 0x800072B4000749B7,
-    0x800072BC800072B8, 0x0008A4E1800072C0, 0x0004CCE30006A4E1,
-    0x800072C40005F8E3, 0x00187CEF800072C8, 0x000135A30005F4EF,
-    0x000505A10003A5A2, 0x0007BDA300051DA0, 0x800072D0800072CC,
-    0x800072D8800072D4, 0x000675A5000665A6, 0x0006E5AB800072DC,
-    0x800072E0000659AA, 0x0006D1AF800072E4, 0x0006E9AD0006E1AE,
-    0x007DED8F0006F5AC, 0x00001D8100001D82, 0x800072E800001D80,
-    0x800072F0800072EC, 0x0000BC84001C00BE, 0x00009887800072F8,
-    0x800072FC001F2C86, 0x8000730480007300, 0x0007459F80007308,
-    0x0009E99D00096D9E, 0x000D099B000B359C, 0x000F3599000DF59A,
-    0x0071E5970070CD98, 0x00748D950071E596, 0x0075F5930075E594,
-    0x0079F19100772592, 0x8000730C0079F990, 0x007E658280007310,
-    0x007A1984007A0983, 0x00761D860074C585, 0x001C6C8400761587,
-    0x00774D8A80007314, 0x0072358C00722D8B, 0x000CD58E0073258D,
-    0x00250485000E1D8F, 0x000E919280007318, 0x000A71940008B593,
-    0x0004DD96000AFD95, 0x8000731C00067D97, 0x0006559A80007320,
-    0x0002759C0000FD9B, 0x001C30BE80007324, 0x002AA1B380007328,
-    0x002FA9B1002801B2, 0x800071A4002E85B0, 0x002401B5001DD884,
-    0x8000732C002BA9B4, 0x8000733480007330, 0x8000733C80007338,
-    0x8000734480007340, 0x001EC5A380007348, 0x0002E1A1001EC5A2,
-    0x001A01A78000734C, 0x001801A5001AF5A6, 0x001001AB001E05A4,
-    0x0016C9A90016D5AA, 0x002EF5AF001401A8, 0x0012A5AD00138DAE,
-    0x00032DA80012F9AC, 0x000355AA000355A9, 0x0003B5AC0003ADAB,
-    0x0003B5AE0003BDAD, 0x800073500003ADAF, 0x8000735880007354,
-    0x000019808000735C, 0x0000058200001981, 0x000265B880007360,
-    0x000201BA000201B9, 0x000201BC000201BB, 0x000231BE000209BD,
-    0x0003D5B0000231BF, 0x0003D5B20003DDB1, 0x000275B400026DB3,
-    0x000275B600027DB5, 0x0006B8E100025DB7, 0x000600E380007368,
-    0x8000736C001CC4BE, 0x001C38860000F0EF, 0x80007370001CBC87,
-    0x80007374000014E3, 0x001F10860001D4E1, 0x00076CEF001F8487,
-    0x80007378001FF084, 0x00036DB500036DB6, 0x00038DB3001F84BE,
-    0x800073808000737C, 0x0002E9BF80007384, 0x0002E9BD0002E1BE,
-    0x000349BB000351BC, 0x80007388000341BA, 0x800072F48000738C,
-    0x8000739080007364, 0x8000739880007394, 0x800073A08000739C,
-    0x001D85A8800073A4, 0x001D9DAA001D95A9, 0x001D8DAC001D95AB,
-    0x001E3DAE001E35AD, 0x001E05B0001E19AF, 0x001E2DB2001E25B1,
-    0x001E2DB4001E39B3, 0x001EFDB6001EF5B5, 0x001ECDB8001EF5B7,
-    0x001EDDBA001ED5B9, 0x001ECDBC001ED5BB, 0x001F59BE001EB9BD,
-    0x001D35B3001F59BF, 0x800073AC800073A8, 0x800073B4800073B0,
-    0x001D25B5001E95B6, 0x001EC1BB800073B8, 0x800073BC001EE1BA,
-    0x001E11BF800073C0, 0x001ECDBD001EC9BE, 0x800073C4001ED9BC,
-    0x800073C8001D71A3, 0x001ED1A6800073CC, 0x800073D4800073D0,
-    0x001EA5AA001D61A5, 0x800073D8001E85AB, 0x001E8DAE800073DC,
-    0x001E9DAC001E55AF, 0x001E85A8001E89AD, 0x001EA9AA001E95A9,
-    0x001ED1AC001EA5AB, 0x001EC1AE001EC1AD, 0x0001E980001EC1AF,
-    0x0001FD820001E981, 0x800073E4800073E0, 0x800073EC800073E8,
-    0x0001C5B8800073F0, 0x0001C5BA0001CDB9, 0x0001CDBC0001DDBB,
-    0x0001F5BE0001FDBD, 0x001E99B000019DBF, 0x000161B2001E99B1,
-    0x000161B4000161B3, 0x000161B6000161B5, 0x001E21B3000161B7,
-    0x800073F8800073F4, 0x80007400800073FC, 0x001E59B5001E41B6,
-    0x001E5DBB00009084, 0x80007404001E5DBA, 0x0001A5BF80007408,
-    0x0001B5BD0001A5BE, 0x8000740C0001B5BC, 0x80007410001E65A3,
-    0x001E05A680007414, 0x8000741C80007418, 0x001E19AA001E1DA5,
-    0x80007420001E19AB, 0x0001E1AE80007424, 0x0001F1AC0001E1AF,
-    0x0002E9A80001F1AD, 0x000325AA0002E9A9, 0x000325AC00033DAB,
-    0x000325AE00032DAD, 0x80007428000399AF, 0x800074308000742C,
-    0x00000D8080007434, 0x0000B1820000B181, 0x000395B880007438,
-    0x0003A1BA00039DB9, 0x000071BC0003A1BB, 0x000071BE000071BD,
-    0x0003E9B0000071BF, 0x0003F9B20003E9B1, 0x0003DDB40003C5B3,
-    0x0003B1B60003B5B5, 0x0003ADB30003B1B7, 0x800074408000743C,
-    0x8000744C80007448, 0x0003ADB50003ADB6, 0x00032DBB80007450,
-    0x0004088600032DBA, 0x000371BF0003DC87, 0x000311BD000319BE,
-    0x8000745400032DBC, 0x800074580003E9A3, 0x0003E9A68000745C,
-    0x8000746480007460, 0x000369AA0003E9A5, 0x80007468000369AB,
-    0x00035DAE8000746C, 0x000369AC000335AF, 0x80007470000355AD,
-    0x001FADA5001FADA6, 0x001F4DA380007474, 0x8000747C80007478,
-    0x001E29AF80007480, 0x001E29AD001E21AE, 0x001F89AB001F91AC,
-    0x80007484001F81AA, 0x8000748C80007488, 0x000544BE80007490,
-    0x8000749400048084, 0x80007498001EF8E3, 0x8000749C001EF4E1,
-    0x800074A0001E50EF, 0x00000D80800074A4, 0x0000ED820000E581,
-    0x001E00E1800074A8, 0x001F0CE3001C30E1, 0x800074AC001FBCE3,
-    0x0004B886800074B0, 0x000720BE00055C87, 0x800074BC800074B8,
-    0x800074C4800074C0, 0x001ECCEF800074C8, 0x00015DB3001F3CEF,
-    0x800074D0800074CC, 0x800074D8800074D4, 0x000111B5000119B6,
-    0x00010DBB800074DC, 0x800074E000010DBA, 0x0000E9BF800074E4,
-    0x0000F5BD0000E1BE, 0x00068DA80000FDBC, 0x000695AA000695A9,
-    0x0006D9AC0006CDAB, 0x0006CDAE0006D9AD, 0x800074B40006C5AF,
-    0x800074EC800074E8, 0x800074F4800074F0, 0x800074FC800074F8,
-    0x00077DB880007500, 0x00070DBA000705B9, 0x000709BC00071DBB,
-    0x000731BE000735BD, 0x000689B0000731BF, 0x000769B2000689B1,
-    0x000765B4000761B3, 0x000765B600076DB5, 0x00061DA3000745B7,
-    0x8000750880007504, 0x800075108000750C, 0x000651A5000659A6,
-    0x00064DAB80007514, 0x8000751800064DAA, 0x0007A9AF8000751C,
-    0x0007B5AD0007A1AE, 0x000069800007BDAC, 0x0000058200006981,
-    0x8000752480007520, 0x0003A88780007528, 0x8000752C00009C86,
-    0x8000753480007530, 0x8000753C80007538, 0x8000754480007540,
-    0x0006ADB380007548, 0x800075508000754C, 0x8000755880007554,
-    0x0006ADB50006D1B6, 0x0006FDBB8000755C, 0x800075600006F5BA,
-    0x000119BF80007564, 0x0006D5BD000111BE, 0x800075680006D5BC,
-    0x800075708000756C, 0x8000757880007574, 0x8000757C0006D8E3,
-    0x800075800005ECE1, 0x800075840004D4EF, 0x00001D8080007588,
-    0x0000D98200000981, 0x000625A88000758C, 0x000625AA00062DA9,
-    0x000625AC00063DAB, 0x000651AE000651AD, 0x80007590000651AF,
-    0x000024BE00013C84, 0x000F108680007598, 0x8000759C0000C487,
-    0x0001F5B8800075A0, 0x0001F5BA0001FDB9, 0x000159BC000149BB,
-    0x000149BE000159BD, 0x000631B0000141BF, 0x000601B2000639B1,
-    0x0001D5B4000601B3, 0x0001D5B60001DDB5, 0x800075A40001CDB7,
-    0x800075AC800075A8, 0x800075B00005A1A3, 0x0005DDA60005A1A5,
-    0x800075B8800075B4, 0x0005F9AA800075BC, 0x0005D9AC0005F1AB,
-    0x00021DAE0005D9AD, 0x0002B1B6000215AF, 0x800075C4800075C0,
-    0x800075C80002A1B5, 0x800075CC0002B1B3, 0x000371BE800075D0,
-    0x000361BC000379BF, 0x00028DBA000361BD, 0x800075D4000371BB,
-    0x00027DB0800075D8, 0x000381B2000381B1, 0x000381B4000381B3,
-    0x0003B1B6000389B5, 0x0003E1B80003B1B7, 0x0003E1BA0003E1B9,
-    0x0003E1BC0003E1BB, 0x0003D9BE0003EDBD, 0x000DE0860003D5BF,
-    0x800075DC000CE887, 0x800075E4800075E0, 0x80007444800075E8,
-    0x000235A8800075EC, 0x00020DAA000205A9, 0x00021DAC000205AB,
-    0x00020DAE000205AD, 0x0003E5A3000205AF, 0x800075F4800075F0,
-    0x800075FC800075F8, 0x0003F5A50003E5A6, 0x000225AB80007600,
-    0x000228840003D9AA, 0x00022DAF000D20BE, 0x000235AD000225AE,
-    0x80007608000235AC, 0x800076108000760C, 0x000CDCBE80007614,
-    0x80007618000190E1, 0x8000761C0000E8E3, 0x8000762480007620,
-    0x00007D8000003CEF, 0x0000658200006581, 0x8000762C80007628,
-    0x8000763480007630, 0x8000763C80007638, 0x8000764480007640,
-    0x800076480006C4EF, 0x003FA0BF000624E1, 0x8000764C0006F0E3,
-    0x80007650000DBC84, 0x8000765880007654, 0x800076608000765C,
-    0x80007664000271B3, 0x0001D1B60001D1B5, 0x8000766C80007668,
-    0x0001C9BA80007670, 0x0001B5BC0001C9BB, 0x0001A9BE0001BDBD,
-    0x000DADA80001A5BF, 0x000DB5AA000DBDA9, 0x000E15AC000E0DAB,
-    0x000E09AE000E1DAD, 0x80007604000E01AF, 0x000FC981000FC182,
-    0x80007674000FF180, 0x0000988680007678, 0x000E3DB800036087,
-    0x000E85BA000E8DB9, 0x000E85BC000E9DBB, 0x000F5DBE000E89BD,
-    0x000E41B0000F55BF, 0x000E2DB2000E41B1, 0x000E21B4000E25B3,
-    0x000E25B6000E2DB5, 0x000D3DA3000E05B7, 0x800076808000767C,
-    0x8000768880007684, 0x000E9DA5000E9DA6, 0x000E85AB8000768C,
-    0x80007690000E85AA, 0x000EE9AF80007694, 0x000EF1AD000EE5AE,
-    0x80007698000EF9AC, 0x8000769C000F8DB3, 0x000FB1B6800076A0,
-    0x800076A8800076A4, 0x000F9DBA000F85B5, 0x800076AC000F65BB,
-    0x000F7DBE800076B0, 0x000F7DBC000F65BF, 0x000E49A3000F75BD,
-    0x800076B8800076B4, 0x800076C0800076BC, 0x000E41A5000E75A6,
-    0x000EA1AB800076C4, 0x800076C8000E59AA, 0x000EA1AF800076CC,
-    0x000EB1AD000EB9AE, 0x800076D0000EB9AC, 0x800076D8800076D4,
-    0x000C84EF800076DC, 0x0000158100003182, 0x000C58E300001580,
-    0x000EC0E1800076E0, 0x000138BE00012084, 0x0000D886800076E8,
-    0x000EC1A800005C87, 0x000EC1AA000EC1A9, 0x000EC9AC000EC1AB,
-    0x000EFDAE000EF5AD, 0x0000A084000EF5AF, 0x800076F0800076EC,
-    0x800076F8800076F4, 0x80007700800076FC, 0x000E61B880007704,
-    0x000E71BA000E75B9, 0x000E11BC000E71BB, 0x000E11BE000E11BD,
-    0x000E9DB0000E0DBF, 0x000E6DB2000E65B1, 0x000E7DB4000E65B3,
-    0x000E6DB6000E65B5, 0x000235A8000E65B7, 0x000281AA000281A9,
-    0x000281AC000281AB, 0x0002B1AE000289AD, 0x01DD28BF0002B1AF,
-    0x8000770C80007708, 0x800077100005E084, 0x8000771880007714,
-    0x000349B88000771C, 0x000359BA000349B9, 0x000379BC000351BB,
-    0x000319BE000379BD, 0x0002A1B0000319BF, 0x0002A5B20002ADB1,
-    0x0002A9B40002B9B3, 0x000295B600029DB5, 0x000209B3000379B7,
-    0x8000772480007720, 0x8000772C80007728, 0x000209B5000209B6,
-    0x000269BB80007730, 0x80007734000265BA, 0x000259BF80007738,
-    0x00026DBD00026DBE, 0x8000773C000271BC, 0x8000774000024DA3,
-    0x00024DA680007744, 0x8000774C80007748, 0x000221AA00024DA5,
-    0x8000775000022DAB, 0x000229AE80007594, 0x000235AC00021DAF,
-    0x000275A8000229AD, 0x0002A1AA0002A1A9, 0x0002A1AC0002A1AB,
-    0x0002A1AE0002A1AD, 0x0001F5800002A1AF, 0x0001C1820001FD81,
-    0x800077580005A0BE, 0x000574878000775C, 0x0001CDB800051C86,
-    0x0001D5BA0001D5B9, 0x0001FDBC0001E5BB, 0x000191BE000195BD,
-    0x000145B0000191BF, 0x000159B200014DB1, 0x000149B4000159B3,
-    0x0001FDB6000149B5, 0x800077600001F5B7, 0x000484BE80007764,
-    0x80007768000004EF, 0x800077708000776C, 0x80007774000E34EF,
-    0x80007778000F7CE1, 0x000000EC000E08E3, 0x8000777C000194E1,
-    0x800077800000C0E3, 0x8000778880007784, 0x8000778C000265B3,
-    0x00026DB6000265B5, 0x8000779480007790, 0x0001CDBA80007798,
-    0x0001C1BC0001C5BB, 0x0001C1BE0001C1BD, 0x0005A9A30001C1BF,
-    0x8000779C80007754, 0x800077A4800077A0, 0x0005A9A50005A1A6,
-    0x000609AB800077A8, 0x800077AC000601AA, 0x00060DAF800077B0,
-    0x00060DAD00060DAE, 0x800077B400060DAC, 0x00001D8100001D82,
-    0x800077B800001D80, 0x800077C0800077BC, 0x0000C086800077C4,
-    0x800077C800033487, 0x800077D0800077CC, 0x800077D8800077D4,
-    0x0007B1A8800077DC, 0x000709AA0007B1A9, 0x000701AC000709AB,
-    0x000701AE000701AD, 0x0003A88400073DAF, 0x800077E4800077E0,
-    0x800077EC800077E8, 0x800077F4800077F0, 0x000725B8800077F8,
-    0x0000C5BA0000CDB9, 0x0000CDBC0000DDBB, 0x0000F5BE0000FDBD,
-    0x000749B000009DBF, 0x000729B200075DB1, 0x000739B4000729B3,
-    0x000725B6000725B5, 0x000681B300071DB7, 0x80007800800077FC,
-    0x8000780880007804, 0x000689B5000681B6, 0x000699BB8000780C,
-    0x8000781000068DBA, 0x0006E1BF80007814, 0x0006FDBD0006FDBE,
-    0x800078180006FDBC, 0x800078208000781C, 0x0000318080007824,
-    0x0000298200003D81, 0x80007828000488EF, 0x8000782C000604E1,
-    0x800078300006DCE3, 0x80007834000174BE, 0x000601A380007838,
-    0x000028868000783C, 0x8000784000014487, 0x000609A5000601A6,
-    0x000619AB80007844, 0x8000784800060DAA, 0x000661AF8000784C,
-    0x00067DAD00067DAE, 0x8000785000067DAC, 0x800078540001B5B3,
-    0x0001ADB680007858, 0x800078608000785C, 0x00014DBA0001A5B5,
-    0x80007864000155BB, 0x00014DBE80007868, 0x00014DBC0000F1BF,
-    0x000269A8000145BD, 0x000279AA000269A9, 0x000269AC000279AB,
-    0x0002BDAE000269AD, 0x003C00840002B5AF, 0x800078708000786C,
-    0x8000787880007874, 0x800078808000787C, 0x000351B880007884,
-    0x000361BA000359B9, 0x000311BC000361BB, 0x000311BE000311BD,
-    0x0002CDB0000311BF, 0x0002DDB20002D5B1, 0x0002CDB40002D5B3,
-    0x000371B6000371B5, 0x80007888000371B7, 0x800078908000788C,
-    0x800078940002FDA3, 0x0002E5A60002EDA5, 0x8000789C80007898,
-    0x000205AA800078A0, 0x000205AC00021DAB, 0x000205AE00020DAD,
-    0x000065820003B9AF, 0x000055800002E4EF, 0x0003148700006D81,
-    0x003C54BE003D7486, 0x800076E4800078A8, 0x800078B0800078AC,
-    0x800078B8800078B4, 0x800078BC0002C4E3, 0x00027DA800016CE1,
-    0x000299AA000281A9, 0x0002B9AC000291AB, 0x0002D9AE0002B9AD,
-    0x800078C00002D5AF, 0x800078C8800078C4, 0x800078D0800078CC,
-    0x800078D8800078D4, 0x000171B8800078DC, 0x000171BA000171B9,
-    0x0001D1BC000171BB, 0x0001D1BE0001D1BD, 0x0002ADB00001CDBF,
-    0x0002B1B20002B1B1, 0x00029DB4000285B3, 0x000151B6000151B5,
-    0x0006A4E1000151B7, 0x0007B4E3800078E0, 0x003D1884003D80BE,
-    0x800078E8800078E4, 0x800078F0800078EC, 0x800078F8800078F4,
-    0x80007900800078FC, 0x0006C8EF80007904, 0x0000098180007908,
-    0x8000790C00007980, 0x000160E300001982, 0x0001A0E180007910,
-    0x00006CEF80007918, 0x003CC8868000791C, 0x80007920003D0C87,
-    0x8000792880007924, 0x003D91A38000792C, 0x80007914800078A4,
-    0x8000793480007930, 0x003DB1A5003DB9A6, 0x003DF5AB80007938,
-    0x8000793C003DE9AA, 0x003E15AF80007940, 0x003DD1AD003DD9AE,
-    0x80007944003DE5AC, 0x80007948003EADB3, 0x003EA9B68000794C,
-    0x8000795480007950, 0x003F49BA003EADB5, 0x80007958003F49BB,
-    0x003F49BE8000795C, 0x003F49BC003F49BF, 0x003E69A3003F49BD,
-    0x8000796480007960, 0x8000796C80007968, 0x003E69A5003E6DA6,
-    0x003F8DAB80007970, 0x80007974003F8DAA, 0x003F8DAF80007978,
-    0x003F8DAD003F8DAE, 0x8000797C003F8DAC, 0x003D05BA80007980,
-    0x003D09BC003D19BB, 0x003D35BE003D3DBD, 0x00004D80003EF9BF,
-    0x00002D8200002581, 0x000038BE003D7DB3, 0x003D55B6003D5DB5,
-    0x003FDDA800003084, 0x003FE1AA003FE1A9, 0x003FE1AC003FE1AB,
-    0x003E39AE003FEDAD, 0x80007984003E39AF, 0x0000888680007988,
-    0x8000798C0001AC87, 0x8000799480007990, 0x003ED9B880007998,
-    0x003EE9BA003ED9B9, 0x003EF9BC003EE9BB, 0x003EE9BE003EF9BD,
-    0x003E49B0003EE9BF, 0x003E59B2003E49B1, 0x003E49B4003E59B3,
-    0x003EFDB6003E49B5, 0x003E5DA8003EE9B7, 0x003E79AA003E61A9,
-    0x003E69AC003E79AB, 0x000195AE003E69AD, 0x0001A08400018DAF,
-    0x800079A08000799C, 0x800079A8800079A4, 0x800079B0800079AC,
-    0x0001A9B8800079B4, 0x00014DBA0001A9B9, 0x00015DBC000145BB,
-    0x000145BE000145BD, 0x0001FDB0000175BF, 0x0001C5B20001CDB1,
-    0x0001B9B40001ADB3, 0x0001ADB60001B9B5, 0x003DF1B30001A5B7,
-    0x000344BE800079B8, 0x800079BC000444BE, 0x003DE1B5003DE9B6,
-    0x000209BB800079C0, 0x800079C4000239BA, 0x00027DBF800079C8,
-    0x000279BD00027DBE, 0x800079CC000211BC, 0x800079D0003DB5A3,
-    0x003DADA6800079D4, 0x800079DC800079D8, 0x00027DAA003DA5A5,
-    0x800079E000024DAB, 0x000239AE800079E4, 0x000255AC000239AF,
-    0x0000118000023DAD, 0x0000118200001181, 0x800079E800060CBE,
-    0x800079F0800079EC, 0x800079F800026CEF, 0x800079FC000194E1,
-    0x80007A0000011CE3, 0x0004188680007A04, 0x80007A0800037087,
-    0x80007A1080007A0C, 0x80007A140003F9B3, 0x00037DB60003E9B5,
-    0x0007ACBE80007A18, 0x000359BA80007A1C, 0x000349BC000359BB,
-    0x0000FDBE000349BD, 0x000225A80000EDBF, 0x00023DAA000235A9,
-    0x00022DAC000235AB, 0x0002B1AE0002B1AD, 0x000968BE0002B1AF,
-    0x80007A2480007A20, 0x000AF8BE80007A28, 0x80007A3080007A2C,
-    0x00016DB880007A34, 0x000101BA000105B9, 0x000105BC000101BB,
-    0x000131BE00010DBD, 0x0002D1B0000131BF, 0x0002D1B20002D1B1,
-    0x000175B40002D1B3, 0x000165B6000165B5, 0x003F4CE1000155B7,
-    0x003E08E3000168E3, 0x80007A38003E7CE1, 0x80007A4080007A3C,
-    0x000D2CBE80007A44, 0x80007A4C80007A48, 0x000C3CBE80007A50,
-    0x003ECCEF0001BCEF, 0x0002F5A380007A54, 0x0000158100002D82,
-    0x80007A5800001D80, 0x0002E5A5000271A6, 0x000255AB80007A5C,
-    0x80007A60000255AA, 0x0001E1AF0005A084, 0x000245AD0001F1AE,
-    0x0006A1A8000245AC, 0x0006B9AA0006ADA9, 0x0006D1AC0006B9AB,
-    0x0006D1AE0006D1AD, 0x80007A640006CDAF, 0x0003088680007A68,
-    0x80007A6C00030887, 0x80007A7480007A70, 0x0007D1B880007A78,
-    0x0007D1BA0007D1B9, 0x0007F5BC0007D1BB, 0x0007F5BE0007FDBD,
-    0x0006BDB00007E1BF, 0x00074DB2000681B1, 0x00075DB4000745B3,
-    0x000741B6000741B5, 0x000631B3000741B7, 0x80007A8080007A7C,
-    0x80007A8880007A84, 0x000659B5000651B6, 0x000645BB80007A8C,
-    0x80007A90000679BA, 0x000785BF80007A94, 0x000791BD000799BE,
-    0x80007A98000799BC, 0x80007A9C000675A3, 0x000615A680007AA0,
-    0x80007AA880007AA4, 0x00063DAA00061DA5, 0x80007AAC000601AB,
-    0x0007DDAE80007AB0, 0x0007DDAC0007C1AF, 0x0005A0E10007D5AD,
-    0x0004F4E380007AB4, 0x80007ABC80007AB8, 0x80007AC480007AC0,
-    0x0000098180007AC8, 0x80007ACC00003980, 0x80007AD000001D82,
-    0x0005CCEF80007AD4, 0x000175B30002B084, 0x0000E88680007AD8,
-    0x80007ADC00012C87, 0x000165B50001D1B6, 0x0001FDBB80007AE0,
-    0x80007AE40001F5BA, 0x0001B9BF80007AE8, 0x0001D5BD0001B5BE,
-    0x0001A5A80001D5BC, 0x0001BDAA0001B5A9, 0x0001ADAC0001B5AB,
-    0x000111AE000111AD, 0x0001AC84000111AF, 0x80007AF080007AEC,
-    0x80007AF880007AF4, 0x80007B0080007AFC, 0x000119B880007B04,
-    0x000125BA00012DB9, 0x0000DDBC0000C9BB, 0x0000CDBE0000C5BD,
-    0x000171B00000C5BF, 0x000141B2000179B1, 0x000131B4000141B3,
-    0x000135B600013DB5, 0x80007B08000129B7, 0x80007B1080007B0C,
-    0x80007B14000239A3, 0x00029DA6000229A5, 0x80007B1C80007B18,
-    0x0002B9AA80007B20, 0x000299AC0002B1AB, 0x0002F9AE000299AD,
-    0x80007B240002F5AF, 0x000CE88680007B28, 0x80007B2C00036887,
-    0x80007B3480007B30, 0x00003D8080007B38, 0x0000198200000981,
-    0x80007B4080007B3C, 0x80007B4C80007B48, 0x80007B5480007B50,
-    0x000690EF800079F4, 0x80007B5C80007B58, 0x80007B6480007B60,
-    0x000CEC8480007B68, 0x80007B7080007B6C, 0x0001E0E180007B74,
-    0x000108E380007B78, 0x00039DB380007B7C, 0x80007B8480007B80,
-    0x80007B8C80007B88, 0x0003B5B5000371B6, 0x000355BB80007B90,
-    0x80007B94000355BA, 0x0000F5BF80007B98, 0x000345BD0000F5BE,
-    0x000245A8000345BC, 0x000255AA000255A9, 0x00027DAC000265AB,
-    0x0002B1AE0002B1AD, 0x00001D810002B1AF, 0x80007B9C00001D80,
-    0x80007BA00000AD82, 0x80007BA880007BA4, 0x00015DB880007BAC,
-    0x00016DBA000165B9, 0x000115BC00010DBB, 0x00010DBE000105BD,
-    0x0002D1B0000105BF, 0x0002D1B20002D1B1, 0x000171B40002D1B3,
-    0x000171B6000171B5, 0x80007BB000016DB7, 0x80007BB8002BD0E1,
-    0x000F90860017D8E3, 0x80007BBC000D5C87, 0x80007BC00007E8EF,
-    0x80007BC4000194E1, 0x000BB0BE000080E3, 0x80007BCC80007BC8,
-    0x00029DA3003F60EF, 0x80007BD480007BD0, 0x80007BDC80007BD8,
-    0x0002B5A5000271A6, 0x000255AB80007BE0, 0x80007BE4000255AA,
-    0x0001F5AF80007BE8, 0x000245AD0001F5AE, 0x000E75A8000245AC,
-    0x000E75AA000E7DA9, 0x000E9DAC000E8DAB, 0x000E8DAE000E85AD,
-    0x80007BB4000E85AF, 0x80007BF080007BEC, 0x80007BF880007BF4,
-    0x80007C0080007BFC, 0x000E8DB880007C04, 0x000E95BA000E95B9,
-    0x000EBDBC000EA5BB, 0x000F69BE000F7DBD, 0x000EFDB0000F69BF,
-    0x000EC1B2000EC1B1, 0x000EC5B4000EC1B3, 0x000EBDB6000EC9B5,
-    0x000EFDB3000EB5B7, 0x80007C0C80007C08, 0x80007C1480007C10,
-    0x000E15B5000E05B6, 0x000E05BB80007C18, 0x80007C1C000E39BA,
-    0x000E0DBF80007C20, 0x000E15BD000E01BE, 0x00006D82000E15BC,
-    0x00005580000EB9A3, 0x000E41A600006581, 0x80007C2800011884,
-    0x000E7DAA000E51A5, 0x00040086000E41AB, 0x000E45AE00007887,
-    0x000E51AC000E49AF, 0x0025298F000E51AD, 0x80007C2C0003F4E1,
-    0x80007C30000154E3, 0x80007C34000DA4E3, 0x80007C38000D4CE1,
-    0x80007C3C0008FCEF, 0x80007C4480007C40, 0x80007C4C80007C48,
-    0x0008619F000864EF, 0x00094D9D00094D9E, 0x00350D9B0034359C,
-    0x0037B9990035519A, 0x0031419700316D98, 0x003DB99500338996,
-    0x003FA993003D4D94, 0x00397D910039AD92, 0x0015A5B100397D90,
-    0x01E989B30015A5B0, 0x01E855B5001441B2, 0x80007C5001E989B4,
-    0x80007C5880007C54, 0x80007C6080007C5C, 0x000329A380007C64,
-    0x000D99A1000C81A2, 0x0001E1A1000F39A0, 0x0005CDA380007C68,
-    0x000481A500034DA2, 0x0019D5A70005DDA4, 0x001DE1A90019DDA6,
-    0x001CE5AB00188DA8, 0x0011D1AD001DE9AA, 0x001401AF001001AC,
-    0x00262982001151AE, 0x80007C6C003A0583, 0x003CC98680007C70,
-    0x003A1D84003E3587, 0x003EF18A003AC585, 0x80007C740030C98B,
-    0x0034D18E80007C78, 0x0032198C0036318F, 0x00361D9200320D8D,
-    0x0003448400373D93, 0x000B71960004C484, 0x000A5D94000CD197,
-    0x000E419A000A5595, 0x80007C7C000F519B, 0x80007C8480007C80,
-    0x0000E19C80007C88, 0x80007C2480007C8C, 0x80007C9880007C90,
-    0x0000ED8180007C9C, 0x80007B440000E180, 0x000D29A800001982,
-    0x000D09AA000D3DA9, 0x000D05AC000D05AB, 0x000D7DAE000D0DAD,
-    0x000D0DB0000D75AF, 0x000D15B2000D15B1, 0x000D3DB4000D25B3,
-    0x0002F1B60002F1B5, 0x0002D1B80002F1B7, 0x0002D1BA0002D1B9,
-    0x000275BC0002D1BB, 0x000275BE00027DBD, 0x000CA1B300026DBF,
-    0x0002B48780007CA0, 0x80007CA400052C86, 0x000CADB5000CADB6,
-    0x000CF1BB80007CA8, 0x80007CAC000CE9BA, 0x000315BF80007CB0,
-    0x000CD1BD000CD9BE, 0x80007CB4000CE1BC, 0x80007CB8000CE5A3,
-    0x000CE9A680007CBC, 0x80007CC480007CC0, 0x000CADAA000CE9A5,
-    0x80007CC8000CB5AB, 0x000C9DAE80007CCC, 0x000CA5AC000351AF,
-    0x80007CD0000C95AD, 0x80007CD880007CD4, 0x80007CE080007CDC,
-    0x80007CE880007CE4, 0x80007CEC000A54EF, 0x80007CF0000E2CE1,
-    0x80007CF4000168E3, 0x0000718100007982, 0x0002D1B300007980,
-    0x80007CFC80007CF8, 0x80007D0880007D00, 0x0002F9B50002FDB6,
-    0x0002D9BB80007D0C, 0x0004C8860002D1BA, 0x000255BF0004AC87,
-    0x000265BD000255BE, 0x000511A3000265BC, 0x80007D0480007C94,
-    0x80007D1480007D10, 0x000539A500053DA6, 0x000519AB80007D18,
-    0x80007D1C000511AA, 0x000595AF80007D20, 0x0005A5AD000595AE,
-    0x80007D240005A5AC, 0x80007D28000741B3, 0x000779B680007D2C,
-    0x80007D3480007D30, 0x000729BA000741B5, 0x80007D38000729BB,
-    0x000719BE80007D3C, 0x00072DBC000719BF, 0x000609A8000725BD,
-    0x00078DAA000609A9, 0x00078DAC00079DAB, 0x0007B5AE0007BDAD,
-    0x80007D400007DDAF, 0x80007D4880007D44, 0x80007D5080007D4C,
-    0x80007D5880007D54, 0x000749B880007D5C, 0x000759BA000749B9,
-    0x000749BC000759BB, 0x000771BE000775BD, 0x0007A9B000076DBF,
-    0x0007BDB20007A9B1, 0x000799B40007B5B3, 0x000779B6000799B5,
-    0x000605A3000779B7, 0x80007D6480007D60, 0x80007D6C80007D68,
-    0x000605A500063DA6, 0x00066DAB80007D70, 0x0002A88400066DAA,
-    0x00065DAF80007D78, 0x000661AD00065DAE, 0x00004D80000669AC,
-    0x0000558200005581, 0x80007D74000141B3, 0x000171B6000149B5,
-    0x0000C08680007D7C, 0x000121BA0001A887, 0x000125BC000121BB,
-    0x000111BE00012DBD, 0x000651A8000111BF, 0x000185AA000659A9,
-    0x00018DAC00019DAB, 0x0001B5AE0001BDAD, 0x0000A0840001DDAF,
-    0x80007D8480007D80, 0x80007D8C80007D88, 0x80007D9480007D90,
-    0x000169B880007D98, 0x000179BA000169B9, 0x000169BC000179BB,
-    0x0000DDBE000169BD, 0x0001A9B00000D5BF, 0x0001BDB20001A9B1,
-    0x000199B40001B5B3, 0x000159B6000199B5, 0x00040DBA000159B7,
-    0x0007CDB800042DBB, 0x00047DBE000405B9, 0x000435BC000465BF,
-    0x0007B9B2000475BD, 0x00071DB00007B9B3, 0x000799B6000715B1,
-    0x000799B4000799B7, 0x000751AA000799B5, 0x000691A800075DAB,
-    0x00077DAE000745A9, 0x000779AC000765AF, 0x80007D9C000775AD,
-    0x80007DA480007DA0, 0x80007DAC80007DA8, 0x80007DB480007DB0,
-    0x80007DBC80007DB8, 0x001C00BE80007DC0, 0x80007DC4000281A3,
-    0x0002B1A6000289A5, 0x80007DCC80007DC8, 0x0002E1AA80007DD0,
-    0x0002E5AC0002E1AB, 0x0002D1AE0002EDAD, 0x80007DD40002D1AF,
-    0x0000118100002582, 0x80007DD800001D80, 0x80007DDC0001ACE1,
-    0x80007DE000016CE3, 0x80007DEC80007DE8, 0x001CB08700029CEF,
-    0x80007DF0001CD486, 0x00022DA880007DF4, 0x0002ADAA000259A9,
-    0x0002BDAC0002A5AB, 0x0002ADAE0002A5AD, 0x001DA0840002A5AF,
-    0x80007DFC80007DF8, 0x80007E0480007E00, 0x001CB4BE80007E08,
-    0x00013DB880007E0C, 0x0001CDBA0001C5B9, 0x0001CDBC0001DDBB,
-    0x0001FDBE0001F5BD, 0x0002DDB000019DBF, 0x00010DB2000105B1,
-    0x00011DB4000105B3, 0x00010DB6000105B5, 0x000694E1000105B7,
-    0x0007D4E30000BCE3, 0x80007E1000067CE1, 0x80007E1880007E14,
-    0x80007E2080007E1C, 0x80007E2880007E24, 0x001F78BE80007E2C,
-    0x000EF0EF000AFCEF, 0x80007E3480007E30, 0x80007E3C80007E38,
-    0x80007E40000269B3, 0x000269B6000279B5, 0x80007E4880007E44,
-    0x000255BA80007E4C, 0x000195BC00018DBB, 0x000195BE00019DBD,
-    0x80007DE400018DBF, 0x80007E50001DA1A3, 0x001DA1A680007E54,
-    0x80007E5C80007E58, 0x001D9DAA001DB1A5, 0x00030886001E45AB,
-    0x001E5DAE00037C87, 0x001E5DAC001E45AF, 0x001FA180001E55AD,
-    0x00004182001FA181, 0x80007E60001EB5B3, 0x001F71B6001F71B5,
-    0x80007E6880007E64, 0x001F5DBA80007E6C, 0x001FE1BC001F51BB,
-    0x001FE1BE001FE1BD, 0x001E65A8001FE1BF, 0x001E71AA001E6DA9,
-    0x001FB5AC001E71AB, 0x001FB5AE001FBDAD, 0x80007E70001FADAF,
-    0x80007E7880007E74, 0x80007E8080007E7C, 0x80007E8880007E84,
-    0x001F69B880007E8C, 0x001F75BA001F69B9, 0x001F15BC001F0DBB,
-    0x001F09BE001F1DBD, 0x001FD5B0001F01BF, 0x001F79B2001FD9B1,
-    0x001F69B4001F79B3, 0x001F59B6001F69B5, 0x001FF9A3001F59B7,
-    0x80007E9480007E90, 0x80007E9C80007E98, 0x001E3DA5001E3DA6,
-    0x001E1DAB80007EA0, 0x80007EA4001E11AA, 0x001EADAF80007EA8,
-    0x001EADAD001EADAE, 0x80007EAC001EADAC, 0x80007EB480007EB0,
-    0x00001D8080007EB8, 0x0000398200000981, 0x80007EC080007EBC,
-    0x80007EC40001A884, 0x0000D08680007EC8, 0x80007ECC0001FC87,
-    0x000195A880007ED0, 0x0001D9AA00019DA9, 0x0001F1AC0001D1AB,
-    0x000139AE0001F1AD, 0x0000A084000139AF, 0x80007ED880007ED4,
-    0x80007EE080007EDC, 0x80007EE880007EE4, 0x0000CDB880007EEC,
-    0x0000D5BA0000D5B9, 0x0000FDBC0000E5BB, 0x00009DBE000095BD,
-    0x000149B0000095BF, 0x000159B2000149B1, 0x000149B4000159B3,
-    0x0000FDB6000149B5, 0x000251B30000F5B7, 0x80007EF480007EF0,
-    0x80007EFC80007EF8, 0x000279B5000271B6, 0x00028DBB80007F00,
-    0x80007F0400028DBA, 0x00028DBF80007F08, 0x00028DBD00028DBE,
-    0x01F43C8500028DBC, 0x80007F0C000215A3, 0x000235A680007F10,
-    0x80007F1880007F14, 0x0002C9AA00023DA5, 0x80007F1C0002C9AB,
-    0x0002C9AE80007F20, 0x0002C9AC0002C9AF, 0x000019820002C9AD,
-    0x0000198080007F24, 0x80007F2800001981, 0x80007F3080007F2C,
-    0x000400BE80007F38, 0x80007F4080007F3C, 0x80007F4880007F44,
-    0x80007F5080007F4C, 0x0004E08680007F54, 0x80007F5800035487,
-    0x80007F6080007F5C, 0x80007F6400028CEF, 0x80007F6C80007F68,
-    0x80007F70000154E3, 0x80007F74000184E1, 0x80007F7C80007F78,
-    0x00026DA880007F80, 0x000275AA00027DA9, 0x0002B9AC0002A9AB,
-    0x0002A9AE0002B9AD, 0x0005AC840002A9AF, 0x80007F8880007F84,
-    0x80007F9080007F8C, 0x80007F9880007F94, 0x00016DB880007F9C,
-    0x000105BA00010DB9, 0x000109BC000119BB, 0x000135BE00013DBD,
-    0x0002D9B00001D9BF, 0x00016DB20002D9B1, 0x00017DB4000165B3,
-    0x000165B6000165B5, 0x0000F4E3000155B7, 0x0001B8E1001F68E1,
-    0x0000A980001EBCE3, 0x000005820000A981, 0x00042CBE80007FA0,
-    0x80007FAC80007FA8, 0x001EF8EF80007FB0, 0x80007FB880007FB4,
-    0x0004E086001E7CEF, 0x80007FBC00057487, 0x80007FC00002B1B3,
-    0x000275B600027DB5, 0x80007FC880007FC4, 0x000251BA80007FCC,
-    0x0001F5BC000251BB, 0x0001F5BE0001FDBD, 0x80007F340001EDBF,
-    0x0005B1AA80007FA4, 0x000615AC0005B1AB, 0x000615AE00061DAD,
-    0x80007FD000060DAF, 0x80007FD880007FD4, 0x80007FDC000551A3,
-    0x000595A600059DA5, 0x0006E9B380007FE0, 0x80007FE880007FE4,
-    0x80007FF080007FEC, 0x0006E9B50006E1B6, 0x0006ADBB80007FF4,
-    0x80007FF80006ADBA, 0x000691BF80007FFC, 0x0006A1BD0006A9BE,
-    0x00061DA80006ADBC, 0x000621AA000621A9, 0x000621AC000621AB,
-    0x000621AE000621AD, 0x80008000000621AF, 0x8000800880008004,
-    0x000019808000800C, 0x0000058200001981, 0x000635B880008010,
-    0x000635BA00063DB9, 0x0007F9BC0007E9BB, 0x0007E9BE0007F9BD,
-    0x000629B00007E9BF, 0x000639B2000629B1, 0x00062DB4000635B3,
-    0x000611B6000615B5, 0x0007A9A300060DB7, 0x000184BE00028C84,
-    0x8000801880008014, 0x0007A9A50007A1A6, 0x0007EDAB8000801C,
-    0x000088860007EDAA, 0x0007D1AF00017887, 0x0007E1AD0007E9AE,
-    0x800080200007EDAC, 0x800080240006E9B3, 0x000699B680008028,
-    0x800080308000802C, 0x000681BA0006E5B5, 0x80008034000681BB,
-    0x000159BE80008038, 0x000159BC000145BF, 0x000625A8000159BD,
-    0x000625AA00062DA9, 0x000629AC000639AB, 0x000649AE00065DAD,
-    0x0001AC84000641AF, 0x800080408000803C, 0x8000804880008044,
-    0x800080508000804C, 0x000189B880008054, 0x000199BA000189B9,
-    0x0001B9BC000191BB, 0x000179BE0001B9BD, 0x000601B0000179BF,
-    0x0001C9B2000601B1, 0x0001D9B40001C9B3, 0x0001C5B60001C5B5,
-    0x0005ADA30001B9B7, 0x8000805C80008058, 0x8000806480008060,
-    0x0005A1A50005DDA6, 0x0005C5AB80008068, 0x8000806C0005C5AA,
-    0x000201AF80008070, 0x00021DAD00021DAE, 0x8000807400021DAC,
-    0x8000807C80008078, 0x0000198180008080, 0x8000808800001180,
-    0x000324BE00001182, 0x000CC8868000808C, 0x8000809000034C87,
-    0x8000809880008094, 0x000130E38000809C, 0x0001B8E1800080A0,
-    0x800080A4000CC084, 0x800080AC800080A8, 0x000CE8BE800080B0,
-    0x800080B8800080B4, 0x000228EF800080BC, 0x800080C4800080C0,
-    0x000381B3800080C8, 0x800080D0800080CC, 0x800080D8800080D4,
-    0x000395B5000381B6, 0x000345BB800080DC, 0x800080E0000389BA,
-    0x000345BF800080E4, 0x000355BD000345BE, 0x000211A8000355BC,
-    0x000249AA000219A9, 0x00025DAC000245AB, 0x000241AE000241AD,
-    0x00000D80000275AF, 0x0000118200001181, 0x000C64BE800080E8,
-    0x800080F0800080EC, 0x000145B8800080F8, 0x000145BA00014DB9,
-    0x00014DBC00015DBB, 0x000175BE00017DBD, 0x00020DB000011DBF,
-    0x0002C1B20002C1B1, 0x0002C1B40002C1B3, 0x0002C1B60002C1B5,
-    0x000658E10002C1B7, 0x00009CE300075CE1, 0x000C30860006BCE3,
-    0x800080FC000D5C87, 0x8000810480008100, 0x8000810C80008108,
-    0x000000EE80008110, 0x000048EF80008114, 0x800081180006E4EF,
-    0x000289A38000811C, 0x00029DA580008120, 0x8000812880008124,
-    0x8000812C000289A6, 0x00024DAB80008130, 0x00025DAD000281AA,
-    0x00024DAF00025DAC, 0x000E59A800024DAE, 0x000E79AA000E61A9,
-    0x000E69AC000E79AB, 0x000E9DAE000E69AD, 0x800080F4000E8DAF,
-    0x8000813880008134, 0x800081408000813C, 0x8000814880008144,
-    0x000EFDB88000814C, 0x000E81BA000E81B9, 0x000E81BC000E81BB,
-    0x000EB1BE000E89BD, 0x000EF5B0000EB1BF, 0x000EF5B2000EFDB1,
-    0x000ED9B4000EC9B3, 0x000ECDB6000ED9B5, 0x000EF1B3000EC5B7,
-    0x8000815480008150, 0x8000815C80008158, 0x000EE1B5000EE9B6,
-    0x000E39BB80008160, 0x0003E484000E31BA, 0x000E1DBF80008164,
-    0x000E1DBD000E1DBE, 0x00005D82000E21BC, 0x00004580000EB5A3,
-    0x000EADA600005D81, 0x8000816C80008168, 0x000E75AA000EA5A5,
-    0x0000E886000E7DAB, 0x000E59AE0000E087, 0x000E65AC000E59AF,
-    0x000E3DA8000E59AD, 0x000E95AA000E9DA9, 0x000EB5AC000EADAB,
-    0x000EB5AE000EBDAD, 0x80008170000EA9AF, 0x8000817880008174,
-    0x800081808000817C, 0x8000818880008184, 0x000169B88000818C,
-    0x000179BA000169B9, 0x000169BC000179BB, 0x0001DDBE000169BD,
-    0x000EA1B00001D5BF, 0x000EA5B2000EA1B1, 0x000EA9B4000EB9B3,
-    0x000E9DB6000E95B5, 0x000E1DB3000159B7, 0x8000819480008190,
-    0x8000819C80008198, 0x000E3DB5000E3DB6, 0x000E65BB800081A0,
-    0x800081A4000E65BA, 0x000E51BF800081A8, 0x000E6DBD000E6DBE,
-    0x000E65A5000E71BC, 0x800081B0800081AC, 0x800081B4000E65A6,
-    0x000E45A3800081B8, 0x000E35AD800081BC, 0x000E09AF000E29AC,
-    0x800081C0000E35AE, 0x000E3DAB80008084, 0x800081C4000E3DAA,
-    0x00001D8100001D82, 0x800081C800001D80, 0x800081D0800081CC,
-    0x800081D8000350BE, 0x800081E0800081DC, 0x800081E8800081E4,
-    0x800081F0800081EC, 0x0003C084800081F4, 0x000E10EF800081F8,
-    0x0004F08600040084, 0x800081FC00035487, 0x8000820480008200,
-    0x8000820C80008208, 0x000E00E180008210, 0x000EE4E380008214,
-    0x8000821C80008218, 0x8000822480008220, 0x8000822C80008228,
-    0x8000823480008230, 0x000275A880008238, 0x000389AA00027DA9,
-    0x000399AC000389AB, 0x000389AE000399AD, 0x0003F9B0000389AF,
-    0x00034DB20003F9B1, 0x00035DB4000345B3, 0x000345B6000345B5,
-    0x00034DB8000375B7, 0x000321BA000321B9, 0x000325BC00033DBB,
-    0x000309BE000329BD, 0x00019D800000FDBF, 0x00001D8200000981,
-    0x0003F4EF8000823C, 0x000550BE80008240, 0x000370E380008248,
-    0x0001D8E18000824C, 0x8000825480008250, 0x8000825C80008258,
-    0x0004E08680008260, 0x8000826400028887, 0x8000826800026DB3,
-    0x000275B600027DB5, 0x800082708000826C, 0x000251BA80008274,
-    0x0002F5BC000251BB, 0x0002F5BE0002FDBD, 0x800082780002EDBF,
-    0x800082808000827C, 0x8000828880008284, 0x800082908000828C,
-    0x0004EC84000E54EF, 0x80008294000E2CE1, 0x80008298000114E3,
-    0x800082A08000829C, 0x0002E1A3800082A4, 0x0000158100002D82,
-    0x800082A800001D80, 0x0002F1A50002F9A6, 0x0002DDAB800082AC,
-    0x800082B00002DDAA, 0x000261AF0004B0BE, 0x000271AD000279AE,
-    0x000621A8000279AC, 0x000621AA000621A9, 0x000621AC000621AB,
-    0x000621AE000621AD, 0x80008244000621AF, 0x00000086800082B4,
-    0x800082B80003A087, 0x800082C0800082BC, 0x00063DB8800082C4,
-    0x0006CDBA0006C5B9, 0x0006DDBC0006C5BB, 0x0006CDBE0006C5BD,
-    0x000625B00006C5BF, 0x000635B200063DB1, 0x000615B400060DB3,
-    0x000615B600061DB5, 0x0006A9B3000609B7, 0x800082CC800082C8,
-    0x800082D4800082D0, 0x0006A9B50006A1B6, 0x0006F9BB800082D8,
-    0x800082DC0006E1BA, 0x0006D9BF800082E0, 0x0006D9BD0006D1BE,
-    0x800082E40006E1BC, 0x800082E80006EDA3, 0x0006E5A6800082EC,
-    0x800082F4800082F0, 0x0006A5AA0006EDA5, 0x800082F80006BDAB,
-    0x000695AE800082FC, 0x0006A5AC00069DAF, 0x00061DA800069DAD,
-    0x000625AA000621A9, 0x000625AC00063DAB, 0x000625AE00062DAD,
-    0x80008300000655AF, 0x8000830880008304, 0x800083108000830C,
-    0x8000831880008314, 0x0006D1B88000831C, 0x0006D1BA0006D1B9,
-    0x0006E1BC0006E5BB, 0x0006E1BE0006E1BD, 0x000631B00006E1BF,
-    0x000609B2000639B1, 0x0006F1B4000609B3, 0x0006F1B60006F1B5,
-    0x0006ADB30006F1B7, 0x0000B58100001582, 0x800083200000B580,
-    0x0006ADB50006ADB6, 0x0006FDBB000010BE, 0x00016C840006F5BA,
-    0x000635BF80008328, 0x0006DDBD000635BE, 0x8000832C0006DDBC,
-    0x0006FDA50006FDA6, 0x0006FDA380008324, 0x01FC008680008330,
-    0x000665AF00016087, 0x00068DAD000665AE, 0x0006ADAB00068DAC,
-    0x800081D40006A5AA, 0x8000833880008334, 0x000231B38000833C,
-    0x000295B580008340, 0x8000834880008344, 0x8000834C000295B6,
-    0x000281BB80008350, 0x0002F1BD0002B1BA, 0x0002F5BF000299BC,
-    0x00021CE30002F5BE, 0x000188E10007A4EF, 0x800083540003C484,
-    0x8000835C80008358, 0x8000836480008360, 0x8000836C80008368,
-    0x000228EF80008370, 0x80008374000628E3, 0x0002B9AA000634E1,
-    0x0002B9A80002B9AB, 0x00021DAE0002B9A9, 0x00021DAC000205AF,
-    0x0003ECBE000215AD, 0x8000837C80008378, 0x8000838480008380,
-    0x0001B1810001B982, 0x0003E1BA0001B980, 0x000221B80003E1BB,
-    0x0003E1BE000221B9, 0x0003F1BC0003E1BF, 0x00024DB20003F1BD,
-    0x000275B0000225B3, 0x000225B6000245B1, 0x000231B400022DB7,
-    0x000339A3000231B5, 0x8000838C80008388, 0x8000839880008390,
-    0x00039DA500039DA6, 0x000389AB8000839C, 0x800083A00003B9AA,
-    0x0003FDAF800083A4, 0x0003F9AD0003FDAE, 0x800083A8000391AC,
-    0x01FDC886000291B3, 0x000155B601FD4087, 0x800083B0800083AC,
-    0x000179BA00015DB5, 0x800083B4000145BB, 0x000125BE800083B8,
-    0x00015DBC00012DBF, 0x800083BC00012DBD, 0x800083C4800083C0,
-    0x800083C801FDC084, 0x0000FCEF800083CC, 0x000190E1000688EF,
-    0x000010E30006FCE1, 0x800083D000061CE3, 0x01FDF0BE800083D4,
-    0x000215A3800083D8, 0x800083E0800083DC, 0x800083E8800083E4,
-    0x0001D9A50001D1A6, 0x0001C1AB800083EC, 0x800083F00001FDAA,
-    0x0001A9AF800083F4, 0x0001A9AD0001A1AE, 0x800083940001D9AC,
-    0x00001D8100001D82, 0x800083F800001D80, 0x80008400800083FC,
-    0x01FF31A880008404, 0x01FF01AA01FF39A9, 0x01FF71AC01FF01AB,
-    0x01FF61AE01FF61AD, 0x01FF2DB001FF55AF, 0x01FFCDB201FFC5B1,
-    0x01FFC1B401FFC5B3, 0x01FFE9B601FFC1B5, 0x01FFDDB801FFE9B7,
-    0x01FFA1BA01FFA1B9, 0x01FFA1BC01FFA1BB, 0x01FF91BE01FFA9BD,
-    0x01FFA9B301FF91BF, 0x0003288680008408, 0x8000840C00031C87,
-    0x01FFD9B501FFC1B6, 0x01FF21BB80008410, 0x8000841401FFC9BA,
-    0x01FF25BF80008418, 0x01FF31BD01FF25BE, 0x8000841C01FF39BC,
-    0x8000842001FFEDA3, 0x01FF85A680008424, 0x8000842C80008428,
-    0x01FF8DAA01FF9DA5, 0x8000843001FF65AB, 0x01FF61AE80008434,
-    0x01FF7DAC01FF61AF, 0x000185A801FF75AD, 0x000195AA000195A9,
-    0x0001BDAC0001A5AB, 0x0001D1AE0001D5AD, 0x800084380001D1AF,
-    0x800084408000843C, 0x8000844880008444, 0x800084508000844C,
-    0x000071B880008454, 0x000071BA000071B9, 0x0000D1BC000071BB,
-    0x0000D1BE0000D1BD, 0x0001B5B00000D1BF, 0x000181B20001BDB1,
-    0x000051B4000181B3, 0x000051B6000051B5, 0x80008458000051B7,
-    0x800084608000845C, 0x00000D8080008464, 0x00003D8200003581,
-    0x8000846C80008468, 0x0002A48480008470, 0x00019086000120BE,
-    0x8000847800011487, 0x00015DB38000847C, 0x8000848080008474,
-    0x8000848880008484, 0x000175B5000115B6, 0x000139BB8000848C,
-    0x80008490000131BA, 0x0000D5BF80008494, 0x000115BD0000D5BE,
-    0x00029DA3000115BC, 0x8000849C80008498, 0x800084A0000400BE,
-    0x0002B5A50002D5A6, 0x0002F9AB800084A4, 0x800084A80002F1AA,
-    0x000315AF800084AC, 0x0002D5AD000315AE, 0x0001E8E10002D5AC,
-    0x000214E3800084B0, 0x800084B400034084, 0x800084BC800084B8,
-    0x800084C4800084C0, 0x800084CC800084C8, 0x800084D4800084D0,
-    0x00028CEF800084D8, 0x000339A8800084DC, 0x000341AA000339A9,
-    0x000341AC000341AB, 0x000371AE000349AD, 0x0000D580000371AF,
-    0x000005820000D981, 0x0005B485800084E0, 0x800084EC800084E8,
-    0x000095B8800084F0, 0x000095BA00009DB9, 0x0000B5BC0000ADBB,
-    0x0000B5BE0000BDBD, 0x0000E5B00000ADBF, 0x0000F9B20000EDB1,
-    0x0000D9B40000F1B3, 0x0000BDB60000D9B5, 0x01FD5DB60000ADB7,
-    0x01FD5DB401FD65B7, 0x01FD35B201FD55B5, 0x01FD01B001FD45B3,
-    0x01FD8DBE01FD3DB1, 0x01FD8DBC01FD95BF, 0x01FD5DBA01FD85BD,
-    0x01FD5DB801FD4DBB, 0x0007E08601FD55B9, 0x800084F400051887,
-    0x800084FC800084F8, 0x8000850000071484, 0x01FD6DAE80008504,
-    0x01FD69AC01FD71AF, 0x01FEA5AA01FD69AD, 0x01FED5A801FE89AB,
-    0x000324E101FEBDA9, 0x001FDCE301FED4E1, 0x000028E301FE10E3,
-    0x000108E180008508, 0x0000F4EF8000850C, 0x0007F4BE80008510,
-    0x8000851880008514, 0x001F7CEF8000851C, 0x000209B301FEFCEF,
-    0x8000852480008520, 0x8000852C80008528, 0x000209B5000209B6,
-    0x0001E9BB80008530, 0x800085340001E1BA, 0x0001DDBF80008538,
-    0x0001E5BD0001DDBE, 0x01F8259F0001EDBC, 0x01F9F19D01F8259E,
-    0x01FC319B01FCA19C, 0x01F1159901FDE19A, 0x01F0219701F01998,
-    0x01F4259501F54D96, 0x01E9619301F42D94, 0x01E83D9101E83D92,
-    0x01EC618F01EEBD90, 0x8000853C800084E4, 0x8000854480008540,
-    0x00003D8100003D82, 0x8000854800003D80, 0x800085508000854C,
-    0x0001D08680008554, 0x8000855800030087, 0x000549A38000855C,
-    0x8000856480008560, 0x8000856C80008568, 0x000549A5000549A6,
-    0x0006A9AB80008570, 0x800085740006A1AA, 0x00069DAF80008578,
-    0x0006A5AD00069DAE, 0x8000857C0006ADAC, 0x8000858480008580,
-    0x8000858C80008588, 0x8000859480008590, 0x8000859C80008598,
-    0x800085A4800085A0, 0x000535A3800085A8, 0x0004F9A10004E9A2,
-    0x01ED218201F931A0, 0x800085AC01EF8583, 0x01E84186800085B0,
-    0x01E91D8401F51587, 0x01F52D8A01E91585, 0x800085B401F44D8B,
-    0x01F03D8E800085B8, 0x01F1398C01FCC18F, 0x01FD619201F1398D,
-    0x800085BC01FDE993, 0x01F97596800085C0, 0x01F8A19401F89997,
-    0x0005619A01F97D95, 0x00002C840005799B, 0x00003981000C00BE,
-    0x0004BD9C00005980, 0x0001C5A100003182, 0x000D89A3800085C8,
-    0x000D01A50001E1A2, 0x0009B9A7000D91A4, 0x001401A9000801A6,
-    0x001599AB00090DA8, 0x0011B1AD00158DAA, 0x001C01AF001421AC,
-    0x001DCDB1001115AE, 0x001801B3001DC5B0, 0x001939B5001C55B2,
-    0x000F58860019F5B4, 0x800085CC0000FC87, 0x800085D4800085D0,
-    0x800085DC800085D8, 0x800085E4800085E0, 0x0004B5A8800085E8,
-    0x0004B5AA0004BDA9, 0x00055DAC000545AB, 0x000541AE000545AD,
-    0x800085EC000541AF, 0x800085F4800085F0, 0x800085FC800085F8,
-    0x8000860480008600, 0x00058DB880008608, 0x000595BA000595B9,
-    0x0005BDBC0005A5BB, 0x000215BE00021DBD, 0x0005F5B000020DBF,
-    0x0005CDB20005F9B1, 0x0005C1B40005C1B3, 0x0005BDB60005C1B5,
-    0x8000860C0005B5B7, 0x8000861480008610, 0x80008618000431B3,
-    0x0005C9B60005D9B5, 0x800086208000861C, 0x00053DBA80008624,
-    0x000515BC000531BB, 0x000509BE000519BD, 0x80008628000509BF,
-    0x000585A5000595A6, 0x00046DA38000862C, 0x000C90BE80008630,
-    0x000555AF80008638, 0x000545AD000555AE, 0x00056DAB000549AC,
-    0x8000863C000561AA, 0x000379A880008640, 0x00038DAA000379A9,
-    0x00039DAC000385AB, 0x000385AE000385AD, 0x0001E1800003B5AF,
-    0x0001FD820001E981, 0x80008644000D6084, 0x000D248780008648,
-    0x00039DB8000C9C86, 0x0003ADBA0003A5B9, 0x0003BDBC0003A5BB,
-    0x000255BE00025DBD, 0x0003CDB0000275BF, 0x0003ADB20003A5B1,
-    0x0003BDB40003A5B3, 0x0003ADB60003A5B5, 0x8000864C0003A5B7,
-    0x00030DB380008650, 0x00031DB580008654, 0x80008658000CDCBE,
-    0x8000865C00030DB6, 0x000369BB80008660, 0x000379BD000301BA,
-    0x000369BF000371BC, 0x000698E1000371BE, 0x00068CE380008664,
-    0x000044EF80008668, 0x800086708000866C, 0x000098E380008674,
-    0x000180E180008678, 0x800086808000867C, 0x0006A8EF80008684,
-    0x800085C480008688, 0x8000868C80008634, 0x8000869480008690,
-    0x0000918100009982, 0x000DA5A800009980, 0x000DCDAA000DC5A9,
-    0x000DDDAC000DC5AB, 0x000DCDAE000DC5AD, 0x000DBDB0000DC5AF,
-    0x000E6DB2000E65B1, 0x000E7DB4000E65B3, 0x000E6DB6000E65B5,
-    0x000E5DB8000E65B7, 0x000EEDBA000EE5B9, 0x000EE1BC000EE5BB,
-    0x000ED9BE000EE9BD, 0x000D65B3000ED9BF, 0x8000869C80008698,
-    0x800086A4800086A0, 0x000D75B5000D65B6, 0x000E81BB0003D084,
-    0x00000886000D59BA, 0x000E81BF00008C87, 0x000E91BD000E99BE,
-    0x800086A8000E99BC, 0x800086AC000D21A3, 0x000D21A6800086B0,
-    0x800086B8800086B4, 0x000D1DAA000D31A5, 0x800086BC000EC5AB,
-    0x000EDDAE800086C0, 0x000EDDAC000EC5AF, 0x000FDDA8000ED5AD,
-    0x000FF5AA000FFDA9, 0x000E25AC000E3DAB, 0x000E9DAE000E95AD,
-    0x800086C4000E91AF, 0x800086CC800086C8, 0x800086D4800086D0,
-    0x800086DC800086D8, 0x000EA5B8800086E0, 0x000EB9BA000EA9B9,
-    0x000EADBC000EB9BB, 0x000155BE00015DBD, 0x000EF9B000014DBF,
-    0x000ECDB2000EC5B1, 0x000EBDB4000EA9B3, 0x000EA1B6000EA1B5,
-    0x000F7DB3000EA1B7, 0x800086E8800086E4, 0x800086F0800086EC,
-    0x000EADB5000E15B6, 0x000E01BB800086F4, 0x800086F8000E3DBA,
-    0x000E65BF800086FC, 0x000E7DBD000E79BE, 0x80008700000E19BC,
-    0x80008704000F39A3, 0x000E51A680008708, 0x800087108000870C,
-    0x000E79AA000EE9A5, 0x0001A484000E45AB, 0x000E3DAE0001A4BE,
-    0x000E5DAC000E21AF, 0x0002C1A8000E39AD, 0x0002C5AA0002CDA9,
-    0x0002C9AC0002D9AB, 0x0002F5AE0002FDAD, 0x0003ED8000033DAF,
-    0x0003F1820003F181, 0x0004908680008718, 0x8000871C00037487,
-    0x0003C9B880008720, 0x0003D9BA0003C9B9, 0x0003F9BC0003D1BB,
-    0x000399BE0003F9BD, 0x000345B0000395BF, 0x000345B200034DB1,
-    0x000345B400035DB3, 0x000345B600034DB5, 0x00024DB30003F9B7,
-    0x8000872880008724, 0x8000872C00030C84, 0x000241B500027DB6,
-    0x000389BB80008730, 0x80008734000381BA, 0x000389BF80008738,
-    0x000389BD000381BE, 0x8000873C000391BC, 0x80008740000209A3,
-    0x000239A680008744, 0x8000874C80008748, 0x0003C5AA000205A5,
-    0x800087500003CDAB, 0x0003C5AE80008754, 0x0003D5AC0003CDAF,
-    0x800087580003CDAD, 0x800087608000875C, 0x8000876880008764,
-    0x8000876C0001A4E1, 0x800087700000E4E3, 0x8000877880008774,
-    0x8000877C000038EF, 0x8000878880008780, 0x00006980000530BE,
-    0x00007D8200007581, 0x8000878C000171B3, 0x0001C9B60001DDB5,
-    0x0004C08680008790, 0x0001C5BA0005E887, 0x0001B9BC0001ADBB,
-    0x0001ADBE0001B9BD, 0x000EA4E10001A5BF, 0x000104E3000C44E1,
-    0x80008794000EB0E3, 0x8000879C80008798, 0x0005A884800087A0,
-    0x800087A8800087A4, 0x800087B0800087AC, 0x000ED4EF800087B4,
-    0x000F01AB000F9CEF, 0x800087B8000F69AA, 0x000F01AF800087BC,
-    0x000F11AD000F19AE, 0x000F65A3000F19AC, 0x800087C4800087C0,
-    0x800087CC800087C8, 0x000F75A5000F65A6, 0x00062DA8800087D0,
-    0x000635AA00063DA9, 0x000699AC000689AB, 0x000685AE000685AD,
-    0x800087840006B5AF, 0x800087D8800087D4, 0x800087E0800087DC,
-    0x800087E8800087E4, 0x000741B8800087EC, 0x000741BA000741B9,
-    0x000741BC000741BB, 0x000779BE00074DBD, 0x0006CDB0000779BF,
-    0x0006A1B20006A5B1, 0x0006A5B40006A1B3, 0x000691B60006ADB5,
-    0x0006BDA3000691B7, 0x0000B58100001582, 0x800087F00000B580,
-    0x000611A5000605A6, 0x000661AB000390BE, 0x800087F4000609AA,
-    0x000669AF800087F8, 0x000675AD000661AE, 0x800087FC000675AC,
-    0x0000C886000799B3, 0x000749B600031C87, 0x8000880480008800,
-    0x000749BA000749B5, 0x80008808000749BB, 0x000739BE8000880C,
-    0x000749BC000739BF, 0x000639A8000749BD, 0x000655AA000655A9,
-    0x000685AC00069DAB, 0x000685AE00068DAD, 0x800088100006B9AF,
-    0x8000881880008814, 0x800088208000881C, 0x8000882880008824,
-    0x0006ADB88000882C, 0x0006B5BA0006B5B9, 0x000179BC000169BB,
-    0x000169BE000179BD, 0x0006C9B0000169BF, 0x0006D9B20006C9B1,
-    0x0006B9B40006D9B3, 0x00069DB60006B9B5, 0x0006DDA3000695B7,
-    0x8000883480008830, 0x8000883C80008838, 0x00060DA500060DA6,
-    0x00060DAB80008714, 0x8000884000060DAA, 0x00067DAF80008844,
-    0x00060DAD00067DAE, 0x8000884800060DAC, 0x800088508000884C,
-    0x00000D8080008854, 0x00003D8200003581, 0x8000885C80008858,
-    0x0001648480008860, 0x001F5086000034BE, 0x8000886800016887,
-    0x800088708000886C, 0x0003048480008874, 0x00068CE3001C00BE,
-    0x000630E180008878, 0x0007E4EF8000887C, 0x8000888480008880,
-    0x8000888C80008888, 0x8000889480008890, 0x000295B380008898,
-    0x800088A08000889C, 0x800088A8800088A4, 0x0002B1B50002B9B6,
-    0x00034DBB800088AC, 0x800088B000034DBA, 0x00034DBF800088B4,
-    0x00034DBD00034DBE, 0x800088B800034DBC, 0x800088C0800088BC,
-    0x800088C4001D88BE, 0x00019D8100019D82, 0x00020DA800019D80,
-    0x00027DAA000265A9, 0x000259AC000275AB, 0x000399AE000259AD,
-    0x0003E9B0000399AF, 0x0003F9B20003E9B1, 0x0003E9B40003F9B3,
-    0x00035DB60003E9B5, 0x00036DB8000355B7, 0x000341BA000375B9,
-    0x000331BC000341BB, 0x000331BE000331BD, 0x800088C8000331BF,
-    0x800088D0800088CC, 0x800088D8001D0484, 0x001CB886000000EF,
-    0x800088DC001D3087, 0x800088E00000DCE3, 0x800088E40001FCE1,
-    0x800088EC800088E8, 0x0003ADA6800088F0, 0x800088F8800088F4,
-    0x800088FC0003A5A5, 0x80008900000381A3, 0x000259AE80008904,
-    0x000259AC000259AF, 0x000259AA000259AD, 0x80008908000259AB,
-    0x800089108000890C, 0x8000891880008914, 0x800089208000891C,
-    0x8000892880008924, 0x001CAC840001F8EF, 0x8000892C0006DCE1,
-    0x8000893000061CE3, 0x8000893880008934, 0x0001C1B38000893C,
-    0x001FEC8680008940, 0x000000F5001CF087, 0x0001CDB50001C9B6,
-    0x0001D1BB80008948, 0x001F2CBE0001C9BA, 0x0001B5BF8000894C,
-    0x0001C1BD0001C9BE, 0x001DF9A80001C9BC, 0x001E09AA001DF9A9,
-    0x001E05AC001E09AB, 0x001E3DAE001E09AD, 0x800088D4001E35AF,
-    0x001F9581001FE182, 0x80008944001F9580, 0x8000895480008950,
-    0x001E11B880008958, 0x001E21BA001E19B9, 0x001ED1BC001E21BB,
-    0x001ED1BE001ED1BD, 0x001E4DB0001ECDBF, 0x001E5DB2001E55B1,
-    0x001E4DB4001E55B3, 0x001E31B6001E31B5, 0x001E81A3001E31B7,
-    0x800089608000895C, 0x8000896880008964, 0x001E8DA5001E89A6,
-    0x001E91AB8000896C, 0x80008970001E89AA, 0x001EF5AF80008974,
-    0x001E81AD001E89AE, 0x80008978001E89AC, 0x8000897C001FEDB3,
-    0x001FE9B680008980, 0x8000898880008984, 0x001FA9BA001FE9B5,
-    0x8000898C001FB5BB, 0x001F89BE80008990, 0x001FA1BC001F89BF,
-    0x001E11A8001FA1BD, 0x001E21AA001E25A9, 0x001E25AC001E3DAB,
-    0x001E25AE001E2DAD, 0x80008994001E59AF, 0x8000899C80008998,
-    0x800089A4800089A0, 0x0000BD810000B982, 0x001EC9B80000BD80,
-    0x001ED9BA001EC9B9, 0x001EC9BC001ED9BB, 0x001EE9BE001EFDBD,
-    0x001E29B0001EE9BF, 0x001E31B2001E35B1, 0x001E09B4001E31B3,
-    0x001EF9B6001E09B5, 0x800089A8001EF9B7, 0x001EA1A3800089AC,
-    0x001EA5A5800089B0, 0x800089B8800089B4, 0x80008864001EA5A6,
-    0x001EF9AB0000A884, 0x001EEDAD001EE5AA, 0x001EC5AF001EEDAC,
-    0x0001F5B3001EC5AE, 0x00006C86800089BC, 0x800089C000017C87,
-    0x0001E5B5000151B6, 0x00017DBB800089C4, 0x800089C8000175BA,
-    0x000139BF800089CC, 0x000155BD000135BE, 0x800089D0000155BC,
-    0x000400BE800089D4, 0x0001C0E100040084, 0x0001ACE3800089D8,
-    0x800089DC001EE8EF, 0x0002D8EF001DCCE1, 0x800089E0001E34E3,
-    0x800089E8800089E4, 0x800089F0800089EC, 0x00027DA3800089F4,
-    0x00026DA5800089F8, 0x80008A00800089FC, 0x80008A040002D9A6,
-    0x0002F5AB80008A08, 0x0002DDAD0002FDAA, 0x0002B1AF0002DDAC,
-    0x001D6DA90002BDAE, 0x0002B1AB001D59A8, 0x000215AD0002B9AA,
-    0x000215AF0002A9AC, 0x00052CBE000215AE, 0x80008A1080008A0C,
-    0x00000D8080008A14, 0x00019D8200019581, 0x000229B980008A18,
-    0x0003E9BB00021DB8, 0x0003F9BD000221BA, 0x0003E9BF0003F1BC,
-    0x00027DB10003F1BE, 0x000245B3000275B0, 0x000239B5000245B2,
-    0x00022DB7000235B4, 0x0002C5A9000231B6, 0x0002D5AB00029DA8,
-    0x0002FDAD0002DDAA, 0x00011DAF0002FDAC, 0x80008A1C00011DAE,
-    0x0005A08780008A20, 0x80008A280005AC86, 0x80008A3080008A2C,
-    0x000115B9000000CF, 0x000129BB000109B8, 0x0001F5BD000115BA,
-    0x0001F5BF000131BC, 0x000141B10001F5BE, 0x00015DB3000171B0,
-    0x00014DB500015DB2, 0x00013DB700014DB4, 0x80008A3400013DB6,
-    0x80008A3C80008A38, 0x80008A40001E9CEF, 0x80008A4880008A44,
-    0x00076C84001FE8EF, 0x80008A4C001EA4E1, 0x80008A50001EB0E3,
-    0x80008A54001E64E1, 0x000211B3001E0CE3, 0x80008A5C80008A58,
-    0x00065C8480008A60, 0x000231B5000239B6, 0x0001C1BB80008A64,
-    0x80008A680001F1BA, 0x0001CDBF80008A6C, 0x0001CDBD0001CDBE,
-    0x80008A240001CDBC, 0x80008A7480008A70, 0x80008A7C80008A78,
-    0x80008A8480008A80, 0x000711A880008A88, 0x000711AA000711A9,
-    0x0007B5AC0007ADAB, 0x0007A9AE0007BDAD, 0x0007D9B00007A9AF,
-    0x0007F9B20007EDB1, 0x000795B40007F5B3, 0x000795B600079DB5,
-    0x000789B8000781B7, 0x00006DBA000789B9, 0x00007DBC000065BB,
-    0x00006DBE000065BD, 0x00001982000065BF, 0x00009580000555A3,
-    0x00057DA600001981, 0x80008A9080008A8C, 0x0006B5AA000575A5,
-    0x80008A94000685AB, 0x000689AE80008A98, 0x000689AC000689AF,
-    0x0007F9B3000689AD, 0x0003E88680008A9C, 0x80008AA00003BC87,
-    0x00075DB500075DB6, 0x000749BB80008AA4, 0x80008AA800077DBA,
-    0x000739BF80008AAC, 0x000741BD000731BE, 0x000639A300074DBC,
-    0x80008AB480008AB0, 0x80008ABC80008AB8, 0x00069DA500069DA6,
-    0x000689AB80008AC0, 0x80008AC40006BDAA, 0x0006F9AF80008AC8,
-    0x000681AD0006F1AE, 0x80008ACC00068DAC, 0x80008AD480008AD0,
-    0x80008ADC80008AD8, 0x80008AE480008AE0, 0x80008AEC80008AE8,
-    0x80008AF480008AF0, 0x00001D8080008AF8, 0x0000D98200000981,
-    0x00012DA880008AFC, 0x000159AA000141A9, 0x000179AC000151AB,
-    0x00009DAE000179AD, 0x80008B00000095AF, 0x0001888600016CBE,
-    0x80008B080000AC87, 0x80008B1080008B0C, 0x00008DB880008B14,
-    0x00009DBA000095B9, 0x00008DBC000095BB, 0x0000BDBE0000B5BD,
-    0x0000EDB00000B5BF, 0x0000F5B20000F5B1, 0x0000DDB40000C5B3,
-    0x0000BDB60000B5B5, 0x0002D1B30000B5B7, 0x000370BE80008B18,
-    0x80008B1C000C44BE, 0x0002F1B50002F9B6, 0x000331BB80008B20,
-    0x80008B24000329BA, 0x000371BF80008B28, 0x000311BD000319BE,
-    0x80008B2C000321BC, 0x80008B30000295A3, 0x0002BDA680008B34,
-    0x80008B3C80008B38, 0x00036DAA0002B5A5, 0x80008B40000375AB,
-    0x00035DAE80008B44, 0x000365AC000335AF, 0x000331A8000355AD,
-    0x000331AA000331A9, 0x000095AC000331AB, 0x000085AE000085AD,
-    0x80008B480000B5AF, 0x80008B5080008B4C, 0x80008B5880008B54,
-    0x80008B6080008B5C, 0x000061B880008B64, 0x000061BA000061B9,
-    0x000061BC000061BB, 0x000061BE000061BD, 0x0000CDB0000061BF,
-    0x0000A1B20000A5B1, 0x0000A5B40000A1B3, 0x000091B60000ADB5,
-    0x00000D80000091B7, 0x0000118200001581, 0x80008B6C80008B68,
-    0x80008B70000048EF, 0x000D6C8480008B78, 0x80008B7C000694E3,
-    0x80008B800001FCE1, 0x000DB88680008B84, 0x0002F9B30002C487,
-    0x80008B8C80008B88, 0x80008B9480008B90, 0x00015DB5000145B6,
-    0x000121BB000F98BE, 0x80008B98000149BA, 0x000129BF80008B9C,
-    0x000135BD000121BE, 0x80008BA0000135BC, 0x80008BA40006B0E1,
-    0x80008BA8000790E3, 0x80008BB080008BAC, 0x80008BB880008BB4,
-    0x80008BC080008BBC, 0x000000E580008BC4, 0x80008BCC80008BC8,
-    0x000195AE0007E4EF, 0x000181AC00019DAF, 0x0001FDAA000181AD,
-    0x80008BD0000195AB, 0x0001F1A680008BD4, 0x80008BDC80008BD8,
-    0x000FF5820001E9A5, 0x00001D8000024DA3, 0x80008B7400001581,
-    0x80008BE880008BE0, 0x0003B08680008BE4, 0x80008BEC00039487,
-    0x000F3DA880008BF0, 0x000F19AA000F0DA9, 0x000F75AC000F19AB,
-    0x000F75AE000F7DAD, 0x000F29B0000F6DAF, 0x000FC9B2000F29B1,
-    0x000FC1B4000FC5B3, 0x000FC1B6000FC1B5, 0x000FC1B8000FC1B7,
-    0x000FC1BA000FC1B9, 0x000FC5BC000FC1BB, 0x000FA1BE000FCDBD,
-    0x000FB1B3000FA1BF, 0x80008BF880008BF4, 0x80008C0080008BFC,
-    0x000FDDB5000FDDB6, 0x000FCDBB80008C04, 0x80008C08000FCDBA,
-    0x000F25BF80008C0C, 0x000F31BD000F25BE, 0x80008C10000F3DBC,
-    0x80008C14000FF5A3, 0x000F99A680008C18, 0x80008C2080008C1C,
-    0x000F89AA000F99A5, 0x80008C24000F89AB, 0x000F61AE80008C28,
-    0x000F79AC000F61AF, 0x000109A8000F75AD, 0x000115AA00011DA9,
-    0x000199AC000189AB, 0x000189AE000199AD, 0x80008C2C000181AF,
-    0x80008C3480008C30, 0x80008C3C80008C38, 0x80008C4480008C40,
-    0x0001ADB880008C48, 0x00006DBA000065B9, 0x00007DBC000065BB,
-    0x00006DBE000065BD, 0x0001C9B0000065BF, 0x0001A9B20001C9B1,
-    0x0001BDB40001A5B3, 0x0001A1B60001A1B5, 0x0000BD80000195B7,
-    0x0000198200000981, 0x80008C5080008C4C, 0x80008C5400029884,
-    0x0001808680008C58, 0x80008C5C00010487, 0x80008C6480008C60,
-    0x80008B0480008C68, 0x0001DDB380008C6C, 0x80008C7480008C70,
-    0x80008C7C80008C78, 0x000151B5000159B6, 0x00014DBB80008C80,
-    0x80008C8400014DBA, 0x000125BF80008C88, 0x000131BD000139BE,
-    0x80008C8C00013DBC, 0x0003D0BE000EF4EF, 0x80008C9480008C90,
-    0x80008C98000248E3, 0x80008C9C0001A4E1, 0x80008CA00002D4EF,
-    0x80008CA880008CA4, 0x80008CAC000BF4E3, 0x80008CB0000D8CE1,
-    0x000255A380008CB4, 0x0002D9A500038C84, 0x80008CBC80008CB8,
-    0x80008CC00002D1A6, 0x0002C5AB80008CC8, 0x0002B9AD0002C5AA,
-    0x0002ADAF0002B5AC, 0x00031DA80002B1AE, 0x000325AA00032DA9,
-    0x000325AC00033DAB, 0x00035DAE000329AD, 0x0000F580000355AF,
-    0x0000C9820000FD81, 0x80008CD080008CCC, 0x00036C8780008CD4,
-    0x0000D9B800041C86, 0x0000E1BA0000E5B9, 0x0000E1BC0000E1BB,
-    0x000099BE0000EDBD, 0x00032DB0000099BF, 0x000335B2000335B1,
-    0x00031DB4000305B3, 0x0000F5B60000F5B5, 0x80008CD80000EDB7,
-    0x80008CE080008CDC, 0x80008CE40002A1B3, 0x0002A5B60002A1B5,
-    0x80008CEC80008CE8, 0x0002E5BA80008CF0, 0x0002EDBC0002FDBB,
-    0x0002D5BE0002DDBD, 0x0002FDA600013DBF, 0x80008CF880008CF4,
-    0x80008CFC0002F9A5, 0x80008D000002F9A3, 0x00028DAE80008D04,
-    0x0002B5AC000165AF, 0x0002BDAA000285AD, 0x80008D080002A5AB,
-    0x000004E380008D0C, 0x000180E1000EA4E1, 0x80008D10000EBCE3,
-    0x80008D1880008D14, 0x0005E8BE80008D1C, 0x80008D2480008D20,
-    0x000E14EF80008D28, 0x0004F0BE80008D2C, 0x80008D30000F48EF,
-    0x80008D3C80008D38, 0x0000658180008D40, 0x80008D4400007980,
-    0x0005AC8400001D82, 0x0004C88780008D48, 0x80008D4C00044C86,
-    0x80008D5480008D50, 0x000665A880008D58, 0x0006B9AA0006A9A9,
-    0x000711AC0006B9AB, 0x000711AE000711AD, 0x80008D3400070DAF,
-    0x80008D6080008D5C, 0x80008D6880008D64, 0x80008D7080008D6C,
-    0x000739B880008D74, 0x0007D9BA000739B9, 0x0007C9BC0007D9BB,
-    0x0007FDBE0007C9BD, 0x00077DB00007F5BF, 0x00074DB2000745B1,
-    0x000739B4000729B3, 0x000721B6000725B5, 0x000631B3000721B7,
-    0x80008D7C80008D78, 0x80008D8480008D80, 0x000799B5000791B6,
-    0x000781BB80008D88, 0x80008D8C0007B9BA, 0x0007E1BF80008D90,
-    0x0007FDBD0007FDBE, 0x80008D94000799BC, 0x80008D98000675A3,
-    0x0007D5A680008D9C, 0x0003948480008DA0, 0x0007FDAA0007DDA5,
-    0x80008DA40007C5AB, 0x0007B9AE80008DA8, 0x0007DDAC0007A5AF,
-    0x0006C5A80007B9AD, 0x0006C5AA0006CDA9, 0x0006F5AC0006D5AB,
-    0x0006EDAE0006E5AD, 0x0001E9800006E5AF, 0x0001FD820001E981,
-    0x0001308680008DAC, 0x80008DB000016487, 0x0001C5B880008DB4,
-    0x0001C5BA0001CDB9, 0x0001CDBC0001DDBB, 0x0001F5BE0001FDBD,
-    0x00069DB000019DBF, 0x000141B2000141B1, 0x000141B4000141B3,
-    0x000141B6000141B5, 0x00064DB3000141B7, 0x80008DBC80008DB8,
-    0x80008DC480008DC0, 0x00067DB5000665B6, 0x000185BB80008DC8,
-    0x80008DCC000659BA, 0x000185BF80008DD0, 0x000195BD000185BE,
-    0x80008DD4000195BC, 0x80008CC4000609A3, 0x000621A680008DD8,
-    0x80008DE080008DDC, 0x00061DAA000639A5, 0x80008DE40001C1AB,
-    0x0001C1AE80008DE8, 0x0001D1AC0001C1AF, 0x80008DEC0001D1AD,
-    0x80008DF480008DF0, 0x80008DFC80008DF8, 0x80008E0480008E00,
-    0x80008E0C80008E08, 0x000364BE80008E10, 0x80008E1C80008E18,
-    0x003D30BF80008E20, 0x003C008580008E24, 0x00038C840007B4E3,
-    0x00004D800006E8E1, 0x00007D8200007581, 0x003C4485002A8CEF,
-    0x80008E28000210E1, 0x80008E2C002EA8E3, 0x003D588600072CEF,
-    0x80008E3000030C87, 0x80008E3880008E34, 0x80008E4080008E3C,
-    0x80008E4880008E44, 0x00024DA880008E4C, 0x000391AA000389A9,
-    0x000335AC000391AB, 0x000335AE00033DAD, 0x00035DB000032DAF,
-    0x000365B200036DB1, 0x00031DB400030DB3, 0x000301B6000301B5,
-    0x00030DB8000335B7, 0x0000EDBA0000E5B9, 0x0000FDBC0000E5BB,
-    0x0000EDBE0000E5BD, 0x80008E500000E5BF, 0x000231B380008E54,
-    0x000295B580008E58, 0x80008E6080008E5C, 0x003DBC84000295B6,
-    0x000281BB80008E64, 0x0002F1BD0002B1BA, 0x0002EDBF000299BC,
-    0x80008E680002F1BE, 0x003FC4BE80008E6C, 0x80008E70003FC484,
-    0x80008E74000384EF, 0x80008E7C80008E78, 0x80008E8000033CE3,
-    0x00002D80000044E1, 0x0000118200001581, 0x003FA08680008E88,
-    0x80008E8C003D5887, 0x80008E9480008E90, 0x80008E9C80008E98,
-    0x003E2CBE0006DCEF, 0x80008EA0000650E1, 0x80008EA4000164E3,
-    0x80008EAC80008EA8, 0x0002F9A380008EB0, 0x80008EB880008EB4,
-    0x80008EC080008EBC, 0x00025DA500025DA6, 0x000249AB80008EC4,
-    0x80008EC8000279AA, 0x000225AF80008ECC, 0x000239AD000239AE,
-    0x00246D90000251AC, 0x00280D9200280591, 0x002D8D9400289D93,
-    0x002CB196002C3595, 0x003001980031BD97, 0x0031BD9A00301599,
-    0x0034999C00342D9B, 0x0038399E00398D9D, 0x80008E140038319F,
-    0x003E21B5003E35B6, 0x003E65B380008E84, 0x80008ED480008ED0,
-    0x003EFDBF80008ED8, 0x003EFDBD003EFDBE, 0x003EFDBB003EFDBC,
-    0x80008EDC003EFDBA, 0x000945A90024398F, 0x000CA9AB00089DA8,
-    0x001001AD000C01AA, 0x001035AF000D1DAC, 0x0000E1A10010BDAE,
-    0x000401A380008EE0, 0x000495A50000E1A2, 0x0008B9A7000495A4,
-    0x003959A0000801A6, 0x003CE9A2003D0DA1, 0x80008EE4003C25A3,
-    0x000025810001DD82, 0x0014C1B100002580, 0x001801B3001401B0,
-    0x0018D5B50014D5B2, 0x80008EE80018D5B4, 0x0025458280008EEC,
-    0x80008EF000256983, 0x0029818680008EF8, 0x00291984002C8987,
-    0x002D018A00291985, 0x00078886002D8D8B, 0x0031018E00018C87,
-    0x0030998C0030A18F, 0x0035419200311D8D, 0x80008EFC0035ED93,
-    0x00395D9680008F00, 0x0038E1940038CD97, 0x003CA99A00395D95,
-    0x80008F04003D459B, 0x80008F0C80008F08, 0x003D919C80008F10,
-    0x003FADA880008F14, 0x003FBDAA003FB5A9, 0x003FF9AC003FB1AB,
-    0x003F19AE003FE5AD, 0x80008F18003F19AF, 0x80008F2080008F1C,
-    0x80008F2880008F24, 0x80008F3080008F2C, 0x003ECDB880008F34,
-    0x003ED1BA003ED1B9, 0x003E91BC003EE9BB, 0x003E89BE003E91BD,
-    0x003F69B0003E85BF, 0x003F79B2003F69B1, 0x003F69B4003F79B3,
-    0x003EF9B6003F69B5, 0x003C35A3003EF9B7, 0x80008F3C80008F38,
-    0x80008F4480008F40, 0x003C71A5003C65A6, 0x003CADAB80008F48,
-    0x80008F4C003CADAA, 0x003CADAF80008F50, 0x003CADAD003CADAE,
-    0x0003C884003CADAC, 0x80008F54003DC9B3, 0x003DC9B680008F58,
-    0x80008F6080008F5C, 0x003DD5BA003DC9B5, 0x00036CBE003DDDBB,
-    0x003DB9BE80008F68, 0x003DC5BC003DB9BF, 0x003DFDA8003DB9BD,
-    0x003D05AA003D0DA9, 0x003D05AC003D1DAB, 0x003D31AE003D0DAD,
-    0x0000ED80003D2DAF, 0x0000F1820000F181, 0x0004908680008F6C,
-    0x80008F7000036487, 0x003D29B880008F74, 0x00029DBA003D29B9,
-    0x00028DBC000295BB, 0x00027DBE0002ADBD, 0x003D55B000020DBF,
-    0x003D55B2003D5DB1, 0x003D39B4003D29B3, 0x003D29B6003D39B5,
-    0x003C8DA3003D21B7, 0x0004C48480008F78, 0x80008F8080008F7C,
-    0x003C8DA5003C8DA6, 0x003C99AB80008F84, 0x80008F88003C91AA,
-    0x003CFDAF80008F8C, 0x003CFDAD003CFDAE, 0x80008F90003C81AC,
-    0x80008F94000209B3, 0x000221B680008F98, 0x80008FA080008F9C,
-    0x0002C1BA000231B5, 0x80008FA40002C1BB, 0x0002C1BE80008FA8,
-    0x0002D1BC0002C1BF, 0x80008FAC0002D1BD, 0x0002C5A380008FB0,
-    0x0002FDA580008FB4, 0x80008FBC80008FB8, 0x80008FC00002EDA6,
-    0x00020DAB80008FC4, 0x00021DAD00020DAA, 0x00020DAF00021DAC,
-    0x003E9CE300020DAE, 0x003E50E1003FA4E1, 0x00000D80003EFCE3,
-    0x00003D8200003581, 0x80008FCC80008FC8, 0x000420BE80008FD0,
-    0x003EA8EF80008FD8, 0x80008FE080008FDC, 0x00050DA8003E4CEF,
-    0x000545AA00054DA9, 0x00057DAC000565AB, 0x00056DAE000565AD,
-    0x80008F64000561AF, 0x0003088680008FD4, 0x80008FE400038C87,
-    0x80008FEC80008FE8, 0x000645B880008FF0, 0x000645BA00064DB9,
-    0x00064DBC00065DBB, 0x000675BE00067DBD, 0x0005C1B0000615BF,
-    0x0005C1B20005C1B1, 0x0005C1B40005C1B3, 0x0005C1B60005C1B5,
-    0x0006EDA80005C1B7, 0x000631AA000629A9, 0x000795AC000631AB,
-    0x000789AE000799AD, 0x80008FF4000781AF, 0x80008EF480008FF8,
-    0x8000900080008FFC, 0x8000900880009004, 0x0007ADB88000900C,
-    0x00074DBA000745B9, 0x000745BC000741BB, 0x000779BE00074DBD,
-    0x0007C9B0000779BF, 0x0007A9B20007C9B1, 0x0007BDB40007A5B3,
-    0x0007A1B60007A1B5, 0x0007B9B3000795B7, 0x8000901480009010,
-    0x8000901C80009018, 0x00061DB5000619B6, 0x00060DBB80009020,
-    0x8000902400060DBA, 0x000665BF80009028, 0x000671BD000679BE,
-    0x8000902C00067DBC, 0x800090300007FDA3, 0x00065DA680009034,
-    0x8000903C80009038, 0x000649AA000659A5, 0x80009040000649AB,
-    0x00063DAE0001A084, 0x000639AC000621AF, 0x00066DA8000635AD,
-    0x000671AA000671A9, 0x0001A5AC0001BDAB, 0x0001A5AE0001ADAD,
-    0x0000F58000013DAF, 0x0000C1820000FD81, 0x00019886000164BE,
-    0x800090480000AC87, 0x0000CDB88000904C, 0x0000D5BA0000D5B9,
-    0x0000FDBC0000E5BB, 0x000091BE000095BD, 0x000145B0000091BF,
-    0x000159B2000149B1, 0x000149B4000159B3, 0x0000FDB6000149B5,
-    0x0005F9B30000F5B7, 0x8000905480009050, 0x8000905C80009058,
-    0x00022DB500022DB6, 0x00028DBB80009060, 0x80009064000281BA,
-    0x00028DBF80009068, 0x00028DBD00028DBE, 0x8000906C00028DBC,
-    0x800090700005BDA3, 0x000269A680009074, 0x8000907C80009078,
-    0x0002C5AA000269A5, 0x800090800002C9AB, 0x0002C9AE80009084,
-    0x0002C9AC0002C9AF, 0x000085AA0002C9AD, 0x000331A800008DAB,
-    0x00008DAE000331A9, 0x000095AC0000BDAF, 0x0003E4BE00008DAD,
-    0x8000908C80009088, 0x8000909480009090, 0x8000909C80009098,
-    0x00006DBA800090A0, 0x00006DB8000075BB, 0x00006DBE000065B9,
-    0x00006DBC0001D5BF, 0x0000A9B2000065BD, 0x0000C5B00000A9B3,
-    0x000099B60000ADB1, 0x0000ADB4000099B7, 0x800090A40000A5B5,
-    0x800090AC800090A8, 0x00001D82800090B0, 0x00001D80000784EF,
-    0x0002AC8400001581, 0x800090B8000774E3, 0x800090BC00041CE1,
-    0x000DB886800090C0, 0x000145B3000D3087, 0x800090C8800090C4,
-    0x800090D0800090CC, 0x000155B5000145B6, 0x000121BB800090D4,
-    0x800090D8000149BA, 0x000129BF800090DC, 0x000135BD000121BE,
-    0x800090E0000135BC, 0x800090E8800090E4, 0x800090F0800090EC,
-    0x0000CCEF800090F4, 0x000190E100068CEF, 0x0000D4E300070CE1,
-    0x800090F80006B4E3, 0x80009100800090FC, 0x8000910880009104,
-    0x0001CDA38000910C, 0x0001DDA5000D9084, 0x8000911480009110,
-    0x800091180001CDA6, 0x0001A9AB8000911C, 0x0001BDAD0001C1AA,
-    0x0001A1AF0001BDAC, 0x800090B40001A9AE, 0x000FED8100001582,
-    0x80009120000FED80, 0x80009124000258BE, 0x000F09A880009128,
-    0x000F1DAA000F15A9, 0x000F79AC000F69AB, 0x000F69AE000F79AD,
-    0x000F29B0000F61AF, 0x000FC9B2000F29B1, 0x000FD9B4000FC9B3,
-    0x000FC9B6000FD9B5, 0x000FD1B8000FC9B7, 0x000FD1BA000FD1B9,
-    0x000FB1BC000FD1BB, 0x000FADBE000FB1BD, 0x8000912C000FA5BF,
-    0x000FC1B5000FC9B6, 0x000FA5B380009130, 0x0003088780009134,
-    0x000F2DBF00033C86, 0x000F2DBD000F2DBE, 0x000FDDBB000F2DBC,
-    0x80009138000FDDBA, 0x000FF5A380009044, 0x800091408000913C,
-    0x8000914880009144, 0x000F91A5000F99A6, 0x000F8DAB8000914C,
-    0x80009150000F8DAA, 0x000F7DAF80009154, 0x000F7DAD000F7DAE,
-    0x00011DA8000F7DAC, 0x000185AA000185A9, 0x0001A9AC000199AB,
-    0x0001D1AE0001D1AD, 0x800091580001D1AF, 0x800091608000915C,
-    0x8000916880009164, 0x800091708000916C, 0x000071B880009174,
-    0x000071BA000071B9, 0x0000D1BC000071BB, 0x0000D1BE0000D1BD,
-    0x0001B1B00000D1BF, 0x000181B20001B9B1, 0x000051B4000181B3,
-    0x000051B6000051B5, 0x80009178000051B7, 0x800091808000917C,
-    0x00000D8080009184, 0x00003D8200003581, 0x8000918C80009188,
-    0x0002A48480009190, 0x00019086000124BE, 0x8000919800011487,
-    0x00014DB38000919C, 0x800091A080009194, 0x800091A8800091A4,
-    0x000161B5000111B6, 0x00013DBB800091AC, 0x800091B0000135BA,
-    0x0000D5BF800091B4, 0x000115BD0000D5BE, 0x800091B8000115BC,
-    0x800091BC000289A3, 0x0002D5A6000340BE, 0x800091C4800091C0,
-    0x0002F1AA0002A5A5, 0x800091C80002F9AB, 0x000311AE800091CC,
-    0x0002D1AC000311AF, 0x0001E8E10002D1AD, 0x000214E3800091D0,
-    0x800091D40003D484, 0x800091DC800091D8, 0x800091E000041084,
-    0x800091E8800091E4, 0x800091F0800091EC, 0x00028CEF800091F4,
-    0x000339A8800091F8, 0x000341AA000339A9, 0x000341AC000341AB,
-    0x000371AE000349AD, 0x0000DD80000371AF, 0x0000198200000981,
-    0x80009200800091FC, 0x0003088780009208, 0x000095B800041C86,
-    0x0000A9BA00009DB9, 0x0000B9BC0000A9BB, 0x0000A9BE0000B9BD,
-    0x0000E1B00000A9BF, 0x0000FDB20000EDB1, 0x0000D9B40000F5B3,
-    0x0000BDB60000D9B5, 0x000F58EF0000ADB7, 0x800092108000920C,
-    0x80009214000544BE, 0x8000921C80009218, 0x8000922480009220,
-    0x8000922C80009228, 0x000E00E380009230, 0x000C90E180009234,
-    0x00029DB680009238, 0x800092408000923C, 0x8000924400029DB5,
-    0x8000924800029DB3, 0x000149BE8000924C, 0x00017DBC000149BF,
-    0x000171BA000175BD, 0x8000925000017DBB, 0x8000925880009254,
-    0x8000925C000EA4E1, 0x0004DCBE000E50E3, 0x8000926000002CE3,
-    0x0006ECBE000118E1, 0x0019E0BE0000F0EF, 0x0000758080009264,
-    0x0000758200007D81, 0x000209A3000150EF, 0x8000926C80009268,
-    0x8000927880009270, 0x000209A5000209A6, 0x0001E9AB8000927C,
-    0x000448860001E5AA, 0x0001DDAF0004DC87, 0x0001E1AD0001DDAE,
-    0x0006B5A80001E9AC, 0x000701AA000701A9, 0x000705AC00071DAB,
-    0x000705AE00070DAD, 0x8000927400073DAF, 0x8000928480009280,
-    0x8000928C80009288, 0x8000929480009290, 0x000719B880009298,
-    0x000731BA000719B9, 0x0007D1BC000731BB, 0x0007C1BE0007DDBD,
-    0x000745B00007F1BF, 0x000745B200074DB1, 0x00074DB4000759B3,
-    0x000731B6000731B5, 0x000789B300072DB7, 0x800092A08000929C,
-    0x800092A480009204, 0x00078DB500078DB6, 0x000789BB800092A8,
-    0x800092AC000781BA, 0x0007F5BF800092B0, 0x00078DBD000789BE,
-    0x800092B4000791BC, 0x800092B80007CDA3, 0x0007C9A6800092BC,
-    0x800092C4800092C0, 0x0007C5AA0007C9A5, 0x800092C80007CDAB,
-    0x0007CDAE800092CC, 0x0007D5AC0007B1AF, 0x0006D1A80007C9AD,
-    0x00012DAA0006D9A9, 0x00013DAC000125AB, 0x00012DAE000125AD,
-    0x800092D0000125AF, 0x800092D8800092D4, 0x00001980800092DC,
-    0x0000058200001981, 0x000189B8800092E0, 0x00019DBA000189B9,
-    0x0001B9BC000195BB, 0x000079BE0001B9BD, 0x00015DB0000079BF,
-    0x0001E1B20001E1B1, 0x0001EDB40001FDB3, 0x0001D5B60001DDB5,
-    0x0002C8840001BDB7, 0x800092E40001A1B3, 0x0001A1B6800092E8,
-    0x800092F0800092EC, 0x00019DBA0001B1B5, 0x0001A886000121BB,
-    0x00012DBE0001E087, 0x000131BC00011DBF, 0x0001E5A300012DBD,
-    0x800092F8800092F4, 0x80009300800092FC, 0x0001F5A50001E5A6,
-    0x000165AB80009304, 0x800093080001D9AA, 0x000159AF8000930C,
-    0x000169AD000169AE, 0x80009310000175AC, 0x80009314000274E3,
-    0x800093180001BCE1, 0x800093208000931C, 0x8000932880009324,
-    0x800093308000932C, 0x8000933880009334, 0x8000933C00076CEF,
-    0x8000934480009340, 0x00001D8100001D82, 0x8000934800001D80,
-    0x800093508000934C, 0x0003E48480009358, 0x00032C878000935C,
-    0x80009360001CCC86, 0x8000936880009364, 0x800093708000936C,
-    0x8000937880009374, 0x800093808000937C, 0x8000938880009384,
-    0x00032DA88000938C, 0x000331AA000331A9, 0x000379AC000369AB,
-    0x000369AE000379AD, 0x000329B0000361AF, 0x0000C9B2000329B1,
-    0x0000C1B40000C5B3, 0x0000C1B60000C1B5, 0x0000C1B80000C1B7,
-    0x0000C1BA0000C1B9, 0x0000C1BC0000C1BB, 0x0000C5BE0000CDBD,
-    0x00074CE30000BDBF, 0x0007ACE180009390, 0x8000939880009394,
-    0x800093A08000939C, 0x800093A8800093A4, 0x800093B0800093AC,
-    0x0006C4EF800093B4, 0x800093BC800093B8, 0x00007981800093C0,
-    0x0002D1B300007180, 0x000299B500007182, 0x001D1484800093C8,
-    0x001CC086000291B6, 0x000285BB001DE887, 0x000179BD000285BA,
-    0x00016DBF000175BC, 0x800093CC00016DBE, 0x800093D4800093D0,
-    0x001CDC84800093D8, 0x800093E0800093DC, 0x001CE8BE0006F4EF,
-    0x800093E4000654E1, 0x001FF08400003CE3, 0x800093EC800093E8,
-    0x000259A3800093F0, 0x800093F8800093F4, 0x80009400800093FC,
-    0x000211A5000219A6, 0x00020DAB80009404, 0x8000940800020DAA,
-    0x0001E5AF8000940C, 0x0001F1AD0001E5AE, 0x001E5DA80001FDAC,
-    0x001E79AA001E61A9, 0x001E69AC001E79AB, 0x001E99AE001E69AD,
-    0x800093C4001E99AF, 0x8000941480009410, 0x8000941C80009418,
-    0x8000942480009420, 0x001F51B880009428, 0x001F51BA001F51B9,
-    0x001F71BC001F51BB, 0x001F71BE001F71BD, 0x001EE9B0001F71BF,
-    0x001EF9B2001EE9B1, 0x001EE9B4001EF9B3, 0x001F7DB6001EE9B5,
-    0x001EF1B3001F71B7, 0x0000158100002D82, 0x800093540000B180,
-    0x001EE1B5001EE9B6, 0x001E2DBB8000942C, 0x80009430001E2DBA,
-    0x001E2DBF80009434, 0x001E2DBD001E2DBE, 0x80009438001E2DBC,
-    0x00006886001EB5A3, 0x001EADA600004C87, 0x800094408000943C,
-    0x001E69AA001EA5A5, 0x80009444001E69AB, 0x001E69AE80009448,
-    0x001E69AC001E69AF, 0x001E55A8001E69AD, 0x001E8DAA001E85A9,
-    0x001E9DAC001E85AB, 0x001E81AE001E81AD, 0x8000944C001EB5AF,
-    0x8000945480009450, 0x8000945C80009458, 0x8000946480009460,
-    0x001EA9B880009468, 0x00016DBA001EA9B9, 0x00017DBC000165BB,
-    0x00016DBE000165BD, 0x001ED5B0000165BF, 0x001ED5B2001EDDB1,
-    0x001EB9B4001EA9B3, 0x001EA9B6001EB9B5, 0x001E05B3001EA1B7,
-    0x800094708000946C, 0x8000947880009474, 0x001E15B5001E01B6,
-    0x001E19BB8000947C, 0x80009480001E11BA, 0x001E7DBF80009484,
-    0x001E7DBD001E7DBE, 0x80009488001E01BC, 0x8000948C001E41A3,
-    0x001E45A680009490, 0x8000949880009494, 0x001E55AA001E51A5,
-    0x8000949C001E5DAB, 0x001E39AE800094A0, 0x001E45AC001E39AF,
-    0x0002C1A8001E39AD, 0x0002D1AA0002D1A9, 0x0002FDAC0002E5AB,
-    0x00033DAE000335AD, 0x800094A8000335AF, 0x800094AC000268BE,
-    0x800094B4800094B0, 0x800094BC800094B8, 0x0003D1B8800094C0,
-    0x0003E1BA0003D9B9, 0x000391BC0003E1BB, 0x000395BE00039DBD,
-    0x00034DB0000389BF, 0x00035DB2000355B1, 0x00034DB4000355B3,
-    0x0003F1B60003F1B5, 0x000359800003F1B7, 0x0003658200036D81,
-    0x800094C4000255B3, 0x0003B9B6000279B5, 0x00040086800094C8,
-    0x000385BA0003E487, 0x000385BC00039DBB, 0x000385BE00038DBD,
-    0x000219A3000339BF, 0x800094D0800094CC, 0x000BAC85800094D4,
-    0x000235A50003F5A6, 0x0003D1AB800094D8, 0x800094DC0003C9AA,
-    0x000375AF800094E0, 0x0003C1AD0003C9AE, 0x800094E40003C9AC,
-    0x800094EC800094E8, 0x0003CCEF800094F0, 0x800094F4001F60E1,
-    0x001E80E3001E54E3, 0x0001F8E1800094F8, 0x800094FC001A7CEF,
-    0x8000950480009500, 0x00007D8080009508, 0x00001D8200001581,
-    0x8000950C00024084, 0x000554BE80009510, 0x0004C08680009518,
-    0x8000951C00054887, 0x8000952480009520, 0x8000952C80009528,
-    0x8000953480009530, 0x8000953C80009538, 0x8000954480009540,
-    0x8000954C80009548, 0x0005A484001EC4EF, 0x80009550001EDCE1,
-    0x80009554001E2CE3, 0x8000955C80009558, 0x0001EDB380009560,
-    0x8000956880009564, 0x800095708000956C, 0x0001EDB50001EDB6,
-    0x000125BB80009574, 0x80009578000179BA, 0x000119BF8000957C,
-    0x000135BD000129BE, 0x000621A8000135BC, 0x000625AA00062DA9,
-    0x000625AC00063DAB, 0x000625AE00062DAD, 0x8000951400068DAF,
-    0x0007C58100001582, 0x800095800007C580, 0x80009584000338BE,
-    0x000745B880009588, 0x000745BA00074DB9, 0x000775BC000755BB,
-    0x00076DBE000765BD, 0x0006FDB0000765BF, 0x0006A1B20006A1B1,
-    0x0006ADB40006A1B3, 0x000695B600069DB5, 0x0006ADA300077DB7,
-    0x000028868000958C, 0x8000959000034C87, 0x0006ADA50006ADA6,
-    0x000665AB80009594, 0x80009598000639AA, 0x000659AF8000959C,
-    0x000675AD000669AE, 0x800095A0000675AC, 0x0007A5B3800095A4,
-    0x0007A5B5800095A8, 0x800095B0800095AC, 0x800095B40007A5B6,
-    0x000761BB800094A4, 0x000761BD000769BA, 0x000761BF000779BC,
-    0x00065DA8000769BE, 0x000665AA00066DA9, 0x000665AC00067DAB,
-    0x000665AE00066DAD, 0x800095B800069DAF, 0x800095C0800095BC,
-    0x800095C8800095C4, 0x800095D0800095CC, 0x0006F5B8800095D4,
-    0x0006F5BA0006FDB9, 0x000699BC000689BB, 0x000689BE000699BD,
-    0x0006E5B0000681BF, 0x0006E5B20006EDB1, 0x0006E5B40006FDB3,
-    0x0006D1B60006EDB5, 0x0006EDA30006CDB7, 0x800095DC800095D8,
-    0x800095E4800095E0, 0x0006EDA50006EDA6, 0x000629AB800095E8,
-    0x800095EC000621AA, 0x000629AF800095F0, 0x000629AD000621AE,
-    0x00005D80000631AC, 0x0000658200006D81, 0x000198BE00018084,
-    0x0000D887800095F8, 0x800095FC000F7C86, 0x8000960480009600,
-    0x8000960C80009608, 0x8000961480009610, 0x00023DA880009618,
-    0x000299AA000281A9, 0x0002B9AC000291AB, 0x0002D9AE0002B9AD,
-    0x8000961C0002D9AF, 0x8000962480009620, 0x80009628000EA4BF,
-    0x800096308000962C, 0x000371B880009634, 0x000371BA000371B9,
-    0x0003D5BC000371BB, 0x0003D5BE0003DDBD, 0x0002ADB00003CDBF,
-    0x0002B1B20002B1B1, 0x000351B400028DB3, 0x000351B6000351B5,
-    0x000211B3000351B7, 0x8000963C80009638, 0x0003508480009640,
-    0x000231B5000239B6, 0x000275BB80009644, 0x80009648000269BA,
-    0x000395BF8000964C, 0x000259BD000395BE, 0x80009650000265BC,
-    0x80009654000255A3, 0x00027DA6000380BE, 0x8000965C80009658,
-    0x00022DAA000275A5, 0x80009660000231AB, 0x0003D1AE80009668,
-    0x000221AC0003D1AF, 0x0001D58000021DAD, 0x0001E1820001DD81,
-    0x8000966C0002E9B3, 0x0002E9B60002F9B5, 0x8000967480009670,
-    0x0002D5BA80009678, 0x000115BC00010DBB, 0x000109BE00011DBD,
-    0x000D2086000109BF, 0x8000967C00028887, 0x8000968480009680,
-    0x0006F8EF80009688, 0x0007D8E18000968C, 0x000634E380009690,
-    0x8000969880009694, 0x800096A08000969C, 0x800096A8800096A4,
-    0x800096AC0000ACEF, 0x800096B4800096B0, 0x800096BC800096B8,
-    0x000C6C84000150EF, 0x800096C00006DCE1, 0x800096C400061CE3,
-    0x800096C8000068E3, 0x000225A300012CE1, 0x0000198100002D82,
-    0x800096CC00007580, 0x000235A5000225A6, 0x0001C1AB00080085,
-    0x800096D0000219AA, 0x0001C5AF000C24BE, 0x0001D1AD0001C5AE,
-    0x000DD5A80001D9AC, 0x000DEDAA000DE5A9, 0x000DFDAC000DE5AB,
-    0x000DEDAE000DE5AD, 0x80009664000DE5AF, 0x00030886800096D4,
-    0x800096D800038C87, 0x800096E0800096DC, 0x000EC5B8800096E4,
-    0x000EC5BA000ECDB9, 0x000EF9BC000ED1BB, 0x000E99BE000EF9BD,
-    0x000D9DB0000E95BF, 0x000E4DB2000E45B1, 0x000E41B4000E45B3,
-    0x000E41B6000E41B5, 0x000F01B8000E41B7, 0x000F05BA000F0DB9,
-    0x000F05BC000F1DBB, 0x000F31BE000F0DBD, 0x000FF5B0000F2DBF,
-    0x000FCDB2000FF9B1, 0x000FC1B4000FC1B3, 0x000FC1B6000FC1B5,
-    0x000E2DA8000FC1B7, 0x000FD5AA000FD5A9, 0x000FF5AC000FEDAB,
-    0x000F95AE000F9DAD, 0x800096E8000F8DAF, 0x800095F4800096EC,
-    0x800096F4800096F0, 0x800096FC800096F8, 0x000E1DB380009700,
-    0x8000970880009704, 0x800097108000970C, 0x000E3DB5000E5DB6,
-    0x000E7DBB80009714, 0x80009718000E71BA, 0x000E45BF8000971C,
-    0x000E55BD000E45BE, 0x80009720000E59BC, 0x80009724000E59A3,
-    0x000E19A680009728, 0x800097308000972C, 0x000E35AA000E79A5,
-    0x80009734000E39AB, 0x000E01AE80009738, 0x000E1DAC000E01AF,
-    0x0001B1A8000E11AD, 0x0001B1AA0001B1A9, 0x000171AC0001B1AB,
-    0x000175AE000171AD, 0x0000CD8000016DAF, 0x0000198200000981,
-    0x800097408000973C, 0x000170BE00017884, 0x0000E9B880009748,
-    0x000089BA0000E9B9, 0x00009DBC000085BB, 0x000081BE000081BD,
-    0x000115B00000B5BF, 0x000115B200011DB1, 0x0000F9B40000E9B3,
-    0x0000E9B60000F9B5, 0x000239B30000E1B7, 0x000488868000974C,
-    0x8000975000037887, 0x0002F9B50002F1B6, 0x0002D9BB80009754,
-    0x800097580002D1BA, 0x00033DBF8000975C, 0x00033DBD00033DBE,
-    0x800097600002C1BC, 0x8000976400027DA3, 0x0002B5A680009768,
-    0x800097708000976C, 0x000295AA0002BDA5, 0x8000977400029DAB,
-    0x000379AE80009778, 0x000285AC000379AF, 0x001AC182000379AD,
-    0x8000977C001AC983, 0x001EB58680009780, 0x001E318400122987,
-    0x0012A18A001EBD85, 0x800097840012BD8B, 0x0016918E80009788,
-    0x00137D8C00173D8F, 0x000A51920016318D, 0x8000978C000AE593,
-    0x000EF196006CDCEF, 0x000A7594000EED97, 0x0002099A000E3995,
-    0x800097900002C59B, 0x8000979880009794, 0x0002DD9C000F3CE1,
-    0x000F219F0076A8E3, 0x000BA59D000FE19E, 0x000B819B000B2D9C,
-    0x0017399900169D9A, 0x0012E59700179198, 0x0013299500133D96,
-    0x001F159300139194, 0x001F9191001F1592, 0x001B018F001B7190,
-    0x800097A08000979C, 0x800097A8800097A4, 0x800097AC000F60EF,
-    0x000528BE800097B0, 0x00052484000FE4E3, 0x00002981000CA8E1,
-    0x800097B800002580, 0x00050C8700002182, 0x800097BC0004A486,
-    0x800097C4800097C0, 0x0000D8EF800097C8, 0x000190E1000E88EF,
-    0x00000CE3000EFCE1, 0x800097CC000E1CE3, 0x800097D4800097D0,
-    0x800097DC800097D8, 0x00070484800097E0, 0x800097E4000704BE,
-    0x800097EC800097E8, 0x800097F4800097F0, 0x800097FC800097F8,
-    0x0003D5A380009800, 0x000311A10003D5A2, 0x80009744000FC5A0,
-    0x80009804800097B4, 0x8000980C80009808, 0x8000981480009810,
-    0x00059DA880009818, 0x000639AA000621A9, 0x000619AC000631AB,
-    0x000679AE000619AD, 0x00060DB0000675AF, 0x000611B2000611B1,
-    0x00063DB4000625B3, 0x0006D5B60006DDB5, 0x0006CDB80006F5B7,
-    0x000681BA0006ADB9, 0x000681BC000681BB, 0x000731BE000731BD,
-    0x000691B3000731BF, 0x0000198100001982, 0x8000981C00009180,
-    0x0006B1B50006B9B6, 0x0006F5BB80009820, 0x800098240006E9BA,
-    0x000639BF80009828, 0x0006D1BD0006D9BE, 0x8000982C0006E5BC,
-    0x0000E8860006D5A3, 0x0006FDA600008487, 0x8000983480009830,
-    0x0006ADAA0006F5A5, 0x800098380006B1AB, 0x00069DAE8000983C,
-    0x0006A1AC00067DAF, 0x80009840000695AD, 0x006929B580009844,
-    0x006C51B3006801B4, 0x006D19B1006D11B2, 0x80009848007071B0,
-    0x800098508000984C, 0x8000985880009854, 0x800098608000985C,
-    0x007801A780009864, 0x007D69A5007DE9A6, 0x0001F9A3007C01A4,
-    0x000159A1000151A2, 0x007169AF80009868, 0x007001AD007161AE,
-    0x007551AB007549AC, 0x0079F5A9007401AA, 0x0006F9A800795DA8,
-    0x000605AA00060DA9, 0x000605AC00061DAB, 0x00062DAE000625AD,
-    0x8000986C000625AF, 0x8000987480009870, 0x8000987C80009878,
-    0x8000988480009880, 0x000625B880009888, 0x000625BA00062DB9,
-    0x0001DDBC0001C5BB, 0x0001CDBE0001C5BD, 0x00065DB00001C5BF,
-    0x00062DB2000625B1, 0x000625B400063DB3, 0x000625B600062DB5,
-    0x00058DB300061DB7, 0x800098908000988C, 0x80009898000340BE,
-    0x00058DB50005A5B6, 0x0005E1BB8000989C, 0x0003A484000599BA,
-    0x0005E9BF800098A0, 0x0005E9BD0005E1BE, 0x000049820005F9BC,
-    0x0003F5800005C9A3, 0x0005E1A600004981, 0x800098A8800098A4,
-    0x0005DDAA0005C9A5, 0x007CC8860005A5AB, 0x0005A5AE0003FC87,
-    0x0005BDAC0005ADAF, 0x00022DA80005ADAD, 0x00025DAA000245A9,
-    0x000279AC000255AB, 0x000399AE000279AD, 0x800098AC000399AF,
-    0x800098B4800098B0, 0x800098BC800098B8, 0x800098C4800098C0,
-    0x00039DB8800098C8, 0x00036DBA000365B9, 0x00037DBC000365BB,
-    0x00036DBE000365BD, 0x0003C1B0000365BF, 0x0003C1B20003C1B1,
-    0x0003C1B40003C1B3, 0x0003C1B60003C1B5, 0x800098CC0003C1B7,
-    0x800098D4800098D0, 0x800098DC800098D8, 0x800098E00003E0EF,
-    0x007D68BE800098E4, 0x800098E800031CE3, 0x800098EC0001FCE1,
-    0x800098F4800098F0, 0x000395B3800098F8, 0x80009900800098FC,
-    0x007D1C8480009908, 0x0003B1B50003B9B6, 0x000245BB8000990C,
-    0x007CA886000245BA, 0x000245BF007DA487, 0x000255BD000245BE,
-    0x00061CE1000255BC, 0x000650E3000794E1, 0x0000AD800006F4E3,
-    0x0000D1820000D181, 0x8000991480009910, 0x8000991C80009918,
-    0x8000992480009920, 0x00068CEF80009928, 0x000315A3000624EF,
-    0x800099308000992C, 0x8000993880009934, 0x000331A5000339A6,
-    0x0002C5AB007C9084, 0x8000993C0002C5AA, 0x0002C5AF80009940,
-    0x0002D5AD0002C5AE, 0x007E11A80002D5AC, 0x007E25AA007E19A9,
-    0x007E49AC007E59AB, 0x007EBDAE007EB5AD, 0x80009904007EB5AF,
-    0x8000994880009944, 0x800099508000994C, 0x8000995880009954,
-    0x007F4DB88000995C, 0x007F51BA007F55B9, 0x007F79BC007F51BB,
-    0x007F6DBE007F65BD, 0x007ECDB0007F65BF, 0x007EDDB2007ED5B1,
-    0x007ECDB4007ED5B3, 0x007F7DB6007F75B5, 0x007EADB3007F75B7,
-    0x8000996480009960, 0x8000996C80009968, 0x007EC1B5007E35B6,
-    0x007E11BB80009970, 0x0003E0BE007E09BA, 0x007FB1BF80009974,
-    0x007E01BD007E09BE, 0x00005D82007E09BC, 0x00004580007EE9A3,
-    0x007E71A600005D81, 0x8000997C80009978, 0x007E4DAA007E85A5,
-    0x00004886007E55AB, 0x007E4DAE00006C87, 0x007E4DAC007FF5AF,
-    0x007EB5B0007E45AD, 0x007EC1B2007EC1B1, 0x007EC1B4007EC1B3,
-    0x007EF5B6007EFDB5, 0x000109B8000139B7, 0x000119BA000109B9,
-    0x000109BC000119BB, 0x0001BDBE000109BD, 0x800099800001B5BF,
-    0x8000998880009984, 0x8000998C80009894, 0x8000999480009990,
-    0x007E2DA880009998, 0x007EF5AA007E39A9, 0x007ED5AC007ECDAB,
-    0x007ED5AE007EDDAD, 0x007E1DB3007ECDAF, 0x800099A08000999C,
-    0x800099A8800099A4, 0x007E3DB5007E3DB6, 0x007E65BB800099AC,
-    0x800099B0007E11BA, 0x007E59BF800099B4, 0x007E75BD007E65BE,
-    0x800099B8007E75BC, 0x800099BC007E59A3, 0x007E79A6800099C0,
-    0x800099C8800099C4, 0x007E55AA007E79A5, 0x800099CC007E21AB,
-    0x007E21AE800099D0, 0x007E31AC007E1DAF, 0x0002D5A8007E31AD,
-    0x000349AA0002DDA9, 0x000359AC000349AB, 0x00034DAE000345AD,
-    0x800099D4000345AF, 0x800099DC800099D8, 0x0003F4BE800099E0,
-    0x800099EC800099E8, 0x000381B8800099F0, 0x000385BA00038DB9,
-    0x000389BC000399BB, 0x0003B5BE0003BDBD, 0x00033DB000007DBF,
-    0x0003E1B20003E1B1, 0x0003E1B40003E1B3, 0x0003D1B60003E9B5,
-    0x0000A5800003D1B7, 0x0000A5820000AD81, 0x800099F40003CDB3,
-    0x0003C9B60003DDB5, 0x0002E084800099F8, 0x0003F5BA800099FC,
-    0x000339BC000329BB, 0x000329BE000339BD, 0x007F80E3000321BF,
-    0x007EB8E180009A00, 0x0003E08780009A04, 0x80009A0800047486,
-    0x80009A1080009A0C, 0x80009A1880009A14, 0x007F28EF80009A1C,
-    0x80009A2480009A20, 0x80009A2C80009A28, 0x80009A3480009A30,
-    0x0003D8EF80009A38, 0x80009A4080009A3C, 0x000390E380009A44,
-    0x0001D8E180009A48, 0x0000158080009A4C, 0x0000058200001981,
-    0x00054CBE80009A50, 0x000209A380009A58, 0x000219A5000440BE,
-    0x80009A6080009A5C, 0x00026C8400020DA6, 0x0002EDAB80009A64,
-    0x0002FDAD000231AA, 0x0002E5AF0002FDAC, 0x000440860002EDAE,
-    0x80009A680005B487, 0x80009A7080009A6C, 0x80009A7880009A74,
-    0x80009A7C0001CCEF, 0x80009A80007ED4E1, 0x80009A84007E60E3,
-    0x80009A8C80009A88, 0x0001FDB380009A90, 0x80009A9880009A94,
-    0x80009AA080009A9C, 0x000145B5000171B6, 0x00013DBB80009AA4,
-    0x80009AA800013DBA, 0x00011DBF80009AAC, 0x00011DBD00011DBE,
-    0x000635A800011DBC, 0x000635AA00063DA9, 0x00068DAC00069DAB,
-    0x0006DDAE0006B1AD, 0x80009A540006D1AF, 0x80009AB480009AB0,
-    0x00001D8080009AB8, 0x0000BD8200000981, 0x0006BDB880009ABC,
-    0x00074DBA000745B9, 0x00075DBC000745BB, 0x000741BE000741BD,
-    0x0006B9B0000741BF, 0x000689B20006B9B1, 0x000699B4000689B3,
-    0x00068DB6000699B5, 0x0006BDA3000685B7, 0x0002808480009AC0,
-    0x80009AC880009AC4, 0x000605A5000631A6, 0x00067DAB80009ACC,
-    0x0000088600067DAA, 0x00065DAF00036087, 0x00065DAD00065DAE,
-    0x80009AD000065DAC, 0x80009AD400079DB3, 0x00075DB680009AD8,
-    0x80009AE080009ADC, 0x000779BA000755B5, 0x80009AE4000745BB,
-    0x000745BE80009AE8, 0x000751BC00074DBF, 0x000699B000074DBD,
-    0x0006A1B20006A5B1, 0x0006A5B40006BDB3, 0x0006A5B60006ADB5,
-    0x0006E9B80006D9B7, 0x0006F9BA0006E9B9, 0x0006D9BC0006F1BB,
-    0x00011DBE0006D9BD, 0x800099E4000115BF, 0x80009AF080009AEC,
-    0x80009AF880009AF4, 0x80009B0080009AFC, 0x000625A880009B04,
-    0x000631AA000631A9, 0x0006F9AC000609AB, 0x0006E9AE0006F9AD,
-    0x0006D9A30006E9AF, 0x80009B0C80009B08, 0x80009B1480009B10,
-    0x000611A5000619A6, 0x000601AB80009B18, 0x80009B1C00063DAA,
-    0x000609AF80009B20, 0x000609AD000601AE, 0x80009B24000615AC,
-    0x00001D8100001D82, 0x80009B2800001D80, 0x80009B3080009B2C,
-    0x0001AC84000010BE, 0x0001088780009B38, 0x80009B3C000F4C86,
-    0x80009B4480009B40, 0x000291A880009B48, 0x0002C9AA000299A9,
-    0x0002DDAC0002C5AB, 0x0002C1AE0002C1AD, 0x80009B4C0002F5AF,
-    0x80009B5480009B50, 0x0000008D80009B58, 0x80009B6080009B5C,
-    0x0003C5B880009B64, 0x0003C5BA0003CDB9, 0x0003CDBC0003DDBB,
-    0x0003F5BE0003FDBD, 0x00028DB000039DBF, 0x000321B2000321B1,
-    0x000321B4000321B3, 0x000321B6000321B5, 0x000241B3000321B7,
-    0x80009B6C80009B68, 0x80009B7480009B70, 0x000255B5000241B6,
-    0x0003E5BB80009B78, 0x80009B7C000249BA, 0x0003E5BF80009B80,
-    0x0003F5BD0003E5BE, 0x80009B840003F5BC, 0x80009B88000608E1,
-    0x80009B8C0007BCE3, 0x80009B9480009B90, 0x00003D8080009B98,
-    0x0000198200000981, 0x80009BA080009B9C, 0x000CF4BE0003F884,
-    0x0000F0E3000744EF, 0x0001B8E180009BA8, 0x000CF086000D0484,
-    0x80009BAC0002A487, 0x80009BB480009BB0, 0x80009BB8000C28BE,
-    0x000014EF80009BBC, 0x80009BC480009BC0, 0x000381A380009BC8,
-    0x80009BD080009BCC, 0x80009BD880009BD4, 0x000395A5000381A6,
-    0x000225AB80009BDC, 0x80009BE0000389AA, 0x000225AF80009BE4,
-    0x000235AD000225AE, 0x80009BE8000235AC, 0x80009BF080009BEC,
-    0x80009BF880009BF4, 0x80009C0080009BFC, 0x80009C04000628EF,
-    0x80009C0800063CE1, 0x00001980000134E3, 0x00007D8200000981,
-    0x000235B380009C0C, 0x80009C1880009C10, 0x80009C1C000FC0BE,
-    0x000225B50001F5B6, 0x0001D5BB80009C20, 0x80009C240001C9BA,
-    0x000179BF80009C28, 0x0001C5BD000171BE, 0x000DE5A80001C5BC,
-    0x000DFDAA000DF5A9, 0x000DEDAC000DF5AB, 0x000E3DAE000E35AD,
-    0x80009BA4000E35AF, 0x0000888680009C14, 0x80009C2C00034087,
-    0x80009C3480009C30, 0x000EDDB880009C38, 0x000EE5BA000EEDB9,
-    0x000E99BC000E8DBB, 0x000E8DBE000E85BD, 0x000E4DB0000E85BF,
-    0x000E51B2000E51B1, 0x000EF5B4000E51B3, 0x000EF5B6000EFDB5,
-    0x000D75A3000EEDB7, 0x80009C4080009C3C, 0x80009C4880009C44,
-    0x000D65A5000EB5A6, 0x000E95AB80009C4C, 0x80009C50000E89AA,
-    0x000E39AF80009C54, 0x000E85AD000E31AE, 0x80009C58000E85AC,
-    0x80009C5C000FD9B3, 0x000FFDB680009C60, 0x80009C6880009C64,
-    0x000FDDBA000FF5B5, 0x80009C6C000FA1BB, 0x000FA1BE80009C70,
-    0x000FB1BC000FA9BF, 0x000E81B8000FB1BD, 0x000E81BA000E81B9,
-    0x000E81BC000E81BB, 0x000E81BE000E81BD, 0x000E1DB0000EBDBF,
-    0x000E61B2000E61B1, 0x000E61B4000E61B3, 0x000E51B6000E69B5,
-    0x000E45A8000E51B7, 0x000E55AA000E5DA9, 0x000E79AC000E6DAB,
-    0x000E6DAE000E79AD, 0x80009C74000E65AF, 0x0000118100001982,
-    0x80009C7800001D80, 0x80009C7C80009B34, 0x000E9DA380009C80,
-    0x0000288600028C84, 0x80009C8800014887, 0x000EB1A5000EB9A6,
-    0x000EE5AB80009C8C, 0x80009C90000E99AA, 0x000EEDAF80009C94,
-    0x000EF5AD000EE5AE, 0x80009C98000EF5AC, 0x80009C9C000E19B3,
-    0x0001E1B680009CA0, 0x80009CA880009CA4, 0x0001DDBA000E31B5,
-    0x80009CAC000161BB, 0x000161BE80009CB0, 0x000171BC000169BF,
-    0x000DC9A8000171BD, 0x000DD5AA000DD5A9, 0x000DFDAC000DE5AB,
-    0x00022DAE000225AD, 0x80009CB4000225AF, 0x80009CBC80009CB8,
-    0x0000008C80009CC0, 0x80009CC880009CC4, 0x000281B880009CCC,
-    0x000285BA00028DB9, 0x000289BC000299BB, 0x0002B5BE0002BDBD,
-    0x00025DB000033DBF, 0x0002E1B20002E1B1, 0x0002E1B40002E1B3,
-    0x0002D1B60002E9B5, 0x000D5DA30002D1B7, 0x80009CD480009CD0,
-    0x80009CDC80009CD8, 0x000D75A50002A5A6, 0x000225AB80009CE0,
-    0x80009CE4000299AA, 0x00022DAF80009CE8, 0x000235AD000225AE,
-    0x00001181000235AC, 0x000EC8EF00006980, 0x80009CEC00001982,
-    0x80009CF880009CF0, 0x0003EC8480009CFC, 0x0004A887000468BE,
-    0x000D04E10004CC86, 0x000E1CE380009D00, 0x000291A880009D04,
-    0x0002A1AA000299A9, 0x0002D1AC0002A1AB, 0x0002D5AE0002DDAD,
-    0x80009D080002C9AF, 0x80009D1080009D0C, 0x80009D1880009D14,
-    0x80009D2080009D1C, 0x000179B880009D24, 0x0001CDBA000179B9,
-    0x0001DDBC0001C5BB, 0x0001C5BE0001C5BD, 0x0002B9B00001F5BF,
-    0x000285B200028DB1, 0x000155B400014DB3, 0x000155B600015DB5,
-    0x80009D28000149B7, 0x80009D3080009D2C, 0x0005108480009D34,
-    0x000EACEF000418BE, 0x000190E1000E40EF, 0x00003CE3000EACE1,
-    0x80009D38000EB0E3, 0x80009D4080009D3C, 0x000269B380009D44,
-    0x80009D4C80009D48, 0x80009D5480009D50, 0x000219B500021DB6,
-    0x000209BB80009D58, 0x80009D5C00023DBA, 0x0001D9BF80009D60,
-    0x0001D9BD0001D1BE, 0x0005A9A3000211BC, 0x80009D6480009CF4,
-    0x80009D6C80009D68, 0x0005D9A50005DDA6, 0x0005C9AB80009D70,
-    0x000308860005FDAA, 0x000619AF00037C87, 0x000619AD000611AE,
-    0x00079D800005D1AC, 0x0007A5820007AD81, 0x80009D740006B9B3,
-    0x00077DB6000775B5, 0x80009D7C80009D78, 0x000759BA80009D80,
-    0x000749BC000759BB, 0x0007F9BE000749BD, 0x000659A80007F9BF,
-    0x00067DAA000659A9, 0x0007B1AC000671AB, 0x0007B1AE0007B1AD,
-    0x80009D840007B1AF, 0x80009D8C80009D88, 0x80009D9480009D90,
-    0x80009D9C80009D98, 0x000771B880009DA0, 0x000771BA000771B9,
-    0x000715BC000771BB, 0x000715BE00071DBD, 0x0007D1B000070DBF,
-    0x0007D1B20007D1B1, 0x000771B40007D1B3, 0x000771B6000771B5,
-    0x0007F5A3000771B7, 0x80009DA880009DA4, 0x80009DB080009DAC,
-    0x000639A5000631A6, 0x000615AB80009DB4, 0x80009DB8000615AA,
-    0x0006B5AF80009DBC, 0x000605AD0006B5AE, 0x80009DC0000605AC,
-    0x80009DC880009DC4, 0x0000198080009DCC, 0x0000058200001981,
-    0x0001AC8480009DD0, 0x80009C840001ACBE, 0x0000D08680009DD4,
-    0x80009DD80001E487, 0x00018DA880009DDC, 0x000195AA000195A9,
-    0x0001CDAC0001DDAB, 0x0001FDAE0001F5AD, 0x0000A08400013DAF,
-    0x80009DE480009DE0, 0x80009DEC80009DE8, 0x80009DF480009DF0,
-    0x0000C9B880009DF8, 0x0000D9BA0000C9B9, 0x0000F9BC0000D1BB,
-    0x000099BE0000F9BD, 0x000145B0000099BF, 0x000145B200014DB1,
-    0x000145B400015DB3, 0x000145B600014DB5, 0x00021DB30000F9B7,
-    0x80009E0080009DFC, 0x80009E0880009E04, 0x000245B5000275B6,
-    0x000289BB80009E0C, 0x80009E10000281BA, 0x000289BF80009E14,
-    0x000289BD000281BE, 0x001A5C85000291BC, 0x80009E18000259A3,
-    0x000231A680009E1C, 0x80009E2480009E20, 0x0002C5AA000201A5,
-    0x80009E280002CDAB, 0x0002C5AE80009E2C, 0x0002D5AC0002CDAF,
-    0x001C00BE0002CDAD, 0x80009E3480009E30, 0x00000D80001C0084,
-    0x00003D8200003581, 0x80009E3C80009E38, 0x80009E4880009E40,
-    0x80009E5080009E4C, 0x80009E5880009E54, 0x001CE08680009E5C,
-    0x80009E6000030487, 0x0002A8EF80009E64, 0x80009E680006C8EF,
-    0x000150E380009E6C, 0x000180E1000614E3, 0x80009E7000057CE1,
-    0x80009E7880009E74, 0x000279A880009E7C, 0x0002ADAA000279A9,
-    0x0002BDAC0002A5AB, 0x0002ADAE0002A5AD, 0x001CACBE0002A5AF,
-    0x80009E8480009E80, 0x80009E8C80009E88, 0x80009E9480009E90,
-    0x000109B880009E98, 0x00011DBA000109B9, 0x000139BC000115BB,
-    0x0001DDBE000139BD, 0x0002DDB00001D5BF, 0x000161B2000161B1,
-    0x000161B4000161B3, 0x000151B6000169B5, 0x000774E1000151B7,
-    0x000608E30000F4E3, 0x80009E9C0001BCE1, 0x80009EA480009EA0,
-    0x80009EA8001C1484, 0x80009EB080009EAC, 0x000000EC80009EB8,
-    0x0006D4EF000028EF, 0x0000618080009EBC, 0x0000618200006181,
-    0x80009EC0000269B3, 0x000269B6000279B5, 0x001CC08680009EC4,
-    0x000255BA001CE887, 0x0001F9BC0001E9BB, 0x0001E9BE0001F9BD,
-    0x80009E440001E1BF, 0x80009EC880009EB4, 0x80009ECC001DA9A3,
-    0x001DA9A6001DB9A5, 0x80009ED480009ED0, 0x001D95AA80009ED8,
-    0x001E39AC001E29AB, 0x001E29AE001E39AD, 0x001EF1B3001E21AF,
-    0x80009EE080009EDC, 0x80009EE880009EE4, 0x001EE1B5001EE9B6,
-    0x001EBDBB80009EEC, 0x80009EF0001EBDBA, 0x001E8DBF80009EF4,
-    0x001E8DBD001E8DBE, 0x001E29A8001E8DBC, 0x001E39AA001E29A9,
-    0x001E29AC001E39AB, 0x001E51AE001E29AD, 0x80009EF8001E51AF,
-    0x80009F0080009EFC, 0x80009F0880009F04, 0x80009F1080009F0C,
-    0x001FEDB880009F14, 0x001FFDBA001FF5B9, 0x001FEDBC001FF5BB,
-    0x001F4DBE001F45BD, 0x001E39B0001F45BF, 0x001E19B2001E39B1,
-    0x001E09B4001E19B3, 0x001FDDB6001E09B5, 0x001FB1A3001FD5B7,
-    0x0000158100002D82, 0x80009F1800001D80, 0x001FA1A5001FA9A6,
-    0x001FFDAB80009F1C, 0x80009F20001FFDAA, 0x001FCDAF0001A0BE,
-    0x001FCDAD001FCDAE, 0x80009F24001FCDAC, 0x0000E886001E95B3,
-    0x000141B600012C87, 0x80009F2C80009F28, 0x000145BA000155B5,
-    0x80009F30000121BB, 0x000121BE80009F34, 0x000131BC000129BF,
-    0x001E2DA8000131BD, 0x001E59AA001E39A9, 0x000199AC001E51AB,
-    0x000185AE000185AD, 0x0001E0840001BDAF, 0x80009F3C80009F38,
-    0x80009F4480009F40, 0x80009F4C80009F48, 0x000161B880009F50,
-    0x000161BA000161B9, 0x000161BC000161BB, 0x000161BE000161BD,
-    0x0001C5B0000161BF, 0x0001A1B20001A1B1, 0x0001A1B40001A1B3,
-    0x000191B60001A9B5, 0x001DD1A3000191B7, 0x80009F5880009F54,
-    0x80009F6080009F5C, 0x000211A5000205A6, 0x000265AB80009F64,
-    0x80009F68000201AA, 0x00026DAF80009F6C, 0x000275AD000265AE,
-    0x80009F70000275AC, 0x80009F7880009F74, 0x80009F8080009F7C,
-    0x80009F8880009F84, 0x0000398080009F8C, 0x0000058200003981,
-    0x00033CBE80009F90, 0x0003448780009F98, 0x0004CC840004DC86,
-    0x80009FA080009F9C, 0x80009FA40003E5B3, 0x000331B6000331B5,
-    0x80009FAC80009FA8, 0x00031DBA80009FB0, 0x00030DBC000315BB,
-    0x0000B1BE0000B1BD, 0x80009FB40000B1BF, 0x80009FB8000194E1,
-    0x80009FBC00017CE3, 0x80009FC480009FC0, 0x80009FCC80009FC8,
-    0x80009FD480009FD0, 0x80009FDC80009FD8, 0x80009FE480009FE0,
-    0x000255A8000274EF, 0x000261AA00025DA9, 0x0002B5AC000261AB,
-    0x0002B5AE0002BDAD, 0x000568BE0002ADAF, 0x80009FEC80009FE8,
-    0x80009FF480009FF0, 0x0000AD810000AD82, 0x000169B80000AD80,
-    0x000109BA000169B9, 0x000119BC000109BB, 0x000109BE000119BD,
-    0x0002D5B0000109BF, 0x0002D5B20002DDB1, 0x000179B4000169B3,
-    0x000169B6000179B5, 0x001E78E1000161B7, 0x001E04E3001F5CE1,
-    0x80009FF8001F10E3, 0x8000A00080009FFC, 0x8000A00C8000A008,
-    0x8000A0148000A010, 0x0004F8BE8000A018, 0x00006CEF8000A01C,
-    0x0002A9A3001E64EF, 0x000448868000A020, 0x8000A02400050487,
-    0x00027DA500027DA6, 0x000259AB8000A028, 0x8000A02C000251AA,
-    0x0001FDAF8000A030, 0x0001FDAD0001FDAE, 0x000675A8000241AC,
-    0x000681AA000681A9, 0x00068DAC00069DAB, 0x0006BDAE0006B5AD,
-    0x8000A0040006DDAF, 0x8000A0388000A034, 0x8000A0408000A03C,
-    0x8000A0488000A044, 0x0006ADB88000A04C, 0x0006BDBA0006B5B9,
-    0x000775BC0006B5BB, 0x000775BE00077DBD, 0x0006A9B000076DBF,
-    0x0006BDB20006B5B1, 0x0006ADB40006B5B3, 0x00069DB6000695B5,
-    0x000609B3000695B7, 0x8000A0548000A050, 0x8000A05C8000A058,
-    0x000605B5000635B6, 0x000671BB8000A060, 0x8000A06400066DBA,
-    0x000651BF8000A068, 0x000661BD000669BE, 0x00001982000669BC,
-    0x0000658000064DA3, 0x000671A600001981, 0x8000A0708000A06C,
-    0x000629AA000641A5, 0x0003A484000635AB, 0x00062DAE0001E4BF,
-    0x00062DAC000615AF, 0x00016C85000625AD, 0x000008848000A078,
-    0x000C008680009F94, 0x8000A07C00006087, 0x0006BDA88000A080,
-    0x00068DAA000685A9, 0x00069DAC000685AB, 0x0006FDAE000681AD,
-    0x00069DB00006EDAF, 0x0006A5B20006ADB1, 0x000179B4000169B3,
-    0x000169B6000179B5, 0x000159B8000169B7, 0x0001EDBA000159B9,
-    0x0001FDBC0001E5BB, 0x0001E5BE0001E5BD, 0x000625B30001D5BF,
-    0x8000A074000F08BE, 0x8000A0888000A084, 0x000635B5000621B6,
-    0x000649BB8000A08C, 0x8000A090000679BA, 0x00019DBF8000A094,
-    0x00019DBD00019DBE, 0x8000A098000651BC, 0x8000A09C000661A3,
-    0x000665A68000A0A0, 0x8000A0A88000A0A4, 0x00063DAA000671A5,
-    0x8000A0AC00060DAB, 0x0001D9AE8000A0B0, 0x000615AC0001D9AF,
-    0x000114E10001D9AD, 0x0036D0E300011CE3, 0x000C9084000104E1,
-    0x8000A0B88000A0B4, 0x006A218F8000A0BC, 0x8000A0C48000A0C0,
-    0x8000A0CC8000A0C8, 0x0037FCEF000148EF, 0x0012CD978000A0D0,
-    0x0013099500126596, 0x001671930016E194, 0x006AF19100171992,
-    0x0006619F006A5990, 0x001A619D001B099E, 0x001B199B001A699C,
-    0x001E5999001E519A, 0x000610E3001F0998, 0x0006ECE18000A0D4,
-    0x0002B1A38000A0D8, 0x0007D9A1000399A2, 0x003A9DB3000699A0,
-    0x003801B1003A85B2, 0x000768EF003E01B0, 0x0026BDB58000A0DC,
-    0x003001AB002401B4, 0x003645A9003619AA, 0x003EF1AF00364DA8,
-    0x0033C5AD003C01AE, 0x000E65A30032FDAC, 0x000361A1000E65A2,
-    0x000BE1A78000A0E0, 0x000A79A5000A71A6, 0x006B0182000F89A4,
-    0x000D0886006B0183, 0x00173186000D4C87, 0x006BB1840017B987,
-    0x0012618A00165185, 0x8000A0E80013058B, 0x001F118E8000A0EC,
-    0x0013B58C001F118F, 0x001A2992001E498D, 0x000CC4BE001B4193,
-    0x00075196000F8084, 0x001B519400075197, 0x0006219A001A3195,
-    0x8000A0F00002699B, 0x0001C5818000A0F4, 0x0003599C0001F980,
-    0x000F48BE00019182, 0x8000A0FC8000A0F8, 0x8000A1048000A100,
-    0x8000A10C8000A108, 0x8000A1100006D4EF, 0x8000A114000674E1,
-    0x8000A118000154E3, 0x8000A1208000A11C, 0x8000A1288000A124,
-    0x8000A1308000A12C, 0x8000A1340002D1B3, 0x000245B6000245B5,
-    0x8000A13C8000A138, 0x000229BA8000A140, 0x00023DBC000225BB,
-    0x000221BE000221BD, 0x000D1DA3000215BF, 0x8000A1488000A144,
-    0x8000A1508000A14C, 0x000D89A5000D89A6, 0x000DE9AB8000A158,
-    0x00030886000DE5AA, 0x000DD9AF00034C87, 0x000DEDAD000DEDAE,
-    0x00004980000DF1AC, 0x0000598200004981, 0x8000A15C000F25B3,
-    0x000FD1B6000FD1B5, 0x8000A1648000A160, 0x000FFDBA8000A168,
-    0x000FD1BC000FF5BB, 0x000FC9BE000FDDBD, 0x8000A0E4000FC9BF,
-    0x8000A16C8000A154, 0x8000A1748000A170, 0x8000A17C8000A178,
-    0x000F89A88000A180, 0x000FC1AA000F89A9, 0x000F35AC000FC1AB,
-    0x000F35AE000F3DAD, 0x000F55B0000F2DAF, 0x000F69B2000F5DB1,
-    0x000F61B4000F65B3, 0x000F61B6000F61B5, 0x000F21B8000F61B7,
-    0x000F21BA000F21B9, 0x000F2DBC000F3DBB, 0x000F1DBE000F15BD,
-    0x000E69A30000F9BF, 0x8000A1888000A184, 0x00009C848000A18C,
-    0x000E9DA5000E9DA6, 0x000EB9AB8000A190, 0x8000A194000EB1AA,
-    0x000E85AF8000A198, 0x000E91AD000E85AE, 0x8000A19C000E9DAC,
-    0x8000A1A48000A1A0, 0x8000A1AC8000A1A8, 0x8000A1B48000A1B0,
-    0x00003D808000A1B8, 0x0000198200000981, 0x8000A1C08000A1BC,
-    0x0001D8878000A1C4, 0x00019DA800040086, 0x0001D5AA0001D5A9,
-    0x0001F5AC0001EDAB, 0x000115AE00011DAD, 0x8000A1C800010DAF,
-    0x8000A1D08000A1CC, 0x8000A1D88000A1D4, 0x8000A1E08000A1DC,
-    0x0000EDB88000A1E4, 0x00008DBA000085B9, 0x00009DBC000085BB,
-    0x00008DBE000085BD, 0x000175B0000085BF, 0x000175B200017DB1,
-    0x0000F9B40000E9B3, 0x0000E5B60000F9B5, 0x00025DB30000D5B7,
-    0x8000A1EC8000A1E8, 0x8000A1F48000A1F0, 0x00027DB500029DB6,
-    0x0002B9BB8000A1F8, 0x8000A1FC0002B1BA, 0x00033DBF8000A200,
-    0x00033DBD00033DBE, 0x0004CCBE0002A1BC, 0x8000A204000219A3,
-    0x0002D9A68000A208, 0x8000A2108000A20C, 0x0002F5AA000239A5,
-    0x8000A2140002FDAB, 0x000379AE8000A218, 0x0002E5AC000379AF,
-    0x00050CBE000379AD, 0x8000A2208000A21C, 0x8000A224000610BE,
-    0x8000A22C8000A228, 0x000035808000A230, 0x0000358200003D81,
-    0x0004B0868000A238, 0x8000A23C00033487, 0x8000A2448000A240,
-    0x000270EF8000A248, 0x8000A2508000A24C, 0x8000A2588000A254,
-    0x8000A2608000A25C, 0x8000A2688000A264, 0x000194E18000A26C,
-    0x0001F8E38000A270, 0x000130E18000A274, 0x000E54E3000E14E1,
-    0x8000A278000F0CE3, 0x8000A2808000A27C, 0x0005A8848000A284,
-    0x8000A28C8000A288, 0x8000A2948000A290, 0x000E40EF8000A298,
-    0x0002E9B3000E8CEF, 0x8000A2A08000A29C, 0x8000A2AC8000A2A8,
-    0x0002E9B50002E1B6, 0x000169BB8000A2B0, 0x8000A2B4000161BA,
-    0x000159BF8000A2B8, 0x000169BD000169BE, 0x0000B580000171BC,
-    0x0000C1820000C181, 0x0004B08600040484, 0x8000A2BC00050887,
-    0x000E31AC8000A2C0, 0x000E01AE000E39AD, 0x000EB5A8000E01AF,
-    0x000EA5AA000EADA9, 0x000E19B4000E31AB, 0x000E25B6000E2DB5,
-    0x000E11B0000DD9B7, 0x000E15B2000E1DB1, 0x000DCDBC000E09B3,
-    0x000DD5BE000DD5BD, 0x000DE9B8000D69BF, 0x000DDDBA000DE9B9,
-    0x0005A9A3000DD5BB, 0x8000A2C88000A2C4, 0x8000A2D08000A2CC,
-    0x0005A9A50005A1A6, 0x000629AB8000A2D4, 0x8000A2D8000621AA,
-    0x000619AF8000A2DC, 0x000629AD000629AE, 0x8000A2E0000631AC,
-    0x8000A2E40006F9B3, 0x00069DB68000A2E8, 0x8000A2F08000A2EC,
-    0x0006B9BA000699B5, 0x8000A2F40006B9BB, 0x000699BE8000A2F8,
-    0x000699BC000699BF, 0x8000A234000699BD, 0x8000A2FC8000A2A4,
-    0x8000A3048000A300, 0x0001918100019982, 0x00060DA800019980,
-    0x000615AA00061DA9, 0x000675AC00066DAB, 0x000669AE000679AD,
-    0x000619B0000669AF, 0x000629B2000619B1, 0x000639B4000629B3,
-    0x00062DB6000639B5, 0x0007F1B8000621B7, 0x0007F1BA0007F1B9,
-    0x0007D5BC0007F1BB, 0x0007D5BE0007DDBD, 0x0007BDA30007CDBF,
-    0x8000A30C8000A308, 0x8000A3148000A310, 0x0007DDA50007D9A6,
-    0x0007FDAB8000A318, 0x003C00860007FDAA, 0x0007DDAF00017C87,
-    0x0007DDAD0007DDAE, 0x8000A31C0007DDAC, 0x8000A320000619B3,
-    0x0006C9B68000A324, 0x8000A32C8000A328, 0x000169BA000631B5,
-    0x8000A330000169BB, 0x000159BE8000A334, 0x000169BC000159BF,
-    0x0006CDA8000161BD, 0x0006D5AA0006D5A9, 0x0006FDAC0006E5AB,
-    0x000605AE00060DAD, 0x8000A338000199AF, 0x8000A3408000A33C,
-    0x8000A3488000A344, 0x8000A3508000A34C, 0x00018DB88000A354,
-    0x000191BA000191B9, 0x0001BDBC0001A5BB, 0x000155BE00015DBD,
-    0x0001E9B000014DBF, 0x0001F9B20001E9B1, 0x0001D9B40001F1B3,
-    0x0001B9B60001D9B5, 0x0003CCBE0001B5B7, 0x000565A500059DA6,
-    0x00054DA38000A358, 0x8000A3608000A35C, 0x00020DAF8000A364,
-    0x000235AD00020DAE, 0x00023DAB00023DAC, 0x8000A36800023DAA,
-    0x8000A3708000A36C, 0x8000A3788000A374, 0x000019808000A37C,
-    0x0000058200001981, 0x000260848000A380, 0x000348878000A388,
-    0x8000A38C003CCC86, 0x8000A3948000A390, 0x8000A39C8000A398,
-    0x003CC0848000A3A0, 0x8000A3A4000000A6, 0x8000A3A80001A4E1,
-    0x8000A3AC0002ECE3, 0x8000A3B48000A3B0, 0x8000A3B8000254EF,
-    0x8000A3C08000A3BC, 0x8000A3C88000A3C4, 0x8000A3D08000A3CC,
-    0x8000A3D4000385B3, 0x000381B6000381B5, 0x8000A3DC8000A3D8,
-    0x000349BA8000A3E0, 0x000359BC000349BB, 0x000349BE000359BD,
-    0x000688E1000349BF, 0x000694E30007ACE1, 0x8000A3E400063CE3,
-    0x8000A3EC8000A3E8, 0x000035808000A3F0, 0x0000358200003D81,
-    0x003C9086003D60BE, 0x00074CEF003DF887, 0x000209A3000744EF,
-    0x8000A3FC8000A3F8, 0x8000A4048000A400, 0x00020DA500020DA6,
-    0x0002C5AB8000A408, 0x8000A40C0002C5AA, 0x0002C5AF8000A410,
-    0x0002D5AD0002C5AE, 0x000511BF0002D5AC, 0x00050DBD00050DBE,
-    0x0005F9BB0005E1BC, 0x0005CDB90005CDBA, 0x0005E9B70005D1B8,
-    0x000581B5000589B6, 0x0005A1B30005B1B4, 0x0005B1B10005B9B2,
-    0x00067DAF0005B9B0, 0x00067DAD00067DAE, 0x00067DAB00067DAC,
-    0x00067DA900067DAA, 0x8000A41400067DA8, 0x8000A41C8000A418,
-    0x8000A4248000A420, 0x8000A42C8000A428, 0x003E65A88000A430,
-    0x003E7DAA003E75A9, 0x003E6DAC003E75AB, 0x003E9DAE003E95AD,
-    0x8000A434003E8DAF, 0x8000A43C8000A438, 0x8000A4448000A440,
-    0x8000A44C8000A448, 0x003E81B88000A450, 0x003E81BA003E81B9,
-    0x003E8DBC003E9DBB, 0x003EBDBE003EB5BD, 0x003EF9B0003F4DBF,
-    0x003ECDB2003EC5B1, 0x003EC1B4003EC5B3, 0x003EC1B6003EC1B5,
-    0x8000A384003EC1B7, 0x8000A454003EF1B3, 0x003E11B68000A3F4,
-    0x8000A45C8000A458, 0x003E39BA003EE1B5, 0x8000A460003E05BB,
-    0x003E01BE0003E084, 0x003E1DBC003E01BF, 0x003EB5A3003E15BD,
-    0x0000518100006D82, 0x000398BE00005980, 0x003EA5A5003E55A6,
-    0x003E41AB8000A464, 0x00004886003E7DAA, 0x003E45AF00006887,
-    0x003E51AD003E45AE, 0x003E45A8003E59AC, 0x003E45AA003E4DA9,
-    0x003E4DAC003E5DAB, 0x003E75AE003E7DAD, 0x8000A468003EFDAF,
-    0x00016C848000A46C, 0x8000A4748000A470, 0x8000A47C8000A478,
-    0x000165B88000A480, 0x000165BA00016DB9, 0x000165BC00017DBB,
-    0x000165BE00016DBD, 0x003E85B00001D5BF, 0x003E85B2003E8DB1,
-    0x003E85B4003E9DB3, 0x003E85B6003E8DB5, 0x003ECDB300015DB7,
-    0x8000A4888000A484, 0x8000A4908000A48C, 0x003EC5B5003EFDB6,
-    0x003E49BB8000A494, 0x8000A498003E41BA, 0x003E49BF8000A49C,
-    0x003E49BD003E41BE, 0x8000A4A0003E51BC, 0x8000A4A4003E89A3,
-    0x003EB9A68000A4A8, 0x8000A4B08000A4AC, 0x003E05AA003E81A5,
-    0x8000A4B4003E0DAB, 0x003E05AE8000A4B8, 0x003E15AC003E0DAF,
-    0x000321A8003E0DAD, 0x000325AA000321A9, 0x000325AC00033DAB,
-    0x000325AE00032DAD, 0x8000A4BC000399AF, 0x000264BE8000A4C0,
-    0x8000A4C88000A4C4, 0x8000A4D08000A4CC, 0x00038DB88000A4D8,
-    0x000391BA000391B9, 0x0003BDBC0003A5BB, 0x00007DBE000075BD,
-    0x0003E9B0000071BF, 0x0003F9B20003E9B1, 0x0003D9B40003F1B3,
-    0x0003B9B60003D9B5, 0x0000B5810003B5B7, 0x0003A5B30000B580,
-    0x0003A5B50000B582, 0x8000A4E08000A4DC, 0x000400860003A5B6,
-    0x000325BB0003D487, 0x00032DBD000325BA, 0x000315BF000325BC,
-    0x0003EDA3000315BE, 0x8000A4E88000A4E4, 0x8000A4EC00116085,
-    0x0003EDA50003EDA6, 0x00036DAB8000A4F0, 0x8000A4F400036DAA,
-    0x00035DAF8000A4F8, 0x000365AD00035DAE, 0x0001BCE300036DAC,
-    0x003E20E1003F70E1, 0x8000A4FC003F94E3, 0x8000A5048000A500,
-    0x8000A50C8000A508, 0x8000A5148000A510, 0x000118EF8000A518,
-    0x001338BF8000A51C, 0x8000A520003EE8EF, 0x000504848000A524,
-    0x000068E3000504BE, 0x0001A8E18000A528, 0x0003B0EF8000A52C,
-    0x8000A5348000A530, 0x00001D808000A538, 0x0000DD8200000981,
-    0x003EC9AA8000A53C, 0x003E95A8003EC9AB, 0x003EF9AE003ECDA9,
-    0x003ECDAC003EF9AF, 0x00EFA8BF003EC5AD, 0x000488868000A540,
-    0x8000A5480004AC87, 0x8000A5508000A54C, 0x003DD9BA8000A554,
-    0x003DC5B8003DD9BB, 0x003D91BE003DD9B9, 0x003DE1BC003D91BF,
-    0x003E49B2003DE1BD, 0x003E49B0003E49B3, 0x003E49B6003E49B1,
-    0x003E49B4003E49B7, 0x00063DA8003E49B5, 0x000691AA000691A9,
-    0x0006D9AC0006C9AB, 0x0006C5AE0006C5AD, 0x8000A5440006F5AF,
-    0x8000A55C8000A558, 0x8000A5648000A560, 0x8000A56C8000A568,
-    0x00076DB88000A570, 0x000709BA000771B9, 0x000705BC000709BB,
-    0x00073DBE000709BD, 0x00068DB0000735BF, 0x00077DB2000775B1,
-    0x00076DB4000775B3, 0x00075DB6000755B5, 0x000619B3000755B7,
-    0x8000A5788000A574, 0x8000A5808000A57C, 0x000651B5000645B6,
-    0x0007B1BB8000A584, 0x8000A588000649BA, 0x000791BF8000A58C,
-    0x0007A1BD0007A9BE, 0x8000A5900007A9BC, 0x8000A59400065DA3,
-    0x000601A68000A598, 0x8000A5A08000A59C, 0x00060DAA000615A5,
-    0x8000A5A40007F5AB, 0x0007EDAE8000A5A8, 0x0007EDAC0007D5AF,
-    0x0001ED800007E5AD, 0x0000198200000981, 0x8000A5B08000A5AC,
-    0x0001C4BE0001D084, 0x000080868000A5B4, 0x8000A5B800012087,
-    0x8000A5C08000A5BC, 0x8000A5C88000A5C4, 0x8000A5D08000A5CC,
-    0x8000A5D40005FCE1, 0x8000A5D8000414E3, 0x8000A5E08000A5DC,
-    0x8000A5E88000A5E4, 0x8000A5F08000A5EC, 0x8000A5F88000A5F4,
-    0x8000A6008000A5FC, 0x000631A80004C8EF, 0x000631AA000631A9,
-    0x0001B5AC000631AB, 0x0001B5AE0001BDAD, 0x0001A8840001ADAF,
-    0x8000A6088000A604, 0x8000A6108000A60C, 0x8000A6188000A614,
-    0x000169B88000A61C, 0x000109BA000169B9, 0x00011DBC000105BB,
-    0x000101BE000101BD, 0x0001D5B0000135BF, 0x0001D5B20001DDB1,
-    0x000179B4000169B3, 0x000169B6000179B5, 0x0003BD81000161B7,
-    0x0005B1B30003BD80, 0x000235B500001982, 0x000354BE8000A620,
-    0x8000A628000235B6, 0x000211BB8000A62C, 0x0002B5BD000219BA,
-    0x0002B5BF000209BC, 0x8000A6300002B5BE, 0x00026DA500026DA6,
-    0x0005E9A3000C0084, 0x000CF8868000A634, 0x0002EDAF00037C87,
-    0x0002EDAD0002EDAE, 0x000249AB000251AC, 0x8000A638000241AA,
-    0x8000A6408000A63C, 0x8000A6488000A644, 0x000120E38000A64C,
-    0x0001A0E18000A650, 0x000240EF8000A654, 0x8000A65C8000A658,
-    0x8000A6648000A660, 0x8000A66C8000A668, 0x8000A4D48000A670,
-    0x8000A6740003F5B3, 0x0003EDB68000A678, 0x8000A6808000A67C,
-    0x0003B1BA0003E5B5, 0x8000A6840003B1BB, 0x000391BE8000A688,
-    0x0003A1BC000391BF, 0x000229A80003A1BD, 0x000239AA000229A9,
-    0x000229AC000239AB, 0x000259AE000229AD, 0x0001CD80000255AF,
-    0x0000198200000981, 0x8000A6908000A68C, 0x000DB487000D7CBE,
-    0x000231B8000C1C86, 0x000231BA000231B9, 0x0002F9BC0002E9BB,
-    0x0002E9BE0002F9BD, 0x000231B00002E9BF, 0x000231B2000231B1,
-    0x000211B4000231B3, 0x000211B6000211B5, 0x8000A698000211B7,
-    0x8000A6A08000A69C, 0x8000A6A88000A6A4, 0x8000A6B08000A6AC,
-    0x8000A6B400007CEF, 0x8000A6B8000614E1, 0x000000EC0000A8E3,
-    0x8000A6C08000A6BC, 0x0002F1A38000A6C4, 0x8000A6CC8000A6C8,
-    0x8000A6D48000A6D0, 0x0002E1A50002E9A6, 0x0002B5AB8000A6D8,
-    0x8000A6DC0002B5AA, 0x000295AF8000A6E0, 0x0002A5AD000295AE,
-    0x000E21A80002A5AC, 0x000E49AA000E21A9, 0x000E69AC000E59AB,
-    0x000E99AE000E69AD, 0x8000A694000E99AF, 0x8000A6E88000A6E4,
-    0x8000A6F08000A6EC, 0x8000A6F88000A6F4, 0x000EF5B88000A6FC,
-    0x000EF5BA000EFDB9, 0x000E9DBC000E85BB, 0x000E81BE000E81BD,
-    0x000EE9B0000EB5BF, 0x000EFDB2000EE9B1, 0x000ED5B4000EF1B3,
-    0x000ED5B6000EDDB5, 0x000EC1B3000ECDB7, 0x0000B58100001582,
-    0x8000A7000000B580, 0x000EE1B5000EE9B6, 0x000E2DBB000010BE,
-    0x00036484000E2DBA, 0x000E19BF8000A704, 0x000E25BD000E11BE,
-    0x8000A708000E29BC, 0x00008886000E85A3, 0x000EADA600036C87,
-    0x8000A7108000A70C, 0x000E69AA000EA5A5, 0x8000A714000E69AB,
-    0x000E55AE8000A718, 0x000E6DAC000E5DAF, 0x000E89B3000E61AD,
-    0x8000A7208000A71C, 0x8000A7288000A724, 0x000E89B5000E81B6,
-    0x000E95BB8000A72C, 0x8000A730000E8DBA, 0x000EF9BF8000A734,
-    0x000E85BD000EF1BE, 0x8000A738000E85BC, 0x8000A7408000A73C,
-    0x000D10E18000A744, 0x000D04E38000A748, 0x8000A7508000A74C,
-    0x000CA4EF8000A754, 0x8000A75C8000A758, 0x8000A7648000A760,
-    0x000E01A88000A768, 0x000E01AA000E01A9, 0x000E01AC000E01AB,
-    0x000E01AE000E01AD, 0x0000DD80000E3DAF, 0x0000198200000981,
-    0x8000A7708000A76C, 0x000074BE00013C84, 0x000E1DB88000A778,
-    0x000E25BA000E2DB9, 0x0001F9BC0001E9BB, 0x0001E9BE0001F9BD,
-    0x000E49B00001E9BF, 0x000E51B2000E51B1, 0x000E35B4000E51B3,
-    0x000E35B6000E3DB5, 0x000D89A3000E2DB7, 0x0004A8868000A77C,
-    0x8000A78000033C87, 0x000D89A5000D81A6, 0x000D95AB8000A784,
-    0x8000A788000D8DAA, 0x000DF9AF8000A78C, 0x000D85AD000DF1AE,
-    0x8000A790000D85AC, 0x0003488400029DB3, 0x000355B6000440BE,
-    0x8000A7988000A794, 0x000371BA0002B1B5, 0x8000A79C000379BB,
-    0x000331BE8000A7A0, 0x000351BC000331BF, 0x000391B0000351BD,
-    0x0003A1B2000399B1, 0x0003D1B40003A1B3, 0x0003D1B60003D1B5,
-    0x0003F5B80003D1B7, 0x0003C9BA0003F9B9, 0x0003DDBC0003C9BB,
-    0x0003CDBE0003C5BD, 0x8000A7A40003C5BF, 0x8000A7AC8000A7A8,
-    0x001AEC858000A7B0, 0x8000A7B48000A624, 0x000221A88000A7B8,
-    0x000231AA000231A9, 0x00021DAC000205AB, 0x0003F1AE0003F1AD,
-    0x8000A7BC0003F1AF, 0x8000A7C48000A7C0, 0x000050EF8000A7C8,
-    0x8000A7D08000A7CC, 0x000084E38000A7D4, 0x0001F8E18000A7D8,
-    0x000015808000A7DC, 0x0000058200001981, 0x000399A38000A7E0,
-    0x000468868000A7E8, 0x8000A7EC00056087, 0x0003B5A5000251A6,
-    0x00027DAB8000A7F0, 0x8000A7F4000275AA, 0x000235AF8000A7F8,
-    0x000255AD000235AE, 0x8000A7FC000255AC, 0x8000A8048000A800,
-    0x8000A80C8000A808, 0x8000A8148000A810, 0x00046CBE0001F8EF,
-    0x8000A818000E80E1, 0x8000A81C000114E3, 0x8000A8248000A820,
-    0x8000A82C8000A828, 0x8000A8348000A830, 0x8000A8380001DDB3,
-    0x0001F5B60001FDB5, 0x8000A8408000A83C, 0x0001B1BA8000A844,
-    0x00019DBC000185BB, 0x00013DBE000135BD, 0x0005BDA8000135BF,
-    0x000619AA000601A9, 0x000639AC000611AB, 0x000659AE000639AD,
-    0x8000A7E4000649AF, 0x00001D8100001D82, 0x8000A84800001D80,
-    0x8000A8508000A84C, 0x00069DB88000A854, 0x00074DBA000745B9,
-    0x00075DBC000745BB, 0x000741BE000741BD, 0x000639B0000741BF,
-    0x000605B200060DB1, 0x0006B5B40006ADB3, 0x0006B5B60006BDB5,
-    0x000691A30006A5B7, 0x0000288600028C84, 0x8000A8580003C087,
-    0x0006B1A50006B9A6, 0x0006C9AB8000A85C, 0x8000A8600006FDAA,
-    0x000679AF8000A864, 0x000679AD000671AE, 0x8000A8680006D1AC,
-    0x8000A86C000799B3, 0x000751B68000A870, 0x8000A8788000A874,
-    0x00076DBA0007B1B5, 0x8000A87C000775BB, 0x000759BE8000A880,
-    0x00076DBC000745BF, 0x8000A884000765BD, 0x8000A88C8000A888,
-    0x8000A8948000A890, 0x8000A8980005A0EF, 0x8000A8A08000A89C,
-    0x8000A8A400057CE3, 0x8000A8A800053CE1, 0x8000A8B08000A8AC,
-    0x00061DA38000A8B4, 0x0000158100002D82, 0x8000A8B800001D80,
-    0x000635A50006D5A6, 0x0006F1AB8000A8BC, 0x8000A8C00006E9AA,
-    0x0006C1AF00012084, 0x0006E1AD0006DDAE, 0x0006C5A80006E9AC,
-    0x0006D5AA0006D5A9, 0x0006FDAC0006E5AB, 0x00061DAE000615AD,
-    0x0001E8BE000615AF, 0x000008868000A8C8, 0x8000A8CC00008887,
-    0x8000A8D48000A8D0, 0x00067DB88000A8D8, 0x00060DBA000605B9,
-    0x000601BC000605BB, 0x000639BE000609BD, 0x00066DB0000639BF,
-    0x00067DB2000675B1, 0x000659B4000675B3, 0x00064DB6000645B5,
-    0x000291A8000645B7, 0x0002A1AA000299A9, 0x0002D1AC0002A1AB,
-    0x0002D5AE0002DDAD, 0x8000A8DC0002C9AF, 0x8000A8E48000A8E0,
-    0x8000A8E800C524BF, 0x8000A8F08000A8EC, 0x000359B88000A8F4,
-    0x000369BA000359B9, 0x000361BC000365BB, 0x000361BE000361BD,
-    0x0002B9B0000361BF, 0x000285B200028DB1, 0x000375B400036DB3,
-    0x000375B600037DB5, 0x000229B300036DB7, 0x8000A8FC8000A8F8,
-    0x000384848000A900, 0x000259B500025DB6, 0x000249BB8000A904,
-    0x8000A77400027DBA, 0x0003B9BF8000A908, 0x0003B9BD0003B1BE,
-    0x8000A90C000251BC, 0x8000A91000026DA3, 0x000219A68000A914,
-    0x8000A91C8000A918, 0x000239AA00021DA5, 0x8000A92000020DAB,
-    0x0003F5AE8000A924, 0x000215AC0003FDAF, 0x0002A5A80003FDAD,
-    0x0002BDAA0002B5A9, 0x0002ADAC0002B5AB, 0x00013DAE000135AD,
-    0x00001181000135AF, 0x8000A92800000D80, 0x8000A92C00001982,
-    0x001DB4BE8000A930, 0x0001EDB88000A938, 0x000181BA000185B9,
-    0x000185BC000181BB, 0x0001B1BE00018DBD, 0x000151B00001B1BF,
-    0x000151B2000151B1, 0x0001F5B4000151B3, 0x0001F5B60001FDB5,
-    0x0000D8EF0001E1B7, 0x001DA8868000A93C, 0x8000A940001DDC87,
-    0x8000A9488000A944, 0x8000A94C0006C4EF, 0x8000A9500006F4E1,
-    0x0000CCE3000138E3, 0x000128E18000A954, 0x00027DB58000A958,
-    0x8000A9608000A95C, 0x8000A9640001EDB6, 0x00026DB38000A968,
-    0x0001DDBD8000A96C, 0x000161BF0001DDBC, 0x8000A970000169BE,
-    0x0001CDBB8000A974, 0x001DB5A30001CDBA, 0x8000A9788000A934,
-    0x8000A9808000A97C, 0x001DA5A5001E35A6, 0x001E15AB8000A984,
-    0x8000A988001E15AA, 0x001EB9AF8000A98C, 0x001E05AD001EB1AE,
-    0x8000A990001E05AC, 0x8000A9988000A994, 0x000019808000A99C,
-    0x0000058200001981, 0x0003EC848000A9A0, 0x8000A9A4000368BE,
-    0x0001D0868000A9A8, 0x8000A9AC0000B487, 0x001FBDA88000A9B0,
-    0x001F19AA001F01A9, 0x001F09AC001F19AB, 0x001F31AE001F35AD,
-    0x8000A9B4001F29AF, 0x8000A9BC8000A9B8, 0x8000A9C48000A9C0,
-    0x8000A9CC8000A9C8, 0x0000C9B88000A9D0, 0x0000D9BA0000C9B9,
-    0x0000F9BC0000D1BB, 0x000099BE0000F9BD, 0x001F5DB0000099BF,
-    0x001F21B2001F21B1, 0x001F25B4001F3DB3, 0x001F09B6001F29B5,
-    0x001CC0EF001F09B7, 0x8000A9D40019C0E1, 0x8000A9D80009D0E3,
-    0x8000A9E08000A9DC, 0x000125A38000A9E4, 0x001D15A1000399A2,
-    0x001DB0E3001D0DA0, 0x001E88E18000A9E8, 0x001ED1B3001060EF,
-    0x8000A9F08000A9EC, 0x8000A9F88000A9F4, 0x001EC1B5001EF5B6,
-    0x001EA5BB8000A9FC, 0x8000AA00001ED9BA, 0x001EADBF8000AA04,
-    0x001EA9BD001EA1BE, 0x001E11A3001EB5BC, 0x8000AA0C8000AA08,
-    0x0024198F8000AA10, 0x001E01A5001E35A6, 0x001E65AB000710BE,
-    0x00040084001E19AA, 0x001E6DAF8000AA18, 0x001E69AD001E61AE,
-    0x0014DD9B001E75AC, 0x001709990015E59A, 0x001F799F00111198,
-    0x0019F99D0019799E, 0x002D6193001B099C, 0x00296D9100289192,
-    0x00117D9700296D90, 0x002CE99500117596, 0x00042086002D7994,
-    0x0027818200039487, 0x002A9584002A8D83, 0x002E9986002B6185,
-    0x0004A884002E9987, 0x002F398A8000AA1C, 0x0012BD8C0012AD8B,
-    0x0014F98E0012E98D, 0x8000AA2000161D8F, 0x001755928000AA24,
-    0x001A719400189993, 0x001CF196001AF995, 0x8000AA28001E7D97,
-    0x001E559A8000AA2C, 0x0002759C0000D19B, 0x8000AA348000AA30,
-    0x00F685AF8000AA38, 0x00F2F1AD00F401AE, 0x00F001AB00F289AC,
-    0x000E9DA9000E85AA, 0x000A75A7000E9DA8, 0x000801A5000A85A6,
-    0x0006B9A3000601A4, 0x0002D1A10006B1A2, 0x000345808000A8C4,
-    0x0003458200034D81, 0x8000AA408000AA3C, 0x8000AA488000AA44,
-    0x8000AA508000AA4C, 0x00FC01B58000AA54, 0x00FAF5B300FBC1B4,
-    0x00F7BDB100FAF5B2, 0x00036DA800F6F5B0, 0x000375AA00037DA9,
-    0x0003B5AC0003ADAB, 0x0003B5AE0003BDAD, 0x8000AA580003ADAF,
-    0x8000AA608000AA5C, 0x8000AA688000AA64, 0x8000AA708000AA6C,
-    0x000275B88000AA74, 0x000201BA000201B9, 0x000201BC000201BB,
-    0x000231BE000209BD, 0x0003D5B0000231BF, 0x0003D5B20003DDB1,
-    0x000261B4000265B3, 0x000261B6000261B5, 0x001E10E1000261B7,
-    0x0001F8E38000AA78, 0x00001580000400BE, 0x0000398200000981,
-    0x8000AA808000AA7C, 0x000488868000AA88, 0x8000AA8C00050C87,
-    0x000124EF8000AA90, 0x8000AA988000AA94, 0x8000AAA08000AA9C,
-    0x8000AAA88000AAA4, 0x8000AAAC0000B0EF, 0x8000AAB48000AAB0,
-    0x8000AAB800005CE3, 0x8000AABC000184E1, 0x8000AAC48000AAC0,
-    0x00047DA38000AAC8, 0x8000AA848000AA14, 0x8000AAD08000AACC,
-    0x0004BDA50004BDA6, 0x000499AB8000AAD4, 0x8000AAD8000491AA,
-    0x000525AF8000AADC, 0x000531AD000525AE, 0x8000AAE0000481AC,
-    0x8000AAE4000605B3, 0x000635B68000AAE8, 0x8000AAF08000AAEC,
-    0x00061DBA00063DB5, 0x0003E0840006A9BB, 0x0006A9BE0000E0BE,
-    0x0006A9BC0006A9BF, 0x00052DA80006A9BD, 0x0006C9AA0006A1A9,
-    0x0006F1AC0006C9AB, 0x0006EDAE0006F1AD, 0x0000ED800006E5AF,
-    0x0000F1820000F181, 0x000090868000AAF4, 0x8000AAF80000F087,
-    0x0007E1B88000AAFC, 0x0007E1BA0007E1B9, 0x0007E1BC0007E1BB,
-    0x0007E1BE0007E1BD, 0x00069DB00007E1BF, 0x000639B2000629B1,
-    0x000629B4000639B3, 0x000619B6000629B5, 0x000741A3000619B7,
-    0x8000AB048000AB00, 0x8000AB0C8000AB08, 0x000779A5000771A6,
-    0x0007EDAB8000AB10, 0x8000AB14000759AA, 0x0007EDAF8000AB18,
-    0x0007EDAD0007EDAE, 0x8000AB1C0007EDAC, 0x8000AB200006F1B3,
-    0x000695B68000AB24, 0x8000AB2C8000AB28, 0x0006B1BA00069DB5,
-    0x8000AB300006BDBB, 0x000699BE8000AB34, 0x0006A5BC000685BF,
-    0x000625A800069DBD, 0x00063DAA000635A9, 0x000651AC000631AB,
-    0x000651AE000651AD, 0x8000AB38000651AF, 0x00012C848000AB3C,
-    0x8000AB448000AB40, 0x8000AB4C8000AB48, 0x00063DB88000AB50,
-    0x0001D9BA0001C9B9, 0x0001F9BC0001D1BB, 0x000199BE0001F9BD,
-    0x000631B0000195BF, 0x000635B200063DB1, 0x000619B400060DB3,
-    0x00060DB6000605B5, 0x0005B5A3000605B7, 0x8000AB588000AB54,
-    0x8000AB608000AB5C, 0x0005D9A50005D1A6, 0x0005F9AB000358BE,
-    0x000C88860005F5AA, 0x0005C1AF0003BC87, 0x0005D9AD0005DDAE,
-    0x00004D810005E1AC, 0x00034DB300004D80, 0x00035DB500005D82,
-    0x8000AB6C8000AB68, 0x8000AB7000034DB6, 0x000329BB8000AB74,
-    0x00033DBD000341BA, 0x000321BF00033DBC, 0x000239A8000329BE,
-    0x00038DAA000239A9, 0x00039DAC000385AB, 0x000385AE000385AD,
-    0x000CEC840003B5AF, 0x000CE0BE8000AB78, 0x8000AB808000AB7C,
-    0x8000AB888000AB84, 0x000361B88000AB8C, 0x000365BA000361B9,
-    0x000365BC00037DBB, 0x000365BE00036DBD, 0x0003CDB00000D9BF,
-    0x0003A1B20003A5B1, 0x0003A5B40003A1B3, 0x000391B60003ADB5,
-    0x8000AB90000391B7, 0x8000AB988000AB94, 0x0003F4EF8000AB9C,
-    0x8000ABA48000ABA0, 0x000098E38000ABA8, 0x000180E18000ABAC,
-    0x8000ABB48000ABB0, 0x8000ABBC8000ABB8, 0x0007F5BA8000ABC0,
-    0x8000ABC40007FDBB, 0x000641BE8000ABC8, 0x0007E5BC000641BF,
-    0x8000ABCC000655BD, 0x8000ABD0000719B3, 0x0007D9B68000ABD8,
-    0x8000ABE08000ABDC, 0x000798E10007D9B5, 0x0007B8E30001FCE3,
-    0x0000DD8000063CE1, 0x0000F1820000F581, 0x8000ABE88000ABE4,
-    0x8000ABF08000ABEC, 0x8000ABF88000ABF4, 0x0006CCEF000120EF,
-    0x000CE0868000ABFC, 0x0002CDA3000D0C87, 0x0002DDA58000AC00,
-    0x8000AC088000AC04, 0x8000AC0C0002CDA6, 0x0002A9AB8000AC10,
-    0x0002BDAD0002C1AA, 0x0002A1AF0002BDAC, 0x000E55A80002A9AE,
-    0x000E69AA000E5DA9, 0x000EB5AC000E69AB, 0x000EA9AE000EB9AD,
-    0x8000ABD4000EA9AF, 0x8000AC188000AC14, 0x8000AC208000AC1C,
-    0x8000AC288000AC24, 0x000F5DB88000AC2C, 0x000F65BA000F61B9,
-    0x000F65BC000F7DBB, 0x000F65BE000F6DBD, 0x000ED9B0000F15BF,
-    0x000F69B2000ED9B1, 0x000F79B4000F69B3, 0x000F6DB6000F79B5,
-    0x000EE1B3000F65B7, 0x8000AC348000AC30, 0x8000AC3C8000AC38,
-    0x000E3DB5000E39B6, 0x000E1DBB8000AC40, 0x00032484000E1DBA,
-    0x000FBDBF00002CBE, 0x000FBDBD000FBDBE, 0x00005D82000FBDBC,
-    0x00004580000EA5A3, 0x000E7DA600005D81, 0x8000AC488000AC44,
-    0x000E59AA000E79A5, 0x00040086000E59AB, 0x000FF9AE00035C87,
-    0x000FF9AC000FF9AF, 0x8000AC4C000FF9AD, 0x8000AC548000AC50,
-    0x8000AC5C8000AC58, 0x8000AC60000D44EF, 0x8000AC688000AC64,
-    0x8000AC6C000D94E3, 0x8000AC70000D1CE1, 0x8000AC788000AC74,
-    0x000E39B38000AC7C, 0x8000AC848000AC80, 0x8000AC8C8000AC88,
-    0x000EF9B5000EF1B6, 0x000EDDBB8000AC90, 0x8000AC94000ED5BA,
-    0x000E91BF8000AC98, 0x000EADBD000EADBE, 0x000EEDA8000EB1BC,
-    0x000EF1AA000EF1A9, 0x000E31AC000E11AB, 0x000E31AE000E31AD,
-    0x8000AC9C000E31AF, 0x8000ACA48000ACA0, 0x00001D808000ACA8,
-    0x0003B98200000981, 0x0001E9B88000ACAC, 0x0001F9BA0001E9B9,
-    0x0001E9BC0001F9BB, 0x00015DBE0001E9BD, 0x000E51B0000155BF,
-    0x000E25B2000E59B1, 0x000E29B4000E39B3, 0x000E1DB6000E15B5,
-    0x8000ACB00001D9B7, 0x000204BE00029C84, 0x8000ACB8000D75A3,
-    0x000DBDA6000DB5A5, 0x8000ACC08000ACBC, 0x000D99AA8000ACC4,
-    0x000DFDAC000D91AB, 0x000DE1AE000DE1AD, 0x000351B6000DDDAF,
-    0x000380848000ACC8, 0x8000ACCC000341B5, 0x0004D886000351B3,
-    0x000359BE0002A887, 0x000359BC000359BF, 0x000351BA000359BD,
-    0x8000ACD000035DBB, 0x000385A38000ACD4, 0x8000ACDC8000ACD8,
-    0x8000ACE48000ACE0, 0x000395A5000385A6, 0x000389AB8000ACE8,
-    0x8000AB64000385AA, 0x00038DAF8000ACEC, 0x00038DAD00038DAE,
-    0x8000ACF000038DAC, 0x00F5B4858000ACF4, 0x8000ACFC8000ACF8,
-    0x8000AD048000AD00, 0x8000AD0C8000AD08, 0x8000AD148000AD10,
-    0x000009818000AD18, 0x8000AD1C00001980, 0x8000AD200001FD82,
-    0x000248BE000125B3, 0x000141B68000AD28, 0x8000AD308000AD2C,
-    0x0001A5BA000159B5, 0x8000AD340001ADBB, 0x0001ADBE8000AD38,
-    0x0001B5BC000199BF, 0x000560860001ADBD, 0x8000AD3C0002F487,
-    0x0001A0E18000AD40, 0x00034CE38000AD44, 0x8000AD4C8000AD48,
-    0x000398EF8000AD50, 0x8000AD588000AD54, 0x8000AD608000AD5C,
-    0x8000AD688000AD64, 0x8000AD708000AD6C, 0x8000AD7400044084,
-    0x8000AD7C8000AD78, 0x8000AD800001C8EF, 0x8000AD84000E7CE1,
-    0x00000D80000114E3, 0x00007D8200007581, 0x8000AD8C8000AD88,
-    0x8000AD988000AD90, 0x8000AD9C0001EDA3, 0x000189A6000191A5,
-    0x0004C086000414BE, 0x00016DAA00044487, 0x00017DAC000165AB,
-    0x000165AE000165AD, 0x8000ACB4000151AF, 0x8000ADA08000AD24,
-    0x8000AD948000ADA4, 0x8000ADAC8000ADA8, 0x000631A88000ADB0,
-    0x000609AA000639A9, 0x000675AC000609AB, 0x000641AE000665AD,
-    0x0006C5B0000641AF, 0x0006C5B20006CDB1, 0x0006CDB40006DDB3,
-    0x000699B60006F9B5, 0x0006ADB8000699B7, 0x0006BDBA0006B5B9,
-    0x0006ADBC0006B5BB, 0x00069DBE000695BD, 0x00078DA8000695BF,
-    0x000791AA000791A9, 0x0007BDAC0007A5AB, 0x0007D5AE0007DDAD,
-    0x8000ADB40007C9AF, 0x8000ADBC8000ADB8, 0x8000ADC48000ADC0,
-    0x8000ADCC8000ADC8, 0x000741B88000ADD0, 0x000741BA000741B9,
-    0x000745BC00075DBB, 0x000779BE00074DBD, 0x0007BDB0000775BF,
-    0x00078DB2000785B1, 0x000781B4000781B3, 0x000781B6000781B5,
-    0x000619B3000781B7, 0x8000ADD88000ADD4, 0x8000ADE08000ADDC,
-    0x000635B500065DB6, 0x000641BB8000ADE4, 0x8000ADE8000679BA,
-    0x000645BF8000ADEC, 0x000655BD000645BE, 0x8000ADF0000659BC,
-    0x8000ADF400065DA3, 0x000619A68000ADF8, 0x8000AE008000ADFC,
-    0x00063DAA000671A5, 0x00012484000605AB, 0x000601AE000124BE,
-    0x00061DAC000601AF, 0x000649A8000611AD, 0x000181AA000655A9,
-    0x000185AC000181AB, 0x000185AE00018DAD, 0x0000ED8000013DAF,
-    0x0000F1820000F181, 0x000010868000AE08, 0x8000AE0C00009487,
-    0x0000C9B88000AE10, 0x0000D9BA0000C9B9, 0x0000F9BC0000D1BB,
-    0x000099BE0000F9BD, 0x000145B0000095BF, 0x000145B200014DB1,
-    0x000145B400015DB3, 0x000145B600014DB5, 0x00075DBA0000F9B7,
-    0x000745B800076DBB, 0x0007B5BE00075DB9, 0x000775BC0007BDBF,
-    0x000711B20007BDBD, 0x000709B0000711B3, 0x000775B6000705B1,
-    0x000701B400077DB7, 0x000745AA000701B5, 0x000691A800074DAB,
-    0x000749AE00074DA9, 0x000755AC000749AF, 0x8000AE1400074DAD,
-    0x8000AE1C8000AE18, 0x8000AE248000AE20, 0x8000AE2C8000AE28,
-    0x000289B38000AE30, 0x8000AE388000AE34, 0x8000AE408000AE3C,
-    0x00028DB500028DB6, 0x000209BB8000AE44, 0x8000AE48000201BA,
-    0x000209BF8000AE4C, 0x000209BD000201BE, 0x8000AE50000211BC,
-    0x8000AE540002CDA3, 0x0002C9A68000AE58, 0x8000AE608000AE5C,
-    0x000245AA0002C9A5, 0x8000AE6400024DAB, 0x000245AE8000AE68,
-    0x000255AC00024DAF, 0x0002ADA800024DAD, 0x0002B5AA0002BDA9,
-    0x000175AC00016DAB, 0x000175AE00017DAD, 0x0001CD8000016DAF,
-    0x0000198200000981, 0x8000AE708000AE6C, 0x00FDB4BE00023484,
-    0x0001E9B88000AE78, 0x000189BA0001E9B9, 0x00019DBC000185BB,
-    0x000181BE000181BD, 0x000115B00001B5BF, 0x000115B200011DB1,
-    0x0001F9B40001E9B3, 0x0001E9B60001F9B5, 0x0006D8E10001E1B7,
-    0x000614E30007C4E1, 0x8000AE7C0007D4E3, 0x8000AE8000066CE1,
-    0x00FC2C840000E4E3, 0x00FCA4BE8000AE84, 0x00FD9086000038EF,
-    0x00069CEF00FDD487, 0x8000AE8800064CEF, 0x00FCC8848000AE8C,
-    0x8000AE90000271B3, 0x0001B1B60001B1B5, 0x8000AE988000AE94,
-    0x00019DBA8000AE9C, 0x00018DBC000195BB, 0x000171BE000171BD,
-    0x00FDBDA3000171BF, 0x8000AEA08000AE74, 0x8000AEA88000AEA4,
-    0x00FE7DA500FE7DA6, 0x00FE59AB8000AEAC, 0x8000AEB000FE51AA,
-    0x00FEBDAF8000AEB4, 0x00FEBDAD00FEBDAE, 0x8000AEB800FE41AC,
-    0x8000AEC08000AEBC, 0x8000AEC88000AEC4, 0x8000AED08000AECC,
-    0x00000D808000AED4, 0x00001D8200001581, 0x8000AEDC8000AED8,
-    0x8000AEE48000AEE0, 0x00FFBDA88000AEE8, 0x00FF11AA00FF15A9,
-    0x00FF51AC00FF29AB, 0x00FF49AE00FF51AD, 0x0000008600FF41AF,
-    0x8000AEEC00036487, 0x8000AEF48000AEF0, 0x8000AEFC8000AEF8,
-    0x000081B88000AF00, 0x000081BA000081B9, 0x000081BC000081BB,
-    0x000081BE000081BD, 0x00FF01B0000081BF, 0x0000E1B200FF01B1,
-    0x0000E1B40000E1B3, 0x0000D1B60000E9B5, 0x8000AF040000D1B7,
-    0x8000AF0C8000AF08, 0x8000AF148000AF10, 0x8000AF1800FEBCE1,
-    0x8000AF1C00FC38E3, 0x8000AF248000AF20, 0x8000AF2800FCECEF,
-    0x8000AF308000AF2C, 0x00FED9B38000AF34, 0x8000AF3C8000AF38,
-    0x8000AF448000AF40, 0x00FE99B500FE91B6, 0x00FE85BB8000AF48,
-    0x8000AF4C00FE85BA, 0x00016DBF8000AF50, 0x000175BD000165BE,
-    0x00FE19A3000175BC, 0x00002886000184BE, 0x8000AF5F00015C87,
-    0x00FE59A500FE51A6, 0x00FE45AB8000AF63, 0x8000AF6700FE45AA,
-    0x0001ADAF8000AF6B, 0x0001B5AD0001A5AE, 0x00009D800001B5AC,
-    0x0000A5820000AD81, 0x8000AF6F000161B3, 0x000161B6000161B5,
-    0x8000AF778000AF73, 0x0000EDBA8000AF7B, 0x0000FDBC0000E5BB,
-    0x0000E1BE0000E5BD, 0x00027DA80000DDBF, 0x0002A5AA0002ADA9,
-    0x0002A5AC0002BDAB, 0x0002A5AE0002ADAD, 0x0004EC8400031DAF,
-    0x8000AF838000AF7F, 0x8000AF8B8000AF87, 0x8000AF938000AF8F,
-    0x000371B88000AF97, 0x000371BA000371B9, 0x000315BC000371BB,
-    0x000315BE00031DBD, 0x000365B000030DBF, 0x000365B200036DB1,
-    0x000369B400037DB3, 0x00035DB6000355B5, 0x8000AF9B000355B7,
-    0x8000AFA38000AF9F, 0x8000AE04000221A3, 0x000221A6000221A5,
-    0x8000AFAB8000AFA7, 0x0003ADAA8000AFAF, 0x0003BDAC0003A5AB,
-    0x0003A1AE0003A5AD, 0x8000AFB300039DAF, 0x8000AFBB8000AFB7,
-    0x00001D808000AFBF, 0x0000398200000981, 0x8000AFC78000AFC3,
-    0x8000AFD38000AFCF, 0x0004D086000424BE, 0x8000AFD700032087,
-    0x000291A88000AFDB, 0x0002A1AA000299A9, 0x0002C1AC0002A1AB,
-    0x0002F9AE0002C9AD, 0x8000AFDF0002F9AF, 0x8000AFE78000AFE3,
-    0x8000AFEF8000AFEB, 0x8000AFF78000AFF3, 0x0001C5B88000AFFB,
-    0x0001D5BA0001CDB9, 0x0001F5BC0001EDBB, 0x00019DBE000195BD,
-    0x000145B0000195BF, 0x000145B200014DB1, 0x000145B400015DB3,
-    0x000145B600014DB5, 0x00FE28E10001FDB7, 0x00FE38E38000AFFF,
-    0x00041C84000480BE, 0x8000B0078000B003, 0x8000B00F8000B00B,
-    0x8000B0178000B013, 0x8000B01F8000B01B, 0x00FEE8EF8000B023,
-    0x000000F68000B027, 0x8000B02B0000B8EF, 0x000009818000B02F,
-    0x8000B03300001980, 0x8000B03700003D82, 0x000748868000B03F,
-    0x8000B0430004FC87, 0x8000B04700FE2CE3, 0x0005A9B3000144E1,
-    0x8000B03B8000AFCB, 0x8000B04F8000B04B, 0x0005C9B50005C9B6,
-    0x000609BB8000B053, 0x8000B057000601BA, 0x000609BF8000B05B,
-    0x000609BD000601BE, 0x8000B05F000611BC, 0x8000B0630005EDA3,
-    0x00058DA68000B067, 0x8000B06F8000B06B, 0x000645AA00058DA5,
-    0x8000B07300064DAB, 0x000645AE8000B077, 0x000655AC00064DAF,
-    0x000665A800064DAD, 0x000671AA00066DA9, 0x0007B1AC000671AB,
-    0x0007B1AE0007B1AD, 0x0003A8840007B1AF, 0x8000B07F8000B07B,
-    0x8000B0878000B083, 0x8000B08F8000B08B, 0x000761B88000B093,
-    0x000761BA000761B9, 0x000761BC000761BB, 0x000765BE000761BD,
-    0x0007D1B0000715BF, 0x0007D1B20007D1B1, 0x000775B40007D1B3,
-    0x000761B6000761B5, 0x000690E1000761B7, 0x000538E38000B097,
-    0x00000D808000B09B, 0x00003D8200003581, 0x8000B0A38000B09F,
-    0x000128BE8000B0A7, 0x000090868000B0AF, 0x00052CEF00010487,
-    0x0007B9B38000B0B3, 0x8000B0B78000B0AB, 0x8000B0BF8000B0BB,
-    0x000679B5000671B6, 0x000655BB8000B0C3, 0x8000B0C7000655BA,
-    0x0006E5BF8000B0CB, 0x000645BD0006F1BE, 0x000779A3000645BC,
-    0x8000B0D38000B0CF, 0x8000B0DB8000B0D7, 0x0006B9A50006B1A6,
-    0x000695AB8000B0DF, 0x8000B0E3000695AA, 0x000625AF8000B0E7,
-    0x000685AD000631AE, 0x8000B0EB000685AC, 0x8000B0EF00063DB3,
-    0x0001CDB68000B0F3, 0x8000B0FB8000B0F7, 0x0001A1BA0001CDB5,
-    0x8000B0FF0001ADBB, 0x0001A9BE8000B103, 0x0001B5BC00019DBF,
-    0x0005E5A80001A9BD, 0x0005F1AA0005F5A9, 0x000201AC000201AB,
-    0x000231AE000209AD, 0x000C00BE000231AF, 0x8000B10B8000B107,
-    0x00001D808000B10F, 0x0000B98200000981, 0x0002E9B88000B113,
-    0x0002F9BA0002E9B9, 0x0002E9BC0002F9BB, 0x00035DBE0002E9BD,
-    0x000221B0000351BF, 0x000225B200022DB1, 0x000229B4000239B3,
-    0x000215B600021DB5, 0x8000B1170002D9B7, 0x000380848000AF54,
-    0x8000B11F000575A3, 0x000285A6000285A5, 0x8000B1278000B123,
-    0x0002E9AA8000B12B, 0x0002FDAC0002E5AB, 0x0002E1AE0002E1AD,
-    0x0001E8E10002D5AF, 0x000250E38000B12F, 0x000C5086000CC084,
-    0x8000B13300037087, 0x8000B13B8000B137, 0x8000B1438000B13F,
-    0x8000B14B8000B147, 0x000220EF8000B14F, 0x00022DA88000B153,
-    0x000241AA000239A9, 0x000241AC000241AB, 0x000271AE000249AD,
-    0x8000B157000271AF, 0x000CA4BE8000B15B, 0x8000B1638000B15F,
-    0x8000B16B8000B167, 0x000289B88000B16F, 0x00029DBA000289B9,
-    0x0002B9BC000295BB, 0x00017DBE0002B9BD, 0x0002E5B0000175BF,
-    0x0002E5B20002EDB1, 0x0002EDB40002FDB3, 0x0002D5B60002DDB5,
-    0x000768E10002BDB7, 0x000608E3000134E3, 0x8000B17300067CE1,
-    0x8000B17B8000B177, 0x000009818000B17F, 0x8000B18300003980,
-    0x8000B18700001D82, 0x0006CCEF000064EF, 0x8000B1938000B18F,
-    0x8000B19700028DB3, 0x00028DB68000B19B, 0x8000B1A38000B19F,
-    0x000265BA00028DB5, 0x000CC88600026DBB, 0x00026DBE000CE087,
-    0x000275BC00025DBF, 0x8000B11B00026DBD, 0x8000B1A78000B18B,
-    0x8000B1AF8000B1AB, 0x8000B1B78000B1B3, 0x000EB1A88000B1BB,
-    0x000E89AA000EB9A9, 0x000EF5AC000E89AB, 0x000EE5AE000EEDAD,
-    0x000EA5B0000EDDAF, 0x000EA5B2000EADB1, 0x000EA5B4000EBDB3,
-    0x000F79B6000EA9B5, 0x000F4DB8000F79B7, 0x000F5DBA000F55B9,
-    0x000F4DBC000F55BB, 0x000FEDBE000FE5BD, 0x8000B1BF000FE5BF,
-    0x000D0DAA8000B1C3, 0x000D1DAC000D05AB, 0x000D05AE000D05AD,
-    0x8000B1C7000D35AF, 0x8000B1CF8000B1CB, 0x8000B1D3000DE5A3,
-    0x000DE5A6000DE5A5, 0x000E61B38000B1D7, 0x8000B1DF8000B1DB,
-    0x8000B1E78000B1E3, 0x000E1DB5000E0DB6, 0x000E09BB8000B1EB,
-    0x8000B1EF000E01BA, 0x000E0DBF8000B1F3, 0x000E09BD000E01BE,
-    0x000FA1A3000E11BC, 0x0001E58100001582, 0x8000B1F70001E580,
-    0x000FDDA5000FCDA6, 0x000FC9AB000190BE, 0x0002A084000FC1AA,
-    0x000FCDAF8000B1FB, 0x000FC9AD000FC1AE, 0x8000B1FF000FD1AC,
-    0x0000E886000E8DB3, 0x000E89B600012C87, 0x8000B2078000B203,
-    0x000171BA000E89B5, 0x8000B20B000171BB, 0x0001F1BE8000B20F,
-    0x000161BC0001F1BF, 0x000E3DA8000161BD, 0x000E41AA000E45A9,
-    0x000E45AC000E41AB, 0x000191AE000E4DAD, 0x00018C84000191AF,
-    0x8000B2178000B213, 0x8000B21F8000B21B, 0x8000B2278000B223,
-    0x000169B88000B22B, 0x000175BA000175B9, 0x000119BC000109BB,
-    0x000109BE000119BD, 0x0001F1B0000101BF, 0x0001F1B20001F1B1,
-    0x000171B40001F1B3, 0x000159B6000171B5, 0x000DC9A3000159B7,
-    0x8000B2338000B22F, 0x8000B237000340BE, 0x000DCDA5000DCDA6,
-    0x000235AB8000B23B, 0x8000B23F000235AA, 0x0002B5AF8000B243,
-    0x000225AD0002B5AE, 0x8000B247000225AC, 0x8000B24F8000B24B,
-    0x8000B2578000B253, 0x8000B25F8000B25B, 0x000039808000B263,
-    0x0000058200003981, 0x0004D0868000B267, 0x8000B26F00034487,
-    0x0004CC848000B273, 0x8000B27B8000B277, 0x000264EF8000B27F,
-    0x8000B2878000B283, 0x000110E38000B28B, 0x0001F8E18000B28F,
-    0x8000B2978000B293, 0x8000B29F8000B29B, 0x000375BB8000B2A3,
-    0x8000B2A7000375BA, 0x0000D9BF8000B2AB, 0x000365BD0000D1BE,
-    0x0003B9B3000365BC, 0x8000B2B38000B2AF, 0x8000B2BB8000B2B7,
-    0x000395B5000351B6, 0x000241A88000B2BF, 0x000251AA000251A9,
-    0x00027DAC000265AB, 0x0002B1AE0002B1AD, 0x000568BE0002B1AF,
-    0x8000B2C78000B2C3, 0x00000D808000B2CB, 0x0000AD820000A581,
-    0x000169B88000B2CF, 0x000109BA000169B9, 0x000119BC000109BB,
-    0x000109BE000119BD, 0x0002D1B0000109BF, 0x0002D1B20002D1B1,
-    0x000175B40002D1B3, 0x000169B600017DB5, 0x000E70E1000161B7,
-    0x0000B0E3000F4CE1, 0x8000B2D3000F50E3, 0x00E748EF8000B2D7,
-    0x000228E100049885, 0x001370E38000B2DF, 0x000678BE8000B2E3,
-    0x00005CEF8000B2E7, 0x000299A3000E6CEF, 0x0007A8868000B2EB,
-    0x8000B2EF00048087, 0x0002B5A5000271A6, 0x000255AB8000B2F3,
-    0x8000B2F7000255AA, 0x0001F9AF8000B2FB, 0x000245AD0001F1AE,
-    0x000679A8000245AC, 0x00068DAA000679A9, 0x00069DAC000685AB,
-    0x000681AE000681AD, 0x8000B2DB000681AF, 0x8000B3038000B2FF,
-    0x8000B30B8000B307, 0x8000B3138000B30F, 0x0006A5B88000B317,
-    0x0006A5BA0006ADB9, 0x0006ADBC0006BDBB, 0x000771BE000771BD,
-    0x0006C5B0000771BF, 0x0006D9B20006C9B1, 0x0006C9B40006D9B3,
-    0x0006B1B60006C9B5, 0x000605B30006B1B7, 0x8000B31F8000B31B,
-    0x8000B3278000B323, 0x000615B5000601B6, 0x00060DBB8000B32B,
-    0x8000B32F000601BA, 0x00060DBF8000B333, 0x00060DBD00060DBE,
-    0x8000B33700060DBC, 0x8000B33B000641A3, 0x000645A68000B33F,
-    0x8000B3478000B343, 0x000645AA000651A5, 0x00012484000649AB,
-    0x000649AE8000B34F, 0x000649AC000649AF, 0x0008459F000649AD,
-    0x00090D9D00090D9E, 0x000D699B000B899C, 0x000F8999000D119A,
-    0x00F1259700F10D98, 0x00F5FD9500F12596, 0x00F7999300F50D94,
-    0x00F9319100F99992, 0x00FD298F00F93990, 0x00019D8100019D82,
-    0x8000B35300018180, 0x0001ECEF00048CEF, 0x000208E18000B357,
-    0x000130E30005F4E3, 0x001C0086000504E1, 0x8000B35B00018887,
-    0x0001A1A18000B35F, 0x001D09A38000B363, 0x001CF9A5000369A2,
-    0x00191DA7001D11A4, 0x001521A900191DA6, 0x001579AB0018B5A8,
-    0x001191AD001521AA, 0x002C01AF001001AC, 0x002D5DB1001089AE,
-    0x0029E9B3002DE5B0, 0x002801B5002801B2, 0x8000B3670029EDB4,
-    0x8000B36F8000B36B, 0x8000B3778000B373, 0x000319A38000B37B,
-    0x0005FDA10004D9A2, 0x00FE69820005FDA0, 0x8000B37F00FA0183,
-    0x00F4B9868000B383, 0x00FA118400F63587, 0x00F6F58A00FAB185,
-    0x8000B38700F2258B, 0x000CA18E8000B38B, 0x00F23D8C000E318F,
-    0x000E399200F22D8D, 0x000344840008C193, 0x000B1596001CC0BE,
-    0x000A5D9400065597, 0x0006659A000A5595, 0x8000B38F0007499B,
-    0x8000B3978000B393, 0x0000919C8000B39B, 0x000209B88000B39F,
-    0x000219BA000209B9, 0x000239BC000211BB, 0x000259BE000239BD,
-    0x000285B0000255BF, 0x000285B200028DB1, 0x000285B400029DB3,
-    0x000285B600028DB5, 0x000565A8000239B7, 0x00057DAA000569A9,
-    0x000501AC000509AB, 0x000531AE000509AD, 0x8000B3A3000531AF,
-    0x8000B26B8000B34B, 0x8000B3AB8000B3A7, 0x0001F5810001E582,
-    0x0004B5B30001F580, 0x8000B3AF001D8C84, 0x8000B3B78000B3B3,
-    0x0004C9B50004C9B6, 0x000309BB8000B3BF, 0x001D2886000301BA,
-    0x000309BF001D2087, 0x000309BD000301BE, 0x8000B3C3000311BC,
-    0x8000B3C70004F1A3, 0x00048DA6001CC0BE, 0x8000B3CF8000B3CB,
-    0x000345AA00048DA5, 0x001FA0BE00034DAB, 0x000345AE8000B3D3,
-    0x000355AC00034DAF, 0x8000B3D700034DAD, 0x8000B3DF8000B3DB,
-    0x8000B3E78000B3E3, 0x8000B3EF8000B3EB, 0x8000B3F30006D4EF,
-    0x8000B3F7000674E1, 0x8000B3FB000154E3, 0x8000B4038000B3FF,
-    0x00029DB38000B407, 0x8000B40F8000B40B, 0x8000B4178000B413,
-    0x0002A5B5000251B6, 0x000205BB8000B41B, 0x8000B41F000205BA,
-    0x0001FDBF8000B423, 0x000215BD0001FDBE, 0x001D5DA3000215BC,
-    0x001FE58100001582, 0x8000B427001FE580, 0x001D65A5001D91A6,
-    0x001DC5AB8000B42F, 0x00030886001DC5AA, 0x001E3DAF00037C87,
-    0x001DD5AD001E3DAE, 0x8000B433001DD5AC, 0x8000B437001F0DB3,
-    0x001F3DB68000B43B, 0x8000B4438000B43F, 0x001FE9BA001F05B5,
-    0x8000B447001FE9BB, 0x001FE9BE8000B44B, 0x001FE9BC001FD5BF,
-    0x8000B3BB001FE9BD, 0x8000B44F8000B42B, 0x8000B4578000B453,
-    0x8000B45F8000B45B, 0x001FBDA88000B463, 0x001FE1AA001FE5A9,
-    0x001FE9AC001FF9AB, 0x001F39AE001FD9AD, 0x001F49B0001F39AF,
-    0x001F59B2001F49B1, 0x001F49B4001F59B3, 0x001F71B6001F75B5,
-    0x001F55B8001F6DB7, 0x001F55BA001F5DB9, 0x001F39BC001F29BB,
-    0x001F29BE001F39BD, 0x001E49A3001F21BF, 0x8000B46B8000B467,
-    0x8000B4738000B46F, 0x001E41A5001E79A6, 0x001EADAB8000B477,
-    0x8000B47B001EADAA, 0x001E91AF8000B47F, 0x001EADAD001EADAE,
-    0x8000B483001EADAC, 0x8000B48B8000B487, 0x00001D808000B48F,
-    0x0000398200000981, 0x8000B4978000B493, 0x8000B49B0001A884,
-    0x000400868000B49F, 0x8000B4A30001D487, 0x00019DA88000B4A7,
-    0x0001C1AA0001C5A9, 0x0001C5AC0001C1AB, 0x0001F1AE0001CDAD,
-    0x8000B4AB0001F1AF, 0x8000B4B38000B4AF, 0x8000B4BB8000B4B7,
-    0x8000B4C38000B4BF, 0x0000CDB88000B4C7, 0x0000D5BA0000D5B9,
-    0x0000FDBC0000E5BB, 0x00009DBE000095BD, 0x000169B0000095BF,
-    0x000179B2000169B1, 0x000169B4000179B3, 0x0000FDB6000169B5,
-    0x000249B30000F5B7, 0x8000B4CF8000B4CB, 0x8000B4D3000480BE,
-    0x00024DB500024DB6, 0x0002ADBB8000B4D7, 0x8000B4DB0002ADBA,
-    0x0002ADBF8000B4DF, 0x0002ADBD0002ADBE, 0x8000B4E30002ADBC,
-    0x8000B4E700020DA3, 0x000209A68000B4EB, 0x8000B4F38000B4EF,
-    0x0002E9AA000209A5, 0x8000B4F70002E9AB, 0x0002E9AE8000B4FB,
-    0x0002E9AC0002E9AF, 0x8000B4FF0002E9AD, 0x000504BE8000B503,
-    0x8000B507000614BE, 0x8000B5138000B50F, 0x000021808000B517,
-    0x0000398200002181, 0x0004B08600032084, 0x8000B51B00035487,
-    0x8000B5238000B51F, 0x8000B52B8000B527, 0x8000B5338000B52F,
-    0x8000B537000220EF, 0x8000B53F8000B53B, 0x8000B543000258E3,
-    0x8000B547000184E1, 0x8000B54F8000B54B, 0x000130E18000B553,
-    0x00007CE3001E34E1, 0x8000B557001E18E3, 0x8000B55F8000B55B,
-    0x0005A8848000B563, 0x8000B56B8000B567, 0x8000B5738000B56F,
-    0x001EA8EF8000B577, 0x00023DB3001E88EF, 0x000468868000B57F,
-    0x000000D40005DC87, 0x0002FDB50002FDB6, 0x0002D9BB8000B583,
-    0x0006A4BE0002D1BA, 0x000161BF8000B587, 0x00017DBD00017DBE,
-    0x001DFDBC0002C1BC, 0x001D65BE001D65BD, 0x001DF9B8001D7DBF,
-    0x001DE9BA001DF9B9, 0x001DA9B4001DE9BB, 0x001D81B6001DA9B5,
-    0x001E51B0001D81B7, 0x001E51B2001E51B1, 0x001EBDAC001E51B3,
-    0x001E11AE001E41AD, 0x001F3DA8001E11AF, 0x001E85AA001E85A9,
-    0x0000CD80001EA5AB, 0x0000DD820000D581, 0x8000B58F8000B58B,
-    0x8000B5978000B593, 0x00057DA38000B59B, 0x8000B5A38000B59F,
-    0x8000B5AB8000B5A7, 0x0005BDA50005BDA6, 0x000599AB8000B5AF,
-    0x8000B5B3000591AA, 0x000621AF8000B5B7, 0x00063DAD00063DAE,
-    0x8000B5BB000581AC, 0x8000B5BF0006D9B3, 0x000691B68000B5C3,
-    0x8000B5CB8000B5C7, 0x0006ADBA0006E1B5, 0x8000B5CF0006B5BB,
-    0x000691BE8000B5D3, 0x0006A1BC00069DBF, 0x8000B50B0006A1BD,
-    0x8000B5D78000B57B, 0x00000D808000B5DB, 0x0001918200019181,
-    0x000635A88000B5DF, 0x000635AA00063DA9, 0x00060DAC000615AB,
-    0x00067DAE000675AD, 0x00060DB0000675AF, 0x000611B2000615B1,
-    0x000631B4000611B3, 0x000629B600063DB5, 0x000619B8000629B7,
-    0x0007F9BA000619B9, 0x0007E9BC0007F9BB, 0x0007DDBE0007E9BD,
-    0x00079DA30007D5BF, 0x8000B5E78000B5E3, 0x8000B5EB00028484,
-    0x0007A5A50007D5A6, 0x0007F1AB8000B5EF, 0x000C00860007E9AA,
-    0x0007D9AF00017C87, 0x0007E5AD0007D5AE, 0x8000B5F30007E5AC,
-    0x8000B5F7000629B3, 0x0006A5B68000B5FB, 0x8000B6038000B5FF,
-    0x000165BA0006ADB5, 0x8000B60700016DBB, 0x00016DBE8000B60B,
-    0x000175BC00015DBF, 0x0006C5A800016DBD, 0x0006E5AA0006D9A9,
-    0x000661AC000661AB, 0x000661AE000661AD, 0x8000B60F000661AF,
-    0x8000B6178000B613, 0x8000B61F8000B61B, 0x8000B6278000B623,
-    0x000189B88000B62B, 0x00019DBA000189B9, 0x0001B9BC000195BB,
-    0x000151BE0001B9BD, 0x0001E5B0000151BF, 0x0001E5B20001EDB1,
-    0x0001EDB40001FDB3, 0x0001D5B60001DDB5, 0x0005F5A60001BDB7,
-    0x8000B6338000B62F, 0x8000B6370005FDA5, 0x8000B63B000579A3,
-    0x00023DAE8000B63F, 0x000225AC00020DAF, 0x000235AA00023DAD,
-    0x8000B64300023DAB, 0x8000B64B8000B647, 0x8000B6538000B64F,
-    0x000015808000B657, 0x0000158200001D81, 0x000260848000B65F,
-    0x000348878000B663, 0x8000B667000CCC86, 0x8000B66F8000B66B,
-    0x8000B6778000B673, 0x000D28868000B67B, 0x000000A1000C4087,
-    0x8000B67F0001A4E1, 0x8000B6830002ECE3, 0x8000B68B8000B687,
-    0x8000B68F00073CEF, 0x8000B6978000B693, 0x8000B69F8000B69B,
-    0x8000B6A78000B6A3, 0x8000B6AB000385B3, 0x000381B6000381B5,
-    0x8000B6B38000B6AF, 0x000349BA8000B6B7, 0x000359BC000349BB,
-    0x000349BE000359BD, 0x000650E1000349BF, 0x000014E30007ACE1,
-    0x8000B6BB00063CE3, 0x8000B6C38000B6BF, 0x00003580000D5484,
-    0x0000058200003D81, 0x000D7CBE8000B6C7, 0x00074CEF8000B6CF,
-    0x000209A300074CEF, 0x8000B6D78000B6D3, 0x8000B6DF8000B6DB,
-    0x00020DA500020DA6, 0x0002C5AB8000B6E3, 0x8000B6E70002C5AA,
-    0x0002C5AF8000B6EB, 0x0002D5AD0002C5AE, 0x0005D5BC0002D5AC,
-    0x0005E1BE0005DDBD, 0x0005BDB80005E1BF, 0x0005D1BA0005D5B9,
-    0x0005A9B40005D1BB, 0x0005B5B60005B5B5, 0x000605B0000585B7,
-    0x000605B200060DB1, 0x000665AC0005B9B3, 0x000665AE00066DAD,
-    0x0006BDA800067DAF, 0x0006A5AA0006ADA9, 0x8000B6EF00067DAB,
-    0x8000B6F78000B6F3, 0x8000B6FF8000B6FB, 0x8000B7078000B703,
-    0x000E5DA88000B70B, 0x000E79AA000E61A9, 0x000E69AC000E79AB,
-    0x000E99AE000E69AD, 0x8000B70F000E99AF, 0x8000B7178000B713,
-    0x8000B71F8000B71B, 0x8000B7278000B723, 0x000E81B88000B72B,
-    0x000E81BA000E81B9, 0x000E8DBC000E9DBB, 0x000EBDBE000EB5BD,
-    0x000EF9B0000F4DBF, 0x000ECDB2000EC5B1, 0x000EC1B4000EC5B3,
-    0x000EC1B6000EC1B5, 0x8000B65B000EC1B7, 0x8000B72F000EF5B3,
-    0x000EEDB68000B6CB, 0x8000B7378000B733, 0x000E39BA000EE5B5,
-    0x0003E484000E05BB, 0x000E01BE0003E4BE, 0x000E1DBC000E01BF,
-    0x000EB1A3000E15BD, 0x00005D8100005D82, 0x8000B73B00004580,
-    0x000EA1A5000EA9A6, 0x000E41AB8000B73F, 0x00004886000E7DAA,
-    0x000E45AF00006887, 0x000E51AD000E45AE, 0x000E35A8000E59AC,
-    0x000E41AA000E41A9, 0x000E41AC000E41AB, 0x000E71AE000E49AD,
-    0x8000B743000E71AF, 0x00016C848000B747, 0x8000B74F8000B74B,
-    0x8000B7578000B753, 0x000161B88000B75B, 0x000161BA000161B9,
-    0x000161BC000161BB, 0x000161BE000161BD, 0x000E81B0000161BF,
-    0x000E81B2000E81B1, 0x000E81B4000E81B3, 0x000E81B6000E81B5,
-    0x000EC9B3000E81B7, 0x8000B7638000B75F, 0x8000B76B8000B767,
-    0x000EC9B5000EC9B6, 0x000E45BB8000B76F, 0x8000B773000E45BA,
-    0x000E45BF8000B777, 0x000E55BD000E45BE, 0x8000B77B000E55BC,
-    0x8000B77F000E8DA3, 0x000E8DA68000B783, 0x8000B78B8000B787,
-    0x000E01AA000E8DA5, 0x8000B78F000E01AB, 0x000E01AE8000B793,
-    0x000E11AC000E01AF, 0x0002DDA8000E11AD, 0x00032DAA000325A9,
-    0x000321AC000321AB, 0x000321AE000321AD, 0x8000B797000321AF,
-    0x8000B79F8000B79B, 0x8000B7A78000B7A3, 0x8000B7AF000278BE,
-    0x000389B88000B7B3, 0x00039DBA000389B9, 0x0003B9BC000395BB,
-    0x000079BE0003B9BD, 0x0003E5B0000079BF, 0x0003E5B20003EDB1,
-    0x0003EDB40003FDB3, 0x0003D5B60003DDB5, 0x0000AD810003BDB7,
-    0x0003ADB30000AD80, 0x0003A1B50000BD82, 0x8000B7BB8000B7B7,
-    0x000400860003A9B6, 0x000321BB0003D487, 0x000321BD000329BA,
-    0x000311BF000339BC, 0x0003E5A3000321BE, 0x8000B7C38000B7BF,
-    0x8000B7C70030F085, 0x0003E9A50003E1A6, 0x000369AB8000B7CB,
-    0x8000B7CF000361AA, 0x000359AF8000B7D3, 0x000369AD000369AE,
-    0x0001B8E3000371AC, 0x000F98E1000F70E1, 0x8000B7D7000E00E3,
-    0x8000B7DF8000B7DB, 0x8000B7E78000B7E3, 0x8000B7EF8000B7EB,
-    0x000164EF8000B7F3, 0x005D30BF8000B7F7, 0x8000B7FB000EF4EF,
-    0x000504848000B7FF, 0x8000B803000504BE, 0x8000B8070001A4E1,
-    0x8000B80B000064E3, 0x8000B8138000B80F, 0x000015800003A0EF,
-    0x0000058200001D81, 0x000EC9AA8000B817, 0x000E95A8000EC9AB,
-    0x000EF9AE000ECDA9, 0x000ECDAC000EF9AF, 0x00040086000EC5AD,
-    0x8000B81F00048487, 0x8000B8278000B823, 0x8000B82F8000B82B,
-    0x000DD1BA8000B833, 0x000DCDB8000DDDBB, 0x000DEDBE000DC5B9,
-    0x000DC5BC000D95BF, 0x000E49B2000DE5BD, 0x000E49B0000E49B3,
-    0x000E49B6000E49B1, 0x000E49B4000E49B7, 0x000639A8000E49B5,
-    0x00069DAA000639A9, 0x0006D5AC000695AB, 0x0006C9AE0006DDAD,
-    0x8000B81B0006C1AF, 0x8000B83B8000B837, 0x8000B8438000B83F,
-    0x8000B84B8000B847, 0x000769B88000B84F, 0x00077DBA000769B9,
-    0x000715BC00070DBB, 0x000701BE000705BD, 0x000689B000073DBF,
-    0x000779B2000689B1, 0x000769B4000779B3, 0x000759B6000769B5,
-    0x000615B3000759B7, 0x8000B8578000B853, 0x8000B85F8000B85B,
-    0x00065DB500065DB6, 0x00064DBB8000B863, 0x8000B86700064DBA,
-    0x0007ADBF8000B86B, 0x0007ADBD0007ADBE, 0x8000B86F0007ADBC,
-    0x8000B873000651A3, 0x000619A68000B877, 0x8000B87F8000B87B,
-    0x000609AA000619A5, 0x8000B883000609AB, 0x0007E9AE8000B887,
-    0x0007E9AC0007E9AF, 0x0006A9A80007E9AD, 0x0006B9AA0006A9A9,
-    0x0006A9AC0006B9AB, 0x00013DAE0006A9AD, 0x0001F980000135AF,
-    0x0001C5820001CD81, 0x00027CBE00027884, 0x000178878000B88B,
-    0x0001D1B800009C86, 0x0001E1BA0001D9B9, 0x000191BC0001E1BB,
-    0x000195BE00019DBD, 0x00014DB0000189BF, 0x00015DB2000155B1,
-    0x00014DB4000155B3, 0x0001F1B60001F1B5, 0x000639B50001F1B7,
-    0x8000B8938000B88F, 0x8000B897000631B6, 0x000629B38000B89B,
-    0x000199BD8000B89F, 0x00012DBF000191BC, 0x8000B8A3000191BE,
-    0x000189BB8000B7AB, 0x000671A3000191BA, 0x8000B8AB8000B8A7,
-    0x8000B8B38000B8AF, 0x000661A5000669A6, 0x0001D1AB8000B8B7,
-    0x8000B8BB0001C9AA, 0x000175AF8000B8BF, 0x0001C1AD0001C9AE,
-    0x8000B8C30001C9AC, 0x0007BCEF8000B8C7, 0x8000B8CF8000B8CB,
-    0x8000B8D78000B8D3, 0x8000B8DF8000B8DB, 0x8000B8E78000B8E3,
-    0x000688E18000B8EB, 0x000760E38000B8EF, 0x000075808000B8F3,
-    0x0000198200007981, 0x8000B8F700024084, 0x000314BE8000B8FF,
-    0x003CC0868000B903, 0x8000B90700034487, 0x8000B90F8000B90B,
-    0x8000B9178000B913, 0x8000B91F8000B91B, 0x000228EF8000B923,
-    0x8000B92B8000B927, 0x8000B9338000B92F, 0x003CEC848000B937,
-    0x8000B93F8000B93B, 0x000194E18000B943, 0x0002E4E38000B947,
-    0x000351B38000B94B, 0x8000B9538000B94F, 0x8000B95B8000B957,
-    0x000345B5000365B6, 0x000331BB8000B95F, 0x8000B963000329BA,
-    0x0000F1BF8000B967, 0x000311BD000319BE, 0x000285A8000321BC,
-    0x000299AA00028DA9, 0x00028DAC000299AB, 0x0002DDAE0002A5AD,
-    0x0001FD800002D5AF, 0x0001C5820001CD81, 0x003C9086003D64BE,
-    0x8000B96F003DB887, 0x000151B8003C6084, 0x000151BA000151B9,
-    0x000171BC000151BB, 0x000171BE000171BD, 0x0002ADB0000171BF,
-    0x0002B5B20002B5B1, 0x00029DB4000285B3, 0x00017DB6000175B5,
-    0x8000B973000171B7, 0x8000B97B8000B977, 0x8000B9838000B97F,
-    0x8000B98B8000B987, 0x003CECBE000054EF, 0x8000B98F00061CE1,
-    0x8000B993000090E3, 0x8000B99B8000B997, 0x000251A38000B99F,
-    0x003F44848000B9A3, 0x8000B9AB8000B9A7, 0x000245A5000265A6,
-    0x000231AB8000B9AF, 0x8000B9B3000229AA, 0x0001F1AF8000B9B7,
-    0x000211AD000219AE, 0x003E55A8000221AC, 0x003E61AA003E59A9,
-    0x003E61AC003E61AB, 0x003E61AE003E61AD, 0x8000B96B003E61AF,
-    0x8000B9BF8000B9BB, 0x8000B9C78000B9C3, 0x8000B9CF8000B9CB,
-    0x003EE9B88000B9D3, 0x003EFDBA003EE9B9, 0x003E95BC003EF1BB,
-    0x003E95BE003E9DBD, 0x003EE1B0003E8DBF, 0x003EE5B2003EE1B1,
-    0x003EE5B4003EFDB3, 0x003EE5B6003EEDB5, 0x003EE9B3003ED9B7,
-    0x0000B58100001582, 0x8000B9D70000B580, 0x003EE9B5003EE1B6,
-    0x003E25BB8000B9DB, 0x0000A886003E25BA, 0x003E29BF0000A487,
-    0x003E29BD003E21BE, 0x8000B9DF003E31BC, 0x8000B9E3003EADA3,
-    0x003EA5A68000B9E7, 0x8000B9EF8000B9EB, 0x003E61AA003EADA5,
-    0x8000B9F3003E61AB, 0x003E65AE8000B9F7, 0x003E75AC003E6DAF,
-    0x003E39A8003E6DAD, 0x003E8DAA003E85A9, 0x003E9DAC003E85AB,
-    0x003E8DAE003E85AD, 0x8000B9FB003E85AF, 0x8000BA038000B9FF,
-    0x8000BA0B8000BA07, 0x8000BA138000BA0F, 0x003E9DB88000BA17,
-    0x003EA5BA003EADB9, 0x000179BC000169BB, 0x000169BE000179BD,
-    0x003EC9B0000169BF, 0x003EDDB2003ED5B1, 0x003ECDB4003ED5B3,
-    0x003EB5B6003EB1B5, 0x8000BA1B003EADB7, 0x003E05B5003E15B6,
-    0x003E15B38000BA1F, 0x8000BA278000BA23, 0x003E65BF8000BA2B,
-    0x003E11BD003E19BE, 0x003E01BB003E19BC, 0x8000B8FB003E1DBA,
-    0x003E45A38000BA2F, 0x8000BA378000BA33, 0x8000BA3F8000BA3B,
-    0x003E55A5003E45A6, 0x003E51AB8000BA43, 0x8000BA47003E4DAA,
-    0x003E35AF00002CBE, 0x003E41AD003E49AE, 0x0003E982003E49AC,
-    0x0003D980003FD0EF, 0x000350870003D181, 0x8000BA4F0004B486,
-    0x8000BA578000BA53, 0x8000BA5F8000BA5B, 0x8000BA678000BA63,
-    0x8000BA6B003E84E3, 0x00028DB3003E44E1, 0x000344848000BA6F,
-    0x8000BA73000440BE, 0x000281B50002BDB6, 0x000349BB8000BA77,
-    0x8000BA7B000341BA, 0x000349BF8000BA7F, 0x000349BD000341BE,
-    0x8000BA83000351BC, 0x8000BA8B8000BA87, 0x8000BA8F000580BE,
-    0x8000BA978000BA93, 0x000261A88000BA9B, 0x000265AA00026DA9,
-    0x000269AC000279AB, 0x000255AE00025DAD, 0x0003E5B000039DAF,
-    0x0003E5B20003EDB1, 0x0003E5B40003FDB3, 0x0003E5B60003EDB5,
-    0x00036DB800035DB7, 0x000371BA000371B9, 0x00034DBC000371BB,
-    0x000331BE000331BD, 0x8000BA9F000331BF, 0x8000BAA78000BAA3,
-    0x003EC8E38000BAAB, 0x0001A8E100051C84, 0x000034EF8000BAAF,
-    0x8000BAB78000BAB3, 0x8000BAC38000BABF, 0x8000BACB8000BAC7,
-    0x00005D808000BACF, 0x0000618200006181, 0x8000BAD3000381A3,
-    0x0003B1A600038DA5, 0x0004C08600079884, 0x00024DAA00054887,
-    0x00025DAC000245AB, 0x00024DAE000245AD, 0x8000BAD7000245AF,
-    0x8000BADF8000BADB, 0x8000BAE78000BAE3, 0x8000BAEF8000BAEB,
-    0x000620BE003EC4EF, 0x8000BAF3003E8CE1, 0x8000BAF7003E80E3,
-    0x8000BAFF8000BAFB, 0x0001E1B38000BB03, 0x8000BB0B8000BB07,
-    0x8000BB138000BB0F, 0x0001F1B50001F9B6, 0x0001B1BB8000BB17,
-    0x8000BB1B0001A1BA, 0x000171BF8000BB1F, 0x000191BD000199BE,
-    0x000619A80001A1BC, 0x000629AA000619A9, 0x000639AC000629AB,
-    0x000629AE000639AD, 0x8000BABB000625AF, 0x0000198100001182,
-    0x8000BB230007C180, 0x8000BB2B8000BB27, 0x0006D1B88000BB2F,
-    0x0006D1BA0006D1B9, 0x000751BC0006D1BB, 0x000751BE000751BD,
-    0x000625B0000751BF, 0x000635B2000635B1, 0x00061DB4000605B3,
-    0x0006FDB60006F5B5, 0x0006A1A30006F5B7, 0x0000AC868000BB33,
-    0x8000BB3700034C87, 0x0006B1A50006B9A6, 0x0006F1AB8000BB3B,
-    0x8000BB3F0006E1AA, 0x000631AF8000BB43, 0x0006D1AD0006D9AE,
-    0x8000BB470006E1AC, 0x8000BB4B0007F1B3, 0x000781B68000BB4F,
-    0x8000BB578000BB53, 0x000769BA0007E1B5, 0x8000BB5B000769BB,
-    0x000769BE8000BB5F, 0x000769BC00075DBF, 0x00062DA8000769BD,
-    0x00063DAA000635A9, 0x00062DAC000635AB, 0x000699AE000645AD,
-    0x8000BB63000699AF, 0x8000BB6B8000BB67, 0x8000BB738000BB6F,
-    0x8000BB7B8000BB77, 0x0006F9B88000BB7F, 0x00068DBA0006F9B9,
-    0x00069DBC000685BB, 0x000685BE000685BD, 0x0006E9B00006B5BF,
-    0x0006F9B20006E9B1, 0x0006E9B40006F9B3, 0x0006C9B60006DDB5,
-    0x0006B5A30006C9B7, 0x8000BB878000BB83, 0x8000BB8F8000BB8B,
-    0x0006A5A50006C5A6, 0x00062DAB8000BB93, 0x8000BB9700062DAA,
-    0x000619AF8000BB9F, 0x00062DAD00062DAE, 0x0000518000062DAC,
-    0x0000618200005981, 0x8000BB9B8000BA4B, 0x000178878000BBA3,
-    0x8000BBA7000C0086, 0x8000BBAF8000BBAB, 0x8000BBB78000BBB3,
-    0x8000BBBF8000BBBB, 0x8000BBC78000BBC3, 0x000299B38000BBCB,
-    0x0002B9B500039484, 0x8000BBD38000BBCF, 0x8000BBD70002B1B6,
-    0x0002FDBB8000BBDB, 0x0002D9BD0002E1BA, 0x000319BF0002EDBC,
-    0x000618E30002D1BE, 0x000644E18000BBDF, 0x8000BBE78000BBE3,
-    0x8000BBEB000190E1, 0x8000BBEF000240E3, 0x8000BBF78000BBF3,
-    0x0006C4EF000224EF, 0x8000BBFF8000BBFB, 0x000205A98000BC03,
-    0x000215AB0002B9A8, 0x00023DAD00021DAA, 0x000251AF00023DAC,
-    0x0000D98000025DAE, 0x000005820000D981, 0x0003FCBE8000BC07,
-    0x000CA8878000BC0F, 0x0003F1B9000D9C86, 0x0003F1BB0003E9B8,
-    0x000355BD0003F9BA, 0x000355BF0003E9BC, 0x000235B1000355BE,
-    0x000201B3000229B0, 0x0003D1B5000235B2, 0x0003D1B7000219B4,
-    0x000D44850003D9B6, 0x8000BC178000BC13, 0x8000BC1B00039DA3,
-    0x0003B5A60003BDA5, 0x8000BC238000BC1F, 0x0003E5AA8000BC27,
-    0x0003E9AC0003F9AB, 0x0003D5AE0003DDAD, 0x0002ADB600021DAF,
-    0x8000BC2F8000BC2B, 0x8000BC330002A5B5, 0x8000BC370002B5B3,
-    0x000175BE8000BC3B, 0x000169BC00017DBF, 0x000289BA000169BD,
-    0x8000BC3F000289BB, 0x000294E18000BC43, 0x00123CE30006DCEF,
-    0x8000BC4B8000BC47, 0x8000BC5A8000BC56, 0x000F6C840007E8EF,
-    0x8000BC5E0006DCE1, 0x8000BC620006A8E3, 0x003E7CEF0006F8E3,
-    0x000221A300012CE1, 0x8000BC66000000F1, 0x8000BC6E8000BC6A,
-    0x000231A5000239A6, 0x00021DAB8000BC72, 0x8000BC7600021DAA,
-    0x0001E9AF8000BC7A, 0x0001FDAD0001E1AE, 0x0024218F0001FDAC,
-    0x8000BC7E8000BC0B, 0x000FF9800000C4BE, 0x00001182000FF981,
-    0x000000868000BC82, 0x8000BC8600038487, 0x8000BC8E8000BC8A,
-    0x8000BC968000BC92, 0x0008DD9F8000BC9A, 0x0009F19D0008659E,
-    0x0034719B0035799C, 0x003089990035E19A, 0x0030E1970030D198,
-    0x003CED95003D7596, 0x00391993003CED94, 0x0038F9910038F192,
-    0x0014A5B100269990, 0x006809B30014A5B0, 0x006805B50015E9B2,
-    0x8000BC9E006819B4, 0x8000BCA68000BCA2, 0x8000BCAE8000BCAA,
-    0x000D61A38000BCB2, 0x000C99A1000C8DA2, 0x0000C5A1000D71A0,
-    0x0004C9A38000BCB6, 0x000439A50000D1A2, 0x0018F9A70004D1A4,
-    0x001C01A9001801A6, 0x001C6DAB001879A8, 0x0010D1AD001CCDAA,
-    0x001401AF001001AC, 0x002521820010D9AE, 0x8000BCBA0027E183,
-    0x003879868000BCBE, 0x00399984003D9587, 0x003DA58A00399985,
-    0x8000BCC2003C358B, 0x0031E18E8000BCC6, 0x0031B98C0034818F,
-    0x003561920031B18D, 0x8000BCCA00347193, 0x0009CD968000BCCE,
-    0x0008E194000C9997, 0x000D419A00097D95, 0x8000BCD2000D5D9B,
-    0x8000BCDA8000BCD6, 0x000CA99C8000BCDE, 0x000F35A88000BCE2,
-    0x000E81AA000E81A9, 0x000E8DAC000E9DAB, 0x000EBDAE000EB5AD,
-    0x00000D80000ED9AF, 0x00001D8200001581, 0x8000BCEA8000BCE6,
-    0x0000BC848000BCEE, 0x000E55B8000400BE, 0x000E61BA000E5DB9,
-    0x000E19BC000E7DBB, 0x000E09BE000E19BD, 0x000EA9B0000E01BF,
-    0x000EB1B2000EB5B1, 0x000E89B4000EB1B3, 0x000E75B6000E89B5,
-    0x000D09B3000E6DB7, 0x000488868000BCF6, 0x00045C840003C887,
-    0x000D05B5000D35B6, 0x000D71BB8000BCFA, 0x8000BCFE000D6DBA,
-    0x000D4DBF8000BD02, 0x000D65BD000D4DBE, 0x8000BD06000D65BC,
-    0x8000BD0A000D4DA3, 0x000D71A68000BD0E, 0x8000BD168000BD12,
-    0x000D29AA000D41A5, 0x8000BD1A000D35AB, 0x000D09AE8000BD1E,
-    0x000D21AC000D09AF, 0x0002A9A8000D21AD, 0x0002B9AA0002A9A9,
-    0x0002A9AC0002B9AB, 0x000201AE0002A9AD, 0x8000BD22000201AF,
-    0x8000BD2A8000BD26, 0x8000BD328000BD2E, 0x8000BD3A8000BD36,
-    0x0003EDB88000BD3E, 0x000385BA00038DB9, 0x0003BDBC0003A5BB,
-    0x0003ADBE0003A5BD, 0x000241B00003A1BF, 0x000241B2000241B1,
-    0x000235B4000241B3, 0x0003F5B60003FDB5, 0x000D25B70003D5B7,
-    0x000D11B5000D25B6, 0x000D15B3000D19B4, 0x000EE5B1000D15B2,
-    0x000D55BF000EE5B0, 0x000D4DBD000D4DBE, 0x000D69BB000D4DBC,
-    0x000D75B9000D65BA, 0x8000BD42000D79B8, 0x8000BD4600050CBE,
-    0x8000BD4E8000BD4A, 0x8000BD568000BD52, 0x000E95AF8000BD5A,
-    0x000E85AD000E95AE, 0x000E95AB000E85AC, 0x000EADA9000E95AA,
-    0x000F24E1000ED1A8, 0x000E38E38000BD5E, 0x000038E38000BD66,
-    0x0001A0E18000BD6A, 0x0000E4EF8000BD6E, 0x0004E4848000BD72,
-    0x000410868000BD76, 0x00014CEF00049487, 0x00009D818000BD7A,
-    0x0003F9B300009D80, 0x0003E9B500004982, 0x8000BD828000BD7E,
-    0x8000BD860003E1B6, 0x00034DBB8000BD8A, 0x00035DBD00034DBA,
-    0x0002F1BF00035DBC, 0x8000BCF2000339BE, 0x8000BD8E8000BD62,
-    0x8000BD968000BD92, 0x8000BD9E8000BD9A, 0x000625A88000BDA2,
-    0x000601AA000639A9, 0x000671AC000601AB, 0x000675AE00067DAD,
-    0x000619B0000669AF, 0x000625B200062DB1, 0x0006F5B40006EDB3,
-    0x0006F5B60006FDB5, 0x0006D5B80006EDB7, 0x0006D5BA0006DDB9,
-    0x000739BC00072DBB, 0x000729BE000739BD, 0x0006A9B3000729BF,
-    0x8000BDAA8000BDA6, 0x8000BDB28000BDAE, 0x0006D9B50006DDB6,
-    0x0006C9BB8000BDB6, 0x8000BDBA0006FDBA, 0x000619BF8000BDBE,
-    0x000619BD000611BE, 0x8000BDC20006D1BC, 0x8000BDC60006EDA3,
-    0x000699A68000BDCA, 0x000390BE8000BDCE, 0x0006B9AA00069DA5,
-    0x8000BDD200068DAB, 0x000655AE8000BDD6, 0x000695AC00065DAF,
-    0x00068DA800065DAD, 0x000695AA000691A9, 0x0006BDAC0006A9AB,
-    0x0006CDAE0006A9AD, 0x0001E9800006C5AF, 0x0001FD820001E981,
-    0x000190868000BDDA, 0x8000BDDE0001B487, 0x0001C5B88000BDE2,
-    0x0001C5BA0001CDB9, 0x0001CDBC0001DDBB, 0x0001F5BE0001FDBD,
-    0x000681B000019DBF, 0x000161B2000681B1, 0x000161B4000161B3,
-    0x000161B6000161B5, 0x00061DB3000161B7, 0x8000BDEA8000BDE6,
-    0x8000BDF28000BDEE, 0x000635B5000629B6, 0x000645BB8000BDF6,
-    0x8000BDFA000645BA, 0x0001A5BF8000BDFE, 0x0001B5BD0001A5BE,
-    0x8000BE020001B5BC, 0x8000BE06000659A3, 0x00066DA68000BE0A,
-    0x8000BE128000BE0E, 0x000601AA000671A5, 0x8000BE16000601AB,
-    0x0001E1AE8000BE1A, 0x0001F1AC0001E1AF, 0x0002EDA80001F1AD,
-    0x000331AA000329A9, 0x000395AC000331AB, 0x000395AE00039DAD,
-    0x00020C8400038DAF, 0x8000BE228000BE1E, 0x8000BE2A8000BE26,
-    0x8000BE328000BE2E, 0x0003A9B88000BE36, 0x00006DBA0003A9B9,
-    0x00007DBC000065BB, 0x00006DBE000065BD, 0x0003FDB0000065BF,
-    0x0003C5B20003CDB1, 0x0003B9B40003ADB3, 0x0003ADB60003B9B5,
-    0x0000BD800003A5B7, 0x0000198200000981, 0x8000BE3E8000BE3A,
-    0x8000BE460003D8BE, 0x001C00868000BE4A, 0x8000BE4E0003D487,
-    0x8000BE568000BE52, 0x8000BE5E8000BE5A, 0x000369BB8000BE62,
-    0x8000BE66000359BA, 0x00031DBF8000BE6A, 0x000319BD00031DBE,
-    0x0003D9B3000371BC, 0x8000BE728000BE6E, 0x8000BE7A8000BE76,
-    0x00037DB500037DB6, 0x000609AF8000BE7E, 0x0007A5AD0007C1AE,
-    0x0007B5AB0007A5AC, 0x8000BE820007B5AA, 0x001D6C848000BE86,
-    0x000711A5000719A6, 0x000701A38000BE8A, 0x8000BE928000BE8E,
-    0x8000BE9A8000BE96, 0x001D08BE8000BE9E, 0x8000BEA2001C0084,
-    0x8000BEA6000338EF, 0x8000BEAE8000BEAA, 0x8000BEB6000354E3,
-    0x00002980000044E1, 0x0000198200002981, 0x000628E18000BEBA,
-    0x000630E3000794E1, 0x8000BEBE00078CE3, 0x001C38868000BEC2,
-    0x001CE8BE001D5C87, 0x8000BECA8000BEC6, 0x8000BED28000BECE,
-    0x000198EF8000BED6, 0x0002F9A3000600EF, 0x8000BEDE8000BEDA,
-    0x8000BEE68000BEE2, 0x00025DA500025DA6, 0x000249AB8000BEEA,
-    0x8000BEEE000279AA, 0x00023DAF8000BEF2, 0x000239AD00023DAE,
-    0x001E99A8000251AC, 0x001EA1AA001E99A9, 0x001EA1AC001EA1AB,
-    0x001EA1AE001EA1AD, 0x8000BEB2001EA1AF, 0x8000BEFA8000BEF6,
-    0x8000BF028000BEFE, 0x8000BF0A8000BF06, 0x001F5DB88000BF0E,
-    0x001F7DBA001F69B9, 0x001F6DBC001F75BB, 0x001F1DBE001F15BD,
-    0x001EA9B0001F0DBF, 0x001E8DB2001EA9B1, 0x001E9DB4001E85B3,
-    0x001E85B6001E85B5, 0x001E29B3001F65B7, 0x8000BF168000BF12,
-    0x8000BF1E8000BF1A, 0x001E29B5001E21B6, 0x001E6DBB8000BF22,
-    0x00000886001E6DBA, 0x001E41BF00002C87, 0x001E51BD001E59BE,
-    0x00004980001E59BC, 0x0000598200004981, 0x8000BF26001E65A3,
-    0x001E6DA6001E65A5, 0x8000BF2A8000BE42, 0x001E21AA8000BF2E,
-    0x001E15AC001E21AB, 0x001E15AE001E1DAD, 0x001E69A8001E0DAF,
-    0x001E99AA001E79A9, 0x001EB5AC001E99AB, 0x001EB5AE001EBDAD,
-    0x8000BF32001EADAF, 0x8000BF3A8000BF36, 0x8000BF428000BF3E,
-    0x8000BF4A8000BF46, 0x000165B88000BF4E, 0x000165BA00016DB9,
-    0x000165BC00017DBB, 0x000165BE00016DBD, 0x001ED5B00001D9BF,
-    0x001EA1B2001EA1B1, 0x001EA1B4001EA1B3, 0x001E91B6001EA9B5,
-    0x001E11B3001E91B7, 0x8000BF568000BF52, 0x8000BF5E8000BF5A,
-    0x001E3DB5001E3DB6, 0x001E65BB8000BF62, 0x8000BF66001E11BA,
-    0x001E6DBF8000BF6A, 0x001E75BD001E65BE, 0x8000BF6E001E75BC,
-    0x8000BF72001E55A3, 0x001E79A68000BF76, 0x8000BF7E8000BF7A,
-    0x001E55AA001E79A5, 0x8000BF82001E21AB, 0x001E21AE8000BF86,
-    0x001E31AC001E29AF, 0x0002D1A8001E31AD, 0x0002EDAA0002E5A9,
-    0x000335AC00032DAB, 0x000335AE00033DAD, 0x8000BF8A00032DAF,
-    0x8000BF968000BF8E, 0x8000BF9A000260BE, 0x8000BFA28000BF9E,
-    0x0003E9B88000BFA6, 0x000389BA0003E9B9, 0x00039DBC000385BB,
-    0x000381BE000381BD, 0x000355B00003B5BF, 0x000355B200035DB1,
-    0x0003F9B40003E9B3, 0x0003E9B60003F9B5, 0x00036D800003E1B7,
-    0x0000AD820000A581, 0x8000BFAA000269B3, 0x0003B1B60003B1B5,
-    0x0002E0848000BFAE, 0x00039DBA8000BFB2, 0x00038DBC000395BB,
-    0x000331BE000331BD, 0x000225A3000331BF, 0x8000BFBA8000BFB6,
-    0x8000BFC28000BFBE, 0x0003FDA50003FDA6, 0x0003D9AB8000BFC6,
-    0x000448860003D1AA, 0x00037DAF0003AC87, 0x00037DAD00037DAE,
-    0x8000BFCA0003C1AC, 0x8000BFD28000BFCE, 0x0003A0E38000BFD6,
-    0x0001A8E18000BFDA, 0x0003F4EF8000BFDE, 0x8000BFE68000BFE2,
-    0x8000BFEE8000BFEA, 0x8000BFF68000BFF2, 0x000079808000BFFA,
-    0x00001D8200007981, 0x001E60E300024084, 0x001FE0E18000BFFE,
-    0x001ED0EF000550BE, 0x0004C8868000C006, 0x8000C00A00054087,
-    0x8000C0128000C00E, 0x8000C01A8000C016, 0x8000C0228000C01E,
-    0x8000C02A8000C026, 0x8000C0328000C02E, 0x0005A484001EC4EF,
-    0x8000C036001E24E1, 0x8000C03A001EF0E3, 0x8000C0428000C03E,
-    0x0001F5B38000C046, 0x8000C04E8000C04A, 0x8000C0568000C052,
-    0x0001E5B5000145B6, 0x000129BB8000C05A, 0x8000C05E000121BA,
-    0x0000F1BF8000C062, 0x000129BD000129BE, 0x000629A8000131BC,
-    0x00063DAA000635A9, 0x00062DAC000635AB, 0x000685AE000685AD,
-    0x8000C0020006B9AF, 0x00001D8100001D82, 0x8000C06600001D80,
-    0x8000C06E8000C06A, 0x00074DB8000330BE, 0x000755BA000751B9,
-    0x000779BC000769BB, 0x000769BE000779BD, 0x0006A5B0000769BF,
-    0x0006A5B20006ADB1, 0x0006ADB40006BDB3, 0x00077DB6000775B5,
-    0x0006B5A3000775B7, 0x000028868000C072, 0x8000C07600034C87,
-    0x0006A5A5000605A6, 0x000669AB8000C07A, 0x8000C07E000661AA,
-    0x0007B1AF8000C082, 0x000669AD000669AE, 0x8000C086000671AC,
-    0x0007BDB38000C08A, 0x0007ADB58000C08E, 0x8000C0968000C092,
-    0x8000BF9200075DB6, 0x000779BB8000C09A, 0x000769BD000761BA,
-    0x000749BF000761BC, 0x000665A8000755BE, 0x00067DAA000675A9,
-    0x00066DAC000675AB, 0x00069DAE000695AD, 0x8000C09E000695AF,
-    0x8000C0A68000C0A2, 0x8000C0AE8000C0AA, 0x8000C0B68000C0B2,
-    0x0006FDB88000C0BA, 0x000681BA000681B9, 0x000681BC000681BB,
-    0x0006B1BE000689BD, 0x0006EDB00006B1BF, 0x0006FDB20006F5B1,
-    0x0006D9B40006F5B3, 0x0006CDB60006C5B5, 0x0006F5A30006C5B7,
-    0x8000C0C28000C0BE, 0x8000C0CA8000C0C6, 0x0006E5A5000615A6,
-    0x000631AB8000C0CE, 0x8000C0D2000629AA, 0x000601AF8000C0D6,
-    0x000621AD00061DAE, 0x00006D80000629AC, 0x0000198200000981,
-    0x8000C0DE8000C0DA, 0x000190BE00019884, 0x000F40868000C0E6,
-    0x8000C0EA0000F487, 0x8000C0F28000C0EE, 0x8000C0FA8000C0F6,
-    0x000291A88000C0FE, 0x0002A1AA000299A9, 0x0002D1AC0002A1AB,
-    0x0002D5AE0002D1AD, 0x8000C1020002C9AF, 0x8000C10A8000C106,
-    0x8000C10E00032084, 0x8000C1168000C112, 0x000379B88000C11A,
-    0x0003CDBA000379B9, 0x0003DDBC0003C5BB, 0x0003C5BE0003C5BD,
-    0x0002B9B00003F5BF, 0x000349B2000281B1, 0x000359B4000349B3,
-    0x000349B6000359B5, 0x000229B3000349B7, 0x8000C1228000C11E,
-    0x000C9CBE8000C126, 0x000259B5000251B6, 0x000245BB8000C12A,
-    0x8000C12E00027DBA, 0x00039DBF8000C132, 0x00039DBD00039DBE,
-    0x8000C13600039DBC, 0x8000C13A00026DA3, 0x000215A68000C13E,
-    0x8000C1468000C142, 0x000239AA00021DA5, 0x8000C14A000201AB,
-    0x0003D9AE8000C14E, 0x0003D9AC0003D9AF, 0x000190E10003D9AD,
-    0x0000CCE38000C156, 0x0001C980000D1884, 0x0001F5820001FD81,
-    0x000C2CBE8000C15A, 0x8000C1628000C15E, 0x8000C16A8000C166,
-    0x000018EF8000C16E, 0x000CE0868000C172, 0x8000C1760002F487,
-    0x8000C17E8000C17A, 0x8000C1868000C182, 0x8000C18A0006F8EF,
-    0x8000C18E000794E1, 0x8000C192000630E3, 0x8000C19A8000C196,
-    0x8000C1A28000C19E, 0x8000C1AA8000C1A6, 0x8000C1B28000C1AE,
-    0x8000C1BA8000C1B6, 0x00700085000158EF, 0x8000C1BE0006D4E1,
-    0x0000258000061CE3, 0x0000118200002D81, 0x00022DB38000C1C6,
-    0x000CE887000000F1, 0x8000C1CA000F6C86, 0x00022DB500022DB6,
-    0x0001C9BB8000C1CE, 0x8000C1D20001C1BA, 0x0001CDBF8000C1D6,
-    0x0001CDBD0001CDBE, 0x000DDDA80001CDBC, 0x000DE5AA000DEDA9,
-    0x000DE5AC000DFDAB, 0x000DE5AE000DEDAD, 0x8000C152000E3DAF,
-    0x8000C1DA8000C1C2, 0x8000C1E28000C1DE, 0x8000C1EA8000C1E6,
-    0x000ECDB88000C1EE, 0x000EE1BA000ED9B9, 0x000E91BC000EE1BB,
-    0x000E95BE000E9DBD, 0x000E45B0000E89BF, 0x000E59B2000E4DB1,
-    0x000E49B4000E59B3, 0x000EFDB6000E49B5, 0x000D6DA3000EF5B7,
-    0x8000C1F68000C1F2, 0x8000C1FE8000C1FA, 0x000D6DA5000D6DA6,
-    0x000E89AB8000C202, 0x8000C206000E81AA, 0x000E8DAF8000C20A,
-    0x000E8DAD000E8DAE, 0x000FEDB5000E8DAC, 0x8000C2128000C20E,
-    0x8000C216000FFDB6, 0x000FFDB38000C21A, 0x000FB5BD8000C21E,
-    0x000F89BF000FB5BC, 0x8000C222000FB5BE, 0x000FB5BB8000C0E2,
-    0x000E05A8000FB5BA, 0x000E2DAA000E19A9, 0x000E3DAC000E25AB,
-    0x000E2DAE000E25AD, 0x00000D80000E21AF, 0x00001D8200001581,
-    0x8000C22A8000C226, 0x000230848000C22E, 0x000E35B8000130BE,
-    0x000E35BA000E3DB9, 0x000ED5BC000ECDBB, 0x000ED5BE000EDDBD,
-    0x000E29B0000ECDBF, 0x000E39B2000E29B1, 0x000E29B4000E39B3,
-    0x000E11B6000E15B5, 0x000EA5A3000E0DB7, 0x000728868000C236,
-    0x8000C23A00014C87, 0x000EB5A5000EA5A6, 0x000EEDAB8000C23E,
-    0x8000C242000EEDAA, 0x000ED1AF8000C246, 0x000EEDAD000EEDAE,
-    0x8000C24A000EEDAC, 0x8000C24E000E19B3, 0x0001F9B68000C252,
-    0x8000C25A8000C256, 0x0001D5BA000E31B5, 0x8000C25E0001DDBB,
-    0x000179BE8000C262, 0x0001C5BC000179BF, 0x000DBDA8000179BD,
-    0x000DD5AA000DD5A9, 0x000DFDAC000DE5AB, 0x000235AE00023DAD,
-    0x8000C26600022DAF, 0x8000C26E8000C26A, 0x8000C2768000C272,
-    0x8000C27E8000C27A, 0x0002E9B88000C282, 0x00028DBA0002E9B9,
-    0x000281BC000285BB, 0x000285BE000281BD, 0x000255B00002ADBF,
-    0x000255B200025DB1, 0x0002F9B40002E9B3, 0x0002E9B60002F9B5,
-    0x000D5DA30002E1B7, 0x8000C28A8000C286, 0x8000C2928000C28E,
-    0x000D75A50002BDA6, 0x000299AB8000C296, 0x8000C29A000291AA,
-    0x00023DAF8000C29E, 0x00023DAD00023DAE, 0x00005D80000281AC,
-    0x0000658200006D81, 0x8000C2A6000351B3, 0x000319B6000379B5,
-    0x00036CBE0002DC84, 0x00033DBA8000C2AA, 0x000319BC000335BB,
-    0x0000D9BE000319BD, 0x00029DA30000D9BF, 0x000568868000C2AE,
-    0x8000C2B20002BC87, 0x0002B5A50002D5A6, 0x0002F9AB8000C2B6,
-    0x000220840002F1AA, 0x000115AF8000C2BA, 0x0002D5AD000115AE,
-    0x8000C2BE0002D5AC, 0x000F20EF8000C2C2, 0x8000C2CA8000C2C6,
-    0x8000C2D28000C2CE, 0x8000C2DA8000C2D6, 0x8000C2E28000C2DE,
-    0x000C88E18000C2E6, 0x000F44E38000C2EA, 0x8000C2F28000C2EE,
-    0x8000C2FA8000C2F6, 0x00041CBE8000C2FE, 0x00005CEF8000C302,
-    0x000E74E1000E44EF, 0x000080E3000EFCE1, 0x00000D80000EB0E3,
-    0x0000718200007581, 0x8000C30A8000C306, 0x8000C30E0000C1B3,
-    0x0000FDB68000C316, 0x8000C31E8000C31A, 0x0000ADBA0000C1B5,
-    0x0004C8860000B5BB, 0x0000ADBE0004A087, 0x0000ADBC000095BF,
-    0x000585A80000A5BD, 0x000595AA000595A9, 0x0005BDAC0005A5AB,
-    0x0005D1AE0005D5AD, 0x8000C2A20005D1AF, 0x8000C3228000C312,
-    0x8000C32A8000C326, 0x8000C3328000C32E, 0x000671B88000C336,
-    0x000671BA000671B9, 0x0006D5BC000671BB, 0x0006D5BE0006DDBD,
-    0x0005B5B00006CDBF, 0x000581B20005BDB1, 0x000651B4000581B3,
-    0x000651B6000651B5, 0x0007EDB8000651B7, 0x0007F5BA0007F1B9,
-    0x000715BC00070DBB, 0x000715BE00071DBD, 0x0007EDB0000709BF,
-    0x0007FDB20007F5B1, 0x0007EDB40007F5B3, 0x0007DDB60007D5B5,
-    0x000611A80007D5B7, 0x0007CDAA000621A9, 0x0007CDAC0007DDAB,
-    0x000799AE0007F9AD, 0x8000C33A000799AF, 0x8000C3428000C33E,
-    0x8000C34A8000C346, 0x8000C2328000C34E, 0x000605A38000C352,
-    0x8000C35A8000C356, 0x8000C3628000C35E, 0x000605A5000639A6,
-    0x000671AB8000C366, 0x8000C36A000669AA, 0x000651AF8000C36E,
-    0x000661AD000669AE, 0x8000C372000669AC, 0x00001D8100001D82,
-    0x8000C37600001D80, 0x8000C37E8000C37A, 0x0001AC84000194BE,
-    0x0001E8878000C382, 0x8000C386007C0086, 0x8000C38E8000C38A,
-    0x00064DA88000C392, 0x000651AA000651A9, 0x0001BDAC0001A5AB,
-    0x0001ADAE0001A5AD, 0x8000C3960001A5AF, 0x8000C39E8000C39A,
-    0x8000C3A68000C3A2, 0x8000C3AE8000C3AA, 0x0000C5B88000C3B2,
-    0x0000C5BA0000CDB9, 0x0000CDBC0000DDBB, 0x0000F5BE0000FDBD,
-    0x0001DDB000009DBF, 0x00010DB2000105B1, 0x00011DB4000105B3,
-    0x00010DB6000105B5, 0x0005D9B3000101B7, 0x8000C3BA8000C3B6,
-    0x8000C3C28000C3BE, 0x000235B5000225B6, 0x0002C1BB8000C3C6,
-    0x8000C3CA000219BA, 0x0002C1BF8000C3CE, 0x0002D1BD0002D9BE,
-    0x8000C3D20002D9BC, 0x8000C3D600059DA3, 0x000261A68000C3DA,
-    0x8000C3E28000C3DE, 0x00025DAA000271A5, 0x8000C3E6000285AB,
-    0x00029DAE8000C3EA, 0x00029DAC000285AF, 0x0000E980000295AD,
-    0x000005820000E981, 0x0003DCBE8000C3EE, 0x8000C3F600035884,
-    0x007C00868000C3FA, 0x8000C3FE0003EC87, 0x8000C4068000C402,
-    0x8000C40E8000C40A, 0x8000C4168000C412, 0x8000C41E8000C41A,
-    0x8000C4268000C422, 0x0002B0EF8000C42A, 0x000194E18000C42E,
-    0x00026CE38000C432, 0x8000C43A8000C436, 0x8000C4428000C43E,
-    0x000279B38000C446, 0x8000C44E8000C44A, 0x8000C4568000C452,
-    0x0002B9B50002B1B6, 0x000295BB8000C45A, 0x8000C45E000295BA,
-    0x000139BF007D6484, 0x000285BD000131BE, 0x00006982000285BC,
-    0x0000598000075CEF, 0x8000C46600005181, 0x007D68878000C46A,
-    0x007D2CBE007CFC86, 0x8000C4728000C46E, 0x007C3C848000C476,
-    0x8000C47A0007B8E3, 0x8000C47E0004ECE1, 0x8000C4868000C482,
-    0x8000C48A007FC084, 0x00003CEF8000C48E, 0x000650E10006F4EF,
-    0x0000D8E30006B4E1, 0x8000C492000694E3, 0x8000C49A8000C496,
-    0x0002B9A38000C49E, 0x8000C4A68000C4A2, 0x8000C4AE8000C4AA,
-    0x000279A5000271A6, 0x000255AB8000C4B2, 0x8000C4B6000255AA,
-    0x0001F9AF8000C4BA, 0x000245AD0001F1AE, 0x007E99A8000245AC,
-    0x007EA1AA007E99A9, 0x007EA1AC007EA1AB, 0x007EA1AE007EA1AD,
-    0x8000C4BE007EA1AF, 0x8000C4C68000C4C2, 0x000019808000C4CA,
-    0x0000058200001981, 0x007EADB88000C4CE, 0x007EBDBA007EB5B9,
-    0x007EADBC007EB5BB, 0x007F51BE007F51BD, 0x007EA1B0007F51BF,
-    0x007EA1B2007EA1B1, 0x007EA1B4007EA1B3, 0x007EA5B6007EA1B5,
-    0x8000C3F2007E99B7, 0x0003C0BE007E2DB3, 0x007E25B68000C462,
-    0x8000C4D68000C4D2, 0x007E61BA007E2DB5, 0x00000886007E61BB,
-    0x007E61BE00009C87, 0x007E71BC007E61BF, 0x007E69A3007E71BD,
-    0x8000C4DE8000C4DA, 0x8000C4E68000C4E2, 0x007E69A5007E61A6,
-    0x007E25AB8000C4EA, 0x8000C4EE007E25AA, 0x007E25AF8000C4F2,
-    0x007E35AD007E25AE, 0x007E51A8007E35AC, 0x007E6DAA007E51A9,
-    0x007E7DAC007E65AB, 0x007E99AE007E6DAD, 0x8000C4F6007E99AF,
-    0x8000C4FE8000C4FA, 0x8000C5068000C502, 0x8000C50E8000C50A,
-    0x007E89B88000C512, 0x007E99BA007E89B9, 0x007EB9BC007E91BB,
-    0x000179BE007EB9BD, 0x007EE9B0000179BF, 0x007EFDB2007EF5B1,
-    0x007EC5B4007EDDB3, 0x007EC5B6007ECDB5, 0x007EE5B3007EB9B7,
-    0x8000C51A8000C516, 0x8000C5228000C51E, 0x007EF5B5007EEDB6,
-    0x007E31BB8000C526, 0x8000C52A007E2DBA, 0x007E09BF8000C52E,
-    0x007E09BD007E01BE, 0x8000C532007E29BC, 0x8000C536007EA1A3,
-    0x007EA9A68000C53A, 0x000190BE8000C53E, 0x007E69AA007EB1A5,
-    0x0007EC85007E75AB, 0x007E45AE8000C546, 0x007E6DAC007E4DAF,
-    0x000299A8007E4DAD, 0x0002C1AA0002C1A9, 0x0002C1AC0002C1AB,
-    0x0002F1AE0002C9AD, 0x000584860002F1AF, 0x8000C54A00037887,
-    0x0003F9808000C54E, 0x0003ED820003F981, 0x0003C5B88000C552,
-    0x0003C5BA0003CDB9, 0x0003CDBC0003DDBB, 0x0003F5BE0003FDBD,
-    0x000341B000039DBF, 0x000341B2000341B1, 0x000341B4000341B3,
-    0x000341B6000341B5, 0x000249B3000341B7, 0x8000C55A8000C556,
-    0x8000C5628000C55E, 0x000249B5000249B6, 0x000385BB8000C566,
-    0x8000C56A000385BA, 0x000385BF8000C56E, 0x000395BD000385BE,
-    0x8000C572000395BC, 0x8000C57600020DA3, 0x00020DA68000C57A,
-    0x8000C5828000C57E, 0x0003C1AA00020DA5, 0x8000C5860003C1AB,
-    0x0003C1AE8000C58A, 0x0003D1AC0003C1AF, 0x007E60E30003D1AD,
-    0x007E4CE1007FBCE1, 0x0001A0E1007E30E3, 0x007E84E30000E8E1,
-    0x0007ACBE0017C4E3, 0x000008EF8000C58E, 0x000140EF001C7CEF,
-    0x8000C5968000C592, 0x8000C59A007EA4EF, 0x000404848000C59E,
-    0x8000C5A20006C0BE, 0x00003D8100003D82, 0x8000C5A600003D80,
-    0x8000C5AE8000C5AA, 0x000750868000C5B6, 0x8000C5BA00058487,
-    0x007DC1BB8000C5BE, 0x007DA1B9007DA9BA, 0x007DC9BF007D91B8,
-    0x007DD5BD007DC1BE, 0x007E1DB3007DD5BC, 0x007E1DB1007E1DB2,
-    0x007DA1B7007E1DB0, 0x007DB1B5007DB9B6, 0x007EBDAB007DB9B4,
-    0x007EBDA9007EBDAA, 0x007E7DAF007EBDA8, 0x007E7DAD007E7DAE,
-    0x8000C5C2007E7DAC, 0x8000C5CA8000C5C6, 0x8000C5D28000C5CE,
-    0x8000C5DA8000C5D6, 0x006C05820006B0BE, 0x8000C542006C1983,
-    0x0068F9868000C5B2, 0x006819840074C987, 0x0074018A00681985,
-    0x8000C5DE0074B98B, 0x0070B58E8000C5E2, 0x0070D98C007CC18F,
-    0x007C619200701D8D, 0x8000C5E6007C6193, 0x007875968000C5EA,
-    0x0078A19400793197, 0x0004E19A00787D95, 0x8000C5EE0004E99B,
-    0x8000C5F68000C5F2, 0x0004ED9C8000C5FA, 0x000089A18000C5FE,
-    0x000C01A38000C602, 0x000C15A5000081A2, 0x0008B9A7000CB5A4,
-    0x0009F9A9000801A6, 0x001425AB000881A8, 0x001001AD00142DAA,
-    0x001199AF001421AC, 0x001CE1B100109DAE, 0x001801B3001C01B0,
-    0x0018F5B5001C65B2, 0x8000C6060018F5B4, 0x8000C60E8000C60A,
-    0x8000C6168000C612, 0x000019808000C61A, 0x0000058200001981,
-    0x000729A98000C61E, 0x000735AB0004E9A8, 0x000799AD000721AA,
-    0x000789AF000791AC, 0x007845A0000791AE, 0x0005A5A20005A5A1,
-    0x0001B8840005B9A3, 0x000034878000C626, 0x0007ADB9000F9C86,
-    0x00067DBB0007ADB8, 0x000611BD000669BA, 0x000609BF00066DBC,
-    0x0007C9B1000619BE, 0x0007A9B30007F9B0, 0x0007BDB50007C1B2,
-    0x0007A1B70007BDB4, 0x0007ADB30007A9B6, 0x8000C62E8000C62A,
-    0x8000C63200000C84, 0x00071DB500071DB6, 0x000709BB006D998F,
-    0x8000C636000739BA, 0x00077DBF8000C63A, 0x000779BD00077DBE,
-    0x0071E198000711BC, 0x007D3D9A00714199, 0x007DF59C007DED9B,
-    0x0079F99E00792D9D, 0x006DFD900079F99F, 0x0069ED920069E591,
-    0x00752D9400690593, 0x007565960075D595, 0x00042DA300711D97,
-    0x000DC0BE000DC884, 0x8000C6428000C63E, 0x00049DA500049DA6,
-    0x000489AB8000C646, 0x8000C64A0004B9AA, 0x0004FDAF8000C64E,
-    0x0004F9AD0004FDAE, 0x8000C652000491AC, 0x8000C656000539B3,
-    0x0005B9B68000C65A, 0x8000C6628000C65E, 0x000595BA0005B9B5,
-    0x8000C66600059DBB, 0x000239BE8000C66A, 0x000251BC000239BF,
-    0x0005EDA8000251BD, 0x0005F5AA0005FDA9, 0x000575AC00056DAB,
-    0x000575AE00057DAD, 0x8000C66E00056DAF, 0x8000C6768000C672,
-    0x8000C67E8000C67A, 0x8000C6868000C682, 0x0002B1B88000C68A,
-    0x0002B5BA0002B1B9, 0x000209BC000249BB, 0x000239BE000209BD,
-    0x000515B0000239BF, 0x0002C1B200051DB1, 0x0002B9B40002C1B3,
-    0x000291B60002B9B5, 0x00047DA3000291B7, 0x0001B58100001582,
-    0x8000C68E0001B580, 0x0004FDA50004FDA6, 0x0004D9AB8000C696,
-    0x000D28860004D1AA, 0x00037DAF000D1C87, 0x000315AD00037DAE,
-    0x8000C69A000315AC, 0x8000C6A28000C69E, 0x8000C6A60002C5B3,
-    0x0002F5B60002FDB5, 0x8000C6AE8000C6AA, 0x000175BA8000C6B2,
-    0x00015DBC000145BB, 0x000131BE000135BD, 0x0000A0E3000131BF,
-    0x0006E0E10006DCE3, 0x8000C6B60007B4E1, 0x8000C6BE8000C6BA,
-    0x000C8086000F90BE, 0x8000C6C2000C1887, 0x000604EF8000C6C6,
-    0x8000C6CA00066CEF, 0x000249A38000C6CE, 0x8000C6D68000C6D2,
-    0x8000C6DE8000C6DA, 0x000271A5000279A6, 0x0001C9AB8000C6E2,
-    0x8000C6E60001F9AA, 0x0001BDAF8000C6EA, 0x0001B9AD0001BDAE,
-    0x000D35A80001D1AC, 0x000D81AA000D81A9, 0x000DB9AC000D9DAB,
-    0x000E15AE000DB9AD, 0x8000C692000E0DAF, 0x8000C6F28000C6EE,
-    0x00001D808000C6F6, 0x0000B98200000981, 0x000E29B88000C6FA,
-    0x000EE9BA000E29B9, 0x000EF9BC000EE9BB, 0x000EE9BE000EF9BD,
-    0x000E7DB0000EE9BF, 0x000E45B2000E4DB1, 0x000E39B4000E2DB3,
-    0x000E2DB6000E39B5, 0x8000C6FE000E25B7, 0x8000C6228000C702,
-    0x8000C70A8000C706, 0x8000C7128000C70E, 0x000FBDA88000C716,
-    0x000FEDAA000FE5A9, 0x000FE1AC000FE5AB, 0x000FD9AE000FE9AD,
-    0x000F51B0000FD9AF, 0x000F51B2000F51B1, 0x000F7DB4000F65B3,
-    0x000F65B6000F6DB5, 0x000F29B8000F19B7, 0x000F39BA000F29B9,
-    0x000F19BC000F31BB, 0x0000DDBE000F19BD, 0x000E45B30000D5BF,
-    0x8000C71E8000C71A, 0x8000C7268000C722, 0x000E49B5000E49B6,
-    0x000EB5BB8000C72A, 0x8000C72E000EB5BA, 0x000E89BF8000C732,
-    0x000E91BD000E99BE, 0x8000C736000EA5BC, 0x8000C73A000E01A3,
-    0x000E0DA68000C73E, 0x8000C7468000C742, 0x000EF1AA000E0DA5,
-    0x8000C74A000EF1AB, 0x000EDDAE8000C74E, 0x000EE1AC000ECDAF,
-    0x00010DA8000ED5AD, 0x00011DAA000115A9, 0x00010DAC000115AB,
-    0x000091AE000091AD, 0x8000C752000091AF, 0x8000C75A8000C756,
-    0x8000C7628000C75E, 0x0003B1810003B982, 0x000099B80003B980,
-    0x0000A5BA0000ADB9, 0x000075BC00006DBB, 0x000075BE00007DBD,
-    0x0000F1B000006DBF, 0x0000C1B20000F9B1, 0x0000B1B40000C1B3,
-    0x0000B5B60000BDB5, 0x0002AC840000A9B7, 0x8000C76A8000C766,
-    0x0004B0868000C76E, 0x8000C7760002D887, 0x8000C77E8000C77A,
-    0x8000C7868000C782, 0x000374BE8000C78A, 0x8000C7928000C78E,
-    0x00035DBB8000C796, 0x8000C79A000355BA, 0x000319BF8000C79E,
-    0x000375BD000315BE, 0x0002F5B3000375BC, 0x8000C7A68000C7A2,
-    0x8000C7AE8000C7AA, 0x0002E5B5000371B6, 0x000215A38000C7B2,
-    0x8000C7BA8000C7B6, 0x8000C7C28000C7BE, 0x000205A5000391A6,
-    0x0003BDAB8000C7C6, 0x8000C7CA0003B5AA, 0x0003F9AF8000C7CE,
-    0x000395AD0003F5AE, 0x8000C7D2000395AC, 0x00001D8100001D82,
-    0x8000C7D600001D80, 0x8000C7DE8000C7DA, 0x00036C848000C7E6,
-    0x000378878000C7EA, 0x8000C7EE0004EC86, 0x8000C7F68000C7F2,
-    0x000EB8E18000C7FA, 0x000318E3000C98E3, 0x8000C7FE000C28E1,
-    0x8000C8068000C802, 0x0004EC848000C80A, 0x8000C8128000C80E,
-    0x8000C81A8000C816, 0x000E5CEF000A30EF, 0x000161B38000C81E,
-    0x8000C8268000C822, 0x8000C82E8000C82A, 0x000171B5000179B6,
-    0x000129BB8000C832, 0x8000C836000159BA, 0x00011DBF8000C83A,
-    0x00011DBD00011DBE, 0x8000C83E000131BC, 0x8000C8468000C842,
-    0x000019808000C84A, 0x0000398200001981, 0x8000C84E000F6CEF,
-    0x000460BE000E94E1, 0x8000C856000FB8E3, 0x8000C85E8000C85A,
-    0x0004E086000000EF, 0x0001E9A300048C87, 0x0001F9A500070084,
-    0x8000C8668000C862, 0x8000C86A0001F1A6, 0x0001A1AB8000C86E,
-    0x000195AD0001D1AA, 0x000195AF0001B9AC, 0x000619A8000195AE,
-    0x000629AA000619A9, 0x000639AC000629AB, 0x00062DAE000639AD,
-    0x8000C7E2000625AF, 0x8000C8728000C852, 0x8000C87A8000C876,
-    0x8000C8828000C87E, 0x0006C5B88000C886, 0x0006C5BA0006CDB9,
-    0x0006CDBC0006DDBB, 0x0006F5BE0006FDBD, 0x00065DB000069DBF,
-    0x000625B200062DB1, 0x000619B4000609B3, 0x000609B6000619B5,
-    0x8000C88A000601B7, 0x8000C8928000C88E, 0x8000C89A8000C896,
-    0x8000C8A28000C89E, 0x00067DA88000C772, 0x000659AA000641A9,
-    0x000649AC000659AB, 0x0007A9AE000649AD, 0x0007D9B00007A9AF,
-    0x0007EDB20007D9B1, 0x0007FDB40007E5B3, 0x0007E5B60007E5B5,
-    0x0007A9B8000799B7, 0x0007B9BA0007A9B9, 0x000799BC0007B1BB,
-    0x000779BE000799BD, 0x0007F1B3000779BF, 0x8000C8AA8000C8A6,
-    0x8000C8B28000C8AE, 0x0007E1B50007E9B6, 0x00063DBB8000C8B6,
-    0x8000C8BA00063DBA, 0x000601BF8000C8BE, 0x000611BD000619BE,
-    0x00006D82000619BC, 0x000055800007B5A3, 0x0007ADA600006581,
-    0x8000C8C200019CBE, 0x000679AA0007A5A5, 0x001C0086000679AB,
-    0x00065DAE0001CC87, 0x00065DAC000645AF, 0x000631A8000655AD,
-    0x000631AA000631A9, 0x000649AC000659AB, 0x00067DAE000675AD,
-    0x8000C8C60001B5AF, 0x8000C8CE8000C8CA, 0x8000C8D68000C8D2,
-    0x8000C8DE8000C8DA, 0x00014DB88000C8E2, 0x00015DBA000155B9,
-    0x0000D5BC000155BB, 0x0000D5BE0000DDBD, 0x0001CDB00000CDBF,
-    0x0001DDB20001D5B1, 0x0001CDB40001D5B3, 0x00017DB6000175B5,
-    0x0005B9B3000175B7, 0x8000C8EA8000C8E6, 0x8000C8F28000C8EE,
-    0x0005C1B50005F5B6, 0x000211BB8000C8F6, 0x8000C8FA000209BA,
-    0x0002B1BF8000C8FE, 0x000201BD000209BE, 0x8000C902000209BC,
-    0x8000C9060005FDA3, 0x0005B1A68000C90A, 0x8000C9128000C90E,
-    0x00024DAA000585A5, 0x8000C916000255AB, 0x00024DAE8000C91A,
-    0x00024DAC0002F5AF, 0x8000C91E000245AD, 0x8000C9268000C922,
-    0x8000C92E8000C92A, 0x8000C9360003D8BE, 0x000029808000C93A,
-    0x0000398200002981, 0x001C108600032084, 0x8000C93E00038487,
-    0x8000C9468000C942, 0x8000C94E8000C94A, 0x8000C9568000C952,
-    0x8000C95A000268E3, 0x8000C95E0001ACE1, 0x8000C962000288EF,
-    0x8000C96A8000C966, 0x8000C9728000C96E, 0x8000C97A8000C976,
-    0x8000C9828000C97E, 0x0007E0E38000C986, 0x000710E18000C98A,
-    0x0007E8EF8000C98E, 0x8000C9968000C992, 0x000019808000C99A,
-    0x0000058200001981, 0x0002E9B38000C99E, 0x001CE886001D0C84,
-    0x8000C9A6001D8087, 0x0002E9B50002E1B6, 0x000169BB8000C9AA,
-    0x001D24BE000161BA, 0x000159BF8000C9AE, 0x000169BD000169BE,
-    0x8000C9B2000171BC, 0x8000C9BA8000C9B6, 0x8000C9C28000C9BE,
-    0x000034EF8000C9C6, 0x000678E10006CCEF, 0x000674E3000750E1,
-    0x8000C9CA00069CE3, 0x8000C9D28000C9CE, 0x000269A38000C9D6,
-    0x001C84848000C9DA, 0x8000C9DE001C84BE, 0x000269A5000261A6,
-    0x0001E9AB8000C9E2, 0x8000C9E60001E1AA, 0x0001D9AF8000C9EA,
-    0x0001E9AD0001E9AE, 0x001E99A80001F1AC, 0x001EB1AA001E99A9,
-    0x001ED1AC001EB1AB, 0x001ED1AE001ED1AD, 0x8000C9EE001ED1AF,
-    0x8000C9F68000C9F2, 0x8000C9FE8000C9FA, 0x8000CA068000CA02,
-    0x001EBDB88000CA0A, 0x001F4DBA001F45B9, 0x001F5DBC001F45BB,
-    0x001F45BE001F41BD, 0x001EB1B0001F79BF, 0x001E89B2001EBDB1,
-    0x001E9DB4001E85B3, 0x001E8DB6001E85B5, 0x8000C932001E85B7,
-    0x8000CA0E001E3DB3, 0x001E55B68000C9A2, 0x8000CA168000CA12,
-    0x001E71BA001E5DB5, 0x00000886001E7DBB, 0x001E5DBE00002887,
-    0x001E59BC001E45BF, 0x001E79A3001E55BD, 0x0000F58100004D82,
-    0x8000CA1A0000F580, 0x001E19A5001E11A6, 0x001E39AB8000CA1E,
-    0x8000CA22001E35AA, 0x001E01AF8000CA26, 0x001E11AD001E19AE,
-    0x001E51A8001E1DAC, 0x001E65AA001E59A9, 0x001E95AC001E8DAB,
-    0x001E95AE001E9DAD, 0x8000CA2A001E8DAF, 0x8000CA328000CA2E,
-    0x8000CA3A8000CA36, 0x8000CA428000CA3E, 0x001EA9B88000CA46,
-    0x000169BA001EA9B9, 0x000179BC000169BB, 0x000169BE000179BD,
-    0x001EF5B0000169BF, 0x001EDDB2001ED5B1, 0x001ECDB4001ED5B3,
-    0x001EA9B6001EB1B5, 0x001EEDB3001EA1B7, 0x8000CA4E8000CA4A,
-    0x8000CA568000CA52, 0x001E1DB5001E1DB6, 0x001E11BB8000CA5A,
-    0x8000CA5E001E31BA, 0x001E75BF8000CA62, 0x001E01BD001E09BE,
-    0x8000CA66001E09BC, 0x8000CA6A001EA9A3, 0x001E59A68000CA6E,
-    0x8000CA768000CA72, 0x001E75AA001E59A5, 0x8000CA7A001E55AB,
-    0x001E4DAE8000CA7E, 0x001E4DAC001E31AF, 0x000299A8001E45AD,
-    0x0002D5AA0002D5A9, 0x0002FDAC0002E5AB, 0x00033DAE000335AD,
-    0x00026CBE000335AF, 0x8000CA8A8000CA86, 0x8000CA928000CA8E,
-    0x8000CA9A8000CA96, 0x0003D1B88000CA9E, 0x0003E1BA0003D9B9,
-    0x000391BC0003E1BB, 0x000395BE00039DBD, 0x00034DB0000389BF,
-    0x00035DB2000355B1, 0x00034DB4000355B3, 0x0003F1B60003F1B5,
-    0x000369810003F1B7, 0x000255B300035D80, 0x00027DB500036182,
-    0x8000CAA68000CAA2, 0x0002E4840003BDB6, 0x000399BB8000CAAA,
-    0x000389BD000381BA, 0x00033DBF000381BC, 0x00021DA3000381BE,
-    0x000398878000CAAE, 0x8000CAB200046C86, 0x000235A50003F5A6,
-    0x0003D1AB8000CAB6, 0x8000CABA0003C9AA, 0x000375AF8000CABE,
-    0x0003C1AD0003C9AE, 0x001ED8E30003C9AC, 0x001E58E1001FACE1,
-    0x0001A0E1001E58E3, 0x0003A8E38000CAC2, 0x8000CACA8000CAC6,
-    0x0003F0EF8000CACE, 0x001E04EF8000CAD2, 0x8000CADA8000CAD6,
-    0x8000CADE001EACEF, 0x00051C848000CAE2, 0x00000D80000500BE,
-    0x00003D8200003581, 0x8000CAEA8000CAE6, 0x0004C8868000CAEE,
-    0x8000CAF60005A087, 0x8000CAFE8000CAFA, 0x001DD5BA8000CB02,
-    0x001DC1B8001DDDBB, 0x001D11BE001DC1B9, 0x001DF1BC001D11BF,
-    0x001DADB2001DF1BD, 0x001DADB0001DB5B3, 0x001D9DB6001DA5B1,
-    0x001DA5B4001DF5B7, 0x001E69AA001D95B5, 0x001E69A8001E69AB,
-    0x001E69AE001E69A9, 0x001E69AC001E69AF, 0x8000CB06001E69AD,
-    0x8000CB0E8000CB0A, 0x8000CB168000CB12, 0x8000CB1E8000CB1A,
-    0x000625A88000CB22, 0x000625AA00062DA9, 0x000625AC00063DAB,
-    0x000625AE00062DAD, 0x8000CAF2000689AF, 0x8000CB2A8000CB26,
-    0x8000CB328000CB2E, 0x8000CB3A8000CB36, 0x00075DB88000CB3E,
-    0x00076DBA000765B9, 0x000765BC000761BB, 0x000719BE00076DBD,
-    0x0006F9B0000719BF, 0x0006ADB20006C1B1, 0x0006ADB40006BDB3,
-    0x000695B600069DB5, 0x0006B5B5000769B7, 0x8000CB468000CB42,
-    0x8000CB4A0006B5B6, 0x0006B5B38000CB4E, 0x000671BD8000CB52,
-    0x000641BF000661BC, 0x8000CA82000671BE, 0x00061DBB8000CB56,
-    0x0006EDA3000625BA, 0x0000198100001982, 0x8000CB5A00007980,
-    0x0006EDA50006EDA6, 0x000645AB8000CB5E, 0x0003A48400067DAA,
-    0x000619AF0003A4BE, 0x000629AD000629AE, 0x000671A8000639AC,
-    0x000671AA000671A9, 0x0006F1AC000671AB, 0x0006F1AE0006F1AD,
-    0x8000CB620006F1AF, 0x000188868000CB66, 0x8000CB6A0001AC87,
-    0x8000CB728000CB6E, 0x000175B88000CB76, 0x000175BA00017DB9,
-    0x0001D9BC0001C9BB, 0x0001CDBE0001C5BD, 0x000691B00001C1BF,
-    0x000691B2000691B1, 0x000155B4000691B3, 0x000155B600015DB5,
-    0x0006F9B300014DB7, 0x8000CB7E8000CB7A, 0x8000CB868000CB82,
-    0x000679B5000671B6, 0x000655BB8000CB8A, 0x8000CB8E000655BA,
-    0x000199BF8000CB92, 0x000645BD000191BE, 0x8000CB96000645BC,
-    0x8000CB9A0006BDA3, 0x000635A68000CB9E, 0x8000CBA68000CBA2,
-    0x000611AA00063DA5, 0x8000CBAA000611AB, 0x0001D5AE8000CBAE,
-    0x000601AC0001DDAF, 0x000325A8000601AD, 0x00033DAA000335A9,
-    0x00032DAC000335AB, 0x000391AE000391AD, 0x8000CBB2000391AF,
-    0x8000CBBA8000CBB6, 0x8000CBC28000CBBE, 0x8000CBCA8000CBC6,
-    0x000399B88000CBCE, 0x0003A5BA0003ADB9, 0x000075BC00006DBB,
-    0x000075BE00007DBD, 0x0003F1B000006DBF, 0x0003C1B20003F9B1,
-    0x0003B1B40003C1B3, 0x0003B5B60003BDB5, 0x0000B5800003A9B7,
-    0x0000B5820000BD81, 0x00031087000348BE, 0x8000CBD6000C3486,
-    0x000CACBE0002E8EF, 0x0002A484000194E1, 0x8000CBDA000214E3,
-    0x8000CBE28000CBDE, 0x0003F5B38000CBE6, 0x8000CBEE8000CBEA,
-    0x8000CBF2000A7485, 0x0003E5B5000351B6, 0x00037DBB8000CBF6,
-    0x8000CBFA000375BA, 0x000339BF8000CBFE, 0x000355BD000335BE,
-    0x8000CC02000355BC, 0x8000CC0A8000CC06, 0x8000CC0E0006B8EF,
-    0x8000CC168000CC12, 0x8000CC1E8000CC1A, 0x8000CC268000CC22,
-    0x000000CC8000CC2A, 0x8000CC2E0007E0E1, 0x8000CC32000664E3,
-    0x00027DA38000CC36, 0x00026DA500024084, 0x8000CC3E8000CC3A,
-    0x000CC0860002D9A6, 0x0002F5AB000DAC87, 0x0002DDAD0002FDAA,
-    0x0002B1AF0002DDAC, 0x0000C5800002BDAE, 0x0000D1820000D181,
-    0x8000CC4A8000CC46, 0x8000CC528000CC4E, 0x000CEC84000188EF,
-    0x8000CC560006D4E1, 0x8000CC5A000154E3, 0x8000CC628000CC5E,
-    0x000175B38000CC66, 0x8000CC6E8000CC6A, 0x8000CC768000CC72,
-    0x000165B5000111B6, 0x000109BB8000CC7A, 0x8000CC7E000139BA,
-    0x0000F9BF8000CC82, 0x0000F9BD0000F1BE, 0x000E99A8000111BC,
-    0x000EBDAA000EB5A9, 0x000EADAC000EB5AB, 0x000ED1AE000ED1AD,
-    0x8000CC42000ECDAF, 0x8000CC8A8000CC86, 0x8000CC928000CC8E,
-    0x8000CC9A8000CC96, 0x000F59B88000CC9E, 0x000F69BA000F59B9,
-    0x000F79BC000F69BB, 0x000F61BE000F79BD, 0x000EBDB0000F1DBF,
-    0x000E85B2000E8DB1, 0x000F75B4000F6DB3, 0x000F69B6000F7DB5,
-    0x000E35A3000F69B7, 0x8000CCA68000CCA2, 0x8000CCAE8000CCAA,
-    0x000E25A5000E51A6, 0x000E49AB0003D4BE, 0x8000CCB2000E79AA,
-    0x000FB9AF8000CCB6, 0x000FB9AD000FB1AE, 0x00005182000E51AC,
-    0x000041808000CBD2, 0x0003188700004181, 0x8000CCBA00007486,
-    0x8000CCC28000CCBE, 0x8000CCCA8000CCC6, 0x8000CCD28000CCCE,
-    0x8000CCDA8000CCD6, 0x000E8DA88000CCDE, 0x000E9DAA000E95A9,
-    0x000EB9AC000E95AB, 0x000EADAE000EA5AD, 0x8000CCE2000EA5AF,
-    0x8000CCEA8000CCE6, 0x8000CCF28000CCEE, 0x8000CCFA8000CCF6,
-    0x000165B88000CCFE, 0x000165BA00016DB9, 0x000165BC00017DBB,
-    0x000165BE00016DBD, 0x000EDDB00001D9BF, 0x000EA1B2000EA1B1,
-    0x000EA1B4000EA1B3, 0x000E91B6000EA9B5, 0x000E15B3000E91B7,
-    0x8000CD068000CD02, 0x8000CD0E8000CD0A, 0x000E31B5000E25B6,
-    0x000E65BB8000CD12, 0x8000CD16000E19BA, 0x000E6DBF8000CD1A,
-    0x000E75BD000E65BE, 0x00001582000E75BC, 0x00006180000E51A3,
-    0x000E61A600006181, 0x000190BE8000CD1E, 0x000E5DAA000E75A5,
-    0x0001A484000E21AB, 0x000E21AE8000CD26, 0x000E31AC000E29AF,
-    0x0002ADA8000E31AD, 0x0002D5AA0002DDA9, 0x0002F5AC0002EDAB,
-    0x0002F5AE0002FDAD, 0x8000CD2A0002EDAF, 0x8000CD328000CD2E,
-    0x8000CD3A8000CD36, 0x8000CD428000CD3E, 0x000375B88000CD46,
-    0x000375BA00037DB9, 0x0003D9BC0003C9BB, 0x0003C9BE0003D9BD,
-    0x000295B00003C1BF, 0x000295B200029DB1, 0x000355B400034DB3,
-    0x000355B600035DB5, 0x00025DB300034DB7, 0x0003D8878000CD4A,
-    0x8000CD4E00042C86, 0x00027DB500027DB6, 0x000259BB8000CD52,
-    0x8000CD56000251BA, 0x000399BF8000CD5A, 0x000399BD000391BE,
-    0x8000CD5E000241BC, 0x00056886000219A3, 0x000239A600030087,
-    0x8000CD668000CD62, 0x000215AA000239A5, 0x8000CD6A00021DAB,
-    0x0003D5AE8000CD6E, 0x000205AC0003DDAF, 0x8000CD720003DDAD,
-    0x8000CD7A8000CD76, 0x8000CD828000CD7E, 0x8000CD86000034EF,
-    0x00052CBE8000CD8A, 0x8000CD8E0000DCE3, 0x8000CD960001E0E1,
-    0x8000CD9E8000CD9A, 0x000FCCE18000CDA2, 0x000E14E300024884,
-    0x0000498000054484, 0x0000718200004981, 0x8000CDAA8000CDA6,
-    0x8000CDB28000CDAE, 0x8000CDBA8000CDB6, 0x000F20EF8000CDBE,
-    0x8000CDC68000CDC2, 0x8000CDCE8000CDCA, 0x8000CDD68000CDD2,
-    0x8000CDDE8000CDDA, 0x0004ACBE0001F8EF, 0x8000CDE2000EDCE1,
-    0x8000CDE6000E1CE3, 0x8000CDEE8000CDEA, 0x0001E5B38000CDF2,
-    0x0000158100002D82, 0x8000CDF600001D80, 0x0001F5B50001E5B6,
-    0x000165BB8000CDFA, 0x8000CDFE0001D9BA, 0x00015DBF8000CE06,
-    0x000165BD00015DBE, 0x0005EDA800016DBC, 0x00062DAA000625A9,
-    0x00063DAC000625AB, 0x00062DAE000625AD, 0x8000CD92000625AF,
-    0x000088868000CE02, 0x8000CE0A0003A087, 0x8000CE128000CE0E,
-    0x000699B88000CE16, 0x0006A5BA0006ADB9, 0x000755BC00074DBB,
-    0x000755BE00075DBD, 0x00065DB000074DBF, 0x0006F9B20006E1B1,
-    0x0006D9B40006F1B3, 0x0006B9B60006D9B5, 0x0006A5A30006A9B7,
-    0x8000CE1E8000CE1A, 0x8000CE268000CE22, 0x0006B5A50006A5A6,
-    0x000625AB8000CE2A, 0x8000CE2E000699AA, 0x00061DAF8000CE32,
-    0x000625AD00061DAE, 0x8000CE3600062DAC, 0x8000CD220007FDB3,
-    0x0007B1B68000CE3A, 0x8000CE428000CE3E, 0x00078DBA000781B5,
-    0x8000CE46000795BB, 0x000771BE8000CE4A, 0x00078DBC000771BF,
-    0x000631A8000781BD, 0x000631AA000631A9, 0x00064DAC000631AB,
-    0x00067DAE000675AD, 0x8000CE4E000675AF, 0x8000CE568000CE52,
-    0x8000CE5E8000CE5A, 0x8000CE668000CE62, 0x0006D1B88000CE6A,
-    0x0006D1BA0006D1B9, 0x0006E1BC0006E5BB, 0x0006E1BE0006E1BD,
-    0x00060DB00006E1BF, 0x00061DB2000615B1, 0x0006F1B4000611B3,
-    0x0006F1B60006F1B5, 0x0006B9A30006F1B7, 0x0000B58100001582,
-    0x8000CE6E0000B580, 0x0006C5A50006F5A6, 0x0006D1AB000190BE,
-    0x0002A0840006C9AA, 0x000635AF003C00BF, 0x0006C5AD000635AE,
-    0x000084850006C9AC, 0x003F08860001F1B3, 0x000155B600009C87,
-    0x8000CE7A8000CE76, 0x000171BA0001E1B5, 0x8000CE7E000179BB,
-    0x000131BE8000CE82, 0x000151BC000139BF, 0x0005EDA8000151BD,
-    0x00023DAA000235A9, 0x00022DAC000235AB, 0x000291AE000291AD,
-    0x8000CE86000291AF, 0x8000CE8E8000CE8A, 0x000211A28000CE92,
-    0x000691A0000211A3, 0x0002A9B80002F5A1, 0x000369BA0002A9B9,
-    0x000379BC000369BB, 0x000369BE000379BD, 0x0002F1B0000369BF,
-    0x0002C1B20002F9B1, 0x0002B1B40002C1B3, 0x0002ADB60002B9B5,
-    0x0002B5A30002A5B7, 0x8000CE96001A8D8F, 0x8000CE9E8000CE9A,
-    0x0002A5A5000211A6, 0x00023DAB8000CEA2, 0x8000CEA6000235AA,
-    0x00027DAF8000CEAA, 0x000215AD000275AE, 0x0006059E000215AC,
-    0x000AE99C0006E59F, 0x000E459A000BED9D, 0x000E5998000A659B,
-    0x0012C596000E5199, 0x001275940013BD97, 0x0016D1920012CD95,
-    0x001AF59000167193, 0x8000CEAE00167591, 0x003D44840000D0E1,
-    0x8000CEB2002410E3, 0x8000CEBA8000CEB6, 0x8000CEC28000CEBE,
-    0x8000CECA8000CEC6, 0x8000CED28000CECE, 0x8000CEDA8000CED6,
-    0x8000CEDE0024CCEF, 0x003D00BE000610E3, 0x000065800007D8E1,
-    0x0000198200006D81, 0x003CC0868000CEE6, 0x8000CEEA003C4487,
-    0x8000CEF28000CEEE, 0x8000CEF60006ACEF, 0x002B85B18000CEFA,
-    0x002AA1B3002801B0, 0x002791B5002B9DB2, 0x0000FCEF002401B4,
-    0x000190E10006C4EF, 0x000010E3000774E1, 0x8000CEFE0006F0E3,
-    0x003CF4848000CF02, 0x003F95A18000CF06, 0x003F21A38000CF0A,
-    0x003BD5A5003F49A2, 0x003401A7003BD5A4, 0x003741A9003B01A6,
-    0x0033E9AB0037E1A8, 0x002C01AD003001AA, 0x002FA1AF0033C1AC,
-    0x003E89B0002FF9AE, 0x003E99B2003E89B1, 0x003EB9B4003E91B3,
-    0x003F19B6003EB9B5, 0x003F35B8003F0DB7, 0x003F35BA003F3DB9,
-    0x003F15BC003F0DBB, 0x003F15BE003F1DBD, 0x8000CF0E003F0DBF,
-    0x8000CF168000CF12, 0x8000CF1E8000CF1A, 0x8000CF228000CE72,
-    0x003DB5A88000CF26, 0x003D81AA003DBDA9, 0x003E51AC003D81AB,
-    0x003E51AE003E51AD, 0x003D69B3003E51AF, 0x8000CF2A8000CEE2,
-    0x8000CF328000CF2E, 0x003EB9B5003EB1B6, 0x003E2DBB8000CF36,
-    0x8000CF3A003E2DBA, 0x003E1DBF8000CF3E, 0x003E25BD003E1DBE,
-    0x8000CF42003E2DBC, 0x8000CF46003D2DA3, 0x003EF5A68000CF4A,
-    0x000390BE8000CF4E, 0x003E69AA003EFDA5, 0x8000CF52003E69AB,
-    0x003E59AE8000CF56, 0x003E69AC003E59AF, 0x003E5DA8003E61AD,
-    0x003E65AA003E6DA9, 0x003E69AC003E7DAB, 0x003EA9AE003E69AD,
-    0x0001E980003EA9AF, 0x0001FD820001E981, 0x000190868000CF5A,
-    0x8000CF5E0001B487, 0x00011DB88000CF62, 0x0001C5BA0001CDB9,
-    0x0001CDBC0001DDBB, 0x0001F5BE0001FDBD, 0x003EDDB000019DBF,
-    0x000129B2003EA1B1, 0x000139B4000129B3, 0x000129B6000139B5,
-    0x003EEDB3000125B7, 0x8000CF6A8000CF66, 0x8000CF728000CF6E,
-    0x003EE1B5003EE9B6, 0x003E65BB8000CF76, 0x8000CF7A003E19BA,
-    0x0001FDBF0001A084, 0x0001FDBD0001FDBE, 0x8000CF7E0001FDBC,
-    0x8000CF82003EA9A3, 0x003EADA68000CF86, 0x8000CF8E8000CF8A,
-    0x003E5DAA003EA5A5, 0x8000CF92003E21AB, 0x0001B9AE8000CF96,
-    0x0001B9AC0001B9AF, 0x0002EDA80001B9AD, 0x000331AA000329A9,
-    0x000395AC000331AB, 0x000395AE00039DAD, 0x000400BE00038DAF,
-    0x8000CF9E8000CF9A, 0x8000CFA68000CFA2, 0x8000CFAE8000CFAA,
-    0x0003ADB88000CFB2, 0x000061BA000061B9, 0x000061BC000061BB,
-    0x000061BE000061BD, 0x0003FDB0000061BF, 0x0003A9B20003CDB1,
-    0x0003BDB40003A5B3, 0x0003A1B60003A1B5, 0x00000D80000395B7,
-    0x00001D8200001581, 0x8000CFBA8000CFB6, 0x8000CFC68000CFBE,
-    0x000420868000CFCA, 0x8000CFCE0003E487, 0x8000CFD68000CFD2,
-    0x8000CFDE8000CFDA, 0x0003F9B38000CFE2, 0x8000CFEA8000CFE6,
-    0x8000CFF28000CFEE, 0x00035DB500035DB6, 0x000349BB8000CFF6,
-    0x8000CFFA000379BA, 0x000325BF8000CFFE, 0x000331BD000339BE,
-    0x003E78E100033DBC, 0x003EA0E38000D002, 0x8000D00A8000D006,
-    0x8000D0128000D00E, 0x00052C848000D016, 0x0004A4BE8000D01A,
-    0x8000D0228000D01E, 0x000118EF8000D026, 0x000075808000D02A,
-    0x0000198200007981, 0x8000D02E0004C484, 0x0005948700038CEF,
-    0x8000D03600047C86, 0x8000D03A0000D8E3, 0x8000D03E000044E1,
-    0x8000D0468000D042, 0x002FA58B8000D04A, 0x8000D04E002FA58A,
-    0x0033058F8000D052, 0x0033198D0033198E, 0x0027AD83002EC58C,
-    0x8000D0560027AD82, 0x002F2D878000D05A, 0x002B9985002B3186,
-    0x003FE19B002B3584, 0x8000D05E003F0D9A, 0x003C00E18000D062,
-    0x003EC0E38000D066, 0x0037C193003FF99C, 0x003E58EF00375592,
-    0x003B59978000D06A, 0x003B3D95003BD596, 0x8000CFC20036DD94,
-    0x8000D06E8000D032, 0x8000D0768000D072, 0x8000D07E8000D07A,
-    0x0006B9A88000D082, 0x000681AA0006B9A9, 0x000681AC000681AB,
-    0x000681AE000681AD, 0x000685B0000681AF, 0x000685B200068DB1,
-    0x00068DB400069DB3, 0x0006B5B60006BDB5, 0x00076DB8000759B7,
-    0x00077DBA000775B9, 0x00076DBC000775BB, 0x000755BE000751BD,
-    0x8000D08600074DBF, 0x000515AA8000D08A, 0x000551AC000525AB,
-    0x000555AE00055DAD, 0x8000D08E000549AF, 0x8000D0968000D092,
-    0x8000D09A000595A3, 0x000531A6000531A5, 0x000669B38000D09E,
-    0x0000798100007982, 0x00005CBE00006580, 0x000669B5000661B6,
-    0x000629BB8000D0A2, 0x0000C886000621BA, 0x000619BF0000AC87,
-    0x000629BD000629BE, 0x0007A9A3000631BC, 0x8000D0AA8000D0A6,
-    0x8000D0B28000D0AE, 0x0007A9A50007A1A6, 0x0007E9AB8000D0B6,
-    0x8000D0BA0007E1AA, 0x0007D9AF8000D0BE, 0x0007E9AD0007E9AE,
-    0x8000D0C20007F1AC, 0x8000D0C6000635B3, 0x00062DB68000D0CA,
-    0x8000D0D28000D0CE, 0x0006C9BA000625B5, 0x8000D0D60006C9BB,
-    0x0006A5BE8000D0DA, 0x0006B1BC000699BF, 0x0006E1A80006B1BD,
-    0x0006F9AA0006EDA9, 0x0006E9AC0006F9AB, 0x000611AE0006E9AD,
-    0x8000D0DE000611AF, 0x8000D0E68000D0E2, 0x8000D0EE8000D0EA,
-    0x8000D0F68000D0F2, 0x0001D1B88000D0FA, 0x0001D1BA0001D1B9,
-    0x000151BC0001D1BB, 0x000179BE000151BD, 0x000649B0000179BF,
-    0x000621B2000649B1, 0x000625B4000621B3, 0x0001F1B6000619B5,
-    0x000571A30001F1B7, 0x8000D1028000D0FE, 0x8000D1060002CCBE,
-    0x000561A5000569A6, 0x00058DAB8000D10A, 0x0003E48400058DAA,
-    0x0005DDAF8000D10E, 0x0005F5AD0005E1AE, 0x000051800005F5AC,
-    0x0000658200005981, 0x8000D116000341A3, 0x000371A6000349A5,
-    0x000C40868000D11A, 0x000321AA00031487, 0x000325AC000321AB,
-    0x000311AE00032DAD, 0x000385A8000311AF, 0x000385AA00038DA9,
-    0x00038DAC00039DAB, 0x0003B5AE0003BDAD, 0x8000D11E0003DDAF,
-    0x8000D1268000D122, 0x8000D12E8000D12A, 0x8000D1368000D132,
-    0x000369B88000D13A, 0x000379BA000369B9, 0x000369BC000379BB,
-    0x0000DDBE000369BD, 0x0003A9B00000D5BF, 0x0003BDB20003A9B1,
-    0x000399B40003B5B3, 0x000359B6000399B5, 0x00020DB3000359B7,
-    0x8000D1428000D13E, 0x8000D14A8000D146, 0x000205B500023DB6,
-    0x00026DBB8000D14E, 0x8000D15200026DBA, 0x00025DBF8000D156,
-    0x000261BD00025DBE, 0x8000D15A000269BC, 0x8000D1628000D15E,
-    0x000368E38000D166, 0x0001A8E18000D16A, 0x000394EF8000D16E,
-    0x8000D1768000D172, 0x000019808000D17A, 0x0000058200001981,
-    0x0006C4E18000D17E, 0x0006D4E300073CE1, 0x000290840006BCE3,
-    0x000D14BE8000D186, 0x8000D18E8000D18A, 0x8000D1968000D192,
-    0x8000D19E8000D19A, 0x000190EF8000D1A2, 0x000C608600066CEF,
-    0x8000D1A6000DAC87, 0x8000D1AA00016DB3, 0x000175B600017DB5,
-    0x8000D1B28000D1AE, 0x000151BA8000D1B6, 0x0000F5BC000151BB,
-    0x0000F5BE0000FDBD, 0x000EA1A30000EDBF, 0x8000D1BA8000D182,
-    0x8000D1C28000D1BE, 0x000EB1A5000EB9A6, 0x000E9DAB8000D1C6,
-    0x8000D1CA000E9DAA, 0x000F21AF8000D1CE, 0x000F31AD000F39AE,
-    0x8000D1D2000F39AC, 0x8000D1DA8000D1D6, 0x8000D1E28000D1DE,
-    0x8000D1E6000DD4E3, 0x00001582000C24E1, 0x00003180000D04EF,
-    0x8000D1EA00001D81, 0x000330BE8000D1EE, 0x000ED1A800033084,
-    0x000E29AA000ED9A9, 0x000E39AC000E29AB, 0x000E29AE000E39AD,
-    0x8000D1F2000E29AF, 0x00000C860000F884, 0x8000D1F600008C87,
-    0x8000D1FE8000D1FA, 0x000FEDB88000D202, 0x000F8DBA000F85B9,
-    0x000F9DBC000F85BB, 0x000F85BE000F81BD, 0x000E59B0000FB9BF,
-    0x000FEDB2000E59B1, 0x000FFDB4000FE5B3, 0x000FE5B6000FE5B5,
-    0x000E95A8000FD5B7, 0x000EA9AA000E99A9, 0x000EB9AC000EA9AB,
-    0x000EA9AE000EB9AD, 0x8000D206000EA9AF, 0x8000D20E8000D20A,
-    0x8000D2168000D212, 0x8000D21E8000D21A, 0x00015DB88000D222,
-    0x000165BA00016DB9, 0x000115BC00010DBB, 0x000115BE00011DBD,
-    0x000EA9B000010DBF, 0x000EB5B2000EB5B1, 0x000E95B4000E8DB3,
-    0x000175B600017DB5, 0x000E21B300016DB7, 0x8000D22A8000D226,
-    0x8000D2328000D22E, 0x000E31B5000E39B6, 0x000E71BB8000D236,
-    0x8000D23A000E6DBA, 0x0001B9BF8000D23E, 0x000E59BD000E51BE,
-    0x8000D242000E61BC, 0x8000D246000E65A3, 0x000E7DA68000D24A,
-    0x8000D2528000D24E, 0x000E29AA000E75A5, 0x8000D256000E35AB,
-    0x000E15AE8000D25A, 0x000E25AC0001FDAF, 0x0002B5A9000E1DAD,
-    0x0002B5AB0002B5A8, 0x000335AD0002B5AA, 0x000335AF000335AC,
-    0x0003C980000335AE, 0x000005820003C981, 0x000264BE8000D25E,
-    0x000378878000D266, 0x0003E9B900041C86, 0x000389BB0003D9B8,
-    0x00039DBD0003E1BA, 0x000381BF00039DBC, 0x000355B1000389BE,
-    0x000355B3000355B0, 0x0003F9B5000355B2, 0x0003E9B70003F1B4,
-    0x000300840003F1B6, 0x0003A9B50003A1B6, 0x000229B30005C084,
-    0x8000D26E8000D26A, 0x000329BF8000D272, 0x000395BD000321BE,
-    0x000385BB000395BC, 0x8000D276000385BA, 0x0003E1A58000D27A,
-    0x8000D2828000D27E, 0x8000D2860003E9A6, 0x000261A38000D28A,
-    0x0003DDAD8000D28E, 0x000361AF0003DDAC, 0x8000D292000369AE,
-    0x0003CDAB8000D296, 0x8000D29A0003CDAA, 0x8000D2A28000D29E,
-    0x8000D2AA8000D2A6, 0x8000D2B28000D2AE, 0x8000D2B60003D4EF,
-    0x8000D2BA000194E1, 0x8000D2BE000198E3, 0x8000D2C68000D2C2,
-    0x000079808000D2CA, 0x0000198200007981, 0x8000D2CE0001D5B3,
-    0x0001E5B60001EDB5, 0x0004C086000214BE, 0x0001B9BA00054087,
-    0x0001ADBC0001B9BB, 0x00018DBE000185BD, 0x000E58E100007DBF,
-    0x0001E4E38000D2D6, 0x8000D2DE8000D2DA, 0x8000D2E68000D2E2,
-    0x0004E8BE8000D2EA, 0x8000D2F28000D2EE, 0x8000D2FA8000D2F6,
-    0x000138EF8000D2FE, 0x8000D3068000D302, 0x8000D30A00015DA3,
-    0x00016DA68000D30E, 0x8000D1128000D312, 0x000131AA000165A5,
-    0x8000D316000131AB, 0x000105AE8000D31A, 0x000125AC0000F5AF,
-    0x8000D2D200010DAD, 0x8000D3228000D31E, 0x8000D32A8000D326,
-    0x8000D3328000D32E, 0x000009818000D336, 0x8000D33A00003980,
-    0x8000D33E00001D82, 0x8000D3468000D342, 0x0000E0868000D34A,
-    0x8000D34E00037487, 0x0005F4EF8000D352, 0x8000D35A8000D356,
-    0x0004F0E38000D35E, 0x00044CE18000D362, 0x8000D36A8000D366,
-    0x8000D3728000D36E, 0x0006B5A88000D376, 0x000611AA000611A9,
-    0x000629AC000611AB, 0x00065DAE000629AD, 0x0000E484000655AF,
-    0x8000D37E8000D37A, 0x8000D3868000D382, 0x8000D38E8000D38A,
-    0x0007CDB88000D392, 0x0007DDBA0007D5B9, 0x0007F1BC0007D5BB,
-    0x0007E9BE0007FDBD, 0x000631B00007E9BF, 0x000635B2000631B1,
-    0x000605B400061DB3, 0x0007FDB60007F5B5, 0x000685A80007F5B7,
-    0x000699AA000689A9, 0x000689AC000699AB, 0x0006BDAE000689AD,
-    0x8000D3960006B5AF, 0x8000D39E8000D39A, 0x00000D808000D3A2,
-    0x0000B1820000B181, 0x0006A9B88000D3A6, 0x00014DBA0006A9B9,
-    0x00015DBC000145BB, 0x00014DBE000145BD, 0x0006CDB0000145BF,
-    0x0006DDB20006D5B1, 0x0006B1B40006D1B3, 0x0006B5B60006BDB5,
-    0x000611B30006A9B7, 0x8000D3AE8000D3AA, 0x8000D3B60001CCBE,
-    0x000601B5000609B6, 0x000611BB8000D3BA, 0x00000886000609BA,
-    0x000679BF00009C87, 0x000605BD000675BE, 0x8000D3BE000609BC,
-    0x8000D3C2000655A3, 0x00064DA68000D3C6, 0x8000D3CE8000D3CA,
-    0x00064DAA000645A5, 0x8000D3D2000655AB, 0x000631AE8000D3D6,
-    0x00064DAC00063DAF, 0x000239A8000641AD, 0x000241AA000239A9,
-    0x000245AC000241AB, 0x000245AE00024DAD, 0x001CACBE000279AF,
-    0x8000D3DA001DECBE, 0x8000D3E28000D3DE, 0x8000D3EA8000D3E6,
-    0x0002F1B88000D3EE, 0x0002F1BA0002F1B9, 0x000371BC0002F1BB,
-    0x000371BE000371BD, 0x00020DB0000371BF, 0x00021DB2000215B1,
-    0x00020DB4000215B3, 0x0002D1B60002D1B5, 0x8000D3F20002D1B7,
-    0x0002C1B38000D3F6, 0x0002C5B58000D3FA, 0x8000D4028000D3FE,
-    0x8000D4060002C5B6, 0x0002D9BB8000D40A, 0x0002C9BD0002C1BA,
-    0x00021DBF0002C1BC, 0x8000D40E0002C1BE, 0x8000D4168000D412,
-    0x8000D41A000285A3, 0x000281A6000281A5, 0x0003EC848000D41E,
-    0x000285AA8000D426, 0x000285AC00029DAB, 0x000285AE00028DAD,
-    0x000291A8000259AF, 0x0002C9AA000299A9, 0x0002DDAC0002C5AB,
-    0x0002C1AE0002C1AD, 0x001D00860002F5AF, 0x8000D42A001D6487,
-    0x0001F9808000D42E, 0x0001E5820001E581, 0x0001C5B88000D432,
-    0x0001DDBA0001C9B9, 0x0001F1BC0001D5BB, 0x000199BE0001F1BD,
-    0x00028DB0000199BF, 0x000145B200014DB1, 0x000145B400015DB3,
-    0x000145B600014DB5, 0x8000D4360001FDB7, 0x0000FCEF0006FCE1,
-    0x8000D43A000710E3, 0x8000D4428000D43E, 0x001C6C848000D446,
-    0x8000D44E8000D44A, 0x0001E8E18000D452, 0x000108E38000D456,
-    0x000201B30006E4EF, 0x8000D45E8000D45A, 0x8000D4668000D462,
-    0x000215B5000201B6, 0x0001C9BB8000D46A, 0x8000D46E000209BA,
-    0x0001C9BF8000D262, 0x0001C9BD0001C1BE, 0x001DC1A30001D1BC,
-    0x8000D4728000D422, 0x8000D47A8000D476, 0x001DD5A5001DC1A6,
-    0x001E09AB8000D47E, 0x8000D482001DC9AA, 0x001E09AF8000D486,
-    0x001E09AD001E01AE, 0x00006D80001E11AC, 0x00001D8200001581,
-    0x8000D48A001C94EF, 0x8000D4928000D48E, 0x000340868000D496,
-    0x8000D49A0003C487, 0x8000D4A28000D49E, 0x8000D4A6001FC4E1,
-    0x001E65A8001C8CE3, 0x001FA5AA001FADA9, 0x001F11AC001FB1AB,
-    0x001F0DAE001F05AD, 0x8000D4AA001F05AF, 0x8000D4B28000D4AE,
-    0x8000D4BA8000D4B6, 0x8000D4C28000D4BE, 0x001F11B88000D4C6,
-    0x001F21BA001F19B9, 0x0000DDBC0000C5BB, 0x0000C5BE0000C5BD,
-    0x001F7DB00000F5BF, 0x001F41B2001F41B1, 0x001F45B4001F41B3,
-    0x001F31B6001F31B5, 0x001E2DB3001F31B7, 0x8000D4CE8000D4CA,
-    0x8000D4D68000D4D2, 0x001E99B5001E85B6, 0x001E85BB8000D4DA,
-    0x8000D4DE001EB9BA, 0x001EF5BF8000D4E2, 0x001E95BD001E81BE,
-    0x8000D4E6001E95BC, 0x8000D4EA001E69A3, 0x001EC1A68000D4EE,
-    0x8000D4F68000D4F2, 0x001EFDAA001EDDA5, 0x8000D4FA001EC1AB,
-    0x001EC5AE8000D4FE, 0x001ED1AC001EB1AF, 0x0001DDA8001ED1AD,
-    0x0001E5AA0001EDA9, 0x0001E9AC0001FDAB, 0x00013DAE000135AD,
-    0x0000F980000135AF, 0x0000C5820000CD81, 0x000178BE00016084,
-    0x0000AC878000D506, 0x0000D1B800019C86, 0x0000E1BA0000D9B9,
-    0x000091BC0000E1BB, 0x000095BE00009DBD, 0x00014DB0000089BF,
-    0x00015DB2000155B1, 0x00014DB4000155B3, 0x0000F1B60000F1B5,
-    0x00026DB30000F1B7, 0x8000D50E8000D50A, 0x8000D5168000D512,
-    0x000261B50002B5B6, 0x000291BB8000D51A, 0x8000D51E000289BA,
-    0x000335BF8000D522, 0x000281BD000289BE, 0x00000090000289BC,
-    0x8000D526000229A3, 0x0002F1A68000D52A, 0x8000D5328000D52E,
-    0x0002CDAA000225A5, 0x8000D5360002D5AB, 0x0002CDAE8000D53A,
-    0x0002CDAC000371AF, 0x000335A80002C5AD, 0x000081AA000081A9,
-    0x000081AC000081AB, 0x0000B1AE000089AD, 0x8000D53E0000B1AF,
-    0x8000D5468000D542, 0x8000D54E8000D54A, 0x8000D5568000D552,
-    0x000069B88000D55A, 0x000079BA000069B9, 0x000069BC000079BB,
-    0x0001D9BE000069BD, 0x0000A9B00001D9BF, 0x0000BDB20000A9B1,
-    0x000099B40000B5B3, 0x000059B6000099B5, 0x8000D55E000059B7,
-    0x8000D5668000D562, 0x00001D818000D56A, 0x8000D56E00001D80,
-    0x0002AC8400000582, 0x0002C0870005E8BE, 0x8000D57600050C86,
-    0x8000D57E8000D57A, 0x000149B38000D582, 0x000000D28000D586,
-    0x8000D58E8000D58A, 0x000149B5000149B6, 0x00012DBB8000D592,
-    0x8000D59600012DBA, 0x00011DBF8000D59A, 0x000121BD00011DBE,
-    0x001EB8E1000129BC, 0x001E38E38000D59E, 0x8000D5A68000D5A2,
-    0x8000D5AE8000D5AA, 0x00046C848000D5B2, 0x8000D5B600046CBE,
-    0x8000D5BE8000D5BA, 0x001F64EF8000D5C2, 0x8000D5C600077084,
-    0x0004E8868000D5CA, 0x8000D5CE0004C087, 0x8000D5D2000068E3,
-    0x00001D820001A4E1, 0x00001D800000ACEF, 0x8000D3B200001581,
-    0x8000D5DA8000D5D6, 0x0007A4E18000D5DE, 0x000538E38000D572,
-    0x8000D5E28000D5E6, 0x8000D5EE8000D5EA, 0x8000D5F68000D5F2,
-    0x8000D5FE8000D5FA, 0x8000D6068000D602, 0x0005E4EF8000D60A,
-    0x000649A38000D60E, 0x8000D6168000D612, 0x8000D61E8000D61A,
-    0x000649A5000649A6, 0x00062DAB8000D622, 0x8000D62600062DAA,
-    0x00061DAF8000D62A, 0x000621AD00061DAE, 0x0007B1A8000629AC,
-    0x0007B1AA0007B1A9, 0x0007D5AC0007B1AB, 0x0007C9AE0007D9AD,
-    0x8000D62E0007C1AF, 0x8000D6368000D632, 0x8000D63E8000D63A,
-    0x8000D6468000D642, 0x000769B88000D64A, 0x000779BA000769B9,
-    0x000711BC000779BB, 0x000709BE000711BD, 0x000789B0000705BF,
-    0x00076DB2000789B1, 0x000761B4000761B3, 0x000761B6000761B5,
-    0x000639B3000761B7, 0x0001B58100001582, 0x8000D64E0001B580,
-    0x00065DB5000659B6, 0x00064DBB0001D4BE, 0x8000D65200064DBA,
-    0x0006A5BF8000D656, 0x0006B5BD0006A5BE, 0x8000D65A0006B9BC,
-    0x0000E88600067DA3, 0x00061DA60001CC87, 0x8000D6628000D65E,
-    0x000609AA000619A5, 0x8000D666000609AB, 0x0006E1AE8000D66A,
-    0x0006FDAC0006E1AF, 0x00018DA80006F1AD, 0x000191AA000191A9,
-    0x0001B1AC000191AB, 0x0001B1AE0001B1AD, 0x0001A8840001B1AF,
-    0x8000D6728000D66E, 0x8000D67A8000D676, 0x8000D6828000D67E,
-    0x0001A9B88000D686, 0x000169BA0001A9B9, 0x000179BC000169BB,
-    0x000169BE000179BD, 0x0001D1B0000169BF, 0x0001D5B20001D1B1,
-    0x0001B9B40001A9B3, 0x0001A9B60001B9B5, 0x000219B30001A1B7,
-    0x8000D68E8000D68A, 0x8000D6968000D692, 0x000239B5000231B6,
-    0x000215BB8000D69A, 0x8000D69E000215BA, 0x00027DBF8000D6A2,
-    0x00027DBD00027DBE, 0x8000D6A6000201BC, 0x8000D6AA00025DA3,
-    0x000275A68000D6AE, 0x8000D6B68000D6B2, 0x000251AA00027DA5,
-    0x8000D6BA000251AB, 0x000239AE8000D6BE, 0x000245AC000239AF,
-    0x0000D980000239AD, 0x0000E5820000ED81, 0x8000D6C60003C0BE,
-    0x000354870002E0EF, 0x8000D6CA000CBC86, 0x8000D6CE000114E3,
-    0x8000D6D20001FCE1, 0x8000D6DA8000D6D6, 0x8000D6E28000D6DE,
-    0x8000D6E6000391B3, 0x0003BDB6000C4084, 0x8000D6EE8000D6EA,
-    0x00034DBA0003B5B5, 0x8000D6F2000355BB, 0x00034DBE8000D6F6,
-    0x00034DBC0000F1BF, 0x00022DA8000345BD, 0x00025DAA000245A9,
-    0x000279AC000255AB, 0x0002BDAE000279AD, 0x000CACBE0002B5AF,
-    0x8000D6FE8000D6FA, 0x8000D7068000D702, 0x8000D70E8000D70A,
-    0x000151B88000D712, 0x000161BA000159B9, 0x000111BC000161BB,
-    0x000115BE00011DBD, 0x0002CDB0000109BF, 0x0002DDB20002D5B1,
-    0x0002CDB40002D5B3, 0x000171B6000171B5, 0x0006B4E1000171B7,
-    0x000638E38000D502, 0x8000D71A8000D716, 0x8000D71E000644E1,
-    0x8000D722000630E3, 0x8000D72A8000D726, 0x000F3CBE000048EF,
-    0x0006E4EF8000D72E, 0x000295A38000D736, 0x0000618100007182,
-    0x8000D73A00006980, 0x0002B1A50002B9A6, 0x000251AB8000D73E,
-    0x000CC886000249AA, 0x0001F5AF000D5C87, 0x000241AD000249AE,
-    0x8000D732000249AC, 0x8000D7468000D742, 0x000C94E18000D74A,
-    0x000CC4E38000D74E, 0x8000D7568000D752, 0x000DF0EF8000D75A,
-    0x8000D7628000D75E, 0x8000D76A8000D766, 0x000EE9B38000D76E,
-    0x8000D7768000D772, 0x8000D77E8000D77A, 0x000EEDB5000EEDB6,
-    0x000EB5BB8000D782, 0x8000D786000EADBA, 0x000F55BF8000D78A,
-    0x000E91BD000E99BE, 0x000E11A8000EA5BC, 0x000E21AA000E25A9,
-    0x000E25AC000E3DAB, 0x000E25AE000E2DAD, 0x8000D78E000E5DAF,
-    0x8000D7968000D792, 0x000019808000D79A, 0x0000058200001981,
-    0x000FF1B88000D79E, 0x000FF1BA000FF1B9, 0x000F5DBC000FF1BB,
-    0x000F41BE000F45BD, 0x000E29B0000F7DBF, 0x000E31B2000E31B1,
-    0x000E1DB4000E05B3, 0x000FD1B6000FD1B5, 0x000FA9A3000FD1B7,
-    0x0001188500028C84, 0x00017485000004BF, 0x000FADA5000FADA6,
-    0x000FF5AB8000D7A6, 0x00078886000FEDAA, 0x000E15AF00005887,
-    0x000FD1AD000FD9AE, 0x8000D7AA000FE5AC, 0x8000D7AE000E99B3,
-    0x00015DB68000D7B2, 0x8000D7BA8000D7B6, 0x000141BA000155B5,
-    0x8000D7BE000141BB, 0x000145BE8000D7C2, 0x000155BC00014DBF,
-    0x000E3DA800014DBD, 0x000E55AA000E55A9, 0x000199AC000189AB,
-    0x00018DAE000199AD, 0x8000D7C6000181AF, 0x8000D7CE8000D7CA,
-    0x8000D7D68000D7D2, 0x8000D7DE8000D7DA, 0x000189B88000D7E2,
-    0x000199BA000189B9, 0x0001B9BC000191BB, 0x000179BE0001B9BD,
-    0x0001C1B0000179BF, 0x0001C5B20001C1B1, 0x0001C5B40001DDB3,
-    0x0001C5B60001CDB5, 0x000DDDA30001B9B7, 0x8000D7EA8000D7E6,
-    0x8000D7F28000D7EE, 0x000211A5000219A6, 0x000205AB8000D7F6,
-    0x8000D7FA000205AA, 0x000209AF001A798F, 0x000209AD000201AE,
-    0x001E9D94000211AC, 0x00124D9600119D95, 0x001AD5900012ED97,
-    0x001E5992001ABD91, 0x000A659C001E5993, 0x000DED9E000A3D9D,
-    0x00159D98000E699F, 0x00164D9A00167599, 0x000184E100098D9B,
-    0x03FEDCE38000D7FE, 0x000E81A000038084, 0x00010DA2000FBDA1,
-    0x03E801B100029DA3, 0x03EA69B303F649B0, 0x03EE91B503EA8DB2,
-    0x03EC60EF03EC01B4, 0x03FED1A98000D802, 0x03F2CDAB03FEC1A8,
-    0x03F391AD03FFBDAA, 0x03F6F9AF03F2DDAC, 0x000401A103F401AE,
-    0x000679A38000D806, 0x03FAD1A50006CDA2, 0x03FC01A703F801A4,
-    0x001B558203FAC1A6, 0x8000D80A001B8183, 0x001F15868000D80E,
-    0x001BF184001FC987, 0x0010F98A001F1D85, 0x000524BE0013218B,
-    0x0017318E8000D816, 0x0013998C0017018F, 0x0008B1920014D18D,
-    0x00046886000B4193, 0x000F559600057C87, 0x000B0194000F4597,
-    0x000E499A000CD995, 0x8000D81A0000F19B, 0x000079818000D81E,
-    0x0003799C00019180, 0x8000D82200007182, 0x00058084000E74E1,
-    0x8000D826000098E3, 0x8000D82A8000D6C2, 0x8000D8328000D82E,
-    0x8000D83A8000D836, 0x8000D8428000D83E, 0x8000D84A8000D846,
-    0x000289B3000EF4EF, 0x8000D8528000D84E, 0x8000D85A8000D856,
-    0x00028DB500028DB6, 0x000245BB8000D85E, 0x8000D862000245BA,
-    0x000245BF8000D866, 0x000255BD000245BE, 0x00051DA8000255BC,
-    0x000541AA000545A9, 0x000545AC000541AB, 0x000571AE00054DAD,
-    0x8000D812000571AF, 0x8000D86E8000D86A, 0x00000D808000D872,
-    0x0000B1820000B181, 0x000645B88000D876, 0x000645BA00064DB9,
-    0x00064DBC00065DBB, 0x0006DDBE00067DBD, 0x0005C1B00006E1BF,
-    0x0005C1B20005C1B1, 0x0005C1B40005C1B3, 0x0005C1B60005C1B5,
-    0x0005C9A30005C1B7, 0x8000D87E8000D87A, 0x8000D882000300BE,
-    0x0005CDA50005CDA6, 0x000505AB8000D886, 0x00000886000505AA,
-    0x000505AF0003DC87, 0x000515AD000505AE, 0x8000D88A000515AC,
-    0x8000D88E0007A9B3, 0x000749B68000D892, 0x8000D89A8000D896,
-    0x000745BA00074DB5, 0x8000D89E00074DBB, 0x000745BE8000D8A2,
-    0x000755BC00074DBF, 0x00066DA300074DBD, 0x8000D8AA8000D8A6,
-    0x8000D8B28000D8AE, 0x000689A500068DA6, 0x000689AB8000D8B6,
-    0x8000D8BA000681AA, 0x000689AF8000D8BE, 0x000689AD000681AE,
-    0x8000D8C2000691AC, 0x8000D8CA8000D8C6, 0x8000D8D28000D8CE,
-    0x8000D8DA8000D8D6, 0x8000D8DE00048CEF, 0x8000D8E2000654E1,
-    0x00000D800005F0E3, 0x0000DD820000D581, 0x0006ADA88000D8E6,
-    0x0006E5AA0006EDA9, 0x000601AC00060DAB, 0x000605AE00060DAD,
-    0x8000D8EA00063DAF, 0x0001E4848000D8EE, 0x03FF10860001E8BE,
-    0x8000D8F600009487, 0x00060DB88000D8FA, 0x000615BA000615B9,
-    0x00063DBC000625BB, 0x0001F1BE0001F1BD, 0x000645B00001F1BF,
-    0x000645B200064DB1, 0x000645B400065DB3, 0x00063DB600064DB5,
-    0x00029DA8000635B7, 0x0002C1AA0002C5A9, 0x0002C5AC0002C1AB,
-    0x0002F1AE0002CDAD, 0x8000D8FE0002F1AF, 0x8000D9068000D902,
-    0x000328E18000D90A, 0x8000D9128000D90E, 0x000305B88000D916,
-    0x000305BA00030DB9, 0x000305BC00031DBB, 0x000339BE00030DBD,
-    0x000365B0000335BF, 0x000365B200036DB1, 0x00036DB400037DB3,
-    0x000355B600035DB5, 0x000249B300033DB7, 0x8000D91E8000D91A,
-    0x8000D9268000D922, 0x00024DB500024DB6, 0x0003A9BB8000D92A,
-    0x8000D92E0003A1BA, 0x000399BF8000D932, 0x0003A9BD0003A9BE,
-    0x8000D9360003B1BC, 0x000201A38000D93A, 0x000205A58000D93E,
-    0x8000D9468000D942, 0x8000D7A2000205A6, 0x0003E1AB8000D94A,
-    0x0003E1AD0003E9AA, 0x0003D1AF0003F9AC, 0x8000D94E0003E1AE,
-    0x8000D9568000D952, 0x8000D95E8000D95A, 0x000244BE0001A4E1,
-    0x000029800000A4E3, 0x0000398200002981, 0x8000D966000078EF,
-    0x8000D96A0002B884, 0x000161B38000D96E, 0x8000D9768000D972,
-    0x8000D97E8000D97A, 0x000171B5000179B6, 0x0001E5BB8000D982,
-    0x03FD4886000159BA, 0x0001EDBF03FD4087, 0x0001F5BD0001E5BE,
-    0x8000D9860001F5BC, 0x8000D98E8000D98A, 0x8000D99203FC4084,
-    0x8000D99A8000D996, 0x8000D99E0001E8EF, 0x03F3B4AC0006F4E1,
-    0x8000D9A2000634E3, 0x03FCB0848000D9A6, 0x8000D9AE8000D9AA,
-    0x0001E1A50001E9A6, 0x0001F1A38000D9B2, 0x8000D9BA8000D9B6,
-    0x00017DAF8000D9BE, 0x000165AD000175AE, 0x000175AB000165AC,
-    0x8000D9C20001C9AA, 0x03FDB9A88000D9C6, 0x03FE29AA03FDB9A9,
-    0x03FE39AC03FE29AB, 0x03FE2DAE03FE39AD, 0x8000D96203FE25AF,
-    0x0000198100001182, 0x8000D9CA03FFD180, 0x000334BE8000D9CE,
-    0x03FE81B88000D9D2, 0x03FE85BA03FE81B9, 0x03FE89BC03FE99BB,
-    0x03FF59BE03FEB1BD, 0x03FE5DB003FF59BF, 0x03FEE1B203FEE1B1,
-    0x03FEE1B403FEE1B3, 0x03FED1B603FEE9B5, 0x03FFB1A803FED1B7,
-    0x03FFB5AA03FFBDA9, 0x03FFD9AC03FFC9AB, 0x03FFC9AE03FFD9AD,
-    0x0000008603FFC1AF, 0x8000D9D600032887, 0x8000D9DE8000D9DA,
-    0x8000D9E68000D9E2, 0x03FF75B88000D9EA, 0x03FF09BA03FF79B9,
-    0x03FF05BC03FF01BB, 0x03FF05BE03FF0DBD, 0x03FF61B003FF25BF,
-    0x03FF61B203FF61B1, 0x03FF65B403FF7DB3, 0x03FF59B603FF6DB5,
-    0x03FE3DB303FF4DB7, 0x8000D9F28000D9EE, 0x8000D9FA8000D9F6,
-    0x03FE51B503FE59B6, 0x03FEA5BB8000D9FE, 0x8000DA0203FEA5BA,
-    0x03FEA9BF8000DA06, 0x03FEA9BD03FEA1BE, 0x8000DA0A03FEB5BC,
-    0x03FCA8EF8000DA0E, 0x8000DA168000DA12, 0x8000DA1E8000DA1A,
-    0x8000DA268000DA22, 0x8000DA2E8000DA2A, 0x03FED8E18000DA32,
-    0x03FC5CE38000DA36, 0x03FEBDA38000DA3A, 0x0000F58100001582,
-    0x8000DA3E0000F980, 0x03FED1A503FED9A6, 0x03FE25AB000190BE,
-    0x0000088603FE25AA, 0x03FE29AF00017087, 0x03FE29AD03FE21AE,
-    0x8000DA4603FE35AC, 0x8000DA4A000175B3, 0x00016DB68000DA4E,
-    0x8000DA568000DA52, 0x0000EDBA000165B5, 0x8000DA5A0000F5BB,
-    0x0000DDBE8000DA5E, 0x0000E5BC0000B5BF, 0x0002A9A80000D5BD,
-    0x0002B9AA0002A9A9, 0x0002A9AC0002B9AB, 0x00031DAE0002A9AD,
-    0x0004ACBE000315AF, 0x8000DA668000DA62, 0x8000DA6A0005A0BE,
-    0x8000DA728000DA6E, 0x000311B88000DA76, 0x000311BA000311B9,
-    0x000331BC000311BB, 0x000331BE000331BD, 0x00036DB0000331BF,
-    0x000375B2000375B1, 0x00035DB4000345B3, 0x00033DB6000335B5,
-    0x000231A3000331B7, 0x8000DA7E8000DA7A, 0x8000DA868000DA82,
-    0x000221A5000229A6, 0x0003B1AB8000DA8A, 0x8000DA8E0003A9AA,
-    0x0003F1AF8000DA92, 0x000391AD000399AE, 0x0003C0840003A1AC,
-    0x8000DA9A8000DA96, 0x8000DAA28000DA9E, 0x00003D8100003D82,
-    0x8000DAA600003D80, 0x8000DAAE8000DAAA, 0x8000DABA8000DAB6,
-    0x0004D8868000D8F2, 0x0002C1A80004BC87, 0x0002C5AA0002CDA9,
-    0x0002C9AC0002D9AB, 0x0002F5AE0002FDAD, 0x8000DABE000139AF,
-    0x8000DAC68000DAC2, 0x8000DACE8000DACA, 0x8000DAD68000DAD2,
-    0x0001D1B88000DADA, 0x0001E1BA0001D9B9, 0x000195BC0001E1BB,
-    0x000195BE00019DBD, 0x000149B000018DBF, 0x00015DB2000155B1,
-    0x00014DB4000155B3, 0x0001F1B60001F1B5, 0x0000E4EF0001F1B7,
-    0x8000DAE28000DADE, 0x8000DAE600050C84, 0x8000DAEE8000DAEA,
-    0x8000DAF200015CEF, 0x8000DAF603FED4E1, 0x03FE18E303FE30E3,
-    0x000100E18000DAFA, 0x8000DB028000DAFE, 0x8000DB0A8000DB06,
-    0x8000DB0E000201B3, 0x000231B600020DB5, 0x8000DB168000DB12,
-    0x0001C1BA8000DB1A, 0x0001C5BC0001DDBB, 0x0001C5BE0001CDBD,
-    0x0005CDA3000179BF, 0x8000DB1E8000DAB2, 0x8000DB228000DB26,
-    0x0005C1A50005FDA6, 0x000611AB8000DB2A, 0x0003088600060DAA,
-    0x0006B5AF00037C87, 0x000601AD000609AE, 0x0007A180000609AC,
-    0x000041820007A181, 0x8000DB2E00076DB3, 0x000775B600077DB5,
-    0x8000DB368000DB32, 0x000751BA8000DB3A, 0x0007F5BC000751BB,
-    0x0007E9BE0007FDBD, 0x00066DA80007E9BF, 0x0007A9AA000675A9,
-    0x0007B9AC0007A9AB, 0x0007A9AE0007B9AD, 0x8000DB3E0007A9AF,
-    0x8000DB468000DB42, 0x8000DB4E8000DB4A, 0x8000DB568000DB52,
-    0x000769B88000DB5A, 0x000775BA000771B9, 0x000715BC00070DBB,
-    0x000709BE00071DBD, 0x0007D9B0000701BF, 0x00076DB20007D9B1,
-    0x000761B4000765B3, 0x000761B6000761B5, 0x000621A300075DB7,
-    0x8000DB628000DB5E, 0x8000DB6A8000DB66, 0x000631A5000639A6,
-    0x00061DAB8000DB6E, 0x8000DB7200061DAA, 0x0006A5AF8000DB76,
-    0x0006B1AD0006A5AE, 0x8000DB7A0006B9AC, 0x8000DB7E00059CEF,
-    0x00000D808000DB82, 0x00003D8200003581, 0x8000DB8A8000DB86,
-    0x0001A4848000DB8E, 0x8000DB920001A4BE, 0x8000DB960005A4E3,
-    0x00018DA80006CCE1, 0x0001D9AA000199A9, 0x0001F1AC0001D1AB,
-    0x000139AE0001F1AD, 0x00000086000139AF, 0x8000DB9A00016487,
-    0x8000DBA28000DB9E, 0x8000DBAA8000DBA6, 0x0000CDB88000DBAE,
-    0x0000D5BA0000D5B9, 0x0000FDBC0000E5BB, 0x000091BE000095BD,
-    0x000149B0000091BF, 0x000159B2000149B1, 0x000149B4000159B3,
-    0x0000FDB6000149B5, 0x000251B30000F5B7, 0x8000DBB68000DBB2,
-    0x8000DBBE8000DBBA, 0x000279B5000271B6, 0x00028DBB8000DBC2,
-    0x8000DBC600028DBA, 0x00028DBF8000DBCA, 0x00028DBD00028DBE,
-    0x00126C8500028DBC, 0x8000DBCE000215A3, 0x000235A68000DBD2,
-    0x8000DBDA8000DBD6, 0x0002C9AA00023DA5, 0x8000DBDE0002C9AB,
-    0x0002C9AE8000DBE2, 0x0002C9AC0002C9AF, 0x000380840002C9AD,
-    0x8000DBEA8000DBE6, 0x8000DBEE000C00BE, 0x8000DBF68000DBF2,
-    0x8000DBFE8000DBFA, 0x8000DC0A8000DC06, 0x8000DC128000DC0E,
-    0x8000DC1A8000DC16, 0x000059808000DC1E, 0x0000698200005981,
-    0x8000DC220002B8EF, 0x000318878000DC26, 0x8000DC2A000CFC86,
-    0x8000DC328000DC2E, 0x8000DC3A8000DC36, 0x8000DA4200014CE1,
-    0x000279A800026CE3, 0x0002A5AA0002ADA9, 0x0002A5AC0002BDAB,
-    0x0002A5AE0002ADAD, 0x000DA884000119AF, 0x8000DC428000DC3E,
-    0x8000DC4A8000DC46, 0x8000DC528000DC4E, 0x000115B88000DC56,
-    0x000121BA00011DB9, 0x0001D5BC000121BB, 0x0001C9BE0001DDBD,
-    0x000169B00001C9BF, 0x000179B2000169B1, 0x000159B4000171B3,
-    0x000131B6000159B5, 0x000760E1000131B7, 0x000614E38000DC5A,
-    0x8000DC5E000C00BE, 0x8000DC620001A4E1, 0x8000DC660000E4E3,
-    0x8000DC6E8000DC6A, 0x8000DC76000038EF, 0x00076CEF8000DC7A,
-    0x000061808000DC7E, 0x0000618200006181, 0x8000DC82000261B3,
-    0x000261B6000261B5, 0x000CC0868000DC86, 0x0001E1BA000CE887,
-    0x0001E1BC0001E1BB, 0x0001D1BE0001E9BD, 0x000DADA30001D1BF,
-    0x8000DC728000DC02, 0x8000DC8E8000DC8A, 0x000DADA5000DADA6,
-    0x000E2DAB8000DC92, 0x8000DC96000E2DAA, 0x000E1DAF8000DC9A,
-    0x000E25AD000E1DAE, 0x8000DC9E000E2DAC, 0x8000DCA2000EE9B3,
-    0x000EE9B68000DCA6, 0x8000DCAE8000DCAA, 0x000EB5BA000EE9B5,
-    0x8000DCB2000E85BB, 0x000E9DBE8000DCB6, 0x000E9DBC000E85BF,
-    0x000E25A8000E95BD, 0x000E25AA000E2DA9, 0x000E25AC000E3DAB,
-    0x000E25AE000E2DAD, 0x8000DCBA000E4DAF, 0x8000DCC28000DCBE,
-    0x8000DCCA8000DCC6, 0x8000DCD28000DCCE, 0x000FEDB88000DCD6,
-    0x000FFDBA000FF5B9, 0x000FEDBC000FF5BB, 0x000F49BE000F51BD,
-    0x000E35B0000F45BF, 0x000E0DB2000E05B1, 0x000E1DB4000E05B3,
-    0x000FDDB6000E05B5, 0x000FADA3000FD5B7, 0x0000198100001982,
-    0x8000DCDA0001B180, 0x000FADA5000FADA6, 0x000FC1AB8000DCDE,
-    0x0002A084000FF1AA, 0x000FC1AF0001A0BE, 0x000FD1AD000FD9AE,
-    0x8000DCE2000FD9AC, 0x0000E886000EF9B3, 0x00015DB600012C87,
-    0x8000DCEA8000DCE6, 0x000149BA000E81B5, 0x8000DCEE000149BB,
-    0x00013DBE8000DCF2, 0x000139BC000121BF, 0x000E31A8000135BD,
-    0x000E35AA000E31A9, 0x000E4DAC000E5DAB, 0x000189AE000199AD,
-    0x0001E484000181AF, 0x8000DCFA8000DCF6, 0x8000DD028000DCFE,
-    0x8000DD0A8000DD06, 0x0001ADB88000DD0E, 0x00016DBA000165B9,
-    0x00017DBC000165BB, 0x00016DBE000165BD, 0x0001C9B0000165BF,
-    0x0001A9B20001C9B1, 0x0001BDB40001A5B3, 0x0001A1B60001A1B5,
-    0x000DBDA3000195B7, 0x8000DD168000DD12, 0x8000DD1E8000DD1A,
-    0x000DC5A5000219A6, 0x00020DAB8000DD22, 0x8000DD2600020DAA,
-    0x000265AF8000DD2A, 0x000271AD000279AE, 0x8000DD2E00027DAC,
-    0x8000DD368000DD32, 0x8000DD3E8000DD3A, 0x8000DD468000DD42,
-    0x000039808000DD4A, 0x0000058200003981, 0x00033CBE8000DD4E,
-    0x000344878000DD56, 0x8000DD5A0004DC86, 0x0003E1B38000DD5E,
-    0x000335B50004C084, 0x8000DD668000DD62, 0x8000DD6A000335B6,
-    0x000311BB8000DD6E, 0x0000B5BD000319BA, 0x0000B5BF000309BC,
-    0x8000DD720000B5BE, 0x8000DD76000194E1, 0x8000DD7A00017CE3,
-    0x8000DD828000DD7E, 0x8000DD8A8000DD86, 0x8000DD928000DD8E,
-    0x8000DD9A8000DD96, 0x8000DDA28000DD9E, 0x000255A8000274EF,
-    0x000261AA00025DA9, 0x0002B5AC000261AB, 0x0002B5AE0002BDAD,
-    0x000568BE0002ADAF, 0x8000DDAA8000DDA6, 0x8000DDB28000DDAE,
-    0x0000AD810000AD82, 0x000169B80000AD80, 0x000109BA000169B9,
-    0x000119BC000109BB, 0x000109BE000119BD, 0x0002D5B0000109BF,
-    0x0002D5B20002DDB1, 0x000179B4000169B3, 0x000169B6000179B5,
-    0x000F44E1000161B7, 0x000EB8E30000BCE3, 0x8000DDB6000E74E1,
-    0x8000DDBE8000DDBA, 0x8000DDCA8000DDC6, 0x8000DDD28000DDCE,
-    0x000730BE8000DDD6, 0x000ECCEF000060EF, 0x0002A9A38000DDDA,
-    0x0004E8868000DDDE, 0x8000DDE200050487, 0x00027DA500027DA6,
-    0x000259AB8000DDE6, 0x8000DDEA000251AA, 0x0001FDAF8000DDEE,
-    0x0001FDAD0001FDAE, 0x000675A8000241AC, 0x00068DAA000685A9,
-    0x000689AC000699AB, 0x0006BDAE0006B5AD, 0x8000DDC20006DDAF,
-    0x8000DDF68000DDF2, 0x8000DDFE8000DDFA, 0x8000DE068000DE02,
-    0x0006ADB88000DE0A, 0x0006BDBA0006B5B9, 0x000775BC0006B5BB,
-    0x000775BE00077DBD, 0x0006A9B000076DBF, 0x0006BDB20006B5B1,
-    0x0006ADB40006B5B3, 0x00069DB6000695B5, 0x000605B3000695B7,
-    0x8000DE128000DE0E, 0x8000DE1A8000DE16, 0x000601B5000635B6,
-    0x000671BB8000DE1E, 0x8000DE2200066DBA, 0x000651BF8000DE26,
-    0x000661BD000669BE, 0x00001982000669BC, 0x00006580000641A3,
-    0x000671A600001981, 0x8000DE2E8000DE2A, 0x000629AA000645A5,
-    0x00032484000635AB, 0x00062DAE8000DE32, 0x00062DAC000615AF,
-    0x0006A5B3000625AD, 0x000148868000DE36, 0x8000DE3A00010C87,
-    0x0006B5B50006A9B6, 0x0006C9BB8000DE3E, 0x8000DE420006F9BA,
-    0x00011DBF8000DE46, 0x00011DBD00011DBE, 0x8000DE4A0006D1BC,
-    0x8000DE528000DE4E, 0x0005CCE38000DE56, 0x00053CE18000DE5A,
-    0x0004B8EF8000DE5E, 0x8000DE668000DE62, 0x8000DE6E8000DE6A,
-    0x8000DE768000DE72, 0x00061DA88000DE7A, 0x00062DAA000625A9,
-    0x00063DAC000625AB, 0x00065DAE000629AD, 0x8000DE7E00064DAF,
-    0x8000DE868000DE82, 0x8000DE8E8000DE8A, 0x8000DE968000DE92,
-    0x0001F9B88000DE9A, 0x00014DBA0001F9B9, 0x00015DBC000145BB,
-    0x000145BE000145BD, 0x00063DB0000175BF, 0x000605B200060DB1,
-    0x0001D9B40001C9B3, 0x0001C9B60001D9B5, 0x0003B9800001C9B7,
-    0x000019820003B981, 0x8000DE9E0005A9A3, 0x0005A5A60005B9A5,
-    0x8000DEA600028484, 0x0005F5AA0003A8BE, 0x0005DDAC0005C5AB,
-    0x000211AE000211AD, 0x000301B6000211AF, 0x000344848000DEAA,
-    0x8000DEAE000361B5, 0x0003D887000349B3, 0x000309BE001CF886,
-    0x000335BC000309BF, 0x00033DBA000309BD, 0x8000DEB2000321BB,
-    0x0003CDB08000DEB6, 0x0003D5B20003D1B1, 0x0003F9B40003E9B3,
-    0x0003E9B60003F9B5, 0x0003D9B80003E9B7, 0x000309BA0003D9B9,
-    0x000319BC000309BB, 0x000309BE000319BD, 0x8000DEBA000309BF,
-    0x002169B58000DEBE, 0x8000DEC68000DEC2, 0x8000DECE8000DECA,
-    0x0003E5A88000DD52, 0x0003F5AA0003F5A9, 0x0003DDAC0003C5AB,
-    0x0003BDAE0003B5AD, 0x00021DA30003B5AF, 0x8000DED68000DED2,
-    0x001D1CBE8000DEDA, 0x000235A5000255A6, 0x000275AB8000DEDE,
-    0x8000DEE2000269AA, 0x00025DAF8000DEE6, 0x00025DAD00025DAE,
-    0x8000DEEA000261AC, 0x8000DEF28000DEEE, 0x00038CEF8000DEF6,
-    0x8000DEFE8000DEFA, 0x00036CE38000DF02, 0x000078E18000DF06,
-    0x8000DF0E8000DF0A, 0x001C58868000DF16, 0x0000E580001DF087,
-    0x0000F9820000E981, 0x8000DF1E8000DF1A, 0x8000DF268000DF22,
-    0x8000DF2A0006D4EF, 0x8000DF2E000674E1, 0x8000DF32000154E3,
-    0x8000DF3A8000DF36, 0x000191B38000DF3E, 0x8000DF468000DF42,
-    0x8000DF4A001DC484, 0x0001A5B5000115B6, 0x000135BB8000DF4E,
-    0x8000DF52000135BA, 0x0000FDBF8000DF56, 0x0000FDBD0000FDBE,
-    0x001E45A9000125BC, 0x001E71AB001E25A8, 0x001ED1AD001E5DAA,
-    0x001ED5AF001E69AC, 0x8000DF12001ED9AE, 0x8000DF5E8000DF5A,
-    0x8000DF668000DF62, 0x8000DF6E8000DF6A, 0x001F01B98000DF72,
-    0x001F01BB001F69B8, 0x001F05BD001F09BA, 0x001F05BF001F19BC,
-    0x001EB5B1001F05BE, 0x001F6DB3001EB5B0, 0x001F7DB5001EB5B2,
-    0x001F51B7001F7DB4, 0x001ED1A3001F61B6, 0x0000158100002D82,
-    0x8000DF7600001D80, 0x001EE5A5001E55A6, 0x001E75AB8000DF7A,
-    0x8000DF7E001E75AA, 0x001FBDAF000360BE, 0x001FBDAD001FBDAE,
-    0x8000DF82001E65AC, 0x00006886001F55B3, 0x001F71B600036C87,
-    0x8000DF8A8000DF86, 0x001F55BA001F45B5, 0x8000DF8E001F5DBB,
-    0x001F39BE8000DF92, 0x001F45BC001F25BF, 0x8000DF96001F39BD,
-    0x8000DF9E8000DF9A, 0x8000DFA68000DFA2, 0x8000DFAE8000DFAA,
-    0x8000DFB2001CF4EF, 0x8000DFB6001D60E1, 0x8000DFBA001D74E3,
-    0x8000DFC28000DFBE, 0x001ED1A38000DFC6, 0x8000DFCE8000DFCA,
-    0x8000DFD68000DFD2, 0x001EC1A5001EF5A6, 0x001ED9AB8000DFDA,
-    0x8000DFDE001ED1AA, 0x001EA1AF8000DFE2, 0x001EBDAD001EBDAE,
-    0x001EF5A8001EC1AC, 0x001E19AA001E09A9, 0x001E09AC001E19AB,
-    0x001E31AE001E35AD, 0x00000D80001E2DAF, 0x00001D8200001581,
-    0x8000DFEA8000DFE6, 0x0001FC848000DFEE, 0x001E2DB80001F0BE,
-    0x0001E1BA0001E1B9, 0x0001E1BC0001E1BB, 0x0001E1BE0001E1BD,
-    0x001E55B00001E1BF, 0x001E55B2001E5DB1, 0x001E39B4001E29B3,
-    0x001E25B6001E25B5, 0x0002D5A8001E15B7, 0x0002E1AA0002DDA9,
-    0x000311AC0002E1AB, 0x000311AE000311AD, 0x8000DFF6000311AF,
-    0x8000DFFE8000DFFA, 0x8000E0068000E002, 0x8000E00E8000E00A,
-    0x00037DB88000E012, 0x000301BA000301B9, 0x000301BC000301BB,
-    0x000331BE000309BD, 0x000371B0000331BF, 0x000371B2000371B1,
-    0x000359B400034DB3, 0x00034DB6000345B5, 0x00042086000345B7,
-    0x000261B30003A087, 0x000391B5000480BE, 0x8000E01A8000E016,
-    0x8000DEA2000399B6, 0x0003BDBB0003B884, 0x000391BD0003BDBA,
-    0x000389BF0003ADBC, 0x000229A3000395BE, 0x8000E0228000E01E,
-    0x000694858000E026, 0x0003D9A50003D1A6, 0x0003F5AB8000E02A,
-    0x8000E02E0003F5AA, 0x0003C1AF8000E032, 0x0003D9AD0003DDAE,
-    0x8000E0360003E5AC, 0x8000E03E8000E03A, 0x8000E0468000E042,
-    0x8000E04E8000E04A, 0x000011818000E052, 0x8000E05600000D80,
-    0x8000E05A00001982, 0x8000E0668000E05E, 0x8000E06E8000E06A,
-    0x8000E072000194E1, 0x8000E0760000E4E3, 0x000478868000E07A,
-    0x8000E07E00028087, 0x8000E0868000E082, 0x8000E08E8000E08A,
-    0x8000E0968000E092, 0x8000E09A000038EF, 0x8000E0A28000E09E,
-    0x8000E0AA8000E0A6, 0x8000E0B28000E0AE, 0x00046CBE0001C8EF,
-    0x0005A084001EC4E1, 0x8000E0B6001E70E3, 0x8000E0BE8000E0BA,
-    0x8000E0C68000E0C2, 0x0001F9B50001F1B6, 0x0001F9B30020F485,
-    0x0000218100002982, 0x00013DBF00001D80, 0x000121BD000129BE,
-    0x000169BB000129BC, 0x8000E0CA000161BA, 0x000621A88000E0CE,
-    0x000621AA000621A9, 0x000621AC000621AB, 0x000621AE000621AD,
-    0x8000E062000621AF, 0x000C008600012CBF, 0x8000E0D200036087,
-    0x8000E0DA8000E0D6, 0x000741B88000E0DE, 0x000741BA000741B9,
-    0x000741BC000741BB, 0x000741BE000741BD, 0x0006F5B0000741BF,
-    0x0006ADB20006FDB1, 0x0006BDB40006A5B3, 0x000691B60006A9B5,
-    0x0006A9A3000691B7, 0x8000E0E68000E0E2, 0x8000E0EE8000E0EA,
-    0x0006A9A50006A1A6, 0x000639AB8000E0F2, 0x8000E0F6000631AA,
-    0x00066DAF8000E0FA, 0x000671AD000679AE, 0x8000E0FE000679AC,
-    0x8000E102000785B3, 0x000785B68000E106, 0x8000E10E8000E10A,
-    0x000789BA00078DB5, 0x8000E112000789BB, 0x000789BE8000E116,
-    0x000789BC000775BF, 0x8000E11A000789BD, 0x00057CEF0005C0E1,
-    0x8000E11E0004D4E3, 0x8000E1268000E122, 0x8000E12E8000E12A,
-    0x8000E1368000E132, 0x0005A8E18000E13A, 0x00058CE38000E13E,
-    0x000601A3000CF8EF, 0x0000658100007182, 0x0001808400006580,
-    0x000609A5000601A6, 0x00060DAB000010BE, 0x000FE88600060DAA,
-    0x0006F1AF00001887, 0x00060DAD00060DAE, 0x0006B9A800060DAC,
-    0x0006C9AA0006B9A9, 0x0006C1AC0006C5AB, 0x0006F9AE0006C1AD,
-    0x8000E1460006F9AF, 0x8000E14E8000E14A, 0x8000E1568000E152,
-    0x8000E15E8000E15A, 0x00064DB88000E162, 0x00065DBA000655B9,
-    0x000661BC000655BB, 0x000661BE000661BD, 0x000689B0000661BF,
-    0x000699B2000689B1, 0x000689B4000699B3, 0x00067DB6000675B5,
-    0x8000E166000675B7, 0x000285B600028DB5, 0x03E8F18F000C4084,
-    0x8000E16E8000E16A, 0x0002E1BC000221B3, 0x0002E5BE0002EDBD,
-    0x0032C8AC0002F9BF, 0x000291BA8000E172, 0x001F019B000291BB,
-    0x001E4D99001E4D9A, 0x0004F19F00133D98, 0x001A799D001AB99E,
-    0x001641930018819C, 0x03EAF99100149192, 0x0012619703EA4D90,
-    0x0010819500127196, 0x003801B100164194, 0x003A81B3003EE1B0,
-    0x002691B5003A89B2, 0x8000E176003BD1B4, 0x000281A38000E17A,
-    0x000669A1000011A2, 0x8000E17E00068DA0, 0x8000E1828000DFF2,
-    0x0002F1A18000E186, 0x000ECDA38000E18A, 0x000801A5000ECDA2,
-    0x000BC5A7000F51A4, 0x0036C1A9000AFDA6, 0x003001AB003401A8,
-    0x0032CDAD003641AA, 0x003EFDAF0032D5AC, 0x03E9E182003C01AE,
-    0x8000E18E03EB2583, 0x001731868000E192, 0x03EB898400172987,
-    0x0011D98A0015E985, 0x8000E1960013218B, 0x001F358E8000E19A,
-    0x0013298C001F3D8F, 0x001E019200124D8D, 0x8000E19E00199993,
-    0x0005F196000C0084, 0x001B799400077597, 0x0007A19A001BB195,
-    0x8000E1A20001D19B, 0x0000F1818000E1A6, 0x0003599C00001980,
-    0x8000E1AA0000F982, 0x000DC0BE8000E1AE, 0x000C3086000028EF,
-    0x8000E1B6000D5C87, 0x8000E1BA0006C4EF, 0x8000E1BE0006E4E1,
-    0x8000E1C2000614E3, 0x8000E1C6000184E1, 0x0003FDA30000FCE3,
-    0x8000E1CE8000E1CA, 0x8000E1D2000F4C84, 0x000351A5000359A6,
-    0x00034DAB8000E1D6, 0x8000E1DA00034DAA, 0x000325AF8000E1DE,
-    0x000331AD000339AE, 0x8000E14200033DAC, 0x8000E1E28000E1B2,
-    0x8000E1EA8000E1E6, 0x8000E1F28000E1EE, 0x000D81A88000E1F6,
-    0x000D81AA000D81A9, 0x000D81AC000D81AB, 0x000D81AE000D81AD,
-    0x000E41B0000D81AF, 0x000E51B2000E51B1, 0x000EC9B4000E79B3,
-    0x000EE9B6000EC9B5, 0x000EDDB8000EE5B7, 0x000F61BA000EA1B9,
-    0x000F59BC000F61BB, 0x000F31BE000F59BD, 0x000D29B3000F15BF,
-    0x8000E1FE8000E1FA, 0x8000E2068000E202, 0x000D29B5000D21B6,
-    0x000EB5BB8000E20A, 0x8000E20E000EA5BA, 0x000E2DBF8000E212,
-    0x000E8DBD000E2DBE, 0x00002982000EA5BC, 0x00001D80000D6DA3,
-    0x000D65A600001581, 0x8000E21A8000E216, 0x000EE1AA000D6DA5,
-    0x8000E21E000EF1AB, 0x000E69AE0003A4BE, 0x000EE1AC000E69AF,
-    0x000E99A8000EC9AD, 0x000EA1AA000EA5A9, 0x000EA9AC000EBDAB,
-    0x000ED1AE000ED1AD, 0x00008086000ECDAF, 0x8000E22600010487,
-    0x8000E22E8000E22A, 0x8000E2368000E232, 0x00015DB88000E23A,
-    0x0001C1BA0001C1B9, 0x0001C1BC0001C1BB, 0x0001F1BE0001C9BD,
-    0x000EBDB00001F1BF, 0x000E85B2000E8DB1, 0x00017DB4000165B3,
-    0x00016DB6000165B5, 0x000E29B3000165B7, 0x8000E23E8000E222,
-    0x8000E2468000E242, 0x000E21B5000E51B6, 0x000E49BB8000E24A,
-    0x8000E24E000E79BA, 0x0001A1BF8000E252, 0x0001B1BD0001B9BE,
-    0x8000E256000E51BC, 0x8000E25A000E6DA3, 0x000E15A68000E25E,
-    0x8000E2668000E262, 0x000E3DAA000E65A5, 0x8000E26A000E0DAB,
-    0x0001FDAE8000E26E, 0x000E15AC0001E5AF, 0x000138E10001F5AD,
-    0x000FD4E38000E272, 0x8000E27A8000E276, 0x8000E27E000358BE,
-    0x00000D808000E282, 0x00001D8200001581, 0x8000E28A8000E286,
-    0x0002D4EF8000E28E, 0x0004E0868000E296, 0x000400BE0003E087,
-    0x8000E29A00037DB3, 0x000365B600036DB5, 0x8000E2A28000E29E,
-    0x0003F9BA8000E2A6, 0x0003D9BC0003F1BB, 0x0003B9BE0003D9BD,
-    0x000EBDAA0003B5BF, 0x000EBDA8000E01AB, 0x000E01AE000EB5A9,
-    0x000E11AC000E09AF, 0x8000E2AA000E11AD, 0x8000E2B28000E2AE,
-    0x8000E2BA8000E2B6, 0x8000E2C28000E2BE, 0x000FD1BA8000E2C6,
-    0x000E25B8000FD1BB, 0x000FF1BE000FCDB9, 0x000FC1BC000FF1BF,
-    0x000E21B2000FC1BD, 0x000E41B0000E2DB3, 0x000E29B6000E41B1,
-    0x000E35B4000E1DB7, 0x000229A8000E29B5, 0x000239AA000229A9,
-    0x000229AC000239AB, 0x000291AE000229AD, 0x8000E2CA000291AF,
-    0x8000E2D28000E2CE, 0x8000E2DA8000E2D6, 0x002B30BF8000E2DE,
-    0x000299B88000E2E2, 0x0002A5BA0002ADB9, 0x00015DBC000149BB,
-    0x000141BE000145BD, 0x0002F1B0000141BF, 0x0002C1B20002F9B1,
-    0x0002B1B40002C1B3, 0x0002B5B60002BDB5, 0x000E08E10002A9B7,
-    0x000ED4E3000CF4E1, 0x8000E2E6000F90E3, 0x8000E2EE8000E2EA,
-    0x000011818000E2F2, 0x8000E2F600000D80, 0x8000E2FA00001982,
-    0x0001ACEF8000E2FE, 0x0002F1A3000ECCEF, 0x8000E30A8000E306,
-    0x8000E30E00058084, 0x0002E1A50002E9A6, 0x00027DAB8000E312,
-    0x0004C886000275AA, 0x000239AF0004EC87, 0x000255AD000235AE,
-    0x00063DA8000255AC, 0x000655AA000655A9, 0x000675AC00066DAB,
-    0x0006BDAE0006B5AD, 0x8000E3020006B5AF, 0x8000E31A8000E316,
-    0x8000E3228000E31E, 0x8000E32A8000E326, 0x00076DB88000E32E,
-    0x00077DBA000775B9, 0x0007C1BC000771BB, 0x0007C1BE0007C1BD,
-    0x0006CDB00007C1BF, 0x0006DDB20006D5B1, 0x0006CDB40006D5B3,
-    0x00075DB6000755B5, 0x0006DDB3000755B7, 0x8000E3368000E332,
-    0x8000E33E8000E33A, 0x0006FDB5000635B6, 0x000611BB8000E342,
-    0x8000E346000609BA, 0x000791BF8000E34A, 0x000601BD000609BE,
-    0x8000E34E000609BC, 0x8000E352000699A3, 0x000671A68000E356,
-    0x8000E35E8000E35A, 0x00064DAA0006B9A5, 0x8000E362000655AB,
-    0x00064DAE8000E366, 0x00064DAC0007D5AF, 0x0001ED80000645AD,
-    0x0000198200000981, 0x8000E36E8000E36A, 0x0001C4BE0001D084,
-    0x000080868000E372, 0x8000E37600012087, 0x8000E37E8000E37A,
-    0x8000E3868000E382, 0x8000E38E8000E38A, 0x8000E3968000E392,
-    0x8000E39E8000E39A, 0x8000E3A200050CE1, 0x8000E3A6000480E3,
-    0x8000E3AE8000E3AA, 0x8000E3B20005DCEF, 0x8000E3BA8000E3B6,
-    0x000645A88000E3BE, 0x000659AA00064DA9, 0x0001B5AC000651AB,
-    0x0001B5AE0001BDAD, 0x0001AC840001ADAF, 0x8000E3C68000E3C2,
-    0x8000E3CE8000E3CA, 0x8000E3D68000E3D2, 0x000169B88000E3DA,
-    0x000109BA000169B9, 0x00011DBC000105BB, 0x000101BE000101BD,
-    0x0001D5B0000135BF, 0x0001D5B20001DDB1, 0x000179B4000169B3,
-    0x000169B6000179B5, 0x0003BD81000161B7, 0x0005D9B30003BD80,
-    0x000235B500001982, 0x000344BE8000E3DE, 0x8000E3E6000235B6,
-    0x000211BB8000E3EA, 0x0002B5BD000219BA, 0x0002B5BF000209BC,
-    0x8000E3EE0002B5BE, 0x00026DA500026DA6, 0x000581A3003C0084,
-    0x003CF8868000E3F2, 0x0002EDAF00037C87, 0x0002EDAD0002EDAE,
-    0x000249AB000251AC, 0x8000E3F6000241AA, 0x8000E3FE8000E3FA,
-    0x8000E4068000E402, 0x000120E38000E40A, 0x0001A0E18000E40E,
-    0x000240EF8000E412, 0x8000E41A8000E416, 0x8000E4228000E41E,
-    0x8000E42A8000E426, 0x0003F1B38000E42E, 0x8000E4368000E432,
-    0x8000E2928000E43A, 0x0003E1B50003E9B6, 0x0003ADBB8000E43E,
-    0x8000E4420003ADBA, 0x000395BF8000E446, 0x0003A1BD000395BE,
-    0x000229A80003ADBC, 0x000239AA000229A9, 0x000229AC000239AB,
-    0x000259AE000229AD, 0x0001CD80000259AF, 0x0000198200000981,
-    0x8000E44E8000E44A, 0x003DB487003D7CBE, 0x000235B8003C1C86,
-    0x000235BA00023DB9, 0x0002F9BC0002E9BB, 0x0002E9BE0002F9BD,
-    0x000229B00002E9BF, 0x000239B2000229B1, 0x000211B4000235B3,
-    0x000211B6000211B5, 0x8000E45600020DB7, 0x8000E45E8000E45A,
-    0x8000E4668000E462, 0x8000E46E8000E46A, 0x8000E47200007CEF,
-    0x8000E476000654E1, 0x0032CCAC0000A8E3, 0x8000E47E8000E47A,
-    0x8000E4868000E482, 0x0002F1A50002F9A6, 0x0002E1A38000E48A,
-    0x8000E4928000E48E, 0x000285AF8000E496, 0x0002B1AD000285AE,
-    0x0002BDAB0002BDAC, 0x8000E49A0002BDAA, 0x003E21A88000E49E,
-    0x003E49AA003E21A9, 0x003E69AC003E59AB, 0x003E99AE003E69AD,
-    0x8000E452003E99AF, 0x8000E4A68000E4A2, 0x8000E4AE8000E4AA,
-    0x8000E4B68000E4B2, 0x003EF5B88000E4BA, 0x003EF5BA003EFDB9,
-    0x003E99BC003E89BB, 0x003E85BE003E99BD, 0x003EE9B0003EB5BF,
-    0x003EFDB2003EE9B1, 0x003ED5B4003EF1B3, 0x003ED5B6003EDDB5,
-    0x003EC1B3003ECDB7, 0x0000B58100001582, 0x8000E4BE0000B580,
-    0x003EE1B5003EE9B6, 0x003E2DBB000010BE, 0x00036484003E2DBA,
-    0x003E19BF8000E4C2, 0x003E25BD003E11BE, 0x8000E4C6003E29BC,
-    0x0000A886003E85A3, 0x003EADA600036C87, 0x8000E4CE8000E4CA,
-    0x003E69AA003EA5A5, 0x8000E4D2003E69AB, 0x003E55AE8000E4D6,
-    0x003E6DAC003E5DAF, 0x003E89B3003E61AD, 0x8000E4DE8000E4DA,
-    0x8000E4E68000E4E2, 0x003E89B5003E81B6, 0x003E95BB8000E4EA,
-    0x8000E4EE003E8DBA, 0x003EFDBF8000E4F2, 0x003E85BD003EF5BE,
-    0x8000E4F6003E85BC, 0x8000E4FE8000E4FA, 0x8000E5068000E502,
-    0x003C94EF8000E50A, 0x003D20E18000E50E, 0x003D34E38000E512,
-    0x8000E51A8000E516, 0x8000E5228000E51E, 0x003E01A88000E526,
-    0x003E01AA003E01A9, 0x003E01AC003E01AB, 0x003E01AE003E01AD,
-    0x0000DD80003E3DAF, 0x0000198200000981, 0x8000E52E8000E52A,
-    0x000074BE00013C84, 0x003E1DB88000E536, 0x003E35BA003E3DB9,
-    0x0001F9BC0001E9BB, 0x0001E9BE0001F9BD, 0x003E49B00001E9BF,
-    0x003E51B2003E51B1, 0x003E31B4003E51B3, 0x003E2DB6003E39B5,
-    0x003D89A3003E25B7, 0x0004A8868000E53A, 0x8000E53E00033C87,
-    0x003D89A5003D81A6, 0x003D95AB8000E542, 0x8000E546003D8DAA,
-    0x003DFDAF8000E54A, 0x003D85AD003DF5AE, 0x8000E54E003D85AC,
-    0x0003488400029DB3, 0x000355B6000440BE, 0x8000E5568000E552,
-    0x000371BA0002B1B5, 0x8000E55A000379BB, 0x000331BE8000E55E,
-    0x000351BC000331BF, 0x0003D1B8000351BD, 0x0003D5BA0003D1B9,
-    0x0003FDBC0003E9BB, 0x0003EDBE0003E5BD, 0x0003B1B00003E5BF,
-    0x000381B20003B9B1, 0x0003F1B4000381B3, 0x0003F1B60003F1B5,
-    0x000201A80003F1B7, 0x000211AA000211A9, 0x00023DAC000225AB,
-    0x0003D1AE0003D1AD, 0x8000E5620003D1AF, 0x8000E5668000E3E2,
-    0x8000E56E8000E56A, 0x8000E5768000E572, 0x8000E57E8000E57A,
-    0x8000E5868000E582, 0x000050EF8000E58A, 0x8000E5928000E58E,
-    0x000084E38000E596, 0x0001F8E18000E59A, 0x00002D808000E59E,
-    0x0000158200001181, 0x000399A38000E5A6, 0x8000E5AE8000E5AA,
-    0x8000E5B68000E5B2, 0x0003B5A5000251A6, 0x00027DAB8000E5BA,
-    0x8000E5BE000275AA, 0x000235AF8000E5C2, 0x000255AD000235AE,
-    0x0004A086000255AC, 0x8000E5C60005A887, 0x8000E5CE8000E5CA,
-    0x8000E5D68000E5D2, 0x00042CBE0001F8EF, 0x8000E5DA003E80E1,
-    0x8000E5DE000114E3, 0x8000E5E68000E5E2, 0x8000E5EE8000E5EA,
-    0x8000E5F68000E5F2, 0x8000E5FA0001DDB3, 0x0001F5B60001FDB5,
-    0x8000E6028000E5FE, 0x0001B1BA8000E606, 0x00019DBC000185BB,
-    0x000111BE000115BD, 0x0005BDA8000111BF, 0x000619AA000601A9,
-    0x000639AC000611AB, 0x000659AE000639AD, 0x8000E5A2000649AF,
-    0x0000198100001182, 0x8000E60A0007C980, 0x000230848000E60E,
-    0x0006B9B88000E612, 0x000749BA0006B9B9, 0x000759BC000749BB,
-    0x00074DBE000759BD, 0x000639B0000745BF, 0x000605B200060DB1,
-    0x000699B400068DB3, 0x000689B6000699B5, 0x000691A3000689B7,
-    0x000028868000E616, 0x8000E61A0003FC87, 0x0006B1A50006B9A6,
-    0x0006C9AB8000E61E, 0x8000E6220006FDAA, 0x00065DAF8000E626,
-    0x000659AD00065DAE, 0x8000E62A0006D1AC, 0x8000E62E00078DB3,
-    0x00078DB68000E632, 0x8000E63A8000E636, 0x000769BA00078DB5,
-    0x8000E63E000769BB, 0x000769BE8000E642, 0x000769BC00075DBF,
-    0x000649A3000769BD, 0x8000E64A8000E646, 0x8000E6528000E64E,
-    0x000649A5000649A6, 0x0006ADAB8000E656, 0x8000E65A0006ADAA,
-    0x000699AF8000E65E, 0x0006ADAD0006ADAE, 0x8000E6620006ADAC,
-    0x8000E66A8000E666, 0x8000E6728000E66E, 0x8000E6760005E0EF,
-    0x8000E67E8000E67A, 0x0001248400053CE3, 0x0000298000057CE1,
-    0x0000198200002981, 0x0006BDA8000134BE, 0x0006C1AA0006C1A9,
-    0x0006C1AC0006C1AB, 0x0006F1AE0006C9AD, 0x000000860006F1AF,
-    0x8000E68600008487, 0x8000E68E8000E68A, 0x8000E6968000E692,
-    0x000679B88000E69A, 0x000609BA000679B9, 0x000619BC000609BB,
-    0x000605BE000605BD, 0x000669B000063DBF, 0x000679B2000669B1,
-    0x000669B4000679B3, 0x000649B600065DB5, 0x000285A8000649B7,
-    0x000295AA000295A9, 0x0002BDAC0002A5AB, 0x0002D1AE0002D5AD,
-    0x8000E69E0002D1AF, 0x000328848000E6A2, 0x8000E6A6000989B5,
-    0x8000E6AE8000E6AA, 0x00035DB88000E6B2, 0x000365BA00036DB9,
-    0x000315BC00030DBB, 0x000315BE00031DBD, 0x0002B5B000030DBF,
-    0x000281B20002BDB1, 0x000375B4000281B3, 0x000375B600037DB5,
-    0x00021DB300036DB7, 0x8000E6BA8000E6B6, 0x000C1CBE8000E6BE,
-    0x000235B5000255B6, 0x000279BB8000E6C2, 0x8000E6C6000271BA,
-    0x0003B9BF8000E6CA, 0x000255BD0003B1BE, 0x8000E6CE000255BC,
-    0x8000E6D2000259A3, 0x000211A68000E6D6, 0x8000E6DE8000E6DA,
-    0x000235AA000271A5, 0x8000E53200023DAB, 0x0003F5AE8000E6E2,
-    0x000211AC0003FDAF, 0x0002A9A8000211AD, 0x0002B9AA0002A9A9,
-    0x0002A9AC0002B9AB, 0x000139AE0002A9AD, 0x00000D80000139AF,
-    0x00001D8200001581, 0x8000E6EA8000E6E6, 0x000D68878000E6EE,
-    0x0001E9B8000D1C86, 0x000189BA0001E9B9, 0x00019DBC000185BB,
-    0x000181BE000181BD, 0x000151B00001B1BF, 0x000151B2000151B1,
-    0x0001F5B4000151B3, 0x0001F5B60001FDB5, 0x000004EF0001E1B7,
-    0x8000E6F6000C4884, 0x8000E6FA000C80BE, 0x8000E7028000E6FE,
-    0x8000E7060006C4EF, 0x8000E70A00063CE1, 0x0000C8E3000184E3,
-    0x000128E18000E70E, 0x8000E7168000E712, 0x8000E71E8000E71A,
-    0x8000E72200027DB3, 0x000265B600026DB5, 0x8000E72A8000E726,
-    0x0001D9BA8000E72E, 0x0001C9BC0001D9BB, 0x00017DBE0001C9BD,
-    0x000DB1A3000175BF, 0x8000E7328000E6F2, 0x8000E73A8000E736,
-    0x000DA1A5000DA9A6, 0x000E15AB8000E73E, 0x8000E742000E15AA,
-    0x000EB9AF8000E746, 0x000E05AD000EB1AE, 0x8000E74A000E05AC,
-    0x8000E7528000E74E, 0x00001D808000E756, 0x0000398200000981,
-    0x8000E75E8000E75A, 0x0000C8868000E762, 0x8000E76600030C87,
-    0x8000E76E8000E76A, 0x000F09A88000E772, 0x000F1DAA000F15A9,
-    0x000F31AC000F15AB, 0x000F2DAE000F3DAD, 0x0003A884000F21AF,
-    0x8000E77A8000E776, 0x8000E7828000E77E, 0x8000E78A8000E786,
-    0x0000CDB88000E78E, 0x0000D5BA0000D5B9, 0x0000FDBC0000E5BB,
-    0x00009DBE000095BD, 0x000F21B0000095BF, 0x000F21B2000F21B1,
-    0x000F2DB4000F3DB3, 0x000F1DB6000F15B5, 0x000E95B30000F5B7,
-    0x8000E7968000E792, 0x8000E79E8000E79A, 0x000EB9B5000EBDB6,
-    0x000EE5BB8000E7A2, 0x8000E7A6000EE5BA, 0x000ED1BF8000E7AA,
-    0x000EE9BD000EE9BE, 0x000D28EF000EF5BC, 0x8000E7B28000E7AE,
-    0x8000E7BA8000E7B6, 0x8000E7C28000E7BE, 0x00003D808000E7C6,
-    0x00001D8200000981, 0x000D48E38000E7CA, 0x000EC0E18000E7CE,
-    0x000E15A300017CBE, 0x000028868000E7D6, 0x8000E7DA00014C87,
-    0x000E39A5000E3DA6, 0x000E65AB8000E7DE, 0x8000E7E2000E65AA,
-    0x000E51AF8000E7E6, 0x000E69AD000E69AE, 0x8000E7EA000E75AC,
-    0x8000E7EE0001BDB3, 0x000179B68000E7F2, 0x8000E7FA8000E7F6,
-    0x000155BA000179B5, 0x8000E7FE00015DBB, 0x0000F9BE8000E802,
-    0x000145BC0000F9BF, 0x000271A80000F9BD, 0x000271AA000271A9,
-    0x0002B5AC000271AB, 0x0002B5AE0002BDAD, 0x0004EC840002ADAF,
-    0x8000E80A8000E806, 0x8000E8128000E80E, 0x8000E81A8000E816,
-    0x000369B88000E81E, 0x00030DBA000369B9, 0x00031DBC000305BB,
-    0x00030DBE000305BD, 0x0002D5B0000305BF, 0x0002D5B20002DDB1,
-    0x000379B4000369B3, 0x000369B6000379B5, 0x8000E822000361B7,
-    0x8000E82A8000E826, 0x8000E82E0002F5A3, 0x000231A6000231A5,
-    0x8000E8368000E832, 0x00021DAA8000E83A, 0x00020DAC000215AB,
-    0x0003B1AE0003B1AD, 0x000061800003B1AF, 0x0000058200006181,
-    0x0004F0868000E83E, 0x000410BE00036087, 0x8000E84A8000E846,
-    0x8000E8528000E84E, 0x8000E85A8000E856, 0x8000E8628000E85E,
-    0x000285A88000E682, 0x000295AA000295A9, 0x0002BDAC0002A5AB,
-    0x0002D1AE0002D5AD, 0x8000E8660002D1AF, 0x8000E86E8000E86A,
-    0x8000E8768000E872, 0x8000E87E8000E87A, 0x000175B88000E882,
-    0x000175BA00017DB9, 0x0001D5BC0001CDBB, 0x0001C9BE0001DDBD,
-    0x0002B5B00001C1BF, 0x000281B20002BDB1, 0x000155B4000281B3,
-    0x000155B600015DB5, 0x000EA4E100014DB7, 0x000E1CE300040884,
-    0x00049CBE8000E886, 0x8000E88E8000E88A, 0x8000E8968000E892,
-    0x8000E89E8000E89A, 0x000075818000E8A2, 0x000ECCEF00007580,
-    0x8000E8A600007582, 0x8000E8AE8000E8AA, 0x8000E8B60004C484,
-    0x00005CEF8000E8BA, 0x000190E18000E8BE, 0x000158E38000E8C2,
-    0x8000E8CA8000E8C6, 0x000758868000E8CE, 0x00059DB30004EC87,
-    0x8000E8B28000E842, 0x8000E8D68000E8D2, 0x0005B5B50005D5B6,
-    0x0005F9BB8000E8DA, 0x8000E8DE0005F1BA, 0x000619BF8000E8E2,
-    0x0005D5BD000611BE, 0x8000E8E60005D5BC, 0x8000E8EA0005D9A3,
-    0x000591A68000E8EE, 0x8000E8F68000E8F2, 0x0005B5AA0005F1A5,
-    0x8000E8FA0005BDAB, 0x000655AE8000E8FE, 0x000591AC00065DAF,
-    0x00065DA8000591AD, 0x00066DAA000665A9, 0x00067DAC000665AB,
-    0x00066DAE000665AD, 0x0000A0BE000665AF, 0x8000E9068000E902,
-    0x8000E90E8000E90A, 0x8000E9168000E912, 0x00074DB88000E91A,
-    0x000751BA000751B9, 0x000771BC000751BB, 0x000771BE000771BD,
-    0x0007C1B0000771BF, 0x0007C1B20007C1B1, 0x0007C1B40007C1B3,
-    0x0007C1B60007C1B5, 0x8000E91E0007C1B7, 0x8000E926000694E1,
-    0x0001A98000053CE3, 0x000019820001A981, 0x000080868000E92A,
-    0x8000E92E00010487, 0x8000E9368000E932, 0x8000E93E8000E93A,
-    0x0007A5B300051CEF, 0x8000E9428000E922, 0x8000E94A8000E946,
-    0x0007B5B50007A5B6, 0x000645BB8000E94E, 0x8000E952000645BA,
-    0x000645BF8000E956, 0x000655BD000645BE, 0x000765A3000655BC,
-    0x8000E95E8000E95A, 0x8000E9668000E962, 0x000775A5000765A6,
-    0x000685AB8000E96A, 0x8000E96E000685AA, 0x000685AF8000E972,
-    0x000695AD000685AE, 0x8000E976000695AC, 0x8000E97A000619B3,
-    0x000621B68000E97E, 0x8000E9868000E982, 0x0001F1BA000631B5,
-    0x8000E98A0001F9BB, 0x0001B1BE8000E98E, 0x0001D1BC0001BDBF,
-    0x0005D1A80001D1BD, 0x0005D5AA0005D1A9, 0x0005FDAC0005E9AB,
-    0x0005EDAE0005E5AD, 0x0003F980000211AF, 0x0003C5820003CD81,
-    0x001C00BE00032084, 0x000304878000E996, 0x000219B8001CBC86,
-    0x000225BA00022DB9, 0x0002F9BC0002E9BB, 0x0002E9BE0002F9BD,
-    0x000271B00002E9BF, 0x000241B2000279B1, 0x000231B4000241B3,
-    0x000235B600023DB5, 0x8000E7D2000229B7, 0x8000E99E8000E99A,
-    0x8000E9A2000551A3, 0x000569A6000579A5, 0x8000E9AA8000E9A6,
-    0x0002B9AA8000E9AE, 0x000299AC0002B1AB, 0x0002F9AE000299AD,
-    0x0001F8E10002F5AF, 0x0002F4E38000E9B2, 0x8000E9B6001CC084,
-    0x8000E9BE8000E9BA, 0x8000E9C68000E9C2, 0x8000E9CE8000E9CA,
-    0x8000E9D68000E9D2, 0x000230EF8000E9DA, 0x000209A88000E9DE,
-    0x000221AA000219A9, 0x000221AC000221AB, 0x000221AE000221AD,
-    0x8000E9E2000255AF, 0x8000E9EA8000E9E6, 0x8000E9F28000E9EE,
-    0x0000AD810000AD82, 0x0002D1B80000AD80, 0x0002E1BA0002D9B9,
-    0x000291BC0002E1BB, 0x000295BE00029DBD, 0x00022DB0000289BF,
-    0x000235B2000235B1, 0x00021DB4000205B3, 0x0002F1B60002F1B5,
-    0x000748E10002F1B7, 0x00061CE3000658E3, 0x8000E9F600066CE1,
-    0x8000E9FE8000E9FA, 0x80001B68001C10BE, 0x000005B580001B6C,
-    0x003B30E180001B70, 0x0006D0EF0001ECEF, 0x0002E9B30032BCE1,
-    0x80001B740002DCAC, 0x001D0CAC000005B5, 0x0002E9B50002E1B6,
-    0x0002B1BB80001B78, 0x001D28AC0002A9BA, 0x000275BF80001B7C,
-    0x000291BD000299BE, 0x8000E9920002A1BC, 0x001FD5B5001D70E1,
-    0x001D48E1001D6CE1, 0x0002B4E1001FF9B5, 0x001E15A8000258AC,
-    0x001EA1AA001EA1A9, 0x001EADAC001EBDAB, 0x001E9DAE001E95AD,
-    0x001E91B0001EF9AF, 0x001E91B2001E91B1, 0x001EB1B4001E91B3,
-    0x001EB1B6001EB1B5, 0x001E91B8001EB1B7, 0x001E95BA001E91B9,
-    0x001F55BC001F4DBB, 0x001F55BE001F5DBD, 0x001DDDAC001F4DBF,
-    0x001DE5AE001DEDAD, 0x000005B5001D09AF, 0x001DD5AA80001B80,
-    0x0002BCE1001DCDAB, 0x001D9DA6001D95A5, 0x000005B50002C8AC,
-    0x0002CCE180001B84, 0x001E49B3001D95A3, 0x80001B880002DCAC,
-    0x0002BCE1000005B5, 0x001E45B5001E75B6, 0x001E35BB0002E0AC,
-    0x000005B5001E35BA, 0x001E15BF80001B8C, 0x001E25BD001E15BE,
-    0x001F89A3001E25BC, 0x000380820025358F, 0x0002D8AC0002ACE1,
-    0x001F85A5001FB5A6, 0x001FF5AB000035B5, 0x80001B90001FF5AA,
-    0x001FD5AF80000B9D, 0x001FE5AD001FD5AE, 0x0019259E001FE5AC,
-    0x001BB59C001FC59F, 0x0015719A0019499D, 0x0011F5980015699B,
-    0x001385960017B599, 0x002D499400114597, 0x0028F592002C7195,
-    0x00297990002FA593, 0x00000DAC00297991, 0x000011AE000015AD,
-    0x800009B1800006BE, 0x800006C980000C96, 0x800006F1800006DE,
-    0x800006F980000CC6, 0x001CF5A280000CC9, 0x001DBDA00001B5A3,
-    0x0005E9BC001DB5A1, 0x0005F9BE000051BD, 0x0005C1B80005F5BF,
-    0x0005F9BA0005CDB9, 0x00003DB480000709, 0x0005E9B6000401B5,
-    0x00006DB00005E5B7, 0x0000D5B2000055B1, 0x00263182000055B3,
-    0x80001B94002A0583, 0x002CC186000280CC, 0x002A1D84002E3587,
-    0x002F3D8A002B9985, 0x80000CE10010C18B, 0x0014D98E80000CE4,
-    0x0012198C0016158F, 0x00167D920012918D, 0x80000CE70018B993,
-    0x001CD19680000B0A, 0x001A5D94001E5197, 0x001E699A001A3995,
-    0x80000CD2001F219B, 0x80001B98000515B8, 0x0002799C000214CC,
-    0x0005FDB880001B9C, 0x0002A8CC80001BA4, 0x023484CC80000D1D,
-    0x0234FCCC02338CCD, 0x0225A0CC023398CD, 0x0232BCCC0223ACCD,
-    0x80001BA80234B0CC, 0x0232BCCD80000E46, 0x000541B80230CCCC,
-    0x80000EA280000E49, 0x80000CCC80000EFF, 0x80000A2080000CD5,
-    0x80000CED000318CC, 0x80000CFF80000CF3, 0x80000D08023494CC,
-    0x80000D0E80000D0B, 0x03CEB48580001BAC, 0x001F3C8780000D89,
-    0x80001BB080000D17, 0x80000D2380000D8C, 0x0000A88D80000D20,
-    0x80000D260004A1B8, 0x001F088980000D29, 0x0000B88B80000D2C,
-    0x80000D2F000288CC, 0x80000D5080000D38, 0x0000C49180000D53,
-    0x80000BA90004DDB8, 0x80000D5F80000D5C, 0x80000D6280001BA0,
-    0x80000D7180000D65, 0x0000E99A80000D74, 0x022C5DBB0222F99B,
-    0x022D79B9001FA0E1, 0x800007F8001F50E3, 0x00002DBD0002E8CC,
-    0x022F6DB3000411B8, 0x00000DB180000E31, 0x022D91B780000E37,
-    0x000015B580000EB7, 0x80000038001ED8EF, 0x02253DBB80000070,
-    0x000011BD8000002A, 0x000019BF8000003C, 0x8000004200068DBE,
-    0x0000BCBF0000C0BC, 0x800000480000D0BE, 0x8000005880000052,
-    0x8000072C000060BA, 0x8000072E800007A2, 0x800007398000005E,
-    0x800007498000001E, 0x80000769800007E7, 0x800007740006E9B8,
-    0x800007768000004E, 0x8000078F800022AC, 0x000689A18000006C,
-    0x022199A30215F5A0, 0x000175A500017DA2, 0x0006EDA700017DA4,
-    0x0215D1A90215F9A6, 0x800007AF0215D9A8, 0x800007C50006A5AA,
-    0x800007FC800003A9, 0x0006CD9180000821, 0x0006E19300019590,
-    0x0006ED9502201192, 0x0001B9970006E594, 0x0006259900018596,
-    0x00018D9B00018D98, 0x0215159D0006219A, 0x00067D9F0006059C,
-    0x0000A9810006099E, 0x0220A5830000C180, 0x00057D850000BD82,
-    0x0000A18700057184, 0x00001D890006D586, 0x0006C58B00009988,
-    0x0001F18D0006C58A, 0x0001F18F0006BD8C, 0x8000085D8000083E,
-    0x80000883800003AB, 0x800008C9800003AE, 0x80000922800022B8,
-    0x8000095C800022BF, 0x8000095E80000960, 0x8000086C800022D6,
-    0x80000802800022DD, 0x80000847800000C8, 0x8000097B80000D77,
-    0x80000E588000000E, 0x800006B6800006B9, 0x800022C9800000CB,
-    0x800005DC800003B4, 0x8000063D80000617, 0x8000067180000654,
-    0x800003A380000688, 0x800006A98000232B, 0x8000039F800000DA,
-    0x8000079180000906, 0x800007F4800007F0, 0x80000994800006CE,
-    0x8000080880000804, 0x8000099880000BA0, 0x8000077180000AB9,
-    0x800007998000078D, 0x800007AC800007A4, 0x80000E28800007DB,
-    0x80000FF2800009AE, 0x800006F380000717, 0x8000101680001009,
-    0x800006F58000101E, 0x023A4CCC80000DE9, 0x02384CCC023A34CD,
-    0x02260CCC023FCCCD, 0x022644CC022718CC, 0x023B60CC022724CD,
-    0x022524CC023AC0CD, 0x02395CCC0224D0CD, 0x023B74CC0239A0CD,
-    0x80000E10023AACCD, 0x800004C980002327, 0x800004A4800004D5,
-    0x8000056E800004B1, 0x022A38CC800004BD, 0x80000574022ABCCD,
-    0x023950CC800005C1, 0x0220E8CC0221F8CC, 0x80000602022610CD,
-    0x80000622800004C3, 0x0236C4CD800004CF, 0x800005C70236D4CC,
-    0x023BA0CD80000628, 0x8000065C0224A4CC, 0x800007198000049F,
-    0x023478CC8000071C, 0x800007B380000726, 0x800007B6000151B8,
-    0x023018CD800007BA, 0x023630CC0231D4CC, 0x023E64CD80000729,
-    0x022E58CC023E80CC, 0x800007D2022F74CD, 0x8000086E800007D5,
-    0x023240CC80000871, 0x023D08CC0233F8CD, 0x023310CC023268CD,
-    0x02334CCC0230C4CD, 0x023070CC023254CD, 0x0230D4CC023044CD,
-    0x0236B0CC0230A0CD, 0x02307CCC023D78CC, 0x023F44CC0231C4CD,
-    0x023E2CCC023C84CD, 0x8000087B023904CC, 0x023C50CC80000962,
-    0x80000965023C24CD, 0x8000087E80000969, 0x8000097580000972,
-    0x800022FB800003BC, 0x80002303800022FF, 0x8000230F80002307,
-    0x0003B4CC8000230B, 0x000D74CC80002317, 0x000D54CC8000231F,
-    0x000D54CC80002323, 0x000D70CC8000232F, 0x8000233780002333,
-    0x800023558000233E, 0x023880CD8000235C, 0x000D18CC0238C8CC,
-    0x000D78CC80002366, 0x000D48CC8000236A, 0x000D58CC8000236E,
-    0x000D48CC80002372, 0x000D78CC8000237E, 0x000BC5808000237A,
-    0x0008AC94000BC981, 0x00081C92000F91B8, 0x0009F490000FB1B8,
-    0x000F95B8000FB1B8, 0x000FFDB8022F8189, 0x000F4589022FF98B,
-    0x000BD98E000BE18D, 0x000F9DB8000BCD8F, 0x000FADB80009FC94,
-    0x000FBDB802267993, 0x000FADB800096890, 0x000F9DB800091892,
-    0x000019B80002ECCC, 0x8000238A02215D9B, 0x000B819E000B819D,
-    0x800007B1000C719F, 0x0002F8CC800007D0, 0x000005B8021891A3,
-    0x0002E4CC8000238E, 0x8000239200003DB8, 0x000019B80002ECCC,
-    0x8000239602229DAB, 0x022EC5AE000CA1AD, 0x0002E4CC000CA1AF,
-    0x8000239A00000DB8, 0x000005B80002FCCC, 0x0002E4CC8000239E,
-    0x800023A200003DB8, 0x0002F8CC022105B9, 0x00000DB8022271BB,
-    0x000CD5BE022C89BD, 0x00010080800023A6, 0x0007088200040081,
-    0x00061084000039B8, 0x00030C8600051485, 0x0004A48800034487,
-    0x007B208A00780089, 0x0001C08C0002C0CC, 0x0002E4CC800023CA,
-    0x800023CE00007DB8, 0x00794892007AC491, 0x0002E4CC007C0093,
-    0x800023AA00000DB8, 0x000005B80002DCCC, 0x000019B8800023AE,
-    0x800023B20002E4CC, 0x000005B80002ECCC, 0x000009B8800023B6,
-    0x800023D20002E4CC, 0x000005B80002FCCC, 0x000031B8800023D6,
-    0x0001F0A80002E4CC, 0x800023BA0001F0A9, 0x000009B80001B0AB,
-    0x800023BE0002E4CC, 0x000009B800029CCC, 0x800023C200019CB1,
-    0x0002FCCC000015B8, 0x800023C6000118B5, 0x0002DCCC000039B8,
-    0x00011CBA800023DA, 0x007FF0BC000180BB, 0x007DF8BE007EF4BD,
-    0x022E3D8000737CBF, 0x022E3582022E3981, 0x022E2D84022E3183,
-    0x022E2586022E2985, 0x022E1D88022E2187, 0x022E158A022E1989,
-    0x000005B80002FCCC, 0x0002E4CC800023DE, 0x800023E200007DB8,
-    0x0002E4CC000E7191, 0x800023E600001DB8, 0x000031AA000031A9,
-    0x00044D8C000E6197, 0x8000241880002414, 0x000575AF0004519C,
-    0x800003E8800003CF, 0x0004B5B50002D882, 0x0005859F0004B5B4,
-    0x000EE5A40004B5B6, 0x000F95A6000F91A5, 0x0005B5B380000401,
-    0x022F61BC8000041A, 0x000270BE00022DBD, 0x0072419E00023C82,
-    0x00723999022E81AF, 0x000250BE8000042D, 0x000FCDB4000FD9B3,
-    0x000FF5B680000440, 0x00721D9E000FE9B7, 0x8000046680000453,
-    0x000FCDBC00721199, 0x000FB5BE0073B599, 0x000E45BE000378BE,
-    0x000879BC021AADBF, 0x00087DBA000879BD, 0x00087DB8000E5DBB,
-    0x00087DB6000875B9, 0x000865B4000845B7, 0x000875B2000875B5,
-    0x00080DB000087DB3, 0x000869AE00087DB1, 0x000851AC000869AF,
-    0x000841AA000851AD, 0x00085DA8000841AB, 0x000999A6000851A9,
-    0x000999A4000999A7, 0x000995A2000999A5, 0x000E39A000099DA3,
-    0x0009599E00099DA1, 0x000EB19C0009CD9F, 0x800004790009599D,
-    0x0003ECE30009459B, 0x800004F28000048C, 0x0003E0EF800000DC,
-    0x023048CC80000112, 0x00032882023064CC, 0x8000011F8000011B,
-    0x8000012780000123, 0x8000012F8000012B, 0x0003E8EF80000133,
-    0x800000E00003D8EF, 0x000354E3800000E5, 0x005FF1B000025CE3,
-    0x800000EF800000EA, 0x000965800072BD9E, 0x02287D8202288D81,
-    0x0009798402287583, 0x0009698600097985, 0x0009598800096987,
-    0x0009A98A00095989, 0x0009B98C0009A98B, 0x0009A98E0009B98D,
-    0x02282D90800000F4, 0x0228D5920008B991, 0x000E799402244D93,
-    0x0009519602299595, 0x0009719802261D97, 0x0009719A0009F199,
-    0x0009519C0229319B, 0x800000FE800000F9, 0x8000010880000103,
-    0x800001168000010D, 0x800004EF800003CC, 0x800005E18000058E,
-    0x000939A88000061C, 0x8000065980000642, 0x8000068D80000676,
-    0x800003EB800003D2, 0x8000041D80000404, 0x8000044380000430,
-    0x8000046980000456, 0x8000048F8000047C, 0x8000019F800004F5,
-    0x800001A7800001A3, 0x800001AF800001AB, 0x800001B7800001B3,
-    0x00089D80800001BB, 0x0008AD820008A581, 0x800001BF0008A583,
-    0x0008A9860008B985, 0x800001C30008A987, 0x0008298A00089D89,
-    0x000CED8C000CF18B, 0x0228F98E000CE98D, 0x0008A5900008818F,
-    0x000959920008A191, 0x800001C702270D93, 0x0008999600097595,
-    0x800001CF800001CB, 0x0008AD9A00095599, 0x0009499C0009559B,
-    0x800001D300094D9D, 0x000849A0800001D7, 0x000859A2000849A1,
-    0x000845A4800001DB, 0x000841A6800001DF, 0x000875A8800001E3,
-    0x0009A1AA800001E7, 0x000C29AC000871AB, 0x800001EB000811AD,
-    0x000C61B0000941AF, 0x800001EF00082DB1, 0x02294DB40009D5B3,
-    0x000835B6000835B5, 0x000815B800082DB7, 0x800001F30008D5B9,
-    0x021C45BC000815BB, 0x000805BE000F41BD, 0x00084580000BBDBF,
-    0x800001FB800001F7, 0x80000203800001FF, 0x000A6586000F8985,
-    0x000B1188000AE187, 0x8000093B022BED89, 0x800006A180002348,
-    0x800006AE800006AB, 0x00080990800023FE, 0x0008199200080991,
-    0x0008099400081993, 0x0008319600080995, 0x0008E19800083197,
-    0x0008E19A0008E199, 0x0008E99C0008E19B, 0x0008D19E0008E99D,
-    0x000B65A00008D19F, 0x00083DA2000835A1, 0x000611A4000831A3,
-    0x000851A6000855A5, 0x00087DA8000851A7, 0x000859AA000621A9,
-    0x000AA1AC000859AB, 0x000F25AE000B11AD, 0x000BF1B0000AA5AF,
-    0x000835B200083DB1, 0x000BF5B400099DB3, 0x0009B9B6000981B5,
-    0x000989B80009B9B7, 0x000985BA0226E5B9, 0x0224B1BC00099DBB,
-    0x000A85BE000BCDBD, 0x00742C82000BEDBF, 0x000B90A380002408,
-    0x000B90A5800032CC, 0x000A4D8780003B38, 0x000A7D8980002B98,
-    0x80002BBC000A7D88, 0x00088DAD80002BC8, 0x0008F9AF00088DAC,
-    0x80002BE080002BD8, 0x80002BEC000881B0, 0x80002BFC80002BF4,
-    0x000A399780002C08, 0x80002C7B80002C63, 0x80002C9B80002C8B,
-    0x80002CBB80002CAB, 0x022B459F000B94BC, 0x80002CDF80002CDB,
-    0x000B288380002CE3, 0x80002D3F80002D17, 0x80002D5780002D4F,
-    0x80002D6780002D5B, 0x000B388B000B0C88, 0x80002D8780002D83,
-    0x80002DA380002D9F, 0x000AF1B180002DAB, 0x80002DBB000AA1B0,
-    0x000ABDB580002E57, 0x000AADB7000ABDB4, 0x000A9DB9000AADB6,
-    0x000A6DBB000A9DB8, 0x000A7DBD000A6DBA, 0x000A6DBF000A7DBC,
-    0x007A6082000A6DBE, 0x80000794007A6883, 0x007A4C8680002E6B,
-    0x80002EAB007A5087, 0x80002EC780002EBF, 0x800007D980002ECB,
-    0x80002EFE80002EDB, 0x80002F3680002F2A, 0x007AFC9280002F56,
-    0x80002F6680002F62, 0x80002F6A007AF891, 0x80002F8680002F82,
-    0x80002FBE80002FB2, 0x80002FCE80002FC6, 0x80002FED80002FDE,
-    0x8000302180002FF5, 0x007A2CA280003025, 0x800030318000302D,
-    0x8000304180003039, 0x8000306D80003051, 0x800030F9800030E5,
-    0x007A2CA8007A44AB, 0x800030FD007A20A9, 0x800031818000317D,
-    0x007A10B280003195, 0x800031990079D8B3, 0x800031A9800031A1,
-    0x0079D0B4800031B1, 0x800031BD0079A4B5, 0x80003203800031F3,
-    0x8000329480003217, 0x800032B0800032A0, 0x025589A0800032BC,
-    0x025571A2025571A1, 0x00747DA4007465A3, 0x00746DA6007465A5,
-    0x00745DA8007465A7, 0x007B51AA007B51A9, 0x007B95AC007BB1AB,
-    0x007BEDAE007BF1AD, 0x007BA9B0007BD9AF, 0x007AD5B2007AD5B1,
-    0x007AF5B4007AEDB3, 0x007AE5B6007AF9B5, 0x007AA5B8007A9DB7,
-    0x007AA5BA007AADB9, 0x007AA5BC007ABDBB, 0x007A9DBE007A95BD,
-    0x800032C8007A89BF, 0x800032D4800032D0, 0x800032DC800032D8,
-    0x80003398800032E0, 0x800033EC800033E0, 0x800033F4800033F0,
-    0x80003400800033F8, 0x8000341880003404, 0x8000343C80003434,
-    0x800034808000347C, 0x8000348C80003488, 0x8000349880003494,
-    0x800034C0800034BC, 0x800034E0800034DC, 0x800034E8800034E4,
-    0x80003520800034FC, 0x0079F98080003524, 0x007E7D82007E3981,
-    0x007E6D84007E7583, 0x007E9D86007E9585, 0x007EAD88007E9587,
-    0x007EBD8A007EB589, 0x007EAD8C007EB58B, 0x007E918E007E918D,
-    0x007EF190007E918F, 0x007EF192007EF191, 0x007E9194007EF193,
-    0x007E9196007E9195, 0x007EB198007E9197, 0x007EB19A007EB199,
-    0x007E919C007EB19B, 0x007E919E007E919D, 0x007E71A0007E919F,
-    0x007E71A2007E71A1, 0x007F81A4007E71A3, 0x007FC9A6007FB5A5,
-    0x007FC1A8007FC1A7, 0x007FD5AA007FDDA9, 0x007F25AC007F21AB,
-    0x007F1DAE007F29AD, 0x007F09B0007F05AF, 0x007CCDB2007F2DB1,
-    0x007CDDB4007CC5B3, 0x007355B6007CC5B5, 0x007745B8007345B7,
-    0x007349BA007379B9, 0x007359BC007705BB, 0x00771DBE007329BD,
-    0x800035280076FDBF, 0x8000354C0077DC88, 0x0077D08D8000358C,
-    0x024B28CC80003590, 0x800035B0024B5CCD, 0x0077C083800035C0,
-    0x0077E085800035C8, 0x02497CCC800035E8, 0x8000360C0246B4CD,
-    0x8000365880003650, 0x800036788000366C, 0x8000368880003680,
-    0x800036908000368C, 0x800036AC800036A8, 0x800036D0800036C8,
-    0x800036DC800036D8, 0x0008A8BE800036E8, 0x800036F4800036F0,
-    0x80003700800036F8, 0x800037A480003704, 0x007768A1800037B8,
-    0x800037F0800037EC, 0x007808A5800037F8, 0x800038240078E0A4,
-    0x800038348000382C, 0x007810BB8000383C, 0x0079D8BD007824BA,
-    0x800038780079C8BC, 0x800038848000387C, 0x00787CB380003888,
-    0x8000388C007844B2, 0x007804B7800038B8, 0x007B1D8000780CB6,
-    0x007B2582007B2D81, 0x007B5184007B3183, 0x007B5586007B5D85,
-    0x007B7588007B4D87, 0x007B498A007B7989, 0x007B598C007B498B,
-    0x007B498E007B598D, 0x007BC190007B418F, 0x007BC592007BC191,
-    0x007BC594007BDD93, 0x007BC596007BCD95, 0x007BCD98007BFD97,
-    0x007BDD9A007BD599, 0x007BCD9C007BD59B, 0x007BB59E007BBD9D,
-    0x000499A00004719F, 0x0004A9A2000499A1, 0x0004BDA40004A9A3,
-    0x0004ADA60004A5A5, 0x00049DA80004A5A7, 0x0004E1AA0004E1A9,
-    0x0004E5AC0004FDAB, 0x0004E5AE0004EDAD, 0x0004A5B00004DDAF,
-    0x0004A5B20004ADB1, 0x0004A5B40004BDB3, 0x0004A5B60004ADB5,
-    0x000469B8000459B7, 0x000479BA000469B9, 0x000469BC000479BB,
-    0x000459BE000469BD, 0x0005B58000044DBF, 0x0005B5820005BD81,
-    0x0005DD840005C583, 0x0005CD860005C585, 0x0005C5880005FD87,
-    0x0005C58A0005CD89, 0x0005C58C0005DD8B, 0x0005C58E0005CD8D,
-    0x0004C5900004BD8F, 0x0004C5920004CD91, 0x0004C5940004DD93,
-    0x0004C5960004CD95, 0x0004C5980004FD97, 0x0004C59A0004CD99,
-    0x0004C59C0004DD9B, 0x0004C59E0004CD9D, 0x000445A00004BD9F,
-    0x000445A200044DA1, 0x000445A400045DA3, 0x000445A600044DA5,
-    0x000445A800047DA7, 0x000445AA00044DA9, 0x000445AC00045DAB,
-    0x000445AE00044DAD, 0x0006C9B00006BDAF, 0x0006DDB20006C9B1,
-    0x0006F5B40006EDB3, 0x0006EDB60006F9B5, 0x0006A5B800069DB7,
-    0x0006A5BA0006ADB9, 0x0006A5BC0006BDBB, 0x0006A5BE0006ADBD,
-    0x0005658000069DBF, 0x0005658200056D81, 0x0005658400057D83,
-    0x0005658600056D85, 0x0005A98800059987, 0x0005B98A0005A989,
-    0x0005A98C0005B98B, 0x00059D8E0005A98D, 0x0005ED900005958F,
-    0x0005F5920005FD91, 0x0005959400058D93, 0x0005859600058D95,
-    0x000585980005BD97, 0x00059D9A00059599, 0x0001F59C0001ED9B,
-    0x0001F59E0001FD9D, 0x000115A00001ED9F, 0x000115A200011DA1,
-    0x000135A400012DA3, 0x000135A600013DA5, 0x000115A800012DA7,
-    0x000115AA00011DA9, 0x000141AC00014DAB, 0x000141AE000141AD,
-    0x000C01B0000141AF, 0x000ECDB2000EC5B1, 0x000EDDB4000EC5B3,
-    0x000ECDB6000EC5B5, 0x000EFDB8000EC5B7, 0x000ECDBA000EC5B9,
-    0x000EDDBC000EC5BB, 0x000ECDBE000EC5BD, 0x00737D93000EC5BF,
-    0x0073719100723D92, 0x0252D19702536190, 0x0072599500721596,
-    0x025D619B0073E194, 0x0252E19900722D9A, 0x0072299F00721D98,
-    0x0072AD9D0253A99E, 0x0073618300722D9C, 0x0073598100735D82,
-    0x0073698700734580, 0x0073698500729186, 0x007FE18B00737584,
-    0x00736D8900730D8A, 0x0072A58F00076588, 0x007FFD8D00075D8E,
-    0x800038BC00730D8C, 0x800038D8800038C4, 0x800038F8800038DC,
-    0x8000390480003900, 0x8000394880003928, 0x800039608000395C,
-    0x8000399480003968, 0x800039A880003998, 0x0072E1A3800039CC,
-    0x025219A1025299A2, 0x025FD1A70079E5A0, 0x0072FDA5025209A6,
-    0x800039D80072F5A4, 0x0072A1A9007281AA, 0x800039E802521DA8,
-    0x80003A0880003A00, 0x80003A1480003A0C, 0x80003A2480003A20,
-    0x80003A4880003A2C, 0x80003A5080003A4C, 0x80003A6C80003A68,
-    0x80003A8880003A74, 0x80003A9080003A8C, 0x80003A9C80003A98,
-    0x80003AA480003AA0, 0x80003AB880003AA8, 0x80003AD480003AD0,
-    0x80003AE480003AE0, 0x80003AEC80003AE8, 0x80003AF880003AF4,
-    0x80003B0880003AFC, 0x80003B1880003B10, 0x0000BD8080003B1C,
-    0x0000CD820000C581, 0x0000DD840000C583, 0x0000CD860000C585,
-    0x0000FD880000C587, 0x0000CD8A0000C589, 0x0000DD8C0000C58B,
-    0x0000CD8E0000C58D, 0x0000BD900000C58F, 0x0000159200001D91,
-    0x0000359400002D93, 0x80003B2400003D95, 0x80003B2C80003B28,
-    0x80003B3C80003B34, 0x80002D6B8000243C, 0x8000252480002D6F,
-    0x00021C80800024B2, 0x0007248200019C81, 0x00041C8400050883,
-    0x00198086001A1485, 0x001C0088025D2180, 0x0010008A001E2489,
-    0x0011F48C00122C8B, 0x002B388E0028008D, 0x800024B6002A3C8F,
-    0x0000059900744CE3, 0x00001D9E800024BA, 0x00026082000204E3,
-    0x800024BE00003D99, 0x00027CBE0002F4EF, 0x0002688200000D9E,
-    0x800024C200080CEF, 0x0002F8E300000599, 0x0000F19E000274BE,
-    0x00026C82000214E3, 0x00000D99800024C6, 0x0000359E000244BE,
-    0x0000059900026882, 0x0002CCEF800024CA, 0x0000159E000274BE,
-    0x00026CE30002D8EF, 0x800024CE000129B6, 0x000214E300027C82,
-    0x00027CBE00001D99, 0x0002C0EF0000159E, 0x0000059900026882,
-    0x0002BCEF800024D2, 0x000109B9000101BA, 0x0002ACA600765DB8,
-    0x0003D8A40001E8E3, 0x800024D6000280E3, 0x800024DE800024DA,
-    0x800024E20003C4A1, 0x0002E0BE0005A8AF, 0x800024E60002ECBE,
-    0x000400A80002E4EF, 0x0002C0E30002E8EF, 0x800024EA0002F4BE,
-    0x0005D8B20005C8B5, 0x0002C8BE0002D0E3, 0x800024EE0002C0E3,
-    0x800024F20002C4EF, 0x0002CCEF000394EF, 0x0005F8B80005E8BB,
-    0x00729486800024F6, 0x0002448200024082, 0x800024FA00023882,
-    0x800024FE00020082, 0x0000B19980002502, 0x0000BD9900006C8F,
-    0x00035D9E0072CC8D, 0x0000A99E0072D08B, 0x80003A1080002506,
-    0x0000DD9E00008897, 0x0000C9990000CC95, 0x000000940000D893,
-    0x8000250A00000491, 0x0000F19E0003289F, 0x0000D19900036C9D,
-    0x8000251500037C9B, 0x00027CBE0000A099, 0x000260820001259E,
-    0x0002F0EF00013D99, 0x000274BE0001259E, 0x000220E300025482,
-    0x0001619E000274BE, 0x00026C8200016599, 0x0001599900015D9A,
-    0x0001299F000200BE, 0x000250BE0001259E, 0x00012D9900027882,
-    0x0001C59A0001CD99, 0x00011D9E0001CD9A, 0x000260BE0001F59F,
-    0x0001E99E00025C82, 0x00029C820001E99F, 0x00013D99000294BE,
-    0x8000251900024482, 0x00012D9E00016D9A, 0x000021990001759F,
-    0x000218BE800000BC, 0x0000219A00015D99, 0x000000CE00025082,
-    0x80002530800000C1, 0x0001399E8000251D, 0x80002A6C00004D9F,
-    0x000264BE0000759E, 0x0000119900024CBE, 0x0000299E00026082,
-    0x0005D8BE80002607, 0x000000DE000000D9, 0x0005D8828000266A,
-    0x000538E30005F882, 0x00056CE3000540E3, 0x00059CE300057CE3,
-    0x0005CCE30005B0E3, 0x000578EF0005ECE3, 0x000504EF00050CEF,
-    0x0005F4EF00053CEF, 0x0005A8EF000584EF, 0x0007DD9B0005B8EF,
-    0x0007AD9E0007A199, 0x0007ED9F0005B8E3, 0x0003C09D000248BE,
-    0x000795990007E99C, 0x000260820007919E, 0x800026AD00001D99,
-    0x0000099E000274BE, 0x000310AB0002D0EF, 0x0002E8E300001999,
-    0x000394AF000210E3, 0x000358AD00027482, 0x00027CBE800026D9,
-    0x000260820000399E, 0x8000270300001D99, 0x000234EF000320A6,
-    0x000204BB0002E4EF, 0x000204B9000264E3, 0x0002C8BF000268BE,
-    0x00020CBD000230E3, 0x0003A8B300006D9E, 0x0003E8B1000318EF,
-    0x000270B700027482, 0x0003A0B50002C0EF, 0x800027280001FD99,
-    0x0000099E000274BE, 0x00026C82000214E3, 0x00000D998000273D,
-    0x00026C82000240BE, 0x000005990000319E, 0x000270BE80002745,
-    0x0000159E00021482, 0x000268E30002D8EF, 0x000274E300000999,
-    0x00027CBE80002761, 0x00020CE30000199E, 0x8000276900025482,
-    0x000210EF0000059E, 0x0002B8EF00000D99, 0x0002D0EF0000FD99,
-    0x00029CBE00026C82, 0x80002771000458EF, 0x000264BE0000059E,
-    0x00023CE30002E0E3, 0x8000278300003D99, 0x0002ECEF0002CCEF,
-    0x0000799900025482, 0x00026C820004A0E3, 0x0000619E000204BE,
-    0x8000278F000274E3, 0x000200E3000260BE, 0x0001B88200001D9E,
-    0x0002F4EF80002797, 0x800027AD00018882, 0x800027EB000198BE,
-    0x0001AC820002C0EF, 0x00001481800027FC, 0x0003E1990003D999,
-    0x0001FCE30003D19E, 0x000248BE00005C84, 0x0002C0EF00004086,
-    0x0003C19E0003B888, 0x0003C19900020CBE, 0x0003F48F00023C82,
-    0x0001B0E30003C19A, 0x000274E38000281E, 0x00034C9500038892,
-    0x0003A99E80002829, 0x0000359A0003A19F, 0x0002E09B00035498,
-    0x0002A4EF00002999, 0x0002F4EF00028CBE, 0x0000C59E80002879,
-    0x000278820000D19F, 0x000210E30002BCEF, 0x0000319A0000E999,
-    0x0002C4EF000254BE, 0x8000288D00021082, 0x0000CD9E0003C882,
-    0x00023CAF00001D9F, 0x000204B1000210E3, 0x000250BE000208B0,
-    0x00000999000228B2, 0x800028AF0003BCBE, 0x0000399E0000359A,
-    0x800028BD00000D9F, 0x0002FCBD80002905, 0x000145BF80002913,
-    0x0001ED9900038082, 0x0003FCE30003A0E3, 0x0001FD9E0001F99A,
-    0x00003D990001E19F, 0x0000059E000244BE, 0x0002608280002A70,
-    0x80002A7400001D99, 0x0002DCEF000318EF, 0x000260E300000D9E,
-    0x000214E3000200BE, 0x80002A7800001D99, 0x0002D4EF00027C82,
-    0x0000319E000244BE, 0x000204E300026C82, 0x0002CCEF0000CD99,
-    0x0002DCEF000270BE, 0x000200E30000DD9E, 0x0002688200003199,
-    0x000000E4000260BE, 0x000000E60000219E, 0x80002A7C000000E7,
-    0x002D519C000000E9, 0x0002C4EF002D519D, 0x000000EE80002A88,
-    0x00000DB3800026E8, 0x001C01B1001FC5B2, 0x001FC1B780001B18,
-    0x001FD1B5001FD9B6, 0x001FFDBB001FD9B4, 0x001FA9B9001FFDBA,
-    0x001FDDBF001FF9B8, 0x001FDDBD001FDDBE, 0x800019F4001FDDBC,
-    0x80001B1C800019F8, 0x800019FC80001B20, 0x80001B2480001A00,
-    0x80001B2C80001B28, 0x80001B3480001B30, 0x80001A4480001B38,
-    0x80001AA480001A58, 0x80001A0480001AA8, 0x80001AAC02798CE3,
-    0x80001A08027A80E1, 0x80001AB080001A0C, 0x80001AB880001AB4,
-    0x80001AC080001ABC, 0x80001AC880001AC4, 0x80001ACC02798CEF,
-    0x80001A1480001A10, 0x80001AD080001A18, 0x80001A1C80001AD4,
-    0x00495CEF80001A20, 0x004064E1004644EF, 0x004170E303ABA0E1,
-    0x80001AD80054F8E3, 0x80001AE080001ADC, 0x80001A2480001AE4,
-    0x80001AE880001A28, 0x80001A2C80001AEC, 0x80001AF080001A30,
-    0x80001AA080001AF4, 0x80001A4080001A3C, 0x80001B4080001B3C,
-    0x80001B4880001B44, 0x0001318080001B4C, 0x00015D8200015581,
-    0x0001258400017183, 0x00019D8600019585, 0x0001B58800018D87,
-    0x0001ED8A0001E589, 0x0001E58C0001FD8B, 0x00021D8E0002158D,
-    0x00017D900001058F, 0x0001119200010D91, 0x0001319400011193,
-    0x0001319600013195, 0x0001119800013197, 0x0001E19A00011D99,
-    0x0001C19C0001E19B, 0x0001C19E0001C19D, 0x000149A00001C19F,
-    0x000151A2000149A1, 0x000171A4000151A3, 0x000171A6000171A5,
-    0x000151A8000171A7, 0x000151AA000151A9, 0x0003B1AC000151AB,
-    0x0003B1AE0003B1AD, 0x0003D1B00003B1AF, 0x0003D1B20003D1B1,
-    0x0003F1B40003D1B3, 0x0003F1B60003F1B5, 0x0003D1B80003F1B7,
-    0x0003D1BA0003D1B9, 0x0003B1BC0003D1BB, 0x0003B1BE0003B1BD,
-    0x80001B500003B1BF, 0x80001A4880001B54, 0x80001B5C80001B58,
-    0x80001A5080001A4C, 0x80001A5C80001A54, 0x80001A6480001A60,
-    0x80001A6C80001A68, 0x80001A7480001A70, 0x80001A7C80001A78,
-    0x80001A8480001A80, 0x80001A8C80001A88, 0x80001B6080001A90,
-    0x80001A9480001B64, 0x80001A9C80001A98, 0x80001AFC80001AF8,
-    0x80001B0480001B00, 0x0000E98680001B08, 0x0000E9840000E987,
-    0x0000C9820000E985, 0x0000C9800000C983, 0x00007D8E0000C981,
-    0x00000D8C80001B0C, 0x0000C98A0000758D, 0x0000C9880000C98B,
-    0x0001A1960000C989, 0x0001F9940001A197, 0x006DCD920001F995,
-    0x80001B10006D9993, 0x0001559E80001B14, 0x00014D9C006DB19F,
-    0x006DB19A00015D9D, 0x000189980001559B, 0x0366C8E100018999,
-    0x031F5CE3034F70E1, 0x033BD4E10363C8E3, 0x032BC4E3030400E1,
-    0x80001A340333C4E3, 0x033274EF80001A38, 0x80002BA40332E8EF,
-    0x030528EF80002E5B, 0x03F654E10367ECEF, 0x03E6FCE303DACCE1,
-    0x038C84E103CB94E3, 0x03A844E30390A8E1, 0x80002BA8038784E3,
-    0x03A898EF80002BC0, 0x80002BAC038C3CEF, 0x03CABCEF8000349C,
-    0x020F74E1033FACEF, 0x02322CE3026BD0E1, 0x02C4ECE1021350E3,
-    0x02F24CE3023A20E1, 0x80002BB402DB40E3, 0x02E13CEF80002B9C,
-    0x80002EE6020B70EF, 0x0212F8EF80002E6F, 0x034F04EF026FFCEF,
-    0x800019F002E18CE1, 0x029F84E102EB98E3, 0x02A38CE3029660E1,
-    0x80000207029A08E3, 0x028B34EF8000020D, 0x034FA4E30286B8EF,
-    0x03535CE180000213, 0x80000219029A8CEF, 0x800002258000021F,
-    0x800002318000022B, 0x8000023D80000237, 0x8000024980000243,
-    0x800002558000024F, 0x8000033980000333, 0x800003458000033F,
-    0x800003518000034B, 0x8000035D80000357, 0x8000038180000375,
-    0x8000038D80000387, 0x8000039980000393, 0x800003638000037B,
-    0x8000025B8000036C, 0x8000026780000273, 0x80000279800002BB,
-    0x8000026180000297, 0x8000026D80000291, 0x800002D3800002A3,
-    0x800002EB800002F1, 0x8000032D800002DF, 0x800002CD800002C1,
-    0x800002D9800002E5, 0x800002AF80000303, 0x80000321800002F7,
-    0x8000029D80000309, 0x800002B58000027F, 0x800002FD800002C7,
-    0x8000032780000285, 0x8000030F800002A9, 0x800003158000028B,
-    0x0001F1808000031B, 0x0001F1820001F181, 0x000035840001F183,
-    0x0015A58600003D85, 0x00DA249E00003187, 0x80002FF180002E0F,
-    0x80000786800035D0, 0x8000051480000505, 0x0000719080000523,
-    0x0000719200007191, 0x0000AD9400007193, 0x00035D9600035595,
-    0x00036D9800035597, 0x00037D9A00037599, 0x00036D9C0003759B,
-    0x00035D9E0003559D, 0x000649A00003559F, 0x000599A2000581A1,
-    0x0005C1A40005CDA3, 0x000519A60005C1A5, 0x000535A800050DA7,
-    0x000579AA000561A9, 0x000691AC000579AB, 0x000395AE000691AD,
-    0x0003F5B000038DAF, 0x0003F5B20003FDB1, 0x000395B400038DB3,
-    0x000395B600039DB5, 0x0003B5B800038DB7, 0x0003B5BA0003BDB9,
-    0x000395BC00038DBB, 0x000395BE00039DBD, 0x8000054180000532,
-    0x800005538000054A, 0x800005658000055C, 0x8000059A80000591,
-    0x800005A6800005A3, 0x800005AC800005A9, 0x80004DC480003B64,
-    0x80006954800056F4, 0x80007BB480007284, 0x8000A07480008E14,
-    0x8000BC0B8000A9A4, 0x8000CE728000C542, 0x8000E0D28000D7A2,
-    0x8000AF588000BC4F, 0x80002BF88000A624, 0x80002BA080002C8F,
-    0x8000324780002BE4, 0x8000355480003207, 0x8000340C80002DDB,
-    0x8000356880003890, 0x8000359880002D23, 0x80003918800034A4,
-    0x80002D8B80002C5F, 0x80003A3080003300, 0x80002CB380002C23,
-    0x80002FC2800032A4, 0x80002D5F80002DCF, 0x80002F3280002F3E,
-    0x80002C1F800034F4, 0x80002D7B800038A4, 0x800005AF80002ED3,
-    0x800005B5800005B2, 0x00031588800005B8, 0x0003158A00031D89,
-    0x0003358C00032D8B, 0x0003358E00033D8D, 0x0004C18000032D8F,
-    0x0004DD820004C181, 0x0003358400032D83, 0x0003358600033D85,
-    0x0003159800032D87, 0x0003159A00031D99, 0x0003359C00032D9B,
-    0x0003359E00033D9D, 0x00035590800005E4, 0x0003559200035D91,
-    0x0003359400032D93, 0x0003359600033D95, 0x000311A800032D97,
-    0x000311AA000311A9, 0x000331AC000311AB, 0x000331AE000331AD,
-    0x0003D1A0000331AF, 0x0003D1A20003D1A1, 0x000331A40003D1A3,
-    0x000331A6000331A5, 0x000311B8000331A7, 0x000311BA000311B9,
-    0x000331BC000311BB, 0x000331BE000331BD, 0x000351B0000331BF,
-    0x000351B2000351B1, 0x000331B4000351B3, 0x000331B6000331B5,
-    0x800005E7000331B7, 0x800005ED800005EA, 0x800005F3800005F0,
-    0x800005F9800005F6, 0x800005FF800005FC, 0x800004E08000061F,
-    0x800005D28000057F, 0x800006338000060D, 0x800006678000064A,
-    0x800006958000067E, 0x800003F5800003DC, 0x800007118000040E,
-    0x8000084080000843, 0x800025388000073E, 0x800025A280002571,
-    0x800027B8800025C7, 0x800027C7800027C3, 0x800027CF800027CB,
-    0x8000296D8000292E, 0x8000299980002995, 0x800029C6800029BE,
-    0x800029F4800029E2, 0x800029FC800029F8, 0x80002A2F80002A14,
-    0x80002C1880002A4A, 0x8000253C003880F0, 0x8000255680002549,
-    0x8000257580002563, 0x8000259480002582, 0x800025B6800025A6,
-    0x800025D5800025CB, 0x800025ED800025E3, 0x00012980800025FA,
-    0x0001398200012981, 0x0001298400013983, 0x0001598600012985,
-    0x0001698800015987, 0x0001798A00016989, 0x0001698C0001798B,
-    0x0001598E0001698D, 0x000129900001598F, 0x0004019200012991,
-    0x0001259400013D93, 0x00055D9600055595, 0x00056D9800055597,
-    0x00057D9A00057599, 0x00056D9C0005759B, 0x00053D9E0005359D,
-    0x0005CDA00005359F, 0x0005DDA20005D5A1, 0x0001E9A40005D5A3,
-    0x000119A60001E9A5, 0x000129A8000119A7, 0x000529AA0005F9A9,
-    0x000539AC000529AB, 0x000135AE000539AD, 0x000155B000012DAF,
-    0x000155B200015DB1, 0x000175B400016DB3, 0x000175B600017DB5,
-    0x000155B800016DB7, 0x000155BA00015DB9, 0x000565BC00057DBB,
-    0x000565BE00056DBD, 0x0006EDB100013DBF, 0x0006FDB30006EDB0,
-    0x0006EDB50006FDB2, 0x00071DB70006EDB4, 0x00072DB900071DB6,
-    0x00073DBB00072DB8, 0x00072DBD00073DBA, 0x00071DBF00072DBC,
-    0x0006EDA100071DBE, 0x0006FDA30006EDA0, 0x0006EDA50006FDA2,
-    0x00069DA70006EDA4, 0x0006ADA900069DA6, 0x0006BDAB0006ADA8,
-    0x0006ADAD0006BDAA, 0x00069DAF0006ADAC, 0x0006AD9100069DAE,
-    0x0006BD930006AD90, 0x0006AD950006BD92, 0x0006FD970006AD94,
-    0x0006CD990006FD96, 0x0006DD9B0006CD98, 0x0006CD9D0006DD9A,
-    0x00061D9F0006CD9C, 0x00020D8100061D9E, 0x00021D8300020D80,
-    0x00020D8500021D82, 0x00023D8700020D84, 0x00020D8900023D86,
-    0x00021D8B00020D88, 0x00020D8D00021D8A, 0x0006DD8F00020D8C,
-    0x001BE98A0006DD8E, 0x000D7588001A9D8B, 0x001CBD8E0064AD89,
-    0x001A8D8C001B198F, 0x00081182001BF18D, 0x0064358000081183,
-    0x000D498600646181, 0x000D498400645187, 0x000DE59A000D4985,
-    0x000DD598001B759B, 0x000DE19E000DED99, 0x000DF19C000DE19F,
-    0x000DE592000DF19D, 0x001B8990000DED93, 0x000DE596000DED91,
-    0x000DF594000DED97, 0x000D55AA000DED95, 0x000D45A8000D5DAB,
-    0x000D75AE000D5DA9, 0x000D45AC000D7DAF, 0x000D61A2000D7DAD,
-    0x000D11A0000D61A3, 0x0008C1A6000D11A1, 0x0008D9A4000D7DA7,
-    0x0003D1BA0008D9A5, 0x0003C1B80003D1BB, 0x0003F1BE0003C1B9,
-    0x0003C1BC0003F1BF, 0x000D15B20003C1BD, 0x000D05B0000D1DB3,
-    0x0003F1B6000D1DB1, 0x000D05B40003F1B7, 0x800025DF000CA1B5,
-    0x8000268C8000261F, 0x800026DD800026CB, 0x8000272C80002707,
-    0x8000260B80002741, 0x8000266E80002615, 0x8000262380002675,
-    0x800026308000267F, 0x8000274980002637, 0x8000276D80002765,
-    0x8000278B8000277F, 0x800027A280002793, 0x80002647800027DD,
-    0x800026B18000265A, 0x80002690800026BB, 0x800026CF800026A0,
-    0x8000281A800026EF, 0x800028918000284E, 0x80002A4E800028E6,
-    0x80002A6080002A5C, 0x800026F980002A64, 0x800027188000270B,
-    0x8000274D80002730, 0x8000277580002757, 0x800027B18000279B,
-    0x800027BC800027A6, 0x800027E1800027D3, 0x8000281080002800,
-    0x8000282D800027EF, 0x800028478000283D, 0x8000285280002822,
-    0x8000287D80002862, 0x800028958000286F, 0x800028C8800028C1,
-    0x800028D2800028A5, 0x800028B3800028DC, 0x8000290980002917,
-    0x80002924800028EA, 0x800028FB800028F1, 0x8000293F80002932,
-    0x8000295380002949, 0x800025908000295D, 0x8000297E80002971,
-    0x8000299D80002985, 0x800029B1800029A4, 0x800029D4800029CA,
-    0x80002A00800029E6, 0x80002A1880002A0D, 0x80002A3380002A22,
-    0x80002A5280002A3A, 0x800004E5800003C2, 0x800005D780000584,
-    0x8000063880000612, 0x8000066C8000064F, 0x8000069A80000683,
-    0x800003FB800003E2, 0x8000042780000414, 0x8000044D8000043A,
-    0x8000047380000460, 0x8000049980000486, 0x8000050E800004FF,
-    0x8000052C8000051D, 0x800008658000053B, 0x800006C080000826,
-    0x8000091F800007FE, 0x8000082C80000935, 0x800008338000082F,
-    0x80002FE280000723, 0x80002EDF800031D0, 0x8000324B800031C1,
-    0x8000090880000929, 0x800008D180000E64, 0x8000073080000885,
-    0x800006FB8000074B, 0x800008968000080C, 0x8000090B8000092C,
-    0x80000E7480000E68, 0x8000089B800008DD, 0x8000088880000714,
-    0x800006FE80000754, 0x80000E78800007A8, 0x80000829800008E4,
-    0x8000085A8000089E, 0x80000E7C80000917, 0x80000816800008E7,
-    0x800008EA800008A1, 0x800008CB80000819, 0x800008EE800008A4,
-    0x800008CE8000081D, 0x800008A8001710E3, 0x800008FF800008F9,
-    0x8000088C8000078A, 0x8000070280000758, 0x800009438000093F,
-    0x800009388000094B, 0x8000091A0017D8EF, 0x0013A4E380000E80,
-    0x001388E30013A8E3, 0x00127CE3001330E3, 0x001238E3001248E3,
-    0x0012ACEF0013E8E3, 0x001294EF001280EF, 0x00131CEF0012C8EF,
-    0x00135CEF001304EF, 0x800008F60013B0EF, 0x8000090E8000092F,
-    0x800008D480000E6C, 0x8000075F80000890, 0x8000091180000932,
-    0x800008D780000E70, 0x8000076280000893, 0x80000765800008AF,
-    0x00111CE3800007EB, 0x00167CE30011E4E3, 0x0010B8E300167CE3,
-    0x00117CE300108CE3, 0x001068EF00113CE3, 0x001054EF001040EF,
-    0x0011E8EF001198EF, 0x001114EF001138EF, 0x001064E30016ACEF,
-    0x00109CE30010B0E3, 0x001020E30010F0E3, 0x001008E3001038E3,
-    0x001334EF00106CE3, 0x0010FCEF001304EF, 0x0010B4EF0010D8EF,
-    0x001080EF00109CEF, 0x003E2D80001054EF, 0x003E3D82003E3581,
-    0x003E2D84003E3583, 0x003E5D86003E5585, 0x003E6D88003E5587,
-    0x003E7D8A003E7589, 0x003E6D8C003E758B, 0x003E5D8E003E558D,
-    0x003E2D90003E558F, 0x003E3D92003E3591, 0x003E2D94003E3593,
-    0x003EDD96003ED595, 0x003EED98003ED597, 0x003EFD9A003EF599,
-    0x003EED9C003EF59B, 0x003EDD9E003ED59D, 0x003E2DA0003ED59F,
-    0x003E3DA2003E35A1, 0x003E2DA4003E35A3, 0x003E51A6003E55A5,
-    0x003E71A8003E51A7, 0x003E71AA003E71A9, 0x003E51AC003E71AB,
-    0x003E51AE003E51AD, 0x003E31B0003E51AF, 0x003E35B2003E3DB1,
-    0x003F95B4003F8DB3, 0x003F95B6003F9DB5, 0x003FB5B8003F8DB7,
-    0x003FB5BA003FBDB9, 0x003F95BC003F8DBB, 0x003F95BE003F9DBD,
-    0x003E7580003F8DBF, 0x003E7582003E7D81, 0x0003198400030983,
-    0x0003098600031985, 0x0003398800030987, 0x0003098A00033989,
-    0x0003198C0003098B, 0x0003098E0003198D, 0x000379900003098F,
-    0x0003099200037991, 0x0003199400030993, 0x0003099600031995,
-    0x0003399800030997, 0x0003099A00033999, 0x0003199C0003099B,
-    0x0003099E0003199D, 0x0003F9A00003099F, 0x000309A20003F9A1,
-    0x000319A4000309A3, 0x000309A6000319A5, 0x000339A8000309A7,
-    0x000309AA000339A9, 0x000319AC000309AB, 0x000309AE000319AD,
-    0x000379B0000309AF, 0x000309B2000379B1, 0x000319B4000309B3,
-    0x000309B6000319B5, 0x000339B8000309B7, 0x000309BA000339B9,
-    0x000319BC000309BB, 0x000309BE000319BD, 0x000B0891000309BF,
-    0x80000810800006CB, 0x800006D780000813, 0x80000823800006D3,
-    0x8000086980000706, 0x800008788000070B, 0x8000073680000733,
-    0x800008B8800008AC, 0x800008BE800008BB, 0x800008DA800008C2,
-    0x800008F2800008E0, 0x8000092480000778, 0x800007838000077B,
-    0x8000079F80000956, 0x800007BF800007CD, 0x800004DB800006C3,
-    0x800005CD8000057A, 0x8000062E80000608, 0x8000066280000645,
-    0x8000069080000679, 0x800003EF800003D6, 0x8000042180000408,
-    0x8000044780000434, 0x8000046D8000045A, 0x8000049380000480,
-    0x80000508800004F9, 0x8000052680000517, 0x8000054480000535,
-    0x800005568000054D, 0x800005688000055F, 0x8000059D80000594,
-    0x80000F238000085F, 0x003E618D80000F29, 0x003E618F003E598C,
-    0x003E4189003E698E, 0x003E418B003E5988, 0x003E1D85003E498A,
-    0x003E6D87003E1D84, 0x003E1D81003E6D86, 0x003E0D83003E1D80,
-    0x003EE19D003E0D82, 0x003EE19F003ED99C, 0x003EC199003EE99E,
-    0x003EC19B003ED998, 0x003EE195003EC99A, 0x003EE197003E1994,
-    0x003E0191003EE996, 0x003E0193003E1990, 0x003E61AD003E0992,
-    0x003E61AF003E59AC, 0x003E41A9003E69AE, 0x003E41AB003E59A8,
-    0x003E61A5003E49AA, 0x003E61A7003E19A4, 0x003E01A1003E69A6,
-    0x003E01A3003E19A0, 0x003DE1BD003E09A2, 0x003DE1BF003DD9BC,
-    0x003DC1B9003DE9BE, 0x003DC1BB003DD9B8, 0x003DE1B5003DC9BA,
-    0x003DE1B7003E19B4, 0x003E01B1003DE9B6, 0x003E01B3003E19B0,
-    0x80003B4C003E09B2, 0x0051ADB280000AB6, 0x0051D5B40051E5B3,
-    0x80003B4880000B1F, 0x000025B880003B54, 0x80000D11000031B9,
-    0x0000299C80003B58, 0x0000119E0000299D, 0x80000D680000119F,
-    0x800031F780003880, 0x80003360800038A8, 0x80002DBF80002BC4,
-    0x80002E2380002EF2, 0x800030DD80002EEE, 0x80003668800034D8,
-    0x800037DC800037A0, 0x00E5C0B880003810, 0x00E388BA00E000B9,
-    0x00EEACBC00E274BB, 0x00EC98BE00EF94BD, 0x001C00B000E148BF,
-    0x001DA8B2001FA4B1, 0x0018B0B4001AACB3, 0x00E400B6001974B5,
-    0x000AC0A800E7BCB7, 0x001400AA0009D0A9, 0x0015D0AC0017CCAB,
-    0x001208AE001000AD, 0x80003924001190AF, 0x8000327C0000A5A9,
-    0x0000F8A480003308, 0x000FF8A60001D4A5, 0x800033B4000800A7,
-    0x800037688000344C, 0x80003A7C80003950, 0x80002D9380002BE8,
-    0x005C109D80003290, 0x0003E89F800033D8, 0x80003ACC8000379C,
-    0x8000338080002F9A, 0x8000378C0028E49A, 0x8000311180003820,
-    0x800037E8800036C4, 0x8000320B8000300D, 0x8000342480003320,
-    0x80002C2F80003930, 0x80002D3380002CC7, 0x8000328880003179,
-    0x800034F8800033D4, 0x800037AC80003794, 0x800039F8800038D0,
-    0x80003AC880003AAC, 0x8000357480003534, 0x8000374C8000362C,
-    0x8000385480003984, 0x8000302980002EA7, 0x800036A4800035E4,
-    0x8000353C80003408, 0x800039F0800038A0, 0x80002F7680002EA3,
-    0x0003A0AD80003280, 0x8000336880003334, 0x800036448000361C,
-    0x80002D2F800039B0, 0x00218D82800036B0, 0x00024D8000024983,
-    0x0002498600024581, 0x00024984002EF187, 0x00218D8A00024985,
-    0x002ECD880002458B, 0x00025D8E00027189, 0x00025D8C0002458F,
-    0x00023D920002558D, 0x00023D900002A593, 0x0003519600023591,
-    0x0002B19400035197, 0x0003719A0002B195, 0x000361980003719B,
-    0x0003559E00036199, 0x0003619C0003599F, 0x0003B9A20003619D,
-    0x0003B9A00003C5A3, 0x0003DDA60003B9A1, 0x0003DDA40003C5A7,
-    0x0003FDAA0003D5A5, 0x0003FDA80003C5AB, 0x0003DDAE0003F5A9,
-    0x0003DDAC0003C5AF, 0x0003BDB20003D5AD, 0x0003BDB0000345B3,
-    0x000359B60003B5B1, 0x00258DB4000359B7, 0x000379BA000359B5,
-    0x000379B8000379BB, 0x000359BE000379B9, 0x000359BC000359BF,
-    0x0002CD88000359BD, 0x0002DD8A0002D589, 0x0002CD8C0002D58B,
-    0x0002F18E0002F18D, 0x002F71800002F18F, 0x00029D8200029581,
-    0x00028D8400029583, 0x0002FD860002F585, 0x000F4D980002F587,
-    0x000F5D9A000F5599, 0x0003B99C000F559B, 0x000F799E000F499D,
-    0x00029190000F799F, 0x0002919200029191, 0x000C019400029193,
-    0x000F7D96000F7595, 0x000FC9A8000F7597, 0x000FD9AA000FC9A9,
-    0x000FC9AC000FD9AB, 0x000FF9AE000FC9AD, 0x000F89A0000FF9AF,
-    0x000F99A2000F89A1, 0x000F89A4000F99A3, 0x000FF9A6000F89A5,
-    0x000F45B8000FF9A7, 0x000F45BA000F4DB9, 0x000F45BC000F5DBB,
-    0x000F45BE000F4DBD, 0x000F89B0000F7DBF, 0x000F85B2002071B1,
-    0x000F85B4000F9DB3, 0x000F85B6000F8DB5, 0x80002CCF000F7DB7,
-    0x80003630800035A0, 0x800039B880003764, 0x8000311580003870,
-    0x80002BD080003860, 0x8000308180002F4A, 0x800034B880003438,
-    0x8000354480002D53, 0x8000306580002C3F, 0x800032F480002BB0,
-    0x8000361880003191, 0x80002E9380002DA7, 0x8000376C80003504,
-    0x800032C080003844, 0x80003125800032EC, 0x8000314580003724,
-    0x80002C04800034B4, 0x80002CCB80002C87, 0x800036048000325C,
-    0x80003850800036FC, 0x80002D3B80003964, 0x8000301D80002DEF,
-    0x80003384800031AD, 0x800039A080003548, 0x80003AD880003A84,
-    0x800031EF80003AF0, 0x800038F4800032AC, 0x800030C580002FE9,
-    0x80003171800030E1, 0x800033C080003374, 0x8000359480003470,
-    0x8000369C80003638, 0x0001A580800038E4, 0x0001A18200293581,
-    0x0001A1840001A183, 0x0001A1860001A185, 0x0001E1880001A187,
-    0x0056A58A0001E189, 0x0001FD8C0001E58B, 0x0001ED8E0001E58D,
-    0x00019D900001E58F, 0x0001AD920001A591, 0x0001BD940001A593,
-    0x0001AD960001A595, 0x00019D980001A597, 0x00036D9A00036599,
-    0x00037D9C0003659B, 0x00036D9E0003659D, 0x00039DA00003659F,
-    0x0003A9A20001E1A1, 0x0003B9A40003A9A3, 0x0003A9A60003B9A5,
-    0x000399A80003A9A7, 0x00018DAA000399A9, 0x0003F5AC0003EDAB,
-    0x0003F5AE0003FDAD, 0x000395B00003EDAF, 0x000395B200039DB1,
-    0x0003B5B40003ADB3, 0x0003B5B60003BDB5, 0x000395B80003ADB7,
-    0x000395BA00039DB9, 0x000375BC00036DBB, 0x000375BE00037DBD,
-    0x80003774000D15BF, 0x800039888000390C, 0x80002D1B80002CEF,
-    0x800033B080002E03, 0x80003009800037FC, 0x8000313D80003079,
-    0x800035E0800032B8, 0x80002C148000342C, 0x80002FAE80002E5F,
-    0x8000344880003085, 0x8000367480003464, 0x8000396C80003694,
-    0x800039FC800039EC, 0x80002C3380003A38, 0x8000395880003588,
-    0x8000309D800039D4, 0x80002C5780002BF0, 0x80002F6E80002F4E,
-    0x800033CC8000319D, 0x80003780800034D4, 0x800031DF80003920,
-    0x80002D13800039AC, 0x800032378000322B, 0x8000335080003318,
-    0x800034B080003454, 0x8000361480003530, 0x800030F180003A44,
-    0x80002C47800039BC, 0x8000333880002F92, 0x80003055800038E8,
-    0x8000323F800030A5, 0x80002CF7800039C4, 0x0004FC8280002DDF,
-    0x800031C880002F7E, 0x800032688000321B, 0x80003450800032FC,
-    0x80003660800034C4, 0x8000380C80003800, 0x80002D63800039E0,
-    0x80002DE380003354, 0x80003474800033D0, 0x800039C880003790,
-    0x80003ADC80003AB4, 0x8000333080003233, 0x800035CC800036CC,
-    0x80003420800035F0, 0x80002CD3800033A8, 0x80002C0C8000386C,
-    0x80002CFB80003734, 0x80002FF980002CEB, 0x800035FC80003119,
-    0x8000330C80002F3A, 0x800038F0800031EB, 0x80003015800039B4,
-    0x80002E3B80002C67, 0x800031DB80002E8F, 0x8000342880002CD7,
-    0x80003550800034EC, 0x800035708000356C, 0x80003A0480003584,
-    0x80003798800035F4, 0x800039108000385C, 0x80003A588000393C,
-    0x80003A6080003A5C, 0x0007858080003AC4, 0x0007858200078D81,
-    0x0007858400079D83, 0x0007858600078D85, 0x002F89880007BD87,
-    0x002F958A00078189, 0x00079D8C0007858B, 0x800039340007858D,
-    0x0007F590800039D0, 0x0007F19280002C37, 0x80002C6F80002C5B,
-    0x000799960007E995, 0x0007A99800079997, 0x0007B99A0007A999,
-    0x0007A99C0007B99B, 0x002F299E0007A99D, 0x000761A080002D2B,
-    0x00076DA280002D37, 0x80002E2B80002D77, 0x000775A600077DA5,
-    0x80002E4F80002E47, 0x000759AA80002E8B, 0x000749AC000759AB,
-    0x000081AE000749AD, 0x000085B0000081AF, 0x000099B2000089B1,
-    0x000055B40000B1B3, 0x000049B6000059B5, 0x000041B8000041B7,
-    0x0001E9BA000041B9, 0x002CA5BC0001E9BB, 0x0001F5BE0001FDBD,
-    0x80002E970001EDBF, 0x8000309180002F7A, 0x800030C1800030B9,
-    0x800030D900E6C9BA, 0x800031B580003185, 0x80003260800031E3,
-    0x8000334080003324, 0x800033C480003370, 0x80003458800033E4,
-    0x8000355C80003538, 0x8000356080003558, 0x8000357C80003564,
-    0x800035A480003580, 0x800035D4800035B8, 0x8000364880003640,
-    0x800036848000365C, 0x80003708800036D4, 0x8000381880003770,
-    0x8000386480003830, 0x800038AC80003868, 0x80003908800038B0,
-    0x80003A28800039E4, 0x8000308980003A40, 0x800036EC8000EBD1,
-    0x80002CC380002BB8, 0x80002C2B80002C83, 0x80002C9780002C6B,
-    0x80002D4380002D27, 0x80002E2780002E17, 0x80002E9F80002E43,
-    0x80002EF680002EEA, 0x80002F2280002F12, 0x000581BF80003011,
-    0x000571BD000579BE, 0x00057DBB000579BC, 0x00057DB900057DBA,
-    0x002D05B700057DB8, 0x000571B5000579B6, 0x000561B3000579B4,
-    0x000571B1000579B2, 0x000501AF000579B0, 0x000571AD000579AE,
-    0x000561AB000579AC, 0x0029F1A9000579AA, 0x000545A7000575A8,
-    0x000575A5000545A6, 0x000565A3000575A4, 0x000575A1000565A2,
-    0x0005859F000575A0, 0x0004719D0004799E, 0x0006BD9B0004799C,
-    0x000475990004659A, 0x0004459700047598, 0x0004759500044596,
-    0x0004659300047594, 0x0038CD9100046592, 0x0004098F00047190,
-    0x0004098D0004018E, 0x0004698B0004718C, 0x000469890004618A,
-    0x0004498700047188, 0x0004498500044186, 0x00046D8300260D84,
-    0x00046D8100046D82, 0x00022D8500046D80, 0x00023D8700022D84,
-    0x00020D8100023D86, 0x00023D8300020D80, 0x0002698D00023D82,
-    0x0002798F0039AD8C, 0x00020D890002618E, 0x00027D8B00020D88,
-    0x0002259500027D8A, 0x0007BD9700022594, 0x0039C99100022596,
-    0x0002059300020190, 0x0002E19D00020592, 0x0002ED9F0002199C,
-    0x000201990002ED9E, 0x0002019B00021998, 0x00021DA50002099A,
-    0x000461A700021DA4, 0x00021DA100022DA6, 0x00020DA300021DA0,
-    0x000219AD00020DA2, 0x80003019000211AC, 0x000219A980003049,
-    0x000209AB000211A8, 0x000211B5000211AA, 0x000211B7000209B4,
-    0x000211B1000219B6, 0x000211B3000269B0, 0x000209BD000219B2,
-    0x000219BF000451BC, 0x00022DB9000201BE, 0x00021DBB000365B8,
-    0x8000306900021DBA, 0x800030B180003099, 0x800030B5800030A1,
-    0x8000314D800030F5, 0x8000316180003159, 0x8000320F80003189,
-    0x8000336480003223, 0x8000338C8000335C, 0x800033E88000351C,
-    0x800034688000341C, 0x800034A880003484, 0x800034D0800034CC,
-    0x80003500800034F0, 0x8000351080003514, 0x800035BC80003540,
-    0x80003624800035EC, 0x8000373880003654, 0x800037D480003760,
-    0x800038288000381C, 0x8000384C80003848, 0x8000387480003858,
-    0x8000391C800038EC, 0x8000397480003954, 0x80003A18800039C0,
-    0x80003A9480003A34, 0x000A2CF08000EB68, 0x8000EB958000EB63,
-    0x80002B0880002AE0, 0x8000EC1780002B0C, 0x8000ECBC8000EC44,
-    0x80003B3080003B20, 0x8000085480000849, 0x0001A18080000857,
-    0x0001AD82001C9981, 0x0001BD840001A583, 0x0001AD860001A585,
-    0x001AF988001CB987, 0x001C818A00019D89, 0x001C9D8C0001918B,
-    0x0039DD8E0001F58D, 0x00018D900001F98F, 0x00019D9200019591,
-    0x00018D9400019593, 0x0001BD960001B595, 0x00018D980001B597,
-    0x8000084C00019599, 0x8000095980000850, 0x80000FD980000FE8,
-    0x80000FED80000FDE, 0x8000104280000FE3, 0x80001078800010C1,
-    0x8000105180001026, 0x800010668000105E, 0x800010B980001094,
-    0x800010B4800010AF, 0x800010A8800010A1, 0x80000FFA80000FF5,
-    0x8000100C80000FFF, 0x8000102180001019, 0x8000103380001029,
-    0x8000103D80001038, 0x8000104C80001047, 0x8000106180001054,
-    0x8000106E80001069, 0x001F79B280001073, 0x001AFDB0000285B3,
-    0x00029DB6000271B1, 0x00029DB4000285B7, 0x0019C5BA000295B5,
-    0x001FA5B80002BDBB, 0x0002ADBE0002B9B9, 0x001FA9BC001FA1BF,
-    0x00020DA2001A21BD, 0x001989A0000215A3, 0x001995A6000205A1,
-    0x00020DA4000209A7, 0x000209AA000205A5, 0x000209A8000209AB,
-    0x000205AE000209A9, 0x000209AC00020DAF, 0x0017D992001F15AD,
-    0x0002AD900002AD93, 0x001AF99600193591, 0x0002B59400182197,
-    0x001F959A00326595, 0x0002999800029D9B, 0x0002F19E00029999,
-    0x001F819C0002F19F, 0x001F3D820002819D, 0x001F3D80001D6D83,
-    0x0002D586001F3581, 0x001D75840002DD87, 0x0002F18A0002DD85,
-    0x0002E588001FFD8B, 0x0002ED8E001FFD89, 0x0002ED8C0002D58F,
-    0x0003F5800002E58D, 0x0003F5820003FD81, 0x00027D8400026583,
-    0x0002698600026585, 0x800010808000107B, 0x8000108F8000108A,
-    0x8000109C80001097, 0x8000102E800010BC, 0x8000105980001011,
-    0x8000100480001085, 0x0002299400023993, 0x0002D99600022995,
-    0x800017D40002D997, 0x800017E0800017DD, 0x800017DA800017E6,
-    0x800017E30002FD9D, 0x000211A00002F19F, 0x00467DA2004665A1,
-    0x000209A4000219A3, 0x000239A6000209A5, 0x000209A8000239A7,
-    0x000205AA004625A9, 0x000205AC00021DAB, 0x000205AE00020DAD,
-    0x000205B000027DAF, 0x000205B200020DB1, 0x000205B400021DB3,
-    0x000205B600020DB5, 0x000201B8800017D7, 0x000201BA000201B9,
-    0x000201BC000201BB, 0x00020DBE8000180A, 0x800017EC8000180D,
-    0x800017EF800017E9, 0x025179B0800017F2, 0x800017FB02517DB1,
-    0x800017FE800017F8, 0x80001804800017F5, 0x8000181080001801,
-    0x8000181C80001816, 0x025139A480001819, 0x0001EDA602513DA5,
-    0x000155A00001E9A7, 0x00015DA2000151A1, 0x000169AC000159A3,
-    0x0001D1AE00016DAD, 0x0001D5A80001D5AF, 0x000171AA0001D1A9,
-    0x0000B594000175AB, 0x0000B1960000B195, 0x0000A1900000B597,
-    0x0000AD920000A591, 0x00009D9C0000A993, 0x0001A99E0000999D,
-    0x000089980001AD9F, 0x0000859A00008D99, 0x0000DD840000819B,
-    0x0000E9860000D985, 0x0000DD800000ED87, 0x0000C1820000D981,
-    0x0000CD8C0000C583, 0x0000D98E0000C98D, 0x0000D5880000DD8F,
-    0x0000D18A0000D189, 0x8000181F0000D58B, 0x8000182580001822,
-    0x8000183780001813, 0x8000183F80001834, 0x8000183180001845,
-    0x8000184B80001842, 0x800010D880001770, 0x8000110580001128,
-    0x8000111480001119, 0x800011238000111E, 0x8000110A00FEADBA,
-    0x800010DD80001848, 0x800010F6800010E2, 0x8000114F8000110F,
-    0x8000116080001154, 0x8000118580001176, 0x800011958000118A,
-    0x800011C2800011AF, 0x80001218800011E6, 0x800012258000121D,
-    0x8000124380001234, 0x800012488000182B, 0x8000127380001250,
-    0x8000184E8000129A, 0x800012C6800012A6, 0x800012D0800012CB,
-    0x8000131D8000130A, 0x8000134180001329, 0x80001411800013BB,
-    0x8000142F8000142A, 0x8000145380001442, 0x8000146A80001465,
-    0x000201B580001491, 0x00020DB7000205B4, 0x0005EDB1000209B6,
-    0x000219B30005E9B0, 0x000229BD00021DB2, 0x000225BF00022DBC,
-    0x000235B9000221BE, 0x000231BB000231B8, 0x000539A5000235BA,
-    0x000595A700053DA4, 0x000521A1000591A6, 0x000521A3000525A0,
-    0x000595AD000525A2, 0x000595AF000591AC, 0x0005ADA9000591AE,
-    0x00058DAB0005A9A8, 0x0005E595000589AA, 0x0005E1970005E194,
-    0x0005F1910005E596, 0x0005FD930005F590, 0x0005CD9D0005F992,
-    0x0005D99F0005C99C, 0x0005D9990005DD9E, 0x0005D59B0005DD98,
-    0x0251F0D90005D19A, 0x0005E1870005E584, 0x00072D810005E986,
-    0x0005C58300072D80, 0x0005A18D0251E0D9, 0x0005A18F0005D98C,
-    0x0005C1890005A98E, 0x0005C18B0005D988, 0x00020D930005C98A,
-    0x800014AE80001499, 0x00020197800014D5, 0x0002159500021996,
-    0x0002399B00021194, 0x00023D990002319A, 0x00022D9F00023D98,
-    0x0252699D0002299E, 0x800014DA0002259C, 0x8000153980001534,
-    0x8000154A8000153E, 0x8000155B80001556, 0x8000156F80001563,
-    0x8000158E80001589, 0x8000159B80001596, 0x800015A5800015A0,
-    0x0002C1B3800015AA, 0x0002B5B10002C5B2, 0x0002D9B70002B1B0,
-    0x0002DDB50002DDB6, 0x0002F1BB0002D9B4, 0x0002F9B90002F5BA,
-    0x0002E5BF0002E5B8, 0x0002EDBD0002E1BE, 0x0002E5A30002E9BC,
-    0x0002E9A10002E1A2, 0x0002F9A70002EDA0, 0x0002F1A50002FDA6,
-    0x0002D1AB0002F5A4, 0x0002C5A90002D5AA, 0x0002C9AF0002C1A8,
-    0x0002CDAD0002CDAE, 0x800015AF0002C9AC, 0x800015C7800015C2,
-    0x800015FD800015E3, 0x8000162380001617, 0x800016408000163B,
-    0x8000166E8000164D, 0x800016AB80001691, 0x800016BC800016B7,
-    0x800016E5800016C4, 0x8000170E800016FF, 0x800017308000172B,
-    0x8000173D80001738, 0x8000175580001750, 0x800017878000177B,
-    0x800017A280001793, 0x800017C4800017BF, 0x800012F6800012E5,
-    0x8000008080001773, 0x800000948000008A, 0x800000A88000009E,
-    0x00FDA891800000B2, 0x800010F1800010EC, 0x8000116C800010FB,
-    0x8000117B80001171, 0x800011E1800011DC, 0x8000122A8000120E,
-    0x800012398000122F, 0x8000170480001648, 0x8000172180001709,
-    0x8000179D80001798, 0x800010E7800017B5, 0x8000118080001100,
-    0x0001118080001213, 0x0001118200011181, 0x0001318400011183,
-    0x0001318600013185, 0x0001118800013187, 0x0001118A00011189,
-    0x0001718C0001118B, 0x0001718E0001718D, 0x000111900001718F,
-    0x0001119200011191, 0x0001319400011193, 0x0001319600013195,
-    0x0001119800013197, 0x0001119A00011199, 0x0001F19C0001119B,
-    0x0001F19E0001F19D, 0x000111A00001F19F, 0x000111A2000111A1,
-    0x000135A4000111A3, 0x000CF9A600013DA5, 0x000CF1A8000131A7,
-    0x000119AA000CC1A9, 0x000CC5AC000119AB, 0x000CC1AE00010DAD,
-    0x00017DB0000CC1AF, 0x000C45B2000105B1, 0x000C41B4000109B3,
-    0x000111B6000C41B5, 0x000C7DB8000111B7, 0x000C49BA000135B9,
-    0x000EA5BC000C45BB, 0x000E9DBE000EADBD, 0x000FFDA8000E95BF,
-    0x000D01AA000D01A9, 0x000D51AC000D01AB, 0x000D51AE000D51AD,
-    0x000FFDA0000D51AF, 0x000FCDA2000FC5A1, 0x000FDDA4000FC5A3,
-    0x000FCDA6000FC5A5, 0x000D31B8000FC5A7, 0x000D31BA000D31B9,
-    0x000251BC000D31BB, 0x000251BE000251BD, 0x000D31B0000251BF,
-    0x000D31B2000D31B1, 0x000D11B4000D31B3, 0x000D11B6000D11B5,
-    0x000FF588000D11B7, 0x000FF58A000FFD89, 0x000F158C000F0D8B,
-    0x000F158E000F1D8D, 0x000FF980000F0D8F, 0x000FC982000FF981,
-    0x000FD584000FFD83, 0x000FD586000FDD85, 0x000F7D98000FCD87,
-    0x000F0D9A000F0599, 0x000F1D9C000F059B, 0x000F0D9E000F059D,
-    0x000F7590000F059F, 0x000F7592000F7D91, 0x000F5994000F4D93,
-    0x000F4996000F5995, 0x000C6D80000F4997, 0x00030982000C7981,
-    0x0003198400030983, 0x0003058600030D85, 0x0003F58800033D87,
-    0x0003F18A00031589, 0x0003798C0003F18B, 0x0003758E0003D58D,
-    0x0003C59000036D8F, 0x0003A1920003A191, 0x0003BD9400033D93,
-    0x0003219600032195, 0x000E3D98000E0597, 0x000E099A00039599,
-    0x000E199C0003999B, 0x000E099E000E199D, 0x000379A0000E099F,
-    0x000E0DA2000E05A1, 0x000E1DA4000E05A3, 0x000E05A6000369A5,
-    0x000E11A8000E31A7, 0x000E15AA000E11A9, 0x000E79AC000E6DAB,
-    0x000E69AE000E79AD, 0x000E19B0000E69AF, 0x000401B2000E19B1,
-    0x000E35B4000E2DB3, 0x000E35B6000E3DB5, 0x000E15B8000E2DB7,
-    0x000CEDBA000E19B9, 0x000CFDBC000CE5BB, 0x000CBDBE000CB5BD,
-    0x000B4D80000CB5BF, 0x000B5582000B5581, 0x000B6184000B6D83,
-    0x000B6186000B6185, 0x00042188000B6187, 0x0004398A00042989,
-    0x0004298C0004398B, 0x00041D8E0001358D, 0x00046D900004158F,
-    0x0004759200047591, 0x0004159400040D93, 0x00015D9600041D95,
-    0x0004399800040997, 0x00040D9A00016D99, 0x00041D9C0004059B,
-    0x0001759E0004059D, 0x000475A000091D9F, 0x000471A20009F1A1,
-    0x00046DA4000901A3, 0x000189A6000909A5, 0x0001B5A8000945A7,
-    0x0001B5AA0001BDA9, 0x0004B1AC000BD1AB, 0x0004F1AE0004ADAD,
-    0x0004B1B00001B9AF, 0x0001E9B20001C9B1, 0x0001F9B40001E9B3,
-    0x0001E9B60001F9B5, 0x0001D9B80001E9B7, 0x0001A9BA0001D9B9,
-    0x0001B9BC0001A9BB, 0x0001A9BE0001B9BD, 0x800016360001A9BF,
-    0x8000175A80001726, 0x8000123E800017BA, 0x800013968000135B,
-    0x80001508800013A9, 0x8000152080001514, 0x8000148980001484,
-    0x800014CD800014C8, 0x800014F9800014F4, 0x8000136580001360,
-    0x800013B3800013AE, 0x800012BE800012B9, 0x800012928000128D,
-    0x800012DA800012D5, 0x800014228000141D, 0x8000145D80001458,
-    0x800013798000136D, 0x800013918000138C, 0x800013D38000133C,
-    0x800011428000144E, 0x800011B48000119A, 0x800011C7800011BB,
-    0x800011F2800011EB, 0x80001278800011F9, 0x00F42C91800012B2,
-    0x80001322800012AB, 0x800013168000130F, 0x800013468000134D,
-    0x800013C080001354, 0x8000137E80001416, 0x8000139B80001372,
-    0x80001434800013A2, 0x0003598080001447, 0x0003698200035981,
-    0x0003798400036983, 0x0003698600037985, 0x0003598800036987,
-    0x0003A98A00035989, 0x0003298C0003A98B, 0x0003B58E0003BD8D,
-    0x0003D5900003AD8F, 0x0003799200036991, 0x0003699400037993,
-    0x0003199600036995, 0x0003299800031997, 0x0003399A00032999,
-    0x0003299C0003399B, 0x0003199E0003299D, 0x0003E9A00003199F,
-    0x0003F9A20003E9A1, 0x0003E9A40003F9A3, 0x000319A60003E9A5,
-    0x0003B9A8000319A7, 0x000325AA00032DA9, 0x000325AC00033DAB,
-    0x000339AE0003C9AD, 0x0003D9B0000339AF, 0x0003D5B20003D1B1,
-    0x000B1DB4000B35B3, 0x000B0DB6000B05B5, 0x000361B80003A1B7,
-    0x000B11BA000361B9, 0x000369BC000B11BB, 0x8000146F00036DBD,
-    0x0002FDB080001476, 0x0001F1B20001FDB1, 0x0000D1B40001F1B3,
-    0x0000D5B60000D1B5, 0x0000F1B80000D1B7, 0x0000F1BA0000F1B9,
-    0x0000D1BC0000F1BB, 0x0000DDBE0000D5BD, 0x000295A00000D5BF,
-    0x000295A200029DA1, 0x0002B1A4000291A3, 0x0002BDA60002B5A5,
-    0x000289A80002B5A7, 0x000299AA000289A9, 0x000285AC00029DAB,
-    0x000281AE000281AD, 0x00021D90000281AF, 0x0002299200022591,
-    0x0002399400022993, 0x0002299600023995, 0x0002199800022997,
-    0x0002159A00021D99, 0x0002799C0002699B, 0x0002699E0002799D,
-    0x8000147D0002699F, 0x800013CE8000149E, 0x800014BA800014B3,
-    0x800014ED800014DF, 0x80001543800014E6, 0x8000157B80001574,
-    0x8000161080001602, 0x800015E880001609, 0x800016288000161C,
-    0x8000167A80001673, 0x800016528000168A, 0x8000169680001660,
-    0x800016598000169D, 0x8000174980001742, 0x800016F1800016EA,
-    0x800016D7800016D0, 0x800017138000171A, 0x80001165800017A7,
-    0x800011A1800011A8, 0x800011CE800011D5, 0x8000120080001207,
-    0x8000125580001286, 0x8000132E8000127F, 0x80001385800013C7,
-    0x800015F68000143B, 0x8000178C8000162F, 0x800017AE80001780,
-    0x80001582800016B0, 0x800014C1800016F8, 0x800016C9800015BB,
-    0x800015EF800016A4, 0x02A7D4D9800015B4, 0x8000129F8000125C,
-    0x8000154F80001667, 0x8000133580001159, 0x8000140A800016DE,
-    0x8000113980001568, 0x8000168180001130, 0x800012ED800013D8,
-    0x80001762800014A5, 0x0001C580800013E1, 0x0001C5820001CD81,
-    0x00FED5840001B983, 0x0001C98600FCF185, 0x800013E80001C987,
-    0x800012FB80001263, 0x8000240C800003A5, 0x8000069F80002410,
-    0x00FEC4F0800000C6, 0x800024AA800006B4, 0x800022B4800024AE,
-    0x800007E5800022B0, 0x800009818000097D, 0x800024A680002440,
-    0x8000243880002434, 0x800024208000241C, 0x8000242880002424,
-    0x800024308000242C, 0x800007E1800007DD, 0x800000D6800000D0,
-    0x800003A7800003A1, 0x800000D2800007DF, 0x800006BC800000D4,
-    0x800014FE80000078, 0x000119B08000007C, 0x000129B2000119B1,
-    0x000139B4000129B3, 0x000129B6000139B5, 0x000119B8000129B7,
-    0x0000C9BA000119B9, 0x0000D9BC0000C9BB, 0x8000009080000086,
-    0x000289AE80001503, 0x000289AC000289AF, 0x000269AA000289AD,
-    0x000269A8000269AB, 0x000249A6000269A9, 0x000249A4000249A7,
-    0x000269A2000249A5, 0x000269A0000269A3, 0x0002FDBE000269A1,
-    0x0002F9BC000285BF, 0x0002E1BA0002E5BD, 0x0002F1B80002E1BB,
-    0x0002C1B60002F1B9, 0x00FD8DB40002C1B7, 0x0002E9B200FD81B5,
-    0x0002E9B00002E9B3, 0x0002998E0002E9B1, 0x0002998C0002998F,
-    0x0002998A0002998D, 0x000299880002998B, 0x00FD5D8600029989,
-    0x00FD51840002A587, 0x00FD458200FD5185, 0x00FD558000FD4D83,
-    0x0002899E00FD4D81, 0x00028D9C0002899F, 0x0002ED9A0002899D,
-    0x0002ED980002959B, 0x0002D1960002E599, 0x0002C1940002D197,
-    0x0002D1920002C195, 0x8000009A0002D193, 0x00022D908000150F,
-    0x00023D9200023591, 0x0295699400023593, 0x00025D9600022995,
-    0x00026D9800025597, 0x800000A400B62999, 0x800000AE8000151B,
-    0x800000B880001527, 0x800010C68000152C, 0x8000118F8000114C,
-    0x8000122280001192, 0x800012978000124D, 0x800012DF800012C3,
-    0x800012EA800012E2, 0x8000130780001304, 0x800013B88000136A,
-    0x00B685B080001427, 0x0002DDB20002D5B1, 0x0002C9B40002D5B3,
-    0x00B609B600B619B5, 0x0002C1B80002C1B7, 0x0002C1BA0002C1B9,
-    0x0002C1BC0002C1BB, 0x0002C1BE0002C1BD, 0x8000146200B9D9BF,
-    0x800014968000148E, 0x80001531800014D2, 0x8000159380001560,
-    0x800015CC8000112D, 0x800016C180001645, 0x800015CF80001735,
-    0x800015D9800015D4, 0x800006A7800015DE, 0x0007F5BC800006B2,
-    0x0007E9BA8000175F, 0x0007EDB80007E9BB, 0x0007D1B60007E5B9,
-    0x000C15B40007D1B7, 0x000C0DB20007DDB5, 0x00F115B0000C09B3,
-    0x00004DAE000C01B1, 0x0007A5AC00F169AF, 0x0007BDAA000041AD,
-    0x0007B1A80007B9AB, 0x000789A60007B1A9, 0x00078DA400078DA7,
-    0x000795A2000785A5, 0x000799A0000791A3, 0x0007619E000799A1,
-    0x0007519C0007659F, 0x0007499A00076D9D, 0x00074D9800074D9B,
-    0x0007759600074599, 0x0007799400077197, 0x0007619200077995,
-    0x0007759000076593, 0x00070D8E00076D91, 0x0007018C0007098F,
-    0x0007198A0007018D, 0x00071D8800071D8B, 0x0007258600071589,
-    0x0007298400072187, 0x0007318200072985, 0x0007858000073583,
-    0x0006418000073D81, 0x02A0198202A01581, 0x02A0058402A01983,
-    0x02A0298602A02585, 0x02A0398802A00587, 0x02A06D8A02A06189,
-    0x02A0758C02A0698B, 0x0006E58E0006E98D, 0x000605900006798F,
-    0x00060D9200060191, 0x0006199400060593, 0x0006159600061995,
-    0x00062D9800061197, 0x0006399A00063599, 0x0006219C00063D9B,
-    0x00062D9E0006219D, 0x0006D5A00006299F, 0x0006D1A20006DDA1,
-    0x0006C9A40006D5A3, 0x0006C5A60006C9A5, 0x0006FDA80006C1A7,
-    0x0006C9AA0006C5A9, 0x0006D5AC0006C9AB, 0x0006D1AE0006DDAD,
-    0x0006ADB00006D1AF, 0x0006B9B20006B5B1, 0x0006A1B40006BDB3,
-    0x0006ADB60006A1B5, 0x000695B80006A9B7, 0x000691BA00069DB9,
-    0x000689BC000695BB, 0x000685BE000689BD, 0x00079DA1000681BF,
-    0x0007A9A300079DA0, 0x000215A500B38DA2, 0x800000CE000215A4,
-    0x000789A900B395A6, 0x000799AB000781A8, 0x800000D8000781AA,
-    0x800007E3800003B2, 0x0007F9B18000097F, 0x800009830007F1B0,
-    0x800023F60007F1B2, 0x0007A5B70007A1B4, 0x00049DB90007A5B6,
-    0x000545BB00049DB8, 0x000545BD000545BA, 0x000545BF000545BC,
-    0x0007FD81000545BE, 0x00070D8300BCED80, 0x800023FA00070D82,
-    0x0007158700071D84, 0x00B3ED8980002A80, 0x00B3E58B00073588,
-    0x0007E98D00B3E18A, 0x0007E18F00B3F98C, 0x0007F1910007E58E,
-    0x800025340007E990, 0x0007E99500B2B992, 0x0007E5970290AD94,
-    0x00B36D9900079996, 0x0007B59B0007B598, 0x0007959D00B37D9A,
-    0x00076D9F0007959C, 0x00F0A4820007959E, 0x8000258C8000256D,
-    0x800025C38000259E, 0x800029DE800029C2, 0x80002787800029F0,
-    0x80002A6880002A84, 0x0285C4CC80002528, 0x8000252C0285F8CC,
-    0x0000E1A580000985, 0x800009880000D1A4, 0x000019A10000E5A6,
-    0x029329A3000069A0, 0x0000B1AD0000CDA2, 0x02A010D90000D1AC,
-    0x000801A90000B9AE, 0x000B79AB0000D9A8, 0x00D739950000CDAA,
-    0x00D7419700D73194, 0x8000099100D73196, 0x00D729938000098E,
-    0x8000098B00D71192, 0x800022E700D7799C, 0x800022EF800023EA,
-    0x00D7619B800022F7, 0x00D7998500D7719A, 0x00D7A98700D78D84,
-    0x800023EE00D79186, 0x00D7E583800023F2, 0x00D7918D00D7E582,
-    0x00D7618F00D7898C, 0x80000DB700D7958E, 0x00D7918B80000DBA,
-    0x000E358F00D7998A, 0x0009A98D00BC718E, 0x00BD918B0009218C,
-    0x00BD89890009898A, 0x000E118700BD9188, 0x000985850009BD86,
-    0x0009A58300098584, 0x00090581000E2D82, 0x0009059F80001778,
-    0x00BDC19D000C419E, 0x029E299B000C4D9C, 0x00BD259900096D9A,
-    0x00BD159700BD2598, 0x00BD059500BD1596, 0x029C099300BD0594,
-    0x029CF191029C0192, 0x00BCA9AF00BD1990, 0x00BC91AD00BCA5AE,
-    0x00BC99AB00BC9DAC, 0x00BD7DA900BC95AA, 0x00B909A700B941A8,
-    0x00B909A500BD4DA6, 0x00BD15A300B931A4, 0x00B935A100B92DA2,
-    0x000999BF00090DA0, 0x000959BD000EFDBE, 0x00095DBB000969BC,
-    0x00C8F1B900BD01BA, 0x00BC91B700B059B8, 0x00BC81B500B065B6,
-    0x00C8F9B300BCE9B4, 0x00BD3DB100BCCDB2, 0x00BD1D8000BCCDB0,
-    0x00BD7182029D6981, 0x00BD6D8400BC3583, 0x00BC598600BDE985,
-    0x029CE18800BC1D87, 0x029CE18A00BC1589, 0x0293718C00BC258B,
-    0x00BCAD8E00BC258D, 0x00BC5990029DA18F, 0x029C999200B76D91,
-    0x00BC7194029C1193, 0x00BC7D9600BC7D95, 0x02916198029C8197,
-    0x00BC219A029C9599, 0x000F4D9C0008319B, 0x000F319E0008159D,
-    0x000F11A0000F2D9F, 0x0008F5A20008B1A1, 0x0008A1A400080DA3,
-    0x00D505A6000F09A5, 0x000EA5A8000F55A7, 0x000EA5AA000EADA9,
-    0x000EA5AC000EBDAB, 0x000EA5AE000EADAD, 0x000EE5B0000E9DAF,
-    0x00D1A5B200D1ADB1, 0x00D1A1B400D485B3, 0x00D5D9B600D1A1B5,
-    0x00D5E9B800D5D9B7, 0x00D5F9BA00D5E9B9, 0x00D5E9BC00D5F9BB,
-    0x00D5B9BE00D5E9BD, 0x00D6498000D5B9BF, 0x00D6598200D64981,
-    0x00D6498400D65983, 0x00D2658600D26D85, 0x00D2A58800D29D87,
-    0x00D6758A00D2AD89, 0x00D6B58C00D6AD8B, 0x00D6B58E00D6BD8D,
-    0x00D2D19000D2B18F, 0x00D2D19200D2D191, 0x00D2F19400D2D193,
-    0x00D2F19600D2F195, 0x00D2D19800D2F197, 0x00D2D19A00D2D199,
-    0x00D6E19C00D2D19B, 0x000D559E000D5D9D, 0x00DA61A0000D459F,
-    0x00DA89A200D94DA1, 0x00DA95A400C541A3, 0x00DA95A600DA9DA5,
-    0x00DAB5A800DA8DA7, 0x00DAB1AA00DAE5A9, 0x00DA91AC00DAB1AB,
-    0x00DA91AE00DA91AD, 0x00DAF1B000DA91AF, 0x00DA9DB200DA95B1,
-    0x00DAE5B400DAB1B3, 0x00DA5DB600DA55B5, 0x00DA75B800DA4DB7,
-    0x00DA2DBA00DA25B9, 0x00DA25BC00DA3DBB, 0x00D9DDBE00D9D5BD,
-    0x00DC589080000D80, 0x80000CCF00050891, 0x80003B5C80000D95,
-    0x80000D9880000D9B, 0x80000CDB80000CD8, 0x80000CDE80001BB4,
-    0x80000DA780000CEA, 0x0006109E0007209D, 0x80000CFC0015849F,
-    0x80000CF080000CF6, 0x80000D8680000D83, 0x02AC3CD980000D05,
-    0x80000DAA80000D56, 0x80000D1480000DAD, 0x80003B508000EADC,
-    0x8000EAE180000D1A, 0x8000EAE680000D6B, 0x80000D3580000D32,
-    0x8000EAEB80000D3E, 0x80000D4780000D44, 0x80000DA480000D4A,
-    0x80003B6080000D9E, 0x80000D5980000DA1, 0x80000D6E80002404,
-    0x80000D9280000D8F, 0x80000BB280000D7D, 0x80000BB880000BB5,
-    0x8000EAF58000EAF0, 0x8000EA078000EA02, 0x000005BA0002E482,
-    0x0005B8F08000EA0C, 0x000414B20003C4B1, 0x800022F3002EF8AF,
-    0x80000A0200024082, 0x80000A3500020891, 0x00003DA700029084,
-    0x00D7249900052C84, 0x00FF209B8000EA11, 0x00021491000258F0,
-    0x0002688D00029484, 0x00BCCCCC000075A7, 0x8000EA160002D492,
-    0x00024CF00003DC87, 0x00020C9100027496, 0x0002508D000035BE,
-    0x8000EA1B0002C08C, 0x8000EA2000001D97, 0x000240F00003CCB9,
-    0x0002CC920002B491, 0x8000EA258000EA2A, 0x8000EA2F000009B0,
-    0x00026091000248F0, 0x0002B49600029CA4, 0x000005AF00002DBA,
-    0x000254918000EA34, 0x000029BD000240F0, 0x000394B80002CC96,
-    0x8000EA390003B0B9, 0x000E50A5000005AF, 0x000240F00003CCB5,
-    0x8000EA3E0002A491, 0x0002C4A48000EA43, 0x8000EA4C00000DB0,
-    0x8000EA628000EA55, 0x000AF1818000EA6F, 0x000101838000EA7C,
-    0x00061985000AF982, 0x000615870007E984, 0x000635898000EA89,
-    0x0006358B00063588, 0x0006158D0006358A, 0x0006158F0006158C,
-    0x000601910006158E, 0x0006019300067590, 0x00B8ED9500060992,
-    0x00060D9700061994, 0x0006019900060D96, 0x0006019B00063998,
-    0x0006019D0006099A, 0x0006019F0006199C, 0x000601A10006099E,
-    0x00060DA30006F9A0, 0x00B3B9A500019DA2, 0x000609A700061DA4,
-    0x000639A9000611A6, 0x000609AB000631A8, 0x000619AD000631AA,
-    0x000609AF000611AC, 0x8000EA96000611AE, 0x000675B3000671B0,
-    0x000615B5000675B2, 0x000615B7000615B4, 0x000635B9000615B6,
-    0x8000EAB9000635B8, 0x8000EA9F000635BA, 0x8000EAAC8000EAC2,
-    0x00026C850003E580, 0x0001F4870012E590, 0x0002E19F0002C086,
-    0x8000EACF0002F19E, 0x0003E1BB00000086, 0x80000AD700000088,
-    0x0003C9BF0000008A, 0x0003C9BD0003C1BE, 0x001940950003D5BC,
-    0x001B489700180094, 0x001D5091001A4C96, 0x001F5893001C0090,
-    0x002D609D001E5C92, 0x0011C89F002C009C, 0x00157099002E6C9E,
-    0x0017789B00140098, 0x80000CC300167C9A, 0x80000DE6000000A0,
-    0x000000A580000DEC, 0x000000A780000DF2, 0x000000A980000DFE,
-    0x80000E01000000A8, 0x80000E0780000E04, 0x80000E1980000E0A,
-    0x80000E2280000E1F, 0x000000B380000E25, 0x80000E4F8000231B,
-    0x80000E8480000E5E, 0x0003E5A3000000B6, 0x000211A1000219A2,
-    0x80000E87000219A0, 0x000000BF000000BC, 0x80000E960003E1A4,
-    0x80000E99000000C0, 0x8000231380000EA5, 0x000000C780000EBA,
-    0x029814CC80000EBD, 0x00026880029810CD, 0x00BF50CC00133081,
-    0x02981CCC80000EDB, 0x80000EDE029E50CD, 0x80000EF980000EE4,
-    0x80000F0580000EFC, 0x80000F08000000D4, 0x80000F0E80000F0B,
-    0x80000F1780000F14, 0x800022EB80000F1A, 0x80000F20000000DC,
-    0x0013A5B080000F2F, 0x0000C9B20000B1B1, 0x0000D5B40013DDB3,
-    0x0013F9B60013CDB5, 0x0013C9B80013F9B7, 0x0000E1BA0013D5B9,
-    0x0013C9BC0000E1BB, 0x0000E1BE0000E5BD, 0x0029C0BA0000E1BF,
-    0x002EC8B880003B44, 0x80000F44002800B9, 0x80000F6B80000F5F,
-    0x80000F3B80000F59, 0x80000F5C80000F50, 0x80003B40000000FA,
-    0x800017CE80000F62, 0x00A88195000000FE, 0x00A88D9700A46D94,
-    0x00A4299100A47D96, 0x00A41D9300A5D190, 0x00A45D9D00D01592,
-    0x00A5699F00A51D9C, 0x00A5E59900A4D19E, 0x00A4419B00D01D98,
-    0x00A5A9850284519A, 0x00A1958700A1E584, 0x00A1C98100A1E586,
-    0x00A1C58300A1D980, 0x00A4418D00A5F982, 0x00A4598F00A47D8C,
-    0x00A4718900A4458E, 0x00A4698B00A59188, 0x00A119B500A4758A,
-    0x00A115B700A109B4, 0x028891B100A529B6, 0x00A5D1B3028549B0,
-    0x00A4A1BD02856DB2, 0x00A4B9BF00A541BC, 0x00A559B900A4A5BE,
-    0x00A165BB00A115B8, 0x028A81A500A115BA, 0x00A55DA700A5CDA4,
-    0x028511A100A5DDA6, 0x028511A300A5D5A0, 0x00A5C1AD00A5EDA2,
-    0x00A5CDAF0285B9AC, 0x00A5E9A900A5C5AE, 0x028529AB028469A8,
-    0x00A5198000AED5AA, 0x00A5218200A52581, 0x00A4D58400A53D83,
-    0x00D1118600A52D85, 0x00A5498800A51987, 0x00A5598A00A9A589,
-    0x00D1198C00A9A98B, 0x0285558E00A4E18D, 0x00A4799000A5458F,
-    0x00A5B59200A53991, 0x00A4519400A40D93, 0x00A4699680001807,
-    0x00A4699802849597, 0x00A4799A028B2599, 0x0285ED9C00A4F99B,
-    0x00AF519E00A46D9D, 0x0284DDA00284AD9F, 0x00A4A1A200A4A5A1,
-    0x02844DA400A4A9A3, 0x028469A6028995A5, 0x00A02DA800A4D5A7,
-    0x00A40DAA00A03DA9, 0x00A011AC00A031AB, 0x00A011AE00A45DAD,
-    0x00A425B000A061AF, 0x00A5C1B200A5C5B1, 0x00A5C9B400A5DDB3,
-    0x00A5F1B600A5F5B5, 0x00A425B800A5EDB7, 0x00D1E1BA00A5DDB9,
-    0x00A599BC00A5E9BB, 0x00A589BE00A975BD, 0x00D2298000A979BF,
-    0x0286658200A7D181, 0x00A7298400A67583, 0x00A6E58600A66985,
-    0x00A7218800A75D87, 0x00A7198A0287E589, 0x00A7398C0287E58B,
-    0x00A7298E0288758D, 0x0286DD9000A7A98F, 0x00AC619200A75D91,
-    0x02870D9402879D93, 0x00A7719600A77595, 0x0287BD9800A77997,
-    0x0287999A028A6599, 0x00A3FD9C00A7259B, 0x00A7DD9E800017D1,
-    0x800003B900A3E19F, 0x00A321A2800003B6, 0x800004EC800003C9,
-    0x00A691A600A695A5, 0x800005DE8000058B, 0x00A681AA00A685A9,
-    0x00A775AC00A69DAB, 0x00D2B1AE80000619, 0x00A6A9B000A6B9AF,
-    0x00A6B9B200AA45B1, 0x00D2F9B400AA49B3, 0x0286B5B600A701B5,
-    0x00A7D9B800A6A5B7, 0x8000063F00A699B9, 0x8000065600A7ADBB,
-    0x00A7C9BE028735BD, 0x00A60980028735BF, 0x00A6198202894581,
-    0x8000067300A69983, 0x00AD318600A60D85, 0x02867D880286CD87,
-    0x00A6018A00A60589, 0x0286ED8C00A6098B, 0x0286C98E028B358D,
-    0x00A2CD9000A6758F, 0x00A6ED9200A2DD91, 0x00A2F19400A2D193,
-    0x00A2F19600A6BD95, 0x00A6859800A28197, 0x00A7619A00A76599,
-    0x00A7699C00A77D9B, 0x00A7519E00A7559D, 0x00A645A000A74D9F,
-    0x00D381A200A7BDA1, 0x00A7F9A400A789A3, 0x00A7E9A600AB15A5,
-    0x00D389A800AB19A7, 0x0287C5AA00A671A9, 0x00A689AC00A7D5AB,
-    0x00A745AE00A7C9AD, 0x00A6C1B000A6FDAF, 0x00A6F9B2028605B1,
-    0x00A6D9B4028605B3, 0x00A6C9B6028995B5, 0x02877DB800A649B7,
-    0x00ADC1BA00A6FDB9, 0x0286ADBC02863DBB, 0x00A6D1BE00A6D5BD,
-    0x00A0219100A6D9BF, 0x00A0399300A01D90, 0x8000068A00A02592,
-    0x00A01D9700A1D194, 0x00ACA19900D41596, 0x00ACAD9B00A04D98,
-    0x8000013700A05D9A, 0x00A0419F00D41D9C, 0x028C41810280519E,
-    0x00A1018302819980, 0x00A5C9850281BD82, 0x00A5C58700A5D984,
-    0x00A189898000013B, 0x8000013F00A5C588, 0x00A0718D00A5C58A,
-    0x00A0698F80000143, 0x00A1A1B100A0758E, 0x00A1ADB30281D9B0,
-    0x028C91B500A1A5B2, 0x00A1D1B7028149B4, 0x00A539B902816DB6,
-    0x00A535BB00A529B8, 0x00A159BD80000147, 0x8000014B00A515BC,
-    0x00A0BDA100A515BE, 0x00A189A300A1FDA0, 0x028111A500A031A2,
-    0x028111A700A1D5A4, 0x028EA1A900A1EDA6, 0x00A17DAB00A1EDA8,
-    0x00A1E9AD00A1FDAA, 0x028129AF028069AC, 0x00A0E58000AAD5AE,
-    0x00A1018200A10581, 0x00A1098400A11D83, 0x00A131868000014F,
-    0x8000015780000153, 0x00D5218A8000015B, 0x00A1598C00A1298B,
-    0x00A1498E00ADB58D, 0x00D5699000ADB98F, 0x028125928000015F,
-    0x00A0699400A13593, 0x00A1A59600A12995, 0x00A0619800A01D97,
-    0x00A0599A0280A599, 0x00A0799C0280A59B, 0x00A0699E028F359D,
-    0x02811DA000A0E99F, 0x00ABA1A200A09DA1, 0x0280CDA402805DA3,
-    0x00A0B1A600A0B5A5, 0x02807DA800A0B9A7, 0x028059AA028DA5A9,
-    0x00A43DAC00A0E5AB, 0x00A01DAE00A42DAD, 0x00A461B000A421AF,
-    0x00A461B200A02DB1, 0x00A035B400A411B3, 0x00A1D1B600A1D5B5,
-    0x00A1F9B800A1CDB7, 0x00A1C1BA00A1C5B9, 0x00A035BC00A1DDBB,
-    0x00D5F1BE00A1CDBD, 0x00A2698000A1F9BF, 0x00A2798200AE8581,
-    0x00D6398400AE8983, 0x0282758600A3C185, 0x00A3198800A26587,
-    0x00A2D58A00A25989, 0x00A3318C00A36D8B, 0x00A3098E0283F58D,
-    0x00A349900283F58F, 0x00A35992028C0591, 0x0282CD9400A3D993,
-    0x00A8719600A34D95, 0x02833D9802838D97, 0x00A3419A00A34599,
-    0x0283AD9C00A3499B, 0x0283899E028E759D, 0x00A70DA000A3359F,
-    0x00A32DA200A71DA1, 0x00A731A400A711A3, 0x00A731A600A37DA5,
-    0x00A345A800A741A7, 0x00A2A1AA00A2A5A9, 0x00A2A9AC00A2BDAB,
-    0x00A291AE00A295AD, 0x00A305B000A28DAF, 0x00D6C1B200A2FDB1,
-    0x00A2B9B400A2C9B3, 0x00A2A9B600AE55B5, 0x00D6C9B800AE59B7,
-    0x028285BA00A331B9, 0x00A3C9BC00A295BB, 0x00A205BE00A289BD,
-    0x00A2018000A3BDBF, 0x00A239820282C581, 0x00A219840282C583,
-    0x00A20986028D5585, 0x0283BD8800A28987, 0x00A9018A00A23D89,
-    0x02826D8C0282FD8B, 0x00A2118E00A2158D, 0x02829D9000A2198F,
-    0x0282B992028F4591, 0x00A6DD9400A20593, 0x00A2FD9600A6CD95,
-    0x00A6C19800A6C197, 0x00A6C19A00A28D99, 0x00A2959C00A6B19B,
-    0x00A3719E00A3759D, 0x00A399A000A36D9F, 0x00A3A1A200A3A5A1,
-    0x00A255A400A3BDA3, 0x00D791A600A3ADA5, 0x00A3C9A800A399A7,
-    0x00A3D9AA00AF25A9, 0x00D799AC00AF29AB, 0x0283D5AE00A261AD,
-    0x00A2F9B000A3C5AF, 0x00A335B200A3B9B1, 0x00A2D1B400A28DB3,
-    0x00A2E9B6028215B5, 0x00A2E9B8028215B7, 0x00A2F9BA028DA5B9,
-    0x02836DBC00A279BB, 0x00A9D1BE00A2EDBD, 0x00A0819D02822DBF,
-    0x00A08D9F00AC6D9C, 0x00AC299900AC7D9E, 0x00AC1D9B00ADD198,
-    0x00AC019500D8159A, 0x00AC199700AC3D94, 0x00AC319100AC0596,
-    0x00AC299300ADD190, 0x00ADA98D00AC3592, 0x00A9958F00A9E58C,
-    0x00A9C98900A9E58E, 0x00A9C58B00A9D988, 0x0280618500ADF98A,
-    0x00AD2187028DB984, 0x00AD5181028D9D86, 0x00AD5D83028D2980,
-    0x00A919BD00AD5582, 0x00A915BF00A909BC, 0x028091B900AD29BE,
-    0x00ADD1BB028D49B8, 0x00AD81B5028D6DBA, 0x00AD8DB7028DF9B4,
-    0x00ADA9B100AD85B6, 0x028D69B3028C29B0, 0x028281AD00A695B2,
-    0x00AD5DAF00ADCDAC, 0x028D11A900ADDDAE, 0x028D11AB00ADD5A8,
-    0x00AC9DA500ADEDAA, 0x00ADA9A700ADDDA4, 0x00AD25A100AC11A6,
-    0x00AC81A300D8DDA0, 0x00A8A180028C91A2, 0x00A8A18200ACED81,
-    0x00ACF58400A8D183, 0x00AD118600AD1585, 0x00AD398800AD0D87,
-    0x00AD018A00AD0589, 0x00ACF58C00AD1D8B, 0x00D9318E00AD0D8D,
-    0x00AD299000AD398F, 0x00AD399200A1C591, 0x00D9799400A1C993,
-    0x028D359600AC8195, 0x00AC599800AD2597, 0x00AD959A00AD1999,
-    0x00AC719C00AC2D9B, 0x00AC499E028CB59D, 0x00AC89A0028CB59F,
-    0x00AC99A20283C5A1, 0x028D0DA400AC19A3, 0x00A7B1A600AC8DA5,
-    0x028CFDA8028C4DA7, 0x00AC81AA00AC85A9, 0x028C6DAC00AC89AB,
-    0x028C49AE0281B5AD, 0x00A84DB000ACF5AF, 0x00AC6DB200A85DB1,
-    0x00A871B400A851B3, 0x00A871B600AC3DB5, 0x00AC05B800A801B7,
-    0x00ADE1BA00ADE5B9, 0x00ADE9BC00ADFDBB, 0x00ADD1BE00ADD5BD,
-    0x00AFC58000ADCDBF, 0x00DA018200AE3D81, 0x00AE798400AE0983,
-    0x00AE698600A29585, 0x00DA098800A29987, 0x028E458A00AFF189,
-    0x00AF098C00AE558B, 0x00AEC58E00AE498D, 0x00AF419000AF7D8F,
-    0x00AF7992028F8591, 0x00AF5994028F8593, 0x00AF499602801595,
-    0x028EFD9800AFC997, 0x00A4419A00AF7D99, 0x028F2D9C028FBD9B,
-    0x00AF519E00AF559D, 0x028F5DA000AF599F, 0x028F79A2028285A1,
-    0x0014B9A400AFC5A3, 0x80000163001415A5, 0x000401A880000167,
-    0x00AF55AA00142DA9, 0x0006B1AC001421AB, 0x0006B1AE001421AD,
-    0x0006D1B002BB79AF, 0x001461B2001459B1, 0x001461B4001461B3,
-    0x0006C1B6001461B5, 0x0006C1B800AC9DB7, 0x02BB25BA02BB21B9,
-    0x02BB5DBC001455BB, 0x001439BE001449BD, 0x00AED980001439BF,
-    0x000721820015C981, 0x00A9A58402BAA983, 0x02BAD18600A9D185,
-    0x000711880015CD87, 0x00AE798A02BA8D89, 0x0015E58C02BA958B,
-    0x0015E18E028FB58D, 0x000779900015CD8F, 0x02BAF59202BAF191,
-    0x0014519402BA8D93, 0x0007199600145D95, 0x00A9A99802BAED97,
-    0x0007219A00146999, 0x02BAB99C00147D9B, 0x02BAC19E02BADD9D,
-    0x02BA3DA002BACD9F, 0x000529A202BA31A1, 0x00AE6DA4001505A3,
-    0x000799A6001519A5, 0x0007A9A8001509A7, 0x0007B9AA02BA61A9,
-    0x001519AC001531AB, 0x001509AE001519AD, 0x0007D9B0001509AF,
-    0x0007E9B200AD85B1, 0x02BA1DB402BA09B3, 0x02BA75B600156DB5,
-    0x001521B8001561B7, 0x00AE31BA001521B9, 0x0007D9BC001521BB,
-    0x00A94DBE02BA51BD, 0x80000DEF00A939BF, 0x80000DF5003984AF,
-    0x80000DFB8000016B, 0x80000E0D8000016F, 0x80000E1300000006,
-    0x80000E4080000ED8, 0x80000E5280000173, 0x80000E8A80000177,
-    0x80000E9C8000017B, 0x80000EA88000017F, 0x80000EE180000183,
-    0x80000EE780000187, 0x80000EEA8000018B, 0x80000EF080000EED,
-    0x80000EF68000018F, 0x80000F0280000193, 0x80000F1D80000F11,
-    0x80000F2C80000F26, 0x80000F3E80000F80, 0x0034008880000FA7,
-    0x0010E48600017489, 0x0038A88480000197, 0x8000019B0039AC85,
-    0x800006D080002444, 0x8000070E800007CA, 0x8000079C80000796,
-    0x800007C78000077F, 0x800007518000074E, 0x800006E08000075C,
-    0x800026E18000250E, 0x80002DB380002F2E, 0x8000383880002ECF,
-    0x00A871A880002C00, 0x00A871AA00A871A9, 0x00A875AC00A871AB,
-    0x02896DAE02899DAD, 0x00A859A0028965AF, 0x00A84DA200A859A1,
-    0x028945A40289A5A3, 0x00A851A602894DA5, 0x028935B800A851A7,
-    0x00A821BA02893DB9, 0x00A821BC00A821BB, 0x00A821BE00A821BD,
-    0x00A809B000A821BF, 0x00A819B200A809B1, 0x00A809B400A819B3,
-    0x00A81DB600A809B5, 0x02BC3D880289F5B7, 0x00129D8A02BC3189,
-    0x0288F4CC0012918B, 0x00A8FD8E0287C0CD, 0x00AF69800289158F,
-    0x0001E1820012A981, 0x02BC79840012BD83, 0x02BC018602BC1D85,
-    0x00A8A59802BC0D87, 0x0289BD9A02894D99, 0x00A8B99C0289B59B,
-    0x00A8A99E00A8B99D, 0x0289959000A8A99F, 0x00A8819202899D91,
-    0x00A8819400A88193, 0x00A8819600A88195, 0x0002858700A88197,
-    0x0002898500117D86, 0x0011798300116184, 0x0002658100117182,
-    0x0011F58F00117D80, 0x0011F98D0002B18E, 0x0002BD8B0002BD8C,
-    0x001145890011F18A, 0x000221970002B588, 0x00110D9500021D96,
-    0x0011119300021D94, 0x00027D9100116192, 0x00021D9F00118590,
-    0x0011399D0011399E, 0x0011399B0002159C, 0x000211990002059A,
-    0x00115DA700021998, 0x00FC71A5000201A6, 0x000211A30011F5A4,
-    0x0011F1A10011F9A2, 0x80000DE30011F9A0, 0x00115DAD800031CC,
-    0x001159AB00115DAC, 0x001159A9001151AA, 0x800033B800022DA8,
-    0x8000305980002D07, 0x800031A580002C9F, 0x8000362080002CF3,
-    0x80002EB380003894, 0x8000336C80002DEB, 0x8000313580003109,
-    0x8000312980003914, 0x028008CC80003101, 0x028280CC0280E4CD,
-    0x02875CCC028298CD, 0x028170CC02BE8CCC, 0x0284A0CC0281BCCD,
-    0x02872CCC0284FCCD, 0x028508CC028714CD, 0x028438CC028580CD,
-    0x02B95CCC028454CD, 0x02B928CC02B958CD, 0x02B948CC02B9D0CC,
-    0x02B95CCC02BE80CC, 0x02BEF4CC02B97CCD, 0x02B8C0CC02BEF0CD,
-    0x028298CC02BEE0CD, 0x02BC70CC02BDC4CD, 0x8000357802BCFCCD,
-    0x80002DD7800035B4, 0x800034A08000334C, 0x80002E3780002D0B,
-    0x02B918CC8000394C, 0x003C259002B974CC, 0x02B9F4CC003C3991,
-    0x02B9E4CC02B9E4CC, 0x8000248802B944CC, 0x003C49888000244C,
-    0x8000246A80002456, 0x8000247480002492, 0x00002D828000249C,
-    0x0000298000003583, 0x00002D8600002981, 0x00002D84003CE587,
-    0x0281B19900002585, 0x0281BD9B0281BD98, 0x002E999D0281B19A,
-    0x002E959F003EB59C, 0x028191910095E19E, 0x002F2593003C6590,
-    0x002F21950092E592, 0x02819197003C7D94, 0x002EB1890281F596,
-    0x0281FD8B002E9D88, 0x0281918D003C758A, 0x002F518F0281E98C,
-    0x002EB181002F5D8E, 0x0281C1830281AD80, 0x0281C985003C5D82,
-    0x02B4F98700952584, 0x009245B9002EA986, 0x002E59BB009231B8,
-    0x028139BD028145BA, 0x028121BF003CA5BC, 0x002E1DB10095CDBE,
-    0x002E6DB3028109B0, 0x002E7DB5002E6DB2, 0x02811DB700956DB4,
-    0x002E75A9003C95B6, 0x009689AB002E75A8, 0x028115AD003CD5AA,
-    0x002E61AF003CF5AC, 0x002E75A1028111AE, 0x02812DA3003CE5A0,
-    0x002E6DA5003CE5A2, 0x002E55A7003CE5A4, 0x002FD580002E55A6,
-    0x002FD18202B58581, 0x003D2984002FFD83, 0x0280A5860280A185,
-    0x002E21880280DD87, 0x003D698A002E2D89, 0x0093F98C02809D8B,
-    0x003D718E002E398D, 0x02808990002E2D8F, 0x0280F1920280ED91,
-    0x0280ED940280FD93, 0x003FF9960280E195, 0x00949D98002FD597,
-    0x003D699A002FE999, 0x003D799C002FF99B, 0x003D699E0280B19D,
-    0x002F29A0002FE19F, 0x002F39A2002F29A1, 0x003D89A4002F39A3,
-    0x003DB9A60097D5A5, 0x02806DA8028059A7, 0x028005AA002F1DA9,
-    0x002F71AC002F11AB, 0x009461AE002F71AD, 0x003DE9B0002F71AF,
-    0x00937DB2028061B1, 0x028019B4009309B3, 0x003DE9B6002F05B5,
-    0x0094B1B8028075B7, 0x002F3DBA02805DB9, 0x002F29BC02B56DBB,
-    0x003DC1BE002F05BD, 0x0283BD80028049BF, 0x002D09820283C581,
-    0x003E5184002D0583, 0x0090D1860283A585, 0x003E6988002D1187,
-    0x0283E18A002D3589, 0x0283898C0283858B, 0x0283858E0283858D,
-    0x003CE1900283898F, 0x0097B592002CCD91, 0x003E5194002CC193,
-    0x003E5196002CC195, 0x003E719802839997, 0x002CC19A002CF999,
-    0x002CC19C002CC19B, 0x003E619E002CC19D, 0x003EA1A000943D9F,
-    0x028345A2028341A1, 0x02833DA4002C35A3, 0x002C59A6002C29A5,
-    0x009779A8002C59A7, 0x003E81AA002C69A9, 0x009005AC028309AB,
-    0x028371AE009071AD, 0x003EF1B0002C6DAF, 0x009799B202836DB1,
-    0x002C05B4028375B3, 0x002C01B602B655B5, 0x003ED9B8002C2DB7,
-    0x028355BA028351B9, 0x002DF1BC02832DBB, 0x003EB9BE002DFDBD,
-    0x0020DD8002834DBF, 0x00205D8200204181, 0x8000246000206D83,
-    0x0020758600271985, 0x0020E18800204587, 0x0020F18A002B1989,
-    0x0020D58C0020CD8B, 0x0020498E0020DD8D, 0x002091900020F58F,
-    0x00209D9200203D91, 0x0020259400202593, 0x0020099600201995,
-    0x0020059800200197, 0x0020AD9A00201999, 0x002C899C0020A19B,
-    0x00322D9E00D6759D, 0x8000247E0031099F, 0x0020DDA20020C1A1,
-    0x00207DA48000305D, 0x80002BD480002DC7, 0x80002BCC0020C5A7,
-    0x002071AA002B99A9, 0x002055AC00204DAB, 0x0020C9AE00205DAD,
-    0x002011B0002075AF, 0x00201DB20020BDB1, 0x0020A5B480002BDC,
-    0x002089B6002099B5, 0x8000EAFA002081B7, 0x80002C27002099B9,
-    0x80002C3B002021BB, 0x80002C4B80002C43, 0x0027DD8080002C4F,
-    0x00275D8200274181, 0x0027FD8400276D83, 0x0027758600201985,
-    0x0027E18800274587, 0x80002C53002C1989, 0x0027D58C0027CD8B,
-    0x0027498E0027DD8D, 0x002791900027F58F, 0x00279D9200273D91,
-    0x0027259400272593, 0x0027099600271995, 0x0027059800270197,
-    0x0027AD9A00271999, 0x80002A8C0027A19B, 0x80002C738000EB0E,
-    0x80002C9380002C77, 0x0027DDA20027C1A1, 0x8000EAFF0027EDA3,
-    0x0027F5A6002099A5, 0x002761A80027C5A7, 0x80002A90002C99A9,
-    0x002755AC00274DAB, 0x0027C9AE00275DAD, 0x002711B0002775AF,
-    0x00271DB20027BDB1, 0x0027A5B40027A5B3, 0x002789B6002799B5,
-    0x002785B8002781B7, 0x00272DBA002799B9, 0x80002C7F002721BB,
-    0x80002CAF8000EB09, 0x0000004180002CB7, 0x002671B10026D1B2,
-    0x00264DB70026DDB0, 0x002655B5002645B6, 0x0026EDBB002669B4,
-    0x002655B90026E1BA, 0x80002CBF80002C10, 0x8000ECE4003461BE,
-    0x80002CE7002AC5BC, 0x00260DA1002611A2, 0x002609A780002A98,
-    0x80002D0380002CFF, 0x80002D0F0026B1A4, 0x002D55A90026BDAA,
-    0x0026B9AF0026ADA8, 0x002691AD002605AE, 0x00000061002699AC,
-    0x0026F19100265192, 0x0026CD9780002A9C, 0x80002D4B80002D47,
-    0x00266D9B0026E994, 0x0026D59980002D73, 0x0037C59F80002D7F,
-    0x00D0B99D80002D8F, 0x80002D9B80002D97, 0x8000EB1D00269182,
-    0x002689878000339C, 0x80002DB780002DAF, 0x0026018B8000EB22,
-    0x002DD58980002DC3, 0x0026398F80002DCB, 0x0026118D0026858E,
-    0x0034818080002DD3, 0x0034858200348181, 0x00348D8400349D83,
-    0x0034AD860034B585, 0x0034D1880034F587, 0x0032398A00353589,
-    0x80002DE780002DFB, 0x80002DF780002DF3, 0x0032719080002DFF,
-    0x0032219200321991, 0x0032FD9400323993, 0x00334D960032ED95,
-    0x0033419800334197, 0x00332D9A00334199, 0x003CED9C003CF59B,
-    0x003A6D9E003A799D, 0x003AB5A0003A519F, 0x003ACDA2003AB9A1,
-    0x003AC5A4003AC1A3, 0x003AF5A6003AC9A5, 0x003A05A8003AE9A7,
-    0x003A1DAA003A15A9, 0x009289AC009309AB, 0x003A05AE003A0DAD,
-    0x00974DB080002E07, 0x00936DB200975DB1, 0x009771B4009751B3,
-    0x009771B600933DB5, 0x009305B8009701B7, 0x0092E1BA0092E5B9,
-    0x0092E9BC0092FDBB, 0x0092D1BE0092D5BD, 0x0092C5800092CDBF,
-    0x00E7018200933D81, 0x0093798400930983, 0x00936986009F9585,
-    0x00E70988009F9987, 0x003BE98A0092F189, 0x003BF58C00F8758B,
-    0x003BF58E003BFD8D, 0x009225B0003BED8F, 0x02B33DB202B3CDB1,
-    0x009239B402B335B3, 0x009229B6009239B5, 0x009219B8009229B7,
-    0x80002E0B009219B9, 0x80002E1B80002E13, 0x80002E2F80002E1F,
-    0x000C80A180002E33, 0x000B80A3000D84A0, 0x001490A5000800A2,
-    0x001198A7000A94A4, 0x80002E3F001000A6, 0x003B7DAA001324A8,
-    0x003B6DAC003B75AB, 0x80002E6780002E63, 0x80002E5380002E4B,
-    0x80002E7780002EAF, 0x80002E7B000000F2, 0x80002E7380002E7F,
-    0x80002E8380002E87, 0x8000EB2780002E9B, 0x80002EBB80002EB7,
-    0x80002ED780002EC3, 0x000B998080002EFA, 0x000BC582000BA581,
-    0x000BF984000BED83, 0x000BE98600C45185, 0x000BD988000BE987,
-    0x00C4658A000BD989, 0x000C418C000C058B, 0x00C4698E000C418D,
-    0x000C3D90000C458F, 0x000CD992000CC591, 0x000CA194000CE593,
-    0x000CAD9600369995, 0x000C9D98000CA597, 0x000D599A000D4999,
-    0x000D499C00C5AD9B, 0x000D399E0097F59D, 0x000DD5A0000D319F,
-    0x00371DA2000DDDA1, 0x000D1DA4000DEDA3, 0x00C4E5A600C541A5,
-    0x00C579A800C419A7, 0x000D31AA000D31A9, 0x000D15AC00C099AB,
-    0x000D15AE00C439AD, 0x000D7DB0000D09AF, 0x000D09B2000D05B1,
-    0x000D19B4000D0DB3, 0x000D0DB6000D19B5, 0x000D3DB8000D05B7,
-    0x000D05BA000D01B9, 0x000EDDBC000D11BB, 0x000ECDBE000EC5BD,
-    0x000F3D80000EC5BF, 0x000E6182000FB981, 0x000E6184000E6183,
-    0x000E6D86000E6185, 0x000E598800C4FD87, 0x000E898A000E5989,
-    0x00C5AD8C000E998B, 0x000E858E000E8D8D, 0x00C2C590000EFD8F,
-    0x000E8192000E8191, 0x000E8594000E9D93, 0x000E8596000E8D95,
-    0x000E8598000EBD97, 0x000E859A000E8D99, 0x000E859C000E9D9B,
-    0x000E819E0035819D, 0x0003E1A0000E819F, 0x0003E1A20003E1A1,
-    0x000311A40003E1A3, 0x000311A6000311A5, 0x000335A8000311A7,
-    0x00030DAA000305A9, 0x000319AC000309AB, 0x000309AE000319AD,
-    0x000375B000C199AF, 0x000375B200037DB1, 0x0002D9B40002C9B3,
-    0x0002D5B600E99DB5, 0x00E9B5B80002CDB7, 0x0002F1BA0002F1B9,
-    0x0002D1BC0002F1BB, 0x0002D1BE0002D1BD, 0x0034E9A60002D1BF,
-    0x0034E9A40034E9A7, 0x00EBC1A20034E9A5, 0x0034CDA00034C9A3,
-    0x00FCEDAE0034C5A1, 0x00FCE1AC0034DDAF, 0x00F901AA00F949AD,
-    0x0034C9A80034C1AB, 0x00FCC5B6009521A9, 0x0035C1B400EAADB7,
-    0x0035D1B20035C1B5, 0x00EDD9B00035D1B3, 0x8000EB2C0034A1B1,
-    0x80002F068000EB31, 0x0035E9BA80002F0A, 0x0035E9B80035E9BB,
-    0x80002F0E0035E9B9, 0x80002F1680002F02, 0x00F6858280002AA0,
-    0x00367D8080002AA4, 0x80002F1E00367581, 0x8000EB3680002F26,
-    0x80002F4680002F42, 0x8000EB3B80002F52, 0x0036399680002F5A,
-    0x0095859400F97597, 0x0036219200363995, 0x00EA699000E70993,
-    0x00342D9E00363191, 0x00342D9C0034359F, 0x0034319A0034259D,
-    0x0036059800C0799B, 0x000000C000345999, 0x8000303D80002F5E,
-    0x80002F7280002AA8, 0x80002F8A80002F8E, 0x80002F968000EB45,
-    0x80002FA28000EB4A, 0x80002FA680002F9E, 0x80002FBA80002FB6,
-    0x80002FCA80002AAC, 0x80002FD680002FD2, 0x80002AB080002FDA,
-    0x80002AB48000EB4F, 0x8000300180002FFD, 0x8000ED1680003005,
-    0x800036E08000EB59, 0x80002AB880003035, 0x8000EC4E8000EB7C,
-    0x8000304D80003045, 0x8000306180002ABC, 0x8000307580003071,
-    0x8000308D8000307D, 0x80002AC080002AC4, 0x800030958000EB5E,
-    0x800030AD800030A9, 0x800030D1800030BD, 0x800030CD800030C9,
-    0x800030E9800030D5, 0x80003105800030ED, 0x8000311D8000310D,
-    0x8000EB6D80003131, 0x800031218000312D, 0x8000314180003139,
-    0x8000EB7280003155, 0x000000438000315D, 0x8000314980003151,
-    0x8000316580002AC8, 0x800031758000316D, 0x80002ACC80003169,
-    0x8000EB778000318D, 0x800031FB800031B9, 0x80002AD8800031D7,
-    0x80002AD080002AD4, 0x80002CA780002CA3, 0x800031E7800031FF,
-    0x80002B3C800036B4, 0x8000322780003213, 0x8000EB908000321F,
-    0x8000323B80002ADC, 0x800032588000322F, 0x800032648000EB9A,
-    0x8000326C80003243, 0x8000327480003270, 0x8000328480003278,
-    0x8000328C8000EB9F, 0x8000329880002AE4, 0x8000329C8000EBA4,
-    0x800032A880002AE8, 0x800032C4800032B4, 0x8000EB408000EBA9,
-    0x800032E48000EBAE, 0x800032F08000EBB3, 0x800032E8800032F8,
-    0x8000331480003304, 0x800033288000331C, 0x00C009A38000332C,
-    0x0006ADA10006ADA2, 0x0006A9A70006ADA0, 0x0006A9A50006A1A6,
-    0x00C031AB0006B1A4, 0x000691A9000699AA, 0x0006FDAF00DD25A8,
-    0x0006FDAD0006FDAE, 0x000695B30006FDAC, 0x00C921B1000695B2,
-    0x0006B5B700C055B0, 0x000685B50006B5B6, 0x000695BB000685B4,
-    0x000685B9000695BA, 0x000175BF000685B8, 0x000685BD000175BE,
-    0x0006BD83000685BC, 0x0006BD810006BD82, 0x0006DD870006BD80,
-    0x0006DD850006DD86, 0x0006FD8B0006DD84, 0x0006FD890006FD8A,
-    0x0006D98F0006FD88, 0x0006DD8D00C8E18E, 0x0006B1930006DD8C,
-    0x0006A5910006A992, 0x0006AD9700EDB190, 0x0006AD950006AD96,
-    0x00066D9B0006A994, 0x00066D9900066D9A, 0x00066D9F00066D98,
-    0x00066D9D00066D9E, 0x0002158200066D9C, 0x0002058000021D83,
-    0x0002358600021D81, 0x0002058400023D87, 0x0002118A00023D85,
-    0x00C335880002118B, 0x0002718E00020189, 0x0002018C00C3118F,
-    0x00020D920002018D, 0x00020D9000021593, 0x00020D9600020591,
-    0x00020D9400023597, 0x0002059A00020595, 0x00CB319800020D9B,
-    0x0002059E00DE4599, 0x0002159C00020D9F, 0x00C3A9A200020D9D,
-    0x0002F5A0000201A3, 0x000201A600020DA1, 0x000211A400DE41A7,
-    0x00023DAA000211A5, 0x00023DA8000205AB, 0x00021DAE000235A9,
-    0x00021DAC000205AF, 0x00027DB2000215AD, 0x00027DB0000205B3,
-    0x000219B6000275B1, 0x00021DB4000219B7, 0x000239BA00C6B5B5,
-    0x000239B8000239BB, 0x000219BE000239B9, 0x000219BC000219BF,
-    0x0003F180000219BD, 0x00C5BD8200C22981, 0x0003E9840003F983,
-    0x001C01860003E985, 0x001F2588001F1D87, 0x001F258A001F2D89,
-    0x001F218C00DF198B, 0x001F218E001F218D, 0x001F6190001F218F,
-    0x001F6192001F6191, 0x00DF6994001F6193, 0x001F6D96001F6595,
-    0x001F5D98001F6597, 0x001F2D9A001F2599, 0x001F3D9C001F259B,
-    0x001F2D9E001F259D, 0x001FDDA0001F259F, 0x001FE9A200DFC5A1,
-    0x001FF9A4001FE9A3, 0x001FE9A6001FF9A5, 0x001FD9A8001FE9A7,
-    0x001F29AA001FD9A9, 0x001F39AC001F29AB, 0x001F35AE001F3DAD,
-    0x00DF89B0001F2DAF, 0x001F51B2001F51B1, 0x001F71B4001F51B3,
-    0x001F71B6001F71B5, 0x00C7D1B8001F71B7, 0x001F5DBA001F55B9,
-    0x001F4DBC001F55BB, 0x001F3DBE001F35BD, 0x80003310001F35BF,
-    0x8000334880003344, 0x8000335880002AEC, 0x8000333C8000EBBD,
-    0x8000EBC280003378, 0x8000337C8000EBC7, 0x8000338880003390,
-    0x8000339480002AF0, 0x800033A0800033A4, 0x8000EB04800033AC,
-    0x8000EBCC800033BC, 0x8000EBD6800033C8, 0x800033FC800033DC,
-    0x800034108000EBDB, 0x8000EBE080003414, 0x800034308000EBE5,
-    0x80002AF480003440, 0x80002AF880003444, 0x800034608000345C,
-    0x800034788000346C, 0x8000349080002AFC, 0x800034AC8000EBEA,
-    0x8000EB548000EBEF, 0x8000EBF4800034C8, 0x8000EBFE8000EBF9,
-    0x80002B0480002B00, 0x8000EC038000EC08, 0x8000EC128000EC0D,
-    0x8000350C80003508, 0x80002B1080003518, 0x8000EC1C80002B14,
-    0x80002B188000352C, 0x00C161808000EC21, 0x0001D9820001C981,
-    0x0001C9840001D983, 0x0001F9860001C985, 0x00C145880001FD87,
-    0x0001CD8A00C1F989, 0x0001DD8C0001C58B, 0x0001CD8E0001C58D,
-    0x00C12D9000D4A18F, 0x0001B5920001BD91, 0x0003519400DC1993,
-    0x00DA8D9600035195, 0x00036D9800035597, 0x0003799A00DC0599,
-    0x0003699C0003799B, 0x0003659E0003699D, 0x0003A5A000035D9F,
-    0x0003A5A20003ADA1, 0x0003A5A40003BDA3, 0x0003A5A60003ADA5,
-    0x0003E5A80003DDA7, 0x0003E5AA0003EDA9, 0x0003E5AC0003FDAB,
-    0x0003E5AE0003EDAD, 0x0003A5B00003DDAF, 0x0003A5B20003ADB1,
-    0x0003A5B40003BDB3, 0x0003A5B60003ADB5, 0x000365B800035DB7,
-    0x00DC91BA00036DB9, 0x000361BC000361BB, 0x000361BE000361BD,
-    0x8000EC26000361BF, 0x8000359C8000EC2B, 0x800035A880002B1C,
-    0x8000EC30800035AC, 0x800035C48000EC35, 0x8000EC3A80002B20,
-    0x800035DC800035D8, 0x8000EC3F80002B24, 0x80002B28800035F8,
-    0x8000360080003608, 0x8000EC4980003610, 0x80002B2C80003628,
-    0x8000363C80003634, 0x80002B308000364C, 0x8000EC588000EC53,
-    0x8000EC5D80002B34, 0x8000EC6280003664, 0x8000367C80003670,
-    0x8000EC6C8000EC67, 0x8000EC7180003698, 0x8000EB81800036A0,
-    0x800036B880002B38, 0x80002B40800036BC, 0x80002F1A800036C0,
-    0x8000EC7B8000EC76, 0x8000EB8B8000EB86, 0x800038FC800036E4,
-    0x8000371080002B44, 0x800037148000370C, 0x8000371880002D1F,
-    0x800037208000371C, 0x8000EC8080003728, 0x000000808000372C,
-    0x8000374480003740, 0x8000374880003730, 0x8000EBB88000375C,
-    0x800037508000373C, 0x8000375880003754, 0x8000EC8F8000EC85,
-    0x80002B488000EC8A, 0x8000377C80003778, 0x8000ECA380003784,
-    0x8000EC9480003788, 0x80002B5080002B4C, 0x8000EC9E8000EC99,
-    0x800037A880002B54, 0x800037B4800037B0, 0x800037BC800037C0,
-    0x800037C4800037C8, 0x800037D0800037CC, 0x80002B58800037D8,
-    0x800037E4800037E0, 0x800037F480002B5C, 0x800038048000ECA8,
-    0x80002B6080003808, 0x80002A9480003814, 0x8000ECB28000ECAD,
-    0x80002B6880002B64, 0x8000ECB780003840, 0x8000389C80003898,
-    0x800038C0800038B4, 0x8000EB138000ECC1, 0x800038C8800038CC,
-    0x8000EB18800038D4, 0x00079580800038E0, 0x0007958200079D81,
-    0x00F7BD840007AD83, 0x0007B1860007B185, 0x000791880007B187,
-    0x0007918A00079189, 0x0007F18C0007918B, 0x0007F18E0007F18D,
-    0x00C4DD900007F18F, 0x00F7A99200C4E591, 0x0007859400079D93,
-    0x0007859600078D95, 0x000785980007BD97, 0x0007859A00078D99,
-    0x0007859C00079D9B, 0x0007859E00078D9D, 0x000285A000C4D99F,
-    0x000285A200028DA1, 0x000285A400029DA3, 0x000285A600028DA5,
-    0x000285A80002BDA7, 0x000285AA00028DA9, 0x000285AC00029DAB,
-    0x000285AE00028DAD, 0x000285B00002FDAF, 0x000285B200028DB1,
-    0x000285B400029DB3, 0x000285B600028DB5, 0x000285B80002BDB7,
-    0x000285BA00028DB9, 0x000285BC00029DBB, 0x00C445BE00028DBD,
-    0x000639A700C441BF, 0x000639A5000631A6, 0x000639A3000621A4,
-    0x000639A1000631A2, 0x000679AF000601A0, 0x000679AD000671AE,
-    0x000679AB000661AC, 0x000679A9000671AA, 0x000635B7000601A8,
-    0x000625B5000635B6, 0x000639B300D0C5B4, 0x000639B1000631B2,
-    0x0006F1BF000601B0, 0x0006E1BD0006E9BE, 0x00C5F9BB0006E9BC,
-    0x000605B90006F5BA, 0x00073D87000605B8, 0x00073D8500073D86,
-    0x00073D8300073D84, 0x00073D8100073D82, 0x0007798F00073D80,
-    0x0007798D0007718E, 0x001EC18B0007618C, 0x00077D8900077D8A,
-    0x0007399700077D88, 0x0007399500073196, 0x0007399300072194,
-    0x0007399100073192, 0x00D8659F00070190, 0x0006F99D0006F19E,
-    0x0006F99B0006E19C, 0x000705990006F19A, 0x0000008000DD8998,
-    0x8000ECCB8000ECC6, 0x800039388000392C, 0x8000ECD080003940,
-    0x8000397080003944, 0x800039808000397C, 0x8000398C80003978,
-    0x8000ECD580003990, 0x80002B6C8000399C, 0x8000ECDA800039A4,
-    0x800039DC80002B70, 0x800039F480002FAA, 0x8000ECE98000ECDF,
-    0x80002B7880002B74, 0x8000ECEE80003A1C, 0x8000ECF380002B7C,
-    0x0000FD8080003A3C, 0x0000758200007D81, 0x0001958400018D83,
-    0x0001958600019D85, 0x0001B58800018D87, 0x00C6698A0001BD89,
-    0x0001918C0001B18B, 0x0001918E0001918D, 0x0001F1900001918F,
-    0x0001F1920001F191, 0x000191940001F193, 0x00019D9600FE8D95,
-    0x0001A99800FEB597, 0x0001B99A0001A999, 0x00FE8D9C0001B99B,
-    0x8000ECF80001AD9D, 0x80002B8080003A54, 0x80003A7080003A64,
-    0x80003A8080003A78, 0x8000ECFD80002B84, 0x80003ABC80003AB0,
-    0x80002B8C80002B88, 0x8000ED0280003AC0, 0x8000ED0780002B90,
-    0x8000ED118000ED0C, 0x80003B0080002B94, 0x80003B0C80003B04,
-    0x8000ED1B80003B14, 0x000000D8000000D7, 0x000000DA000000D9,
-    0x000000DC000000DB, 0x000000DE000000DD, 0x000000E0000000DF,
-    0x000000E2000000E1, 0x000000E4000000E3, 0x000000E6000000E5,
-    0x000000E8000000E7, 0x000000EA000000E9, 0x000000EC000000EB,
-    0x000000EE000000ED, 0x000000F0000000EF, 0x000000F2000000F1,
-    0x000000F4000000F3, 0x000000F6000000F5, 0x000000F8000000F7,
-    0x000000FA000000F9, 0x000000FC000000FB, 0x000000FE000000FD,
-    0xCC200020000000FF, 0xCC200083CC200081, 0xCC200085CC200084,
-    0xCC200087CC200086, 0x0080CC88CC200088, 0xCC200081CC88CC20,
-    0x008ACC200082CD88, 0x0093CC20008BCC20, 0xCC200080CC93CC20,
-    0xCD93CC200081CC93, 0xCC200094CC200082, 0xCC94CC200080CC94,
-    0x0082CD94CC200081, 0x00A8CC2000A7CC20, 0x0082CD2000B3CC20,
-    0x008BD9200085CD20, 0xD98CD920008CD920, 0xD920008DD9200091,
-    0x008ED9200091D98D, 0xD9200091D98ED920, 0x0091D98FD920008F,
-    0xD990D9200090D920, 0xD9200091D9200091, 0x0092D92000B0D991,
-    0x82E320009982E320, 0x210021210021009A, 0x002400230022003F,
-    0x0028002700260025, 0x2930312800293128, 0x3128002931312800,
-    0x0029333128002932, 0x3531280029343128, 0x2800293631280029,
-    0x2938312800293731, 0x3228002939312800, 0x2800293032280029,
-    0x2800293428002933, 0x2800293628002935, 0x2800293828002937,
-    0x2800294128002939, 0x2800294328002942, 0x2800294528002944,
-    0x2800294728002946, 0x2800294928002948, 0x2800294B2800294A,
-    0x2800294D2800294C, 0x2800294F2800294E, 0x2800295128002950,
-    0x2800295328002952, 0x2800295528002954, 0x2800295728002956,
-    0x2800295928002958, 0x280029612800295A, 0x2800296328002962,
-    0x2800296528002964, 0x2800296728002966, 0x2800296928002968,
-    0x2800296B2800296A, 0x2800296D2800296C, 0x2800296F2800296E,
-    0x2800297128002970, 0x2800297328002972, 0x2800297528002974,
-    0x2800297728002976, 0x2800297928002978, 0x298084E12800297A,
-    0x2800298284E12800, 0x84E12800298384E1, 0x298684E128002985,
-    0x2800298784E12800, 0x84E12800298984E1, 0x298C84E12800298B,
-    0x2800298E84E12800, 0x84E12800298F84E1, 0x299184E128002990,
-    0x2800299284E12800, 0xB8E428002980B8E4, 0x2989B8E428002983,
-    0x2800299DB9E42800, 0xBAE42800298CBAE4, 0x29A3BBE428002994,
-    0x28002981BCE42800, 0x85E528002991BCE4, 0x29AD85E5280029AB,
-    0x280029B48AE52800, 0x8DE5280029818DE5, 0x298D90E528002994,
-    0x280029BC91E52800, 0x9CE52800299B9BE5, 0x29A6ADE52800299F,
-    0x280029A597E62800, 0x9CE6280029889CE6, 0x29A89CE628002989,
-    0x280029AAA0E62800, 0x81E7280029B4B0E6, 0x29B989E7280029AB,
-    0x280029A39BE72800, 0xA5E7280029BEA4E7, 0x29ADA5E72800299D,
-    0x280029AA87E82800, 0xB2E8280029B387E8, 0x2987B3E8280029A1,
-    0x2800299187E92800, 0x82EB28002980B0EA, 0x29A48BEB28002998,
-    0x280029BC9DEB2800, 0xB0EB28002988A7EB, 0x29AC82EC28002994,
-    0x2800298495EC2800, 0x002984A0ECA498EC, 0x29849BEDA498EC28,
-    0x280029909EEC2800, 0xB0EC280029BCA3EC, 0x29B4B9EC280029A8,
-    0x2800298083ED2800, 0x95ED2800298C8CED, 0x2B002A0029002998,
-    0x2E002E002D002C00, 0x002F002E2E2E002E, 0x002E30002C300030,
-    0xE73000338481E230, 0x002C31003100B982, 0x3031003031002E31,
-    0x00A597E63031002E, 0x303100889CE63031, 0x3100313100B982E7,
-    0xA597E63131002E31, 0x3100889CE6313100, 0x00323100B982E731,
-    0x97E63231002E3231, 0x00889CE6323100A5, 0x333100B982E73231,
-    0xE63331002E333100, 0xB982E7333100A597, 0x002E343100343100,
-    0x343100A597E63431, 0x3100353100B982E7, 0xA597E63531002E35,
-    0x3100B982E7353100, 0x3631002E36310036, 0x82E7363100A597E6,
-    0x2E373100373100B9, 0x3100A597E6373100, 0x00383100B982E737,
-    0x97E63831002E3831, 0x00B982E7383100A5, 0x31002E3931003931,
-    0xE7393100A597E639, 0x008481E23100B982, 0x310030318481E231,
-    0x81E23100328481E2, 0x348481E231003384, 0x3100358481E23100,
-    0x81E23100368481E2, 0x388481E231003784, 0x3100398481E23100,
-    0x889CE63100A597E6, 0x003200B982E73100, 0x3032002E32002C32,
-    0xE63032002E303200, 0xB982E7303200A597, 0x97E6313200313200,
-    0x00B982E7313200A5, 0xA597E63232003232, 0x3200B982E7323200,
-    0x00A597E633320033, 0x343200B982E73332, 0x3200A597E6343200,
-    0x00353200B982E734, 0x363200A597E63532, 0x3200A597E6363200,
-    0x00A597E637320037, 0xA597E63832003832, 0x97E6393200393200,
-    0x00338481E23200A5, 0xE63200358481E232, 0x00889CE63200A597,
-    0x33003300B982E732, 0x003033002E33002C, 0x313300A597E63033,
-    0x3300A597E6313300, 0x0034330033330032, 0x3733003633003533,
-    0x3300393300383300, 0x81E23300348481E2, 0x388481E233003584,
-    0xE63300A597E63300, 0x00B982E73300889C, 0x002E34002C340034,
-    0x3234003134003034, 0x3400343400333400, 0x0037340036340035,
-    0xE234003934003834, 0xA597E63400358481, 0xE73400889CE63400,
-    0x002C35003500B982, 0xE235003035002E35, 0x8481E23500368481,
-    0x3500A597E6350038, 0xB982E73500889CE6, 0x2E36002C36003600,
-    0xE63600A597E63600, 0x00B982E73600889C, 0x002E37002C370037,
-    0xE63700388481E237, 0x00889CE63700A597, 0x38003800B982E737,
-    0x97E638002E38002C, 0x3800889CE63800A5, 0x2C39003900B982E7,
-    0xA597E639002E3900, 0xE73900889CE63900, 0x3D3A3A003A00B982,
-    0x3D003D003C003B00, 0x003E003D3D3D003D, 0x003F3F00213F003F,
-    0x4100554100410040, 0x420042006D9588E2, 0x4300444300430071,
-    0x6B9588E243002E6F, 0x44004A4400440067, 0xBDC544007A44005A,
-    0x46004500BEC54400, 0x4700470058414600, 0x5047007A48470042,
-    0x4800480079470061, 0x0067480056480050, 0x0049490049007A48,
-    0x49004A4900494949, 0x0058490056490055, 0x4B00424B004B004A,
-    0x4C004C004D4B004B, 0x6A4C0044544C004A, 0x4D004D00B7C24C00,
-    0x00444D00434D0042, 0x0061504D007A484D, 0x574D00564D00524D,
-    0x4E004E00A9CE4D00, 0x006F4E006A4E004A, 0x500048500050004F,
-    0x5000565050004D50, 0x6150004554500052, 0x0073520052005100,
-    0x004D530044530053, 0x0054007653005353, 0x007A4854004C4554,
-    0x5600560055004D54, 0x4956004949560049, 0x6D9588E256004949,
-    0x5A57004357005700, 0x4958005800625700, 0x5A00590049495800,
-    0x5E005D005C005B00, 0x6100610060005F00, 0x00632F61002E6D2E,
-    0x00BECA6100732F61, 0x0063007261620062, 0x00752F63006F2F63,
-    0x63006363006C6163, 0x326D63006D630064, 0x64006400336D6300,
-    0x006C640061640042, 0x6400326D64006D64, 0xC564007A6400336D,
-    0x65005665006500BE, 0x0066660066006772, 0x006C666600696666,
-    0x6D66006C66006966, 0x68006C6167006700, 0x0061680061506800,
-    0x6969690069690069, 0x69006E69006A6900, 0x6B006A0078690076,
-    0x007A486B00416B00, 0x6B00566B0061506B, 0x6B006C61636B0057,
-    0x006D6B006C6B0067, 0x00336D6B00326D6B, 0x6C00A9CE6B00746B,
-    0x6C006D6C006A6C00, 0x786C00676F6C006E, 0x6D006D00B7C26C00,
-    0x00416D00336D0032, 0x626D00576D00566D, 0x006C696D00676D00,
-    0x6D6D006D6D006C6D, 0x6F6D00336D6D0032, 0x88E26D00736D006C,
-    0x739588E26D007395, 0x6E00416E006E0032, 0x00576E00566E0046,
-    0x736E006D6E006A6E, 0x007000566F006F00, 0x004170002E6D2E70,
-    0x5770005670004670, 0x7100737000637000, 0x7200646172007200,
-    0x7200739588E26461, 0x0032739588E26461, 0x0074730072730073,
-    0x6976007600750074, 0x6969760069697600, 0x6978007800770069,
-    0x7A00790069697800, 0x7E007D007C007B00, 0xC200A3C200A2C200,
-    0x00ACC200A6C200A5, 0x0046B0C20043B0C2, 0x81C30080C300B7C2,
-    0xC30083C30082C300, 0x0086C30085C30084, 0x89C30088C30087C3,
-    0xC3008BC3008AC300, 0x008EC3008DC3008C, 0x92C30091C3008FC3,
-    0xC30094C30093C300, 0x0099C30096C30095, 0x9CC3009BC3009AC3,
-    0xC300A0C3009DC300, 0x00A3C300A2C300A1, 0xA6C300A5C300A4C3,
-    0xC300A8C300A7C300, 0x00ABC300AAC300A9, 0xAEC300ADC300ACC3,
-    0xC300B0C300AFC300, 0x00B3C300B2C300B1, 0xB6C300B5C300B4C3,
-    0xC300B9C300B8C300, 0x00BCC300BBC300BA, 0x80C400BFC300BDC3,
-    0xC40082C40081C400, 0x0085C40084C40083, 0x88C40087C40086C4,
-    0xC4008AC40089C400, 0x008DC4008CC4008B, 0x92C4008FC4008EC4,
-    0xC40094C40093C400, 0x0097C40096C40095, 0x9AC40099C40098C4,
-    0xC4009CC4009BC400, 0x009FC4009EC4009D, 0xA2C400A1C400A0C4,
-    0xC400A4C400A3C400, 0x00A7C400A6C400A5, 0xAAC400A9C400A8C4,
-    0xC400ACC400ABC400, 0x00AFC400AEC400AD, 0xB4C400B1C400B0C4,
-    0xC400B6C400B5C400, 0x00BAC400B9C400B7, 0xBDC400BCC400BBC4,
-    0xC50083C500BEC400, 0x0086C50085C50084, 0x8BC50088C50087C5,
-    0xC5008DC5008CC500, 0x0090C5008FC5008E, 0x94C50093C50091C5,
-    0xC50096C50095C500, 0x0099C50098C50097, 0x9CC5009BC5009AC5,
-    0xC5009EC5009DC500, 0x00A1C500A0C5009F, 0xA4C500A3C500A2C5,
-    0xC500A8C500A5C500, 0x00ABC500AAC500A9, 0xAEC500ADC500ACC5,
-    0xC500B0C500AFC500, 0x00B3C500B2C500B1, 0xB6C500B5C500B4C5,
-    0xC500B8C500B7C500, 0x00BBC500BAC500B9, 0xBEC500BDC500BCC5,
-    0xC60090C6008EC600, 0x00ABC600A1C600A0, 0x80C700B0C600AFC6,
-    0xC70082C70081C700, 0x008FC7008EC7008D, 0x92C70091C70090C7,
-    0xC70094C70093C700, 0x0097C70096C70095, 0x9AC70099C70098C7,
-    0xC7009CC7009BC700, 0x00A0C7009FC7009E, 0xA3C700A2C700A1C7,
-    0xC700A7C700A6C700, 0x00AAC700A9C700A8, 0xADC700ACC700ABC7,
-    0xC700AFC700AEC700, 0x00B5C700B4C700B0, 0xBAC700B9C700B8C7,
-    0xC700BCC700BBC700, 0x00BFC700BEC700BD, 0x82C80081C80080C8,
-    0xC80084C80083C800, 0x0087C80086C80085, 0x8AC80089C80088C8,
-    0xC8008CC8008BC800, 0x008FC8008EC8008D, 0x92C80091C80090C8,
-    0xC80094C80093C800, 0x0097C80096C80095, 0x9AC80099C80098C8,
-    0xC8009EC8009BC800, 0x00A6C800A2C8009F, 0xA9C800A8C800A7C8,
-    0xC800ABC800AAC800, 0x00AEC800ADC800AC, 0xB1C800B0C800AFC8,
-    0xC800B3C800B2C800, 0x0091C90090C900B7, 0x94C90093C90092C9,
-    0xC90096C90095C900, 0x0099C90098C90097, 0x9EC9009CC9009BC9,
-    0xC900A0C9009FC900, 0x00A3C900A2C900A1, 0xA6C900A5C900A4C9,
-    0xC900A8C900A7C900, 0x00ABC900AAC900A9, 0xAEC900ADC900ACC9,
-    0xC900B0C900AFC900, 0x00B3C900B2C900B1, 0xB6C900B5C900B4C9,
-    0xC900B8C900B7C900, 0x00BBC900BAC900B9, 0x80CA00BEC900BDC9,
-    0xCA0082CA0081CA00, 0x0088CA0084CA0083, 0x8BCA008ACA0089CA,
-    0xCA008DCA008CCA00, 0x0090CA008FCA008E, 0x95CA0092CA0091CA,
-    0xCA0099CA0098CA00, 0x009DCA009CCA009B, 0xA2CA00A1CA009FCA,
-    0xCA00A4CA00A3CA00, 0x00A7CA00A6CA00A5, 0xAACA00A9CA00A8CA,
-    0xCA00B9CA00ABCA00, 0x91CB0090CB006EBC, 0xCC0081CC0080CC00,
-    0xCE0093CC0081CC88, 0x0089CE0088CE0086, 0x8ECE008CCE008ACE,
-    0xCE0090CE008FCE00, 0x0093CE0092CE0091, 0x96CE0095CE0094CE,
-    0xCE0098CE0097CE00, 0x009BCE009ACE0099, 0x9ECE009DCE009CCE,
-    0xCE00A0CE009FCE00, 0x00A4CE00A3CE00A1, 0xA7CE00A6CE00A5CE,
-    0xCE00A9CE00A8CE00, 0x00ACCE00ABCE00AA, 0xAFCE00AECE00ADCE,
-    0xCE00B1CE00B0CE00, 0x00B4CE00B3CE00B2, 0xB7CE00B6CE00B5CE,
-    0xCE00B9CE00B8CE00, 0x00BCCE00BBCE00BA, 0x0046BCCE0041BCCE,
-    0x0057BCCE0056BCCE, 0x006CBCCE0067BCCE, 0x0073BCCE006DBCCE,
-    0xBFCE00BECE00BDCE, 0xCF0081CF0080CF00, 0x0084CF0083CF0082,
-    0x87CF0086CF0085CF, 0xCF0089CF0088CF00, 0x008CCF008BCF008A,
-    0x9CCF008ECF008DCF, 0xD00080D0009DCF00, 0x0087D00083D00081,
-    0x8ED0008DD0008CD0, 0xD000B0D00099D000, 0x00B3D000B2D000B1,
-    0xB6D000B5D000B4D0, 0xD000B8D000B7D000, 0x00BBD000BAD000B9,
-    0xBED000BDD000BCD0, 0xD10080D100BFD000, 0x0083D10082D10081,
-    0x86D10085D10084D1, 0xD10088D10087D100, 0x008CD1008BD1008A,
-    0x90D1008ED1008DD1, 0xD10093D10091D100, 0x0097D10096D10095,
-    0x9DD1009CD10098D1, 0xD1009FD1009ED100, 0x0091D200B7D100B6,
-    0xB1D200AFD200ABD2, 0xD30082D30081D300, 0x0091D30090D3008F,
-    0x96D30093D30092D3, 0xD30099D30097D300, 0x009CD3009BD3009A,
-    0x9FD3009ED3009DD3, 0xD300A3D300A2D300, 0x00A6D300A5D300A4,
-    0xAAD300A9D300A7D3, 0xD300ACD300ABD300, 0x00AFD300AED300AD,
-    0xB2D300B1D300B0D3, 0xD300B4D300B3D300, 0x00B9D300B8D300B5,
-    0xD5B4D50082D6A5D5, 0xD500ABD5B4D500A5, 0xB6D5B4D500ADD5B4,
-    0x90D700B6D5BED500, 0x90D700B7D690D700, 0x00BCD690D700B8D6,
-    0x0091D7009CD790D7, 0xD691D700BCD691D7, 0xD692D70092D700BF,
-    0xD693D70093D700BC, 0xD694D70094D700BC, 0xD700B9D695D700BC,
-    0xBCD696D700BCD695, 0x99D700BCD698D700, 0x00BCD699D700B4D6,
-    0x009BD700BCD69AD7, 0xD69BD700BCD69BD7, 0xD69CD7009CD700BF,
-    0xD69ED7009DD700BC, 0xD700BCD6A0D700BC, 0xD700A2D700BCD6A1,
-    0xBCD6A4D700BCD6A3, 0xA6D700BFD6A4D700, 0x00BCD6A7D700BCD6,
-    0x00BCD6A8D700A8D7, 0xD6A9D700BCD6A9D7, 0xBCD6A9D70081D7BC,
-    0x0081D7A9D70082D7, 0x00AAD70082D7A9D7, 0xD6B2D700BCD6AAD7,
-    0x00A2D800A1D800B7, 0xA5D800A4D800A3D8, 0xA7D8A6D800A6D800,
-    0xA6D800ACD8A6D800, 0x00AED8A6D800ADD8, 0xD8A6D800B1D8A6D8,
-    0xD80085D9A6D800B2, 0x87D9A6D80086D9A6, 0xA6D80088D9A6D800,
-    0x008AD9A6D80089D9, 0xDBA6D80086DBA6D8, 0xD80088DBA6D80087,
-    0x95DBA6D80090DBA6, 0x83D9A7D800A7D800, 0xD9A7D800B1D8A8D8,
-    0xA7D80087D984D984, 0x00B4D9A7D8008BD9, 0x00ACD8A8D800A8D8,
-    0xD8A8D800ADD8A8D8, 0xAED8A8D8008AD9AD, 0x008AD9AED8A8D800,
-    0xD8A8D800B1D8A8D8, 0xD80085D9A8D800B2, 0x87D9A8D80086D9A8,
-    0xA8D80089D9A8D800, 0xAAD800A9D8008AD9, 0xAAD800ACD8AAD800,
-    0xD8AAD80085D9ACD8, 0xACD8AAD80089D9AC, 0x00ADD8AAD8008AD9,
-    0xD800ACD8ADD8AAD8, 0xAAD80085D9ADD8AA, 0xD9AED8AAD800AED8,
-    0x89D9AED8AAD80085, 0x008AD9AED8AAD800, 0xD8AAD800B1D8AAD8,
-    0xD80085D9AAD800B2, 0xAAD800ACD885D9AA, 0xD9AAD800ADD885D9,
-    0x85D9AAD800AED885, 0xD985D9AAD80089D9, 0xD80086D9AAD8008A,
-    0x89D9AAD80087D9AA, 0xABD8008AD9AAD800, 0xABD800ACD8ABD800,
-    0x00B2D8ABD800B1D8, 0xD9ABD80085D9ABD8, 0xD80087D9ABD80086,
-    0x8AD9ABD80089D9AB, 0xADD8ACD800ACD800, 0x0089D9ADD8ACD800,
-    0xD8008AD9ADD8ACD8, 0x84D9ACD82084D9AC, 0xD80087D984D9A7D8,
-    0x85D9ACD80085D9AC, 0xD985D9ACD800ADD8, 0x8AD985D9ACD80089,
-    0xACD80089D9ACD800, 0xADD800ADD8008AD9, 0xD9ACD8ADD800ACD8,
-    0xD80085D9ADD8008A, 0xADD80089D985D9AD, 0xD9ADD8008AD985D9,
-    0xD8008AD9ADD80089, 0xD800ACD8AED800AE, 0x85D9AED800ADD8AE,
-    0xAED80089D9AED800, 0xB0D800AFD8008AD9, 0xB1D800B0D9B0D800,
-    0xD988D9B3D8B1D800, 0xD800B0D9B1D80084, 0x0084D9A7D88CDBB1,
-    0xB3D800B3D800B2D8, 0xD8ACD8B3D800ACD8, 0x89D9ACD8B3D800AD,
-    0xB3D800ADD8B3D800, 0xD8B3D800ACD8ADD8, 0x89D9AED8B3D800AE,
-    0x008AD9AED8B3D800, 0xD9B3D800B1D8B3D8, 0xACD885D9B3D80085,
-    0x00ADD885D9B3D800, 0xD80085D985D9B3D8, 0x89D9B3D80087D9B3,
-    0xB4D8008AD9B3D800, 0xB4D800ACD8B4D800, 0xD8B4D8008AD9ACD8,
-    0x85D9ADD8B4D800AD, 0x008AD9ADD8B4D800, 0xD8B4D800AED8B4D8,
-    0xD80085D9B4D800B1, 0xB4D800AED885D9B4, 0xD9B4D80085D985D9,
-    0xD80089D9B4D80087, 0xD800B5D8008AD9B4, 0xADD8B5D800ADD8B5,
-    0xD9ADD8B5D800ADD8, 0xD800AED8B5D8008A, 0x84D9B5D800B1D8B5,
-    0xD9B5D80085D9B9D8, 0x84D9B5D80089D984, 0xD984D9A7D82089D9,
-    0x84D9B9D82087D984, 0xD888D92087D98AD9, 0xB5D80085D984D9B3,
-    0xD9B5D80092DB84D9, 0x85D985D9B5D80085, 0xB5D80089D9B5D800,
-    0xB6D800B6D8008AD9, 0x00ADD8B6D800ACD8, 0xD80089D9ADD8B6D8,
-    0xB6D8008AD9ADD8B6, 0xD9AED8B6D800AED8, 0xD800B1D8B6D80085,
-    0x89D9B6D80085D9B6, 0xB7D8008AD9B6D800, 0xB7D800ADD8B7D800,
-    0xD885D9B7D80085D9, 0x85D985D9B7D800AD, 0x008AD985D9B7D800,
-    0xD9B7D80089D9B7D8, 0xD9B8D800B8D8008A, 0xD8B9D800B9D80085,
-    0x85D9ACD8B9D800AC, 0xD98AD984D9B9D800, 0xD80085D9B9D80087,
-    0xB9D80085D985D9B9, 0xD9B9D80089D985D9, 0x89D9B9D8008AD985,
-    0xBAD8008AD9B9D800, 0xBAD800ACD8BAD800, 0xD985D9BAD80085D9,
-    0x89D985D9BAD80085, 0x008AD985D9BAD800, 0xD9BAD80089D9BAD8,
-    0xD9008BD980D9008A, 0x8ED980D9008ED980, 0x008FD980D90091D9,
-    0xD90091D98FD980D9, 0x90D980D90090D980, 0x0091D980D90091D9,
-    0x0081D90092D980D9, 0xD881D900ACD881D9, 0xD900AED881D900AD,
-    0x81D90085D9AED881, 0xD985D981D90085D9, 0xD90089D981D9008A,
-    0xD90082D9008AD981, 0x84D982D900ADD882, 0x0085D982D90092DB,
-    0xD900ADD885D982D9, 0x82D90085D985D982, 0xD982D9008AD985D9,
-    0xD9008AD982D90089, 0xD900A7D883D90083, 0xADD883D900ACD883,
-    0x83D900AED883D900, 0x0085D983D90084D9, 0xD90085D985D983D9,
-    0x83D9008AD985D983, 0x008AD983D90089D9, 0x00A2D884D90084D9,
-    0xD884D900A3D884D9, 0xD900A7D884D900A5, 0xACD884D900ACD884,
-    0xD9ACD884D900ACD8, 0x8AD9ACD884D90085, 0x84D900ADD884D900,
-    0xD884D90085D9ADD8, 0xADD884D90089D9AD, 0x00AED884D9008AD9,
-    0xD90085D9AED884D9, 0x85D984D90085D984, 0xD985D984D900ADD8,
-    0xD90087D984D9008A, 0x8AD984D90089D984, 0xA7D885D90085D900,
-    0x85D900ACD885D900, 0xD885D900ADD8ACD8, 0xACD885D900AED8AC,
-    0xD9ACD885D90085D9, 0xD900ADD885D9008A, 0x85D900ACD8ADD885,
-    0xD885D90085D9ADD8, 0x85D900AFD885D9AD, 0xD885D9008AD9ADD8,
-    0xACD8AED885D900AE, 0x0085D9AED885D900, 0xD9008AD9AED885D9,
-    0x85D985D90085D985, 0x0089D985D9008AD9, 0x0086D9008AD985D9,
-    0xD886D900ACD886D9, 0xACD886D900ADD8AC, 0xD9ACD886D90085D9,
-    0x8AD9ACD886D90089, 0x86D900ADD886D900, 0xD886D90085D9ADD8,
-    0xADD886D90089D9AD, 0x00AED886D9008AD9, 0xD886D900B1D886D9,
-    0xD90085D986D900B2, 0x86D90089D985D986, 0xD986D9008AD985D9,
-    0xD90087D986D90086, 0x8AD986D90089D986, 0xACD887D90087D900,
-    0x87D90085D987D900, 0xD987D900ACD885D9, 0x89D987D90085D985,
-    0x87D9008AD987D900, 0x88D90088D900B0D9, 0xD90085D984D9B3D8,
-    0xD90089D900B4D988, 0xD9008AD900B0D989, 0xACD88AD900ACD88A,
-    0x00ADD88AD9008AD9, 0xD9008AD9ADD88AD9, 0xB1D88AD900AED88A,
-    0x8AD900B2D88AD900, 0xD985D98AD90085D9, 0x8AD985D98AD90085,
-    0x8AD90086D98AD900, 0x0089D98AD90087D9, 0xD98AD9008AD98AD9,
-    0x00AFD900AED900B4, 0xBAD900B9D900B1D9, 0xD900BED900BBD900,
-    0x0083DA0080DA00BF, 0x87DA0086DA0084DA, 0xDA008CDA0088DA00,
-    0x0091DA008EDA008D, 0xA4DA00A1DA0098DA, 0xDA00A9DA00A6DA00,
-    0x00B1DA00AFDA00AD, 0xBBDA00BADA00B3DA, 0xDB0080DB00BEDA00,
-    0x0085DB0082DB0081, 0x87DB0087DB0086DB, 0x89DB0088DB00B4D9,
-    0xDB008CDB008BDB00, 0x0093DB0092DB0090, 0xE000BCA4E095A4E0,
-    0xA4E000BCA4E096A4, 0x9CA4E000BCA4E097, 0xE0A1A4E000BCA4E0,
-    0xA4E0A2A4E000BCA4, 0xA4E000A9A4E000BC, 0xAFA4E000BCA4E0AB,
-    0x00B1A4E000BCA4E0, 0xE0A1A6E000B4A4E0, 0xA6E0A2A6E000BCA6,
-    0xBCA6E0AFA6E000BC, 0x8CA7E0008BA7E000, 0x00BCA8E096A8E000,
-    0xE000BCA8E097A8E0, 0xA8E000BCA8E09CA8, 0xB2A8E000BCA8E0AB,
-    0xE0B8A8E000BCA8E0, 0xACE0A1ACE000BCA8, 0xBCACE0A2ACE000BC,
-    0x8BADE00088ADE000, 0x94AEE0008CADE000, 0x8BAFE0008AAFE000,
-    0x88B1E0008CAFE000, 0x87B3E00080B3E000, 0x8AB3E00088B3E000,
-    0x8AB5E0008BB3E000, 0x8CB5E0008BB5E000, 0x9CB7E0009AB7E000,
-    0x9EB7E0009DB7E000, 0x00B2B8E08DB9E000, 0xE00099BAE0ABBAE0,
-    0xBBE000A1BAE0ABBA, 0x8BBCE000B2BAE08D, 0x00B5BEE080BDE000,
-    0xE000B7BEE082BDE0, 0xBDE000B7BEE08CBD, 0x96BDE000B7BEE091,
-    0xE09BBDE000B7BEE0, 0xBDE0B1BDE000B7BE, 0xB4BDE0B1BDE000B2,
-    0x0080BEE0B1BDE000, 0xE000B5BEE090BEE0, 0xBEE000B7BEE092BE,
-    0xA1BEE000B7BEE09C, 0xE0A6BEE000B7BEE0, 0xBEE0ABBEE000B7BE,
-    0xB1BDE0B2BEE000B7, 0xE0B2BEE00080BEE0, 0xBDE0B3BEE00080BE,
-    0xB3BEE00080BEE0B1, 0x00A680E10080BEE0, 0x008084E1009C83E1,
-    0x008284E1008184E1, 0x008484E1008384E1, 0x008684E1008584E1,
-    0x008884E1008784E1, 0x008A84E1008984E1, 0x008C84E1008B84E1,
-    0x008E84E1008D84E1, 0x009084E1008F84E1, 0x009284E1009184E1,
-    0x009584E1009484E1, 0x009C84E1009A84E1, 0x009E84E1009D84E1,
-    0x00A184E100A084E1, 0x00A384E100A284E1, 0x00A984E100A784E1,
-    0x00AC84E100AB84E1, 0x00AE84E100AD84E1, 0x00B284E100AF84E1,
-    0x008085E100B684E1, 0x008C85E1008785E1, 0x009885E1009785E1,
-    0x00A085E1009985E1, 0x00A285E100A185E1, 0x00A485E100A385E1,
-    0x00A685E100A585E1, 0x00A885E100A785E1, 0x00AA85E100A985E1,
-    0x00AC85E100AB85E1, 0x00AE85E100AD85E1, 0x00B085E100AF85E1,
-    0x00B285E100B185E1, 0x00B485E100B385E1, 0x008486E100B585E1,
-    0x008886E1008586E1, 0x009286E1009186E1, 0x009E86E1009486E1,
-    0x00AA86E100A186E1, 0x00AD86E100AC86E1, 0x00B186E100B086E1,
-    0x00B386E100B286E1, 0x00B586E100B486E1, 0x008887E1008787E1,
-    0x008E87E1008C87E1, 0x009787E1009387E1, 0x009D87E1009987E1,
-    0x00B187E1009F87E1, 0x0086ACE100B287E1, 0x008AACE10088ACE1,
-    0x008EACE1008CACE1, 0x00BBACE10092ACE1, 0x0080ADE100BDACE1,
-    0x0083ADE10081ADE1, 0x0096B4E10082B4E1, 0x009CB4E10097B4E1,
-    0x00A5B4E1009DB4E1, 0x0085B6E100BBB5E1, 0x0080B8E10091B6E1,
-    0x0082B8E10081B8E1, 0x0084B8E10083B8E1, 0x0086B8E10085B8E1,
-    0x0088B8E10087B8E1, 0x008AB8E10089B8E1, 0x008CB8E1008BB8E1,
-    0x008EB8E1008DB8E1, 0x0090B8E1008FB8E1, 0x0092B8E10091B8E1,
-    0x0094B8E10093B8E1, 0x0096B8E10095B8E1, 0x0098B8E10097B8E1,
-    0x009AB8E10099B8E1, 0x009CB8E1009BB8E1, 0x009EB8E1009DB8E1,
-    0x00A0B8E1009FB8E1, 0x00A2B8E100A1B8E1, 0x00A4B8E100A3B8E1,
-    0x00A6B8E100A5B8E1, 0x00A8B8E100A7B8E1, 0x00AAB8E100A9B8E1,
-    0x00ACB8E100ABB8E1, 0x00AEB8E100ADB8E1, 0x00B0B8E100AFB8E1,
-    0x00B2B8E100B1B8E1, 0x00B4B8E100B3B8E1, 0x00B6B8E100B5B8E1,
-    0x00B8B8E100B7B8E1, 0x00BAB8E100B9B8E1, 0x00BCB8E100BBB8E1,
-    0x00BEB8E100BDB8E1, 0x0080B9E100BFB8E1, 0x0082B9E10081B9E1,
-    0x0084B9E10083B9E1, 0x0086B9E10085B9E1, 0x0088B9E10087B9E1,
-    0x008AB9E10089B9E1, 0x008CB9E1008BB9E1, 0x008EB9E1008DB9E1,
-    0x0090B9E1008FB9E1, 0x0092B9E10091B9E1, 0x0094B9E10093B9E1,
-    0x0096B9E10095B9E1, 0x0098B9E10097B9E1, 0x009AB9E10099B9E1,
-    0x009CB9E1009BB9E1, 0x009EB9E1009DB9E1, 0x00A0B9E1009FB9E1,
-    0x00A2B9E100A1B9E1, 0x00A4B9E100A3B9E1, 0x00A6B9E100A5B9E1,
-    0x00A8B9E100A7B9E1, 0x00AAB9E100A9B9E1, 0x00ACB9E100ABB9E1,
-    0x00AEB9E100ADB9E1, 0x00B0B9E100AFB9E1, 0x00B2B9E100B1B9E1,
-    0x00B4B9E100B3B9E1, 0x00B6B9E100B5B9E1, 0x00B8B9E100B7B9E1,
-    0x00BAB9E100B9B9E1, 0x00BCB9E100BBB9E1, 0x00BEB9E100BDB9E1,
-    0x0080BAE100BFB9E1, 0x0082BAE10081BAE1, 0x0084BAE10083BAE1,
-    0x0086BAE10085BAE1, 0x0088BAE10087BAE1, 0x008ABAE10089BAE1,
-    0x008CBAE1008BBAE1, 0x008EBAE1008DBAE1, 0x0090BAE1008FBAE1,
-    0x0092BAE10091BAE1, 0x0094BAE10093BAE1, 0x0096BAE10095BAE1,
-    0x0098BAE10097BAE1, 0x00A0BAE10099BAE1, 0x00A2BAE100A1BAE1,
-    0x00A4BAE100A3BAE1, 0x00A6BAE100A5BAE1, 0x00A8BAE100A7BAE1,
-    0x00AABAE100A9BAE1, 0x00ACBAE100ABBAE1, 0x00AEBAE100ADBAE1,
-    0x00B0BAE100AFBAE1, 0x00B2BAE100B1BAE1, 0x00B4BAE100B3BAE1,
-    0x00B6BAE100B5BAE1, 0x00B8BAE100B7BAE1, 0x00BABAE100B9BAE1,
-    0x00BCBAE100BBBAE1, 0x00BEBAE100BDBAE1, 0x0080BBE100BFBAE1,
-    0x0082BBE10081BBE1, 0x0084BBE10083BBE1, 0x0086BBE10085BBE1,
-    0x0088BBE10087BBE1, 0x008ABBE10089BBE1, 0x008CBBE1008BBBE1,
-    0x008EBBE1008DBBE1, 0x0090BBE1008FBBE1, 0x0092BBE10091BBE1,
-    0x0094BBE10093BBE1, 0x0096BBE10095BBE1, 0x0098BBE10097BBE1,
-    0x009ABBE10099BBE1, 0x009CBBE1009BBBE1, 0x009EBBE1009DBBE1,
-    0x00A0BBE1009FBBE1, 0x00A2BBE100A1BBE1, 0x00A4BBE100A3BBE1,
-    0x00A6BBE100A5BBE1, 0x00A8BBE100A7BBE1, 0x00AABBE100A9BBE1,
-    0x00ACBBE100ABBBE1, 0x00AEBBE100ADBBE1, 0x00B0BBE100AFBBE1,
-    0x00B2BBE100B1BBE1, 0x00B4BBE100B3BBE1, 0x00B6BBE100B5BBE1,
-    0x00B8BBE100B7BBE1, 0x0080BCE100B9BBE1, 0x0082BCE10081BCE1,
-    0x0084BCE10083BCE1, 0x0086BCE10085BCE1, 0x0088BCE10087BCE1,
-    0x008ABCE10089BCE1, 0x008CBCE1008BBCE1, 0x008EBCE1008DBCE1,
-    0x0090BCE1008FBCE1, 0x0092BCE10091BCE1, 0x0094BCE10093BCE1,
-    0x0098BCE10095BCE1, 0x009ABCE10099BCE1, 0x009CBCE1009BBCE1,
-    0x00A0BCE1009DBCE1, 0x00A2BCE100A1BCE1, 0x00A4BCE100A3BCE1,
-    0x00A6BCE100A5BCE1, 0x00A8BCE100A7BCE1, 0x00AABCE100A9BCE1,
-    0x00ACBCE100ABBCE1, 0x00AEBCE100ADBCE1, 0x00B0BCE100AFBCE1,
-    0x00B2BCE100B1BCE1, 0x00B4BCE100B3BCE1, 0x00B6BCE100B5BCE1,
-    0x00B8BCE100B7BCE1, 0x00BABCE100B9BCE1, 0x00BCBCE100BBBCE1,
-    0x00BEBCE100BDBCE1, 0x0080BDE100BFBCE1, 0x0082BDE10081BDE1,
-    0x0084BDE10083BDE1, 0x0088BDE10085BDE1, 0x008ABDE10089BDE1,
-    0x008CBDE1008BBDE1, 0x0090BDE1008DBDE1, 0x0092BDE10091BDE1,
-    0x0094BDE10093BDE1, 0x0096BDE10095BDE1, 0x0099BDE10097BDE1,
-    0x009DBDE1009BBDE1, 0x00A0BDE1009FBDE1, 0x00A2BDE100A1BDE1,
-    0x00A4BDE100A3BDE1, 0x00A6BDE100A5BDE1, 0x00A8BDE100A7BDE1,
-    0x00AABDE100A9BDE1, 0x00ACBDE100ABBDE1, 0x00AEBDE100ADBDE1,
-    0x00B0BDE100AFBDE1, 0x00B4BDE100B2BDE1, 0x00B8BDE100B6BDE1,
-    0x00BCBDE100BABDE1, 0x0081BEE10080BEE1, 0x0083BEE10082BEE1,
-    0x0085BEE10084BEE1, 0x0087BEE10086BEE1, 0x0089BEE10088BEE1,
-    0x008BBEE1008ABEE1, 0x008DBEE1008CBEE1, 0x008FBEE1008EBEE1,
-    0x0091BEE10090BEE1, 0x0093BEE10092BEE1, 0x0095BEE10094BEE1,
-    0x0097BEE10096BEE1, 0x0099BEE10098BEE1, 0x009BBEE1009ABEE1,
-    0x009DBEE1009CBEE1, 0x009FBEE1009EBEE1, 0x00A1BEE100A0BEE1,
-    0x00A3BEE100A2BEE1, 0x00A5BEE100A4BEE1, 0x00A7BEE100A6BEE1,
-    0x00A9BEE100A8BEE1, 0x00ABBEE100AABEE1, 0x00ADBEE100ACBEE1,
-    0x00AFBEE100AEBEE1, 0x00B1BEE100B0BEE1, 0x00B3BEE100B2BEE1,
-    0x00B6BEE100B4BEE1, 0x00B8BEE100B7BEE1, 0x00BABEE100B9BEE1,
-    0x0082BFE100BCBEE1, 0x0084BFE10083BFE1, 0x0087BFE10086BFE1,
-    0x008ABFE10088BFE1, 0x0090BFE1008CBFE1, 0x0092BFE10091BFE1,
-    0x0097BFE10096BFE1, 0x0099BFE10098BFE1, 0x00A0BFE1009ABFE1,
-    0x00A2BFE100A1BFE1, 0x00A5BFE100A4BFE1, 0x00A7BFE100A6BFE1,
-    0x00A9BFE100A8BFE1, 0x00ACBFE100AABFE1, 0x00B3BFE100B2BFE1,
-    0x00B6BFE100B4BFE1, 0x00B8BFE100B7BFE1, 0x00BCBFE100BABFE1,
-    0x009380E2009080E2, 0xE2B280E2009480E2, 0x80E2B280E200B280,
-    0xB280E200B280E2B2, 0x80E2B280E2B280E2, 0xB580E2B580E200B2,
-    0xE2B580E2B580E200, 0xE200A982E200B580, 0xE2009186E2009086,
-    0xE2009386E2009286, 0xE2009B86E2009A86, 0xE2008D87E200AE86,
-    0xE2008F87E2008E87, 0xE2008488E2008288, 0xE2008988E2008788,
-    0xE2009188E2008C88, 0xE200A488E2009288, 0x88E2AB88E200A688,
-    0xAB88E2AB88E200AB, 0xE2AB88E200AB88E2, 0xAB88E2AB88E2AB88,
-    0x00AE88E2AE88E200, 0x88E2AE88E2AE88E2, 0x89E2008189E200AE,
-    0x89E2008789E20084, 0x89E200A089E20089, 0x89E200AD89E200A2,
-    0x89E200AF89E200AE, 0x89E200B189E200B0, 0x89E200B589E200B4,
-    0x8AE200B989E200B8, 0x8AE200818AE20080, 0x8AE200858AE20084,
-    0x8AE200898AE20088, 0x8AE200AD8AE200AC, 0x8BE200AF8AE200AE,
-    0x8BE200A18BE200A0, 0x8BE200A38BE200A2, 0x8BE200AB8BE200AA,
-    0x94E200AD8BE200AC, 0x97E200A096E20082, 0xA6E20085A6E2008B,
-    0x00B8CC9DABE20086, 0x00A1B5E200B1B1E2, 0x008280E3008180E3,
-    0x008980E3008880E3, 0x008B80E3008A80E3, 0x008D80E3008C80E3,
-    0x008F80E3008E80E3, 0x009180E3009080E3, 0x009480E3009280E3,
-    0x009580E3539480E3, 0x80E389B8E49480E3, 0x8CBAE49480E30095,
-    0xE59480E3009580E3, 0x80E3009580E39D8B, 0x009580E389AEE594,
-    0x80E39389E69480E3, 0x9795E69480E30095, 0xE69480E3009580E3,
-    0x80E3009580E3AC9C, 0x009580E3B982E794, 0x80E3979BE79480E3,
-    0x80E3009580E30095, 0x81E3009780E30096, 0x81E3008E81E3008C,
-    0x81E3009281E30090, 0x81E3009681E30094, 0x81E3009A81E30098,
-    0x81E3009E81E3009C, 0x81E300A281E300A0, 0x81E300A781E300A5,
-    0x81E300B081E300A9, 0x81E300B381E300B1, 0x81E300B681E300B4,
-    0x81E300B981E300B7, 0x8B81E3BB81E300BA, 0xBD81E300BC81E300,
-    0x008A82E38882E300, 0x009982E3009482E3, 0x009E82E3009A82E3,
-    0x00A282E300A182E3, 0x83E39183E3A282E3, 0xA282E3008883E3BC,
-    0x82E39583E3AB83E3, 0xB383E3A282E300A1, 0xE300A282E39A83E3,
-    0xAB83E3BC83E3A282, 0xA482E300A382E300, 0xE38B83E3A482E300,
-    0x82E300B082E3B383, 0x008183E3B383E3A4, 0x00A682E300A582E3,
-    0x83E3A982E3A682E3, 0x82E300A782E300B3, 0xB982E3A882E300A8,
-    0x83E3BC83E3AF82E3, 0xBC83E3A882E30089, 0xE300BC83E3AB82E3,
-    0xE300AA82E300A982, 0xB982E3B383E3AA82, 0xE3BC83E3AA82E300,
-    0xE300AB82E300A083, 0xAA83E3A482E3AB82, 0xE3A983E3AB82E300,
-    0x82E3008883E38383, 0xE3AA83E3AD83E3AB, 0xE300AC82E300BC83,
-    0xB383E3AD83E3AC82, 0xE3B383E3AC82E300, 0xE300AD82E3009E83,
-    0xAA83E3A583E3AD82, 0xE3AD82E300BC83E3, 0x83E3AD82E300AD83,
-    0xE3A983E3B082E3AD, 0x83E3AD82E300A083, 0xE3BC83E3A183E3AD,
-    0x82E300AB83E38883, 0xE3AF83E3AD83E3AD, 0x82E3008883E38383,
-    0xAC82E3AE82E300AE, 0xE38B83E3AE82E300, 0x83E3AE82E300BC83,
-    0x00BC83E38083E3AB, 0xE3AF82E300AF82E3, 0xA482E3BC82E3AB83,
-    0xE3AF82E300AD83E3, 0x8D83E3BC83E3AD83, 0xB082E300B082E300,
-    0xE300A083E3A983E3, 0xA083E3A983E3B082, 0xE300B383E38883E3,
-    0x83E3B182E300B182, 0xB282E300B982E3BC, 0xB382E300B382E300,
-    0xE3B382E300B382E3, 0x83E3B382E3008883, 0xB382E3008A83E3AB,
-    0xE3009D83E3BC83E3, 0xE300B582E300B482, 0xAF82E3A482E3B582,
-    0xE3B582E300AB83E3, 0xBC83E38183E3B383, 0x00B682E300A083E3,
-    0xE3B782E300B782E3, 0xB082E3B383E3AA83, 0xB982E300B882E300,
-    0xBB82E300BA82E300, 0xE3B383E3BB82E300, 0x83E3BB82E3008183,
-    0xBC82E3008883E3B3, 0xBE82E300BD82E300, 0x8083E300BF82E300,
-    0xE3BC83E38083E300, 0xE3008183E300B982, 0xE3008383E3008283,
-    0xE3008583E3008483, 0xE3008783E3008683, 0x83E300B782E38783,
-    0xB383E38883E30088, 0x8983E3008983E300, 0x008A83E300AB83E3,
-    0xE3008E83E38A83E3, 0xE3008C83E3008B83, 0xE3008E83E3008D83,
-    0x8883E38383E38E83, 0x8F83E3008F83E300, 0xE3008483E3A482E3,
-    0x83E39083E3009083, 0x00AB83E3AC83E3BC, 0xE39183E3009183E3,
-    0xB383E3BB82E3BC83, 0xE39183E3008883E3, 0x83E3008483E3BC83,
-    0x83E3009383E30092, 0x9483E300AB83E393, 0xE3A282E39483E300,
-    0xAB83E38883E3B982, 0xE3AF82E39483E300, 0x82E39483E300AB83,
-    0x83E3009583E300B3, 0xE3A983E3A182E395, 0x83E3008983E38383,
-    0xE3BC83E3A382E395, 0x83E39583E3008883, 0x9683E300B383E3A9,
-    0xE38383E39683E300, 0xAB83E3A782E3B782, 0x9883E3009783E300,
-    0xE3AF82E39883E300, 0xAB83E3BC83E3BF82, 0xE3AB83E39883E300,
-    0xE3009983E3008483, 0xBF82E3BC83E39983, 0x9A83E3009A83E300,
-    0xE39A83E300BD82E3, 0x83E3009283E38B83, 0x00B982E3B383E39A,
-    0x82E3BC83E39A83E3, 0x83E3009B83E300B8, 0x9B83E300B383E39B,
-    0xE300AB83E3BC83E3, 0xB383E3BC83E39B83, 0x9C83E3009C83E300,
-    0xE3008883E3AB83E3, 0x82E39D83E3009D83, 0x008883E3B383E3A4,
-    0x83E3B383E39D83E3, 0x83E3009E83E30089, 0xE3AF82E3A482E39E,
-    0x82E39E83E300AD83, 0x9E83E300AB83E3A4, 0xE3008F83E38383E3,
-    0xAF82E3AB83E39E83, 0xE3B383E39E83E300, 0xB383E3A783E3B782,
-    0x9F83E3009F83E300, 0x83E3AD83E3AF82E3, 0xAA83E39F83E300B3,
-    0xE3AA83E39F83E300, 0xAB83E3BC83E39083, 0xA183E300A083E300,
-    0x00AC82E3A183E300, 0x83E3AC82E3A183E3, 0xA183E300B383E388,
-    0x83E38883E3BC83E3, 0x83E300A283E300AB, 0x83E300A483E300A3,
-    0x008983E3BC83E3A4, 0x83E3BC83E3A483E3, 0x83E300A583E300AB,
-    0xA282E3A683E300A6, 0x00A783E300B383E3, 0x00A983E300A883E3,
-    0xE3AA83E300AA83E3, 0xAB83E38883E38383, 0x00A983E3AA83E300,
-    0xE3AB83E300AB83E3, 0x83E300BC83E39483, 0xE39683E3BC83E3AB,
-    0xE300AC83E300AB83, 0x83E300A083E3AC83, 0xE38883E3B383E3AC,
-    0x83E300B383E3B282, 0x83E300AF83E300AD, 0x008883E38383E3AF,
-    0x00B183E300B083E3, 0x00B383E300B283E3, 0x00B783E300B483E3,
-    0x00B983E300B883E3, 0x00BB83E300BA83E3, 0x00BE83E300BC83E3,
-    0x00B992E3009E92E3, 0x009F93E300BB92E3, 0x00AE9BE3009594E3,
-    0x00819EE300BC9BE3, 0x00A2A1E300AFA0E3, 0x0087A3E300BCA1E3,
-    0x009CA4E300A3A3E3, 0x00AEA8E300BAA4E3, 0x00A4ABE300ACA9E3,
-    0x0099ACE30088ACE3, 0x009DAEE30089ADE3, 0x008EB1E30098B0E3,
-    0x0096B6E300B3B4E3, 0x00B8BAE300ACBAE3, 0x00BCBFE3009BBCE3,
-    0x009880E4008880E4, 0x008681E400B980E4, 0x00A383E4009682E4,
-    0x008288E400AF84E4, 0x00A08AE400A788E4, 0x00B48CE400818CE4,
-    0x00958FE400998DE4, 0x008B90E400998FE4, 0x00AB94E400AB91E4,
-    0x00A195E4009D95E4, 0x009797E400AB95E4, 0x00B598E400B997E4,
-    0x00879BE400BE9AE4, 0x00A6A7E40095A6E4, 0x00B6A9E400AEA9E4,
-    0x00B3ACE400B2AAE4, 0x008EB3E4008EAFE4, 0x00B8B3E400ADB3E4,
-    0x0080B8E40096B5E4, 0x0083B8E40081B8E4, 0x008AB8E40089B8E4,
-    0x008DB8E4008BB8E4, 0x00A6B8E40099B8E4, 0x00ADB8E400A8B8E4,
-    0x00B6B8E400B2B8E4, 0x00B9B8E400B8B8E4, 0x00BFB8E400BDB8E4,
-    0x0099B9E40081B9E4, 0x0082BAE4009DB9E4, 0x0086BAE40085BAE4,
-    0x0094BAE4008CBAE4, 0x00A4BAE400A0BAE4, 0x00BABAE400AEBAE4,
-    0x008CBBE40080BBE4, 0xE5A4BBE400A4BBE4, 0xE40081BCE4008C92,
-    0xE400A0BDE40091BC, 0xE40086BEE40080BE, 0xE400AEBEE4008BBE,
-    0xE500BFBEE400BBBE, 0xE500AB80E5008280, 0xE5009982E500BA81,
-    0xE5009A83E5008F83, 0xE500AA84E500A783, 0xE5008085E500BF84,
-    0xE5008D85E5008585, 0xE500A485E5009485, 0xE500A785E500A585,
-    0xE500A985E500A885, 0xE500AD85E500AB85, 0xE5008086E500B785,
-    0xE5008D86E5008286, 0xE5009586E5009286, 0xE5009786E5009686,
-    0xE500A486E5009986, 0xE500AC86E500AB86, 0xE500B786E500B586,
-    0xE5008C87E5008987, 0xE5009E87E5009C87, 0xE500B587E500A087,
-    0xE5008388E5008088, 0xE5009788E5008788, 0xE500A988E5009D88,
-    0xE500BB88E500BA88, 0xE5008D89E5008689, 0xE500B789E500B289,
-    0xE5009B8AE500898A, 0xE500B38AE500A38A, 0xE500878BE500B48A,
-    0xE500928BE500898B, 0xE500A48BE5009E8B, 0xE500B98BE500B58B,
-    0xE500858CE500BA8B, 0xE500958CE500868C, 0xE5009A8CE500978C,
-    0xE500BB8CE500B88C, 0xE500818DE500BF8C, 0xE500858DE500848D,
-    0xE500918DE500898D, 0xE5009A8DE500948D, 0xE500A98DE5009C8D,
-    0xE500B38DE500B08D, 0xE500BD8DE500B58D, 0xE500828EE500BF8D,
-    0xE500838FE500B68E, 0xE5008A8FE500888F, 0xE5009F8FE5008C8F,
-    0xE500A58FE500A38F, 0xE500AF8FE500AB8F, 0xE500B38FE500B18F,
-    0xE5008890E5008690, 0xE5008F90E5008D90, 0xE500B890E5009D90,
-    0xE5008291E500B990, 0xE500A891E5008891, 0xE500A292E5009E92,
-    0xE500B693E500BD92, 0xE5008F95E5009094, 0xE5009595E5009395,
-    0xE5008496E500A395, 0xE5009996E5008796, 0xE500AB96E5009D96,
-    0xE500B696E500B396, 0xE5008297E5008097, 0xE5008698E500A297,
-    0xE500A899E5009199, 0xE500979BE500B499, 0xE500B99BE5009B9B,
-    0xE500979CE500969C, 0xE500B09CE5009F9C, 0xE5008E9FE5008B9E,
-    0xE5008DA0E500B49F, 0xE500B2A0E500B1A0, 0xE5009AA1E50080A1,
-    0xE500A8A2E5009EA1, 0xE500B3A2E500ACA2, 0xE5009FA3E50098A3,
-    0xE500AEA3E500ABA3, 0xE500B2A3E500B0A3, 0xE50082A4E500B7A3,
-    0xE5008AA4E50086A4, 0xE5009AA4E50095A4, 0xE500A2A4E5009CA4,
-    0xADE6A7A4E500A7A4, 0xA5E500A9A4E500A3, 0xA5E50088A5E50084,
-    0xA5E50094A5E50091, 0xA7E500B3A5E500A2, 0xA8E500ACA7E50098,
-    0xA9E500A7A8E5009B, 0xAAE500A6A9E500A2, 0xACE50088ACE500B5,
-    0xADE500BEACE500A8, 0xADE50097ADE50090, 0xAEE50080AEE500A6,
-    0xAFE50097AEE50085, 0xAFE50098AFE50083, 0xAFE500AEAFE500A7,
-    0xAFE500B8AFE500B3, 0xB0E50086B0E500BF, 0xB0E500A2B0E5008F,
-    0xB1E500BFB0E500B8, 0xB1E500A2B1E500A0, 0xB1E500A5B1E500A4,
-    0xB2E500B1B1E500AE, 0xB4E50080B3E5008D, 0xB5E50083B5E50099,
-    0xB5E500ABB5E50090, 0xB6E500BCB5E500AE, 0xB7E500BAB6E500B2,
-    0xB7E500A1B7E5009B, 0xB7E500A5B7E500A2, 0xB7E500B1B7E500A6,
-    0xB8E500BEB7E500BD, 0xB9E500BDB8E500A8, 0xB9E500B2B9E500A9,
-    0xB4B9E5009088E6B3, 0xBCB9E500BAB9E500, 0xA6BAE500BFB9E500,
-    0xB3BAE500B0BAE500, 0x89BBE500B6BAE500, 0x92BBE5008ABBE500,
-    0x99BBE50093BBE500, 0xB4BBE500ACBBE500, 0x84BCE500BEBBE500,
-    0x93BCE5008BBCE500, 0x90BDE500A2BCE500, 0xA1BDE50093BDE500,
-    0xA9BDE500A2BDE500, 0xB3BDE500ABBDE500, 0x8CBEE5008BBEE500,
-    0x9ABEE50097BEE500, 0xADBEE500A9BEE500, 0x8DBFE50083BFE500,
-    0xB5BFE50097BFE500, 0x9280E600B9BFE500, 0xB581E6009C80E600,
-    0x9482E6008182E600, 0x9883E6008783E600, 0x8884E600A183E600,
-    0x8885E6008485E600, 0x8E85E6008C85E600, 0xA885E600A085E600,
-    0x8E86E600BA85E600, 0xA486E6009086E600, 0xB286E600AF86E600,
-    0xB287E6009E87E600, 0x8088E600B687E600, 0x9088E6008888E600,
-    0xAE88E6009B88E600, 0xB688E600B488E600, 0x9389E6008B89E600,
-    0x958AE6009D89E600, 0x898BE600B18AE600, 0x938BE6008F8BE600,
-    0xBC8BE600948BE600, 0x878CE600BE8BE600, 0x908DE600BD8CE600,
-    0xA88DE600958DE600, 0x838EE600BB8DE600, 0xA98EE600A08EE600,
-    0x858FE600848FE600, 0x9C90E600A48FE600, 0x9291E600A290E600,
-    0xB791E600A991E600, 0x9A92E600BE91E600, 0x8493E6009D92E600,
-    0xB494E600AF94E600, 0x9695E6008F95E600, 0xB895E600AC95E600,
-    0x9796E6008796E600, 0xA496E6009996E600, 0xB996E600B096E600,
-    0xA097E6008597E600, 0xA397E600A297E600, 0x8E98E600A597E600,
-    0x009398E600BBB2E6, 0xE5AD98E600A098E6, 0xE6008999E6008C92,
-    0xE600889AE600B499, 0xE6009C9AE600919A, 0xE600869BE600B49A,
-    0xE600B49BE600B09B, 0xE600809CE600B89B, 0xE600899CE600889C,
-    0xE6009B9CE600979C, 0xE600A89CE600A19C, 0xE600939DE6008E9D,
-    0xE6009E9DE600969D, 0xE600859EE600BB9D, 0xE600B39FE600979E,
-    0xE60097A0E600BA9F, 0xE600AAA0E6009FA0, 0x9ABCE48FBCE5AAA0,
-    0x0092A1E600BEA4E7, 0x0085A2E60081A2E6, 0x00A8A2E6008EA2E6,
-    0x0082A5E60094A4E6, 0x00AAA7E600A3A6E6, 0x0093A8E60082A8E6,
-    0x0093ABE600A8AAE6, 0x0084ACE6009BABE6, 0x00A1ACE600A0ACE6,
-    0x00A2ADE60094ADE6, 0x00B2ADE600A3ADE6, 0x00B9ADE600B7ADE6,
-    0x00AEAEE6009FAEE6, 0x00BAAEE600B3AEE6, 0x008BAFE600BBAEE6,
-    0x0094AFE6008DAFE6, 0x008FB0E6009BAFE6, 0x00B4B0E60094B0E6,
-    0x00A7B1E6008EB1E6, 0x00BFB2E60088B2E6, 0x008DB3E6008CB3E6,
-    0x00A8B3E600A5B3E6, 0x009BB4E60096B4E6, 0x00B4B4E6009EB4E6,
-    0x0081B5E600BEB4E6, 0x00AAB5E600A9B5E6, 0x00B8B5E600B7B5E6,
-    0x008BB7E60085B6E6, 0x00AAB7E6009AB7E6, 0x009AB8E600B9B7E6,
-    0x00AEB9E600AFB8E6, 0x009CBAE60080BAE6, 0x0087BBE600BABAE6,
-    0x0091BBE6008BBBE6, 0x008FBCE6009BBBE6, 0x00A2BCE60094BCE6,
-    0x00AEBDE600A3BCE6, 0x00ABBFE60086BFE6, 0x009B80E700BEBFE6,
-    0x00B980E7009E80E7, 0x00AB81E7008A81E7, 0x00B781E700B081E7,
-    0x009982E700BD81E7, 0x008883E700AD82E7, 0x00A184E7009983E7,
-    0x008985E7008585E7, 0x009C86E700AE85E7, 0x009087E7008E87E7,
-    0x009B88E7009088E7, 0x00AA88E700A888E7, 0x00B588E700AB88E7,
-    0x00BB88E700B688E7, 0x008789E700BF88E7, 0x009989E7009089E7,
-    0x00A289E7009B89E7, 0x00808AE700B989E7, 0x00AC8AE700958AE7,
-    0x00808BE700AF8AE7, 0x00AA8CE700BC8BE7, 0x00BA8DE700B58DE7,
-    0x00878EE700848EE7, 0x008B8EE700898EE7, 0x00B28EE700A58EE7,
-    0x008690E7009E8FE7, 0x00A290E7008990E7, 0x009C91E7008791E7,
-    0x00B191E700A991E7, 0x008992E7008592E7, 0x008A93E7009892E7,
-    0x00A693E7009C93E7, 0x009894E7008694E7, 0x00A494E7009F94E7,
-    0x00B094E700A894E7, 0x00B394E700B294E7, 0x00BB94E700B794E7,
-    0x009995E700BE94E7, 0x00B095E700A595E7, 0x009296E7008B96E7,
-    0x009098E700A297E7, 0x009F98E7009D98E7, 0x00A999E7008299E7,
-    0x00BD99E700B699E7, 0x00BF9AE700AE9AE7, 0x009B9BE7008A9BE7,
-    0x00A79BE700A39BE7, 0x00B49BE700AE9BE7, 0x009E9CE700819CE7,
-    0x00809DE7009F9CE7, 0x008B9EE7008A9DE7, 0x009B9FE700A79EE7,
-    0x00B39FE700A29FE7, 0x00ABA1E7008EA1E7, 0x0091A2E7008CA2E7,
-    0x008CA3E7008AA3E7, 0x00AAA4E700BBA3E7, 0x00BCA4E700BAA4E7,
-    0x0088A5E700BEA4E7, 0x0090A5E70089A5E7, 0x009DA5E70096A5E7,
-    0x00A5A5E7009EA5E7, 0x0081A6E700BFA5E7, 0x008EA6E7008DA6E7,
-    0x00AEA6E7008FA6E7, 0x00BEA6E700B8A6E7, 0x0098A7E7008AA7E7,
-    0x009CA8E700ABA7E7, 0x008AA9E70080A9E7, 0x00B4A9E7008FA9E7,
-    0x0081AAE700BAA9E7, 0x008BABE700B1AAE7, 0x00B9ABE700AEABE7,
-    0x008FAEE700A0ACE7, 0x0086AFE70080AFE7, 0x00BEB0E70089AFE7,
-    0x00B3B1E700A0B1E7, 0x0092B2E700BBB1E7, 0x0092B3E700BEB2E7,
-    0x00A3B3E70096B3E7, 0x00A8B3E700A7B3E7, 0x0080B4E700B8B3E7,
-    0x00A2B4E70090B4E7, 0x0082B5E700AFB4E7, 0x00A3B5E7009BB5E7,
-    0x00BEB6E700A0B6E7, 0x00B4B7E70087B7E7, 0x0089B8E70082B8E7,
-    0x0081B9E700B7B8E7, 0x00B6BCE70085B9E7, 0x0091BDE700BEBCE7,
-    0x00B9BDE700B2BDE7, 0x0085BEE700BABDE7, 0x0095BEE7008ABEE7,
-    0x00BDBEE7009ABEE7, 0x008180E800BABFE7, 0x008C80E8008580E8,
-    0x00B380E8009280E8, 0x00A081E8008681E8, 0x00B081E800AF81E8,
-    0x00BF81E800BE81E8, 0x008B82E8008982E8, 0x00B282E800AD82E8,
-    0x00BE84E8008384E8, 0x00A387E8009887E8, 0x00AA87E800A887E8,
-    0x00B387E800AD87E8, 0x008188E800BC87E8, 0x008C88E8008488E8,
-    0x009B88E8009888E8, 0x00AE89E8009F88E8, 0x00B289E800AF89E8,
-    0x00B989E800B889E8, 0x00918AE8008B8AE8, 0x00B18AE8009D8AE8,
-    0x00BD8AE800B38AE8, 0x00A68BE800A58BE8, 0x00A38CE8009D8CE8,
-    0x00928DE800B68CE8, 0x00A38DE800938DE8, 0x00BD8EE800AD8EE8,
-    0x008A8FE800898FE8, 0x009C8FE8008C8FE8, 0x00AF8FE800A78FE8,
-    0x00BD90E800B18FE8, 0x009791E8008991E8, 0x00B193E800AE93E8,
-    0x00BC93E800B393E8, 0x00A495E8009694E8, 0x00BA97E8008D97E8,
-    0x009298E8008698E8, 0x00BF98E800AD98E8, 0x009099E8008D99E8,
-    0x00A799E8009C99E8, 0x00AB99E800A999E8, 0x00A99AE800889AE8,
-    0x008E9CE800A29BE8, 0x00AB9DE800A89CE8, 0x00869EE800B99DE8,
-    0x00A19FE800BA9EE8, 0x009FA0E80081A0E8, 0x008CA1E80080A1E8,
-    0x00A3A1E800A0A1E8, 0x008FA3E80082A3E8, 0x009EA3E80097A3E8,
-    0x00B8A3E800A1A3E8, 0x0090A4E800BAA3E8, 0x00A4A5E80081A5E8,
-    0x0086A6E800BEA5E8, 0x0096A6E8008BA6E8, 0x00A3A7E80092A7E8,
-    0x00A0AAE80080A8E8, 0x00BFAAE800AAAAE8, 0x0092ABE8008BABE8,
-    0x00ADABE80096ABE8, 0x00BEABE800B8ABE8, 0x00B9ACE80081ACE8,
-    0x0080AEE80098ADE8, 0x00B7B0E8008AAEE8, 0x0088B1E80086B1E8,
-    0x00B8B1E80095B1E8, 0x00A1B2E8009DB2E8, 0x00ABB2E800A9B2E8,
-    0x0082B3E80081B3E8, 0x0088B3E80087B3E8, 0x0088B4E80093B3E8,
-    0x00A4B5E8009BB4E8, 0x00B7B5E800B0B5E8, 0x00BCB6E800B3B6E8,
-    0x00AFB7E8008BB7E8, 0x00ABBAE800B0B7E8, 0x0094BBE8008ABBE8,
-    0x00AABCE800A6BCE8, 0x00BBBCE800B8BCE8, 0x009BBEE800A2BDE8,
-    0x00B0BEE8009EBEE8, 0x00B6BEE800B5BEE8, 0x00B880E900A380E9,
-    0x00A981E9008A81E9, 0x00BC81E900B281E9, 0x009182E9008F82E9,
-    0x008E83E9009482E9, 0x00B183E9009E83E9, 0x009184E900BD83E9,
-    0x008985E9009B84E9, 0x00AA85E9008D85E9, 0x00B486E9009986E9,
-    0x008C87E9008687E9, 0x009187E9008F87E9, 0x00B888E900B488E9,
-    0x00BC89E900B689E9, 0x00988BE900978BE9, 0x008A8DE900848CE9,
-    0x009590E900B98FE9, 0x008096E900B795E9, 0x00AD96E9008B96E9,
-    0x009C98E900B796E9, 0x008B99E900AE98E9, 0x00B599E9008D99E9,
-    0x00BC99E900B899E9, 0x00A39AE900869AE9, 0x00B79AE900B69AE9,
-    0x00B99AE900B89AE9, 0x00A29BE900839BE9, 0x00A89BE900A39BE9,
-    0x00B79BE900B69BE9, 0x00B29CE900A39CE9, 0x00919DE900889DE9,
-    0x009E9DE900969DE9, 0x00A99DE900A29DE9, 0x009B9FE9008B9FE9,
-    0x00AD9FE900A09FE9, 0x00BF9FE900B39FE9, 0x0085A0E90081A0E9,
-    0x0098A0E9008BA0E9, 0x00BBA0E900A9A0E9, 0x00A8A2E9009EA1E9,
-    0x009FA3E9009BA3E9, 0x00AFA3E900A2A3E9, 0x00A8A4E900BCA3E9,
-    0x0096A6E900A9A4E9, 0x00A7A6E90099A6E9, 0x0082A7E900ACA6E9,
-    0x00BEA7E900B1A7E9, 0x00A8AAE900AAA9E9, 0x009FABE90098ABE9,
-    0x00A5ACE90092ACE9, 0x00B2ACE900AFACE9, 0x009AADE900BCACE9,
-    0x0080B1E900AFADE9, 0x00A5B3E90097B1E9, 0x00A7B5E900BDB3E9,
-    0x00BAB7E900B4B6E9, 0x00B5B9E9009EB8E9, 0x0097BAE900BFB9E9,
-    0x00A5BAE9009FBAE9, 0x0083BBE900BBBAE9, 0x008EBBE9008DBBE9,
-    0x00B9BBE90091BBE9, 0x00BEBBE900BDBBE9, 0x008EBCE90085BCE9,
-    0x0093BCE9008FBCE9, 0x00A0BCE90096BCE9, 0x0083BDE900BBBCE9,
-    0x0092BDE9008ABDE9, 0x008EBEE9008DBEE9, 0x009FBEE9009CBEE9,
-    0x009199EA00A0BEE9, 0x00A79CEA00899AEA, 0x008E9EEA00AF9DEA,
-    0x0092ADEA00B7ACEA, 0x00A7ADEA00A6ADEA, 0x0081B0EA0080B0EA,
-    0x0083B0EA0082B0EA, 0x0085B0EA0084B0EA, 0x0087B0EA0086B0EA,
-    0x0089B0EA0088B0EA, 0x008BB0EA008AB0EA, 0x008DB0EA008CB0EA,
-    0x008FB0EA008EB0EA, 0x0091B0EA0090B0EA, 0x0093B0EA0092B0EA,
-    0x0095B0EA0094B0EA, 0x0097B0EA0096B0EA, 0x0099B0EA0098B0EA,
-    0x009BB0EA009AB0EA, 0x009DB0EA009CB0EA, 0x009FB0EA009EB0EA,
-    0x00A1B0EA00A0B0EA, 0x00A3B0EA00A2B0EA, 0x00A5B0EA00A4B0EA,
-    0x00A7B0EA00A6B0EA, 0x00A9B0EA00A8B0EA, 0x00ABB0EA00AAB0EA,
-    0x00ADB0EA00ACB0EA, 0x00AFB0EA00AEB0EA, 0x00B1B0EA00B0B0EA,
-    0x00B3B0EA00B2B0EA, 0x00B5B0EA00B4B0EA, 0x00B7B0EA00B6B0EA,
-    0x00B9B0EA00B8B0EA, 0x00BBB0EA00BAB0EA, 0x00BDB0EA00BCB0EA,
-    0x00BFB0EA00BEB0EA, 0x0081B1EA0080B1EA, 0x0083B1EA0082B1EA,
-    0x0085B1EA0084B1EA, 0x0087B1EA0086B1EA, 0x0089B1EA0088B1EA,
-    0x008BB1EA008AB1EA, 0x008DB1EA008CB1EA, 0x008FB1EA008EB1EA,
-    0x0091B1EA0090B1EA, 0x0093B1EA0092B1EA, 0x0095B1EA0094B1EA,
-    0x0097B1EA0096B1EA, 0x0099B1EA0098B1EA, 0x009BB1EA009AB1EA,
-    0x009DB1EA009CB1EA, 0x009FB1EA009EB1EA, 0x00A1B1EA00A0B1EA,
-    0x00A3B1EA00A2B1EA, 0x00A5B1EA00A4B1EA, 0x00A7B1EA00A6B1EA,
-    0x00A9B1EA00A8B1EA, 0x00ABB1EA00AAB1EA, 0x00ADB1EA00ACB1EA,
-    0x00AFB1EA00AEB1EA, 0x00B1B1EA00B0B1EA, 0x00B3B1EA00B2B1EA,
-    0x00B5B1EA00B4B1EA, 0x00B7B1EA00B6B1EA, 0x00B9B1EA00B8B1EA,
-    0x00BBB1EA00BAB1EA, 0x00BDB1EA00BCB1EA, 0x00BFB1EA00BEB1EA,
-    0x0081B2EA0080B2EA, 0x0083B2EA0082B2EA, 0x0085B2EA0084B2EA,
-    0x0087B2EA0086B2EA, 0x0089B2EA0088B2EA, 0x008BB2EA008AB2EA,
-    0x008DB2EA008CB2EA, 0x008FB2EA008EB2EA, 0x0091B2EA0090B2EA,
-    0x0093B2EA0092B2EA, 0x0095B2EA0094B2EA, 0x0097B2EA0096B2EA,
-    0x0099B2EA0098B2EA, 0x009BB2EA009AB2EA, 0x009DB2EA009CB2EA,
-    0x009FB2EA009EB2EA, 0x00A1B2EA00A0B2EA, 0x00A3B2EA00A2B2EA,
-    0x00A5B2EA00A4B2EA, 0x00A7B2EA00A6B2EA, 0x00A9B2EA00A8B2EA,
-    0x00ABB2EA00AAB2EA, 0x00ADB2EA00ACB2EA, 0x00AFB2EA00AEB2EA,
-    0x00B1B2EA00B0B2EA, 0x00B3B2EA00B2B2EA, 0x00B5B2EA00B4B2EA,
-    0x00B7B2EA00B6B2EA, 0x00B9B2EA00B8B2EA, 0x00BBB2EA00BAB2EA,
-    0x00BDB2EA00BCB2EA, 0x00BFB2EA00BEB2EA, 0x0081B3EA0080B3EA,
-    0x0083B3EA0082B3EA, 0x0085B3EA0084B3EA, 0x0087B3EA0086B3EA,
-    0x0089B3EA0088B3EA, 0x008BB3EA008AB3EA, 0x008DB3EA008CB3EA,
-    0x008FB3EA008EB3EA, 0x0091B3EA0090B3EA, 0x0093B3EA0092B3EA,
-    0x0095B3EA0094B3EA, 0x0097B3EA0096B3EA, 0x0099B3EA0098B3EA,
-    0x009BB3EA009AB3EA, 0x009DB3EA009CB3EA, 0x009FB3EA009EB3EA,
-    0x00A1B3EA00A0B3EA, 0x00A3B3EA00A2B3EA, 0x00A5B3EA00A4B3EA,
-    0x00A7B3EA00A6B3EA, 0x00A9B3EA00A8B3EA, 0x00ABB3EA00AAB3EA,
-    0x00ADB3EA00ACB3EA, 0x00AFB3EA00AEB3EA, 0x00B1B3EA00B0B3EA,
-    0x00B3B3EA00B2B3EA, 0x00B5B3EA00B4B3EA, 0x00B7B3EA00B6B3EA,
-    0x00B9B3EA00B8B3EA, 0x00BBB3EA00BAB3EA, 0x00BDB3EA00BCB3EA,
-    0x00BFB3EA00BEB3EA, 0x0081B4EA0080B4EA, 0x0083B4EA0082B4EA,
-    0x0085B4EA0084B4EA, 0x0087B4EA0086B4EA, 0x0089B4EA0088B4EA,
-    0x008BB4EA008AB4EA, 0x008DB4EA008CB4EA, 0x008FB4EA008EB4EA,
-    0x0091B4EA0090B4EA, 0x0093B4EA0092B4EA, 0x0095B4EA0094B4EA,
-    0x0097B4EA0096B4EA, 0x0099B4EA0098B4EA, 0x009BB4EA009AB4EA,
-    0x009DB4EA009CB4EA, 0x009FB4EA009EB4EA, 0x00A1B4EA00A0B4EA,
-    0x00A3B4EA00A2B4EA, 0x00A5B4EA00A4B4EA, 0x00A7B4EA00A6B4EA,
-    0x00A9B4EA00A8B4EA, 0x00ABB4EA00AAB4EA, 0x00ADB4EA00ACB4EA,
-    0x00AFB4EA00AEB4EA, 0x00B1B4EA00B0B4EA, 0x00B3B4EA00B2B4EA,
-    0x00B5B4EA00B4B4EA, 0x00B7B4EA00B6B4EA, 0x00B9B4EA00B8B4EA,
-    0x00BBB4EA00BAB4EA, 0x00BDB4EA00BCB4EA, 0x00BFB4EA00BEB4EA,
-    0x0081B5EA0080B5EA, 0x0083B5EA0082B5EA, 0x0085B5EA0084B5EA,
-    0x0087B5EA0086B5EA, 0x0089B5EA0088B5EA, 0x008BB5EA008AB5EA,
-    0x008DB5EA008CB5EA, 0x008FB5EA008EB5EA, 0x0091B5EA0090B5EA,
-    0x0093B5EA0092B5EA, 0x0095B5EA0094B5EA, 0x0097B5EA0096B5EA,
-    0x0099B5EA0098B5EA, 0x009BB5EA009AB5EA, 0x009DB5EA009CB5EA,
-    0x009FB5EA009EB5EA, 0x00A1B5EA00A0B5EA, 0x00A3B5EA00A2B5EA,
-    0x00A5B5EA00A4B5EA, 0x00A7B5EA00A6B5EA, 0x00A9B5EA00A8B5EA,
-    0x00ABB5EA00AAB5EA, 0x00ADB5EA00ACB5EA, 0x00AFB5EA00AEB5EA,
-    0x00B1B5EA00B0B5EA, 0x00B3B5EA00B2B5EA, 0x00B5B5EA00B4B5EA,
-    0x00B7B5EA00B6B5EA, 0x00B9B5EA00B8B5EA, 0x00BBB5EA00BAB5EA,
-    0x00BDB5EA00BCB5EA, 0x00BFB5EA00BEB5EA, 0x0081B6EA0080B6EA,
-    0x0083B6EA0082B6EA, 0x0085B6EA0084B6EA, 0x0087B6EA0086B6EA,
-    0x0089B6EA0088B6EA, 0x008BB6EA008AB6EA, 0x008DB6EA008CB6EA,
-    0x008FB6EA008EB6EA, 0x0091B6EA0090B6EA, 0x0093B6EA0092B6EA,
-    0x0095B6EA0094B6EA, 0x0097B6EA0096B6EA, 0x0099B6EA0098B6EA,
-    0x009BB6EA009AB6EA, 0x009DB6EA009CB6EA, 0x009FB6EA009EB6EA,
-    0x00A1B6EA00A0B6EA, 0x00A3B6EA00A2B6EA, 0x00A5B6EA00A4B6EA,
-    0x00A7B6EA00A6B6EA, 0x00A9B6EA00A8B6EA, 0x00ABB6EA00AAB6EA,
-    0x00ADB6EA00ACB6EA, 0x00AFB6EA00AEB6EA, 0x00B1B6EA00B0B6EA,
-    0x00B3B6EA00B2B6EA, 0x00B5B6EA00B4B6EA, 0x00B7B6EA00B6B6EA,
-    0x00B9B6EA00B8B6EA, 0x00BBB6EA00BAB6EA, 0x00BDB6EA00BCB6EA,
-    0x00BFB6EA00BEB6EA, 0x0081B7EA0080B7EA, 0x0083B7EA0082B7EA,
-    0x0085B7EA0084B7EA, 0x0087B7EA0086B7EA, 0x0089B7EA0088B7EA,
-    0x008BB7EA008AB7EA, 0x008DB7EA008CB7EA, 0x008FB7EA008EB7EA,
-    0x0091B7EA0090B7EA, 0x0093B7EA0092B7EA, 0x0095B7EA0094B7EA,
-    0x0097B7EA0096B7EA, 0x0099B7EA0098B7EA, 0x009BB7EA009AB7EA,
-    0x009DB7EA009CB7EA, 0x009FB7EA009EB7EA, 0x00A1B7EA00A0B7EA,
-    0x00A3B7EA00A2B7EA, 0x00A5B7EA00A4B7EA, 0x00A7B7EA00A6B7EA,
-    0x00A9B7EA00A8B7EA, 0x00ABB7EA00AAB7EA, 0x00ADB7EA00ACB7EA,
-    0x00AFB7EA00AEB7EA, 0x00B1B7EA00B0B7EA, 0x00B3B7EA00B2B7EA,
-    0x00B5B7EA00B4B7EA, 0x00B7B7EA00B6B7EA, 0x00B9B7EA00B8B7EA,
-    0x00BBB7EA00BAB7EA, 0x00BDB7EA00BCB7EA, 0x00BFB7EA00BEB7EA,
-    0x0081B8EA0080B8EA, 0x0083B8EA0082B8EA, 0x0085B8EA0084B8EA,
-    0x0087B8EA0086B8EA, 0x0089B8EA0088B8EA, 0x008BB8EA008AB8EA,
-    0x008DB8EA008CB8EA, 0x008FB8EA008EB8EA, 0x0091B8EA0090B8EA,
-    0x0093B8EA0092B8EA, 0x0095B8EA0094B8EA, 0x0097B8EA0096B8EA,
-    0x0099B8EA0098B8EA, 0x009BB8EA009AB8EA, 0x009DB8EA009CB8EA,
-    0x009FB8EA009EB8EA, 0x00A1B8EA00A0B8EA, 0x00A3B8EA00A2B8EA,
-    0x00A5B8EA00A4B8EA, 0x00A7B8EA00A6B8EA, 0x00A9B8EA00A8B8EA,
-    0x00ABB8EA00AAB8EA, 0x00ADB8EA00ACB8EA, 0x00AFB8EA00AEB8EA,
-    0x00B1B8EA00B0B8EA, 0x00B3B8EA00B2B8EA, 0x00B5B8EA00B4B8EA,
-    0x00B7B8EA00B6B8EA, 0x00B9B8EA00B8B8EA, 0x00BBB8EA00BAB8EA,
-    0x00BDB8EA00BCB8EA, 0x00BFB8EA00BEB8EA, 0x0081B9EA0080B9EA,
-    0x0083B9EA0082B9EA, 0x0085B9EA0084B9EA, 0x0087B9EA0086B9EA,
-    0x0089B9EA0088B9EA, 0x008BB9EA008AB9EA, 0x008DB9EA008CB9EA,
-    0x008FB9EA008EB9EA, 0x0091B9EA0090B9EA, 0x0093B9EA0092B9EA,
-    0x0095B9EA0094B9EA, 0x0097B9EA0096B9EA, 0x0099B9EA0098B9EA,
-    0x009BB9EA009AB9EA, 0x009DB9EA009CB9EA, 0x009FB9EA009EB9EA,
-    0x00A1B9EA00A0B9EA, 0x00A3B9EA00A2B9EA, 0x00A5B9EA00A4B9EA,
-    0x00A7B9EA00A6B9EA, 0x00A9B9EA00A8B9EA, 0x00ABB9EA00AAB9EA,
-    0x00ADB9EA00ACB9EA, 0x00AFB9EA00AEB9EA, 0x00B1B9EA00B0B9EA,
-    0x00B3B9EA00B2B9EA, 0x00B5B9EA00B4B9EA, 0x00B7B9EA00B6B9EA,
-    0x00B9B9EA00B8B9EA, 0x00BBB9EA00BAB9EA, 0x00BDB9EA00BCB9EA,
-    0x00BFB9EA00BEB9EA, 0x0081BAEA0080BAEA, 0x0083BAEA0082BAEA,
-    0x0085BAEA0084BAEA, 0x0087BAEA0086BAEA, 0x0089BAEA0088BAEA,
-    0x008BBAEA008ABAEA, 0x008DBAEA008CBAEA, 0x008FBAEA008EBAEA,
-    0x0091BAEA0090BAEA, 0x0093BAEA0092BAEA, 0x0095BAEA0094BAEA,
-    0x0097BAEA0096BAEA, 0x0099BAEA0098BAEA, 0x009BBAEA009ABAEA,
-    0x009DBAEA009CBAEA, 0x009FBAEA009EBAEA, 0x00A1BAEA00A0BAEA,
-    0x00A3BAEA00A2BAEA, 0x00A5BAEA00A4BAEA, 0x00A7BAEA00A6BAEA,
-    0x00A9BAEA00A8BAEA, 0x00ABBAEA00AABAEA, 0x00ADBAEA00ACBAEA,
-    0x00AFBAEA00AEBAEA, 0x00B1BAEA00B0BAEA, 0x00B3BAEA00B2BAEA,
-    0x00B5BAEA00B4BAEA, 0x00B7BAEA00B6BAEA, 0x00B9BAEA00B8BAEA,
-    0x00BBBAEA00BABAEA, 0x00BDBAEA00BCBAEA, 0x00BFBAEA00BEBAEA,
-    0x0081BBEA0080BBEA, 0x0083BBEA0082BBEA, 0x0085BBEA0084BBEA,
-    0x0087BBEA0086BBEA, 0x0089BBEA0088BBEA, 0x008BBBEA008ABBEA,
-    0x008DBBEA008CBBEA, 0x008FBBEA008EBBEA, 0x0091BBEA0090BBEA,
-    0x0093BBEA0092BBEA, 0x0095BBEA0094BBEA, 0x0097BBEA0096BBEA,
-    0x0099BBEA0098BBEA, 0x009BBBEA009ABBEA, 0x009DBBEA009CBBEA,
-    0x009FBBEA009EBBEA, 0x00A1BBEA00A0BBEA, 0x00A3BBEA00A2BBEA,
-    0x00A5BBEA00A4BBEA, 0x00A7BBEA00A6BBEA, 0x00A9BBEA00A8BBEA,
-    0x00ABBBEA00AABBEA, 0x00ADBBEA00ACBBEA, 0x00AFBBEA00AEBBEA,
-    0x00B1BBEA00B0BBEA, 0x00B3BBEA00B2BBEA, 0x00B5BBEA00B4BBEA,
-    0x00B7BBEA00B6BBEA, 0x00B9BBEA00B8BBEA, 0x00BBBBEA00BABBEA,
-    0x00BDBBEA00BCBBEA, 0x00BFBBEA00BEBBEA, 0x0081BCEA0080BCEA,
-    0x0083BCEA0082BCEA, 0x0085BCEA0084BCEA, 0x0087BCEA0086BCEA,
-    0x0089BCEA0088BCEA, 0x008BBCEA008ABCEA, 0x008DBCEA008CBCEA,
-    0x008FBCEA008EBCEA, 0x0091BCEA0090BCEA, 0x0093BCEA0092BCEA,
-    0x0095BCEA0094BCEA, 0x0097BCEA0096BCEA, 0x0099BCEA0098BCEA,
-    0x009BBCEA009ABCEA, 0x009DBCEA009CBCEA, 0x009FBCEA009EBCEA,
-    0x00A1BCEA00A0BCEA, 0x00A3BCEA00A2BCEA, 0x00A5BCEA00A4BCEA,
-    0x00A7BCEA00A6BCEA, 0x00A9BCEA00A8BCEA, 0x00ABBCEA00AABCEA,
-    0x00ADBCEA00ACBCEA, 0x00AFBCEA00AEBCEA, 0x00B1BCEA00B0BCEA,
-    0x00B3BCEA00B2BCEA, 0x00B5BCEA00B4BCEA, 0x00B7BCEA00B6BCEA,
-    0x00B9BCEA00B8BCEA, 0x00BBBCEA00BABCEA, 0x00BDBCEA00BCBCEA,
-    0x00BFBCEA00BEBCEA, 0x0081BDEA0080BDEA, 0x0083BDEA0082BDEA,
-    0x0085BDEA0084BDEA, 0x0087BDEA0086BDEA, 0x0089BDEA0088BDEA,
-    0x008BBDEA008ABDEA, 0x008DBDEA008CBDEA, 0x008FBDEA008EBDEA,
-    0x0091BDEA0090BDEA, 0x0093BDEA0092BDEA, 0x0095BDEA0094BDEA,
-    0x0097BDEA0096BDEA, 0x0099BDEA0098BDEA, 0x009BBDEA009ABDEA,
-    0x009DBDEA009CBDEA, 0x009FBDEA009EBDEA, 0x00A1BDEA00A0BDEA,
-    0x00A3BDEA00A2BDEA, 0x00A5BDEA00A4BDEA, 0x00A7BDEA00A6BDEA,
-    0x00A9BDEA00A8BDEA, 0x00ABBDEA00AABDEA, 0x00ADBDEA00ACBDEA,
-    0x00AFBDEA00AEBDEA, 0x00B1BDEA00B0BDEA, 0x00B3BDEA00B2BDEA,
-    0x00B5BDEA00B4BDEA, 0x00B7BDEA00B6BDEA, 0x00B9BDEA00B8BDEA,
-    0x00BBBDEA00BABDEA, 0x00BDBDEA00BCBDEA, 0x00BFBDEA00BEBDEA,
-    0x0081BEEA0080BEEA, 0x0083BEEA0082BEEA, 0x0085BEEA0084BEEA,
-    0x0087BEEA0086BEEA, 0x0089BEEA0088BEEA, 0x008BBEEA008ABEEA,
-    0x008DBEEA008CBEEA, 0x008FBEEA008EBEEA, 0x0091BEEA0090BEEA,
-    0x0093BEEA0092BEEA, 0x0095BEEA0094BEEA, 0x0097BEEA0096BEEA,
-    0x0099BEEA0098BEEA, 0x009BBEEA009ABEEA, 0x009DBEEA009CBEEA,
-    0x009FBEEA009EBEEA, 0x00A1BEEA00A0BEEA, 0x00A3BEEA00A2BEEA,
-    0x00A5BEEA00A4BEEA, 0x00A7BEEA00A6BEEA, 0x00A9BEEA00A8BEEA,
-    0x00ABBEEA00AABEEA, 0x00ADBEEA00ACBEEA, 0x00AFBEEA00AEBEEA,
-    0x00B1BEEA00B0BEEA, 0x00B3BEEA00B2BEEA, 0x00B5BEEA00B4BEEA,
-    0x00B7BEEA00B6BEEA, 0x00B9BEEA00B8BEEA, 0x00BBBEEA00BABEEA,
-    0x00BDBEEA00BCBEEA, 0x00BFBEEA00BEBEEA, 0x0081BFEA0080BFEA,
-    0x0083BFEA0082BFEA, 0x0085BFEA0084BFEA, 0x0087BFEA0086BFEA,
-    0x0089BFEA0088BFEA, 0x008BBFEA008ABFEA, 0x008DBFEA008CBFEA,
-    0x008FBFEA008EBFEA, 0x0091BFEA0090BFEA, 0x0093BFEA0092BFEA,
-    0x0095BFEA0094BFEA, 0x0097BFEA0096BFEA, 0x0099BFEA0098BFEA,
-    0x009BBFEA009ABFEA, 0x009DBFEA009CBFEA, 0x009FBFEA009EBFEA,
-    0x00A1BFEA00A0BFEA, 0x00A3BFEA00A2BFEA, 0x00A5BFEA00A4BFEA,
-    0x00A7BFEA00A6BFEA, 0x00A9BFEA00A8BFEA, 0x00ABBFEA00AABFEA,
-    0x00ADBFEA00ACBFEA, 0x00AFBFEA00AEBFEA, 0x00B1BFEA00B0BFEA,
-    0x00B3BFEA00B2BFEA, 0x00B5BFEA00B4BFEA, 0x00B7BFEA00B6BFEA,
-    0x00B9BFEA00B8BFEA, 0x00BBBFEA00BABFEA, 0x00BDBFEA00BCBFEA,
-    0x00BFBFEA00BEBFEA, 0x008180EB008080EB, 0x008380EB008280EB,
-    0x008580EB008480EB, 0x008780EB008680EB, 0x008980EB008880EB,
-    0x008B80EB008A80EB, 0x008D80EB008C80EB, 0x008F80EB008E80EB,
-    0x009180EB009080EB, 0x009380EB009280EB, 0x009580EB009480EB,
-    0x009780EB009680EB, 0x009980EB009880EB, 0x009B80EB009A80EB,
-    0x009D80EB009C80EB, 0x009F80EB009E80EB, 0x00A180EB00A080EB,
-    0x00A380EB00A280EB, 0x00A580EB00A480EB, 0x00A780EB00A680EB,
-    0x00A980EB00A880EB, 0x00AB80EB00AA80EB, 0x00AD80EB00AC80EB,
-    0x00AF80EB00AE80EB, 0x00B180EB00B080EB, 0x00B380EB00B280EB,
-    0x00B580EB00B480EB, 0x00B780EB00B680EB, 0x00B980EB00B880EB,
-    0x00BB80EB00BA80EB, 0x00BD80EB00BC80EB, 0x00BF80EB00BE80EB,
-    0x008181EB008081EB, 0x008381EB008281EB, 0x008581EB008481EB,
-    0x008781EB008681EB, 0x008981EB008881EB, 0x008B81EB008A81EB,
-    0x008D81EB008C81EB, 0x008F81EB008E81EB, 0x009181EB009081EB,
-    0x009381EB009281EB, 0x009581EB009481EB, 0x009781EB009681EB,
-    0x009981EB009881EB, 0x009B81EB009A81EB, 0x009D81EB009C81EB,
-    0x009F81EB009E81EB, 0x00A181EB00A081EB, 0x00A381EB00A281EB,
-    0x00A581EB00A481EB, 0x00A781EB00A681EB, 0x00A981EB00A881EB,
-    0x00AB81EB00AA81EB, 0x00AD81EB00AC81EB, 0x00AF81EB00AE81EB,
-    0x00B181EB00B081EB, 0x00B381EB00B281EB, 0x00B581EB00B481EB,
-    0x00B781EB00B681EB, 0x00B981EB00B881EB, 0x00BB81EB00BA81EB,
-    0x00BD81EB00BC81EB, 0x00BF81EB00BE81EB, 0x008182EB008082EB,
-    0x008382EB008282EB, 0x008582EB008482EB, 0x008782EB008682EB,
-    0x008982EB008882EB, 0x008B82EB008A82EB, 0x008D82EB008C82EB,
-    0x008F82EB008E82EB, 0x009182EB009082EB, 0x009382EB009282EB,
-    0x009582EB009482EB, 0x009782EB009682EB, 0x009982EB009882EB,
-    0x009B82EB009A82EB, 0x009D82EB009C82EB, 0x009F82EB009E82EB,
-    0x00A182EB00A082EB, 0x00A382EB00A282EB, 0x00A582EB00A482EB,
-    0x00A782EB00A682EB, 0x00A982EB00A882EB, 0x00AB82EB00AA82EB,
-    0x00AD82EB00AC82EB, 0x00AF82EB00AE82EB, 0x00B182EB00B082EB,
-    0x00B382EB00B282EB, 0x00B582EB00B482EB, 0x00B782EB00B682EB,
-    0x00B982EB00B882EB, 0x00BB82EB00BA82EB, 0x00BD82EB00BC82EB,
-    0x00BF82EB00BE82EB, 0x008183EB008083EB, 0x008383EB008283EB,
-    0x008583EB008483EB, 0x008783EB008683EB, 0x008983EB008883EB,
-    0x008B83EB008A83EB, 0x008D83EB008C83EB, 0x008F83EB008E83EB,
-    0x009183EB009083EB, 0x009383EB009283EB, 0x009583EB009483EB,
-    0x009783EB009683EB, 0x009983EB009883EB, 0x009B83EB009A83EB,
-    0x009D83EB009C83EB, 0x009F83EB009E83EB, 0x00A183EB00A083EB,
-    0x00A383EB00A283EB, 0x00A583EB00A483EB, 0x00A783EB00A683EB,
-    0x00A983EB00A883EB, 0x00AB83EB00AA83EB, 0x00AD83EB00AC83EB,
-    0x00AF83EB00AE83EB, 0x00B183EB00B083EB, 0x00B383EB00B283EB,
-    0x00B583EB00B483EB, 0x00B783EB00B683EB, 0x00B983EB00B883EB,
-    0x00BB83EB00BA83EB, 0x00BD83EB00BC83EB, 0x00BF83EB00BE83EB,
-    0x008184EB008084EB, 0x008384EB008284EB, 0x008584EB008484EB,
-    0x008784EB008684EB, 0x008984EB008884EB, 0x008B84EB008A84EB,
-    0x008D84EB008C84EB, 0x008F84EB008E84EB, 0x009184EB009084EB,
-    0x009384EB009284EB, 0x009584EB009484EB, 0x009784EB009684EB,
-    0x009984EB009884EB, 0x009B84EB009A84EB, 0x009D84EB009C84EB,
-    0x009F84EB009E84EB, 0x00A184EB00A084EB, 0x00A384EB00A284EB,
-    0x00A584EB00A484EB, 0x00A784EB00A684EB, 0x00A984EB00A884EB,
-    0x00AB84EB00AA84EB, 0x00AD84EB00AC84EB, 0x00AF84EB00AE84EB,
-    0x00B184EB00B084EB, 0x00B384EB00B284EB, 0x00B584EB00B484EB,
-    0x00B784EB00B684EB, 0x00B984EB00B884EB, 0x00BB84EB00BA84EB,
-    0x00BD84EB00BC84EB, 0x00BF84EB00BE84EB, 0x008185EB008085EB,
-    0x008385EB008285EB, 0x008585EB008485EB, 0x008785EB008685EB,
-    0x008985EB008885EB, 0x008B85EB008A85EB, 0x008D85EB008C85EB,
-    0x008F85EB008E85EB, 0x009185EB009085EB, 0x009385EB009285EB,
-    0x009585EB009485EB, 0x009785EB009685EB, 0x009985EB009885EB,
-    0x009B85EB009A85EB, 0x009D85EB009C85EB, 0x009F85EB009E85EB,
-    0x00A185EB00A085EB, 0x00A385EB00A285EB, 0x00A585EB00A485EB,
-    0x00A785EB00A685EB, 0x00A985EB00A885EB, 0x00AB85EB00AA85EB,
-    0x00AD85EB00AC85EB, 0x00AF85EB00AE85EB, 0x00B185EB00B085EB,
-    0x00B385EB00B285EB, 0x00B585EB00B485EB, 0x00B785EB00B685EB,
-    0x00B985EB00B885EB, 0x00BB85EB00BA85EB, 0x00BD85EB00BC85EB,
-    0x00BF85EB00BE85EB, 0x008186EB008086EB, 0x008386EB008286EB,
-    0x008586EB008486EB, 0x008786EB008686EB, 0x008986EB008886EB,
-    0x008B86EB008A86EB, 0x008D86EB008C86EB, 0x008F86EB008E86EB,
-    0x009186EB009086EB, 0x009386EB009286EB, 0x009586EB009486EB,
-    0x009786EB009686EB, 0x009986EB009886EB, 0x009B86EB009A86EB,
-    0x009D86EB009C86EB, 0x009F86EB009E86EB, 0x00A186EB00A086EB,
-    0x00A386EB00A286EB, 0x00A586EB00A486EB, 0x00A786EB00A686EB,
-    0x00A986EB00A886EB, 0x00AB86EB00AA86EB, 0x00AD86EB00AC86EB,
-    0x00AF86EB00AE86EB, 0x00B186EB00B086EB, 0x00B386EB00B286EB,
-    0x00B586EB00B486EB, 0x00B786EB00B686EB, 0x00B986EB00B886EB,
-    0x00BB86EB00BA86EB, 0x00BD86EB00BC86EB, 0x00BF86EB00BE86EB,
-    0x008187EB008087EB, 0x008387EB008287EB, 0x008587EB008487EB,
-    0x008787EB008687EB, 0x008987EB008887EB, 0x008B87EB008A87EB,
-    0x008D87EB008C87EB, 0x008F87EB008E87EB, 0x009187EB009087EB,
-    0x009387EB009287EB, 0x009587EB009487EB, 0x009787EB009687EB,
-    0x009987EB009887EB, 0x009B87EB009A87EB, 0x009D87EB009C87EB,
-    0x009F87EB009E87EB, 0x00A187EB00A087EB, 0x00A387EB00A287EB,
-    0x00A587EB00A487EB, 0x00A787EB00A687EB, 0x00A987EB00A887EB,
-    0x00AB87EB00AA87EB, 0x00AD87EB00AC87EB, 0x00AF87EB00AE87EB,
-    0x00B187EB00B087EB, 0x00B387EB00B287EB, 0x00B587EB00B487EB,
-    0x00B787EB00B687EB, 0x00B987EB00B887EB, 0x00BB87EB00BA87EB,
-    0x00BD87EB00BC87EB, 0x00BF87EB00BE87EB, 0x008188EB008088EB,
-    0x008388EB008288EB, 0x008588EB008488EB, 0x008788EB008688EB,
-    0x008988EB008888EB, 0x008B88EB008A88EB, 0x008D88EB008C88EB,
-    0x008F88EB008E88EB, 0x009188EB009088EB, 0x009388EB009288EB,
-    0x009588EB009488EB, 0x009788EB009688EB, 0x009988EB009888EB,
-    0x009B88EB009A88EB, 0x009D88EB009C88EB, 0x009F88EB009E88EB,
-    0x00A188EB00A088EB, 0x00A388EB00A288EB, 0x00A588EB00A488EB,
-    0x00A788EB00A688EB, 0x00A988EB00A888EB, 0x00AB88EB00AA88EB,
-    0x00AD88EB00AC88EB, 0x00AF88EB00AE88EB, 0x00B188EB00B088EB,
-    0x00B388EB00B288EB, 0x00B588EB00B488EB, 0x00B788EB00B688EB,
-    0x00B988EB00B888EB, 0x00BB88EB00BA88EB, 0x00BD88EB00BC88EB,
-    0x00BF88EB00BE88EB, 0x008189EB008089EB, 0x008389EB008289EB,
-    0x008589EB008489EB, 0x008789EB008689EB, 0x008989EB008889EB,
-    0x008B89EB008A89EB, 0x008D89EB008C89EB, 0x008F89EB008E89EB,
-    0x009189EB009089EB, 0x009389EB009289EB, 0x009589EB009489EB,
-    0x009789EB009689EB, 0x009989EB009889EB, 0x009B89EB009A89EB,
-    0x009D89EB009C89EB, 0x009F89EB009E89EB, 0x00A189EB00A089EB,
-    0x00A389EB00A289EB, 0x00A589EB00A489EB, 0x00A789EB00A689EB,
-    0x00A989EB00A889EB, 0x00AB89EB00AA89EB, 0x00AD89EB00AC89EB,
-    0x00AF89EB00AE89EB, 0x00B189EB00B089EB, 0x00B389EB00B289EB,
-    0x00B589EB00B489EB, 0x00B789EB00B689EB, 0x00B989EB00B889EB,
-    0x00BB89EB00BA89EB, 0x00BD89EB00BC89EB, 0x00BF89EB00BE89EB,
-    0x00818AEB00808AEB, 0x00838AEB00828AEB, 0x00858AEB00848AEB,
-    0x00878AEB00868AEB, 0x00898AEB00888AEB, 0x008B8AEB008A8AEB,
-    0x008D8AEB008C8AEB, 0x008F8AEB008E8AEB, 0x00918AEB00908AEB,
-    0x00938AEB00928AEB, 0x00958AEB00948AEB, 0x00978AEB00968AEB,
-    0x00998AEB00988AEB, 0x009B8AEB009A8AEB, 0x009D8AEB009C8AEB,
-    0x009F8AEB009E8AEB, 0x00A18AEB00A08AEB, 0x00A38AEB00A28AEB,
-    0x00A58AEB00A48AEB, 0x00A78AEB00A68AEB, 0x00A98AEB00A88AEB,
-    0x00AB8AEB00AA8AEB, 0x00AD8AEB00AC8AEB, 0x00AF8AEB00AE8AEB,
-    0x00B18AEB00B08AEB, 0x00B38AEB00B28AEB, 0x00B58AEB00B48AEB,
-    0x00B78AEB00B68AEB, 0x00B98AEB00B88AEB, 0x00BB8AEB00BA8AEB,
-    0x00BD8AEB00BC8AEB, 0x00BF8AEB00BE8AEB, 0x00818BEB00808BEB,
-    0x00838BEB00828BEB, 0x00858BEB00848BEB, 0x00878BEB00868BEB,
-    0x00898BEB00888BEB, 0x008B8BEB008A8BEB, 0x008D8BEB008C8BEB,
-    0x008F8BEB008E8BEB, 0x00918BEB00908BEB, 0x00938BEB00928BEB,
-    0x00958BEB00948BEB, 0x00978BEB00968BEB, 0x00998BEB00988BEB,
-    0x009B8BEB009A8BEB, 0x009D8BEB009C8BEB, 0x009F8BEB009E8BEB,
-    0x00A18BEB00A08BEB, 0x00A38BEB00A28BEB, 0x00A58BEB00A48BEB,
-    0x00A78BEB00A68BEB, 0x00A98BEB00A88BEB, 0x00AB8BEB00AA8BEB,
-    0x00AD8BEB00AC8BEB, 0x00AF8BEB00AE8BEB, 0x00B18BEB00B08BEB,
-    0x00B38BEB00B28BEB, 0x00B58BEB00B48BEB, 0x00B78BEB00B68BEB,
-    0x00B98BEB00B88BEB, 0x00BB8BEB00BA8BEB, 0x00BD8BEB00BC8BEB,
-    0x00BF8BEB00BE8BEB, 0x00818CEB00808CEB, 0x00838CEB00828CEB,
-    0x00858CEB00848CEB, 0x00878CEB00868CEB, 0x00898CEB00888CEB,
-    0x008B8CEB008A8CEB, 0x008D8CEB008C8CEB, 0x008F8CEB008E8CEB,
-    0x00918CEB00908CEB, 0x00938CEB00928CEB, 0x00958CEB00948CEB,
-    0x00978CEB00968CEB, 0x00998CEB00988CEB, 0x009B8CEB009A8CEB,
-    0x009D8CEB009C8CEB, 0x009F8CEB009E8CEB, 0x00A18CEB00A08CEB,
-    0x00A38CEB00A28CEB, 0x00A58CEB00A48CEB, 0x00A78CEB00A68CEB,
-    0x00A98CEB00A88CEB, 0x00AB8CEB00AA8CEB, 0x00AD8CEB00AC8CEB,
-    0x00AF8CEB00AE8CEB, 0x00B18CEB00B08CEB, 0x00B38CEB00B28CEB,
-    0x00B58CEB00B48CEB, 0x00B78CEB00B68CEB, 0x00B98CEB00B88CEB,
-    0x00BB8CEB00BA8CEB, 0x00BD8CEB00BC8CEB, 0x00BF8CEB00BE8CEB,
-    0x00818DEB00808DEB, 0x00838DEB00828DEB, 0x00858DEB00848DEB,
-    0x00878DEB00868DEB, 0x00898DEB00888DEB, 0x008B8DEB008A8DEB,
-    0x008D8DEB008C8DEB, 0x008F8DEB008E8DEB, 0x00918DEB00908DEB,
-    0x00938DEB00928DEB, 0x00958DEB00948DEB, 0x00978DEB00968DEB,
-    0x00998DEB00988DEB, 0x009B8DEB009A8DEB, 0x009D8DEB009C8DEB,
-    0x009F8DEB009E8DEB, 0x00A18DEB00A08DEB, 0x00A38DEB00A28DEB,
-    0x00A58DEB00A48DEB, 0x00A78DEB00A68DEB, 0x00A98DEB00A88DEB,
-    0x00AB8DEB00AA8DEB, 0x00AD8DEB00AC8DEB, 0x00AF8DEB00AE8DEB,
-    0x00B18DEB00B08DEB, 0x00B38DEB00B28DEB, 0x00B58DEB00B48DEB,
-    0x00B78DEB00B68DEB, 0x00B98DEB00B88DEB, 0x00BB8DEB00BA8DEB,
-    0x00BD8DEB00BC8DEB, 0x00BF8DEB00BE8DEB, 0x00818EEB00808EEB,
-    0x00838EEB00828EEB, 0x00858EEB00848EEB, 0x00878EEB00868EEB,
-    0x00898EEB00888EEB, 0x008B8EEB008A8EEB, 0x008D8EEB008C8EEB,
-    0x008F8EEB008E8EEB, 0x00918EEB00908EEB, 0x00938EEB00928EEB,
-    0x00958EEB00948EEB, 0x00978EEB00968EEB, 0x00998EEB00988EEB,
-    0x009B8EEB009A8EEB, 0x009D8EEB009C8EEB, 0x009F8EEB009E8EEB,
-    0x00A18EEB00A08EEB, 0x00A38EEB00A28EEB, 0x00A58EEB00A48EEB,
-    0x00A78EEB00A68EEB, 0x00A98EEB00A88EEB, 0x00AB8EEB00AA8EEB,
-    0x00AD8EEB00AC8EEB, 0x00AF8EEB00AE8EEB, 0x00B18EEB00B08EEB,
-    0x00B38EEB00B28EEB, 0x00B58EEB00B48EEB, 0x00B78EEB00B68EEB,
-    0x00B98EEB00B88EEB, 0x00BB8EEB00BA8EEB, 0x00BD8EEB00BC8EEB,
-    0x00BF8EEB00BE8EEB, 0x00818FEB00808FEB, 0x00838FEB00828FEB,
-    0x00858FEB00848FEB, 0x00878FEB00868FEB, 0x00898FEB00888FEB,
-    0x008B8FEB008A8FEB, 0x008D8FEB008C8FEB, 0x008F8FEB008E8FEB,
-    0x00918FEB00908FEB, 0x00938FEB00928FEB, 0x00958FEB00948FEB,
-    0x00978FEB00968FEB, 0x00998FEB00988FEB, 0x009B8FEB009A8FEB,
-    0x009D8FEB009C8FEB, 0x009F8FEB009E8FEB, 0x00A18FEB00A08FEB,
-    0x00A38FEB00A28FEB, 0x00A58FEB00A48FEB, 0x00A78FEB00A68FEB,
-    0x00A98FEB00A88FEB, 0x00AB8FEB00AA8FEB, 0x00AD8FEB00AC8FEB,
-    0x00AF8FEB00AE8FEB, 0x00B18FEB00B08FEB, 0x00B38FEB00B28FEB,
-    0x00B58FEB00B48FEB, 0x00B78FEB00B68FEB, 0x00B98FEB00B88FEB,
-    0x00BB8FEB00BA8FEB, 0x00BD8FEB00BC8FEB, 0x00BF8FEB00BE8FEB,
-    0x008190EB008090EB, 0x008390EB008290EB, 0x008590EB008490EB,
-    0x008790EB008690EB, 0x008990EB008890EB, 0x008B90EB008A90EB,
-    0x008D90EB008C90EB, 0x008F90EB008E90EB, 0x009190EB009090EB,
-    0x009390EB009290EB, 0x009590EB009490EB, 0x009790EB009690EB,
-    0x009990EB009890EB, 0x009B90EB009A90EB, 0x009D90EB009C90EB,
-    0x009F90EB009E90EB, 0x00A190EB00A090EB, 0x00A390EB00A290EB,
-    0x00A590EB00A490EB, 0x00A790EB00A690EB, 0x00A990EB00A890EB,
-    0x00AB90EB00AA90EB, 0x00AD90EB00AC90EB, 0x00AF90EB00AE90EB,
-    0x00B190EB00B090EB, 0x00B390EB00B290EB, 0x00B590EB00B490EB,
-    0x00B790EB00B690EB, 0x00B990EB00B890EB, 0x00BB90EB00BA90EB,
-    0x00BD90EB00BC90EB, 0x00BF90EB00BE90EB, 0x008191EB008091EB,
-    0x008391EB008291EB, 0x008591EB008491EB, 0x008791EB008691EB,
-    0x008991EB008891EB, 0x008B91EB008A91EB, 0x008D91EB008C91EB,
-    0x008F91EB008E91EB, 0x009191EB009091EB, 0x009391EB009291EB,
-    0x009591EB009491EB, 0x009791EB009691EB, 0x009991EB009891EB,
-    0x009B91EB009A91EB, 0x009D91EB009C91EB, 0x009F91EB009E91EB,
-    0x00A191EB00A091EB, 0x00A391EB00A291EB, 0x00A591EB00A491EB,
-    0x00A791EB00A691EB, 0x00A991EB00A891EB, 0x00AB91EB00AA91EB,
-    0x00AD91EB00AC91EB, 0x00AF91EB00AE91EB, 0x00B191EB00B091EB,
-    0x00B391EB00B291EB, 0x00B591EB00B491EB, 0x00B791EB00B691EB,
-    0x00B991EB00B891EB, 0x00BB91EB00BA91EB, 0x00BD91EB00BC91EB,
-    0x00BF91EB00BE91EB, 0x008192EB008092EB, 0x008392EB008292EB,
-    0x008592EB008492EB, 0x008792EB008692EB, 0x008992EB008892EB,
-    0x008B92EB008A92EB, 0x008D92EB008C92EB, 0x008F92EB008E92EB,
-    0x009192EB009092EB, 0x009392EB009292EB, 0x009592EB009492EB,
-    0x009792EB009692EB, 0x009992EB009892EB, 0x009B92EB009A92EB,
-    0x009D92EB009C92EB, 0x009F92EB009E92EB, 0x00A192EB00A092EB,
-    0x00A392EB00A292EB, 0x00A592EB00A492EB, 0x00A792EB00A692EB,
-    0x00A992EB00A892EB, 0x00AB92EB00AA92EB, 0x00AD92EB00AC92EB,
-    0x00AF92EB00AE92EB, 0x00B192EB00B092EB, 0x00B392EB00B292EB,
-    0x00B592EB00B492EB, 0x00B792EB00B692EB, 0x00B992EB00B892EB,
-    0x00BB92EB00BA92EB, 0x00BD92EB00BC92EB, 0x00BF92EB00BE92EB,
-    0x008193EB008093EB, 0x008393EB008293EB, 0x008593EB008493EB,
-    0x008793EB008693EB, 0x008993EB008893EB, 0x008B93EB008A93EB,
-    0x008D93EB008C93EB, 0x008F93EB008E93EB, 0x009193EB009093EB,
-    0x009393EB009293EB, 0x009593EB009493EB, 0x009793EB009693EB,
-    0x009993EB009893EB, 0x009B93EB009A93EB, 0x009D93EB009C93EB,
-    0x009F93EB009E93EB, 0x00A193EB00A093EB, 0x00A393EB00A293EB,
-    0x00A593EB00A493EB, 0x00A793EB00A693EB, 0x00A993EB00A893EB,
-    0x00AB93EB00AA93EB, 0x00AD93EB00AC93EB, 0x00AF93EB00AE93EB,
-    0x00B193EB00B093EB, 0x00B393EB00B293EB, 0x00B593EB00B493EB,
-    0x00B793EB00B693EB, 0x00B993EB00B893EB, 0x00BB93EB00BA93EB,
-    0x00BD93EB00BC93EB, 0x00BF93EB00BE93EB, 0x008194EB008094EB,
-    0x008394EB008294EB, 0x008594EB008494EB, 0x008794EB008694EB,
-    0x008994EB008894EB, 0x008B94EB008A94EB, 0x008D94EB008C94EB,
-    0x008F94EB008E94EB, 0x009194EB009094EB, 0x009394EB009294EB,
-    0x009594EB009494EB, 0x009794EB009694EB, 0x009994EB009894EB,
-    0x009B94EB009A94EB, 0x009D94EB009C94EB, 0x009F94EB009E94EB,
-    0x00A194EB00A094EB, 0x00A394EB00A294EB, 0x00A594EB00A494EB,
-    0x00A794EB00A694EB, 0x00A994EB00A894EB, 0x00AB94EB00AA94EB,
-    0x00AD94EB00AC94EB, 0x00AF94EB00AE94EB, 0x00B194EB00B094EB,
-    0x00B394EB00B294EB, 0x00B594EB00B494EB, 0x00B794EB00B694EB,
-    0x00B994EB00B894EB, 0x00BB94EB00BA94EB, 0x00BD94EB00BC94EB,
-    0x00BF94EB00BE94EB, 0x008195EB008095EB, 0x008395EB008295EB,
-    0x008595EB008495EB, 0x008795EB008695EB, 0x008995EB008895EB,
-    0x008B95EB008A95EB, 0x008D95EB008C95EB, 0x008F95EB008E95EB,
-    0x009195EB009095EB, 0x009395EB009295EB, 0x009595EB009495EB,
-    0x009795EB009695EB, 0x009995EB009895EB, 0x009B95EB009A95EB,
-    0x009D95EB009C95EB, 0x009F95EB009E95EB, 0x00A195EB00A095EB,
-    0x00A395EB00A295EB, 0x00A595EB00A495EB, 0x00A795EB00A695EB,
-    0x00A995EB00A895EB, 0x00AB95EB00AA95EB, 0x00AD95EB00AC95EB,
-    0x00AF95EB00AE95EB, 0x00B195EB00B095EB, 0x00B395EB00B295EB,
-    0x00B595EB00B495EB, 0x00B795EB00B695EB, 0x00B995EB00B895EB,
-    0x00BB95EB00BA95EB, 0x00BD95EB00BC95EB, 0x00BF95EB00BE95EB,
-    0x008196EB008096EB, 0x008396EB008296EB, 0x008596EB008496EB,
-    0x008796EB008696EB, 0x008996EB008896EB, 0x008B96EB008A96EB,
-    0x008D96EB008C96EB, 0x008F96EB008E96EB, 0x009196EB009096EB,
-    0x009396EB009296EB, 0x009596EB009496EB, 0x009796EB009696EB,
-    0x009996EB009896EB, 0x009B96EB009A96EB, 0x009D96EB009C96EB,
-    0x009F96EB009E96EB, 0x00A196EB00A096EB, 0x00A396EB00A296EB,
-    0x00A596EB00A496EB, 0x00A796EB00A696EB, 0x00A996EB00A896EB,
-    0x00AB96EB00AA96EB, 0x00AD96EB00AC96EB, 0x00AF96EB00AE96EB,
-    0x00B196EB00B096EB, 0x00B396EB00B296EB, 0x00B596EB00B496EB,
-    0x00B796EB00B696EB, 0x00B996EB00B896EB, 0x00BB96EB00BA96EB,
-    0x00BD96EB00BC96EB, 0x00BF96EB00BE96EB, 0x008197EB008097EB,
-    0x008397EB008297EB, 0x008597EB008497EB, 0x008797EB008697EB,
-    0x008997EB008897EB, 0x008B97EB008A97EB, 0x008D97EB008C97EB,
-    0x008F97EB008E97EB, 0x009197EB009097EB, 0x009397EB009297EB,
-    0x009597EB009497EB, 0x009797EB009697EB, 0x009997EB009897EB,
-    0x009B97EB009A97EB, 0x009D97EB009C97EB, 0x009F97EB009E97EB,
-    0x00A197EB00A097EB, 0x00A397EB00A297EB, 0x00A597EB00A497EB,
-    0x00A797EB00A697EB, 0x00A997EB00A897EB, 0x00AB97EB00AA97EB,
-    0x00AD97EB00AC97EB, 0x00AF97EB00AE97EB, 0x00B197EB00B097EB,
-    0x00B397EB00B297EB, 0x00B597EB00B497EB, 0x00B797EB00B697EB,
-    0x00B997EB00B897EB, 0x00BB97EB00BA97EB, 0x00BD97EB00BC97EB,
-    0x00BF97EB00BE97EB, 0x008198EB008098EB, 0x008398EB008298EB,
-    0x008598EB008498EB, 0x008798EB008698EB, 0x008998EB008898EB,
-    0x008B98EB008A98EB, 0x008D98EB008C98EB, 0x008F98EB008E98EB,
-    0x009198EB009098EB, 0x009398EB009298EB, 0x009598EB009498EB,
-    0x009798EB009698EB, 0x009998EB009898EB, 0x009B98EB009A98EB,
-    0x009D98EB009C98EB, 0x009F98EB009E98EB, 0x00A198EB00A098EB,
-    0x00A398EB00A298EB, 0x00A598EB00A498EB, 0x00A798EB00A698EB,
-    0x00A998EB00A898EB, 0x00AB98EB00AA98EB, 0x00AD98EB00AC98EB,
-    0x00AF98EB00AE98EB, 0x00B198EB00B098EB, 0x00B398EB00B298EB,
-    0x00B598EB00B498EB, 0x00B798EB00B698EB, 0x00B998EB00B898EB,
-    0x00BB98EB00BA98EB, 0x00BD98EB00BC98EB, 0x00BF98EB00BE98EB,
-    0x008199EB008099EB, 0x008399EB008299EB, 0x008599EB008499EB,
-    0x008799EB008699EB, 0x008999EB008899EB, 0x008B99EB008A99EB,
-    0x008D99EB008C99EB, 0x008F99EB008E99EB, 0x009199EB009099EB,
-    0x009399EB009299EB, 0x009599EB009499EB, 0x009799EB009699EB,
-    0x009999EB009899EB, 0x009B99EB009A99EB, 0x009D99EB009C99EB,
-    0x009F99EB009E99EB, 0x00A199EB00A099EB, 0x00A399EB00A299EB,
-    0x00A599EB00A499EB, 0x00A799EB00A699EB, 0x00A999EB00A899EB,
-    0x00AB99EB00AA99EB, 0x00AD99EB00AC99EB, 0x00AF99EB00AE99EB,
-    0x00B199EB00B099EB, 0x00B399EB00B299EB, 0x00B599EB00B499EB,
-    0x00B799EB00B699EB, 0x00B999EB00B899EB, 0x00BB99EB00BA99EB,
-    0x00BD99EB00BC99EB, 0x00BF99EB00BE99EB, 0x00819AEB00809AEB,
-    0x00839AEB00829AEB, 0x00859AEB00849AEB, 0x00879AEB00869AEB,
-    0x00899AEB00889AEB, 0x008B9AEB008A9AEB, 0x008D9AEB008C9AEB,
-    0x008F9AEB008E9AEB, 0x00919AEB00909AEB, 0x00939AEB00929AEB,
-    0x00959AEB00949AEB, 0x00979AEB00969AEB, 0x00999AEB00989AEB,
-    0x009B9AEB009A9AEB, 0x009D9AEB009C9AEB, 0x009F9AEB009E9AEB,
-    0x00A19AEB00A09AEB, 0x00A39AEB00A29AEB, 0x00A59AEB00A49AEB,
-    0x00A79AEB00A69AEB, 0x00A99AEB00A89AEB, 0x00AB9AEB00AA9AEB,
-    0x00AD9AEB00AC9AEB, 0x00AF9AEB00AE9AEB, 0x00B19AEB00B09AEB,
-    0x00B39AEB00B29AEB, 0x00B59AEB00B49AEB, 0x00B79AEB00B69AEB,
-    0x00B99AEB00B89AEB, 0x00BB9AEB00BA9AEB, 0x00BD9AEB00BC9AEB,
-    0x00BF9AEB00BE9AEB, 0x00819BEB00809BEB, 0x00839BEB00829BEB,
-    0x00859BEB00849BEB, 0x00879BEB00869BEB, 0x00899BEB00889BEB,
-    0x008B9BEB008A9BEB, 0x008D9BEB008C9BEB, 0x008F9BEB008E9BEB,
-    0x00919BEB00909BEB, 0x00939BEB00929BEB, 0x00959BEB00949BEB,
-    0x00979BEB00969BEB, 0x00999BEB00989BEB, 0x009B9BEB009A9BEB,
-    0x009D9BEB009C9BEB, 0x009F9BEB009E9BEB, 0x00A19BEB00A09BEB,
-    0x00A39BEB00A29BEB, 0x00A59BEB00A49BEB, 0x00A79BEB00A69BEB,
-    0x00A99BEB00A89BEB, 0x00AB9BEB00AA9BEB, 0x00AD9BEB00AC9BEB,
-    0x00AF9BEB00AE9BEB, 0x00B19BEB00B09BEB, 0x00B39BEB00B29BEB,
-    0x00B59BEB00B49BEB, 0x00B79BEB00B69BEB, 0x00B99BEB00B89BEB,
-    0x00BB9BEB00BA9BEB, 0x00BD9BEB00BC9BEB, 0x00BF9BEB00BE9BEB,
-    0x00819CEB00809CEB, 0x00839CEB00829CEB, 0x00859CEB00849CEB,
-    0x00879CEB00869CEB, 0x00899CEB00889CEB, 0x008B9CEB008A9CEB,
-    0x008D9CEB008C9CEB, 0x008F9CEB008E9CEB, 0x00919CEB00909CEB,
-    0x00939CEB00929CEB, 0x00959CEB00949CEB, 0x00979CEB00969CEB,
-    0x00999CEB00989CEB, 0x009B9CEB009A9CEB, 0x009D9CEB009C9CEB,
-    0x009F9CEB009E9CEB, 0x00A19CEB00A09CEB, 0x00A39CEB00A29CEB,
-    0x00A59CEB00A49CEB, 0x00A79CEB00A69CEB, 0x00A99CEB00A89CEB,
-    0x00AB9CEB00AA9CEB, 0x00AD9CEB00AC9CEB, 0x00AF9CEB00AE9CEB,
-    0x00B19CEB00B09CEB, 0x00B39CEB00B29CEB, 0x00B59CEB00B49CEB,
-    0x00B79CEB00B69CEB, 0x00B99CEB00B89CEB, 0x00BB9CEB00BA9CEB,
-    0x00BD9CEB00BC9CEB, 0x00BF9CEB00BE9CEB, 0x00819DEB00809DEB,
-    0x00839DEB00829DEB, 0x00859DEB00849DEB, 0x00879DEB00869DEB,
-    0x00899DEB00889DEB, 0x008B9DEB008A9DEB, 0x008D9DEB008C9DEB,
-    0x008F9DEB008E9DEB, 0x00919DEB00909DEB, 0x00939DEB00929DEB,
-    0x00959DEB00949DEB, 0x00979DEB00969DEB, 0x00999DEB00989DEB,
-    0x009B9DEB009A9DEB, 0x009D9DEB009C9DEB, 0x009F9DEB009E9DEB,
-    0x00A19DEB00A09DEB, 0x00A39DEB00A29DEB, 0x00A59DEB00A49DEB,
-    0x00A79DEB00A69DEB, 0x00A99DEB00A89DEB, 0x00AB9DEB00AA9DEB,
-    0x00AD9DEB00AC9DEB, 0x00AF9DEB00AE9DEB, 0x00B19DEB00B09DEB,
-    0x00B39DEB00B29DEB, 0x00B59DEB00B49DEB, 0x00B79DEB00B69DEB,
-    0x00B99DEB00B89DEB, 0x00BB9DEB00BA9DEB, 0x00BD9DEB00BC9DEB,
-    0x00BF9DEB00BE9DEB, 0x00819EEB00809EEB, 0x00839EEB00829EEB,
-    0x00859EEB00849EEB, 0x00879EEB00869EEB, 0x00899EEB00889EEB,
-    0x008B9EEB008A9EEB, 0x008D9EEB008C9EEB, 0x008F9EEB008E9EEB,
-    0x00919EEB00909EEB, 0x00939EEB00929EEB, 0x00959EEB00949EEB,
-    0x00979EEB00969EEB, 0x00999EEB00989EEB, 0x009B9EEB009A9EEB,
-    0x009D9EEB009C9EEB, 0x009F9EEB009E9EEB, 0x00A19EEB00A09EEB,
-    0x00A39EEB00A29EEB, 0x00A59EEB00A49EEB, 0x00A79EEB00A69EEB,
-    0x00A99EEB00A89EEB, 0x00AB9EEB00AA9EEB, 0x00AD9EEB00AC9EEB,
-    0x00AF9EEB00AE9EEB, 0x00B19EEB00B09EEB, 0x00B39EEB00B29EEB,
-    0x00B59EEB00B49EEB, 0x00B79EEB00B69EEB, 0x00B99EEB00B89EEB,
-    0x00BB9EEB00BA9EEB, 0x00BD9EEB00BC9EEB, 0x00BF9EEB00BE9EEB,
-    0x00819FEB00809FEB, 0x00839FEB00829FEB, 0x00859FEB00849FEB,
-    0x00879FEB00869FEB, 0x00899FEB00889FEB, 0x008B9FEB008A9FEB,
-    0x008D9FEB008C9FEB, 0x008F9FEB008E9FEB, 0x00919FEB00909FEB,
-    0x00939FEB00929FEB, 0x00959FEB00949FEB, 0x00979FEB00969FEB,
-    0x00999FEB00989FEB, 0x009B9FEB009A9FEB, 0x009D9FEB009C9FEB,
-    0x009F9FEB009E9FEB, 0x00A19FEB00A09FEB, 0x00A39FEB00A29FEB,
-    0x00A59FEB00A49FEB, 0x00A79FEB00A69FEB, 0x00A99FEB00A89FEB,
-    0x00AB9FEB00AA9FEB, 0x00AD9FEB00AC9FEB, 0x00AF9FEB00AE9FEB,
-    0x00B19FEB00B09FEB, 0x00B39FEB00B29FEB, 0x00B59FEB00B49FEB,
-    0x00B79FEB00B69FEB, 0x00B99FEB00B89FEB, 0x00BB9FEB00BA9FEB,
-    0x00BD9FEB00BC9FEB, 0x00BF9FEB00BE9FEB, 0x0081A0EB0080A0EB,
-    0x0083A0EB0082A0EB, 0x0085A0EB0084A0EB, 0x0087A0EB0086A0EB,
-    0x0089A0EB0088A0EB, 0x008BA0EB008AA0EB, 0x008DA0EB008CA0EB,
-    0x008FA0EB008EA0EB, 0x0091A0EB0090A0EB, 0x0093A0EB0092A0EB,
-    0x0095A0EB0094A0EB, 0x0097A0EB0096A0EB, 0x0099A0EB0098A0EB,
-    0x009BA0EB009AA0EB, 0x009DA0EB009CA0EB, 0x009FA0EB009EA0EB,
-    0x00A1A0EB00A0A0EB, 0x00A3A0EB00A2A0EB, 0x00A5A0EB00A4A0EB,
-    0x00A7A0EB00A6A0EB, 0x00A9A0EB00A8A0EB, 0x00ABA0EB00AAA0EB,
-    0x00ADA0EB00ACA0EB, 0x00AFA0EB00AEA0EB, 0x00B1A0EB00B0A0EB,
-    0x00B3A0EB00B2A0EB, 0x00B5A0EB00B4A0EB, 0x00B7A0EB00B6A0EB,
-    0x00B9A0EB00B8A0EB, 0x00BBA0EB00BAA0EB, 0x00BDA0EB00BCA0EB,
-    0x00BFA0EB00BEA0EB, 0x0081A1EB0080A1EB, 0x0083A1EB0082A1EB,
-    0x0085A1EB0084A1EB, 0x0087A1EB0086A1EB, 0x0089A1EB0088A1EB,
-    0x008BA1EB008AA1EB, 0x008DA1EB008CA1EB, 0x008FA1EB008EA1EB,
-    0x0091A1EB0090A1EB, 0x0093A1EB0092A1EB, 0x0095A1EB0094A1EB,
-    0x0097A1EB0096A1EB, 0x0099A1EB0098A1EB, 0x009BA1EB009AA1EB,
-    0x009DA1EB009CA1EB, 0x009FA1EB009EA1EB, 0x00A1A1EB00A0A1EB,
-    0x00A3A1EB00A2A1EB, 0x00A5A1EB00A4A1EB, 0x00A7A1EB00A6A1EB,
-    0x00A9A1EB00A8A1EB, 0x00ABA1EB00AAA1EB, 0x00ADA1EB00ACA1EB,
-    0x00AFA1EB00AEA1EB, 0x00B1A1EB00B0A1EB, 0x00B3A1EB00B2A1EB,
-    0x00B5A1EB00B4A1EB, 0x00B7A1EB00B6A1EB, 0x00B9A1EB00B8A1EB,
-    0x00BBA1EB00BAA1EB, 0x00BDA1EB00BCA1EB, 0x00BFA1EB00BEA1EB,
-    0x0081A2EB0080A2EB, 0x0083A2EB0082A2EB, 0x0085A2EB0084A2EB,
-    0x0087A2EB0086A2EB, 0x0089A2EB0088A2EB, 0x008BA2EB008AA2EB,
-    0x008DA2EB008CA2EB, 0x008FA2EB008EA2EB, 0x0091A2EB0090A2EB,
-    0x0093A2EB0092A2EB, 0x0095A2EB0094A2EB, 0x0097A2EB0096A2EB,
-    0x0099A2EB0098A2EB, 0x009BA2EB009AA2EB, 0x009DA2EB009CA2EB,
-    0x009FA2EB009EA2EB, 0x00A1A2EB00A0A2EB, 0x00A3A2EB00A2A2EB,
-    0x00A5A2EB00A4A2EB, 0x00A7A2EB00A6A2EB, 0x00A9A2EB00A8A2EB,
-    0x00ABA2EB00AAA2EB, 0x00ADA2EB00ACA2EB, 0x00AFA2EB00AEA2EB,
-    0x00B1A2EB00B0A2EB, 0x00B3A2EB00B2A2EB, 0x00B5A2EB00B4A2EB,
-    0x00B7A2EB00B6A2EB, 0x00B9A2EB00B8A2EB, 0x00BBA2EB00BAA2EB,
-    0x00BDA2EB00BCA2EB, 0x00BFA2EB00BEA2EB, 0x0081A3EB0080A3EB,
-    0x0083A3EB0082A3EB, 0x0085A3EB0084A3EB, 0x0087A3EB0086A3EB,
-    0x0089A3EB0088A3EB, 0x008BA3EB008AA3EB, 0x008DA3EB008CA3EB,
-    0x008FA3EB008EA3EB, 0x0091A3EB0090A3EB, 0x0093A3EB0092A3EB,
-    0x0095A3EB0094A3EB, 0x0097A3EB0096A3EB, 0x0099A3EB0098A3EB,
-    0x009BA3EB009AA3EB, 0x009DA3EB009CA3EB, 0x009FA3EB009EA3EB,
-    0x00A1A3EB00A0A3EB, 0x00A3A3EB00A2A3EB, 0x00A5A3EB00A4A3EB,
-    0x00A7A3EB00A6A3EB, 0x00A9A3EB00A8A3EB, 0x00ABA3EB00AAA3EB,
-    0x00ADA3EB00ACA3EB, 0x00AFA3EB00AEA3EB, 0x00B1A3EB00B0A3EB,
-    0x00B3A3EB00B2A3EB, 0x00B5A3EB00B4A3EB, 0x00B7A3EB00B6A3EB,
-    0x00B9A3EB00B8A3EB, 0x00BBA3EB00BAA3EB, 0x00BDA3EB00BCA3EB,
-    0x00BFA3EB00BEA3EB, 0x0081A4EB0080A4EB, 0x0083A4EB0082A4EB,
-    0x0085A4EB0084A4EB, 0x0087A4EB0086A4EB, 0x0089A4EB0088A4EB,
-    0x008BA4EB008AA4EB, 0x008DA4EB008CA4EB, 0x008FA4EB008EA4EB,
-    0x0091A4EB0090A4EB, 0x0093A4EB0092A4EB, 0x0095A4EB0094A4EB,
-    0x0097A4EB0096A4EB, 0x0099A4EB0098A4EB, 0x009BA4EB009AA4EB,
-    0x009DA4EB009CA4EB, 0x009FA4EB009EA4EB, 0x00A1A4EB00A0A4EB,
-    0x00A3A4EB00A2A4EB, 0x00A5A4EB00A4A4EB, 0x00A7A4EB00A6A4EB,
-    0x00A9A4EB00A8A4EB, 0x00ABA4EB00AAA4EB, 0x00ADA4EB00ACA4EB,
-    0x00AFA4EB00AEA4EB, 0x00B1A4EB00B0A4EB, 0x00B3A4EB00B2A4EB,
-    0x00B5A4EB00B4A4EB, 0x00B7A4EB00B6A4EB, 0x00B9A4EB00B8A4EB,
-    0x00BBA4EB00BAA4EB, 0x00BDA4EB00BCA4EB, 0x00BFA4EB00BEA4EB,
-    0x0081A5EB0080A5EB, 0x0083A5EB0082A5EB, 0x0085A5EB0084A5EB,
-    0x0087A5EB0086A5EB, 0x0089A5EB0088A5EB, 0x008BA5EB008AA5EB,
-    0x008DA5EB008CA5EB, 0x008FA5EB008EA5EB, 0x0091A5EB0090A5EB,
-    0x0093A5EB0092A5EB, 0x0095A5EB0094A5EB, 0x0097A5EB0096A5EB,
-    0x0099A5EB0098A5EB, 0x009BA5EB009AA5EB, 0x009DA5EB009CA5EB,
-    0x009FA5EB009EA5EB, 0x00A1A5EB00A0A5EB, 0x00A3A5EB00A2A5EB,
-    0x00A5A5EB00A4A5EB, 0x00A7A5EB00A6A5EB, 0x00A9A5EB00A8A5EB,
-    0x00ABA5EB00AAA5EB, 0x00ADA5EB00ACA5EB, 0x00AFA5EB00AEA5EB,
-    0x00B1A5EB00B0A5EB, 0x00B3A5EB00B2A5EB, 0x00B5A5EB00B4A5EB,
-    0x00B7A5EB00B6A5EB, 0x00B9A5EB00B8A5EB, 0x00BBA5EB00BAA5EB,
-    0x00BDA5EB00BCA5EB, 0x00BFA5EB00BEA5EB, 0x0081A6EB0080A6EB,
-    0x0083A6EB0082A6EB, 0x0085A6EB0084A6EB, 0x0087A6EB0086A6EB,
-    0x0089A6EB0088A6EB, 0x008BA6EB008AA6EB, 0x008DA6EB008CA6EB,
-    0x008FA6EB008EA6EB, 0x0091A6EB0090A6EB, 0x0093A6EB0092A6EB,
-    0x0095A6EB0094A6EB, 0x0097A6EB0096A6EB, 0x0099A6EB0098A6EB,
-    0x009BA6EB009AA6EB, 0x009DA6EB009CA6EB, 0x009FA6EB009EA6EB,
-    0x00A1A6EB00A0A6EB, 0x00A3A6EB00A2A6EB, 0x00A5A6EB00A4A6EB,
-    0x00A7A6EB00A6A6EB, 0x00A9A6EB00A8A6EB, 0x00ABA6EB00AAA6EB,
-    0x00ADA6EB00ACA6EB, 0x00AFA6EB00AEA6EB, 0x00B1A6EB00B0A6EB,
-    0x00B3A6EB00B2A6EB, 0x00B5A6EB00B4A6EB, 0x00B7A6EB00B6A6EB,
-    0x00B9A6EB00B8A6EB, 0x00BBA6EB00BAA6EB, 0x00BDA6EB00BCA6EB,
-    0x00BFA6EB00BEA6EB, 0x0081A7EB0080A7EB, 0x0083A7EB0082A7EB,
-    0x0085A7EB0084A7EB, 0x0087A7EB0086A7EB, 0x0089A7EB0088A7EB,
-    0x008BA7EB008AA7EB, 0x008DA7EB008CA7EB, 0x008FA7EB008EA7EB,
-    0x0091A7EB0090A7EB, 0x0093A7EB0092A7EB, 0x0095A7EB0094A7EB,
-    0x0097A7EB0096A7EB, 0x0099A7EB0098A7EB, 0x009BA7EB009AA7EB,
-    0x009DA7EB009CA7EB, 0x009FA7EB009EA7EB, 0x00A1A7EB00A0A7EB,
-    0x00A3A7EB00A2A7EB, 0x00A5A7EB00A4A7EB, 0x00A7A7EB00A6A7EB,
-    0x00A9A7EB00A8A7EB, 0x00ABA7EB00AAA7EB, 0x00ADA7EB00ACA7EB,
-    0x00AFA7EB00AEA7EB, 0x00B1A7EB00B0A7EB, 0x00B3A7EB00B2A7EB,
-    0x00B5A7EB00B4A7EB, 0x00B7A7EB00B6A7EB, 0x00B9A7EB00B8A7EB,
-    0x00BBA7EB00BAA7EB, 0x00BDA7EB00BCA7EB, 0x00BFA7EB00BEA7EB,
-    0x0081A8EB0080A8EB, 0x0083A8EB0082A8EB, 0x0085A8EB0084A8EB,
-    0x0087A8EB0086A8EB, 0x0089A8EB0088A8EB, 0x008BA8EB008AA8EB,
-    0x008DA8EB008CA8EB, 0x008FA8EB008EA8EB, 0x0091A8EB0090A8EB,
-    0x0093A8EB0092A8EB, 0x0095A8EB0094A8EB, 0x0097A8EB0096A8EB,
-    0x0099A8EB0098A8EB, 0x009BA8EB009AA8EB, 0x009DA8EB009CA8EB,
-    0x009FA8EB009EA8EB, 0x00A1A8EB00A0A8EB, 0x00A3A8EB00A2A8EB,
-    0x00A5A8EB00A4A8EB, 0x00A7A8EB00A6A8EB, 0x00A9A8EB00A8A8EB,
-    0x00ABA8EB00AAA8EB, 0x00ADA8EB00ACA8EB, 0x00AFA8EB00AEA8EB,
-    0x00B1A8EB00B0A8EB, 0x00B3A8EB00B2A8EB, 0x00B5A8EB00B4A8EB,
-    0x00B7A8EB00B6A8EB, 0x00B9A8EB00B8A8EB, 0x00BBA8EB00BAA8EB,
-    0x00BDA8EB00BCA8EB, 0x00BFA8EB00BEA8EB, 0x0081A9EB0080A9EB,
-    0x0083A9EB0082A9EB, 0x0085A9EB0084A9EB, 0x0087A9EB0086A9EB,
-    0x0089A9EB0088A9EB, 0x008BA9EB008AA9EB, 0x008DA9EB008CA9EB,
-    0x008FA9EB008EA9EB, 0x0091A9EB0090A9EB, 0x0093A9EB0092A9EB,
-    0x0095A9EB0094A9EB, 0x0097A9EB0096A9EB, 0x0099A9EB0098A9EB,
-    0x009BA9EB009AA9EB, 0x009DA9EB009CA9EB, 0x009FA9EB009EA9EB,
-    0x00A1A9EB00A0A9EB, 0x00A3A9EB00A2A9EB, 0x00A5A9EB00A4A9EB,
-    0x00A7A9EB00A6A9EB, 0x00A9A9EB00A8A9EB, 0x00ABA9EB00AAA9EB,
-    0x00ADA9EB00ACA9EB, 0x00AFA9EB00AEA9EB, 0x00B1A9EB00B0A9EB,
-    0x00B3A9EB00B2A9EB, 0x00B5A9EB00B4A9EB, 0x00B7A9EB00B6A9EB,
-    0x00B9A9EB00B8A9EB, 0x00BBA9EB00BAA9EB, 0x00BDA9EB00BCA9EB,
-    0x00BFA9EB00BEA9EB, 0x0081AAEB0080AAEB, 0x0083AAEB0082AAEB,
-    0x0085AAEB0084AAEB, 0x0087AAEB0086AAEB, 0x0089AAEB0088AAEB,
-    0x008BAAEB008AAAEB, 0x008DAAEB008CAAEB, 0x008FAAEB008EAAEB,
-    0x0091AAEB0090AAEB, 0x0093AAEB0092AAEB, 0x0095AAEB0094AAEB,
-    0x0097AAEB0096AAEB, 0x0099AAEB0098AAEB, 0x009BAAEB009AAAEB,
-    0x009DAAEB009CAAEB, 0x009FAAEB009EAAEB, 0x00A1AAEB00A0AAEB,
-    0x00A3AAEB00A2AAEB, 0x00A5AAEB00A4AAEB, 0x00A7AAEB00A6AAEB,
-    0x00A9AAEB00A8AAEB, 0x00ABAAEB00AAAAEB, 0x00ADAAEB00ACAAEB,
-    0x00AFAAEB00AEAAEB, 0x00B1AAEB00B0AAEB, 0x00B3AAEB00B2AAEB,
-    0x00B5AAEB00B4AAEB, 0x00B7AAEB00B6AAEB, 0x00B9AAEB00B8AAEB,
-    0x00BBAAEB00BAAAEB, 0x00BDAAEB00BCAAEB, 0x00BFAAEB00BEAAEB,
-    0x0081ABEB0080ABEB, 0x0083ABEB0082ABEB, 0x0085ABEB0084ABEB,
-    0x0087ABEB0086ABEB, 0x0089ABEB0088ABEB, 0x008BABEB008AABEB,
-    0x008DABEB008CABEB, 0x008FABEB008EABEB, 0x0091ABEB0090ABEB,
-    0x0093ABEB0092ABEB, 0x0095ABEB0094ABEB, 0x0097ABEB0096ABEB,
-    0x0099ABEB0098ABEB, 0x009BABEB009AABEB, 0x009DABEB009CABEB,
-    0x009FABEB009EABEB, 0x00A1ABEB00A0ABEB, 0x00A3ABEB00A2ABEB,
-    0x00A5ABEB00A4ABEB, 0x00A7ABEB00A6ABEB, 0x00A9ABEB00A8ABEB,
-    0x00ABABEB00AAABEB, 0x00ADABEB00ACABEB, 0x00AFABEB00AEABEB,
-    0x00B1ABEB00B0ABEB, 0x00B3ABEB00B2ABEB, 0x00B5ABEB00B4ABEB,
-    0x00B7ABEB00B6ABEB, 0x00B9ABEB00B8ABEB, 0x00BBABEB00BAABEB,
-    0x00BDABEB00BCABEB, 0x00BFABEB00BEABEB, 0x0081ACEB0080ACEB,
-    0x0083ACEB0082ACEB, 0x0085ACEB0084ACEB, 0x0087ACEB0086ACEB,
-    0x0089ACEB0088ACEB, 0x008BACEB008AACEB, 0x008DACEB008CACEB,
-    0x008FACEB008EACEB, 0x0091ACEB0090ACEB, 0x0093ACEB0092ACEB,
-    0x0095ACEB0094ACEB, 0x0097ACEB0096ACEB, 0x0099ACEB0098ACEB,
-    0x009BACEB009AACEB, 0x009DACEB009CACEB, 0x009FACEB009EACEB,
-    0x00A1ACEB00A0ACEB, 0x00A3ACEB00A2ACEB, 0x00A5ACEB00A4ACEB,
-    0x00A7ACEB00A6ACEB, 0x00A9ACEB00A8ACEB, 0x00ABACEB00AAACEB,
-    0x00ADACEB00ACACEB, 0x00AFACEB00AEACEB, 0x00B1ACEB00B0ACEB,
-    0x00B3ACEB00B2ACEB, 0x00B5ACEB00B4ACEB, 0x00B7ACEB00B6ACEB,
-    0x00B9ACEB00B8ACEB, 0x00BBACEB00BAACEB, 0x00BDACEB00BCACEB,
-    0x00BFACEB00BEACEB, 0x0081ADEB0080ADEB, 0x0083ADEB0082ADEB,
-    0x0085ADEB0084ADEB, 0x0087ADEB0086ADEB, 0x0089ADEB0088ADEB,
-    0x008BADEB008AADEB, 0x008DADEB008CADEB, 0x008FADEB008EADEB,
-    0x0091ADEB0090ADEB, 0x0093ADEB0092ADEB, 0x0095ADEB0094ADEB,
-    0x0097ADEB0096ADEB, 0x0099ADEB0098ADEB, 0x009BADEB009AADEB,
-    0x009DADEB009CADEB, 0x009FADEB009EADEB, 0x00A1ADEB00A0ADEB,
-    0x00A3ADEB00A2ADEB, 0x00A5ADEB00A4ADEB, 0x00A7ADEB00A6ADEB,
-    0x00A9ADEB00A8ADEB, 0x00ABADEB00AAADEB, 0x00ADADEB00ACADEB,
-    0x00AFADEB00AEADEB, 0x00B1ADEB00B0ADEB, 0x00B3ADEB00B2ADEB,
-    0x00B5ADEB00B4ADEB, 0x00B7ADEB00B6ADEB, 0x00B9ADEB00B8ADEB,
-    0x00BBADEB00BAADEB, 0x00BDADEB00BCADEB, 0x00BFADEB00BEADEB,
-    0x0081AEEB0080AEEB, 0x0083AEEB0082AEEB, 0x0085AEEB0084AEEB,
-    0x0087AEEB0086AEEB, 0x0089AEEB0088AEEB, 0x008BAEEB008AAEEB,
-    0x008DAEEB008CAEEB, 0x008FAEEB008EAEEB, 0x0091AEEB0090AEEB,
-    0x0093AEEB0092AEEB, 0x0095AEEB0094AEEB, 0x0097AEEB0096AEEB,
-    0x0099AEEB0098AEEB, 0x009BAEEB009AAEEB, 0x009DAEEB009CAEEB,
-    0x009FAEEB009EAEEB, 0x00A1AEEB00A0AEEB, 0x00A3AEEB00A2AEEB,
-    0x00A5AEEB00A4AEEB, 0x00A7AEEB00A6AEEB, 0x00A9AEEB00A8AEEB,
-    0x00ABAEEB00AAAEEB, 0x00ADAEEB00ACAEEB, 0x00AFAEEB00AEAEEB,
-    0x00B1AEEB00B0AEEB, 0x00B3AEEB00B2AEEB, 0x00B5AEEB00B4AEEB,
-    0x00B7AEEB00B6AEEB, 0x00B9AEEB00B8AEEB, 0x00BBAEEB00BAAEEB,
-    0x00BDAEEB00BCAEEB, 0x00BFAEEB00BEAEEB, 0x0081AFEB0080AFEB,
-    0x0083AFEB0082AFEB, 0x0085AFEB0084AFEB, 0x0087AFEB0086AFEB,
-    0x0089AFEB0088AFEB, 0x008BAFEB008AAFEB, 0x008DAFEB008CAFEB,
-    0x008FAFEB008EAFEB, 0x0091AFEB0090AFEB, 0x0093AFEB0092AFEB,
-    0x0095AFEB0094AFEB, 0x0097AFEB0096AFEB, 0x0099AFEB0098AFEB,
-    0x009BAFEB009AAFEB, 0x009DAFEB009CAFEB, 0x009FAFEB009EAFEB,
-    0x00A1AFEB00A0AFEB, 0x00A3AFEB00A2AFEB, 0x00A5AFEB00A4AFEB,
-    0x00A7AFEB00A6AFEB, 0x00A9AFEB00A8AFEB, 0x00ABAFEB00AAAFEB,
-    0x00ADAFEB00ACAFEB, 0x00AFAFEB00AEAFEB, 0x00B1AFEB00B0AFEB,
-    0x00B3AFEB00B2AFEB, 0x00B5AFEB00B4AFEB, 0x00B7AFEB00B6AFEB,
-    0x00B9AFEB00B8AFEB, 0x00BBAFEB00BAAFEB, 0x00BDAFEB00BCAFEB,
-    0x00BFAFEB00BEAFEB, 0x0081B0EB0080B0EB, 0x0083B0EB0082B0EB,
-    0x0085B0EB0084B0EB, 0x0087B0EB0086B0EB, 0x0089B0EB0088B0EB,
-    0x008BB0EB008AB0EB, 0x008DB0EB008CB0EB, 0x008FB0EB008EB0EB,
-    0x0091B0EB0090B0EB, 0x0093B0EB0092B0EB, 0x0095B0EB0094B0EB,
-    0x0097B0EB0096B0EB, 0x0099B0EB0098B0EB, 0x009BB0EB009AB0EB,
-    0x009DB0EB009CB0EB, 0x009FB0EB009EB0EB, 0x00A1B0EB00A0B0EB,
-    0x00A3B0EB00A2B0EB, 0x00A5B0EB00A4B0EB, 0x00A7B0EB00A6B0EB,
-    0x00A9B0EB00A8B0EB, 0x00ABB0EB00AAB0EB, 0x00ADB0EB00ACB0EB,
-    0x00AFB0EB00AEB0EB, 0x00B1B0EB00B0B0EB, 0x00B3B0EB00B2B0EB,
-    0x00B5B0EB00B4B0EB, 0x00B7B0EB00B6B0EB, 0x00B9B0EB00B8B0EB,
-    0x00BBB0EB00BAB0EB, 0x00BDB0EB00BCB0EB, 0x00BFB0EB00BEB0EB,
-    0x0081B1EB0080B1EB, 0x0083B1EB0082B1EB, 0x0085B1EB0084B1EB,
-    0x0087B1EB0086B1EB, 0x0089B1EB0088B1EB, 0x008BB1EB008AB1EB,
-    0x008DB1EB008CB1EB, 0x008FB1EB008EB1EB, 0x0091B1EB0090B1EB,
-    0x0093B1EB0092B1EB, 0x0095B1EB0094B1EB, 0x0097B1EB0096B1EB,
-    0x0099B1EB0098B1EB, 0x009BB1EB009AB1EB, 0x009DB1EB009CB1EB,
-    0x009FB1EB009EB1EB, 0x00A1B1EB00A0B1EB, 0x00A3B1EB00A2B1EB,
-    0x00A5B1EB00A4B1EB, 0x00A7B1EB00A6B1EB, 0x00A9B1EB00A8B1EB,
-    0x00ABB1EB00AAB1EB, 0x00ADB1EB00ACB1EB, 0x00AFB1EB00AEB1EB,
-    0x00B1B1EB00B0B1EB, 0x00B3B1EB00B2B1EB, 0x00B5B1EB00B4B1EB,
-    0x00B7B1EB00B6B1EB, 0x00B9B1EB00B8B1EB, 0x00BBB1EB00BAB1EB,
-    0x00BDB1EB00BCB1EB, 0x00BFB1EB00BEB1EB, 0x0081B2EB0080B2EB,
-    0x0083B2EB0082B2EB, 0x0085B2EB0084B2EB, 0x0087B2EB0086B2EB,
-    0x0089B2EB0088B2EB, 0x008BB2EB008AB2EB, 0x008DB2EB008CB2EB,
-    0x008FB2EB008EB2EB, 0x0091B2EB0090B2EB, 0x0093B2EB0092B2EB,
-    0x0095B2EB0094B2EB, 0x0097B2EB0096B2EB, 0x0099B2EB0098B2EB,
-    0x009BB2EB009AB2EB, 0x009DB2EB009CB2EB, 0x009FB2EB009EB2EB,
-    0x00A1B2EB00A0B2EB, 0x00A3B2EB00A2B2EB, 0x00A5B2EB00A4B2EB,
-    0x00A7B2EB00A6B2EB, 0x00A9B2EB00A8B2EB, 0x00ABB2EB00AAB2EB,
-    0x00ADB2EB00ACB2EB, 0x00AFB2EB00AEB2EB, 0x00B1B2EB00B0B2EB,
-    0x00B3B2EB00B2B2EB, 0x00B5B2EB00B4B2EB, 0x00B7B2EB00B6B2EB,
-    0x00B9B2EB00B8B2EB, 0x00BBB2EB00BAB2EB, 0x00BDB2EB00BCB2EB,
-    0x00BFB2EB00BEB2EB, 0x0081B3EB0080B3EB, 0x0083B3EB0082B3EB,
-    0x0085B3EB0084B3EB, 0x0087B3EB0086B3EB, 0x0089B3EB0088B3EB,
-    0x008BB3EB008AB3EB, 0x008DB3EB008CB3EB, 0x008FB3EB008EB3EB,
-    0x0091B3EB0090B3EB, 0x0093B3EB0092B3EB, 0x0095B3EB0094B3EB,
-    0x0097B3EB0096B3EB, 0x0099B3EB0098B3EB, 0x009BB3EB009AB3EB,
-    0x009DB3EB009CB3EB, 0x009FB3EB009EB3EB, 0x00A1B3EB00A0B3EB,
-    0x00A3B3EB00A2B3EB, 0x00A5B3EB00A4B3EB, 0x00A7B3EB00A6B3EB,
-    0x00A9B3EB00A8B3EB, 0x00ABB3EB00AAB3EB, 0x00ADB3EB00ACB3EB,
-    0x00AFB3EB00AEB3EB, 0x00B1B3EB00B0B3EB, 0x00B3B3EB00B2B3EB,
-    0x00B5B3EB00B4B3EB, 0x00B7B3EB00B6B3EB, 0x00B9B3EB00B8B3EB,
-    0x00BBB3EB00BAB3EB, 0x00BDB3EB00BCB3EB, 0x00BFB3EB00BEB3EB,
-    0x0081B4EB0080B4EB, 0x0083B4EB0082B4EB, 0x0085B4EB0084B4EB,
-    0x0087B4EB0086B4EB, 0x0089B4EB0088B4EB, 0x008BB4EB008AB4EB,
-    0x008DB4EB008CB4EB, 0x008FB4EB008EB4EB, 0x0091B4EB0090B4EB,
-    0x0093B4EB0092B4EB, 0x0095B4EB0094B4EB, 0x0097B4EB0096B4EB,
-    0x0099B4EB0098B4EB, 0x009BB4EB009AB4EB, 0x009DB4EB009CB4EB,
-    0x009FB4EB009EB4EB, 0x00A1B4EB00A0B4EB, 0x00A3B4EB00A2B4EB,
-    0x00A5B4EB00A4B4EB, 0x00A7B4EB00A6B4EB, 0x00A9B4EB00A8B4EB,
-    0x00ABB4EB00AAB4EB, 0x00ADB4EB00ACB4EB, 0x00AFB4EB00AEB4EB,
-    0x00B1B4EB00B0B4EB, 0x00B3B4EB00B2B4EB, 0x00B5B4EB00B4B4EB,
-    0x00B7B4EB00B6B4EB, 0x00B9B4EB00B8B4EB, 0x00BBB4EB00BAB4EB,
-    0x00BDB4EB00BCB4EB, 0x00BFB4EB00BEB4EB, 0x0081B5EB0080B5EB,
-    0x0083B5EB0082B5EB, 0x0085B5EB0084B5EB, 0x0087B5EB0086B5EB,
-    0x0089B5EB0088B5EB, 0x008BB5EB008AB5EB, 0x008DB5EB008CB5EB,
-    0x008FB5EB008EB5EB, 0x0091B5EB0090B5EB, 0x0093B5EB0092B5EB,
-    0x0095B5EB0094B5EB, 0x0097B5EB0096B5EB, 0x0099B5EB0098B5EB,
-    0x009BB5EB009AB5EB, 0x009DB5EB009CB5EB, 0x009FB5EB009EB5EB,
-    0x00A1B5EB00A0B5EB, 0x00A3B5EB00A2B5EB, 0x00A5B5EB00A4B5EB,
-    0x00A7B5EB00A6B5EB, 0x00A9B5EB00A8B5EB, 0x00ABB5EB00AAB5EB,
-    0x00ADB5EB00ACB5EB, 0x00AFB5EB00AEB5EB, 0x00B1B5EB00B0B5EB,
-    0x00B3B5EB00B2B5EB, 0x00B5B5EB00B4B5EB, 0x00B7B5EB00B6B5EB,
-    0x00B9B5EB00B8B5EB, 0x00BBB5EB00BAB5EB, 0x00BDB5EB00BCB5EB,
-    0x00BFB5EB00BEB5EB, 0x0081B6EB0080B6EB, 0x0083B6EB0082B6EB,
-    0x0085B6EB0084B6EB, 0x0087B6EB0086B6EB, 0x0089B6EB0088B6EB,
-    0x008BB6EB008AB6EB, 0x008DB6EB008CB6EB, 0x008FB6EB008EB6EB,
-    0x0091B6EB0090B6EB, 0x0093B6EB0092B6EB, 0x0095B6EB0094B6EB,
-    0x0097B6EB0096B6EB, 0x0099B6EB0098B6EB, 0x009BB6EB009AB6EB,
-    0x009DB6EB009CB6EB, 0x009FB6EB009EB6EB, 0x00A1B6EB00A0B6EB,
-    0x00A3B6EB00A2B6EB, 0x00A5B6EB00A4B6EB, 0x00A7B6EB00A6B6EB,
-    0x00A9B6EB00A8B6EB, 0x00ABB6EB00AAB6EB, 0x00ADB6EB00ACB6EB,
-    0x00AFB6EB00AEB6EB, 0x00B1B6EB00B0B6EB, 0x00B3B6EB00B2B6EB,
-    0x00B5B6EB00B4B6EB, 0x00B7B6EB00B6B6EB, 0x00B9B6EB00B8B6EB,
-    0x00BBB6EB00BAB6EB, 0x00BDB6EB00BCB6EB, 0x00BFB6EB00BEB6EB,
-    0x0081B7EB0080B7EB, 0x0083B7EB0082B7EB, 0x0085B7EB0084B7EB,
-    0x0087B7EB0086B7EB, 0x0089B7EB0088B7EB, 0x008BB7EB008AB7EB,
-    0x008DB7EB008CB7EB, 0x008FB7EB008EB7EB, 0x0091B7EB0090B7EB,
-    0x0093B7EB0092B7EB, 0x0095B7EB0094B7EB, 0x0097B7EB0096B7EB,
-    0x0099B7EB0098B7EB, 0x009BB7EB009AB7EB, 0x009DB7EB009CB7EB,
-    0x009FB7EB009EB7EB, 0x00A1B7EB00A0B7EB, 0x00A3B7EB00A2B7EB,
-    0x00A5B7EB00A4B7EB, 0x00A7B7EB00A6B7EB, 0x00A9B7EB00A8B7EB,
-    0x00ABB7EB00AAB7EB, 0x00ADB7EB00ACB7EB, 0x00AFB7EB00AEB7EB,
-    0x00B1B7EB00B0B7EB, 0x00B3B7EB00B2B7EB, 0x00B5B7EB00B4B7EB,
-    0x00B7B7EB00B6B7EB, 0x00B9B7EB00B8B7EB, 0x00BBB7EB00BAB7EB,
-    0x00BDB7EB00BCB7EB, 0x00BFB7EB00BEB7EB, 0x0081B8EB0080B8EB,
-    0x0083B8EB0082B8EB, 0x0085B8EB0084B8EB, 0x0087B8EB0086B8EB,
-    0x0089B8EB0088B8EB, 0x008BB8EB008AB8EB, 0x008DB8EB008CB8EB,
-    0x008FB8EB008EB8EB, 0x0091B8EB0090B8EB, 0x0093B8EB0092B8EB,
-    0x0095B8EB0094B8EB, 0x0097B8EB0096B8EB, 0x0099B8EB0098B8EB,
-    0x009BB8EB009AB8EB, 0x009DB8EB009CB8EB, 0x009FB8EB009EB8EB,
-    0x00A1B8EB00A0B8EB, 0x00A3B8EB00A2B8EB, 0x00A5B8EB00A4B8EB,
-    0x00A7B8EB00A6B8EB, 0x00A9B8EB00A8B8EB, 0x00ABB8EB00AAB8EB,
-    0x00ADB8EB00ACB8EB, 0x00AFB8EB00AEB8EB, 0x00B1B8EB00B0B8EB,
-    0x00B3B8EB00B2B8EB, 0x00B5B8EB00B4B8EB, 0x00B7B8EB00B6B8EB,
-    0x00B9B8EB00B8B8EB, 0x00BBB8EB00BAB8EB, 0x00BDB8EB00BCB8EB,
-    0x00BFB8EB00BEB8EB, 0x0081B9EB0080B9EB, 0x0083B9EB0082B9EB,
-    0x0085B9EB0084B9EB, 0x0087B9EB0086B9EB, 0x0089B9EB0088B9EB,
-    0x008BB9EB008AB9EB, 0x008DB9EB008CB9EB, 0x008FB9EB008EB9EB,
-    0x0091B9EB0090B9EB, 0x0093B9EB0092B9EB, 0x0095B9EB0094B9EB,
-    0x0097B9EB0096B9EB, 0x0099B9EB0098B9EB, 0x009BB9EB009AB9EB,
-    0x009DB9EB009CB9EB, 0x009FB9EB009EB9EB, 0x00A1B9EB00A0B9EB,
-    0x00A3B9EB00A2B9EB, 0x00A5B9EB00A4B9EB, 0x00A7B9EB00A6B9EB,
-    0x00A9B9EB00A8B9EB, 0x00ABB9EB00AAB9EB, 0x00ADB9EB00ACB9EB,
-    0x00AFB9EB00AEB9EB, 0x00B1B9EB00B0B9EB, 0x00B3B9EB00B2B9EB,
-    0x00B5B9EB00B4B9EB, 0x00B7B9EB00B6B9EB, 0x00B9B9EB00B8B9EB,
-    0x00BBB9EB00BAB9EB, 0x00BDB9EB00BCB9EB, 0x00BFB9EB00BEB9EB,
-    0x0081BAEB0080BAEB, 0x0083BAEB0082BAEB, 0x0085BAEB0084BAEB,
-    0x0087BAEB0086BAEB, 0x0089BAEB0088BAEB, 0x008BBAEB008ABAEB,
-    0x008DBAEB008CBAEB, 0x008FBAEB008EBAEB, 0x0091BAEB0090BAEB,
-    0x0093BAEB0092BAEB, 0x0095BAEB0094BAEB, 0x0097BAEB0096BAEB,
-    0x0099BAEB0098BAEB, 0x009BBAEB009ABAEB, 0x009DBAEB009CBAEB,
-    0x009FBAEB009EBAEB, 0x00A1BAEB00A0BAEB, 0x00A3BAEB00A2BAEB,
-    0x00A5BAEB00A4BAEB, 0x00A7BAEB00A6BAEB, 0x00A9BAEB00A8BAEB,
-    0x00ABBAEB00AABAEB, 0x00ADBAEB00ACBAEB, 0x00AFBAEB00AEBAEB,
-    0x00B1BAEB00B0BAEB, 0x00B3BAEB00B2BAEB, 0x00B5BAEB00B4BAEB,
-    0x00B7BAEB00B6BAEB, 0x00B9BAEB00B8BAEB, 0x00BBBAEB00BABAEB,
-    0x00BDBAEB00BCBAEB, 0x00BFBAEB00BEBAEB, 0x0081BBEB0080BBEB,
-    0x0083BBEB0082BBEB, 0x0085BBEB0084BBEB, 0x0087BBEB0086BBEB,
-    0x0089BBEB0088BBEB, 0x008BBBEB008ABBEB, 0x008DBBEB008CBBEB,
-    0x008FBBEB008EBBEB, 0x0091BBEB0090BBEB, 0x0093BBEB0092BBEB,
-    0x0095BBEB0094BBEB, 0x0097BBEB0096BBEB, 0x0099BBEB0098BBEB,
-    0x009BBBEB009ABBEB, 0x009DBBEB009CBBEB, 0x009FBBEB009EBBEB,
-    0x00A1BBEB00A0BBEB, 0x00A3BBEB00A2BBEB, 0x00A5BBEB00A4BBEB,
-    0x00A7BBEB00A6BBEB, 0x00A9BBEB00A8BBEB, 0x00ABBBEB00AABBEB,
-    0x00ADBBEB00ACBBEB, 0x00AFBBEB00AEBBEB, 0x00B1BBEB00B0BBEB,
-    0x00B3BBEB00B2BBEB, 0x00B5BBEB00B4BBEB, 0x00B7BBEB00B6BBEB,
-    0x00B9BBEB00B8BBEB, 0x00BBBBEB00BABBEB, 0x00BDBBEB00BCBBEB,
-    0x00BFBBEB00BEBBEB, 0x0081BCEB0080BCEB, 0x0083BCEB0082BCEB,
-    0x0085BCEB0084BCEB, 0x0087BCEB0086BCEB, 0x0089BCEB0088BCEB,
-    0x008BBCEB008ABCEB, 0x008DBCEB008CBCEB, 0x008FBCEB008EBCEB,
-    0x0091BCEB0090BCEB, 0x0093BCEB0092BCEB, 0x0095BCEB0094BCEB,
-    0x0097BCEB0096BCEB, 0x0099BCEB0098BCEB, 0x009BBCEB009ABCEB,
-    0x009DBCEB009CBCEB, 0x009FBCEB009EBCEB, 0x00A1BCEB00A0BCEB,
-    0x00A3BCEB00A2BCEB, 0x00A5BCEB00A4BCEB, 0x00A7BCEB00A6BCEB,
-    0x00A9BCEB00A8BCEB, 0x00ABBCEB00AABCEB, 0x00ADBCEB00ACBCEB,
-    0x00AFBCEB00AEBCEB, 0x00B1BCEB00B0BCEB, 0x00B3BCEB00B2BCEB,
-    0x00B5BCEB00B4BCEB, 0x00B7BCEB00B6BCEB, 0x00B9BCEB00B8BCEB,
-    0x00BBBCEB00BABCEB, 0x00BDBCEB00BCBCEB, 0x00BFBCEB00BEBCEB,
-    0x0081BDEB0080BDEB, 0x0083BDEB0082BDEB, 0x0085BDEB0084BDEB,
-    0x0087BDEB0086BDEB, 0x0089BDEB0088BDEB, 0x008BBDEB008ABDEB,
-    0x008DBDEB008CBDEB, 0x008FBDEB008EBDEB, 0x0091BDEB0090BDEB,
-    0x0093BDEB0092BDEB, 0x0095BDEB0094BDEB, 0x0097BDEB0096BDEB,
-    0x0099BDEB0098BDEB, 0x009BBDEB009ABDEB, 0x009DBDEB009CBDEB,
-    0x009FBDEB009EBDEB, 0x00A1BDEB00A0BDEB, 0x00A3BDEB00A2BDEB,
-    0x00A5BDEB00A4BDEB, 0x00A7BDEB00A6BDEB, 0x00A9BDEB00A8BDEB,
-    0x00ABBDEB00AABDEB, 0x00ADBDEB00ACBDEB, 0x00AFBDEB00AEBDEB,
-    0x00B1BDEB00B0BDEB, 0x00B3BDEB00B2BDEB, 0x00B5BDEB00B4BDEB,
-    0x00B7BDEB00B6BDEB, 0x00B9BDEB00B8BDEB, 0x00BBBDEB00BABDEB,
-    0x00BDBDEB00BCBDEB, 0x00BFBDEB00BEBDEB, 0x0081BEEB0080BEEB,
-    0x0083BEEB0082BEEB, 0x0085BEEB0084BEEB, 0x0087BEEB0086BEEB,
-    0x0089BEEB0088BEEB, 0x008BBEEB008ABEEB, 0x008DBEEB008CBEEB,
-    0x008FBEEB008EBEEB, 0x0091BEEB0090BEEB, 0x0093BEEB0092BEEB,
-    0x0095BEEB0094BEEB, 0x0097BEEB0096BEEB, 0x0099BEEB0098BEEB,
-    0x009BBEEB009ABEEB, 0x009DBEEB009CBEEB, 0x009FBEEB009EBEEB,
-    0x00A1BEEB00A0BEEB, 0x00A3BEEB00A2BEEB, 0x00A5BEEB00A4BEEB,
-    0x00A7BEEB00A6BEEB, 0x00A9BEEB00A8BEEB, 0x00ABBEEB00AABEEB,
-    0x00ADBEEB00ACBEEB, 0x00AFBEEB00AEBEEB, 0x00B1BEEB00B0BEEB,
-    0x00B3BEEB00B2BEEB, 0x00B5BEEB00B4BEEB, 0x00B7BEEB00B6BEEB,
-    0x00B9BEEB00B8BEEB, 0x00BBBEEB00BABEEB, 0x00BDBEEB00BCBEEB,
-    0x00BFBEEB00BEBEEB, 0x0081BFEB0080BFEB, 0x0083BFEB0082BFEB,
-    0x0085BFEB0084BFEB, 0x0087BFEB0086BFEB, 0x0089BFEB0088BFEB,
-    0x008BBFEB008ABFEB, 0x008DBFEB008CBFEB, 0x008FBFEB008EBFEB,
-    0x0091BFEB0090BFEB, 0x0093BFEB0092BFEB, 0x0095BFEB0094BFEB,
-    0x0097BFEB0096BFEB, 0x0099BFEB0098BFEB, 0x009BBFEB009ABFEB,
-    0x009DBFEB009CBFEB, 0x009FBFEB009EBFEB, 0x00A1BFEB00A0BFEB,
-    0x00A3BFEB00A2BFEB, 0x00A5BFEB00A4BFEB, 0x00A7BFEB00A6BFEB,
-    0x00A9BFEB00A8BFEB, 0x00ABBFEB00AABFEB, 0x00ADBFEB00ACBFEB,
-    0x00AFBFEB00AEBFEB, 0x00B1BFEB00B0BFEB, 0x00B3BFEB00B2BFEB,
-    0x00B5BFEB00B4BFEB, 0x00B7BFEB00B6BFEB, 0x00B9BFEB00B8BFEB,
-    0x00BBBFEB00BABFEB, 0x00BDBFEB00BCBFEB, 0x00BFBFEB00BEBFEB,
-    0x008180EC008080EC, 0x008380EC008280EC, 0x008580EC008480EC,
-    0x008780EC008680EC, 0x008980EC008880EC, 0x008B80EC008A80EC,
-    0x008D80EC008C80EC, 0x008F80EC008E80EC, 0x009180EC009080EC,
-    0x009380EC009280EC, 0x009580EC009480EC, 0x009780EC009680EC,
-    0x009980EC009880EC, 0x009B80EC009A80EC, 0x009D80EC009C80EC,
-    0x009F80EC009E80EC, 0x00A180EC00A080EC, 0x00A380EC00A280EC,
-    0x00A580EC00A480EC, 0x00A780EC00A680EC, 0x00A980EC00A880EC,
-    0x00AB80EC00AA80EC, 0x00AD80EC00AC80EC, 0x00AF80EC00AE80EC,
-    0x00B180EC00B080EC, 0x00B380EC00B280EC, 0x00B580EC00B480EC,
-    0x00B780EC00B680EC, 0x00B980EC00B880EC, 0x00BB80EC00BA80EC,
-    0x00BD80EC00BC80EC, 0x00BF80EC00BE80EC, 0x008181EC008081EC,
-    0x008381EC008281EC, 0x008581EC008481EC, 0x008781EC008681EC,
-    0x008981EC008881EC, 0x008B81EC008A81EC, 0x008D81EC008C81EC,
-    0x008F81EC008E81EC, 0x009181EC009081EC, 0x009381EC009281EC,
-    0x009581EC009481EC, 0x009781EC009681EC, 0x009981EC009881EC,
-    0x009B81EC009A81EC, 0x009D81EC009C81EC, 0x009F81EC009E81EC,
-    0x00A181EC00A081EC, 0x00A381EC00A281EC, 0x00A581EC00A481EC,
-    0x00A781EC00A681EC, 0x00A981EC00A881EC, 0x00AB81EC00AA81EC,
-    0x00AD81EC00AC81EC, 0x00AF81EC00AE81EC, 0x00B181EC00B081EC,
-    0x00B381EC00B281EC, 0x00B581EC00B481EC, 0x00B781EC00B681EC,
-    0x00B981EC00B881EC, 0x00BB81EC00BA81EC, 0x00BD81EC00BC81EC,
-    0x00BF81EC00BE81EC, 0x008182EC008082EC, 0x008382EC008282EC,
-    0x008582EC008482EC, 0x008782EC008682EC, 0x008982EC008882EC,
-    0x008B82EC008A82EC, 0x008D82EC008C82EC, 0x008F82EC008E82EC,
-    0x009182EC009082EC, 0x009382EC009282EC, 0x009582EC009482EC,
-    0x009782EC009682EC, 0x009982EC009882EC, 0x009B82EC009A82EC,
-    0x009D82EC009C82EC, 0x009F82EC009E82EC, 0x00A182EC00A082EC,
-    0x00A382EC00A282EC, 0x00A582EC00A482EC, 0x00A782EC00A682EC,
-    0x00A982EC00A882EC, 0x00AB82EC00AA82EC, 0x00AD82EC00AC82EC,
-    0x00AF82EC00AE82EC, 0x00B182EC00B082EC, 0x00B382EC00B282EC,
-    0x00B582EC00B482EC, 0x00B782EC00B682EC, 0x00B982EC00B882EC,
-    0x00BB82EC00BA82EC, 0x00BD82EC00BC82EC, 0x00BF82EC00BE82EC,
-    0x008183EC008083EC, 0x008383EC008283EC, 0x008583EC008483EC,
-    0x008783EC008683EC, 0x008983EC008883EC, 0x008B83EC008A83EC,
-    0x008D83EC008C83EC, 0x008F83EC008E83EC, 0x009183EC009083EC,
-    0x009383EC009283EC, 0x009583EC009483EC, 0x009783EC009683EC,
-    0x009983EC009883EC, 0x009B83EC009A83EC, 0x009D83EC009C83EC,
-    0x009F83EC009E83EC, 0x00A183EC00A083EC, 0x00A383EC00A283EC,
-    0x00A583EC00A483EC, 0x00A783EC00A683EC, 0x00A983EC00A883EC,
-    0x00AB83EC00AA83EC, 0x00AD83EC00AC83EC, 0x00AF83EC00AE83EC,
-    0x00B183EC00B083EC, 0x00B383EC00B283EC, 0x00B583EC00B483EC,
-    0x00B783EC00B683EC, 0x00B983EC00B883EC, 0x00BB83EC00BA83EC,
-    0x00BD83EC00BC83EC, 0x00BF83EC00BE83EC, 0x008184EC008084EC,
-    0x008384EC008284EC, 0x008584EC008484EC, 0x008784EC008684EC,
-    0x008984EC008884EC, 0x008B84EC008A84EC, 0x008D84EC008C84EC,
-    0x008F84EC008E84EC, 0x009184EC009084EC, 0x009384EC009284EC,
-    0x009584EC009484EC, 0x009784EC009684EC, 0x009984EC009884EC,
-    0x009B84EC009A84EC, 0x009D84EC009C84EC, 0x009F84EC009E84EC,
-    0x00A184EC00A084EC, 0x00A384EC00A284EC, 0x00A584EC00A484EC,
-    0x00A784EC00A684EC, 0x00A984EC00A884EC, 0x00AB84EC00AA84EC,
-    0x00AD84EC00AC84EC, 0x00AF84EC00AE84EC, 0x00B184EC00B084EC,
-    0x00B384EC00B284EC, 0x00B584EC00B484EC, 0x00B784EC00B684EC,
-    0x00B984EC00B884EC, 0x00BB84EC00BA84EC, 0x00BD84EC00BC84EC,
-    0x00BF84EC00BE84EC, 0x008185EC008085EC, 0x008385EC008285EC,
-    0x008585EC008485EC, 0x008785EC008685EC, 0x008985EC008885EC,
-    0x008B85EC008A85EC, 0x008D85EC008C85EC, 0x008F85EC008E85EC,
-    0x009185EC009085EC, 0x009385EC009285EC, 0x009585EC009485EC,
-    0x009785EC009685EC, 0x009985EC009885EC, 0x009B85EC009A85EC,
-    0x009D85EC009C85EC, 0x009F85EC009E85EC, 0x00A185EC00A085EC,
-    0x00A385EC00A285EC, 0x00A585EC00A485EC, 0x00A785EC00A685EC,
-    0x00A985EC00A885EC, 0x00AB85EC00AA85EC, 0x00AD85EC00AC85EC,
-    0x00AF85EC00AE85EC, 0x00B185EC00B085EC, 0x00B385EC00B285EC,
-    0x00B585EC00B485EC, 0x00B785EC00B685EC, 0x00B985EC00B885EC,
-    0x00BB85EC00BA85EC, 0x00BD85EC00BC85EC, 0x00BF85EC00BE85EC,
-    0x008186EC008086EC, 0x008386EC008286EC, 0x008586EC008486EC,
-    0x008786EC008686EC, 0x008986EC008886EC, 0x008B86EC008A86EC,
-    0x008D86EC008C86EC, 0x008F86EC008E86EC, 0x009186EC009086EC,
-    0x009386EC009286EC, 0x009586EC009486EC, 0x009786EC009686EC,
-    0x009986EC009886EC, 0x009B86EC009A86EC, 0x009D86EC009C86EC,
-    0x009F86EC009E86EC, 0x00A186EC00A086EC, 0x00A386EC00A286EC,
-    0x00A586EC00A486EC, 0x00A786EC00A686EC, 0x00A986EC00A886EC,
-    0x00AB86EC00AA86EC, 0x00AD86EC00AC86EC, 0x00AF86EC00AE86EC,
-    0x00B186EC00B086EC, 0x00B386EC00B286EC, 0x00B586EC00B486EC,
-    0x00B786EC00B686EC, 0x00B986EC00B886EC, 0x00BB86EC00BA86EC,
-    0x00BD86EC00BC86EC, 0x00BF86EC00BE86EC, 0x008187EC008087EC,
-    0x008387EC008287EC, 0x008587EC008487EC, 0x008787EC008687EC,
-    0x008987EC008887EC, 0x008B87EC008A87EC, 0x008D87EC008C87EC,
-    0x008F87EC008E87EC, 0x009187EC009087EC, 0x009387EC009287EC,
-    0x009587EC009487EC, 0x009787EC009687EC, 0x009987EC009887EC,
-    0x009B87EC009A87EC, 0x009D87EC009C87EC, 0x009F87EC009E87EC,
-    0x00A187EC00A087EC, 0x00A387EC00A287EC, 0x00A587EC00A487EC,
-    0x00A787EC00A687EC, 0x00A987EC00A887EC, 0x00AB87EC00AA87EC,
-    0x00AD87EC00AC87EC, 0x00AF87EC00AE87EC, 0x00B187EC00B087EC,
-    0x00B387EC00B287EC, 0x00B587EC00B487EC, 0x00B787EC00B687EC,
-    0x00B987EC00B887EC, 0x00BB87EC00BA87EC, 0x00BD87EC00BC87EC,
-    0x00BF87EC00BE87EC, 0x008188EC008088EC, 0x008388EC008288EC,
-    0x008588EC008488EC, 0x008788EC008688EC, 0x008988EC008888EC,
-    0x008B88EC008A88EC, 0x008D88EC008C88EC, 0x008F88EC008E88EC,
-    0x009188EC009088EC, 0x009388EC009288EC, 0x009588EC009488EC,
-    0x009788EC009688EC, 0x009988EC009888EC, 0x009B88EC009A88EC,
-    0x009D88EC009C88EC, 0x009F88EC009E88EC, 0x00A188EC00A088EC,
-    0x00A388EC00A288EC, 0x00A588EC00A488EC, 0x00A788EC00A688EC,
-    0x00A988EC00A888EC, 0x00AB88EC00AA88EC, 0x00AD88EC00AC88EC,
-    0x00AF88EC00AE88EC, 0x00B188EC00B088EC, 0x00B388EC00B288EC,
-    0x00B588EC00B488EC, 0x00B788EC00B688EC, 0x00B988EC00B888EC,
-    0x00BB88EC00BA88EC, 0x00BD88EC00BC88EC, 0x00BF88EC00BE88EC,
-    0x008189EC008089EC, 0x008389EC008289EC, 0x008589EC008489EC,
-    0x008789EC008689EC, 0x008989EC008889EC, 0x008B89EC008A89EC,
-    0x008D89EC008C89EC, 0x008F89EC008E89EC, 0x009189EC009089EC,
-    0x009389EC009289EC, 0x009589EC009489EC, 0x009789EC009689EC,
-    0x009989EC009889EC, 0x009B89EC009A89EC, 0x009D89EC009C89EC,
-    0x009F89EC009E89EC, 0x00A189EC00A089EC, 0x00A389EC00A289EC,
-    0x00A589EC00A489EC, 0x00A789EC00A689EC, 0x00A989EC00A889EC,
-    0x00AB89EC00AA89EC, 0x00AD89EC00AC89EC, 0x00AF89EC00AE89EC,
-    0x00B189EC00B089EC, 0x00B389EC00B289EC, 0x00B589EC00B489EC,
-    0x00B789EC00B689EC, 0x00B989EC00B889EC, 0x00BB89EC00BA89EC,
-    0x00BD89EC00BC89EC, 0x00BF89EC00BE89EC, 0x00818AEC00808AEC,
-    0x00838AEC00828AEC, 0x00858AEC00848AEC, 0x00878AEC00868AEC,
-    0x00898AEC00888AEC, 0x008B8AEC008A8AEC, 0x008D8AEC008C8AEC,
-    0x008F8AEC008E8AEC, 0x00918AEC00908AEC, 0x00938AEC00928AEC,
-    0x00958AEC00948AEC, 0x00978AEC00968AEC, 0x00998AEC00988AEC,
-    0x009B8AEC009A8AEC, 0x009D8AEC009C8AEC, 0x009F8AEC009E8AEC,
-    0x00A18AEC00A08AEC, 0x00A38AEC00A28AEC, 0x00A58AEC00A48AEC,
-    0x00A78AEC00A68AEC, 0x00A98AEC00A88AEC, 0x00AB8AEC00AA8AEC,
-    0x00AD8AEC00AC8AEC, 0x00AF8AEC00AE8AEC, 0x00B18AEC00B08AEC,
-    0x00B38AEC00B28AEC, 0x00B58AEC00B48AEC, 0x00B78AEC00B68AEC,
-    0x00B98AEC00B88AEC, 0x00BB8AEC00BA8AEC, 0x00BD8AEC00BC8AEC,
-    0x00BF8AEC00BE8AEC, 0x00818BEC00808BEC, 0x00838BEC00828BEC,
-    0x00858BEC00848BEC, 0x00878BEC00868BEC, 0x00898BEC00888BEC,
-    0x008B8BEC008A8BEC, 0x008D8BEC008C8BEC, 0x008F8BEC008E8BEC,
-    0x00918BEC00908BEC, 0x00938BEC00928BEC, 0x00958BEC00948BEC,
-    0x00978BEC00968BEC, 0x00998BEC00988BEC, 0x009B8BEC009A8BEC,
-    0x009D8BEC009C8BEC, 0x009F8BEC009E8BEC, 0x00A18BEC00A08BEC,
-    0x00A38BEC00A28BEC, 0x00A58BEC00A48BEC, 0x00A78BEC00A68BEC,
-    0x00A98BEC00A88BEC, 0x00AB8BEC00AA8BEC, 0x00AD8BEC00AC8BEC,
-    0x00AF8BEC00AE8BEC, 0x00B18BEC00B08BEC, 0x00B38BEC00B28BEC,
-    0x00B58BEC00B48BEC, 0x00B78BEC00B68BEC, 0x00B98BEC00B88BEC,
-    0x00BB8BEC00BA8BEC, 0x00BD8BEC00BC8BEC, 0x00BF8BEC00BE8BEC,
-    0x00818CEC00808CEC, 0x00838CEC00828CEC, 0x00858CEC00848CEC,
-    0x00878CEC00868CEC, 0x00898CEC00888CEC, 0x008B8CEC008A8CEC,
-    0x008D8CEC008C8CEC, 0x008F8CEC008E8CEC, 0x00918CEC00908CEC,
-    0x00938CEC00928CEC, 0x00958CEC00948CEC, 0x00978CEC00968CEC,
-    0x00998CEC00988CEC, 0x009B8CEC009A8CEC, 0x009D8CEC009C8CEC,
-    0x009F8CEC009E8CEC, 0x00A18CEC00A08CEC, 0x00A38CEC00A28CEC,
-    0x00A58CEC00A48CEC, 0x00A78CEC00A68CEC, 0x00A98CEC00A88CEC,
-    0x00AB8CEC00AA8CEC, 0x00AD8CEC00AC8CEC, 0x00AF8CEC00AE8CEC,
-    0x00B18CEC00B08CEC, 0x00B38CEC00B28CEC, 0x00B58CEC00B48CEC,
-    0x00B78CEC00B68CEC, 0x00B98CEC00B88CEC, 0x00BB8CEC00BA8CEC,
-    0x00BD8CEC00BC8CEC, 0x00BF8CEC00BE8CEC, 0x00818DEC00808DEC,
-    0x00838DEC00828DEC, 0x00858DEC00848DEC, 0x00878DEC00868DEC,
-    0x00898DEC00888DEC, 0x008B8DEC008A8DEC, 0x008D8DEC008C8DEC,
-    0x008F8DEC008E8DEC, 0x00918DEC00908DEC, 0x00938DEC00928DEC,
-    0x00958DEC00948DEC, 0x00978DEC00968DEC, 0x00998DEC00988DEC,
-    0x009B8DEC009A8DEC, 0x009D8DEC009C8DEC, 0x009F8DEC009E8DEC,
-    0x00A18DEC00A08DEC, 0x00A38DEC00A28DEC, 0x00A58DEC00A48DEC,
-    0x00A78DEC00A68DEC, 0x00A98DEC00A88DEC, 0x00AB8DEC00AA8DEC,
-    0x00AD8DEC00AC8DEC, 0x00AF8DEC00AE8DEC, 0x00B18DEC00B08DEC,
-    0x00B38DEC00B28DEC, 0x00B58DEC00B48DEC, 0x00B78DEC00B68DEC,
-    0x00B98DEC00B88DEC, 0x00BB8DEC00BA8DEC, 0x00BD8DEC00BC8DEC,
-    0x00BF8DEC00BE8DEC, 0x00818EEC00808EEC, 0x00838EEC00828EEC,
-    0x00858EEC00848EEC, 0x00878EEC00868EEC, 0x00898EEC00888EEC,
-    0x008B8EEC008A8EEC, 0x008D8EEC008C8EEC, 0x008F8EEC008E8EEC,
-    0x00918EEC00908EEC, 0x00938EEC00928EEC, 0x00958EEC00948EEC,
-    0x00978EEC00968EEC, 0x00998EEC00988EEC, 0x009B8EEC009A8EEC,
-    0x009D8EEC009C8EEC, 0x009F8EEC009E8EEC, 0x00A18EEC00A08EEC,
-    0x00A38EEC00A28EEC, 0x00A58EEC00A48EEC, 0x00A78EEC00A68EEC,
-    0x00A98EEC00A88EEC, 0x00AB8EEC00AA8EEC, 0x00AD8EEC00AC8EEC,
-    0x00AF8EEC00AE8EEC, 0x00B18EEC00B08EEC, 0x00B38EEC00B28EEC,
-    0x00B58EEC00B48EEC, 0x00B78EEC00B68EEC, 0x00B98EEC00B88EEC,
-    0x00BB8EEC00BA8EEC, 0x00BD8EEC00BC8EEC, 0x00BF8EEC00BE8EEC,
-    0x00818FEC00808FEC, 0x00838FEC00828FEC, 0x00858FEC00848FEC,
-    0x00878FEC00868FEC, 0x00898FEC00888FEC, 0x008B8FEC008A8FEC,
-    0x008D8FEC008C8FEC, 0x008F8FEC008E8FEC, 0x00918FEC00908FEC,
-    0x00938FEC00928FEC, 0x00958FEC00948FEC, 0x00978FEC00968FEC,
-    0x00998FEC00988FEC, 0x009B8FEC009A8FEC, 0x009D8FEC009C8FEC,
-    0x009F8FEC009E8FEC, 0x00A18FEC00A08FEC, 0x00A38FEC00A28FEC,
-    0x00A58FEC00A48FEC, 0x00A78FEC00A68FEC, 0x00A98FEC00A88FEC,
-    0x00AB8FEC00AA8FEC, 0x00AD8FEC00AC8FEC, 0x00AF8FEC00AE8FEC,
-    0x00B18FEC00B08FEC, 0x00B38FEC00B28FEC, 0x00B58FEC00B48FEC,
-    0x00B78FEC00B68FEC, 0x00B98FEC00B88FEC, 0x00BB8FEC00BA8FEC,
-    0x00BD8FEC00BC8FEC, 0x00BF8FEC00BE8FEC, 0x008190EC008090EC,
-    0x008390EC008290EC, 0x008590EC008490EC, 0x008790EC008690EC,
-    0x008990EC008890EC, 0x008B90EC008A90EC, 0x008D90EC008C90EC,
-    0x008F90EC008E90EC, 0x009190EC009090EC, 0x009390EC009290EC,
-    0x009590EC009490EC, 0x009790EC009690EC, 0x009990EC009890EC,
-    0x009B90EC009A90EC, 0x009D90EC009C90EC, 0x009F90EC009E90EC,
-    0x00A190EC00A090EC, 0x00A390EC00A290EC, 0x00A590EC00A490EC,
-    0x00A790EC00A690EC, 0x00A990EC00A890EC, 0x00AB90EC00AA90EC,
-    0x00AD90EC00AC90EC, 0x00AF90EC00AE90EC, 0x00B190EC00B090EC,
-    0x00B390EC00B290EC, 0x00B590EC00B490EC, 0x00B790EC00B690EC,
-    0x00B990EC00B890EC, 0x00BB90EC00BA90EC, 0x00BD90EC00BC90EC,
-    0x00BF90EC00BE90EC, 0x008191EC008091EC, 0x008391EC008291EC,
-    0x008591EC008491EC, 0x008791EC008691EC, 0x008991EC008891EC,
-    0x008B91EC008A91EC, 0x008D91EC008C91EC, 0x008F91EC008E91EC,
-    0x009191EC009091EC, 0x009391EC009291EC, 0x009591EC009491EC,
-    0x009791EC009691EC, 0x009991EC009891EC, 0x009B91EC009A91EC,
-    0x009D91EC009C91EC, 0x009F91EC009E91EC, 0x00A191EC00A091EC,
-    0x00A391EC00A291EC, 0x00A591EC00A491EC, 0x00A791EC00A691EC,
-    0x00A991EC00A891EC, 0x00AB91EC00AA91EC, 0x00AD91EC00AC91EC,
-    0x00AF91EC00AE91EC, 0x00B191EC00B091EC, 0x00B391EC00B291EC,
-    0x00B591EC00B491EC, 0x00B791EC00B691EC, 0x00B991EC00B891EC,
-    0x00BB91EC00BA91EC, 0x00BD91EC00BC91EC, 0x00BF91EC00BE91EC,
-    0x008192EC008092EC, 0x008392EC008292EC, 0x008592EC008492EC,
-    0x008792EC008692EC, 0x008992EC008892EC, 0x008B92EC008A92EC,
-    0x008D92EC008C92EC, 0x008F92EC008E92EC, 0x009192EC009092EC,
-    0x009392EC009292EC, 0x009592EC009492EC, 0x009792EC009692EC,
-    0x009992EC009892EC, 0x009B92EC009A92EC, 0x009D92EC009C92EC,
-    0x009F92EC009E92EC, 0x00A192EC00A092EC, 0x00A392EC00A292EC,
-    0x00A592EC00A492EC, 0x00A792EC00A692EC, 0x00A992EC00A892EC,
-    0x00AB92EC00AA92EC, 0x00AD92EC00AC92EC, 0x00AF92EC00AE92EC,
-    0x00B192EC00B092EC, 0x00B392EC00B292EC, 0x00B592EC00B492EC,
-    0x00B792EC00B692EC, 0x00B992EC00B892EC, 0x00BB92EC00BA92EC,
-    0x00BD92EC00BC92EC, 0x00BF92EC00BE92EC, 0x008193EC008093EC,
-    0x008393EC008293EC, 0x008593EC008493EC, 0x008793EC008693EC,
-    0x008993EC008893EC, 0x008B93EC008A93EC, 0x008D93EC008C93EC,
-    0x008F93EC008E93EC, 0x009193EC009093EC, 0x009393EC009293EC,
-    0x009593EC009493EC, 0x009793EC009693EC, 0x009993EC009893EC,
-    0x009B93EC009A93EC, 0x009D93EC009C93EC, 0x009F93EC009E93EC,
-    0x00A193EC00A093EC, 0x00A393EC00A293EC, 0x00A593EC00A493EC,
-    0x00A793EC00A693EC, 0x00A993EC00A893EC, 0x00AB93EC00AA93EC,
-    0x00AD93EC00AC93EC, 0x00AF93EC00AE93EC, 0x00B193EC00B093EC,
-    0x00B393EC00B293EC, 0x00B593EC00B493EC, 0x00B793EC00B693EC,
-    0x00B993EC00B893EC, 0x00BB93EC00BA93EC, 0x00BD93EC00BC93EC,
-    0x00BF93EC00BE93EC, 0x008194EC008094EC, 0x008394EC008294EC,
-    0x008594EC008494EC, 0x008794EC008694EC, 0x008994EC008894EC,
-    0x008B94EC008A94EC, 0x008D94EC008C94EC, 0x008F94EC008E94EC,
-    0x009194EC009094EC, 0x009394EC009294EC, 0x009594EC009494EC,
-    0x009794EC009694EC, 0x009994EC009894EC, 0x009B94EC009A94EC,
-    0x009D94EC009C94EC, 0x009F94EC009E94EC, 0x00A194EC00A094EC,
-    0x00A394EC00A294EC, 0x00A594EC00A494EC, 0x00A794EC00A694EC,
-    0x00A994EC00A894EC, 0x00AB94EC00AA94EC, 0x00AD94EC00AC94EC,
-    0x00AF94EC00AE94EC, 0x00B194EC00B094EC, 0x00B394EC00B294EC,
-    0x00B594EC00B494EC, 0x00B794EC00B694EC, 0x00B994EC00B894EC,
-    0x00BB94EC00BA94EC, 0x00BD94EC00BC94EC, 0x00BF94EC00BE94EC,
-    0x008195EC008095EC, 0x008395EC008295EC, 0x008595EC008495EC,
-    0x008795EC008695EC, 0x008995EC008895EC, 0x008B95EC008A95EC,
-    0x008D95EC008C95EC, 0x008F95EC008E95EC, 0x009195EC009095EC,
-    0x009395EC009295EC, 0x009595EC009495EC, 0x009795EC009695EC,
-    0x009995EC009895EC, 0x009B95EC009A95EC, 0x009D95EC009C95EC,
-    0x009F95EC009E95EC, 0x00A195EC00A095EC, 0x00A395EC00A295EC,
-    0x00A595EC00A495EC, 0x00A795EC00A695EC, 0x00A995EC00A895EC,
-    0x00AB95EC00AA95EC, 0x00AD95EC00AC95EC, 0x00AF95EC00AE95EC,
-    0x00B195EC00B095EC, 0x00B395EC00B295EC, 0x00B595EC00B495EC,
-    0x00B795EC00B695EC, 0x00B995EC00B895EC, 0x00BB95EC00BA95EC,
-    0x00BD95EC00BC95EC, 0x00BF95EC00BE95EC, 0x008196EC008096EC,
-    0x008396EC008296EC, 0x008596EC008496EC, 0x008796EC008696EC,
-    0x008996EC008896EC, 0x008B96EC008A96EC, 0x008D96EC008C96EC,
-    0x008F96EC008E96EC, 0x009196EC009096EC, 0x009396EC009296EC,
-    0x009596EC009496EC, 0x009796EC009696EC, 0x009996EC009896EC,
-    0x009B96EC009A96EC, 0x009D96EC009C96EC, 0x009F96EC009E96EC,
-    0x00A196EC00A096EC, 0x00A396EC00A296EC, 0x00A596EC00A496EC,
-    0x00A796EC00A696EC, 0x00A996EC00A896EC, 0x00AB96EC00AA96EC,
-    0x00AD96EC00AC96EC, 0x00AF96EC00AE96EC, 0x00B196EC00B096EC,
-    0x00B396EC00B296EC, 0x00B596EC00B496EC, 0x00B796EC00B696EC,
-    0x00B996EC00B896EC, 0x00BB96EC00BA96EC, 0x00BD96EC00BC96EC,
-    0x00BF96EC00BE96EC, 0x008197EC008097EC, 0x008397EC008297EC,
-    0x008597EC008497EC, 0x008797EC008697EC, 0x008997EC008897EC,
-    0x008B97EC008A97EC, 0x008D97EC008C97EC, 0x008F97EC008E97EC,
-    0x009197EC009097EC, 0x009397EC009297EC, 0x009597EC009497EC,
-    0x009797EC009697EC, 0x009997EC009897EC, 0x009B97EC009A97EC,
-    0x009D97EC009C97EC, 0x009F97EC009E97EC, 0x00A197EC00A097EC,
-    0x00A397EC00A297EC, 0x00A597EC00A497EC, 0x00A797EC00A697EC,
-    0x00A997EC00A897EC, 0x00AB97EC00AA97EC, 0x00AD97EC00AC97EC,
-    0x00AF97EC00AE97EC, 0x00B197EC00B097EC, 0x00B397EC00B297EC,
-    0x00B597EC00B497EC, 0x00B797EC00B697EC, 0x00B997EC00B897EC,
-    0x00BB97EC00BA97EC, 0x00BD97EC00BC97EC, 0x00BF97EC00BE97EC,
-    0x008198EC008098EC, 0x008398EC008298EC, 0x008598EC008498EC,
-    0x008798EC008698EC, 0x008998EC008898EC, 0x008B98EC008A98EC,
-    0x008D98EC008C98EC, 0x008F98EC008E98EC, 0x009198EC009098EC,
-    0x009398EC009298EC, 0x009598EC009498EC, 0x009798EC009698EC,
-    0x009998EC009898EC, 0x009B98EC009A98EC, 0x009D98EC009C98EC,
-    0x009F98EC009E98EC, 0x00A198EC00A098EC, 0x00A398EC00A298EC,
-    0x00A598EC00A498EC, 0x00A798EC00A698EC, 0x00A998EC00A898EC,
-    0x00AB98EC00AA98EC, 0x00AD98EC00AC98EC, 0x00AF98EC00AE98EC,
-    0x00B198EC00B098EC, 0x00B398EC00B298EC, 0x00B598EC00B498EC,
-    0x00B798EC00B698EC, 0x00B998EC00B898EC, 0x00BB98EC00BA98EC,
-    0x00BD98EC00BC98EC, 0x00BF98EC00BE98EC, 0x008199EC008099EC,
-    0x008399EC008299EC, 0x008599EC008499EC, 0x008799EC008699EC,
-    0x008999EC008899EC, 0x008B99EC008A99EC, 0x008D99EC008C99EC,
-    0x008F99EC008E99EC, 0x009199EC009099EC, 0x009399EC009299EC,
-    0x009599EC009499EC, 0x009799EC009699EC, 0x009999EC009899EC,
-    0x009B99EC009A99EC, 0x009D99EC009C99EC, 0x009F99EC009E99EC,
-    0x00A199EC00A099EC, 0x00A399EC00A299EC, 0x00A599EC00A499EC,
-    0x00A799EC00A699EC, 0x00A999EC00A899EC, 0x00AB99EC00AA99EC,
-    0x00AD99EC00AC99EC, 0x00AF99EC00AE99EC, 0x00B199EC00B099EC,
-    0x00B399EC00B299EC, 0x00B599EC00B499EC, 0x00B799EC00B699EC,
-    0x00B999EC00B899EC, 0x00BB99EC00BA99EC, 0x00BD99EC00BC99EC,
-    0x00BF99EC00BE99EC, 0x00819AEC00809AEC, 0x00839AEC00829AEC,
-    0x00859AEC00849AEC, 0x00879AEC00869AEC, 0x00899AEC00889AEC,
-    0x008B9AEC008A9AEC, 0x008D9AEC008C9AEC, 0x008F9AEC008E9AEC,
-    0x00919AEC00909AEC, 0x00939AEC00929AEC, 0x00959AEC00949AEC,
-    0x00979AEC00969AEC, 0x00999AEC00989AEC, 0x009B9AEC009A9AEC,
-    0x009D9AEC009C9AEC, 0x009F9AEC009E9AEC, 0x00A19AEC00A09AEC,
-    0x00A39AEC00A29AEC, 0x00A59AEC00A49AEC, 0x00A79AEC00A69AEC,
-    0x00A99AEC00A89AEC, 0x00AB9AEC00AA9AEC, 0x00AD9AEC00AC9AEC,
-    0x00AF9AEC00AE9AEC, 0x00B19AEC00B09AEC, 0x00B39AEC00B29AEC,
-    0x00B59AEC00B49AEC, 0x00B79AEC00B69AEC, 0x00B99AEC00B89AEC,
-    0x00BB9AEC00BA9AEC, 0x00BD9AEC00BC9AEC, 0x00BF9AEC00BE9AEC,
-    0x00819BEC00809BEC, 0x00839BEC00829BEC, 0x00859BEC00849BEC,
-    0x00879BEC00869BEC, 0x00899BEC00889BEC, 0x008B9BEC008A9BEC,
-    0x008D9BEC008C9BEC, 0x008F9BEC008E9BEC, 0x00919BEC00909BEC,
-    0x00939BEC00929BEC, 0x00959BEC00949BEC, 0x00979BEC00969BEC,
-    0x00999BEC00989BEC, 0x009B9BEC009A9BEC, 0x009D9BEC009C9BEC,
-    0x009F9BEC009E9BEC, 0x00A19BEC00A09BEC, 0x00A39BEC00A29BEC,
-    0x00A59BEC00A49BEC, 0x00A79BEC00A69BEC, 0x00A99BEC00A89BEC,
-    0x00AB9BEC00AA9BEC, 0x00AD9BEC00AC9BEC, 0x00AF9BEC00AE9BEC,
-    0x00B19BEC00B09BEC, 0x00B39BEC00B29BEC, 0x00B59BEC00B49BEC,
-    0x00B79BEC00B69BEC, 0x00B99BEC00B89BEC, 0x00BB9BEC00BA9BEC,
-    0x00BD9BEC00BC9BEC, 0x00BF9BEC00BE9BEC, 0x00819CEC00809CEC,
-    0x00839CEC00829CEC, 0x00859CEC00849CEC, 0x00879CEC00869CEC,
-    0x00899CEC00889CEC, 0x008B9CEC008A9CEC, 0x008D9CEC008C9CEC,
-    0x008F9CEC008E9CEC, 0x00919CEC00909CEC, 0x00939CEC00929CEC,
-    0x00959CEC00949CEC, 0x00979CEC00969CEC, 0x00999CEC00989CEC,
-    0x009B9CEC009A9CEC, 0x009D9CEC009C9CEC, 0x009F9CEC009E9CEC,
-    0x00A19CEC00A09CEC, 0x00A39CEC00A29CEC, 0x00A59CEC00A49CEC,
-    0x00A79CEC00A69CEC, 0x00A99CEC00A89CEC, 0x00AB9CEC00AA9CEC,
-    0x00AD9CEC00AC9CEC, 0x00AF9CEC00AE9CEC, 0x00B19CEC00B09CEC,
-    0x00B39CEC00B29CEC, 0x00B59CEC00B49CEC, 0x00B79CEC00B69CEC,
-    0x00B99CEC00B89CEC, 0x00BB9CEC00BA9CEC, 0x00BD9CEC00BC9CEC,
-    0x00BF9CEC00BE9CEC, 0x00819DEC00809DEC, 0x00839DEC00829DEC,
-    0x00859DEC00849DEC, 0x00879DEC00869DEC, 0x00899DEC00889DEC,
-    0x008B9DEC008A9DEC, 0x008D9DEC008C9DEC, 0x008F9DEC008E9DEC,
-    0x00919DEC00909DEC, 0x00939DEC00929DEC, 0x00959DEC00949DEC,
-    0x00979DEC00969DEC, 0x00999DEC00989DEC, 0x009B9DEC009A9DEC,
-    0x009D9DEC009C9DEC, 0x009F9DEC009E9DEC, 0x00A19DEC00A09DEC,
-    0x00A39DEC00A29DEC, 0x00A59DEC00A49DEC, 0x00A79DEC00A69DEC,
-    0x00A99DEC00A89DEC, 0x00AB9DEC00AA9DEC, 0x00AD9DEC00AC9DEC,
-    0x00AF9DEC00AE9DEC, 0x00B19DEC00B09DEC, 0x00B39DEC00B29DEC,
-    0x00B59DEC00B49DEC, 0x00B79DEC00B69DEC, 0x00B99DEC00B89DEC,
-    0x00BB9DEC00BA9DEC, 0x00BD9DEC00BC9DEC, 0x00BF9DEC00BE9DEC,
-    0x00819EEC00809EEC, 0x00839EEC00829EEC, 0x00859EEC00849EEC,
-    0x00879EEC00869EEC, 0x00899EEC00889EEC, 0x008B9EEC008A9EEC,
-    0x008D9EEC008C9EEC, 0x008F9EEC008E9EEC, 0x00919EEC00909EEC,
-    0x00939EEC00929EEC, 0x00959EEC00949EEC, 0x00979EEC00969EEC,
-    0x00999EEC00989EEC, 0x009B9EEC009A9EEC, 0x009D9EEC009C9EEC,
-    0x009F9EEC009E9EEC, 0x00A19EEC00A09EEC, 0x00A39EEC00A29EEC,
-    0x00A59EEC00A49EEC, 0x00A79EEC00A69EEC, 0x00A99EEC00A89EEC,
-    0x00AB9EEC00AA9EEC, 0x00AD9EEC00AC9EEC, 0x00AF9EEC00AE9EEC,
-    0x00B19EEC00B09EEC, 0x00B39EEC00B29EEC, 0x00B59EEC00B49EEC,
-    0x00B79EEC00B69EEC, 0x00B99EEC00B89EEC, 0x00BB9EEC00BA9EEC,
-    0x00BD9EEC00BC9EEC, 0x00BF9EEC00BE9EEC, 0x00819FEC00809FEC,
-    0x00839FEC00829FEC, 0x00859FEC00849FEC, 0x00879FEC00869FEC,
-    0x00899FEC00889FEC, 0x008B9FEC008A9FEC, 0x008D9FEC008C9FEC,
-    0x008F9FEC008E9FEC, 0x00919FEC00909FEC, 0x00939FEC00929FEC,
-    0x00959FEC00949FEC, 0x00979FEC00969FEC, 0x00999FEC00989FEC,
-    0x009B9FEC009A9FEC, 0x009D9FEC009C9FEC, 0x009F9FEC009E9FEC,
-    0x00A19FEC00A09FEC, 0x00A39FEC00A29FEC, 0x00A59FEC00A49FEC,
-    0x00A79FEC00A69FEC, 0x00A99FEC00A89FEC, 0x00AB9FEC00AA9FEC,
-    0x00AD9FEC00AC9FEC, 0x00AF9FEC00AE9FEC, 0x00B19FEC00B09FEC,
-    0x00B39FEC00B29FEC, 0x00B59FEC00B49FEC, 0x00B79FEC00B69FEC,
-    0x00B99FEC00B89FEC, 0x00BB9FEC00BA9FEC, 0x00BD9FEC00BC9FEC,
-    0x00BF9FEC00BE9FEC, 0x0081A0EC0080A0EC, 0x0083A0EC0082A0EC,
-    0x0085A0EC0084A0EC, 0x0087A0EC0086A0EC, 0x0089A0EC0088A0EC,
-    0x008BA0EC008AA0EC, 0x008DA0EC008CA0EC, 0x008FA0EC008EA0EC,
-    0x0091A0EC0090A0EC, 0x0093A0EC0092A0EC, 0x0095A0EC0094A0EC,
-    0x0097A0EC0096A0EC, 0x0099A0EC0098A0EC, 0x009BA0EC009AA0EC,
-    0x009DA0EC009CA0EC, 0x009FA0EC009EA0EC, 0x00A1A0EC00A0A0EC,
-    0x00A3A0EC00A2A0EC, 0x00A5A0EC00A4A0EC, 0x00A7A0EC00A6A0EC,
-    0x00A9A0EC00A8A0EC, 0x00ABA0EC00AAA0EC, 0x00ADA0EC00ACA0EC,
-    0x00AFA0EC00AEA0EC, 0x00B1A0EC00B0A0EC, 0x00B3A0EC00B2A0EC,
-    0x00B5A0EC00B4A0EC, 0x00B7A0EC00B6A0EC, 0x00B9A0EC00B8A0EC,
-    0x00BBA0EC00BAA0EC, 0x00BDA0EC00BCA0EC, 0x00BFA0EC00BEA0EC,
-    0x0081A1EC0080A1EC, 0x0083A1EC0082A1EC, 0x0085A1EC0084A1EC,
-    0x0087A1EC0086A1EC, 0x0089A1EC0088A1EC, 0x008BA1EC008AA1EC,
-    0x008DA1EC008CA1EC, 0x008FA1EC008EA1EC, 0x0091A1EC0090A1EC,
-    0x0093A1EC0092A1EC, 0x0095A1EC0094A1EC, 0x0097A1EC0096A1EC,
-    0x0099A1EC0098A1EC, 0x009BA1EC009AA1EC, 0x009DA1EC009CA1EC,
-    0x009FA1EC009EA1EC, 0x00A1A1EC00A0A1EC, 0x00A3A1EC00A2A1EC,
-    0x00A5A1EC00A4A1EC, 0x00A7A1EC00A6A1EC, 0x00A9A1EC00A8A1EC,
-    0x00ABA1EC00AAA1EC, 0x00ADA1EC00ACA1EC, 0x00AFA1EC00AEA1EC,
-    0x00B1A1EC00B0A1EC, 0x00B3A1EC00B2A1EC, 0x00B5A1EC00B4A1EC,
-    0x00B7A1EC00B6A1EC, 0x00B9A1EC00B8A1EC, 0x00BBA1EC00BAA1EC,
-    0x00BDA1EC00BCA1EC, 0x00BFA1EC00BEA1EC, 0x0081A2EC0080A2EC,
-    0x0083A2EC0082A2EC, 0x0085A2EC0084A2EC, 0x0087A2EC0086A2EC,
-    0x0089A2EC0088A2EC, 0x008BA2EC008AA2EC, 0x008DA2EC008CA2EC,
-    0x008FA2EC008EA2EC, 0x0091A2EC0090A2EC, 0x0093A2EC0092A2EC,
-    0x0095A2EC0094A2EC, 0x0097A2EC0096A2EC, 0x0099A2EC0098A2EC,
-    0x009BA2EC009AA2EC, 0x009DA2EC009CA2EC, 0x009FA2EC009EA2EC,
-    0x00A1A2EC00A0A2EC, 0x00A3A2EC00A2A2EC, 0x00A5A2EC00A4A2EC,
-    0x00A7A2EC00A6A2EC, 0x00A9A2EC00A8A2EC, 0x00ABA2EC00AAA2EC,
-    0x00ADA2EC00ACA2EC, 0x00AFA2EC00AEA2EC, 0x00B1A2EC00B0A2EC,
-    0x00B3A2EC00B2A2EC, 0x00B5A2EC00B4A2EC, 0x00B7A2EC00B6A2EC,
-    0x00B9A2EC00B8A2EC, 0x00BBA2EC00BAA2EC, 0x00BDA2EC00BCA2EC,
-    0x00BFA2EC00BEA2EC, 0x0081A3EC0080A3EC, 0x0083A3EC0082A3EC,
-    0x0085A3EC0084A3EC, 0x0087A3EC0086A3EC, 0x0089A3EC0088A3EC,
-    0x008BA3EC008AA3EC, 0x008DA3EC008CA3EC, 0x008FA3EC008EA3EC,
-    0x0091A3EC0090A3EC, 0x0093A3EC0092A3EC, 0x0095A3EC0094A3EC,
-    0x0097A3EC0096A3EC, 0x0099A3EC0098A3EC, 0x009BA3EC009AA3EC,
-    0x009DA3EC009CA3EC, 0x009FA3EC009EA3EC, 0x00A1A3EC00A0A3EC,
-    0x00A3A3EC00A2A3EC, 0x00A5A3EC00A4A3EC, 0x00A7A3EC00A6A3EC,
-    0x00A9A3EC00A8A3EC, 0x00ABA3EC00AAA3EC, 0x00ADA3EC00ACA3EC,
-    0x00AFA3EC00AEA3EC, 0x00B1A3EC00B0A3EC, 0x00B3A3EC00B2A3EC,
-    0x00B5A3EC00B4A3EC, 0x00B7A3EC00B6A3EC, 0x00B9A3EC00B8A3EC,
-    0x00BBA3EC00BAA3EC, 0xECBCA3EC00BCA3EC, 0xEC00BDA3EC00989D,
-    0xEC00BFA3EC00BEA3, 0xEC0081A4EC0080A4, 0xEC0083A4EC0082A4,
-    0xEC0085A4EC0084A4, 0xEC0087A4EC0086A4, 0xEC0089A4EC0088A4,
-    0xEC008BA4EC008AA4, 0xEC008DA4EC008CA4, 0xEC008FA4EC008EA4,
-    0xEC0091A4EC0090A4, 0xEC0093A4EC0092A4, 0xEC0095A4EC0094A4,
-    0xEC0097A4EC0096A4, 0xEC0099A4EC0098A4, 0xEC009BA4EC009AA4,
-    0xEC009DA4EC009CA4, 0xEC009FA4EC009EA4, 0xEC00A1A4EC00A0A4,
-    0xEC00A3A4EC00A2A4, 0xEC00A5A4EC00A4A4, 0xEC00A7A4EC00A6A4,
-    0xEC00A9A4EC00A8A4, 0xEC00ABA4EC00AAA4, 0xEC00ADA4EC00ACA4,
-    0xEC00AFA4EC00AEA4, 0xEC00B1A4EC00B0A4, 0xEC00B3A4EC00B2A4,
-    0xEC00B5A4EC00B4A4, 0xEC00B7A4EC00B6A4, 0xEC00B9A4EC00B8A4,
-    0xEC00BBA4EC00BAA4, 0xEC00BDA4EC00BCA4, 0xEC00BFA4EC00BEA4,
-    0xEC0081A5EC0080A5, 0xEC0083A5EC0082A5, 0xEC0085A5EC0084A5,
-    0xEC0087A5EC0086A5, 0xEC0089A5EC0088A5, 0xEC008BA5EC008AA5,
-    0xEC008DA5EC008CA5, 0xEC008FA5EC008EA5, 0xEC0091A5EC0090A5,
-    0xEC0093A5EC0092A5, 0xEC0095A5EC0094A5, 0xEC0097A5EC0096A5,
-    0xEC0099A5EC0098A5, 0xEC009BA5EC009AA5, 0xEC009DA5EC009CA5,
-    0xEC009FA5EC009EA5, 0xEC00A1A5EC00A0A5, 0xEC00A3A5EC00A2A5,
-    0xEC00A5A5EC00A4A5, 0xEC00A7A5EC00A6A5, 0xEC00A9A5EC00A8A5,
-    0xEC00ABA5EC00AAA5, 0xEC00ADA5EC00ACA5, 0xEC00AFA5EC00AEA5,
-    0xEC00B1A5EC00B0A5, 0xEC00B3A5EC00B2A5, 0xEC00B5A5EC00B4A5,
-    0xEC00B7A5EC00B6A5, 0xEC00B9A5EC00B8A5, 0xEC00BBA5EC00BAA5,
-    0xEC00BDA5EC00BCA5, 0xEC00BFA5EC00BEA5, 0xEC0081A6EC0080A6,
-    0xEC0083A6EC0082A6, 0xEC0085A6EC0084A6, 0xEC0087A6EC0086A6,
-    0xEC0089A6EC0088A6, 0xEC008BA6EC008AA6, 0xEC008DA6EC008CA6,
-    0xEC008FA6EC008EA6, 0xEC0091A6EC0090A6, 0xEC0093A6EC0092A6,
-    0xEC0095A6EC0094A6, 0xEC0097A6EC0096A6, 0xEC0099A6EC0098A6,
-    0xEC009BA6EC009AA6, 0xEC009DA6EC009CA6, 0xEC009FA6EC009EA6,
-    0xEC00A1A6EC00A0A6, 0xEC00A3A6EC00A2A6, 0xEC00A5A6EC00A4A6,
-    0xEC00A7A6EC00A6A6, 0xEC00A9A6EC00A8A6, 0xEC00ABA6EC00AAA6,
-    0xEC00ADA6EC00ACA6, 0xEC00AFA6EC00AEA6, 0xEC00B1A6EC00B0A6,
-    0xEC00B3A6EC00B2A6, 0xEC00B5A6EC00B4A6, 0xEC00B7A6EC00B6A6,
-    0xEC00B9A6EC00B8A6, 0xEC00BBA6EC00BAA6, 0xEC00BDA6EC00BCA6,
-    0xEC00BFA6EC00BEA6, 0xEC0081A7EC0080A7, 0xEC0083A7EC0082A7,
-    0xEC0085A7EC0084A7, 0xEC0087A7EC0086A7, 0xEC0089A7EC0088A7,
-    0xEC008BA7EC008AA7, 0xEC008DA7EC008CA7, 0xEC008FA7EC008EA7,
-    0xEC0091A7EC0090A7, 0xEC0093A7EC0092A7, 0xEC0095A7EC0094A7,
-    0xEC0097A7EC0096A7, 0xEC0099A7EC0098A7, 0xEC009BA7EC009AA7,
-    0xEC009DA7EC009CA7, 0xEC009FA7EC009EA7, 0xEC00A1A7EC00A0A7,
-    0xEC00A3A7EC00A2A7, 0xEC00A5A7EC00A4A7, 0xEC00A7A7EC00A6A7,
-    0xEC00A9A7EC00A8A7, 0xEC00ABA7EC00AAA7, 0xEC00ADA7EC00ACA7,
-    0xEC00AFA7EC00AEA7, 0xEC00B1A7EC00B0A7, 0xEC00B3A7EC00B2A7,
-    0xEC00B5A7EC00B4A7, 0xEC00B7A7EC00B6A7, 0xEC00B9A7EC00B8A7,
-    0xEC00BBA7EC00BAA7, 0xEC00BDA7EC00BCA7, 0xEC00BFA7EC00BEA7,
-    0xEC0081A8EC0080A8, 0xEC0083A8EC0082A8, 0xEC0085A8EC0084A8,
-    0xEC0087A8EC0086A8, 0xEC0089A8EC0088A8, 0xEC008BA8EC008AA8,
-    0xEC008DA8EC008CA8, 0xEC008FA8EC008EA8, 0xEC0091A8EC0090A8,
-    0xEC0093A8EC0092A8, 0xEC0095A8EC0094A8, 0xEC0097A8EC0096A8,
-    0xEC0099A8EC0098A8, 0xEC009BA8EC009AA8, 0xEC009DA8EC009CA8,
-    0xEC009FA8EC009EA8, 0xEC00A1A8EC00A0A8, 0xEC00A3A8EC00A2A8,
-    0xEC00A5A8EC00A4A8, 0xEC00A7A8EC00A6A8, 0xEC00A9A8EC00A8A8,
-    0xEC00ABA8EC00AAA8, 0xEC00ADA8EC00ACA8, 0xEC00AFA8EC00AEA8,
-    0xEC00B1A8EC00B0A8, 0xEC00B3A8EC00B2A8, 0xEC00B5A8EC00B4A8,
-    0xEC00B7A8EC00B6A8, 0xEC00B9A8EC00B8A8, 0xEC00BBA8EC00BAA8,
-    0xEC00BDA8EC00BCA8, 0xEC00BFA8EC00BEA8, 0xEC0081A9EC0080A9,
-    0xEC0083A9EC0082A9, 0xEC0085A9EC0084A9, 0xEC0087A9EC0086A9,
-    0xEC0089A9EC0088A9, 0xEC008BA9EC008AA9, 0xEC008DA9EC008CA9,
-    0xEC008FA9EC008EA9, 0xEC0091A9EC0090A9, 0xEC0093A9EC0092A9,
-    0xEC0095A9EC0094A9, 0xEC0097A9EC0096A9, 0xEC0099A9EC0098A9,
-    0xEC009BA9EC009AA9, 0xEC009DA9EC009CA9, 0xEC009FA9EC009EA9,
-    0xEC00A1A9EC00A0A9, 0xEC00A3A9EC00A2A9, 0xEC00A5A9EC00A4A9,
-    0xEC00A7A9EC00A6A9, 0xEC00A9A9EC00A8A9, 0xEC00ABA9EC00AAA9,
-    0xEC00ADA9EC00ACA9, 0xEC00AFA9EC00AEA9, 0xEC00B1A9EC00B0A9,
-    0xEC00B3A9EC00B2A9, 0xEC00B5A9EC00B4A9, 0xEC00B7A9EC00B6A9,
-    0xEC00B9A9EC00B8A9, 0xEC00BBA9EC00BAA9, 0xEC00BDA9EC00BCA9,
-    0xEC00BFA9EC00BEA9, 0xEC0081AAEC0080AA, 0xEC0083AAEC0082AA,
-    0xEC0085AAEC0084AA, 0xEC0087AAEC0086AA, 0xEC0089AAEC0088AA,
-    0xEC008BAAEC008AAA, 0xEC008DAAEC008CAA, 0xEC008FAAEC008EAA,
-    0xEC0091AAEC0090AA, 0xEC0093AAEC0092AA, 0xEC0095AAEC0094AA,
-    0xEC0097AAEC0096AA, 0xEC0099AAEC0098AA, 0xEC009BAAEC009AAA,
-    0xEC009DAAEC009CAA, 0xEC009FAAEC009EAA, 0xEC00A1AAEC00A0AA,
-    0xEC00A3AAEC00A2AA, 0xEC00A5AAEC00A4AA, 0xEC00A7AAEC00A6AA,
-    0xEC00A9AAEC00A8AA, 0xEC00ABAAEC00AAAA, 0xEC00ADAAEC00ACAA,
-    0xEC00AFAAEC00AEAA, 0xEC00B1AAEC00B0AA, 0xEC00B3AAEC00B2AA,
-    0xEC00B5AAEC00B4AA, 0xEC00B7AAEC00B6AA, 0xEC00B9AAEC00B8AA,
-    0xEC00BBAAEC00BAAA, 0xEC00BDAAEC00BCAA, 0xEC00BFAAEC00BEAA,
-    0xEC0081ABEC0080AB, 0xEC0083ABEC0082AB, 0xEC0085ABEC0084AB,
-    0xEC0087ABEC0086AB, 0xEC0089ABEC0088AB, 0xEC008BABEC008AAB,
-    0xEC008DABEC008CAB, 0xEC008FABEC008EAB, 0xEC0091ABEC0090AB,
-    0xEC0093ABEC0092AB, 0xEC0095ABEC0094AB, 0xEC0097ABEC0096AB,
-    0xEC0099ABEC0098AB, 0xEC009BABEC009AAB, 0xEC009DABEC009CAB,
-    0xEC009FABEC009EAB, 0xEC00A1ABEC00A0AB, 0xEC00A3ABEC00A2AB,
-    0xEC00A5ABEC00A4AB, 0xEC00A7ABEC00A6AB, 0xEC00A9ABEC00A8AB,
-    0xEC00ABABEC00AAAB, 0xEC00ADABEC00ACAB, 0xEC00AFABEC00AEAB,
-    0xEC00B1ABEC00B0AB, 0xEC00B3ABEC00B2AB, 0xEC00B5ABEC00B4AB,
-    0xEC00B7ABEC00B6AB, 0xEC00B9ABEC00B8AB, 0xEC00BBABEC00BAAB,
-    0xEC00BDABEC00BCAB, 0xEC00BFABEC00BEAB, 0xEC0081ACEC0080AC,
-    0xEC0083ACEC0082AC, 0xEC0085ACEC0084AC, 0xEC0087ACEC0086AC,
-    0xEC0089ACEC0088AC, 0xEC008BACEC008AAC, 0xEC008DACEC008CAC,
-    0xEC008FACEC008EAC, 0xEC0091ACEC0090AC, 0xEC0093ACEC0092AC,
-    0xEC0095ACEC0094AC, 0xEC0097ACEC0096AC, 0xEC0099ACEC0098AC,
-    0xEC009BACEC009AAC, 0xEC009DACEC009CAC, 0xEC009FACEC009EAC,
-    0xEC00A1ACEC00A0AC, 0xEC00A3ACEC00A2AC, 0xEC00A5ACEC00A4AC,
-    0xEC00A7ACEC00A6AC, 0xEC00A9ACEC00A8AC, 0xEC00ABACEC00AAAC,
-    0xEC00ADACEC00ACAC, 0xEC00AFACEC00AEAC, 0xEC00B1ACEC00B0AC,
-    0xEC00B3ACEC00B2AC, 0xEC00B5ACEC00B4AC, 0xEC00B7ACEC00B6AC,
-    0xEC00B9ACEC00B8AC, 0xEC00BBACEC00BAAC, 0xEC00BDACEC00BCAC,
-    0xEC00BFACEC00BEAC, 0xEC0081ADEC0080AD, 0xEC0083ADEC0082AD,
-    0xEC0085ADEC0084AD, 0xEC0087ADEC0086AD, 0xEC0089ADEC0088AD,
-    0xEC008BADEC008AAD, 0xEC008DADEC008CAD, 0xEC008FADEC008EAD,
-    0xEC0091ADEC0090AD, 0xEC0093ADEC0092AD, 0xEC0095ADEC0094AD,
-    0xEC0097ADEC0096AD, 0xEC0099ADEC0098AD, 0xEC009BADEC009AAD,
-    0xEC009DADEC009CAD, 0xEC009FADEC009EAD, 0xEC00A1ADEC00A0AD,
-    0xEC00A3ADEC00A2AD, 0xEC00A5ADEC00A4AD, 0xEC00A7ADEC00A6AD,
-    0xEC00A9ADEC00A8AD, 0xEC00ABADEC00AAAD, 0xEC00ADADEC00ACAD,
-    0xEC00AFADEC00AEAD, 0xEC00B1ADEC00B0AD, 0xEC00B3ADEC00B2AD,
-    0xEC00B5ADEC00B4AD, 0xEC00B7ADEC00B6AD, 0xEC00B9ADEC00B8AD,
-    0xEC00BBADEC00BAAD, 0xEC00BDADEC00BCAD, 0xEC00BFADEC00BEAD,
-    0xEC0081AEEC0080AE, 0xEC0083AEEC0082AE, 0xEC0085AEEC0084AE,
-    0xEC0087AEEC0086AE, 0xEC0089AEEC0088AE, 0xEC008BAEEC008AAE,
-    0xEC008DAEEC008CAE, 0xEC008FAEEC008EAE, 0xEC0091AEEC0090AE,
-    0xEC0093AEEC0092AE, 0xEC0095AEEC0094AE, 0xEC0097AEEC0096AE,
-    0xEC0099AEEC0098AE, 0xEC009BAEEC009AAE, 0xEC009DAEEC009CAE,
-    0xEC009FAEEC009EAE, 0xEC00A1AEEC00A0AE, 0xEC00A3AEEC00A2AE,
-    0xEC00A5AEEC00A4AE, 0xEC00A7AEEC00A6AE, 0xEC00A9AEEC00A8AE,
-    0xEC00ABAEEC00AAAE, 0xEC00ADAEEC00ACAE, 0xEC00AFAEEC00AEAE,
-    0xEC00B1AEEC00B0AE, 0xEC00B3AEEC00B2AE, 0xEC00B5AEEC00B4AE,
-    0xEC00B7AEEC00B6AE, 0xEC00B9AEEC00B8AE, 0xEC00BBAEEC00BAAE,
-    0xEC00BDAEEC00BCAE, 0xEC00BFAEEC00BEAE, 0xEC0081AFEC0080AF,
-    0xEC0083AFEC0082AF, 0xEC0085AFEC0084AF, 0xEC0087AFEC0086AF,
-    0xEC0089AFEC0088AF, 0xEC008BAFEC008AAF, 0xEC008DAFEC008CAF,
-    0xEC008FAFEC008EAF, 0xEC0091AFEC0090AF, 0xEC0093AFEC0092AF,
-    0xEC0095AFEC0094AF, 0xEC0097AFEC0096AF, 0xEC0099AFEC0098AF,
-    0xEC009BAFEC009AAF, 0xEC009DAFEC009CAF, 0xEC009FAFEC009EAF,
-    0xEC00A1AFEC00A0AF, 0xEC00A3AFEC00A2AF, 0xEC00A5AFEC00A4AF,
-    0xEC00A7AFEC00A6AF, 0xEC00A9AFEC00A8AF, 0xEC00ABAFEC00AAAF,
-    0xEC00ADAFEC00ACAF, 0xEC00AFAFEC00AEAF, 0xEC00B1AFEC00B0AF,
-    0xEC00B3AFEC00B2AF, 0xEC00B5AFEC00B4AF, 0xEC00B7AFEC00B6AF,
-    0xEC00B9AFEC00B8AF, 0xEC00BBAFEC00BAAF, 0xEC00BDAFEC00BCAF,
-    0xEC00BFAFEC00BEAF, 0xEC0081B0EC0080B0, 0xEC0083B0EC0082B0,
-    0xEC0085B0EC0084B0, 0xEC0087B0EC0086B0, 0xEC0089B0EC0088B0,
-    0xEC008BB0EC008AB0, 0xEC008DB0EC008CB0, 0xEC008FB0EC008EB0,
-    0xEC0091B0EC0090B0, 0xEC0093B0EC0092B0, 0xEC0095B0EC0094B0,
-    0xEC0097B0EC0096B0, 0xEC0099B0EC0098B0, 0xEC009BB0EC009AB0,
-    0xEC009DB0EC009CB0, 0xEC009FB0EC009EB0, 0xEC00A1B0EC00A0B0,
-    0xEC00A3B0EC00A2B0, 0xEC00A5B0EC00A4B0, 0xEC00A7B0EC00A6B0,
-    0xEC00A9B0EC00A8B0, 0xEC00ABB0EC00AAB0, 0xEC00ADB0EC00ACB0,
-    0xEC00AFB0EC00AEB0, 0xEC00B1B0EC00B0B0, 0xEC00B3B0EC00B2B0,
-    0xEC00B5B0EC00B4B0, 0xEC00B7B0EC00B6B0, 0xB3EAB8B0EC00B8B0,
-    0xB0EC00B9B0EC00A0, 0xB0EC00BBB0EC00BA, 0xB0EC00BDB0EC00BC,
-    0xB1EC00BFB0EC00BE, 0xB1EC0081B1EC0080, 0xB1EC0083B1EC0082,
-    0xB1EC0085B1EC0084, 0xB1EC0087B1EC0086, 0xB1EC0089B1EC0088,
-    0xB1EC008BB1EC008A, 0xB1EC008DB1EC008C, 0xB1EC008FB1EC008E,
-    0xB1EC0091B1EC0090, 0xB1EC0093B1EC0092, 0xB1EC0095B1EC0094,
-    0xB1EC0097B1EC0096, 0xB1EC0099B1EC0098, 0xB1EC009BB1EC009A,
-    0xB1EC009DB1EC009C, 0xB1EC009FB1EC009E, 0xB1EC00A1B1EC00A0,
-    0xB1EC00A3B1EC00A2, 0xB1EC00A5B1EC00A4, 0xB1EC00A7B1EC00A6,
-    0xB1EC00A9B1EC00A8, 0xB1EC00ABB1EC00AA, 0xB1EC00ADB1EC00AC,
-    0xB1EC00AFB1EC00AE, 0xB1EC00B1B1EC00B0, 0xB1EC00B3B1EC00B2,
-    0xB1EC00B5B1EC00B4, 0xB1EC00B7B1EC00B6, 0xB1EC00B9B1EC00B8,
-    0xB1EC00BBB1EC00BA, 0xB1EC00BDB1EC00BC, 0xB2EC00BFB1EC00BE,
-    0xB2EC0081B2EC0080, 0xB2EC0083B2EC0082, 0xB2EC0085B2EC0084,
-    0xB2EC0087B2EC0086, 0xB2EC0089B2EC0088, 0xB2EC008BB2EC008A,
-    0xB2EC008DB2EC008C, 0xB2EC008FB2EC008E, 0xB2EC0091B2EC0090,
-    0xB2EC0093B2EC0092, 0xB2EC0095B2EC0094, 0xB2EC0097B2EC0096,
-    0xB2EC0099B2EC0098, 0xB2EC009BB2EC009A, 0xB2EC009DB2EC009C,
-    0xB2EC009FB2EC009E, 0xB2EC00A1B2EC00A0, 0xB2EC00A3B2EC00A2,
-    0xB2EC00A5B2EC00A4, 0xB2EC00A7B2EC00A6, 0xB2EC00A9B2EC00A8,
-    0xB2EC00ABB2EC00AA, 0xB2EC00ADB2EC00AC, 0xB2EC00AFB2EC00AE,
-    0xB2EC00B1B2EC00B0, 0xB2EC00B3B2EC00B2, 0xB2EC00B5B2EC00B4,
-    0xB2EC00B7B2EC00B6, 0xB2EC00B9B2EC00B8, 0xB2EC00BBB2EC00BA,
-    0xB2EC00BDB2EC00BC, 0xB3EC00BFB2EC00BE, 0xB3EC0081B3EC0080,
-    0xB3EC0083B3EC0082, 0xB3EC0085B3EC0084, 0xB3EC0087B3EC0086,
-    0xB3EC0089B3EC0088, 0xB3EC008BB3EC008A, 0xB3EC008DB3EC008C,
-    0xB3EC008FB3EC008E, 0xB3EC0091B3EC0090, 0xB3EC0093B3EC0092,
-    0xB3EC0095B3EC0094, 0xB3EC0097B3EC0096, 0xB3EC0099B3EC0098,
-    0xB3EC009BB3EC009A, 0xB3EC009DB3EC009C, 0xB3EC009FB3EC009E,
-    0xB3EC00A1B3EC00A0, 0xB3EC00A3B3EC00A2, 0xB3EC00A5B3EC00A4,
-    0xB3EC00A7B3EC00A6, 0xB3EC00A9B3EC00A8, 0xB3EC00ABB3EC00AA,
-    0xB3EC00ADB3EC00AC, 0xB3EC00AFB3EC00AE, 0xB3EC00B1B3EC00B0,
-    0xB3EC00B3B3EC00B2, 0xB3EC00B5B3EC00B4, 0xB3EC00B7B3EC00B6,
-    0xB3EC00B9B3EC00B8, 0xB3EC00BBB3EC00BA, 0xB3EC00BDB3EC00BC,
-    0xB4EC00BFB3EC00BE, 0xB4EC0081B4EC0080, 0xB4EC0083B4EC0082,
-    0xB4EC0085B4EC0084, 0xB4EC0087B4EC0086, 0xB4EC0089B4EC0088,
-    0xB4EC008BB4EC008A, 0xB4EC008DB4EC008C, 0xB4EC008FB4EC008E,
-    0xB4EC0091B4EC0090, 0xB4EC0093B4EC0092, 0xB4EC0095B4EC0094,
-    0xB4EC0097B4EC0096, 0xB4EC0099B4EC0098, 0xB4EC009BB4EC009A,
-    0xB4EC009DB4EC009C, 0xB4EC009FB4EC009E, 0xB4EC00A1B4EC00A0,
-    0xB4EC00A3B4EC00A2, 0xB4EC00A5B4EC00A4, 0xB4EC00A7B4EC00A6,
-    0xB4EC00A9B4EC00A8, 0xB4EC00ABB4EC00AA, 0xB4EC00ADB4EC00AC,
-    0xB4EC00AFB4EC00AE, 0xB4EC00B1B4EC00B0, 0xB4EC00B3B4EC00B2,
-    0xB4EC00B5B4EC00B4, 0xB4EC00B7B4EC00B6, 0xB4EC00B9B4EC00B8,
-    0xB4EC00BBB4EC00BA, 0xB4EC00BDB4EC00BC, 0xB5EC00BFB4EC00BE,
-    0xB5EC0081B5EC0080, 0xB5EC0083B5EC0082, 0xB5EC0085B5EC0084,
-    0xB5EC0087B5EC0086, 0xB5EC0089B5EC0088, 0xB5EC008BB5EC008A,
-    0xB5EC008DB5EC008C, 0xB5EC008FB5EC008E, 0xB5EC0091B5EC0090,
-    0xB5EC0093B5EC0092, 0xB5EC0095B5EC0094, 0xB5EC0097B5EC0096,
-    0xB5EC0099B5EC0098, 0xB5EC009BB5EC009A, 0xB5EC009DB5EC009C,
-    0xB5EC009FB5EC009E, 0xB5EC00A1B5EC00A0, 0xB5EC00A3B5EC00A2,
-    0xB5EC00A5B5EC00A4, 0xB5EC00A7B5EC00A6, 0xB5EC00A9B5EC00A8,
-    0xB5EC00ABB5EC00AA, 0xB5EC00ADB5EC00AC, 0xB5EC00AFB5EC00AE,
-    0xB5EC00B1B5EC00B0, 0xB5EC00B3B5EC00B2, 0xB5EC00B5B5EC00B4,
-    0xB5EC00B7B5EC00B6, 0xB5EC00B9B5EC00B8, 0xB5EC00BBB5EC00BA,
-    0xB5EC00BDB5EC00BC, 0xB6EC00BFB5EC00BE, 0xB6EC0081B6EC0080,
-    0xB6EC0083B6EC0082, 0xB6EC0085B6EC0084, 0xB6EC0087B6EC0086,
-    0xB6EC0089B6EC0088, 0xB6EC008BB6EC008A, 0xB6EC008DB6EC008C,
-    0xB6EC008FB6EC008E, 0xB6EC0091B6EC0090, 0xB6EC0093B6EC0092,
-    0xB6EC0095B6EC0094, 0xB6EC0097B6EC0096, 0xB6EC0099B6EC0098,
-    0xB6EC009BB6EC009A, 0xB6EC009DB6EC009C, 0xB6EC009FB6EC009E,
-    0xB6EC00A1B6EC00A0, 0xB6EC00A3B6EC00A2, 0xB6EC00A5B6EC00A4,
-    0xB6EC00A7B6EC00A6, 0xB6EC00A9B6EC00A8, 0xB6EC00ABB6EC00AA,
-    0xB6EC00ADB6EC00AC, 0xB6EC00AFB6EC00AE, 0xB6EC00B1B6EC00B0,
-    0xB6EC00B3B6EC00B2, 0xB6EC00B5B6EC00B4, 0xB6EC00B7B6EC00B6,
-    0xB6EC00B9B6EC00B8, 0xB6EC00BBB6EC00BA, 0xB6EC00BDB6EC00BC,
-    0xB7EC00BFB6EC00BE, 0xB7EC0081B7EC0080, 0xB7EC0083B7EC0082,
-    0xB7EC0085B7EC0084, 0xB7EC0087B7EC0086, 0xB7EC0089B7EC0088,
-    0xB7EC008BB7EC008A, 0xB7EC008DB7EC008C, 0xB7EC008FB7EC008E,
-    0xB7EC0091B7EC0090, 0xB7EC0093B7EC0092, 0xB7EC0095B7EC0094,
-    0xB7EC0097B7EC0096, 0xB7EC0099B7EC0098, 0xB7EC009BB7EC009A,
-    0xB7EC009DB7EC009C, 0xB7EC009FB7EC009E, 0xB7EC00A1B7EC00A0,
-    0xB7EC00A3B7EC00A2, 0xB7EC00A5B7EC00A4, 0xB7EC00A7B7EC00A6,
-    0xB7EC00A9B7EC00A8, 0xB7EC00ABB7EC00AA, 0xB7EC00ADB7EC00AC,
-    0xB7EC00AFB7EC00AE, 0xB7EC00B1B7EC00B0, 0xB7EC00B3B7EC00B2,
-    0xB7EC00B5B7EC00B4, 0xB7EC00B7B7EC00B6, 0xB7EC00B9B7EC00B8,
-    0xB7EC00BBB7EC00BA, 0xB7EC00BDB7EC00BC, 0xB8EC00BFB7EC00BE,
-    0xB8EC0081B8EC0080, 0xB8EC0083B8EC0082, 0xB8EC0085B8EC0084,
-    0xB8EC0087B8EC0086, 0xB8EC0089B8EC0088, 0xB8EC008BB8EC008A,
-    0xB8EC008DB8EC008C, 0xB8EC008FB8EC008E, 0xB8EC0091B8EC0090,
-    0xB8EC0093B8EC0092, 0xB8EC0095B8EC0094, 0xB8EC0097B8EC0096,
-    0xB8EC0099B8EC0098, 0xB8EC009BB8EC009A, 0xB8EC009DB8EC009C,
-    0xB8EC009FB8EC009E, 0xB8EC00A1B8EC00A0, 0xB8EC00A3B8EC00A2,
-    0xB8EC00A5B8EC00A4, 0xB8EC00A7B8EC00A6, 0xB8EC00A9B8EC00A8,
-    0xB8EC00ABB8EC00AA, 0xB8EC00ADB8EC00AC, 0xB8EC00AFB8EC00AE,
-    0xB8EC00B1B8EC00B0, 0xB8EC00B3B8EC00B2, 0xB8EC00B5B8EC00B4,
-    0xB8EC00B7B8EC00B6, 0xB8EC00B9B8EC00B8, 0xB8EC00BBB8EC00BA,
-    0xB8EC00BDB8EC00BC, 0xB9EC00BFB8EC00BE, 0xB9EC0081B9EC0080,
-    0xB9EC0083B9EC0082, 0xB9EC0085B9EC0084, 0xB9EC0087B9EC0086,
-    0xB9EC0089B9EC0088, 0xB9EC008BB9EC008A, 0xB9EC008DB9EC008C,
-    0xB9EC008FB9EC008E, 0xB9EC0091B9EC0090, 0xB9EC0093B9EC0092,
-    0xB9EC0095B9EC0094, 0xB9EC0097B9EC0096, 0xB9EC0099B9EC0098,
-    0xB9EC009BB9EC009A, 0xB9EC009DB9EC009C, 0xB9EC009FB9EC009E,
-    0xB9EC00A1B9EC00A0, 0xB9EC00A3B9EC00A2, 0xB9EC00A5B9EC00A4,
-    0xB9EC00A7B9EC00A6, 0xB9EC00A9B9EC00A8, 0xB9EC00ABB9EC00AA,
-    0xB9EC00ADB9EC00AC, 0xB9EC00AFB9EC00AE, 0xB9EC00B1B9EC00B0,
-    0xB9EC00B3B9EC00B2, 0xB9EC00B5B9EC00B4, 0xB9EC00B7B9EC00B6,
-    0xB9EC00B9B9EC00B8, 0xB9EC00BBB9EC00BA, 0xB9EC00BDB9EC00BC,
-    0xBAEC00BFB9EC00BE, 0xBAEC0081BAEC0080, 0xBAEC0083BAEC0082,
-    0xBAEC0085BAEC0084, 0xBAEC0087BAEC0086, 0xBAEC0089BAEC0088,
-    0xBAEC008BBAEC008A, 0xBAEC008DBAEC008C, 0xBAEC008FBAEC008E,
-    0xBAEC0091BAEC0090, 0xBAEC0093BAEC0092, 0xBAEC0095BAEC0094,
-    0xBAEC0097BAEC0096, 0xBAEC0099BAEC0098, 0xBAEC009BBAEC009A,
-    0xBAEC009DBAEC009C, 0xBAEC009FBAEC009E, 0xBAEC00A1BAEC00A0,
-    0xBAEC00A3BAEC00A2, 0xBAEC00A5BAEC00A4, 0xBAEC00A7BAEC00A6,
-    0xBAEC00A9BAEC00A8, 0xBAEC00ABBAEC00AA, 0xBAEC00ADBAEC00AC,
-    0xBAEC00AFBAEC00AE, 0xBAEC00B1BAEC00B0, 0xBAEC00B3BAEC00B2,
-    0xBAEC00B5BAEC00B4, 0xBAEC00B7BAEC00B6, 0xBAEC00B9BAEC00B8,
-    0xBAEC00BBBAEC00BA, 0xBAEC00BDBAEC00BC, 0xBBEC00BFBAEC00BE,
-    0xBBEC0081BBEC0080, 0xBBEC0083BBEC0082, 0xBBEC0085BBEC0084,
-    0xBBEC0087BBEC0086, 0xBBEC0089BBEC0088, 0xBBEC008BBBEC008A,
-    0xBBEC008DBBEC008C, 0xBBEC008FBBEC008E, 0xBBEC0091BBEC0090,
-    0xBBEC0093BBEC0092, 0xBBEC0095BBEC0094, 0xBBEC0097BBEC0096,
-    0xBBEC0099BBEC0098, 0xBBEC009BBBEC009A, 0xBBEC009DBBEC009C,
-    0xBBEC009FBBEC009E, 0xBBEC00A1BBEC00A0, 0xBBEC00A3BBEC00A2,
-    0xBBEC00A5BBEC00A4, 0xBBEC00A7BBEC00A6, 0xBBEC00A9BBEC00A8,
-    0xBBEC00ABBBEC00AA, 0xBBEC00ADBBEC00AC, 0xBBEC00AFBBEC00AE,
-    0xBBEC00B1BBEC00B0, 0xBBEC00B3BBEC00B2, 0xBBEC00B5BBEC00B4,
-    0xBBEC00B7BBEC00B6, 0xBBEC00B9BBEC00B8, 0xBBEC00BBBBEC00BA,
-    0xBBEC00BDBBEC00BC, 0xBCEC00BFBBEC00BE, 0xBCEC0081BCEC0080,
-    0xBCEC0083BCEC0082, 0xBCEC0085BCEC0084, 0xBCEC0087BCEC0086,
-    0xBCEC0089BCEC0088, 0xBCEC008BBCEC008A, 0xBCEC008DBCEC008C,
-    0xBCEC008FBCEC008E, 0xBCEC0091BCEC0090, 0xBCEC0093BCEC0092,
-    0xBCEC0095BCEC0094, 0xBCEC0097BCEC0096, 0xBCEC0099BCEC0098,
-    0xBCEC009BBCEC009A, 0xBCEC009DBCEC009C, 0xBCEC009FBCEC009E,
-    0xBCEC00A1BCEC00A0, 0xBCEC00A3BCEC00A2, 0xBCEC00A5BCEC00A4,
-    0xBCEC00A7BCEC00A6, 0xBCEC00A9BCEC00A8, 0xBCEC00ABBCEC00AA,
-    0xBCEC00ADBCEC00AC, 0xBCEC00AFBCEC00AE, 0xBCEC00B1BCEC00B0,
-    0xBCEC00B3BCEC00B2, 0xBCEC00B5BCEC00B4, 0xBCEC00B7BCEC00B6,
-    0xBCEC00B9BCEC00B8, 0xBCEC00BBBCEC00BA, 0xBCEC00BDBCEC00BC,
-    0xBDEC00BFBCEC00BE, 0xBDEC0081BDEC0080, 0xBDEC0083BDEC0082,
-    0xBDEC0085BDEC0084, 0xBDEC0087BDEC0086, 0xBDEC0089BDEC0088,
-    0xBDEC008BBDEC008A, 0xBDEC008DBDEC008C, 0xBDEC008FBDEC008E,
-    0xBDEC0091BDEC0090, 0xBDEC0093BDEC0092, 0xBDEC0095BDEC0094,
-    0xBDEC0097BDEC0096, 0xBDEC0099BDEC0098, 0xBDEC009BBDEC009A,
-    0xBDEC009DBDEC009C, 0xBDEC009FBDEC009E, 0xBDEC00A1BDEC00A0,
-    0xBDEC00A3BDEC00A2, 0xBDEC00A5BDEC00A4, 0xBDEC00A7BDEC00A6,
-    0xBDEC00A9BDEC00A8, 0xBDEC00ABBDEC00AA, 0xBDEC00ADBDEC00AC,
-    0xBDEC00AFBDEC00AE, 0xBDEC00B1BDEC00B0, 0xBDEC00B3BDEC00B2,
-    0xBDEC00B5BDEC00B4, 0xBDEC00B7BDEC00B6, 0xBDEC00B9BDEC00B8,
-    0xBDEC00BBBDEC00BA, 0xBDEC00BDBDEC00BC, 0xBEEC00BFBDEC00BE,
-    0xBEEC0081BEEC0080, 0xBEEC0083BEEC0082, 0xBEEC0085BEEC0084,
-    0xBEEC0087BEEC0086, 0xBEEC0089BEEC0088, 0xBEEC008BBEEC008A,
-    0xBEEC008DBEEC008C, 0xBEEC008FBEEC008E, 0xBEEC0091BEEC0090,
-    0xBEEC0093BEEC0092, 0xBEEC0095BEEC0094, 0xBEEC0097BEEC0096,
-    0xBEEC0099BEEC0098, 0xBEEC009BBEEC009A, 0xBEEC009DBEEC009C,
-    0xBEEC009FBEEC009E, 0xBEEC00A1BEEC00A0, 0xBEEC00A3BEEC00A2,
-    0xBEEC00A5BEEC00A4, 0xBEEC00A7BEEC00A6, 0xBEEC00A9BEEC00A8,
-    0xBEEC00ABBEEC00AA, 0xBEEC00ADBEEC00AC, 0xBEEC00AFBEEC00AE,
-    0xBEEC00B1BEEC00B0, 0xBEEC00B3BEEC00B2, 0xBEEC00B5BEEC00B4,
-    0xBEEC00B7BEEC00B6, 0xBEEC00B9BEEC00B8, 0xBEEC00BBBEEC00BA,
-    0xBEEC00BDBEEC00BC, 0xBFEC00BFBEEC00BE, 0xBFEC0081BFEC0080,
-    0xBFEC0083BFEC0082, 0xBFEC0085BFEC0084, 0xBFEC0087BFEC0086,
-    0xBFEC0089BFEC0088, 0xBFEC008BBFEC008A, 0xBFEC008DBFEC008C,
-    0xBFEC008FBFEC008E, 0xBFEC0091BFEC0090, 0xBFEC0093BFEC0092,
-    0xBFEC0095BFEC0094, 0xBFEC0097BFEC0096, 0xBFEC0099BFEC0098,
-    0xBFEC009BBFEC009A, 0xBFEC009DBFEC009C, 0xBFEC009FBFEC009E,
-    0xBFEC00A1BFEC00A0, 0xBFEC00A3BFEC00A2, 0xBFEC00A5BFEC00A4,
-    0xBFEC00A7BFEC00A6, 0xBFEC00A9BFEC00A8, 0xBFEC00ABBFEC00AA,
-    0xBFEC00ADBFEC00AC, 0xBFEC00AFBFEC00AE, 0xBFEC00B1BFEC00B0,
-    0xBFEC00B3BFEC00B2, 0xBFEC00B5BFEC00B4, 0xBFEC00B7BFEC00B6,
-    0xBFEC00B9BFEC00B8, 0xBFEC00BBBFEC00BA, 0xBFEC00BDBFEC00BC,
-    0x80ED00BFBFEC00BE, 0x80ED008180ED0080, 0x80ED008380ED0082,
-    0x80ED008580ED0084, 0x80ED008780ED0086, 0x80ED008980ED0088,
-    0x80ED008B80ED008A, 0x80ED008D80ED008C, 0x80ED008F80ED008E,
-    0x80ED009180ED0090, 0x80ED009380ED0092, 0x80ED009580ED0094,
-    0x80ED009780ED0096, 0x80ED009980ED0098, 0x80ED009B80ED009A,
-    0x80ED009D80ED009C, 0x80ED009F80ED009E, 0x80ED00A180ED00A0,
-    0x80ED00A380ED00A2, 0x80ED00A580ED00A4, 0x80ED00A780ED00A6,
-    0x80ED00A980ED00A8, 0x80ED00AB80ED00AA, 0x80ED00AD80ED00AC,
-    0x80ED00AF80ED00AE, 0x80ED00B180ED00B0, 0x80ED00B380ED00B2,
-    0x80ED00B580ED00B4, 0x80ED00B780ED00B6, 0x80ED00B980ED00B8,
-    0x80ED00BB80ED00BA, 0x80ED00BD80ED00BC, 0x81ED00BF80ED00BE,
-    0x81ED008181ED0080, 0x81ED008381ED0082, 0x81ED008581ED0084,
-    0x81ED008781ED0086, 0x81ED008981ED0088, 0x81ED008B81ED008A,
-    0x81ED008D81ED008C, 0x81ED008F81ED008E, 0x81ED009181ED0090,
-    0x81ED009381ED0092, 0x81ED009581ED0094, 0x81ED009781ED0096,
-    0x81ED009981ED0098, 0x81ED009B81ED009A, 0x81ED009D81ED009C,
-    0x81ED009F81ED009E, 0x81ED00A181ED00A0, 0x81ED00A381ED00A2,
-    0x81ED00A581ED00A4, 0x81ED00A781ED00A6, 0x81ED00A981ED00A8,
-    0x81ED00AB81ED00AA, 0x81ED00AD81ED00AC, 0x81ED00AF81ED00AE,
-    0x81ED00B181ED00B0, 0x81ED00B381ED00B2, 0x81ED00B581ED00B4,
-    0x81ED00B781ED00B6, 0x81ED00B981ED00B8, 0x81ED00BB81ED00BA,
-    0x81ED00BD81ED00BC, 0x82ED00BF81ED00BE, 0x82ED008182ED0080,
-    0x82ED008382ED0082, 0x82ED008582ED0084, 0x82ED008782ED0086,
-    0x82ED008982ED0088, 0x82ED008B82ED008A, 0x82ED008D82ED008C,
-    0x82ED008F82ED008E, 0x82ED009182ED0090, 0x82ED009382ED0092,
-    0x82ED009582ED0094, 0x82ED009782ED0096, 0x82ED009982ED0098,
-    0x82ED009B82ED009A, 0x82ED009D82ED009C, 0x82ED009F82ED009E,
-    0x82ED00A182ED00A0, 0x82ED00A382ED00A2, 0x82ED00A582ED00A4,
-    0x82ED00A782ED00A6, 0x82ED00A982ED00A8, 0x82ED00AB82ED00AA,
-    0x82ED00AD82ED00AC, 0x82ED00AF82ED00AE, 0x82ED00B182ED00B0,
-    0x82ED00B382ED00B2, 0x82ED00B582ED00B4, 0x82ED00B782ED00B6,
-    0x82ED00B982ED00B8, 0x82ED00BB82ED00BA, 0x82ED00BD82ED00BC,
-    0x83ED00BF82ED00BE, 0x83ED008183ED0080, 0x83ED008383ED0082,
-    0x83ED008583ED0084, 0x83ED008783ED0086, 0x83ED008983ED0088,
-    0x83ED008B83ED008A, 0x83ED008D83ED008C, 0x83ED008F83ED008E,
-    0x83ED009183ED0090, 0x83ED009383ED0092, 0x83ED009583ED0094,
-    0x83ED009783ED0096, 0x83ED009983ED0098, 0x83ED009B83ED009A,
-    0x83ED009D83ED009C, 0x83ED009F83ED009E, 0x83ED00A183ED00A0,
-    0x83ED00A383ED00A2, 0x83ED00A583ED00A4, 0x83ED00A783ED00A6,
-    0x83ED00A983ED00A8, 0x83ED00AB83ED00AA, 0x83ED00AD83ED00AC,
-    0x83ED00AF83ED00AE, 0x83ED00B183ED00B0, 0x83ED00B383ED00B2,
-    0x83ED00B583ED00B4, 0x83ED00B783ED00B6, 0x83ED00B983ED00B8,
-    0x83ED00BB83ED00BA, 0x83ED00BD83ED00BC, 0x84ED00BF83ED00BE,
-    0x84ED008184ED0080, 0x84ED008384ED0082, 0x84ED008584ED0084,
-    0x84ED008784ED0086, 0x84ED008984ED0088, 0x84ED008B84ED008A,
-    0x84ED008D84ED008C, 0x84ED008F84ED008E, 0x84ED009184ED0090,
-    0x84ED009384ED0092, 0x84ED009584ED0094, 0x84ED009784ED0096,
-    0x84ED009984ED0098, 0x84ED009B84ED009A, 0x84ED009D84ED009C,
-    0x84ED009F84ED009E, 0x84ED00A184ED00A0, 0x84ED00A384ED00A2,
-    0x84ED00A584ED00A4, 0x84ED00A784ED00A6, 0x84ED00A984ED00A8,
-    0x84ED00AB84ED00AA, 0x84ED00AD84ED00AC, 0x84ED00AF84ED00AE,
-    0x84ED00B184ED00B0, 0x84ED00B384ED00B2, 0x84ED00B584ED00B4,
-    0x84ED00B784ED00B6, 0x84ED00B984ED00B8, 0x84ED00BB84ED00BA,
-    0x84ED00BD84ED00BC, 0x85ED00BF84ED00BE, 0x85ED008185ED0080,
-    0x85ED008385ED0082, 0x85ED008585ED0084, 0x85ED008785ED0086,
-    0x85ED008985ED0088, 0x85ED008B85ED008A, 0x85ED008D85ED008C,
-    0x85ED008F85ED008E, 0x85ED009185ED0090, 0x85ED009385ED0092,
-    0x85ED009585ED0094, 0x85ED009785ED0096, 0x85ED009985ED0098,
-    0x85ED009B85ED009A, 0x85ED009D85ED009C, 0x85ED009F85ED009E,
-    0x85ED00A185ED00A0, 0x85ED00A385ED00A2, 0x85ED00A585ED00A4,
-    0x85ED00A785ED00A6, 0x85ED00A985ED00A8, 0x85ED00AB85ED00AA,
-    0x85ED00AD85ED00AC, 0x85ED00AF85ED00AE, 0x85ED00B185ED00B0,
-    0x85ED00B385ED00B2, 0x85ED00B585ED00B4, 0x85ED00B785ED00B6,
-    0x85ED00B985ED00B8, 0x85ED00BB85ED00BA, 0x85ED00BD85ED00BC,
-    0x86ED00BF85ED00BE, 0x86ED008186ED0080, 0x86ED008386ED0082,
-    0x86ED008586ED0084, 0x86ED008786ED0086, 0x86ED008986ED0088,
-    0x86ED008B86ED008A, 0x86ED008D86ED008C, 0x86ED008F86ED008E,
-    0x86ED009186ED0090, 0x86ED009386ED0092, 0x86ED009586ED0094,
-    0x86ED009786ED0096, 0x86ED009986ED0098, 0x86ED009B86ED009A,
-    0x86ED009D86ED009C, 0x86ED009F86ED009E, 0x86ED00A186ED00A0,
-    0x86ED00A386ED00A2, 0x86ED00A586ED00A4, 0x86ED00A786ED00A6,
-    0x86ED00A986ED00A8, 0x86ED00AB86ED00AA, 0x86ED00AD86ED00AC,
-    0x86ED00AF86ED00AE, 0x86ED00B186ED00B0, 0x86ED00B386ED00B2,
-    0x86ED00B586ED00B4, 0x86ED00B786ED00B6, 0x86ED00B986ED00B8,
-    0x86ED00BB86ED00BA, 0x86ED00BD86ED00BC, 0x87ED00BF86ED00BE,
-    0x87ED008187ED0080, 0x87ED008387ED0082, 0x87ED008587ED0084,
-    0x87ED008787ED0086, 0x87ED008987ED0088, 0x87ED008B87ED008A,
-    0x87ED008D87ED008C, 0x87ED008F87ED008E, 0x87ED009187ED0090,
-    0x87ED009387ED0092, 0x87ED009587ED0094, 0x87ED009787ED0096,
-    0x87ED009987ED0098, 0x87ED009B87ED009A, 0x87ED009D87ED009C,
-    0x87ED009F87ED009E, 0x87ED00A187ED00A0, 0x87ED00A387ED00A2,
-    0x87ED00A587ED00A4, 0x87ED00A787ED00A6, 0x87ED00A987ED00A8,
-    0x87ED00AB87ED00AA, 0x87ED00AD87ED00AC, 0x87ED00AF87ED00AE,
-    0x87ED00B187ED00B0, 0x87ED00B387ED00B2, 0x87ED00B587ED00B4,
-    0x87ED00B787ED00B6, 0x87ED00B987ED00B8, 0x87ED00BB87ED00BA,
-    0x87ED00BD87ED00BC, 0x88ED00BF87ED00BE, 0x88ED008188ED0080,
-    0x88ED008388ED0082, 0x88ED008588ED0084, 0x88ED008788ED0086,
-    0x88ED008988ED0088, 0x88ED008B88ED008A, 0x88ED008D88ED008C,
-    0x88ED008F88ED008E, 0x88ED009188ED0090, 0x88ED009388ED0092,
-    0x88ED009588ED0094, 0x88ED009788ED0096, 0x88ED009988ED0098,
-    0x88ED009B88ED009A, 0x88ED009D88ED009C, 0x88ED009F88ED009E,
-    0x88ED00A188ED00A0, 0x88ED00A388ED00A2, 0x88ED00A588ED00A4,
-    0x88ED00A788ED00A6, 0x88ED00A988ED00A8, 0x88ED00AB88ED00AA,
-    0x88ED00AD88ED00AC, 0x88ED00AF88ED00AE, 0x88ED00B188ED00B0,
-    0x88ED00B388ED00B2, 0x88ED00B588ED00B4, 0x88ED00B788ED00B6,
-    0x88ED00B988ED00B8, 0x88ED00BB88ED00BA, 0x88ED00BD88ED00BC,
-    0x89ED00BF88ED00BE, 0x89ED008189ED0080, 0x89ED008389ED0082,
-    0x89ED008589ED0084, 0x89ED008789ED0086, 0x89ED008989ED0088,
-    0x89ED008B89ED008A, 0x89ED008D89ED008C, 0x89ED008F89ED008E,
-    0x89ED009189ED0090, 0x89ED009389ED0092, 0x89ED009589ED0094,
-    0x89ED009789ED0096, 0x89ED009989ED0098, 0x89ED009B89ED009A,
-    0x89ED009D89ED009C, 0x89ED009F89ED009E, 0x89ED00A189ED00A0,
-    0x89ED00A389ED00A2, 0x89ED00A589ED00A4, 0x89ED00A789ED00A6,
-    0x89ED00A989ED00A8, 0x89ED00AB89ED00AA, 0x89ED00AD89ED00AC,
-    0x89ED00AF89ED00AE, 0x89ED00B189ED00B0, 0x89ED00B389ED00B2,
-    0x89ED00B589ED00B4, 0x89ED00B789ED00B6, 0x89ED00B989ED00B8,
-    0x89ED00BB89ED00BA, 0x89ED00BD89ED00BC, 0x8AED00BF89ED00BE,
-    0x8AED00818AED0080, 0x8AED00838AED0082, 0x8AED00858AED0084,
-    0x8AED00878AED0086, 0x8AED00898AED0088, 0x8AED008B8AED008A,
-    0x8AED008D8AED008C, 0x8AED008F8AED008E, 0x8AED00918AED0090,
-    0x8AED00938AED0092, 0x8AED00958AED0094, 0x8AED00978AED0096,
-    0x8AED00998AED0098, 0x8AED009B8AED009A, 0x8AED009D8AED009C,
-    0x8AED009F8AED009E, 0x8AED00A18AED00A0, 0x8AED00A38AED00A2,
-    0x8AED00A58AED00A4, 0x8AED00A78AED00A6, 0x8AED00A98AED00A8,
-    0x8AED00AB8AED00AA, 0x8AED00AD8AED00AC, 0x8AED00AF8AED00AE,
-    0x8AED00B18AED00B0, 0x8AED00B38AED00B2, 0x8AED00B58AED00B4,
-    0x8AED00B78AED00B6, 0x8AED00B98AED00B8, 0x8AED00BB8AED00BA,
-    0x8AED00BD8AED00BC, 0x8BED00BF8AED00BE, 0x8BED00818BED0080,
-    0x8BED00838BED0082, 0x8BED00858BED0084, 0x8BED00878BED0086,
-    0x8BED00898BED0088, 0x8BED008B8BED008A, 0x8BED008D8BED008C,
-    0x8BED008F8BED008E, 0x8BED00918BED0090, 0x8BED00938BED0092,
-    0x8BED00958BED0094, 0x8BED00978BED0096, 0x8BED00998BED0098,
-    0x8BED009B8BED009A, 0x8BED009D8BED009C, 0x8BED009F8BED009E,
-    0x8BED00A18BED00A0, 0x8BED00A38BED00A2, 0x8BED00A58BED00A4,
-    0x8BED00A78BED00A6, 0x8BED00A98BED00A8, 0x8BED00AB8BED00AA,
-    0x8BED00AD8BED00AC, 0x8BED00AF8BED00AE, 0x8BED00B18BED00B0,
-    0x8BED00B38BED00B2, 0x8BED00B58BED00B4, 0x8BED00B78BED00B6,
-    0x8BED00B98BED00B8, 0x8BED00BB8BED00BA, 0x8BED00BD8BED00BC,
-    0x8CED00BF8BED00BE, 0x8CED00818CED0080, 0x8CED00838CED0082,
-    0x8CED00858CED0084, 0x8CED00878CED0086, 0x8CED00898CED0088,
-    0x8CED008B8CED008A, 0x8CED008D8CED008C, 0x8CED008F8CED008E,
-    0x8CED00918CED0090, 0x8CED00938CED0092, 0x8CED00958CED0094,
-    0x8CED00978CED0096, 0x8CED00998CED0098, 0x8CED009B8CED009A,
-    0x8CED009D8CED009C, 0x8CED009F8CED009E, 0x8CED00A18CED00A0,
-    0x8CED00A38CED00A2, 0x8CED00A58CED00A4, 0x8CED00A78CED00A6,
-    0x8CED00A98CED00A8, 0x8CED00AB8CED00AA, 0x8CED00AD8CED00AC,
-    0x8CED00AF8CED00AE, 0x8CED00B18CED00B0, 0x8CED00B38CED00B2,
-    0x8CED00B58CED00B4, 0x8CED00B78CED00B6, 0x8CED00B98CED00B8,
-    0x8CED00BB8CED00BA, 0x8CED00BD8CED00BC, 0x8DED00BF8CED00BE,
-    0x8DED00818DED0080, 0x8DED00838DED0082, 0x8DED00858DED0084,
-    0x8DED00878DED0086, 0x8DED00898DED0088, 0x8DED008B8DED008A,
-    0x8DED008D8DED008C, 0x8DED008F8DED008E, 0x8DED00918DED0090,
-    0x8DED00938DED0092, 0x8DED00958DED0094, 0x8DED00978DED0096,
-    0x8DED00998DED0098, 0x8DED009B8DED009A, 0x8DED009D8DED009C,
-    0x8DED009F8DED009E, 0x8DED00A18DED00A0, 0x8DED00A38DED00A2,
-    0x8DED00A58DED00A4, 0x8DED00A78DED00A6, 0x8DED00A98DED00A8,
-    0x8DED00AB8DED00AA, 0x8DED00AD8DED00AC, 0x8DED00AF8DED00AE,
-    0x8DED00B18DED00B0, 0x8DED00B38DED00B2, 0x8DED00B58DED00B4,
-    0x8DED00B78DED00B6, 0x8DED00B98DED00B8, 0x8DED00BB8DED00BA,
-    0x8DED00BD8DED00BC, 0x8EED00BF8DED00BE, 0x8EED00818EED0080,
-    0x8EED00838EED0082, 0x8EED00858EED0084, 0x8EED00878EED0086,
-    0x8EED00898EED0088, 0x8EED008B8EED008A, 0x8EED008D8EED008C,
-    0x8EED008F8EED008E, 0x8EED00918EED0090, 0x8EED00938EED0092,
-    0x8EED00958EED0094, 0x8EED00978EED0096, 0x8EED00998EED0098,
-    0x8EED009B8EED009A, 0x8EED009D8EED009C, 0x8EED009F8EED009E,
-    0x8EED00A18EED00A0, 0x8EED00A38EED00A2, 0x8EED00A58EED00A4,
-    0x8EED00A78EED00A6, 0x8EED00A98EED00A8, 0x8EED00AB8EED00AA,
-    0x8EED00AD8EED00AC, 0x8EED00AF8EED00AE, 0x8EED00B18EED00B0,
-    0x8EED00B38EED00B2, 0x8EED00B58EED00B4, 0x8EED00B78EED00B6,
-    0x8EED00B98EED00B8, 0x8EED00BB8EED00BA, 0x8EED00BD8EED00BC,
-    0x8FED00BF8EED00BE, 0x8FED00818FED0080, 0x8FED00838FED0082,
-    0x8FED00858FED0084, 0x8FED00878FED0086, 0x8FED00898FED0088,
-    0x8FED008B8FED008A, 0x8FED008D8FED008C, 0x8FED008F8FED008E,
-    0x8FED00918FED0090, 0x8FED00938FED0092, 0x8FED00958FED0094,
-    0x8FED00978FED0096, 0x8FED00998FED0098, 0x8FED009B8FED009A,
-    0x8FED009D8FED009C, 0x8FED009F8FED009E, 0x8FED00A18FED00A0,
-    0x8FED00A38FED00A2, 0x8FED00A58FED00A4, 0x8FED00A78FED00A6,
-    0x8FED00A98FED00A8, 0x8FED00AB8FED00AA, 0x8FED00AD8FED00AC,
-    0x8FED00AF8FED00AE, 0x8FED00B18FED00B0, 0x8FED00B38FED00B2,
-    0x8FED00B58FED00B4, 0x8FED00B78FED00B6, 0x8FED00B98FED00B8,
-    0x8FED00BB8FED00BA, 0x8FED00BD8FED00BC, 0x90ED00BF8FED00BE,
-    0x90ED008190ED0080, 0x90ED008390ED0082, 0x90ED008590ED0084,
-    0x90ED008790ED0086, 0x90ED008990ED0088, 0x90ED008B90ED008A,
-    0x90ED008D90ED008C, 0x90ED008F90ED008E, 0x90ED009190ED0090,
-    0x90ED009390ED0092, 0x90ED009590ED0094, 0x90ED009790ED0096,
-    0x90ED009990ED0098, 0x90ED009B90ED009A, 0x90ED009D90ED009C,
-    0x90ED009F90ED009E, 0x90ED00A190ED00A0, 0x90ED00A390ED00A2,
-    0x90ED00A590ED00A4, 0x90ED00A790ED00A6, 0x90ED00A990ED00A8,
-    0x90ED00AB90ED00AA, 0x90ED00AD90ED00AC, 0x90ED00AF90ED00AE,
-    0x90ED00B190ED00B0, 0x90ED00B390ED00B2, 0x90ED00B590ED00B4,
-    0x90ED00B790ED00B6, 0x90ED00B990ED00B8, 0x90ED00BB90ED00BA,
-    0x90ED00BD90ED00BC, 0x91ED00BF90ED00BE, 0x91ED008191ED0080,
-    0x91ED008391ED0082, 0x91ED008591ED0084, 0x91ED008791ED0086,
-    0x91ED008991ED0088, 0x91ED008B91ED008A, 0x91ED008D91ED008C,
-    0x91ED008F91ED008E, 0x91ED009191ED0090, 0x91ED009391ED0092,
-    0x91ED009591ED0094, 0x91ED009791ED0096, 0x91ED009991ED0098,
-    0x91ED009B91ED009A, 0x91ED009D91ED009C, 0x91ED009F91ED009E,
-    0x91ED00A191ED00A0, 0x91ED00A391ED00A2, 0x91ED00A591ED00A4,
-    0x91ED00A791ED00A6, 0x91ED00A991ED00A8, 0x91ED00AB91ED00AA,
-    0x91ED00AD91ED00AC, 0x91ED00AF91ED00AE, 0x91ED00B191ED00B0,
-    0x91ED00B391ED00B2, 0x91ED00B591ED00B4, 0x91ED00B791ED00B6,
-    0x91ED00B991ED00B8, 0x91ED00BB91ED00BA, 0x91ED00BD91ED00BC,
-    0x92ED00BF91ED00BE, 0x92ED008192ED0080, 0x92ED008392ED0082,
-    0x92ED008592ED0084, 0x92ED008792ED0086, 0x92ED008992ED0088,
-    0x92ED008B92ED008A, 0x92ED008D92ED008C, 0x92ED008F92ED008E,
-    0x92ED009192ED0090, 0x92ED009392ED0092, 0x92ED009592ED0094,
-    0x92ED009792ED0096, 0x92ED009992ED0098, 0x92ED009B92ED009A,
-    0x92ED009D92ED009C, 0x92ED009F92ED009E, 0x92ED00A192ED00A0,
-    0x92ED00A392ED00A2, 0x92ED00A592ED00A4, 0x92ED00A792ED00A6,
-    0x92ED00A992ED00A8, 0x92ED00AB92ED00AA, 0x92ED00AD92ED00AC,
-    0x92ED00AF92ED00AE, 0x92ED00B192ED00B0, 0x92ED00B392ED00B2,
-    0x92ED00B592ED00B4, 0x92ED00B792ED00B6, 0x92ED00B992ED00B8,
-    0x92ED00BB92ED00BA, 0x92ED00BD92ED00BC, 0x93ED00BF92ED00BE,
-    0x93ED008193ED0080, 0x93ED008393ED0082, 0x93ED008593ED0084,
-    0x93ED008793ED0086, 0x93ED008993ED0088, 0x93ED008B93ED008A,
-    0x93ED008D93ED008C, 0x93ED008F93ED008E, 0x93ED009193ED0090,
-    0x93ED009393ED0092, 0x93ED009593ED0094, 0x93ED009793ED0096,
-    0x93ED009993ED0098, 0x93ED009B93ED009A, 0x93ED009D93ED009C,
-    0x93ED009F93ED009E, 0x93ED00A193ED00A0, 0x93ED00A393ED00A2,
-    0x93ED00A593ED00A4, 0x93ED00A793ED00A6, 0x93ED00A993ED00A8,
-    0x93ED00AB93ED00AA, 0x93ED00AD93ED00AC, 0x93ED00AF93ED00AE,
-    0x93ED00B193ED00B0, 0x93ED00B393ED00B2, 0x93ED00B593ED00B4,
-    0x93ED00B793ED00B6, 0x93ED00B993ED00B8, 0x93ED00BB93ED00BA,
-    0x93ED00BD93ED00BC, 0x94ED00BF93ED00BE, 0x94ED008194ED0080,
-    0x94ED008394ED0082, 0x94ED008594ED0084, 0x94ED008794ED0086,
-    0x94ED008994ED0088, 0x94ED008B94ED008A, 0x94ED008D94ED008C,
-    0x94ED008F94ED008E, 0x94ED009194ED0090, 0x94ED009394ED0092,
-    0x94ED009594ED0094, 0x94ED009794ED0096, 0x94ED009994ED0098,
-    0x94ED009B94ED009A, 0x94ED009D94ED009C, 0x94ED009F94ED009E,
-    0x94ED00A194ED00A0, 0x94ED00A394ED00A2, 0x94ED00A594ED00A4,
-    0x94ED00A794ED00A6, 0x94ED00A994ED00A8, 0x94ED00AB94ED00AA,
-    0x94ED00AD94ED00AC, 0x94ED00AF94ED00AE, 0x94ED00B194ED00B0,
-    0x94ED00B394ED00B2, 0x94ED00B594ED00B4, 0x94ED00B794ED00B6,
-    0x94ED00B994ED00B8, 0x94ED00BB94ED00BA, 0x94ED00BD94ED00BC,
-    0x95ED00BF94ED00BE, 0x95ED008195ED0080, 0x95ED008395ED0082,
-    0x95ED008595ED0084, 0x95ED008795ED0086, 0x95ED008995ED0088,
-    0x95ED008B95ED008A, 0x95ED008D95ED008C, 0x95ED008F95ED008E,
-    0x95ED009195ED0090, 0x95ED009395ED0092, 0x95ED009595ED0094,
-    0x95ED009795ED0096, 0x95ED009995ED0098, 0x95ED009B95ED009A,
-    0x95ED009D95ED009C, 0x95ED009F95ED009E, 0x95ED00A195ED00A0,
-    0x95ED00A395ED00A2, 0x95ED00A595ED00A4, 0x95ED00A795ED00A6,
-    0x95ED00A995ED00A8, 0x95ED00AB95ED00AA, 0x95ED00AD95ED00AC,
-    0x95ED00AF95ED00AE, 0x95ED00B195ED00B0, 0x95ED00B395ED00B2,
-    0x95ED00B595ED00B4, 0x95ED00B795ED00B6, 0x95ED00B995ED00B8,
-    0x95ED00BB95ED00BA, 0x95ED00BD95ED00BC, 0x96ED00BF95ED00BE,
-    0x96ED008196ED0080, 0x96ED008396ED0082, 0x96ED008596ED0084,
-    0x96ED008796ED0086, 0x96ED008996ED0088, 0x96ED008B96ED008A,
-    0x96ED008D96ED008C, 0x96ED008F96ED008E, 0x96ED009196ED0090,
-    0x96ED009396ED0092, 0x96ED009596ED0094, 0x96ED009796ED0096,
-    0x96ED009996ED0098, 0x96ED009B96ED009A, 0x96ED009D96ED009C,
-    0x96ED009F96ED009E, 0x96ED00A196ED00A0, 0x96ED00A396ED00A2,
-    0x96ED00A596ED00A4, 0x96ED00A796ED00A6, 0x96ED00A996ED00A8,
-    0x96ED00AB96ED00AA, 0x96ED00AD96ED00AC, 0x96ED00AF96ED00AE,
-    0x96ED00B196ED00B0, 0x96ED00B396ED00B2, 0x96ED00B596ED00B4,
-    0x96ED00B796ED00B6, 0x96ED00B996ED00B8, 0x96ED00BB96ED00BA,
-    0x96ED00BD96ED00BC, 0x97ED00BF96ED00BE, 0x97ED008197ED0080,
-    0x97ED008397ED0082, 0x97ED008597ED0084, 0x97ED008797ED0086,
-    0x97ED008997ED0088, 0x97ED008B97ED008A, 0x97ED008D97ED008C,
-    0x97ED008F97ED008E, 0x97ED009197ED0090, 0x97ED009397ED0092,
-    0x97ED009597ED0094, 0x97ED009797ED0096, 0x97ED009997ED0098,
-    0x97ED009B97ED009A, 0x97ED009D97ED009C, 0x97ED009F97ED009E,
-    0x97ED00A197ED00A0, 0x97ED00A397ED00A2, 0x97ED00A597ED00A4,
-    0x97ED00A797ED00A6, 0x97ED00A997ED00A8, 0x97ED00AB97ED00AA,
-    0x97ED00AD97ED00AC, 0x97ED00AF97ED00AE, 0x97ED00B197ED00B0,
-    0x97ED00B397ED00B2, 0x97ED00B597ED00B4, 0x97ED00B797ED00B6,
-    0x97ED00B997ED00B8, 0x97ED00BB97ED00BA, 0x97ED00BD97ED00BC,
-    0x98ED00BF97ED00BE, 0x98ED008198ED0080, 0x98ED008398ED0082,
-    0x98ED008598ED0084, 0x98ED008798ED0086, 0x98ED008998ED0088,
-    0x98ED008B98ED008A, 0x98ED008D98ED008C, 0x98ED008F98ED008E,
-    0x98ED009198ED0090, 0x98ED009398ED0092, 0x98ED009598ED0094,
-    0x98ED009798ED0096, 0x98ED009998ED0098, 0x98ED009B98ED009A,
-    0x98ED009D98ED009C, 0x98ED009F98ED009E, 0x98ED00A198ED00A0,
-    0x98ED00A398ED00A2, 0x98ED00A598ED00A4, 0x98ED00A798ED00A6,
-    0x98ED00A998ED00A8, 0x98ED00AB98ED00AA, 0x98ED00AD98ED00AC,
-    0x98ED00AF98ED00AE, 0x98ED00B198ED00B0, 0x98ED00B398ED00B2,
-    0x98ED00B598ED00B4, 0x98ED00B798ED00B6, 0x98ED00B998ED00B8,
-    0x98ED00BB98ED00BA, 0x98ED00BD98ED00BC, 0x99ED00BF98ED00BE,
-    0x99ED008199ED0080, 0x99ED008399ED0082, 0x99ED008599ED0084,
-    0x99ED008799ED0086, 0x99ED008999ED0088, 0x99ED008B99ED008A,
-    0x99ED008D99ED008C, 0x99ED008F99ED008E, 0x99ED009199ED0090,
-    0x99ED009399ED0092, 0x99ED009599ED0094, 0x99ED009799ED0096,
-    0x99ED009999ED0098, 0x99ED009B99ED009A, 0x99ED009D99ED009C,
-    0x99ED009F99ED009E, 0x99ED00A199ED00A0, 0x99ED00A399ED00A2,
-    0x99ED00A599ED00A4, 0x99ED00A799ED00A6, 0x99ED00A999ED00A8,
-    0x99ED00AB99ED00AA, 0x99ED00AD99ED00AC, 0x99ED00AF99ED00AE,
-    0x99ED00B199ED00B0, 0x99ED00B399ED00B2, 0x99ED00B599ED00B4,
-    0x99ED00B799ED00B6, 0x99ED00B999ED00B8, 0x99ED00BB99ED00BA,
-    0x99ED00BD99ED00BC, 0x9AED00BF99ED00BE, 0x9AED00819AED0080,
-    0x9AED00839AED0082, 0x9AED00859AED0084, 0x9AED00879AED0086,
-    0x9AED00899AED0088, 0x9AED008B9AED008A, 0x9AED008D9AED008C,
-    0x9AED008F9AED008E, 0x9AED00919AED0090, 0x9AED00939AED0092,
-    0x9AED00959AED0094, 0x9AED00979AED0096, 0x9AED00999AED0098,
-    0x9AED009B9AED009A, 0x9AED009D9AED009C, 0x9AED009F9AED009E,
-    0x9AED00A19AED00A0, 0x9AED00A39AED00A2, 0x9AED00A59AED00A4,
-    0x9AED00A79AED00A6, 0x9AED00A99AED00A8, 0x9AED00AB9AED00AA,
-    0x9AED00AD9AED00AC, 0x9AED00AF9AED00AE, 0x9AED00B19AED00B0,
-    0x9AED00B39AED00B2, 0x9AED00B59AED00B4, 0x9AED00B79AED00B6,
-    0x9AED00B99AED00B8, 0x9AED00BB9AED00BA, 0x9AED00BD9AED00BC,
-    0x9BED00BF9AED00BE, 0x9BED00819BED0080, 0x9BED00839BED0082,
-    0x9BED00859BED0084, 0x9BED00879BED0086, 0x9BED00899BED0088,
-    0x9BED008B9BED008A, 0x9BED008D9BED008C, 0x9BED008F9BED008E,
-    0x9BED00919BED0090, 0x9BED00939BED0092, 0x9BED00959BED0094,
-    0x9BED00979BED0096, 0x9BED00999BED0098, 0x9BED009B9BED009A,
-    0x9BED009D9BED009C, 0x9BED009F9BED009E, 0x9BED00A19BED00A0,
-    0x9BED00A39BED00A2, 0x9BED00A59BED00A4, 0x9BED00A79BED00A6,
-    0x9BED00A99BED00A8, 0x9BED00AB9BED00AA, 0x9BED00AD9BED00AC,
-    0x9BED00AF9BED00AE, 0x9BED00B19BED00B0, 0x9BED00B39BED00B2,
-    0x9BED00B59BED00B4, 0x9BED00B79BED00B6, 0x9BED00B99BED00B8,
-    0x9BED00BB9BED00BA, 0x9BED00BD9BED00BC, 0x9CED00BF9BED00BE,
-    0x9CED00819CED0080, 0x9CED00839CED0082, 0x9CED00859CED0084,
-    0x9CED00879CED0086, 0x9CED00899CED0088, 0x9CED008B9CED008A,
-    0x9CED008D9CED008C, 0x9CED008F9CED008E, 0x9CED00919CED0090,
-    0x9CED00939CED0092, 0x9CED00959CED0094, 0x9CED00979CED0096,
-    0x9CED00999CED0098, 0x9CED009B9CED009A, 0x9CED009D9CED009C,
-    0x9CED009F9CED009E, 0x9CED00A19CED00A0, 0x9CED00A39CED00A2,
-    0x9CED00A59CED00A4, 0x9CED00A79CED00A6, 0x9CED00A99CED00A8,
-    0x9CED00AB9CED00AA, 0x9CED00AD9CED00AC, 0x9CED00AF9CED00AE,
-    0x9CED00B19CED00B0, 0x9CED00B39CED00B2, 0x9CED00B59CED00B4,
-    0x9CED00B79CED00B6, 0x9CED00B99CED00B8, 0x9CED00BB9CED00BA,
-    0x9CED00BD9CED00BC, 0x9DED00BF9CED00BE, 0x9DED00819DED0080,
-    0x9DED00839DED0082, 0x9DED00859DED0084, 0x9DED00879DED0086,
-    0x9DED00899DED0088, 0x9DED008B9DED008A, 0x9DED008D9DED008C,
-    0x9DED008F9DED008E, 0x9DED00919DED0090, 0x9DED00939DED0092,
-    0x9DED00959DED0094, 0x9DED00979DED0096, 0x9DED00999DED0098,
-    0x9DED009B9DED009A, 0x9DED009D9DED009C, 0x9DED009F9DED009E,
-    0x9DED00A19DED00A0, 0x9DED00A39DED00A2, 0x9DED00A59DED00A4,
-    0x9DED00A79DED00A6, 0x9DED00A99DED00A8, 0x9DED00AB9DED00AA,
-    0x9DED00AD9DED00AC, 0x9DED00AF9DED00AE, 0x9DED00B19DED00B0,
-    0x9DED00B39DED00B2, 0x9DED00B59DED00B4, 0x9DED00B79DED00B6,
-    0x9DED00B99DED00B8, 0x9DED00BB9DED00BA, 0x9DED00BD9DED00BC,
-    0x9EED00BF9DED00BE, 0x9EED00819EED0080, 0x9EED00839EED0082,
-    0x9EED00859EED0084, 0x9EED00879EED0086, 0x9EED00899EED0088,
-    0x9EED008B9EED008A, 0x9EED008D9EED008C, 0x9EED008F9EED008E,
-    0x9EED00919EED0090, 0x9EED00939EED0092, 0x9EED00959EED0094,
-    0x9EED00979EED0096, 0x9EED00999EED0098, 0x9EED009B9EED009A,
-    0x9EED009D9EED009C, 0x9EED009F9EED009E, 0x9EED00A19EED00A0,
-    0x91F000A39EED00A2, 0x009C8291F0009A82, 0x8491F000AB8291F0,
-    0xF000AF8491F000AE, 0x8C8D91F0008B8D91, 0x91F000BB9291F000,
-    0x00BE9291F000BC92, 0x9691F000BA9691F0, 0xF000B8A491F000BB,
-    0x00A5859DF097859D, 0xA5859DF098859DF0, 0x859DF098859DF000,
-    0x9DF000AE859DF0A5, 0x9DF0A5859DF09885, 0xF098859DF000AF85,
-    0x00B0859DF0A5859D, 0xA5859DF098859DF0, 0x859DF000B1859DF0,
-    0x859DF0A5859DF098, 0x9DF0B9869DF000B2, 0xF0B9869DF000A585,
-    0x00AE859DF0A5859D, 0xA5859DF0B9869DF0, 0x869DF000AF859DF0,
-    0x9DF000A5859DF0BA, 0x9DF0A5859DF0BA86, 0xF0BA869DF000AE85,
-    0x00AF859DF0A5859D, 0xBC9DF00084BC9DF0, 0xF00086BC9DF00085,
-    0x8ABC9DF00088BC9D, 0xA0F0009EBC9DF000, 0x009C94A0F000A284,
-    0x95A0F000A594A0F0, 0xF000BA98A0F0008B, 0x9EA3A0F00084A0A0,
-    0xA0F000ACA8A0F000, 0x00A493A1F000A3AD, 0x9BA1F000A89AA1F0,
-    0xF00088A7A1F000AA, 0x8BB4A1F00098ACA1, 0xA1F000A4B7A1F000,
-    0x008386A2F000A6B7, 0x8CA2F0009F86A2F0, 0xF000949BA2F000B1,
-    0x8AA1A2F00084A1A2, 0xA2F0008CACA2F000, 0x008A80A3F000B1AF,
-    0x8DA3F000B88AA3F0, 0xF000938EA3F0009F, 0x838FA3F0009C8EA3,
-    0xA3F000958FA3F000, 0x00A39AA3F000AD91, 0xAAA3F000A7A2A3F0,
-    0xF000BAABA3F0008D, 0x9EB4A3F000BCB2A3, 0xA3F00091BBA3F000,
-    0x008EBEA3F0009EBD, 0x8BA4F000A389A4F0, 0xF000AB8EA4F000AE,
-    0xB59CA4F0008898A4, 0xA4F00094A0A4F000, 0x0092B2A4F000B6B0,
-    0xBEA4F000A1BEA4F0, 0xF0008481A5F000B8, 0xB383A5F000B283A5,
-    0xA5F0009984A5F000, 0x008989A5F000B384, 0x98A5F0009D90A5F0,
-    0xF0009A9AA5F000A6, 0xBCA5A5F000859BA5, 0xA5F000A7AAA5F000,
-    0x0080B2A5F000ABAE, 0xBEA5F00090B3A5F0, 0xF0009A87A6F00086,
-    0x8789A6F000A888A6, 0xA6F000998BA6F000, 0x009A93A6F000BE8C,
-    0x96A6F000A394A6F0, 0xF000A79EA6F000A8, 0xBCACA6F000B59EA6,
-    0xA6F000B6B0A6F000, 0x00ABB5A6F00095B3, 0xBEA6F000ACBCA6F0,
-    0xF0009283A7F000B1, 0xA799A7F0008A8FA7, 0xA7F000AEA2A7F000,
-    0x00A8B2A7F000A6A5, 0xBCA7F00093BBA7F0, 0xF0009297A8F000AF,
-    0xAE9CA8F000AD97A8, 0xA8F000BAAFA8F000, 0x008585A9F000B7B5,
-    0x88A9F0009F87A9F0, 0xF0008A90A9F0009A, 0xB696A9F0009692A9,
-    0xAAF000B0ACA9F000, 0x008584AAF0008E83, 0x8AAAF0008E88AAF0,
-    0xF000928EAAF00091, 0x00000000008098AA, 0x000084000002B800,
-    0x8000000180000000, 0x000005B80002FCCC, 0x0002E4CC80002381,
-    0x8000237500003DB8, 0x000005B80002ECCC, 0x00003CCC80002385,
-    0x8000099E0000BCCD, 0x800009A4800009A1, 0x0000198000001D83,
-    0x80001E3700001981, 0x80001E4780001E2F, 0x0000318080001E3F,
-    0x800009A700003181, 0x0003583D00003189, 0x80000A490003483C,
-    0x000085010003683E, 0x00008D0300010184, 0x0000950500008902,
-    0x00009D0700009104, 0x0000A90900009906, 0x0000AD0B0000A108,
-    0x0000B90D0000A50A, 0x0000BD0F0000BD0C, 0x0000C5110000B90E,
-    0x0000CD130000C110, 0x0000D5150000C912, 0x0000DD170000D114,
-    0x0000E5190000D916, 0x0000ED1B0000E118, 0x0000F51D0000E91A,
-    0x0000FD1F0000F11C, 0x000938610000F91E, 0x000E406380001E5F,
-    0x000F28650008F062, 0x000FD867000E8064, 0x000C7069000FF066,
-    0x000CF06B000C3068, 0x000D486D000CD86A, 0x0012106F000D1C6C,
-    0x80000A4F000D6C6E, 0x0013A87300130C70, 0x001028750013EC72,
-    0x0011D07700104C74, 0x0016D07900101476, 0x80001E570010F478,
-    0x80001E6F0016F47A, 0x00017D7F80001E67, 0x0002804180000C98,
-    0x0000184380000BF0, 0x000048450000C042, 0x0006904700006044,
-    0x0006284900018446, 0x0007384B00016C48, 0x0007704D0007F04A,
-    0x0004F04F0007444C, 0x800009AA0004CC4E, 0x000AA05300052C50,
-    0x000A005500051052, 0x0008D057000A5054, 0x00084859000B1056,
-    0x000011840008B858, 0x0000F58000083C5A, 0x80000BEA00009D81,
-    0x0002698380001E27, 0x0002358100024582, 0x00036D8700023580,
-    0x0000398100035586, 0x80000C2900026584, 0x0003D5890004018A,
-    0x0000098F0003ED88, 0x80000C3B80000BBA, 0x80000C410000098C,
-    0x0000099100033189, 0x80001E4F0000D8CC, 0x80001BB780001E77,
-    0x0003658080000A55, 0x00003D8600030D81, 0x000210CC00031D83,
-    0x00002982000184CD, 0x000681A3000374CC, 0x0002B18080001BBF,
-    0x80001BC70002B181, 0x0000C1A500002981, 0x0003E8CC80001BCF,
-    0x80000A5B000260CD, 0x80000A610000F1A8, 0x80000A6D80000A67,
-    0x0000698180001BD7, 0x000111820004DCCC, 0x80000A7380001BDF,
-    0x80001BF780001BE7, 0x80001BEF80001BFF, 0x0004A0CC800009B6,
-    0x000108CC0004DCCD, 0x80001E9F800009B9, 0x002B00E100011186,
-    0x024888E300277CE0, 0x80001E9702380CE2, 0x80001EAF00019187,
-    0x0001D98080001EA7, 0x0002C4CC00003581, 0x80001E8F026444EA,
-    0x0267F8EF80001C07, 0x0001F18200071181, 0x00005589028D0CF0,
-    0x80001C0F0001E581, 0x00006D800001E187, 0x0000798300000D81,
-    0x0000798180000A7F, 0x000138CC00019580, 0x0001C18C000114CD,
-    0x800009BF80000A85, 0x001794C3000115A3, 0x001784C500149CC2,
-    0x0017ECC7001750C4, 0x80001E87001280C6, 0x0016D0CB80000A91,
-    0x001614CD0016E0CA, 0x0020BCCF80000C47, 0x0024CCD100199CCE,
-    0x0024E8D30025D8D0, 0x80000C4D000151B1, 0x80001EB70007DDA7,
-    0x0027C8D90024BCD6, 0x0027F8DB0022D4D8, 0x0007ED8780001C27,
-    0x0004F8CC80000A8B, 0x80001E1F80001C17, 0x800009B30006C987,
-    0x800009BC0007DD91, 0x80000A7980001C1F, 0x80001E7F80001C2F,
-    0x0006F98C80000C9E, 0x0006D58080000C1D, 0x0002C4CC0006C181,
-    0x00005182000504CD, 0x0006AD8000077583, 0x000735860006B981,
-    0x0000418400072987, 0x0000EDA780000A97, 0x0007E98880001C37,
-    0x80000A9D0007CD89, 0x0007DD8C0007C18F, 0x80000BFC80000AA3,
-    0x00009DB000065DA7, 0x80000AAF80000AA9, 0x80001C3F000641A3,
-    0x80000C8F80001C4F, 0x80001C5780001C47, 0x80001C5F000671AD,
-    0x0003F8CC800009C2, 0x800009C500032CCD, 0x800009C80000E5A3,
-    0x80000ABB00004D8C, 0x80000AC10000F1A7, 0x00009D8700063DB1,
-    0x00071DA800009586, 0x80000AC700008984, 0x0000D5810000A982,
-    0x80000AD3000701AD, 0x0001998200003D91, 0x00050CCD800009CB,
-    0x00003D81000508CC, 0x0001BD8800018587, 0x80001EBF80001C6F,
-    0x0001918C80000BC0, 0x80000C5980000C53, 0x0001F58080001EC7,
-    0x0001818200018581, 0x0001898400019D83, 0x0001A98680000ACD,
-    0x00006D8800005D87, 0x80001C6700004D89, 0x00005D8C000210CC,
-    0x80000AD900000D82, 0x0000E9B00000498F, 0x80000C0280001C77,
-    0x00012580000129A3, 0x80001C7F00015581, 0x0001A4CC000135A7,
-    0x80000ADF000210CD, 0x0000358180001C87, 0x000109AE80000AE5,
-    0x0001E8CC80000AF1, 0x80001C8F000228CD, 0x000005840000F1A3,
-    0x80000AEB80001C97, 0x0000B5A880001CA7, 0x00005D8180001C9F,
-    0x0001FCCC80001CAF, 0x80001CB70001C0CD, 0x0003858180001CBF,
-    0x0003E58180000C23, 0x0003E98780000AF7, 0x0003A58C800009CE,
-    0x80000B0380001CC7, 0x80001CCF800009AD, 0x0003DD8100039981,
-    0x000124CD0003BD8C, 0x000210CC000120CC, 0x0000F981000230CD,
-    0x0000558000005187, 0x80000AFD00005981, 0x0000418C00004983,
-    0x80001CD780001CDF, 0x00007D81800009D1, 0x000320CC00007180,
-    0x00037DA30003B0CD, 0x000311A3800009D4, 0x00012D8180001EE7,
-    0x000331A70000FDB1, 0x80001EDF000341AD, 0x80001EF700036DA3,
-    0x80001CE780001EEF, 0x0000358000036DA7, 0x00036DB100001D81,
-    0x80000CAA00002D88, 0x00007181000095A7, 0x0000718300005980,
-    0x000051810000F5A3, 0x80001CEF000081AD, 0x00005D89000309B1,
-    0x000138CC00003D84, 0x0000198100019D84, 0x0000658400000980,
-    0x0007D0CC80001CF7, 0x80001CFF0007F0CC, 0x00064CCC00008991,
-    0x000670CC000604CD, 0x000740CC000674CD, 0x0007A8CC000F2D9B,
-    0x00000D860007ACCD, 0x000F4184000F5587, 0x000C018280000B12,
-    0x00013580000F5583, 0x80000CB60001D981, 0x0000ED8C0000E98F,
-    0x80001ED780000CA4, 0x000F6D880000F58B, 0x80000B1800000D89,
-    0x0000258280000BC6, 0x0000418180000C5F, 0x80001F0F80000C65,
-    0x80001F1F80001F07, 0x80001F2780001F17, 0x0000298080001EFF,
-    0x80000C0E00002981, 0x0000798480001D07, 0x00014D8180001D0F,
-    0x80000B210001A180, 0x000238CC000FFDA3, 0x80001D17000348CD,
-    0x00027CCD00005981, 0x80000B2D000DA4CC, 0x000FC9A880000C6B,
-    0x800009DA00003987, 0x80000B0C000FA189, 0x80000CB000001191,
-    0x80000BA2800009DD, 0x000FB98080000C71, 0x80001ECF000FB981,
-    0x80000C08000F8D83, 0x0000058480001D1F, 0x80000B2780001D27,
-    0x80000B3380001D2F, 0x00002187000F9D81, 0x0002A8CC000FB587,
-    0x000CB4CC0002E8CD, 0x0000CDA6000CDCCD, 0x80001D470000C5A7,
-    0x000F898C000F818F, 0x80000B390000E5A3, 0x80000B4580001D37,
-    0x000005870000C9B1, 0x80001D3F80001D4F, 0x80001D5700003D87,
-    0x80000B3F80000C83, 0x0000F981000F05A3, 0x80001D5F0003A8CC,
-    0x0000498C80000B51, 0x80001D670000F1A3, 0x80000B4B80000C89,
-    0x80001D77000095A7, 0x0003ACCC80001D6F, 0x800009E00000C8CD,
-    0x800009E300006987, 0x000079820000BD81, 0x0001B987800009E6,
-    0x0000A59180000B57, 0x80001D9700001D81, 0x0000058880000B5D,
-    0x0000919B80001D9F, 0x800009E980000B63, 0x80000BE40001918C,
-    0x80000BD280000BD8, 0x0000458380000BDE, 0x0001B9810001B982,
-    0x80001F370001A580, 0x80000B6900007186, 0x0000158B00004984,
-    0x0000398900003D8A, 0x0000F98F00004588, 0x80000BCC80000B6F,
-    0x000151A70000058C, 0x80000C77000149A6, 0x000159A3000079B0,
-    0x0000A9800000A98C, 0x000095810000A981, 0x000171AD00009580,
-    0x80000BAB80000C7D, 0x80001F4700008D84, 0x000021A380001F3F,
-    0x80001F4F80001F57, 0x00006D8180001F5F, 0x000105B100006D80,
-    0x80001F2F000039A4, 0x80000B7580001D7F, 0x80001D8F000005A8,
-    0x000009AD80001D87, 0x0003BD810001B0CC, 0x0003CD830003BD80,
-    0x80001DAF80001DA7, 0x0001BCCC80001DB7, 0x0003E989000184CD,
-    0x0002D98100011CCC, 0x000138CD0002C580, 0x000268CC00013CCC,
-    0x00004D83000244CD, 0x0000D98780001DBF, 0x0000458000002D87,
-    0x0000058200001181, 0x80001DCF80000B7B, 0x80001DD780001DC7,
-    0x00001588000238CC, 0x00006D800000E188, 0x000284CC00003981,
-    0x000345A3000150CD, 0x80001DDF00013984, 0x0003DCCC80001DE7,
-    0x80001F67000248CD, 0x80000B81800009EC, 0x80000CBC80001F7F,
-    0x80001DEF00006D81, 0x0000D1A380000B87, 0x80001F6F80001F77,
-    0x00007581000488CC, 0x80000B8A00007580, 0x000400CC0000B5A3,
-    0x80001DF7000254CD, 0x00014D8480000B90, 0x80001DFF80000B96,
-    0x000ED0CC80001E07, 0x000500CC00012CCC, 0x800009EF00055CCD,
-    0x000E98CC800009F2, 0x000FCCCC0000C181, 0x000108CC000E3CCC,
-    0x000FD4CC000E9CCD, 0x000E3CCC000F78CD, 0x0001E581000E38CD,
-    0x0001E5830001E580, 0x800009D70001E582, 0x0001E1870001E584,
-    0x0001A58980001E3B, 0x0007C9810001D988, 0x80001E4B80001E33,
-    0x000134CC80001E43, 0x0000DD80800009F8, 0x80000A4C0000E981,
-    0x0000CD800000FD83, 0x000711810000F981, 0x80001E5B80001E63,
-    0x000128CC0000D189, 0x80001E6B80001E73, 0x000134CC80000BF3,
-    0x00005D8080000BED, 0x000101A300006581, 0x0000718100006183,
-    0x80001E2B00004980, 0x80000BBD80000C2C, 0x80000C3E00005589,
-    0x80001E5380000C44, 0x80001BBB80001E7B, 0x00001D8280000A58,
-    0x0007198000077983, 0x0000218600071981, 0x0007918400002987,
-    0x0000698A800009F5, 0x000069880006D9B1, 0x80000A520007F589,
-    0x0007D98C0007DD8F, 0x800009FB80000C9B, 0x800009FE80001E23,
-    0x000741800007A191, 0x0000058700074181, 0x0007918280001BC3,
-    0x80001BD380001BCB, 0x00078586000695A3, 0x000290CC0000EDA7,
-    0x0000E9B10005E0CD, 0x80000A5E0000C1A3, 0x80000A6A80000A64,
-    0x80000A0480000A70, 0x80001BDB000765A5, 0x000755A80003B8CC,
-    0x80000A7680001BE3, 0x80001BFB80001BEB, 0x80001BF380001C03,
-    0x0000ADA380000A07, 0x0006258C80000A0A, 0x00004D8C80000A0D,
-    0x00006D8280001EA3, 0x0006018200063D81, 0x80001E9B00006581,
-    0x80001EB300006587, 0x0001AD8700061187, 0x000250CC80001EAB,
-    0x0001E1810002C4CD, 0x0001E1830001C980, 0x0001FD8000008991,
-    0x80001E930001D581, 0x0000358900019D8C, 0x0000758080000A7C,
-    0x00002D8600005D81, 0x00007D8400003587, 0x0001858280001C0B,
-    0x00019D8000007D83, 0x80001C1300019181, 0x0000E18C0000E98F,
-    0x80000A8880000A82, 0x00000D8880000A10, 0x000139A70000F989,
-    0x80000A9480001E8B, 0x000125A380000C4A, 0x000059B080000C50,
-    0x0000058280001E83, 0x000115AD80001EBB, 0x80001C2B80000CA1,
-    0x80000A8E00000586, 0x80001C2380001C1B, 0x00010D80000069A7,
-    0x0000318700010181, 0x000119B1000049A3, 0x80001C33000210CC,
-    0x80000A9A80000C20, 0x0001D4CC000051AD, 0x000041A80001F8CD,
-    0x000138CD80001C3B, 0x0003ED8100013CCC, 0x80000AA680000AA0,
-    0x80000BFF00000D8C, 0x0003C58180000AAC, 0x80000AB2000268CC,
-    0x80001C430003C582, 0x000029870003C984, 0x80001C5300003186,
-    0x80000C9200006D82, 0x0000618780001C4B, 0x80001C6380001C5B,
-    0x0003A8CC80001E0F, 0x0000F9810002ACCD, 0x80000A1300002D88,
-    0x80000A1980000A16, 0x80000ABE0000358C, 0x0000D5B180000AC4,
-    0x80001C7380000ACA, 0x80000BC380001EC3, 0x80000C56000130CC,
-    0x80001ECB80000C5C, 0x80000AD0000104CC, 0x0000A5A780001C6B,
-    0x00008DA380000ADC, 0x0000B980000214CC, 0x000085A70000B981,
-    0x0000658280000C1A, 0x0000358C80001C7B, 0x001CECCC80000C05,
-    0x80001C830003FCCD, 0x80000AE20000ADAE, 0x0006C0CC80001C8B,
-    0x0001BDB00006D0CD, 0x80000AE8000790CC, 0x001D18CC0001C981,
-    0x00001184000220CD, 0x0006B8CD80000AF4, 0x80001C9B0006ACCC,
-    0x0000298180000AEE, 0x80001CA380001CAB, 0x001D40CC0001EDA3,
-    0x80001CB3000210CD, 0x0000098180001CBB, 0x001D40CC80001CC3,
-    0x0001D9A80002D0CD, 0x000790CC80000C26, 0x000660CC00011CCC,
-    0x80000AFA000664CD, 0x80001CCB80000A22, 0x80000B0600012991,
-    0x80001CD3001FCD81, 0x80001CE380000B00, 0x0006C4CC80001CDB,
-    0x000031810006C0CD, 0x80000A250000D980, 0x0001B98280000A28,
-    0x0001B98000014583, 0x00015D860001B981, 0x00015D8480001EEB,
-    0x0000C0CD80001EE3, 0x001C01880000F0CC, 0x80001EFB00017989,
-    0x0001618C0001558F, 0x001EDD8180001EF3, 0x80001C93001F1581,
-    0x001F818C001F5D81, 0x000318CC001F4187, 0x001F4D80000368CD,
-    0x80000A2B001FA581, 0x001F158C001FA983, 0x80000A2E001ED18C,
-    0x001F9580001FB587, 0x00001181001F9981, 0x001F8580001FBD83,
-    0x80001CEB001F8981, 0x80000CAD00003D84, 0x80001CF3001F9989,
-    0x80000B0F00000588, 0x00000D8080001CFB, 0x80001D030000FD81,
-    0x001E79A3001FD9A3, 0x80000B15000085A3, 0x001E61A7001F35AD,
-    0x80000CB3001FA9A7, 0x80000CB900000D84, 0x80000A31001F35A3,
-    0x00007184001F89AD, 0x001E3DB1001F21A7, 0x80000CA7001F85B1,
-    0x80000B1B80001EDB, 0x001CB4CC80000BC9, 0x80000C62001CB0CD,
-    0x80000C68001F41B1, 0x80001F1380000BA5, 0x80001F2380001F0B,
-    0x001EB98080001F1B, 0x00012182001EB981, 0x0001618400013583,
-    0x0001898680001F2B, 0x000191880001BD87, 0x80001ED30001E989,
-    0x0001E58C0001FD8B, 0x80001F0300000D82, 0x0001B9830001DD8F,
-    0x0001BD810001AD91, 0x0001A1800001BD80, 0x80000C0B0001A181,
-    0x00003D8480000C11, 0x0001958980001D0B, 0x001ECD810001819B,
-    0x0002FCCC001ECD80, 0x0000E5810002F0CD, 0x0000ED8180001D13,
-    0x000190CC0000A5A3, 0x00001D8700025CCD, 0x0000F9A880000B24,
-    0x80000B3080001D1B, 0x80000C7480000C6E, 0x0000058480001D23,
-    0x0000D5A380001D2B, 0x80001D3380000B2A, 0x0001F4CC00005181,
-    0x80000B360001D0CD, 0x80001D4B00003D87, 0x80001D3B80000B3C,
-    0x0000058780000B48, 0x80001D4380001D53, 0x0003D9810003F187,
-    0x0000918100039981, 0x80000C8680001D5B, 0x000124CC0003F98C,
-    0x0003F9810003B987, 0x000264CC80000B42, 0x0003998C0003C982,
-    0x0003D18780001D63, 0x0006DD810003918F, 0x000400CC0003F591,
-    0x000065870003ECCD, 0x80000B5480001E13, 0x80000C8C80001D6B,
-    0x000118CC80000B4E, 0x80001D7B00003988, 0x00035CCC80001D73,
-    0x80000A370000258C, 0x0000FDB100012CCC, 0x80000A3A000339A3,
-    0x80001D9B80000A3D, 0x000374CC000359A7, 0x000351A300000988,
-    0x80000B6680001DA3, 0x00009DA7000D6D83, 0x00009DA30000ADA6,
-    0x000128CC00030DB1, 0x0000B5A780000BE7, 0x80000BDB0000B5A6,
-    0x0007D581000130CC, 0x000128CC80000BD5, 0x80001F3B80000BE1,
-    0x000089AD80000B6C, 0x0001CD8080000B72, 0x000144CC0001DD81,
-    0x0001F584001E9CCD, 0x000158CC80000BCF, 0x0001ED8000014CCD,
-    0x0000B5830001FD81, 0x0001DD8C80000C7A, 0x001E08CC80000C80,
-    0x000138CC0006F18C, 0x80001F4B00013CCD, 0x0006058100001189,
-    0x80001F5B80001F43, 0x80001F6380001F53, 0x0000218000003D83,
-    0x0000398000003981, 0x80001F3300002181, 0x000128CC80001D83,
-    0x80000B780006D981, 0x000124CC0006FD81, 0x80001D8B80001D93,
-    0x0001218080001DAB, 0x000005A300013581, 0x80001DBB80001DB3,
-    0x00058CCC80001DC3, 0x00031CB7000240CD, 0x0007CD8780000B7E,
-    0x80001DCB80001DD3, 0x000088CD80001E17, 0x0005DD8C000090CC,
-    0x80000A1C000085A3, 0x80001DDB0002E0CC, 0x00004D810007D588,
-    0x80000B5A00004D80, 0x80000B6080001DE3, 0x80000A4000073991,
-    0x80000BAE0007C588, 0x80001DEB0007C98A, 0x0007E58700003D9B,
-    0x0007858180001F6B, 0x80000A4300078580, 0x0006D5820007F982,
-    0x0006C98000004583, 0x0006E1860006DD81, 0x0000518480000B84,
-    0x0006958A80001F83, 0x000079880006998B, 0x80000CBF00069D89,
-    0x00076D8C0007598F, 0x000380CC80001DF3, 0x000711B0000224CD,
-    0x80001F7B80000A46, 0x80001F7380001E1B, 0x00008D8080000B8D,
-    0x0003ECCC00009D81, 0x80001DFB0004E8CD, 0x000749A380000B93,
-    0x80001E0380000B99, 0x80001E0B0007BDA3, 0x800007EA8000001B,
-    0x0007A5A48000000B, 0x8000058A800004EB, 0x000761A880000003,
-    0x8000006580000E60, 0x8000091E800003C8, 0x00012D80000769AD,
-    0x00015D8200013D81, 0x0001618400015183, 0x800004AC800004B8,
-    0x0001AD8800016187, 0x0001BD8A0001B589, 0x800005BC0015298F,
-    0x000278CC80000C2F, 0x0001DD810005D8CD, 0x80000BF600007181,
-    0x000019840001FD82, 0x0001ED8780000C35, 0x000434CC80000C32,
-    0x80000BF9000430CD, 0x80000C380015699F, 0x0004F8CD0001C98C,
-    0x000149B10002F0CC, 0x0000D58180000721, 0x001599A0000115A3,
-    0x8000074680000876, 0x000428CC0000F584, 0x800008C600042CCD,
-    0x00017DA8000179AF, 0x000169AA80000DB2, 0x000125B480000955,
-    0x000109A3000129B5, 0x800006EA80000C14, 0x000105B2800006EE,
-    0x0000F5BC00013DB3, 0x0000D9BE0000FDBD, 0x000101B88000083B,
-    0x000101BA000101B9, 0x800007438000073C, 0x00039DB300000C86,
-    0x800008B6000389B2, 0x8000076C0003BD80, 0x800009158000076F,
-    0x800006E7800006E4, 0x0003858980000838, 0x0003A1BF0007E4CC,
-    0x80000D0180000C17, 0x000CE4CD0000658C, 0x00004189000C24CC,
-    0x0000458B00005588, 0x0003B5850000458A, 0x000395870003B584,
-    0x80000D3A00038186, 0x80000D4C80000D40, 0x8000001300002C98,
-    0x00068CCD0000C8A6, 0x000069850006A8CC, 0x8000003180000017,
-    0x0003F0CC80000069, 0x8000003580000007, 0x000095B180000CF8,
-    0x000095B380000D79, 0x80000DBC000095B2, 0x80000DC780000DBF,
-    0x80000DAF80000DC2, 0x800006A680000075, 0x8000099B80000025,
-    0x000355BF80002217, 0x8000221380000DCA, 0x8000215B8000220F,
-    0x000D659400006CB8, 0x0001AD9C00020092, 0x0001899A0001A59D,
-    0x000199980001899B, 0x000620CC00018999, 0x000640CC000604CD,
-    0x00073CCC00065CCD, 0x0007BCCC000738CD, 0x000F018000007585,
-    0x80002163000F0D81, 0x00000585000199BA, 0x800021538000216B,
-    0x000F2981000185BE, 0x8000215F000F2580, 0x00002985000F2182,
-    0x000011850001A5B4, 0x000F89B380002167, 0x0001C9B1000F85B2,
-    0x000F81B7000C01B0, 0x0001EDB50001EDB6, 0x000165810001EDB4,
-    0x0001618300016580, 0x000B3CCC000FB5B8, 0x000FE18000017984,
-    0x8000216F000FDD81, 0x0004C8CC80002157, 0x0000AD850004B8CD,
-    0x8000221B00001585, 0x0019E8CC80002233, 0x00015DA40019B4CD,
-    0x00024DA280000DCD, 0x000155A0000FF1A3, 0x80000882000FDDA1,
-    0x8000093E80000971, 0x8000096F80001FE7, 0x8000097A80001FEF,
-    0x800008B480001FEB, 0x80001FA700000D93, 0x000CD58480001FF3,
-    0x00006994000EE985, 0x80001FAF000EDD87, 0x80001FB70002B499,
-    0x80001FAB80001FBF, 0x80001FB380002237, 0x80001FBB0001749F,
-    0x000E8D80000D1891, 0x000E8586000E8181, 0x000E898400038C95,
-    0x0000118200024497, 0x0000D1800000B8A9, 0x80001FC30000ED81,
-    0x80001FDF80000DD0, 0x00008983000F5985, 0x000E4581000D34A1,
-    0x80001FE3000E4980, 0x000F6185000134A5, 0x800020170014F0CC,
-    0x0003C8CC0005C4B9, 0x0000DD800003DCCD, 0x8000201F0000C181,
-    0x00001185000590BF, 0x0000F5800007ECB1, 0x8000219B0000C181,
-    0x8000202700068CB5, 0x000EDD80000640B7, 0x000228CC000EE981,
-    0x000ECD800002B4CD, 0x00002985000EF981, 0x0001758000018583,
-    0x0001F1800001B181, 0x800021A30001D581, 0x000005858000202F,
-    0x00019980800021AB, 0x00007D820001BD81, 0x0001E1940001D593,
-    0x8000219300000D85, 0x000009808000201B, 0x8000202300001981,
-    0x00002D940001AD93, 0x00000D858000219F, 0x000005858000202B,
-    0x80002033800021A7, 0x0000198200002985, 0x0000F18000003585,
-    0x800021AF0000E181, 0x8000223B80002197, 0x8000225B00000585,
-    0x0001ED8000017583, 0x0001F0CC00012981, 0x80000DD30001B0CD,
-    0x8000225380002257, 0x8000205780000E2A, 0x00003D808000205F,
-    0x8000206700000D81, 0x00002D808000206F, 0x0000158200001D81,
-    0x00012CCD8000205B, 0x8000206B80002063, 0x0003C58880002073,
-    0x000290CC8000229F, 0x0003ED840002BCCD, 0x0003F98680000DD6,
-    0x0003818080002097, 0x00003D800003FD81, 0x0000498100003581,
-    0x0001DCCD00004180, 0x8000209F00004182, 0x800020A38000209B,
-    0x00039D94000130CD, 0x0001F0CD80002287, 0x0003B98100010CCC,
-    0x0003A1830003C580, 0x0000AD800003A593, 0x00009D800000D581,
-    0x800022830000A981, 0x0000C18180000DD9, 0x0000B5800000C980,
-    0x8000227F00008181, 0x000370CC00046983, 0x8000227B0003B4CD,
-    0x80000E2D00013CCD, 0x00028CCD00040193, 0x00000D800002F4CC,
-    0x0006D99400003581, 0x800020CF800020CB, 0x000108CC800020D3,
-    0x0000118100011CCD, 0x800022A300002980, 0x0001918000001982,
-    0x000568CD00019181, 0x000910CC000294CC, 0x000E58CC001628CD,
-    0x000D14CC000E78CD, 0x000C28CC000AD8CD, 0x0017E0CC000D8CCD,
-    0x000B38CC000AE0CC, 0x000011850008F4CD, 0x0007518080000DDC,
-    0x800021DB00075181, 0x00000585000E90CD, 0x000ED8CC800021E3,
-    0x800021EB0001ECCD, 0x00001985000E00CD, 0x000E3CCD000FF0CD,
-    0x000168CC000E54CD, 0x800021D300016CCD, 0x0007999480000864,
-    0x00016180003BC0CC, 0x000029850000D981, 0x000178CC000E64CD,
-    0x0007AD8100017CCD, 0x000065850007AD80, 0x0001518000079D82,
-    0x0007E19400015181, 0x000179840000C0CD, 0x000161860007C593,
-    0x00012188800021DF, 0x800021E700000D85, 0x0000D8CC000118CD,
-    0x0007DD800000B4CD, 0x800020930007CD81, 0x800020C700000985,
-    0x800020F7800021EF, 0x800020FF800021D7, 0x8000210F80002107,
-    0x800020FB80002103, 0x8000210B800022A7, 0x0002B8CC80002113,
-    0x0000658000031CCD, 0x00007D8200007581, 0x0000098580002117,
-    0x800021FB00004185, 0x0006998080000E30, 0x0000198500072181,
-    0x8000220300007D83, 0x800021F700005985, 0x0000CD80800021F3,
-    0x80001F870000D981, 0x0000E98400001185, 0x0000E58680001F8F,
-    0x000035808000213B, 0x80001F9700000D81, 0x8000214300001D85,
-    0x0000058580001F9F, 0x000055808000214B, 0x00007D8200006D81,
-    0x00000D9400004593, 0x8000213300000D85, 0x0000098080001F8B,
-    0x80001F9300001181, 0x8000213F00001D85, 0x0000058580001F9B,
-    0x0001E98080002147, 0x000005820001F181, 0x0000098580001FA3,
-    0x8000214F00000985, 0x00016D8080002137, 0x0000198200017981,
-    0x800022070001A583, 0x0000058500007585, 0x800021FF8000220B,
-    0x0002C8CC8000211B, 0x80000E330002DCCD, 0x0000398080001FC7,
-    0x80001FCF00003981, 0x80001FCB80001FD7, 0x00001D8080001FD3,
-    0x80001FDB00000D81, 0x0000C5808000211F, 0x0000C0CD0000D581,
-    0x0000D580000224CC, 0x000039850000C581, 0x8000221F0000C983,
-    0x0000D58080000E36, 0x0000058500000981, 0x80001FF780002227,
-    0x000009808000217B, 0x80001FFF00001181, 0x0000B99400009993,
-    0x0000658580002007, 0x8000218300003D85, 0x8000200F00005D93,
-    0x000570CD00001185, 0x003C0194000574CC, 0x800021738000218B,
-    0x8000217F000128CD, 0x0000198580002187, 0x800021778000218F,
-    0x800022238000222F, 0x0000258580002123, 0x0002F8CC00003985,
-    0x00013CCD0003C4CD, 0x0003958180000E39, 0x0001C4CD00038D80,
-    0x000051850003A582, 0x000128CC00004985, 0x000138CD00012CCD,
-    0x003E698000013CCC, 0x80002243003E6981, 0x003C38CD8000223F,
-    0x003CD181003C54CC, 0x000148CC003E9D93, 0x000134CD0002C8CD,
-    0x003EB99480002247, 0x003EA18080000DDF, 0x003EA182003EA181,
-    0x8000224F003C8D88, 0x00002D800000AD85, 0x003FD58500002181,
-    0x0000ED808000208F, 0x0000A5860000F181, 0x0000A9848000203F,
-    0x000128CD000124CD, 0x003EF98880002047, 0x8000204F003FF185,
-    0x003FC58580002043, 0x000110CD000130CD, 0x0001DD800006F4CD,
-    0x0006BCCD0001E981, 0x0006E0CC000670CD, 0x000668CC00015CCD,
-    0x000664CD000690CD, 0x0007ACCC000678CD, 0x0007E8CC0007A8CD,
-    0x003F4D820007C8CD, 0x000235800002FD83, 0x8000204B0002E981,
-    0x003D058080002053, 0x8000224B00017D85, 0x000114CD80002127,
-    0x0001ED8180000EB0, 0x00013CCD0001E180, 0x00012CCD0001CD82,
-    0x00016D8100017582, 0x0001658000015980, 0x8000207F0000C581,
-    0x000138CD000124CD, 0x0000F9810000F182, 0x0000298000015980,
-    0x000118CC00007181, 0x00012CCD00017CCD, 0x00001D8180002087,
-    0x8000208300001D80, 0x8000226B8000208B, 0x003DE4CC000124CD,
-    0x0000E8CD000F5CCD, 0x0001D58000010CCC, 0x000099820001C981,
-    0x80002003003FE583, 0x800020138000200B, 0x80000EAA8000222B,
-    0x800020378000226F, 0x8000203B8000212B, 0x003F4D8080000EB3,
-    0x80002077003F4181, 0x8000207B80002263, 0x003F25808000225F,
-    0x003F9193003F2981, 0x80000EAD00007D83, 0x0002D8CC00004194,
-    0x800022670002ACCD, 0x0000098000005593, 0x80000E3C00001D81,
-    0x0000419480002277, 0x00009D80800020A7, 0x00001D800000A181,
-    0x0000A58400001181, 0x0000A586800020AF, 0x0000F188800020B7,
-    0x0000D9810000D182, 0x000025800000CD80, 0x0000058200002981,
-    0x800020AB800020BF, 0x800020BB800020B3, 0x000011940000B193,
-    0x00001D80800020C3, 0x0000258000001581, 0x00003D8200002D81,
-    0x0000ED8080002273, 0x000015820000D181, 0x003DD08100008183,
-    0x000220CC8000212F, 0x000278850001DCCD, 0x80000EB68000228B,
-    0x0003188980002293, 0x00002D80800020D7, 0x0000098000003581,
-    0x800020DF00006D81, 0x0000B191800021BB, 0x003BDD930000A990,
-    0x0000A59500040192, 0x800020E7003B9594, 0x000009850000A196,
-    0x800021C300004193, 0x00000585800020EF, 0x0000F594800021CB,
-    0x0000B980800021B3, 0x0000E5820000DD81, 0x800020E3800020DB,
-    0x000009800000E985, 0x0000058500003581, 0x800020EB800021BF,
-    0x800021C700001D85, 0x00000585800020F3, 0x0005C1B1800021CF,
-    0x800021B70003C5B0, 0x0005D9B50003C5B2, 0x800022970003CDB4,
-    0x0003FDB900003985, 0x8000228F8000229B, 0x80000E5480000E42,
-    0x80001FFB80000E9E, 0x80000E5A80000E8C, 0x80000E92000288CC,
-    0x000190CD003EE0CD, 0x80000DF7000124CC, 0x80000EC880000E15,
-    0x0000B18080000F6D, 0x003EE4CD003ED4CD, 0x003EBCCC80000F73,
-    0x0003D180003EB8CD, 0x0003FD820003ED81, 0x80000EC500006986,
-    0x00013CCD00039D85, 0x0002E8CC80000EBF, 0x0000E588003F3CCD,
-    0x80000F79000118CD, 0x80000EC200004188, 0x00012CCD80000F64,
-    0x0000D58080000F88, 0x0000E18680000F8E, 0x80000ECE0000E986,
-    0x80000F94000124CD, 0x0000D188000110CD, 0x0002F8CC00002988,
-    0x0002F8CD80000ED4, 0x80000F9A000124CC, 0x003EA0CC00032D85,
-    0x0003D580003E6CCD, 0x000301820003CD81, 0x000364CC0003F983,
-    0x80000ECB0002CCCD, 0x00020CCC000124CD, 0x00001181000208CD,
-    0x003EB0CC00039CCC, 0x003EC4CC003EB4CD, 0x003E80CC003EC0CD,
-    0x80000FA0003EB8CD, 0x003F98CC000284CC, 0x003E20CC003E50CD,
-    0x003F0CCD003EA0CD, 0x003F78CD003F30CD, 0x00005988003F04CD,
-    0x0001F8CD80000FB5, 0x80000ED10001C4CC, 0x80000FC180000FBB,
-    0x000108CC000214CC, 0x000005880001C8CD, 0x80000FCD80000FC7,
-    0x80000FAF000228CC, 0x0000D18600000D88, 0x0000CD8400011D80,
-    0x00023CCC0000F588, 0x000244CC00013584, 0x80000F7000012986,
-    0x80000F7600016588, 0x80000F37000544A7, 0x0000ED880000ED8B,
-    0x0000258800000D81, 0x000270CC00006586, 0x000230CC000274CD,
-    0x80000F310005D8CD, 0x0000398080000F7C, 0x0005E0CC80000F34,
-    0x000128CC80000F67, 0x00002586000114CD, 0x80000E8F00005588,
-    0x80000E4B00308486, 0x0007FD860007D580, 0x000224CC80000F8B,
-    0x80000F9100003D88, 0x00003D8880000F49, 0x000248CC80000F97,
-    0x80000EF2000278CD, 0x0005C09780000F9D, 0x0005189500057096,
-    0x0005589300006980, 0x0000658800003981, 0x00005986003CF890,
-    0x000045840005A89E, 0x0002D4CC80000F46, 0x00005D800005B49A,
-    0x0004ECA70005AC98, 0x0002D8CC00001188, 0x0004BCA30002DCCD,
-    0x000260CC80000FA3, 0x80000F4C80000FB8, 0x000408AD80000FBE,
-    0x0004F4AB80000FC4, 0x0000058800022CCC, 0x0003E8B780000FCA,
-    0x000428B500041CB6, 0x000428B3000200CC, 0x00007D880000798B,
-    0x00004186000474B0, 0x000075840003A4BE, 0x80000FD000001D88,
-    0x0003DCCC00034CBA, 0x0002A8830003FCB8, 0x80000FB200000D88,
-    0x0002948700001588, 0x80000F40000238CC, 0x0002DC8B000204CC,
-    0x80000F5200000D8F, 0x000220CC0000198F, 0x0002F08D80000F55,
-    0x0003209800007588, 0x80000F8200032C99, 0x00024CCC80000E1B,
-    0x000224CC00037096, 0x80000F8500003D88, 0x0000058800022CCC,
-    0x000224CC80000FA9, 0x80000FAC00000D88, 0x0003D4A80000F587,
-    0x80000FD30003C4A9, 0x800010C8000260D9, 0x800010D1800010CB,
-    0x0000319300003594, 0x800010CE000268D9, 0x0000059400024CD9,
-    0x00002195800010D4, 0x8000114600002994, 0x800018398000176A,
-    0x000260D9800017C8, 0x0003D8B58000182D, 0x000035940003F0B4,
-    0x80001850000258D9, 0x000254D900000594, 0x8000182700000D94,
-    0x0000C88A000174E0, 0x0000C888000015BC, 0x8000187D000280E0,
-    0x0002ECA40000A081, 0x00005CA80002C8A4, 0x8000188F00000DBC,
-    0x000005BC000284E0, 0x0002F8A480001893, 0x0003CCB00002F4E0,
-    0x800018530000D095, 0x0002C8A60003E0B3, 0x0000D892000260A7,
-    0x0000C1BE8000185A, 0x0000C19780001861, 0x8000186F80001868,
-    0x003F78CD80001876, 0x000D80BE003FBCCC, 0x000C78BC80001881,
-    0x000CF4BA000D80BD, 0x000CF4B8000C54BB, 0x800018AC80001888,
-    0x800018B0000CB8B7, 0x0003A0B2000C8CB5, 0x80001897000CA0B3,
-    0x000270AE000340B1, 0x000305B8000364AF, 0x8000189E000348AD,
-    0x000344A8800018A5, 0x800018D00003D8A9, 0x000268A40003D8A7,
-    0x000335B6000388A5, 0x0002C89200033DB7, 0x00015D9900033DB5,
-    0x0001659B00014D98, 0x0001659D0001699A, 0x0001859F0001659C,
-    0x00029C870001999E, 0x0000A5960000ADBE, 0x000534CC0000BD97,
-    0x0038B8CC00378CCD, 0x0001959F0038ACCD, 0x0001999D00001DB6,
-    0x0001B1B30001F59C, 0x800018D7000278AE, 0x00003D99800018B4,
-    0x0000199B800018BB, 0x800018C20000099A, 0x000248E0800018C9,
-    0x00025CAC0000099E, 0x800018F00002F4AD, 0x800018F4800018EC,
-    0x00036087800018DE, 0x000254AF800018E5, 0x00000597000011BE,
-    0x0002ACE0800018F8, 0x0003F886800018FC, 0x0002B4E00003EC87,
-    0x000248AF80001904, 0x00023CE0000290AE, 0x8000190000000DBE,
-    0x0002D8E000001997, 0x0000119600038486, 0x00006D9D000C00BF,
-    0x800019080000619C, 0x000250B300024CB1, 0x8000190C00000D95,
-    0x0002C8E000039C86, 0x00000582000204B3, 0x000250B380001918,
-    0x8000191C00000D95, 0x8000191480001910, 0x00038C860002B4E0,
-    0x00000DBE0003DC87, 0x0000799600006995, 0x0002E8B480001920,
-    0x00000597000250B5, 0x0002D4E080001928, 0x000019BE0002F4B4,
-    0x800019240002A0E0, 0x00038C990002D4E0, 0x0000058A0002D4B7,
-    0x800019308000192C, 0x000234B70000158A, 0x8000193400001D8F,
-    0x000005B380001938, 0x000005B38000193C, 0x0000099C80001951,
-    0x800019430000099D, 0x0000058C8000194A, 0x8000196380001958,
-    0x800019718000196A, 0x800019780000499F, 0x8000195C8000197F,
-    0x8000198D80001986, 0x800019C5800019CF, 0x800019D6800019E0,
-    0x0001C98380001994, 0x800019A20001F981, 0x800019B0800019A9,
-    0x800019BE800019B7, 0x0002B4808000199B, 0x000208E1000388A5,
-    0x800019E700001DAE, 0x0001F58D000009BC, 0x000200E1800019EB,
-    0x001094E300019592, 0x0001899700004593, 0x0001788700001485,
-    0x80003B6300040086, 0x80003B6B80003B67, 0x80003B7380003B6F,
-    0x00688CEF0000799D, 0x80003B770001A19C, 0x000099A280003B7B,
-    0x80003B8380003B7F, 0x80003B8B80003B87, 0x80003B8F000089A7,
-    0x000149A980003B93, 0x0000A9AC80003B97, 0x80003B9F80003B9B,
-    0x000125B380003BA3, 0x80003BAB80003BA7, 0x000121B780003BAF,
-    0x000131B5000139B6, 0x80003BB780003BB3, 0x000111B90000F9B9,
-    0x80003BBB000119B8, 0x80003BC380003BBF, 0x0001B08080003BC7,
-    0x80003BCB00028884, 0x000354840001C883, 0x80003BCF00045C84,
-    0x0003DD8000055C84, 0x0000318200002D81, 0x80003BD700023CBE,
-    0x0003888780003BDB, 0x00032DB300049C86, 0x80003BE380003BDF,
-    0x00051CBE000400BE, 0x0003D9B50003D1B6, 0x0003F9BB80003BE7,
-    0x01D3589A0003F1BA, 0x0003C5BF0007E084, 0x0003DDBD0003D9BE,
-    0x001800BE0003DDBC, 0x00030DA6000305A5, 0x00181C8480003BEB,
-    0x80003BF380003BEF, 0x000301AC0003F1A3, 0x000305AE000301AD,
-    0x021BA4AC000319AF, 0x00032DAA021ADCAD, 0x00194CBE000325AB,
-    0x80003BF7001A48BE, 0x021AE0B4001B80BE, 0x021E30B6021DD4B5,
-    0x80003BFB000208EF, 0x021A38BA0001A0E1, 0x000000BA0002F8E3,
-    0x021CF4BE021C64BD, 0x01D30491021000BF, 0x000144E001F6E490,
-    0x80003BFF023E0892, 0x80003C0780003C03, 0x001CACBE80003C0B,
-    0x80003C1380003C0F, 0x80003C1B80003C17, 0x80003C2380003C1F,
-    0x006DB08180003C27, 0x00521C8300018080, 0x009A20850077D882,
-    0x00CF088700BE9084, 0x01376C8900E68C86, 0x017E188B012CE088,
-    0x01A5E08D0113D88A, 0x01EBFC8F015AF08C, 0x0017D5B0018F588E,
-    0x006B8DB2006801B1, 0x006B49B4006B9DB3, 0x80003C2B006F95B5,
-    0x80003C2F000170E0, 0x80003C3780003C33, 0x0000198080003C3B,
-    0x0000058200001981, 0x80003C4780003C3F, 0x000749A20002A5A1,
-    0x000641A4000741A3, 0x001BDDA6001BD5A5, 0x001C01A8001AC1A7,
-    0x001F49AA001FE1A9, 0x0013F5AC001001AB, 0x001401AE0013ADAD,
-    0x000631A80017F9AF, 0x00064DAA000631A9, 0x00064DAC000645AB,
-    0x00068DAE000699AD, 0x00038086000685AF, 0x80003C4B00031887,
-    0x80003C5380003C4F, 0x80003C5B80003C57, 0x00076DB880003C5F,
-    0x00077DBA000775B9, 0x00076DBC000775BB, 0x0007F9BE0007CDBD,
-    0x0006A9B00007F9BF, 0x000779B2000685B1, 0x000769B4000779B3,
-    0x00075DB6000769B5, 0x80003BD3000755B7, 0x80003C630006C1B3,
-    0x00061DB680003C43, 0x80003C6B80003C67, 0x000669BA0006C1B5,
-    0x80003C6F000645BB, 0x0007A9BE80003C73, 0x0007A9BC0007A9BF,
-    0x000685A30007A9BD, 0x80003C7B80003C77, 0x80003C8380003C7F,
-    0x000685A5000659A6, 0x000601AB80003C87, 0x80003C8B00062DAA,
-    0x0007EDAF80003C8F, 0x0007EDAD0007EDAE, 0x0006C1A80007EDAC,
-    0x000125AA00012DA9, 0x000125AC00013DAB, 0x000125AE00012DAD,
-    0x80003C93000195AF, 0x80003C9B80003C97, 0x80003CA380003C9F,
-    0x0000BD810000BD82, 0x00019DB80000BD80, 0x0001A5BA0001ADB9,
-    0x000075BC00006DBB, 0x000075BE00007DBD, 0x0001F5B000006DBF,
-    0x0001C1B20001FDB1, 0x0001B5B40001C1B3, 0x0001B5B60001BDB5,
-    0x80003CA70001ADB7, 0x80003CAF80003CAB, 0x80003CB30001A1B3,
-    0x0001A1B60001A1B5, 0x0001808680003CB7, 0x00013DBA0001C487,
-    0x000119BC000135BB, 0x00017DBE000119BD, 0x0001EDA3000175BF,
-    0x80003CBF80003CBB, 0x80003CC780003CC3, 0x0001EDA50001EDA6,
-    0x000179AB80003CCB, 0x80003CCF000171AA, 0x000139AF80003CD3,
-    0x000155AD000131AE, 0x80003CD7000155AC, 0x80003CDF80003CDB,
-    0x80003CE780003CE3, 0x80003CEB0001ACE1, 0x80003CEF000660E3,
-    0x80003CF780003CF3, 0x80003CFB000654EF, 0x001AB4BE80003CFF,
-    0x80003D0780003D03, 0x001C688680003D0B, 0x80003D0F00038C87,
-    0x80003D1780003D13, 0x0000398080003D1B, 0x0000058200003981,
-    0x80003D2780003D1F, 0x80003D2F80003D2B, 0x00031DA880003D33,
-    0x000341AA000341A9, 0x000341AC000341AB, 0x000371AE000349AD,
-    0x001D2084000371AF, 0x80003D3B80003D37, 0x80003D4380003D3F,
-    0x80003D4B80003D47, 0x0000E9B880003D4F, 0x0000FDBA0000F5B9,
-    0x000091BC0000F1BB, 0x000089BE000091BD, 0x0000E1B0000089BF,
-    0x0000E1B20000E1B1, 0x0000E1B40000E1B3, 0x0000D9B60000EDB5,
-    0x00070CE30000D9B7, 0x000130E1000720E1, 0x80003D53000708E3,
-    0x80003D5B80003D57, 0x80003D6380003D5F, 0x80003D6B80003D67,
-    0x000798EF80003D6F, 0x80003D7780003D73, 0x000289B3000788EF,
-    0x80003D7F80003D7B, 0x80003D83001A80BE, 0x000289B5000289B6,
-    0x000165BB80003D87, 0x80003D8B000165BA, 0x000169BF80003D8F,
-    0x000175BD000165BE, 0x00063DB7000175BC, 0x00063DB500063DB6,
-    0x000635B3000621B4, 0x000601B1000635B2, 0x000659BF000609B0,
-    0x000659BD000651BE, 0x00066DBB00064DBC, 0x000671B9000679BA,
-    0x00009D80000679B8, 0x0000A5820000AD81, 0x80003D9B80003D97,
-    0x80003DA380003D9F, 0x000671AF80003DA7, 0x00066DAD000669AE,
-    0x000681AB00066DAC, 0x000691A9000699AA, 0x80003D23000699A8,
-    0x80003DAB80003D93, 0x80003DAF001DC5A3, 0x001DC5A6001DC5A5,
-    0x0003008680003DB3, 0x001E29AA00036487, 0x001E39AC001E29AB,
-    0x001E29AE001E39AD, 0x001E39B3001E25AF, 0x80003DBB80003DB7,
-    0x80003DC380003DBF, 0x001EFDB5001EFDB6, 0x001ED9BB80003DC7,
-    0x80003DCB001ED1BA, 0x001F69BF80003DCF, 0x001F69BD001F61BE,
-    0x001EF1A8001EC1BC, 0x001EF1AA001EF1A9, 0x001E35AC001EF1AB,
-    0x001E35AE001E3DAD, 0x80003DD3001E2DAF, 0x80003DDB80003DD7,
-    0x80003DE380003DDF, 0x80003DEB80003DE7, 0x001FE5B880003DEF,
-    0x001FE5BA001FEDB9, 0x001FE9BC001FF9BB, 0x001F91BE001F91BD,
-    0x001E55B0001F8DBF, 0x001E55B2001E5DB1, 0x001FE5B4001FFDB3,
-    0x001FE5B6001FEDB5, 0x001F79A3001FDDB7, 0x80003DF780003DF3,
-    0x80003DFF80003DFB, 0x001FBDA5001FBDA6, 0x001F99AB80003E03,
-    0x00008886001F91AA, 0x001E29AF00014C87, 0x001E29AD001E21AE,
-    0x00004980001F81AC, 0x0000598200004981, 0x80003E07001E99B3,
-    0x000151B6001E89B5, 0x80003E0F80003E0B, 0x00012DBA80003E13,
-    0x00013DBC000125BB, 0x000125BE000125BD, 0x001E4DA8000115BF,
-    0x001E5DAA001E55A9, 0x001E4DAC001E55AB, 0x000189AE00019DAD,
-    0x0001AC84000181AF, 0x80003E1B80003E17, 0x80003E2380003E1F,
-    0x80003E2B80003E27, 0x000165B880003E2F, 0x000165BA00016DB9,
-    0x000165BC00017DBB, 0x000165BE00016DBD, 0x0001A5B00000D9BF,
-    0x0001A5B20001ADB1, 0x0001ADB40001BDB3, 0x000195B600019DB5,
-    0x001DD5A300015DB7, 0x80003E3780003E33, 0x80003E3F80003E3B,
-    0x001DC5A500021DA6, 0x000269AB80003E43, 0x80003E47000261AA,
-    0x000259AF80003E4B, 0x000269AD000269AE, 0x80003E4F000271AC,
-    0x80003E5780003E53, 0x80003E5F80003E5B, 0x80003E6780003E63,
-    0x0000398080003E6B, 0x0000058200003981, 0x80003E7780003E6F,
-    0x0003408780003E7B, 0x00044C8400045C86, 0x80003E8380003E7F,
-    0x000120E300050484, 0x0001A8E180003E87, 0x001A94EF80003E8B,
-    0x80003E9380003E8F, 0x80003E9B80003E97, 0x80003EA380003E9F,
-    0x0003A1B380003EA7, 0x80003EAF80003EAB, 0x80003EB780003EB3,
-    0x00037DB500037DB6, 0x000359BB80003EBB, 0x80003EBF000351BA,
-    0x0000FDBF80003EC3, 0x0000FDBD0000FDBE, 0x000251A8000341BC,
-    0x000269AA000259A9, 0x0002B5AC000269AB, 0x0002B5AE0002BDAD,
-    0x0007A8840002ADAF, 0x80003ECB80003EC7, 0x0000A98280003ECF,
-    0x0000A98080003ED3, 0x000169B80000A981, 0x000109BA000169B9,
-    0x000119BC000109BB, 0x000109BE000119BD, 0x0002D5B0000109BF,
-    0x0002D5B20002DDB1, 0x000179B4000169B3, 0x000169B6000179B5,
-    0x0001B8E1000161B7, 0x001F38E3001FD4E1, 0x80003ED7001B0CE3,
-    0x80003EDF80003EDB, 0x80003EEB80003EE7, 0x80003EF380003EEF,
-    0x00093CBE80003EF7, 0x001B84EF80003EFB, 0x0002A1A3001B84EF,
-    0x0004E88680003EFF, 0x80003F030005BC87, 0x00027DA500027DA6,
-    0x000259AB80003F07, 0x80003F0B000251AA, 0x0001FDAF80003F0F,
-    0x0001FDAD0001FDAE, 0x000685B3000241AC, 0x80003F1380003EE3,
-    0x80003F1B80003F17, 0x00068DB5000689B6, 0x000691BB80003F1F,
-    0x80003F23000689BA, 0x0006F5BF80003F27, 0x000681BD000689BE,
-    0x80003F2B000689BC, 0x80003F3380003F2F, 0x80003F3B80003F37,
-    0x80003F4380003F3F, 0x80003F47001D84EF, 0x80003F4B000400E1,
-    0x000011800004FCE3, 0x0000058200001D81, 0x0006F1A880003F4F,
-    0x00060DAA0006F1A9, 0x000605AC000605AB, 0x00063DAE000609AD,
-    0x80003F53000635AF, 0x0000888680003F57, 0x80003F5B00031087,
-    0x80003F6380003F5F, 0x000611B880003F67, 0x000621BA000619B9,
-    0x0007F5BC000621BB, 0x0007F5BE0007FDBD, 0x00064DB00007E9BF,
-    0x00065DB2000655B1, 0x00064DB4000655B3, 0x000631B6000631B5,
-    0x000785A3000631B7, 0x80003F6F80003F6B, 0x80003F7780003F73,
-    0x00078DA5000789A6, 0x000791AB80003F7B, 0x80003F7F000789AA,
-    0x0007F5AF80003F83, 0x000781AD000789AE, 0x80003F87000789AC,
-    0x80003F8B0006E1B3, 0x0006E5B680003F8F, 0x80003F9780003F93,
-    0x0006BDBA0006F1B5, 0x80003F9B00068DBB, 0x00017DBE80003F9F,
-    0x000695BC000165BF, 0x00061DA8000175BD, 0x00062DAA000625A9,
-    0x00063DAC000625AB, 0x00065DAE000621AD, 0x80003FA300064DAF,
-    0x80003FAB80003FA7, 0x80003FB380003FAF, 0x0003B1810003B982,
-    0x0001EDB80003B980, 0x00018DBA000185B9, 0x00019DBC000185BB,
-    0x00018DBE000185BD, 0x00063DB0000185BF, 0x000605B200060DB1,
-    0x0001FDB40001E5B3, 0x0001E5B60001E5B5, 0x0005A5A30001D5B7,
-    0x80003FBB80003FB7, 0x80003FC780003FBF, 0x0005B5A50005A1A6,
-    0x0005C9AB80003FCB, 0x000C08860005F9AA, 0x000221AF00031C87,
-    0x000231AD000239AE, 0x80003FCF0005D1AC, 0x80003FD30002F1B3,
-    0x000355B680003FD7, 0x80003FDF80003FDB, 0x000371BA0002E1B5,
-    0x80003FE3000379BB, 0x000331BE80003FE7, 0x000351BC00033DBF,
-    0x000225A8000351BD, 0x00023DAA000235A9, 0x00022DAC000235AB,
-    0x000391AE000391AD, 0x80003FEB000391AF, 0x80003FF380003FEF,
-    0x000000AC80003FF7, 0x80003FFF80003FFB, 0x000399B880004003,
-    0x0003A5BA0003ADB9, 0x000375BC00036DBB, 0x000375BE00037DBD,
-    0x0003F1B000036DBF, 0x0003C1B20003F9B1, 0x0003B1B40003C1B3,
-    0x0003B5B60003BDB5, 0x800040070003A9B7, 0x8000400F8000400B,
-    0x8000401780004013, 0x8000401F8000401B, 0x000C68BE001AA8EF,
-    0x80004023000194E1, 0x00000D8000061CE3, 0x0000718200007181,
-    0x0003F5A380004027, 0x8000402F8000402B, 0x8000403700024C84,
-    0x0003E5A5000251A6, 0x00027DAB8000403B, 0x000CC886000275AA,
-    0x000239AF000D2C87, 0x000255AD000235AE, 0x000650E1000255AC,
-    0x000714E38000403F, 0x80004043000CC084, 0x8000404B80004047,
-    0x800040538000404F, 0x8000405B80004057, 0x800040638000405F,
-    0x001B78EF80004067, 0x8000406B000FF8BE, 0x800040738000406F,
-    0x8000407700018DB3, 0x00018DB6000199B5, 0x80003E738000407B,
-    0x000185BA8000407F, 0x000155BC00014DBB, 0x000155BE00015DBD,
-    0x000E41A3000149BF, 0x8000408380004033, 0x8000408B80004087,
-    0x000E55A5000E41A6, 0x000E81AB8000408F, 0x80004093000E49AA,
-    0x000E85AF80004097, 0x000E91AD000E99AE, 0x00006D80000E99AC,
-    0x00001D8200000981, 0x000918EF8000409B, 0x800040A38000409F,
-    0x000D3CE3800040A7, 0x000C58E1800040AB, 0x0000D086800040AF,
-    0x800040B30003BC87, 0x000E91A8800040B7, 0x000EC9AA000E99A9,
-    0x000EDDAC000EC5AB, 0x000EC1AE000EC1AD, 0x0000E884000EF5AF,
-    0x800040BF800040BB, 0x800040C7800040C3, 0x800040CF800040CB,
-    0x000FC1B8800040D3, 0x000FC1BA000FC1B9, 0x000FC1BC000FC1BB,
-    0x000FC1BE000FC1BD, 0x000E8DB0000FF5BF, 0x000E4DB2000E45B1,
-    0x000E5DB4000E45B3, 0x000E41B6000E41B5, 0x000E51A8000E41B7,
-    0x000E8DAA000E59A9, 0x000E85AC000E9DAB, 0x000E85AE000E8DAD,
-    0x800040D7000EBDAF, 0x800040DF800040DB, 0x800040E7800040E3,
-    0x800040EF800040EB, 0x000E91B8800040F3, 0x000EADBA000E99B9,
-    0x00015DBC000145BB, 0x000145BE000145BD, 0x000EC5B0000175BF,
-    0x000EC5B2000ECDB1, 0x000EC5B4000EDDB3, 0x000EBDB6000EB5B5,
-    0x000E05B3000EB5B7, 0x800040FB800040F7, 0x80004103800040FF,
-    0x000E0DB5000E0DB6, 0x000E09BB80004107, 0x8000410B000E01BA,
-    0x000E71BF8000410F, 0x000E09BD000E01BE, 0x00006D82000E11BC,
-    0x00005580000E41A3, 0x000E49A600006581, 0x8000411B80004117,
-    0x000E45AA000E49A5, 0x00004886000E4DAB, 0x000E45AE00007887,
-    0x000E55AC000E35AF, 0x00025DA8000E4DAD, 0x000265AA000261A9,
-    0x00026DAC000275AB, 0x0002B1AE0002B1AD, 0x0004EC840002B1AF,
-    0x800041238000411F, 0x8000412B80004127, 0x800041338000412F,
-    0x000371B880004137, 0x000371BA000371B9, 0x0003D5BC000371BB,
-    0x0003D5BE0003DDBD, 0x0002D1B00003CDBF, 0x0002D1B20002D1B1,
-    0x000351B40002D1B3, 0x000351B6000351B5, 0x8000413B000351B7,
-    0x8000413F0002E9B3, 0x000235B6000480BE, 0x8000414780004143,
-    0x000211BA0002E1B5, 0x8000414B000211BB, 0x000391BE8000414F,
-    0x000201BC000391BF, 0x80004153000201BD, 0x0002A5A380004157,
-    0x0002ADA58000415B, 0x800041638000415F, 0x80004167000279A6,
-    0x00025DAB8000416B, 0x00024DAD00025DAA, 0x0003DDAF00024DAC,
-    0x0002D5A80003DDAE, 0x00012DAA0002DDA9, 0x00013DAC000125AB,
-    0x00012DAE000125AD, 0x8000416F000125AF, 0x8000417780004173,
-    0x8000417F8000417B, 0x8000418B80004187, 0x000185B88000418F,
-    0x00019DBA000189B9, 0x0001B1BC000195BB, 0x000079BE0001B1BD,
-    0x00015DB0000079BF, 0x0001E1B20001E5B1, 0x0001E9B40001F9B3,
-    0x0001D5B60001DDB5, 0x000EF0E10001BDB7, 0x000E14E380004193,
-    0x0000BD8180004197, 0x8000419B0000BD80, 0x000400860000AD82,
-    0x8000419F0005EC87, 0x800041A7800041A3, 0x000E60EF800041AB,
-    0x800041B3800041AF, 0x007D5C85800041B7, 0x800041BF800041BB,
-    0x800041C3000164E3, 0x800041C70001B4E1, 0x800041CB000EA0EF,
-    0x0005F88480003FC3, 0x800041D3800041CF, 0x000625B3800041D7,
-    0x800041DB80004183, 0x800041E3800041DF, 0x000635B5000625B6,
-    0x0006A1BB800041E7, 0x800041EB000619BA, 0x000699BF800041EF,
-    0x0006ADBD0006ADBE, 0x00006D820006B5BC, 0x00005580000430EF,
-    0x00035CBE00006581, 0x0000F886800041F3, 0x800041F700036C87,
-    0x800041FF800041FB, 0x8000420780004203, 0x8000420B000444E3,
-    0x0006A5A300078CE1, 0x800042138000420F, 0x8000421B80004217,
-    0x0006B5A50006A5A6, 0x000621AB8000421F, 0x80004223000699AA,
-    0x000619AF80004227, 0x00062DAD00062DAE, 0x8000422B000635AC,
-    0x8000422F0007F9B3, 0x000749B680004233, 0x8000423B80004237,
-    0x000751BA000751B5, 0x8000423F000751BB, 0x000735BE80004243,
-    0x000745BC000739BF, 0x000635A800073DBD, 0x000635AA00063DA9,
-    0x00069DAC000689AB, 0x0006B1AE00068DAD, 0x800042470006B1AF,
-    0x8000424F8000424B, 0x00000D8080004253, 0x0000B1820000B181,
-    0x0006A1B880004257, 0x0006BDBA0006ADB9, 0x0006ADBC0006B5BB,
-    0x000155BE00015DBD, 0x0006D1B000014DBF, 0x0006D5B20006D1B1,
-    0x0006B5B40006ADB3, 0x0006A9B60006B9B5, 0x0006BDA30006A5B7,
-    0x8000425F8000425B, 0x000190BE00028484, 0x000615A500060DA6,
-    0x000615AB80004267, 0x00000886000615AA, 0x00067DAF00017C87,
-    0x000679AD000671AE, 0x8000426B000601AC, 0x8000426F0001ADB3,
-    0x0001A9B680004273, 0x8000427B80004277, 0x00016DBA0001ADB5,
-    0x8000427F000175BB, 0x00015DBE80004283, 0x000165BC000135BF,
-    0x000261A8000155BD, 0x000261AA000261A9, 0x000261AC000261AB,
-    0x00029DAE00026DAD, 0x80004287000295AF, 0x8000428F8000428B,
-    0x8000429780004293, 0x8000429F8000429B, 0x000295B8800042A3,
-    0x0002A1BA00029DB9, 0x000371BC0002A1BB, 0x000371BE000371BD,
-    0x0002EDB0000371BF, 0x0002F5B20002F5B1, 0x0002DDB40002C5B3,
-    0x0002B1B60002B5B5, 0x800042A70002B1B7, 0x800042AF800042AB,
-    0x800042B30002E5A3, 0x0002E1A60002E5A5, 0x800042BB800042B7,
-    0x000225AA800042BF, 0x00022DAC00023DAB, 0x000215AE00021DAD,
-    0x800042C300027DAF, 0x800042CB800042C7, 0x00001580007C4084,
-    0x0000058200001D81, 0x007CECBE800042CF, 0x007D4886800042D7,
-    0x800042DB00030887, 0x800042E3800042DF, 0x00029DA8800042E7,
-    0x0002C1AA0002C5A9, 0x0002C5AC0002C1AB, 0x0002F1AE0002CDAD,
-    0x800042EB0002F1AF, 0x800042F3800042EF, 0x000000C9800042F7,
-    0x800042FF800042FB, 0x0001C1B880004303, 0x0001C1BA0001C1B9,
-    0x0001CDBC0001DDBB, 0x0001FDBE0001F5BD, 0x000141B000019DBF,
-    0x000141B2000141B1, 0x000141B4000141B3, 0x000141B6000141B5,
-    0x000638E1000141B7, 0x000668E380004307, 0x8000430F8000430B,
-    0x8000431780004313, 0x007CACBE007D9484, 0x8000431F8000431B,
-    0x007FB8BE80004323, 0x000110EF80004327, 0x8000432F8000432B,
-    0x8000433780004333, 0x000190E18000433B, 0x000044E38000433F,
-    0x00002D8180004347, 0x000080EF00002D80, 0x8000434B00003982,
-    0x800041138000434F, 0x007FB0E180004353, 0x007CD4E3800042D3,
-    0x8000434380004357, 0x0000B8868000435B, 0x8000435F0002D487,
-    0x8000436780004363, 0x8000436F8000436B, 0x007C20EF80004373,
-    0x007D89B380004377, 0x8000437F8000437B, 0x8000438780004383,
-    0x007D8DB5007D8DB6, 0x007E45BB8000438B, 0x8000438F007E45BA,
-    0x007E45BF80004393, 0x007E55BD007E45BE, 0x007D49A3007E55BC,
-    0x8000439B80004397, 0x800043A38000439F, 0x007D4DA5007D4DA6,
-    0x007E85AB800043A7, 0x800043AB007E85AA, 0x007E85AF800043AF,
-    0x007E95AD007E85AE, 0x00006D82007E95AC, 0x00005580007E31B3,
-    0x007FF1B600006581, 0x800043B30003DC84, 0x007FCDBA007E21B5,
-    0x00040086007FD1BB, 0x007F75BE0000D487, 0x007FC1BC007F79BF,
-    0x007FE5A8007FC1BD, 0x007FFDAA007FF5A9, 0x007FEDAC007FF5AB,
-    0x007E3DAE007E35AD, 0x800043B7007E35AF, 0x800043BF800043BB,
-    0x800043C7800043C3, 0x800043CF800043CB, 0x007ED9B8800043D3,
-    0x007EE1BA007EE1B9, 0x007EE5BC007EE1BB, 0x007E99BE007EE9BD,
-    0x007E4DB0007E99BF, 0x007E51B2007E51B1, 0x007EF5B4007E51B3,
-    0x007EE9B6007EF9B5, 0x007F75A3007EE9B7, 0x800043DB800043D7,
-    0x800043E3800043DF, 0x007F65A5007EB5A6, 0x007E95AB800043E7,
-    0x800043EB007E89AA, 0x007E3DAF800043EF, 0x007E85AD007E31AE,
-    0x800043F3007E85AC, 0x800043F7007E31B3, 0x0001C5B6800043FB,
-    0x80004403800043FF, 0x0001C9BA0001D1B5, 0x800044070001C9BB,
-    0x0001CDBE8000440B, 0x0001C9BC0001B1BF, 0x007DDDA80001C9BD,
-    0x007DFDAA007DF5A9, 0x00021DAC007DF1AB, 0x000239AE000201AD,
-    0x8000440F000239AF, 0x8000441780004413, 0x000005828000441B,
-    0x000011808000441F, 0x000211B800001181, 0x000221BA000219B9,
-    0x0002D5BC000221BB, 0x0002D5BE0002DDBD, 0x000249B00002CDBF,
-    0x000259B2000249B1, 0x00024DB4000259B3, 0x000231B6000231B5,
-    0x000300BE000231B7, 0x00028884007D71A3, 0x000285A6000480BE,
-    0x8000442B80004427, 0x000289AA000291A5, 0x00042886000289AB,
-    0x00028DAE00030C87, 0x000289AC0002F1AF, 0x8000442F000289AD,
-    0x0005CC850003C884, 0x800044330003E5B3, 0x0003EDB60003E5B5,
-    0x8000443B80004437, 0x00036DBA8000443F, 0x00037DBC000365BB,
-    0x000365BE000365BD, 0x80004443000355BF, 0x000400BF80004447,
-    0x8000444B000225A3, 0x00022DA6000225A5, 0x800044538000444F,
-    0x0002ADAA80004457, 0x0002BDAC0002A5AB, 0x0002A5AE0002A5AD,
-    0x8000445B000295AF, 0x800044638000445F, 0x8000446B80004467,
-    0x8000446F0003CCE3, 0x800044730001ACE1, 0x8000447700031CEF,
-    0x8000447F8000447B, 0x8000448780004483, 0x007F70E18000448B,
-    0x007E04E380004263, 0x800044978000448F, 0x8000449B000194E1,
-    0x0000D980000158E3, 0x000025820000D981, 0x8000449F007EE4EF,
-    0x007EE0EF800044A3, 0x000101B3800044A7, 0x0007E886800044AB,
-    0x800044AF00042C87, 0x000105B5000135B6, 0x0000F1BB800044B3,
-    0x800044B70000E1BA, 0x0000B1BF800044BB, 0x0000D9BD0000D1BE,
-    0x800044230000E1BC, 0x800044BF80004493, 0x000401A1800044C3,
-    0x000019A3000411A0, 0x0006B5A80004C5A2, 0x0006E9AA0006BDA9,
-    0x0006EDAC0006FDAB, 0x0006DDAE0006D5AD, 0x000745B000073DAF,
-    0x000755B2000755B1, 0x000775B400076DB3, 0x00071DB6000715B5,
-    0x00072DB8000715B7, 0x000731BA000735B9, 0x000719BC00070DBB,
-    0x000709BE000719BD, 0x000641A3000709BF, 0x800044CB800044C7,
-    0x800044D3800044CF, 0x000645A5000675A6, 0x0007B1AB800044D7,
-    0x006DC98F0007A1AA, 0x0007F1AF800044DB, 0x000799AD000791AE,
-    0x0075DD950007A1AC, 0x0070CD970074C194, 0x00685D9100732D96,
-    0x0074D99300685590, 0x0078E59D00694D92, 0x00076D9F007B1D9C,
-    0x007F1D990078B99E, 0x007CA19B00705598, 0x006C6182007CF99A,
-    0x800044DF00698583, 0x00751186800044E3, 0x0069958400751187,
-    0x00743D8A00688D85, 0x800044E70072C58B, 0x007EDD8E800044EB,
-    0x00713D8C007D318F, 0x007D199200719D8D, 0x800044EF007ABD93,
-    0x0006D596000990EF, 0x00795D9400055197, 0x0005719A00790995,
-    0x800044F30005BD9B, 0x800044FB800044F7, 0x0001799C0005A8E1,
-    0x000185A10008B8E3, 0x000DA9A3800044FF, 0x000801A5000C01A2,
-    0x0009A9A7000D39A4, 0x0015A5A90009BDA6, 0x001401AB001401A8,
-    0x0011B9AD0015FDAA, 0x001C01AF0011B1AC, 0x001CFDB10010A1AE,
-    0x0019E9B3001DE5B0, 0x002401B5001801B2, 0x0000148400197DB4,
-    0x00008D80000014BE, 0x00006D8200009581, 0x000F508680004507,
-    0x8000450B00006487, 0x0007D5BA8000450F, 0x0007C1B80007DDBB,
-    0x000431BE0007C1B9, 0x0007F1BC000431BF, 0x0007ADB20007F1BD,
-    0x0007ADB00007B5B3, 0x00079DB60007A5B1, 0x0007A5B40007F5B7,
-    0x000769AA000795B5, 0x000769A8000769AB, 0x000769AE000769A9,
-    0x000769AC000769AF, 0x80004513000769AD, 0x8000451B80004517,
-    0x800045238000451F, 0x8000452B80004527, 0x000511A88000452F,
-    0x000539AA00051DA9, 0x00052DAC000539AB, 0x000549AE000551AD,
-    0x80004533000541AF, 0x8000453B80004537, 0x800045438000453F,
-    0x8000454B80004547, 0x00025DB88000454F, 0x0003C1BA000269B9,
-    0x0003F9BC0003C1BB, 0x000391BE0003F9BD, 0x000509B00003B5BF,
-    0x0002E1B2000509B1, 0x000275B40002E1B3, 0x000275B600027DB5,
-    0x0004B1B300026DB7, 0x000D04BE00020084, 0x8000455780004553,
-    0x0004A5B50004D1B6, 0x0004CDBB8000455B, 0x8000455F0004CDBA,
-    0x0003B9BF80004563, 0x000335BD0003B1BE, 0x80004567000335BC,
-    0x8000456B0004F5A3, 0x000495A68000456F, 0x8000457B80004577,
-    0x000489AA0004E1A5, 0x000DA887000489AB, 0x0003F5AE000CCC86,
-    0x000371AC0003FDAF, 0x000650E1000371AD, 0x000040E3000734E1,
-    0x0000D180000758E3, 0x0000D5820000DD81, 0x800045838000457F,
-    0x8000458B80004587, 0x800045938000458F, 0x00009CEF80004597,
-    0x8000459B0007C8EF, 0x000235B38000459F, 0x0001B5B5800045A3,
-    0x800045AB800045A7, 0x000CECBE0001B5B6, 0x000191BB800045AF,
-    0x000151BD000199BA, 0x000151BF000189BC, 0x000DEDA8000159BE,
-    0x000DF5AA000DFDA9, 0x000E75AC000E6DAB, 0x000E75AE000E7DAD,
-    0x80004573000E6DAF, 0x800045B7800045B3, 0x800045BF800045BB,
-    0x800045C7800045C3, 0x000EF5B8800045CB, 0x000EF5BA000EFDB9,
-    0x000F41BC000F41BB, 0x000F71BE000F49BD, 0x000E15B0000F71BF,
-    0x000E15B2000E1DB1, 0x000ED5B4000ECDB3, 0x000ED5B6000EDDB5,
-    0x000D7DA3000ECDB7, 0x800045D3800045CF, 0x800045DB800045D7,
-    0x000EFDA5000EFDA6, 0x000ED9AB800045DF, 0x0002A884000ED1AA,
-    0x000E19AF800045E3, 0x000E19AD000E11AE, 0x00004D80000EC1AC,
-    0x0000558200005581, 0x800045E7000F51B3, 0x000F71B6000F71B5,
-    0x0000E086800045EB, 0x000F5DBA00030487, 0x000F35BC000F2DBB,
-    0x000F2DBE000F39BD, 0x000E55A8000F25BF, 0x000E95AA000E5DA9,
-    0x000EB5AC000EADAB, 0x000EB5AE000EBDAD, 0x800045EF000EADAF,
-    0x800045F7800045F3, 0x800045FF800045FB, 0x8000460780004603,
-    0x000169B88000460B, 0x000179BA000169B9, 0x000169BC000179BB,
-    0x0001DDBE000169BD, 0x000ED5B00001D5BF, 0x000EADB2000EA5B1,
-    0x000EA5B4000EA1B3, 0x000E9DB6000EADB5, 0x000E1DA3000159B7,
-    0x800046138000460F, 0x8000461780004503, 0x000E3DA5000E3DA6,
-    0x000E61AB8000461B, 0x8000461F000E11AA, 0x000E69AF80004623,
-    0x000E75AD000E61AE, 0x80004627000E79AC, 0x8000462F8000462B,
-    0x8000463780004633, 0x8000463F8000463B, 0x00000D8080004643,
-    0x00001D8200001581, 0x8000464B80004647, 0x000178848000464F,
-    0x000F14E3000014BE, 0x000DE0E180004657, 0x0004B08600030084,
-    0x8000465B00031487, 0x800046638000465F, 0x8000466B80004667,
-    0x000FC0EF8000466F, 0x8000467780004673, 0x8000467F8000467B,
-    0x8000468780004683, 0x8000468B00036DB3, 0x000375B600037DB5,
-    0x800046938000468F, 0x000351BA80004697, 0x0003F5BC000351BB,
-    0x0003E9BE0003FDBD, 0x8000469B0003E1BF, 0x800046A38000469F,
-    0x800046AB800046A7, 0x800046B3800046AF, 0x000271A8800046B7,
-    0x000389AA000279A9, 0x000399AC000389AB, 0x000389AE000399AD,
-    0x0003F9B0000389AF, 0x00034DB20003F9B1, 0x000341B4000345B3,
-    0x000371B6000349B5, 0x000321B8000371B7, 0x000321BA000321B9,
-    0x000321BC000321BB, 0x000321BE000321BD, 0x00019D80000321BF,
-    0x0000118200001181, 0x000014EF00050484, 0x800046BF800046BB,
-    0x0000C8E3000510BE, 0x0001D0E1800046C7, 0x800046CF800046CB,
-    0x800046D7800046D3, 0x000279AA800046DB, 0x0004E886000279AB,
-    0x0002C1AE00056087, 0x0002DDAC0002C9AF, 0x800046DF0002D5AD,
-    0x800046E3000245A3, 0x00025DA6800046E7, 0x800046EF800046EB,
-    0x800046F3000255A5, 0x800046FB800046F7, 0x80004703800046FF,
-    0x8000470B80004707, 0x0004ACBE000E84EF, 0x8000470F000ED0E1,
-    0x80004713000114E3, 0x8000471B80004717, 0x0001DDB38000471F,
-    0x8000472780004723, 0x8000472F8000472B, 0x0001FDB50001FDB6,
-    0x0001DDBB80004737, 0x0004E0840001DDBA, 0x0001A1BF8000473B,
-    0x0001BDBD0001BDBE, 0x000605A80001BDBC, 0x000615AA00061DA9,
-    0x000635AC00062DAB, 0x000635AE00063DAD, 0x800046C3000629AF,
-    0x0007E5810007F582, 0x800047330007E580, 0x00001C868000473F,
-    0x000625B80003B087, 0x0006CDBA0006C5B9, 0x0006DDBC0006C5BB,
-    0x0006CDBE0006C5BD, 0x000659B00006C5BF, 0x000629B2000659B1,
-    0x000639B4000629B3, 0x000625B6000625B5, 0x00069DA300061DB7,
-    0x8000474780004743, 0x8000474F8000474B, 0x0006BDA50006BDA6,
-    0x00069DAB80004753, 0x8000475700069DAA, 0x0006E1AF8000475B,
-    0x0006FDAD0006FDAE, 0x8000475F0006FDAC, 0x800047630007F5B3,
-    0x0007EDB680004767, 0x8000476F8000476B, 0x00078DBA0007E5B5,
-    0x80004773000791BB, 0x000775BE80004777, 0x000781BC00077DBF,
-    0x000625A800077DBD, 0x000639AA000629A9, 0x000629AC000639AB,
-    0x000679AE000629AD, 0x8000477B000675AF, 0x800047838000477F,
-    0x8000478B80004787, 0x800047938000478F, 0x0006D5B880004797,
-    0x0006E1BA0006DDB9, 0x0006E5BC0006FDBB, 0x0006E5BE0006EDBD,
-    0x00060DB0000699BF, 0x000611B2000611B1, 0x0006F5B40006EDB3,
-    0x0006F5B60006FDB5, 0x0006B1A30006EDB7, 0x0000158100002D82,
-    0x800046530000B180, 0x0006A1A50006A9A6, 0x0006D5AB8000479B,
-    0x8000479F0006C9AA, 0x000639AF000168BE, 0x000639AD000631AE,
-    0x800047A70006C5AC, 0x000068860001F1B3, 0x00015DB600017C87,
-    0x800047AF800047AB, 0x000149BA000155B5, 0x800047B3000149BB,
-    0x00013DBE800047B7, 0x000139BC000121BF, 0x800047BB000135BD,
-    0x00030484800047BF, 0x800047C3001C80BE, 0x800047C700061CE1,
-    0x002E08BF000600E3, 0x800047CF800047CB, 0x800047D30007C8EF,
-    0x800047DB800047D7, 0x800047E3800047DF, 0x00027DA3800047E7,
-    0x0002D9A5800047EB, 0x800047F3800047EF, 0x800047F70002D1A6,
-    0x0002C5AB800047FB, 0x0002B9AD0002C5AA, 0x0002ADAF0002B5AC,
-    0x00056DA90002B1AE, 0x00020DAB000559A8, 0x00021DAD0002B9AA,
-    0x000205AF00021DAC, 0x001D68BE00020DAE, 0x80004803800047FF,
-    0x00001D8080004807, 0x0001998200000981, 0x0003E1B98000480B,
-    0x0003E1BB000229B8, 0x0003E1BD0003E9BA, 0x0003E1BF0003F9BC,
-    0x00024DB10003E9BE, 0x000221B300024DB0, 0x000225B500022DB2,
-    0x000211B7000239B4, 0x0002C1A8000225B6, 0x0002D1AA0002D1A9,
-    0x0002FDAC0002E5AB, 0x00011DAE000115AD, 0x8000480F00010DAF,
-    0x8000481B80004817, 0x800048238000481F, 0x8000482B80004827,
-    0x000105B88000482F, 0x00011DBA000109B9, 0x000131BC000115BB,
-    0x0001FDBE000131BD, 0x000179B00001F5BF, 0x000141B2000141B1,
-    0x000145B400015DB3, 0x000145B600014DB5, 0x001DA08600013DB7,
-    0x80004833001DC487, 0x800048370000D8EF, 0x8000483F8000483B,
-    0x001C6C840006C4EF, 0x800048430006F4E1, 0x800048470006B0E3,
-    0x8000484B000194E1, 0x000219B3000644E3, 0x800048538000484F,
-    0x002CEC8580004857, 0x000235B50001D5B6, 0x0001C5BB8000485B,
-    0x8000485F0001FDBA, 0x0001C9BF80004863, 0x0001C9BD0001C1BE,
-    0x001DD9A30001D5BC, 0x8000486780004813, 0x8000486F8000486B,
-    0x001DF5A5001E15A6, 0x001E05AB80004873, 0x80004877001E3DAA,
-    0x001E09AF8000487B, 0x001E09AD001E01AE, 0x00006980001E15AC,
-    0x0000058200006981, 0x800048838000487F, 0x0003108780004887,
-    0x8000488B00037C86, 0x800048938000488F, 0x8000489B80004897,
-    0x800048A38000489F, 0x001FA5A8800048A7, 0x001FA5AA001FADA9,
-    0x001FA5AC001FBDAB, 0x001FA5AE001FADAD, 0x800048AB001F15AF,
-    0x800048B3800048AF, 0x800048BB800048B7, 0x800048C3800048BF,
-    0x001F0DB8800048C7, 0x001F21BA001F19B9, 0x0000D9BC001F21BB,
-    0x0000C9BE0000D9BD, 0x001F71B00000C1BF, 0x001F71B2001F71B1,
-    0x001F41B4001F45B3, 0x001F3DB6001F4DB5, 0x001E2DB3001F35B7,
-    0x800048CF800048CB, 0x800048D7800048D3, 0x001E2DB5001E2DB6,
-    0x001EB5BB800048DB, 0x800048DF001EB5BA, 0x001E89BF800048E3,
-    0x001E91BD001E85BE, 0x00002982001EA5BC, 0x00001D80001E69A3,
-    0x001E69A600001581, 0x800048EB800048E7, 0x001EF1AA001E69A5,
-    0x800048EF001EF1AB, 0x001EC1AE0001E084, 0x001EE1AC001ECDAF,
-    0x0001D5A8001ED5AD, 0x0001EDAA0001E5A9, 0x0001FDAC0001E5AB,
-    0x0001EDAE0001E5AD, 0x0001A0BE0001E5AF, 0x800048F3800047A3,
-    0x00001086800048F7, 0x800048FB00017487, 0x00017DB8800048FF,
-    0x0000C1BA0000C1B9, 0x0000C1BC0000C1BB, 0x0000F1BE0000C9BD,
-    0x00019DB00000F1BF, 0x00014DB2000145B1, 0x00015DB4000145B3,
-    0x00014DB6000145B5, 0x80004903000145B7, 0x8000490B80004907,
-    0x800049138000490F, 0x001EECEF000234EF, 0x001DF0E180004917,
-    0x001E30E30002E0E3, 0x8000491B000184E1, 0x800049238000491F,
-    0x000225B380004927, 0x8000492B00000094, 0x800049338000492F,
-    0x000235B5000225B6, 0x0002C1BB80004937, 0x8000493B000219BA,
-    0x0002C1BF8000493F, 0x0002D1BD0002D9BE, 0x800049430002D9BC,
-    0x8000494B80004947, 0x8000494F0002E9A3, 0x0002E9A60002F9A5,
-    0x8000495780004953, 0x0002D5AA8000495B, 0x000215AC00020DAB,
-    0x000215AE00021DAD, 0x0000618000020DAF, 0x0000058200006181,
-    0x800049678000495F, 0x000414BE00040084, 0x0004C0868000496B,
-    0x8000496F00035087, 0x8000497780004973, 0x8000497F8000497B,
-    0x0002ADA880004983, 0x0002B5AA0002BDA9, 0x000115AC00010DAB,
-    0x000115AE00011DAD, 0x0004EC8400010DAF, 0x8000498B80004987,
-    0x800049938000498F, 0x8000499B80004997, 0x00011DB88000499F,
-    0x000125BA00012DB9, 0x0001D5BC0001CDBB, 0x0001C9BE0001DDBD,
-    0x00017DB00001C1BF, 0x00015DB2000155B1, 0x00014DB4000155B3,
-    0x000135B600013DB5, 0x001E18E100012DB7, 0x001E38E3800049A3,
-    0x800049AB800049A7, 0x800049B3800049AF, 0x800049BB800049B7,
-    0x000460BE800049BF, 0x00007581800049C3, 0x001FE0EF00007580,
-    0x800049C700006D82, 0x0004E886800049CB, 0x800049CF00057C87,
-    0x800049D7000190E1, 0x800049DB0000A0E3, 0x800049E3800049DF,
-    0x800049E70000B4EF, 0x800049EF800049EB, 0x000605B5800049F3,
-    0x800049D380004963, 0x800049F7000605B6, 0x0005A5B3800049FB,
-    0x000619BD800049FF, 0x000661BF000611BC, 0x80004A03000611BE,
-    0x000609BB80004A07, 0x0005FDA3000621BA, 0x80004A0F80004A0B,
-    0x80004A1780004A13, 0x00065DA500065DA6, 0x000651AB80004A1B,
-    0x80004A1F000679AA, 0x000639AF80004A23, 0x000641AD000649AE,
-    0x000651A8000649AC, 0x000661AA000659A9, 0x000661AC000661AB,
-    0x000661AE000661AD, 0x80004A27000661AF, 0x80004A2F80004A2B,
-    0x00000D8080004A33, 0x0001B1820001B181, 0x00074DB880004A37,
-    0x00075DBA000755B9, 0x00074DBC000755BB, 0x00077DBE000775BD,
-    0x0007C5B0000771BF, 0x0007C5B20007CDB1, 0x0007C5B40007DDB3,
-    0x0007C5B60007CDB5, 0x0007E9B3000779B7, 0x80004A3F80004A3B,
-    0x0001D8BE0001C084, 0x0007E9B50007E1B6, 0x000609BB80004A43,
-    0x00008886000601BA, 0x000609BF0001AC87, 0x000609BD000601BE,
-    0x80004A47000611BC, 0x80004A4B0007ADA3, 0x0007A5A680004A4F,
-    0x80004A5780004A53, 0x000645AA0007ADA5, 0x80004A5B00064DAB,
-    0x000645AE80004A5F, 0x000655AC00064DAF, 0x000659A800064DAD,
-    0x00066DAA000665A9, 0x000661AC000665AB, 0x000661AE000661AD,
-    0x0001AC84000661AF, 0x80004A6780004A63, 0x80004A6F80004A6B,
-    0x80004A7780004A73, 0x000191B880004A7B, 0x0001A1BA000199B9,
-    0x000171BC0001A1BB, 0x000171BE000171BD, 0x0001F1B0000171BF,
-    0x0001F5B20001F1B1, 0x0001C5B40001DDB3, 0x0001B1B60001B1B5,
-    0x0005E5B30001B1B7, 0x80004A8380004A7F, 0x80004A8B80004A87,
-    0x0005E9B50005E1B6, 0x000235BB80004A8F, 0x80004A93000235BA,
-    0x000275BF80004A97, 0x000209BD000201BE, 0x80004A9B000221BC,
-    0x80004A9F0005A1A3, 0x0005A5A680004AA3, 0x80004AAB80004AA7,
-    0x000271AA0005ADA5, 0x80004AAF000271AB, 0x000245AE000328BE,
-    0x000265AC000231AF, 0x0000D58000024DAD, 0x0000E1820000DD81,
-    0x000120E380004AB7, 0x0001A8E180004ABB, 0x000234EF80004ABF,
-    0x000C088680004AC3, 0x000000AC00036887, 0x80004ACB80004AC7,
-    0x00038DB380004ACF, 0x80004AD780004AD3, 0x80004ADB000C8084,
-    0x000381B50003BDB6, 0x00034DBB80004ADF, 0x80004AE300034DBA,
-    0x00034DBF80004AE7, 0x00034DBD00034DBE, 0x000241A800034DBC,
-    0x000245AA00024DA9, 0x000249AC000259AB, 0x000275AE00027DAD,
-    0x000D68BE0002B9AF, 0x80004AEF80004AEB, 0x000C6C8480004AF3,
-    0x80004AFB80004AF7, 0x00014DB880004AFF, 0x000155BA000155B9,
-    0x00017DBC000165BB, 0x000111BE000115BD, 0x0002C9B0000111BF,
-    0x0002D9B20002C9B1, 0x0002C9B40002D9B3, 0x00017DB60002C9B5,
-    0x000778E1000175B7, 0x0006B8E3000098E3, 0x80004B0300065CE1,
-    0x80004B0B80004B07, 0x80004B1380004B0F, 0x80004B1B80004B17,
-    0x000000EC80004B1F, 0x0006E8EF00005CEF, 0x0000698180004B27,
-    0x000285A300006180, 0x000289A500006182, 0x80004B2F80004B2B,
-    0x000C40860002B5A6, 0x000245AB000CC487, 0x000245AD000245AA,
-    0x000245AF000245AC, 0x000E8DA8000245AE, 0x000E95AA000E91A9,
-    0x000EA5AC000EA9AB, 0x000EA5AE000EADAD, 0x80004B23000ED9AF,
-    0x80004B3780004B33, 0x80004B3F80004B3B, 0x80004B4780004B43,
-    0x000F75B880004B4B, 0x000F75BA000F7DB9, 0x000FC5BC000FDDBB,
-    0x000FC5BE000FCDBD, 0x000EA9B0000FFDBF, 0x000EB5B2000EB5B1,
-    0x000E9DB4000E85B3, 0x000F51B6000F51B5, 0x000E1DB3000F51B7,
-    0x80004B5380004B4F, 0x80004B5B80004B57, 0x000E2DB5000E2DB6,
-    0x000E71BB80004B5F, 0x80004B63000E6DBA, 0x000F95BF80004B67,
-    0x000E51BD000E59BE, 0x80004B6B000E61BC, 0x80004B6F000E59A3,
-    0x000E69A680004B73, 0x80004B7B80004B77, 0x000E29AA000E69A5,
-    0x80004B7F000E35AB, 0x000E1DAE80004B83, 0x000E25AC000FD1AF,
-    0x000EBDA8000E15AD, 0x000ED1AA000ED1A9, 0x000139AC000129AB,
-    0x000129AE000139AD, 0x00000D80000129AF, 0x00001D8200001581,
-    0x80004B8B80004B87, 0x000274BE80004B8F, 0x0001EDB880004B93,
-    0x000181BA000185B9, 0x000185BC000181BB, 0x0001B1BE00018DBD,
-    0x000159B00001B1BF, 0x0001EDB2000159B1, 0x0001FDB40001E5B3,
-    0x0001E5B60001E5B5, 0x80004B970001D5B7, 0x0001A1B50001A9B6,
-    0x000E49B380004B9B, 0x0000388680004B9F, 0x000131BF0001DC87,
-    0x000121BD000129BE, 0x00018DBB000129BC, 0x80004AB300018DBA,
-    0x000E19A380004BA3, 0x80004BAB80004BA7, 0x80004BB380004BAF,
-    0x0001F1A50001F9A6, 0x0001DDAB80004BB7, 0x80004BBB0001DDAA,
-    0x000161AF80004BBF, 0x000171AD000179AE, 0x80004BC3000179AC,
-    0x80004BC7000FDCEF, 0x80004BCF80004BCB, 0x80004BD780004BD3,
-    0x80004BDF80004BDB, 0x80004BE780004BE3, 0x80004BEF80004BEB,
-    0x80004BF3000EE8E3, 0x00001180000E8CE1, 0x0000118200001181,
-    0x80004BF700024084, 0x80004BFF80004BFB, 0x0004C086000310BE,
-    0x80004C0700034487, 0x80004C0F80004C0B, 0x80004C1780004C13,
-    0x000224EF80004C1B, 0x80004C2380004C1F, 0x80004C2B80004C27,
-    0x80004C3380004C2F, 0x0004EC8480004C37, 0x80004C3F80004C3B,
-    0x0002ECE380004C43, 0x000138E180004C47, 0x000355B380004C4B,
-    0x80004C5380004C4F, 0x80004C5B80004C57, 0x00031DB500031DB6,
-    0x000309BB80004C5F, 0x80004C63000339BA, 0x0000FDBF80004C67,
-    0x0000F9BD0000FDBE, 0x000281A8000311BC, 0x00029DAA000289A9,
-    0x0002D5AC0002B1AB, 0x0002D5AE0002DDAD, 0x0001CD800002CDAF,
-    0x0000198200000981, 0x80004C6F80004C6B, 0x000570BE80004C77,
-    0x000159B880004C7B, 0x000169BA000159B9, 0x000179BC000169BB,
-    0x000169BE000179BD, 0x0002BDB0000165BF, 0x000285B200028DB1,
-    0x000179B400016DB3, 0x000169B6000179B5, 0x00042086000169B7,
-    0x80004C7F00050887, 0x80004C8780004C83, 0x80004C8F80004C8B,
-    0x0004EC8400005CEF, 0x80004C93000E5CE1, 0x80004C97000E8CE3,
-    0x80004C9F80004C9B, 0x000255A380004CA3, 0x80004CAB80004CA7,
-    0x80004CB380004CAF, 0x00021DA500021DA6, 0x000209AB80004CB7,
-    0x80004CBB000239AA, 0x0001FDAF80004CBF, 0x0001F9AD0001FDAE,
-    0x000669A8000211AC, 0x000679AA000669A9, 0x000669AC000679AB,
-    0x00069DAE000669AD, 0x80004C73000695AF, 0x80004CC780004CC3,
-    0x80004CCF80004CCB, 0x80004CD780004CD3, 0x0006F5B880004CDB,
-    0x00068DBA0006F9B9, 0x00069DBC000685BB, 0x000685BE000685BD,
-    0x0006EDB00006B5BF, 0x0006FDB20006F5B1, 0x0006EDB40006F5B3,
-    0x0006D1B60006D1B5, 0x0006F1B30006D1B7, 0x0000B58100001582,
-    0x80004CDF0000B580, 0x0006E1B50006E9B6, 0x000631BB0003D4BE,
-    0x80004CE3000629BA, 0x000615BF80004CE7, 0x000621BD000629BE,
-    0x80004CEB000629BC, 0x0000C8860006B5A3, 0x0006ADA60000C887,
-    0x80004CF380004CEF, 0x00066DAA0006A5A5, 0x80004CF7000675AB,
-    0x00066DAE80004CFB, 0x00066DAC000651AF, 0x000635A8000665AD,
-    0x000681AA000639A9, 0x000681AC000681AB, 0x000681AE000681AD,
-    0x80004CFF0006B5AF, 0x80004D0780004D03, 0x80004D0F80004D0B,
-    0x80004D1780004D13, 0x00069DB880004D1B, 0x0006A5BA0006ADB9,
-    0x000179BC000169BB, 0x000169BE000179BD, 0x0006D1B0000169BF,
-    0x0006D1B20006D1B1, 0x0006B5B40006D1B3, 0x0006B5B60006BDB5,
-    0x000609B30006ADB7, 0x80004D2380004D1F, 0x80004D2B80004D27,
-    0x000609B5000601B6, 0x000615BB80004D2F, 0x80004D33000615BA,
-    0x000679BF80004D37, 0x000605BD000671BE, 0x80004D3B000605BC,
-    0x80004D3F00064DA3, 0x000645A680004C03, 0x80004D4780004D43,
-    0x000651AA00064DA5, 0x80004D4B000651AB, 0x000635AE80004D4F,
-    0x000641AC00063DAF, 0x0003E981000641AD, 0x000388840003DD80,
-    0x003CB0860003E182, 0x000354BE00028887, 0x80004D5B80004D57,
-    0x80004D6380004D5F, 0x80004D6B80004D67, 0x80004D7380004D6F,
-    0x0006F0E380004D77, 0x000630E180004D7B, 0x80004D7F003C0084,
-    0x80004D8780004D83, 0x80004D8F80004D8B, 0x80004D93003D3485,
-    0x0007B0EF80004D97, 0x80004D9F80004D9B, 0x80004DA780004DA3,
-    0x80004DAF80004DAB, 0x80004DB3003CC4BE, 0x00019D8100019D82,
-    0x00020DA800019D80, 0x00027DAA000265A9, 0x000259AC000275AB,
-    0x000399AE000259AD, 0x0003E9B0000399AF, 0x0003F9B20003E9B1,
-    0x0003E9B40003F9B3, 0x00035DB60003E9B5, 0x00036DB8000355B7,
-    0x000375BA000375B9, 0x00035DBC000345BB, 0x000331BE000335BD,
-    0x80004DB7000329BF, 0x80004DBF80004DBB, 0x000360E180004DC7,
-    0x000244E30003F4EF, 0x80004DCF80004DCB, 0x0037F4EF00033CE3,
-    0x80004DD30001FCE1, 0x80004DDB80004DD7, 0x003F648680004DDF,
-    0x00213485003D6887, 0x80004DE3000399B3, 0x0003B5B60003BDB5,
-    0x80004DEB80004DE7, 0x000241BA80004DEF, 0x000241BC000241BB,
-    0x000241BE000241BD, 0x80004DF3000241BF, 0x80004DFB80004DF7,
-    0x80004E0380004DFF, 0x80004E0B80004E07, 0x003C68840001F8EF,
-    0x80004E0F00063CE1, 0x80004E1300061CE3, 0x80004E1B80004E17,
-    0x000315A380004E1F, 0x80004E2780004E23, 0x80004E2F80004E2B,
-    0x000331A5000339A6, 0x0002CDAB80004E37, 0x003E64BE0002CDAA,
-    0x0002CDAF80004E3B, 0x0002CDAD0002CDAE, 0x003E1DA80002CDAC,
-    0x003E2DAA003E25A9, 0x003E3DAC003E25AB, 0x003E2DAE003E25AD,
-    0x80004DC3003E25AF, 0x003FE581003FF582, 0x80004E33003FE580,
-    0x0004008680004E3F, 0x003E2DB800039C87, 0x003E31BA003E31B9,
-    0x003ED1BC003E31BB, 0x003ED1BE003ED1BD, 0x003E5DB0003ECDBF,
-    0x003E39B2003E21B1, 0x003E29B4003E39B3, 0x003E19B6003E29B5,
-    0x003EA5B3003E15B7, 0x80004E4780004E43, 0x80004E4F80004E4B,
-    0x003EB5B5003EA5B6, 0x003EE5BB80004E53, 0x80004E57003E99BA,
-    0x003EEDBF80004E5B, 0x003EEDBD003EEDBE, 0x80004E5F003EEDBC,
-    0x80004E63003EE1A3, 0x003EE1A680004E67, 0x80004E6F80004E6B,
-    0x003EDDAA003EF1A5, 0x80004E73003EA1AB, 0x003EA9AE80004E77,
-    0x003EA9AC003EA9AF, 0x0025058F003EA9AD, 0x80004E7B003E25B3,
-    0x003E29B680004E7F, 0x80004E8780004E83, 0x003E7DBA003E29B5,
-    0x80004E8B003E45BB, 0x003EB5BE80004E8F, 0x003E5DBC003EBDBF,
-    0x00387D9F003EBDBD, 0x0038F19D0039659E, 0x0035599B0034519C,
-    0x0030F1990035519A, 0x0031659700310D98, 0x002D59950030C196,
-    0x002CE193002D4D94, 0x002959910029D992, 0x0019A5B40028F190,
-    0x80004E930018DDB5, 0x001591B000000884, 0x001801B2001501B1,
-    0x00000D800019BDB3, 0x0003A5820003AD81, 0x000061A380004E97,
-    0x003D99A1003D1DA2, 0x0005C5A4003C71A0, 0x000801A6000441A5,
-    0x80004D530009E1A7, 0x0001E9A20001F5A1, 0x001001AC0005DDA3,
-    0x0011D9AE0011C5AD, 0x000865A8001039AF, 0x000DD9AA000C01A9,
-    0x002E318A000D09AB, 0x80004E9B0033A18B, 0x0033B98E80004E9F,
-    0x00321D8C0036118F, 0x0026258200324D8D, 0x000364BE002BA983,
-    0x002FA18600046084, 0x002A1984002E1587, 0x003E619A002A9185,
-    0x0004C886003EED9B, 0x80004EA70003DC87, 0x0003899C003E5CE1,
-    0x00366992003E00E3, 0x80004EAB00377993, 0x003B4D96003E18EF,
-    0x003BC194003FB997, 0x003D9DA9003A5D95, 0x003DB9AB003D89A8,
-    0x003DA9AD003DADAA, 0x003DC9AF003DA1AC, 0x0004E8BE003DA1AE,
-    0x80004EB380004EAF, 0x80004EBB80004EB7, 0x80004EC380004EBF,
-    0x003D59B980004EC7, 0x003D79BB003D51B8, 0x003D69BD003D65BA,
-    0x003D1DBF003D61BC, 0x003D81B1003D61BE, 0x003D69B3003DB9B0,
-    0x003D79B5003D89B2, 0x003D69B7003D71B4, 0x003C21A3003D71B6,
-    0x80004ECF80004ECB, 0x80004ED780004ED3, 0x003C2DA5003C2DA6,
-    0x003C41AB80004EDB, 0x80004EDF003C79AA, 0x003CB9AF80004EE3,
-    0x003CB9AD003CB1AE, 0x80004EE7003C59AC, 0x00037DB380004EEB,
-    0x0003D9B580004EEF, 0x80004EF780004EF3, 0x80004EFB0003D1B6,
-    0x0003C5BB80004EFF, 0x0003B9BD0003C5BA, 0x0003ADBF0003B5BC,
-    0x80004F030003B1BE, 0x80004F0B80004F07, 0x0000158000035CEF,
-    0x0000318200001D81, 0x0004EC84003E8CEF, 0x80004F0F003EFCE1,
-    0x80004F17003E1CE3, 0x80004F1B000194E1, 0x0003F5A3000020E3,
-    0x0004E88780004F1F, 0x80004F2300046C86, 0x000351A5000359A6,
-    0x00034DAB80004F27, 0x80004F2B00034DAA, 0x000325AF80004F2F,
-    0x000331AD000339AE, 0x80004EA300033DAC, 0x80004F3380004F13,
-    0x80004F3B80004F37, 0x80004F4380004F3F, 0x000671A880004F47,
-    0x00068DAA00064DA9, 0x00069DAC000685AB, 0x00068DAE000685AD,
-    0x0006FDB0000685AF, 0x000741B2000741B1, 0x000741B4000741B3,
-    0x000771B6000749B5, 0x000721B8000771B7, 0x000725BA000721B9,
-    0x000729BC000739BB, 0x00071DBE000715BD, 0x000625B30007FDBF,
-    0x80004F4F80004F4B, 0x80004F5780004F53, 0x000635B5000625B6,
-    0x0007A5BB80004F5B, 0x80004F5F000619BA, 0x0007ADBF80004F63,
-    0x0007B5BD0007A5BE, 0x80004F670007B5BC, 0x80004F6B000661A3,
-    0x000661A680004F6F, 0x80004F7780004F73, 0x00065DAA000671A5,
-    0x80004F7B0007E1AB, 0x0007E1AE80004F7F, 0x0007F1AC0007E9AF,
-    0x0006B1A80007F1AD, 0x0006CDAA0006BDA9, 0x0006CDAC0006DDAB,
-    0x0006F5AE0006FDAD, 0x0001F980000115AF, 0x0001C5820001CD81,
-    0x00009086000264BE, 0x80004F8700010087, 0x0001D1B880004F8B,
-    0x0001E1BA0001D9B9, 0x000191BC0001E1BB, 0x000195BE00019DBD,
-    0x00016DB0000189BF, 0x00017DB2000175B1, 0x00016DB4000175B3,
-    0x0001F1B60001F1B5, 0x000645B30001F1B7, 0x80004F8F80004F83,
-    0x80004F9780004F93, 0x000645B500067DB6, 0x0001B1BB80004F9B,
-    0x80004F9F0001A9BA, 0x000135BF80004FA3, 0x0001A1BD0001A9BE,
-    0x80004FA70001A9BC, 0x80004FAB000601A3, 0x000639A680004FAF,
-    0x80004FB780004FB3, 0x0001EDAA000601A5, 0x80004FBB0001F5AB,
-    0x0001EDAE80004FBF, 0x0001EDAC000171AF, 0x000128E10001E5AD,
-    0x000250E380004FC3, 0x80004FCB80004FC7, 0x80004FD380004FCF,
-    0x80004FDB80004FD7, 0x80004FE380004FDF, 0x0000718180004FE7,
-    0x00029CEF00001980, 0x80004FEB00007982, 0x0002C88480004FEF,
-    0x80004FF70003C5B3, 0x0003C5B60003C5B5, 0x000CC086000310BE,
-    0x0003A9BA000C4487, 0x0003BDBC0003A5BB, 0x0003A1BE0003A1BD,
-    0x000611AE000395BF, 0x000601AC000619AF, 0x0006A5AA000601AD,
-    0x003945A8000611AB, 0x80004FFB003971A9, 0x8000500380004FFF,
-    0x8000500B80005007, 0x800050138000500F, 0x0007EDBE80005017,
-    0x0007D1BC00074DBF, 0x0007DDBA0007E5BD, 0x000635B80007C1BB,
-    0x000639B60007D5B9, 0x000625B400060DB7, 0x000631B2000639B5,
-    0x000651B000063DB3, 0x000239A8000651B1, 0x000281AA000239A9,
-    0x000281AC000281AB, 0x0002B1AE000289AD, 0x000D6C840002B1AF,
-    0x000D60BE8000501B, 0x800050238000501F, 0x8000502B80005027,
-    0x00014DB88000502F, 0x00015DBA000155B9, 0x00014DBC000155BB,
-    0x00017DBE000175BD, 0x0002A1B0000175BF, 0x0002A5B20002ADB1,
-    0x0002A9B40002B9B3, 0x000295B600029DB5, 0x000654E1000179B7,
-    0x000638E30007D4E1, 0x800050330007B0E3, 0x8000503B80005037,
-    0x000CE4848000503F, 0x8000504780005043, 0x8000504F8000504B,
-    0x00006CEF80005053, 0x000249A30007E4EF, 0x8000505B80005057,
-    0x800050678000505F, 0x000249A5000249A6, 0x000229AB8000506B,
-    0x000C4886000225AA, 0x000219AF000CDC87, 0x00022DAD00022DAE,
-    0x000E51A8000231AC, 0x000EADAA000EA5A9, 0x000EBDAC000EA5AB,
-    0x000EADAE000EA5AD, 0x000FE580000EA5AF, 0x000FE582000FED81,
-    0x8000506F80005063, 0x8000507780005073, 0x000FD5B88000507B,
-    0x000FD5BA000FDDB9, 0x000FF9BC000FE9BB, 0x000FE9BE000FF9BD,
-    0x000EDDB0000FE9BF, 0x000F45B2000F41B1, 0x000F45B4000F5DB3,
-    0x000F45B6000F4DB5, 0x000E25B3000FEDB7, 0x800050838000507F,
-    0x8000508B80005087, 0x000E35B5000E25B6, 0x000F85BB8000508F,
-    0x80005093000E19BA, 0x000F89BF80005097, 0x000F89BD000F81BE,
-    0x8000509B000F91BC, 0x8000509F000E61A3, 0x000E61A6800050A3,
-    0x800050AB800050A7, 0x000E5DAA000E71A5, 0x800050AF000FC1AB,
-    0x000FC5AE800050B3, 0x000FD5AC000FCDAF, 0x000ED1A8000FCDAD,
-    0x000139AA000ED9A9, 0x000129AC000139AB, 0x00019DAE000129AD,
-    0x800050B7000195AF, 0x800050BF800050BB, 0x00000D80800050C3,
-    0x0000BD820000B581, 0x000195B8800050C7, 0x0001A1BA00019DB9,
-    0x000071BC0001A1BB, 0x000071BE000071BD, 0x0001EDB0000071BF,
-    0x0001F5B20001F5B1, 0x0001DDB40001C5B3, 0x0001B1B60001B5B5,
-    0x800050CB0001B1B7, 0x0001B9B3800050CF, 0x0001A9B50002C0BE,
-    0x800050D7800050D3, 0x000180860001A1B6, 0x000139BB0001C487,
-    0x000119BD000121BA, 0x000179BF000129BC, 0x0001F1A3000111BE,
-    0x80004FF3800050DB, 0x800050E3800050DF, 0x0001E1A50001E9A6,
-    0x000171AB800050E7, 0x800050EB000169AA, 0x000131AF800050EF,
-    0x000151AD000159AE, 0x800050F3000161AC, 0x800050FB800050F7,
-    0x0001A8E1800050FF, 0x000240E380005103, 0x001528BF80005107,
-    0x000274EF8000510B, 0x800051138000510F, 0x8000511B80005117,
-    0x800051238000511F, 0x00038884000F74E3, 0x80005127000E34E1,
-    0x8000512F8000512B, 0x00000D8080005133, 0x0000118200001581,
-    0x8000513B80005137, 0x8000513F000FA4EF, 0x000319A880005147,
-    0x000345AA000341A9, 0x00034DAC00035DAB, 0x000375AE00037DAD,
-    0x0015A48500009DAF, 0x0004088600037CBE, 0x8000514B00031C87,
-    0x800051538000514F, 0x0000F5B880005157, 0x0000F5BA0000FDB9,
-    0x000081BC00008DBB, 0x000081BE000081BD, 0x0000E5B0000081BF,
-    0x0000E5B20000EDB1, 0x0000EDB40000F9B3, 0x0000D5B60000D1B5,
-    0x8000515B0000CDB7, 0x800051638000515F, 0x800051670002C1B3,
-    0x0002F1B60002C9B5, 0x8000516F8000516B, 0x00012DBA80005173,
-    0x00013DBC000125BB, 0x000125BE000121BD, 0x0002A9A6000119BF,
-    0x8000517700024884, 0x8000517B000291A5, 0x8000517F000299A3,
-    0x00017DAE80005183, 0x000165AC000141AF, 0x000175AA000179AD,
-    0x8000518700017DAB, 0x8000518F8000518B, 0x8000519780005193,
-    0x0000E4EF8000519B, 0x800051A38000519F, 0x000E10E3800051A7,
-    0x000E58E1800051AB, 0x00002D80800051AF, 0x0000158200001181,
-    0x000175B30005B0BE, 0x00051487800051B7, 0x800051BB00046C86,
-    0x000165B50000D5B6, 0x0000FDBB800051BF, 0x800051C30000F5BA,
-    0x0000A1BF800051C7, 0x0000DDBD0000BDBE, 0x00067DA80000DDBC,
-    0x0006BDAA0006B5A9, 0x0006ADAC0006B5AB, 0x00071DAE000715AD,
-    0x800051B3000715AF, 0x800051CF800051CB, 0x800051D7800051D3,
-    0x800051DF800051DB, 0x000739B8800051E3, 0x0007C9BA000739B9,
-    0x0007D9BC0007C9BB, 0x0007CDBE0007D9BD, 0x000771B00007C5BF,
-    0x000749B2000779B1, 0x000739B4000749B3, 0x000721B6000725B5,
-    0x000635A3000721B7, 0x800051EB800051E7, 0x800051F3800051EF,
-    0x000625A5000795A6, 0x0007BDAB800051F7, 0x800051FB0007B5AA,
-    0x0007E1AF800051FF, 0x00079DAD0007FDAE, 0x8000520300079DAC,
-    0x8000520B80005207, 0x800052138000520F, 0x00003D8100003D82,
-    0x8000521700003D80, 0x8000521F8000521B, 0x0003A4BE0003A084,
-    0x0000F88680005223, 0x0006C5A80000E087, 0x0006D5AA0006D5A9,
-    0x0006FDAC0006E5AB, 0x000131AE000131AD, 0x80005227000131AF,
-    0x8000522F8000522B, 0x8000523780005233, 0x8000523F8000523B,
-    0x0001DDB880005243, 0x0001E5BA0001EDB9, 0x000195BC00018DBB,
-    0x000195BE00019DBD, 0x000151B000018DBF, 0x000151B2000151B1,
-    0x0001F5B4000151B3, 0x0001F5B60001FDB5, 0x00065DB30001EDB7,
-    0x8000524B80005247, 0x800052538000524F, 0x000675B50001B1B6,
-    0x000195BB80005257, 0x8000525B000195BA, 0x000139BF8000525F,
-    0x000185BD000131BE, 0x00062DA5000185BC, 0x8000526780005263,
-    0x8000526B0001E9A6, 0x000605A38000526F, 0x0001DDAD80005273,
-    0x000161AF0001DDAC, 0x80005277000169AE, 0x0001CDAB80005143,
-    0x8000527B0001CDAA, 0x0003C4848000527F, 0x8000528300F400BF,
-    0x00003D8100003D82, 0x8000528700003D80, 0x8000528F8000528B,
-    0x80005297000360BE, 0x8000529F8000529B, 0x001CE086800052A3,
-    0x00070CEF00030087, 0x800052AB800052A7, 0x800052B3800052AF,
-    0x800052BB800052B7, 0x800052C3800052BF, 0x0006C0E1800052C7,
-    0x00072CE3800052CB, 0x800052D3800052CF, 0x800052DB800052D7,
-    0x800052E3800052DF, 0x800052EB800052E7, 0x000381A8800052EF,
-    0x000381AA000381A9, 0x000381AC000381AB, 0x000381AE000381AD,
-    0x000345B0000381AF, 0x000345B200034DB1, 0x00034DB400035DB3,
-    0x000375B600037DB5, 0x000329B8000319B7, 0x00033DBA000335B9,
-    0x000301BC000331BB, 0x0000FDBE0000F5BD, 0x000229B30000EDBF,
-    0x800052F7800052F3, 0x800052FF800052FB, 0x000229B5000221B6,
-    0x0002A9BB001D5084, 0x800053070002A1BA, 0x000299BF8000530B,
-    0x0002A9BD0002A9BE, 0x00004D810002B1BC, 0x00039CEF00004D80,
-    0x001CF08600005D82, 0x001D10BE001D7887, 0x800053138000530F,
-    0x8000531B80005317, 0x000190E18000531F, 0x000360E380005323,
-    0x8000532B80005327, 0x800053338000532F, 0x8000533B80005337,
-    0x800053438000533F, 0x001CEC84000194EF, 0x80005347000670E1,
-    0x8000534B000150E3, 0x800053538000534F, 0x0002E9A380005357,
-    0x8000535F8000535B, 0x8000536780005363, 0x0002E9A50002E1A6,
-    0x000269AB8000536B, 0x8000536F000261AA, 0x000259AF001CA8BE,
-    0x000269AD000269AE, 0x001E31A8000271AC, 0x001E31AA001E31A9,
-    0x001E5DAC001E31AB, 0x001E4DAE001E45AD, 0x80005303001E45AF,
-    0x001FF581001FCD82, 0x80005373001FFD80, 0x00001C8680005377,
-    0x001EC5B80003F887, 0x001EC5BA001ECDB9, 0x001EC5BC001EDDBB,
-    0x001EC5BE001ECDBD, 0x001E3DB0001F59BF, 0x001E0DB2001E05B1,
-    0x001E1DB4001E05B3, 0x001E05B6001E01B5, 0x001EB9B3001EFDB7,
-    0x8000537F8000537B, 0x8000538780005383, 0x001ED5B5001EC5B6,
-    0x001EC1BB8000538B, 0x8000538F001EF9BA, 0x001EC5BF80005393,
-    0x001ED1BD001ED9BE, 0x80005397001ED9BC, 0x8000539B001EFDA3,
-    0x001E81A68000539F, 0x800053A7800053A3, 0x001EBDAA001E91A5,
-    0x800053AB001E85AB, 0x001E9DAE800053AF, 0x001E9DAC001E81AF,
-    0x001E29A8001E95AD, 0x001E55AA001E29A9, 0x001E75AC001E6DAB,
-    0x001E75AE001E7DAD, 0x800053B3001E69AF, 0x800053BB800053B7,
-    0x800053C3800053BF, 0x800053CB800053C7, 0x001EE9B8800053CF,
-    0x001EFDBA001EF5B9, 0x001EEDBC001EF5BB, 0x001E91BE001E91BD,
-    0x001E1DB0001E91BF, 0x001EEDB2001EE5B1, 0x001EFDB4001EE5B3,
-    0x001EEDB6001EE5B5, 0x001EDDB3001EE5B7, 0x800053D7800053D3,
-    0x800053DF800053DB, 0x001EFDB5001EFDB6, 0x001E21BB00015884,
-    0x000028BE001ED9BA, 0x001E21BF800053E7, 0x001E31BD001E39BE,
-    0x00004D81001E39BC, 0x001E95A300004D80, 0x001EB5A500005D82,
-    0x800053EB800053E3, 0x80005293001EB5A6, 0x001E69AB800053EF,
-    0x001E79AD001E91AA, 0x001E69AF001E71AC, 0x00040086001E71AE,
-    0x000281B300034487, 0x000281B5800053F3, 0x800053FB800053F7,
-    0x00000088000281B6, 0x0002E1BB800053FF, 0x0002E5BD0002EDBA,
-    0x0002D1BF0002F9BC, 0x800054030002E5BE, 0x0003448480005407,
-    0x000248E100038CBE, 0x000200E38000540B, 0x8000540F001FFCEF,
-    0x0002DCEF001E3CE1, 0x80005413001F60E3, 0x8000541B80005417,
-    0x000205A98000541F, 0x000205AB0002B1A8, 0x00020DAD000205AA,
-    0x000235AF000205AC, 0x00056C84000235AE, 0x8000542780005423,
-    0x8000542F8000542B, 0x8000543780005433, 0x0003E9B98000543B,
-    0x0003F9BB0003E1B8, 0x0003E9BD0003E1BA, 0x00035DBF0003E1BC,
-    0x000229B10003E1BE, 0x00023DB3000225B0, 0x000219B5000221B2,
-    0x0003D9B700022DB4, 0x8000543F000211B6, 0x8000544780005443,
-    0x8000544B000385A3, 0x000385A6000385A5, 0x800054578000544F,
-    0x0003E9AA8000545B, 0x0003FDAC0003E5AB, 0x0003E1AE0003E1AD,
-    0x000049800003D5AF, 0x0000558200005581, 0x000498BE0002A9A3,
-    0x000141A6000141A5, 0x0005E0868000545F, 0x00012DAA0005E087,
-    0x000111AC000139AB, 0x00010DAE000105AD, 0x80005463000105AF,
-    0x8000546B80005467, 0x8000546F0000DCEF, 0x8000547780005473,
-    0x00072C84001E88EF, 0x8000547B001EECE1, 0x8000547F001E1CE3,
-    0x80005483000194E1, 0x000225B3000030E3, 0x8000548700E66085,
-    0x8000548F8000548B, 0x0001CDB50001CDB6, 0x0001B5BB80005493,
-    0x800054970001A1BA, 0x000189BF8000549B, 0x000189BD000181BE,
-    0x8000545300019DBC, 0x800054A38000549F, 0x800054AB800054A7,
-    0x800054B3800054AF, 0x000701A8800054B7, 0x000775AA000701A9,
-    0x0007B5AC0007ADAB, 0x0007A9AE0007BDAD, 0x0007D9B00007A9AF,
-    0x0007F9B20007EDB1, 0x000799B40007F5B3, 0x000789B6000799B5,
-    0x000789B8000781B7, 0x00006DBA000789B9, 0x00007DBC000065BB,
-    0x00006DBE000065BD, 0x00000981000065BF, 0x800054BB00009980,
-    0x800054BF00001D82, 0x800054C7800054C3, 0x00050DA8800054CB,
-    0x00054DAA00057DA9, 0x0006A5AC0006A1AB, 0x00069DAE0006ADAD,
-    0x000685B00006FDAF, 0x0006A9B2000691B1, 0x0006A5B40006BDB3,
-    0x0006A5B60006ADB5, 0x000649B8000679B7, 0x000659BA000649B9,
-    0x000649BC000659BB, 0x0007F9BE000649BD, 0x00065DB30007F9BF,
-    0x00022886800054CF, 0x800054D300012C87, 0x00069DB500069DB6,
-    0x000689BB800054D7, 0x800054DB0006BDBA, 0x0006FDBF800054DF,
-    0x000681BD0006FDBE, 0x800054E300068DBC, 0x800054E7000619A3,
-    0x0006D9A6800054EB, 0x800054F3800054EF, 0x0006F9AA0006D9A5,
-    0x800054F70006CDAB, 0x0006B9AE800054FB, 0x0006C9AC0006B9AF,
-    0x000119A80006C5AD, 0x00008DAA000119A9, 0x00008DAC00009DAB,
-    0x0000B5AE0000BDAD, 0x800054FF0000DDAF, 0x8000550780005503,
-    0x8000550F8000550B, 0x8000551780005513, 0x000069B88000551B,
-    0x000079BA000069B9, 0x000069BC000079BB, 0x0003DDBE000069BD,
-    0x0000A9B00003D5BF, 0x0000BDB20000A9B1, 0x000099B40000B5B3,
-    0x000059B6000099B5, 0x00022CBE000059B7, 0x800055238000551F,
-    0x8000552B80005527, 0x800055378000552F, 0x00002D808000553B,
-    0x00003D8200003581, 0x000C90868000553F, 0x800055430002E887,
-    0x000345B380005547, 0x8000554F8000554B, 0x8000555780005553,
-    0x000345B500037DB6, 0x00032DBB8000555B, 0x8000555F00032DBA,
-    0x00031DBF80005563, 0x000321BD00031DBE, 0x0003CDAB000329BC,
-    0x800055670003CDAA, 0x0003FDAF8000556B, 0x0003C1AD0003FDAE,
-    0x0003A5A30003C9AC, 0x800055738000556F, 0x8000557B80005577,
-    0x0003A5A500039DA6, 0x800055838000557F, 0x8000558B80005587,
-    0x800055938000558F, 0x00003D8100003D82, 0x8000559700003D80,
-    0x8000559F8000559B, 0x000CD08600036084, 0x800055A700033087,
-    0x000244BE800055AB, 0x800055B3800055AF, 0x000000E1800055B7,
-    0x0006A8E3800055BB, 0x000CEC84800055BF, 0x0003D0EF800055C3,
-    0x800055CB800055C7, 0x800055D3800055CF, 0x00016DB3800055D7,
-    0x800055DF800055DB, 0x800055E7800055E3, 0x000165B5000111B6,
-    0x00013DBB800055EB, 0x800055EF000135BA, 0x0001FDBF800055F3,
-    0x000115BD0001FDBE, 0x800055F7000115BC, 0x800055FB0006FCE1,
-    0x800055FF00073CE3, 0x8000560780005603, 0x8000560F8000560B,
-    0x80005617000C6CBE, 0x8000561F8000561B, 0x8000562780005623,
-    0x00005D81000614EF, 0x0001E5A300005D80, 0x0001EDA500006D82,
-    0x8000562F8000562B, 0x000CA887000199A6, 0x0001B5AB000CE486,
-    0x00019DAD0001BDAA, 0x000175AF00019DAC, 0x000E19A8000175AE,
-    0x000E25AA000E19A9, 0x000E2DAC000E35AB, 0x000E51AE000E51AD,
-    0x800055A3000E51AF, 0x8000563380005613, 0x8000563B80005637,
-    0x800056438000563F, 0x000EEDB880005647, 0x000EF5BA000EF5B9,
-    0x000E95BC000E8DBB, 0x000E95BE000E9DBD, 0x000E31B0000E8DBF,
-    0x000E01B2000E39B1, 0x000EF9B4000E01B3, 0x000EDDB6000EF9B5,
-    0x000E79A8000ED5B7, 0x000F8DAA000E79A9, 0x000F9DAC000F85AB,
-    0x000F85AE000F81AD, 0x8000564B000FB9AF, 0x800056538000564F,
-    0x8000565B80005657, 0x800056638000565F, 0x000F91B880005667,
-    0x000FA1BA000F99B9, 0x000F51BC000FA1BB, 0x000F49BE000F5DBD,
-    0x000FCDB0000F49BF, 0x000FDDB2000FD5B1, 0x000FCDB4000FD5B3,
-    0x000FB1B6000FB1B5, 0x000E05B3000FB1B7, 0x8000566F8000566B,
-    0x8000567780005673, 0x000E15B5000E01B6, 0x000E11BB8000567B,
-    0x8000567F000E09BA, 0x000E75BF0001A084, 0x000E01BD000E09BE,
-    0x00006982000E09BC, 0x00005980000E41A3, 0x000E45A600005181,
-    0x80005683000158BE, 0x000E4DAA000E51A5, 0x0000C886000E55AB,
-    0x000E4DAE0001AC87, 0x000E4DAC000E31AF, 0x80005687000E45AD,
-    0x8000568F8000568B, 0x8000569780005693, 0x800055338000569B,
-    0x000E09A88000569F, 0x000E19AA000E09A9, 0x000E09AC000E19AB,
-    0x000E61AE000E61AD, 0x0001EDB0000195AF, 0x0001FDB20001F5B1,
-    0x0001EDB40001F5B3, 0x00017DB6000175B5, 0x00014DB8000175B7,
-    0x00015DBA000155B9, 0x00014DBC000155BB, 0x0000F1BE0000F1BD,
-    0x800056A30000F1BF, 0x800056AB800056A7, 0x800056B3800056AF,
-    0x800056B7000EE8E3, 0x000400BE000E34E1, 0x800056BB000FDCEF,
-    0x800056C3800056BF, 0x800056CB800056C7, 0x000DF1B3800056CF,
-    0x800056D7800056D3, 0x800056DF800056DB, 0x000DE1B5000D81B6,
-    0x000251BB800056E3, 0x800056E7000249BA, 0x0002D1BF800056EB,
-    0x000241BD000249BE, 0x000D31A3000249BC, 0x00038884800056EF,
-    0x800056FB800056F7, 0x000D21A5000D41A6, 0x000291AB800056FF,
-    0x80005703000289AA, 0x000211AF80005707, 0x000281AD000289AE,
-    0x0000A980000289AC, 0x00004D820000A981, 0x001264E18000570B,
-    0x000B08E300024CE3, 0x8000570F0001ACE1, 0x001618EF00023CEF,
-    0x0006508680005713, 0x8000571700034887, 0x000281A88000571B,
-    0x000281AA000281A9, 0x000281AC000281AB, 0x000285AE000281AD,
-    0x8000571F00011DAF, 0x8000572780005723, 0x8000572F8000572B,
-    0x8000573780005733, 0x000175B80005B884, 0x000175BA00017DB9,
-    0x000119BC000109BB, 0x000109BE000119BD, 0x000165B0000101BF,
-    0x000165B200016DB1, 0x000169B400017DB3, 0x000155B600015DB5,
-    0x8000573B00014DB7, 0x800057438000573F, 0x8000574B80005747,
-    0x000034EF8000574F, 0x000E5CE1000EC0EF, 0x000094E3000F4CE1,
-    0x80005753000E9CE3, 0x00007D8100006582, 0x8000575700007D80,
-    0x0007C4BE8000575B, 0x8000575F000245B3, 0x00024DB6000245B5,
-    0x0006408680005767, 0x0001EDBA00049087, 0x0001FDBC0001E5BB,
-    0x0001E1BE0001E5BD, 0x0008959F0001D5BF, 0x000DDD9D0008059E,
-    0x000C319B000C3D9C, 0x0010ED99000DF59A, 0x0011D59700103D98,
-    0x001405950015B196, 0x0019B5930015E594, 0x0018399100183192,
-    0x001C618F001C0D90, 0x0006D5B3800056F3, 0x0006FDB58000576B,
-    0x8000576F80005763, 0x80005773000691B6, 0x000695BB80005777,
-    0x000755BD000695BA, 0x000755BF000755BC, 0x8000577B000755BE,
-    0x00068DAA8000577F, 0x00069DAC000685AB, 0x000681AE000685AD,
-    0x0000A8840006B5AF, 0x8000578780005783, 0x8000578B000525A3,
-    0x00052DA6000525A5, 0x800057938000578F, 0x8000579B80005797,
-    0x800057A38000579F, 0x800057AB800057A7, 0x800057B3800057AF,
-    0x800057BB800057B7, 0x0005A9A3800057BF, 0x0004D9A10004D1A2,
-    0x001D2182000565A0, 0x800057C3001D3583, 0x00189586800057C7,
-    0x001919840014DD87, 0x00150D8A00191985, 0x800057CB0014B18B,
-    0x0011C58E800057CF, 0x0010CD8C000CD58F, 0x000D619200111D8D,
-    0x00004CBE000D7593, 0x0009C596800057D7, 0x000C8D940004CD97,
-    0x0005419A00095D95, 0x00FFC8860005419B, 0x0000598100004C87,
-    0x0004219C00007980, 0x0001C5A100005182, 0x00FC01A3800057DB,
-    0x00FD15A50001D9A2, 0x00F9B9A700FDB5A4, 0x00F809A900F801A6,
-    0x00F505AB00F995A8, 0x00F1B1AD00F5ADAA, 0x00F01DAF00F001AC,
-    0x00EDE1B100F1ADAE, 0x00E801B300EC01B0, 0x00E95DB500ECFDB2,
-    0x800057DF00E9F5B4, 0x800057E7800057E3, 0x800057EF800057EB,
-    0x800057F7800057F3, 0x800057FF800057FB, 0x00048DA880005803,
-    0x000495AA000495A9, 0x0004BDAC0004A5AB, 0x0004F9AE0004D9AD,
-    0x00FC6C840004F1AF, 0x8000580B80005807, 0x800058138000580F,
-    0x8000581B80005817, 0x000579B88000581F, 0x0005CDBA0005C5B9,
-    0x0005DDBC0005C5BB, 0x0005CDBE0005C5BD, 0x000499B00005EDBF,
-    0x000549B2000499B1, 0x000559B4000549B3, 0x000549B6000559B5,
-    0x00041DA3000549B7, 0x8000582300FCCCBE, 0x8000582B80005827,
-    0x000435A5000459A6, 0x00045DAB8000582F, 0x8000583300045DAA,
-    0x00059DAF80005837, 0x00059DAD00059DAE, 0x8000583B00059DAC,
-    0x8000583F00029DB3, 0x000251B680005847, 0x8000584F8000584B,
-    0x000249BA0002B9B5, 0x00FD4886000255BB, 0x000235BE00FCC087,
-    0x000245BC00023DBF, 0x0004DDA800023DBD, 0x000351AA000349A9,
-    0x000375AC00036DAB, 0x00039DAE000395AD, 0x00018D800003B9AF,
-    0x0001E1820001E581, 0x8000585780005853, 0x8000585F8000585B,
-    0x000395B880005863, 0x000395BA00039DB9, 0x0003B5BC0003ADBB,
-    0x0003B5BE0003BDBD, 0x0003C9B0000255BF, 0x0003D5B20003D5B1,
-    0x0003B5B40003ADB3, 0x0003B5B60003BDB5, 0x800058670003ADB7,
-    0x0003D1A38000586B, 0x0003F5A58000586F, 0x8000587780005873,
-    0x8000587B00031DA6, 0x000319AB8000587F, 0x000371AD000305AA,
-    0x000371AF000309AC, 0x000728E1000379AE, 0x0006E4E380005883,
-    0x000628E180005887, 0x000168E38000588B, 0x800058938000588F,
-    0x000058EF80005897, 0x8000589F8000589B, 0x0006C8EF800058A3,
-    0x00FD4DA8800058A7, 0x00FDD1AA00FD81A9, 0x00FDF1AC00FDD1AB,
-    0x00FDF1AE00FDF1AD, 0x8000584300FDF1AF, 0x0000198100001182,
-    0x800058AB00FFD180, 0x00037484800058AF, 0x00FE75B8000374BE,
-    0x00FE75BA00FE7DB9, 0x00FEDDBC00FEC5BB, 0x00FEC5BE00FEC5BD,
-    0x00FD91B000FEF5BF, 0x00FD91B200FD91B1, 0x00FE55B400FD91B3,
-    0x00FE55B600FE5DB5, 0x00FD59B300FE4DB7, 0x0000AC86800058B3,
-    0x800058B700034487, 0x00FD79B500FD71B6, 0x00FD55BB800058BB,
-    0x800058BF00FD55BA, 0x00FE99BF800058C3, 0x00FD45BD00FE91BE,
-    0x800058C700FD45BC, 0x800058CB00FD1DA3, 0x00FD35A6800058CF,
-    0x800058D7800058D3, 0x00FD11AA00FD3DA5, 0x800058DB00FD11AB,
-    0x00FED5AE800058DF, 0x00FD01AC00FEDDAF, 0x00FFCDA800FD01AD,
-    0x00FFD1AA00FFD1A9, 0x00FE21AC00FE21AB, 0x00FE21AE00FE21AD,
-    0x800058E300FE21AF, 0x800058EB800058E7, 0x800058F3800058EF,
-    0x800058FB800058F7, 0x00FE8DB8800058FF, 0x00FE95BA00FE91B9,
-    0x00FEB5BC00FEADBB, 0x00FEA9BE00FEB9BD, 0x00FEE1B000FEA9BF,
-    0x00FEE5B200FEE1B1, 0x00FEE9B400FEF9B3, 0x00FEDDB600FED5B5,
-    0x00FF59B300FEB9B7, 0x800057D380005903, 0x8000590B80005907,
-    0x00FEA9B500FEA1B6, 0x00FE25BB8000590F, 0x8000591300FE25BA,
-    0x00FE11BF80005917, 0x00FE2DBD00FE2DBE, 0x0000198200FE31BC,
-    0x0000658000FF1DA3, 0x00FEE5A600001981, 0x8000591F8000591B,
-    0x00FE61AA00FEEDA5, 0x0001648400FE61AB, 0x00FE69AE000020BE,
-    0x00FE75AC00FE55AF, 0x8000592700FE69AD, 0x8000592B00FE34E3,
-    0x0004B08600FD7CE1, 0x8000592F00030887, 0x8000593780005933,
-    0x000324848000593B, 0x8000593F00042484, 0x8000594300FE48EF,
-    0x0002E9B380005947, 0x000444BE8000594B, 0x8000594F000540BE,
-    0x000299B500029DB6, 0x000289BB80005953, 0x800059570002BDBA,
-    0x000359BF8000595B, 0x000359BD000351BE, 0x00021DA9000291BC,
-    0x000225AB000215A8, 0x000259AD000225AA, 0x00024DAF000255AC,
-    0x000664BE000251AE, 0x800059638000595F, 0x8000596B80005967,
-    0x800059738000596F, 0x0003F9B980005977, 0x00034DBB0003F1B8,
-    0x000341BD0003F1BA, 0x000371BF000359BC, 0x000209B1000341BE,
-    0x0003C9B300023DB0, 0x0003D9B5000201B2, 0x0003C9B70003D1B4,
-    0x000190E10003D1B6, 0x0000F0E38000597B, 0x800059838000597F,
-    0x00003D8100003D82, 0x8000598700003D80, 0x8000598F8000598B,
-    0x8000599B80005997, 0x00002CEF8000599F, 0x0003E9A3800059A3,
-    0x0004E886800059A7, 0x800059AB00056087, 0x000399A500039DA6,
-    0x000389AB800059AF, 0x800059B30003BDAA, 0x000259AF800059B7,
-    0x000259AD000251AE, 0x800059BB000391AC, 0x800059C3800059BF,
-    0x800059CB800059C7, 0x800059D3800059CF, 0x0004A084000158EF,
-    0x800059D700FF54E1, 0x800059DB000184E3, 0x800059E3800059DF,
-    0x0001D9B3800059E7, 0x0019CC85800059EB, 0x800059F3800059EF,
-    0x0001F9B50001F1B6, 0x0001A5BB800059F7, 0x800059FB0001D9BA,
-    0x00019DBF800059FF, 0x0001A5BD00019DBE, 0x000601A80001ADBC,
-    0x000611AA00060DA9, 0x000631AC000611AB, 0x000629AE000631AD,
-    0x80005993000625AF, 0x0007C1810007C982, 0x80005A030007F180,
-    0x00001C8680005A07, 0x0006EDB80003FC87, 0x0006FDBA0006F5B9,
-    0x0006EDBC0006F5BB, 0x000755BE000751BD, 0x00065DB000074DBF,
-    0x000639B2000621B1, 0x000619B4000631B3, 0x0006DDB6000619B5,
-    0x000699A30006D5B7, 0x80005A0F80005A0B, 0x80005A1780005A13,
-    0x0006B9A50006B1A6, 0x0006E5AB80005A1B, 0x80005A1F000699AA,
-    0x0006DDAF80005A23, 0x0006E5AD0006DDAE, 0x80005A270006EDAC,
-    0x80005A2B0007F1B3, 0x000781B680005A2F, 0x80005A3780005A33,
-    0x00078DBA000799B5, 0x80005A3B000765BB, 0x00077DBE80005A3F,
-    0x00077DBC000765BF, 0x00062DA8000775BD, 0x00063DAA000635A9,
-    0x000655AC000631AB, 0x000645AE000645AD, 0x80005A43000675AF,
-    0x80005A4B80005A47, 0x80005A5380005A4F, 0x80005A5B80005A57,
-    0x0006E9B880005A5F, 0x0006FDBA0006F5B9, 0x0006EDBC0006F5BB,
-    0x000695BE000691BD, 0x00060DB000068DBF, 0x0006EDB20006E5B1,
-    0x0006FDB40006E5B3, 0x0006EDB60006E5B5, 0x0006B5A30006E5B7,
-    0x80005A6780005A63, 0x80005A6F80005A6B, 0x0006DDA50006C5A6,
-    0x000621AB80005923, 0x80005A730006C9AA, 0x000621AF80005A77,
-    0x000631AD000639AE, 0x00004980000639AC, 0x0000598200004981,
-    0x80005A7B000145B3, 0x000145B6000145B5, 0x0000408680005A7F,
-    0x000129BA00006487, 0x00013DBC000125BB, 0x000121BE000121BD,
-    0x80005A83000115BF, 0x0003048480005A87, 0x000628E3000C00BE,
-    0x000610E10002E0EF, 0x000690EF80005A8B, 0x80005A8F00023CE3,
-    0x80005A930001D4E1, 0x80005A9B80005A97, 0x80005AA380005A9F,
-    0x0002C1A380005AA7, 0x0002C1A580005AAB, 0x80005AB380005AAF,
-    0x80005AB70002C1A6, 0x0002A1AB80005ABB, 0x0002A5AD0002ADAA,
-    0x000291AF0002B9AC, 0x000205A90002A5AE, 0x000205AB0002B1A8,
-    0x00020DAD000205AA, 0x000235AF000205AC, 0x000CA884000235AE,
-    0x80005AC380005ABF, 0x80005ACB80005AC7, 0x80005AD380005ACF,
-    0x0003E9B980005AD7, 0x0003F9BB0003E1B8, 0x0003E9BD0003E1BA,
-    0x00035DBF0003E1BC, 0x000229B10003E1BE, 0x00023DB3000225B0,
-    0x000219B5000221B2, 0x0003D9B700022DB4, 0x0002ADA8000211B6,
-    0x0002D5AA0002D5A9, 0x000115AC00010DAB, 0x000131AE000109AD,
-    0x80005ADB00012DAF, 0x80005AE780005ADF, 0x80005AEF80005AEB,
-    0x80005AF780005AF3, 0x000121B880005AFB, 0x0001EDBA000121B9,
-    0x0001FDBC0001E5BB, 0x0001EDBE0001E5BD, 0x000155B00001E5BF,
-    0x000155B200015DB1, 0x000135B400012DB3, 0x00012DB6000139B5,
-    0x00013D80000125B7, 0x0000AD820000A581, 0x80005AFF0007D4EF,
-    0x80005B0780005B03, 0x000C2CBE000728EF, 0x80005B0B0006F4E1,
-    0x80005B0F000790E3, 0x80005B13000194E1, 0x00021DB300060CE3,
-    0x000D448780005B17, 0x80005B1B000D4C86, 0x0001DDB50001C9B6,
-    0x0001D9BB80005B1F, 0x80005B230001D1BA, 0x0001BDBF80005B27,
-    0x0001B9BD0001B1BE, 0x80005AE30001C1BC, 0x80005B2F80005B2B,
-    0x80005B3780005B33, 0x80005B3F80005B3B, 0x000F09A880005B43,
-    0x000F19AA000F09A9, 0x000F09AC000F19AB, 0x000FA9AE000F09AD,
-    0x000FD9B0000FA9AF, 0x000FF9B2000FEDB1, 0x000F95B4000FF5B3,
-    0x000F85B6000F85B5, 0x000F8DB8000FB5B7, 0x000061BA000061B9,
-    0x000061BC000061BB, 0x000061BE000061BD, 0x000D5DA3000061BF,
-    0x0000158100002D82, 0x80005B4700001D80, 0x000E9DA5000E89A6,
-    0x000E99AB80005B4B, 0x80005B4F000E91AA, 0x000EFDAF80005B53,
-    0x000EF9AD000EF1AE, 0x80005B57000E81AC, 0x0003E886000FF5B3,
-    0x000FEDB60003BC87, 0x80005B5F80005B5B, 0x000F79BA000FE5B5,
-    0x80005B63000F4DBB, 0x000F35BE80005B67, 0x000F49BC000F25BF,
-    0x000E31A3000F45BD, 0x80005B6F80005B6B, 0x80005B7780005B73,
-    0x000E21A5000E29A6, 0x000E89AB80005B7B, 0x80005B7F000EBDAA,
-    0x000EE1AF80005B83, 0x000E81AD000EF1AE, 0x80005B87000E8DAC,
-    0x80005B8F80005B8B, 0x80005B9780005B93, 0x80005B9F80005B9B,
-    0x80005BA780005BA3, 0x80005BAF80005BAB, 0x00000D8080005BB3,
-    0x0000DD820000D581, 0x000141A880005BB7, 0x000151AA000151A9,
-    0x00017DAC000165AB, 0x000091AE000091AD, 0x80005BBB000091AF,
-    0x0001648480005BBF, 0x00019086000164BE, 0x80005BC700008487,
-    0x000091B880005BCB, 0x000091BA000091B9, 0x0000B1BC000091BB,
-    0x0000B1BE0000B1BD, 0x0000F1B00000B1BF, 0x0000C1B20000F9B1,
-    0x0000B1B40000C1B3, 0x0000B1B60000B1B5, 0x0002D9B30000B1B7,
-    0x000370BE80005BCF, 0x80005BD3000444BE, 0x0002F5B5000311B6,
-    0x00033DBB80005BD7, 0x80005BDB000335BA, 0x000375BF80005BDF,
-    0x000315BD000375BE, 0x80005BE3000315BC, 0x80005BE700029DA3,
-    0x000355A680005BEB, 0x80005BF380005BEF, 0x000371AA0002B1A5,
-    0x80005BF7000379AB, 0x000331AE80005BFB, 0x000351AC000331AF,
-    0x000339A8000351AD, 0x00008DAA000339A9, 0x00008DAC00009DAB,
-    0x0000B5AE0000BDAD, 0x80005BFF0000DDAF, 0x80005C0780005C03,
-    0x80005C0F80005C0B, 0x80005C1780005C13, 0x000069B880005C1B,
-    0x000079BA000069B9, 0x000069BC000079BB, 0x0001D9BE000069BD,
-    0x0000A9B00001D9BF, 0x0000BDB20000A9B1, 0x000099B40000B5B3,
-    0x000059B6000099B5, 0x80005C1F000059B7, 0x80005C2780005C23,
-    0x000044EF80005C2B, 0x0005988680005C2F, 0x000098E30002E487,
-    0x0001F8E10002A884, 0x0000398080005C37, 0x00002D8200003981,
-    0x000145B380005C3B, 0x80005C4380005C3F, 0x80005C4B80005C47,
-    0x000145B500017DB6, 0x00012DBB80005C4F, 0x80005C5300012DBA,
-    0x00011DBF80005C57, 0x000121BD00011DBE, 0x80005C5B000129BC,
-    0x80005C5F000E50E1, 0x80005C63000F3CE3, 0x80005C6B80005C67,
-    0x80005C7380005C6F, 0x000000E380005C77, 0x80005C7F80005C7B,
-    0x0005F48480005C83, 0x0001A9AE000EE4EF, 0x00019DAC0001A9AF,
-    0x000199AA000195AD, 0x80005C87000199AB, 0x0001C9A680005C8B,
-    0x80005C9380005C8F, 0x0007FD820001F1A5, 0x0007FD800001F1A3,
-    0x80005C330007F581, 0x80005C9B80005C97, 0x80005CA380005C9F,
-    0x0003B88680005CA7, 0x00070DA800039087, 0x00076DAA000719A9,
-    0x00077DAC000765AB, 0x000765AE000765AD, 0x00072DB0000755AF,
-    0x0007C1B20007C5B1, 0x0007C5B40007DDB3, 0x0007C5B60007CDB5,
-    0x0007C5B80007FDB7, 0x0007D9BA0007C9B9, 0x0007A9BC0007D9BB,
-    0x00079DBE0007A9BD, 0x0007C5B3000795BF, 0x80005CAF80005CAB,
-    0x80005CB780005CB3, 0x0007D5B50007C5B6, 0x000721BB80005CBB,
-    0x80005CBF0007C9BA, 0x000729BF80005CC3, 0x000729BD000721BE,
-    0x80005CC7000735BC, 0x80005CCB000781A3, 0x000781A680005CCF,
-    0x80005CD780005CD3, 0x00078DAA000791A5, 0x80005CDB000765AB,
-    0x000765AE80005CDF, 0x000771AC00076DAF, 0x000115A800076DAD,
-    0x000181AA000181A9, 0x000181AC000181AB, 0x0001B1AE000189AD,
-    0x80005CE30001B1AF, 0x80005CEB80005CE7, 0x80005CF380005CEF,
-    0x80005CFB80005CF7, 0x000065B880005CFF, 0x000065BA00006DB9,
-    0x000065BC00007DBB, 0x000065BE00006DBD, 0x0001A1B00000DDBF,
-    0x0001A5B20001ADB1, 0x0001A9B40001B9B3, 0x000195B600019DB5,
-    0x80005D0300005DB7, 0x00001D8100001D82, 0x80005D0700001D80,
-    0x80005D0F80005D0B, 0x0002AC84000158BE, 0x0001088780005D13,
-    0x80005D1700018C86, 0x80005D1B80005BC3, 0x000149B380005D1F,
-    0x80005D2780005D23, 0x80005D2F80005D2B, 0x000149B5000149B6,
-    0x000129BB80005D33, 0x80005D37000125BA, 0x000119BF80005D3B,
-    0x00012DBD00012DBE, 0x0003D8BE000131BC, 0x0006F8EF80005D3F,
-    0x80005D4780005D43, 0x0002E0EF80005D4B, 0x000194E180005D4F,
-    0x000274E380005D53, 0x000598E180005D57, 0x00070CE380005D5B,
-    0x80005D6380005D5F, 0x0002C1A380005D67, 0x0002C1A500038C84,
-    0x80005D6F80005D6B, 0x80005D730002C1A6, 0x0002A1AB80005D77,
-    0x0002A5AD0002ADAA, 0x000291AF0002B9AC, 0x000331A80002A5AE,
-    0x000335AA00033DA9, 0x000359AC000349AB, 0x000349AE000359AD,
-    0x0000C580000341AF, 0x0000198200000981, 0x80005D7F80005D7B,
-    0x00036C8780005D87, 0x0000EDB8001C1C86, 0x0000F5BA0000F1B9,
-    0x000095BC00008DBB, 0x000081BE000081BD, 0x000309B00000BDBF,
-    0x0000EDB2000309B1, 0x0000E1B40000E1B3, 0x0000E5B60000E1B5,
-    0x80005D8B0000D9B7, 0x80005D9380005D8F, 0x80005D970002B1B3,
-    0x0002D5B60002DDB5, 0x80005D9F80005D9B, 0x0002C1BA80005DA3,
-    0x000135BC0002C1BB, 0x000129BE000139BD, 0x00028DA6000129BF,
-    0x80005DAB80005DA7, 0x80005DAF000285A5, 0x80005DB30002E9A3,
-    0x000171AE80005DB7, 0x00016DAC000171AF, 0x000299AA000161AD,
-    0x80005DBB000299AB, 0x80005DC380005DBF, 0x80005DC7000684E1,
-    0x80005DCB000624E3, 0x80005DCF000194E1, 0x001DE8BE00002CE3,
-    0x80005DD780005DD3, 0x001DFC840000C8EF, 0x80005DDB001CF0BE,
-    0x80005DDF000748EF, 0x80005DE780005DE3, 0x00001D8180005DEB,
-    0x80005DEF00001D80, 0x001C408600000582, 0x80005DF7001CC487,
-    0x80005DFF80005DFB, 0x80005E0780005E03, 0x001EB5A880005E0B,
-    0x001F0DAA001F05A9, 0x001F01AC001F05AB, 0x001F39AE001F09AD,
-    0x80005DF3001F39AF, 0x80005E1380005E0F, 0x80005E1B80005E17,
-    0x80005E2380005E1F, 0x001FC9B880005E27, 0x001FD1BA001FD5B9,
-    0x001FF9BC001FE9BB, 0x001F99BE001FEDBD, 0x001F25B0001F99BF,
-    0x001F39B2001F2DB1, 0x001F2DB4001F35B3, 0x001F05B6001F0DB5,
-    0x001F85B3001FFDB7, 0x80005E2F80005E2B, 0x80005E3780005E33,
-    0x001F89B5001F89B6, 0x001FE9BB80005E3B, 0x80005E3F001FE1BA,
-    0x001FC9BF80005E43, 0x001FE1BD001FE9BE, 0x80005E47001FEDBC,
-    0x80005E4B001FC1A3, 0x001FCDA680005E4F, 0x80005E5780005E53,
-    0x001FA5AA001FCDA5, 0x80005E5B001FADAB, 0x001FADAE80005E5F,
-    0x001FA9AC001F8DAF, 0x001EE9A8001FA5AD, 0x001EF9AA001EE9A9,
-    0x001EE9AC001EF9AB, 0x00013DAE001EE9AD, 0x0001F980000135AF,
-    0x0001C5820001CD81, 0x80005E6300026084, 0x0001688780005E67,
-    0x0001D1B800009C86, 0x0001E1BA0001D9B9, 0x000191BC0001E1BB,
-    0x000195BE00019DBD, 0x00014DB0000189BF, 0x00015DB2000155B1,
-    0x00014DB4000155B3, 0x0001F1B60001F1B5, 0x001E71B30001F1B7,
-    0x80005E6F80005E6B, 0x80005E7780005E73, 0x001E61B5001E69B6,
-    0x000191BB80005E7B, 0x80005E7F000189BA, 0x000135BF80005E83,
-    0x000181BD000189BE, 0x80005E87000189BC, 0x001E39A380005D83,
-    0x001E29A580005E8B, 0x80005E9380005E8F, 0x80005E97001E21A6,
-    0x0001D9AB80005E9B, 0x0001C9AD0001C1AA, 0x00017DAF0001C1AC,
-    0x80005E9F0001C1AE, 0x80005EA780005EA3, 0x80005EAF80005EAB,
-    0x80005EB780005EB3, 0x80005EBF80005EBB, 0x80005EC780005EC3,
-    0x80005ECF80005ECB, 0x80005EDB80005ED7, 0x001EE4E1000370BE,
-    0x001FE8E300024884, 0x0000798000040084, 0x0000698200007981,
-    0x0004C08680005EDF, 0x80005EE300034487, 0x80005EEB80005EE7,
-    0x001F20EF80005EEF, 0x80005EF780005EF3, 0x80005EFF80005EFB,
-    0x80005F03000248EF, 0x80005F0B80005F07, 0x0004E8BE80005F0F,
-    0x80005F1780005F13, 0x80005F1F80005F1B, 0x80005F23000190E1,
-    0x000349B30002E8E3, 0x80005F2B80005F27, 0x80005F3380005F2F,
-    0x000349B5000349B6, 0x00032DBB80005F37, 0x80005F3B000325BA,
-    0x000315BF80005F3F, 0x000321BD000315BE, 0x000235A8000329BC,
-    0x000281AA000281A9, 0x000281AC000281AB, 0x0002B1AE000289AD,
-    0x0001FD800002B1AF, 0x0001C5820001CD81, 0x0004908680005F47,
-    0x80005F4B00050487, 0x000149B800047084, 0x000159BA000149B9,
-    0x000149BC000159BB, 0x000179BE000149BD, 0x0002A1B0000179BF,
-    0x0002BDB20002A9B1, 0x000291B40002B5B3, 0x000179B6000291B5,
-    0x80005F4F000179B7, 0x80005F5780005F53, 0x80005F5F80005F5B,
-    0x80005F6780005F63, 0x0004E8BE001ED0EF, 0x80005F6B001E5CE1,
-    0x80005F6F000090E3, 0x80005F7780005F73, 0x000249A380005F7B,
-    0x80005F8380005F7F, 0x80005F8B80005F87, 0x000249A5000249A6,
-    0x00022DAB80005F8F, 0x80005F93000225AA, 0x000215AF80005F97,
-    0x000221AD000215AE, 0x000635A8000229AC, 0x000655AA00063DA9,
-    0x00067DAC000665AB, 0x00066DAE000665AD, 0x80005F43000661AF,
-    0x80005F9F80005F9B, 0x00000D8080005FA3, 0x0000B1820000B181,
-    0x0006E9B880005FA7, 0x0006F9BA0006E9B9, 0x000695BC0006F5BB,
-    0x000695BE00069DBD, 0x0006E1B000068DBF, 0x0006E1B20006E1B1,
-    0x0006E5B40006FDB3, 0x0006E5B60006EDB5, 0x0006DDB30006D9B7,
-    0x80005FAF80005FAB, 0x80005FB780005FB3, 0x0006F5B50006E5B6,
-    0x000625BB80005FBB, 0x00009886000625BA, 0x000629BF0000AC87,
-    0x000629BD000621BE, 0x80005FBF000635BC, 0x80005FC3000699A3,
-    0x0006A1A680005FC7, 0x80005FCF80005FCB, 0x000661AA0006B1A5,
-    0x80005FD3000661AB, 0x000665AE80005FD7, 0x000671AC00066DAF,
-    0x00062DA800066DAD, 0x000689AA00063DA9, 0x000699AC000689AB,
-    0x000689AE000699AD, 0x80005FDB000689AF, 0x80005FE380005FDF,
-    0x80005FEB80005FE7, 0x80005FF380005FEF, 0x00068DB880005FF7,
-    0x000695BA000695B9, 0x0006BDBC0006A5BB, 0x000171BE000171BD,
-    0x0006F9B0000171BF, 0x0006D9B20006CDB1, 0x0006C9B40006D9B3,
-    0x0006BDB60006C9B5, 0x000601B30006B5B7, 0x80005FFF80005FFB,
-    0x8000600780006003, 0x000611B5000619B6, 0x000609BB8000600B,
-    0x8000600F00063DBA, 0x00060DBF80006013, 0x00060DBD00060DBE,
-    0x8000601700060DBC, 0x80005ED3000645A3, 0x00065DA68000601B,
-    0x000058848000601F, 0x000679AA000655A5, 0x000168BE00064DAB,
-    0x000649AE80006023, 0x000649AC000649AF, 0x0003C180000649AD,
-    0x0003DD820003C981, 0x800060270002CDA3, 0x0002CDA60002D9A5,
-    0x000D80868000602B, 0x0002C5AA00039487, 0x000315AC00030DAB,
-    0x000315AE00031DAD, 0x00179CE100030DAF, 0x000314E3000618E3,
-    0x000218EF000634E1, 0x800060338000602F, 0x0002D0E380006037,
-    0x000250E18000603B, 0x800060438000603F, 0x0025E4EF00062CEF,
-    0x00024DAC80006047, 0x000251AE000251AD, 0x000201A8000265AF,
-    0x000259AA000209A9, 0x000DEC84000255AB, 0x8000604F8000604B,
-    0x000DF8BE80006053, 0x8000605B80006057, 0x000351BC8000605F,
-    0x000361BE000359BD, 0x0003EDB8000361BF, 0x000351BA000351B9,
-    0x0003CDB4000351BB, 0x0003DDB60003D5B5, 0x00021DB00003D5B7,
-    0x0003DDB20003D5B1, 0x0000C8E30003D5B3, 0x0001B8E180006063,
-    0x000F548480006067, 0x8000606F8000606B, 0x0002E9A180006073,
-    0x0003A5A3000615A0, 0x000020EF00030DA2, 0x8000607B80006077,
-    0x800060838000607F, 0x0026348580006087, 0x8000608B00038DB3,
-    0x00038DB6000399B5, 0x000FC0868000608F, 0x000385BA000D9487,
-    0x000255BC00024DBB, 0x000255BE00025DBD, 0x8000609700024DBF,
-    0x8000609F8000609B, 0x800060A7800060A3, 0x00EBDD8F800060AB,
-    0x000FE0BE0006C4EF, 0x800060AF00068CE1, 0x0000F580000680E3,
-    0x0000E5820000E581, 0x001F6D99800060B3, 0x001BC99B001FC598,
-    0x001A759D001A799A, 0x0007F99F001BC59C, 0x001669910006199E,
-    0x00176D9300EBB190, 0x0013D1950017CD92, 0x0012519700126994,
-    0x00E455830013CD96, 0x8000609300E44182, 0x001D8D87800060B7,
-    0x0018398500189186, 0x001C118B00199584, 0x800060BB001C058A,
-    0x0015858F800060BF, 0x0010398D0010998E, 0x00145193001D958C,
-    0x800060C300144592, 0x00096197800060C7, 0x00099D9500087596,
-    0x000C419B0015F994, 0x800060CB000DAD9A, 0x800060D3800060CF,
-    0x800060DB800060D7, 0x000D6DA1000CF19C, 0x000401A3800060DF,
-    0x000499A5000085A2, 0x003819A7000491A4, 0x003809A90005C5A6,
-    0x003C89AB003829A8, 0x003001AD003C01AA, 0x003029AF003C1DAC,
-    0x0034E1B100309DAE, 0x002801B3003401B0, 0x0028D9B50034FDB2,
-    0x800060E30028D1B4, 0x800060EB800060E7, 0x800060F3800060EF,
-    0x00001D80800060F7, 0x0001D98200000981, 0x000F11A8800060FB,
-    0x000F5DAA000F19A9, 0x000F4DAC000F55AB, 0x000F75AE000F71AD,
-    0x800060FF000F6DAF, 0x0000888680006107, 0x8000610B00011087,
-    0x800061138000610F, 0x000E4DB880006117, 0x000E51BA000E51B9,
-    0x000E65BC000E51BB, 0x000E65BE000E6DBD, 0x000F1DB0000E1DBF,
-    0x000FC1B2000FC1B1, 0x000FC5B4000FC1B3, 0x000FC5B6000FCDB5,
-    0x000FD5B3000E79B7, 0x8000611F8000611B, 0x8000612780006123,
-    0x000FC5B5000FF1B6, 0x000F05BB8000612B, 0x8000612F000FD9BA,
-    0x000F09BF80006133, 0x000F15BD000F01BE, 0x80006137000F15BC,
-    0x8000613B000F91A3, 0x000FB5A68000613F, 0x8000614780006143,
-    0x000F9DAA000F81A5, 0x8000614B000F41AB, 0x000F45AE8000614F,
-    0x000F51AC000F4DAF, 0x000D81A8000F51AD, 0x000D81AA000D81A9,
-    0x000D81AC000D81AB, 0x000DB1AE000D81AD, 0x80006153000DA1AF,
-    0x8000615B80006157, 0x800061638000615F, 0x0000BD810000B982,
-    0x000235B80000BD80, 0x0002CDBA00023DB9, 0x00028DBC000295BB,
-    0x0002BDBE0002B5BD, 0x00026DB00002B5BF, 0x000245B200024DB1,
-    0x00023DB4000225B3, 0x000215B600021DB5, 0x8000616700020DB7,
-    0x000D01B38000616B, 0x000D01B58000616F, 0x0003948480006177,
-    0x00042CBE000D09B6, 0x000281BB8000617B, 0x0002F9BD0002A1BA,
-    0x0002D1BF000299BC, 0x8000617F0002F1BE, 0x8000618780006183,
-    0x8000618B000D45A3, 0x000D4DA6000D45A5, 0x0004E0868000618F,
-    0x0002E5AA0004A487, 0x0002DDAC0002C5AB, 0x0002B5AE0002BDAD,
-    0x000225A8000295AF, 0x00023DAA000235A9, 0x00022DAC000235AB,
-    0x000291AE000291AD, 0x80006193000291AF, 0x8000619B80006197,
-    0x000000CC8000619F, 0x800061A7800061A3, 0x000299B8800061AB,
-    0x0002A5BA0002ADB9, 0x000175BC00016DBB, 0x000175BE00017DBD,
-    0x0002F1B000016DBF, 0x0002C1B20002F9B1, 0x0002B1B40002C1B3,
-    0x0002B5B60002BDB5, 0x800061AF0002A9B7, 0x800061B3000E84E3,
-    0x800061B7000EF4E1, 0x800061BF800061BB, 0x00052084800061C3,
-    0x800061CB800061C7, 0x800061D3800061CF, 0x800061D7000EECEF,
-    0x0002F5B3800061DB, 0x0004E886800061DF, 0x00045CBE00048C87,
-    0x0002E5B5000251B6, 0x00027DBB800061E7, 0x800061EB000275BA,
-    0x000239BF800061EF, 0x000255BD000235BE, 0x000535A3000255BC,
-    0x800061E380006173, 0x800061F7800061F3, 0x000525A5000591A6,
-    0x0005BDAB800061FB, 0x800061FF0005B5AA, 0x0005F9AF80006203,
-    0x000595AD0005F5AE, 0x0007F980000595AC, 0x00078D820007F981,
-    0x8000620700068DB3, 0x000689B600069DB5, 0x8000620F8000620B,
-    0x00074DBA80006213, 0x00075DBC000745BB, 0x000741BE000741BD,
-    0x80006217000741BF, 0x800061038000621B, 0x800062238000621F,
-    0x8000622B80006227, 0x000635A88000622F, 0x000669AA000601A9,
-    0x000679AC000669AB, 0x000665AE000665AD, 0x0007E5B000079DAF,
-    0x0007E5B20007EDB1, 0x0007E9B40007F9B3, 0x000759B60007E9B5,
-    0x000771B8000755B7, 0x000771BA000771B9, 0x000755BC000771BB,
-    0x000755BE00075DBD, 0x0007C1A300074DBF, 0x8000623780006233,
-    0x8000623F8000623B, 0x0007D1A50007C5A6, 0x000609AB80006243,
-    0x80006247000601AA, 0x00060DAF8000624B, 0x00060DAD00060DAE,
-    0x00006980000611AC, 0x0000058200006981, 0x000198BE8000624F,
-    0x8000625300019884, 0x003C008680006257, 0x8000625B0001C487,
-    0x800062638000625F, 0x8000626B80006267, 0x00065DA88000626F,
-    0x000665AA00066DA9, 0x000181AC000181AB, 0x0001B1AE000189AD,
-    0x800062730001B1AF, 0x8000627B80006277, 0x800062838000627F,
-    0x8000628B80006287, 0x000155B88000628F, 0x000155BA00015DB9,
-    0x0000D9BC0000C9BB, 0x0000C9BE0000D9BD, 0x0001B1B00000C1BF,
-    0x000189B20001B9B1, 0x000171B4000189B3, 0x000175B6000171B5,
-    0x0005EDB300016DB7, 0x8000629780006293, 0x8000629F8000629B,
-    0x000209B5000209B6, 0x00027DBB800062A3, 0x800062A7000275BA,
-    0x0002B5BF800062AB, 0x00025DBD0002B5BE, 0x000260BE00025DBC,
-    0x800062AF0005A9A3, 0x00024DA6800062B3, 0x800062BB800062B7,
-    0x000231AA00024DA5, 0x800062BF000239AB, 0x0002F1AE0003E084,
-    0x000219AC0002F1AF, 0x003C88BE000219AD, 0x00032DAA800062C7,
-    0x00033DAC000325AB, 0x00032DAE000325AD, 0x0000F580000325AF,
-    0x0000C1820000FD81, 0x800062CB0003C1A3, 0x0003C1A60003C1A5,
-    0x003C6086800062CF, 0x800062D300035487, 0x800062DB800062D7,
-    0x800062DF0002A8E3, 0x800062E30001A4E1, 0x800062E700025CEF,
-    0x800062EF800062EB, 0x800062F7800062F3, 0x800062FF800062FB,
-    0x8000630300078CE3, 0x8000630700043CE1, 0x8000630F8000630B,
-    0x0002208480006313, 0x8000631B80006317, 0x800063238000631F,
-    0x8000632700076CEF, 0x00022DB38000632B, 0x8000632F003D4484,
-    0x8000633B80006337, 0x00022DB500022DB6, 0x0002C9BB8000633F,
-    0x800063430002C1BA, 0x0002C9BF80006347, 0x0002C9BD0002C1BE,
-    0x000778E10002D1BC, 0x000638E30000C0E3, 0x0000A98000065CE1,
-    0x0000D1820000A981, 0x8000634F8000634B, 0x003CA4BE80006353,
-    0x8000635B80006357, 0x000690EF00001CEF, 0x003C60868000635F,
-    0x80006363003D0487, 0x80006367000171B3, 0x000109B6000119B5,
-    0x8000636F8000636B, 0x000101BA80006373, 0x000101BC000101BB,
-    0x000101BE000101BD, 0x003E85A8000101BF, 0x003EB5AA003EB5A9,
-    0x003EDDAC003EC5AB, 0x003EC1AE003EC5AD, 0x80006377003EFDAF,
-    0x8000637F8000637B, 0x8000638780006383, 0x8000638F8000638B,
-    0x003F51B880006393, 0x003F51BA003F51B9, 0x003F75BC003F51BB,
-    0x003F75BE003F7DBD, 0x003E89B0003F6DBF, 0x003E99B2003E89B1,
-    0x003E89B4003E99B3, 0x003F79B6003E89B5, 0x800062C3003F75B7,
-    0x80006397003E39A3, 0x003E41A680006333, 0x8000639F8000639B,
-    0x003E49AA003E51A5, 0x800063A3003E49AB, 0x003E49AE800063A7,
-    0x003E49AC003E49AF, 0x00004980003E49AD, 0x0000518200005181,
-    0x800063AB003F91B3, 0x003F45B6003FB9B5, 0x00004086800063AF,
-    0x003F45BA00030487, 0x003F4DBC003F5DBB, 0x003F29BE003F3DBD,
-    0x003E4DA8003F21BF, 0x003E55AA003E55A9, 0x003E7DAC003E65AB,
-    0x003EB9AE003E89AD, 0x800063B3003EB9AF, 0x800063BB800063B7,
-    0x800063C3800063BF, 0x800063CB800063C7, 0x000161B8800063CF,
-    0x000161BA000161B9, 0x000161BC000161BB, 0x000161BE000161BD,
-    0x003ECDB0000161BF, 0x003ED5B2003ED5B1, 0x003EA1B4003EA5B3,
-    0x003E99B6003EA9B5, 0x003EDDA3003E99B7, 0x800063D7800063D3,
-    0x800063DF800063DB, 0x003EF5A5003E09A6, 0x003E11AB800063E3,
-    0x800063E7003E09AA, 0x003E6DAF800063EB, 0x003E71AD003E65AE,
-    0x800063EF003E01AC, 0x800063F7800063F3, 0x800063FF800063FB,
-    0x8000640780006403, 0x000039808000640B, 0x0000058200003981,
-    0x0001B8BE8000640F, 0x800064170001B884, 0x0002ADA88000641B,
-    0x0002D5AA0002D5A9, 0x000335AC00032DAB, 0x000335AE00033DAD,
-    0x8000641F00032DAF, 0x8000642780006423, 0x8000642F8000642B,
-    0x8000643780006433, 0x0003E9B88000643B, 0x000389BA0003E9B9,
-    0x000399BC000389BB, 0x000389BE000399BD, 0x000355B0000389BF,
-    0x000355B200035DB1, 0x0003F9B40003E9B3, 0x0003E9B60003F9B5,
-    0x00025DB30003E1B7, 0x000428868000643F, 0x800064430003CC87,
-    0x0003BDB50003BDB6, 0x000399BB80006447, 0x8000644B000391BA,
-    0x00033DBF8000644F, 0x00033DBD00033DBE, 0x00140085000381BC,
-    0x80006453000219A3, 0x0003F9A680006457, 0x8000645F8000645B,
-    0x0003D5AA0003F9A5, 0x800064630003DDAB, 0x000379AE80006467,
-    0x0003C5AC000379AF, 0x003E54E3000379AD, 0x003ED0E1003FDCE1,
-    0x8000646B003E3CE3, 0x8000646F00001CEF, 0x8000647780006473,
-    0x8000647B0000C0E3, 0x003E60EF0001D4E1, 0x800064878000647F,
-    0x00006180003F44EF, 0x00007D8200006D81, 0x0004F08600050084,
-    0x000510BE00059C87, 0x8000648F8000648B, 0x8000649780006493,
-    0x8000649F8000649B, 0x800064A7800064A3, 0x003D89B8800064AB,
-    0x003D99BA003D89B9, 0x003DB9BC003D91BB, 0x003DD1BE003DB9BD,
-    0x003E05B0003DD1BF, 0x003E05B2003E0DB1, 0x003E05B4003E1DB3,
-    0x003E05B6003E0DB5, 0x003E9DA8003DB9B7, 0x003EA5AA003EADA9,
-    0x003EA5AC003EBDAB, 0x003EA5AE003EADAD, 0x0004AC84003E7DAF,
-    0x800064AF0004ACBE, 0x800064B7800064B3, 0x800064BF800064BB,
-    0x0005F9A8800064C3, 0x000629AA0005F9A9, 0x000639AC000629AB,
-    0x000629AE000639AD, 0x80006483000629AF, 0x800064CB800064C7,
-    0x800064D3800064CF, 0x800064DB800064D7, 0x00068DB8800064DF,
-    0x000691BA000691B9, 0x0006BDBC0006A5BB, 0x000741BE000745BD,
-    0x000659B0000741BF, 0x0006EDB2000659B1, 0x0006EDB40006FDB3,
-    0x0006D5B60006DDB5, 0x0006A1B30006B5B7, 0x800064E7800064E3,
-    0x800064EF800064EB, 0x0006B1B50006B9B6, 0x00069DBB800064F7,
-    0x800064F300069DBA, 0x000619BF80006413, 0x000629BD000629BE,
-    0x00005182000639BC, 0x000041800006E5A3, 0x0006FDA600004181,
-    0x800064FF800064FB, 0x0006D9AA0006F5A5, 0x000048860006D9AB,
-    0x00066DAE00006C87, 0x00067DAC00065DAF, 0x000639A800066DAD,
-    0x000669AA000659A9, 0x000679AC000669AB, 0x000669AE000679AD,
-    0x80006503000669AF, 0x8000650B80006507, 0x800065138000650F,
-    0x8000651B80006517, 0x000165B88000651F, 0x000165BA00016DB9,
-    0x000165BC00017DBB, 0x000165BE00016DBD, 0x000619B00001D9BF,
-    0x000681B2000619B1, 0x000681B4000681B3, 0x000681B6000681B5,
-    0x0006E1B3000681B7, 0x8000652780006523, 0x8000652F8000652B,
-    0x0006F1B50006F9B6, 0x0006DDBB80006533, 0x800065370006DDBA,
-    0x000645BF8000653B, 0x000655BD000645BE, 0x8000653F000655BC,
-    0x800065430006A5A3, 0x0006BDA680006547, 0x8000654F8000654B,
-    0x000699AA0006B5A5, 0x80006553000699AB, 0x000601AE80006557,
-    0x000611AC000601AF, 0x00029DA8000611AD, 0x0002D1AA0002D1A9,
-    0x000335AC00032DAB, 0x000335AE00033DAD, 0x8000655B00032DAF,
-    0x000264BE8000655F, 0x8000656B80006567, 0x800065738000656F,
-    0x0003E9B880006577, 0x000389BA0003E9B9, 0x00039DBC000385BB,
-    0x000381BE000381BD, 0x000355B00003B5BF, 0x000355B200035DB1,
-    0x0003F9B40003E9B3, 0x0003E9B60003F9B5, 0x00036D800003E1B7,
-    0x0000AD820000A581, 0x8000657B000255B3, 0x0003B1B60003B1B5,
-    0x0002E0848000657F, 0x00039DBA80006583, 0x00038DBC000395BB,
-    0x000331BE000331BD, 0x000219A3000331BF, 0x0069708580006587,
-    0x8000658F8000658B, 0x0003FDA50003FDA6, 0x0003D9AB80006593,
-    0x000C48860003D1AA, 0x00037DAF0003BC87, 0x00037DAD00037DAE,
-    0x800065970003C1AC, 0x8000659F8000659B, 0x0006B0EF800065A3,
-    0x800065A70003C4EF, 0x00067CE3800065AB, 0x0007B8E1000394E3,
-    0x800065AF0001DCE1, 0x800065B7800065B3, 0x800065BF800065BB,
-    0x00024484800065C3, 0x00000D80000D58BE, 0x00003D8200003581,
-    0x800065CB800065C7, 0x000CC886800065CF, 0x800065D7000D5C87,
-    0x800065DF800065DB, 0x800065E7800065E3, 0x800065EF800065EB,
-    0x800065F7800065F3, 0x800065FF800065FB, 0x000DAC84000680EF,
-    0x800066030006ECE1, 0x8000660700065CE3, 0x8000660F8000660B,
-    0x0001F5B380006613, 0x8000661B80006617, 0x800066238000661F,
-    0x0001E5B5000145B6, 0x00012DBB80006627, 0x8000662B000121BA,
-    0x0000F5BF8000662F, 0x000125BD0000F5BE, 0x000E2DA800012DBC,
-    0x000E3DAA000E35A9, 0x000E2DAC000E35AB, 0x000E85AE000E85AD,
-    0x800065D3000EB9AF, 0x8000663780006633, 0x000019808000663B,
-    0x0000058200001981, 0x000F59B88000663F, 0x000F79BA000F59B9,
-    0x000F69BC000F79BB, 0x000F19BE000F69BD, 0x000EA5B0000F19BF,
-    0x000EB9B2000EA9B1, 0x000F71B4000EB1B3, 0x000F71B6000F71B5,
-    0x000EBDA5000F71B7, 0x80006647000388BE, 0x80006643000E1DA6,
-    0x000EADA38000664B, 0x000E7DAD8000664F, 0x000FADAF000E75AC,
-    0x80006563000FADAE, 0x000E75AB80006653, 0x000FB9B3000E79AA,
-    0x0000688680006657, 0x8000665B00030C87, 0x000F59B5000F51B6,
-    0x000F75BB8000665F, 0x80006663000F75BA, 0x000F45BF80006667,
-    0x000F51BD000F45BE, 0x000E71A8000F65BC, 0x000E71AA000E71A9,
-    0x000E91AC000E71AB, 0x000E91AE000E91AD, 0x8000666B000E91AF,
-    0x800066738000666F, 0x8000667B80006677, 0x800066838000667F,
-    0x000E85B880006687, 0x000E85BA000E8DB9, 0x000E8DBC000E9DBB,
-    0x000EB5BE000EBDBD, 0x000EF1B0000179BF, 0x000EF1B2000EF1B1,
-    0x000EC1B4000EC5B3, 0x000EC1B6000EC1B5, 0x000EF9A3000EC1B7,
-    0x8000668F8000668B, 0x8000669780006693, 0x000E19A5000E11A6,
-    0x000E35AB8000669B, 0x8000669F000E35AA, 0x000E05AF800066A3,
-    0x000E11AD000E05AE, 0x00000D80000E25AC, 0x00001D8200001581,
-    0x800066AB800066A7, 0x00019484800066AF, 0x00074086000194BE,
-    0x800066B70000E487, 0x800066BF800066BB, 0x800066C7800066C3,
-    0x000289A8800066CB, 0x000295AA000291A9, 0x0002D5AC0002B9AB,
-    0x0002C5AE0002C5AD, 0x800066CF0002F5AF, 0x800066D7800066D3,
-    0x00037CBE800066DB, 0x800066E3800066DF, 0x00037DB8800066E7,
-    0x0003C1BA0003C1B9, 0x0003C1BC0003C1BB, 0x0003F1BE0003C9BD,
-    0x00028DB00003F1BF, 0x00034DB2000345B1, 0x00035DB4000345B3,
-    0x00034DB6000345B5, 0x00021DB3000345B7, 0x800066EF800066EB,
-    0x800066F7800066F3, 0x00025DB5000245B6, 0x000381BB800066FB,
-    0x800066FF000249BA, 0x000381BF80006703, 0x000391BD000399BE,
-    0x80006707000399BC, 0x8000670B000259A3, 0x000201A68000670F,
-    0x8000671780006713, 0x00020DAA000219A5, 0x8000671B0003C5AB,
-    0x0003DDAE8000671F, 0x0003DDAC0003C5AF, 0x0001D9800003D5AD,
-    0x0001E5820001ED81, 0x80006727000EB8EF, 0x000298840001C4E1,
-    0x8000672B0000D4E3, 0x8000672F0004ECBE, 0x80006733000008EF,
-    0x80006737000FC4E1, 0x00050086000E64E3, 0x8000673B00056887,
-    0x8000673F0002BDB3, 0x0002A5B60002ADB5, 0x8000674780006743,
-    0x000141BA8000674B, 0x000145BC000141BB, 0x000145BE00014DBD,
-    0x8000674F0001F9BF, 0x8000675780006753, 0x8000675F8000675B,
-    0x8000676780006763, 0x00046C840001E0EF, 0x8000676B000ED4E1,
-    0x8000676F000E1CE3, 0x8000677780006773, 0x000231A38000677B,
-    0x000724848000677F, 0x8000678780006783, 0x000221A5000229A6,
-    0x0001CDAB8000678B, 0x8000678F0001CDAA, 0x000175AF80006797,
-    0x0001C1AD0001C9AE, 0x0005E9A80001C9AC, 0x0005F9AA0005E9A9,
-    0x0005E9AC0005F9AB, 0x000639AE0005E9AD, 0x80006723000639AF,
-    0x0007F5810007CD82, 0x800067930007FD80, 0x000318868000679B,
-    0x0006D1B800039087, 0x0006E1BA0006D9B9, 0x000691BC0006E1BB,
-    0x000695BE00069DBD, 0x000649B0000689BF, 0x00065DB2000649B1,
-    0x00064DB4000655B3, 0x0006F1B60006F1B5, 0x0007E1B00006F1B7,
-    0x000709B20007E1B1, 0x000719B4000709B3, 0x000709B6000719B5,
-    0x000739B8000709B7, 0x000719BA00070DB9, 0x000709BC000719BB,
-    0x00077DBE000709BD, 0x8000679F000771BF, 0x800067A3800066B3,
-    0x800067AB800067A7, 0x800067B3800067AF, 0x0007F1A8800067B7,
-    0x0007C1AA0007C5A9, 0x0007C9AC0007DDAB, 0x0007A9AE0007BDAD,
-    0x000629B30007A1AF, 0x800067BF800067BB, 0x800067C7800067C3,
-    0x000621B500065DB6, 0x000645BB800067CB, 0x800067CF000645BA,
-    0x0006BDBF800067D3, 0x0006BDBD0006BDBE, 0x800067D70006BDBC,
-    0x800067DB00066DA3, 0x000619A6800067DF, 0x800067E7800067E3,
-    0x000601AA000665A5, 0x800067EB000601AB, 0x0006F9AE800067EF,
-    0x0006F9AC0006F9AF, 0x00066DA80006F9AD, 0x000149AA0001B1A9,
-    0x00015DAC000145AB, 0x00014DAE000145AD, 0x800067F3000145AF,
-    0x00001D8100001D82, 0x800067F700001D80, 0x800067FF800067FB,
-    0x000081B8000154BE, 0x000085BA00008DB9, 0x000089BC000099BB,
-    0x0000B5BE0000BDBD, 0x00013DB000007DBF, 0x0000E1B20000E1B1,
-    0x0000E1B40000E1B3, 0x0000D1B60000E9B5, 0x000305BB0000D1B7,
-    0x00022C860002D9BA, 0x00030DBF00032C87, 0x000315BD000305BE,
-    0x0002E1B3000315BC, 0x8000680B80006807, 0x8000680F00032C84,
-    0x0002F5B50002E5B6, 0x00029DAA80006813, 0x80006817000341AB,
-    0x000341AE8000681B, 0x000351AC000349AF, 0x8000681F000351AD,
-    0x800068230002A5A3, 0x0002A1A680006827, 0x8000682F8000682B,
-    0x0000A9A90002B1A5, 0x0000B5AB0000A1A8, 0x0000DDAD0000BDAA,
-    0x0000F1AF0000DDAC, 0x001C2CBE0000FDAE, 0x8000683780006833,
-    0x8000683F8000683B, 0x8000684780006843, 0x00007DB98000684B,
-    0x0001CDBB00007DB8, 0x0001DDBD0001CDBA, 0x0001CDBF0001DDBC,
-    0x000095B10001CDBE, 0x00004DB3000089B0, 0x00005DB5000095B2,
-    0x00004DB700005DB4, 0x8000684F00004DB6, 0x8000685780006853,
-    0x8000685F8000685B, 0x8000686780006863, 0x000039808000686B,
-    0x0000058200003981, 0x800068778000686F, 0x0002F8878000687B,
-    0x0004B8E1001D7C86, 0x000640E3001CC8BE, 0x800068838000687F,
-    0x8000688B80006887, 0x800068938000688F, 0x8000689B80006897,
-    0x800068A38000689F, 0x0007B0EF800068A7, 0x800068AF800068AB,
-    0x800068B7800068B3, 0x800068BB0000C8EF, 0x800068C3800068BF,
-    0x800068C7000640EF, 0x800068CB0006FCE1, 0x800068CF00061CE3,
-    0x800068D3000194E1, 0x00001180000664E3, 0x0000298200001981,
-    0x800068D70001FDB3, 0x000195B600019DB5, 0x800068DF800068DB,
-    0x000185BA001D6C84, 0x000189BC000199BB, 0x000151BE000151BD,
-    0x001E31A3000151BF, 0x800068E380006873, 0x800068EB800068E7,
-    0x001E51A5001E59A6, 0x001E55AB800068EF, 0x00030886001E49AA,
-    0x001E9DAF00036C87, 0x001E9DAD001E9DAE, 0x800068F3001E45AC,
-    0x800068F7001F09B3, 0x001F09B6800068FB, 0x80006903800068FF,
-    0x001F05BA001F09B5, 0x80006907001F0DBB, 0x001F05BE8000690B,
-    0x001F15BC001F09BF, 0x001FE5B0001F0DBD, 0x001FE5B2001FEDB1,
-    0x001FE5B4001FFDB3, 0x001F19B6001FE9B5, 0x001F29B8001F19B7,
-    0x001F3DBA001F35B9, 0x001F11BC001F35BB, 0x001F09BE001F1DBD,
-    0x8000690F001F05BF, 0x8000680380006913, 0x8000691B80006917,
-    0x800069238000691F, 0x001FD1A880006927, 0x001FA5AA001FD1A9,
-    0x001FA5AC001FBDAB, 0x001FA5AE001FADAD, 0x001E4DA3001F9DAF,
-    0x8000692F8000692B, 0x8000693780006933, 0x001E4DA5001E4DA6,
-    0x001E49AB8000693B, 0x8000693F001E41AA, 0x001E4DAF80006943,
-    0x001E49AD001E41AE, 0x00000D80001E51AC, 0x00001D8200001581,
-    0x8000694B80006947, 0x0001B4848000694F, 0x0001E8BF0001B4BE,
-    0x0007488680006957, 0x0006BC8400003487, 0x8000695F8000695B,
-    0x00018DA80006B4BE, 0x000195AA000195A9, 0x0001D9AC0001CDAB,
-    0x0001CDAE0001D9AD, 0x800069630001C5AF, 0x8000696B80006967,
-    0x800069738000696F, 0x8000697B80006977, 0x00011DB88000697F,
-    0x0000C1BA0000C1B9, 0x0000C1BC0000C1BB, 0x0000F1BE0000C9BD,
-    0x000189B00000F1BF, 0x000129B2000189B1, 0x00013DB4000129B3,
-    0x00012DB6000125B5, 0x00026DBB000125B7, 0x8000698300026DBA,
-    0x0002C1BF80006987, 0x0002DDBD0002D9BE, 0x00023DB30002DDBC,
-    0x8000698F8000698B, 0x0006FC8480006993, 0x000271B5000279B6,
-    0x000229AA80006997, 0x8000699B000229AB, 0x00029DAE8000699F,
-    0x000299AC000285AF, 0x800069A3000299AD, 0x800069A7000279A3,
-    0x00023DA6800069AB, 0x800069B3800069AF, 0x00276D82000235A5,
-    0x0005A886002A8D83, 0x002E998600036C87, 0x002E1184002FCD87,
-    0x0012218A002E9985, 0x800069B70012298B, 0x0016918E800069BB,
-    0x0012118C00161D8F, 0x001A71920016118D, 0x800069BF001AE593,
-    0x001EF196007694EF, 0x001A4994001EE597, 0x0002299A001E1995,
-    0x800069C70002E19B, 0x800069CF800069CB, 0x0002F19C001220E1,
-    0x001F219F001620E3, 0x001B999D001FC19E, 0x001BA19B001B2D9C,
-    0x001739990017BD9A, 0x001389970017B198, 0x0013A9950013A996,
-    0x002F3993002E7594, 0x002FB191002F3192, 0x0026AD8F002B3590,
-    0x000000D2001F78E3, 0x00019982001F5CE1, 0x0001F180800069D3,
-    0x0007A8BE0001F181, 0x800069DB800069D7, 0x0006BC84800069DF,
-    0x800069E3001F2CEF, 0x001EA4E1800069E7, 0x001E7CE30000CCE3,
-    0x800069EB0001BCE1, 0x800069F3800069EF, 0x800069F700069C84,
-    0x800069FB00066CBE, 0x80006A03800069FF, 0x001E04EF000010EF,
-    0x80006A0B80006A07, 0x80006A1380006A0F, 0x80006A1B80006A17,
-    0x80006A2380006A1F, 0x00003D8080006A27, 0x0007C98200000981,
-    0x0003F9A380006A2B, 0x000331A10003F1A2, 0x007149B0001FCDA0,
-    0x007C09B2007C01B1, 0x007801B4007D21B3, 0x80006953007815B5,
-    0x80006A2F800069C3, 0x00000C86000E28BE, 0x80006A3300038C87,
-    0x80006A3B80006A37, 0x80006A4380006A3F, 0x000055A200005DA1,
-    0x000C01A400007DA3, 0x000CF5A6000C15A5, 0x000821A8000801A7,
-    0x007401AA0008C5A9, 0x007401AC007409AB, 0x007079AE00751DAD,
-    0x0005E5A8007079AF, 0x0005F1AA0005F1A9, 0x000535AC00052DAB,
-    0x000535AE00053DAD, 0x80006A4700052DAF, 0x80006A4F80006A4B,
-    0x80006A5780006A53, 0x80006A5F80006A5B, 0x0006FDB880006A63,
-    0x000685BA00068DB9, 0x0006B9BC0006ADBB, 0x0006ADBE0006B9BD,
-    0x000555B00006A5BF, 0x000555B200055DB1, 0x0006FDB40006E5B3,
-    0x0006D1B60006D1B5, 0x000479B30006D1B7, 0x80006A6B80006A67,
-    0x80006A7380006A6F, 0x0004BDB50004BDB6, 0x000499BB80006A77,
-    0x80006A7B000491BA, 0x000715BF80006A7F, 0x000731BD000739BE,
-    0x80006A83000481BC, 0x80006A8700043DA3, 0x0004F9A680006A8B,
-    0x80006A9380006A8F, 0x0004D5AA0004F9A5, 0x80006A970004DDAB,
-    0x00077DAE80006A9B, 0x0004C5AC000751AF, 0x000769A8000775AD,
-    0x000779AA000769A9, 0x0006F5AC0006D9AB, 0x0006F5AE0006FDAD,
-    0x0000C9800006E5AF, 0x000005820000C981, 0x000F708680006A9F,
-    0x80006AA700003487, 0x00067DB880006AAB, 0x00060DBA000605B9,
-    0x00061DBC000605BB, 0x00060DBE000605BD, 0x00069DB0000605BF,
-    0x00067DB2000675B1, 0x000651B4000675B3, 0x000655B600065DB5,
-    0x0004F1B300064DB7, 0x80006AB380006AAF, 0x80006ABB80006AB7,
-    0x00057DB500057DB6, 0x0005B1BB80006ABF, 0x80006AC3000559BA,
-    0x000599BF80006AC7, 0x0005A1BD000595BE, 0x80006ACB0005A9BC,
-    0x80006ACF0004B5A3, 0x000539A680006AD3, 0x80006ADB80006AD7,
-    0x00051DAA000539A5, 0x80006ADF0005F5AB, 0x0005D1AE80006AE3,
-    0x0005EDAC0005DDAF, 0x0002B9A90005E5AD, 0x0002C9AB0002B1A8,
-    0x000235AD0002B1AA, 0x000235AF000235AC, 0x80006AE7000235AE,
-    0x80006AEF80006AEB, 0x80006AF780006AF3, 0x80006AFF80006AFB,
-    0x0003E9B980006B03, 0x000389BB0003D9B8, 0x00039DBD0003E1BA,
-    0x000381BF00039DBC, 0x000255B1000389BE, 0x000255B3000255B0,
-    0x0003F9B5000255B2, 0x0003E9B70003F1B4, 0x00033DB30003F1B6,
-    0x80006B0B80006B07, 0x80006B1780006B0F, 0x0003B9B50003B1B6,
-    0x000395BB80006B1B, 0x000C8886000395BA, 0x000239BF000DA087,
-    0x000385BD000231BE, 0x80006B1F000385BC, 0x80006B23000379A3,
-    0x0003F5A680006B27, 0x80006B2F80006B2B, 0x0003D1AA0003FDA5,
-    0x80006B330003D1AB, 0x000275AE80006B37, 0x0003C1AC00027DAF,
-    0x000085800003C1AD, 0x0000858200008D81, 0x0007B0E30006D0EF,
-    0x000708E10004F4E3, 0x000638EF0004ECE1, 0x000CA4BE000420EF,
-    0x80006B3F80006B3B, 0x80006B43000184E1, 0x80006B470006DCE3,
-    0x000DC08480006B4B, 0x80006B4F0001C9B3, 0x0001C9B60001D9B5,
-    0x80006B5780006B53, 0x0001C5BA80006B5B, 0x0001B9BC0001ADBB,
-    0x0001ADBE0001B9BD, 0x000E2DB00001A5BF, 0x000E41B2000E45B1,
-    0x000E45B4000E41B3, 0x000E71B6000E4DB5, 0x000E81B8000E71B7,
-    0x000E81BA000E81B9, 0x000E81BC000E81BB, 0x000E81BE000E81BD,
-    0x80006B13000E81BF, 0x80006B6380006B5F, 0x80006AA380006B67,
-    0x80006B6F80006B6B, 0x000DD9A880006B73, 0x000E6DAA000DD9A9,
-    0x000E7DAC000E65AB, 0x000E65AE000E65AD, 0x000E85A3000E55AF,
-    0x0000158100002D82, 0x80006B7700001D80, 0x000E95A5000E85A6,
-    0x000EE1AB80006B7B, 0x80006B7F000E89AA, 0x000EE9AF80006B83,
-    0x000EF5AD000EE1AE, 0x80006B87000EF5AC, 0x00006886000F85B3,
-    0x000F85B600031C87, 0x80006B8F80006B8B, 0x000F8DBA000F91B5,
-    0x80006B93000F49BB, 0x000F31BE80006B97, 0x000F49BC000F31BF,
-    0x000E11A8000F45BD, 0x000E49AA000E19A9, 0x000E5DAC000E45AB,
-    0x000E41AE000E41AD, 0x80006B9B000E75AF, 0x80006BA380006B9F,
-    0x80006BAB80006BA7, 0x80006BB380006BAF, 0x000EA1B880006BB7,
-    0x00010DBA000EA1B9, 0x00011DBC000105BB, 0x000101BE000101BD,
-    0x000E0DB0000101BF, 0x000ED9B2000EC9B1, 0x000EB1B4000ED5B3,
-    0x000EA9B6000EB1B5, 0x000EC1A3000EA1B7, 0x80006BBF80006BBB,
-    0x80006BC780006BC3, 0x000ED5A5000EC1A6, 0x000E0DAB80006BCB,
-    0x80006BCF000EC9AA, 0x000E75AF80006BD3, 0x000E01AD000E75AE,
-    0x80006BD7000E0DAC, 0x80006BDF80006BDB, 0x00000D8080006BE3,
-    0x00003D8200003581, 0x80006BEB80006BE7, 0x0001848480006BEF,
-    0x00073086000184BE, 0x80006BF70000F887, 0x0002C1A880006BFB,
-    0x0002D1AA0002D1A9, 0x0002FDAC0002E5AB, 0x00033DAE000335AD,
-    0x80006BFF000335AF, 0x80006C0780006C03, 0x80006C0F80006C0B,
-    0x80006C1780006C13, 0x0003D1B880006C1B, 0x0003E1BA0003D9B9,
-    0x000391BC0003E1BB, 0x000391BE000391BD, 0x00034DB0000391BF,
-    0x00035DB2000355B1, 0x00034DB4000355B3, 0x0003F1B60003F1B5,
-    0x0003B1BB0003F1B7, 0x80006C1F0003A9BA, 0x000315BF000384BE,
-    0x0003A1BD0003A9BE, 0x000279B30003A9BC, 0x80006C2780006C23,
-    0x80006C2F80006C2B, 0x000255B5000395B6, 0x0003EDAA80006C33,
-    0x80006C370003F5AB, 0x0003EDAE80006C3B, 0x0003EDAC000351AF,
-    0x80006C3F0003E5AD, 0x80006C4300023DA3, 0x0003D1A680006C47,
-    0x80006C4F80006C4B, 0x80006C53000211A5, 0x0000158100002182,
-    0x000004EF00001580, 0x80006C5700029484, 0x0000D8E380006C5B,
-    0x0001F8E180006C5F, 0x80006C6B80006C67, 0x80006C7380006C6F,
-    0x0005608680006C77, 0x80006C7B00050887, 0x80006C7F00016DB3,
-    0x000175B600017DB5, 0x80006C8780006C83, 0x000151BA80006C8B,
-    0x0001F9BC000151BB, 0x0001D1BE0001D1BD, 0x80006C8F0001D1BF,
-    0x80006C930001A5A3, 0x0001BDA680006C97, 0x80006C9F80006C9B,
-    0x000199AA0001B5A5, 0x80006CA3000199AB, 0x000119AE80006CA7,
-    0x000131AC000119AF, 0x80006CAB000119AD, 0x80006CAF000EF4E1,
-    0x000000F4000E14E3, 0x80006CB3000C7CE1, 0x80006CB7000A50E3,
-    0x000420BE80006CBB, 0x80006CBF000D30EF, 0x0004348480006CC3,
-    0x000639B0000EACEF, 0x000681B2000639B1, 0x0006B9B40006A9B3,
-    0x0006A9B60006B9B5, 0x0006E9B80006A1B7, 0x0006C9BA0006E9B9,
-    0x0006DDBC0006C5BB, 0x0006C1BE0006C1BD, 0x80006C6300073DBF,
-    0x00001D8100001D82, 0x80006CC700001D80, 0x80006CCF80006CCB,
-    0x000599A880006BF3, 0x000649AA000599A9, 0x000659AC000649AB,
-    0x000649AE000659AD, 0x0007ADA8000649AF, 0x0007BDAA0007B5A9,
-    0x0007ADAC0007B5AB, 0x0007C9AE0007DDAD, 0x80006CD30007C1AF,
-    0x0003888680006CD7, 0x80006CDB00031087, 0x80006CE380006CDF,
-    0x00076DB880006CE7, 0x000701BA000705B9, 0x000731BC000715BB,
-    0x000729BE000731BD, 0x000781B0000729BF, 0x000769B2000781B1,
-    0x00077DB4000765B3, 0x000761B6000761B5, 0x000635B3000755B7,
-    0x80006CEF80006CEB, 0x80006CF780006CF3, 0x000625B500065DB6,
-    0x000645BB80006CFB, 0x80006CFF000645BA, 0x0006A5BF80006D03,
-    0x0006B1BD0006B9BE, 0x80006D070006BDBC, 0x80006D0B000671A3,
-    0x000619A680006D0F, 0x80006D1780006D13, 0x000601AA000661A5,
-    0x80006D1B000601AB, 0x0006FDAE80006D1F, 0x0006F9AC0006E1AF,
-    0x000125A80006F5AD, 0x00013DAA000135A9, 0x00012DAC000135AB,
-    0x000091AE000091AD, 0x80006D23000091AF, 0x80006D2B80006D27,
-    0x80006D3380006D2F, 0x0003BD810003BD82, 0x000099B80003BD80,
-    0x0000A5BA0000ADB9, 0x000075BC00006DBB, 0x000075BE00007DBD,
-    0x0000F1B000006DBF, 0x0000C1B20000F9B1, 0x0000B1B40000C1B3,
-    0x0000B5B60000BDB5, 0x80006D370000A9B7, 0x80006D3F80006D3B,
-    0x00021CBE00028084, 0x007CF88680006D47, 0x0003AC840002CC87,
-    0x80006D4F80006D4B, 0x80006D5780006D53, 0x80006D5F80006D5B,
-    0x0002F5B380006D63, 0x80006D6B80006D67, 0x80006D6F00000092,
-    0x0002E5B5000351B6, 0x00037DBB80006D73, 0x80006D77000375BA,
-    0x000339BF80006D7B, 0x000355BD000335BE, 0x000235A3000355BC,
-    0x80006D8380006D7F, 0x80006D8B80006D87, 0x000225A5000391A6,
-    0x0003BDAB80006D8F, 0x80006D930003B5AA, 0x0003F9AF80006D97,
-    0x000395AD0003F5AE, 0x0003C0BE000395AC, 0x80006D9F80006D9B,
-    0x00000D8080006DA3, 0x00003D8200003581, 0x80006DAB80006DA7,
-    0x007CC88680006DAF, 0x80006DB700030087, 0x80006DBF80006DBB,
-    0x80006DC780006DC3, 0x80006DCF80006DCB, 0x80006DD780006DD3,
-    0x80006DDF80006DDB, 0x007CEC840003F4EF, 0x80006DE3000194E1,
-    0x80006DE7000330E3, 0x80006DEF80006DEB, 0x000169B380006DF3,
-    0x80006DFB80006DF7, 0x80006E0380006DFF, 0x000169B5000161B6,
-    0x000129BB80006E07, 0x80006E0B000121BA, 0x00011DBF80006E0F,
-    0x000125BD00011DBE, 0x80006E1300012DBC, 0x80006E1B80006E17,
-    0x80006E1F0001A5A3, 0x0001ADA60001A5A5, 0x007C8086007C54BE,
-    0x0001EDAA007CEC87, 0x0001E1AC0001E5AB, 0x0001D1AE0001E9AD,
-    0x80006E270001D1AF, 0x007F048400069CE1, 0x0006D4E3000624E3,
-    0x000430E180006E2B, 0x000794EF80006E2F, 0x0000618100007582,
-    0x80006E3300006980, 0x80006E3B80006E37, 0x007E8DB80006ECEF,
-    0x007E95BA007E95B9, 0x007EBDBC007EA5BB, 0x007ED1BE007ED1BD,
-    0x007E65B0007ED1BF, 0x007E79B2007E6DB1, 0x007E59B4007E71B3,
-    0x007EBDB6007E59B5, 0x007E55A8007EB5B7, 0x007E61AA007E61A9,
-    0x007E61AC007E61AB, 0x007E61AE007E61AD, 0x80006E3F007E61AF,
-    0x80006D4380006DB3, 0x80006E2380006E43, 0x80006E4B80006E47,
-    0x007E71A880006E4F, 0x007E71AA007E71A9, 0x007F91AC007E71AB,
-    0x007F91AE007F91AD, 0x80006E53007F91AF, 0x80006E5B80006E57,
-    0x80006E6380006E5F, 0x80006E6B80006E67, 0x007F85B880006E6F,
-    0x007F85BA007F8DB9, 0x007F8DBC007F9DBB, 0x007FB5BE007FBDBD,
-    0x007FF1B0007F5DBF, 0x007FF1B2007FF1B1, 0x007FC1B4007FC5B3,
-    0x007FC1B6007FC1B5, 0x007FF9B3007FC1B7, 0x80006E7780006E73,
-    0x80006E7F80006E7B, 0x007E19B5007E11B6, 0x007E35BB80006E83,
-    0x80006E87007E35BA, 0x007E05BF80006E8B, 0x007E11BD007E05BE,
-    0x00001582007E25BC, 0x00006180007FBDA3, 0x007E55A600006181,
-    0x000190BE80006E8F, 0x007E71AA007E5DA5, 0x80006E93007E71AB,
-    0x007E41AE80006E97, 0x007E61AC007E41AF, 0x007E41A8007E55AD,
-    0x007E55AA007E51A9, 0x007E65AC007E7DAB, 0x0001F9AE007E6DAD,
-    0x000000860001F1AF, 0x80006E9B00013487, 0x80006EA380006E9F,
-    0x80006EAB80006EA7, 0x000175B880006EAF, 0x000175BA00017DB9,
-    0x0000D9BC0000C9BB, 0x0000C9BE0000D9BD, 0x000195B00000C1BF,
-    0x000195B200019DB1, 0x000155B400014DB3, 0x000155B600015DB5,
-    0x007DDDB300014DB7, 0x80006EB780006EB3, 0x80006EBF80006EBB,
-    0x007DEDB5007DEDB6, 0x000259BB80006EC3, 0x80006EC7000251BA,
-    0x000299BF80006ECB, 0x000299BD000291BE, 0x80006ECF000241BC,
-    0x80006ED3007D99A3, 0x007DA9A680006ED7, 0x80006EDF80006EDB,
-    0x000215AA007DA9A5, 0x80006EE300021DAB, 0x0002D5AE80006EE7,
-    0x000205AC0002DDAF, 0x80006EEB0002DDAD, 0x80006EF380006EEF,
-    0x00001D8080006EF7, 0x0000398200000981, 0x80006EFF80006EFB,
-    0x80006F07000400BE, 0x80006F0F80006F0B, 0x80006F1780006F13,
-    0x00038C8480006F1B, 0x0003088780006F1F, 0x80006F230004EC86,
-    0x80006F27000288EF, 0x00046C8480006F2B, 0x80006F2F000234E3,
-    0x80006F33000154E1, 0x80006F3B80006F37, 0x00026DA880006F3F,
-    0x0002A5AA0002ADA9, 0x0002A5AC0002BDAB, 0x0002A5AE0002ADAD,
-    0x0004ACBE000119AF, 0x80006F4780006F43, 0x80006F4F80006F4B,
-    0x80006F5780006F53, 0x00010DB880006F5B, 0x000111BA000111B9,
-    0x00013DBC000125BB, 0x0001DDBE0001D5BD, 0x000169B00001D5BF,
-    0x000179B2000169B1, 0x000159B4000171B3, 0x000139B6000159B5,
-    0x00022DB3000135B7, 0x80006F6380006F5F, 0x80006F6B80006F67,
-    0x00022DB500022DB6, 0x0001ADBB80006F6F, 0x80006F770001ADBA,
-    0x00019DBF80006F7B, 0x0001A5BD00019DBE, 0x00004D800001ADBC,
-    0x0000558200005581, 0x80006F7F00006CEF, 0x007EBCEF007FECEF,
-    0x007F10E180006F83, 0x007ED8E3007FD4E3, 0x80006F87007FECE1,
-    0x80006F8B007E4CE1, 0x80006F03007E24E3, 0x000605AB80006F73,
-    0x00060DAD000605AA, 0x000635AF000605AC, 0x00030086000635AE,
-    0x000585A300032887, 0x000585A580006F8F, 0x80006F9780006F93,
-    0x0006F1B3000585A6, 0x80006F9F80006F9B, 0x80006FA780006FA3,
-    0x0006E1B50006E9B6, 0x0006BDBB80006FAB, 0x80006FAF0006BDBA,
-    0x000689BF80006FB3, 0x000689BD000681BE, 0x000629A8000695BC,
-    0x000639AA000629A9, 0x000629AC000639AB, 0x00065DAE000629AD,
-    0x80006FB700064DAF, 0x80006FBF80006FBB, 0x80006FC780006FC3,
-    0x80006FCF80006FCB, 0x0007E9B880006FD3, 0x0007F9BA0007E9B9,
-    0x0007E9BC0007F9BB, 0x00075DBE0007E9BD, 0x000639B0000755BF,
-    0x000601B2000639B1, 0x000605B400061DB3, 0x000605B600060DB5,
-    0x0007B1A30007D9B7, 0x0000158100002D82, 0x80006FD700001D80,
-    0x0007A1A50007A9A6, 0x0007FDAB80006FDB, 0x80006FDF0007FDAA,
-    0x0007C9AF0002A484, 0x0007C9AD0007C1AE, 0x0001CCBE0007D5AC,
-    0x80006FE3000695B3, 0x0006A9B680006FE7, 0x80006FEF80006FEB,
-    0x000159BA0006ADB5, 0x0000C886000121BB, 0x000121BE00010C87,
-    0x000131BC000129BF, 0x000629A8000131BD, 0x000659AA000629A9,
-    0x000661AC000651AB, 0x0001B5AE00066DAD, 0x0001E0840001A9AF,
-    0x80006FF780006FF3, 0x80006FFF80006FFB, 0x8000700780007003,
-    0x000161B88000700B, 0x000161BA000161B9, 0x000161BC000161BB,
-    0x000161BE000161BD, 0x0001D9B0000161BF, 0x0001A1B20001A1B1,
-    0x0001A1B40001A1B3, 0x000191B60001A9B5, 0x0005D1A3000191B7,
-    0x800070138000700F, 0x8000701B80007017, 0x0005E9A50005EDA6,
-    0x000265AB8000701F, 0x8000702300021DAA, 0x00026DAF80007027,
-    0x000275AD000265AE, 0x8000702B000275AC, 0x800070338000702F,
-    0x8000703B80007037, 0x800070438000703F, 0x00003D8080007047,
-    0x0000198200000981, 0x8000704F8000704B, 0x00033CBE80007057,
-    0x000C00868000705B, 0x8000705F00031487, 0x800070630003D5B3,
-    0x000335B600033DB5, 0x8000706B80007067, 0x000311BA000A8CBF,
-    0x0000B5BC000311BB, 0x0000B5BE0000BDBD, 0x8000706F0000ADBF,
-    0x80007073000174E3, 0x800070770001BCE1, 0x8000707F8000707B,
-    0x8000708780007083, 0x8000708F8000708B, 0x8000709780007093,
-    0x8000709B000274EF, 0x00024DA88000709F, 0x000251AA000251A9,
-    0x0002B9AC0002A9AB, 0x0002A9AE0002B9AD, 0x000D6C840002A9AF,
-    0x800070A7800070A3, 0x800070AF800070AB, 0x800070B7800070B3,
-    0x00016DB8000D74BE, 0x00010DBA000105B9, 0x00011DBC000105BB,
-    0x00010DBE000105BD, 0x0002D9B0000105BF, 0x00016DB20002D9B1,
-    0x00017DB4000165B3, 0x000165B6000165B5, 0x0001B8E1000155B7,
-    0x0000B4E3000750E1, 0x0000A98000077CE3, 0x0000398200000981,
-    0x800070BF800070BB, 0x800070CB800070C7, 0x000000EE800070CF,
-    0x000000EF800070D3, 0x000C6086000664EF, 0x000291A3000CE487,
-    0x000279A5800070D7, 0x800070DF800070DB, 0x800070E3000271A6,
-    0x000255AB800070E7, 0x0001F9AD000255AA, 0x0001E9AF0001F1AC,
-    0x800070C30001F1AE, 0x800070EB80007053, 0x800070F3800070EF,
-    0x800070FB800070F7, 0x000E55A8800070FF, 0x000EA1AA000E59A9,
-    0x000EADAC000EBDAB, 0x000EF9AE000E95AD, 0x000E91B0000EF5AF,
-    0x000E91B2000E91B1, 0x000EB1B4000E91B3, 0x000EB1B6000EB1B5,
-    0x000E91B8000EB1B7, 0x000E95BA000E9DB9, 0x000F59BC000F49BB,
-    0x000F49BE000F59BD, 0x000E09B3000F49BF, 0x8000710780007103,
-    0x8000710F8000710B, 0x000E05B5000E35B6, 0x000E75BB80007113,
-    0x80007117000E75BA, 0x000E55BF8000711B, 0x000E65BD000E55BE,
-    0x8000711F000E65BC, 0x80007123000E4DA3, 0x000E71A680007127,
-    0x8000712F8000712B, 0x000E31AA000E41A5, 0x0003A484000E31AB,
-    0x000E11AE0003A4BE, 0x000E21AC000E11AF, 0x000EA5A8000E21AD,
-    0x000EA5AA000EADA9, 0x000EDDAC000EB9AB, 0x000EC1AE000EC1AD,
-    0x0001ED80000EFDAF, 0x0001F1820001F181, 0x0001908680007133,
-    0x800071370001B487, 0x0001C9B88000713B, 0x0001D9BA0001C9B9,
-    0x0001F9BC0001D1BB, 0x000199BE0001F9BD, 0x000E85B0000195BF,
-    0x000165B200016DB1, 0x000165B400017DB3, 0x000165B600016DB5,
-    0x000E2DB30001F9B7, 0x800071438000713F, 0x8000714B80007147,
-    0x000E55B5000E41B6, 0x0001A9BB8000714F, 0x80007153000E41BA,
-    0x0001A9BF80007157, 0x0001A9BD0001A1BE, 0x8000715B0001B1BC,
-    0x8000715F000E69A3, 0x000E05A680007163, 0x8000716B80007167,
-    0x000E05AA000E11A5, 0x8000716F0001EDAB, 0x0001E5AE80007173,
-    0x0001F5AC0001EDAF, 0x000339A80001EDAD, 0x00038DAA000339A9,
-    0x00039DAC000385AB, 0x000385AE000385AD, 0x800071770003B5AF,
-    0x8000717F8000717B, 0x8000718780007183, 0x8000718F8000718B,
-    0x000061B880007193, 0x000061BA000061B9, 0x000061BC000061BB,
-    0x000061BE000061BD, 0x0003CDB0000061BF, 0x0003A1B20003A5B1,
-    0x0003A5B40003A1B3, 0x000391B60003ADB5, 0x00000D80000391B7,
-    0x0000118200001181, 0x0002F4EF80007197, 0x8000719F8000719B,
-    0x000214E300031CBE, 0x000180E100028884, 0x800071AB800071A7,
-    0x0003C887800071AF, 0x000301BB00043C86, 0x800071B3000369BA,
-    0x000309BF800071B7, 0x000315BD000301BE, 0x000365B3000315BC,
-    0x800071BF800071BB, 0x800071C7800071C3, 0x000375B5000365B6,
-    0x800071CF800071CB, 0x800071D7800071D3, 0x800071DB000289A3,
-    0x000289A6000299A5, 0x00022C84800071DF, 0x000285AA800071E3,
-    0x0002F9AC0002EDAB, 0x0002EDAE0002F9AD, 0x800071E70002E5AF,
-    0x000544BE800071EB, 0x800071EF00054C84, 0x800071F7800071F3,
-    0x800071FF800071FB, 0x8000720780007203, 0x000019808000720B,
-    0x0000058200001981, 0x000F18E18000720F, 0x000EB8E3000E5CE1,
-    0x80007217000174E3, 0x8000721F8000721B, 0x0004008680007223,
-    0x8000722700056487, 0x8000722F8000722B, 0x000E7CEF80007233,
-    0x000175B30001A8EF, 0x8000723B80007237, 0x800072438000723F,
-    0x000115B5000131B6, 0x00011DBB80007247, 0x8000724B000115BA,
-    0x0000E1BF8000724F, 0x0000FDBD0000FDBE, 0x800072130000FDBC,
-    0x8000725780007253, 0x800071A38000725B, 0x800072638000725F,
-    0x00068DA880007267, 0x00069DAA000695A9, 0x0006FDAC0006E5AB,
-    0x0006D1AE0006D1AD, 0x0006B1B00006D1AF, 0x000749B20006B9B1,
-    0x000759B4000749B3, 0x000745B6000745B5, 0x000721B8000779B7,
-    0x000739BA000721B9, 0x000729BC000739BB, 0x000719BE000729BD,
-    0x000635A3000719BF, 0x8000726F8000726B, 0x8000727780007273,
-    0x000655A5000671A6, 0x00065DAB8000727B, 0x8000727F000655AA,
-    0x0007A1AF0003A0BE, 0x0007BDAD0007BDAE, 0x000051800007BDAC,
-    0x0000618200005981, 0x0000F485000755B3, 0x000775B600077DB5,
-    0x001C008680007283, 0x00072DBA0001E487, 0x00073DBC000725BB,
-    0x000725BE000725BD, 0x000691A8000715BF, 0x000691AA000691A9,
-    0x0006B9AC000691AB, 0x0006A9AE0006B9AD, 0x800072870006A9AF,
-    0x8000728F8000728B, 0x000121A280007293, 0x000539A0000135A3,
-    0x000149B8000411A1, 0x00015DBA000149B9, 0x00014DBC000155BB,
-    0x0001D1BE0001D1BD, 0x0006A5B00001D1BF, 0x0006A5B20006ADB1,
-    0x0006ADB40006BDB3, 0x000695B600069DB5, 0x000619A3000179B7,
-    0x8000729700799D8F, 0x8000729F8000729B, 0x000631A5000639A6,
-    0x000669AB800072A3, 0x800072A7000661AA, 0x000659AF800072AB,
-    0x000669AD000669AE, 0x0008899E000671AC, 0x0009C99C0005C19F,
-    0x000DA19A0009C99D, 0x000DB198000CBD9B, 0x0071A196000DB999,
-    0x0075459400714597, 0x0075A1920071B195, 0x0079E5900075B593,
-    0x00725D8A0079CD91, 0x800072AF0072458B, 0x000E758E001C80BE,
-    0x0072B98C000E658F, 0x007A3982000E758D, 0x800072B3007A3983,
-    0x00767186800072B7, 0x007A098400769987, 0x00076D9A00766D85,
-    0x800072BB0002559B, 0x800072C3800072BF, 0x0002599C000090E1,
-    0x000F0992001A08E3, 0x800072C7000A6593, 0x00067596001638EF,
-    0x000A7D9400067597, 0x00168DA9000B3595, 0x001001AB001685A8,
-    0x001271AD001631AA, 0x002EB9AF0012B5AC, 0x00027DA1002C01AE,
-    0x001EA1A3800072CB, 0x001AB1A5001EA9A2, 0x001BE5A7001FF5A4,
-    0x0003CC84001AB1A6, 0x800072CF001C4C84, 0x800072D7800072D3,
-    0x800072DF800072DB, 0x002801B1800072E3, 0x002A8DB3002ED9B0,
-    0x0026F5B5002AE9B2, 0x001D7084002401B4, 0x0001FD80800072E7,
-    0x00001D8200001581, 0x800072EB001C80BE, 0x0002B887800072EF,
-    0x800072F7001D3C86, 0x800072FF800072FB, 0x8000730780007303,
-    0x8000730F8000730B, 0x8000731780007313, 0x8000731F8000731B,
-    0x000380E380007323, 0x0001A0E180007327, 0x000394EF8000732B,
-    0x800073338000732F, 0x8000733B80007337, 0x800073438000733F,
-    0x8000734B80007347, 0x8000734F00063CE1, 0x800073530006D0E3,
-    0x80007357000730E1, 0x000039800006ACE3, 0x00001D8200001581,
-    0x8000735B00061CEF, 0x001FB8BE8000735F, 0x0002E9B30001E8EF,
-    0x001CC48780007367, 0x8000736B001CEC86, 0x000251B5000259B6,
-    0x00024DBB8000736F, 0x8000737300024DBA, 0x0001B9BF80007377,
-    0x0001D1BD0001D9BE, 0x001D29A30001F5BC, 0x80007363800072F3,
-    0x8000737F8000737B, 0x001D91A5001D99A6, 0x001D8DAB80007383,
-    0x80007387001D8DAA, 0x001E79AF8000738B, 0x001E11AD001E19AE,
-    0x8000738F001E35AC, 0x80007393001F6DB3, 0x001F65B680007397,
-    0x8000739F8000739B, 0x001F21BA001F6DB5, 0x800073A3001F21BB,
-    0x001F25BE800073A7, 0x001F31BC001F29BF, 0x001FA1A8001F31BD,
-    0x001FA1AA001FA1A9, 0x001FA1AC001FA1AB, 0x001FA1AE001FA1AD,
-    0x800073AB001FA1AF, 0x800073B3800073AF, 0x800073BB800073B7,
-    0x800073C3800073BF, 0x001FADB8800073C7, 0x001FBDBA001FB5B9,
-    0x001FADBC001FB5BB, 0x001F51BE001F55BD, 0x001FA1B0001F4DBF,
-    0x001FA1B2001FA1B1, 0x001FA5B4001FA1B3, 0x001F9DB6001FA9B5,
-    0x001E29A3001F95B7, 0x0000198100001982, 0x800073CB0001B180,
-    0x001E29A5001E21A6, 0x001E65AB800073CF, 0x00008886001E65AA,
-    0x001E6DAF0001FC87, 0x001E75AD001E61AE, 0x800073D3001E75AC,
-    0x800073D7001E99B3, 0x00015DB6800073DB, 0x800073E3800073DF,
-    0x000149BA001EB1B5, 0x800073E7000149BB, 0x00013DBE800073EB,
-    0x000139BC000121BF, 0x001E45A8000135BD, 0x001E55AA001E55A9,
-    0x001E7DAC001E65AB, 0x000189AE000199AD, 0x0000AC84000181AF,
-    0x800073F3800073EF, 0x800073FB800073F7, 0x80007403800073FF,
-    0x0001ADB880007407, 0x00016DBA000165B9, 0x00017DBC000165BB,
-    0x00016DBE000165BD, 0x0001C9B0000165BF, 0x0001A9B20001C9B1,
-    0x0001BDB40001A5B3, 0x0001A1B60001A1B5, 0x001C69B8000195B7,
-    0x001CC1BA001CA1B9, 0x001CC1BC001CC1BB, 0x001CC1BE001CC1BD,
-    0x001F89B0001CC1BF, 0x001C21B2001F89B1, 0x001C1DB4001C05B3,
-    0x001C75B6001C7DB5, 0x001F61A8001C6DB7, 0x001F61AA001F61A9,
-    0x001FD9AC001F61AB, 0x001FC9AE001FD9AD, 0x8000740B001FC1AF,
-    0x800074138000740F, 0x8000741B80007417, 0x800074238000741F,
-    0x8000742B80007427, 0x000400BE8000742F, 0x80007433001D51A3,
-    0x000295A6001D79A5, 0x8000743B80007437, 0x000281AA8000743F,
-    0x0002F1AC000281AB, 0x0002F5AE0002FDAD, 0x800074470002E9AF,
-    0x8000744B0001F4E1, 0x00004D8000017CE3, 0x00007D8200007581,
-    0x0004C0868000744F, 0x800074530004A487, 0x8000745B80007457,
-    0x800074638000745F, 0x000249A800028CEF, 0x00025DAA000249A9,
-    0x000279AC000255AB, 0x0002BDAE000279AD, 0x0005A0840002B5AF,
-    0x8000746B80007467, 0x0004BCBE8000746F, 0x8000747780007473,
-    0x00012DB88000747B, 0x00013DBA000135B9, 0x00012DBC000135BB,
-    0x0001DDBE0001D5BD, 0x0002CDB00001CDBF, 0x0002DDB20002D5B1,
-    0x0002CDB40002D5B3, 0x00011DB6000115B5, 0x001E84E1000115B7,
-    0x001F5CE3001F6CE3, 0x8000747F001E58E1, 0x8000748780007483,
-    0x8000748F8000748B, 0x8000749780007493, 0x000000E98000749B,
-    0x001E4CEF001EF4EF, 0x000295B38000749F, 0x800074A7800074A3,
-    0x800074AF800074AB, 0x0002B1B50002B9B6, 0x000251BB800074B7,
-    0x0004C886000249BA, 0x000191BF0004AC87, 0x000241BD000249BE,
-    0x80007443000249BC, 0x800074BB000551A3, 0x00057DA6800074B3,
-    0x800074C3800074BF, 0x00058DAA000575A5, 0x800074C7000595AB,
-    0x00058DAE800074CB, 0x00058DAC000655AF, 0x00078980000585AD,
-    0x0007998200078981, 0x800074CF000691B3, 0x0006A9B60006B9B5,
-    0x800074D7800074D3, 0x00074DBA800074DB, 0x00075DBC000745BB,
-    0x000741BE000741BD, 0x000641A8000741BF, 0x000655AA00064DA9,
-    0x00067DAC000665AB, 0x00066DAE000665AD, 0x800074DF000665AF,
-    0x800074E7800074E3, 0x800074EF800074EB, 0x800074F7800074F3,
-    0x000759B8800074FB, 0x000769BA000759B9, 0x000779BC000769BB,
-    0x000765BE000779BD, 0x0007C5B0000719BF, 0x0007C5B20007CDB1,
-    0x0007C9B40007D9B3, 0x000769B60007C9B5, 0x0007DDA3000769B7,
-    0x80007503800074FF, 0x8000750B80007507, 0x0007F5A50007E5A6,
-    0x000609AB8000750F, 0x80007513000601AA, 0x00060DAF80007517,
-    0x00060DAD00060DAE, 0x00006D80000611AC, 0x0000198200000981,
-    0x8000751F8000751B, 0x000190BE00019884, 0x0000C08680007523,
-    0x800075270001E487, 0x8000752F8000752B, 0x8000753780007533,
-    0x00065DA88000753B, 0x00019DAA00018DA9, 0x0001B9AC000195AB,
-    0x0001C9AE0001B9AD, 0x0000A0840001C1AF, 0x800075438000753F,
-    0x8000754B80007547, 0x800075538000754F, 0x000179B880007557,
-    0x0000CDBA000179B9, 0x0000DDBC0000C5BB, 0x0000C5BE0000C5BD,
-    0x000181B00000F5BF, 0x000149B2000181B1, 0x000159B4000149B3,
-    0x000149B6000159B5, 0x000215B3000149B7, 0x8000755F8000755B,
-    0x8000756780007563, 0x000231B5000239B6, 0x000245BB8000756B,
-    0x8000756F000245BA, 0x00029DBF80007573, 0x00029DBD00029DBE,
-    0x003E7C8500029DBC, 0x80007577000251A3, 0x00027DA68000757B,
-    0x800075838000757F, 0x000201AA000275A5, 0x80007587000201AB,
-    0x0002D9AE8000758B, 0x0002D9AC0002D9AF, 0x0000E9800002D9AD,
-    0x000005820000E981, 0x000C00BE8000758F, 0x0003A88780007597,
-    0x8000759B000CBC86, 0x800075A38000759F, 0x800075AB800075A7,
-    0x800075B3800075AF, 0x800075BB800075B7, 0x800075C3800075BF,
-    0x000160E3800075C7, 0x0001A0E1800075CB, 0x000288EF800075CF,
-    0x800075D7800075D3, 0x800075DF800075DB, 0x800075E7800075E3,
-    0x000269A8800075EB, 0x000279AA000269A9, 0x000269AC000279AB,
-    0x0002BDAE000269AD, 0x000CACBE0002B5AF, 0x800075F3800075EF,
-    0x00001D80800075F7, 0x0000A98200000981, 0x000151B8800075FB,
-    0x000161BA000159B9, 0x000119BC000161BB, 0x00010DBE000119BD,
-    0x0002CDB0000105BF, 0x0002DDB20002D5B1, 0x0002CDB40002D5B3,
-    0x000171B6000171B5, 0x0000C4E3000171B7, 0x000678E100077CE1,
-    0x800075FF0006BCE3, 0x000CB886000D1884, 0x000F2CBE000D3C87,
-    0x8000760B80007607, 0x000010EF8000760F, 0x8000761780007613,
-    0x8000761B000674EF, 0x800076238000761F, 0x800076270002BDB3,
-    0x0002A5B60002ADB5, 0x8000762F8000762B, 0x000245BA80007633,
-    0x000245BC00025DBB, 0x000245BE00024DBD, 0x800075930001F9BF,
-    0x000D75A6000D7DA5, 0x8000763780007603, 0x8000763F8000763B,
-    0x000D95AC000D6DA3, 0x000D95AE000D9DAD, 0x80007643000E29AF,
-    0x000D95AA80007647, 0x000EE5B3000D8DAB, 0x8000764F8000764B,
-    0x8000765780007653, 0x000EF5B5000EE5B6, 0x000EA1BB8000765B,
-    0x8000765F000ED9BA, 0x000EA5BF80007663, 0x000EB1BD000EB9BE,
-    0x000E15A8000EB9BC, 0x000E2DAA000E25A9, 0x000E3DAC000E25AB,
-    0x000E2DAE000E25AD, 0x00000D80000E25AF, 0x00001D8200001581,
-    0x8000766B80007667, 0x000330848000766F, 0x000E29B880007673,
-    0x000E39BA000E29B9, 0x000E29BC000E39BB, 0x000FFDBE000E29BD,
-    0x000E5DB0000FF5BF, 0x000E2DB2000E25B1, 0x000E3DB4000E25B3,
-    0x000E25B6000E21B5, 0x000FA5A3000E19B7, 0x0001288680007677,
-    0x8000767B00014C87, 0x000FB5A5000FA5A6, 0x000FE1AB8000767F,
-    0x80007683000F99AA, 0x000FE5AF80007687, 0x000FF1AD000FF9AE,
-    0x8000768B000FF9AC, 0x8000768F000EE9B3, 0x000E91B680007693,
-    0x8000769B80007697, 0x000EB1BA000EE5B5, 0x8000769F000EB9BB,
-    0x000161BE800076A3, 0x000E99BC000161BF, 0x000E25A8000E99BD,
-    0x000E25AA000E2DA9, 0x000E29AC000E39AB, 0x000E5DAE000E55AD,
-    0x800076A7000E55AF, 0x800076AF800076AB, 0x800076B7800076B3,
-    0x800076BF800076BB, 0x0001F5B8800076C3, 0x000181BA000181B9,
-    0x000181BC000181BB, 0x0001B1BE000189BD, 0x000E31B00001B1BF,
-    0x000E09B2000E39B1, 0x0001E1B4000E09B3, 0x0001E1B60001E1B5,
-    0x000DADA30001CDB7, 0x800076CB800076C7, 0x800076D3800076CF,
-    0x000DA1A5000DD5A6, 0x000DFDAB800076D7, 0x800076DB000DF5AA,
-    0x000225AF800076DF, 0x000DDDAD000225AE, 0x00005D80000DDDAC,
-    0x0000658200006D81, 0x00039CBE000351B3, 0x000319B6000379B5,
-    0x0002E084800076E7, 0x00033DBA800076EB, 0x000319BC000335BB,
-    0x0003D9BE000319BD, 0x000385A80003D9BF, 0x000395AA000395A9,
-    0x0003BDAC0003A5AB, 0x0003D1AE0003D5AD, 0x000400860003D1AF,
-    0x0033A0BF00033487, 0x800076F3800076EF, 0x800076FB800076F7,
-    0x000371B8800076FF, 0x000371BA000371B9, 0x0000D5BC000371BB,
-    0x0000D5BE0000DDBD, 0x0003B5B00000CDBF, 0x000381B20003BDB1,
-    0x000351B4000381B3, 0x000351B6000351B5, 0x0003A8EF000351B7,
-    0x8000770780007703, 0x00021C848000770B, 0x800077138000770F,
-    0x00052CBE80007717, 0x8000771F8000771B, 0x000340E380007723,
-    0x000028E180007727, 0x00025DA38000772B, 0x800077338000772F,
-    0x8000773B80007737, 0x000275A5000215A6, 0x000239AB8000773F,
-    0x80007743000231AA, 0x0002D5AF80007747, 0x000215AD0002D5AE,
-    0x000128E3000215AC, 0x000E08E1000F00E1, 0x0000FD80000E08E3,
-    0x0000398200000981, 0x8000774F8000774B, 0x8000775B80007757,
-    0x000EE8EF8000775F, 0x8000776780007763, 0x00016DB3000E64EF,
-    0x0004E8868000776B, 0x8000776F00050C87, 0x00016DB500016DB6,
-    0x0000EDBB80007773, 0x800077770000EDBA, 0x0000D5BF8000777B,
-    0x0000E1BD0000E9BE, 0x00065DA80000E9BC, 0x0006A5AA000661A9,
-    0x0006A5AC0006BDAB, 0x0006A5AE0006ADAD, 0x80007753000719AF,
-    0x800077838000777F, 0x8000778B80007787, 0x800077938000778F,
-    0x000775B880007797, 0x00070DBA000779B9, 0x00071DBC000705BB,
-    0x000705BE000705BD, 0x000769B0000735BF, 0x00077DB2000769B1,
-    0x00076DB4000775B3, 0x000751B6000751B5, 0x00062DA3000751B7,
-    0x8000779F8000779B, 0x800077A7800077A3, 0x00062DA500062DA6,
-    0x0007ADAB800077AB, 0x800077AF0007ADAA, 0x000795AF800077B3,
-    0x0007A1AD0007A9AE, 0x00000D800007A9AC, 0x00001D8200001581,
-    0x800077BB800077B7, 0x00035484800077BF, 0x800077C300005CBE,
-    0x0000E886800077C7, 0x800077CB00034C87, 0x800077D3800077CF,
-    0x800077DB800077D7, 0x800077DF000404E3, 0x800077E30005B4E1,
-    0x800077EB800077E7, 0x800077F3800077EF, 0x800077FB800077F7,
-    0x80007803800077FF, 0x800078070004ECEF, 0x0006EDB38000780B,
-    0x800078138000780F, 0x8000781B80007817, 0x0006E1B5000691B6,
-    0x00068DBB8000781F, 0x8000782300068DBA, 0x000141BF80007827,
-    0x000151BD000159BE, 0x000625A8000159BC, 0x000625AA00062DA9,
-    0x000629AC000639AB, 0x000649AE000651AD, 0x0000CD80000641AF,
-    0x0000198200000981, 0x8000782F8000782B, 0x000034BE00012C84,
-    0x0001FDB880007837, 0x000141BA000141B9, 0x000141BC000141BB,
-    0x000171BE000149BD, 0x000609B0000171BF, 0x0001CDB2000609B1,
-    0x0001DDB40001C5B3, 0x0001CDB60001C5B5, 0x003CA0860001C5B7,
-    0x8000783B00034487, 0x8000783F0005A1A3, 0x0005DDA60005ADA5,
-    0x8000784780007843, 0x0005C1AA003C28BE, 0x000215AC0005C1AB,
-    0x000215AE00021DAD, 0x000341B600020DAF, 0x8000784F8000784B,
-    0x800078530002B1B5, 0x800078570002A1B3, 0x000345BE8000785B,
-    0x000375BC00034DBF, 0x000365BA00034DBD, 0x8000785F00036DBB,
-    0x8000786780007863, 0x800076E38000786B, 0x800078738000786F,
-    0x8000787B80007877, 0x000245A88000787F, 0x00025DAA000255A9,
-    0x00024DAC000255AB, 0x0003B1AE0003B5AD, 0x0003D5B00003ADAF,
-    0x0003D5B20003DDB1, 0x0003F5B40003EDB3, 0x0003F5B60003FDB5,
-    0x0003D9B80003EDB7, 0x0003ADBA0003D9B9, 0x0003BDBC0003A5BB,
-    0x0003A5BE0003A5BD, 0x0003F5A3000395BF, 0x8000788780007883,
-    0x8000788F8000788B, 0x0003E5A5000215A6, 0x000239AB80007893,
-    0x80007897000231AA, 0x000219AF8000789B, 0x000219AD000211AE,
-    0x00006980000221AC, 0x0000058200006981, 0x800078A78000789F,
-    0x800078AF800078AB, 0x00026C8400001CEF, 0x800078B3000194E1,
-    0x800078B70000C8E3, 0x003C5886800078BB, 0x800078BF003D5087,
-    0x800078C7800078C3, 0x800078CB003D8484, 0x800078D3800078CF,
-    0x003C6CBE0001B8EF, 0x800078D7000674E1, 0x800078DB000164E3,
-    0x800078E3800078DF, 0x0001D1B3800078E7, 0x800078EF800078EB,
-    0x800078F7800078F3, 0x0001F9B5000191B6, 0x0001BDBB800078FB,
-    0x800078FF0001BDBA, 0x000175BF80007903, 0x000185BD000175BE,
-    0x003DBDA8000185BC, 0x003E19AA003E0DA9, 0x003E31AC003E11AB,
-    0x003E51AE003E51AD, 0x800078A3003E4DAF, 0x00001D8100001D82,
-    0x8000790700001D80, 0x8000790F8000790B, 0x003ED5B880007913,
-    0x003ED5BA003EDDB9, 0x003F59BC003F49BB, 0x003F49BE003F59BD,
-    0x003E39B0003F41BF, 0x003E11B2003E39B1, 0x003EF1B4003E11B3,
-    0x003EF5B6003EF1B5, 0x003E91A3003EEDB7, 0x0000288680007917,
-    0x8000791B0003C087, 0x003EB9A5003ED1A6, 0x003EFDAB8000791F,
-    0x80007923003EFDAA, 0x003E35AF80007927, 0x003EC5AD003E35AE,
-    0x8000792B003EC5AC, 0x8000792F003F9DB3, 0x003FA5B680007933,
-    0x8000793B80007937, 0x003F69BA003FADB5, 0x8000793F003F75BB,
-    0x003F59BE80007943, 0x003F6DBC003F45BF, 0x80007947003F65BD,
-    0x8000794F8000794B, 0x003C60E380007953, 0x003D00E180007957,
-    0x003DFCEF8000795B, 0x800079638000795F, 0x8000796B80007967,
-    0x800079738000796F, 0x003E19A380007977, 0x0000198100001982,
-    0x8000797B00007180, 0x003E29A5003E21A6, 0x003EF1AB8000797F,
-    0x00012484003EEDAA, 0x003EC1AF000124BE, 0x003EE1AD003EDDAE,
-    0x003ED1A8003EE9AC, 0x003ED1AA003ED1A9, 0x003EE1AC003EE5AB,
-    0x003E19AE003EE1AD, 0x00000086003E19AF, 0x8000798700008487,
-    0x8000798F8000798B, 0x8000799780007993, 0x003E7DB88000799B,
-    0x003E01BA003E01B9, 0x003E01BC003E01BB, 0x003E31BE003E09BD,
-    0x003E69B0003E31BF, 0x003E7DB2003E75B1, 0x003E59B4003E75B3,
-    0x003E4DB6003E45B5, 0x000285A8003E45B7, 0x000295AA000295A9,
-    0x0002BDAC0002A5AB, 0x0002D1AE0002D5AD, 0x8000799F0002D1AF,
-    0x800079A7800079A3, 0x800079AB01E624BF, 0x800079B3800079AF,
-    0x000355B8800079B7, 0x00036DBA000359B9, 0x00037DBC000365BB,
-    0x00036DBE000365BD, 0x0002B5B0000365BF, 0x000281B20002BDB1,
-    0x000371B4000281B3, 0x000371B6000371B5, 0x00021DB3000371B7,
-    0x800079BF800079BB, 0x00038884800079C3, 0x000235B5000255B6,
-    0x000279BB80007833, 0x800079C7000271BA, 0x0003B5BF800079CB,
-    0x000255BD0003B5BE, 0x800079CF000255BC, 0x800079D3000259A3,
-    0x000211A6800079D7, 0x800079DF800079DB, 0x000235AA000271A5,
-    0x800079E300023DAB, 0x0003F1AE800079E7, 0x000211AC0003F1AF,
-    0x0002A9A8000211AD, 0x0002B9AA0002A9A9, 0x0002A9AC0002B9AB,
-    0x000139AE0002A9AD, 0x0001CD80000139AF, 0x0000198200000981,
-    0x800079EF800079EB, 0x800079F70005B8BE, 0x0001E9B8800079FB,
-    0x000189BA0001E9B9, 0x00019DBC000185BB, 0x000181BE000181BD,
-    0x000149B00001B5BF, 0x00015DB2000155B1, 0x00014DB4000155B3,
-    0x0001F1B60001F1B5, 0x000014EF0001F1B7, 0x0005A886800079FF,
-    0x80007A030005DC87, 0x80007A0700045884, 0x80007A0B003EC4EF,
-    0x80007A0F003EC4E1, 0x0000C8E3003E30E3, 0x000128E180007A13,
-    0x00027DB680007A17, 0x80007A1F80007A1B, 0x80007A23000275B5,
-    0x80007A27000265B3, 0x0001DDBE80007A2B, 0x0001DDBC000161BF,
-    0x0001D9BA0001D5BD, 0x80007A2F0001C5BB, 0x0005B1A380007A33,
-    0x80007A37800079F3, 0x80007A3F80007A3B, 0x0005A1A50005A9A6,
-    0x000611AB80007A43, 0x80007A4700060DAA, 0x0006B5AF80007A4B,
-    0x000601AD000609AE, 0x80007A4F000609AC, 0x80007A5780007A53,
-    0x0000198080007A5B, 0x0000058200001981, 0x00036CBE80007A5F,
-    0x0000C88680007A63, 0x80007A6700031C87, 0x80007A6F80007A6B,
-    0x0007B9A880007A73, 0x00070DAA0007B9A9, 0x000709AC00071DAB,
-    0x000731AE000735AD, 0x0003A884000729AF, 0x80007A7B80007A77,
-    0x80007A8380007A7F, 0x80007A8B80007A87, 0x0000C9B880007A8F,
-    0x0000D9BA0000C9B9, 0x0000F9BC0000D1BB, 0x000099BE0000F9BD,
-    0x00075DB0000099BF, 0x000721B2000721B1, 0x000729B400073DB3,
-    0x000701B6000729B5, 0x000685B3000701B7, 0x80007A9780007A93,
-    0x80007A9F80007A9B, 0x000681B50006B5B6, 0x0006E5BB80007AA3,
-    0x80007AA7000699BA, 0x0006EDBF80007AAB, 0x0006E9BD0006EDBE,
-    0x80007AAF0006F5BC, 0x80007AB780007AB3, 0x80007ABF80007ABB,
-    0x80007AC780007AC3, 0x80007ACB000490EF, 0x80007ACF0006DCE1,
-    0x000029800005ECE3, 0x0000118200001581, 0x000605A300017CBE,
-    0x0000288680007AD7, 0x80007ADB00014C87, 0x000601A5000635A6,
-    0x000665AB80007ADF, 0x80007AE3000619AA, 0x00066DAF80007AE7,
-    0x000669AD00066DAE, 0x80007AEB000675AC, 0x80007AEF0001BDB3,
-    0x000179B680007AF3, 0x80007AFB80007AF7, 0x000155BA000179B5,
-    0x80007AFF00015DBB, 0x0000F9BE80007B03, 0x000145BC0000F9BF,
-    0x000271A80000F9BD, 0x000271AA000271A9, 0x0002B5AC000271AB,
-    0x0002B5AE0002BDAD, 0x000CEC840002ADAF, 0x80007B0B80007B07,
-    0x80007B1380007B0F, 0x80007B1B80007B17, 0x000369B880007B1F,
-    0x000309BA000369B9, 0x000319BC000309BB, 0x000309BE000319BD,
-    0x0002D5B0000309BF, 0x0002D5B20002DDB1, 0x000379B4000369B3,
-    0x000369B6000379B5, 0x80007B23000361B7, 0x80007B2B80007B27,
-    0x80007B2F0002F5A3, 0x000231A6000231A5, 0x80007B3780007B33,
-    0x00021DAA80007B3B, 0x00020DAC000215AB, 0x0003B1AE0003B1AD,
-    0x000061800003B1AF, 0x0000058200006181, 0x000CF08680007B3F,
-    0x000C10BE00036087, 0x8000798380007B47, 0x80007B4F80007B4B,
-    0x80007B5780007B53, 0x80007B5F80007B5B, 0x000285A880007B63,
-    0x000295AA000295A9, 0x0002BDAC0002A5AB, 0x0002D1AE0002D5AD,
-    0x80007B670002D1AF, 0x80007B6F80007B6B, 0x80007B7780007B73,
-    0x80007B7F80007B7B, 0x000175B880007B83, 0x000175BA00017DB9,
-    0x0001D5BC0001CDBB, 0x0001C9BE0001DDBD, 0x0002B5B00001C1BF,
-    0x000281B20002BDB1, 0x000155B4000281B3, 0x000155B600015DB5,
-    0x000610E100014DB7, 0x00060CE3000C4884, 0x000C988480007B87,
-    0x80007B8F80007B8B, 0x80007B9780007B93, 0x80007B9F80007B9B,
-    0x0000758180007BA3, 0x000120EF00007580, 0x80007BA700007582,
-    0x80007BAF80007BAB, 0x000EB485000CC0BE, 0x00005CEF000210E1,
-    0x000190E1001600E3, 0x000158E380007BB7, 0x80007BBB00073CEF,
-    0x0008008680007BBF, 0x000D9DB3000CB887, 0x80007BC380007B43,
-    0x80007BCB80007BC7, 0x000DB5B5000DD5B6, 0x000DF9BB80007BCF,
-    0x80007BD3000DF1BA, 0x000E19BF80007BD7, 0x000DD5BD000E11BE,
-    0x80007BDB000DD5BC, 0x80007BDF000DD9A3, 0x000D91A680007BE3,
-    0x80007BEB80007BE7, 0x000DB5AA000DF1A5, 0x80007BEF000DBDAB,
-    0x000E55AE80007BF3, 0x000D91AC000E5DAF, 0x000E5DA8000D91AD,
-    0x000E61AA000E61A9, 0x000E61AC000E61AB, 0x000E61AE000E61AD,
-    0x80007BF7000E61AF, 0x80007BFF80007BFB, 0x80007C0780007C03,
-    0x80007C0F80007C0B, 0x000F4DB880007C13, 0x000F51BA000F51B9,
-    0x000F71BC000F51BB, 0x000F71BE000F71BD, 0x000FC1B0000F71BF,
-    0x000FC1B2000FC1B1, 0x000FC1B4000FC1B3, 0x000FC1B6000FC1B5,
-    0x000FE9B3000FC1B7, 0x000180BE80007C17, 0x80007BB380007C1B,
-    0x000FE9B5000FE1B6, 0x000E05BB80007C1F, 0x80007C27000E05BA,
-    0x000E05BF80007C23, 0x000E15BD000E05BE, 0x00004D81000E15BC,
-    0x000D68EF00004180, 0x0007F08600005182, 0x80007C2B0001D487,
-    0x80007C3380007C2F, 0x80007C3700017084, 0x000EE0E180007C3B,
-    0x000D48E380007C3F, 0x000F69A380007C43, 0x80007C4B80007C47,
-    0x80007C5380007C4F, 0x000F69A5000F61A6, 0x000E85AB80007C57,
-    0x80007C5B000E85AA, 0x000E85AF80007C5F, 0x000E95AD000E85AE,
-    0x80007C63000E95AC, 0x80007C67000E31B3, 0x0001C1B680007C6B,
-    0x80007C7380007C6F, 0x0001CDBA0001D1B5, 0x80007C770001A5BB,
-    0x0001A5BE80007C7B, 0x0001B1BC0001ADBF, 0x0026DD8F0001B1BD,
-    0x80007C7F000DF1A3, 0x000201A680007C83, 0x80007C8B80007C87,
-    0x00020DAA000211A5, 0x80007C8F000265AB, 0x000265AE000420BE,
-    0x000271AC00026DAF, 0x000CA19F000271AD, 0x000A699D000A799E,
-    0x0036B19B0008D19C, 0x0034D1990036759A, 0x00326D970032E198,
-    0x003F359500326596, 0x003E7193003E6194, 0x003A7191003B3592,
-    0x00009580003A7990, 0x0000A18200009D81, 0x000244EF80007C97,
-    0x80007C9B000F74E1, 0x0001D4E3000F1CE3, 0x0001E0E180007C9F,
-    0x0002B5A300015CEF, 0x000EDDA1000041A2, 0x0003A9B5000EB9A0,
-    0x0004C08480007CA3, 0x0005F0860003A1B6, 0x000385B30004E487,
-    0x000371BD80007CA7, 0x000341BF000369BC, 0x80007CAB000371BE,
-    0x000371BB80007AD3, 0x0027DD82000379BA, 0x000784BE003B0183,
-    0x003F11860006C0BE, 0x003B1184003F1987, 0x003E7D8A003A5D85,
-    0x80007CAF0033258B, 0x0035B98E80007CB3, 0x0033358C0037158F,
-    0x0037A1920033818D, 0x0019C4BE0009D993, 0x000DB19680007CB7,
-    0x000B7994000F5197, 0x000F059A000B6195, 0x80007CBB0001919B,
-    0x80007CC380007CBF, 0x00037D9C000000DD, 0x000F48E180007CC7,
-    0x000E1CE380007CCB, 0x80007CD380007CCF, 0x80007CDB80007CD7,
-    0x001741B180007CDF, 0x01EBA9B30017A1B0, 0x01EC01B501E801B2,
-    0x000E8CEF01EB11B4, 0x001FC5A980007CE3, 0x001001AB001C01A8,
-    0x0013D9AD001F91AA, 0x0017D9AF0013D1AC, 0x0002F1A1001305AE,
-    0x0007C9A380007CE7, 0x001801A50002C1A2, 0x001BA1A7000765A4,
-    0x000529A8001BF9A6, 0x000555AA00055DA9, 0x000579AC00056DAB,
-    0x00056DAE000579AD, 0x80007C93000561AF, 0x80007CEF80007CEB,
-    0x00000D8080007CF3, 0x0000B1820000B181, 0x000589B880007CF7,
-    0x00059DBA000589B9, 0x0005B9BC000595BB, 0x000651BE0005B9BD,
-    0x0005E5B0000651BF, 0x0005E5B20005EDB1, 0x0005EDB40005FDB3,
-    0x0005D5B60005DDB5, 0x0005DDA30005BDB7, 0x80007CFF80007CFB,
-    0x80007D0300000C84, 0x0005F1A50005F9A6, 0x000529AB80007D07,
-    0x00009886000521AA, 0x000519AF0000A087, 0x000529AD000529AE,
-    0x80007D0B000531AC, 0x80007D0F000661B3, 0x000621B680007D13,
-    0x80007D1B80007D17, 0x0007ADBA000601B5, 0x80007D1F00078DBB,
-    0x00078DBE80007D23, 0x000795BC000771BF, 0x0005B9BE00078DBD,
-    0x0005B9BC0005B9BF, 0x0005B9BA0005B9BD, 0x0005B9B80005B9BB,
-    0x000549B60005B9B9, 0x00057DB4000549B7, 0x000579B2000575B5,
-    0x000515B0000579B3, 0x00055DAE00057DB1, 0x000545AC00056DAF,
-    0x000AA5AA00055DAD, 0x000ABDA800055DAB, 0x80007D27000AADA9,
-    0x80007D2F80007D2B, 0x80007D3780007D33, 0x80007D3F80007D3B,
-    0x00070DA880007D43, 0x00072DAA00071DA9, 0x00074DAC000749AB,
-    0x0006B1AE000765AD, 0x80007D470006B1AF, 0x80007D4F80007D4B,
-    0x80007D5780007D53, 0x80007D5F80007D5B, 0x000655B880007D63,
-    0x000655BA00065DB9, 0x000635BC000671BB, 0x0001F1BE0001F1BD,
-    0x0006ADB00001F1BF, 0x000685B200068DB1, 0x00068DB400069DB3,
-    0x000675B6000671B5, 0x0004A5A300066DB7, 0x0000158100002D82,
-    0x80007D6700001D80, 0x0004C5A50004E5A6, 0x000549AB80007D6B,
-    0x80007D6F000569AA, 0x0005B5AF80007D77, 0x000549AD000549AE,
-    0x001C6086000551AC, 0x80007D7B00030887, 0x80007D7F000285B3,
-    0x000281B6000281B5, 0x80007D8780007D83, 0x000309BA80007D8B,
-    0x000319BC000309BB, 0x000309BE000319BD, 0x000255AC000309BF,
-    0x000261AE00025DAD, 0x00020DA8000261AF, 0x000251AA000255A9,
-    0x0003AC84000251AB, 0x80007D9380007D8F, 0x001DFC8480007D97,
-    0x80007D9F80007D9B, 0x000371BC80007DA3, 0x000371BE000371BD,
-    0x000371B8000371BF, 0x000371BA000371B9, 0x000391B4000371BB,
-    0x000391B6000391B5, 0x000391B0000391B7, 0x000391B2000391B1,
-    0x80007DA7000391B3, 0x80007DAF80007DAB, 0x80007DB780007DB3,
-    0x80007DBB0001A4E1, 0x001C68BE000180E3, 0x80007DC380007DBF,
-    0x80007DC70006D8EF, 0x80007DCF80007DCB, 0x000389A380007DD3,
-    0x0000158100002D82, 0x80007DD700001D80, 0x00038DA500038DA6,
-    0x000205AB80007DDB, 0x80007DDF000205AA, 0x000205AF80007DE7,
-    0x000215AD000205AE, 0x001C2086000215AC, 0x80007DEB001DC487,
-    0x80007DF380007DEF, 0x80007DFB80007DF7, 0x80007DFF00066CEF,
-    0x80007E0300076CE1, 0x80007E07000774E3, 0x80007E0F80007E0B,
-    0x000191B380007E13, 0x80007E1B80007E17, 0x80007E2380007E1F,
-    0x0001B9B50001B1B6, 0x000155BB80007E27, 0x80007E2B000149BA,
-    0x0000F5BF80007E2F, 0x000145BD0000F5BE, 0x001E51A3000145BC,
-    0x80007E3380007DE3, 0x80007E3B80007E37, 0x001E79A5001E71A6,
-    0x001E95AB80007E3F, 0x80007E43001E89AA, 0x001F35AF80007E47,
-    0x001E85AD001F35AE, 0x00006D80001E85AC, 0x00001D8200001581,
-    0x80007E4B0019FCEF, 0x80007E5380007E4F, 0x0000C08680007E57,
-    0x80007E5B0003AC87, 0x80007E6380007E5F, 0x80007E67001C2CE1,
-    0x001EADA8001CCCE3, 0x001ED9AA001ECDA9, 0x001EF1AC001ED1AB,
-    0x001E3DAE001EF1AD, 0x0000EC84001E35AF, 0x80007E6F80007E6B,
-    0x80007E7780007E73, 0x80007E7F80007E7B, 0x001FD1B880007E83,
-    0x001FEDBA001FDDB9, 0x001FE1BC001FE5BB, 0x001FE1BE001FE1BD,
-    0x001E4DB0001FE1BF, 0x001E51B2001E51B1, 0x001FF1B4001E51B3,
-    0x001FF1B6001FF1B5, 0x001E6DA8001FF1B7, 0x001E85AA001E8DA9,
-    0x001E85AC001E9DAB, 0x001EB9AE001E8DAD, 0x80007E87001EB5AF,
-    0x80007E8F80007E8B, 0x80007E9780007E93, 0x80007E9F80007E9B,
-    0x001E9DB880007EA3, 0x001EA5BA001EADB9, 0x00015DBC000145BB,
-    0x000145BE000145BD, 0x001ED1B0000175BF, 0x001ED1B2001ED1B1,
-    0x001EB5B4001ED1B3, 0x001EB5B6001EBDB5, 0x001E0DB3001EADB7,
-    0x80007EAB80007EA7, 0x80007EB380007EAF, 0x001E0DB5001E0DB6,
-    0x001E15BB80007EB7, 0x80007EBB001E15BA, 0x001E79BF80007EBF,
-    0x001E05BD001E71BE, 0x00006D82001E05BC, 0x00005580001E49A3,
-    0x001E49A600006581, 0x80007EC700019CBE, 0x001E51AA001E49A5,
-    0x00007C87001E51AB, 0x001E35AE00004C86, 0x001E41AC001E3DAF,
-    0x00025DA8001E41AD, 0x000265AA00026DA9, 0x000269AC00027DAB,
-    0x0002B1AE0002B1AD, 0x0004EC840002B1AF, 0x80007ECF80007ECB,
-    0x80007ED780007ED3, 0x80007EDF80007EDB, 0x000371B880007EE3,
-    0x000371BA000371B9, 0x0003D5BC000371BB, 0x0003D5BE0003DDBD,
-    0x0002D1B00003CDBF, 0x0002D1B20002D1B1, 0x000351B40002D1B3,
-    0x000351B6000351B5, 0x0002EDB3000351B7, 0x80007EEB80007EE7,
-    0x80007EEF000480BE, 0x0002E1B5000231B6, 0x000215BB80007EF3,
-    0x80007EF7000215BA, 0x000395BF80007EFB, 0x000205BD000395BE,
-    0x80007EFF000205BC, 0x80007F030002A9A3, 0x000275A680007F07,
-    0x80007F0F80007F0B, 0x000251AA0002A5A5, 0x80007F13000251AB,
-    0x0003D1AE80007F17, 0x000241AC0003D1AF, 0x0002D9A8000241AD,
-    0x000121AA000121A9, 0x000121AC000121AB, 0x000121AE000121AD,
-    0x80007F1B000121AF, 0x80007F2380007F1F, 0x80007F27000420BE,
-    0x80007F2F80007F2B, 0x00018DB880007F37, 0x000191BA000191B9,
-    0x0001BDBC0001A5BB, 0x00007DBE000075BD, 0x0001E5B0000075BF,
-    0x0001F9B20001EDB1, 0x0001D9B40001F1B3, 0x0001B9B60001D9B5,
-    0x001E18E10001B5B7, 0x001F28E380007F3B, 0x0000A58180007F3F,
-    0x80007F430000A580, 0x000400860000A582, 0x80007F470005F487,
-    0x80007F4F80007F4B, 0x001E60EF80007F53, 0x80007F5B80007F57,
-    0x01F4F08580007F5F, 0x80007F6780007F63, 0x80007F6B000164E3,
-    0x80007F6F0001A4E1, 0x80007F730000C8EF, 0x80007D7380007F77,
-    0x80007F7B0005FC84, 0x000629B380007F7F, 0x80007F8380007F33,
-    0x80007F8B80007F87, 0x000629B5000621B6, 0x0006A9BB80007F8F,
-    0x80007F930006A1BA, 0x00069DBF80007F97, 0x0006A5BD00069DBE,
-    0x0007B4E10006ADBC, 0x000478E380007F9B, 0x00007D8080007F9F,
-    0x0000158200001181, 0x0000C08680007FA3, 0x80007FA700035887,
-    0x80007FAF80007FAB, 0x00040CEF80007FB3, 0x0006A9A380007FB7,
-    0x80007FBF80007FBB, 0x80007FC780007FC3, 0x0006A9A50006A1A6,
-    0x000629AB80007FCB, 0x80007FCF000621AA, 0x00061DAF80007FD3,
-    0x000625AD00061DAE, 0x80007FD700062DAC, 0x80007FDB000749B3,
-    0x00077DB680007FDF, 0x80007FE780007FE3, 0x00075DBA000775B5,
-    0x80007FEB000725BB, 0x000721BE80007FEF, 0x00073DBC00072DBF,
-    0x00063DA8000731BD, 0x000685AA000681A9, 0x0006B9AC000695AB,
-    0x0006A9AE0006B9AD, 0x80007FF30006A9AF, 0x80007FFB80007FF7,
-    0x8000800380007FFF, 0x0000B1810000B982, 0x0006ADB80000B980,
-    0x0006B5BA0006BDB9, 0x00015DBC000145BB, 0x000145BE000145BD,
-    0x0006DDB0000175BF, 0x0006ADB20006A5B1, 0x0006A1B40006A1B3,
-    0x000699B60006ADB5, 0x00060DA3000695B7, 0x8000800B80008007,
-    0x000298848000800F, 0x000631A5000639A6, 0x000661AB00019CBE,
-    0x00000886000619AA, 0x000669AF00017C87, 0x000675AD000665AE,
-    0x80008017000679AC, 0x8000801B0001B5B3, 0x000155B68000801F,
-    0x8000802780008023, 0x000171BA0001A1B5, 0x8000802B000179BB,
-    0x000131BE8000802F, 0x000151BC00013DBF, 0x000269A8000151BD,
-    0x000279AA000269A9, 0x00026DAC000279AB, 0x000291AE000291AD,
-    0x80008033000291AF, 0x8000803B80008037, 0x800080438000803F,
-    0x8000804B80008047, 0x000299B88000804F, 0x0002A5BA0002ADB9,
-    0x000375BC00036DBB, 0x000375BE00037DBD, 0x0002F1B000036DBF,
-    0x0002C1B20002F9B1, 0x0002B1B40002C1B3, 0x0002B5B60002BDB5,
-    0x800080530002A9B7, 0x8000805B80008057, 0x8000805F0002FDA3,
-    0x00021DA60002E9A5, 0x8000806780008063, 0x000239AA8000806B,
-    0x000219AC000231AB, 0x000279AE000219AD, 0x8000806F000275AF,
-    0x8000807780008073, 0x00001980000C0084, 0x0000398200000981,
-    0x8000807F8000807B, 0x8000808B80008087, 0x8000808F000CE0BE,
-    0x000D588680008093, 0x0002C1A800030C87, 0x0002C5AA0002CDA9,
-    0x0002C9AC0002D9AB, 0x0002F5AE0002FDAD, 0x8000809700013DAF,
-    0x8000809F8000809B, 0x800080A7800080A3, 0x800080AF800080AB,
-    0x0001C5B8000C30BE, 0x0001D9BA0001CDB9, 0x0001F1BC0001D1BB,
-    0x000199BE0001F1BD, 0x000145B0000199BF, 0x000145B200014DB1,
-    0x000145B400015DB3, 0x000145B600014DB5, 0x000638E10001FDB7,
-    0x00066CE3800080B3, 0x800080BB800080B7, 0x800080C3800080BF,
-    0x000DA884800080C7, 0x800080CF800080CB, 0x000FB0BE800080D3,
-    0x00066CEF800080D7, 0x800080DF800080DB, 0x800080E380007EC3,
-    0x000050E3800080E7, 0x0001A0E1800080EB, 0x000084EF800080EF,
-    0x000D2087800080F7, 0x00003D80000F4C86, 0x0000258200002181,
-    0x000DCDB3800080FB, 0x800080F380008083, 0x80008103800080FF,
-    0x000DC1B5000DFDB6, 0x000E09BB80008107, 0x8000810B000E01BA,
-    0x000E09BF8000810F, 0x000E09BD000E01BE, 0x80008113000E11BC,
-    0x80008117000C98E3, 0x8000811B000FFCE1, 0x800081238000811F,
-    0x8000812B80008127, 0x800081338000812F, 0x8000813B80008137,
-    0x8000813F000C60EF, 0x000D4DA380008143, 0x8000814B80008147,
-    0x800081538000814F, 0x000D41A5000D7DA6, 0x000E89AB80008157,
-    0x8000815B000E81AA, 0x000E89AF8000815F, 0x000E89AD000E81AE,
-    0x00006D82000E91AC, 0x00005580000E35B3, 0x000FF5B600006581,
-    0x800081630003DC84, 0x000FD1BA000FFDB5, 0x00040086000FD1BB,
-    0x000F7DBE0000DC87, 0x000FC1BC000F65BF, 0x000FE5A8000F79BD,
-    0x000FF9AA000FEDA9, 0x000E31AC000FF9AB, 0x000E31AE000E31AD,
-    0x80008167000E31AF, 0x8000816F8000816B, 0x8000817780008173,
-    0x8000817F8000817B, 0x000EDDB880008183, 0x000EE1BA000EE1B9,
-    0x000EE5BC000EE1BB, 0x000E99BE000EE9BD, 0x000E51B0000E95BF,
-    0x000E51B2000E51B1, 0x000EFDB4000EE9B3, 0x000EEDB6000EE5B5,
-    0x000F71A3000EE5B7, 0x8000818B80008187, 0x800081938000818F,
-    0x000EB9A5000EB1A6, 0x000E95AB80008197, 0x8000819B000E95AA,
-    0x000E21AF8000819F, 0x000E3DAD000E39AE, 0x800081A3000E85AC,
-    0x800081A70001C9B3, 0x0001F9B6800081AB, 0x800081B3800081AF,
-    0x0001A9BA0001C1B5, 0x800081B70001B5BB, 0x0001ADBE800081BB,
-    0x0001ADBC000195BF, 0x000DE5A80001A5BD, 0x000205AA000DF9A9,
-    0x00020DAC00021DAB, 0x000235AE00023DAD, 0x800081BF00025DAF,
-    0x800081C7800081C3, 0x00001980800081CB, 0x0000058200001981,
-    0x00022DB8800081CF, 0x000231BA000235B9, 0x0002D5BC000231BB,
-    0x0002D5BE0002DDBD, 0x000229B00002CDBF, 0x00023DB2000235B1,
-    0x00022DB4000235B3, 0x00021DB6000215B5, 0x800081D7000215B7,
-    0x800081DB00029C84, 0x800081DF000281A3, 0x0002B1A6000289A5,
-    0x00042086800081E3, 0x0002E1AA0003D487, 0x0002E5AC0002FDAB,
-    0x0002E5AE0002EDAD, 0x0003F5B60002DDAF, 0x01FD8C85000344BE,
-    0x800081E70003FDB5, 0x800081EB0003FDB3, 0x00037DBE800081EF,
-    0x000365BC00034DBF, 0x000375BA00037DBD, 0x800081F300037DBB,
-    0x800081FB800081F7, 0x00020484800081FF, 0x80008203000225A3,
-    0x00022DA6000225A5, 0x8000820B80008207, 0x0002ADAA8000820F,
-    0x0002BDAC0002A5AB, 0x0002A5AE0002A5AD, 0x80008213000295AF,
-    0x8000821B80008217, 0x800082238000821F, 0x800082270003C0E3,
-    0x8000822B0001ACE1, 0x8000822F000358EF, 0x00000D8080008233,
-    0x0000ED820000E581, 0x000F60E180008237, 0x000E38E3000140E3,
-    0x8000823B000E78E1, 0x000594BE8000823F, 0x0004008680008247,
-    0x8000824B00056487, 0x800082538000824F, 0x000E7CEF000EFCEF,
-    0x000151B380008257, 0x800080138000825B, 0x800082638000825F,
-    0x000109B500010DB6, 0x0000E1BB80008267, 0x8000826B000119BA,
-    0x0000D1BF8000826F, 0x0000E9BD0000E9BE, 0x800081D30000F9BC,
-    0x8000827380008243, 0x8000827B80008277, 0x800082838000827F,
-    0x0006A1A880008287, 0x0006B9AA0006ADA9, 0x0006E1AC0006B1AB,
-    0x0006E5AE0006EDAD, 0x0006BDB00006C5AF, 0x00074DB2000745B1,
-    0x00074DB400075DB3, 0x00077DB6000775B5, 0x000725B800071DB7,
-    0x000725BA00072DB9, 0x000729BC00073DBB, 0x000711BE000715BD,
-    0x000611A3000711BF, 0x8000828F8000828B, 0x8000829780008293,
-    0x000649A500064DA6, 0x0007A1AB8000829B, 0x8000829F000659AA,
-    0x000791AF800082A3, 0x0007A9AD0007A9AE, 0x00000D800007B9AC,
-    0x00001D8200001581, 0x800082AB800082A7, 0x00039484800082AF,
-    0x800082B3000394BE, 0x0000C886800082B7, 0x800082BB00008C87,
-    0x800082C3800082BF, 0x00069DA8800082C7, 0x0006A5AA0006ADA9,
-    0x0006ADAC0006BDAB, 0x0006D5AE0006D1AD, 0x800082CB0006CDAF,
-    0x800082D3800082CF, 0x800082DB800082D7, 0x800082E3800082DF,
-    0x00015DB8800082E7, 0x0001C1BA0001C1B9, 0x0001C1BC0001C1BB,
-    0x0001F1BE0001C9BD, 0x0006BDB00001F1BF, 0x000685B200068DB1,
-    0x00017DB4000165B3, 0x00016DB6000165B5, 0x00062DB3000165B7,
-    0x800082EF800082EB, 0x800082F7800082F3, 0x000625B5000651B6,
-    0x000649BB800082FB, 0x800082FF000679BA, 0x0001A1BF80008303,
-    0x0001B1BD0001B9BE, 0x80008307000651BC, 0x8000830B000669A3,
-    0x000615A68000830F, 0x8000831780008313, 0x00063DAA000661A5,
-    0x8000831B00060DAB, 0x0001FDAE8000831F, 0x000615AC0001E5AF,
-    0x0007D5BA0001F5AD, 0x0007C1B80007DDBB, 0x000431BE0007C1B9,
-    0x0007F1BC000431BF, 0x0007ADB20007F1BD, 0x0007ADB00007B5B3,
-    0x00079DB60007A5B1, 0x0007A5B40007F5B7, 0x000769AA000795B5,
-    0x000769A8000769AB, 0x000769AE000769A9, 0x000769AC000769AF,
-    0x0003B980000769AD, 0x0003858200038D81, 0x01FC50860003A884,
-    0x000334BE00030887, 0x000265A880008327, 0x00027DAA000275A9,
-    0x00026DAC000275AB, 0x0003BDAE0003B5AD, 0x8000832B0003B5AF,
-    0x800083338000832F, 0x8000833B80008337, 0x800083438000833F,
-    0x000351B880008347, 0x000361BA000359B9, 0x000311BC000361BB,
-    0x000315BE00031DBD, 0x0003CDB0000309BF, 0x0003DDB20003D5B1,
-    0x0003CDB40003D5B3, 0x000371B6000371B5, 0x8000834B000371B7,
-    0x0003FDB38000834F, 0x0003EDB580008353, 0x8000835B80008357,
-    0x8000835F00023DB6, 0x000219BB80008363, 0x000209BD000201BA,
-    0x0002BDBF000201BC, 0x80008367000201BE, 0x01FDC4848000836B,
-    0x8000836F01FCC0BE, 0x8000837780008373, 0x8000837B0003DCEF,
-    0x8000837F000194E1, 0x80008383000310E3, 0x0000ED810000ED82,
-    0x000638E10000ED80, 0x000140E300073CE1, 0x80008387000694E3,
-    0x8000838F8000838B, 0x01FC808680008397, 0x8000839B01FDE887,
-    0x800083A38000839F, 0x00019CEF800083A7, 0x000335A30006DCEF,
-    0x800083AF800083AB, 0x800083B7800083B3, 0x000325A50002F5A6,
-    0x0002D1AB800083BB, 0x800083BF0002C9AA, 0x000275AF800083C3,
-    0x0002C1AD0002C9AE, 0x800083930002C9AC, 0x800083CB800083C7,
-    0x01FDA4EF800083CF, 0x800083D7800083D3, 0x01FD78E3800083DB,
-    0x01FC40E1800083DF, 0x800083E7800083E3, 0x800083EF800083EB,
-    0x01FEE5B3800083F3, 0x0000158100002D82, 0x800083F700001D80,
-    0x01FEF5B501FEE5B6, 0x01FF61BB800083FB, 0x800083FF01FED9BA,
-    0x01FF69BF0003E484, 0x01FF75BD01FF61BE, 0x01FEFDA801FF79BC,
-    0x01FE2DAA01FE25A9, 0x01FE3DAC01FE25AB, 0x01FE2DAE01FE25AD,
-    0x00002CBE01FE25AF, 0x0000888680008403, 0x800084070000EC87,
-    0x8000840F8000840B, 0x01FF81B880008413, 0x01FF99BA01FF81B9,
-    0x01FF89BC01FF99BB, 0x01FFB1BE01FFB5BD, 0x01FE5DB001FFADBF,
-    0x01FFEDB201FFE5B1, 0x01FFE5B401FFE1B3, 0x01FFD9B601FFEDB5,
-    0x01FFA5A301FFD9B7, 0x8000841B80008417, 0x800084238000841F,
-    0x01FFB5A501FFA5A6, 0x01FE21AB80008427, 0x8000842B01FF99AA,
-    0x01FE29AF8000842F, 0x01FE35AD01FE21AE, 0x8000843301FE39AC,
-    0x8000843701FEE9B3, 0x01FE95B68000843B, 0x800084438000843F,
-    0x01FEB1BA01FEE9B5, 0x8000844701FEB9BB, 0x000175BE8000844B,
-    0x01FE91BC00017DBF, 0x01FE1DA801FE91BD, 0x01FE25AA01FE2DA9,
-    0x01FE25AC01FE3DAB, 0x01FE51AE01FE51AD, 0x8000844F01FE51AF,
-    0x8000845780008453, 0x8000845F8000845B, 0x8000846780008463,
-    0x0001D9B88000846B, 0x0001EDBA0001D9B9, 0x0001E1BC0001E1BB,
-    0x0001E1BE0001E1BD, 0x01FE31B00001E1BF, 0x01FE01B201FE39B1,
-    0x0001F5B401FE01B3, 0x0001F5B60001FDB5, 0x01FDADA30001E9B7,
-    0x00034CBE8000846F, 0x8000847B80008477, 0x01FDADA501FDD1A6,
-    0x01FDFDAB8000847F, 0x8000848301FDF5AA, 0x000239AF80008487,
-    0x01FDD5AD000231AE, 0x0003F98001FDD5AC, 0x00004D820003F981,
-    0x8000848B00207485, 0x0004D4840003D884, 0x000440868000848F,
-    0x8000849300035487, 0x8000849B80008497, 0x800084A38000849F,
-    0x000231A80005F0BE, 0x000231AA000231A9, 0x000395AC000231AB,
-    0x000395AE00039DAD, 0x800084A700038DAF, 0x800084AF800084AB,
-    0x00077C84800084B3, 0x800084BB800084B7, 0x0003A9B8800084BF,
-    0x000369BA0003A9B9, 0x000379BC000369BB, 0x000369BE000379BD,
-    0x0003FDB0000369BF, 0x0003C5B20003CDB1, 0x0003B9B40003ADB3,
-    0x0003ADB60003B9B5, 0x800084C30003A5B7, 0x800084CB800084C7,
-    0x800084D3800084CF, 0x0003E8EF800084D7, 0x000190E1800084DB,
-    0x00036CE3800084DF, 0x00002580800084E7, 0x00001D8200002981,
-    0x0003F9B3800084EB, 0x00076886800084EF, 0x800084F300055C87,
-    0x00025DB500025DB6, 0x000249BB800084F7, 0x800084FB000279BA,
-    0x00023DBF800084FF, 0x000239BD00023DBE, 0x80008503000251BC,
-    0x000640BE01FE3CE1, 0x800085070001F0E3, 0x8000850F8000850B,
-    0x8000851780008513, 0x8000851F8000851B, 0x8000852380008323,
-    0x8000852B80008527, 0x8000852F0001E4EF, 0x000371A380008533,
-    0x0002D5A580008537, 0x8000853F8000853B, 0x800085430002D5A6,
-    0x0002C1AB80008547, 0x0002B1AD0002F1AA, 0x0002B5AF0002D9AC,
-    0x01FCDCE10002B5AE, 0x000450E3800084E3, 0x0007F580000488EF,
-    0x00003D8200000981, 0x000124848000854B, 0x800085538000854F,
-    0x8000855B80008557, 0x001CE4EF00044CE1, 0x00065DB3000474E3,
-    0x000C00868000855F, 0x800085630003E087, 0x000675B5000605B6,
-    0x000611BB80008567, 0x8000856B000609BA, 0x0006D5BF8000856F,
-    0x000601BD0006D5BE, 0x00068DA8000609BC, 0x000695AA000695A9,
-    0x0006BDAC0006A5AB, 0x0006C5AE0006C5AD, 0x800085730006F5AF,
-    0x8000857B80008577, 0x800085838000857F, 0x8000858B80008587,
-    0x000675B88000858F, 0x000675BA00067DB9, 0x0007D5BC0007CDBB,
-    0x0007D5BE0007DDBD, 0x00068DB00007CDBF, 0x00069DB2000695B1,
-    0x000651B4000695B3, 0x000651B6000651B5, 0x00071DA3000651B7,
-    0x8000859301EC158F, 0x8000859B80008597, 0x000735A5000745A6,
-    0x000751AB8000859F, 0x800085A3000749AA, 0x000795AF800085A7,
-    0x000741AD000795AE, 0x01F9459E000749AC, 0x01FD919C01F9A59F,
-    0x01FD559A01F94D9D, 0x01F1059801FD5D9B, 0x01F2A59601FE9599,
-    0x01F56D9401F16197, 0x01E9E19201F53595, 0x01EC559001F68593,
-    0x001DB1B101E95D91, 0x001989B3001DA5B0, 0x002401B5001C41B2,
-    0x800085AB0019F5B4, 0x800085B3800085AF, 0x0003A1810003A982,
-    0x000515A200006980, 0x000615A0000205A3, 0x0001C5A10005BDA1,
-    0x000DCDA3800085B7, 0x000801A50001E5A2, 0x0009D1A7000DDDA4,
-    0x001401A90009D9A6, 0x0014C5AB0008A5A8, 0x0011D1AD0015CDAA,
-    0x001C01AF001001AC, 0x01EF1182001175AE, 0x800085BB01EF2983,
-    0x01F5E186800085BF, 0x01EA398401F61187, 0x01F6798A01EA8D85,
-    0x000CA4BE01F1E18B, 0x01FDF98E800085C7, 0x01F23D8C01FE318F,
-    0x01FE499201F2618D, 0x000C088701FE9D93, 0x01FAB196000C6C86,
-    0x01FA59940005F997, 0x0006619A01FA5995, 0x800085CB0006C19B,
-    0x800085D3800085CF, 0x0001C99C800085D7, 0x0005ADA8800085DB,
-    0x0005CDAA0005B9A9, 0x0005CDAC0005DDAB, 0x0005F5AE0005FDAD,
-    0x800085DF00051DAF, 0x800085E7800085E3, 0x800085EF800085EB,
-    0x800085F7800085F3, 0x000575B8800085FB, 0x000509BA00057DB9,
-    0x00051DBC000509BB, 0x000501BE000501BD, 0x000571B000053DBF,
-    0x000571B2000571B1, 0x000551B4000571B3, 0x000551B6000551B5,
-    0x000445B300054DB7, 0x80008603800085FF, 0x8000860B80008607,
-    0x000445B500047DB6, 0x0004B5BB8000860F, 0x800086130004B5BA,
-    0x000495BF80008617, 0x0004A5BD000495BE, 0x8000861B0004A5BC,
-    0x8000861F000401A3, 0x000439A680008623, 0x8000862B80008627,
-    0x0004F1AA000401A5, 0x8000862F0004F1AB, 0x0004D1AE000DEC84,
-    0x0004E1AC0004D1AF, 0x0006D0E10004E1AD, 0x000728E3000C0C84,
-    0x80008637000C00BE, 0x000CB886000344EF, 0x8000863B000D2C87,
-    0x8000863F000194E3, 0x800084730001FCE1, 0x0006C8EF80008643,
-    0x8000864B80008647, 0x00038DB38000864F, 0x00038DB580008653,
-    0x8000865B80008657, 0x8000865F0003B5B6, 0x000341BB80008663,
-    0x000341BD000349BA, 0x000341BF000359BC, 0x000C45A3000349BE,
-    0x80008633800085C3, 0x8000866B80008667, 0x000C45A5000C7DA6,
-    0x000C89AB8000866F, 0x80008673000C81AA, 0x000C89AF80008677,
-    0x000C89AD000C81AE, 0x000F1580000C91AC, 0x000F2182000F1D81,
-    0x8000867B000E21B3, 0x000E25B6000E21B5, 0x800086838000867F,
-    0x000EC1BA80008687, 0x000EC1BC000EC1BB, 0x000EC1BE000EC1BD,
-    0x000EADA8000EC1BF, 0x000EE5AA000EEDA9, 0x000EE5AC000EFDAB,
-    0x000E39AE000EE9AD, 0x8000868B000E39AF, 0x800086938000868F,
-    0x00001D8080008697, 0x0001BD8200000981, 0x000FCDB88000869B,
-    0x000FD5BA000FD5B9, 0x000FFDBC000FE5BB, 0x000F91BE000F95BD,
-    0x000E49B0000F91BF, 0x000E59B2000E49B1, 0x000E49B4000E59B3,
-    0x000FFDB6000E49B5, 0x000F6DA3000FF5B7, 0x000184BE8000869F,
-    0x800086AB800086A7, 0x000F6DA5000F69A6, 0x000F8DAB800086AF,
-    0x00008886000F8DAA, 0x000F8DAF00016C87, 0x000F8DAD000F8DAE,
-    0x800086B3000F8DAC, 0x800086B7000EEDB3, 0x000E91B6800086BB,
-    0x800086C3800086BF, 0x000EB5BA000EE1B5, 0x800086C7000EBDBB,
-    0x00017DBE800086CB, 0x000E99BC000165BF, 0x000E11A8000E99BD,
-    0x000E21AA000E25A9, 0x000E2DAC000E39AB, 0x000E5DAE000E55AD,
-    0x0000A884000E51AF, 0x800086D3800086CF, 0x800086DB800086D7,
-    0x800086E3800086DF, 0x0001EDB8800086E7, 0x000195BA000195B9,
-    0x0001B5BC0001ADBB, 0x00017DBE000175BD, 0x000E35B0000175BF,
-    0x000E09B2000E3DB1, 0x0001FDB4000E09B3, 0x0001E5B60001E5B5,
-    0x000DA9A30001D5B7, 0x800086EF800086EB, 0x800086F7800086F3,
-    0x000DA5A5000DD5A6, 0x000DF9AB800086FB, 0x800086FF000DF1AA,
-    0x000221AF80008703, 0x000DDDAD000239AE, 0x00000D80000DDDAC,
-    0x00001D8200001581, 0x8000870B80008707, 0x000390878000870F,
-    0x0004ECBE00047C86, 0x8000871B80008717, 0x800087238000871F,
-    0x8000872B80008727, 0x000E2DB28000872F, 0x000E2DB0000E35B3,
-    0x000E2DB6000E25B1, 0x000E2DB4000F9DB7, 0x000FBDBA000E25B5,
-    0x000FA5B8000F8DBB, 0x000F2DBE000FBDB9, 0x000F95BC000F15BF,
-    0x80008733000F25BD, 0x8000873B80008737, 0x800087438000873F,
-    0x8000874B80008747, 0x000EA5AA8000874F, 0x000EC5A8000EB5AB,
-    0x000E9DAE000EDDA9, 0x000EA5AC000E55AF, 0x00020DA8000E95AD,
-    0x000215AA000215A9, 0x000259AC00024DAB, 0x00024DAE000259AD,
-    0x0005A884000245AF, 0x8000875780008753, 0x0004B8848000875B,
-    0x800087638000875F, 0x0002FDB880008767, 0x000141BA000141B9,
-    0x000141BC000141BB, 0x000171BE000149BD, 0x000209B0000171BF,
-    0x0002CDB2000209B1, 0x0002DDB40002C5B3, 0x0002CDB60002C5B5,
-    0x000FD4E10002C5B7, 0x000EF4E3000E10E3, 0x8000876B000E0CE1,
-    0x800087738000876F, 0x8000877B80008777, 0x000424BE8000877F,
-    0x000000EC80008787, 0x000ECCEF000044EF, 0x000065828000878B,
-    0x000055800002D9B3, 0x00029DB600006D81, 0x800087938000878F,
-    0x000289BA000295B5, 0x0004A886000289BB, 0x00025DBE0004E087,
-    0x00025DBC000245BF, 0x00051DA3000255BD, 0x8000878380008713,
-    0x8000879B80008797, 0x000551A5000559A6, 0x00054DAB8000879F,
-    0x800087A300054DAA, 0x000581AF800087A7, 0x000591AD000599AE,
-    0x800087AB000599AC, 0x800087AF000629B3, 0x000629B6800087B3,
-    0x800087BB800087B7, 0x0006A5BA000629B5, 0x800087BF0006ADBB,
-    0x0006A5BE800087C3, 0x0006B5BC0006ADBF, 0x0006E5A80006ADBD,
-    0x0006E5AA0006EDA9, 0x0006E5AC0006FDAB, 0x0006E5AE0006EDAD,
-    0x800087C700065DAF, 0x800087CF800087CB, 0x800087D7800087D3,
-    0x800087DF800087DB, 0x0007E9B8800087E3, 0x0007FDBA0007E9B9,
-    0x0007EDBC0007F5BB, 0x00074DBE000745BD, 0x000625B0000745BF,
-    0x000625B200062DB1, 0x000625B400063DB3, 0x000625B600062DB5,
-    0x00076DA30007D9B7, 0x0000158100002D82, 0x800087E700001D80,
-    0x00076DA500076DA6, 0x0007E9AB800087EB, 0x800087EF0007E1AA,
-    0x0007E9AF0001A0BE, 0x0007E9AD0007E1AE, 0x800087F30007F1AC,
-    0x0000E886000691B3, 0x000141B600012C87, 0x800087FB800087F7,
-    0x00014DBA000151B5, 0x800087FF000125BB, 0x000125BE80008803,
-    0x000131BC00012DBF, 0x0001ADB0000131BD, 0x0001C1B20001C5B1,
-    0x0001C5B40001C1B3, 0x0001F1B60001CDB5, 0x000101B80001F1B7,
-    0x000101BA000101B9, 0x000101BC000101BB, 0x000101BE000101BD,
-    0x80008807000101BF, 0x8000880F8000880B, 0x800086A380008813,
-    0x8000881B80008817, 0x00064DA88000881F, 0x00063DAA000659A9,
-    0x0001FDAC000635AB, 0x0001E5AE0001E5AD, 0x0005D5A30001D5AF,
-    0x8000882780008823, 0x8000882F8000882B, 0x000215A5000205A6,
-    0x000261AB80008833, 0x80008837000209AA, 0x000269AF8000883B,
-    0x000275AD000261AE, 0x8000883F000275AC, 0x8000884780008843,
-    0x8000884F8000884B, 0x80008853000564E1, 0x000011800005ECE3,
-    0x0000118200001181, 0x800088570006F4EF, 0x8000885F8000885B,
-    0x0002C484000330BE, 0x0003E1B380008867, 0x000351B5001CC084,
-    0x8000886F8000886B, 0x80008873000359B6, 0x000371BB80008877,
-    0x0000B5BD000379BA, 0x0000B5BF000369BC, 0x8000887B0000B5BE,
-    0x8000887F000194E1, 0x001C708600025CE3, 0x8000888300034487,
-    0x8000888B80008887, 0x800088938000888F, 0x8000889B80008897,
-    0x800088A38000889F, 0x000255A8000278EF, 0x000261AA00025DA9,
-    0x0002D1AC000261AB, 0x0002D1AE0002D1AD, 0x800088A70002D1AF,
-    0x800088AF800088AB, 0x800088B7800088B3, 0x800088BF800088BB,
-    0x000169B8800088C3, 0x000109BA000169B9, 0x000119BC000109BB,
-    0x000109BE000119BD, 0x0002B5B0000105BF, 0x0002B5B20002BDB1,
-    0x000179B4000169B3, 0x000169B6000179B5, 0x0007C4E1000161B7,
-    0x0006B8E30006A4E3, 0x00000D8000067CE1, 0x00003D8200003581,
-    0x800088CB800088C7, 0x001DE0BE800088CF, 0x800088DB800088D7,
-    0x0006E8EF000060EF, 0x0002A9A3800088DF, 0x800088E7800088E3,
-    0x800088EF800088EB, 0x000219A5000211A6, 0x000239AB800088F3,
-    0x001C4886000231AA, 0x0001FDAF001CCC87, 0x0001FDAD0001FDAE,
-    0x001E85A8000221AC, 0x001E91AA001E91A9, 0x001EB5AC001EADAB,
-    0x001ED1AE001ED5AD, 0x800088D3001EC5AF, 0x800088FB800088F7,
-    0x80008903800088FF, 0x8000890B80008907, 0x001F61B88000890F,
-    0x001F61BA001F61B9, 0x001F61BC001F61BB, 0x001F61BE001F61BD,
-    0x001EBDB0001F61BF, 0x001E8DB2001E85B1, 0x001E9DB4001E85B3,
-    0x001E8DB6001E85B5, 0x001E19B3001E85B7, 0x8000891780008913,
-    0x8000891F8000891B, 0x001E3DB5001E55B6, 0x001E41BB80008923,
-    0x80008927001E79BA, 0x001E41BF8000892B, 0x001E51BD001E59BE,
-    0x8000892F001E59BC, 0x80008933001E5DA3, 0x001E11A680008937,
-    0x8000893F8000893B, 0x001E3DAA001E79A5, 0x0003A484001E05AB,
-    0x001E1DAE0003A8BE, 0x001E1DAC001E05AF, 0x001EADA8001E15AD,
-    0x001EB5AA001EB5A9, 0x001ED9AC001EC9AB, 0x001EC9AE001ED9AD,
-    0x0001ED80001EC1AF, 0x0001F1820001F181, 0x0000908680008943,
-    0x8000894700017487, 0x0001C9B88000894B, 0x0001D9BA0001C9B9,
-    0x0001F9BC0001D1BB, 0x000199BE0001F9BD, 0x000145B0000195BF,
-    0x000145B200014DB1, 0x000145B400015DB3, 0x000145B600014DB5,
-    0x001E3DB30001F9B7, 0x800089538000894F, 0x8000895B80008957,
-    0x001E51B5001E59B6, 0x000189BB8000895F, 0x80008963000181BA,
-    0x000189BF80008967, 0x000189BD000181BE, 0x8000896B000191BC,
-    0x001E75A38000896F, 0x001E19A580008973, 0x8000897B80008977,
-    0x80008863001E11A6, 0x0001C1AB8000897F, 0x0001C1AD0001C9AA,
-    0x0001C1AF0001D9AC, 0x800089830001C9AE, 0x8000898B80008987,
-    0x800089938000898F, 0x8000899700021884, 0x8000899F8000899B,
-    0x800089A7800089A3, 0x800089AF800089AB, 0x800089BB800089B7,
-    0x00006980000370BE, 0x0000798200006981, 0x0004F08600040084,
-    0x800089BF00037487, 0x800089C3001F0CEF, 0x800089C7001EACE1,
-    0x800089CB001EF0E3, 0x800089D3800089CF, 0x800089DB800089D7,
-    0x800089E3800089DF, 0x000254EF800089E7, 0x800089EF800089EB,
-    0x0002A4E3800089F3, 0x000180E1800089F7, 0x800089FF800089FB,
-    0x80008A0780008A03, 0x000345B380008A0B, 0x80008A1380008A0F,
-    0x80008A1B80008A17, 0x000355B5000345B6, 0x000321BB80008A1F,
-    0x0004A0BE000349BA, 0x000329BF80008A27, 0x000335BD000321BE,
-    0x000239A8000339BC, 0x00028DAA000239A9, 0x00029DAC000285AB,
-    0x000285AE000285AD, 0x0001ED800002B5AF, 0x0001F5820001F581,
-    0x0004908680008A2B, 0x80008A2F00050487, 0x000145B880008A33,
-    0x000145BA00014DB9, 0x000149BC00015DBB, 0x000179BE000149BD,
-    0x0002CDB0000179BF, 0x0002ADB20002A5B1, 0x0002A5B40002A1B3,
-    0x00029DB60002ADB5, 0x80008A3700017DB7, 0x80008A3F80008A3B,
-    0x80008A4780008A43, 0x80008A4F80008A4B, 0x0004EC84000188EF,
-    0x80008A53001E54E1, 0x80008A57000154E3, 0x80008A5F80008A5B,
-    0x000245A380008A63, 0x80008A6B80008A67, 0x80008A7380008A6F,
-    0x000255A5000245A6, 0x000221AB80008A77, 0x80008A7B000249AA,
-    0x000229AF80008A7F, 0x000235AD000221AE, 0x000635A8000239AC,
-    0x000651AA00063DA9, 0x000679AC00066DAB, 0x00066DAE000665AD,
-    0x80008A23000665AF, 0x80008A8780008A83, 0x0000198080008A8B,
-    0x0000058200001981, 0x0006E5B880008A8F, 0x0006F9BA0006E9B9,
-    0x0006E9BC0006F9BB, 0x00069DBE0006E9BD, 0x00061DB0000695BF,
-    0x0006EDB20006E5B1, 0x0006FDB40006E5B3, 0x0006E1B60006E1B5,
-    0x0006D9B30006E1B7, 0x80008A930003D0BE, 0x80008A9B80008A97,
-    0x0006F1B50006E5B6, 0x000621BB80008A9F, 0x000098860006D9BA,
-    0x000625BF00037887, 0x000631BD000639BE, 0x80008AA3000639BC,
-    0x80008AA700069DA3, 0x0006A1A680008AAB, 0x80008AB380008AAF,
-    0x00069DAA0006B5A5, 0x80008AB7000665AB, 0x00067DAE80008ABB,
-    0x00067DAC000661AF, 0x0007EDA8000675AD, 0x000631AA000629A9,
-    0x000691AC000631AB, 0x000691AE000691AD, 0x80008ABF000691AF,
-    0x80008AC780008AC3, 0x80008ACF80008ACB, 0x80008AD780008AD3,
-    0x000685B880008ADB, 0x000685BA00068DB9, 0x00068DBC00069DBB,
-    0x0006B5BE0006BDBD, 0x0006F1B0000179BF, 0x0006F1B20006F1B1,
-    0x0006C1B40006C5B3, 0x0006C1B60006C1B5, 0x0006B9B30006C1B7,
-    0x80008AE380008ADF, 0x80008AEB80008AE7, 0x000619B5000611B6,
-    0x000635BB80008AEF, 0x80008AF3000635BA, 0x000605BF80008AF7,
-    0x000611BD000605BE, 0x000641A5000625BC, 0x80008AFF80008AFB,
-    0x00001581000649A6, 0x0006E1A300007980, 0x000649AD00001582,
-    0x00065DAF00067DAC, 0x0001348400065DAE, 0x00066DAB800089B3,
-    0x0003CCBE00066DAA, 0x000295B380008B07, 0x0002D9B580008B0B,
-    0x80008B1380008B0F, 0x000C80860002D1B6, 0x0002C5BB00033887,
-    0x000315BD0002C5BA, 0x000315BF000315BC, 0x80008B17000315BE,
-    0x000658EF80008B1B, 0x80008B1F00034084, 0x80008B2780008B23,
-    0x80008B2F80008B2B, 0x80008B3780008B33, 0x000638E180008B3B,
-    0x000624E380008B3F, 0x000249AC000C70BE, 0x00025DAE000249AD,
-    0x00021DA8000255AF, 0x000205AA000205A9, 0x000CA88400025DAB,
-    0x80008B4780008B43, 0x000DBCBE80008B4B, 0x80008B5380008B4F,
-    0x00034DBC80008B57, 0x000355BE000355BD, 0x0003E9B8000365BF,
-    0x00035DBA0003E9B9, 0x0003C9B4000355BB, 0x0003D9B60003C9B5,
-    0x000219B00003D9B7, 0x0003D9B2000219B1, 0x80008B5B0003D9B3,
-    0x80008B5F0000E4E3, 0x80008B630001BCE1, 0x00003D8100003D82,
-    0x80008B6700003D80, 0x80008B6F80008B6B, 0x80008B7B80008B77,
-    0x80008B7F0003CCEF, 0x0003DDA380008B83, 0x000CE88680008B87,
-    0x80008B8B000D6087, 0x000391A5000399A6, 0x00038DAB80008B8F,
-    0x80008B9300038DAA, 0x00025DAF80008B97, 0x00025DAD00025DAE,
-    0x80008B9B00025DAC, 0x80008BA380008B9F, 0x80008BAB80008BA7,
-    0x80008BB380008BAF, 0x000CBC840001E0EF, 0x80008BB700068CE1,
-    0x80008BBB00061CE3, 0x80008BC380008BBF, 0x0001D5B380008BC7,
-    0x80008BCF80008BCB, 0x80008BD780008BD3, 0x000199B5000191B6,
-    0x0001BDBB80008BDB, 0x80008BDF0001BDBA, 0x00019DBF80008BE7,
-    0x00019DBD00019DBE, 0x000E05A800019DBC, 0x000E1DAA000E09A9,
-    0x000E51AC000E15AB, 0x000E49AE000E51AD, 0x80008B73000E49AF,
-    0x000FF581000FCD82, 0x80008BE3000FFD80, 0x00001C8680008BEB,
-    0x000EE9B80003B087, 0x000EFDBA000EE9B9, 0x000EEDBC000EF5BB,
-    0x000F51BE000F55BD, 0x000E39B0000F4DBF, 0x000E09B2000E39B1,
-    0x000E19B4000E09B3, 0x000E0DB6000E19B5, 0x000E95A3000ED9B7,
-    0x80008BF380008BEF, 0x80008BFB80008BF7, 0x000ED9A5000ED1A6,
-    0x000EFDAB80008BFF, 0x80008C03000EFDAA, 0x000EDDAF80008C07,
-    0x000EDDAD000EDDAE, 0x80008C0B000EDDAC, 0x80008C0F000FFDB3,
-    0x000F81B680008C13, 0x80008C1B80008C17, 0x000F8DBA000F99B5,
-    0x80008C1F000F65BB, 0x000F7DBE80008C23, 0x000F7DBC000F65BF,
-    0x000E2DA8000F75BD, 0x000E31AA000E35A9, 0x000E55AC000E31AB,
-    0x000E45AE000E45AD, 0x80008C27000E75AF, 0x80008C2F80008C2B,
-    0x80008C3780008C33, 0x80008C3F80008C3B, 0x000EE9B880008C43,
-    0x000EFDBA000EF5B9, 0x000EEDBC000EF5BB, 0x000E91BE000E91BD,
-    0x000E0DB0000E91BF, 0x000EEDB2000EE5B1, 0x000EFDB4000EE5B3,
-    0x000EEDB6000EE5B5, 0x000EB9A3000EE5B7, 0x0000158100002D82,
-    0x80008C4700001D80, 0x000EDDA5000EC5A6, 0x000E21AB80008C4B,
-    0x80008C4F000EC9AA, 0x000E21AF00012CBE, 0x000E31AD000E39AE,
-    0x80008C57000E39AC, 0x000145B5000155B6, 0x000155B380008C53,
-    0x0000788680008C5B, 0x000139BF00005C87, 0x000125BD000131BE,
-    0x000131BB000125BC, 0x80008B03000159BA, 0x80008C6380008C5F,
-    0x0003048480008C67, 0x80008C6B000289A3, 0x000289A6000299A5,
-    0x003928BF80008C6F, 0x000285AA80008C73, 0x0002F9AC0002EDAB,
-    0x0002EDAE0002F9AD, 0x000258E30002E5AF, 0x000188E1000EC0EF,
-    0x80008C7B80008C77, 0x80008C8380008C7F, 0x80008C8B80008C87,
-    0x80008C9380008C8F, 0x000228EF80008C97, 0x80008C9B000E28E3,
-    0x0002B5A9000E44E1, 0x000201AB000D69A8, 0x000201AD000209AA,
-    0x000231AF000219AC, 0x000400BE000201AE, 0x80008CA380008C9F,
-    0x80008CAB80008CA7, 0x80008CB380008CAF, 0x0003E5B980008CB7,
-    0x0003E5BB0003E5B8, 0x0003E5BD0003E5BA, 0x0003E5BF0003E5BC,
-    0x000221B10003E5BE, 0x000225B3000249B0, 0x000229B5000225B2,
-    0x000215B7000221B4, 0x0002C1A8000215B6, 0x0002F5AA0002D1A9,
-    0x000115AC00010DAB, 0x000105AE000105AD, 0x80008CBB000139AF,
-    0x80008CC780008CBF, 0x80008CCF80008CCB, 0x80008CD780008CD3,
-    0x00012DB880008CDB, 0x0001EDBA00013DB9, 0x0001FDBC0001E5BB,
-    0x0001EDBE0001E5BD, 0x00012DB00001E5BF, 0x00013DB2000135B1,
-    0x00012DB4000135B3, 0x00011DB6000115B5, 0x00013D80000115B7,
-    0x0000AD820000A581, 0x0005B0860000D8EF, 0x80008CDF0005D887,
-    0x00046C84000FD4EF, 0x80008CE3000EF4E1, 0x80008CE7000F1CE3,
-    0x80008CEB000194E1, 0x00023DB3000E0CE3, 0x80008CF380008CEF,
-    0x80008CFB80008CF7, 0x0001DDB50001C5B6, 0x0001B1BB80008CFF,
-    0x80008D030001A9BA, 0x000191BF80008D07, 0x0001A1BD0001A9BE,
-    0x80008CC30001A9BC, 0x000645AA80008D0B, 0x000645AC00065DAB,
-    0x000645AE00064DAD, 0x80008D0F00067DAF, 0x80008D1780008D13,
-    0x80008D1B0005D1A3, 0x000629A6000631A5, 0x80008D2380008D1F,
-    0x00001D8100001D82, 0x80008D2700001D80, 0x80008D2F80008D2B,
-    0x80008D33000394BE, 0x0003488680008D37, 0x80008D3B0003CC87,
-    0x80008D4380008D3F, 0x0007A9A880008D47, 0x0007B9AA0007A9A9,
-    0x0007A9AC0007B9AB, 0x000701AE0007A9AD, 0x80008D4B000735AF,
-    0x80008D5380008D4F, 0x80008D5B80008D57, 0x80008D6380008D5F,
-    0x00072DB880008D67, 0x0000E9BA0000C1B9, 0x0000F9BC0000E9BB,
-    0x0000E5BE0000E5BD, 0x000751B000009DBF, 0x00072DB200075DB1,
-    0x00073DB4000725B3, 0x000725B6000725B5, 0x000631B3000715B7,
-    0x80008D6F80008D6B, 0x80008D7780008D73, 0x000621B5000629B6,
-    0x000699BB80008D7B, 0x80008D7F000695BA, 0x0006E1BF80008D83,
-    0x0006F1BD0006F9BE, 0x80008D870006F9BC, 0x80008D8B000675A3,
-    0x00066DA680008D8F, 0x80008D9780008D93, 0x0006D1AA000665A5,
-    0x80008D9B0006DDAB, 0x0006BDAE80008D9F, 0x0006BDAC0006A5AF,
-    0x0001E5A80006B5AD, 0x0001FDAA0001F5A9, 0x0001EDAC0001F5AB,
-    0x00013DAE000135AD, 0x0000F980000135AF, 0x0000C5820000CD81,
-    0x000178BE00016084, 0x0000AC8780008DA7, 0x0000D1B800019086,
-    0x0000E1BA0000D9B9, 0x000091BC0000E1BB, 0x000095BE00009DBD,
-    0x00014DB0000089BF, 0x00015DB2000155B1, 0x00014DB4000155B3,
-    0x0000F1B60000F1B5, 0x000275B30000F1B7, 0x80008DAF80008DAB,
-    0x80008DB780008DB3, 0x000265B50002B5B6, 0x000291BB80008DBB,
-    0x80008DBF000289BA, 0x000335BF80008DC3, 0x000281BD000289BE,
-    0x80008DC7000289BC, 0x80008DCB000231A3, 0x0002F1A60003C084,
-    0x80008DD380008DCF, 0x0002CDAA000221A5, 0x80008DD70002D5AB,
-    0x0002CDAE80008DDB, 0x0002CDAC000371AF, 0x00008DAB0002C5AD,
-    0x00008DA900008DAA, 0x0000BDAF000339A8, 0x000085AD0000BDAE,
-    0x000000AA00008DAC, 0x80008DDF000000AB, 0x80008DE780008DE3,
-    0x80008DEF80008DEB, 0x00007DBB80008DF3, 0x00007DB900007DBA,
-    0x0001DDBF00007DB8, 0x0001DDBD0001DDBE, 0x0000B9B30001DDBC,
-    0x0000A1B10000B1B2, 0x00005DB70000ADB0, 0x000095B500005DB6,
-    0x80008DF7000095B4, 0x80008DFF80008DFB, 0x0000358180008E03,
-    0x80008E0700000D80, 0x003DACBE00003582, 0x80008E0F80008E0B,
-    0x80008E17003D6885, 0x003C388680008E1B, 0x000149B30002C087,
-    0x000000D080008E1F, 0x80008E2780008E23, 0x000149B5000149B6,
-    0x000129BB80008E2B, 0x80008E2F000125BA, 0x00011DBF80008E33,
-    0x000121BD00011DBE, 0x003634E3000129BC, 0x0002B0E100060CE1,
-    0x80008E3700063CE3, 0x80008E3F80008E3B, 0x003F2C8480008E43,
-    0x80008E47003FA0BE, 0x00377CEF80008E4B, 0x80008E5380008E4F,
-    0x80008E57000118EF, 0x003E688680008E5B, 0x80008E5F003FCC87,
-    0x80008E6780008E63, 0x80008E6B000094EF, 0x80008E6F000194E1,
-    0x80008E73000050E3, 0x003FE981003FE982, 0x003E09A3003FF180,
-    0x80008E130024018F, 0x80008E7B80008E77, 0x003E09A5003E09A6,
-    0x003E69AB80008E7F, 0x80008E83003E65AA, 0x003E5DAF80008E87,
-    0x003E61AD003E5DAE, 0x0038619E003E69AC, 0x0034C19C0038759F,
-    0x0035A59A0039B59D, 0x003079980034759B, 0x002D219600317199,
-    0x002C6D9400318597, 0x00287992002C6595, 0x00245190002DA593,
-    0x00140DB100287991, 0x001801B3001405B0, 0x001879B5001455B2,
-    0x80008E8B0018B5B4, 0x80008E9380008E8F, 0x80008E9B80008E97,
-    0x003CD1A280008E9F, 0x003975A0000125A3, 0x0000C9A1003CBDA1,
-    0x000401A380008EA3, 0x00041DA50000E5A2, 0x0008F9A70004F5A4,
-    0x000C01A9000801A6, 0x000CCDAB00086DA8, 0x0010D1AD000CCDAA,
-    0x0011D5AF001001AC, 0x002505820010D9AE, 0x80008EA700252D83,
-    0x002CC18680008EAB, 0x00291D84002D1187, 0x002D298A00291585,
-    0x80008EAF002C658B, 0x00301D8E80008EB3, 0x0030D98C0034C18F,
-    0x0035619200311D8D, 0x80008EB70035CD93, 0x0039619680008EBB,
-    0x0038A19400389997, 0x003D619A00397D95, 0x80008EBF003D059B,
-    0x80008EC780008EC3, 0x0000E99C80008ECB, 0x80008ED380008ECF,
-    0x80008EDB80008ED7, 0x80008DA380008EDF, 0x80008EE780008EE3,
-    0x003E91A880008EEB, 0x003EEDAA003E91A9, 0x003EE1AC003EE1AB,
-    0x003ED1AE003EE9AD, 0x003E51B0003ED1AF, 0x003E51B2003E51B1,
-    0x003E79B4003E51B3, 0x003E6DB6003E79B5, 0x003E21B8003E65B7,
-    0x003E39BA003E21B9, 0x003E29BC003E39BB, 0x003E11BE003E15BD,
-    0x00039980003E0DBF, 0x0000058200039981, 0x000354BE80008EEF,
-    0x80008EF7003DB0E1, 0x00023884003E40E3, 0x80008EFF80008EFB,
-    0x80008F03003FF4EF, 0x0004988680008F07, 0x003D05B300031C87,
-    0x80008F0B00040884, 0x80008F1380008F0F, 0x003D09B5003D09B6,
-    0x003DF5BB80008F17, 0x80008F1B003DF5BA, 0x003DDDBF80008F1F,
-    0x003DE5BD003DDDBE, 0x80008F23003DE5BC, 0x003DCDA380008F27,
-    0x003DC1A50004C4BE, 0x80008F2F80008F2B, 0x80008F33003DC1A6,
-    0x003D3DAB80008F37, 0x003D2DAD003D3DAA, 0x003D15AF003D2DAC,
-    0x000269B6003D15AE, 0x80008F3F80008F3B, 0x80008F43000269B5,
-    0x80008F47000249B3, 0x0003A9BE80008F4B, 0x0003C1BC0003A9BF,
-    0x0003F9BA0003C1BD, 0x80008F4F0003F9BB, 0x00032DA880008F53,
-    0x000395AA00039DA9, 0x0003B5AC0003ADAB, 0x0003B5AE0003BDAD,
-    0x00000D800003D9AF, 0x00001D8200001581, 0x80008F5B80008F57,
-    0x0005B48780008F5F, 0x000231B800049C86, 0x000235BA000231B9,
-    0x0002D5BC0002CDBB, 0x0002D5BE0002DDBD, 0x000269B00002C9BF,
-    0x000241B2000269B1, 0x000239B4000241B3, 0x000211B6000239B5,
-    0x80008F67000211B7, 0x80008F6B003E34E3, 0x80008F6F003EACE1,
-    0x80008F7300037CEF, 0x80008F7B80008F77, 0x80008F7F000394E3,
-    0x80008F83003E7CE1, 0x80008F87003EE8EF, 0x80008F8F80008F8B,
-    0x000355A380008F93, 0x000375A580008F97, 0x80008F9F80008F9B,
-    0x80008FA3000375A6, 0x0002E5AB80008FA7, 0x0002DDAD0002E5AA,
-    0x0002B5AF0002DDAC, 0x000619A80002B5AE, 0x000621AA000621A9,
-    0x000625AC00063DAB, 0x000655AE00065DAD, 0x80008F6300064DAF,
-    0x80008FAF80008FAB, 0x80008FB780008FB3, 0x80008FBF80008FBB,
-    0x0006E5B880008FC3, 0x000681BA000681B9, 0x000689BC00069DBB,
-    0x0006A1BE000689BD, 0x00063DB00006A1BF, 0x000605B200060DB1,
-    0x0006F5B40006EDB3, 0x0006E1B60006E1B5, 0x0006A9B30006DDB7,
-    0x0000158100002D82, 0x80008FC700001D80, 0x0006ADB50006DDB6,
-    0x0006C9BB80008FCB, 0x80008FCF0006F9BA, 0x000625BF0003E084,
-    0x000639BD000631BE, 0x000388BE0006D1BC, 0x80008EF30006EDA3,
-    0x000699A680008FD3, 0x80008FDB80008FD7, 0x0006BDAA0006E9A5,
-    0x0000488600068DAB, 0x000675AE00006C87, 0x000695AC000661AF,
-    0x000681A800067DAD, 0x000699AA00068DA9, 0x00068DAC000695AB,
-    0x0006B1AE0006B5AD, 0x80008FDF0006ADAF, 0x80008FE780008FE3,
-    0x80008FEF80008FEB, 0x80008FF780008FF3, 0x0006A5B880008FFB,
-    0x000161BA000161B9, 0x000161BC000161BB, 0x000161BE000161BD,
-    0x0006D9B0000161BF, 0x0006A9B20006D9B1, 0x0006BDB40006A9B3,
-    0x0006A1B60006A1B5, 0x000611B300069DB7, 0x8000900380008FFF,
-    0x8000900B80009007, 0x000605B5000635B6, 0x00061DBB8000900F,
-    0x8000901300061DBA, 0x000665BF80009017, 0x00067DBD000679BE,
-    0x8000901B00067DBC, 0x8000901F000655A3, 0x000671A680009023,
-    0x8000902B80009027, 0x000659AA000641A5, 0x8000902F000659AB,
-    0x00063DAE80009033, 0x000639AC000621AF, 0x0002D5A8000639AD,
-    0x000329AA0002DDA9, 0x000339AC000329AB, 0x000329AE000339AD,
-    0x80009037000329AF, 0x8000903F8000903B, 0x8000904B80009047,
-    0x0003B8BE8000904F, 0x0003EDB880009053, 0x000381BA000385B9,
-    0x000385BC000381BB, 0x0003B1BE00038DBD, 0x000359B00003B1BF,
-    0x0003EDB2000359B1, 0x0003FDB40003E5B3, 0x0003E5B60003E5B5,
-    0x0000A1800003D5B7, 0x0000A1820000A181, 0x80009057000C80BE,
-    0x8000905B00029884, 0x000C00868000905F, 0x800090630003F487,
-    0x8000906B80009067, 0x800090738000906F, 0x0003E1B30003B084,
-    0x8000907B80009077, 0x800090838000907F, 0x0003F1B50003F9B6,
-    0x0003DDBB80009087, 0x8000908B0003DDBA, 0x000361BF8000908F,
-    0x000371BD000379BE, 0x80009093000379BC, 0x8000909B80009097,
-    0x8000909F00022DA3, 0x000235A600023DA5, 0x800090A7800090A3,
-    0x000211AA800090AB, 0x0002B5AC000211AB, 0x0002B5AE0002BDAD,
-    0x0003C4E30002ADAF, 0x0001B8E1000710E3, 0x0000718000076CE1,
-    0x0000058200007181, 0x000CC086800090AF, 0x800090B7000D5487,
-    0x0003B0EF800090BB, 0x800090BF000700EF, 0x800090C7800090C3,
-    0x800090CF800090CB, 0x800090D7800090D3, 0x800090DF800090DB,
-    0x000DA0840001A4EF, 0x800090E3000680E1, 0x800090E7000110E3,
-    0x800090EF800090EB, 0x0001F5B3800090F3, 0x800090FB800090F7,
-    0x80009103800090FF, 0x000149B500014DB6, 0x000151BB80009107,
-    0x8000910B000149BA, 0x000139BF8000910F, 0x000141BD000131BE,
-    0x000E2DA8000149BC, 0x000E3DAA000E35A9, 0x000E81AC000E31AB,
-    0x000E85AE000E8DAD, 0x800090B3000EB5AF, 0x8000911780009113,
-    0x000019808000911B, 0x0000058200001981, 0x000E9DB88000911F,
-    0x000EA5BA000EADB9, 0x000F55BC000F4DBB, 0x000F49BE000F5DBD,
-    0x000ECDB0000F41BF, 0x000EDDB2000ED5B1, 0x000EB5B4000ED5B3,
-    0x000EB5B6000EBDB5, 0x000EB5A3000EADB7, 0x80009123000388BE,
-    0x8000912B80009127, 0x000E09A5000E0DA6, 0x000E11AB8000912F,
-    0x00000886000E09AA, 0x000E79AF00036C87, 0x000E01AD000E71AE,
-    0x80009133000E09AC, 0x8000913B80009137, 0x8000913F000FB5B3,
-    0x000F5DB6000F55B5, 0x8000914380009043, 0x000F79BA80009147,
-    0x000F69BC000F79BB, 0x000F49BE000F5DBD, 0x000E69A8000F49BF,
-    0x000E71AA000E69A9, 0x000E91AC000E71AB, 0x000E91AE000E91AD,
-    0x8000914B000E91AF, 0x800091538000914F, 0x8000915B80009157,
-    0x800091638000915F, 0x000E85B880009167, 0x000E85BA000E8DB9,
-    0x000E8DBC000E9DBB, 0x000EB5BE000EBDBD, 0x000EF1B0000179BF,
-    0x000EF1B2000EF1B1, 0x000EC1B4000EC5B3, 0x000EC1B6000EC1B5,
-    0x000EF9A3000EC1B7, 0x8000916F8000916B, 0x8000917780009173,
-    0x000E19A5000E11A6, 0x000E35AB8000917B, 0x8000917F000E35AA,
-    0x000E05AF80009183, 0x000E11AD000E05AE, 0x00000D80000E25AC,
-    0x00001D8200001581, 0x8000918B80009187, 0x000194848000918F,
-    0x00074086000194BE, 0x800091970000F487, 0x8000919F8000919B,
-    0x800091A7800091A3, 0x00028DA8800091AB, 0x000295AA000295A9,
-    0x0002D5AC0002CDAB, 0x0002C9AE0002DDAD, 0x800091AF0002FDAF,
-    0x800091B7800091B3, 0x001400BF800091BB, 0x800091C3800091BF,
-    0x00037DB8800091C7, 0x0003C1BA0003C1B9, 0x0003C1BC0003C1BB,
-    0x0003F1BE0003C9BD, 0x000285B00003F1BF, 0x00034DB2000345B1,
-    0x00035DB4000345B3, 0x00034DB6000345B5, 0x00021DB3000345B7,
-    0x800091CF800091CB, 0x800091D7800091D3, 0x00025DB500025DB6,
-    0x000381BB800091DB, 0x800091DF000241BA, 0x000381BF800091E3,
-    0x000391BD000399BE, 0x800091E7000399BC, 0x800091EB000259A3,
-    0x000219A6800091EF, 0x800091F7800091F3, 0x000205AA000219A5,
-    0x800091FB0003C5AB, 0x0003DDAE800091FF, 0x0003DDAC0003C5AF,
-    0x800092070003D5AD, 0x000204840000CCE3, 0x0001C9800001BCE1,
-    0x0001F5820001FD81, 0x8000920B000510BE, 0x8000920F000428BE,
-    0x8000921780009213, 0x8000921B000010EF, 0x0004E0868000921F,
-    0x800092230002F487, 0x000FFCE380009227, 0x000FE0E18000922B,
-    0x000F14EF8000922F, 0x8000923780009233, 0x8000923F8000923B,
-    0x8000924780009243, 0x8000924F8000924B, 0x8000925780009253,
-    0x8000925F8000925B, 0x8000926780009263, 0x001204850001E0EF,
-    0x8000926B000EDCE1, 0x00002980000E1CE3, 0x0000058200001D81,
-    0x000231B38000926F, 0x0005CC8480009277, 0x8000927F8000927B,
-    0x000221B5000229B6, 0x0001CDBB80009283, 0x800092870001CDBA,
-    0x000175BF8000928B, 0x0001C1BD0001C9BE, 0x0005E9A80001C9BC,
-    0x0005F9AA0005E9A9, 0x0005E9AC0005F9AB, 0x000639AE0005E9AD,
-    0x80009203000639AF, 0x0000888680009273, 0x8000928F00030087,
-    0x8000929780009293, 0x0006D1B88000929B, 0x0006E1BA0006D9B9,
-    0x000691BC0006E1BB, 0x000695BE00069DBD, 0x000649B0000689BF,
-    0x00065DB2000649B1, 0x00064DB4000655B3, 0x0006F1B60006F1B5,
-    0x000571A30006F1B7, 0x800092A38000929F, 0x800092AB800092A7,
-    0x000561A5000569A6, 0x00068DAB800092AF, 0x800092B300068DAA,
-    0x000635AF800092B7, 0x000681AD000689AE, 0x800092BB000689AC,
-    0x0007E1B3800092BF, 0x0007E1B5800092C3, 0x800092CB800092C7,
-    0x800091930007E5B6, 0x0007BDBB800092CF, 0x000795BD0007A1BA,
-    0x000795BF0007A9BC, 0x000601A8000795BE, 0x000621AA000625A9,
-    0x000621AC000621AB, 0x000625AE00062DAD, 0x800092D3000655AF,
-    0x00001D8100001D82, 0x800092D700001D80, 0x800092DF800092DB,
-    0x000639B8000130BE, 0x0006C9BA000639B9, 0x0006D9BC0006C9BB,
-    0x0006C9BE0006D9BD, 0x00062DB00006C9BF, 0x000635B2000631B1,
-    0x000619B4000609B3, 0x000609B6000619B5, 0x0006A9A3000609B7,
-    0x001F288600028C84, 0x800092E700014487, 0x0006A9A50006ADA6,
-    0x0006F5AB800092EB, 0x800092EF0006E9AA, 0x0006DDAF800092F3,
-    0x0006DDAD0006DDAE, 0x800092F70006E1AC, 0x800092FB000615B3,
-    0x00063DB6800092FF, 0x8000930780009303, 0x0001D9BA000635B5,
-    0x8000930B0001D9BB, 0x00017DBE8000930F, 0x00017DBC000165BF,
-    0x0005C5A8000175BD, 0x0005D9AA0005C9A9, 0x0005F9AC0005D1AB,
-    0x000229AE0005F9AD, 0x80009313000229AF, 0x8000931B80009317,
-    0x0000008C8000931F, 0x8000932780009323, 0x0002EDB88000932B,
-    0x00028DBA000285B9, 0x000285BC000281BB, 0x0002BDBE00028DBD,
-    0x000259B000037DBF, 0x0002EDB2000259B1, 0x0002FDB40002E5B3,
-    0x0002E5B60002E5B5, 0x000551A30002D5B7, 0x800093338000932F,
-    0x8000933B80009337, 0x000571A5000579A6, 0x00029DAB8000933F,
-    0x8000934300029DAA, 0x000221AF80009347, 0x000231AD000239AE,
-    0x00001182000239AC, 0x000065808000934B, 0x8000934F00000981,
-    0x80009357000398BE, 0x000324848000935B, 0x000368878000935F,
-    0x80009363001C8C86, 0x8000936B80009367, 0x800093738000936F,
-    0x0002A1B380009377, 0x0002D1B5001CC084, 0x8000937F8000937B,
-    0x800093830002D5B6, 0x0002C1BB80009387, 0x000111BD0002F5BA,
-    0x000111BF0002D9BC, 0x8000938B000119BE, 0x800093938000938F,
-    0x8000939B80009397, 0x800093A38000939F, 0x800093A70006B8EF,
-    0x800093AB00049CE1, 0x800093AF000674E3, 0x800093B7800093B3,
-    0x0000F980800093BB, 0x000005820000F981, 0x001C58BE800093BF,
-    0x00005CEF001F5884, 0x000190E1000640EF, 0x00003CE30006FCE1,
-    0x800093C7000674E3, 0x001C9886800093CB, 0x000269A3001CF487,
-    0x800093CF001F0CBE, 0x800093D7800093D3, 0x000219A500021DA6,
-    0x000209AB800093DB, 0x800093DF00023DAA, 0x0001D9AF800093E3,
-    0x0001D9AD0001D1AE, 0x001D91A8000211AC, 0x001DA1AA001D99A9,
-    0x001DD1AC001DA1AB, 0x001DD5AE001DDDAD, 0x80009353001DC9AF,
-    0x800093E7800093C3, 0x800093EF800093EB, 0x800093F7800093F3,
-    0x001E79B8800093FB, 0x001ECDBA001E79B9, 0x001EDDBC001EC5BB,
-    0x001EC5BE001EC5BD, 0x001DB9B0001EF5BF, 0x001D85B2001D8DB1,
-    0x001E55B4001E4DB3, 0x001E55B6001E5DB5, 0x001FCDB8001E49B7,
-    0x001FDDBA001FD5B9, 0x001FF1BC001FD5BB, 0x001FE9BE001FFDBD,
-    0x001FA5B0001FE9BF, 0x001FB1B2001FB1B1, 0x001F95B4001F8DB3,
-    0x001FFDB6001FF5B5, 0x001E19A8001FF5B7, 0x001E2DAA001E19A9,
-    0x001E25AC001E3DAB, 0x001E25AE001E2DAD, 0x800093FF001FDDAF,
-    0x8000940780009403, 0x8000940F8000940B, 0x80009413800092E3,
-    0x001FE5B380009417, 0x8000941F8000941B, 0x8000942780009423,
-    0x001FEDB5001FEDB6, 0x001E35BB8000942B, 0x8000942F001E21BA,
-    0x001E71BF80009433, 0x001E19BD001E11BE, 0x00006982001E25BC,
-    0x00005980001FA1A3, 0x001FA9A600005181, 0x8000943B80009437,
-    0x001E65AA001FA9A5, 0x00040086001E71AB, 0x001E55AE0001EC87,
-    0x001E61AC001E35AF, 0x001E31A8001E5DAD, 0x001E41AA001E31A9,
-    0x001E41AC001E41AB, 0x001E71AE001E49AD, 0x8000943F001E71AF,
-    0x8000944780009443, 0x8000944F8000944B, 0x8000945780009453,
-    0x000129B88000945B, 0x000139BA000129B9, 0x00012DBC000135BB,
-    0x0000D1BE0000D1BD, 0x0001C9B00000D1BF, 0x0001D9B20001C9B1,
-    0x0001C9B40001D9B3, 0x000119B60001C9B5, 0x001DC9B3000119B7,
-    0x800094638000945F, 0x8000946B80009467, 0x001DC9B5001DC9B6,
-    0x00020DBB8000946F, 0x8000947300020DBA, 0x00020DBF80009477,
-    0x00020DBD00020DBE, 0x8000947B00020DBC, 0x8000947F001D8DA3,
-    0x001D8DA680009483, 0x8000948B80009487, 0x000249AA001D8DA5,
-    0x8000948F000249AB, 0x000249AE80009493, 0x000249AC000249AF,
-    0x00000D80000249AD, 0x0000118200001181, 0x0002CCEF80009497,
-    0x8000949F8000949B, 0x00022CE300029084, 0x0001D8E1000328BE,
-    0x00041086800094A7, 0x800094AB0003D487, 0x000361B3800094AF,
-    0x800094B7800094B3, 0x800094BF800094BB, 0x000371B5000379B6,
-    0x00035DBB800094C3, 0x800094C700035DBA, 0x0000E1BF800094CB,
-    0x0000F1BD0000F9BE, 0x0002A1A30000F9BC, 0x800094D3800094CF,
-    0x800094DB800094D7, 0x0002B1A50002B9A6, 0x00029DAB800094DF,
-    0x800094E300029DAA, 0x000121AF800094E7, 0x000131AD000139AE,
-    0x800094EB000139AC, 0x001F64EF800094EF, 0x800094F7800094F3,
-    0x800094FF800094FB, 0x00000D8080009503, 0x0000158200001181,
-    0x001CE0E180009507, 0x001F88E38000950B, 0x000280848000950F,
-    0x0005488700058CBE, 0x8000951700042C86, 0x001EA4EF8000951B,
-    0x001E00E1001EF4EF, 0x001E2CE3001F50E1, 0x8000951F001EB0E3,
-    0x8000952780009523, 0x8000952F8000952B, 0x0004848480009533,
-    0x80009537000171B3, 0x000115B600011DB5, 0x8000953F8000953B,
-    0x000101BA80009543, 0x0000F5BC000101BB, 0x0000F5BE0000FDBD,
-    0x0006ADA80000EDBF, 0x0006B5AA0006BDA9, 0x0006D9AC0006C9AB,
-    0x0006C9AE0006D9AD, 0x800095470006C1AF, 0x8000954F8000954B,
-    0x8000955780009553, 0x8000955F8000955B, 0x00076DB880009563,
-    0x00070DBA000705B9, 0x000701BC000701BB, 0x000701BE000701BD,
-    0x000689B0000701BF, 0x00076DB2000689B1, 0x00077DB4000765B3,
-    0x000765B6000765B5, 0x800094A3000755B7, 0x80009567000639A3,
-    0x00065DA680009513, 0x8000956F8000956B, 0x000649AA000655A5,
-    0x80009573000649AB, 0x0007BDAE80009577, 0x0007BDAC0007A5AF,
-    0x00006D800007B5AD, 0x0000198200000981, 0x8000957F8000957B,
-    0x8000958300039CBE, 0x0000408680009587, 0x8000958B00006087,
-    0x800095938000958F, 0x8000959B80009597, 0x000691A88000959F,
-    0x0006B9AA000691A9, 0x0006A9AC0006B9AB, 0x0006D9AE0006A9AD,
-    0x800095A30006D9AF, 0x800095AB800095A7, 0x800095B3800095AF,
-    0x800095BB800095B7, 0x000171B8800095BF, 0x000171BA000171B9,
-    0x0001D9BC000171BB, 0x0001C1BE0001C1BD, 0x0006B1B00001F5BF,
-    0x000689B20006B9B1, 0x000151B4000689B3, 0x000151B6000151B5,
-    0x000631B3000151B7, 0x800095C7800095C3, 0x800095CF800095CB,
-    0x000621B5000629B6, 0x00067DBB800095D3, 0x800095D7000675BA,
-    0x000195BF800095DB, 0x00065DBD000195BE, 0x800095DF00065DBC,
-    0x800095E3000675A3, 0x00066DA6800095E7, 0x800095EF800095EB,
-    0x000631AA000665A5, 0x0001AC84000639AB, 0x0001D1AE0001A0BE,
-    0x000619AC0001D1AF, 0x0002DDA8000619AD, 0x0002E5AA0002EDA9,
-    0x0002E5AC0002FDAB, 0x0002E5AE0002EDAD, 0x800095F700033DAF,
-    0x800095FF800095FB, 0x80009603000C64BE, 0x8000960B80009607,
-    0x0003C9B88000960F, 0x0003D9BA0003C9B9, 0x0003F9BC0003D1BB,
-    0x000399BE0003F9BD, 0x000345B0000395BF, 0x000345B200034DB1,
-    0x000345B400035DB3, 0x000345B600034DB5, 0x000355810003F9B7,
-    0x000265B300034980, 0x000265B500035582, 0x8000961780009613,
-    0x000C0086000265B6, 0x000381BB0003E487, 0x000381BD000389BA,
-    0x000381BF000399BC, 0x00022DA3000389BE, 0x8000961F8000961B,
-    0x8000962780009623, 0x00022DA500022DA6, 0x0003C9AB8000962B,
-    0x8000962F0003C1AA, 0x0003C9AF80009633, 0x0003C9AD0003C1AE,
-    0x0006D8E30003D1AC, 0x00069CE10007ACE1, 0x8000963700069CE3,
-    0x8000963B000D5884, 0x800096438000963F, 0x8000964B80009647,
-    0x00011CEF8000964F, 0x8000965780009653, 0x00007980000640EF,
-    0x0000118200001581, 0x8000965B000C0084, 0x8000965F0000ACE3,
-    0x800096670001A4E1, 0x000CC886000070EF, 0x8000966B000DA487,
-    0x800096738000966F, 0x0005C9BA80009677, 0x0005A5B80005C9BB,
-    0x0005F9BE0005CDB9, 0x0005CDBC0005F9BF, 0x0005A5B20005C5BD,
-    0x000611B00005ADB3, 0x0005ADB6000611B1, 0x0005B5B400059DB7,
-    0x000661AA0005ADB5, 0x00069DA8000661AB, 0x000661AE000695A9,
-    0x000671AC000661AF, 0x8000967B000671AD, 0x800096838000967F,
-    0x8000968B80009687, 0x000CB0BE8000968F, 0x000E21A880009693,
-    0x000E21AA000E21A9, 0x000E25AC000E3DAB, 0x000E25AE000E2DAD,
-    0x80009663000E89AF, 0x8000969B80009697, 0x800096A38000969F,
-    0x800096AB800096A7, 0x000F51B8800096AF, 0x000F55BA000F5DB9,
-    0x000F75BC000F6DBB, 0x000F75BE000F7DBD, 0x000EF9B0000F69BF,
-    0x000EA1B2000EA1B1, 0x000EA1B4000EA1B3, 0x000E91B6000EA9B5,
-    0x000EA9B3000E91B7, 0x800095F3800096B3, 0x800096BB800096B7,
-    0x000EADB5000EADB6, 0x000E65BB800096BF, 0x800096C3000E3DBA,
-    0x000E6DBF800096C7, 0x000E75BD000E65BE, 0x00001982000E75BC,
-    0x00006580000EEDA3, 0x000EE9A600001981, 0x800096CF800096CB,
-    0x000E79AA000EE9A5, 0x800096D3000E21AB, 0x000E21AE800096D7,
-    0x000E31AC000E29AF, 0x000E61A8000E31AD, 0x000E7DAA000E75A9,
-    0x000E6DAC000E75AB, 0x000EFDAE000EF5AD, 0x00018086000EF5AF,
-    0x800096DB0001A487, 0x800096E3800096DF, 0x800096EB800096E7,
-    0x000175B8800096EF, 0x000175BA00017DB9, 0x0001DDBC0001C9BB,
-    0x0001C5BE0001C5BD, 0x000E8DB00001F5BF, 0x000E9DB2000E95B1,
-    0x000155B4000E91B3, 0x000155B600015DB5, 0x000EF5B300014DB7,
-    0x800096F7800096F3, 0x800096FF800096FB, 0x000EE5B5000E75B6,
-    0x000E51BB80009703, 0x80009707000E49BA, 0x000199BF8000970B,
-    0x000E45BD000191BE, 0x8000970F000E49BC, 0x80009713000EB1A3,
-    0x000E31A680009717, 0x8000971F8000971B, 0x000E0DAA000EA1A5,
-    0x80009723000E15AB, 0x0001D5AE80009727, 0x000E0DAC0001DDAF,
-    0x0002EDA8000E01AD, 0x000325AA00032DA9, 0x000325AC00033DAB,
-    0x000325AE00032DAD, 0x8000972B000399AF, 0x800097338000972F,
-    0x8000973B80009737, 0x000274BE8000973F, 0x00038DB880009747,
-    0x000391BA000391B9, 0x0003BDBC0003A5BB, 0x00007DBE000075BD,
-    0x0003E9B0000075BF, 0x0003F9B20003E9B1, 0x0003D9B40003F1B3,
-    0x0003B9B60003D9B5, 0x0000AD800003B5B7, 0x0000BD820000B581,
-    0x8000974B0003A1B3, 0x0003A1B60003A1B5, 0x0002E0848000974F,
-    0x000321BA80009753, 0x000321BC000321BB, 0x000311BE000329BD,
-    0x0003EDA3000311BF, 0x006FB4850004C085, 0x8000975B80009757,
-    0x0003EDA50003EDA6, 0x00036DAB8000975F, 0x0005488600036DAA,
-    0x00035DAF00036C87, 0x000365AD00035DAE, 0x8000976300036DAC,
-    0x006C58EF000E00E3, 0x80009767000FB4E1, 0x8000976F8000976B,
-    0x0003A9A180009773, 0x0003C1A3000FFDA0, 0x0003E0E100030DA2,
-    0x0003E4E3000FE0EF, 0x8000977B80009777, 0x000404848000977F,
-    0x80009783000404BE, 0x80009787000394EF, 0x8000978F8000978B,
-    0x800097930003D4E3, 0x80009797000054E1, 0x8000979F8000979B,
-    0x00000D80800097A3, 0x00001D8200001581, 0x800097AB800097A7,
-    0x001B918F800097AF, 0x0007E084000E9CEF, 0x800097B7000ED4E1,
-    0x800097BB000EF0E3, 0x00071886800097BF, 0x0017E59900049087,
-    0x000BE99B0017E598, 0x000AFD9D000B399A, 0x000F559F000BF19C,
-    0x001F2991000F5D9E, 0x001F2593001BCD90, 0x0013D195001FCD92,
-    0x0017199700132994, 0x0010E18C00137596, 0x00102D8E0010258D,
-    0x80009743000CF98F, 0x0014398A800097B3, 0x0018E9840014958B,
-    0x0018E58600180585, 0x800097C30014F187, 0x001C3182800097C7,
-    0x0004A99C001C1583, 0x800097CF800097CB, 0x800097D7800097D3,
-    0x0004D19A800097DB, 0x000D4D9400047D9B, 0x0008E5960008E595,
-    0x800097DF0008B597, 0x000C5992800097E3, 0x000091A1000C5993,
-    0x007C01A3800097E7, 0x007CB9A5000099A2, 0x007899A7007C91A4,
-    0x007981A9007DE1A6, 0x007489AB0078A1A8, 0x007001AD007401AA,
-    0x007085AF007459AC, 0x006CE1B100709DAE, 0x006801B3006C01B0,
-    0x0068F5B5006C1DB2, 0x800097EB0068F5B4, 0x800097F3800097EF,
-    0x800097FB800097F7, 0x80009803800097FF, 0x8000980B80009807,
-    0x00073DA88000980F, 0x000795AA000795A9, 0x0007DDAC0007A9AB,
-    0x0007C5AE0007C5AD, 0x8000981300061DAF, 0x00001D8100001D82,
-    0x8000981700001D80, 0x8000981F8000981B, 0x000619B8000160BE,
-    0x000629BA000619B9, 0x000621BC000625BB, 0x000621BE000621BD,
-    0x000671B0000621BF, 0x000671B2000671B1, 0x00064DB4000671B3,
-    0x00063DB6000635B5, 0x00071DB300062DB7, 0x0000288680009827,
-    0x8000982B0000A887, 0x000755B5000745B6, 0x0006B5BB8000982F,
-    0x800098330006B5BA, 0x0006F1BF80009837, 0x0006A5BD000689BE,
-    0x8000983B0006A5BC, 0x8000983F000759A3, 0x000701A680009843,
-    0x8000984B80009847, 0x0006F1AA000711A5, 0x8000984F0006F1AB,
-    0x0006CDAE80009853, 0x0006E1AC0006B5AF, 0x0005A9A80006E1AD,
-    0x0005BDAA0005B5A9, 0x000225AC00023DAB, 0x000251AE000251AD,
-    0x80009857000251AF, 0x8000985F8000985B, 0x00033C8480009863,
-    0x8000986B80009867, 0x0002A5B88000986F, 0x0002A5BA0002ADB9,
-    0x0002A5BC0002BDBB, 0x0002A5BE0002ADBD, 0x000231B000037DBF,
-    0x000219B2000231B1, 0x0002F5B4000219B3, 0x0002A5B60002ADB5,
-    0x8000987300029DB7, 0x8000987B80009877, 0x8000987F000539B3,
-    0x0002DDB60002A1B5, 0x8000988780009883, 0x0002F9BA8000988B,
-    0x0002C1BC0002F9BB, 0x00023DBE0002C1BD, 0x8000988F000265BF,
-    0x0002FDA5000281A6, 0x000565A380009897, 0x007CD886007C58BE,
-    0x000239AF007C9C87, 0x00029DAD000261AE, 0x0002A5AB00029DAC,
-    0x8000989B0002A5AA, 0x000285A88000989F, 0x000295AA000295A9,
-    0x0002BDAC0002A5AB, 0x0002D1AE0002D5AD, 0x000185810002D1AF,
-    0x800098A300018580, 0x800098A70001ED82, 0x800098AF800098AB,
-    0x000175B8800098B3, 0x000175BA00017DB9, 0x0001D5BC0001CDBB,
-    0x0001C9BE0001DDBD, 0x0002B5B00001C1BF, 0x000281B20002BDB1,
-    0x000151B4000281B3, 0x000151B6000151B5, 0x800098B7000151B7,
-    0x800098BF800098BB, 0x000630E1800098C3, 0x000628E3000764E1,
-    0x007D20840006C4E3, 0x00006CEF800098C7, 0x800098CB000618EF,
-    0x800098D3800098CF, 0x00025DB3800098D7, 0x800098DB007C48BE,
-    0x800098E3800098DF, 0x000275B5000215B6, 0x000239BB800098E7,
-    0x800098EB000231BA, 0x0001D5BF800098EF, 0x000215BD0001D5BE,
-    0x007D9DA3000215BC, 0x800098F380009893, 0x800098FB800098F7,
-    0x007DB5A5007DD5A6, 0x007DF9AB800098FF, 0x80009903007DF1AA,
-    0x007E15AF80009907, 0x007DD5AD007E15AE, 0x00004D80007DD5AC,
-    0x0000558200005581, 0x8000990B007EB1B3, 0x007F6DB6007F65B5,
-    0x000340868000990F, 0x007FE9BA00030487, 0x007FF9BC007FE9BB,
-    0x007FDDBE007FF9BD, 0x80009913007FCDBF, 0x8000982380009917,
-    0x8000991F8000991B, 0x8000992780009923, 0x007E6DA88000992B,
-    0x007E55AA007E5DA9, 0x007F81AC007F85AB, 0x007F81AE007F81AD,
-    0x007F41B0007F81AF, 0x007F41B2007F41B1, 0x007F75B4007F41B3,
-    0x007F6DB6007F65B5, 0x007F5DB8007F65B7, 0x007F25BA007F2DB9,
-    0x007F2DBC007F3DBB, 0x007F15BE007F1DBD, 0x007FFDA30000F5BF,
-    0x800099338000992F, 0x8000993B80009937, 0x007E29A5007E21A6,
-    0x007EA5AB8000993F, 0x80009943007EA5AA, 0x007E81AF80009947,
-    0x007EB5AD007E91AE, 0x8000994B007EB5AC, 0x800099538000994F,
-    0x8000995B80009957, 0x800099638000995F, 0x00003D8080009967,
-    0x0000198200000981, 0x8000996F8000996B, 0x0001B0BE0001B884,
-    0x0001ADA880009973, 0x0001D5AA0001D5A9, 0x000115AC00010DAB,
-    0x000109AE000119AD, 0x00040086000109AF, 0x8000997700010487,
-    0x8000997F8000997B, 0x8000998780009983, 0x0000EDB88000998B,
-    0x00008DBA000085B9, 0x00009DBC000085BB, 0x00008DBE000085BD,
-    0x000179B0000085BF, 0x0000EDB2000179B1, 0x0000FDB40000E5B3,
-    0x0000E5B60000E5B5, 0x00025DB30000D5B7, 0x800099938000998F,
-    0x8000999B80009997, 0x00029DB5000299B6, 0x0002BDBB8000999F,
-    0x800099A30002BDBA, 0x000321BF800099A7, 0x000331BD000339BE,
-    0x000328BE000339BC, 0x800099AB000219A3, 0x0002DDA6800099AF,
-    0x800099B7800099B3, 0x0002F9AA0002D9A5, 0x800099BB0002F9AB,
-    0x00037DAE800099BF, 0x00037DAC000365AF, 0x0004C8BE000375AD,
-    0x800099C7800099C3, 0x800099CB0005C4BE, 0x800099D3800099CF,
-    0x00003D80800099D7, 0x0000198200000981, 0x800099DF800099DB,
-    0x800099E700033884, 0x000278EF800099EB, 0x00044886800099EF,
-    0x800099F300035487, 0x800099FB800099F7, 0x80009A03800099FF,
-    0x80009A0B80009A07, 0x000254E380009A0F, 0x000140E180009A13,
-    0x80009A1B80009A17, 0x80009A1F007F24E3, 0x80009A23007C9CE1,
-    0x80009A2B80009A27, 0x00056C8480009A2F, 0x80009A3780009A33,
-    0x80009A3F80009A3B, 0x80009A43007F18EF, 0x0002F1B380009A47,
-    0x80009A4F80009A4B, 0x80009A5B80009A57, 0x0002E1B50002E9B6,
-    0x000171BB80009A5F, 0x0005A087000169BA, 0x000139BF0004CC86,
-    0x000151BD000159BE, 0x007F40E1000161BC, 0x007E38E380009A63,
-    0x0000D9820004C084, 0x0000A580000014EF, 0x80009A670000D981,
-    0x80009A6B0000C0E3, 0x80009A6F0001D4E1, 0x007E84EF80009A73,
-    0x0001CDAA80009A77, 0x80009A7B0001D5AB, 0x0001FDAE80009A7F,
-    0x0001C5AC00019DAF, 0x80009A830001F5AD, 0x80009A87000255A3,
-    0x00024DA680009A8B, 0x80009A9380009A8F, 0x800099E3000245A5,
-    0x80009A9780009A53, 0x80009A9F80009A9B, 0x80009AA780009AA3,
-    0x000699A880009AAB, 0x0006EDAA000699A9, 0x0006E5AC0006FDAB,
-    0x0006E5AE0006EDAD, 0x0006A5B00006DDAF, 0x0006A5B20006ADB1,
-    0x0006ADB40006B9B3, 0x000751B6000755B5, 0x000775B800074DB7,
-    0x000775BA00077DB9, 0x000759BC000749BB, 0x000749BE000759BD,
-    0x000645B3000745BF, 0x80009AB380009AAF, 0x80009ABB80009AB7,
-    0x00064DB500064DB6, 0x000649BB80009ABF, 0x00000886000641BA,
-    0x0007B1BF00008C87, 0x00064DBD000649BE, 0x00005D82000651BC,
-    0x00004580000601A3, 0x000609A600005D81, 0x80009ACB80009AC7,
-    0x000605AA000609A5, 0x80009ACF00060DAB, 0x00060DAE80009AD3,
-    0x000615AC0007F5AF, 0x00064DA8000609AD, 0x000655AA000655A9,
-    0x0006B1AC000689AB, 0x0006A9AE0006BDAD, 0x80009AC30006A9AF,
-    0x80009ADB80009AD7, 0x80009AE380009ADF, 0x80009AEB80009AE7,
-    0x000149B880009AEF, 0x000159BA000149B9, 0x000149BC000159BB,
-    0x0001DDBE000149BD, 0x0006DDB00001D5BF, 0x0006A5B20006ADB1,
-    0x000699B400068DB3, 0x00068DB6000699B5, 0x0006DDB3000685B7,
-    0x80009AF780009AF3, 0x80009AFF80009AFB, 0x000639B500063DB6,
-    0x000669BB80009B03, 0x80009B07000619BA, 0x00065DBF80009B0B,
-    0x000659BD00065DBE, 0x80009B0F000671BC, 0x80009B13000699A3,
-    0x000679A680009B17, 0x80009B1F80009B1B, 0x00065DAA00067DA5,
-    0x80009B2300062DAB, 0x000619AE80009B27, 0x000635AC000619AF,
-    0x0002D5A800061DAD, 0x0002E1AA0002DDA9, 0x000335AC0002E1AB,
-    0x000335AE00033DAD, 0x0003CD8000032DAF, 0x0000198200000981,
-    0x80009B2F80009B2B, 0x000374BE00021884, 0x0003E9B880009B37,
-    0x000389BA0003E9B9, 0x00039DBC000385BB, 0x000381BE000381BD,
-    0x000355B00003B5BF, 0x000355B200035DB1, 0x0003F9B40003E9B3,
-    0x0003E9B60003F9B5, 0x000C20860003E1B7, 0x80009B3B00031087,
-    0x80009B4380009B3F, 0x80009B4B80009B47, 0x80009B4F00025CEF,
-    0x80009B53000014E1, 0x000CF8BE000288E3, 0x80009B5B80009B57,
-    0x0003F9BB80009B5F, 0x000D80BE0003F1BA, 0x00035DBF80009B63,
-    0x00035DBD00035DBE, 0x000209B30003E1BC, 0x80009B6B80009B67,
-    0x80009B7380009B6F, 0x0003DDB50003DDB6, 0x0006A9A880009B77,
-    0x0006B9AA0006A9A9, 0x0006A9AC0006B9AB, 0x00051DAE0006A9AD,
-    0x80009B7B000515AF, 0x80009B8380009B7F, 0x80009B8B80009B87,
-    0x80009B9380009B8F, 0x000519B880009B97, 0x000525BA00052DB9,
-    0x0005D9BC0005C9BB, 0x0005CDBE0005C5BD, 0x00056DB00005C5BF,
-    0x000575B2000575B1, 0x00055DB4000545B3, 0x000535B600053DB5,
-    0x0006F4E1000529B7, 0x000614E3000754E1, 0x000009810006BCE3,
-    0x80009B9B0000A980, 0x000DEC8400003982, 0x000C888780009B9F,
-    0x80009BA7000C0C86, 0x0007CCEF80009BAB, 0x000329A30007C4EF,
-    0x80009BB380009BAF, 0x80009BBB80009BB7, 0x0002FDA50002FDA6,
-    0x0002D9AB80009BBF, 0x80009BC30002D1AA, 0x00027DAF80009BC7,
-    0x00027DAD00027DAE, 0x000E3DA80002C1AC, 0x000E85AA000E8DA9,
-    0x000E85AC000E9DAB, 0x000EB9AE000E8DAD, 0x80009BA3000EB5AF,
-    0x80009BCF80009BCB, 0x80009BD780009BD3, 0x80009BDF80009BDB,
-    0x000EBDB880009BE3, 0x000F6DBA000F65B9, 0x000F7DBC000F65BB,
-    0x000F6DBE000F65BD, 0x000ED5B0000F65BF, 0x000ED5B2000EDDB1,
-    0x000E95B4000EA1B3, 0x000E95B6000E9DB5, 0x000E0DB3000E8DB7,
-    0x80009BEB80009BE7, 0x80009BF380009BEF, 0x000E0DB5000E0DB6,
-    0x000E19BB80009BF7, 0x80009BFB000E11BA, 0x000E59BF80009B33,
-    0x000E75BD000E51BE, 0x80009BFF000E01BC, 0x80009C03000E49A3,
-    0x000E49A680009C07, 0x80009C0F80009C0B, 0x000E55AA000E49A5,
-    0x0003A484000E5DAB, 0x000E15AE80009C13, 0x000E45AC000E1DAF,
-    0x000EB1A8000E31AD, 0x000ECDAA000EB1A9, 0x000EDDAC000EC5AB,
-    0x000EC5AE000EC5AD, 0x0001ED80000EF5AF, 0x0001F1820001F181,
-    0x0001908680009C17, 0x80009C1B0001B487, 0x0001C5B880009C1F,
-    0x0001C5BA0001CDB9, 0x0001CDBC0001DDBB, 0x000199BE0001FDBD,
-    0x000E8DB0000195BF, 0x000141B2000141B1, 0x000141B4000141B3,
-    0x000141B6000141B5, 0x000E45B3000141B7, 0x80009C2780009C23,
-    0x80009C2F80009C2B, 0x000E55B5000E45B6, 0x000185BB80009C33,
-    0x80009C37000E49BA, 0x000185BF80009C3B, 0x000195BD000185BE,
-    0x80009C3F000195BC, 0x80009C43000E01A3, 0x000E01A680009C47,
-    0x80009C4F80009C4B, 0x000E0DAA000E11A5, 0x80009C530001C1AB,
-    0x0001C1AE80009C57, 0x0001D1AC0001C1AF, 0x00032DA80001D1AD,
-    0x000335AA00033DA9, 0x000399AC000389AB, 0x000389AE000399AD,
-    0x80009C5B000381AF, 0x80009C6380009C5F, 0x80009C6B80009C67,
-    0x80009C7380009C6F, 0x0003ADB880009C77, 0x00006DBA000065B9,
-    0x00007DBC000065BB, 0x00006DBE000065BD, 0x0003C9B0000065BF,
-    0x0003A9B20003C9B1, 0x0003BDB40003A5B3, 0x0003A1B60003A1B5,
-    0x0000BD80000395B7, 0x0000198200000981, 0x80009C7F80009C7B,
-    0x80009C870003D8BE, 0x0002AC8480009C8B, 0x0003E88780009C8F,
-    0x80009C9300040C86, 0x80009C9B80009C97, 0x0003FDB380009C9F,
-    0x80009CA780009CA3, 0x80009CAF80009CAB, 0x000351B5000359B6,
-    0x00034DBB80009CB3, 0x80009CB700034DBA, 0x000325BF80009CBB,
-    0x000331BD000339BE, 0x80009CBF00033DBC, 0x80009CC780009CC3,
-    0x000F58EF80009CCB, 0x80009CD380009CCF, 0x000E90E380009CD7,
-    0x000FB0E180009CDB, 0x80009CE380009CDF, 0x80009CEB80009CE7,
-    0x0000758080009CEF, 0x0000758200007D81, 0x00033CEF00051884,
-    0x000514BE80009CF7, 0x0003D0E380009CFB, 0x000040E180009CFF,
-    0x0004D08680009D03, 0x80009D0700055887, 0x80009D0F80009D0B,
-    0x80009D1780009D13, 0x80009D1F80009D1B, 0x80009D2780009D23,
-    0x0004EC84000FACEF, 0x80009D2B000E10E1, 0x80009D2F000154E3,
-    0x80009D3780009D33, 0x0002FDA380009D3B, 0x80009D4380009D3F,
-    0x80009D4B80009D47, 0x000251A5000259A6, 0x00024DAB80009D4F,
-    0x80009D5300024DAA, 0x000225AF80009D57, 0x000231AD000239AE,
-    0x000699A800023DAC, 0x0006ADAA000699A9, 0x0006A5AC0006BDAB,
-    0x0006A5AE0006ADAD, 0x80009CF30006D9AF, 0x0000198100001182,
-    0x80009D5B0007C180, 0x000370BE80009D5F, 0x000749B880009D63,
-    0x00075DBA000749B9, 0x00074DBC000755BB, 0x000775BE000771BD,
-    0x0006A9B000076DBF, 0x0006B9B20006A9B1, 0x000699B40006B1B3,
-    0x000779B6000699B5, 0x000635B5000779B7, 0x80009D6B80009D67,
-    0x00003086000635B6, 0x00063DB300035C87, 0x00067DBD80009D6F,
-    0x000645BF000675BC, 0x80009C83000645BE, 0x000675BB80009D73,
-    0x000665A3000675BA, 0x80009D7B80009D77, 0x80009D8380009D7F,
-    0x00066DA500066DA6, 0x00062DAB80009D87, 0x80009D8B00062DAA,
-    0x00061DAF80009D8F, 0x000625AD00061DAE, 0x000655A800062DAC,
-    0x00066DAA000659A9, 0x000669AC000661AB, 0x000699AE000669AD,
-    0x80009D93000699AF, 0x80009D9B80009D97, 0x80009DA380009D9F,
-    0x80009DAB80009DA7, 0x0006F9B880009DAF, 0x00068DBA0006F9B9,
-    0x00069DBC000685BB, 0x000685BE000685BD, 0x0006E9B00006B5BF,
-    0x0006F9B20006E9B1, 0x0006E9B40006F9B3, 0x0006C9B60006DDB5,
-    0x0006E5B30006C9B7, 0x80009DB780009DB3, 0x80009DBF80009DBB,
-    0x0006E1B50006E9B6, 0x00062DBB80009DC3, 0x80009DC700062DBA,
-    0x000619BF80009DCB, 0x00062DBD00062DBE, 0x0000158200062DBC,
-    0x000061800006A1A3, 0x0006ADA600006181, 0x000190BE80009DCF,
-    0x000669AA0006A5A5, 0x0001A484000669AB, 0x000669AE80009DD7,
-    0x000669AC00065DAF, 0x000285A8000669AD, 0x000295AA00028DA9,
-    0x0002D5AC0002B9AB, 0x0002D5AE0002DDAD, 0x001C80860002CDAF,
-    0x80009DDB00036487, 0x80009DDF000360BE, 0x80009DE780009DE3,
-    0x000375B880009DEB, 0x000375BA00037DB9, 0x0003D9BC0003C9BB,
-    0x0003C9BE0003D9BD, 0x0002BDB00003C1BF, 0x000285B200028DB1,
-    0x000355B400034DB3, 0x000355B600035DB5, 0x00021DB300034DB7,
-    0x0003088480009DEF, 0x80009DF780009DF3, 0x00025DB500025DB6,
-    0x000249BB80009DFB, 0x80009DFF000279BA, 0x000399BF80009E03,
-    0x000399BD000391BE, 0x000000B0000251BC, 0x80009E07000259A3,
-    0x000219A680009E0B, 0x80009E1380009E0F, 0x00023DAA000219A5,
-    0x80009E1700020DAB, 0x0003D5AE80009E1B, 0x000215AC0003DDAF,
-    0x80009E1F0003DDAD, 0x80009E2780009E23, 0x80009E2B0006ACEF,
-    0x80009E3380009E2F, 0x001CE8BE80009E37, 0x80009E3F80009E3B,
-    0x80009E4B80009E47, 0x80009E4F000780E1, 0x00005D80000660E3,
-    0x0000618200006181, 0x80009E5300017DB3, 0x000165B600016DB5,
-    0x001D208680009E57, 0x0001F9BA001D5887, 0x0001D9BC0001F1BB,
-    0x0001B1BE0001D9BD, 0x0000A0EF0001B1BF, 0x80009E5F80009E5B,
-    0x80009E6780009E63, 0x80009E6F80009E6B, 0x001C6C8400015CEF,
-    0x80009E730006CCE1, 0x000048E300061CE3, 0x000130E180009E77,
-    0x0001F1A380009E7B, 0x0014008580009E7F, 0x80009E8780009E83,
-    0x0001E1A50001E9A6, 0x00017DAB80009E8B, 0x80009E8F000175AA,
-    0x00013DAF80009E93, 0x000155AD00013DAE, 0x001DEDA8000155AC,
-    0x001E39AA001E2DA9, 0x001E29AC001E39AB, 0x001E9DAE001E29AD,
-    0x80009E43001E91AF, 0x80009E9B80009E97, 0x80009EA380009E9F,
-    0x0000F1810000F982, 0x001EA9B80000FD80, 0x001F49BA001EA9B9,
-    0x001F59BC001F49BB, 0x001F4DBE001F45BD, 0x001EF1B0001F45BF,
-    0x001EC1B2001EF9B1, 0x001EB9B4001EC1B3, 0x001EADB6001EB9B5,
-    0x001F11B0001EA5B7, 0x001F11B2001F11B1, 0x001F25B4001F25B3,
-    0x001F5DB6001F29B5, 0x001F71B8001F51B7, 0x001F41BA001F79B9,
-    0x000095BC001F41BB, 0x000095BE00009DBD, 0x80009EA700008DBF,
-    0x80009EAB80009DD3, 0x80009EB380009EAF, 0x0003F88680009EB7,
-    0x001FCDA80000D087, 0x001FD1AA001FD5A9, 0x001FADAC001FBDAB,
-    0x001F71AE001F71AD, 0x001E39B3001F71AF, 0x80009EBF80009EBB,
-    0x80009EC780009EC3, 0x001E45B5001E91B6, 0x001EB5BB80009ECB,
-    0x80009ECF001EB5BA, 0x001E8DBF80009ED3, 0x001E91BD001E81BE,
-    0x80009ED7001EA5BC, 0x80009EDB001E7DA3, 0x001ED5A680009EDF,
-    0x80009EE780009EE3, 0x001EF1AA001E01A5, 0x80009EEB001EF1AB,
-    0x001EC5AE80009EEF, 0x001EE1AC001EC9AF, 0x000155A8001ED5AD,
-    0x000081AA000081A9, 0x000081AC000081AB, 0x0000B1AE000089AD,
-    0x80009EF30000B1AF, 0x80009EFB80009EF7, 0x80009F0380009EFF,
-    0x80009F0B80009F07, 0x000065B880009F0F, 0x000065BA00006DB9,
-    0x000065BC00007DBB, 0x000065BE00006DBD, 0x0000A1B00003DDBF,
-    0x0000A5B20000ADB1, 0x0000A9B40000B9B3, 0x000095B600009DB5,
-    0x80009F1300005DB7, 0x00001D8100001D82, 0x80009F1700001D80,
-    0x80009F1F80009F1B, 0x80009F27000214BE, 0x80009F2B0002A884,
-    0x80009F3380009F2F, 0x80009F3B80009F37, 0x000349B30000008F,
-    0x0004E88680009F3F, 0x80009F430002AC87, 0x000349B5000349B6,
-    0x000329BB80009F47, 0x80009F4B000325BA, 0x000319BF80009F4F,
-    0x00032DBD00032DBE, 0x80009F53000331BC, 0x80009F5700038DA3,
-    0x00038DA680009F5B, 0x80009F6380009F5F, 0x0003E1AA00038DA5,
-    0x80009F670003EDAB, 0x0003E9AE80009F6B, 0x0003F5AC0003DDAF,
-    0x000384BE0003E9AD, 0x80009F7380009F6F, 0x80009F7B80009F77,
-    0x80009F8380009F7F, 0x00003D8080009F87, 0x0000198200000981,
-    0x80009F8F80009F8B, 0x0003348480009F97, 0x00012DB380009F9B,
-    0x0002CC8780009F9F, 0x80009FA300054C86, 0x00012DB500012DB6,
-    0x000149BB80009FA7, 0x80009FAB000179BA, 0x0001BDBF80009FAF,
-    0x0001B9BD0001BDBE, 0x001F78E1000151BC, 0x001FD0E380009FB3,
-    0x80009FBB80009FB7, 0x80009FBF000194E1, 0x80009FC3000368E3,
-    0x80009FCB80009FC7, 0x80009FCF000388EF, 0x001FACEF80009FD3,
-    0x80009FDB80009FD7, 0x80009FE380009FDF, 0x80009FEB80009FE7,
-    0x80009FF380009FEF, 0x80009FF7001E44EF, 0x80009FFB001EDCE1,
-    0x80009FFF001E1CE3, 0x8000A00B8000A007, 0x000069818000A00F,
-    0x0001E5A300006580, 0x0001E5A500007D82, 0x000414848000A013,
-    0x000428BE0001E5A6, 0x000181AB8000A017, 0x000171AD0001B1AA,
-    0x000175AF000199AC, 0x000621A8000175AE, 0x000625AA00062DA9,
-    0x000625AC00063DAB, 0x00065DAE00062DAD, 0x80009F9300064DAF,
-    0x000308868000A003, 0x8000A01B00038C87, 0x8000A0238000A01F,
-    0x0006E9B88000A027, 0x00068DBA0006E9B9, 0x00069DBC000685BB,
-    0x00068DBE000685BD, 0x00063DB0000685BF, 0x000605B200060DB1,
-    0x0006F9B40006EDB3, 0x0006EDB60006F9B5, 0x0007CDB00006E5B7,
-    0x0007D1B20007D5B1, 0x0007F5B40007EDB3, 0x0007E9B60007FDB5,
-    0x0007DDB80007E9B7, 0x00072DBA000725B9, 0x00073DBC000725BB,
-    0x00072DBE000725BD, 0x8000A02B000725BF, 0x80009F238000A02F,
-    0x8000A0378000A033, 0x8000A03F8000A03B, 0x000615A88000A043,
-    0x0007EDAA000619A9, 0x0007EDAC0007FDAB, 0x0007D5AE0007DDAD,
-    0x000605B30007B9AF, 0x8000A04B8000A047, 0x8000A0538000A04F,
-    0x000605B500063DB6, 0x000671BB8000A057, 0x8000A05B000669BA,
-    0x000659BF8000A05F, 0x000659BD000651BE, 0x8000A063000665BC,
-    0x8000A067000641A3, 0x000679A68000A06B, 0x0001B4848000A06F,
-    0x00062DAA000641A5, 0x000180BE000635AB, 0x000615AE8000A077,
-    0x000621AC00061DAF, 0x000635A800061DAD, 0x000635AA00063DA9,
-    0x000675AC000659AB, 0x0001ADAE0001A5AD, 0x0000E9800001A5AF,
-    0x0000FD820000E981, 0x000F3086000124BF, 0x8000A07B0000E487,
-    0x0000C5B88000A07F, 0x0000C5BA0000CDB9, 0x0000CDBC0000DDBB,
-    0x0000F5BE0000FDBD, 0x0001DDB000009DBF, 0x00012DB2000125B1,
-    0x000121B4000121B3, 0x000121B6000121B5, 0x0002C1BB000121B7,
-    0x8000A083000239BA, 0x0002C5BF8000A087, 0x0002D5BD0002C5BE,
-    0x00059DB30002D9BC, 0x8000A08F8000A08B, 0x0000008C8000A093,
-    0x0005DDB5000205B6, 0x00027DAA8000A097, 0x8000A09B000285AB,
-    0x000281AE8000A09F, 0x00029DAC000281AF, 0x8000A0A3000291AD,
-    0x8000A0A70005D9A3, 0x000241A68000A0AB, 0x8000A0B38000A0AF,
-    0x006A9182000599A5, 0x8000A0B7006A9183, 0x0016B9868000A0BB,
-    0x001611840017E987, 0x0012A18A00169985, 0x8000A0BF0012A18B,
-    0x001E918E8000A0C3, 0x0013998C001E598F, 0x001A7192001E118D,
-    0x8000A0C7001AE593, 0x0006F1960024E8EF, 0x001AE59400060597,
-    0x0002299A00061995, 0x8000A0CB0002C59B, 0x8000A0D38000A0CF,
-    0x0002DD9C001B28E1, 0x0007219F000F20E3, 0x001B359D0007C19E,
-    0x001BA19B001B2D9C, 0x001F3999001FF99A, 0x00120197001FB198,
-    0x0013299500132196, 0x0017199300165194, 0x0017B19100173192,
-    0x006B518F006B2990, 0x000D04840007ACE3, 0x00000D8000071CE1,
-    0x00003D8200003581, 0x8000A0DB8000A0D7, 0x000D20BE8000A0DF,
-    0x8000A0EB8000A0E7, 0x000C588600074CEF, 0x8000A0EF000D6087,
-    0x8000A0F78000A0F3, 0x000F5C848000A0FB, 0x000008EF8000A0FF,
-    0x000190E1000684EF, 0x0000E0E30006B0E1, 0x8000A103000664E3,
-    0x8000A10B8000A107, 0x8000A1138000A10F, 0x000F00BE8000A117,
-    0x8000A11B000E4084, 0x8000A1238000A11F, 0x8000A12B8000A127,
-    0x8000A1338000A12F, 0x0003D5A38000A137, 0x000331A10003D5A2,
-    0x8000A07300072DA0, 0x8000A13B8000A0E3, 0x8000A1438000A13F,
-    0x0000918100009982, 0x000D4DA800009980, 0x000D95AA000D9DA9,
-    0x000E3DAC000E25AB, 0x000E11AE000E11AD, 0x000E7DB0000E11AF,
-    0x000E05B2000E0DB1, 0x000E39B4000E2DB3, 0x000E2DB6000E39B5,
-    0x000EE9B8000E25B7, 0x000EC1BA000EE9B9, 0x000EB9BC000EC1BB,
-    0x000E95BE000E9DBD, 0x000D3DB3000E8DBF, 0x8000A14B8000A147,
-    0x8000A1538000A14F, 0x000E95B5000EB1B6, 0x000EE9BB8000A157,
-    0x00008886000E99BA, 0x000EDDBF0000E487, 0x000ED9BD000EDDBE,
-    0x8000A15B000EF1BC, 0x000184BE000D79A3, 0x000EF5A600068084,
-    0x8000A1638000A15F, 0x000EDDAA000ED1A5, 0x8000A167000EADAB,
-    0x000E99AE8000A16B, 0x000EB5AC000E99AF, 0x003505B2000E9DAD,
-    0x00306DB0003419B3, 0x8000A16F00350DB1, 0x002801B48000A173,
-    0x8000A17700293DB5, 0x8000A17F8000A17B, 0x8000A1878000A183,
-    0x8000A18F8000A18B, 0x000145A28000A193, 0x8000A1970001D5A3,
-    0x0005CDA600014DA1, 0x000401A4003801A7, 0x003C01AA00057DA5,
-    0x003949A8003D45AB, 0x003175AE0039E5A9, 0x003D5DAC00317DAF,
-    0x000EEDA8003001AD, 0x000E09AA000EF5A9, 0x000E19AC000E09AB,
-    0x000E0DAE000E19AD, 0x8000A19B000E2DAF, 0x8000A1A38000A19F,
-    0x8000A1AB8000A1A7, 0x8000A1B38000A1AF, 0x000E1DB88000A1B7,
-    0x000E2DBA000E25B9, 0x000E3DBC000E25BB, 0x0001D5BE0001DDBD,
-    0x000E55B00001CDBF, 0x000E55B2000E5DB1, 0x000E35B4000E2DB3,
-    0x000E2DB6000E25B5, 0x000D81B3000E25B7, 0x8000A1BF8000A1BB,
-    0x8000A1CB8000A1C7, 0x000D91B5000D99B6, 0x000D99BB000454BE,
-    0x00048886000D91BA, 0x000DE1BF0003BC87, 0x000DF9BD000DF1BE,
-    0x00004982000D81BC, 0x0003F580000DC5A3, 0x000DDDA600004981,
-    0x8000A1D38000A1CF, 0x000DD5AA000DD5A5, 0x8000A1D7000DDDAB,
-    0x000DB5AE8000A1DB, 0x000DC5AC000DA5AF, 0x00021DA8000DBDAD,
-    0x00025DAA000245A9, 0x000279AC000255AB, 0x000389AE000279AD,
-    0x8000A1DF000389AF, 0x8000A1E78000A1E3, 0x0005FC848000A1EB,
-    0x8000A1F38000A1EF, 0x000389B88000A1F7, 0x00036DBA000365B9,
-    0x00037DBC000365BB, 0x00036DBE000365BD, 0x0003C1B0000365BF,
-    0x0003C1B20003C1B1, 0x0003C1B40003C1B3, 0x0003C1B60003C1B5,
-    0x8000A1FB0003C1B7, 0x8000A2038000A1FF, 0x8000A20B8000A207,
-    0x8000A20F0001A4E1, 0x000468BE000EC0E3, 0x8000A2178000A213,
-    0x8000A21B00011CEF, 0x8000A2238000A21F, 0x000395B38000A227,
-    0x8000A22F8000A22B, 0x8000A23B8000A237, 0x0003B1B50003B9B6,
-    0x000245BB8000A23F, 0x0004A886000245BA, 0x000245BF0005A487,
-    0x000255BD000245BE, 0x000EE0E1000255BC, 0x000E08E3000C2CE1,
-    0x0000AD80000ED4E3, 0x0000D1820000D181, 0x8000A2478000A243,
-    0x8000A24F8000A24B, 0x8000A2578000A253, 0x000E88EF8000A25B,
-    0x000315A3000E2CEF, 0x002BC4858000A25F, 0x8000A2678000A263,
-    0x000331A5000339A6, 0x0002C5AB8000A26B, 0x8000A26F0002C5AA,
-    0x0002C5AF8000A273, 0x0002D5AD0002C5AE, 0x00060DA80002D5AC,
-    0x00065DAA000615A9, 0x000679AC000655AB, 0x0006B9AE000679AD,
-    0x8000A2330006B9AF, 0x8000A27B8000A277, 0x8000A2838000A27F,
-    0x8000A28B8000A287, 0x00074DB88000A28F, 0x000751BA000755B9,
-    0x000779BC000769BB, 0x00076DBE000765BD, 0x0006C9B0000765BF,
-    0x0006DDB20006C9B1, 0x0006CDB40006D5B3, 0x00077DB6000775B5,
-    0x0006D5B3000775B7, 0x8000A2978000A293, 0x8000A29F8000A29B,
-    0x0006F1B50006F9B6, 0x00060DBB8000A2A3, 0x0000088600060DBA,
-    0x0007B1BF00002C87, 0x000601BD000609BE, 0x00005182000609BC,
-    0x00004180000691A3, 0x0006BDA600004181, 0x8000A2AB8000A2A7,
-    0x000649AA0006B5A5, 0x8000A2AF000649AB, 0x00064DAE8000A2B3,
-    0x00064DAC0007F5AF, 0x0006B1B0000645AD, 0x0006CDB20006B1B1,
-    0x0006C1B40006C1B3, 0x0006F1B60006C9B5, 0x000105B80006F1B7,
-    0x000105BA00010DB9, 0x000105BC00011DBB, 0x000105BE00010DBD,
-    0x8000A2B70001B9BF, 0x8000A2BF8000A2BB, 0x8000A2C78000A2C3,
-    0x8000A1C38000A2CB, 0x00062DA88000A2CF, 0x000635AA000635A9,
-    0x0006D1AC0006F1AB, 0x0006D1AE0006D1AD, 0x0006DDB30006D1AF,
-    0x8000A2D78000A2D3, 0x8000A2DF8000A2DB, 0x000639B5000631B6,
-    0x000615BB8000A2E3, 0x8000A2E7000615BA, 0x00066DBF8000A2EB,
-    0x000675BD000665BE, 0x8000A2EF000679BC, 0x8000A2F3000699A3,
-    0x000675A68000A2F7, 0x8000A2FF8000A2FB, 0x000651AA00067DA5,
-    0x8000A303000651AB, 0x000621AE8000A307, 0x00063DAC000629AF,
-    0x0002D5A8000631AD, 0x0002E1AA0002DDA9, 0x000351AC0002E1AB,
-    0x000351AE000351AD, 0x8000A30B000351AF, 0x0003ECBE8000A30F,
-    0x8000A31B8000A317, 0x8000A3238000A31F, 0x0003E9B88000A327,
-    0x000389BA0003E9B9, 0x00039DBC000385BB, 0x000381BE000381BD,
-    0x000331B00003B5BF, 0x000335B2000331B1, 0x0003F9B40003E9B3,
-    0x0003E9B60003F9B5, 0x00036D800003E1B7, 0x0000AD820000A581,
-    0x8000A32B000265B3, 0x0003DDB60003D5B5, 0x0002E0848000A32F,
-    0x0003F9BA8000A333, 0x0003EDBC0003F9BB, 0x000331BE000331BD,
-    0x0003EC87000331BF, 0x000000B2003C6486, 0x8000A33B8000A337,
-    0x8000A33F000408E3, 0x8000A3430006ECE1, 0x8000A347000600EF,
-    0x8000A34F8000A34B, 0x8000A3578000A353, 0x8000A35F8000A35B,
-    0x8000A3678000A363, 0x0003A0E38000A36B, 0x0001A8E18000A36F,
-    0x0003F0EF8000A373, 0x00001D8100001D82, 0x8000A37700001D80,
-    0x8000A37F8000A37B, 0x003D4CBE8000A387, 0x0003A1A38000A38B,
-    0x000211A5003C40BE, 0x8000A3938000A38F, 0x00026C84000219A6,
-    0x00023DAB8000A397, 0x0002F5AD00023DAA, 0x0002F5AF000229AC,
-    0x003C40860002F5AE, 0x8000A39B003DB487, 0x8000A3A38000A39F,
-    0x8000A3AB8000A3A7, 0x8000A3AF000644EF, 0x8000A3B30006D4E1,
-    0x8000A3B700070CE3, 0x8000A3BF8000A3BB, 0x0001F5B38000A3C3,
-    0x8000A3CB8000A3C7, 0x8000A3D38000A3CF, 0x0001E5B5000145B6,
-    0x000131BB8000A3D7, 0x8000A3DB000129BA, 0x00011DBF8000A3DF,
-    0x000119BD00011DBE, 0x003E2DA8000125BC, 0x003E3DAA003E35A9,
-    0x003E2DAC003E35AB, 0x003E85AE003E85AD, 0x8000A383003EB9AF,
-    0x8000A3E78000A3E3, 0x000019808000A3EB, 0x0000058200001981,
-    0x003EB9B88000A3EF, 0x003F49BA003EB9B9, 0x003F59BC003F49BB,
-    0x003F4DBE003F59BD, 0x003EADB0003F41BF, 0x003EB1B2003EB5B1,
-    0x003E99B4003E8DB3, 0x003E89B6003E99B5, 0x003EB5A3003E89B7,
-    0x8000A3F300028C84, 0x8000A3FB8000A3F7, 0x003EA5A5003E05A6,
-    0x003E71AB8000A3FF, 0x00000886003E69AA, 0x003E5DAF00036887,
-    0x003E59AD003E5DAE, 0x8000A403003E65AC, 0x8000A407003F91B3,
-    0x003F59B68000A40B, 0x8000A4138000A40F, 0x003F75BA003FB9B5,
-    0x8000A417003F7DBB, 0x003F41BE8000A41B, 0x003F59BC003F41BF,
-    0x003E95B0003F55BD, 0x003EA9B2003E9DB1, 0x003EA1B4003EA5B3,
-    0x003EA1B6003EA1B5, 0x003EE5B8003EA1B7, 0x003EE5BA003EEDB9,
-    0x003EEDBC003EFDBB, 0x003ED5BE003EDDBD, 0x8000A41F000119BF,
-    0x8000A4278000A423, 0x8000A42F8000A42B, 0x8000A4338000A313,
-    0x003E5DA88000A437, 0x003E3DAA003E21A9, 0x003E15AC003E35AB,
-    0x003EF5AE003EFDAD, 0x003ED5A3003EEDAF, 0x8000A43F8000A43B,
-    0x8000A4478000A443, 0x003EFDA5003E1DA6, 0x003E39AB8000A44B,
-    0x8000A44F003E31AA, 0x003E05AF8000A453, 0x003E11AD003E05AE,
-    0x00001181003E1DAC, 0x8000A45700000D80, 0x8000A45B00001982,
-    0x000194848000A45F, 0x00074086000010BE, 0x8000A46700010087,
-    0x8000A46F8000A46B, 0x8000A4778000A473, 0x00028DA88000A47B,
-    0x000295AA000295A9, 0x0002D9AC0002CDAB, 0x0002CDAE0002D9AD,
-    0x8000A47F0002C5AF, 0x8000A4878000A483, 0x0000008C8000A48B,
-    0x8000A4938000A48F, 0x00031DB88000A497, 0x0003C1BA0003C1B9,
-    0x0003C1BC0003C1BB, 0x0003F1BE0003C9BD, 0x000289B00003F1BF,
-    0x000329B2000289B1, 0x000339B4000329B3, 0x000329B6000339B5,
-    0x00021DB3000325B7, 0x8000A49F8000A49B, 0x8000A4A78000A4A3,
-    0x000251B5000259B6, 0x00024DBB8000A4AB, 0x8000A4AF00024DBA,
-    0x0003FDBF8000A4B3, 0x0003FDBD0003FDBE, 0x8000A4B70003FDBC,
-    0x8000A4BF8000A4BB, 0x003E0CE18000A4C3, 0x003FA8E38000A4C7,
-    0x00003D818000A4CB, 0x003F50EF00003180, 0x8000A4CF00001D82,
-    0x000518860004F8BE, 0x00020C8400034C87, 0x8000A4D70000CCE3,
-    0x8000A4DB0001BCE1, 0x8000A4E38000A4DF, 0x8000A4EB8000A4E7,
-    0x8000A4EF00052C84, 0x8000A4F78000A4F3, 0x8000A4FB000010EF,
-    0x0003DDA38000A4FF, 0x8000A5078000A503, 0x8000A50F8000A50B,
-    0x000391A5000399A6, 0x00038DAB8000A513, 0x8000A51700038DAA,
-    0x00023DAF8000A51B, 0x00023DAD00023DAE, 0x8000A51F00023DAC,
-    0x8000A5278000A523, 0x8000A52F8000A52B, 0x8000A5378000A533,
-    0x8000A53B003E28EF, 0x8000A53F003E3CE1, 0x00002980000138E3,
-    0x0000118200001581, 0x000231B38000A547, 0x0004C0840004C8BE,
-    0x8000A54F8000A54B, 0x000221B5000229B6, 0x0001CDBB8000A553,
-    0x8000A5570001CDBA, 0x000175BF8000A55B, 0x0001C1BD0001C9BE,
-    0x0005E9A80001C9BC, 0x0005F9AA0005E9A9, 0x0005E9AC0005F9AB,
-    0x000639AE0005E9AD, 0x8000A4D3000639AF, 0x000088868000A543,
-    0x8000A55F00034087, 0x8000A5678000A563, 0x0006D1B88000A56B,
-    0x0006E1BA0006D9B9, 0x000691BC0006E1BB, 0x000691BE000691BD,
-    0x000649B0000691BF, 0x00065DB2000649B1, 0x00064DB4000655B3,
-    0x0006F1B60006F1B5, 0x000571A30006F1B7, 0x8000A5738000A56F,
-    0x8000A57B8000A577, 0x000561A5000569A6, 0x00068DAB8000A57F,
-    0x8000A58300068DAA, 0x000635AF8000A587, 0x000681AD000689AE,
-    0x8000A58B000689AC, 0x8000A58F0007D1B3, 0x0007F1B68000A593,
-    0x8000A59B8000A597, 0x0007D1BA0007C1B5, 0x8000A59F0007DDBB,
-    0x0007B9BE8000A5A3, 0x0007C5BC0007B9BF, 0x000669B80007B9BD,
-    0x000689BA000669B9, 0x000699BC000689BB, 0x000689BE000699BD,
-    0x000611B0000689BF, 0x000615B200061DB1, 0x000679B4000669B3,
-    0x000669B6000679B5, 0x000649A8000661B7, 0x00065DAA000655A9,
-    0x00064DAC000655AB, 0x000671AE000675AD, 0x8000A463000671AF,
-    0x00001D8100001D82, 0x8000A5A700001D80, 0x8000A5AF8000A5AB,
-    0x000695A3000170BE, 0x000028868000A5B7, 0x8000A5BB00014887,
-    0x000685A50006B5A6, 0x000699AB8000A5BF, 0x8000A5C3000695AA,
-    0x0006FDAF8000A5C7, 0x0006FDAD0006FDAE, 0x8000A5CB000681AC,
-    0x8000A5CF000615B3, 0x00063DB68000A5D3, 0x8000A5DB8000A5D7,
-    0x0001D9BA000635B5, 0x8000A5DF0001D9BB, 0x00017DBE8000A5E3,
-    0x00017DBC000165BF, 0x0005BDA8000175BD, 0x0005D9AA0005C9A9,
-    0x0005F9AC0005D1AB, 0x000229AE0005F9AD, 0x8000A5E7000229AF,
-    0x8000A5EF8000A5EB, 0x0000008C8000A5F3, 0x8000A5FB8000A5F7,
-    0x0002EDB88000A5FF, 0x000281BA000285B9, 0x000285BC000281BB,
-    0x0002B1BE00028DBD, 0x000259B00002B1BF, 0x0002EDB2000259B1,
-    0x0002FDB40002E5B3, 0x0002E5B60002E5B5, 0x000551A30002D5B7,
-    0x8000A6078000A603, 0x8000A60F8000A60B, 0x000571A5000579A6,
-    0x00029DAB8000A613, 0x8000A61700029DAA, 0x000221AF8000A61B,
-    0x000231AD000239AE, 0x00006D81000239AC, 0x8000A61F00006D80,
-    0x000C5CBE00000582, 0x8000A62B8000A627, 0x0003EC840006D0EF,
-    0x8000A62F00051CE1, 0x8000A6330007FCE3, 0x000CD8868000A637,
-    0x000285A8000CBC87, 0x000295AA000295A9, 0x0002BDAC0002A5AB,
-    0x0002D1AE0002D5AD, 0x8000A63B0002D1AF, 0x8000A6438000A63F,
-    0x8000A64B8000A647, 0x8000A6538000A64F, 0x000175B88000A657,
-    0x000175BA00017DB9, 0x0001D9BC0001C9BB, 0x0001C9BE0001D9BD,
-    0x0002B5B00001C1BF, 0x000281B20002BDB1, 0x000151B4000281B3,
-    0x000151B6000151B5, 0x8000A65B000151B7, 0x8000A65F000D0484,
-    0x000C1CBE8000A663, 0x00001CEF8000A667, 0x000190E10006A0EF,
-    0x000670E3000744E1, 0x8000A66B000698E3, 0x8000A6738000A66F,
-    0x00025DB38000A677, 0x8000A67F8000A67B, 0x8000A6878000A683,
-    0x000275B5000215B6, 0x000239BB8000A68B, 0x8000A68F000231BA,
-    0x0001D5BF8000A697, 0x000215BD0001D5BE, 0x000D9DA3000215BC,
-    0x8000A6938000A623, 0x8000A69F8000A69B, 0x000DB5A5000DD5A6,
-    0x000DF9AB8000A6A3, 0x00030886000DF1AA, 0x000E15AF00036087,
-    0x000DD5AD000E15AE, 0x000F9180000DD5AC, 0x000FA182000F9981,
-    0x8000A6A7000EA5B3, 0x000F79B6000EA1B5, 0x8000A6AF8000A6AB,
-    0x000F45BA8000A6B3, 0x000F45BC000F5DBB, 0x000F45BE000F4DBD,
-    0x000E55A8000FFDBF, 0x000E61AA000E5DA9, 0x000E65AC000E7DAB,
-    0x000FBDAE000FB5AD, 0x8000A6B7000FB5AF, 0x8000A6BF8000A6BB,
-    0x8000A6C78000A6C3, 0x8000A6CF8000A6CB, 0x000F55B88000A6D3,
-    0x000F7DBA000F75B9, 0x000F6DBC000F75BB, 0x000F11BE000F11BD,
-    0x000FCDB0000F11BF, 0x000FDDB2000FD5B1, 0x000FCDB4000FD5B3,
-    0x000F71B6000F75B5, 0x000FE9A3000F71B7, 0x8000A6DB8000A6D7,
-    0x8000A6E38000A6DF, 0x000FEDA5000E35A6, 0x000E11AB8000A6E7,
-    0x8000A6EB000E09AA, 0x000EB1AF8000A6EF, 0x000E01AD000E09AE,
-    0x8000A6F3000E09AC, 0x00001D8100001D82, 0x8000A6F700001D80,
-    0x8000A6FF8000A6FB, 0x0001AC84000194BE, 0x0001E0878000A703,
-    0x8000A7070000CC86, 0x8000A70F8000A70B, 0x000E6DA88000A5B3,
-    0x000199AA000189A9, 0x0001C1AC000191AB, 0x0001F9AE0001C9AD,
-    0x0000A0840001F9AF, 0x8000A7178000A713, 0x8000A71F8000A71B,
-    0x8000A7278000A723, 0x0000C5B88000A72B, 0x0000C5BA0000CDB9,
-    0x0000CDBC0000DDBB, 0x0000F5BE0000FDBD, 0x000141B000009DBF,
-    0x000141B2000141B1, 0x000141B4000141B3, 0x000141B6000141B5,
-    0x000211B3000141B7, 0x8000A7338000A72F, 0x8000A73B8000A737,
-    0x000249B5000249B6, 0x000285BB8000A73F, 0x8000A743000285BA,
-    0x000285BF8000A747, 0x000295BD000285BE, 0x001A3C85000295BC,
-    0x8000A74B000255A3, 0x00020DA68000A74F, 0x8000A7578000A753,
-    0x0002C1AA00020DA5, 0x8000A75B0002C1AB, 0x0002C1AE8000A75F,
-    0x0002D1AC0002C1AF, 0x000019820002D1AD, 0x000019808000A763,
-    0x8000A76700001981, 0x8000A76F8000A76B, 0x000400BE8000A777,
-    0x8000A77F8000A77B, 0x8000A7878000A783, 0x8000A78F8000A78B,
-    0x000EE8EF8000A793, 0x0004E8868000A797, 0x8000A79B00035087,
-    0x000280EF8000A79F, 0x000194E18000A7A3, 0x000158E38000A7A7,
-    0x000E00E38000A7AB, 0x000D68E18000A7AF, 0x000271A88000A7B3,
-    0x000271AA000271A9, 0x0002B9AC0002A9AB, 0x0002A9AE0002B9AD,
-    0x0005AC840002A9AF, 0x8000A7BB8000A7B7, 0x8000A7C38000A7BF,
-    0x8000A7CB8000A7C7, 0x00016DB88000A7CF, 0x000105BA00010DB9,
-    0x000109BC000119BB, 0x000135BE00013DBD, 0x0002D9B00001D9BF,
-    0x00016DB20002D9B1, 0x00017DB4000165B3, 0x000165B6000165B5,
-    0x000F60E1000155B7, 0x000E1CE30000F4E3, 0x8000A7D30001BCE1,
-    0x0000318100003982, 0x8000A7D700003D80, 0x8000A7DB000428BE,
-    0x00073CBE8000A7DF, 0x000ED0EF000010EF, 0x0004E0868000A7E7,
-    0x8000A7EB0004C887, 0x8000A7EF0002B5B3, 0x000275B600027DB5,
-    0x8000A7F78000A7F3, 0x000251BA8000A7FB, 0x0001F5BC000251BB,
-    0x0001F5BE0001FDBD, 0x8000A7730001EDBF, 0x0005B1AA8000A7E3,
-    0x000615AC0005B1AB, 0x000615AE00061DAD, 0x8000A7FF00060DAF,
-    0x8000A8078000A803, 0x8000A80B000555A3, 0x000595A600059DA5,
-    0x0006E9B38000A80F, 0x8000A8178000A813, 0x8000A81F8000A81B,
-    0x0006E9B50006E1B6, 0x0006B1BB8000A823, 0x8000A8270006A1BA,
-    0x00069DBF8000A82B, 0x0006A5BD000691BE, 0x00061DA80006A9BC,
-    0x000621AA000621A9, 0x000621AC000621AB, 0x000621AE000621AD,
-    0x8000A82F000621AF, 0x8000A8378000A833, 0x8000A83F8000A83B,
-    0x8000A8478000A843, 0x0007E5B88000A84B, 0x0007E5BA0007EDB9,
-    0x0007E5BC0007FDBB, 0x0007E5BE0007EDBD, 0x000625B000074DBF,
-    0x00063DB2000635B1, 0x000615B4000631B3, 0x00060DB6000619B5,
-    0x0007A9A3000601B7, 0x0001B58100001582, 0x8000A84F0001B580,
-    0x0007A9A50007A1A6, 0x0007F1AB8000A853, 0x0002A0840007E1AA,
-    0x0007DDAF8000A857, 0x0007E5AD0007D1AE, 0x8000A85B0007E9AC,
-    0x0000E886000695B3, 0x0006B5B600011887, 0x8000A8638000A85F,
-    0x000149BA0006BDB5, 0x8000A867000155BB, 0x000139BE8000A86B,
-    0x000145BC000139BF, 0x00063DA8000139BD, 0x000659AA00064DA9,
-    0x000671AC000651AB, 0x0001B9AE000671AD, 0x0001AC840001B9AF,
-    0x8000A8738000A86F, 0x8000A87B8000A877, 0x8000A8838000A87F,
-    0x0001A9B88000A887, 0x000169BA0001A9B9, 0x000179BC000169BB,
-    0x000169BE000179BD, 0x0001C9B0000169BF, 0x0001D5B20001D5B1,
-    0x0001B9B40001A9B3, 0x0001A9B60001B9B5, 0x0005D1A30001A1B7,
-    0x8000A88F8000A88B, 0x8000A8978000A893, 0x0005F9A50005F1A6,
-    0x000211AB8000A89B, 0x8000A89F00020DAA, 0x00027DAF8000A8A3,
-    0x00027DAD00027DAE, 0x8000A8A7000201AC, 0x8000A8AF8000A8AB,
-    0x000031818000A8B3, 0x8000A8B700000D80, 0x8000A8BB00003982,
-    0x000324BE8000A8BF, 0x8000A8CB8000A8C7, 0x001CD8868000A8CF,
-    0x8000A8D300034C87, 0x8000A8DB8000A8D7, 0x000120E3001CC084,
-    0x0001C8E18000A8DF, 0x000230EF8000A8E3, 0x8000A8EB8000A8E7,
-    0x8000A8F38000A8EF, 0x8000A8FB8000A8F7, 0x000395B38000A8FF,
-    0x8000A9078000A903, 0x8000A90F8000A90B, 0x0003B1B50003B9B6,
-    0x000351BB8000A913, 0x8000A917000349BA, 0x0000F5BF8000A91B,
-    0x000341BD000349BE, 0x00022DA8000349BC, 0x00025DAA000245A9,
-    0x000279AC000255AB, 0x0002BDAE000279AD, 0x001D68BE0002B5AF,
-    0x8000A9238000A91F, 0x00001D808000A927, 0x0000A98200000981,
-    0x000151B88000A92B, 0x000161BA000159B9, 0x000111BC000161BB,
-    0x000111BE000111BD, 0x0002CDB0000111BF, 0x0002DDB20002D5B1,
-    0x0002CDB40002D5B3, 0x000171B6000171B5, 0x000658E1000171B7,
-    0x0000B0E3000754E1, 0x8000A92F0006BCE3, 0x001C18868000A937,
-    0x8000A93B001D5487, 0x8000A9438000A93F, 0x001CF8BE8000A947,
-    0x0006F0EF8000A94B, 0x000295A30006E0EF, 0x8000A9538000A94F,
-    0x8000A95B8000A957, 0x0002B1A50002B9A6, 0x000251AB8000A95F,
-    0x8000A963000249AA, 0x0001F5AF8000A967, 0x000241AD000249AE,
-    0x001E6DA8000249AC, 0x001E7DAA001E75A9, 0x001E95AC001E8DAB,
-    0x001E8DAE001E9DAD, 0x8000A933001E81AF, 0x8000A96F8000A96B,
-    0x8000A9778000A973, 0x8000A97F8000A97B, 0x001E89B88000A983,
-    0x001E99BA001E89B9, 0x001EB9BC001E91BB, 0x001F7DBE001EB9BD,
-    0x001EC5B0001F75BF, 0x001EC5B2001ECDB1, 0x001EC5B4001EDDB3,
-    0x001EC5B6001ECDB5, 0x001EF5B3001EB9B7, 0x8000A98B8000A987,
-    0x8000A9938000A98F, 0x001E1DB5001E1DB6, 0x001E09BB8000A997,
-    0x8000A99B001E01BA, 0x001E09BF8000A99F, 0x001E09BD001E01BE,
-    0x00006D82001E11BC, 0x00005580001EB1A3, 0x001E59A600006581,
-    0x0001D0BF00039884, 0x001E45AA001E59A5, 0x00040086001E4DAB,
-    0x001E45AE00019887, 0x001E55AC001E4DAF, 0x8000A9A7001E4DAD,
-    0x000024848000A9AB, 0x8000A9B38000A9AF, 0x8000A8C38000A9B7,
-    0x001EB5A88000A9A3, 0x001E8DAA001E85A9, 0x001E9DAC001E85AB,
-    0x001EFDAE001E81AD, 0x001E8DB0001EF5AF, 0x001E95B2001E95B1,
-    0x001EBDB4001EA5B3, 0x000171B6000171B5, 0x000151B8000171B7,
-    0x000151BA000151B9, 0x0001F5BC000151BB, 0x0001F5BE0001FDBD,
-    0x001E25B30001EDBF, 0x8000A9BB000708BE, 0x8000A9C38000A9BF,
-    0x001E35B5001E21B6, 0x001E71BB8000A9C7, 0x8000A9CB001E69BA,
-    0x000195BF8000A9CF, 0x001E51BD001E59BE, 0x8000A9D3001E61BC,
-    0x8000A9D7001E61A3, 0x001E65A68000A9DB, 0x8000A9E38000A9DF,
-    0x001E2DAA001E71A5, 0x8000A9E7001E35AB, 0x001E1DAE8000A9EB,
-    0x001E25AC0001D1AF, 0x001A54E1001E15AD, 0x000A5CE38000A9EF,
-    0x8000A9F78000A9F3, 0x8000A9FF8000A9FB, 0x8000AA078000AA03,
-    0x8000AA0B0005A8BE, 0x002A318F8000AA0F, 0x00FBC4EF8000AA17,
-    0x002E61938000AA1B, 0x002AD991002F1D92, 0x0012D997002A4990,
-    0x00130D9500127596, 0x001B1D9B002EC194, 0x001649990016419A,
-    0x001E719F00170D98, 0x001A719D001B119E, 0x0002A1A3001A799C,
-    0x001F95A100039DA2, 0x000188E3001E89A0, 0x000068E1001E08E3,
-    0x0001DCEF001EFCE1, 0x00FE89B5001F7CEF, 0x00FA01B300FC01B4,
-    0x00F801B100FA95B2, 0x00F6E1AF00F611B0, 0x00F225AD00F401AE,
-    0x00F001AB00F2EDAC, 0x000ED5A9000FE9AA, 0x000C01A7000EDDA8,
-    0x000AEDA5000AC9A6, 0x0006E1A3000801A4, 0x0003E5A10006F1A2,
-    0x002B01828000AA1F, 0x8000AA23002B0183, 0x002F31868000AA27,
-    0x002B8984002F8987, 0x0012758A002E4585, 0x000508860013058B,
-    0x0017118E00056C87, 0x0013B18C0017118F, 0x001A3D9200165D8D,
-    0x0005C884001B4193, 0x001F519600070084, 0x001B5194001F5197,
-    0x001E7D9A001A3995, 0x8000AA2B00027D9B, 0x000159818000AA2F,
-    0x0003599C00015580, 0x0007C8BE00015182, 0x8000AA378000AA33,
-    0x8000AA3F8000AA3B, 0x8000AA478000AA43, 0x8000AA4B001ED4EF,
-    0x8000AA4F001E24E1, 0x8000AA53000168E3, 0x8000AA5B8000AA57,
-    0x000269BB8000AA5F, 0x8000AA63000259BA, 0x00021DBF8000AA67,
-    0x000219BD00021DBE, 0x0002EDB3000271BC, 0x8000AA6F8000AA6B,
-    0x8000AA778000AA73, 0x0002EDB500027DB6, 0x00050DA38000AA7B,
-    0x8000AA7F8000AA13, 0x8000AA838000AA87, 0x00050DA500059DA6,
-    0x000589AB8000AA8B, 0x000308860005B9AA, 0x0005FDAF00037C87,
-    0x0005F9AD0005FDAE, 0x0007B180000591AC, 0x000041820007B981,
-    0x8000AA8F0006B1B3, 0x000731B6000759B5, 0x8000AA978000AA93,
-    0x0007E1BA8000AA9B, 0x0007E1BC0007E1BB, 0x0007E1BE0007E1BD,
-    0x0006B5A80007DDBF, 0x000675AA0006B9A9, 0x00079DAC000785AB,
-    0x0007F5AE0007FDAD, 0x8000AA9F000719AF, 0x8000AAA78000AAA3,
-    0x8000AAAF8000AAAB, 0x8000AAB78000AAB3, 0x000775B88000AABB,
-    0x000775BA00077DB9, 0x00071DBC000705BB, 0x000731BE000731BD,
-    0x000769B0000731BF, 0x000779B2000769B1, 0x000769B4000779B3,
-    0x000751B6000755B5, 0x0007FDA300074DB7, 0x8000AAC38000AABF,
-    0x8000AACB8000AAC7, 0x000615A500067DA6, 0x0006ADAB8000AACF,
-    0x8000AAD30006ADAA, 0x000691AF8000AAD7, 0x0006ADAD0006ADAE,
-    0x8000AADB0006ADAC, 0x8000AAE38000AADF, 0x00001D808000AAE7,
-    0x0000398200000981, 0x8000AAEF8000AAEB, 0x0000C8868000AAF7,
-    0x8000AAFB0001A487, 0x8000AB038000AAFF, 0x00018DA88000AB07,
-    0x0001D9AA000199A9, 0x0001F1AC0001D1AB, 0x000139AE0001F1AD,
-    0x0000A084000139AF, 0x8000AB0F8000AB0B, 0x8000AB178000AB13,
-    0x8000AB1F8000AB1B, 0x0000CDB88000AB23, 0x0000D5BA0000D5B9,
-    0x0000FDBC0000E5BB, 0x00009DBE000095BD, 0x000149B0000095BF,
-    0x000159B2000149B1, 0x000149B4000159B3, 0x0000FDB6000149B5,
-    0x000405BA0000F5B7, 0x0007E1B8000409BB, 0x000409BE0007E1B9,
-    0x000409BC000409BF, 0x00078DB2000409BD, 0x0007B5B00007E5B3,
-    0x0007E5B6000785B1, 0x0007F1B40007EDB7, 0x00074DAA0007F1B5,
-    0x000749A8000755AB, 0x0007DDAE000749A9, 0x00074DAC0007C5AF,
-    0x8000AB270007D5AD, 0x8000AB2F8000AB2B, 0x8000AB378000AB33,
-    0x8000AB3F8000AB3B, 0x0002D1B38000AB43, 0x8000AB4B8000AB47,
-    0x8000AB4F000C00BE, 0x0002F9B50002F1B6, 0x00020DBB8000AB53,
-    0x8000AB5700020DBA, 0x00020DBF8000AB5B, 0x00020DBD00020DBE,
-    0x0000158200020DBC, 0x00006180000295A3, 0x0002B5A600006181,
-    0x8000AB678000AB5F, 0x000249AA0002BDA5, 0x000CC886000249AB,
-    0x000249AE000CAC87, 0x000249AC000249AF, 0x000265A8000249AD,
-    0x00027DAA000275A9, 0x00026DAC000275AB, 0x0002B1AE0002B1AD,
-    0x000DA0840002B1AF, 0x8000AB6F8000AB6B, 0x8000AB778000AB73,
-    0x8000AB7F8000AB7B, 0x000131B88000AB83, 0x000131BA000131B9,
-    0x0001D5BC000131BB, 0x0001C9BE0001DDBD, 0x0002D1B00001C1BF,
-    0x0002D1B20002D1B1, 0x000111B40002D1B3, 0x000111B6000111B5,
-    0x0006B0E1000111B7, 0x0006D0E38000AB87, 0x8000AB8B000F4084,
-    0x8000AB8F0001A4E1, 0x8000AB930006C0E3, 0x8000AB9B8000AB97,
-    0x000000EC0006D4EF, 0x000764EF8000AB9F, 0x8000ABA78000ABA3,
-    0x8000ABAF8000ABAB, 0x8000ABB30002B9B3, 0x00027DB60002A9B5,
-    0x8000ABBB8000ABB7, 0x000259BA8000ABBF, 0x000249BC000259BB,
-    0x000199BE000249BD, 0x000D75A3000199BF, 0x8000ABC38000AB63,
-    0x8000ABCB8000ABC7, 0x000D65A5000DB1A6, 0x000D95AB8000ABCF,
-    0x0003A4BE000D95AA, 0x000E55AF8000ABD3, 0x000D85AD000E55AE,
-    0x00004D80000D85AC, 0x0000558200005581, 0x8000ABD7000F65B3,
-    0x000F6DB6000F65B5, 0x000340868000ABDB, 0x000FEDBA00031487,
-    0x000FE9BC000FFDBB, 0x000FDDBE000FD5BD, 0x000E59A8000FD5BF,
-    0x000FA1AA000FA1A9, 0x000FA1AC000FA1AB, 0x000FA1AE000FA1AD,
-    0x8000ABDF000FA1AF, 0x8000ABE78000ABE3, 0x8000ABEF8000ABEB,
-    0x8000ABF78000ABF3, 0x000F01B88000ABFB, 0x000F01BA000F01B9,
-    0x000F0DBC000F1DBB, 0x000F3DBE000F35BD, 0x000F65B00000D5BF,
-    0x000F71B2000F75B1, 0x000F55B4000F4DB3, 0x000F4DB6000F5DB5,
-    0x000E29A3000F41B7, 0x8000AC038000ABFF, 0x8000AC0B8000AC07,
-    0x000E29A5000E21A6, 0x000EB1AB8000AC0F, 0x8000AC13000EA1AA,
-    0x000E99AF8000AC17, 0x000E99AD000E91AE, 0x8000AC1B000EA5AC,
-    0x8000AC238000AC1F, 0x000D24EF8000AC27, 0x8000AC2F8000AC2B,
-    0x000ED0E38000AC33, 0x000E18E18000AC37, 0x000015808000AC3B,
-    0x0000058200001981, 0x0001D1A88000AC3F, 0x000129AA0001D9A9,
-    0x000139AC000129AB, 0x000129AE000139AD, 0x000168BE000129AF,
-    0x0001C8868000AC47, 0x8000AC4B0000AC87, 0x8000AC538000AC4F,
-    0x0000EDB88000AC57, 0x00008DBA000085B9, 0x00009DBC000085BB,
-    0x000081BE000081BD, 0x000159B0000081BF, 0x0000EDB2000159B1,
-    0x0000FDB40000E5B3, 0x0000E5B60000E5B5, 0x0002A1B30000D5B7,
-    0x8000AC5F8000AC5B, 0x8000AC678000AC63, 0x0002B1B50002B9B6,
-    0x00029DBB8000AC6B, 0x8000AC6F00029DBA, 0x000321BF8000AC73,
-    0x000331BD000339BE, 0x00F5F885000339BC, 0x8000AC770002E5A3,
-    0x0002FDA68000AC7B, 0x8000AC838000AC7F, 0x0002D9AA0002F5A5,
-    0x8000AC870002D9AB, 0x00037DAE8000AC8B, 0x00037DAC000365AF,
-    0x0000E9B8000375AD, 0x000069BA0000E9B9, 0x000079BC000069BB,
-    0x00006DBE000065BD, 0x0000B1B0000065BF, 0x000081B20000B9B1,
-    0x0000F9B4000081B3, 0x0000EDB60000F9B5, 0x000365A80000E5B7,
-    0x00037DAA000375A9, 0x00036DAC000375AB, 0x0000D1AE0000D1AD,
-    0x8000AC8F0000D1AF, 0x8000AC978000AC93, 0x8000AC9B8000AAF3,
-    0x8000ACA38000AC9F, 0x0001FD808000ACA7, 0x0000198200000981,
-    0x8000ACAF8000ACAB, 0x8000ACB7000244BE, 0x0002AC848000ACBB,
-    0x0002F8878000ACBF, 0x8000ACC30005AC86, 0x8000ACCB8000ACC7,
-    0x0002F5B38000ACCF, 0x8000ACD78000ACD3, 0x8000ACDF8000ACDB,
-    0x0002E5B5000151B6, 0x00017DBB8000ACE3, 0x8000ACE7000175BA,
-    0x00013DBF8000ACEB, 0x000155BD00013DBE, 0x000F70E1000155BC,
-    0x000EB8E38000ACEF, 0x0000C8EF0005C084, 0x8000ACF78000ACF3,
-    0x000E3CE38000ACFB, 0x0001D0E18000ACFF, 0x000774848000AD03,
-    0x000168EF8000AD07, 0x8000AD0F8000AD0B, 0x0001CDA6000279A5,
-    0x000021808000AD13, 0x0007DD8200002181, 0x0001C9AC000269A3,
-    0x0001A1AE0001C9AD, 0x8000AD170001A1AF, 0x0001E9AA8000AD1B,
-    0x8000ACB30001E1AB, 0x000240BE8000AD1F, 0x000330868000AD23,
-    0x8000AD2700033087, 0x000739A88000AD2B, 0x00070DAA000739A9,
-    0x000705AC00071DAB, 0x000705AE00070DAD, 0x000709B0000779AF,
-    0x000719B2000709B1, 0x000739B4000711B3, 0x0007DDB6000739B5,
-    0x0007F9B80007C9B7, 0x0007CDBA0007F9B9, 0x0007C5BC0007C1BB,
-    0x0007B9BE0007C9BD, 0x0007A5B30007B1BF, 0x8000AD338000AD2F,
-    0x8000AD3B8000AD37, 0x0007ADB50007ADB6, 0x0007EDBB8000AD3F,
-    0x8000AD430007EDBA, 0x0007DDBF8000AD47, 0x0007E5BD0007DDBE,
-    0x8000AD4B0007EDBC, 0x8000AD4F0007E1A3, 0x0007E9A68000AD53,
-    0x8000AD5B8000AD57, 0x0007A9AA0007E9A5, 0x8000AD5F0007A9AB,
-    0x000799AE8000AD63, 0x0007A9AC000799AF, 0x8000AD670007A1AD,
-    0x8000AD6F8000AD6B, 0x8000AD778000AD73, 0x8000AD7F8000AD7B,
-    0x000011818000AD83, 0x8000AD8700000D80, 0x8000AD8B00001982,
-    0x0001E4BE8000AD8F, 0x0000E0868000AD93, 0x8000AD9700011887,
-    0x8000AD9F8000AD9B, 0x8000ADA78000ADA3, 0x8000ADAB0004B0EF,
-    0x8000ADAF0006D4E1, 0x8000ADB300045CE3, 0x8000ADBB8000ADB7,
-    0x000199A88000ADBF, 0x00010DAA000199A9, 0x00011DAC000105AB,
-    0x000105AE000105AD, 0x00018884000135AF, 0x8000ADC78000ADC3,
-    0x8000ADCF8000ADCB, 0x8000ADD78000ADD3, 0x0000C1B88000ADDB,
-    0x0000C1BA0000C1B9, 0x0000C1BC0000C1BB, 0x0000C1BE0000C1BD,
-    0x00014DB00000C1BF, 0x000121B2000125B1, 0x000125B4000121B3,
-    0x000111B600012DB5, 0x8000ADDF000111B7, 0x00028DB38000ADE3,
-    0x00029DB58000ADE7, 0x8000ADEF8000ADEB, 0x8000ADF300028DB6,
-    0x0002E9BB8000ADF7, 0x0002FDBD000281BA, 0x0002E1BF0002FDBC,
-    0x8000ADFB0002E9BE, 0x0002C5A50002D5A6, 0x0002D5A3000308BE,
-    0x0000198100002D82, 0x0002B9AF00007980, 0x0002A5AD0002B1AE,
-    0x0002B1AB0002A5AC, 0x8000ADFF0002D9AA, 0x000234EF8000AE07,
-    0x000318878000AE0B, 0x8000AE0F00FC2C86, 0x8000AE178000AE13,
-    0x8000AE1F8000AE1B, 0x8000AE278000AE23, 0x000100E38000AE2B,
-    0x0001E0E18000AE2F, 0x0003B5BA8000AE33, 0x8000AE370003BDBB,
-    0x000379BE8000AE3B, 0x0003A5BC000379BF, 0x8000AC43000379BD,
-    0x8000AE3F0003B5B3, 0x000391B68000AE43, 0x8000AE4B8000AE47,
-    0x000249A80003A5B5, 0x000259AA000249A9, 0x000249AC000259AB,
-    0x000271AE000275AD, 0x00FD68BE00026DAF, 0x8000AE4F00FFA4BE,
-    0x8000AE578000AE53, 0x8000AE5F8000AE5B, 0x0002F9B88000AE63,
-    0x000149BA0002F9B9, 0x00015DBC000149BB, 0x000141BE000141BD,
-    0x000215B000017DBF, 0x000215B200021DB1, 0x0002D9B40002C9B3,
-    0x0002C9B60002D9B5, 0x000620E30002C9B7, 0x000180E10006B0E1,
-    0x00000D80000610E3, 0x00003D8200003581, 0x8000AE6B8000AE67,
-    0x8000AE778000AE6F, 0x000070EF8000AE7B, 0x8000AE838000AE7F,
-    0x00FD8C8400014CEF, 0x000299A38000AE87, 0x000289A58000AE8B,
-    0x8000AE938000AE8F, 0x00FCC0860002BDA6, 0x000291AB00FCE487,
-    0x000255AD000299AA, 0x000255AF000289AC, 0x00FE7DA8000255AE,
-    0x00FE99AA00FE81A9, 0x00FE89AC00FE99AB, 0x00FEB9AE00FE89AD,
-    0x8000AE7300FEB9AF, 0x8000AE9B8000AE97, 0x8000AEA38000AE9F,
-    0x8000AEAB8000AEA7, 0x00FEB5B88000AEAF, 0x00FF61BA00FEBDB9,
-    0x00FF61BC00FF61BB, 0x00FF61BE00FF61BD, 0x00FEC9B000FF61BF,
-    0x00FEDDB200FEC9B1, 0x00FEB9B400FED1B3, 0x00FE91B600FEB9B5,
-    0x00FE11B300FE91B7, 0x8000AEB78000AEB3, 0x8000AEBF8000AEBB,
-    0x00FE01B500FE09B6, 0x00FE0DBB8000AEC3, 0x8000AEC700FE0DBA,
-    0x00FE7DBF8000AECB, 0x00FE05BD00FE7DBE, 0x8000AECF00FE09BC,
-    0x8000AED300FE55A3, 0x00FE4DA68000AED7, 0x8000AEDF8000AEDB,
-    0x00FE49AA00FE45A5, 0x0003288400FE49AB, 0x00FE39AE8000AEE3,
-    0x00FE4DAC00FE39AF, 0x0001CD8000FE41AD, 0x0001D1820001D181,
-    0x8000AEE700FEB9B3, 0x00FED5B600FED1B5, 0x000160868000AEEB,
-    0x00010DBA00016087, 0x00011DBC000105BB, 0x00010DBE000105BD,
-    0x8000AEEF000105BF, 0x8000AEF78000AEF3, 0x8000AEFF8000AEFB,
-    0x8000AF0300FDC0E1, 0x8000AF0700FC8CE3, 0x8000AF0F8000AF0B,
-    0x8000AF1300FCB4EF, 0x8000AF1B8000AF17, 0x00FE29A88000AF1F,
-    0x00FE3DAA00FE29A9, 0x00FE55AC00FE35AB, 0x00FE4DAE00FE59AD,
-    0x8000AF2300FE45AF, 0x8000AF2B8000AF27, 0x8000AF338000AF2F,
-    0x8000AF3B8000AF37, 0x000149B88000AF3F, 0x000159BA000149B9,
-    0x000179BC000151BB, 0x000119BE000179BD, 0x0001C5B0000115BF,
-    0x0001C5B20001CDB1, 0x0001C5B40001DDB3, 0x0001C5B60001CDB5,
-    0x8000AF43000179B7, 0x8000AF4B8000AF47, 0x8000AF4F00FDB9A3,
-    0x00FDD5A600FDD1A5, 0x8000AF5E0003D084, 0x00020DAA0002E0BE,
-    0x00021DAC000205AB, 0x00020DAE000205AD, 0x00004981000205AF,
-    0x000309A300004180, 0x000319A500005D82, 0x8000AF668000AF62,
-    0x0004C086000311A6, 0x00030DAB0003E487, 0x000301AD00030DAA,
-    0x000301AF00031DAC, 0x0003E1B0000309AE, 0x0003E1B20003E1B1,
-    0x0003E5B40003FDB3, 0x0003E5B60003EDB5, 0x000305B800033DB7,
-    0x000305BA00030DB9, 0x000305BC00031DBB, 0x000305BE00030DBD,
-    0x8000AF6A0000BDBF, 0x8000AF728000AF6E, 0x8000AE038000AF76,
-    0x8000AF7E8000AF7A, 0x0003F1A88000AF82, 0x0003A9AA0003F9A9,
-    0x0003B9AC0003A9AB, 0x0003A9AE0003B9AD, 0x000241B30003A5AF,
-    0x8000AF8A8000AF86, 0x8000AF928000AF8E, 0x000251B5000259B6,
-    0x000245BB8000AF96, 0x8000AF9A000245BA, 0x000249BF8000AF9E,
-    0x000249BD000241BE, 0x8000AFA2000255BC, 0x8000AFAA8000AFA6,
-    0x00038CEF8000AFAE, 0x8000AFB68000AFB2, 0x00036CE38000AFBA,
-    0x000050E18000AFBE, 0x8000AFC68000AFC2, 0x8000AFCE000578BE,
-    0x0000E58000027084, 0x0000F9820000E981, 0x0004B0868000AFD2,
-    0x8000AFD600055487, 0x8000AFDA00FE40EF, 0x8000AFDE00FE5CE1,
-    0x8000AFE2000154E3, 0x8000AFEA8000AFE6, 0x000199B38000AFEE,
-    0x8000AFF68000AFF2, 0x8000AFFE8000AFFA, 0x00011DB500011DB6,
-    0x00013DBB8000B002, 0x8000B00600013DBA, 0x0000E1BF8000B00A,
-    0x0000F1BD0000F9BE, 0x000621A80000F9BC, 0x000651AA000651A9,
-    0x0006D5AC0006CDAB, 0x0006D5AE0006DDAD, 0x8000AFCA0006C9AF,
-    0x8000B0128000B00E, 0x8000B01A8000B016, 0x8000B0228000B01E,
-    0x00076DB88000B026, 0x00070DBA000705B9, 0x00071DBC000705BB,
-    0x000701BE000701BD, 0x0006B9B0000701BF, 0x00076DB20006B9B1,
-    0x00077DB4000765B3, 0x000765B6000765B5, 0x0006D9A3000755B7,
-    0x8000B02E8000B02A, 0x8000B0368000B032, 0x00065DA500065DA6,
-    0x00067DAB00029C84, 0x000324BE00067DAA, 0x0007A1AF8000B03A,
-    0x0007B1AD0007B9AE, 0x000049800007B9AC, 0x0000598200004981,
-    0x8000B03E000755B3, 0x000769B600077DB5, 0x000040868000B042,
-    0x000755BA00035487, 0x000739BC000729BB, 0x000729BE000739BD,
-    0x000699A3000721BF, 0x8000B04A8000B046, 0x8000B0528000B04E,
-    0x0006B1A50006A5A6, 0x0006E5AB8000B056, 0x8000B05A000699AA,
-    0x0006EDAF8000B05E, 0x0006F5AD0006E5AE, 0x000538E10006F5AC,
-    0x000424E38000B062, 0x8000B06A8000B066, 0x8000B0728000B06E,
-    0x8000B07A8000B076, 0x8000B0828000B07E, 0x8000B08A8000B086,
-    0x0004F4EF8000B08E, 0x0006F9A88000B092, 0x000611AA000609A9,
-    0x000639AC00062DAB, 0x00062DAE000625AD, 0x8000B096000625AF,
-    0x8000B09E8000B09A, 0x000019808000B0A2, 0x0000058200001981,
-    0x0001E5B88000B0A6, 0x0001E5BA0001EDB9, 0x0001E5BC0001FDBB,
-    0x0001E5BE0001EDBD, 0x00065DB0000159BF, 0x000621B2000621B1,
-    0x000621B4000621B3, 0x000611B6000629B5, 0x0002D1A8000611B7,
-    0x00030DAA0002D9A9, 0x00031DAC000305AB, 0x000305AE000305AD,
-    0x000264BE000335AF, 0x8000B0B28000B0AE, 0x8000B0BA8000B0B6,
-    0x8000B0C28000B0BE, 0x000325B88000B0C6, 0x000325BA00032DB9,
-    0x000325BC00033DBB, 0x0003F9BE000329BD, 0x00034DB00003F9BF,
-    0x000325B2000321B1, 0x000325B400033DB3, 0x000325B600032DB5,
-    0x000385B300031DB7, 0x8000B0CA0002C884, 0x8000B0CE00030884,
-    0x000395B5000385B6, 0x0003E5BB8000B0D2, 0x000C0886000389BA,
-    0x0003E9BF00039C87, 0x0003E9BD0003E1BE, 0x000AEC850003F1BC,
-    0x0003CDA38000AF53, 0x0003DDA58000B0D6, 0x8000B0DE8000B0DA,
-    0x8000B0E20003CDA6, 0x0003ADAB8000B0E6, 0x0003A1AD0003C1AA,
-    0x0003A1AF0003B9AC, 0x8000B0EA0003A9AE, 0x8000B0F28000B0EE,
-    0x0003C8EF8000B0F6, 0x8000B0FE8000B0FA, 0x0003B4E38000B102,
-    0x0001D0E18000B106, 0x00000D808000B10A, 0x00007D8200007581,
-    0x8000B1128000B10E, 0x000165B38000B116, 0x000165B5000204BE,
-    0x8000B1228000B11E, 0x000D4086000165B6, 0x0001F5BB000D5887,
-    0x0001A5BD0001FDBA, 0x000199BF0001B9BC, 0x8000B1260001A5BE,
-    0x8000B12E8000B12A, 0x8000B132000C0084, 0x8000B13A8000B136,
-    0x8000B13E0001CCEF, 0x8000B14200062CE1, 0x000000EC000120E3,
-    0x8000B14A8000B146, 0x8000B1528000B14E, 0x8000B1560001E9A3,
-    0x0001E9A68000B15A, 0x8000B1628000B15E, 0x000171AA0001E9A5,
-    0x8000B166000179AB, 0x000129AE8000B16A, 0x000135AC000115AF,
-    0x000E25A8000129AD, 0x000E25AA000E2DA9, 0x000E25AC000E3DAB,
-    0x000E25AE000E2DAD, 0x8000B11A000E95AF, 0x8000B1728000B16E,
-    0x8000B17A8000B176, 0x00009D8100009D82, 0x000F45B800009D80,
-    0x000F45BA000F4DB9, 0x000F49BC000F59BB, 0x000F7DBE000F49BD,
-    0x000EF1B0000F71BF, 0x000EA5B2000EF9B1, 0x000EA9B4000EB9B3,
-    0x000E9DB6000E95B5, 0x000FD5A8000F7DB7, 0x000F09AA000FDDA9,
-    0x000F19AC000F09AB, 0x000F0DAE000F05AD, 0x8000B17E000F05AF,
-    0x8000B1868000B182, 0x8000B18A0003A0BE, 0x000378878000B18E,
-    0x000F15B800001086, 0x000F21BA000F1DB9, 0x0000F5BC000F21BB,
-    0x0000F5BE0000FDBD, 0x000F41B00000EDBF, 0x000F5DB2000F4DB1,
-    0x000F4DB4000F55B3, 0x000F31B6000F35B5, 0x8000B192000F31B7,
-    0x8000B196000CB0EF, 0x8000B19E8000B19A, 0x8000B1A68000B1A2,
-    0x8000B1AE8000B1AA, 0x8000B1B68000B1B2, 0x8000B1BE8000B1BA,
-    0x8000B1C2000DA8E3, 0x000EC1B3000D8CE1, 0x8000B1CA8000B1C6,
-    0x8000B1D28000B1CE, 0x000ED1B5000EC5B6, 0x000EC9BB8000B1D6,
-    0x8000B1DA000EC5BA, 0x000EB1BF8000B1DE, 0x000EC1BD000EC9BE,
-    0x000E01A3000EC9BC, 0x8000B1E68000B1E2, 0x8000B1EE8000B1EA,
-    0x000E11A5000E05A6, 0x000E09AB8000B1F2, 0x8000B1F6000E05AA,
-    0x000E71AF0002AC84, 0x000E01AD000E09AE, 0x00005180000E09AC,
-    0x0000618200005981, 0x0001CCBE0000C5B3, 0x0000CDB60000C5B5,
-    0x000740868000B1FE, 0x0000C9BA00011487, 0x0000D9BC0000C9BB,
-    0x0000C9BE0000D9BD, 0x00030DAB0000C5BF, 0x000309A900030DAA,
-    0x00030DAF0002B9A8, 0x00030DAD00030DAE, 0x000360BE00030DAC,
-    0x8000B2068000B202, 0x8000B20E8000B20A, 0x8000B2168000B212,
-    0x000329BB000530BE, 0x000319B9000319BA, 0x0003DDBF000301B8,
-    0x0003DDBD0003DDBE, 0x00034DB3000331BC, 0x00034DB100034DB2,
-    0x000339B700034DB0, 0x000341B5000331B6, 0x8000B21A00034DB4,
-    0x000399A5000391A6, 0x000399A38000B21E, 0x8000B2268000B222,
-    0x000399AF8000B22A, 0x000385AD000395AE, 0x000395AB000385AC,
-    0x8000B0AA000395AA, 0x8000B2328000B22E, 0x8000B23A8000B236,
-    0x8000B2428000B23E, 0x8000B24A8000B246, 0x8000B2528000B24E,
-    0x8000B25A8000B256, 0x00001D808000B25E, 0x0001FD8200000981,
-    0x000708BF8000B262, 0x8000B26E8000B266, 0x8000B272000024EF,
-    0x8000B27600029484, 0x0002E0878000B27A, 0x001800BE00050486,
-    0x8000B2828000B27E, 0x8000B286000190E1, 0x8000B28A000080E3,
-    0x8000B2928000B28E, 0x8000B296000165B3, 0x00016DB6000165B5,
-    0x8000B29E8000B29A, 0x0001E5BA8000B2A2, 0x0001E9BC0001F9BB,
-    0x0001C5BE0001E9BD, 0x8000B2A60001BDBF, 0x8000B2AE8000B2AA,
-    0x8000B2B20019C4BE, 0x8000B2BA8000B2B6, 0x8000B2BE0001C8EF,
-    0x8000B2C2000EDCE1, 0x8000B2C6000E30E3, 0x8000B2CE8000B2CA,
-    0x0000798000043084, 0x0000158200001581, 0x8000B2D20001E5A3,
-    0x0001EDA60001E5A5, 0x000640868000B2D6, 0x000165AA00079087,
-    0x000169AC000179AB, 0x000145AE000169AD, 0x0005DDA800013DAF,
-    0x000621AA000621A9, 0x000621AC000621AB, 0x000621AE000621AD,
-    0x8000B26A00069DAF, 0x8000B2DE8000B2DA, 0x8000B2E200013084,
-    0x8000B2EA8000B2E6, 0x00068DB88000B2EE, 0x00069DBA000695B9,
-    0x00068DBC000695BB, 0x0006BDBE0006B5BD, 0x0006EDB00006B5BF,
-    0x0006F1B20006F1B1, 0x0006CDB40006F1B3, 0x0006BDB60006B5B5,
-    0x000789A80006B5B7, 0x000791AA000795A9, 0x0007BDAC000791AB,
-    0x0007A1AE0007A5AD, 0x8000B2F20007DDAF, 0x8000B2FA8000B2F6,
-    0x8000B3028000B2FE, 0x8000B30A8000B306, 0x000749B88000B30E,
-    0x00075DBA000755B9, 0x000771BC000755BB, 0x000769BE00077DBD,
-    0x0007A5B0000769BF, 0x0007B9B20007ADB1, 0x000791B40007B1B3,
-    0x000779B6000791B5, 0x8000B312000779B7, 0x8000B31A8000B316,
-    0x0005C8EF8000B31E, 0x8000B3220009C0E1, 0x000424E30019C0E3,
-    0x0006D0E18000B326, 0x00039DA20028DCEF, 0x000519A0000115A3,
-    0x000659B300058DA1, 0x8000B32E8000B32A, 0x8000B3368000B332,
-    0x000675B5000665B6, 0x000629BB8000B33A, 0x8000B33E000621BA,
-    0x000615BF8000B342, 0x000625BD000615BE, 0x000699A300062DBC,
-    0x8000B34600FC998F, 0x8000B3528000B34E, 0x0006B5A50006A5A6,
-    0x0006E9AB8000B356, 0x001F28860006E1AA, 0x0006D5AF00009C87,
-    0x0006E5AD0006D5AE, 0x00096D9E0006EDAC, 0x000B0D9C00077D9F,
-    0x000DF19A0009ED9D, 0x00F0E598000D399B, 0x00F1E196000F0D99,
-    0x00F5C59400F1E997, 0x00F71D9200F30D95, 0x00F9FD9000F5F193,
-    0x00FF1D8200F9ED91, 0x8000B35A00FA0183, 0x00F615868000B35E,
-    0x00FA098400F63987, 0x00F0C18A00F4F185, 0x8000B36200F2258B,
-    0x000CE18E8000B366, 0x00F2358C000E158F, 0x000EAD9200F3018D,
-    0x8000B36A00088193, 0x0004E9968000B36E, 0x000A799400067597,
-    0x0006D19A000AF195, 0x8000B3720000C99B, 0x00031D818000B376,
-    0x0002599C00031D80, 0x001001AB0003F582, 0x00168DA9001695AA,
-    0x002E95AF00168DA8, 0x0012FDAD002C01AE, 0x001EA5A3001299AC,
-    0x00028DA1001EA5A2, 0x001A19A78000B1FA, 0x001801A5001A95A6,
-    0x8000B37A001F9DA4, 0x8000B3828000B37E, 0x8000B38A8000B386,
-    0x8000B3928000B38E, 0x002AE5B38000B396, 0x002FADB1002AE5B2,
-    0x8000B39A002EE5B0, 0x002401B58000B39E, 0x000329A8002A11B4,
-    0x00033DAA000335A9, 0x00032DAC000335AB, 0x0003BDAE0003B5AD,
-    0x8000B3A20003B5AF, 0x8000B3AA8000B3A6, 0x00001D808000B3AE,
-    0x0000B98200000981, 0x00024DB88000B3B2, 0x00026DBA00025DB9,
-    0x000219BC000209BB, 0x000209BE000219BD, 0x0003CDB0000201BF,
-    0x0003DDB20003D5B1, 0x0003CDB40003D5B3, 0x00027DB6000275B5,
-    0x8000B3B6000275B7, 0x8000B3BE001DC884, 0x8000B3C2001F0CBE,
-    0x000648EF8000B3C6, 0x000658E1000758EF, 0x000194E3000698E1,
-    0x001C1086000600E3, 0x001EF4BE001D7C87, 0x8000B3CE8000B3CA,
-    0x0003F5B5000365B6, 0x0003E5B38000B3D2, 0x8000B3DA8000B3D6,
-    0x0002E1BF8000B3DE, 0x000351BD000359BE, 0x000341BB000359BC,
-    0x8000B3E2000359BA, 0x8000B34A8000B3E6, 0x8000B3EA8000B3BA,
-    0x8000B3F28000B3EE, 0x8000B3FA8000B3F6, 0x001DADA88000B3FE,
-    0x001DB5AA001DB5A9, 0x001E3DAC001E25AB, 0x001E15AE001E1DAD,
-    0x001E0DB0001E75AF, 0x001E25B2001E2DB1, 0x001E95B4001E8DB3,
-    0x001E95B6001E9DB5, 0x001EB5B8001E8DB7, 0x001EB5BA001EBDB9,
-    0x001E85BC001E9DBB, 0x001F5DBE001F55BD, 0x001D1DB3001F55BF,
-    0x8000B4068000B402, 0x8000B40E8000B40A, 0x001E95B5001EBDB6,
-    0x001EC9BB8000B412, 0x8000B416001EE9BA, 0x001E79BF8000B41A,
-    0x001E79BD001E71BE, 0x00002982001ED1BC, 0x00001D80001D59A3,
-    0x001EF9A600001581, 0x8000B4228000B41E, 0x001EADAA001ED1A5,
-    0x8000B426001E8DAB, 0x001E35AE0003E084, 0x001E95AC001E3DAF,
-    0x001E89A8001E3DAD, 0x001E9DAA001E95A9, 0x001ED1AC001EB1AB,
-    0x001ECDAE001ED9AD, 0x8000B42A001EC5AF, 0x000088868000B42E,
-    0x8000B43200016C87, 0x8000B43A8000B436, 0x00015DB88000B43E,
-    0x0001C1BA0001C1B9, 0x0001C1BC0001C1BB, 0x0001F1BE0001C9BD,
-    0x001EBDB00001F1BF, 0x001E95B2001E9DB1, 0x00017DB4000165B3,
-    0x00016DB6000165B5, 0x001D2DAA000165B7, 0x8000B442001D35AB,
-    0x001D2DAE8000B446, 0x001D2DAC001C95AF, 0x00018C84001D25AD,
-    0x8000B44A001D91A3, 0x001D11A68000B44E, 0x8000B4568000B452,
-    0x001E55B3001D81A5, 0x8000B45E8000B45A, 0x8000B4668000B462,
-    0x001E19B5001E19B6, 0x001E19BB8000B46A, 0x8000B46E001E39BA,
-    0x0001E1BF8000B472, 0x0001F1BD0001F9BE, 0x8000B476001E01BC,
-    0x8000B47E8000B47A, 0x8000B482001D99A3, 0x001DD5A6001DD5A5,
-    0x8000B48A8000B486, 0x001DF5AA8000B48E, 0x001DCDAC001DD5AB,
-    0x000235AE00023DAD, 0x0000658000022DAF, 0x00001D8200001581,
-    0x8000B49200040084, 0x000370878000B496, 0x8000B49E0004FC86,
-    0x8000B4A68000B4A2, 0x8000B4AE8000B4AA, 0x8000B4B68000B4B2,
-    0x0004C8BE8000B4BA, 0x8000B4C28000B4BE, 0x8000B4CA8000B4C6,
-    0x8000B4D28000B4CE, 0x8000B4D6001FDCEF, 0x8000B4DA000194E1,
-    0x8000B4DE00011CE3, 0x8000B4E68000B4E2, 0x000375BB8000B4EA,
-    0x000648BE000369BA, 0x00031DBF8000B4EE, 0x00031DBD00031DBE,
-    0x0003D5B3000365BC, 0x8000B4F68000B4F2, 0x8000B4FE8000B4FA,
-    0x0003C5B500037DB6, 0x000225A800057084, 0x00023DAA000235A9,
-    0x00022DAC000235AB, 0x000295AE00029DAD, 0x000015820002B5AF,
-    0x0001D9808000B502, 0x000000C400000981, 0x8000B50E8000B506,
-    0x0002A9B88000B512, 0x000149BA0002A9B9, 0x000159BC000149BB,
-    0x000145BE000145BD, 0x0002CDB000017DBF, 0x0002D1B20002D1B1,
-    0x0002B9B40002A9B3, 0x0002A9B60002B9B5, 0x001E28E10002A1B7,
-    0x000174E3001C34E1, 0x8000B516001E18E3, 0x0004B8868000B51A,
-    0x0007388400055487, 0x8000B5228000B51E, 0x0007ACBE8000B526,
-    0x001ECCEF8000B52A, 0x0002F5A3001AC8EF, 0x8000B5328000B52E,
-    0x8000B53A8000B536, 0x0002E5A500025DA6, 0x000255AB8000B53E,
-    0x8000B542000249AA, 0x00023DAF8000B546, 0x00023DAD00023DAE,
-    0x000661A8000245AC, 0x000661AA000661A9, 0x000661AC000661AB,
-    0x000661AE000661AD, 0x8000B50A000661AF, 0x8000B54E8000B54A,
-    0x8000B5568000B552, 0x8000B55E8000B55A, 0x0006F1B88000B562,
-    0x0006F1BA0006F1B9, 0x00069DBC0006F1BB, 0x0006B1BE0006B1BD,
-    0x0006E5B00006B1BF, 0x0006E5B20006EDB1, 0x0006E5B40006FDB3,
-    0x0006D9B60006EDB5, 0x0006E9B30006D5B7, 0x8000B56A8000B566,
-    0x8000B5728000B56E, 0x0006E9B50006E1B6, 0x000629BB8000B576,
-    0x8000B57A000621BA, 0x000629BF8000B57E, 0x000629BD000621BE,
-    0x00005D82000631BC, 0x000045800006ADA3, 0x0006A5A600005D81,
-    0x8000B5868000B582, 0x000665AA0006ADA5, 0x000C008600066DAB,
-    0x000665AE00034087, 0x000675AC00066DAF, 0x8000B58A00066DAD,
-    0x8000B58E00057CEF, 0x8000B5968000B592, 0x8000B59E8000B59A,
-    0x8000B5A68000B5A2, 0x8000B5AE8000B5AA, 0x8000B5B68000B5B2,
-    0x8000B5BA000568E3, 0x0006D1B3000578E1, 0x8000B5C28000B5BE,
-    0x8000B5CA8000B5C6, 0x0006FDB50006FDB6, 0x0006A9BB8000B5CE,
-    0x8000B5D20006A1BA, 0x000699BF8000B5D6, 0x0006A9BD0006A9BE,
-    0x0006B9A80006B1BC, 0x000619AA0006B9A9, 0x000635AC000619AB,
-    0x000635AE00063DAD, 0x8000B5DA000629AF, 0x00001D8100001D82,
-    0x8000B5DE00001D80, 0x8000B5E68000B5E2, 0x0001E9B88000B5EE,
-    0x0001F9BA0001E9B9, 0x0001E9BC0001F9BB, 0x00015DBE0001E9BD,
-    0x000625B0000155BF, 0x000625B200062DB1, 0x00062DB400063DB3,
-    0x000615B600061DB5, 0x000C80860001D9B7, 0x8000B5F20002E487,
-    0x8000B5F600059DA3, 0x0005B1A60005B1A5, 0x8000B5FE8000B5FA,
-    0x0005EDAA8000B602, 0x0005FDAC0005E5AB, 0x0005E5AE0005E5AD,
-    0x00034DB60005D5AF, 0x0003C4848000B606, 0x8000B60A000345B5,
-    0x8000B60E00028DB3, 0x000349BE8000B612, 0x000349BC000349BF,
-    0x000369BA000349BD, 0x8000B616000369BB, 0x8000B61E8000B61A,
-    0x000381A5000389A6, 0x000249A38000B622, 0x8000B62A8000B626,
-    0x00038DAF8000B62E, 0x00038DAD00038DAE, 0x0003ADAB00038DAC,
-    0x8000B49A0003ADAA, 0x8000B6368000B632, 0x003DB4858000B63A,
-    0x8000B6428000B63E, 0x8000B64A8000B646, 0x00003D808000B64E,
-    0x00001D8200000981, 0x0003B0BE8000B652, 0x000338878000B656,
-    0x8000B65E000CDC86, 0x8000B6668000B662, 0x8000B66E8000B66A,
-    0x8000B67200065CEF, 0x000CE8BE8000B676, 0x8000B67A000694E3,
-    0x8000B67E0001DCE1, 0x8000B6868000B682, 0x000151B38000B68A,
-    0x8000B6928000B68E, 0x8000B69A8000B696, 0x000179B5000171B6,
-    0x00013DBB8000B69E, 0x8000B6A200013DBA, 0x0001FDBF8000B6A6,
-    0x000105BD0001F5BE, 0x8000B6AA000105BC, 0x8000B6B28000B6AE,
-    0x8000B6B6000C4084, 0x8000B6BE8000B6BA, 0x8000B6C20007ACEF,
-    0x8000B6C6000624E1, 0x000C9086000740E3, 0x8000B6CE000CFC87,
-    0x000059808000B6D2, 0x0000618200006581, 0x8000B6D60001DDA3,
-    0x0001FDA60001F5A5, 0x8000B6DE8000B6DA, 0x0001B1AA8000B6E2,
-    0x000189AC0001B1AB, 0x000179AE000189AD, 0x8000B5EA000171AF,
-    0x8000B6E68000B65A, 0x8000B6CA8000B6EA, 0x8000B6F28000B6EE,
-    0x000D9DA88000B6F6, 0x000E39AA000E2DA9, 0x000E11AC000E31AB,
-    0x000E7DAE000E11AD, 0x000E19B0000E6DAF, 0x000E31B2000E19B1,
-    0x000ED1B4000E31B3, 0x000ECDB6000ED9B5, 0x000EFDB8000EC5B7,
-    0x000EA9BA000ED9B9, 0x000EBDBC000EA9BB, 0x000EADBE000EA5BD,
-    0x000F81A8000EA5BF, 0x000F81AA000F81A9, 0x000F81AC000F81AB,
-    0x000F85AE000F8DAD, 0x8000B6FA000FB5AF, 0x8000B7028000B6FE,
-    0x8000B70A8000B706, 0x8000B7128000B70E, 0x000F9DB88000B716,
-    0x000FA5BA000FADB9, 0x000F55BC000F4DBB, 0x000F49BE000F5DBD,
-    0x000FD1B0000F49BF, 0x000FD1B2000FD1B1, 0x000FB5B4000FD1B3,
-    0x000FB5B6000FBDB5, 0x000E09B3000FADB7, 0x8000B71E8000B71A,
-    0x8000B7268000B722, 0x000E09B5000E0DB6, 0x000E15BB8000B72A,
-    0x8000B72E000E15BA, 0x000E79BF8000B732, 0x000E05BD000E71BE,
-    0x00001582000E05BC, 0x00006180000E4DA3, 0x000E49A600006181,
-    0x000110BE8000B736, 0x000E51AA000E4DA5, 0x00012484000E51AB,
-    0x000E35AE8000B73E, 0x000E41AC000E3DAF, 0x000E39A8000E41AD,
-    0x000E59AA000E39A9, 0x000E71AC000E51AB, 0x000191AE000E71AD,
-    0x00000086000191AF, 0x8000B74200008487, 0x8000B74A8000B746,
-    0x8000B7528000B74E, 0x000175B88000B756, 0x000175BA00017DB9,
-    0x0000D9BC0000C9BB, 0x0000C9BE0000D9BD, 0x0001F5B00000C1BF,
-    0x0001F5B20001FDB1, 0x000155B400014DB3, 0x000155B600015DB5,
-    0x000F4DBA00014DB7, 0x000F4DB8000F55BB, 0x000F7DBE000F45B9,
-    0x000F45BC000FB5BF, 0x000F01B2000F75BD, 0x000F31B0000F01B3,
-    0x000F01B6000F31B1, 0x000F11B4000F0DB7, 0x000E99AA000F11B5,
-    0x000EE5A8000F45AB, 0x000F41AE000E99A9, 0x000F51AC000F41AF,
-    0x8000B75A000F51AD, 0x8000B7628000B75E, 0x8000B76A8000B766,
-    0x8000B7728000B76E, 0x000D51B38000B776, 0x8000B77E8000B77A,
-    0x8000B7868000B782, 0x000D79B5000D71B6, 0x0002B9BB8000B78A,
-    0x8000B78E0002B1BA, 0x000219BF8000B792, 0x000219BD000211BE,
-    0x8000B7960002A1BC, 0x8000B79A000D15A3, 0x000D35A68000B79E,
-    0x8000B7A68000B7A2, 0x0002F5AA000D3DA5, 0x0003E8840002FDAB,
-    0x000255AE8000B7AE, 0x0002E5AC00025DAF, 0x00026DA800025DAD,
-    0x0002A5AA0002ADA9, 0x0002A5AC0002BDAB, 0x0002A5AE0002ADAD,
-    0x0001ED8000017DAF, 0x0001F1820001F181, 0x8000B7B20005A0BE,
-    0x000568878000B7B6, 0x0001C9B800051C86, 0x0001D9BA0001C9B9,
-    0x0001F9BC0001D1BB, 0x000199BE0001F9BD, 0x000105B0000195BF,
-    0x000105B200010DB1, 0x000105B400011DB3, 0x000105B600010DB5,
-    0x000F64E10001F9B7, 0x000E14E3000F9CE1, 0x8000B7BA000FD4E3,
-    0x8000B7BE000E3CE1, 0x0004ACBE0000E4E3, 0x8000B7C68000B7C2,
-    0x8000B7CA00000CEF, 0x000E60EF8000B7CE, 0x8000B7D2000FB4EF,
-    0x000444848000B7D6, 0x8000B7DA000261B3, 0x000261B6000261B5,
-    0x8000B7E28000B7DE, 0x00018DBA8000B7E6, 0x00019DBC000185BB,
-    0x00018DBE000185BD, 0x0005ADA3000185BF, 0x8000B7EA8000B7AA,
-    0x8000B7F28000B7EE, 0x0005ADA50005ADA6, 0x000649AB8000B7F6,
-    0x8000B7FA000641AA, 0x000649AF8000B7FE, 0x000649AD000641AE,
-    0x8000B802000651AC, 0x8000B80A8000B806, 0x00001D808000B80E,
-    0x0000398200000981, 0x8000B8168000B812, 0x0000C8868000B81A,
-    0x8000B81E00032087, 0x8000B8268000B822, 0x00066DA88000B82A,
-    0x0007BDAA0007B5A9, 0x000709AC00071DAB, 0x000731AE000731AD,
-    0x0003A88400072DAF, 0x8000B8328000B82E, 0x8000B83A8000B836,
-    0x8000B8428000B83E, 0x0000CDB88000B846, 0x0000D5BA0000D5B9,
-    0x0000FDBC0000E5BB, 0x00009DBE000095BD, 0x000755B0000095BF,
-    0x00072DB2000725B1, 0x00072DB400073DB3, 0x00071DB6000715B5,
-    0x8000B84A0000F5B7, 0x8000B84E0006BCE1, 0x8000B8520005F4E3,
-    0x8000B85A8000B856, 0x8000B8628000B85E, 0x8000B86A8000B866,
-    0x8000B8728000B86E, 0x8000B87A8000B876, 0x000675B30004E8EF,
-    0x0000158100002D82, 0x8000B87E00001D80, 0x0006C1B50006F1B6,
-    0x0006A1BB8000B882, 0x8000B8860006D1BA, 0x000691BF00016CBE,
-    0x0006A9BD0006A9BE, 0x0006B5A30006B9BC, 0x000028868000B88E,
-    0x8000B89200014C87, 0x000601A5000631A6, 0x000661AB8000B896,
-    0x8000B89A000611AA, 0x000651AF8000B89E, 0x000669AD000669AE,
-    0x8000B8A2000679AC, 0x8000B8A60001BDB3, 0x000179B68000B8AA,
-    0x8000B8B28000B8AE, 0x000155BA000179B5, 0x8000B8B600015DBB,
-    0x0000F9BE8000B8BA, 0x000145BC0000E5BF, 0x000271A80000FDBD,
-    0x000271AA000271A9, 0x0002B5AC000271AB, 0x0002B5AE0002BDAD,
-    0x003CEC840002ADAF, 0x8000B8C28000B8BE, 0x8000B8CA8000B8C6,
-    0x8000B8D28000B8CE, 0x00035DB88000B8D6, 0x00036DBA000365B9,
-    0x00037DBC000365BB, 0x00036DBE000365BD, 0x0002D5B0000365BF,
-    0x0002D5B20002DDB1, 0x000379B400036DB3, 0x00036DB6000365B5,
-    0x8000B73A000365B7, 0x8000B8DE8000B8DA, 0x8000B8E20002F5A3,
-    0x000231A6000231A5, 0x8000B8EA8000B8E6, 0x00021DAA8000B8EE,
-    0x00020DAC000215AB, 0x0003B1AE0003B5AD, 0x000065800003ADAF,
-    0x0000198200000981, 0x8000B8F68000B8F2, 0x003C10BE8000B8FE,
-    0x003CC0868000B902, 0x8000B90600032087, 0x8000B90E8000B90A,
-    0x8000B9168000B912, 0x000285A88000B91A, 0x000295AA000295A9,
-    0x0002BDAC0002A5AB, 0x0002D1AE0002D5AD, 0x8000B91E0002D1AF,
-    0x8000B9268000B922, 0x8000B92E8000B92A, 0x8000B9368000B932,
-    0x000175B88000B93A, 0x000175BA00017DB9, 0x0001D9BC0001C9BB,
-    0x0001C5BE0001C5BD, 0x0002B5B00001FDBF, 0x000281B20002BDB1,
-    0x000155B4000281B3, 0x000155B600015DB5, 0x0006A4E100014DB7,
-    0x000650E38000B93E, 0x003C1C84003C80BE, 0x8000B942003F80BE,
-    0x8000B94A8000B946, 0x8000B9528000B94E, 0x8000B95A8000B956,
-    0x0006E0EF8000B95E, 0x00007D818000B962, 0x8000B96600007180,
-    0x8000B96E00000582, 0x00004CEF8000B972, 0x000190E18000B976,
-    0x000158E38000B97A, 0x8000B9828000B97E, 0x003F58868000B986,
-    0x003D9DB3003CFC87, 0x8000B96A8000B8FA, 0x8000B98E8000B98A,
-    0x003DB5B5003DD5B6, 0x003DF9BB8000B992, 0x8000B996003DF1BA,
-    0x003E19BF8000B99A, 0x003DD5BD003E11BE, 0x8000B99E003DD5BC,
-    0x8000B9A2003DD9A3, 0x003D91A68000B9A6, 0x8000B9AE8000B9AA,
-    0x003DB5AA003DF1A5, 0x8000B9B2003DBDAB, 0x003E55AE8000B9B6,
-    0x003D91AC003E5DAF, 0x003E55A8003D91AD, 0x003E61AA003E59A9,
-    0x003E61AC003E61AB, 0x003E61AE003E61AD, 0x0003A884003E61AF,
-    0x8000B9BE8000B9BA, 0x8000B9C68000B9C2, 0x8000B9CE8000B9CA,
-    0x003F45B88000B9D2, 0x003F55BA003F5DB9, 0x003F75BC003F6DBB,
-    0x003F75BE003F7DBD, 0x003FC1B0003F6DBF, 0x003FC1B2003FC1B1,
-    0x003FC1B4003FC1B3, 0x003FC1B6003FC1B5, 0x0001B980003FC1B7,
-    0x000005820001B981, 0x003E80E18000B9D6, 0x003D28E38000B9DE,
-    0x000080868000B9E2, 0x003D08EF00010487, 0x8000B9EA8000B9E6,
-    0x8000B9F28000B9EE, 0x003FA9B38000B9F6, 0x8000B9FA8000B9DA,
-    0x8000BA028000B9FE, 0x003FA9B5003FA1B6, 0x003E45BB8000BA06,
-    0x8000BA0A003E45BA, 0x003E45BF8000BA0E, 0x003E55BD003E45BE,
-    0x003F69A3003E55BC, 0x8000BA168000BA12, 0x8000BA1E8000BA1A,
-    0x003F69A5003F61A6, 0x003E85AB8000BA22, 0x8000BA26003E85AA,
-    0x003E85AF8000BA2A, 0x003E95AD003E85AE, 0x8000BA2E003E95AC,
-    0x8000BA32003E19B3, 0x003E21B68000BA36, 0x8000BA3E8000BA3A,
-    0x0001F1BA003E31B5, 0x8000BA420001F9BB, 0x0001B1BE8000BA46,
-    0x0001D1BC0001BDBF, 0x003DD1A80001D1BD, 0x003DD5AA003DD1A9,
-    0x003DFDAC003DE9AB, 0x003DEDAE003DE5AD, 0x0003F980000211AF,
-    0x0003C5820003CD81, 0x000400BE00032484, 0x0003AC878000BA4E,
-    0x000219B80004BC86, 0x000225BA00022DB9, 0x0002F9BC0002E9BB,
-    0x0002E9BE0002F9BD, 0x000271B00002E9BF, 0x000241B2000279B1,
-    0x000231B4000241B3, 0x000235B600023DB5, 0x003D6DA5000229B7,
-    0x8000BA568000BA52, 0x8000BA5A003D7DA6, 0x003D45A38000B88A,
-    0x00028DAD8000BA5E, 0x0002E1AF00028DAC, 0x000000AC0002EDAE,
-    0x0002A5AB8000BA62, 0x0001F8E10002ADAA, 0x0002F4E38000BA66,
-    0x8000BA6A0004C084, 0x8000BA728000BA6E, 0x8000BA7A8000BA76,
-    0x8000BA828000BA7E, 0x8000BA8A8000BA86, 0x000230EF8000BA8E,
-    0x000215A88000BA92, 0x000225AA000219A9, 0x000225AC00023DAB,
-    0x000225AE00022DAD, 0x8000BA96000255AF, 0x8000BA9E8000BA9A,
-    0x8000BAA68000BAA2, 0x0004B0848000BAAA, 0x0002D1B88000BAAE,
-    0x0002E1BA0002D9B9, 0x000291BC0002E1BB, 0x000295BE00029DBD,
-    0x00022DB0000289BF, 0x000235B2000235B1, 0x00021DB4000205B3,
-    0x0002F1B60002F1B5, 0x003F78E10002F1B7, 0x003E08E3000134E3,
-    0x00000981003E6CE1, 0x8000BAB20000A980, 0x8000BAB600003D82,
-    0x8000BAC28000BABE, 0x8000BAC6000420BE, 0x003ECCEF003ED8EF,
-    0x8000BACE8000BACA, 0x0004E8860002E9B3, 0x0002E9B60004C087,
-    0x8000BAD68000BAD2, 0x0002ADBA0002E9B5, 0x8000BADA0002B5BB,
-    0x00029DBE8000BADE, 0x0002A5BC000271BF, 0x8000BA4A000295BD,
-    0x8000BAE28000BABA, 0x8000BAEA8000BAE6, 0x8000BAF28000BAEE,
-    0x000615A88000BAF6, 0x0006A1AA0006A1A9, 0x0006ADAC0006BDAB,
-    0x0006FDAE000681AD, 0x000695B00006EDAF, 0x000695B200069DB1,
-    0x0006B5B40006ADB3, 0x0006B5B60006BDB5, 0x000695B80006ADB7,
-    0x000749BA000699B9, 0x000759BC000749BB, 0x000749BE000759BD,
-    0x0005DDAC000749BF, 0x0005E5AE0005EDAD, 0x8000BAFA000509AF,
-    0x0005D5AA8000BAFE, 0x8000BB020005CDAB, 0x000591A6000591A5,
-    0x8000BB0A8000BB06, 0x8000BB128000BB0E, 0x000649B3000591A3,
-    0x8000BB1A8000BB16, 0x8000BB228000BB1E, 0x000645B5000661B6,
-    0x000639BB8000BB26, 0x000064BE000631BA, 0x000619BF8000BB2A,
-    0x000619BD000611BE, 0x000789A3000621BC, 0x0001D9810001D982,
-    0x8000BB2E0001C180, 0x000785A50007A1A6, 0x0007F9AB8000BB32,
-    0x000108860007F1AA, 0x0007D9AF00017C87, 0x0007D9AD0007D1AE,
-    0x8000BB360007E1AC, 0x8000BB3A0006F5B3, 0x000685B68000BB3E,
-    0x8000BB468000BB42, 0x00068DBA00069DB5, 0x8000BB4A00016DBB,
-    0x000165BE8000BB4E, 0x000175BC00016DBF, 0x000625A800016DBD,
-    0x000639AA00062DA9, 0x000651AC000631AB, 0x000641AE000641AD,
-    0x8000BB52000675AF, 0x8000BB5A8000BB56, 0x8000BB628000BB5E,
-    0x8000BB6A8000BB66, 0x000155B88000BB6E, 0x000165BA000165B9,
-    0x000165BC00017DBB, 0x00011DBE00016DBD, 0x00060DB0000115BF,
-    0x0001E5B20001EDB1, 0x0001E5B40001FDB3, 0x0001E5B60001EDB5,
-    0x0005B1A300016DB7, 0x8000BB768000BB72, 0x000C58BE000348BE,
-    0x0005D9A50005C1A6, 0x000229AB8000BB7A, 0x8000BB7E0005C9AA,
-    0x000229AF8000BB82, 0x000229AD000221AE, 0x8000BB86000231AC,
-    0x8000BB8E8000BB8A, 0x000019808000BB92, 0x0000058200001981,
-    0x0003AC848000BB96, 0x000318878000BB9E, 0x8000BBA2000CCC86,
-    0x8000BBAA8000BBA6, 0x000391A88000BBAE, 0x0003C9AA000399A9,
-    0x0003DDAC0003C5AB, 0x0003C1AE0003C1AD, 0x8000BBB20003F5AF,
-    0x8000BBBA8000BBB6, 0x8000BBC28000BBBE, 0x8000BBCA8000BBC6,
-    0x00037DB88000BBCE, 0x0000C1BA0000C1B9, 0x0000F9BC0000D1BB,
-    0x000099BE0000F9BD, 0x00038DB0000099BF, 0x00034DB2000345B1,
-    0x00035DB4000345B3, 0x00034DB6000345B5, 0x000241B3000345B7,
-    0x8000BBD68000BBD2, 0x8000BBDA000F04BF, 0x000255B5000241B6,
-    0x000281BB8000BBDE, 0x8000BBE2000249BA, 0x000281BF8000BBE6,
-    0x000291BD000299BE, 0x8000BBEA000299BC, 0x8000BBF28000BBEE,
-    0x0003A4EF8000BBF6, 0x8000BBFE8000BBFA, 0x000310E38000BC02,
-    0x000058E18000BC06, 0x00002980000D2084, 0x00001D8200002981,
-    0x000650E18000BC0E, 0x0006A0E30007A0E1, 0x000C9485000750E3,
-    0x001B48EF8000BC12, 0x000224E18000BC16, 0x001A70E38000BC1A,
-    0x8000BC228000BC1E, 0x0001A8EF8000BC26, 0x000FA0860006E8EF,
-    0x000289A3000D0C87, 0x00029DA58000BC2A, 0x8000BC328000BC2E,
-    0x8000BC36000289A6, 0x000249AB8000BC3A, 0x000259AD000281AA,
-    0x000249AF000251AC, 0x000E65A8000251AE, 0x000E7DAA000E75A9,
-    0x000E6DAC000E75AB, 0x000EBDAE000EB5AD, 0x8000BC0A000EB5AF,
-    0x8000BC428000BC3E, 0x8000BC4A8000BC46, 0x8000BC598000BC55,
-    0x000F5DB88000BC5D, 0x000F65BA000F6DB9, 0x000F15BC000F0DBB,
-    0x000F15BE000F1DBD, 0x000ECDB0000F09BF, 0x000EDDB2000ED5B1,
-    0x000ECDB4000ED5B3, 0x000F71B6000F71B5, 0x000EF5B3000F6DB7,
-    0x8000BC658000BC61, 0x8000BC6D8000BC69, 0x000EE5B5000E35B6,
-    0x000E11BB8000BC71, 0x8000BC75000E09BA, 0x000FB5BF8000BC79,
-    0x000E01BD000E09BE, 0x00001582000E09BC, 0x00006180000EB1A3,
-    0x000E71A600006181, 0x000110BE8000BC7D, 0x000E4DAA000EA1A5,
-    0x8000BC85000E55AB, 0x000E4DAE00012084, 0x000E4DAC000FF1AF,
-    0x0008459F000E45AD, 0x00090D9D00090D9E, 0x0035759B000B999C,
-    0x003799990035119A, 0x0031259700310D98, 0x003D819500312596,
-    0x003F8993003D0D94, 0x00393D9100381592, 0x0025658F00393D90,
-    0x00040086000DACEF, 0x8000BC890001A087, 0x8000BC918000BC8D,
-    0x8000BC950001E8EF, 0x8000BC99000234E1, 0x8000BC9D000134E3,
-    0x8000BCA1000C08E3, 0x0001A1A1000D08E1, 0x000509A38000BCA5,
-    0x0004CDA500036DA2, 0x00191DA7000511A4, 0x001DA1A900191DA6,
-    0x001CE9AB001939A8, 0x001191AD001DA9AA, 0x001401AF001001AC,
-    0x0015F9B1001151AE, 0x0069E9B30015E5B0, 0x006C01B5006801B2,
-    0x8000BCA9006979B4, 0x8000BCB18000BCAD, 0x8000BCB98000BCB5,
-    0x000309A38000BCBD, 0x000DFDA1000D19A2, 0x00262582000DFDA0,
-    0x8000BCC1003A0583, 0x003CA9868000BCC5, 0x003A1D84003E3587,
-    0x003E898A003B3D85, 0x8000BCC90032258B, 0x0034B18E8000BCCD,
-    0x00323D8C0036318F, 0x0036259200322D8D, 0x000344840008D193,
-    0x000B15960004C0BE, 0x000A5D94000E5597, 0x000E659A000A5595,
-    0x8000BCD1000E899B, 0x8000BCD98000BCD5, 0x0000819C8000BCDD,
-    0x0002B5B88000BCE1, 0x0002B5BA0002BDB9, 0x000219BC000209BB,
-    0x000209BE000219BD, 0x000D75B0000201BF, 0x000D49B2000D7DB1,
-    0x000295B4000D49B3, 0x000295B600029DB5, 0x000DBDA800028DB7,
-    0x000D55AA000D51A9, 0x000D7DAC000D69AB, 0x000D6DAE000D65AD,
-    0x8000BC81000D11AF, 0x00001D810001ED82, 0x8000BCE500001D80,
-    0x8000BB9A8000BCE9, 0x000C9DB3000570BE, 0x8000BCED00058C84,
-    0x8000BCF98000BCF5, 0x000CB5B5000CA5B6, 0x000CF9BB8000BCFD,
-    0x00052886000CF1BA, 0x000319BF00052087, 0x000CDDBD000311BE,
-    0x8000BD01000CDDBC, 0x8000BD05000CD9A3, 0x000CE1A68000BD09,
-    0x8000BD118000BD0D, 0x000CB5AA000CF1A5, 0x8000BD15000CBDAB,
-    0x000355AE8000BD19, 0x000C99AC00035DAF, 0x8000BD1D000C99AD,
-    0x8000BD258000BD21, 0x8000BD2D8000BD29, 0x8000BD358000BD31,
-    0x8000BD390001BCEF, 0x8000BD3D000E7CE1, 0x8000BD41000150E3,
-    0x8000BD498000BD45, 0x000295B38000BD4D, 0x8000BD558000BD51,
-    0x8000BD5D8000BD59, 0x0002B9B50002BDB6, 0x000239BB8000BD65,
-    0x0004C886000261BA, 0x000219BF0004AC87, 0x000219BD000211BE,
-    0x000555A3000221BC, 0x0007C5810007D582, 0x8000BD690007C580,
-    0x000579A500057DA6, 0x0005F9AB8000BD6D, 0x8000BD710005A1AA,
-    0x0005D9AF8000BD75, 0x0005D9AD0005D1AE, 0x8000BD790005E1AC,
-    0x8000BD7D000759B3, 0x00071DB68000BD81, 0x8000BD898000BD85,
-    0x000709BA000715B5, 0x8000BD8D000709BB, 0x0007F9BE8000BD91,
-    0x0007F9BC0007F9BF, 0x8000BCF10007F9BD, 0x8000BD958000BD61,
-    0x8000BD9D8000BD99, 0x8000BDA58000BDA1, 0x0007ADA88000BDA9,
-    0x0007B5AA0007B5A9, 0x0007F9AC0007EDAB, 0x0007EDAE0007F9AD,
-    0x0007A9B00007E5AF, 0x000749B20007A9B1, 0x000759B4000749B3,
-    0x000749B6000759B5, 0x000779B8000749B7, 0x000741BA000745B9,
-    0x000745BC00075DBB, 0x000745BE00074DBD, 0x00061DA3000739BF,
-    0x8000BDB18000BDAD, 0x8000BDB98000BDB5, 0x000651A5000659A6,
-    0x00064DAB8000BDBD, 0x8000BDC100064DAA, 0x0006BDAF8000BDC5,
-    0x0006BDAD0006BDAE, 0x00006D800006BDAC, 0x0000198200000981,
-    0x8000BDCD8000BDC9, 0x000190BE00019884, 0x001C00868000BDD1,
-    0x8000BDD50001C487, 0x8000BDDD8000BDD9, 0x8000BDE58000BDE1,
-    0x00065DA88000BDE9, 0x000195AA000195A9, 0x0001BDAC0001A5AB,
-    0x0001D1AE0001D5AD, 0x8000BDED0001D1AF, 0x8000BDF58000BDF1,
-    0x8000BDFD8000BDF9, 0x8000BE058000BE01, 0x000159B88000BE09,
-    0x0000CDBA000159B9, 0x0000DDBC0000C5BB, 0x0000C5BE0000C5BD,
-    0x0001B5B00000F5BF, 0x000181B20001BDB1, 0x000179B4000181B3,
-    0x000169B6000179B5, 0x00021DB3000169B7, 0x8000BE118000BE0D,
-    0x8000BE15001C80BE, 0x000235B5000255B6, 0x000279BB8000BE19,
-    0x8000BE1D000271BA, 0x0002BDBF8000BE21, 0x000255BD0002BDBE,
-    0x8000BE25000255BC, 0x8000BE29000259A3, 0x000211A68000BE2D,
-    0x8000BE358000BE31, 0x000235AA000271A5, 0x8000BE3900023DAB,
-    0x0002F9AE8000BE3D, 0x000211AC0002F9AF, 0x8000BE45000211AD,
-    0x001D04BE8000BE49, 0x8000BE4D001E00BE, 0x8000BE558000BE51,
-    0x0000F9808000BE59, 0x0000C5820000CD81, 0x001CB08600032084,
-    0x8000BE5D00039487, 0x8000BE658000BE61, 0x8000BE6D8000BE69,
-    0x00026CE38000BE71, 0x0001A0E18000BE75, 0x000214EF8000BE79,
-    0x8000BE818000BE7D, 0x8000BE898000BE85, 0x8000BE918000BE8D,
-    0x0006F0E18000BE95, 0x00004CE3000634E1, 0x8000BE99000618E3,
-    0x8000BEA18000BE9D, 0x00003D808000BEA5, 0x0000198200000981,
-    0x8000BEAD8000BEA9, 0x000098EF001DBC84, 0x000231B3000738EF,
-    0x001DD887000000D1, 0x8000BEB5001C6C86, 0x000221B5000229B6,
-    0x0002CDBB8000BEB9, 0x8000BEBD0002CDBA, 0x0002CDBF8000BEC1,
-    0x0002CDBD0002CDBE, 0x00065DB20002CDBC, 0x00060DB0000665B3,
-    0x00067DB6000655B1, 0x00067DB400059DB7, 0x00058DBA000675B5,
-    0x0005A5B80005CDBB, 0x0005C5BE000585B9, 0x0005D5BC0005C9BF,
-    0x8000BEC50005CDBD, 0x8000BECD8000BEC9, 0x8000BED58000BED1,
-    0x8000BEDD8000BED9, 0x0006B5AA8000BEE1, 0x000701A80006BDAB,
-    0x000681AE0006BDA9, 0x000699AC00068DAF, 0x001D71A3000695AD,
-    0x8000BEE98000BEE5, 0x8000BEF18000BEED, 0x001D61A5001D69A6,
-    0x001D8DAB8000BEF5, 0x8000BEF9001D8DAA, 0x001D8DAF8000BEFD,
-    0x001D8DAD001D8DAE, 0x8000BF01001D8DAC, 0x8000BF05001ED5B3,
-    0x001ECDB68000BF09, 0x8000BF118000BF0D, 0x001EA1BA001EC5B5,
-    0x8000BF15001EA1BB, 0x001EA5BE8000BF19, 0x001EB1BC001EA9BF,
-    0x8000BE41001EB1BD, 0x8000BF1D8000BEB1, 0x0000F98000030484,
-    0x000011820000F981, 0x001E21A88000BF21, 0x001E39AA001E21A9,
-    0x001E29AC001E39AB, 0x001E01AE001E29AD, 0x001E01B0001E01AF,
-    0x001E01B2001E01B1, 0x001E05B4001E01B3, 0x001E3DB6001E09B5,
-    0x001E0DB8001E35B7, 0x001E1DBA001E15B9, 0x001E0DBC001E15BB,
-    0x001FEDBE001FE5BD, 0x001F91A3001FE5BF, 0x000128868000BF25,
-    0x8000BF2900014887, 0x001F81A5001F89A6, 0x001FE5AB8000BF2D,
-    0x8000BF31001FE5AA, 0x001FEDAF8000BF35, 0x001FF5AD001FE1AE,
-    0x8000BF39001FF5AC, 0x8000BF3D001E2DB3, 0x001E95B68000BF41,
-    0x8000BF498000BF45, 0x001EB1BA001E9DB5, 0x8000BF4D001EB1BB,
-    0x000175BE8000BF51, 0x001EA1BC00017DBF, 0x001ED1A8001EA1BD,
-    0x001EE1AA001ED9A9, 0x001E51AC001EE1AB, 0x001E51AE001E51AD,
-    0x8000BF55001E51AF, 0x8000BF5D8000BF59, 0x8000BF658000BF61,
-    0x8000BF6D8000BF69, 0x0001DDB88000BF71, 0x0001E5BA0001EDB9,
-    0x000199BC00018DBB, 0x00018DBE000199BD, 0x001E31B0000185BF,
-    0x001E31B2001E31B1, 0x0001F5B4001E31B3, 0x0001F5B60001FDB5,
-    0x001D69A30001EDB7, 0x8000BF798000BF75, 0x8000BF818000BF7D,
-    0x001DD9A5001DD1A6, 0x001DF5AB8000BF85, 0x8000BF89001DF5AA,
-    0x000239AF8000BF8D, 0x001DE5AD000231AE, 0x00006981001DE5AC,
-    0x000400BE00005980, 0x8000BF9500006182, 0x8000BF9D8000BF99,
-    0x0003EC848000BFA1, 0x000388878000BFA5, 0x8000BFA90004EC86,
-    0x8000BFB18000BFAD, 0x000385A88000BFB5, 0x000395AA000395A9,
-    0x0003BDAC0003A5AB, 0x0003D1AE0003D5AD, 0x8000BFB90003D1AF,
-    0x8000BFC18000BFBD, 0x8000BFC98000BFC5, 0x8000BFD18000BFCD,
-    0x000371B88000BFD5, 0x000371BA000371B9, 0x0000D5BC000371BB,
-    0x0000D5BE0000DDBD, 0x0003B5B00000CDBF, 0x000381B20003BDB1,
-    0x000351B4000381B3, 0x000351B6000351B5, 0x001E54E1000351B7,
-    0x000194E3001FACE1, 0x000360E3001E28E3, 0x000008E18000BFD9,
-    0x000390EF8000BFDD, 0x8000BFE58000BFE1, 0x8000BFED8000BFE9,
-    0x001F4CEF8000BFF1, 0x00025DB3001F4CEF, 0x8000BFF98000BFF5,
-    0x8000C0058000BFFD, 0x000275B5000215B6, 0x000239BB8000C009,
-    0x00052484000231BA, 0x0002D5BF0004E0BE, 0x000215BD0002D5BE,
-    0x001D91B8000215BC, 0x001DA1BA001D99B9, 0x001DD1BC001DA1BB,
-    0x001DD5BE001DDDBD, 0x001E09B0001DC9BF, 0x001E19B2001E09B1,
-    0x001E09B4001E19B3, 0x001DBDB6001E09B5, 0x001EA9A8001DB5B7,
-    0x001EB9AA001EA9A9, 0x001EA9AC001EB9AB, 0x001E79AE001EA9AD,
-    0x0000A580001E79AF, 0x0000A5820000AD81, 0x0004D0868000C00D,
-    0x8000C0110004F887, 0x8000BF918000C015, 0x8000C0198000C001,
-    0x8000C0218000C01D, 0x8000C0298000C025, 0x000671A88000C02D,
-    0x000671AA000671A9, 0x000655AC000671AB, 0x00064DAE000645AD,
-    0x00063DB0000645AF, 0x0006EDB20006E5B1, 0x0006FDB40006E5B3,
-    0x0006EDB60006E5B5, 0x0006DDB80006E5B7, 0x000775BA000771B9,
-    0x000759BC000749BB, 0x000749BE000759BD, 0x0006D9B3000749BF,
-    0x8000C0358000C031, 0x8000C03D8000C039, 0x0006FDB50006E5B6,
-    0x000601BB8000C041, 0x8000C0450006D9BA, 0x000601BF8000C049,
-    0x000611BD000619BE, 0x8000C04D000619BC, 0x8000C0510005D9A3,
-    0x000591A68000C055, 0x8000C05D8000C059, 0x0005B5AA0005F1A5,
-    0x8000C0610005BDAB, 0x000551AE8000C065, 0x000591AC000551AF,
-    0x000759A3000591AD, 0x0000198100001982, 0x8000C0690001E180,
-    0x00077DA5000765A6, 0x000781AB8000C06D, 0x0002A084000759AA,
-    0x000781AF0001ACBE, 0x000791AD000799AE, 0x8000C071000799AC,
-    0x0000E8860006A9B3, 0x000159B600012C87, 0x8000C0798000C075,
-    0x000175BA0006A1B5, 0x8000C07D00017DBB, 0x0001F9BE8000C081,
-    0x000165BC0001F9BF, 0x0006D1A80001F9BD, 0x000665AA0006D9A9,
-    0x00066DAC000675AB, 0x000195AE00019DAD, 0x0001EC8400018DAF,
-    0x8000C0898000C085, 0x8000C0918000C08D, 0x8000C0998000C095,
-    0x000169B88000C09D, 0x000109BA000169B9, 0x00011DBC000105BB,
-    0x000101BE000101BD, 0x0001F5B0000135BF, 0x0001F5B20001FDB1,
-    0x000179B4000169B3, 0x000169B6000179B5, 0x8000C0A1000161B7,
-    0x8000C0A98000C0A5, 0x8000C0AD0005E1A3, 0x000211A60005E9A5,
-    0x8000C0B58000C0B1, 0x00023DAA8000C0B9, 0x00022DAC000235AB,
-    0x0002B1AE0002B1AD, 0x8000C0BD0002B1AF, 0x000304BE8000C0C1,
-    0x8000C0C5000C0084, 0x8000C0CD8000C0C9, 0x00000D808000C0D1,
-    0x00001D8200001581, 0x8000C0D98000C0D5, 0x000344878000C0DD,
-    0x0003E9B3000CDC86, 0x8000C0E98000C0E5, 0x8000C0F18000C0ED,
-    0x00033DB500033DB6, 0x000319BB8000C0F5, 0x8000C0F9000311BA,
-    0x0000B9BF8000C0FD, 0x0000B9BD0000B1BE, 0x8000C101000301BC,
-    0x8000C105000194E1, 0x8000C10900017CE3, 0x8000C1118000C10D,
-    0x8000C1198000C115, 0x8000C1218000C11D, 0x8000C1298000C125,
-    0x8000C1318000C12D, 0x000255A800024CEF, 0x000261AA00025DA9,
-    0x0002B5AC000261AB, 0x0002B5AE0002BDAD, 0x000D68BE0002ADAF,
-    0x8000C1398000C135, 0x8000C1418000C13D, 0x0000AD810000AD82,
-    0x000161B80000AD80, 0x000109BA000161B9, 0x000101BC000109BB,
-    0x000101BE000101BD, 0x0002D5B0000101BF, 0x0002D5B20002DDB1,
-    0x000175B400016DB3, 0x000169B600017DB5, 0x000668E1000161B7,
-    0x0000B0E30007F0E1, 0x8000C1450006F4E3, 0x8000C14D8000C149,
-    0x8000C1598000C155, 0x8000C1618000C15D, 0x000CFCBE8000C165,
-    0x00006CEF8000C169, 0x0002A9A3000628EF, 0x000C68868000C16D,
-    0x8000C171000D0487, 0x00027DA500027DA6, 0x000259AB8000C175,
-    0x8000C179000251AA, 0x0001F9AF8000C17D, 0x0001F9AD0001F1AE,
-    0x000E85A8000241AC, 0x000E85AA000E8DA9, 0x000E8DAC000E9DAB,
-    0x000EB5AE000EBDAD, 0x8000C151000EDDAF, 0x8000C1858000C181,
-    0x8000C18D8000C189, 0x8000C1958000C191, 0x000EADB88000C199,
-    0x000EBDBA000EB5B9, 0x000F75BC000EB5BB, 0x000F75BE000F7DBD,
-    0x000EA9B0000F6DBF, 0x000EBDB2000EB5B1, 0x000EADB4000EB5B3,
-    0x000E9DB6000E95B5, 0x000E0DB3000E95B7, 0x8000C1A18000C19D,
-    0x8000C1A98000C1A5, 0x000E05B5000E3DB6, 0x000E71BB8000C1AD,
-    0x8000C1B1000E6DBA, 0x000E51BF8000C1B5, 0x000E61BD000E69BE,
-    0x00001982000E69BC, 0x00006580000E49A3, 0x000E79A600001981,
-    0x8000C1BD8000C1B9, 0x000E29AA000E41A5, 0x0003BC84000E35AB,
-    0x000E2DAE8000C1C1, 0x000E2DAC000E15AF, 0x000E9DA8000E25AD,
-    0x000EADAA000EA5A9, 0x000EBDAC000EA5AB, 0x000EDDAE000EA1AD,
-    0x00012086000ECDAF, 0x8000C1C500016487, 0x8000C1CD8000C1C9,
-    0x8000C1D58000C1D1, 0x000179B88000C1D9, 0x0001CDBA000179B9,
-    0x0001DDBC0001C5BB, 0x0001C5BE0001C5BD, 0x000EBDB00001F5BF,
-    0x000E85B2000E8DB1, 0x000159B4000149B3, 0x000149B6000159B5,
-    0x000E2DB5000149B7, 0x8000C1E18000C1DD, 0x8000C1E5000E39B6,
-    0x000E3DB38000C1E9, 0x000185BD8000C1ED, 0x000185BF000E49BC,
-    0x8000C1F1000185BE, 0x000E51BB8000C0E1, 0x000E65A3000E61BA,
-    0x8000C1F98000C1F5, 0x8000C2018000C1FD, 0x000E75A5000E61A6,
-    0x000E09AB8000C205, 0x8000C209000E39AA, 0x0001DDAF8000C20D,
-    0x0001DDAD0001DDAE, 0x8000C211000E11AC, 0x000FD0EF8000C215,
-    0x8000C21D8000C219, 0x8000C2258000C221, 0x8000C22D8000C229,
-    0x8000C235000368BE, 0x000E54E18000C239, 0x000E64E38000C23D,
-    0x000059808000C241, 0x0000698200005981, 0x0004F08600038C84,
-    0x8000C24500031487, 0x8000C24D8000C249, 0x8000C2558000C251,
-    0x8000C25D8000C259, 0x8000C2658000C261, 0x8000C26D8000C269,
-    0x8000C2758000C271, 0x8000C27D8000C279, 0x8000C2858000C281,
-    0x000389A88000C289, 0x000399AA000389A9, 0x000389AC000399AB,
-    0x00033DAE000389AD, 0x000351B0000335AF, 0x000355B2000351B1,
-    0x000315B400037DB3, 0x000315B600031DB5, 0x00033DB800030DB7,
-    0x000305BA00030DB9, 0x0000F5BC0000EDBB, 0x0000E9BE0000F9BD,
-    0x8000C28D0000E9BF, 0x8000C2958000C291, 0x8000C29900023DB3,
-    0x000225B600022DB5, 0x000568BE8000C29D, 0x0002B9BA8000C2A5,
-    0x0002ADBC0002B9BB, 0x0002FDBE000285BD, 0x000049800002F5BF,
-    0x0000558200005581, 0x00038CEF00050084, 0x00057487000418BE,
-    0x00033CE30004FC86, 0x000050E18000C2A9, 0x8000C2B18000C2AD,
-    0x8000C2B98000C2B5, 0x8000C2C18000C2BD, 0x8000C2C98000C2C5,
-    0x8000C2D18000C2CD, 0x8000C2D98000C2D5, 0x000628BE000ED4EF,
-    0x8000C2DD000E74E1, 0x8000C2E1000154E3, 0x8000C2E98000C2E5,
-    0x0002F1A38000C2ED, 0x8000C2F58000C2F1, 0x8000C2FD8000C2F9,
-    0x0002E1A50002E9A6, 0x000275AB8000C301, 0x8000C305000275AA,
-    0x000239AF8000C309, 0x000249AD000231AE, 0x000629A8000261AC,
-    0x00063DAA000629A9, 0x000649AC000631AB, 0x000679AE000649AD,
-    0x8000C2A1000679AF, 0x0007C58100001582, 0x8000C30D0007C580,
-    0x8000C31100036884, 0x0006C9B88000C315, 0x0006D9BA0006C9B9,
-    0x0006C9BC0006D9BB, 0x000759BE0006C9BD, 0x000609B0000759BF,
-    0x000619B2000609B1, 0x000609B4000619B3, 0x0006F9B6000609B5,
-    0x0006B5B30006F9B7, 0x0000AC868000C319, 0x8000C31D00034087,
-    0x0006C1B50006C9B6, 0x0006CDBB8000C321, 0x8000C3250006CDBA,
-    0x0006CDBF8000C329, 0x0006CDBD0006CDBE, 0x8000C32D0006CDBC,
-    0x8000C3310006F1A3, 0x00068DA68000C335, 0x8000C33D8000C339,
-    0x000689AA000685A5, 0x8000C341000689AB, 0x000689AE8000C345,
-    0x000689AC000689AF, 0x000625A8000689AD, 0x000665AA000661A9,
-    0x000665AC00067DAB, 0x000665AE00066DAD, 0x8000C34900069DAF,
-    0x8000C3518000C34D, 0x8000C3598000C355, 0x8000C3618000C35D,
-    0x0006F5B88000C365, 0x0006F5BA0006FDB9, 0x000699BC000689BB,
-    0x000689BE000699BD, 0x0006E5B0000681BF, 0x0006E5B20006EDB1,
-    0x0006E5B40006FDB3, 0x0006D1B60006EDB5, 0x8000C3690006CDB7,
-    0x0006FDB50006FDB6, 0x0006FDB38000C36D, 0x8000C3758000C371,
-    0x000639BF8000C379, 0x000639BD000631BE, 0x000639BB000621BC,
-    0x8000C231000631BA, 0x0006ADA38000C37D, 0x0000558100007982,
-    0x00015C8400005580, 0x0006ADA50006ADA6, 0x000669AB8000C385,
-    0x007F4886000661AA, 0x000669AF0000E487, 0x000669AD000661AE,
-    0x8000C389000671AC, 0x8000C38D0007DCEF, 0x8000C3958000C391,
-    0x8000C39D8000C399, 0x8000C3A58000C3A1, 0x8000C3A90003A084,
-    0x8000C3AD007924BF, 0x8000C3B1000768E3, 0x0002D1B3000674E1,
-    0x000304BE8000C3B5, 0x8000C3B9007D8084, 0x0002F1B50002F9B6,
-    0x0002B5BB8000C3BD, 0x8000C3C10002A9BA, 0x000351BF8000C3C5,
-    0x000291BD000299BE, 0x000205A90002A5BC, 0x000215AB0002B9A8,
-    0x00023DAD00021DAA, 0x000251AF00023DAC, 0x007D68BE00025DAE,
-    0x8000C3CD8000C3C9, 0x8000C3D58000C3D1, 0x8000C3DD8000C3D9,
-    0x0003F1B98000C3E1, 0x0003F1BB0003E9B8, 0x000351BD0003F9BA,
-    0x00034DBF0003E9BC, 0x000235B1000351BE, 0x000201B3000229B0,
-    0x0003D1B5000235B2, 0x0003D1B7000219B4, 0x000029820003D9B6,
-    0x00001D80000395A3, 0x0003BDA600001581, 0x8000C3E98000C3E5,
-    0x0003EDAA0003B5A5, 0x8000C3ED0003F1AB, 0x0003DDAE8000C3F5,
-    0x0003E1AC000215AF, 0x007D60860003D5AD, 0x000141B3007D7887,
-    0x000141B5007F0084, 0x8000C3FD8000C3F9, 0x8000C401000149B6,
-    0x000141BB8000C405, 0x000149BD00014DBA, 0x000129BF000145BC,
-    0x8000C409000139BE, 0x8000C40D0006DCEF, 0x8000C4158000C411,
-    0x8000C41D8000C419, 0x007E2C84000630EF, 0x8000C4210006E0E1,
-    0x8000C425000188E3, 0x8000C429000074E3, 0x0001C9A300012CE1,
-    0x8000C4318000C42D, 0x8000C43500797085, 0x0001C9A50001C1A6,
-    0x0001C9AB8000C439, 0x8000C43D0001C5AA, 0x0001A1AF8000C441,
-    0x0001C1AD0001B1AE, 0x007DDDA80001CDAC, 0x007E01AA007E05A9,
-    0x007E01AC007E01AB, 0x007E01AE007E01AD, 0x8000C3F1007E01AF,
-    0x8000C4498000C445, 0x8000C4518000C44D, 0x00009D8100009D82,
-    0x007E2DB800009D80, 0x007EE1BA007EE1B9, 0x007EE1BC007EE1BB,
-    0x007EE1BE007EE1BD, 0x007E41B0007EE1BF, 0x007E59B2007E4DB1,
-    0x007E35B4007E55B3, 0x007E25B6007E25B5, 0x007FADA8007E15B7,
-    0x007FD1AA007FD1A9, 0x007FF5AC007FEDAB, 0x007F11AE007F15AD,
-    0x8000C455007F11AF, 0x8000C45D8000C459, 0x0003F0868000C461,
-    0x8000C4650000B887, 0x007F11B88000C469, 0x007F21BA007F19B9,
-    0x0000F5BC007F21BB, 0x0000F5BE0000FDBD, 0x007F71B00000EDBF,
-    0x007F71B2007F71B1, 0x007F41B4007F45B3, 0x007F3DB6007F4DB5,
-    0x007E59B3007F35B7, 0x8000C4718000C46D, 0x8000C4798000C475,
-    0x007E7DB5007E95B6, 0x007EB5BB8000C47D, 0x8000C481007EB5BA,
-    0x007E89BF8000C485, 0x007E91BD007E85BE, 0x8000C489007EA5BC,
-    0x8000C48D007E1DA3, 0x007ED1A68000C491, 0x8000C4998000C495,
-    0x007EF1AA007E39A5, 0x8000C49D007EF1AB, 0x007EC1AE8000C4A1,
-    0x007EE1AC007ECDAF, 0x0000ADB0007ED5AD, 0x0000C1B20000C5B1,
-    0x0000C5B40000C1B3, 0x0000F1B60000CDB5, 0x000061B80000F1B7,
-    0x000061BA000061B9, 0x000065BC00007DBB, 0x000065BE00006DBD,
-    0x8000C4A50003BDBF, 0x8000C4AD8000C4A9, 0x8000C4B18000C381,
-    0x8000C4B98000C4B5, 0x000159A88000C4BD, 0x0000EDAA000159A9,
-    0x0000FDAC0000E5AB, 0x0000E5AE0000E5AD, 0x8000C4C10000D5AF,
-    0x00001D8100001D82, 0x8000C4C500001D80, 0x8000C4CD8000C4C9,
-    0x00048086000254BE, 0x8000C4D50002EC87, 0x8000C4DD8000C4D9,
-    0x8000C4E58000C4E1, 0x007E74E3000378BE, 0x007DB8E18000C4E9,
-    0x8000C4F18000C4ED, 0x8000C4F98000C4F5, 0x8000C5018000C4FD,
-    0x8000C5098000C505, 0x007EC0EF8000C50D, 0x8000C5158000C511,
-    0x000385B38000C519, 0x8000C5218000C51D, 0x8000C5298000C525,
-    0x000395B5000385B6, 0x0003E1BB8000C52D, 0x000624BE000389BA,
-    0x0003E9BF8000C531, 0x0003F5BD0003E1BE, 0x000029820003F5BC,
-    0x00001D800003C1A3, 0x0003C1A600001581, 0x8000C5398000C535,
-    0x0003CDAA0003D1A5, 0x8000C53D0003A5AB, 0x0003A5AE0005E085,
-    0x0003B1AC0003ADAF, 0x0003F8E10003B1AD, 0x001F5CE38000C545,
-    0x0003F4EF8000C549, 0x00073C868000C54D, 0x007E7CE30002AC87,
-    0x000180E18000C551, 0x8000C5598000C555, 0x0013E4EF8000C55D,
-    0x000171B38000C561, 0x8000C5698000C565, 0x8000C5718000C56D,
-    0x000161B5000111B6, 0x000139BB8000C575, 0x8000C579000131BA,
-    0x000119BF8000C57D, 0x000119BD000111BE, 0x8000C581000121BC,
-    0x8000C5898000C585, 0x8000C5918000C58D, 0x8000C5998000C595,
-    0x8000C59D007FC4EF, 0x8000C5A1007EFCE1, 0x00000D80007FF0E3,
-    0x00007D8200007581, 0x8000C5A98000C5A5, 0x0001F9A38000C5AD,
-    0x0001E9A5000800BE, 0x8000C5B98000C5B5, 0x0005A884000199A6,
-    0x0001B1AB8000C5BD, 0x000191AD0001B9AA, 0x000191AF0001A9AC,
-    0x000629A8000199AE, 0x000639AA000629A9, 0x000629AC000639AB,
-    0x000655AE000651AD, 0x8000C54100064DAF, 0x8000C5C100012084,
-    0x0001E1A38000C5B1, 0x000419A1000499A2, 0x000639B80005F1A0,
-    0x0006CDBA000639B9, 0x0006DDBC0006C5BB, 0x0006C5BE0006C5BD,
-    0x000635B00006F1BF, 0x000635B200063DB1, 0x000615B400060DB3,
-    0x000615B600061DB5, 0x006CA18F000609B7, 0x00012886000791B3,
-    0x0007A1B600037C87, 0x8000C5C98000C5C5, 0x0007E5BA0007B1B5,
-    0x8000C5CD0007EDBB, 0x0007EDBE8000C5D1, 0x0007F5BC0007DDBF,
-    0x0078F99F0007EDBD, 0x0079719D0078E59E, 0x007CF19B007C219C,
-    0x00719999007D619A, 0x0070E1970070D998, 0x00746D9500747996,
-    0x0069AD9300746D94, 0x0068799100687192, 0x006D2182006E1D90,
-    0x8000C5D5006FE583, 0x00684D868000C5D9, 0x0069998400759587,
-    0x0075A58A00699985, 0x8000C5DD0075B98B, 0x0070718E8000C5E1,
-    0x0071398C007C818F, 0x007D61920071398D, 0x8000C5E5007DA993,
-    0x007969968000C5E9, 0x0078E19400049997, 0x0005419A00797D95,
-    0x8000C5ED0005C99B, 0x8000C5F58000C5F1, 0x0000A99C8000C5F9,
-    0x000D81A38000C5FD, 0x0001A9A10001A9A2, 0x0009B1A78000C601,
-    0x000D4DA5000801A6, 0x001591AB000D99A4, 0x001401A9001585AA,
-    0x00101DAF000971A8, 0x0011BDAD0011E9AE, 0x001801B3001001AC,
-    0x001DC1B1001CF1B2, 0x8000C4D1001DC9B0, 0x0018E1B58000C605,
-    0x8000C6090019FDB4, 0x8000C6118000C60D, 0x00001D808000C615,
-    0x0003FD8200000981, 0x000515A38000C619, 0x000C88868000C61D,
-    0x8000C62500033C87, 0x000535A5000525A6, 0x000569AB8000C629,
-    0x8000C62D000561AA, 0x000559AF8000C631, 0x000569AD000569AE,
-    0x8000C635000571AC, 0x000C04848000C639, 0x8000C6418000C63D,
-    0x8000C64500066CE1, 0x8000C649007BECE3, 0x8000C6518000C64D,
-    0x8000C6550006A8EF, 0x8000C65D8000C659, 0x00058DA98000C661,
-    0x0005ADAB000581A8, 0x0005A1AD000599AA, 0x0005A5AF0005B9AC,
-    0x000D68840005A9AE, 0x8000C6698000C665, 0x8000C6718000C66D,
-    0x8000C6798000C675, 0x000549B9000CF4BE, 0x000559BB000541B8,
-    0x000549BD000541BA, 0x000571BF000541BC, 0x0005A9B100057DBE,
-    0x0005B9B30005A1B0, 0x000599B50005A1B2, 0x000579B70005A9B4,
-    0x0004D5A8000591B6, 0x0004EDAA0004DDA9, 0x000385AC000325AB,
-    0x0003B1AE00038DAD, 0x8000C67D0003B1AF, 0x8000C6858000C681,
-    0x000019808000C689, 0x0000058200001981, 0x000231B88000C68D,
-    0x000235BA000231B9, 0x000285BC0002C1BB, 0x0002BDBE0002B5BD,
-    0x000269B00002B5BF, 0x000241B2000269B1, 0x000239B4000241B3,
-    0x000211B6000239B5, 0x000CA086000211B7, 0x8000C695000D4C87,
-    0x0006A4EF8000C699, 0x8000C6A18000C69D, 0x000194E30007CCEF,
-    0x000138E10006A4E1, 0x8000C6A500065CE3, 0x8000C6AD8000C6A9,
-    0x8000C6B58000C6B1, 0x00042DB38000C6B9, 0x000345B58000C6BD,
-    0x8000C6C58000C6C1, 0x8000C6C9000345B6, 0x0002E5BB8000C6CD,
-    0x0002DDBD0002E5BA, 0x0002B5BF0002DDBC, 0x8000C6210002B5BE,
-    0x8000C6D18000C691, 0x8000C6D98000C6D5, 0x8000C6E18000C6DD,
-    0x000E31A88000C6E5, 0x000E01AA000E39A9, 0x000E71AC000E01AB,
-    0x000E75AE000E71AD, 0x000E19B0000E6DAF, 0x000E2DB2000E25B1,
-    0x000E21B4000E25B3, 0x000E21B6000E21B5, 0x000EC5B8000E21B7,
-    0x000EC5BA000ECDB9, 0x000EC5BC000EDDBB, 0x000F59BE000ECDBD,
-    0x000EA9B3000F59BF, 0x8000C6ED8000C6E9, 0x8000C6F58000C6F1,
-    0x000ED9B5000ED1B6, 0x000EC1BB8000C6F9, 0x8000C6FD000EFDBA,
-    0x000EC5BF00002CBE, 0x000ED1BD000EC5BE, 0x00006982000ED9BC,
-    0x00005980000EEDA3, 0x000E95A600005181, 0x8000C7058000C701,
-    0x000EB9AA000E9DA5, 0x0000C886000E85AB, 0x000E81AE0000AC87,
-    0x000E9DAC000E81AF, 0x8000C709000E95AD, 0x8000C70D000E91B3,
-    0x000EA5B68000C711, 0x8000C7198000C715, 0x000EE1BA000EADB5,
-    0x8000C71D000EE1BB, 0x000EE9BE8000C721, 0x000EF5BC000ED5BF,
-    0x000AA5A3000EE9BD, 0x8000C7298000C725, 0x8000C7318000C72D,
-    0x000DCDA5000DCDA6, 0x000C6DAB8000C735, 0x8000C739000C6DAA,
-    0x000C3DAF8000C73D, 0x000C55AD000C3DAE, 0x000E09A8000C55AC,
-    0x000E1DAA000E09A9, 0x000E21AC000E15AB, 0x000E21AE000E21AD,
-    0x8000C741000E21AF, 0x8000C7498000C745, 0x8000C7518000C74D,
-    0x8000C7598000C755, 0x0001CDB88000C75D, 0x0001DDBA0001D5B9,
-    0x0001CDBC0001D5BB, 0x000151BE000151BD, 0x000E21B0000151BF,
-    0x000E25B2000E21B1, 0x000E29B4000E39B3, 0x000E1DB6000E15B5,
-    0x8000C7610001F5B7, 0x8000C7698000C765, 0x8000C76D000D99A3,
-    0x000DADA6000DA5A5, 0x0003EC840002DCBE, 0x000DE9AA8000C775,
-    0x000DFDAC000DE9AB, 0x000DE1AE000DE1AD, 0x00004580000DDDAF,
-    0x0000598200004D81, 0x8000C779000345A3, 0x000341A6000341A5,
-    0x0004C0868000C77D, 0x00032DAA00030087, 0x00033DAC000325AB,
-    0x000325AE000325AD, 0x000259A8000315AF, 0x000381AA000385A9,
-    0x000385AC000381AB, 0x0003B1AE00038DAD, 0x8000C7810003B1AF,
-    0x8000C7898000C785, 0x8000C7918000C78D, 0x8000C7998000C795,
-    0x000365B88000C79D, 0x000365BA00036DB9, 0x000365BC00037DBB,
-    0x000365BE00036DBD, 0x0003A5B00000DDBF, 0x0003A5B20003ADB1,
-    0x0003ADB40003BDB3, 0x000395B600039DB5, 0x000209B300035DB7,
-    0x8000C7A58000C7A1, 0x8000C7AD8000C7A9, 0x00020DB500020DB6,
-    0x000269BB8000C7B1, 0x8000C7B5000261BA, 0x000259BF8000C7B9,
-    0x000269BD000269BE, 0x8000C7BD000271BC, 0x8000C7C58000C7C1,
-    0x0001A0E18000C7C9, 0x000368E38000C7CD, 0x000015818000C7D1,
-    0x00038CEF00001580, 0x8000C7D500001582, 0x8000C7DD8000C7D9,
-    0x000E18E1000570BE, 0x000F38E3000F94E1, 0x00029484000FD4E3,
-    0x8000C7E98000C7E5, 0x8000C7F18000C7ED, 0x8000C7F98000C7F5,
-    0x8000C8018000C7FD, 0x0001ECEF8000C805, 0x00046086000FF8EF,
-    0x000195B300050487, 0x000175B50005CC84, 0x8000C80D8000C809,
-    0x8000C811000175B6, 0x000151BB8000C815, 0x0000F5BD000159BA,
-    0x0000F5BF000149BC, 0x000625A80000F5BE, 0x000655AA000655A9,
-    0x0006B5AC0006ADAB, 0x0006B5AE0006BDAD, 0x8000C7E10006ADAF,
-    0x8000C81D8000C819, 0x8000C8258000C821, 0x8000C82D8000C829,
-    0x000769B88000C831, 0x000709BA000769B9, 0x000719BC000709BB,
-    0x00070DBE000719BD, 0x0006D5B0000701BF, 0x0006D5B20006DDB1,
-    0x000779B4000769B3, 0x000769B6000779B5, 0x0006DDA3000761B7,
-    0x8000C8398000C835, 0x8000C8418000C83D, 0x00063DA500063DA6,
-    0x000619AB8000C845, 0x8000C849000611AA, 0x0007BDAF8000C84D,
-    0x0007BDAD0007BDAE, 0x00005D80000601AC, 0x0000658200006D81,
-    0x0003D8BE000751B3, 0x000771B6000771B5, 0x0000E0868000C851,
-    0x00075DBA00031487, 0x000729BC000739BB, 0x00071DBE000715BD,
-    0x000695A80000D9BF, 0x000695AA00069DA9, 0x0006B5AC0006ADAB,
-    0x0006B5AE0006BDAD, 0x8000C8550006ADAF, 0x8000C85D8000C859,
-    0x8000C8658000C861, 0x8000C86D8000C869, 0x000171B88000C871,
-    0x000171BA000171B9, 0x0001D5BC000171BB, 0x0001D5BE0001DDBD,
-    0x0006D5B00001CDBF, 0x0006B1B20006B1B1, 0x000695B400068DB3,
-    0x000151B6000151B5, 0x8000C875000151B7, 0x8000C879000619A3,
-    0x000639A68000C87D, 0x8000C8818000C771, 0x000615AA000639A5,
-    0x8000C885000671AB, 0x000655AE8000C889, 0x000661AC000191AF,
-    0x8000C88D00065DAD, 0x8000C8958000C891, 0x8000C89D8000C899,
-    0x8000C8A58000C8A1, 0x8000C8AD8000C8A9, 0x8000C8B58000C8B1,
-    0x000019808000C8B9, 0x0000058200001981, 0x000280848000C8BD,
-    0x00030C87000380BE, 0x000668E1001CAC86, 0x000798E38000C8C5,
-    0x8000C8CD8000C8C9, 0x0007ACEF8000C8D1, 0x8000C8D98000C8D5,
-    0x8000C8E18000C8DD, 0x8000C8E98000C8E5, 0x8000C8F18000C8ED,
-    0x8000C8F5000365B3, 0x00036DB6000365B5, 0x8000C8FD8000C8F9,
-    0x0003E9BA8000C901, 0x0003FDBC0003E5BB, 0x0003D1BE0003EDBD,
-    0x8000C9050003D1BF, 0x8000C90D8000C909, 0x8000C9158000C911,
-    0x8000C91D8000C919, 0x000381A88000C921, 0x000381AA000381A9,
-    0x000381AC000381AB, 0x000381AE000381AD, 0x000341B0000381AF,
-    0x000345B200034DB1, 0x000379B4000355B3, 0x000319B6000379B5,
-    0x000329B8000319B7, 0x000339BA000329B9, 0x000329BC000339BB,
-    0x000319BE000329BD, 0x00001981000319BF, 0x000221A300001180,
-    0x000221A500002D82, 0x8000C9298000C925, 0x8000C92D000229A6,
-    0x0002A1AB8000C935, 0x0002A9AD0002ADAA, 0x000295AF0002B9AC,
-    0x00024C84000295AE, 0x001D6487001D48BE, 0x000340E3001CEC86,
-    0x0001A0E18000C939, 0x00039CEF8000C93D, 0x8000C9458000C941,
-    0x8000C94D8000C949, 0x8000C9558000C951, 0x8000C95D8000C959,
-    0x8000C9658000C961, 0x8000C96D8000C969, 0x8000C9758000C971,
-    0x001DA8840001B4EF, 0x8000C97900067CE1, 0x8000C97D000670E3,
-    0x8000C9858000C981, 0x0001E1B38000C989, 0x8000C9918000C98D,
-    0x8000C9998000C995, 0x0001F5B50001E5B6, 0x0001A1BB8000C99D,
-    0x001CE4BE0001D9BA, 0x0001ADBF8000C9A5, 0x0001B1BD0001A5BE,
-    0x001E15A80001B9BC, 0x001E29AA001E19A9, 0x001E3DAC001E29AB,
-    0x001E2DAE001E25AD, 0x8000C931001E25AF, 0x001FF981001FF982,
-    0x8000C9A1001FE180, 0x00001C868000C9A9, 0x001EC1B80003B087,
-    0x001EC1BA001EC1B9, 0x001EC1BC001EC1BB, 0x001EC1BE001EC1BD,
-    0x001E5DB0001EC1BF, 0x001E2DB2001E25B1, 0x001E21B4001E25B3,
-    0x001E19B6001E29B5, 0x001EA1A3001E19B7, 0x8000C9B18000C9AD,
-    0x8000C9B98000C9B5, 0x001EB5A5001EA5A6, 0x001EE1AB8000C9BD,
-    0x8000C9C1001E99AA, 0x001EEDAF8000C9C5, 0x001EF1AD001EE5AE,
-    0x8000C9C9001EF9AC, 0x8000C9CD001F99B3, 0x001FBDB68000C9D1,
-    0x8000C9D98000C9D5, 0x001F99BA001FB5B5, 0x8000C9DD001F99BB,
-    0x001F79BE8000C9E1, 0x001F79BC001F79BF, 0x001E25A8001F79BD,
-    0x001E55AA001E51A9, 0x001E79AC001E69AB, 0x001E69AE001E79AD,
-    0x8000C9E5001E69AF, 0x8000C9ED8000C9E9, 0x8000C9F58000C9F1,
-    0x8000C9FD8000C9F9, 0x001ED9B88000CA01, 0x001EF9BA001EEDB9,
-    0x001EE9BC001EF9BB, 0x001E9DBE001EE9BD, 0x001E19B0001E95BF,
-    0x001EE9B2001E19B1, 0x001EF9B4001EE9B3, 0x001EE9B6001EF9B5,
-    0x001EDDA3001EE9B7, 0x0000158100002982, 0x8000CA0500001D80,
-    0x001EF1A5001EF9A6, 0x001EDDAB8000CA09, 0x8000C8C1001EDDAA,
-    0x001E3DAF8000CA0D, 0x001E3DAD001E3DAE, 0x0002C884001E3DAC,
-    0x00010CBE000141B3, 0x000141B68000CA15, 0x8000CA1D8000CA19,
-    0x00014DBA000151B5, 0x00004886000125BB, 0x000125BE00015087,
-    0x000131BC00012DBF, 0x8000CA21000131BD, 0x000304848000CA25,
-    0x8000CA29000480BE, 0x8000CA2D001FA8EF, 0x003128BF8000CA31,
-    0x8000CA35001F74E3, 0x8000CA39001E38E1, 0x8000CA418000CA3D,
-    0x8000CA498000CA45, 0x0002CDA38000CA4D, 0x0002DDA58000CA51,
-    0x8000CA598000CA55, 0x8000CA5D0002CDA6, 0x0002A9AB8000CA61,
-    0x0002BDAD0002C1AA, 0x0002A1AF0002BDAC, 0x0002B5A90002A9AE,
-    0x000201AB001D69A8, 0x000201AD000209AA, 0x000231AF000219AC,
-    0x00056C84000201AE, 0x8000CA698000CA65, 0x8000CA718000CA6D,
-    0x00019D8100019D82, 0x0003E5B900019D80, 0x0003E5BB0003E5B8,
-    0x0003E5BD0003E5BA, 0x0003E5BF0003E5BC, 0x000221B10003E5BE,
-    0x000225B3000249B0, 0x000229B5000225B2, 0x000215B7000221B4,
-    0x0002CDA8000215B6, 0x0002D1AA0002D1A9, 0x000115AC00010DAB,
-    0x000101AE000105AD, 0x8000CA75000101AF, 0x8000CA7D8000CA79,
-    0x0004F8BE8000CA85, 0x8000CA8D8000CA89, 0x000115B88000CA91,
-    0x000129BA00011DB9, 0x0001F5BC000129BB, 0x0001F5BE0001FDBD,
-    0x000149B00001EDBF, 0x00015DB2000155B1, 0x00014DB4000155B3,
-    0x00013DB6000135B5, 0x0005A086000131B7, 0x8000CA950005C887,
-    0x8000CA990000BCEF, 0x8000CAA18000CA9D, 0x00072C84001E8CEF,
-    0x8000CAA5001EF0E1, 0x8000CAA9001E1CE3, 0x8000CAAD000194E1,
-    0x000219B300006CE3, 0x8000CAB58000CAB1, 0x000800848000CAB9,
-    0x0001DDB50001C5B6, 0x0001D1BB8000CABD, 0x8000CAC10001CDBA,
-    0x0001B1BF8000CAC5, 0x0001C1BD0001C9BE, 0x0005D9A30001C9BC,
-    0x8000CAC98000CA81, 0x8000CAD18000CACD, 0x00061DA5000605A6,
-    0x000611AB8000CAD5, 0x8000CAD900060DAA, 0x000671AF8000CADD,
-    0x000601AD000609AE, 0x8000CAE1000609AC, 0x00001D8100001D82,
-    0x8000CAE500001D80, 0x8000CAED8000CAE9, 0x000340868000CAF1,
-    0x8000CAF50003C487, 0x8000CAFD8000CAF9, 0x8000CB058000CB01,
-    0x0007ADA88000CB09, 0x0007B1AA0007B1A9, 0x000719AC0007B1AB,
-    0x00070DAE000705AD, 0x8000CB0D000705AF, 0x8000CB158000CB11,
-    0x8000CB1D8000CB19, 0x8000CB258000CB21, 0x00072DB88000CB29,
-    0x0000C5BA0000CDB9, 0x0000CDBC0000DDBB, 0x0000F5BE0000FDBD,
-    0x000749B000009DBF, 0x000751B2000755B1, 0x000739B4000729B3,
-    0x000725B6000739B5, 0x000639B3000715B7, 0x8000CB318000CB2D,
-    0x8000CB398000CB35, 0x000691B5000685B6, 0x000691BB8000CB3D,
-    0x8000CB4100068DBA, 0x0006FDBF8000CB45, 0x0006FDBD0006FDBE,
-    0x8000CB49000685BC, 0x8000CB4D00067DA3, 0x0006C1A68000CB51,
-    0x8000CB598000CB55, 0x0006C9AA0006D5A5, 0x8000CB5D0006D5AB,
-    0x0006B9AE00016CBE, 0x0006C1AC0006B9AF, 0x0001E9A80006B9AD,
-    0x0001F9AA0001E9A9, 0x0001E9AC0001F9AB, 0x000139AE0001E9AD,
-    0x0000F580000139AF, 0x0000C1820000FD81, 0x000010868000CB61,
-    0x8000CB6500017487, 0x0000CDB88000CA11, 0x0000D5BA0000D5B9,
-    0x0000FDBC0000E5BB, 0x000091BE000095BD, 0x000149B0000091BF,
-    0x000159B2000149B1, 0x000149B4000159B3, 0x0000FDB6000149B5,
-    0x0006F4EF0000F5B7, 0x8000CB6D8000CB69, 0x000230EF8000CB71,
-    0x8000CB798000CB75, 0x0007E0E38000CB7D, 0x000180E18000CB81,
-    0x000628E38000CB85, 0x0005C8E18000CB89, 0x000231B38000CB8D,
-    0x8000CB958000CB91, 0x8000CB9900000096, 0x000221B5000229B6,
-    0x0002CDBB8000CB9D, 0x8000CBA10002CDBA, 0x0002CDBF8000CBA5,
-    0x0002CDBD0002CDBE, 0x8000CBA90002CDBC, 0x8000CBB18000CBAD,
-    0x8000CBB50002FDA3, 0x0002E5A60002EDA5, 0x8000CBBD8000CBB9,
-    0x000201AA8000CBC1, 0x000201AC000201AB, 0x000201AE000201AD,
-    0x00000D80000201AF, 0x00001D8200001581, 0x8000CBC98000CBC5,
-    0x000C54BE8000CBCD, 0x000CC0868000CBD5, 0x8000CBD900032487,
-    0x8000CBE18000CBDD, 0x8000CBE98000CBE5, 0x0002B9A88000CBED,
-    0x000101AA000101A9, 0x000105AC000101AB, 0x000105AE00010DAD,
-    0x000DA884000139AF, 0x8000CBF58000CBF1, 0x8000CBFD8000CBF9,
-    0x8000CC058000CC01, 0x00012DB88000CC09, 0x0001CDBA0001C5B9,
-    0x0001C1BC0001C5BB, 0x0001F9BE0001C9BD, 0x00014DB00001F9BF,
-    0x000155B2000151B1, 0x000139B4000129B3, 0x000125B6000125B5,
-    0x000618E1000115B7, 0x000738E38000CC0D, 0x8000CC158000CC11,
-    0x8000CC19000C58BE, 0x000F6C848000CC1D, 0x000F60BE8000CC21,
-    0x8000CC298000CC25, 0x000670EF8000CC2D, 0x000015808000CC31,
-    0x00002D8200001981, 0x000660E3000FCC84, 0x0001A0E18000CC35,
-    0x000074EF8000CC39, 0x000CC8868000CC3D, 0x8000CC45000CFC87,
-    0x8000CC4D8000CC49, 0x000E09A38000CC51, 0x8000CC418000CBD1,
-    0x8000CC598000CC55, 0x000E0DA5000E0DA6, 0x000E15AB8000CC5D,
-    0x8000CC61000E09AA, 0x000E61AF8000CC65, 0x000E7DAD000E7DAE,
-    0x8000CC69000E01AC, 0x8000CC6D000EA9B3, 0x000EA9B68000CC71,
-    0x8000CC798000CC75, 0x000F49BA000EA9B5, 0x8000CC7D000F49BB,
-    0x000F49BE8000CC81, 0x000F49BC000F49BF, 0x000E5DA8000F49BD,
-    0x000E65AA000E6DA9, 0x000E65AC000E7DAB, 0x000E65AE000E6DAD,
-    0x8000CC85000FB9AF, 0x8000CC8D8000CC89, 0x8000CC958000CC91,
-    0x8000CC9D8000CC99, 0x000F51B88000CCA1, 0x000F55BA000F5DB9,
-    0x000F7DBC000F69BB, 0x000F6DBE000F65BD, 0x000FC9B0000F61BF,
-    0x000FD9B2000FC9B1, 0x000FC9B4000FD9B3, 0x000F7DB6000FC9B5,
-    0x8000CCA5000F71B7, 0x000F11B6000F11B5, 0x000011808000CCA9,
-    0x0000298200001981, 0x000F1DBC000F15B3, 0x000F61BE000F61BD,
-    0x8000CCAD000F7DBF, 0x000F15BA8000CCB1, 0x000FADA3000F09BB,
-    0x0000E8868000CCB5, 0x8000CCB90001C887, 0x000FADA5000FADA6,
-    0x000E4DAB8000CCBD, 0x8000CCC1000E4DAA, 0x000E4DAF8000CCC5,
-    0x000E4DAD000E4DAE, 0x000E71A8000E4DAC, 0x000E71AA000E71A9,
-    0x00019DAC000E71AB, 0x000185AE000185AD, 0x0000ECBE0001B5AF,
-    0x8000CCCD8000CCC9, 0x8000CCD58000CCD1, 0x8000CCDD8000CCD9,
-    0x000161B88000CCE1, 0x000161BA000161B9, 0x000161BC000161BB,
-    0x000161BE000161BD, 0x0001CDB0000161BF, 0x0001A1B20001A5B1,
-    0x0001A5B40001A1B3, 0x000191B60001ADB5, 0x000DF9B3000191B7,
-    0x8000CCE98000CCE5, 0x8000CCF18000CCED, 0x000215B5000205B6,
-    0x000261BB8000CCF5, 0x8000CCF9000209BA, 0x000269BF8000CCFD,
-    0x000275BD000261BE, 0x8000CD01000275BC, 0x8000CD05000DBDA3,
-    0x000241A68000CD09, 0x8000CD118000CD0D, 0x00024DAA000251A5,
-    0x8000CD15000225AB, 0x000225AE8000CD19, 0x000231AC00022DAF,
-    0x0000ED81000231AD, 0x0001D0EF0000ED80, 0x8000CD1D00001D82,
-    0x000438868000CD25, 0x8000CD2900035087, 0x8000CD318000CD2D,
-    0x0001E0E18000CD35, 0x000F64E38000CD39, 0x8000CD418000CD3D,
-    0x8000CD498000CD45, 0x8000CD4D000391B3, 0x00037DB60003B9B5,
-    0x8000CD558000CD51, 0x000359BA8000CD59, 0x000349BC000359BB,
-    0x0000FDBE000349BD, 0x000245A80000F5BF, 0x000255AA000255A9,
-    0x00027DAC000265AB, 0x0002B1AE0002B1AD, 0x000568BE0002B1AF,
-    0x8000CD618000CD5D, 0x8000CD698000CD65, 0x8000CD718000CD6D,
-    0x00015DB88000CD75, 0x000165BA00016DB9, 0x000119BC00010DBB,
-    0x00010DBE000119BD, 0x0002D1B0000105BF, 0x0002D1B20002D1B1,
-    0x000175B40002D1B3, 0x000175B600017DB5, 0x000F78E100016DB7,
-    0x000EB8E3000E34E3, 0x8000CD79000E7CE1, 0x8000CD818000CD7D,
-    0x8000CD898000CD85, 0x8000CD958000CD8D, 0x8000CD9D8000CD99,
-    0x000ED4EF000EE4EF, 0x00029DA38000CDA1, 0x0000698100006182,
-    0x00059C8400005180, 0x0002B5A5000271A6, 0x000255AB8000CDA5,
-    0x00044886000255AA, 0x0001F9AF0004CC87, 0x000245AD0001F1AE,
-    0x000695A8000245AC, 0x0006ADAA0006A5A9, 0x0006BDAC0006A5AB,
-    0x0006A5AE0006A1AD, 0x8000CD910006DDAF, 0x8000CDAD8000CDA9,
-    0x8000CDB58000CDB1, 0x8000CDBD8000CDB9, 0x00076DB88000CDC1,
-    0x00077DBA000775B9, 0x00076DBC000775BB, 0x0007CDBE0007C5BD,
-    0x0006A5B00007C5BF, 0x0006B9B20006ADB1, 0x000691B40006B1B3,
-    0x00075DB6000691B5, 0x000625B3000755B7, 0x8000CDC98000CDC5,
-    0x8000CDD18000CDCD, 0x000635B5000621B6, 0x000669BB8000CDD5,
-    0x8000CDD9000661BA, 0x000655BF8000CDDD, 0x000665BD000655BE,
-    0x8000CDE100066DBC, 0x8000CDE5000661A3, 0x000665A68000CDE9,
-    0x8000CDF18000CDED, 0x000625AA000671A5, 0x8000CDF500062DAB,
-    0x000611AE8000CDF9, 0x000629AC000611AF, 0x0006A9A8000621AD,
-    0x0006B9AA0006A9A9, 0x0006CDAC0006B1AB, 0x000131AE000131AD,
-    0x0001C980000131AF, 0x000005820001C981, 0x000278BE8000CDFD,
-    0x8000CE0100027884, 0x0001DDB88000CE05, 0x0001E5BA0001EDB9,
-    0x000199BC00018DBB, 0x00018DBE000199BD, 0x000151B0000185BF,
-    0x000151B2000151B1, 0x0001F5B4000151B3, 0x0001F5B60001FDB5,
-    0x000631B30001EDB7, 0x000028868000CE09, 0x8000CE0D0001DC87,
-    0x000645B50001B1B6, 0x000195BB8000CE11, 0x8000CE15000195BA,
-    0x000139BF8000CE19, 0x000185BD000131BE, 0x8000CD21000185BC,
-    0x8000CE1D000675A3, 0x0001F5A68000CE21, 0x8000CE298000CE25,
-    0x0001D1AA000601A5, 0x8000CE2D0001D1AB, 0x000175AE8000CE31,
-    0x0001C1AC00017DAF, 0x8000CE350001C1AD, 0x8000CE3D8000CE39,
-    0x8000CE458000CE41, 0x8000CE4D8000CE49, 0x8000CE558000CE51,
-    0x8000CE5D8000CE59, 0x8000CE658000CE61, 0x8000CE6D8000CE69,
-    0x003C0085000370BE, 0x000244840006C4E1, 0x0000618000078CE3,
-    0x0000618200006181, 0x003C4485000368EF, 0x8000CE75000210E1,
-    0x003D5086002BD8E3, 0x8000CE790003F487, 0x8000CE7D0007A4EF,
-    0x8000CE81000240EF, 0x8000CE898000CE85, 0x8000CE918000CE8D,
-    0x003C3C848000CE95, 0x8000CE9D8000CE99, 0x8000CEA58000CEA1,
-    0x8000CEA90002ECE3, 0x000351B300012CE1, 0x8000CEB18000CEAD,
-    0x8000CEB98000CEB5, 0x000371B5000379B6, 0x000339BB8000CEBD,
-    0x8000CEC1000331BA, 0x0000F5BF8000CEC5, 0x000315BD0000F5BE,
-    0x00023DA8000315BC, 0x000299AA000281A9, 0x0002B9AC000291AB,
-    0x0002D1AE0002B9AD, 0x003FA8840002D1AF, 0x8000CEC9003FA8BE,
-    0x8000CED18000CECD, 0x8000CED98000CED5, 0x000151B88000CEDD,
-    0x000151BA000151B9, 0x000171BC000151BB, 0x000171BE000171BD,
-    0x0002B5B0000171BF, 0x000281B20002BDB1, 0x000171B4000281B3,
-    0x000171B6000171B5, 0x0000B580000171B7, 0x0000B5820000BD81,
-    0x003FB0868000CEE5, 0x8000CEE9003CE087, 0x003E2CBE000054EF,
-    0x8000CEED000654E1, 0x8000CEF1000088E3, 0x8000CEF98000CEF5,
-    0x000251A38000CEFD, 0x0026D8BF8000CF01, 0x8000CF098000CF05,
-    0x000271A5000279A6, 0x000239AB8000CF0D, 0x8000CF11000231AA,
-    0x0001F5AF8000CF15, 0x000215AD0001F5AE, 0x00242D90000215AC,
-    0x00280D9200280591, 0x002D85940028D993, 0x002CC596002C3595,
-    0x003001980031B597, 0x0035959A00301599, 0x00346D9C0034299B,
-    0x00383D9E0035999D, 0x8000CE710038359F, 0x003EC5B5003ED5B6,
-    0x003ED1B38000CEE1, 0x8000CF1D8000CF19, 0x003ED9BF8000CF21,
-    0x003EC5BD003ED5BE, 0x003ED5BB003EC5BC, 0x8000CF25003EE9BA,
-    0x0009E5A900245D8F, 0x000C05AB000815A8, 0x001001AD000C0DAA,
-    0x001041AF000C01AC, 0x0000E1A10010BDAE, 0x000401A38000CF29,
-    0x00049DA50000E1A2, 0x0008B9A7000495A4, 0x0039F5A0000801A6,
-    0x003CE9A2003D05A1, 0x8000CF2D003DF5A3, 0x8000CF358000CF31,
-    0x0014C1B18000CF39, 0x001801B3001401B0, 0x0018D5B500147DB2,
-    0x8000CF3D0018D5B4, 0x002521828000CF41, 0x8000CF4500252983,
-    0x0029C5868000CF49, 0x00291984002C8187, 0x002D018A00291985,
-    0x8000CF4D002DF98B, 0x0031018E8000CF55, 0x0030918C0034898F,
-    0x0035459200311D8D, 0x0007E88600351993, 0x003959960001EC87,
-    0x0034F59400386197, 0x003CA19A00395995, 0x8000CF59003D459B,
-    0x00007D818000CF5D, 0x003C419C00007D80, 0x003FE9A800005582,
-    0x003F09AA003FFDA9, 0x003F1DAC003F05AB, 0x003F0DAE003F05AD,
-    0x8000CF61003F05AF, 0x8000CF698000CF65, 0x8000CF718000CF6D,
-    0x8000CF798000CF75, 0x003F0DB88000CF7D, 0x003F15BA003F15B9,
-    0x003F3DBC003F25BB, 0x003EF5BE003EFDBD, 0x003F7DB0003EEDBF,
-    0x003F41B2003F41B1, 0x003F41B4003F41B3, 0x003F3DB6003F4DB5,
-    0x003C81A3003F35B7, 0x8000CF858000CF81, 0x8000CF8D8000CF89,
-    0x003C95A5003C85A6, 0x003C85AB8000CF91, 0x8000CF95003CB9AA,
-    0x003C89AF8000CF99, 0x003C95AD003C85AE, 0x0003C884003C95AC,
-    0x8000CF9D003DEDB3, 0x003DE9B68000CFA1, 0x8000CFA98000CFA5,
-    0x003DADBA003DE9B5, 0x8000CFAD003DB5BB, 0x003D9DBE8000CFB1,
-    0x003DA5BC000221BF, 0x003D0DA8003D95BD, 0x003D21AA003D1DA9,
-    0x003D25AC003D3DAB, 0x003D25AE003D2DAD, 0x00000D80003D59AF,
-    0x00001D8200001581, 0x8000CFB98000CFB5, 0x8000CFC58000CFBD,
-    0x0002B9B80003B8BE, 0x000219BA000261B9, 0x000209BC000219BB,
-    0x000221BE000209BD, 0x003D2DB0000221BF, 0x003D35B2003D35B1,
-    0x003D1DB4003D05B3, 0x0002A1B60002A1B5, 0x003CA9A30002A1B7,
-    0x000528868000CFC9, 0x8000CFCD0002EC87, 0x003CADA5003CADA6,
-    0x003CF1AB8000CFD1, 0x8000CFD5003CE9AA, 0x000365AF8000CFD9,
-    0x003CD1AD003CD9AE, 0x8000CFDD003CE1AC, 0x8000CFE1000229B3,
-    0x0002F9B68000CFE5, 0x8000CFED8000CFE9, 0x0002D5BA0002F9B5,
-    0x8000CFF10002DDBB, 0x000179BE8000CFF5, 0x0002C5BC000179BF,
-    0x8000CFF9000179BD, 0x0002E5A38000CFFD, 0x000235A58000D001,
-    0x8000D0098000D005, 0x8000D00D000235A6, 0x000211AB8000D011,
-    0x0001B5AD000219AA, 0x0001B5AF000209AC, 0x003EF0E30001B5AE,
-    0x003E40E1003FACE1, 0x8000D015003F7CE3, 0x8000D01D8000D019,
-    0x00000D808000D021, 0x0000118200001181, 0x003EA8EF8000D025,
-    0x8000D02D8000D029, 0x00052DA8003E20EF, 0x000565AA00056DA9,
-    0x0005B5AC0005ADAB, 0x0005B5AE0005BDAD, 0x8000CFC10005ADAF,
-    0x0003E0BE0003E884, 0x000310868000D031, 0x8000D03500039887,
-    0x000669B88000D039, 0x000601BA000669B9, 0x000605BC000601BB,
-    0x000631BE00060DBD, 0x0005D5B0000631BF, 0x0005D5B20005DDB1,
-    0x000679B4000669B3, 0x000669B6000679B5, 0x000639A8000661B7,
-    0x000799AA000781A9, 0x0007B9AC000791AB, 0x0007D9AE0007B9AD,
-    0x8000D03D0007D5AF, 0x8000CF518000D041, 0x8000D0498000D045,
-    0x8000D0518000D04D, 0x000755B88000D055, 0x000769BA00075DB9,
-    0x000701BC000769BB, 0x000701BE000701BD, 0x0007ADB0000701BF,
-    0x0007B1B20007B1B1, 0x00079DB4000785B3, 0x00077DB6000775B5,
-    0x000611B3000771B7, 0x8000D05D8000D059, 0x8000D0658000D061,
-    0x000631B5000639B6, 0x000675BB8000D069, 0x8000D06D000669BA,
-    0x0006B1BF8000D071, 0x000651BD000659BE, 0x8000D075000665BC,
-    0x8000D079000655A3, 0x00067DA68000D07D, 0x8000D0858000D081,
-    0x00062DAA000675A5, 0x8000D089000631AB, 0x00061DAE8000D08D,
-    0x000621AC0006F5AF, 0x0001B9A8000615AD, 0x000129AA0001B9A9,
-    0x00013DAC000129AB, 0x00012DAE000125AD, 0x8000D091000125AF,
-    0x00001D8100001D82, 0x8000D09500001D80, 0x8000D09D8000D099,
-    0x000081B8000170BE, 0x000085BA00008DB9, 0x000089BC000099BB,
-    0x0000B5BE0000BDBD, 0x00015DB000007DBF, 0x0000E1B20000E1B1,
-    0x0000E1B40000E1B3, 0x0000D1B60000E9B5, 0x0002C8840000D1B7,
-    0x000338870002A5B3, 0x0002A1B600022886, 0x8000D0A98000D0A5,
-    0x00029DBA0002B1B5, 0x00036CBE000321BB, 0x000321BE8000D0AD,
-    0x000331BC000329BF, 0x0002E1A3000331BD, 0x8000D0B58000D0B1,
-    0x8000D0BD8000D0B9, 0x0002F5A50002E5A6, 0x000365AB8000D0C1,
-    0x8000D0C50002D9AA, 0x00036DAF8000D0C9, 0x000375AD000365AE,
-    0x000099A9000375AC, 0x0000CDAB000091A8, 0x0000DDAD0000A1AA,
-    0x0000C5AF0000DDAC, 0x000D2CBE0000CDAE, 0x8000D0D18000D0CD,
-    0x8000D0D98000D0D5, 0x8000D0E18000D0DD, 0x0001C1B98000D0E5,
-    0x0001C1BB000079B8, 0x0001C1BD0001C9BA, 0x0001C5BF0001D9BC,
-    0x00008DB10001C5BE, 0x000041B300008DB0, 0x000041B5000049B2,
-    0x000041B7000059B4, 0x8000D0E9000049B6, 0x8000D0F18000D0ED,
-    0x8000D0F98000D0F5, 0x8000D0FD000750EF, 0x000FF8BE8000D101,
-    0x8000D105000774E3, 0x0000198000047CE1, 0x0000798200000981,
-    0x8000D10D8000D109, 0x8000D115000169B3, 0x00011DB60002C484,
-    0x8000D11D8000D119, 0x000109BA000115B5, 0x000DE886000109BB,
-    0x0001DDBE000DE887, 0x0001DDBC0001C5BF, 0x8000D1210001D5BD,
-    0x8000D1298000D125, 0x0000D4EF8000D12D, 0x8000D1358000D131,
-    0x0003B0E300014CEF, 0x000180E10006B4E1, 0x8000D139000198E3,
-    0x8000D1418000D13D, 0x8000D1498000D145, 0x0001E5A38000D14D,
-    0x000199A5000DC084, 0x8000D1558000D151, 0x8000D159000191A6,
-    0x000185AB8000D15D, 0x000159AD000185AA, 0x000149AF000151AC,
-    0x8000D111000151AE, 0x8000D1658000D161, 0x8000D16D8000D169,
-    0x8000D1758000D171, 0x000F69A88000D179, 0x000F71AA000F71A9,
-    0x000FB5AC000FADAB, 0x000FB5AE000FBDAD, 0x000FD9B0000FA9AF,
-    0x000FFDB2000FF5B1, 0x000FEDB4000FF5B3, 0x000F9DB6000F95B5,
-    0x000FB9B8000F89B7, 0x000F8DBA000F85B9, 0x000079BC000069BB,
-    0x000069BE000079BD, 0x00009D81000069BF, 0x8000D17D00009D80,
-    0x8000D18100000582, 0x8000D1898000D185, 0x000380868000D18D,
-    0x8000D1910003F487, 0x8000D1998000D195, 0x8000D1A18000D19D,
-    0x000F99B38000D0A1, 0x8000D1A98000D1A5, 0x8000D1B18000D1AD,
-    0x000F5DB5000F5DB6, 0x000F51BB8000D1B5, 0x8000D1B9000F7DBA,
-    0x000F31BF8000D1BD, 0x000F41BD000F49BE, 0x000E59A3000F49BC,
-    0x8000D1C58000D1C1, 0x8000D1CD8000D1C9, 0x000E9DA5000E9DA6,
-    0x000E91AB8000D1D1, 0x8000D1D5000EBDAA, 0x000EF1AF8000D1D9,
-    0x000E81AD000E89AE, 0x8000D1DD000E89AC, 0x8000D1E58000D1E1,
-    0x000019808000D1E9, 0x0000058200001981, 0x0001A0848000D1ED,
-    0x0001E8878000D1F1, 0x8000D1F500040086, 0x8000D1FD8000D1F9,
-    0x000115A88000D201, 0x000115AA00011DA9, 0x00013DAC000125AB,
-    0x00012DAE000125AD, 0x8000D205000125AF, 0x8000D20D8000D209,
-    0x8000D2158000D211, 0x8000D21D8000D219, 0x0000C9B88000D221,
-    0x0000D9BA0000C9B9, 0x0000F9BC0000D1BB, 0x000099BE0000F9BD,
-    0x000125B0000099BF, 0x000125B200012DB1, 0x00012DB400013DB3,
-    0x000115B600011DB5, 0x8000D2250000F9B7, 0x8000D22D8000D229,
-    0x8000D231000291B3, 0x0002A9B60002B9B5, 0x8000D2398000D235,
-    0x0002EDBA8000D23D, 0x0002FDBC0002E5BB, 0x0002E5BE0002E5BD,
-    0x0002F1A60002D5BF, 0x8000D2458000D241, 0x8000D2490002E1A5,
-    0x8000D24D0002C9A3, 0x0002BDAE8000D251, 0x0002A5AC00028DAF,
-    0x0002B5AA0002BDAD, 0x8000D2550002BDAB, 0x0000F9808000D259,
-    0x000005820000F981, 0x0003C8BE8000D25D, 0x8000D26500031884,
-    0x8000D26D8000D269, 0x8000D2758000D271, 0x8000D27D8000D279,
-    0x000418868000D281, 0x8000D28500039C87, 0x8000D28D8000D289,
-    0x8000D2958000D291, 0x0002B0EF8000D299, 0x000194E18000D29D,
-    0x00026CE38000D2A1, 0x8000D2A98000D2A5, 0x8000D2B18000D2AD,
-    0x000FD0EF8000D2B5, 0x8000D2BD8000D2B9, 0x8000D2C58000D2C1,
-    0x8000D2C90001ACE1, 0x0000318000006CE3, 0x00001D8200003D81,
-    0x000E6CE3000EF4EF, 0x000F2CE18000D2CD, 0x000239B3000570BE,
-    0x0004E88600050C84, 0x000000DC00056087, 0x0002F9B50002F1B6,
-    0x0002D5BB8000D2D5, 0x8000D2D90002D5BA, 0x000175BF8000D2DD,
-    0x0002C5BD000175BE, 0x8000D2E10002C5BC, 0x8000D2E5000EF4E1,
-    0x8000D2E9000E50E3, 0x8000D2F18000D2ED, 0x8000D2F98000D2F5,
-    0x8000D3018000D2FD, 0x8000D3098000D305, 0x8000D3118000D30D,
-    0x000265A6000FE4EF, 0x8000D3198000D315, 0x8000D31D00026DA5,
-    0x8000D3210002ADA3, 0x0001E1AE8000D325, 0x000251AC0001E1AF,
-    0x000241AA000251AD, 0x8000D329000241AB, 0x000699A88000D32D,
-    0x0006A9AA000699A9, 0x0006B9AC0006A9AB, 0x0006A9AE0006B9AD,
-    0x8000D3310006A9AF, 0x00001D8100001D82, 0x8000D33500001D80,
-    0x8000D33D8000D339, 0x0006ADB8000374BE, 0x0006BDBA0006B5B9,
-    0x0006ADBC0006B5BB, 0x000751BE000751BD, 0x0006A1B0000751BF,
-    0x0006A1B20006A1B1, 0x0006A1B40006A1B3, 0x0006A5B60006A1B5,
-    0x8000D261000699B7, 0x0003C484000625B3, 0x00063DB68000D2D1,
-    0x8000D3458000D341, 0x000661BA000635B5, 0x00000886000661BB,
-    0x000661BE00008887, 0x000671BC000661BF, 0x000661A3000671BD,
-    0x8000D34D8000D349, 0x8000D3558000D351, 0x000671A5000679A6,
-    0x000625AB8000D359, 0x8000D35D000625AA, 0x000625AF8000D361,
-    0x000635AD000625AE, 0x00065DA8000635AC, 0x000665AA00066DA9,
-    0x000699AC00068DAB, 0x00068DAE000685AD, 0x8000D365000685AF,
-    0x8000D36D8000D369, 0x8000D3758000D371, 0x8000D37D8000D379,
-    0x000685B88000D381, 0x000699BA00068DB9, 0x00068DBC000695BB,
-    0x0006A5BE0006ADBD, 0x0006FDB000017DBF, 0x0006CDB20006C5B1,
-    0x0006DDB40006C5B3, 0x0006CDB60006C5B5, 0x0006EDB30006C5B7,
-    0x8000D3898000D385, 0x8000D3918000D38D, 0x000611B5000605B6,
-    0x000601BB8000D395, 0x8000D399000639BA, 0x000601BF8000D39D,
-    0x000611BD000619BE, 0x8000D3A1000619BC, 0x8000D3A50006A9A3,
-    0x000641A68000D3A9, 0x000194848000D3AD, 0x00067DAA000655A5,
-    0x0001A0BE000645AB, 0x00065DAE8000D3B5, 0x00065DAC000645AF,
-    0x000299A8000655AD, 0x0002C1AA0002C1A9, 0x0002C1AC0002C1AB,
-    0x0002F1AE0002C9AD, 0x0003ED810002F1AF, 0x8000D3B90003ED80,
-    0x001C90860003F982, 0x8000D3BD00037487, 0x0003C5B88000D3C1,
-    0x0003C5BA0003CDB9, 0x0003CDBC0003DDBB, 0x0003F5BE0003FDBD,
-    0x000341B000039DBF, 0x000341B2000341B1, 0x000341B4000341B3,
-    0x000341B6000341B5, 0x000249B3000341B7, 0x8000D3C98000D3C5,
-    0x8000D3D18000D3CD, 0x000249B5000249B6, 0x000385BB8000D3D5,
-    0x8000D3D9000385BA, 0x000385BF8000D3DD, 0x000395BD000385BE,
-    0x8000D3E1000395BC, 0x8000D3E500020DA3, 0x00020DA68000D3E9,
-    0x8000D3F18000D3ED, 0x0003C1AA00020DA5, 0x8000D3F50003C1AB,
-    0x0003C1AE8000D3F9, 0x0003D1AC0003C1AF, 0x000198E30003D1AD,
-    0x000658E10007A4E1, 0x0001A0E1000658E3, 0x0000D0E38000D3FD,
-    0x8000D4058000D401, 0x000038EF8000D409, 0x0001F4EF8000D40D,
-    0x8000D4158000D411, 0x000079800006F8EF, 0x00001D8200001581,
-    0x8000D419001D0084, 0x001D10BE8000D41D, 0x001CC0868000D425,
-    0x8000D429001DAC87, 0x8000D4318000D42D, 0x8000D4398000D435,
-    0x0005C5BB8000D43D, 0x0005A9B90005A1BA, 0x0005CDBF000591B8,
-    0x0005C9BD0005CDBE, 0x00061DB30005D5BC, 0x00061DB100061DB2,
-    0x0005A1B700061DB0, 0x0005BDB50005BDB6, 0x0006BDAB0005BDB4,
-    0x0006BDA90006BDAA, 0x00067DAF0006BDA8, 0x00067DAD00067DAE,
-    0x8000D44100067DAC, 0x8000D4498000D445, 0x8000D4518000D44D,
-    0x8000D4598000D455, 0x001DEDA88000D45D, 0x001E31AA001E2DA9,
-    0x001E95AC001E31AB, 0x001E95AE001E9DAD, 0x8000D421001E8DAF,
-    0x8000D4658000D461, 0x8000D46D8000D469, 0x8000D4758000D471,
-    0x001EA9B88000D479, 0x001F5DBA001EA9B9, 0x001F71BC001F51BB,
-    0x001F75BE001F71BD, 0x001EFDB0001F69BF, 0x001EC5B2001ECDB1,
-    0x001EB9B4001EADB3, 0x001EADB6001EB9B5, 0x001EB9B3001EA5B7,
-    0x8000D4818000D47D, 0x8000D3B18000D485, 0x001E1DB5001E1DB6,
-    0x001E09BB8000D489, 0x8000D48D001E39BA, 0x001E7DBF0003E084,
-    0x001E79BD001E7DBE, 0x00006982001E11BC, 0x00004580001EFDA3,
-    0x001E59A600005181, 0x8000D49100039CBE, 0x001E7DAA001E59A5,
-    0x00004886001E4DAB, 0x001E39AE00006C87, 0x001E55AC001E39AF,
-    0x001E5DA8001E3DAD, 0x001E65AA001E6DA9, 0x001E65AC001E7DAB,
-    0x001E65AE001E6DAD, 0x8000D495001EFDAF, 0x8000D49D8000D499,
-    0x8000D4A58000D4A1, 0x8000D4AD8000D4A9, 0x000169B88000D4B1,
-    0x000179BA000169B9, 0x000169BC000179BB, 0x0001DDBE000169BD,
-    0x001E85B00001D5BF, 0x001E85B2001E8DB1, 0x001E85B4001E9DB3,
-    0x001E85B6001E8DB5, 0x001EEDB3000159B7, 0x8000D4B98000D4B5,
-    0x8000D4C18000D4BD, 0x001EEDB5001EEDB6, 0x001E49BB8000D4C5,
-    0x8000D4C9001E41BA, 0x001E49BF8000D4CD, 0x001E49BD001E41BE,
-    0x8000D4D1001E51BC, 0x8000D4D5001EA9A3, 0x001EA9A68000D4D9,
-    0x8000D4E18000D4DD, 0x001E05AA001EA9A5, 0x8000D4E5001E0DAB,
-    0x001E05AE8000D4E9, 0x001E15AC001E0DAF, 0x000321A8001E0DAD,
-    0x000321AA000321A9, 0x000321AC000321AB, 0x000321AE000321AD,
-    0x8000D4ED000321AF, 0x8000D4F58000D4F1, 0x8000D4F9000260BE,
-    0x8000D5058000D4FD, 0x000389B88000D509, 0x00039DBA000389B9,
-    0x0003B9BC000395BB, 0x000079BE0003B9BD, 0x0003E5B0000079BF,
-    0x0003E5B20003EDB1, 0x0003EDB40003FDB3, 0x0003D5B60003DDB5,
-    0x0000A9800003BDB7, 0x0000BD820000B581, 0x8000D50D0003A5B3,
-    0x0003ADB60003A5B5, 0x0002E0848000D511, 0x00032DBA8000D515,
-    0x00033DBC000325BB, 0x000325BE000325BD, 0x0003E9A3000315BF,
-    0x000468868000D519, 0x8000D51D00038087, 0x0003E9A50003E1A6,
-    0x000369AB8000D521, 0x8000D525000361AA, 0x000359AF8000D529,
-    0x000369AD000369AE, 0x8000D52D000371AC, 0x8000D5358000D531,
-    0x8000D53D8000D539, 0x8000D541001F3CE1, 0x8000D545001E40E3,
-    0x8000D54D8000D549, 0x8000D551001E8CEF, 0x8000D5598000D555,
-    0x8000D5618000D55D, 0x0000118100002582, 0x8000D56500001D80,
-    0x8000D5690003E4E3, 0x8000D56D0001ACE1, 0x000264840003B0EF,
-    0x0004D086000560BE, 0x8000D57500054487, 0x8000D57D8000D579,
-    0x8000D5858000D581, 0x8000D58D8000D589, 0x8000D5958000D591,
-    0x0005A4840001B0EF, 0x8000D599001EDCE1, 0x8000D59D000110E3,
-    0x8000D5A58000D5A1, 0x000151B38000D5A9, 0x8000D5B18000D5AD,
-    0x8000D5B98000D5B5, 0x00017DB5000111B6, 0x00010DBB8000D5BD,
-    0x8000D5C100010DBA, 0x0000FDBF8000D5C5, 0x0000FDBD0000FDBE,
-    0x000639A80000FDBC, 0x000699AA000639A9, 0x0006B5AC000691AB,
-    0x0006C9AE0006D1AD, 0x8000D5710006C1AF, 0x8000D5CD8000D5C9,
-    0x00000D808000D5D1, 0x0000B1820000B181, 0x000761B88000D5D5,
-    0x000761BA000761B9, 0x000765BC00077DBB, 0x000765BE00076DBD,
-    0x000689B000071DBF, 0x000769B2000689B1, 0x000779B4000769B3,
-    0x000769B6000779B5, 0x000611A3000765B7, 0x8000D5DD8000D5D9,
-    0x8000D5E1000380BE, 0x00063DA5000651A6, 0x00064DAB8000D5E5,
-    0x0000088600064DAA, 0x0007BDAF00037C87, 0x0007BDAD0007BDAE,
-    0x8000D5E90007BDAC, 0x000749B38000D5ED, 0x000759B58000D5F1,
-    0x8000D5F98000D5F5, 0x8000D5FD000751B6, 0x000741BB8000D501,
-    0x000745BD000775BA, 0x000745BF000745BC, 0x000679A8000745BE,
-    0x000689AA000679A9, 0x000699AC000689AB, 0x000689AE000699AD,
-    0x8000D601000689AF, 0x8000D6098000D605, 0x8000D6118000D60D,
-    0x8000D6198000D615, 0x00068DB88000D61D, 0x000695BA000695B9,
-    0x0006BDBC0006A5BB, 0x000171BE000171BD, 0x0006F9B0000171BF,
-    0x0006D9B20006CDB1, 0x0006C9B40006D9B3, 0x0006BDB60006C9B5,
-    0x000601A30006B5B7, 0x8000D6258000D621, 0x8000D62D8000D629,
-    0x000611A5000619A6, 0x000609AB8000D631, 0x8000D63500063DAA,
-    0x00060DAF8000D639, 0x00060DAD00060DAE, 0x8000D63D00060DAC,
-    0x8000D6458000D641, 0x000019808000D649, 0x0000058200001981,
-    0x0001AC848000D64D, 0x0000E8870001ACBE, 0x8000D655000F4C86,
-    0x8000D65D8000D659, 0x00029DA88000D661, 0x0002CDAA0002C5A9,
-    0x0002C5AC0002C1AB, 0x0002F9AE0002CDAD, 0x8000D66500033DAF,
-    0x8000D66D8000D669, 0x000C3CBE8000D671, 0x8000D6798000D675,
-    0x0003C9B88000D67D, 0x0003D9BA0003C9B9, 0x0003F9BC0003D1BB,
-    0x000399BE0003F9BD, 0x000345B0000399BF, 0x000345B200034DB1,
-    0x000345B400035DB3, 0x000345B600034DB5, 0x000245B30003F9B7,
-    0x8000D6858000D681, 0x8000D68D8000D689, 0x00024DB500024DB6,
-    0x000389BB8000D691, 0x8000D695000381BA, 0x000389BF8000D699,
-    0x000389BD000381BE, 0x8000D69D000391BC, 0x8000D6A1000201A3,
-    0x000209A68000D6A5, 0x8000D6AD8000D6A9, 0x0003C5AA000209A5,
-    0x8000D6B10003CDAB, 0x0003C5AE8000D6B5, 0x0003D5AC0003CDAF,
-    0x0001ED800003CDAD, 0x0000118200001581, 0x8000D6B900020084,
-    0x8000D6BD0001A4E1, 0x8000D6C50000ECE3, 0x8000D6CD8000D6C9,
-    0x8000D6D1000030EF, 0x8000D6D98000D6D5, 0x000CE0868000D6DD,
-    0x8000D6E10002F487, 0x8000D6E98000D6E5, 0x0006CCEF8000D6ED,
-    0x0007B0E18000D6F1, 0x000610E38000D6F5, 0x8000D6FD8000D6F9,
-    0x8000D7058000D701, 0x8000D70D8000D709, 0x8000D7158000D711,
-    0x8000D71D8000D719, 0x8000D7258000D721, 0x000F18850001E8EF,
-    0x8000D7290006CCE1, 0x0000298000061CE3, 0x0000058200001D81,
-    0x000239B38000D72D, 0x000F6886000DCC84, 0x03D2D0E1000CFC87,
-    0x0001F9B50001F1B6, 0x0001D9BB8000D735, 0x000CE4BE0001D1BA,
-    0x00017DBF8000D739, 0x00017DBD00017DBE, 0x000DF1A80001C1BC,
-    0x000DF1AA000DF1A9, 0x000E31AC000DF1AB, 0x000E31AE000E31AD,
-    0x8000D6C1000E31AF, 0x8000D73D8000D731, 0x8000D7458000D741,
-    0x8000D74D8000D749, 0x000EE9B88000D751, 0x000E89BA000EE9B9,
-    0x000E9DBC000E85BB, 0x000E81BE000E81BD, 0x000E55B0000EB5BF,
-    0x000E55B2000E5DB1, 0x000EF9B4000EE9B3, 0x000EE9B6000EF9B5,
-    0x000D79A3000EE1B7, 0x8000D7598000D755, 0x8000D7618000D75D,
-    0x000EB9A5000EB1A6, 0x000E99AB8000D765, 0x8000D769000E91AA,
-    0x000E3DAF8000D76D, 0x000E3DAD000E3DAE, 0x8000D771000E81AC,
-    0x8000D775000FEDB3, 0x000FE9B68000D779, 0x8000D7818000D77D,
-    0x000FADBA000FE9B5, 0x8000D651000FB5BB, 0x000F95BE8000D785,
-    0x000FADBC000F99BF, 0x000E21A8000FA1BD, 0x000E21AA000E21A9,
-    0x000E25AC000E3DAB, 0x000E25AE000E2DAD, 0x8000D789000E55AF,
-    0x8000D7918000D78D, 0x00001D808000D795, 0x0000BD8200000981,
-    0x000E39B88000D799, 0x000EC9BA000E39B9, 0x000ED9BC000EC9BB,
-    0x000EC9BE000ED9BD, 0x000E2DB0000EC9BF, 0x000E3DB2000E35B1,
-    0x000E15B4000E31B3, 0x000E09B6000E19B5, 0x000EA9A3000E09B7,
-    0x000280848000D79D, 0x00040085000180BE, 0x000EADA5000EADA6,
-    0x000EF1AB8000D7A5, 0x00072886000EE9AA, 0x000EDDAF00001887,
-    0x000EE5AD000ED1AE, 0x8000D7A9000EE9AC, 0x8000D7AD0001EDB3,
-    0x0001E5B68000D7B1, 0x8000D7B98000D7B5, 0x000165BA0001EDB5,
-    0x8000D7BD00016DBB, 0x00016DBE8000D7C1, 0x000175BC00015DBF,
-    0x000DDDA800016DBD, 0x000221AA000DE9A9, 0x000221AC000221AB,
-    0x000221AE000221AD, 0x8000D7C5000221AF, 0x8000D7CD8000D7C9,
-    0x000211A28000D7D1, 0x000EA9A0000211A3, 0x000289B80002D5A1,
-    0x00029DBA000289B9, 0x0002BDBC000295BB, 0x00037DBE000375BD,
-    0x0002E5B0000375BF, 0x0002E5B20002EDB1, 0x0002EDB40002FDB3,
-    0x0002D5B60002DDB5, 0x0002A9A30002BDB7, 0x8000D7D5001AC58F,
-    0x8000D7DD8000D7D9, 0x0002A9A50002A1A6, 0x000229AB8000D7E1,
-    0x8000D7E5000221AA, 0x000219AF8000D7E9, 0x000229AD000229AE,
-    0x000E259E000231AC, 0x000AE99C000EE59F, 0x0016459A000A159D,
-    0x00165998000A459B, 0x0012719600165199, 0x0012559400138197,
-    0x001ED1920012ED95, 0x001AB590001ED993, 0x001FA984001E5591,
-    0x001F2586001F2585, 0x8000D7ED0013B187, 0x001B59828000D7F1,
-    0x0012458C001B5183, 0x0017A58E0017A58D, 0x01FB39800017B98F,
-    0x0013398A8000D7F5, 0x000B99940013AD8B, 0x000F6996000FA195,
-    0x8000D7F9000FDD97, 0x000BC992000F9CEF, 0x0003459C000B7D93,
-    0x8000D7FD000E78E3, 0x8000D801000C98E1, 0x0003919A00027884,
-    0x000004E100035D9B, 0x03EADCE3000588BE, 0x8000D8098000D805,
-    0x0000ECEF8000D80D, 0x000E70E1000E8CEF, 0x000030E3000EFCE1,
-    0x00002181000E78E3, 0x03EDB4EF00003980, 0x000209B300002982,
-    0x000468868000D815, 0x8000D81900052C87, 0x00020DB500020DB6,
-    0x0001C5BB8000D81D, 0x8000D8210001C5BA, 0x00017DBF8000D825,
-    0x0001D5BD00017DBE, 0x8000D7A10001D5BC, 0x8000D8298000D811,
-    0x8000D8318000D82D, 0x8000D8398000D835, 0x0005BDA88000D83D,
-    0x0005E1AA0005E5A9, 0x0005E5AC0005E1AB, 0x0005D1AE0005EDAD,
-    0x000661B00005D1AF, 0x000661B2000661B1, 0x0006D9B4000661B3,
-    0x0006C9B60006D9B5, 0x0006C9B80006C1B7, 0x000779BA0006C9B9,
-    0x000745BC000779BB, 0x000711BE000725BD, 0x000549A300070DBF,
-    0x8000D8458000D841, 0x8000D84D8000D849, 0x00054DA500054DA6,
-    0x000685AB8000D851, 0x8000D855000685AA, 0x00063DAF8000D859,
-    0x000695AD00063DAE, 0x8000D85D000695AC, 0x8000D8658000D861,
-    0x8000D86D8000D869, 0x8000D8758000D871, 0x000039808000D879,
-    0x0000058200003981, 0x0003B8BE8000D87D, 0x8000D8850003B884,
-    0x0006C5A88000D889, 0x0006D5AA0006D5A9, 0x0006FDAC0006E5AB,
-    0x000115AE00011DAD, 0x8000D88100010DAF, 0x000188868000D88D,
-    0x8000D89100011C87, 0x8000D8998000D895, 0x0001E9B88000D89D,
-    0x000189BA0001E9B9, 0x000199BC000189BB, 0x000189BE000199BD,
-    0x000175B0000189BF, 0x000175B200017DB1, 0x0001F9B40001E9B3,
-    0x0001E9B60001F9B5, 0x00065DB30001E1B7, 0x8000D8A58000D8A1,
-    0x0001BC848000D8A9, 0x000675B500019DB6, 0x0001B9BB8000D8AD,
-    0x8000D8B10001B1BA, 0x00013DBF8000D8B5, 0x00013DBD00013DBE,
-    0x8000D8B90001A1BC, 0x8000D8BD000619A3, 0x0001D9A68000D8C1,
-    0x8000D8C98000D8C5, 0x0001F5AA000631A5, 0x8000D8CD0001FDAB,
-    0x000179AE8000D8D1, 0x0001E5AC000179AF, 0x0002F9A8000179AD,
-    0x00032DAA0002F9A9, 0x000325AC00033DAB, 0x000325AE00032DAD,
-    0x8000D8D5000399AF, 0x8000D8DD8000D8D9, 0x00000D808000D8E1,
-    0x0000B1820000B181, 0x000395B88000D8E5, 0x0003A1BA00039DB9,
-    0x000071BC0003A1BB, 0x000071BE000071BD, 0x0003E9B0000071BF,
-    0x0003F5B20003E9B1, 0x0003DDB40003C5B3, 0x0003B1B60003B5B5,
-    0x0003CCBE0003B1B7, 0x8000D8ED8000D8E9, 0x8000D8F98000D8F5,
-    0x8000D9018000D8FD, 0x8000D9050002E4EF, 0x8000D909000194E1,
-    0x8000D90D00012CE3, 0x000318878000D911, 0x00034DBB03FC1C86,
-    0x8000D91500034DBA, 0x000311BF8000D919, 0x000371BD000379BE,
-    0x0003C5B3000341BC, 0x8000D91D03FCE884, 0x8000D9258000D921,
-    0x0003CDB50003CDB6, 0x03FC01A48000D929, 0x03FFD9A603FF25A5,
-    0x8000D92D03F801A7, 0x00031DA2000315A1, 0x03F711AC00023DA3,
-    0x03F32DAE03F001AD, 0x03FB11A803F301AF, 0x03F401AA03FB99A9,
-    0x03E801B403F71DAB, 0x03FCC4BE03EB65B5, 0x03EC01B003FCC484,
-    0x03EF4DB203EF55B1, 0x8000D93103EE71B3, 0x8000D9398000D935,
-    0x8000D9418000D93D, 0x8000D9498000D945, 0x000650E18000D94D,
-    0x00016CE3000434E1, 0x8000D9510006E8E3, 0x8000D9598000D955,
-    0x000035808000D95D, 0x0000358200003D81, 0x8000D9698000D965,
-    0x0001B0EF8000D96D, 0x0002E5A30006C0EF, 0x03FCE8868000D971,
-    0x8000D97503FD7C87, 0x0002EDA50002EDA6, 0x00026DAB8000D979,
-    0x8000D97D00026DAA, 0x000231AF8000D981, 0x000251AD000259AE,
-    0x03FE8DA8000261AC, 0x03FE9DAA03FE95A9, 0x03FEB1AC03FE95AB,
-    0x03FEA9AE03FEBDAD, 0x8000D96103FEA9AF, 0x8000D9898000D985,
-    0x8000D9918000D98D, 0x8000D9998000D995, 0x03FF25B88000D99D,
-    0x03FF25BA03FF2DB9, 0x03FF25BC03FF3DBB, 0x03FF25BE03FF2DBD,
-    0x03FEA9B003FFCDBF, 0x03FE81B203FEA9B1, 0x03FE81B403FE81B3,
-    0x03FF69B603FE89B5, 0x8000D9A103FF1DB7, 0x8000D9A503FC20E1,
-    0x8000D9A903FC78E3, 0x8000D9B18000D9AD, 0x8000D9B98000D9B5,
-    0x8000D9C18000D9BD, 0x00001D808000D9C5, 0x0000718200007181,
-    0x03FE55B303FD0CEF, 0x8000D9CD8000D9C9, 0x8000D9D1000040BE,
-    0x03FE79B503FE7DB6, 0x03FE2DBB8000D9D5, 0x0000E88603FE2DBA,
-    0x03FE0DBF0000AC87, 0x03FE15BD03FE05BE, 0x03FF95A303FE15BC,
-    0x8000D9DD8000D9D9, 0x8000D9E58000D9E1, 0x03FFB9A503FFBDA6,
-    0x03FFEDAB8000D9E9, 0x8000D9ED03FFEDAA, 0x03FFCDAF8000D9F1,
-    0x03FFD5AD03FFC5AE, 0x8000D9F503FFD5AC, 0x8000D9F903FEF1B3,
-    0x03FE8DB68000D9FD, 0x8000DA058000DA01, 0x000145BA03FE8DB5,
-    0x8000DA0900014DBB, 0x000145BE8000DA0D, 0x000155BC00014DBF,
-    0x03FE2DA800014DBD, 0x03FE3DAA03FE35A9, 0x03FE41AC03FE49AB,
-    0x03FE71AE03FE49AD, 0x8000DA1103FE71AF, 0x8000DA198000DA15,
-    0x8000DA218000DA1D, 0x8000DA298000DA25, 0x000149B88000DA2D,
-    0x00015DBA000155B9, 0x00014DBC000155BB, 0x00017DBE000175BD,
-    0x0001C5B0000175BF, 0x0001C5B20001CDB1, 0x0001C5B40001DDB3,
-    0x0001C5B60001CDB5, 0x03FDB5A300017DB7, 0x8000DA358000DA31,
-    0x8000DA390003C484, 0x03FDC9A503FDC9A6, 0x000209AB8000DA3D,
-    0x8000DA45000201AA, 0x000209AF0002ECBE, 0x000209AD000201AE,
-    0x00004980000211AC, 0x0000558200005581, 0x8000DA49000345A3,
-    0x000345A6000345A5, 0x0004C0868000DA4D, 0x000329AA00031487,
-    0x00033DAC000325AB, 0x000321AE000321AD, 0x8000DA51000315AF,
-    0x8000DA598000DA55, 0x8000DA618000DA5D, 0x8000DA698000DA65,
-    0x00027DA88000DA6D, 0x0003A1AA0003A1A9, 0x0003A1AC0003A1AB,
-    0x000391AE0003A9AD, 0x000381B0000391AF, 0x000385B200038DB1,
-    0x000389B4000399B3, 0x0003B5B60003BDB5, 0x000345B800037DB7,
-    0x000345BA00034DB9, 0x000345BC00035DBB, 0x000345BE00034DBD,
-    0x8000D8F10000F9BF, 0x8000DA7100020DB3, 0x00020DB68000DA75,
-    0x8000DA7D8000DA79, 0x000261BA00020DB5, 0x8000DA8100026DBB,
-    0x000269BE8000DA85, 0x000275BC00025DBF, 0x8000DA89000269BD,
-    0x8000DA918000DA8D, 0x8000DA998000DA95, 0x8000DA9D0001A4E1,
-    0x8000DAA1000340E3, 0x8000DAA98000DAA5, 0x000015800003B8EF,
-    0x0000058200001D81, 0x000280848000DAAD, 0x000518870005C8BE,
-    0x8000DAB500042C86, 0x8000DABD8000DAB9, 0x8000DAC10001A8EF,
-    0x8000DAC503FE74E1, 0x8000DAC903FEF0E3, 0x8000DAD18000DACD,
-    0x8000DAD98000DAD5, 0x8000DAE18000DADD, 0x8000DAE5000191B3,
-    0x000151B6000151B5, 0x8000DAED8000DAE9, 0x00017DBA8000DAF1,
-    0x00016DBC000175BB, 0x0000F5BE0000FDBD, 0x000629A80000E9BF,
-    0x000655AA000655A9, 0x000695AC00068DAB, 0x000695AE00069DAD,
-    0x8000DAB100068DAF, 0x8000DAF98000DAF5, 0x8000DB018000DAFD,
-    0x8000DB098000DB05, 0x00076DB88000DB0D, 0x00070DBA000705B9,
-    0x00071DBC000705BB, 0x000701BE000705BD, 0x0006F5B000073DBF,
-    0x0006F5B20006FDB1, 0x00077DB4000765B3, 0x000761B6000761B5,
-    0x000510E3000755B7, 0x0004F0E18000DB11, 0x00001D808000DB15,
-    0x0000398200000981, 0x8000DB1D8000DB19, 0x0003E0BE0003E884,
-    0x0005CCEF8000DB21, 0x000038878000DB25, 0x00069DA300001886,
-    0x8000DB2D8000DB29, 0x8000DB358000DB31, 0x00065DA500065DA6,
-    0x000679AB8000DB39, 0x8000DB3D000671AA, 0x0007E5AF8000DB41,
-    0x0007F1AD0007F9AE, 0x000691A8000661AC, 0x000691AA000691A9,
-    0x0006B9AC0006ADAB, 0x0006ADAE0006A5AD, 0x8000DB450006A5AF,
-    0x8000DB4D8000DB49, 0x8000DB558000DB51, 0x8000DB5D8000DB59,
-    0x000165B88000DB61, 0x000165BA00016DB9, 0x000165BC00017DBB,
-    0x000165BE00016DBD, 0x0006DDB00001D9BF, 0x0006ADB20006A5B1,
-    0x0006A1B40006A5B3, 0x000699B60006A9B5, 0x000619B3000699B7,
-    0x8000DB698000DB65, 0x8000DB718000DB6D, 0x000631B5000625B6,
-    0x000661BB8000DB75, 0x8000DB79000619BA, 0x00066DBF8000DB7D,
-    0x000671BD000665BE, 0x8000DB81000679BC, 0x8000DB8500065DA3,
-    0x000661A68000DB89, 0x000198848000DB8D, 0x00065DAA000675A5,
-    0x0001A4BE000625AB, 0x000621AE8000DB95, 0x00063DAC000629AF,
-    0x0002C5A9000635AD, 0x0002C5AB0002B1A8, 0x0002CDAD0002C5AA,
-    0x0002F5AF0002C5AC, 0x8000DB990002F5AE, 0x8000DBA18000DB9D,
-    0x8000DBA98000DBA5, 0x8000DBB18000DBAD, 0x0003C9B98000DBB5,
-    0x0003D9BB0003C1B8, 0x0003F9BD0003C1BA, 0x000399BF0003C9BC,
-    0x000345B10003F1BE, 0x000345B3000345B0, 0x000345B5000345B2,
-    0x000345B7000345B4, 0x00034980000345B6, 0x00035D8200034981,
-    0x000CDCBE000245B3, 0x000245B6000245B5, 0x000C00868000DBB9,
-    0x000389BA0003E487, 0x000399BC000389BB, 0x000389BE000399BD,
-    0x000209A3000389BF, 0x8000DBC18000DBBD, 0x8000DBC98000DBC5,
-    0x000209A5000209A6, 0x0003C5AB8000DBCD, 0x8000DBD10003C5AA,
-    0x0003C5AF8000DBD5, 0x0003D5AD0003C5AE, 0x8000DBD90003D5AC,
-    0x8000DBE18000DBDD, 0x000000EF8000DA41, 0x8000DBE98000DBE5,
-    0x0000E8E38000DBED, 0x00018CE18000DBF1, 0x8000DBF98000DBF5,
-    0x8000DC058000DBFD, 0x00006D808000DC09, 0x00007D8200007581,
-    0x000CF08600024084, 0x8000DC0D000D7887, 0x8000DC158000DC11,
-    0x8000DC1D8000DC19, 0x8000DC258000DC21, 0x8000DC2D8000DC29,
-    0x8000DC358000DC31, 0x8000DC3D8000DC39, 0x8000DC458000DC41,
-    0x8000DC4D8000DC49, 0x000CEC840001CCEF, 0x8000DC51000630E1,
-    0x8000DC55000118E3, 0x8000DC5D8000DC59, 0x0001E5B38000DC61,
-    0x000F84848000DC65, 0x8000DC6D8000DC69, 0x0001F5B50001E5B6,
-    0x00017DBB8000DC75, 0x000CA0BE0001D9BA, 0x000121BF8000DC79,
-    0x000131BD000139BE, 0x000DEDA8000139BC, 0x000E2DAA000E25A9,
-    0x000E3DAC000E25AB, 0x000E2DAE000E25AD, 0x8000DC01000E25AF,
-    0x000FE581000FF582, 0x8000DC71000FE980, 0x000098868000DC7D,
-    0x000EADB800030C87, 0x000F4DBA000F45B9, 0x000F45BC000F45BB,
-    0x000F45BE000F4DBD, 0x000E5DB0000F79BF, 0x000EADB2000EF9B1,
-    0x000EBDB4000EA5B3, 0x000EA5B6000EA5B5, 0x8000DC81000E95B7,
-    0x8000DC85000CECEF, 0x8000DC8D8000DC89, 0x8000DC958000DC91,
-    0x0000E8BE8000DC99, 0x8000DCA18000DC9D, 0x8000DCA98000DCA5,
-    0x8000DCAD000DE8E3, 0x000EE5A3000C04E1, 0x8000DCB58000DCB1,
-    0x8000DCBD8000DCB9, 0x000EF5A5000EE5A6, 0x000E7DAB8000DCC1,
-    0x8000DCC5000ED9AA, 0x000E21AF8000DCC9, 0x000E31AD000E39AE,
-    0x000E39A8000E39AC, 0x000E51AA000E39A9, 0x000E71AC000E51AB,
-    0x000E71AE000E71AD, 0x8000DCCD000E71AF, 0x8000DCD58000DCD1,
-    0x000019808000DCD9, 0x0000058200001981, 0x000ED1B88000DCDD,
-    0x000ED1BA000ED1B9, 0x000EE1BC000EE5BB, 0x000EE1BE000EE1BD,
-    0x000E11B0000EE1BF, 0x000E11B2000E11B1, 0x000EF1B4000E11B3,
-    0x000EF1B6000EF1B5, 0x000ED9B3000EF1B7, 0x000028868000DCE5,
-    0x8000DCE90000B887, 0x000EF9B5000EF1B6, 0x000ED5BB8000DCED,
-    0x8000DCF1000ED5BA, 0x000E35BF8000DCF5, 0x000EC5BD000E35BE,
-    0x8000DCF9000EC5BC, 0x8000DCFD000E9DA3, 0x000EB5A68000DD01,
-    0x8000DD098000DD05, 0x000E91AA000EBDA5, 0x8000DD0D000E91AB,
-    0x000E71AE8000DD11, 0x000E81AC000E71AF, 0x000DDDA8000E81AD,
-    0x00023DAA000DE9A9, 0x000299AC00028DAB, 0x000289AE000299AD,
-    0x0004ACBE000289AF, 0x8000DD198000DD15, 0x8000DD1D00032084,
-    0x8000DD258000DD21, 0x000289B88000DD29, 0x000299BA000289B9,
-    0x0002B9BC000291BB, 0x000379BE0002B9BD, 0x0002F9B0000379BF,
-    0x0002CDB20002F9B1, 0x0002DDB40002C5B3, 0x0002C1B60002C5B5,
-    0x0002B5B30002B9B7, 0x8000DD318000DD2D, 0x8000DD398000DD35,
-    0x000211B5000219B6, 0x00023DBB8000DD3D, 0x8000DD4100023DBA,
-    0x000201BF8000DD45, 0x000211BD000219BE, 0x8000DD49000219BC,
-    0x8000DD4D0002F1A3, 0x00025DA68000DD55, 0x8000DD5D8000DD59,
-    0x000279AA000255A5, 0x00054886000279AB, 0x00025DAE0004AC87,
-    0x00025DAC000245AF, 0x000285A8000255AD, 0x000295AA000295A9,
-    0x0002BDAC0002A5AB, 0x0002D1AE0002D5AD, 0x8000DD610002D1AF,
-    0x8000DD698000DD65, 0x0001F1818000DD6D, 0x8000DB9100019980,
-    0x000179B80001F582, 0x0001CDBA000179B9, 0x0001DDBC0001C5BB,
-    0x0001C5BE0001C5BD, 0x0002B5B00001F5BF, 0x000281B20002BDB1,
-    0x000155B4000281B3, 0x000149B600015DB5, 0x8000DD71000149B7,
-    0x8000DD798000DD75, 0x000510840001D4EF, 0x8000DD818000DD7D,
-    0x0005ECBE000E8CEF, 0x8000DD85000EECE1, 0x8000DD89000E1CE3,
-    0x8000DD8D000194E1, 0x00025DB3000E64E3, 0x8000DD958000DD91,
-    0x8000DD9D8000DD99, 0x000275B5000215B6, 0x000239BB8000DDA1,
-    0x8000DDA5000231BA, 0x0001D9BF8000DDA9, 0x000215BD0001D1BE,
-    0x00059DA3000215BC, 0x8000DDAD8000DD51, 0x8000DDB58000DDB1,
-    0x0005B5A50005D5A6, 0x0005F9AB8000DDB9, 0x8000DDBD0005F1AA,
-    0x000619AF8000DDC1, 0x0005D5AD000611AE, 0x000051800005D5AC,
-    0x0000618200005981, 0x8000DDC5000695B3, 0x000771B6000771B5,
-    0x000340868000DDC9, 0x0007EDBA00035487, 0x0007E5BC0007E5BB,
-    0x0007D1BE0007EDBD, 0x8000DDCD0007CDBF, 0x8000DDD58000DDD1,
-    0x8000DDDD8000DDD9, 0x000440EF8000DDE1, 0x000730E18000DDE5,
-    0x000494E38000DDE9, 0x8000DDF18000DDED, 0x8000DDF98000DDF5,
-    0x000619A38000DDFD, 0x8000DE058000DE01, 0x8000DE0D8000DE09,
-    0x0007FDA50007FDA6, 0x000769AB8000DE11, 0x8000DE15000761AA,
-    0x000741AF8000DE19, 0x000761AD00075DAE, 0x0007CDA8000769AC,
-    0x0007D1AA0007D1A9, 0x000635AC00062DAB, 0x000635AE00063DAD,
-    0x8000DE1D00069DAF, 0x8000DE258000DE21, 0x000019808000DE29,
-    0x0000058200001981, 0x000689B88000DE2D, 0x000699BA000689B9,
-    0x0006B9BC000691BB, 0x000151BE0006B9BD, 0x0006E5B0000151BF,
-    0x0006FDB20006EDB1, 0x0006D9B40006F5B3, 0x0006C1B60006C5B5,
-    0x0001D1A80006B9B7, 0x000109AA0001D9A9, 0x000119AC000109AB,
-    0x000109AE000119AD, 0x00016084000109AF, 0x0000A88700017CBE,
-    0x8000DE3500018C86, 0x8000DE3D8000DE39, 0x000109B88000DE41,
-    0x000119BA000109B9, 0x000139BC000111BB, 0x0000F9BE000139BD,
-    0x00017DB00000F9BF, 0x000145B2000141B1, 0x000145B400015DB3,
-    0x000145B600014DB5, 0x8000DE45000139B7, 0x8000DE4D8000DE49,
-    0x8000DE5100028DB3, 0x000295B600029DB5, 0x8000DE598000DE55,
-    0x0002B5BA8000DE5D, 0x00029DBC000289BB, 0x00028DBE000285BD,
-    0x0002CDA6000285BF, 0x8000DE658000DE61, 0x8000DE690002C5A5,
-    0x8000DE6D0002D5A3, 0x0002D5AE8000DE71, 0x0002C5AC0002DDAF,
-    0x0002EDAA0002DDAD, 0x0003F4840002D1AB, 0x000331A88000DE75,
-    0x000331AA000331A9, 0x000091AC000331AB, 0x000091AE000091AD,
-    0x8000DE7900008DAF, 0x8000DE818000DE7D, 0x8000DE898000DE85,
-    0x8000DE918000DE8D, 0x0000BDB88000DE95, 0x00006DBA000065B9,
-    0x00007DBC000065BB, 0x00006DBE000065BD, 0x0000FDB0000065BF,
-    0x0000CDB20000C5B1, 0x0000B9B40000A9B3, 0x0000A1B60000A5B5,
-    0x0001BD800000A1B7, 0x0000198200000981, 0x8000DE9D8000DE99,
-    0x001D1487000258BE, 0x001D6C84001D9C86, 0x8000DEA58000DCE1,
-    0x001CBCBE8000DEA9, 0x8000DEB18000DEAD, 0x0002F9B38000DEB5,
-    0x8000DEBD8000DEB9, 0x8000DEC58000DEC1, 0x000159B5000151B6,
-    0x000141BB001FDCBE, 0x8000DEC9000179BA, 0x000121BF8000DECD,
-    0x00013DBD00013DBE, 0x000770E1000159BC, 0x000608E38000DED1,
-    0x0000CCEF8000DED5, 0x8000DEDD8000DED9, 0x000010E38000DEE1,
-    0x0001D0E18000DEE5, 0x001C90868000DEE9, 0x0006E0EF001CF487,
-    0x000279A38000DEED, 0x8000DEF58000DEF1, 0x8000DEFD8000DEF9,
-    0x0001D9A50001D1A6, 0x0001C1AB8000DF01, 0x8000DF050001F9AA,
-    0x0001A1AF8000DF09, 0x0001BDAD0001BDAE, 0x8000DEA10001D9AC,
-    0x001FE58100001582, 0x8000DF0D001FE580, 0x8000DF158000DF11,
-    0x001F09A88000DF19, 0x001F1DAA001F09A9, 0x001F0DAC001F15AB,
-    0x001F71AE001F71AD, 0x001F11B0001F71AF, 0x001F25B2001F2DB1,
-    0x001FDDB4001FC9B3, 0x001FC1B6001FC1B5, 0x001FC5B8001FC1B7,
-    0x001FD5BA001FC9B9, 0x001FB9BC001FA9BB, 0x001FA9BE001FB9BD,
-    0x001FB5B3001FA1BF, 0x8000DF218000DF1D, 0x8000DF298000DF25,
-    0x001FA5B5001FD1B6, 0x001FC9BB8000DF2D, 0x8000DF31001FF5BA,
-    0x001F25BF8000DF35, 0x001F3DBD001F39BE, 0x8000DF39001FD1BC,
-    0x8000DF418000DF3D, 0x8000DF498000DF45, 0x8000DF4D001F60E1,
-    0x8000DF51001CB4E3, 0x8000DF598000DF55, 0x8000DF5D001D00EF,
-    0x8000DF658000DF61, 0x001E35A38000DF69, 0x8000DF718000DF6D,
-    0x8000DF798000DF75, 0x001E25A5001E51A6, 0x001E49AB8000DF7D,
-    0x0002A884001E75AA, 0x001EA5AF8000DF81, 0x001EBDAD001EB9AE,
-    0x00004D80001E51AC, 0x0000558200005581, 0x8000DF850001C9B3,
-    0x0001C9B60001D9B5, 0x0000A0868000DF89, 0x0001C5BA00010487,
-    0x0001B5BC0001ADBB, 0x0001B5BE0001BDBD, 0x000199A80001ADBF,
-    0x00010DAA000199A9, 0x00011DAC000105AB, 0x000105AE000105AD,
-    0x8000DF8D000135AF, 0x8000DF958000DF91, 0x8000DF9D8000DF99,
-    0x8000DFA58000DFA1, 0x000125B88000DFA9, 0x000139BA00012DB9,
-    0x000129BC000139BB, 0x0000DDBE000129BD, 0x00014DB00000D5BF,
-    0x00012DB2000125B1, 0x00013DB4000125B3, 0x000121B6000125B5,
-    0x8000DFAD00011DB7, 0x8000DFB58000DFB1, 0x8000DFB9000289A3,
-    0x000289A6000299A5, 0x8000DFC18000DFBD, 0x000285AA8000DFC5,
-    0x0002F5AC0002EDAB, 0x0002F5AE0002FDAD, 0x8000DFC90002EDAF,
-    0x8000DFD18000DFCD, 0x8000DFD500024084, 0x8000DFDD8000DFD9,
-    0x00000D808000DFE1, 0x00001D8200001581, 0x8000DFE98000DFE5,
-    0x0003B4878000DFED, 0x0003CCBE0004DC86, 0x8000DFF98000DFF5,
-    0x0002E8EF8000DFFD, 0x8000E0058000E001, 0x0002FCE38000E009,
-    0x0001D0E18000E00D, 0x8000E0158000E011, 0x8000E01D8000E019,
-    0x000361B38000E021, 0x8000E0298000E025, 0x8000E0318000E02D,
-    0x000371B5000379B6, 0x00035DBB8000E035, 0x8000E03900035DBA,
-    0x0000E1BF8000E03D, 0x0000FDBD0000FDBE, 0x0002B9B00000FDBC,
-    0x000109B20002B9B1, 0x000119B4000109B3, 0x000105B6000105B5,
-    0x000105B800013DB7, 0x00016DBA000165B9, 0x000161BC000165BB,
-    0x000161BE000161BD, 0x00075C85000161BF, 0x8000E0458000E041,
-    0x8000DE318000E049, 0x8000E0518000E04D, 0x000231A88000E055,
-    0x0002C9AA000239A9, 0x0002D9AC0002C9AB, 0x0002C9AE0002D9AD,
-    0x0005CC840002C9AF, 0x8000E059001E80E1, 0x001E38E1001EBCE3,
-    0x0000A0E38000E05D, 0x8000E065000410BE, 0x001E0CEF8000E069,
-    0x8000E0718000E06D, 0x8000E0798000E075, 0x000261A3001E74EF,
-    0x0000418100005182, 0x8000E07D00009180, 0x000271A5000279A6,
-    0x00025DAB8000E081, 0x0004C88600025DAA, 0x0001E1AF00053C87,
-    0x0001FDAD0001FDAE, 0x000685A80001FDAC, 0x000685AA00068DA9,
-    0x000689AC000699AB, 0x0006BDAE000689AD, 0x8000E0610006B1AF,
-    0x8000E0898000E085, 0x8000E0918000E08D, 0x8000E0998000E095,
-    0x00069DB88000E09D, 0x0006A5BA0006ADB9, 0x000719BC000709BB,
-    0x00070DBE000719BD, 0x0006D1B0000705BF, 0x0006D1B20006D1B1,
-    0x0006B5B40006D1B3, 0x0006B5B60006BDB5, 0x00060DB30006ADB7,
-    0x8000E0A58000E0A1, 0x8000E0AD8000E0A9, 0x000601B5000609B6,
-    0x000615BB8000E0B1, 0x8000E0B5000615BA, 0x000679BF8000E0B9,
-    0x000605BD000671BE, 0x8000E0BD000605BC, 0x8000E0C10004A0E1,
-    0x00000D8000055CE3, 0x00003D8200003581, 0x8000E0C98000E0C5,
-    0x000360848000E0CD, 0x0001F0BF00006CBE, 0x8000E0D100009085,
-    0x00078DA3000598EF, 0x000C008600000884, 0x8000E0D500008C87,
-    0x000781A5000789A6, 0x000795AB8000E0D9, 0x8000E0DD000795AA,
-    0x0007F9AF8000E0E1, 0x000785AD0007F1AE, 0x8000E0E5000785AC,
-    0x8000E0E90006E9B3, 0x0006E9B68000E0ED, 0x8000E0F58000E0F1,
-    0x000145BA0006E9B5, 0x8000E0F900014DBB, 0x000145BE8000E0FD,
-    0x000155BC00014DBF, 0x000621A800014DBD, 0x000625AA000621A9,
-    0x000625AC00063DAB, 0x000649AE00062DAD, 0x8000E101000641AF,
-    0x8000E1098000E105, 0x8000E1118000E10D, 0x8000E1198000E115,
-    0x000149B88000E11D, 0x000159BA000149B9, 0x000179BC000151BB,
-    0x000119BE000179BD, 0x0001C5B0000115BF, 0x0001C5B20001CDB1,
-    0x0001C5B40001DDB3, 0x0001C5B60001CDB5, 0x8000E121000179B7,
-    0x8000E1298000E125, 0x8000E12D0005A1A3, 0x0005A1A60005A1A5,
-    0x03EA218F8000E131, 0x00020DAA8000E135, 0x00021DAC000205AB,
-    0x00020DAE000205AD, 0x0012ED97000205AF, 0x0011459500126596,
-    0x0016719300169D94, 0x03EB799100155592, 0x0006619F03EA5990,
-    0x001A459D0005359E, 0x0019559B001A699C, 0x001E5999001E459A,
-    0x000060E1001D4598, 0x003E4CE38000E139, 0x0002B1A38000E13D,
-    0x0007B5A10001D5A2, 0x003801B1000689A0, 0x003A95B3003E09B0,
-    0x0026B5B5003A8DB2, 0x003A68EF002401B4, 0x0036C9A9000C30BE,
-    0x003001AB0036C1A8, 0x0033CDAD0037E1AA, 0x003EF9AF0032F5AC,
-    0x000219A1003C01AE, 0x000E6DA38000E145, 0x000AF5A5000E6DA2,
-    0x000B85A7000801A4, 0x000D0086000A99A6, 0x03EB0982000D4487,
-    0x03EA318403EB0983, 0x0017398600141585, 0x000DA0840017CD87,
-    0x0010258A8000E149, 0x00139D8C00130D8B, 0x001F198E001C0D8D,
-    0x8000DFF1001F0D8F, 0x00186D92000300EF, 0x001B7D94001B4993,
-    0x0007599600040595, 0x8000E14D00074997, 0x0006459A8000E151,
-    0x0003519C00002D9B, 0x8000E155000060E3, 0x0001A58000006CE1,
-    0x0001058200010181, 0x8000E159000C00BE, 0x8000E1618000E15D,
-    0x8000E165000188EF, 0x8000E169000654E1, 0x8000E16D000154E3,
-    0x8000E1758000E171, 0x0002E9B38000E179, 0x8000E1818000E17D,
-    0x8000E1898000E185, 0x000299B500029DB6, 0x000289BB8000E18D,
-    0x8000E1910002BDBA, 0x000259BF8000E195, 0x000259BD000251BE,
-    0x000D29A3000291BC, 0x8000E19D8000E199, 0x8000E1A58000E1A1,
-    0x000D59A5000D5DA6, 0x000D49AB8000E1A9, 0x8000E1AD000D7DAA,
-    0x000D99AF8000E1B5, 0x000D99AD000D91AE, 0x00005180000D51AC,
-    0x0000618200005981, 0x8000E1B9000F2DB3, 0x000FC9B6000F2DB5,
-    0x000340868000E1BD, 0x000FC9BA00030887, 0x000FC1BC000FC5BB,
-    0x000FC1BE000FC1BD, 0x8000E141000FC1BF, 0x8000E1C18000E1B1,
-    0x8000E1C98000E1C5, 0x8000E1D18000E1CD, 0x000E45A88000E1D5,
-    0x000FC9AA000F81A9, 0x000FC9AC000FC9AB, 0x000F2DAE000F25AD,
-    0x000F61B0000F25AF, 0x000F79B2000F6DB1, 0x000F69B4000F79B3,
-    0x000F1DB6000F69B5, 0x000F2DB8000F15B7, 0x000F35BA000F35B9,
-    0x000F1DBC000F05BB, 0x0000F1BE0000F1BD, 0x000E61A30000F1BF,
-    0x0001C4848000E1D9, 0x8000E1E18000E1DD, 0x000E61A5000E85A6,
-    0x000E89AB8000E1E5, 0x8000E1E9000E85AA, 0x000E8DAF8000E1ED,
-    0x000E8DAD000E8DAE, 0x8000E1F1000E8DAC, 0x8000E1F98000E1F5,
-    0x8000E2018000E1FD, 0x8000E2098000E205, 0x8000E2118000E20D,
-    0x00001D8100001D82, 0x8000E21500001D80, 0x8000E21D8000E219,
-    0x0001ADA80001B4BE, 0x0001D5AA0001D5A9, 0x00011DAC000105AB,
-    0x00010DAE000105AD, 0x00018086000105AF, 0x8000E22500011887,
-    0x8000E22D8000E229, 0x8000E2358000E231, 0x000085B88000E239,
-    0x000085BA00008DB9, 0x000085BC00009DBB, 0x000085BE00008DBD,
-    0x00017DB00000BDBF, 0x0000E5B20000E1B1, 0x0000EDB40000FDB3,
-    0x0000D5B60000DDB5, 0x00025DB30000BDB7, 0x8000E2418000E23D,
-    0x8000E2498000E245, 0x000295B5000285B6, 0x000325BB8000E24D,
-    0x8000E2510002B9BA, 0x000319BF8000E255, 0x000329BD000329BE,
-    0x0004CCBE000331BC, 0x8000E259000219A3, 0x0002C1A68000E25D,
-    0x8000E2658000E261, 0x0002FDAA0002D1A5, 0x8000E269000361AB,
-    0x00036DAE8000E26D, 0x000375AC00035DAF, 0x000329A800036DAD,
-    0x000339AA000329A9, 0x000329AC000339AB, 0x00009DAE000329AD,
-    0x8000E271000095AF, 0x8000E2798000E275, 0x8000E2818000E27D,
-    0x0001AD810001A982, 0x000099B80001AD80, 0x0000A5BA0000ADB9,
-    0x000075BC00006DBB, 0x000075BE00007DBD, 0x0000EDB000006DBF,
-    0x0000F5B20000F5B1, 0x0000DDB40000C5B3, 0x0000B5B60000BDB5,
-    0x000178E10000A9B7, 0x000010E3000E1CE1, 0x8000E285000E3CE3,
-    0x000414BE8000E289, 0x0002AC848000E28D, 0x000578878000E295,
-    0x8000E29900050C86, 0x0000BCEF8000E29D, 0x0002F1B3000E48EF,
-    0x8000E2A58000E2A1, 0x8000E2AD8000E2A9, 0x0002E1B50002E9B6,
-    0x000171BB8000E2B1, 0x8000E2B5000169BA, 0x000139BF0004A084,
-    0x000151BD000159BE, 0x8000E2B9000161BC, 0x8000E2BD00048C84,
-    0x8000E2C100380084, 0x8000E2C98000E2C5, 0x000E89AA8000E2CD,
-    0x000EB9A8000E89AB, 0x000EFDAE000EB1A9, 0x000EF9AC000EE1AF,
-    0x000EADB2000EF5AD, 0x000E91B0000E65B3, 0x000E7DB6000EA5B1,
-    0x000E7DB4000E65B7, 0x000E5DBA000E75B5, 0x000E5DB8000DE5BB,
-    0x000DE1BE000E55B9, 0x000DFDBC000DE9BF, 0x0005B1A3000DF5BD,
-    0x8000E2D18000E291, 0x8000E2D98000E2D5, 0x0005A1A50005A9A6,
-    0x000631AB8000E2DD, 0x8000E2E1000629AA, 0x000679AF8000E2E5,
-    0x000611AD000619AE, 0x8000E2E9000621AC, 0x8000E2F18000E2ED,
-    0x00001D808000E2F5, 0x0000398200000981, 0x8000E2FD8000E2F9,
-    0x0000C8868000E301, 0x8000E30500030C87, 0x8000E30D8000E309,
-    0x0007A5A88000E311, 0x0007BDAA0007B5A9, 0x0007D9AC0007C9AB,
-    0x0007C9AE0007D9AD, 0x0000A0BE0007C1AF, 0x8000E3198000E315,
-    0x8000E3218000E31D, 0x8000E3298000E325, 0x0000CDB88000E32D,
-    0x0000D5BA0000D5B9, 0x0000FDBC0000E5BB, 0x00009DBE000095BD,
-    0x000789B0000095BF, 0x000761B2000765B1, 0x000765B400077DB3,
-    0x000765B600076DB5, 0x000635B30000F5B7, 0x8000E3358000E331,
-    0x8000E33D8000E339, 0x000651B5000659B6, 0x0006A1BB8000E341,
-    0x8000E34500064DBA, 0x0006A9BF8000E349, 0x0006A9BD0006A1BE,
-    0x8000E34D0006B5BC, 0x0005F0EF8000E351, 0x8000E3598000E355,
-    0x8000E3618000E35D, 0x00003D808000E365, 0x00001D8200000981,
-    0x0006C8E18000E369, 0x000448E38000E36D, 0x0006B5A38000E371,
-    0x000028868000E375, 0x8000E37900014087, 0x0006D1A50006D9A6,
-    0x000621AB8000E37D, 0x8000E3810006CDAA, 0x000629AF8000E385,
-    0x000629AD000621AE, 0x8000E389000635AC, 0x8000E38D0001BDB3,
-    0x000179B68000E391, 0x8000E3998000E395, 0x000155BA000179B5,
-    0x8000E39D00015DBB, 0x0000F9BE8000E3A1, 0x000145BC0000F9BF,
-    0x000271A80000F9BD, 0x000271AA000271A9, 0x0002B5AC000271AB,
-    0x0002B5AE0002BDAD, 0x003CACBE0002ADAF, 0x8000E3A98000E3A5,
-    0x8000E3B18000E3AD, 0x8000E3B98000E3B5, 0x000369B88000E3BD,
-    0x000309BA000369B9, 0x00031DBC000309BB, 0x00030DBE000305BD,
-    0x0002D5B0000305BF, 0x0002D5B20002DDB1, 0x000379B4000369B3,
-    0x000369B6000379B5, 0x8000E3C1000361B7, 0x8000E3C98000E3C5,
-    0x8000E3CD0002F5A3, 0x000231A6000231A5, 0x8000E3D58000E3D1,
-    0x00021DAA8000E3D9, 0x00020DAC000215AB, 0x0003B1AE0003B1AD,
-    0x000218EF0003B1AF, 0x00006D8100001582, 0x8000E3DD00006D80,
-    0x003CF8868000E3E5, 0x8000E3E900031887, 0x8000E3F18000E3ED,
-    0x00076CE38000E3F5, 0x000168E18000E221, 0x000285A88000E3F9,
-    0x000295AA000295A9, 0x0002BDAC0002A5AB, 0x0002D1AE0002D5AD,
-    0x8000E3FD0002D1AF, 0x8000E4058000E401, 0x8000E40D8000E409,
-    0x8000E4158000E411, 0x000175B88000E419, 0x000175BA00017DB9,
-    0x0001D5BC0001CDBB, 0x0001C9BE0001DDBD, 0x0002B5B00001C1BF,
-    0x000281B20002BDB1, 0x000155B4000281B3, 0x000155B600015DB5,
-    0x0006B4E100014DB7, 0x0006F4E38000E41D, 0x003D18848000E421,
-    0x8000E4298000E425, 0x8000E4318000E42D, 0x8000E4398000E435,
-    0x8000E4418000E43D, 0x000658EF8000E445, 0x00007D818000E449,
-    0x8000E44D00007180, 0x8000E45500000582, 0x000180EF8000E459,
-    0x000190E1003C54BE, 0x00067CE38000E45D, 0x8000E4658000E461,
-    0x003CD8868000E469, 0x003D9DA3003CF487, 0x8000E4518000E3E1,
-    0x8000E4718000E46D, 0x003DB5A5003DD5A6, 0x003DF9AB8000E475,
-    0x8000E479003DF1AA, 0x003E19AF8000E47D, 0x003DD5AD003E11AE,
-    0x8000E481003DD5AC, 0x8000E485003EA1B3, 0x003EADB68000E489,
-    0x8000E4918000E48D, 0x003F49BA003EB1B5, 0x8000E495003F49BB,
-    0x003F49BE8000E499, 0x003F49BC003F49BF, 0x003E55A8003F49BD,
-    0x003E6DAA003E65A9, 0x003E7DAC003E65AB, 0x003FB9AE003E69AD,
-    0x8000E49D003FB9AF, 0x8000E4A58000E4A1, 0x8000E4AD8000E4A9,
-    0x8000E4B58000E4B1, 0x003F55B88000E4B9, 0x003F55BA003F5DB9,
-    0x003F75BC003F6DBB, 0x003F75BE003F7DBD, 0x003FC9B0003F6DBF,
-    0x003FD9B2003FC9B1, 0x003FC9B4003FD9B3, 0x003F7DB6003FC9B5,
-    0x00001582003F71B7, 0x0001B180003FE1A3, 0x003FEDA60001B181,
-    0x0001D0BE8000E4BD, 0x003E09AA003FF1A5, 0x0001E484003E09AB,
-    0x003E09AE8000E4C1, 0x003E09AC003E09AF, 0x003CDDB3003E09AD,
-    0x0000E8868000E4C5, 0x8000E4C90001CC87, 0x003CF5B5003C95B6,
-    0x003CB9BB8000E4CD, 0x8000E4D1003CB1BA, 0x003F59BF8000E4D5,
-    0x003C95BD003F51BE, 0x003E51A8003C95BC, 0x003E6DAA003E51A9,
-    0x003E61AC003E61AB, 0x000195AE003E61AD, 0x0001A08400018DAF,
-    0x8000E4DD8000E4D9, 0x8000E4E58000E4E1, 0x8000E4ED8000E4E9,
-    0x0001A9B88000E4F1, 0x000169BA0001A9B9, 0x000179BC000169BB,
-    0x000169BE000179BD, 0x0001FDB0000169BF, 0x0001C5B20001CDB1,
-    0x0001B9B40001ADB3, 0x0001ADB60001B9B5, 0x003DE5B30001A5B7,
-    0x8000E4F98000E4F5, 0x8000E5018000E4FD, 0x003DE9B5003DE1B6,
-    0x000209BB8000E505, 0x8000E509000239BA, 0x00027DBF8000E50D,
-    0x000279BD00027DBE, 0x8000E511000211BC, 0x8000E515003DA1A3,
-    0x003DA5A68000E519, 0x8000E5218000E51D, 0x00027DAA003DADA5,
-    0x8000E52500024DAB, 0x000239AE8000E529, 0x000255AC000239AF,
-    0x0000E98000023DAD, 0x00001D820000E981, 0x0002E4EF0003C0BE,
-    0x000354878000E52D, 0x000110E300043C86, 0x0001F8E18000E535,
-    0x8000E53D8000E539, 0x8000E5458000E541, 0x8000E54D8000E549,
-    0x8000E5558000E551, 0x8000E559000391B3, 0x00037DB60003B9B5,
-    0x8000E5618000E55D, 0x000359BA8000E565, 0x000349BC000359BB,
-    0x0000FDBE000349BD, 0x000245A80000F5BF, 0x000255AA000255A9,
-    0x00027DAC000265AB, 0x0002B1AE0002B1AD, 0x00056C840002B1AF,
-    0x8000E56D8000E569, 0x8000E5758000E571, 0x000570BE8000E579,
-    0x00015DB88000E57D, 0x000165BA00016DB9, 0x000119BC00010DBB,
-    0x00010DBE000119BD, 0x0002D1B0000105BF, 0x0002D1B20002D1B1,
-    0x000175B40002D1B3, 0x000175B600017DB5, 0x003F40E100016DB7,
-    0x003E08E30000BCE3, 0x8000E581003E6CE1, 0x8000E5898000E585,
-    0x8000E5918000E58D, 0x8000E5998000E595, 0x00076CBE8000E59D,
-    0x003E9CEF000054EF, 0x00029DA38000E5A5, 0x0000618100006982,
-    0x8000E5A900006980, 0x0002B5A5000271A6, 0x000255AB8000E5AD,
-    0x0004C886000255AA, 0x0001F9AF0004EC87, 0x000245AD0001F1AE,
-    0x0006A5A8000245AC, 0x0006B9AA0006A9A9, 0x0006A9AC0006B9AB,
-    0x0006D9AE0006A9AD, 0x8000E5A10006D9AF, 0x8000E5B58000E5B1,
-    0x8000E5BD8000E5B9, 0x8000E5C58000E5C1, 0x000771B88000E5C9,
-    0x000775BA000771B9, 0x0007C5BC0007DDBB, 0x0007C5BE0007CDBD,
-    0x0006A9B00007FDBF, 0x0006B5B20006B5B1, 0x000695B400068DB3,
-    0x000751B6000751B5, 0x000631B3000751B7, 0x8000E5D18000E5CD,
-    0x8000E5D98000E5D5, 0x000621B5000629B6, 0x000671BB8000E5DD,
-    0x8000E5E100066DBA, 0x000795BF8000E5E5, 0x000659BD000651BE,
-    0x8000E5E9000661BC, 0x8000E5ED000675A3, 0x00066DA68000E5F1,
-    0x8000E5F98000E5F5, 0x000629AA000665A5, 0x8000E5FD000635AB,
-    0x000615AE8000E601, 0x000625AC0007D1AF, 0x00000D8000061DAD,
-    0x00001D8200001581, 0x8000E6098000E605, 0x0001DC848000E60D,
-    0x000080868000E611, 0x8000E61500012087, 0x8000E61D8000E619,
-    0x8000E6258000E621, 0x8000E62D8000E629, 0x8000E631000474E3,
-    0x8000E6350005C8E1, 0x8000E63D8000E639, 0x8000E6458000E641,
-    0x8000E64D8000E649, 0x8000E6558000E651, 0x8000E6590004B4EF,
-    0x00063DA88000E65D, 0x000655AA000655A9, 0x0001B9AC0001A9AB,
-    0x0001A9AE0001B9AD, 0x0001AC840001A9AF, 0x8000E6658000E661,
-    0x8000E66D8000E669, 0x8000E6758000E671, 0x00016DB88000E679,
-    0x000101BA000105B9, 0x000105BC000101BB, 0x000131BE00010DBD,
-    0x0001D9B0000131BF, 0x00016DB20001D9B1, 0x00017DB4000165B3,
-    0x000165B6000165B5, 0x0003BD81000155B7, 0x0005D5B30003BD80,
-    0x000239B500001982, 0x000354BE8000E67D, 0x8000E685000231B6,
-    0x000215BB8000E689, 0x0002B9BD000215BA, 0x0002A9BF0002B1BC,
-    0x8000E68D0002B1BE, 0x000261A5000269A6, 0x00058DA3000C0084,
-    0x000CF8868000E691, 0x0002F1AF00037C87, 0x0002E1AD0002E9AE,
-    0x00024DAB0002E9AC, 0x8000E69500024DAA, 0x8000E69D8000E699,
-    0x8000E6A58000E6A1, 0x000120E38000E6A9, 0x0001A0E18000E6AD,
-    0x000240EF8000E6B1, 0x8000E6B98000E6B5, 0x8000E6C18000E6BD,
-    0x8000E6C98000E6C5, 0x0003F1B38000E6CD, 0x8000E5318000E6D1,
-    0x8000E6D98000E6D5, 0x0003E1B50003E9B6, 0x0003B5BB8000E6DD,
-    0x8000E6E10003B5BA, 0x000395BF8000E6E5, 0x0003A5BD000395BE,
-    0x000229A80003A5BC, 0x000239AA000229A9, 0x000229AC000239AB,
-    0x000259AE000229AD, 0x0001CD80000255AF, 0x0000198200000981,
-    0x8000E6ED8000E6E9, 0x000DB487000D7CBE, 0x000231B8000C1C86,
-    0x000235BA00023DB9, 0x0002F9BC0002E9BB, 0x0002E9BE0002F9BD,
-    0x000231B00002E9BF, 0x000231B2000231B1, 0x000211B4000231B3,
-    0x000211B6000211B5, 0x8000E6F5000211B7, 0x8000E6FD8000E6F9,
-    0x8000E7058000E701, 0x8000E70D8000E709, 0x8000E7110006D4EF,
-    0x8000E715000654E1, 0x00150CAC0000A4E3, 0x8000E71D8000E719,
-    0x0002F1A38000E721, 0x8000E7298000E725, 0x8000E7318000E72D,
-    0x0002E1A50002E9A6, 0x0002B5AB8000E735, 0x8000E7390002B5AA,
-    0x000295AF8000E73D, 0x0002A5AD000295AE, 0x000E21A80002A5AC,
-    0x000E49AA000E21A9, 0x000E69AC000E59AB, 0x000E99AE000E69AD,
-    0x8000E6F1000E99AF, 0x8000E7458000E741, 0x8000E74D8000E749,
-    0x8000E7558000E751, 0x000EF5B88000E759, 0x000EF5BA000EFDB9,
-    0x000E9DBC000E89BB, 0x000E85BE000E85BD, 0x000EE9B0000EB5BF,
-    0x000EFDB2000EE9B1, 0x000ED5B4000EF1B3, 0x000ED5B6000EDDB5,
-    0x000EC1B3000ECDB7, 0x0000B58100001582, 0x8000E75D0000B580,
-    0x000EE1B5000EE9B6, 0x000E2DBB000010BE, 0x00036484000E2DBA,
-    0x000E19BF8000E761, 0x000E25BD000E11BE, 0x8000E765000E29BC,
-    0x00008886000E85A3, 0x000EADA600036C87, 0x8000E76D8000E769,
-    0x000E69AA000EA5A5, 0x8000E771000E69AB, 0x000E55AE8000E775,
-    0x000E6DAC000E5DAF, 0x000E89B3000E61AD, 0x8000E77D8000E779,
-    0x8000E7858000E781, 0x000E89B5000E81B6, 0x000E95BB8000E789,
-    0x8000E78D000E8DBA, 0x000EF9BF8000E791, 0x000E85BD000EF1BE,
-    0x8000E795000E85BC, 0x8000E79D8000E799, 0x000D0CE38000E7A1,
-    0x000D18E18000E7A5, 0x000CACEF8000E7A9, 0x8000E7B18000E7AD,
-    0x8000E7B98000E7B5, 0x8000E7C18000E7BD, 0x000E01A88000E7C5,
-    0x000E01AA000E01A9, 0x000E01AC000E01AB, 0x000E01AE000E01AD,
-    0x0000DD80000E3DAF, 0x0000198200000981, 0x8000E7CD8000E7C9,
-    0x000074BE00013C84, 0x000E1DB88000E7D5, 0x000E25BA000E2DB9,
-    0x0001F9BC0001E9BB, 0x0001E9BE0001F9BD, 0x000E49B00001E9BF,
-    0x000E51B2000E51B1, 0x000E35B4000E51B3, 0x000E35B6000E3DB5,
-    0x000D89A3000E2DB7, 0x0004AC868000E7D9, 0x8000E7DD00033C87,
-    0x000D89A5000D81A6, 0x000D95AB8000E7E1, 0x8000E7E5000D8DAA,
-    0x000DF9AF8000E7E9, 0x000D85AD000DF1AE, 0x8000E7ED000D85AC,
-    0x0003488400029DB3, 0x000355B6000440BE, 0x8000E7F58000E7F1,
-    0x000371BA0002B1B5, 0x8000E7F9000379BB, 0x000331BE8000E7FD,
-    0x000351BC000331BF, 0x000391B0000351BD, 0x0003A1B2000399B1,
-    0x0003D1B40003A1B3, 0x0003D1B60003D1B5, 0x0003F5B80003D1B7,
-    0x0003CDBA0003F9B9, 0x0003DDBC0003C5BB, 0x0003CDBE0003C5BD,
-    0x8000E8010003C5BF, 0x8000E8098000E805, 0x00197C858000E80D,
-    0x8000E8158000E811, 0x000221A88000E681, 0x000231AA000231A9,
-    0x00021DAC000205AB, 0x0003F1AE0003F1AD, 0x8000E8190003F1AF,
-    0x8000E8218000E81D, 0x000050EF8000E825, 0x8000E82D8000E829,
-    0x000084E38000E831, 0x0001F8E18000E835, 0x000015808000E839,
-    0x0000058200001981, 0x000399A38000E83D, 0x000468868000E845,
-    0x8000E84900056087, 0x0003B5A5000251A6, 0x00027DAB8000E84D,
-    0x8000E851000275AA, 0x000235AF8000E855, 0x000255AD000235AE,
-    0x8000E859000255AC, 0x8000E8618000E85D, 0x8000E8698000E865,
-    0x8000E8718000E86D, 0x00046CBE0001F8EF, 0x8000E875000E80E1,
-    0x8000E879000114E3, 0x8000E8818000E87D, 0x8000E8898000E885,
-    0x8000E8918000E88D, 0x8000E8950001DDB3, 0x0001F5B60001FDB5,
-    0x8000E89D8000E899, 0x0001B1BA8000E8A1, 0x00019DBC000185BB,
-    0x00013DBE000135BD, 0x000605A9000135BF, 0x000615AB0005B9A8,
-    0x00063DAD00061DAA, 0x00064DAF00063DAC, 0x8000E84100065DAE,
-    0x00001D8100001D82, 0x8000E8A500001D80, 0x8000E8AD8000E8A9,
-    0x000741B98000E8B1, 0x000741BB00069DB8, 0x000745BD000749BA,
-    0x000745BF000759BC, 0x000609B1000745BE, 0x0006A9B300063DB0,
-    0x0006B9B5000601B2, 0x0006ADB70006B1B4, 0x000691A30006B1B6,
-    0x0000288600028C84, 0x8000E8B50003C087, 0x0006B1A50006B9A6,
-    0x0006C9AB8000E8B9, 0x8000E8BD0006FDAA, 0x000679AF8000E8C1,
-    0x000679AD000671AE, 0x8000E8C50006D1AC, 0x8000E8C9000799B3,
-    0x000751B68000E8CD, 0x8000E8D58000E8D1, 0x00076DBA0007B1B5,
-    0x8000E8D9000775BB, 0x000759BE8000E8DD, 0x00076DBC000745BF,
-    0x8000E8E1000765BD, 0x8000E8E98000E8E5, 0x8000E8F18000E8ED,
-    0x8000E8F98000E8F5, 0x8000E8FD0005A8EF, 0x8000E901000534E1,
-    0x8000E905000574E3, 0x8000E90D8000E909, 0x00061DA38000E911,
-    0x0000158100002D82, 0x8000E91500001D80, 0x000635A50006D5A6,
-    0x0006F1AB8000E919, 0x8000E91D0006E9AA, 0x0006C1AF00012884,
-    0x0006E1AD0006DDAE, 0x0006C5A80006E9AC, 0x0006D5AA0006D5A9,
-    0x0006FDAC0006E5AB, 0x00061DAE000615AD, 0x0001ECBE000615AF,
-    0x000008868000E925, 0x8000E92900002087, 0x8000E9318000E92D,
-    0x00067DB88000E935, 0x00060DBA000605B9, 0x000601BC000605BB,
-    0x000639BE000609BD, 0x00066DB0000639BF, 0x00067DB2000675B1,
-    0x000659B4000675B3, 0x00064DB6000645B5, 0x000291A8000645B7,
-    0x0002A1AA000299A9, 0x0002D1AC0002A1AB, 0x0002D5AE0002DDAD,
-    0x8000E9390002C9AF, 0x8000E9418000E93D, 0x8000E945001E2CBF,
-    0x8000E94D8000E949, 0x000355B88000E951, 0x000369BA00035DB9,
-    0x000361BC000365BB, 0x000361BE000361BD, 0x0002B9B0000361BF,
-    0x000285B200028DB1, 0x000375B400036DB3, 0x000375B600037DB5,
-    0x8000E95500036DB7, 0x000221B38000E959, 0x000251B58000E95D,
-    0x8000E96100038884, 0x8000E7D1000255B6, 0x000241BB001C28BE,
-    0x0003B1BD000275BA, 0x0003B1BF000259BC, 0x000269A30003B9BE,
-    0x8000E9698000E965, 0x8000E9718000E96D, 0x000219A500021DA6,
-    0x000209AB8000E975, 0x8000E97900023DAA, 0x0003F9AF8000E97D,
-    0x0003F9AD0003F1AE, 0x0002A5A8000211AC, 0x0002BDAA0002B5A9,
-    0x0002ADAC0002B5AB, 0x00013DAE000135AD, 0x001C2CBE000135AF,
-    0x8000E9858000E981, 0x8000E98D8000E989, 0x001D68878000E995,
-    0x000185B8001D1C86, 0x000185BA00018DB9, 0x00018DBC00019DBB,
-    0x0001B5BE0001BDBD, 0x000151B0000079BF, 0x000151B2000151B1,
-    0x0001F1B4000151B3, 0x0001F5B60001F1B5, 0x0000D8EF0001E5B7,
-    0x0000A5810000B582, 0x8000E9990000A580, 0x8000E9A18000E99D,
-    0x8000E9A50006C4EF, 0x8000E9A90006F4E1, 0x0000CCE3000138E3,
-    0x000128E18000E9AD, 0x8000E9B58000E9B1, 0x000275B50001E5B6,
-    0x000265B3001C4084, 0x8000E9BD8000E9B9, 0x000165BF8000E9C1,
-    0x0001D5BD000165BE, 0x0001C5BB0001D5BC, 0x8000E9C50001C5BA,
-    0x001DB5A38000E9C9, 0x8000E9CD8000E991, 0x8000E9D58000E9D1,
-    0x001DA5A5001E35A6, 0x001E15AB8000E9D9, 0x8000E9DD001E15AA,
-    0x001EB5AF8000E9E1, 0x001E05AD001EB5AE, 0x8000E9E5001E05AC,
-    0x8000E9ED8000E9E9, 0x00000D808000E9F1, 0x00003D8200003581,
-    0x8000E9F98000E9F5, 0x000005B58000E9FD, 0x0002B4E180001B67,
-    0x000005B50002D8AC, 0x001FB5A880001B6B, 0x001F1DAA001F15A9,
-    0x001F31AC001F15AB, 0x001F2DAE001F3DAD, 0x0002B4E1001F21AF,
-    0x000005B50002D8AC, 0x0002B4E180001B6F, 0x000005B50002D8AC,
-    0x0000D1B880001B73, 0x0000E1BA0000D9B9, 0x000091BC0000E1BB,
-    0x000091BE000091BD, 0x001F21B0000091BF, 0x001F31B2001F31B1,
-    0x001F09B4001F31B3, 0x0000F1B6001F09B5, 0x0002B4E10000F1B7,
-    0x001D54EF0002DCAC, 0x80001B7700001DB5, 0x0002D0AC0002BCE1,
-    0x000125A3000029B5, 0x001D15A1000391A2, 0x001E80E1001D0DA0,
-    0x001DBCE380001B7B, 0x001ED5B30002C4E1, 0x0002D8AC000009B5,
-    0x0002BCE180001B7F, 0x001EF9B5001EFDB6, 0x001EB5BB0002E0AC,
-    0x000005B5001EA5BA, 0x001E8DBF80001B83, 0x001EA5BD001E8DBE,
-    0x001E15A3001EA5BC, 0x0002D0AC0002BCE1, 0x0025698F000011B5,
-    0x001E39A5001E3DA6, 0x001E75AB80001B87, 0x0002B4E1001E65AA,
-    0x001E4DAF000178AC, 0x001E65AD001E4DAE, 0x0014DD9B001E65AC,
-    0x001701990015E59A, 0x001F719F0010E198, 0x0019F99D0019799E,
-    0x002DE593001B019C, 0x00296D91002F1192, 0x0011F99700296D90,
-    0x002C85950011F196, 0x000025B5002D9994, 0x0026F182000294E1,
-    0x002A9584002A8D83, 0x002E1D86002CE185, 0x0002B4AC002E7987,
-    0x002F558A80001B8B, 0x0012398C0012298B, 0x0014F18E0013198D,
-    0x000005B500161D8F, 0x0017559280001B8F, 0x001A719400189193,
-    0x001CF996001AF995, 0x0003E082001E7D97, 0x001E559A80000B9C,
-    0x0002759C0000D99B, 0x80000CC580000C95, 0x000401AC000A89B8,
-    0x000681AE000685AD, 0x00037CCC000210CC, 0x80001B9380000CC8,
-    0x80000CE380000CE0, 0x80000B0980000CE6, 0x80000CD10232A581,
-    0x0006959A8000E921, 0x000691BC0223559B, 0x000691BE0000B1BD,
-    0x0006E9B80006ADBF, 0x000691BA000695B9, 0x0006C1B480001B97,
-    0x0006CDB60006CDB5, 0x0006FDB00006DDB7, 0x00009DB20006E5B1,
-    0x03C754850006E5B3, 0x0000FC8780001B9B, 0x80001BA3000A01B8,
-    0x80001BA780000D1C, 0x0001A48D000CE58C, 0x80000E450002F0CC,
-    0x0000148980000E48, 0x00000C8B000A19B8, 0x80000EFE80000EA1,
-    0x80000CD480000CCB, 0x0001C09180000A1F, 0x000A6DB880000CEC,
-    0x80000CFE80000CF2, 0x80000D0A80000D07, 0x80001BAB80000D0D,
-    0x80000D1680000D88, 0x001EA4E180001BAF, 0x001E54E380000D8B,
-    0x000228CC007325AF, 0x80000D1F80000D22, 0x80000D2880000D25,
-    0x000280CC80000D2B, 0x80000D2E0003B894, 0x001E60EF00021492,
-    0x80000D3700020090, 0x000AF1B880000D4F, 0x80000BA880000D52,
-    0x000B298980000D5B, 0x80001B9F80000D5E, 0x0001C4BF000130BC,
-    0x80000D610001ECBE, 0x000AC5B80002ECCC, 0x800009B0000144BA,
-    0x800006CA80000D64, 0x800006F2800006DF, 0x800006FA80000D70,
-    0x8000070A80000D73, 0x800007188000072D, 0x8000072F800007F9,
-    0x8000073A80000E30, 0x8000074A80000E36, 0x8000076A000079B5,
-    0x80000775022A71B7, 0x80000777000061B1, 0x8000079002290DB3,
-    0x800007A30000E9BD, 0x800007B0800007FD, 0x800007C6022B19B9,
-    0x80000822021459BB, 0x8000083F8000095D, 0x8000085E80000EB6,
-    0x8000088480000039, 0x800008CA80000071, 0x800009238000002B,
-    0x8000095F8000003D, 0x8000096180000043, 0x8000086D80000848,
-    0x8000080380000049, 0x8000097C80000053, 0x80000D7680000059,
-    0x02220DBB8000005F, 0x023614CC80000E57, 0x000065BF8000001F,
-    0x00006DBD000011BE, 0x00016980800007E8, 0x0001718200017581,
-    0x000769840221DD83, 0x0007818600078185, 0x0001A18800017187,
-    0x0007B58A0001AD89, 0x0007958C00078D8B, 0x8000004F0001A58D,
-    0x0001E1900001A58F, 0x0221B1920007ED91, 0x00078D940007FD93,
-    0x0001F19600065195, 0x0001D1980001D997, 0x00067D9A00067599,
-    0x0006699C0001D59B, 0x0006519E0214999D, 0x021475A00006519F,
-    0x000149A20006B9A1, 0x000121A4022D85A3, 0x021475A600012DA5,
-    0x0211A1A80006B5A7, 0x000695AA021451A9, 0x0235BCCC800022AB,
-    0x8000006D02353CCD, 0x800003AC800003AA, 0x800022B7800003AF,
-    0x800022D5800022BE, 0x800000C9800022DC, 0x0006A1B88000000F,
-    0x800006B7800006BA, 0x800022C8800000CC, 0x800005DD800003B5,
-    0x000205BA80000618, 0x0005E9B80002D5BB, 0x0002C9BE0005FDB9,
-    0x00020DBC0217E5BF, 0x0005E1B2000201BD, 0x00058DB00005E1B3,
-    0x0005E5B600059DB1, 0x0005F5B40005E9B7, 0x00053DAA0005EDB5,
-    0x00053DA80005C1AB, 0x0005CDAE000535A9, 0x0005CDAC0005F5AF,
-    0x00053DA20005C5AD, 0x0002D5A0000505A3, 0x00051DA6000539A1,
-    0x00051DA4000505A7, 0x800006BF000515A5, 0x8000063E0005C19B,
-    0x0005DD9E80000655, 0x0002519C0004319F, 0x800006720005D59D,
-    0x800003A480000689, 0x800000DB8000232A, 0x80000792800003A0,
-    0x800007F5800007F1, 0x8000080580000993, 0x80000B9F80000809,
-    0x80000AB880000997, 0x8000078E80000772, 0x800007A58000079A,
-    0x800009AD800007AD, 0x8000100880000FF1, 0x023304CC80001015,
-    0x023320CC0230B0CD, 0x023010CC0230E0CD, 0x023058CC023068CC,
-    0x023168CC02308CCD, 0x0231D4CC023114CD, 0x022070CC023614CD,
-    0x023630CC0227D0CD, 0x023C0CCC023124CC, 0x023F18CC023F68CD,
-    0x023D18CC023F34CD, 0x023C44CC0232C0CD, 0x023238CC023918CC,
-    0x023288CC0232A8CD, 0x02337CCC0233F8CD, 0x023328CC023340CD,
-    0x0221C8CC023308CD, 0x0226ACCC022624CD, 0x022598CC023840CC,
-    0x02241CCC023AC8CD, 0x023B84CC022410CD, 0x0225ACCC023BA8CD,
-    0x0238ACCC023ABCCD, 0x0238E0CC0227F8CC, 0x8000101D023874CD,
-    0x80000DE8800006F6, 0x022A20CD80000E0F, 0x800006AA022B1CCC,
-    0x0220A4CC80002326, 0x022638CC0227A0CD, 0x023A38CC800004CA,
-    0x023910CC023B3CCD, 0x0003A0CC023E60CD, 0x022CEDBD000D3DBE,
-    0x022355BB800004D6, 0x02223DB980000907, 0x800007DC800006CF,
-    0x80000E27800004A5, 0x8000056F800004B2, 0x800004BE800006F4,
-    0x000CBDAF80000575, 0x000CA5AD022E99AE, 0x0222C5AB800005C2,
-    0x800004C480000603, 0x800004D080000623, 0x80000629800005C8,
-    0x021801A38000065D, 0x8000071A800004A0, 0x000C5D9F8000071D,
-    0x000C459D000C519E, 0x0221599B80000727, 0x800007B4800007B2,
-    0x800007BB800007B7, 0x800007D18000072A, 0x02262D93800007D3,
-    0x8000086F800007D6, 0x000C058F80000872, 0x000C658D000C718E,
-    0x0220458B8000087C, 0x022F358980000963, 0x8000096A80000966,
-    0x800009738000087F, 0x800003BD80000976, 0x000C5D81800022FA,
-    0x00010080000C6180, 0x0004008200001881, 0x00071084800022FE,
-    0x0008E48600061485, 0x00052C8800023C87, 0x0078008A0005A489,
-    0x0006908C80002302, 0x8000230E80002306, 0x000011B88000230A,
-    0x007A4892007BC491, 0x8000231600794C93, 0x0009F0960002E8CC,
-    0x0009C090000039B8, 0x0009F0928000231E, 0x0009B4940002F8CC,
-    0x8000232200000DB8, 0x000005B800021CCC, 0x0002E4CC8000232E,
-    0x8000233200001DB8, 0x8000233D80002336, 0x00088CA880002354,
-    0x8000235B007BE4A9, 0x0002E4CC00047CAB, 0x8000236500000DB8,
-    0x80002369000294CC, 0x000009B8007B94B1, 0x0002FCCC000015B8,
-    0x8000236D0008E4B5, 0x800023710002D8CC, 0x0005E8BA000005B8,
-    0x007C00BC000484BB, 0x007DF8BE007FF4BD, 0x023A09800072FCBF,
-    0x023A0182023A0D81, 0x023A1984023A0583, 0x023A1186023A1D85,
-    0x023A2988023A1587, 0x023A218A023A2D89, 0x023A398C023A258B,
-    0x023A318E000284CC, 0x8000237D023A358F, 0x80002379000FE991,
-    0x000011B80002CCCC, 0x0002DCCC80002389, 0x000029B8000FE597,
-    0x0002F8CC000031B8, 0x00001DB88000238D, 0x8000239100023CCC,
-    0x000284CC00022CCC, 0x8000239980002395, 0x000F2DA4000284CC,
-    0x000F5DA6000F55A5, 0x023AA9A88000239D, 0x0000D1B8023AADA9,
-    0x000035B8800023A1, 0x000350CC800023A5, 0x00032CCC023AB5AF,
-    0x800023C9000314CC, 0x000F1DB4000F05B3, 0x000F09B6000300CC,
-    0x007D68BE000F09B7, 0x000075B800006DB8, 0x000F29BC000348CC,
-    0x000F25BE800023CD, 0x000EE587000350CC, 0x0001A5B8023AE8CC,
-    0x0235F0CC0001B9B8, 0x022280CC0233A4CD, 0x023694CD02257CCD,
-    0x023BECCD022904CC, 0x0001F9B8023A64CC, 0x000ED5890001C1B8,
-    0x000EB197000ED588, 0x800023A900000DB8, 0x800023AD00001DB8,
-    0x800023B100000DB8, 0x023B699F800023B5, 0x800023D5800023D1,
-    0x0002C4CC000031B8, 0x800023B90002F8CC, 0x0000E5B8800023BD,
-    0x800023C5800023C1, 0x0000E9B80000E5B8, 0x0233CCCC0000E9B8,
-    0x000025B8800023D9, 0x000021B80233E8CC, 0x800023E1800023DD,
-    0x0002D8CC000029B8, 0x000D4DB7000011B8, 0x000E35B5000D4DB6,
-    0x000061B8000E35B4, 0x000F19B1800023E5, 0x0003F5BF000E29B0,
-    0x000D59BD000D51BE, 0x000D5DBB000C35BC, 0x000D5DB9000D5DBA,
-    0x000ABD80000D5DB8, 0x00041582000AC581, 0x00057CCC000AC983,
-    0x000AD586000ADD85, 0x000550CC000ACD87, 0x0008B98A000AF189,
-    0x0008118C00080D8B, 0x0227B58E0008158D, 0x00086990000AE58F,
-    0x0008719200086D91, 0x000524CC02246D93, 0x000A859600081D95,
-    0x000528CC00051CCC, 0x000A899A00083D99, 0x0008119C00080D9B,
-    0x00052CCC0008159D, 0x000A65A0000520CC, 0x000A65A2000A6DA1,
-    0x0004B1A40005BCCC, 0x0004B1A60002E8CC, 0x000481A8000749B8,
-    0x000899AA80002413, 0x0008B9AC000A5DAB, 0x800024170004A9AD,
-    0x0008D1B0000835AF, 0x800003D00004F1B1, 0x022829B400088DB3,
-    0x000A21B6000A21B5, 0x000B21B8000A21B7, 0x800003E9000825B9,
-    0x021D0DBC000B29BB, 0x000B31BE000FDDBD, 0x000ADD80000B31BF,
-    0x0001B1AA0001A1A9, 0x8000041B80000402, 0x0009998600091985,
-    0x0009E18800099587, 0x8000042E02252589, 0x8000045480000441,
-    0x8000047A80000467, 0x000AAD908000048D, 0x00056D9200054991,
-    0x0005619400056193, 0x0005659600056D95, 0x0005359800051197,
-    0x0005359A00053D99, 0x0005159C00050D9B, 0x0005159E00051D9D,
-    0x0009A9A00005099F, 0x0005F5A20005FDA1, 0x000505A4000501A3,
-    0x000505A600050DA5, 0x000505A800053DA7, 0x000505AA00050DA9,
-    0x000989AC000519AB, 0x0009B5AE0005A9AD, 0x000901B00009FDAF,
-    0x000575B200057DB1, 0x000969B4000501B3, 0x000505B6000501B5,
-    0x000505B800053DB7, 0x000501BA0225E1B9, 0x0225D1BC000501BB,
-    0x000979BE00093DBD, 0x0005308300095DBF, 0x800004F3000778A1,
-    0x000778A4800000DD, 0x8000011300074CA5, 0x00047C8B8000011C,
-    0x8000012480000120, 0x8000012800044888, 0x800001308000012C,
-    0x800000E180000134, 0x00070CB2800000E6, 0x800000EB000700B3,
-    0x0007E8B6800000F0, 0x800000F50007E0B7, 0x0007E0BA800000FA,
-    0x000784BC00079CBB, 0x800000FF000790BD, 0x0004CCA3007ED19F,
-    0x8000010980000104, 0x8000010E00040083, 0x000420A500047485,
-    0x00044C8880000117, 0x800004F0800003CD, 0x00072DAF8000058F,
-    0x000721AD0007B08D, 0x800005E2000729AC, 0x800006438000061D,
-    0x8000065A000765B0, 0x8000068E80000677, 0x800003EC800003D3,
-    0x8000041E80000405, 0x8000044480000431, 0x8000046A80000457,
-    0x022BDD8200047CBC, 0x000AFD80022BE583, 0x00099986022BF981,
-    0x000AE18400099987, 0x0009898A000AE185, 0x000989880009898B,
-    0x0009858E00098989, 0x0008E18C8000047D, 0x022BB99200098D8D,
-    0x022B999002274193, 0x000BC996000BC591, 0x000DC59402277597,
-    0x000BC99A0224D195, 0x000BC998022AD99B, 0x8000049000077999,
-    0x000BFD9C800004F6, 0x800001A4800001A0, 0x800001AC800001A8,
-    0x800001B4800001B0, 0x800001BC800001B8, 0x007F3D9C800001C0,
-    0x000911A8007F15B3, 0x800001C8800001C4, 0x800001D0800001CC,
-    0x800001D4007ED882, 0x800001DC800001D8, 0x800001E4800001E0,
-    0x800001EC800001E8, 0x800001F4800001F0, 0x800001FC800001F8,
-    0x8000020480000200, 0x800023478000093C, 0x0254F5A0800006A2,
-    0x0254E9A20254D9A1, 0x0075EDA40075F5A3, 0x007695A600769DA5,
-    0x007685A80076BDA7, 0x007CC1AA007CEDA9, 0x007D21AC007D21AB,
-    0x007D15AE007D1DAD, 0x007D75B0007D0DAF, 0x007D75B2007D7DB1,
-    0x007E5DB4007E45B3, 0x007E45B6007E4DB5, 0x007E49B8007E79B7,
-    0x007E5DBA007E55B9, 0x007E4DBC007E55BB, 0x007E3DBE007E35BD,
-    0x80000795007E2DBF, 0x800007DA800006AC, 0x007F6DB4800006AF,
-    0x000015B6000011B5, 0x0258A1BC800023FD, 0x80002407007831BD,
-    0x80003B37800032CB, 0x80002BBB80002B97, 0x80002BD780002BC7,
-    0x80002BEB80002BDF, 0x80002BFB80002BF3, 0x80002C6280002C07,
-    0x80002C8A80002C7A, 0x80002CAA80002C9A, 0x80002CDA80002CBA,
-    0x80002CE280002CDE, 0x007F558080002D16, 0x007F6D82007F6581,
-    0x007F9584007F7583, 0x007F8986007F9D85, 0x007F8588007F8187,
-    0x007F858A007F8D89, 0x007FF58C007FED8B, 0x007FE98E007FFD8D,
-    0x007F9990007FE98F, 0x007FA992007F9991, 0x007FB994007FA993,
-    0x007FA996007FB995, 0x007F9998007FA997, 0x007E199A007E5199,
-    0x007E0D9C007E199B, 0x007EFD9E007EF59D, 0x007E0DA0007EF59F,
-    0x007E1DA2007E15A1, 0x007E0DA4007E15A3, 0x007E29A6007E39A5,
-    0x007E19A8007E29A7, 0x007E6DAA007E65A9, 0x007E7DAC007E65AB,
-    0x007E6DAE007E65AD, 0x007E1DB0007E65AF, 0x007E2DB2007E25B1,
-    0x007E3DB4007E25B3, 0x007759B6007E25B5, 0x0079FDB80075D5B7,
-    0x0075F9BA0075E9B9, 0x0075E9BC0079D9BB, 0x0075D1BE0075D5BD,
-    0x007631800075D9BF, 0x0076498200763D81, 0x0076418400764583,
-    0x0074F98600764D85, 0x0002A18800767D87, 0x0076598A00764D89,
-    0x0076498C007AB98B, 0x0002B18E007A9D8D, 0x02569190007A1D8F,
-    0x0076819200762991, 0x0076D9940076CD93, 0x0076C9960076E995,
-    0x0076A1980259D197, 0x0076F19A025A6599, 0x0076D19C025A4D9B,
-    0x0256819E007A619D, 0x007D75A00076D59F, 0x025735A2025A8DA1,
-    0x007621A4007609A3, 0x025A89A600762DA5, 0x025AB9A8025A99A7,
-    0x00775DAA007675A9, 0x80002D4E80002D3E, 0x80002D5A80002D56,
-    0x80002D8280002D66, 0x80002D9E80002D86, 0x80002DAA80002DA2,
-    0x80002E5680002DBA, 0x80002EAA80002E6A, 0x80002EC680002EBE,
-    0x80002EDA80002ECA, 0x80002F2980002EFD, 0x000A699980002F35,
-    0x80002F6180002F55, 0x80002F6980002F65, 0x80002F8580002F81,
-    0x007AFC8280002FB1, 0x80002FBD007AE883, 0x007B708680002FC5,
-    0x80002FCD007B7087, 0x80002FEC80002FDD, 0x8000302080002FF4,
-    0x8000302C80003024, 0x8000303880003030, 0x007C409280003040,
-    0x8000306C80003050, 0x800030E4007BE491, 0x800030FC800030F8,
-    0x800031808000317C, 0x007CE8A8007CD4AB, 0x80003194007CE0A9,
-    0x800031A080003198, 0x007CB0A2800031A8, 0x800031BC800031B0,
-    0x024940CC800031F2, 0x024AFCCC024900CD, 0x80003202024B68CD,
-    0x8000329380003216, 0x800032AF8000329F, 0x800032C7800032BB,
-    0x007CA8B2800032CF, 0x800032D3007CA8B3, 0x800032DB800032D7,
-    0x007C44B4800032DF, 0x00077D80007C58B5, 0x0007998200078581,
-    0x0007898400079983, 0x0007BD8600078985, 0x00078D880007B587,
-    0x0007ED8A0007E589, 0x0007FD8C0007E58B, 0x0007ED8E0007E58D,
-    0x00079D900007E58F, 0x0007A5920007AD91, 0x0007A5940007BD93,
-    0x00015D9600015595, 0x00016D9800015597, 0x00017D9A00017599,
-    0x00016D9C0001759B, 0x00015D9E0001559D, 0x0001ADA00001559F,
-    0x0001BDA20001B5A1, 0x0001ADA40001B5A3, 0x0001DDA60001D5A5,
-    0x0001EDA80001D5A7, 0x0001FDAA0001F5A9, 0x0001EDAC0001F5AB,
-    0x0001DDAE0001D5AD, 0x0001ADB00001D5AF, 0x0001BDB20001B5B1,
-    0x0001ADB40001B5B3, 0x00015DB6000155B5, 0x00016DB8000155B7,
-    0x00017DBA000175B9, 0x00016DBC000175BB, 0x00015DBE000155BD,
-    0x80003397000155BF, 0x800033EB800033DF, 0x800033F3800033EF,
-    0x800033FF800033F7, 0x8000341780003403, 0x8000343B80003433,
-    0x8000347F8000347B, 0x8000348B80003487, 0x8000349780003493,
-    0x800034BF800034BB, 0x800034DF800034DB, 0x800034E7800034E3,
-    0x8000351F800034FB, 0x8000352780003523, 0x8000358B8000354B,
-    0x800035AF8000358F, 0x800035C7800035BF, 0x8000360B800035E7,
-    0x800036578000364F, 0x800036778000366B, 0x800036878000367F,
-    0x8000368F8000368B, 0x800036AB800036A7, 0x800036CF800036C7,
-    0x800036DB800036D7, 0x800036EF800036E7, 0x800036F7800036F3,
-    0x80003703800036FF, 0x800037B7800037A3, 0x800037EF800037EB,
-    0x000C999E800037F7, 0x8000382B80003823, 0x0003AD8080003833,
-    0x0003BD820003B581, 0x0003AD840003B583, 0x0003DD860003D585,
-    0x0003ED880003D587, 0x0003FD8A0003F589, 0x0003ED8C0003F58B,
-    0x0003DD8E0003D58D, 0x0003AD900003D58F, 0x0003B1920003B191,
-    0x000C01940003B193, 0x000E5D96000E5595, 0x000E6D98000E5597,
-    0x000E7D9A000E7599, 0x000E6D9C000E759B, 0x000E5D9E000E559D,
-    0x000EADA0000E559F, 0x000EBDA2000EB5A1, 0x000EADA4000EB5A3,
-    0x000EDDA6000ED5A5, 0x000EEDA8000ED5A7, 0x000EFDAA000EF5A9,
-    0x000EEDAC000EF5AB, 0x000EDDAE000ED5AD, 0x000EADB0000ED5AF,
-    0x000EBDB2000EB5B1, 0x000EADB4000EB5B3, 0x000E5DB6000E55B5,
-    0x000E6DB8000E55B7, 0x000E7DBA000E75B9, 0x000E6DBC000E75BB,
-    0x000E5DBE000E55BD, 0x000469BC000E55BF, 0x00047DBE000475BD,
-    0x000445B8000475BF, 0x000445BA00044DB9, 0x000465B4000479BB,
-    0x000465B600046DB5, 0x00041DB000047DB7, 0x000405B200040DB1,
-    0x000455AC00047DB3, 0x000455AE00045DAD, 0x000455A8000465AF,
-    0x000455AA00045DA9, 0x000591A400044DAB, 0x000591A6000591A5,
-    0x000591A0000591A7, 0x000591A2000591A1, 0x0005519C000591A3,
-    0x0005519E0005519D, 0x000551980005519F, 0x0005519A00055199,
-    0x000511940005519B, 0x0005119600051195, 0x0007459000051197,
-    0x0007459200074D91, 0x0007258C00051193, 0x0007258E00072D8D,
-    0x0007058800073D8F, 0x0007058A00070D89, 0x0007258400073D8B,
-    0x0007258600072D85, 0x0007458000073D87, 0x0007458200074D81,
-    0x8000387700073D83, 0x800038838000387B, 0x8000388B80003887,
-    0x800038BB800038B7, 0x800038D7800038C3, 0x800038F7800038DB,
-    0x80003903800038FF, 0x8000394780003927, 0x8000395F8000395B,
-    0x8000399380003967, 0x800039A780003997, 0x800039D7800039CB,
-    0x800039FF800039E7, 0x80003A0B80003A07, 0x80003A1380003A0F,
-    0x80003A2380003A1F, 0x80003A4780003A2B, 0x80003A4F80003A4B,
-    0x80003A6B80003A67, 0x80003A8780003A73, 0x80003A8F80003A8B,
-    0x80003A9B80003A97, 0x80003AA380003A9F, 0x80003AB780003AA7,
-    0x80003AD380003ACF, 0x80003AE380003ADF, 0x80003AEB80003AE7,
-    0x80003AF780003AF3, 0x80003B0780003AFB, 0x80003B1780003B0F,
-    0x80003B2380003B1B, 0x80003B2B80003B27, 0x0001AD8080003B33,
-    0x0001BD820001B581, 0x0001AD840001B583, 0x0001DD860001D585,
-    0x0001ED880001D587, 0x0001FD8A0001F589, 0x0001ED8C0001F58B,
-    0x0001DD8E0001D58D, 0x0001AD900001D58F, 0x0001BD920001B591,
-    0x0001AD940001B593, 0x80003B3B00000D95, 0x80002D6A8000243B,
-    0x8000252380002D6E, 0x0000059900027882, 0x0002D082800024B1,
-    0x00008080000298BE, 0x0006E48200009881, 0x001BD48400048883,
-    0x001F1886001A9485, 0x001EC4880000E99E, 0x0013A88A00100089,
-    0x0028008C0011AC8B, 0x002AB88E002BB48D, 0x007474E30029BC8F,
-    0x0002B8E30002A0E3, 0x800024B500001D99, 0x0000099E000274BE,
-    0x00026C820002D8EF, 0x00000D99800024B9, 0x0071E8EF000204E3,
-    0x0002D4EF000270EF, 0x000304E300008CBE, 0x00031CE3800024BD,
-    0x000320E3800024C1, 0x00034CE3800024C5, 0x0003E4EF800024C9,
-    0x000384EF800024CD, 0x0003D8EF800024D1, 0x0003ECEF800024D5,
-    0x0003C8E3800024D9, 0x0004A0EF0003DCE3, 0x8000383B0003E8E3,
-    0x800024DD0003A8E3, 0x000000D70004A4EF, 0x800024E100034CEF,
-    0x000320EF000400E3, 0x000478E300033CEF, 0x800024E5000344EF,
-    0x800024ED800024E9, 0x000458EF800024F1, 0x00045CEF800024F5,
-    0x000480E3000480E3, 0x000498E3800024F9, 0x800024FD000474EF,
-    0x8000250580002501, 0x8000251480002509, 0x000488E3000460EF,
-    0x800000BD80002518, 0x8000252F800000C2, 0x80002A6B8000251C,
-    0x8000266980002606, 0x800026D8800026AC, 0x000444EF0004A4E3,
-    0x024B098080002702, 0x0002ED9E00014886, 0x00027CBE00027082,
-    0x00011D990001119E, 0x0001408F00025C82, 0x00016C8D00012999,
-    0x0001788B000260BE, 0x000218820001399E, 0x0002D497000208BE,
-    0x0002D49500015199, 0x0002D4930001599E, 0x0002D491000258BE,
-    0x0002B89F00025482, 0x0002A89D00000096, 0x0002B89B00017199,
-    0x0002B4990001699E, 0x0002988200015999, 0x0001AD9E000294A6,
-    0x00026882000278A4, 0x000278A10001B99E, 0x000284AF0001B199,
-    0x000258BE80002727, 0x00025CBE00027082, 0x00018D998000273C,
-    0x00027CBE000290A8, 0x0002B0B50001F19E, 0x000250BE0001F199,
-    0x000285B6000290B2, 0x00095DB800018599, 0x00028DBA000285B9,
-    0x0004F4BB80002744, 0x0001C59E00025C82, 0x00027CBE0004F4B8,
-    0x0002608200065D9E, 0x000270BE00066599, 0x0006719E00026C82,
-    0x00027CBE00069D99, 0x000260820006A59E, 0x00027CBE0006A999,
-    0x00021C820006B19E, 0x0006C999000214BE, 0x00024C8200024CBE,
-    0x0006D99E0006C19A, 0x0002BC820006D99F, 0x0006F5990002B0BE,
-    0x000288BE0005DC82, 0x0006E59E0006C99A, 0x0002B8820006D19F,
-    0x000631990002A882, 0x0002FCE30002F4E3, 0x0006259F0006399A,
-    0x0006219E0006259E, 0x000218BE0006219F, 0x0006159A00061D99,
-    0x0006719900020CBE, 0x0002E0EF000224EF, 0x0006019E0006719A,
-    0x000214E30006099F, 0x00022CE300061599, 0x0006019E00027C82,
-    0x0002F0EF80002760, 0x00027CBE00060599, 0x8000276800060D9E,
-    0x0002ACEF00027482, 0x0007D19B00039482, 0x000250BE0003F0E3,
-    0x0007CD9F0007E999, 0x00025C9D000234E3, 0x0007C59E0007C99C,
-    0x000260820002ECEF, 0x000270BE0007C599, 0x0007D19E0002C4EF,
-    0x000274E300021082, 0x0007A59900026CBE, 0x0007BD9E000254BE,
-    0x000268820007A599, 0x0007919E00020CE3, 0x0002688200079999,
-    0x0002CCEF0002D8EF, 0x00029082000288BE, 0x000004BB000298BE,
-    0x000000B90007799E, 0x00005CBF00062999, 0x000004BD0002F4E3,
-    0x0003B8B30006399E, 0x000378B1000228EF, 0x00003CB700027C82,
-    0x0003FCB500060199, 0x000348AB0006099E, 0x0002B8E3000200BE,
-    0x00031CAF00062199, 0x000340AD00027C82, 0x00060D9E000244BE,
-    0x00019CBE00026C82, 0x0001808200013199, 0x00027CEF0002A8A6,
-    0x0001A4E30001BC82, 0x0001E4E30001D8E3, 0x0001E8E30002D19E,
-    0x00027CBE0002E599, 0x0002C0EF0002ED9E, 0x00026CBE00000999,
-    0x00010CEF80002770, 0x000128EF000110EF, 0x0001B4E30000119E,
-    0x00002D9900021CE3, 0x00027C8200026CBE, 0x00003D9E00025082,
-    0x0000E599000200E3, 0x000268BE80002782, 0x0002A0EF0002FCEF,
-    0x0000FD9E0002C0EF, 0x000200E30000F199, 0x000290828000278E,
-    0x00027CBE80002796, 0x00020CE30000199E, 0x00000D9900025882,
-    0x0002E4EF00020482, 0x0002D8EF800027AC, 0x0000159E000274BE,
-    0x00026CE300016599, 0x00019D9A0001C999, 0x0001DD9E000264BE,
-    0x00020CE30001819F, 0x0001959900025882, 0x0002F4EF0003B4BE,
-    0x0002C4EF0001959A, 0x002A45B00003AC82, 0x0002C8810003F482,
-    0x000230BE0003F8E3, 0x0003E4E3800027EA, 0x000238BE0002C484,
-    0x0002C0EF00021086, 0x00005D9E00023088, 0x0000B19A0000559F,
-    0x0002748F0002C4EF, 0x00003D9E00004199, 0x00028C820000C59F,
-    0x0003D89500021892, 0x0000C99E800027FB, 0x0000F59A0000C99F,
-    0x0003609B0003F898, 0x0003A8820000FD99, 0x800028288000281D,
-    0x8000288C80002878, 0x000220BE0002C4E3, 0x800028AE000250BE,
-    0x0002E0E3800028BC, 0x00003D9980002904, 0x0002DCEF00007D9E,
-    0x8000291200001D9A, 0x0003ACAF0002C4EF, 0x001C00B100021082,
-    0x00002999000394B0, 0x00004D9E001C4CB2, 0x80002A6F0000659F,
-    0x80002A73000268E3, 0x0000099E000274BE, 0x001C3CBD0002F0EF,
-    0x001FCDBF00026C82, 0x80002A77001FFD99, 0x0000099E000274BE,
-    0x00026C82000214E3, 0x00000D9980002A7B, 0x0000359E000244BE,
-    0x0000059900026882, 0x000270BE80002A87, 0x0000159E800026E7,
-    0x00026CE30002D8EF, 0x800019F3800019EF, 0x80001B17001DB0E3,
-    0x80001B1B800019F7, 0x800019FB80001B1F, 0x80001B2780001B23,
-    0x0002B8EF80001B2B, 0x80001B3380001B2F, 0x0000C5B1000280EF,
-    0x0000D9B3800019FF, 0x0000C9B50000C5B2, 0x0000F9B70000C1B4,
-    0x000061B90000C1B6, 0x0000C1BB000021B8, 0x0000C5BD0000C9BA,
-    0x00007DBF0000D9BC, 0x80001A030000C5BE, 0x80001A4380001B37,
-    0x80001A0B80001A07, 0x80001A5780001A0F, 0x80001AA3037810EF,
-    0x80001AA7034DDCE1, 0x80001AAB037958E3, 0x80001AB380001AAF,
-    0x0001C98080001AB7, 0x0001D9820001C981, 0x0001C9840001D983,
-    0x0004018600027585, 0x0005258800051D87, 0x00053D8A00053589,
-    0x0005758C00056D8B, 0x00056D8E0005658D, 0x0001A9900001DD8F,
-    0x0001BD920001B591, 0x0001AD940001B593, 0x00035D9600035595,
-    0x00036D9800035597, 0x00037D9A00037599, 0x0003519C0003719B,
-    0x0003519E0003519D, 0x0003B5A000034D9F, 0x0003B5A20003BDA1,
-    0x0003C1A40003C5A3, 0x0003C5A60003C1A5, 0x0003C5A80003FDA7,
-    0x0003C5AA0003CDA9, 0x0003C9AC0003D9AB, 0x00033DAE000335AD,
-    0x00034DB0000335AF, 0x00035DB2000355B1, 0x00034DB4000355B3,
-    0x00037DB6000375B5, 0x00034DB8000375B7, 0x00035DBA000355B9,
-    0x00034DBC000355BB, 0x00033DBE000335BD, 0x80001ABB000331BF,
-    0x80001AC380001ABF, 0x80001ACB80001AC7, 0x000210F080001ACF,
-    0x80001AD780001AD3, 0x80001ADF80001ADB, 0x80001AE30002B482,
-    0x80001AEB80001AE7, 0x0036F59C80001AEF, 0x80001AF30036FD9D,
-    0x80001A9F00026891, 0x00DD5C9E80001A3B, 0x80001B3B80001A3F,
-    0x80001B4380001B3F, 0x00DEEDBA80001B47, 0x80001B4F80001B4B,
-    0x80001B530002F0F0, 0x80001B5780001A47, 0x80001A4B80001B5B,
-    0x80001A5380001A4F, 0x80001A5F80001A5B, 0x80001A6780001A63,
-    0x80001A6F80001A6B, 0x80001A7780001A73, 0x80001A7F80001A7B,
-    0x0002249180001A83, 0x80001A8B80001A87, 0x80001B5F80001A8F,
-    0x80001A9380001B63, 0x80001A9700038C82, 0x80001AF780001A9B,
-    0x80001AFF80001AFB, 0x02D2DCE10007DCF0, 0x02C654E380001B03,
-    0x80001B0B80001B07, 0x80001B1380001B0F, 0x80002E5A80002BA3,
-    0x80002BBF80002BA7, 0x8000349B80002BAB, 0x02E0A4EF80002BB3,
-    0x028714E380002B9B, 0x029D8CE102F390E1, 0x02AB8CE102F790E3,
-    0x035300E302B690E1, 0x80001A1302A280E3, 0x034BB4EF80002EE5,
-    0x028B24EF02B3BCEF, 0x8000020880002E6E, 0x00DF21BA029724EF,
-    0x800002148000020E, 0x800002208000021A, 0x8000022C80000226,
-    0x8000023880000232, 0x800002448000023E, 0x800002508000024A,
-    0x8000033480000256, 0x027880E18000033A, 0x026A14E3021E00E3,
-    0x020FA0E1021310E1, 0x020E84E380000340, 0x80000346023F2CEF,
-    0x023B34EF023B50E1, 0x8000034C02223CE3, 0x02120CEF021F28EF,
-    0x80001A1B80000352, 0x8000035880001A1F, 0x80001A238000035E,
-    0x8000037680001A27, 0x8000038880000382, 0x800003948000038E,
-    0x8000037C8000039A, 0x8000036D80000364, 0x80001A2B8000025C,
-    0x8000027480001A2F, 0x80001A3380000268, 0x800002BC80001A37,
-    0x800002988000027A, 0x8000029280000262, 0x800002A48000026E,
-    0x800002F2800002D4, 0x00065180800002EC, 0x0006518200065181,
-    0x0006718400065183, 0x0006758600067D85, 0x0006558800066D87,
-    0x0007AD8A00065D89, 0x0007A18C0007A18B, 0x0007A18E0007A18D,
-    0x8000032E800002E0, 0x00147D92800002C2, 0x0007CD9400140193,
-    0x0007FD960007F595, 0x0007CD980007F597, 0x00142D9A0007D599,
-    0x0007CD9C0007D99B, 0x00078D9E0007919D, 0x000125990014299F,
-    0x0001259B00012598, 0x0001259D0001259A, 0x80001A170001259C,
-    0x0006C1910001259E, 0x0006D5930006F190, 0x00013D950006DD92,
-    0x0001259700013D94, 0x0006E58900012596, 0x0006E58B0006E588,
-    0x0006E58D0006E58A, 0x0006E58F0006E58C, 0x0006E5810006E58E,
-    0x0006E58300061D80, 0x0006E5850006E582, 0x0006E5870006E584,
-    0x0003A5B90006E586, 0x0003A5BB0003A5B8, 0x0003A5BD0003A5BA,
-    0x0003A5BF0003A5BC, 0x0003A5B10003A5BE, 0x0003A5B30003A5B0,
-    0x0003A5B50003A5B2, 0x0003A5B70003A5B4, 0x0001B1A90003A5B6,
-    0x0001B1AB0001A9A8, 0x00016DAD0001B9AA, 0x00011DAF0001A9AC,
-    0x0001A9A100011DAE, 0x0001B1A30001D9A0, 0x000191A50001A1A2,
-    0x000191A70001A9A4, 0x800002CE000199A6, 0x800002DA800002E6,
-    0x800002B080000304, 0x80000322800002F8, 0x8000029E8000030A,
-    0x800002B680000280, 0x800002FE800002C8, 0x8000032880000286,
-    0x80000310800002AA, 0x800003168000028C, 0x80002E0E8000031C,
-    0x800035CF80002FF0, 0x8000050680000787, 0x8000052480000515,
-    0x8000054280000533, 0x800005548000054B, 0x000DCC828000055D,
-    0x8000059280000566, 0x800005A48000059B, 0x800005A7036E48E3,
-    0x800005AA036E4CE1, 0x800005AD0301F4EF, 0x80004DC380003B63,
-    0x80006953800056F3, 0x80007BB380007283, 0x8000A07380008E13,
-    0x8000BC0A8000A9A3, 0x8000CE718000C541, 0x8000E0D18000D7A1,
-    0x8000AF578000BC4E, 0x80002BF78000A623, 0x80002B9F80002C8E,
-    0x8000324680002BE3, 0x8000355380003206, 0x8000340B80002DDA,
-    0x800035678000388F, 0x8000359780002D22, 0x80003917800034A3,
-    0x80002D8A80002C5E, 0x80003A2F800032FF, 0x80002CB280002C22,
-    0x80002FC1800032A3, 0x80002D5E80002DCE, 0x80002F3180002F3D,
-    0x80002C1E800034F3, 0x800038A30003EC91, 0x80002ED280002D7A,
-    0x800005B3800005B0, 0x033FD4E1800005B6, 0x032FDCE3031A58E1,
-    0x03F214E1030E50E3, 0x03EA40E303D340E1, 0x800005B903C340E3,
-    0x03EB78EF800005E5, 0x800005E803F25CEF, 0x03DE44EF800005EB,
-    0x038BB4E1032598EF, 0x03A27CE303973CE1, 0x004130E1038B98E3,
-    0x007F1CE303AC50E1, 0x800005EE004688E3, 0x007B38EF800005F1,
-    0x800005F40041E0EF, 0x038ACCEF800005F7, 0x0018B580038720EF,
-    0x000B15820018C981, 0x000EE184000B2D83, 0x000EE186000EE185,
-    0x000EDD880018FD87, 0x001BE98A001A9989, 0x001BC58C001D2D8B,
-    0x00107D8E001BF18D, 0x001BE990001AE18F, 0x000F4D92000F4591,
-    0x000F5D94000F4593, 0x000F4D96000F4595, 0x000F7D98000F4597,
-    0x000F4D9A000F4599, 0x000F599C001A999B, 0x000F499E000F599D,
-    0x000FB9A0000F499F, 0x000FC9A2000FB9A1, 0x000BB5A4000FC9A3,
-    0x000BB5A6000BBDA5, 0x000FEDA8000FD5A7, 0x000FFDAA000FF5A9,
-    0x000FEDAC000FF5AB, 0x000FD1AE000FD5AD, 0x000FB1B0000FD1AF,
-    0x000FB1B2000FB1B1, 0x000F71B4000FB1B3, 0x000F69B6000F71B5,
-    0x000F01B8000F69B7, 0x000F19BA000F01B9, 0x0001F1BC000F19BB,
-    0x0001F1BE0001F1BD, 0x800005FA0001F1BF, 0x80000600800005FD,
-    0x800004E180000620, 0x800005D380000580, 0x800006348000060E,
-    0x800006688000064B, 0x800006968000067F, 0x800003F6800003DD,
-    0x800007128000040F, 0x8000084180000844, 0x800025378000073F,
-    0x800025A180002570, 0x800027B7800025C6, 0x800027C6800027C2,
-    0x800027CE800027CA, 0x8000296C8000292D, 0x8000299880002994,
-    0x800029C5800029BD, 0x800029F3800029E1, 0x800029FB800029F7,
-    0x80002A2E80002A13, 0x80002A4900D501BA, 0x8000253B80002C17,
-    0x8000255580002548, 0x8000257480002562, 0x8000259380002581,
-    0x800025B5800025A5, 0x800025D4800025CA, 0x800025EC800025E2,
-    0x8000260A800025F9, 0x8000266D80002614, 0x8000262280002674,
-    0x000311808000267E, 0x0003118200031181, 0x0003318400031183,
-    0x0003318600033185, 0x0003118800033187, 0x0003118A00031189,
-    0x0003718C0003118B, 0x0003718E0003718D, 0x000311900003718F,
-    0x0004019200031191, 0x00030D9400031593, 0x00065D9600065595,
-    0x00066D9800065597, 0x00067D9A00067599, 0x00066D9C0006759B,
-    0x00063D9E0006359D, 0x0006CDA00006359F, 0x0006DDA20006D5A1,
-    0x0003F1A40006D5A3, 0x0003F1A60003F1A5, 0x0003D1A80003F1A7,
-    0x000629AA0006F9A9, 0x000639AC000629AB, 0x0003CDAE000639AD,
-    0x0003BDB00003C5AF, 0x00034DB2000345B1, 0x00035DB4000345B3,
-    0x00034DB6000345B5, 0x00037DB8000345B7, 0x000341BA000345B9,
-    0x000665BC00067DBB, 0x000665BE00066DBD, 0x000F15A9000351BF,
-    0x000F4DAB000F01A8, 0x000F45AD000F4DAA, 0x000DA9AF000F45AC,
-    0x000B75A1000DB1AE, 0x000BE9A3000B61A0, 0x000BE1A5000B4DA2,
-    0x000F39A7000BF9A4, 0x000D99B900085DA6, 0x000DA9BB000D91B8,
-    0x000DB9BD000D91BA, 0x8000262F000DB1BC, 0x000DD9B1000DB1BE,
-    0x000DA9B3000DD1B0, 0x000DB9B5000DD1B2, 0x000DA9B7000DB1B4,
-    0x80002636000DB1B6, 0x8000265980002646, 0x800026BA800026B0,
-    0x8000269F8000268F, 0x000DE581800026CE, 0x000DE583000DE580,
-    0x000DE585000DE582, 0x000DE187000DE584, 0x000D459900182D86,
-    0x000D459B000DB998, 0x000D499D000D459A, 0x000D719F000D459C,
-    0x000D8191000D7D9E, 0x000D8193000DB990, 0x000D8195000D8992,
-    0x000D8197000D9994, 0x800025DE000D8996, 0x8000268B8000261E,
-    0x800026DC800026CA, 0x8000272B80002706, 0x800026EE80002740,
-    0x8000270A800026F8, 0x8000272F80002717, 0x800027568000274C,
-    0x8000274880002774, 0x8000276C80002764, 0x8000278A8000277E,
-    0x800027A180002792, 0x8000279A800027DC, 0x800027A5800027B0,
-    0x800027D2800027BB, 0x800027FF800027E0, 0x800028198000280F,
-    0x800028908000284D, 0x80002A4D800028E5, 0x80002A5F80002A5B,
-    0x800027EE80002A63, 0x8000283C8000282C, 0x8000282180002846,
-    0x8000286180002851, 0x8000286E8000287C, 0x800028C080002894,
-    0x800028A4800028C7, 0x800028DB800028D1, 0x80002916800028B2,
-    0x800028E980002908, 0x800028F080002923, 0x80002931800028FA,
-    0x800029488000293E, 0x8000295C80002952, 0x8000297D80002970,
-    0x8000299C80002984, 0x800029B0800029A3, 0x800029D3800029C9,
-    0x800029FF800029E5, 0x80002A1780002A0C, 0x80002A3280002A21,
-    0x80002A5180002A39, 0x800004E6800003C3, 0x800005D880000585,
-    0x8000063980000613, 0x8000066D80000650, 0x8000069B80000684,
-    0x800003FC800003E3, 0x8000042880000415, 0x8000044E8000043B,
-    0x8000047480000461, 0x8000049A80000487, 0x8000050F80000500,
-    0x8000052D8000051E, 0x800008668000053C, 0x800006C180000827,
-    0x80000920800007FF, 0x80000936001338E3, 0x800008308000082D,
-    0x8000072480000834, 0x800031CF80002FE1, 0x800031C080002EDE,
-    0x8000092A8000324A, 0x800009090013E4EF, 0x800008D280000E63,
-    0x8000073180000886, 0x800006FC8000074C, 0x800008978000080D,
-    0x8000090C8000092D, 0x80000E7380000E67, 0x8000089C800008DE,
-    0x8000088980000715, 0x800006FF80000755, 0x8000258F800007A9,
-    0x800008E580000E77, 0x8000089F8000082A, 0x800009188000085B,
-    0x800008E880000E7B, 0x800008A280000817, 0x8000081A800008EB,
-    0x800008A5800008CC, 0x8000081E800008EF, 0x800008A9800008CF,
-    0x80000900800008FA, 0x8000088D8000078B, 0x8000070380000759,
-    0x8000094480000940, 0x800009398000094C, 0x80000E7F8000091B,
-    0x80000930800008F7, 0x80000E6B8000090F, 0x80000891800008D5,
-    0x8000093380000760, 0x80000E6F80000912, 0x80000894800008D8,
-    0x800008B080000763, 0x001240E380000766, 0x0013F4E30013F0E3,
-    0x000D08E30013D8E3, 0x0012ACE30012E8E3, 0x000DC4EF001264E3,
-    0x001234EF000DD8EF, 0x001240EF001258EF, 0x0012C8EF0012D4EF,
-    0x800007EC001320EF, 0x80000811800006CC, 0x800006D880000814,
-    0x80000824800006D4, 0x8000086A80000707, 0x800008798000070C,
-    0x8000073780000734, 0x800008B9800008AD, 0x0010D0E3800008BC,
-    0x0010A8E30010B0E3, 0x00104CE3001178E3, 0x0010E8E3001030E3,
-    0x001084EF0010FCE3, 0x00102CEF001098EF, 0x001004EF001020EF,
-    0x001070EF001008EF, 0x001398E300104CEF, 0x001030E3001030E3,
-    0x00137CE3001020E3, 0x001358E3001340E3, 0x0013D4EF001380E3,
-    0x001380EF0013A8EF, 0x001344EF00137CEF, 0x001318EF001324EF,
-    0x003CED800010ECEF, 0x003CFD82003CF581, 0x003CED84003CF583,
-    0x003D1D86003D1585, 0x003D2D88003D1587, 0x003D3D8A003D3589,
-    0x003D2D8C003D358B, 0x003D1D8E003D158D, 0x003D6D90003D158F,
-    0x003D7D92003D7591, 0x003D6D94003D7593, 0x003D1D96003D1595,
-    0x003D2D98003D1597, 0x003D3D9A003D3599, 0x003D2D9C003D359B,
-    0x003D1D9E003D159D, 0x003DEDA0003D159F, 0x003DFDA2003DF5A1,
-    0x003DEDA4003DF5A3, 0x003D1DA6003D15A5, 0x003D2DA8003D15A7,
-    0x003D3DAA003D35A9, 0x003D2DAC003D35AB, 0x003D1DAE003D15AD,
-    0x003D6DB0003D15AF, 0x003D7DB2003D75B1, 0x003D6DB4003D75B3,
-    0x003D11B6003D15B5, 0x003D31B8003D11B7, 0x003D31BA003D31B9,
-    0x003D11BC003D31BB, 0x003D11BE003D11BD, 0x003CF180003D11BF,
-    0x003CF582003CFD81, 0x003F1584003F0D83, 0x003F1586003F1D85,
-    0x003F3588003F0D87, 0x003F358A003F3D89, 0x003F158C003F0D8B,
-    0x003F158E003F1D8D, 0x003F7590003F0D8F, 0x003F7592003F7D91,
-    0x003F1594003F0D93, 0x003F0996003F1995, 0x003F3998003F0997,
-    0x003F099A003F3999, 0x003F199C003F099B, 0x003F099E003F199D,
-    0x003FF9A0003F099F, 0x003F09A2003FF9A1, 0x003F19A4003F09A3,
-    0x003F09A6003F19A5, 0x003F39A8003F09A7, 0x003F09AA003F39A9,
-    0x003F19AC003F09AB, 0x003F09AE003F19AD, 0x003F79B0003F09AF,
-    0x003F09B2003F79B1, 0x003F19B4003F09B3, 0x003F09B6003F19B5,
-    0x003F39B8003F09B7, 0x003F09BA003F39B9, 0x003F19BC003F09BB,
-    0x003F09BE003F19BD, 0x003CF980003F09BF, 0x003D4982003CF981,
-    0x003D5984003D4983, 0x003D4986003D5985, 0x003D7988003D4987,
-    0x003D498A003D7989, 0x003D598C003D498B, 0x003D498E003D598D,
-    0x003D3990003D498F, 0x00040192003D3991, 0x0006559400064D93,
-    0x0006559600065D95, 0x0006759800064D97, 0x0006759A00067D99,
-    0x0006559C00064D9B, 0x0006559E00065D9D, 0x0006B5A000064D9F,
-    0x0006B5A20006BDA1, 0x0006D5A40006CDA3, 0x0006D5A60006DDA5,
-    0x0006F5A80006CDA7, 0x0006F5AA0006FDA9, 0x0006D5AC0006CDAB,
-    0x0006D5AE0006DDAD, 0x0006B5B00006CDAF, 0x0006B5B20006BDB1,
-    0x000655B400064DB3, 0x000655B600065DB5, 0x000675B800064DB7,
-    0x000675BA00067DB9, 0x000655BC00064DBB, 0x000655BE00065DBD,
-    0x003F7DAC00064DBF, 0x003FADAE003FA5AD, 0x003F7DA8003FA5AF,
-    0x003F6DAA003F65A9, 0x003F1DA4003F65AB, 0x003F4DA6003F45A5,
-    0x003F1DA0003F45A7, 0x003F0DA2003F05A1, 0x003F9DBC003F05A3,
-    0x003FADBE003FA5BD, 0x003F9DB8003FA5BF, 0x003F8DBA003F85B9,
-    0x003FDDB4003F85BB, 0x003FADB6003FA5B5, 0x003FDDB0003FA5B7,
-    0x003FCDB2003FC5B1, 0x003A618C003FC5B3, 0x003A618E003A618D,
-    0x003A4188003A618F, 0x003A418A003A4189, 0x003A6184003A418B,
-    0x003A6186003A6185, 0x003A0180003A6187, 0x003A0182003A0181,
-    0x00385D9C003A0183, 0x003FED9E003FE59D, 0x00385D98003FE59F,
-    0x00384D9A00384599, 0x0038B9940038459B, 0x00386D9600386595,
-    0x003A019000386597, 0x003A0192003A0191, 0x800008C3003A0193,
-    0x800008E1800008DB, 0x80000779800008F3, 0x8000077C80000925,
-    0x8000095780000784, 0x800007CE800007A0, 0x800006C4800007C0,
-    0x8000057B800004DC, 0x80000609800005CE, 0x800006468000062F,
-    0x8000067A80000663, 0x800003D780000691, 0x80000409800003F0,
-    0x8000043580000422, 0x8000045B80000448, 0x800004818000046E,
-    0x800004FA80000494, 0x8000051880000509, 0x8000053680000527,
-    0x8000054E80000545, 0x8000056080000557, 0x8000059580000569,
-    0x800008608000059E, 0x80000F2880000F22, 0x005CDDB280003B4B,
-    0x80000AB5005CF1B3, 0x80003B4780000B1E, 0x005239B480003B53,
-    0x80003B5780000D10, 0x000021B880000D67, 0x8000387F000021B9,
-    0x800038A7800031F6, 0x0035309A8000335F, 0x80002DBE80002BC3,
-    0x80002E2280002EF1, 0x80002EED00031C9F, 0x0001F19C0029249D,
-    0x0001F59E0001F19D, 0x800030DC0001CD9F, 0x80003667800034D7,
-    0x800037DB8000379F, 0x800039238000380F, 0x800033078000327B,
-    0x8000344B800033B3, 0x8000394F80003767, 0x80002BE780003A7B,
-    0x8000328F80002D92, 0x8000379B800033D7, 0x80002F9980003ACB,
-    0x8000378B8000337F, 0x800031108000381F, 0x800037E7800036C3,
-    0x8000320A8000300C, 0x800034238000331F, 0x80002C2E8000392F,
-    0x80002D3280002CC6, 0x8000328780003178, 0x800034F7800033D3,
-    0x800037AB80003793, 0x800039F7800038CF, 0x80003AAB00028CAD,
-    0x8000353380003AC7, 0x8000362B80003573, 0x8000374B000111A9,
-    0x001C00B080003983, 0x001800B2001E04B1, 0x001B10B400190CB3,
-    0x00E518B600E400B5, 0x00E000B800E71CB7, 0x00E220BA00E124B9,
-    0x00ED34BC00EC00BB, 0x00EF38BE00EE34BD, 0x8000385300E374BF,
-    0x800008BF80002EA6, 0x000350A480003028, 0x000E58A6000C00A5,
-    0x000A60A8000800A7, 0x001568AA001400A9, 0x001750AC0016ECAB,
-    0x0012B0AE001174AD, 0x800035E30013F0AF, 0x80003407800036A3,
-    0x8000389F8000353B, 0x80002EA2800039EF, 0x8000327F80002F75,
-    0x8000336780003333, 0x800036438000361B, 0x80002D2E800039AF,
-    0x80002CD2800036AF, 0x8000359F80002CCE, 0x800037638000362F,
-    0x8000386F800039B7, 0x8000385F80003114, 0x80002F4980002BCF,
-    0x8000343780003080, 0x80002D52800034B7, 0x0003F18080003543,
-    0x002F2D820003F181, 0x0003ED840003F583, 0x0002118600021185,
-    0x002FDD88002FE187, 0x002F6D8A00022989, 0x0002298C00023D8B,
-    0x00020D8E0002058D, 0x00027D900002058F, 0x00020D9200020591,
-    0x00021D9400020593, 0x00020D9600020595, 0x00023D9800020597,
-    0x00020D9A00020599, 0x00021D9C0002059B, 0x00020D9E0002059D,
-    0x0002FDA00002059F, 0x00020DA2000205A1, 0x00021DA4000205A3,
-    0x00020DA6000205A5, 0x00023DA8000205A7, 0x00020DAA000205A9,
-    0x00021DAC000205AB, 0x00020DAE000205AD, 0x00027DB0000205AF,
-    0x00020DB2000205B1, 0x00224DB4000205B3, 0x000209B6000219B5,
-    0x000239B8000209B7, 0x00020DBA000205B9, 0x00021DBC000205BB,
-    0x000201BE000205BD, 0x000EC595000201BF, 0x000EC597000EC594,
-    0x000EA591000EC596, 0x000EA593000EA590, 0x000EC19D000EA592,
-    0x000EC19F000D699C, 0x000EE599000EC99E, 0x000EE59B000EE598,
-    0x000EC585000EE59A, 0x000EC587000EC584, 0x000E1181000EC586,
-    0x000E158300209580, 0x000EC58D000E1982, 0x000EC58F000EC58C,
-    0x000EE589000EC58E, 0x000EE58B000EE588, 0x0001BDB5000EE58A,
-    0x0001CDB70001BDB4, 0x0020BDB10001CDB6, 0x0001ADB30001B9B0,
-    0x0001FDBD0001ADB2, 0x0001CDBF0001FDBC, 0x0001FDB90001CDBE,
-    0x0001EDBB0001FDB8, 0x0001C1A50001EDBA, 0x0001C1A7000E39A4,
-    0x000E21A10001C9A6, 0x000E21A3000E39A0, 0x0001C1AD000E29A2,
-    0x0001C1AF0001F9AC, 0x0001E1A90001C9AE, 0x0001E1AB0001F9A8,
-    0x80002C3E0001E9AA, 0x80002BAF80003064, 0x80003190800032F3,
-    0x80002DA680003617, 0x8000350380002E92, 0x800038438000376B,
-    0x800032EB800032BF, 0x8000372380003124, 0x800034B380003144,
-    0x80002C8680002C03, 0x8000325B80002CCA, 0x800036FB80003603,
-    0x800039638000384F, 0x80002DEE80002D3A, 0x800031AC8000301C,
-    0x8000354780003383, 0x80003A838000399F, 0x80003AEF80003AD7,
-    0x800032AB800031EE, 0x80002FE8800038F3, 0x800030E0800030C4,
-    0x8000337380003170, 0x8000346F800033BF, 0x8000363780003593,
-    0x800038E38000369B, 0x8000390B80003773, 0x80002CEE80003987,
-    0x80002E0280002D1A, 0x800037FB800033AF, 0x8000307880003008,
-    0x800032B78000313C, 0x8000342B800035DF, 0x0003698080002C13,
-    0x000365820028F181, 0x0003658400037D83, 0x0003658600036D85,
-    0x0003A58800039D87, 0x002AD58A0003AD89, 0x0003A18C0003A18B,
-    0x0003A18E0003A18D, 0x0003E1900003A18F, 0x0003E1920003E191,
-    0x0003E1940003E193, 0x0003E1960003E195, 0x0003A1980003E197,
-    0x0003A19A0003A199, 0x0003A19C0003A19B, 0x0003A19E0003A19D,
-    0x000361A00003A19F, 0x00036DA20003ADA1, 0x00037DA4000365A3,
-    0x00036DA6000365A5, 0x00035DA8000365A7, 0x000C41AA0001A5A9,
-    0x0001B9AC0001A9AB, 0x0001A9AE0001B9AD, 0x0001D9B00001A9AF,
-    0x0001E9B20001D9B1, 0x0001F9B40001E9B3, 0x0001E9B60001F9B5,
-    0x0001D9B80001E9B7, 0x0001A9BA0001D9B9, 0x0001B9BC0001A9BB,
-    0x0001A9BE0001B9BD, 0x80002E5E000FB1BF, 0x8000308480002FAD,
-    0x8000346380003447, 0x8000369380003673, 0x800039EB8000396B,
-    0x80003A37800039FB, 0x8000358780002C32, 0x800039D380003957,
-    0x80002BEF8000309C, 0x80002F4D80002C56, 0x8000319C80002F6D,
-    0x800034D3800033CB, 0x8000391F8000377F, 0x800039AB800031DE,
-    0x8000322A80002D12, 0x8000331780003236, 0x800034538000334F,
-    0x8000352F800034AF, 0x80003A4380003613, 0x800039BB800030F0,
-    0x80002F9180002C46, 0x800038E780003337, 0x800030A480003054,
-    0x800039C38000323E, 0x80002DDE80002CF6, 0x800031C780002F7D,
-    0x800032678000321A, 0x8000344F800032FB, 0x8000365F800034C3,
-    0x8000380B800037FF, 0x80002D62800039DF, 0x80002DE280003353,
-    0x0007A997800033CF, 0x0007A9950007A196, 0x0007899300079194,
-    0x003AED9100078192, 0x0007E59F00078D90, 0x0007959D0007E59E,
-    0x000AC19B0007959C, 0x000789990007819A, 0x00072D8700079198,
-    0x00072D8500072D86, 0x000731830024B584, 0x0007018100070982,
-    0x0007ED8F00070980, 0x0007ED8D0007ED8E, 0x00070D8B0007ED8C,
-    0x00070D8900070D8A, 0x0029FDB700070D88, 0x000191B5000199B6,
-    0x000181B3000199B4, 0x000191B1000199B2, 0x00019DBF000199B0,
-    0x00019DBD00019DBE, 0x00019DBB00019DBC, 0x00019DB900019DBA,
-    0x000725A700019DB8, 0x000715A5000725A6, 0x000705A3000715A4,
-    0x000715A1000705A2, 0x0001E1AF000715A0, 0x000711AD000719AE,
-    0x000701AB000719AC, 0x0024B9A9000719AA, 0x80003473000715A8,
-    0x800039C78000378F, 0x80003ADB80003AB3, 0x8000332F80003232,
-    0x800035CB800036CB, 0x8000341F800035EF, 0x8000386B800033A7,
-    0x8000373380002C0B, 0x80002CEA80002CFA, 0x8000311880002FF8,
-    0x80002F39800035FB, 0x800031EA8000330B, 0x800039B3800038EF,
-    0x80002C6680003014, 0x80002E8E80002E3A, 0x80002CD6800031DA,
-    0x800034EB80003427, 0x8000356B8000354F, 0x800035838000356F,
-    0x800035F380003A03, 0x8000385B80003797, 0x8000393B8000390F,
-    0x80003A5B80003A57, 0x80003AC380003A5F, 0x800039CF80003933,
-    0x80002C5A80002C36, 0x80002D2A80002C6E, 0x80002D7680002D36,
-    0x80002E4680002E2A, 0x80002E8A80002E4E, 0x80002F7980002E96,
-    0x800030B880003090, 0x00034180800030C0, 0x0003418200034181,
-    0x0003418400034183, 0x0003418600034185, 0x002BC98800034187,
-    0x002BD18A00034589, 0x0003598C0003498B, 0x800030D80003598D,
-    0x0003319080003184, 0x00033D92800031B4, 0x8000325F800031E2,
-    0x0003259600032D95, 0x0003E5980003DD97, 0x0003E59A0003ED99,
-    0x0003E59C0003FD9B, 0x002BE19E0003ED9D, 0x00031DA080003323,
-    0x000329A28000333F, 0x800033C38000336F, 0x000331A6000331A5,
-    0x80003457800033E3, 0x000305AA80003537, 0x000305AC00031DAB,
-    0x000305AE00030DAD, 0x000305B000037DAF, 0x000305B200030DB1,
-    0x000305B400031DB3, 0x000305B600030DB5, 0x000305B800033DB7,
-    0x000305BA00030DB9, 0x002DE5BC00031DBB, 0x000301BE000301BD,
-    0x8000355B000301BF, 0x8000355F80003557, 0x8000357B80003563,
-    0x800035A38000357F, 0x800035D3800035B7, 0x800036478000363F,
-    0x800036838000365B, 0x80003707800036D3, 0x800038178000376F,
-    0x800038638000382F, 0x800038AB80003867, 0x80003907800038AF,
-    0x80003A27800039E3, 0x8000308880003A3F, 0x800036EB8000EBD0,
-    0x80002CC280002BB7, 0x80002C2A80002C82, 0x80002C9680002C6A,
-    0x80002D4280002D26, 0x80002E2680002E16, 0x80002E9E80002E42,
-    0x80002EF580002EE9, 0x80002F2180002F11, 0x8000301880003010,
-    0x8000306880003048, 0x800030B080003098, 0x800030B4800030A0,
-    0x8000314C800030F4, 0x8000316080003158, 0x8000320E80003188,
-    0x8000336380003222, 0x8000338B8000335B, 0x001F29998000351B,
-    0x001F399B001F2198, 0x001F299D001F219A, 0x001F259F001F219C,
-    0x00273991001F259E, 0x001F3D93001F2990, 0x001F2D95001F3D92,
-    0x00198197001F2D94, 0x001EDD89001F1D96, 0x001EED8B001ED588,
-    0x001F518D001EED8A, 0x001F518F00275D8C, 0x001E5D81001F598E,
-    0x001E2583001E5D80, 0x001E0D85001E2982, 0x001EF587001E2584,
-    0x001F15B9001EF986, 0x001F15BB001D59B8, 0x001F31BD001F15BA,
-    0x001F31BF001DE5BC, 0x001FD9B1001F39BE, 0x001F29B3001FD1B0,
-    0x001F39B5001FD1B2, 0x001F29B7001F31B4, 0x001FA1A9001F31B6,
-    0x001FA1AB001F99A8, 0x001FA1AD001FA9AA, 0x800033E7001FB9AC,
-    0x001FA5A18000341B, 0x001FA5A3001FA5A0, 0x001FA5A5001FA5A2,
-    0x001D5DA7001FA5A4, 0x001EAD80001FA5A6, 0x001EBD82001EB581,
-    0x001EAD84001EB583, 0x001EDD86001ED585, 0x001EED88001ED587,
-    0x001EF98A001CD989, 0x001EE58C001F158B, 0x001EE58E001EED8D,
-    0x001EA590001EDD8F, 0x0014E992001C3191, 0x001EBD94001EA593,
-    0x0018759600328595, 0x001E9198001A2597, 0x001FE59A001E9199,
-    0x001FE19C001E959B, 0x0001A19E001E899D, 0x001CFDA00001A19F,
-    0x00038DA2000385A1, 0x00039DA4000385A3, 0x001C19A6000385A5,
-    0x0003B9A8000389A7, 0x000389AA0003B9A9, 0x000399AC000389AB,
-    0x000395AE001E95AD, 0x001879B000038DAF, 0x001EF9B20003F1B1,
-    0x0003EDB40003F5B3, 0x00039DB6000395B5, 0x001EA5B8000395B7,
-    0x001C49BA0003A9B9, 0x001EB9BC0003BDBB, 0x0003ADBE001715BD,
-    0x80003467001EB1BF, 0x800034A780003483, 0x800034CF800034CB,
-    0x800034FF800034EF, 0x8000350F80003513, 0x800035BB8000353F,
-    0x80003623800035EB, 0x8000373780003653, 0x800037D38000375F,
-    0x800038278000381B, 0x8000384B80003847, 0x8000387380003857,
-    0x8000391B800038EB, 0x80003953000320F0, 0x800039BF80003973,
-    0x80003A3380003A17, 0x0000D98080003A93, 0x0000D582001DD181,
-    0x0000F9840000ED83, 0x0000E9860000F985, 0x001F3588001DC187,
-    0x001DC98A0000D189, 0x001DD58C0000D58B, 0x003A458E0000C98D,
-    0x00002D900000558F, 0x00003D9200003591, 0x00002D9400003593,
-    0x0001D1960001D195, 0x0001F1980001D197, 0x8000EB670001F199,
-    0x8000EB948000EB62, 0x80002B0780002ADF, 0x000F189180002B0B,
-    0x8000EC438000EC16, 0x80003B1F8000ECBB, 0x8000084A80003B2F,
-    0x0000398080000855, 0x0000158200003981, 0x0000758400006D83,
-    0x0000718600007D85, 0x8000084D80000858, 0x8000095A80000851,
-    0x80000FD880000FE7, 0x80000FDD00003D9D, 0x80000FEC0000899F,
-    0x8000104180000FE2, 0x0000219400002193, 0x0000199600002D95,
-    0x00078DA800001997, 0x000799AA00452DA9, 0x000789AC000799AB,
-    0x0007B9AE000789AD, 0x00007DA00007B9AF, 0x004469A2004551A1,
-    0x000401A4000051A3, 0x0007BDA60007B5A5, 0x000785B80007B5A7,
-    0x000785BA00078DB9, 0x000785BC00079DBB, 0x000781BE800010C0,
-    0x0007C9B080001077, 0x0007D9B20007C9B1, 0x0007C9B40007D9B3,
-    0x0007B9B60007C9B5, 0x8000105D80001025, 0x8000109380001065,
-    0x800010AE800010B8, 0x800010A0800010B3, 0x80000FF4800010A7,
-    0x80000FFE80000FF9, 0x800010188000100B, 0x8000102880001020,
-    0x8000103780001032, 0x800010468000103C, 0x800010538000104B,
-    0x8000106880001060, 0x800010728000106D, 0x8000107F8000107A,
-    0x8000108E80001089, 0x8000109B80001096, 0x8000102D800010BB,
-    0x8000105880001010, 0x8000100380001084, 0x800017DC800017D3,
-    0x800017E5800017DF, 0x800017E2800017D9, 0x80001809800017D6,
-    0x800017EB8000180C, 0x800017EE800017E8, 0x025374D9800017F1,
-    0x800017F7800017FA, 0x800017F4800017FD, 0x8000180080001803,
-    0x800018158000180F, 0x800018188000181B, 0x800018218000181E,
-    0x00015D808000182A, 0x8000182400016581, 0x0001798400016983,
-    0x0001758680001812, 0x0001558800016D87, 0x0001558A00015D89,
-    0x0001B58C0001AD8B, 0x0001B58E0001BD8D, 0x0001D5900001AD8F,
-    0x0001D1920001D191, 0x0001C9940001D593, 0x0001C5960001CD95,
-    0x0001FD980001C197, 0x0001C99A0001F999, 0x0001D19C0001CD9B,
-    0x0001DD9E0001D59D, 0x000125A00001D99F, 0x000121A2000121A1,
-    0x000139A4000125A3, 0x000135A600013DA5, 0x00010DA8000131A7,
-    0x000119AA000109A9, 0x000101AC00011DAB, 0x00010DAE000105AD,
-    0x000175B0000109AF, 0x000171B2000171B1, 0x000169B4000175B3,
-    0x000165B600016DB5, 0x00015DB8000161B7, 0x000129BA000159B9,
-    0x000131BC00012DBB, 0x00013DBE000135BD, 0x80001836000139BF,
-    0x0253E1B080001833, 0x8000183E0253E5B1, 0x8000183080001844,
-    0x8000184A80001841, 0x800010D78000176F, 0x8000110480001127,
-    0x8000111380001118, 0x0002B9A28000111D, 0x0002B1A00002BDA3,
-    0x0002ADA60002B5A1, 0x0253A1A40002A9A7, 0x000369AA0253A5A5,
-    0x000295A800036DAB, 0x0001DDAE000291A9, 0x000371AC0001D9AF,
-    0x00026992000375AD, 0x00026D9000026D93, 0x00027D9600026991,
-    0x0002719400027997, 0x0002519A00027595, 0x000245980002559B,
-    0x0002459E00024199, 0x0002499C0002419F, 0x00020D8200024D9D,
-    0x00020D8000020983, 0x0002058600020981, 0x0002098400020187,
-    0x00023D8A00020D85, 0x000231880002398B, 0x0002158E00023589,
-    0x0002198C0002118F, 0x00011DA300021D8D, 0x000111A1000119A2,
-    0x000131A7000115A0, 0x000139A5000135A6, 0x000109AB00013DA4,
-    0x00010DA900010DAA, 0x000101AF000109A8, 0x000115AD000105AE,
-    0x000179B3000111AC, 0x00017DB100017DB2, 0x0000D1B7000179B0,
-    0x000165B50000D5B6, 0x0000E5BB000161B4, 0x0000EDB90000E1BA,
-    0x0000F9BF0000EDB8, 0x0000F1BD0000FDBE, 0x800011220000F5BC,
-    0x8000184780001109, 0x800010E1800010DC, 0x025028D9800010F5,
-    0x8000114E8000110E, 0x8000115F80001153, 0x800010508000184D,
-    0x8000118480001175, 0x0002D59380001189, 0x800011AE80001194,
-    0x0001D197800011C1, 0x0002CD950002C196, 0x0001F19B0002C994,
-    0x0001ED990001E19A, 0x0001E59F0001ED98, 0x0252319D0001E19E,
-    0x000F8D800001ED9C, 0x000F9D82000F9581, 0x000F8D84000F8983,
-    0x000FBD86000FB585, 0x000F9588000FB587, 0x000F958A000F9D89,
-    0x000FF98C000FE98B, 0x0001398E000FF98D, 0x000149900001398F,
-    0x0001599200014991, 0x0001499400015993, 0x0001799600014995,
-    0x0001499800017997, 0x0001599A00014999, 0x0001499C0001599B,
-    0x0001399E0001499D, 0x0001C9A00001399F, 0x0001D9A20001C9A1,
-    0x0001C9A40001D9A3, 0x0001F9A60001C9A5, 0x0001C9A80001F9A7,
-    0x0001D9AA0001C9A9, 0x0001C9AC0001D9AB, 0x000139AE0001C9AD,
-    0x000149B0000139AF, 0x000159B2000149B1, 0x000149B4000159B3,
-    0x000179B6000149B5, 0x000149B8000179B7, 0x000159BA000149B9,
-    0x000149BC000159BB, 0x000139BE000149BD, 0x800011E5000139BF,
-    0x8000121C80001217, 0x8000123380001224, 0x8000124780001242,
-    0x800012728000124F, 0x800012A580001299, 0x800012CA800012C5,
-    0x80001309800012CF, 0x800013288000131C, 0x800013BA80001340,
-    0x8000142980001410, 0x800014418000142E, 0x8000146480001452,
-    0x8000149080001469, 0x800014AD80001498, 0x800014D9800014D4,
-    0x8000153880001533, 0x800015498000153D, 0x8000155A80001555,
-    0x8000156E80001562, 0x8000158D80001588, 0x8000159A80001595,
-    0x800015A48000159F, 0x800015AE800015A9, 0x800015C6800015C1,
-    0x800015FC800015E2, 0x8000162280001616, 0x8000163F8000163A,
-    0x8000166D8000164C, 0x800016AA80001690, 0x800016BB800016B6,
-    0x800016E4800016C3, 0x0003C980800016FE, 0x0003D9820003C981,
-    0x0003C9840003D983, 0x0003F9860003C985, 0x0003C9880003F987,
-    0x0003D98A0003C989, 0x0003C98C0003D98B, 0x0001398E0003C98D,
-    0x000149900001398F, 0x0001599200014991, 0x0001499400015993,
-    0x0001799600014995, 0x0001499800017997, 0x0001599A00014999,
-    0x0001499C0001599B, 0x0001399E0001499D, 0x0001C9A00001399F,
-    0x0001D9A20001C9A1, 0x0001C9A40001D9A3, 0x000D0DA60001C9A5,
-    0x000D25A80001FDA7, 0x0001CDAA000D39A9, 0x000D05AC0001C5AB,
-    0x000D01AE0001D9AD, 0x0001B1B0000D01AF, 0x0003C9B20001B1B1,
-    0x0003D5B40001B5B3, 0x0001A5B60003DDB5, 0x0003F1B800015DB7,
-    0x0003FDBA000161B9, 0x00037DBC0003F5BB, 0x000365BE000365BD,
-    0x8000170D00035DBF, 0x8000172F8000172A, 0x8000173C80001737,
-    0x800017548000174F, 0x800017868000177A, 0x800017A180001792,
-    0x800017C3800017BE, 0x800012F5800012E4, 0x8000008180001772,
-    0x800000958000008B, 0x800000A98000009F, 0x800010EB800000B3,
-    0x800010FA800010F0, 0x800011708000116B, 0x800011DB8000117A,
-    0x8000120D800011E0, 0x8000122E80001229, 0x8000164780001238,
-    0x8000170880001703, 0x8000179780001720, 0x800017B48000179C,
-    0x800010FF800010E6, 0x800012128000117F, 0x80001635800013CD,
-    0x8000175980001725, 0x8000123D800017B9, 0x800013958000135A,
-    0x80001507800013A8, 0x8000151F80001513, 0x8000148880001483,
-    0x800014CC800014C7, 0x800014F8800014F3, 0x0005B9828000135F,
-    0x0005AD800005B983, 0x0005C5860005B981, 0x0005D9840005CD87,
-    0x0007AD8A0005CD85, 0x0007A1880007B58B, 0x0007358E0005C589,
-    0x0007398C00073D8F, 0x0007DD920007A98D, 0x0007759000076D93,
-    0x000761960007D591, 0x0007F994000BD997, 0x000BF59A00077195,
-    0x000BF9980007CD9B, 0x000BD99E0007C999, 0x000BD99C000BC19F,
-    0x000B49A2000BD99D, 0x000725A0000B49A3, 0x000581A6000B49A1,
-    0x000599A4000589A7, 0x000589AA00072DA5, 0x000589A8000589AB,
-    0x00058DAE000589A9, 0x00058DAC0005F5AF, 0x000775B2000585AD,
-    0x00058DB0000589B3, 0x000589B6000585B1, 0x000589B4000589B7,
-    0x00058DBA000589B5, 0x000589B8000595BB, 0x00058DBE000585B9,
-    0x00058DBC000575BF, 0x00048580000585BD, 0x00049D8200048D81,
-    0x0004B98400049583, 0x0004A9860004B985, 0x000499880004A987,
-    0x0004E18A0004E189, 0x0004E18C0004E18B, 0x0004E58E00068D8D,
-    0x0004A5900004DD8F, 0x0004BD920004AD91, 0x0004AD940004B593,
-    0x0006E59600065595, 0x0006719800065197, 0x0006759A0006A599,
-    0x0006559C00064D9B, 0x0006BD9E00065D9D, 0x000629A0000AE99F,
-    0x000625A2000A31A1, 0x000621A400048DA3, 0x000651A6000485A5,
-    0x00066DA80004D9A7, 0x00067DAA000675A9, 0x000409AC000419AB,
-    0x0006B9AE000465AD, 0x0006E9B0000601AF, 0x000611B2000601B1,
-    0x000631B4000611B3, 0x000631B6000631B5, 0x000611B8000631B7,
-    0x000611BA000611B9, 0x0003F1BC000611BB, 0x0003F5BE0003F1BD,
-    0x800013640003EDBF, 0x00FE8084800013AD, 0x800012B8800013B2,
-    0x8000128C800012BD, 0x800012D480001291, 0x8000141C800012D9,
-    0x8000145780001421, 0x8000136C8000145C, 0x8000138B80001378,
-    0x8000133B80001390, 0x8000144D800013D2, 0x8000119980001141,
-    0x800011BA800011B3, 0x800011EA800011C6, 0x800011F8800011F1,
-    0x800012B180001277, 0x80001321800012AA, 0x800013158000130E,
-    0x800013458000134C, 0x800013BF80001353, 0x8000137D80001415,
-    0x8000139A80001371, 0x80001433800013A1, 0x8000146E80001446,
-    0x8000147C80001475, 0x800014B28000149D, 0x800014DE800014B9,
-    0x800014E5800014EC, 0x8000157380001542, 0x800016018000157A,
-    0x800016088000160F, 0x8000161B800015E7, 0x0001158080001627,
-    0x0001158200011D81, 0x0001358400012D83, 0x0001358600013D85,
-    0x0001158800012D87, 0x0001158A00011D89, 0x0004E18C00016D8B,
-    0x0001718E0001718D, 0x000111900001718F, 0x0004A1920004A191,
-    0x0004A1940004A193, 0x0004A1960004A195, 0x000161980004A197,
-    0x0001659A00016199, 0x0001659C00017D9B, 0x0001659E00016D9D,
-    0x0001A5A000015D9F, 0x0001A5A20001ADA1, 0x0001A5A40001BDA3,
-    0x0001A5A60001ADA5, 0x000471A80001DDA7, 0x0001E1AA0001E1A9,
-    0x0001E1AC0001E1AB, 0x000185AE00018DAD, 0x000411B00001FDAF,
-    0x00041DB2000409B1, 0x0006E9B40006A9B3, 0x000699B60006E9B5,
-    0x0001DDB80001E5B7, 0x00068DBA0001A5B9, 0x0001A5BC000685BB,
-    0x800016720001A1BD, 0x8000168980001679, 0x8000165F80001651,
-    0x8000169C80001695, 0x8000174180001658, 0x800016E980001748,
-    0x800016CF800016F0, 0x80001719800016D6, 0x800017A680001712,
-    0x00FE299580001164, 0x00FE399700FE2194, 0x00FE0D9100FE2196,
-    0x00FE399300FE0D90, 0x00FEE59D00FE0192, 0x00FEE59F00FEE59C,
-    0x00FE059900FEE59E, 0x00FE059B00FE0198, 0x00FE19A500FE059A,
-    0x00FE29A700FE1DA4, 0x00FE01A100FE11A6, 0x00FE0DA300FE19A0,
-    0x00FE0DAD00FE09A2, 0x00FE1DAF00FE09AC, 0x00FE15A900FE1DAE,
-    0x00FE11AB00FE15A8, 0x00FE05B500FE15AA, 0x00FE01B700FE05B4,
-    0x00FE69B100FE05B6, 0x00FE65B300FE61B0, 0x00FE3DBD00FE65B2,
-    0x00FE0DBF00FE39BC, 0x00FE21B900FE0DBE, 0x00FE21BB00FE39B8,
-    0x00FFE98000FE29BA, 0x00FFF98200FFE981, 0x00FFE58400FFF983,
-    0x00FFE18600FFED85, 0x00FFDD8800FFE187, 0x00FED58A00FEDD89,
-    0x0001358C00012D8B, 0x0001358E00013D8D, 0x800011A700012D8F,
-    0x00015D92800011A0, 0x00014D9400015593, 0x00017D9600017595,
-    0x0001499800017597, 0x0001599A00014999, 0x0001499C0001599B,
-    0x0001459E00014D9D, 0x0000C5A000003D9F, 0x00038DA2000385A1,
-    0x00039DA4000385A3, 0x000381A6000385A5, 0x000381A8000381A7,
-    0x000381AA000381A9, 0x000381AC000381AB, 0x000381AE000381AD,
-    0x000381B0000381AF, 0x000381B2000381B1, 0x00FF31B4000381B3,
-    0x000389B600FF3DB5, 0x0003B9B8000389B7, 0x000389BA0003B9B9,
-    0x000391BC000389BB, 0x000391BE0003ADBD, 0x800011D4000391BF,
-    0x80001206800011CD, 0x80001285800011FF, 0x80001254000EBDA7,
-    0x8000132D8000127E, 0x80001384800013C6, 0x800015F58000143A,
-    0x8000178B8000162E, 0x800017AD8000177F, 0x80001581800016AF,
-    0x800014C0800016F7, 0x800016C8800015BA, 0x800015EE800016A3,
-    0x8000125B800015B3, 0x800016668000129E, 0x800011588000154E,
-    0x800016DD80001334, 0x8000156780001409, 0x8000112F80001138,
-    0x800013D780001680, 0x800014A4800012EC, 0x800013E080001761,
-    0x80001262800013E7, 0x800003A6800012FA, 0x8000240F8000240B,
-    0x800000C7800006A0, 0x800024A9800006B5, 0x800022B3800024AD,
-    0x800007E6800022AF, 0x800009828000097E, 0x800024A58000243F,
-    0x8000243780002433, 0x000171808000241B, 0x0001718200017181,
-    0x00FDD58400016983, 0x0001658600FD8985, 0x8000241F00019D87,
-    0x8000242780002423, 0x8000242F8000242B, 0x800007E2800007DE,
-    0x800000D7800000D1, 0x800003A8800003A2, 0x800000D3800007E0,
-    0x800006BD800000D5, 0x800014FD80000079, 0x800000878000007D,
-    0x8000150280000091, 0x8000150E8000009B, 0x8000151A800000A5,
-    0x80001526800000AF, 0x8000152B800000B9, 0x8000114B800010C5,
-    0x800011918000118E, 0x8000124C80001221, 0x800012C280001296,
-    0x800012E1800012DE, 0x000145B0800012E9, 0x000145B200014DB1,
-    0x000145B400015DB3, 0x000145B600014DB5, 0x000145B800017DB7,
-    0x000145BA00014DB9, 0x000145BC00015DBB, 0x8000130680001303,
-    0x800006A880001369, 0x000AE0F0800006B3, 0x800013B78000112C,
-    0x8000146180001426, 0x800014958000148D, 0x80001530800014D1,
-    0x800015928000155F, 0x80001644800015CB, 0x00069598800016C0,
-    0x8000173400B18199, 0x800015CE8000175E, 0x80001777800015D3,
-    0x0006D990800015D8, 0x0006A9920006D991, 0x0292A9940006A993,
-    0x0006B5960006BD95, 0x800015DD0006AD97, 0x800000D9800000CF,
-    0x800007E4800003B3, 0x800023F580000980, 0x80002A7F800023F9,
-    0x8000256C80002533, 0x8000259D8000258B, 0x800029C1800025C2,
-    0x000629B8800029DD, 0x000639BA000629B9, 0x000629BC000639BB,
-    0x0007F9BE000629BD, 0x00B121B000B879BF, 0x000635B200063DB1,
-    0x000611B400060DB3, 0x00B24DB600B245B5, 0x00B9B980000619B7,
-    0x0006218200062181, 0x0006218400062183, 0x80002786800029EF,
-    0x0006198800062987, 0x00B0F58A00B0F989, 0x00B0ED8C00B0F18B,
-    0x0007D98E0007D58D, 0x0007F9900007DD8F, 0x00B0F5920007F991,
-    0x0293899480002A83, 0x0007959600079195, 0x0007B99800079597,
-    0x00B3DD9A00B3D599, 0x0007999C0007899B, 0x0007899E0007999D,
-    0x0006C1A00006419F, 0x00B321A20006C1A1, 0x000191A40006C5A3,
-    0x00B241A6000191A5, 0x0006EDA880002A67, 0x0006FDAA0006F5A9,
-    0x800025270006F5AB, 0x800009848000252B, 0x00069DB080000987,
-    0x0006ADB20006A5B1, 0x0006B9B480000990, 0x0006B5B68000098D,
-    0x000695B80006ADB7, 0x000695BA00069DB9, 0x000675BC00066DBB,
-    0x000675BE00067DBD, 0x02A3918200066DBF, 0x00059D8002A39183,
-    0x02A3B98602A39D81, 0x02A38D8402A38D87, 0x02A3958A02A3B585,
-    0x02A3B18802A3918B, 0x00023D8E02A39989, 0x02A38D8C0005AD8F,
-    0x0005D9920002318D, 0x0005D1900005D993, 0x0005F1960005D591,
-    0x0005C5940005F597, 0x0005C59A0005FD95, 0x0005C9980005C19B,
-    0x0005D99E0005C999, 0x0005DD9C0005DD9F, 0x00052DA20005D59D,
-    0x000521A0000529A3, 0x000521A6000521A1, 0x000535A4000525A7,
-    0x000515AA00052DA5, 0x000519A800051DAB, 0x00050DAE000519A9,
-    0x000501AC000541AF, 0x000539B2000501AD, 0x00053DB000053DB3,
-    0x00052DB6000535B1, 0x000521B4000529B7, 0x0002F1BA000521B5,
-    0x000515B80002F5BB, 0x0002E5BE0002FDB9, 0x0002E9BC0002E1BF,
-    0x00D51D920002E9BD, 0x8000098A00D52593, 0x00D52196800022E6,
-    0x00D53D9400D52197, 0x00D5199A00D53595, 0x02A0E4D900D5199B,
-    0x0285C0CC800023E9, 0x00D5719C0285FCCC, 0x00D7A18202A0DCD9,
-    0x800022EE00D7AD83, 0x00D7A986800022F6, 0x00D7A98400D7A987,
-    0x00D5618A00D7A985, 0x02A084D900D5618B, 0x00D5618E800023ED,
-    0x00D5718C00D5618F, 0x800023F100D5718D, 0x80000DB980000DB6,
-    0x80000CCE80000D7F, 0x80003B5B80000D94, 0x80000D9780000D9A,
-    0x80000CDA80000CD7, 0x80000CDD80001BB3, 0x80000DA680000CE9,
-    0x00020DA20287B5BD, 0x000271A0029311A3, 0x0000DDA6000271A1,
-    0x00020DA480000CFB, 0x0000A9AA0000D5A5, 0x0000C5A8000BE5AB,
-    0x000071AE000801A9, 0x0000B5AC80000CF5, 0x000C218B000091AD,
-    0x000C2989000C258A, 0x000C358F000C1188, 0x000C3D8D000C318E,
-    0x000C3983000C3D8C, 0x000C3181000C3D82, 0x000C2D87000C3980,
-    0x000C2585000C2986, 0x000C719B000C2584, 0x000C7999000C759A,
-    0x000C659F000C6198, 0x000C6D9D000C619E, 0x000C4993000C6D9C,
-    0x000C4191000C4D92, 0x000C5D97000C4990, 0x000C5595000C5996,
-    0x000CC9AB000C5594, 0x000CC1A9000CCDAA, 0x00FEB5AF000CB1A8,
-    0x000CD5AD000CD9AE, 0x000C99A3000CD5AC, 0x000C91A1000C9DA2,
-    0x000C8DA7000C99A0, 0x000C85A5000C89A6, 0x000C41BB000C85A4,
-    0x000C95B9000C99BA, 0x0288D9BF000C99B8, 0x80000D8280000CEF,
-    0x000CB9B3000C5DBC, 0x000CB1B1000CBDB2, 0x000CA5B700FEC9B0,
-    0x000CA9B5000CA5B6, 0x000BC981000CA5B4, 0x000A358380000D85,
-    0x000A4585000D8982, 0x000D9187000A4584, 0x00BF9589000A5D86,
-    0x00BF818B00BF9588, 0x000A798D000A698A, 0x000DB58F000B8D8C,
-    0x029DB59100BC818E, 0x029DA99300BCB190, 0x00BCAD95029D9192,
-    0x00BFD59700BCAD94, 0x00BFC59900BFD596, 0x029FB19B00BFC598,
-    0x00BEA99D000BA99A, 0x000BB59F000D759C, 0x00BF01A1000D799E,
-    0x00B195A3000ADDA0, 0x00BD81A500BD85A2, 0x00BD81A700BD99A4,
-    0x00BD89A900B1A9A6, 0x00BD99AB00BD85A8, 0x00BDF1AD00BD89AA,
-    0x00BDE9AF00BD8DAC, 0x00BF21B100BDF5AE, 0x00C949B300BD9DB0,
-    0x00BDA1B500BDADB2, 0x00BDA1B700BDBDB4, 0x00C969B900B149B6,
-    0x000ADDBB00B14DB8, 0x000AC9BD00B1CDBA, 0x000B19BF000AB9BC,
-    0x00BEFD80000D6DBE, 0x00BE6582029ED581, 0x00BE918400BEC583,
-    0x00BEA58600BE9D85, 0x0291AD8800BE8D87, 0x0292298A00BEE589,
-    0x0292118C00BEB58B, 0x00B22D8E00BE958D, 0x00BEE990029EC58F,
-    0x0292419200B5B191, 0x00BED594029FF193, 0x00BEE19600BEE595,
-    0x0292659802924D97, 0x00BEB99A02927D99, 0x000EE59C0008899B,
-    0x80000D0400089D9D, 0x000E1DA0000EE59F, 0x0009ADA2000845A1,
-    0x00082DA40009A5A3, 0x00D605A6000E35A5, 0x000E09A8000E39A7,
-    0x000E19AA000E09A9, 0x000E09AC000E19AB, 0x000FADAE000E09AD,
-    0x000F8DB0000FA5AF, 0x00D39DB200D395B1, 0x00D389B400D485B3,
-    0x00D6D9B600D389B5, 0x00D6E9B800D6D9B7, 0x00D6F9BA00D6E9B9,
-    0x00D6E9BC00D6F9BB, 0x00D6B9BE00D6E9BD, 0x00D7498000D6B9BF,
-    0x00D7598200D74981, 0x00D7498400D75983, 0x00D27D8600D27585,
-    0x00D24D8800D27587, 0x00D7758A00D25589, 0x00D7B58C00D7AD8B,
-    0x00D7B58E00D7BD8D, 0x00D2399000D2498F, 0x00D2C99200D23991,
-    0x00D2D99400D2C993, 0x00D2C99600D2D995, 0x00D2F99800D2C997,
-    0x00D2C99A00D2F999, 0x00D7E19C00D2C59B, 0x000EC99E000EE99D,
-    0x00D9D1A0000EC99F, 0x00DBC1A200DBC1A1, 0x00DBC5A400DBDDA3,
-    0x00DBC5A600DBCDA5, 0x00DB25A800DBFDA7, 0x00DBCDAA00DB65A9,
-    0x00DBDDAC00DBC5AB, 0x00DBC1AE00DBC1AD, 0x00DB09B000DB79AF,
-    0x00DB19B200DB09B1, 0x00DB09B400DB19B3, 0x00DEC1B600D8B5B5,
-    0x00DFE5B800DFDDB7, 0x00DFFDBA00DFF5B9, 0x00DFB5BC00DFADBB,
-    0x00DFADBE00DFA5BD, 0x800022F280000D55, 0x00DC289080000DA9,
-    0x80000A0100000491, 0x80000D1380000DAC, 0x80003B4F8000EADB,
-    0x8000EAE080000D19, 0x0012609E80000D6A, 0x8000EAE50011EC9F,
-    0x80000A340000889D, 0x80000D3480000D31, 0x8000EAEA80000D3D,
-    0x80000D4680000D43, 0x80000DA380000D49, 0x80003B5F80000D9D,
-    0x80000D5880000DA0, 0x80000D6D80002403, 0x80000D9180000D8E,
-    0x80000BB180000D7C, 0x80000BB780000BB4, 0x8000EAF48000EAEF,
-    0x8000EA068000EA01, 0x8000EA108000EA0B, 0x00FD14B100020891,
-    0x000079A700029084, 0x8000EA15000B74B2, 0x000240F00003CC87,
-    0x0002F08C00022C91, 0x000039BE00026C8D, 0x8000EA1A00BC40CC,
-    0x8000EA1F00000D97, 0x00024CF00003DCB9, 0x00024C9100299CAF,
-    0x000A858100003892, 0x0001198300DC38A5, 0x000A8185000A8582,
-    0x000A8D87000A9984, 0x000ABD898000EA29, 0x000A8D8B000ABD88,
-    0x000A9D8D000A8D8A, 0x000A8D8F000A9D8C, 0x000A9591000A8D8E,
-    0x00059D93000AFD90, 0x00BB8995000AB992, 0x0007219700058D94,
-    0x0001659900045996, 0x0001719B00069998, 0x0001519D0001799A,
-    0x0001519F0001699C, 0x0001B9A10001599E, 0x0001CDA30001A9A0,
-    0x00B265A50001CDA2, 0x0001C9A70001DDA4, 0x0001F9A90001D1A6,
-    0x0001C9AB0001F1A8, 0x0001D9AD0001F1AA, 0x0001C9AF0001D1AC,
-    0x8000EA240001D1AE, 0x0001B5B30001B1B0, 0x000155B50001B5B2,
-    0x000155B7000155B4, 0x000175B9000155B6, 0x8000EA2E000175B8,
-    0x00024CF0000175BA, 0x0001F49100DC2C99, 0x0001908200DC409B,
-    0x0003D4B9000319B0, 0x000005AF00003CB8, 0x00006C848000EA33,
-    0x0003C4B5000228F0, 0x8000EA38000214F0, 0x0003D9BD0003EDBA,
-    0x8000EA428000EA3D, 0x000048920000E48D, 0x0002F0A400006DAF,
-    0x0002BC968000EA4B, 0x8000EA540002F896, 0x8000EA6E8000EA61,
-    0x8000EA7B00029C96, 0x00024C918000EA88, 0x00004DB00002AC91,
-    0x8000EAB88000EA95, 0x001079908000EA9E, 0x8000EAAB8000EAC1,
-    0x00026CA48000EACE, 0x0003D88580000AD6, 0x0002188700000086,
-    0x0000258000034C86, 0x0000008100000082, 0x002DC09D00000080,
-    0x0011689F002C009C, 0x0015D099002ECC9E, 0x0017D89B00140098,
-    0x0019E0950016DC9A, 0x001BE89700180094, 0x001DF091001AEC96,
-    0x001FF893001C0090, 0x0001399E001EFC92, 0x000000ED0001499F,
-    0x000191BC000171BB, 0x000195BE000191BD, 0x0298ACCC00018DBF,
-    0x80000CC2029EA4CD, 0x00BEDCCC80000DE5, 0x80000DF180000DEB,
-    0x80000E0080000DFD, 0x80000E0680000E03, 0x80000E1880000E09,
-    0x80000E2180000E1E, 0x8000231A80000E24, 0x80000E5D80000E4E,
-    0x029EF8CC80000E83, 0x80000E86029EFCCD, 0x80000E9880000E95,
-    0x8000231280000EA4, 0x000000CB80000EB9, 0x80000EDA80000EBC,
-    0x80000EE380000EDD, 0x80000EFB80000EF8, 0x80000F0780000F04,
-    0x80000F0D80000F0A, 0x80000F13000000C0, 0x80000F1980000F16,
-    0x80000F2E80000F1F, 0x80000F4380003B43, 0x80000F6A80000F5E,
-    0x000185A480000F58, 0x000191A2800022EA, 0x0001A1A000019DA3,
-    0x00A425BD0001A1A1, 0x00A431BF00A42DBC, 0x00A825B900A43DBE,
-    0x00A439BB00A405B8, 0x00A429B500A405BA, 0x00A425B700A6A5B4,
-    0x028835B100A831B6, 0x00A4E9B3028825B0, 0x00A4C5AD028825B2,
-    0x00A4F1AF0285F9AC, 0x00A4F9A900A4FDAE, 0x028851AB0284ADA8,
-    0x028841A500AFA9AA, 0x00A87DA700A4FDA4, 0x028BFDA100A4CDA6,
-    0x028879A300A4E5A0, 0x00A4019D00A4BDA2, 0x00A4359F00A44D9C,
-    0x00A8B19900A4059E, 0x00A4F59B00D01D98, 0x00A8359502844D9A,
-    0x00A8019700A4D594, 0x00A4D19100A4D596, 0x00A4D19300A65590,
-    0x00A4898D00D03D92, 0x00A4918F00A4858C, 0x00A4F58900A49D8E,
-    0x00A4E18B00A4FD88, 0x00A8D58500A4ED8A, 0x00A4C98700A4F584,
-    0x00A4F98100A4F586, 0x00A4F58300A67580, 0x00A5418000A8E182,
-    0x00A5598200A54D81, 0x00A7F18400A55583, 0x00D1998600A57585,
-    0x00A5518800A57587, 0x00A5518A00A9B189, 0x00D1B98C00A9858B,
-    0x0285E98E00A9158D, 0x00A5899000A5518F, 0x00A5C19200A5C591,
-    0x00A5C19400A5F193, 0x00A5999680000F3A, 0x00A5F99802895D97,
-    0x00A5C99A02894599, 0x0285899C00A9799B, 0x00AE8D9E00A5DD9D,
-    0x02843DA00289759F, 0x00A539A200A501A1, 0x028981A400A535A3,
-    0x028981A6028991A5, 0x00A721A800A54DA7, 0x00A9B5AA00A5ADA9,
-    0x00A5A1AC00A5A1AB, 0x00A5A1AE00A981AD, 0x00A5E9B000A59DAF,
-    0x00A5F9B200A5E1B1, 0x00A591B400A5F5B3, 0x00A589B600A59DB5,
-    0x00A701B800A585B7, 0x00D169BA00A585B9, 0x00A581BC00A585BB,
-    0x00A581BE00A961BD, 0x00D2898000A955BF, 0x0286D98200AA2581,
-    0x00A6D98400A66183, 0x00A6918600A69585, 0x00A6B18800A6A187,
-    0x00A6E98A0289A989, 0x00A6A98C028A2D8B, 0x00A6998E028A158D,
-    0x0286B99000AA298F, 0x00ADBD9200A6ED91, 0x0287ED94028A4593,
-    0x00A6E99600A6D195, 0x028A719800A6E597, 0x028A719A028A6199,
-    0x00A4F19C00A6BD9B, 0x00AA659E80000F4F, 0x80000F5B00A6719F,
-    0x00A691A280003B3F, 0x800017CD80000F61, 0x00A6A9A600A6B1A5,
-    0x800017D080001806, 0x00A6F9AA00A6EDA9, 0x00A451AC00A6F5AB,
-    0x00D239AE800003BA, 0x00A6B1B000A6D5AF, 0x00A6B1B200AA51B1,
-    0x00D259B400AA65B3, 0x028609B600AAF5B5, 0x00A629B800A6B1B7,
-    0x800003B700A665B9, 0x800003CA00A651BB, 0x00A639BE028979BD,
-    0x00A79980028AFDBF, 0x00A7A982028B2581, 0x800004ED00AB1983,
-    0x00ACED8600A7BD85, 0x02869D88028B1587, 0x00A7998A00A7A189,
-    0x028B218C00A7958B, 0x028B218E028B318D, 0x00A5C19000A7ED8F,
-    0x00AB559200A74D91, 0x00A7419400A74193, 0x00A7419600AB6195,
-    0x00A7499800A77D97, 0x00A7599A00A74199, 0x00A7319C00A7559B,
-    0x00A7299E00A73D9D, 0x00A561A000A7259F, 0x00D309A200A7E5A1,
-    0x00A7E1A400A7E5A3, 0x00A7E1A600AB01A5, 0x00D329A800AB35A7,
-    0x028779AA00AB85A9, 0x00A779AC00A7C1AB, 0x00A731AE00A735AD,
-    0x00A751B000A701AF, 0x00A709B2028849B1, 0x00A749B4028BCDB3,
-    0x00A779B6028BF5B5, 0x028719B800ABC9B7, 0x00AC1DBA00A74DB9,
-    0x02864DBC028BE5BB, 0x00A749BE00A771BD, 0x00A009B900A745BF,
-    0x00A005BB00A285B8, 0x00AC25BD8000058C, 0x800005DF00A005BC,
-    0x00A0A5B100A005BE, 0x00A091B3028199B0, 0x028C35B500A09DB2,
-    0x00A0E9B7028C25B4, 0x028C61A9028C25B6, 0x00AC5DAB00A0DDA8,
-    0x00A0F9AD00A0EDAA, 0x028C51AF0280ADAC, 0x00A0E1A100ABA9AE,
-    0x00A0D5A300A0ADA0, 0x028FFDA500A0E5A2, 0x028C79A700A0E5A4,
-    0x00AC159900A0BDA6, 0x00AC219B00A0F598, 0x8000061A00A0F59A,
-    0x00A0F59F00D41D9C, 0x00A0E99102804D9E, 0x00A0F19300A0E590,
-    0x8000064000A0FD92, 0x00A0D19700A25594, 0x00ACF58900D43D96,
-    0x8000065700A0D588, 0x00A0F58D00A0D58A, 0x00A0E18F80000674,
-    0x028CE58100A0ED8E, 0x00A03983028CF580, 0x00A0F985028CF582,
-    0x00A0F58700A27584, 0x00A129808000068B, 0x00A1398200A12181,
-    0x00A1518400A13583, 0x00A1498680000138, 0x800001408000013C,
-    0x00D5A98A80000144, 0x00A1418C00A1458B, 0x00A1418E00ADA18D,
-    0x00D5C99000AD958F, 0x0281999280000148, 0x00A1999400A12193,
-    0x00A1D19600A1D595, 0x00A1F19800A1E197, 0x00A1A99A028EE999,
-    0x00A1E99C028D6D9B, 0x00A1D99E028D559D, 0x028179A000AD699F,
-    0x00AA7DA200A12DA1, 0x02802DA4028D85A3, 0x00A129A600A111A5,
-    0x028DB1A800A125A7, 0x028DB1AA028DA1A9, 0x00A331AC00A17DAB,
-    0x00ADA5AE00A1BDAD, 0x00A1D1B000A1B1AF, 0x00A1D1B200ADF1B1,
-    0x00A1F9B400A1EDB3, 0x00A1E9B600A1F1B5, 0x00A1A1B800A1E5B7,
-    0x00A1B9BA00A1ADB9, 0x00A311BC00A1B5BB, 0x00D579BE00A195BD,
-    0x00A2718000A195BF, 0x00A2718200AE9181, 0x00D6998400AEA583,
-    0x0282C98600AE3585, 0x00A2E98800A27187, 0x00A2A18A00A2A589,
-    0x00A2A18C00A2918B, 0x00A2F98E028DB98D, 0x00A2D990028E3D8F,
-    0x00A2E992028E6591, 0x0282A99400AE5993, 0x00A9AD9600A2FD95,
-    0x0283DD98028E5597, 0x00A2D99A00A2E199, 0x028E619C00A2D59B,
-    0x028E619E028E719D, 0x00A001A000A2AD9F, 0x00AE95A200A28DA1,
-    0x00A281A400A281A3, 0x00A281A600AEA1A5, 0x00A289A800A2BDA7,
-    0x00A299AA00A281A9, 0x00A2F1AC00A295AB, 0x00A2E9AE00A2FDAD,
-    0x00A021B000A2E5AF, 0x00D649B200A2A5B1, 0x00A2A1B400A2A5B3,
-    0x00A2A1B600AE41B5, 0x00D669B800AE75B7, 0x028239BA00AEC5B9,
-    0x00A239BC00A281BB, 0x00A271BE00A275BD, 0x00A3918000A241BF,
-    0x00A3C982028C8981, 0x00A38984028F0D83, 0x00A3B986028F3585,
-    0x0283D98800AF0987, 0x00A8DD8A00A38D89, 0x02828D8C028F258B,
-    0x00A3898E00A3B18D, 0x028F519000A3858F, 0x028F5192028F4191,
-    0x00A1D19400A39D93, 0x00AF459600A35D95, 0x00A3719800A35197,
-    0x00A3719A00AF5199, 0x00A3599C00A34D9B, 0x00A3499E00A3519D,
-    0x00A3C1A000A3459F, 0x00A3D9A200A3CDA1, 0x00A171A400A3D5A3,
-    0x00D719A600A3F5A5, 0x00A3D1A800A3F5A7, 0x00A3D1AA00AF31A9,
-    0x00D739AC00AF05AB, 0x028369AE00AF95AD, 0x00A309B000A3D1AF,
-    0x00A341B200A345B1, 0x00A341B400A371B3, 0x00A319B6028C59B5,
-    0x00A379B8028FDDB7, 0x00A349BA028FC5B9, 0x028309BC00AFF9BB,
-    0x00A80DBE00A35DBD, 0x00AC85B5028FF5BF, 0x00ACB1B7028DB9B4,
-    0x00ACB9B100ACBDB6, 0x028011B3028CEDB0, 0x00AC29BD00A7E9B2,
-    0x00AC25BF00AEA5BC, 0x028035B900A031BE, 0x00ACE9BB028025B8,
-    0x00ACC1A5028025BA, 0x00ACF5A700AC8DA4, 0x00A071A100ACC5A6,
-    0x00AC35A300D8DDA0, 0x028041AD028C8DA2, 0x00A07DAF00ACFDAC,
-    0x0283FDA900ACCDAE, 0x028079AB00ACE5A8, 0x00ACC99500ACBDAA,
-    0x00ACD19700ACC594, 0x00ACB59100ACDD96, 0x00ACA19300ACBD90,
-    0x00A0359D00ACAD92, 0x00A0019F00ACD59C, 0x00ACD19900ACD59E,
-    0x00ACD19B00AE5598, 0x0280C58500D83D9A, 0x00AC19870280D584,
-    0x00AC55810280D586, 0x00AC6183028D6980, 0x00A0D58D00AC6D82,
-    0x00ACC98F00ACF58C, 0x00ACF98900ACF58E, 0x00ACF58B00AE7588,
-    0x00AD118000A0E18A, 0x00AD118200A13181, 0x00AD398400AD2D83,
-    0x00AD298600AD3185, 0x00AD618800AD2587, 0x00AD798A00AD6D89,
-    0x00AFD18C00AD758B, 0x00D9B98E00AD558D, 0x00AD319000AD558F,
-    0x00AD319200A1D191, 0x00D9D99400A1E593, 0x028D899600A17595,
-    0x00ADA99800AD3197, 0x00ADE19A00ADE599, 0x00ADE19C00ADD19B,
-    0x00ADB99E0282F99D, 0x00AD19A002817D9F, 0x00AD29A20281A5A1,
-    0x028D69A400A199A3, 0x00A66DA600AD3DA5, 0x028C1DA8028195A7,
-    0x00AD19AA00AD21A9, 0x0281A1AC00AD15AB, 0x0281A1AE0281B1AD,
-    0x00AF41B000AD6DAF, 0x00A1D5B200ADCDB1, 0x00ADC1B400ADC1B3,
-    0x00ADC1B600A1E1B5, 0x00ADC9B800ADFDB7, 0x00ADD9BA00ADC1B9,
-    0x00ADB1BC00ADD5BB, 0x00ADA9BE00ADBDBD, 0x00ACE18000ADA5BF,
-    0x00DA898200AE6581, 0x00AE618400AE6583, 0x00AE618600A28185,
-    0x00DAA98800A2B587, 0x028EF98A00A20589, 0x00AEF98C00AE418B,
-    0x00AEB18E00AEB58D, 0x00AED19000AE818F, 0x00AE89920281C991,
-    0x00AEC99402824D93, 0x00AEF99602827595, 0x028E999800A24997,
-    0x00A59D9A00AECD99, 0x028FCD9C0282659B, 0x00AEC99E00AEF19D,
-    0x028291A000AEC59F, 0x028291A2028281A1, 0x00140DA400AE5DA3,
-    0x8000014C0015BDA5, 0x000401A880000150, 0x00AD61AA001589A9,
-    0x0006B1AC001599AB, 0x0006B1AE00158DAD, 0x0006D1B002B8A5AF,
-    0x001589B20015F9B1, 0x001599B4001589B3, 0x0006C1B6001599B5,
-    0x0006C1B800AD11B7, 0x02B8F5BA02B8FDB9, 0x02BFBDBC0015BDBB,
-    0x0015A1BE0015A1BD, 0x00A329800015A1BF, 0x0007218200146581,
-    0x00A8398402BAE183, 0x02BF118600A83985, 0x0007118800147987,
-    0x00AFCD8A02BAD589, 0x0014598C02B92D8B, 0x0014558E028FD98D,
-    0x000779900014458F, 0x02BF499202BAB991, 0x00142D9402B94593,
-    0x000719960014D595, 0x00A85D9802BAA997, 0x0007219A0014E199,
-    0x02BF799C0014E59B, 0x02B9799E02BA859D, 0x02BF81A002BA899F,
-    0x000529A202BA79A1, 0x00AC59A40014A1A3, 0x000799A60014A1A5,
-    0x0007A9A80014A5A7, 0x0007B9AA02B9BDA9, 0x0014F1AC001491AB,
-    0x0014F1AE0014F1AD, 0x0007D9B00014F1AF, 0x0007E9B200AC09B1,
-    0x02B9CDB402B9D5B3, 0x02BE95B6001485B5, 0x0014B9B8001489B7,
-    0x00A3C1BA0014B9B9, 0x0007D9BC00148DBB, 0x00A8D1BE02BA19BD,
-    0x80000DEE00A8D1BF, 0x80000DF480000154, 0x80000DFA80000158,
-    0x80000E0C8000015C, 0x80000E1200000013, 0x80000E3F80000ED7,
-    0x80000E5180000160, 0x80000E8980000164, 0x80000E9B80000168,
-    0x80000EA78000016C, 0x0002508080000170, 0x80000EE000004881,
-    0x80000EE680000174, 0x80000EE980000178, 0x80000EEF80000EEC,
-    0x80000EF58000017C, 0x80000F0180000180, 0x80000F1C80000F10,
-    0x80000F2B80000F25, 0x80000F3D80000F7F, 0x8000018480000FA6,
-    0x8000018C80000188, 0x003DB0B880000190, 0x003AB8BA003800B9,
-    0x0000E5B080000194, 0x001355B200135DB1, 0x001371B4000095B3,
-    0x0000B1B60000B5B5, 0x000091B80000B1B7, 0x00134DBA00009DB9,
-    0x000071BC001345BB, 0x001349BE001359BD, 0x00AB89A8001349BF,
-    0x00ABA9AA00AB89A9, 0x00A8E5AC00ABA9AB, 0x0289C5AE0289E1AD,
-    0x00AB81A00289FDAF, 0x00A8DDA200AB81A1, 0x0289EDA40289D9A3,
-    0x00A8C1A60289D5A5, 0x02899DB800A8C1A7, 0x00A8B1BA0289A5B9,
-    0x00ABD9BC00A8B1BB, 0x00ABF9BE00ABD9BD, 0x00A899B000ABF9BF,
-    0x00ABE1B200A899B1, 0x00ABD1B400ABE1B3, 0x00A88DB600ABD1B5,
-    0x02B98188028989B7, 0x0012058A02BC7989, 0x0287F8CC0012198B,
-    0x00A86D8E028738CD, 0x00AE9D800289698F, 0x0001E18200122181,
-    0x02B9B98400122583, 0x02BFB98602BC4585, 0x00A8359802BC4987,
-    0x0289159A02893199, 0x00A8299C02892D9B, 0x00AB519E00A8299D,
-    0x02893D9000AB519F, 0x00A8119202890591, 0x00AB799400A81193,
-    0x00AB599600AB7995, 0x00028D8700AB5997, 0x000281850011C986,
-    0x0011D5830011C584, 0x00029D810011D582, 0x0011FD8F0011A580,
-    0x0011F18D0002B98E, 0x0002A58B0002B58C, 0x0011E9890011E18A,
-    0x000229970002AD88, 0x0011E99500022596, 0x0011CD9300021594,
-    0x00027D910011CD92, 0x0002E59F0011BD90, 0x0011C59D0011D19E,
-    0x0011D59B0002ED9C, 0x0002099900020D9A, 0x001165A700021198,
-    0x00FA3DA5000209A6, 0x000219A3001141A4, 0x00115DA100115DA2,
-    0x80000DE200115DA0, 0x000C25AD80000198, 0x000C31AB000C29AC,
-    0x000FEDA9000CC1AA, 0x8000019C000225A8, 0x800006D180002443,
-    0x8000070F800007CB, 0x8000079D80000797, 0x800007C880000780,
-    0x800007528000074F, 0x800006E18000075D, 0x800026E08000250D,
-    0x0280A8CC80002F2D, 0x02BD20CC02824CCD, 0x028A40CC02BD4CCD,
-    0x028198CC02BEC8CC, 0x028710CC02803CCD, 0x0287A0CC028750CD,
-    0x028430CC02878CCD, 0x028418CC028424CD, 0x02B938CC028484CD,
-    0x02B918CC02B908CD, 0x02BEB0CC02B91CCD, 0x02B950CC02B958CD,
-    0x80002DB202B960CC, 0x02BED0CC003AE0AF, 0x02BDF0CC02BED4CD,
-    0x02BC50CC02BD84CD, 0x80002ECE02BFACCD, 0x80002BFF80003837,
-    0x800033B7800031CB, 0x8000305880002D06, 0x02B884CC80002C9E,
-    0x02B884CC02B8D8CC, 0x02B914CC02B8F8CC, 0x02B920CC02B94CCC,
-    0x0013088602B970CC, 0x0038C884800031A4, 0x80002CF2003ACC85,
-    0x800038938000361F, 0x80002DEA80002EB2, 0x800031088000336B,
-    0x8000391380003134, 0x0034008880003128, 0x002F55B1003A7C89,
-    0x002F55B3028549B0, 0x002F71B5002F55B2, 0x0281F5B700983DB4,
-    0x00930DB9003C35B6, 0x002F4DBB00930DB8, 0x0281C1BD028425BA,
-    0x028239BF003C05BC, 0x002F79A10094D9BE, 0x028251A3003C45A0,
-    0x002F6DA5003C45A2, 0x002F1DA7003C45A4, 0x002F2DA9002F1DA6,
-    0x0097A5AB002F2DA8, 0x028269AD003C75AA, 0x002F29AF003C55AC,
-    0x02817591028261AE, 0x002F0D93003CC590, 0x002F19950093B192,
-    0x02816997003CDD94, 0x0281559902849596, 0x0281559B0282A598,
-    0x002F9D9D0284AD9A, 0x002F8D9F003E159C, 0x002FA5810097759E,
-    0x028139830284CD80, 0x0282D185003CFD82, 0x02B4358700943184,
-    0x002F9989002FB586, 0x0281158B002F8988, 0x0282F98D003CD58A,
-    0x002F798F0284F58C, 0x002E6980002F818E, 0x002E658202B5E981,
-    0x003D2984002E7583, 0x028519860280E985, 0x002E5D8802831587,
-    0x003D698A002EA589, 0x00920D8C0280D98B, 0x003D718E002EB18D,
-    0x02854990002EB58F, 0x028349920280B591, 0x028551940280B993,
-    0x003FF9960280A995, 0x0096A998002E7197, 0x003D699A002E5199,
-    0x003D799C002E559B, 0x003D699E02836D9D, 0x002EC1A0002E419F,
-    0x002EC1A2002EC1A1, 0x003D89A4002EC1A3, 0x003DB9A6009659A5,
-    0x0283BDA8028385A7, 0x0284E5AA002EF5A9, 0x002EE9AC002EF9AB,
-    0x009991AE002EE9AD, 0x003DE9B0002EDDAF, 0x0092E1B2028029B1,
-    0x0285D9B40092E1B3, 0x003DE9B6002EB1B5, 0x009505B802802DB7,
-    0x002E81BA0283E5B9, 0x002E9DBC02B501BB, 0x003DC1BE002E8DBD,
-    0x02860180028001BF, 0x002D758202800D81, 0x003E5184002D8D83,
-    0x009125860283E185, 0x003E6988002D9987, 0x0286218A002DAD89,
-    0x0280318C0283DD8B, 0x0286398E0283C18D, 0x003CE1900283C18F,
-    0x00958192002D6991, 0x003E5194002D7993, 0x003E5196002D6D95,
-    0x003E719802804597, 0x002D299A002D5999, 0x002D399C002D299B,
-    0x003E619E002D399D, 0x003EA1A00095B19F, 0x028095A202809DA1,
-    0x0287DDA4002DDDA3, 0x002DC1A6002DC1A5, 0x009A89A8002DC1A7,
-    0x003E81AA002DC5A9, 0x009199AC028341AB, 0x0286B1AE009199AD,
-    0x003EF1B0002DD9AF, 0x00962DB2028335B1, 0x002DB9B40280CDB3,
-    0x002DB5B602B639B5, 0x003ED9B8002DA5B7, 0x0286E9BA028319B9,
-    0x002D8DBC0280E5BB, 0x003EB9BE002D75BD, 0x00270D80028309BF,
-    0x0020A18200209D81, 0x800031000020B983, 0x0020ED8600275985,
-    0x002739880020B187, 0x0027198A00277D89, 0x0027098C0027198B,
-    0x0020C18E0027098D, 0x0027699000271D8F, 0x0027659200275591,
-    0x0020AD940020F193, 0x0020FD960020F595, 0x0020D9980020ED97,
-    0x0027759A00277199, 0x0031E19C0027499B, 0x0031E59E00D7819D,
-    0x800035770031DD9F, 0x002021A200201DA1, 0x0027D9A4800035B3,
-    0x8000334B80002DD6, 0x8000349F002031A7, 0x002799AA0027FDA9,
-    0x002789AC002799AB, 0x002041AE002789AD, 0x0027E9B000279DAF,
-    0x0027E5B20027D5B1, 0x00202DB480002D0A, 0x00207DB6002075B5,
-    0x80002E3600206DB7, 0x8000394B0027F1B9, 0x800024870027C9BB,
-    0x800024558000244B, 0x00200D8080002469, 0x0027A18200279D81,
-    0x002059840027B983, 0x0027ED8600205985, 0x002039880027B187,
-    0x8000249100207D89, 0x0020098C0020198B, 0x0027C18E0020098D,
-    0x0020699000201D8F, 0x0020659200205591, 0x0027AD940027F193,
-    0x0027FD960027F595, 0x0027D9980027ED97, 0x0020759A00207199,
-    0x800024730020499B, 0x8000245F8000249B, 0x8000305C8000247D,
-    0x002721A200271DA1, 0x80002DC6002739A3, 0x00276DA60020D9A5,
-    0x0020B9A8002731A7, 0x80002BD30020FDA9, 0x002089AC002099AB,
-    0x002741AE002089AD, 0x0020E9B000209DAF, 0x0020E5B20020D5B1,
-    0x00272DB4002771B3, 0x00277DB6002775B5, 0x002759B800276DB7,
-    0x0020F5BA0020F1B9, 0x80002BCB0020C9BB, 0x8000EAF980002BDB,
-    0x0037B58080002C26, 0x0037F9820037ED81, 0x00311D8400361D83,
-    0x0031898600310D85, 0x0031E9880031AD87, 0x0031D58A0031F189,
-    0x80002C4280002C3A, 0x80002C4E80002C4A, 0x0030559080002C52,
-    0x0030699200305991, 0x0030159400306993, 0x003FC596003FCD95,
-    0x00396D9800395997, 0x0039459A00397199, 0x00394D9C0039599B,
-    0x0039359E0039319D, 0x0039E5A00039219F, 0x0039F5A20039F1A1,
-    0x003935A400392DA3, 0x003935A600393DA5, 0x003B05A8003921A7,
-    0x003B6DAA003B65A9, 0x0091D1AC009DB1AB, 0x003B75AE003B7DAD,
-    0x009341B080002A8B, 0x009DD5B20091CDB1, 0x0091C1B40091C1B3,
-    0x0091C1B6009DE1B5, 0x0091C9B80091FDB7, 0x0091D9BA0091C1B9,
-    0x0091B1BC0091D5BB, 0x0091A9BE0091BDBD, 0x00223DAD0091A5BF,
-    0x002229AF00223DAC, 0x002249A90025F5AE, 0x8000EB0D00220DA8,
-    0x80002C7200222DAA, 0x002585A700226DA4, 0x0025A9A180002C76,
-    0x8000EAFE80002C92, 0x80002A8F002595A2, 0x80002C7E0034D5BC,
-    0x002245B90034D1BE, 0x00227DBB8000EB08, 0x0025C1B5002241BA,
-    0x0025D9B7002599B4, 0x002261B10025C9B6, 0x80002CAE00225DB0,
-    0x0022BD8D002251B2, 0x0022A98F80002CB6, 0x0022C9890025758E,
-    0x0022AD8B80002C0F, 0x8000ECE380002CBE, 0x0025058780002CE6,
-    0x80002CFE80002A97, 0x80002D0E80002D02, 0x00D2359D00251582,
-    0x0034699F80002A9B, 0x0022C59980002D46, 0x0022FD9B80002D4A,
-    0x80002D7E80002D72, 0x0025599700251994, 0x0022E19180002D8E,
-    0x80002D9A80002D96, 0x0091E1800022D192, 0x00E7898200936581,
-    0x0093618400936583, 0x00936186009F8185, 0x00E7A988009FB587,
-    0x0039F98A009F0589, 0x0039E58C00F9F18B, 0x0039E58E0039ED8D,
-    0x02B3B5B10039DD8F, 0x02B3A9B30092B1B0, 0x0092ADB502B391B2,
-    0x0091D5B70092ADB4, 0x0091C5B90091D5B6, 0x8000EB1C0091C5B8,
-    0x80002DAE8000339B, 0x8000EB2180002DB6, 0x0004718080002DC2,
-    0x0004718200047181, 0x0003B98400047183, 0x0003ED860003E585,
-    0x0003DD880003E587, 0x00394DAA80002DCA, 0x00395DAC003945AB,
-    0x80002DFA80002DD2, 0x0003419080002DE6, 0x0012C8A800035591,
-    0x80002DF680002DF2, 0x80002E0680002DFE, 0x000800A280002E0A,
-    0x000DE8A0000B8CA3, 0x001000A6000FECA1, 0x001400A40011F4A7,
-    0x000CA9800017C8A5, 0x000CFD82000CF581, 0x000CED84000CF583,
-    0x000D358600C5D185, 0x000D1588000D2D87, 0x00C5E18A000D1D89,
-    0x000D8D8C000D118B, 0x00C5F58E000E2D8D, 0x000E6D90000E1D8F,
-    0x000E7192000E7191, 0x000E1994000E0D93, 0x000E0D960037C995,
-    0x000EE598000E2597, 0x000EF59A000EF199, 0x000ED19C00C7159B,
-    0x000EDD9E00E8819D, 0x000E31A0000ED19F, 0x000B45A2000E31A1,
-    0x000E55A4000E35A3, 0x00C549A600C719A5, 0x00C731A800C549A7,
-    0x000E7DAA000E69A9, 0x000E6DAC00CE55AB, 0x000E59AE00C56DAD,
-    0x000E35B0000E51AF, 0x000E31B2000E3DB1, 0x000FB9B4000E35B3,
-    0x000FA9B6000FB9B5, 0x000F99B8000FA9B7, 0x000F4DBA000F99B9,
-    0x000F49BC000F59BB, 0x000F35BE000F49BD, 0x80002E12000F2DBF,
-    0x80002E1E80002E1A, 0x000AF1BA80002E2E, 0x000AB9B8000AFDBB,
-    0x00FE25B6000AD1B9, 0x000ABDB400EAE5B7, 0x000AB1B2000AB5B5,
-    0x00E121B0000AB9B3, 0x00FE0DAE000A2DB1, 0x00FE01AC0036B1AF,
-    0x00FA4DAA00FA65AD, 0x0036B5A80036A5AB, 0x003685A600E9E9A9,
-    0x0036B5A400368DA7, 0x00EC49A200368DA5, 0x0036A9A00036ADA3,
-    0x00361D9E0036A9A1, 0x00361D9C0036419F, 0x0036019A0036159D,
-    0x0036119800C2C59B, 0x0036259600361199, 0x00E9659400FAC197,
-    0x0036D19200362D95, 0x00EAC19000E73993, 0x80002E3200374191,
-    0x80002E6280002E3E, 0x80002E4A80002E66, 0x80002EAE80002E52,
-    0x80002E7A80002E76, 0x80002E7280002E7E, 0x00F5E58280002E86,
-    0x00370D8080002E82, 0x000DD58000370581, 0x000DD582000DDD81,
-    0x0003158400030D83, 0x0003118600031D85, 0x00032D8800C58587,
-    0x0003ED8A0003E589, 0x00C7D18C0003E58B, 0x0003E98E0003F98D,
-    0x00C701900003E98F, 0x0003959200039D91, 0x0003B5940003AD93,
-    0x0003B5960003BD95, 0x000161980003A997, 0x0001619A00016199,
-    0x0001619C0001619B, 0x00016D9E0035F59D, 0x00019DA00001659F,
-    0x0001ADA20001A5A1, 0x0001BDA40001A5A3, 0x0001ADA60001A5A5,
-    0x00019DA80001A5A7, 0x0001EDAA0001E5A9, 0x0001F9AC0001E9AB,
-    0x0001E9AE0001F9AD, 0x000195B000CC55AF, 0x000195B200019DB1,
-    0x0001B5B40001ADB3, 0x0001B1B600E9D9B5, 0x00E9F5B80001B1B7,
-    0x00019DBA000195B9, 0x00018DBC000195BB, 0x00017DBE000175BD,
-    0x80002E9A000175BF, 0x80002EB68000EB26, 0x80002EC280002EBA,
-    0x80002EF980002ED6, 0x8000EB308000EB2B, 0x80002F0980002F05,
-    0x80002F0180002F0D, 0x80002A9F80002F15, 0x80002F1D80002AA3,
-    0x8000EB3580002F25, 0x80002F4580002F41, 0x8000EB3A80002F51,
-    0x80002F5D80002F59, 0x80002AA78000303C, 0x80002F8D80002F71,
-    0x8000EB4480002F89, 0x8000EB4980002F95, 0x80002F9D80002FA1,
-    0x80002FB580002FA5, 0x80002AAB80002FB9, 0x80002FD180002FC9,
-    0x80002FD980002FD5, 0x8000EB4E80002AAF, 0x80002FFC80002AB3,
-    0x8000300480003000, 0x8000EB588000ED15, 0x80003034800036DF,
-    0x8000EB7B80002AB7, 0x800030448000EC4D, 0x80002ABB8000304C,
-    0x8000307080003060, 0x000000FF80003074, 0x8000308C8000307C,
-    0x80002ABF80002AC3, 0x800030948000EB5D, 0x800030AC800030A8,
-    0x800030D0800030BC, 0x800030CC800030C8, 0x800030E8800030D4,
-    0x80003104800030EC, 0x8000311C8000310C, 0x8000EB6C80003130,
-    0x800031208000312C, 0x8000314080003138, 0x8000EB7180003154,
-    0x0000005A8000315C, 0x8000314880003150, 0x8000316480002AC7,
-    0x800031748000316C, 0x80002ACB80003168, 0x8000EB768000318C,
-    0x800031FA800031B8, 0x80002AD7800031D6, 0x80002ACF80002AD3,
-    0x80002CA680002CA2, 0x800031E6800031FE, 0x80002B3B800036B3,
-    0x8000322680003212, 0x8000EB8F8000321E, 0x8000323A80002ADB,
-    0x800032578000322E, 0x800032638000EB99, 0x8000326B80003242,
-    0x800032738000326F, 0x000149B880003277, 0x000159BA000149B9,
-    0x000149BC000159BB, 0x000179BE000149BD, 0x00C319B0000179BF,
-    0x000119B200C815B1, 0x000109B4000119B3, 0x000179B6000109B5,
-    0x00C289A8000179B7, 0x000155AA00015DA9, 0x000171AC00C0B1AB,
-    0x000171AE000171AD, 0x000161A0000171AF, 0x000161A2000161A1,
-    0x00017DA400C099A3, 0x00016DA6000165A5, 0x00065D98000165A7,
-    0x00066D9A00066599, 0x00067D9C0006659B, 0x00066D9E0006659D,
-    0x00EEED900006619F, 0x0006619200066D91, 0x0006619400066193,
-    0x00066D9600066595, 0x00066D8800066597, 0x00067D8A00067589,
-    0x00066D8C0006758B, 0x00C8458E0006158D, 0x00066D800006198F,
-    0x00067D8200067581, 0x00066D8400067583, 0x00065D8600065585,
-    0x0002C59A00065587, 0x00CBA9980002CD9B, 0x0002C59E00C1E599,
-    0x0002D59C0002CD9F, 0x00028D920002CD9D, 0x00028D9000029593,
-    0x00028D9600028591, 0x00028D940002F597, 0x0002D18A00028595,
-    0x00C075880002D18B, 0x0002F18E0002C189, 0x0002C18C00C0198F,
-    0x000215820002C18D, 0x0002058000021D83, 0x0002F58600021D81,
-    0x000205840002FD87, 0x0001B9BA0002FD85, 0x0001B9B80001B9BB,
-    0x0001D9BE0001B9B9, 0x0001D9BC0001D9BF, 0x0001BDB20001D9BD,
-    0x0001BDB0000185B3, 0x000199B60001B5B1, 0x00019DB4000199B7,
-    0x00023DAA00C791B5, 0x00023DA80001C5AB, 0x0001DDAE000235A9,
-    0x0001DDAC0001C5AF, 0x00C0B1A20001D5AD, 0x000235A0000201A3,
-    0x000201A600020DA1, 0x000211A400DC01A7, 0x80003283000211A5,
-    0x8000328B8000EB9E, 0x8000329780002AE3, 0x8000329B8000EBA3,
-    0x800032A780002AE7, 0x800032C3800032B3, 0x8000EB3F8000EBA8,
-    0x800032E38000EBAD, 0x800032EF8000EBB2, 0x800032E7800032F7,
-    0x8000331380003303, 0x800033278000331B, 0x8000EBB78000332B,
-    0x800033438000330F, 0x80002AEB80003347, 0x8000EBBC80003357,
-    0x800033778000333B, 0x8000EBC68000EBC1, 0x8000338F8000337B,
-    0x80002AEF80003387, 0x800033A380003393, 0x800033AB8000339F,
-    0x800033BB8000EB03, 0x800033C78000EBCB, 0x800033DB8000EBD5,
-    0x8000EBDA800033FB, 0x800034138000340F, 0x8000EBE48000EBDF,
-    0x000000B88000342F, 0x80002AF38000343F, 0x80002AF780003443,
-    0x8000345F8000345B, 0x800034778000346B, 0x8000348F80002AFB,
-    0x800034AB8000EBE9, 0x8000EB538000EBEE, 0x8000EBF3800034C7,
-    0x8000EBFD8000EBF8, 0x80002B0380002AFF, 0x8000EC070000004D,
-    0x8000EC0C8000EC02, 0x800035078000EC11, 0x800035178000350B,
-    0x80002B1380002B0F, 0x8000352B8000EC1B, 0x8000EC2080002B17,
-    0x8000EC2A8000EC25, 0x80002B1B8000359B, 0x800035AB800035A7,
-    0x8000EC348000EC2F, 0x80002B1F800035C3, 0x800035D78000EC39,
-    0x80002B23800035DB, 0x800035F78000EC3E, 0x8000360780002B27,
-    0x8000360F800035FF, 0x800036278000EC48, 0x8000363380002B2B,
-    0x8000364B8000363B, 0x8000EC5280002B2F, 0x80002B338000EC57,
-    0x800036638000EC5C, 0x8000366F8000EC61, 0x8000EC668000367B,
-    0x00012DBE8000EC6B, 0x00012DBC000135BF, 0x00010DBA000125BD,
-    0x00D7EDB8000135BB, 0x000131B6000105B9, 0x000121B4000131B7,
-    0x000131B2000121B5, 0x00C339B0000131B3, 0x001EB5AE001EC1B1,
-    0x001EB9AC001EBDAF, 0x001EB9AA001EBDAD, 0x001E85A8001EB9AB,
-    0x001EB5A6001EB9A9, 0x001EA5A4001EBDA7, 0x001EB5A2001EBDA5,
-    0x001EB9A0001EBDA3, 0x001E399E00C1A5A1, 0x001E399C001E399F,
-    0x001E399A001E399D, 0x001E3998001E399B, 0x001E3996001E3999,
-    0x00DE5194001E3997, 0x001E3D92001E3995, 0x001E3D90001E2593,
-    0x001EBD8E001E3591, 0x001EBD8C001E458F, 0x001EA18A001EB58D,
-    0x001EB18800DE218B, 0x001E8186001EB189, 0x001EB184001E8187,
-    0x00DA8982001EB185, 0x001EA980001EA183, 0x00025D8D00DF8181,
-    0x00D7E18F00025D8C, 0x00C0ED8900026D8E, 0x00024D8B00C00988,
-    0x0002698500024D8A, 0x0002658700026184, 0x0002098100026186,
-    0x0002798300C02D80, 0x0002C99D00020182, 0x0002C59F0002C19C,
-    0x00DD49990002C59E, 0x0002D99B00022D98, 0x000211950002219A,
-    0x00021D9700020994, 0x0002159100DFFD96, 0x00DD759300C05190,
-    0x000245AD00021592, 0x000245AF000245AC, 0x000225A9000245AE,
-    0x000225AB000225A8, 0x000205A5000225AA, 0x000205A7000205A4,
-    0x000225A1000205A6, 0x000225A3000225A0, 0x0001C1BD000225A2,
-    0x0001C1BF000239BC, 0x000225B90001C9BE, 0x000221BB000225B8,
-    0x000205B500DDEDBA, 0x000205B7000205B4, 0x000225B1000205B6,
-    0x000225B3000225B0, 0x80003697000225B2, 0x8000369F8000EC70,
-    0x80002B378000EB80, 0x800036BB800036B7, 0x800036BF80002B3F,
-    0x8000EC7580002F19, 0x8000EB858000EC7A, 0x800036E38000EB8A,
-    0x80002B43800038FB, 0x8000370B8000370F, 0x80002D1E80003713,
-    0x8000371B80003717, 0x800037278000371F, 0x8000372B8000EC7F,
-    0x800037438000373F, 0x800037478000372F, 0x8000373B8000375B,
-    0x800037538000374F, 0x8000EC8480003757, 0x8000EC898000EC8E,
-    0x8000377780002B47, 0x800037838000377B, 0x800037878000ECA2,
-    0x80002B4B8000EC93, 0x8000EC9880002B4F, 0x80002B538000EC9D,
-    0x800037AF800037A7, 0x800037BF800037B3, 0x800037C7800037BB,
-    0x800037CB800037C3, 0x800037D7800037CF, 0x000000BE80002B57,
-    0x0007FD80800037DF, 0x0002018200020181, 0x00F1FD8400020183,
-    0x00020D8600020585, 0x00023D8800020587, 0x00020D8A00020589,
-    0x00021D8C0002058B, 0x00020D8E0002058D, 0x00C615900002058F,
-    0x00FC419200C61D91, 0x0002119400027193, 0x0002119600021195,
-    0x0002319800021197, 0x0002319A00023199, 0x0002119C0002319B,
-    0x0002119E0002119D, 0x0002EDA000C6219F, 0x0002FDA20002F5A1,
-    0x0002EDA40002F5A3, 0x00021DA6000215A5, 0x00022DA8000215A7,
-    0x00023DAA000235A9, 0x00022DAC000235AB, 0x00021DAE000215AD,
-    0x00026DB0000215AF, 0x00027DB2000275B1, 0x00026DB4000275B3,
-    0x00021DB6000215B5, 0x00022DB8000215B7, 0x00023DBA000235B9,
-    0x00022DBC000235BB, 0x00C681BE000215BD, 0x0006E9BC00C685BF,
-    0x0006F9BE0006E9BD, 0x0006CDB80006F9BF, 0x0006FDBA0006F5B9,
-    0x00D0F1B400C75DBB, 0x0006FDB60006F5B5, 0x0006F1B00006F5B7,
-    0x0006F1B20006F1B1, 0x0006F1AC0006F1B3, 0x0006F1AE0006F1AD,
-    0x0006F1A80006F1AF, 0x0006F1AA0006F1A9, 0x0006F1A40006F1AB,
-    0x0006F1A60006F1A5, 0x0006F1A00006F1A7, 0x0006F1A20006F1A1,
-    0x0006F59C0006F1A3, 0x0006F59E0006FD9D, 0x00DCA99800DAB19F,
-    0x0006F59A0006FD99, 0x0006F9940006ED9B, 0x0006C9960006F995,
-    0x0006F9900006C997, 0x0006E9920006F991, 0x0006F98C0006E993,
-    0x0006898E0006F98D, 0x0006FD880006898F, 0x0006ED8A0006E589,
-    0x0006FD84001E118B, 0x0006CD860006C585, 0x0006FD800006C587,
-    0x0006ED820006E581, 0x000000800006E583, 0x80002B5B800037E3,
-    0x8000ECA7800037F3, 0x8000380780003803, 0x8000381380002B5F,
-    0x8000ECAC80002A93, 0x80002B638000ECB1, 0x8000383F80002B67,
-    0x800038978000ECB6, 0x800038B38000389B, 0x8000ECC0800038BF,
-    0x800038CB8000EB12, 0x800038D3800038C7, 0x800038DF8000EB17,
-    0x8000ECCA8000ECC5, 0x800039378000392B, 0x8000ECCF8000393F,
-    0x8000396F80003943, 0x8000397F8000397B, 0x8000398B80003977,
-    0x8000ECD48000398F, 0x80002B6B8000399B, 0x8000ECD9800039A3,
-    0x800039DB80002B6F, 0x800039F380002FA9, 0x8000ECE88000ECDE,
-    0x80002B7780002B73, 0x8000ECED80003A1B, 0x8000ECF280002B7B,
-    0x8000ECF780003A3B, 0x80002B7F80003A53, 0x80003A6F80003A63,
-    0x0001E58080003A77, 0x0001E5820001ED81, 0x0001E5840001FD83,
-    0x0001E5860001ED85, 0x00005D8800006587, 0x00C5558A0000A589,
-    0x0000B98C0000A98B, 0x0000A98E0000B98D, 0x0000D9900000A98F,
-    0x0000E9920000D991, 0x0000F9940000E993, 0x0000F59600FF9D95,
-    0x0000D19800FF8597, 0x0000D19A0000D199, 0x00FF9D9C0000D19B,
-    0x80003A7F0000B59D, 0x8000ECFC80002B83, 0x80003ABB80003AAF,
-    0x80002B8B80002B87, 0x8000ED0180003ABF, 0x8000ED0680002B8F,
-    0x8000ED108000ED0B, 0x80003AFF80002B93, 0x80003B0B80003B03,
-    0x8000ED1A80003B13, 0x000000F2000000F1, 0x000000F4000000F3,
-    0x000000F6000000F5, 0x000000F8000000F7, 0x000000FA000000F9,
-    0x000000FC000000FB, 0x000000FE000000FD, 0x20002000000000FF,
-    0x200083CC200081CC, 0x200085CC200084CC, 0x200087CC200086CC,
-    0x80CC88CC200088CC, 0x200081CC88CC2000, 0x8ACC200082CD88CC,
-    0x93CC20008BCC2000, 0x200080CC93CC2000, 0x93CC200081CC93CC,
-    0x200094CC200082CD, 0x94CC200080CC94CC, 0x82CD94CC200081CC,
-    0xA8CC2000A7CC2000, 0x82CD2000B3CC2000, 0x8BD9200085CD2000,
-    0x8CD920008CD92000, 0x20008DD9200091D9, 0x8ED9200091D98DD9,
-    0x200091D98ED92000, 0x91D98FD920008FD9, 0x90D9200090D92000,
-    0x200091D9200091D9, 0x92D92000B0D991D9, 0xE320009982E32000,
-    0x0021210021009A82, 0x2400230022003F21, 0x2800270026002500,
-    0x3031280029312800, 0x2800293131280029, 0x2933312800293231,
-    0x3128002934312800, 0x0029363128002935, 0x3831280029373128,
-    0x2800293931280029, 0x0029303228002932, 0x0029342800293328,
-    0x0029362800293528, 0x0029382800293728, 0x0029412800293928,
-    0x0029432800294228, 0x0029452800294428, 0x0029472800294628,
-    0x0029492800294828, 0x00294B2800294A28, 0x00294D2800294C28,
-    0x00294F2800294E28, 0x0029512800295028, 0x0029532800295228,
-    0x0029552800295428, 0x0029572800295628, 0x0029592800295828,
-    0x0029612800295A28, 0x0029632800296228, 0x0029652800296428,
-    0x0029672800296628, 0x0029692800296828, 0x00296B2800296A28,
-    0x00296D2800296C28, 0x00296F2800296E28, 0x0029712800297028,
-    0x0029732800297228, 0x0029752800297428, 0x0029772800297628,
-    0x0029792800297828, 0x8084E12800297A28, 0x00298284E1280029,
-    0xE12800298384E128, 0x8684E12800298584, 0x00298784E1280029,
-    0xE12800298984E128, 0x8C84E12800298B84, 0x00298E84E1280029,
-    0xE12800298F84E128, 0x9184E12800299084, 0x00299284E1280029,
-    0xE428002980B8E428, 0x89B8E428002983B8, 0x00299DB9E4280029,
-    0xE42800298CBAE428, 0xA3BBE428002994BA, 0x002981BCE4280029,
-    0xE528002991BCE428, 0xAD85E5280029AB85, 0x0029B48AE5280029,
-    0xE5280029818DE528, 0x8D90E5280029948D, 0x0029BC91E5280029,
-    0xE52800299B9BE528, 0xA6ADE52800299F9C, 0x0029A597E6280029,
-    0xE6280029889CE628, 0xA89CE6280029899C, 0x0029AAA0E6280029,
-    0xE7280029B4B0E628, 0xB989E7280029AB81, 0x0029A39BE7280029,
-    0xE7280029BEA4E728, 0xADA5E72800299DA5, 0x0029AA87E8280029,
-    0xE8280029B387E828, 0x87B3E8280029A1B2, 0x00299187E9280029,
-    0xEB28002980B0EA28, 0xA48BEB2800299882, 0x0029BC9DEB280029,
-    0xEB28002988A7EB28, 0xAC82EC28002994B0, 0x00298495EC280029,
-    0x2984A0ECA498EC28, 0x849BEDA498EC2800, 0x0029909EEC280029,
-    0xEC280029BCA3EC28, 0xB4B9EC280029A8B0, 0x00298083ED280029,
-    0xED2800298C8CED28, 0x002A002900299895, 0x002E002D002C002B,
-    0x2F002E2E2E002E2E, 0x2E30002C30003000, 0x3000338481E23000,
-    0x2C31003100B982E7, 0x31003031002E3100, 0xA597E63031002E30,
-    0x3100889CE6303100, 0x00313100B982E730, 0x97E63131002E3131,
-    0x00889CE6313100A5, 0x323100B982E73131, 0xE63231002E323100,
-    0x889CE6323100A597, 0x3100B982E7323100, 0x3331002E33310033,
-    0x82E7333100A597E6, 0x2E343100343100B9, 0x3100A597E6343100,
-    0x00353100B982E734, 0x97E63531002E3531, 0x00B982E7353100A5,
-    0x31002E3631003631, 0xE7363100A597E636, 0x373100373100B982,
-    0x00A597E63731002E, 0x383100B982E73731, 0xE63831002E383100,
-    0xB982E7383100A597, 0x002E393100393100, 0x393100A597E63931,
-    0x8481E23100B982E7, 0x0030318481E23100, 0xE23100328481E231,
-    0x8481E23100338481, 0x00358481E2310034, 0xE23100368481E231,
-    0x8481E23100378481, 0x00398481E2310038, 0x9CE63100A597E631,
-    0x3200B982E7310088, 0x32002E32002C3200, 0x3032002E30320030,
-    0x82E7303200A597E6, 0xE6313200313200B9, 0xB982E7313200A597,
-    0x97E6323200323200, 0x00B982E7323200A5, 0xA597E63332003332,
-    0x3200B982E7333200, 0x00A597E634320034, 0x353200B982E73432,
-    0x3200A597E6353200, 0x00A597E636320036, 0xA597E63732003732,
-    0x97E6383200383200, 0xE6393200393200A5, 0x338481E23200A597,
-    0x3200358481E23200, 0x889CE63200A597E6, 0x003300B982E73200,
-    0x3033002E33002C33, 0x3300A597E6303300, 0x00A597E631330031,
-    0x3433003333003233, 0x3300363300353300, 0x0039330038330037,
-    0xE23300348481E233, 0x8481E23300358481, 0x3300A597E6330038,
-    0xB982E73300889CE6, 0x2E34002C34003400, 0x3400313400303400,
-    0x0034340033340032, 0x3734003634003534, 0x3400393400383400,
-    0x97E63400358481E2, 0x3400889CE63400A5, 0x2C35003500B982E7,
-    0x35003035002E3500, 0x81E23500368481E2, 0x00A597E635003884,
-    0x82E73500889CE635, 0x36002C36003600B9, 0x3600A597E636002E,
-    0xB982E73600889CE6, 0x2E37002C37003700, 0x3700388481E23700,
-    0x889CE63700A597E6, 0x003800B982E73700, 0xE638002E38002C38,
-    0x00889CE63800A597, 0x39003900B982E738, 0x97E639002E39002C,
-    0x3900889CE63900A5, 0x3A3A003A00B982E7, 0x003D003C003B003D,
-    0x3E003D3D3D003D3D, 0x3F3F00213F003F00, 0x0055410041004000,
-    0x0042006D9588E241, 0x0044430043007142, 0x9588E243002E6F43,
-    0x004A44004400676B, 0xC544007A44005A44, 0x004500BEC54400BD,
-    0x0047005841460046, 0x47007A4847004247, 0x0048007947006150,
-    0x6748005648005048, 0x49490049007A4800, 0x004A490049494900,
-    0x5849005649005549, 0x00424B004B004A00, 0x004C004D4B004B4B,
-    0x4C0044544C004A4C, 0x004D00B7C24C006A, 0x444D00434D00424D,
-    0x61504D007A484D00, 0x4D00564D00524D00, 0x004E00A9CE4D0057,
-    0x6F4E006A4E004A4E, 0x0048500050004F00, 0x00565050004D5050,
-    0x5000455450005250, 0x7352005200510061, 0x4D53004453005300,
-    0x5400765300535300, 0x7A4854004C455400, 0x00560055004D5400,
-    0x5600494956004956, 0x9588E25600494949, 0x570043570057006D,
-    0x580058006257005A, 0x0059004949580049, 0x005D005C005B005A,
-    0x00610060005F005E, 0x632F61002E6D2E61, 0xBECA6100732F6100,
-    0x6300726162006200, 0x752F63006F2F6300, 0x006363006C616300,
-    0x6D63006D63006463, 0x006400336D630032, 0x6C64006164004264,
-    0x00326D64006D6400, 0x64007A6400336D64, 0x005665006500BEC5,
-    0x6666006600677265, 0x6C66660069666600, 0x66006C6600696600,
-    0x006C61670067006D, 0x6168006150680068, 0x6969006969006900,
-    0x006E69006A690069, 0x006A007869007669, 0x7A486B00416B006B,
-    0x00566B0061506B00, 0x006C61636B00576B, 0x6D6B006C6B00676B,
-    0x336D6B00326D6B00, 0x00A9CE6B00746B00, 0x006D6C006A6C006C,
-    0x6C00676F6C006E6C, 0x006D00B7C26C0078, 0x416D00336D00326D,
-    0x6D00576D00566D00, 0x6C696D00676D0062, 0x6D006D6D006C6D00,
-    0x6D00336D6D00326D, 0xE26D00736D006C6F, 0x9588E26D00739588,
-    0x00416E006E003273, 0x576E00566E00466E, 0x6E006D6E006A6E00,
-    0x7000566F006F0073, 0x4170002E6D2E7000, 0x7000567000467000,
-    0x0073700063700057, 0x0064617200720071, 0x00739588E2646172,
-    0x32739588E2646172, 0x7473007273007300, 0x7600760075007400,
-    0x6976006969760069, 0x7800780077006969, 0x0079006969780069,
-    0x007D007C007B007A, 0xA5C200A3C200A2C2, 0xC200ACC200A6C200,
-    0xC20046B0C20043B0, 0x0081C30080C300B7, 0x84C30083C30082C3,
-    0xC30086C30085C300, 0x0089C30088C30087, 0x8CC3008BC3008AC3,
-    0xC3008EC3008DC300, 0x0092C30091C3008F, 0x95C30094C30093C3,
-    0xC30099C30096C300, 0x009CC3009BC3009A, 0xA1C300A0C3009DC3,
-    0xC300A3C300A2C300, 0x00A6C300A5C300A4, 0xA9C300A8C300A7C3,
-    0xC300ABC300AAC300, 0x00AEC300ADC300AC, 0xB1C300B0C300AFC3,
-    0xC300B3C300B2C300, 0x00B6C300B5C300B4, 0xBAC300B9C300B8C3,
-    0xC300BCC300BBC300, 0x0080C400BFC300BD, 0x83C40082C40081C4,
-    0xC40085C40084C400, 0x0088C40087C40086, 0x8BC4008AC40089C4,
-    0xC4008DC4008CC400, 0x0092C4008FC4008E, 0x95C40094C40093C4,
-    0xC40097C40096C400, 0x009AC40099C40098, 0x9DC4009CC4009BC4,
-    0xC4009FC4009EC400, 0x00A2C400A1C400A0, 0xA5C400A4C400A3C4,
-    0xC400A7C400A6C400, 0x00AAC400A9C400A8, 0xADC400ACC400ABC4,
-    0xC400AFC400AEC400, 0x00B4C400B1C400B0, 0xB7C400B6C400B5C4,
-    0xC400BAC400B9C400, 0x00BDC400BCC400BB, 0x84C50083C500BEC4,
-    0xC50086C50085C500, 0x008BC50088C50087, 0x8EC5008DC5008CC5,
-    0xC50090C5008FC500, 0x0094C50093C50091, 0x97C50096C50095C5,
-    0xC50099C50098C500, 0x009CC5009BC5009A, 0x9FC5009EC5009DC5,
-    0xC500A1C500A0C500, 0x00A4C500A3C500A2, 0xA9C500A8C500A5C5,
-    0xC500ABC500AAC500, 0x00AEC500ADC500AC, 0xB1C500B0C500AFC5,
-    0xC500B3C500B2C500, 0x00B6C500B5C500B4, 0xB9C500B8C500B7C5,
-    0xC500BBC500BAC500, 0x00BEC500BDC500BC, 0xA0C60090C6008EC6,
-    0xC600ABC600A1C600, 0x0080C700B0C600AF, 0x8DC70082C70081C7,
-    0xC7008FC7008EC700, 0x0092C70091C70090, 0x95C70094C70093C7,
-    0xC70097C70096C700, 0x009AC70099C70098, 0x9EC7009CC7009BC7,
-    0xC700A0C7009FC700, 0x00A3C700A2C700A1, 0xA8C700A7C700A6C7,
-    0xC700AAC700A9C700, 0x00ADC700ACC700AB, 0xB0C700AFC700AEC7,
-    0xC700B5C700B4C700, 0x00BAC700B9C700B8, 0xBDC700BCC700BBC7,
-    0xC800BFC700BEC700, 0x0082C80081C80080, 0x85C80084C80083C8,
-    0xC80087C80086C800, 0x008AC80089C80088, 0x8DC8008CC8008BC8,
-    0xC8008FC8008EC800, 0x0092C80091C80090, 0x95C80094C80093C8,
-    0xC80097C80096C800, 0x009AC80099C80098, 0x9FC8009EC8009BC8,
-    0xC800A6C800A2C800, 0x00A9C800A8C800A7, 0xACC800ABC800AAC8,
-    0xC800AEC800ADC800, 0x00B1C800B0C800AF, 0xB7C800B3C800B2C8,
-    0xC90091C90090C900, 0x0094C90093C90092, 0x97C90096C90095C9,
-    0xC90099C90098C900, 0x009EC9009CC9009B, 0xA1C900A0C9009FC9,
-    0xC900A3C900A2C900, 0x00A6C900A5C900A4, 0xA9C900A8C900A7C9,
-    0xC900ABC900AAC900, 0x00AEC900ADC900AC, 0xB1C900B0C900AFC9,
-    0xC900B3C900B2C900, 0x00B6C900B5C900B4, 0xB9C900B8C900B7C9,
-    0xC900BBC900BAC900, 0x0080CA00BEC900BD, 0x83CA0082CA0081CA,
-    0xCA0088CA0084CA00, 0x008BCA008ACA0089, 0x8ECA008DCA008CCA,
-    0xCA0090CA008FCA00, 0x0095CA0092CA0091, 0x9BCA0099CA0098CA,
-    0xCA009DCA009CCA00, 0x00A2CA00A1CA009F, 0xA5CA00A4CA00A3CA,
-    0xCA00A7CA00A6CA00, 0x00AACA00A9CA00A8, 0xBCCA00B9CA00ABCA,
-    0x0091CB0090CB006E, 0x88CC0081CC0080CC, 0x86CE0093CC0081CC,
-    0xCE0089CE0088CE00, 0x008ECE008CCE008A, 0x91CE0090CE008FCE,
-    0xCE0093CE0092CE00, 0x0096CE0095CE0094, 0x99CE0098CE0097CE,
-    0xCE009BCE009ACE00, 0x009ECE009DCE009C, 0xA1CE00A0CE009FCE,
-    0xCE00A4CE00A3CE00, 0x00A7CE00A6CE00A5, 0xAACE00A9CE00A8CE,
-    0xCE00ACCE00ABCE00, 0x00AFCE00AECE00AD, 0xB2CE00B1CE00B0CE,
-    0xCE00B4CE00B3CE00, 0x00B7CE00B6CE00B5, 0xBACE00B9CE00B8CE,
-    0xCE00BCCE00BBCE00, 0xCE0046BCCE0041BC, 0xCE0057BCCE0056BC,
-    0xCE006CBCCE0067BC, 0xCE0073BCCE006DBC, 0x00BFCE00BECE00BD,
-    0x82CF0081CF0080CF, 0xCF0084CF0083CF00, 0x0087CF0086CF0085,
-    0x8ACF0089CF0088CF, 0xCF008CCF008BCF00, 0x009CCF008ECF008D,
-    0x81D00080D0009DCF, 0xD00087D00083D000, 0x008ED0008DD0008C,
-    0xB1D000B0D00099D0, 0xD000B3D000B2D000, 0x00B6D000B5D000B4,
-    0xB9D000B8D000B7D0, 0xD000BBD000BAD000, 0x00BED000BDD000BC,
-    0x81D10080D100BFD0, 0xD10083D10082D100, 0x0086D10085D10084,
-    0x8AD10088D10087D1, 0xD1008CD1008BD100, 0x0090D1008ED1008D,
-    0x95D10093D10091D1, 0xD10097D10096D100, 0x009DD1009CD10098,
-    0xB6D1009FD1009ED1, 0xD20091D200B7D100, 0x00B1D200AFD200AB,
-    0x8FD30082D30081D3, 0xD30091D30090D300, 0x0096D30093D30092,
-    0x9AD30099D30097D3, 0xD3009CD3009BD300, 0x009FD3009ED3009D,
-    0xA4D300A3D300A2D3, 0xD300A6D300A5D300, 0x00AAD300A9D300A7,
-    0xADD300ACD300ABD3, 0xD300AFD300AED300, 0x00B2D300B1D300B0,
-    0xB5D300B4D300B3D3, 0xD500B9D300B8D300, 0xA5D5B4D50082D6A5,
-    0xB4D500ABD5B4D500, 0x00B6D5B4D500ADD5, 0x0090D700B6D5BED5,
-    0xD690D700B7D690D7, 0xD700BCD690D700B8, 0xD70091D7009CD790,
-    0xBFD691D700BCD691, 0xBCD692D70092D700, 0xBCD693D70093D700,
-    0xBCD694D70094D700, 0x95D700B9D695D700, 0x00BCD696D700BCD6,
-    0xD699D700BCD698D7, 0xD700BCD699D700B4, 0xD7009BD700BCD69A,
-    0xBFD69BD700BCD69B, 0xBCD69CD7009CD700, 0xBCD69ED7009DD700,
-    0xA1D700BCD6A0D700, 0xA3D700A2D700BCD6, 0x00BCD6A4D700BCD6,
-    0xD6A6D700BFD6A4D7, 0xD700BCD6A7D700BC, 0xD700BCD6A8D700A8,
-    0xBCD6A9D700BCD6A9, 0xD7BCD6A9D70081D7, 0xD70081D7A9D70082,
-    0xD700AAD70082D7A9, 0xB7D6B2D700BCD6AA, 0xD800A2D800A1D800,
-    0x00A5D800A4D800A3, 0x00A7D8A6D800A6D8, 0xD8A6D800ACD8A6D8,
-    0xD800AED8A6D800AD, 0xB2D8A6D800B1D8A6, 0xA6D80085D9A6D800,
-    0x0087D9A6D80086D9, 0xD9A6D80088D9A6D8, 0xD8008AD9A6D80089,
-    0x87DBA6D80086DBA6, 0xA6D80088DBA6D800, 0x0095DBA6D80090DB,
-    0xD883D9A7D800A7D8, 0x84D9A7D800B1D8A8, 0xD9A7D80087D984D9,
-    0xD800B4D9A7D8008B, 0xD800ACD8A8D800A8, 0xADD8A8D800ADD8A8,
-    0x00AED8A8D8008AD9, 0xD8008AD9AED8A8D8, 0xB2D8A8D800B1D8A8,
-    0xA8D80085D9A8D800, 0x0087D9A8D80086D9, 0xD9A8D80089D9A8D8,
-    0x00AAD800A9D8008A, 0xD8AAD800ACD8AAD8, 0xACD8AAD80085D9AC,
-    0xD9ACD8AAD80089D9, 0xD800ADD8AAD8008A, 0xAAD800ACD8ADD8AA,
-    0xD8AAD80085D9ADD8, 0x85D9AED8AAD800AE, 0x0089D9AED8AAD800,
-    0xD8008AD9AED8AAD8, 0xB2D8AAD800B1D8AA, 0xAAD80085D9AAD800,
-    0xD9AAD800ACD885D9, 0x85D9AAD800ADD885, 0xD985D9AAD800AED8,
-    0x8AD985D9AAD80089, 0xAAD80086D9AAD800, 0x0089D9AAD80087D9,
-    0x00ABD8008AD9AAD8, 0xD8ABD800ACD8ABD8, 0xD800B2D8ABD800B1,
-    0x86D9ABD80085D9AB, 0xABD80087D9ABD800, 0x008AD9ABD80089D9,
-    0x00ADD8ACD800ACD8, 0xD80089D9ADD8ACD8, 0xACD8008AD9ADD8AC,
-    0xD884D9ACD82084D9, 0xACD80087D984D9A7, 0xD885D9ACD80085D9,
-    0x89D985D9ACD800AD, 0x008AD985D9ACD800, 0xD9ACD80089D9ACD8,
-    0xD8ADD800ADD8008A, 0x8AD9ACD8ADD800AC, 0xADD80085D9ADD800,
-    0xD9ADD80089D985D9, 0x89D9ADD8008AD985, 0xAED8008AD9ADD800,
-    0xAED800ACD8AED800, 0x0085D9AED800ADD8, 0xD9AED80089D9AED8,
-    0x00B0D800AFD8008A, 0x00B1D800B0D9B0D8, 0x84D988D9B3D8B1D8,
-    0xB1D800B0D9B1D800, 0xD80084D9A7D88CDB, 0xD8B3D800B3D800B2,
-    0xADD8ACD8B3D800AC, 0x0089D9ACD8B3D800, 0xD8B3D800ADD8B3D8,
-    0xAED8B3D800ACD8AD, 0x0089D9AED8B3D800, 0xD8008AD9AED8B3D8,
-    0x85D9B3D800B1D8B3, 0x00ACD885D9B3D800, 0xD800ADD885D9B3D8,
-    0xB3D80085D985D9B3, 0x0089D9B3D80087D9, 0x00B4D8008AD9B3D8,
-    0xD8B4D800ACD8B4D8, 0xADD8B4D8008AD9AC, 0x0085D9ADD8B4D800,
-    0xD8008AD9ADD8B4D8, 0xB1D8B4D800AED8B4, 0xB4D80085D9B4D800,
-    0xD9B4D800AED885D9, 0x87D9B4D80085D985, 0xB4D80089D9B4D800,
-    0xB5D800B5D8008AD9, 0xD8ADD8B5D800ADD8, 0x8AD9ADD8B5D800AD,
-    0xB5D800AED8B5D800, 0xD884D9B5D800B1D8, 0x84D9B5D80085D9B9,
-    0xD984D9B5D80089D9, 0x84D984D9A7D82089, 0xD984D9B9D82087D9,
-    0xB3D888D92087D98A, 0xD9B5D80085D984D9, 0x85D9B5D80092DB84,
-    0x0085D985D9B5D800, 0xD9B5D80089D9B5D8, 0xD8B6D800B6D8008A,
-    0xD800ADD8B6D800AC, 0xB6D80089D9ADD8B6, 0xD8B6D8008AD9ADD8,
-    0x85D9AED8B6D800AE, 0xB6D800B1D8B6D800, 0x0089D9B6D80085D9,
-    0x00B7D8008AD9B6D8, 0xD9B7D800ADD8B7D8, 0xADD885D9B7D80085,
-    0x0085D985D9B7D800, 0xD8008AD985D9B7D8, 0x8AD9B7D80089D9B7,
-    0x85D9B8D800B8D800, 0xACD8B9D800B9D800, 0x0085D9ACD8B9D800,
-    0x87D98AD984D9B9D8, 0xB9D80085D9B9D800, 0xD9B9D80085D985D9,
-    0x85D9B9D80089D985, 0x0089D9B9D8008AD9, 0x00BAD8008AD9B9D8,
-    0xD9BAD800ACD8BAD8, 0x85D985D9BAD80085, 0x0089D985D9BAD800,
-    0xD8008AD985D9BAD8, 0x8AD9BAD80089D9BA, 0x80D9008BD980D900,
-    0xD98ED980D9008ED9, 0xD9008FD980D90091, 0x80D90091D98FD980,
-    0xD990D980D90090D9, 0xD90091D980D90091, 0xD90081D90092D980,
-    0xADD881D900ACD881, 0x81D900AED881D900, 0xD981D90085D9AED8,
-    0x8AD985D981D90085, 0x81D90089D981D900, 0x82D90082D9008AD9,
-    0xDB84D982D900ADD8, 0xD90085D982D90092, 0x82D900ADD885D982,
-    0xD982D90085D985D9, 0x89D982D9008AD985, 0x83D9008AD982D900,
-    0x83D900A7D883D900, 0x00ADD883D900ACD8, 0xD983D900AED883D9,
-    0xD90085D983D90084, 0x83D90085D985D983, 0xD983D9008AD985D9,
-    0xD9008AD983D90089, 0xD900A2D884D90084, 0xA5D884D900A3D884,
-    0x84D900A7D884D900, 0xD8ACD884D900ACD8, 0x85D9ACD884D900AC,
-    0x008AD9ACD884D900, 0xD884D900ADD884D9, 0xADD884D90085D9AD,
-    0xD9ADD884D90089D9, 0xD900AED884D9008A, 0x84D90085D9AED884,
-    0xD885D984D90085D9, 0x8AD985D984D900AD, 0x84D90087D984D900,
-    0x008AD984D90089D9, 0x00A7D885D90085D9, 0xD885D900ACD885D9,
-    0xACD885D900ADD8AC, 0xD9ACD885D900AED8, 0x8AD9ACD885D90085,
-    0x85D900ADD885D900, 0xD885D900ACD8ADD8, 0xADD885D90085D9AD,
-    0xD885D900AFD885D9, 0xAED885D9008AD9AD, 0x00ACD8AED885D900,
-    0xD90085D9AED885D9, 0x85D9008AD9AED885, 0xD985D985D90085D9,
-    0xD90089D985D9008A, 0xD90086D9008AD985, 0xACD886D900ACD886,
-    0xD9ACD886D900ADD8, 0x89D9ACD886D90085, 0x008AD9ACD886D900,
-    0xD886D900ADD886D9, 0xADD886D90085D9AD, 0xD9ADD886D90089D9,
-    0xD900AED886D9008A, 0xB2D886D900B1D886, 0x86D90085D986D900,
-    0xD986D90089D985D9, 0x86D986D9008AD985, 0x86D90087D986D900,
-    0x008AD986D90089D9, 0x00ACD887D90087D9, 0xD987D90085D987D9,
-    0x85D987D900ACD885, 0x0089D987D90085D9, 0xD987D9008AD987D9,
-    0xD888D90088D900B0, 0x88D90085D984D9B3, 0x89D90089D900B4D9,
-    0x8AD9008AD900B0D9, 0xD9ACD88AD900ACD8, 0xD900ADD88AD9008A,
-    0x8AD9008AD9ADD88A, 0x00B1D88AD900AED8, 0xD98AD900B2D88AD9,
-    0x85D985D98AD90085, 0x008AD985D98AD900, 0xD98AD90086D98AD9,
-    0xD90089D98AD90087, 0xB4D98AD9008AD98A, 0xD900AFD900AED900,
-    0x00BAD900B9D900B1, 0xBFD900BED900BBD9, 0xDA0083DA0080DA00,
-    0x0087DA0086DA0084, 0x8DDA008CDA0088DA, 0xDA0091DA008EDA00,
-    0x00A4DA00A1DA0098, 0xADDA00A9DA00A6DA, 0xDA00B1DA00AFDA00,
-    0x00BBDA00BADA00B3, 0x81DB0080DB00BEDA, 0xDB0085DB0082DB00,
-    0xD987DB0087DB0086, 0x0089DB0088DB00B4, 0x90DB008CDB008BDB,
-    0xE00093DB0092DB00, 0xA4E000BCA4E095A4, 0x97A4E000BCA4E096,
-    0xE09CA4E000BCA4E0, 0xA4E0A1A4E000BCA4, 0xBCA4E0A2A4E000BC,
-    0xABA4E000A9A4E000, 0xE0AFA4E000BCA4E0, 0xE000B1A4E000BCA4,
-    0xA6E0A1A6E000B4A4, 0xBCA6E0A2A6E000BC, 0x00BCA6E0AFA6E000,
-    0x008CA7E0008BA7E0, 0xE000BCA8E096A8E0, 0xA8E000BCA8E097A8,
-    0xABA8E000BCA8E09C, 0xE0B2A8E000BCA8E0, 0xA8E0B8A8E000BCA8,
-    0xBCACE0A1ACE000BC, 0x00BCACE0A2ACE000, 0x008BADE00088ADE0,
-    0x0094AEE0008CADE0, 0x008BAFE0008AAFE0, 0x0088B1E0008CAFE0,
-    0x0087B3E00080B3E0, 0x008AB3E00088B3E0, 0x008AB5E0008BB3E0,
-    0x008CB5E0008BB5E0, 0x009CB7E0009AB7E0, 0x009EB7E0009DB7E0,
-    0xE000B2B8E08DB9E0, 0xBAE00099BAE0ABBA, 0x8DBBE000A1BAE0AB,
-    0x008BBCE000B2BAE0, 0xE000B5BEE080BDE0, 0xBDE000B7BEE082BD,
-    0x91BDE000B7BEE08C, 0xE096BDE000B7BEE0, 0xBEE09BBDE000B7BE,
-    0xB2BDE0B1BDE000B7, 0x00B4BDE0B1BDE000, 0xE00080BEE0B1BDE0,
-    0xBEE000B5BEE090BE, 0x9CBEE000B7BEE092, 0xE0A1BEE000B7BEE0,
-    0xBEE0A6BEE000B7BE, 0xB7BEE0ABBEE000B7, 0xE0B1BDE0B2BEE000,
-    0xBEE0B2BEE00080BE, 0xB1BDE0B3BEE00080, 0xE0B3BEE00080BEE0,
-    0xE100A680E10080BE, 0xE1008084E1009C83, 0xE1008284E1008184,
-    0xE1008484E1008384, 0xE1008684E1008584, 0xE1008884E1008784,
-    0xE1008A84E1008984, 0xE1008C84E1008B84, 0xE1008E84E1008D84,
-    0xE1009084E1008F84, 0xE1009284E1009184, 0xE1009584E1009484,
-    0xE1009C84E1009A84, 0xE1009E84E1009D84, 0xE100A184E100A084,
-    0xE100A384E100A284, 0xE100A984E100A784, 0xE100AC84E100AB84,
-    0xE100AE84E100AD84, 0xE100B284E100AF84, 0xE1008085E100B684,
-    0xE1008C85E1008785, 0xE1009885E1009785, 0xE100A085E1009985,
-    0xE100A285E100A185, 0xE100A485E100A385, 0xE100A685E100A585,
-    0xE100A885E100A785, 0xE100AA85E100A985, 0xE100AC85E100AB85,
-    0xE100AE85E100AD85, 0xE100B085E100AF85, 0xE100B285E100B185,
-    0xE100B485E100B385, 0xE1008486E100B585, 0xE1008886E1008586,
-    0xE1009286E1009186, 0xE1009E86E1009486, 0xE100AA86E100A186,
-    0xE100AD86E100AC86, 0xE100B186E100B086, 0xE100B386E100B286,
-    0xE100B586E100B486, 0xE1008887E1008787, 0xE1008E87E1008C87,
-    0xE1009787E1009387, 0xE1009D87E1009987, 0xE100B187E1009F87,
-    0xE10086ACE100B287, 0xE1008AACE10088AC, 0xE1008EACE1008CAC,
-    0xE100BBACE10092AC, 0xE10080ADE100BDAC, 0xE10083ADE10081AD,
-    0xE10096B4E10082B4, 0xE1009CB4E10097B4, 0xE100A5B4E1009DB4,
-    0xE10085B6E100BBB5, 0xE10080B8E10091B6, 0xE10082B8E10081B8,
-    0xE10084B8E10083B8, 0xE10086B8E10085B8, 0xE10088B8E10087B8,
-    0xE1008AB8E10089B8, 0xE1008CB8E1008BB8, 0xE1008EB8E1008DB8,
-    0xE10090B8E1008FB8, 0xE10092B8E10091B8, 0xE10094B8E10093B8,
-    0xE10096B8E10095B8, 0xE10098B8E10097B8, 0xE1009AB8E10099B8,
-    0xE1009CB8E1009BB8, 0xE1009EB8E1009DB8, 0xE100A0B8E1009FB8,
-    0xE100A2B8E100A1B8, 0xE100A4B8E100A3B8, 0xE100A6B8E100A5B8,
-    0xE100A8B8E100A7B8, 0xE100AAB8E100A9B8, 0xE100ACB8E100ABB8,
-    0xE100AEB8E100ADB8, 0xE100B0B8E100AFB8, 0xE100B2B8E100B1B8,
-    0xE100B4B8E100B3B8, 0xE100B6B8E100B5B8, 0xE100B8B8E100B7B8,
-    0xE100BAB8E100B9B8, 0xE100BCB8E100BBB8, 0xE100BEB8E100BDB8,
-    0xE10080B9E100BFB8, 0xE10082B9E10081B9, 0xE10084B9E10083B9,
-    0xE10086B9E10085B9, 0xE10088B9E10087B9, 0xE1008AB9E10089B9,
-    0xE1008CB9E1008BB9, 0xE1008EB9E1008DB9, 0xE10090B9E1008FB9,
-    0xE10092B9E10091B9, 0xE10094B9E10093B9, 0xE10096B9E10095B9,
-    0xE10098B9E10097B9, 0xE1009AB9E10099B9, 0xE1009CB9E1009BB9,
-    0xE1009EB9E1009DB9, 0xE100A0B9E1009FB9, 0xE100A2B9E100A1B9,
-    0xE100A4B9E100A3B9, 0xE100A6B9E100A5B9, 0xE100A8B9E100A7B9,
-    0xE100AAB9E100A9B9, 0xE100ACB9E100ABB9, 0xE100AEB9E100ADB9,
-    0xE100B0B9E100AFB9, 0xE100B2B9E100B1B9, 0xE100B4B9E100B3B9,
-    0xE100B6B9E100B5B9, 0xE100B8B9E100B7B9, 0xE100BAB9E100B9B9,
-    0xE100BCB9E100BBB9, 0xE100BEB9E100BDB9, 0xE10080BAE100BFB9,
-    0xE10082BAE10081BA, 0xE10084BAE10083BA, 0xE10086BAE10085BA,
-    0xE10088BAE10087BA, 0xE1008ABAE10089BA, 0xE1008CBAE1008BBA,
-    0xE1008EBAE1008DBA, 0xE10090BAE1008FBA, 0xE10092BAE10091BA,
-    0xE10094BAE10093BA, 0xE10096BAE10095BA, 0xE10098BAE10097BA,
-    0xE100A0BAE10099BA, 0xE100A2BAE100A1BA, 0xE100A4BAE100A3BA,
-    0xE100A6BAE100A5BA, 0xE100A8BAE100A7BA, 0xE100AABAE100A9BA,
-    0xE100ACBAE100ABBA, 0xE100AEBAE100ADBA, 0xE100B0BAE100AFBA,
-    0xE100B2BAE100B1BA, 0xE100B4BAE100B3BA, 0xE100B6BAE100B5BA,
-    0xE100B8BAE100B7BA, 0xE100BABAE100B9BA, 0xE100BCBAE100BBBA,
-    0xE100BEBAE100BDBA, 0xE10080BBE100BFBA, 0xE10082BBE10081BB,
-    0xE10084BBE10083BB, 0xE10086BBE10085BB, 0xE10088BBE10087BB,
-    0xE1008ABBE10089BB, 0xE1008CBBE1008BBB, 0xE1008EBBE1008DBB,
-    0xE10090BBE1008FBB, 0xE10092BBE10091BB, 0xE10094BBE10093BB,
-    0xE10096BBE10095BB, 0xE10098BBE10097BB, 0xE1009ABBE10099BB,
-    0xE1009CBBE1009BBB, 0xE1009EBBE1009DBB, 0xE100A0BBE1009FBB,
-    0xE100A2BBE100A1BB, 0xE100A4BBE100A3BB, 0xE100A6BBE100A5BB,
-    0xE100A8BBE100A7BB, 0xE100AABBE100A9BB, 0xE100ACBBE100ABBB,
-    0xE100AEBBE100ADBB, 0xE100B0BBE100AFBB, 0xE100B2BBE100B1BB,
-    0xE100B4BBE100B3BB, 0xE100B6BBE100B5BB, 0xE100B8BBE100B7BB,
-    0xE10080BCE100B9BB, 0xE10082BCE10081BC, 0xE10084BCE10083BC,
-    0xE10086BCE10085BC, 0xE10088BCE10087BC, 0xE1008ABCE10089BC,
-    0xE1008CBCE1008BBC, 0xE1008EBCE1008DBC, 0xE10090BCE1008FBC,
-    0xE10092BCE10091BC, 0xE10094BCE10093BC, 0xE10098BCE10095BC,
-    0xE1009ABCE10099BC, 0xE1009CBCE1009BBC, 0xE100A0BCE1009DBC,
-    0xE100A2BCE100A1BC, 0xE100A4BCE100A3BC, 0xE100A6BCE100A5BC,
-    0xE100A8BCE100A7BC, 0xE100AABCE100A9BC, 0xE100ACBCE100ABBC,
-    0xE100AEBCE100ADBC, 0xE100B0BCE100AFBC, 0xE100B2BCE100B1BC,
-    0xE100B4BCE100B3BC, 0xE100B6BCE100B5BC, 0xE100B8BCE100B7BC,
-    0xE100BABCE100B9BC, 0xE100BCBCE100BBBC, 0xE100BEBCE100BDBC,
-    0xE10080BDE100BFBC, 0xE10082BDE10081BD, 0xE10084BDE10083BD,
-    0xE10088BDE10085BD, 0xE1008ABDE10089BD, 0xE1008CBDE1008BBD,
-    0xE10090BDE1008DBD, 0xE10092BDE10091BD, 0xE10094BDE10093BD,
-    0xE10096BDE10095BD, 0xE10099BDE10097BD, 0xE1009DBDE1009BBD,
-    0xE100A0BDE1009FBD, 0xE100A2BDE100A1BD, 0xE100A4BDE100A3BD,
-    0xE100A6BDE100A5BD, 0xE100A8BDE100A7BD, 0xE100AABDE100A9BD,
-    0xE100ACBDE100ABBD, 0xE100AEBDE100ADBD, 0xE100B0BDE100AFBD,
-    0xE100B4BDE100B2BD, 0xE100B8BDE100B6BD, 0xE100BCBDE100BABD,
-    0xE10081BEE10080BE, 0xE10083BEE10082BE, 0xE10085BEE10084BE,
-    0xE10087BEE10086BE, 0xE10089BEE10088BE, 0xE1008BBEE1008ABE,
-    0xE1008DBEE1008CBE, 0xE1008FBEE1008EBE, 0xE10091BEE10090BE,
-    0xE10093BEE10092BE, 0xE10095BEE10094BE, 0xE10097BEE10096BE,
-    0xE10099BEE10098BE, 0xE1009BBEE1009ABE, 0xE1009DBEE1009CBE,
-    0xE1009FBEE1009EBE, 0xE100A1BEE100A0BE, 0xE100A3BEE100A2BE,
-    0xE100A5BEE100A4BE, 0xE100A7BEE100A6BE, 0xE100A9BEE100A8BE,
-    0xE100ABBEE100AABE, 0xE100ADBEE100ACBE, 0xE100AFBEE100AEBE,
-    0xE100B1BEE100B0BE, 0xE100B3BEE100B2BE, 0xE100B6BEE100B4BE,
-    0xE100B8BEE100B7BE, 0xE100BABEE100B9BE, 0xE10082BFE100BCBE,
-    0xE10084BFE10083BF, 0xE10087BFE10086BF, 0xE1008ABFE10088BF,
-    0xE10090BFE1008CBF, 0xE10092BFE10091BF, 0xE10097BFE10096BF,
-    0xE10099BFE10098BF, 0xE100A0BFE1009ABF, 0xE100A2BFE100A1BF,
-    0xE100A5BFE100A4BF, 0xE100A7BFE100A6BF, 0xE100A9BFE100A8BF,
-    0xE100ACBFE100AABF, 0xE100B3BFE100B2BF, 0xE100B6BFE100B4BF,
-    0xE100B8BFE100B7BF, 0xE200BCBFE100BABF, 0xE2009380E2009080,
-    0x80E2B280E2009480, 0xB280E2B280E200B2, 0xE2B280E200B280E2,
-    0xB280E2B280E2B280, 0x00B580E2B580E200, 0x80E2B580E2B580E2,
-    0x86E200A982E200B5, 0x86E2009186E20090, 0x86E2009386E20092,
-    0x86E2009B86E2009A, 0x87E2008D87E200AE, 0x88E2008F87E2008E,
-    0x88E2008488E20082, 0x88E2008988E20087, 0x88E2009188E2008C,
-    0x88E200A488E20092, 0xAB88E2AB88E200A6, 0xE2AB88E2AB88E200,
-    0x88E2AB88E200AB88, 0x00AB88E2AB88E2AB, 0xE200AE88E2AE88E2,
-    0xAE88E2AE88E2AE88, 0x8489E2008189E200, 0x8989E2008789E200,
-    0xA289E200A089E200, 0xAE89E200AD89E200, 0xB089E200AF89E200,
-    0xB489E200B189E200, 0xB889E200B589E200, 0x808AE200B989E200,
-    0x848AE200818AE200, 0x888AE200858AE200, 0xAC8AE200898AE200,
-    0xAE8AE200AD8AE200, 0xA08BE200AF8AE200, 0xA28BE200A18BE200,
-    0xAA8BE200A38BE200, 0xAC8BE200AB8BE200, 0x8294E200AD8BE200,
-    0x8B97E200A096E200, 0x86A6E20085A6E200, 0xE200B8CC9DABE200,
-    0xE300A1B5E200B1B1, 0xE3008280E3008180, 0xE3008980E3008880,
-    0xE3008B80E3008A80, 0xE3008D80E3008C80, 0xE3008F80E3008E80,
-    0xE3009180E3009080, 0xE3009480E3009280, 0xE3009580E3539480,
-    0x9580E389B8E49480, 0xE38CBAE49480E300, 0x8BE59480E3009580,
-    0x9480E3009580E39D, 0xE3009580E389AEE5, 0x9580E39389E69480,
-    0xE39795E69480E300, 0x9CE69480E3009580, 0x9480E3009580E3AC,
-    0xE3009580E3B982E7, 0x9580E3979BE79480, 0x9680E3009580E300,
-    0x8C81E3009780E300, 0x9081E3008E81E300, 0x9481E3009281E300,
-    0x9881E3009681E300, 0x9C81E3009A81E300, 0xA081E3009E81E300,
-    0xA581E300A281E300, 0xA981E300A781E300, 0xB181E300B081E300,
-    0xB481E300B381E300, 0xB781E300B681E300, 0xBA81E300B981E300,
-    0x008B81E3BB81E300, 0x00BD81E300BC81E3, 0xE3008A82E38882E3,
-    0xE3009982E3009482, 0xE3009E82E3009A82, 0xE300A282E300A182,
-    0xBC83E39183E3A282, 0xE3A282E3008883E3, 0xA182E39583E3AB83,
-    0xE3B383E3A282E300, 0x82E300A282E39A83, 0x00AB83E3BC83E3A2,
-    0x00A482E300A382E3, 0x83E38B83E3A482E3, 0xA482E300B082E3B3,
-    0xE3008183E3B383E3, 0xE300A682E300A582, 0xB383E3A982E3A682,
-    0xA882E300A782E300, 0xE3B982E3A882E300, 0x8983E3BC83E3AF82,
-    0xE3BC83E3A882E300, 0x82E300BC83E3AB82, 0x82E300AA82E300A9,
-    0x00B982E3B383E3AA, 0x83E3BC83E3AA82E3, 0x82E300AB82E300A0,
-    0x00AA83E3A482E3AB, 0x83E3A983E3AB82E3, 0xAB82E3008883E383,
-    0x83E3AA83E3AD83E3, 0x82E300AC82E300BC, 0x00B383E3AD83E3AC,
-    0x83E3B383E3AC82E3, 0x82E300AD82E3009E, 0xE3AA83E3A583E3AD,
-    0x83E3AD82E300BC83, 0xAD83E3AD82E300AD, 0x83E3A983E3B082E3,
-    0xAD83E3AD82E300A0, 0x83E3BC83E3A183E3, 0xAD82E300AB83E388,
-    0x83E3AF83E3AD83E3, 0xAE82E3008883E383, 0x00AC82E3AE82E300,
-    0x83E38B83E3AE82E3, 0xAB83E3AE82E300BC, 0xE300BC83E38083E3,
-    0x83E3AF82E300AF82, 0xE3A482E3BC82E3AB, 0x83E3AF82E300AD83,
-    0x008D83E3BC83E3AD, 0xE3B082E300B082E3, 0x82E300A083E3A983,
-    0xE3A083E3A983E3B0, 0x82E300B383E38883, 0xBC83E3B182E300B1,
-    0x00B282E300B982E3, 0xE3B382E300B382E3, 0x83E3B382E300B382,
-    0xAB83E3B382E30088, 0xE3B382E3008A83E3, 0x82E3009D83E3BC83,
-    0x82E300B582E300B4, 0xE3AF82E3A482E3B5, 0x83E3B582E300AB83,
-    0xE3BC83E38183E3B3, 0xE300B682E300A083, 0x83E3B782E300B782,
-    0x00B082E3B383E3AA, 0x00B982E300B882E3, 0x00BB82E300BA82E3,
-    0x83E3B383E3BB82E3, 0xB383E3BB82E30081, 0x00BC82E3008883E3,
-    0x00BE82E300BD82E3, 0x008083E300BF82E3, 0x82E3BC83E38083E3,
-    0x83E3008183E300B9, 0x83E3008383E30082, 0x83E3008583E30084,
-    0x83E3008783E30086, 0x8883E300B782E387, 0x00B383E38883E300,
-    0xE38983E3008983E3, 0xE3008A83E300AB83, 0x83E3008E83E38A83,
-    0x83E3008C83E3008B, 0x83E3008E83E3008D, 0x008883E38383E38E,
-    0xE38F83E3008F83E3, 0x83E3008483E3A482, 0xBC83E39083E30090,
-    0xE300AB83E3AC83E3, 0x83E39183E3009183, 0xE3B383E3BB82E3BC,
-    0x83E39183E3008883, 0x9283E3008483E3BC, 0x9383E3009383E300,
-    0x009483E300AB83E3, 0x82E3A282E39483E3, 0x00AB83E38883E3B9,
-    0x83E3AF82E39483E3, 0xB382E39483E300AB, 0x9583E3009583E300,
-    0x83E3A983E3A182E3, 0x9583E3008983E383, 0x83E3BC83E3A382E3,
-    0xA983E39583E30088, 0x009683E300B383E3, 0x82E38383E39683E3,
-    0x00AB83E3A782E3B7, 0x009883E3009783E3, 0x82E3AF82E39883E3,
-    0x00AB83E3BC83E3BF, 0x83E3AB83E39883E3, 0x83E3009983E30084,
-    0x00BF82E3BC83E399, 0xE39A83E3009A83E3, 0x83E39A83E300BD82,
-    0x9A83E3009283E38B, 0xE300B982E3B383E3, 0xB882E3BC83E39A83,
-    0x9B83E3009B83E300, 0xE39B83E300B383E3, 0x83E300AB83E3BC83,
-    0x00B383E3BC83E39B, 0xE39C83E3009C83E3, 0x83E3008883E3AB83,
-    0xA482E39D83E3009D, 0xE3008883E3B383E3, 0x8983E3B383E39D83,
-    0x9E83E3009E83E300, 0x83E3AF82E3A482E3, 0xA482E39E83E300AD,
-    0xE39E83E300AB83E3, 0x83E3008F83E38383, 0x00AF82E3AB83E39E,
-    0x82E3B383E39E83E3, 0x00B383E3A783E3B7, 0xE39F83E3009F83E3,
-    0xB383E3AD83E3AF82, 0x00AA83E39F83E300, 0x83E3AA83E39F83E3,
-    0x00AB83E3BC83E390, 0x00A183E300A083E3, 0xE300AC82E3A183E3,
-    0x8883E3AC82E3A183, 0xE3A183E300B383E3, 0xAB83E38883E3BC83,
-    0xA383E300A283E300, 0xA483E300A483E300, 0xE3008983E3BC83E3,
-    0xAB83E3BC83E3A483, 0xA683E300A583E300, 0xE3A282E3A683E300,
-    0xE300A783E300B383, 0xE300A983E300A883, 0x83E3AA83E300AA83,
-    0x00AB83E38883E383, 0xE300A983E3AA83E3, 0x83E3AB83E300AB83,
-    0xAB83E300BC83E394, 0x83E39683E3BC83E3, 0x83E300AC83E300AB,
-    0xAC83E300A083E3AC, 0x82E38883E3B383E3, 0xAD83E300B383E3B2,
-    0xAF83E300AF83E300, 0xE3008883E38383E3, 0xE300B183E300B083,
-    0xE300B383E300B283, 0xE300B783E300B483, 0xE300B983E300B883,
-    0xE300BB83E300BA83, 0xE300BE83E300BC83, 0xE300B992E3009E92,
-    0xE3009F93E300BB92, 0xE300AE9BE3009594, 0xE300819EE300BC9B,
-    0xE300A2A1E300AFA0, 0xE30087A3E300BCA1, 0xE3009CA4E300A3A3,
-    0xE300AEA8E300BAA4, 0xE300A4ABE300ACA9, 0xE30099ACE30088AC,
-    0xE3009DAEE30089AD, 0xE3008EB1E30098B0, 0xE30096B6E300B3B4,
-    0xE300B8BAE300ACBA, 0xE400BCBFE3009BBC, 0xE4009880E4008880,
-    0xE4008681E400B980, 0xE400A383E4009682, 0xE4008288E400AF84,
-    0xE400A08AE400A788, 0xE400B48CE400818C, 0xE400958FE400998D,
-    0xE4008B90E400998F, 0xE400AB94E400AB91, 0xE400A195E4009D95,
-    0xE4009797E400AB95, 0xE400B598E400B997, 0xE400879BE400BE9A,
-    0xE400A6A7E40095A6, 0xE400B6A9E400AEA9, 0xE400B3ACE400B2AA,
-    0xE4008EB3E4008EAF, 0xE400B8B3E400ADB3, 0xE40080B8E40096B5,
-    0xE40083B8E40081B8, 0xE4008AB8E40089B8, 0xE4008DB8E4008BB8,
-    0xE400A6B8E40099B8, 0xE400ADB8E400A8B8, 0xE400B6B8E400B2B8,
-    0xE400B9B8E400B8B8, 0xE400BFB8E400BDB8, 0xE40099B9E40081B9,
-    0xE40082BAE4009DB9, 0xE40086BAE40085BA, 0xE40094BAE4008CBA,
-    0xE400A4BAE400A0BA, 0xE400BABAE400AEBA, 0xE4008CBBE40080BB,
-    0x92E5A4BBE400A4BB, 0xBCE40081BCE4008C, 0xBEE400A0BDE40091,
-    0xBEE40086BEE40080, 0xBEE400AEBEE4008B, 0x80E500BFBEE400BB,
-    0x81E500AB80E50082, 0x83E5009982E500BA, 0x83E5009A83E5008F,
-    0x84E500AA84E500A7, 0x85E5008085E500BF, 0x85E5008D85E50085,
-    0x85E500A485E50094, 0x85E500A785E500A5, 0x85E500A985E500A8,
-    0x85E500AD85E500AB, 0x86E5008086E500B7, 0x86E5008D86E50082,
-    0x86E5009586E50092, 0x86E5009786E50096, 0x86E500A486E50099,
-    0x86E500AC86E500AB, 0x87E500B786E500B5, 0x87E5008C87E50089,
-    0x87E5009E87E5009C, 0x88E500B587E500A0, 0x88E5008388E50080,
-    0x88E5009788E50087, 0x88E500A988E5009D, 0x89E500BB88E500BA,
-    0x89E5008D89E50086, 0x8AE500B789E500B2, 0x8AE5009B8AE50089,
-    0x8AE500B38AE500A3, 0x8BE500878BE500B4, 0x8BE500928BE50089,
-    0x8BE500A48BE5009E, 0x8BE500B98BE500B5, 0x8CE500858CE500BA,
-    0x8CE500958CE50086, 0x8CE5009A8CE50097, 0x8CE500BB8CE500B8,
-    0x8DE500818DE500BF, 0x8DE500858DE50084, 0x8DE500918DE50089,
-    0x8DE5009A8DE50094, 0x8DE500A98DE5009C, 0x8DE500B38DE500B0,
-    0x8DE500BD8DE500B5, 0x8EE500828EE500BF, 0x8FE500838FE500B6,
-    0x8FE5008A8FE50088, 0x8FE5009F8FE5008C, 0x8FE500A58FE500A3,
-    0x8FE500AF8FE500AB, 0x90E500B38FE500B1, 0x90E5008890E50086,
-    0x90E5008F90E5008D, 0x90E500B890E5009D, 0x91E5008291E500B9,
-    0x92E500A891E50088, 0x92E500A292E5009E, 0x94E500B693E500BD,
-    0x95E5008F95E50090, 0x95E5009595E50093, 0x96E5008496E500A3,
-    0x96E5009996E50087, 0x96E500AB96E5009D, 0x97E500B696E500B3,
-    0x97E5008297E50080, 0x99E5008698E500A2, 0x99E500A899E50091,
-    0x9BE500979BE500B4, 0x9CE500B99BE5009B, 0x9CE500979CE50096,
-    0x9EE500B09CE5009F, 0x9FE5008E9FE5008B, 0xA0E5008DA0E500B4,
-    0xA1E500B2A0E500B1, 0xA1E5009AA1E50080, 0xA2E500A8A2E5009E,
-    0xA3E500B3A2E500AC, 0xA3E5009FA3E50098, 0xA3E500AEA3E500AB,
-    0xA3E500B2A3E500B0, 0xA4E50082A4E500B7, 0xA4E5008AA4E50086,
-    0xA4E5009AA4E50095, 0xA4E500A2A4E5009C, 0xA3ADE6A7A4E500A7,
-    0x84A5E500A9A4E500, 0x91A5E50088A5E500, 0xA2A5E50094A5E500,
-    0x98A7E500B3A5E500, 0x9BA8E500ACA7E500, 0xA2A9E500A7A8E500,
-    0xB5AAE500A6A9E500, 0xA8ACE50088ACE500, 0x90ADE500BEACE500,
-    0xA6ADE50097ADE500, 0x85AEE50080AEE500, 0x83AFE50097AEE500,
-    0xA7AFE50098AFE500, 0xB3AFE500AEAFE500, 0xBFAFE500B8AFE500,
-    0x8FB0E50086B0E500, 0xB8B0E500A2B0E500, 0xA0B1E500BFB0E500,
-    0xA4B1E500A2B1E500, 0xAEB1E500A5B1E500, 0x8DB2E500B1B1E500,
-    0x99B4E50080B3E500, 0x90B5E50083B5E500, 0xAEB5E500ABB5E500,
-    0xB2B6E500BCB5E500, 0x9BB7E500BAB6E500, 0xA2B7E500A1B7E500,
-    0xA6B7E500A5B7E500, 0xBDB7E500B1B7E500, 0xA8B8E500BEB7E500,
-    0xA9B9E500BDB8E500, 0xB3B9E500B2B9E500, 0x00B4B9E5009088E6,
-    0x00BCB9E500BAB9E5, 0x00A6BAE500BFB9E5, 0x00B3BAE500B0BAE5,
-    0x0089BBE500B6BAE5, 0x0092BBE5008ABBE5, 0x0099BBE50093BBE5,
-    0x00B4BBE500ACBBE5, 0x0084BCE500BEBBE5, 0x0093BCE5008BBCE5,
-    0x0090BDE500A2BCE5, 0x00A1BDE50093BDE5, 0x00A9BDE500A2BDE5,
-    0x00B3BDE500ABBDE5, 0x008CBEE5008BBEE5, 0x009ABEE50097BEE5,
-    0x00ADBEE500A9BEE5, 0x008DBFE50083BFE5, 0x00B5BFE50097BFE5,
-    0x009280E600B9BFE5, 0x00B581E6009C80E6, 0x009482E6008182E6,
-    0x009883E6008783E6, 0x008884E600A183E6, 0x008885E6008485E6,
-    0x008E85E6008C85E6, 0x00A885E600A085E6, 0x008E86E600BA85E6,
-    0x00A486E6009086E6, 0x00B286E600AF86E6, 0x00B287E6009E87E6,
-    0x008088E600B687E6, 0x009088E6008888E6, 0x00AE88E6009B88E6,
-    0x00B688E600B488E6, 0x009389E6008B89E6, 0x00958AE6009D89E6,
-    0x00898BE600B18AE6, 0x00938BE6008F8BE6, 0x00BC8BE600948BE6,
-    0x00878CE600BE8BE6, 0x00908DE600BD8CE6, 0x00A88DE600958DE6,
-    0x00838EE600BB8DE6, 0x00A98EE600A08EE6, 0x00858FE600848FE6,
-    0x009C90E600A48FE6, 0x009291E600A290E6, 0x00B791E600A991E6,
-    0x009A92E600BE91E6, 0x008493E6009D92E6, 0x00B494E600AF94E6,
-    0x009695E6008F95E6, 0x00B895E600AC95E6, 0x009796E6008796E6,
-    0x00A496E6009996E6, 0x00B996E600B096E6, 0x00A097E6008597E6,
-    0x00A397E600A297E6, 0xE68E98E600A597E6, 0xE6009398E600BBB2,
-    0x92E5AD98E600A098, 0x99E6008999E6008C, 0x9AE600889AE600B4,
-    0x9AE6009C9AE60091, 0x9BE600869BE600B4, 0x9BE600B49BE600B0,
-    0x9CE600809CE600B8, 0x9CE600899CE60088, 0x9CE6009B9CE60097,
-    0x9DE600A89CE600A1, 0x9DE600939DE6008E, 0x9DE6009E9DE60096,
-    0x9EE600859EE600BB, 0x9FE600B39FE60097, 0xA0E60097A0E600BA,
-    0xA0E600AAA0E6009F, 0xE79ABCE48FBCE5AA, 0xE60092A1E600BEA4,
-    0xE60085A2E60081A2, 0xE600A8A2E6008EA2, 0xE60082A5E60094A4,
-    0xE600AAA7E600A3A6, 0xE60093A8E60082A8, 0xE60093ABE600A8AA,
-    0xE60084ACE6009BAB, 0xE600A1ACE600A0AC, 0xE600A2ADE60094AD,
-    0xE600B2ADE600A3AD, 0xE600B9ADE600B7AD, 0xE600AEAEE6009FAE,
-    0xE600BAAEE600B3AE, 0xE6008BAFE600BBAE, 0xE60094AFE6008DAF,
-    0xE6008FB0E6009BAF, 0xE600B4B0E60094B0, 0xE600A7B1E6008EB1,
-    0xE600BFB2E60088B2, 0xE6008DB3E6008CB3, 0xE600A8B3E600A5B3,
-    0xE6009BB4E60096B4, 0xE600B4B4E6009EB4, 0xE60081B5E600BEB4,
-    0xE600AAB5E600A9B5, 0xE600B8B5E600B7B5, 0xE6008BB7E60085B6,
-    0xE600AAB7E6009AB7, 0xE6009AB8E600B9B7, 0xE600AEB9E600AFB8,
-    0xE6009CBAE60080BA, 0xE60087BBE600BABA, 0xE60091BBE6008BBB,
-    0xE6008FBCE6009BBB, 0xE600A2BCE60094BC, 0xE600AEBDE600A3BC,
-    0xE600ABBFE60086BF, 0xE7009B80E700BEBF, 0xE700B980E7009E80,
-    0xE700AB81E7008A81, 0xE700B781E700B081, 0xE7009982E700BD81,
-    0xE7008883E700AD82, 0xE700A184E7009983, 0xE7008985E7008585,
-    0xE7009C86E700AE85, 0xE7009087E7008E87, 0xE7009B88E7009088,
-    0xE700AA88E700A888, 0xE700B588E700AB88, 0xE700BB88E700B688,
-    0xE7008789E700BF88, 0xE7009989E7009089, 0xE700A289E7009B89,
-    0xE700808AE700B989, 0xE700AC8AE700958A, 0xE700808BE700AF8A,
-    0xE700AA8CE700BC8B, 0xE700BA8DE700B58D, 0xE700878EE700848E,
-    0xE7008B8EE700898E, 0xE700B28EE700A58E, 0xE7008690E7009E8F,
-    0xE700A290E7008990, 0xE7009C91E7008791, 0xE700B191E700A991,
-    0xE7008992E7008592, 0xE7008A93E7009892, 0xE700A693E7009C93,
-    0xE7009894E7008694, 0xE700A494E7009F94, 0xE700B094E700A894,
-    0xE700B394E700B294, 0xE700BB94E700B794, 0xE7009995E700BE94,
-    0xE700B095E700A595, 0xE7009296E7008B96, 0xE7009098E700A297,
-    0xE7009F98E7009D98, 0xE700A999E7008299, 0xE700BD99E700B699,
-    0xE700BF9AE700AE9A, 0xE7009B9BE7008A9B, 0xE700A79BE700A39B,
-    0xE700B49BE700AE9B, 0xE7009E9CE700819C, 0xE700809DE7009F9C,
-    0xE7008B9EE7008A9D, 0xE7009B9FE700A79E, 0xE700B39FE700A29F,
-    0xE700ABA1E7008EA1, 0xE70091A2E7008CA2, 0xE7008CA3E7008AA3,
-    0xE700AAA4E700BBA3, 0xE700BCA4E700BAA4, 0xE70088A5E700BEA4,
-    0xE70090A5E70089A5, 0xE7009DA5E70096A5, 0xE700A5A5E7009EA5,
-    0xE70081A6E700BFA5, 0xE7008EA6E7008DA6, 0xE700AEA6E7008FA6,
-    0xE700BEA6E700B8A6, 0xE70098A7E7008AA7, 0xE7009CA8E700ABA7,
-    0xE7008AA9E70080A9, 0xE700B4A9E7008FA9, 0xE70081AAE700BAA9,
-    0xE7008BABE700B1AA, 0xE700B9ABE700AEAB, 0xE7008FAEE700A0AC,
-    0xE70086AFE70080AF, 0xE700BEB0E70089AF, 0xE700B3B1E700A0B1,
-    0xE70092B2E700BBB1, 0xE70092B3E700BEB2, 0xE700A3B3E70096B3,
-    0xE700A8B3E700A7B3, 0xE70080B4E700B8B3, 0xE700A2B4E70090B4,
-    0xE70082B5E700AFB4, 0xE700A3B5E7009BB5, 0xE700BEB6E700A0B6,
-    0xE700B4B7E70087B7, 0xE70089B8E70082B8, 0xE70081B9E700B7B8,
-    0xE700B6BCE70085B9, 0xE70091BDE700BEBC, 0xE700B9BDE700B2BD,
-    0xE70085BEE700BABD, 0xE70095BEE7008ABE, 0xE700BDBEE7009ABE,
-    0xE8008180E800BABF, 0xE8008C80E8008580, 0xE800B380E8009280,
-    0xE800A081E8008681, 0xE800B081E800AF81, 0xE800BF81E800BE81,
-    0xE8008B82E8008982, 0xE800B282E800AD82, 0xE800BE84E8008384,
-    0xE800A387E8009887, 0xE800AA87E800A887, 0xE800B387E800AD87,
-    0xE8008188E800BC87, 0xE8008C88E8008488, 0xE8009B88E8009888,
-    0xE800AE89E8009F88, 0xE800B289E800AF89, 0xE800B989E800B889,
-    0xE800918AE8008B8A, 0xE800B18AE8009D8A, 0xE800BD8AE800B38A,
-    0xE800A68BE800A58B, 0xE800A38CE8009D8C, 0xE800928DE800B68C,
-    0xE800A38DE800938D, 0xE800BD8EE800AD8E, 0xE8008A8FE800898F,
-    0xE8009C8FE8008C8F, 0xE800AF8FE800A78F, 0xE800BD90E800B18F,
-    0xE8009791E8008991, 0xE800B193E800AE93, 0xE800BC93E800B393,
-    0xE800A495E8009694, 0xE800BA97E8008D97, 0xE8009298E8008698,
-    0xE800BF98E800AD98, 0xE8009099E8008D99, 0xE800A799E8009C99,
-    0xE800AB99E800A999, 0xE800A99AE800889A, 0xE8008E9CE800A29B,
-    0xE800AB9DE800A89C, 0xE800869EE800B99D, 0xE800A19FE800BA9E,
-    0xE8009FA0E80081A0, 0xE8008CA1E80080A1, 0xE800A3A1E800A0A1,
-    0xE8008FA3E80082A3, 0xE8009EA3E80097A3, 0xE800B8A3E800A1A3,
-    0xE80090A4E800BAA3, 0xE800A4A5E80081A5, 0xE80086A6E800BEA5,
-    0xE80096A6E8008BA6, 0xE800A3A7E80092A7, 0xE800A0AAE80080A8,
-    0xE800BFAAE800AAAA, 0xE80092ABE8008BAB, 0xE800ADABE80096AB,
-    0xE800BEABE800B8AB, 0xE800B9ACE80081AC, 0xE80080AEE80098AD,
-    0xE800B7B0E8008AAE, 0xE80088B1E80086B1, 0xE800B8B1E80095B1,
-    0xE800A1B2E8009DB2, 0xE800ABB2E800A9B2, 0xE80082B3E80081B3,
-    0xE80088B3E80087B3, 0xE80088B4E80093B3, 0xE800A4B5E8009BB4,
-    0xE800B7B5E800B0B5, 0xE800BCB6E800B3B6, 0xE800AFB7E8008BB7,
-    0xE800ABBAE800B0B7, 0xE80094BBE8008ABB, 0xE800AABCE800A6BC,
-    0xE800BBBCE800B8BC, 0xE8009BBEE800A2BD, 0xE800B0BEE8009EBE,
-    0xE900B6BEE800B5BE, 0xE900B880E900A380, 0xE900A981E9008A81,
-    0xE900BC81E900B281, 0xE9009182E9008F82, 0xE9008E83E9009482,
-    0xE900B183E9009E83, 0xE9009184E900BD83, 0xE9008985E9009B84,
-    0xE900AA85E9008D85, 0xE900B486E9009986, 0xE9008C87E9008687,
-    0xE9009187E9008F87, 0xE900B888E900B488, 0xE900BC89E900B689,
-    0xE900988BE900978B, 0xE9008A8DE900848C, 0xE9009590E900B98F,
-    0xE9008096E900B795, 0xE900AD96E9008B96, 0xE9009C98E900B796,
-    0xE9008B99E900AE98, 0xE900B599E9008D99, 0xE900BC99E900B899,
-    0xE900A39AE900869A, 0xE900B79AE900B69A, 0xE900B99AE900B89A,
-    0xE900A29BE900839B, 0xE900A89BE900A39B, 0xE900B79BE900B69B,
-    0xE900B29CE900A39C, 0xE900919DE900889D, 0xE9009E9DE900969D,
-    0xE900A99DE900A29D, 0xE9009B9FE9008B9F, 0xE900AD9FE900A09F,
-    0xE900BF9FE900B39F, 0xE90085A0E90081A0, 0xE90098A0E9008BA0,
-    0xE900BBA0E900A9A0, 0xE900A8A2E9009EA1, 0xE9009FA3E9009BA3,
-    0xE900AFA3E900A2A3, 0xE900A8A4E900BCA3, 0xE90096A6E900A9A4,
-    0xE900A7A6E90099A6, 0xE90082A7E900ACA6, 0xE900BEA7E900B1A7,
-    0xE900A8AAE900AAA9, 0xE9009FABE90098AB, 0xE900A5ACE90092AC,
-    0xE900B2ACE900AFAC, 0xE9009AADE900BCAC, 0xE90080B1E900AFAD,
-    0xE900A5B3E90097B1, 0xE900A7B5E900BDB3, 0xE900BAB7E900B4B6,
-    0xE900B5B9E9009EB8, 0xE90097BAE900BFB9, 0xE900A5BAE9009FBA,
-    0xE90083BBE900BBBA, 0xE9008EBBE9008DBB, 0xE900B9BBE90091BB,
-    0xE900BEBBE900BDBB, 0xE9008EBCE90085BC, 0xE90093BCE9008FBC,
-    0xE900A0BCE90096BC, 0xE90083BDE900BBBC, 0xE90092BDE9008ABD,
-    0xE9008EBEE9008DBE, 0xE9009FBEE9009CBE, 0xEA009199EA00A0BE,
-    0xEA00A79CEA00899A, 0xEA008E9EEA00AF9D, 0xEA0092ADEA00B7AC,
-    0xEA00A7ADEA00A6AD, 0xEA0081B0EA0080B0, 0xEA0083B0EA0082B0,
-    0xEA0085B0EA0084B0, 0xEA0087B0EA0086B0, 0xEA0089B0EA0088B0,
-    0xEA008BB0EA008AB0, 0xEA008DB0EA008CB0, 0xEA008FB0EA008EB0,
-    0xEA0091B0EA0090B0, 0xEA0093B0EA0092B0, 0xEA0095B0EA0094B0,
-    0xEA0097B0EA0096B0, 0xEA0099B0EA0098B0, 0xEA009BB0EA009AB0,
-    0xEA009DB0EA009CB0, 0xEA009FB0EA009EB0, 0xEA00A1B0EA00A0B0,
-    0xEA00A3B0EA00A2B0, 0xEA00A5B0EA00A4B0, 0xEA00A7B0EA00A6B0,
-    0xEA00A9B0EA00A8B0, 0xEA00ABB0EA00AAB0, 0xEA00ADB0EA00ACB0,
-    0xEA00AFB0EA00AEB0, 0xEA00B1B0EA00B0B0, 0xEA00B3B0EA00B2B0,
-    0xEA00B5B0EA00B4B0, 0xEA00B7B0EA00B6B0, 0xEA00B9B0EA00B8B0,
-    0xEA00BBB0EA00BAB0, 0xEA00BDB0EA00BCB0, 0xEA00BFB0EA00BEB0,
-    0xEA0081B1EA0080B1, 0xEA0083B1EA0082B1, 0xEA0085B1EA0084B1,
-    0xEA0087B1EA0086B1, 0xEA0089B1EA0088B1, 0xEA008BB1EA008AB1,
-    0xEA008DB1EA008CB1, 0xEA008FB1EA008EB1, 0xEA0091B1EA0090B1,
-    0xEA0093B1EA0092B1, 0xEA0095B1EA0094B1, 0xEA0097B1EA0096B1,
-    0xEA0099B1EA0098B1, 0xEA009BB1EA009AB1, 0xEA009DB1EA009CB1,
-    0xEA009FB1EA009EB1, 0xEA00A1B1EA00A0B1, 0xEA00A3B1EA00A2B1,
-    0xEA00A5B1EA00A4B1, 0xEA00A7B1EA00A6B1, 0xEA00A9B1EA00A8B1,
-    0xEA00ABB1EA00AAB1, 0xEA00ADB1EA00ACB1, 0xEA00AFB1EA00AEB1,
-    0xEA00B1B1EA00B0B1, 0xEA00B3B1EA00B2B1, 0xEA00B5B1EA00B4B1,
-    0xEA00B7B1EA00B6B1, 0xEA00B9B1EA00B8B1, 0xEA00BBB1EA00BAB1,
-    0xEA00BDB1EA00BCB1, 0xEA00BFB1EA00BEB1, 0xEA0081B2EA0080B2,
-    0xEA0083B2EA0082B2, 0xEA0085B2EA0084B2, 0xEA0087B2EA0086B2,
-    0xEA0089B2EA0088B2, 0xEA008BB2EA008AB2, 0xEA008DB2EA008CB2,
-    0xEA008FB2EA008EB2, 0xEA0091B2EA0090B2, 0xEA0093B2EA0092B2,
-    0xEA0095B2EA0094B2, 0xEA0097B2EA0096B2, 0xEA0099B2EA0098B2,
-    0xEA009BB2EA009AB2, 0xEA009DB2EA009CB2, 0xEA009FB2EA009EB2,
-    0xEA00A1B2EA00A0B2, 0xEA00A3B2EA00A2B2, 0xEA00A5B2EA00A4B2,
-    0xEA00A7B2EA00A6B2, 0xEA00A9B2EA00A8B2, 0xEA00ABB2EA00AAB2,
-    0xEA00ADB2EA00ACB2, 0xEA00AFB2EA00AEB2, 0xEA00B1B2EA00B0B2,
-    0xEA00B3B2EA00B2B2, 0xEA00B5B2EA00B4B2, 0xEA00B7B2EA00B6B2,
-    0xEA00B9B2EA00B8B2, 0xEA00BBB2EA00BAB2, 0xEA00BDB2EA00BCB2,
-    0xEA00BFB2EA00BEB2, 0xEA0081B3EA0080B3, 0xEA0083B3EA0082B3,
-    0xEA0085B3EA0084B3, 0xEA0087B3EA0086B3, 0xEA0089B3EA0088B3,
-    0xEA008BB3EA008AB3, 0xEA008DB3EA008CB3, 0xEA008FB3EA008EB3,
-    0xEA0091B3EA0090B3, 0xEA0093B3EA0092B3, 0xEA0095B3EA0094B3,
-    0xEA0097B3EA0096B3, 0xEA0099B3EA0098B3, 0xEA009BB3EA009AB3,
-    0xEA009DB3EA009CB3, 0xEA009FB3EA009EB3, 0xEA00A1B3EA00A0B3,
-    0xEA00A3B3EA00A2B3, 0xEA00A5B3EA00A4B3, 0xEA00A7B3EA00A6B3,
-    0xEA00A9B3EA00A8B3, 0xEA00ABB3EA00AAB3, 0xEA00ADB3EA00ACB3,
-    0xEA00AFB3EA00AEB3, 0xEA00B1B3EA00B0B3, 0xEA00B3B3EA00B2B3,
-    0xEA00B5B3EA00B4B3, 0xEA00B7B3EA00B6B3, 0xEA00B9B3EA00B8B3,
-    0xEA00BBB3EA00BAB3, 0xEA00BDB3EA00BCB3, 0xEA00BFB3EA00BEB3,
-    0xEA0081B4EA0080B4, 0xEA0083B4EA0082B4, 0xEA0085B4EA0084B4,
-    0xEA0087B4EA0086B4, 0xEA0089B4EA0088B4, 0xEA008BB4EA008AB4,
-    0xEA008DB4EA008CB4, 0xEA008FB4EA008EB4, 0xEA0091B4EA0090B4,
-    0xEA0093B4EA0092B4, 0xEA0095B4EA0094B4, 0xEA0097B4EA0096B4,
-    0xEA0099B4EA0098B4, 0xEA009BB4EA009AB4, 0xEA009DB4EA009CB4,
-    0xEA009FB4EA009EB4, 0xEA00A1B4EA00A0B4, 0xEA00A3B4EA00A2B4,
-    0xEA00A5B4EA00A4B4, 0xEA00A7B4EA00A6B4, 0xEA00A9B4EA00A8B4,
-    0xEA00ABB4EA00AAB4, 0xEA00ADB4EA00ACB4, 0xEA00AFB4EA00AEB4,
-    0xEA00B1B4EA00B0B4, 0xEA00B3B4EA00B2B4, 0xEA00B5B4EA00B4B4,
-    0xEA00B7B4EA00B6B4, 0xEA00B9B4EA00B8B4, 0xEA00BBB4EA00BAB4,
-    0xEA00BDB4EA00BCB4, 0xEA00BFB4EA00BEB4, 0xEA0081B5EA0080B5,
-    0xEA0083B5EA0082B5, 0xEA0085B5EA0084B5, 0xEA0087B5EA0086B5,
-    0xEA0089B5EA0088B5, 0xEA008BB5EA008AB5, 0xEA008DB5EA008CB5,
-    0xEA008FB5EA008EB5, 0xEA0091B5EA0090B5, 0xEA0093B5EA0092B5,
-    0xEA0095B5EA0094B5, 0xEA0097B5EA0096B5, 0xEA0099B5EA0098B5,
-    0xEA009BB5EA009AB5, 0xEA009DB5EA009CB5, 0xEA009FB5EA009EB5,
-    0xEA00A1B5EA00A0B5, 0xEA00A3B5EA00A2B5, 0xEA00A5B5EA00A4B5,
-    0xEA00A7B5EA00A6B5, 0xEA00A9B5EA00A8B5, 0xEA00ABB5EA00AAB5,
-    0xEA00ADB5EA00ACB5, 0xEA00AFB5EA00AEB5, 0xEA00B1B5EA00B0B5,
-    0xEA00B3B5EA00B2B5, 0xEA00B5B5EA00B4B5, 0xEA00B7B5EA00B6B5,
-    0xEA00B9B5EA00B8B5, 0xEA00BBB5EA00BAB5, 0xEA00BDB5EA00BCB5,
-    0xEA00BFB5EA00BEB5, 0xEA0081B6EA0080B6, 0xEA0083B6EA0082B6,
-    0xEA0085B6EA0084B6, 0xEA0087B6EA0086B6, 0xEA0089B6EA0088B6,
-    0xEA008BB6EA008AB6, 0xEA008DB6EA008CB6, 0xEA008FB6EA008EB6,
-    0xEA0091B6EA0090B6, 0xEA0093B6EA0092B6, 0xEA0095B6EA0094B6,
-    0xEA0097B6EA0096B6, 0xEA0099B6EA0098B6, 0xEA009BB6EA009AB6,
-    0xEA009DB6EA009CB6, 0xEA009FB6EA009EB6, 0xEA00A1B6EA00A0B6,
-    0xEA00A3B6EA00A2B6, 0xEA00A5B6EA00A4B6, 0xEA00A7B6EA00A6B6,
-    0xEA00A9B6EA00A8B6, 0xEA00ABB6EA00AAB6, 0xEA00ADB6EA00ACB6,
-    0xEA00AFB6EA00AEB6, 0xEA00B1B6EA00B0B6, 0xEA00B3B6EA00B2B6,
-    0xEA00B5B6EA00B4B6, 0xEA00B7B6EA00B6B6, 0xEA00B9B6EA00B8B6,
-    0xEA00BBB6EA00BAB6, 0xEA00BDB6EA00BCB6, 0xEA00BFB6EA00BEB6,
-    0xEA0081B7EA0080B7, 0xEA0083B7EA0082B7, 0xEA0085B7EA0084B7,
-    0xEA0087B7EA0086B7, 0xEA0089B7EA0088B7, 0xEA008BB7EA008AB7,
-    0xEA008DB7EA008CB7, 0xEA008FB7EA008EB7, 0xEA0091B7EA0090B7,
-    0xEA0093B7EA0092B7, 0xEA0095B7EA0094B7, 0xEA0097B7EA0096B7,
-    0xEA0099B7EA0098B7, 0xEA009BB7EA009AB7, 0xEA009DB7EA009CB7,
-    0xEA009FB7EA009EB7, 0xEA00A1B7EA00A0B7, 0xEA00A3B7EA00A2B7,
-    0xEA00A5B7EA00A4B7, 0xEA00A7B7EA00A6B7, 0xEA00A9B7EA00A8B7,
-    0xEA00ABB7EA00AAB7, 0xEA00ADB7EA00ACB7, 0xEA00AFB7EA00AEB7,
-    0xEA00B1B7EA00B0B7, 0xEA00B3B7EA00B2B7, 0xEA00B5B7EA00B4B7,
-    0xEA00B7B7EA00B6B7, 0xEA00B9B7EA00B8B7, 0xEA00BBB7EA00BAB7,
-    0xEA00BDB7EA00BCB7, 0xEA00BFB7EA00BEB7, 0xEA0081B8EA0080B8,
-    0xEA0083B8EA0082B8, 0xEA0085B8EA0084B8, 0xEA0087B8EA0086B8,
-    0xEA0089B8EA0088B8, 0xEA008BB8EA008AB8, 0xEA008DB8EA008CB8,
-    0xEA008FB8EA008EB8, 0xEA0091B8EA0090B8, 0xEA0093B8EA0092B8,
-    0xEA0095B8EA0094B8, 0xEA0097B8EA0096B8, 0xEA0099B8EA0098B8,
-    0xEA009BB8EA009AB8, 0xEA009DB8EA009CB8, 0xEA009FB8EA009EB8,
-    0xEA00A1B8EA00A0B8, 0xEA00A3B8EA00A2B8, 0xEA00A5B8EA00A4B8,
-    0xEA00A7B8EA00A6B8, 0xEA00A9B8EA00A8B8, 0xEA00ABB8EA00AAB8,
-    0xEA00ADB8EA00ACB8, 0xEA00AFB8EA00AEB8, 0xEA00B1B8EA00B0B8,
-    0xEA00B3B8EA00B2B8, 0xEA00B5B8EA00B4B8, 0xEA00B7B8EA00B6B8,
-    0xEA00B9B8EA00B8B8, 0xEA00BBB8EA00BAB8, 0xEA00BDB8EA00BCB8,
-    0xEA00BFB8EA00BEB8, 0xEA0081B9EA0080B9, 0xEA0083B9EA0082B9,
-    0xEA0085B9EA0084B9, 0xEA0087B9EA0086B9, 0xEA0089B9EA0088B9,
-    0xEA008BB9EA008AB9, 0xEA008DB9EA008CB9, 0xEA008FB9EA008EB9,
-    0xEA0091B9EA0090B9, 0xEA0093B9EA0092B9, 0xEA0095B9EA0094B9,
-    0xEA0097B9EA0096B9, 0xEA0099B9EA0098B9, 0xEA009BB9EA009AB9,
-    0xEA009DB9EA009CB9, 0xEA009FB9EA009EB9, 0xEA00A1B9EA00A0B9,
-    0xEA00A3B9EA00A2B9, 0xEA00A5B9EA00A4B9, 0xEA00A7B9EA00A6B9,
-    0xEA00A9B9EA00A8B9, 0xEA00ABB9EA00AAB9, 0xEA00ADB9EA00ACB9,
-    0xEA00AFB9EA00AEB9, 0xEA00B1B9EA00B0B9, 0xEA00B3B9EA00B2B9,
-    0xEA00B5B9EA00B4B9, 0xEA00B7B9EA00B6B9, 0xEA00B9B9EA00B8B9,
-    0xEA00BBB9EA00BAB9, 0xEA00BDB9EA00BCB9, 0xEA00BFB9EA00BEB9,
-    0xEA0081BAEA0080BA, 0xEA0083BAEA0082BA, 0xEA0085BAEA0084BA,
-    0xEA0087BAEA0086BA, 0xEA0089BAEA0088BA, 0xEA008BBAEA008ABA,
-    0xEA008DBAEA008CBA, 0xEA008FBAEA008EBA, 0xEA0091BAEA0090BA,
-    0xEA0093BAEA0092BA, 0xEA0095BAEA0094BA, 0xEA0097BAEA0096BA,
-    0xEA0099BAEA0098BA, 0xEA009BBAEA009ABA, 0xEA009DBAEA009CBA,
-    0xEA009FBAEA009EBA, 0xEA00A1BAEA00A0BA, 0xEA00A3BAEA00A2BA,
-    0xEA00A5BAEA00A4BA, 0xEA00A7BAEA00A6BA, 0xEA00A9BAEA00A8BA,
-    0xEA00ABBAEA00AABA, 0xEA00ADBAEA00ACBA, 0xEA00AFBAEA00AEBA,
-    0xEA00B1BAEA00B0BA, 0xEA00B3BAEA00B2BA, 0xEA00B5BAEA00B4BA,
-    0xEA00B7BAEA00B6BA, 0xEA00B9BAEA00B8BA, 0xEA00BBBAEA00BABA,
-    0xEA00BDBAEA00BCBA, 0xEA00BFBAEA00BEBA, 0xEA0081BBEA0080BB,
-    0xEA0083BBEA0082BB, 0xEA0085BBEA0084BB, 0xEA0087BBEA0086BB,
-    0xEA0089BBEA0088BB, 0xEA008BBBEA008ABB, 0xEA008DBBEA008CBB,
-    0xEA008FBBEA008EBB, 0xEA0091BBEA0090BB, 0xEA0093BBEA0092BB,
-    0xEA0095BBEA0094BB, 0xEA0097BBEA0096BB, 0xEA0099BBEA0098BB,
-    0xEA009BBBEA009ABB, 0xEA009DBBEA009CBB, 0xEA009FBBEA009EBB,
-    0xEA00A1BBEA00A0BB, 0xEA00A3BBEA00A2BB, 0xEA00A5BBEA00A4BB,
-    0xEA00A7BBEA00A6BB, 0xEA00A9BBEA00A8BB, 0xEA00ABBBEA00AABB,
-    0xEA00ADBBEA00ACBB, 0xEA00AFBBEA00AEBB, 0xEA00B1BBEA00B0BB,
-    0xEA00B3BBEA00B2BB, 0xEA00B5BBEA00B4BB, 0xEA00B7BBEA00B6BB,
-    0xEA00B9BBEA00B8BB, 0xEA00BBBBEA00BABB, 0xEA00BDBBEA00BCBB,
-    0xEA00BFBBEA00BEBB, 0xEA0081BCEA0080BC, 0xEA0083BCEA0082BC,
-    0xEA0085BCEA0084BC, 0xEA0087BCEA0086BC, 0xEA0089BCEA0088BC,
-    0xEA008BBCEA008ABC, 0xEA008DBCEA008CBC, 0xEA008FBCEA008EBC,
-    0xEA0091BCEA0090BC, 0xEA0093BCEA0092BC, 0xEA0095BCEA0094BC,
-    0xEA0097BCEA0096BC, 0xEA0099BCEA0098BC, 0xEA009BBCEA009ABC,
-    0xEA009DBCEA009CBC, 0xEA009FBCEA009EBC, 0xEA00A1BCEA00A0BC,
-    0xEA00A3BCEA00A2BC, 0xEA00A5BCEA00A4BC, 0xEA00A7BCEA00A6BC,
-    0xEA00A9BCEA00A8BC, 0xEA00ABBCEA00AABC, 0xEA00ADBCEA00ACBC,
-    0xEA00AFBCEA00AEBC, 0xEA00B1BCEA00B0BC, 0xEA00B3BCEA00B2BC,
-    0xEA00B5BCEA00B4BC, 0xEA00B7BCEA00B6BC, 0xEA00B9BCEA00B8BC,
-    0xEA00BBBCEA00BABC, 0xEA00BDBCEA00BCBC, 0xEA00BFBCEA00BEBC,
-    0xEA0081BDEA0080BD, 0xEA0083BDEA0082BD, 0xEA0085BDEA0084BD,
-    0xEA0087BDEA0086BD, 0xEA0089BDEA0088BD, 0xEA008BBDEA008ABD,
-    0xEA008DBDEA008CBD, 0xEA008FBDEA008EBD, 0xEA0091BDEA0090BD,
-    0xEA0093BDEA0092BD, 0xEA0095BDEA0094BD, 0xEA0097BDEA0096BD,
-    0xEA0099BDEA0098BD, 0xEA009BBDEA009ABD, 0xEA009DBDEA009CBD,
-    0xEA009FBDEA009EBD, 0xEA00A1BDEA00A0BD, 0xEA00A3BDEA00A2BD,
-    0xEA00A5BDEA00A4BD, 0xEA00A7BDEA00A6BD, 0xEA00A9BDEA00A8BD,
-    0xEA00ABBDEA00AABD, 0xEA00ADBDEA00ACBD, 0xEA00AFBDEA00AEBD,
-    0xEA00B1BDEA00B0BD, 0xEA00B3BDEA00B2BD, 0xEA00B5BDEA00B4BD,
-    0xEA00B7BDEA00B6BD, 0xEA00B9BDEA00B8BD, 0xEA00BBBDEA00BABD,
-    0xEA00BDBDEA00BCBD, 0xEA00BFBDEA00BEBD, 0xEA0081BEEA0080BE,
-    0xEA0083BEEA0082BE, 0xEA0085BEEA0084BE, 0xEA0087BEEA0086BE,
-    0xEA0089BEEA0088BE, 0xEA008BBEEA008ABE, 0xEA008DBEEA008CBE,
-    0xEA008FBEEA008EBE, 0xEA0091BEEA0090BE, 0xEA0093BEEA0092BE,
-    0xEA0095BEEA0094BE, 0xEA0097BEEA0096BE, 0xEA0099BEEA0098BE,
-    0xEA009BBEEA009ABE, 0xEA009DBEEA009CBE, 0xEA009FBEEA009EBE,
-    0xEA00A1BEEA00A0BE, 0xEA00A3BEEA00A2BE, 0xEA00A5BEEA00A4BE,
-    0xEA00A7BEEA00A6BE, 0xEA00A9BEEA00A8BE, 0xEA00ABBEEA00AABE,
-    0xEA00ADBEEA00ACBE, 0xEA00AFBEEA00AEBE, 0xEA00B1BEEA00B0BE,
-    0xEA00B3BEEA00B2BE, 0xEA00B5BEEA00B4BE, 0xEA00B7BEEA00B6BE,
-    0xEA00B9BEEA00B8BE, 0xEA00BBBEEA00BABE, 0xEA00BDBEEA00BCBE,
-    0xEA00BFBEEA00BEBE, 0xEA0081BFEA0080BF, 0xEA0083BFEA0082BF,
-    0xEA0085BFEA0084BF, 0xEA0087BFEA0086BF, 0xEA0089BFEA0088BF,
-    0xEA008BBFEA008ABF, 0xEA008DBFEA008CBF, 0xEA008FBFEA008EBF,
-    0xEA0091BFEA0090BF, 0xEA0093BFEA0092BF, 0xEA0095BFEA0094BF,
-    0xEA0097BFEA0096BF, 0xEA0099BFEA0098BF, 0xEA009BBFEA009ABF,
-    0xEA009DBFEA009CBF, 0xEA009FBFEA009EBF, 0xEA00A1BFEA00A0BF,
-    0xEA00A3BFEA00A2BF, 0xEA00A5BFEA00A4BF, 0xEA00A7BFEA00A6BF,
-    0xEA00A9BFEA00A8BF, 0xEA00ABBFEA00AABF, 0xEA00ADBFEA00ACBF,
-    0xEA00AFBFEA00AEBF, 0xEA00B1BFEA00B0BF, 0xEA00B3BFEA00B2BF,
-    0xEA00B5BFEA00B4BF, 0xEA00B7BFEA00B6BF, 0xEA00B9BFEA00B8BF,
-    0xEA00BBBFEA00BABF, 0xEA00BDBFEA00BCBF, 0xEB00BFBFEA00BEBF,
-    0xEB008180EB008080, 0xEB008380EB008280, 0xEB008580EB008480,
-    0xEB008780EB008680, 0xEB008980EB008880, 0xEB008B80EB008A80,
-    0xEB008D80EB008C80, 0xEB008F80EB008E80, 0xEB009180EB009080,
-    0xEB009380EB009280, 0xEB009580EB009480, 0xEB009780EB009680,
-    0xEB009980EB009880, 0xEB009B80EB009A80, 0xEB009D80EB009C80,
-    0xEB009F80EB009E80, 0xEB00A180EB00A080, 0xEB00A380EB00A280,
-    0xEB00A580EB00A480, 0xEB00A780EB00A680, 0xEB00A980EB00A880,
-    0xEB00AB80EB00AA80, 0xEB00AD80EB00AC80, 0xEB00AF80EB00AE80,
-    0xEB00B180EB00B080, 0xEB00B380EB00B280, 0xEB00B580EB00B480,
-    0xEB00B780EB00B680, 0xEB00B980EB00B880, 0xEB00BB80EB00BA80,
-    0xEB00BD80EB00BC80, 0xEB00BF80EB00BE80, 0xEB008181EB008081,
-    0xEB008381EB008281, 0xEB008581EB008481, 0xEB008781EB008681,
-    0xEB008981EB008881, 0xEB008B81EB008A81, 0xEB008D81EB008C81,
-    0xEB008F81EB008E81, 0xEB009181EB009081, 0xEB009381EB009281,
-    0xEB009581EB009481, 0xEB009781EB009681, 0xEB009981EB009881,
-    0xEB009B81EB009A81, 0xEB009D81EB009C81, 0xEB009F81EB009E81,
-    0xEB00A181EB00A081, 0xEB00A381EB00A281, 0xEB00A581EB00A481,
-    0xEB00A781EB00A681, 0xEB00A981EB00A881, 0xEB00AB81EB00AA81,
-    0xEB00AD81EB00AC81, 0xEB00AF81EB00AE81, 0xEB00B181EB00B081,
-    0xEB00B381EB00B281, 0xEB00B581EB00B481, 0xEB00B781EB00B681,
-    0xEB00B981EB00B881, 0xEB00BB81EB00BA81, 0xEB00BD81EB00BC81,
-    0xEB00BF81EB00BE81, 0xEB008182EB008082, 0xEB008382EB008282,
-    0xEB008582EB008482, 0xEB008782EB008682, 0xEB008982EB008882,
-    0xEB008B82EB008A82, 0xEB008D82EB008C82, 0xEB008F82EB008E82,
-    0xEB009182EB009082, 0xEB009382EB009282, 0xEB009582EB009482,
-    0xEB009782EB009682, 0xEB009982EB009882, 0xEB009B82EB009A82,
-    0xEB009D82EB009C82, 0xEB009F82EB009E82, 0xEB00A182EB00A082,
-    0xEB00A382EB00A282, 0xEB00A582EB00A482, 0xEB00A782EB00A682,
-    0xEB00A982EB00A882, 0xEB00AB82EB00AA82, 0xEB00AD82EB00AC82,
-    0xEB00AF82EB00AE82, 0xEB00B182EB00B082, 0xEB00B382EB00B282,
-    0xEB00B582EB00B482, 0xEB00B782EB00B682, 0xEB00B982EB00B882,
-    0xEB00BB82EB00BA82, 0xEB00BD82EB00BC82, 0xEB00BF82EB00BE82,
-    0xEB008183EB008083, 0xEB008383EB008283, 0xEB008583EB008483,
-    0xEB008783EB008683, 0xEB008983EB008883, 0xEB008B83EB008A83,
-    0xEB008D83EB008C83, 0xEB008F83EB008E83, 0xEB009183EB009083,
-    0xEB009383EB009283, 0xEB009583EB009483, 0xEB009783EB009683,
-    0xEB009983EB009883, 0xEB009B83EB009A83, 0xEB009D83EB009C83,
-    0xEB009F83EB009E83, 0xEB00A183EB00A083, 0xEB00A383EB00A283,
-    0xEB00A583EB00A483, 0xEB00A783EB00A683, 0xEB00A983EB00A883,
-    0xEB00AB83EB00AA83, 0xEB00AD83EB00AC83, 0xEB00AF83EB00AE83,
-    0xEB00B183EB00B083, 0xEB00B383EB00B283, 0xEB00B583EB00B483,
-    0xEB00B783EB00B683, 0xEB00B983EB00B883, 0xEB00BB83EB00BA83,
-    0xEB00BD83EB00BC83, 0xEB00BF83EB00BE83, 0xEB008184EB008084,
-    0xEB008384EB008284, 0xEB008584EB008484, 0xEB008784EB008684,
-    0xEB008984EB008884, 0xEB008B84EB008A84, 0xEB008D84EB008C84,
-    0xEB008F84EB008E84, 0xEB009184EB009084, 0xEB009384EB009284,
-    0xEB009584EB009484, 0xEB009784EB009684, 0xEB009984EB009884,
-    0xEB009B84EB009A84, 0xEB009D84EB009C84, 0xEB009F84EB009E84,
-    0xEB00A184EB00A084, 0xEB00A384EB00A284, 0xEB00A584EB00A484,
-    0xEB00A784EB00A684, 0xEB00A984EB00A884, 0xEB00AB84EB00AA84,
-    0xEB00AD84EB00AC84, 0xEB00AF84EB00AE84, 0xEB00B184EB00B084,
-    0xEB00B384EB00B284, 0xEB00B584EB00B484, 0xEB00B784EB00B684,
-    0xEB00B984EB00B884, 0xEB00BB84EB00BA84, 0xEB00BD84EB00BC84,
-    0xEB00BF84EB00BE84, 0xEB008185EB008085, 0xEB008385EB008285,
-    0xEB008585EB008485, 0xEB008785EB008685, 0xEB008985EB008885,
-    0xEB008B85EB008A85, 0xEB008D85EB008C85, 0xEB008F85EB008E85,
-    0xEB009185EB009085, 0xEB009385EB009285, 0xEB009585EB009485,
-    0xEB009785EB009685, 0xEB009985EB009885, 0xEB009B85EB009A85,
-    0xEB009D85EB009C85, 0xEB009F85EB009E85, 0xEB00A185EB00A085,
-    0xEB00A385EB00A285, 0xEB00A585EB00A485, 0xEB00A785EB00A685,
-    0xEB00A985EB00A885, 0xEB00AB85EB00AA85, 0xEB00AD85EB00AC85,
-    0xEB00AF85EB00AE85, 0xEB00B185EB00B085, 0xEB00B385EB00B285,
-    0xEB00B585EB00B485, 0xEB00B785EB00B685, 0xEB00B985EB00B885,
-    0xEB00BB85EB00BA85, 0xEB00BD85EB00BC85, 0xEB00BF85EB00BE85,
-    0xEB008186EB008086, 0xEB008386EB008286, 0xEB008586EB008486,
-    0xEB008786EB008686, 0xEB008986EB008886, 0xEB008B86EB008A86,
-    0xEB008D86EB008C86, 0xEB008F86EB008E86, 0xEB009186EB009086,
-    0xEB009386EB009286, 0xEB009586EB009486, 0xEB009786EB009686,
-    0xEB009986EB009886, 0xEB009B86EB009A86, 0xEB009D86EB009C86,
-    0xEB009F86EB009E86, 0xEB00A186EB00A086, 0xEB00A386EB00A286,
-    0xEB00A586EB00A486, 0xEB00A786EB00A686, 0xEB00A986EB00A886,
-    0xEB00AB86EB00AA86, 0xEB00AD86EB00AC86, 0xEB00AF86EB00AE86,
-    0xEB00B186EB00B086, 0xEB00B386EB00B286, 0xEB00B586EB00B486,
-    0xEB00B786EB00B686, 0xEB00B986EB00B886, 0xEB00BB86EB00BA86,
-    0xEB00BD86EB00BC86, 0xEB00BF86EB00BE86, 0xEB008187EB008087,
-    0xEB008387EB008287, 0xEB008587EB008487, 0xEB008787EB008687,
-    0xEB008987EB008887, 0xEB008B87EB008A87, 0xEB008D87EB008C87,
-    0xEB008F87EB008E87, 0xEB009187EB009087, 0xEB009387EB009287,
-    0xEB009587EB009487, 0xEB009787EB009687, 0xEB009987EB009887,
-    0xEB009B87EB009A87, 0xEB009D87EB009C87, 0xEB009F87EB009E87,
-    0xEB00A187EB00A087, 0xEB00A387EB00A287, 0xEB00A587EB00A487,
-    0xEB00A787EB00A687, 0xEB00A987EB00A887, 0xEB00AB87EB00AA87,
-    0xEB00AD87EB00AC87, 0xEB00AF87EB00AE87, 0xEB00B187EB00B087,
-    0xEB00B387EB00B287, 0xEB00B587EB00B487, 0xEB00B787EB00B687,
-    0xEB00B987EB00B887, 0xEB00BB87EB00BA87, 0xEB00BD87EB00BC87,
-    0xEB00BF87EB00BE87, 0xEB008188EB008088, 0xEB008388EB008288,
-    0xEB008588EB008488, 0xEB008788EB008688, 0xEB008988EB008888,
-    0xEB008B88EB008A88, 0xEB008D88EB008C88, 0xEB008F88EB008E88,
-    0xEB009188EB009088, 0xEB009388EB009288, 0xEB009588EB009488,
-    0xEB009788EB009688, 0xEB009988EB009888, 0xEB009B88EB009A88,
-    0xEB009D88EB009C88, 0xEB009F88EB009E88, 0xEB00A188EB00A088,
-    0xEB00A388EB00A288, 0xEB00A588EB00A488, 0xEB00A788EB00A688,
-    0xEB00A988EB00A888, 0xEB00AB88EB00AA88, 0xEB00AD88EB00AC88,
-    0xEB00AF88EB00AE88, 0xEB00B188EB00B088, 0xEB00B388EB00B288,
-    0xEB00B588EB00B488, 0xEB00B788EB00B688, 0xEB00B988EB00B888,
-    0xEB00BB88EB00BA88, 0xEB00BD88EB00BC88, 0xEB00BF88EB00BE88,
-    0xEB008189EB008089, 0xEB008389EB008289, 0xEB008589EB008489,
-    0xEB008789EB008689, 0xEB008989EB008889, 0xEB008B89EB008A89,
-    0xEB008D89EB008C89, 0xEB008F89EB008E89, 0xEB009189EB009089,
-    0xEB009389EB009289, 0xEB009589EB009489, 0xEB009789EB009689,
-    0xEB009989EB009889, 0xEB009B89EB009A89, 0xEB009D89EB009C89,
-    0xEB009F89EB009E89, 0xEB00A189EB00A089, 0xEB00A389EB00A289,
-    0xEB00A589EB00A489, 0xEB00A789EB00A689, 0xEB00A989EB00A889,
-    0xEB00AB89EB00AA89, 0xEB00AD89EB00AC89, 0xEB00AF89EB00AE89,
-    0xEB00B189EB00B089, 0xEB00B389EB00B289, 0xEB00B589EB00B489,
-    0xEB00B789EB00B689, 0xEB00B989EB00B889, 0xEB00BB89EB00BA89,
-    0xEB00BD89EB00BC89, 0xEB00BF89EB00BE89, 0xEB00818AEB00808A,
-    0xEB00838AEB00828A, 0xEB00858AEB00848A, 0xEB00878AEB00868A,
-    0xEB00898AEB00888A, 0xEB008B8AEB008A8A, 0xEB008D8AEB008C8A,
-    0xEB008F8AEB008E8A, 0xEB00918AEB00908A, 0xEB00938AEB00928A,
-    0xEB00958AEB00948A, 0xEB00978AEB00968A, 0xEB00998AEB00988A,
-    0xEB009B8AEB009A8A, 0xEB009D8AEB009C8A, 0xEB009F8AEB009E8A,
-    0xEB00A18AEB00A08A, 0xEB00A38AEB00A28A, 0xEB00A58AEB00A48A,
-    0xEB00A78AEB00A68A, 0xEB00A98AEB00A88A, 0xEB00AB8AEB00AA8A,
-    0xEB00AD8AEB00AC8A, 0xEB00AF8AEB00AE8A, 0xEB00B18AEB00B08A,
-    0xEB00B38AEB00B28A, 0xEB00B58AEB00B48A, 0xEB00B78AEB00B68A,
-    0xEB00B98AEB00B88A, 0xEB00BB8AEB00BA8A, 0xEB00BD8AEB00BC8A,
-    0xEB00BF8AEB00BE8A, 0xEB00818BEB00808B, 0xEB00838BEB00828B,
-    0xEB00858BEB00848B, 0xEB00878BEB00868B, 0xEB00898BEB00888B,
-    0xEB008B8BEB008A8B, 0xEB008D8BEB008C8B, 0xEB008F8BEB008E8B,
-    0xEB00918BEB00908B, 0xEB00938BEB00928B, 0xEB00958BEB00948B,
-    0xEB00978BEB00968B, 0xEB00998BEB00988B, 0xEB009B8BEB009A8B,
-    0xEB009D8BEB009C8B, 0xEB009F8BEB009E8B, 0xEB00A18BEB00A08B,
-    0xEB00A38BEB00A28B, 0xEB00A58BEB00A48B, 0xEB00A78BEB00A68B,
-    0xEB00A98BEB00A88B, 0xEB00AB8BEB00AA8B, 0xEB00AD8BEB00AC8B,
-    0xEB00AF8BEB00AE8B, 0xEB00B18BEB00B08B, 0xEB00B38BEB00B28B,
-    0xEB00B58BEB00B48B, 0xEB00B78BEB00B68B, 0xEB00B98BEB00B88B,
-    0xEB00BB8BEB00BA8B, 0xEB00BD8BEB00BC8B, 0xEB00BF8BEB00BE8B,
-    0xEB00818CEB00808C, 0xEB00838CEB00828C, 0xEB00858CEB00848C,
-    0xEB00878CEB00868C, 0xEB00898CEB00888C, 0xEB008B8CEB008A8C,
-    0xEB008D8CEB008C8C, 0xEB008F8CEB008E8C, 0xEB00918CEB00908C,
-    0xEB00938CEB00928C, 0xEB00958CEB00948C, 0xEB00978CEB00968C,
-    0xEB00998CEB00988C, 0xEB009B8CEB009A8C, 0xEB009D8CEB009C8C,
-    0xEB009F8CEB009E8C, 0xEB00A18CEB00A08C, 0xEB00A38CEB00A28C,
-    0xEB00A58CEB00A48C, 0xEB00A78CEB00A68C, 0xEB00A98CEB00A88C,
-    0xEB00AB8CEB00AA8C, 0xEB00AD8CEB00AC8C, 0xEB00AF8CEB00AE8C,
-    0xEB00B18CEB00B08C, 0xEB00B38CEB00B28C, 0xEB00B58CEB00B48C,
-    0xEB00B78CEB00B68C, 0xEB00B98CEB00B88C, 0xEB00BB8CEB00BA8C,
-    0xEB00BD8CEB00BC8C, 0xEB00BF8CEB00BE8C, 0xEB00818DEB00808D,
-    0xEB00838DEB00828D, 0xEB00858DEB00848D, 0xEB00878DEB00868D,
-    0xEB00898DEB00888D, 0xEB008B8DEB008A8D, 0xEB008D8DEB008C8D,
-    0xEB008F8DEB008E8D, 0xEB00918DEB00908D, 0xEB00938DEB00928D,
-    0xEB00958DEB00948D, 0xEB00978DEB00968D, 0xEB00998DEB00988D,
-    0xEB009B8DEB009A8D, 0xEB009D8DEB009C8D, 0xEB009F8DEB009E8D,
-    0xEB00A18DEB00A08D, 0xEB00A38DEB00A28D, 0xEB00A58DEB00A48D,
-    0xEB00A78DEB00A68D, 0xEB00A98DEB00A88D, 0xEB00AB8DEB00AA8D,
-    0xEB00AD8DEB00AC8D, 0xEB00AF8DEB00AE8D, 0xEB00B18DEB00B08D,
-    0xEB00B38DEB00B28D, 0xEB00B58DEB00B48D, 0xEB00B78DEB00B68D,
-    0xEB00B98DEB00B88D, 0xEB00BB8DEB00BA8D, 0xEB00BD8DEB00BC8D,
-    0xEB00BF8DEB00BE8D, 0xEB00818EEB00808E, 0xEB00838EEB00828E,
-    0xEB00858EEB00848E, 0xEB00878EEB00868E, 0xEB00898EEB00888E,
-    0xEB008B8EEB008A8E, 0xEB008D8EEB008C8E, 0xEB008F8EEB008E8E,
-    0xEB00918EEB00908E, 0xEB00938EEB00928E, 0xEB00958EEB00948E,
-    0xEB00978EEB00968E, 0xEB00998EEB00988E, 0xEB009B8EEB009A8E,
-    0xEB009D8EEB009C8E, 0xEB009F8EEB009E8E, 0xEB00A18EEB00A08E,
-    0xEB00A38EEB00A28E, 0xEB00A58EEB00A48E, 0xEB00A78EEB00A68E,
-    0xEB00A98EEB00A88E, 0xEB00AB8EEB00AA8E, 0xEB00AD8EEB00AC8E,
-    0xEB00AF8EEB00AE8E, 0xEB00B18EEB00B08E, 0xEB00B38EEB00B28E,
-    0xEB00B58EEB00B48E, 0xEB00B78EEB00B68E, 0xEB00B98EEB00B88E,
-    0xEB00BB8EEB00BA8E, 0xEB00BD8EEB00BC8E, 0xEB00BF8EEB00BE8E,
-    0xEB00818FEB00808F, 0xEB00838FEB00828F, 0xEB00858FEB00848F,
-    0xEB00878FEB00868F, 0xEB00898FEB00888F, 0xEB008B8FEB008A8F,
-    0xEB008D8FEB008C8F, 0xEB008F8FEB008E8F, 0xEB00918FEB00908F,
-    0xEB00938FEB00928F, 0xEB00958FEB00948F, 0xEB00978FEB00968F,
-    0xEB00998FEB00988F, 0xEB009B8FEB009A8F, 0xEB009D8FEB009C8F,
-    0xEB009F8FEB009E8F, 0xEB00A18FEB00A08F, 0xEB00A38FEB00A28F,
-    0xEB00A58FEB00A48F, 0xEB00A78FEB00A68F, 0xEB00A98FEB00A88F,
-    0xEB00AB8FEB00AA8F, 0xEB00AD8FEB00AC8F, 0xEB00AF8FEB00AE8F,
-    0xEB00B18FEB00B08F, 0xEB00B38FEB00B28F, 0xEB00B58FEB00B48F,
-    0xEB00B78FEB00B68F, 0xEB00B98FEB00B88F, 0xEB00BB8FEB00BA8F,
-    0xEB00BD8FEB00BC8F, 0xEB00BF8FEB00BE8F, 0xEB008190EB008090,
-    0xEB008390EB008290, 0xEB008590EB008490, 0xEB008790EB008690,
-    0xEB008990EB008890, 0xEB008B90EB008A90, 0xEB008D90EB008C90,
-    0xEB008F90EB008E90, 0xEB009190EB009090, 0xEB009390EB009290,
-    0xEB009590EB009490, 0xEB009790EB009690, 0xEB009990EB009890,
-    0xEB009B90EB009A90, 0xEB009D90EB009C90, 0xEB009F90EB009E90,
-    0xEB00A190EB00A090, 0xEB00A390EB00A290, 0xEB00A590EB00A490,
-    0xEB00A790EB00A690, 0xEB00A990EB00A890, 0xEB00AB90EB00AA90,
-    0xEB00AD90EB00AC90, 0xEB00AF90EB00AE90, 0xEB00B190EB00B090,
-    0xEB00B390EB00B290, 0xEB00B590EB00B490, 0xEB00B790EB00B690,
-    0xEB00B990EB00B890, 0xEB00BB90EB00BA90, 0xEB00BD90EB00BC90,
-    0xEB00BF90EB00BE90, 0xEB008191EB008091, 0xEB008391EB008291,
-    0xEB008591EB008491, 0xEB008791EB008691, 0xEB008991EB008891,
-    0xEB008B91EB008A91, 0xEB008D91EB008C91, 0xEB008F91EB008E91,
-    0xEB009191EB009091, 0xEB009391EB009291, 0xEB009591EB009491,
-    0xEB009791EB009691, 0xEB009991EB009891, 0xEB009B91EB009A91,
-    0xEB009D91EB009C91, 0xEB009F91EB009E91, 0xEB00A191EB00A091,
-    0xEB00A391EB00A291, 0xEB00A591EB00A491, 0xEB00A791EB00A691,
-    0xEB00A991EB00A891, 0xEB00AB91EB00AA91, 0xEB00AD91EB00AC91,
-    0xEB00AF91EB00AE91, 0xEB00B191EB00B091, 0xEB00B391EB00B291,
-    0xEB00B591EB00B491, 0xEB00B791EB00B691, 0xEB00B991EB00B891,
-    0xEB00BB91EB00BA91, 0xEB00BD91EB00BC91, 0xEB00BF91EB00BE91,
-    0xEB008192EB008092, 0xEB008392EB008292, 0xEB008592EB008492,
-    0xEB008792EB008692, 0xEB008992EB008892, 0xEB008B92EB008A92,
-    0xEB008D92EB008C92, 0xEB008F92EB008E92, 0xEB009192EB009092,
-    0xEB009392EB009292, 0xEB009592EB009492, 0xEB009792EB009692,
-    0xEB009992EB009892, 0xEB009B92EB009A92, 0xEB009D92EB009C92,
-    0xEB009F92EB009E92, 0xEB00A192EB00A092, 0xEB00A392EB00A292,
-    0xEB00A592EB00A492, 0xEB00A792EB00A692, 0xEB00A992EB00A892,
-    0xEB00AB92EB00AA92, 0xEB00AD92EB00AC92, 0xEB00AF92EB00AE92,
-    0xEB00B192EB00B092, 0xEB00B392EB00B292, 0xEB00B592EB00B492,
-    0xEB00B792EB00B692, 0xEB00B992EB00B892, 0xEB00BB92EB00BA92,
-    0xEB00BD92EB00BC92, 0xEB00BF92EB00BE92, 0xEB008193EB008093,
-    0xEB008393EB008293, 0xEB008593EB008493, 0xEB008793EB008693,
-    0xEB008993EB008893, 0xEB008B93EB008A93, 0xEB008D93EB008C93,
-    0xEB008F93EB008E93, 0xEB009193EB009093, 0xEB009393EB009293,
-    0xEB009593EB009493, 0xEB009793EB009693, 0xEB009993EB009893,
-    0xEB009B93EB009A93, 0xEB009D93EB009C93, 0xEB009F93EB009E93,
-    0xEB00A193EB00A093, 0xEB00A393EB00A293, 0xEB00A593EB00A493,
-    0xEB00A793EB00A693, 0xEB00A993EB00A893, 0xEB00AB93EB00AA93,
-    0xEB00AD93EB00AC93, 0xEB00AF93EB00AE93, 0xEB00B193EB00B093,
-    0xEB00B393EB00B293, 0xEB00B593EB00B493, 0xEB00B793EB00B693,
-    0xEB00B993EB00B893, 0xEB00BB93EB00BA93, 0xEB00BD93EB00BC93,
-    0xEB00BF93EB00BE93, 0xEB008194EB008094, 0xEB008394EB008294,
-    0xEB008594EB008494, 0xEB008794EB008694, 0xEB008994EB008894,
-    0xEB008B94EB008A94, 0xEB008D94EB008C94, 0xEB008F94EB008E94,
-    0xEB009194EB009094, 0xEB009394EB009294, 0xEB009594EB009494,
-    0xEB009794EB009694, 0xEB009994EB009894, 0xEB009B94EB009A94,
-    0xEB009D94EB009C94, 0xEB009F94EB009E94, 0xEB00A194EB00A094,
-    0xEB00A394EB00A294, 0xEB00A594EB00A494, 0xEB00A794EB00A694,
-    0xEB00A994EB00A894, 0xEB00AB94EB00AA94, 0xEB00AD94EB00AC94,
-    0xEB00AF94EB00AE94, 0xEB00B194EB00B094, 0xEB00B394EB00B294,
-    0xEB00B594EB00B494, 0xEB00B794EB00B694, 0xEB00B994EB00B894,
-    0xEB00BB94EB00BA94, 0xEB00BD94EB00BC94, 0xEB00BF94EB00BE94,
-    0xEB008195EB008095, 0xEB008395EB008295, 0xEB008595EB008495,
-    0xEB008795EB008695, 0xEB008995EB008895, 0xEB008B95EB008A95,
-    0xEB008D95EB008C95, 0xEB008F95EB008E95, 0xEB009195EB009095,
-    0xEB009395EB009295, 0xEB009595EB009495, 0xEB009795EB009695,
-    0xEB009995EB009895, 0xEB009B95EB009A95, 0xEB009D95EB009C95,
-    0xEB009F95EB009E95, 0xEB00A195EB00A095, 0xEB00A395EB00A295,
-    0xEB00A595EB00A495, 0xEB00A795EB00A695, 0xEB00A995EB00A895,
-    0xEB00AB95EB00AA95, 0xEB00AD95EB00AC95, 0xEB00AF95EB00AE95,
-    0xEB00B195EB00B095, 0xEB00B395EB00B295, 0xEB00B595EB00B495,
-    0xEB00B795EB00B695, 0xEB00B995EB00B895, 0xEB00BB95EB00BA95,
-    0xEB00BD95EB00BC95, 0xEB00BF95EB00BE95, 0xEB008196EB008096,
-    0xEB008396EB008296, 0xEB008596EB008496, 0xEB008796EB008696,
-    0xEB008996EB008896, 0xEB008B96EB008A96, 0xEB008D96EB008C96,
-    0xEB008F96EB008E96, 0xEB009196EB009096, 0xEB009396EB009296,
-    0xEB009596EB009496, 0xEB009796EB009696, 0xEB009996EB009896,
-    0xEB009B96EB009A96, 0xEB009D96EB009C96, 0xEB009F96EB009E96,
-    0xEB00A196EB00A096, 0xEB00A396EB00A296, 0xEB00A596EB00A496,
-    0xEB00A796EB00A696, 0xEB00A996EB00A896, 0xEB00AB96EB00AA96,
-    0xEB00AD96EB00AC96, 0xEB00AF96EB00AE96, 0xEB00B196EB00B096,
-    0xEB00B396EB00B296, 0xEB00B596EB00B496, 0xEB00B796EB00B696,
-    0xEB00B996EB00B896, 0xEB00BB96EB00BA96, 0xEB00BD96EB00BC96,
-    0xEB00BF96EB00BE96, 0xEB008197EB008097, 0xEB008397EB008297,
-    0xEB008597EB008497, 0xEB008797EB008697, 0xEB008997EB008897,
-    0xEB008B97EB008A97, 0xEB008D97EB008C97, 0xEB008F97EB008E97,
-    0xEB009197EB009097, 0xEB009397EB009297, 0xEB009597EB009497,
-    0xEB009797EB009697, 0xEB009997EB009897, 0xEB009B97EB009A97,
-    0xEB009D97EB009C97, 0xEB009F97EB009E97, 0xEB00A197EB00A097,
-    0xEB00A397EB00A297, 0xEB00A597EB00A497, 0xEB00A797EB00A697,
-    0xEB00A997EB00A897, 0xEB00AB97EB00AA97, 0xEB00AD97EB00AC97,
-    0xEB00AF97EB00AE97, 0xEB00B197EB00B097, 0xEB00B397EB00B297,
-    0xEB00B597EB00B497, 0xEB00B797EB00B697, 0xEB00B997EB00B897,
-    0xEB00BB97EB00BA97, 0xEB00BD97EB00BC97, 0xEB00BF97EB00BE97,
-    0xEB008198EB008098, 0xEB008398EB008298, 0xEB008598EB008498,
-    0xEB008798EB008698, 0xEB008998EB008898, 0xEB008B98EB008A98,
-    0xEB008D98EB008C98, 0xEB008F98EB008E98, 0xEB009198EB009098,
-    0xEB009398EB009298, 0xEB009598EB009498, 0xEB009798EB009698,
-    0xEB009998EB009898, 0xEB009B98EB009A98, 0xEB009D98EB009C98,
-    0xEB009F98EB009E98, 0xEB00A198EB00A098, 0xEB00A398EB00A298,
-    0xEB00A598EB00A498, 0xEB00A798EB00A698, 0xEB00A998EB00A898,
-    0xEB00AB98EB00AA98, 0xEB00AD98EB00AC98, 0xEB00AF98EB00AE98,
-    0xEB00B198EB00B098, 0xEB00B398EB00B298, 0xEB00B598EB00B498,
-    0xEB00B798EB00B698, 0xEB00B998EB00B898, 0xEB00BB98EB00BA98,
-    0xEB00BD98EB00BC98, 0xEB00BF98EB00BE98, 0xEB008199EB008099,
-    0xEB008399EB008299, 0xEB008599EB008499, 0xEB008799EB008699,
-    0xEB008999EB008899, 0xEB008B99EB008A99, 0xEB008D99EB008C99,
-    0xEB008F99EB008E99, 0xEB009199EB009099, 0xEB009399EB009299,
-    0xEB009599EB009499, 0xEB009799EB009699, 0xEB009999EB009899,
-    0xEB009B99EB009A99, 0xEB009D99EB009C99, 0xEB009F99EB009E99,
-    0xEB00A199EB00A099, 0xEB00A399EB00A299, 0xEB00A599EB00A499,
-    0xEB00A799EB00A699, 0xEB00A999EB00A899, 0xEB00AB99EB00AA99,
-    0xEB00AD99EB00AC99, 0xEB00AF99EB00AE99, 0xEB00B199EB00B099,
-    0xEB00B399EB00B299, 0xEB00B599EB00B499, 0xEB00B799EB00B699,
-    0xEB00B999EB00B899, 0xEB00BB99EB00BA99, 0xEB00BD99EB00BC99,
-    0xEB00BF99EB00BE99, 0xEB00819AEB00809A, 0xEB00839AEB00829A,
-    0xEB00859AEB00849A, 0xEB00879AEB00869A, 0xEB00899AEB00889A,
-    0xEB008B9AEB008A9A, 0xEB008D9AEB008C9A, 0xEB008F9AEB008E9A,
-    0xEB00919AEB00909A, 0xEB00939AEB00929A, 0xEB00959AEB00949A,
-    0xEB00979AEB00969A, 0xEB00999AEB00989A, 0xEB009B9AEB009A9A,
-    0xEB009D9AEB009C9A, 0xEB009F9AEB009E9A, 0xEB00A19AEB00A09A,
-    0xEB00A39AEB00A29A, 0xEB00A59AEB00A49A, 0xEB00A79AEB00A69A,
-    0xEB00A99AEB00A89A, 0xEB00AB9AEB00AA9A, 0xEB00AD9AEB00AC9A,
-    0xEB00AF9AEB00AE9A, 0xEB00B19AEB00B09A, 0xEB00B39AEB00B29A,
-    0xEB00B59AEB00B49A, 0xEB00B79AEB00B69A, 0xEB00B99AEB00B89A,
-    0xEB00BB9AEB00BA9A, 0xEB00BD9AEB00BC9A, 0xEB00BF9AEB00BE9A,
-    0xEB00819BEB00809B, 0xEB00839BEB00829B, 0xEB00859BEB00849B,
-    0xEB00879BEB00869B, 0xEB00899BEB00889B, 0xEB008B9BEB008A9B,
-    0xEB008D9BEB008C9B, 0xEB008F9BEB008E9B, 0xEB00919BEB00909B,
-    0xEB00939BEB00929B, 0xEB00959BEB00949B, 0xEB00979BEB00969B,
-    0xEB00999BEB00989B, 0xEB009B9BEB009A9B, 0xEB009D9BEB009C9B,
-    0xEB009F9BEB009E9B, 0xEB00A19BEB00A09B, 0xEB00A39BEB00A29B,
-    0xEB00A59BEB00A49B, 0xEB00A79BEB00A69B, 0xEB00A99BEB00A89B,
-    0xEB00AB9BEB00AA9B, 0xEB00AD9BEB00AC9B, 0xEB00AF9BEB00AE9B,
-    0xEB00B19BEB00B09B, 0xEB00B39BEB00B29B, 0xEB00B59BEB00B49B,
-    0xEB00B79BEB00B69B, 0xEB00B99BEB00B89B, 0xEB00BB9BEB00BA9B,
-    0xEB00BD9BEB00BC9B, 0xEB00BF9BEB00BE9B, 0xEB00819CEB00809C,
-    0xEB00839CEB00829C, 0xEB00859CEB00849C, 0xEB00879CEB00869C,
-    0xEB00899CEB00889C, 0xEB008B9CEB008A9C, 0xEB008D9CEB008C9C,
-    0xEB008F9CEB008E9C, 0xEB00919CEB00909C, 0xEB00939CEB00929C,
-    0xEB00959CEB00949C, 0xEB00979CEB00969C, 0xEB00999CEB00989C,
-    0xEB009B9CEB009A9C, 0xEB009D9CEB009C9C, 0xEB009F9CEB009E9C,
-    0xEB00A19CEB00A09C, 0xEB00A39CEB00A29C, 0xEB00A59CEB00A49C,
-    0xEB00A79CEB00A69C, 0xEB00A99CEB00A89C, 0xEB00AB9CEB00AA9C,
-    0xEB00AD9CEB00AC9C, 0xEB00AF9CEB00AE9C, 0xEB00B19CEB00B09C,
-    0xEB00B39CEB00B29C, 0xEB00B59CEB00B49C, 0xEB00B79CEB00B69C,
-    0xEB00B99CEB00B89C, 0xEB00BB9CEB00BA9C, 0xEB00BD9CEB00BC9C,
-    0xEB00BF9CEB00BE9C, 0xEB00819DEB00809D, 0xEB00839DEB00829D,
-    0xEB00859DEB00849D, 0xEB00879DEB00869D, 0xEB00899DEB00889D,
-    0xEB008B9DEB008A9D, 0xEB008D9DEB008C9D, 0xEB008F9DEB008E9D,
-    0xEB00919DEB00909D, 0xEB00939DEB00929D, 0xEB00959DEB00949D,
-    0xEB00979DEB00969D, 0xEB00999DEB00989D, 0xEB009B9DEB009A9D,
-    0xEB009D9DEB009C9D, 0xEB009F9DEB009E9D, 0xEB00A19DEB00A09D,
-    0xEB00A39DEB00A29D, 0xEB00A59DEB00A49D, 0xEB00A79DEB00A69D,
-    0xEB00A99DEB00A89D, 0xEB00AB9DEB00AA9D, 0xEB00AD9DEB00AC9D,
-    0xEB00AF9DEB00AE9D, 0xEB00B19DEB00B09D, 0xEB00B39DEB00B29D,
-    0xEB00B59DEB00B49D, 0xEB00B79DEB00B69D, 0xEB00B99DEB00B89D,
-    0xEB00BB9DEB00BA9D, 0xEB00BD9DEB00BC9D, 0xEB00BF9DEB00BE9D,
-    0xEB00819EEB00809E, 0xEB00839EEB00829E, 0xEB00859EEB00849E,
-    0xEB00879EEB00869E, 0xEB00899EEB00889E, 0xEB008B9EEB008A9E,
-    0xEB008D9EEB008C9E, 0xEB008F9EEB008E9E, 0xEB00919EEB00909E,
-    0xEB00939EEB00929E, 0xEB00959EEB00949E, 0xEB00979EEB00969E,
-    0xEB00999EEB00989E, 0xEB009B9EEB009A9E, 0xEB009D9EEB009C9E,
-    0xEB009F9EEB009E9E, 0xEB00A19EEB00A09E, 0xEB00A39EEB00A29E,
-    0xEB00A59EEB00A49E, 0xEB00A79EEB00A69E, 0xEB00A99EEB00A89E,
-    0xEB00AB9EEB00AA9E, 0xEB00AD9EEB00AC9E, 0xEB00AF9EEB00AE9E,
-    0xEB00B19EEB00B09E, 0xEB00B39EEB00B29E, 0xEB00B59EEB00B49E,
-    0xEB00B79EEB00B69E, 0xEB00B99EEB00B89E, 0xEB00BB9EEB00BA9E,
-    0xEB00BD9EEB00BC9E, 0xEB00BF9EEB00BE9E, 0xEB00819FEB00809F,
-    0xEB00839FEB00829F, 0xEB00859FEB00849F, 0xEB00879FEB00869F,
-    0xEB00899FEB00889F, 0xEB008B9FEB008A9F, 0xEB008D9FEB008C9F,
-    0xEB008F9FEB008E9F, 0xEB00919FEB00909F, 0xEB00939FEB00929F,
-    0xEB00959FEB00949F, 0xEB00979FEB00969F, 0xEB00999FEB00989F,
-    0xEB009B9FEB009A9F, 0xEB009D9FEB009C9F, 0xEB009F9FEB009E9F,
-    0xEB00A19FEB00A09F, 0xEB00A39FEB00A29F, 0xEB00A59FEB00A49F,
-    0xEB00A79FEB00A69F, 0xEB00A99FEB00A89F, 0xEB00AB9FEB00AA9F,
-    0xEB00AD9FEB00AC9F, 0xEB00AF9FEB00AE9F, 0xEB00B19FEB00B09F,
-    0xEB00B39FEB00B29F, 0xEB00B59FEB00B49F, 0xEB00B79FEB00B69F,
-    0xEB00B99FEB00B89F, 0xEB00BB9FEB00BA9F, 0xEB00BD9FEB00BC9F,
-    0xEB00BF9FEB00BE9F, 0xEB0081A0EB0080A0, 0xEB0083A0EB0082A0,
-    0xEB0085A0EB0084A0, 0xEB0087A0EB0086A0, 0xEB0089A0EB0088A0,
-    0xEB008BA0EB008AA0, 0xEB008DA0EB008CA0, 0xEB008FA0EB008EA0,
-    0xEB0091A0EB0090A0, 0xEB0093A0EB0092A0, 0xEB0095A0EB0094A0,
-    0xEB0097A0EB0096A0, 0xEB0099A0EB0098A0, 0xEB009BA0EB009AA0,
-    0xEB009DA0EB009CA0, 0xEB009FA0EB009EA0, 0xEB00A1A0EB00A0A0,
-    0xEB00A3A0EB00A2A0, 0xEB00A5A0EB00A4A0, 0xEB00A7A0EB00A6A0,
-    0xEB00A9A0EB00A8A0, 0xEB00ABA0EB00AAA0, 0xEB00ADA0EB00ACA0,
-    0xEB00AFA0EB00AEA0, 0xEB00B1A0EB00B0A0, 0xEB00B3A0EB00B2A0,
-    0xEB00B5A0EB00B4A0, 0xEB00B7A0EB00B6A0, 0xEB00B9A0EB00B8A0,
-    0xEB00BBA0EB00BAA0, 0xEB00BDA0EB00BCA0, 0xEB00BFA0EB00BEA0,
-    0xEB0081A1EB0080A1, 0xEB0083A1EB0082A1, 0xEB0085A1EB0084A1,
-    0xEB0087A1EB0086A1, 0xEB0089A1EB0088A1, 0xEB008BA1EB008AA1,
-    0xEB008DA1EB008CA1, 0xEB008FA1EB008EA1, 0xEB0091A1EB0090A1,
-    0xEB0093A1EB0092A1, 0xEB0095A1EB0094A1, 0xEB0097A1EB0096A1,
-    0xEB0099A1EB0098A1, 0xEB009BA1EB009AA1, 0xEB009DA1EB009CA1,
-    0xEB009FA1EB009EA1, 0xEB00A1A1EB00A0A1, 0xEB00A3A1EB00A2A1,
-    0xEB00A5A1EB00A4A1, 0xEB00A7A1EB00A6A1, 0xEB00A9A1EB00A8A1,
-    0xEB00ABA1EB00AAA1, 0xEB00ADA1EB00ACA1, 0xEB00AFA1EB00AEA1,
-    0xEB00B1A1EB00B0A1, 0xEB00B3A1EB00B2A1, 0xEB00B5A1EB00B4A1,
-    0xEB00B7A1EB00B6A1, 0xEB00B9A1EB00B8A1, 0xEB00BBA1EB00BAA1,
-    0xEB00BDA1EB00BCA1, 0xEB00BFA1EB00BEA1, 0xEB0081A2EB0080A2,
-    0xEB0083A2EB0082A2, 0xEB0085A2EB0084A2, 0xEB0087A2EB0086A2,
-    0xEB0089A2EB0088A2, 0xEB008BA2EB008AA2, 0xEB008DA2EB008CA2,
-    0xEB008FA2EB008EA2, 0xEB0091A2EB0090A2, 0xEB0093A2EB0092A2,
-    0xEB0095A2EB0094A2, 0xEB0097A2EB0096A2, 0xEB0099A2EB0098A2,
-    0xEB009BA2EB009AA2, 0xEB009DA2EB009CA2, 0xEB009FA2EB009EA2,
-    0xEB00A1A2EB00A0A2, 0xEB00A3A2EB00A2A2, 0xEB00A5A2EB00A4A2,
-    0xEB00A7A2EB00A6A2, 0xEB00A9A2EB00A8A2, 0xEB00ABA2EB00AAA2,
-    0xEB00ADA2EB00ACA2, 0xEB00AFA2EB00AEA2, 0xEB00B1A2EB00B0A2,
-    0xEB00B3A2EB00B2A2, 0xEB00B5A2EB00B4A2, 0xEB00B7A2EB00B6A2,
-    0xEB00B9A2EB00B8A2, 0xEB00BBA2EB00BAA2, 0xEB00BDA2EB00BCA2,
-    0xEB00BFA2EB00BEA2, 0xEB0081A3EB0080A3, 0xEB0083A3EB0082A3,
-    0xEB0085A3EB0084A3, 0xEB0087A3EB0086A3, 0xEB0089A3EB0088A3,
-    0xEB008BA3EB008AA3, 0xEB008DA3EB008CA3, 0xEB008FA3EB008EA3,
-    0xEB0091A3EB0090A3, 0xEB0093A3EB0092A3, 0xEB0095A3EB0094A3,
-    0xEB0097A3EB0096A3, 0xEB0099A3EB0098A3, 0xEB009BA3EB009AA3,
-    0xEB009DA3EB009CA3, 0xEB009FA3EB009EA3, 0xEB00A1A3EB00A0A3,
-    0xEB00A3A3EB00A2A3, 0xEB00A5A3EB00A4A3, 0xEB00A7A3EB00A6A3,
-    0xEB00A9A3EB00A8A3, 0xEB00ABA3EB00AAA3, 0xEB00ADA3EB00ACA3,
-    0xEB00AFA3EB00AEA3, 0xEB00B1A3EB00B0A3, 0xEB00B3A3EB00B2A3,
-    0xEB00B5A3EB00B4A3, 0xEB00B7A3EB00B6A3, 0xEB00B9A3EB00B8A3,
-    0xEB00BBA3EB00BAA3, 0xEB00BDA3EB00BCA3, 0xEB00BFA3EB00BEA3,
-    0xEB0081A4EB0080A4, 0xEB0083A4EB0082A4, 0xEB0085A4EB0084A4,
-    0xEB0087A4EB0086A4, 0xEB0089A4EB0088A4, 0xEB008BA4EB008AA4,
-    0xEB008DA4EB008CA4, 0xEB008FA4EB008EA4, 0xEB0091A4EB0090A4,
-    0xEB0093A4EB0092A4, 0xEB0095A4EB0094A4, 0xEB0097A4EB0096A4,
-    0xEB0099A4EB0098A4, 0xEB009BA4EB009AA4, 0xEB009DA4EB009CA4,
-    0xEB009FA4EB009EA4, 0xEB00A1A4EB00A0A4, 0xEB00A3A4EB00A2A4,
-    0xEB00A5A4EB00A4A4, 0xEB00A7A4EB00A6A4, 0xEB00A9A4EB00A8A4,
-    0xEB00ABA4EB00AAA4, 0xEB00ADA4EB00ACA4, 0xEB00AFA4EB00AEA4,
-    0xEB00B1A4EB00B0A4, 0xEB00B3A4EB00B2A4, 0xEB00B5A4EB00B4A4,
-    0xEB00B7A4EB00B6A4, 0xEB00B9A4EB00B8A4, 0xEB00BBA4EB00BAA4,
-    0xEB00BDA4EB00BCA4, 0xEB00BFA4EB00BEA4, 0xEB0081A5EB0080A5,
-    0xEB0083A5EB0082A5, 0xEB0085A5EB0084A5, 0xEB0087A5EB0086A5,
-    0xEB0089A5EB0088A5, 0xEB008BA5EB008AA5, 0xEB008DA5EB008CA5,
-    0xEB008FA5EB008EA5, 0xEB0091A5EB0090A5, 0xEB0093A5EB0092A5,
-    0xEB0095A5EB0094A5, 0xEB0097A5EB0096A5, 0xEB0099A5EB0098A5,
-    0xEB009BA5EB009AA5, 0xEB009DA5EB009CA5, 0xEB009FA5EB009EA5,
-    0xEB00A1A5EB00A0A5, 0xEB00A3A5EB00A2A5, 0xEB00A5A5EB00A4A5,
-    0xEB00A7A5EB00A6A5, 0xEB00A9A5EB00A8A5, 0xEB00ABA5EB00AAA5,
-    0xEB00ADA5EB00ACA5, 0xEB00AFA5EB00AEA5, 0xEB00B1A5EB00B0A5,
-    0xEB00B3A5EB00B2A5, 0xEB00B5A5EB00B4A5, 0xEB00B7A5EB00B6A5,
-    0xEB00B9A5EB00B8A5, 0xEB00BBA5EB00BAA5, 0xEB00BDA5EB00BCA5,
-    0xEB00BFA5EB00BEA5, 0xEB0081A6EB0080A6, 0xEB0083A6EB0082A6,
-    0xEB0085A6EB0084A6, 0xEB0087A6EB0086A6, 0xEB0089A6EB0088A6,
-    0xEB008BA6EB008AA6, 0xEB008DA6EB008CA6, 0xEB008FA6EB008EA6,
-    0xEB0091A6EB0090A6, 0xEB0093A6EB0092A6, 0xEB0095A6EB0094A6,
-    0xEB0097A6EB0096A6, 0xEB0099A6EB0098A6, 0xEB009BA6EB009AA6,
-    0xEB009DA6EB009CA6, 0xEB009FA6EB009EA6, 0xEB00A1A6EB00A0A6,
-    0xEB00A3A6EB00A2A6, 0xEB00A5A6EB00A4A6, 0xEB00A7A6EB00A6A6,
-    0xEB00A9A6EB00A8A6, 0xEB00ABA6EB00AAA6, 0xEB00ADA6EB00ACA6,
-    0xEB00AFA6EB00AEA6, 0xEB00B1A6EB00B0A6, 0xEB00B3A6EB00B2A6,
-    0xEB00B5A6EB00B4A6, 0xEB00B7A6EB00B6A6, 0xEB00B9A6EB00B8A6,
-    0xEB00BBA6EB00BAA6, 0xEB00BDA6EB00BCA6, 0xEB00BFA6EB00BEA6,
-    0xEB0081A7EB0080A7, 0xEB0083A7EB0082A7, 0xEB0085A7EB0084A7,
-    0xEB0087A7EB0086A7, 0xEB0089A7EB0088A7, 0xEB008BA7EB008AA7,
-    0xEB008DA7EB008CA7, 0xEB008FA7EB008EA7, 0xEB0091A7EB0090A7,
-    0xEB0093A7EB0092A7, 0xEB0095A7EB0094A7, 0xEB0097A7EB0096A7,
-    0xEB0099A7EB0098A7, 0xEB009BA7EB009AA7, 0xEB009DA7EB009CA7,
-    0xEB009FA7EB009EA7, 0xEB00A1A7EB00A0A7, 0xEB00A3A7EB00A2A7,
-    0xEB00A5A7EB00A4A7, 0xEB00A7A7EB00A6A7, 0xEB00A9A7EB00A8A7,
-    0xEB00ABA7EB00AAA7, 0xEB00ADA7EB00ACA7, 0xEB00AFA7EB00AEA7,
-    0xEB00B1A7EB00B0A7, 0xEB00B3A7EB00B2A7, 0xEB00B5A7EB00B4A7,
-    0xEB00B7A7EB00B6A7, 0xEB00B9A7EB00B8A7, 0xEB00BBA7EB00BAA7,
-    0xEB00BDA7EB00BCA7, 0xEB00BFA7EB00BEA7, 0xEB0081A8EB0080A8,
-    0xEB0083A8EB0082A8, 0xEB0085A8EB0084A8, 0xEB0087A8EB0086A8,
-    0xEB0089A8EB0088A8, 0xEB008BA8EB008AA8, 0xEB008DA8EB008CA8,
-    0xEB008FA8EB008EA8, 0xEB0091A8EB0090A8, 0xEB0093A8EB0092A8,
-    0xEB0095A8EB0094A8, 0xEB0097A8EB0096A8, 0xEB0099A8EB0098A8,
-    0xEB009BA8EB009AA8, 0xEB009DA8EB009CA8, 0xEB009FA8EB009EA8,
-    0xEB00A1A8EB00A0A8, 0xEB00A3A8EB00A2A8, 0xEB00A5A8EB00A4A8,
-    0xEB00A7A8EB00A6A8, 0xEB00A9A8EB00A8A8, 0xEB00ABA8EB00AAA8,
-    0xEB00ADA8EB00ACA8, 0xEB00AFA8EB00AEA8, 0xEB00B1A8EB00B0A8,
-    0xEB00B3A8EB00B2A8, 0xEB00B5A8EB00B4A8, 0xEB00B7A8EB00B6A8,
-    0xEB00B9A8EB00B8A8, 0xEB00BBA8EB00BAA8, 0xEB00BDA8EB00BCA8,
-    0xEB00BFA8EB00BEA8, 0xEB0081A9EB0080A9, 0xEB0083A9EB0082A9,
-    0xEB0085A9EB0084A9, 0xEB0087A9EB0086A9, 0xEB0089A9EB0088A9,
-    0xEB008BA9EB008AA9, 0xEB008DA9EB008CA9, 0xEB008FA9EB008EA9,
-    0xEB0091A9EB0090A9, 0xEB0093A9EB0092A9, 0xEB0095A9EB0094A9,
-    0xEB0097A9EB0096A9, 0xEB0099A9EB0098A9, 0xEB009BA9EB009AA9,
-    0xEB009DA9EB009CA9, 0xEB009FA9EB009EA9, 0xEB00A1A9EB00A0A9,
-    0xEB00A3A9EB00A2A9, 0xEB00A5A9EB00A4A9, 0xEB00A7A9EB00A6A9,
-    0xEB00A9A9EB00A8A9, 0xEB00ABA9EB00AAA9, 0xEB00ADA9EB00ACA9,
-    0xEB00AFA9EB00AEA9, 0xEB00B1A9EB00B0A9, 0xEB00B3A9EB00B2A9,
-    0xEB00B5A9EB00B4A9, 0xEB00B7A9EB00B6A9, 0xEB00B9A9EB00B8A9,
-    0xEB00BBA9EB00BAA9, 0xEB00BDA9EB00BCA9, 0xEB00BFA9EB00BEA9,
-    0xEB0081AAEB0080AA, 0xEB0083AAEB0082AA, 0xEB0085AAEB0084AA,
-    0xEB0087AAEB0086AA, 0xEB0089AAEB0088AA, 0xEB008BAAEB008AAA,
-    0xEB008DAAEB008CAA, 0xEB008FAAEB008EAA, 0xEB0091AAEB0090AA,
-    0xEB0093AAEB0092AA, 0xEB0095AAEB0094AA, 0xEB0097AAEB0096AA,
-    0xEB0099AAEB0098AA, 0xEB009BAAEB009AAA, 0xEB009DAAEB009CAA,
-    0xEB009FAAEB009EAA, 0xEB00A1AAEB00A0AA, 0xEB00A3AAEB00A2AA,
-    0xEB00A5AAEB00A4AA, 0xEB00A7AAEB00A6AA, 0xEB00A9AAEB00A8AA,
-    0xEB00ABAAEB00AAAA, 0xEB00ADAAEB00ACAA, 0xEB00AFAAEB00AEAA,
-    0xEB00B1AAEB00B0AA, 0xEB00B3AAEB00B2AA, 0xEB00B5AAEB00B4AA,
-    0xEB00B7AAEB00B6AA, 0xEB00B9AAEB00B8AA, 0xEB00BBAAEB00BAAA,
-    0xEB00BDAAEB00BCAA, 0xEB00BFAAEB00BEAA, 0xEB0081ABEB0080AB,
-    0xEB0083ABEB0082AB, 0xEB0085ABEB0084AB, 0xEB0087ABEB0086AB,
-    0xEB0089ABEB0088AB, 0xEB008BABEB008AAB, 0xEB008DABEB008CAB,
-    0xEB008FABEB008EAB, 0xEB0091ABEB0090AB, 0xEB0093ABEB0092AB,
-    0xEB0095ABEB0094AB, 0xEB0097ABEB0096AB, 0xEB0099ABEB0098AB,
-    0xEB009BABEB009AAB, 0xEB009DABEB009CAB, 0xEB009FABEB009EAB,
-    0xEB00A1ABEB00A0AB, 0xEB00A3ABEB00A2AB, 0xEB00A5ABEB00A4AB,
-    0xEB00A7ABEB00A6AB, 0xEB00A9ABEB00A8AB, 0xEB00ABABEB00AAAB,
-    0xEB00ADABEB00ACAB, 0xEB00AFABEB00AEAB, 0xEB00B1ABEB00B0AB,
-    0xEB00B3ABEB00B2AB, 0xEB00B5ABEB00B4AB, 0xEB00B7ABEB00B6AB,
-    0xEB00B9ABEB00B8AB, 0xEB00BBABEB00BAAB, 0xEB00BDABEB00BCAB,
-    0xEB00BFABEB00BEAB, 0xEB0081ACEB0080AC, 0xEB0083ACEB0082AC,
-    0xEB0085ACEB0084AC, 0xEB0087ACEB0086AC, 0xEB0089ACEB0088AC,
-    0xEB008BACEB008AAC, 0xEB008DACEB008CAC, 0xEB008FACEB008EAC,
-    0xEB0091ACEB0090AC, 0xEB0093ACEB0092AC, 0xEB0095ACEB0094AC,
-    0xEB0097ACEB0096AC, 0xEB0099ACEB0098AC, 0xEB009BACEB009AAC,
-    0xEB009DACEB009CAC, 0xEB009FACEB009EAC, 0xEB00A1ACEB00A0AC,
-    0xEB00A3ACEB00A2AC, 0xEB00A5ACEB00A4AC, 0xEB00A7ACEB00A6AC,
-    0xEB00A9ACEB00A8AC, 0xEB00ABACEB00AAAC, 0xEB00ADACEB00ACAC,
-    0xEB00AFACEB00AEAC, 0xEB00B1ACEB00B0AC, 0xEB00B3ACEB00B2AC,
-    0xEB00B5ACEB00B4AC, 0xEB00B7ACEB00B6AC, 0xEB00B9ACEB00B8AC,
-    0xEB00BBACEB00BAAC, 0xEB00BDACEB00BCAC, 0xEB00BFACEB00BEAC,
-    0xEB0081ADEB0080AD, 0xEB0083ADEB0082AD, 0xEB0085ADEB0084AD,
-    0xEB0087ADEB0086AD, 0xEB0089ADEB0088AD, 0xEB008BADEB008AAD,
-    0xEB008DADEB008CAD, 0xEB008FADEB008EAD, 0xEB0091ADEB0090AD,
-    0xEB0093ADEB0092AD, 0xEB0095ADEB0094AD, 0xEB0097ADEB0096AD,
-    0xEB0099ADEB0098AD, 0xEB009BADEB009AAD, 0xEB009DADEB009CAD,
-    0xEB009FADEB009EAD, 0xEB00A1ADEB00A0AD, 0xEB00A3ADEB00A2AD,
-    0xEB00A5ADEB00A4AD, 0xEB00A7ADEB00A6AD, 0xEB00A9ADEB00A8AD,
-    0xEB00ABADEB00AAAD, 0xEB00ADADEB00ACAD, 0xEB00AFADEB00AEAD,
-    0xEB00B1ADEB00B0AD, 0xEB00B3ADEB00B2AD, 0xEB00B5ADEB00B4AD,
-    0xEB00B7ADEB00B6AD, 0xEB00B9ADEB00B8AD, 0xEB00BBADEB00BAAD,
-    0xEB00BDADEB00BCAD, 0xEB00BFADEB00BEAD, 0xEB0081AEEB0080AE,
-    0xEB0083AEEB0082AE, 0xEB0085AEEB0084AE, 0xEB0087AEEB0086AE,
-    0xEB0089AEEB0088AE, 0xEB008BAEEB008AAE, 0xEB008DAEEB008CAE,
-    0xEB008FAEEB008EAE, 0xEB0091AEEB0090AE, 0xEB0093AEEB0092AE,
-    0xEB0095AEEB0094AE, 0xEB0097AEEB0096AE, 0xEB0099AEEB0098AE,
-    0xEB009BAEEB009AAE, 0xEB009DAEEB009CAE, 0xEB009FAEEB009EAE,
-    0xEB00A1AEEB00A0AE, 0xEB00A3AEEB00A2AE, 0xEB00A5AEEB00A4AE,
-    0xEB00A7AEEB00A6AE, 0xEB00A9AEEB00A8AE, 0xEB00ABAEEB00AAAE,
-    0xEB00ADAEEB00ACAE, 0xEB00AFAEEB00AEAE, 0xEB00B1AEEB00B0AE,
-    0xEB00B3AEEB00B2AE, 0xEB00B5AEEB00B4AE, 0xEB00B7AEEB00B6AE,
-    0xEB00B9AEEB00B8AE, 0xEB00BBAEEB00BAAE, 0xEB00BDAEEB00BCAE,
-    0xEB00BFAEEB00BEAE, 0xEB0081AFEB0080AF, 0xEB0083AFEB0082AF,
-    0xEB0085AFEB0084AF, 0xEB0087AFEB0086AF, 0xEB0089AFEB0088AF,
-    0xEB008BAFEB008AAF, 0xEB008DAFEB008CAF, 0xEB008FAFEB008EAF,
-    0xEB0091AFEB0090AF, 0xEB0093AFEB0092AF, 0xEB0095AFEB0094AF,
-    0xEB0097AFEB0096AF, 0xEB0099AFEB0098AF, 0xEB009BAFEB009AAF,
-    0xEB009DAFEB009CAF, 0xEB009FAFEB009EAF, 0xEB00A1AFEB00A0AF,
-    0xEB00A3AFEB00A2AF, 0xEB00A5AFEB00A4AF, 0xEB00A7AFEB00A6AF,
-    0xEB00A9AFEB00A8AF, 0xEB00ABAFEB00AAAF, 0xEB00ADAFEB00ACAF,
-    0xEB00AFAFEB00AEAF, 0xEB00B1AFEB00B0AF, 0xEB00B3AFEB00B2AF,
-    0xEB00B5AFEB00B4AF, 0xEB00B7AFEB00B6AF, 0xEB00B9AFEB00B8AF,
-    0xEB00BBAFEB00BAAF, 0xEB00BDAFEB00BCAF, 0xEB00BFAFEB00BEAF,
-    0xEB0081B0EB0080B0, 0xEB0083B0EB0082B0, 0xEB0085B0EB0084B0,
-    0xEB0087B0EB0086B0, 0xEB0089B0EB0088B0, 0xEB008BB0EB008AB0,
-    0xEB008DB0EB008CB0, 0xEB008FB0EB008EB0, 0xEB0091B0EB0090B0,
-    0xEB0093B0EB0092B0, 0xEB0095B0EB0094B0, 0xEB0097B0EB0096B0,
-    0xEB0099B0EB0098B0, 0xEB009BB0EB009AB0, 0xEB009DB0EB009CB0,
-    0xEB009FB0EB009EB0, 0xEB00A1B0EB00A0B0, 0xEB00A3B0EB00A2B0,
-    0xEB00A5B0EB00A4B0, 0xEB00A7B0EB00A6B0, 0xEB00A9B0EB00A8B0,
-    0xEB00ABB0EB00AAB0, 0xEB00ADB0EB00ACB0, 0xEB00AFB0EB00AEB0,
-    0xEB00B1B0EB00B0B0, 0xEB00B3B0EB00B2B0, 0xEB00B5B0EB00B4B0,
-    0xEB00B7B0EB00B6B0, 0xEB00B9B0EB00B8B0, 0xEB00BBB0EB00BAB0,
-    0xEB00BDB0EB00BCB0, 0xEB00BFB0EB00BEB0, 0xEB0081B1EB0080B1,
-    0xEB0083B1EB0082B1, 0xEB0085B1EB0084B1, 0xEB0087B1EB0086B1,
-    0xEB0089B1EB0088B1, 0xEB008BB1EB008AB1, 0xEB008DB1EB008CB1,
-    0xEB008FB1EB008EB1, 0xEB0091B1EB0090B1, 0xEB0093B1EB0092B1,
-    0xEB0095B1EB0094B1, 0xEB0097B1EB0096B1, 0xEB0099B1EB0098B1,
-    0xEB009BB1EB009AB1, 0xEB009DB1EB009CB1, 0xEB009FB1EB009EB1,
-    0xEB00A1B1EB00A0B1, 0xEB00A3B1EB00A2B1, 0xEB00A5B1EB00A4B1,
-    0xEB00A7B1EB00A6B1, 0xEB00A9B1EB00A8B1, 0xEB00ABB1EB00AAB1,
-    0xEB00ADB1EB00ACB1, 0xEB00AFB1EB00AEB1, 0xEB00B1B1EB00B0B1,
-    0xEB00B3B1EB00B2B1, 0xEB00B5B1EB00B4B1, 0xEB00B7B1EB00B6B1,
-    0xEB00B9B1EB00B8B1, 0xEB00BBB1EB00BAB1, 0xEB00BDB1EB00BCB1,
-    0xEB00BFB1EB00BEB1, 0xEB0081B2EB0080B2, 0xEB0083B2EB0082B2,
-    0xEB0085B2EB0084B2, 0xEB0087B2EB0086B2, 0xEB0089B2EB0088B2,
-    0xEB008BB2EB008AB2, 0xEB008DB2EB008CB2, 0xEB008FB2EB008EB2,
-    0xEB0091B2EB0090B2, 0xEB0093B2EB0092B2, 0xEB0095B2EB0094B2,
-    0xEB0097B2EB0096B2, 0xEB0099B2EB0098B2, 0xEB009BB2EB009AB2,
-    0xEB009DB2EB009CB2, 0xEB009FB2EB009EB2, 0xEB00A1B2EB00A0B2,
-    0xEB00A3B2EB00A2B2, 0xEB00A5B2EB00A4B2, 0xEB00A7B2EB00A6B2,
-    0xEB00A9B2EB00A8B2, 0xEB00ABB2EB00AAB2, 0xEB00ADB2EB00ACB2,
-    0xEB00AFB2EB00AEB2, 0xEB00B1B2EB00B0B2, 0xEB00B3B2EB00B2B2,
-    0xEB00B5B2EB00B4B2, 0xEB00B7B2EB00B6B2, 0xEB00B9B2EB00B8B2,
-    0xEB00BBB2EB00BAB2, 0xEB00BDB2EB00BCB2, 0xEB00BFB2EB00BEB2,
-    0xEB0081B3EB0080B3, 0xEB0083B3EB0082B3, 0xEB0085B3EB0084B3,
-    0xEB0087B3EB0086B3, 0xEB0089B3EB0088B3, 0xEB008BB3EB008AB3,
-    0xEB008DB3EB008CB3, 0xEB008FB3EB008EB3, 0xEB0091B3EB0090B3,
-    0xEB0093B3EB0092B3, 0xEB0095B3EB0094B3, 0xEB0097B3EB0096B3,
-    0xEB0099B3EB0098B3, 0xEB009BB3EB009AB3, 0xEB009DB3EB009CB3,
-    0xEB009FB3EB009EB3, 0xEB00A1B3EB00A0B3, 0xEB00A3B3EB00A2B3,
-    0xEB00A5B3EB00A4B3, 0xEB00A7B3EB00A6B3, 0xEB00A9B3EB00A8B3,
-    0xEB00ABB3EB00AAB3, 0xEB00ADB3EB00ACB3, 0xEB00AFB3EB00AEB3,
-    0xEB00B1B3EB00B0B3, 0xEB00B3B3EB00B2B3, 0xEB00B5B3EB00B4B3,
-    0xEB00B7B3EB00B6B3, 0xEB00B9B3EB00B8B3, 0xEB00BBB3EB00BAB3,
-    0xEB00BDB3EB00BCB3, 0xEB00BFB3EB00BEB3, 0xEB0081B4EB0080B4,
-    0xEB0083B4EB0082B4, 0xEB0085B4EB0084B4, 0xEB0087B4EB0086B4,
-    0xEB0089B4EB0088B4, 0xEB008BB4EB008AB4, 0xEB008DB4EB008CB4,
-    0xEB008FB4EB008EB4, 0xEB0091B4EB0090B4, 0xEB0093B4EB0092B4,
-    0xEB0095B4EB0094B4, 0xEB0097B4EB0096B4, 0xEB0099B4EB0098B4,
-    0xEB009BB4EB009AB4, 0xEB009DB4EB009CB4, 0xEB009FB4EB009EB4,
-    0xEB00A1B4EB00A0B4, 0xEB00A3B4EB00A2B4, 0xEB00A5B4EB00A4B4,
-    0xEB00A7B4EB00A6B4, 0xEB00A9B4EB00A8B4, 0xEB00ABB4EB00AAB4,
-    0xEB00ADB4EB00ACB4, 0xEB00AFB4EB00AEB4, 0xEB00B1B4EB00B0B4,
-    0xEB00B3B4EB00B2B4, 0xEB00B5B4EB00B4B4, 0xEB00B7B4EB00B6B4,
-    0xEB00B9B4EB00B8B4, 0xEB00BBB4EB00BAB4, 0xEB00BDB4EB00BCB4,
-    0xEB00BFB4EB00BEB4, 0xEB0081B5EB0080B5, 0xEB0083B5EB0082B5,
-    0xEB0085B5EB0084B5, 0xEB0087B5EB0086B5, 0xEB0089B5EB0088B5,
-    0xEB008BB5EB008AB5, 0xEB008DB5EB008CB5, 0xEB008FB5EB008EB5,
-    0xEB0091B5EB0090B5, 0xEB0093B5EB0092B5, 0xEB0095B5EB0094B5,
-    0xEB0097B5EB0096B5, 0xEB0099B5EB0098B5, 0xEB009BB5EB009AB5,
-    0xEB009DB5EB009CB5, 0xEB009FB5EB009EB5, 0xEB00A1B5EB00A0B5,
-    0xEB00A3B5EB00A2B5, 0xEB00A5B5EB00A4B5, 0xEB00A7B5EB00A6B5,
-    0xEB00A9B5EB00A8B5, 0xEB00ABB5EB00AAB5, 0xEB00ADB5EB00ACB5,
-    0xEB00AFB5EB00AEB5, 0xEB00B1B5EB00B0B5, 0xEB00B3B5EB00B2B5,
-    0xEB00B5B5EB00B4B5, 0xEB00B7B5EB00B6B5, 0xEB00B9B5EB00B8B5,
-    0xEB00BBB5EB00BAB5, 0xEB00BDB5EB00BCB5, 0xEB00BFB5EB00BEB5,
-    0xEB0081B6EB0080B6, 0xEB0083B6EB0082B6, 0xEB0085B6EB0084B6,
-    0xEB0087B6EB0086B6, 0xEB0089B6EB0088B6, 0xEB008BB6EB008AB6,
-    0xEB008DB6EB008CB6, 0xEB008FB6EB008EB6, 0xEB0091B6EB0090B6,
-    0xEB0093B6EB0092B6, 0xEB0095B6EB0094B6, 0xEB0097B6EB0096B6,
-    0xEB0099B6EB0098B6, 0xEB009BB6EB009AB6, 0xEB009DB6EB009CB6,
-    0xEB009FB6EB009EB6, 0xEB00A1B6EB00A0B6, 0xEB00A3B6EB00A2B6,
-    0xEB00A5B6EB00A4B6, 0xEB00A7B6EB00A6B6, 0xEB00A9B6EB00A8B6,
-    0xEB00ABB6EB00AAB6, 0xEB00ADB6EB00ACB6, 0xEB00AFB6EB00AEB6,
-    0xEB00B1B6EB00B0B6, 0xEB00B3B6EB00B2B6, 0xEB00B5B6EB00B4B6,
-    0xEB00B7B6EB00B6B6, 0xEB00B9B6EB00B8B6, 0xEB00BBB6EB00BAB6,
-    0xEB00BDB6EB00BCB6, 0xEB00BFB6EB00BEB6, 0xEB0081B7EB0080B7,
-    0xEB0083B7EB0082B7, 0xEB0085B7EB0084B7, 0xEB0087B7EB0086B7,
-    0xEB0089B7EB0088B7, 0xEB008BB7EB008AB7, 0xEB008DB7EB008CB7,
-    0xEB008FB7EB008EB7, 0xEB0091B7EB0090B7, 0xEB0093B7EB0092B7,
-    0xEB0095B7EB0094B7, 0xEB0097B7EB0096B7, 0xEB0099B7EB0098B7,
-    0xEB009BB7EB009AB7, 0xEB009DB7EB009CB7, 0xEB009FB7EB009EB7,
-    0xEB00A1B7EB00A0B7, 0xEB00A3B7EB00A2B7, 0xEB00A5B7EB00A4B7,
-    0xEB00A7B7EB00A6B7, 0xEB00A9B7EB00A8B7, 0xEB00ABB7EB00AAB7,
-    0xEB00ADB7EB00ACB7, 0xEB00AFB7EB00AEB7, 0xEB00B1B7EB00B0B7,
-    0xEB00B3B7EB00B2B7, 0xEB00B5B7EB00B4B7, 0xEB00B7B7EB00B6B7,
-    0xEB00B9B7EB00B8B7, 0xEB00BBB7EB00BAB7, 0xEB00BDB7EB00BCB7,
-    0xEB00BFB7EB00BEB7, 0xEB0081B8EB0080B8, 0xEB0083B8EB0082B8,
-    0xEB0085B8EB0084B8, 0xEB0087B8EB0086B8, 0xEB0089B8EB0088B8,
-    0xEB008BB8EB008AB8, 0xEB008DB8EB008CB8, 0xEB008FB8EB008EB8,
-    0xEB0091B8EB0090B8, 0xEB0093B8EB0092B8, 0xEB0095B8EB0094B8,
-    0xEB0097B8EB0096B8, 0xEB0099B8EB0098B8, 0xEB009BB8EB009AB8,
-    0xEB009DB8EB009CB8, 0xEB009FB8EB009EB8, 0xEB00A1B8EB00A0B8,
-    0xEB00A3B8EB00A2B8, 0xEB00A5B8EB00A4B8, 0xEB00A7B8EB00A6B8,
-    0xEB00A9B8EB00A8B8, 0xEB00ABB8EB00AAB8, 0xEB00ADB8EB00ACB8,
-    0xEB00AFB8EB00AEB8, 0xEB00B1B8EB00B0B8, 0xEB00B3B8EB00B2B8,
-    0xEB00B5B8EB00B4B8, 0xEB00B7B8EB00B6B8, 0xEB00B9B8EB00B8B8,
-    0xEB00BBB8EB00BAB8, 0xEB00BDB8EB00BCB8, 0xEB00BFB8EB00BEB8,
-    0xEB0081B9EB0080B9, 0xEB0083B9EB0082B9, 0xEB0085B9EB0084B9,
-    0xEB0087B9EB0086B9, 0xEB0089B9EB0088B9, 0xEB008BB9EB008AB9,
-    0xEB008DB9EB008CB9, 0xEB008FB9EB008EB9, 0xEB0091B9EB0090B9,
-    0xEB0093B9EB0092B9, 0xEB0095B9EB0094B9, 0xEB0097B9EB0096B9,
-    0xEB0099B9EB0098B9, 0xEB009BB9EB009AB9, 0xEB009DB9EB009CB9,
-    0xEB009FB9EB009EB9, 0xEB00A1B9EB00A0B9, 0xEB00A3B9EB00A2B9,
-    0xEB00A5B9EB00A4B9, 0xEB00A7B9EB00A6B9, 0xEB00A9B9EB00A8B9,
-    0xEB00ABB9EB00AAB9, 0xEB00ADB9EB00ACB9, 0xEB00AFB9EB00AEB9,
-    0xEB00B1B9EB00B0B9, 0xEB00B3B9EB00B2B9, 0xEB00B5B9EB00B4B9,
-    0xEB00B7B9EB00B6B9, 0xEB00B9B9EB00B8B9, 0xEB00BBB9EB00BAB9,
-    0xEB00BDB9EB00BCB9, 0xEB00BFB9EB00BEB9, 0xEB0081BAEB0080BA,
-    0xEB0083BAEB0082BA, 0xEB0085BAEB0084BA, 0xEB0087BAEB0086BA,
-    0xEB0089BAEB0088BA, 0xEB008BBAEB008ABA, 0xEB008DBAEB008CBA,
-    0xEB008FBAEB008EBA, 0xEB0091BAEB0090BA, 0xEB0093BAEB0092BA,
-    0xEB0095BAEB0094BA, 0xEB0097BAEB0096BA, 0xEB0099BAEB0098BA,
-    0xEB009BBAEB009ABA, 0xEB009DBAEB009CBA, 0xEB009FBAEB009EBA,
-    0xEB00A1BAEB00A0BA, 0xEB00A3BAEB00A2BA, 0xEB00A5BAEB00A4BA,
-    0xEB00A7BAEB00A6BA, 0xEB00A9BAEB00A8BA, 0xEB00ABBAEB00AABA,
-    0xEB00ADBAEB00ACBA, 0xEB00AFBAEB00AEBA, 0xEB00B1BAEB00B0BA,
-    0xEB00B3BAEB00B2BA, 0xEB00B5BAEB00B4BA, 0xEB00B7BAEB00B6BA,
-    0xEB00B9BAEB00B8BA, 0xEB00BBBAEB00BABA, 0xEB00BDBAEB00BCBA,
-    0xEB00BFBAEB00BEBA, 0xEB0081BBEB0080BB, 0xEB0083BBEB0082BB,
-    0xEB0085BBEB0084BB, 0xEB0087BBEB0086BB, 0xEB0089BBEB0088BB,
-    0xEB008BBBEB008ABB, 0xEB008DBBEB008CBB, 0xEB008FBBEB008EBB,
-    0xEB0091BBEB0090BB, 0xEB0093BBEB0092BB, 0xEB0095BBEB0094BB,
-    0xEB0097BBEB0096BB, 0xEB0099BBEB0098BB, 0xEB009BBBEB009ABB,
-    0xEB009DBBEB009CBB, 0xEB009FBBEB009EBB, 0xEB00A1BBEB00A0BB,
-    0xEB00A3BBEB00A2BB, 0xEB00A5BBEB00A4BB, 0xEB00A7BBEB00A6BB,
-    0xEB00A9BBEB00A8BB, 0xEB00ABBBEB00AABB, 0xEB00ADBBEB00ACBB,
-    0xEB00AFBBEB00AEBB, 0xEB00B1BBEB00B0BB, 0xEB00B3BBEB00B2BB,
-    0xEB00B5BBEB00B4BB, 0xEB00B7BBEB00B6BB, 0xEB00B9BBEB00B8BB,
-    0xEB00BBBBEB00BABB, 0xEB00BDBBEB00BCBB, 0xEB00BFBBEB00BEBB,
-    0xEB0081BCEB0080BC, 0xEB0083BCEB0082BC, 0xEB0085BCEB0084BC,
-    0xEB0087BCEB0086BC, 0xEB0089BCEB0088BC, 0xEB008BBCEB008ABC,
-    0xEB008DBCEB008CBC, 0xEB008FBCEB008EBC, 0xEB0091BCEB0090BC,
-    0xEB0093BCEB0092BC, 0xEB0095BCEB0094BC, 0xEB0097BCEB0096BC,
-    0xEB0099BCEB0098BC, 0xEB009BBCEB009ABC, 0xEB009DBCEB009CBC,
-    0xEB009FBCEB009EBC, 0xEB00A1BCEB00A0BC, 0xEB00A3BCEB00A2BC,
-    0xEB00A5BCEB00A4BC, 0xEB00A7BCEB00A6BC, 0xEB00A9BCEB00A8BC,
-    0xEB00ABBCEB00AABC, 0xEB00ADBCEB00ACBC, 0xEB00AFBCEB00AEBC,
-    0xEB00B1BCEB00B0BC, 0xEB00B3BCEB00B2BC, 0xEB00B5BCEB00B4BC,
-    0xEB00B7BCEB00B6BC, 0xEB00B9BCEB00B8BC, 0xEB00BBBCEB00BABC,
-    0xEB00BDBCEB00BCBC, 0xEB00BFBCEB00BEBC, 0xEB0081BDEB0080BD,
-    0xEB0083BDEB0082BD, 0xEB0085BDEB0084BD, 0xEB0087BDEB0086BD,
-    0xEB0089BDEB0088BD, 0xEB008BBDEB008ABD, 0xEB008DBDEB008CBD,
-    0xEB008FBDEB008EBD, 0xEB0091BDEB0090BD, 0xEB0093BDEB0092BD,
-    0xEB0095BDEB0094BD, 0xEB0097BDEB0096BD, 0xEB0099BDEB0098BD,
-    0xEB009BBDEB009ABD, 0xEB009DBDEB009CBD, 0xEB009FBDEB009EBD,
-    0xEB00A1BDEB00A0BD, 0xEB00A3BDEB00A2BD, 0xEB00A5BDEB00A4BD,
-    0xEB00A7BDEB00A6BD, 0xEB00A9BDEB00A8BD, 0xEB00ABBDEB00AABD,
-    0xEB00ADBDEB00ACBD, 0xEB00AFBDEB00AEBD, 0xEB00B1BDEB00B0BD,
-    0xEB00B3BDEB00B2BD, 0xEB00B5BDEB00B4BD, 0xEB00B7BDEB00B6BD,
-    0xEB00B9BDEB00B8BD, 0xEB00BBBDEB00BABD, 0xEB00BDBDEB00BCBD,
-    0xEB00BFBDEB00BEBD, 0xEB0081BEEB0080BE, 0xEB0083BEEB0082BE,
-    0xEB0085BEEB0084BE, 0xEB0087BEEB0086BE, 0xEB0089BEEB0088BE,
-    0xEB008BBEEB008ABE, 0xEB008DBEEB008CBE, 0xEB008FBEEB008EBE,
-    0xEB0091BEEB0090BE, 0xEB0093BEEB0092BE, 0xEB0095BEEB0094BE,
-    0xEB0097BEEB0096BE, 0xEB0099BEEB0098BE, 0xEB009BBEEB009ABE,
-    0xEB009DBEEB009CBE, 0xEB009FBEEB009EBE, 0xEB00A1BEEB00A0BE,
-    0xEB00A3BEEB00A2BE, 0xEB00A5BEEB00A4BE, 0xEB00A7BEEB00A6BE,
-    0xEB00A9BEEB00A8BE, 0xEB00ABBEEB00AABE, 0xEB00ADBEEB00ACBE,
-    0xEB00AFBEEB00AEBE, 0xEB00B1BEEB00B0BE, 0xEB00B3BEEB00B2BE,
-    0xEB00B5BEEB00B4BE, 0xEB00B7BEEB00B6BE, 0xEB00B9BEEB00B8BE,
-    0xEB00BBBEEB00BABE, 0xEB00BDBEEB00BCBE, 0xEB00BFBEEB00BEBE,
-    0xEB0081BFEB0080BF, 0xEB0083BFEB0082BF, 0xEB0085BFEB0084BF,
-    0xEB0087BFEB0086BF, 0xEB0089BFEB0088BF, 0xEB008BBFEB008ABF,
-    0xEB008DBFEB008CBF, 0xEB008FBFEB008EBF, 0xEB0091BFEB0090BF,
-    0xEB0093BFEB0092BF, 0xEB0095BFEB0094BF, 0xEB0097BFEB0096BF,
-    0xEB0099BFEB0098BF, 0xEB009BBFEB009ABF, 0xEB009DBFEB009CBF,
-    0xEB009FBFEB009EBF, 0xEB00A1BFEB00A0BF, 0xEB00A3BFEB00A2BF,
-    0xEB00A5BFEB00A4BF, 0xEB00A7BFEB00A6BF, 0xEB00A9BFEB00A8BF,
-    0xEB00ABBFEB00AABF, 0xEB00ADBFEB00ACBF, 0xEB00AFBFEB00AEBF,
-    0xEB00B1BFEB00B0BF, 0xEB00B3BFEB00B2BF, 0xEB00B5BFEB00B4BF,
-    0xEB00B7BFEB00B6BF, 0xEB00B9BFEB00B8BF, 0xEB00BBBFEB00BABF,
-    0xEB00BDBFEB00BCBF, 0xEC00BFBFEB00BEBF, 0xEC008180EC008080,
-    0xEC008380EC008280, 0xEC008580EC008480, 0xEC008780EC008680,
-    0xEC008980EC008880, 0xEC008B80EC008A80, 0xEC008D80EC008C80,
-    0xEC008F80EC008E80, 0xEC009180EC009080, 0xEC009380EC009280,
-    0xEC009580EC009480, 0xEC009780EC009680, 0xEC009980EC009880,
-    0xEC009B80EC009A80, 0xEC009D80EC009C80, 0xEC009F80EC009E80,
-    0xEC00A180EC00A080, 0xEC00A380EC00A280, 0xEC00A580EC00A480,
-    0xEC00A780EC00A680, 0xEC00A980EC00A880, 0xEC00AB80EC00AA80,
-    0xEC00AD80EC00AC80, 0xEC00AF80EC00AE80, 0xEC00B180EC00B080,
-    0xEC00B380EC00B280, 0xEC00B580EC00B480, 0xEC00B780EC00B680,
-    0xEC00B980EC00B880, 0xEC00BB80EC00BA80, 0xEC00BD80EC00BC80,
-    0xEC00BF80EC00BE80, 0xEC008181EC008081, 0xEC008381EC008281,
-    0xEC008581EC008481, 0xEC008781EC008681, 0xEC008981EC008881,
-    0xEC008B81EC008A81, 0xEC008D81EC008C81, 0xEC008F81EC008E81,
-    0xEC009181EC009081, 0xEC009381EC009281, 0xEC009581EC009481,
-    0xEC009781EC009681, 0xEC009981EC009881, 0xEC009B81EC009A81,
-    0xEC009D81EC009C81, 0xEC009F81EC009E81, 0xEC00A181EC00A081,
-    0xEC00A381EC00A281, 0xEC00A581EC00A481, 0xEC00A781EC00A681,
-    0xEC00A981EC00A881, 0xEC00AB81EC00AA81, 0xEC00AD81EC00AC81,
-    0xEC00AF81EC00AE81, 0xEC00B181EC00B081, 0xEC00B381EC00B281,
-    0xEC00B581EC00B481, 0xEC00B781EC00B681, 0xEC00B981EC00B881,
-    0xEC00BB81EC00BA81, 0xEC00BD81EC00BC81, 0xEC00BF81EC00BE81,
-    0xEC008182EC008082, 0xEC008382EC008282, 0xEC008582EC008482,
-    0xEC008782EC008682, 0xEC008982EC008882, 0xEC008B82EC008A82,
-    0xEC008D82EC008C82, 0xEC008F82EC008E82, 0xEC009182EC009082,
-    0xEC009382EC009282, 0xEC009582EC009482, 0xEC009782EC009682,
-    0xEC009982EC009882, 0xEC009B82EC009A82, 0xEC009D82EC009C82,
-    0xEC009F82EC009E82, 0xEC00A182EC00A082, 0xEC00A382EC00A282,
-    0xEC00A582EC00A482, 0xEC00A782EC00A682, 0xEC00A982EC00A882,
-    0xEC00AB82EC00AA82, 0xEC00AD82EC00AC82, 0xEC00AF82EC00AE82,
-    0xEC00B182EC00B082, 0xEC00B382EC00B282, 0xEC00B582EC00B482,
-    0xEC00B782EC00B682, 0xEC00B982EC00B882, 0xEC00BB82EC00BA82,
-    0xEC00BD82EC00BC82, 0xEC00BF82EC00BE82, 0xEC008183EC008083,
-    0xEC008383EC008283, 0xEC008583EC008483, 0xEC008783EC008683,
-    0xEC008983EC008883, 0xEC008B83EC008A83, 0xEC008D83EC008C83,
-    0xEC008F83EC008E83, 0xEC009183EC009083, 0xEC009383EC009283,
-    0xEC009583EC009483, 0xEC009783EC009683, 0xEC009983EC009883,
-    0xEC009B83EC009A83, 0xEC009D83EC009C83, 0xEC009F83EC009E83,
-    0xEC00A183EC00A083, 0xEC00A383EC00A283, 0xEC00A583EC00A483,
-    0xEC00A783EC00A683, 0xEC00A983EC00A883, 0xEC00AB83EC00AA83,
-    0xEC00AD83EC00AC83, 0xEC00AF83EC00AE83, 0xEC00B183EC00B083,
-    0xEC00B383EC00B283, 0xEC00B583EC00B483, 0xEC00B783EC00B683,
-    0xEC00B983EC00B883, 0xEC00BB83EC00BA83, 0xEC00BD83EC00BC83,
-    0xEC00BF83EC00BE83, 0xEC008184EC008084, 0xEC008384EC008284,
-    0xEC008584EC008484, 0xEC008784EC008684, 0xEC008984EC008884,
-    0xEC008B84EC008A84, 0xEC008D84EC008C84, 0xEC008F84EC008E84,
-    0xEC009184EC009084, 0xEC009384EC009284, 0xEC009584EC009484,
-    0xEC009784EC009684, 0xEC009984EC009884, 0xEC009B84EC009A84,
-    0xEC009D84EC009C84, 0xEC009F84EC009E84, 0xEC00A184EC00A084,
-    0xEC00A384EC00A284, 0xEC00A584EC00A484, 0xEC00A784EC00A684,
-    0xEC00A984EC00A884, 0xEC00AB84EC00AA84, 0xEC00AD84EC00AC84,
-    0xEC00AF84EC00AE84, 0xEC00B184EC00B084, 0xEC00B384EC00B284,
-    0xEC00B584EC00B484, 0xEC00B784EC00B684, 0xEC00B984EC00B884,
-    0xEC00BB84EC00BA84, 0xEC00BD84EC00BC84, 0xEC00BF84EC00BE84,
-    0xEC008185EC008085, 0xEC008385EC008285, 0xEC008585EC008485,
-    0xEC008785EC008685, 0xEC008985EC008885, 0xEC008B85EC008A85,
-    0xEC008D85EC008C85, 0xEC008F85EC008E85, 0xEC009185EC009085,
-    0xEC009385EC009285, 0xEC009585EC009485, 0xEC009785EC009685,
-    0xEC009985EC009885, 0xEC009B85EC009A85, 0xEC009D85EC009C85,
-    0xEC009F85EC009E85, 0xEC00A185EC00A085, 0xEC00A385EC00A285,
-    0xEC00A585EC00A485, 0xEC00A785EC00A685, 0xEC00A985EC00A885,
-    0xEC00AB85EC00AA85, 0xEC00AD85EC00AC85, 0xEC00AF85EC00AE85,
-    0xEC00B185EC00B085, 0xEC00B385EC00B285, 0xEC00B585EC00B485,
-    0xEC00B785EC00B685, 0xEC00B985EC00B885, 0xEC00BB85EC00BA85,
-    0xEC00BD85EC00BC85, 0xEC00BF85EC00BE85, 0xEC008186EC008086,
-    0xEC008386EC008286, 0xEC008586EC008486, 0xEC008786EC008686,
-    0xEC008986EC008886, 0xEC008B86EC008A86, 0xEC008D86EC008C86,
-    0xEC008F86EC008E86, 0xEC009186EC009086, 0xEC009386EC009286,
-    0xEC009586EC009486, 0xEC009786EC009686, 0xEC009986EC009886,
-    0xEC009B86EC009A86, 0xEC009D86EC009C86, 0xEC009F86EC009E86,
-    0xEC00A186EC00A086, 0xEC00A386EC00A286, 0xEC00A586EC00A486,
-    0xEC00A786EC00A686, 0xEC00A986EC00A886, 0xEC00AB86EC00AA86,
-    0xEC00AD86EC00AC86, 0xEC00AF86EC00AE86, 0xEC00B186EC00B086,
-    0xEC00B386EC00B286, 0xEC00B586EC00B486, 0xEC00B786EC00B686,
-    0xEC00B986EC00B886, 0xEC00BB86EC00BA86, 0xEC00BD86EC00BC86,
-    0xEC00BF86EC00BE86, 0xEC008187EC008087, 0xEC008387EC008287,
-    0xEC008587EC008487, 0xEC008787EC008687, 0xEC008987EC008887,
-    0xEC008B87EC008A87, 0xEC008D87EC008C87, 0xEC008F87EC008E87,
-    0xEC009187EC009087, 0xEC009387EC009287, 0xEC009587EC009487,
-    0xEC009787EC009687, 0xEC009987EC009887, 0xEC009B87EC009A87,
-    0xEC009D87EC009C87, 0xEC009F87EC009E87, 0xEC00A187EC00A087,
-    0xEC00A387EC00A287, 0xEC00A587EC00A487, 0xEC00A787EC00A687,
-    0xEC00A987EC00A887, 0xEC00AB87EC00AA87, 0xEC00AD87EC00AC87,
-    0xEC00AF87EC00AE87, 0xEC00B187EC00B087, 0xEC00B387EC00B287,
-    0xEC00B587EC00B487, 0xEC00B787EC00B687, 0xEC00B987EC00B887,
-    0xEC00BB87EC00BA87, 0xEC00BD87EC00BC87, 0xEC00BF87EC00BE87,
-    0xEC008188EC008088, 0xEC008388EC008288, 0xEC008588EC008488,
-    0xEC008788EC008688, 0xEC008988EC008888, 0xEC008B88EC008A88,
-    0xEC008D88EC008C88, 0xEC008F88EC008E88, 0xEC009188EC009088,
-    0xEC009388EC009288, 0xEC009588EC009488, 0xEC009788EC009688,
-    0xEC009988EC009888, 0xEC009B88EC009A88, 0xEC009D88EC009C88,
-    0xEC009F88EC009E88, 0xEC00A188EC00A088, 0xEC00A388EC00A288,
-    0xEC00A588EC00A488, 0xEC00A788EC00A688, 0xEC00A988EC00A888,
-    0xEC00AB88EC00AA88, 0xEC00AD88EC00AC88, 0xEC00AF88EC00AE88,
-    0xEC00B188EC00B088, 0xEC00B388EC00B288, 0xEC00B588EC00B488,
-    0xEC00B788EC00B688, 0xEC00B988EC00B888, 0xEC00BB88EC00BA88,
-    0xEC00BD88EC00BC88, 0xEC00BF88EC00BE88, 0xEC008189EC008089,
-    0xEC008389EC008289, 0xEC008589EC008489, 0xEC008789EC008689,
-    0xEC008989EC008889, 0xEC008B89EC008A89, 0xEC008D89EC008C89,
-    0xEC008F89EC008E89, 0xEC009189EC009089, 0xEC009389EC009289,
-    0xEC009589EC009489, 0xEC009789EC009689, 0xEC009989EC009889,
-    0xEC009B89EC009A89, 0xEC009D89EC009C89, 0xEC009F89EC009E89,
-    0xEC00A189EC00A089, 0xEC00A389EC00A289, 0xEC00A589EC00A489,
-    0xEC00A789EC00A689, 0xEC00A989EC00A889, 0xEC00AB89EC00AA89,
-    0xEC00AD89EC00AC89, 0xEC00AF89EC00AE89, 0xEC00B189EC00B089,
-    0xEC00B389EC00B289, 0xEC00B589EC00B489, 0xEC00B789EC00B689,
-    0xEC00B989EC00B889, 0xEC00BB89EC00BA89, 0xEC00BD89EC00BC89,
-    0xEC00BF89EC00BE89, 0xEC00818AEC00808A, 0xEC00838AEC00828A,
-    0xEC00858AEC00848A, 0xEC00878AEC00868A, 0xEC00898AEC00888A,
-    0xEC008B8AEC008A8A, 0xEC008D8AEC008C8A, 0xEC008F8AEC008E8A,
-    0xEC00918AEC00908A, 0xEC00938AEC00928A, 0xEC00958AEC00948A,
-    0xEC00978AEC00968A, 0xEC00998AEC00988A, 0xEC009B8AEC009A8A,
-    0xEC009D8AEC009C8A, 0xEC009F8AEC009E8A, 0xEC00A18AEC00A08A,
-    0xEC00A38AEC00A28A, 0xEC00A58AEC00A48A, 0xEC00A78AEC00A68A,
-    0xEC00A98AEC00A88A, 0xEC00AB8AEC00AA8A, 0xEC00AD8AEC00AC8A,
-    0xEC00AF8AEC00AE8A, 0xEC00B18AEC00B08A, 0xEC00B38AEC00B28A,
-    0xEC00B58AEC00B48A, 0xEC00B78AEC00B68A, 0xEC00B98AEC00B88A,
-    0xEC00BB8AEC00BA8A, 0xEC00BD8AEC00BC8A, 0xEC00BF8AEC00BE8A,
-    0xEC00818BEC00808B, 0xEC00838BEC00828B, 0xEC00858BEC00848B,
-    0xEC00878BEC00868B, 0xEC00898BEC00888B, 0xEC008B8BEC008A8B,
-    0xEC008D8BEC008C8B, 0xEC008F8BEC008E8B, 0xEC00918BEC00908B,
-    0xEC00938BEC00928B, 0xEC00958BEC00948B, 0xEC00978BEC00968B,
-    0xEC00998BEC00988B, 0xEC009B8BEC009A8B, 0xEC009D8BEC009C8B,
-    0xEC009F8BEC009E8B, 0xEC00A18BEC00A08B, 0xEC00A38BEC00A28B,
-    0xEC00A58BEC00A48B, 0xEC00A78BEC00A68B, 0xEC00A98BEC00A88B,
-    0xEC00AB8BEC00AA8B, 0xEC00AD8BEC00AC8B, 0xEC00AF8BEC00AE8B,
-    0xEC00B18BEC00B08B, 0xEC00B38BEC00B28B, 0xEC00B58BEC00B48B,
-    0xEC00B78BEC00B68B, 0xEC00B98BEC00B88B, 0xEC00BB8BEC00BA8B,
-    0xEC00BD8BEC00BC8B, 0xEC00BF8BEC00BE8B, 0xEC00818CEC00808C,
-    0xEC00838CEC00828C, 0xEC00858CEC00848C, 0xEC00878CEC00868C,
-    0xEC00898CEC00888C, 0xEC008B8CEC008A8C, 0xEC008D8CEC008C8C,
-    0xEC008F8CEC008E8C, 0xEC00918CEC00908C, 0xEC00938CEC00928C,
-    0xEC00958CEC00948C, 0xEC00978CEC00968C, 0xEC00998CEC00988C,
-    0xEC009B8CEC009A8C, 0xEC009D8CEC009C8C, 0xEC009F8CEC009E8C,
-    0xEC00A18CEC00A08C, 0xEC00A38CEC00A28C, 0xEC00A58CEC00A48C,
-    0xEC00A78CEC00A68C, 0xEC00A98CEC00A88C, 0xEC00AB8CEC00AA8C,
-    0xEC00AD8CEC00AC8C, 0xEC00AF8CEC00AE8C, 0xEC00B18CEC00B08C,
-    0xEC00B38CEC00B28C, 0xEC00B58CEC00B48C, 0xEC00B78CEC00B68C,
-    0xEC00B98CEC00B88C, 0xEC00BB8CEC00BA8C, 0xEC00BD8CEC00BC8C,
-    0xEC00BF8CEC00BE8C, 0xEC00818DEC00808D, 0xEC00838DEC00828D,
-    0xEC00858DEC00848D, 0xEC00878DEC00868D, 0xEC00898DEC00888D,
-    0xEC008B8DEC008A8D, 0xEC008D8DEC008C8D, 0xEC008F8DEC008E8D,
-    0xEC00918DEC00908D, 0xEC00938DEC00928D, 0xEC00958DEC00948D,
-    0xEC00978DEC00968D, 0xEC00998DEC00988D, 0xEC009B8DEC009A8D,
-    0xEC009D8DEC009C8D, 0xEC009F8DEC009E8D, 0xEC00A18DEC00A08D,
-    0xEC00A38DEC00A28D, 0xEC00A58DEC00A48D, 0xEC00A78DEC00A68D,
-    0xEC00A98DEC00A88D, 0xEC00AB8DEC00AA8D, 0xEC00AD8DEC00AC8D,
-    0xEC00AF8DEC00AE8D, 0xEC00B18DEC00B08D, 0xEC00B38DEC00B28D,
-    0xEC00B58DEC00B48D, 0xEC00B78DEC00B68D, 0xEC00B98DEC00B88D,
-    0xEC00BB8DEC00BA8D, 0xEC00BD8DEC00BC8D, 0xEC00BF8DEC00BE8D,
-    0xEC00818EEC00808E, 0xEC00838EEC00828E, 0xEC00858EEC00848E,
-    0xEC00878EEC00868E, 0xEC00898EEC00888E, 0xEC008B8EEC008A8E,
-    0xEC008D8EEC008C8E, 0xEC008F8EEC008E8E, 0xEC00918EEC00908E,
-    0xEC00938EEC00928E, 0xEC00958EEC00948E, 0xEC00978EEC00968E,
-    0xEC00998EEC00988E, 0xEC009B8EEC009A8E, 0xEC009D8EEC009C8E,
-    0xEC009F8EEC009E8E, 0xEC00A18EEC00A08E, 0xEC00A38EEC00A28E,
-    0xEC00A58EEC00A48E, 0xEC00A78EEC00A68E, 0xEC00A98EEC00A88E,
-    0xEC00AB8EEC00AA8E, 0xEC00AD8EEC00AC8E, 0xEC00AF8EEC00AE8E,
-    0xEC00B18EEC00B08E, 0xEC00B38EEC00B28E, 0xEC00B58EEC00B48E,
-    0xEC00B78EEC00B68E, 0xEC00B98EEC00B88E, 0xEC00BB8EEC00BA8E,
-    0xEC00BD8EEC00BC8E, 0xEC00BF8EEC00BE8E, 0xEC00818FEC00808F,
-    0xEC00838FEC00828F, 0xEC00858FEC00848F, 0xEC00878FEC00868F,
-    0xEC00898FEC00888F, 0xEC008B8FEC008A8F, 0xEC008D8FEC008C8F,
-    0xEC008F8FEC008E8F, 0xEC00918FEC00908F, 0xEC00938FEC00928F,
-    0xEC00958FEC00948F, 0xEC00978FEC00968F, 0xEC00998FEC00988F,
-    0xEC009B8FEC009A8F, 0xEC009D8FEC009C8F, 0xEC009F8FEC009E8F,
-    0xEC00A18FEC00A08F, 0xEC00A38FEC00A28F, 0xEC00A58FEC00A48F,
-    0xEC00A78FEC00A68F, 0xEC00A98FEC00A88F, 0xEC00AB8FEC00AA8F,
-    0xEC00AD8FEC00AC8F, 0xEC00AF8FEC00AE8F, 0xEC00B18FEC00B08F,
-    0xEC00B38FEC00B28F, 0xEC00B58FEC00B48F, 0xEC00B78FEC00B68F,
-    0xEC00B98FEC00B88F, 0xEC00BB8FEC00BA8F, 0xEC00BD8FEC00BC8F,
-    0xEC00BF8FEC00BE8F, 0xEC008190EC008090, 0xEC008390EC008290,
-    0xEC008590EC008490, 0xEC008790EC008690, 0xEC008990EC008890,
-    0xEC008B90EC008A90, 0xEC008D90EC008C90, 0xEC008F90EC008E90,
-    0xEC009190EC009090, 0xEC009390EC009290, 0xEC009590EC009490,
-    0xEC009790EC009690, 0xEC009990EC009890, 0xEC009B90EC009A90,
-    0xEC009D90EC009C90, 0xEC009F90EC009E90, 0xEC00A190EC00A090,
-    0xEC00A390EC00A290, 0xEC00A590EC00A490, 0xEC00A790EC00A690,
-    0xEC00A990EC00A890, 0xEC00AB90EC00AA90, 0xEC00AD90EC00AC90,
-    0xEC00AF90EC00AE90, 0xEC00B190EC00B090, 0xEC00B390EC00B290,
-    0xEC00B590EC00B490, 0xEC00B790EC00B690, 0xEC00B990EC00B890,
-    0xEC00BB90EC00BA90, 0xEC00BD90EC00BC90, 0xEC00BF90EC00BE90,
-    0xEC008191EC008091, 0xEC008391EC008291, 0xEC008591EC008491,
-    0xEC008791EC008691, 0xEC008991EC008891, 0xEC008B91EC008A91,
-    0xEC008D91EC008C91, 0xEC008F91EC008E91, 0xEC009191EC009091,
-    0xEC009391EC009291, 0xEC009591EC009491, 0xEC009791EC009691,
-    0xEC009991EC009891, 0xEC009B91EC009A91, 0xEC009D91EC009C91,
-    0xEC009F91EC009E91, 0xEC00A191EC00A091, 0xEC00A391EC00A291,
-    0xEC00A591EC00A491, 0xEC00A791EC00A691, 0xEC00A991EC00A891,
-    0xEC00AB91EC00AA91, 0xEC00AD91EC00AC91, 0xEC00AF91EC00AE91,
-    0xEC00B191EC00B091, 0xEC00B391EC00B291, 0xEC00B591EC00B491,
-    0xEC00B791EC00B691, 0xEC00B991EC00B891, 0xEC00BB91EC00BA91,
-    0xEC00BD91EC00BC91, 0xEC00BF91EC00BE91, 0xEC008192EC008092,
-    0xEC008392EC008292, 0xEC008592EC008492, 0xEC008792EC008692,
-    0xEC008992EC008892, 0xEC008B92EC008A92, 0xEC008D92EC008C92,
-    0xEC008F92EC008E92, 0xEC009192EC009092, 0xEC009392EC009292,
-    0xEC009592EC009492, 0xEC009792EC009692, 0xEC009992EC009892,
-    0xEC009B92EC009A92, 0xEC009D92EC009C92, 0xEC009F92EC009E92,
-    0xEC00A192EC00A092, 0xEC00A392EC00A292, 0xEC00A592EC00A492,
-    0xEC00A792EC00A692, 0xEC00A992EC00A892, 0xEC00AB92EC00AA92,
-    0xEC00AD92EC00AC92, 0xEC00AF92EC00AE92, 0xEC00B192EC00B092,
-    0xEC00B392EC00B292, 0xEC00B592EC00B492, 0xEC00B792EC00B692,
-    0xEC00B992EC00B892, 0xEC00BB92EC00BA92, 0xEC00BD92EC00BC92,
-    0xEC00BF92EC00BE92, 0xEC008193EC008093, 0xEC008393EC008293,
-    0xEC008593EC008493, 0xEC008793EC008693, 0xEC008993EC008893,
-    0xEC008B93EC008A93, 0xEC008D93EC008C93, 0xEC008F93EC008E93,
-    0xEC009193EC009093, 0xEC009393EC009293, 0xEC009593EC009493,
-    0xEC009793EC009693, 0xEC009993EC009893, 0xEC009B93EC009A93,
-    0xEC009D93EC009C93, 0xEC009F93EC009E93, 0xEC00A193EC00A093,
-    0xEC00A393EC00A293, 0xEC00A593EC00A493, 0xEC00A793EC00A693,
-    0xEC00A993EC00A893, 0xEC00AB93EC00AA93, 0xEC00AD93EC00AC93,
-    0xEC00AF93EC00AE93, 0xEC00B193EC00B093, 0xEC00B393EC00B293,
-    0xEC00B593EC00B493, 0xEC00B793EC00B693, 0xEC00B993EC00B893,
-    0xEC00BB93EC00BA93, 0xEC00BD93EC00BC93, 0xEC00BF93EC00BE93,
-    0xEC008194EC008094, 0xEC008394EC008294, 0xEC008594EC008494,
-    0xEC008794EC008694, 0xEC008994EC008894, 0xEC008B94EC008A94,
-    0xEC008D94EC008C94, 0xEC008F94EC008E94, 0xEC009194EC009094,
-    0xEC009394EC009294, 0xEC009594EC009494, 0xEC009794EC009694,
-    0xEC009994EC009894, 0xEC009B94EC009A94, 0xEC009D94EC009C94,
-    0xEC009F94EC009E94, 0xEC00A194EC00A094, 0xEC00A394EC00A294,
-    0xEC00A594EC00A494, 0xEC00A794EC00A694, 0xEC00A994EC00A894,
-    0xEC00AB94EC00AA94, 0xEC00AD94EC00AC94, 0xEC00AF94EC00AE94,
-    0xEC00B194EC00B094, 0xEC00B394EC00B294, 0xEC00B594EC00B494,
-    0xEC00B794EC00B694, 0xEC00B994EC00B894, 0xEC00BB94EC00BA94,
-    0xEC00BD94EC00BC94, 0xEC00BF94EC00BE94, 0xEC008195EC008095,
-    0xEC008395EC008295, 0xEC008595EC008495, 0xEC008795EC008695,
-    0xEC008995EC008895, 0xEC008B95EC008A95, 0xEC008D95EC008C95,
-    0xEC008F95EC008E95, 0xEC009195EC009095, 0xEC009395EC009295,
-    0xEC009595EC009495, 0xEC009795EC009695, 0xEC009995EC009895,
-    0xEC009B95EC009A95, 0xEC009D95EC009C95, 0xEC009F95EC009E95,
-    0xEC00A195EC00A095, 0xEC00A395EC00A295, 0xEC00A595EC00A495,
-    0xEC00A795EC00A695, 0xEC00A995EC00A895, 0xEC00AB95EC00AA95,
-    0xEC00AD95EC00AC95, 0xEC00AF95EC00AE95, 0xEC00B195EC00B095,
-    0xEC00B395EC00B295, 0xEC00B595EC00B495, 0xEC00B795EC00B695,
-    0xEC00B995EC00B895, 0xEC00BB95EC00BA95, 0xEC00BD95EC00BC95,
-    0xEC00BF95EC00BE95, 0xEC008196EC008096, 0xEC008396EC008296,
-    0xEC008596EC008496, 0xEC008796EC008696, 0xEC008996EC008896,
-    0xEC008B96EC008A96, 0xEC008D96EC008C96, 0xEC008F96EC008E96,
-    0xEC009196EC009096, 0xEC009396EC009296, 0xEC009596EC009496,
-    0xEC009796EC009696, 0xEC009996EC009896, 0xEC009B96EC009A96,
-    0xEC009D96EC009C96, 0xEC009F96EC009E96, 0xEC00A196EC00A096,
-    0xEC00A396EC00A296, 0xEC00A596EC00A496, 0xEC00A796EC00A696,
-    0xEC00A996EC00A896, 0xEC00AB96EC00AA96, 0xEC00AD96EC00AC96,
-    0xEC00AF96EC00AE96, 0xEC00B196EC00B096, 0xEC00B396EC00B296,
-    0xEC00B596EC00B496, 0xEC00B796EC00B696, 0xEC00B996EC00B896,
-    0xEC00BB96EC00BA96, 0xEC00BD96EC00BC96, 0xEC00BF96EC00BE96,
-    0xEC008197EC008097, 0xEC008397EC008297, 0xEC008597EC008497,
-    0xEC008797EC008697, 0xEC008997EC008897, 0xEC008B97EC008A97,
-    0xEC008D97EC008C97, 0xEC008F97EC008E97, 0xEC009197EC009097,
-    0xEC009397EC009297, 0xEC009597EC009497, 0xEC009797EC009697,
-    0xEC009997EC009897, 0xEC009B97EC009A97, 0xEC009D97EC009C97,
-    0xEC009F97EC009E97, 0xEC00A197EC00A097, 0xEC00A397EC00A297,
-    0xEC00A597EC00A497, 0xEC00A797EC00A697, 0xEC00A997EC00A897,
-    0xEC00AB97EC00AA97, 0xEC00AD97EC00AC97, 0xEC00AF97EC00AE97,
-    0xEC00B197EC00B097, 0xEC00B397EC00B297, 0xEC00B597EC00B497,
-    0xEC00B797EC00B697, 0xEC00B997EC00B897, 0xEC00BB97EC00BA97,
-    0xEC00BD97EC00BC97, 0xEC00BF97EC00BE97, 0xEC008198EC008098,
-    0xEC008398EC008298, 0xEC008598EC008498, 0xEC008798EC008698,
-    0xEC008998EC008898, 0xEC008B98EC008A98, 0xEC008D98EC008C98,
-    0xEC008F98EC008E98, 0xEC009198EC009098, 0xEC009398EC009298,
-    0xEC009598EC009498, 0xEC009798EC009698, 0xEC009998EC009898,
-    0xEC009B98EC009A98, 0xEC009D98EC009C98, 0xEC009F98EC009E98,
-    0xEC00A198EC00A098, 0xEC00A398EC00A298, 0xEC00A598EC00A498,
-    0xEC00A798EC00A698, 0xEC00A998EC00A898, 0xEC00AB98EC00AA98,
-    0xEC00AD98EC00AC98, 0xEC00AF98EC00AE98, 0xEC00B198EC00B098,
-    0xEC00B398EC00B298, 0xEC00B598EC00B498, 0xEC00B798EC00B698,
-    0xEC00B998EC00B898, 0xEC00BB98EC00BA98, 0xEC00BD98EC00BC98,
-    0xEC00BF98EC00BE98, 0xEC008199EC008099, 0xEC008399EC008299,
-    0xEC008599EC008499, 0xEC008799EC008699, 0xEC008999EC008899,
-    0xEC008B99EC008A99, 0xEC008D99EC008C99, 0xEC008F99EC008E99,
-    0xEC009199EC009099, 0xEC009399EC009299, 0xEC009599EC009499,
-    0xEC009799EC009699, 0xEC009999EC009899, 0xEC009B99EC009A99,
-    0xEC009D99EC009C99, 0xEC009F99EC009E99, 0xEC00A199EC00A099,
-    0xEC00A399EC00A299, 0xEC00A599EC00A499, 0xEC00A799EC00A699,
-    0xEC00A999EC00A899, 0xEC00AB99EC00AA99, 0xEC00AD99EC00AC99,
-    0xEC00AF99EC00AE99, 0xEC00B199EC00B099, 0xEC00B399EC00B299,
-    0xEC00B599EC00B499, 0xEC00B799EC00B699, 0xEC00B999EC00B899,
-    0xEC00BB99EC00BA99, 0xEC00BD99EC00BC99, 0xEC00BF99EC00BE99,
-    0xEC00819AEC00809A, 0xEC00839AEC00829A, 0xEC00859AEC00849A,
-    0xEC00879AEC00869A, 0xEC00899AEC00889A, 0xEC008B9AEC008A9A,
-    0xEC008D9AEC008C9A, 0xEC008F9AEC008E9A, 0xEC00919AEC00909A,
-    0xEC00939AEC00929A, 0xEC00959AEC00949A, 0xEC00979AEC00969A,
-    0xEC00999AEC00989A, 0xEC009B9AEC009A9A, 0xEC009D9AEC009C9A,
-    0xEC009F9AEC009E9A, 0xEC00A19AEC00A09A, 0xEC00A39AEC00A29A,
-    0xEC00A59AEC00A49A, 0xEC00A79AEC00A69A, 0xEC00A99AEC00A89A,
-    0xEC00AB9AEC00AA9A, 0xEC00AD9AEC00AC9A, 0xEC00AF9AEC00AE9A,
-    0xEC00B19AEC00B09A, 0xEC00B39AEC00B29A, 0xEC00B59AEC00B49A,
-    0xEC00B79AEC00B69A, 0xEC00B99AEC00B89A, 0xEC00BB9AEC00BA9A,
-    0xEC00BD9AEC00BC9A, 0xEC00BF9AEC00BE9A, 0xEC00819BEC00809B,
-    0xEC00839BEC00829B, 0xEC00859BEC00849B, 0xEC00879BEC00869B,
-    0xEC00899BEC00889B, 0xEC008B9BEC008A9B, 0xEC008D9BEC008C9B,
-    0xEC008F9BEC008E9B, 0xEC00919BEC00909B, 0xEC00939BEC00929B,
-    0xEC00959BEC00949B, 0xEC00979BEC00969B, 0xEC00999BEC00989B,
-    0xEC009B9BEC009A9B, 0xEC009D9BEC009C9B, 0xEC009F9BEC009E9B,
-    0xEC00A19BEC00A09B, 0xEC00A39BEC00A29B, 0xEC00A59BEC00A49B,
-    0xEC00A79BEC00A69B, 0xEC00A99BEC00A89B, 0xEC00AB9BEC00AA9B,
-    0xEC00AD9BEC00AC9B, 0xEC00AF9BEC00AE9B, 0xEC00B19BEC00B09B,
-    0xEC00B39BEC00B29B, 0xEC00B59BEC00B49B, 0xEC00B79BEC00B69B,
-    0xEC00B99BEC00B89B, 0xEC00BB9BEC00BA9B, 0xEC00BD9BEC00BC9B,
-    0xEC00BF9BEC00BE9B, 0xEC00819CEC00809C, 0xEC00839CEC00829C,
-    0xEC00859CEC00849C, 0xEC00879CEC00869C, 0xEC00899CEC00889C,
-    0xEC008B9CEC008A9C, 0xEC008D9CEC008C9C, 0xEC008F9CEC008E9C,
-    0xEC00919CEC00909C, 0xEC00939CEC00929C, 0xEC00959CEC00949C,
-    0xEC00979CEC00969C, 0xEC00999CEC00989C, 0xEC009B9CEC009A9C,
-    0xEC009D9CEC009C9C, 0xEC009F9CEC009E9C, 0xEC00A19CEC00A09C,
-    0xEC00A39CEC00A29C, 0xEC00A59CEC00A49C, 0xEC00A79CEC00A69C,
-    0xEC00A99CEC00A89C, 0xEC00AB9CEC00AA9C, 0xEC00AD9CEC00AC9C,
-    0xEC00AF9CEC00AE9C, 0xEC00B19CEC00B09C, 0xEC00B39CEC00B29C,
-    0xEC00B59CEC00B49C, 0xEC00B79CEC00B69C, 0xEC00B99CEC00B89C,
-    0xEC00BB9CEC00BA9C, 0xEC00BD9CEC00BC9C, 0xEC00BF9CEC00BE9C,
-    0xEC00819DEC00809D, 0xEC00839DEC00829D, 0xEC00859DEC00849D,
-    0xEC00879DEC00869D, 0xEC00899DEC00889D, 0xEC008B9DEC008A9D,
-    0xEC008D9DEC008C9D, 0xEC008F9DEC008E9D, 0xEC00919DEC00909D,
-    0xEC00939DEC00929D, 0xEC00959DEC00949D, 0xEC00979DEC00969D,
-    0xEC00999DEC00989D, 0xEC009B9DEC009A9D, 0xEC009D9DEC009C9D,
-    0xEC009F9DEC009E9D, 0xEC00A19DEC00A09D, 0xEC00A39DEC00A29D,
-    0xEC00A59DEC00A49D, 0xEC00A79DEC00A69D, 0xEC00A99DEC00A89D,
-    0xEC00AB9DEC00AA9D, 0xEC00AD9DEC00AC9D, 0xEC00AF9DEC00AE9D,
-    0xEC00B19DEC00B09D, 0xEC00B39DEC00B29D, 0xEC00B59DEC00B49D,
-    0xEC00B79DEC00B69D, 0xEC00B99DEC00B89D, 0xEC00BB9DEC00BA9D,
-    0xEC00BD9DEC00BC9D, 0xEC00BF9DEC00BE9D, 0xEC00819EEC00809E,
-    0xEC00839EEC00829E, 0xEC00859EEC00849E, 0xEC00879EEC00869E,
-    0xEC00899EEC00889E, 0xEC008B9EEC008A9E, 0xEC008D9EEC008C9E,
-    0xEC008F9EEC008E9E, 0xEC00919EEC00909E, 0xEC00939EEC00929E,
-    0xEC00959EEC00949E, 0xEC00979EEC00969E, 0xEC00999EEC00989E,
-    0xEC009B9EEC009A9E, 0xEC009D9EEC009C9E, 0xEC009F9EEC009E9E,
-    0xEC00A19EEC00A09E, 0xEC00A39EEC00A29E, 0xEC00A59EEC00A49E,
-    0xEC00A79EEC00A69E, 0xEC00A99EEC00A89E, 0xEC00AB9EEC00AA9E,
-    0xEC00AD9EEC00AC9E, 0xEC00AF9EEC00AE9E, 0xEC00B19EEC00B09E,
-    0xEC00B39EEC00B29E, 0xEC00B59EEC00B49E, 0xEC00B79EEC00B69E,
-    0xEC00B99EEC00B89E, 0xEC00BB9EEC00BA9E, 0xEC00BD9EEC00BC9E,
-    0xEC00BF9EEC00BE9E, 0xEC00819FEC00809F, 0xEC00839FEC00829F,
-    0xEC00859FEC00849F, 0xEC00879FEC00869F, 0xEC00899FEC00889F,
-    0xEC008B9FEC008A9F, 0xEC008D9FEC008C9F, 0xEC008F9FEC008E9F,
-    0xEC00919FEC00909F, 0xEC00939FEC00929F, 0xEC00959FEC00949F,
-    0xEC00979FEC00969F, 0xEC00999FEC00989F, 0xEC009B9FEC009A9F,
-    0xEC009D9FEC009C9F, 0xEC009F9FEC009E9F, 0xEC00A19FEC00A09F,
-    0xEC00A39FEC00A29F, 0xEC00A59FEC00A49F, 0xEC00A79FEC00A69F,
-    0xEC00A99FEC00A89F, 0xEC00AB9FEC00AA9F, 0xEC00AD9FEC00AC9F,
-    0xEC00AF9FEC00AE9F, 0xEC00B19FEC00B09F, 0xEC00B39FEC00B29F,
-    0xEC00B59FEC00B49F, 0xEC00B79FEC00B69F, 0xEC00B99FEC00B89F,
-    0xEC00BB9FEC00BA9F, 0xEC00BD9FEC00BC9F, 0xEC00BF9FEC00BE9F,
-    0xEC0081A0EC0080A0, 0xEC0083A0EC0082A0, 0xEC0085A0EC0084A0,
-    0xEC0087A0EC0086A0, 0xEC0089A0EC0088A0, 0xEC008BA0EC008AA0,
-    0xEC008DA0EC008CA0, 0xEC008FA0EC008EA0, 0xEC0091A0EC0090A0,
-    0xEC0093A0EC0092A0, 0xEC0095A0EC0094A0, 0xEC0097A0EC0096A0,
-    0xEC0099A0EC0098A0, 0xEC009BA0EC009AA0, 0xEC009DA0EC009CA0,
-    0xEC009FA0EC009EA0, 0xEC00A1A0EC00A0A0, 0xEC00A3A0EC00A2A0,
-    0xEC00A5A0EC00A4A0, 0xEC00A7A0EC00A6A0, 0xEC00A9A0EC00A8A0,
-    0xEC00ABA0EC00AAA0, 0xEC00ADA0EC00ACA0, 0xEC00AFA0EC00AEA0,
-    0xEC00B1A0EC00B0A0, 0xEC00B3A0EC00B2A0, 0xEC00B5A0EC00B4A0,
-    0xEC00B7A0EC00B6A0, 0xEC00B9A0EC00B8A0, 0xEC00BBA0EC00BAA0,
-    0xEC00BDA0EC00BCA0, 0xEC00BFA0EC00BEA0, 0xEC0081A1EC0080A1,
-    0xEC0083A1EC0082A1, 0xEC0085A1EC0084A1, 0xEC0087A1EC0086A1,
-    0xEC0089A1EC0088A1, 0xEC008BA1EC008AA1, 0xEC008DA1EC008CA1,
-    0xEC008FA1EC008EA1, 0xEC0091A1EC0090A1, 0xEC0093A1EC0092A1,
-    0xEC0095A1EC0094A1, 0xEC0097A1EC0096A1, 0xEC0099A1EC0098A1,
-    0xEC009BA1EC009AA1, 0xEC009DA1EC009CA1, 0xEC009FA1EC009EA1,
-    0xEC00A1A1EC00A0A1, 0xEC00A3A1EC00A2A1, 0xEC00A5A1EC00A4A1,
-    0xEC00A7A1EC00A6A1, 0xEC00A9A1EC00A8A1, 0xEC00ABA1EC00AAA1,
-    0xEC00ADA1EC00ACA1, 0xEC00AFA1EC00AEA1, 0xEC00B1A1EC00B0A1,
-    0xEC00B3A1EC00B2A1, 0xEC00B5A1EC00B4A1, 0xEC00B7A1EC00B6A1,
-    0xEC00B9A1EC00B8A1, 0xEC00BBA1EC00BAA1, 0xEC00BDA1EC00BCA1,
-    0xEC00BFA1EC00BEA1, 0xEC0081A2EC0080A2, 0xEC0083A2EC0082A2,
-    0xEC0085A2EC0084A2, 0xEC0087A2EC0086A2, 0xEC0089A2EC0088A2,
-    0xEC008BA2EC008AA2, 0xEC008DA2EC008CA2, 0xEC008FA2EC008EA2,
-    0xEC0091A2EC0090A2, 0xEC0093A2EC0092A2, 0xEC0095A2EC0094A2,
-    0xEC0097A2EC0096A2, 0xEC0099A2EC0098A2, 0xEC009BA2EC009AA2,
-    0xEC009DA2EC009CA2, 0xEC009FA2EC009EA2, 0xEC00A1A2EC00A0A2,
-    0xEC00A3A2EC00A2A2, 0xEC00A5A2EC00A4A2, 0xEC00A7A2EC00A6A2,
-    0xEC00A9A2EC00A8A2, 0xEC00ABA2EC00AAA2, 0xEC00ADA2EC00ACA2,
-    0xEC00AFA2EC00AEA2, 0xEC00B1A2EC00B0A2, 0xEC00B3A2EC00B2A2,
-    0xEC00B5A2EC00B4A2, 0xEC00B7A2EC00B6A2, 0xEC00B9A2EC00B8A2,
-    0xEC00BBA2EC00BAA2, 0xEC00BDA2EC00BCA2, 0xEC00BFA2EC00BEA2,
-    0xEC0081A3EC0080A3, 0xEC0083A3EC0082A3, 0xEC0085A3EC0084A3,
-    0xEC0087A3EC0086A3, 0xEC0089A3EC0088A3, 0xEC008BA3EC008AA3,
-    0xEC008DA3EC008CA3, 0xEC008FA3EC008EA3, 0xEC0091A3EC0090A3,
-    0xEC0093A3EC0092A3, 0xEC0095A3EC0094A3, 0xEC0097A3EC0096A3,
-    0xEC0099A3EC0098A3, 0xEC009BA3EC009AA3, 0xEC009DA3EC009CA3,
-    0xEC009FA3EC009EA3, 0xEC00A1A3EC00A0A3, 0xEC00A3A3EC00A2A3,
-    0xEC00A5A3EC00A4A3, 0xEC00A7A3EC00A6A3, 0xEC00A9A3EC00A8A3,
-    0xEC00ABA3EC00AAA3, 0xEC00ADA3EC00ACA3, 0xEC00AFA3EC00AEA3,
-    0xEC00B1A3EC00B0A3, 0xEC00B3A3EC00B2A3, 0xEC00B5A3EC00B4A3,
-    0xEC00B7A3EC00B6A3, 0xEC00B9A3EC00B8A3, 0xEC00BBA3EC00BAA3,
-    0x9DECBCA3EC00BCA3, 0xA3EC00BDA3EC0098, 0xA4EC00BFA3EC00BE,
-    0xA4EC0081A4EC0080, 0xA4EC0083A4EC0082, 0xA4EC0085A4EC0084,
-    0xA4EC0087A4EC0086, 0xA4EC0089A4EC0088, 0xA4EC008BA4EC008A,
-    0xA4EC008DA4EC008C, 0xA4EC008FA4EC008E, 0xA4EC0091A4EC0090,
-    0xA4EC0093A4EC0092, 0xA4EC0095A4EC0094, 0xA4EC0097A4EC0096,
-    0xA4EC0099A4EC0098, 0xA4EC009BA4EC009A, 0xA4EC009DA4EC009C,
-    0xA4EC009FA4EC009E, 0xA4EC00A1A4EC00A0, 0xA4EC00A3A4EC00A2,
-    0xA4EC00A5A4EC00A4, 0xA4EC00A7A4EC00A6, 0xA4EC00A9A4EC00A8,
-    0xA4EC00ABA4EC00AA, 0xA4EC00ADA4EC00AC, 0xA4EC00AFA4EC00AE,
-    0xA4EC00B1A4EC00B0, 0xA4EC00B3A4EC00B2, 0xA4EC00B5A4EC00B4,
-    0xA4EC00B7A4EC00B6, 0xA4EC00B9A4EC00B8, 0xA4EC00BBA4EC00BA,
-    0xA4EC00BDA4EC00BC, 0xA5EC00BFA4EC00BE, 0xA5EC0081A5EC0080,
-    0xA5EC0083A5EC0082, 0xA5EC0085A5EC0084, 0xA5EC0087A5EC0086,
-    0xA5EC0089A5EC0088, 0xA5EC008BA5EC008A, 0xA5EC008DA5EC008C,
-    0xA5EC008FA5EC008E, 0xA5EC0091A5EC0090, 0xA5EC0093A5EC0092,
-    0xA5EC0095A5EC0094, 0xA5EC0097A5EC0096, 0xA5EC0099A5EC0098,
-    0xA5EC009BA5EC009A, 0xA5EC009DA5EC009C, 0xA5EC009FA5EC009E,
-    0xA5EC00A1A5EC00A0, 0xA5EC00A3A5EC00A2, 0xA5EC00A5A5EC00A4,
-    0xA5EC00A7A5EC00A6, 0xA5EC00A9A5EC00A8, 0xA5EC00ABA5EC00AA,
-    0xA5EC00ADA5EC00AC, 0xA5EC00AFA5EC00AE, 0xA5EC00B1A5EC00B0,
-    0xA5EC00B3A5EC00B2, 0xA5EC00B5A5EC00B4, 0xA5EC00B7A5EC00B6,
-    0xA5EC00B9A5EC00B8, 0xA5EC00BBA5EC00BA, 0xA5EC00BDA5EC00BC,
-    0xA6EC00BFA5EC00BE, 0xA6EC0081A6EC0080, 0xA6EC0083A6EC0082,
-    0xA6EC0085A6EC0084, 0xA6EC0087A6EC0086, 0xA6EC0089A6EC0088,
-    0xA6EC008BA6EC008A, 0xA6EC008DA6EC008C, 0xA6EC008FA6EC008E,
-    0xA6EC0091A6EC0090, 0xA6EC0093A6EC0092, 0xA6EC0095A6EC0094,
-    0xA6EC0097A6EC0096, 0xA6EC0099A6EC0098, 0xA6EC009BA6EC009A,
-    0xA6EC009DA6EC009C, 0xA6EC009FA6EC009E, 0xA6EC00A1A6EC00A0,
-    0xA6EC00A3A6EC00A2, 0xA6EC00A5A6EC00A4, 0xA6EC00A7A6EC00A6,
-    0xA6EC00A9A6EC00A8, 0xA6EC00ABA6EC00AA, 0xA6EC00ADA6EC00AC,
-    0xA6EC00AFA6EC00AE, 0xA6EC00B1A6EC00B0, 0xA6EC00B3A6EC00B2,
-    0xA6EC00B5A6EC00B4, 0xA6EC00B7A6EC00B6, 0xA6EC00B9A6EC00B8,
-    0xA6EC00BBA6EC00BA, 0xA6EC00BDA6EC00BC, 0xA7EC00BFA6EC00BE,
-    0xA7EC0081A7EC0080, 0xA7EC0083A7EC0082, 0xA7EC0085A7EC0084,
-    0xA7EC0087A7EC0086, 0xA7EC0089A7EC0088, 0xA7EC008BA7EC008A,
-    0xA7EC008DA7EC008C, 0xA7EC008FA7EC008E, 0xA7EC0091A7EC0090,
-    0xA7EC0093A7EC0092, 0xA7EC0095A7EC0094, 0xA7EC0097A7EC0096,
-    0xA7EC0099A7EC0098, 0xA7EC009BA7EC009A, 0xA7EC009DA7EC009C,
-    0xA7EC009FA7EC009E, 0xA7EC00A1A7EC00A0, 0xA7EC00A3A7EC00A2,
-    0xA7EC00A5A7EC00A4, 0xA7EC00A7A7EC00A6, 0xA7EC00A9A7EC00A8,
-    0xA7EC00ABA7EC00AA, 0xA7EC00ADA7EC00AC, 0xA7EC00AFA7EC00AE,
-    0xA7EC00B1A7EC00B0, 0xA7EC00B3A7EC00B2, 0xA7EC00B5A7EC00B4,
-    0xA7EC00B7A7EC00B6, 0xA7EC00B9A7EC00B8, 0xA7EC00BBA7EC00BA,
-    0xA7EC00BDA7EC00BC, 0xA8EC00BFA7EC00BE, 0xA8EC0081A8EC0080,
-    0xA8EC0083A8EC0082, 0xA8EC0085A8EC0084, 0xA8EC0087A8EC0086,
-    0xA8EC0089A8EC0088, 0xA8EC008BA8EC008A, 0xA8EC008DA8EC008C,
-    0xA8EC008FA8EC008E, 0xA8EC0091A8EC0090, 0xA8EC0093A8EC0092,
-    0xA8EC0095A8EC0094, 0xA8EC0097A8EC0096, 0xA8EC0099A8EC0098,
-    0xA8EC009BA8EC009A, 0xA8EC009DA8EC009C, 0xA8EC009FA8EC009E,
-    0xA8EC00A1A8EC00A0, 0xA8EC00A3A8EC00A2, 0xA8EC00A5A8EC00A4,
-    0xA8EC00A7A8EC00A6, 0xA8EC00A9A8EC00A8, 0xA8EC00ABA8EC00AA,
-    0xA8EC00ADA8EC00AC, 0xA8EC00AFA8EC00AE, 0xA8EC00B1A8EC00B0,
-    0xA8EC00B3A8EC00B2, 0xA8EC00B5A8EC00B4, 0xA8EC00B7A8EC00B6,
-    0xA8EC00B9A8EC00B8, 0xA8EC00BBA8EC00BA, 0xA8EC00BDA8EC00BC,
-    0xA9EC00BFA8EC00BE, 0xA9EC0081A9EC0080, 0xA9EC0083A9EC0082,
-    0xA9EC0085A9EC0084, 0xA9EC0087A9EC0086, 0xA9EC0089A9EC0088,
-    0xA9EC008BA9EC008A, 0xA9EC008DA9EC008C, 0xA9EC008FA9EC008E,
-    0xA9EC0091A9EC0090, 0xA9EC0093A9EC0092, 0xA9EC0095A9EC0094,
-    0xA9EC0097A9EC0096, 0xA9EC0099A9EC0098, 0xA9EC009BA9EC009A,
-    0xA9EC009DA9EC009C, 0xA9EC009FA9EC009E, 0xA9EC00A1A9EC00A0,
-    0xA9EC00A3A9EC00A2, 0xA9EC00A5A9EC00A4, 0xA9EC00A7A9EC00A6,
-    0xA9EC00A9A9EC00A8, 0xA9EC00ABA9EC00AA, 0xA9EC00ADA9EC00AC,
-    0xA9EC00AFA9EC00AE, 0xA9EC00B1A9EC00B0, 0xA9EC00B3A9EC00B2,
-    0xA9EC00B5A9EC00B4, 0xA9EC00B7A9EC00B6, 0xA9EC00B9A9EC00B8,
-    0xA9EC00BBA9EC00BA, 0xA9EC00BDA9EC00BC, 0xAAEC00BFA9EC00BE,
-    0xAAEC0081AAEC0080, 0xAAEC0083AAEC0082, 0xAAEC0085AAEC0084,
-    0xAAEC0087AAEC0086, 0xAAEC0089AAEC0088, 0xAAEC008BAAEC008A,
-    0xAAEC008DAAEC008C, 0xAAEC008FAAEC008E, 0xAAEC0091AAEC0090,
-    0xAAEC0093AAEC0092, 0xAAEC0095AAEC0094, 0xAAEC0097AAEC0096,
-    0xAAEC0099AAEC0098, 0xAAEC009BAAEC009A, 0xAAEC009DAAEC009C,
-    0xAAEC009FAAEC009E, 0xAAEC00A1AAEC00A0, 0xAAEC00A3AAEC00A2,
-    0xAAEC00A5AAEC00A4, 0xAAEC00A7AAEC00A6, 0xAAEC00A9AAEC00A8,
-    0xAAEC00ABAAEC00AA, 0xAAEC00ADAAEC00AC, 0xAAEC00AFAAEC00AE,
-    0xAAEC00B1AAEC00B0, 0xAAEC00B3AAEC00B2, 0xAAEC00B5AAEC00B4,
-    0xAAEC00B7AAEC00B6, 0xAAEC00B9AAEC00B8, 0xAAEC00BBAAEC00BA,
-    0xAAEC00BDAAEC00BC, 0xABEC00BFAAEC00BE, 0xABEC0081ABEC0080,
-    0xABEC0083ABEC0082, 0xABEC0085ABEC0084, 0xABEC0087ABEC0086,
-    0xABEC0089ABEC0088, 0xABEC008BABEC008A, 0xABEC008DABEC008C,
-    0xABEC008FABEC008E, 0xABEC0091ABEC0090, 0xABEC0093ABEC0092,
-    0xABEC0095ABEC0094, 0xABEC0097ABEC0096, 0xABEC0099ABEC0098,
-    0xABEC009BABEC009A, 0xABEC009DABEC009C, 0xABEC009FABEC009E,
-    0xABEC00A1ABEC00A0, 0xABEC00A3ABEC00A2, 0xABEC00A5ABEC00A4,
-    0xABEC00A7ABEC00A6, 0xABEC00A9ABEC00A8, 0xABEC00ABABEC00AA,
-    0xABEC00ADABEC00AC, 0xABEC00AFABEC00AE, 0xABEC00B1ABEC00B0,
-    0xABEC00B3ABEC00B2, 0xABEC00B5ABEC00B4, 0xABEC00B7ABEC00B6,
-    0xABEC00B9ABEC00B8, 0xABEC00BBABEC00BA, 0xABEC00BDABEC00BC,
-    0xACEC00BFABEC00BE, 0xACEC0081ACEC0080, 0xACEC0083ACEC0082,
-    0xACEC0085ACEC0084, 0xACEC0087ACEC0086, 0xACEC0089ACEC0088,
-    0xACEC008BACEC008A, 0xACEC008DACEC008C, 0xACEC008FACEC008E,
-    0xACEC0091ACEC0090, 0xACEC0093ACEC0092, 0xACEC0095ACEC0094,
-    0xACEC0097ACEC0096, 0xACEC0099ACEC0098, 0xACEC009BACEC009A,
-    0xACEC009DACEC009C, 0xACEC009FACEC009E, 0xACEC00A1ACEC00A0,
-    0xACEC00A3ACEC00A2, 0xACEC00A5ACEC00A4, 0xACEC00A7ACEC00A6,
-    0xACEC00A9ACEC00A8, 0xACEC00ABACEC00AA, 0xACEC00ADACEC00AC,
-    0xACEC00AFACEC00AE, 0xACEC00B1ACEC00B0, 0xACEC00B3ACEC00B2,
-    0xACEC00B5ACEC00B4, 0xACEC00B7ACEC00B6, 0xACEC00B9ACEC00B8,
-    0xACEC00BBACEC00BA, 0xACEC00BDACEC00BC, 0xADEC00BFACEC00BE,
-    0xADEC0081ADEC0080, 0xADEC0083ADEC0082, 0xADEC0085ADEC0084,
-    0xADEC0087ADEC0086, 0xADEC0089ADEC0088, 0xADEC008BADEC008A,
-    0xADEC008DADEC008C, 0xADEC008FADEC008E, 0xADEC0091ADEC0090,
-    0xADEC0093ADEC0092, 0xADEC0095ADEC0094, 0xADEC0097ADEC0096,
-    0xADEC0099ADEC0098, 0xADEC009BADEC009A, 0xADEC009DADEC009C,
-    0xADEC009FADEC009E, 0xADEC00A1ADEC00A0, 0xADEC00A3ADEC00A2,
-    0xADEC00A5ADEC00A4, 0xADEC00A7ADEC00A6, 0xADEC00A9ADEC00A8,
-    0xADEC00ABADEC00AA, 0xADEC00ADADEC00AC, 0xADEC00AFADEC00AE,
-    0xADEC00B1ADEC00B0, 0xADEC00B3ADEC00B2, 0xADEC00B5ADEC00B4,
-    0xADEC00B7ADEC00B6, 0xADEC00B9ADEC00B8, 0xADEC00BBADEC00BA,
-    0xADEC00BDADEC00BC, 0xAEEC00BFADEC00BE, 0xAEEC0081AEEC0080,
-    0xAEEC0083AEEC0082, 0xAEEC0085AEEC0084, 0xAEEC0087AEEC0086,
-    0xAEEC0089AEEC0088, 0xAEEC008BAEEC008A, 0xAEEC008DAEEC008C,
-    0xAEEC008FAEEC008E, 0xAEEC0091AEEC0090, 0xAEEC0093AEEC0092,
-    0xAEEC0095AEEC0094, 0xAEEC0097AEEC0096, 0xAEEC0099AEEC0098,
-    0xAEEC009BAEEC009A, 0xAEEC009DAEEC009C, 0xAEEC009FAEEC009E,
-    0xAEEC00A1AEEC00A0, 0xAEEC00A3AEEC00A2, 0xAEEC00A5AEEC00A4,
-    0xAEEC00A7AEEC00A6, 0xAEEC00A9AEEC00A8, 0xAEEC00ABAEEC00AA,
-    0xAEEC00ADAEEC00AC, 0xAEEC00AFAEEC00AE, 0xAEEC00B1AEEC00B0,
-    0xAEEC00B3AEEC00B2, 0xAEEC00B5AEEC00B4, 0xAEEC00B7AEEC00B6,
-    0xAEEC00B9AEEC00B8, 0xAEEC00BBAEEC00BA, 0xAEEC00BDAEEC00BC,
-    0xAFEC00BFAEEC00BE, 0xAFEC0081AFEC0080, 0xAFEC0083AFEC0082,
-    0xAFEC0085AFEC0084, 0xAFEC0087AFEC0086, 0xAFEC0089AFEC0088,
-    0xAFEC008BAFEC008A, 0xAFEC008DAFEC008C, 0xAFEC008FAFEC008E,
-    0xAFEC0091AFEC0090, 0xAFEC0093AFEC0092, 0xAFEC0095AFEC0094,
-    0xAFEC0097AFEC0096, 0xAFEC0099AFEC0098, 0xAFEC009BAFEC009A,
-    0xAFEC009DAFEC009C, 0xAFEC009FAFEC009E, 0xAFEC00A1AFEC00A0,
-    0xAFEC00A3AFEC00A2, 0xAFEC00A5AFEC00A4, 0xAFEC00A7AFEC00A6,
-    0xAFEC00A9AFEC00A8, 0xAFEC00ABAFEC00AA, 0xAFEC00ADAFEC00AC,
-    0xAFEC00AFAFEC00AE, 0xAFEC00B1AFEC00B0, 0xAFEC00B3AFEC00B2,
-    0xAFEC00B5AFEC00B4, 0xAFEC00B7AFEC00B6, 0xAFEC00B9AFEC00B8,
-    0xAFEC00BBAFEC00BA, 0xAFEC00BDAFEC00BC, 0xB0EC00BFAFEC00BE,
-    0xB0EC0081B0EC0080, 0xB0EC0083B0EC0082, 0xB0EC0085B0EC0084,
-    0xB0EC0087B0EC0086, 0xB0EC0089B0EC0088, 0xB0EC008BB0EC008A,
-    0xB0EC008DB0EC008C, 0xB0EC008FB0EC008E, 0xB0EC0091B0EC0090,
-    0xB0EC0093B0EC0092, 0xB0EC0095B0EC0094, 0xB0EC0097B0EC0096,
-    0xB0EC0099B0EC0098, 0xB0EC009BB0EC009A, 0xB0EC009DB0EC009C,
-    0xB0EC009FB0EC009E, 0xB0EC00A1B0EC00A0, 0xB0EC00A3B0EC00A2,
-    0xB0EC00A5B0EC00A4, 0xB0EC00A7B0EC00A6, 0xB0EC00A9B0EC00A8,
-    0xB0EC00ABB0EC00AA, 0xB0EC00ADB0EC00AC, 0xB0EC00AFB0EC00AE,
-    0xB0EC00B1B0EC00B0, 0xB0EC00B3B0EC00B2, 0xB0EC00B5B0EC00B4,
-    0xB0EC00B7B0EC00B6, 0xA0B3EAB8B0EC00B8, 0xBAB0EC00B9B0EC00,
-    0xBCB0EC00BBB0EC00, 0xBEB0EC00BDB0EC00, 0x80B1EC00BFB0EC00,
-    0x82B1EC0081B1EC00, 0x84B1EC0083B1EC00, 0x86B1EC0085B1EC00,
-    0x88B1EC0087B1EC00, 0x8AB1EC0089B1EC00, 0x8CB1EC008BB1EC00,
-    0x8EB1EC008DB1EC00, 0x90B1EC008FB1EC00, 0x92B1EC0091B1EC00,
-    0x94B1EC0093B1EC00, 0x96B1EC0095B1EC00, 0x98B1EC0097B1EC00,
-    0x9AB1EC0099B1EC00, 0x9CB1EC009BB1EC00, 0x9EB1EC009DB1EC00,
-    0xA0B1EC009FB1EC00, 0xA2B1EC00A1B1EC00, 0xA4B1EC00A3B1EC00,
-    0xA6B1EC00A5B1EC00, 0xA8B1EC00A7B1EC00, 0xAAB1EC00A9B1EC00,
-    0xACB1EC00ABB1EC00, 0xAEB1EC00ADB1EC00, 0xB0B1EC00AFB1EC00,
-    0xB2B1EC00B1B1EC00, 0xB4B1EC00B3B1EC00, 0xB6B1EC00B5B1EC00,
-    0xB8B1EC00B7B1EC00, 0xBAB1EC00B9B1EC00, 0xBCB1EC00BBB1EC00,
-    0xBEB1EC00BDB1EC00, 0x80B2EC00BFB1EC00, 0x82B2EC0081B2EC00,
-    0x84B2EC0083B2EC00, 0x86B2EC0085B2EC00, 0x88B2EC0087B2EC00,
-    0x8AB2EC0089B2EC00, 0x8CB2EC008BB2EC00, 0x8EB2EC008DB2EC00,
-    0x90B2EC008FB2EC00, 0x92B2EC0091B2EC00, 0x94B2EC0093B2EC00,
-    0x96B2EC0095B2EC00, 0x98B2EC0097B2EC00, 0x9AB2EC0099B2EC00,
-    0x9CB2EC009BB2EC00, 0x9EB2EC009DB2EC00, 0xA0B2EC009FB2EC00,
-    0xA2B2EC00A1B2EC00, 0xA4B2EC00A3B2EC00, 0xA6B2EC00A5B2EC00,
-    0xA8B2EC00A7B2EC00, 0xAAB2EC00A9B2EC00, 0xACB2EC00ABB2EC00,
-    0xAEB2EC00ADB2EC00, 0xB0B2EC00AFB2EC00, 0xB2B2EC00B1B2EC00,
-    0xB4B2EC00B3B2EC00, 0xB6B2EC00B5B2EC00, 0xB8B2EC00B7B2EC00,
-    0xBAB2EC00B9B2EC00, 0xBCB2EC00BBB2EC00, 0xBEB2EC00BDB2EC00,
-    0x80B3EC00BFB2EC00, 0x82B3EC0081B3EC00, 0x84B3EC0083B3EC00,
-    0x86B3EC0085B3EC00, 0x88B3EC0087B3EC00, 0x8AB3EC0089B3EC00,
-    0x8CB3EC008BB3EC00, 0x8EB3EC008DB3EC00, 0x90B3EC008FB3EC00,
-    0x92B3EC0091B3EC00, 0x94B3EC0093B3EC00, 0x96B3EC0095B3EC00,
-    0x98B3EC0097B3EC00, 0x9AB3EC0099B3EC00, 0x9CB3EC009BB3EC00,
-    0x9EB3EC009DB3EC00, 0xA0B3EC009FB3EC00, 0xA2B3EC00A1B3EC00,
-    0xA4B3EC00A3B3EC00, 0xA6B3EC00A5B3EC00, 0xA8B3EC00A7B3EC00,
-    0xAAB3EC00A9B3EC00, 0xACB3EC00ABB3EC00, 0xAEB3EC00ADB3EC00,
-    0xB0B3EC00AFB3EC00, 0xB2B3EC00B1B3EC00, 0xB4B3EC00B3B3EC00,
-    0xB6B3EC00B5B3EC00, 0xB8B3EC00B7B3EC00, 0xBAB3EC00B9B3EC00,
-    0xBCB3EC00BBB3EC00, 0xBEB3EC00BDB3EC00, 0x80B4EC00BFB3EC00,
-    0x82B4EC0081B4EC00, 0x84B4EC0083B4EC00, 0x86B4EC0085B4EC00,
-    0x88B4EC0087B4EC00, 0x8AB4EC0089B4EC00, 0x8CB4EC008BB4EC00,
-    0x8EB4EC008DB4EC00, 0x90B4EC008FB4EC00, 0x92B4EC0091B4EC00,
-    0x94B4EC0093B4EC00, 0x96B4EC0095B4EC00, 0x98B4EC0097B4EC00,
-    0x9AB4EC0099B4EC00, 0x9CB4EC009BB4EC00, 0x9EB4EC009DB4EC00,
-    0xA0B4EC009FB4EC00, 0xA2B4EC00A1B4EC00, 0xA4B4EC00A3B4EC00,
-    0xA6B4EC00A5B4EC00, 0xA8B4EC00A7B4EC00, 0xAAB4EC00A9B4EC00,
-    0xACB4EC00ABB4EC00, 0xAEB4EC00ADB4EC00, 0xB0B4EC00AFB4EC00,
-    0xB2B4EC00B1B4EC00, 0xB4B4EC00B3B4EC00, 0xB6B4EC00B5B4EC00,
-    0xB8B4EC00B7B4EC00, 0xBAB4EC00B9B4EC00, 0xBCB4EC00BBB4EC00,
-    0xBEB4EC00BDB4EC00, 0x80B5EC00BFB4EC00, 0x82B5EC0081B5EC00,
-    0x84B5EC0083B5EC00, 0x86B5EC0085B5EC00, 0x88B5EC0087B5EC00,
-    0x8AB5EC0089B5EC00, 0x8CB5EC008BB5EC00, 0x8EB5EC008DB5EC00,
-    0x90B5EC008FB5EC00, 0x92B5EC0091B5EC00, 0x94B5EC0093B5EC00,
-    0x96B5EC0095B5EC00, 0x98B5EC0097B5EC00, 0x9AB5EC0099B5EC00,
-    0x9CB5EC009BB5EC00, 0x9EB5EC009DB5EC00, 0xA0B5EC009FB5EC00,
-    0xA2B5EC00A1B5EC00, 0xA4B5EC00A3B5EC00, 0xA6B5EC00A5B5EC00,
-    0xA8B5EC00A7B5EC00, 0xAAB5EC00A9B5EC00, 0xACB5EC00ABB5EC00,
-    0xAEB5EC00ADB5EC00, 0xB0B5EC00AFB5EC00, 0xB2B5EC00B1B5EC00,
-    0xB4B5EC00B3B5EC00, 0xB6B5EC00B5B5EC00, 0xB8B5EC00B7B5EC00,
-    0xBAB5EC00B9B5EC00, 0xBCB5EC00BBB5EC00, 0xBEB5EC00BDB5EC00,
-    0x80B6EC00BFB5EC00, 0x82B6EC0081B6EC00, 0x84B6EC0083B6EC00,
-    0x86B6EC0085B6EC00, 0x88B6EC0087B6EC00, 0x8AB6EC0089B6EC00,
-    0x8CB6EC008BB6EC00, 0x8EB6EC008DB6EC00, 0x90B6EC008FB6EC00,
-    0x92B6EC0091B6EC00, 0x94B6EC0093B6EC00, 0x96B6EC0095B6EC00,
-    0x98B6EC0097B6EC00, 0x9AB6EC0099B6EC00, 0x9CB6EC009BB6EC00,
-    0x9EB6EC009DB6EC00, 0xA0B6EC009FB6EC00, 0xA2B6EC00A1B6EC00,
-    0xA4B6EC00A3B6EC00, 0xA6B6EC00A5B6EC00, 0xA8B6EC00A7B6EC00,
-    0xAAB6EC00A9B6EC00, 0xACB6EC00ABB6EC00, 0xAEB6EC00ADB6EC00,
-    0xB0B6EC00AFB6EC00, 0xB2B6EC00B1B6EC00, 0xB4B6EC00B3B6EC00,
-    0xB6B6EC00B5B6EC00, 0xB8B6EC00B7B6EC00, 0xBAB6EC00B9B6EC00,
-    0xBCB6EC00BBB6EC00, 0xBEB6EC00BDB6EC00, 0x80B7EC00BFB6EC00,
-    0x82B7EC0081B7EC00, 0x84B7EC0083B7EC00, 0x86B7EC0085B7EC00,
-    0x88B7EC0087B7EC00, 0x8AB7EC0089B7EC00, 0x8CB7EC008BB7EC00,
-    0x8EB7EC008DB7EC00, 0x90B7EC008FB7EC00, 0x92B7EC0091B7EC00,
-    0x94B7EC0093B7EC00, 0x96B7EC0095B7EC00, 0x98B7EC0097B7EC00,
-    0x9AB7EC0099B7EC00, 0x9CB7EC009BB7EC00, 0x9EB7EC009DB7EC00,
-    0xA0B7EC009FB7EC00, 0xA2B7EC00A1B7EC00, 0xA4B7EC00A3B7EC00,
-    0xA6B7EC00A5B7EC00, 0xA8B7EC00A7B7EC00, 0xAAB7EC00A9B7EC00,
-    0xACB7EC00ABB7EC00, 0xAEB7EC00ADB7EC00, 0xB0B7EC00AFB7EC00,
-    0xB2B7EC00B1B7EC00, 0xB4B7EC00B3B7EC00, 0xB6B7EC00B5B7EC00,
-    0xB8B7EC00B7B7EC00, 0xBAB7EC00B9B7EC00, 0xBCB7EC00BBB7EC00,
-    0xBEB7EC00BDB7EC00, 0x80B8EC00BFB7EC00, 0x82B8EC0081B8EC00,
-    0x84B8EC0083B8EC00, 0x86B8EC0085B8EC00, 0x88B8EC0087B8EC00,
-    0x8AB8EC0089B8EC00, 0x8CB8EC008BB8EC00, 0x8EB8EC008DB8EC00,
-    0x90B8EC008FB8EC00, 0x92B8EC0091B8EC00, 0x94B8EC0093B8EC00,
-    0x96B8EC0095B8EC00, 0x98B8EC0097B8EC00, 0x9AB8EC0099B8EC00,
-    0x9CB8EC009BB8EC00, 0x9EB8EC009DB8EC00, 0xA0B8EC009FB8EC00,
-    0xA2B8EC00A1B8EC00, 0xA4B8EC00A3B8EC00, 0xA6B8EC00A5B8EC00,
-    0xA8B8EC00A7B8EC00, 0xAAB8EC00A9B8EC00, 0xACB8EC00ABB8EC00,
-    0xAEB8EC00ADB8EC00, 0xB0B8EC00AFB8EC00, 0xB2B8EC00B1B8EC00,
-    0xB4B8EC00B3B8EC00, 0xB6B8EC00B5B8EC00, 0xB8B8EC00B7B8EC00,
-    0xBAB8EC00B9B8EC00, 0xBCB8EC00BBB8EC00, 0xBEB8EC00BDB8EC00,
-    0x80B9EC00BFB8EC00, 0x82B9EC0081B9EC00, 0x84B9EC0083B9EC00,
-    0x86B9EC0085B9EC00, 0x88B9EC0087B9EC00, 0x8AB9EC0089B9EC00,
-    0x8CB9EC008BB9EC00, 0x8EB9EC008DB9EC00, 0x90B9EC008FB9EC00,
-    0x92B9EC0091B9EC00, 0x94B9EC0093B9EC00, 0x96B9EC0095B9EC00,
-    0x98B9EC0097B9EC00, 0x9AB9EC0099B9EC00, 0x9CB9EC009BB9EC00,
-    0x9EB9EC009DB9EC00, 0xA0B9EC009FB9EC00, 0xA2B9EC00A1B9EC00,
-    0xA4B9EC00A3B9EC00, 0xA6B9EC00A5B9EC00, 0xA8B9EC00A7B9EC00,
-    0xAAB9EC00A9B9EC00, 0xACB9EC00ABB9EC00, 0xAEB9EC00ADB9EC00,
-    0xB0B9EC00AFB9EC00, 0xB2B9EC00B1B9EC00, 0xB4B9EC00B3B9EC00,
-    0xB6B9EC00B5B9EC00, 0xB8B9EC00B7B9EC00, 0xBAB9EC00B9B9EC00,
-    0xBCB9EC00BBB9EC00, 0xBEB9EC00BDB9EC00, 0x80BAEC00BFB9EC00,
-    0x82BAEC0081BAEC00, 0x84BAEC0083BAEC00, 0x86BAEC0085BAEC00,
-    0x88BAEC0087BAEC00, 0x8ABAEC0089BAEC00, 0x8CBAEC008BBAEC00,
-    0x8EBAEC008DBAEC00, 0x90BAEC008FBAEC00, 0x92BAEC0091BAEC00,
-    0x94BAEC0093BAEC00, 0x96BAEC0095BAEC00, 0x98BAEC0097BAEC00,
-    0x9ABAEC0099BAEC00, 0x9CBAEC009BBAEC00, 0x9EBAEC009DBAEC00,
-    0xA0BAEC009FBAEC00, 0xA2BAEC00A1BAEC00, 0xA4BAEC00A3BAEC00,
-    0xA6BAEC00A5BAEC00, 0xA8BAEC00A7BAEC00, 0xAABAEC00A9BAEC00,
-    0xACBAEC00ABBAEC00, 0xAEBAEC00ADBAEC00, 0xB0BAEC00AFBAEC00,
-    0xB2BAEC00B1BAEC00, 0xB4BAEC00B3BAEC00, 0xB6BAEC00B5BAEC00,
-    0xB8BAEC00B7BAEC00, 0xBABAEC00B9BAEC00, 0xBCBAEC00BBBAEC00,
-    0xBEBAEC00BDBAEC00, 0x80BBEC00BFBAEC00, 0x82BBEC0081BBEC00,
-    0x84BBEC0083BBEC00, 0x86BBEC0085BBEC00, 0x88BBEC0087BBEC00,
-    0x8ABBEC0089BBEC00, 0x8CBBEC008BBBEC00, 0x8EBBEC008DBBEC00,
-    0x90BBEC008FBBEC00, 0x92BBEC0091BBEC00, 0x94BBEC0093BBEC00,
-    0x96BBEC0095BBEC00, 0x98BBEC0097BBEC00, 0x9ABBEC0099BBEC00,
-    0x9CBBEC009BBBEC00, 0x9EBBEC009DBBEC00, 0xA0BBEC009FBBEC00,
-    0xA2BBEC00A1BBEC00, 0xA4BBEC00A3BBEC00, 0xA6BBEC00A5BBEC00,
-    0xA8BBEC00A7BBEC00, 0xAABBEC00A9BBEC00, 0xACBBEC00ABBBEC00,
-    0xAEBBEC00ADBBEC00, 0xB0BBEC00AFBBEC00, 0xB2BBEC00B1BBEC00,
-    0xB4BBEC00B3BBEC00, 0xB6BBEC00B5BBEC00, 0xB8BBEC00B7BBEC00,
-    0xBABBEC00B9BBEC00, 0xBCBBEC00BBBBEC00, 0xBEBBEC00BDBBEC00,
-    0x80BCEC00BFBBEC00, 0x82BCEC0081BCEC00, 0x84BCEC0083BCEC00,
-    0x86BCEC0085BCEC00, 0x88BCEC0087BCEC00, 0x8ABCEC0089BCEC00,
-    0x8CBCEC008BBCEC00, 0x8EBCEC008DBCEC00, 0x90BCEC008FBCEC00,
-    0x92BCEC0091BCEC00, 0x94BCEC0093BCEC00, 0x96BCEC0095BCEC00,
-    0x98BCEC0097BCEC00, 0x9ABCEC0099BCEC00, 0x9CBCEC009BBCEC00,
-    0x9EBCEC009DBCEC00, 0xA0BCEC009FBCEC00, 0xA2BCEC00A1BCEC00,
-    0xA4BCEC00A3BCEC00, 0xA6BCEC00A5BCEC00, 0xA8BCEC00A7BCEC00,
-    0xAABCEC00A9BCEC00, 0xACBCEC00ABBCEC00, 0xAEBCEC00ADBCEC00,
-    0xB0BCEC00AFBCEC00, 0xB2BCEC00B1BCEC00, 0xB4BCEC00B3BCEC00,
-    0xB6BCEC00B5BCEC00, 0xB8BCEC00B7BCEC00, 0xBABCEC00B9BCEC00,
-    0xBCBCEC00BBBCEC00, 0xBEBCEC00BDBCEC00, 0x80BDEC00BFBCEC00,
-    0x82BDEC0081BDEC00, 0x84BDEC0083BDEC00, 0x86BDEC0085BDEC00,
-    0x88BDEC0087BDEC00, 0x8ABDEC0089BDEC00, 0x8CBDEC008BBDEC00,
-    0x8EBDEC008DBDEC00, 0x90BDEC008FBDEC00, 0x92BDEC0091BDEC00,
-    0x94BDEC0093BDEC00, 0x96BDEC0095BDEC00, 0x98BDEC0097BDEC00,
-    0x9ABDEC0099BDEC00, 0x9CBDEC009BBDEC00, 0x9EBDEC009DBDEC00,
-    0xA0BDEC009FBDEC00, 0xA2BDEC00A1BDEC00, 0xA4BDEC00A3BDEC00,
-    0xA6BDEC00A5BDEC00, 0xA8BDEC00A7BDEC00, 0xAABDEC00A9BDEC00,
-    0xACBDEC00ABBDEC00, 0xAEBDEC00ADBDEC00, 0xB0BDEC00AFBDEC00,
-    0xB2BDEC00B1BDEC00, 0xB4BDEC00B3BDEC00, 0xB6BDEC00B5BDEC00,
-    0xB8BDEC00B7BDEC00, 0xBABDEC00B9BDEC00, 0xBCBDEC00BBBDEC00,
-    0xBEBDEC00BDBDEC00, 0x80BEEC00BFBDEC00, 0x82BEEC0081BEEC00,
-    0x84BEEC0083BEEC00, 0x86BEEC0085BEEC00, 0x88BEEC0087BEEC00,
-    0x8ABEEC0089BEEC00, 0x8CBEEC008BBEEC00, 0x8EBEEC008DBEEC00,
-    0x90BEEC008FBEEC00, 0x92BEEC0091BEEC00, 0x94BEEC0093BEEC00,
-    0x96BEEC0095BEEC00, 0x98BEEC0097BEEC00, 0x9ABEEC0099BEEC00,
-    0x9CBEEC009BBEEC00, 0x9EBEEC009DBEEC00, 0xA0BEEC009FBEEC00,
-    0xA2BEEC00A1BEEC00, 0xA4BEEC00A3BEEC00, 0xA6BEEC00A5BEEC00,
-    0xA8BEEC00A7BEEC00, 0xAABEEC00A9BEEC00, 0xACBEEC00ABBEEC00,
-    0xAEBEEC00ADBEEC00, 0xB0BEEC00AFBEEC00, 0xB2BEEC00B1BEEC00,
-    0xB4BEEC00B3BEEC00, 0xB6BEEC00B5BEEC00, 0xB8BEEC00B7BEEC00,
-    0xBABEEC00B9BEEC00, 0xBCBEEC00BBBEEC00, 0xBEBEEC00BDBEEC00,
-    0x80BFEC00BFBEEC00, 0x82BFEC0081BFEC00, 0x84BFEC0083BFEC00,
-    0x86BFEC0085BFEC00, 0x88BFEC0087BFEC00, 0x8ABFEC0089BFEC00,
-    0x8CBFEC008BBFEC00, 0x8EBFEC008DBFEC00, 0x90BFEC008FBFEC00,
-    0x92BFEC0091BFEC00, 0x94BFEC0093BFEC00, 0x96BFEC0095BFEC00,
-    0x98BFEC0097BFEC00, 0x9ABFEC0099BFEC00, 0x9CBFEC009BBFEC00,
-    0x9EBFEC009DBFEC00, 0xA0BFEC009FBFEC00, 0xA2BFEC00A1BFEC00,
-    0xA4BFEC00A3BFEC00, 0xA6BFEC00A5BFEC00, 0xA8BFEC00A7BFEC00,
-    0xAABFEC00A9BFEC00, 0xACBFEC00ABBFEC00, 0xAEBFEC00ADBFEC00,
-    0xB0BFEC00AFBFEC00, 0xB2BFEC00B1BFEC00, 0xB4BFEC00B3BFEC00,
-    0xB6BFEC00B5BFEC00, 0xB8BFEC00B7BFEC00, 0xBABFEC00B9BFEC00,
-    0xBCBFEC00BBBFEC00, 0xBEBFEC00BDBFEC00, 0x8080ED00BFBFEC00,
-    0x8280ED008180ED00, 0x8480ED008380ED00, 0x8680ED008580ED00,
-    0x8880ED008780ED00, 0x8A80ED008980ED00, 0x8C80ED008B80ED00,
-    0x8E80ED008D80ED00, 0x9080ED008F80ED00, 0x9280ED009180ED00,
-    0x9480ED009380ED00, 0x9680ED009580ED00, 0x9880ED009780ED00,
-    0x9A80ED009980ED00, 0x9C80ED009B80ED00, 0x9E80ED009D80ED00,
-    0xA080ED009F80ED00, 0xA280ED00A180ED00, 0xA480ED00A380ED00,
-    0xA680ED00A580ED00, 0xA880ED00A780ED00, 0xAA80ED00A980ED00,
-    0xAC80ED00AB80ED00, 0xAE80ED00AD80ED00, 0xB080ED00AF80ED00,
-    0xB280ED00B180ED00, 0xB480ED00B380ED00, 0xB680ED00B580ED00,
-    0xB880ED00B780ED00, 0xBA80ED00B980ED00, 0xBC80ED00BB80ED00,
-    0xBE80ED00BD80ED00, 0x8081ED00BF80ED00, 0x8281ED008181ED00,
-    0x8481ED008381ED00, 0x8681ED008581ED00, 0x8881ED008781ED00,
-    0x8A81ED008981ED00, 0x8C81ED008B81ED00, 0x8E81ED008D81ED00,
-    0x9081ED008F81ED00, 0x9281ED009181ED00, 0x9481ED009381ED00,
-    0x9681ED009581ED00, 0x9881ED009781ED00, 0x9A81ED009981ED00,
-    0x9C81ED009B81ED00, 0x9E81ED009D81ED00, 0xA081ED009F81ED00,
-    0xA281ED00A181ED00, 0xA481ED00A381ED00, 0xA681ED00A581ED00,
-    0xA881ED00A781ED00, 0xAA81ED00A981ED00, 0xAC81ED00AB81ED00,
-    0xAE81ED00AD81ED00, 0xB081ED00AF81ED00, 0xB281ED00B181ED00,
-    0xB481ED00B381ED00, 0xB681ED00B581ED00, 0xB881ED00B781ED00,
-    0xBA81ED00B981ED00, 0xBC81ED00BB81ED00, 0xBE81ED00BD81ED00,
-    0x8082ED00BF81ED00, 0x8282ED008182ED00, 0x8482ED008382ED00,
-    0x8682ED008582ED00, 0x8882ED008782ED00, 0x8A82ED008982ED00,
-    0x8C82ED008B82ED00, 0x8E82ED008D82ED00, 0x9082ED008F82ED00,
-    0x9282ED009182ED00, 0x9482ED009382ED00, 0x9682ED009582ED00,
-    0x9882ED009782ED00, 0x9A82ED009982ED00, 0x9C82ED009B82ED00,
-    0x9E82ED009D82ED00, 0xA082ED009F82ED00, 0xA282ED00A182ED00,
-    0xA482ED00A382ED00, 0xA682ED00A582ED00, 0xA882ED00A782ED00,
-    0xAA82ED00A982ED00, 0xAC82ED00AB82ED00, 0xAE82ED00AD82ED00,
-    0xB082ED00AF82ED00, 0xB282ED00B182ED00, 0xB482ED00B382ED00,
-    0xB682ED00B582ED00, 0xB882ED00B782ED00, 0xBA82ED00B982ED00,
-    0xBC82ED00BB82ED00, 0xBE82ED00BD82ED00, 0x8083ED00BF82ED00,
-    0x8283ED008183ED00, 0x8483ED008383ED00, 0x8683ED008583ED00,
-    0x8883ED008783ED00, 0x8A83ED008983ED00, 0x8C83ED008B83ED00,
-    0x8E83ED008D83ED00, 0x9083ED008F83ED00, 0x9283ED009183ED00,
-    0x9483ED009383ED00, 0x9683ED009583ED00, 0x9883ED009783ED00,
-    0x9A83ED009983ED00, 0x9C83ED009B83ED00, 0x9E83ED009D83ED00,
-    0xA083ED009F83ED00, 0xA283ED00A183ED00, 0xA483ED00A383ED00,
-    0xA683ED00A583ED00, 0xA883ED00A783ED00, 0xAA83ED00A983ED00,
-    0xAC83ED00AB83ED00, 0xAE83ED00AD83ED00, 0xB083ED00AF83ED00,
-    0xB283ED00B183ED00, 0xB483ED00B383ED00, 0xB683ED00B583ED00,
-    0xB883ED00B783ED00, 0xBA83ED00B983ED00, 0xBC83ED00BB83ED00,
-    0xBE83ED00BD83ED00, 0x8084ED00BF83ED00, 0x8284ED008184ED00,
-    0x8484ED008384ED00, 0x8684ED008584ED00, 0x8884ED008784ED00,
-    0x8A84ED008984ED00, 0x8C84ED008B84ED00, 0x8E84ED008D84ED00,
-    0x9084ED008F84ED00, 0x9284ED009184ED00, 0x9484ED009384ED00,
-    0x9684ED009584ED00, 0x9884ED009784ED00, 0x9A84ED009984ED00,
-    0x9C84ED009B84ED00, 0x9E84ED009D84ED00, 0xA084ED009F84ED00,
-    0xA284ED00A184ED00, 0xA484ED00A384ED00, 0xA684ED00A584ED00,
-    0xA884ED00A784ED00, 0xAA84ED00A984ED00, 0xAC84ED00AB84ED00,
-    0xAE84ED00AD84ED00, 0xB084ED00AF84ED00, 0xB284ED00B184ED00,
-    0xB484ED00B384ED00, 0xB684ED00B584ED00, 0xB884ED00B784ED00,
-    0xBA84ED00B984ED00, 0xBC84ED00BB84ED00, 0xBE84ED00BD84ED00,
-    0x8085ED00BF84ED00, 0x8285ED008185ED00, 0x8485ED008385ED00,
-    0x8685ED008585ED00, 0x8885ED008785ED00, 0x8A85ED008985ED00,
-    0x8C85ED008B85ED00, 0x8E85ED008D85ED00, 0x9085ED008F85ED00,
-    0x9285ED009185ED00, 0x9485ED009385ED00, 0x9685ED009585ED00,
-    0x9885ED009785ED00, 0x9A85ED009985ED00, 0x9C85ED009B85ED00,
-    0x9E85ED009D85ED00, 0xA085ED009F85ED00, 0xA285ED00A185ED00,
-    0xA485ED00A385ED00, 0xA685ED00A585ED00, 0xA885ED00A785ED00,
-    0xAA85ED00A985ED00, 0xAC85ED00AB85ED00, 0xAE85ED00AD85ED00,
-    0xB085ED00AF85ED00, 0xB285ED00B185ED00, 0xB485ED00B385ED00,
-    0xB685ED00B585ED00, 0xB885ED00B785ED00, 0xBA85ED00B985ED00,
-    0xBC85ED00BB85ED00, 0xBE85ED00BD85ED00, 0x8086ED00BF85ED00,
-    0x8286ED008186ED00, 0x8486ED008386ED00, 0x8686ED008586ED00,
-    0x8886ED008786ED00, 0x8A86ED008986ED00, 0x8C86ED008B86ED00,
-    0x8E86ED008D86ED00, 0x9086ED008F86ED00, 0x9286ED009186ED00,
-    0x9486ED009386ED00, 0x9686ED009586ED00, 0x9886ED009786ED00,
-    0x9A86ED009986ED00, 0x9C86ED009B86ED00, 0x9E86ED009D86ED00,
-    0xA086ED009F86ED00, 0xA286ED00A186ED00, 0xA486ED00A386ED00,
-    0xA686ED00A586ED00, 0xA886ED00A786ED00, 0xAA86ED00A986ED00,
-    0xAC86ED00AB86ED00, 0xAE86ED00AD86ED00, 0xB086ED00AF86ED00,
-    0xB286ED00B186ED00, 0xB486ED00B386ED00, 0xB686ED00B586ED00,
-    0xB886ED00B786ED00, 0xBA86ED00B986ED00, 0xBC86ED00BB86ED00,
-    0xBE86ED00BD86ED00, 0x8087ED00BF86ED00, 0x8287ED008187ED00,
-    0x8487ED008387ED00, 0x8687ED008587ED00, 0x8887ED008787ED00,
-    0x8A87ED008987ED00, 0x8C87ED008B87ED00, 0x8E87ED008D87ED00,
-    0x9087ED008F87ED00, 0x9287ED009187ED00, 0x9487ED009387ED00,
-    0x9687ED009587ED00, 0x9887ED009787ED00, 0x9A87ED009987ED00,
-    0x9C87ED009B87ED00, 0x9E87ED009D87ED00, 0xA087ED009F87ED00,
-    0xA287ED00A187ED00, 0xA487ED00A387ED00, 0xA687ED00A587ED00,
-    0xA887ED00A787ED00, 0xAA87ED00A987ED00, 0xAC87ED00AB87ED00,
-    0xAE87ED00AD87ED00, 0xB087ED00AF87ED00, 0xB287ED00B187ED00,
-    0xB487ED00B387ED00, 0xB687ED00B587ED00, 0xB887ED00B787ED00,
-    0xBA87ED00B987ED00, 0xBC87ED00BB87ED00, 0xBE87ED00BD87ED00,
-    0x8088ED00BF87ED00, 0x8288ED008188ED00, 0x8488ED008388ED00,
-    0x8688ED008588ED00, 0x8888ED008788ED00, 0x8A88ED008988ED00,
-    0x8C88ED008B88ED00, 0x8E88ED008D88ED00, 0x9088ED008F88ED00,
-    0x9288ED009188ED00, 0x9488ED009388ED00, 0x9688ED009588ED00,
-    0x9888ED009788ED00, 0x9A88ED009988ED00, 0x9C88ED009B88ED00,
-    0x9E88ED009D88ED00, 0xA088ED009F88ED00, 0xA288ED00A188ED00,
-    0xA488ED00A388ED00, 0xA688ED00A588ED00, 0xA888ED00A788ED00,
-    0xAA88ED00A988ED00, 0xAC88ED00AB88ED00, 0xAE88ED00AD88ED00,
-    0xB088ED00AF88ED00, 0xB288ED00B188ED00, 0xB488ED00B388ED00,
-    0xB688ED00B588ED00, 0xB888ED00B788ED00, 0xBA88ED00B988ED00,
-    0xBC88ED00BB88ED00, 0xBE88ED00BD88ED00, 0x8089ED00BF88ED00,
-    0x8289ED008189ED00, 0x8489ED008389ED00, 0x8689ED008589ED00,
-    0x8889ED008789ED00, 0x8A89ED008989ED00, 0x8C89ED008B89ED00,
-    0x8E89ED008D89ED00, 0x9089ED008F89ED00, 0x9289ED009189ED00,
-    0x9489ED009389ED00, 0x9689ED009589ED00, 0x9889ED009789ED00,
-    0x9A89ED009989ED00, 0x9C89ED009B89ED00, 0x9E89ED009D89ED00,
-    0xA089ED009F89ED00, 0xA289ED00A189ED00, 0xA489ED00A389ED00,
-    0xA689ED00A589ED00, 0xA889ED00A789ED00, 0xAA89ED00A989ED00,
-    0xAC89ED00AB89ED00, 0xAE89ED00AD89ED00, 0xB089ED00AF89ED00,
-    0xB289ED00B189ED00, 0xB489ED00B389ED00, 0xB689ED00B589ED00,
-    0xB889ED00B789ED00, 0xBA89ED00B989ED00, 0xBC89ED00BB89ED00,
-    0xBE89ED00BD89ED00, 0x808AED00BF89ED00, 0x828AED00818AED00,
-    0x848AED00838AED00, 0x868AED00858AED00, 0x888AED00878AED00,
-    0x8A8AED00898AED00, 0x8C8AED008B8AED00, 0x8E8AED008D8AED00,
-    0x908AED008F8AED00, 0x928AED00918AED00, 0x948AED00938AED00,
-    0x968AED00958AED00, 0x988AED00978AED00, 0x9A8AED00998AED00,
-    0x9C8AED009B8AED00, 0x9E8AED009D8AED00, 0xA08AED009F8AED00,
-    0xA28AED00A18AED00, 0xA48AED00A38AED00, 0xA68AED00A58AED00,
-    0xA88AED00A78AED00, 0xAA8AED00A98AED00, 0xAC8AED00AB8AED00,
-    0xAE8AED00AD8AED00, 0xB08AED00AF8AED00, 0xB28AED00B18AED00,
-    0xB48AED00B38AED00, 0xB68AED00B58AED00, 0xB88AED00B78AED00,
-    0xBA8AED00B98AED00, 0xBC8AED00BB8AED00, 0xBE8AED00BD8AED00,
-    0x808BED00BF8AED00, 0x828BED00818BED00, 0x848BED00838BED00,
-    0x868BED00858BED00, 0x888BED00878BED00, 0x8A8BED00898BED00,
-    0x8C8BED008B8BED00, 0x8E8BED008D8BED00, 0x908BED008F8BED00,
-    0x928BED00918BED00, 0x948BED00938BED00, 0x968BED00958BED00,
-    0x988BED00978BED00, 0x9A8BED00998BED00, 0x9C8BED009B8BED00,
-    0x9E8BED009D8BED00, 0xA08BED009F8BED00, 0xA28BED00A18BED00,
-    0xA48BED00A38BED00, 0xA68BED00A58BED00, 0xA88BED00A78BED00,
-    0xAA8BED00A98BED00, 0xAC8BED00AB8BED00, 0xAE8BED00AD8BED00,
-    0xB08BED00AF8BED00, 0xB28BED00B18BED00, 0xB48BED00B38BED00,
-    0xB68BED00B58BED00, 0xB88BED00B78BED00, 0xBA8BED00B98BED00,
-    0xBC8BED00BB8BED00, 0xBE8BED00BD8BED00, 0x808CED00BF8BED00,
-    0x828CED00818CED00, 0x848CED00838CED00, 0x868CED00858CED00,
-    0x888CED00878CED00, 0x8A8CED00898CED00, 0x8C8CED008B8CED00,
-    0x8E8CED008D8CED00, 0x908CED008F8CED00, 0x928CED00918CED00,
-    0x948CED00938CED00, 0x968CED00958CED00, 0x988CED00978CED00,
-    0x9A8CED00998CED00, 0x9C8CED009B8CED00, 0x9E8CED009D8CED00,
-    0xA08CED009F8CED00, 0xA28CED00A18CED00, 0xA48CED00A38CED00,
-    0xA68CED00A58CED00, 0xA88CED00A78CED00, 0xAA8CED00A98CED00,
-    0xAC8CED00AB8CED00, 0xAE8CED00AD8CED00, 0xB08CED00AF8CED00,
-    0xB28CED00B18CED00, 0xB48CED00B38CED00, 0xB68CED00B58CED00,
-    0xB88CED00B78CED00, 0xBA8CED00B98CED00, 0xBC8CED00BB8CED00,
-    0xBE8CED00BD8CED00, 0x808DED00BF8CED00, 0x828DED00818DED00,
-    0x848DED00838DED00, 0x868DED00858DED00, 0x888DED00878DED00,
-    0x8A8DED00898DED00, 0x8C8DED008B8DED00, 0x8E8DED008D8DED00,
-    0x908DED008F8DED00, 0x928DED00918DED00, 0x948DED00938DED00,
-    0x968DED00958DED00, 0x988DED00978DED00, 0x9A8DED00998DED00,
-    0x9C8DED009B8DED00, 0x9E8DED009D8DED00, 0xA08DED009F8DED00,
-    0xA28DED00A18DED00, 0xA48DED00A38DED00, 0xA68DED00A58DED00,
-    0xA88DED00A78DED00, 0xAA8DED00A98DED00, 0xAC8DED00AB8DED00,
-    0xAE8DED00AD8DED00, 0xB08DED00AF8DED00, 0xB28DED00B18DED00,
-    0xB48DED00B38DED00, 0xB68DED00B58DED00, 0xB88DED00B78DED00,
-    0xBA8DED00B98DED00, 0xBC8DED00BB8DED00, 0xBE8DED00BD8DED00,
-    0x808EED00BF8DED00, 0x828EED00818EED00, 0x848EED00838EED00,
-    0x868EED00858EED00, 0x888EED00878EED00, 0x8A8EED00898EED00,
-    0x8C8EED008B8EED00, 0x8E8EED008D8EED00, 0x908EED008F8EED00,
-    0x928EED00918EED00, 0x948EED00938EED00, 0x968EED00958EED00,
-    0x988EED00978EED00, 0x9A8EED00998EED00, 0x9C8EED009B8EED00,
-    0x9E8EED009D8EED00, 0xA08EED009F8EED00, 0xA28EED00A18EED00,
-    0xA48EED00A38EED00, 0xA68EED00A58EED00, 0xA88EED00A78EED00,
-    0xAA8EED00A98EED00, 0xAC8EED00AB8EED00, 0xAE8EED00AD8EED00,
-    0xB08EED00AF8EED00, 0xB28EED00B18EED00, 0xB48EED00B38EED00,
-    0xB68EED00B58EED00, 0xB88EED00B78EED00, 0xBA8EED00B98EED00,
-    0xBC8EED00BB8EED00, 0xBE8EED00BD8EED00, 0x808FED00BF8EED00,
-    0x828FED00818FED00, 0x848FED00838FED00, 0x868FED00858FED00,
-    0x888FED00878FED00, 0x8A8FED00898FED00, 0x8C8FED008B8FED00,
-    0x8E8FED008D8FED00, 0x908FED008F8FED00, 0x928FED00918FED00,
-    0x948FED00938FED00, 0x968FED00958FED00, 0x988FED00978FED00,
-    0x9A8FED00998FED00, 0x9C8FED009B8FED00, 0x9E8FED009D8FED00,
-    0xA08FED009F8FED00, 0xA28FED00A18FED00, 0xA48FED00A38FED00,
-    0xA68FED00A58FED00, 0xA88FED00A78FED00, 0xAA8FED00A98FED00,
-    0xAC8FED00AB8FED00, 0xAE8FED00AD8FED00, 0xB08FED00AF8FED00,
-    0xB28FED00B18FED00, 0xB48FED00B38FED00, 0xB68FED00B58FED00,
-    0xB88FED00B78FED00, 0xBA8FED00B98FED00, 0xBC8FED00BB8FED00,
-    0xBE8FED00BD8FED00, 0x8090ED00BF8FED00, 0x8290ED008190ED00,
-    0x8490ED008390ED00, 0x8690ED008590ED00, 0x8890ED008790ED00,
-    0x8A90ED008990ED00, 0x8C90ED008B90ED00, 0x8E90ED008D90ED00,
-    0x9090ED008F90ED00, 0x9290ED009190ED00, 0x9490ED009390ED00,
-    0x9690ED009590ED00, 0x9890ED009790ED00, 0x9A90ED009990ED00,
-    0x9C90ED009B90ED00, 0x9E90ED009D90ED00, 0xA090ED009F90ED00,
-    0xA290ED00A190ED00, 0xA490ED00A390ED00, 0xA690ED00A590ED00,
-    0xA890ED00A790ED00, 0xAA90ED00A990ED00, 0xAC90ED00AB90ED00,
-    0xAE90ED00AD90ED00, 0xB090ED00AF90ED00, 0xB290ED00B190ED00,
-    0xB490ED00B390ED00, 0xB690ED00B590ED00, 0xB890ED00B790ED00,
-    0xBA90ED00B990ED00, 0xBC90ED00BB90ED00, 0xBE90ED00BD90ED00,
-    0x8091ED00BF90ED00, 0x8291ED008191ED00, 0x8491ED008391ED00,
-    0x8691ED008591ED00, 0x8891ED008791ED00, 0x8A91ED008991ED00,
-    0x8C91ED008B91ED00, 0x8E91ED008D91ED00, 0x9091ED008F91ED00,
-    0x9291ED009191ED00, 0x9491ED009391ED00, 0x9691ED009591ED00,
-    0x9891ED009791ED00, 0x9A91ED009991ED00, 0x9C91ED009B91ED00,
-    0x9E91ED009D91ED00, 0xA091ED009F91ED00, 0xA291ED00A191ED00,
-    0xA491ED00A391ED00, 0xA691ED00A591ED00, 0xA891ED00A791ED00,
-    0xAA91ED00A991ED00, 0xAC91ED00AB91ED00, 0xAE91ED00AD91ED00,
-    0xB091ED00AF91ED00, 0xB291ED00B191ED00, 0xB491ED00B391ED00,
-    0xB691ED00B591ED00, 0xB891ED00B791ED00, 0xBA91ED00B991ED00,
-    0xBC91ED00BB91ED00, 0xBE91ED00BD91ED00, 0x8092ED00BF91ED00,
-    0x8292ED008192ED00, 0x8492ED008392ED00, 0x8692ED008592ED00,
-    0x8892ED008792ED00, 0x8A92ED008992ED00, 0x8C92ED008B92ED00,
-    0x8E92ED008D92ED00, 0x9092ED008F92ED00, 0x9292ED009192ED00,
-    0x9492ED009392ED00, 0x9692ED009592ED00, 0x9892ED009792ED00,
-    0x9A92ED009992ED00, 0x9C92ED009B92ED00, 0x9E92ED009D92ED00,
-    0xA092ED009F92ED00, 0xA292ED00A192ED00, 0xA492ED00A392ED00,
-    0xA692ED00A592ED00, 0xA892ED00A792ED00, 0xAA92ED00A992ED00,
-    0xAC92ED00AB92ED00, 0xAE92ED00AD92ED00, 0xB092ED00AF92ED00,
-    0xB292ED00B192ED00, 0xB492ED00B392ED00, 0xB692ED00B592ED00,
-    0xB892ED00B792ED00, 0xBA92ED00B992ED00, 0xBC92ED00BB92ED00,
-    0xBE92ED00BD92ED00, 0x8093ED00BF92ED00, 0x8293ED008193ED00,
-    0x8493ED008393ED00, 0x8693ED008593ED00, 0x8893ED008793ED00,
-    0x8A93ED008993ED00, 0x8C93ED008B93ED00, 0x8E93ED008D93ED00,
-    0x9093ED008F93ED00, 0x9293ED009193ED00, 0x9493ED009393ED00,
-    0x9693ED009593ED00, 0x9893ED009793ED00, 0x9A93ED009993ED00,
-    0x9C93ED009B93ED00, 0x9E93ED009D93ED00, 0xA093ED009F93ED00,
-    0xA293ED00A193ED00, 0xA493ED00A393ED00, 0xA693ED00A593ED00,
-    0xA893ED00A793ED00, 0xAA93ED00A993ED00, 0xAC93ED00AB93ED00,
-    0xAE93ED00AD93ED00, 0xB093ED00AF93ED00, 0xB293ED00B193ED00,
-    0xB493ED00B393ED00, 0xB693ED00B593ED00, 0xB893ED00B793ED00,
-    0xBA93ED00B993ED00, 0xBC93ED00BB93ED00, 0xBE93ED00BD93ED00,
-    0x8094ED00BF93ED00, 0x8294ED008194ED00, 0x8494ED008394ED00,
-    0x8694ED008594ED00, 0x8894ED008794ED00, 0x8A94ED008994ED00,
-    0x8C94ED008B94ED00, 0x8E94ED008D94ED00, 0x9094ED008F94ED00,
-    0x9294ED009194ED00, 0x9494ED009394ED00, 0x9694ED009594ED00,
-    0x9894ED009794ED00, 0x9A94ED009994ED00, 0x9C94ED009B94ED00,
-    0x9E94ED009D94ED00, 0xA094ED009F94ED00, 0xA294ED00A194ED00,
-    0xA494ED00A394ED00, 0xA694ED00A594ED00, 0xA894ED00A794ED00,
-    0xAA94ED00A994ED00, 0xAC94ED00AB94ED00, 0xAE94ED00AD94ED00,
-    0xB094ED00AF94ED00, 0xB294ED00B194ED00, 0xB494ED00B394ED00,
-    0xB694ED00B594ED00, 0xB894ED00B794ED00, 0xBA94ED00B994ED00,
-    0xBC94ED00BB94ED00, 0xBE94ED00BD94ED00, 0x8095ED00BF94ED00,
-    0x8295ED008195ED00, 0x8495ED008395ED00, 0x8695ED008595ED00,
-    0x8895ED008795ED00, 0x8A95ED008995ED00, 0x8C95ED008B95ED00,
-    0x8E95ED008D95ED00, 0x9095ED008F95ED00, 0x9295ED009195ED00,
-    0x9495ED009395ED00, 0x9695ED009595ED00, 0x9895ED009795ED00,
-    0x9A95ED009995ED00, 0x9C95ED009B95ED00, 0x9E95ED009D95ED00,
-    0xA095ED009F95ED00, 0xA295ED00A195ED00, 0xA495ED00A395ED00,
-    0xA695ED00A595ED00, 0xA895ED00A795ED00, 0xAA95ED00A995ED00,
-    0xAC95ED00AB95ED00, 0xAE95ED00AD95ED00, 0xB095ED00AF95ED00,
-    0xB295ED00B195ED00, 0xB495ED00B395ED00, 0xB695ED00B595ED00,
-    0xB895ED00B795ED00, 0xBA95ED00B995ED00, 0xBC95ED00BB95ED00,
-    0xBE95ED00BD95ED00, 0x8096ED00BF95ED00, 0x8296ED008196ED00,
-    0x8496ED008396ED00, 0x8696ED008596ED00, 0x8896ED008796ED00,
-    0x8A96ED008996ED00, 0x8C96ED008B96ED00, 0x8E96ED008D96ED00,
-    0x9096ED008F96ED00, 0x9296ED009196ED00, 0x9496ED009396ED00,
-    0x9696ED009596ED00, 0x9896ED009796ED00, 0x9A96ED009996ED00,
-    0x9C96ED009B96ED00, 0x9E96ED009D96ED00, 0xA096ED009F96ED00,
-    0xA296ED00A196ED00, 0xA496ED00A396ED00, 0xA696ED00A596ED00,
-    0xA896ED00A796ED00, 0xAA96ED00A996ED00, 0xAC96ED00AB96ED00,
-    0xAE96ED00AD96ED00, 0xB096ED00AF96ED00, 0xB296ED00B196ED00,
-    0xB496ED00B396ED00, 0xB696ED00B596ED00, 0xB896ED00B796ED00,
-    0xBA96ED00B996ED00, 0xBC96ED00BB96ED00, 0xBE96ED00BD96ED00,
-    0x8097ED00BF96ED00, 0x8297ED008197ED00, 0x8497ED008397ED00,
-    0x8697ED008597ED00, 0x8897ED008797ED00, 0x8A97ED008997ED00,
-    0x8C97ED008B97ED00, 0x8E97ED008D97ED00, 0x9097ED008F97ED00,
-    0x9297ED009197ED00, 0x9497ED009397ED00, 0x9697ED009597ED00,
-    0x9897ED009797ED00, 0x9A97ED009997ED00, 0x9C97ED009B97ED00,
-    0x9E97ED009D97ED00, 0xA097ED009F97ED00, 0xA297ED00A197ED00,
-    0xA497ED00A397ED00, 0xA697ED00A597ED00, 0xA897ED00A797ED00,
-    0xAA97ED00A997ED00, 0xAC97ED00AB97ED00, 0xAE97ED00AD97ED00,
-    0xB097ED00AF97ED00, 0xB297ED00B197ED00, 0xB497ED00B397ED00,
-    0xB697ED00B597ED00, 0xB897ED00B797ED00, 0xBA97ED00B997ED00,
-    0xBC97ED00BB97ED00, 0xBE97ED00BD97ED00, 0x8098ED00BF97ED00,
-    0x8298ED008198ED00, 0x8498ED008398ED00, 0x8698ED008598ED00,
-    0x8898ED008798ED00, 0x8A98ED008998ED00, 0x8C98ED008B98ED00,
-    0x8E98ED008D98ED00, 0x9098ED008F98ED00, 0x9298ED009198ED00,
-    0x9498ED009398ED00, 0x9698ED009598ED00, 0x9898ED009798ED00,
-    0x9A98ED009998ED00, 0x9C98ED009B98ED00, 0x9E98ED009D98ED00,
-    0xA098ED009F98ED00, 0xA298ED00A198ED00, 0xA498ED00A398ED00,
-    0xA698ED00A598ED00, 0xA898ED00A798ED00, 0xAA98ED00A998ED00,
-    0xAC98ED00AB98ED00, 0xAE98ED00AD98ED00, 0xB098ED00AF98ED00,
-    0xB298ED00B198ED00, 0xB498ED00B398ED00, 0xB698ED00B598ED00,
-    0xB898ED00B798ED00, 0xBA98ED00B998ED00, 0xBC98ED00BB98ED00,
-    0xBE98ED00BD98ED00, 0x8099ED00BF98ED00, 0x8299ED008199ED00,
-    0x8499ED008399ED00, 0x8699ED008599ED00, 0x8899ED008799ED00,
-    0x8A99ED008999ED00, 0x8C99ED008B99ED00, 0x8E99ED008D99ED00,
-    0x9099ED008F99ED00, 0x9299ED009199ED00, 0x9499ED009399ED00,
-    0x9699ED009599ED00, 0x9899ED009799ED00, 0x9A99ED009999ED00,
-    0x9C99ED009B99ED00, 0x9E99ED009D99ED00, 0xA099ED009F99ED00,
-    0xA299ED00A199ED00, 0xA499ED00A399ED00, 0xA699ED00A599ED00,
-    0xA899ED00A799ED00, 0xAA99ED00A999ED00, 0xAC99ED00AB99ED00,
-    0xAE99ED00AD99ED00, 0xB099ED00AF99ED00, 0xB299ED00B199ED00,
-    0xB499ED00B399ED00, 0xB699ED00B599ED00, 0xB899ED00B799ED00,
-    0xBA99ED00B999ED00, 0xBC99ED00BB99ED00, 0xBE99ED00BD99ED00,
-    0x809AED00BF99ED00, 0x829AED00819AED00, 0x849AED00839AED00,
-    0x869AED00859AED00, 0x889AED00879AED00, 0x8A9AED00899AED00,
-    0x8C9AED008B9AED00, 0x8E9AED008D9AED00, 0x909AED008F9AED00,
-    0x929AED00919AED00, 0x949AED00939AED00, 0x969AED00959AED00,
-    0x989AED00979AED00, 0x9A9AED00999AED00, 0x9C9AED009B9AED00,
-    0x9E9AED009D9AED00, 0xA09AED009F9AED00, 0xA29AED00A19AED00,
-    0xA49AED00A39AED00, 0xA69AED00A59AED00, 0xA89AED00A79AED00,
-    0xAA9AED00A99AED00, 0xAC9AED00AB9AED00, 0xAE9AED00AD9AED00,
-    0xB09AED00AF9AED00, 0xB29AED00B19AED00, 0xB49AED00B39AED00,
-    0xB69AED00B59AED00, 0xB89AED00B79AED00, 0xBA9AED00B99AED00,
-    0xBC9AED00BB9AED00, 0xBE9AED00BD9AED00, 0x809BED00BF9AED00,
-    0x829BED00819BED00, 0x849BED00839BED00, 0x869BED00859BED00,
-    0x889BED00879BED00, 0x8A9BED00899BED00, 0x8C9BED008B9BED00,
-    0x8E9BED008D9BED00, 0x909BED008F9BED00, 0x929BED00919BED00,
-    0x949BED00939BED00, 0x969BED00959BED00, 0x989BED00979BED00,
-    0x9A9BED00999BED00, 0x9C9BED009B9BED00, 0x9E9BED009D9BED00,
-    0xA09BED009F9BED00, 0xA29BED00A19BED00, 0xA49BED00A39BED00,
-    0xA69BED00A59BED00, 0xA89BED00A79BED00, 0xAA9BED00A99BED00,
-    0xAC9BED00AB9BED00, 0xAE9BED00AD9BED00, 0xB09BED00AF9BED00,
-    0xB29BED00B19BED00, 0xB49BED00B39BED00, 0xB69BED00B59BED00,
-    0xB89BED00B79BED00, 0xBA9BED00B99BED00, 0xBC9BED00BB9BED00,
-    0xBE9BED00BD9BED00, 0x809CED00BF9BED00, 0x829CED00819CED00,
-    0x849CED00839CED00, 0x869CED00859CED00, 0x889CED00879CED00,
-    0x8A9CED00899CED00, 0x8C9CED008B9CED00, 0x8E9CED008D9CED00,
-    0x909CED008F9CED00, 0x929CED00919CED00, 0x949CED00939CED00,
-    0x969CED00959CED00, 0x989CED00979CED00, 0x9A9CED00999CED00,
-    0x9C9CED009B9CED00, 0x9E9CED009D9CED00, 0xA09CED009F9CED00,
-    0xA29CED00A19CED00, 0xA49CED00A39CED00, 0xA69CED00A59CED00,
-    0xA89CED00A79CED00, 0xAA9CED00A99CED00, 0xAC9CED00AB9CED00,
-    0xAE9CED00AD9CED00, 0xB09CED00AF9CED00, 0xB29CED00B19CED00,
-    0xB49CED00B39CED00, 0xB69CED00B59CED00, 0xB89CED00B79CED00,
-    0xBA9CED00B99CED00, 0xBC9CED00BB9CED00, 0xBE9CED00BD9CED00,
-    0x809DED00BF9CED00, 0x829DED00819DED00, 0x849DED00839DED00,
-    0x869DED00859DED00, 0x889DED00879DED00, 0x8A9DED00899DED00,
-    0x8C9DED008B9DED00, 0x8E9DED008D9DED00, 0x909DED008F9DED00,
-    0x929DED00919DED00, 0x949DED00939DED00, 0x969DED00959DED00,
-    0x989DED00979DED00, 0x9A9DED00999DED00, 0x9C9DED009B9DED00,
-    0x9E9DED009D9DED00, 0xA09DED009F9DED00, 0xA29DED00A19DED00,
-    0xA49DED00A39DED00, 0xA69DED00A59DED00, 0xA89DED00A79DED00,
-    0xAA9DED00A99DED00, 0xAC9DED00AB9DED00, 0xAE9DED00AD9DED00,
-    0xB09DED00AF9DED00, 0xB29DED00B19DED00, 0xB49DED00B39DED00,
-    0xB69DED00B59DED00, 0xB89DED00B79DED00, 0xBA9DED00B99DED00,
-    0xBC9DED00BB9DED00, 0xBE9DED00BD9DED00, 0x809EED00BF9DED00,
-    0x829EED00819EED00, 0x849EED00839EED00, 0x869EED00859EED00,
-    0x889EED00879EED00, 0x8A9EED00899EED00, 0x8C9EED008B9EED00,
-    0x8E9EED008D9EED00, 0x909EED008F9EED00, 0x929EED00919EED00,
-    0x949EED00939EED00, 0x969EED00959EED00, 0x989EED00979EED00,
-    0x9A9EED00999EED00, 0x9C9EED009B9EED00, 0x9E9EED009D9EED00,
-    0xA09EED009F9EED00, 0xA29EED00A19EED00, 0x8291F000A39EED00,
-    0xF0009C8291F0009A, 0xAE8491F000AB8291, 0x91F000AF8491F000,
-    0x008C8D91F0008B8D, 0x9291F000BB9291F0, 0xF000BE9291F000BC,
-    0xBB9691F000BA9691, 0x9DF000B8A491F000, 0xF000A5859DF09785,
-    0x00A5859DF098859D, 0xA5859DF098859DF0, 0x859DF000AE859DF0,
-    0x859DF0A5859DF098, 0x9DF098859DF000AF, 0xF000B0859DF0A585,
-    0xF0A5859DF098859D, 0x98859DF000B1859D, 0xB2859DF0A5859DF0,
-    0x859DF0B9869DF000, 0x9DF0B9869DF000A5, 0xF000AE859DF0A585,
-    0xF0A5859DF0B9869D, 0xBA869DF000AF859D, 0x869DF000A5859DF0,
-    0x859DF0A5859DF0BA, 0x9DF0BA869DF000AE, 0xF000AF859DF0A585,
-    0x85BC9DF00084BC9D, 0x9DF00086BC9DF000, 0x008ABC9DF00088BC,
-    0x84A0F0009EBC9DF0, 0xF0009C94A0F000A2, 0x8B95A0F000A594A0,
-    0xA0F000BA98A0F000, 0x009EA3A0F00084A0, 0xADA0F000ACA8A0F0,
-    0xF000A493A1F000A3, 0xAA9BA1F000A89AA1, 0xA1F00088A7A1F000,
-    0x008BB4A1F00098AC, 0xB7A1F000A4B7A1F0, 0xF0008386A2F000A6,
-    0xB18CA2F0009F86A2, 0xA2F000949BA2F000, 0x008AA1A2F00084A1,
-    0xAFA2F0008CACA2F0, 0xF0008A80A3F000B1, 0x9F8DA3F000B88AA3,
-    0xA3F000938EA3F000, 0x00838FA3F0009C8E, 0x91A3F000958FA3F0,
-    0xF000A39AA3F000AD, 0x8DAAA3F000A7A2A3, 0xA3F000BAABA3F000,
-    0x009EB4A3F000BCB2, 0xBDA3F00091BBA3F0, 0xF0008EBEA3F0009E,
-    0xAE8BA4F000A389A4, 0xA4F000AB8EA4F000, 0x00B59CA4F0008898,
-    0xB0A4F00094A0A4F0, 0xF00092B2A4F000B6, 0xB8BEA4F000A1BEA4,
-    0xA5F0008481A5F000, 0x00B383A5F000B283, 0x84A5F0009984A5F0,
-    0xF0008989A5F000B3, 0xA698A5F0009D90A5, 0xA5F0009A9AA5F000,
-    0x00BCA5A5F000859B, 0xAEA5F000A7AAA5F0, 0xF00080B2A5F000AB,
-    0x86BEA5F00090B3A5, 0xA6F0009A87A6F000, 0x008789A6F000A888,
-    0x8CA6F000998BA6F0, 0xF0009A93A6F000BE, 0xA896A6F000A394A6,
-    0xA6F000A79EA6F000, 0x00BCACA6F000B59E, 0xB3A6F000B6B0A6F0,
-    0xF000ABB5A6F00095, 0xB1BEA6F000ACBCA6, 0xA7F0009283A7F000,
-    0x00A799A7F0008A8F, 0xA5A7F000AEA2A7F0, 0xF000A8B2A7F000A6,
-    0xAFBCA7F00093BBA7, 0xA8F0009297A8F000, 0x00AE9CA8F000AD97,
-    0xB5A8F000BAAFA8F0, 0xF0008585A9F000B7, 0x9A88A9F0009F87A9,
-    0xA9F0008A90A9F000, 0x00B696A9F0009692, 0x83AAF000B0ACA9F0,
-    0xF0008584AAF0008E, 0x918AAAF0008E88AA, 0xAAF000928EAAF000,
-    0x0000000000008098, 0x0000FC000002CC00, 0x0003243D00031C3E,
-    0x0002E4CC00032C3C, 0x8000215D00000DB8, 0x000005B80002DCCC,
-    0x0002E4CC80002151, 0x8000216100001DB8, 0x800008B380000662,
-    0x800008B600010591, 0x0001458C0001558F, 0x0001598A80001E0B,
-    0x00015988800008B9, 0x0000D1860000AD89, 0x0000C9840000D987,
-    0x0000258280001E07, 0x000041800000E183, 0x0000CD8000003D81,
-    0x80001E130000E981, 0x80001E0F00000583, 0x0000D580800008BC,
-    0x800009100000F181, 0x80001E1F00003589, 0x8000091380001E1B,
-    0x80001E2380001E27, 0x00006D8180000A90, 0x00007D8300006D80,
-    0x0000798000001184, 0x000125A800007981, 0x0000598980000A30,
-    0x00000D8480000A2D, 0x80001E03000161A5, 0x0001FD81000189A3,
-    0x800008BF0000E8CC, 0x800008C280000A51, 0x80000A5A80000A12,
-    0x000EF079000F2C7A, 0x000EA8770009A878, 0x000EB07500097876,
-    0x000EA07300095474, 0x80000A5D000E7472, 0x00098C6F000E2C70,
-    0x0009CC6D0009986E, 0x0009006B0009E46C, 0x000934690008AC6A,
-    0x000910670008D868, 0x0008B0650008F466, 0x0008D8630008D064,
-    0x000830610008C062, 0x0000058280001DFF, 0x80001E2B80001E17,
-    0x0000198680001CC3, 0x00085559000C015A, 0x00082D5700082D58,
-    0x00093D550008C156, 0x0009F1530009F554, 0x00061D5100098552,
-    0x0004394F00099550, 0x0004F54D0004554E, 0x0005154B0005214C,
-    0x00058D490005614A, 0x0005D1470005E548, 0x0005F9450005C946,
-    0x000401430005E544, 0x0001CD4100023942, 0x0002418180000916,
-    0x0000058700023D80, 0x80001CCB80001CC7, 0x0000598C80001CCF,
-    0x0000058280000919, 0x8000091F8000091C, 0x0000C8CC00003181,
-    0x0000C4CC00001D87, 0x800009220000C0CD, 0x0000158C80001CD3,
-    0x80001CD70001C4CC, 0x0000458100001181, 0x80001CDB80000925,
-    0x80001CE300003587, 0x80001CDF80001CE7, 0x00008982800008CB,
-    0x00000980800008CE, 0x80001E3F00001581, 0x80001E3B00001983,
-    0x000089A380001E47, 0x80001E3780001E43, 0x80001CEB00003989,
-    0x8000092E80001CEF, 0x800009310006B5A7, 0x000049800000CDB1,
-    0x800008D400005581, 0x0000C5B10000A1AD, 0x0000718080001E33,
-    0x8000093700007181, 0x80000A6000008DA7, 0x80001E4B80000A63,
-    0x0001F186000095A3, 0x000304CD0001D581, 0x80001CFB00017CCC,
-    0x80001CF380000934, 0x00016D8080001CF7, 0x0000058700016181,
-    0x80000A4B80001CFF, 0x8000093A00075981, 0x8000093D80001D03,
-    0x027150EF02955CF0, 0x80000A3980000940, 0x8000094680000943,
-    0x80001D07026A30EA, 0x0001F8CD80001D0F, 0x80000A870001E4CC,
-    0x024FB0E380001D0B, 0x002978E1023C00E2, 0x000144CC0025ECE0,
-    0x00074982000154CD, 0x0022B4DB000150CC, 0x002400D980001D13,
-    0x80001D17002358D8, 0x002228D5000BA8D6, 0x0020B0D30023A0D4,
-    0x001EF4D10021B8D2, 0x001C00CF001D40D0, 0x0017C8CD0016D4CE,
-    0x0015E0CB8000067E, 0x0015DCC9001598CA, 0x0012E0C7001400C8,
-    0x0011E8C50013E0C6, 0x000EF0C3001000C4, 0x800008D7000F74C2,
-    0x800008C8800008DA, 0x800006B10006E991, 0x800006D180001E2F,
-    0x800008D1800008DD, 0x8000092B000751A7, 0x80000A938000094C,
-    0x800006F4800006DA, 0x800007088000094F, 0x0006358080000952,
-    0x0000518200063981, 0x000071840006ED83, 0x0006E9868000071D,
-    0x0006A9880006E587, 0x800009580006B189, 0x0006918C800008E0,
-    0x80001E4F80001D1F, 0x000795B00006958F, 0x0007B9820007AD81,
-    0x0007A58480000A15, 0x0007A18680000A66, 0x80000A690007D987,
-    0x8000095580001E53, 0x0007F98C80001D1B, 0x0007418080000735,
-    0x8000095E00074181, 0x80001D2380000737, 0x80000A3C0000C1A3,
-    0x8000096180001D27, 0x0007EDA80000F5A7, 0x0000218180001D2B,
-    0x800009648000076D, 0x8000096A0007E5AD, 0x0006518280001D2F,
-    0x000005840001A591, 0x0001D98180001D33, 0x0001E18C00064D87,
-    0x80001D3B80000967, 0x000675880001F58C, 0x80001D370001DD81,
-    0x0006518C0001C581, 0x0006B98080000787, 0x00015D820006B981,
-    0x0001658400016983, 0x0001658680001D3F, 0x00010D8800013587,
-    0x80001D4300011D89, 0x0001FD8C00003981, 0x80001D4700001D87,
-    0x00012DB00001E98F, 0x000151A70006E9A3, 0x800007D90000D981,
-    0x00014DA3000601A7, 0x80000A4E0000BDAD, 0x80000970000159A3,
-    0x000611AE800008E6, 0x80001D4B0000A9A7, 0x80001D4F80000976,
-    0x80001D5780000973, 0x000135B1000169A3, 0x0000798080001D53,
-    0x000149A800007581, 0x800008E900009DA3, 0x800006870000DDB1,
-    0x800007F3800008EC, 0x0003B98780001E63, 0x000180CD80001E5F,
-    0x0003BD83000164CC, 0x0003B1810000E191, 0x80001E6B0003BD80,
-    0x80001D5B80001E67, 0x80000A990003918C, 0x80001D6380001D5F,
-    0x000E4984000E259B, 0x0000218600006D81, 0x0003B980000E5D87,
-    0x000C01820003A181, 0x0000818C000E5D83, 0x8000097F80001D67,
-    0x000E65880000858F, 0x80000A9F0000A589, 0x000345A70000A98B,
-    0x000375800000A583, 0x000351A300008981, 0x000365800000ED84,
-    0x0002BCCC00009981, 0x000361AD00017CCD, 0x0000898000008989,
-    0x000121810000D181, 0x0006E8CC80000A96, 0x000298CC000698CD,
-    0x80001E5B000358CD, 0x00030DB180000982, 0x80000A18000EFDA3,
-    0x80000A6C00005984, 0x000ED5A800014581, 0x80001E7780000A6F,
-    0x000180CC80001E73, 0x000000CC0007FCCD, 0x0007C8CC80001E7F,
-    0x0007F8CC0007CCCD, 0x000160CC000620CC, 0x0007A8CC00063CCD,
-    0x000760CC0007ECCD, 0x000740CC000764CD, 0x0001718400074CCC,
-    0x0001B18000004983, 0x80001E7B0001B181, 0x000670CD00017D81,
-    0x0001D8CD000674CC, 0x0000E982000024CC, 0x0001458800003D89,
-    0x0001858180001E83, 0x80001E6F00018580, 0x00000584000604CC,
-    0x0002658180000A42, 0x80001D6B00027D80, 0x0000398180001D6F,
-    0x8000098880000823, 0x8000098E80001D73, 0x0000258780000A72,
-    0x0000098480000A75, 0x80001D7B0000B1A3, 0x8000098B0000D4CC,
-    0x0000E8CD80001D7F, 0x80001D8B0000F4CC, 0x80001D8380000994,
-    0x00011CCD80001D8F, 0x000019870000E0CC, 0x800008F200001187,
-    0x8000097C000FB991, 0x80000A9C00001187, 0x800009E5800008F5,
-    0x80001E5780001D93, 0x800007F880000A3F, 0x8000082580000A7E,
-    0x80001D7780000997, 0x8000083F80000991, 0x8000099A000FED81,
-    0x00006D8180001D87, 0x8000085300001D8C, 0x00005D81000FF187,
-    0x800009A080001D97, 0x000FC18C000F8982, 0x000F918780001D9B,
-    0x80000A81000FC98F, 0x8000099D000F51B1, 0x80001D9F80001DA3,
-    0x800008FB80000860, 0x80000901800008FE, 0x00000581800009A6,
-    0x800009A980001DB3, 0x80001DB700000D88, 0x80000904800009AC,
-    0x80000A2700002584, 0x000009810000F5A3, 0x80000A2100001D80,
-    0x80000A1E000F0DA7, 0x0000A5A30000098C, 0x0000398080000A24,
-    0x0000CDA700002D81, 0x000165870000D1A6, 0x800009AF80001E8B,
-    0x80000A1B800009B2, 0x80000A7B80000A78, 0x80001E9300002D89,
-    0x80001E9B80001E8F, 0x80001E970001458C, 0x0000198000001D83,
-    0x800009F700001981, 0x0000698080001E9F, 0x80001E8700006981,
-    0x800009B580001DA7, 0x80001DAB80001DAF, 0x000DE0CC80000862,
-    0x0001E980000DF4CD, 0x000005830001E981, 0x000109A780001DBB,
-    0x80001DBF000105A6, 0x0001E5A380000876, 0x80001DC780001DC3,
-    0x80001DCF800009B8, 0x000125AD80001DCB, 0x8000088180001DD3,
-    0x00002D800000F1A3, 0x0000058700002581, 0x80001DDB80001DD7,
-    0x80001EA38000088A, 0x000149B180000907, 0x80001EAF800009BB,
-    0x80001DDF80000AA2, 0x80001EAB8000090D, 0x000071A380001EA7,
-    0x00029D8100003588, 0x0002AD8300029D80, 0x002841870002AD82,
-    0x800009BE0002A584, 0x000325890002A186, 0x0003258B0002E588,
-    0x80001DE30003258A, 0x00032D8F0003258C, 0x00035991800009C1,
-    0x80001DE7800009C4, 0x0000498180001DEB, 0x800008EF000120CC,
-    0x0003B1800003F1A3, 0x0003299B0003B181, 0x000148CC80001DEF,
-    0x000800CC00013CCD, 0x00015CCC000A78CD, 0x0003D9A3000A90CC,
-    0x80000A48000AF4CC, 0x000AB0CC0003C1A4, 0x000B9CCC000AC0CD,
-    0x80001DF30003FDA8, 0x0003EDAD00196D94, 0x000100CD0004B8CC,
-    0x80001DF700026CCC, 0x00021CCC0003A5B0, 0x0003458000036D87,
-    0x0003458200034D81, 0x80001DFB000E1D82, 0x000CA8CC0000758C,
-    0x00035588000BF4CD, 0x000298CD80000000, 0x000140CC00033CCC,
-    0x000ED8CC000FECCD, 0x000E70CC000EE8CC, 0x000EBCCC000E8CCC,
-    0x000FA4CC000FA0CD, 0x000E60CC000FD0CC, 0x000E60CC000E50CD,
-    0x000F88CC000F74CD, 0x000DC0CC000F8CCD, 0x000454CC000BBCCD,
-    0x8000001A0002B9A3, 0x000DE0CD8000000A, 0x000C74CC000504CC,
-    0x000074CC000D88CD, 0x000114CC000018CC, 0x000E98CC000120CD,
-    0x000114CC00016CCC, 0x80000482000058CC, 0x00013CCD80000521,
-    0x000CB0CC000138CC, 0x0000D0CD000CACCD, 0x800000020003A0CC,
-    0x000C38CD80000C13, 0x0010D180000CDCCC, 0x0002118080000064,
-    0x0002158200021181, 0x0002358400022D83, 0x8000044F8000035F,
-    0x00020D8800023587, 0x000C30CD00021589, 0x000124CC000FF0CC,
-    0x000E6CCD000104CD, 0x00061981000FACCC, 0x80000443000F5582,
-    0x000F498700063D82, 0x0006258780000553, 0x000984CC0008BD91,
-    0x800008B9000898CD, 0x800008BF000F7188, 0x800008C20006718C,
-    0x00087D86000F558C, 0x00087D8480000A54, 0x0008458280000A33,
-    0x000FA18000087183, 0x0005CDA3000FA181, 0x0008E58C0008F18F,
-    0x000630CC00002981, 0x0008158800089CCD, 0x000FEDA300080589,
-    0x0006C1A3800008C5, 0x000F05A7800008C8, 0x000835B0800008D1,
-    0x0006C4CC800008E0, 0x0004ACCC0006F0CD, 0x800008E30007DCCD,
-    0x000254CC000F15AE, 0x00068DB1000704CD, 0x0006A5B10006058A,
-    0x800008EF00056988, 0x00039CCC000871A3, 0x0000A5810004BCCD,
-    0x0005718100057982, 0x0005518700057980, 0x000851A80000B984,
-    0x00056587800008F2, 0x00030CCC800008F5, 0x00000581000558CD,
-    0x8000090480000A57, 0x800008F88000090A, 0x8000092880000913,
-    0x0003D4CC0005458C, 0x8000092B00035CCD, 0x0004958100079588,
-    0x0004A98300049580, 0x8000094900049182, 0x0004B1870004B184,
-    0x0004918980000726, 0x8000078300048988, 0x000264CC00078D8A,
-    0x8000096D000260CD, 0x000509A780000BB6, 0x80000979000505A6,
-    0x0005E5A38000097C, 0x800009858000088C, 0x8000099A80000991,
-    0x000525AD800009A3, 0x800009A9800009A6, 0x800009CA80000ACC,
-    0x80000ACF800009CD, 0x00043DA3800009D0, 0x80000AD880000AD5,
-    0x000549B1800009D3, 0x80000ADE80000A2A, 0x800009D680000AE1,
-    0x80000AF680000AED, 0x80000B0580000B08, 0x0004F5A0800009DC,
-    0x80000B1A0000918C, 0x000104CC000238CC, 0x80000B23000100CD,
-    0x000571A880000B2C, 0x000E1DAA800009E5, 0x800009EB800009E8,
-    0x001E958580000B47, 0x001EF585000569AF, 0x0005C1B2800009EE,
-    0x0005F9B40005C1B3, 0x0015D5880005F9B5, 0x0005DDB880000B50,
-    0x000B91BA0005B1B9, 0x0005A9BC800009F4, 0x00026DBE00027DBD,
-    0x800009F780000B59, 0x80000B6280000B5F, 0x800009FD800009FA,
-    0x80000A0080000A45, 0x0006F4CD80000A03, 0x800006CD0000A0CC,
-    0x800007E18000076F, 0x80000A3F80000A36, 0x800009D9800006CA,
-    0x80000A8480000A06, 0x80000A8A800009E2, 0x80000A9080000A8D,
-    0x80000A9C80000A93, 0x800022A3001E5985, 0x80000AA8001E2185,
-    0x800022A7800009DF, 0x000F398080000AB1, 0x00032182000F4581,
-    0x00032184000F9983, 0x00039D8600032D85, 0x000D558800039187,
-    0x0003BD8A000D5989, 0x000CC98C000DE58B, 0x000B2D8E000CC98D,
-    0x0003F1900003A18F, 0x000AA192000AE991, 0x00038194000AB993,
-    0x00025D960003A995, 0x00024198800009C7, 0x0007F59A0007FD99,
-    0x0002699C0007CD9B, 0x0002659E0006A19D, 0x80000B6580000B5C,
-    0x0014798880000AB4, 0x001F458580000AB7, 0x000198A680000ABA,
-    0x80000AC080000ABD, 0x80000B3880000AFF, 0x0001D8CD80000B3E,
-    0x80000B4A000150CC, 0x8000001680000012, 0x8000006880000030,
-    0x8000003480000006, 0x80000BC080000B7A, 0x0001A0B880000BC3,
-    0x0012258880000BCB, 0x80000C0A80000BC6, 0x80000BD180000BCE,
-    0x001D698080000BB3, 0x80000B77001D7D81, 0x80000BD40015A583,
-    0x800000740015D185, 0x8000063D0015A187, 0x001DA18A001D5989,
-    0x001DBD8C80000C90, 0x0016518E80000708, 0x0016959080000735,
-    0x001DD19280000876, 0x0017F99480000024, 0x0017899680000BE3,
-    0x0017D198800008AD, 0x001DF59A80000BE6, 0x0017959C80000BE9,
-    0x00184D9E80000BEC, 0x001D21A080000C60, 0x001801A280000C63,
-    0x0018F9A480000C66, 0x001D3DA680001F8F, 0x001D0DA88000200F,
-    0x001D19AA8000200B, 0x00180DAC80001FB3, 0x0019B1AE80001FBB,
-    0x0019CDB080001FC3, 0x001991B280001FAB, 0x00194DB480001FB7,
-    0x001905B680001FBF, 0x00190DB800002D85, 0x00014985001AFDB9,
-    0x80001FC7001AD1BB, 0x80001FAF001AB1BD, 0x00141DBB0018BDBF,
-    0x0015F9B98000088A, 0x001C79BF80002017, 0x0015E1BD80000BF5,
-    0x001C7DB380000BF8, 0x00001980001C79B2, 0x80001EC700006581,
-    0x80000BFB0015EDB6, 0x80001F930015FDB4, 0x80001ECF001571AA,
-    0x80001EE7001569A8, 0x80000BF20015FDAE, 0x80001EEF001561AC,
-    0x80001EEB001229A2, 0x80001EF3001229A0, 0x80001ED7001C25A6,
-    0x80000C0100122DA4, 0x80001F970013659A, 0x80001ECB00128198,
-    0x80001EBF0012AD9E, 0x80001FD30012AD9C, 0x00001985001C9D92,
-    0x80001FDB001C8590, 0x0000798500130D96, 0x80001FE300130D94,
-    0x80001FCB0013ED8A, 0x000009850013E988, 0x80001FD70013898E,
-    0x80001ED300139D8C, 0x80001EDB001CC182, 0x80001EFF00114D80,
-    0x80001FDF0013E986, 0x001FB58100100184, 0x0003F185001E28CD,
-    0x80001FE7001FB582, 0x001FDD87001FA984, 0x001FED89001FDD86,
-    0x001FFD8B00003585, 0x001A50CD001FFD8A, 0x001FD98F001D70CC,
-    0x001FA991001FE18E, 0x001FA593001FA190, 0x80001FCF001EBCCD,
-    0x001FA197001FA594, 0x8000202B001FA996, 0x000648CC001F9998,
-    0x001C699D0004ACCD, 0x001C659F001C7D9C, 0x80000C07001E28CD,
-    0x80001F9B001CA1A0, 0x8000203F001CA1A2, 0x001CDDA7001CA5A4,
-    0x001CFDA9001CA9A6, 0x000180CC001E5CCD, 0x8000203B000244CD,
-    0x001C39AF001CD9AC, 0x001C45B1001CC1AE, 0x001C45B3000038CD,
-    0x001C41B5001C45B2, 0x0002D1B7001F50CD, 0x00039982000014CD,
-    0x0003D980001C7DB8, 0x80000C5A0003A581, 0x80001F1F001C69BC,
-    0x80001F2700023985, 0x80001F2F001D08CC, 0x001DB984001BD8CC,
-    0x001DB186001DBD85, 0x001D8D88001DB187, 0x001D998A001D8989,
-    0x001D818C001D9D8B, 0x001F0CCC0012398D, 0x80001F230017B18F,
-    0x80001F2B0015A591, 0x80001F33001A5D93, 0x80000C0D001BB595,
-    0x80000C10001B8597, 0x80000C36001B8999, 0x0018C8CD001BA59B,
-    0x00139D9E001D98CC, 0x00135DA080000C39, 0x001FC5A280001F9F,
-    0x001D25A480001F3F, 0x0010D1A680001F47, 0x001771A880001F43,
-    0x001D11AA80001F4B, 0x00143DAC0006F585, 0x001D2DAE80000C3F,
-    0x8000205F0006DD85, 0x001D75B2001D79B1, 0x0010F9B4001D71B3,
-    0x001D19B680000C45, 0x001641B8001D19B7, 0x0012EDBA80000C48,
-    0x001FB5BC00085180, 0x001EEDBE80001FA3, 0x001AFD8180002053,
-    0x001B558300193892, 0x001B918580000C42, 0x00016985001B6984,
-    0x0001B185001B9986, 0x80001EC3001BA988, 0x8000204F001BB98A,
-    0x80000C5D001BA98C, 0x001B8D9C001B9D8E, 0x0005D980001B959D,
-    0x001BED980005AD81, 0x001B9D9A001B9599, 0x0005F180001B959B,
-    0x0005818200058581, 0x80001F6380001F5B, 0x0005298000048D85,
-    0x001CB1B300051D81, 0x001B2DB10004B9B2, 0x0004D5B70004B9B0,
-    0x001B35B5001B2DB6, 0x001B75A4001B35B4, 0x80000C4E80001F6B,
-    0x001941A00005A9B8, 0x001CF9A2001059A1, 0x00045580001D0DA3,
-    0x80000BFE00049181, 0x80001EDF00055583, 0x0005B1800005D985,
-    0x0005A1800005AD81, 0x80001EE300048181, 0x0005018000003D83,
-    0x80000C5100000981, 0x0001FD8180001F07, 0x80001F0F0004F580,
-    0x0004D98500000582, 0x80000C04001CF1B2, 0x80000C54001365B0,
-    0x001BDDBD001BDDBE, 0x001BC1BB80001FA7, 0x00007980001BDDBA,
-    0x80001F0300005581, 0x80001EF7001B9DA6, 0x80001EFB001B99A4,
-    0x80001FF3001B85A2, 0x80001FFB001B91A0, 0x80001F0B001B85AE,
-    0x8000200300135DAC, 0x80001FEB0013C9AA, 0x80001F13001BA1A8,
-    0x80000C0A001C9D96, 0x80001FF7001C9994, 0x80001F1700122992,
-    0x0000098500122990, 0x80001FFF0016CD9E, 0x00000985001B659C,
-    0x80002007001D359A, 0x80001F77001D9198, 0x80001FEF00172986,
-    0x80001F1B00172584, 0x8000206F0014ED82, 0x80001F7F0014CD80,
-    0x80001F870013118E, 0x80001F7B0013058C, 0x80001F830011558A,
-    0x0019D98000116988, 0x000051850019D981, 0x0019FD840019E983,
-    0x000759800019E585, 0x0006B58000067981, 0x000288CC00067181,
-    0x0006AD840002ACCD, 0x00069D8600083D88, 0x00032981000590CD,
-    0x0002A19400030580, 0x000401850003CD82, 0x8000201300028593,
-    0x000158CC000640CD, 0x0000E185000220CD, 0x000259940007F593,
-    0x000088CC0005B0CD, 0x00070980000178CD, 0x8000201B0006E981,
-    0x000021850003C583, 0x000461800004F585, 0x0007298000059D81,
-    0x000230CC0006C981, 0x8000202300030CCD, 0x000C00CC000684CD,
-    0x001935B0000CACCD, 0x001931B20000F0CD, 0x00192DB4000178CD,
-    0x0006C9B600077CCD, 0x0003F4CC0000D0CD, 0x0006F1BA000CE4CD,
-    0x0003A1940003A593, 0x0006D9BE00074CCD, 0x001CA9840006D1BF,
-    0x00075D8600074985, 0x0004598100075187, 0x000150CC00045580,
-    0x00075D8C00017CCD, 0x0007AD8E00007585, 0x000755880007A18F,
-    0x00074D8A00075989, 0x00067994001FC0CC, 0x0006C99600038D95,
-    0x8000202700025597, 0x00064D9200060D91, 0x001CCD9C00064193,
-    0x0004159E00040D9D, 0x0005F198000F619F, 0x0004219A0002FD99,
-    0x000465A400041D9B, 0x000321A6000FCDA5, 0x000425A0000375A7,
-    0x00003585000399A1, 0x00055980000449A3, 0x0004A1820004A581,
-    0x000291A88000202F, 0x000549AA000FC1A9, 0x00001185000385AB,
-    0x000474CD000DACB5, 0x80002037000D9CB7, 0x000100CC000DF8B1,
-    0x000198CC000104CD, 0x8000204300019CCD, 0x80000BE0000D04BF,
-    0x0003E0CC000DBCB9, 0x000B418000011CCD, 0x000B5982000B4D81,
-    0x000A998080000C3C, 0x80001F3700086181, 0x000BA580000D8583,
-    0x000BA58000085981, 0x80001F3B000BB981, 0x000BA5808000204B,
-    0x000BC580000B9981, 0x80000C4B00083981, 0x80001F5300005583,
-    0x000C358080002047, 0x000C0D80000C2D81, 0x000CCD82000C1581,
-    0x0000419400007593, 0x000B8D8080000C57, 0x000805810008F981,
-    0x000DA580000B7180, 0x80001F6F00080181, 0x000A6988000B6984,
-    0x000A1186000CAD86, 0x000A3984000C9D88, 0x000D718080001F73,
-    0x000A3180000D5581, 0x000D29820008C981, 0x000D098000005593,
-    0x80001F8B000D2D81, 0x0000759400008594, 0x000D1180000DA585,
-    0x000D3994000D1D81, 0x00002182000D2585, 0x000D0980000D6D93,
-    0x000C2D80000D0581, 0x80001F8F00098981, 0x80000BE300019D83,
-    0x000258CD000CCD85, 0x0000298000023CCC, 0x8000201F00003581,
-    0x0009AD8480001F97, 0x00099D8680000BE9, 0x0000498080002033,
-    0x0000618200005581, 0x80000C5A0008D583, 0x000D40CC00085985,
-    0x0002E0CC000FBCCD, 0x0000DD930002FCCD, 0x000D59940008F593,
-    0x0002D0CC000AB594, 0x0000AD8000013CCD, 0x0000D9880000B181,
-    0x0000AD82000BED85, 0x0000B580000CDD83, 0x000E49800000A981,
-    0x000B2180000E5581, 0x000B01860009D981, 0x000B29848000205B,
-    0x000E298000000983, 0x0001FD82000E3581, 0x000C8194000C8593,
-    0x0003F8CC80000C5D, 0x8000205700039CCD, 0x0002CCCC000C7D93,
-    0x000170CC00025CCD, 0x000C2994000144CD, 0x000F8CCD80000BEF,
-    0x000FDCCD80002067, 0x0003F18080001F57, 0x000F28CD0003E981,
-    0x00002582000E98CD, 0x0003C980000E58CD, 0x80001F5F0003D181,
-    0x0000458580001F67, 0x0008C980000E20CD, 0x00021D82000A3181,
-    0x0000C585000FD583, 0x0000558080002063, 0x0001C0CC00006D81,
-    0x0002E0CC0001A4CD, 0x00007D80000294CD, 0x00004D8200004581,
-    0x000975808000206B, 0x00068180000A0181, 0x00096D8400068181,
-    0x000EA98680002073, 0x000375888000207B, 0x0000358580000C69,
-    0x0009458080000C6C, 0x0000ED82000A3181, 0x0001B5810006BD83,
-    0x0001718100025593, 0x000281940006CD93, 0x0002759480000C6F,
-    0x0002C8CD80000C72, 0x0004CD880002F4CC, 0x0006E182000F3585,
-    0x0006F18000026583, 0x000138CC0006F181, 0x80000C7500013CCD,
-    0x80000C7B80000C78, 0x80000C8180000C7E, 0x0005118880000C84,
-    0x000100CC0002FCCC, 0x80000C87000104CD, 0x000178CD80000C8A,
-    0x000040CD00011CCC, 0x80000C8D0001CCCC, 0x80000C9680000C93,
-    0x80000BDA80000BD7, 0x000138CC80000BDD, 0x80000CF9000104CD,
-    0x80000CFF80000CFC, 0x0001FCCC80000D02, 0x0005418F000104CD,
-    0x80000D0880000D05, 0x80000D1180000D0E, 0x000160CD80000D14,
-    0x000110CC00017CCC, 0x0004FCCC00016CCD, 0x003BA98880000D17,
-    0x80000D1D80000D1A, 0x00000D8680000D20, 0x80000D2680000D23,
-    0x80000DC880000DC5, 0x80000C9F80000C9C, 0x80000DCE80000CA5,
-    0x80000DAD000C40CD, 0x80000DDA80000DD7, 0x0000298880000DE0,
-    0x80001F4F80000CB4, 0x80001FA7000158BE, 0x8000207780000C66,
-    0x80000C420001ACBA, 0x000178B70001ACB8, 0x000140B50004BCB6,
-    0x000150B380000D0B, 0x80000CA280000C99, 0x000205AF000428B0,
-    0x003C4DAD000271AE, 0x003C51AB000261AC, 0x00024DA9000275AA,
-    0x003C01A7000251A8, 0x000269A5000261A6, 0x000199A3000195A4,
-    0x0001A5A1000195A2, 0x0001519F0001A5A0, 0x00013D9D0001459E,
-    0x0001FD9B0001F19C, 0x000765990001C19A, 0x000181970001B198,
-    0x0001A5950001BD96, 0x0000899300077194, 0x00076D9100076D92,
-    0x0007098F00008990, 0x0007758D0007018E, 0x0007658B0007758C,
-    0x00074D890007798A, 0x00076D8700074D88, 0x000761850000C986,
-    0x0007858300076984, 0x0007958100078582, 0x000E59B000079580,
-    0x000EDDB200011DB1, 0x000F91B480000CA8, 0x80000CAB0009DDB5,
-    0x80000CAE000651B7, 0x000679BA000EF1B9, 0x80000DE380000CB1,
-    0x000661BE000661BD, 0x00064DA0000661BF, 0x000659A200062D80,
-    0x000645A40006A586, 0x000641A680000CB7, 0x00067DA80006BD86,
-    0x000659AA00060988, 0x000645AC80000CBA, 0x000631AE00069988,
-    0x000C959080000CBD, 0x0005C192000F0591, 0x0009A994000D0993,
-    0x000E2D9600093D95, 0x00070D980000AD88, 0x00077D9A00065986,
-    0x00076D9C00065984, 0x0006B19E80000CC0, 0x80000CC300060180,
-    0x000E1D8200071C81, 0x80000CC680000DE6, 0x80000CC900071485,
-    0x80000CCF80000CCC, 0x80000CD20004BC89, 0x80000DF580000DF2,
-    0x80000CD580000DF8, 0x80000CD80007598F, 0x80000DFB80000CDB,
-    0x0013B8CC80000CE1, 0x000FF4CD80000CE4, 0x80000CE70001A0CC,
-    0x00017CCC80000E01, 0x80000DEF80000CED, 0x001780CD80000CF3,
-    0x0003AD84000178CC, 0x00058D8680000CF6, 0x000694CC80000D29,
-    0x00011CCC000690CD, 0x0006ACCD0017E4CD, 0x0013C0CC0006A8CC,
-    0x000381880013A4CC, 0x000A6CCC000224CC, 0x00013CCC0003898B,
-    0x000644CC000F68CD, 0x000658CC000640CD, 0x0012E0CC000794CC,
-    0x80000D2F80000D2C, 0x80000D3580000D32, 0x80000D3B80000D38,
-    0x80000D4180000D3E, 0x80000D4A80000D44, 0x80000D5080000D4D,
-    0x80000D53000770CC, 0x000288CC80000D56, 0x80000D5C80000D59,
-    0x80000D6280000D5F, 0x001338CC80000D65, 0x80000D6B80000D68,
-    0x80000D6E003ED588, 0x0007CCCC0006DCCC, 0x80000D710007C8CD,
-    0x80000D740002BC96, 0x80000D77001364CC, 0x00040CCC000798CD,
-    0x000768CC00049CCD, 0x00077CCD000744CD, 0x00020483000710CD,
-    0x80000D7A0006D4CC, 0x0002188780000D7D, 0x80000D80000388CC,
-    0x0002C08B80000D83, 0x80000D8980000D86, 0x80000D8F80000D8C,
-    0x0002DC8D80000D92, 0x80000D9880000D95, 0x80000D9B000139B2,
-    0x80000D9E000135B0, 0x000264B5000155B6, 0x80000DA1003F95B4,
-    0x80000DA400016DBA, 0x80000DA7000169B8, 0x80000DAA000161BE,
-    0x003E88D900016DBC, 0x80000DC2000159A2, 0x80000DB0000181A0,
-    0x80000DB3000171A6, 0x80000DB600017DA4, 0x80000DB9000149AA,
-    0x80000DBC000145A8, 0x80000DBF000141AE, 0x80000CEA00014DAC,
-    0x80000D4780000CF0, 0x80000DCB003ED180, 0x80000DD480000DD1,
-    0x80000DE980000DDD, 0x80000DEC003EC58A, 0x80000E0480000DFE,
-    0x80000E07003EDD8E, 0x80000E0A003ED98C, 0x80000E0D003EB192,
-    0x80000E10003EA190, 0x80000E13003F5596, 0x80000CDE003EA594,
-    0x80000E16003F7D9A, 0x80000E19003F7D98, 0x80000E1C003F3D9E,
-    0x80000E1F003F5D9C, 0x80000E22003FDDA2, 0x80000E25003FC1A0,
-    0x80000E28003FF1A6, 0x80000E2B003FC1A4, 0x80000E2E003FC9AA,
-    0x80000E31003FC5A8, 0x80000E34003FC1AE, 0x80000E37003FCDAC,
-    0x80000E3A003F45B2, 0x80000E3D003FB1B0, 0x80000E40003F45B6,
-    0x80000E43003F55B4, 0x80000E46003F4DBA, 0x80000E49003F4DB8,
-    0x80000E4C003F4DBE, 0x003EC980003F4DBC, 0x80000E4F003AED81,
-    0x80000E52003EDD83, 0x80000E55003ECD85, 0x80000E58003EFD87,
-    0x80000E5B003ECD89, 0x80000E5E003EDD8B, 0x80000E61003ECD8D,
-    0x003A8D9080000E69, 0x003A999280000E6C, 0x0001419480000E6F,
-    0x003AB99680000E72, 0x0001759880000E75, 0x0001719A003D1099,
-    0x003A9D9C80000E78, 0x003A899E80000E7B, 0x00019DA080000E7E,
-    0x003A7DA280000E81, 0x003CA1A480000E84, 0x003C55A680000E87,
-    0x000181A880000E8A, 0x000181AA003DC0A9, 0x003F81AC80000E8D,
-    0x003C55AE80000E90, 0x003C31B080000E93, 0x003C3DB280000EAA,
-    0x003FD9B480000EAD, 0x0001F5B680000EB0, 0x003FC5B880000EB3,
-    0x0001D9BA80000EB6, 0x0001C9BC80000EB9, 0x0001B9BE80000EBC,
-    0x80000EC280000EBF, 0x80000EC500025582, 0x80000ECD00024580,
-    0x80000ED000027D86, 0x80000ED300024584, 0x80000ED600025D8A,
-    0x80000ED900025D88, 0x80000EDC0002BD8E, 0x80000EDF0002BD8C,
-    0x80000EE20002DD92, 0x80000E640002DD90, 0x003EFCA70002FD96,
-    0x80000FBC0002FD94, 0x80000FBF0002DD9A, 0x80000FC50002DD98,
-    0x0002E4D90002BD9E, 0x80000FC20002BD9C, 0x80000FC800025DA2,
-    0x8000103A00025DA0, 0x8000165E00027DA6, 0x8000172D00027DA4,
-    0x800016BC00025DAA, 0x8000172100025DA8, 0x800017440003A5AE,
-    0x0003E9B30003B5AC, 0x0003D1B10003E5B2, 0x000395B700004D94,
-    0x0003E1B50003E9B6, 0x00038DBB0003EDB4, 0x0003BDB90003B9BA,
-    0x00037DBF0003A1B8, 0x000385BD000371BE, 0x000258D9000389BC,
-    0x0001819400016593, 0x003FCCCC8000171B, 0x000171940002CCE0,
-    0x00036C8100016D95, 0x00023CD9003AB4CC, 0x003BF8CC80001771,
-    0x003A84CC003A34CC, 0x000244D90002C4E0, 0x0002D8A40001F594,
-    0x003744CC80001783, 0x000254D90001ED94, 0x00004DBC00005194,
-    0x0002E0A4000214E0, 0x00034C9500034892, 0x000035BC80001787,
-    0x0002158080001747, 0x0002258200021981, 0x00022D8400023983,
-    0x0002558600025185, 0x00027D8800024987, 0x0002458A00024189,
-    0x00024D8C0002598B, 0x0001B18E0001B58D, 0x0001D9900001AD8F,
-    0x0001E1920001E591, 0x0001E9940001FD93, 0x0001919600019595,
-    0x0002CCA40000A9BC, 0x800017558000174E, 0x800017638000175C,
-    0x800017758000176A, 0x0002D4A48000177C, 0x0002A8E00002E4A5,
-    0x00000DBE0002B0A7, 0x0004FCB30004A8B0, 0x00006197800017A0,
-    0x000370AE000338AD, 0x000258A800031CAF, 0x0002D0A60002B0A9,
-    0x800017A4000240A7, 0x8000178B000394B5, 0x8000179200002CB7,
-    0x00032CB200032CB1, 0x00004CBC0003DCB3, 0x000174BE000164BD,
-    0x000090B880001799, 0x000088BA0004E0A8, 0x0007FD98000044BB,
-    0x0007619A00076199, 0x0007619C0007619B, 0x0007619E0007619D,
-    0x000328870007619F, 0x800017C400000DB3, 0x00001DB6800017CB,
-    0x800017AF800017A8, 0x0000AD9F800017B6, 0x0000999D800017BD,
-    0x0002BCE00000859C, 0x000270AD000031BE, 0x000055990002F0AC,
-    0x0000559B800017E4, 0x000009960000559A, 0x800017E00001F197,
-    0x000250870000399E, 0x800017D2800017E8, 0x00038492800017D9,
-    0x000258AF0002B0E0, 0x800017EC00003D97, 0x00038C860002A8E0,
-    0x000224AF0003F487, 0x000005BE0002F4AE, 0x0007D9B6800017F0,
-    0x000019970007E5B7, 0x800017F80007DDB5, 0x0002B4E00002C8AE,
-    0x00039086800017F4, 0x0000659D000039BE, 0x0002C0E00000659C,
-    0x0007D1B8000250B1, 0x800017FC0000FD96, 0x00038CBF0002D8E0,
-    0x000248B300001995, 0x0003848680001800, 0x00020CB30002F0E0,
-    0x8000180C00000D82, 0x0005D48A00024CB3, 0x0004A48800000995,
-    0x8000180480001810, 0x0003988680001808, 0x0002D8E00003BC87,
-    0x0002F4B4000035BE, 0x80001814000264B5, 0x8000181C00000D97,
-    0x0002CCE00002E8B4, 0x0000399500003196, 0x800018180003FDBE,
-    0x0002D8E000038C99, 0x80001820000204B7, 0x00003D8A000224B7,
-    0x0000318A0000458C, 0x000148E080001828, 0x8000182C000015B3,
-    0x80001830000C018F, 0x8000184500007DB3, 0x00000D9C80001837,
-    0x000220E10000199D, 0x8000184C8000183E, 0x8000185E80001857,
-    0x0003BCA580001865, 0x0002A480000089AE, 0x8000186C0000459F,
-    0x80001873000EC885, 0x80001850000F4884, 0x00029CEF8000187A,
-    0x80001881000FD884, 0x000260E3800018C3, 0x0002C0E1000F6CBE,
-    0x800018D4800018B9, 0x80001888800018CA, 0x800018960008B0BE,
-    0x800018A48000189D, 0x800018B2800018AB, 0x000994BE0060E885,
-    0x000AA8BE8000188F, 0x000BA4BE800018DB, 0x006178BF80002397,
-    0x8000239F8000239B, 0x800023A7800023A3, 0x000B448400090C84,
-    0x800023AB00015183, 0x0001A1810014DCBE, 0x800023AF001710BE,
-    0x800023B3000AA484, 0x0011BCBE004400BF, 0x800023B70001558D,
-    0x0010CCBE000000CF, 0x00012D92800023BB, 0x00019193800023BF,
-    0x00149484800023C3, 0x00146C8400012597, 0x0001AD9D800023C7,
-    0x0001F19C00166C84, 0x0010F8840011FC84, 0x00014DA7800023CB,
-    0x000FB486800023CF, 0x800023D3000C2C87, 0x000ED08700015DA2,
-    0x800023D700093C86, 0x800023DB00012DA9, 0x0012B08400017DAC,
-    0x0010F4BE800023DF, 0x0013C0BE800023E3, 0x800023EB800023E7,
-    0x001D98BE000141B3, 0x000145B600014DB5, 0x000151B8000171B7,
-    0x800023EF000151B9, 0x800023F3001DA084, 0x000125B9800023F7,
-    0x000CFC80800023FB, 0x00038882800023FF, 0x0003D08400024483,
-    0x8000240780002403, 0x8000240F8000240B, 0x8000241780002413,
-    0x8000241F8000241B, 0x8000242780002423, 0x00007182021B5C8F,
-    0x0000618000007183, 0x8000242B00006181, 0x0000618400003587,
-    0x8000242F00000D85, 0x8000190F80002433, 0x80003D4380003D3F,
-    0x80003D4B80003D47, 0x80003D4F0000198D, 0x80003D5780003D53,
-    0x80003D5F80003D5B, 0x80003D6780003D63, 0x80003D6F80003D6B,
-    0x80003D7780003D73, 0x0218A4AC80003D7B, 0x0000D1BC022D0CAD,
-    0x80003D8380003D7F, 0x0219C8B280003D87, 0x0218A0B480003D8B,
-    0x021EB0B6021C00B5, 0x021FE0B880003D8F, 0x021DA4BA021FE0B9,
-    0x0212D0BC021000BB, 0x0212F8BE0213D4BD, 0x000FEDA00211D4BF,
-    0x000FF1A2000FF1A1, 0x000E11A4000FF1A3, 0x000E39A6000E05A5,
-    0x000E19A8000E35A7, 0x000E65AA000E61A9, 0x000E55AC000E79AB,
-    0x000EBDAE000E49AD, 0x000EE1B0000EADAF, 0x000EFDB2000EEDB1,
-    0x000E95B4000EF1B3, 0x000EB1B6000E9DB5, 0x000E95B8000EADB7,
-    0x00016DBA000E99B9, 0x000169BC000179BB, 0x000159BE000169BD,
-    0x01F49490000159BF, 0x023F9C9201D28491, 0x80003D9780003D93,
-    0x80003D9F80003D9B, 0x00003D8080003DA3, 0x0000198200000981,
-    0x80003DAB80003DA7, 0x80003DB3000D7CE0, 0x000CEC8080003DB7,
-    0x00757C820069C481, 0x00BC50840050BC83, 0x00E70C86009B4485,
-    0x0129708800CDDC87, 0x0110E88A01356489, 0x015BE48C017F6C8B,
-    0x018CF88E01A1D08D, 0x001221AD01EB988F, 0x001401AF001001AC,
-    0x001E15A90012E1AE, 0x001F49AB001C01A8, 0x001801A5001E15AA,
-    0x001A0DA70006B5A4, 0x00025DA1001A0DA6, 0x000619A380001824,
-    0x80003DBB000401A2, 0x0011448480003DBF, 0x80003DC780003DC3,
-    0x80003DCF80003DCB, 0x006801B580003DD3, 0x80003DD7006A59B4,
-    0x001645B180003DDB, 0x006A59B3001655B0, 0x000211AF006801B2,
-    0x000201AD000209AE, 0x000211AB000209AC, 0x0002EDA90002EDAA,
-    0x80003DDF0002EDA8, 0x80003DE780003DE3, 0x80003DEF80003DEB,
-    0x80003DF780003DF3, 0x000359BF80003DFB, 0x000359BD000351BE,
-    0x00020DBB000341BC, 0x00021DB9000219BA, 0x00025DB700021DB8,
-    0x000241B5000249B6, 0x000251B3000249B4, 0x000261B1000269B2,
-    0x000379B3000269B0, 0x80003E0380003DFF, 0x80003E0B80003E07,
-    0x00039DB500039DB6, 0x0003B9BB80003E0F, 0x80003E130003B1BA,
-    0x000399BF80003E17, 0x000399BD000391BE, 0x000015820003A1BC,
-    0x0000618000033DA3, 0x0003D9A600006181, 0x80003E2380003E1B,
-    0x0003F5AA0003D9A5, 0x0004C8860003FDAB, 0x0003D5AE0005E087,
-    0x0003E5AC0003DDAF, 0x0001C1B30003DDAD, 0x80003E2B80003E27,
-    0x80003E3380003E2F, 0x0001D5B50001C5B6, 0x0001C9BB80003E37,
-    0x80003E3B0001F9BA, 0x000119BF80003E3F, 0x000119BD000111BE,
-    0x001340BE0001D1BC, 0x80003E4780003E43, 0x000F8CEF80003E4B,
-    0x80003E5380003E4F, 0x000F74E380003E57, 0x000E78E180003E5B,
-    0x80003E6380003E5F, 0x80003E6B80003E67, 0x0005B9A880003E6F,
-    0x000649AA0005B9A9, 0x00065DAC000649AB, 0x00064DAE000645AD,
-    0x80003E1F000645AF, 0x80003E7780003E73, 0x80003E7F80003E7B,
-    0x80003E8780003E83, 0x0006F5B880003E8B, 0x0006F5BA0006FDB9,
-    0x000759BC000749BB, 0x000749BE000759BD, 0x00063DB0000741BF,
-    0x000605B200060DB1, 0x0006D5B40006CDB3, 0x0006D5B60006DDB5,
-    0x0006C1A30006CDB7, 0x0000B1810000B982, 0x80003E8F0000B980,
-    0x0006D5A50006C5A6, 0x0006C9AB80003E93, 0x000008860006F9AA,
-    0x000619AF00030087, 0x000619AD000611AE, 0x80003E970006D1AC,
-    0x80003E9B00074DB3, 0x00074DB680003E9F, 0x80003EA780003EA3,
-    0x000725BA00074DB5, 0x80003EAB00072DBB, 0x000729BE80003EAF,
-    0x000735BC000729BF, 0x000659A8000729BD, 0x000681AA000681A9,
-    0x000681AC000681AB, 0x0006B1AE000689AD, 0x80003EB30006B1AF,
-    0x80003EBB80003EB7, 0x80003EC380003EBF, 0x80003ECB80003EC7,
-    0x0006B1B880003ECF, 0x0006B1BA0006B1B9, 0x000175BC0006B1BB,
-    0x000175BE00017DBD, 0x0006A5B000016DBF, 0x0006A5B20006ADB1,
-    0x0006A9B40006B9B3, 0x00069DB60006A9B5, 0x0009C4EF000691B7,
-    0x80003ED780003ED3, 0x80003EDF80003EDB, 0x80003EE780003EE3,
-    0x80003EEF80003EEB, 0x0000158100006182, 0x000934E300001980,
-    0x00083CE180003EF3, 0x000649A380003EF7, 0x001F688680003EFB,
-    0x80003F030000CC87, 0x000649A5000649A6, 0x000629AB80003F07,
-    0x80003F0B000621AA, 0x00062DAF80003F0F, 0x00062DAD00062DAE,
-    0x00029DA8000631AC, 0x0002C1AA0002C5A9, 0x0002C5AC0002C1AB,
-    0x0002F1AE0002CDAD, 0x80003F130002F1AF, 0x80003F1B80003F17,
-    0x80003F2380003F1F, 0x80003F2B80003F27, 0x0003C5B880003DAF,
-    0x0003C5BA0003CDB9, 0x0003C5BC0003DDBB, 0x0003C5BE0003CDBD,
-    0x000341B00003FDBF, 0x000341B2000341B1, 0x000341B4000341B3,
-    0x000341B6000341B5, 0x000249B3000341B7, 0x80003F3380003F2F,
-    0x80003F370017C0BE, 0x00024DB500024DB6, 0x000385BB80003F3B,
-    0x80003F3F000385BA, 0x000385BF80003F43, 0x000395BD000385BE,
-    0x80003F47000395BC, 0x80003F4B00020DA3, 0x000209A680003F4F,
-    0x80003F5780003F53, 0x0003C1AA000209A5, 0x80003F5B0003C1AB,
-    0x0003C1AE80003F5F, 0x0003D1AC0003C1AF, 0x00000D800003D1AD,
-    0x0000118200001581, 0x80003F6780003F63, 0x80003F6B0001A4E1,
-    0x80003F73000A04E3, 0x80003F7B80003F77, 0x80003F7F000B34EF,
-    0x80003F8780003F83, 0x001D648680003F8B, 0x80003F8F00028887,
-    0x80003F930002A5B3, 0x00017DB6000175B5, 0x80003F9B80003F97,
-    0x000159BA80003F9F, 0x000149BC000159BB, 0x0001F1BE0001F1BD,
-    0x80003FA30001EDBF, 0x80003FAB80003FA7, 0x80003FB380003FAF,
-    0x80003FBB80003FB7, 0x00166084000BBCEF, 0x80003FBF0006D4E1,
-    0x80003FC3000AB8E3, 0x80003FCB80003FC7, 0x000229A380003FCF,
-    0x0000218100002982, 0x00119C8400001180, 0x0001F9A50001F1A6,
-    0x0001D5AB80003FD3, 0x80003FD70001D5AA, 0x000161AF80003FDB,
-    0x00017DAD00017DAE, 0x001DD1A80001C5AC, 0x001DE1AA001DD9A9,
-    0x001E31AC001DE1AB, 0x001E31AE001E31AD, 0x80003F6F001E31AF,
-    0x0000888680003FDF, 0x80003FE30003AC87, 0x80003FEB80003FE7,
-    0x001EEDB880003FEF, 0x001E81BA001E85B9, 0x001E85BC001E81BB,
-    0x001EB1BE001E8DBD, 0x001E51B0001EB1BF, 0x001E51B2001E51B1,
-    0x001EF9B4001EE9B3, 0x001EE5B6001EE5B5, 0x001FA5A9001ED5B7,
-    0x001FB9AB001FA5A8, 0x001FA9AD001FA5AA, 0x001FD1AF001FA1AC,
-    0x000EE8BE001FA1AE, 0x80003FF780003FF3, 0x80003FFF80003FFB,
-    0x8000400780004003, 0x001F55B98000400B, 0x001F61BB001F55B8,
-    0x001F65BD001F55BA, 0x001F65BF001F65BC, 0x001FBDB1001F65BE,
-    0x001F8DB3001FB5B0, 0x001F75B5001F8DB2, 0x001F75B7001F75B4,
-    0x001E29B3001F75B6, 0x800040138000400F, 0x8000401B80004017,
-    0x001E2DB5001E2DB6, 0x001E7DBB8000401F, 0x80004023001E75BA,
-    0x001EB5BF80004027, 0x001E5DBD001EB5BE, 0x8000402B001E5DBC,
-    0x8000402F001E6DA3, 0x001E69A680004033, 0x8000403B80004037,
-    0x001E31AA001E69A5, 0x8000403F001E39AB, 0x001EF1AE80004043,
-    0x001E19AC001EF1AF, 0x0000ED80001E19AD, 0x0000198200000981,
-    0x8000404B80004047, 0x8000405780004053, 0x000D2CBE00158CEF,
-    0x8000405B001E74E1, 0x00001086001498E3, 0x8000405F00017487,
-    0x0001F9B380004063, 0x8000406B80004067, 0x800040738000406F,
-    0x00017DB500017DB6, 0x000159BB80004077, 0x8000407B000151BA,
-    0x0000FDBF8000407F, 0x0000FDBD0000FDBE, 0x000231A8000141BC,
-    0x000231AA000231A9, 0x0002B5AC000231AB, 0x0002B5AE0002BDAD,
-    0x80003EFF0002ADAF, 0x80004083000FA8BE, 0x000F7CBE80004087,
-    0x8000408F8000408B, 0x000369B880004093, 0x000309BA000369B9,
-    0x00031DBC000305BB, 0x000301BE000301BD, 0x0002D5B0000335BF,
-    0x0002D5B20002DDB1, 0x000379B4000369B3, 0x000369B6000379B5,
-    0x0002B9A3000361B7, 0x8000409B80004097, 0x800040A38000409F,
-    0x00023DA500023DA6, 0x000219AB800040A7, 0x800040AB000211AA,
-    0x0003BDAF800040AF, 0x0003BDAD0003BDAE, 0x00000D80000201AC,
-    0x00001D8200001581, 0x800040B7800040B3, 0x00031887800040BB,
-    0x800040C30004FC86, 0x800040CB800040C7, 0x800040D3800040CF,
-    0x800040DB800040D7, 0x000295A8800040DF, 0x0002A1AA00029DA9,
-    0x0002ADAC0002B5AB, 0x0002D1AE0002D1AD, 0x800040E30002D1AF,
-    0x800040EB800040E7, 0x800040F3800040EF, 0x800040FB800040F7,
-    0x00017DB8800040FF, 0x0001C1BA0001C1B9, 0x0001C1BC0001C1BB,
-    0x0001F1BE0001C9BD, 0x0002B1B00001F1BF, 0x000281B20002B9B1,
-    0x000159B4000281B3, 0x00014DB6000159B5, 0x007048BE000145B7,
-    0x8000410780004103, 0x8000410F8000410B, 0x00167CEF80004113,
-    0x000190E1001670EF, 0x0016D8E3001EFCE1, 0x800041170016E4E3,
-    0x8000411F8000411B, 0x00007D8080004123, 0x0000158200001581,
-    0x80004127000265B3, 0x00021DB6000269B5, 0x000760868000412B,
-    0x000239BA0004A487, 0x000219BC000231BB, 0x0001D1BE000219BD,
-    0x0005A9A30001D1BF, 0x8000412F800040BF, 0x8000413780004133,
-    0x0005A5A50005D1A6, 0x0005FDAB8000413B, 0x8000413F0005F5AA,
-    0x00061DAF80004143, 0x0005D5AD00061DAE, 0x800041470005D5AC,
-    0x8000414B0006ADB3, 0x0006A5B68000414F, 0x8000415780004153,
-    0x000741BA0006ADB5, 0x8000415B000741BB, 0x000741BE8000415F,
-    0x000751BC000741BF, 0x000661A8000751BD, 0x000661AA000661A9,
-    0x000661AC000661AB, 0x000661AE000661AD, 0x80004163000661AF,
-    0x8000416B80004167, 0x800041738000416F, 0x8000417B80004177,
-    0x000745B88000417F, 0x000745BA00074DB9, 0x000745BC00075DBB,
-    0x00076DBE000765BD, 0x0007C1B0000765BF, 0x0007C1B20007C1B1,
-    0x0007C1B40007C1B3, 0x0007C1B60007C1B5, 0x0007E9A30007C1B7,
-    0x8000418780004183, 0x8000418F8000418B, 0x0007E9A50007E1A6,
-    0x000605AB80004193, 0x80004197000605AA, 0x000605AF8000419B,
-    0x000615AD000605AE, 0x00005D80000615AC, 0x0000658200006D81,
-    0x8000419F000151B3, 0x000119B6000179B5, 0x0000C086800041A3,
-    0x00013DBA00017C87, 0x000119BC000135BB, 0x0001D9BE000119BD,
-    0x800041A70001D9BF, 0x800041AF800041AB, 0x000638E1800041B3,
-    0x0010E4E3800041B7, 0x0075E8BE800041BB, 0x001060EF800041BF,
-    0x800041C7800041C3, 0x800041CF800041CB, 0x0001DDA3800041D3,
-    0x800041DB800041D7, 0x800041E3800041DF, 0x0001F5A5000195A6,
-    0x0001B9AB800041E7, 0x800041EB0001B1AA, 0x000155AF800041EF,
-    0x000195AD000155AE, 0x800041F3000195AC, 0x800041FB800041F7,
-    0x00008182800041FF, 0x000099808000404F, 0x00022DA800009981,
-    0x000235AA000235A9, 0x00021DAC000205AB, 0x000271AE000275AD,
-    0x000215B0000271AF, 0x000221B200021DB1, 0x0002F1B4000221B3,
-    0x0002F1B60002F1B5, 0x0002D1B80002F1B7, 0x0002D1BA0002D1B9,
-    0x000375BC0002D1BB, 0x000375BE00037DBD, 0x8000420300036DBF,
-    0x0009848480004207, 0x8000420B000309A3, 0x00030DA6000319A5,
-    0x8000421780004213, 0x000305AA8000421B, 0x000305AC00031DAB,
-    0x000305AE00030DAD, 0x000C6086000379AF, 0x8000421F00036C87,
-    0x8000422780004223, 0x8000422B0001ACE1, 0x8000422F00137CE3,
-    0x8000423780004233, 0x8000423B0013D4EF, 0x800042438000423F,
-    0x00025DA880004247, 0x00028DAA000285A9, 0x00029DAC000285AB,
-    0x000285AE000281AD, 0x8000424B0002B9AF, 0x800042538000424F,
-    0x8000425B80004257, 0x800042638000425F, 0x000291B880004267,
-    0x0002A1BA000299B9, 0x000151BC0002A1BB, 0x000151BE000151BD,
-    0x0002CDB0000151BF, 0x0002DDB20002D5B1, 0x0002CDB40002D5B3,
-    0x0002B1B60002B1B5, 0x000744E10002B1B7, 0x0012ECE30012ECE3,
-    0x8000426B000674E1, 0x800042738000426F, 0x0000058280004277,
-    0x000031808000427B, 0x8000428300003181, 0x001200EF0012F4EF,
-    0x000C608680004287, 0x8000428B000CE487, 0x8000428F000249B3,
-    0x00024DB6000259B5, 0x8000429780004293, 0x000245BA8000429B,
-    0x000245BC00025DBB, 0x000245BE00024DBD, 0x000E9DA8000239BF,
-    0x000EADAA000EA5A9, 0x000EDDAC000EC5AB, 0x000EC1AE000EC5AD,
-    0x8000427F000EFDAF, 0x800042A38000429F, 0x800042AB800042A7,
-    0x800042B3800042AF, 0x000F69B8800042B7, 0x000F79BA000F69B9,
-    0x000F69BC000F79BB, 0x000FC9BE000F69BD, 0x000E85B0000FC9BF,
-    0x000E99B2000E8DB1, 0x000E89B4000E99B3, 0x000F59B6000E89B5,
-    0x000E25B3000F59B7, 0x800042BF800042BB, 0x800042C7800042C3,
-    0x000E55B5000E45B6, 0x000E49BB800042CB, 0x800042CF000E41BA,
-    0x000E4DBF800042D3, 0x000E4DBD000E4DBE, 0x800042D7000E4DBC,
-    0x800042DB000E61A3, 0x000E01A6800042DF, 0x800042E7800042E3,
-    0x000E05AA000E11A5, 0x800042EB000E0DAB, 0x000E09AE800042EF,
-    0x000E09AC000E09AF, 0x000EB1A8000E09AD, 0x000EC1AA000EB9A9,
-    0x000ECDAC000EDDAB, 0x000EFDAE000EF5AD, 0x0001F580000139AF,
-    0x0001C1820001FD81, 0x00013086800042F3, 0x800042F700017487,
-    0x0001CDB8800042FB, 0x0001D5BA0001D5B9, 0x0001FDBC0001E5BB,
-    0x000191BE000195BD, 0x000149B0000191BF, 0x000159B2000149B1,
-    0x000149B4000159B3, 0x0001FDB6000149B5, 0x000D58E10001F5B7,
-    0x001C74E3800042FF, 0x8000430780004303, 0x8000430F8000430B,
-    0x8000431780004313, 0x8000431F8000431B, 0x8000432780004323,
-    0x001C90EF8000432B, 0x000E09B38000432F, 0x8000433780004333,
-    0x8000433F8000433B, 0x000E05B5000E35B6, 0x0001CDBB80004343,
-    0x800043470001CDBA, 0x0001CDBF8000434B, 0x0001CDBD0001CDBE,
-    0x8000434F0001CDBC, 0x800043538000420F, 0x80004357000DC5A3,
-    0x000DF9A6000DC9A5, 0x800043638000435B, 0x000201AA80004367,
-    0x000201AC000201AB, 0x000201AE000201AD, 0x00004980000201AF,
-    0x0000558200005581, 0x0004F0860006CC84, 0x8000436B00031487,
-    0x800043738000436F, 0x8000437B80004377, 0x800043838000437F,
-    0x8000438B80004387, 0x800043938000438F, 0x8000439B80004397,
-    0x800043A38000439F, 0x800043AB800043A7, 0x00038DA8800043AF,
-    0x000395AA00039DA9, 0x00033DAC000329AB, 0x00032DAE000325AD,
-    0x00035DB0000325AF, 0x00036DB2000365B1, 0x00037DB4000365B3,
-    0x00036DB6000365B5, 0x000325B8000361B7, 0x000325BA00032DB9,
-    0x00032DBC00033DBB, 0x000315BE00031DBD, 0x800043B30000FDBF,
-    0x800043BB800043B7, 0x800043BF000231B3, 0x000289B6000299B5,
-    0x0006A484800043C3, 0x0002B5BA800043C7, 0x000295BC00028DBB,
-    0x000295BE00029DBD, 0x00005D8000028DBF, 0x0000058200006181,
-    0x001724EF800043CB, 0x00057487800043D3, 0x0016D0E30004FC86,
-    0x000040E1800043D7, 0x800043DF800043DB, 0x800043E7800043E3,
-    0x800043EF800043EB, 0x800043F7800043F3, 0x800043FF800043FB,
-    0x8000440780004403, 0x007C00BE0012A4EF, 0x8000440B000E7CE1,
-    0x8000440F0013F4E3, 0x8000441780004413, 0x0002FDA38000441B,
-    0x800044238000441F, 0x8000442B80004427, 0x000255A5000245A6,
-    0x000241AB8000442F, 0x80004433000279AA, 0x000241AF80004437,
-    0x000251AD000259AE, 0x000621A8000259AC, 0x000651AA000621A9,
-    0x000671AC000651AB, 0x000671AE000671AD, 0x800043CF000671AF,
-    0x0007C58100001582, 0x8000443B0007C580, 0x00001C868000443F,
-    0x0006D1B80003A087, 0x0006D1BA0006D1B9, 0x000751BC0006D1BB,
-    0x000751BE000751BD, 0x000611B000074DBF, 0x000611B2000611B1,
-    0x0006F1B4000611B3, 0x0006F1B60006F1B5, 0x0006D9B30006F1B7,
-    0x8000444780004443, 0x8000444B007F4084, 0x0006F9B50006F1B6,
-    0x0006D5BB8000444F, 0x800044530006D5BA, 0x000635BF80004457,
-    0x0006C5BD000635BE, 0x8000445B0006C5BC, 0x8000445F00069DA3,
-    0x0006B5A680004463, 0x8000446B80004467, 0x000691AA0006BDA5,
-    0x8000446F000691AB, 0x000671AE80004473, 0x000681AC000671AF,
-    0x000655A8000681AD, 0x000665AA000661A9, 0x000665AC00067DAB,
-    0x000665AE00066DAD, 0x8000447700069DAF, 0x8000447F8000447B,
-    0x8000448780004483, 0x8000448F8000448B, 0x0006F5B880004493,
-    0x0006F5BA0006FDB9, 0x000699BC000689BB, 0x000689BE000699BD,
-    0x0006E5B0000681BF, 0x0006E5B20006EDB1, 0x0006E5B40006FDB3,
-    0x0006D1B60006EDB5, 0x800044970006CDB7, 0x8000449B001044E3,
-    0x8000449F0005C4E1, 0x00003D8100003D82, 0x800044A300003D80,
-    0x800044AB800044A7, 0x800044B7800044B3, 0x00010487000FECEF,
-    0x0006ADB3007F1C86, 0x800044BF800044BB, 0x800044C7800044C3,
-    0x0006ADB50006ADB6, 0x000669BB800044CB, 0x800044CF000661BA,
-    0x000669BF800044D3, 0x000669BD000661BE, 0x00056DA3000671BC,
-    0x8000435F800044D7, 0x800044DF800044DB, 0x00056DA500056DA6,
-    0x0005A9AB800044E3, 0x800044E70005A1AA, 0x0005A9AF007E6084,
-    0x0005A9AD0005A1AE, 0x800044EB0005B1AC, 0x007F4884800044EF,
-    0x800044F30002DDB3, 0x0002F5B60002FDB5, 0x800044FB800044F7,
-    0x0002A5BA800044FF, 0x0002A9BC0002B9BB, 0x000295BE00029DBD,
-    0x000209A900035DBF, 0x000215AB000201A8, 0x00023DAD00021DAA,
-    0x000251AF00023DAC, 0x0078E48500025DAE, 0x8000450780004503,
-    0x8000450F8000450B, 0x0001B1810001A582, 0x0003F1B90001B980,
-    0x0003F1BB0003E9B8, 0x000351BD0003F9BA, 0x000349BF0003E9BC,
-    0x000235B1000355BE, 0x000201B3000229B0, 0x0003D1B5000235B2,
-    0x0003D1B7000219B4, 0x000391A30003D9B6, 0x8000451780004513,
-    0x8000451B00788084, 0x0003B1A50003B9A6, 0x0003F5AB80004523,
-    0x800045270003E9AA, 0x000211AF8000452B, 0x0003D1AD0003D9AE,
-    0x007D60860003E5AC, 0x00028DB3007D6887, 0x00029DB58000452F,
-    0x8000453780004533, 0x8000453B00028DB6, 0x00016DBB8000453F,
-    0x000161BD000281BA, 0x000151BF000179BC, 0x80004543000161BE,
-    0x000F54EF80004547, 0x8000454F8000454B, 0x8000455780004553,
-    0x00620485000CC4EF, 0x8000455B0006E0E1, 0x000190E1000B34E3,
-    0x0007E4E38000455F, 0x000205A380004563, 0x8000456B80004567,
-    0x800045738000456F, 0x000215A5000205A6, 0x0001E5AB80004577,
-    0x8000457B000209AA, 0x0001D9AF8000457F, 0x0001E9AD0001E9AE,
-    0x007DADA80001F1AC, 0x007DCDAA007DB5A9, 0x007DDDAC007DC5AB,
-    0x007DC5AE007DC5AD, 0x8000451F007DF5AF, 0x00001D8100001D82,
-    0x8000458300001D80, 0x8000458B80004587, 0x007E11B88000458F,
-    0x007E29BA007E19B9, 0x007EF5BC007E29BB, 0x007EF5BE007EFDBD,
-    0x007D8DB0007EEDBF, 0x007E65B2007E61B1, 0x007E6DB4007E7DB3,
-    0x007E39B6007E5DB5, 0x007F9DA8007E39B7, 0x007FA5AA007FADA9,
-    0x007FC5AC007FC1AB, 0x007FF9AE007FCDAD, 0x00004086007FF9AF,
-    0x8000459300033487, 0x8000459B80004597, 0x800045A38000459F,
-    0x007F7DB8800045A7, 0x007F01BA007F01B9, 0x007F01BC007F01BB,
-    0x007F31BE007F09BD, 0x007F61B0007F31BF, 0x007F65B2007F61B1,
-    0x007F65B4007F7DB3, 0x007F51B6007F6DB5, 0x007E2DB3007F49B7,
-    0x800045AF800045AB, 0x800045B7800045B3, 0x007E4DB5007E4DB6,
-    0x007EA5BB800045BB, 0x800045BF007EA5BA, 0x007EA9BF800045C3,
-    0x007EA9BD007EA1BE, 0x800045C7007EB1BC, 0x800045CB007E69A3,
-    0x007E09A6800045CF, 0x800045D7800045D3, 0x007EE1AA007E09A5,
-    0x800045DB007EE1AB, 0x007EE5AE800045DF, 0x007EF5AC007EEDAF,
-    0x800045E3007EEDAD, 0x800045EB800045E7, 0x00000D80800045EF,
-    0x00003D8200003581, 0x800045F7800045F3, 0x00000086800045FB,
-    0x8000460300016487, 0x8000460B80004607, 0x007108EF8000460F,
-    0x8000461780004613, 0x8000461F8000461B, 0x8000462780004623,
-    0x8000462B007884EF, 0x8000462F000254E1, 0x0078C0E3007DB4E3,
-    0x007E5CE180004633, 0x000355B880004637, 0x000361BA00035DB9,
-    0x000391BC000361BB, 0x000391BE000391BD, 0x00032DB0000391BF,
-    0x000335B2000335B1, 0x00031DB4000305B3, 0x000371B6000375B5,
-    0x0002E9A8000371B7, 0x0002F9AA0002E9A9, 0x0002E9AC0002F9AB,
-    0x00035DAE0002E9AD, 0x800045FF000355AF, 0x8000463F8000463B,
-    0x8000464780004643, 0x8000464F8000464B, 0x000231B3800044AF,
-    0x8000465780004653, 0x8000465F8000465B, 0x000221B5000229B6,
-    0x0003B1BB80004663, 0x800046670003A9BA, 0x0003F1BF8000466B,
-    0x000391BD000399BE, 0x00006D820003A1BC, 0x00005580000275A3,
-    0x00026DA600006581, 0x8000467780004673, 0x0003EDAA000265A5,
-    0x000648860003F5AB, 0x0003DDAE0004FC87, 0x0003E5AC0003B5AF,
-    0x0002B9AA0003D5AD, 0x0002B9A80002B9AB, 0x0002DDAE0002B9A9,
-    0x0002DDAC0002C5AF, 0x8000467B0002D5AD, 0x800046838000467F,
-    0x8000468B80004687, 0x800046938000468F, 0x000135BA80004697,
-    0x000125B800013DBB, 0x0001DDBE00013DB9, 0x0001DDBC0001C5BF,
-    0x000281B20001D5BD, 0x0002BDB000028DB3, 0x000115B60002B5B1,
-    0x00028DB400011DB7, 0x0074D8EF000281B5, 0x8000469B007F30E1,
-    0x000000E5007E38E3, 0x800046A38000469F, 0x800046A70072E4EF,
-    0x800046AB007EFCE1, 0x007F8CE3007D30E3, 0x000100E1800046AF,
-    0x800046B300666CEF, 0x0007C8BE800046B7, 0x800046BB000275B3,
-    0x000211B6000211B5, 0x800046C3800046BF, 0x00023DBA800046C7,
-    0x000211BC000235BB, 0x00020DBE00021DBD, 0x006C6D8F000201BF,
-    0x800046CB8000466F, 0x00000D80800046CF, 0x00003D8200003581,
-    0x800046D7800046D3, 0x800046DF800046DB, 0x00019086800046E3,
-    0x800046E70003C087, 0x0078759F800046EB, 0x007A419D0078659E,
-    0x007C719B007CA19C, 0x00710999007E519A, 0x0070659700705598,
-    0x00744D9500727196, 0x0076519300744D94, 0x006859910068E592,
-    0x001C85B1006A4190, 0x001801B3001C85B0, 0x0019A5B5001DADB2,
-    0x800046EF0018B5B4, 0x0000B884800046F3, 0x800046FB800046F7,
-    0x0006D1A3800046FF, 0x000499A1000489A2, 0x0000C5A10006C1A0,
-    0x000C01A380004703, 0x000C05A50000E5A2, 0x0008F9A7000CF5A4,
-    0x001401A9000801A6, 0x0014D9AB000849A8, 0x0010D1AD0014CDAA,
-    0x001191AF001001AC, 0x800047070010CDAE, 0x0004DDAA8000470B,
-    0x0004F1AC0004D5AB, 0x0004EDAE0004FDAD, 0x8000470F0004E1AF,
-    0x8000471780004713, 0x8000471B000495A3, 0x0004F1A60004F1A5,
-    0x000661B38000471F, 0x8000472780004723, 0x8000472F8000472B,
-    0x000671B5000679B6, 0x00065DBB80004733, 0x8000473700065DBA,
-    0x0006D1BF8000473B, 0x0006EDBD0006EDBE, 0x006DED820006F5BC,
-    0x8000473F006BB983, 0x0077B18680004743, 0x0069B9840075B587,
-    0x0075858A00690D85, 0x800047470073B18B, 0x0070E18E8000474B,
-    0x0071BD8C007FB18F, 0x007DE1920071B58D, 0x000088BE007D2993,
-    0x0079E59680004753, 0x007BD1940007A997, 0x0005C19A0079FD95,
-    0x000F48860005D99B, 0x000059810000D887, 0x0007999C00007980,
-    0x0004D5A800005182, 0x000529AA0004DDA9, 0x000539AC000529AB,
-    0x000529AE000539AD, 0x80004757000529AF, 0x000C60BE8000475B,
-    0x800047638000475F, 0x8000476B80004767, 0x000581B8003EB4BF,
-    0x000599BA000581B9, 0x0005B1BC000591BB, 0x000251BE0005B1BD,
-    0x000559B0000251BF, 0x0005E1B2000559B1, 0x0005E9B40005F9B3,
-    0x0005D1B60005D5B5, 0x8000476F0005C9B7, 0x8000477780004773,
-    0x8000477B0005ADA3, 0x0005B5A60005BDA5, 0x800047838000477F,
-    0x000591AA80004787, 0x000539AC000591AB, 0x000521AE000521AD,
-    0x000265B600051DAF, 0x8000478F8000478B, 0x80004793000251B5,
-    0x800047970002E9B3, 0x0003C5BE8000479B, 0x0003EDBC0003CDBF,
-    0x000235BA0003CDBD, 0x8000479F0003F5BB, 0x00032DA8800047A3,
-    0x000335AA00033DA9, 0x00038DAC000395AB, 0x0003B1AE0003B1AD,
-    0x800047A70003B1AF, 0x800047AF800047AB, 0x00000981800047B3,
-    0x800047B700001980, 0x000205B80000B982, 0x000205BA00020DB9,
-    0x00020DBC00021DBB, 0x000235BE00023DBD, 0x0003A5B00002D5BF,
-    0x00026DB2000265B1, 0x000255B400024DB3, 0x000249B600025DB5,
-    0x000770E1000241B7, 0x007A94E3800047BB, 0x800047C3000DC084,
-    0x000C38860000A4EF, 0x000CEC84000CBC87, 0x800047C7000114E3,
-    0x800047CB000184E1, 0x000688EF800047CF, 0x800047D7800047D3,
-    0x800047DF800047DB, 0x800047E30003F1A3, 0x00037DA6000349A5,
-    0x800047EB800047E7, 0x00032DAA800047EF, 0x0002F5AC0002EDAB,
-    0x0002DDAE0002D5AD, 0x000DCDA80002D5AF, 0x000E2DAA000E25A9,
-    0x000E21AC000E21AB, 0x000E21AE000E21AD, 0x800047BF000E9DAF,
-    0x800047F7800047F3, 0x800047FF800047FB, 0x8000480780004803,
-    0x000E91B88000480B, 0x000E95BA000E9DB9, 0x000EADBC000EB5BB,
-    0x000F7DBE000F75BD, 0x000EE5B0000F75BF, 0x000EF9B2000EEDB1,
-    0x000ED1B4000EF1B3, 0x000EBDB6000ED1B5, 0x000EA5B3000EB5B7,
-    0x800048138000480F, 0x8000481B80004817, 0x000EA9B5000EA1B6,
-    0x000E29BB8000481F, 0x80004823000E21BA, 0x000E15BF80004827,
-    0x000E25BD000E15BE, 0x8000482B000E2DBC, 0x8000474F000EE1A3,
-    0x000EE5A6000384BE, 0x800048338000482F, 0x000E65AA000EEDA5,
-    0x00004886000E6DAB, 0x000E51AE00006C87, 0x000E69AC000E51AF,
-    0x000E91A8000E61AD, 0x000EB9AA000EADA9, 0x000EADAC000EB9AB,
-    0x000EDDAE000ED5AD, 0x00018980000ED5AF, 0x0001E98200018981,
-    0x8000483B80004837, 0x800048438000483F, 0x000175B880004847,
-    0x000175BA00017DB9, 0x0001D9BC0001C9BB, 0x0001C9BE0001D9BD,
-    0x000EB1B00001C1BF, 0x000E89B2000EB9B1, 0x000155B4000E89B3,
-    0x000155B600015DB5, 0x000E31B300014DB7, 0x8000484F8000484B,
-    0x8000485780004853, 0x000E25B5000E55B6, 0x000E7DBB8000485B,
-    0x8000485F000E75BA, 0x000199BF80004863, 0x000E5DBD000191BE,
-    0x80004867000E5DBC, 0x8000486B000E75A3, 0x000E11A68000486F,
-    0x8000487780004873, 0x000E31AA000E61A5, 0x8000487B000E39AB,
-    0x0001D5AE8000487F, 0x000E19AC0001DDAF, 0x0002D9A8000E19AD,
-    0x000339AA000329A9, 0x000329AC000339AB, 0x00039DAE000329AD,
-    0x80004883000395AF, 0x8000488B80004887, 0x800048938000488F,
-    0x80004897000278BE, 0x000395B88000489B, 0x0003A1BA00039DB9,
-    0x000071BC0003A1BB, 0x000071BE000071BD, 0x0003EDB0000071BF,
-    0x0003F5B20003F5B1, 0x0003DDB40003C5B3, 0x0003B1B60003B5B5,
-    0x0000B1800003B1B7, 0x0000B5820000BD81, 0x800048A30003BDB3,
-    0x0003A5B60003ADB5, 0x0002E084800048A7, 0x000325BA800048AB,
-    0x00032DBC00033DBB, 0x000315BE00031DBD, 0x0003F1A300037DBF,
-    0x00046886800048AF, 0x800048B30003A887, 0x0003E1A50003E9A6,
-    0x000371AB800048B7, 0x800048BB000369AA, 0x000331AF800048BF,
-    0x000351AD000359AE, 0x800048C3000361AC, 0x800048CB800048C7,
-    0x800048D3800048CF, 0x800048DB800048D7, 0x000000C8800048DF,
-    0x800048E7800048E3, 0x800048EF800048EB, 0x800048F7800048F3,
-    0x800048FF800048FB, 0x00024484000ED8E3, 0x0001A0E1000CBCE1,
-    0x0003E8E380004903, 0x00003D8180004907, 0x0003A4EF00003180,
-    0x8000490B00001D82, 0x000574BE000FA4EF, 0x0004208680004913,
-    0x8000491700050887, 0x8000491F8000491B, 0x8000492780004923,
-    0x0005A0840001A8EF, 0x8000492B000E24E1, 0x8000492F000E1CE3,
-    0x8000493780004933, 0x000165B38000493B, 0x800049438000493F,
-    0x8000494B80004947, 0x000109B5000109B6, 0x0000E5BB8000494F,
-    0x800049530000E5BA, 0x0000E5BF80004957, 0x0000F5BD0000E5BE,
-    0x000691A80000F5BC, 0x0006ADAA000699A9, 0x0006C1AC0006C9AB,
-    0x0006F1AE0006C9AD, 0x8000490F0006F1AF, 0x8000495F8000495B,
-    0x8000496780004963, 0x8000496F8000496B, 0x000765B880004973,
-    0x000765BA00076DB9, 0x000765BC00077DBB, 0x000765BE00076DBD,
-    0x000761B000071DBF, 0x000761B2000761B1, 0x000761B4000761B3,
-    0x000761B6000761B5, 0x000625A3000761B7, 0x0000198100001982,
-    0x800049770000B180, 0x000649A5000649A6, 0x0007A5AB8000497B,
-    0x0003A4BE0007A5AA, 0x0007A5AF8000497F, 0x0007B5AD0007A5AE,
-    0x000060860007B5AC, 0x8000498300031887, 0x80004987000741B3,
-    0x000741B6000741B5, 0x8000498B8000489F, 0x00075DBA8000498F,
-    0x00074DBC000755BB, 0x000731BE000731BD, 0x00067DA8000731BF,
-    0x000685AA00068DA9, 0x000685AC00069DAB, 0x0006B1AE00068DAD,
-    0x800049930006ADAF, 0x8000499B80004997, 0x800049A38000499F,
-    0x800049AB800049A7, 0x0006A9B8800049AF, 0x00016DBA0006A9B9,
-    0x00017DBC000165BB, 0x00016DBE000165BD, 0x0006D5B0000165BF,
-    0x0006D5B20006DDB1, 0x0006B9B40006A9B3, 0x0006A9B60006B9B5,
-    0x00060DA30006A1B7, 0x800049B7800049B3, 0x800049BF800049BB,
-    0x00060DA500060DA6, 0x000619AB800049C3, 0x800049C7000611AA,
-    0x00067DAF800049CB, 0x00067DAD00067DAE, 0x800049CF000601AC,
-    0x800049D7800049D3, 0x800049DF800049DB, 0x00003D8100003D82,
-    0x800049E300003D80, 0x800049EB800049E7, 0x0001B8BE0001A084,
-    0x001F5886800049F3, 0x0002C9A80000BC87, 0x0002E9AA0002D5A9,
-    0x000331AC0002E9AB, 0x000331AE000331AD, 0x800049F7000331AF,
-    0x800049FF800049FB, 0x80004A0780004A03, 0x80004A0F80004A0B,
-    0x0003DDB880004A13, 0x0003E5BA0003EDB9, 0x000395BC00038DBB,
-    0x000395BE00039DBD, 0x000351B000038DBF, 0x000351B2000351B1,
-    0x0003F5B4000351B3, 0x0003F5B60003FDB5, 0x000261B30003EDB7,
-    0x80004A1B80004A17, 0x80004A1F001C80BE, 0x0003B9B50003B1B6,
-    0x000395BB80004A23, 0x80004A27000395BA, 0x000339BF80004A2B,
-    0x000385BD000331BE, 0x80004A2F000385BC, 0x80004A33000225A3,
-    0x0003F5A680004A37, 0x002C5C8580004A3B, 0x0003D1AA0003FDA5,
-    0x80004A3F0003D1AB, 0x000375AE80004A43, 0x0003C1AC00037DAF,
-    0x80004A470003C1AD, 0x0002048480004A4B, 0x000018EF001D0884,
-    0x80004A5380004A4F, 0x0000DCE380004A57, 0x0001F0E180004A5B,
-    0x80004A6780004A63, 0x80004A6F80004A6B, 0x0000598080004A73,
-    0x00006D8200005981, 0x000608E380004A77, 0x000788E180004A7B,
-    0x0006E0EF80004A7F, 0x001CC88680004A83, 0x80004A87001D4087,
-    0x80004A8F80004A8B, 0x80004A9780004A93, 0x80004A9F80004A9B,
-    0x80004AA780004AA3, 0x80004AAF80004AAB, 0x001CACBE0001CCEF,
-    0x80004AB300063CE1, 0x80004AB700061CE3, 0x80004ABF80004ABB,
-    0x0001E5B380004AC3, 0x80004ACB80004AC7, 0x80004AD780004AD3,
-    0x0001F5B50001E5B6, 0x000165BB000000F7, 0x001F48860001D9BA,
-    0x000159BF001C6087, 0x000169BD000169BE, 0x001DF5A8000175BC,
-    0x001E2DAA001E25A9, 0x001E3DAC001E25AB, 0x001E2DAE001E25AD,
-    0x80004A5F001E25AF, 0x001FED81001FED82, 0x80004ACF001F9180,
-    0x80004ADF80004ADB, 0x001E89B880004AE3, 0x001E9DBA001E89B9,
-    0x001EBDBC001EA5BB, 0x001F5DBE001F55BD, 0x001E5DB0001F49BF,
-    0x001EE1B2001EE1B1, 0x001EEDB4001EFDB3, 0x001ED5B6001EDDB5,
-    0x001EA5A3001EBDB7, 0x80004AEB80004AE7, 0x80004AF380004AEF,
-    0x001EB5A5001EA5A6, 0x001E25AB80004AF7, 0x80004AFB001E99AA,
-    0x001E19AF80004AFF, 0x001E29AD001E29AE, 0x80004B03001E35AC,
-    0x001FFDB380004B07, 0x001FEDB580004B0B, 0x80004B1380004B0F,
-    0x800049EF001F91B6, 0x001F99BB80004B17, 0x001F89BD001F89BA,
-    0x001F79BF001F81BC, 0x001E2DA8001F81BE, 0x001E3DAA001E35A9,
-    0x001E2DAC001E35AB, 0x001E4DAE001E59AD, 0x80004B1B001E41AF,
-    0x80004B2380004B1F, 0x80004B2B80004B27, 0x80004B3380004B2F,
-    0x001ECDB880004B37, 0x001EDDBA001ED5B9, 0x001EF9BC001ED5BB,
-    0x001EEDBE001EE5BD, 0x001E05B0001EE5BF, 0x001E1DB2001E15B1,
-    0x001E0DB4001E15B3, 0x001EFDB6001EF5B5, 0x001EB5A3001EF5B7,
-    0x0000B58100001582, 0x80004B3B0000B580, 0x001EA5A5001ED9A6,
-    0x001ED1AB000190BE, 0x0002A084001EC1AA, 0x001E31AF80004B43,
-    0x001EC1AD001EC9AE, 0x80004B47001EC9AC, 0x000768860001F1B3,
-    0x000155B600011C87, 0x80004B4F80004B4B, 0x000171BA0001E1B5,
-    0x80004B53000179BB, 0x000131BE80004B57, 0x000151BC000131BF,
-    0x001DCDA8000151BD, 0x00023DAA001DD5A9, 0x00022DAC000235AB,
-    0x000291AE000291AD, 0x80004B5B000291AF, 0x80004B6380004B5F,
-    0x80004B670000008B, 0x80004B6F80004B6B, 0x000299B880004B73,
-    0x000369BA0002A1B9, 0x000379BC000369BB, 0x000369BE000379BD,
-    0x0002F1B0000369BF, 0x0002C1B20002F9B1, 0x0002B1B40002C1B3,
-    0x0002B5B60002B1B5, 0x0002B5A30002A9B7, 0x80004B7B80004B77,
-    0x80004B8380004B7F, 0x0002A5A5000211A6, 0x00023DAB80004B87,
-    0x80004B8B000235AA, 0x000275AF80004B8F, 0x000215AD000275AE,
-    0x80004B93000215AC, 0x80004B9B80004B97, 0x00000D8080004B9F,
-    0x00003D8200003581, 0x80004BA780004BA3, 0x0003248480004BAB,
-    0x000490860003A0BE, 0x80004BB30004D887, 0x0002D5B380004BB7,
-    0x0004C48480004BBB, 0x80004BBF0004C4BE, 0x0002F1B50002F9B6,
-    0x000111BB80004BC3, 0x80004BC700010DBA, 0x0001B5BF80004BCB,
-    0x000101BD000109BE, 0x80004BCF000109BC, 0x80004BD780004BD3,
-    0x80004BDF80004BDB, 0x80004BE780004BE3, 0x80004BEB001F28EF,
-    0x80004BEF001C84E1, 0x80004BF3001E74E3, 0x80004BFB80004BF7,
-    0x80004C0380004BFF, 0x80004C0B80004C07, 0x80004C0F0000E8EF,
-    0x80004C1780004C13, 0x00076C84001EC4EF, 0x80004C1B001ED4E1,
-    0x80004C23001EB0E3, 0x000770BE001E44E1, 0x000215A3000004E3,
-    0x0004E88680004C27, 0x002AEC8500046C87, 0x000231A5000239A6,
-    0x0001D1AB80004C2B, 0x80004C2F0001CDAA, 0x000175AF80004C33,
-    0x0001C1AD0001C9AE, 0x0005ADA80001C9AC, 0x0005DDAA0005C5A9,
-    0x0005F9AC0005D5AB, 0x000639AE0005F9AD, 0x80004BAF000639AF,
-    0x0007B18100078D82, 0x80004C1F0007B980, 0x80004C3B80004C37,
-    0x0006D1B880004C3F, 0x0006E1BA0006D9B9, 0x000691BC0006E1BB,
-    0x000691BE000691BD, 0x000649B0000691BF, 0x00065DB2000655B1,
-    0x00064DB4000655B3, 0x0006F1B60006F1B5, 0x0007F1B80006F1B7,
-    0x0007F1BA0007F1B9, 0x000711BC0007F1BB, 0x000711BE000711BD,
-    0x0007E5B000070DBF, 0x0007E5B20007EDB1, 0x0007E5B40007FDB3,
-    0x0007D9B60007EDB5, 0x000639A80007D5B7, 0x0007CDAA000639A9,
-    0x0007CDAC0007DDAB, 0x0007F5AE0007FDAD, 0x80004C4300079DAF,
-    0x80004C4B80004C47, 0x80004C5380004C4F, 0x80004C5780004B3F,
-    0x000605B380004C5B, 0x80004C6380004C5F, 0x80004C6B80004C67,
-    0x000605B500063DB6, 0x000669BB80004C6F, 0x80004C73000661BA,
-    0x000669BF80004C77, 0x000669BD000661BE, 0x80004C7B000671BC,
-    0x80004C7F000641A3, 0x000679A680004C83, 0x80004C8B80004C87,
-    0x000625AA000641A5, 0x80004C8F00062DAB, 0x000625AE80004C93,
-    0x000635AC00062DAF, 0x000631A800062DAD, 0x000655AA000641A9,
-    0x00067DAC000665AB, 0x0001A1AE0001A1AD, 0x0000ED800001A1AF,
-    0x0000F1820000F181, 0x000C008680004C97, 0x80004C9B00011487,
-    0x0000C9B880004C9F, 0x0000D9BA0000C9B9, 0x0000F9BC0000D1BB,
-    0x000099BE0000F9BD, 0x000101B0000095BF, 0x000101B2000101B1,
-    0x000101B4000101B3, 0x000105B600010DB5, 0x0005DDB30000F9B7,
-    0x80004CA780004CA3, 0x80004CAF80004CAB, 0x0005F5B5000221B6,
-    0x0002C5BB80004CB3, 0x80004CB70002C5BA, 0x0002C9BF000368BE,
-    0x0002D5BD0002C5BE, 0x80004CBB0002D5BC, 0x80004CBF000599A3,
-    0x000265A680004CC3, 0x80004CCB80004CC7, 0x000281AA0005B1A5,
-    0x80004CCF000281AB, 0x000281AE80004CD3, 0x000291AC00028DAF,
-    0x80004CD7000291AD, 0x80004CDF80004CDB, 0x80004CE780004CE3,
-    0x80004CEF80004CEB, 0x00003D8080004CF3, 0x0000198200000981,
-    0x80004CFB80004CF7, 0x80004D0300037884, 0x000C408680004D07,
-    0x80004D0B00039887, 0x80004D1380004D0F, 0x80004D17000200EF,
-    0x80004D1F80004D1B, 0x80004D23000240E3, 0x80004D27000144E1,
-    0x80004D2F80004D2B, 0x0002B9B380004D33, 0x80004D3B80004D37,
-    0x80004D4380004D3F, 0x00011DB500011DB6, 0x000109BB80004D47,
-    0x80004D4B000139BA, 0x000165BF80004D4F, 0x000171BD000179BE,
-    0x80004D5300017DBC, 0x00075CEF80004D57, 0x80004D5F80004D5B,
-    0x00003D8100003D82, 0x80004D6300003D80, 0x80004D6B80004D67,
-    0x0004A0E1000D2084, 0x000780E3000D3CBE, 0x000C208680004D73,
-    0x80004D77000D4C87, 0x80004D7B0006CCEF, 0x80004D8380004D7F,
-    0x000CEC84000668EF, 0x000CE0BE00074CE1, 0x80004D870006B4E3,
-    0x80004D8B000664E1, 0x80004D8F00060CE3, 0x000271A380004D93,
-    0x0001D5A5000FE084, 0x80004D9B80004D97, 0x80004D9F0001D5A6,
-    0x0001C1AB80004DA3, 0x0001B9AD0001F1AA, 0x0001ADAF0001B5AC,
-    0x80004CFF0001B1AE, 0x80004DA780004D6F, 0x80004DAF80004DAB,
-    0x80004DB780004DB3, 0x000E99A880004DBB, 0x000EE1AA000E99A9,
-    0x000EE1AC000EE1AB, 0x000EE5AE000EEDAD, 0x000EA9B0000EDDAF,
-    0x000EBDB2000EB5B1, 0x000EADB4000EB5B3, 0x000F51B6000F51B5,
-    0x000F71B8000F51B7, 0x000F41BA000F79B9, 0x000F31BC000F41BB,
-    0x000F35BE000F31BD, 0x000E49B3000F2DBF, 0x80004DC380004DBF,
-    0x80004DCB80004DC7, 0x000E49B5000E4DB6, 0x000E51BB80004DCF,
-    0x80004DD3000E4DBA, 0x000FB5BF80004DD7, 0x000E41BD000E49BE,
-    0x00001582000E49BC, 0x00006180000E0DA3, 0x000E09A600006181,
-    0x80004DE380004DDB, 0x000E09AA000E0DA5, 0x00004886000E15AB,
-    0x000E0DAE0000DC87, 0x000E0DAC000FF1AF, 0x000E7DA8000E05AD,
-    0x000E99AA000E8DA9, 0x000E89AC000E99AB, 0x000EA1AE000EA5AD,
-    0x80004DDF000EDDAF, 0x80004DEB80004DE7, 0x80004DF380004DEF,
-    0x80004DFB80004DF7, 0x000155B880004DFF, 0x000155BA00015DB9,
-    0x0001D9BC0001C9BB, 0x0001C9BE0001D9BD, 0x000EA5B00001C1BF,
-    0x000EB1B2000EB1B1, 0x000E95B4000E8DB3, 0x000175B600017DB5,
-    0x000E11B300016DB7, 0x80004E0780004E03, 0x80004E0F80004E0B,
-    0x000E01B5000E25B6, 0x000E75BB80004E13, 0x80004E17000E61BA,
-    0x0001B9BF80004E1B, 0x000E59BD000E51BE, 0x80004E1F000E65BC,
-    0x80004E23000E55A3, 0x000E61A680004E27, 0x80004E2F80004E2B,
-    0x000E25AA000E45A5, 0x80004E33000E31AB, 0x000E15AE80004E37,
-    0x000E21AC0001FDAF, 0x0002D9A8000E1DAD, 0x000321AA000321A9,
-    0x000321AC000321AB, 0x000321AE000321AD, 0x80004E3B000321AF,
-    0x80004E4380004E3F, 0x0000198080004E47, 0x0000058200001981,
-    0x000389B880004E4B, 0x00039DBA000389B9, 0x0003B9BC000395BB,
-    0x000079BE0003B9BD, 0x0003E5B0000079BF, 0x0003E5B20003EDB1,
-    0x0003EDB40003FDB3, 0x0003D5B60003DDB5, 0x0003CCBE0003BDB7,
-    0x0003288780004E53, 0x80004E5700042C86, 0x80004E5F80004E5B,
-    0x80004E630002E4EF, 0x80004E67000054E1, 0x80004E6B000F70E3,
-    0x80004E7380004E6F, 0x000349BB80004E77, 0x00240085000341BA,
-    0x000379BF80004E7B, 0x000349BD000349BE, 0x0003C9B3000351BC,
-    0x80004E8380004E7F, 0x0004BC8480004E87, 0x0003C9B50003C1B6,
-    0x000D05AC80004E8B, 0x000D05AE000D0DAD, 0x000E7DA8000D1DAF,
-    0x000EA1AA000EB9A9, 0x80004E8F000EA1AB, 0x80004E9780004E93,
-    0x80004E9F80004E9B, 0x80004EA780004EA3, 0x000DCDBC80004EAB,
-    0x000DDDBE000DD5BD, 0x000D3DB8000DD5BF, 0x000DDDBA000DD5B9,
-    0x000D2DB4000DD5BB, 0x000D31B6000D31B5, 0x000D6DB0000D05B7,
-    0x000D39B2000D5DB1, 0x000E18E1000D35B3, 0x000E38E3000F5CE1,
-    0x80004EAF000EF0E3, 0x80004EB780004EB3, 0x0000358080004EBB,
-    0x0000358200003D81, 0x80004EC30005E4BE, 0x000EE8EF80004EC7,
-    0x0002E9A3000E30EF, 0x0004688680004ECB, 0x80004ECF0004EC87,
-    0x0002E9A50002E1A6, 0x000269AB80004ED3, 0x80004ED7000261AA,
-    0x000259AF80004EDB, 0x000269AD000269AE, 0x000685A8000271AC,
-    0x00069DAA000695A9, 0x0006B1AC000695AB, 0x0006A9AE0006BDAD,
-    0x80004EBF0006A9AF, 0x80004EE380004EDF, 0x80004EEB80004EE7,
-    0x80004EF380004EEF, 0x000765B880004EF7, 0x000765BA00076DB9,
-    0x000765BC00077DBB, 0x000765BE00076DBD, 0x0006D9B00007CDBF,
-    0x0006A5B20006D9B1, 0x000699B4000689B3, 0x000685B6000699B5,
-    0x000615B300075DB7, 0x80004E4F80004EFB, 0x80004F0380004EFF,
-    0x000639B500063DB6, 0x00061DBB80004F07, 0x80004F0B00061DBA,
-    0x00065DBF80004F0F, 0x00065DBD00065DBE, 0x80004F13000671BC,
-    0x80004F17000651A3, 0x000679A680004F1B, 0x80004F2380004F1F,
-    0x000659AA00067DA5, 0x80004F27000659AB, 0x000619AE80004F2B,
-    0x000635AC000619AF, 0x0006B1A8000619AD, 0x0006CDAA0006B1A9,
-    0x0006CDAC0006DDAB, 0x0006F5AE0006FDAD, 0x0001F980000139AF,
-    0x0001C5820001CD81, 0x000278BE00026084, 0x0001AC8780004F2F,
-    0x0001D1B800019C86, 0x0001E1BA0001D9B9, 0x000191BC0001E1BB,
-    0x000195BE00019DBD, 0x000149B0000189BF, 0x000159B2000149B1,
-    0x000149B4000159B3, 0x0001FDB6000149B5, 0x000645B30001F1B7,
-    0x80004F3780004F33, 0x80004F3F80004F3B, 0x000645B500067DB6,
-    0x00018DBB80004F43, 0x80004F4700018DBA, 0x00018DBF80004F4B,
-    0x00018DBD00018DBE, 0x80004F4F00018DBC, 0x80004F53000601A3,
-    0x000639A680004F57, 0x80004F5F80004F5B, 0x0001C9AA000601A5,
-    0x80004F630001C9AB, 0x0001C9AE80004F67, 0x0001C9AC0001C9AF,
-    0x000385A80001C9AD, 0x000385AA00038DA9, 0x00038DAC00039DAB,
-    0x0003B5AE0003BDAD, 0x80004F6B0003DDAF, 0x80004F7380004F6F,
-    0x80004F7B80004F77, 0x80004F8380004F7F, 0x000069B880004F87,
-    0x000079BA000069B9, 0x000069BC000079BB, 0x0000D9BE000069BD,
-    0x0003A9B00000D9BF, 0x0003BDB20003A9B1, 0x000399B40003B5B3,
-    0x000059B6000399B5, 0x80004F8B000059B7, 0x0002D8EF80004F8F,
-    0x0000098180004F93, 0x80004F9700001980, 0x0002AC8400003982,
-    0x00036CBE80004F9B, 0x000194E180004FA3, 0x000208E380004FA7,
-    0x00034DB380004FAB, 0x003CE88680004FAF, 0x80004FB300037C87,
-    0x000345B500037DB6, 0x00032DBB80004FB7, 0x80004FBB00032DBA,
-    0x00031DBF80004FBF, 0x000321BD00031DBE, 0x80004FC3000329BC,
-    0x001BC8EF0007E4E3, 0x80004FC7000798E1, 0x80004FCF80004FCB,
-    0x0003A9A180004FD3, 0x00021DA3000605A0, 0x0013FCE100030DA2,
-    0x000EFCE3000660EF, 0x80004FD7003C7084, 0x0002C5A380004FDB,
-    0x0002CDA5003C04BE, 0x80004FE380004FDF, 0x80004FE70002F5A6,
-    0x0002A5AB80004FEB, 0x0002A9AD0002A5AA, 0x000295AF0002A1AC,
-    0x80004FEF000295AE, 0x80004FF780004FF3, 0x80004FFF80004FFB,
-    0x001B9D8F80005003, 0x003FE8BE000054EF, 0x80005007000674E1,
-    0x00002580000090E3, 0x0000058200001D81, 0x000FF1998000500B,
-    0x000BE99B000FE598, 0x000AB99D000B399A, 0x0007DD9F000BE59C,
-    0x001729910007DD9E, 0x00173D93001BCD90, 0x0013D1950017CD92,
-    0x000F199700132994, 0x000005A100135196, 0x000401A380004F9F,
-    0x00040DA5000055A2, 0x000819A7000411A4, 0x0009E5A9000801A6,
-    0x000C09AB0008B5A8, 0x001001AD000C01AA, 0x0010D5AF000DA9AC,
-    0x001401B100101DAE, 0x001801B3001401B0, 0x001805B50014DDB2,
-    0x8000500F001811B4, 0x0000C08680005013, 0x800050170003E887,
-    0x8000501F8000501B, 0x8000502780005023, 0x003E2DA88000502B,
-    0x003E35AA003E31A9, 0x003E41AC003E49AB, 0x003E71AE003E49AD,
-    0x0000EC84003E71AF, 0x800050338000502F, 0x8000503B80005037,
-    0x800050438000503F, 0x003F99B880005047, 0x003FADBA003FA5B9,
-    0x003F55BC003F4DBB, 0x003F55BE003F5DBD, 0x003FE1B0003F4DBF,
-    0x003FFDB2003FF5B1, 0x003FEDB4003FF5B3, 0x003FBDB6003FB5B5,
-    0x002DAD8B003FA9B7, 0x8000504B002D258A, 0x0035E58F8000504F,
-    0x0031398D0031258E, 0x002509830031FD8C, 0x8000505300252182,
-    0x002DED8780005057, 0x002919850029E186, 0x003D619B0029F584,
-    0x8000505B003DCD9A, 0x800050638000505F, 0x8000506B80005067,
-    0x00355193003C699C, 0x8000506F00354592, 0x0038519780005073,
-    0x0039799500397196, 0x003F3DB300341594, 0x0000198100001982,
-    0x800050770000E180, 0x003F49B5003F49B6, 0x003EB1BB8000507B,
-    0x00000886003EA5BA, 0x003EF1BF00008C87, 0x003EA1BD003EA9BE,
-    0x80005083003EA9BC, 0x80005087003F79A3, 0x003F0DA68000508B,
-    0x800050938000508F, 0x003EE1AA003F0DA5, 0x80005097003EF5AB,
-    0x003EEDAE8000509B, 0x003EEDAC003EB5AF, 0x00025DA8003EE5AD,
-    0x000265AA00026DA9, 0x000265AC00027DAB, 0x000265AE00026DAD,
-    0x0004EC840002B9AF, 0x800050A38000509F, 0x800050AB800050A7,
-    0x800050B3800050AF, 0x000369B8800050B7, 0x000301BA000369B9,
-    0x000301BC000301BB, 0x000301BE000301BD, 0x0002C9B0000335BF,
-    0x0002D9B20002C9B1, 0x0002C9B40002D9B3, 0x000371B60002C9B5,
-    0x0002EDB3000371B7, 0x800050BB00054884, 0x800050C3800050BF,
-    0x0002EDB50002EDB6, 0x00020DBB800050C7, 0x800050CB00020DBA,
-    0x00020DBF800050CF, 0x00020DBD00020DBE, 0x800050D300020DBC,
-    0x800050D70002A9A3, 0x0002A9A6800050DB, 0x800050E3800050DF,
-    0x000249AA0002A9A5, 0x800050E7000249AB, 0x000249AE800050EB,
-    0x000249AC000249AF, 0x0002C1A8000249AD, 0x0002C5AA0002CDA9,
-    0x0002C9AC0002D9AB, 0x0002F5AE0002FDAD, 0x0001F980000139AF,
-    0x0001C5820001CD81, 0x800050F30007E084, 0x800050FB800050F7,
-    0x0001D1B80005F4BE, 0x0001E1BA0001D9B9, 0x000191BC0001E1BB,
-    0x000195BE00019DBD, 0x000149B0000189BF, 0x000159B2000149B1,
-    0x000149B4000159B3, 0x0001F1B6000149B5, 0x003F50E10001F1B7,
-    0x003EA4E3800050FF, 0x0005B086000444BE, 0x800051030005DC87,
-    0x8000510B80005107, 0x800051138000510F, 0x00F78C8580005117,
-    0x003E0CEF8000511B, 0x800051238000511F, 0x8000512B80005127,
-    0x800051338000512F, 0x80005137000164E3, 0x8000513B0001A4E1,
-    0x8000513F003EBCEF, 0x8000514780005143, 0x8000514F8000514B,
-    0x0005CDB380005153, 0x8000507F800050EF, 0x8000515B80005157,
-    0x0005C1B50005FDB6, 0x00060DBB8000515F, 0x0003088600060DBA,
-    0x00060DBF00038C87, 0x00060DBD00060DBE, 0x0000518200060DBC,
-    0x0007B580000589A3, 0x0005B9A600004181, 0x8000516780005163,
-    0x000649AA000585A5, 0x8000516B000649AB, 0x000649AE8000516F,
-    0x000649AC000649AF, 0x0007B1A8000649AD, 0x0007B1AA0007B1A9,
-    0x000709AC000709AB, 0x000739AE000709AD, 0x80005173000739AF,
-    0x8000517B80005177, 0x800051838000517F, 0x8000518B80005187,
-    0x000711B88000518F, 0x000729BA00071DB9, 0x00073DBC000725BB,
-    0x00072DBE000725BD, 0x000749B0000725BF, 0x00075DB2000749B1,
-    0x000731B4000751B3, 0x000731B6000731B5, 0x000639B3000731B7,
-    0x8000519780005193, 0x8000519F8000519B, 0x000681B5000689B6,
-    0x00068DBB800051A3, 0x800051A700068DBA, 0x0006F5BF800051AB,
-    0x000685BD0006F5BE, 0x800051AF000689BC, 0x800051B300067DA3,
-    0x0006CDA6800051B7, 0x800051BF800051BB, 0x0006C9AA0006C5A5,
-    0x800051C30006C9AB, 0x0006B1AE800051C7, 0x0006CDAC0006B1AF,
-    0x0001D5A80006C1AD, 0x0001E1AA0001DDA9, 0x000131AC0001E1AB,
-    0x000131AE000131AD, 0x0000C980000131AF, 0x000005820000C981,
-    0x000178BE800051CB, 0x0000AC87800051D3, 0x0000DDB800019C86,
-    0x0000E5BA0000EDB9, 0x000099BC00008DBB, 0x00008DBE000099BD,
-    0x000151B0000085BF, 0x000151B2000151B1, 0x0000F5B4000151B3,
-    0x0000F5B60000FDB5, 0x000269B30000EDB7, 0x800051DB800051D7,
-    0x800051E3800051DF, 0x0002B9B50002B1B6, 0x000295BB800051E7,
-    0x800051EB000295BA, 0x000339BF800051EF, 0x000285BD000331BE,
-    0x800051F3000285BC, 0x00F20C8500022DA3, 0x0002F5A6800051F7,
-    0x800051FF800051FB, 0x0002D1AA0002FDA5, 0x800052030002D1AB,
-    0x000375AE80005207, 0x0002C1AC00037DAF, 0x000335A80002C1AD,
-    0x000081AA000081A9, 0x000081AC000081AB, 0x0000B1AE000089AD,
-    0x8000520B0000B1AF, 0x800052138000520F, 0x8000521B80005217,
-    0x800052238000521F, 0x000069B880005227, 0x000079BA000069B9,
-    0x000069BC000079BB, 0x0001D9BE000069BD, 0x0000A1B00001D9BF,
-    0x0000BDB20000A9B1, 0x000099B40000B5B3, 0x000059B6000099B5,
-    0x8000522B000059B7, 0x800052338000522F, 0x00001D8180005237,
-    0x8000523B00001D80, 0x0002AC8400000582, 0x000D3887000260BE,
-    0x80005243000DAC86, 0x8000524B80005247, 0x000149B38000524F,
-    0x000D04BE80005253, 0x8000525B80005257, 0x000149B5000149B6,
-    0x00012DBB8000525F, 0x80005263000125BA, 0x00011DBF80005267,
-    0x000121BD00011DBE, 0x000608E1000129BC, 0x00071CE30007B4E3,
-    0x0006ACE300043CE1, 0x0001A8E18000526B, 0x0000E8EF8000526F,
-    0x80005273000F28BE, 0x8000527B80005277, 0x0006CCEF000670EF,
-    0x800052838000527F, 0x0001C1A380005287, 0x0001C1A5000000F3,
-    0x8000528F8000528B, 0x800052930001C1A6, 0x0001A5AB80005297,
-    0x0001A9AD0001ADAA, 0x000195AF0001A1AC, 0x8000523F000195AE,
-    0x8000529F8000529B, 0x00001D80800052A3, 0x00009D8200009581,
-    0x000F11A8800051CF, 0x000F11AA000F11A9, 0x000F75AC000F11AB,
-    0x000F69AE000F79AD, 0x000F29B0000F61AF, 0x000FDDB2000F29B1,
-    0x000FF9B4000FD1B3, 0x000FE9B6000FF9B5, 0x000FA1B8000FE1B7,
-    0x000FBDBA000FA1B9, 0x000F91BC000FB5BB, 0x00007DBE000F91BD,
-    0x800052A7000075BF, 0x0002C884800052AB, 0x800052AF000FB5B3,
-    0x000FD5B6000FD1B5, 0x00030086800052B3, 0x000FC1BA00030087,
-    0x000F25BC000FC1BB, 0x000F11BE000F29BD, 0x000FF9A3000F11BF,
-    0x800052BB800052B7, 0x800052C3800052BF, 0x000F9DA5000F99A6,
-    0x000F8DAB800052C7, 0x800052CB000F8DAA, 0x000F5DAF800052CF,
-    0x000F65AD000F5DAE, 0x00019DA8000F69AC, 0x0001C1AA0001C1A9,
-    0x0001C1AC0001C1AB, 0x0001F1AE0001C9AD, 0x800052D30001F1AF,
-    0x800052DB800052D7, 0x800052E3800052DF, 0x800052EB800052E7,
-    0x0000C1B8800052EF, 0x0000C1BA0000C1B9, 0x0000C1BC0000C1BB,
-    0x0000C1BE0000C1BD, 0x000041B00000C1BF, 0x000041B2000041B1,
-    0x000041B4000041B3, 0x000041B6000041B5, 0x800052F3000041B7,
-    0x800052FB800052F7, 0x80005303800052FF, 0x8000530B80005307,
-    0x000011818000530F, 0x8000531300000D80, 0x8000531700001982,
-    0x0002B4848000531B, 0x000109B380005323, 0x0001E8868000531F,
-    0x8000532700016087, 0x000109B5000109B6, 0x0000C5BB8000532B,
-    0x8000532F0000C5BA, 0x0000C5BF80005333, 0x0000D5BD0000C5BE,
-    0x0002C9A30000D5BC, 0x8000533B80005337, 0x000000808000533F,
-    0x0002C9A50002C9A6, 0x000305AB80005343, 0x80005347000305AA,
-    0x000305AF8000534B, 0x000315AD000305AE, 0x000190E1000315AC,
-    0x000214E38000534F, 0x800053530003E084, 0x8000535B80005357,
-    0x800053638000535F, 0x8000536B80005367, 0x800053738000536F,
-    0x000238EF80005377, 0x00033DA88000537B, 0x000355AA000355A9,
-    0x00037DAC000365AB, 0x000091AE000091AD, 0x8000537F000091AF,
-    0x00001D8100001982, 0x8000538300001D80, 0x0003F4BE80005387,
-    0x0000A1B88000538B, 0x0000A1BA0000A1B9, 0x0000A1BC0000A1BB,
-    0x0000A5BE0000A1BD, 0x0000FDB000007DBF, 0x0000C5B20000CDB1,
-    0x0000A1B40000A9B3, 0x0000A1B60000A1B5, 0x0005A0860000A1B7,
-    0x800053930002C887, 0x8000539700054084, 0x8000539B0001ACE1,
-    0x8000539F0000B4E3, 0x800053A7800053A3, 0x800053AB000E70EF,
-    0x800053B3800053AF, 0x000169BB800053B7, 0x800053BB000159BA,
-    0x000105BF800053BF, 0x00011DBD000105BE, 0x0002BDB3000171BC,
-    0x800053C7800053C3, 0x800053CF800053CB, 0x000295B5000171B6,
-    0x800053D7800053D3, 0x800053DB000EB0E1, 0x000EA0E3000114E3,
-    0x000C3CE1800053DF, 0x000E7CEF800053E3, 0x0004E8840005ECBE,
-    0x800053EB800053E7, 0x800053F3800053EF, 0x00021DA3000E0CEF,
-    0x0000158100002D82, 0x0006108400007580, 0x000235A50001D1A6,
-    0x0001C9AB800053F7, 0x800053FB0001F9AA, 0x0001A5AF80005403,
-    0x0001BDAD0001A5AE, 0x000709A80001D1AC, 0x000711AA000715A9,
-    0x000735AC00072DAB, 0x000729AE00073DAD, 0x800053FF000729AF,
-    0x8000540B80005407, 0x000390868000538F, 0x8000540F00030887,
-    0x00073DB880005413, 0x0007C1BA0007C5B9, 0x0007CDBC0007D9BB,
-    0x0007E9BE0007F9BD, 0x000759B00007E9BF, 0x000721B2000759B1,
-    0x000715B4000721B3, 0x000715B600071DB5, 0x000799B300070DB7,
-    0x8000541B80005417, 0x800054238000541F, 0x0007BDB50007BDB6,
-    0x00079DBB80005427, 0x8000542B00079DBA, 0x0007D9BF8000542F,
-    0x0007F5BD0007D1BE, 0x800054330007F5BC, 0x800054370007DDA3,
-    0x0007F9A68000543B, 0x800054438000543F, 0x0007D9AA0007F9A5,
-    0x800054470007D9AB, 0x000795AE8000544B, 0x0007B1AC00079DAF,
-    0x0006EDA80007B1AD, 0x000121AA000121A9, 0x00012DAC000135AB,
-    0x000191AE000191AD, 0x8000544F000191AF, 0x8000545780005453,
-    0x8000545F8000545B, 0x0000BD810000BD82, 0x000199B80000BD80,
-    0x0001A5BA0001ADB9, 0x000075BC00006DBB, 0x000075BE00007DBD,
-    0x0001F1B000006DBF, 0x0001C1B20001F9B1, 0x0001B1B40001C1B3,
-    0x0001B5B60001BDB5, 0x800054630001A9B7, 0x8000546B80005467,
-    0x0002DCBE0001A5B3, 0x00011DB60001A9B5, 0x000190868000546F,
-    0x000139BA0001C487, 0x000119BC000131BB, 0x000179BE000119BD,
-    0x0001E9A3000175BF, 0x8000547780005473, 0x8000547F8000547B,
-    0x0001E5A5000151A6, 0x00017DAB80005483, 0x80005487000175AA,
-    0x000139AF8000548B, 0x000155AD000135AE, 0x8000548F000155AC,
-    0x8000549780005493, 0x0007D0E38000549B, 0x0001A0E18000549F,
-    0x0002F0EF800054A3, 0x0003F0BE800054A7, 0x800054AF800054AB,
-    0x800054B7800054B3, 0x800054BF800054BB, 0x001CE886800054C3,
-    0x800054C70003C087, 0x800054CF800054CB, 0x00003D80800054D3,
-    0x0000198200000981, 0x800054DB800054D7, 0x800054E7800054E3,
-    0x00035DAC800054EB, 0x000365AE00036DAD, 0x000355A800008DAF,
-    0x000355AA00035DA9, 0x0003AC8400034DAB, 0x800054F3800054EF,
-    0x800054FB800054F7, 0x80005503800054FF, 0x000085BC80005507,
-    0x00009DBE000095BD, 0x0000E5B8000095BF, 0x0000E5BA0000EDB9,
-    0x0000FDB400009DBB, 0x0000C5B60000C1B5, 0x0000F5B00000DDB7,
-    0x0000F5B20000FDB1, 0x8000550B0000E9B3, 0x800055138000550F,
-    0x800055170002C1B3, 0x0002F1B60002C9B5, 0x8000551F8000551B,
-    0x00012DBA80005523, 0x00013DBC000125BB, 0x000125BE000121BD,
-    0x0002A9A6000119BF, 0x8000552B80005527, 0x8000552F000291A5,
-    0x80005533000299A3, 0x00017DAE80005537, 0x000165AC000141AF,
-    0x000175AA000179AD, 0x8000553B00017DAB, 0x800055438000553F,
-    0x00001D8100001D82, 0x0006CCEF00000D80, 0x8000554B80005547,
-    0x0006BCE30000F4EF, 0x0006B0E100018CE1, 0x001DBCBE000614E3,
-    0x001C188680005553, 0x0001BDB3001D6C87, 0x8000555B80005557,
-    0x800055638000555F, 0x000171B5000179B6, 0x00015DBB80005567,
-    0x8000556B00015DBA, 0x0000E1BF8000556F, 0x0000FDBD0000FDBE,
-    0x001E75A80000FDBC, 0x001E75AA001E7DA9, 0x001EB9AC001EA9AB,
-    0x001EA9AE001EB9AD, 0x8000554F001EA9AF, 0x8000557780005573,
-    0x800054DF8000557B, 0x800055838000557F, 0x001F65B880005587,
-    0x001F01BA001F01B9, 0x001F09BC001F1DBB, 0x001F21BE001F09BD,
-    0x001ED9B0001F21BF, 0x001F69B2001ED9B1, 0x001F79B4001F69B3,
-    0x001F65B6001F65B5, 0x001EFDA3001F5DB7, 0x8000558F8000558B,
-    0x8000559780005593, 0x001E31A5001E39A6, 0x001E1DAB8000559B,
-    0x8000559F001E1DAA, 0x001FA1AF800055A3, 0x001FBDAD001FBDAE,
-    0x800055A7001FBDAC, 0x800055AF800055AB, 0x00001D80800055B3,
-    0x0000398200000981, 0x800055BB800055B7, 0x0003A0BE0003A884,
-    0x00005086800055BF, 0x800055C300007487, 0x001EB5A8800055C7,
-    0x001EC1AA001EC1A9, 0x001EC1AC001EC1AB, 0x001EF1AE001EC9AD,
-    0x800055CB001EF1AF, 0x800055D3800055CF, 0x800055DB800055D7,
-    0x800055E3800055DF, 0x0001CDB8800055E7, 0x0001D5BA0001D5B9,
-    0x0001FDBC0001E5BB, 0x00019DBE000195BD, 0x000145B0000195BF,
-    0x000145B200014DB1, 0x000149B4000159B3, 0x0001FDB6000149B5,
-    0x001E49B30001F5B7, 0x800055EF800055EB, 0x800055F7800055F3,
-    0x001E49B5001E49B6, 0x000189BB800055FB, 0x800055FF000181BA,
-    0x00018DBF80005603, 0x00018DBD00018DBE, 0x80005607000191BC,
-    0x8000560B001E0DA3, 0x001E0DA68000560F, 0x8000561780005613,
-    0x0001C5AA001E0DA5, 0x8000561B0001CDAB, 0x0001C9AE8000561F,
-    0x0001D5AC0001C9AF, 0x0002F9A80001C9AD, 0x000329AA0002F9A9,
-    0x000339AC000329AB, 0x000329AE000339AD, 0x80005623000329AF,
-    0x8000562B80005627, 0x80005633000260BE, 0x8000563B80005637,
-    0x0003EDB88000563F, 0x000381BA000385B9, 0x000385BC000381BB,
-    0x0003B1BE00038DBD, 0x000359B00003B1BF, 0x0003EDB2000359B1,
-    0x0003FDB40003E5B3, 0x0003E5B60003E5B5, 0x0000A1800003D5B7,
-    0x0000A1820000A181, 0x800056430003ADB3, 0x0003B5B60003BDB5,
-    0x0002E08480005647, 0x000391BA8000564B, 0x000335BC000391BB,
-    0x000335BE00033DBD, 0x0003E1A300032DBF, 0x000408868000564F,
-    0x000000B400038C87, 0x0003F1A50003F9A6, 0x0003DDAB80005653,
-    0x0004E4BE0003DDAA, 0x000361AF80005657, 0x000371AD000379AE,
-    0x8000565B000379AC, 0x800056638000565F, 0x8000566B80005667,
-    0x800056738000566F, 0x8000567B80005677, 0x800056838000567F,
-    0x8000568B80005687, 0x800056938000568F, 0x0000798080005697,
-    0x00001D8200007981, 0x000384EF00024084, 0x000508878000569B,
-    0x0003D0E30004FC86, 0x000180E1800056A3, 0x800056AB800056A7,
-    0x800056B3800056AF, 0x800056BB800056B7, 0x800056C3800056BF,
-    0x800056C7001EC4EF, 0x800056CF800056CB, 0x0005A88400018CEF,
-    0x800056D3001E04E1, 0x800056D7000158E3, 0x800056DB001FE4E1,
-    0x0001F9B3001F0CE3, 0x800056E3800056DF, 0x800056EB800056E7,
-    0x000149B5000149B6, 0x00010DBB800056EF, 0x800056F3000139BA,
-    0x0000FDBF800056F7, 0x0000FDBD0000FDBE, 0x0006DDB00000FDBC,
-    0x000709B20006E9B1, 0x000719B4000709B3, 0x000709B6000719B5,
-    0x000739B8000709B7, 0x000709BA000739B9, 0x000719BC000709BB,
-    0x000709BE000719BD, 0x8000569F000709BF, 0x800056FB8000562F,
-    0x00000D80800056FF, 0x0000D1820000D181, 0x000651A880005703,
-    0x000651AA000651A9, 0x0006E1AC000651AB, 0x0006DDAE0006E9AD,
-    0x0006B9A30006B9AF, 0x8000570B80005707, 0x8000570F000380BE,
-    0x000609A5000609A6, 0x00064DAB80005713, 0x00000886000679AA,
-    0x0007BDAF00036087, 0x0007BDAD0007BDAE, 0x800057170007BDAC,
-    0x8000571B0007B5B3, 0x0007ADB68000571F, 0x8000572780005723,
-    0x000769BA0007A5B5, 0x8000572B000775BB, 0x000759BE8000572F,
-    0x00076DBC000745BF, 0x000665A8000765BD, 0x000679AA000669A9,
-    0x000669AC000679AB, 0x000699AE000669AD, 0x80005733000699AF,
-    0x8000573B80005737, 0x800057438000573F, 0x8000574B80005747,
-    0x000689B88000574F, 0x000699BA000689B9, 0x0006B9BC000691BB,
-    0x000111BE0006B9BD, 0x0006E9B0000111BF, 0x0006FDB20006F5B1,
-    0x0006D9B40006F5B3, 0x0006C1B60006C5B5, 0x0006F1A30006B9B7,
-    0x8000575780005753, 0x8000575F8000575B, 0x0006E1A50006E9A6,
-    0x000631AB80005763, 0x8000576700062DAA, 0x000601AF8000576B,
-    0x000621AD00061DAE, 0x8000576F000629AC, 0x00001D8100001D82,
-    0x8000577300001D80, 0x8000577B80005777, 0x0001B884000188BE,
-    0x0000E88780005783, 0x80005787000F2C86, 0x8000578F8000578B,
-    0x00029DA880005793, 0x0002C1AA0002C5A9, 0x0002C5AC0002C1AB,
-    0x0002F1AE0002CDAD, 0x800057970002F1AF, 0x8000579F8000579B,
-    0x800057A7800057A3, 0x800057AF800057AB, 0x0003C5B8800057B3,
-    0x0003C5BA0003CDB9, 0x0003CDBC0003DDBB, 0x0003F5BE0003FDBD,
-    0x000341B000039DBF, 0x000341B2000341B1, 0x000341B4000341B3,
-    0x000341B6000341B5, 0x000259B6000341B7, 0x800057BB800057B7,
-    0x800057BF000259B5, 0x800057C300025DB3, 0x000391BE800057C7,
-    0x000381BC000391BF, 0x000391BA000381BD, 0x800057CB000391BB,
-    0x000215A5800057CF, 0x800057D7800057D3, 0x800057DB000215A6,
-    0x000211A3800057DF, 0x0003CDAD800057E3, 0x0003DDAF0003CDAC,
-    0x800057E70003DDAE, 0x0003DDAB800057EB, 0x0001DD800003DDAA,
-    0x0001FD820001F581, 0x00021CBE00028C84, 0x800057F30001A4E1,
-    0x800057F70000ECE3, 0x800057FB000CECBE, 0x800057FF000008EF,
-    0x8000580780005803, 0x000D60868000580B, 0x8000580F0002A887,
-    0x800058130002B9B3, 0x000179B60002A9B5, 0x8000581B80005817,
-    0x000145BA8000581F, 0x000145BC00015DBB, 0x000145BE00014DBD,
-    0x000275A30001F9BF, 0x8000582780005823, 0x8000582F8000582B,
-    0x000265A50001B5A6, 0x000191AB80005833, 0x80005837000189AA,
-    0x000135AF8000583B, 0x000181AD000189AE, 0x000710E1000189AC,
-    0x000694E3000680E1, 0x8000583F00067CE3, 0x8000584780005843,
-    0x8000584F8000584B, 0x000FA581000FB582, 0x000C3C84000FAD80,
-    0x0007E8EF00E2FCBF, 0x000E2DB0000188EF, 0x000E3DB2000E35B1,
-    0x000E2DB4000E35B3, 0x000E91B6000E91B5, 0x000EB9B8000E8DB7,
-    0x000E91BA000EB9B9, 0x000EF5BC000E91BB, 0x000EE9BE000EFDBD,
-    0x80005853000EE9BF, 0x8000577F800057EF, 0x8000585B80005857,
-    0x000378868000585F, 0x000D85A80003EC87, 0x000D9DAA000D95A9,
-    0x000D8DAC000D95AB, 0x000E5DAE000E55AD, 0x000F99A8000E55AF,
-    0x000FA1AA000FA5A9, 0x000FA5AC000FBDAB, 0x000FA5AE000FADAD,
-    0x80005863000FD5AF, 0x8000586B80005867, 0x800058738000586F,
-    0x8000587B80005877, 0x000F5DB88000587F, 0x000F65BA000F6DB9,
-    0x000F05BC000F1DBB, 0x000F05BE000F0DBD, 0x000FB1B0000F3DBF,
-    0x000F89B2000FB9B1, 0x000F75B4000F89B3, 0x000F75B6000F7DB5,
-    0x000E29B3000F6DB7, 0x8000588780005883, 0x8000588F8000588B,
-    0x000E2DB5000E2DB6, 0x000E7DBB80005893, 0x80005897000E75BA,
-    0x000EB9BF8000589B, 0x000E5DBD000EB1BE, 0x8000589F000E5DBC,
-    0x800058A3000E6DA3, 0x000E69A6800058A7, 0x800058AF800058AB,
-    0x000E31AA000E69A5, 0x800058B3000E39AB, 0x000EF5AE800058B7,
-    0x000E19AC000EFDAF, 0x000121A8000E19AD, 0x000121AA000121A9,
-    0x000121AC000121AB, 0x000121AE000121AD, 0x800058BB000121AF,
-    0x800058C3800058BF, 0x00001980800058C7, 0x0000058200001981,
-    0x000089B8800058CB, 0x00009DBA000089B9, 0x0000B9BC000095BB,
-    0x000079BE0000B9BD, 0x0000E5B0000079BF, 0x0000E5B20000EDB1,
-    0x0000EDB40000FDB3, 0x0000D5B60000DDB5, 0x0002C8840000BDB7,
-    0x0002C4BE0002ADB3, 0x0002A5B6800058D3, 0x800058DB800058D7,
-    0x000325BA0002ADB5, 0x0004088600032DBB, 0x00032DBE00030C87,
-    0x000335BC00031DBF, 0x0002E9A300032DBD, 0x800058E3800058DF,
-    0x800058EB800058E7, 0x0002E9A50002E1A6, 0x000369AB800058EF,
-    0x800058F3000361AA, 0x000359AF800058F7, 0x000369AD000369AE,
-    0x001A8582000371AC, 0x800058FB001ADD83, 0x001E8D86800058FF,
-    0x001E918400128987, 0x0012A18A001E9985, 0x800059030012358B,
-    0x0016918E80005907, 0x00137D8C0017298F, 0x000AF1920016918D,
-    0x8000590B000AE593, 0x000E9196000FC4EF, 0x000A3594000E4597,
-    0x0002A99A000E9995, 0x8000590F0002A19B, 0x8000591780005913,
-    0x0002AD9C002228E1, 0x001B058F001B58E3, 0x8000591F8000591B,
-    0x8000592780005923, 0x8000592F8000592B, 0x00003D8080005933,
-    0x0000198200000981, 0x8000593B80005937, 0x0004F0BE0004F884,
-    0x000F019F80005943, 0x000BA59D000F019E, 0x000B619B000B6D9C,
-    0x001779990016C99A, 0x0012A59700177198, 0x001369950013F596,
-    0x001F6D9300137194, 0x001F7191001F7192, 0x0004A086001B2590,
-    0x8000594700052487, 0x8000594F8000594B, 0x8000595780005953,
-    0x8000595F8000595B, 0x8000596780005963, 0x0003E5A38000596B,
-    0x0003F1A10003F1A2, 0x000E98E1000F2DA0, 0x000E28E3000ED8E3,
-    0x000190E1000C38E1, 0x0003C8E38000596F, 0x8000597780005973,
-    0x000318EF8000597B, 0x800059838000597F, 0x000E24EF000E1CEF,
-    0x00EC05B280005987, 0x00F161B000ED0DB3, 0x800058CF00EC0DB1,
-    0x00E801B48000593F, 0x8000598B00E81DB5, 0x800059938000598F,
-    0x8000599B80005997, 0x800059A38000599F, 0x000055A2800059A7,
-    0x800059AB0000F1A3, 0x00FCBDA600005DA1, 0x00FC01A400F801A7,
-    0x00F401AA00FC1DA5, 0x00F829A800F401AB, 0x00F075AE00F8ADA9,
-    0x00F409AC00F069AF, 0x00068DA800F001AD, 0x000691AA000691A9,
-    0x0006D9AC0006C9AB, 0x0006C9AE0006D9AD, 0x000099800006C1AF,
-    0x0000E18200009981, 0x000078BE00006084, 0x00003887800059AF,
-    0x00062DB80000BC86, 0x0007CDBA000635B9, 0x0007DDBC0007C5BB,
-    0x0007CDBE0007C5BD, 0x000689B00007C5BF, 0x000661B2000689B1,
-    0x000659B4000661B3, 0x000631B6000659B5, 0x000719B3000615B7,
-    0x800059B7800059B3, 0x800059BF800059BB, 0x000751B5000759B6,
-    0x00074DBB800059C3, 0x800059C700074DBA, 0x00079DBF800059CB,
-    0x0007B5BD00079DBE, 0x800059CF0007B5BC, 0x800059D300075DA3,
-    0x00071DA6800059D7, 0x800059DF800059DB, 0x000709AA000715A5,
-    0x800059E3000709AB, 0x0007D9AE800059E7, 0x0007F1AC0007D9AF,
-    0x0006BDA80007F1AD, 0x000605AA00060DA9, 0x000605AC00061DAB,
-    0x000605AE00060DAD, 0x800059EB00063DAF, 0x800059F3800059EF,
-    0x800059FB800059F7, 0x80005A03800059FF, 0x000631B880005A07,
-    0x0001E9BA000639B9, 0x0001F9BC0001E9BB, 0x0001E9BE0001F9BD,
-    0x000645B00001E9BF, 0x000645B200064DB1, 0x00062DB4000639B3,
-    0x00061DB6000615B5, 0x00058DB3000611B7, 0x80005A0F80005A0B,
-    0x80005A1780005A13, 0x00058DB500058DB6, 0x000589BB80005A1B,
-    0x00036CBE000581BA, 0x0005D1BF80005A23, 0x0005EDBD0005E9BE,
-    0x0003CC84000591BC, 0x0003C8870005C9A3, 0x0005C9A600FC8C86,
-    0x80005A2B80005A27, 0x0005C5AA0005C9A5, 0x80005A2F0005CDAB,
-    0x0005ADAE80005A33, 0x0005D5AC000595AF, 0x00008D800005A9AD,
-    0x0000958200009581, 0x80005A3B80005A37, 0x80005A4380005A3F,
-    0x0003A5A880005A47, 0x0003A5AA0003ADA9, 0x0003ADAC0003BDAB,
-    0x000395AE00039DAD, 0x000389B00003FDAF, 0x00039DB2000389B1,
-    0x0003B9B4000395B3, 0x000379B60003B9B5, 0x000349B8000379B7,
-    0x000359BA000349B9, 0x000349BC000359BB, 0x0000FDBE000349BD,
-    0x80005A4B0000F5BF, 0x0003D8EF80005A4F, 0x80005A5780005A53,
-    0x80005A5F80005A5B, 0x00FD6C8480005A63, 0x80005A6B80005A67,
-    0x000194E180005A6F, 0x000368E380005A73, 0x00024DB380005A77,
-    0x80005A7F80005A7B, 0x80005A8780005A83, 0x000245B500027DB6,
-    0x00022DBB80005A8B, 0x00FD20BE00022DBA, 0x00021DBF80005A93,
-    0x000221BD00021DBE, 0x0000CD80000229BC, 0x0000D5820000D581,
-    0x000608E180005A97, 0x000124E30007A4E1, 0x80005A9B0006B8E3,
-    0x0001E0EF80005A9F, 0x80005AA300066CEF, 0x80005AAB80005AA7,
-    0x00FC608680005AAF, 0x0002C5A300FD0C87, 0x0002CDA500FC8084,
-    0x80005AB780005AB3, 0x80005ABB0002F5A6, 0x0002A5AB80005ABF,
-    0x0002A9AD0002A5AA, 0x000295AF0002A1AC, 0x00FE99B0000295AE,
-    0x00FF09B200FE99B1, 0x00FF1DB400FF09B3, 0x00FF0DB600FF05B5,
-    0x00FF3DB800FF05B7, 0x00FF01BA00FF01B9, 0x00FF05BC00FF01BB,
-    0x00FF7DBE00FF09BD, 0x80005A8F00FF75BF, 0x80005AC780005AC3,
-    0x80005ACF80005ACB, 0x80005A1F80005AD3, 0x00FE4DA880005AD7,
-    0x00FE31AA00FE31A9, 0x00FE1DAC00FE05AB, 0x00FEE9AE00FEF9AD,
-    0x00FED9B300FEE9AF, 0x80005ADF80005ADB, 0x80005AE780005AE3,
-    0x00FEF5B500FE19B6, 0x00FE3DBB80005AEB, 0x80005AEF00FE3DBA,
-    0x00FFA1BF80005AF3, 0x00FFBDBD00FFB9BE, 0x0000198200FFBDBC,
-    0x0000658000FE9DA3, 0x00FE5DA600001981, 0x80005AFB80005AF7,
-    0x00FE79AA00FEB1A5, 0x0003A0BE00FE79AB, 0x00FFFDAE80005AFF,
-    0x00FFF9AC00FFE5AF, 0x00FE6DA800FFF9AD, 0x00FEA9AA00FE89A9,
-    0x00FEA1AC00FEA9AB, 0x00FEA1AE00FEA1AD, 0x0001808600FEA1AF,
-    0x80005B0300011487, 0x80005B0B80005B07, 0x80005B1380005B0F,
-    0x000115B880005B17, 0x0001CDBA000119B9, 0x0001DDBC0001C5BB,
-    0x0001C5BE0001C5BD, 0x00FEB5B00001F5BF, 0x00FE89B200FEBDB1,
-    0x000135B400FE89B3, 0x000135B600013DB5, 0x00FE21B300012DB7,
-    0x80005B1F80005B1B, 0x80005B2780005B23, 0x00FE29B500FE21B6,
-    0x00FE79BB80005B2B, 0x80005B2F00FE71BA, 0x0001F9BF80005B33,
-    0x00FE5DBD0001F1BE, 0x80005B3700FE5DBC, 0x80005B3B00FE65A3,
-    0x00FE65A680005B3F, 0x80005B4780005B43, 0x00FE35AA00FE6DA5,
-    0x80005B4B00FE3DAB, 0x0001B5AE80005B4F, 0x00FE19AC0001BDAF,
-    0x000329A800FE19AD, 0x000339AA000329A9, 0x000329AC000339AB,
-    0x00039DAE000329AD, 0x80005B53000395AF, 0x80005B5B80005B57,
-    0x00000D8080005B5F, 0x0000BD820000B581, 0x000395B880005B63,
-    0x0003A1BA00039DB9, 0x000071BC0003A1BB, 0x000071BE000071BD,
-    0x0003EDB0000071BF, 0x0003F5B20003F5B1, 0x0003DDB40003C5B3,
-    0x0003B1B60003B5B5, 0x80005B670003B1B7, 0x0002848480005B6B,
-    0x80005B7300030CBE, 0x0004588680005B77, 0x80005B7B0003D087,
-    0x80005B8380005B7F, 0x80005B8B80005B87, 0x80005B9380005B8F,
-    0x0003F1B380005B97, 0x0031708580005B9B, 0x80005BA380005B9F,
-    0x0003E1B50003E9B6, 0x000371BB80005BA7, 0x80005BAB000369BA,
-    0x000331BF80005BAF, 0x000351BD000359BE, 0x80005BB3000361BC,
-    0x80005BBB80005BB7, 0x80005BBF00023DA3, 0x000225A600022DA5,
-    0x80005BC780005BC3, 0x0002A5AA80005BCB, 0x0002ADAC0002BDAB,
-    0x000295AE00029DAD, 0x80005BCF0002FDAF, 0x0000118100001982,
-    0x000324EF00001980, 0x80005BD780005BD3, 0x0003F8E380005BDB,
-    0x000050E100022C84, 0x0004D086000560BE, 0x80005BE30005F887,
-    0x80005BEB80005BE7, 0x80005BEF000125B3, 0x00013DB680005BF3,
-    0x80005BFB80005BF7, 0x000119BA000135B5, 0x80005BFF000119BB,
-    0x000161BE80005C03, 0x00017DBC000169BF, 0x80005C07000171BD,
-    0x80005C0B00FE1CE1, 0x80005C0F00FE10E3, 0x80005C1300FEE4E3,
-    0x80005C1700FF88E1, 0x80005C1B00FE74EF, 0x80005C2380005C1F,
-    0x80005C2B80005C27, 0x0006B9B00001ECEF, 0x0006CDB20006B9B1,
-    0x0006C1B40006C1B3, 0x0006F1B60006C9B5, 0x000721B80006F1B7,
-    0x000721BA000721B9, 0x00072DBC00073DBB, 0x00071DBE000715BD,
-    0x80005BDF00077DBF, 0x80005C3380005C2F, 0x80005C3B80005C37,
-    0x80005C3F80005B6F, 0x0006F1A880005C43, 0x0006C9AA0006F1A9,
-    0x0006D9AC0006C9AB, 0x0006C9AE0006D9AD, 0x000621A30006C9AF,
-    0x0000198100001982, 0x80005C470000D180, 0x000631A5000639A6,
-    0x00061DAB80005C4B, 0x0002A08400061DAA, 0x00066DAF0003A0BE,
-    0x000675AD000665AE, 0x80005C4F000679AC, 0x00006886000745B3,
-    0x000745B600031087, 0x80005C5780005C53, 0x00074DBA000751B5,
-    0x80005C5B000725BB, 0x000729BE80005C5F, 0x000731BC000715BF,
-    0x000679A800072DBD, 0x000689AA000679A9, 0x00069DAC000685AB,
-    0x000681AE000681AD, 0x80005C630006B5AF, 0x80005C6B80005C67,
-    0x80005C7380005C6F, 0x80005C7B80005C77, 0x0006B1B880005C7F,
-    0x0006B1BA0006B1B9, 0x000119BC000109BB, 0x000109BE000119BD,
-    0x0006CDB0000105BF, 0x0006A1B20006A5B1, 0x0006A9B40006BDB3,
-    0x000691B6000695B5, 0x000601A3000691B7, 0x80005C8780005C83,
-    0x80005C8F80005C8B, 0x000615A5000601A6, 0x000661AB80005C93,
-    0x80005C97000609AA, 0x000651AF80005C9B, 0x000669AD00066DAE,
-    0x80005C9F000675AC, 0x80005CA780005CA3, 0x80005CAF80005CAB,
-    0x00003D8100003D82, 0x80005CB300003D80, 0x80005CBB80005CB7,
-    0x000198BE00018084, 0x000F188680005CC3, 0x0002D5A80000EC87,
-    0x0002E1AA0002DDA9, 0x000331AC0002E1AB, 0x000331AE000331AD,
-    0x80005CC7000331AF, 0x80005CCF80005CCB, 0x80005CD780005CD3,
-    0x80005CDF80005CDB, 0x0003DDB880005CE3, 0x0003E5BA0003EDB9,
-    0x000395BC00038DBB, 0x000395BE00039DBD, 0x000351B000038DBF,
-    0x000351B2000351B1, 0x0003F5B4000351B3, 0x0003F5B60003FDB5,
-    0x0003B5BB0003EDB7, 0x80005CE70003B5BA, 0x000319BF000C80BE,
-    0x0003A5BD000311BE, 0x000249B30003A5BC, 0x80005CEF80005CEB,
-    0x80005CF780005CF3, 0x000399B5000391B6, 0x0003F1AA80005CFB,
-    0x80005CFF0003F1AB, 0x000355AE80005D03, 0x0003E1AC00035DAF,
-    0x80005D070003E1AD, 0x80005D0B00020DA3, 0x0003D5A680005D0F,
-    0x80005D1780005D13, 0x80005D1B0003DDA5, 0x0002448480005D1F,
-    0x00000D80000D0484, 0x00003D8200003581, 0x80005D2780005D23,
-    0x80005D3380005D2B, 0x80005D3B80005D37, 0x80005D4380005D3F,
-    0x80005D4B80005D47, 0x80005D4F000194E1, 0x000CF0860000B0E3,
-    0x80005D530002A487, 0x80005D5B80005D57, 0x80005D6380005D5F,
-    0x80005D6B80005D67, 0x80005D7380005D6F, 0x0001A5B3000074EF,
-    0x80005D7B80005D77, 0x80005D8380005D7F, 0x0001B5B50001A5B6,
-    0x000129BB80005D87, 0x80005D8B000199BA, 0x000101BF80005D8F,
-    0x000111BD000119BE, 0x80005D93000119BC, 0x80005D970006ACE1,
-    0x000750E1000174E3, 0x0006ACE380005D9B, 0x000F4086000000F7,
-    0x00067CEF000CD887, 0x00003180000CE0BE, 0x00002D8200003D81,
-    0x000ECDB80001A8EF, 0x000EE1BA000EE5B9, 0x000EE5BC000EE1BB,
-    0x000F19BE000EEDBD, 0x000E1DB0000F19BF, 0x000E8DB2000EADB1,
-    0x000E9DB4000E85B3, 0x000E85B6000E85B5, 0x000DB5A8000EF5B7,
-    0x000E6DAA000E65A9, 0x000E7DAC000E65AB, 0x000E6DAE000E65AD,
-    0x80005CBF000E65AF, 0x80005DA380005D2F, 0x80005D9F80005DA7,
-    0x80005DAF80005DAB, 0x000EA5A380005DB3, 0x80005DBB80005DB7,
-    0x80005DC380005DBF, 0x000EB5A5000EA5A6, 0x000E29AB80005DC7,
-    0x80005DCB000E99AA, 0x000E01AF80005DCF, 0x000E11AD000E19AE,
-    0x80005DD3000E19AC, 0x80005DD7000FF5B3, 0x000FEDB680005DDB,
-    0x80005DE380005DDF, 0x000FB9BA000FE5B5, 0x80005DE7000FB9BB,
-    0x000F99BE80005DEB, 0x000F99BC000F99BF, 0x000E29A8000F99BD,
-    0x000E39AA000E29A9, 0x000E29AC000E39AB, 0x000E59AE000E29AD,
-    0x80005DEF000E55AF, 0x80005DF780005DF3, 0x80005DFF80005DFB,
-    0x0000B1810000B982, 0x000E39B80000B980, 0x000EC1BA000EC1B9,
-    0x000EC1BC000EC1BB, 0x000EF1BE000EC9BD, 0x000E39B0000EF1BF,
-    0x000E09B2000E39B1, 0x000E19B4000E09B3, 0x000E09B6000E19B5,
-    0x000EB1A3000E09B7, 0x80005E0780005E03, 0x0002988480005E0B,
-    0x000EA1A5000EA9A6, 0x000EFDAB000190BE, 0x00000886000EFDAA,
-    0x000EDDAF00017C87, 0x000EDDAD000EDDAE, 0x80005E13000EDDAC,
-    0x80005E170001B1B3, 0x000155B680005E1B, 0x80005E2380005E1F,
-    0x000171BA0001A1B5, 0x80005E27000179BB, 0x000131BE80005E2B,
-    0x000151BC00013DBF, 0x000269A8000151BD, 0x00027DAA000269A9,
-    0x00026DAC000275AB, 0x000291AE000291AD, 0x80005E2F000291AF,
-    0x80005E3780005E33, 0x80005E3F80005E3B, 0x80005E4780005E43,
-    0x000299B880005E4B, 0x0002A5BA0002ADB9, 0x000375BC00036DBB,
-    0x000375BE00037DBD, 0x0002F1B000036DBF, 0x0002C1B20002F9B1,
-    0x0002B1B40002C1B3, 0x0002B5B60002BDB5, 0x0002F5A30002A9B7,
-    0x80005E5380005E4F, 0x80005E5B80005E57, 0x0002E5A5000211A6,
-    0x00023DAB80005E5F, 0x80005E63000235AA, 0x000279AF80005E67,
-    0x000215AD000275AE, 0x80005E6B000215AC, 0x80005E7380005E6F,
-    0x0000198080005E77, 0x0000058200001981, 0x0003AC8480005E7B,
-    0x0004A887000468BE, 0x80005E830004CC86, 0x80005E8B80005E87,
-    0x0002C1A880005E8F, 0x0002C5AA0002CDA9, 0x0002C9AC0002D9AB,
-    0x0002F5AE0002FDAD, 0x80005E9300013DAF, 0x80005E9B80005E97,
-    0x80005EA380005E9F, 0x80005EAB80005EA7, 0x0001C9B80004F4BE,
-    0x0001D9BA0001C9B9, 0x0001F9BC0001D1BB, 0x000199BE0001F9BD,
-    0x000145B0000199BF, 0x000145B200014DB1, 0x000145B400015DB3,
-    0x000145B600014DB5, 0x00024DB30001F9B7, 0x80005EB380005EAF,
-    0x80005EBB80005EB7, 0x000241B500027DB6, 0x000189BB80005EBF,
-    0x80005EC3000181BA, 0x000189BF80005EC7, 0x000189BD000181BE,
-    0x80005ECB000191BC, 0x80005ED380005ECF, 0x80005ED70000A8EF,
-    0x000ECCEF80005EDB, 0x000EA8E1000FA8EF, 0x000EF4E3000C04E1,
-    0x80005EDF000EF4E3, 0x80005EE3000E0CE1, 0x0005CDA300007CE3,
-    0x80005EE780005E7F, 0x0003DC8480005EEB, 0x0005C1A50005FDA6,
-    0x000609AB80005EEF, 0x00030886000601AA, 0x000609AF00036487,
-    0x000609AD000601AE, 0x0007B980000611AC, 0x0000458200004D81,
-    0x80005EF780005EF3, 0x80005EFF80005EFB, 0x80005F0780005F03,
-    0x80005F0F80005F0B, 0x80005F1780005F13, 0x80005F1F80005F1B,
-    0x000671A880005E0F, 0x000671AA000671A9, 0x0007B1AC000671AB,
-    0x0007B1AE0007B1AD, 0x80005F230007B1AF, 0x80005F2B80005F27,
-    0x80005F3380005F2F, 0x80005F3B80005F37, 0x000769B880005F3F,
-    0x000709BA000769B9, 0x000719BC000709BB, 0x00070DBE000705BD,
-    0x0007D1B000072DBF, 0x0007D1B20007D1B1, 0x000779B4000769B3,
-    0x000769B6000779B5, 0x0007F9B3000761B7, 0x80005F4780005F43,
-    0x80005F4F80005F4B, 0x000639B5000631B6, 0x000615BB80005F53,
-    0x80005F57000615BA, 0x0006BDBF80005F5B, 0x0006BDBD0006BDBE,
-    0x00006D82000605BC, 0x000055800007BDA3, 0x000675A600006581,
-    0x000190BE00019884, 0x000651AA00067DA5, 0x80005F5F000651AB,
-    0x0006F9AE80005F63, 0x000641AC0006F9AF, 0x0001CDA80006F9AD,
-    0x0001DDAA0001D5A9, 0x0001CDAC0001D5AB, 0x0001FDAE0001F5AD,
-    0x000000860001F5AF, 0x80005F6700013887, 0x80005F6F80005F6B,
-    0x80005F7780005F73, 0x000171B880005F7B, 0x000171BA000171B9,
-    0x0000D5BC000171BB, 0x0000D5BE0000DDBD, 0x00018DB00000CDBF,
-    0x00019DB2000195B1, 0x000151B4000195B3, 0x000151B6000151B5,
-    0x000255B3000151B7, 0x80005F8380005F7F, 0x80005F8B80005F87,
-    0x000245B5000275B6, 0x000251BB80005F8F, 0x80005F93000249BA,
-    0x000295BF80005F97, 0x000241BD000295BE, 0x00000090000249BC,
-    0x80005F9B000211A3, 0x000231A680005F9F, 0x80005FA780005FA3,
-    0x00020DAA000201A5, 0x80005FAB000215AB, 0x0002D1AE80005FAF,
-    0x00020DAC0002D1AF, 0x80005FB3000205AD, 0x80005FBB80005FB7,
-    0x80005FBF003E9885, 0x00003D8100003D82, 0x80005FC300003D80,
-    0x80005FCB80005FC7, 0x001C7CBE80005FD3, 0x80005FDB80005FD7,
-    0x001C008480005FDF, 0x0003048780005FE3, 0x80005FE7001CEC86,
-    0x80005FEB00029CEF, 0x80005FF380005FEF, 0x80005FF7000234E3,
-    0x80005FFB000154E1, 0x8000600380005FFF, 0x00026DA880006007,
-    0x0002A5AA0002ADA9, 0x0002A5AC0002BDAB, 0x0002A5AE0002ADAD,
-    0x001CEC84000119AF, 0x8000600F8000600B, 0x8000601780006013,
-    0x8000601F8000601B, 0x000115B880006023, 0x000121BA00011DB9,
-    0x0001D1BC000121BB, 0x0001D1BE0001D1BD, 0x000169B00001CDBF,
-    0x000179B2000169B1, 0x000159B4000171B3, 0x000131B6000159B5,
-    0x000774E1000131B7, 0x000188E380006027, 0x0000D4E38000602B,
-    0x0001A8E18000602F, 0x000000EF80006033, 0x8000603B80006037,
-    0x80006043001C64BE, 0x00067CEF80006047, 0x0007DCEF8000604B,
-    0x00006D8100006D82, 0x001CF08600005180, 0x8000604F001D7887,
-    0x8000605780006053, 0x8000605F8000605B, 0x000700E380006063,
-    0x000448E180006067, 0x80005FCF8000606B, 0x8000606F8000603F,
-    0x80006073001DA1B3, 0x001DA1B6001DA1B5, 0x8000607B80006077,
-    0x001E21BA8000607F, 0x001E21BC001E21BB, 0x001E11BE001E29BD,
-    0x001DEDA3001E11BF, 0x8000608780006083, 0x8000608F8000608B,
-    0x001DEDA5001DEDA6, 0x001E6DAB80006093, 0x80006097001E6DAA,
-    0x001E5DAF8000609B, 0x001E65AD001E5DAE, 0x001E2DA8001E6DAC,
-    0x001E31AA001E31A9, 0x001E55AC001E31AB, 0x001E41AE001E45AD,
-    0x8000609F001E41AF, 0x00001D8100001D82, 0x800060A300001D80,
-    0x800060AB800060A7, 0x001FEDB8000074BE, 0x001FFDBA001FF5B9,
-    0x001FEDBC001FF5BB, 0x001F55BE001F51BD, 0x001E01B0001F49BF,
-    0x001E19B2001E0DB1, 0x001E09B4001E19B3, 0x001FDDB6001E09B5,
-    0x001FB9B3001FD5B7, 0x800060AF0001C884, 0x800060B7800060B3,
-    0x001FDDB5001FC5B6, 0x001FC9BB800060BB, 0x00008886001FC5BA,
-    0x001FCDBF0001BC87, 0x001FCDBD001FCDBE, 0x800060BF001FCDBC,
-    0x800060C3001FFDA3, 0x001F81A6800060C7, 0x800060CF800060CB,
-    0x001F81AA001F99A5, 0x800060D3001F8DAB, 0x001F89AE800060D7,
-    0x001F89AC001F89AF, 0x001E31A8001F89AD, 0x001E35AA001E31A9,
-    0x001E4DAC001E55AB, 0x001E7DAE001E75AD, 0x0001AC84001E75AF,
-    0x800060DF800060DB, 0x800060E7800060E3, 0x800060EF800060EB,
-    0x000185B8800060F3, 0x000185BA00018DB9, 0x00018DBC00019DBB,
-    0x0001B5BE0001BDBD, 0x001E0DB000017DBF, 0x001E15B2001E11B1,
-    0x0001E1B40001E5B3, 0x0001D9B60001E9B5, 0x001DBDB30001D9B7,
-    0x800060FB800060F7, 0x80006103800060FF, 0x001DC5B5001DF5B6,
-    0x001DD5BB80006107, 0x8000610B001DC9BA, 0x00022DBF8000610F,
-    0x000231BD000225BE, 0x80006113001DC1BC, 0x80006117001DF9A3,
-    0x001DB1A68000611B, 0x80006123000310BE, 0x001D8DAA001D81A5,
-    0x0004E886001D91AB, 0x000261AE0003A887, 0x001D85AC000269AF,
-    0x80006127000275AD, 0x8000612B000134E3, 0x0000D1800001BCE1,
-    0x0000D5820000DD81, 0x800061338000612F, 0x8000613B80006137,
-    0x8000613F0004E084, 0x80006143000218EF, 0x0003DDB380006147,
-    0x8000614F8000614B, 0x80006153000484BE, 0x000391B5000399B6,
-    0x00038DBB80006157, 0x8000615B00038DBA, 0x000341BF8000615F,
-    0x000351BD000359BE, 0x00020DA8000359BC, 0x000215AA000215A9,
-    0x000259AC00024DAB, 0x00024DAE000259AD, 0x00046C84000245AF,
-    0x8000616780006163, 0x000478BE8000616B, 0x800061738000616F,
-    0x0002FDB880006177, 0x000141BA000141B9, 0x000141BC000141BB,
-    0x000171BE000149BD, 0x000209B0000171BF, 0x0002CDB2000209B1,
-    0x0002DDB40002C5B3, 0x0002CDB60002C5B5, 0x001E48E10002C5B7,
-    0x00009CE3001F50E1, 0x8000617B001EBCE3, 0x800061838000617F,
-    0x8000618B80006187, 0x8000619780006193, 0x000000E98000619B,
-    0x0001FCEF8000619F, 0x00005D81001EE8EF, 0x0002D5A300005D80,
-    0x000299A500006D82, 0x800061A7800061A3, 0x00044086000291A6,
-    0x000285AB0004C487, 0x000259AD000285AA, 0x000249AF000251AC,
-    0x8000618F000251AE, 0x800061AF800061AB, 0x800061B7800061B3,
-    0x8000611F800061BB, 0x000679A8800061BF, 0x0006A1AA0006A5A9,
-    0x000691AC0006B1AB, 0x000691AE000691AD, 0x0006F9B000068DAF,
-    0x00068DB20006F9B1, 0x000681B4000685B3, 0x0006B1B6000689B5,
-    0x000745B80006B1B7, 0x000745BA00074DB9, 0x000749BC000759BB,
-    0x0007E9BE000749BD, 0x000609B30007E9BF, 0x800061C7800061C3,
-    0x800061CF800061CB, 0x000639B5000631B6, 0x00061DBB800061D3,
-    0x800061D700061DBA, 0x00066DBF800061DB, 0x000671BD000665BE,
-    0x800061DF000679BC, 0x800061E300064DA3, 0x000675A6800061E7,
-    0x800061EF800061EB, 0x000659AA00067DA5, 0x800061F3000659AB,
-    0x000621AE800061F7, 0x00063DAC000629AF, 0x0006A5A8000635AD,
-    0x0006A5AA0006ADA9, 0x0006C5AC0006B9AB, 0x00011DAE0006CDAD,
-    0x0001C980000115AF, 0x000005820001C981, 0x000C0086800061FB,
-    0x8000620300019487, 0x0001DDB880006207, 0x0001E5BA0001EDB9,
-    0x000199BC00018DBB, 0x00018DBE000199BD, 0x00016DB0000185BF,
-    0x00017DB2000175B1, 0x0001F5B4000171B3, 0x0001F5B60001FDB5,
-    0x00062DB30001EDB7, 0x8000620F8000620B, 0x800061FF80006213,
-    0x00064DB500064DB6, 0x0001B1BB80006217, 0x8000621B0001A9BA,
-    0x000139BF8000621F, 0x0001A5BD000131BE, 0x800062230001A9BC,
-    0x80006227000669A3, 0x000609A68000622B, 0x800062338000622F,
-    0x0001EDAA000609A5, 0x800062370001F5AB, 0x000175AE8000623B,
-    0x0001EDAC00017DAF, 0x00EA118F0001E1AD, 0x8000623F00F258E1,
-    0x800062430016D0E3, 0x8000624700013CE1, 0x8000624B000304E3,
-    0x800062538000624F, 0x800062570003D8EF, 0x8000625F8000625B,
-    0x0006419F0037F8EF, 0x001A559D001B2D9E, 0x001A2D9B001A499C,
-    0x001E3D99001E259A, 0x0012A197001E3D98, 0x00123D9500122596,
-    0x0016319300168194, 0x00EACD9100162D92, 0x003801B100EA6990,
-    0x003A0DB3003E29B0, 0x0026B1B5003AADB2, 0x80006263002401B4,
-    0x0000398080006267, 0x0000058200003981, 0x0002B1A38000626B,
-    0x0007B1A10002ADA2, 0x000351A10006B9A0, 0x000E4DA380006273,
-    0x000AD5A5000E4DA2, 0x000BF5A7000FA1A4, 0x00366DA9000AD5A6,
-    0x003001AB003665A8, 0x0033E9AD003601AA, 0x003ED9AF0032D5AC,
-    0x00EB1182003C01AE, 0x0002048400EB1183, 0x00175186000D80BE,
-    0x00EBAD840017F987, 0x0013758A00174585, 0x000D28860013658B,
-    0x001F558E000D6C87, 0x0013F98C001F458F, 0x001B3D92001F5D8D,
-    0x000DC884001B4193, 0x00075196000CC0BE, 0x001B459400075197,
-    0x0006699A001A3595, 0x8000627700037D9B, 0x8000627F8000627B,
-    0x0003799C80006283, 0x8000628B80006287, 0x8000628F00067CE1,
-    0x800062930006BCE3, 0x80006297000728E3, 0x000F6C840007ECE1,
-    0x8000629B000790EF, 0x800062A38000629F, 0x800062AB800062A7,
-    0x000245B3000674EF, 0x800062B3800062AF, 0x800062BB800062B7,
-    0x000255B5000249B6, 0x000229BB800062BF, 0x800062C3000225BA,
-    0x000219BF800062C7, 0x00022DBD00022DBE, 0x000D85A3000231BC,
-    0x800062CF800062CB, 0x800062D7800062D3, 0x000D95A5000D89A6,
-    0x000DE9AB800062DB, 0x00030886000DE5AA, 0x000DD9AF00034C87,
-    0x000DEDAD000DEDAE, 0x00004D80000DF1AC, 0x0000558200005581,
-    0x800062E3000F21B3, 0x000F25B6000F21B5, 0x800062EB800062E7,
-    0x000FF9BA800062EF, 0x000FD9BC000FF5BB, 0x000FC9BE000FD9BD,
-    0x8000626F000FC9BF, 0x800062F3800062DF, 0x800062FB800062F7,
-    0x80006303800062FF, 0x000E65A880006307, 0x000FC5AA000FCDA9,
-    0x000FC5AC000FDDAB, 0x000F39AE000FC9AD, 0x000F51B0000F35AF,
-    0x000F61B2000F5DB1, 0x000F61B4000F61B3, 0x000F61B6000F61B5,
-    0x000F29B8000F19B7, 0x000F39BA000F29B9, 0x000F19BC000F31BB,
-    0x0000FDBE000F19BD, 0x000E6DA30000F5BF, 0x8000630F8000630B,
-    0x8000631780006313, 0x000E6DA5000E69A6, 0x000EB9AB8000631B,
-    0x8000631F000EB5AA, 0x000E85AF80006323, 0x000E95AD000E85AE,
-    0x80006327000E95AC, 0x8000632F8000632B, 0x8000633780006333,
-    0x8000633F8000633B, 0x00003D8080006343, 0x0000198200000981,
-    0x8000634B80006347, 0x0001B0BE8000634F, 0x0001D1A880006353,
-    0x0001E9AA0001D9A9, 0x000119AC0001E9AB, 0x000109AE000119AD,
-    0x00040086000109AF, 0x8000635700010487, 0x8000635F8000635B,
-    0x8000636780006363, 0x0000EDB88000636B, 0x00008DBA000085B9,
-    0x00009DBC000085BB, 0x00008DBE000085BD, 0x000179B0000085BF,
-    0x0000EDB2000179B1, 0x0000FDB40000E5B3, 0x0000E5B60000E5B5,
-    0x000261B30000D5B7, 0x800063738000636F, 0x8000637B80006377,
-    0x000291B5000299B6, 0x0002BDBB8000637F, 0x800063830002BDBA,
-    0x000321BF80006387, 0x000331BD000339BE, 0x0004CCBE000339BC,
-    0x8000638B000225A3, 0x0002DDA68000638F, 0x8000639780006393,
-    0x0002F9AA0002D5A5, 0x8000639B0002F9AB, 0x00037DAE8000639F,
-    0x00037DAC000365AF, 0x00050CBE000375AD, 0x800063A7800063A3,
-    0x800063AB000600BE, 0x800063B3800063AF, 0x00003980800063B7,
-    0x0000058200003981, 0x800063C3800063BB, 0x800063C700033884,
-    0x0004E086800063CB, 0x000270EF00036487, 0x800063D3800063CF,
-    0x800063DB800063D7, 0x800063E3800063DF, 0x800063EB800063E7,
-    0x000148E1800063EF, 0x00025CE3800063F3, 0x000130E1800063F7,
-    0x000EB8E3000E34E1, 0x800063FB000E54E3, 0x80006403800063FF,
-    0x00066CBE80006407, 0x8000640F8000640B, 0x8000641780006413,
-    0x0000B8EF8000641B, 0x0002EDB3000E88EF, 0x800064238000641F,
-    0x8000642B80006427, 0x0002EDB50002EDB6, 0x00016DBB80006433,
-    0x0005E08400016DBA, 0x00015DBF80006437, 0x000165BD00015DBE,
-    0x000E0DAC00016DBC, 0x000E21AE000E19AD, 0x000EA9A8000E21AF,
-    0x000E1DAA000EA9A9, 0x0000E980000E15AB, 0x0000F9820000E981,
-    0x8000643F8000643B, 0x0004B48780006443, 0x000DE9BC00049C86,
-    0x000DFDBE000DF5BD, 0x000DC1B8000DF5BF, 0x000DF9BA000DCDB9,
-    0x000E39B4000DF9BB, 0x000E05B6000E0DB5, 0x000E31B0000DE5B7,
-    0x000E35B2000E3DB1, 0x0005ADA3000E29B3, 0x8000642F800063BF,
-    0x8000644B80006447, 0x0005ADA50005ADA6, 0x00062DAB8000644F,
-    0x8000645300062DAA, 0x00061DAF80006457, 0x000625AD00061DAE,
-    0x8000645B00062DAC, 0x8000645F0006F9B3, 0x000685B680006463,
-    0x8000646B80006467, 0x000681BA00068DB5, 0x8000646F000681BB,
-    0x000685BE80006473, 0x000691BC000689BF, 0x000631A8000691BD,
-    0x000635AA00063DA9, 0x000641AC000641AB, 0x000641AE000641AD,
-    0x80006477000641AF, 0x8000647F8000647B, 0x8000648780006483,
-    0x8000648F8000648B, 0x0007D9B880006493, 0x0007EDBA0007D9B9,
-    0x0007FDBC0007E5BB, 0x0007EDBE0007E5BD, 0x000601B0000799BF,
-    0x000601B2000601B1, 0x000605B4000601B3, 0x0007E9B6000609B5,
-    0x0007BDA30007E9B7, 0x0000198100001982, 0x800064970001B180,
-    0x0007C9A50007C1A6, 0x0007C5AB8000649B, 0x0002A0840007C5AA,
-    0x0007CDAF8000649F, 0x0007D5AD0007C1AE, 0x800064A30007D5AC,
-    0x003C00860006E9B3, 0x00015DB600011087, 0x800064AB800064A7,
-    0x000149BA0006E9B5, 0x800064AF000149BB, 0x00013DBE800064B3,
-    0x000139BC000121BF, 0x0006DDA8000135BD, 0x000625AA0006E9A9,
-    0x000625AC00063DAB, 0x000189AE000199AD, 0x0001A084000181AF,
-    0x800064BB800064B7, 0x800064C3800064BF, 0x800064CB800064C7,
-    0x0001ADB8800064CF, 0x00016DBA000165B9, 0x00017DBC000165BB,
-    0x00016DBE000165BD, 0x0001C9B0000165BF, 0x0001A9B20001C9B1,
-    0x0001BDB40001A5B3, 0x0001A1B60001A1B5, 0x0005ADA3000195B7,
-    0x800064D7800064D3, 0x800064DF800064DB, 0x0005ADA5000219A6,
-    0x00020DAB800064E3, 0x800064E700020DAA, 0x000265AF800064EB,
-    0x000271AD000279AE, 0x800064EF00027DAC, 0x800064F7800064F3,
-    0x800064FF800064FB, 0x00003D8100003D82, 0x8000650300003D80,
-    0x8000650B80006507, 0x8000651780006513, 0x003CD8868000651B,
-    0x8000651F00034C87, 0x003CC08480006523, 0x000748EF003CC0BE,
-    0x8000652B80006527, 0x000110E38000652F, 0x0001F8E180006533,
-    0x8000653B80006537, 0x800065438000653F, 0x8000654B80006547,
-    0x800065538000654F, 0x80006557000395B3, 0x00037DB60003B9B5,
-    0x8000655F8000655B, 0x000359BA80006563, 0x000349BC000359BB,
-    0x0000FDBE000349BD, 0x00067DA80000F5BF, 0x0006A5AA0006ADA9,
-    0x0006A5AC0006BDAB, 0x0006A5AE0006ADAD, 0x003C68BE000519AF,
-    0x8000656B80006567, 0x00000D808000656F, 0x0000AD820000A581,
-    0x000515B880006573, 0x000521BA00051DB9, 0x0005D5BC000521BB,
-    0x0005C9BE0005DDBD, 0x000569B00005C9BF, 0x000579B2000569B1,
-    0x000559B4000571B3, 0x000531B6000559B5, 0x000758E1000531B7,
-    0x0006F0E30000BCE3, 0x80006577000628E1, 0x003C38868000657B,
-    0x80006583003D5087, 0x8000658B80006587, 0x003FFCBE8000658F,
-    0x0007C0EF0007C0EF, 0x000299A380006593, 0x8000659B80006597,
-    0x800065A38000659F, 0x0002B5A5000271A6, 0x000255AB800065A7,
-    0x800065AB000255AA, 0x0001F9AF800065AF, 0x000245AD0001F1AE,
-    0x003E6DA8000245AC, 0x003E7DAA003E75A9, 0x003E95AC003E8DAB,
-    0x003E81AE003E81AD, 0x800065B3003E81AF, 0x800065BB800065B7,
-    0x800065C3800065BF, 0x800065CB800065C7, 0x003E89B8800065CF,
-    0x003E95BA003E95B9, 0x003EB5BC003EADBB, 0x003F71BE003F71BD,
-    0x003EC5B0003F71BF, 0x003ED9B2003EC9B1, 0x003EC9B4003ED9B3,
-    0x003EB9B6003EC9B5, 0x8000650F003EB9B7, 0x800065D3003EF1B3,
-    0x003E05B68000657F, 0x800065DB800065D7, 0x003E05BA003E19B5,
-    0x800065DF003E09BB, 0x003E09BE800065E3, 0x003E09BC003E09BF,
-    0x003EB5A3003E09BD, 0x0000198100001982, 0x800065E700006180,
-    0x003E5DA5003E41A6, 0x003E4DAB800065EB, 0x0003A484003E41AA,
-    0x003E4DAF800065EF, 0x003E4DAD003E4DAE, 0x003EA1A8003E4DAC,
-    0x003EA1AA003EA1A9, 0x003EADAC003EA1AB, 0x003ED1AE003ED1AD,
-    0x00018086003ED1AF, 0x800065F30001A087, 0x800065FB800065F7,
-    0x80006603800065FF, 0x000171B880006607, 0x000171BA000171B9,
-    0x0001D9BC000171BB, 0x0001C1BE0001C1BD, 0x003EB1B00001F5BF,
-    0x003E81B2003EB9B1, 0x000151B4003E81B3, 0x000151B6000151B5,
-    0x003E29B3000151B7, 0x8000660F8000660B, 0x8000661780006613,
-    0x003E25B5003E51B6, 0x003E7DBB8000661B, 0x8000661F003E75BA,
-    0x000195BF80006623, 0x003E55BD000195BE, 0x80006627003E55BC,
-    0x8000662B003E6DA3, 0x003E15A68000662F, 0x8000663780006633,
-    0x003E31AA003E61A5, 0x8000663B003E39AB, 0x0001D1AE8000663F,
-    0x003E11AC0001D1AF, 0x000329A8003E11AD, 0x000339AA000329A9,
-    0x000329AC000339AB, 0x00039DAE000329AD, 0x80006643000395AF,
-    0x8000664B80006647, 0x800066538000664F, 0x8000665B80006657,
-    0x000395B880006663, 0x0003A1BA00039DB9, 0x000071BC0003A1BB,
-    0x000071BE000071BD, 0x0003EDB0000071BF, 0x0003F5B20003F5B1,
-    0x0003DDB40003C5B3, 0x0003B1B60003B5B5, 0x0000B5810003B1B7,
-    0x0003B9B30000B580, 0x0003A9B50000B582, 0x8000666B80006667,
-    0x0002E4840003A1B6, 0x000339BB8000666F, 0x000319BD000321BA,
-    0x000379BF000329BC, 0x0003F1A3000311BE, 0x0003B88780006673,
-    0x8000667700046C86, 0x0003E1A50003E9A6, 0x000371AB8000667B,
-    0x8000667F000369AA, 0x000331AF80006683, 0x000351AD000359AE,
-    0x0001F8E3000361AC, 0x003E88E1003F64E1, 0x80006687003F94E3,
-    0x8000668F8000668B, 0x8000669780006693, 0x8000669F8000669B,
-    0x003EE8EF800066A3, 0x001330BF800066A7, 0x800066AB003F48EF,
-    0x00050484800066AF, 0x003E18EF000504BE, 0x800066B7800066B3,
-    0x0003D0E3800066BB, 0x000050E1800066BF, 0x00001D80800066C3,
-    0x0000B98200000981, 0x003E39B2800066C7, 0x003E39B0003E39B3,
-    0x003D9DB6003E39B1, 0x003D9DB4003D85B7, 0x003D8DBA003D95B5,
-    0x003DB5B8003DE5BB, 0x003DFDBE003D85B9, 0x003DFDBC003DE5BF,
-    0x800066CB003DF5BD, 0x0004E886800066D3, 0x800066D70004CC87,
-    0x800066DF800066DB, 0x003EB9AA800066E3, 0x003EA9A8003E8DAB,
-    0x003E59AE003EB9A9, 0x003E95AC003E59AF, 0x00063DA8003E59AD,
-    0x000691AA000691A9, 0x0006D5AC0006A9AB, 0x0006C5AE0006C5AD,
-    0x800066CF0006F5AF, 0x800066EB800066E7, 0x800066F3800066EF,
-    0x800066FB800066F7, 0x00076DB8800066FF, 0x000709BA000771B9,
-    0x000705BC000709BB, 0x000705BE00070DBD, 0x00068DB0000739BF,
-    0x00077DB2000775B1, 0x00076DB4000775B3, 0x00075DB6000755B5,
-    0x000619B3000755B7, 0x8000670780006703, 0x8000670F8000670B,
-    0x00065DB5000645B6, 0x0007B1BB80006713, 0x80006717000649BA,
-    0x000791BF8000671B, 0x0007A1BD0007A9BE, 0x8000671F0007A9BC,
-    0x8000672300065DA3, 0x000601A680006727, 0x8000672F8000672B,
-    0x00060DAA000619A5, 0x800067330007F5AB, 0x0007EDAE80006737,
-    0x0007EDAC0007D5AF, 0x0001E9800007E5AD, 0x000005820001E981,
-    0x0001CCBE8000673B, 0x8000673F0001D084, 0x0000808680006743,
-    0x8000674700029887, 0x8000674F8000674B, 0x8000675780006753,
-    0x8000675F8000675B, 0x8000676780006763, 0x8000676F8000676B,
-    0x800067730004E0EF, 0x8000677B80006777, 0x8000677F00043CE3,
-    0x800067830005B8E1, 0x8000678B80006787, 0x000645A88000678F,
-    0x00065DAA000655A9, 0x00064DAC000655AB, 0x0001B1AE0001B1AD,
-    0x0001AC840001B1AF, 0x8000679780006793, 0x8000679F8000679B,
-    0x0003BD810003B982, 0x00015DB80003BD80, 0x000165BA00016DB9,
-    0x000119BC00010DBB, 0x00010DBE000119BD, 0x0001D1B0000105BF,
-    0x0001D1B20001D1B1, 0x000175B40001D1B3, 0x000175B600017DB5,
-    0x800067A300016DB7, 0x0005DDB3800067A7, 0x0005CDB5800067AB,
-    0x00034CBE800067B3, 0x800067B7000239B6, 0x00021DBB800067BB,
-    0x00020DBD00021DBA, 0x0002B1BF00020DBC, 0x800067BF0002B9BE,
-    0x000595A5000261A6, 0x000585A3000C0084, 0x000CF886800067C3,
-    0x0002E9AF00037C87, 0x000255AD0002E1AE, 0x000245AB000255AC,
-    0x800067C7000245AA, 0x800067CF800067CB, 0x800067D7800067D3,
-    0x800067DF800067DB, 0x800067E3000124E3, 0x000CE8BE0001A4E1,
-    0x800067E7000244EF, 0x800067EF800067EB, 0x800067F7800067F3,
-    0x0003EDB3800067FB, 0x80006803800067FF, 0x8000680B80006807,
-    0x0003EDB50003EDB6, 0x0003A9BB8000680F, 0x8000665F0003A1BA,
-    0x0003ADBF80006813, 0x0003A9BD0003ADBE, 0x00021DA80003B1BC,
-    0x000225AA00022DA9, 0x000225AC00023DAB, 0x000225AE00022DAD,
-    0x0001CD8000025DAF, 0x0000198200000981, 0x8000681B80006817,
-    0x000DB48780006823, 0x000231B8000C1C86, 0x000231BA000231B9,
-    0x0002F5BC000231BB, 0x0002F5BE0002FDBD, 0x000225B00002EDBF,
-    0x000225B200022DB1, 0x000229B400023DB3, 0x000219B6000229B5,
-    0x80006827000215B7, 0x8000682F8000682B, 0x8000683780006833,
-    0x8000683F8000683B, 0x80006843000074EF, 0x80006847000674E1,
-    0x000000EC0000A8E3, 0x8000684F8000684B, 0x0002F5A5000000EF,
-    0x8000685780006853, 0x8000685B0002F5A6, 0x0002F5A38000685F,
-    0x0002B1AD80006863, 0x0002B5AF0002A9AC, 0x800068670002B5AE,
-    0x0002B1AB8000686B, 0x000E15A80002B9AA, 0x000E2DAA000E25A9,
-    0x000E45AC000E4DAB, 0x000E65AE000E6DAD, 0x8000681F000E9DAF,
-    0x800068738000686F, 0x8000687B80006877, 0x800068838000687F,
-    0x000EF5B880006887, 0x000EF5BA000EFDB9, 0x000E99BC000E89BB,
-    0x000E89BE000E99BD, 0x000EE5B0000E81BF, 0x000EE5B2000EEDB1,
-    0x000EE9B4000EFDB3, 0x000ED5B6000EDDB5, 0x000EA5B3000ECDB7,
-    0x0000B58100001582, 0x8000688B0000B580, 0x000ECDB5000EEDB6,
-    0x000E29BB000010BE, 0x00036484000E21BA, 0x000E19BF8000688F,
-    0x000E29BD000E2DBE, 0x80006893000E31BC, 0x0000A886000EE1A3,
-    0x000EA9A600036C87, 0x8000689B80006897, 0x000E65AA000E89A5,
-    0x8000689F000E6DAB, 0x000E69AE800068A3, 0x000E75AC000E5DAF,
-    0x000E85B3000E6DAD, 0x800068AB800068A7, 0x800068B3800068AF,
-    0x000E95B5000E85B6, 0x000E89BB800068B7, 0x800068BB000E85BA,
-    0x000EF9BF800068BF, 0x000E81BD000EF1BE, 0x000D18EF000E89BC,
-    0x800068C7800068C3, 0x800068CF800068CB, 0x800068D7800068D3,
-    0x800068DF800068DB, 0x800068E7800068E3, 0x000D2CE3800068EB,
-    0x000D38E1800068EF, 0x000EB5A8800068F3, 0x000E0DAA000E05A9,
-    0x000E1DAC000E05AB, 0x000E0DAE000E05AD, 0x0000DD80000E05AF,
-    0x0000198200000981, 0x800068FB800068F7, 0x000074BE00013C84,
-    0x000E1DB880006903, 0x000E3DBA000E35B9, 0x0001F9BC0001E9BB,
-    0x0001E9BE0001F9BD, 0x000E41B00001E9BF, 0x000E5DB2000E4DB1,
-    0x000E35B4000E55B3, 0x000E29B6000E3DB5, 0x000D85A3000E29B7,
-    0x0004A88680006907, 0x8000690B00033C87, 0x000D95A5000D85A6,
-    0x000D89AB8000690F, 0x80006913000D85AA, 0x000DF9AF80006917,
-    0x000D81AD000DF1AE, 0x8000691B000D89AC, 0x0003488400029DB3,
-    0x000355B6000440BE, 0x800069238000691F, 0x000371BA0002B1B5,
-    0x80006927000379BB, 0x000335BE8000692B, 0x000351BC00033DBF,
-    0x0003D5B8000351BD, 0x0003D5BA0003DDB9, 0x0003F5BC0003EDBB,
-    0x0003F5BE0003FDBD, 0x0003B1B00003E9BF, 0x000381B20003B9B1,
-    0x0003F5B4000381B3, 0x0003F5B60003FDB5, 0x000201A80003EDB7,
-    0x000211AA000211A9, 0x00023DAC000225AB, 0x0003D1AE0003D1AD,
-    0x8000692F0003D1AF, 0x8000693780006933, 0x8000693F8000693B,
-    0x8000694780006943, 0x8000694B800067AF, 0x800069538000694F,
-    0x000050EF80006957, 0x8000695F8000695B, 0x000084E380006963,
-    0x0001F8E180006967, 0x00002D808000696B, 0x0000158200001181,
-    0x000399A380006973, 0x8000697B80006977, 0x800069838000697F,
-    0x0003B5A5000251A6, 0x00027DAB80006987, 0x00044886000275AA,
-    0x000239AF00054087, 0x000255AD000231AE, 0x8000698B000255AC,
-    0x800069938000698F, 0x8000699B80006997, 0x800069A38000699F,
-    0x0064DC850001F8EF, 0x800069A7000E80E1, 0x800069AB000114E3,
-    0x800069B3800069AF, 0x800069BB800069B7, 0x800069C3800069BF,
-    0x800069C70001DDB3, 0x0001F5B60001FDB5, 0x800069CF800069CB,
-    0x0001B1BA800069D3, 0x00019DBC000185BB, 0x00011DBE000115BD,
-    0x0005BDA8000115BF, 0x000619AA000601A9, 0x000639AC000611AB,
-    0x000659AE000639AD, 0x8000696F000649AF, 0x0000198100001182,
-    0x800069D70007C980, 0x00023084800069DB, 0x0006B5B80003F0BE,
-    0x000749BA0006BDB9, 0x000759BC000749BB, 0x00074DBE000759BD,
-    0x000639B0000745BF, 0x000605B200060DB1, 0x000695B400068DB3,
-    0x000695B600069DB5, 0x000691A300068DB7, 0x00002886800069DF,
-    0x800069E30003C087, 0x0006B1A50006B9A6, 0x0006C9AB800069E7,
-    0x800069EB0006FDAA, 0x000659AF800069EF, 0x000659AD000651AE,
-    0x800069F30006D1AC, 0x800069F700078DB3, 0x00078DB6800069FB,
-    0x80006A03800069FF, 0x000769BA00078DB5, 0x80006A07000769BB,
-    0x000769BE80006A0B, 0x000769BC00075DBF, 0x80006A0F000769BD,
-    0x80006A13000514E1, 0x80006A17000554E3, 0x80006A1F80006A1B,
-    0x80006A2780006A23, 0x80006A2F80006A2B, 0x80006A3780006A33,
-    0x80006A3F80006A3B, 0x000609A3000588EF, 0x0000158100002D82,
-    0x80006A4300001D80, 0x000609A5000609A6, 0x0006EDAB80006A47,
-    0x80006A4B0006EDAA, 0x0006D9AF00012484, 0x0006EDAD0006EDAE,
-    0x0006BDA80006EDAC, 0x0006C1AA0006C1A9, 0x0006C1AC0006C1AB,
-    0x000619AE0006C9AD, 0x0001ECBE000619AF, 0x001F088680006A53,
-    0x80006A570000EC87, 0x80006A5F80006A5B, 0x000679B880006A63,
-    0x000609BA000679B9, 0x000619BC000609BB, 0x000605BE000605BD,
-    0x000669B000063DBF, 0x000679B2000669B1, 0x000669B4000679B3,
-    0x000649B600065DB5, 0x000291A8000649B7, 0x0002A1AA000299A9,
-    0x0002D1AC0002A1AB, 0x0002D5AE0002DDAD, 0x80006A670002C9AF,
-    0x0003A88480006A6B, 0x80006A6F006224BF, 0x80006A7780006A73,
-    0x00035DB880006A7B, 0x00036DBA000365B9, 0x00037DBC000365BB,
-    0x00036DBE000365BD, 0x0002B9B0000365BF, 0x000285B200028DB1,
-    0x000379B400036DB3, 0x00036DB6000365B5, 0x800068FF000365B7,
-    0x80006A7F00022DB3, 0x000259B680006A83, 0x80006A8B80006A87,
-    0x000279BA00025DB5, 0x80006A8F00024DBB, 0x0003B9BE80006A93,
-    0x000255BC0003A5BF, 0x000269A30003BDBD, 0x80006A97000388BE,
-    0x80006A9F80006A9B, 0x000219A500021DA6, 0x000209AB80006AA3,
-    0x80006AA700023DAA, 0x0003E1AF80006AAB, 0x0003F9AD0003FDAE,
-    0x000190E1000211AC, 0x000074E380006AAF, 0x80006AB780006AB3,
-    0x001D348780006ABB, 0x001C2CBE001D3C86, 0x80006AC780006AC3,
-    0x80006ACF80006ACB, 0x000060EF80006AD3, 0x00004D8080006AD7,
-    0x00005D8200005581, 0x80006ADB0002F9B3, 0x000179B60002E9B5,
-    0x80006AE380006ADF, 0x00015DBA80006AE7, 0x00014DBC000155BB,
-    0x0001F5BE0001FDBD, 0x80006AEB0001E9BF, 0x80006AF380006AEF,
-    0x80006AFB80006AF7, 0x80006B0380006AFF, 0x80006B07000158EF,
-    0x80006B0B0006D4E1, 0x001C3C84000184E3, 0x80006B1380006B0F,
-    0x000275A380006B17, 0x80006B1F80006B1B, 0x80006B23001ED084,
-    0x000265A50001F5A6, 0x0001D9AB80006B27, 0x80006B2B0001D1AA,
-    0x000165AF80006B33, 0x000171AD000179AE, 0x001DA5A80001C1AC,
-    0x001DBDAA001DB5A9, 0x001DADAC001DB5AB, 0x001E3DAE001E35AD,
-    0x80006ABF001E2DAF, 0x001FF581001FCD82, 0x80006B2F001FFD80,
-    0x00004C8780006B37, 0x001EEDB800019C86, 0x001E85BA001E8DB9,
-    0x001E8DBC001E9DBB, 0x001EB5BE001EBDBD, 0x001E55B0001F59BF,
-    0x001E55B2001E5DB1, 0x001EFDB4001EE5B3, 0x001EE1B6001EE1B5,
-    0x001FA5A8001ED5B7, 0x001FA5AA001FADA9, 0x001FA5AC001FBDAB,
-    0x001FC9AE001FA9AD, 0x80006B3B001FC1AF, 0x80006B4380006B3F,
-    0x80006B4B80006B47, 0x80006B5380006B4F, 0x001F59B880006B57,
-    0x001F61BA001F65B9, 0x001F69BC001F79BB, 0x001F1DBE001F69BD,
-    0x001F81B0001F15BF, 0x001F69B2001F81B1, 0x001F79B4001F69B3,
-    0x001F69B6001F79B5, 0x80006B5B001F69B7, 0x80006B6380006B5F,
-    0x001EA0E180006B67, 0x001D20E3003D8CE1, 0x80006B6B0018A8E3,
-    0x001D14EF80006B6F, 0x0003A5A2001DFCEF, 0x001D19A0000115A3,
-    0x001E6DB3001DB9A1, 0x80006B7780006B73, 0x80006B7F80006B7B,
-    0x001E6DB5001E69B6, 0x001E05BB80006B83, 0x80006B87001E05BA,
-    0x001EFDBF80006B8B, 0x001EFDBD001EFDBE, 0x001EADA3001EFDBC,
-    0x80006B8F0024158F, 0x80006B9780006B93, 0x001EADA5001EA9A6,
-    0x001EC5AB80006B9B, 0x000400BE001EC5AA, 0x001E3DAF80006BA3,
-    0x001E3DAD001E3DAE, 0x00196D9E001E3DAC, 0x001B359C001F459F,
-    0x0015F19A0019ED9D, 0x001135980014CD9B, 0x0011619600173599,
-    0x002DE19400116997, 0x00288592002CC195, 0x00297990002DF193,
-    0x0027918200297991, 0x00048886002A8583, 0x002E158600038C87,
-    0x002A9D84002E1D87, 0x002F0D8A002B6985, 0x80006BA70012A58B,
-    0x0014D98E00046084, 0x0012B58C0016158F, 0x001759920012D18D,
-    0x80006BAB0018B993, 0x001CD19680006BAF, 0x001A7994001E7597,
-    0x001EE59A001AF195, 0x80006BB30000F19B, 0x0003E18180006BB7,
-    0x0002599C0003F980, 0x000801A50003E182, 0x000A7DA70007B5A4,
-    0x0002A1A1000AADA6, 0x00069DA380006A4F, 0x00729DAD00069DA2,
-    0x0076ADAF007285AC, 0x000E91A9007401AE, 0x007001AB000C01A8,
-    0x007C01B5000E15AA, 0x80006BBB007BE9B4, 0x0077B9B180006BBF,
-    0x007ADDB30076F9B0, 0x80006BC3007ADDB2, 0x80006BCB80006BC7,
-    0x80006BD380006BCF, 0x80006BDB80006BD7, 0x001CEDA880006BDF,
-    0x0003A9AA000331A9, 0x0003B9AC0003A9AB, 0x0003A9AE0003B9AD,
-    0x80006BE30003A9AF, 0x80006BEB80006BE7, 0x80006BF380006BEF,
-    0x80006BFB80006BF7, 0x000205B880006BFF, 0x000205BA00020DB9,
-    0x00020DBC00021DBB, 0x000235BE00023DBD, 0x0003D9B00002D5BF,
-    0x000269B20003D9B1, 0x000279B4000269B3, 0x000259B6000279B5,
-    0x00001180000259B7, 0x0000118200001181, 0x80006C030000DCEF,
-    0x80006C0B80006C07, 0x0005E8BE001E44EF, 0x80006C13001EA0E1,
-    0x80006C17001E0CE3, 0x0004F484000190E1, 0x80006C1B001E60E3,
-    0x000361B5000369B6, 0x000371B380006C1F, 0x0004788680006C23,
-    0x0002EDBF0004FC87, 0x0002EDBD0002EDBE, 0x00034DBB0002EDBC,
-    0x80006C2700034DBA, 0x80006B9F80006C2B, 0x80006C2F80006C0F,
-    0x80006C3780006C33, 0x80006C3F80006C3B, 0x00058DA880006C43,
-    0x000661AA00060DA9, 0x000659AC000661AB, 0x0006F1AE000659AD,
-    0x0006A1B00006F1AF, 0x0006A1B20006A1B1, 0x0006A1B40006A1B3,
-    0x0006A1B60006A1B5, 0x000741B80006A1B7, 0x000741BA000741B9,
-    0x000741BC000741BB, 0x000741BE000741BD, 0x0006C9B3000741BF,
-    0x80006C4B80006C47, 0x80006C5380006C4F, 0x0006F1B50006F9B6,
-    0x000645BB80006C57, 0x80006C5B000645BA, 0x000645BF80006C5F,
-    0x000655BD000645BE, 0x80006C63000655BC, 0x80006C6700068DA3,
-    0x0006BDA680006C6B, 0x80006C7380006C6F, 0x000601AA0006B5A5,
-    0x80006C77000601AB, 0x000601AE80006C7B, 0x000611AC000601AF,
-    0x0006A9A8000611AD, 0x0006C9AA0006B5A9, 0x0006C5AC0006C1AB,
-    0x0006E5AE0006EDAD, 0x0001F980000115AF, 0x0001C5820001CD81,
-    0x80006C7F000264BE, 0x0001AC8780006C83, 0x0001D1B800019C86,
-    0x0001E1BA0001D9B9, 0x000191BC0001E1BB, 0x000195BE00019DBD,
-    0x00016DB0000189BF, 0x00017DB2000175B1, 0x00016DB4000175B3,
-    0x0001F1B60001F1B5, 0x000641B30001F1B7, 0x80006C8B80006C87,
-    0x80006C9380006C8F, 0x00064DB500066DB6, 0x0001B1BB80006C97,
-    0x000184840001A9BA, 0x000135BF80006C9B, 0x0001A1BD0001A9BE,
-    0x80006C9F0001A9BC, 0x80006CA3000605A3, 0x000629A680006CA7,
-    0x80006CAF80006CAB, 0x0001EDAA000609A5, 0x80006CB30001F5AB,
-    0x0001EDAE80006CB7, 0x0001EDAC000171AF, 0x000335A80001E5AD,
-    0x000381AA000381A9, 0x000381AC000381AB, 0x0003B5AE000389AD,
-    0x80006CBB0003DDAF, 0x80006CC380006CBF, 0x80006CCB80006CC7,
-    0x80006CD380006CCF, 0x000069B880006CD7, 0x000079BA000069B9,
-    0x000069BC000079BB, 0x0000DDBE000069BD, 0x0003A9B00000D5BF,
-    0x0003BDB20003A9B1, 0x000399B40003B5B3, 0x000059B6000399B5,
-    0x000344BE000059B7, 0x80006CDF80006CDB, 0x000031810002F4EF,
-    0x80006CE300000D80, 0x80006CE700003982, 0x000C088680006CEB,
-    0x80006CF300030C87, 0x80006CF7000190E1, 0x00030DBB000104E3,
-    0x80006CFB00030DBA, 0x00033DBF80006CFF, 0x000301BD00033DBE,
-    0x000369B3000309BC, 0x80006D0780006D03, 0x80006D0F80006D0B,
-    0x000369B5000369B6, 0x0007B9AA80006D13, 0x80006D170007B9AB,
-    0x000619AE80006D1B, 0x000619AC000619AF, 0x005C28BF000619AD,
-    0x80006D1F000785A3, 0x00079DA680006D23, 0x80006D2B80006D27,
-    0x80006D2F000795A5, 0x0002C1A380006D33, 0x0002C1A500024084,
-    0x80006D3B80006D37, 0x80006D3F0002C1A6, 0x0002A5AB80006D43,
-    0x0002A9AD0002A5AA, 0x000295AF0002A1AC, 0x000640E1000295AE,
-    0x000620E3000794E1, 0x80006D470007B0E3, 0x80006D4F80006D4B,
-    0x00003D8080006D53, 0x0000198200000981, 0x80006D5B80006D57,
-    0x00006CEF000DB8BE, 0x000109B30006E4EF, 0x000CE88680006D63,
-    0x80006D67000D1087, 0x00013DB500013DB6, 0x0000E9BB80006D6B,
-    0x80006D6F000119BA, 0x0000C5BF80006D73, 0x0000D1BD0000D9BE,
-    0x000EB1A80000F1BC, 0x000EC1AA000EBDA9, 0x000EF5AC000EC1AB,
-    0x000EF5AE000EFDAD, 0x80006D5F000EEDAF, 0x80006D7B80006D77,
-    0x80006D8380006D7F, 0x80006D8B80006D87, 0x000F05B880006D8F,
-    0x000F11BA000F11B9, 0x000F35BC000F2DBB, 0x000FCDBE000FC5BD,
-    0x000E9DB0000FC5BF, 0x000F65B2000F6DB1, 0x000F5DB4000F45B3,
-    0x000F41B6000F41B5, 0x000E49A3000F3DB7, 0x80006D9780006D93,
-    0x80006D9F80006D9B, 0x000E7DA5000E7DA6, 0x000FA9AB80006DA3,
-    0x80006DA7000E59AA, 0x000F85AF80006DAB, 0x000F91AD000F99AE,
-    0x80006DAF000FB1AC, 0x80006DB780006DB3, 0x80006DBF80006DBB,
-    0x00003D8100003D82, 0x80006DC300003D80, 0x80006DCB80006DC7,
-    0x000070BE00036084, 0x0000F88680006DCF, 0x80006DD300034C87,
-    0x80006DD7000DB4E1, 0x80006DDB000C50E3, 0x80006DE380006DDF,
-    0x80006DEB80006DE7, 0x80006DF380006DEF, 0x80006DFB80006DF7,
-    0x80006E0380006DFF, 0x000E9DB3000C70EF, 0x80006E0B80006E07,
-    0x80006E1380006E0F, 0x000EB5B5000171B6, 0x000155BB80006E17,
-    0x80006E1B000155BA, 0x0001F9BF80006E1F, 0x000145BD0001F1BE,
-    0x000E45A8000145BC, 0x000E55AA000E55A9, 0x000E7DAC000E65AB,
-    0x0001B1AE0001B1AD, 0x80006E230001B1AF, 0x80006E2B80006E27,
-    0x00000D8080006E2F, 0x0003BD820003B581, 0x00015DB880006E33,
-    0x000165BA00016DB9, 0x000115BC00010DBB, 0x000115BE00011DBD,
-    0x0001D1B000010DBF, 0x0001D1B20001D1B1, 0x000175B40001D1B3,
-    0x000175B600017DB5, 0x80006E3700016DB7, 0x0002908480006E3B,
-    0x80006E43000DD1A3, 0x00023DA6000DF9A5, 0x0003ACBE80006E47,
-    0x000219AA80006E4B, 0x000209AC000219AB, 0x0002BDAE000209AD,
-    0x00035DB60002B5AF, 0x80006E5380006E4F, 0x80006E57000355B5,
-    0x0004F886000345B3, 0x000359BE00030887, 0x000359BC000321BF,
-    0x000359BA000359BD, 0x80006E5B000359BB, 0x80006CEF80006E5F,
-    0x80006E6780006E63, 0x80006E6F80006E6B, 0x80006E7780006E73,
-    0x0003A1A880006E7B, 0x0003B9AA0003ADA9, 0x0003A9AC0003B9AB,
-    0x000399AE0003A9AD, 0x0003E9B0000399AF, 0x0003F9B20003E9B1,
-    0x0003E9B40003F9B3, 0x000391B6000391B5, 0x0003B1B8000391B7,
-    0x000381BA0003B9B9, 0x000351BC000381BB, 0x000351BE000351BD,
-    0x000211A3000351BF, 0x80006E8380006E7F, 0x80006E8B80006E87,
-    0x000201A5000209A6, 0x00020DAB80006E8F, 0x80006E9300020DAA,
-    0x000275AF80006E97, 0x00020DAD00020DAE, 0x80006E9B00020DAC,
-    0x000390EF80006E9F, 0x0000098180006EA3, 0x80006EA700001980,
-    0x80006EAB00003D82, 0x000564BE00026884, 0x000190E180006EB3,
-    0x000360E380006EB7, 0x0004A08680006EBB, 0x80006EBF00050887,
-    0x80006EC300058484, 0x80006ECB80006EC7, 0x80006ECF000194EF,
-    0x80006ED3000E74E1, 0x000000ED000150E3, 0x80006EDB80006ED7,
-    0x0001E5B380006EDF, 0x80006EE780006EE3, 0x80006EEF80006EEB,
-    0x000191B5000185B6, 0x000169BB80006EF3, 0x80006EF7000161BA,
-    0x000169BF80006EFB, 0x000169BD000161BE, 0x000611A8000171BC,
-    0x00062DAA000619A9, 0x000659AC000649AB, 0x000661AE000645AD,
-    0x80006EAF000661AF, 0x80006F0380006EFF, 0x80006F0B80006F07,
-    0x80006F1380006F0F, 0x000749B880006F17, 0x000759BA000749B9,
-    0x000779BC000751BB, 0x000719BE000779BD, 0x0006E5B0000719BF,
-    0x0006E5B20006EDB1, 0x0006E5B40006FDB3, 0x0006E5B60006EDB5,
-    0x0006A5A3000779B7, 0x0000B58100001582, 0x80006F1B0000B580,
-    0x0006D1A50006C5A6, 0x000629AB000350BE, 0x80006F1F000621AA,
-    0x000629AF80006F23, 0x000629AD000621AE, 0x80006F27000631AC,
-    0x000068860007B9B3, 0x000745B600036C87, 0x80006F2F80006F2B,
-    0x000749BA00074DB5, 0x80006F33000751BB, 0x000735BE80006F37,
-    0x000741BC00073DBF, 0x80006F3B000741BD, 0x000498EF80006F3F,
-    0x80006F4780006F43, 0x80006F4F80006F4B, 0x80006F5780006F53,
-    0x80006F5F80006F5B, 0x000570E180006F63, 0x00044CE380006F67,
-    0x00063DA380006F6B, 0x80006F7380006F6F, 0x80006F7B80006F77,
-    0x0006C9A50006C1A6, 0x0006D5AB80006F7F, 0x80006F830006CDAA,
-    0x0006B9AF80006F87, 0x0006C5AD0006B1AE, 0x0006EDA80006C5AC,
-    0x0006F5AA0006F1A9, 0x000601AC000605AB, 0x000601AE000601AD,
-    0x0000C98000063DAF, 0x000005820000C981, 0x0001FCBE80006F8B,
-    0x80006F930001E084, 0x00061DB880006F97, 0x000625BA00062DB9,
-    0x0001F9BC0001E9BB, 0x0001E9BE0001F9BD, 0x000649B00001E9BF,
-    0x000651B2000651B1, 0x000635B4000651B3, 0x000635B600063DB5,
-    0x0002ADA800062DB7, 0x0002DDAA0002C5A9, 0x0002F9AC0002D5AB,
-    0x000319AE0002F9AD, 0x80006F9B000319AF, 0x80006FA380006F9F,
-    0x80006FAB80006FA7, 0x80006FB380006FAF, 0x00036DB880006FB7,
-    0x000375BA000371B9, 0x000319BC000309BB, 0x000309BE000319BD,
-    0x00036DB0000301BF, 0x00037DB2000375B1, 0x00036DB4000375B3,
-    0x00035DB6000355B5, 0x000269B5000355B7, 0x80006FBF80006FBB,
-    0x007C3086000261B6, 0x00024DB30003A487, 0x0003B9BD80006FC3,
-    0x000389BF0003B1BC, 0x80006FC70003B1BE, 0x0003A9BB80006E3F,
-    0x000215A30003B1BA, 0x80006FCF80006FCB, 0x0071D48580006FD3,
-    0x000231A5000239A6, 0x0003F1AB80006FD7, 0x80006FDB0003E9AA,
-    0x0003D1AF80006FDF, 0x0003E1AD0003E9AE, 0x80006FE30003E9AC,
-    0x80006FEB80006FE7, 0x80006FF380006FEF, 0x80006FF700005CEF,
-    0x000260BE80006FFB, 0x0002A484000080E3, 0x000021800001FCE1,
-    0x0000258200002D81, 0x000161B380007003, 0x007D648780007007,
-    0x8000700B007D6C86, 0x000171B5000179B6, 0x00015DBB8000700F,
-    0x8000701300015DBA, 0x0001C1BF80007017, 0x0001F1BD0001E9BE,
-    0x8000701B0001F9BC, 0x800070238000701F, 0x8000702B80007027,
-    0x800070338000702F, 0x007C2C840001F8EF, 0x8000703700063CE1,
-    0x8000703B000630E3, 0x800070438000703F, 0x8000704B80007047,
-    0x800070538000704F, 0x800070570001EDA3, 0x0001F5A60001FDA5,
-    0x8000705F8000705B, 0x0001D1AA80007063, 0x000175AC0001D1AB,
-    0x000165AE00017DAD, 0x007DB9A800014DAF, 0x007E29AA007DB9A9,
-    0x007E39AC007E29AB, 0x007E29AE007E39AD, 0x80006FFF007E29AF,
-    0x0000198100001182, 0x80007067007FC980, 0x000334BE8000706B,
-    0x007E8DB88000706F, 0x007EA1BA007E99B9, 0x007F51BC007EA1BB,
-    0x007F51BE007F51BD, 0x007E59B0007F51BF, 0x007EEDB2007E59B1,
-    0x007EEDB4007EE5B3, 0x007EBDB6007EC5B5, 0x007FBDA8007EB5B7,
-    0x007FD9AA007FC1A9, 0x007FF9AC007FD1AB, 0x007F19AE007FF9AD,
-    0x00036087007F19AF, 0x8000707300000486, 0x8000707B80007077,
-    0x800070838000707F, 0x007F79B880007087, 0x007F09BA007F79B9,
-    0x007F19BC007F09BB, 0x007F05BE007F05BD, 0x007F69B0007F3DBF,
-    0x007F79B2007F69B1, 0x007F69B4007F79B3, 0x007F49B6007F5DB5,
-    0x8000708B007F49B7, 0x8000708F007CDCEF, 0x8000709780007093,
-    0x8000709F8000709B, 0x800070A7800070A3, 0x800070AF800070AB,
-    0x800070B7800070B3, 0x800070BB007C0CE3, 0x007E11B3007DB4E1,
-    0x800070C3800070BF, 0x800070CB800070C7, 0x007E31B5007E39B6,
-    0x007EEDBB800070CF, 0x800070D3007EEDBA, 0x007ED9BF800070D7,
-    0x007EEDBD007EEDBE, 0x007ED1A3007EEDBC, 0x00028884800070DB,
-    0x800070E3000190BE, 0x007EF1A5007EF9A6, 0x007E2DAB800070E7,
-    0x00000886007E2DAA, 0x007E19AF00014C87, 0x007E2DAD007E2DAE,
-    0x0000A980007E2DAC, 0x00004D820000A981, 0x800070EB00017DB3,
-    0x000165B600016DB5, 0x800070F3800070EF, 0x0000E5BA800070F7,
-    0x0000E5BC0000FDBB, 0x0000E5BE0000EDBD, 0x0002A9A80000DDBF,
-    0x0002B9AA0002A9A9, 0x0002A9AC0002B9AB, 0x000319AE0002A9AD,
-    0x0004ACBE000315AF, 0x800070FF800070FB, 0x800071030005E0BE,
-    0x8000710B80007107, 0x00036DB88000710F, 0x000371BA000375B9,
-    0x000319BC000309BB, 0x000309BE000319BD, 0x00036DB0000301BF,
-    0x00037DB2000375B1, 0x00036DB4000375B3, 0x00035DB6000355B5,
-    0x80007113000355B7, 0x000231A5000239A6, 0x000221A380007117,
-    0x8000711F8000711B, 0x000381AF80007123, 0x0003B1AD0003B9AE,
-    0x0003A1AB0003B9AC, 0x80006F8F0003B9AA, 0x0003C484000674BE,
-    0x8000712B80007127, 0x800071338000712F, 0x8000713B80007137,
-    0x00000D808000713F, 0x00001D8200001581, 0x8000714780007143,
-    0x0004B8878000714B, 0x000291A80004DC86, 0x0002D9AA000299A9,
-    0x0002F1AC0002D1AB, 0x00013DAE0002F1AD, 0x80007153000135AF,
-    0x8000715B80007157, 0x800071638000715F, 0x8000716B80007167,
-    0x0001DDB88000716F, 0x000181BA000181B9, 0x000181BC000181BB,
-    0x000181BE000181BD, 0x000151B0000181BF, 0x000151B2000151B1,
-    0x0001F1B4000151B3, 0x0001EDB60001F9B5, 0x0000F4EF0001E5B7,
-    0x8000717780007173, 0x8000717B000800BE, 0x000518848000717F,
-    0x8000718300015CEF, 0x80007187007EF0E1, 0x000028E3007E1CE3,
-    0x000104E18000718B, 0x0029FC858000718F, 0x000219B380007193,
-    0x000231B580007197, 0x8000719F8000719B, 0x800071A3000239B6,
-    0x0001DDBB800071A7, 0x0001CDBD0001DDBA, 0x000175BF0001CDBC,
-    0x0005D1A300017DBE, 0x800071AB8000714F, 0x800071B3800071AF,
-    0x0005F9A50005F1A6, 0x000615AB800071B7, 0x800071BB000615AA,
-    0x0006BDAF800071BF, 0x000605AD0006B5AE, 0x00005180000605AC,
-    0x0000618200005981, 0x800071C3000765B3, 0x00076DB6000765B5,
-    0x00034086800071C7, 0x0007EDBA0003C487, 0x0007E1BC0007E5BB,
-    0x0007D9BE0007E1BD, 0x0007A1A80007D5BF, 0x0007A1AA0007A1A9,
-    0x0007A1AC0007A1AB, 0x0007A1AE0007A1AD, 0x800071CB0007A1AF,
-    0x800071D3800071CF, 0x800071DB800071D7, 0x800071E3800071DF,
-    0x000705B8800071E7, 0x00071DBA000709B9, 0x000731BC000715BB,
-    0x0000D1BE000731BD, 0x000765B00000D1BF, 0x000779B200076DB1,
-    0x000751B4000779B3, 0x00074DB600075DB5, 0x800071EB000741B7,
-    0x800071F3800071EF, 0x0005C8EF800071F7, 0x800071FF800071FB,
-    0x0005F0E380007203, 0x000698E180007207, 0x8000720F8000720B,
-    0x8000721780007213, 0x000669A38000721B, 0x800072238000721F,
-    0x8000722B80007227, 0x000669A5000661A6, 0x0006E9AB00019084,
-    0x8000722F0006E1AA, 0x0006D9AF80007233, 0x0006EDAD0006D5AE,
-    0x0001D1A80006EDAC, 0x000129AA0001D9A9, 0x000139AC000129AB,
-    0x000129AE000139AD, 0x0000E980000129AF, 0x0000F9820000E981,
-    0x0000108680007237, 0x8000723B00017487, 0x0000DDB88000723F,
-    0x0000E5BA0000E1B9, 0x0000E5BC0000FDBB, 0x0000E5BE0000EDBD,
-    0x000159B0000099BF, 0x0000E9B2000159B1, 0x0000F9B40000E9B3,
-    0x0000EDB60000F9B5, 0x0002A1B30000E5B7, 0x8000724780007243,
-    0x8000724F8000724B, 0x0002B1B50002B9B6, 0x00029DBB80007253,
-    0x8000725700029DBA, 0x00033DBF8000725B, 0x00033DBD00033DBE,
-    0x002C048500033DBC, 0x8000725F0002E5A3, 0x0002FDA680007263,
-    0x8000726B80007267, 0x0002D9AA0002F5A5, 0x8000726F0002D9AB,
-    0x000379AE80007273, 0x000379AC000379AF, 0x00039084000379AD,
-    0x8000727B80007277, 0x8000727F000CC084, 0x8000728780007283,
-    0x8000728F8000728B, 0x8000729780007293, 0x8000729B800070DF,
-    0x800072A7800072A3, 0x00006580800072AB, 0x0000798200006D81,
-    0x800072B3800072AF, 0x00036C870007A0EF, 0x800072B7000C5C86,
-    0x800072BB000274E3, 0x800072BF000144E1, 0x800072C7800072C3,
-    0x00026DA8800072CB, 0x0002B1AA0002A9A9, 0x000115AC0002B1AB,
-    0x000115AE00011DAD, 0x000D68BE00010DAF, 0x800072D3800072CF,
-    0x800072DB800072D7, 0x800072E3800072DF, 0x00012DB8800072E7,
-    0x0001C1BA0001C1B9, 0x0001C5BC0001DDBB, 0x0001C5BE0001CDBD,
-    0x00017DB00001FDBF, 0x000145B200014DB1, 0x00013DB4000125B3,
-    0x000121B6000121B5, 0x000764E1000115B7, 0x000600E3800072EB,
-    0x0007D4EF800072EF, 0x800072F7800072F3, 0x0000ECE3800072FB,
-    0x0001D0E1800072FF, 0x0000718180007303, 0x00076CEF00001980,
-    0x8000730700006D82, 0x0001C1B60001C1B5, 0x8000730B000C4C84,
-    0x8000731780007313, 0x0001CDBC000265B3, 0x0001A5BE0001ADBD,
-    0x000CD0860001B9BF, 0x0001E5BA000CB487, 0x8000729F0001D5BB,
-    0x8000731B8000730F, 0x800073238000731F, 0x8000732B80007327,
-    0x000E95A88000732F, 0x000E95AA000E9DA9, 0x000EE1AC000EE1AB,
-    0x000ED1AE000ED5AD, 0x000EB5B0000ECDAF, 0x000EB5B2000EBDB1,
-    0x000F79B4000F69B3, 0x000F69B6000F79B5, 0x000F59B8000F69B7,
-    0x000FEDBA000F59B9, 0x000FE5BC000FE1BB, 0x000FE5BE000FEDBD,
-    0x000E3DB3000FD9BF, 0x8000733780007333, 0x8000733F8000733B,
-    0x000E49B5000E75B6, 0x000E59BB80007343, 0x80007347000E51BA,
-    0x000F9DBF8000734B, 0x000F9DBD000F9DBE, 0x8000734F000E41BC,
-    0x80007353000E79A3, 0x000E31A680007357, 0x8000735F8000735B,
-    0x000E15AA000E0DA5, 0x80007363000E1DAB, 0x000FD9AE80007367,
-    0x000E05AC000FD9AF, 0x000CB9A3000FD9AD, 0x8000736F8000736B,
-    0x8000737780007373, 0x000F1DA5000F1DA6, 0x000F09AB8000737B,
-    0x0001A8BE000F39AA, 0x000F65AF8000737F, 0x000F71AD000F79AE,
-    0x00004D80000F11AC, 0x0000558200005581, 0x80007383000145B3,
-    0x00014DB6000145B5, 0x0000C08680007387, 0x0001EDBA00011487,
-    0x0001FDBC0001E5BB, 0x0001E5BE0001E5BD, 0x000E55A80001D5BF,
-    0x000181AA000E5DA9, 0x000181AC000181AB, 0x000181AE000181AD,
-    0x00018C84000181AF, 0x8000738F8000738B, 0x8000739780007393,
-    0x8000739F8000739B, 0x000109B8800073A3, 0x00011DBA000109B9,
-    0x000139BC000115BB, 0x0000DDBE000139BD, 0x000165B00000D5BF,
-    0x000165B200016DB1, 0x00016DB400017DB3, 0x000155B600015DB5,
-    0x800073A700013DB7, 0x800073AF800073AB, 0x800073B3000205A3,
-    0x00020DA6000205A5, 0x800073BB800073B7, 0x0002ADAA800073BF,
-    0x0002BDAC0002A5AB, 0x0002A5AE0002A5AD, 0x800073C3000295AF,
-    0x000304BE800073C7, 0x800073CB00040084, 0x800073D3800073CF,
-    0x800073DB800073D7, 0x800073E3800073DF, 0x00001980800073E7,
-    0x0000058200001981, 0x00042086800073EB, 0x800073F300031487,
-    0x000220EF800073F7, 0x800073FF800073FB, 0x000110E380007403,
-    0x0001D0E180007407, 0x8000740F8000740B, 0x8000741780007413,
-    0x0000C1BB8000741B, 0x8000741F000379BA, 0x0000CDBF80007423,
-    0x0000D1BD0000C5BE, 0x000345B30000D9BC, 0x8000742B80007427,
-    0x800074338000742F, 0x000355B5000345B6, 0x000251A880007437,
-    0x0002ADAA000259A9, 0x0002BDAC0002A5AB, 0x0002ADAE0002A5AD,
-    0x000568BE0002A5AF, 0x8000743F8000743B, 0x8000744780007443,
-    0x8000744F8000744B, 0x000101B880007453, 0x000119BA000101B9,
-    0x000109BC000119BB, 0x000139BE000109BD, 0x0002DDB0000139BF,
-    0x00016DB2000165B1, 0x000161B4000165B3, 0x000159B6000169B5,
-    0x000E68E1000159B7, 0x000E54E3000F40E1, 0x0000AD80000F10E3,
-    0x0000398200000981, 0x8000745B80007457, 0x8000746780007463,
-    0x000678BE8000746B, 0x000068EF8000746F, 0x000265A3000E64EF,
-    0x0007888680007473, 0x00061C840004A887, 0x000275A5000265A6,
-    0x0001E1AB80007477, 0x8000747B000259AA, 0x0001EDAF8000747F,
-    0x0001F1AD0001E5AE, 0x000689A80001F9AC, 0x00069DAA000689A9,
-    0x00068DAC000695AB, 0x0006B5AE0006B1AD, 0x8000745F0006ADAF,
-    0x8000748780007483, 0x8000748F8000748B, 0x8000749780007493,
-    0x0006B1B88000749B, 0x00076DBA0006B9B9, 0x00077DBC000765BB,
-    0x00076DBE000765BD, 0x0006D5B0000765BF, 0x0006D5B20006DDB1,
-    0x0006BDB40006A5B3, 0x000691B6000691B5, 0x000615B3000691B7,
-    0x800074A38000749F, 0x800074AB800074A7, 0x000605B5000631B6,
-    0x000619BB800074AF, 0x800074B3000611BA, 0x000655BF800074B7,
-    0x000671BD000679BE, 0x800074BB000601BC, 0x800074BF000651A3,
-    0x000675A6800074C3, 0x800074CB800074C7, 0x000655AA000641A5,
-    0x001E488600065DAB, 0x00063DAE00003C87, 0x000645AC000611AF,
-    0x0008699F000635AD, 0x00090D9D0009299E, 0x000D2D9B000BA19C,
-    0x000FA199000D359A, 0x007101970071B598, 0x0075DD9500739196,
-    0x0077B19300750D94, 0x00793D9100780592, 0x007D718F00793D90,
-    0x800074D30005DCE1, 0x007194E1000510E3, 0x000C8CE3800074D7,
-    0x0001B581800074DB, 0x0078B4EF0001A980, 0x800074DF0001B582,
-    0x800074E7800074E3, 0x0001A1A10004A4EF, 0x001DA9A3800074EB,
-    0x001C39A5001C01A2, 0x0019C1A7001DA5A4, 0x0015A5A900199DA6,
-    0x001485AB001401A8, 0x001191AD0015A5AA, 0x002C01AF001001AC,
-    0x002D51B1001089AE, 0x0029E9B3002DE5B0, 0x002801B5002801B2,
-    0x800074EF0029E9B4, 0x800074F7800074F3, 0x800074FF800074FB,
-    0x000331A380007503, 0x0005FDA10004D9A2, 0x007E31820005FDA0,
-    0x80007507007A0583, 0x007491868000750B, 0x007A1D8400763587,
-    0x0076D58A007B2D85, 0x8000750F0070918B, 0x000C898E80007513,
-    0x0072198C000E158F, 0x000E7D920072958D, 0x000344840008E993,
-    0x000B3996001C40BE, 0x000A5D9400065597, 0x0006659A000A7195,
-    0x800075170007499B, 0x8000751F8000751B, 0x0000B99C80007523,
-    0x0002A5B080007527, 0x0002A5B20002ADB1, 0x0002A5B40002BDB3,
-    0x0002A5B60002ADB5, 0x000229B8000219B7, 0x000239BA000229B9,
-    0x000219BC000231BB, 0x000279BE000219BD, 0x800073EF000275BF,
-    0x8000752F8000752B, 0x00001D8080007533, 0x0001D18200000981,
-    0x000545A880007537, 0x000545AA00054DA9, 0x000535AC000559AB,
-    0x000511AE000529AD, 0x0004ADB3000511AF, 0x8000753B001D0C84,
-    0x8000754780007543, 0x0004DDB50004C9B6, 0x000309BB8000754B,
-    0x001C8886000301BA, 0x000309BF001D0087, 0x000309BD000301BE,
-    0x8000754F000311BC, 0x800075530004E9A3, 0x00048DA680007557,
-    0x8000755F8000755B, 0x000345AA000499A5, 0x8000756300034DAB,
-    0x000345AE80007567, 0x000355AC00034DAF, 0x8000756B00034DAD,
-    0x800075738000756F, 0x8000757B80007577, 0x800075838000757F,
-    0x800075870006D4EF, 0x8000758B000674E1, 0x8000758F00066CE3,
-    0x8000759780007593, 0x00029DB38000759B, 0x800075A38000759F,
-    0x800075AB800075A7, 0x0002ADB50002ADB6, 0x000201BB001CD0BE,
-    0x001FCC86000239BA, 0x0001F9BF001C2C87, 0x000211BD000219BE,
-    0x001D49A8000219BC, 0x001D55AA001D51A9, 0x001D65AC001D7DAB,
-    0x001D9DAE001D6DAD, 0x8000753F001DC5AF, 0x001FE581001FF582,
-    0x800075AF001FE580, 0x800075B7800075B3, 0x001E6DB8800075BB,
-    0x001E25BA001E2DB9, 0x001E25BC001E3DBB, 0x001E9DBE001E29BD,
-    0x001DBDB0001E95BF, 0x001D8DB2001D85B1, 0x001D9DB4001D85B3,
-    0x001E75B6001E7DB5, 0x001DDDA3001E55B7, 0x800075C3800075BF,
-    0x800075CB800075C7, 0x001DEDA5001DEDA6, 0x001D41AB800075CF,
-    0x800075D3001D79AA, 0x001EB9AF800075D7, 0x001D51AD001D59AE,
-    0x800075DB001D59AC, 0x800075DF001FD9B3, 0x001F9DB6800075E3,
-    0x800075EB800075E7, 0x001F89BA001F95B5, 0x800075EF001F89BB,
-    0x001F79BE800075F3, 0x001F79BC001F79BF, 0x001E0DA8001F79BD,
-    0x001E15AA001E15A9, 0x001E59AC001E4DAB, 0x001E4DAE001E59AD,
-    0x800075F7001E45AF, 0x800075FF800075FB, 0x8000760780007603,
-    0x8000760F8000760B, 0x001ED9B880007613, 0x001EE1BA001EE5B9,
-    0x001EE5BC001EFDBB, 0x001EE5BE001EEDBD, 0x001E09B0001E99BF,
-    0x001EE9B2001E09B1, 0x001EF9B4001EE9B3, 0x001EE9B6001EF9B5,
-    0x001E9DA3001EE9B7, 0x8000761B80007617, 0x00029084800074CF,
-    0x001ED1A5001ED9A6, 0x001ECDAB000158BE, 0x8000761F001ECDAA,
-    0x001E3DAF80007623, 0x001E3DAD001E3DAE, 0x0000FD80001E3DAC,
-    0x0000418200004181, 0x800076270001F1B3, 0x000155B600015DB5,
-    0x000040868000762B, 0x000141BA00006487, 0x000131BC000141BB,
-    0x000135BE00013DBD, 0x001E58E3000129BF, 0x001CF8E18000762F,
-    0x8000763300030084, 0x8000763700024CEF, 0x8000763F8000763B,
-    0x80007643000238E3, 0x001E84EF0001D4E1, 0x8000764B80007647,
-    0x800076538000764F, 0x000275A380007657, 0x0002D9A58000765B,
-    0x800076638000765F, 0x800076670002D1A6, 0x0002C5AB8000766B,
-    0x0002B9AD0002C5AA, 0x0002ADAF0002B5AC, 0x0002B9A90002B1AE,
-    0x00020DAB0002B1A8, 0x00021DAD0002B1AA, 0x000205AF00021DAC,
-    0x000400BE00020DAE, 0x800076738000766F, 0x8000767B80007677,
-    0x800076838000767F, 0x0003E1B980007687, 0x0003E1BB000229B8,
-    0x0003E1BD0003E9BA, 0x0003E1BF0003F9BC, 0x00024DB10003E9BE,
-    0x000221B300024DB0, 0x000225B500022DB2, 0x000211B7000239B4,
-    0x0002ADA8000225B6, 0x0002D5AA0002D5A9, 0x000101AC000109AB,
-    0x000105AE00010DAD, 0x8000768B00013DAF, 0x8000769780007693,
-    0x8000769F8000769B, 0x800076A7800076A3, 0x000119B8800076AB,
-    0x00012DBA000125B9, 0x0001F9BC0001E9BB, 0x0001E9BE0001F9BD,
-    0x000149B00001E9BF, 0x000151B2000151B1, 0x000131B4000151B3,
-    0x000135B6000131B5, 0x00013980000129B7, 0x0000A98200013981,
-    0x800076AF0000E8EF, 0x800076B7800076B3, 0x00046C84001EB8EF,
-    0x800076BB001E94E1, 0x800076BF001EB8E3, 0x800076C3001E44E1,
-    0x00021DB300003CE3, 0x00058887800076C7, 0x800076CB0005EC86,
-    0x0001C9B50001CDB6, 0x0001D5BB800076CF, 0x800076D30001CDBA,
-    0x0001B5BF800076D7, 0x0001C5BD0001B5BE, 0x8000768F0001C5BC,
-    0x000621AA800076DB, 0x000629AC000639AB, 0x000659AE000629AD,
-    0x800076DF000659AF, 0x800076E7800076E3, 0x800076EB0005F1A3,
-    0x000621A6000625A5, 0x00000D80800076EF, 0x00001D8200001581,
-    0x800076F7800076F3, 0x800076FF800076FB, 0x00034086000394BE,
-    0x800077030003C487, 0x8000770B80007707, 0x800077138000770F,
-    0x0007A5A880007717, 0x0007A5AA0007ADA9, 0x0007A5AC0007BDAB,
-    0x0007A5AE0007ADAD, 0x8000771B00070DAF, 0x800077238000771F,
-    0x8000772B80007727, 0x800077338000772F, 0x000729B880007737,
-    0x0000CDBA000729B9, 0x0000F5BC0000EDBB, 0x0000E9BE0000FDBD,
-    0x000775B00000E1BF, 0x000741B200077DB1, 0x000749B400075DB3,
-    0x000729B6000731B5, 0x00062DB3000721B7, 0x8000773F8000773B,
-    0x8000774780007743, 0x00062DB500062DB6, 0x0006B9BB8000774B,
-    0x8000774F0006B1BA, 0x0006F5BF80007753, 0x000689BD00068DBE,
-    0x80007757000695BC, 0x8000775B000669A3, 0x000669A68000775F,
-    0x8000776780007763, 0x0006F5AA000669A5, 0x8000776B0006FDAB,
-    0x0006C9AE0001A084, 0x0006D1AC0006B1AF, 0x0001E9A80006CDAD,
-    0x0001F9AA0001E9A9, 0x0001E9AC0001F9AB, 0x000139AE0001E9AD,
-    0x0000F580000139AF, 0x0000C1820000FD81, 0x00019C86000164BE,
-    0x800077730000A087, 0x0000CDB880007777, 0x0000D5BA0000D5B9,
-    0x0000FDBC0000E5BB, 0x000091BE000095BD, 0x000149B0000091BF,
-    0x000159B2000149B1, 0x000149B4000159B3, 0x0000FDB6000149B5,
-    0x000271B30000F5B7, 0x8000777F8000777B, 0x8000778780007783,
-    0x000261B5000269B6, 0x00028DBB8000778B, 0x8000778F00028DBA,
-    0x00028DBF80007793, 0x00028DBD00028DBE, 0x0000009200028DBC,
-    0x80007797000235A3, 0x00022DA68000779B, 0x800077A38000779F,
-    0x0002C9AA000225A5, 0x800077A70002C9AB, 0x0002C9AE800077AB,
-    0x0002C9AC0002C9AF, 0x000089AB0002C9AD, 0x000335A9000081AA,
-    0x0000B9AF000335A8, 0x000089AD000089AE, 0x000000AA000091AC,
-    0x800077AF000C28BE, 0x800077B7800077B3, 0x800077BF800077BB,
-    0x000079BB800077C3, 0x000079B9000071BA, 0x0001D9BF000061B8,
-    0x0001D9BD0001D1BE, 0x0000B5B3000061BC, 0x0000A1B10000A9B2,
-    0x000059B70000C1B0, 0x000091B5000099B6, 0x800077C70000A5B4,
-    0x800077CF800077CB, 0x00000981800077D3, 0x800077D700001980,
-    0x0002AC8400003982, 0x0002C087800077DB, 0x800077E3000D0C86,
-    0x800077EB800077E7, 0x000145B3800077EF, 0x800077F7800077F3,
-    0x800077FF800077FB, 0x000155B5000145B6, 0x000129BB80007803,
-    0x80007807000149BA, 0x000119BF8000780B, 0x00012DBD00012DBE,
-    0x8000780F000131BC, 0x80007813000650E1, 0x8000781700063CE3,
-    0x8000781F8000781B, 0x000D288480007823, 0x000C24BE80007827,
-    0x8000782F8000782B, 0x8000783780007833, 0x8000783B000118EF,
-    0x000CE8868000783F, 0x80007843000C8C87, 0x800078470000A8E3,
-    0x000011820001A4E1, 0x000039800000B0EF, 0x8000784B00003581,
-    0x8000785780007853, 0x000E05A38000785B, 0x8000784F800077DF,
-    0x800078638000785F, 0x000E15A5000E05A6, 0x000E69AB80007867,
-    0x8000786B000E09AA, 0x000E59AF8000786F, 0x000E6DAD000E6DAE,
-    0x80007873000E71AC, 0x80007877000F7DB3, 0x000F19B68000787B,
-    0x800078838000787F, 0x000F3DBA000F19B5, 0x80007887000F0DBB,
-    0x000FF9BE8000788B, 0x000F15BC000FE5BF, 0x000FB1A8000FFDBD,
-    0x000FB1AA000FB1A9, 0x000FD5AC000FB1AB, 0x000FD5AE000FDDAD,
-    0x8000788F000FCDAF, 0x8000789780007893, 0x8000789F8000789B,
-    0x800078A7800078A3, 0x000F5DB8800078AB, 0x000F65BA000F61B9,
-    0x000F65BC000F7DBB, 0x000F19BE000F6DBD, 0x000FBDB0000F19BF,
-    0x000F85B2000F8DB1, 0x000F79B4000F6DB3, 0x000F6DB6000F65B5,
-    0x000E39A3000F65B7, 0x800078B3800078AF, 0x800078BB800078B7,
-    0x000E5DA5000E5DA6, 0x000E49AB00029C84, 0x0001E4BE000E79AA,
-    0x000EA1AF800078BF, 0x000EB9AD000EBDAE, 0x00004980000E51AC,
-    0x0000598200004981, 0x800078C3000EB1B3, 0x000145B6000145B5,
-    0x0000C086800078C7, 0x000129BA0001E487, 0x00013DBC000125BB,
-    0x000121BE000121BD, 0x000E75A8000115BF, 0x000E75AA000E7DA9,
-    0x000181AC000181AB, 0x0001B1AE000189AD, 0x0001A0840001B1AF,
-    0x800078CB8000776F, 0x800078D3800078CF, 0x800078DB800078D7,
-    0x000165B8800078DF, 0x000165BA00016DB9, 0x000165BC00017DBB,
-    0x000165BE00016DBD, 0x0001A1B00000D9BF, 0x0001A5B20001ADB1,
-    0x0001A9B40001B9B3, 0x000195B600019DB5, 0x800078E300015DB7,
-    0x800078EB800078E7, 0x800078F3800078EF, 0x000E44EF800078F7,
-    0x000D88E1800078FB, 0x000F18E3800078FF, 0x8000790780007903,
-    0x8000790F8000790B, 0x000DBDA380007913, 0x8000791B80007917,
-    0x800079238000791F, 0x000249A5000249A6, 0x000229AB80007927,
-    0x8000792B000225AA, 0x000219AF000320BE, 0x00022DAD00022DAE,
-    0x0000D180000231AC, 0x0000E1820000D981, 0x000204EF0004C084,
-    0x0003548780007933, 0x000110E300043C86, 0x0001D0E180007937,
-    0x8000793F8000793B, 0x8000794780007943, 0x8000794F8000794B,
-    0x8000795780007953, 0x8000795B000365B3, 0x00036DB6000365B5,
-    0x800079638000795F, 0x0000E9BA80007967, 0x0000F9BC0000E9BB,
-    0x0000E9BE0000F9BD, 0x000251A80000E9BF, 0x0002A1AA000259A9,
-    0x0002A1AC0002A1AB, 0x0002A1AE0002A1AD, 0x000568BE0002A1AF,
-    0x8000796F8000796B, 0x8000797780007973, 0x8000797F8000797B,
-    0x00016DB880007983, 0x00017DBA000175B9, 0x00016DBC000175BB,
-    0x000111BE000111BD, 0x000161B0000111BF, 0x000161B2000161B1,
-    0x000161B4000161B3, 0x000165B6000161B5, 0x000EB4E1000159B7,
-    0x000E38E3000088E3, 0x80007987000E5CE1, 0x8000798F8000798B,
-    0x8000799780007993, 0x800079A38000799B, 0x0004FCBE800079A7,
-    0x000EE8EF000064EF, 0x000269A3800079AB, 0x0000698100006182,
-    0x0004848400005180, 0x000269A5000261A6, 0x0001E5AB800079AF,
-    0x000448860001E5AA, 0x0001E5AF0004CC87, 0x0001F5AD0001E5AE,
-    0x000695A80001F5AC, 0x0006A5AA0006A1A9, 0x0006A5AC0006BDAB,
-    0x0006A5AE0006ADAD, 0x8000799F0006DDAF, 0x800079B7800079B3,
-    0x800079BF800079BB, 0x800079C7800079C3, 0x00076DB8800079CB,
-    0x00077DBA000775B9, 0x00076DBC000775BB, 0x0007CDBE0007C5BD,
-    0x0006A5B00007C5BF, 0x0006B9B20006ADB1, 0x000691B40006B1B3,
-    0x00075DB6000691B5, 0x00062DB3000755B7, 0x800079D3800079CF,
-    0x800079DB800079D7, 0x00062DB500062DB6, 0x000669BB800079DF,
-    0x800079E3000661BA, 0x000655BF800079E7, 0x000665BD000655BE,
-    0x800079EB00066DBC, 0x800079EF000669A3, 0x000669A6800079F3,
-    0x800079FB800079F7, 0x000625AA000669A5, 0x800079FF00062DAB,
-    0x000611AE80007A03, 0x000629AC000611AF, 0x0006A9A8000621AD,
-    0x0006B5AA0006A9A9, 0x000135AC0006D1AB, 0x000135AE00013DAD,
-    0x0001CD8000012DAF, 0x0000198200000981, 0x80007A0B80007A07,
-    0x0001788780007A0F, 0x0001E9B800009C86, 0x000189BA0001E9B9,
-    0x00019DBC000185BB, 0x000181BE000181BD, 0x000155B00001B5BF,
-    0x000155B200015DB1, 0x0001F9B40001E9B3, 0x0001E9B60001F9B5,
-    0x0001A5B50001E1B7, 0x80007A1780007A13, 0x80007A1B0001A5B6,
-    0x000625B380007A1F, 0x000125BD80007A23, 0x000125BF000199BC,
-    0x80007A27000125BE, 0x000181BB8000792F, 0x00067DA3000189BA,
-    0x80007A2F80007A2B, 0x80007A3780007A33, 0x0001FDA50001FDA6,
-    0x0001D9AB80007A3B, 0x80007A3F0001D1AA, 0x00017DAF80007A43,
-    0x00017DAD00017DAE, 0x80007A470001C1AC, 0x0007D0EF80007A4B,
-    0x80007A5380007A4F, 0x80007A5B80007A57, 0x80007A6380007A5F,
-    0x80007A6B80007A67, 0x0006B4E180007A6F, 0x0006A4E380007A73,
-    0x0000758080007A77, 0x0000198200007981, 0x80007A7B00024084,
-    0x000314BE80007A83, 0x003CC08680007A87, 0x80007A8B00034487,
-    0x80007A9380007A8F, 0x80007A9B80007A97, 0x80007AA380007A9F,
-    0x000228EF80007AA7, 0x80007AAF80007AAB, 0x80007AB780007AB3,
-    0x003CEC8480007ABB, 0x80007AC380007ABF, 0x000194E180007AC7,
-    0x0002E4E380007ACB, 0x000351B380007ACF, 0x80007AD780007AD3,
-    0x80007ADF80007ADB, 0x000345B5000365B6, 0x000331BB80007AE3,
-    0x80007AE7000329BA, 0x0000F1BF80007AEB, 0x000311BD000319BE,
-    0x000285A8000321BC, 0x000299AA00028DA9, 0x00028DAC000299AB,
-    0x0002DDAE0002A5AD, 0x0001FD800002D5AF, 0x0001C5820001CD81,
-    0x003C9086003D64BE, 0x80007AF3003D0487, 0x000151B880007AF7,
-    0x000151BA000151B9, 0x000171BC000151BB, 0x000171BE000171BD,
-    0x0002ADB0000171BF, 0x0002B5B20002B5B1, 0x00029DB4000285B3,
-    0x00017DB6000175B5, 0x80007AFB000171B7, 0x80007B0380007AFF,
-    0x80007B0B80007B07, 0x80007B1380007B0F, 0x003CECBE0006D4EF,
-    0x80007B17000654E1, 0x80007B1B00066CE3, 0x80007B2380007B1F,
-    0x000251A380007B27, 0x80007B2F80007B2B, 0x80007B3780007B33,
-    0x000245A5000265A6, 0x000231AB80007B3B, 0x80007B3F000229AA,
-    0x0001F1AF80007B43, 0x000211AD000219AE, 0x003E55A8000221AC,
-    0x003E61AA003E59A9, 0x003E61AC003E61AB, 0x003E61AE003E61AD,
-    0x80007AEF003E61AF, 0x80007B4B80007B47, 0x00000D8080007B4F,
-    0x0000B1820000B181, 0x003EE9B880007B53, 0x003EFDBA003EE9B9,
-    0x003E95BC003EF1BB, 0x003E95BE003E9DBD, 0x003EE1B0003E8DBF,
-    0x003EE5B2003EE1B1, 0x003EE5B4003EFDB3, 0x003EE5B6003EEDB5,
-    0x003EE9B3003ED9B7, 0x80007B5B80007B57, 0x80007B5F0003D084,
-    0x003EE9B5003EE1B6, 0x003E25BB80007B63, 0x00009886003E25BA,
-    0x003E29BF0000A487, 0x003E29BD003E21BE, 0x80007B67003E31BC,
-    0x80007B6B003EADA3, 0x003EA5A680007B6F, 0x80007B7780007B73,
-    0x003E61AA003EADA5, 0x80007B7B003E61AB, 0x003E65AE80007B7F,
-    0x003E75AC003E6DAF, 0x003E39A8003E6DAD, 0x003E89AA003E39A9,
-    0x003E99AC003E89AB, 0x003E89AE003E99AD, 0x80007B83003E89AF,
-    0x80007B8B80007B87, 0x80007B9380007B8F, 0x80007B9B80007B97,
-    0x003E8DB880007B9F, 0x003E95BA003E95B9, 0x003EBDBC003EA5BB,
-    0x000171BE000171BD, 0x003EF9B0000171BF, 0x003ED9B2003ECDB1,
-    0x003EC9B4003ED9B3, 0x003EBDB6003EC9B5, 0x80007BA3003EB5B7,
-    0x003E01B5003E09B6, 0x003E11B380007BA7, 0x80007BAF80007BAB,
-    0x003E1DBF80007BB3, 0x003E1DBD003E1DBE, 0x003E19BB003E1DBC,
-    0x80007A7F003E2DBA, 0x003E41A380007BB7, 0x80007BBF80007BBB,
-    0x80007BC780007BC3, 0x003E51A5003E59A6, 0x003E49AB80007BCB,
-    0x00016484003E7DAA, 0x003E4DAF00002CBE, 0x003E4DAD003E4DAE,
-    0x0003DD82003E4DAC, 0x0003C580003FD0EF, 0x000350870003DD81,
-    0x80007BD30004B486, 0x80007BDB80007BD7, 0x80007BE380007BDF,
-    0x0000008C80007BE7, 0x80007BEB003E84E3, 0x000289B3003E44E1,
-    0x0003448480007BEF, 0x80007BF3000440BE, 0x00028DB500028DB6,
-    0x000345BB80007BF7, 0x80007BFB000345BA, 0x000345BF80007BFF,
-    0x000355BD000345BE, 0x000255AC000355BC, 0x000261AE00025DAD,
-    0x00020DA8000261AF, 0x000251AA000255A9, 0x0004A884000251AB,
-    0x80007C0780007C03, 0x0005BCBE80007C0B, 0x80007C1380007C0F,
-    0x00035DBC80007C17, 0x00030DBE000379BD, 0x000351B8000305BF,
-    0x00034DBA000359B9, 0x0003D1B4000345BB, 0x0003D1B60003D1B5,
-    0x0003D1B00003D1B7, 0x0003D1B20003D1B1, 0x80007C1B0003D1B3,
-    0x80007C2380007C1F, 0x80007C2B80007C27, 0x80007C2F0001A4E1,
-    0x80007C330000E4E3, 0x80007C3B80007C37, 0x80007C43000038EF,
-    0x80007C4B80007C47, 0x0000618080007C4F, 0x0000618200006181,
-    0x80007C53000385A3, 0x000381A6000381A5, 0x0004C08600079884,
-    0x000249AA00054887, 0x000259AC000249AB, 0x000249AE000259AD,
-    0x80007C57000249AF, 0x80007C5F80007C5B, 0x80007C6780007C63,
-    0x80007C6F80007C6B, 0x000620BE003EC4EF, 0x80007C73003E8CE1,
-    0x80007C77000130E3, 0x80007C7F80007C7B, 0x0001E1B380007C83,
-    0x80007C8B80007C87, 0x80007C9380007C8F, 0x0001F1B50001F9B6,
-    0x0001BDBB80007C97, 0x80007C9B0001B5BA, 0x000159BF80007C9F,
-    0x000195BD000151BE, 0x000619A8000195BC, 0x000629AA000619A9,
-    0x000639AC000629AB, 0x000621AE000639AD, 0x80007C3F000659AF,
-    0x00001D8100001D82, 0x80007CA300001D80, 0x80007CAB80007CA7,
-    0x0006F9B880007CAF, 0x000749BA0006F9B9, 0x000759BC000749BB,
-    0x000749BE000759BD, 0x000631B0000745BF, 0x000601B2000639B1,
-    0x0006D5B4000601B3, 0x0006D5B60006DDB5, 0x0006A1A30006C9B7,
-    0x0000AC8680007CB3, 0x80007CB700034C87, 0x0006B1A50006B9A6,
-    0x0006FDAB80007CBB, 0x80007CBF0006F5AA, 0x000619AF80007CC3,
-    0x0006D5AD000611AE, 0x80007CC70006D5AC, 0x80007CCB0007FDB3,
-    0x000751B680007CCF, 0x80007CD780007CD3, 0x00076DBA00078DB5,
-    0x80007CDB000775BB, 0x000759BE80007CDF, 0x00076DBC000745BF,
-    0x000631A8000765BD, 0x000631AA000631A9, 0x000641AC000631AB,
-    0x00069DAE000695AD, 0x80007CE3000695AF, 0x80007CEB80007CE7,
-    0x80007CF380007CEF, 0x80007CFB80007CF7, 0x0006FDB880007CFF,
-    0x000681BA000681B9, 0x000681BC000681BB, 0x0006B1BE000689BD,
-    0x0006EDB00006B1BF, 0x0006FDB20006F5B1, 0x0006D9B40006F5B3,
-    0x0006CDB60006C5B5, 0x0006B9A30006C5B7, 0x80007D0780007D03,
-    0x80007D0F80007D0B, 0x0006C9A5000615A6, 0x000631AB80007D13,
-    0x80007D17000629AA, 0x000601AF80007D1B, 0x000621AD00061DAE,
-    0x00005D80000629AC, 0x0000658200006D81, 0x80007D1F80007BCF,
-    0x0001788780007D23, 0x80007D27000C0086, 0x80007D2F80007D2B,
-    0x80007D3780007D33, 0x80007D3F80007D3B, 0x80007D4780007D43,
-    0x000295B380007D4B, 0x0002BDB500039484, 0x80007D5380007D4F,
-    0x80007D570002DDB6, 0x0002F1BB80007D5B, 0x0002DDBD0002F9BA,
-    0x00031DBF0002DDBC, 0x000618E300031DBE, 0x000644E180007D5F,
-    0x80007D6780007D63, 0x80007D6B000190E1, 0x80007D6F000240E3,
-    0x80007D7780007D73, 0x0006C4EF000220EF, 0x80007D7F80007D7B,
-    0x000211A980007D83, 0x000221AB000201A8, 0x000251AD000211AA,
-    0x000255AF000239AC, 0x0000DD80000255AE, 0x0000198200000981,
-    0x80007D8B80007D87, 0x000CA8870003F8BE, 0x0003F5B9000D9C86,
-    0x0003F5BB0003F5B8, 0x000359BD0003F5BA, 0x000349BF000351BC,
-    0x000239B1000351BE, 0x000205B3000231B0, 0x0003D5B5000205B2,
-    0x0003D5B70003D5B4, 0x80007D930003D5B6, 0x80007D9B80007D97,
-    0x80007D9F000391A3, 0x0003D9A60003B9A5, 0x80007DA780007DA3,
-    0x0003FDAA80007DAB, 0x0003D9AC0003F5AB, 0x000219AE0003D9AD,
-    0x0002B1B6000219AF, 0x80007DB380007DAF, 0x80007DB70002A1B5,
-    0x80007DBB0002B1B3, 0x000179BE80007DBF, 0x00016DBC000145BF,
-    0x00028DBA000179BD, 0x80007DC3000175BB, 0x0006D8EF80007DC7,
-    0x80007DCB0023D0E1, 0x80007DCF001238E3, 0x80007DD780007DD3,
-    0x000F6C840007D4EF, 0x80007DDB0006DCE1, 0x0006FCE300071CE3,
-    0x000128E180007DDF, 0x000225A301E4A8EF, 0x80007DE3000000F0,
-    0x80007DEB80007DE7, 0x000235A5000225A6, 0x0001E1AB80007DEF,
-    0x80007DF3000219AA, 0x0001D1AF80007DF7, 0x0001EDAD0001EDAE,
-    0x0024218F0001F9AC, 0x80007DFB80007D8F, 0x000FF9800000C4BE,
-    0x00001182000FF981, 0x0000008680007DFF, 0x80007E0300038487,
-    0x80007E0B80007E07, 0x80007E1380007E0F, 0x0008D19F80007E17,
-    0x0009F19D0008659E, 0x0034719B0035799C, 0x003175990035E19A,
-    0x0030659700305598, 0x003C31950031C196, 0x0039D993003C6D94,
-    0x0038599100384592, 0x0014A5B10039F190, 0x01E809B30014A5B0,
-    0x01E805B50015E9B2, 0x80007E1B01E819B4, 0x80007E2380007E1F,
-    0x80007E2B80007E27, 0x000D61A380007E2F, 0x000C99A1000C89A2,
-    0x0000C5A1000D71A0, 0x0004C9A380007E33, 0x000439A50000D1A2,
-    0x0018F9A70004D1A4, 0x001C01A9001801A6, 0x001C6DAB001879A8,
-    0x0010D1AD001CCDAA, 0x001401AF001001AC, 0x002521820010D9AE,
-    0x80007E3700388983, 0x0038B98680007E3B, 0x00393984003D1587,
-    0x003D798A00392D85, 0x80007E3F0030818B, 0x00301D8E80007E43,
-    0x00313D8C0034818F, 0x003561920031358D, 0x80007E4700347193,
-    0x0009C19680007E4B, 0x0008E194000C9997, 0x000D419A00097D95,
-    0x80007E4F000D599B, 0x80007E5780007E53, 0x000CA99C80007E5B,
-    0x000E81A880007E5F, 0x000E81AA000E81A9, 0x000E8DAC000E9DAB,
-    0x000EBDAE000EB5AD, 0x00000D80000ED9AF, 0x00001D8200001581,
-    0x80007E6780007E63, 0x0000BC8480007E6B, 0x000E55B8000400BE,
-    0x000E61BA000E5DB9, 0x000E19BC000E79BB, 0x000E09BE000E19BD,
-    0x000EA9B0000E01BF, 0x000EB1B2000EB5B1, 0x000E89B4000EB1B3,
-    0x000E71B6000E89B5, 0x000D09B3000E6DB7, 0x0004888680007E73,
-    0x00045C840003C887, 0x000D05B5000D35B6, 0x000D71BB80007E77,
-    0x80007E7B000D6DBA, 0x000D4DBF80007E7F, 0x000D65BD000D4DBE,
-    0x80007E83000D65BC, 0x80007E87000D4DA3, 0x000D71A680007E8B,
-    0x80007E9380007E8F, 0x000D29AA000D41A5, 0x80007E97000D35AB,
-    0x000D09AE80007E9B, 0x000D21AC000D09AF, 0x0002A9A8000D21AD,
-    0x0002B9AA0002A9A9, 0x0002A9AC0002B9AB, 0x000201AE0002A9AD,
-    0x80007E9F000201AF, 0x80007EA780007EA3, 0x80007EAF80007EAB,
-    0x80007EB780007EB3, 0x0003EDB880007EBB, 0x000385BA00038DB9,
-    0x0003BDBC0003A5BB, 0x0003ADBE0003A5BD, 0x000241B00003A1BF,
-    0x000241B2000241B1, 0x000235B4000241B3, 0x0003F5B60003FDB5,
-    0x000D45B70003D5B7, 0x000D29B5000D21B6, 0x000D05B3000D15B4,
-    0x000D19B1000D11B2, 0x000D59BF000EE1B0, 0x000D41BD000D51BE,
-    0x000D51BB000D49BC, 0x000D69B9000D6DBA, 0x80007EBF000D71B8,
-    0x80007EC300050CBE, 0x80007ECB80007EC7, 0x80007ED380007ECF,
-    0x000E99AF80007ED7, 0x000E99AD000E91AE, 0x000E99AB000E81AC,
-    0x000E99A9000E91AA, 0x000F24E1000EA9A8, 0x000E38E380007EDB,
-    0x000038E380007EE3, 0x0001A0E180007EE7, 0x0000E4EF80007EEB,
-    0x0004E48480007EEF, 0x0004108680007EF3, 0x00014CEF00049487,
-    0x00009D8180007EF7, 0x0003F9B300009D80, 0x0003E9B500004982,
-    0x80007EFF80007EFB, 0x80007F030003E1B6, 0x00034DBB80007F07,
-    0x00035DBD00034DBA, 0x0002F1BF00035DBC, 0x80007E6F000339BE,
-    0x80007F0B80007EDF, 0x80007F1380007F0F, 0x80007F1B80007F17,
-    0x000625A880007F1F, 0x000601AA000639A9, 0x000671AC000601AB,
-    0x000675AE00067DAD, 0x000619B0000669AF, 0x000625B200062DB1,
-    0x0006F5B40006EDB3, 0x0006F5B60006FDB5, 0x0006D5B80006EDB7,
-    0x0006D5BA0006DDB9, 0x000739BC00072DBB, 0x000729BE000739BD,
-    0x0006A9B3000729BF, 0x80007F2780007F23, 0x80007F2F80007F2B,
-    0x0006D9B50006DDB6, 0x0006C9BB80007F33, 0x80007F370006FDBA,
-    0x000619BF80007F3B, 0x000619BD000611BE, 0x80007F3F0006D1BC,
-    0x80007F430006EDA3, 0x000699A680007F47, 0x000390BE80007F4B,
-    0x0006B9AA00069DA5, 0x80007F4F00068DAB, 0x000655AE80007F53,
-    0x000695AC00065DAF, 0x00068DA800065DAD, 0x000695AA000691A9,
-    0x0006BDAC0006A9AB, 0x0006CDAE0006A9AD, 0x0001E9800006C5AF,
-    0x0001FD820001E981, 0x0001908680007F57, 0x80007F5B0001B487,
-    0x0001C5B880007F5F, 0x0001C5BA0001CDB9, 0x0001CDBC0001DDBB,
-    0x0001F5BE0001FDBD, 0x000681B000019DBF, 0x000161B2000681B1,
-    0x000161B4000161B3, 0x000161B6000161B5, 0x00061DB3000161B7,
-    0x80007F6780007F63, 0x80007F6F80007F6B, 0x000635B5000629B6,
-    0x000645BB80007F73, 0x80007F77000645BA, 0x0001A5BF80007F7B,
-    0x0001B5BD0001A5BE, 0x80007F7F0001B5BC, 0x80007F83000659A3,
-    0x00066DA680007F87, 0x80007F8F80007F8B, 0x000601AA000671A5,
-    0x80007F93000601AB, 0x0001E1AE80007F97, 0x0001F1AC0001E1AF,
-    0x0002EDA80001F1AD, 0x000331AA000329A9, 0x000395AC000331AB,
-    0x000395AE00039DAD, 0x00020C8400038DAF, 0x80007F9F80007F9B,
-    0x80007FA780007FA3, 0x80007FAF80007FAB, 0x0003A9B880007FB3,
-    0x00006DBA0003A9B9, 0x00007DBC000065BB, 0x00006DBE000065BD,
-    0x0003FDB0000065BF, 0x0003C5B20003CDB1, 0x0003B9B40003ADB3,
-    0x0003ADB60003B9B5, 0x0000BD800003A5B7, 0x0000198200000981,
-    0x80007FBB80007FB7, 0x80007FC30003D8BE, 0x001C008680007FC7,
-    0x80007FCB0003D487, 0x80007FD380007FCF, 0x80007FDB80007FD7,
-    0x000369BB80007FDF, 0x80007FE3000359BA, 0x00031DBF80007FE7,
-    0x000319BD00031DBE, 0x0003D9B3000371BC, 0x80007FEF80007FEB,
-    0x80007FF780007FF3, 0x00037DB500037DB6, 0x000609AF80007FFB,
-    0x0007A5AD0007C1AE, 0x0007B5AB0007A5AC, 0x80007FFF0007B5AA,
-    0x001D6C8480008003, 0x000711A5000719A6, 0x000701A380008007,
-    0x8000800F8000800B, 0x8000801780008013, 0x001D08BE8000801B,
-    0x8000801F001C0084, 0x80008023000338EF, 0x8000802B80008027,
-    0x80008033000354E3, 0x00002980000044E1, 0x0000198200002981,
-    0x000628E180008037, 0x000630E3000794E1, 0x8000803B00078CE3,
-    0x001C38868000803F, 0x001CE8BE001D5C87, 0x8000804780008043,
-    0x8000804F8000804B, 0x000198EF80008053, 0x0002F9A3000600EF,
-    0x8000805B80008057, 0x800080638000805F, 0x00025DA500025DA6,
-    0x000249AB80008067, 0x8000806B000279AA, 0x00023DAF8000806F,
-    0x000239AD00023DAE, 0x001E99A8000251AC, 0x001EA1AA001E99A9,
-    0x001EA1AC001EA1AB, 0x001EA1AE001EA1AD, 0x8000802F001EA1AF,
-    0x8000807780008073, 0x8000807F8000807B, 0x8000808780008083,
-    0x001F5DB88000808B, 0x001F7DBA001F69B9, 0x001F6DBC001F75BB,
-    0x001F1DBE001F15BD, 0x001EA9B0001F0DBF, 0x001E8DB2001EA9B1,
-    0x001E9DB4001E85B3, 0x001E85B6001E85B5, 0x001E29B3001F65B7,
-    0x800080938000808F, 0x8000809B80008097, 0x001E29B5001E21B6,
-    0x001E6DBB8000809F, 0x00000886001E6DBA, 0x001E41BF00002C87,
-    0x001E51BD001E59BE, 0x00004980001E59BC, 0x0000598200004981,
-    0x800080A3001E65A3, 0x001E6DA6001E65A5, 0x800080A780007FBF,
-    0x001E21AA800080AB, 0x001E15AC001E21AB, 0x001E15AE001E1DAD,
-    0x001E69A8001E0DAF, 0x001E99AA001E79A9, 0x001EB5AC001E99AB,
-    0x001EB5AE001EBDAD, 0x800080AF001EADAF, 0x800080B7800080B3,
-    0x800080BF800080BB, 0x800080C7800080C3, 0x000165B8800080CB,
-    0x000165BA00016DB9, 0x000165BC00017DBB, 0x000165BE00016DBD,
-    0x001ED5B00001D9BF, 0x001EA1B2001EA1B1, 0x001EA1B4001EA1B3,
-    0x001E91B6001EA9B5, 0x001E11B3001E91B7, 0x800080D3800080CF,
-    0x800080DB800080D7, 0x001E3DB5001E3DB6, 0x001E65BB800080DF,
-    0x800080E3001E11BA, 0x001E6DBF800080E7, 0x001E75BD001E65BE,
-    0x800080EB001E75BC, 0x800080EF001E55A3, 0x001E79A6800080F3,
-    0x800080FB800080F7, 0x001E55AA001E79A5, 0x800080FF001E21AB,
-    0x001E21AE80008103, 0x001E31AC001E29AF, 0x0002D1A8001E31AD,
-    0x0002EDAA0002E5A9, 0x000335AC00032DAB, 0x000335AE00033DAD,
-    0x8000810700032DAF, 0x800081138000810B, 0x80008117000260BE,
-    0x8000811F8000811B, 0x0003E9B880008123, 0x000389BA0003E9B9,
-    0x00039DBC000385BB, 0x000381BE000381BD, 0x000355B00003B5BF,
-    0x000355B200035DB1, 0x0003F9B40003E9B3, 0x0003E9B60003F9B5,
-    0x00036D800003E1B7, 0x0000AD820000A581, 0x80008127000269B3,
-    0x0003B1B60003B1B5, 0x0002E0848000812B, 0x00039DBA8000812F,
-    0x00038DBC000395BB, 0x000331BE000331BD, 0x000225A3000331BF,
-    0x8000813780008133, 0x8000813F8000813B, 0x0003FDA50003FDA6,
-    0x0003D9AB80008143, 0x000448860003D1AA, 0x00037DAF0003AC87,
-    0x00037DAD00037DAE, 0x800081470003C1AC, 0x8000814F8000814B,
-    0x0003A0E380008153, 0x0001A8E180008157, 0x0003F4EF8000815B,
-    0x800081638000815F, 0x8000816B80008167, 0x800081738000816F,
-    0x0000798080008177, 0x00001D8200007981, 0x001E60E300024084,
-    0x001FE0E18000817B, 0x001ED0EF000550BE, 0x0004C88680008183,
-    0x8000818700054087, 0x8000818F8000818B, 0x8000819780008193,
-    0x8000819F8000819B, 0x800081A7800081A3, 0x800081AF800081AB,
-    0x0005A484001EC4EF, 0x800081B3001E24E1, 0x800081B7001EF0E3,
-    0x800081BF800081BB, 0x0001F5B3800081C3, 0x800081CB800081C7,
-    0x800081D3800081CF, 0x0001E5B5000145B6, 0x000129BB800081D7,
-    0x800081DB000121BA, 0x0000F1BF800081DF, 0x000129BD000129BE,
-    0x000629A8000131BC, 0x00063DAA000635A9, 0x00062DAC000635AB,
-    0x000685AE000685AD, 0x8000817F0006B9AF, 0x00001D8100001D82,
-    0x800081E300001D80, 0x800081EB800081E7, 0x00074DB8000330BE,
-    0x000755BA000751B9, 0x000779BC000769BB, 0x000769BE000779BD,
-    0x0006A5B0000769BF, 0x0006A5B20006ADB1, 0x0006ADB40006BDB3,
-    0x00077DB6000775B5, 0x0006B5A3000775B7, 0x00002886800081EF,
-    0x800081F300034C87, 0x0006A5A5000605A6, 0x000669AB800081F7,
-    0x800081FB000661AA, 0x0007B1AF800081FF, 0x000669AD000669AE,
-    0x80008203000671AC, 0x0007BDB380008207, 0x0007ADB58000820B,
-    0x800082138000820F, 0x8000810F00075DB6, 0x000779BB80008217,
-    0x000769BD000761BA, 0x000749BF000761BC, 0x000665A8000755BE,
-    0x00067DAA000675A9, 0x00066DAC000675AB, 0x00069DAE000695AD,
-    0x8000821B000695AF, 0x800082238000821F, 0x8000822B80008227,
-    0x800082338000822F, 0x0006FDB880008237, 0x000681BA000681B9,
-    0x000681BC000681BB, 0x0006B1BE000689BD, 0x0006EDB00006B1BF,
-    0x0006FDB20006F5B1, 0x0006D9B40006F5B3, 0x0006CDB60006C5B5,
-    0x0006F5A30006C5B7, 0x8000823F8000823B, 0x8000824780008243,
-    0x0006E5A5000615A6, 0x000631AB8000824B, 0x8000824F000629AA,
-    0x000601AF80008253, 0x000621AD00061DAE, 0x00006D80000629AC,
-    0x0000198200000981, 0x8000825B80008257, 0x000190BE00019884,
-    0x000F408680008263, 0x800082670000F487, 0x8000826F8000826B,
-    0x8000827780008273, 0x000291A88000827B, 0x0002A1AA000299A9,
-    0x0002D1AC0002A1AB, 0x0002D5AE0002D1AD, 0x8000827F0002C9AF,
-    0x8000828780008283, 0x8000828B00032084, 0x800082938000828F,
-    0x000379B880008297, 0x0003CDBA000379B9, 0x0003DDBC0003C5BB,
-    0x0003C5BE0003C5BD, 0x0002B9B00003F5BF, 0x000349B2000281B1,
-    0x000359B4000349B3, 0x000349B6000359B5, 0x000229B3000349B7,
-    0x8000829F8000829B, 0x000C9CBE800082A3, 0x000259B5000251B6,
-    0x000245BB800082A7, 0x800082AB00027DBA, 0x00039DBF800082AF,
-    0x00039DBD00039DBE, 0x800082B300039DBC, 0x800082B700026DA3,
-    0x000215A6800082BB, 0x800082C3800082BF, 0x000239AA00021DA5,
-    0x800082C7000201AB, 0x0003D9AE800082CB, 0x0003D9AC0003D9AF,
-    0x000190E10003D9AD, 0x0000CCE3800082D3, 0x0001C980000D1884,
-    0x0001F5820001FD81, 0x000C2CBE800082D7, 0x800082DF800082DB,
-    0x800082E7800082E3, 0x000018EF800082EB, 0x000CE086800082EF,
-    0x800082F30002F487, 0x800082FB800082F7, 0x80008303800082FF,
-    0x800083070006F8EF, 0x8000830B000794E1, 0x8000830F000630E3,
-    0x8000831780008313, 0x8000831F8000831B, 0x8000832780008323,
-    0x8000832F8000832B, 0x8000833780008333, 0x01F00085000158EF,
-    0x8000833B0006D4E1, 0x0000258000061CE3, 0x0000118200002D81,
-    0x00022DB380008343, 0x000CE887000000F1, 0x80008347000F6C86,
-    0x00022DB500022DB6, 0x0001C9BB8000834B, 0x8000834F0001C1BA,
-    0x0001CDBF80008353, 0x0001CDBD0001CDBE, 0x000DDDA80001CDBC,
-    0x000DE5AA000DEDA9, 0x000DE5AC000DFDAB, 0x000DE5AE000DEDAD,
-    0x800082CF000E3DAF, 0x800083578000833F, 0x8000835F8000835B,
-    0x8000836780008363, 0x000ECDB88000836B, 0x000EE1BA000ED9B9,
-    0x000E91BC000EE1BB, 0x000E95BE000E9DBD, 0x000E45B0000E89BF,
-    0x000E59B2000E4DB1, 0x000E49B4000E59B3, 0x000EFDB6000E49B5,
-    0x000D6DA3000EF5B7, 0x800083738000836F, 0x8000837B80008377,
-    0x000D6DA5000D6DA6, 0x000E89AB8000837F, 0x80008383000E81AA,
-    0x000E8DAF80008387, 0x000E8DAD000E8DAE, 0x000FEDB5000E8DAC,
-    0x8000838F8000838B, 0x80008393000FFDB6, 0x000FFDB380008397,
-    0x000FB5BD8000839B, 0x000F89BF000FB5BC, 0x8000839F000FB5BE,
-    0x000FB5BB8000825F, 0x000E05A8000FB5BA, 0x000E2DAA000E19A9,
-    0x000E3DAC000E25AB, 0x000E2DAE000E25AD, 0x00000D80000E21AF,
-    0x00001D8200001581, 0x800083A7800083A3, 0x00023084800083AB,
-    0x000E35B8000130BE, 0x000E35BA000E3DB9, 0x000ED5BC000ECDBB,
-    0x000ED5BE000EDDBD, 0x000E29B0000ECDBF, 0x000E39B2000E29B1,
-    0x000E29B4000E39B3, 0x000E11B6000E15B5, 0x000EA5A3000E0DB7,
-    0x00072886800083B3, 0x800083B700014C87, 0x000EB5A5000EA5A6,
-    0x000EEDAB800083BB, 0x800083BF000EEDAA, 0x000ED1AF800083C3,
-    0x000EEDAD000EEDAE, 0x800083C7000EEDAC, 0x800083CB000E19B3,
-    0x0001F9B6800083CF, 0x800083D7800083D3, 0x0001D5BA000E31B5,
-    0x800083DB0001DDBB, 0x000179BE800083DF, 0x0001C5BC000179BF,
-    0x000DBDA8000179BD, 0x000DD5AA000DD5A9, 0x000DFDAC000DE5AB,
-    0x000235AE00023DAD, 0x800083E300022DAF, 0x800083EB800083E7,
-    0x800083F3800083EF, 0x800083FB800083F7, 0x0002E9B8800083FF,
-    0x00028DBA0002E9B9, 0x000281BC000285BB, 0x000285BE000281BD,
-    0x000255B00002ADBF, 0x000255B200025DB1, 0x0002F9B40002E9B3,
-    0x0002E9B60002F9B5, 0x000D5DA30002E1B7, 0x8000840780008403,
-    0x8000840F8000840B, 0x000D75A50002BDA6, 0x000299AB80008413,
-    0x80008417000291AA, 0x00023DAF8000841B, 0x00023DAD00023DAE,
-    0x00005D80000281AC, 0x0000658200006D81, 0x80008423000351B3,
-    0x000319B6000379B5, 0x00036CBE0002DC84, 0x00033DBA80008427,
-    0x000319BC000335BB, 0x0000D9BE000319BD, 0x00029DA30000D9BF,
-    0x000568868000842B, 0x8000842F0002BC87, 0x0002B5A50002D5A6,
-    0x0002F9AB80008433, 0x000220840002F1AA, 0x000115AF80008437,
-    0x0002D5AD000115AE, 0x8000843B0002D5AC, 0x000F20EF8000843F,
-    0x8000844780008443, 0x8000844F8000844B, 0x8000845780008453,
-    0x8000845F8000845B, 0x000C88E180008463, 0x000F44E380008467,
-    0x8000846F8000846B, 0x8000847780008473, 0x00041CBE8000847B,
-    0x00005CEF8000847F, 0x000E74E1000E44EF, 0x000080E3000EFCE1,
-    0x00000D80000EB0E3, 0x0000718200007581, 0x8000848780008483,
-    0x8000848B0000C1B3, 0x0000FDB680008493, 0x8000849B80008497,
-    0x0000ADBA0000C1B5, 0x0004C8860000B5BB, 0x0000ADBE0004A087,
-    0x0000ADBC000095BF, 0x000585A80000A5BD, 0x000595AA000595A9,
-    0x0005BDAC0005A5AB, 0x0005D1AE0005D5AD, 0x8000841F0005D1AF,
-    0x8000849F8000848F, 0x800084A7800084A3, 0x800084AF800084AB,
-    0x000671B8800084B3, 0x000671BA000671B9, 0x0006D5BC000671BB,
-    0x0006D5BE0006DDBD, 0x0005B5B00006CDBF, 0x000581B20005BDB1,
-    0x000651B4000581B3, 0x000651B6000651B5, 0x0007EDB8000651B7,
-    0x0007F5BA0007F1B9, 0x000715BC00070DBB, 0x000715BE00071DBD,
-    0x0007EDB0000709BF, 0x0007FDB20007F5B1, 0x0007EDB40007F5B3,
-    0x0007DDB60007D5B5, 0x000611A80007D5B7, 0x0007CDAA000621A9,
-    0x0007CDAC0007DDAB, 0x000799AE0007F9AD, 0x800084B7000799AF,
-    0x800084BF800084BB, 0x800084C7800084C3, 0x800083AF800084CB,
-    0x000605A3800084CF, 0x800084D7800084D3, 0x800084DF800084DB,
-    0x000605A5000639A6, 0x000671AB800084E3, 0x800084E7000669AA,
-    0x000651AF800084EB, 0x000661AD000669AE, 0x800084EF000669AC,
-    0x00001D8100001D82, 0x800084F300001D80, 0x800084FB800084F7,
-    0x0001AC84000194BE, 0x0001E887800084FF, 0x8000850301FC0086,
-    0x8000850B80008507, 0x00064DA88000850F, 0x000651AA000651A9,
-    0x0001BDAC0001A5AB, 0x0001ADAE0001A5AD, 0x800085130001A5AF,
-    0x8000851B80008517, 0x800085238000851F, 0x8000852B80008527,
-    0x0000C5B88000852F, 0x0000C5BA0000CDB9, 0x0000CDBC0000DDBB,
-    0x0000F5BE0000FDBD, 0x0001DDB000009DBF, 0x00010DB2000105B1,
-    0x00011DB4000105B3, 0x00010DB6000105B5, 0x0005D9B3000101B7,
-    0x8000853780008533, 0x8000853F8000853B, 0x000235B5000225B6,
-    0x0002C1BB80008543, 0x80008547000219BA, 0x0002C1BF8000854B,
-    0x0002D1BD0002D9BE, 0x8000854F0002D9BC, 0x8000855300059DA3,
-    0x000261A680008557, 0x8000855F8000855B, 0x00025DAA000271A5,
-    0x80008563000285AB, 0x00029DAE80008567, 0x00029DAC000285AF,
-    0x0000E980000295AD, 0x000005820000E981, 0x0003DCBE8000856B,
-    0x8000857300035884, 0x01FC008680008577, 0x8000857B0003EC87,
-    0x800085838000857F, 0x8000858B80008587, 0x800085938000858F,
-    0x8000859B80008597, 0x800085A38000859F, 0x0002B0EF800085A7,
-    0x000194E1800085AB, 0x00026CE3800085AF, 0x800085B7800085B3,
-    0x800085BF800085BB, 0x000279B3800085C3, 0x800085CB800085C7,
-    0x800085D3800085CF, 0x0002B9B50002B1B6, 0x000295BB800085D7,
-    0x800085DB000295BA, 0x000139BF01FD6484, 0x000285BD000131BE,
-    0x00006982000285BC, 0x0000598000075CEF, 0x800085E300005181,
-    0x01FD6887800085E7, 0x01FD2CBE01FCFC86, 0x800085EF800085EB,
-    0x01FC3C84800085F3, 0x800085F70007B8E3, 0x800085FB0004ECE1,
-    0x80008603800085FF, 0x8000860701FFC084, 0x00003CEF8000860B,
-    0x000650E10006F4EF, 0x0000D8E30006B4E1, 0x8000860F000694E3,
-    0x8000861780008613, 0x0002B9A38000861B, 0x800086238000861F,
-    0x8000862B80008627, 0x000279A5000271A6, 0x000255AB8000862F,
-    0x80008633000255AA, 0x0001F9AF80008637, 0x000245AD0001F1AE,
-    0x01FE99A8000245AC, 0x01FEA1AA01FE99A9, 0x01FEA1AC01FEA1AB,
-    0x01FEA1AE01FEA1AD, 0x8000863B01FEA1AF, 0x800086438000863F,
-    0x0000198080008647, 0x0000058200001981, 0x01FEADB88000864B,
-    0x01FEBDBA01FEB5B9, 0x01FEADBC01FEB5BB, 0x01FF51BE01FF51BD,
-    0x01FEA1B001FF51BF, 0x01FEA1B201FEA1B1, 0x01FEA1B401FEA1B3,
-    0x01FEA5B601FEA1B5, 0x8000856F01FE99B7, 0x0003C0BE01FE2DB3,
-    0x01FE25B6800085DF, 0x800086538000864F, 0x01FE61BA01FE2DB5,
-    0x0000088601FE61BB, 0x01FE61BE00009C87, 0x01FE71BC01FE61BF,
-    0x01FE69A301FE71BD, 0x8000865B80008657, 0x800086638000865F,
-    0x01FE69A501FE61A6, 0x01FE25AB80008667, 0x8000866B01FE25AA,
-    0x01FE25AF8000866F, 0x01FE35AD01FE25AE, 0x01FE51A801FE35AC,
-    0x01FE6DAA01FE51A9, 0x01FE7DAC01FE65AB, 0x01FE99AE01FE6DAD,
-    0x8000867301FE99AF, 0x8000867B80008677, 0x800086838000867F,
-    0x8000868B80008687, 0x01FE89B88000868F, 0x01FE99BA01FE89B9,
-    0x01FEB9BC01FE91BB, 0x000179BE01FEB9BD, 0x01FEE9B0000179BF,
-    0x01FEFDB201FEF5B1, 0x01FEC5B401FEDDB3, 0x01FEC5B601FECDB5,
-    0x01FEE5B301FEB9B7, 0x8000869780008693, 0x8000869F8000869B,
-    0x01FEF5B501FEEDB6, 0x01FE31BB800086A3, 0x800086A701FE2DBA,
-    0x01FE09BF800086AB, 0x01FE09BD01FE01BE, 0x800086AF01FE29BC,
-    0x800086B301FEA1A3, 0x01FEA9A6800086B7, 0x000190BE800086BB,
-    0x01FE69AA01FEB1A5, 0x0007E0BF01FE75AB, 0x01FE45AE0007E885,
-    0x01FE6DAC01FE4DAF, 0x000299A801FE4DAD, 0x0002C1AA0002C1A9,
-    0x0002C1AC0002C1AB, 0x0002F1AE0002C9AD, 0x000584860002F1AF,
-    0x800086C300034487, 0x0003F980800086C7, 0x0003ED820003F981,
-    0x0003C5B8800086CB, 0x0003C5BA0003CDB9, 0x0003CDBC0003DDBB,
-    0x0003F5BE0003FDBD, 0x000341B000039DBF, 0x000341B2000341B1,
-    0x000341B4000341B3, 0x000341B6000341B5, 0x000249B3000341B7,
-    0x800086D3800086CF, 0x800086DB800086D7, 0x000249B5000249B6,
-    0x000385BB800086DF, 0x800086E3000385BA, 0x000385BF800086E7,
-    0x000395BD000385BE, 0x800086EB000395BC, 0x800086EF00020DA3,
-    0x00020DA6800086F3, 0x800086FB800086F7, 0x0003C1AA00020DA5,
-    0x800086FF0003C1AB, 0x0003C1AE80008703, 0x0003D1AC0003C1AF,
-    0x01FE60E30003D1AD, 0x01FE4CE101FFBCE1, 0x0001A0E101FE30E3,
-    0x01FE84E30000ECE1, 0x0007ACBE001CC8E3, 0x000008EF80008707,
-    0x000140EF001BFCEF, 0x8000870F8000870B, 0x8000871301FEA4EF,
-    0x0004048480008717, 0x8000871B0007C484, 0x800087238000871F,
-    0x0000398080008727, 0x0000058200003981, 0x000750868000872B,
-    0x8000873300058487, 0x01FDC1BB80008737, 0x01FDA1B901FDA9BA,
-    0x01FDC9BF01FD91B8, 0x01FDD5BD01FDC1BE, 0x01FE1DB301FDD5BC,
-    0x01FE1DB101FE1DB2, 0x01FDA1B701FE1DB0, 0x01FDB1B501FDB9B6,
-    0x01FEBDAB01FDB9B4, 0x01FEBDA901FEBDAA, 0x01FE7DAF01FEBDA8,
-    0x01FE7DAD01FE7DAE, 0x8000873B01FE7DAC, 0x800087438000873F,
-    0x8000874B80008747, 0x000674BE8000874F, 0x01EC058280008753,
-    0x800086BF01EC1983, 0x01E8F9868000872F, 0x01E8198401F4C987,
-    0x01F4018A01E81985, 0x8000875701F4B98B, 0x01F0B58E8000875B,
-    0x01F0D98C01FCC18F, 0x01FC619201F01D8D, 0x8000875F01FC6193,
-    0x01F8759680008763, 0x01F8A19401F93197, 0x0004E19A01F87D95,
-    0x800087670004E99B, 0x8000876F8000876B, 0x0004ED9C80008773,
-    0x000089A180008777, 0x000C01A38000877B, 0x000C15A5000081A2,
-    0x0008B9A7000CB5A4, 0x0009C9A9000801A6, 0x0014ADAB000881A8,
-    0x001001AD0014ADAA, 0x001055AF001595AC, 0x001CE1B100109DAE,
-    0x001801B3001C01B0, 0x0018F5B5001CF5B2, 0x8000877F0018F5B4,
-    0x8000878780008783, 0x8000878F8000878B, 0x00001D8080008793,
-    0x0001C98200000981, 0x00072DA980008797, 0x00078DAB00072DA8,
-    0x00079DAD000731AA, 0x000785AF00079DAC, 0x01F845A000078DAE,
-    0x0005A5A20005A5A1, 0x8000879B0005B9A3, 0x000F9886800087A3,
-    0x000665B900003087, 0x000671BB0007A9B8, 0x000615BD000679BA,
-    0x00060DBF000615BC, 0x0007CDB100060DBE, 0x0007A1B30007CDB0,
-    0x0007A5B50007ADB2, 0x000791B70007B9B4, 0x0007BDB30007A5B6,
-    0x800087AB800087A7, 0x800087AF00000C84, 0x000711B5000719B6,
-    0x00070DBB01ED998F, 0x800087B300070DBA, 0x000765BF800087B7,
-    0x000771BD000779BE, 0x01F1E19800077DBC, 0x01FD3D9A01F14199,
-    0x01FDF59C01FDED9B, 0x01F9F99E01F92D9D, 0x01EDFD9001F9F99F,
-    0x01E9ED9201E9E591, 0x01F52D9401E90593, 0x01F5659601F5D595,
-    0x00043DA301F11D97, 0x000344BE000DC884, 0x800087BF800087BB,
-    0x000491A5000499A6, 0x00048DAB800087C3, 0x800087C700048DAA,
-    0x0004E5AF800087CB, 0x0004F1AD0004F9AE, 0x800087CF0004FDAC,
-    0x800087D30005A5B3, 0x0005BDB6800087D7, 0x800087DF800087DB,
-    0x00059DBA0005B5B5, 0x800087E300024DBB, 0x00023DBE800087E7,
-    0x000255BC00021DBF, 0x0005F9A8000235BD, 0x000569AA0005F9A9,
-    0x000579AC000569AB, 0x000569AE000579AD, 0x800087EB000569AF,
-    0x800087F3800087EF, 0x800087FB800087F7, 0x80008803800087FF,
-    0x0002B5B880008807, 0x00024DBA0002B9B9, 0x00020DBC000215BB,
-    0x00023DBE000235BD, 0x00051DB0000235BF, 0x0002C5B20002CDB1,
-    0x0002BDB40002A5B3, 0x000295B600029DB5, 0x0004E1A300028DB7,
-    0x0001B58100001582, 0x8000880B0001B580, 0x0004F1A50004F9A6,
-    0x000309AB80008813, 0x800088170004D9AA, 0x000359AF8000881B,
-    0x000371AD000379AE, 0x000D4086000311AC, 0x8000881F000D7487,
-    0x800088230002C5B3, 0x0002F5B60002FDB5, 0x8000882B80008827,
-    0x000175BA8000882F, 0x00015DBC000145BB, 0x000131BE000135BD,
-    0x0006DCE3000131BF, 0x0006E0E10006DCE3, 0x800088330007B4E1,
-    0x8000883B80008837, 0x000C8086000C54BE, 0x8000883F000C1887,
-    0x0006B0EF80008843, 0x8000884700066CEF, 0x000249A38000884B,
-    0x800088538000884F, 0x8000885B80008857, 0x000271A5000279A6,
-    0x0001C9AB8000885F, 0x800088630001F9AA, 0x0001BDAF80008867,
-    0x0001B9AD0001BDAE, 0x000D35A80001D1AC, 0x000D81AA000D81A9,
-    0x000DB9AC000D9DAB, 0x000E15AE000DB9AD, 0x8000880F000E0DAF,
-    0x8000886F8000886B, 0x00001D8080008873, 0x0000B98200000981,
-    0x000E29B880008877, 0x000EE9BA000E29B9, 0x000EF9BC000EE9BB,
-    0x000EE9BE000EF9BD, 0x000E7DB0000EE9BF, 0x000E45B2000E4DB1,
-    0x000E39B4000E2DB3, 0x000E2DB6000E39B5, 0x8000887B000E25B7,
-    0x8000879F8000887F, 0x8000888780008883, 0x8000888F8000888B,
-    0x000FBDA880008893, 0x000F85AA000F8DA9, 0x000FE1AC000FE5AB,
-    0x000FD9AE000FE9AD, 0x000F51B0000FD9AF, 0x000F51B2000F51B1,
-    0x000F7DB4000F65B3, 0x000F65B6000F6DB5, 0x000F29B8000F19B7,
-    0x000F39BA000F29B9, 0x000F19BC000F31BB, 0x0000DDBE000F19BD,
-    0x000E2DB30000D5BF, 0x8000889B80008897, 0x800088A38000889F,
-    0x000E49B5000E49B6, 0x000EB5BB800088A7, 0x800088AB000EB5BA,
-    0x000E89BF800088AF, 0x000E91BD000E99BE, 0x800088B3000EA5BC,
-    0x800088B7000E69A3, 0x000E0DA6800088BB, 0x800088C3800088BF,
-    0x000EF1AA000E0DA5, 0x800088C7000EF1AB, 0x000EDDAE800088CB,
-    0x000EE1AC000ECDAF, 0x00010DA8000ED5AD, 0x00011DAA000115A9,
-    0x00010DAC000115AB, 0x000091AE000091AD, 0x800088CF000091AF,
-    0x800088D7800088D3, 0x800088DF800088DB, 0x0003B1810003B982,
-    0x000099B80003B980, 0x0000A5BA0000ADB9, 0x000075BC00006DBB,
-    0x000075BE00007DBD, 0x0000F1B000006DBF, 0x0000C1B20000F9B1,
-    0x0000B1B40000C1B3, 0x0000B5B60000BDB5, 0x0002AC840000A9B7,
-    0x800088E7800088E3, 0x00021CBE800088EB, 0x0004B886800088F3,
-    0x800088F70002CC87, 0x800088FF800088FB, 0x000374BE80008903,
-    0x8000890B80008907, 0x00035DBB8000890F, 0x80008913000355BA,
-    0x000319BF80008917, 0x000375BD000315BE, 0x0002F5B3000375BC,
-    0x8000891F8000891B, 0x8000892780008923, 0x0002E5B5000371B6,
-    0x000215A38000892B, 0x800089338000892F, 0x8000893B80008937,
-    0x000205A5000391A6, 0x0003BDAB8000893F, 0x800089430003B5AA,
-    0x0003F9AF80008947, 0x000395AD0003F5AE, 0x8000894B000395AC,
-    0x800089538000894F, 0x0000198080008957, 0x0000058200001981,
-    0x00036C848000895B, 0x0003788780008963, 0x800089670004EC86,
-    0x8000896F8000896B, 0x000EB8E180008973, 0x000318E3000C98E3,
-    0x80008977000C2CE1, 0x8000897F8000897B, 0x0004EC8480008983,
-    0x8000898B80008987, 0x800089938000898F, 0x0003D0EF000FC4EF,
-    0x000165B380008997, 0x8000899F8000899B, 0x800089A7800089A3,
-    0x000175B5000161B6, 0x00012DBB800089AB, 0x800089AF00012DBA,
-    0x000101BF800089B3, 0x000111BD000119BE, 0x800089B7000119BC,
-    0x800089BF800089BB, 0x00001D80800089C3, 0x0000318200003581,
-    0x800089C7000EF8EF, 0x800089CB000EFCE1, 0x800089D3000FB8E3,
-    0x800089DB800089D7, 0x000408BE000000EE, 0x0001EDA3800089DF,
-    0x0001FDA500070084, 0x800089E7800089E3, 0x0004C0860001E9A6,
-    0x0001A5AB0004AC87, 0x000199AD0001A5AA, 0x000189AF000191AC,
-    0x000615A8000191AE, 0x00062DAA000625A9, 0x00063DAC000625AB,
-    0x000621AE000621AD, 0x8000895F000621AF, 0x800089EB800089CF,
-    0x800089F3800089EF, 0x800089FB800089F7, 0x0006C9B8800089FF,
-    0x0006D9BA0006C9B9, 0x0006F9BC0006D1BB, 0x000699BE0006F9BD,
-    0x000629B0000699BF, 0x00060DB2000629B1, 0x00061DB4000605B3,
-    0x000605B6000605B5, 0x80008A030006F9B7, 0x80008A0B80008A07,
-    0x80008A1380008A0F, 0x80008A1B80008A17, 0x000645A8800088EF,
-    0x00065DAA000655A9, 0x00064DAC000655AB, 0x0007ADAE0007A5AD,
-    0x0007DDB00007A5AF, 0x0007E1B20007E1B1, 0x0007E1B40007E1B3,
-    0x00079DB60007E9B5, 0x0007ADB8000795B7, 0x0007B5BA0007B5B9,
-    0x00079DBC000785BB, 0x00077DBE000775BD, 0x0007F5B3000775BF,
-    0x80008A2380008A1F, 0x80008A2B80008A27, 0x0007E5B5000605B6,
-    0x000605BB80008A2F, 0x80008A33000639BA, 0x00060DBF80008A37,
-    0x000615BD000605BE, 0x00001582000615BC, 0x000061800007B1A3,
-    0x000641A600006181, 0x000190BE80008A3B, 0x00067DAA0007A1A5,
-    0x80008A3F000641AB, 0x000641AE80008A43, 0x000651AC000649AF,
-    0x000631A8000651AD, 0x000631AA000631A9, 0x000645AC000649AB,
-    0x0001B1AE00064DAD, 0x001C00860001B1AF, 0x80008A4700010487,
-    0x80008A4F80008A4B, 0x80008A5780008A53, 0x000151B880008A5B,
-    0x000151BA000151B9, 0x0000D9BC0000C9BB, 0x0000C9BE0000D9BD,
-    0x0001D1B00000C1BF, 0x0001D1B20001D1B1, 0x000171B40001D1B3,
-    0x000171B6000171B5, 0x0005B9B3000171B7, 0x80008A6380008A5F,
-    0x80008A6B80008A67, 0x0005CDB50005CDB6, 0x000215BB80008A6F,
-    0x80008A73000215BA, 0x0002B5BF80008A77, 0x000205BD0002B5BE,
-    0x80008A7B000205BC, 0x80008A7F0005FDA3, 0x000589A680008A83,
-    0x80008A8B80008A87, 0x000251AA000589A5, 0x80008A8F000251AB,
-    0x0002F1AE80008A93, 0x000241AC0002F1AF, 0x80008A97000241AD,
-    0x80008A9F80008A9B, 0x80008AA780008AA3, 0x0003D4BE80008AAB,
-    0x00002D8080008AB3, 0x00003D8200003581, 0x001C308600032084,
-    0x80008AB700038487, 0x80008ABF80008ABB, 0x80008AC780008AC3,
-    0x80008ACF80008ACB, 0x80008AD30002B4EF, 0x80008ADB80008AD7,
-    0x80008ADF000258E3, 0x80008AE3000154E1, 0x80008AEB80008AE7,
-    0x80008AF380008AEF, 0x80008AFB80008AF7, 0x0007E0E380008AFF,
-    0x000718E180008B03, 0x0007E8EF80008B07, 0x80008B0F80008B0B,
-    0x00001D8080008B13, 0x00007D8200000981, 0x0002EDB380008B17,
-    0x001D048480008B1B, 0x80008B2780008B23, 0x0002EDB50002EDB6,
-    0x00016DBB001D14BE, 0x001CC88600016DBA, 0x00015DBF001DA487,
-    0x000165BD00015DBE, 0x80008B2B00016DBC, 0x80008B3380008B2F,
-    0x80008B3B80008B37, 0x000038EF80008B3F, 0x000658E10006A4EF,
-    0x000674E3000750E1, 0x80008B4300069CE3, 0x001CF0BE80008B47,
-    0x00026DA380008B4B, 0x001C848480008B4F, 0x80008B53001FE084,
-    0x00026DA500026DA6, 0x0001EDAB80008B57, 0x80008B5B0001EDAA,
-    0x0001DDAF80008B5F, 0x0001E5AD0001DDAE, 0x80008AAF0001EDAC,
-    0x80008B6380008B1F, 0x80008B6B80008B67, 0x80008B7380008B6F,
-    0x001EBDA880008B77, 0x001E95AA001E9DA9, 0x001EF5AC001EEDAB,
-    0x001EF5AE001EFDAD, 0x001E99B0001EEDAF, 0x001EA9B2001E99B1,
-    0x001EA1B4001EA5B3, 0x001EA1B6001EA1B5, 0x001F61B8001EA1B7,
-    0x001F61BA001F61B9, 0x001F65BC001F61BB, 0x001F5DBE001F69BD,
-    0x001E3DB3001F55BF, 0x80008B7F80008B7B, 0x80008B8780008B83,
-    0x001E5DB5001E5DB6, 0x001E7DBB80008B8B, 0x00000886001E7DBA,
-    0x001E45BF00008C87, 0x001E51BD001E45BE, 0x00005D82001E5DBC,
-    0x00004580001E79A3, 0x001E19A600005D81, 0x80008B9780008B93,
-    0x001E39AA001E19A5, 0x80008B9B001E39AB, 0x001E01AE80008B9F,
-    0x001E19AC001E01AF, 0x001E4DA8001E15AD, 0x001E89AA001E55A9,
-    0x001E99AC001E89AB, 0x001E89AE001E99AD, 0x80008B8F001E89AF,
-    0x80008BA780008BA3, 0x80008BAF80008BAB, 0x80008BB780008BB3,
-    0x000149B880008BBB, 0x000159BA000149B9, 0x000149BC000159BB,
-    0x0001DDBE000149BD, 0x001ED1B00001D5BF, 0x001ED5B2001ED1B1,
-    0x001EA1B4001EA9B3, 0x001E91B6001EA9B5, 0x001E01B3001E91B7,
-    0x80008BC380008BBF, 0x80008BCB80008BC7, 0x001E11B5001E19B6,
-    0x001E15BB80008BCF, 0x80008BD3001E15BA, 0x001E6DBF80008BD7,
-    0x001E7DBD001E65BE, 0x80008BDB001E01BC, 0x80008BDF001E45A3,
-    0x001E5DA680008BE3, 0x80008BEB80008BE7, 0x001E51AA001E55A5,
-    0x80008BEF001E51AB, 0x001E21AE80008BF3, 0x001E45AC001E29AF,
-    0x0002D5A8001E39AD, 0x0002E1AA0002DDA9, 0x000335AC0002E1AB,
-    0x000335AE00033DAD, 0x0003CD8000032DAF, 0x0000198200000981,
-    0x80008BFB80008BF7, 0x0003AC87000378BE, 0x0003E9B800049C86,
-    0x000389BA0003E9B9, 0x00039DBC000385BB, 0x000381BE000381BD,
-    0x000355B00003B5BF, 0x000355B200035DB1, 0x0003F9B40003E9B3,
-    0x0003E9B60003F9B5, 0x0004CCBE0003E1B7, 0x80008C0780008C03,
-    0x80008C0F80008C0B, 0x80008C1780008C13, 0x80008C1B000250EF,
-    0x80008C1F000014E1, 0x0005F8BE001F70E3, 0x80008C2780008C23,
-    0x0003F9BB80008C2B, 0x000784BE0003F1BA, 0x00035DBF80008C2F,
-    0x00035DBD00035DBE, 0x000209B30003E1BC, 0x80008C3780008C33,
-    0x80008C3F80008C3B, 0x0003DDB50003DDB6, 0x001E6DA880008C43,
-    0x001EA5AA001EADA9, 0x001EA5AC001EBDAB, 0x001EA5AE001EADAD,
-    0x80008C47001D19AF, 0x80008C4F80008C4B, 0x80008C5780008C53,
-    0x80008C5F80008C5B, 0x001D0DB880008C63, 0x001D11BA001D11B9,
-    0x001D3DBC001D25BB, 0x001DDDBE001DD5BD, 0x001D69B0001DD5BF,
-    0x001D79B2001D69B1, 0x001D59B4001D71B3, 0x001D39B6001D59B5,
-    0x001EF0E1001D35B7, 0x001E50E3001F2CE1, 0x00000981001ED4E3,
-    0x80008C670000A980, 0x0005F48400003982, 0x0004888780008C6B,
-    0x80008C7300040C86, 0x001EF8EF80008C77, 0x000329A3001E18EF,
-    0x80008C7F80008C7B, 0x80008C8780008C83, 0x0002FDA50002FDA6,
-    0x0002D9AB80008C8B, 0x80008C8F0002D1AA, 0x00027DAF80008C93,
-    0x00027DAD00027DAE, 0x000685A80002C1AC, 0x000685AA00068DA9,
-    0x000685AC00069DAB, 0x0006B9AE00068DAD, 0x80008C6F0006B5AF,
-    0x80008C9B80008C97, 0x80008CA380008C9F, 0x80008CAB80008CA7,
-    0x0006BDB880008CAF, 0x00076DBA000765B9, 0x00077DBC000765BB,
-    0x00076DBE000765BD, 0x0006D5B0000765BF, 0x0006D5B20006DDB1,
-    0x000695B40006A1B3, 0x000695B600069DB5, 0x000615B500068DB7,
-    0x80008CB780008CB3, 0x80008CBB000615B6, 0x000615B380008CBF,
-    0x00066DBD80008CC3, 0x000641BF000619BC, 0x80008BFF000649BE,
-    0x000601BB80008CC7, 0x00064DA3000609BA, 0x80008CCF80008CCB,
-    0x80008CD780008CD3, 0x00064DA500064DA6, 0x000659AB80008CDB,
-    0x0003A484000651AA, 0x000619AF80008CDF, 0x000635AD000611AE,
-    0x0006B1A8000641AC, 0x0006CDAA0006B1A9, 0x0006DDAC0006C5AB,
-    0x0006C5AE0006C5AD, 0x0001ED800006F5AF, 0x0001F5820001F581,
-    0x0001908680008CE3, 0x80008CE70001B487, 0x0001C5B880008CEB,
-    0x0001D9BA0001CDB9, 0x0001CDBC0001D9BB, 0x00019DBE0001E1BD,
-    0x00068DB0000195BF, 0x000141B2000141B1, 0x000141B4000141B3,
-    0x000141B6000141B5, 0x000645B3000141B7, 0x80008CF380008CEF,
-    0x80008CFB80008CF7, 0x000655B5000645B6, 0x000185BB80008CFF,
-    0x80008D03000649BA, 0x000185BF80008D07, 0x000195BD000185BE,
-    0x80008D0B000195BC, 0x80008D0F000601A3, 0x000601A680008D13,
-    0x80008D1B80008D17, 0x00060DAA000611A5, 0x80008D1F0001C1AB,
-    0x0001C1AE80008D23, 0x0001D1AC0001C1AF, 0x00032DA80001D1AD,
-    0x000335AA00033DA9, 0x000399AC000389AB, 0x000389AE000399AD,
-    0x80008D27000381AF, 0x80008D2F80008D2B, 0x80008D3780008D33,
-    0x80008D3F80008D3B, 0x0003ADB880008D43, 0x00006DBA000065B9,
-    0x00007DBC000065BB, 0x00006DBE000065BD, 0x0003C9B0000065BF,
-    0x0003A9B20003C9B1, 0x0003BDB40003A5B3, 0x0003A1B60003A1B5,
-    0x0000BD80000395B7, 0x0000198200000981, 0x80008D4B80008D47,
-    0x80008D5780008D53, 0x000C008600029084, 0x80008D5B0003E487,
-    0x80008D6380008D5F, 0x80008D6B80008D67, 0x0003FDB380008D6F,
-    0x80008D7780008D73, 0x80008D7F80008D7B, 0x000351B5000359B6,
-    0x00034DBB80008D83, 0x80008D8700034DBA, 0x000325BF80008D8B,
-    0x000331BD000339BE, 0x80008D8F00033DBC, 0x80008D9780008D93,
-    0x80008D9B000774EF, 0x80008DA380008D9F, 0x000D6C8480008DA7,
-    0x000D6CBE80008DAB, 0x80008DB380008DAF, 0x80008DB7000710E1,
-    0x0000718000068CE3, 0x00001D8200007D81, 0x000760EF000C0084,
-    0x000D748780008DBB, 0x0003ECE3000CFC86, 0x000040E180008DC3,
-    0x80008DCB80008DC7, 0x80008DD380008DCF, 0x80008DDB80008DD7,
-    0x80008DE380008DDF, 0x80008DEB80008DE7, 0x80008DF380008DEF,
-    0x000CE8BE0007D4EF, 0x80008DF7000610E1, 0x80008DFB000154E3,
-    0x80008E0380008DFF, 0x0002FDA380008E07, 0x80008E0F80008E0B,
-    0x80008E1780008E13, 0x000251A5000259A6, 0x00024DAB80008E1B,
-    0x80008E1F00024DAA, 0x000225AF80008E23, 0x000231AD000239AE,
-    0x000E95A800023DAC, 0x000EA1AA000E9DA9, 0x000EA1AC000EA1AB,
-    0x000EA1AE000EA1AD, 0x80008DBF000EA1AF, 0x0000198100001182,
-    0x80008E27000FC180, 0x80008E2F80008E2B, 0x000F45B880008E33,
-    0x000F59BA000F4DB9, 0x000F49BC000F59BB, 0x000F7DBE000F49BD,
-    0x000EA5B0000F71BF, 0x000EA5B2000EADB1, 0x000EADB4000EBDB3,
-    0x000E95B6000E9DB5, 0x000E29B3000F7DB7, 0x0000288680008E37,
-    0x80008E3B00034487, 0x000E29B5000E21B6, 0x000E69BB80008E3F,
-    0x80008E43000E61BA, 0x000E59BF80008E47, 0x000E69BD000E69BE,
-    0x80008E4B000E71BC, 0x000E61A380008D4F, 0x000E61A580008E4F,
-    0x80008E5780008E53, 0x80008E5B000E69A6, 0x000E21AB80008E5F,
-    0x000E21AD000E29AA, 0x000E11AF000E39AC, 0x000E65A8000E21AE,
-    0x000E79AA000E69A9, 0x000E69AC000E79AB, 0x000E99AE000E69AD,
-    0x80008E63000E99AF, 0x80008E6B80008E67, 0x80008E7380008E6F,
-    0x80008E7B80008E77, 0x000EF9B880008E7F, 0x000E8DBA000EF9B9,
-    0x000E9DBC000E85BB, 0x000E85BE000E85BD, 0x000EE9B0000EB5BF,
-    0x000EF9B2000EE9B1, 0x000EE9B4000EF9B3, 0x000EC9B6000EDDB5,
-    0x000EF1B3000EC9B7, 0x80008E8780008E83, 0x80008E8F80008E8B,
-    0x000EE1B5000EE9B6, 0x000E2DBB80008E93, 0x80008E97000E2DBA,
-    0x000E19BF80008E9B, 0x000E2DBD000E2DBE, 0x00006D82000E2DBC,
-    0x00005580000EB5A3, 0x000EADA600006581, 0x000190BE00019884,
-    0x000E69AA000EA5A5, 0x00074886000E69AB, 0x000E69AE0000FC87,
-    0x000E69AC000E5DAF, 0x00023DA8000E69AD, 0x000299AA000281A9,
-    0x0002B9AC000291AB, 0x0002D9AE0002B9AD, 0x80008EA30002D1AF,
-    0x80008EAB80008EA7, 0x80008EAF003800BF, 0x80008EB780008EB3,
-    0x000371B880008EBB, 0x000371BA000371B9, 0x0003D5BC000371BB,
-    0x0003D5BE0003DDBD, 0x0002B5B00003CDBF, 0x000289B20002BDB1,
-    0x000351B4000289B3, 0x000351B6000351B5, 0x000211B3000351B7,
-    0x80008EC380008EBF, 0x00039C8480008EC7, 0x000231B5000239B6,
-    0x000279BB80008ECB, 0x80008ECF000271BA, 0x000395BF80008ED3,
-    0x00025DBD000395BE, 0x80008ED700025DBC, 0x80008EDB000255A3,
-    0x00027DA680008EDF, 0x80008EE780008EE3, 0x000235AA000275A5,
-    0x80008EEB00023DAB, 0x0003D1AE80008EEF, 0x000219AC0003D1AF,
-    0x80008EF3000219AD, 0x80008EFB80008EF7, 0x00000D8080008EFF,
-    0x0000318200003181, 0x00052CBE80008F03, 0x80008F0B80008F07,
-    0x80008F1780008F13, 0x80008F1F80008F1B, 0x000FF8E180008F23,
-    0x000E78E380008F27, 0x0004F08600054484, 0x80008F2B00029887,
-    0x80008F3380008F2F, 0x80008F3B80008F37, 0x80008F4380008F3F,
-    0x000EA8EF80008F47, 0x80008F4F80008F4B, 0x80008F530000A4EF,
-    0x80008F5B80008F57, 0x80008F6380008F5F, 0x0004ACBE000EC4EF,
-    0x80008F67000EE0E1, 0x80008F6B000EF0E3, 0x80008F6F000070E3,
-    0x0001E9B300012CE1, 0x0000158100002D82, 0x80008F7300001D80,
-    0x0001EDB50001EDB6, 0x00016DBB80008F77, 0x80008F7B00016DBA,
-    0x000131BF80008F83, 0x000151BD000159BE, 0x000621A8000161BC,
-    0x000621AA000621A9, 0x000625AC00063DAB, 0x000625AE00062DAD,
-    0x80008F0F000699AF, 0x0000888680008F7F, 0x80008F870003A087,
-    0x80008F8F80008F8B, 0x000695B880008F93, 0x0006A1BA00069DB9,
-    0x000751BC0006A1BB, 0x000751BE000751BD, 0x0006E9B0000751BF,
-    0x0006F5B20006E9B1, 0x0006DDB40006C5B3, 0x0006B1B60006B5B5,
-    0x0006A9A30006B1B7, 0x80008F9B80008F97, 0x80008FA380008F9F,
-    0x0006ADA50006ADA6, 0x00062DAB80008FA7, 0x80008FAB00062DAA,
-    0x000671AF80008FAF, 0x000611AD000619AE, 0x80008FB3000621AC,
-    0x80008E9F0007F1B3, 0x000795B680008FB7, 0x80008FBF80008FBB,
-    0x000781BA0007E1B5, 0x80008FC3000781BB, 0x000781BE80008FC7,
-    0x000791BC000781BF, 0x00062DA8000791BD, 0x00063DAA000635A9,
-    0x00062DAC000635AB, 0x000641AE000651AD, 0x80008FCB000641AF,
-    0x80008FD380008FCF, 0x80008FDB80008FD7, 0x80008FE380008FDF,
-    0x0006C9B880008FE7, 0x0006D9BA0006C9B9, 0x0006C9BC0006D9BB,
-    0x0006E9BE0006FDBD, 0x000601B00006E5BF, 0x000601B2000601B1,
-    0x000601B4000601B3, 0x000605B6000601B5, 0x0006B5A30006FDB7,
-    0x0000B58100001582, 0x80008FEB0000B580, 0x0006A5A50006D1A6,
-    0x0006C5AB000190BE, 0x0002A0840006C5AA, 0x0006C5AF80008FF3,
-    0x0006D5AD0006C5AE, 0x000178850006D5AC, 0x00008C850001F1B3,
-    0x000155B680008FF7, 0x80008FFF80008FFB, 0x000171BA0001E1B5,
-    0x003FC886000179BB, 0x000131BE0000BC87, 0x000151BC000139BF,
-    0x0005EDA8000151BD, 0x00023DAA000235A9, 0x00022DAC000235AB,
-    0x000291AE000291AD, 0x80009003000291AF, 0x8000900B80009007,
-    0x000211A28000900F, 0x000601A0000211A3, 0x0002A9B80002F5A1,
-    0x000369BA0002A9B9, 0x000379BC000369BB, 0x000369BE000379BD,
-    0x0002F1B0000369BF, 0x0002C1B20002F9B1, 0x0002B1B40002C1B3,
-    0x0002ADB60002B9B5, 0x0002B5A30002A5B7, 0x80009013001A8D8F,
-    0x8000901B80009017, 0x0002A5A5000211A6, 0x00023DAB8000901F,
-    0x80009023000235AA, 0x00027DAF80009027, 0x000215AD000275AE,
-    0x0006059E000215AC, 0x000AE99C0006E59F, 0x000FF19A000A219D,
-    0x000ED998000A659B, 0x0012C596000ED999, 0x0012759400138D97,
-    0x0016D1920012CD95, 0x001AF59000167193, 0x8000902B00167591,
-    0x003CC4840000D8E1, 0x8000902F0032C8E3, 0x8000903780009033,
-    0x8000903F8000903B, 0x8000904780009043, 0x8000904F8000904B,
-    0x8000905780009053, 0x8000905B003768EF, 0x003C80BE0007C4E3,
-    0x000065800004BCE1, 0x0000198200006D81, 0x003C408680009063,
-    0x80009067003DD887, 0x8000906F8000906B, 0x800090730006ACEF,
-    0x002B85B180009077, 0x002AA1B3002801B0, 0x002791B5002B9DB2,
-    0x0000FCEF002401B4, 0x000190E10006C4EF, 0x000010E30006ECE1,
-    0x8000907B0006F0E3, 0x003F74848000907F, 0x003FC9A180009083,
-    0x003ED5A380009087, 0x003BD5A5003FC1A2, 0x003401A7003801A4,
-    0x0037B9A9003B61A6, 0x0033EDAB0037E1A8, 0x002C01AD003001AA,
-    0x002F91AF0033DDAC, 0x002C258B002FF9AE, 0x80008FEF002C258A,
-    0x0031318F8000905F, 0x0030198D0030118E, 0x00242D83002D758C,
-    0x8000908B00242D82, 0x002CAD878000908F, 0x002819850028B186,
-    0x003C619B0028B584, 0x80009093003C8D9A, 0x8000909B80009097,
-    0x800090A38000909F, 0x00344193003C799C, 0x800090A70034D592,
-    0x00384997800090AB, 0x0038BD9500385596, 0x003D2DA800349194,
-    0x003D5DAA003D49A9, 0x003EB1AC003D55AB, 0x003EB1AE003EB1AD,
-    0x800090AF003EB1AF, 0x800090B7800090B3, 0x800090BF800090BB,
-    0x0001B1810001B982, 0x003FD5B80001B980, 0x003FD5BA003FDDB9,
-    0x003FF5BC003FEDBB, 0x003FF5BE003FFDBD, 0x003E69B0003FEDBF,
-    0x003E79B2003E69B1, 0x003E59B4003E71B3, 0x003FF9B6003E59B5,
-    0x003CD5B3003FEDB7, 0x800090C7800090C3, 0x0001D8BE800090CB,
-    0x003F39B5003F31B6, 0x003FADBB800090CF, 0x00008886003FADBA,
-    0x003F9DBF0001AC87, 0x003FA5BD003F9DBE, 0x800090D3003FADBC,
-    0x800090D7003C91A3, 0x003F75A6800090DB, 0x800090E3800090DF,
-    0x003FE9AA003F7DA5, 0x800090E7003FE9AB, 0x003FD9AE800090EB,
-    0x003FE9AC003FD9AF, 0x003EDDA8003FE1AD, 0x003EE5AA003EEDA9,
-    0x003EE9AC003EFDAB, 0x003E29AE003EE9AD, 0x0001A484003E29AF,
-    0x800090F3800090EF, 0x800090FB800090F7, 0x80009103800090FF,
-    0x000199B880009107, 0x0001A5BA0001ADB9, 0x000175BC00016DBB,
-    0x000175BE00017DBD, 0x003E5DB000016DBF, 0x0001C1B2003E21B1,
-    0x0001B9B40001C1B3, 0x0001A9B60001B9B5, 0x003D6DB30001A9B7,
-    0x8000910F8000910B, 0x8000911780009113, 0x003D61B5003D69B6,
-    0x003DE5BB8000911B, 0x8000911F003D99BA, 0x00027DBF000368BE,
-    0x000215BD00027DBE, 0x80009123000215BC, 0x80009127003D29A3,
-    0x003D2DA68000912B, 0x800091338000912F, 0x003DDDAA003D25A5,
-    0x80009137003DA1AB, 0x000239AE8000913B, 0x000251AC000239AF,
-    0x0000D980000251AD, 0x0000E5820000ED81, 0x000120E380009143,
-    0x0001A8E180009147, 0x000274EF8000914B, 0x00036C878000914F,
-    0x8000915300048C86, 0x8000915B80009157, 0x8000915F000530BE,
-    0x000385B380009163, 0x000389B580009167, 0x8000916F8000916B,
-    0x800091730003B5B6, 0x000345BB80009177, 0x000345BD000345BA,
-    0x000345BF000345BC, 0x000241A8000345BE, 0x000245AA00024DA9,
-    0x000249AC000259AB, 0x000275AE00027DAD, 0x0005EC840002B9AF,
-    0x8000917F8000917B, 0x0004B88480009183, 0x8000918B80009187,
-    0x00014DB88000918F, 0x000155BA000155B9, 0x000111BC000165BB,
-    0x000115BE00011DBD, 0x0002C9B0000109BF, 0x0002D9B20002C9B1,
-    0x0002C9B40002D9B3, 0x00017DB60002C9B5, 0x003E28E3000175B7,
-    0x003EF0E1003E9CE3, 0x80009193003C7CE1, 0x8000919B80009197,
-    0x800091A38000919F, 0x800091AB800091A7, 0x000058EF800091B3,
-    0x800091B7003E04EF, 0x00006581800091BB, 0x000285A300006580,
-    0x000289A500006582, 0x800091C3800091BF, 0x0004C0860002B5A6,
-    0x000245AB00054487, 0x000245AD000245AA, 0x000245AF000245AC,
-    0x00069DB3000245AE, 0x800091C7800091AF, 0x800091CF800091CB,
-    0x0006A9B50006A1B6, 0x0006EDBB800091D3, 0x800091D70006E1BA,
-    0x0006D5BF800091DB, 0x0006E1BD0006D5BE, 0x800091DF0006E9BC,
-    0x800091E7800091E3, 0x0004E4E1800091EB, 0x000568E3800091EF,
-    0x800091F7800091F3, 0x00042CEF800091FB, 0x80009203800091FF,
-    0x8000920B80009207, 0x00060DA88000920F, 0x000615AA000611A9,
-    0x000621AC000629AB, 0x000621AE000621AD, 0x00000D8000065DAF,
-    0x00001D8200001581, 0x8000921780009213, 0x000034BE8000921B,
-    0x0007EDB88000921F, 0x0007F1BA0007F1B9, 0x000741BC0007F1BB,
-    0x000741BE000741BD, 0x000625B0000741BF, 0x00063DB2000629B1,
-    0x000611B4000635B3, 0x0007DDB6000611B5, 0x00079DA30007D5B7,
-    0x0001288680009223, 0x8000922700015C87, 0x0007A9A50007A1A6,
-    0x0007EDAB8000922B, 0x8000922F0007E1AA, 0x0007D5AF80009233,
-    0x0007E1AD0007D5AE, 0x800092370007E9AC, 0x8000923B0006F9B3,
-    0x000159B68000923F, 0x8000924780009243, 0x00017DBA000685B5,
-    0x8000924B00014DBB, 0x000135BE8000924F, 0x000155BC00013DBF,
-    0x00062DA800013DBD, 0x000635AA00063DA9, 0x000649AC000659AB,
-    0x000195AE00019DAD, 0x8000925300018DAF, 0x8000925B80009257,
-    0x800092638000925F, 0x8000926B80009267, 0x00019DB88000926F,
-    0x0001ADBA0001A5B9, 0x0001A1BC0001A1BB, 0x0001A1BE0001A1BD,
-    0x0001FDB00001A1BF, 0x0001C5B20001CDB1, 0x0001B5B40001ADB3,
-    0x0001B5B60001BDB5, 0x0005BDA30001A9B7, 0x8000927780009273,
-    0x8000927F8000927B, 0x0005C1A500021DA6, 0x000209AB80009283,
-    0x80009287000239AA, 0x000279AF8000928B, 0x000279AD000271AE,
-    0x000304BE000211AC, 0x0003C48480009293, 0x00004D8080009297,
-    0x0000758200007D81, 0x8000929F8000929B, 0x800092A7800092A3,
-    0x000C5086800092AB, 0x800092AF0003E487, 0x000385A8800092B3,
-    0x000395AA000395A9, 0x0003BDAC0003A5AB, 0x0003D1AE0003D5AD,
-    0x000000A80003D1AF, 0x800092BB800092B7, 0x800092C3800092BF,
-    0x800092CB800092C7, 0x000371B8800092CF, 0x000371BA000371B9,
-    0x0000D5BC000371BB, 0x0000D5BE0000DDBD, 0x0003B5B00000CDBF,
-    0x000381B20003BDB1, 0x000351B4000381B3, 0x000351B6000351B5,
-    0x0003B8EF000351B7, 0x800092D7800092D3, 0x000DDC84800092DB,
-    0x800092E3800092DF, 0x000CACBE800092E7, 0x800092EF800092EB,
-    0x000368E3800092F3, 0x000028E1800092F7, 0x00025DB3800092FB,
-    0x000C688680009303, 0x80009307000D0C87, 0x000275B5000215B6,
-    0x000239BB8000930B, 0x8000930F000231BA, 0x0002D5BF80009313,
-    0x000215BD0002D5BE, 0x000678E1000215BC, 0x0006A8E380009317,
-    0x0000BD80000C0484, 0x0000A5820000A181, 0x8000931F8000931B,
-    0x8000932780009323, 0x8000932F8000932B, 0x0001F8EF80009333,
-    0x8000933B80009337, 0x0002D5A38000933F, 0x0002FDA580009343,
-    0x800093478000913F, 0x8000934B00029DA6, 0x0002B1AB8000934F,
-    0x00029DAD0002B9AA, 0x00025DAF00029DAC, 0x000E5DB300025DAE,
-    0x80009353800092FF, 0x8000935B80009357, 0x000E7DB5000E7DB6,
-    0x000E85BB8000935F, 0x80009363000E59BA, 0x000E85BF80009367,
-    0x000E95BD000E85BE, 0x00006980000E95BC, 0x0000058200006981,
-    0x0000DCBE8000936B, 0x8000936F000D54E3, 0x80009373000C34E1,
-    0x00004886000DBCEF, 0x800093770003CC87, 0x8000937F8000937B,
-    0x000EC5A880009383, 0x000ED5AA000EC9A9, 0x000EF5AC000EEDAB,
-    0x000EF5AE000EFDAD, 0x80009387000EE5AF, 0x8000938F8000938B,
-    0x8000939780009393, 0x8000939F8000939B, 0x000FC5B8800093A3,
-    0x000FC5BA000FCDB9, 0x000FC9BC000FD9BB, 0x000FFDBE000FC9BD,
-    0x000E9DB0000FF5BF, 0x000E41B2000E41B1, 0x000E41B4000E41B3,
-    0x000E41B6000E41B5, 0x000F5DA3000E41B7, 0x800093AB800093A7,
-    0x800093B3800093AF, 0x000F7DA5000F7DA6, 0x000F85AB800093B7,
-    0x800093BB000F59AA, 0x000F85AF800093BF, 0x000F95AD000F85AE,
-    0x800093C3000F95AC, 0x800093C7000E19B3, 0x000ED5B6800093CB,
-    0x800093D3800093CF, 0x000EC5BA000E31B5, 0x800093D7000ECDBB,
-    0x000EC5BE800093DB, 0x000ED5BC000EC9BF, 0x000EC5A8000ECDBD,
-    0x000ED5AA000ED5A9, 0x000EFDAC000EE5AB, 0x000E05AE000E11AD,
-    0x0000F980000E3DAF, 0x0000C5820000CD81, 0x00003CBE00012484,
-    0x0000FC87800093E3, 0x000E0DB800079C86, 0x000E29BA000E1DB9,
-    0x0001D9BC000E29BB, 0x0001C9BE0001D9BD, 0x000E45B00001C1BF,
-    0x000E45B2000E4DB1, 0x000E45B4000E5DB3, 0x000E39B6000E49B5,
-    0x000D5DA3000E39B7, 0x800093EB800093E7, 0x800093F3800093EF,
-    0x000D75A5000D91A6, 0x000D89AB800093F7, 0x800093FB000D81AA,
-    0x000D8DAF800093FF, 0x000D89AD000D81AE, 0x0003C484000D91AC,
-    0x80009403000231B3, 0x0002F1B680009407, 0x8000940F8000940B,
-    0x0002CDBA000221B5, 0x800094130002D5BB, 0x0002CDBE80009417,
-    0x0002CDBC000371BF, 0x0002E1A80002C5BD, 0x0002FDAA0002F5A9,
-    0x0002EDAC0002F5AB, 0x00023DAE000235AD, 0x8000941B000235AF,
-    0x0013A4BF8000941F, 0x8000942780009423, 0x8000942F8000942B,
-    0x0003D1B880009433, 0x0003E1BA0003D9B9, 0x000391BC0003E1BB,
-    0x000395BE00039DBD, 0x00024DB0000389BF, 0x00025DB2000255B1,
-    0x00024DB4000255B3, 0x0003F1B60003F1B5, 0x000375A30003F1B7,
-    0x8000943B80009437, 0x800094438000943F, 0x000365A50003B5A6,
-    0x000391AB80009447, 0x8000944B000389AA, 0x000235AF0005A0BE,
-    0x000381AD000389AE, 0x00005D81000389AC, 0x000141B300005580,
-    0x000141B500006582, 0x8000945780009453, 0x0004C086000141B6,
-    0x000125BB00052487, 0x000129BD000125BA, 0x000115BF000121BC,
-    0x000ED0E1000115BE, 0x000E14E38000945B, 0x8000945F0004C084,
-    0x8000946780009463, 0x0007E8848000946B, 0x8000946F0004ECBE,
-    0x8000947780009473, 0x000110EF8000947B, 0x8000947F0006B084,
-    0x8000948780009483, 0x8000948F8000948B, 0x8000928F000068E3,
-    0x800094930001ACE1, 0x800094970000ACEF, 0x8000949F8000949B,
-    0x800094A7800094A3, 0x000609A3800094AB, 0x800094AF8000944F,
-    0x800094B7800094B3, 0x000609A5000609A6, 0x00066DAB800094BB,
-    0x800094BF00066DAA, 0x00065DAF800094C3, 0x000661AD00065DAE,
-    0x00004D80000669AC, 0x0000558200005581, 0x000584EF800094C7,
-    0x0003D487800094CB, 0x000430E30000FC86, 0x0007CCE1800094CF,
-    0x800094D7800094D3, 0x800094DF800094DB, 0x000535A8800094E3,
-    0x000681AA000681A9, 0x000681AC000681AB, 0x0006B1AE000689AD,
-    0x0000E4BE0006B1AF, 0x800094EB800094E7, 0x800094F3800094EF,
-    0x800094FB800094F7, 0x00066DB8800094FF, 0x00067DBA000675B9,
-    0x00066DBC000675BB, 0x0007D1BE0007D1BD, 0x0006A9B00007D1BF,
-    0x0006BDB20006A9B1, 0x000699B40006B5B3, 0x000659B6000699B5,
-    0x000655A8000659B7, 0x00068DAA00065DA9, 0x000685AC00069DAB,
-    0x000685AE00068DAD, 0x800095030006B5AF, 0x8000950B80009507,
-    0x800095138000950F, 0x8000951B80009517, 0x0006A1B88000951F,
-    0x0006A5BA0006ADB9, 0x0006A5BC0006BDBB, 0x0006A5BE0006ADBD,
-    0x0006CDB0000155BF, 0x0006DDB20006D5B1, 0x0006B5B40006ADB3,
-    0x0006B5B60006BDB5, 0x000605B30006ADB7, 0x0000158100002D82,
-    0x8000952300001D80, 0x00060DB500060DB6, 0x000611BB80009527,
-    0x8000952B000609BA, 0x000679BF80009533, 0x000679BD000671BE,
-    0x80009537000609BC, 0x00006886000641A3, 0x000649A600005487,
-    0x8000953F8000953B, 0x00064DAA000649A5, 0x80009543000655AB,
-    0x000635AE80009547, 0x00064DAC00063DAF, 0x00022DA800063DAD,
-    0x00025DAA000249A9, 0x00027DAC000269AB, 0x00026DAE000265AD,
-    0x001CACBE000265AF, 0x8000954F8000954B, 0x8000955780009553,
-    0x8000955F8000955B, 0x000361B880009563, 0x000361BA000361B9,
-    0x000361BC000361BB, 0x000361BE000361BD, 0x00021DB0000361BF,
-    0x0002C1B20002C1B1, 0x0002C1B40002C1B3, 0x0002C1B60002C1B5,
-    0x800095670002C1B7, 0x8000956B0002D1B3, 0x0002E1B6001D4084,
-    0x800095738000956F, 0x0002DDBA0002F1B5, 0x80009577000201BB,
-    0x000201BE8000957B, 0x000211BC000201BF, 0x8000957F000211BD,
-    0x00029DA380009583, 0x0002BDA580009587, 0x8000958F8000958B,
-    0x800095930002ADA6, 0x00024DAB80009597, 0x00025DAD000291AA,
-    0x00024DAF00025DAC, 0x0002ADA800024DAE, 0x0002DDAA0002C5A9,
-    0x0002F9AC0002D5AB, 0x00013DAE0002F9AD, 0x8000959B000135AF,
-    0x800095A7800095A3, 0x800095AF800095AB, 0x800095B7800095B3,
-    0x0001D9B8800095BB, 0x0001E5BA0001EDB9, 0x000195BC00018DBB,
-    0x00018DBE000199BD, 0x00014DB0000185BF, 0x00015DB2000155B1,
-    0x00014DB4000155B3, 0x0001F1B60001F5B5, 0x0006F8E10001E9B7,
-    0x000614E3800095BF, 0x0000B182001C44BE, 0x000179800001CCEF,
-    0x001F6C8400017981, 0x800095C3000614E3, 0x800095C70001E8E1,
-    0x0006E4EF800095CB, 0x000215B3800095CF, 0x001DE886800095D3,
-    0x0017C085001D9C87, 0x000231B5000239B6, 0x0001D1BB800095D7,
-    0x800095DB0001C9BA, 0x000171BF800093DF, 0x0001C1BD0001C9BE,
-    0x001DD5A30001C9BC, 0x800095DF8000959F, 0x800095E7800095E3,
-    0x001DF1A5001DF9A6, 0x001E11AB800095EB, 0x800095EF001E09AA,
-    0x001EB1AF800095F3, 0x001E01AD001E09AE, 0x800095F7001E09AC,
-    0x800095FF800095FB, 0x00000D80001C94EF, 0x00003D8200003581,
-    0x8000960780009603, 0x8000960F8000960B, 0x8000961780009613,
-    0x8000961B001FA4E1, 0x001FA5A80019D0E3, 0x001FA5AA001FADA9,
-    0x001FA5AC001FBDAB, 0x001FA5AE001FADAD, 0x00038086001F19AF,
-    0x8000961F00030887, 0x8000962780009623, 0x8000962F8000962B,
-    0x0000CDB880009633, 0x0000D5BA0000D5B9, 0x0000FDBC0000E5BB,
-    0x00009DBE000095BD, 0x001F49B0000095BF, 0x001F5DB2001F55B1,
-    0x001F35B4001F51B3, 0x001F11B6001F29B5, 0x001E2DB30000F5B7,
-    0x8000963B80009637, 0x800096438000963F, 0x001E2DB5001E2DB6,
-    0x001E91BB80009647, 0x8000964B001E8DBA, 0x001EEDBF8000964F,
-    0x001E85BD001EF1BE, 0x80009653001E89BC, 0x80009657001E69A3,
-    0x001E69A68000965B, 0x800096638000965F, 0x001EC9AA001E69A5,
-    0x80009667001ED5AB, 0x001EB5AE8000966B, 0x001ECDAC001EA9AF,
-    0x0001EDA8001EC1AD, 0x0001F1AA0001F5A9, 0x00013DAC000125AB,
-    0x00012DAE000125AD, 0x8000966F000125AF, 0x00001D8100001D82,
-    0x8000967300001D80, 0x8000967B80009677, 0x000081B8000174BE,
-    0x000085BA00008DB9, 0x000089BC000099BB, 0x0000B5BE0000BDBD,
-    0x00015DB000007DBF, 0x0000E1B20000E1B1, 0x0000E1B40000E1B3,
-    0x0000D1B60000E9B5, 0x000279B30000D1B7, 0x0003288780009683,
-    0x8000968700022C86, 0x0002B5B50002A5B6, 0x000325BB8000968B,
-    0x8000968F000299BA, 0x00032DBF80009693, 0x000335BD000325BE,
-    0x80009697000335BC, 0x8000969B00023DA3, 0x0002E1A68000969F,
-    0x800096A7800096A3, 0x0002DDAA0002F1A5, 0x800096AB000361AB,
-    0x000361AE800096AF, 0x000371AC000369AF, 0x000085A8000371AD,
-    0x000095AA000095A9, 0x0000BDAC0000A5AB, 0x0000D1AE0000D5AD,
-    0x800096B30000D1AF, 0x800096BB800096B7, 0x800096C3800096BF,
-    0x000000AE800096C7, 0x000071B8800096CB, 0x000075BA000071B9,
-    0x0001D5BC0001CDBB, 0x0001D5BE0001DDBD, 0x0000B5B00001CDBF,
-    0x000081B20000BDB1, 0x000051B4000081B3, 0x000051B6000051B5,
-    0x800096CF000051B7, 0x800096D7800096D3, 0x800096DF800096DB,
-    0x800096E7800096E3, 0x00003580800096EB, 0x0000358200003D81,
-    0x0005FCBE0002A084, 0x0002C487800096F3, 0x00015DB300051C86,
-    0x800096F7000B2085, 0x800096FF800096FB, 0x000175B5000115B6,
-    0x000139BB80009703, 0x80009707000131BA, 0x0001D5BF8000970B,
-    0x000115BD0001D5BE, 0x001E50E1000115BC, 0x001F14E38000970F,
-    0x8000971780009713, 0x8000971F8000971B, 0x8000972780009723,
-    0x8000972F8000972B, 0x8000973300046084, 0x001ECCEF80009737,
-    0x8000973F8000973B, 0x000400BE80009743, 0x8000974700070084,
-    0x8000974B000090EF, 0x800097538000974F, 0x80009757001E54E3,
-    0x8000952F000154E1, 0x000011810007F982, 0x000414E100001D80,
-    0x000534E3800096EF, 0x8000975F8000975B, 0x0003388680009763,
-    0x8000976700035487, 0x8000976F8000976B, 0x8000977780009773,
-    0x0005F0EF8000977B, 0x00065DA38000977F, 0x8000978780009783,
-    0x8000978F8000978B, 0x000675A5000615A6, 0x000639AB80009793,
-    0x80009797000631AA, 0x0006D5AF8000979B, 0x000615AD0006D5AE,
-    0x0007B5A8000615AC, 0x0007C5AA0007CDA9, 0x0007CDAC0007DDAB,
-    0x0007F5AE0007FDAD, 0x8000979F000719AF, 0x800097A7800097A3,
-    0x800097AF800097AB, 0x800097B7800097B3, 0x000771B8800097BB,
-    0x000775BA000771B9, 0x000715BC00070DBB, 0x000701BE000705BD,
-    0x00076DB0000701BF, 0x00077DB2000775B1, 0x00076DB4000775B3,
-    0x00075DB6000755B5, 0x00064DB3000755B7, 0x800097C3800097BF,
-    0x800097CB800097C7, 0x000645B500067DB6, 0x0006B1BB0001D4BE,
-    0x800097CF0006A9BA, 0x000691BF800097D3, 0x0006A1BD0006A9BE,
-    0x000051820006A9BC, 0x00004180000609A3, 0x000639A600004181,
-    0x800097DB800097D7, 0x0006EDAA000601A5, 0x0000C8860006F5AB,
-    0x0006EDAE0001EC87, 0x0006EDAC0006D5AF, 0x000195A80006E5AD,
-    0x0001ADAA0001A5A9, 0x0001BDAC0001A5AB, 0x0001ADAE0001A5AD,
-    0x0001A8840001A1AF, 0x800097E3800097DF, 0x800097EB800097E7,
-    0x800097F3800097EF, 0x000165B8800097F7, 0x000165BA00016DB9,
-    0x000165BC00017DBB, 0x000165BE00016DBD, 0x0001A5B00000D9BF,
-    0x0001A5B20001ADB1, 0x0001ADB40001BDB3, 0x000195B600019DB5,
-    0x000225B300015DB7, 0x800097FF800097FB, 0x8000980780009803,
-    0x000235B5000225B6, 0x000269BB8000980B, 0x8000980F000261BA,
-    0x000259BF80009813, 0x000269BD000269BE, 0x80009817000271BC,
-    0x8000981B000261A3, 0x000261A68000981F, 0x8000982780009823,
-    0x000225AA000271A5, 0x8000982B00022DAB, 0x00022DAE8000982F,
-    0x000235AC00021DAF, 0x00000D8000022DAD, 0x0000118200001581,
-    0x8000983780009833, 0x8000983B0002F0EF, 0x800098430003D0BE,
-    0x80009847000114E3, 0x8000984B0001DCE1, 0x800098538000984F,
-    0x8000985B80009857, 0x000CE88600039DB3, 0x00037DB600030087,
-    0x800098638000985F, 0x000359BA000375B5, 0x80009867000359BB,
-    0x0000FDBE8000986B, 0x0000FDBC0000E5BF, 0x00022DA80000F5BD,
-    0x000251AA000251A9, 0x0002B9AC0002A9AB, 0x0002A9AE0002B9AD,
-    0x000DAC840002A9AF, 0x800098738000986F, 0x8000987B80009877,
-    0x800098838000987F, 0x00016DB880009887, 0x000101BA000105B9,
-    0x000105BC000101BB, 0x000131BE00010DBD, 0x0002D9B0000131BF,
-    0x00016DB20002D9B1, 0x00017DB4000165B3, 0x000165B6000165B5,
-    0x0006A4E1000155B7, 0x000634E38000967F, 0x8000988F8000988B,
-    0x80009893000194E1, 0x80009897000640E3, 0x8000989F8000989B,
-    0x000C3CBE000048EF, 0x000688EF800098A3, 0x000299A3800098A7,
-    0x0000758100001582, 0x800098AB00007980, 0x000271A5000279A6,
-    0x00025DAB800098B3, 0x000CC88600025DAA, 0x0001E1AF000D5C87,
-    0x0001F1AD0001F9AE, 0x800098AF0001F9AC, 0x800098BB800098B7,
-    0x000CC0E1800098BF, 0x000D28E3800098C3, 0x800098CB800098C7,
-    0x000D60EF800098CF, 0x800098D7800098D3, 0x800098DF800098DB,
-    0x000EF1B3800098E3, 0x800098EB800098E7, 0x800098F3800098EF,
-    0x000EE1B5000EE9B6, 0x000E89BB800098F7, 0x800098FB000EBDBA,
-    0x000F5DBF800098FF, 0x000F5DBD000F5DBE, 0x000E21A8000E91BC,
-    0x000E39AA000E2DA9, 0x000E29AC000E39AB, 0x000E59AE000E29AD,
-    0x80009903000E49AF, 0x8000990B80009907, 0x00000D808000990F,
-    0x0001B1820001B181, 0x000FF9B880009913, 0x000F45BA000FF9B9,
-    0x000F49BC000F5DBB, 0x000F7DBE000F75BD, 0x000E39B0000F75BF,
-    0x000E05B2000E0DB1, 0x000FD9B4000FC9B3, 0x000FC9B6000FD9B5,
-    0x000FB1A3000FC9B7, 0x8000991B80009917, 0x00019CBE00028484,
-    0x000FA1A5000FA9A6, 0x000FC9AB80009923, 0x00078886000FFDAA,
-    0x000E1DAF00005C87, 0x000E1DAD000E1DAE, 0x80009927000FD1AC,
-    0x8000992B00014DB3, 0x00014DB68000992F, 0x8000993780009933,
-    0x00014DBA000149B5, 0x8000993B000155BB, 0x00014DBE8000993F,
-    0x00014DBC000131BF, 0x000E5DA8000145BD, 0x000181AA000181A9,
-    0x000185AC000181AB, 0x0001B9AE000189AD, 0x800099430001B9AF,
-    0x8000994B80009947, 0x800099538000994F, 0x8000995B80009957,
-    0x000191B88000995F, 0x0001A1BA000199B9, 0x000171BC0001A1BB,
-    0x000171BE000171BD, 0x0001CDB0000171BF, 0x0001DDB20001D5B1,
-    0x0001CDB40001D5B3, 0x0001B1B60001B1B5, 0x000209A30001B1B7,
-    0x8000996780009963, 0x8000996F8000996B, 0x00020DA5000209A6,
-    0x000211AB80009973, 0x80009977000209AA, 0x000275AF001A258F,
-    0x000201AD000209AE, 0x001EFD94000209AC, 0x00124D9600119D95,
-    0x001A5D9000121597, 0x001D8D92001B4991, 0x000A659C001E5993,
-    0x000DED9E000A0D9D, 0x00159D98000E699F, 0x0016519A00167199,
-    0x8000997B00098D9B, 0x8000997F000BD4E1, 0x000E81A0007ED8E3,
-    0x00010DA2000FBDA1, 0x006801B100029DA3, 0x006A69B3007649B0,
-    0x006E91B5006A8DB2, 0x80009983006C01B4, 0x007ED1A9006C64EF,
-    0x0072CDAB007EC1A8, 0x007391AD007FBDAA, 0x0076F9AF0072DDAC,
-    0x000401A1007401AE, 0x000679A380009987, 0x007AD1A50006CDA2,
-    0x007C01A7007801A4, 0x001B0D83007AC1A6, 0x8000998B001B0D82,
-    0x001FAD87000500BE, 0x001CCD85001F1186, 0x0013258B001A0184,
-    0x000230840010FD8A, 0x0017198F80009993, 0x0014D58D0017318E,
-    0x000B45930013658C, 0x000468860008B592, 0x000F419700057C87,
-    0x000CDD95000F5196, 0x0000F59B000B3594, 0x80009997000E4D9A,
-    0x0001C9808000999B, 0x0000718200019181, 0x8000999F00037D9C,
-    0x00058084000E74E1, 0x800099A3000090E3, 0x800099A78000983F,
-    0x800099AF800099AB, 0x800099B7800099B3, 0x800099BF800099BB,
-    0x800099C7800099C3, 0x000299B6000054EF, 0x800099CF800099CB,
-    0x800099D3000299B5, 0x800099D700029DB3, 0x000251BE800099DB,
-    0x000241BC000251BF, 0x000251BA000241BD, 0x800099DF000251BB,
-    0x00051DA8800099E3, 0x000541AA000545A9, 0x000545AC000541AB,
-    0x000571AE00054DAD, 0x8000998F000571AF, 0x800099EB800099E7,
-    0x00000D80800099EF, 0x0000B1820000B181, 0x000645B8800099F3,
-    0x000645BA00064DB9, 0x00064DBC00065DBB, 0x0006D9BE00067DBD,
-    0x0005C1B00006D5BF, 0x0005C1B20005C1B1, 0x0005C1B40005C1B3,
-    0x0005C1B60005C1B5, 0x0005C9A30005C1B7, 0x800099FB800099F7,
-    0x800099FF000300BE, 0x0005CDA50005CDA6, 0x000505AB80009A03,
-    0x00000886000505AA, 0x000505AF0003DC87, 0x000515AD000505AE,
-    0x80009A07000515AC, 0x80009A0B0007A9B3, 0x000749B680009A0F,
-    0x80009A1780009A13, 0x000745BA00074DB5, 0x80009A1B00074DBB,
-    0x000745BE80009A1F, 0x000755BC00074DBF, 0x00066DA300074DBD,
-    0x80009A2780009A23, 0x80009A2F80009A2B, 0x000689A500068DA6,
-    0x000689AB80009A33, 0x80009A37000681AA, 0x000689AF80009A3B,
-    0x000689AD000681AE, 0x80009A3F000691AC, 0x80009A4780009A43,
-    0x80009A4F80009A4B, 0x80009A5780009A53, 0x80009A5B0004A8EF,
-    0x80009A5F000654E1, 0x00000D80000454E3, 0x0000DD820000D581,
-    0x0006ADA880009A63, 0x0006E5AA0006EDA9, 0x000601AC00060DAB,
-    0x000605AE00060DAD, 0x80009A6700063DAF, 0x0001E48480009A6B,
-    0x007F10860001E8BE, 0x80009A7300009487, 0x000611B880009A77,
-    0x000615BA00061DB9, 0x00062DBC000635BB, 0x0001F1BE0001F1BD,
-    0x000645B00001F1BF, 0x000645B200064DB1, 0x000645B400065DB3,
-    0x00063DB600064DB5, 0x00029DA8000635B7, 0x0002C1AA0002C5A9,
-    0x0002C5AC0002C1AB, 0x0002F1AE0002CDAD, 0x80009A7B0002F1AF,
-    0x80009A8380009A7F, 0x0000008C80009A87, 0x80009A8F80009A8B,
-    0x000305B880009A93, 0x000305BA00030DB9, 0x000305BC00031DBB,
-    0x000339BE00030DBD, 0x000365B0000335BF, 0x000365B200036DB1,
-    0x00036DB400037DB3, 0x000355B600035DB5, 0x000249B300033DB7,
-    0x80009A9B80009A97, 0x80009AA380009A9F, 0x00024DB500024DB6,
-    0x0003A9BB80009AA7, 0x80009AAB0003A1BA, 0x000399BF80009AAF,
-    0x0003A9BD0003A9BE, 0x80009AB30003B1BC, 0x000201A380009AB7,
-    0x000205A580009ABB, 0x80009AC380009ABF, 0x8000991F000205A6,
-    0x0003E1AB80009AC7, 0x0003E1AD0003E9AA, 0x0003D1AF0003F9AC,
-    0x80009ACB0003E1AE, 0x80009AD380009ACF, 0x80009ADB80009AD7,
-    0x000244BE0001A4E1, 0x000029800000A4E3, 0x0000398200002981,
-    0x80009AE3000078EF, 0x80009AE70002B884, 0x000161B380009AEB,
-    0x80009AF380009AEF, 0x80009AFB80009AF7, 0x000171B5000179B6,
-    0x0001E5BB80009AFF, 0x007D4886000159BA, 0x0001EDBF007D4087,
-    0x0001F5BD0001E5BE, 0x80009B030001F5BC, 0x80009B0B80009B07,
-    0x80009B0F007C4084, 0x80009B1780009B13, 0x80009B1B0001E8EF,
-    0x000000EA0006F4E1, 0x80009B1F000634E3, 0x007CB08480009B23,
-    0x80009B2B80009B27, 0x0001E1A50001E9A6, 0x0001F1A380009B2F,
-    0x80009B3780009B33, 0x00017DAF80009B3B, 0x000165AD000175AE,
-    0x000175AB000165AC, 0x80009B3F0001C9AA, 0x007DB9A880009B43,
-    0x007E29AA007DB9A9, 0x007E39AC007E29AB, 0x007E2DAE007E39AD,
-    0x80009ADF007E25AF, 0x0000198100001182, 0x80009B47007FD180,
-    0x000334BE80009B4B, 0x007E81B880009B4F, 0x007E85BA007E81B9,
-    0x007E89BC007E99BB, 0x007F59BE007EB1BD, 0x007E5DB0007F59BF,
-    0x007EE1B2007EE1B1, 0x007EE1B4007EE1B3, 0x007ED1B6007EE9B5,
-    0x007FB1A8007ED1B7, 0x007FB5AA007FBDA9, 0x007FD9AC007FC9AB,
-    0x007FC9AE007FD9AD, 0x00000086007FC1AF, 0x80009B5300032887,
-    0x80009B5B80009B57, 0x80009B6380009B5F, 0x007F75B880009B67,
-    0x007F09BA007F79B9, 0x007F05BC007F01BB, 0x007F05BE007F0DBD,
-    0x007F61B0007F25BF, 0x007F61B2007F61B1, 0x007F65B4007F7DB3,
-    0x007F59B6007F6DB5, 0x007E3DB3007F4DB7, 0x80009B6F80009B6B,
-    0x80009B7780009B73, 0x007E51B5007E59B6, 0x007EA5BB80009B7B,
-    0x80009B7F007EA5BA, 0x007EA9BF80009B83, 0x007EA9BD007EA1BE,
-    0x80009B87007EB5BC, 0x007CA8EF80009B8B, 0x80009B9380009B8F,
-    0x80009B9B80009B97, 0x80009BA380009B9F, 0x80009BAB80009BA7,
-    0x007ED8E180009BAF, 0x007C5CE380009BB3, 0x007EBDA380009BB7,
-    0x0000F58100001582, 0x80009BBB0000F980, 0x007ED1A5007ED9A6,
-    0x007E25AB000190BE, 0x00000886007E25AA, 0x007E29AF00017087,
-    0x007E29AD007E21AE, 0x80009BC3007E35AC, 0x80009BC7000175B3,
-    0x00016DB680009BCB, 0x80009BD380009BCF, 0x0000EDBA000165B5,
-    0x80009BD70000F5BB, 0x0000DDBE80009BDB, 0x0000E5BC0000B5BF,
-    0x0002A9A80000D5BD, 0x0002B9AA0002A9A9, 0x0002A9AC0002B9AB,
-    0x00031DAE0002A9AD, 0x0004ACBE000315AF, 0x80009BE380009BDF,
-    0x80009BE70005A0BE, 0x80009BEF80009BEB, 0x000311B880009BF3,
-    0x000311BA000311B9, 0x000331BC000311BB, 0x000331BE000331BD,
-    0x00036DB0000331BF, 0x000375B2000375B1, 0x00035DB4000345B3,
-    0x00033DB6000335B5, 0x000231A3000331B7, 0x80009BFB80009BF7,
-    0x80009C0380009BFF, 0x000221A5000229A6, 0x0003B1AB80009C07,
-    0x80009C0B0003A9AA, 0x0003F1AF80009C0F, 0x000391AD000399AE,
-    0x0003C0840003A1AC, 0x80009C1780009C13, 0x80009C1F80009C1B,
-    0x00003D8100003D82, 0x80009C2300003D80, 0x80009C2B80009C27,
-    0x80009C3780009C33, 0x0004D88680009A6F, 0x0002C1A80004BC87,
-    0x0002C5AA0002CDA9, 0x0002C9AC0002D9AB, 0x0002F5AE0002FDAD,
-    0x80009C3B000139AF, 0x80009C4380009C3F, 0x80009C4B80009C47,
-    0x80009C5380009C4F, 0x0001D1B880009C57, 0x0001E1BA0001D9B9,
-    0x000195BC0001E1BB, 0x000195BE00019DBD, 0x000149B000018DBF,
-    0x00015DB2000155B1, 0x00014DB4000155B3, 0x0001F1B60001F1B5,
-    0x0000E4EF0001F1B7, 0x80009C5F80009C5B, 0x80009C6300050C84,
-    0x80009C6B80009C67, 0x80009C6F00015CEF, 0x80009C73007ED4E1,
-    0x007E18E3007E30E3, 0x000100E180009C77, 0x80009C7F80009C7B,
-    0x80009C8780009C83, 0x80009C8B000201B3, 0x000231B600020DB5,
-    0x80009C9380009C8F, 0x0001C1BA80009C97, 0x0001C5BC0001DDBB,
-    0x0001C5BE0001CDBD, 0x0005CDA3000179BF, 0x80009C9B80009C2F,
-    0x80009C9F80009CA3, 0x0005C1A50005FDA6, 0x000611AB80009CA7,
-    0x0003088600060DAA, 0x0006B5AF00037C87, 0x000601AD000609AE,
-    0x0007A180000609AC, 0x000041820007A181, 0x80009CAB00076DB3,
-    0x000775B600077DB5, 0x80009CB380009CAF, 0x000751BA80009CB7,
-    0x0007F5BC000751BB, 0x0007E9BE0007FDBD, 0x00066DA80007E9BF,
-    0x0007A9AA000675A9, 0x0007B9AC0007A9AB, 0x0007A9AE0007B9AD,
-    0x80009CBB0007A9AF, 0x80009CC380009CBF, 0x80009CCB80009CC7,
-    0x80009CD380009CCF, 0x000769B880009CD7, 0x000775BA000771B9,
-    0x000715BC00070DBB, 0x000709BE00071DBD, 0x0007D9B0000701BF,
-    0x00076DB20007D9B1, 0x000761B4000765B3, 0x000761B6000761B5,
-    0x000621A300075DB7, 0x80009CDF80009CDB, 0x80009CE780009CE3,
-    0x000631A5000639A6, 0x00061DAB80009CEB, 0x80009CEF00061DAA,
-    0x0006A5AF80009CF3, 0x0006B1AD0006A5AE, 0x80009CF70006B9AC,
-    0x80009CFB00059CEF, 0x00000D8080009CFF, 0x00003D8200003581,
-    0x80009D0780009D03, 0x0001A48480009D0B, 0x80009D0F0001A4BE,
-    0x80009D130005A4E3, 0x00018DA80006CCE1, 0x0001D9AA000199A9,
-    0x0001F1AC0001D1AB, 0x000139AE0001F1AD, 0x00000086000139AF,
-    0x80009D1700016487, 0x80009D1F80009D1B, 0x80009D2780009D23,
-    0x0000CDB880009D2B, 0x0000D5BA0000D5B9, 0x0000FDBC0000E5BB,
-    0x000091BE000095BD, 0x000149B0000091BF, 0x000159B2000149B1,
-    0x000149B4000159B3, 0x0000FDB6000149B5, 0x000251B30000F5B7,
-    0x80009D3380009D2F, 0x80009D3B80009D37, 0x000279B5000271B6,
-    0x00028DBB80009D3F, 0x80009D4300028DBA, 0x00028DBF80009D47,
-    0x00028DBD00028DBE, 0x00126C8500028DBC, 0x80009D4B000215A3,
-    0x000235A680009D4F, 0x80009D5780009D53, 0x0002C9AA00023DA5,
-    0x80009D5B0002C9AB, 0x0002C9AE80009D5F, 0x0002C9AC0002C9AF,
-    0x000380840002C9AD, 0x80009D6780009D63, 0x80009D6B000C00BE,
-    0x80009D7380009D6F, 0x80009D7B80009D77, 0x80009D8780009D83,
-    0x80009D8F80009D8B, 0x80009D9780009D93, 0x0000598080009D9B,
-    0x0000698200005981, 0x80009D9F0002B8EF, 0x0003188780009DA3,
-    0x80009DA7000CFC86, 0x80009DAF80009DAB, 0x80009DB780009DB3,
-    0x80009BBF00014CE1, 0x000279A800026CE3, 0x0002A5AA0002ADA9,
-    0x0002A5AC0002BDAB, 0x0002A5AE0002ADAD, 0x000DA884000119AF,
-    0x80009DBF80009DBB, 0x80009DC780009DC3, 0x80009DCF80009DCB,
-    0x000115B880009DD3, 0x000121BA00011DB9, 0x0001D5BC000121BB,
-    0x0001C9BE0001DDBD, 0x000169B00001C9BF, 0x000179B2000169B1,
-    0x000159B4000171B3, 0x000131B6000159B5, 0x000760E1000131B7,
-    0x000614E380009DD7, 0x80009DDB000C00BE, 0x80009DDF0001A4E1,
-    0x80009DE30000E4E3, 0x80009DEB80009DE7, 0x80009DF3000038EF,
-    0x00076CEF80009DF7, 0x0000618080009DFB, 0x0000618200006181,
-    0x80009DFF000261B3, 0x000261B6000261B5, 0x000CC08680009E03,
-    0x0001E1BA000CE887, 0x0001E1BC0001E1BB, 0x0001D1BE0001E9BD,
-    0x000DADA30001D1BF, 0x80009DEF80009D7F, 0x80009E0B80009E07,
-    0x000DADA5000DADA6, 0x000E2DAB80009E0F, 0x80009E13000E2DAA,
-    0x000E1DAF80009E17, 0x000E25AD000E1DAE, 0x80009E1B000E2DAC,
-    0x80009E1F000EE9B3, 0x000EE9B680009E23, 0x80009E2B80009E27,
-    0x000EB5BA000EE9B5, 0x80009E2F000E85BB, 0x000E9DBE80009E33,
-    0x000E9DBC000E85BF, 0x000E25A8000E95BD, 0x000E25AA000E2DA9,
-    0x000E25AC000E3DAB, 0x000E25AE000E2DAD, 0x80009E37000E4DAF,
-    0x80009E3F80009E3B, 0x80009E4780009E43, 0x80009E4F80009E4B,
-    0x000FEDB880009E53, 0x000FFDBA000FF5B9, 0x000FEDBC000FF5BB,
-    0x000F49BE000F51BD, 0x000E35B0000F45BF, 0x000E0DB2000E05B1,
-    0x000E1DB4000E05B3, 0x000FDDB6000E05B5, 0x000FADA3000FD5B7,
-    0x0000198100001982, 0x80009E570001B180, 0x000FADA5000FADA6,
-    0x000FC1AB80009E5B, 0x0002A084000FF1AA, 0x000FC1AF0001A0BE,
-    0x000FD1AD000FD9AE, 0x80009E5F000FD9AC, 0x0000E886000EF9B3,
-    0x00015DB600012C87, 0x80009E6780009E63, 0x000149BA000E81B5,
-    0x80009E6B000149BB, 0x00013DBE80009E6F, 0x000139BC000121BF,
-    0x000E31A8000135BD, 0x000E35AA000E31A9, 0x000E4DAC000E5DAB,
-    0x000189AE000199AD, 0x0001E484000181AF, 0x80009E7780009E73,
-    0x80009E7F80009E7B, 0x80009E8780009E83, 0x0001ADB880009E8B,
-    0x00016DBA000165B9, 0x00017DBC000165BB, 0x00016DBE000165BD,
-    0x0001C9B0000165BF, 0x0001A9B20001C9B1, 0x0001BDB40001A5B3,
-    0x0001A1B60001A1B5, 0x000DBDA3000195B7, 0x80009E9380009E8F,
-    0x80009E9B80009E97, 0x000DC5A5000219A6, 0x00020DAB80009E9F,
-    0x80009EA300020DAA, 0x000265AF80009EA7, 0x000271AD000279AE,
-    0x80009EAB00027DAC, 0x80009EB380009EAF, 0x80009EBB80009EB7,
-    0x80009EC380009EBF, 0x0000398080009EC7, 0x0000058200003981,
-    0x00033CBE80009ECB, 0x0003448780009ED3, 0x80009ED70004DC86,
-    0x0003E1B380009EDB, 0x000335B50004C084, 0x80009EE380009EDF,
-    0x80009EE7000335B6, 0x000311BB80009EEB, 0x0000B5BD000319BA,
-    0x0000B5BF000309BC, 0x80009EEF0000B5BE, 0x80009EF3000194E1,
-    0x80009EF700017CE3, 0x80009EFF80009EFB, 0x80009F0780009F03,
-    0x80009F0F80009F0B, 0x80009F1780009F13, 0x80009F1F80009F1B,
-    0x000255A8000274EF, 0x000261AA00025DA9, 0x0002B5AC000261AB,
-    0x0002B5AE0002BDAD, 0x000568BE0002ADAF, 0x80009F2780009F23,
-    0x80009F2F80009F2B, 0x0000AD810000AD82, 0x000169B80000AD80,
-    0x000109BA000169B9, 0x000119BC000109BB, 0x000109BE000119BD,
-    0x0002D5B0000109BF, 0x0002D5B20002DDB1, 0x000179B4000169B3,
-    0x000169B6000179B5, 0x000F44E1000161B7, 0x000EB8E30000BCE3,
-    0x80009F33000E74E1, 0x80009F3B80009F37, 0x80009F4780009F43,
-    0x80009F4F80009F4B, 0x000730BE80009F53, 0x000ECCEF000060EF,
-    0x0002A9A380009F57, 0x0004E88680009F5B, 0x80009F5F00050487,
-    0x00027DA500027DA6, 0x000259AB80009F63, 0x80009F67000251AA,
-    0x0001FDAF80009F6B, 0x0001FDAD0001FDAE, 0x000675A8000241AC,
-    0x00068DAA000685A9, 0x000689AC000699AB, 0x0006BDAE0006B5AD,
-    0x80009F3F0006DDAF, 0x80009F7380009F6F, 0x80009F7B80009F77,
-    0x80009F8380009F7F, 0x0006ADB880009F87, 0x0006BDBA0006B5B9,
-    0x000775BC0006B5BB, 0x000775BE00077DBD, 0x0006A9B000076DBF,
-    0x0006BDB20006B5B1, 0x0006ADB40006B5B3, 0x00069DB6000695B5,
-    0x000605B3000695B7, 0x80009F8F80009F8B, 0x80009F9780009F93,
-    0x000601B5000635B6, 0x000671BB80009F9B, 0x80009F9F00066DBA,
-    0x000651BF80009FA3, 0x000661BD000669BE, 0x00001982000669BC,
-    0x00006580000641A3, 0x000671A600001981, 0x80009FAB80009FA7,
-    0x000629AA000645A5, 0x00032484000635AB, 0x00062DAE80009FAF,
-    0x00062DAC000615AF, 0x0006A5B3000625AD, 0x0001488680009FB3,
-    0x80009FB700010C87, 0x0006B5B50006A9B6, 0x0006C9BB80009FBB,
-    0x80009FBF0006F9BA, 0x00011DBF80009FC3, 0x00011DBD00011DBE,
-    0x80009FC70006D1BC, 0x80009FCF80009FCB, 0x0005CCE380009FD3,
-    0x00053CE180009FD7, 0x0004B8EF80009FDB, 0x80009FE380009FDF,
-    0x80009FEB80009FE7, 0x80009FF380009FEF, 0x00061DA880009FF7,
-    0x00062DAA000625A9, 0x00063DAC000625AB, 0x00065DAE000629AD,
-    0x80009FFB00064DAF, 0x8000A00380009FFF, 0x8000A00B8000A007,
-    0x8000A0138000A00F, 0x0001F9B88000A017, 0x00014DBA0001F9B9,
-    0x00015DBC000145BB, 0x000145BE000145BD, 0x00063DB0000175BF,
-    0x000605B200060DB1, 0x0001D9B40001C9B3, 0x0001C9B60001D9B5,
-    0x0003B9800001C9B7, 0x000019820003B981, 0x8000A01B0005A9A3,
-    0x0005A5A60005B9A5, 0x8000A02300028484, 0x0005F5AA0003A8BE,
-    0x0005DDAC0005C5AB, 0x000211AE000211AD, 0x000301B6000211AF,
-    0x000344848000A027, 0x8000A02B000361B5, 0x0003D887000349B3,
-    0x000309BE001CF886, 0x000335BC000309BF, 0x00033DBA000309BD,
-    0x8000A02F000321BB, 0x0003CDB08000A033, 0x0003D5B20003D1B1,
-    0x0003F9B40003E9B3, 0x0003E9B60003F9B5, 0x0003D9B80003E9B7,
-    0x000309BA0003D9B9, 0x000319BC000309BB, 0x000309BE000319BD,
-    0x8000A037000309BF, 0x000000B38000A03B, 0x8000A0438000A03F,
-    0x8000A04B8000A047, 0x0003E5A880009ECF, 0x0003F5AA0003F5A9,
-    0x0003DDAC0003C5AB, 0x0003BDAE0003B5AD, 0x00021DA30003B5AF,
-    0x8000A0538000A04F, 0x001D1CBE8000A057, 0x000235A5000255A6,
-    0x000275AB8000A05B, 0x8000A05F000269AA, 0x00025DAF8000A063,
-    0x00025DAD00025DAE, 0x8000A067000261AC, 0x8000A06F8000A06B,
-    0x00038CEF8000A073, 0x8000A07B8000A077, 0x00036CE38000A07F,
-    0x000078E18000A083, 0x8000A08B8000A087, 0x001C58868000A093,
-    0x0000E580001DF087, 0x0000F9820000E981, 0x8000A09B8000A097,
-    0x8000A0A38000A09F, 0x8000A0A70006D4EF, 0x8000A0AB000674E1,
-    0x8000A0AF000154E3, 0x8000A0B78000A0B3, 0x000191B38000A0BB,
-    0x8000A0C38000A0BF, 0x8000A0C7001DC484, 0x0001A5B5000115B6,
-    0x000135BB8000A0CB, 0x8000A0CF000135BA, 0x0000FDBF8000A0D3,
-    0x0000FDBD0000FDBE, 0x001E45A9000125BC, 0x001E71AB001E25A8,
-    0x001ED1AD001E5DAA, 0x001ED5AF001E69AC, 0x8000A08F001ED9AE,
-    0x8000A0DB8000A0D7, 0x8000A0E38000A0DF, 0x8000A0EB8000A0E7,
-    0x001F01B98000A0EF, 0x001F01BB001F69B8, 0x001F05BD001F09BA,
-    0x001F05BF001F19BC, 0x001EB5B1001F05BE, 0x001F6DB3001EB5B0,
-    0x001F7DB5001EB5B2, 0x001F51B7001F7DB4, 0x001ED1A3001F61B6,
-    0x0000158100002D82, 0x8000A0F300001D80, 0x001EE5A5001E55A6,
-    0x001E75AB8000A0F7, 0x8000A0FB001E75AA, 0x001FBDAF000360BE,
-    0x001FBDAD001FBDAE, 0x8000A0FF001E65AC, 0x00006886001F55B3,
-    0x001F71B600036C87, 0x8000A1078000A103, 0x001F55BA001F45B5,
-    0x8000A10B001F5DBB, 0x001F39BE8000A10F, 0x001F45BC001F25BF,
-    0x8000A113001F39BD, 0x8000A11B8000A117, 0x8000A1238000A11F,
-    0x8000A12B8000A127, 0x8000A12F001CF4EF, 0x8000A133001D60E1,
-    0x8000A137001D74E3, 0x8000A13F8000A13B, 0x001ED1A38000A143,
-    0x8000A14B8000A147, 0x8000A1538000A14F, 0x001EC1A5001EF5A6,
-    0x001ED9AB8000A157, 0x8000A15B001ED1AA, 0x001EA1AF8000A15F,
-    0x001EBDAD001EBDAE, 0x001EF5A8001EC1AC, 0x001E19AA001E09A9,
-    0x001E09AC001E19AB, 0x001E31AE001E35AD, 0x00000D80001E2DAF,
-    0x00001D8200001581, 0x8000A1678000A163, 0x0001FC848000A16B,
-    0x001E2DB80001F0BE, 0x0001E1BA0001E1B9, 0x0001E1BC0001E1BB,
-    0x0001E1BE0001E1BD, 0x001E55B00001E1BF, 0x001E55B2001E5DB1,
-    0x001E39B4001E29B3, 0x001E25B6001E25B5, 0x0002D5A8001E15B7,
-    0x0002E1AA0002DDA9, 0x000311AC0002E1AB, 0x000311AE000311AD,
-    0x8000A173000311AF, 0x8000A17B8000A177, 0x8000A1838000A17F,
-    0x8000A18B8000A187, 0x00037DB88000A18F, 0x000301BA000301B9,
-    0x000301BC000301BB, 0x000331BE000309BD, 0x000371B0000331BF,
-    0x000371B2000371B1, 0x000359B400034DB3, 0x00034DB6000345B5,
-    0x00042086000345B7, 0x000261B30003A087, 0x000391B5000480BE,
-    0x8000A1978000A193, 0x8000A01F000399B6, 0x0003BDBB0003B884,
-    0x000391BD0003BDBA, 0x000389BF0003ADBC, 0x000229A3000395BE,
-    0x8000A19F8000A19B, 0x000694858000A1A3, 0x0003D9A50003D1A6,
-    0x0003F5AB8000A1A7, 0x8000A1AB0003F5AA, 0x0003C1AF8000A1AF,
-    0x0003D9AD0003DDAE, 0x8000A1B30003E5AC, 0x8000A1BB8000A1B7,
-    0x8000A1C38000A1BF, 0x8000A1CB8000A1C7, 0x000011818000A1CF,
-    0x8000A1D300000D80, 0x8000A1D700001982, 0x8000A1E38000A1DB,
-    0x8000A1EB8000A1E7, 0x8000A1EF000194E1, 0x8000A1F30000E4E3,
-    0x000478868000A1F7, 0x8000A1FB00028087, 0x8000A2038000A1FF,
-    0x8000A20B8000A207, 0x8000A2138000A20F, 0x8000A217000038EF,
-    0x8000A21F8000A21B, 0x8000A2278000A223, 0x8000A22F8000A22B,
-    0x00046CBE0001C8EF, 0x0005A084001EC4E1, 0x8000A233001E70E3,
-    0x8000A23B8000A237, 0x8000A2438000A23F, 0x0001F9B50001F1B6,
-    0x0001F9B30020F485, 0x0000218100002982, 0x00013DBF00001D80,
-    0x000121BD000129BE, 0x000169BB000129BC, 0x8000A247000161BA,
-    0x000621A88000A24B, 0x000621AA000621A9, 0x000621AC000621AB,
-    0x000621AE000621AD, 0x8000A1DF000621AF, 0x000C008600012CBF,
-    0x8000A24F00036087, 0x8000A2578000A253, 0x000741B88000A25B,
-    0x000741BA000741B9, 0x000741BC000741BB, 0x000741BE000741BD,
-    0x0006F5B0000741BF, 0x0006ADB20006FDB1, 0x0006BDB40006A5B3,
-    0x000691B60006A9B5, 0x0006A9A3000691B7, 0x8000A2638000A25F,
-    0x8000A26B8000A267, 0x0006A9A50006A1A6, 0x000639AB8000A26F,
-    0x8000A273000631AA, 0x00066DAF8000A277, 0x000671AD000679AE,
-    0x8000A27B000679AC, 0x8000A27F000785B3, 0x000785B68000A283,
-    0x8000A28B8000A287, 0x000789BA00078DB5, 0x8000A28F000789BB,
-    0x000789BE8000A293, 0x000789BC000775BF, 0x8000A297000789BD,
-    0x00057CEF0005C0E1, 0x8000A29B0004D4E3, 0x8000A2A38000A29F,
-    0x8000A2AB8000A2A7, 0x8000A2B38000A2AF, 0x0005A8E18000A2B7,
-    0x00058CE38000A2BB, 0x000601A3000CF8EF, 0x0000658100007182,
-    0x0001808400006580, 0x000609A5000601A6, 0x00060DAB000010BE,
-    0x000FE88600060DAA, 0x0006F1AF00001887, 0x00060DAD00060DAE,
-    0x0006B9A800060DAC, 0x0006C9AA0006B9A9, 0x0006C1AC0006C5AB,
-    0x0006F9AE0006C1AD, 0x8000A2C30006F9AF, 0x8000A2CB8000A2C7,
-    0x8000A2D38000A2CF, 0x8000A2DB8000A2D7, 0x00064DB88000A2DF,
-    0x00065DBA000655B9, 0x000661BC000655BB, 0x000661BE000661BD,
-    0x000689B0000661BF, 0x000699B2000689B1, 0x000689B4000699B3,
-    0x00067DB6000675B5, 0x8000A2E3000675B7, 0x000285B600028DB5,
-    0x0068F18F000C4084, 0x8000A2EB8000A2E7, 0x0002E1BC000221B3,
-    0x0002E5BE0002EDBD, 0x002634850002F9BF, 0x000291BA8000A2EF,
-    0x001F019B000291BB, 0x001E4D99001E4D9A, 0x0004F19F00133D98,
-    0x001A799D001AB99E, 0x001641930018819C, 0x006AF99100149192,
-    0x00126197006A4D90, 0x0010819500127196, 0x003801B100164194,
-    0x003A81B3003EE1B0, 0x002691B5003A89B2, 0x8000A2F3003BD1B4,
-    0x000281A38000A2F7, 0x000669A1000011A2, 0x8000A2FB00068DA0,
-    0x8000A2FF8000A16F, 0x0002F1A18000A303, 0x000ECDA38000A307,
-    0x000801A5000ECDA2, 0x000BC5A7000F51A4, 0x0036C1A9000AFDA6,
-    0x003001AB003401A8, 0x0032CDAD003641AA, 0x003EFDAF0032D5AC,
-    0x0069E182003C01AE, 0x8000A30B006B2583, 0x001731868000A30F,
-    0x006B898400172987, 0x0011D98A0015E985, 0x8000A3130013218B,
-    0x001F358E8000A317, 0x0013298C001F3D8F, 0x001E019200124D8D,
-    0x8000A31B00199993, 0x0005F196000C0084, 0x001B799400077597,
-    0x0007A19A001BB195, 0x8000A31F0001D19B, 0x0000F1818000A323,
-    0x0003599C00001980, 0x8000A3270000F982, 0x000DC0BE8000A32B,
-    0x000C3086000028EF, 0x8000A333000D5C87, 0x8000A3370006C4EF,
-    0x8000A33B0006E4E1, 0x8000A33F000614E3, 0x8000A343000184E1,
-    0x0003FDA30000FCE3, 0x8000A34B8000A347, 0x8000A34F000F4C84,
-    0x000351A5000359A6, 0x00034DAB8000A353, 0x8000A35700034DAA,
-    0x000325AF8000A35B, 0x000331AD000339AE, 0x8000A2BF00033DAC,
-    0x8000A35F8000A32F, 0x8000A3678000A363, 0x8000A36F8000A36B,
-    0x000D81A88000A373, 0x000D81AA000D81A9, 0x000D81AC000D81AB,
-    0x000D81AE000D81AD, 0x000E41B0000D81AF, 0x000E51B2000E51B1,
-    0x000EC9B4000E79B3, 0x000EE9B6000EC9B5, 0x000EDDB8000EE5B7,
-    0x000F61BA000EA1B9, 0x000F59BC000F61BB, 0x000F31BE000F59BD,
-    0x000D29B3000F15BF, 0x8000A37B8000A377, 0x8000A3838000A37F,
-    0x000D29B5000D21B6, 0x000EB5BB8000A387, 0x8000A38B000EA5BA,
-    0x000E2DBF8000A38F, 0x000E8DBD000E2DBE, 0x00002982000EA5BC,
-    0x00001D80000D6DA3, 0x000D65A600001581, 0x8000A3978000A393,
-    0x000EE1AA000D6DA5, 0x8000A39B000EF1AB, 0x000E69AE0003A4BE,
-    0x000EE1AC000E69AF, 0x000E99A8000EC9AD, 0x000EA1AA000EA5A9,
-    0x000EA9AC000EBDAB, 0x000ED1AE000ED1AD, 0x00008086000ECDAF,
-    0x8000A3A300010487, 0x8000A3AB8000A3A7, 0x8000A3B38000A3AF,
-    0x00015DB88000A3B7, 0x0001C1BA0001C1B9, 0x0001C1BC0001C1BB,
-    0x0001F1BE0001C9BD, 0x000EBDB00001F1BF, 0x000E85B2000E8DB1,
-    0x00017DB4000165B3, 0x00016DB6000165B5, 0x000E29B3000165B7,
-    0x8000A3BB8000A39F, 0x8000A3C38000A3BF, 0x000E21B5000E51B6,
-    0x000E49BB8000A3C7, 0x8000A3CB000E79BA, 0x0001A1BF8000A3CF,
-    0x0001B1BD0001B9BE, 0x8000A3D3000E51BC, 0x8000A3D7000E6DA3,
-    0x000E15A68000A3DB, 0x8000A3E38000A3DF, 0x000E3DAA000E65A5,
-    0x8000A3E7000E0DAB, 0x0001FDAE8000A3EB, 0x000E15AC0001E5AF,
-    0x000138E10001F5AD, 0x000FD4E38000A3EF, 0x8000A3F78000A3F3,
-    0x8000A3FB000358BE, 0x00000D808000A3FF, 0x00001D8200001581,
-    0x8000A4078000A403, 0x0002D4EF8000A40B, 0x0004E0868000A413,
-    0x000400BE0003E087, 0x8000A41700037DB3, 0x000365B600036DB5,
-    0x8000A41F8000A41B, 0x0003F9BA8000A423, 0x0003D9BC0003F1BB,
-    0x0003B9BE0003D9BD, 0x000EBDAA0003B5BF, 0x000EBDA8000E01AB,
-    0x000E01AE000EB5A9, 0x000E11AC000E09AF, 0x8000A427000E11AD,
-    0x8000A42F8000A42B, 0x8000A4378000A433, 0x8000A43F8000A43B,
-    0x000FD1BA8000A443, 0x000E25B8000FD1BB, 0x000FF1BE000FCDB9,
-    0x000FC1BC000FF1BF, 0x000E21B2000FC1BD, 0x000E41B0000E2DB3,
-    0x000E29B6000E41B1, 0x000E35B4000E1DB7, 0x000229A8000E29B5,
-    0x000239AA000229A9, 0x000229AC000239AB, 0x000291AE000229AD,
-    0x8000A447000291AF, 0x8000A44F8000A44B, 0x8000A4578000A453,
-    0x002B30BF8000A45B, 0x000299B88000A45F, 0x0002A5BA0002ADB9,
-    0x00015DBC000149BB, 0x000141BE000145BD, 0x0002F1B0000141BF,
-    0x0002C1B20002F9B1, 0x0002B1B40002C1B3, 0x0002B5B60002BDB5,
-    0x000E08E10002A9B7, 0x000ED4E3000CF4E1, 0x8000A463000F90E3,
-    0x8000A46B8000A467, 0x000011818000A46F, 0x8000A47300000D80,
-    0x8000A47700001982, 0x0001ACEF8000A47B, 0x0002F1A3000ECCEF,
-    0x8000A4878000A483, 0x8000A48B00058084, 0x0002E1A50002E9A6,
-    0x00027DAB8000A48F, 0x0004C886000275AA, 0x000239AF0004EC87,
-    0x000255AD000235AE, 0x00063DA8000255AC, 0x000655AA000655A9,
-    0x000675AC00066DAB, 0x0006BDAE0006B5AD, 0x8000A47F0006B5AF,
-    0x8000A4978000A493, 0x8000A49F8000A49B, 0x8000A4A78000A4A3,
-    0x00076DB88000A4AB, 0x00077DBA000775B9, 0x0007C1BC000771BB,
-    0x0007C1BE0007C1BD, 0x0006CDB00007C1BF, 0x0006DDB20006D5B1,
-    0x0006CDB40006D5B3, 0x00075DB6000755B5, 0x0006DDB3000755B7,
-    0x8000A4B38000A4AF, 0x8000A4BB8000A4B7, 0x0006FDB5000635B6,
-    0x000611BB8000A4BF, 0x8000A4C3000609BA, 0x000791BF8000A4C7,
-    0x000601BD000609BE, 0x8000A4CB000609BC, 0x8000A4CF000699A3,
-    0x000671A68000A4D3, 0x8000A4DB8000A4D7, 0x00064DAA0006B9A5,
-    0x8000A4DF000655AB, 0x00064DAE8000A4E3, 0x00064DAC0007D5AF,
-    0x0001ED80000645AD, 0x0000198200000981, 0x8000A4EB8000A4E7,
-    0x0001C4BE0001D084, 0x000080868000A4EF, 0x8000A4F300012087,
-    0x8000A4FB8000A4F7, 0x8000A5038000A4FF, 0x8000A50B8000A507,
-    0x8000A5138000A50F, 0x8000A51B8000A517, 0x8000A51F00050CE1,
-    0x8000A523000480E3, 0x8000A52B8000A527, 0x8000A52F0005DCEF,
-    0x8000A5378000A533, 0x000645A88000A53B, 0x000659AA00064DA9,
-    0x0001B5AC000651AB, 0x0001B5AE0001BDAD, 0x0001AC840001ADAF,
-    0x8000A5438000A53F, 0x8000A54B8000A547, 0x8000A5538000A54F,
-    0x000169B88000A557, 0x000109BA000169B9, 0x00011DBC000105BB,
-    0x000101BE000101BD, 0x0001D5B0000135BF, 0x0001D5B20001DDB1,
-    0x000179B4000169B3, 0x000169B6000179B5, 0x0003BD81000161B7,
-    0x0005D9B30003BD80, 0x000235B500001982, 0x000344BE8000A55B,
-    0x8000A563000235B6, 0x000211BB8000A567, 0x0002B5BD000219BA,
-    0x0002B5BF000209BC, 0x8000A56B0002B5BE, 0x00026DA500026DA6,
-    0x000581A3003C0084, 0x003CF8868000A56F, 0x0002EDAF00037C87,
-    0x0002EDAD0002EDAE, 0x000249AB000251AC, 0x8000A573000241AA,
-    0x8000A57B8000A577, 0x8000A5838000A57F, 0x000120E38000A587,
-    0x0001A0E18000A58B, 0x000240EF8000A58F, 0x8000A5978000A593,
-    0x8000A59F8000A59B, 0x8000A5A78000A5A3, 0x0003F1B38000A5AB,
-    0x8000A5B38000A5AF, 0x8000A40F8000A5B7, 0x0003E1B50003E9B6,
-    0x0003ADBB8000A5BB, 0x8000A5BF0003ADBA, 0x000395BF8000A5C3,
-    0x0003A1BD000395BE, 0x000229A80003ADBC, 0x000239AA000229A9,
-    0x000229AC000239AB, 0x000259AE000229AD, 0x0001CD80000259AF,
-    0x0000198200000981, 0x8000A5CB8000A5C7, 0x003DB487003D7CBE,
-    0x000235B8003C1C86, 0x000235BA00023DB9, 0x0002F9BC0002E9BB,
-    0x0002E9BE0002F9BD, 0x000229B00002E9BF, 0x000239B2000229B1,
-    0x000211B4000235B3, 0x000211B6000211B5, 0x8000A5D300020DB7,
-    0x8000A5DB8000A5D7, 0x8000A5E38000A5DF, 0x8000A5EB8000A5E7,
-    0x8000A5EF00007CEF, 0x8000A5F3000654E1, 0x000000EC0000A8E3,
-    0x8000A5FB8000A5F7, 0x8000A6038000A5FF, 0x0002F1A50002F9A6,
-    0x0002E1A38000A607, 0x8000A60F8000A60B, 0x000285AF8000A613,
-    0x0002B1AD000285AE, 0x0002BDAB0002BDAC, 0x8000A6170002BDAA,
-    0x003E21A88000A61B, 0x003E49AA003E21A9, 0x003E69AC003E59AB,
-    0x003E99AE003E69AD, 0x8000A5CF003E99AF, 0x8000A6238000A61F,
-    0x8000A62B8000A627, 0x8000A6338000A62F, 0x003EF5B88000A637,
-    0x003EF5BA003EFDB9, 0x003E99BC003E89BB, 0x003E85BE003E99BD,
-    0x003EE9B0003EB5BF, 0x003EFDB2003EE9B1, 0x003ED5B4003EF1B3,
-    0x003ED5B6003EDDB5, 0x003EC1B3003ECDB7, 0x0000B58100001582,
-    0x8000A63B0000B580, 0x003EE1B5003EE9B6, 0x003E2DBB000010BE,
-    0x00036484003E2DBA, 0x003E19BF8000A63F, 0x003E25BD003E11BE,
-    0x8000A643003E29BC, 0x0000A886003E85A3, 0x003EADA600036C87,
-    0x8000A64B8000A647, 0x003E69AA003EA5A5, 0x8000A64F003E69AB,
-    0x003E55AE8000A653, 0x003E6DAC003E5DAF, 0x003E89B3003E61AD,
-    0x8000A65B8000A657, 0x8000A6638000A65F, 0x003E89B5003E81B6,
-    0x003E95BB8000A667, 0x8000A66B003E8DBA, 0x003EFDBF8000A66F,
-    0x003E85BD003EF5BE, 0x8000A673003E85BC, 0x8000A67B8000A677,
-    0x8000A6838000A67F, 0x003C94EF8000A687, 0x003D20E18000A68B,
-    0x003D34E38000A68F, 0x8000A6978000A693, 0x8000A69F8000A69B,
-    0x003E01A88000A6A3, 0x003E01AA003E01A9, 0x003E01AC003E01AB,
-    0x003E01AE003E01AD, 0x0000DD80003E3DAF, 0x0000198200000981,
-    0x8000A6AB8000A6A7, 0x000074BE00013C84, 0x003E1DB88000A6B3,
-    0x003E35BA003E3DB9, 0x0001F9BC0001E9BB, 0x0001E9BE0001F9BD,
-    0x003E49B00001E9BF, 0x003E51B2003E51B1, 0x003E31B4003E51B3,
-    0x003E2DB6003E39B5, 0x003D89A3003E25B7, 0x0004A8868000A6B7,
-    0x8000A6BB00033C87, 0x003D89A5003D81A6, 0x003D95AB8000A6BF,
-    0x8000A6C3003D8DAA, 0x003DFDAF8000A6C7, 0x003D85AD003DF5AE,
-    0x8000A6CB003D85AC, 0x0003488400029DB3, 0x000355B6000440BE,
-    0x8000A6D38000A6CF, 0x000371BA0002B1B5, 0x8000A6D7000379BB,
-    0x000331BE8000A6DB, 0x000351BC000331BF, 0x0003D1B8000351BD,
-    0x0003D5BA0003D1B9, 0x0003FDBC0003E9BB, 0x0003EDBE0003E5BD,
-    0x0003B1B00003E5BF, 0x000381B20003B9B1, 0x0003F1B4000381B3,
-    0x0003F1B60003F1B5, 0x000201A80003F1B7, 0x000211AA000211A9,
-    0x00023DAC000225AB, 0x0003D1AE0003D1AD, 0x8000A6DF0003D1AF,
-    0x8000A6E38000A55F, 0x8000A6EB8000A6E7, 0x8000A6F38000A6EF,
-    0x8000A6FB8000A6F7, 0x8000A7038000A6FF, 0x000050EF8000A707,
-    0x8000A70F8000A70B, 0x000084E38000A713, 0x0001F8E18000A717,
-    0x00002D808000A71B, 0x0000158200001181, 0x000399A38000A723,
-    0x8000A72B8000A727, 0x8000A7338000A72F, 0x0003B5A5000251A6,
-    0x00027DAB8000A737, 0x8000A73B000275AA, 0x000235AF8000A73F,
-    0x000255AD000235AE, 0x0004A086000255AC, 0x8000A7430005A887,
-    0x8000A74B8000A747, 0x8000A7538000A74F, 0x00042CBE0001F8EF,
-    0x8000A757003E80E1, 0x8000A75B000114E3, 0x8000A7638000A75F,
-    0x8000A76B8000A767, 0x8000A7738000A76F, 0x8000A7770001DDB3,
-    0x0001F5B60001FDB5, 0x8000A77F8000A77B, 0x0001B1BA8000A783,
-    0x00019DBC000185BB, 0x000111BE000115BD, 0x0005BDA8000111BF,
-    0x000619AA000601A9, 0x000639AC000611AB, 0x000659AE000639AD,
-    0x8000A71F000649AF, 0x0000198100001182, 0x8000A7870007C980,
-    0x000230848000A78B, 0x0006B9B88000A78F, 0x000749BA0006B9B9,
-    0x000759BC000749BB, 0x00074DBE000759BD, 0x000639B0000745BF,
-    0x000605B200060DB1, 0x000699B400068DB3, 0x000689B6000699B5,
-    0x000691A3000689B7, 0x000028868000A793, 0x8000A7970003FC87,
-    0x0006B1A50006B9A6, 0x0006C9AB8000A79B, 0x8000A79F0006FDAA,
-    0x00065DAF8000A7A3, 0x000659AD00065DAE, 0x8000A7A70006D1AC,
-    0x8000A7AB00078DB3, 0x00078DB68000A7AF, 0x8000A7B78000A7B3,
-    0x000769BA00078DB5, 0x8000A7BB000769BB, 0x000769BE8000A7BF,
-    0x000769BC00075DBF, 0x000649A3000769BD, 0x8000A7C78000A7C3,
-    0x8000A7CF8000A7CB, 0x000649A5000649A6, 0x0006ADAB8000A7D3,
-    0x8000A7D70006ADAA, 0x000699AF8000A7DB, 0x0006ADAD0006ADAE,
-    0x8000A7DF0006ADAC, 0x8000A7E78000A7E3, 0x8000A7EF8000A7EB,
-    0x8000A7F30005E0EF, 0x8000A7FB8000A7F7, 0x0001248400053CE3,
-    0x0000298000057CE1, 0x0000198200002981, 0x0006BDA8000134BE,
-    0x0006C1AA0006C1A9, 0x0006C1AC0006C1AB, 0x0006F1AE0006C9AD,
-    0x000000860006F1AF, 0x8000A80300008487, 0x8000A80B8000A807,
-    0x8000A8138000A80F, 0x000679B88000A817, 0x000609BA000679B9,
-    0x000619BC000609BB, 0x000605BE000605BD, 0x000669B000063DBF,
-    0x000679B2000669B1, 0x000669B4000679B3, 0x000649B600065DB5,
-    0x000285A8000649B7, 0x000295AA000295A9, 0x0002BDAC0002A5AB,
-    0x0002D1AE0002D5AD, 0x8000A81B0002D1AF, 0x000328848000A81F,
-    0x8000A82300E904BF, 0x8000A82B8000A827, 0x00035DB88000A82F,
-    0x000365BA00036DB9, 0x000315BC00030DBB, 0x000315BE00031DBD,
-    0x0002B5B000030DBF, 0x000281B20002BDB1, 0x000375B4000281B3,
-    0x000375B600037DB5, 0x00021DB300036DB7, 0x8000A8378000A833,
-    0x000C1CBE8000A83B, 0x000235B5000255B6, 0x000279BB8000A83F,
-    0x8000A843000271BA, 0x0003B9BF8000A847, 0x000255BD0003B1BE,
-    0x8000A84B000255BC, 0x8000A84F000259A3, 0x000211A68000A853,
-    0x8000A85B8000A857, 0x000235AA000271A5, 0x8000A6AF00023DAB,
-    0x0003F5AE8000A85F, 0x000211AC0003FDAF, 0x0002A9A8000211AD,
-    0x0002B9AA0002A9A9, 0x0002A9AC0002B9AB, 0x000139AE0002A9AD,
-    0x00000D80000139AF, 0x00001D8200001581, 0x8000A8678000A863,
-    0x000D68878000A86B, 0x0001E9B8000D1C86, 0x000189BA0001E9B9,
-    0x00019DBC000185BB, 0x000181BE000181BD, 0x000151B00001B1BF,
-    0x000151B2000151B1, 0x0001F5B4000151B3, 0x0001F5B60001FDB5,
-    0x000004EF0001E1B7, 0x8000A873000C4884, 0x8000A877000C80BE,
-    0x8000A87F8000A87B, 0x8000A8830006C4EF, 0x8000A88700063CE1,
-    0x0000C8E3000184E3, 0x000128E18000A88B, 0x8000A8938000A88F,
-    0x8000A89B8000A897, 0x8000A89F00027DB3, 0x000265B600026DB5,
-    0x8000A8A78000A8A3, 0x0001D9BA8000A8AB, 0x0001C9BC0001D9BB,
-    0x00017DBE0001C9BD, 0x000DB1A3000175BF, 0x8000A8AF8000A86F,
-    0x8000A8B78000A8B3, 0x000DA1A5000DA9A6, 0x000E15AB8000A8BB,
-    0x8000A8BF000E15AA, 0x000EB9AF8000A8C3, 0x000E05AD000EB1AE,
-    0x8000A8C7000E05AC, 0x8000A8CF8000A8CB, 0x00001D808000A8D3,
-    0x0000398200000981, 0x8000A8DB8000A8D7, 0x0000C8868000A8DF,
-    0x8000A8E300030C87, 0x8000A8EB8000A8E7, 0x000F09A88000A8EF,
-    0x000F1DAA000F15A9, 0x000F31AC000F15AB, 0x000F2DAE000F3DAD,
-    0x0003A884000F21AF, 0x8000A8F78000A8F3, 0x8000A8FF8000A8FB,
-    0x8000A9078000A903, 0x0000CDB88000A90B, 0x0000D5BA0000D5B9,
-    0x0000FDBC0000E5BB, 0x00009DBE000095BD, 0x000F21B0000095BF,
-    0x000F21B2000F21B1, 0x000F2DB4000F3DB3, 0x000F1DB6000F15B5,
-    0x000E95B30000F5B7, 0x8000A9138000A90F, 0x8000A91B8000A917,
-    0x000EB9B5000EBDB6, 0x000EE5BB8000A91F, 0x8000A923000EE5BA,
-    0x000ED1BF8000A927, 0x000EE9BD000EE9BE, 0x000D28EF000EF5BC,
-    0x8000A92F8000A92B, 0x8000A9378000A933, 0x8000A93F8000A93B,
-    0x00003D808000A943, 0x00001D8200000981, 0x000D48E38000A947,
-    0x000EC0E18000A94B, 0x000E15A300017CBE, 0x000028868000A953,
-    0x8000A95700014C87, 0x000E39A5000E3DA6, 0x000E65AB8000A95B,
-    0x8000A95F000E65AA, 0x000E51AF8000A963, 0x000E69AD000E69AE,
-    0x8000A967000E75AC, 0x8000A96B0001BDB3, 0x000179B68000A96F,
-    0x8000A9778000A973, 0x000155BA000179B5, 0x8000A97B00015DBB,
-    0x0000F9BE8000A97F, 0x000145BC0000F9BF, 0x000271A80000F9BD,
-    0x000271AA000271A9, 0x0002B5AC000271AB, 0x0002B5AE0002BDAD,
-    0x0004EC840002ADAF, 0x8000A9878000A983, 0x8000A98F8000A98B,
-    0x8000A9978000A993, 0x000369B88000A99B, 0x00030DBA000369B9,
-    0x00031DBC000305BB, 0x00030DBE000305BD, 0x0002D5B0000305BF,
-    0x0002D5B20002DDB1, 0x000379B4000369B3, 0x000369B6000379B5,
-    0x8000A99F000361B7, 0x8000A9A78000A9A3, 0x8000A9AB0002F5A3,
-    0x000231A6000231A5, 0x8000A9B38000A9AF, 0x00021DAA8000A9B7,
-    0x00020DAC000215AB, 0x0003B1AE0003B1AD, 0x000061800003B1AF,
-    0x0000058200006181, 0x0004F0868000A9BB, 0x000410BE00036087,
-    0x8000A9C78000A9C3, 0x8000A9CF8000A9CB, 0x8000A9D78000A9D3,
-    0x8000A9DF8000A9DB, 0x000285A88000A7FF, 0x000295AA000295A9,
-    0x0002BDAC0002A5AB, 0x0002D1AE0002D5AD, 0x8000A9E30002D1AF,
-    0x8000A9EB8000A9E7, 0x8000A9F38000A9EF, 0x8000A9FB8000A9F7,
-    0x000175B88000A9FF, 0x000175BA00017DB9, 0x0001D5BC0001CDBB,
-    0x0001C9BE0001DDBD, 0x0002B5B00001C1BF, 0x000281B20002BDB1,
-    0x000155B4000281B3, 0x000155B600015DB5, 0x000EA4E100014DB7,
-    0x000E1CE300040884, 0x00049CBE8000AA03, 0x8000AA0B8000AA07,
-    0x8000AA138000AA0F, 0x8000AA1B8000AA17, 0x000075818000AA1F,
-    0x000ECCEF00007580, 0x8000AA2300007582, 0x8000AA2B8000AA27,
-    0x8000AA330004C484, 0x00005CEF8000AA37, 0x000190E18000AA3B,
-    0x000158E38000AA3F, 0x8000AA478000AA43, 0x000758868000AA4B,
-    0x00059DB30004EC87, 0x8000AA2F8000A9BF, 0x8000AA538000AA4F,
-    0x0005B5B50005D5B6, 0x0005F9BB8000AA57, 0x8000AA5B0005F1BA,
-    0x000619BF8000AA5F, 0x0005D5BD000611BE, 0x8000AA630005D5BC,
-    0x8000AA670005D9A3, 0x000591A68000AA6B, 0x8000AA738000AA6F,
-    0x0005B5AA0005F1A5, 0x8000AA770005BDAB, 0x000655AE8000AA7B,
-    0x000591AC00065DAF, 0x00065DA8000591AD, 0x00066DAA000665A9,
-    0x00067DAC000665AB, 0x00066DAE000665AD, 0x0000A0BE000665AF,
-    0x8000AA838000AA7F, 0x8000AA8B8000AA87, 0x8000AA938000AA8F,
-    0x00074DB88000AA97, 0x000751BA000751B9, 0x000771BC000751BB,
-    0x000771BE000771BD, 0x0007C1B0000771BF, 0x0007C1B20007C1B1,
-    0x0007C1B40007C1B3, 0x0007C1B60007C1B5, 0x8000AA9B0007C1B7,
-    0x8000AAA3000694E1, 0x0001A98000053CE3, 0x000019820001A981,
-    0x000080868000AAA7, 0x8000AAAB00010487, 0x8000AAB38000AAAF,
-    0x8000AABB8000AAB7, 0x0007A5B300051CEF, 0x8000AABF8000AA9F,
-    0x8000AAC78000AAC3, 0x0007B5B50007A5B6, 0x000645BB8000AACB,
-    0x8000AACF000645BA, 0x000645BF8000AAD3, 0x000655BD000645BE,
-    0x000765A3000655BC, 0x8000AADB8000AAD7, 0x8000AAE38000AADF,
-    0x000775A5000765A6, 0x000685AB8000AAE7, 0x8000AAEB000685AA,
-    0x000685AF8000AAEF, 0x000695AD000685AE, 0x8000AAF3000695AC,
-    0x8000AAF7000619B3, 0x000621B68000AAFB, 0x8000AB038000AAFF,
-    0x0001F1BA000631B5, 0x8000AB070001F9BB, 0x0001B1BE8000AB0B,
-    0x0001D1BC0001BDBF, 0x0005D1A80001D1BD, 0x0005D5AA0005D1A9,
-    0x0005FDAC0005E9AB, 0x0005EDAE0005E5AD, 0x0003F980000211AF,
-    0x0003C5820003CD81, 0x001C00BE00032084, 0x000304878000AB13,
-    0x000219B8001CBC86, 0x000225BA00022DB9, 0x0002F9BC0002E9BB,
-    0x0002E9BE0002F9BD, 0x000271B00002E9BF, 0x000241B2000279B1,
-    0x000231B4000241B3, 0x000235B600023DB5, 0x8000A94F000229B7,
-    0x8000AB1B8000AB17, 0x8000AB1F000551A3, 0x000569A6000579A5,
-    0x8000AB278000AB23, 0x0002B9AA8000AB2B, 0x000299AC0002B1AB,
-    0x0002F9AE000299AD, 0x0001F8E10002F5AF, 0x0002F4E38000AB2F,
-    0x8000AB33001CC084, 0x8000AB3B8000AB37, 0x8000AB438000AB3F,
-    0x8000AB4B8000AB47, 0x8000AB538000AB4F, 0x000230EF8000AB57,
-    0x000209A88000AB5B, 0x000221AA000219A9, 0x000221AC000221AB,
-    0x000221AE000221AD, 0x8000AB5F000255AF, 0x8000AB678000AB63,
-    0x8000AB6F8000AB6B, 0x0000AD810000AD82, 0x0002D1B80000AD80,
-    0x0002E1BA0002D9B9, 0x000291BC0002E1BB, 0x000295BE00029DBD,
-    0x00022DB0000289BF, 0x000235B2000235B1, 0x00021DB4000205B3,
-    0x0002F1B60002F1B5, 0x000748E10002F1B7, 0x00061CE3000658E3,
-    0x8000AB7300066CE1, 0x8000AB7B8000AB77, 0x001E0086001C10BE,
-    0x8000AB83001C2487, 0x8000AB8B8000AB87, 0x0006D0EF0001ECEF,
-    0x0002E9B38000AB8F, 0x8000AB978000AB93, 0x8000AB9F8000AB9B,
-    0x0002E9B50002E1B6, 0x0002B1BB8000ABA3, 0x8000ABA70002A9BA,
-    0x000275BF8000ABAB, 0x000291BD000299BE, 0x8000AB0F0002A1BC,
-    0x00030484000148BE, 0x8000ABAF8000AB7F, 0x8000ABB78000ABB3,
-    0x001E15A88000ABBB, 0x001EA1AA001EA1A9, 0x001EADAC001EBDAB,
-    0x001E9DAE001E95AD, 0x001E91B0001EF9AF, 0x001E91B2001E91B1,
-    0x001EB1B4001E91B3, 0x001EB1B6001EB1B5, 0x001E91B8001EB1B7,
-    0x001E95BA001E91B9, 0x001F55BC001F4DBB, 0x001F55BE001F5DBD,
-    0x001DDDAC001F4DBF, 0x001DE5AE001DEDAD, 0x8000ABBF001D09AF,
-    0x001DD5AA8000ABC3, 0x8000ABC7001DCDAB, 0x001D9DA6001D95A5,
-    0x8000ABCF8000ABCB, 0x8000ABD78000ABD3, 0x001E49B3001D95A3,
-    0x8000ABDF8000ABDB, 0x8000ABE78000ABE3, 0x001E45B5001E75B6,
-    0x001E35BB8000ABEB, 0x00040086001E35BA, 0x001E15BF00000C87,
-    0x001E25BD001E15BE, 0x001F89A3001E25BC, 0x8000ABF30025358F,
-    0x8000ABFB8000ABF7, 0x001F85A5001FB5A6, 0x001FF5AB8000ABFF,
-    0x8000AC03001FF5AA, 0x001FD5AF8000AC07, 0x001FE5AD001FD5AE,
-    0x0019259E001FE5AC, 0x001BB59C001FC59F, 0x0015719A0019499D,
-    0x0011F5980015699B, 0x001385960017B599, 0x002D499400114597,
-    0x0028F592002C7195, 0x00297990002FA593, 0x00014D8000297991,
-    0x0001558200015581, 0x0025B4E18000AC0B, 0x000444E3000160EF,
-    0x00076C848000AC0F, 0x001578EF000214E3, 0x001CF5A200035CE1,
-    0x001DBDA00001B5A3, 0x001E71B3001DB5A1, 0x8000AC178000AC13,
-    0x8000AC1F8000AC1B, 0x001E1DB5001E1DB6, 0x001E09BB8000AC23,
-    0x8000AC27001E39BA, 0x0001E9BF8000AC2B, 0x0001E9BD0001E1BE,
-    0x00263182001E11BC, 0x8000AC2F002A0583, 0x002CC1860004C4BE,
-    0x002A1D84002E3587, 0x002F3D8A002B9985, 0x8000ABEF0010C18B,
-    0x0014D98E8000AC33, 0x0012198C0016158F, 0x00167D920012918D,
-    0x8000AC370018B993, 0x001CD1968000AC3B, 0x001A5D94001E5197,
-    0x001E699A001A3995, 0x8000AC3F001F219B, 0x8000AC478000AC43,
-    0x0002799C8000AC4B, 0x00028DA18000AC4F, 0x00068DA38000AC53,
-    0x000801A50000A5A2, 0x000AEDA7000685A4, 0x000EA5A9000A99A6,
-    0x000FEDAB000C01A8, 0x00F2B5AD000EA5AA, 0x00F401AF00F2B5AC,
-    0x00F619B100F3A1AE, 0x00FACDB300F6E5B0, 0x00FC01B500FACDB2,
-    0x8000AC5700FAC5B4, 0x000065808000AC5B, 0x0000358200003D81,
-    0x0004D0868000AC63, 0x8000AC670004BC87, 0x001CA9A800053084,
-    0x001CB9AA001CA9A9, 0x001CD5AC001CB1AB, 0x001CD5AE001CDDAD,
-    0x8000AC6B001CCDAF, 0x8000AC738000AC6F, 0x8000AC7B8000AC77,
-    0x0004F0BE8000AC7F, 0x0003D5B88000AC83, 0x0003E5BA0003E5B9,
-    0x0003E5BC0003FDBB, 0x00039DBE0003EDBD, 0x001CBDB0000395BF,
-    0x001C85B2001C8DB1, 0x000365B400037DB3, 0x000365B600036DB5,
-    0x000198E30003EDB7, 0x001E30E1001FA0E1, 0x8000AC87001F50E3,
-    0x8000AC8F8000AC8B, 0x8000AC978000AC93, 0x8000AC9F8000AC9B,
-    0x00014CEF8000ACA3, 0x8000ACAB8000ACA7, 0x001C71A3001ED8EF,
-    0x8000ACB38000ACAF, 0x8000ACBB8000ACB7, 0x001C1DA5001C1DA6,
-    0x001C09AB8000ACBF, 0x8000ACC3001C39AA, 0x0003E9AF8000ACC7,
-    0x0003E9AD0003E1AE, 0x000511B3001C11AC, 0x8000ACCB8000AC5F,
-    0x8000ACCF8000ACD3, 0x000531B5000539B6, 0x0005E9BB8000ACD7,
-    0x000308860005E1BA, 0x0005D9BF00038C87, 0x0005E9BD0005E9BE,
-    0x000049820005F1BC, 0x0007A180000555A3, 0x00057DA60007A181,
-    0x8000ACDF8000ACDB, 0x0005A5AA000575A5, 0x8000ACE30005ADAB,
-    0x0005ADAE8000ACE7, 0x0005B5AC00059DAF, 0x00066DA80005ADAD,
-    0x0007A9AA00067DA9, 0x0007B9AC0007A9AB, 0x0007A9AE0007B9AD,
-    0x8000ACEB0007A9AF, 0x8000ACF38000ACEF, 0x8000ACFB8000ACF7,
-    0x8000AD038000ACFF, 0x00076DB88000AD07, 0x000771BA000771B9,
-    0x000719BC000709BB, 0x000709BE000719BD, 0x0007DDB0000701BF,
-    0x00077DB2000775B1, 0x00076DB4000775B3, 0x00075DB6000755B5,
-    0x000621B3000755B7, 0x8000AD0F8000AD0B, 0x8000AD178000AD13,
-    0x000631B5000639B6, 0x0006B1BB8000AD1B, 0x8000AD1F000619BA,
-    0x000691BF8000AD23, 0x0006A1BD0006A9BE, 0x8000AD270006A9BC,
-    0x8000AD2B000665A3, 0x00067DA68000AD2F, 0x8000AD378000AD33,
-    0x00065DAA000675A5, 0x8000AD3B0006F5AB, 0x0006EDAE00012084,
-    0x0006EDAC0006D5AF, 0x00018DA80006E5AD, 0x0001DDAA000199A9,
-    0x0001F9AC0001D5AB, 0x000139AE0001F9AD, 0x0000F580000139AF,
-    0x0000C1820000FD81, 0x000010860001E4BE, 0x8000AD4300008487,
-    0x0000CDB88000AD47, 0x0000D5BA0000D5B9, 0x0000FDBC0000E5BB,
-    0x000091BE000095BD, 0x000149B0000091BF, 0x000159B2000149B1,
-    0x000149B4000159B3, 0x0000FDB6000149B5, 0x00044DBA0000F5B7,
-    0x000425B800046DBB, 0x000465BE000445B9, 0x000475BC000469BF,
-    0x0007EDB200046DBD, 0x0007EDB00007F5B3, 0x0007DDB60007E5B1,
-    0x0007E5B400043DB7, 0x000751AA0007D5B5, 0x000741A8000759AB,
-    0x0007DDAE000741A9, 0x000771AC0007C9AF, 0x8000AD4B000771AD,
-    0x8000AD538000AD4F, 0x8000AD5B8000AD57, 0x8000AD638000AD5F,
-    0x0002D5B38000AD67, 0x8000AD6F8000AD6B, 0x8000AD778000AD73,
-    0x0002F1B50002F9B6, 0x00020DBB8000AD7B, 0x8000AD7F00020DBA,
-    0x00020DBF8000AD83, 0x00020DBD00020DBE, 0x8000AD8700020DBC,
-    0x0003C884000291A3, 0x0002BDA6000C40BE, 0x8000AD8F8000AD8B,
-    0x000249AA0002B5A5, 0x8000AD93000249AB, 0x000249AE8000AD97,
-    0x000249AC000249AF, 0x0002B9A8000249AD, 0x000169AA0002B9A9,
-    0x000179AC000169AB, 0x000169AE000179AD, 0x000C2CBE000169AF,
-    0x8000AD9F8000AD9B, 0x00001D808000ADA3, 0x0000BD8200000981,
-    0x0001EDB88000ADA7, 0x00019DBA000185B9, 0x0001B9BC000195BB,
-    0x00007DBE0001B9BD, 0x000119B0000075BF, 0x0001EDB2000119B1,
-    0x0001FDB40001E5B3, 0x0001E5B60001E5B5, 0x000768E10001D5B7,
-    0x000764E38000ADAB, 0x8000ADB78000ADB3, 0x8000ADBB0001A4E1,
-    0x000C2C840000E0E3, 0x000FA8BE8000ADBF, 0x000D90860000F8EF,
-    0x000678EF000DD487, 0x8000ADC78000ADC3, 0x00F9F0858000ADCB,
-    0x8000ADCF0001ADB3, 0x0001B5B60001BDB5, 0x8000ADD78000ADD3,
-    0x000191BA8000ADDB, 0x000175BC000191BB, 0x000175BE00017DBD,
-    0x000E61A300016DBF, 0x8000ADDF8000ADAF, 0x8000ADE78000ADE3,
-    0x000E71A5000E79A6, 0x000E5DAB8000ADEB, 0x8000ADEF000E5DAA,
-    0x000EA1AF8000ADF3, 0x000EB1AD000EB9AE, 0x8000ADF7000EB9AC,
-    0x8000ADFF8000ADFB, 0x8000AE078000AE03, 0x8000AE0F8000AE0B,
-    0x00003D808000AE13, 0x0000198200000981, 0x8000AE1B8000AE17,
-    0x0003B8878000AE1F, 0x000F0DA80000DC86, 0x000F11AA000F15A9,
-    0x000F2DAC000F29AB, 0x000F5DAE000F55AD, 0x0003A084000F4DAF,
-    0x8000AE278000AE23, 0x8000AE2F8000AE2B, 0x8000AE378000AE33,
-    0x0000EDB88000AE3B, 0x00008DBA000085B9, 0x00009DBC000085BB,
-    0x00008DBE000085BD, 0x000F3DB0000085BF, 0x000F0DB2000F05B1,
-    0x0000FDB40000E5B3, 0x0000E5B60000E5B5, 0x8000AE3F0000D5B7,
-    0x8000AE478000AE43, 0x8000AE4F8000AE4B, 0x8000AE53000EBCE1,
-    0x8000AE57000DC4E3, 0x8000AE5F8000AE5B, 0x8000AE63000D18EF,
-    0x8000AE6B8000AE67, 0x000ED9B38000AE6F, 0x8000AE778000AE73,
-    0x8000AE7F8000AE7B, 0x000EE5B5000E95B6, 0x000E81BB8000AE83,
-    0x8000AE87000EB9BA, 0x000161BF8000AE8B, 0x000171BD000179BE,
-    0x000E19A3000E99BC, 0x000028868000AE93, 0x8000AE9700015C87,
-    0x000E25A5000E55A6, 0x000E41AB8000AE9B, 0x8000AE9F000E79AA,
-    0x0001A1AF8000AEA3, 0x0001B1AD0001B9AE, 0x00009D80000E59AC,
-    0x0000A5820000AD81, 0x8000AEA7000161B3, 0x000161B6000161B5,
-    0x8000AEAF8000AEAB, 0x0000EDBA8000AEB3, 0x0000FDBC0000E5BB,
-    0x0000E1BE0000E5BD, 0x0002A5A80000DDBF, 0x0002A5AA0002ADA9,
-    0x0002A5AC0002BDAB, 0x0002A5AE0002ADAD, 0x0004EC8400031DAF,
-    0x8000AEBB8000AEB7, 0x8000AEC38000AEBF, 0x8000AECB8000AEC7,
-    0x000371B88000AECF, 0x000371BA000371B9, 0x000315BC000371BB,
-    0x000315BE00031DBD, 0x000365B000030DBF, 0x000365B200036DB1,
-    0x000369B400037DB3, 0x00035DB6000355B5, 0x8000AED3000355B7,
-    0x8000AEDB8000AED7, 0x8000AD3F000221A3, 0x000221A6000221A5,
-    0x8000AEE38000AEDF, 0x0003ADAA8000AEE7, 0x0003BDAC0003A5AB,
-    0x0003A1AE0003A5AD, 0x8000AEEB00039DAF, 0x8000AEF38000AEEF,
-    0x000019808000AEF7, 0x0000058200001981, 0x000428BE8000AEFB,
-    0x0004C8868000AF03, 0x8000AF0700032887, 0x8000AF0F8000AF0B,
-    0x000285A88000AF13, 0x000295AA000295A9, 0x0002BDAC0002A5AB,
-    0x0002C5AE0002C5AD, 0x8000AF170002FDAF, 0x8000AF1F8000AF1B,
-    0x8000AF278000AF23, 0x8000AF2F8000AF2B, 0x0001C1B88000AF33,
-    0x0001C1BA0001C1B9, 0x0001F1BC0001D1BB, 0x000199BE0001F1BD,
-    0x000285B0000199BF, 0x000141B2000141B1, 0x000141B4000141B3,
-    0x000141B6000141B5, 0x000E28E1000141B7, 0x000E38E38000AF37,
-    0x00041C84000480BE, 0x8000AF3F8000AF3B, 0x8000AF478000AF43,
-    0x8000AF4F8000AF4B, 0x8000AF578000AF53, 0x000E60EF8000AF5B,
-    0x000015818000AF5F, 0x000E98EF00001580, 0x8000AF6300001582,
-    0x8000AF6B8000AF67, 0x000740868000AF73, 0x8000AF770004E487,
-    0x000194E18000AF7B, 0x000148E38000AF7F, 0x00059DB38000AF83,
-    0x8000AF6F8000AEFF, 0x8000AF8B8000AF87, 0x0005B5B50005C5B6,
-    0x000605BB8000AF8F, 0x8000AF930005C1BA, 0x000605BF8000AF97,
-    0x000615BD000605BE, 0x8000AF9B000615BC, 0x8000AF9F0005D9A3,
-    0x000581A68000AFA3, 0x8000AFAB8000AFA7, 0x000585AA0005F1A5,
-    0x8000AFAF000641AB, 0x000641AE8000AFB3, 0x000651AC000641AF,
-    0x000665A8000651AD, 0x00067DAA000675A9, 0x00066DAC000675AB,
-    0x0007BDAE0007B5AD, 0x0003A8840007B5AF, 0x8000AFBB8000AFB7,
-    0x8000AFC38000AFBF, 0x8000AFCB8000AFC7, 0x00075DB88000AFCF,
-    0x00076DBA000765B9, 0x00077DBC000765BB, 0x00076DBE000765BD,
-    0x0007CDB0000761BF, 0x0007DDB20007D5B1, 0x0007CDB40007D5B3,
-    0x00076DB6000771B5, 0x00000D80000765B7, 0x00001D8200001581,
-    0x8000AFD78000AFD3, 0x000500EF8000AFDB, 0x0006B0E100011CBE,
-    0x000518E38000AFE3, 0x000090868000AFE7, 0x8000AFEB00011487,
-    0x0007B5B38000AFEF, 0x8000AFF38000AFDF, 0x8000AFFB8000AFF7,
-    0x0007A5B5000675B6, 0x000651BB8000AFFF, 0x8000B003000649BA,
-    0x0006F5BF8000B007, 0x000641BD000649BE, 0x000775A3000649BC,
-    0x8000B00F8000B00B, 0x8000B0178000B013, 0x000765A50006B5A6,
-    0x000691AB8000B01B, 0x8000B01F000689AA, 0x000635AF8000B023,
-    0x000681AD000689AE, 0x8000B027000689AC, 0x8000B02B000629B3,
-    0x0001DDB68000B02F, 0x8000B0378000B033, 0x0001C9BA000621B5,
-    0x8000B03B0001C9BB, 0x0001BDBE8000B03F, 0x0001B9BC0001A1BF,
-    0x0005DDA80001B5BD, 0x0005E5AA0005EDA9, 0x0005EDAC0005F9AB,
-    0x000209AE000219AD, 0x00000D80000201AF, 0x00001D8200001581,
-    0x8000B0478000B043, 0x00FC00BE8000B04B, 0x00022DB88000B053,
-    0x0002E1BA0002E1B9, 0x0002E1BC0002E1BB, 0x0002E1BE0002E1BD,
-    0x000249B00002E1BF, 0x000229B2000249B1, 0x00023DB4000225B3,
-    0x000221B6000221B5, 0x00038884000215B7, 0x8000B05B8000B057,
-    0x8000B05F000561A3, 0x000295A6000569A5, 0x8000B0678000B063,
-    0x000281AA8000AE8F, 0x0002F1AC000281AB, 0x0002F5AE0002FDAD,
-    0x0001E8E10002E9AF, 0x000258E38000B06B, 0x00FC508600FCC084,
-    0x8000B06F00037087, 0x8000B0778000B073, 0x8000B07F8000B07B,
-    0x8000B0878000B083, 0x000238EF8000B08B, 0x000225A88000B08F,
-    0x000231AA000229A9, 0x000259AC000231AB, 0x000249AE000259AD,
-    0x8000B093000241AF, 0x00FCA4BE8000B097, 0x8000B09F8000B09B,
-    0x8000B0A78000B0A3, 0x0002EDB88000B0AB, 0x000281BA000285B9,
-    0x000285BC000281BB, 0x0002B1BE00028DBD, 0x000209B00002B1BF,
-    0x0002EDB2000209B1, 0x0002FDB40002E5B3, 0x0002E5B60002E5B5,
-    0x000770E10002D5B7, 0x000608E3000134E3, 0x8000B0AF000674E1,
-    0x8000B0B78000B0B3, 0x000039818000B0BB, 0x8000B0C300003180,
-    0x8000B0C700003182, 0x0006CCEF000050EF, 0x8000B0CF8000B0CB,
-    0x8000B0D30002FDB3, 0x00029DB68000B0D7, 0x8000B0DF8000B0DB,
-    0x000289BA000295B5, 0x00FCC886000289BB, 0x00027DBE00FCE087,
-    0x00027DBC000265BF, 0x8000B04F000275BD, 0x8000B0E38000B0BF,
-    0x8000B0EB8000B0E7, 0x8000B0F38000B0EF, 0x00FE1DA88000B0F7,
-    0x00FEA1AA00FEA1A9, 0x00FEADAC00FEBDAB, 0x00FEFDAE00FE95AD,
-    0x00FE9DB000FEE5AF, 0x00FEADB200FEA5B1, 0x00FEBDB400FEA5B3,
-    0x00FEADB600FEA5B5, 0x00FF41B800FEA1B7, 0x00FF45BA00FF41B9,
-    0x00FF45BC00FF5DBB, 0x00FF45BE00FF4DBD, 0x00FD0DAE00FFEDBF,
-    0x00FD0DAC00FD15AF, 0x00FDF9AA00FD05AD, 0x8000B0FB00FDF9AB,
-    0x00FDEDA68000B0FF, 0x8000B1078000B103, 0x8000B10B00FDE5A5,
-    0x8000B10F00FD8DA3, 0x00FE49B38000B113, 0x8000B11B8000B117,
-    0x8000B1238000B11F, 0x00FE45B500FE75B6, 0x00FE01BB8000B127,
-    0x8000B12B00FE39BA, 0x00FE01BF00036484, 0x00FE11BD00FE19BE,
-    0x00FF89A300FE19BC, 0x0001D5810001ED82, 0x00019CBE0001DD80,
-    0x00FF85A500FFB5A6, 0x00FFC1AB8000B12F, 0x0000888600FFF9AA,
-    0x00FFC1AF00017C87, 0x00FFD1AD00FFD9AE, 0x8000B13A00FFD9AC,
-    0x8000B13E00FEF9B3, 0x00FE9DB68000B142, 0x8000B14A8000B146,
-    0x000169BA00FE95B5, 0x8000B14E000169BB, 0x000169BE8000B152,
-    0x000169BC000169BF, 0x00FE29A8000169BD, 0x00FE35AA00FE35A9,
-    0x00FE59AC00FE4DAB, 0x00FE4DAE00FE59AD, 0x00018C8400FE45AF,
-    0x8000B15A8000B156, 0x8000B1628000B15E, 0x8000B16A8000B166,
-    0x000165B88000B16E, 0x000179BA00016DB9, 0x00016DBC000175BB,
-    0x000111BE000111BD, 0x0001E9B0000111BF, 0x0001F9B20001E9B1,
-    0x0001E9B40001F9B3, 0x000161B60001E9B5, 0x00FDBDA3000161B7,
-    0x8000B1768000B172, 0x8000B17A000348BE, 0x00FDD1A500FDD9A6,
-    0x00022DAB8000B17E, 0x8000B18200022DAA, 0x00022DAF8000B186,
-    0x00022DAD00022DAE, 0x8000B18A00022DAC, 0x8000B1928000B18E,
-    0x00000D808000B196, 0x00003D8200003581, 0x8000B19E8000B19A,
-    0x0004C8868000B1A2, 0x8000B1AA00035C87, 0x8000B1B28000B1AE,
-    0x0004CC848000B1B6, 0x8000B1BE8000B1BA, 0x000120E38000B1C2,
-    0x0001A8E18000B1C6, 0x000254EF8000B1CA, 0x8000B1D28000B1CE,
-    0x8000B1DA8000B1D6, 0x8000B1E28000B1DE, 0x00036DBB8000B1E6,
-    0x8000B1EA00036DBA, 0x00036DBF8000B1EE, 0x00036DBD00036DBE,
-    0x0003ADB300036DBC, 0x8000B1F68000B1F2, 0x8000B1FE8000B1FA,
-    0x0003A1B500039DB6, 0x000241A800057084, 0x000245AA00024DA9,
-    0x000249AC000259AB, 0x000275AE00027DAD, 0x000568BE0002B9AF,
-    0x8000B2068000B202, 0x00001D808000B20A, 0x0000A98200000981,
-    0x000151B88000B20E, 0x000161BA000159B9, 0x000111BC000161BB,
-    0x000111BE000111BD, 0x0002C9B0000111BF, 0x0002D9B20002C9B1,
-    0x0002C9B40002D9B3, 0x00017DB60002C9B5, 0x00FE70E1000175B7,
-    0x000174E300FF4CE1, 0x8000B21200FF50E3, 0x0004B8868000B21A,
-    0x8000B21E00055487, 0x8000B2268000B222, 0x000000E48000B22A,
-    0x000050EF8000B22E, 0x8000B23200FE6CEF, 0x000285A38000B236,
-    0x000289A58000B23A, 0x8000B2428000B23E, 0x8000B2460002B5A6,
-    0x000245AB8000B24A, 0x000245AD000245AA, 0x000245AF000245AC,
-    0x000671A8000245AE, 0x000671AA000671A9, 0x000695AC000671AB,
-    0x000695AE00069DAD, 0x8000B216000689AF, 0x8000B2528000B24E,
-    0x8000B25A8000B256, 0x8000B2628000B25E, 0x000689B88000B266,
-    0x0006ADBA000689B9, 0x0006BDBC0006A5BB, 0x00077DBE0006A5BD,
-    0x0006FDB0000775BF, 0x0006C1B20006C1B1, 0x0006C1B40006C1B3,
-    0x0006C1B60006C1B5, 0x0006F9B30006C1B7, 0x8000B26E8000B26A,
-    0x8000B2768000B272, 0x00061DB500061DB6, 0x000605BB8000B27A,
-    0x8000B27E000639BA, 0x000605BF8000B282, 0x000615BD000605BE,
-    0x00006D82000615BC, 0x000055800006BDA3, 0x000659A600006581,
-    0x8000B286000054BE, 0x00067DAA000659A5, 0x000C0086000641AB,
-    0x000641AE00035C87, 0x000651AC000641AF, 0x8000B28A000651AD,
-    0x8000B28E0004DCEF, 0x8000B2968000B292, 0x8000B29E8000B29A,
-    0x8000B2A68000B2A2, 0x8000B2AE8000B2AA, 0x8000B2B68000B2B2,
-    0x8000B2BA0005A4E3, 0x0006E5B300054CE1, 0x8000B2C28000B2BE,
-    0x8000B2CA8000B2C6, 0x0006F5B50006E1B6, 0x0006B5BB8000B2CE,
-    0x8000B2D20006A9BA, 0x000155BF8000B2D6, 0x000699BD000691BE,
-    0x000615A80006A5BC, 0x00062DAA000625A9, 0x00063DAC000625AB,
-    0x000659AE000621AD, 0x8000B2DA000659AF, 0x8000B2E28000B2DE,
-    0x00000D808000B2E6, 0x0003BD820003B581, 0x0001F1B88000B2EA,
-    0x0001F1BA0001F1B9, 0x000155BC0001F1BB, 0x000155BE00015DBD,
-    0x00062DB000014DBF, 0x000631B2000631B1, 0x000615B400060DB3,
-    0x0001D1B60001D1B5, 0x8000B2EE0001D1B7, 0x000290848000B2F2,
-    0x00020CBE0005A9A3, 0x0005ADA60005B9A5, 0x8000B2FE8000B2FA,
-    0x0005E5AA8000B302, 0x0005E9AC0005F9AB, 0x0005DDAE0005D5AD,
-    0x000359B6000219AF, 0x00038C848000B306, 0x8000B30A000359B5,
-    0x000CD886000359B3, 0x000359BE0002AC87, 0x000345BC000321BF,
-    0x000355BA00035DBD, 0x8000B30E00035DBB, 0x8000B3168000B312,
-    0x00039DA500039DA6, 0x00039DA38000B31A, 0x8000B3228000B31E,
-    0x0003E5AF8000B326, 0x000399AD00039DAE, 0x000399AB000381AC,
-    0x8000B1A6000391AA, 0x8000B32E8000B32A, 0x00088C858000B332,
-    0x8000B33A8000B336, 0x8000B3428000B33E, 0x8000B34A8000B346,
-    0x8000B3528000B34E, 0x00000D808000B356, 0x0001FD820001F581,
-    0x8000B35E8000B35A, 0x00012DB38000B362, 0x000141B500024CBE,
-    0x8000B36E8000B36A, 0x00029887000165B6, 0x0001BDBB000D0486,
-    0x0001ADBD0001A1BA, 0x0001F9BF0001ADBC, 0x8000B372000181BE,
-    0x8000B37A8000B376, 0x8000B3828000B37E, 0x8000B38600034CE3,
-    0x8000B38A0001ACE1, 0x8000B38E00038CEF, 0x8000B3968000B392,
-    0x8000B39E8000B39A, 0x8000B3A68000B3A2, 0x8000B3AE8000B3AA,
-    0x8000B3B2000C4084, 0x8000B3BA8000B3B6, 0x8000B3BE0001B0EF,
-    0x8000B3C2000624E1, 0x00007982000740E3, 0x000079808000B3C6,
-    0x8000B3CA00007981, 0x8000B3D28000B3CE, 0x8000B3DA0001E9A3,
-    0x0001A1A6000185A5, 0x000CC0868000B3DE, 0x000165AA000C4487,
-    0x000169AC000179AB, 0x000145AE000169AD, 0x8000B2F600013DAF,
-    0x0003C0BE8000B366, 0x8000B3D68000B3E2, 0x8000B3EA8000B3E6,
-    0x000E3DA88000B3EE, 0x000E0DAA000E05A9, 0x000E61AC000E69AB,
-    0x000E45AE000E4DAD, 0x000ECDB0000EBDAF, 0x000ED1B2000ED1B1,
-    0x000EEDB4000ED1B3, 0x000E9DB6000E95B5, 0x000EB1B8000E91B7,
-    0x000EB1BA000EB1B9, 0x000E91BC000EB1BB, 0x000E91BE000E91BD,
-    0x000F95A8000E91BF, 0x000FC9AA000F9DA9, 0x000FD9AC000FC9AB,
-    0x000FC9AE000FD9AD, 0x8000B3F2000FC5AF, 0x8000B3FA8000B3F6,
-    0x8000B4028000B3FE, 0x8000B40A8000B406, 0x000F45B88000B40E,
-    0x000F59BA000F4DB9, 0x000F49BC000F59BB, 0x000F71BE000F75BD,
-    0x000F81B0000F69BF, 0x000F85B2000F8DB1, 0x000F85B4000F9DB3,
-    0x000F85B6000F8DB5, 0x000E41B3000F7DB7, 0x8000B4168000B412,
-    0x8000B41E8000B41A, 0x000E51B5000E59B6, 0x000E49BB8000B422,
-    0x8000B426000E45BA, 0x000E49BF8000B42A, 0x000E49BD000E41BE,
-    0x8000B42E000E51BC, 0x8000B432000E05A3, 0x000E1DA68000B436,
-    0x8000B43E8000B43A, 0x000E01AA000E15A5, 0x8000B442000E0DAB,
-    0x000E05AE00012084, 0x000E15AC000E0DAF, 0x000E51A8000E0DAD,
-    0x000185AA000E59A9, 0x000189AC000199AB, 0x000139AE000189AD,
-    0x0000F580000139AF, 0x0000C1820000FD81, 0x000010860001E4BE,
-    0x8000B44A00009487, 0x0000CDB88000B44E, 0x0000D5BA0000D5B9,
-    0x0000FDBC0000E5BB, 0x000091BE000095BD, 0x000149B0000091BF,
-    0x000159B2000149B1, 0x000149B4000159B3, 0x0000FDB6000149B5,
-    0x000F69BA0000F5B7, 0x000F59B8000F69BB, 0x000FB9BE000F51B9,
-    0x000FB9BC000FB9BF, 0x000F0DB2000FB9BD, 0x000F0DB0000F15B3,
-    0x000F79B6000F05B1, 0x000F01B4000F79B7, 0x000F4DAA000F79B5,
-    0x000F49A8000F55AB, 0x000F4DAE000F45A9, 0x000F4DAC000F75AF,
-    0x8000B452000F45AD, 0x8000B45A8000B456, 0x8000B4628000B45E,
-    0x8000B46A8000B466, 0x00028DB38000B46E, 0x8000B4768000B472,
-    0x8000B47E8000B47A, 0x000281B5000289B6, 0x00020DBB8000B482,
-    0x8000B48600020DBA, 0x00020DBF8000B48A, 0x00020DBD00020DBE,
-    0x8000B48E00020DBC, 0x8000B4920002C9A3, 0x0002CDA68000B496,
-    0x8000B49E8000B49A, 0x000249AA0002C5A5, 0x8000B4A2000249AB,
-    0x000249AE8000B4A6, 0x000249AC000249AF, 0x0002B9A8000249AD,
-    0x000169AA0002B9A9, 0x000179AC000169AB, 0x000169AE000179AD,
-    0x00000D80000169AF, 0x00001D8200001581, 0x8000B4AE8000B4AA,
-    0x0005B4BE8000B4B2, 0x0001EDB88000B4BA, 0x000181BA000185B9,
-    0x000185BC000181BB, 0x0001B1BE00018DBD, 0x000119B00001B1BF,
-    0x0001EDB2000119B1, 0x0001FDB40001E5B3, 0x0001E5B60001E5B5,
-    0x000EBCE10001D5B7, 0x000E14E3000FC4E1, 0x00029084000FD4E3,
-    0x0006CCEF000E6CE1, 0x000F7CE10000E0E3, 0x00137CE38000B4BE,
-    0x000490860000F8EF, 0x000148EF00041487, 0x8000B4C2000E48EF,
-    0x000704848000B4C6, 0x8000B4CA0001ADB3, 0x0001B5B60001BDB5,
-    0x8000B4D28000B4CE, 0x000191BA8000B4D6, 0x000175BC000191BB,
-    0x000175BE00017DBD, 0x000661A300016DBF, 0x8000B4B600FC098F,
-    0x8000B4DE8000B4DA, 0x000671A5000679A6, 0x00065DAB8000B4E2,
-    0x8000B4E600065DAA, 0x0006A1AF8000B4EA, 0x0006B1AD0006B9AE,
-    0x0008619E0006B9AC, 0x000D559C00089D9F, 0x000C519A000AF59D,
-    0x00F05198000C519B, 0x00F2C59600F11999, 0x00F46D9400F06197,
-    0x00F8619200F41995, 0x00FE519000F6E593, 0x002CB1B100F87991,
-    0x002889B3002CA5B0, 0x0029B1B5002DE1B2, 0x8000B4EE002891B4,
-    0x8000B4F68000B4F2, 0x8000B4FE8000B4FA, 0x0004A1A28000B502,
-    0x0004B9A0000265A3, 0x0000C9A10004B9A1, 0x001CC9A38000B506,
-    0x001D15A5001C01A2, 0x0018F9A7001CC5A4, 0x001401A9001801A6,
-    0x001441AB0019C9A8, 0x0010D1AD0014CDAA, 0x002C01AF001001AC,
-    0x00F5758A0010CDAE, 0x8000B50A00F3A18B, 0x00F0558E8000B50E,
-    0x00F11D8C000D158F, 0x00FD2D8200F1158D, 0x8000B51200FF0D83,
-    0x00F7A1868000B516, 0x00F93D8400F53187, 0x0005459A00F92D85,
-    0x8000B51A0005559B, 0x8000B5228000B51E, 0x0003899C0006FCE1,
-    0x000D65920004DCE3, 0x0001A0BE000C7993, 0x0009A9960004ACEF,
-    0x000BC19400057597, 0x0006BDA800095D95, 0x000649AA000621A9,
-    0x000641AC000649AB, 0x000679AE000649AD, 0x0000E980000679AF,
-    0x0000F9820000E981, 0x0001F0868000B52A, 0x8000B52E0000B487,
-    0x000165B88000B532, 0x000165BA00016DB9, 0x0000D9BC000109BB,
-    0x0000F1BE0000D9BD, 0x0001EDB00000F1BF, 0x0001F5B20001F5B1,
-    0x0001DDB40001C5B3, 0x000165B600016DB5, 0x0005C1B300015DB7,
-    0x8000B53A8000B536, 0x8000B5428000B53E, 0x0005C9B50005C9B6,
-    0x000231BB8000B546, 0x8000B54A000229BA, 0x0002A9BF8000B54E,
-    0x000211BD000219BE, 0x8000B552000221BC, 0x8000B556000585A3,
-    0x00058DA68000B55A, 0x8000B5628000B55E, 0x00026DAA00058DA5,
-    0x8000B566000275AB, 0x00025DAE8000B56A, 0x000265AC0002EDAF,
-    0x0000A9B0000255AD, 0x0000C9B20000A9B1, 0x0000DDB40000C5B3,
-    0x0000C1B60000C1B5, 0x0000CDB80000F5B7, 0x00006DBA000065B9,
-    0x00007DBC000065BB, 0x000061BE000061BD, 0x8000B56E000061BF,
-    0x8000B5768000B572, 0x8000B4468000B57A, 0x8000B5828000B57E,
-    0x00034DA88000B586, 0x000355AA00035DA9, 0x0000F9AC0000E9AB,
-    0x0000E9AE0000F9AD, 0x00000D800000E1AF, 0x00001D8200001581,
-    0x8000B58E8000B58A, 0x00025CBE8000B592, 0x0002AC848000B59A,
-    0x001D24878000B59E, 0x8000B5A2001DAC86, 0x8000B5AA8000B5A6,
-    0x0002FDB38000B5AE, 0x8000B5B68000B5B2, 0x8000B5BE8000B5BA,
-    0x000151B5000159B6, 0x00014DBB8000B5C2, 0x8000B5C600014DBA,
-    0x000125BF8000B5CA, 0x000131BD000139BE, 0x8000B5CE00013DBC,
-    0x8000B5D200065CE3, 0x001DDC84000624E1, 0x8000B5DA8000B5D6,
-    0x001DE8BE8000B5DE, 0x0024A0BF8000B5E2, 0x8000B5EA8000B5E6,
-    0x8000B5EE000024EF, 0x8000B5F68000B5F2, 0x000275A38000B5FA,
-    0x0001D9A58000B5FE, 0x8000B60A8000B602, 0x001C40860001D1A6,
-    0x0001C5AB001C6487, 0x0001B9AD0001C5AA, 0x0001ADAF0001B5AC,
-    0x8000B5960001B1AE, 0x001FC581001FD182, 0x8000B606001FC980,
-    0x8000B6128000B60E, 0x001F3DA88000B616, 0x001F19AA001F01A9,
-    0x001F09AC001F19AB, 0x001F7DAE001F09AD, 0x001F0DB0001F75AF,
-    0x001F15B2001F15B1, 0x001F3DB4001F25B3, 0x001FD1B6001FD5B5,
-    0x001FF9B8001FC9B7, 0x001FCDBA001FF9B9, 0x001FC1BC001FC1BB,
-    0x001FC5BE001FCDBD, 0x001FB1B3001FBDBF, 0x8000B61E8000B61A,
-    0x8000B6268000B622, 0x001FA1B5001FA9B6, 0x001FF1BB8000B62A,
-    0x8000B62E001FE9BA, 0x001F31BF8000B632, 0x001FD1BD001FD9BE,
-    0x8000B636001FE1BC, 0x8000B63A001FF5A3, 0x001FEDA68000B526,
-    0x8000B6428000B63E, 0x001FADAA001FE5A5, 0x8000B646001FB5AB,
-    0x001F9DAE8000B64A, 0x001FA5AC001F75AF, 0x001ED9A8001F95AD,
-    0x000111AA000111A9, 0x000195AC000111AB, 0x000195AE00019DAD,
-    0x8000B64E00018DAF, 0x8000B6568000B652, 0x8000B65E8000B65A,
-    0x8000B6668000B662, 0x0001ADB88000B66A, 0x00006DBA000065B9,
-    0x00007DBC000065BB, 0x00006DBE000065BD, 0x0001FDB0000065BF,
-    0x0001C5B20001CDB1, 0x0001B9B40001ADB3, 0x0001A1B60001A1B5,
-    0x0000BD80000195B7, 0x0000198200000981, 0x8000B6728000B66E,
-    0x000118BE00029884, 0x000180868000B67A, 0x8000B67E0001C487,
-    0x8000B6868000B682, 0x8000B68E8000B68A, 0x0001D9B38000B692,
-    0x8000B6968000B676, 0x8000B69E8000B69A, 0x00015DB500015DB6,
-    0x000149BB8000B6A2, 0x8000B6A6000179BA, 0x000125BF8000B6AA,
-    0x000139BD00013DBE, 0x8000B6AE000151BC, 0x8000B6B68000B6B2,
-    0x8000B6BA000215A3, 0x000291A6000291A5, 0x8000B6C28000B6BE,
-    0x0002B5AA8000B6C6, 0x00029DAC000285AB, 0x0002F1AE0002F5AD,
-    0x0001F8E10002E9AF, 0x000214E38000B6CA, 0x8000B6CE00038C84,
-    0x8000B6D68000B6D2, 0x00003D808000B6DA, 0x0000198200000981,
-    0x8000B6E28000B6DE, 0x000220EF8000B6EA, 0x00031DA88000B6EE,
-    0x000325AA00032DA9, 0x000325AC00033DAB, 0x000325AE00032DAD,
-    0x0003ECBE000355AF, 0x000408868000B6F2, 0x8000B6F600030C87,
-    0x8000B6FE8000B6FA, 0x0000D1B88000B702, 0x0000E9BA0000D9B9,
-    0x000095BC0000E9BB, 0x00008DBE000085BD, 0x00032DB0000085BF,
-    0x000335B2000335B1, 0x00031DB4000305B3, 0x0000FDB60000F5B5,
-    0x8000B7060000F5B7, 0x8000B70E8000B70A, 0x8000B7120002A1B3,
-    0x0002A1B60002A1B5, 0x8000B71A8000B716, 0x0002E5BA8000B71E,
-    0x0002EDBC0002FDBB, 0x0002D5BE0002DDBD, 0x001E70E300013DBF,
-    0x001CE0E18000B722, 0x8000B72A8000B726, 0x8000B7328000B72E,
-    0x00052C848000B736, 0x8000B73E8000B73A, 0x001E90EF8000B742,
-    0x8000B74A8000B746, 0x0005C0BE8000B74E, 0x8000B752001EB0E1,
-    0x8000B75A001E0CE3, 0x8000B75E001E6CE1, 0x0000D980000018E3,
-    0x000021820000D981, 0x8000B7620000CCEF, 0x0004F0BE8000B766,
-    0x00026DA3001EE4EF, 0x0004E8878000B76A, 0x8000B76E00046C86,
-    0x00026DA500026DA6, 0x000231AB8000B772, 0x8000B776000229AA,
-    0x0001F1AF8000B77A, 0x000211AD000219AE, 0x000675A8000221AC,
-    0x0006ADAA000679A9, 0x0006A5AC0006BDAB, 0x0006A5AE0006ADAD,
-    0x8000B75600071DAF, 0x8000B7828000B77E, 0x8000B78A8000B786,
-    0x8000B7928000B78E, 0x000711B88000B796, 0x000711BA000711B9,
-    0x000735BC000711BB, 0x000735BE00073DBD, 0x000765B0000729BF,
-    0x000779B200076DB1, 0x000751B4000771B3, 0x00073DB6000751B5,
-    0x000625B3000731B7, 0x8000B79E8000B79A, 0x8000B7A68000B7A2,
-    0x00062DB500062DB6, 0x0007A9BB8000B7AA, 0x8000B7AE0007A1BA,
-    0x000795BF8000B7B2, 0x0007A5BD000795BE, 0x8000B7B60007ADBC,
-    0x8000B7BE8000B7BA, 0x8000B7C2000669A3, 0x000661A6000661A5,
-    0x0003AC848000B6E6, 0x0007EDAA8000B7CA, 0x0007E1AC0007E5AB,
-    0x0007D9AE0007E9AD, 0x000691A80007D9AF, 0x0006B9AA00069DA9,
-    0x0006D5AC0006B9AB, 0x0006C5AE0006C5AD, 0x0001E5800006F5AF,
-    0x0001F9820001ED81, 0x000190868000B7C6, 0x8000B7CE0001B487,
-    0x0001C5B88000B7D2, 0x0001C5BA0001CDB9, 0x0001CDBC0001DDBB,
-    0x0001F5BE0001FDBD, 0x00068DB000019DBF, 0x000141B2000141B1,
-    0x000141B4000141B3, 0x000141B6000141B5, 0x000631B3000141B7,
-    0x8000B7DA8000B7D6, 0x8000B7E28000B7DE, 0x00065DB5000645B6,
-    0x000185BB8000B7E6, 0x8000B7EA000649BA, 0x000185BF8000B7EE,
-    0x000195BD000185BE, 0x8000B7F2000195BC, 0x8000B7F6000675A3,
-    0x000601A68000B7FA, 0x8000B8028000B7FE, 0x00060DAA000619A5,
-    0x8000B8060001C1AB, 0x0001C1AE8000B80A, 0x0001D1AC0001C1AF,
-    0x0002E5A80001D1AD, 0x000335AA00033DA9, 0x000399AC000389AB,
-    0x000389AE000399AD, 0x8000B80E000381AF, 0x8000B8168000B812,
-    0x8000B81E8000B81A, 0x8000B8268000B822, 0x0003ADB88000B82A,
-    0x00006DBA000065B9, 0x00007DBC000065BB, 0x00006DBE000065BD,
-    0x0003C9B0000065BF, 0x0003A9B20003C9B1, 0x0003BDB40003A5B3,
-    0x0003A1B60003A1B5, 0x0000BD80000395B7, 0x00001D8200000981,
-    0x0002F0EF8000B82E, 0x000318BE8000B832, 0x000214E38000B83A,
-    0x000180E10002A884, 0x8000B8428000B83E, 0x0003C8878000B846,
-    0x0003FDB3000C1C86, 0x8000B84E8000B84A, 0x8000B8568000B852,
-    0x000351B5000359B6, 0x00034DBB8000B85A, 0x8000B85E00034DBA,
-    0x000325BF8000B862, 0x000331BD000339BE, 0x8000B86600033DBC,
-    0x8000B86E8000B86A, 0x8000B872000231A3, 0x000295A600029DA5,
-    0x00022C848000B876, 0x000281AA8000B87A, 0x0002F1AC000281AB,
-    0x0002F5AE0002FDAD, 0x8000B87E0002E9AF, 0x000D44BE8000B882,
-    0x000620E3000D4C84, 0x0007E8E18000B886, 0x000618EF8000B88A,
-    0x8000B8928000B88E, 0x00000D808000B896, 0x0000FD820000F581,
-    0x8000B89E8000B89A, 0x000D48878000B8A2, 0x8000B8AA000C2C86,
-    0x8000B8B28000B8AE, 0x000CACBE0006DCEF, 0x8000B8B6000670E1,
-    0x8000B8BA00015CE3, 0x8000B8C28000B8BE, 0x000171B38000B8C6,
-    0x8000B8CE8000B8CA, 0x8000B8D68000B8D2, 0x00011DB500011DB6,
-    0x000109BB8000B8DA, 0x8000B8DE000139BA, 0x0000F9BF8000B8E2,
-    0x0000F9BD0000F1BE, 0x000E9DA8000111BC, 0x000EB9AA000EA9A9,
-    0x000ED5AC000EB9AB, 0x000ED5AE000EDDAD, 0x8000B8A6000ECDAF,
-    0x8000B8EA8000B8E6, 0x8000B8F28000B8EE, 0x8000B8FA8000B8F6,
-    0x000F65B88000B8FE, 0x000F0DBA000F05B9, 0x000F0DBC000F19BB,
-    0x000F31BE000F31BD, 0x000EBDB0000F31BF, 0x000E85B2000E8DB1,
-    0x000F75B4000F6DB3, 0x000F69B6000F7DB5, 0x000E31A3000F61B7,
-    0x8000B9068000B902, 0x8000B90E8000B90A, 0x000E5DA5000E5DA6,
-    0x000E49AB8000B912, 0x8000B916000E79AA, 0x000FB9AF8000B91A,
-    0x000FB9AD000FB1AE, 0x00004D80000E51AC, 0x0000558200005581,
-    0x8000B91E000F55B3, 0x000F6DB6000F65B5, 0x8000B8368000B922,
-    0x000F2DBA8000B926, 0x000F3DBC000F25BB, 0x000F25BE000F25BD,
-    0x000E91A8000F15BF, 0x000E91AA000E91A9, 0x000EA1AC000EA5AB,
-    0x000EA1AE000EA1AD, 0x00018086000EA1AF, 0x8000B92A0002C487,
-    0x8000B9328000B92E, 0x8000B93A8000B936, 0x000169B88000B93E,
-    0x000179BA000169B9, 0x000169BC000179BB, 0x0001DDBE000169BD,
-    0x000EA5B00001D5BF, 0x000EA5B2000EADB1, 0x000EADB4000EBDB3,
-    0x000E95B6000E9DB5, 0x000E19A3000159B7, 0x8000B9468000B942,
-    0x8000B94E8000B94A, 0x000E29A5000E21A6, 0x000E69AB8000B952,
-    0x8000B956000E61AA, 0x000E59AF8000B95A, 0x000E69AD000E69AE,
-    0x8000B95E000E71AC, 0x8000B9668000B962, 0x8000B96E8000B96A,
-    0x8000B9768000B972, 0x00003D808000B97A, 0x0000198200000981,
-    0x8000B9828000B97E, 0x0001B0BE0001B884, 0x0002ADA88000B98A,
-    0x0002D5AA0002D5A9, 0x000339AC0002E9AB, 0x000329AE000339AD,
-    0x8000B98E000329AF, 0x8000B9968000B992, 0x8000B99E8000B99A,
-    0x8000B9A68000B9A2, 0x0003EDB88000B9AA, 0x00038DBA000385B9,
-    0x00039DBC000385BB, 0x00038DBE000385BD, 0x000359B0000385BF,
-    0x0003EDB2000359B1, 0x0003FDB40003E5B3, 0x0003E5B60003E5B5,
-    0x00025DB30003D5B7, 0x0003C8878000B9AE, 0x8000B9B200042C86,
-    0x0003B1B50003B9B6, 0x00039DBB8000B9B6, 0x8000B9BA00039DBA,
-    0x000321BF8000B9BE, 0x000331BD000339BE, 0x00300085000339BC,
-    0x8000B9C2000219A3, 0x0003FDA68000B9C6, 0x8000B9CE8000B9CA,
-    0x0003D9AA0003F5A5, 0x8000B9D20003D9AB, 0x00037DAE8000B9D6,
-    0x00037DAC000365AF, 0x00020C84000375AD, 0x8000B9DE8000B9DA,
-    0x8000B9E2000500BE, 0x8000B9E6000014EF, 0x8000B9EE8000B9EA,
-    0x8000B9F20000C0E3, 0x8000B9FA000014E1, 0x000498868000B9FE,
-    0x00005D8000051087, 0x0000618200006581, 0x8000BA068000BA02,
-    0x8000BA0E8000BA0A, 0x8000BA12000EECEF, 0x8000BA16000FC0E1,
-    0x8000BA1A000E5CE3, 0x8000BA228000BA1E, 0x8000BA2A8000BA26,
-    0x8000BA328000BA2E, 0x8000BA3A8000BA36, 0x8000BA428000BA3E,
-    0x0005B484000EC4EF, 0x8000BA46000E30E1, 0x8000BA4A000EF0E3,
-    0x8000BA528000BA4E, 0x0001E5B38000BA56, 0x8000BA5E8000BA5A,
-    0x8000BA6A8000BA62, 0x0001F5B50001E5B6, 0x000165BB0004D0BE,
-    0x8000BA6E0001D9BA, 0x00015DBF8000BA72, 0x000165BD00015DBE,
-    0x0005FDA800016DBC, 0x00062DAA000625A9, 0x00063DAC000625AB,
-    0x00062DAE000625AD, 0x8000B9F6000625AF, 0x0007E1810007E982,
-    0x8000BA660007E980, 0x000098868000BA76, 0x000695B80003B087,
-    0x0006A1BA00069DB9, 0x000755BC0006A1BB, 0x000749BE00075DBD,
-    0x00065DB0000745BF, 0x0006F9B20006E1B1, 0x0006D9B40006F1B3,
-    0x0006B1B60006D9B5, 0x0006A5A30006B1B7, 0x8000BA7E8000BA7A,
-    0x8000BA868000BA82, 0x0006B5A50006A5A6, 0x000625AB8000BA8A,
-    0x8000BA8E000699AA, 0x00061DAF8000BA92, 0x000625AD00061DAE,
-    0x000795B500062DAC, 0x8000BA9A8000BA96, 0x8000BA9E000795B6,
-    0x000795B38000BAA2, 0x00079DBD8000BAA6, 0x00076DBF00079DBC,
-    0x8000BAAA00076DBE, 0x00078DBB8000B986, 0x000631A800078DBA,
-    0x000645AA000631A9, 0x000645AC00065DAB, 0x000679AE00064DAD,
-    0x8000BAAE000675AF, 0x8000BAB68000BAB2, 0x8000BABE8000BABA,
-    0x8000BAC68000BAC2, 0x0006D1B88000BACA, 0x0006D1BA0006D1B9,
-    0x0006E1BC0006E5BB, 0x0006E1BE0006E1BD, 0x000611B00006E1BF,
-    0x000611B2000611B1, 0x0006F1B4000611B3, 0x0006F1B60006F1B5,
-    0x0006CDA30006F1B7, 0x0000198100001982, 0x8000BACE0000B180,
-    0x0006CDA50006CDA6, 0x0006D5AB8000BAD2, 0x0002A0840006D5AA,
-    0x000635AF0001A0BE, 0x0006C5AD000635AE, 0x8000BADA0006C5AC,
-    0x003F6886000629B3, 0x000159B600012C87, 0x8000BAE28000BADE,
-    0x00017DBA000159B5, 0x8000BAE600014DBB, 0x000139BE8000BAEA,
-    0x000155BC000125BF, 0x0005D1A8000139BD, 0x0005E5AA0005EDA9,
-    0x000295AC000239AB, 0x000295AE00029DAD, 0x00036CBE00028DAF,
-    0x8000BAF28000BAEE, 0x8000BAF60003E084, 0x8000BAFE8000BAFA,
-    0x0002B1B88000BB02, 0x00036DBA000365B9, 0x00037DBC000365BB,
-    0x00036DBE000365BD, 0x0002FDB0000365BF, 0x0002C5B20002CDB1,
-    0x0002B9B40002A9B3, 0x0002ADB60002A5B5, 0x00056DA30002A1B7,
-    0x8000BB0A8000BB06, 0x8000BB128000BB0E, 0x00021DA500021DA6,
-    0x000209AB8000BB16, 0x8000BB1A000239AA, 0x000261AF8000BB1E,
-    0x00027DAD00027DAE, 0x8000BB22000211AC, 0x8000BB2A8000BB26,
-    0x8000BB328000BB2E, 0x8000BB3A8000BB36, 0x000039808000BB3E,
-    0x0000058200003981, 0x003C90868000BB42, 0x8000BB4A003CD887,
-    0x0002D9B38000BB4E, 0x003CC4848000BB52, 0x8000BB56003CC4BE,
-    0x0002F5B5000131B6, 0x000119BB8000BB5A, 0x8000BB5E000111BA,
-    0x0001BDBF8000BB62, 0x0001BDBD0001BDBE, 0x8000BB66000101BC,
-    0x0006BCEF8000BB6A, 0x8000BB728000BB6E, 0x8000BB76003F9884,
-    0x8000BB7E8000BB7A, 0x8000BB868000BB82, 0x0007E4E18000BB8A,
-    0x000668E38000BB8E, 0x8000BB968000BB92, 0x8000BB9E8000BB9A,
-    0x8000BBA20000D8EF, 0x8000BBAA8000BBA6, 0x003F68BE0006B8EF,
-    0x8000BBAE000684E1, 0x8000BBB200061CE3, 0x003E7484000194E1,
-    0x000219A3000004E3, 0x003CE8868000BBBA, 0x00357485003C6C87,
-    0x000235A50001F1A6, 0x0001D9AB8000BBBE, 0x8000BBC20001D1AA,
-    0x00017DAF8000BBC6, 0x00017DAD00017DAE, 0x003DC1A80001C1AC,
-    0x003DD1AA003DD1A9, 0x003DFDAC003DE5AB, 0x003E31AE003E31AD,
-    0x8000BB46003E2DAF, 0x003F8181003F9582, 0x8000BBB6003FB580,
-    0x8000BBCE8000BBCA, 0x003EE9B88000BBD2, 0x003E89BA003EE9B9,
-    0x003E99BC003E89BB, 0x003E89BE003E99BD, 0x003E55B0003E89BF,
-    0x003E55B2003E5DB1, 0x003EF9B4003EE9B3, 0x003EE9B6003EF9B5,
-    0x003FEDB8003EE1B7, 0x003FF5BA003FF1B9, 0x003F15BC003F0DBB,
-    0x003F09BE003F1DBD, 0x003FE5B0003F09BF, 0x003FF9B2003FE9B1,
-    0x003FEDB4003FF5B3, 0x003FDDB6003FD5B5, 0x003E39A8003FD5B7,
-    0x003FCDAA003E39A9, 0x003FCDAC003FDDAB, 0x003FF5AE003FFDAD,
-    0x8000BBD6003F9DAF, 0x8000BBDE8000BBDA, 0x8000BBE68000BBE2,
-    0x8000BAD68000BBEA, 0x003E05B38000BBEE, 0x8000BBF68000BBF2,
-    0x8000BBFE8000BBFA, 0x003E05B5003E3DB6, 0x003E6DBB8000BC02,
-    0x8000BC06003E61BA, 0x003E51BF8000BC0A, 0x003E61BD003E69BE,
-    0x8000BC0E003E6DBC, 0x8000BC12003E41A3, 0x003E79A68000BC16,
-    0x8000BC1E8000BC1A, 0x003E25AA003E41A5, 0x8000BC22003E29AB,
-    0x003E2DAE8000BC26, 0x003E29AC003E15AF, 0x003E45A8003E25AD,
-    0x003E61AA003E59A9, 0x0001A5AC003E61AB, 0x0001A5AE0001ADAD,
-    0x0000F58000017DAF, 0x0000C1820000FD81, 0x000400868000BC2A,
-    0x8000BC2E00011487, 0x0000CDB88000BC32, 0x0000D5BA0000D5B9,
-    0x0000FDBC0000E5BB, 0x000091BE000095BD, 0x000105B0000091BF,
-    0x000105B200010DB1, 0x000105B400011DB3, 0x0000FDB6000109B5,
-    0x003DE9B30000F5B7, 0x8000BC3A8000BC36, 0x8000BC428000BC3E,
-    0x00022DB500022DB6, 0x0002C9BB8000BC46, 0x8000BC4A0002C1BA,
-    0x0002CDBF000368BE, 0x0002C9BD0002C1BE, 0x8000BC4E0002D1BC,
-    0x8000BC52003DADA3, 0x000269A68000BC56, 0x8000BC5E8000BC5A,
-    0x000285AA000269A5, 0x8000BC6200028DAB, 0x000285AE8000BC66,
-    0x000295AC000289AF, 0x8000BC6A00028DAD, 0x8000BC728000BC6E,
-    0x8000BC7A8000BC76, 0x8000BC828000BC7E, 0x00000D808000BC86,
-    0x00001D8200001581, 0x8000BC8E8000BC8A, 0x0003E4878000BC92,
-    0x8000BC9A00043C86, 0x8000BCA28000BC9E, 0x000204EF8000BCA6,
-    0x8000BCAE8000BCAA, 0x003F48E38000BCB2, 0x000140E18000BCB6,
-    0x8000BCBE8000BCBA, 0x8000BCC68000BCC2, 0x0002BDB38000BCCA,
-    0x8000BCD28000BCCE, 0x8000BCDA8000BCD6, 0x000111B5000119B6,
-    0x000161BB8000BCDE, 0x8000BCE200010DBA, 0x000169BF8000BCE6,
-    0x000175BD000161BE, 0x8000BCEA000175BC, 0x8000BCEE003F9CEF,
-    0x8000BCF68000BCF2, 0x00003D8100003D82, 0x8000BCFA00003D80,
-    0x8000BD028000BCFE, 0x000538BE00052084, 0x8000BD0A003F48E3,
-    0x00042086003CACE1, 0x8000BD0E00054C87, 0x8000BD12003ED4EF,
-    0x8000BD1A8000BD16, 0x0004EC84003E68EF, 0x0004E0BE003F4CE1,
-    0x8000BD1E003EB4E3, 0x8000BD22003E64E1, 0x8000BD26003EF8E3,
-    0x000275A38000BD2A, 0x0001D9A50007E084, 0x8000BD328000BD2E,
-    0x8000BD360001D1A6, 0x0001A9AB8000BD3A, 0x0001BDAD0001C5AA,
-    0x0001A1AF0001BDAC, 0x8000BC960001A9AE, 0x8000BD3E8000BD06,
-    0x8000BD468000BD42, 0x8000BD4E8000BD4A, 0x0006E1A88000BD52,
-    0x0006E1AA0006E1A9, 0x0006E5AC0006E1AB, 0x0006E5AE0006EDAD,
-    0x0006ADB00006D9AF, 0x0006BDB20006B5B1, 0x0006ADB40006B5B3,
-    0x000751B6000751B5, 0x000771B8000751B7, 0x000741BA000779B9,
-    0x000731BC000741BB, 0x000735BE00073DBD, 0x000649B300072DBF,
-    0x8000BD5A8000BD56, 0x8000BD628000BD5E, 0x00064DB500064DB6,
-    0x000651BB8000BD66, 0x8000BD6A000649BA, 0x0007B5BF8000BD6E,
-    0x000641BD000649BE, 0x00001582000649BC, 0x0000618000060DA3,
-    0x000609A600006181, 0x8000BD7A8000BD72, 0x00060DAA000609A5,
-    0x00004886000615AB, 0x00060DAE0000DC87, 0x00060DAC0007F1AF,
-    0x000679A8000605AD, 0x000699AA00068DA9, 0x000689AC000699AB,
-    0x0006A1AE0006A5AD, 0x8000BD760006DDAF, 0x8000BD828000BD7E,
-    0x8000BD8A8000BD86, 0x8000BD928000BD8E, 0x000155B88000BD96,
-    0x000155BA00015DB9, 0x0001D9BC0001C9BB, 0x0001C9BE0001D9BD,
-    0x0006A5B00001C1BF, 0x0006B1B20006B1B1, 0x000695B400068DB3,
-    0x000175B600017DB5, 0x000611B300016DB7, 0x8000BD9E8000BD9A,
-    0x8000BDA68000BDA2, 0x000601B5000625B6, 0x000675BB8000BDAA,
-    0x8000BDAE000661BA, 0x0001B9BF8000BDB2, 0x000659BD000651BE,
-    0x8000BDB6000665BC, 0x8000BDBA000655A3, 0x000661A68000BDBE,
-    0x8000BDC68000BDC2, 0x000625AA000645A5, 0x8000BDCA000631AB,
-    0x000615AE8000BDCE, 0x000621AC0001FDAF, 0x0002D9A800061DAD,
-    0x000321AA000321A9, 0x000321AC000321AB, 0x000321AE000321AD,
-    0x8000BDD2000321AF, 0x8000BDDA8000BDD6, 0x000019808000BDDE,
-    0x0000058200001981, 0x000389B88000BDE2, 0x00039DBA000389B9,
-    0x0003B9BC000395BB, 0x000079BE0003B9BD, 0x0003E5B0000079BF,
-    0x0003E5B20003EDB1, 0x0003EDB40003FDB3, 0x0003D5B60003DDB5,
-    0x0003CCBE0003BDB7, 0x8000BDEA003B40E1, 0x8000BDEE0016D8E3,
-    0x8000BDF68000BDF2, 0x8000BDFA0002E4EF, 0x8000BDFE000054E1,
-    0x8000BE02000770E3, 0x8000BE0A8000BE06, 0x000349BB000A60EF,
-    0x000DE886000341BA, 0x000379BF00036C87, 0x000349BD000349BE,
-    0x0003C9B3000351BC, 0x8000BE128000BE0E, 0x000CEC848000BE16,
-    0x0003C9B50003C1B6, 0x000509AC8000BE1A, 0x000519AE000509AD,
-    0x0006BDA8000511AF, 0x0006A5AA0006ADA9, 0x8000BE1E000519AB,
-    0x8000BE268000BE22, 0x8000BE358000BE31, 0x8000BE3D8000BE39,
-    0x0005D1BC8000BE41, 0x0005D1BE0005D1BD, 0x0005D1B80005D1BF,
-    0x0005D1BA0005D1B9, 0x000535B40005D1BB, 0x000501B600053DB5,
-    0x000559B0000501B7, 0x000531B2000535B1, 0x000618E1000531B3,
-    0x000638E3000720E1, 0x8000BE450006F0E3, 0x8000BE4D8000BE49,
-    0x00003D808000BE51, 0x0000198200000981, 0x8000BE598000BE55,
-    0x0006E8EF8000BE61, 0x0002E9A3000634EF, 0x000FE8868000BE65,
-    0x8000BE69000CB887, 0x0002E9A50002E1A6, 0x000269AB8000BE6D,
-    0x000F64BE000261AA, 0x000259AF8000BE71, 0x000269AD000269AE,
-    0x001C01A9000271AC, 0x001C1DAB001859A8, 0x001001AD001C01AA,
-    0x00119DAF001D41AC, 0x000005A100101DAE, 0x000401A38000BDE6,
-    0x000405A500007DA2, 0x001839A7000411A4, 0x00092DA0001801A6,
-    0x000C4DA2000C45A1, 0x8000BE5D000D6DA3, 0x8000BE798000BE75,
-    0x001401B18000BE7D, 0x006801B3001401B0, 0x0068F5B5001489B2,
-    0x8000BE81006811B4, 0x00384D908000BE85, 0x0038999200388591,
-    0x003C4D9400394993, 0x003DB596003CB595, 0x0030819800307D97,
-    0x00345D9A00300D99, 0x0034819C00348D9B, 0x0008999E00084D9D,
-    0x8000BE8900088D9F, 0x000E69B5000E6DB6, 0x000E45B38000BE8D,
-    0x8000BE958000BE91, 0x000FC1BF8000BE99, 0x000E05BD000E19BE,
-    0x000E15BB000E05BC, 0x8000BE9D000E4DBA, 0x002509820024F58F,
-    0x8000BEA10039C983, 0x0038D5868000BEA5, 0x00391984003DC987,
-    0x003D018A00390D85, 0x8000BEA9003C098B, 0x0031998E8000BEAD,
-    0x0031D98C0035C18F, 0x0035619200311D8D, 0x8000BEB100357593,
-    0x000961968000BEB5, 0x0009A19400083997, 0x000D619A00097D95,
-    0x8000BEB9000D619B, 0x0000D5818000BEBD, 0x000C599C0000D580,
-    0x000F09A80000D582, 0x000F1DAA000F15A9, 0x000F31AC000F15AB,
-    0x000F29AE000F3DAD, 0x8000BEC1000F29AF, 0x8000BEC98000BEC5,
-    0x00003CBE00012484, 0x000718868000BED1, 0x000EEDB80000FC87,
-    0x000EFDBA000EF5B9, 0x000EEDBC000EF5BB, 0x000E4DBE000E45BD,
-    0x000F59B0000E45BF, 0x000F01B2000F01B1, 0x000F0DB4000F01B3,
-    0x000EDDB6000ED5B5, 0x000C95A3000ED5B7, 0x8000BED98000BED5,
-    0x8000BEE18000BEDD, 0x000CB9A5000CBDA6, 0x000CC5AB00039C84,
-    0x8000BEE5000C9DAA, 0x000D11AF8000BEE9, 0x000CD5AD000CC9AE,
-    0x8000BEED000CD5AC, 0x8000BEF1000D95B3, 0x000205B68000BEF5,
-    0x8000BEFD8000BEF9, 0x000219BA000DB1B5, 0x8000BF01000219BB,
-    0x0002A5BE8000BF05, 0x0002B9BC0002ADBF, 0x000D35A80002B1BD,
-    0x000D59AA000D41A9, 0x000D7DAC000D51AB, 0x0002C1AE0002C1AD,
-    0x8000BF090002C1AF, 0x8000BF0D0004AC84, 0x8000BF158000BF11,
-    0x8000BF1D8000BF19, 0x0003C9B88000BF21, 0x0003E1BA0003C9B9,
-    0x000399BC0003E1BB, 0x0003B1BE000399BD, 0x000299B00003B1BF,
-    0x000229B2000299B1, 0x000225B4000221B3, 0x000225B600022DB5,
-    0x000CD1A3000205B7, 0x000504BE8000BF25, 0x8000BF2D8000BF29,
-    0x000CF5A5000341A6, 0x00035DAB8000BF31, 0x8000BF3500035DAA,
-    0x0003E9AF8000BF39, 0x0003F5AD0003E1AE, 0x00005D800003FDAC,
-    0x0000658200006D81, 0x8000BF41000145B3, 0x000171B6000145B5,
-    0x0004C0868000BF45, 0x000121BA00029887, 0x000125BC000121BB,
-    0x000111BE00012DBD, 0x8000BF49000111BF, 0x8000BF518000BF4D,
-    0x8000BF598000BF55, 0x000E88EF8000BF5D, 0x000F70E18000BF61,
-    0x000F48E38000BF65, 0x8000BF6D8000BF69, 0x8000BF758000BF71,
-    0x8000BF790004B0BE, 0x8000BF818000BF7D, 0x8000BF898000BF85,
-    0x8000BF8D000E04E1, 0x8000BF91000060E3, 0x8000BF998000BF95,
-    0x8000BF9D0000A4EF, 0x0007D5810007C582, 0x0005B5A80007DD80,
-    0x000601AA000601A9, 0x000601AC000601AB, 0x000635AE00063DAD,
-    0x8000BFA100065DAF, 0x8000BECD8000BF3D, 0x8000BFA98000BFA5,
-    0x000098868000BFAD, 0x0006EDB80003FC87, 0x0006FDBA0006F5B9,
-    0x0006EDBC0006F5BB, 0x000755BE000751BD, 0x000629B000074DBF,
-    0x00063DB2000629B1, 0x000619B4000635B3, 0x0006D9B6000619B5,
-    0x000689A30006D5B7, 0x8000BFB58000BFB1, 0x8000BFBD8000BFB9,
-    0x000689A50006BDA6, 0x0006EDAB8000BFC1, 0x8000BFC50006EDAA,
-    0x0006DDAF8000BFC9, 0x0006E1AD0006DDAE, 0x8000BFCD0006E9AC,
-    0x8000BFD1000745B3, 0x00075DB68000BFD5, 0x8000BFDD8000BFD9,
-    0x00077DBA000755B5, 0x8000BFE1000741BB, 0x000745BE8000BFE5,
-    0x000755BC00074DBF, 0x000651A800074DBD, 0x000689AA000659A9,
-    0x000699AC000689AB, 0x000689AE000699AD, 0x8000BFE9000689AF,
-    0x8000BFF18000BFED, 0x8000BFF98000BFF5, 0x8000C0018000BFFD,
-    0x000689B88000C005, 0x000699BA000689B9, 0x0006B1BC000695BB,
-    0x0006B1BE0006B1BD, 0x0006FDB00006ADBF, 0x0006C5B20006C1B1,
-    0x0006C5B40006DDB3, 0x0006B9B60006CDB5, 0x000601A30006B9B7,
-    0x8000C00D8000C009, 0x8000C0158000C011, 0x000611A5000619A6,
-    0x000605AB8000C019, 0x0002A884000639AA, 0x000609AF0001A0BE,
-    0x000609AD000601AE, 0x00004D80000611AC, 0x0000558200005581,
-    0x8000C021000141B3, 0x000171B6000149B5, 0x000040868000C025,
-    0x000121BA00010487, 0x000125BC000121BB, 0x000111BE00012DBD,
-    0x000285A8000111BF, 0x000285AA00028DA9, 0x00028DAC00029DAB,
-    0x0002B5AE0002BDAD, 0x006800BF0002DDAF, 0x8000C02D8000C029,
-    0x8000C0358000C031, 0x8000C03D8000C039, 0x000369B88000C041,
-    0x000379BA000369B9, 0x000369BC000379BB, 0x0003DDBE000369BD,
-    0x0002A9B00003D5BF, 0x0002BDB20002A9B1, 0x000299B40002B5B3,
-    0x000359B6000299B5, 0x00020DA3000359B7, 0x8000C0498000C045,
-    0x8000C0518000C04D, 0x000205A500023DA6, 0x00026DAB8000C055,
-    0x8000C05900026DAA, 0x00025DAF8000C05D, 0x000261AD00025DAE,
-    0x8000C061000269AC, 0x8000C0698000C065, 0x8000C0718000C06D,
-    0x8000C0798000C075, 0x00000D808000C07D, 0x00001D8200001581,
-    0x8000C0858000C081, 0x001C64BE8000C089, 0x0002D1A88000C091,
-    0x00012DAA0002D9A9, 0x00013DAC000125AB, 0x00012DAE000125AD,
-    0x8000C095000125AF, 0x8000C09D8000C099, 0x8000C0A58000C0A1,
-    0x8000C0AD8000C0A9, 0x000181B88000C0B1, 0x000181BA000181B9,
-    0x000181BC000181BB, 0x000181BE000181BD, 0x00015DB0000181BF,
-    0x0001E1B20001E1B1, 0x0001E1B40001E1B3, 0x0001D1B60001E9B5,
-    0x8000C0B50001D1B7, 0x0000DCEF00028884, 0x8000C0BD8000C0B9,
-    0x0006DCEF000664EF, 0x0006D8E18000C0C1, 0x00071CE30007A8E3,
-    0x000190E100040CE1, 0x000000E38000C0C5, 0x0001E5B38000C0C9,
-    0x001D88878000C0CD, 0x8000C0D1001DEC86, 0x0001F5B50001E5B6,
-    0x000165BB8000C0D5, 0x8000C0D90001D9BA, 0x00016DBF8000C0DD,
-    0x000175BD000165BE, 0x8000C08D000175BC, 0x001E25A5001E35A6,
-    0x001E35A38000C0E1, 0x8000C0E98000C0E5, 0x001EBDAF8000C0ED,
-    0x001EA5AD001EB5AE, 0x001EB5AB001EA5AC, 0x8000C0F1001E09AA,
-    0x000079808000C01D, 0x0000198200000981, 0x8000C0F98000C0F5,
-    0x000390BE00039884, 0x000340868000C0FD, 0x8000C1010003C487,
-    0x8000C1098000C105, 0x8000C1118000C10D, 0x001E7DA88000C115,
-    0x001FADAA001FA5A9, 0x001FBDAC001FA5AB, 0x001FADAE001FA5AD,
-    0x8000C119001FA5AF, 0x8000C1218000C11D, 0x8000C1298000C125,
-    0x8000C1318000C12D, 0x001F11B88000C135, 0x001F21BA001F19B9,
-    0x0000DDBC001F21BB, 0x0000C1BE0000C1BD, 0x001FDDB00000C1BF,
-    0x001F65B2001F65B1, 0x001F6DB4001F7DB3, 0x001F45B6001F41B5,
-    0x001E25B3001F39B7, 0x8000C13D8000C139, 0x8000C1458000C141,
-    0x001E35B5001E25B6, 0x001EA1BB8000C149, 0x8000C14D001E19BA,
-    0x001E85BF8000C151, 0x001EA9BD001EA9BE, 0x8000C155001EB1BC,
-    0x8000C159001E61A3, 0x001E61A68000C15D, 0x8000C1658000C161,
-    0x001E5DAA001E71A5, 0x8000C169001EE5AB, 0x001EEDAE0001A084,
-    0x001EF5AC001EC1AF, 0x0001E9A8001EEDAD, 0x0001F9AA0001E9A9,
-    0x0001E9AC0001F9AB, 0x000139AE0001E9AD, 0x0000F580000139AF,
-    0x0000C1820000FD81, 0x00019086000164BE, 0x8000C1710000A087,
-    0x0000CDB88000C175, 0x0000D5BA0000D5B9, 0x0000FDBC0000E5BB,
-    0x000091BE000095BD, 0x000149B0000091BF, 0x000159B2000149B1,
-    0x000149B4000159B3, 0x0000FDB6000149B5, 0x000271B30000F5B7,
-    0x8000C17D8000C179, 0x8000C1858000C181, 0x000261B5000269B6,
-    0x00028DBB8000C189, 0x8000C18D00028DBA, 0x00028DBF8000C191,
-    0x00028DBD00028DBE, 0x00735C8500028DBC, 0x8000C195000235A3,
-    0x00022DA68000C199, 0x8000C1A18000C19D, 0x0002C9AA000225A5,
-    0x8000C1A50002C9AB, 0x0002C9AE8000C1A9, 0x0002C9AC0002C9AF,
-    0x000339A80002C9AD, 0x00008DAA000339A9, 0x00009DAC000085AB,
-    0x000085AE000085AD, 0x000428BE0000B5AF, 0x8000C1B18000C1AD,
-    0x8000C1B98000C1B5, 0x8000C1C18000C1BD, 0x000069B88000C1C5,
-    0x000079BA000069B9, 0x000069BC000079BB, 0x0001D9BE000069BD,
-    0x0000CDB00001D9BF, 0x0000A5B20000ADB1, 0x0000A9B40000B9B3,
-    0x000095B600009DB5, 0x8000C1C9000059B7, 0x8000C1D18000C1CD,
-    0x00001D818000C1D5, 0x8000C1D900001D80, 0x0002AC8400000582,
-    0x0002C4878000C1E1, 0x8000C1E500050C86, 0x8000C1ED8000C1E9,
-    0x000145B38000C1F1, 0x8000C1F98000C1F5, 0x8000C2018000C1FD,
-    0x000155B5000145B6, 0x000129BB8000C205, 0x8000C209000149BA,
-    0x000119BF8000C20D, 0x00012DBD00012DBE, 0x001F80EF000131BC,
-    0x8000C211001E50E1, 0x8000C2150001C8E3, 0x8000C21D8000C219,
-    0x00052C848000C221, 0x000424BE8000C225, 0x001E24E38000C229,
-    0x001C40E18000C22D, 0x8000C23100011CEF, 0x0004E8868000C235,
-    0x0000ACE300048C87, 0x0001A0E18000C239, 0x0000B4EF8000C23D,
-    0x00001D8100001D82, 0x8000C24100001D80, 0x8000C2498000C245,
-    0x8000C16D8000C251, 0x8000C255000601A3, 0x000601A68000C1DD,
-    0x8000C2598000C24D, 0x00060DAA000611A5, 0x8000C25D00066DAB,
-    0x000669AE8000C261, 0x000675AC00065DAF, 0x000775B3000669AD,
-    0x8000C2698000C265, 0x8000C2718000C26D, 0x000765B5000711B6,
-    0x00073DBB8000C275, 0x8000C279000735BA, 0x0007E1BF8000C27D,
-    0x000715BD0007FDBE, 0x0007A5A8000715BC, 0x0007BDAA0007B5A9,
-    0x0007ADAC0007B5AB, 0x0007D1AE0007D1AD, 0x8000C2810007D1AF,
-    0x8000C2898000C285, 0x8000C2918000C28D, 0x8000C2998000C295,
-    0x00075DB88000C29D, 0x000765BA000761B9, 0x000711BC000779BB,
-    0x00070DBE000711BD, 0x0007B1B0000701BF, 0x000781B20007B9B1,
-    0x000779B4000781B3, 0x00076DB6000765B5, 0x000635A3000765B7,
-    0x8000C2A58000C2A1, 0x8000C2AD8000C2A9, 0x000625A5000651A6,
-    0x00067DAB8000C2B1, 0x8000C2B5000675AA, 0x0006A1AF8000C2B9,
-    0x000655AD0006BDAE, 0x00005D80000655AC, 0x0000658200006D81,
-    0x000198BE00018084, 0x0001D8878000C2BD, 0x8000C2C10000FC86,
-    0x8000C2C98000C2C5, 0x8000C2D18000C2CD, 0x8000C2D98000C2D5,
-    0x00067DA88000C2DD, 0x000191AA00018DA9, 0x0001BDAC0001A5AB,
-    0x0001D1AE0001D1AD, 0x0001E0840001D1AF, 0x8000C2E58000C2E1,
-    0x8000C2ED8000C2E9, 0x8000C2F58000C2F1, 0x000171B88000C2F9,
-    0x000171BA000171B9, 0x0000D5BC000171BB, 0x0000D5BE0000DDBD,
-    0x0001B1B00000CDBF, 0x000181B20001B9B1, 0x000151B4000181B3,
-    0x000151B6000151B5, 0x000219B3000151B7, 0x8000C3018000C2FD,
-    0x8000C3098000C305, 0x000235B5000251B6, 0x00027DBB8000C30D,
-    0x8000C311000275BA, 0x000295BF8000C315, 0x000255BD000295BE,
-    0x8000C319000255BC, 0x8000C31D00025DA3, 0x000215A68000C321,
-    0x8000C3298000C325, 0x000231AA000271A5, 0x000324BE000239AB,
-    0x0002D1AE8000C331, 0x000211AC0002D1AF, 0x0000CD80000211AD,
-    0x0000D1820000D181, 0x8000C3350003E5B3, 0x0003EDB60003E5B5,
-    0x8000C33D8000C339, 0x00030DBA8000C341, 0x00031DBC000305BB,
-    0x00030DBE000305BD, 0x000C4086000305BF, 0x8000C34500030887,
-    0x000194E18000C349, 0x000168E38000C34D, 0x8000C3558000C351,
-    0x000278EF8000C359, 0x8000C3618000C35D, 0x8000C3698000C365,
-    0x000261A88000C36D, 0x000261AA000261A9, 0x000261AC000261AB,
-    0x000261AE000261AD, 0x000D6C84000261AF, 0x8000C3758000C371,
-    0x8000C37D8000C379, 0x8000C3858000C381, 0x000149B88000C389,
-    0x000159BA000149B9, 0x000179BC000151BB, 0x000119BE000179BD,
-    0x0002C5B0000115BF, 0x0002C5B20002CDB1, 0x0002C5B40002DDB3,
-    0x0002C5B60002CDB5, 0x000630E3000179B7, 0x0006BCE100042CE1,
-    0x000640E10006BCE3, 0x0000C4E38000C38D, 0x8000C3958000C391,
-    0x000000EF8000C399, 0x0006E0EF8000C3A1, 0x000DF0BE8000C3A5,
-    0x0002A9A30007E4EF, 0x0000698100006182, 0x8000C3A900005180,
-    0x0002A9A50002A1A6, 0x000249AB8000C3AD, 0x000C4886000241AA,
-    0x000249AF000CCC87, 0x000249AD000241AE, 0x8000C39D000251AC,
-    0x8000C3B58000C3B1, 0x8000C32D8000C3B9, 0x8000C3C18000C3BD,
-    0x000E4DA88000C3C5, 0x000EB5AA000EB5A9, 0x000E95AC000E8DAB,
-    0x000EF5AE000EFDAD, 0x000E9DB0000EE9AF, 0x000EADB2000EA5B1,
-    0x000EA1B4000EA1B3, 0x000EA1B6000EA1B5, 0x000F45B8000EA1B7,
-    0x000F45BA000F4DB9, 0x000F45BC000F5DBB, 0x000F45BE000F4DBD,
-    0x000E1DB3000FE9BF, 0x8000C3CD8000C3C9, 0x8000C3D58000C3D1,
-    0x000E3DB5000E5DB6, 0x000E41BB8000C3D9, 0x8000C3DD000E79BA,
-    0x000E45BF8000C3E1, 0x000E55BD000E45BE, 0x8000C3E5000E59BC,
-    0x8000C3E9000E59A3, 0x000E19A68000C3ED, 0x8000C3F58000C3F1,
-    0x000E3DAA000E79A5, 0x8000C3F9000E05AB, 0x000E01AE8000C3FD,
-    0x000E1DAC000E01AF, 0x000EADA8000E11AD, 0x000ED9AA000EB1A9,
-    0x000115AC000ED1AB, 0x000115AE00011DAD, 0x00000D80000109AF,
-    0x00001D8200001581, 0x8000C4058000C401, 0x000274848000C409,
-    0x0001EDB8000274BE, 0x000181BA000185B9, 0x000185BC000181BB,
-    0x0001B1BE00018DBD, 0x000179B00001B1BF, 0x0001EDB2000179B1,
-    0x0001FDB40001E5B3, 0x0001E5B60001E5B5, 0x000E51B30001D5B7,
-    0x0001A8868000C40D, 0x8000C4110001CC87, 0x00019DB500019DB6,
-    0x0001BDBB8000C415, 0x8000C4190001BDBA, 0x000121BF8000C41D,
-    0x000131BD000139BE, 0x8000C421000139BC, 0x8000C425000E15A3,
-    0x0001D9A68000C429, 0x8000C4318000C42D, 0x0001F9AA0001D9A5,
-    0x8000C4350001F9AB, 0x00017DAE8000C439, 0x00017DAC000165AF,
-    0x000190E1000175AD, 0x000254E38000C43D, 0x8000C4458000C441,
-    0x8000C44D8000C449, 0x8000C4558000C451, 0x8000C45D8000C459,
-    0x8000C4658000C461, 0x00029CEF8000C469, 0x000011808000C46D,
-    0x0000118200001181, 0x8000C47100038884, 0x8000C4798000C475,
-    0x00048086000380BE, 0x8000C4810003C487, 0x8000C4898000C485,
-    0x8000C4918000C48D, 0x000311A88000C495, 0x000321AA000319A9,
-    0x000351AC000321AB, 0x000355AE00035DAD, 0x8000C499000349AF,
-    0x8000C4A18000C49D, 0x8000C4A98000C4A5, 0x8000C4B18000C4AD,
-    0x0000DDB88000C4B5, 0x0000EDBA0000E5B9, 0x0000FDBC0000E5BB,
-    0x0000EDBE0000E5BD, 0x000339B00000E1BF, 0x000305B200030DB1,
-    0x0000FDB40000E9B3, 0x0000EDB60000E5B5, 0x8000C4B90000E5B7,
-    0x0002A1B38000C4BD, 0x0002D1B58000C4C1, 0x8000C4C98000C4C5,
-    0x000000C10002D5B6, 0x0002C1BB8000C4CD, 0x000135BD0002F5BA,
-    0x000129BF0002D9BC, 0x8000C4D1000131BE, 0x000289A500028DA6,
-    0x0002F9A300024084, 0x8000C4D98000C4D5, 0x000171AF8000C4DD,
-    0x00016DAD000169AE, 0x000299AB000281AC, 0x8000C4E10002ADAA,
-    0x8000C4E98000C4E5, 0x000584BE000E74E1, 0x0004DC840000A4E3,
-    0x8000C4F1000F4CE1, 0x0000D980000EACE3, 0x000029820000D981,
-    0x8000C4F5000EECEF, 0x8000C4FD8000C4F9, 0x0001B1B3000078EF,
-    0x0004688600048CBE, 0x8000C50100051C87, 0x0001A1B50001A9B6,
-    0x00014DBB8000C505, 0x8000C50900014DBA, 0x00014DBF8000C50D,
-    0x00014DBD00014DBE, 0x00070DB800014DBC, 0x000711BA000711B9,
-    0x000751BC000729BB, 0x000741BE000741BD, 0x000689B0000741BF,
-    0x000699B2000689B1, 0x000689B4000699B3, 0x000739B6000689B5,
-    0x000629A8000739B7, 0x000639AA000629A9, 0x000629AC000639AB,
-    0x0006F9AE000629AD, 0x8000C4ED0006F9AF, 0x8000C5118000C47D,
-    0x8000C5198000C515, 0x8000C5218000C51D, 0x0006F1A38000C525,
-    0x8000C52D8000C529, 0x8000C5358000C531, 0x0006E1A50006E9A6,
-    0x00060DAB8000C539, 0x8000C53D00060DAA, 0x00060DAF8000C541,
-    0x00060DAD00060DAE, 0x8000C54500060DAC, 0x8000C54D8000C549,
-    0x00001D808000C551, 0x0000398200000981, 0x8000C5598000C555,
-    0x000048868000C55D, 0x8000C5610000CC87, 0x8000C5698000C565,
-    0x000699A88000C56D, 0x0006BDAA0006B5A9, 0x0006C1AC0006B5AB,
-    0x0006F1AE0006C9AD, 0x8000C5710006F1AF, 0x8000C5798000C575,
-    0x8000C5818000C57D, 0x8000C5898000C585, 0x0001CDB88000C58D,
-    0x0001D5BA0001D5B9, 0x0001FDBC0001E5BB, 0x00019DBE000195BD,
-    0x000109B0000195BF, 0x000119B2000109B1, 0x000109B4000119B3,
-    0x0001FDB6000109B5, 0x000635B30001F5B7, 0x8000C5958000C591,
-    0x8000C59D8000C599, 0x000649B5000649B6, 0x0001CDBB8000C5A1,
-    0x8000C5A50001CDBA, 0x0001CDBF8000C5A9, 0x0001CDBD0001CDBE,
-    0x8000C5AD0001CDBC, 0x8000C5B1000671A3, 0x00060DA68000C5B5,
-    0x8000C5BD8000C5B9, 0x000189AA00060DA5, 0x8000C5C1000189AB,
-    0x000189AE8000C5C5, 0x000189AC000189AF, 0x0002F1A8000189AD,
-    0x0002F1AA0002F1A9, 0x000331AC0002F1AB, 0x000331AE000331AD,
-    0x0003C980000331AF, 0x000005820003C981, 0x8000C5D18000C5C9,
-    0x8000C5D500027884, 0x0003DDB88000C5D9, 0x0003E5BA0003EDB9,
-    0x000399BC00038DBB, 0x00038DBE000399BD, 0x000351B0000385BF,
-    0x000351B2000351B1, 0x0003F5B4000351B3, 0x0003F5B60003FDB5,
-    0x000279B30003EDB7, 0x0003C8878000C5DD, 0x8000C5E1007C6C86,
-    0x0003B9B50003B1B6, 0x000395BB8000C5E5, 0x0002E0BE000395BA,
-    0x000339BF8000C5E9, 0x000385BD000331BE, 0x8000C5ED000385BC,
-    0x8000C5F100023DA3, 0x0003F5A68000C5F5, 0x8000C5FD8000C5F9,
-    0x0003D1AA0003FDA5, 0x8000C6010003D1AB, 0x000375AE8000C605,
-    0x0003C1AC00037DAF, 0x8000C6090003C1AD, 0x8000C6118000C60D,
-    0x8000C6198000C615, 0x8000C6218000C61D, 0x8000C6298000C625,
-    0x8000C6318000C62D, 0x8000C6398000C635, 0x8000C6458000C641,
-    0x00006980007D30BE, 0x0000798200006981, 0x007CF08600020084,
-    0x000308EF007D2487, 0x000190E18000C649, 0x00034CE38000C64D,
-    0x8000C6558000C651, 0x8000C65D8000C659, 0x0001BDB38000C661,
-    0x8000C6698000C665, 0x8000C6718000C66D, 0x000109B5000109B6,
-    0x00016DBB8000C675, 0x8000C67900016DBA, 0x00015DBF8000C67D,
-    0x000161BD00015DBE, 0x8000C681000169BC, 0x8000C6898000C685,
-    0x8000C6918000C68D, 0x0006A8EF8000C695, 0x000798E10006E0EF,
-    0x0006B4E3000634E1, 0x8000C69900066CE3, 0x8000C6A18000C69D,
-    0x007F09B88000C6A5, 0x007F19BA007F09B9, 0x007F09BC007F19BB,
-    0x007F39BE007F09BD, 0x007EE9B0007F39BF, 0x007EFDB2007EE9B1,
-    0x007ED9B4007EF5B3, 0x007F39B6007ED9B5, 0x007E71A8007F39B7,
-    0x007E75AA007E7DA9, 0x007EC1AC007EC9AB, 0x007EF5AE007EC9AD,
-    0x8000C6A9007E9DAF, 0x8000C5CD8000C63D, 0x007F898000026484,
-    0x00001982007F8981, 0x007EBDA3000370BE, 0x000028868000C6AD,
-    0x8000C6B100034C87, 0x007E09A5007E09A6, 0x007E6DAB8000C6B5,
-    0x8000C6B9007E6DAA, 0x007E5DAF8000C6BD, 0x007E61AD007E5DAE,
-    0x8000C6C1007E69AC, 0x8000C6C5007F99B3, 0x007FA1B68000C6C9,
-    0x8000C6D18000C6CD, 0x007F9DBA007FB1B5, 0x8000C6D5007F61BB,
-    0x007F61BE8000C6D9, 0x007F71BC007F69BF, 0x007E35A8007F71BD,
-    0x007E55AA007E3DA9, 0x007E7DAC007E69AB, 0x007E6DAE007E65AD,
-    0x8000C6DD007E65AF, 0x8000C6E58000C6E1, 0x8000C6ED8000C6E9,
-    0x8000C6F58000C6F1, 0x007E81B88000C6F9, 0x007E81BA007E81B9,
-    0x007E81BC007E81BB, 0x007EA9BE007E81BD, 0x007E1DB0007EA9BF,
-    0x007EE1B2007EE1B1, 0x007EE1B4007EE1B3, 0x007ED1B6007EE9B5,
-    0x007EDDA3007ED1B7, 0x8000C7018000C6FD, 0x8000C7098000C705,
-    0x007EF5A5007EE5A6, 0x007E25AB8000C70D, 0x8000C711007ED9AA,
-    0x007E2DAF8000C715, 0x007E35AD007E25AE, 0x00006980007E35AC,
-    0x0000058200006981, 0x000198BE8000C719, 0x0018C0EF00019884,
-    0x000228E10007D885, 0x000434E38000C721, 0x000650868000C725,
-    0x8000C7290000E487, 0x000291A88000C72D, 0x0002A1AA000299A9,
-    0x0002D1AC0002A1AB, 0x0002D5AE0002DDAD, 0x8000C7310002C9AF,
-    0x8000C7398000C735, 0x8000C7418000C73D, 0x8000C7498000C745,
-    0x000375B88000C74D, 0x000375BA00037DB9, 0x0003D9BC0003C9BB,
-    0x0003C9BE0003D9BD, 0x0002B9B00003C1BF, 0x000285B200028DB1,
-    0x000355B400034DB3, 0x000355B600035DB5, 0x000229B300034DB7,
-    0x8000C7510005C884, 0x0006DC848000C755, 0x000259B500025DB6,
-    0x000249BB8000C759, 0x8000C75D00027DBA, 0x000399BF8000C761,
-    0x000399BD000391BE, 0x8000C765000251BC, 0x8000C76900026DA3,
-    0x000219A68000C76D, 0x8000C7758000C771, 0x000239AA00021DA5,
-    0x8000C77900020DAB, 0x0003D5AE8000C77D, 0x000215AC0003DDAF,
-    0x000190E10003DDAD, 0x0000CCE38000C781, 0x00001580000444BE,
-    0x00003D8200000981, 0x006B058F8000C785, 0x000760BE8000C789,
-    0x8000C7958000C791, 0x000010EF8000C799, 0x007741978000C79D,
-    0x0073499500731996, 0x006F659300737194, 0x006F7191006F7192,
-    0x007F019F006B0D90, 0x007AED9D007F019E, 0x007B619B007B6D9C,
-    0x00776D990076219A, 0x0007208600777198, 0x8000C7A100024487,
-    0x000315A38000C7A5, 0x0003F1A10003F1A2, 0x00080084007F79A0,
-    0x8000C7AD8000C7A9, 0x8000C7B58000C7B1, 0x8000C7BD8000C7B9,
-    0x000795808000C7C1, 0x0007958200079D81, 0x8000C7C98000C7C5,
-    0x00015CEF007E10EF, 0x007EA8E18000C7CD, 0x000194E3007EB8E3,
-    0x8000C7D1007F78E1, 0x8000C7D98000C7D5, 0x8000C71D8000C7DD,
-    0x0074098A8000C78D, 0x0070118C0074518B, 0x00700D8E0070198D,
-    0x8000C7E10071498F, 0x006C0D828000C7E5, 0x00681984006C7D83,
-    0x00680D8600681185, 0x8000C7E900740187, 0x0004219A8000C7ED,
-    0x0004D59C0004299B, 0x8000C7F58000C7F1, 0x8000C7FD8000C7F9,
-    0x007C79928000C801, 0x007DF594007C6D93, 0x0078199600781195,
-    0x000081A100789997, 0x0000E9A38000C805, 0x000C99A5000089A2,
-    0x000801A7000C91A4, 0x00087DA9000C31A6, 0x001489AB000885A8,
-    0x001001AD001401AA, 0x00108DAF001425AC, 0x001CC1B100109DAE,
-    0x001DA5B3001C01B0, 0x001855B5001C59B2, 0x00017886001855B4,
-    0x8000C80900031487, 0x8000C8118000C80D, 0x8000C8198000C815,
-    0x8000C8218000C81D, 0x0004E5A88000C825, 0x0004F1AA0004E9A9,
-    0x000731AC0004F1AB, 0x000731AE000731AD, 0x8000C829000731AF,
-    0x8000C8318000C82D, 0x8000C8398000C835, 0x8000C8418000C83D,
-    0x0007EDB88000C845, 0x000781BA000785B9, 0x000785BC000781BB,
-    0x0007B1BE00078DBD, 0x000751B00007B1BF, 0x000751B2000751B1,
-    0x0007F9B4000751B3, 0x0007E5B60007F9B5, 0x000479B30007D5B7,
-    0x8000C84D8000C849, 0x8000C8558000C851, 0x0007B9B50007B1B6,
-    0x000795BB8000C859, 0x8000C85D000795BA, 0x00073DBF8000C861,
-    0x000785BD00073DBE, 0x00001982000785BC, 0x0000658000043DA3,
-    0x0007F5A600001981, 0x8000C8698000C865, 0x0007D1AA0007FDA5,
-    0x000F64BE0007D1AB, 0x000779AE8000C871, 0x0007C1AC000779AF,
-    0x00052DA80007C1AD, 0x00053DAA000535A9, 0x000599AC000531AB,
-    0x0005BDAE000599AD, 0x000C00860005B5AF, 0x8000C8750003B087,
-    0x8000C87D8000C879, 0x8000C8858000C881, 0x0005B1B88000C889,
-    0x0005B1BA0005B1B9, 0x000271BC0005B1BB, 0x000219BE000271BD,
-    0x0005CDB0000219BF, 0x0005DDB20005D5B1, 0x0005CDB40005D5B3,
-    0x0005A9B60005B9B5, 0x0005B5B30005A1B7, 0x8000C8918000C88D,
-    0x8000C8998000C895, 0x000511B5000519B6, 0x000511BB8000C89D,
-    0x8000C8A1000509BA, 0x00057DBF8000C8A5, 0x000501BD000509BE,
-    0x000C4884000509BC, 0x8000C8A90005F1A3, 0x00055DA68000C8AD,
-    0x8000C8B58000C8B1, 0x00054DAA000555A5, 0x8000C8B9000555AB,
-    0x00054DAE8000C8BD, 0x00054DAC000539AF, 0x000365A8000545AD,
-    0x0003E9AA000381A9, 0x0003F9AC0003E9AB, 0x0003E9AE0003F9AD,
-    0x8000C8C10003E9AF, 0x8000C8C98000C8C5, 0x8000C8D18000C8CD,
-    0x8000C8D98000C8D5, 0x000279B8000CB0BE, 0x0002CDBA000279B9,
-    0x0002DDBC0002C5BB, 0x0002C5BE0002C5BD, 0x000399B00002F5BF,
-    0x000249B2000399B1, 0x000259B4000249B3, 0x000249B6000259B5,
-    0x000618E1000249B7, 0x000708E300046CE1, 0x0000A581000718E3,
-    0x8000C8E10000A580, 0x000C80860000A582, 0x8000C8E5000D0487,
-    0x8000C8ED8000C8E9, 0x000668EF8000C8F1, 0x8000C8F5000648EF,
-    0x001BF0858000C8F9, 0x8000C9018000C8FD, 0x8000C9098000C905,
-    0x8000C90D0000CCEF, 0x8000C911000124E1, 0x8000C91500065CE3,
-    0x8000C91D8000C919, 0x000CE1A38000C921, 0x8000C8DD8000C86D,
-    0x8000C9298000C925, 0x000CF1A5000CF9A6, 0x000CDDAB8000C92D,
-    0x8000C931000CDDAA, 0x000D1DAF8000C935, 0x000D1DAD000D1DAE,
-    0x8000C939000D1DAC, 0x8000C93D000E69B3, 0x000E69B68000C941,
-    0x8000C9498000C945, 0x000EE9BA000E69B5, 0x0000A484000EE9BB,
-    0x000ED9BE0000A4BE, 0x000EE9BC000ED9BF, 0x000D6DA8000EE1BD,
-    0x000EA5AA000EADA9, 0x000EA5AC000EBDAB, 0x000EA5AE000EADAD,
-    0x0000ED80000E19AF, 0x0000F1820000F181, 0x000090868000C94D,
-    0x8000C95100001487, 0x000E15B88000C955, 0x000E21BA000E1DB9,
-    0x000FF1BC000E21BB, 0x000FF1BE000FF1BD, 0x000E69B0000FF1BF,
-    0x000E79B2000E69B1, 0x000E59B4000E71B3, 0x000E31B6000E59B5,
-    0x000F2DA3000E31B7, 0x8000C95D8000C959, 0x8000C9658000C961,
-    0x000F2DA5000F2DA6, 0x000FADAB8000C969, 0x8000C96D000FADAA,
-    0x000F9DAF8000C971, 0x000FA5AD000F9DAE, 0x8000C975000FADAC,
-    0x000DB1BA8000C979, 0x000C71BC000DB1BB, 0x000C71BE000C71BD,
-    0x8000C97D000C71BF, 0x8000C9858000C981, 0x8000C989000D8DB3,
-    0x000D95B6000D9DB5, 0x000E19A88000C98D, 0x000E25AA000E2DA9,
-    0x000E25AC000E3DAB, 0x000E5DAE000E55AD, 0x8000C991000E51AF,
-    0x8000C9998000C995, 0x8000C9A18000C99D, 0x8000C9A98000C9A5,
-    0x000E3DB88000C9AD, 0x0001D9BA0001C9B9, 0x0001F9BC0001D1BB,
-    0x000199BE0001F9BD, 0x000E31B0000195BF, 0x000E31B2000E31B1,
-    0x000E19B4000E0DB3, 0x000E0DB6000E05B5, 0x000DADB3000E05B7,
-    0x8000C9B58000C9B1, 0x00039C848000C9B9, 0x000DADB5000DD5B6,
-    0x000DF5BB000350BE, 0x00048886000DF5BA, 0x000DC1BF0003BC87,
-    0x000DD9BD000DDDBE, 0x8000C9C1000DE5BC, 0x8000C9C5000DE9A3,
-    0x000D91A68000C9C9, 0x8000C9D18000C9CD, 0x000DB1AA000DE9A5,
-    0x8000C9D5000DB1AB, 0x000D99AE8000C9D9, 0x000DA1AC000D85AF,
-    0x000239A8000D9DAD, 0x00038DAA000239A9, 0x00039DAC000385AB,
-    0x000385AE000385AD, 0x0000A5800003B5AF, 0x0000A5820000AD81,
-    0x8000C9E18000C9DD, 0x8000C9E98000C9E5, 0x000361B88000C9ED,
-    0x000361BA000361B9, 0x000361BC000361BB, 0x000361BE000361BD,
-    0x0003CDB0000361BF, 0x0003ADB20003A5B1, 0x000395B400038DB3,
-    0x000389B600039DB5, 0x8000C9F1000381B7, 0x8000C9F98000C9F5,
-    0x0003F4EF8000C9FD, 0x8000CA058000CA01, 0x000320E38000CA09,
-    0x000180E18000CA0D, 0x8000CA158000CA11, 0x8000CA1D8000CA19,
-    0x000245B38000CA21, 0x8000CA298000CA25, 0x8000CA3100050484,
-    0x000255B5000245B6, 0x000221BB8000CA35, 0x8000CA39000249BA,
-    0x000219BF000520BE, 0x000219BD000211BE, 0x000FB8E1000239BC,
-    0x000F58E30001FCE3, 0x0000AD80000E34E1, 0x0000D1820000D181,
-    0x0004E8BE8000CA3D, 0x8000CA458000CA41, 0x8000CA4D8000CA49,
-    0x000E7CEF000120EF, 0x0002C5A38000CA51, 0x000468868000CA55,
-    0x8000CA5900050087, 0x0002D5A50002C5A6, 0x0002A1AB8000CA5D,
-    0x8000CA610002C9AA, 0x000299AF8000CA65, 0x000299AD000291AE,
-    0x000641A80002B9AC, 0x000645AA00064DA9, 0x000649AC000659AB,
-    0x000675AE00067DAD, 0x8000CA2D0006B1AF, 0x8000CA6D8000CA69,
-    0x8000CA758000CA71, 0x8000CA7D8000CA79, 0x000755B88000CA81,
-    0x000755BA00075DB9, 0x00077DBC000769BB, 0x00076DBE000765BD,
-    0x0006D1B0000765BF, 0x0006D1B20006D1B1, 0x000771B40006D1B3,
-    0x000771B6000771B5, 0x8000CA85000771B7, 0x8000CA890006C9B3,
-    0x0006F9B68000CA8D, 0x8000C9BD8000CA91, 0x000611BA0006C5B5,
-    0x8000CA95000611BB, 0x0007B1BE8000CA99, 0x000601BC0007B1BF,
-    0x00068DA3000601BD, 0x0000618100006982, 0x00039C8400005980,
-    0x000681A50006BDA6, 0x000655AB8000CA9D, 0x00004886000655AA,
-    0x0007F5AF00006887, 0x000645AD0007F5AE, 0x000651A8000645AC,
-    0x000661AA000659A9, 0x0006B5AC000691AB, 0x0006B5AE0006BDAD,
-    0x8000CAA10006ADAF, 0x8000CAA98000CAA5, 0x8000CAB18000CAAD,
-    0x8000CAB98000CAB5, 0x00016DB88000CABD, 0x00017DBA000175B9,
-    0x00016DBC000175BB, 0x0001D1BE0001D5BD, 0x0006DDB00001D1BF,
-    0x0006B9B20006A1B1, 0x000699B40006B1B3, 0x00015DB6000699B5,
-    0x0006E9B3000155B7, 0x8000CAC58000CAC1, 0x8000CACD8000CAC9,
-    0x00063DB500063DB6, 0x000665BB8000CAD1, 0x8000CAD5000619BA,
-    0x00065DBF8000CAD9, 0x000665BD00065DBE, 0x8000CADD00066DBC,
-    0x8000CAE10006ADA3, 0x000679A68000CAE5, 0x8000CAED8000CAE9,
-    0x00065DAA000679A5, 0x8000CAF1000621AB, 0x000619AE8000CAF5,
-    0x000629AC000619AF, 0x000321A8000621AD, 0x000321AA000321A9,
-    0x000321AC000321AB, 0x000321AE000321AD, 0x8000CAF9000321AF,
-    0x8000CB018000CAFD, 0x8000CB098000CB05, 0x8000CB11000278BE,
-    0x000389B88000CB15, 0x00039DBA000389B9, 0x0003B9BC000395BB,
-    0x000079BE0003B9BD, 0x0003E5B0000079BF, 0x0003E5B20003EDB1,
-    0x0003EDB40003FDB3, 0x0003D5B60003DDB5, 0x0000AD810003BDB7,
-    0x0003A1B30000AD80, 0x0003A1B50000BD82, 0x8000CB1D8000CB19,
-    0x0002E4840003A9B6, 0x000321BB8000CB21, 0x000321BD000329BA,
-    0x000311BF000339BC, 0x0003E9A3000321BE, 0x001C68868000CB25,
-    0x8000CB2900038087, 0x0003E9A50003E1A6, 0x000369AB8000CB2D,
-    0x8000CB31000361AA, 0x000359AF8000CB35, 0x000369AD000369AE,
-    0x8000CB39000371AC, 0x8000CB418000CB3D, 0x8000CB498000CB45,
-    0x8000CB4D000648E3, 0x8000CB5100072CE1, 0x8000CB550006B0EF,
-    0x8000CB5D8000CB59, 0x8000CB658000CB61, 0x8000CB6D8000CB69,
-    0x0000158100002182, 0x0003E0E300001980, 0x0001A8E18000CB71,
-    0x0003ACEF8000CB75, 0x000264848000CB79, 0x001CD086001D60BE,
-    0x8000CB81001D4487, 0x8000CB898000CB85, 0x8000CB918000CB8D,
-    0x8000CB998000CB95, 0x8000CBA18000CB9D, 0x001DA8840006C4EF,
-    0x8000CBA5000674E1, 0x8000CBA900016CE3, 0x8000CBB18000CBAD,
-    0x000155B38000CBB5, 0x8000CBBD8000CBB9, 0x8000CBC58000CBC1,
-    0x00017DB5000111B6, 0x00010DBB8000CBC9, 0x8000CBCD00010DBA,
-    0x0000FDBF8000CBD1, 0x0000FDBD0000FDBE, 0x001E39A80000FDBC,
-    0x001E9DAA001E39A9, 0x001EB5AC001E95AB, 0x001EC9AE001ED1AD,
-    0x8000CB7D001EC1AF, 0x8000CBD98000CBD5, 0x00000D808000CBDD,
-    0x0000B1820000B181, 0x001F61B88000CBE1, 0x001F65BA001F6DB9,
-    0x001F65BC001F7DBB, 0x001F65BE001F6DBD, 0x001E89B0001F1DBF,
-    0x001F69B2001E89B1, 0x001F79B4001F69B3, 0x001F69B6001F79B5,
-    0x001E15A3001F65B7, 0x8000CBE98000CBE5, 0x8000CBED000380BE,
-    0x001E3DA5001E51A6, 0x001E4DAB8000CBF1, 0x8000CBF5001E4DAA,
-    0x001FBDAF8000CB0D, 0x001FBDAD001FBDAE, 0x8000CBF9001FBDAC,
-    0x00006886001F45B3, 0x001F5DB600031C87, 0x8000CC018000CBFD,
-    0x001F79BA001F55B5, 0x8000CC05001F4DBB, 0x001F49BE8000CC09,
-    0x001F49BC001F49BF, 0x001E79A8001F49BD, 0x001E89AA001E79A9,
-    0x001E99AC001E89AB, 0x001E89AE001E99AD, 0x8000CC0D001E89AF,
-    0x8000CC158000CC11, 0x8000CC1D8000CC19, 0x8000CC258000CC21,
-    0x001E8DB88000CC29, 0x001E95BA001E95B9, 0x001EBDBC001EA5BB,
-    0x000171BE000171BD, 0x001EF9B0000171BF, 0x001ED9B2001ECDB1,
-    0x001EC9B4001ED9B3, 0x001EBDB6001EC9B5, 0x001E01A3001EB5B7,
-    0x8000CC318000CC2D, 0x8000CC398000CC35, 0x001E11A5001E19A6,
-    0x001E09AB8000CC3D, 0x8000CC41001E3DAA, 0x001E0DAF8000CC45,
-    0x001E0DAD001E0DAE, 0x8000CC49001E0DAC, 0x8000CC518000CC4D,
-    0x000019808000CC55, 0x0000058200001981, 0x0001AC848000CC59,
-    0x0000E8870001ACBE, 0x8000CC6100074C86, 0x8000CC698000CC65,
-    0x00029DA88000CC6D, 0x0002C9AA0002ADA9, 0x0002C9AC0002D9AB,
-    0x0002F5AE0002FDAD, 0x8000CC7100033DAF, 0x8000CC798000CC75,
-    0x00037CBE8000CC7D, 0x8000CC858000CC81, 0x0003C9B88000CC89,
-    0x0003D9BA0003C9B9, 0x0003F9BC0003D1BB, 0x000399BE0003F9BD,
-    0x000345B0000399BF, 0x000345B200034DB1, 0x000345B400035DB3,
-    0x000345B600034DB5, 0x000241B30003F9B7, 0x8000CC918000CC8D,
-    0x8000CC998000CC95, 0x000241B500027DB6, 0x000389BB8000CC9D,
-    0x8000CCA1000381BA, 0x000389BF8000CCA5, 0x000389BD000381BE,
-    0x8000CCA9000391BC, 0x8000CCAD000205A3, 0x000239A68000CCB1,
-    0x8000CCB98000CCB5, 0x0003C5AA000205A5, 0x8000CCBD0003CDAB,
-    0x0003C5AE8000CCC1, 0x0003D5AC0003CDAF, 0x0001ED800003CDAD,
-    0x00002D8200001D81, 0x001E84EF00028C84, 0x8000CCC50001C4E1,
-    0x001E58E30000ECE3, 0x001FF8E18000CCC9, 0x8000CCD1000030EF,
-    0x8000CCD98000CCD5, 0x000560868000CCDD, 0x8000CCE100056887,
-    0x8000CCE50002B5B3, 0x00017DB6000175B5, 0x8000CCED8000CCE9,
-    0x00015DBA8000CCF1, 0x00014DBC000155BB, 0x0001F1BE0001F1BD,
-    0x8000CCF50001F1BF, 0x8000CCFD8000CCF9, 0x8000CD058000CD01,
-    0x8000CD0D8000CD09, 0x00046C840001E8EF, 0x8000CD11001ECCE1,
-    0x8000CD15001E1CE3, 0x8000CD1D8000CD19, 0x000239A38000CD21,
-    0x8000CD298000CD25, 0x000000F48000CD2D, 0x0001F9A50001F1A6,
-    0x0001D9AB8000CD31, 0x8000CD350001D1AA, 0x00017DAF8000CD39,
-    0x00017DAD00017DAE, 0x0005F1A80001C1AC, 0x0005F1AA0005F1A9,
-    0x000631AC0005F1AB, 0x000631AE000631AD, 0x8000CCCD000631AF,
-    0x0007C1810007C982, 0x0003F8BE0007F180, 0x000388878000CD3D,
-    0x0006E9B800031C86, 0x000689BA0006E9B9, 0x00069DBC000685BB,
-    0x000681BE000681BD, 0x000655B00006B5BF, 0x000655B200065DB1,
-    0x0006F9B40006E9B3, 0x0006E9B60006F9B5, 0x0007B1A80006E1B7,
-    0x0007B1AA0007B1A9, 0x0007D9AC0007B1AB, 0x0007C5AE0007C5AD,
-    0x8000CD410007FDAF, 0x8000CD498000CD45, 0x8000CC5D8000CD4D,
-    0x8000CD558000CD51, 0x000769B88000CD59, 0x00077DBA000775B9,
-    0x00076DBC000775BB, 0x000711BE000711BD, 0x000785B0000711BF,
-    0x00076DB2000765B1, 0x00077DB4000765B3, 0x00076DB6000765B5,
-    0x000639B3000765B7, 0x8000CD618000CD5D, 0x8000CD698000CD65,
-    0x000651B5000645B6, 0x0006A1BB8000CD6D, 0x8000CD71000641BA,
-    0x0006A1BF8000CD75, 0x0006B1BD0006B9BE, 0x8000CD790006B9BC,
-    0x8000CD7D00067DA3, 0x000601A68000CD81, 0x8000CD898000CD85,
-    0x000605AA000615A5, 0x8000CD8D0006E5AB, 0x0006FDAE8000CD91,
-    0x0006FDAC0006E5AF, 0x000121A80006F5AD, 0x000121AA000121A9,
-    0x000125AC000121AB, 0x000125AE00012DAD, 0x8000CD95000099AF,
-    0x8000CD9D8000CD99, 0x00001D808000CDA1, 0x0003B98200000981,
-    0x00008DB88000CDA5, 0x000091BA000091B9, 0x0000BDBC0000A5BB,
-    0x00007DBE000075BD, 0x0000E9B0000075BF, 0x0000F9B20000E9B1,
-    0x0000D9B40000F1B3, 0x0000B9B60000D9B5, 0x8000CDA90000B5B7,
-    0x0002A1B30002C084, 0x0002A5B50002C0BE, 0x8000CDB58000CDB1,
-    0x000204860002A5B6, 0x000325BB00036087, 0x00032DBD000325BA,
-    0x000315BF000325BC, 0x0002E9A3000315BE, 0x8000CDBD8000CDB9,
-    0x8000CDC58000CDC1, 0x0002EDA50002EDA6, 0x00036DAB8000CDC9,
-    0x8000CDCD00036DAA, 0x00035DAF8000CDD1, 0x000365AD00035DAE,
-    0x000099A900036DAC, 0x0000C1AB000091A8, 0x0000C1AD0000C9AA,
-    0x0000F1AF0000D9AC, 0x0003FCBE0000C1AE, 0x8000CDD98000CDD5,
-    0x8000CDE18000CDDD, 0x8000CDE98000CDE5, 0x0001C5B98000CDED,
-    0x0001D9BB0001C5B8, 0x0001C9BD0001C5BA, 0x0001FDBF0001C1BC,
-    0x000045B10001C1BE, 0x000045B3000089B0, 0x000045B5000045B2,
-    0x000045B7000045B4, 0x8000CDF1000045B6, 0x8000CDF98000CDF5,
-    0x8000CDFD0007C4EF, 0x8000CE058000CE01, 0x0002A4848000CE09,
-    0x8000CE118000CE0D, 0x8000CE198000CE15, 0x8000CE210004C0E1,
-    0x00006D81000620E3, 0x00010DB300006D80, 0x00011DB500007D82,
-    0x8000CE298000CE25, 0x000DE08600010DB6, 0x0001CDBB000DE087,
-    0x0001DDBD000101BA, 0x0001CDBF0001DDBC, 0x8000CE2D0001CDBE,
-    0x00037CEF8000CE31, 0x8000CE398000CE35, 0x8000CE418000CE3D,
-    0x8000CE45000154EF, 0x8000CE490006FCE1, 0x000194E100061CE3,
-    0x0003ACE38000CE4D, 0x8000CE558000CE51, 0x000185A5000195A6,
-    0x000195A3000DC084, 0x8000CE5D8000CE59, 0x000155AF8000CE61,
-    0x000145AD000155AE, 0x000155AB000145AC, 0x8000CE65000199AA,
-    0x8000CE1D8000CE69, 0x8000CE718000CE6D, 0x8000CE798000CE75,
-    0x8000CE818000CE7D, 0x000F19A88000CE85, 0x000F71AA000F71A9,
-    0x000FB5AC000FADAB, 0x000FA9AE000FB9AD, 0x000FD9B0000FA9AF,
-    0x000FE1B2000FE5B1, 0x000FEDB4000FF9B3, 0x000F8DB6000F95B5,
-    0x000FBDB8000F85B7, 0x000F95BA000F9DB9, 0x000079BC000069BB,
-    0x000069BE000079BD, 0x00009D81000069BF, 0x8000CE8900009D80,
-    0x00031CBE00000582, 0x8000CE918000CE8D, 0x000380868000CE95,
-    0x8000CE990003F487, 0x8000CEA18000CE9D, 0x8000CEA98000CEA5,
-    0x000F99B38000CEAD, 0x8000CEB18000CDAD, 0x8000CEB98000CEB5,
-    0x000F5DB5000F59B6, 0x000F41BB8000CEBD, 0x8000CEC1000F7DBA,
-    0x000F31BF8000CEC5, 0x000F4DBD000F49BE, 0x000E59A3000F55BC,
-    0x8000CECD8000CEC9, 0x8000CED58000CED1, 0x000E9DA5000E99A6,
-    0x000E81AB8000CED9, 0x8000CEDD000EBDAA, 0x000EF1AF8000CEE1,
-    0x000E8DAD000E89AE, 0x8000CEE5000E95AC, 0x8000CEED8000CEE9,
-    0x00001D808000CEF1, 0x0000398200000981, 0x8000CEF98000CEF5,
-    0x0001B0BE0001A484, 0x000400868000CEFD, 0x8000CF010001F487,
-    0x000119A88000CF05, 0x000121AA000119A9, 0x000121AC000121AB,
-    0x000121AE000121AD, 0x8000CF09000121AF, 0x8000CF118000CF0D,
-    0x8000CF198000CF15, 0x8000CF218000CF1D, 0x0000CDB88000CF25,
-    0x0000D5BA0000D5B9, 0x0000FDBC0000E5BB, 0x00009DBE000095BD,
-    0x000129B0000095BF, 0x000139B2000129B1, 0x000119B4000131B3,
-    0x0000FDB6000119B5, 0x8000CF290000F5B7, 0x8000CF318000CF2D,
-    0x8000CF350002A5B3, 0x0002ADB60002A5B5, 0x8000CF3D8000CF39,
-    0x0002E1BA8000CF41, 0x0002E1BC0002E1BB, 0x0002D1BE0002E9BD,
-    0x0002F5A60002D1BF, 0x8000CF498000CF45, 0x8000CF4D0002FDA5,
-    0x8000CF510002FDA3, 0x000289AE8000CF55, 0x0002B9AC000289AF,
-    0x0002B9AA0002B1AD, 0x8000CF590002B9AB, 0x00000D808000CF5D,
-    0x00001D8200001581, 0x8000CF658000CF61, 0x000414BE8000CF69,
-    0x8000CF7100039C84, 0x8000CF798000CF75, 0x8000CF818000CF7D,
-    0x0004B8868000CF85, 0x8000CF8900039887, 0x8000CF918000CF8D,
-    0x0002B0EF8000CF95, 0x8000CF9D8000CF99, 0x00026CE38000CFA1,
-    0x000180E18000CFA5, 0x8000CFAD8000CFA9, 0x8000CFB58000CFB1,
-    0x8000CFBD8000CFB9, 0x8000CFC58000CFC1, 0x0001A8E18000CFC9,
-    0x00006CE38000CFCD, 0x000011818000CFD1, 0x0000B8EF00003980,
-    0x8000CFD500001982, 0x0004B4BE8000CFD9, 0x000F78E18000CFE1,
-    0x000E78E30005CC84, 0x000000DC8000CFE5, 0x8000CFED8000CFE9,
-    0x8000CFF58000CFF1, 0x000448868000CFF9, 0x8000CFFD00056087,
-    0x000F20EF0004FC84, 0x8000D0058000D001, 0x8000D009000EFCE1,
-    0x00071CBE000E50E3, 0x8000D0118000D00D, 0x8000D0198000D015,
-    0x8000D0218000D01D, 0x8000D0298000D025, 0x8000D0318000D02D,
-    0x8000D0350001D4EF, 0x8000D03D8000D039, 0x8000D04100026DB3,
-    0x000275B600027DB5, 0x8000D0498000D045, 0x000251BA003800BF,
-    0x0001F1BC000251BB, 0x0001F1BE0001F1BD, 0x0005A1A30001EDBF,
-    0x0007D5810007E582, 0x0001F4850007DD80, 0x0005B1A50005B9A6,
-    0x00059DAB8000D051, 0x003E888600059DAA, 0x000621AF00038C87,
-    0x00063DAD00063DAE, 0x8000D05500063DAC, 0x8000D0590006EDB3,
-    0x0006E5B68000D05D, 0x8000D0658000D061, 0x0006A9BA0006EDB5,
-    0x8000D0690006A9BB, 0x0006ADBE8000D06D, 0x0006ADBC000691BF,
-    0x8000CF6D0006A5BD, 0x8000D0718000CFDD, 0x8000D07500008084,
-    0x8000D07D8000D079, 0x000601A88000D081, 0x000601AA000601A9,
-    0x000601AC000601AB, 0x000679AE000601AD, 0x000609B0000679AF,
-    0x00061DB2000609B1, 0x00060DB4000615B3, 0x000635B6000631B5,
-    0x000615B800062DB7, 0x000615BA00061DB9, 0x0007F5BC0007EDBB,
-    0x0007F5BE0007FDBD, 0x0007A9A30007E9BF, 0x8000D0898000D085,
-    0x8000D04D8000D08D, 0x0007A9A50007A1A6, 0x0007EDAB8000D091,
-    0x8000D0950007EDAA, 0x0007D5AF8000D099, 0x0007E1AD0007E9AE,
-    0x8000D09D0007E9AC, 0x8000D0A1000611B3, 0x000631B68000D0A5,
-    0x8000D0AD8000D0A9, 0x00061DBA000601B5, 0x8000D0B10006E5BB,
-    0x0006E9BE8000D0B5, 0x0006F1BC0006E9BF, 0x0006BDA80006E9BD,
-    0x0006DDAA0006C9A9, 0x0006CDAC0006D5AB, 0x0006FDAE0006F5AD,
-    0x0000C9800006EDAF, 0x000005820000C981, 0x003FBCBE8000D0B9,
-    0x000048878000D0C1, 0x000669B8003E1C86, 0x000679BA000669B9,
-    0x000669BC000679BB, 0x00061DBE000615BD, 0x00069DB0000615BF,
-    0x000661B2000665B1, 0x000669B4000679B3, 0x000659B6000669B5,
-    0x000555A3000659B7, 0x8000D0C5001A058F, 0x8000D0C9003E80BE,
-    0x000545A5000575A6, 0x0005A1AB8000D0CD, 0x8000D0D1000559AA,
-    0x0005ADAF8000D0D5, 0x0005ADAD0005ADAE, 0x0006859E0005B5AC,
-    0x000A699C0006459F, 0x000EC19A000A799D, 0x000EF598000AE59B,
-    0x00124196000E7D99, 0x0016C5940012B997, 0x001675920012F595,
-    0x001A7D9000167D93, 0x8000D0D9001A1D91, 0x003F04BE8000D0DD,
-    0x000094E1003F4CBE, 0x003B2CE30000A4E1, 0x8000D0E1000000E3,
-    0x0033FCEF8000D0E5, 0x000231A20001E4EF, 0x00062DA0000231A3,
-    0x000289B30007D1A1, 0x8000D0ED8000D0E9, 0x8000D0F58000D0F1,
-    0x00028DB500028DB6, 0x000271BB8000D0F9, 0x8000D0FD00026DBA,
-    0x0003F5BF8000D101, 0x000261BD000269BE, 0x000670E1000269BC,
-    0x00066CE30007A4E1, 0x8000D105000658E3, 0x8000D10D8000D109,
-    0x002801B18000D111, 0x002BBDB3002E21B0, 0x002711B5002B0DB2,
-    0x000650EF002401B4, 0x003765A90006FCEF, 0x003349AB003765A8,
-    0x00320DAD003001AA, 0x002F79AF003345AC, 0x000271A1002C01AE,
-    0x003F45A38000D115, 0x003B51A5003F4DA2, 0x003401A7003801A4,
-    0x0005D9BC003BEDA6, 0x0005F1BE0005D9BD, 0x0005B5B80005F1BF,
-    0x0005D1BA0005D1B9, 0x0005A9B40005CDBB, 0x0005B5B60005B5B5,
-    0x000609B000058DB7, 0x0005B9B2000609B1, 0x000669AC0005B9B3,
-    0x000679AE000669AD, 0x0006A9A8000679AF, 0x000679AA0006A9A9,
-    0x8000D119000679AB, 0x8000D1218000D11D, 0x000019808000D125,
-    0x0000058200001981, 0x003C1DA88000D129, 0x003C41AA003C45A9,
-    0x003C45AC003C41AB, 0x003C71AE003C4DAD, 0x8000D0BD003C71AF,
-    0x8000D13100022C84, 0x000014868000D12D, 0x8000D1350003AC87,
-    0x003D5DB88000D139, 0x003D65BA003D6DB9, 0x003D19BC003D0DBB,
-    0x003D0DBE003D19BD, 0x003CE9B0003D05BF, 0x003CFDB2003CF5B1,
-    0x003CEDB4003CF5B3, 0x003D71B6003D71B5, 0x003CC9A3003D71B7,
-    0x8000D1418000D13D, 0x8000D14500000C84, 0x003CCDA5003CCDA6,
-    0x003C31AB8000D149, 0x8000D14D003C2DAA, 0x003DB5AF8000D151,
-    0x003C21AD003C29AE, 0x8000D155003C29AC, 0x8000D159003ED5B3,
-    0x003EFDB68000D15D, 0x8000D1658000D161, 0x003E41BA003EF5B5,
-    0x8000D169003E41BB, 0x003E3DBE8000D16D, 0x003E3DBC003E25BF,
-    0x003CB5A8003E35BD, 0x003F19AA003F0DA9, 0x003F39AC003F11AB,
-    0x003F99AE003F39AD, 0x8000D171003F99AF, 0x8000D1798000D175,
-    0x8000D1818000D17D, 0x8000D1898000D185, 0x003FADB88000D18D,
-    0x003E5DBA003E55B9, 0x003E4DBC003E55BB, 0x003EB9BE003E31BD,
-    0x003FC1B0003EB9BF, 0x003FADB2003FC1B1, 0x003FBDB4003FA5B3,
-    0x003FA5B6003FA5B5, 0x8000D191003F95B7, 0x003D099B8000D195,
-    0x000031820038E19A, 0x00001980003D119C, 0x8000D19900001981,
-    0x003589938000D1A1, 0x0039759500351D92, 0x00392D9700358194,
-    0x000760860039BD96, 0x002DE98B00008487, 0x00317D8D002D558A,
-    0x0031498F002D098C, 0x8000D1A50031FD8E, 0x0025ED838000D1A9,
-    0x0029F5850025ED82, 0x0029558700259584, 0x003C11A30029F586,
-    0x000304848000D1AD, 0x8000D1B100054084, 0x003C31A5003C39A6,
-    0x003C85AB8000D1B5, 0x000560BE003C85AA, 0x003CE1AF8000D1B9,
-    0x003CF1AD003CF9AE, 0x8000D1BD003CF9AC, 0x8000D1C58000D1C1,
-    0x8000D1C9000271B3, 0x0002C9B60002DDB5, 0x8000D1D18000D1CD,
-    0x0002C5BA8000D1D5, 0x0002B9BC0002ADBB, 0x0002ADBE0002B9BD,
-    0x0002B9A90002A5BF, 0x00020DAB0002ADA8, 0x000201AD0002B1AA,
-    0x000231AF00021DAC, 0x0006EC84000201AE, 0x8000D1DD8000D1D9,
-    0x8000D1E58000D1E1, 0x8000D1ED8000D1E9, 0x0003E5B98000D1F1,
-    0x0003E5BB0003E5B8, 0x0003E5BD0003E5BA, 0x0003E5BF0003E5BC,
-    0x000221B10003E5BE, 0x000225B3000249B0, 0x000229B5000225B2,
-    0x000215B7000221B4, 0x00033DA3000215B6, 0x8000D1F98000D1F5,
-    0x8000D2018000D1FD, 0x000391A5000385A6, 0x0003E1AB8000D205,
-    0x8000D209000389AA, 0x0003E9AF8000D211, 0x0003F5AD0003E1AE,
-    0x000051800003F5AC, 0x0000618200005981, 0x8000D2150002A9A3,
-    0x000145A6000145A5, 0x0004C0868000D219, 0x00012DAA0005E487,
-    0x000121AC000125AB, 0x000119AE000129AD, 0x8000D21D000119AF,
-    0x8000D2258000D221, 0x003FE0EF8000D229, 0x0007D0BE8000D22D,
-    0x00013CE3003F68EF, 0x000180E1003C00E1, 0x8000D231000198E3,
-    0x8000D2398000D235, 0x000225B38000D23D, 0x8000D2458000D241,
-    0x8000D24D8000D249, 0x0001C9B50001C9B6, 0x0001A9BB8000D251,
-    0x8000D2550001A1BA, 0x000195BF8000D259, 0x0001A5BD000195BE,
-    0x0005A5A80001ADBC, 0x0005EDAA0005E5A9, 0x000601AC000609AB,
-    0x000635AE000609AD, 0x8000D20D00065DAF, 0x8000D2618000D25D,
-    0x8000D2698000D265, 0x8000D2718000D26D, 0x0006EDB88000D275,
-    0x0006FDBA0006F5B9, 0x0006EDBC0006F5BB, 0x000751BE000751BD,
-    0x000625B0000751BF, 0x000639B200062DB1, 0x000611B4000631B3,
-    0x0006DDB6000611B5, 0x0000B9800006D5B7, 0x000005820000B981,
-    0x8000D27D8000D279, 0x0003A8878000D281, 0x0001348400003C86,
-    0x8000D2898000D285, 0x8000D2918000D28D, 0x8000D2998000D295,
-    0x8000D2A18000D29D, 0x0038BCEF8000D2A5, 0x8000D2AD8000D2A9,
-    0x8000D2B58000D2B1, 0x8000D2BD8000D2B9, 0x8000D2C58000D2C1,
-    0x000728E18000D2C9, 0x0004E0E38000D2CD, 0x00062DA88000D2D1,
-    0x000655AA000655A9, 0x000695AC00068DAB, 0x000695AE00069DAD,
-    0x8000D2D5000689AF, 0x8000D2DD8000D2D9, 0x00001D808000D2E1,
-    0x0000BD8200000981, 0x00069DB88000D2E5, 0x0006A5BA0006ADB9,
-    0x000179BC000169BB, 0x000169BE000179BD, 0x0006D1B0000169BF,
-    0x0006D1B20006D1B1, 0x0006B5B40006D1B3, 0x0006B5B60006BDB5,
-    0x0006DDB30006ADB7, 0x000F28868000D2E9, 0x8000D2F100014487,
-    0x00061DB500061DB6, 0x000615BB8000D2F5, 0x8000D2F9000615BA,
-    0x000679BF8000D2FD, 0x000605BD000671BE, 0x8000D301000605BC,
-    0x8000D305000699A3, 0x000659A68000D309, 0x8000D3118000D30D,
-    0x000651AA000659A5, 0x8000D315000651AB, 0x000635AE8000D319,
-    0x000641AC00063DAF, 0x000295A8000641AD, 0x0002A1AA0002A1A9,
-    0x0002A1AC0002A1AB, 0x0002A1AE0002A1AD, 0x8000D31D0002A1AF,
-    0x8000D3258000D321, 0x8000D32D8000D329, 0x8000D3358000D331,
-    0x0002A5B88000D339, 0x0002A5BA0002ADB9, 0x0002ADBC0002B9BB,
-    0x00037DBE000375BD, 0x0002A1B0000375BF, 0x0002A5B20002A1B1,
-    0x0002A5B40002BDB3, 0x0002A5B60002ADB5, 0x000229B300029DB7,
-    0x8000D33D0003C8BE, 0x8000D3458000D341, 0x000229B5000221B6,
-    0x000265BB8000D349, 0x8000D34D000265BA, 0x000269BF8000D351,
-    0x000269BD000261BE, 0x00001982000271BC, 0x0000658000026DA3,
-    0x000265A600001981, 0x8000D3598000D355, 0x000221AA00026DA5,
-    0x000D4886000221AB, 0x000225AE00035C87, 0x000235AC00022DAF,
-    0x8000D36100022DAD, 0x8000D3698000D365, 0x8000D3718000D36D,
-    0x8000D375000190E1, 0x8000D3790007ACE3, 0x8000D3818000D37D,
-    0x000000CC00064CEF, 0x8000D3898000D385, 0x000281B58000D38D,
-    0x000D04848000D391, 0x8000D395000285B6, 0x0002F1B38000D399,
-    0x000141BD8000D39D, 0x000141BF000289BC, 0x8000D3A1000149BE,
-    0x000291BB8000D19D, 0x8000D3A50002A5BA, 0x8000D3AD8000D3A9,
-    0x000C9CBE8000D3B1, 0x8000D3B98000D3B5, 0x001F9C8500074CEF,
-    0x8000D3BD0006DCE1, 0x00000D800007B0E3, 0x0000718200007181,
-    0x000269A38000D3C1, 0x8000D3C98000D3C5, 0x8000D3D58000D3D1,
-    0x000219A500021DA6, 0x000209AB8000D3D9, 0x000F488600023DAA,
-    0x0001D9AF000D0087, 0x0001D9AD0001D1AE, 0x000D91A8000211AC,
-    0x000DA1AA000D99A9, 0x000DD1AC000DA1AB, 0x000DD5AE000DDDAD,
-    0x8000D35D000DC9AF, 0x8000D3DD8000D3CD, 0x8000D3E58000D3E1,
-    0x8000D3ED8000D3E9, 0x000E75B88000D3F1, 0x000E75BA000E7DB9,
-    0x000EDDBC000EC5BB, 0x000EC5BE000EC5BD, 0x000DB9B0000EF5BF,
-    0x000D85B2000D8DB1, 0x000E55B4000E4DB3, 0x000E55B6000E5DB5,
-    0x000F89B3000E4DB7, 0x8000D3F98000D3F5, 0x8000D4018000D3FD,
-    0x000F89B5000F8DB6, 0x000F95BB8000D405, 0x8000D409000F8DBA,
-    0x000FE1BF8000D40D, 0x000F85BD000FF1BE, 0x8000D411000F85BC,
-    0x8000D415000CB8EF, 0x8000D41D8000D419, 0x8000D4258000D421,
-    0x8000D42D8000D429, 0x000011810001C582, 0x8000D43100001D80,
-    0x8000D435000E44E3, 0x000E01A8000F6CE1, 0x000E01AA000E01A9,
-    0x000E01AC000E01AB, 0x000E39AE000E0DAD, 0x8000D439000E39AF,
-    0x000124BE00022C84, 0x0000D0868000D441, 0x8000D4450001F087,
-    0x000E31B88000D449, 0x000ECDBA000EC5B9, 0x000EDDBC000EC5BB,
-    0x000EC1BE000EC1BD, 0x000E49B0000EF5BF, 0x000E51B2000E51B1,
-    0x000E35B4000E51B3, 0x000E2DB6000E25B5, 0x000E89A3000E21B7,
-    0x8000D4518000D44D, 0x8000D4598000D455, 0x000E89A5000E8DA6,
-    0x000E95AB8000D45D, 0x8000D461000E8DAA, 0x000EE1AF8000D465,
-    0x000E85AD000EF1AE, 0x8000D469000E85AC, 0x8000D46D000E19B3,
-    0x0001E1B68000D471, 0x8000D4798000D475, 0x0001DDBA0001F1B5,
-    0x8000D47D000161BB, 0x000161BE8000D481, 0x000171BC000169BF,
-    0x000DCDA8000171BD, 0x000DD5AA000DD5A9, 0x00023DAC000225AB,
-    0x00022DAE000225AD, 0x8000D485000225AF, 0x8000D48D8000D489,
-    0x8000D4958000D491, 0x8000D49D8000D499, 0x000281B88000D4A1,
-    0x000285BA00028DB9, 0x000285BC00029DBB, 0x000285BE00028DBD,
-    0x00025DB00002BDBF, 0x0002E1B20002E1B1, 0x0002E1B40002E1B3,
-    0x0002D1B60002E9B5, 0x000D5DA30002D1B7, 0x8000D4A98000D4A5,
-    0x000294840003C0BE, 0x0002B5A50002A5A6, 0x000225AB8000D4B1,
-    0x00042886000299AA, 0x00022DAF0004A087, 0x000235AD000225AE,
-    0x00004581000235AC, 0x0003F5B30000F980, 0x000359B500004582,
-    0x8000D4B98000D4B5, 0x8000D4BD000351B6, 0x000345BB8000D4C1,
-    0x000339BD000345BA, 0x00032DBF000335BC, 0x0002EDB8000331BE,
-    0x00012DBA000125B9, 0x00013DBC000125BB, 0x00012DBE000125BD,
-    0x000289B0000125BF, 0x0002E9B2000289B1, 0x0002FDB40002E5B3,
-    0x0002E1B60002E1B5, 0x00026DA80002D5B7, 0x000275AA00027DA9,
-    0x0002D9AC0002C9AB, 0x0002C9AE0002D9AD, 0x8000D4C50002C1AF,
-    0x8000D4CD8000D4C9, 0x8000D4D58000D4D1, 0x8000D2ED8000D4D9,
-    0x8000D4E18000D4DD, 0x8000D4E98000D4E5, 0x8000D4F18000D4ED,
-    0x000048EF8000D4F5, 0x000E50E1000EF4EF, 0x00009CE3000F54E1,
-    0x8000D4F9000EF4E3, 0x8000D5018000D4FD, 0x0002FDA38000D505,
-    0x8000D50D8000D509, 0x00045CBE8000D511, 0x000251A5000259A6,
-    0x00024DAB8000D515, 0x8000D51900024DAA, 0x000225AF0005A084,
-    0x000231AD000239AE, 0x8000D4AD00023DAC, 0x0007D5810007ED82,
-    0x8000D51D0007C180, 0x000338868000D521, 0x0006BDA800035C87,
-    0x0006F1AA0006F1A9, 0x0006D5AC0006CDAB, 0x00075DAE000755AD,
-    0x00072DB0000755AF, 0x00073DB2000735B1, 0x0007D1B4000731B3,
-    0x0007D1B60007D1B5, 0x0007F1B80007D1B7, 0x0007D9BA0007F1B9,
-    0x0007C9BC0007D9BB, 0x0007BDBE0007C9BD, 0x000659B30007B5BF,
-    0x8000D5298000D525, 0x8000D5318000D52D, 0x00067DB50007F5B6,
-    0x0007D1BB8000D535, 0x8000D5390007C9BA, 0x000735BF8000D53D,
-    0x0007C5BD000735BE, 0x000440E30007C9BC, 0x0007E0E18000D541,
-    0x8000D5498000D545, 0x8000D5518000D54D, 0x8000D5598000D555,
-    0x8000D5618000D55D, 0x000484EF8000D565, 0x8000D56D8000D569,
-    0x0007D9A38000D571, 0x8000D5798000D575, 0x8000D5818000D57D,
-    0x0007FDA5000675A6, 0x000651AB8000D585, 0x8000D589000649AA,
-    0x0006B5AF0002AC84, 0x000645AD0006B5AE, 0x00005180000649AC,
-    0x0000618200005981, 0x0001D0BE0001C5B3, 0x0001C5B60001C5B5,
-    0x0000A0868000D58D, 0x0001ADBA00011487, 0x0001A1BC0001A1BB,
-    0x0001A1BE0001A1BD, 0x000195A80001A1BF, 0x000101AA000101A9,
-    0x000101AC000101AB, 0x000131AE000109AD, 0x8000D591000131AF,
-    0x8000D5998000D595, 0x8000D5A18000D59D, 0x8000D5A98000D5A5,
-    0x000129B88000D5AD, 0x00013DBA000129B9, 0x00012DBC000135BB,
-    0x0000D1BE0000D1BD, 0x000125B00000D1BF, 0x000139B2000129B1,
-    0x000129B4000139B3, 0x000119B6000129B5, 0x00000081000119B7,
-    0x8000D5B58000D5B1, 0x8000D5B9000285A3, 0x000285A6000285A5,
-    0x8000D5C18000D5BD, 0x0002EDAA8000D5C5, 0x0002E1AC0002E1AB,
-    0x0002E1AE0002E1AD, 0x8000D5C90002E1AF, 0x8000D5D18000D5CD,
-    0x8000D5D500020084, 0x8000D5DD8000D5D9, 0x8000D5E58000D5E1,
-    0x8000D5ED8000D5E9, 0x00001D808000D5F1, 0x0000FD8200000981,
-    0x000329B38000D5F5, 0x8000D6018000D5F9, 0x8000D6050003C0BE,
-    0x000329B5000321B6, 0x0000A9BB8000D609, 0x001C08860000A1BA,
-    0x000095BF00036087, 0x0000A5BD000095BE, 0x8000D60D0000ADBC,
-    0x8000D6158000D611, 0x00028CE38000D619, 0x0001A0E18000D61D,
-    0x000294EF8000D621, 0x8000D6298000D625, 0x8000D6318000D62D,
-    0x8000D6398000D635, 0x000105B08000D63D, 0x000119B200010DB1,
-    0x000131B4000111B3, 0x000159B6000131B5, 0x000169B8000159B7,
-    0x00017DBA000175B9, 0x00016DBC000175BB, 0x00015DBE000155BD,
-    0x001D0C84000155BF, 0x8000D6458000D641, 0x8000D6498000D43D,
-    0x8000D6518000D64D, 0x0002C1A88000D655, 0x0002C1AA0002C1A9,
-    0x0002C1AC0002C1AB, 0x0002C1AE0002C1AD, 0x001C0CBE0002C1AF,
-    0x8000D659000734E1, 0x8000D65D000670E3, 0x8000D661000194E1,
-    0x8000D6650000A0E3, 0x0016B8BF8000D669, 0x8000D671000004EF,
-    0x8000D6798000D675, 0x000269A3000694EF, 0x0000698100006182,
-    0x8000D67D00005580, 0x000269A5000261A6, 0x0001E9AB8000D681,
-    0x001CC8860001E1AA, 0x0001D5AF001D0087, 0x0001E5AD0001D5AE,
-    0x001E89A80001EDAC, 0x001E95AA001E95A9, 0x001ED5AC001ECDAB,
-    0x001ED5AE001EDDAD, 0x8000D66D001EC9AF, 0x8000D6898000D685,
-    0x8000D6918000D68D, 0x8000D6998000D695, 0x001F05B88000D69D,
-    0x001F19BA001F0DB9, 0x001F09BC001F19BB, 0x001FD1BE001F09BD,
-    0x001EBDB0001FCDBF, 0x001E81B2001E81B1, 0x001E81B4001E81B3,
-    0x001E81B6001E81B5, 0x001E1DB3001E81B7, 0x8000D6A58000D6A1,
-    0x8000D6AD8000D6A9, 0x001E5DB5001E5DB6, 0x001E45BB8000D6B1,
-    0x8000D6B5001E79BA, 0x001E45BF8000D6B9, 0x001E55BD001E45BE,
-    0x8000D6BD001E55BC, 0x8000D6C58000D6C1, 0x001D80E38000D6C9,
-    0x001C68E18000D6CD, 0x001D98EF8000D6D1, 0x0000198100001182,
-    0x8000D6D500003180, 0x000370848000D6D9, 0x001F9DA300007CBE,
-    0x0000C8868000D6DD, 0x8000D6E100014C87, 0x001FDDA5001FDDA6,
-    0x001FC5AB8000D6E5, 0x8000D6E9001FF9AA, 0x001FC5AF8000D6ED,
-    0x001FD5AD001FC5AE, 0x8000D6F1001FD5AC, 0x8000D6F5001EFDB3,
-    0x000175B68000D6F9, 0x8000D7018000D6FD, 0x000151BA001E81B5,
-    0x8000D705000151BB, 0x0001F5BE8000D709, 0x000141BC0001FDBF,
-    0x001E31A8000141BD, 0x001E31AA001E31A9, 0x001E4DAC001E5DAB,
-    0x0001B1AE0001B1AD, 0x8000D70D0001B1AF, 0x8000D7158000D711,
-    0x8000D71D8000D719, 0x8000D7258000D721, 0x00015DB88000D729,
-    0x000165BA00016DB9, 0x000115BC00010DBB, 0x000115BE00011DBD,
-    0x0001D1B0000109BF, 0x0001D1B20001D1B1, 0x000175B40001D1B3,
-    0x000175B600017DB5, 0x8000D72D00016DB7, 0x8000D73100020884,
-    0x8000D735001DB5A3, 0x00023DA6001DC9A5, 0x8000D73D8000D739,
-    0x000219AA8000D741, 0x000209AC000219AB, 0x0002BDAE000209AD,
-    0x000065800002B5AF, 0x00001D8200001581, 0x8000D745000308BE,
-    0x000308878000D749, 0x8000D7510004FC86, 0x8000D7598000D755,
-    0x8000D7618000D75D, 0x8000D7698000D765, 0x000190E18000D76D,
-    0x00029CE38000D771, 0x8000D7798000D775, 0x8000D7818000D77D,
-    0x0004AC848000D785, 0x8000D78D8000D789, 0x8000D7958000D791,
-    0x00020CEF8000D799, 0x0003E9B38000D79D, 0x8000D7A58000D7A1,
-    0x8000D7AD8000D7A9, 0x000399B500039DB6, 0x000389BB8000D7B1,
-    0x000564BE0003BDBA, 0x000359BF00056484, 0x000359BD000351BE,
-    0x000259B0000391BC, 0x000265B200026DB1, 0x0002B5B40002ADB3,
-    0x0002B5B60002BDB5, 0x000295B80002ADB7, 0x000295BA00029DB9,
-    0x000139BC00012DBB, 0x000125BE000125BD, 0x8000D7B500011DBF,
-    0x8000D7C18000D7B9, 0x00018D808000D7C5, 0x00001D8200001581,
-    0x000271A88000D5FD, 0x000241AA000279A9, 0x000231AC000241AB,
-    0x000235AE00023DAD, 0x0004A486000229AF, 0x000700BE00050487,
-    0x8000D7C900070884, 0x8000D7D18000D7CD, 0x8000D7D5000198EF,
-    0x8000D7D9001E54E1, 0x8000D7DD000144E3, 0x8000D7E58000D7E1,
-    0x8000D7ED8000D7E9, 0x0002E1A38000D7F1, 0x000291A58000D7F5,
-    0x8000D7FD8000D7F9, 0x8000D801000295A6, 0x000281AB8000D805,
-    0x000251AD0002B5AA, 0x000251AF000299AC, 0x000661A8000259AE,
-    0x000661AA000661A9, 0x000661AC000661AB, 0x000661AE000661AD,
-    0x8000D7BD000661AF, 0x8000D80D8000D809, 0x8000D8158000D811,
-    0x8000D81D8000D819, 0x0006EDB88000D821, 0x0006FDBA0006F5B9,
-    0x000695BC0006F1BB, 0x000695BE00069DBD, 0x0006E5B000068DBF,
-    0x0006F9B20006E9B1, 0x0006E9B40006F9B3, 0x0006DDB60006E9B5,
-    0x0000BD800006D5B7, 0x0000198200000981, 0x8000D8298000D825,
-    0x8000D8318000D82D, 0x8000D83500048CEF, 0x8000D8390004E0E1,
-    0x0000B08600041CE3, 0x8000D83D00031487, 0x0006A9B38000D841,
-    0x8000D8498000D845, 0x8000D8518000D84D, 0x0006A9B50006A1B6,
-    0x00066DBB8000D855, 0x8000D859000661BA, 0x00066DBF8000D85D,
-    0x00066DBD00066DBE, 0x000769A300066DBC, 0x8000D8658000D861,
-    0x8000D86D8000D869, 0x000769A5000761A6, 0x0007ADAB8000D871,
-    0x8000D8750007A1AA, 0x0007ADAF8000D879, 0x0007ADAD0007ADAE,
-    0x8000D87D0007ADAC, 0x8000D8810006D5B3, 0x0006F1B68000D885,
-    0x8000D88D8000D889, 0x0006D5BA0006C5B5, 0x8000D8910006DDBB,
-    0x0006B5BE8000D895, 0x0006C5BC0006B9BF, 0x0006CDA90006BDBD,
-    0x00061DAB0006CDA8, 0x000631AD00061DAA, 0x000629AF00060DAC,
-    0x0000C980000635AE, 0x000005820000C981, 0x00007CBE8000D899,
-    0x000088878000D8A1, 0x000625B9000F1C86, 0x0001E1BB000625B8,
-    0x0001E1BD0001E9BA, 0x0001E1BF0001F9BC, 0x000659B10001E9BE,
-    0x000629B3000651B0, 0x00063DB5000651B2, 0x000621B7000631B4,
-    0x8000D8A5000629B6, 0x8000D8A9000595A3, 0x0005B1A600030C84,
-    0x8000D8B18000D8AD, 0x000595AA000585A5, 0x8000D8B500059DAB,
-    0x0005F5AE8000D8B9, 0x000585AC0005F9AF, 0x8000D8BD0005FDAD,
-    0x0002E1B38000D8C1, 0x0002EDB58000D8C5, 0x8000D8CD8000D8C9,
-    0x8000D8D1000299B6, 0x0002B5BB8000D8D5, 0x00029DBD0002BDBA,
-    0x000371BF00029DBC, 0x000CCCBE000379BE, 0x8000D8DD8000D8D9,
-    0x8000D8E58000D8E1, 0x8000D8ED8000D8E9, 0x000239A88000D8F1,
-    0x000201AA000205A9, 0x00020DAC000215AB, 0x000271AE000271AD,
-    0x000211B0000271AF, 0x000221B2000219B1, 0x0003D5B4000221B3,
-    0x0003D5B60003DDB5, 0x0003FDB80003CDB7, 0x0003C5BA0003CDB9,
-    0x0003B9BC0003ADBB, 0x0003ADBE0003B9BD, 0x8000D8F50003A5BF,
-    0x8000D8FD8000D8F9, 0x0001B8E18000D901, 0x000684E38000D905,
-    0x8000D74D8000D909, 0x000060EF000C6884, 0x00003D808000D911,
-    0x0000258200002181, 0x0003E9A38000D915, 0x000CE8868000D919,
-    0x8000D91D000D6087, 0x0003E5A5000391A6, 0x0003BDAB8000D921,
-    0x8000D9250003B5AA, 0x000279AF8000D929, 0x000395AD000271AE,
-    0x8000D92D000395AC, 0x000CC0BE8000D931, 0x8000D935000FC885,
-    0x8000D93D8000D939, 0x000F60840006ECEF, 0x8000D94100063CE1,
-    0x8000D945000750E3, 0x8000D94D8000D949, 0x8000D9558000D951,
-    0x8000D95D8000D959, 0x8000D9610001C5B3, 0x0001F1B60001CDB5,
-    0x8000D9698000D965, 0x0001ADBA8000D96D, 0x0001A5BC0001A1BB,
-    0x000199BE0001ADBD, 0x000DB5A8000199BF, 0x000E01AA000E05A9,
-    0x000E09AC000E19AB, 0x000E35AE000E3DAD, 0x8000D90D000E5DAF,
-    0x00001D8100001D82, 0x8000D97100001D80, 0x8000D9798000D975,
-    0x000ECDB8000170BE, 0x000EDDBA000ED5B9, 0x000ECDBC000ED5BB,
-    0x000F5DBE000F55BD, 0x000E25B0000F55BF, 0x000E3DB2000E29B1,
-    0x000E11B4000E35B3, 0x000EFDB6000E11B5, 0x8000D981000EF5B7,
-    0x000CF0EF000D1CE1, 0x0004008603E914E3, 0x8000D98500035C87,
-    0x8000D98D8000D989, 0x8000D9958000D991, 0x000F14E18000D999,
-    0x000CA4E38000D99D, 0x000EC9A303E9E4EF, 0x8000D9A58000D9A1,
-    0x8000D9AD8000D9A9, 0x000EC1A5000EFDA6, 0x000EADAB8000D9B1,
-    0x8000D97D000EA1AA, 0x000E95AF8000D9B5, 0x000EA1AD000E95AE,
-    0x000E2DA8000EA9AC, 0x000E4DAA000E39A9, 0x000E4DAC000E5DAB,
-    0x000E75AE000E7DAD, 0x8000D9B9000E9DAF, 0x8000D9C18000D9BD,
-    0x8000D9C98000D9C5, 0x8000D9D18000D9CD, 0x000EFDB88000D9D5,
-    0x000E8DBA000E85B9, 0x000E81BC000E85BB, 0x000EB9BE000E89BD,
-    0x000EE5B0000EB9BF, 0x000EFDB2000EF5B1, 0x000ED9B4000EF5B3,
-    0x000ECDB6000EC5B5, 0x000EC5B3000EC5B7, 0x8000D9DD8000D9D9,
-    0x8000D9E58000D9E1, 0x000EC5B5000EFDB6, 0x000E31BB8000D9E9,
-    0x00072886000E21BA, 0x000E01BF00001C87, 0x000E21BD000E1DBE,
-    0x00005D82000E29BC, 0x00004580000E81A3, 0x000EB9A600005D81,
-    0x8000D9F58000D9F1, 0x000E65AA000E81A5, 0x8000D9F9000E75AB,
-    0x000E59AE8000D9FD, 0x000E6DAC000E45AF, 0x0027118F000E65AD,
-    0x8000DA058000DA01, 0x8000DA09000291B3, 0x0002B1B60002B1B5,
-    0x8000DA118000DA0D, 0x0002EDBA8000DA15, 0x0002FDBC0002E5BB,
-    0x0002E5BE0002E5BD, 0x000D619F0002D1BF, 0x000A699D000B419E,
-    0x00170D9B0009119C, 0x001659990016459A, 0x0012459700151198,
-    0x0011119500124596, 0x001E5193001EFD94, 0x001A5591001D0192,
-    0x03E801B1001A5D90, 0x03EA99B303F6B9B0, 0x03EE99B503EA89B2,
-    0x8000DA1903EBFDB4, 0x0004EC848000DA1D, 0x8000DA258000DA21,
-    0x000291A38000DA29, 0x000E2DA1000181A2, 0x000229A1000E9DA0,
-    0x0006CDA300058884, 0x03F801A50006CDA2, 0x03FBFDA70007F9A4,
-    0x03FEC1A903FAF1A6, 0x03F001AB03FC01A8, 0x03F261AD03FE39AA,
-    0x03F6D9AF03F2D5AC, 0x001B2D8403F401AE, 0x001F2986001C7185,
-    0x8000DA2D001F9D87, 0x002609828000DA31, 0x0013058C001B3D83,
-    0x0017298E0014698D, 0x8000DA3500173D8F, 0x0010418A8000D89D,
-    0x000B619400131D8B, 0x000C6996000A4195, 0x8000DA39000F6D97,
-    0x00160592000190EF, 0x0003419C00080193, 0x8000DA3D000ED4E3,
-    0x8000DA410001E4E1, 0x000FED9A8000DA45, 0x8000DA490000499B,
-    0x8000DA518000DA4D, 0x8000DA598000DA55, 0x8000DA610005DCBE,
-    0x0004E8BE000EECEF, 0x8000DA65000EF4E1, 0x00003980000E98E3,
-    0x0000298200003981, 0x00035DA38000DA69, 0x000468868000DA6D,
-    0x8000DA710004EC87, 0x00037DA500037DA6, 0x000329AB8000DA75,
-    0x8000DA79000321AA, 0x00031DAF8000DA7D, 0x000329AD000329AE,
-    0x8000D9ED000331AC, 0x8000DA818000DA5D, 0x8000DA898000DA85,
-    0x8000DA918000DA8D, 0x0005F5A88000DA95, 0x0005C1AA0005FDA9,
-    0x000635AC0005C1AB, 0x000629AE00063DAD, 0x000661B0000621AF,
-    0x0006D1B2000661B1, 0x0006F5B40006D1B3, 0x0006F5B60006FDB5,
-    0x000775B800074DB7, 0x000739BA000755B9, 0x000701BC000705BB,
-    0x0007F9BE000701BD, 0x000569B30007F9BF, 0x8000DA9D8000DA99,
-    0x0001DCBE8000DAA1, 0x00069DB500069DB6, 0x000685BB8000DAA5,
-    0x8000DAA9000685BA, 0x000619BF8000DAAD, 0x000625BD000611BE,
-    0x8000DAB1000625BC, 0x8000DAB500052DA3, 0x0006D9A68000DAB9,
-    0x8000DAC18000DABD, 0x0006C1AA0006D9A5, 0x8000DAC50006C1AB,
-    0x000655AE8000DAC9, 0x000661AC00065DAF, 0x000699A8000661AD,
-    0x0006B5AA0006ADA9, 0x0006C1AC0006C9AB, 0x0006F1AE0006C9AD,
-    0x0001F9800006F1AF, 0x0001C5820001CD81, 0x000090868000DACD,
-    0x8000DAD10001B487, 0x0001CDB88000DAD5, 0x0001E1BA0001D9B9,
-    0x000191BC0001E1BB, 0x000195BE00019DBD, 0x000169B0000189BF,
-    0x000179B2000169B1, 0x000169B4000179B3, 0x0001FDB6000169B5,
-    0x00063DB30001F5B7, 0x8000DADD8000DAD9, 0x8000DAE58000DAE1,
-    0x000649B5000649B6, 0x0001ADBB8000DAE9, 0x8000DAED0001ADBA,
-    0x0001ADBF8000DAF1, 0x0001ADBD0001ADBE, 0x8000DAF50001ADBC,
-    0x8000DAF9000679A3, 0x00060DA68000DAFD, 0x8000DB058000DB01,
-    0x0001E9AA00060DA5, 0x8000DB090001E9AB, 0x0001E9AE8000DB0D,
-    0x0001E9AC0001E9AF, 0x8000DB110001E9AD, 0x8000DB198000DB15,
-    0x8000DB218000DB1D, 0x8000DB298000DB25, 0x000685AC8000DB2D,
-    0x000685AE00068DAD, 0x0006B1A8000699AF, 0x0006A5AA0006ADA9,
-    0x00069DB400069DAB, 0x000769B60006A9B5, 0x000695B0000769B7,
-    0x000695B200069DB1, 0x000759BC00068DB3, 0x0007EDBE000759BD,
-    0x000759B80007E5BF, 0x000749BA000759B9, 0x00000D80000749BB,
-    0x0000118200001181, 0x0000F8EF8000DB31, 0x8000DB398000DB35,
-    0x000214E38000DB41, 0x000180E100028884, 0x8000DB498000DB45,
-    0x03FC18868000DB4D, 0x000301BB03FC9C87, 0x8000DB51000369BA,
-    0x000309BF8000DB55, 0x000315BD000301BE, 0x000365B3000315BC,
-    0x8000DB5D8000DB59, 0x8000DB658000DB61, 0x000375B5000365B6,
-    0x000239A88000DB69, 0x00028DAA000239A9, 0x00029DAC000285AB,
-    0x000285AE000285AD, 0x03FD6C840002B5AF, 0x03FD60BE8000DB6D,
-    0x8000DB758000DB71, 0x8000DB7D8000DB79, 0x000141B88000DB81,
-    0x000141BA000141B9, 0x000141BC000141BB, 0x000141BE000141BD,
-    0x0002CDB0000141BF, 0x0002A1B20002A5B1, 0x0002A5B40002A1B3,
-    0x000291B60002ADB5, 0x000650E1000291B7, 0x00075CE3000434E1,
-    0x8000DB8500073CE3, 0x8000DB8D8000DB89, 0x03FF64848000DB91,
-    0x8000DB998000DB95, 0x8000DBA18000DB9D, 0x0006E0EF8000DBA5,
-    0x000245A30006D0EF, 0x8000DBB18000DBA9, 0x8000DBB98000DBB5,
-    0x000255A5000245A6, 0x000221AB8000DBBD, 0x03FCC886000249AA,
-    0x000229AF03FCFC87, 0x000235AD000221AE, 0x03FE9DA8000235AC,
-    0x03FEB9AA03FEA9A9, 0x03FED9AC03FEB9AB, 0x03FEFDAE03FED9AD,
-    0x03FF918003FEF5AF, 0x03FFED8203FF9181, 0x8000DBC18000DBAD,
-    0x8000DBC98000DBC5, 0x03FF7DB88000DBCD, 0x03FFDDBA03FFD5B9,
-    0x03FFCDBC03FFD5BB, 0x03FFF1BE03FFF1BD, 0x03FE8DB003FFF1BF,
-    0x03FE95B203FE95B1, 0x03FF5DB403FF49B3, 0x03FF4DB603FF45B5,
-    0x03FE31B303FF45B7, 0x8000DBD58000DBD1, 0x8000DBDD8000DBD9,
-    0x03FE51B503FE59B6, 0x03FE51BB8000DBE1, 0x8000DBE503FE49BA,
-    0x03FF81BF8000DBE9, 0x03FF9DBD03FF99BE, 0x8000DBED03FE41BC,
-    0x8000DBF103FE75A3, 0x03FE1DA68000DBF5, 0x8000DBFD8000DBF9,
-    0x03FE0DAA03FE15A5, 0x8000DC0103FE15AB, 0x03FFDDAE8000DC05,
-    0x03FE05AC03FFC5AF, 0x8000DC0903FFD9AD, 0x8000DC118000DC0D,
-    0x000019808000DC15, 0x0000058200001981, 0x0001FCBE8000DC19,
-    0x000088868000DC1D, 0x8000DC2100012887, 0x8000DC298000DC25,
-    0x8000DC318000DC2D, 0x8000DC398000DC35, 0x8000DC418000DC3D,
-    0x8000DC4503FDF4E1, 0x8000DC4903FC6CE3, 0x8000DC518000DC4D,
-    0x8000DC5503FCBCEF, 0x8000DC5D8000DC59, 0x03FE49A88000DC61,
-    0x03FE59AA03FE49A9, 0x0001A1AC03FE51AB, 0x0001A1AE0001A1AD,
-    0x0001AC840001A1AF, 0x8000DC698000DC65, 0x8000DC718000DC6D,
-    0x8000DC798000DC75, 0x000109B88000DC7D, 0x00011DBA000109B9,
-    0x000139BC000115BB, 0x0000D9BE000139BD, 0x000165B00000D9BF,
-    0x000165B200016DB1, 0x00016DB400017DB3, 0x000155B600015DB5,
-    0x0000198100013DB7, 0x03FDD9B300001180, 0x000221B500002982,
-    0x8000DC858000DC81, 0x002D4DB5000229B6, 0x0002A1BB8000DC89,
-    0x0002A1BD0002A9BA, 0x000291BF0002B9BC, 0x00031CBE0002A1BE,
-    0x000279A5000271A6, 0x03FD81A300040084, 0x0004F8868000DC91,
-    0x0002C9AF0003DC87, 0x0002F9AD0002F9AE, 0x0002F9AB0002E1AC,
-    0x8000DC950002F1AA, 0x8000DC9D8000DC99, 0x8000DCA58000DCA1,
-    0x8000DCAD8000DCA9, 0x000220EF8000DCB1, 0x000190E18000DCB5,
-    0x000118E38000DCB9, 0x8000DCC18000DCBD, 0x8000DCC98000DCC5,
-    0x0003F9B38000DCCD, 0x8000DCD58000DCD1, 0x8000DB3D8000DCD9,
-    0x000399B5000391B6, 0x0003B9BB8000DCDD, 0x8000DCE10003B5BA,
-    0x000381BF8000DCE5, 0x00039DBD00039DBE, 0x000231A800039DBC,
-    0x000231AA000231A9, 0x000251AC000231AB, 0x000251AE000251AD,
-    0x000568BE000251AF, 0x8000DCED8000DCE9, 0x0004FC848000DCF1,
-    0x8000DCF98000DCF5, 0x00023DB88000DD01, 0x0002E1BA0002E1B9,
-    0x0002E1BC0002E1BB, 0x0002E1BE0002E1BD, 0x000231B00002E1BF,
-    0x000209B200023DB1, 0x000219B4000209B3, 0x00020DB6000205B5,
-    0x0000AD80000205B7, 0x0000B1820000B181, 0x000048EF8000DD05,
-    0x000550878000DD09, 0x03FE10E30004BC86, 0x03FE58E18000DD0D,
-    0x000000EC8000DD11, 0x8000DD198000DD15, 0x8000DD218000DD1D,
-    0x000289A5000281A6, 0x0002E9A38000DD25, 0x000740848000DD29,
-    0x000291AF8000DD2D, 0x00028DAD00028DAE, 0x0002A9AB00028DAC,
-    0x8000DD310002A5AA, 0x000655A88000DD35, 0x000675AA00067DA9,
-    0x000695AC00068DAB, 0x000695AE00069DAD, 0x8000DCFD00068DAF,
-    0x8000DD3D8000DD39, 0x8000DD458000DD41, 0x8000DD4D8000DD49,
-    0x000685B88000DD51, 0x000685BA00068DB9, 0x0006B9BC000691BB,
-    0x00077DBE0006B9BD, 0x0006F5B0000775BF, 0x0006CDB20006F9B1,
-    0x0006C1B40006C1B3, 0x0006C1B60006C1B5, 0x0006FDB30006C1B7,
-    0x8000DD598000DD55, 0x8000DD618000DD5D, 0x00061DB500061DB6,
-    0x00063DBB8000DD65, 0x8000DD69000631BA, 0x000605BF00002CBE,
-    0x000615BD000605BE, 0x00006982000619BC, 0x000059800006B9A3,
-    0x000659A600005181, 0x8000DD718000DD6D, 0x000675AA000659A5,
-    0x00008886000679AB, 0x000641AE00035C87, 0x00065DAC000641AF,
-    0x000695B3000651AD, 0x8000DD798000DD75, 0x8000DD818000DD7D,
-    0x0006B5B50006A1B6, 0x0006EDBB8000DD85, 0x8000DD890006E5BA,
-    0x0006CDBF8000DD8D, 0x0006E5BD0006CDBE, 0x8000DD910006E9BC,
-    0x8000DD998000DD95, 0x8000DDA18000DD9D, 0x000480EF8000DDA5,
-    0x000520E18000DDA9, 0x000534E38000DDAD, 0x8000DDB58000DDB1,
-    0x8000DDBD8000DDB9, 0x0006BDA88000DDC1, 0x00061DAA000609A9,
-    0x00063DAC000629AB, 0x000621AE000621AD, 0x8000DDC5000621AF,
-    0x8000DDCD8000DDC9, 0x00001D808000DDD1, 0x0003B98200000981,
-    0x0001EDB88000DDD5, 0x0001FDBA0001F5B9, 0x0001EDBC0001F5BB,
-    0x000151BE000151BD, 0x000621B0000151BF, 0x00063DB2000629B1,
-    0x000609B4000631B3, 0x0001DDB6000609B5, 0x8000DDD90001D5B7,
-    0x8000DDE1000388BE, 0x8000DDE5000599A3, 0x0005ADA60005B9A5,
-    0x8000DDED8000DDE9, 0x0005E9AA8000DDF1, 0x0005E5AC0005E1AB,
-    0x0005C1AE0005E9AD, 0x00035DB60005C1AF, 0x000348848000DDF5,
-    0x8000DDF900035DB5, 0x000CF88600035DB3, 0x000331BE0003DC87,
-    0x000321BC000331BF, 0x000331BA000321BD, 0x8000DDFD000331BB,
-    0x0003EDB88000DE01, 0x0003FDBA0003F5B9, 0x0003EDBC0003F5BB,
-    0x00033DBE000335BD, 0x0003E1B0000335BF, 0x0003E1B20003E1B1,
-    0x0003E1B40003E1B3, 0x0003E5B60003E1B5, 0x000219A80003D9B7,
-    0x0003C1AA0003C1A9, 0x0003C1AC0003C1AB, 0x0003F1AE0003C9AD,
-    0x8000DE050003F1AF, 0x8000DE098000DC8D, 0x000CFCBE8000DE0D,
-    0x8000DE158000DE11, 0x000209A38000DE19, 0x8000DE218000DE1D,
-    0x8000DE298000DE25, 0x000209A5000209A6, 0x000265AB8000DE2D,
-    0x8000DE31000265AA, 0x000265AF8000DE35, 0x000275AD000265AE,
-    0x8000DE39000275AC, 0x8000DE418000DE3D, 0x0001A0E18000DE45,
-    0x000690E38000DE49, 0x8000DE558000DE51, 0x000394EF8000DE59,
-    0x000C50868000DE5D, 0x8000DE61000D4487, 0x0000C9808000DE65,
-    0x0000DD820000C981, 0x8000DE6D8000DE69, 0x8000DE758000DE71,
-    0x8000DE790001C8EF, 0x8000DE7D000620E1, 0x00351C8200068CE3,
-    0x8000DE858000DE81, 0x8000DE8D8000DE89, 0x8000DE958000DE91,
-    0x8000DE990001D1B3, 0x00019DB60001F1B5, 0x8000DEA18000DE9D,
-    0x000195BA8000DEA5, 0x000119BC00010DBB, 0x00010DBE000119BD,
-    0x000E09A8000105BF, 0x000E15AA000E15A9, 0x000E35AC000E2DAB,
-    0x000E49AE000E51AD, 0x8000DE4D000E65AF, 0x8000DEAD8000DEA9,
-    0x00001D808000DEB1, 0x0000BD8200000981, 0x000F49B88000DEB5,
-    0x000F59BA000F49B9, 0x000F79BC000F51BB, 0x000F19BE000F79BD,
-    0x000E1DB0000F19BF, 0x000E81B2000E85B1, 0x000E85B4000E81B3,
-    0x000E85B6000E8DB5, 0x000E9DA3000F79B7, 0x000280848000DEB9,
-    0x8000DEBD000300BE, 0x000EBDA5000ED1A6, 0x000E41AB8000DEC1,
-    0x00000886000ED9AA, 0x000E49AF0003DC87, 0x000E55AD000E41AE,
-    0x8000DEC5000E55AC, 0x8000DEC9000FA1B3, 0x000FADB68000DECD,
-    0x8000DED58000DED1, 0x000F75BA000FA5B5, 0x8000DED9000F7DBB,
-    0x000F45BE8000DEDD, 0x000F59BC000F4DBF, 0x000E65A3000F55BD,
-    0x8000DEE58000DEE1, 0x8000DEED8000DEE9, 0x000E61A5000E69A6,
-    0x000EB9AB8000DEF1, 0x8000DEF5000EB1AA, 0x000E89AF8000DEF9,
-    0x000E91AD000E81AE, 0x8000DEFD000E9DAC, 0x8000DF058000DF01,
-    0x8000DF0D8000DF09, 0x8000DF11000D5CEF, 0x8000DF198000DF15,
-    0x8000DF1D000D3CE3, 0x00000D80000D2CE1, 0x0000DD820000D581,
-    0x000EDDA88000DF21, 0x000EEDAA000EE5A9, 0x000EE9AC000EE1AB,
-    0x000E15AE000EE9AD, 0x8000DF25000E0DAF, 0x0001E4848000DF29,
-    0x000710860001E4BE, 0x8000DF3100009487, 0x000E05B88000DF35,
-    0x000E19BA000E0DB9, 0x000E31BC000E11BB, 0x0001FDBE000E31BD,
-    0x000E75B00001F5BF, 0x000E49B2000E7DB1, 0x000E45B4000E45B3,
-    0x000E45B6000E4DB5, 0x00029DA8000E3DB7, 0x0002C1AA0002C5A9,
-    0x0002C5AC0002C1AB, 0x0002F1AE0002CDAD, 0x8000DF390002F1AF,
-    0x8000DF418000DF3D, 0x000328E18000DF45, 0x8000DF4D8000DF49,
-    0x000305B88000DF51, 0x000305BA00030DB9, 0x000309BC00031DBB,
-    0x000339BE000309BD, 0x000365B0000335BF, 0x000365B200036DB1,
-    0x00036DB400037DB3, 0x000355B600035DB5, 0x000249B300033DB7,
-    0x8000DF598000DF55, 0x8000DF618000DF5D, 0x00024DB500024DB6,
-    0x0003A9BB8000DF65, 0x8000DF690003A1BA, 0x000399BF8000DF6D,
-    0x0003A9BD0003A9BE, 0x8000DF710003B1BC, 0x8000DF7500020DA3,
-    0x000209A68000DF79, 0x8000DF818000DF7D, 0x0003E5AA000209A5,
-    0x8000DDDD0003EDAB, 0x0003EDAE8000DF85, 0x0003F5AC0003DDAF,
-    0x8000DF890003EDAD, 0x8000DF918000DF8D, 0x8000DF998000DF95,
-    0x00024CBE0001A4E1, 0x000029800000A4E3, 0x0000398200002981,
-    0x8000DFA1000078EF, 0x8000DFA50002B884, 0x000161B38000DFA9,
-    0x8000DFB18000DFAD, 0x8000DFB98000DFB5, 0x000171B5000179B6,
-    0x0001E5BB8000DFBD, 0x00054886000159BA, 0x0001EDBF00054087,
-    0x0001F5BD0001E5BE, 0x8000DFC10001F5BC, 0x8000DFC98000DFC5,
-    0x8000DFCD00044084, 0x8000DFD58000DFD1, 0x8000DFD90001E8EF,
-    0x000B5CAC000ECCE1, 0x8000DFDD000134E3, 0x0004B0848000DFE1,
-    0x0001E9A58000DFE5, 0x8000DFED8000DFE9, 0x8000DFF10001E1A6,
-    0x0001F9A38000DFF5, 0x00016DAD8000DFF9, 0x000175AF00016DAC,
-    0x8000DFFD00017DAE, 0x00017DAB8000E001, 0x0005B9A80001C1AA,
-    0x000629AA0005B9A9, 0x000639AC000629AB, 0x00062DAE000639AD,
-    0x8000DF9D000625AF, 0x0000198100001182, 0x8000E0050007D980,
-    0x000334BE8000E009, 0x000681B88000E00D, 0x000685BA00068DB9,
-    0x000689BC000699BB, 0x000759BE0006B9BD, 0x00065DB0000759BF,
-    0x0006E1B20006E1B1, 0x0006E1B40006E1B3, 0x0006D1B60006E9B5,
-    0x0007B1A80006D1B7, 0x0007B5AA0007BDA9, 0x0007D9AC0007C9AB,
-    0x0007C9AE0007D9AD, 0x000000860007C1AF, 0x8000E01100032887,
-    0x8000E0198000E015, 0x8000E0218000E01D, 0x000775B88000E025,
-    0x00070DBA000779B9, 0x000701BC000705BB, 0x000705BE000701BD,
-    0x000761B000072DBF, 0x000761B2000761B1, 0x000765B400077DB3,
-    0x000759B600076DB5, 0x00063DB300074DB7, 0x8000E02D8000E029,
-    0x8000E0358000E031, 0x000651B5000659B6, 0x0006A5BB8000E039,
-    0x8000E03D0006A5BA, 0x0006A9BF8000E041, 0x0006A9BD0006A1BE,
-    0x0004A0EF0006B5BC, 0x8000E0498000E045, 0x8000E0518000E04D,
-    0x8000E0598000E055, 0x8000E0618000E05D, 0x8000E0698000E065,
-    0x000454E38000E06D, 0x0006D0E18000E071, 0x0006BDA38000E075,
-    0x0000F58100001582, 0x8000E0790000F980, 0x0006D1A50006D9A6,
-    0x000625AB000190BE, 0x00000886000625AA, 0x000629AF00017087,
-    0x000629AD000621AE, 0x8000E081000635AC, 0x8000E085000175B3,
-    0x00016DB68000E089, 0x8000E0918000E08D, 0x0000EDBA000165B5,
-    0x8000E0950000F5BB, 0x0000DDBE8000E099, 0x0000E5BC0000B1BF,
-    0x0002A9A80000D5BD, 0x0002B9AA0002A9A9, 0x0002A9AC0002B9AB,
-    0x00031DAE0002A9AD, 0x001CACBE000315AF, 0x8000E0A18000E09D,
-    0x8000E0A5001CECBE, 0x8000E0AD8000E0A9, 0x000311B88000E0B1,
-    0x000311BA000311B9, 0x000331BC000311BB, 0x000331BE000331BD,
-    0x00036DB0000331BF, 0x000375B2000375B1, 0x00035DB4000345B3,
-    0x000331B6000331B5, 0x000231A3000331B7, 0x8000E0B98000E0B5,
-    0x8000E0C18000E0BD, 0x000221A5000229A6, 0x0003B1AB8000E0C5,
-    0x8000E0C90003A9AA, 0x0003F5AF8000E0CD, 0x000391AD000399AE,
-    0x0003C0840003A1AC, 0x8000E0D58000E0D1, 0x8000E0DD8000E0D9,
-    0x00003D8100003D82, 0x8000E0E100003D80, 0x8000E0E98000E0E5,
-    0x001CD0868000E0F1, 0x8000E0F5001CB487, 0x0002C1A88000DF2D,
-    0x0002C5AA0002CDA9, 0x0002C9AC0002D9AB, 0x0002F5AE0002FDAD,
-    0x8000E0F9000139AF, 0x8000E1018000E0FD, 0x8000E1098000E105,
-    0x8000E1118000E10D, 0x0001D1B88000E115, 0x0001E1BA0001D9B9,
-    0x000195BC0001E1BB, 0x000195BE00019DBD, 0x000149B000018DBF,
-    0x00015DB2000155B1, 0x00014DB4000155B3, 0x0001F1B60001F1B5,
-    0x0000E4EF0001F1B7, 0x8000E11D8000E119, 0x8000E121001C0084,
-    0x8000E1298000E125, 0x8000E12D00015CEF, 0x8000E1310006D4E1,
-    0x000618E3000630E3, 0x000100E18000E135, 0x8000E13D8000E139,
-    0x8000E1458000E141, 0x8000E149000201B3, 0x000231B600020DB5,
-    0x8000E1518000E14D, 0x0001C1BA8000E155, 0x0001C5BC0001DDBB,
-    0x0001C5BE0001CDBD, 0x001DCDA3000179BF, 0x8000E1598000E0ED,
-    0x8000E15D8000E161, 0x001DC1A5001DFDA6, 0x001E11AB8000E165,
-    0x00030886001E0DAA, 0x001EB5AF00031C87, 0x001E01AD001E09AE,
-    0x001FA180001E09AC, 0x00004182001FA181, 0x8000E169001F6DA3,
-    0x001F75A6001F7DA5, 0x8000E1718000E16D, 0x001F51AA8000E175,
-    0x001FF5AC001F51AB, 0x001FE9AE001FFDAD, 0x001E6DA8001FE9AF,
-    0x001FA9AA001E7DA9, 0x001FB9AC001FA9AB, 0x001FA9AE001FB9AD,
-    0x8000E179001FA9AF, 0x8000E1818000E17D, 0x8000E1898000E185,
-    0x8000E1918000E18D, 0x001F69B88000E195, 0x001F75BA001F71B9,
-    0x001F15BC001F0DBB, 0x001F09BE001F1DBD, 0x001FD9B0001F01BF,
-    0x001F6DB2001FD9B1, 0x001F61B4001F65B3, 0x001F61B6001F61B5,
-    0x8000E199001F5DB7, 0x8000E19D001EB4E1, 0x8000E1A1001DDCE3,
-    0x8000E1A98000E1A5, 0x8000E1B18000E1AD, 0x8000E1B98000E1B5,
-    0x000075808000E1BD, 0x0000758200007D81, 0x001E61B3001DE4EF,
-    0x8000E1C58000E1C1, 0x00019C848000E1C9, 0x001E71B5001E79B6,
-    0x001E5DBB000194BE, 0x8000E1CD001E5DBA, 0x001EE5BF8000E1D1,
-    0x001EF1BD001EE5BE, 0x00018DA8001EF9BC, 0x0001D9AA000199A9,
-    0x0001F1AC0001D1AB, 0x000139AE0001F1AD, 0x00000086000139AF,
-    0x8000E1D500016487, 0x8000E1DD8000E1D9, 0x8000E1E58000E1E1,
-    0x0000D1B88000E1E9, 0x0000D1BA0000D1B9, 0x0000F1BC0000D1BB,
-    0x0000F1BE0000F1BD, 0x000149B00000F1BF, 0x000159B2000149B1,
-    0x000149B4000159B3, 0x0000F9B6000149B5, 0x000251B30000F5B7,
-    0x8000E1F18000E1ED, 0x8000E1F98000E1F5, 0x000279B5000271B6,
-    0x00028DBB8000E1FD, 0x8000E20100028DBA, 0x00028DBF8000E205,
-    0x00028DBD00028DBE, 0x003B24AC00028DBC, 0x8000E209000215A3,
-    0x000235A68000E20D, 0x8000E2158000E211, 0x0002C9AA00023DA5,
-    0x8000E2190002C9AB, 0x0002C9AE8000E21D, 0x0002C9AC0002C9AF,
-    0x8000E07D0002C9AD, 0x8000E2258000E221, 0x8000E22900038084,
-    0x8000E2318000E22D, 0x8000E2398000E235, 0x8000E2458000E241,
-    0x8000E249000400BE, 0x8000E2518000E24D, 0x000059808000E255,
-    0x0000698200005981, 0x8000E25D8000E259, 0x00031487000280EF,
-    0x8000E2610004FC86, 0x8000E265000278E3, 0x8000E269000154E1,
-    0x8000E2718000E26D, 0x00026DA88000E275, 0x0002A5AA0002ADA9,
-    0x0002A5AC0002BDAB, 0x0002A5AE0002ADAD, 0x0005AC84000119AF,
-    0x8000E27D8000E279, 0x0004BCBE8000E281, 0x8000E2898000E285,
-    0x000115B88000E28D, 0x000121BA00011DB9, 0x0001D5BC000121BB,
-    0x0001C9BE0001DDBD, 0x000169B00001C9BF, 0x000179B2000169B1,
-    0x000159B4000171B3, 0x000131B6000159B5, 0x001F68E1000131B7,
-    0x001E14E38000E291, 0x8000E2950004B485, 0x8000E2990001A4E1,
-    0x8000E29D0000DCE3, 0x8000E2A10007A8BF, 0x8000E2A5000008EF,
-    0x001F6CEF8000E2A9, 0x00026DB38000E2B1, 0x0000798100007182,
-    0x0007D48500006580, 0x00026DB500026DB6, 0x0001EDBB8000E2B5,
-    0x000648860001EDBA, 0x0001DDBF00046C87, 0x0001E5BD0001DDBE,
-    0x00060DAB0001EDBC, 0x8000E23D00060DAA, 0x00063DAF00030084,
-    0x000605AD00063DAE, 0x00058DA300060DAC, 0x8000E2B98000E2AD,
-    0x8000E2C18000E2BD, 0x00058DA500058DA6, 0x0006F1B38000E2C5,
-    0x8000E2CD8000E2C9, 0x8000E2D58000E2D1, 0x0006E1B50006E9B6,
-    0x000689BB8000E2D9, 0x8000E2DD000685BA, 0x00068DBF8000E2E1,
-    0x00068DBD00068DBE, 0x000625A800068DBC, 0x000639AA00062DA9,
-    0x000629AC000639AB, 0x000655AE000629AD, 0x8000E2E500064DAF,
-    0x8000E2ED8000E2E9, 0x8000E2F58000E2F1, 0x8000E2FD8000E2F9,
-    0x0007EDB88000E301, 0x0007F1BA0007F1B9, 0x000755BC0007F1BB,
-    0x000741BE000745BD, 0x000601B000077DBF, 0x000619B200060DB1,
-    0x000609B4000619B3, 0x0007DDB6000609B5, 0x0007B1A30007D5B7,
-    0x8000E30503EDED8F, 0x8000E30D8000E309, 0x0007A1A50007A9A6,
-    0x0007C9AB8000E311, 0x8000E3150007C5AA, 0x0007CDAF8000E319,
-    0x0007CDAD0007CDAE, 0x00197D9E0007CDAC, 0x001BB59C0007C59F,
-    0x001D719A00196D9D, 0x00118D98001DC59B, 0x00138596001FB599,
-    0x0015459400114597, 0x03E85D9200144995, 0x03E979900017A593,
-    0x00009D8003E97991, 0x0000A5820000AD81, 0x000454E10000C0BE,
-    0x0038A4E3000040EF, 0x000F6C848000E321, 0x003878EF000214E3,
-    0x0004D9A200027CE1, 0x0005BDA00001BDA3, 0x0001E9B30005ADA1,
-    0x000C00868000E325, 0x8000E32900013087, 0x0001E9B50001E9B6,
-    0x0001B1BB8000E32D, 0x8000E3310001A9BA, 0x000151BF8000E335,
-    0x000191BD000199BE, 0x03EA01830001A1BC, 0x8000E33903EEED82,
-    0x00163D878000E33D, 0x03EB35850014C586, 0x0010C58B03EA1984,
-    0x8000E31D0017018A, 0x001E118F8000E341, 0x0012ED8D001CDD8E,
-    0x0018BD9300121D8C, 0x8000E345001ED592, 0x000655978000E349,
-    0x001A65950004D596, 0x0007099B001A7D94, 0x8000E34D0006759A,
-    0x8000E3558000E351, 0x8000E35D8000E359, 0x0002A5A10002759C,
-    0x000E8DA38000E361, 0x000801A50000A5A2, 0x000ABDA7000EFDA4,
-    0x003681A9000ABDA6, 0x0037E5AB003401A8, 0x0032A1AD003689AA,
-    0x003C01AF0032B5AC, 0x003FD9B10033EDAE, 0x003AE9B3003EE5B0,
-    0x002401B5003801B2, 0x8000E365003A49B4, 0x8000E36D8000E369,
-    0x8000E3758000E371, 0x8000E37D8000E379, 0x0001DD810001DD82,
-    0x0004BDA80001DD80, 0x000321AA0004D9A9, 0x000321AC000321AB,
-    0x00035DAE000329AD, 0x8000E381000355AF, 0x8000E3898000E385,
-    0x8000E391000DE4BE, 0x000D18868000E395, 0x0003EDB8000D9C87,
-    0x0003FDBA0003F5B9, 0x0003EDBC0003F5BB, 0x000255BE00025DBD,
-    0x00032DB000024DBF, 0x000335B2000335B1, 0x00031DB4000305B3,
-    0x0003DDB60003D5B5, 0x000C4C840003D5B7, 0x8000E3990006B4E3,
-    0x8000E39D000748E1, 0x0000ECEF8000E3A1, 0x000650E18000E3A5,
-    0x0001B4E38000E3A9, 0x000C78BE8000E3AD, 0x8000E3B100067CEF,
-    0x0003E9A38000E3B5, 0x8000E3BD8000E3B9, 0x000F5C848000E3C1,
-    0x0003E9A50003E9A6, 0x0003B1AB8000E3C5, 0x8000E3C90003A9AA,
-    0x000351AF8000E3CD, 0x000391AD000399AE, 0x000DD1B30003A1AC,
-    0x8000E3D18000E38D, 0x8000E3D98000E3D5, 0x000DF1B5000DF9B6,
-    0x000E69BB8000E3DD, 0x8000E3E1000E09BA, 0x000E41BF8000E3E5,
-    0x000E51BD000E59BE, 0x8000E3E9000E71BC, 0x8000E3ED000D95A3,
-    0x000DBDA68000E3F1, 0x8000E3F98000E3F5, 0x000E4DAA000DB5A5,
-    0x8000E3FD000E2DAB, 0x000E1DAE8000E401, 0x000E35AC000E05AF,
-    0x000FB1A8000E15AD, 0x000F01AA000FB1A9, 0x000F05AC000F1DAB,
-    0x000F3DAE000F09AD, 0x0000ED80000F31AF, 0x0000F1820000F181,
-    0x000390868000E405, 0x8000E4090000B487, 0x0000C9B88000E40D,
-    0x0000D9BA0000C9B9, 0x0000F9BC0000D1BB, 0x000099BE0000F9BD,
-    0x000F59B0000095BF, 0x000F2DB2000F59B1, 0x000F3DB4000F25B3,
-    0x000F11B6000F29B5, 0x000E89B3000F11B7, 0x8000E4158000E411,
-    0x8000E41D8000E419, 0x000E8DB5000E89B6, 0x000E9DBB8000E421,
-    0x8000E425000E9DBA, 0x000EEDBF8000E429, 0x000EF1BD000EF9BE,
-    0x8000E42D000EF9BC, 0x8000E431000ECDA3, 0x000ECDA68000E435,
-    0x8000E43D8000E439, 0x000ED9AA000EC9A5, 0x8000E441000ED9AB,
-    0x000EBDAE8000E445, 0x000EBDAC000EA9AF, 0x0001EDA8000EB5AD,
-    0x000131AA000129A9, 0x000095AC000131AB, 0x000095AE00009DAD,
-    0x8000E44900008DAF, 0x8000E4518000E44D, 0x8000E4598000E455,
-    0x8000E4618000E45D, 0x0000A9B88000E465, 0x000069BA0000A9B9,
-    0x000079BC000069BB, 0x000069BE000079BD, 0x0000FDB0000069BF,
-    0x0000C5B20000CDB1, 0x0000B9B40000ADB3, 0x0000ADB60000B9B5,
-    0x0003B9800000A5B7, 0x000005820003B981, 0x000208BE8000E469,
-    0x8000E4758000E471, 0x000480868000E479, 0x8000E47D0002E887,
-    0x8000E4858000E481, 0x8000E48D8000E489, 0x0002F9B38000E491,
-    0x8000E4998000E495, 0x8000E4A18000E49D, 0x00035DB500035DB6,
-    0x000349BB8000E4A5, 0x8000E4A9000379BA, 0x00033DBF8000E4AD,
-    0x000339BD00033DBE, 0x000389A6000351BC, 0x8000E4B58000E4B1,
-    0x8000E4B9000389A5, 0x8000E4BD00022DA3, 0x0003E9AE8000E4C1,
-    0x000385AC0003E9AF, 0x0003ADAA0003EDAD, 0x8000E4C500039DAB,
-    0x8000E4CD8000E4C9, 0x00001D8100001D82, 0x8000E4D100001D80,
-    0x8000E4D98000E4D5, 0x0004AC84000494BE, 0x000328878000E4E1,
-    0x8000E4E50005CC86, 0x8000E4ED8000E4E9, 0x000D4DB28000E4F1,
-    0x000D25B0000D6DB3, 0x000D65B6000D45B1, 0x000D75B4000D6DB7,
-    0x000DFDBA000D6DB5, 0x000DADB8000DC5BB, 0x000DDDBE000DF5B9,
-    0x000DDDBC000DC5BF, 0x8000E4F5000DD5BD, 0x8000E4FD8000E4F9,
-    0x8000E5058000E501, 0x8000E50D8000E509, 0x000EB1AA8000E511,
-    0x000EF5A8000EB9AB, 0x000D79AE000EA9A9, 0x000E91AC000D79AF,
-    0x000EA0E1000E91AD, 0x000E18E38000E515, 0x8000E51D8000E519,
-    0x8000E5258000E521, 0x8000E52D8000E529, 0x8000E5358000E531,
-    0x8000E53D8000E539, 0x000EF0EF8000E541, 0x000075808000E545,
-    0x0000198200007981, 0x8000E54900044084, 0x000754BE00033CEF,
-    0x8000E5558000E551, 0x8000E559000E24E3, 0x0007F086000144E1,
-    0x8000E55D0004F487, 0x000615A88000E561, 0x000615AA00061DA9,
-    0x00063DAC000625AB, 0x00062DAE000625AD, 0x8000E4DD000625AF,
-    0x8000E5658000E54D, 0x8000E56D8000E569, 0x8000E5758000E571,
-    0x0006C5B88000E579, 0x0006D9BA0006CDB9, 0x0006CDBC0006D5BB,
-    0x0006E5BE0006EDBD, 0x00065DB000069DBF, 0x000621B2000621B1,
-    0x000621B4000621B3, 0x000611B6000629B5, 0x00069DB3000611B7,
-    0x8000E5818000E57D, 0x8000E58500000C84, 0x0006B5B50006A5B6,
-    0x0006E5BB8000E589, 0x8000E58D000699BA, 0x0006EDBF8000E591,
-    0x0006F5BD0006E5BE, 0x8000E5950006F5BC, 0x8000E5990006D9A3,
-    0x0006E1A68000E59D, 0x8000E5A58000E5A1, 0x0006DDAA0006F1A5,
-    0x8000E5A90006A1AB, 0x0006A1AE8000E5AD, 0x0006B1AC0006A9AF,
-    0x00000D800006B1AD, 0x0000118200001581, 0x8000E5B58000E5B1,
-    0x8000E5B900052CE1, 0x0000ACBE0005ECE3, 0x8000E5C58000E5C1,
-    0x003F9086000508EF, 0x8000E5C900017487, 0x0007B5B38000E5CD,
-    0x8000E5D58000E5D1, 0x8000E5DD8000E5D9, 0x0007A5B5000615B6,
-    0x000631BB8000E5E1, 0x8000E5E5000629BA, 0x0006F5BF8000E5E9,
-    0x000621BD000629BE, 0x000765A8000629BC, 0x00077DAA000775A9,
-    0x00076DAC000775AB, 0x0006DDAE0006D5AD, 0x8000E5ED0006D5AF,
-    0x8000E5F58000E5F1, 0x8000E5FD8000E5F9, 0x8000E6058000E601,
-    0x000655B88000E609, 0x000655BA00065DB9, 0x000675BC00066DBB,
-    0x000675BE00067DBD, 0x0006ADB000066DBF, 0x0006BDB20006B5B1,
-    0x0006ADB40006B5B3, 0x000675B6000671B5, 0x0004F5A300066DB7,
-    0x8000E6118000E60D, 0x8000E6198000E615, 0x0004E5A5000555A6,
-    0x000571AB8000E61D, 0x8000E621000569AA, 0x0005B5AF8000E625,
-    0x000561AD000569AE, 0x00006980000569AC, 0x0000198200006981,
-    0x8000E6290005A5B3, 0x000271B60005A9B5, 0x000320BE0003D484,
-    0x00025DBA8000E631, 0x00024DBC000255BB, 0x0002F1BE0002F1BD,
-    0x8000E6350002EDBF, 0x000348878000E639, 0x8000E46D003CAC86,
-    0x8000E6418000E63D, 0x00057DA8003D90BE, 0x000541AA000545A9,
-    0x00054DAC000555AB, 0x000295AE00029DAD, 0x0002F5B000028DAF,
-    0x0002F5B20002FDB1, 0x000259B4000249B3, 0x000245B6000245B5,
-    0x00024DB8000275B7, 0x000221BA000225B9, 0x000225BC000221BB,
-    0x000211BE00022DBD, 0x0004E9A3000211BF, 0x8000E6498000E645,
-    0x8000E6518000E64D, 0x0004E5A500033DA6, 0x000319AB8000E655,
-    0x003F2CBE000311AA, 0x0003A1AF8000E659, 0x0003BDAD0003BDAE,
-    0x8000E65D000301AC, 0x8000E66100027DB3, 0x000219B68000E665,
-    0x8000E66D8000E669, 0x00023DBA000219B5, 0x8000E67100020DBB,
-    0x0002D5BE8000E675, 0x000215BC0002DDBF, 0x8000E6790002DDBD,
-    0x8000E6818000E67D, 0x8000E685003C0084, 0x000040EF8000E689,
-    0x000648E10006BCEF, 0x000134E3000750E1, 0x8000E68D000670E3,
-    0x8000E6958000E691, 0x000075808000E699, 0x0000758200007D81,
-    0x8000E6A10002F5A3, 0x000291A6000291A5, 0x003F40868000E6A5,
-    0x0002B5AA003CD487, 0x00029DAC000285AB, 0x00025DAE000255AD,
-    0x003D25A8000255AF, 0x003D31AA003D2DA9, 0x003D55AC003D31AB,
-    0x003D55AE003D5DAD, 0x8000E62D003D4DAF, 0x8000E6A98000E69D,
-    0x8000E6B18000E6AD, 0x8000E6B98000E6B5, 0x003DF5B88000E6BD,
-    0x003DF5BA003DFDB9, 0x003E59BC003E49BB, 0x003E49BE003E59BD,
-    0x003D3DB0003E41BF, 0x003D05B2003D0DB1, 0x003DD5B4003DCDB3,
-    0x003DD5B6003DDDB5, 0x003EE5A8003DCDB7, 0x003E0DAA003E05A9,
-    0x003E09AC003E19AB, 0x003F91AE003F91AD, 0x0003A884003F91AF,
-    0x8000E6C58000E6C1, 0x8000E6CD8000E6C9, 0x8000E6D58000E6D1,
-    0x003FA9B88000E6D9, 0x003F49BA003FA9B9, 0x003F59BC003F49BB,
-    0x003F4DBE003F45BD, 0x003FF1B0003F6DBF, 0x003FC5B2003FF9B1,
-    0x003FB9B4003FADB3, 0x003FADB6003FB9B5, 0x8000E6DD003FA5B7,
-    0x8000E6E58000E6E1, 0x8000E6ED8000E6E9, 0x003D80EF8000E6F1,
-    0x003E08E18000E6F5, 0x003D64E38000E6F9, 0x00000D808000E6FD,
-    0x0000718200007181, 0x003FC5B38000E701, 0x8000E7098000E705,
-    0x8000E7118000E70D, 0x003FC1B5003E51B6, 0x003E7DBB8000E715,
-    0x0000C886003E75BA, 0x003E3DBF0001BC87, 0x003E39BD003E3DBE,
-    0x003F05A3003E51BC, 0x8000E71D8000E719, 0x8000E7258000E721,
-    0x003F01A5003E91A6, 0x003EBDAB8000E729, 0x8000E72D003EB5AA,
-    0x003EFDAF8000E731, 0x003EF9AD003EFDAE, 0x8000E735003E91AC,
-    0x8000E7390001EDB3, 0x0001E5B68000E73D, 0x8000E7458000E741,
-    0x000145BA0001EDB5, 0x8000E74900014DBB, 0x000145BE8000E74D,
-    0x000155BC00014DBF, 0x003DD1A800014DBD, 0x000221AA003DD9A9,
-    0x000221AC000221AB, 0x000221AE000221AD, 0x000400BE000221AF,
-    0x8000E7558000E751, 0x8000E75D8000E759, 0x8000E7658000E761,
-    0x000345B88000E769, 0x000345BA00034DB9, 0x000345BC00035DBB,
-    0x00037DBE000349BD, 0x0002C5B0000371BF, 0x0002C5B20002CDB1,
-    0x0002C5B40002DDB3, 0x0002C5B60002CDB5, 0x8000E76D00037DB7,
-    0x0002A1A38000E771, 0x0002A1A58000E775, 0x000380848000E779,
-    0x00052CBE0002A9A6, 0x000201AB8000E5BD, 0x000201AD000209AA,
-    0x000201AF000219AC, 0x000248E3000209AE, 0x0001B8E18000E781,
-    0x000055808000E785, 0x00004D8200005981, 0x000460868000E789,
-    0x8000E78D0004E887, 0x000228EF8000E791, 0x8000E7998000E795,
-    0x000251A88000E79D, 0x00028DAA000259A9, 0x00029DAC000285AB,
-    0x000285AE000285AD, 0x8000E7A10002B5AF, 0x8000E7A98000E7A5,
-    0x000538848000E7AD, 0x8000E7B58000E7B1, 0x000165B88000E7B9,
-    0x000165BA00016DB9, 0x000165BC00017DBB, 0x000165BE00016DBD,
-    0x0002CDB00001DDBF, 0x0002A1B20002A5B1, 0x0002A5B40002A1B3,
-    0x000291B60002ADB5, 0x8000E7BD000291B7, 0x8000E7C58000E7C1,
-    0x8000E7C9000750BE, 0x003E48EF8000E7CD, 0x003E58E1003FF8EF,
-    0x000118E3003F30E1, 0x8000E7D1003E30E3, 0x8000E7D98000E7D5,
-    0x000251B68000E7DD, 0x8000E7E58000E7E1, 0x8000E7E9000241B5,
-    0x00049887000251B3, 0x000235BE0004FC86, 0x000221BC00023DBF,
-    0x00025DBA000221BD, 0x8000E7F1000235BB, 0x8000E77D8000E7F5,
-    0x0007C1810007D582, 0x8000E7ED0007C980, 0x8000E7FD8000E7F9,
-    0x0006E5A88000E801, 0x0006F5AA0006E9A9, 0x0006DDAC0006C5AB,
-    0x0006C1AE0006C1AD, 0x000761B00006C1AF, 0x000761B2000761B1,
-    0x000765B4000761B3, 0x000765B600076DB5, 0x0007E5B80007DDB7,
-    0x0007F9BA0007EDB9, 0x0007C1BC0007F9BB, 0x0007C1BE0007C1BD,
-    0x8000E8050007C1BF, 0x000525AA8000E809, 0x000559AC00054DAB,
-    0x00054DAE000559AD, 0x8000E80D000545AF, 0x8000E8158000E811,
-    0x8000E819000529A3, 0x000529A6000539A5, 0x8000E8218000E81D,
-    0x8000E825000619B3, 0x000625B68000E829, 0x8000E8318000E82D,
-    0x0007C1BA000631B5, 0x8000E8350007C1BB, 0x0007C5BE8000E839,
-    0x0007D1BC0007CDBF, 0x0007DDA30007D1BD, 0x8000E8418000E83D,
-    0x8000E8498000E845, 0x0007F5A50007E1A6, 0x000605AB8000E84D,
-    0x8000E851000605AA, 0x000609AF8000E855, 0x000615AD000601AE,
-    0x00006980000615AC, 0x0000058200006981, 0x8000E85D8000E859,
-    0x8000E86100019084, 0x0000C0868000E865, 0x8000E8690001EC87,
-    0x8000E8718000E86D, 0x8000E8798000E875, 0x000679A88000E87D,
-    0x00019DAA000679A9, 0x00018DAC000195AB, 0x00011DAE000115AD,
-    0x0001E88400013DAF, 0x8000E8858000E881, 0x8000E88D8000E889,
-    0x8000E8958000E891, 0x000119B88000E899, 0x000125BA00012DB9,
-    0x0000D9BC0000C9BB, 0x0000C9BE0000D9BD, 0x000145B00000C1BF,
-    0x000159B200014DB1, 0x000131B4000151B3, 0x000135B600013DB5,
-    0x8000E89D000129B7, 0x00021DB38000E8A1, 0x00020DB5000344BE,
-    0x8000E8A98000E8A5, 0x000C6C8400029DB6, 0x000281BB8000E8AD,
-    0x00028DBD000289BA, 0x0002F1BF000285BC, 0x8000E8B10002FDBE,
-    0x000255A50002C5A6, 0x000245A38000E8B5, 0x8000E8BD8000E8B9,
-    0x0002A9AF8000E8C1, 0x0002D5AD0002A5AE, 0x0002D9AB0002DDAC,
-    0x8000E8C50002D1AA, 0x0000FD828000E8C9, 0x0000D5800003EDB3,
-    0x0003E5B60000F581, 0x8000E8D58000E8D1, 0x000305BA0003EDB5,
-    0x0003608700030DBB, 0x000305BE000C0C86, 0x000315BC00030DBF,
-    0x8000E8D900030DBD, 0x8000E8E18000E8DD, 0x8000E8E98000E8E5,
-    0x8000E8ED0001ACE1, 0x8000E8F100016CE3, 0x8000E8F98000E8F5,
-    0x8000E8FD000218EF, 0x8000E9058000E901, 0x000259A88000E909,
-    0x000261AA000259A9, 0x000261AC000261AB, 0x000261AE000261AD,
-    0x000D68BE000261AF, 0x8000E9118000E90D, 0x8000E9198000E915,
-    0x8000E9218000E91D, 0x000145B88000E925, 0x000151BA000151B9,
-    0x000111BC000169BB, 0x00010DBE000119BD, 0x0002C5B0000105BF,
-    0x0002C5B20002CDB1, 0x0002C5B40002DDB3, 0x0002C5B60002CDB5,
-    0x8000E92900017DB7, 0x8000E92D00077CE1, 0x8000E9310006ACE3,
-    0x8000E935000650E1, 0x8000E9390000D8E3, 0x8000E9458000E941,
-    0x003740E10006D8EF, 0x8000E94D8000E949, 0x000061820007E4EF,
-    0x000059800002ADA3, 0x0002A5A600005581, 0x8000E9558000E951,
-    0x000245AA0002ADA5, 0x000CC88600024DAB, 0x000245AE000CE487,
-    0x000255AC00024DAF, 0x000E95A800024DAD, 0x000EC9AA000E9DA9,
-    0x000ED9AC000EC9AB, 0x000EC9AE000ED9AD, 0x8000E93D000EC5AF,
-    0x8000E95D8000E959, 0x8000E9658000E961, 0x8000E96D8000E969,
-    0x000F69B88000E971, 0x000F79BA000F69B9, 0x000F69BC000F79BB,
-    0x000FD9BE000F69BD, 0x000E81B0000FD5BF, 0x000E81B2000E81B1,
-    0x000E85B4000E9DB3, 0x000E85B6000E8DB5, 0x000E41B3000F59B7,
-    0x8000E9798000E975, 0x8000E9818000E97D, 0x000E51B5000E59B6,
-    0x000E45BB8000E985, 0x8000E989000E45BA, 0x000E49BF8000E98D,
-    0x000E49BD000E41BE, 0x8000E991000E55BC, 0x8000E995000E05A3,
-    0x000E1DA68000E999, 0x8000E9A18000E99D, 0x000E01AA000E15A5,
-    0x8000E9A5000E01AB, 0x000E05AE8000E9A9, 0x000E11AC000E0DAF,
-    0x0001ED82000E0DAD, 0x0001D180000CDCEF, 0x00029C840001E581,
-    0x8000E9AD000194BE, 0x000080868000E9B1, 0x8000E9B500016487,
-    0x8000E9BD8000E9B9, 0x8000E9C1000D04E3, 0x000EF5B3000D88E1,
-    0x8000E9C98000E9C5, 0x8000E9D18000E9CD, 0x000E8DB5000E8DB6,
-    0x000151BB8000E9D5, 0x8000E9D9000149BA, 0x0001F5BF8000E9DD,
-    0x000141BD000149BE, 0x000E25A8000149BC, 0x000E3DAA000E2DA9,
-    0x000E45AC000E35AB, 0x0001BDAE000E4DAD, 0x0001AC840001B5AF,
-    0x8000E9E58000E9E1, 0x8000E9ED8000E9E9, 0x8000E9F58000E9F1,
-    0x000151B88000E9F9, 0x000161BA000159B9, 0x000111BC000161BB,
-    0x000115BE00011DBD, 0x0001CDB0000109BF, 0x0001DDB20001D5B1,
-    0x0001CDB40001D5B3, 0x000171B6000171B5, 0x8000E9FD000171B7,
-    0x8000EA058000EA01, 0x8000EA09000DB9A3, 0x000DC1A6000DC1A5,
-    0x8000EA118000EA0D, 0x000205AA8000EA15, 0x000205AC00021DAB,
-    0x000205AE00020DAD, 0x000061800002B9AF, 0x0000198200006181,
-    0x8000EA1900030CBE, 0x000378878000EA21, 0x8000EA250004FC86,
-    0x8000EA2D8000EA29, 0x8000EA358000EA31, 0x8000EA3D8000EA39,
-    0x8000EA458000EA41, 0x8000EA4D8000EA49, 0x8000EA558000EA51,
-    0x8000EA59000124E3, 0x0004EC840001A4E1, 0x8000EA5D000210EF,
-    0x8000EA658000EA61, 0x8000E8CD8000EA69, 0x0003EDB38000EA6D,
-    0x8000EA758000EA71, 0x8000EA7D8000EA79, 0x0003E1B5000391B6,
-    0x0003BDBB8000EA81, 0x8000EA850003B5BA, 0x000359BF8000EA89,
-    0x000395BD000351BE, 0x00021DA8000395BC, 0x000225AA00022DA9,
-    0x000229AC00023DAB, 0x000251AE000251AD, 0x0001FD80000251AF,
-    0x0001C5820001CD81, 0x8000EA91000564BE, 0x000504878000EA95,
-    0x0002F5B800041C86, 0x0002F5BA0002FDB9, 0x000155BC00014DBB,
-    0x000149BE00015DBD, 0x000231B0000141BF, 0x000201B2000239B1,
-    0x0002D5B4000201B3, 0x0002D5B60002DDB5, 0x8000EA990002CDB7,
-    0x8000EAA18000EA9D, 0x8000EAA500044084, 0x8000EAAD8000EAA9,
-    0x8000EAB1000070EF, 0x8000EAB5000E54E1, 0x000431B50000ACE3,
-    0x8000EABD8000EAB9, 0x8000EAC58000EAC1, 0x8000EACD8000EAC9,
-    0x8000EAD10002E1A3, 0x00029DA60002EDA5, 0x8000EAD98000EAD5,
-    0x0002B9AA8000EADD, 0x000299AC0002B1AB, 0x00025DAE000299AD,
-    0x000655A8000255AF, 0x00066DAA000665A9, 0x00067DAC000665AB,
-    0x00066DAE000665AD, 0x8000EA8D000665AF, 0x8000EAE58000EAE1,
-    0x8000EAED8000EAE9, 0x0000B1810000B982, 0x0006EDB80000B980,
-    0x0006FDBA0006F5B9, 0x000695BC0006F1BB, 0x000695BE00069DBD,
-    0x00061DB000068DBF, 0x0006EDB20006E5B1, 0x0006FDB40006E5B3,
-    0x0006E1B60006E5B5, 0x0006E5B30006D5B7, 0x8000EAF58000EAF1,
-    0x000010BE8000EAF9, 0x0006F5B50006E5B6, 0x000621BB8000EAFD,
-    0x0000E8860006D9BA, 0x000621BF0003D087, 0x000631BD000639BE,
-    0x8000EB01000639BC, 0x8000EB050006A1A3, 0x0006A1A68000EB09,
-    0x8000EB118000EB0D, 0x00069DAA0006B1A5, 0x8000EB15000665AB,
-    0x00067DAE8000EB19, 0x00067DAC000665AF, 0x000681B3000675AD,
-    0x8000EB218000EB1D, 0x8000EB298000EB25, 0x000691B5000685B6,
-    0x000689BB8000EB2D, 0x8000EB31000685BA, 0x0006F1BF8000EB35,
-    0x000681BD000689BE, 0x0005F0E3000689BC, 0x00050CE18000EB39,
-    0x8000EB418000EB3D, 0x8000EB498000EB45, 0x8000EB518000EB4D,
-    0x8000EB598000EB55, 0x0004A4EF8000EB5D, 0x0000DD810000DD82,
-    0x0006B1A80000DD80, 0x000609AA0006B1A9, 0x000619AC000609AB,
-    0x00060DAE000605AD, 0x8000EB61000605AF, 0x8000EB698000EB65,
-    0x00007CBE00013484, 0x001F18868000EB71, 0x000611B800008C87,
-    0x000629BA000619B9, 0x0001F5BC000629BB, 0x0001F5BE0001FDBD,
-    0x000641B00001EDBF, 0x00065DB200064DB1, 0x00064DB4000655B3,
-    0x00063DB6000635B5, 0x8000EB75000635B7, 0x8000EB7D8000EB79,
-    0x8000EB8100058DA3, 0x000589A600059DA5, 0x8000EB898000EB85,
-    0x000589AA8000EB8D, 0x000585AC000585AB, 0x000585AE00058DAD,
-    0x0002A9B60005FDAF, 0x000348848000EB91, 0x8000EB95000295B5,
-    0x8000EB99000299B3, 0x000349BE8000EB9D, 0x000379BC000349BF,
-    0x000379BA000371BD, 0x8000EBA1000379BB, 0x000389B08000EBA5,
-    0x000399B2000389B1, 0x0003B9B4000391B3, 0x0003DDB60003B9B5,
-    0x0003F1B80003D1B7, 0x0003F1BA0003F1B9, 0x0003D1BC0003F1BB,
-    0x0003D1BE0003D1BD, 0x8000EBA90003D1BF, 0x8000EBB18000EBAD,
-    0x8000EBB58000EA1D, 0x8000EBBD8000EBB9, 0x000221A88000EBC1,
-    0x000225AA00022DA9, 0x000229AC000239AB, 0x000215AE00021DAD,
-    0x8000EBC50003F9AF, 0x8000EBCD8000EBC9, 0x0000ACE38000EBD1,
-    0x0001A8E18000EBD5, 0x000068EF8000EBD9, 0x001D9085001D28BE,
-    0x00003D80001C04BF, 0x0000258200002181, 0x00038DA380001C4F,
-    0x000208AC80001C53, 0x80001C5B80001C57, 0x000381A50003BDA6,
-    0x00026DAB80001C5F, 0x80001C6300026DAA, 0x00025DAF80001C67,
-    0x000265AD00025DAE, 0x80001C6B00026DAC, 0x80001C6F00027CE1,
-    0x0002E4AC0002ACE1, 0x80001C7300001DB5, 0x000039B50001F4EF,
-    0x0002BCAC000680E1, 0x0002A8E1000630E3, 0x0002C0E180001C77,
-    0x000075B5001D2CE1, 0x0002FCAC80001C7B, 0x80001C7F0001C9B3,
-    0x0001C5B60001C5B5, 0x000025B580001C83, 0x0001ADBA001D7CE1,
-    0x0001A9BC0001BDBB, 0x000181BE0001A9BD, 0x001DB5A8000181BF,
-    0x001E0DAA001E05A9, 0x001E01AC001E05AB, 0x001E39AE001E09AD,
-    0x0002D0AC001E39AF, 0x0002D0AC001FF5B5, 0x001CADA2001FF1B5,
-    0x001EF5A00002E5A3, 0x001E89B8001C1DA1, 0x001E9DBA001E95B9,
-    0x001E8DBC001E95BB, 0x001F5DBE001F55BD, 0x001E25B0001F55BF,
-    0x001E31B2001E35B1, 0x001E09B4001E31B3, 0x001EB9B6001E09B5,
-    0x001E85A3001EB9B7, 0x0002B4E10024A18F, 0x00000DB50002C8AC,
-    0x001E89A5001E89A6, 0x001EF1AB80001C87, 0x0002B4E1001EE1AA,
-    0x001ECDAF000370AC, 0x001EE5AD001ECDAE, 0x0018E19E001EE5AC,
-    0x0015999C0019DD9F, 0x0014D19A001A759D, 0x0010FD980014C59B,
-    0x0012459600119199, 0x002CC9940010C597, 0x00289992002CC995,
-    0x0026D190002E6593, 0x002D258A0028F991, 0x000005B50013A18B,
-    0x00105D8E80001C8B, 0x0011398C0015158F, 0x002505820011398D,
-    0x0002B4E100270D83, 0x002FA1860002B4AC, 0x00293D84002D1587,
-    0x001D619A00295585, 0x80001C8F001DD99B, 0x80000CC3000005B5,
-    0x0003899C001EF4E1, 0x00157192001D04E3, 0x80003B8300143593,
-    0x00186996001CB0EF, 0x001BC194001FB997, 0x001EBDB800195D95,
-    0x001EC5BA001EC1B9, 0x001ECDBC001EDDBB, 0x800018DF001EFDBD,
-    0x800018E7800018E3, 0x800018EF800018EB, 0x800018F7800018F3,
-    0x800018FF800018FB, 0x8000190780001903, 0x800019138000190B,
-    0x8000191B80001917, 0x800019238000191F, 0x8000192B80001927,
-    0x800019338000192F, 0x8000193B80001937, 0x800019438000193F,
-    0x003BB48580001947, 0x00163C878000194B, 0x8000194F0002F0CC,
-    0x8000195780001953, 0x001E148D0008C5B8, 0x8000195F8000195B,
-    0x001E308980001963, 0x001E308B80001967, 0x8000196B0008DDB8,
-    0x8000196F000298CC, 0x001E6C9180001973, 0x8000197B80001977,
-    0x800019838000197F, 0x8000198B80001987, 0x800019938000198F,
-    0x80000AC680000AC3, 0x023779BC80001C93, 0x023AA1BE023C55BD,
-    0x02367DB8023ABDBF, 0x023709BA0237A9B9, 0x023699B480000AE4,
-    0x02361DB60236F1B5, 0x0236B9B0023669B7, 0x023F65B20236A5B1,
-    0x0232F5AC02368DB3, 0x023105AE023E41AD, 0x8000EB6D80001C97,
-    0x80001CA380001C9B, 0x80000AC980001CA7, 0x0001C4BF001E84BC,
-    0x000819B80001E0BE, 0x0070698C80000AD2, 0x0228D980001EC4BA,
-    0x000259820228D981, 0x022EF184022ED183, 0x022D3986022EF585,
-    0x00024988022D8187, 0x000338CC80000AEA, 0x80000AF00007A4CC,
-    0x80000B0B80000AFC, 0x0002D99080001CAF, 0x0002E9920002D991,
-    0x0002F9940002E993, 0x0002E9960002F995, 0x0002D9980002E997,
-    0x0002A99A0002D999, 0x021B199C0002A99B, 0x0002B59E0002BD9D,
-    0x000255A00002AD9F, 0x000255A200025DA1, 0x000275A400026DA3,
-    0x000275A600027DA5, 0x000255A800026DA7, 0x000255AA00025DA9,
-    0x0001A1AC0001A9AB, 0x0001A1AE0001A1AD, 0x0001E9B0000199AF,
-    0x0001FDB20001E9B1, 0x000199B400018DB3, 0x00018DB6000199B5,
-    0x0001BDB8000185B7, 0x00018DBA000185B9, 0x80000B1480000B89,
-    0x000195BE00019DBD, 0x8000085300018DBF, 0x8000071D0006E59B,
-    0x0238A19E80001CB7, 0x0231E99C0006899F, 0x80000B8F0006E99D,
-    0x80000B1D80000B20, 0x80000B2980000B26, 0x80000B4D80000B35,
-    0x80001C9F800009F1, 0x80000B7480000B71, 0x80001D7780001D2F,
-    0x8000067480001D87, 0x80001DFF800008B0, 0x80001E5780001E2F,
-    0x80001EB780001EB3, 0x80001EBF80001EBB, 0x023B01BA80001EC3,
-    0x0000C9B8023911BB, 0x022499BE023A05B9, 0x0000DDBC022615BF,
-    0x000095B20000D5BD, 0x000099B000009DB3, 0x023BCDB6023AB1B1,
-    0x023AE5B4023A31B7, 0x0000CDAA000081B5, 0x000725A80000E5AB,
-    0x023AC9AE00073DA9, 0x0000FDAC0000F9AF, 0x0007D1A20000F5AD,
-    0x0230C5A00007D1A3, 0x0007C9A60007C9A1, 0x023901A40007C9A7,
-    0x02348D80023909A5, 0x0237058202306D80, 0x0237018402307582,
-    0x02370D860230AD84, 0x023765880230DD86, 0x0237798A0230F188,
-    0x0237518C0230E58A, 0x0237498E02331D8C, 0x023729900233098E,
-    0x0237259202337D90, 0x0237F99402332992, 0x0237F5960232CD94,
-    0x023671980232C196, 0x02367D9A0232C998, 0x0236799C00013D9A,
-    0x02365D9E0232BD9C, 0x0236B5A00232A99E, 0x0236D1A202326DA0,
-    0x0236FDA40001D9A2, 0x0236C5A6023275A4, 0x023635A8023271A6,
-    0x023609AA023DB9A8, 0x02311DAC023DC9AA, 0x0231CDAE023DC1AC,
-    0x023109B0023D2DAE, 0x02311DB2023D49B0, 0x023135B4023C99B2,
-    0x000195B6023CB5B4, 0x0230EDB8023C81B6, 0x0230D9BA023D49B8,
-    0x0230D1BC023D55BA, 0x0230ADBE023C89BC, 0x80001EC7023C75BE,
-    0x80001ECF80001ECB, 0x80001ED780001ED3, 0x80001EDF80001EDB,
-    0x80001EF780001EE3, 0x80001EFF80001EFB, 0x80001F0780001F03,
-    0x80001F0B022B81B8, 0x023F25A580001F0F, 0x022755A7023ED1A4,
-    0x0222A9A1022759A6, 0x023E69A30222FDA0, 0x80001F13000255A2,
-    0x000C99B880001F17, 0x0222D9A980001F1B, 0x80001F37022A6DA8,
-    0x00045595022285AA, 0x0002899700044994, 0x023B1991023EA596,
-    0x023B59930233A990, 0x0227B59D02336592, 0x0227919F0004799C,
-    0x000469990227B19E, 0x023FDD9B023E2198, 0x0004B9850238B59A,
-    0x023505870004B184, 0x023E45810004B186, 0x0236ED83023E8D80,
-    0x0232D18D023F5182, 0x02322D8F0004798C, 0x0232E18980001F3B,
-    0x023B1D8B0232FD88, 0x000DF9B8023B1D8A, 0x80001F53023D4D92,
-    0x80001F6F023D3D90, 0x80001F7780001F73, 0x0230C99B023D5194,
-    0x80001F7B00037D9A, 0x02382CCC80001F7F, 0x023960CC0003699E,
-    0x80001F83023184CD, 0x80001F87023EDD82, 0x80001F8B023ED980,
-    0x80000038023EF586, 0x80000070023EF984, 0x8000002A023EA18A,
-    0x8000003C023EC188, 0x80000042023E9D8E, 0x80000048023EB58C,
-    0x8000005202373DB2, 0x80000058023725B0, 0x8000005E02361DB6,
-    0x8000001E023729B4, 0x80000660023571BA, 0x8000004E0003C9B8,
-    0x8000207F023525BE, 0x8000006C02351DBC, 0x80000341023735A2,
-    0x800003430003ADA0, 0x80000346023799A6, 0x8000208B0237B1A4,
-    0x8000209202376DAA, 0x800020A902377DA8, 0x800020B0023759AE,
-    0x000C7988023679AC, 0x000E098A000C7989, 0x000E198C000E098B,
-    0x000E098E000E198D, 0x0228E588000E098F, 0x0228858A0228DD89,
-    0x0228858C0228E18B, 0x0228858E0228E18D, 0x000E59980228CD8F,
-    0x0229199A000E5999, 0x0229319C0229259B, 0x80000C0A0229259D,
-    0x0229A998800000C8, 0x0229C99A022A3599, 0x0229C19C0229B59B,
-    0x0229B99E02298D9D, 0x000EA1A8022A6D9F, 0x000EA1AA000EA1A9,
-    0x000EA1AC000EA1AB, 0x000EA5AE000EADAD, 0x022DD9A8000EFDAF,
-    0x022DF1AA022D6DA9, 0x022DD1AC022D3DAB, 0x022DC9AE022D1DAD,
-    0x000EA5B8022D05AF, 0x022A31BA000EA9B9, 0x022BD1BC022B8DBB,
-    0x022BD1BE022B9DBD, 0x022845B8022B8DBF, 0x02287DBA022879B9,
-    0x0229FDBC0228B9BB, 0x0000A1BE000FADBD, 0x000F7988000FA5BF,
-    0x022A858A000FE989, 0x022A858C022A918B, 0x8000000E022AB18D,
-    0x8000064E80000651, 0x00020DB8800000CB, 0x8000034C8000209C,
-    0x800005AF80000574, 0x800005EC800005D5, 0x80000609000E4D99,
-    0x80000620022D219B, 0x8000033B022D259D, 0x80002106022DF99F,
-    0x80000641800000DA, 0x8000083C80000337, 0x8000066D80000669,
-    0x000E91A8800008A5, 0x000E91AA000E91A9, 0x000EBDAC000EA5AB,
-    0x000ED1AE000ED5AD, 0x80000689000ECDAF, 0x800008A98000068D,
-    0x80000844800007E7, 0x8000085D80000855, 0x80000EE5800020BE,
-    0x80000EFC0229B9B9, 0x80000F09022949BB, 0x80000F11022959BD,
-    0x80002102800006DC, 0x800004610229B1B1, 0x8000046D022989B3,
-    0x8000043C022999B5, 0x000135B8022989B7, 0x80000449000E69BA,
-    0x800005060233EDB8, 0x80000455000E61BE, 0x000FC0CC000E6DBC,
-    0x8000050C0233F5B2, 0x80000559023369B0, 0x8000059A0233C5B6,
-    0x8000045B0233C9B4, 0x800005BA023365AA, 0x80000467023369A8,
-    0x023898CC02337DAE, 0x8000055F02336DAC, 0x800005C0023D11A2,
-    0x800005F4023D61A0, 0x80000437023371A6, 0x8000071F02332DA4,
-    0x80000722023C799A, 0x8000072F023D0D98, 0x80000864023DC59E,
-    0x80000867023C659C, 0x8000086B023F5992, 0x80000732023CA990,
-    0x80000883023C9596, 0x80000886023C9994, 0x800020C2023E618A,
-    0x80000354023EB588, 0x800020D6023C058E, 0x800020DA000EB58C,
-    0x800020DE02387982, 0x800020E202386580, 0x800020EA02380586,
-    0x022A898A02386984, 0x022A4988022B698B, 0x000C698E022B6989,
-    0x0229798C000C958F, 0x0000E8CC000C6D8D, 0x000C4980800020E6,
-    0x0003D0CC000C4581, 0x0002B8CC800020F2, 0x0229C99A800020FA,
-    0x0229F998022B319B, 0x000CC19E0229DD99, 0x0002C8CC000CCD9F,
-    0x00005DB8000CD59D, 0x0220C0CC022C9593, 0x000005B802232CCD,
-    0x0002E4CC800020FE, 0x022875AA0000DDB8, 0x022FB9A8022B01AB,
-    0x022BDDAE022871A9, 0x022859AC000C3DAF, 0x8000210A000C3DAD,
-    0x0002A8CC022259A3, 0x8000210E0003BC94, 0x800021120002A492,
-    0x022EEDBA000BA090, 0x02284DB8022B49BB, 0x000C61BE022BB1B9,
-    0x022E01BC80002119, 0x000051B8023075BD, 0x8000213780002130,
-    0x800021410002F8CC, 0x000528CC000005B8, 0x0236F58080002145,
-    0x0236FD820236F181, 0x0236E5840236F983, 0x0236ED860236E185,
-    0x0236D5880236E987, 0x0236DD8A0236D189, 0x0007CDB8000C3494,
-    0x000298CC000D4092, 0x000009B8000D0090, 0x80002149000BFD91,
-    0x000005B80002FCCC, 0x000689898000214D, 0x0002E4CC000B9D97,
-    0x8000215900000DB8, 0x0006B9830002F8CC, 0x8000215500000DB8,
-    0x000005B800021CCC, 0x0002E4CC80002165, 0x000B59A400002DB8,
-    0x000B41A6000B45A5, 0x800021690002D8CC, 0x8000216D00000DB8,
-    0x000005B80002FCCC, 0x000298CC80002171, 0x000005B8023649AF,
-    0x0002FCCC80002175, 0x000B39B4000B2DB3, 0x000B29B6000035B8,
-    0x80002179000B25B7, 0x000005B80002ECCC, 0x0004ADBC8000217D,
-    0x0005E1BE0002E8CC, 0x0002E8CC0001F9B8, 0x000005B880002181,
-    0x000009B8800021A5, 0x800021A90002E4CC, 0x000005B80002DCCC,
-    0x0002E4CC80002185, 0x007DB091000019B8, 0x007B389380002189,
-    0x0002E8CC00780092, 0x000005B80001848C, 0x007E5C898000218D,
-    0x0002C0CC0001C888, 0x007F6085007E988A, 0x00049487007C0084,
-    0x00016C8100031086, 0x0000E9B800037480, 0x007000BD00007C82,
-    0x007188BF007684BC, 0x8000219100738CBE, 0x0001C0BB0002D0CC,
-    0x00016CB500019CBA, 0x800021AD00000DB8, 0x007930B100028CCC,
-    0x0077B8B3007AB4B0, 0x000005B8007400B2, 0x0002E4CC800021B1,
-    0x007D9CA9000011B8, 0x0001F8AB000D1CA8, 0x000009B880002195,
-    0x800021990002E4CC, 0x000005B80002FCCC, 0x0002E4CC8000219D,
-    0x00071D870003F9B8, 0x000690A5800021A1, 0x000658A30002F8CC,
-    0x0002E4CC00005DB8, 0x0005C5AF000051B8, 0x0005FDAD0002F4CC,
-    0x000065B8000595AC, 0x00071D8900006DB8, 0x00076D9700071D88,
-    0x800021B5000290CC, 0x007CFD9C800021B9, 0x800021C1800021BD,
-    0x0234959F0005B5B0, 0x000011AA000011A9, 0x00789DAF0006B0BC,
-    0x8000244780002443, 0x8000036700788D9F, 0x8000038000027C82,
-    0x0006908380000399, 0x800003B2007861B3, 0x0070359E00700999,
-    0x800003C500027CBE, 0x0006D88B800003D8, 0x800003FE800003EB,
-    0x000781B70006B488, 0x0007F1B50007F9B6, 0x000334BE0007F9B4,
-    0x000AD1B10003C482, 0x023B71BF0007EDB0, 0x00078DBD000789BE,
-    0x000791BB000785BC, 0x0007A9B90007ADBA, 0x000631800007B5B8,
-    0x023481820234D181, 0x0006258402348183, 0x0006258600062D85,
-    0x0006698800065D87, 0x0006718A00067589, 0x0006598C00064D8B,
-    0x0006558E0006518D, 0x023F759080000411, 0x023AB192023B5591,
-    0x0232B1940236FD93, 0x023B9D96022F6195, 0x023A5D98023BB197,
-    0x0237419A023AF599, 0x000B719C02372D9B, 0x800004240071ED99,
-    0x8000048A0002C8BE, 0x000304E3800000DC, 0x8000011B80000112,
-    0x000344E38000011F, 0x0006D1A80071059E, 0x000380EF80000123,
-    0x8000012700020C82, 0x0003E4EF8000012B, 0x000264E38000012F,
-    0x8000013300715999, 0x800000E00071519E, 0x800000E500717599,
-    0x800000EF800000EA, 0x00025882800000F4, 0x000274EF800000F9,
-    0x0071599E800000FE, 0x0079E18080000103, 0x8000010D80000108,
-    0x8000036480000116, 0x0244118602441D85, 0x0074298802440D87,
-    0x8000048702501989, 0x8000057980000526, 0x800005DA800005B4,
-    0x0079A990800005F1, 0x0079B1920079B591, 0x007A4194007A4593,
-    0x007A7196007A4D95, 0x007A5998007A6D97, 0x007A219A007A2599,
-    0x007A119C007A3D9B, 0x007A099E007A1D9D, 0x0244D5A0007A059F,
-    0x007A01A2007A05A1, 0x0249F5A4007A1DA3, 0x007A39A6007A0DA5,
-    0x007A11A8007A35A7, 0x007A11AA024915A9, 0x02445DAC007A6DAB,
-    0x0244B1AE024545AD, 0x0074C9B0024581AF, 0x007A41B2007A45B1,
-    0x0249B5B4007A5DB3, 0x007A79B6007A4DB5, 0x007A51B8007A75B7,
-    0x007A51BA024955B9, 0x02441DBC007A2DBB, 0x0244F1BE024505BD,
-    0x0079C8830245C1BF, 0x007C20B38000060E, 0x007C64B5007C30B2,
-    0x007E1085007C00B4, 0x8000036A80000625, 0x8000039C80000383,
-    0x800003B5007E5088, 0x007E688D800003C8, 0x800003EE800003DB,
-    0x8000041480000401, 0x80000427007FB4A2, 0x800001378000048D,
-    0x007FBCA98000013B, 0x007FA0AB007FB0A8, 0x800001438000013F,
-    0x8000014B80000147, 0x007F24918000014F, 0x007ED4A180000153,
-    0x80000157007F1092, 0x007E20A58000015B, 0x8000015F007ED0A4,
-    0x8000016780000163, 0x8000016F8000016B, 0x8000017780000173,
-    0x007E5CB38000017B, 0x007FD083007E68B2, 0x007E28B7007FE082,
-    0x007FB087007E5CB6, 0x007E24BB007FAC86, 0x8000017F007E30BA,
-    0x8000018780000183, 0x007FDCBD8000018B, 0x0249D5A0007FCCBC,
-    0x024985A202498DA1, 0x007B29A4007B39A3, 0x007B59A6007B29A5,
-    0x007B71A8007B55A7, 0x007CA1AA007CA5A9, 0x007CA9AC007CA1AB,
-    0x007C95AE007C9DAD, 0x007CF9B0007C89AF, 0x007D31B2007D79B1,
-    0x007DD5B4007DCDB3, 0x007DC9B6007DD9B5, 0x007DC5B8007DC9B7,
-    0x007DDDBA007DD5B9, 0x007DB9BC007DD1BB, 0x007DA1BE007DA5BD,
-    0x8000018F007DA1BF, 0x8000019780000193, 0x800021238000019B,
-    0x00003DB500003DB6, 0x8000063900003DB4, 0x8000064680000643,
-    0x800021DF800021D9, 0x800021E7800021E3, 0x800021EF800021EB,
-    0x800021F7800021F3, 0x800021FF800021FB, 0x8000220780002203,
-    0x8000220F8000220B, 0x8000221780002213, 0x8000221F8000221B,
-    0x8000222780002223, 0x024B29998000222B, 0x024B399B8000222F,
-    0x024B219D0244A99A, 0x80002233024B259C, 0x0247759180002237,
-    0x0247C59302477190, 0x0246A1950247C192, 0x8000223F8000223B,
-    0x007AE989007A9196, 0x0247718B80002243, 0x0247698D02477D8A,
-    0x024E098F02476D8C, 0x007A19810253698E, 0x007A1583007A1180,
-    0x007AD18502469D82, 0x024E9D8780002247, 0x007789B9007AD986,
-    0x007A55BB007A5DB8, 0x02528DBD8000224B, 0x02502DBF025021BC,
-    0x02478DB1025281BE, 0x024695B30247B5B0, 0x007A71B5007A11B2,
-    0x007A69B702464DB4, 0x8000224F007A75B6, 0x024FF1AB024E51A8,
-    0x024621AD024709AA, 0x0247C9AF024019AC, 0x007A55A180002253,
-    0x80002257007A55A0, 0x8000225B007A55A2, 0x8000225F024E61A4,
-    0x0240C580025B89A6, 0x024135820240E981, 0x0241E98402419D83,
-    0x0243BD86024C3D85, 0x024C6188024C2D87, 0x024CBD8A024C7D89,
-    0x024DBD8C024D498B, 0x024DF58E024D858D, 0x024475900249A98F,
-    0x0241B99202478991, 0x0240419402404593, 0x0240699602407995,
-    0x0070699802546997, 0x0240AD9A02545999, 0x02414D9C0240819B,
-    0x0241819E0241E59D, 0x024325A0024C559F, 0x024CC9A2024CB5A1,
-    0x024C05A400708DA3, 0x0254DDA6024DF1A5, 0x0255A5A8024D2DA7,
-    0x007D99AA024931A9, 0x8000226780002263, 0x8000226F8000226B,
-    0x8000227780002273, 0x8000227F8000227B, 0x8000228780002283,
-    0x8000228F8000228B, 0x8000229780002293, 0x8000229F8000229B,
-    0x80001CB380000B0E, 0x800022AB80000B41, 0x0007B180800022AF,
-    0x0007B1820007B181, 0x0007D1840007B183, 0x0007D1860007D185,
-    0x0007F1880007D187, 0x0007F18A0007F189, 0x0007D18C0007F18B,
-    0x0007D18E0007D18D, 0x0007B1900007D18F, 0x0007B1920007B191,
-    0x000655940007B193, 0x0006459600064D95, 0x0006599800066997,
-    0x00061D9A00062199, 0x0006B19C0006E19B, 0x0006919E0006A99D,
-    0x000679A000068D9F, 0x000139A2000129A1, 0x000129A4000139A3,
-    0x000159A6000129A5, 0x000169A8000159A7, 0x000179AA000169A9,
-    0x000169AC000179AB, 0x000159AE000169AD, 0x800022B3000159AF,
-    0x800022BF800022BB, 0x80000AAE80000AAB, 0x800022C7800022C3,
-    0x800022CF800022CB, 0x800022D7800022D3, 0x800022DF800022DB,
-    0x800022E7800022E3, 0x0002E9A0800022EB, 0x0002E5A2800022EF,
-    0x0002E5A40002FDA3, 0x800022F7800022F3, 0x800022FB0002E5A7,
-    0x800022FF0002D9A9, 0x800023030003EDAB, 0x00721DAE00744DAD,
-    0x0074A1B0007449AF, 0x00038DB280002307, 0x8000230F8000230B,
-    0x8000231300039DB5, 0x8000231B80002317, 0x800023238000231F,
-    0x02564DBC80002327, 0x0003B5BE024F11BD, 0x8000232B0003ADBF,
-    0x800023338000232F, 0x8000233B80002337, 0x800023438000233F,
-    0x8000234B80002347, 0x800023538000234F, 0x8000235B80002357,
-    0x800023638000235F, 0x8000236B80002367, 0x800023738000236F,
-    0x8000237B80002377, 0x800023838000237F, 0x8000238B80002387,
-    0x800023938000238F, 0x800032FB80002437, 0x80002BC780003B67,
-    0x0006058F80002BEB, 0x0007318D0007398E, 0x0007218B0007398C,
-    0x000731890007398A, 0x00072D8700070D88, 0x00072D8500072D86,
-    0x0078C5830078D184, 0x0078DD810078D982, 0x0006319F0078DD80,
-    0x0006159D0006299E, 0x0006059B0006159C, 0x000611990006059A,
-    0x0006219700061998, 0x00065D9500065996, 0x0006599300065D94,
-    0x0006599100065192, 0x00066DAF00066190, 0x000635AD00066DAE,
-    0x000625AB000635AC, 0x000635A9000625AA, 0x000605A7000635A8,
-    0x00063DA5000605A6, 0x000635A300063DA4, 0x0006C1A10006C9A2,
-    0x024375BF0006C9A0, 0x0243F1BD024391BE, 0x0243FDBB0243FDBC,
-    0x0243D1B90243C5BA, 0x02442DB702422DB8, 0x000565B50247D1B6,
-    0x000575B3000565B4, 0x00060DB1000575B2, 0x000E998000060DB0,
-    0x000E958280002BF7, 0x000EB18480002C07, 0x000EBD8680002C0F,
-    0x000E898880002C1B, 0x000E858A80002C23, 0x000E818C80002C2B,
-    0x000E8D8E80002C37, 0x000EF99080002C92, 0x000EF59280002CAA,
-    0x000E919480002CBA, 0x000E9D9680002CCA, 0x000EA99880002CDA,
-    0x000EA59A80002CEA, 0x000EA19C80002D0A, 0x000EAD9E80002D0E,
-    0x000E59A080002D12, 0x000E55A280002D46, 0x80002D7E80002D6E,
-    0x80002D8A80002D86, 0x80002DB280002D96, 0x80002DCE80002DB6,
-    0x80002DD2000E4DAB, 0x80002DDA000E51AD, 0x80002E8680002DEA,
-    0x000E21B280002E9A, 0x80002EEE80002EDA, 0x80002EFA80002EF6,
-    0x80002F2D80002F0A, 0x80002F6580002F59, 0x80002F9180002F85,
-    0x80002F9980002F95, 0x000FF98080002FB1, 0x00010D82000FF981,
-    0x0001058400010183, 0x00013D8600010985, 0x0001158800013187,
-    0x00016D8A00011989, 0x0001658C0001618B, 0x00015D8E0001698D,
-    0x000135900001518F, 0x0001CD9200013991, 0x0001C1940001C193,
-    0x0001C1960001C195, 0x0001C1980001C197, 0x0001C19A0001C199,
-    0x0001C99C0001DD9B, 0x0001B99E0001C99D, 0x00014DA00001B99F,
-    0x00015DA2000155A1, 0x00014DA4000155A3, 0x00017DA6000175A5,
-    0x00014DA8000175A7, 0x00015DAA000155A9, 0x00014DAC000155AB,
-    0x0003B1AE0003B1AD, 0x0003D1B00003B1AF, 0x0003D1B20003D1B1,
-    0x0003F1B40003D1B3, 0x0003F1B60003F1B5, 0x0003D1B80003F1B7,
-    0x0003D1BA0003D1B9, 0x0003B1BC0003D1BB, 0x0003B1BE0003B1BD,
-    0x0005E4820003B1BF, 0x80002FE180002FB5, 0x80002FF580002FED,
-    0x8000300D80002FFD, 0x800030248000301C, 0x8000305480003050,
-    0x800030608000305C, 0x8000307080003068, 0x8000309C80003080,
-    0x8000312880003114, 0x800031AC8000312C, 0x800031C4800031B0,
-    0x800031D0800031C8, 0x800031E0800031D8, 0x80003222800031EC,
-    0x8000324680003232, 0x800032CF800032C3, 0x800032EB800032DF,
-    0x800032FF800032F7, 0x8000330780003303, 0x8000330F8000330B,
-    0x8000340F800033C7, 0x8000341F8000341B, 0x8000342780003423,
-    0x800034338000342F, 0x8000346380003447, 0x800034AB8000346B,
-    0x800034B7800034AF, 0x800034C3800034BB, 0x800034EB800034C7,
-    0x8000350B800034EF, 0x800035138000350F, 0x007680BE80003517,
-    0x8000352B000D61AE, 0x8000354F000D6DAC, 0x80003553000D49AA,
-    0x80003557000D45A8, 0x8000357B000D75A6, 0x800035BB000D61A4,
-    0x800035BF000D71A2, 0x800035DF000D0DA0, 0x800035EF000DD1BE,
-    0x800035F7000DDDBC, 0x80003617000DD9BA, 0x8000363B000DD5B8,
-    0x8000367F000DE1B6, 0x80003687000DEDB4, 0x8000369B000D19B2,
-    0x800036A7000D15B0, 0x800036AF000E758E, 0x800036B7000E618C,
-    0x800036BB000E7D8A, 0x800036BF000E7988, 0x800036D7000E4586,
-    0x800036DB000E7184, 0x800036F7000E6D82, 0x800036FF000E6980,
-    0x80003707000DED9E, 0x8000370B000DED9C, 0x80003717000DCD9A,
-    0x8000371F000DC998, 0x80003723000DE196, 0x80003727000E1194,
-    0x8000372F000E0D92, 0x0007A980000E0990, 0x0007B1820007B581,
-    0x0007D9840007CD83, 0x0007C1860007C585, 0x0007C9880007FD87,
-    0x0007D18A0007D589, 0x0001358C00012D8B, 0x0001358E00013D8D,
-    0x0001559000012D8F, 0x0001559200015D91, 0x0001759400016D93,
-    0x0001759600017D95, 0x0001559800016D97, 0x0001559A00015D99,
-    0x0001359C00012D9B, 0x0001359E00013D9D, 0x0001D5A000012D9F,
-    0x0001D5A20001DDA1, 0x0001F5A40001EDA3, 0x0001F5A60001FDA5,
-    0x0001D5A80001EDA7, 0x0001D5AA0001DDA9, 0x000139AC000129AB,
-    0x000129AE000139AD, 0x000159B0000129AF, 0x000169B2000159B1,
-    0x000179B4000169B3, 0x000169B6000179B5, 0x000159B8000169B7,
-    0x000129BA000159B9, 0x000139BC000129BB, 0x000129BE000139BD,
-    0x80003733000129BF, 0x800037E7800037D3, 0x8000381F8000381B,
-    0x8000385380003827, 0x800038638000385B, 0x800038A78000386B,
-    0x800038B3800038AB, 0x800038BB800038B7, 0x800038EB800038E7,
-    0x80003907800038F3, 0x800039278000390B, 0x800039338000392F,
-    0x8000397780003957, 0x8000398F8000398B, 0x800039C380003997,
-    0x800039D7800039C7, 0x00032599800039FB, 0x80003A1780003A07,
-    0x80003A3780003A2F, 0x80003A3F80003A3B, 0x80003A4F80003A43,
-    0x80003A5B80003A53, 0x80003A7B80003A77, 0x80003A9780003A7F,
-    0x80003AA380003A9B, 0x80003ABB80003AB7, 0x80003AC780003ABF,
-    0x80003ACF80003ACB, 0x80003AD780003AD3, 0x80003AFF80003AE7,
-    0x80003B0F80003B03, 0x80003B1780003B13, 0x0003D98080003B1B,
-    0x0003E9820003D981, 0x0003F9840003E983, 0x0003E9860003F985,
-    0x0003D9880003E987, 0x0000718A0003D989, 0x0000518C0000718B,
-    0x0000518E0000518D, 0x000031900000518F, 0x0000319200003191,
-    0x0000D19400003193, 0x80003B230000D195, 0x80003B2B80003B27,
-    0x80003B3F80003B37, 0x80003B4B80003B47, 0x80003B5780003B53,
-    0x80003B6380003B5B, 0x8000246B80003B6B, 0x80002D9E80002D9A,
-    0x800024E180002553, 0x800024E9800024E5, 0x00027CBE800024ED,
-    0x000260820000099E, 0x800024F100000D99, 0x001C0080000218BE,
-    0x001FA082001CEC81, 0x001A54840018BC83, 0x00159886001BD485,
-    0x001000880000659E, 0x0013E88A0011E489, 0x002F148C002C008B,
-    0x0024008E0029F48D, 0x0005D1A00025FC8F, 0x0005D1A20005D1A1,
-    0x0005F1A40005D1A3, 0x0005F1A60005F1A5, 0x0005D1A80005F1A7,
-    0x0005D1AA0005D1A9, 0x000439AC0005D1AB, 0x000421AE000425AD,
-    0x000469B000041DAF, 0x000471B2000475B1, 0x000419B400040DB3,
-    0x000401B6000405B5, 0x000409B800043DB7, 0x000411BA000415B9,
-    0x0004F9BC0004EDBB, 0x0004E1BE0004E5BD, 0x0005D1800004DDBF,
-    0x0005D1820005D181, 0x0005F1840005D183, 0x0005F1860005F185,
-    0x0005D1880005F187, 0x0005D18A0005D189, 0x0005318C0005D18B,
-    0x0005318E0005318D, 0x000551900005318F, 0x0005519200055191,
-    0x0005719400055193, 0x0005719600057195, 0x0005519800057197,
-    0x0005519A00055199, 0x0005319C0005519B, 0x0005319E0005319D,
-    0x000274820005319F, 0x006DE8E3006C108B, 0x00003D99001804E3,
-    0x001E91B60003908F, 0x0018BDB8006C508D, 0x001EB9BA001EA9B9,
-    0x006C3486800024F5, 0x0019C0EF000214BE, 0x001CF8E3001C6CEF,
-    0x000240E300031C9B, 0x000230E300035899, 0x0000759E0002D09F,
-    0x000208EF0002D09D, 0x00028C820003A893, 0x001C34EF0003EC91,
-    0x0002B4EF0003A897, 0x000005990003B895, 0x00028CA8800024F9,
-    0x000264BE0000099E, 0x00020CE300028CAF, 0x0000059900025882,
-    0x0002E4EF800024FD, 0x000264A6000264A1, 0x000264A4000200BE,
-    0x025A358080002501, 0x000140B800010CBB, 0x00026C820000699E,
-    0x0000659900020CE3, 0x0002A4B200024CBE, 0x0000559E0002C0EF,
-    0x0002788200001999, 0x000384BE80002505, 0x0003F8E30002A0B5,
-    0x00020CE30001F19E, 0x0000059900027882, 0x00027CBE80002509,
-    0x000260820000099E, 0x8000250D00000D99, 0x000330EF000260BE,
-    0x0002D8EF00001D9E, 0x0000099900021882, 0x8000251100026CE3,
-    0x000264E3000260BE, 0x00020CE300001D9E, 0x0002D4EF00025C82,
-    0x0002E0EF00000999, 0x000288BE80002515, 0x000260820002A0EF,
-    0x0000ED9E0000F999, 0x0002F8E380002519, 0x000264BE0000099E,
-    0x00025082000204E3, 0x0000099A8000251D, 0x800025210002ECEF,
-    0x000208BE00003199, 0x0002D0EF00020482, 0x00005D9E80002525,
-    0x000274E300006D9F, 0x00022CE380002529, 0x0000159A00006999,
-    0x000118EF8000252D, 0x0000259F000270BE, 0x0003FD9A0000519E,
-    0x0002E4EF00026C82, 0x000268BE00002199, 0x00024C82000204E3,
-    0x0003CD9E80002531, 0x000258BE0000199F, 0x00003D9980002535,
-    0x00020C8200023CBE, 0x8000253900001D9A, 0x0000259E0002D4EF,
-    0x0000099900000D9F, 0x80002544000264E3, 0x8000254800001D9A,
-    0x800000C1800000BC, 0x0000299E00023CE3, 0x00027C820000359F,
-    0x00000999000214EF, 0x8000255F0002A8EF, 0x000284BE0000059E,
-    0x00026082000204E3, 0x0002508200001199, 0x8000254C00024CBE,
-    0x00003D9E80002A9B, 0x8000263600000D99, 0x0000199E000264BE,
-    0x000218820002D8EF, 0x000258E300006D9E, 0x000278E300001599,
-    0x80002699000208BE, 0x00026082000204E3, 0x800026DC00003D99,
-    0x00027CBE0002F4EF, 0x0002A4EF00000D9E, 0x0002CCEF00038082,
-    0x0003FCE380002708, 0x000254BE0001519B, 0x0001ED9900025C82,
-    0x0001419C00018D9F, 0x800027320002CC9D, 0x000204820001D99E,
-    0x0002C0EF00026CBE, 0x0000199E00001599, 0x000244E380002757,
-    0x000075990000799E, 0x0002DCEF000270BE, 0x000278E300003999,
-    0x000234E300026C82, 0x0002DCEF0000099E, 0x0002A4EF8000276C,
-    0x00026082000268BE, 0x0002F0E30000DD99, 0x00022C820003D8BB,
-    0x0000319E000380B9, 0x000270BE000418BF, 0x800027740003F8BD,
-    0x80002790000368B3, 0x0002D0EF000320B1, 0x0000119E000358B7,
-    0x00006999000368B5, 0x00002599000288AB, 0x000250E300020CBE,
-    0x0002D4EF00028CAF, 0x0002D4EF000298AD, 0x000204E300024882,
-    0x0002F8A680002798, 0x00058CBE000400BE, 0x0005F4E300059C82,
-    0x800027A000000D99, 0x0002D8980000059E, 0x000204E300025082,
-    0x00000D990002C49B, 0x0002E4EF0002AC95, 0x000530EF800027B2,
-    0x0002B49200020CBE, 0x00006D9E00021C82, 0x00027CE300001199,
-    0x000348880003008F, 0x00027CBE800027BE, 0x0003A4840000099E,
-    0x0003748600026C82, 0x800027C600001D99, 0x0002D8EF0003E081,
-    0x0000F19E000284BE, 0x00027482000108BD, 0x0002FCE30002A5BF,
-    0x00000599000214E3, 0x000224EF800027DC, 0x00027CBE0002F4EF,
-    0x0002FCB00000319E, 0x0002E4B20002FCB1, 0x000274E300021082,
-    0x000019990000119A, 0x8000281A0002D8AF, 0x00000D9F8000282B,
-    0x000244BE00000D9E, 0x8000284D00027482, 0x0002E8EF80002858,
-    0x0000159A00001D99, 0x0003ECE300038CBE, 0x0001E19E000214E3,
-    0x000005990001E19F, 0x00026C82800028A8, 0x800028BC00003D9A,
-    0x00003D9E000274BE, 0x000334EF00000D9F, 0x00021C820002E8EF,
-    0x00007D99800028DE, 0x800028EC000274E3, 0x000270BE0000059A,
-    0x0000659E00023CE3, 0x000274820000159F, 0x800029340002E4EF,
-    0x0000199980002942, 0x000280BE0002DCEF, 0x0000E99E0000119A,
-    0x00001D990000E99F, 0x00027C820006619E, 0x000200E380002A9F,
-    0x0000359E000244BE, 0x0002CCEF00026C82, 0x80002AA300001D99,
-    0x000274BE0000059E, 0x00026CE300020C82, 0x80002AA7000270E3,
-    0x000204E300026CBE, 0x0000119E00001D99, 0x0002E0EF00025482,
-    0x00000D9E80002AAB, 0x0002B0EF00000999, 0x000268BE0002C0EF,
-    0x00020CE380002AB7, 0x0003F59900026882, 0x0003ED9E000264BE,
-    0x0000008600000085, 0x0000008800000087, 0x8000271700000089,
-    0x0002C4EF80001ABF, 0x0000008E0000008D, 0x000C01B180001AC3,
-    0x000065B300000091, 0x00006DB5000E4DB2, 0x000E49B7000E51B4,
-    0x000E21B90000C9B6, 0x0000F5BB000E71B8, 0x0000C1BD0000F5BA,
-    0x0000C5BF0000DDBC, 0x000000A00000C5BE, 0x000000A2000000A1,
-    0x000000A4000000A3, 0x80001ACB80001AC7, 0x000000A880001ACF,
-    0x80001AD3000000A9, 0x80001AD7000000AB, 0x80001ADB000000AD,
-    0x800019EB80001ADF, 0x000000B2800019FF, 0x80001A4B000000B3,
-    0x000000B6000000B5, 0x80001A5380001A4F, 0x80001A5B80001A57,
-    0x80001A5F000000BB, 0x000000BE80001A63, 0x0001C180000000BF,
-    0x000F5982000F4181, 0x0001D584000F5983, 0x000F9D86000F9585,
-    0x000FB588000F8D87, 0x000FED8A000FE589, 0x000FE58C000FFD8B,
-    0x000E3D8E000E358D, 0x00019D900001ED8F, 0x0001AD920001A591,
-    0x0001A1940001A593, 0x000F51960001A195, 0x000F7998000F5197,
-    0x000F499A000F7999, 0x000F599C000F499B, 0x000F499E000F599D,
-    0x000FB9A0000F499F, 0x000FC1A2000FB9A1, 0x000FC9A4000FC1A3,
-    0x000FF9A6000FC9A5, 0x000FC9A8000FF9A7, 0x000FD9AA000FC9A9,
-    0x000FC9AC000FD9AB, 0x000F31AE000FC9AD, 0x000F59B0000F31AF,
-    0x000F69B2000F59B1, 0x000F79B4000F69B3, 0x000F69B6000F79B5,
-    0x000F59B8000F69B7, 0x000F21BA000F59B9, 0x000F29BC000F21BB,
-    0x000F19BE000F29BD, 0x8000199B000F19BF, 0x80001A678000199F,
-    0x800019A380001A6B, 0x80001A6F800019A7, 0x80001A7780001A73,
-    0x80001A7F80001A7B, 0x80001A8780001A83, 0x80001A8F80001A8B,
-    0x800019AB80001A93, 0x80001A97800019AF, 0x800019B380001A9B,
-    0x80001A47800019B7, 0x800019E7800019E3, 0x80001AE780001AE3,
-    0x80001AEF80001AEB, 0x80001AF780001AF3, 0x800019BB80001AFB,
-    0x800019EF800019BF, 0x800019C380001AFF, 0x80001B03800019C7,
-    0x800019F7800019F3, 0x80001A03800019FB, 0x80001A0B80001A07,
-    0x80001A1380001A0F, 0x800019CB80001A17, 0x80001A1B800019CF,
-    0x800019D380001A1F, 0x80001A23800019D7, 0x80001A2B80001A27,
-    0x80001A3380001A2F, 0x80001B0780001A37, 0x80001A3B80001B0B,
-    0x800019DB80001A3F, 0x001FF192800019DF, 0x0000B19400112993,
-    0x0000B1960000B195, 0x000091980000B197, 0x00113D9A00009199,
-    0x00008D9C0000959B, 0x0000F19E0000F59D, 0x0001A9800011099F,
-    0x0001B9820001A981, 0x0001A9840001B983, 0x000025860001A985,
-    0x0000E5880000DD87, 0x0000E58A0000ED89, 0x0000E58C0000FD8B,
-    0x0000E58E0000ED8D, 0x80001A9F80001A43, 0x80001AA780001AA3,
-    0x80001AAF80001AAB, 0x80001AB780001AB3, 0x80002BD380001ABB,
-    0x80002BD780002E8A, 0x80002BDB80002BEF, 0x80002BE3800034CB,
-    0x80002F1580002BCB, 0x80002E9E027DACE1, 0x8000019F027D0CE3,
-    0x800001AB800001A5, 0x800001B7800001B1, 0x800001C3800001BD,
-    0x800001CF800001C9, 0x800001DB800001D5, 0x800001E1027DF0EF,
-    0x800001ED800001E7, 0x800002D1800002CB, 0x800002DD800002D7,
-    0x800002E9800002E3, 0x800002F5800002EF, 0x800003198000030D,
-    0x800003258000031F, 0x800019978000032B, 0x8000031380000331,
-    0x80000304800002FB, 0x8000020B800001F3, 0x80000253800001FF,
-    0x8000022F80000211, 0x80000229800001F9, 0x8000023B80000205,
-    0x004B44E18000026B, 0x0077B4E303AF1CE1, 0x80000289004B14E3,
-    0x8000027780000283, 0x80000259800002C5, 0x8000027D80000265,
-    0x8000029B80000271, 0x007768EF80000247, 0x8000028F004BE4EF,
-    0x800002A1800002B9, 0x8000021780000235, 0x8000025F8000024D,
-    0x8000021D80000295, 0x80000241800002BF, 0x80000223800002A7,
-    0x800002B3800002AD, 0x036A50E180002E3E, 0x031F5CE30344D8E1,
-    0x030F34E10376C8E3, 0x032EC4E3030BD0E1, 0x80003020030AC4E3,
-    0x033FF8EF800035FF, 0x800008190302E8EF, 0x030DE0EF8000049D,
-    0x03C684E1037368EF, 0x03EB1CE303DF98E1, 0x038554E103C398E3,
-    0x03BEECE303989CE1, 0x800004AC039E80E3, 0x03B63CEF800004BB,
-    0x800004CA038274EF, 0x03F7A4EF800004D9, 0x02174CE103D628EF,
-    0x020A4CE3026730E1, 0x023FCCE1026638E3, 0x02DB44E3020348E1,
-    0x800004E2023A38E3, 0x02FA6CEF800004EB, 0x800004F4022EA4EF,
-    0x021F78EF800004FD, 0x02E704E1027E6CEF, 0x029A1CE302D75CE1,
-    0x035828E102F278E3, 0x03440CE3029F80E1, 0x8000052902AE68E3,
-    0x034338EF80000532, 0x8000053B0358F8EF, 0x02EE24EF8000053E,
-    0x0006CDB802F72CEF, 0x0006DDBA0006D5B9, 0x0006CDBC0006D5BB,
-    0x0006FDBE0006F5BD, 0x000685B00006F5BF, 0x000685B200068DB1,
-    0x000685B400069DB3, 0x0006FDB60006F5B5, 0x000635A80006F5B7,
-    0x000635AA00063DA9, 0x0006C5AC0006DDAB, 0x0006C5AE0006CDAD,
-    0x000635A00006FDAF, 0x000635A200063DA1, 0x000615A400060DA3,
-    0x000615A600061DA5, 0x0006B59800060DA7, 0x0006B59A0006BD99,
-    0x0006D99C0006C99B, 0x0006C99E0006D99D, 0x0006B59080000541,
-    0x0006B5920006BD91, 0x0006959400068D93, 0x0006959600069D95,
-    0x0005318800068D97, 0x0005318A00053189, 0x0006D58C0006CD8B,
-    0x0006D58E0006DD8D, 0x000531800006CD8F, 0x0005318200053181,
-    0x0005118400053183, 0x0005118600051185, 0x0007ED8000051187,
-    0x0007FD820007F581, 0x0007ED840007F583, 0x0017418600063585,
-    0x8000054400063987, 0x80004F9F80003D3F, 0x80006B2F800058CF,
-    0x80007D8F8000745F, 0x0006719080008FEF, 0x0006299200067191,
-    0x0006219400062993, 0x0006F99600062195, 0x0006D1980006F997,
-    0x0006899A0006D199, 0x0006819C0006899B, 0x0001E59E0006819D,
-    0x0007C1A000005D9F, 0x000599A2000581A1, 0x0005D5A40005CDA3,
-    0x000519A60005C5A5, 0x000535A800050DA7, 0x000579AA000561A9,
-    0x0004B1AC000579AB, 0x00009DAE0004B1AD, 0x0000EDB0000095AF,
-    0x0000FDB20000F5B1, 0x0000EDB40000F5B3, 0x0003D9B6000029B5,
-    0x0003E9B80003D9B7, 0x0003F9BA0003E9B9, 0x0003E9BC0003F9BB,
-    0x0003D9BE0003E9BD, 0x8000AB7F8000A24F, 0x8000C71D8000BDE6,
-    0x8000D97D8000D04D, 0x8000BE2A8000E2AD, 0x8000A7FF8000B133,
-    0x80002CBE80002C27, 0x80002C1380002BCF, 0x8000323680003276,
-    0x80002E0A80003583, 0x800038BF8000343B, 0x80002D5280003597,
-    0x800034D3800035C7, 0x80002C8E80003947, 0x8000332F80002DBA,
-    0x80002C5280003A5F, 0x800032D380002CE2, 0x80002DFE80002FF1,
-    0x80002F6D80002D8E, 0x8000352380002F61, 0x800038D380002C4E,
-    0x80002F0280002DAA, 0x8000054A80000547, 0x800005508000054D,
-    0x8000057F8000057C, 0x8000058580000582, 0x8000058B80000588,
-    0x800005918000058E, 0x8000059780000594, 0x000000B8800005B7,
-    0x8000051780000478, 0x800005A58000056A, 0x800005E2800005CB,
-    0x0019F980800005FF, 0x0006A98200172181, 0x000135840006A983,
-    0x0001358600013D85, 0x0001118800175187, 0x00155D8A0017B989,
-    0x0015918C0015898B, 0x0012558E00154D8D, 0x00152D900017F98F,
-    0x0001719200017191, 0x0001119400017193, 0x0001119600011195,
-    0x0001319800011197, 0x0001319A00013199, 0x00012D9C00178D9B,
-    0x00011D9E0001159D, 0x0001EDA00001159F, 0x0001FDA20001F5A1,
-    0x000639A40001F5A3, 0x000629A6000639A5, 0x000121A80001E1A7,
-    0x000121AA000121A9, 0x000121AC000121AB, 0x000121AE000121AD,
-    0x000161B0000121AF, 0x000161B2000161B1, 0x000161B4000161B3,
-    0x000161B6000161B5, 0x000121B8000161B7, 0x000121BA000121B9,
-    0x000121BC000121BB, 0x000121BE000121BD, 0x0001D1BB000121BF,
-    0x0001C1B90001C9BA, 0x0001F1BF0001B1B8, 0x0001C1BD0001C9BE,
-    0x000189B30001C9BC, 0x000189B1000181B2, 0x000189B7000151B0,
-    0x000189B5000181B6, 0x0001DDAB000191B4, 0x0001DDA90001DDAA,
-    0x000129AF00013DA8, 0x00019DAD000121AE, 0x0001F1A300019DAC,
-    0x0001E1A10001E9A2, 0x00011DA70001E9A0, 0x00011DA500011DA6,
-    0x0001D19B00011DA4, 0x0001C1990001C99A, 0x0001119F0001C998,
-    0x0001C19D0001C99E, 0x000191930001C99C, 0x0001819100018992,
-    0x0001F19700018990, 0x0001819500018996, 0x0001D18B00018994,
-    0x0001C1890001C98A, 0x0001F18F0001C988, 0x0001C18D0001C98E,
-    0x003E11830001C98C, 0x003E0181003E0982, 0x0001F187003E0980,
-    0x003E0185003E0986, 0x80000616003E0984, 0x800003748000062D,
-    0x800003A68000038D, 0x800006D38000070D, 0x8000077C800006D7,
-    0x800025A080002567, 0x800025D1800025BF, 0x8000260E800025F6,
-    0x800026BB8000264E, 0x8000270C800026FA, 0x8000275B80002736,
-    0x8000277880002770, 0x800027E780002794, 0x800027F6800027F2,
-    0x800027FE800027FA, 0x8000299C8000295D, 0x8000279C800029C4,
-    0x800027BA800027AE, 0x800027D1800027C2, 0x800028498000280C,
-    0x800029C88000287D, 0x800029F5800029ED, 0x80002A2380002A11,
-    0x80002A2B80002A27, 0x800028C080002A43, 0x80002A5E80002915,
-    0x80002A7D80002A79, 0x80002A8F80002A8B, 0x80002C4780002A93,
-    0x8000256B0004A0F0, 0x8000258580002578, 0x800025A480002592,
-    0x00346CE3800025B1, 0x003414E3003430E3, 0x003500E30033F8E3,
-    0x0034ACE30034F8E3, 0x0035FCEF00349CE3, 0x003430EF003420EF,
-    0x0034B4EF00347CEF, 0x0034F4EF003494EF, 0x800025C300333CEF,
-    0x003520E3800025D5, 0x003584E300357CE3, 0x0035F4E30035E0E3,
-    0x003560EF0035C4E3, 0x003528EF003550EF, 0x800025E5003508EF,
-    0x0035F8EF800025FA, 0x0032B4EF0035D4EF, 0x800026040032D0EF,
-    0x0032FCEF80002612, 0x800026298000261C, 0x003220E38000263A,
-    0x8000269D80002644, 0x003200E3800026A4, 0x8000265200322CE3,
-    0x0033E0E3800026AE, 0x003234E30033C0E3, 0x00333CE300321CE3,
-    0x0033A8E3003354E3, 0x0033ECEF003388E3, 0x003388EF0033D0EF,
-    0x003344EF003360EF, 0x003308EF003320EF, 0x000791B90032ECEF,
-    0x000791BB000789B8, 0x0007B1BD000799BA, 0x0007B1BF000789BC,
-    0x000751B10007B9BE, 0x000751B3000749B0, 0x0007B1B5000759B2,
-    0x0007B1B7000749B4, 0x000711A90007B9B6, 0x000711AB000709A8,
-    0x000731AD000719AA, 0x000731AF000709AC, 0x000751A1000739AE,
-    0x000751A3000749A0, 0x000731A5000759A2, 0x000731A7000749A4,
-    0x00079199000739A6, 0x0007919B00078998, 0x0007B19D0007999A,
-    0x0007B19F0007899C, 0x00010D910007B99E, 0x0001159300010D90,
-    0x0007B19500011592, 0x0007B19700040194, 0x0001B5890007B996,
-    0x0001458B0001B588, 0x0001498D0001458A, 0x00017D8F0001418C,
-    0x00020D810001418E, 0x00021D8300020D80, 0x00020D8500021D82,
-    0x0001BD8700020D84, 0x800026660001BD86, 0x8000268980002676,
-    0x800026EA800026E0, 0x800026CF800026BF, 0x8000271E800026FE,
-    0x8000273A80002728, 0x8000275F80002747, 0x800027868000277C,
-    0x800027CA800027A4, 0x800027D5800027E0, 0x80002802800027EB,
-    0x8000282F80002810, 0x8000281E8000283F, 0x8000286C8000285C,
-    0x8000285180002876, 0x8000289180002881, 0x8000289E800028AC,
-    0x800028F0800028C4, 0x800028D4800028F7, 0x8000290B80002901,
-    0x80002946800028E2, 0x8000291980002938, 0x8000292080002953,
-    0x800029618000292A, 0x800029788000296E, 0x8000298C80002982,
-    0x800029AD800029A0, 0x800029CC800029B4, 0x800029E0800029D3,
-    0x80002A03800029F9, 0x80002A2F80002A15, 0x80002A4780002A3C,
-    0x0001AD8080002A51, 0x0001BD820001B581, 0x0001AD840001B583,
-    0x0001DD860001D585, 0x0001ED880001D587, 0x0001FD8A0001F589,
-    0x0001ED8C0001F58B, 0x0001DD8E0001D58D, 0x0001AD900001D58F,
-    0x0001BD920001B591, 0x0001AD940001B593, 0x00035D9600035595,
-    0x00036D9800035597, 0x00037D9A00037599, 0x00036D9C0003759B,
-    0x00035D9E0003559D, 0x0003ADA00003559F, 0x0003BDA20003B5A1,
-    0x0003ADA40003B5A3, 0x0003DDA60003D5A5, 0x0003EDA80003D5A7,
-    0x0003FDAA0003F5A9, 0x0003EDAC0003F5AB, 0x0003DDAE0003D5AD,
-    0x0003ADB00003D5AF, 0x0003BDB20003B5B1, 0x0003ADB40003B5B3,
-    0x00035DB6000355B5, 0x00036DB8000355B7, 0x00037DBA000375B9,
-    0x00036DBC000375BB, 0x00035DBE000355BD, 0x80002A62000355BF,
-    0x80002A8180002A69, 0x8000047D8000035A, 0x8000056F8000051C,
-    0x800005D0800005AA, 0x80000604800005E7, 0x800006328000061B,
-    0x800003938000037A, 0x800003BF800003AC, 0x800003E5800003D2,
-    0x8000040B800003F8, 0x800004318000041E, 0x800004A680000497,
-    0x800004C4800004B5, 0x80000713800004D3, 0x80000671800006B3,
-    0x800007F580000680, 0x800006BF80000802, 0x800006C6800006C2,
-    0x800030118000072C, 0x80002F0E800031FF, 0x8000327A800031F0,
-    0x800007DB800007FF, 0x8000078F80000C16, 0x8000073C80000739,
-    0x800006FA80000792, 0x8000073F80000691, 0x800007DE80000805,
-    0x80000C1E80000C1A, 0x800007448000079B, 0x800007478000071A,
-    0x800006FD8000079E, 0x000FF09180000859, 0x800007A680000C22,
-    0x8000074E800006BC, 0x800007E4800006F1, 0x800007A980000C26,
-    0x800007518000069B, 0x8000069E800007AC, 0x8000075480000789,
-    0x800006A2800007B0, 0x800007588000078C, 0x800007D2800007CC,
-    0x800007B88000075C, 0x8000082780000701, 0x800008338000082B,
-    0x800007EA80000816, 0x800007BF80000C2A, 0x800007ED8000081D,
-    0x800007C280000C2E, 0x8000082080000763, 0x80000C32800007F0,
-    0x80000766800007C5, 0x800007C880000769, 0x8000068480000664,
-    0x800006958000265F, 0x800006AA80000698, 0x800006B6800006A6,
-    0x8000070A80000705, 0x8000072980000710, 0x800007608000074B,
-    0x8000077580000772, 0x8000078080000778, 0x800007B4800007A2,
-    0x000FA1BE80000808, 0x000F91BC000F99BF, 0x000F81BA000F91BD,
-    0x000F95B8000F81BB, 0x000FA5B6000F95B9, 0x000F55B4000FADB7,
-    0x000F45B2000FADB5, 0x000F55B0000F4DB3, 0x000F25AE000F4DB1,
-    0x000F15AC000F2DAF, 0x000F05AA000F2DAD, 0x000F09A8000F0DAB,
-    0x000F29A6000EF1A9, 0x000F29A4000F29A7, 0x000F49A2000F29A5,
-    0x000F49A0000F49A3, 0x000FA99E000F49A1, 0x000FA99C000FA99F,
-    0x000F899A000FA99D, 0x000F8998000F899B, 0x000FA996000F8999,
-    0x000E5594000FA997, 0x000E4592000FA995, 0x000E5590000E4D93,
-    0x000E258E000E4D91, 0x000E158C000E2D8F, 0x000E058A000E2D8D,
-    0x000E1588000E0D8B, 0x000E2586000E0D89, 0x000E5584000E2D87,
-    0x000E4582000E2D85, 0x000E5580000E4D83, 0x000EA580000E4D81,
-    0x000EA582000EAD81, 0x000EA984000EBD83, 0x000ED986000EA985,
-    0x000EE988000ED987, 0x000EF98A000EE989, 0x000EE98C000EF98B,
-    0x000E118E000EE98D, 0x000EA590000EDD8F, 0x000EA592000EAD91,
-    0x000F9594000EBD93, 0x000EA196000EA195, 0x00036198000EA197,
-    0x0003619A00036199, 0x0003619C0003619B, 0x0003619E0003619D,
-    0x0003A1A00003619F, 0x0003A1A20003A1A1, 0x0003A1A40003A1A3,
-    0x0003A1A60003A1A5, 0x0003E1A80003A1A7, 0x0003E1AA0003E1A9,
-    0x0003E1AC0003E1AB, 0x0003E1AE0003E1AD, 0x0003A1B00003E1AF,
-    0x0003A1B20003A1B1, 0x0003A1B40003A1B3, 0x0003A1B60003A1B5,
-    0x000361B80003A1B7, 0x000361BA000361B9, 0x000361BC000361BB,
-    0x000361BE000361BD, 0x800007FA000361BF, 0x800008138000080B,
-    0x8000085080000847, 0x8000087080000878, 0x8000047380000678,
-    0x8000056580000512, 0x800005C6800005A0, 0x800005FA800005DD,
-    0x8000062880000611, 0x800003878000036E, 0x800003B9800003A0,
-    0x800003DF800003CC, 0x80000405800003F2, 0x8000042B80000418,
-    0x800004A080000491, 0x800004BE800004AF, 0x800004DC800004CD,
-    0x800004EE800004E5, 0x80000500800004F7, 0x800005358000052C,
-    0x80003B6F800006F6, 0x80003B7780003B73, 0x80003B7F80003B7B,
-    0x80003B8B80003B87, 0x80003B9380003B8F, 0x80003B9B80003B97,
-    0x80003BA380003B9F, 0x80003BAB80003BA7, 0x80003BB380003BAF,
-    0x80003BBB80003BB7, 0x80003BC380003BBF, 0x0018508280003BC7,
-    0x0004009980003BCB, 0x80003BD380003BCF, 0x0004149D00050C9A,
-    0x0006189F0002089C, 0x80003BD700071C9E, 0x80003BDF80003BDB,
-    0x80003BE780003BE3, 0x80003BEF80003BEB, 0x80003BF780003BF3,
-    0x80003BFF80003BFB, 0x80003C0780003C03, 0x80003C0F80003C0B,
-    0x80003C1780003C13, 0x80003C1F80003C1B, 0x80003C2780003C23,
-    0x80003C2F80003C2B, 0x80003C3780003C33, 0x80003C3B0000D1A6,
-    0x80003C3F0000DDA4, 0x80003C430000D9A2, 0x80003C4B80003C47,
-    0x80003C4F0000C1AE, 0x80003C530000CDAC, 0x80003C570000E9AA,
-    0x80003C5B0000E5A8, 0x80003C5F0000B5B6, 0x000698AD0000A1B4,
-    0x80003C630000BDB2, 0x80003C67001800AE, 0x80003C6B000065BE,
-    0x80003C6F000091BC, 0x80003C7300008DBA, 0x0006FD80000089B8,
-    0x00070D8280003C77, 0x0007198480003C7B, 0x0007398680003C7F,
-    0x0007058880003C83, 0x0007018A80003C87, 0x00071D8C80003C8B,
-    0x0007098E80003C8F, 0x0007759080003C93, 0x0007719280003C97,
-    0x00076D9480003C9B, 0x0007199680003C9F, 0x0007259880001CAB,
-    0x0007219A80003CA3, 0x0272519C80003CA7, 0x80003CAB0272599D,
-    0x80003CB380003CAF, 0x80003CBF80003CB7, 0x80003CCB80003CC3,
-    0x80003CD380003CCF, 0x80003CDB80003CD7, 0x80003CE380003CDF,
-    0x80003CEB80003CE7, 0x80000B1180003CEF, 0x80000B5680000B8C,
-    0x80003CF380003D33, 0x80003CFB80003CF7, 0x80003D0380003CFF,
-    0x80003D0B80003D07, 0x80003CC780003D0F, 0x80003D1380001CBB,
-    0x80003D1B80003D17, 0x0007A18080003D1F, 0x0007AD8200068D80,
-    0x0007B98400069D82, 0x0007B58600068D84, 0x0007918800068186,
-    0x005C418A00068588, 0x0007858C0006858A, 0x0007818E0006858C,
-    0x0007FD900006858E, 0x0007899200068590, 0x0007959400068592,
-    0x0007919600068D94, 0x0007AD980006BD96, 0x0007B99A00068D98,
-    0x0007A59C00069D9A, 0x0007A19E00068D9C, 0x00075DA00001799E,
-    0x000769A2000189A0, 0x000775A4000199A2, 0x000771A6000189A4,
-    0x00074DA80001B9A6, 0x000759AA000189A8, 0x000745AC000199AA,
-    0x80003D23000189AC, 0x80003D270001F9AE, 0x80003D2B000181B0,
-    0x80000B6880003D2F, 0x80001B1380001B0F, 0x0001ADB980001B17,
-    0x0001BDBB80001B1B, 0x0001ADBD80001B1F, 0x80001B2780001B23,
-    0x00026D80000199BE, 0x0002718280001B2B, 0x00026D8480001B2F,
-    0x0002998680001B33, 0x80001B3B80001B37, 0x80001B4380001B3F,
-    0x80001B470002A98B, 0x80001B4B005D018D, 0x0002C99080001B4F,
-    0x0002C59280001B53, 0x80001B5B80001B57, 0x0002C99680001B5F,
-    0x0002F59880001B63, 0x0002F19A80001B67, 0x0002ED9C80001B6B,
-    0x0002D99E80001B6F, 0x000225A080001B73, 0x000221A280001B77,
-    0x00023DA480001B7B, 0x000229A680001B7F, 0x000215A880001B83,
-    0x02610DAA80001B87, 0x005D81AC005C31AB, 0x005D81AE00020DAD,
-    0x00027DB080001B8B, 0x005D25B2000205B1, 0x000219B4000209B3,
-    0x000215B680001B8F, 0x000231B880001B93, 0x00023DBA80001B97,
-    0x000229BC80001B9B, 0x000225BE80001B9F, 0x0003E18080001BA3,
-    0x0003ED82001ED091, 0x0003F98480001BA7, 0x0003F586005B2D85,
-    0x80001BAB0003ED87, 0x80001BAF0003D189, 0x002BDD9E80001BB3,
-    0x00046D9C0002F99F, 0x0003B9900002F59D, 0x80001BBB80001BB7,
-    0x80001BC380001BBF, 0x0003A59680001BC7, 0x0002D99880001BCB,
-    0x80001BD380001BCF, 0x80001BDB80001BD7, 0x80001BE380001BDF,
-    0x00025DB280001BE7, 0x00025DB0000261B3, 0x000269B6000255B1,
-    0x000275B4000191B7, 0x0001BDBA000269B5, 0x0001A5B8000185BB,
-    0x000181BE0001B9B9, 0x00019DBC00018DBF, 0x80001BEB000195BD,
-    0x026A39B280001BEF, 0x026875B4026BE1B3, 0x80001BF3000215B5,
-    0x0262B1B880001BF7, 0x80001BFB02627DB9, 0x80001BFF00023DA9,
-    0x80001C0780001C03, 0x80001C0F80001C0B, 0x80001C1780001C13,
-    0x80001C1F80001C1B, 0x80001C2780001C23, 0x80001C2F80001C2B,
-    0x80001C3780001C33, 0x80001C3F80001C3B, 0x80001C4780001C43,
-    0x00EE21BA80001C4B, 0x80003226800038AF, 0x8000338F800038D7,
-    0x80002DEE80002BF3, 0x80002E5280002F21, 0x8000310C80002F1D,
-    0x8000369780003507, 0x8000380B800037CF, 0x0000C0A48000383F,
-    0x000788A6000400A5, 0x80003953000800A7, 0x80003337800032AB,
-    0x00F000AC800033E3, 0x00F1FCAE00F2A4AD, 0x000BB0A800F400AF,
-    0x000EB8AA000C00A9, 0x00F994B4000D1CAB, 0x00FFC8B600FC00B5,
-    0x00F6D0B000FDCCB7, 0x00F800B200F580B1, 0x00E7E4BC00FADCB3,
-    0x00EAE8BE00E800BD, 0x00E000B800E134BF, 0x00E3F8BA00E2F4B9,
-    0x00011DB700E400BB, 0x00011DB500011DB6, 0x00011DB300011DB4,
-    0x00011DB100011DB2, 0x00011DBF00011DB0, 0x00011DBD00011DBE,
-    0x00011DBB00011DBC, 0x00011DB900011DBA, 0x001E81A700011DB8,
-    0x001E31A5001E39A6, 0x001E21A3001E39A4, 0x001E31A1001E39A2,
-    0x001E9DAF001E39A0, 0x001E9DAD001E9DAE, 0x001E99AB001E9DAC,
-    0x001EA1A9001EA9AA, 0x001EA197001EB9A8, 0x001F7595001E8996,
-    0x001F7D93001F7994, 0x001F4591001F7192, 0x001EFD9F001F4990,
-    0x001EF99D001EF19E, 0x001EF99B001EE19C, 0x001E9599001EF19A,
-    0x001FAD87001E9598, 0x001F9585001FA186, 0x001F8183001F9984,
-    0x001F9D81001F9982, 0x001F5D8F001FE180, 0x001F758D001F418E,
-    0x001F7D8B001F798C, 0x001F8589001F718A, 0x00027590001F8988,
-    0x0002759200027D91, 0x0002859400029D93, 0x0002859600028D95,
-    0x0001259800011D97, 0x0001259A00012D99, 0x0001259C00013D9B,
-    0x0001259E00012D9D, 0x0002058000011D9F, 0x003F058200020D81,
-    0x0002018400020183, 0x0002018600020185, 0x003CE188003CDD87,
-    0x003F058A00023989, 0x0002158C00020D8B, 0x0002158E00021D8D,
-    0x0001E5B000020D8F, 0x0001E5B20001EDB1, 0x003919B40001FDB3,
-    0x0001E1B60001E1B5, 0x000121B80001E1B7, 0x000121BA000121B9,
-    0x000121BC000121BB, 0x000121BE000121BD, 0x0001E5A0000121BF,
-    0x0001E5A20001EDA1, 0x0001E5A40001FDA3, 0x0001E5A60001EDA5,
-    0x0001A5A800019DA7, 0x0001A5AA0001ADA9, 0x0001A5AC0001BDAB,
-    0x0001A5AE0001ADAD, 0x8000347B00019DAF, 0x8000397F80003797,
-    0x80002C1780003AAB, 0x800032BF80002DC2, 0x800037CB80003407,
-    0x80002FC980003AFB, 0x800037BB800033AF, 0x800031408000384F,
-    0x80003817800036F3, 0x8000323A8000303C, 0x800034538000334F,
-    0x80002C5E8000395F, 0x80002D6280002CF6, 0x800032B7800031A8,
-    0x8000352780003403, 0x800037DB800037C3, 0x80003A27800038FF,
-    0x80003AF780003ADB, 0x800035A380003563, 0x8000377B8000365B,
-    0x80003883800039B3, 0x8000305880002ED6, 0x800036D380003613,
-    0x8000356B80003437, 0x80003A1F800038CF, 0x80002FA580002ED2,
-    0x80003363800032AF, 0x8000364B80003397, 0x800036730005B082,
-    0x80002D5E800039DF, 0x80002D02800036DF, 0x800035CF80002CFE,
-    0x800037938000365F, 0x8000389F800039E7, 0x8000388F80003144,
-    0x80002F7980002BFF, 0x80003467800030B0, 0x80002D82800034E7,
-    0x80002C6E80003573, 0x80002BDF80003094, 0x800031C080003323,
-    0x80002DD680003647, 0x8000353380002EC2, 0x800038738000379B,
-    0x8000331B800032EF, 0x8000375380003154, 0x800034E380003174,
-    0x80002CB680002C33, 0x8000328B80002CFA, 0x8000363300EAFDBA,
-    0x8000387F8000372B, 0x80002D6A80003993, 0x8000304C80002E1E,
-    0x800033B3800031DC, 0x800039CF80003577, 0x80003B0780003AB3,
-    0x8000321E80003B1F, 0x80003923800032DB, 0x800030F480003018,
-    0x800031A080003110, 0x800033EF800033A3, 0x800035C38000349F,
-    0x800036CB80003667, 0x800037A380003913, 0x00019D808000393B,
-    0x0001A982003DE581, 0x0001B9840001A983, 0x0001A9860001B985,
-    0x000199880001A987, 0x003DAD8A00019989, 0x0001F58C0001ED8B,
-    0x0001F58E0001FD8D, 0x000195900001ED8F, 0x0001959200019D91,
-    0x0001B5940001AD93, 0x0001B5960001BD95, 0x000195980001AD97,
-    0x0001959A00019D99, 0x0002799C0002699B, 0x0002699E0002799D,
-    0x000299A00002699F, 0x000295A20001DDA1, 0x0002B5A40002ADA3,
-    0x0002B5A60002BDA5, 0x000295A80002ADA7, 0x0001B1AA00029DA9,
-    0x0002F1AC000291AB, 0x0002F1AE0002F1AD, 0x000291B00002F1AF,
-    0x000291B2000291B1, 0x0002B1B4000291B3, 0x0002B1B60002B1B5,
-    0x000291B80002B1B7, 0x000291BA000291B9, 0x000271BC000291BB,
-    0x000271BE000271BD, 0x000221A1000569BF, 0x000221A3000219A0,
-    0x000221A5000229A2, 0x000221A7000239A4, 0x000261A9000229A6,
-    0x000261AB000219A8, 0x000261AD000269AA, 0x000261AF000279AC,
-    0x003B15B1000269AE, 0x00022DB3000219B0, 0x00023DB500022DB2,
-    0x00022DB700023DB4, 0x00021DB900022DB6, 0x0002E1BB00021DB8,
-    0x0002E1BD0002EDBA, 0x0002E1BF0002F9BC, 0x000521810002E9BE,
-    0x00052183003B6980, 0x0005218500052982, 0x0005218700053984,
-    0x0005618900052986, 0x0005618B00051988, 0x0005618D0005698A,
-    0x0005618F0005798C, 0x000525910005698E, 0x0005259300052590,
-    0x0005259500052592, 0x0005259700052594, 0x0002E59900052596,
-    0x0002E59B0002E598, 0x0002E19D0002E59A, 0x0002E19F0006619C,
-    0x000D84910002E99E, 0x80002D1E800039B7, 0x80002E3280002D4A,
-    0x8000382B800033DF, 0x800030A880003038, 0x800032E78000316C,
-    0x8000345B8000360F, 0x80002E8E80002C43, 0x800030B480002FDD,
-    0x8000349380003477, 0x800036C3800036A3, 0x80003A1B8000399B,
-    0x80003A6780003A2B, 0x800035B780002C62, 0x80003A0380003987,
-    0x80002C1F800030CC, 0x80002F7D80002C86, 0x800031CC80002F9D,
-    0x80003503800033FB, 0x8000394F800037AF, 0x800039DB8000320E,
-    0x8000325A80002D42, 0x8000334780003266, 0x800034838000337F,
-    0x8000355F800034DF, 0x80003A7380003643, 0x800039EB80003120,
-    0x80002FC180002C76, 0x8000391780003367, 0x800030D480003084,
-    0x800039F38000326E, 0x80002E0E80002D26, 0x800031F780002FAD,
-    0x800032978000324A, 0x8000347F8000332B, 0x8000368F800034F3,
-    0x8000383B8000382F, 0x80002D9280003A0F, 0x80002E1280003383,
-    0x800034A3800033FF, 0x800039F7800037BF, 0x80003B0B80003AE3,
-    0x8000335F80003262, 0x800035FB800036FB, 0x8000344F8000361F,
-    0x8000389B800033D7, 0x8000376380002C3B, 0x80002D1A80002D2A,
-    0x8000314880003028, 0x80002F698000362B, 0x8000333B00016482,
-    0x8000391F8000321A, 0x80003044800039E3, 0x80002E6A80002C96,
-    0x8000320A80002EBE, 0x8000345780002D06, 0x8000357F8000351B,
-    0x8000359F8000359B, 0x80003A33800035B3, 0x800037C780003623,
-    0x8000393F8000388B, 0x80003A878000396B, 0x80003A8F80003A8B,
-    0x8000396380003AF3, 0x00017980800039FF, 0x0001898200017981,
-    0x0001998400018983, 0x00018D8600018585, 0x0034318800018587,
-    0x0034358A0001B989, 0x0001958C00018D8B, 0x80002C6600019D8D,
-    0x0001ED9080002C8A, 0x0001F99280002C9E, 0x80002D6680002D5A,
-    0x0001E1960001E195, 0x0001A1980001E197, 0x0001A19A0001A199,
-    0x0001A19C0001A19B, 0x0032F99E0001A19D, 0x000159A080002DA6,
-    0x000155A280002E5A, 0x80002E7E80002E76, 0x00017DA6000175A5,
-    0x80002EC680002EBA, 0x000141AA80002FA9, 0x000141AC000141AB,
-    0x000141AE000141AD, 0x0000F9B0000141AF, 0x00008DB20000F9B1,
-    0x000081B4000081B3, 0x00005DB60000A9B5, 0x000071B8000051B7,
-    0x000049BA000079B9, 0x00310DBC000049BB, 0x0002F1BE0002F1BD,
-    0x00020DA30002F1BF, 0x00020DA100020DA2, 0x00020DA700020DA0,
-    0x00020DA500020DA6, 0x000209AB00020DA4, 0x003741A9000201AA,
-    0x000209AF00020DA8, 0x000209AD000201AE, 0x000209B3000211AC,
-    0x000209B1000201B2, 0x003119B7000271B0, 0x000209B5000201B6,
-    0x000205BB000211B4, 0x000235B9000205BA, 0x000205BF000235B8,
-    0x000215BD000205BE, 0x000D1983000215BC, 0x000D1981000D1182,
-    0x000D3587000D0180, 0x000D0585000D3586, 0x000D158B0034E184,
-    0x000D0589000D158A, 0x000D758F000D0588, 0x000D058D000D758E,
-    0x000D1193000D058C, 0x0024E191000D0992, 0x000D3197000D0590,
-    0x000D0195000D0996, 0x000EC19B000D0994, 0x000D0199000D099A,
-    0x000D0D9F000D0998, 0x000D0D9D000D0D9E, 0x00E581BA000D0D9C,
-    0x800030E8800030C0, 0x80003108800030F0, 0x800031E4800031B4,
-    0x8000328F80003212, 0x8000336F80003353, 0x800033F38000339F,
-    0x8000348780003413, 0x8000358B80003567, 0x8000358F80003587,
-    0x800035AB80003593, 0x800035D3800035AF, 0x80003603800035E7,
-    0x800036778000366F, 0x800036B38000368B, 0x8000373780003703,
-    0x800038478000379F, 0x800038938000385F, 0x800038DB80003897,
-    0x80003937800038DF, 0x80003A5780003A13, 0x800030B880003A6F,
-    0x8000371B8000F2C0, 0x80002CF280002BE7, 0x80002C5A80002CB2,
-    0x80002CC680002C9A, 0x80002D7280002D56, 0x80002E5680002E46,
-    0x80002ECE80002E72, 0x80002F2580002F19, 0x80002F5180002F41,
-    0x8000304880003040, 0x0007ED8080003078, 0x0007FD820007F581,
-    0x0007ED840007F583, 0x00011D8600011585, 0x00012D8800011587,
-    0x0001398A00072989, 0x0001258C00073D8B, 0x0001258E00012D8D,
-    0x0001659000011D8F, 0x000F01920004F591, 0x00017D9400016593,
-    0x000805960036D195, 0x00015198000BED97, 0x0007459A00015199,
-    0x0007419C0001559B, 0x0001399E0001499D, 0x000439A00001399F,
-    0x0001C5A20001CDA1, 0x0001C5A40001DDA3, 0x000425A60001CDA5,
-    0x0001C1A80001C1A7, 0x0001C1AA0001C1A9, 0x0001C1AC0001C1AB,
-    0x0001CDAE0007D5AD, 0x00082DB00001C5AF, 0x0007B9B20001B9B1,
-    0x000155B400014DB3, 0x000155B600015DB5, 0x000765B800014DB7,
-    0x000435BA000171B9, 0x000779BC000175BB, 0x000165BE0008E5BD,
-    0x80003098000771BF, 0x800030E0800030C8, 0x800030E4800030D0,
-    0x8000317C80003124, 0x8000319080003188, 0x8000323E800031B8,
-    0x8000339380003252, 0x800033BB8000338B, 0x800034178000354B,
-    0x800034978000344B, 0x800034D7800034B3, 0x800034FF800034FB,
-    0x8000352F8000351F, 0x8000353F80003543, 0x800035EB8000356F,
-    0x800036538000361B, 0x8000376780003683, 0x800038038000378F,
-    0x800038578000384B, 0x8000387B80003877, 0x0010E0F080003887,
-    0x8000391B800038A3, 0x800039838000394B, 0x800039EF800039A3,
-    0x80003A6380003A47, 0x8000F25780003AC3, 0x8000F2848000F252,
-    0x80002B3780002B0F, 0x8000F30680002B3B, 0x8000F3AB8000F333,
-    0x80003B5F80003B4F, 0x800006EB800006E0, 0x0005B1AD800006EE,
-    0x800006E30005A9AC, 0x0005B1A9800006E7, 0x0005B1AB000589A8,
-    0x0005B5A50005B9AA, 0x0006C9A70005B5A4, 0x0005B5A10005B5A6,
-    0x0005B5A30005B5A0, 0x0005A1BD0005B5A2, 0x0005A1BF0006E9BC,
-    0x000585B90005A9BE, 0x000585BB0007D1B8, 0x0005A9B5000585BA,
-    0x0005B9B70005A1B4, 0x0005C9B10005A1B6, 0x0005B9B30005C1B0,
-    0x0005C18D0005C1B2, 0x0005C18F0020B98C, 0x0005A5890005C98E,
-    0x0005A58B0005A588, 0x000585850005A58A, 0x0005858700058584,
-    0x0005A58100058586, 0x0005A5830005A580, 0x0005B99D0005A582,
-    0x0005B59F0005B19C, 0x0005B9990005B59E, 0x0005A99B0005B198,
-    0x0005BD950005B19A, 0x000A05970005BD94, 0x0020BD9100058D96,
-    0x0005AD930005B990, 0x8000084D0005AD92, 0x80000E9680000EA5,
-    0x80000EC880000E9B, 0x80000F3580000EA0, 0x80000F6B80000FB4,
-    0x80000F4480000F19, 0x80000F5980000F51, 0x80000FAC80000F87,
-    0x80000FA780000FA2, 0x80000F9B80000F94, 0x80000EED80000EE8,
-    0x80000EFF80000EF2, 0x80000F1480000F0C, 0x80000F2680000F1C,
-    0x80000F3080000F2B, 0x80000F3F80000F3A, 0x0007518880000F47,
-    0x0004298A00023989, 0x0004358C00024D8B, 0x0031198E0002518D,
-    0x000205800002558F, 0x0002018200043181, 0x0002018400020183,
-    0x0002018600020185, 0x00022D9800042187, 0x80000F5400023599,
-    0x80000F6180000F5C, 0x80000F6E80000F66, 0x00022D9080000F73,
-    0x00023D9200023591, 0x00022D9400023593, 0x00021D9600021595,
-    0x00FFD18000021597, 0x80000F7D00FFD181, 0x00FFCD8400FFD583,
-    0x0000118680000F82, 0x0001218800000987, 0x0001218A00012189,
-    0x0001218C0001218B, 0x0001218E0001218D, 0x000161900001218F,
-    0x00016D9200016591, 0x0001759400016993, 0x0001719600017195,
-    0x0001499800017597, 0x0001459A00014D99, 0x00015D9C0001419B,
-    0x0001499E0001599D, 0x0001B1A000014D9F, 0x0001BDA20001B5A1,
-    0x0001A5A40001B9A3, 0x0001A1A60001A1A5, 0x000199A80001A5A7,
-    0x000195AA00019DA9, 0x00018DAC000191AB, 0x0001F9AE000189AD,
-    0x000181B00001FDAF, 0x00018DB2000185B1, 0x000195B4000189B3,
-    0x000191B6000191B5, 0x0001A9B8000195B7, 0x0001A5BA0001ADB9,
-    0x0001BDBC0001A1BB, 0x0001A9BE0001B9BD, 0x80000F8A0001ADBF,
-    0x80000FAF80000F8F, 0x80000F0480000F21, 0x80000F7880000F4C,
-    0x800016C780000EF7, 0x800016D3800016D0, 0x800016CD800016D9,
-    0x800016CA800016D6, 0x80001700800016FD, 0x800016DC800016DF,
-    0x800016E5800016E2, 0x800016EE00061091, 0x800016F1800016EB,
-    0x800016F7800016E8, 0x80001703800016F4, 0x8000170F80001709,
-    0x800017128000170C, 0x8000171880001715, 0x800017068000171E,
-    0x800017278000172A, 0x8000174180001732, 0x8000172480001738,
-    0x8000173E80001735, 0x80000FCB80001663, 0x80000FF88000101B,
-    0x800010078000100C, 0x8000101680001011, 0x8000173B80000FFD,
-    0x80000FD580000FD0, 0x8000100280000FE9, 0x8000104780001042,
-    0x8000106980001053, 0x00FD39A880001078, 0x00FD35AA00BB5DA9,
-    0x00FD55AC00FD4DAB, 0x00FD55AE00FD5DAD, 0x00FD01A000FD4DAF,
-    0x00BBF1A200BBE5A1, 0x00FD19A400FD09A3, 0x00FD09A600FD19A5,
-    0x00FD31B800FD09A7, 0x00FD31BA00FD31B9, 0x00FD99BC00FD31BB,
-    0x00FD95BE8000107D, 0x00FD35B080001088, 0x00FD35B200FD3DB1,
-    0x00FD15B400FD0DB3, 0x00FD15B600FD1DB5, 0x800010B5800010A2,
-    0x8000110B800010D9, 0x8000111880001110, 0x8000113680001127,
-    0x00FF55808000113B, 0x00FF558200FF5D81, 0x00FCBD8400FCA583,
-    0x00FDB98600FDB585, 0x8000116680001143, 0x8000118D02A3C0D9,
-    0x800011B980001199, 0x800011BE00FDCD9D, 0x800011C300FDC19F,
-    0x02A3FCD9800011FD, 0x00FDF99400FDE993, 0x00FDE99600FDF995,
-    0x0002759500FDE997, 0x0002719700027194, 0x0002619100027596,
-    0x00026D9300026590, 0x00025D9D00026992, 0x0002A99F0002599C,
-    0x000249990002AD9E, 0x0002459B00024D98, 0x00021D850002419A,
-    0x0002298700021984, 0x0002018100022D86, 0x0002018300020580,
-    0x00020D8D00020582, 0x0002198F0002098C, 0x0002158900021D8E,
-    0x0002118B00021188, 0x800012100002158A, 0x800012348000121C,
-    0x02A129B1800012AE, 0x8000130402A12DB0, 0x800013228000131D,
-    0x8000134680001335, 0x8000135D80001358, 0x8000138C80001384,
-    0x02A0CDA5800013A1, 0x000245A702A0C9A4, 0x000255A1000241A6,
-    0x00025DA3000251A0, 0x000269AD000259A2, 0x000249AF00026DAC,
-    0x00027DA900024DAE, 0x000271AB000279A8, 0x0005DD80000275AA,
-    0x0005ED820005E581, 0x0005F9840005D983, 0x0005E9860005F985,
-    0x0005D9880005E987, 0x00067D8A0005D989, 0x00064D8C0006718B,
-    0x0006BD8E0006B58D, 0x0006CD900006B58F, 0x0006DD920006D591,
-    0x0006CD940006D593, 0x000699960006E995, 0x0006AD9800069597,
-    0x0006B19A0006B599, 0x0006959C00068D9B, 0x0007FD9E0007F59D,
-    0x00070DA00007F59F, 0x000715A200071DA1, 0x000735A400072DA3,
-    0x000735A600073DA5, 0x000715A800072DA7, 0x000715AA00071DA9,
-    0x0001B5AC0001ADAB, 0x0001B5AE0001BDAD, 0x0001D5B00001ADAF,
-    0x0001D5B20001DDB1, 0x0001F5B40001EDB3, 0x0001F5B60001FDB5,
-    0x0001D5B80001EDB7, 0x0001D5BA0001DDB9, 0x0001B5BC0001ADBB,
-    0x0001B5BE0001BDBD, 0x800013C80001ADBF, 0x80001427800013CD,
-    0x800014318000142C, 0x800014498000143D, 0x800014568000144E,
-    0x8000147C80001462, 0x8000148980001481, 0x800014938000148E,
-    0x8000149D80001498, 0x800014B5800014A2, 0x800014D6800014BA,
-    0x800014F0001BB084, 0x800015168000150A, 0x800015338000152E,
-    0x8000156180001540, 0x8000159E80001584, 0x800015AF800015AA,
-    0x800015D8800015B7, 0x80001601800015F2, 0x800016238000161E,
-    0x800016308000162B, 0x8000164880001643, 0x8000167A8000166E,
-    0x8000169580001686, 0x800016B7800016B2, 0x800011E9800011D8,
-    0x8000008080001666, 0x800000948000008A, 0x800000A88000009E,
-    0x80000FDF800000B2, 0x80000FEE80000FE4, 0x800010648000105F,
-    0x0007B9AB8000106E, 0x0007BDA90007BDAA, 0x0007D1AF0007B9A8,
-    0x0007A5AD0007D5AE, 0x00078DA30007A1AC, 0x000781A1000789A2,
-    0x000781A7000785A0, 0x000789A5000785A6, 0x000795BB00078DA4,
-    0x00079DB9000791BA, 0x000789BF00079DB8, 0x000781BD00078DBE,
-    0x0007A9B3000785BC, 0x0007ADB10007ADB2, 0x0007A1B70007A9B0,
-    0x0007B5B50007A5B6, 0x800010CF0007B1B4, 0x80001101800010D4,
-    0x800011228000111D, 0x8000153B8000112C, 0x800015FC800015F7,
-    0x8000168B80001614, 0x800016A880001690, 0x80000FF380000FDA,
-    0x00075D9B80001073, 0x0007519900075D9A, 0x0007759F00075998,
-    0x02A4799D0007719E, 0x000771930007419C, 0x800012C180001106,
-    0x0007659780001529, 0x0007699500076596, 0x0002AD9500076D94,
-    0x0002BD970002AD94, 0x00028D910002BD96, 0x0002BD9300028D90,
-    0x0002ED9D0002BD92, 0x0002FD9F0002ED9C, 0x00028D990002FD9E,
-    0x0002FD9B00028D98, 0x00022D850002FD9A, 0x00023D8700022D84,
-    0x00020D8100023D86, 0x00023D8300020D80, 0x0002ED8D00023D82,
-    0x0002FD8F0002ED8C, 0x00020D890002FD8E, 0x0002FD8B00020D88,
-    0x00050DB50002FD8A, 0x000119B700050DB4, 0x000115B1000101B6,
-    0x000111B3000115B0, 0x0002BDBD000519B2, 0x000285BF0002BDBC,
-    0x000125B9000285BE, 0x000525BB000521B8, 0x00022DA5000525BA,
-    0x000239A700022DA4, 0x00020DA1000629A6, 0x00023DA300020DA0,
-    0x00021DAD00023DA2, 0x0005BDAF00061DAC, 0x000619A9000611AE,
-    0x000201AB000621A8, 0x000D1180000209AA, 0x000D2982000D1181,
-    0x000D2584000D2983, 0x000D2586000D2D85, 0x000D6588000D5D87,
-    0x000D718A000D7589, 0x000D518C000D718B, 0x000D558E000ED98D,
-    0x000D3590000D4D8F, 0x000DCD92000D3D91, 0x000DDD94000DC593,
-    0x00015996000DC595, 0x000DC198000DC197, 0x000DC59A00016199,
-    0x000DC59C000DDD9B, 0x0001719E000DCD9D, 0x000E39A00009999F,
-    0x000E35A200097DA1, 0x000E75A4000A09A3, 0x000195A6000A01A5,
-    0x0001B1A8000B15A7, 0x0001B1AA0001B1A9, 0x000DF5AC000DEDAB,
-    0x000D29AE000DD5AD, 0x000D79B00001BDAF, 0x0001D5B20001CDB1,
-    0x0001F5B40001EDB3, 0x0001F5B60001FDB5, 0x0001D5B80001EDB7,
-    0x0001D5BA0001DDB9, 0x0001B5BC0001ADBB, 0x0001B5BE0001BDBD,
-    0x800016190001ADBF, 0x800016AD8000164D, 0x8000124E80001131,
-    0x8000129C80001289, 0x80001407800013FB, 0x8000137780001413,
-    0x800013BB8000137C, 0x800013E7800013C0, 0x80001253800013EC,
-    0x800012A180001258, 0x800011AC800012A6, 0x800011B100FC10F0,
-    0x8000118580001180, 0x800011CD800011C8, 0x8000131580001310,
-    0x800013508000134B, 0x8000126C80001260, 0x800012848000127F,
-    0x800012C68000122F, 0x8000103580001341, 0x800010A78000108D,
-    0x800010BA800010AE, 0x800010E5800010DE, 0x8000116B800010EC,
-    0x8000119E800011A5, 0x8000120280001215, 0x8000124080001209,
-    0x8000124780001239, 0x80001309800012B3, 0x8000126580001271,
-    0x800012958000128E, 0x8000133A80001327, 0x0009A5B680001362,
-    0x000995B40009ADB7, 0x000CF9B20009ADB5, 0x000991B000098DB3,
-    0x000F29BE000991B1, 0x000F29BC000F29BF, 0x000989BA000F29BD,
-    0x000995B8000989BB, 0x0008D5A6000989B9, 0x0008CDA40008F9A7,
-    0x00082DA2000CA9A5, 0x000CE5A00008D5A3, 0x0009E1AE000801A1,
-    0x000889AC0009E1AF, 0x0008BDAA000889AD, 0x0008B9A80008A5AB,
-    0x000F41960008B5A9, 0x000C3994000B5997, 0x000C3D92000F3195,
-    0x000F5590000F2D93, 0x000B559E000C3591, 0x000B459C000B5D9F,
-    0x000B559A000B5D9D, 0x000B5998000C0D9B, 0x000FA586000C0999,
-    0x000F8584000FAD87, 0x000F9582000FAD85, 0x000FE980000F9D83,
-    0x000F758E000F9D81, 0x000F798C000F7D8F, 0x000C6D8A000C498D,
-    0x000C6188000C558B, 0x00023995000F8589, 0x0002499700023194,
-    0x0002399100023196, 0x0002299300028190, 0x00027D9D00023192,
-    0x00024D9F00027D9C, 0x0002799900024D9E, 0x0002699B00027198,
-    0x00022D850002719A, 0x00023D8700022D84, 0x00020D8100023D86,
-    0x00023D8300020D80, 0x0002E98D00023D82, 0x0002F98F0002718C,
-    0x00020D890002E18E, 0x0002FD8B00020D88, 0x000801B50002FD8A,
-    0x0002F5B7000961B4, 0x000281B1000809B6, 0x000829B3000281B0,
-    0x000239BD000285B2, 0x8000136900023DBC, 0x000235B980001370,
-    0x00080DBB000235B8, 0x0002BDA500080DBA, 0x00024DA70002BDA4,
-    0x0002BDA100024DA6, 0x0002ADA30002BDA0, 0x00029DAD0002ADA2,
-    0x00026DAF000271AC, 0x000279A900026DAE, 0x000269AB0002E1A8,
-    0x80001391000271AA, 0x800013AD800013A6, 0x800013E0800013D2,
-    0x80001436800013D9, 0x8000146E80001467, 0x80001503800014F5,
-    0x800014DB800014FC, 0x8000151B8000150F, 0x8000156D80001566,
-    0x800015458000157D, 0x8000158980001553, 0x8000154C80001590,
-    0x8000163C80001635, 0x800015E4800015DD, 0x800015CA800015C3,
-    0x800016068000160D, 0x800010588000169A, 0x800010948000109B,
-    0x800010C1800010C8, 0x800010F3800010FA, 0x8000114880001179,
-    0x8000122180001172, 0x80001278800012BA, 0x800014E98000132E,
-    0x8000167F80001522, 0x80001673028E30CC, 0x800015A3800016A1,
-    0x800015EB80001475, 0x800014AE800013B4, 0x80001597800015BC,
-    0x800014A7800014E2, 0x800011928000114F, 0x000125808000155A,
-    0x0001258200012D81, 0x0001218400013D83, 0x00012D8600012185,
-    0x0001198800012587, 0x0001698A00011989, 0x0001798C0001698B,
-    0x0001698E0001798D, 0x800014420001698F, 0x000111928000104C,
-    0x0001319400011193, 0x0001319600013195, 0x00010D9800013197,
-    0x00011D9A00011599, 0x00010D9C0001159B, 0x0001F99E0001099D,
-    0x000109A00001F99F, 0x000119A2000109A1, 0x000109A4000119A3,
-    0x000139A6000109A5, 0x000109A8000139A7, 0x000119AA000109A9,
-    0x000109AC000119AB, 0x000179AE000109AD, 0x00010DB0000175AF,
-    0x00011DB2000115B1, 0x0001CDB4000115B3, 0x000105B60001F1B5,
-    0x000105B800013DB7, 0x000105BA00010DB9, 0x00011DBC00011DBB,
-    0x00010DBE000111BD, 0x0007FDB1000105BF, 0x0005A1B30007FDB0,
-    0x000211B50007F1B2, 0x00021DB70005B9B4, 0x00022DB9000219B6,
-    0x00023DBB00022DB8, 0x000229BD00023DBA, 0x000219BF00022DBC,
-    0x000795A1000221BE, 0x000791A3000795A0, 0x00078DA5000795A2,
-    0x0007BDA7000789A4, 0x000789A90007BDA6, 0x000785AB000781A8,
-    0x000781AD000781AA, 0x000781AF000785AC, 0x00072191000789AE,
-    0x00072D9300071990, 0x00073D9500072D92, 0x00072D9700073D94,
-    0x0007199900072D96, 0x0007699B00071D98, 0x0007799D0007119A,
-    0x0007699F0007719C, 0x800012280007719E, 0x800012FD800015D1,
-    0x8000102C8000145B, 0x8000157480001023, 0x800011E0800012CB,
-    0x8000165580001398, 0x800012DB800012D4, 0x800011EE80001156,
-    0x0003FD808000033D, 0x00024D8200024581, 0x000311840003FD83,
-    0x0001D9860006F585, 0x8000243B0001D987, 0x800006378000243F,
-    0x8000064C800000C6, 0x800024DD800024D9, 0x8000208380002087,
-    0x8000088E8000065E, 0x8000246F80000892, 0x80002463800024D5,
-    0x8000244B80002467, 0x800024538000244F, 0x8000245B80002457,
-    0x800006568000245F, 0x800000D08000065A, 0x80000339800000D6,
-    0x800006588000033F, 0x800000D4800000D2, 0x8000007880000654,
-    0x8000007C800013F1, 0x8000009080000086, 0x8000009A800013F6,
-    0x000109B080001402, 0x000119B2000109B1, 0x000109B4000119B3,
-    0x000139B6000109B5, 0x000109B8000139B7, 0x000119BA000109B9,
-    0x000109BC000119BB, 0x800000A48000063F, 0x00AE79808000140E,
-    0x001F6582001F6D81, 0x001F6584001F7D83, 0x800000AE8000064A,
-    0x001F5D88001F6587, 0x00A9318A00A93D89, 0x00A9298C00A9358B,
-    0x001F7D8E001F718D, 0x001EBD90001F798F, 0x00A78192001F6591,
-    0x028071948000141A, 0x001F6996001F7D95, 0x001F5D98001F7197,
-    0x00A9059A00A90199, 0x001F3D9C001F259B, 0x001F2D9E001F259D,
-    0x001F05A0001FFD9F, 0x00A9D1A2001F0DA1, 0x001F65A4001F01A3,
-    0x00A9F1A6000081A5, 0x001F31A8800000B8, 0x001F31AA001F31A9,
-    0x8000141F001F31AB, 0x8000102080000FB9, 0x001F61B08000103F,
-    0x001F61B2001F61B1, 0x001F7DB480001082, 0x001F69B680001085,
-    0x001F59B8001F69B7, 0x001F11BA001F59B9, 0x0000E9BC001F11BB,
-    0x000065BE0000B9BD, 0x0001B58000004DBF, 0x02B3698202B36581,
-    0x02B3798402B36583, 0x02B3998602B39585, 0x02B3518802B36D87,
-    0x02B3A18A02B3AD89, 0x02B3B98C02B3A58B, 0x0001818E00018D8D,
-    0x0001FD900001818F, 0x0001F5920001F991, 0x00019D9400018193,
-    0x00018D9600018195, 0x0001B59800018997, 0x0003299A00032599,
-    0x0003319C00032D9B, 0x00033D9E0003319D, 0x0003C5A00003399F,
-    0x0003C1A20003CDA1, 0x0003D9A40003C5A3, 0x0003D5A60003D9A5,
-    0x0003EDA80003D1A7, 0x0003F9AA0003F5A9, 0x0003E5AC0003F9AB,
-    0x0003E1AE0003EDAD, 0x00039DB00003E1AF, 0x0003A9B20003A5B1,
-    0x0003B1B40003ADB3, 0x0003BDB60003B1B5, 0x000385B80003B9B7,
-    0x000381BA00038DB9, 0x000399BC000385BB, 0x000395BE000399BD,
-    0x001D5DBC000391BF, 0x001D6DBE001D65BD, 0x001D5DB800AC79BF,
-    0x001D4DBA001D45B9, 0x001D65B4001D45BB, 0x00AB45B600AB41B5,
-    0x00A5E1B0001D65B7, 0x001D79B2001D09B1, 0x80001115001D79B3,
-    0x8000118A80001140, 0x800011D2800011B6, 0x800011DD800011D5,
-    0x800011FA800011F7, 0x800012AB8000125D, 0x800013558000131A,
-    0x8000138980001381, 0x80001424800013C5, 0x8000148680001453,
-    0x001DA998800014BF, 0x8000153800A54D99, 0x02828194800015B4,
-    0x001D9996001D8995, 0x001C4D90001D9997, 0x001D9D92001D9591,
-    0x80001628001D9593, 0x800014C7800014C2, 0x800014D1800014CC,
-    0x800000D8800000CE, 0x8000065C8000034A, 0x8000089480000890,
-    0x800021D5800021D1, 0x8000256380002AAF, 0x001010F08000259C,
-    0x800025CD800025BB, 0x800029F1800025F2, 0x80002A1F80002A0D,
-    0x80002AB3800027B6, 0x80002A970291E8CC, 0x8000255B80002557,
-    0x8000089980000896, 0x0000D9A6800008A2, 0x00000DA48000089F,
-    0x00006DA2000029A5, 0x00006DA0029801A3, 0x0000D9AE000065A1,
-    0x0000CDAC8000089C, 0x0000D1AA0000C1AD, 0x0000E5A8000779AB,
-    0x00CBB596000775A9, 0x00CBA59400CBBD97, 0x00CBB59200CBBD95,
-    0x800020BA00CBBD93, 0x800020CA800021C5, 0x00CBE59C800020D2,
-    0x00CB8D9A800021C9, 0x800021CD00CBFD9B, 0x00C529868000EBDD,
-    0x00C5298400C53587, 0x00C5018200C52985, 0x8000EBE200C50983,
-    0x00CBFD8E02B294D9, 0x00CBE58C00CBC58F, 0x00C5058A00CBF58D,
-    0x8000EBE700CBFD8B, 0x800016528000EBEC, 0x800020C68000166B,
-    0x8000EBF1800020CE, 0x0034289E0034109D, 0x0002209000358C9F,
-    0x8000EBF600092C91, 0x8000EC008000EBFB, 0x0036E0968000EC05,
-    0x0001C0908000EC0A, 0x000F88928000EC0F, 0x8000EC14000FCC93,
-    0x000DD896000D9495, 0x8000EC1E8000EC19, 0x8000EC288000EC23,
-    0x8000EC328000EC2D, 0x0008009E8000EC37, 0x8000EC418000EC3C,
-    0x8000EC4B8000EC46, 0x8000EC558000EC50, 0x8000EC5F8000EC5A,
-    0x8000EC698000EC64, 0x8000EC738000EC6E, 0x8000EC7D8000EC78,
-    0x8000EC878000EC82, 0x8000EC918000EC8C, 0x000BC8B28000EC96,
-    0x8000ECA08000EC9B, 0x8000ECAA8000ECA5, 0x8000ECAF002CD8AF,
-    0x8000ECB98000ECB4, 0x8000ECC38000ECBE, 0x8000ECCD8000ECC8,
-    0x001A2D810009B0F0, 0x001A89838000ECD2, 0x001A89850007D982,
-    0x0007C187001A8184, 0x00AC4D89001A9986, 0x00AC718B00AC4188,
-    0x001ABD8D001AA58A, 0x0007E58F001BC98C, 0x029E959100A2F58E,
-    0x029EC59300AC7590, 0x00AC6995029EC592, 0x00AC199700AC6194,
-    0x00AC319900AC6196, 0x0285499B00AC2D98, 0x00AC199D001A459A,
-    0x001A499F001A819C, 0x0295B5A10005699E, 0x029079A3001A19A0,
-    0x029181A5029641A2, 0x0291A9A702919DA4, 0x0291F5A90291B1A6,
-    0x02916DAB02919DA8, 0x02908DAD029119AA, 0x029041AF029159AC,
-    0x029265B102902DAE, 0x029D89B3029DEDB0, 0x029D45B5029DFDB2,
-    0x029C55B7029D8DB4, 0x029C2DB9029CB9B6, 0x001A91BB029C55B8,
-    0x001A8DBD029879BA, 0x001ABDBF001A65BC, 0x000B548700071DBE,
-    0x8000ECDC8000ECD7, 0x0004C1BB0004DDBC, 0x0004D5B90004D9BA,
-    0x0004E9B70004D5B8, 0x0004EDB50004E1B6, 0x000235B3000229B4,
-    0x00023DB1000231B2, 0x0017F5AF001789B0, 0x000249AD000245AE,
-    0x000491AB00048DAC, 0x000499A9000495AA, 0x0004FDA70004C1A8,
-    0x0004F5A50004F9A6, 0x0004E9A30004F5A4, 0x0004E1A10004EDA2,
-    0x00040D9F0004F1A0, 0x0004059D0004099E, 0x0004199B0004059C,
-    0x0004119900041D9A, 0x0004259700041998, 0x00042D9500042196,
-    0x0004319300042D94, 0x0004399100043592, 0x0004BD8F0004C190,
-    0x0004B58D0004B98E, 0x0004A98B0004B58C, 0x0004A1890004AD8A,
-    0x000495870004A988, 0x00049D8500049186, 0x0004818300049D84,
-    0x0004898100048582, 0x0003B1A000049180, 0x0003B1A20003B1A1,
-    0x0003D1A40003B1A3, 0x0003D5A60003D1A5, 0x8000ECE10003CDA7,
-    0x8000ECEB8000ECE6, 0x8000ECF58000ECF0, 0x00A7C4CC8000ECFA,
-    0x8000ECFF00A7E0CC, 0x8000ED098000ED04, 0x8000ED138000ED0E,
-    0x8000ED1D8000ED18, 0x02B638D98000ED22, 0x8000ED2702B7F8D9,
-    0x8000ED318000ED2C, 0x8000ED3B8000ED36, 0x00043D808000ED40,
-    0x0004718200045D81, 0x0003818400047183, 0x0003B5860003BD85,
-    0x000395880003AD87, 0x0003ED8A00039989, 0x0003ED8C0003FD8B,
-    0x0003DD8E0003D58D, 0x0003AD900003D58F, 0x0003BD920003B591,
-    0x000351940003B193, 0x0003519600035195, 0x0003719800035197,
-    0x0003719A00037199, 0x0003519C0003719B, 0x0003519E0003519D,
-    0x00F981800003519F, 0x00F9818200F98181, 0x00F9818400F98183,
-    0x00F97D8600F97585, 0x00F94D8800F97587, 0x00F9BD8A00F95589,
-    0x00F9AD8C00F9B58B, 0x00F9FD8E00F9F58D, 0x00F9399000F9498F,
-    0x00F9E99200F93991, 0x00F9F99400F9E993, 0x00F9E99600F9F995,
-    0x00F9D99800F9E997, 0x00F9A19A00F9D999, 0x00F9A19C00F9A19B,
-    0x000B559E000B5D9D, 0x00C0D9A0000B4D9F, 0x00C089A200C2C5A1,
-    0x00C095A400CEA1A3, 0x00CE0DA600CEA9A5, 0x00C0B5A800C08DA7,
-    0x00CE31AA00C0E5A9, 0x00CE19AC00CE31AB, 0x00CE01AE00CE05AD,
-    0x00CE01B000CE01AF, 0x00C099B200C081B1, 0x00CE15B400C099B3,
-    0x00C05DB600C055B5, 0x00C075B800C04DB7, 0x00C02DBA00C025B9,
-    0x00C025BC00C03DBB, 0x00C1FDBE00C1F5BD, 0x000DD5B08000ED45,
-    0x000DE9B2000DDDB1, 0x000DF9B4000DE9B3, 0x000DE9B6000DF9B5,
-    0x000DD9B8000DE9B7, 0x000F2DBA000C21B9, 0x000E11BC000F25BB,
-    0x000E11BE000E11BD, 0x8000ED4A000E11BF, 0x8000ED548000ED4F,
-    0x8000ED5E8000ED59, 0x8000ED688000ED63, 0x8000ED728000ED6D,
-    0x8000ED7C8000ED77, 0x8000ED868000ED81, 0x8000ED908000ED8B,
-    0x000EB1808000ED95, 0x000EC982000EB181, 0x000ED984000EC983,
-    0x000EC986000ED985, 0x000EF988000EC987, 0x000EC18A000EF989,
-    0x000EC18C000EC18B, 0x000EC18E000EC18D, 0x00013D90000EC18F,
-    0x00008D9200008591, 0x8000ED9A00008593, 0x8000EDA48000ED9F,
-    0x8000EDAE8000EDA9, 0x8000EDB88000EDB3, 0x8000EDC28000EDBD,
-    0x8000EDCC8000EDC7, 0x00FB45B98000EDD1, 0x00FB45BB00FB45B8,
-    0x00FB45BD00FB45BA, 0x00FB45BF00FB45BC, 0x00FB05B100FB45BE,
-    0x00FB05B3000905B0, 0x00FB05B500FB05B2, 0x00FB05B700FB05B4,
-    0x000945A900FB05B6, 0x000945AB0008B9A8, 0x000945AD000945AA,
-    0x000945AF000945AC, 0x001545A1000945AE, 0x001505A3000895A0,
-    0x000885A5001505A2, 0x000881A7001555A4, 0x028B859900FBA1A6,
-    0x0015C19B02933598, 0x0015DD9D0297199A, 0x0008759F0008719C,
-    0x029D05910008759E, 0x0292E99302928D90, 0x0292259502929D92,
-    0x028AFD9700AEB594, 0x00AEC9890293D996, 0x029E0D8B028AF188,
-    0x029FED8D028AF18A, 0x029F218F029E398C, 0x029AD581029F4D8E,
-    0x029F198300AD4D80, 0x029EE18502992182, 0x029EC987029EFD84,
-    0x00024DB0029ED186, 0x00025DB2000255B1, 0x00024DB4000255B3,
-    0x00027DB6000275B5, 0x00024DB8000275B7, 0x00025DBA000255B9,
-    0x000249BC8000EDD6, 0x0002E9BE000249BD, 0x8000EDDB0002E9BF,
-    0x8000EDE58000EDE0, 0x8000EDEF8000EDEA, 0x8000EDF98000EDF4,
-    0x8000EE038000EDFE, 0x80000BBD80000BBA, 0x80000B9880000B80,
-    0x80000B9E80003D37, 0x0002598080000B9B, 0x0002698200025981,
-    0x0002798400026983, 0x0002698600027985, 0x0002598800026987,
-    0x0000858A00025989, 0x0000F18C80001CBF, 0x0000F18E0000F18D,
-    0x000091900000F18F, 0x0000919200009191, 0x00008D9480000ADB,
-    0x80000AE70000B595, 0x80000AF980000BAA, 0x80000B8380000AF3,
-    0x80000B0280000B86, 0x80000BAD80000B53, 0x8000F12180000BB0,
-    0x80000B1780003CBB, 0x80000B6B8000F126, 0x80000B2F8000F12B,
-    0x80000B3B80000B32, 0x80000B448000F130, 0x80000BA180000BA7,
-    0x80000BA480003D3B, 0x80000B6E800022B7, 0x80000B9580000B92,
-    0x80000A0980000B7D, 0x80000A0F80000A0C, 0x8000F13A8000F135,
-    0x8000EE0D8000EE08, 0x8000EE178000EE12, 0x0002FDA78000EE1C,
-    0x8000EE268000EE21, 0x8000EE308000EE2B, 0x8000EE3A8000EE35,
-    0x8000EE448000EE3F, 0x8000EE4E8000EE49, 0x8000EE588000EE53,
-    0x8000EE628000EE5D, 0x8000EE6C8000EE67, 0x8000EE768000EE71,
-    0x8000EE808000EE7B, 0x8000EE8A8000EE85, 0x8000EE948000EE8F,
-    0x8000EE9E8000EE99, 0x8000EEA88000EEA3, 0x8000EEB28000EEAD,
-    0x8000EEBC8000EEB7, 0x000165808000EEC1, 0x0001658200016D81,
-    0x0001658400017D83, 0x0001998600016985, 0x0001A98800019987,
-    0x0001B98A0001A989, 0x0001A98C0001B98B, 0x0001998E0001A98D,
-    0x0001E9900001998F, 0x0001F9920001E991, 0x0001E9940001F993,
-    0x000199960001E995, 0x0001A99800019997, 0x0001B99A0001A999,
-    0x0001A99C0001B99B, 0x0001999E0001A99D, 0x000169A00001999F,
-    0x000179A2000169A1, 0x000169A4000179A3, 0x000055A6000169A5,
-    0x000075A800004DA7, 0x000075AA00007DA9, 0x000055AC00004DAB,
-    0x000055AE00005DAD, 0x000035B000004DAF, 0x000035B200003DB1,
-    0x8000EECB8000EEC6, 0x8000EED58000EED0, 0x8000EEDF8000EEDA,
-    0x8000EEE98000EEE4, 0x8000EEF38000EEEE, 0x8000EEFD8000EEF8,
-    0x0002099E8000EF02, 0x0002099C0002099F, 0x0005ED9A0002099D,
-    0x0005ED9800F2AD9B, 0x0005CD960005E599, 0x0005C1940005D597,
-    0x0005D59202937D95, 0x0005E5900293D993, 0x00058D8E0005DD91,
-    0x0293998C00059D8F, 0x00052D8A0005E18D, 0x00052D880293998B,
-    0x8000EF0700052589, 0x0005058400051587, 0x000519820293D985,
-    0x8000EF0C0292F183, 0x8000EF1100051981, 0x8000EF1B8000EF16,
-    0x0002F1BA8000EF20, 0x0002C1B80003CCF0, 0x0002F1B60002C1B9,
-    0x0002E1B40002F1B7, 0x0002F1B20002E1B5, 0x000285B00002F1B3,
-    0x0002F9AE8000EF25, 0x0002F9AC0290B9AF, 0x0290E5AA0002F9AD,
-    0x00DA2DA80002F9AB, 0x0002C1A60002F5A9, 0x0002F5A40002C1A7,
-    0x0293D1A2029949A5, 0x0002E9A00002EDA3, 0x000260820002E9A1,
-    0x0002089100FD2099, 0x0002988400E16C9B, 0x0002988400019DA7,
-    0x0002E88C00020C91, 0x000195970002508D, 0x0001BDBE00025091,
-    0x8000EF2F0002A092, 0x0002E892000B208D, 0x000874B2000F84B1,
-    0x0002DC96003C00B9, 0x8000EF2A00006DB0, 0x00003DBD8000EF34,
-    0x003CACB900024C91, 0x0002B496003CACB8, 0x00000DAF8000EF39,
-    0x003C00A2000025BA, 0x000005AF00021C91, 0x8000EF438000EF3E,
-    0x003D68A48000EF48, 0x8000EF528000EF4D, 0x8000EF5C8000EF57,
-    0x8000EF668000EF61, 0x8000EF708000EF6B, 0x0002F4968000EF75,
-    0x8000EF7F8000EF7A, 0x8000EF898000EF84, 0x8000EF938000EF8E,
-    0x8000EF9D8000EF98, 0x8000EFA78000EFA2, 0x8000EFB18000EFAC,
-    0x8000EFB6001560A5, 0x003D20F08000EFBB, 0x8000EFC0003DC4F0,
-    0x8000EFCA8000EFC5, 0x8000EFD48000EFCF, 0x8000EFDE8000EFD9,
-    0x000240F00003CCB5, 0x0002C4A400028C91, 0x8000EFE300000DB0,
-    0x8000EFE80003B8B9, 0x8000EFF28000EFED, 0x8000EFFC8000EFF7,
-    0x8000F0068000F001, 0x8000F0108000F00B, 0x8000F01A8000F015,
-    0x8000F0248000F01F, 0x8000F02E8000F029, 0x8000F0388000F033,
-    0x8000F0428000F03D, 0x8000F047003DE4F0, 0x8000F0518000F04C,
-    0x8000F05B8000F056, 0x8000F0658000F060, 0x8000F06F8000F06A,
-    0x8000F0798000F074, 0x8000F0838000F07E, 0x8000F0918000F088,
-    0x8000F0A78000F09A, 0x8000F0C18000F0B4, 0x000005BB8000F0CE,
-    0x8000F0FE8000F0DB, 0x0002F5BF8000F0E4, 0x000015BD0002F5BE,
-    0x003E91A0000015BC, 0x003E95A2003E91A1, 0x003EB5A4003EADA3,
-    0x003EB5A6003EBDA5, 0x003E95A8003EADA7, 0x003E95AA003E9DA9,
-    0x003EF9AC003EE9AB, 0x003EE9AE003EF9AD, 0x003E99B0003EE9AF,
-    0x003EA9B2003E99B1, 0x003EB9B4003EA9B3, 0x003EA9B6003EB9B5,
-    0x003E9DB8003EA5B7, 0x000165BA00016DB9, 0x000165BC00017DBB,
-    0x000165BE00016DBD, 0x0001C58100015DBF, 0x0001C5830001C580,
-    0x0001C5850001C582, 0x0001C5870001C584, 0x0001C5890001C586,
-    0x0001C58B0001C588, 0x0001C58D0001C58A, 0x0001C58F0001C58C,
-    0x000145910001C58E, 0x0001599300014590, 0x0001499500014592,
-    0x0001799700014194, 0x0001499900014196, 0x0001599B00014198,
-    0x0001499D0001419A, 0x0001399F0001419C, 0x8000F1070001419E,
-    0x8000F1148000F0F1, 0x0003688500000083, 0x00023C878000095B,
-    0x80000AA500001086, 0x800020F600002D80, 0x0003B080800020EE,
-    0x8000F13F0010F481, 0x000540918000F144, 0x0007489300040090,
-    0x0009509500064C92, 0x000B589700080094, 0x000D6099000A5C96,
-    0x000F689B000C0098, 0x0012749D000E6C9A, 0x0015789F0011749C,
-    0x8000F1490014009E, 0x8000F1538000F14E, 0x8000F15D8000F158,
-    0x8000F1678000F162, 0x001320A48000F16C, 0x8000F1768000F171,
-    0x8000F1808000F17B, 0x8000F18A8000F185, 0x8000F1948000F18F,
-    0x8000F19E8000F199, 0x0016D0B88000F1A3, 0x001800BA0017DCB9,
-    0x8000F1AD8000F1A8, 0x8000F1B78000F1B2, 0x000000BC8000F1BC,
-    0x8000F1C1000000BD, 0x0003C59E8000F1C6, 0x8000F1CB0003CD9F,
-    0x8000F1D58000F1D0, 0x8000F1DF8000F1DA, 0x800016C18000F1E4,
-    0x800016C4800016FA, 0x80000351001BACAF, 0x800003618000034E,
-    0x02BF75B080000484, 0x02B891B202B891B1, 0x02B88DB402BF75B3,
-    0x02BE41B602BE59B5, 0x02BE4DB802BE7DB7, 0x02BE05BA02BE31B9,
-    0x02BEB1BC02BE11BB, 0x02BEA1BE02BEA5BD, 0x0018808402BEA1BF,
-    0x00020C86001B8485, 0x8000057680000523, 0x800005D7800005B1,
-    0x8000060B800005EE, 0x8000247380000622, 0x001AB0888000087E,
-    0x800007170012E489, 0x000000F080000841, 0x8000084A00002590,
-    0x8000087B8000080F, 0x8000079880000795, 0x800006B9800007BC,
-    0x000335A48000253D, 0x000325A280002710, 0x000335A000032DA3,
-    0x02AC319100032DA1, 0x02AC519302A3A190, 0x02AD759502AC4D92,
-    0x02ADA99702AC8194, 0x02A9B59902AD9196, 0x02A7E59B02ADE998,
-    0x02A0399D02A4199A, 0x02A0159F02A1C59C, 0x02A78D8102A02D9E,
-    0x02A0418302A47180, 0x02A07D8502A1BD82, 0x02A0618702A04584,
-    0x02A0D58902A06D86, 0x02A0858B02A03188, 0x02A1E18D02A0A98A,
-    0x02AC518F02A1498C, 0x02AD31B102A1858E, 0x02A91DB302B4C1B0,
-    0x02A75DB502B589B2, 0x02A091B702A4A1B4, 0x02A08DB902A16DB6,
-    0x02A091BB02A0B5B8, 0x02A005BD02A09DBA, 0x02A055BF02A0E1BC,
-    0x02B4F5A102A079BE, 0x02B4F5A302A0F5A0, 0x02A03DA50090C5A2,
-    0x02A149A702A011A4, 0x02ACC9A902A1E1A6, 0x02ACD9AB02A11DA8,
-    0x0090F1AD02A349AA, 0x02AD9DAF02ACB5AC, 0x02A03D8002AC69AE,
-    0x02A0F18202A09581, 0x02A2B58402AD2583, 0x02AD598602AD2585,
-    0x02ADB58802AD4587, 0x02ACA58A02AC4189, 0x02ACFD8C02AC9D8B,
-    0x02A50D8E02A8A18D, 0x02A0B19002A6F18F, 0x02A1599202A14D91,
-    0x02A1619402A16193, 0x0091519680002F5D, 0x02A1A59802B56197,
-    0x02A0559A02A18999, 0x02A0899C02A0FD9B, 0x02A2DD9E02AD5D9D,
-    0x02ADC1A002AD4D9F, 0x02AD1DA2009185A1, 0x02B5D5A402ACE9A3,
-    0x02B49DA602AC25A5, 0x02A595A802A809A7, 0x02A059AA02A669A9,
-    0x02A1A1AC02A1A5AB, 0x02A189AE02A199AD, 0x02A195B002A185AF,
-    0x02A10DB202A171B1, 0x02A0EDB402A121B3, 0x02A021B602A045B5,
-    0x02A245B802ADF5B7, 0x02ADA9BA02ADD5B9, 0x02AD65BC02ADB5BB,
-    0x02AC75BE02AC91BD, 0x02AFCD8002AC4DBF, 0x02A63D8202AB9181,
-    0x02A3E18402A5C183, 0x02A2098602A21D85, 0x02A2118802A23187,
-    0x0092218A02B61189, 0x02A2F58C02B6118B, 0x02A3058E02A2D98D,
-    0x02A3B99002A3AD8F, 0x02A1ED9202AE6D91, 0x02AE119402AE7D93,
-    0x02AECD9600925595, 0x02B6259802AF3997, 0x02B76D9A02AFD599,
-    0x02A6459C02ABF99B, 0x02A3899E80002DE2, 0x80002EFE02A2759F,
-    0x02A2B9A280003867, 0x800031FB80002C2F, 0x02A25DA602A221A5,
-    0x80002D36800033E7, 0x02A351AA02A335A9, 0x02A115AC02AE85AB,
-    0x02AEF9AE80003088, 0x02AE55B002AEE5AF, 0x02AF45B202AFA1B1,
-    0x02AF1DB402AF7DB3, 0x02A6EDB602AB41B5, 0x02A311B802A511B7,
-    0x80002CCE02A2EDB9, 0x800031D402A2C1BB, 0x0092F1BE02B6C1BD,
-    0x02A3C58002B6C1BF, 0x02A2358202A3E981, 0x80002D2202A29D83,
-    0x02A0BD8602AF3D85, 0x02AF618802AF2D87, 0x02AFBD8A00932589,
-    0x02B7758C02AE498B, 0x02B63D8E02AE858D, 0x02A7759002AAA98F,
-    0x02A2B99202A48991, 0x02A3419402A34593, 0x02A3699602A37995,
-    0x02A3359802A36597, 0x02A3AD9A02A3D199, 0x02A24D9C02A3819B,
-    0x02A2819E02A2E59D, 0x02A025A002AF559F, 0x02AFC9A202AFB5A1,
-    0x02AF05A402AFD5A3, 0x02AE15A602AEF1A5, 0x02AE6DA802AE2DA7,
-    0x02A79DAA02AA31A9, 0x02A241AC02A461AB, 0x02A3A9AE02A3BDAD,
-    0x02A3F1B002A391AF, 0x0093C1B202B7F1B1, 0x02A315B402B7F1B3,
-    0x02A2E5B602A339B5, 0x02A219B802A24DB7, 0x02A04DBA02AFCDB9,
-    0x02AFB1BC02AFDDBB, 0x02AF6DBE0093F5BD, 0x0000000102AE99BF,
-    0x02A0519702AFA194, 0x02AD819102A04D96, 0x02A0319302AD2990,
-    0x8000364F02ADE592, 0x02ABE59F02A1E99C, 0x02A1559902A8199E,
-    0x02A1899B02A0A198, 0x02AB8D8502A1B19A, 0x02AC418702A87184,
-    0x02A1E181800038C3, 0x02A5CD8302B81180, 0x02ACD58D02B95982,
-    0x02AC858F80002EE2, 0x02AC5D8902ACA98E, 0x80002E1A02AC6588,
-    0x02A131B502AC4D8A, 0x02A51DB702B8C1B4, 0x009C91B102B989B6,
-    0x02A1FDB302A0D5B0, 0x02AC8DBD02A009B2, 0x8000339B02ACB5BC,
-    0x02AB7DB902AC9DBE, 0x02ACB1BB02A881B8, 0x02B8F5A580003138,
-    0x02B8F5A702ACF5A4, 0x02ACD9A1009CC5A6, 0x02ACF5A302AD25A0,
-    0x02A0C9AD02ACCDA2, 0x02A0D9AF02AD1DAC, 0x02AC1DA902AF49AE,
-    0x02AD69AB02AC31A8, 0x02AD558002ADC1AA, 0x02ADCD8202ADB181,
-    0x02AC2D8402ADE183, 0x02ACE18680003164, 0x8000315880003943,
-    0x02A1698A80003130, 0x02A1A58C02A1758B, 0x02A0B58E02A0518D,
-    0x02A08D9002A08D8F, 0x02A97D92800035A7, 0x02ACA19402AA8193,
-    0x02AD499602AD5D95, 0x02AD519802AD7197, 0x009D619A02B95199,
-    0x02ADB59C02B9519B, 0x02AC459E02AD999D, 0x02AC79A002ACED9F,
-    0x02AE2DA202A1ADA1, 0x02A1D1A402A1BDA3, 0x02A10DA6009D95A5,
-    0x02B9E5A802A0F9A7, 0x02B8ADAA02A015A9, 0x02A985AC02A439AB,
-    0x02AC49AE02AA79AD, 0x02ADD1B002ADB5AF, 0x02ADF9B202ADE9B1,
-    0x02AD85B402ADF5B3, 0x02AD1DB602AD61B5, 0x02ACDDB802AD31B7,
-    0x02AC11BA02AC75B9, 0x02AE55BC02A1C5BB, 0x02A1B9BE02A1C5BD,
-    0x02A2958002A1A5BF, 0x02A3858202A36181, 0x02A3DD8402A3BD83,
-    0x02AA2D8602A78185, 0x02AFD18802A9D187, 0x02AE398A02AE2D89,
-    0x02AE018C02AE018B, 0x009E318E02BA018D, 0x02AE859002BA018F,
-    0x02AF759202AEA991, 0x02AFA99402AFDD93, 0x02ADFD9602A27D95,
-    0x02A2219802A26D97, 0x02A2FD9A009E6599, 0x02BA359C02A3099B,
-    0x02BB7D9E02A3C59D, 0x02AAB5A002A7E99F, 0x02AF79A202A949A1,
-    0x02AE81A402AE85A3, 0x02AEA9A602AEB9A5, 0x02AEF5A802AEA5A7,
-    0x02AE6DAA02AE11A9, 0x02AF8DAC02AE41AB, 0x02AF41AE02AF25AD,
-    0x02AD65B002A295AF, 0x02A289B202A2F5B1, 0x02A245B402A295B3,
-    0x02A355B602A3B1B5, 0x02A32DB802A36DB7, 0x02AADDBA02A771B9,
-    0x02AF01BC02A921BB, 0x02AEE9BE02AEFDBD, 0x02AF318002AED1BF,
-    0x009F018202BB3181, 0x02AFD58402BB3183, 0x02AE258602AFF985,
-    0x02AED98802AE8D87, 0x02AC8D8A02A30D89, 0x02A3718C02A31D8B,
-    0x02A3AD8E009F358D, 0x02BB059002A2598F, 0x02BA4D9202A2F591,
-    0x02AB659402A6D993, 0x02AEA99602A89995, 0x02AF719802AF5597,
-    0x02AF599A02AF4999, 0x02AF259C02AF559B, 0x02AFBD9E02AFC19D,
-    0x02AEBDA002AF919F, 0x02AE71A202AE15A1, 0x02AC35A402A3A5A3,
-    0x02A3D9A602A3A5A5, 0x02A335A802A3C5A7, 0x02A225AA02A2C1A9,
-    0x02A27DAC02A21DAB, 0x02AB8DAE02A621AD, 0x02AE31B002A871AF,
-    0x02AFD9B202AFCDB1, 0x02AFE1B402AFE1B3, 0x009FD1B602BBE1B5,
-    0x02AF25B802BBE1B7, 0x02AED5BA02AF09B9, 0x02AE09BC02AE7DBB,
-    0x02AC5DBE02A3DDBD, 0x02A4319902A3CDBF, 0x02A4519B02ABA198,
-    0x02A5759D02A44D9A, 0x02A5A99F02A4819C, 0x02A8959102A5919E,
-    0x02A8C59302A87190, 0x02A9A19502A8E992, 0x02A4119702A90994,
-    0x02AF8D8902A9C596, 0x02A8418B02AC7188, 0x02A87D8D02A9BD8A,
-    0x02A8618F02A8458C, 0x0098618102A86D8E, 0x02A50D8302A42580,
-    0x02A5C18502A4F982, 0x02A1ED8702BC3184, 0x02A531B902BD7986,
-    0x02A11DBB02BCC1B8, 0x02AF5DBD02BD89BA, 0x02A891BF02ACA1BC,
-    0x02A489B102A96DBE, 0x02A499B302A95DB0, 0x0098B1B502AB09B2,
-    0x02A5DDB702A4F5B4, 0x02BCF5A902A429B6, 0x02BCF5AB02A8F5A8,
-    0x02A83DAD0098C5AA, 0x02A949AF02A811AC, 0x02A175A102A9E1AE,
-    0x02AF25A302A529A0, 0x02A8F9A502ACD9A2, 0x02A8D5A702A905A4,
-    0x02A9118002A8EDA6, 0x02A9398202A92981, 0x02A9458402A93583,
-    0x02A9DD8602A9A185, 0x02A81D8802A9F187, 0x02A8D18A02A8B589,
-    0x02AA958C02A5058B, 0x02A5798E02A5058D, 0x02A5D59002A5658F,
-    0x02A4C59202A42191, 0x02A49D9402A4FD93, 0x02AD6D9602A0C195,
-    0x02A8919802AE9197, 0x02A9799A02A96D99, 0x02A9419C02A9419B,
-    0x0099719E02BD419D, 0x02A945A002BD419F, 0x02A8B5A202A969A1,
-    0x02A869A402A81DA3, 0x02AA3DA602A5BDA5, 0x02A5E1A802A5ADA7,
-    0x02A53DAA0099A5A9, 0x02BDF5AC02A4C9AB, 0x02BCBDAE02A405AD,
-    0x02ADF5B002A029AF, 0x02A839B202AE09B1, 0x02A9C1B402A9C5B3,
-    0x02A9E9B602A9F9B5, 0x02A9B5B802A9E5B7, 0x02A92DBA02A951B9,
-    0x02A8CDBC02A901BB, 0x02A801BE02A865BD, 0x02A9A58002A5D5BF,
-    0x02A6498202A63581, 0x02A6858402A65583, 0x02A7958602A77185,
-    0x02A7ED8802A7AD87, 0x02AE1D8A02A3B189, 0x02ABC18C02ADE18B,
-    0x02AA298E02AA3D8D, 0x02AA719002AA118F, 0x009A419202BE7191,
-    0x02AA959402BE7193, 0x02AB659602AAB995, 0x02AB999802ABCD97,
-    0x02A9CD9A02A64D99, 0x02A6319C02A65D9B, 0x02A6ED9E009A759D,
-    0x02BEC5A002A7199F, 0x02BF8DA202A735A1, 0x000C85A402A319A3,
-    0x800035E3000CB5A5, 0x02BFF9A880002E06, 0x02BFB5AA02BFB1A9,
-    0x02BA79AC02BF95AB, 0x02BBA1AE02BF35AD, 0x02BB49B002BC3DAF,
-    0x02BDA9B202BD9DB1, 0x02BDB9B402A561B3, 0x02B6D5B602BDA9B5,
-    0x02BAEDB802BD69B7, 0x02BD3DBA02BC65B9, 0x02B6B9BC02BD09BB,
-    0x02BA01BE02BA5DBD, 0x02B7758002BBDDBF, 0x001C0182000D2D81,
-    0x02BE0D8402BE1983, 0x02BBD18602BE2D85, 0x001F218802BE9D87,
-    0x0097718A02BDA589, 0x02BC518C02BC758B, 0x02BC518E02A4998D,
-    0x02B74D9002BC418F, 0x02B3519202BCF191, 0x02BC859402BC9993,
-    0x001F519602BCB195, 0x02BB999802BBF597, 0x001F799A02BA4599,
-    0x02BBB99C000D419B, 0x02BC219E02BDF59D, 0x02B399A002BBD59F,
-    0x02BED1A202BC31A1, 0x02BE8DA402BE99A3, 0x02BB51A602BEADA5,
-    0x02BAB9A802BE1DA7, 0x02BA21AA02BD25A9, 0x02BCD1AC02BCF5AB,
-    0x02BCD1AE02A419AD, 0x02B7CDB002BCC1AF, 0x02BBC5B202BC71B1,
-    0x02BC05B402BD4DB3, 0x02B791B602BC31B5, 0x02BB19B802BB75B7,
-    0x02B79DBA02BAC5B9, 0x001CF9BC000DC5BB, 0x02BEE5BE02BEE1BD,
-    0x80000C0402BEC5BF, 0x80000C4B0000000F, 0x80000C578000337B,
-    0x80002D3A800034CF, 0x8000397B80002E66, 0x8000247B800024B7,
-    0x8000249980002485, 0x02AB18CC800024C1, 0x02ADB9A002AB7CCD,
-    0x02AE4DA202AE45A1, 0x02AE5DA402902DA3, 0x029059A602AE75A5,
-    0x02AC4DA802AEB1A7, 0x02ABA1AA02ABEDA9, 0x00DACDAC02AF01AB,
-    0x800024A302AF09AD, 0x02AF4990800024CB, 0x02AB5D9202ACB191,
-    0x02ACA99402AB5593, 0x02AD959602AB4995, 0x02AD819802AD8597,
-    0x02ADFD9A02AD8999, 0x02AD719C02ADC19B, 0x02AD659E02AD659D,
-    0x02AD218002AD559F, 0x02AD318202AD2D81, 0x02AED58402AECD83,
-    0x02AED5860290B585, 0x0290E5880290DD87, 0x00DBCD8A02AED189,
-    0x02ACDD8C0290F98B, 0x0290D98E02AB658D, 0x02A5C98002AF898F,
-    0x0001098202A08581, 0x0089498402A38D83, 0x02A2798602A24D85,
-    0x02A2498802BAB187, 0x02A9258A02A25989, 0x02AD298C02A2998B,
-    0x02A2ED8E02A2E18D, 0x0001499002A2D98F, 0x02A5B19202A5ED91,
-    0x0001419402A46D93, 0x02A5919600137995, 0x02A2399802A3DD97,
-    0x02AD719A02A5CD99, 0x02A0299C02A2D99B, 0x02A0659E02A0619D,
-    0x02A549A002A0459F, 0x02A491A202A005A1, 0x02A419A402A30DA3,
-    0x02A2F9A602A2CDA5, 0x02A2C9A802BA31A7, 0x02A9A5AA02A2D9A9,
-    0x02A5BDAC02A219AB, 0x02A26DAE02A335AD, 0x02A989B002A259AF,
-    0x02A531B202A56DB1, 0x02A9A5B402A4EDB3, 0x0002D1B60013FDB5,
-    0x02A0FDB802A0C9B7, 0x02A521BA02A0DDB9, 0x0001F1BC02A06DBB,
-    0x0089A1BE02A375BD, 0x02B9AD8102A2A5BF, 0x02A1758302A16580,
-    0x02A1A58502A16582, 0x02A1CD8702AA1984, 0x02A1C58902AE0586,
-    0x02A6818B02A1F188, 0x02A7118D0002258A, 0x0010158F02A6CD8C,
-    0x02A0C19100022D8E, 0x02A6E19302A68D90, 0x02A1E59502A11592,
-    0x02A34D9702AE4D94, 0x02A3599902A30596, 0x02A3E99B02A37998,
-    0x02A0F19D02A6A59A, 0x02A1219F02A76D9C, 0x02B92DA102A7F59E,
-    0x02A1F5A302A1E5A0, 0x02A125A502A1E5A2, 0x02A019A702AA99A4,
-    0x02A145A902A691A6, 0x02A601AB02A171A8, 0x02A791AD02AAE5AA,
-    0x001091AF02A64DAC, 0x02A3D5B102AAC9AE, 0x02A3F1B30001CDB0,
-    0x02A351B502A3D1B2, 0x02A059B702A61DB4, 0x02A1B9B90002DDB6,
-    0x02B945BB008ABDB8, 0x02A18DBD02A18DBA, 0x02A14DBF02A19DBC,
-    0x001019A002AAF1BE, 0x80000BF2001019A1, 0x00000D908000248F,
-    0x800024AD00003D91, 0x02ABA8CC8000308C, 0x80002DF602ABDCCD,
-    0x02A8B0CC80002C03, 0x02ABB8CC02A8ACCD, 0x0003758402AB8CCD,
-    0x0003358600033D85, 0x0003758000004187, 0x0003758200037D81,
-    0x80002BFB00036D83, 0x8000F1E980002C0B, 0x0000698880002C56,
-    0x80002C7280002C6A, 0x0011B98080002C7A, 0x001109820011B981,
-    0x0011198400110983, 0x0011098600111985, 0x0011398800110987,
-    0x00110D8A00113989, 0x00111D8C0011058B, 0x00110D8E0011058D,
-    0x00117D900011058F, 0x00110D9200110591, 0x00111D9400110593,
-    0x0011399600110995, 0x0011099800113997, 0x0011119A00110999,
-    0x0010F99C0011119B, 0x0010E99E0010F99D, 0x02A7B59F0010E99F,
-    0x02A7759D02AC099E, 0x02BFBD9B02A7659C, 0x02A7419902A7759A,
-    0x02A6A19702A19598, 0x02A5A99502A13D96, 0x02A5099302A0E594,
-    0x02A52D9102A52992, 0x02A7F58F02A56590, 0x02A0E18D02A85D8E,
-    0x02A6D18B02A7158C, 0x0016758902A09D8A, 0x02A1418700044D88,
-    0x02A0C18502A09D86, 0x02A7958300046584, 0x02A7AD8102A7A182,
-    0x02A029BF02A86580, 0x02A76DBD00048DBE, 0x02A745BB02A759BC,
-    0x02A72DB902A88DBA, 0x02A7DDB702AC91B8, 0x02BF05B502A7CDB6,
-    0x02A7E9B302A7CDB4, 0x02A639B1008CEDB2, 0x02A541AF0004BDB0,
-    0x02A5F1AD02A00DAE, 0x02A5C5AB02A5D1AC, 0x0016F1A90007DDAA,
-    0x02A1C1A702ACA9A8, 0x02A041A502A01DA6, 0x02A715A302ACA5A4,
-    0x02A679A102A721A2, 0x02A1F98002A0F1A0, 0x0005198202A02581,
-    0x02A1D98400172183, 0x02A6418602A79585, 0x02A9398802A1B587,
-    0x02AFF58A02A69189, 0x80002C7E02AFF18B, 0x008C098E80002C82,
-    0x02BEC99002BEE98F, 0x008C7D9202BEC991, 0x008C6D94008C7593,
-    0x008C1196008C1595, 0x008C5198008C0D97, 0x02BEE19A02BEB199,
-    0x008C459C02BEE19B, 0x008C459E008C4D9D, 0x008CC9A0008C3D9F,
-    0x008CB9A2008CD5A1, 0x02BE19A402BE59A3, 0x008CADA602BE19A5,
-    0x008C9DA8008CA5A7, 0x008CE1AA008CE5A9, 0x008C81AC008CFDAB,
-    0x02BE31AE02BE61AD, 0x008CF5B002BE31AF, 0x008CF5B2008CFDB1,
-    0x008C99B4008C8DB3, 0x008CE9B6008C85B5, 0x02BE69B802BE09B7,
-    0x008CDDBA02BE69B9, 0x008CCDBC008CD5BB, 0x008CB1BE008CB5BD,
-    0x0026B980008CADBF, 0x00244D820026B981, 0x80002ABB00245D83,
-    0x0024458600221585, 0x0024518800245587, 0x0024A18A00222189,
-    0x0024A98C0024A18B, 0x0024598E0024A98D, 0x0024C1900024458F,
-    0x0024CD9200242D91, 0x0024359400241593, 0x0024199600268195,
-    0x0024359800241197, 0x00241D9A00240999, 0x0015519C0024119B,
-    0x00155D9E0037F59D, 0x8000F1FD0015559F, 0x0024CDA2002639A1,
-    0x002451A480002CA2, 0x80002CC280002CA6, 0x8000F1EE0024D5A7,
-    0x002421AA0022A1A9, 0x002429AC002421AB, 0x0024D9AE002429AD,
-    0x002441B00024C5AF, 0x00244DB20024ADB1, 0x0024B5B480002ABF,
-    0x002499B6002601B5, 0x80002CAE002491B7, 0x8000F1F8002489B9,
-    0x80002CDE002491BB, 0x80002C3F80002CE6, 0x0025458280002CEE,
-    0x80002D168000F3D3, 0x80002D2E80002AC7, 0x80002D3200255D87,
-    0x80002ACB80002D3E, 0x80002D760025A98B, 0x0025518E00232989,
-    0x80002D7A00254D8F, 0x0025C5920025A18D, 0x80002DAE80002DA2,
-    0x80002DBE00252591, 0x00253D9400251997, 0x80002DCA80002DC6,
-    0x8000F20C0025199B, 0x800033CB00250199, 0x80002DDE00145D9F,
-    0x0025C5A20036FD9D, 0x8000F21180002DE6, 0x80002DF2002731A1,
-    0x002559A40025DDA7, 0x002529AA80002DFA, 0x0025D9A880002E02,
-    0x0025D1AE0023A9A9, 0x002521AC0025CDAF, 0x002545B2002521AD,
-    0x002549B080002E2A, 0x002591B60025A5B1, 0x0025BDB4002599B7,
-    0x002595BA002709B5, 0x80002E16002599BB, 0x0014D5BE002581B9,
-    0x0014D9BC80002E22, 0x002A41B680002E26, 0x002A6DB4002A49B7,
-    0x002A95B20028D9B5, 0x002A99B0002A4DB3, 0x80002E2E002A75B1,
-    0x80002E3A80002E36, 0x002A45BA80002E42, 0x002A6DB8002A49BB,
-    0x002A1DA6002A51B9, 0x80002E4A002A0DA7, 0x002A15A2002C4DA5,
-    0x80002E4E002A05A3, 0x002A01AE0028E1A1, 0x002AF1AC002A1DAF,
-    0x80002E5E002AF1AD, 0x002A09A8002AF9AB, 0x002AC196002C79A9,
-    0x002AED94002AC997, 0x002A159200285995, 0x002A1990002ACD93,
-    0x80002E62002AF591, 0x80002E9280002E6E, 0x002AC59A80002E96,
-    0x002AED98002AC99B, 0x002A9D86002AD199, 0x002A0984002A8D87,
-    0x002A9582002CCD85, 0x00286180002A8583, 0x002A818E00286181,
-    0x002A718C002A9D8F, 0x80002E7A002A718D, 0x002A8988002A798B,
-    0x001A3180002CF989, 0x001A3182001A3181, 0x001A9D84001A3183,
-    0x001A8D86001A8585, 0x001ABD88001A8587, 0x001A8D8A001A8589,
-    0x80002EDE80002E82, 0x80002EAA80002EA6, 0x00F8199080002EAE,
-    0x00F8259200F81991, 0x00F8259400F83D93, 0x00F8259600F82D95,
-    0x00F8F19800F8D997, 0x00F8C99A00F8F199, 0x00F8D99C00F8C99B,
-    0x00F8C99E00F8D99D, 0x00F839A000F8C99F, 0x00F861A200F839A1,
-    0x00F861A400F861A3, 0x00FB19A600F861A5, 0x00FB29A800FB19A7,
-    0x001A01AA00FB29A9, 0x02B181AC02BC5DAB, 0x001A15AE00DE09AD,
-    0x02B9F5B080002EA2, 0x02BC39B202BA09B1, 0x02BDC1B402BDC5B3,
-    0x02BDE9B602BDF9B5, 0x02BDB5B802BDE5B7, 0x02BD2DBA02BD51B9,
-    0x02BCCDBC02BD01BB, 0x02BC01BE02BC65BD, 0x02BDA58002B1D5BF,
-    0x02B2498202B23581, 0x02B2858402B25583, 0x02B3958602B37185,
-    0x02B3ED8802B3AD87, 0x0019918A02B7B189, 0x00198D8C00DDA58B,
-    0x0019F58E0019FD8D, 0x008371B00019ED8F, 0x02B1C1B202B191B1,
-    0x008365B402B1C1B3, 0x008365B600836DB5, 0x008329B800831DB7,
-    0x80002EB6008335B9, 0x80002ECA80002EB2, 0x80002EE68000F216,
-    0x000580A180002EEA, 0x007800A3000400A0, 0x007C00A500068CA2,
-    0x007000A7007A94A4, 0x80002EF2007E9CA6, 0x00197DAA007124A8,
-    0x00196DAC001975AB, 0x80002F2980002F06, 0x8000F2208000F21B,
-    0x80002F3980002F35, 0x80002F3180002F3D, 0x000000B480002F45,
-    0x80002AD380002ACF, 0x80002F5580002F4D, 0x80002F718000F225,
-    0x80002F8180002F75, 0x0018E5808000F22A, 0x00DA45820018E981,
-    0x80002F8D80002F89, 0x80002AD78000306C, 0x80002FBD80002FA1,
-    0x8000F23480002FB9, 0x8000F23980002FC5, 0x80002FCD80002FD1,
-    0x00F10D9080002FD5, 0x001D3D92001E1191, 0x00F3759400EF3D93,
-    0x001DDD96001DD595, 0x001DED9800E29D97, 0x001DF59A001DF599,
-    0x001DD19C00CCE59B, 0x001D919E001DC19D, 0x001C91A0001D8D9F,
-    0x00F155A20013A5A1, 0x0013F5A40013C1A3, 0x001331A60013EDA5,
-    0x0012B5A8001321A7, 0x00E209AA00FD5DA9, 0x00E5CDAC0012B1AB,
-    0x00E539AE00E219AD, 0x00F681B000129DAF, 0x00128DB20012E1B1,
-    0x00154DB4001555B3, 0x00E5E1B6000BD9B5, 0x000BF5B800F1A9B7,
-    0x000BF5BA000BFDB9, 0x80002FE5000BCDBB, 0x80002ADB80002FE9,
-    0x000B19A180002FF9, 0x000B15A3000B15A0, 0x00C7DDA5000C15A2,
-    0x00C549A7000B35A4, 0x000B09A900C551A6, 0x00C801AB00C7F5A8,
-    0x00C54DAD000B15AA, 0x000B69AF000B71AC, 0x000B21B1000B75AE,
-    0x000B29B3000B19B0, 0x000B21B5000B2DB2, 0x000BDDB7000B3DB4,
-    0x000BF1B9000B25B6, 0x000BCDBB000BEDB8, 0x000B99BD000BF1BA,
-    0x000B85BF000BB1BC, 0x000FE181000B89BE, 0x000FE183000FA980,
-    0x00C5A985000FE982, 0x000FE187000FF984, 0x00099189000FE986,
-    0x00099D8B000FD988, 0x000A318D00C5918A, 0x000A358F000AA58C,
-    0x000A599100C5F98E, 0x000A0593000A5190, 0x001A3995000A6D92,
-    0x000A2997000A3594, 0x000A1999000A3596, 0x00C7519B000A1198,
-    0x00F4859D000A119A, 0x000BF59F000BF19C, 0x00052180000BF59E,
-    0x0005218200052181, 0x0005598400052183, 0x0005418600054D85,
-    0x0005518800C6C587, 0x00058D8A00058589, 0x00C4BD8C0005858B,
-    0x0005C98E0005998D, 0x00C539900004118F, 0x00047D9200047591,
-    0x00046D9400047593, 0x000731960004E995, 0x0007119800073197,
-    0x0007119A00071199, 0x0007D19C0007119B, 0x0007CD9E001BF59D,
-    0x00073DA00007C59F, 0x00074DA2000745A1, 0x00075DA4000745A3,
-    0x00074DA6000745A5, 0x000741A8000741A7, 0x000741AA000741A9,
-    0x00075DAC000745AB, 0x00074DAE000745AD, 0x000739B000C9A5AF,
-    0x0006C5B2000739B1, 0x0006C5B40006DDB3, 0x0006C1B600F521B5,
-    0x00F51DB80006C1B7, 0x0006CDBA0006C5B9, 0x0006DDBC0006C5BB,
-    0x0006CDBE0006C5BD, 0x00053D800006C5BF, 0x0005FD820005F581,
-    0x0005ED840005F583, 0x0001118600011185, 0x0001318800011187,
-    0x0001318A00013189, 0x0001118C0001318B, 0x00CD7D8E0001118D,
-    0x00F469900001158F, 0x0001659200016991, 0x0001659400017D93,
-    0x0001619600016195, 0x0001219800016197, 0x0001219A00012199,
-    0x0001219C0001219B, 0x0001219E0001219D, 0x0001E1A00001219F,
-    0x0001E1A20001E1A1, 0x0001FDA400C589A3, 0x0001EDA60001E5A5,
-    0x00C3A9A80001E5A7, 0x0001D5AA0001DDA9, 0x000131AC00C5A1AB,
-    0x000131AE000131AD, 0x00C5D1B0000131AF, 0x000159B200CEADB1,
-    0x000149B4000159B3, 0x000179B6000149B5, 0x000149B8000179B7,
-    0x000159BA000149B9, 0x000149BC000159BB, 0x000139BE000149BD,
-    0x000000C1000139BF, 0x8000300580003001, 0x80002ADF80003009,
-    0x80002AE38000F23E, 0x800030308000302C, 0x8000F40580003034,
-    0x8000370F8000F248, 0x80002AE780003064, 0x8000F33D8000F26B,
-    0x8000307C80003074, 0x8000309080002AEB, 0x800030A4800030A0,
-    0x800030BC800030AC, 0x80002AEF80002AF3, 0x800030C48000F24D,
-    0x800030DC800030D8, 0x80003100800030EC, 0x800030FC800030F8,
-    0x8000311880003104, 0x800031348000311C, 0x8000314C8000313C,
-    0x8000F25C80003160, 0x800031508000315C, 0x8000317080003168,
-    0x8000F26180003184, 0x800031808000318C, 0x80002AF780003178,
-    0x8000319C80003194, 0x80003198800031A4, 0x800031BC80002AFB,
-    0x800031E88000F266, 0x800032068000322A, 0x00BB21A280002B07,
-    0x000131A0000145A3, 0x00015DA6000131A1, 0x00015DA400BDC1A7,
-    0x000179AA000155A5, 0x000179A8000179AB, 0x000159AE000179A9,
-    0x000159AC000159AF, 0x000139B2000159AD, 0x000139B0000139B3,
-    0x0001D5B6000139B1, 0x0001D9B40001DDB7, 0x0001F5BA00B401B5,
-    0x0001E5B80001FDBB, 0x0001D5BE0001FDB9, 0x0001E5BC0001DDBF,
-    0x007F51820001DDBD, 0x007F4180007F5183, 0x007F7186007F4181,
-    0x007F4184007F7187, 0x007F4D8A007F4185, 0x00BBBD88007F558B,
-    0x007F4D8E007F4589, 0x007F4D8C00BBA98F, 0x0001C992007F458D,
-    0x0001C9900001C993, 0x0001C9960001C991, 0x0001C9940001C997,
-    0x0001C19A0001C995, 0x00B3A1980001C19B, 0x0001C19E00BDC599,
-    0x0001D19C0001C19F, 0x80002B030001D19D, 0x80002CD280002AFF,
-    0x8000322E80002CD6, 0x800036E380003216, 0x8000324280002B6B,
-    0x8000324E80003256, 0x80002B0B8000F27F, 0x8000325E8000326A,
-    0x8000F28980003287, 0x8000327280003293, 0x8000329F8000329B,
-    0x800032A7800032A3, 0x8000F28E800032B3, 0x80002B13800032BB,
-    0x8000F293800032C7, 0x80002B17800032CB, 0x800032E3800032D7,
-    0x8000F298800032F3, 0x8000F29D8000F22F, 0x8000F2A280003313,
-    0x800033278000331F, 0x8000333380003317, 0x8000334B80003343,
-    0x8000335B80003357, 0x8000333F8000F2A7, 0x8000337780003373,
-    0x8000338780002B1B, 0x8000336B8000F2AC, 0x8000F2B1800033A7,
-    0x800033AB8000F2B6, 0x800033B7800033BF, 0x800033C380002B1F,
-    0x0003AD80800033D3, 0x00B5B98200B92181, 0x0003A5840003BD83,
-    0x0003A5860003AD85, 0x0003E5880003DD87, 0x0003E58A0003ED89,
-    0x0003E18C00BF198B, 0x0003E18E0003E18D, 0x0003A1900003E18F,
-    0x0003A1920003A191, 0x00BF69940003A193, 0x0003AD960003A595,
-    0x00039D980003A597, 0x00016D9A00016599, 0x00017D9C0001659B,
-    0x00016D9E0001659D, 0x00019DA00001659F, 0x0001A9A200BFC5A1,
-    0x0001B9A40001A9A3, 0x0001A9A60001B9A5, 0x000199A80001A9A7,
-    0x0001E9AA000199A9, 0x0001F9AC0001E9AB, 0x0001F5AE0001FDAD,
-    0x00BF89B00001EDAF, 0x000191B2000191B1, 0x0001B1B4000191B3,
-    0x0001B1B60001B1B5, 0x00B6C5B80001B1B7, 0x00019DBA000195B9,
-    0x00018DBC000195BB, 0x00017DBE000175BD, 0x800033CF000175BF,
-    0x8000F1F3800033DB, 0x8000F2BB800033EB, 0x8000F2C5800033F7,
-    0x8000342B8000340B, 0x8000343F8000F2CA, 0x8000F2CF80003443,
-    0x8000345F8000F2D4, 0x80002B238000346F, 0x80002B2780003473,
-    0x8000348F8000348B, 0x800034A78000349B, 0x800034BF80002B2B,
-    0x800034DB8000F2D9, 0x8000F2438000F2DE, 0x8000F2E3800034F7,
-    0x8000F2ED8000F2E8, 0x80002B3380002B2F, 0x8000F2F28000F2F7,
-    0x8000F3018000F2FC, 0x8000353B80003537, 0x80002B3F80003547,
-    0x8000F30B80002B43, 0x80002B478000355B, 0x8000F3158000F310,
-    0x800035CB8000F31A, 0x800035D780002B4B, 0x8000F31F800035DB,
-    0x800035F38000F324, 0x8000F32980002B4F, 0x8000360B80003607,
-    0x8000F32E80002B53, 0x0007F5A480003627, 0x0007F5A60007FDA5,
-    0x0007F5A000010DA7, 0x0007F5A20007FDA1, 0x000135AC0007EDA3,
-    0x000135AE00013DAD, 0x000135A800010DAF, 0x000135AA00013DA9,
-    0x000175B400012DAB, 0x000175B600017DB5, 0x000175B000010DB7,
-    0x000175B200017DB1, 0x000131BC00016DB3, 0x000131BE000131BD,
-    0x000135B8000131BF, 0x00BF09BA00013DB9, 0x00071984000131BB,
-    0x0007298600071985, 0x00B9F18000072D87, 0x0007098200071981,
-    0x00070D8C00070983, 0x00071D8E0007158D, 0x00B9D58800B7218F,
-    0x00071D8A00B96989, 0x000701940007158B, 0x00BC059600070195,
-    0x00B9BD9000070597, 0x0007659200076D91, 0x0007399C00BF9193,
-    0x0007359E0007399D, 0x00073D9800070D9F, 0x0007299A00BFAD99,
-    0x80002B570007299B, 0x8000362F80003637, 0x8000F3388000363F,
-    0x80002B5B80003657, 0x8000366B80003663, 0x80002B5F8000367B,
-    0x8000F3478000F342, 0x8000F34C80002B63, 0x8000F35180003693,
-    0x800036AB8000369F, 0x8000F35B8000F356, 0x8000F360800036C7,
-    0x8000F270800036CF, 0x800036E780002B67, 0x80002B6F800036EB,
-    0x80002F49800036EF, 0x8000F36A8000F365, 0x8000F27A8000F275,
-    0x8000392B80003713, 0x8000373F80002B73, 0x800037438000373B,
-    0x8000374780002D4E, 0x8000374F8000374B, 0x8000F36F80003757,
-    0x8000376F8000375B, 0x8000375F80003773, 0x8000378B80003777,
-    0x8000377F8000376B, 0x8000378780003783, 0x8000F37E8000F374,
-    0x80002B778000F379, 0x800037AB800037A7, 0x00036180800037B3,
-    0x0003618200036181, 0x0003618400036183, 0x0003618600036185,
-    0x0003A18800036187, 0x0003A18A0003A189, 0x0003BD8C00791D8B,
-    0x0003AD8E0003A58D, 0x0003DD900003A58F, 0x0003ED920003E591,
-    0x0003FD940003E593, 0x0003ED960003E595, 0x00B209980003E597,
-    0x0003A99A0003D999, 0x0003B99C0003A99B, 0x0003A99E0003B99D,
-    0x000355A000BD019F, 0x000355A200035DA1, 0x000375A400036DA3,
-    0x000375A600037DA5, 0x000355A800036DA7, 0x000355AA00035DA9,
-    0x0001B5AC0001ADAB, 0x0001B5AE0001BDAD, 0x0001D5B00001ADAF,
-    0x0001D5B20001DDB1, 0x00B5A9B40001EDB3, 0x0001F1B60001F1B5,
-    0x0001D1B80001F1B7, 0x0001D1BA0001D1B9, 0x0001CDBC00BB65BB,
-    0x0001BDBE0001B5BD, 0x8000F3920001B5BF, 0x8000F383800037B7,
-    0x80002B7F80002B7B, 0x8000F38D8000F388, 0x800037D780002B83,
-    0x800037E3800037DF, 0x800037EB800037EF, 0x800037F3800037F7,
-    0x800037FF800037FB, 0x80002B8780003807, 0x800038138000380F,
-    0x8000382380002B8B, 0x800038338000F397, 0x80002B8F80003837,
-    0x80002AC380003843, 0x8000F3A18000F39C, 0x80002B9780002B93,
-    0x8000F3A68000386F, 0x800038CB800038C7, 0x800038EF800038E3,
-    0x8000F2028000F3B0, 0x800038F7800038FB, 0x8000F20780003903,
-    0x8000F3B58000390F, 0x8000395B8000F3BA, 0x8000396F80003967,
-    0x800039738000F3BF, 0x800039AB8000399F, 0x800039A7800039AF,
-    0x800039BF800039BB, 0x800039CB8000F3C4, 0x800039D380002B9B,
-    0x000FD1A68000F3C9, 0x000FC1A4000FD1A7, 0x000FD1A2000FC1A5,
-    0x000FA1A0000FD1A3, 0x000FD1AE000FA1A1, 0x000FC1AC000FD1AF,
-    0x000FD1AA000FC1AD, 0x000FE1A8000FD1AB, 0x000151B6000FE1A9,
-    0x000141B4000151B7, 0x000151B2000141B5, 0x000FA1B0000151B3,
-    0x00B5A1BE000FA1B1, 0x000141BC00B5A5BF, 0x000151BA000141BD,
-    0x000161B8000151BB, 0x000FE186000161B9, 0x00859D84000FE187,
-    0x000FC582000FD185, 0x000FD580000FCD83, 0x000F218E000FCD81,
-    0x000FD18C000F218F, 0x000FC18A000FD18D, 0x000FD188000FC18B,
-    0x000F5596000FD189, 0x000F4594000F5D97, 0x0085E192000F5D95,
-    0x00B54590000F5D93, 0x000F559E00B57991, 0x000F459C00B54D9F,
-    0x000F559A000F5D9D, 0x000F6598000F5D9B, 0x80002B9F000F5D99,
-    0x80002FD980003A0B, 0x8000F3CE80003A23, 0x80002BA38000F3D8,
-    0x80003A4B80002BA7, 0x80002BAB8000F3DD, 0x80003A6B8000F3E2,
-    0x80003A838000F3E7, 0x80003A9380002BAF, 0x80003AA780003A9F,
-    0x80002BB380003AAF, 0x80003ADF8000F3EC, 0x80002BB780003AEB,
-    0x80003AEF80002BBB, 0x80002BBF8000F3F1, 0x8000F3FB8000F3F6,
-    0x0000B1808000F400, 0x0000B1820000B181, 0x0000D1840000B183,
-    0x0000D1860000D185, 0x0000F1880000D187, 0x00B46D8A0000F189,
-    0x0000ED8C0000F58B, 0x0000DD8E0000D58D, 0x0000AD900000D58F,
-    0x0000BD920000B591, 0x0000AD940000B593, 0x0000219600842195,
-    0x00001D9800878197, 0x0003ED9A0003E599, 0x0087819C0003E59B,
-    0x80002BC30003F99D, 0x80003B3380003B2F, 0x80003B4380003B3B,
-    0x000000848000F40A, 0x0000008600000085, 0x0000008800000087,
-    0x0000008A00000089, 0x0000008C0000008B, 0x0000008E0000008D,
-    0x000000900000008F, 0x0000009200000091, 0x0000009400000093,
-    0x0000009600000095, 0x0000009800000097, 0x0000009A00000099,
-    0x0000009C0000009B, 0x0000009E0000009D, 0x000000A00000009F,
-    0x000000A2000000A1, 0x000000A4000000A3, 0x000000A6000000A5,
-    0x000000A8000000A7, 0x000000AA000000A9, 0x000000AC000000AB,
-    0x000000AE000000AD, 0x000000B0000000AF, 0x000000B2000000B1,
-    0x000000B4000000B3, 0x000000B6000000B5, 0x000000B8000000B7,
-    0x000000BA000000B9, 0x000000BC000000BB, 0x000000BE000000BD,
-    0x000000C0000000BF, 0x000000C2000000C1, 0x000000C4000000C3,
-    0x000000C6000000C5, 0x000000C8000000C7, 0x000000CA000000C9,
-    0x000000CC000000CB, 0x000000CE000000CD, 0x000000D0000000CF,
-    0x000000D2000000D1, 0x000000D4000000D3, 0x000000D6000000D5,
-    0x000000D8000000D7, 0x000000DA000000D9, 0x000000DC000000DB,
-    0x000000DE000000DD, 0x000000E0000000DF, 0x000000E2000000E1,
-    0x000000E4000000E3, 0x000000E6000000E5, 0x000000E8000000E7,
-    0x000000EA000000E9, 0x000000EC000000EB, 0x000000EE000000ED,
-    0x000000F0000000EF, 0x000000F2000000F1, 0x000000F4000000F3,
-    0x000000F6000000F5, 0x000000F8000000F7, 0x000000FA000000F9,
-    0x000000FC000000FB, 0x000000FE000000FD, 0xCC200020000000FF,
-    0xCC200083CC200081, 0xCC200085CC200084, 0xCC200087CC200086,
-    0x0080CC88CC200088, 0xCC200081CC88CC20, 0x008ACC200082CD88,
-    0x0093CC20008BCC20, 0xCC200080CC93CC20, 0xCD93CC200081CC93,
-    0xCC200094CC200082, 0xCC94CC200080CC94, 0x0082CD94CC200081,
-    0x00A8CC2000A7CC20, 0x0082CD2000B3CC20, 0x008BD92000B9CE20,
-    0xD98CD920008CD920, 0xD920008DD9200091, 0x008ED9200091D98D,
-    0xD9200091D98ED920, 0x0091D98FD920008F, 0xD990D9200090D920,
-    0xD9200091D9200091, 0x0092D92000B0D991, 0x82E320009982E320,
-    0x210021210021009A, 0x002400230022003F, 0x0028002700260025,
-    0x2930312800293128, 0x3128002931312800, 0x0029333128002932,
-    0x3531280029343128, 0x2800293631280029, 0x2938312800293731,
-    0x3228002939312800, 0x2800293032280029, 0x2800293428002933,
-    0x2800293628002935, 0x2800293828002937, 0x2800296128002939,
-    0x2800296328002962, 0x2800296528002964, 0x2800296728002966,
-    0x2800296928002968, 0x2800296B2800296A, 0x2800296D2800296C,
-    0x2800296F2800296E, 0x2800297128002970, 0x2800297328002972,
-    0x2800297528002974, 0x2800297728002976, 0x2800297928002978,
-    0x298084E12800297A, 0x2800298284E12800, 0x84E12800298384E1,
-    0x298684E128002985, 0x2800298784E12800, 0x84E12800298984E1,
-    0x298C84E12800298B, 0x2800298E84E12800, 0x84E12800298F84E1,
-    0x299184E128002990, 0x2800299284E12800, 0xB8E428002980B8E4,
-    0x2989B8E428002983, 0x2800299DB9E42800, 0xBAE42800298CBAE4,
-    0x29A3BBE428002994, 0x28002981BCE42800, 0x85E528002991BCE4,
-    0x29AD85E5280029AB, 0x280029B48AE52800, 0x8DE5280029818DE5,
-    0x298D90E528002994, 0x280029BC91E52800, 0x9CE52800299B9BE5,
-    0x29A6ADE52800299F, 0x280029A597E62800, 0x9CE6280029889CE6,
-    0x29A89CE628002989, 0x280029AAA0E62800, 0x81E7280029B4B0E6,
-    0x29B989E7280029AB, 0x280029A39BE72800, 0xA5E7280029BEA4E7,
-    0x29ADA5E72800299D, 0x280029AA87E82800, 0xB2E8280029B387E8,
-    0x2987B3E8280029A1, 0x2800299187E92800, 0x82EB28002980B0EA,
-    0x29A48BEB28002998, 0x280029BC9DEB2800, 0xB0EB28002988A7EB,
-    0x29AC82EC28002994, 0x2800298495EC2800, 0x002984A0ECA498EC,
-    0x29849BEDA498EC28, 0x280029909EEC2800, 0xB0EC280029BCA3EC,
-    0x29B4B9EC280029A8, 0x2800298083ED2800, 0x95ED2800298C8CED,
-    0x2B002A0029002998, 0x2E002E002D002C00, 0x002F002E2E2E002E,
-    0x002E30002C300030, 0xE73000338481E230, 0x002C31003100B982,
-    0x3031003031002E31, 0x00A597E63031002E, 0x303100889CE63031,
-    0x3100313100B982E7, 0xA597E63131002E31, 0x3100889CE6313100,
-    0x00323100B982E731, 0x97E63231002E3231, 0x00889CE6323100A5,
-    0x333100B982E73231, 0xE63331002E333100, 0xB982E7333100A597,
-    0x002E343100343100, 0x343100A597E63431, 0x3100353100B982E7,
-    0xA597E63531002E35, 0x3100B982E7353100, 0x3631002E36310036,
-    0x82E7363100A597E6, 0x2E373100373100B9, 0x3100A597E6373100,
-    0x00383100B982E737, 0x97E63831002E3831, 0x00B982E7383100A5,
-    0x31002E3931003931, 0xE7393100A597E639, 0x008481E23100B982,
-    0x310030318481E231, 0x81E23100328481E2, 0x348481E231003384,
-    0x3100358481E23100, 0x81E23100368481E2, 0x388481E231003784,
-    0x3100398481E23100, 0x889CE63100A597E6, 0x003200B982E73100,
-    0x3032002E32002C32, 0xE63032002E303200, 0xB982E7303200A597,
-    0x97E6313200313200, 0x00B982E7313200A5, 0xA597E63232003232,
-    0x3200B982E7323200, 0x00A597E633320033, 0x343200B982E73332,
-    0x3200A597E6343200, 0x00353200B982E734, 0x363200A597E63532,
-    0x3200A597E6363200, 0x00A597E637320037, 0xA597E63832003832,
-    0x97E6393200393200, 0x00338481E23200A5, 0xE63200358481E232,
-    0x00889CE63200A597, 0x33003300B982E732, 0x003033002E33002C,
-    0x313300A597E63033, 0x3300A597E6313300, 0x0034330033330032,
-    0x3733003633003533, 0x3300393300383300, 0x81E23300348481E2,
-    0x388481E233003584, 0xE63300A597E63300, 0x00B982E73300889C,
-    0x002E34002C340034, 0x3234003134003034, 0x3400343400333400,
-    0x0037340036340035, 0xE234003934003834, 0xA597E63400358481,
-    0xE73400889CE63400, 0x002C35003500B982, 0xE235003035002E35,
-    0x8481E23500368481, 0x3500A597E6350038, 0xB982E73500889CE6,
-    0x2E36002C36003600, 0xE63600A597E63600, 0x00B982E73600889C,
-    0x002E37002C370037, 0xE63700388481E237, 0x00889CE63700A597,
-    0x38003800B982E737, 0x97E638002E38002C, 0x3800889CE63800A5,
-    0x2C39003900B982E7, 0xA597E639002E3900, 0xE73900889CE63900,
-    0x3D3A3A003A00B982, 0x3D003D003C003B00, 0x003E003D3D3D003D,
-    0x003F3F00213F003F, 0x005D005C005B0040, 0x00610060005F005E,
-    0x632F61002E6D2E61, 0x00756100732F6100, 0x9588E26100BECA61,
-    0x007261620062006D, 0x6F2F630063007162, 0x6C616300752F6300,
-    0x6300646300636300, 0x6D6300326D63006D, 0xE263002E6F630033,
-    0x64006400676B9588, 0x006A640062640061, 0x6D64006D64006C64,
-    0x7A6400336D640032, 0x65006500BEC56400, 0x0066007665006772,
-    0x6600666600786166, 0x66006C6666006966, 0x006D66006C660069,
-    0x6267006C61670067, 0x617067007A686700, 0x6168006800796700,
-    0x6800706800676800, 0x7A68007668006170, 0x6969006969006900,
-    0x006E69006A690069, 0x7869007669007569, 0x00616B006B006A00,
-    0x006C61636B00626B, 0x6B007A686B00676B, 0x006D6B006C6B006B,
-    0x00336D6B00326D6B, 0x6B00746B0061706B, 0x89CF6B00776B0076,
-    0x6D6C006A6C006C00, 0x00676F6C006E6C00, 0x6C00786C0064746C,
-    0x00326D006D00B7C2, 0x626D00616D00336D, 0x6D00646D00636D00,
-    0x696D007A686D0067, 0x006D6D006C6D006C, 0x00336D6D00326D6D,
-    0x0061706D006C6F6D, 0x766D00736D00726D, 0x0089CF6D00776D00,
-    0xE26D00739588E26D, 0x6E006E0032739588, 0x006A6E00666E0061,
-    0x736E006F6E006D6E, 0x6F00776E00766E00, 0x2E70007000766F00,
-    0x6370006170002E6D, 0x7000687000667000, 0x7000767070006D70,
-    0x6574700073700072, 0x7100777000767000, 0x7200646172007200,
-    0x7200739588E26461, 0x0032739588E26461, 0x0064730073007372,
-    0x7373007273006D73, 0x7400767300747300, 0x7A6874006C657400,
-    0x00760075006D7400, 0x7600696976006976, 0x9588E27600696969,
-    0x770062770077006D, 0x780078007A770063, 0x0079006969780069,
-    0x007D007C007B007A, 0x00A3C200A2C2007E, 0xACC200A6C200A5C2,
-    0x66B0C20063B0C200, 0xC3009FC300B7C200, 0x00A2C300A1C300A0,
-    0xA5C300A4C300A3C3, 0xC300A7C300A6C300, 0x00AAC300A9C300A8,
-    0xADC300ACC300ABC3, 0xC300AFC300AEC300, 0x00B2C300B1C300B0,
-    0xB5C300B4C300B3C3, 0xC300B8C300B6C300, 0x00BBC300BAC300B9,
-    0xBEC300BDC300BCC3, 0xC40081C400BFC300, 0x0087C40085C40083,
-    0x8DC4008BC40089C4, 0xC40091C4008FC400, 0x0097C40095C40093,
-    0x9DC4009BC40099C4, 0xC400A1C4009FC400, 0x00A7C400A5C400A3,
-    0xADC400ABC400A9C4, 0xC400B0C400AFC400, 0x00B7C400B5C400B1,
-    0xBEC400BCC400BAC4, 0xC50084C50082C500, 0x008BC50088C50086,
-    0x91C5008FC5008DC5, 0xC50095C50093C500, 0x009BC50099C50097,
-    0xA1C5009FC5009DC5, 0xC500A5C500A3C500, 0x00ABC500A9C500A7,
-    0xB1C500AFC500ADC5, 0xC500B5C500B3C500, 0x00BCC500BAC500B7,
-    0x83C60080C600BEC5, 0xC60088C60085C600, 0x0095C60092C6008C,
-    0x9EC6009AC60099C6, 0xC600A3C600A1C600, 0x00ABC600A8C600A5,
-    0xB4C600B0C600ADC6, 0xC600B9C600B6C600, 0x0080C700BFC600BD,
-    0x8EC70082C70081C7, 0xC70092C70090C700, 0x0098C70096C70094,
-    0x9DC7009CC7009AC7, 0xC700A1C7009FC700, 0x00A7C700A5C700A3,
-    0xADC700ABC700A9C7, 0xC700B0C700AFC700, 0x00BBC700B9C700B5,
-    0x81C800BFC700BDC7, 0xC80085C80083C800, 0x008BC80089C80087,
-    0x91C8008FC8008DC8, 0xC80095C80093C800, 0x009BC80099C80097,
-    0xA3C8009FC8009DC8, 0xC800A7C800A5C800, 0x00ADC800ABC800A9,
-    0xB3C800B1C800AFC8, 0xC800BCC800B7C800, 0x0082C90080C900BF,
-    0x8BC90089C90087C9, 0xC9008FC9008DC900, 0x0092C90091C90090,
-    0x95C90094C90093C9, 0xC90097C90096C900, 0x009BC90099C90098,
-    0x9FC9009EC9009CC9, 0xC900A1C900A0C900, 0x00A4C900A3C900A2,
-    0xA7C900A6C900A5C9, 0xC900A9C900A8C900, 0x00ACC900ABC900AA,
-    0xAFC900AEC900ADC9, 0xC900B1C900B0C900, 0x00B4C900B3C900B2,
-    0xB7C900B6C900B5C9, 0xC900B9C900B8C900, 0x00BDC900BBC900BA,
-    0x81CA0080CA00BEC9, 0xCA0083CA0082CA00, 0x0088CA0087CA0084,
-    0x8BCA008ACA0089CA, 0xCA008DCA008CCA00, 0x0090CA008FCA008E,
-    0x95CA0092CA0091CA, 0xCA0099CA0098CA00, 0x009DCA009CCA009B,
-    0xA1CA009FCA009ECA, 0xCA00A3CA00A2CA00, 0x00A6CA00A5CA00A4,
-    0xA9CA00A8CA00A7CA, 0xCA00ABCA00AACA00, 0x90CB006EBCCA00B9,
-    0xCC0080CC0091CB00, 0xCC0081CC88CC0081, 0x00B3CD00B1CD0093,
-    0xBCCD00BBCD00B7CD, 0xCE0090CE00BDCD00, 0x00AECE00ADCE00AC,
-    0xB1CE00B0CE00AFCE, 0xCE00B3CE00B2CE00, 0x00B6CE00B5CE00B4,
-    0xB9CE00B8CE00B7CE, 0xCE00BBCE00BACE00, 0xBCCE0061BCCE00BC,
-    0xBCCE0067BCCE0066, 0xBCCE006DBCCE006C, 0xBCCE0076BCCE0073,
-    0x00BECE00BDCE0077, 0x81CF0080CF00BFCE, 0xCF0084CF0083CF00,
-    0x0087CF0086CF0085, 0x8ACF0089CF0088CF, 0xCF008CCF008BCF00,
-    0x0097CF008ECF008D, 0x9DCF009BCF0099CF, 0xCF00A1CF009FCF00,
-    0x00A7CF00A5CF00A3, 0xADCF00ABCF00A9CF, 0xCF00B3CF00AFCF00,
-    0x00B0D000BBCF00B8, 0xB3D000B2D000B1D0, 0xD000B5D000B4D000,
-    0x00B8D000B7D000B6, 0xBBD000BAD000B9D0, 0xD000BDD000BCD000,
-    0x0080D100BFD000BE, 0x83D10082D10081D1, 0xD10085D10084D100,
-    0x0088D10087D10086, 0x8BD1008AD10089D1, 0xD1008DD1008CD100,
-    0x0090D1008FD1008E, 0x93D10092D10091D1, 0xD10095D10094D100,
-    0x0098D10097D10096, 0x9BD1009AD10099D1, 0xD1009DD1009CD100,
-    0x00A1D1009FD1009E, 0xA7D100A5D100A3D1, 0xD100ABD100A9D100,
-    0x00B1D100AFD100AD, 0xB7D100B5D100B3D1, 0xD100BBD100B9D100,
-    0x0081D200BFD100BD, 0x8FD2008DD2008BD2, 0xD20093D20091D200,
-    0x0099D20097D20095, 0x9FD2009DD2009BD2, 0xD200A3D200A1D200,
-    0x00A9D200A7D200A5, 0xAFD200ADD200ABD2, 0xD200B3D200B1D200,
-    0x00B9D200B7D200B5, 0xBFD200BDD200BBD2, 0xD30084D30082D300,
-    0x008AD30088D30086, 0x8FD3008ED3008CD3, 0xD30093D30091D300,
-    0x0099D30097D30095, 0x9FD3009DD3009BD3, 0xD300A3D300A1D300,
-    0x00A9D300A7D300A5, 0xAFD300ADD300ABD3, 0xD300B3D300B1D300,
-    0x00B9D300B7D300B5, 0xBFD300BDD300BBD3, 0xD40083D40081D400,
-    0x0089D40087D40085, 0x8FD4008DD4008BD4, 0xD40093D40091D400,
-    0x0099D40097D40095, 0x9FD4009DD4009BD4, 0xD400A3D400A1D400,
-    0x00A9D400A7D400A5, 0xAFD400ADD400ABD4, 0xD500A2D500A1D500,
-    0x00A5D500A4D500A3, 0x00A6D50082D6A5D5, 0xA9D500A8D500A7D5,
-    0xD500ABD500AAD500, 0x00AED500ADD500AC, 0xB1D500B0D500AFD5,
-    0xD500B3D500B2D500, 0xD500A5D5B4D500B4, 0xADD5B4D500ABD5B4,
-    0xB5D500B6D5B4D500, 0xD500B7D500B6D500, 0x00BAD500B9D500B8,
-    0xBDD500BCD500BBD5, 0xB6D5BED500BED500, 0xD60080D600BFD500,
-    0x0083D60082D60081, 0x86D60085D60084D6, 0xB7D690D70090D700,
-    0x90D700B8D690D700, 0x009CD790D700BCD6, 0x00BCD691D70091D7,
-    0x0092D700BFD691D7, 0x0093D700BCD692D7, 0x0094D700BCD693D7,
-    0xD695D700BCD694D7, 0xD700BCD695D700B9, 0xBCD698D700BCD696,
-    0x99D700B4D699D700, 0x00BCD69AD700BCD6, 0x00BCD69BD7009BD7,
-    0x009CD700BFD69BD7, 0x009DD700BCD69CD7, 0xD6A0D700BCD69ED7,
-    0xD700BCD6A1D700BC, 0xD700BCD6A3D700A2, 0xBFD6A4D700BCD6A4,
-    0xA7D700BCD6A6D700, 0xA8D700A8D700BCD6, 0x00BCD6A9D700BCD6,
-    0xD70081D7BCD6A9D7, 0xA9D70082D7BCD6A9, 0x0082D7A9D70081D7,
-    0x00BCD6AAD700AAD7, 0x00A1D800B7D6B2D7, 0xA4D800A3D800A2D8,
-    0xD800A6D800A5D800, 0xACD8A6D800A7D8A6, 0xA6D800ADD8A6D800,
-    0x00B1D8A6D800AED8, 0xD9A6D800B2D8A6D8, 0xD80086D9A6D80085,
-    0x88D9A6D80087D9A6, 0xA6D80089D9A6D800, 0x0086DBA6D8008AD9,
-    0xDBA6D80087DBA6D8, 0xD80090DBA6D80088, 0xD800A7D80095DBA6,
-    0x00B1D8A8D883D9A7, 0x87D984D984D9A7D8, 0xA7D8008BD9A7D800,
-    0xA8D800A8D800B4D9, 0x00ADD8A8D800ACD8, 0xD8008AD9ADD8A8D8,
-    0xAED8A8D800AED8A8, 0x00B1D8A8D8008AD9, 0xD9A8D800B2D8A8D8,
-    0xD80086D9A8D80085, 0x89D9A8D80087D9A8, 0xA9D8008AD9A8D800,
-    0xACD8AAD800AAD800, 0x0085D9ACD8AAD800, 0xD80089D9ACD8AAD8,
-    0xAAD8008AD9ACD8AA, 0xD8ADD8AAD800ADD8, 0x85D9ADD8AAD800AC,
-    0xAAD800AED8AAD800, 0xD8AAD80085D9AED8, 0xAED8AAD80089D9AE,
-    0x00B1D8AAD8008AD9, 0xD9AAD800B2D8AAD8, 0xACD885D9AAD80085,
-    0x00ADD885D9AAD800, 0xD800AED885D9AAD8, 0xAAD80089D985D9AA,
-    0xD9AAD8008AD985D9, 0xD80087D9AAD80086, 0x8AD9AAD80089D9AA,
-    0xACD8ABD800ABD800, 0xABD800B1D8ABD800, 0x0085D9ABD800B2D8,
-    0xD9ABD80086D9ABD8, 0xD80089D9ABD80087, 0xD800ACD8008AD9AB,
-    0xADD8ACD800ADD8AC, 0xD9ADD8ACD80089D9, 0xD82084D9ACD8008A,
-    0xD984D9A7D884D9AC, 0xD80085D9ACD80087, 0xACD800ADD885D9AC,
-    0xD9ACD80089D985D9, 0x89D9ACD8008AD985, 0xADD8008AD9ACD800,
-    0xADD800ACD8ADD800, 0xD9ADD8008AD9ACD8, 0x89D985D9ADD80085,
-    0x008AD985D9ADD800, 0xD9ADD80089D9ADD8, 0xD8AED800AED8008A,
-    0xD800ADD8AED800AC, 0x89D9AED80085D9AE, 0xAFD8008AD9AED800,
-    0xB0D9B0D800B0D800, 0xB3D8B1D800B1D800, 0xD9B1D80084D988D9,
-    0xA7D88CDBB1D800B0, 0xB3D800B2D80084D9, 0xB3D800ACD8B3D800,
-    0xD8B3D800ADD8ACD8, 0xADD8B3D80089D9AC, 0x00ACD8ADD8B3D800,
-    0xD8B3D800AED8B3D8, 0xAED8B3D80089D9AE, 0x00B1D8B3D8008AD9,
-    0xD9B3D80085D9B3D8, 0x85D9B3D800ACD885, 0xD985D9B3D800ADD8,
-    0xD80087D9B3D80085, 0x8AD9B3D80089D9B3, 0xACD8B4D800B4D800,
-    0x008AD9ACD8B4D800, 0xD8B4D800ADD8B4D8, 0xADD8B4D80085D9AD,
-    0x00AED8B4D8008AD9, 0xD9B4D800B1D8B4D8, 0xAED885D9B4D80085,
-    0x0085D985D9B4D800, 0xD9B4D80087D9B4D8, 0xD8008AD9B4D80089,
-    0xD800ADD8B5D800B5, 0xB5D800ADD8ADD8B5, 0xD8B5D8008AD9ADD8,
-    0xD800B1D8B5D800AE, 0x0085D9B9D884D9B5, 0xD80089D984D9B5D8,
-    0xA7D82089D984D9B5, 0xD82087D984D984D9, 0x2087D98AD984D9B9,
-    0x85D984D9B3D888D9, 0x0092DB84D9B5D800, 0xD9B5D80085D9B5D8,
-    0x89D9B5D80085D985, 0xB6D8008AD9B5D800, 0xB6D800ACD8B6D800,
-    0xD9ADD8B6D800ADD8, 0x8AD9ADD8B6D80089, 0xB6D800AED8B6D800,
-    0xD8B6D80085D9AED8, 0xD80085D9B6D800B1, 0x8AD9B6D80089D9B6,
-    0xADD8B7D800B7D800, 0xB7D80085D9B7D800, 0xD9B7D800ADD885D9,
-    0x85D9B7D80085D985, 0x0089D9B7D8008AD9, 0x00B8D8008AD9B7D8,
-    0x00B9D80085D9B8D8, 0xD8B9D800ACD8B9D8, 0x84D9B9D80085D9AC,
-    0xD9B9D80087D98AD9, 0x85D985D9B9D80085, 0x0089D985D9B9D800,
-    0xD8008AD985D9B9D8, 0x8AD9B9D80089D9B9, 0xACD8BAD800BAD800,
-    0xBAD80085D9BAD800, 0xD9BAD80085D985D9, 0x85D9BAD80089D985,
-    0x0089D9BAD8008AD9, 0xD980D9008AD9BAD8, 0xD9008ED980D9008B,
-    0x80D90091D98ED980, 0xD98FD980D9008FD9, 0xD90090D980D90091,
-    0x80D90091D990D980, 0x0092D980D90091D9, 0x00ACD881D90081D9,
-    0xD881D900ADD881D9, 0x85D9AED881D900AE, 0x81D90085D981D900,
-    0xD981D9008AD985D9, 0xD9008AD981D90089, 0xD900ADD882D90082,
-    0x82D90092DB84D982, 0xD885D982D90085D9, 0x85D985D982D900AD,
-    0x008AD985D982D900, 0xD982D90089D982D9, 0xD883D90083D9008A,
-    0xD900ACD883D900A7, 0xAED883D900ADD883, 0x83D90084D983D900,
-    0xD985D983D90085D9, 0x8AD985D983D90085, 0x83D90089D983D900,
-    0x84D90084D9008AD9, 0x00A3D884D900A2D8, 0xD884D900A5D884D9,
-    0xD900ACD884D900A7, 0x84D900ACD8ACD884, 0xD884D90085D9ACD8,
-    0xADD884D9008AD9AC, 0x0085D9ADD884D900, 0xD90089D9ADD884D9,
-    0x84D9008AD9ADD884, 0xD9AED884D900AED8, 0xD90085D984D90085,
-    0x84D900ADD885D984, 0xD984D9008AD985D9, 0xD90089D984D90087,
-    0xD90085D9008AD984, 0xACD885D900A7D885, 0x00ADD8ACD885D900,
-    0xD900AED8ACD885D9, 0x85D90085D9ACD885, 0xD885D9008AD9ACD8,
-    0xACD8ADD885D900AD, 0x0085D9ADD885D900, 0xAFD885D9ADD885D9,
-    0x008AD9ADD885D900, 0xD885D900AED885D9, 0xAED885D900ACD8AE,
-    0xD9AED885D90085D9, 0xD90085D985D9008A, 0x85D9008AD985D985,
-    0x008AD985D90089D9, 0x00ACD886D90086D9, 0xD900ADD8ACD886D9,
-    0x86D90085D9ACD886, 0xD886D90089D9ACD8, 0xADD886D9008AD9AC,
-    0x0085D9ADD886D900, 0xD90089D9ADD886D9, 0x86D9008AD9ADD886,
-    0x00B1D886D900AED8, 0xD986D900B2D886D9, 0x89D985D986D90085,
-    0x008AD985D986D900, 0xD986D90086D986D9, 0xD90089D986D90087,
-    0xD90087D9008AD986, 0x85D987D900ACD887, 0x00ACD885D987D900,
-    0xD90085D985D987D9, 0x8AD987D90089D987, 0x88D900B0D987D900,
-    0xD984D9B3D888D900, 0xD900B4D988D90085, 0xD900B0D989D90089,
-    0xD900ACD88AD9008A, 0x8AD9008AD9ACD88A, 0xD9ADD88AD900ADD8,
-    0xD900AED88AD9008A, 0xB2D88AD900B1D88A, 0x8AD90085D98AD900,
-    0xD98AD90085D985D9, 0x86D98AD9008AD985, 0x8AD90087D98AD900,
-    0x008AD98AD90089D9, 0x00AED900B4D98AD9, 0xB9D900B1D900AFD9,
-    0xD900BBD900BAD900, 0x0080DA00BFD900BE, 0x86DA0084DA0083DA,
-    0xDA0088DA0087DA00, 0x008EDA008DDA008C, 0xA1DA0098DA0091DA,
-    0xDA00A6DA00A4DA00, 0x00AFDA00ADDA00A9, 0xBADA00B3DA00B1DA,
-    0xDB00BEDA00BBDA00, 0x0082DB0081DB0080, 0x87DB0086DB0085DB,
-    0x88DB00B4D987DB00, 0xDB008BDB0089DB00, 0x0092DB0090DB008C,
-    0xA4E095A4E00093DB, 0xBCA4E096A4E000BC, 0x00BCA4E097A4E000,
-    0xE000BCA4E09CA4E0, 0xA4E000BCA4E0A1A4, 0xA9A4E000BCA4E0A2,
-    0x00BCA4E0ABA4E000, 0xE000BCA4E0AFA4E0, 0xE000B4A4E000B1A4,
-    0xA6E000BCA6E0A1A6, 0xAFA6E000BCA6E0A2, 0x008BA7E000BCA6E0,
-    0xE096A8E0008CA7E0, 0xA8E097A8E000BCA8, 0xBCA8E09CA8E000BC,
-    0x00BCA8E0ABA8E000, 0xE000BCA8E0B2A8E0, 0xACE000BCA8E0B8A8,
-    0xA2ACE000BCACE0A1, 0x0088ADE000BCACE0, 0x008CADE0008BADE0,
-    0x008AAFE00094AEE0, 0x008CAFE0008BAFE0, 0x0080B3E00088B1E0,
-    0x0088B3E00087B3E0, 0x008BB3E0008AB3E0, 0x008BB5E0008AB5E0,
-    0x009AB7E0008CB5E0, 0x009DB7E0009CB7E0, 0xE08DB9E0009EB7E0,
-    0xBAE0ABBAE000B2B8, 0xA1BAE0ABBAE00099, 0x00B2BAE08DBBE000,
-    0xE080BDE0008BBCE0, 0xBEE082BDE000B5BE, 0xB7BEE08CBDE000B7,
-    0x00B7BEE091BDE000, 0xE000B7BEE096BDE0, 0xBDE000B7BEE09BBD,
-    0xB1BDE000B2BDE0B1, 0xE0B1BDE000B4BDE0, 0xBEE090BEE00080BE,
-    0xB7BEE092BEE000B5, 0x00B7BEE09CBEE000, 0xE000B7BEE0A1BEE0,
-    0xBEE000B7BEE0A6BE, 0xB2BEE000B7BEE0AB, 0xE00080BEE0B1BDE0,
-    0xBEE00080BEE0B2BE, 0x0080BEE0B1BDE0B3, 0xE10080BEE0B3BEE0,
-    0xE1009083E100A680, 0xE1009283E1009183, 0xE1009483E1009383,
-    0xE1009683E1009583, 0xE1009883E1009783, 0xE1009A83E1009983,
-    0xE1009C83E1009B83, 0xE1009E83E1009D83, 0xE100A083E1009F83,
-    0xE100A283E100A183, 0xE100A483E100A383, 0xE100A683E100A583,
-    0xE100A883E100A783, 0xE100AA83E100A983, 0xE100AC83E100AB83,
-    0xE100AE83E100AD83, 0xE100B083E100AF83, 0xE100B283E100B183,
-    0xE100B483E100B383, 0xE100B683E100B583, 0xE100B883E100B783,
-    0xE100BA83E100B983, 0xE100BE83E100BD83, 0xE1008084E100BF83,
-    0xE1008284E1008184, 0xE1008484E1008384, 0xE1008684E1008584,
-    0xE1008884E1008784, 0xE1008A84E1008984, 0xE1008C84E1008B84,
-    0xE1008E84E1008D84, 0xE1009084E1008F84, 0xE1009284E1009184,
-    0xE1009584E1009484, 0xE1009C84E1009A84, 0xE1009E84E1009D84,
-    0xE100A184E100A084, 0xE100A384E100A284, 0xE100A984E100A784,
-    0xE100AC84E100AB84, 0xE100AE84E100AD84, 0xE100B284E100AF84,
-    0xE1008085E100B684, 0xE1008C85E1008785, 0xE1009885E1009785,
-    0xE100A085E1009985, 0xE100A285E100A185, 0xE100A485E100A385,
-    0xE100A685E100A585, 0xE100A885E100A785, 0xE100AA85E100A985,
-    0xE100AC85E100AB85, 0xE100AE85E100AD85, 0xE100B085E100AF85,
-    0xE100B285E100B185, 0xE100B485E100B385, 0xE1008486E100B585,
-    0xE1008886E1008586, 0xE1009286E1009186, 0xE1009E86E1009486,
-    0xE100AA86E100A186, 0xE100AD86E100AC86, 0xE100B186E100B086,
-    0xE100B386E100B286, 0xE100B586E100B486, 0xE1008887E1008787,
-    0xE1008E87E1008C87, 0xE1009787E1009387, 0xE1009D87E1009987,
-    0xE100B187E1009F87, 0xE100A08EE100B287, 0xE100A28EE100A18E,
-    0xE100A48EE100A38E, 0xE100A68EE100A58E, 0xE100A88EE100A78E,
-    0xE100AA8EE100A98E, 0xE100AC8EE100AB8E, 0xE100AE8EE100AD8E,
-    0xE100B08EE100AF8E, 0xE100B28EE100B18E, 0xE100B48EE100B38E,
-    0xE100B68EE100B58E, 0xE100B88EE100B78E, 0xE100BA8EE100B98E,
-    0xE100BC8EE100BB8E, 0xE100BE8EE100BD8E, 0xE100808FE100BF8E,
-    0xE100828FE100818F, 0xE100848FE100838F, 0xE100868FE100858F,
-    0xE100888FE100878F, 0xE1008A8FE100898F, 0xE1008C8FE1008B8F,
-    0xE1008E8FE1008D8F, 0xE100908FE1008F8F, 0xE100928FE100918F,
-    0xE100948FE100938F, 0xE100968FE100958F, 0xE100988FE100978F,
-    0xE1009A8FE100998F, 0xE1009C8FE1009B8F, 0xE1009E8FE1009D8F,
-    0xE100A08FE1009F8F, 0xE100A28FE100A18F, 0xE100A48FE100A38F,
-    0xE100A68FE100A58F, 0xE100A88FE100A78F, 0xE100AA8FE100A98F,
-    0xE100AC8FE100AB8F, 0xE100AE8FE100AD8F, 0xE100B08FE100AF8F,
-    0xE100B28FE100B18F, 0xE100B48FE100B38F, 0xE10086ACE100B58F,
-    0xE1008AACE10088AC, 0xE1008EACE1008CAC, 0xE100BBACE10092AC,
-    0xE10080ADE100BDAC, 0xE10083ADE10081AD, 0xE10096B4E10082B4,
-    0xE1009CB4E10097B4, 0xE100A5B4E1009DB4, 0xE100BBB5E100B9B5,
-    0xE10085B6E100BDB5, 0xE10091B6E1008EB6, 0xE10083B8E10081B8,
-    0xE10087B8E10085B8, 0xE1008BB8E10089B8, 0xE1008FB8E1008DB8,
-    0xE10093B8E10091B8, 0xE10097B8E10095B8, 0xE1009BB8E10099B8,
-    0xE1009FB8E1009DB8, 0xE100A3B8E100A1B8, 0xE100A7B8E100A5B8,
-    0xE100ABB8E100A9B8, 0xE100AFB8E100ADB8, 0xE100B3B8E100B1B8,
-    0xE100B7B8E100B5B8, 0xE100BBB8E100B9B8, 0xE100BFB8E100BDB8,
-    0xE10083B9E10081B9, 0xE10087B9E10085B9, 0xE1008BB9E10089B9,
-    0xE1008FB9E1008DB9, 0xE10093B9E10091B9, 0xE10097B9E10095B9,
-    0xE1009BB9E10099B9, 0xE1009FB9E1009DB9, 0xE100A3B9E100A1B9,
-    0xE100A7B9E100A5B9, 0xE100ABB9E100A9B9, 0xE100AFB9E100ADB9,
-    0xE100B3B9E100B1B9, 0xE100B7B9E100B5B9, 0xE100BBB9E100B9B9,
-    0xE100BFB9E100BDB9, 0xE10083BAE10081BA, 0xE10087BAE10085BA,
-    0xE1008BBAE10089BA, 0xE1008FBAE1008DBA, 0xE10093BAE10091BA,
-    0xE10096BAE10095BA, 0xE10098BAE10097BA, 0xE100A1BAE10099BA,
-    0xE100A5BAE100A3BA, 0xE100A9BAE100A7BA, 0xE100ADBAE100ABBA,
-    0xE100B1BAE100AFBA, 0xE100B5BAE100B3BA, 0xE100B9BAE100B7BA,
-    0xE100BDBAE100BBBA, 0xE10081BBE100BFBA, 0xE10085BBE10083BB,
-    0xE10089BBE10087BB, 0xE1008DBBE1008BBB, 0xE10091BBE1008FBB,
-    0xE10095BBE10093BB, 0xE10099BBE10097BB, 0xE1009DBBE1009BBB,
-    0xE100A1BBE1009FBB, 0xE100A5BBE100A3BB, 0xE100A9BBE100A7BB,
-    0xE100ADBBE100ABBB, 0xE100B1BBE100AFBB, 0xE100B5BBE100B3BB,
-    0xE100B9BBE100B7BB, 0xE100BDBBE100BBBB, 0xE10080BCE100BFBB,
-    0xE10082BCE10081BC, 0xE10084BCE10083BC, 0xE10086BCE10085BC,
-    0xE10090BCE10087BC, 0xE10092BCE10091BC, 0xE10094BCE10093BC,
-    0xE100A0BCE10095BC, 0xE100A2BCE100A1BC, 0xE100A4BCE100A3BC,
-    0xE100A6BCE100A5BC, 0xE100B0BCE100A7BC, 0xE100B2BCE100B1BC,
-    0xE100B4BCE100B3BC, 0xE100B6BCE100B5BC, 0xE10080BDE100B7BC,
-    0xE10082BDE10081BD, 0xE10084BDE10083BD, 0xE10090BDE10085BD,
-    0xE10092BDE10091BD, 0xE10094BDE10093BD, 0xE10096BDE10095BD,
-    0xE100A0BDE10097BD, 0xE100A2BDE100A1BD, 0xE100A4BDE100A3BD,
-    0xE100A6BDE100A5BD, 0xE100B0BDE100A7BD, 0xE100B4BDE100B2BD,
-    0xE100B8BDE100B6BD, 0xE100BCBDE100BABD, 0xE10081BEE10080BE,
-    0xE10083BEE10082BE, 0xE10085BEE10084BE, 0xE10087BEE10086BE,
-    0xE10091BEE10090BE, 0xE10093BEE10092BE, 0xE10095BEE10094BE,
-    0xE10097BEE10096BE, 0xE100A1BEE100A0BE, 0xE100A3BEE100A2BE,
-    0xE100A5BEE100A4BE, 0xE100A7BEE100A6BE, 0xE100B1BEE100B0BE,
-    0xE100B3BEE100B2BE, 0xE100B6BEE100B4BE, 0xE10082BFE100B7BE,
-    0xE10084BFE10083BF, 0xE10087BFE10086BF, 0xE10091BFE10090BF,
-    0xE10096BFE10092BF, 0xE100A0BFE10097BF, 0xE100A2BFE100A1BF,
-    0xE100A5BFE100A4BF, 0xE100A7BFE100A6BF, 0xE100B3BFE100B2BF,
-    0xE100B6BFE100B4BF, 0xE2009080E200B7BF, 0xE2009480E2009380,
-    0x80E200B280E2B280, 0x00B280E2B280E2B2, 0x80E2B280E2B280E2,
-    0xB580E200B280E2B2, 0xE2B580E200B580E2, 0x82E200B580E2B580,
-    0x86E2008E85E200A9, 0x86E2009086E20084, 0x86E2009286E20091,
-    0x86E2009A86E20093, 0x87E200AE86E2009B, 0x87E2008E87E2008D,
-    0x88E2008288E2008F, 0x88E2008788E20084, 0x88E2008C88E20089,
-    0x88E2009288E20091, 0x88E200A688E200A4, 0xAB88E200AB88E2AB,
-    0xE200AB88E2AB88E2, 0xAB88E2AB88E2AB88, 0xE2AE88E200AB88E2,
-    0x88E2AE88E200AE88, 0x8189E200AE88E2AE, 0x8789E2008489E200,
-    0xA089E2008989E200, 0xAD89E200A289E200, 0xAF89E200AE89E200,
-    0xB189E200B089E200, 0xB589E200B489E200, 0xB989E200B889E200,
-    0x818AE200808AE200, 0x858AE200848AE200, 0x898AE200888AE200,
-    0xAD8AE200AC8AE200, 0xAF8AE200AE8AE200, 0xA18BE200A08BE200,
-    0xA38BE200A28BE200, 0xAB8BE200AA8BE200, 0xAD8BE200AC8BE200,
-    0xA096E2008294E200, 0x85A6E2008B97E200, 0x9DABE20086A6E200,
-    0xE200B0B0E200B8CC, 0xE200B2B0E200B1B0, 0xE200B4B0E200B3B0,
-    0xE200B6B0E200B5B0, 0xE200B8B0E200B7B0, 0xE200BAB0E200B9B0,
-    0xE200BCB0E200BBB0, 0xE200BEB0E200BDB0, 0xE20080B1E200BFB0,
-    0xE20082B1E20081B1, 0xE20084B1E20083B1, 0xE20086B1E20085B1,
-    0xE20088B1E20087B1, 0xE2008AB1E20089B1, 0xE2008CB1E2008BB1,
-    0xE2008EB1E2008DB1, 0xE20090B1E2008FB1, 0xE20092B1E20091B1,
-    0xE20094B1E20093B1, 0xE20096B1E20095B1, 0xE20098B1E20097B1,
-    0xE2009AB1E20099B1, 0xE2009CB1E2009BB1, 0xE2009EB1E2009DB1,
-    0xE200A1B1E2009FB1, 0xE200A6B1E200A5B1, 0xE200AAB1E200A8B1,
-    0xE200B1B1E200ACB1, 0xE200B6B1E200B3B1, 0xE20083B2E20081B2,
-    0xE20087B2E20085B2, 0xE2008BB2E20089B2, 0xE2008FB2E2008DB2,
-    0xE20093B2E20091B2, 0xE20097B2E20095B2, 0xE2009BB2E20099B2,
-    0xE2009FB2E2009DB2, 0xE200A3B2E200A1B2, 0xE200A7B2E200A5B2,
-    0xE200ABB2E200A9B2, 0xE200AFB2E200ADB2, 0xE200B3B2E200B1B2,
-    0xE200B7B2E200B5B2, 0xE200BBB2E200B9B2, 0xE200BFB2E200BDB2,
-    0xE20083B3E20081B3, 0xE20087B3E20085B3, 0xE2008BB3E20089B3,
-    0xE2008FB3E2008DB3, 0xE20093B3E20091B3, 0xE20097B3E20095B3,
-    0xE2009BB3E20099B3, 0xE2009FB3E2009DB3, 0xE200A3B3E200A1B3,
-    0xE200AEB3E200ACB3, 0xE20080B4E200B3B3, 0xE20082B4E20081B4,
-    0xE20084B4E20083B4, 0xE20086B4E20085B4, 0xE20088B4E20087B4,
-    0xE2008AB4E20089B4, 0xE2008CB4E2008BB4, 0xE2008EB4E2008DB4,
-    0xE20090B4E2008FB4, 0xE20092B4E20091B4, 0xE20094B4E20093B4,
-    0xE20096B4E20095B4, 0xE20098B4E20097B4, 0xE2009AB4E20099B4,
-    0xE2009CB4E2009BB4, 0xE2009EB4E2009DB4, 0xE200A0B4E2009FB4,
-    0xE200A2B4E200A1B4, 0xE200A4B4E200A3B4, 0xE200A7B4E200A5B4,
-    0xE300A1B5E200ADB4, 0xE3008280E3008180, 0xE3008980E3008880,
-    0xE3008B80E3008A80, 0xE3008D80E3008C80, 0xE3008F80E3008E80,
-    0xE3009180E3009080, 0xE3009480E3009280, 0xE3009580E3739480,
-    0x9580E389B8E49480, 0xE38CBAE49480E300, 0x8BE59480E3009580,
-    0x9480E3009580E39D, 0xE3009580E389AEE5, 0x9580E39389E69480,
-    0xE39795E69480E300, 0x9CE69480E3009580, 0x9480E3009580E3AC,
-    0xE3009580E3B982E7, 0x9580E3979BE79480, 0x9680E3009580E300,
-    0x8C81E3009780E300, 0x9081E3008E81E300, 0x9481E3009281E300,
-    0x9881E3009681E300, 0x9C81E3009A81E300, 0xA081E3009E81E300,
-    0xA581E300A281E300, 0xA981E300A781E300, 0xB181E300B081E300,
-    0xB481E300B381E300, 0xB781E300B681E300, 0xBA81E300B981E300,
-    0x008B81E3BB81E300, 0x00BD81E300BC81E3, 0xE3008A82E38882E3,
-    0xE3009982E3009482, 0xE3009E82E3009A82, 0xE300A282E300A182,
-    0xBC83E39183E3A282, 0xE3A282E3008883E3, 0xA182E39583E3AB83,
-    0xE3B383E3A282E300, 0x82E300A282E39A83, 0x00AB83E3BC83E3A2,
-    0x00A482E300A382E3, 0x83E38B83E3A482E3, 0xA482E300B082E3B3,
-    0xE3008183E3B383E3, 0xE300A682E300A582, 0xB383E3A982E3A682,
-    0xA882E300A782E300, 0xE3B982E3A882E300, 0x8983E3BC83E3AF82,
-    0xE3BC83E3A882E300, 0x82E300BC83E3AB82, 0x82E300AA82E300A9,
-    0x00B982E3B383E3AA, 0x83E3BC83E3AA82E3, 0x82E300AB82E300A0,
-    0x00AA83E3A482E3AB, 0x83E3A983E3AB82E3, 0xAB82E3008883E383,
-    0x83E3AA83E3AD83E3, 0x82E300AC82E300BC, 0x00B383E3AD83E3AC,
-    0x83E3B383E3AC82E3, 0x82E300AD82E3009E, 0xE3AA83E3A583E3AD,
-    0x83E3AD82E300BC83, 0xAD83E3AD82E300AD, 0x83E3A983E3B082E3,
-    0xAD83E3AD82E300A0, 0x83E3BC83E3A183E3, 0xAD82E300AB83E388,
-    0x83E3AF83E3AD83E3, 0xAE82E3008883E383, 0x00AC82E3AE82E300,
-    0x83E38B83E3AE82E3, 0xAB83E3AE82E300BC, 0xE300BC83E38083E3,
-    0x83E3AF82E300AF82, 0xE3A482E3BC82E3AB, 0x83E3AF82E300AD83,
-    0x008D83E3BC83E3AD, 0xE3B082E300B082E3, 0x82E300A083E3A983,
-    0xE3A083E3A983E3B0, 0x82E300B383E38883, 0xBC83E3B182E300B1,
-    0x00B282E300B982E3, 0xE3B382E300B382E3, 0x83E3B382E300B382,
-    0xAB83E3B382E30088, 0xE3B382E3008A83E3, 0x82E3009D83E3BC83,
-    0x82E300B582E300B4, 0xE3AF82E3A482E3B5, 0x83E3B582E300AB83,
-    0xE3BC83E38183E3B3, 0xE300B682E300A083, 0x83E3B782E300B782,
-    0x00B082E3B383E3AA, 0x00B982E300B882E3, 0x00BB82E300BA82E3,
-    0x83E3B383E3BB82E3, 0xB383E3BB82E30081, 0x00BC82E3008883E3,
-    0x00BE82E300BD82E3, 0x008083E300BF82E3, 0x82E3BC83E38083E3,
-    0x83E3008183E300B9, 0x83E3008383E30082, 0x83E3008583E30084,
-    0x83E3008783E30086, 0x8883E300B782E387, 0x00B383E38883E300,
-    0xE38983E3008983E3, 0xE3008A83E300AB83, 0x83E3008E83E38A83,
-    0x83E3008C83E3008B, 0x83E3008E83E3008D, 0x008883E38383E38E,
-    0xE38F83E3008F83E3, 0x83E3008483E3A482, 0xBC83E39083E30090,
-    0xE300AB83E3AC83E3, 0x83E39183E3009183, 0xE3B383E3BB82E3BC,
-    0x83E39183E3008883, 0x9283E3008483E3BC, 0x9383E3009383E300,
-    0x009483E300AB83E3, 0x82E3A282E39483E3, 0x00AB83E38883E3B9,
-    0x83E3AF82E39483E3, 0xB382E39483E300AB, 0x9583E3009583E300,
-    0x83E3A983E3A182E3, 0x9583E3008983E383, 0x83E3BC83E3A382E3,
-    0xA983E39583E30088, 0x009683E300B383E3, 0x82E38383E39683E3,
-    0x00AB83E3A782E3B7, 0x009883E3009783E3, 0x82E3AF82E39883E3,
-    0x00AB83E3BC83E3BF, 0x83E3AB83E39883E3, 0x83E3009983E30084,
-    0x00BF82E3BC83E399, 0xE39A83E3009A83E3, 0x83E39A83E300BD82,
-    0x9A83E3009283E38B, 0xE300B982E3B383E3, 0xB882E3BC83E39A83,
-    0x9B83E3009B83E300, 0xE39B83E300B383E3, 0x83E300AB83E3BC83,
-    0x00B383E3BC83E39B, 0xE39C83E3009C83E3, 0x83E3008883E3AB83,
-    0xA482E39D83E3009D, 0xE3008883E3B383E3, 0x8983E3B383E39D83,
-    0x9E83E3009E83E300, 0x83E3AF82E3A482E3, 0xA482E39E83E300AD,
-    0xE39E83E300AB83E3, 0x83E3008F83E38383, 0x00AF82E3AB83E39E,
-    0x82E3B383E39E83E3, 0x00B383E3A783E3B7, 0xE39F83E3009F83E3,
-    0xB383E3AD83E3AF82, 0x00AA83E39F83E300, 0x83E3AA83E39F83E3,
-    0x00AB83E3BC83E390, 0x00A183E300A083E3, 0xE300AC82E3A183E3,
-    0x8883E3AC82E3A183, 0xE3A183E300B383E3, 0xAB83E38883E3BC83,
-    0xA383E300A283E300, 0xA483E300A483E300, 0xE3008983E3BC83E3,
-    0xAB83E3BC83E3A483, 0xA683E300A583E300, 0xE3A282E3A683E300,
-    0xE300A783E300B383, 0xE300A983E300A883, 0x83E3AA83E300AA83,
-    0x00AB83E38883E383, 0xE300A983E3AA83E3, 0x83E3AB83E300AB83,
-    0xAB83E300BC83E394, 0x83E39683E3BC83E3, 0x83E300AC83E300AB,
-    0xAC83E300A083E3AC, 0x82E38883E3B383E3, 0xAD83E300B383E3B2,
-    0xAF83E300AF83E300, 0xE3008883E38383E3, 0xE300B183E300B083,
-    0xE300B383E300B283, 0xE300B783E300B483, 0xE300B983E300B883,
-    0xE300BB83E300BA83, 0xE300BE83E300BC83, 0xE300B992E3009E92,
-    0xE3009F93E300BB92, 0xE300AE9BE3009594, 0xE300819EE300BC9B,
-    0xE300A2A1E300AFA0, 0xE30087A3E300BCA1, 0xE3009CA4E300A3A3,
-    0xE300AEA8E300BAA4, 0xE300A4ABE300ACA9, 0xE30099ACE30088AC,
-    0xE3009DAEE30089AD, 0xE3008EB1E30098B0, 0xE30096B6E300B3B4,
-    0xE300B8BAE300ACBA, 0xE400BCBFE3009BBC, 0xE4009880E4008880,
-    0xE4008681E400B980, 0xE400A383E4009682, 0xE4008288E400AF84,
-    0xE400A08AE400A788, 0xE400B48CE400818C, 0xE400958FE400998D,
-    0xE4008B90E400998F, 0xE400AB94E400AB91, 0xE400A195E4009D95,
-    0xE4009797E400AB95, 0xE400B598E400B997, 0xE400879BE400BE9A,
-    0xE400A6A7E40095A6, 0xE400B6A9E400AEA9, 0xE400B3ACE400B2AA,
-    0xE4008EB3E4008EAF, 0xE400B8B3E400ADB3, 0xE40080B8E40096B5,
-    0xE40083B8E40081B8, 0xE4008AB8E40089B8, 0xE4008DB8E4008BB8,
-    0xE400A6B8E40099B8, 0xE400ADB8E400A8B8, 0xE400B6B8E400B2B8,
-    0xE400B9B8E400B8B8, 0xE400BFB8E400BDB8, 0xE40099B9E40081B9,
-    0xE40082BAE4009DB9, 0xE40086BAE40085BA, 0xE40094BAE4008CBA,
-    0xE400A4BAE400A0BA, 0xE400BABAE400AEBA, 0xE4008CBBE40080BB,
-    0x92E5A4BBE400A4BB, 0xBCE40081BCE4008C, 0xBEE400A0BDE40091,
-    0xBEE40086BEE40080, 0xBEE400AEBEE4008B, 0x80E500BFBEE400BB,
-    0x81E500AB80E50082, 0x83E5009982E500BA, 0x83E5009A83E5008F,
-    0x84E500AA84E500A7, 0x85E5008085E500BF, 0x85E5008D85E50085,
-    0x85E500A485E50094, 0x85E500A785E500A5, 0x85E500A985E500A8,
-    0x85E500AD85E500AB, 0x86E5008086E500B7, 0x86E5008D86E50082,
-    0x86E5009586E50092, 0x86E5009786E50096, 0x86E500A486E50099,
-    0x86E500AC86E500AB, 0x87E500B786E500B5, 0x87E5008C87E50089,
-    0x87E5009E87E5009C, 0x88E500B587E500A0, 0x88E5008388E50080,
-    0x88E5009788E50087, 0x88E500A988E5009D, 0x89E500BB88E500BA,
-    0x89E5008D89E50086, 0x8AE500B789E500B2, 0x8AE5009B8AE50089,
-    0x8AE500B38AE500A3, 0x8BE500878BE500B4, 0x8BE500928BE50089,
-    0x8BE500A48BE5009E, 0x8BE500B98BE500B5, 0x8CE500858CE500BA,
-    0x8CE500958CE50086, 0x8CE5009A8CE50097, 0x8CE500BB8CE500B8,
-    0x8DE500818DE500BF, 0x8DE500858DE50084, 0x8DE500918DE50089,
-    0x8DE5009A8DE50094, 0x8DE500A98DE5009C, 0x8DE500B38DE500B0,
-    0x8DE500BD8DE500B5, 0x8EE500828EE500BF, 0x8FE500838FE500B6,
-    0x8FE5008A8FE50088, 0x8FE5009F8FE5008C, 0x8FE500A58FE500A3,
-    0x8FE500AF8FE500AB, 0x90E500B38FE500B1, 0x90E5008890E50086,
-    0x90E5008F90E5008D, 0x90E500B890E5009D, 0x91E5008291E500B9,
-    0x92E500A891E50088, 0x92E500A292E5009E, 0x94E500B693E500BD,
-    0x95E5008F95E50090, 0x95E5009595E50093, 0x96E5008496E500A3,
-    0x96E5009996E50087, 0x96E500AB96E5009D, 0x97E500B696E500B3,
-    0x97E5008297E50080, 0x99E5008698E500A2, 0x99E500A899E50091,
-    0x9BE500979BE500B4, 0x9CE500B99BE5009B, 0x9CE500979CE50096,
-    0x9EE500B09CE5009F, 0x9FE5008E9FE5008B, 0xA0E5008DA0E500B4,
-    0xA1E500B2A0E500B1, 0xA1E5009AA1E50080, 0xA2E500A8A2E5009E,
-    0xA3E500B3A2E500AC, 0xA3E5009FA3E50098, 0xA3E500AEA3E500AB,
-    0xA3E500B2A3E500B0, 0xA4E50082A4E500B7, 0xA4E5008AA4E50086,
-    0xA4E5009AA4E50095, 0xA4E500A2A4E5009C, 0xA3ADE6A7A4E500A7,
-    0x84A5E500A9A4E500, 0x91A5E50088A5E500, 0xA2A5E50094A5E500,
-    0x98A7E500B3A5E500, 0x9BA8E500ACA7E500, 0xA2A9E500A7A8E500,
-    0xB5AAE500A6A9E500, 0xA8ACE50088ACE500, 0x90ADE500BEACE500,
-    0xA6ADE50097ADE500, 0x85AEE50080AEE500, 0x83AFE50097AEE500,
-    0xA7AFE50098AFE500, 0xB3AFE500AEAFE500, 0xBFAFE500B8AFE500,
-    0x8FB0E50086B0E500, 0xB8B0E500A2B0E500, 0xA0B1E500BFB0E500,
-    0xA4B1E500A2B1E500, 0xAEB1E500A5B1E500, 0x8DB2E500B1B1E500,
-    0x99B4E50080B3E500, 0x90B5E50083B5E500, 0xAEB5E500ABB5E500,
-    0xB2B6E500BCB5E500, 0x9BB7E500BAB6E500, 0xA2B7E500A1B7E500,
-    0xA6B7E500A5B7E500, 0xBDB7E500B1B7E500, 0xA8B8E500BEB7E500,
-    0xA9B9E500BDB8E500, 0xB3B9E500B2B9E500, 0x00B4B9E5009088E6,
-    0x00BCB9E500BAB9E5, 0x00A6BAE500BFB9E5, 0x00B3BAE500B0BAE5,
-    0x0089BBE500B6BAE5, 0x0092BBE5008ABBE5, 0x0099BBE50093BBE5,
-    0x00B4BBE500ACBBE5, 0x0084BCE500BEBBE5, 0x0093BCE5008BBCE5,
-    0x0090BDE500A2BCE5, 0x00A1BDE50093BDE5, 0x00A9BDE500A2BDE5,
-    0x00B3BDE500ABBDE5, 0x008CBEE5008BBEE5, 0x009ABEE50097BEE5,
-    0x00ADBEE500A9BEE5, 0x008DBFE50083BFE5, 0x00B5BFE50097BFE5,
-    0x009280E600B9BFE5, 0x00B581E6009C80E6, 0x009482E6008182E6,
-    0x009883E6008783E6, 0x008884E600A183E6, 0x008885E6008485E6,
-    0x008E85E6008C85E6, 0x00A885E600A085E6, 0x008E86E600BA85E6,
-    0x00A486E6009086E6, 0x00B286E600AF86E6, 0x00B287E6009E87E6,
-    0x008088E600B687E6, 0x009088E6008888E6, 0x00AE88E6009B88E6,
-    0x00B688E600B488E6, 0x009389E6008B89E6, 0x00958AE6009D89E6,
-    0x00898BE600B18AE6, 0x00938BE6008F8BE6, 0x00BC8BE600948BE6,
-    0x00878CE600BE8BE6, 0x00908DE600BD8CE6, 0x00A88DE600958DE6,
-    0x00838EE600BB8DE6, 0x00A98EE600A08EE6, 0x00858FE600848FE6,
-    0x009C90E600A48FE6, 0x009291E600A290E6, 0x00B791E600A991E6,
-    0x009A92E600BE91E6, 0x008493E6009D92E6, 0x00B494E600AF94E6,
-    0x009695E6008F95E6, 0x00B895E600AC95E6, 0x009796E6008796E6,
-    0x00A496E6009996E6, 0x00B996E600B096E6, 0x00A097E6008597E6,
-    0x00A397E600A297E6, 0xE68E98E600A597E6, 0xE6009398E600BBB2,
-    0x92E5AD98E600A098, 0x99E6008999E6008C, 0x9AE600889AE600B4,
-    0x9AE6009C9AE60091, 0x9BE600869BE600B4, 0x9BE600B49BE600B0,
-    0x9CE600809CE600B8, 0x9CE600899CE60088, 0x9CE6009B9CE60097,
-    0x9DE600A89CE600A1, 0x9DE600939DE6008E, 0x9DE6009E9DE60096,
-    0x9EE600859EE600BB, 0x9FE600B39FE60097, 0xA0E60097A0E600BA,
-    0xA0E600AAA0E6009F, 0xE79ABCE48FBCE5AA, 0xE60092A1E600BEA4,
-    0xE60085A2E60081A2, 0xE600A8A2E6008EA2, 0xE60082A5E60094A4,
-    0xE600AAA7E600A3A6, 0xE60093A8E60082A8, 0xE60093ABE600A8AA,
-    0xE60084ACE6009BAB, 0xE600A1ACE600A0AC, 0xE600A2ADE60094AD,
-    0xE600B2ADE600A3AD, 0xE600B9ADE600B7AD, 0xE600AEAEE6009FAE,
-    0xE600BAAEE600B3AE, 0xE6008BAFE600BBAE, 0xE60094AFE6008DAF,
-    0xE6008FB0E6009BAF, 0xE600B4B0E60094B0, 0xE600A7B1E6008EB1,
-    0xE600BFB2E60088B2, 0xE6008DB3E6008CB3, 0xE600A8B3E600A5B3,
-    0xE6009BB4E60096B4, 0xE600B4B4E6009EB4, 0xE60081B5E600BEB4,
-    0xE600AAB5E600A9B5, 0xE600B8B5E600B7B5, 0xE6008BB7E60085B6,
-    0xE600AAB7E6009AB7, 0xE6009AB8E600B9B7, 0xE600AEB9E600AFB8,
-    0xE6009CBAE60080BA, 0xE60087BBE600BABA, 0xE60091BBE6008BBB,
-    0xE6008FBCE6009BBB, 0xE600A2BCE60094BC, 0xE600AEBDE600A3BC,
-    0xE600ABBFE60086BF, 0xE7009B80E700BEBF, 0xE700B980E7009E80,
-    0xE700AB81E7008A81, 0xE700B781E700B081, 0xE7009982E700BD81,
-    0xE7008883E700AD82, 0xE700A184E7009983, 0xE7008985E7008585,
-    0xE7009C86E700AE85, 0xE7009087E7008E87, 0xE7009B88E7009088,
-    0xE700AA88E700A888, 0xE700B588E700AB88, 0xE700BB88E700B688,
-    0xE7008789E700BF88, 0xE7009989E7009089, 0xE700A289E7009B89,
-    0xE700808AE700B989, 0xE700AC8AE700958A, 0xE700808BE700AF8A,
-    0xE700AA8CE700BC8B, 0xE700BA8DE700B58D, 0xE700878EE700848E,
-    0xE7008B8EE700898E, 0xE700B28EE700A58E, 0xE7008690E7009E8F,
-    0xE700A290E7008990, 0xE7009C91E7008791, 0xE700B191E700A991,
-    0xE7008992E7008592, 0xE7008A93E7009892, 0xE700A693E7009C93,
-    0xE7009894E7008694, 0xE700A494E7009F94, 0xE700B094E700A894,
-    0xE700B394E700B294, 0xE700BB94E700B794, 0xE7009995E700BE94,
-    0xE700B095E700A595, 0xE7009296E7008B96, 0xE7009098E700A297,
-    0xE7009F98E7009D98, 0xE700A999E7008299, 0xE700BD99E700B699,
-    0xE700BF9AE700AE9A, 0xE7009B9BE7008A9B, 0xE700A79BE700A39B,
-    0xE700B49BE700AE9B, 0xE7009E9CE700819C, 0xE700809DE7009F9C,
-    0xE7008B9EE7008A9D, 0xE7009B9FE700A79E, 0xE700B39FE700A29F,
-    0xE700ABA1E7008EA1, 0xE70091A2E7008CA2, 0xE7008CA3E7008AA3,
-    0xE700AAA4E700BBA3, 0xE700BCA4E700BAA4, 0xE70088A5E700BEA4,
-    0xE70090A5E70089A5, 0xE7009DA5E70096A5, 0xE700A5A5E7009EA5,
-    0xE70081A6E700BFA5, 0xE7008EA6E7008DA6, 0xE700AEA6E7008FA6,
-    0xE700BEA6E700B8A6, 0xE70098A7E7008AA7, 0xE7009CA8E700ABA7,
-    0xE7008AA9E70080A9, 0xE700B4A9E7008FA9, 0xE70081AAE700BAA9,
-    0xE7008BABE700B1AA, 0xE700B9ABE700AEAB, 0xE7008FAEE700A0AC,
-    0xE70086AFE70080AF, 0xE700BEB0E70089AF, 0xE700B3B1E700A0B1,
-    0xE70092B2E700BBB1, 0xE70092B3E700BEB2, 0xE700A3B3E70096B3,
-    0xE700A8B3E700A7B3, 0xE70080B4E700B8B3, 0xE700A2B4E70090B4,
-    0xE70082B5E700AFB4, 0xE700A3B5E7009BB5, 0xE700BEB6E700A0B6,
-    0xE700B4B7E70087B7, 0xE70089B8E70082B8, 0xE70081B9E700B7B8,
-    0xE700B6BCE70085B9, 0xE70091BDE700BEBC, 0xE700B9BDE700B2BD,
-    0xE70085BEE700BABD, 0xE70095BEE7008ABE, 0xE700BDBEE7009ABE,
-    0xE8008180E800BABF, 0xE8008C80E8008580, 0xE800B380E8009280,
-    0xE800A081E8008681, 0xE800B081E800AF81, 0xE800BF81E800BE81,
-    0xE8008B82E8008982, 0xE800B282E800AD82, 0xE800BE84E8008384,
-    0xE800A387E8009887, 0xE800AA87E800A887, 0xE800B387E800AD87,
-    0xE8008188E800BC87, 0xE8008C88E8008488, 0xE8009B88E8009888,
-    0xE800AE89E8009F88, 0xE800B289E800AF89, 0xE800B989E800B889,
-    0xE800918AE8008B8A, 0xE800B18AE8009D8A, 0xE800BD8AE800B38A,
-    0xE800A68BE800A58B, 0xE800A38CE8009D8C, 0xE800928DE800B68C,
-    0xE800A38DE800938D, 0xE800BD8EE800AD8E, 0xE8008A8FE800898F,
-    0xE8009C8FE8008C8F, 0xE800AF8FE800A78F, 0xE800BD90E800B18F,
-    0xE8009791E8008991, 0xE800B193E800AE93, 0xE800BC93E800B393,
-    0xE800A495E8009694, 0xE800BA97E8008D97, 0xE8009298E8008698,
-    0xE800BF98E800AD98, 0xE8009099E8008D99, 0xE800A799E8009C99,
-    0xE800AB99E800A999, 0xE800A99AE800889A, 0xE8008E9CE800A29B,
-    0xE800AB9DE800A89C, 0xE800869EE800B99D, 0xE800A19FE800BA9E,
-    0xE8009FA0E80081A0, 0xE8008CA1E80080A1, 0xE800A3A1E800A0A1,
-    0xE8008FA3E80082A3, 0xE8009EA3E80097A3, 0xE800B8A3E800A1A3,
-    0xE80090A4E800BAA3, 0xE800A4A5E80081A5, 0xE80086A6E800BEA5,
-    0xE80096A6E8008BA6, 0xE800A3A7E80092A7, 0xE800A0AAE80080A8,
-    0xE800BFAAE800AAAA, 0xE80092ABE8008BAB, 0xE800ADABE80096AB,
-    0xE800BEABE800B8AB, 0xE800B9ACE80081AC, 0xE80080AEE80098AD,
-    0xE800B7B0E8008AAE, 0xE80088B1E80086B1, 0xE800B8B1E80095B1,
-    0xE800A1B2E8009DB2, 0xE800ABB2E800A9B2, 0xE80082B3E80081B3,
-    0xE80088B3E80087B3, 0xE80088B4E80093B3, 0xE800A4B5E8009BB4,
-    0xE800B7B5E800B0B5, 0xE800BCB6E800B3B6, 0xE800AFB7E8008BB7,
-    0xE800ABBAE800B0B7, 0xE80094BBE8008ABB, 0xE800AABCE800A6BC,
-    0xE800BBBCE800B8BC, 0xE8009BBEE800A2BD, 0xE800B0BEE8009EBE,
-    0xE900B6BEE800B5BE, 0xE900B880E900A380, 0xE900A981E9008A81,
-    0xE900BC81E900B281, 0xE9009182E9008F82, 0xE9008E83E9009482,
-    0xE900B183E9009E83, 0xE9009184E900BD83, 0xE9008985E9009B84,
-    0xE900AA85E9008D85, 0xE900B486E9009986, 0xE9008C87E9008687,
-    0xE9009187E9008F87, 0xE900B888E900B488, 0xE900BC89E900B689,
-    0xE900988BE900978B, 0xE9008A8DE900848C, 0xE9009590E900B98F,
-    0xE9008096E900B795, 0xE900AD96E9008B96, 0xE9009C98E900B796,
-    0xE9008B99E900AE98, 0xE900B599E9008D99, 0xE900BC99E900B899,
-    0xE900A39AE900869A, 0xE900B79AE900B69A, 0xE900B99AE900B89A,
-    0xE900A29BE900839B, 0xE900A89BE900A39B, 0xE900B79BE900B69B,
-    0xE900B29CE900A39C, 0xE900919DE900889D, 0xE9009E9DE900969D,
-    0xE900A99DE900A29D, 0xE9009B9FE9008B9F, 0xE900AD9FE900A09F,
-    0xE900BF9FE900B39F, 0xE90085A0E90081A0, 0xE90098A0E9008BA0,
-    0xE900BBA0E900A9A0, 0xE900A8A2E9009EA1, 0xE9009FA3E9009BA3,
-    0xE900AFA3E900A2A3, 0xE900A8A4E900BCA3, 0xE90096A6E900A9A4,
-    0xE900A7A6E90099A6, 0xE90082A7E900ACA6, 0xE900BEA7E900B1A7,
-    0xE900A8AAE900AAA9, 0xE9009FABE90098AB, 0xE900A5ACE90092AC,
-    0xE900B2ACE900AFAC, 0xE9009AADE900BCAC, 0xE90080B1E900AFAD,
-    0xE900A5B3E90097B1, 0xE900A7B5E900BDB3, 0xE900BAB7E900B4B6,
-    0xE900B5B9E9009EB8, 0xE90097BAE900BFB9, 0xE900A5BAE9009FBA,
-    0xE90083BBE900BBBA, 0xE9008EBBE9008DBB, 0xE900B9BBE90091BB,
-    0xE900BEBBE900BDBB, 0xE9008EBCE90085BC, 0xE90093BCE9008FBC,
-    0xE900A0BCE90096BC, 0xE90083BDE900BBBC, 0xE90092BDE9008ABD,
-    0xE9008EBEE9008DBE, 0xE9009FBEE9009CBE, 0xEA008199EA00A0BE,
-    0xEA008599EA008399, 0xEA008999EA008799, 0xEA008D99EA008B99,
-    0xEA009199EA008F99, 0xEA009599EA009399, 0xEA009999EA009799,
-    0xEA009D99EA009B99, 0xEA00A199EA009F99, 0xEA00A599EA00A399,
-    0xEA00A999EA00A799, 0xEA00AD99EA00AB99, 0xEA00839AEA00819A,
-    0xEA00879AEA00859A, 0xEA008B9AEA00899A, 0xEA008F9AEA008D9A,
-    0xEA00939AEA00919A, 0xEA00979AEA00959A, 0xEA009B9AEA00999A,
-    0xEA00A59CEA00A39C, 0xEA00A99CEA00A79C, 0xEA00AD9CEA00AB9C,
-    0xEA00B39CEA00AF9C, 0xEA00B79CEA00B59C, 0xEA00BB9CEA00B99C,
-    0xEA00BF9CEA00BD9C, 0xEA00839DEA00819D, 0xEA00879DEA00859D,
-    0xEA008B9DEA00899D, 0xEA008F9DEA008D9D, 0xEA00939DEA00919D,
-    0xEA00979DEA00959D, 0xEA009B9DEA00999D, 0xEA009F9DEA009D9D,
-    0xEA00A39DEA00A19D, 0xEA00A79DEA00A59D, 0xEA00AB9DEA00A99D,
-    0xEA00AF9DEA00AD9D, 0xEA00BC9DEA00BA9D, 0xEA00819EEA00BF9D,
-    0xEA00859EEA00839E, 0xEA008C9EEA00879E, 0xEA00919EEA008E9E,
-    0xEA00949EEA00939E, 0xEA00999EEA00979E, 0xEA009D9EEA009B9E,
-    0xEA00A19EEA009F9E, 0xEA00A59EEA00A39E, 0xEA00A99EEA00A79E,
-    0xEA00B79EEA00B59E, 0xEA00BB9EEA00B99E, 0xEA00BF9EEA00BD9E,
-    0xEA00839FEA00819F, 0xEA008A9FEA00889F, 0xEA00979FEA00919F,
-    0xEA00B69FEA00999F, 0xEA0092ADEA00B7AC, 0xEA00A6ADEA0093AD,
-    0xEA0080B0EA00A7AD, 0xEA0082B0EA0081B0, 0xEA0084B0EA0083B0,
-    0xEA0086B0EA0085B0, 0xEA0088B0EA0087B0, 0xEA008AB0EA0089B0,
-    0xEA008CB0EA008BB0, 0xEA008EB0EA008DB0, 0xEA0090B0EA008FB0,
-    0xEA0092B0EA0091B0, 0xEA0094B0EA0093B0, 0xEA0096B0EA0095B0,
-    0xEA0098B0EA0097B0, 0xEA009AB0EA0099B0, 0xEA009CB0EA009BB0,
-    0xEA009EB0EA009DB0, 0xEA00A0B0EA009FB0, 0xEA00A2B0EA00A1B0,
-    0xEA00A4B0EA00A3B0, 0xEA00A6B0EA00A5B0, 0xEA00A8B0EA00A7B0,
-    0xEA00AAB0EA00A9B0, 0xEA00ACB0EA00ABB0, 0xEA00AEB0EA00ADB0,
-    0xEA00B0B0EA00AFB0, 0xEA00B2B0EA00B1B0, 0xEA00B4B0EA00B3B0,
-    0xEA00B6B0EA00B5B0, 0xEA00B8B0EA00B7B0, 0xEA00BAB0EA00B9B0,
-    0xEA00BCB0EA00BBB0, 0xEA00BEB0EA00BDB0, 0xEA0080B1EA00BFB0,
-    0xEA0082B1EA0081B1, 0xEA0084B1EA0083B1, 0xEA0086B1EA0085B1,
-    0xEA0088B1EA0087B1, 0xEA008AB1EA0089B1, 0xEA008CB1EA008BB1,
-    0xEA008EB1EA008DB1, 0xEA0090B1EA008FB1, 0xEA0092B1EA0091B1,
-    0xEA0094B1EA0093B1, 0xEA0096B1EA0095B1, 0xEA0098B1EA0097B1,
-    0xEA009AB1EA0099B1, 0xEA009CB1EA009BB1, 0xEA009EB1EA009DB1,
-    0xEA00A0B1EA009FB1, 0xEA00A2B1EA00A1B1, 0xEA00A4B1EA00A3B1,
-    0xEA00A6B1EA00A5B1, 0xEA00A8B1EA00A7B1, 0xEA00AAB1EA00A9B1,
-    0xEA00ACB1EA00ABB1, 0xEA00AEB1EA00ADB1, 0xEA00B0B1EA00AFB1,
-    0xEA00B2B1EA00B1B1, 0xEA00B4B1EA00B3B1, 0xEA00B6B1EA00B5B1,
-    0xEA00B8B1EA00B7B1, 0xEA00BAB1EA00B9B1, 0xEA00BCB1EA00BBB1,
-    0xEA00BEB1EA00BDB1, 0xEA0080B2EA00BFB1, 0xEA0082B2EA0081B2,
-    0xEA0084B2EA0083B2, 0xEA0086B2EA0085B2, 0xEA0088B2EA0087B2,
-    0xEA008AB2EA0089B2, 0xEA008CB2EA008BB2, 0xEA008EB2EA008DB2,
-    0xEA0090B2EA008FB2, 0xEA0092B2EA0091B2, 0xEA0094B2EA0093B2,
-    0xEA0096B2EA0095B2, 0xEA0098B2EA0097B2, 0xEA009AB2EA0099B2,
-    0xEA009CB2EA009BB2, 0xEA009EB2EA009DB2, 0xEA00A0B2EA009FB2,
-    0xEA00A2B2EA00A1B2, 0xEA00A4B2EA00A3B2, 0xEA00A6B2EA00A5B2,
-    0xEA00A8B2EA00A7B2, 0xEA00AAB2EA00A9B2, 0xEA00ACB2EA00ABB2,
-    0xEA00AEB2EA00ADB2, 0xEA00B0B2EA00AFB2, 0xEA00B2B2EA00B1B2,
-    0xEA00B4B2EA00B3B2, 0xEA00B6B2EA00B5B2, 0xEA00B8B2EA00B7B2,
-    0xEA00BAB2EA00B9B2, 0xEA00BCB2EA00BBB2, 0xEA00BEB2EA00BDB2,
-    0xEA0080B3EA00BFB2, 0xEA0082B3EA0081B3, 0xEA0084B3EA0083B3,
-    0xEA0086B3EA0085B3, 0xEA0088B3EA0087B3, 0xEA008AB3EA0089B3,
-    0xEA008CB3EA008BB3, 0xEA008EB3EA008DB3, 0xEA0090B3EA008FB3,
-    0xEA0092B3EA0091B3, 0xEA0094B3EA0093B3, 0xEA0096B3EA0095B3,
-    0xEA0098B3EA0097B3, 0xEA009AB3EA0099B3, 0xEA009CB3EA009BB3,
-    0xEA009EB3EA009DB3, 0xEA00A0B3EA009FB3, 0xEA00A2B3EA00A1B3,
-    0xEA00A4B3EA00A3B3, 0xEA00A6B3EA00A5B3, 0xEA00A8B3EA00A7B3,
-    0xEA00AAB3EA00A9B3, 0xEA00ACB3EA00ABB3, 0xEA00AEB3EA00ADB3,
-    0xEA00B0B3EA00AFB3, 0xEA00B2B3EA00B1B3, 0xEA00B4B3EA00B3B3,
-    0xEA00B6B3EA00B5B3, 0xEA00B8B3EA00B7B3, 0xEA00BAB3EA00B9B3,
-    0xEA00BCB3EA00BBB3, 0xEA00BEB3EA00BDB3, 0xEA0080B4EA00BFB3,
-    0xEA0082B4EA0081B4, 0xEA0084B4EA0083B4, 0xEA0086B4EA0085B4,
-    0xEA0088B4EA0087B4, 0xEA008AB4EA0089B4, 0xEA008CB4EA008BB4,
-    0xEA008EB4EA008DB4, 0xEA0090B4EA008FB4, 0xEA0092B4EA0091B4,
-    0xEA0094B4EA0093B4, 0xEA0096B4EA0095B4, 0xEA0098B4EA0097B4,
-    0xEA009AB4EA0099B4, 0xEA009CB4EA009BB4, 0xEA009EB4EA009DB4,
-    0xEA00A0B4EA009FB4, 0xEA00A2B4EA00A1B4, 0xEA00A4B4EA00A3B4,
-    0xEA00A6B4EA00A5B4, 0xEA00A8B4EA00A7B4, 0xEA00AAB4EA00A9B4,
-    0xEA00ACB4EA00ABB4, 0xEA00AEB4EA00ADB4, 0xEA00B0B4EA00AFB4,
-    0xEA00B2B4EA00B1B4, 0xEA00B4B4EA00B3B4, 0xEA00B6B4EA00B5B4,
-    0xEA00B8B4EA00B7B4, 0xEA00BAB4EA00B9B4, 0xEA00BCB4EA00BBB4,
-    0xEA00BEB4EA00BDB4, 0xEA0080B5EA00BFB4, 0xEA0082B5EA0081B5,
-    0xEA0084B5EA0083B5, 0xEA0086B5EA0085B5, 0xEA0088B5EA0087B5,
-    0xEA008AB5EA0089B5, 0xEA008CB5EA008BB5, 0xEA008EB5EA008DB5,
-    0xEA0090B5EA008FB5, 0xEA0092B5EA0091B5, 0xEA0094B5EA0093B5,
-    0xEA0096B5EA0095B5, 0xEA0098B5EA0097B5, 0xEA009AB5EA0099B5,
-    0xEA009CB5EA009BB5, 0xEA009EB5EA009DB5, 0xEA00A0B5EA009FB5,
-    0xEA00A2B5EA00A1B5, 0xEA00A4B5EA00A3B5, 0xEA00A6B5EA00A5B5,
-    0xEA00A8B5EA00A7B5, 0xEA00AAB5EA00A9B5, 0xEA00ACB5EA00ABB5,
-    0xEA00AEB5EA00ADB5, 0xEA00B0B5EA00AFB5, 0xEA00B2B5EA00B1B5,
-    0xEA00B4B5EA00B3B5, 0xEA00B6B5EA00B5B5, 0xEA00B8B5EA00B7B5,
-    0xEA00BAB5EA00B9B5, 0xEA00BCB5EA00BBB5, 0xEA00BEB5EA00BDB5,
-    0xEA0080B6EA00BFB5, 0xEA0082B6EA0081B6, 0xEA0084B6EA0083B6,
-    0xEA0086B6EA0085B6, 0xEA0088B6EA0087B6, 0xEA008AB6EA0089B6,
-    0xEA008CB6EA008BB6, 0xEA008EB6EA008DB6, 0xEA0090B6EA008FB6,
-    0xEA0092B6EA0091B6, 0xEA0094B6EA0093B6, 0xEA0096B6EA0095B6,
-    0xEA0098B6EA0097B6, 0xEA009AB6EA0099B6, 0xEA009CB6EA009BB6,
-    0xEA009EB6EA009DB6, 0xEA00A0B6EA009FB6, 0xEA00A2B6EA00A1B6,
-    0xEA00A4B6EA00A3B6, 0xEA00A6B6EA00A5B6, 0xEA00A8B6EA00A7B6,
-    0xEA00AAB6EA00A9B6, 0xEA00ACB6EA00ABB6, 0xEA00AEB6EA00ADB6,
-    0xEA00B0B6EA00AFB6, 0xEA00B2B6EA00B1B6, 0xEA00B4B6EA00B3B6,
-    0xEA00B6B6EA00B5B6, 0xEA00B8B6EA00B7B6, 0xEA00BAB6EA00B9B6,
-    0xEA00BCB6EA00BBB6, 0xEA00BEB6EA00BDB6, 0xEA0080B7EA00BFB6,
-    0xEA0082B7EA0081B7, 0xEA0084B7EA0083B7, 0xEA0086B7EA0085B7,
-    0xEA0088B7EA0087B7, 0xEA008AB7EA0089B7, 0xEA008CB7EA008BB7,
-    0xEA008EB7EA008DB7, 0xEA0090B7EA008FB7, 0xEA0092B7EA0091B7,
-    0xEA0094B7EA0093B7, 0xEA0096B7EA0095B7, 0xEA0098B7EA0097B7,
-    0xEA009AB7EA0099B7, 0xEA009CB7EA009BB7, 0xEA009EB7EA009DB7,
-    0xEA00A0B7EA009FB7, 0xEA00A2B7EA00A1B7, 0xEA00A4B7EA00A3B7,
-    0xEA00A6B7EA00A5B7, 0xEA00A8B7EA00A7B7, 0xEA00AAB7EA00A9B7,
-    0xEA00ACB7EA00ABB7, 0xEA00AEB7EA00ADB7, 0xEA00B0B7EA00AFB7,
-    0xEA00B2B7EA00B1B7, 0xEA00B4B7EA00B3B7, 0xEA00B6B7EA00B5B7,
-    0xEA00B8B7EA00B7B7, 0xEA00BAB7EA00B9B7, 0xEA00BCB7EA00BBB7,
-    0xEA00BEB7EA00BDB7, 0xEA0080B8EA00BFB7, 0xEA0082B8EA0081B8,
-    0xEA0084B8EA0083B8, 0xEA0086B8EA0085B8, 0xEA0088B8EA0087B8,
-    0xEA008AB8EA0089B8, 0xEA008CB8EA008BB8, 0xEA008EB8EA008DB8,
-    0xEA0090B8EA008FB8, 0xEA0092B8EA0091B8, 0xEA0094B8EA0093B8,
-    0xEA0096B8EA0095B8, 0xEA0098B8EA0097B8, 0xEA009AB8EA0099B8,
-    0xEA009CB8EA009BB8, 0xEA009EB8EA009DB8, 0xEA00A0B8EA009FB8,
-    0xEA00A2B8EA00A1B8, 0xEA00A4B8EA00A3B8, 0xEA00A6B8EA00A5B8,
-    0xEA00A8B8EA00A7B8, 0xEA00AAB8EA00A9B8, 0xEA00ACB8EA00ABB8,
-    0xEA00AEB8EA00ADB8, 0xEA00B0B8EA00AFB8, 0xEA00B2B8EA00B1B8,
-    0xEA00B4B8EA00B3B8, 0xEA00B6B8EA00B5B8, 0xEA00B8B8EA00B7B8,
-    0xEA00BAB8EA00B9B8, 0xEA00BCB8EA00BBB8, 0xEA00BEB8EA00BDB8,
-    0xEA0080B9EA00BFB8, 0xEA0082B9EA0081B9, 0xEA0084B9EA0083B9,
-    0xEA0086B9EA0085B9, 0xEA0088B9EA0087B9, 0xEA008AB9EA0089B9,
-    0xEA008CB9EA008BB9, 0xEA008EB9EA008DB9, 0xEA0090B9EA008FB9,
-    0xEA0092B9EA0091B9, 0xEA0094B9EA0093B9, 0xEA0096B9EA0095B9,
-    0xEA0098B9EA0097B9, 0xEA009AB9EA0099B9, 0xEA009CB9EA009BB9,
-    0xEA009EB9EA009DB9, 0xEA00A0B9EA009FB9, 0xEA00A2B9EA00A1B9,
-    0xEA00A4B9EA00A3B9, 0xEA00A6B9EA00A5B9, 0xEA00A8B9EA00A7B9,
-    0xEA00AAB9EA00A9B9, 0xEA00ACB9EA00ABB9, 0xEA00AEB9EA00ADB9,
-    0xEA00B0B9EA00AFB9, 0xEA00B2B9EA00B1B9, 0xEA00B4B9EA00B3B9,
-    0xEA00B6B9EA00B5B9, 0xEA00B8B9EA00B7B9, 0xEA00BAB9EA00B9B9,
-    0xEA00BCB9EA00BBB9, 0xEA00BEB9EA00BDB9, 0xEA0080BAEA00BFB9,
-    0xEA0082BAEA0081BA, 0xEA0084BAEA0083BA, 0xEA0086BAEA0085BA,
-    0xEA0088BAEA0087BA, 0xEA008ABAEA0089BA, 0xEA008CBAEA008BBA,
-    0xEA008EBAEA008DBA, 0xEA0090BAEA008FBA, 0xEA0092BAEA0091BA,
-    0xEA0094BAEA0093BA, 0xEA0096BAEA0095BA, 0xEA0098BAEA0097BA,
-    0xEA009ABAEA0099BA, 0xEA009CBAEA009BBA, 0xEA009EBAEA009DBA,
-    0xEA00A0BAEA009FBA, 0xEA00A2BAEA00A1BA, 0xEA00A4BAEA00A3BA,
-    0xEA00A6BAEA00A5BA, 0xEA00A8BAEA00A7BA, 0xEA00AABAEA00A9BA,
-    0xEA00ACBAEA00ABBA, 0xEA00AEBAEA00ADBA, 0xEA00B0BAEA00AFBA,
-    0xEA00B2BAEA00B1BA, 0xEA00B4BAEA00B3BA, 0xEA00B6BAEA00B5BA,
-    0xEA00B8BAEA00B7BA, 0xEA00BABAEA00B9BA, 0xEA00BCBAEA00BBBA,
-    0xEA00BEBAEA00BDBA, 0xEA0080BBEA00BFBA, 0xEA0082BBEA0081BB,
-    0xEA0084BBEA0083BB, 0xEA0086BBEA0085BB, 0xEA0088BBEA0087BB,
-    0xEA008ABBEA0089BB, 0xEA008CBBEA008BBB, 0xEA008EBBEA008DBB,
-    0xEA0090BBEA008FBB, 0xEA0092BBEA0091BB, 0xEA0094BBEA0093BB,
-    0xEA0096BBEA0095BB, 0xEA0098BBEA0097BB, 0xEA009ABBEA0099BB,
-    0xEA009CBBEA009BBB, 0xEA009EBBEA009DBB, 0xEA00A0BBEA009FBB,
-    0xEA00A2BBEA00A1BB, 0xEA00A4BBEA00A3BB, 0xEA00A6BBEA00A5BB,
-    0xEA00A8BBEA00A7BB, 0xEA00AABBEA00A9BB, 0xEA00ACBBEA00ABBB,
-    0xEA00AEBBEA00ADBB, 0xEA00B0BBEA00AFBB, 0xEA00B2BBEA00B1BB,
-    0xEA00B4BBEA00B3BB, 0xEA00B6BBEA00B5BB, 0xEA00B8BBEA00B7BB,
-    0xEA00BABBEA00B9BB, 0xEA00BCBBEA00BBBB, 0xEA00BEBBEA00BDBB,
-    0xEA0080BCEA00BFBB, 0xEA0082BCEA0081BC, 0xEA0084BCEA0083BC,
-    0xEA0086BCEA0085BC, 0xEA0088BCEA0087BC, 0xEA008ABCEA0089BC,
-    0xEA008CBCEA008BBC, 0xEA008EBCEA008DBC, 0xEA0090BCEA008FBC,
-    0xEA0092BCEA0091BC, 0xEA0094BCEA0093BC, 0xEA0096BCEA0095BC,
-    0xEA0098BCEA0097BC, 0xEA009ABCEA0099BC, 0xEA009CBCEA009BBC,
-    0xEA009EBCEA009DBC, 0xEA00A0BCEA009FBC, 0xEA00A2BCEA00A1BC,
-    0xEA00A4BCEA00A3BC, 0xEA00A6BCEA00A5BC, 0xEA00A8BCEA00A7BC,
-    0xEA00AABCEA00A9BC, 0xEA00ACBCEA00ABBC, 0xEA00AEBCEA00ADBC,
-    0xEA00B0BCEA00AFBC, 0xEA00B2BCEA00B1BC, 0xEA00B4BCEA00B3BC,
-    0xEA00B6BCEA00B5BC, 0xEA00B8BCEA00B7BC, 0xEA00BABCEA00B9BC,
-    0xEA00BCBCEA00BBBC, 0xEA00BEBCEA00BDBC, 0xEA0080BDEA00BFBC,
-    0xEA0082BDEA0081BD, 0xEA0084BDEA0083BD, 0xEA0086BDEA0085BD,
-    0xEA0088BDEA0087BD, 0xEA008ABDEA0089BD, 0xEA008CBDEA008BBD,
-    0xEA008EBDEA008DBD, 0xEA0090BDEA008FBD, 0xEA0092BDEA0091BD,
-    0xEA0094BDEA0093BD, 0xEA0096BDEA0095BD, 0xEA0098BDEA0097BD,
-    0xEA009ABDEA0099BD, 0xEA009CBDEA009BBD, 0xEA009EBDEA009DBD,
-    0xEA00A0BDEA009FBD, 0xEA00A2BDEA00A1BD, 0xEA00A4BDEA00A3BD,
-    0xEA00A6BDEA00A5BD, 0xEA00A8BDEA00A7BD, 0xEA00AABDEA00A9BD,
-    0xEA00ACBDEA00ABBD, 0xEA00AEBDEA00ADBD, 0xEA00B0BDEA00AFBD,
-    0xEA00B2BDEA00B1BD, 0xEA00B4BDEA00B3BD, 0xEA00B6BDEA00B5BD,
-    0xEA00B8BDEA00B7BD, 0xEA00BABDEA00B9BD, 0xEA00BCBDEA00BBBD,
-    0xEA00BEBDEA00BDBD, 0xEA0080BEEA00BFBD, 0xEA0082BEEA0081BE,
-    0xEA0084BEEA0083BE, 0xEA0086BEEA0085BE, 0xEA0088BEEA0087BE,
-    0xEA008ABEEA0089BE, 0xEA008CBEEA008BBE, 0xEA008EBEEA008DBE,
-    0xEA0090BEEA008FBE, 0xEA0092BEEA0091BE, 0xEA0094BEEA0093BE,
-    0xEA0096BEEA0095BE, 0xEA0098BEEA0097BE, 0xEA009ABEEA0099BE,
-    0xEA009CBEEA009BBE, 0xEA009EBEEA009DBE, 0xEA00A0BEEA009FBE,
-    0xEA00A2BEEA00A1BE, 0xEA00A4BEEA00A3BE, 0xEA00A6BEEA00A5BE,
-    0xEA00A8BEEA00A7BE, 0xEA00AABEEA00A9BE, 0xEA00ACBEEA00ABBE,
-    0xEA00AEBEEA00ADBE, 0xEA00B0BEEA00AFBE, 0xEA00B2BEEA00B1BE,
-    0xEA00B4BEEA00B3BE, 0xEA00B6BEEA00B5BE, 0xEA00B8BEEA00B7BE,
-    0xEA00BABEEA00B9BE, 0xEA00BCBEEA00BBBE, 0xEA00BEBEEA00BDBE,
-    0xEA0080BFEA00BFBE, 0xEA0082BFEA0081BF, 0xEA0084BFEA0083BF,
-    0xEA0086BFEA0085BF, 0xEA0088BFEA0087BF, 0xEA008ABFEA0089BF,
-    0xEA008CBFEA008BBF, 0xEA008EBFEA008DBF, 0xEA0090BFEA008FBF,
-    0xEA0092BFEA0091BF, 0xEA0094BFEA0093BF, 0xEA0096BFEA0095BF,
-    0xEA0098BFEA0097BF, 0xEA009ABFEA0099BF, 0xEA009CBFEA009BBF,
-    0xEA009EBFEA009DBF, 0xEA00A0BFEA009FBF, 0xEA00A2BFEA00A1BF,
-    0xEA00A4BFEA00A3BF, 0xEA00A6BFEA00A5BF, 0xEA00A8BFEA00A7BF,
-    0xEA00AABFEA00A9BF, 0xEA00ACBFEA00ABBF, 0xEA00AEBFEA00ADBF,
-    0xEA00B0BFEA00AFBF, 0xEA00B2BFEA00B1BF, 0xEA00B4BFEA00B3BF,
-    0xEA00B6BFEA00B5BF, 0xEA00B8BFEA00B7BF, 0xEA00BABFEA00B9BF,
-    0xEA00BCBFEA00BBBF, 0xEA00BEBFEA00BDBF, 0xEB008080EB00BFBF,
-    0xEB008280EB008180, 0xEB008480EB008380, 0xEB008680EB008580,
-    0xEB008880EB008780, 0xEB008A80EB008980, 0xEB008C80EB008B80,
-    0xEB008E80EB008D80, 0xEB009080EB008F80, 0xEB009280EB009180,
-    0xEB009480EB009380, 0xEB009680EB009580, 0xEB009880EB009780,
-    0xEB009A80EB009980, 0xEB009C80EB009B80, 0xEB009E80EB009D80,
-    0xEB00A080EB009F80, 0xEB00A280EB00A180, 0xEB00A480EB00A380,
-    0xEB00A680EB00A580, 0xEB00A880EB00A780, 0xEB00AA80EB00A980,
-    0xEB00AC80EB00AB80, 0xEB00AE80EB00AD80, 0xEB00B080EB00AF80,
-    0xEB00B280EB00B180, 0xEB00B480EB00B380, 0xEB00B680EB00B580,
-    0xEB00B880EB00B780, 0xEB00BA80EB00B980, 0xEB00BC80EB00BB80,
-    0xEB00BE80EB00BD80, 0xEB008081EB00BF80, 0xEB008281EB008181,
-    0xEB008481EB008381, 0xEB008681EB008581, 0xEB008881EB008781,
-    0xEB008A81EB008981, 0xEB008C81EB008B81, 0xEB008E81EB008D81,
-    0xEB009081EB008F81, 0xEB009281EB009181, 0xEB009481EB009381,
-    0xEB009681EB009581, 0xEB009881EB009781, 0xEB009A81EB009981,
-    0xEB009C81EB009B81, 0xEB009E81EB009D81, 0xEB00A081EB009F81,
-    0xEB00A281EB00A181, 0xEB00A481EB00A381, 0xEB00A681EB00A581,
-    0xEB00A881EB00A781, 0xEB00AA81EB00A981, 0xEB00AC81EB00AB81,
-    0xEB00AE81EB00AD81, 0xEB00B081EB00AF81, 0xEB00B281EB00B181,
-    0xEB00B481EB00B381, 0xEB00B681EB00B581, 0xEB00B881EB00B781,
-    0xEB00BA81EB00B981, 0xEB00BC81EB00BB81, 0xEB00BE81EB00BD81,
-    0xEB008082EB00BF81, 0xEB008282EB008182, 0xEB008482EB008382,
-    0xEB008682EB008582, 0xEB008882EB008782, 0xEB008A82EB008982,
-    0xEB008C82EB008B82, 0xEB008E82EB008D82, 0xEB009082EB008F82,
-    0xEB009282EB009182, 0xEB009482EB009382, 0xEB009682EB009582,
-    0xEB009882EB009782, 0xEB009A82EB009982, 0xEB009C82EB009B82,
-    0xEB009E82EB009D82, 0xEB00A082EB009F82, 0xEB00A282EB00A182,
-    0xEB00A482EB00A382, 0xEB00A682EB00A582, 0xEB00A882EB00A782,
-    0xEB00AA82EB00A982, 0xEB00AC82EB00AB82, 0xEB00AE82EB00AD82,
-    0xEB00B082EB00AF82, 0xEB00B282EB00B182, 0xEB00B482EB00B382,
-    0xEB00B682EB00B582, 0xEB00B882EB00B782, 0xEB00BA82EB00B982,
-    0xEB00BC82EB00BB82, 0xEB00BE82EB00BD82, 0xEB008083EB00BF82,
-    0xEB008283EB008183, 0xEB008483EB008383, 0xEB008683EB008583,
-    0xEB008883EB008783, 0xEB008A83EB008983, 0xEB008C83EB008B83,
-    0xEB008E83EB008D83, 0xEB009083EB008F83, 0xEB009283EB009183,
-    0xEB009483EB009383, 0xEB009683EB009583, 0xEB009883EB009783,
-    0xEB009A83EB009983, 0xEB009C83EB009B83, 0xEB009E83EB009D83,
-    0xEB00A083EB009F83, 0xEB00A283EB00A183, 0xEB00A483EB00A383,
-    0xEB00A683EB00A583, 0xEB00A883EB00A783, 0xEB00AA83EB00A983,
-    0xEB00AC83EB00AB83, 0xEB00AE83EB00AD83, 0xEB00B083EB00AF83,
-    0xEB00B283EB00B183, 0xEB00B483EB00B383, 0xEB00B683EB00B583,
-    0xEB00B883EB00B783, 0xEB00BA83EB00B983, 0xEB00BC83EB00BB83,
-    0xEB00BE83EB00BD83, 0xEB008084EB00BF83, 0xEB008284EB008184,
-    0xEB008484EB008384, 0xEB008684EB008584, 0xEB008884EB008784,
-    0xEB008A84EB008984, 0xEB008C84EB008B84, 0xEB008E84EB008D84,
-    0xEB009084EB008F84, 0xEB009284EB009184, 0xEB009484EB009384,
-    0xEB009684EB009584, 0xEB009884EB009784, 0xEB009A84EB009984,
-    0xEB009C84EB009B84, 0xEB009E84EB009D84, 0xEB00A084EB009F84,
-    0xEB00A284EB00A184, 0xEB00A484EB00A384, 0xEB00A684EB00A584,
-    0xEB00A884EB00A784, 0xEB00AA84EB00A984, 0xEB00AC84EB00AB84,
-    0xEB00AE84EB00AD84, 0xEB00B084EB00AF84, 0xEB00B284EB00B184,
-    0xEB00B484EB00B384, 0xEB00B684EB00B584, 0xEB00B884EB00B784,
-    0xEB00BA84EB00B984, 0xEB00BC84EB00BB84, 0xEB00BE84EB00BD84,
-    0xEB008085EB00BF84, 0xEB008285EB008185, 0xEB008485EB008385,
-    0xEB008685EB008585, 0xEB008885EB008785, 0xEB008A85EB008985,
-    0xEB008C85EB008B85, 0xEB008E85EB008D85, 0xEB009085EB008F85,
-    0xEB009285EB009185, 0xEB009485EB009385, 0xEB009685EB009585,
-    0xEB009885EB009785, 0xEB009A85EB009985, 0xEB009C85EB009B85,
-    0xEB009E85EB009D85, 0xEB00A085EB009F85, 0xEB00A285EB00A185,
-    0xEB00A485EB00A385, 0xEB00A685EB00A585, 0xEB00A885EB00A785,
-    0xEB00AA85EB00A985, 0xEB00AC85EB00AB85, 0xEB00AE85EB00AD85,
-    0xEB00B085EB00AF85, 0xEB00B285EB00B185, 0xEB00B485EB00B385,
-    0xEB00B685EB00B585, 0xEB00B885EB00B785, 0xEB00BA85EB00B985,
-    0xEB00BC85EB00BB85, 0xEB00BE85EB00BD85, 0xEB008086EB00BF85,
-    0xEB008286EB008186, 0xEB008486EB008386, 0xEB008686EB008586,
-    0xEB008886EB008786, 0xEB008A86EB008986, 0xEB008C86EB008B86,
-    0xEB008E86EB008D86, 0xEB009086EB008F86, 0xEB009286EB009186,
-    0xEB009486EB009386, 0xEB009686EB009586, 0xEB009886EB009786,
-    0xEB009A86EB009986, 0xEB009C86EB009B86, 0xEB009E86EB009D86,
-    0xEB00A086EB009F86, 0xEB00A286EB00A186, 0xEB00A486EB00A386,
-    0xEB00A686EB00A586, 0xEB00A886EB00A786, 0xEB00AA86EB00A986,
-    0xEB00AC86EB00AB86, 0xEB00AE86EB00AD86, 0xEB00B086EB00AF86,
-    0xEB00B286EB00B186, 0xEB00B486EB00B386, 0xEB00B686EB00B586,
-    0xEB00B886EB00B786, 0xEB00BA86EB00B986, 0xEB00BC86EB00BB86,
-    0xEB00BE86EB00BD86, 0xEB008087EB00BF86, 0xEB008287EB008187,
-    0xEB008487EB008387, 0xEB008687EB008587, 0xEB008887EB008787,
-    0xEB008A87EB008987, 0xEB008C87EB008B87, 0xEB008E87EB008D87,
-    0xEB009087EB008F87, 0xEB009287EB009187, 0xEB009487EB009387,
-    0xEB009687EB009587, 0xEB009887EB009787, 0xEB009A87EB009987,
-    0xEB009C87EB009B87, 0xEB009E87EB009D87, 0xEB00A087EB009F87,
-    0xEB00A287EB00A187, 0xEB00A487EB00A387, 0xEB00A687EB00A587,
-    0xEB00A887EB00A787, 0xEB00AA87EB00A987, 0xEB00AC87EB00AB87,
-    0xEB00AE87EB00AD87, 0xEB00B087EB00AF87, 0xEB00B287EB00B187,
-    0xEB00B487EB00B387, 0xEB00B687EB00B587, 0xEB00B887EB00B787,
-    0xEB00BA87EB00B987, 0xEB00BC87EB00BB87, 0xEB00BE87EB00BD87,
-    0xEB008088EB00BF87, 0xEB008288EB008188, 0xEB008488EB008388,
-    0xEB008688EB008588, 0xEB008888EB008788, 0xEB008A88EB008988,
-    0xEB008C88EB008B88, 0xEB008E88EB008D88, 0xEB009088EB008F88,
-    0xEB009288EB009188, 0xEB009488EB009388, 0xEB009688EB009588,
-    0xEB009888EB009788, 0xEB009A88EB009988, 0xEB009C88EB009B88,
-    0xEB009E88EB009D88, 0xEB00A088EB009F88, 0xEB00A288EB00A188,
-    0xEB00A488EB00A388, 0xEB00A688EB00A588, 0xEB00A888EB00A788,
-    0xEB00AA88EB00A988, 0xEB00AC88EB00AB88, 0xEB00AE88EB00AD88,
-    0xEB00B088EB00AF88, 0xEB00B288EB00B188, 0xEB00B488EB00B388,
-    0xEB00B688EB00B588, 0xEB00B888EB00B788, 0xEB00BA88EB00B988,
-    0xEB00BC88EB00BB88, 0xEB00BE88EB00BD88, 0xEB008089EB00BF88,
-    0xEB008289EB008189, 0xEB008489EB008389, 0xEB008689EB008589,
-    0xEB008889EB008789, 0xEB008A89EB008989, 0xEB008C89EB008B89,
-    0xEB008E89EB008D89, 0xEB009089EB008F89, 0xEB009289EB009189,
-    0xEB009489EB009389, 0xEB009689EB009589, 0xEB009889EB009789,
-    0xEB009A89EB009989, 0xEB009C89EB009B89, 0xEB009E89EB009D89,
-    0xEB00A089EB009F89, 0xEB00A289EB00A189, 0xEB00A489EB00A389,
-    0xEB00A689EB00A589, 0xEB00A889EB00A789, 0xEB00AA89EB00A989,
-    0xEB00AC89EB00AB89, 0xEB00AE89EB00AD89, 0xEB00B089EB00AF89,
-    0xEB00B289EB00B189, 0xEB00B489EB00B389, 0xEB00B689EB00B589,
-    0xEB00B889EB00B789, 0xEB00BA89EB00B989, 0xEB00BC89EB00BB89,
-    0xEB00BE89EB00BD89, 0xEB00808AEB00BF89, 0xEB00828AEB00818A,
-    0xEB00848AEB00838A, 0xEB00868AEB00858A, 0xEB00888AEB00878A,
-    0xEB008A8AEB00898A, 0xEB008C8AEB008B8A, 0xEB008E8AEB008D8A,
-    0xEB00908AEB008F8A, 0xEB00928AEB00918A, 0xEB00948AEB00938A,
-    0xEB00968AEB00958A, 0xEB00988AEB00978A, 0xEB009A8AEB00998A,
-    0xEB009C8AEB009B8A, 0xEB009E8AEB009D8A, 0xEB00A08AEB009F8A,
-    0xEB00A28AEB00A18A, 0xEB00A48AEB00A38A, 0xEB00A68AEB00A58A,
-    0xEB00A88AEB00A78A, 0xEB00AA8AEB00A98A, 0xEB00AC8AEB00AB8A,
-    0xEB00AE8AEB00AD8A, 0xEB00B08AEB00AF8A, 0xEB00B28AEB00B18A,
-    0xEB00B48AEB00B38A, 0xEB00B68AEB00B58A, 0xEB00B88AEB00B78A,
-    0xEB00BA8AEB00B98A, 0xEB00BC8AEB00BB8A, 0xEB00BE8AEB00BD8A,
-    0xEB00808BEB00BF8A, 0xEB00828BEB00818B, 0xEB00848BEB00838B,
-    0xEB00868BEB00858B, 0xEB00888BEB00878B, 0xEB008A8BEB00898B,
-    0xEB008C8BEB008B8B, 0xEB008E8BEB008D8B, 0xEB00908BEB008F8B,
-    0xEB00928BEB00918B, 0xEB00948BEB00938B, 0xEB00968BEB00958B,
-    0xEB00988BEB00978B, 0xEB009A8BEB00998B, 0xEB009C8BEB009B8B,
-    0xEB009E8BEB009D8B, 0xEB00A08BEB009F8B, 0xEB00A28BEB00A18B,
-    0xEB00A48BEB00A38B, 0xEB00A68BEB00A58B, 0xEB00A88BEB00A78B,
-    0xEB00AA8BEB00A98B, 0xEB00AC8BEB00AB8B, 0xEB00AE8BEB00AD8B,
-    0xEB00B08BEB00AF8B, 0xEB00B28BEB00B18B, 0xEB00B48BEB00B38B,
-    0xEB00B68BEB00B58B, 0xEB00B88BEB00B78B, 0xEB00BA8BEB00B98B,
-    0xEB00BC8BEB00BB8B, 0xEB00BE8BEB00BD8B, 0xEB00808CEB00BF8B,
-    0xEB00828CEB00818C, 0xEB00848CEB00838C, 0xEB00868CEB00858C,
-    0xEB00888CEB00878C, 0xEB008A8CEB00898C, 0xEB008C8CEB008B8C,
-    0xEB008E8CEB008D8C, 0xEB00908CEB008F8C, 0xEB00928CEB00918C,
-    0xEB00948CEB00938C, 0xEB00968CEB00958C, 0xEB00988CEB00978C,
-    0xEB009A8CEB00998C, 0xEB009C8CEB009B8C, 0xEB009E8CEB009D8C,
-    0xEB00A08CEB009F8C, 0xEB00A28CEB00A18C, 0xEB00A48CEB00A38C,
-    0xEB00A68CEB00A58C, 0xEB00A88CEB00A78C, 0xEB00AA8CEB00A98C,
-    0xEB00AC8CEB00AB8C, 0xEB00AE8CEB00AD8C, 0xEB00B08CEB00AF8C,
-    0xEB00B28CEB00B18C, 0xEB00B48CEB00B38C, 0xEB00B68CEB00B58C,
-    0xEB00B88CEB00B78C, 0xEB00BA8CEB00B98C, 0xEB00BC8CEB00BB8C,
-    0xEB00BE8CEB00BD8C, 0xEB00808DEB00BF8C, 0xEB00828DEB00818D,
-    0xEB00848DEB00838D, 0xEB00868DEB00858D, 0xEB00888DEB00878D,
-    0xEB008A8DEB00898D, 0xEB008C8DEB008B8D, 0xEB008E8DEB008D8D,
-    0xEB00908DEB008F8D, 0xEB00928DEB00918D, 0xEB00948DEB00938D,
-    0xEB00968DEB00958D, 0xEB00988DEB00978D, 0xEB009A8DEB00998D,
-    0xEB009C8DEB009B8D, 0xEB009E8DEB009D8D, 0xEB00A08DEB009F8D,
-    0xEB00A28DEB00A18D, 0xEB00A48DEB00A38D, 0xEB00A68DEB00A58D,
-    0xEB00A88DEB00A78D, 0xEB00AA8DEB00A98D, 0xEB00AC8DEB00AB8D,
-    0xEB00AE8DEB00AD8D, 0xEB00B08DEB00AF8D, 0xEB00B28DEB00B18D,
-    0xEB00B48DEB00B38D, 0xEB00B68DEB00B58D, 0xEB00B88DEB00B78D,
-    0xEB00BA8DEB00B98D, 0xEB00BC8DEB00BB8D, 0xEB00BE8DEB00BD8D,
-    0xEB00808EEB00BF8D, 0xEB00828EEB00818E, 0xEB00848EEB00838E,
-    0xEB00868EEB00858E, 0xEB00888EEB00878E, 0xEB008A8EEB00898E,
-    0xEB008C8EEB008B8E, 0xEB008E8EEB008D8E, 0xEB00908EEB008F8E,
-    0xEB00928EEB00918E, 0xEB00948EEB00938E, 0xEB00968EEB00958E,
-    0xEB00988EEB00978E, 0xEB009A8EEB00998E, 0xEB009C8EEB009B8E,
-    0xEB009E8EEB009D8E, 0xEB00A08EEB009F8E, 0xEB00A28EEB00A18E,
-    0xEB00A48EEB00A38E, 0xEB00A68EEB00A58E, 0xEB00A88EEB00A78E,
-    0xEB00AA8EEB00A98E, 0xEB00AC8EEB00AB8E, 0xEB00AE8EEB00AD8E,
-    0xEB00B08EEB00AF8E, 0xEB00B28EEB00B18E, 0xEB00B48EEB00B38E,
-    0xEB00B68EEB00B58E, 0xEB00B88EEB00B78E, 0xEB00BA8EEB00B98E,
-    0xEB00BC8EEB00BB8E, 0xEB00BE8EEB00BD8E, 0xEB00808FEB00BF8E,
-    0xEB00828FEB00818F, 0xEB00848FEB00838F, 0xEB00868FEB00858F,
-    0xEB00888FEB00878F, 0xEB008A8FEB00898F, 0xEB008C8FEB008B8F,
-    0xEB008E8FEB008D8F, 0xEB00908FEB008F8F, 0xEB00928FEB00918F,
-    0xEB00948FEB00938F, 0xEB00968FEB00958F, 0xEB00988FEB00978F,
-    0xEB009A8FEB00998F, 0xEB009C8FEB009B8F, 0xEB009E8FEB009D8F,
-    0xEB00A08FEB009F8F, 0xEB00A28FEB00A18F, 0xEB00A48FEB00A38F,
-    0xEB00A68FEB00A58F, 0xEB00A88FEB00A78F, 0xEB00AA8FEB00A98F,
-    0xEB00AC8FEB00AB8F, 0xEB00AE8FEB00AD8F, 0xEB00B08FEB00AF8F,
-    0xEB00B28FEB00B18F, 0xEB00B48FEB00B38F, 0xEB00B68FEB00B58F,
-    0xEB00B88FEB00B78F, 0xEB00BA8FEB00B98F, 0xEB00BC8FEB00BB8F,
-    0xEB00BE8FEB00BD8F, 0xEB008090EB00BF8F, 0xEB008290EB008190,
-    0xEB008490EB008390, 0xEB008690EB008590, 0xEB008890EB008790,
-    0xEB008A90EB008990, 0xEB008C90EB008B90, 0xEB008E90EB008D90,
-    0xEB009090EB008F90, 0xEB009290EB009190, 0xEB009490EB009390,
-    0xEB009690EB009590, 0xEB009890EB009790, 0xEB009A90EB009990,
-    0xEB009C90EB009B90, 0xEB009E90EB009D90, 0xEB00A090EB009F90,
-    0xEB00A290EB00A190, 0xEB00A490EB00A390, 0xEB00A690EB00A590,
-    0xEB00A890EB00A790, 0xEB00AA90EB00A990, 0xEB00AC90EB00AB90,
-    0xEB00AE90EB00AD90, 0xEB00B090EB00AF90, 0xEB00B290EB00B190,
-    0xEB00B490EB00B390, 0xEB00B690EB00B590, 0xEB00B890EB00B790,
-    0xEB00BA90EB00B990, 0xEB00BC90EB00BB90, 0xEB00BE90EB00BD90,
-    0xEB008091EB00BF90, 0xEB008291EB008191, 0xEB008491EB008391,
-    0xEB008691EB008591, 0xEB008891EB008791, 0xEB008A91EB008991,
-    0xEB008C91EB008B91, 0xEB008E91EB008D91, 0xEB009091EB008F91,
-    0xEB009291EB009191, 0xEB009491EB009391, 0xEB009691EB009591,
-    0xEB009891EB009791, 0xEB009A91EB009991, 0xEB009C91EB009B91,
-    0xEB009E91EB009D91, 0xEB00A091EB009F91, 0xEB00A291EB00A191,
-    0xEB00A491EB00A391, 0xEB00A691EB00A591, 0xEB00A891EB00A791,
-    0xEB00AA91EB00A991, 0xEB00AC91EB00AB91, 0xEB00AE91EB00AD91,
-    0xEB00B091EB00AF91, 0xEB00B291EB00B191, 0xEB00B491EB00B391,
-    0xEB00B691EB00B591, 0xEB00B891EB00B791, 0xEB00BA91EB00B991,
-    0xEB00BC91EB00BB91, 0xEB00BE91EB00BD91, 0xEB008092EB00BF91,
-    0xEB008292EB008192, 0xEB008492EB008392, 0xEB008692EB008592,
-    0xEB008892EB008792, 0xEB008A92EB008992, 0xEB008C92EB008B92,
-    0xEB008E92EB008D92, 0xEB009092EB008F92, 0xEB009292EB009192,
-    0xEB009492EB009392, 0xEB009692EB009592, 0xEB009892EB009792,
-    0xEB009A92EB009992, 0xEB009C92EB009B92, 0xEB009E92EB009D92,
-    0xEB00A092EB009F92, 0xEB00A292EB00A192, 0xEB00A492EB00A392,
-    0xEB00A692EB00A592, 0xEB00A892EB00A792, 0xEB00AA92EB00A992,
-    0xEB00AC92EB00AB92, 0xEB00AE92EB00AD92, 0xEB00B092EB00AF92,
-    0xEB00B292EB00B192, 0xEB00B492EB00B392, 0xEB00B692EB00B592,
-    0xEB00B892EB00B792, 0xEB00BA92EB00B992, 0xEB00BC92EB00BB92,
-    0xEB00BE92EB00BD92, 0xEB008093EB00BF92, 0xEB008293EB008193,
-    0xEB008493EB008393, 0xEB008693EB008593, 0xEB008893EB008793,
-    0xEB008A93EB008993, 0xEB008C93EB008B93, 0xEB008E93EB008D93,
-    0xEB009093EB008F93, 0xEB009293EB009193, 0xEB009493EB009393,
-    0xEB009693EB009593, 0xEB009893EB009793, 0xEB009A93EB009993,
-    0xEB009C93EB009B93, 0xEB009E93EB009D93, 0xEB00A093EB009F93,
-    0xEB00A293EB00A193, 0xEB00A493EB00A393, 0xEB00A693EB00A593,
-    0xEB00A893EB00A793, 0xEB00AA93EB00A993, 0xEB00AC93EB00AB93,
-    0xEB00AE93EB00AD93, 0xEB00B093EB00AF93, 0xEB00B293EB00B193,
-    0xEB00B493EB00B393, 0xEB00B693EB00B593, 0xEB00B893EB00B793,
-    0xEB00BA93EB00B993, 0xEB00BC93EB00BB93, 0xEB00BE93EB00BD93,
-    0xEB008094EB00BF93, 0xEB008294EB008194, 0xEB008494EB008394,
-    0xEB008694EB008594, 0xEB008894EB008794, 0xEB008A94EB008994,
-    0xEB008C94EB008B94, 0xEB008E94EB008D94, 0xEB009094EB008F94,
-    0xEB009294EB009194, 0xEB009494EB009394, 0xEB009694EB009594,
-    0xEB009894EB009794, 0xEB009A94EB009994, 0xEB009C94EB009B94,
-    0xEB009E94EB009D94, 0xEB00A094EB009F94, 0xEB00A294EB00A194,
-    0xEB00A494EB00A394, 0xEB00A694EB00A594, 0xEB00A894EB00A794,
-    0xEB00AA94EB00A994, 0xEB00AC94EB00AB94, 0xEB00AE94EB00AD94,
-    0xEB00B094EB00AF94, 0xEB00B294EB00B194, 0xEB00B494EB00B394,
-    0xEB00B694EB00B594, 0xEB00B894EB00B794, 0xEB00BA94EB00B994,
-    0xEB00BC94EB00BB94, 0xEB00BE94EB00BD94, 0xEB008095EB00BF94,
-    0xEB008295EB008195, 0xEB008495EB008395, 0xEB008695EB008595,
-    0xEB008895EB008795, 0xEB008A95EB008995, 0xEB008C95EB008B95,
-    0xEB008E95EB008D95, 0xEB009095EB008F95, 0xEB009295EB009195,
-    0xEB009495EB009395, 0xEB009695EB009595, 0xEB009895EB009795,
-    0xEB009A95EB009995, 0xEB009C95EB009B95, 0xEB009E95EB009D95,
-    0xEB00A095EB009F95, 0xEB00A295EB00A195, 0xEB00A495EB00A395,
-    0xEB00A695EB00A595, 0xEB00A895EB00A795, 0xEB00AA95EB00A995,
-    0xEB00AC95EB00AB95, 0xEB00AE95EB00AD95, 0xEB00B095EB00AF95,
-    0xEB00B295EB00B195, 0xEB00B495EB00B395, 0xEB00B695EB00B595,
-    0xEB00B895EB00B795, 0xEB00BA95EB00B995, 0xEB00BC95EB00BB95,
-    0xEB00BE95EB00BD95, 0xEB008096EB00BF95, 0xEB008296EB008196,
-    0xEB008496EB008396, 0xEB008696EB008596, 0xEB008896EB008796,
-    0xEB008A96EB008996, 0xEB008C96EB008B96, 0xEB008E96EB008D96,
-    0xEB009096EB008F96, 0xEB009296EB009196, 0xEB009496EB009396,
-    0xEB009696EB009596, 0xEB009896EB009796, 0xEB009A96EB009996,
-    0xEB009C96EB009B96, 0xEB009E96EB009D96, 0xEB00A096EB009F96,
-    0xEB00A296EB00A196, 0xEB00A496EB00A396, 0xEB00A696EB00A596,
-    0xEB00A896EB00A796, 0xEB00AA96EB00A996, 0xEB00AC96EB00AB96,
-    0xEB00AE96EB00AD96, 0xEB00B096EB00AF96, 0xEB00B296EB00B196,
-    0xEB00B496EB00B396, 0xEB00B696EB00B596, 0xEB00B896EB00B796,
-    0xEB00BA96EB00B996, 0xEB00BC96EB00BB96, 0xEB00BE96EB00BD96,
-    0xEB008097EB00BF96, 0xEB008297EB008197, 0xEB008497EB008397,
-    0xEB008697EB008597, 0xEB008897EB008797, 0xEB008A97EB008997,
-    0xEB008C97EB008B97, 0xEB008E97EB008D97, 0xEB009097EB008F97,
-    0xEB009297EB009197, 0xEB009497EB009397, 0xEB009697EB009597,
-    0xEB009897EB009797, 0xEB009A97EB009997, 0xEB009C97EB009B97,
-    0xEB009E97EB009D97, 0xEB00A097EB009F97, 0xEB00A297EB00A197,
-    0xEB00A497EB00A397, 0xEB00A697EB00A597, 0xEB00A897EB00A797,
-    0xEB00AA97EB00A997, 0xEB00AC97EB00AB97, 0xEB00AE97EB00AD97,
-    0xEB00B097EB00AF97, 0xEB00B297EB00B197, 0xEB00B497EB00B397,
-    0xEB00B697EB00B597, 0xEB00B897EB00B797, 0xEB00BA97EB00B997,
-    0xEB00BC97EB00BB97, 0xEB00BE97EB00BD97, 0xEB008098EB00BF97,
-    0xEB008298EB008198, 0xEB008498EB008398, 0xEB008698EB008598,
-    0xEB008898EB008798, 0xEB008A98EB008998, 0xEB008C98EB008B98,
-    0xEB008E98EB008D98, 0xEB009098EB008F98, 0xEB009298EB009198,
-    0xEB009498EB009398, 0xEB009698EB009598, 0xEB009898EB009798,
-    0xEB009A98EB009998, 0xEB009C98EB009B98, 0xEB009E98EB009D98,
-    0xEB00A098EB009F98, 0xEB00A298EB00A198, 0xEB00A498EB00A398,
-    0xEB00A698EB00A598, 0xEB00A898EB00A798, 0xEB00AA98EB00A998,
-    0xEB00AC98EB00AB98, 0xEB00AE98EB00AD98, 0xEB00B098EB00AF98,
-    0xEB00B298EB00B198, 0xEB00B498EB00B398, 0xEB00B698EB00B598,
-    0xEB00B898EB00B798, 0xEB00BA98EB00B998, 0xEB00BC98EB00BB98,
-    0xEB00BE98EB00BD98, 0xEB008099EB00BF98, 0xEB008299EB008199,
-    0xEB008499EB008399, 0xEB008699EB008599, 0xEB008899EB008799,
-    0xEB008A99EB008999, 0xEB008C99EB008B99, 0xEB008E99EB008D99,
-    0xEB009099EB008F99, 0xEB009299EB009199, 0xEB009499EB009399,
-    0xEB009699EB009599, 0xEB009899EB009799, 0xEB009A99EB009999,
-    0xEB009C99EB009B99, 0xEB009E99EB009D99, 0xEB00A099EB009F99,
-    0xEB00A299EB00A199, 0xEB00A499EB00A399, 0xEB00A699EB00A599,
-    0xEB00A899EB00A799, 0xEB00AA99EB00A999, 0xEB00AC99EB00AB99,
-    0xEB00AE99EB00AD99, 0xEB00B099EB00AF99, 0xEB00B299EB00B199,
-    0xEB00B499EB00B399, 0xEB00B699EB00B599, 0xEB00B899EB00B799,
-    0xEB00BA99EB00B999, 0xEB00BC99EB00BB99, 0xEB00BE99EB00BD99,
-    0xEB00809AEB00BF99, 0xEB00829AEB00819A, 0xEB00849AEB00839A,
-    0xEB00869AEB00859A, 0xEB00889AEB00879A, 0xEB008A9AEB00899A,
-    0xEB008C9AEB008B9A, 0xEB008E9AEB008D9A, 0xEB00909AEB008F9A,
-    0xEB00929AEB00919A, 0xEB00949AEB00939A, 0xEB00969AEB00959A,
-    0xEB00989AEB00979A, 0xEB009A9AEB00999A, 0xEB009C9AEB009B9A,
-    0xEB009E9AEB009D9A, 0xEB00A09AEB009F9A, 0xEB00A29AEB00A19A,
-    0xEB00A49AEB00A39A, 0xEB00A69AEB00A59A, 0xEB00A89AEB00A79A,
-    0xEB00AA9AEB00A99A, 0xEB00AC9AEB00AB9A, 0xEB00AE9AEB00AD9A,
-    0xEB00B09AEB00AF9A, 0xEB00B29AEB00B19A, 0xEB00B49AEB00B39A,
-    0xEB00B69AEB00B59A, 0xEB00B89AEB00B79A, 0xEB00BA9AEB00B99A,
-    0xEB00BC9AEB00BB9A, 0xEB00BE9AEB00BD9A, 0xEB00809BEB00BF9A,
-    0xEB00829BEB00819B, 0xEB00849BEB00839B, 0xEB00869BEB00859B,
-    0xEB00889BEB00879B, 0xEB008A9BEB00899B, 0xEB008C9BEB008B9B,
-    0xEB008E9BEB008D9B, 0xEB00909BEB008F9B, 0xEB00929BEB00919B,
-    0xEB00949BEB00939B, 0xEB00969BEB00959B, 0xEB00989BEB00979B,
-    0xEB009A9BEB00999B, 0xEB009C9BEB009B9B, 0xEB009E9BEB009D9B,
-    0xEB00A09BEB009F9B, 0xEB00A29BEB00A19B, 0xEB00A49BEB00A39B,
-    0xEB00A69BEB00A59B, 0xEB00A89BEB00A79B, 0xEB00AA9BEB00A99B,
-    0xEB00AC9BEB00AB9B, 0xEB00AE9BEB00AD9B, 0xEB00B09BEB00AF9B,
-    0xEB00B29BEB00B19B, 0xEB00B49BEB00B39B, 0xEB00B69BEB00B59B,
-    0xEB00B89BEB00B79B, 0xEB00BA9BEB00B99B, 0xEB00BC9BEB00BB9B,
-    0xEB00BE9BEB00BD9B, 0xEB00809CEB00BF9B, 0xEB00829CEB00819C,
-    0xEB00849CEB00839C, 0xEB00869CEB00859C, 0xEB00889CEB00879C,
-    0xEB008A9CEB00899C, 0xEB008C9CEB008B9C, 0xEB008E9CEB008D9C,
-    0xEB00909CEB008F9C, 0xEB00929CEB00919C, 0xEB00949CEB00939C,
-    0xEB00969CEB00959C, 0xEB00989CEB00979C, 0xEB009A9CEB00999C,
-    0xEB009C9CEB009B9C, 0xEB009E9CEB009D9C, 0xEB00A09CEB009F9C,
-    0xEB00A29CEB00A19C, 0xEB00A49CEB00A39C, 0xEB00A69CEB00A59C,
-    0xEB00A89CEB00A79C, 0xEB00AA9CEB00A99C, 0xEB00AC9CEB00AB9C,
-    0xEB00AE9CEB00AD9C, 0xEB00B09CEB00AF9C, 0xEB00B29CEB00B19C,
-    0xEB00B49CEB00B39C, 0xEB00B69CEB00B59C, 0xEB00B89CEB00B79C,
-    0xEB00BA9CEB00B99C, 0xEB00BC9CEB00BB9C, 0xEB00BE9CEB00BD9C,
-    0xEB00809DEB00BF9C, 0xEB00829DEB00819D, 0xEB00849DEB00839D,
-    0xEB00869DEB00859D, 0xEB00889DEB00879D, 0xEB008A9DEB00899D,
-    0xEB008C9DEB008B9D, 0xEB008E9DEB008D9D, 0xEB00909DEB008F9D,
-    0xEB00929DEB00919D, 0xEB00949DEB00939D, 0xEB00969DEB00959D,
-    0xEB00989DEB00979D, 0xEB009A9DEB00999D, 0xEB009C9DEB009B9D,
-    0xEB009E9DEB009D9D, 0xEB00A09DEB009F9D, 0xEB00A29DEB00A19D,
-    0xEB00A49DEB00A39D, 0xEB00A69DEB00A59D, 0xEB00A89DEB00A79D,
-    0xEB00AA9DEB00A99D, 0xEB00AC9DEB00AB9D, 0xEB00AE9DEB00AD9D,
-    0xEB00B09DEB00AF9D, 0xEB00B29DEB00B19D, 0xEB00B49DEB00B39D,
-    0xEB00B69DEB00B59D, 0xEB00B89DEB00B79D, 0xEB00BA9DEB00B99D,
-    0xEB00BC9DEB00BB9D, 0xEB00BE9DEB00BD9D, 0xEB00809EEB00BF9D,
-    0xEB00829EEB00819E, 0xEB00849EEB00839E, 0xEB00869EEB00859E,
-    0xEB00889EEB00879E, 0xEB008A9EEB00899E, 0xEB008C9EEB008B9E,
-    0xEB008E9EEB008D9E, 0xEB00909EEB008F9E, 0xEB00929EEB00919E,
-    0xEB00949EEB00939E, 0xEB00969EEB00959E, 0xEB00989EEB00979E,
-    0xEB009A9EEB00999E, 0xEB009C9EEB009B9E, 0xEB009E9EEB009D9E,
-    0xEB00A09EEB009F9E, 0xEB00A29EEB00A19E, 0xEB00A49EEB00A39E,
-    0xEB00A69EEB00A59E, 0xEB00A89EEB00A79E, 0xEB00AA9EEB00A99E,
-    0xEB00AC9EEB00AB9E, 0xEB00AE9EEB00AD9E, 0xEB00B09EEB00AF9E,
-    0xEB00B29EEB00B19E, 0xEB00B49EEB00B39E, 0xEB00B69EEB00B59E,
-    0xEB00B89EEB00B79E, 0xEB00BA9EEB00B99E, 0xEB00BC9EEB00BB9E,
-    0xEB00BE9EEB00BD9E, 0xEB00809FEB00BF9E, 0xEB00829FEB00819F,
-    0xEB00849FEB00839F, 0xEB00869FEB00859F, 0xEB00889FEB00879F,
-    0xEB008A9FEB00899F, 0xEB008C9FEB008B9F, 0xEB008E9FEB008D9F,
-    0xEB00909FEB008F9F, 0xEB00929FEB00919F, 0xEB00949FEB00939F,
-    0xEB00969FEB00959F, 0xEB00989FEB00979F, 0xEB009A9FEB00999F,
-    0xEB009C9FEB009B9F, 0xEB009E9FEB009D9F, 0xEB00A09FEB009F9F,
-    0xEB00A29FEB00A19F, 0xEB00A49FEB00A39F, 0xEB00A69FEB00A59F,
-    0xEB00A89FEB00A79F, 0xEB00AA9FEB00A99F, 0xEB00AC9FEB00AB9F,
-    0xEB00AE9FEB00AD9F, 0xEB00B09FEB00AF9F, 0xEB00B29FEB00B19F,
-    0xEB00B49FEB00B39F, 0xEB00B69FEB00B59F, 0xEB00B89FEB00B79F,
-    0xEB00BA9FEB00B99F, 0xEB00BC9FEB00BB9F, 0xEB00BE9FEB00BD9F,
-    0xEB0080A0EB00BF9F, 0xEB0082A0EB0081A0, 0xEB0084A0EB0083A0,
-    0xEB0086A0EB0085A0, 0xEB0088A0EB0087A0, 0xEB008AA0EB0089A0,
-    0xEB008CA0EB008BA0, 0xEB008EA0EB008DA0, 0xEB0090A0EB008FA0,
-    0xEB0092A0EB0091A0, 0xEB0094A0EB0093A0, 0xEB0096A0EB0095A0,
-    0xEB0098A0EB0097A0, 0xEB009AA0EB0099A0, 0xEB009CA0EB009BA0,
-    0xEB009EA0EB009DA0, 0xEB00A0A0EB009FA0, 0xEB00A2A0EB00A1A0,
-    0xEB00A4A0EB00A3A0, 0xEB00A6A0EB00A5A0, 0xEB00A8A0EB00A7A0,
-    0xEB00AAA0EB00A9A0, 0xEB00ACA0EB00ABA0, 0xEB00AEA0EB00ADA0,
-    0xEB00B0A0EB00AFA0, 0xEB00B2A0EB00B1A0, 0xEB00B4A0EB00B3A0,
-    0xEB00B6A0EB00B5A0, 0xEB00B8A0EB00B7A0, 0xEB00BAA0EB00B9A0,
-    0xEB00BCA0EB00BBA0, 0xEB00BEA0EB00BDA0, 0xEB0080A1EB00BFA0,
-    0xEB0082A1EB0081A1, 0xEB0084A1EB0083A1, 0xEB0086A1EB0085A1,
-    0xEB0088A1EB0087A1, 0xEB008AA1EB0089A1, 0xEB008CA1EB008BA1,
-    0xEB008EA1EB008DA1, 0xEB0090A1EB008FA1, 0xEB0092A1EB0091A1,
-    0xEB0094A1EB0093A1, 0xEB0096A1EB0095A1, 0xEB0098A1EB0097A1,
-    0xEB009AA1EB0099A1, 0xEB009CA1EB009BA1, 0xEB009EA1EB009DA1,
-    0xEB00A0A1EB009FA1, 0xEB00A2A1EB00A1A1, 0xEB00A4A1EB00A3A1,
-    0xEB00A6A1EB00A5A1, 0xEB00A8A1EB00A7A1, 0xEB00AAA1EB00A9A1,
-    0xEB00ACA1EB00ABA1, 0xEB00AEA1EB00ADA1, 0xEB00B0A1EB00AFA1,
-    0xEB00B2A1EB00B1A1, 0xEB00B4A1EB00B3A1, 0xEB00B6A1EB00B5A1,
-    0xEB00B8A1EB00B7A1, 0xEB00BAA1EB00B9A1, 0xEB00BCA1EB00BBA1,
-    0xEB00BEA1EB00BDA1, 0xEB0080A2EB00BFA1, 0xEB0082A2EB0081A2,
-    0xEB0084A2EB0083A2, 0xEB0086A2EB0085A2, 0xEB0088A2EB0087A2,
-    0xEB008AA2EB0089A2, 0xEB008CA2EB008BA2, 0xEB008EA2EB008DA2,
-    0xEB0090A2EB008FA2, 0xEB0092A2EB0091A2, 0xEB0094A2EB0093A2,
-    0xEB0096A2EB0095A2, 0xEB0098A2EB0097A2, 0xEB009AA2EB0099A2,
-    0xEB009CA2EB009BA2, 0xEB009EA2EB009DA2, 0xEB00A0A2EB009FA2,
-    0xEB00A2A2EB00A1A2, 0xEB00A4A2EB00A3A2, 0xEB00A6A2EB00A5A2,
-    0xEB00A8A2EB00A7A2, 0xEB00AAA2EB00A9A2, 0xEB00ACA2EB00ABA2,
-    0xEB00AEA2EB00ADA2, 0xEB00B0A2EB00AFA2, 0xEB00B2A2EB00B1A2,
-    0xEB00B4A2EB00B3A2, 0xEB00B6A2EB00B5A2, 0xEB00B8A2EB00B7A2,
-    0xEB00BAA2EB00B9A2, 0xEB00BCA2EB00BBA2, 0xEB00BEA2EB00BDA2,
-    0xEB0080A3EB00BFA2, 0xEB0082A3EB0081A3, 0xEB0084A3EB0083A3,
-    0xEB0086A3EB0085A3, 0xEB0088A3EB0087A3, 0xEB008AA3EB0089A3,
-    0xEB008CA3EB008BA3, 0xEB008EA3EB008DA3, 0xEB0090A3EB008FA3,
-    0xEB0092A3EB0091A3, 0xEB0094A3EB0093A3, 0xEB0096A3EB0095A3,
-    0xEB0098A3EB0097A3, 0xEB009AA3EB0099A3, 0xEB009CA3EB009BA3,
-    0xEB009EA3EB009DA3, 0xEB00A0A3EB009FA3, 0xEB00A2A3EB00A1A3,
-    0xEB00A4A3EB00A3A3, 0xEB00A6A3EB00A5A3, 0xEB00A8A3EB00A7A3,
-    0xEB00AAA3EB00A9A3, 0xEB00ACA3EB00ABA3, 0xEB00AEA3EB00ADA3,
-    0xEB00B0A3EB00AFA3, 0xEB00B2A3EB00B1A3, 0xEB00B4A3EB00B3A3,
-    0xEB00B6A3EB00B5A3, 0xEB00B8A3EB00B7A3, 0xEB00BAA3EB00B9A3,
-    0xEB00BCA3EB00BBA3, 0xEB00BEA3EB00BDA3, 0xEB0080A4EB00BFA3,
-    0xEB0082A4EB0081A4, 0xEB0084A4EB0083A4, 0xEB0086A4EB0085A4,
-    0xEB0088A4EB0087A4, 0xEB008AA4EB0089A4, 0xEB008CA4EB008BA4,
-    0xEB008EA4EB008DA4, 0xEB0090A4EB008FA4, 0xEB0092A4EB0091A4,
-    0xEB0094A4EB0093A4, 0xEB0096A4EB0095A4, 0xEB0098A4EB0097A4,
-    0xEB009AA4EB0099A4, 0xEB009CA4EB009BA4, 0xEB009EA4EB009DA4,
-    0xEB00A0A4EB009FA4, 0xEB00A2A4EB00A1A4, 0xEB00A4A4EB00A3A4,
-    0xEB00A6A4EB00A5A4, 0xEB00A8A4EB00A7A4, 0xEB00AAA4EB00A9A4,
-    0xEB00ACA4EB00ABA4, 0xEB00AEA4EB00ADA4, 0xEB00B0A4EB00AFA4,
-    0xEB00B2A4EB00B1A4, 0xEB00B4A4EB00B3A4, 0xEB00B6A4EB00B5A4,
-    0xEB00B8A4EB00B7A4, 0xEB00BAA4EB00B9A4, 0xEB00BCA4EB00BBA4,
-    0xEB00BEA4EB00BDA4, 0xEB0080A5EB00BFA4, 0xEB0082A5EB0081A5,
-    0xEB0084A5EB0083A5, 0xEB0086A5EB0085A5, 0xEB0088A5EB0087A5,
-    0xEB008AA5EB0089A5, 0xEB008CA5EB008BA5, 0xEB008EA5EB008DA5,
-    0xEB0090A5EB008FA5, 0xEB0092A5EB0091A5, 0xEB0094A5EB0093A5,
-    0xEB0096A5EB0095A5, 0xEB0098A5EB0097A5, 0xEB009AA5EB0099A5,
-    0xEB009CA5EB009BA5, 0xEB009EA5EB009DA5, 0xEB00A0A5EB009FA5,
-    0xEB00A2A5EB00A1A5, 0xEB00A4A5EB00A3A5, 0xEB00A6A5EB00A5A5,
-    0xEB00A8A5EB00A7A5, 0xEB00AAA5EB00A9A5, 0xEB00ACA5EB00ABA5,
-    0xEB00AEA5EB00ADA5, 0xEB00B0A5EB00AFA5, 0xEB00B2A5EB00B1A5,
-    0xEB00B4A5EB00B3A5, 0xEB00B6A5EB00B5A5, 0xEB00B8A5EB00B7A5,
-    0xEB00BAA5EB00B9A5, 0xEB00BCA5EB00BBA5, 0xEB00BEA5EB00BDA5,
-    0xEB0080A6EB00BFA5, 0xEB0082A6EB0081A6, 0xEB0084A6EB0083A6,
-    0xEB0086A6EB0085A6, 0xEB0088A6EB0087A6, 0xEB008AA6EB0089A6,
-    0xEB008CA6EB008BA6, 0xEB008EA6EB008DA6, 0xEB0090A6EB008FA6,
-    0xEB0092A6EB0091A6, 0xEB0094A6EB0093A6, 0xEB0096A6EB0095A6,
-    0xEB0098A6EB0097A6, 0xEB009AA6EB0099A6, 0xEB009CA6EB009BA6,
-    0xEB009EA6EB009DA6, 0xEB00A0A6EB009FA6, 0xEB00A2A6EB00A1A6,
-    0xEB00A4A6EB00A3A6, 0xEB00A6A6EB00A5A6, 0xEB00A8A6EB00A7A6,
-    0xEB00AAA6EB00A9A6, 0xEB00ACA6EB00ABA6, 0xEB00AEA6EB00ADA6,
-    0xEB00B0A6EB00AFA6, 0xEB00B2A6EB00B1A6, 0xEB00B4A6EB00B3A6,
-    0xEB00B6A6EB00B5A6, 0xEB00B8A6EB00B7A6, 0xEB00BAA6EB00B9A6,
-    0xEB00BCA6EB00BBA6, 0xEB00BEA6EB00BDA6, 0xEB0080A7EB00BFA6,
-    0xEB0082A7EB0081A7, 0xEB0084A7EB0083A7, 0xEB0086A7EB0085A7,
-    0xEB0088A7EB0087A7, 0xEB008AA7EB0089A7, 0xEB008CA7EB008BA7,
-    0xEB008EA7EB008DA7, 0xEB0090A7EB008FA7, 0xEB0092A7EB0091A7,
-    0xEB0094A7EB0093A7, 0xEB0096A7EB0095A7, 0xEB0098A7EB0097A7,
-    0xEB009AA7EB0099A7, 0xEB009CA7EB009BA7, 0xEB009EA7EB009DA7,
-    0xEB00A0A7EB009FA7, 0xEB00A2A7EB00A1A7, 0xEB00A4A7EB00A3A7,
-    0xEB00A6A7EB00A5A7, 0xEB00A8A7EB00A7A7, 0xEB00AAA7EB00A9A7,
-    0xEB00ACA7EB00ABA7, 0xEB00AEA7EB00ADA7, 0xEB00B0A7EB00AFA7,
-    0xEB00B2A7EB00B1A7, 0xEB00B4A7EB00B3A7, 0xEB00B6A7EB00B5A7,
-    0xEB00B8A7EB00B7A7, 0xEB00BAA7EB00B9A7, 0xEB00BCA7EB00BBA7,
-    0xEB00BEA7EB00BDA7, 0xEB0080A8EB00BFA7, 0xEB0082A8EB0081A8,
-    0xEB0084A8EB0083A8, 0xEB0086A8EB0085A8, 0xEB0088A8EB0087A8,
-    0xEB008AA8EB0089A8, 0xEB008CA8EB008BA8, 0xEB008EA8EB008DA8,
-    0xEB0090A8EB008FA8, 0xEB0092A8EB0091A8, 0xEB0094A8EB0093A8,
-    0xEB0096A8EB0095A8, 0xEB0098A8EB0097A8, 0xEB009AA8EB0099A8,
-    0xEB009CA8EB009BA8, 0xEB009EA8EB009DA8, 0xEB00A0A8EB009FA8,
-    0xEB00A2A8EB00A1A8, 0xEB00A4A8EB00A3A8, 0xEB00A6A8EB00A5A8,
-    0xEB00A8A8EB00A7A8, 0xEB00AAA8EB00A9A8, 0xEB00ACA8EB00ABA8,
-    0xEB00AEA8EB00ADA8, 0xEB00B0A8EB00AFA8, 0xEB00B2A8EB00B1A8,
-    0xEB00B4A8EB00B3A8, 0xEB00B6A8EB00B5A8, 0xEB00B8A8EB00B7A8,
-    0xEB00BAA8EB00B9A8, 0xEB00BCA8EB00BBA8, 0xEB00BEA8EB00BDA8,
-    0xEB0080A9EB00BFA8, 0xEB0082A9EB0081A9, 0xEB0084A9EB0083A9,
-    0xEB0086A9EB0085A9, 0xEB0088A9EB0087A9, 0xEB008AA9EB0089A9,
-    0xEB008CA9EB008BA9, 0xEB008EA9EB008DA9, 0xEB0090A9EB008FA9,
-    0xEB0092A9EB0091A9, 0xEB0094A9EB0093A9, 0xEB0096A9EB0095A9,
-    0xEB0098A9EB0097A9, 0xEB009AA9EB0099A9, 0xEB009CA9EB009BA9,
-    0xEB009EA9EB009DA9, 0xEB00A0A9EB009FA9, 0xEB00A2A9EB00A1A9,
-    0xEB00A4A9EB00A3A9, 0xEB00A6A9EB00A5A9, 0xEB00A8A9EB00A7A9,
-    0xEB00AAA9EB00A9A9, 0xEB00ACA9EB00ABA9, 0xEB00AEA9EB00ADA9,
-    0xEB00B0A9EB00AFA9, 0xEB00B2A9EB00B1A9, 0xEB00B4A9EB00B3A9,
-    0xEB00B6A9EB00B5A9, 0xEB00B8A9EB00B7A9, 0xEB00BAA9EB00B9A9,
-    0xEB00BCA9EB00BBA9, 0xEB00BEA9EB00BDA9, 0xEB0080AAEB00BFA9,
-    0xEB0082AAEB0081AA, 0xEB0084AAEB0083AA, 0xEB0086AAEB0085AA,
-    0xEB0088AAEB0087AA, 0xEB008AAAEB0089AA, 0xEB008CAAEB008BAA,
-    0xEB008EAAEB008DAA, 0xEB0090AAEB008FAA, 0xEB0092AAEB0091AA,
-    0xEB0094AAEB0093AA, 0xEB0096AAEB0095AA, 0xEB0098AAEB0097AA,
-    0xEB009AAAEB0099AA, 0xEB009CAAEB009BAA, 0xEB009EAAEB009DAA,
-    0xEB00A0AAEB009FAA, 0xEB00A2AAEB00A1AA, 0xEB00A4AAEB00A3AA,
-    0xEB00A6AAEB00A5AA, 0xEB00A8AAEB00A7AA, 0xEB00AAAAEB00A9AA,
-    0xEB00ACAAEB00ABAA, 0xEB00AEAAEB00ADAA, 0xEB00B0AAEB00AFAA,
-    0xEB00B2AAEB00B1AA, 0xEB00B4AAEB00B3AA, 0xEB00B6AAEB00B5AA,
-    0xEB00B8AAEB00B7AA, 0xEB00BAAAEB00B9AA, 0xEB00BCAAEB00BBAA,
-    0xEB00BEAAEB00BDAA, 0xEB0080ABEB00BFAA, 0xEB0082ABEB0081AB,
-    0xEB0084ABEB0083AB, 0xEB0086ABEB0085AB, 0xEB0088ABEB0087AB,
-    0xEB008AABEB0089AB, 0xEB008CABEB008BAB, 0xEB008EABEB008DAB,
-    0xEB0090ABEB008FAB, 0xEB0092ABEB0091AB, 0xEB0094ABEB0093AB,
-    0xEB0096ABEB0095AB, 0xEB0098ABEB0097AB, 0xEB009AABEB0099AB,
-    0xEB009CABEB009BAB, 0xEB009EABEB009DAB, 0xEB00A0ABEB009FAB,
-    0xEB00A2ABEB00A1AB, 0xEB00A4ABEB00A3AB, 0xEB00A6ABEB00A5AB,
-    0xEB00A8ABEB00A7AB, 0xEB00AAABEB00A9AB, 0xEB00ACABEB00ABAB,
-    0xEB00AEABEB00ADAB, 0xEB00B0ABEB00AFAB, 0xEB00B2ABEB00B1AB,
-    0xEB00B4ABEB00B3AB, 0xEB00B6ABEB00B5AB, 0xEB00B8ABEB00B7AB,
-    0xEB00BAABEB00B9AB, 0xEB00BCABEB00BBAB, 0xEB00BEABEB00BDAB,
-    0xEB0080ACEB00BFAB, 0xEB0082ACEB0081AC, 0xEB0084ACEB0083AC,
-    0xEB0086ACEB0085AC, 0xEB0088ACEB0087AC, 0xEB008AACEB0089AC,
-    0xEB008CACEB008BAC, 0xEB008EACEB008DAC, 0xEB0090ACEB008FAC,
-    0xEB0092ACEB0091AC, 0xEB0094ACEB0093AC, 0xEB0096ACEB0095AC,
-    0xEB0098ACEB0097AC, 0xEB009AACEB0099AC, 0xEB009CACEB009BAC,
-    0xEB009EACEB009DAC, 0xEB00A0ACEB009FAC, 0xEB00A2ACEB00A1AC,
-    0xEB00A4ACEB00A3AC, 0xEB00A6ACEB00A5AC, 0xEB00A8ACEB00A7AC,
-    0xEB00AAACEB00A9AC, 0xEB00ACACEB00ABAC, 0xEB00AEACEB00ADAC,
-    0xEB00B0ACEB00AFAC, 0xEB00B2ACEB00B1AC, 0xEB00B4ACEB00B3AC,
-    0xEB00B6ACEB00B5AC, 0xEB00B8ACEB00B7AC, 0xEB00BAACEB00B9AC,
-    0xEB00BCACEB00BBAC, 0xEB00BEACEB00BDAC, 0xEB0080ADEB00BFAC,
-    0xEB0082ADEB0081AD, 0xEB0084ADEB0083AD, 0xEB0086ADEB0085AD,
-    0xEB0088ADEB0087AD, 0xEB008AADEB0089AD, 0xEB008CADEB008BAD,
-    0xEB008EADEB008DAD, 0xEB0090ADEB008FAD, 0xEB0092ADEB0091AD,
-    0xEB0094ADEB0093AD, 0xEB0096ADEB0095AD, 0xEB0098ADEB0097AD,
-    0xEB009AADEB0099AD, 0xEB009CADEB009BAD, 0xEB009EADEB009DAD,
-    0xEB00A0ADEB009FAD, 0xEB00A2ADEB00A1AD, 0xEB00A4ADEB00A3AD,
-    0xEB00A6ADEB00A5AD, 0xEB00A8ADEB00A7AD, 0xEB00AAADEB00A9AD,
-    0xEB00ACADEB00ABAD, 0xEB00AEADEB00ADAD, 0xEB00B0ADEB00AFAD,
-    0xEB00B2ADEB00B1AD, 0xEB00B4ADEB00B3AD, 0xEB00B6ADEB00B5AD,
-    0xEB00B8ADEB00B7AD, 0xEB00BAADEB00B9AD, 0xEB00BCADEB00BBAD,
-    0xEB00BEADEB00BDAD, 0xEB0080AEEB00BFAD, 0xEB0082AEEB0081AE,
-    0xEB0084AEEB0083AE, 0xEB0086AEEB0085AE, 0xEB0088AEEB0087AE,
-    0xEB008AAEEB0089AE, 0xEB008CAEEB008BAE, 0xEB008EAEEB008DAE,
-    0xEB0090AEEB008FAE, 0xEB0092AEEB0091AE, 0xEB0094AEEB0093AE,
-    0xEB0096AEEB0095AE, 0xEB0098AEEB0097AE, 0xEB009AAEEB0099AE,
-    0xEB009CAEEB009BAE, 0xEB009EAEEB009DAE, 0xEB00A0AEEB009FAE,
-    0xEB00A2AEEB00A1AE, 0xEB00A4AEEB00A3AE, 0xEB00A6AEEB00A5AE,
-    0xEB00A8AEEB00A7AE, 0xEB00AAAEEB00A9AE, 0xEB00ACAEEB00ABAE,
-    0xEB00AEAEEB00ADAE, 0xEB00B0AEEB00AFAE, 0xEB00B2AEEB00B1AE,
-    0xEB00B4AEEB00B3AE, 0xEB00B6AEEB00B5AE, 0xEB00B8AEEB00B7AE,
-    0xEB00BAAEEB00B9AE, 0xEB00BCAEEB00BBAE, 0xEB00BEAEEB00BDAE,
-    0xEB0080AFEB00BFAE, 0xEB0082AFEB0081AF, 0xEB0084AFEB0083AF,
-    0xEB0086AFEB0085AF, 0xEB0088AFEB0087AF, 0xEB008AAFEB0089AF,
-    0xEB008CAFEB008BAF, 0xEB008EAFEB008DAF, 0xEB0090AFEB008FAF,
-    0xEB0092AFEB0091AF, 0xEB0094AFEB0093AF, 0xEB0096AFEB0095AF,
-    0xEB0098AFEB0097AF, 0xEB009AAFEB0099AF, 0xEB009CAFEB009BAF,
-    0xEB009EAFEB009DAF, 0xEB00A0AFEB009FAF, 0xEB00A2AFEB00A1AF,
-    0xEB00A4AFEB00A3AF, 0xEB00A6AFEB00A5AF, 0xEB00A8AFEB00A7AF,
-    0xEB00AAAFEB00A9AF, 0xEB00ACAFEB00ABAF, 0xEB00AEAFEB00ADAF,
-    0xEB00B0AFEB00AFAF, 0xEB00B2AFEB00B1AF, 0xEB00B4AFEB00B3AF,
-    0xEB00B6AFEB00B5AF, 0xEB00B8AFEB00B7AF, 0xEB00BAAFEB00B9AF,
-    0xEB00BCAFEB00BBAF, 0xEB00BEAFEB00BDAF, 0xEB0080B0EB00BFAF,
-    0xEB0082B0EB0081B0, 0xEB0084B0EB0083B0, 0xEB0086B0EB0085B0,
-    0xEB0088B0EB0087B0, 0xEB008AB0EB0089B0, 0xEB008CB0EB008BB0,
-    0xEB008EB0EB008DB0, 0xEB0090B0EB008FB0, 0xEB0092B0EB0091B0,
-    0xEB0094B0EB0093B0, 0xEB0096B0EB0095B0, 0xEB0098B0EB0097B0,
-    0xEB009AB0EB0099B0, 0xEB009CB0EB009BB0, 0xEB009EB0EB009DB0,
-    0xEB00A0B0EB009FB0, 0xEB00A2B0EB00A1B0, 0xEB00A4B0EB00A3B0,
-    0xEB00A6B0EB00A5B0, 0xEB00A8B0EB00A7B0, 0xEB00AAB0EB00A9B0,
-    0xEB00ACB0EB00ABB0, 0xEB00AEB0EB00ADB0, 0xEB00B0B0EB00AFB0,
-    0xEB00B2B0EB00B1B0, 0xEB00B4B0EB00B3B0, 0xEB00B6B0EB00B5B0,
-    0xEB00B8B0EB00B7B0, 0xEB00BAB0EB00B9B0, 0xEB00BCB0EB00BBB0,
-    0xEB00BEB0EB00BDB0, 0xEB0080B1EB00BFB0, 0xEB0082B1EB0081B1,
-    0xEB0084B1EB0083B1, 0xEB0086B1EB0085B1, 0xEB0088B1EB0087B1,
-    0xEB008AB1EB0089B1, 0xEB008CB1EB008BB1, 0xEB008EB1EB008DB1,
-    0xEB0090B1EB008FB1, 0xEB0092B1EB0091B1, 0xEB0094B1EB0093B1,
-    0xEB0096B1EB0095B1, 0xEB0098B1EB0097B1, 0xEB009AB1EB0099B1,
-    0xEB009CB1EB009BB1, 0xEB009EB1EB009DB1, 0xEB00A0B1EB009FB1,
-    0xEB00A2B1EB00A1B1, 0xEB00A4B1EB00A3B1, 0xEB00A6B1EB00A5B1,
-    0xEB00A8B1EB00A7B1, 0xEB00AAB1EB00A9B1, 0xEB00ACB1EB00ABB1,
-    0xEB00AEB1EB00ADB1, 0xEB00B0B1EB00AFB1, 0xEB00B2B1EB00B1B1,
-    0xEB00B4B1EB00B3B1, 0xEB00B6B1EB00B5B1, 0xEB00B8B1EB00B7B1,
-    0xEB00BAB1EB00B9B1, 0xEB00BCB1EB00BBB1, 0xEB00BEB1EB00BDB1,
-    0xEB0080B2EB00BFB1, 0xEB0082B2EB0081B2, 0xEB0084B2EB0083B2,
-    0xEB0086B2EB0085B2, 0xEB0088B2EB0087B2, 0xEB008AB2EB0089B2,
-    0xEB008CB2EB008BB2, 0xEB008EB2EB008DB2, 0xEB0090B2EB008FB2,
-    0xEB0092B2EB0091B2, 0xEB0094B2EB0093B2, 0xEB0096B2EB0095B2,
-    0xEB0098B2EB0097B2, 0xEB009AB2EB0099B2, 0xEB009CB2EB009BB2,
-    0xEB009EB2EB009DB2, 0xEB00A0B2EB009FB2, 0xEB00A2B2EB00A1B2,
-    0xEB00A4B2EB00A3B2, 0xEB00A6B2EB00A5B2, 0xEB00A8B2EB00A7B2,
-    0xEB00AAB2EB00A9B2, 0xEB00ACB2EB00ABB2, 0xEB00AEB2EB00ADB2,
-    0xEB00B0B2EB00AFB2, 0xEB00B2B2EB00B1B2, 0xEB00B4B2EB00B3B2,
-    0xEB00B6B2EB00B5B2, 0xEB00B8B2EB00B7B2, 0xEB00BAB2EB00B9B2,
-    0xEB00BCB2EB00BBB2, 0xEB00BEB2EB00BDB2, 0xEB0080B3EB00BFB2,
-    0xEB0082B3EB0081B3, 0xEB0084B3EB0083B3, 0xEB0086B3EB0085B3,
-    0xEB0088B3EB0087B3, 0xEB008AB3EB0089B3, 0xEB008CB3EB008BB3,
-    0xEB008EB3EB008DB3, 0xEB0090B3EB008FB3, 0xEB0092B3EB0091B3,
-    0xEB0094B3EB0093B3, 0xEB0096B3EB0095B3, 0xEB0098B3EB0097B3,
-    0xEB009AB3EB0099B3, 0xEB009CB3EB009BB3, 0xEB009EB3EB009DB3,
-    0xEB00A0B3EB009FB3, 0xEB00A2B3EB00A1B3, 0xEB00A4B3EB00A3B3,
-    0xEB00A6B3EB00A5B3, 0xEB00A8B3EB00A7B3, 0xEB00AAB3EB00A9B3,
-    0xEB00ACB3EB00ABB3, 0xEB00AEB3EB00ADB3, 0xEB00B0B3EB00AFB3,
-    0xEB00B2B3EB00B1B3, 0xEB00B4B3EB00B3B3, 0xEB00B6B3EB00B5B3,
-    0xEB00B8B3EB00B7B3, 0xEB00BAB3EB00B9B3, 0xEB00BCB3EB00BBB3,
-    0xEB00BEB3EB00BDB3, 0xEB0080B4EB00BFB3, 0xEB0082B4EB0081B4,
-    0xEB0084B4EB0083B4, 0xEB0086B4EB0085B4, 0xEB0088B4EB0087B4,
-    0xEB008AB4EB0089B4, 0xEB008CB4EB008BB4, 0xEB008EB4EB008DB4,
-    0xEB0090B4EB008FB4, 0xEB0092B4EB0091B4, 0xEB0094B4EB0093B4,
-    0xEB0096B4EB0095B4, 0xEB0098B4EB0097B4, 0xEB009AB4EB0099B4,
-    0xEB009CB4EB009BB4, 0xEB009EB4EB009DB4, 0xEB00A0B4EB009FB4,
-    0xEB00A2B4EB00A1B4, 0xEB00A4B4EB00A3B4, 0xEB00A6B4EB00A5B4,
-    0xEB00A8B4EB00A7B4, 0xEB00AAB4EB00A9B4, 0xEB00ACB4EB00ABB4,
-    0xEB00AEB4EB00ADB4, 0xEB00B0B4EB00AFB4, 0xEB00B2B4EB00B1B4,
-    0xEB00B4B4EB00B3B4, 0xEB00B6B4EB00B5B4, 0xEB00B8B4EB00B7B4,
-    0xEB00BAB4EB00B9B4, 0xEB00BCB4EB00BBB4, 0xEB00BEB4EB00BDB4,
-    0xEB0080B5EB00BFB4, 0xEB0082B5EB0081B5, 0xEB0084B5EB0083B5,
-    0xEB0086B5EB0085B5, 0xEB0088B5EB0087B5, 0xEB008AB5EB0089B5,
-    0xEB008CB5EB008BB5, 0xEB008EB5EB008DB5, 0xEB0090B5EB008FB5,
-    0xEB0092B5EB0091B5, 0xEB0094B5EB0093B5, 0xEB0096B5EB0095B5,
-    0xEB0098B5EB0097B5, 0xEB009AB5EB0099B5, 0xEB009CB5EB009BB5,
-    0xEB009EB5EB009DB5, 0xEB00A0B5EB009FB5, 0xEB00A2B5EB00A1B5,
-    0xEB00A4B5EB00A3B5, 0xEB00A6B5EB00A5B5, 0xEB00A8B5EB00A7B5,
-    0xEB00AAB5EB00A9B5, 0xEB00ACB5EB00ABB5, 0xEB00AEB5EB00ADB5,
-    0xEB00B0B5EB00AFB5, 0xEB00B2B5EB00B1B5, 0xEB00B4B5EB00B3B5,
-    0xEB00B6B5EB00B5B5, 0xEB00B8B5EB00B7B5, 0xEB00BAB5EB00B9B5,
-    0xEB00BCB5EB00BBB5, 0xEB00BEB5EB00BDB5, 0xEB0080B6EB00BFB5,
-    0xEB0082B6EB0081B6, 0xEB0084B6EB0083B6, 0xEB0086B6EB0085B6,
-    0xEB0088B6EB0087B6, 0xEB008AB6EB0089B6, 0xEB008CB6EB008BB6,
-    0xEB008EB6EB008DB6, 0xEB0090B6EB008FB6, 0xEB0092B6EB0091B6,
-    0xEB0094B6EB0093B6, 0xEB0096B6EB0095B6, 0xEB0098B6EB0097B6,
-    0xEB009AB6EB0099B6, 0xEB009CB6EB009BB6, 0xEB009EB6EB009DB6,
-    0xEB00A0B6EB009FB6, 0xEB00A2B6EB00A1B6, 0xEB00A4B6EB00A3B6,
-    0xEB00A6B6EB00A5B6, 0xEB00A8B6EB00A7B6, 0xEB00AAB6EB00A9B6,
-    0xEB00ACB6EB00ABB6, 0xEB00AEB6EB00ADB6, 0xEB00B0B6EB00AFB6,
-    0xEB00B2B6EB00B1B6, 0xEB00B4B6EB00B3B6, 0xEB00B6B6EB00B5B6,
-    0xEB00B8B6EB00B7B6, 0xEB00BAB6EB00B9B6, 0xEB00BCB6EB00BBB6,
-    0xEB00BEB6EB00BDB6, 0xEB0080B7EB00BFB6, 0xEB0082B7EB0081B7,
-    0xEB0084B7EB0083B7, 0xEB0086B7EB0085B7, 0xEB0088B7EB0087B7,
-    0xEB008AB7EB0089B7, 0xEB008CB7EB008BB7, 0xEB008EB7EB008DB7,
-    0xEB0090B7EB008FB7, 0xEB0092B7EB0091B7, 0xEB0094B7EB0093B7,
-    0xEB0096B7EB0095B7, 0xEB0098B7EB0097B7, 0xEB009AB7EB0099B7,
-    0xEB009CB7EB009BB7, 0xEB009EB7EB009DB7, 0xEB00A0B7EB009FB7,
-    0xEB00A2B7EB00A1B7, 0xEB00A4B7EB00A3B7, 0xEB00A6B7EB00A5B7,
-    0xEB00A8B7EB00A7B7, 0xEB00AAB7EB00A9B7, 0xEB00ACB7EB00ABB7,
-    0xEB00AEB7EB00ADB7, 0xEB00B0B7EB00AFB7, 0xEB00B2B7EB00B1B7,
-    0xEB00B4B7EB00B3B7, 0xEB00B6B7EB00B5B7, 0xEB00B8B7EB00B7B7,
-    0xEB00BAB7EB00B9B7, 0xEB00BCB7EB00BBB7, 0xEB00BEB7EB00BDB7,
-    0xEB0080B8EB00BFB7, 0xEB0082B8EB0081B8, 0xEB0084B8EB0083B8,
-    0xEB0086B8EB0085B8, 0xEB0088B8EB0087B8, 0xEB008AB8EB0089B8,
-    0xEB008CB8EB008BB8, 0xEB008EB8EB008DB8, 0xEB0090B8EB008FB8,
-    0xEB0092B8EB0091B8, 0xEB0094B8EB0093B8, 0xEB0096B8EB0095B8,
-    0xEB0098B8EB0097B8, 0xEB009AB8EB0099B8, 0xEB009CB8EB009BB8,
-    0xEB009EB8EB009DB8, 0xEB00A0B8EB009FB8, 0xEB00A2B8EB00A1B8,
-    0xEB00A4B8EB00A3B8, 0xEB00A6B8EB00A5B8, 0xEB00A8B8EB00A7B8,
-    0xEB00AAB8EB00A9B8, 0xEB00ACB8EB00ABB8, 0xEB00AEB8EB00ADB8,
-    0xEB00B0B8EB00AFB8, 0xEB00B2B8EB00B1B8, 0xEB00B4B8EB00B3B8,
-    0xEB00B6B8EB00B5B8, 0xEB00B8B8EB00B7B8, 0xEB00BAB8EB00B9B8,
-    0xEB00BCB8EB00BBB8, 0xEB00BEB8EB00BDB8, 0xEB0080B9EB00BFB8,
-    0xEB0082B9EB0081B9, 0xEB0084B9EB0083B9, 0xEB0086B9EB0085B9,
-    0xEB0088B9EB0087B9, 0xEB008AB9EB0089B9, 0xEB008CB9EB008BB9,
-    0xEB008EB9EB008DB9, 0xEB0090B9EB008FB9, 0xEB0092B9EB0091B9,
-    0xEB0094B9EB0093B9, 0xEB0096B9EB0095B9, 0xEB0098B9EB0097B9,
-    0xEB009AB9EB0099B9, 0xEB009CB9EB009BB9, 0xEB009EB9EB009DB9,
-    0xEB00A0B9EB009FB9, 0xEB00A2B9EB00A1B9, 0xEB00A4B9EB00A3B9,
-    0xEB00A6B9EB00A5B9, 0xEB00A8B9EB00A7B9, 0xEB00AAB9EB00A9B9,
-    0xEB00ACB9EB00ABB9, 0xEB00AEB9EB00ADB9, 0xEB00B0B9EB00AFB9,
-    0xEB00B2B9EB00B1B9, 0xEB00B4B9EB00B3B9, 0xEB00B6B9EB00B5B9,
-    0xEB00B8B9EB00B7B9, 0xEB00BAB9EB00B9B9, 0xEB00BCB9EB00BBB9,
-    0xEB00BEB9EB00BDB9, 0xEB0080BAEB00BFB9, 0xEB0082BAEB0081BA,
-    0xEB0084BAEB0083BA, 0xEB0086BAEB0085BA, 0xEB0088BAEB0087BA,
-    0xEB008ABAEB0089BA, 0xEB008CBAEB008BBA, 0xEB008EBAEB008DBA,
-    0xEB0090BAEB008FBA, 0xEB0092BAEB0091BA, 0xEB0094BAEB0093BA,
-    0xEB0096BAEB0095BA, 0xEB0098BAEB0097BA, 0xEB009ABAEB0099BA,
-    0xEB009CBAEB009BBA, 0xEB009EBAEB009DBA, 0xEB00A0BAEB009FBA,
-    0xEB00A2BAEB00A1BA, 0xEB00A4BAEB00A3BA, 0xEB00A6BAEB00A5BA,
-    0xEB00A8BAEB00A7BA, 0xEB00AABAEB00A9BA, 0xEB00ACBAEB00ABBA,
-    0xEB00AEBAEB00ADBA, 0xEB00B0BAEB00AFBA, 0xEB00B2BAEB00B1BA,
-    0xEB00B4BAEB00B3BA, 0xEB00B6BAEB00B5BA, 0xEB00B8BAEB00B7BA,
-    0xEB00BABAEB00B9BA, 0xEB00BCBAEB00BBBA, 0xEB00BEBAEB00BDBA,
-    0xEB0080BBEB00BFBA, 0xEB0082BBEB0081BB, 0xEB0084BBEB0083BB,
-    0xEB0086BBEB0085BB, 0xEB0088BBEB0087BB, 0xEB008ABBEB0089BB,
-    0xEB008CBBEB008BBB, 0xEB008EBBEB008DBB, 0xEB0090BBEB008FBB,
-    0xEB0092BBEB0091BB, 0xEB0094BBEB0093BB, 0xEB0096BBEB0095BB,
-    0xEB0098BBEB0097BB, 0xEB009ABBEB0099BB, 0xEB009CBBEB009BBB,
-    0xEB009EBBEB009DBB, 0xEB00A0BBEB009FBB, 0xEB00A2BBEB00A1BB,
-    0xEB00A4BBEB00A3BB, 0xEB00A6BBEB00A5BB, 0xEB00A8BBEB00A7BB,
-    0xEB00AABBEB00A9BB, 0xEB00ACBBEB00ABBB, 0xEB00AEBBEB00ADBB,
-    0xEB00B0BBEB00AFBB, 0xEB00B2BBEB00B1BB, 0xEB00B4BBEB00B3BB,
-    0xEB00B6BBEB00B5BB, 0xEB00B8BBEB00B7BB, 0xEB00BABBEB00B9BB,
-    0xEB00BCBBEB00BBBB, 0xEB00BEBBEB00BDBB, 0xEB0080BCEB00BFBB,
-    0xEB0082BCEB0081BC, 0xEB0084BCEB0083BC, 0xEB0086BCEB0085BC,
-    0xEB0088BCEB0087BC, 0xEB008ABCEB0089BC, 0xEB008CBCEB008BBC,
-    0xEB008EBCEB008DBC, 0xEB0090BCEB008FBC, 0xEB0092BCEB0091BC,
-    0xEB0094BCEB0093BC, 0xEB0096BCEB0095BC, 0xEB0098BCEB0097BC,
-    0xEB009ABCEB0099BC, 0xEB009CBCEB009BBC, 0xEB009EBCEB009DBC,
-    0xEB00A0BCEB009FBC, 0xEB00A2BCEB00A1BC, 0xEB00A4BCEB00A3BC,
-    0xEB00A6BCEB00A5BC, 0xEB00A8BCEB00A7BC, 0xEB00AABCEB00A9BC,
-    0xEB00ACBCEB00ABBC, 0xEB00AEBCEB00ADBC, 0xEB00B0BCEB00AFBC,
-    0xEB00B2BCEB00B1BC, 0xEB00B4BCEB00B3BC, 0xEB00B6BCEB00B5BC,
-    0xEB00B8BCEB00B7BC, 0xEB00BABCEB00B9BC, 0xEB00BCBCEB00BBBC,
-    0xEB00BEBCEB00BDBC, 0xEB0080BDEB00BFBC, 0xEB0082BDEB0081BD,
-    0xEB0084BDEB0083BD, 0xEB0086BDEB0085BD, 0xEB0088BDEB0087BD,
-    0xEB008ABDEB0089BD, 0xEB008CBDEB008BBD, 0xEB008EBDEB008DBD,
-    0xEB0090BDEB008FBD, 0xEB0092BDEB0091BD, 0xEB0094BDEB0093BD,
-    0xEB0096BDEB0095BD, 0xEB0098BDEB0097BD, 0xEB009ABDEB0099BD,
-    0xEB009CBDEB009BBD, 0xEB009EBDEB009DBD, 0xEB00A0BDEB009FBD,
-    0xEB00A2BDEB00A1BD, 0xEB00A4BDEB00A3BD, 0xEB00A6BDEB00A5BD,
-    0xEB00A8BDEB00A7BD, 0xEB00AABDEB00A9BD, 0xEB00ACBDEB00ABBD,
-    0xEB00AEBDEB00ADBD, 0xEB00B0BDEB00AFBD, 0xEB00B2BDEB00B1BD,
-    0xEB00B4BDEB00B3BD, 0xEB00B6BDEB00B5BD, 0xEB00B8BDEB00B7BD,
-    0xEB00BABDEB00B9BD, 0xEB00BCBDEB00BBBD, 0xEB00BEBDEB00BDBD,
-    0xEB0080BEEB00BFBD, 0xEB0082BEEB0081BE, 0xEB0084BEEB0083BE,
-    0xEB0086BEEB0085BE, 0xEB0088BEEB0087BE, 0xEB008ABEEB0089BE,
-    0xEB008CBEEB008BBE, 0xEB008EBEEB008DBE, 0xEB0090BEEB008FBE,
-    0xEB0092BEEB0091BE, 0xEB0094BEEB0093BE, 0xEB0096BEEB0095BE,
-    0xEB0098BEEB0097BE, 0xEB009ABEEB0099BE, 0xEB009CBEEB009BBE,
-    0xEB009EBEEB009DBE, 0xEB00A0BEEB009FBE, 0xEB00A2BEEB00A1BE,
-    0xEB00A4BEEB00A3BE, 0xEB00A6BEEB00A5BE, 0xEB00A8BEEB00A7BE,
-    0xEB00AABEEB00A9BE, 0xEB00ACBEEB00ABBE, 0xEB00AEBEEB00ADBE,
-    0xEB00B0BEEB00AFBE, 0xEB00B2BEEB00B1BE, 0xEB00B4BEEB00B3BE,
-    0xEB00B6BEEB00B5BE, 0xEB00B8BEEB00B7BE, 0xEB00BABEEB00B9BE,
-    0xEB00BCBEEB00BBBE, 0xEB00BEBEEB00BDBE, 0xEB0080BFEB00BFBE,
-    0xEB0082BFEB0081BF, 0xEB0084BFEB0083BF, 0xEB0086BFEB0085BF,
-    0xEB0088BFEB0087BF, 0xEB008ABFEB0089BF, 0xEB008CBFEB008BBF,
-    0xEB008EBFEB008DBF, 0xEB0090BFEB008FBF, 0xEB0092BFEB0091BF,
-    0xEB0094BFEB0093BF, 0xEB0096BFEB0095BF, 0xEB0098BFEB0097BF,
-    0xEB009ABFEB0099BF, 0xEB009CBFEB009BBF, 0xEB009EBFEB009DBF,
-    0xEB00A0BFEB009FBF, 0xEB00A2BFEB00A1BF, 0xEB00A4BFEB00A3BF,
-    0xEB00A6BFEB00A5BF, 0xEB00A8BFEB00A7BF, 0xEB00AABFEB00A9BF,
-    0xEB00ACBFEB00ABBF, 0xEB00AEBFEB00ADBF, 0xEB00B0BFEB00AFBF,
-    0xEB00B2BFEB00B1BF, 0xEB00B4BFEB00B3BF, 0xEB00B6BFEB00B5BF,
-    0xEB00B8BFEB00B7BF, 0xEB00BABFEB00B9BF, 0xEB00BCBFEB00BBBF,
-    0xEB00BEBFEB00BDBF, 0xEC008080EC00BFBF, 0xEC008280EC008180,
-    0xEC008480EC008380, 0xEC008680EC008580, 0xEC008880EC008780,
-    0xEC008A80EC008980, 0xEC008C80EC008B80, 0xEC008E80EC008D80,
-    0xEC009080EC008F80, 0xEC009280EC009180, 0xEC009480EC009380,
-    0xEC009680EC009580, 0xEC009880EC009780, 0xEC009A80EC009980,
-    0xEC009C80EC009B80, 0xEC009E80EC009D80, 0xEC00A080EC009F80,
-    0xEC00A280EC00A180, 0xEC00A480EC00A380, 0xEC00A680EC00A580,
-    0xEC00A880EC00A780, 0xEC00AA80EC00A980, 0xEC00AC80EC00AB80,
-    0xEC00AE80EC00AD80, 0xEC00B080EC00AF80, 0xEC00B280EC00B180,
-    0xEC00B480EC00B380, 0xEC00B680EC00B580, 0xEC00B880EC00B780,
-    0xEC00BA80EC00B980, 0xEC00BC80EC00BB80, 0xEC00BE80EC00BD80,
-    0xEC008081EC00BF80, 0xEC008281EC008181, 0xEC008481EC008381,
-    0xEC008681EC008581, 0xEC008881EC008781, 0xEC008A81EC008981,
-    0xEC008C81EC008B81, 0xEC008E81EC008D81, 0xEC009081EC008F81,
-    0xEC009281EC009181, 0xEC009481EC009381, 0xEC009681EC009581,
-    0xEC009881EC009781, 0xEC009A81EC009981, 0xEC009C81EC009B81,
-    0xEC009E81EC009D81, 0xEC00A081EC009F81, 0xEC00A281EC00A181,
-    0xEC00A481EC00A381, 0xEC00A681EC00A581, 0xEC00A881EC00A781,
-    0xEC00AA81EC00A981, 0xEC00AC81EC00AB81, 0xEC00AE81EC00AD81,
-    0xEC00B081EC00AF81, 0xEC00B281EC00B181, 0xEC00B481EC00B381,
-    0xEC00B681EC00B581, 0xEC00B881EC00B781, 0xEC00BA81EC00B981,
-    0xEC00BC81EC00BB81, 0xEC00BE81EC00BD81, 0xEC008082EC00BF81,
-    0xEC008282EC008182, 0xEC008482EC008382, 0xEC008682EC008582,
-    0xEC008882EC008782, 0xEC008A82EC008982, 0xEC008C82EC008B82,
-    0xEC008E82EC008D82, 0xEC009082EC008F82, 0xEC009282EC009182,
-    0xEC009482EC009382, 0xEC009682EC009582, 0xEC009882EC009782,
-    0xEC009A82EC009982, 0xEC009C82EC009B82, 0xEC009E82EC009D82,
-    0xEC00A082EC009F82, 0xEC00A282EC00A182, 0xEC00A482EC00A382,
-    0xEC00A682EC00A582, 0xEC00A882EC00A782, 0xEC00AA82EC00A982,
-    0xEC00AC82EC00AB82, 0xEC00AE82EC00AD82, 0xEC00B082EC00AF82,
-    0xEC00B282EC00B182, 0xEC00B482EC00B382, 0xEC00B682EC00B582,
-    0xEC00B882EC00B782, 0xEC00BA82EC00B982, 0xEC00BC82EC00BB82,
-    0xEC00BE82EC00BD82, 0xEC008083EC00BF82, 0xEC008283EC008183,
-    0xEC008483EC008383, 0xEC008683EC008583, 0xEC008883EC008783,
-    0xEC008A83EC008983, 0xEC008C83EC008B83, 0xEC008E83EC008D83,
-    0xEC009083EC008F83, 0xEC009283EC009183, 0xEC009483EC009383,
-    0xEC009683EC009583, 0xEC009883EC009783, 0xEC009A83EC009983,
-    0xEC009C83EC009B83, 0xEC009E83EC009D83, 0xEC00A083EC009F83,
-    0xEC00A283EC00A183, 0xEC00A483EC00A383, 0xEC00A683EC00A583,
-    0xEC00A883EC00A783, 0xEC00AA83EC00A983, 0xEC00AC83EC00AB83,
-    0xEC00AE83EC00AD83, 0xEC00B083EC00AF83, 0xEC00B283EC00B183,
-    0xEC00B483EC00B383, 0xEC00B683EC00B583, 0xEC00B883EC00B783,
-    0xEC00BA83EC00B983, 0xEC00BC83EC00BB83, 0xEC00BE83EC00BD83,
-    0xEC008084EC00BF83, 0xEC008284EC008184, 0xEC008484EC008384,
-    0xEC008684EC008584, 0xEC008884EC008784, 0xEC008A84EC008984,
-    0xEC008C84EC008B84, 0xEC008E84EC008D84, 0xEC009084EC008F84,
-    0xEC009284EC009184, 0xEC009484EC009384, 0xEC009684EC009584,
-    0xEC009884EC009784, 0xEC009A84EC009984, 0xEC009C84EC009B84,
-    0xEC009E84EC009D84, 0xEC00A084EC009F84, 0xEC00A284EC00A184,
-    0xEC00A484EC00A384, 0xEC00A684EC00A584, 0xEC00A884EC00A784,
-    0xEC00AA84EC00A984, 0xEC00AC84EC00AB84, 0xEC00AE84EC00AD84,
-    0xEC00B084EC00AF84, 0xEC00B284EC00B184, 0xEC00B484EC00B384,
-    0xEC00B684EC00B584, 0xEC00B884EC00B784, 0xEC00BA84EC00B984,
-    0xEC00BC84EC00BB84, 0xEC00BE84EC00BD84, 0xEC008085EC00BF84,
-    0xEC008285EC008185, 0xEC008485EC008385, 0xEC008685EC008585,
-    0xEC008885EC008785, 0xEC008A85EC008985, 0xEC008C85EC008B85,
-    0xEC008E85EC008D85, 0xEC009085EC008F85, 0xEC009285EC009185,
-    0xEC009485EC009385, 0xEC009685EC009585, 0xEC009885EC009785,
-    0xEC009A85EC009985, 0xEC009C85EC009B85, 0xEC009E85EC009D85,
-    0xEC00A085EC009F85, 0xEC00A285EC00A185, 0xEC00A485EC00A385,
-    0xEC00A685EC00A585, 0xEC00A885EC00A785, 0xEC00AA85EC00A985,
-    0xEC00AC85EC00AB85, 0xEC00AE85EC00AD85, 0xEC00B085EC00AF85,
-    0xEC00B285EC00B185, 0xEC00B485EC00B385, 0xEC00B685EC00B585,
-    0xEC00B885EC00B785, 0xEC00BA85EC00B985, 0xEC00BC85EC00BB85,
-    0xEC00BE85EC00BD85, 0xEC008086EC00BF85, 0xEC008286EC008186,
-    0xEC008486EC008386, 0xEC008686EC008586, 0xEC008886EC008786,
-    0xEC008A86EC008986, 0xEC008C86EC008B86, 0xEC008E86EC008D86,
-    0xEC009086EC008F86, 0xEC009286EC009186, 0xEC009486EC009386,
-    0xEC009686EC009586, 0xEC009886EC009786, 0xEC009A86EC009986,
-    0xEC009C86EC009B86, 0xEC009E86EC009D86, 0xEC00A086EC009F86,
-    0xEC00A286EC00A186, 0xEC00A486EC00A386, 0xEC00A686EC00A586,
-    0xEC00A886EC00A786, 0xEC00AA86EC00A986, 0xEC00AC86EC00AB86,
-    0xEC00AE86EC00AD86, 0xEC00B086EC00AF86, 0xEC00B286EC00B186,
-    0xEC00B486EC00B386, 0xEC00B686EC00B586, 0xEC00B886EC00B786,
-    0xEC00BA86EC00B986, 0xEC00BC86EC00BB86, 0xEC00BE86EC00BD86,
-    0xEC008087EC00BF86, 0xEC008287EC008187, 0xEC008487EC008387,
-    0xEC008687EC008587, 0xEC008887EC008787, 0xEC008A87EC008987,
-    0xEC008C87EC008B87, 0xEC008E87EC008D87, 0xEC009087EC008F87,
-    0xEC009287EC009187, 0xEC009487EC009387, 0xEC009687EC009587,
-    0xEC009887EC009787, 0xEC009A87EC009987, 0xEC009C87EC009B87,
-    0xEC009E87EC009D87, 0xEC00A087EC009F87, 0xEC00A287EC00A187,
-    0xEC00A487EC00A387, 0xEC00A687EC00A587, 0xEC00A887EC00A787,
-    0xEC00AA87EC00A987, 0xEC00AC87EC00AB87, 0xEC00AE87EC00AD87,
-    0xEC00B087EC00AF87, 0xEC00B287EC00B187, 0xEC00B487EC00B387,
-    0xEC00B687EC00B587, 0xEC00B887EC00B787, 0xEC00BA87EC00B987,
-    0xEC00BC87EC00BB87, 0xEC00BE87EC00BD87, 0xEC008088EC00BF87,
-    0xEC008288EC008188, 0xEC008488EC008388, 0xEC008688EC008588,
-    0xEC008888EC008788, 0xEC008A88EC008988, 0xEC008C88EC008B88,
-    0xEC008E88EC008D88, 0xEC009088EC008F88, 0xEC009288EC009188,
-    0xEC009488EC009388, 0xEC009688EC009588, 0xEC009888EC009788,
-    0xEC009A88EC009988, 0xEC009C88EC009B88, 0xEC009E88EC009D88,
-    0xEC00A088EC009F88, 0xEC00A288EC00A188, 0xEC00A488EC00A388,
-    0xEC00A688EC00A588, 0xEC00A888EC00A788, 0xEC00AA88EC00A988,
-    0xEC00AC88EC00AB88, 0xEC00AE88EC00AD88, 0xEC00B088EC00AF88,
-    0xEC00B288EC00B188, 0xEC00B488EC00B388, 0xEC00B688EC00B588,
-    0xEC00B888EC00B788, 0xEC00BA88EC00B988, 0xEC00BC88EC00BB88,
-    0xEC00BE88EC00BD88, 0xEC008089EC00BF88, 0xEC008289EC008189,
-    0xEC008489EC008389, 0xEC008689EC008589, 0xEC008889EC008789,
-    0xEC008A89EC008989, 0xEC008C89EC008B89, 0xEC008E89EC008D89,
-    0xEC009089EC008F89, 0xEC009289EC009189, 0xEC009489EC009389,
-    0xEC009689EC009589, 0xEC009889EC009789, 0xEC009A89EC009989,
-    0xEC009C89EC009B89, 0xEC009E89EC009D89, 0xEC00A089EC009F89,
-    0xEC00A289EC00A189, 0xEC00A489EC00A389, 0xEC00A689EC00A589,
-    0xEC00A889EC00A789, 0xEC00AA89EC00A989, 0xEC00AC89EC00AB89,
-    0xEC00AE89EC00AD89, 0xEC00B089EC00AF89, 0xEC00B289EC00B189,
-    0xEC00B489EC00B389, 0xEC00B689EC00B589, 0xEC00B889EC00B789,
-    0xEC00BA89EC00B989, 0xEC00BC89EC00BB89, 0xEC00BE89EC00BD89,
-    0xEC00808AEC00BF89, 0xEC00828AEC00818A, 0xEC00848AEC00838A,
-    0xEC00868AEC00858A, 0xEC00888AEC00878A, 0xEC008A8AEC00898A,
-    0xEC008C8AEC008B8A, 0xEC008E8AEC008D8A, 0xEC00908AEC008F8A,
-    0xEC00928AEC00918A, 0xEC00948AEC00938A, 0xEC00968AEC00958A,
-    0xEC00988AEC00978A, 0xEC009A8AEC00998A, 0xEC009C8AEC009B8A,
-    0xEC009E8AEC009D8A, 0xEC00A08AEC009F8A, 0xEC00A28AEC00A18A,
-    0xEC00A48AEC00A38A, 0xEC00A68AEC00A58A, 0xEC00A88AEC00A78A,
-    0xEC00AA8AEC00A98A, 0xEC00AC8AEC00AB8A, 0xEC00AE8AEC00AD8A,
-    0xEC00B08AEC00AF8A, 0xEC00B28AEC00B18A, 0xEC00B48AEC00B38A,
-    0xEC00B68AEC00B58A, 0xEC00B88AEC00B78A, 0xEC00BA8AEC00B98A,
-    0xEC00BC8AEC00BB8A, 0xEC00BE8AEC00BD8A, 0xEC00808BEC00BF8A,
-    0xEC00828BEC00818B, 0xEC00848BEC00838B, 0xEC00868BEC00858B,
-    0xEC00888BEC00878B, 0xEC008A8BEC00898B, 0xEC008C8BEC008B8B,
-    0xEC008E8BEC008D8B, 0xEC00908BEC008F8B, 0xEC00928BEC00918B,
-    0xEC00948BEC00938B, 0xEC00968BEC00958B, 0xEC00988BEC00978B,
-    0xEC009A8BEC00998B, 0xEC009C8BEC009B8B, 0xEC009E8BEC009D8B,
-    0xEC00A08BEC009F8B, 0xEC00A28BEC00A18B, 0xEC00A48BEC00A38B,
-    0xEC00A68BEC00A58B, 0xEC00A88BEC00A78B, 0xEC00AA8BEC00A98B,
-    0xEC00AC8BEC00AB8B, 0xEC00AE8BEC00AD8B, 0xEC00B08BEC00AF8B,
-    0xEC00B28BEC00B18B, 0xEC00B48BEC00B38B, 0xEC00B68BEC00B58B,
-    0xEC00B88BEC00B78B, 0xEC00BA8BEC00B98B, 0xEC00BC8BEC00BB8B,
-    0xEC00BE8BEC00BD8B, 0xEC00808CEC00BF8B, 0xEC00828CEC00818C,
-    0xEC00848CEC00838C, 0xEC00868CEC00858C, 0xEC00888CEC00878C,
-    0xEC008A8CEC00898C, 0xEC008C8CEC008B8C, 0xEC008E8CEC008D8C,
-    0xEC00908CEC008F8C, 0xEC00928CEC00918C, 0xEC00948CEC00938C,
-    0xEC00968CEC00958C, 0xEC00988CEC00978C, 0xEC009A8CEC00998C,
-    0xEC009C8CEC009B8C, 0xEC009E8CEC009D8C, 0xEC00A08CEC009F8C,
-    0xEC00A28CEC00A18C, 0xEC00A48CEC00A38C, 0xEC00A68CEC00A58C,
-    0xEC00A88CEC00A78C, 0xEC00AA8CEC00A98C, 0xEC00AC8CEC00AB8C,
-    0xEC00AE8CEC00AD8C, 0xEC00B08CEC00AF8C, 0xEC00B28CEC00B18C,
-    0xEC00B48CEC00B38C, 0xEC00B68CEC00B58C, 0xEC00B88CEC00B78C,
-    0xEC00BA8CEC00B98C, 0xEC00BC8CEC00BB8C, 0xEC00BE8CEC00BD8C,
-    0xEC00808DEC00BF8C, 0xEC00828DEC00818D, 0xEC00848DEC00838D,
-    0xEC00868DEC00858D, 0xEC00888DEC00878D, 0xEC008A8DEC00898D,
-    0xEC008C8DEC008B8D, 0xEC008E8DEC008D8D, 0xEC00908DEC008F8D,
-    0xEC00928DEC00918D, 0xEC00948DEC00938D, 0xEC00968DEC00958D,
-    0xEC00988DEC00978D, 0xEC009A8DEC00998D, 0xEC009C8DEC009B8D,
-    0xEC009E8DEC009D8D, 0xEC00A08DEC009F8D, 0xEC00A28DEC00A18D,
-    0xEC00A48DEC00A38D, 0xEC00A68DEC00A58D, 0xEC00A88DEC00A78D,
-    0xEC00AA8DEC00A98D, 0xEC00AC8DEC00AB8D, 0xEC00AE8DEC00AD8D,
-    0xEC00B08DEC00AF8D, 0xEC00B28DEC00B18D, 0xEC00B48DEC00B38D,
-    0xEC00B68DEC00B58D, 0xEC00B88DEC00B78D, 0xEC00BA8DEC00B98D,
-    0xEC00BC8DEC00BB8D, 0xEC00BE8DEC00BD8D, 0xEC00808EEC00BF8D,
-    0xEC00828EEC00818E, 0xEC00848EEC00838E, 0xEC00868EEC00858E,
-    0xEC00888EEC00878E, 0xEC008A8EEC00898E, 0xEC008C8EEC008B8E,
-    0xEC008E8EEC008D8E, 0xEC00908EEC008F8E, 0xEC00928EEC00918E,
-    0xEC00948EEC00938E, 0xEC00968EEC00958E, 0xEC00988EEC00978E,
-    0xEC009A8EEC00998E, 0xEC009C8EEC009B8E, 0xEC009E8EEC009D8E,
-    0xEC00A08EEC009F8E, 0xEC00A28EEC00A18E, 0xEC00A48EEC00A38E,
-    0xEC00A68EEC00A58E, 0xEC00A88EEC00A78E, 0xEC00AA8EEC00A98E,
-    0xEC00AC8EEC00AB8E, 0xEC00AE8EEC00AD8E, 0xEC00B08EEC00AF8E,
-    0xEC00B28EEC00B18E, 0xEC00B48EEC00B38E, 0xEC00B68EEC00B58E,
-    0xEC00B88EEC00B78E, 0xEC00BA8EEC00B98E, 0xEC00BC8EEC00BB8E,
-    0xEC00BE8EEC00BD8E, 0xEC00808FEC00BF8E, 0xEC00828FEC00818F,
-    0xEC00848FEC00838F, 0xEC00868FEC00858F, 0xEC00888FEC00878F,
-    0xEC008A8FEC00898F, 0xEC008C8FEC008B8F, 0xEC008E8FEC008D8F,
-    0xEC00908FEC008F8F, 0xEC00928FEC00918F, 0xEC00948FEC00938F,
-    0xEC00968FEC00958F, 0xEC00988FEC00978F, 0xEC009A8FEC00998F,
-    0xEC009C8FEC009B8F, 0xEC009E8FEC009D8F, 0xEC00A08FEC009F8F,
-    0xEC00A28FEC00A18F, 0xEC00A48FEC00A38F, 0xEC00A68FEC00A58F,
-    0xEC00A88FEC00A78F, 0xEC00AA8FEC00A98F, 0xEC00AC8FEC00AB8F,
-    0xEC00AE8FEC00AD8F, 0xEC00B08FEC00AF8F, 0xEC00B28FEC00B18F,
-    0xEC00B48FEC00B38F, 0xEC00B68FEC00B58F, 0xEC00B88FEC00B78F,
-    0xEC00BA8FEC00B98F, 0xEC00BC8FEC00BB8F, 0xEC00BE8FEC00BD8F,
-    0xEC008090EC00BF8F, 0xEC008290EC008190, 0xEC008490EC008390,
-    0xEC008690EC008590, 0xEC008890EC008790, 0xEC008A90EC008990,
-    0xEC008C90EC008B90, 0xEC008E90EC008D90, 0xEC009090EC008F90,
-    0xEC009290EC009190, 0xEC009490EC009390, 0xEC009690EC009590,
-    0xEC009890EC009790, 0xEC009A90EC009990, 0xEC009C90EC009B90,
-    0xEC009E90EC009D90, 0xEC00A090EC009F90, 0xEC00A290EC00A190,
-    0xEC00A490EC00A390, 0xEC00A690EC00A590, 0xEC00A890EC00A790,
-    0xEC00AA90EC00A990, 0xEC00AC90EC00AB90, 0xEC00AE90EC00AD90,
-    0xEC00B090EC00AF90, 0xEC00B290EC00B190, 0xEC00B490EC00B390,
-    0xEC00B690EC00B590, 0xEC00B890EC00B790, 0xEC00BA90EC00B990,
-    0xEC00BC90EC00BB90, 0xEC00BE90EC00BD90, 0xEC008091EC00BF90,
-    0xEC008291EC008191, 0xEC008491EC008391, 0xEC008691EC008591,
-    0xEC008891EC008791, 0xEC008A91EC008991, 0xEC008C91EC008B91,
-    0xEC008E91EC008D91, 0xEC009091EC008F91, 0xEC009291EC009191,
-    0xEC009491EC009391, 0xEC009691EC009591, 0xEC009891EC009791,
-    0xEC009A91EC009991, 0xEC009C91EC009B91, 0xEC009E91EC009D91,
-    0xEC00A091EC009F91, 0xEC00A291EC00A191, 0xEC00A491EC00A391,
-    0xEC00A691EC00A591, 0xEC00A891EC00A791, 0xEC00AA91EC00A991,
-    0xEC00AC91EC00AB91, 0xEC00AE91EC00AD91, 0xEC00B091EC00AF91,
-    0xEC00B291EC00B191, 0xEC00B491EC00B391, 0xEC00B691EC00B591,
-    0xEC00B891EC00B791, 0xEC00BA91EC00B991, 0xEC00BC91EC00BB91,
-    0xEC00BE91EC00BD91, 0xEC008092EC00BF91, 0xEC008292EC008192,
-    0xEC008492EC008392, 0xEC008692EC008592, 0xEC008892EC008792,
-    0xEC008A92EC008992, 0xEC008C92EC008B92, 0xEC008E92EC008D92,
-    0xEC009092EC008F92, 0xEC009292EC009192, 0xEC009492EC009392,
-    0xEC009692EC009592, 0xEC009892EC009792, 0xEC009A92EC009992,
-    0xEC009C92EC009B92, 0xEC009E92EC009D92, 0xEC00A092EC009F92,
-    0xEC00A292EC00A192, 0xEC00A492EC00A392, 0xEC00A692EC00A592,
-    0xEC00A892EC00A792, 0xEC00AA92EC00A992, 0xEC00AC92EC00AB92,
-    0xEC00AE92EC00AD92, 0xEC00B092EC00AF92, 0xEC00B292EC00B192,
-    0xEC00B492EC00B392, 0xEC00B692EC00B592, 0xEC00B892EC00B792,
-    0xEC00BA92EC00B992, 0xEC00BC92EC00BB92, 0xEC00BE92EC00BD92,
-    0xEC008093EC00BF92, 0xEC008293EC008193, 0xEC008493EC008393,
-    0xEC008693EC008593, 0xEC008893EC008793, 0xEC008A93EC008993,
-    0xEC008C93EC008B93, 0xEC008E93EC008D93, 0xEC009093EC008F93,
-    0xEC009293EC009193, 0xEC009493EC009393, 0xEC009693EC009593,
-    0xEC009893EC009793, 0xEC009A93EC009993, 0xEC009C93EC009B93,
-    0xEC009E93EC009D93, 0xEC00A093EC009F93, 0xEC00A293EC00A193,
-    0xEC00A493EC00A393, 0xEC00A693EC00A593, 0xEC00A893EC00A793,
-    0xEC00AA93EC00A993, 0xEC00AC93EC00AB93, 0xEC00AE93EC00AD93,
-    0xEC00B093EC00AF93, 0xEC00B293EC00B193, 0xEC00B493EC00B393,
-    0xEC00B693EC00B593, 0xEC00B893EC00B793, 0xEC00BA93EC00B993,
-    0xEC00BC93EC00BB93, 0xEC00BE93EC00BD93, 0xEC008094EC00BF93,
-    0xEC008294EC008194, 0xEC008494EC008394, 0xEC008694EC008594,
-    0xEC008894EC008794, 0xEC008A94EC008994, 0xEC008C94EC008B94,
-    0xEC008E94EC008D94, 0xEC009094EC008F94, 0xEC009294EC009194,
-    0xEC009494EC009394, 0xEC009694EC009594, 0xEC009894EC009794,
-    0xEC009A94EC009994, 0xEC009C94EC009B94, 0xEC009E94EC009D94,
-    0xEC00A094EC009F94, 0xEC00A294EC00A194, 0xEC00A494EC00A394,
-    0xEC00A694EC00A594, 0xEC00A894EC00A794, 0xEC00AA94EC00A994,
-    0xEC00AC94EC00AB94, 0xEC00AE94EC00AD94, 0xEC00B094EC00AF94,
-    0xEC00B294EC00B194, 0xEC00B494EC00B394, 0xEC00B694EC00B594,
-    0xEC00B894EC00B794, 0xEC00BA94EC00B994, 0xEC00BC94EC00BB94,
-    0xEC00BE94EC00BD94, 0xEC008095EC00BF94, 0xEC008295EC008195,
-    0xEC008495EC008395, 0xEC008695EC008595, 0xEC008895EC008795,
-    0xEC008A95EC008995, 0xEC008C95EC008B95, 0xEC008E95EC008D95,
-    0xEC009095EC008F95, 0xEC009295EC009195, 0xEC009495EC009395,
-    0xEC009695EC009595, 0xEC009895EC009795, 0xEC009A95EC009995,
-    0xEC009C95EC009B95, 0xEC009E95EC009D95, 0xEC00A095EC009F95,
-    0xEC00A295EC00A195, 0xEC00A495EC00A395, 0xEC00A695EC00A595,
-    0xEC00A895EC00A795, 0xEC00AA95EC00A995, 0xEC00AC95EC00AB95,
-    0xEC00AE95EC00AD95, 0xEC00B095EC00AF95, 0xEC00B295EC00B195,
-    0xEC00B495EC00B395, 0xEC00B695EC00B595, 0xEC00B895EC00B795,
-    0xEC00BA95EC00B995, 0xEC00BC95EC00BB95, 0xEC00BE95EC00BD95,
-    0xEC008096EC00BF95, 0xEC008296EC008196, 0xEC008496EC008396,
-    0xEC008696EC008596, 0xEC008896EC008796, 0xEC008A96EC008996,
-    0xEC008C96EC008B96, 0xEC008E96EC008D96, 0xEC009096EC008F96,
-    0xEC009296EC009196, 0xEC009496EC009396, 0xEC009696EC009596,
-    0xEC009896EC009796, 0xEC009A96EC009996, 0xEC009C96EC009B96,
-    0xEC009E96EC009D96, 0xEC00A096EC009F96, 0xEC00A296EC00A196,
-    0xEC00A496EC00A396, 0xEC00A696EC00A596, 0xEC00A896EC00A796,
-    0xEC00AA96EC00A996, 0xEC00AC96EC00AB96, 0xEC00AE96EC00AD96,
-    0xEC00B096EC00AF96, 0xEC00B296EC00B196, 0xEC00B496EC00B396,
-    0xEC00B696EC00B596, 0xEC00B896EC00B796, 0xEC00BA96EC00B996,
-    0xEC00BC96EC00BB96, 0xEC00BE96EC00BD96, 0xEC008097EC00BF96,
-    0xEC008297EC008197, 0xEC008497EC008397, 0xEC008697EC008597,
-    0xEC008897EC008797, 0xEC008A97EC008997, 0xEC008C97EC008B97,
-    0xEC008E97EC008D97, 0xEC009097EC008F97, 0xEC009297EC009197,
-    0xEC009497EC009397, 0xEC009697EC009597, 0xEC009897EC009797,
-    0xEC009A97EC009997, 0xEC009C97EC009B97, 0xEC009E97EC009D97,
-    0xEC00A097EC009F97, 0xEC00A297EC00A197, 0xEC00A497EC00A397,
-    0xEC00A697EC00A597, 0xEC00A897EC00A797, 0xEC00AA97EC00A997,
-    0xEC00AC97EC00AB97, 0xEC00AE97EC00AD97, 0xEC00B097EC00AF97,
-    0xEC00B297EC00B197, 0xEC00B497EC00B397, 0xEC00B697EC00B597,
-    0xEC00B897EC00B797, 0xEC00BA97EC00B997, 0xEC00BC97EC00BB97,
-    0xEC00BE97EC00BD97, 0xEC008098EC00BF97, 0xEC008298EC008198,
-    0xEC008498EC008398, 0xEC008698EC008598, 0xEC008898EC008798,
-    0xEC008A98EC008998, 0xEC008C98EC008B98, 0xEC008E98EC008D98,
-    0xEC009098EC008F98, 0xEC009298EC009198, 0xEC009498EC009398,
-    0xEC009698EC009598, 0xEC009898EC009798, 0xEC009A98EC009998,
-    0xEC009C98EC009B98, 0xEC009E98EC009D98, 0xEC00A098EC009F98,
-    0xEC00A298EC00A198, 0xEC00A498EC00A398, 0xEC00A698EC00A598,
-    0xEC00A898EC00A798, 0xEC00AA98EC00A998, 0xEC00AC98EC00AB98,
-    0xEC00AE98EC00AD98, 0xEC00B098EC00AF98, 0xEC00B298EC00B198,
-    0xEC00B498EC00B398, 0xEC00B698EC00B598, 0xEC00B898EC00B798,
-    0xEC00BA98EC00B998, 0xEC00BC98EC00BB98, 0xEC00BE98EC00BD98,
-    0xEC008099EC00BF98, 0xEC008299EC008199, 0xEC008499EC008399,
-    0xEC008699EC008599, 0xEC008899EC008799, 0xEC008A99EC008999,
-    0xEC008C99EC008B99, 0xEC008E99EC008D99, 0xEC009099EC008F99,
-    0xEC009299EC009199, 0xEC009499EC009399, 0xEC009699EC009599,
-    0xEC009899EC009799, 0xEC009A99EC009999, 0xEC009C99EC009B99,
-    0xEC009E99EC009D99, 0xEC00A099EC009F99, 0xEC00A299EC00A199,
-    0xEC00A499EC00A399, 0xEC00A699EC00A599, 0xEC00A899EC00A799,
-    0xEC00AA99EC00A999, 0xEC00AC99EC00AB99, 0xEC00AE99EC00AD99,
-    0xEC00B099EC00AF99, 0xEC00B299EC00B199, 0xEC00B499EC00B399,
-    0xEC00B699EC00B599, 0xEC00B899EC00B799, 0xEC00BA99EC00B999,
-    0xEC00BC99EC00BB99, 0xEC00BE99EC00BD99, 0xEC00809AEC00BF99,
-    0xEC00829AEC00819A, 0xEC00849AEC00839A, 0xEC00869AEC00859A,
-    0xEC00889AEC00879A, 0xEC008A9AEC00899A, 0xEC008C9AEC008B9A,
-    0xEC008E9AEC008D9A, 0xEC00909AEC008F9A, 0xEC00929AEC00919A,
-    0xEC00949AEC00939A, 0xEC00969AEC00959A, 0xEC00989AEC00979A,
-    0xEC009A9AEC00999A, 0xEC009C9AEC009B9A, 0xEC009E9AEC009D9A,
-    0xEC00A09AEC009F9A, 0xEC00A29AEC00A19A, 0xEC00A49AEC00A39A,
-    0xEC00A69AEC00A59A, 0xEC00A89AEC00A79A, 0xEC00AA9AEC00A99A,
-    0xEC00AC9AEC00AB9A, 0xEC00AE9AEC00AD9A, 0xEC00B09AEC00AF9A,
-    0xEC00B29AEC00B19A, 0xEC00B49AEC00B39A, 0xEC00B69AEC00B59A,
-    0xEC00B89AEC00B79A, 0xEC00BA9AEC00B99A, 0xEC00BC9AEC00BB9A,
-    0xEC00BE9AEC00BD9A, 0xEC00809BEC00BF9A, 0xEC00829BEC00819B,
-    0xEC00849BEC00839B, 0xEC00869BEC00859B, 0xEC00889BEC00879B,
-    0xEC008A9BEC00899B, 0xEC008C9BEC008B9B, 0xEC008E9BEC008D9B,
-    0xEC00909BEC008F9B, 0xEC00929BEC00919B, 0xEC00949BEC00939B,
-    0xEC00969BEC00959B, 0xEC00989BEC00979B, 0xEC009A9BEC00999B,
-    0xEC009C9BEC009B9B, 0xEC009E9BEC009D9B, 0xEC00A09BEC009F9B,
-    0xEC00A29BEC00A19B, 0xEC00A49BEC00A39B, 0xEC00A69BEC00A59B,
-    0xEC00A89BEC00A79B, 0xEC00AA9BEC00A99B, 0xEC00AC9BEC00AB9B,
-    0xEC00AE9BEC00AD9B, 0xEC00B09BEC00AF9B, 0xEC00B29BEC00B19B,
-    0xEC00B49BEC00B39B, 0xEC00B69BEC00B59B, 0xEC00B89BEC00B79B,
-    0xEC00BA9BEC00B99B, 0xEC00BC9BEC00BB9B, 0xEC00BE9BEC00BD9B,
-    0xEC00809CEC00BF9B, 0xEC00829CEC00819C, 0xEC00849CEC00839C,
-    0xEC00869CEC00859C, 0xEC00889CEC00879C, 0xEC008A9CEC00899C,
-    0xEC008C9CEC008B9C, 0xEC008E9CEC008D9C, 0xEC00909CEC008F9C,
-    0xEC00929CEC00919C, 0xEC00949CEC00939C, 0xEC00969CEC00959C,
-    0xEC00989CEC00979C, 0xEC009A9CEC00999C, 0xEC009C9CEC009B9C,
-    0xEC009E9CEC009D9C, 0xEC00A09CEC009F9C, 0xEC00A29CEC00A19C,
-    0xEC00A49CEC00A39C, 0xEC00A69CEC00A59C, 0xEC00A89CEC00A79C,
-    0xEC00AA9CEC00A99C, 0xEC00AC9CEC00AB9C, 0xEC00AE9CEC00AD9C,
-    0xEC00B09CEC00AF9C, 0xEC00B29CEC00B19C, 0xEC00B49CEC00B39C,
-    0xEC00B69CEC00B59C, 0xEC00B89CEC00B79C, 0xEC00BA9CEC00B99C,
-    0xEC00BC9CEC00BB9C, 0xEC00BE9CEC00BD9C, 0xEC00809DEC00BF9C,
-    0xEC00829DEC00819D, 0xEC00849DEC00839D, 0xEC00869DEC00859D,
-    0xEC00889DEC00879D, 0xEC008A9DEC00899D, 0xEC008C9DEC008B9D,
-    0xEC008E9DEC008D9D, 0xEC00909DEC008F9D, 0xEC00929DEC00919D,
-    0xEC00949DEC00939D, 0xEC00969DEC00959D, 0xEC00989DEC00979D,
-    0xEC009A9DEC00999D, 0xEC009C9DEC009B9D, 0xEC009E9DEC009D9D,
-    0xEC00A09DEC009F9D, 0xEC00A29DEC00A19D, 0xEC00A49DEC00A39D,
-    0xEC00A69DEC00A59D, 0xEC00A89DEC00A79D, 0xEC00AA9DEC00A99D,
-    0xEC00AC9DEC00AB9D, 0xEC00AE9DEC00AD9D, 0xEC00B09DEC00AF9D,
-    0xEC00B29DEC00B19D, 0xEC00B49DEC00B39D, 0xEC00B69DEC00B59D,
-    0xEC00B89DEC00B79D, 0xEC00BA9DEC00B99D, 0xEC00BC9DEC00BB9D,
-    0xEC00BE9DEC00BD9D, 0xEC00809EEC00BF9D, 0xEC00829EEC00819E,
-    0xEC00849EEC00839E, 0xEC00869EEC00859E, 0xEC00889EEC00879E,
-    0xEC008A9EEC00899E, 0xEC008C9EEC008B9E, 0xEC008E9EEC008D9E,
-    0xEC00909EEC008F9E, 0xEC00929EEC00919E, 0xEC00949EEC00939E,
-    0xEC00969EEC00959E, 0xEC00989EEC00979E, 0xEC009A9EEC00999E,
-    0xEC009C9EEC009B9E, 0xEC009E9EEC009D9E, 0xEC00A09EEC009F9E,
-    0xEC00A29EEC00A19E, 0xEC00A49EEC00A39E, 0xEC00A69EEC00A59E,
-    0xEC00A89EEC00A79E, 0xEC00AA9EEC00A99E, 0xEC00AC9EEC00AB9E,
-    0xEC00AE9EEC00AD9E, 0xEC00B09EEC00AF9E, 0xEC00B29EEC00B19E,
-    0xEC00B49EEC00B39E, 0xEC00B69EEC00B59E, 0xEC00B89EEC00B79E,
-    0xEC00BA9EEC00B99E, 0xEC00BC9EEC00BB9E, 0xEC00BE9EEC00BD9E,
-    0xEC00809FEC00BF9E, 0xEC00829FEC00819F, 0xEC00849FEC00839F,
-    0xEC00869FEC00859F, 0xEC00889FEC00879F, 0xEC008A9FEC00899F,
-    0xEC008C9FEC008B9F, 0xEC008E9FEC008D9F, 0xEC00909FEC008F9F,
-    0xEC00929FEC00919F, 0xEC00949FEC00939F, 0xEC00969FEC00959F,
-    0xEC00989FEC00979F, 0xEC009A9FEC00999F, 0xEC009C9FEC009B9F,
-    0xEC009E9FEC009D9F, 0xEC00A09FEC009F9F, 0xEC00A29FEC00A19F,
-    0xEC00A49FEC00A39F, 0xEC00A69FEC00A59F, 0xEC00A89FEC00A79F,
-    0xEC00AA9FEC00A99F, 0xEC00AC9FEC00AB9F, 0xEC00AE9FEC00AD9F,
-    0xEC00B09FEC00AF9F, 0xEC00B29FEC00B19F, 0xEC00B49FEC00B39F,
-    0xEC00B69FEC00B59F, 0xEC00B89FEC00B79F, 0xEC00BA9FEC00B99F,
-    0xEC00BC9FEC00BB9F, 0xEC00BE9FEC00BD9F, 0xEC0080A0EC00BF9F,
-    0xEC0082A0EC0081A0, 0xEC0084A0EC0083A0, 0xEC0086A0EC0085A0,
-    0xEC0088A0EC0087A0, 0xEC008AA0EC0089A0, 0xEC008CA0EC008BA0,
-    0xEC008EA0EC008DA0, 0xEC0090A0EC008FA0, 0xEC0092A0EC0091A0,
-    0xEC0094A0EC0093A0, 0xEC0096A0EC0095A0, 0xEC0098A0EC0097A0,
-    0xEC009AA0EC0099A0, 0xEC009CA0EC009BA0, 0xEC009EA0EC009DA0,
-    0xEC00A0A0EC009FA0, 0xEC00A2A0EC00A1A0, 0xEC00A4A0EC00A3A0,
-    0xEC00A6A0EC00A5A0, 0xEC00A8A0EC00A7A0, 0xEC00AAA0EC00A9A0,
-    0xEC00ACA0EC00ABA0, 0xEC00AEA0EC00ADA0, 0xEC00B0A0EC00AFA0,
-    0xEC00B2A0EC00B1A0, 0xEC00B4A0EC00B3A0, 0xEC00B6A0EC00B5A0,
-    0xEC00B8A0EC00B7A0, 0xEC00BAA0EC00B9A0, 0xEC00BCA0EC00BBA0,
-    0xEC00BEA0EC00BDA0, 0xEC0080A1EC00BFA0, 0xEC0082A1EC0081A1,
-    0xEC0084A1EC0083A1, 0xEC0086A1EC0085A1, 0xEC0088A1EC0087A1,
-    0xEC008AA1EC0089A1, 0xEC008CA1EC008BA1, 0xEC008EA1EC008DA1,
-    0xEC0090A1EC008FA1, 0xEC0092A1EC0091A1, 0xEC0094A1EC0093A1,
-    0xEC0096A1EC0095A1, 0xEC0098A1EC0097A1, 0xEC009AA1EC0099A1,
-    0xEC009CA1EC009BA1, 0xEC009EA1EC009DA1, 0xEC00A0A1EC009FA1,
-    0xEC00A2A1EC00A1A1, 0xEC00A4A1EC00A3A1, 0xEC00A6A1EC00A5A1,
-    0xEC00A8A1EC00A7A1, 0xEC00AAA1EC00A9A1, 0xEC00ACA1EC00ABA1,
-    0xEC00AEA1EC00ADA1, 0xEC00B0A1EC00AFA1, 0xEC00B2A1EC00B1A1,
-    0xEC00B4A1EC00B3A1, 0xEC00B6A1EC00B5A1, 0xEC00B8A1EC00B7A1,
-    0xEC00BAA1EC00B9A1, 0xEC00BCA1EC00BBA1, 0xEC00BEA1EC00BDA1,
-    0xEC0080A2EC00BFA1, 0xEC0082A2EC0081A2, 0xEC0084A2EC0083A2,
-    0xEC0086A2EC0085A2, 0xEC0088A2EC0087A2, 0xEC008AA2EC0089A2,
-    0xEC008CA2EC008BA2, 0xEC008EA2EC008DA2, 0xEC0090A2EC008FA2,
-    0xEC0092A2EC0091A2, 0xEC0094A2EC0093A2, 0xEC0096A2EC0095A2,
-    0xEC0098A2EC0097A2, 0xEC009AA2EC0099A2, 0xEC009CA2EC009BA2,
-    0xEC009EA2EC009DA2, 0xEC00A0A2EC009FA2, 0xEC00A2A2EC00A1A2,
-    0xEC00A4A2EC00A3A2, 0xEC00A6A2EC00A5A2, 0xEC00A8A2EC00A7A2,
-    0xEC00AAA2EC00A9A2, 0xEC00ACA2EC00ABA2, 0xEC00AEA2EC00ADA2,
-    0xEC00B0A2EC00AFA2, 0xEC00B2A2EC00B1A2, 0xEC00B4A2EC00B3A2,
-    0xEC00B6A2EC00B5A2, 0xEC00B8A2EC00B7A2, 0xEC00BAA2EC00B9A2,
-    0xEC00BCA2EC00BBA2, 0xEC00BEA2EC00BDA2, 0xEC0080A3EC00BFA2,
-    0xEC0082A3EC0081A3, 0xEC0084A3EC0083A3, 0xEC0086A3EC0085A3,
-    0xEC0088A3EC0087A3, 0xEC008AA3EC0089A3, 0xEC008CA3EC008BA3,
-    0xEC008EA3EC008DA3, 0xEC0090A3EC008FA3, 0xEC0092A3EC0091A3,
-    0xEC0094A3EC0093A3, 0xEC0096A3EC0095A3, 0xEC0098A3EC0097A3,
-    0xEC009AA3EC0099A3, 0xEC009CA3EC009BA3, 0xEC009EA3EC009DA3,
-    0xEC00A0A3EC009FA3, 0xEC00A2A3EC00A1A3, 0xEC00A4A3EC00A3A3,
-    0xEC00A6A3EC00A5A3, 0xEC00A8A3EC00A7A3, 0xEC00AAA3EC00A9A3,
-    0xEC00ACA3EC00ABA3, 0xEC00AEA3EC00ADA3, 0xEC00B0A3EC00AFA3,
-    0xEC00B2A3EC00B1A3, 0xEC00B4A3EC00B3A3, 0xEC00B6A3EC00B5A3,
-    0xEC00B8A3EC00B7A3, 0xEC00BAA3EC00B9A3, 0xEC00BCA3EC00BBA3,
-    0xA3EC00989DECBCA3, 0xA3EC00BEA3EC00BD, 0xA4EC0080A4EC00BF,
-    0xA4EC0082A4EC0081, 0xA4EC0084A4EC0083, 0xA4EC0086A4EC0085,
-    0xA4EC0088A4EC0087, 0xA4EC008AA4EC0089, 0xA4EC008CA4EC008B,
-    0xA4EC008EA4EC008D, 0xA4EC0090A4EC008F, 0xA4EC0092A4EC0091,
-    0xA4EC0094A4EC0093, 0xA4EC0096A4EC0095, 0xA4EC0098A4EC0097,
-    0xA4EC009AA4EC0099, 0xA4EC009CA4EC009B, 0xA4EC009EA4EC009D,
-    0xA4EC00A0A4EC009F, 0xA4EC00A2A4EC00A1, 0xA4EC00A4A4EC00A3,
-    0xA4EC00A6A4EC00A5, 0xA4EC00A8A4EC00A7, 0xA4EC00AAA4EC00A9,
-    0xA4EC00ACA4EC00AB, 0xA4EC00AEA4EC00AD, 0xA4EC00B0A4EC00AF,
-    0xA4EC00B2A4EC00B1, 0xA4EC00B4A4EC00B3, 0xA4EC00B6A4EC00B5,
-    0xA4EC00B8A4EC00B7, 0xA4EC00BAA4EC00B9, 0xA4EC00BCA4EC00BB,
-    0xA4EC00BEA4EC00BD, 0xA5EC0080A5EC00BF, 0xA5EC0082A5EC0081,
-    0xA5EC0084A5EC0083, 0xA5EC0086A5EC0085, 0xA5EC0088A5EC0087,
-    0xA5EC008AA5EC0089, 0xA5EC008CA5EC008B, 0xA5EC008EA5EC008D,
-    0xA5EC0090A5EC008F, 0xA5EC0092A5EC0091, 0xA5EC0094A5EC0093,
-    0xA5EC0096A5EC0095, 0xA5EC0098A5EC0097, 0xA5EC009AA5EC0099,
-    0xA5EC009CA5EC009B, 0xA5EC009EA5EC009D, 0xA5EC00A0A5EC009F,
-    0xA5EC00A2A5EC00A1, 0xA5EC00A4A5EC00A3, 0xA5EC00A6A5EC00A5,
-    0xA5EC00A8A5EC00A7, 0xA5EC00AAA5EC00A9, 0xA5EC00ACA5EC00AB,
-    0xA5EC00AEA5EC00AD, 0xA5EC00B0A5EC00AF, 0xA5EC00B2A5EC00B1,
-    0xA5EC00B4A5EC00B3, 0xA5EC00B6A5EC00B5, 0xA5EC00B8A5EC00B7,
-    0xA5EC00BAA5EC00B9, 0xA5EC00BCA5EC00BB, 0xA5EC00BEA5EC00BD,
-    0xA6EC0080A6EC00BF, 0xA6EC0082A6EC0081, 0xA6EC0084A6EC0083,
-    0xA6EC0086A6EC0085, 0xA6EC0088A6EC0087, 0xA6EC008AA6EC0089,
-    0xA6EC008CA6EC008B, 0xA6EC008EA6EC008D, 0xA6EC0090A6EC008F,
-    0xA6EC0092A6EC0091, 0xA6EC0094A6EC0093, 0xA6EC0096A6EC0095,
-    0xA6EC0098A6EC0097, 0xA6EC009AA6EC0099, 0xA6EC009CA6EC009B,
-    0xA6EC009EA6EC009D, 0xA6EC00A0A6EC009F, 0xA6EC00A2A6EC00A1,
-    0xA6EC00A4A6EC00A3, 0xA6EC00A6A6EC00A5, 0xA6EC00A8A6EC00A7,
-    0xA6EC00AAA6EC00A9, 0xA6EC00ACA6EC00AB, 0xA6EC00AEA6EC00AD,
-    0xA6EC00B0A6EC00AF, 0xA6EC00B2A6EC00B1, 0xA6EC00B4A6EC00B3,
-    0xA6EC00B6A6EC00B5, 0xA6EC00B8A6EC00B7, 0xA6EC00BAA6EC00B9,
-    0xA6EC00BCA6EC00BB, 0xA6EC00BEA6EC00BD, 0xA7EC0080A7EC00BF,
-    0xA7EC0082A7EC0081, 0xA7EC0084A7EC0083, 0xA7EC0086A7EC0085,
-    0xA7EC0088A7EC0087, 0xA7EC008AA7EC0089, 0xA7EC008CA7EC008B,
-    0xA7EC008EA7EC008D, 0xA7EC0090A7EC008F, 0xA7EC0092A7EC0091,
-    0xA7EC0094A7EC0093, 0xA7EC0096A7EC0095, 0xA7EC0098A7EC0097,
-    0xA7EC009AA7EC0099, 0xA7EC009CA7EC009B, 0xA7EC009EA7EC009D,
-    0xA7EC00A0A7EC009F, 0xA7EC00A2A7EC00A1, 0xA7EC00A4A7EC00A3,
-    0xA7EC00A6A7EC00A5, 0xA7EC00A8A7EC00A7, 0xA7EC00AAA7EC00A9,
-    0xA7EC00ACA7EC00AB, 0xA7EC00AEA7EC00AD, 0xA7EC00B0A7EC00AF,
-    0xA7EC00B2A7EC00B1, 0xA7EC00B4A7EC00B3, 0xA7EC00B6A7EC00B5,
-    0xA7EC00B8A7EC00B7, 0xA7EC00BAA7EC00B9, 0xA7EC00BCA7EC00BB,
-    0xA7EC00BEA7EC00BD, 0xA8EC0080A8EC00BF, 0xA8EC0082A8EC0081,
-    0xA8EC0084A8EC0083, 0xA8EC0086A8EC0085, 0xA8EC0088A8EC0087,
-    0xA8EC008AA8EC0089, 0xA8EC008CA8EC008B, 0xA8EC008EA8EC008D,
-    0xA8EC0090A8EC008F, 0xA8EC0092A8EC0091, 0xA8EC0094A8EC0093,
-    0xA8EC0096A8EC0095, 0xA8EC0098A8EC0097, 0xA8EC009AA8EC0099,
-    0xA8EC009CA8EC009B, 0xA8EC009EA8EC009D, 0xA8EC00A0A8EC009F,
-    0xA8EC00A2A8EC00A1, 0xA8EC00A4A8EC00A3, 0xA8EC00A6A8EC00A5,
-    0xA8EC00A8A8EC00A7, 0xA8EC00AAA8EC00A9, 0xA8EC00ACA8EC00AB,
-    0xA8EC00AEA8EC00AD, 0xA8EC00B0A8EC00AF, 0xA8EC00B2A8EC00B1,
-    0xA8EC00B4A8EC00B3, 0xA8EC00B6A8EC00B5, 0xA8EC00B8A8EC00B7,
-    0xA8EC00BAA8EC00B9, 0xA8EC00BCA8EC00BB, 0xA8EC00BEA8EC00BD,
-    0xA9EC0080A9EC00BF, 0xA9EC0082A9EC0081, 0xA9EC0084A9EC0083,
-    0xA9EC0086A9EC0085, 0xA9EC0088A9EC0087, 0xA9EC008AA9EC0089,
-    0xA9EC008CA9EC008B, 0xA9EC008EA9EC008D, 0xA9EC0090A9EC008F,
-    0xA9EC0092A9EC0091, 0xA9EC0094A9EC0093, 0xA9EC0096A9EC0095,
-    0xA9EC0098A9EC0097, 0xA9EC009AA9EC0099, 0xA9EC009CA9EC009B,
-    0xA9EC009EA9EC009D, 0xA9EC00A0A9EC009F, 0xA9EC00A2A9EC00A1,
-    0xA9EC00A4A9EC00A3, 0xA9EC00A6A9EC00A5, 0xA9EC00A8A9EC00A7,
-    0xA9EC00AAA9EC00A9, 0xA9EC00ACA9EC00AB, 0xA9EC00AEA9EC00AD,
-    0xA9EC00B0A9EC00AF, 0xA9EC00B2A9EC00B1, 0xA9EC00B4A9EC00B3,
-    0xA9EC00B6A9EC00B5, 0xA9EC00B8A9EC00B7, 0xA9EC00BAA9EC00B9,
-    0xA9EC00BCA9EC00BB, 0xA9EC00BEA9EC00BD, 0xAAEC0080AAEC00BF,
-    0xAAEC0082AAEC0081, 0xAAEC0084AAEC0083, 0xAAEC0086AAEC0085,
-    0xAAEC0088AAEC0087, 0xAAEC008AAAEC0089, 0xAAEC008CAAEC008B,
-    0xAAEC008EAAEC008D, 0xAAEC0090AAEC008F, 0xAAEC0092AAEC0091,
-    0xAAEC0094AAEC0093, 0xAAEC0096AAEC0095, 0xAAEC0098AAEC0097,
-    0xAAEC009AAAEC0099, 0xAAEC009CAAEC009B, 0xAAEC009EAAEC009D,
-    0xAAEC00A0AAEC009F, 0xAAEC00A2AAEC00A1, 0xAAEC00A4AAEC00A3,
-    0xAAEC00A6AAEC00A5, 0xAAEC00A8AAEC00A7, 0xAAEC00AAAAEC00A9,
-    0xAAEC00ACAAEC00AB, 0xAAEC00AEAAEC00AD, 0xAAEC00B0AAEC00AF,
-    0xAAEC00B2AAEC00B1, 0xAAEC00B4AAEC00B3, 0xAAEC00B6AAEC00B5,
-    0xAAEC00B8AAEC00B7, 0xAAEC00BAAAEC00B9, 0xAAEC00BCAAEC00BB,
-    0xAAEC00BEAAEC00BD, 0xABEC0080ABEC00BF, 0xABEC0082ABEC0081,
-    0xABEC0084ABEC0083, 0xABEC0086ABEC0085, 0xABEC0088ABEC0087,
-    0xABEC008AABEC0089, 0xABEC008CABEC008B, 0xABEC008EABEC008D,
-    0xABEC0090ABEC008F, 0xABEC0092ABEC0091, 0xABEC0094ABEC0093,
-    0xABEC0096ABEC0095, 0xABEC0098ABEC0097, 0xABEC009AABEC0099,
-    0xABEC009CABEC009B, 0xABEC009EABEC009D, 0xABEC00A0ABEC009F,
-    0xABEC00A2ABEC00A1, 0xABEC00A4ABEC00A3, 0xABEC00A6ABEC00A5,
-    0xABEC00A8ABEC00A7, 0xABEC00AAABEC00A9, 0xABEC00ACABEC00AB,
-    0xABEC00AEABEC00AD, 0xABEC00B0ABEC00AF, 0xABEC00B2ABEC00B1,
-    0xABEC00B4ABEC00B3, 0xABEC00B6ABEC00B5, 0xABEC00B8ABEC00B7,
-    0xABEC00BAABEC00B9, 0xABEC00BCABEC00BB, 0xABEC00BEABEC00BD,
-    0xACEC0080ACEC00BF, 0xACEC0082ACEC0081, 0xACEC0084ACEC0083,
-    0xACEC0086ACEC0085, 0xACEC0088ACEC0087, 0xACEC008AACEC0089,
-    0xACEC008CACEC008B, 0xACEC008EACEC008D, 0xACEC0090ACEC008F,
-    0xACEC0092ACEC0091, 0xACEC0094ACEC0093, 0xACEC0096ACEC0095,
-    0xACEC0098ACEC0097, 0xACEC009AACEC0099, 0xACEC009CACEC009B,
-    0xACEC009EACEC009D, 0xACEC00A0ACEC009F, 0xACEC00A2ACEC00A1,
-    0xACEC00A4ACEC00A3, 0xACEC00A6ACEC00A5, 0xACEC00A8ACEC00A7,
-    0xACEC00AAACEC00A9, 0xACEC00ACACEC00AB, 0xACEC00AEACEC00AD,
-    0xACEC00B0ACEC00AF, 0xACEC00B2ACEC00B1, 0xACEC00B4ACEC00B3,
-    0xACEC00B6ACEC00B5, 0xACEC00B8ACEC00B7, 0xACEC00BAACEC00B9,
-    0xACEC00BCACEC00BB, 0xACEC00BEACEC00BD, 0xADEC0080ADEC00BF,
-    0xADEC0082ADEC0081, 0xADEC0084ADEC0083, 0xADEC0086ADEC0085,
-    0xADEC0088ADEC0087, 0xADEC008AADEC0089, 0xADEC008CADEC008B,
-    0xADEC008EADEC008D, 0xADEC0090ADEC008F, 0xADEC0092ADEC0091,
-    0xADEC0094ADEC0093, 0xADEC0096ADEC0095, 0xADEC0098ADEC0097,
-    0xADEC009AADEC0099, 0xADEC009CADEC009B, 0xADEC009EADEC009D,
-    0xADEC00A0ADEC009F, 0xADEC00A2ADEC00A1, 0xADEC00A4ADEC00A3,
-    0xADEC00A6ADEC00A5, 0xADEC00A8ADEC00A7, 0xADEC00AAADEC00A9,
-    0xADEC00ACADEC00AB, 0xADEC00AEADEC00AD, 0xADEC00B0ADEC00AF,
-    0xADEC00B2ADEC00B1, 0xADEC00B4ADEC00B3, 0xADEC00B6ADEC00B5,
-    0xADEC00B8ADEC00B7, 0xADEC00BAADEC00B9, 0xADEC00BCADEC00BB,
-    0xADEC00BEADEC00BD, 0xAEEC0080AEEC00BF, 0xAEEC0082AEEC0081,
-    0xAEEC0084AEEC0083, 0xAEEC0086AEEC0085, 0xAEEC0088AEEC0087,
-    0xAEEC008AAEEC0089, 0xAEEC008CAEEC008B, 0xAEEC008EAEEC008D,
-    0xAEEC0090AEEC008F, 0xAEEC0092AEEC0091, 0xAEEC0094AEEC0093,
-    0xAEEC0096AEEC0095, 0xAEEC0098AEEC0097, 0xAEEC009AAEEC0099,
-    0xAEEC009CAEEC009B, 0xAEEC009EAEEC009D, 0xAEEC00A0AEEC009F,
-    0xAEEC00A2AEEC00A1, 0xAEEC00A4AEEC00A3, 0xAEEC00A6AEEC00A5,
-    0xAEEC00A8AEEC00A7, 0xAEEC00AAAEEC00A9, 0xAEEC00ACAEEC00AB,
-    0xAEEC00AEAEEC00AD, 0xAEEC00B0AEEC00AF, 0xAEEC00B2AEEC00B1,
-    0xAEEC00B4AEEC00B3, 0xAEEC00B6AEEC00B5, 0xAEEC00B8AEEC00B7,
-    0xAEEC00BAAEEC00B9, 0xAEEC00BCAEEC00BB, 0xAEEC00BEAEEC00BD,
-    0xAFEC0080AFEC00BF, 0xAFEC0082AFEC0081, 0xAFEC0084AFEC0083,
-    0xAFEC0086AFEC0085, 0xAFEC0088AFEC0087, 0xAFEC008AAFEC0089,
-    0xAFEC008CAFEC008B, 0xAFEC008EAFEC008D, 0xAFEC0090AFEC008F,
-    0xAFEC0092AFEC0091, 0xAFEC0094AFEC0093, 0xAFEC0096AFEC0095,
-    0xAFEC0098AFEC0097, 0xAFEC009AAFEC0099, 0xAFEC009CAFEC009B,
-    0xAFEC009EAFEC009D, 0xAFEC00A0AFEC009F, 0xAFEC00A2AFEC00A1,
-    0xAFEC00A4AFEC00A3, 0xAFEC00A6AFEC00A5, 0xAFEC00A8AFEC00A7,
-    0xAFEC00AAAFEC00A9, 0xAFEC00ACAFEC00AB, 0xAFEC00AEAFEC00AD,
-    0xAFEC00B0AFEC00AF, 0xAFEC00B2AFEC00B1, 0xAFEC00B4AFEC00B3,
-    0xAFEC00B6AFEC00B5, 0xAFEC00B8AFEC00B7, 0xAFEC00BAAFEC00B9,
-    0xAFEC00BCAFEC00BB, 0xAFEC00BEAFEC00BD, 0xB0EC0080B0EC00BF,
-    0xB0EC0082B0EC0081, 0xB0EC0084B0EC0083, 0xB0EC0086B0EC0085,
-    0xB0EC0088B0EC0087, 0xB0EC008AB0EC0089, 0xB0EC008CB0EC008B,
-    0xB0EC008EB0EC008D, 0xB0EC0090B0EC008F, 0xB0EC0092B0EC0091,
-    0xB0EC0094B0EC0093, 0xB0EC0096B0EC0095, 0xB0EC0098B0EC0097,
-    0xB0EC009AB0EC0099, 0xB0EC009CB0EC009B, 0xB0EC009EB0EC009D,
-    0xB0EC00A0B0EC009F, 0xB0EC00A2B0EC00A1, 0xB0EC00A4B0EC00A3,
-    0xB0EC00A6B0EC00A5, 0xB0EC00A8B0EC00A7, 0xB0EC00AAB0EC00A9,
-    0xB0EC00ACB0EC00AB, 0xB0EC00AEB0EC00AD, 0xB0EC00B0B0EC00AF,
-    0xB0EC00B2B0EC00B1, 0xB0EC00B4B0EC00B3, 0xB0EC00B6B0EC00B5,
-    0xB0EC00B8B0EC00B7, 0xB9B0EC00A0B3EAB8, 0xBBB0EC00BAB0EC00,
-    0xBDB0EC00BCB0EC00, 0xBFB0EC00BEB0EC00, 0x81B1EC0080B1EC00,
-    0x83B1EC0082B1EC00, 0x85B1EC0084B1EC00, 0x87B1EC0086B1EC00,
-    0x89B1EC0088B1EC00, 0x8BB1EC008AB1EC00, 0x8DB1EC008CB1EC00,
-    0x8FB1EC008EB1EC00, 0x91B1EC0090B1EC00, 0x93B1EC0092B1EC00,
-    0x95B1EC0094B1EC00, 0x97B1EC0096B1EC00, 0x99B1EC0098B1EC00,
-    0x9BB1EC009AB1EC00, 0x9DB1EC009CB1EC00, 0x9FB1EC009EB1EC00,
-    0xA1B1EC00A0B1EC00, 0xA3B1EC00A2B1EC00, 0xA5B1EC00A4B1EC00,
-    0xA7B1EC00A6B1EC00, 0xA9B1EC00A8B1EC00, 0xABB1EC00AAB1EC00,
-    0xADB1EC00ACB1EC00, 0xAFB1EC00AEB1EC00, 0xB1B1EC00B0B1EC00,
-    0xB3B1EC00B2B1EC00, 0xB5B1EC00B4B1EC00, 0xB7B1EC00B6B1EC00,
-    0xB9B1EC00B8B1EC00, 0xBBB1EC00BAB1EC00, 0xBDB1EC00BCB1EC00,
-    0xBFB1EC00BEB1EC00, 0x81B2EC0080B2EC00, 0x83B2EC0082B2EC00,
-    0x85B2EC0084B2EC00, 0x87B2EC0086B2EC00, 0x89B2EC0088B2EC00,
-    0x8BB2EC008AB2EC00, 0x8DB2EC008CB2EC00, 0x8FB2EC008EB2EC00,
-    0x91B2EC0090B2EC00, 0x93B2EC0092B2EC00, 0x95B2EC0094B2EC00,
-    0x97B2EC0096B2EC00, 0x99B2EC0098B2EC00, 0x9BB2EC009AB2EC00,
-    0x9DB2EC009CB2EC00, 0x9FB2EC009EB2EC00, 0xA1B2EC00A0B2EC00,
-    0xA3B2EC00A2B2EC00, 0xA5B2EC00A4B2EC00, 0xA7B2EC00A6B2EC00,
-    0xA9B2EC00A8B2EC00, 0xABB2EC00AAB2EC00, 0xADB2EC00ACB2EC00,
-    0xAFB2EC00AEB2EC00, 0xB1B2EC00B0B2EC00, 0xB3B2EC00B2B2EC00,
-    0xB5B2EC00B4B2EC00, 0xB7B2EC00B6B2EC00, 0xB9B2EC00B8B2EC00,
-    0xBBB2EC00BAB2EC00, 0xBDB2EC00BCB2EC00, 0xBFB2EC00BEB2EC00,
-    0x81B3EC0080B3EC00, 0x83B3EC0082B3EC00, 0x85B3EC0084B3EC00,
-    0x87B3EC0086B3EC00, 0x89B3EC0088B3EC00, 0x8BB3EC008AB3EC00,
-    0x8DB3EC008CB3EC00, 0x8FB3EC008EB3EC00, 0x91B3EC0090B3EC00,
-    0x93B3EC0092B3EC00, 0x95B3EC0094B3EC00, 0x97B3EC0096B3EC00,
-    0x99B3EC0098B3EC00, 0x9BB3EC009AB3EC00, 0x9DB3EC009CB3EC00,
-    0x9FB3EC009EB3EC00, 0xA1B3EC00A0B3EC00, 0xA3B3EC00A2B3EC00,
-    0xA5B3EC00A4B3EC00, 0xA7B3EC00A6B3EC00, 0xA9B3EC00A8B3EC00,
-    0xABB3EC00AAB3EC00, 0xADB3EC00ACB3EC00, 0xAFB3EC00AEB3EC00,
-    0xB1B3EC00B0B3EC00, 0xB3B3EC00B2B3EC00, 0xB5B3EC00B4B3EC00,
-    0xB7B3EC00B6B3EC00, 0xB9B3EC00B8B3EC00, 0xBBB3EC00BAB3EC00,
-    0xBDB3EC00BCB3EC00, 0xBFB3EC00BEB3EC00, 0x81B4EC0080B4EC00,
-    0x83B4EC0082B4EC00, 0x85B4EC0084B4EC00, 0x87B4EC0086B4EC00,
-    0x89B4EC0088B4EC00, 0x8BB4EC008AB4EC00, 0x8DB4EC008CB4EC00,
-    0x8FB4EC008EB4EC00, 0x91B4EC0090B4EC00, 0x93B4EC0092B4EC00,
-    0x95B4EC0094B4EC00, 0x97B4EC0096B4EC00, 0x99B4EC0098B4EC00,
-    0x9BB4EC009AB4EC00, 0x9DB4EC009CB4EC00, 0x9FB4EC009EB4EC00,
-    0xA1B4EC00A0B4EC00, 0xA3B4EC00A2B4EC00, 0xA5B4EC00A4B4EC00,
-    0xA7B4EC00A6B4EC00, 0xA9B4EC00A8B4EC00, 0xABB4EC00AAB4EC00,
-    0xADB4EC00ACB4EC00, 0xAFB4EC00AEB4EC00, 0xB1B4EC00B0B4EC00,
-    0xB3B4EC00B2B4EC00, 0xB5B4EC00B4B4EC00, 0xB7B4EC00B6B4EC00,
-    0xB9B4EC00B8B4EC00, 0xBBB4EC00BAB4EC00, 0xBDB4EC00BCB4EC00,
-    0xBFB4EC00BEB4EC00, 0x81B5EC0080B5EC00, 0x83B5EC0082B5EC00,
-    0x85B5EC0084B5EC00, 0x87B5EC0086B5EC00, 0x89B5EC0088B5EC00,
-    0x8BB5EC008AB5EC00, 0x8DB5EC008CB5EC00, 0x8FB5EC008EB5EC00,
-    0x91B5EC0090B5EC00, 0x93B5EC0092B5EC00, 0x95B5EC0094B5EC00,
-    0x97B5EC0096B5EC00, 0x99B5EC0098B5EC00, 0x9BB5EC009AB5EC00,
-    0x9DB5EC009CB5EC00, 0x9FB5EC009EB5EC00, 0xA1B5EC00A0B5EC00,
-    0xA3B5EC00A2B5EC00, 0xA5B5EC00A4B5EC00, 0xA7B5EC00A6B5EC00,
-    0xA9B5EC00A8B5EC00, 0xABB5EC00AAB5EC00, 0xADB5EC00ACB5EC00,
-    0xAFB5EC00AEB5EC00, 0xB1B5EC00B0B5EC00, 0xB3B5EC00B2B5EC00,
-    0xB5B5EC00B4B5EC00, 0xB7B5EC00B6B5EC00, 0xB9B5EC00B8B5EC00,
-    0xBBB5EC00BAB5EC00, 0xBDB5EC00BCB5EC00, 0xBFB5EC00BEB5EC00,
-    0x81B6EC0080B6EC00, 0x83B6EC0082B6EC00, 0x85B6EC0084B6EC00,
-    0x87B6EC0086B6EC00, 0x89B6EC0088B6EC00, 0x8BB6EC008AB6EC00,
-    0x8DB6EC008CB6EC00, 0x8FB6EC008EB6EC00, 0x91B6EC0090B6EC00,
-    0x93B6EC0092B6EC00, 0x95B6EC0094B6EC00, 0x97B6EC0096B6EC00,
-    0x99B6EC0098B6EC00, 0x9BB6EC009AB6EC00, 0x9DB6EC009CB6EC00,
-    0x9FB6EC009EB6EC00, 0xA1B6EC00A0B6EC00, 0xA3B6EC00A2B6EC00,
-    0xA5B6EC00A4B6EC00, 0xA7B6EC00A6B6EC00, 0xA9B6EC00A8B6EC00,
-    0xABB6EC00AAB6EC00, 0xADB6EC00ACB6EC00, 0xAFB6EC00AEB6EC00,
-    0xB1B6EC00B0B6EC00, 0xB3B6EC00B2B6EC00, 0xB5B6EC00B4B6EC00,
-    0xB7B6EC00B6B6EC00, 0xB9B6EC00B8B6EC00, 0xBBB6EC00BAB6EC00,
-    0xBDB6EC00BCB6EC00, 0xBFB6EC00BEB6EC00, 0x81B7EC0080B7EC00,
-    0x83B7EC0082B7EC00, 0x85B7EC0084B7EC00, 0x87B7EC0086B7EC00,
-    0x89B7EC0088B7EC00, 0x8BB7EC008AB7EC00, 0x8DB7EC008CB7EC00,
-    0x8FB7EC008EB7EC00, 0x91B7EC0090B7EC00, 0x93B7EC0092B7EC00,
-    0x95B7EC0094B7EC00, 0x97B7EC0096B7EC00, 0x99B7EC0098B7EC00,
-    0x9BB7EC009AB7EC00, 0x9DB7EC009CB7EC00, 0x9FB7EC009EB7EC00,
-    0xA1B7EC00A0B7EC00, 0xA3B7EC00A2B7EC00, 0xA5B7EC00A4B7EC00,
-    0xA7B7EC00A6B7EC00, 0xA9B7EC00A8B7EC00, 0xABB7EC00AAB7EC00,
-    0xADB7EC00ACB7EC00, 0xAFB7EC00AEB7EC00, 0xB1B7EC00B0B7EC00,
-    0xB3B7EC00B2B7EC00, 0xB5B7EC00B4B7EC00, 0xB7B7EC00B6B7EC00,
-    0xB9B7EC00B8B7EC00, 0xBBB7EC00BAB7EC00, 0xBDB7EC00BCB7EC00,
-    0xBFB7EC00BEB7EC00, 0x81B8EC0080B8EC00, 0x83B8EC0082B8EC00,
-    0x85B8EC0084B8EC00, 0x87B8EC0086B8EC00, 0x89B8EC0088B8EC00,
-    0x8BB8EC008AB8EC00, 0x8DB8EC008CB8EC00, 0x8FB8EC008EB8EC00,
-    0x91B8EC0090B8EC00, 0x93B8EC0092B8EC00, 0x95B8EC0094B8EC00,
-    0x97B8EC0096B8EC00, 0x99B8EC0098B8EC00, 0x9BB8EC009AB8EC00,
-    0x9DB8EC009CB8EC00, 0x9FB8EC009EB8EC00, 0xA1B8EC00A0B8EC00,
-    0xA3B8EC00A2B8EC00, 0xA5B8EC00A4B8EC00, 0xA7B8EC00A6B8EC00,
-    0xA9B8EC00A8B8EC00, 0xABB8EC00AAB8EC00, 0xADB8EC00ACB8EC00,
-    0xAFB8EC00AEB8EC00, 0xB1B8EC00B0B8EC00, 0xB3B8EC00B2B8EC00,
-    0xB5B8EC00B4B8EC00, 0xB7B8EC00B6B8EC00, 0xB9B8EC00B8B8EC00,
-    0xBBB8EC00BAB8EC00, 0xBDB8EC00BCB8EC00, 0xBFB8EC00BEB8EC00,
-    0x81B9EC0080B9EC00, 0x83B9EC0082B9EC00, 0x85B9EC0084B9EC00,
-    0x87B9EC0086B9EC00, 0x89B9EC0088B9EC00, 0x8BB9EC008AB9EC00,
-    0x8DB9EC008CB9EC00, 0x8FB9EC008EB9EC00, 0x91B9EC0090B9EC00,
-    0x93B9EC0092B9EC00, 0x95B9EC0094B9EC00, 0x97B9EC0096B9EC00,
-    0x99B9EC0098B9EC00, 0x9BB9EC009AB9EC00, 0x9DB9EC009CB9EC00,
-    0x9FB9EC009EB9EC00, 0xA1B9EC00A0B9EC00, 0xA3B9EC00A2B9EC00,
-    0xA5B9EC00A4B9EC00, 0xA7B9EC00A6B9EC00, 0xA9B9EC00A8B9EC00,
-    0xABB9EC00AAB9EC00, 0xADB9EC00ACB9EC00, 0xAFB9EC00AEB9EC00,
-    0xB1B9EC00B0B9EC00, 0xB3B9EC00B2B9EC00, 0xB5B9EC00B4B9EC00,
-    0xB7B9EC00B6B9EC00, 0xB9B9EC00B8B9EC00, 0xBBB9EC00BAB9EC00,
-    0xBDB9EC00BCB9EC00, 0xBFB9EC00BEB9EC00, 0x81BAEC0080BAEC00,
-    0x83BAEC0082BAEC00, 0x85BAEC0084BAEC00, 0x87BAEC0086BAEC00,
-    0x89BAEC0088BAEC00, 0x8BBAEC008ABAEC00, 0x8DBAEC008CBAEC00,
-    0x8FBAEC008EBAEC00, 0x91BAEC0090BAEC00, 0x93BAEC0092BAEC00,
-    0x95BAEC0094BAEC00, 0x97BAEC0096BAEC00, 0x99BAEC0098BAEC00,
-    0x9BBAEC009ABAEC00, 0x9DBAEC009CBAEC00, 0x9FBAEC009EBAEC00,
-    0xA1BAEC00A0BAEC00, 0xA3BAEC00A2BAEC00, 0xA5BAEC00A4BAEC00,
-    0xA7BAEC00A6BAEC00, 0xA9BAEC00A8BAEC00, 0xABBAEC00AABAEC00,
-    0xADBAEC00ACBAEC00, 0xAFBAEC00AEBAEC00, 0xB1BAEC00B0BAEC00,
-    0xB3BAEC00B2BAEC00, 0xB5BAEC00B4BAEC00, 0xB7BAEC00B6BAEC00,
-    0xB9BAEC00B8BAEC00, 0xBBBAEC00BABAEC00, 0xBDBAEC00BCBAEC00,
-    0xBFBAEC00BEBAEC00, 0x81BBEC0080BBEC00, 0x83BBEC0082BBEC00,
-    0x85BBEC0084BBEC00, 0x87BBEC0086BBEC00, 0x89BBEC0088BBEC00,
-    0x8BBBEC008ABBEC00, 0x8DBBEC008CBBEC00, 0x8FBBEC008EBBEC00,
-    0x91BBEC0090BBEC00, 0x93BBEC0092BBEC00, 0x95BBEC0094BBEC00,
-    0x97BBEC0096BBEC00, 0x99BBEC0098BBEC00, 0x9BBBEC009ABBEC00,
-    0x9DBBEC009CBBEC00, 0x9FBBEC009EBBEC00, 0xA1BBEC00A0BBEC00,
-    0xA3BBEC00A2BBEC00, 0xA5BBEC00A4BBEC00, 0xA7BBEC00A6BBEC00,
-    0xA9BBEC00A8BBEC00, 0xABBBEC00AABBEC00, 0xADBBEC00ACBBEC00,
-    0xAFBBEC00AEBBEC00, 0xB1BBEC00B0BBEC00, 0xB3BBEC00B2BBEC00,
-    0xB5BBEC00B4BBEC00, 0xB7BBEC00B6BBEC00, 0xB9BBEC00B8BBEC00,
-    0xBBBBEC00BABBEC00, 0xBDBBEC00BCBBEC00, 0xBFBBEC00BEBBEC00,
-    0x81BCEC0080BCEC00, 0x83BCEC0082BCEC00, 0x85BCEC0084BCEC00,
-    0x87BCEC0086BCEC00, 0x89BCEC0088BCEC00, 0x8BBCEC008ABCEC00,
-    0x8DBCEC008CBCEC00, 0x8FBCEC008EBCEC00, 0x91BCEC0090BCEC00,
-    0x93BCEC0092BCEC00, 0x95BCEC0094BCEC00, 0x97BCEC0096BCEC00,
-    0x99BCEC0098BCEC00, 0x9BBCEC009ABCEC00, 0x9DBCEC009CBCEC00,
-    0x9FBCEC009EBCEC00, 0xA1BCEC00A0BCEC00, 0xA3BCEC00A2BCEC00,
-    0xA5BCEC00A4BCEC00, 0xA7BCEC00A6BCEC00, 0xA9BCEC00A8BCEC00,
-    0xABBCEC00AABCEC00, 0xADBCEC00ACBCEC00, 0xAFBCEC00AEBCEC00,
-    0xB1BCEC00B0BCEC00, 0xB3BCEC00B2BCEC00, 0xB5BCEC00B4BCEC00,
-    0xB7BCEC00B6BCEC00, 0xB9BCEC00B8BCEC00, 0xBBBCEC00BABCEC00,
-    0xBDBCEC00BCBCEC00, 0xBFBCEC00BEBCEC00, 0x81BDEC0080BDEC00,
-    0x83BDEC0082BDEC00, 0x85BDEC0084BDEC00, 0x87BDEC0086BDEC00,
-    0x89BDEC0088BDEC00, 0x8BBDEC008ABDEC00, 0x8DBDEC008CBDEC00,
-    0x8FBDEC008EBDEC00, 0x91BDEC0090BDEC00, 0x93BDEC0092BDEC00,
-    0x95BDEC0094BDEC00, 0x97BDEC0096BDEC00, 0x99BDEC0098BDEC00,
-    0x9BBDEC009ABDEC00, 0x9DBDEC009CBDEC00, 0x9FBDEC009EBDEC00,
-    0xA1BDEC00A0BDEC00, 0xA3BDEC00A2BDEC00, 0xA5BDEC00A4BDEC00,
-    0xA7BDEC00A6BDEC00, 0xA9BDEC00A8BDEC00, 0xABBDEC00AABDEC00,
-    0xADBDEC00ACBDEC00, 0xAFBDEC00AEBDEC00, 0xB1BDEC00B0BDEC00,
-    0xB3BDEC00B2BDEC00, 0xB5BDEC00B4BDEC00, 0xB7BDEC00B6BDEC00,
-    0xB9BDEC00B8BDEC00, 0xBBBDEC00BABDEC00, 0xBDBDEC00BCBDEC00,
-    0xBFBDEC00BEBDEC00, 0x81BEEC0080BEEC00, 0x83BEEC0082BEEC00,
-    0x85BEEC0084BEEC00, 0x87BEEC0086BEEC00, 0x89BEEC0088BEEC00,
-    0x8BBEEC008ABEEC00, 0x8DBEEC008CBEEC00, 0x8FBEEC008EBEEC00,
-    0x91BEEC0090BEEC00, 0x93BEEC0092BEEC00, 0x95BEEC0094BEEC00,
-    0x97BEEC0096BEEC00, 0x99BEEC0098BEEC00, 0x9BBEEC009ABEEC00,
-    0x9DBEEC009CBEEC00, 0x9FBEEC009EBEEC00, 0xA1BEEC00A0BEEC00,
-    0xA3BEEC00A2BEEC00, 0xA5BEEC00A4BEEC00, 0xA7BEEC00A6BEEC00,
-    0xA9BEEC00A8BEEC00, 0xABBEEC00AABEEC00, 0xADBEEC00ACBEEC00,
-    0xAFBEEC00AEBEEC00, 0xB1BEEC00B0BEEC00, 0xB3BEEC00B2BEEC00,
-    0xB5BEEC00B4BEEC00, 0xB7BEEC00B6BEEC00, 0xB9BEEC00B8BEEC00,
-    0xBBBEEC00BABEEC00, 0xBDBEEC00BCBEEC00, 0xBFBEEC00BEBEEC00,
-    0x81BFEC0080BFEC00, 0x83BFEC0082BFEC00, 0x85BFEC0084BFEC00,
-    0x87BFEC0086BFEC00, 0x89BFEC0088BFEC00, 0x8BBFEC008ABFEC00,
-    0x8DBFEC008CBFEC00, 0x8FBFEC008EBFEC00, 0x91BFEC0090BFEC00,
-    0x93BFEC0092BFEC00, 0x95BFEC0094BFEC00, 0x97BFEC0096BFEC00,
-    0x99BFEC0098BFEC00, 0x9BBFEC009ABFEC00, 0x9DBFEC009CBFEC00,
-    0x9FBFEC009EBFEC00, 0xA1BFEC00A0BFEC00, 0xA3BFEC00A2BFEC00,
-    0xA5BFEC00A4BFEC00, 0xA7BFEC00A6BFEC00, 0xA9BFEC00A8BFEC00,
-    0xABBFEC00AABFEC00, 0xADBFEC00ACBFEC00, 0xAFBFEC00AEBFEC00,
-    0xB1BFEC00B0BFEC00, 0xB3BFEC00B2BFEC00, 0xB5BFEC00B4BFEC00,
-    0xB7BFEC00B6BFEC00, 0xB9BFEC00B8BFEC00, 0xBBBFEC00BABFEC00,
-    0xBDBFEC00BCBFEC00, 0xBFBFEC00BEBFEC00, 0x8180ED008080ED00,
-    0x8380ED008280ED00, 0x8580ED008480ED00, 0x8780ED008680ED00,
-    0x8980ED008880ED00, 0x8B80ED008A80ED00, 0x8D80ED008C80ED00,
-    0x8F80ED008E80ED00, 0x9180ED009080ED00, 0x9380ED009280ED00,
-    0x9580ED009480ED00, 0x9780ED009680ED00, 0x9980ED009880ED00,
-    0x9B80ED009A80ED00, 0x9D80ED009C80ED00, 0x9F80ED009E80ED00,
-    0xA180ED00A080ED00, 0xA380ED00A280ED00, 0xA580ED00A480ED00,
-    0xA780ED00A680ED00, 0xA980ED00A880ED00, 0xAB80ED00AA80ED00,
-    0xAD80ED00AC80ED00, 0xAF80ED00AE80ED00, 0xB180ED00B080ED00,
-    0xB380ED00B280ED00, 0xB580ED00B480ED00, 0xB780ED00B680ED00,
-    0xB980ED00B880ED00, 0xBB80ED00BA80ED00, 0xBD80ED00BC80ED00,
-    0xBF80ED00BE80ED00, 0x8181ED008081ED00, 0x8381ED008281ED00,
-    0x8581ED008481ED00, 0x8781ED008681ED00, 0x8981ED008881ED00,
-    0x8B81ED008A81ED00, 0x8D81ED008C81ED00, 0x8F81ED008E81ED00,
-    0x9181ED009081ED00, 0x9381ED009281ED00, 0x9581ED009481ED00,
-    0x9781ED009681ED00, 0x9981ED009881ED00, 0x9B81ED009A81ED00,
-    0x9D81ED009C81ED00, 0x9F81ED009E81ED00, 0xA181ED00A081ED00,
-    0xA381ED00A281ED00, 0xA581ED00A481ED00, 0xA781ED00A681ED00,
-    0xA981ED00A881ED00, 0xAB81ED00AA81ED00, 0xAD81ED00AC81ED00,
-    0xAF81ED00AE81ED00, 0xB181ED00B081ED00, 0xB381ED00B281ED00,
-    0xB581ED00B481ED00, 0xB781ED00B681ED00, 0xB981ED00B881ED00,
-    0xBB81ED00BA81ED00, 0xBD81ED00BC81ED00, 0xBF81ED00BE81ED00,
-    0x8182ED008082ED00, 0x8382ED008282ED00, 0x8582ED008482ED00,
-    0x8782ED008682ED00, 0x8982ED008882ED00, 0x8B82ED008A82ED00,
-    0x8D82ED008C82ED00, 0x8F82ED008E82ED00, 0x9182ED009082ED00,
-    0x9382ED009282ED00, 0x9582ED009482ED00, 0x9782ED009682ED00,
-    0x9982ED009882ED00, 0x9B82ED009A82ED00, 0x9D82ED009C82ED00,
-    0x9F82ED009E82ED00, 0xA182ED00A082ED00, 0xA382ED00A282ED00,
-    0xA582ED00A482ED00, 0xA782ED00A682ED00, 0xA982ED00A882ED00,
-    0xAB82ED00AA82ED00, 0xAD82ED00AC82ED00, 0xAF82ED00AE82ED00,
-    0xB182ED00B082ED00, 0xB382ED00B282ED00, 0xB582ED00B482ED00,
-    0xB782ED00B682ED00, 0xB982ED00B882ED00, 0xBB82ED00BA82ED00,
-    0xBD82ED00BC82ED00, 0xBF82ED00BE82ED00, 0x8183ED008083ED00,
-    0x8383ED008283ED00, 0x8583ED008483ED00, 0x8783ED008683ED00,
-    0x8983ED008883ED00, 0x8B83ED008A83ED00, 0x8D83ED008C83ED00,
-    0x8F83ED008E83ED00, 0x9183ED009083ED00, 0x9383ED009283ED00,
-    0x9583ED009483ED00, 0x9783ED009683ED00, 0x9983ED009883ED00,
-    0x9B83ED009A83ED00, 0x9D83ED009C83ED00, 0x9F83ED009E83ED00,
-    0xA183ED00A083ED00, 0xA383ED00A283ED00, 0xA583ED00A483ED00,
-    0xA783ED00A683ED00, 0xA983ED00A883ED00, 0xAB83ED00AA83ED00,
-    0xAD83ED00AC83ED00, 0xAF83ED00AE83ED00, 0xB183ED00B083ED00,
-    0xB383ED00B283ED00, 0xB583ED00B483ED00, 0xB783ED00B683ED00,
-    0xB983ED00B883ED00, 0xBB83ED00BA83ED00, 0xBD83ED00BC83ED00,
-    0xBF83ED00BE83ED00, 0x8184ED008084ED00, 0x8384ED008284ED00,
-    0x8584ED008484ED00, 0x8784ED008684ED00, 0x8984ED008884ED00,
-    0x8B84ED008A84ED00, 0x8D84ED008C84ED00, 0x8F84ED008E84ED00,
-    0x9184ED009084ED00, 0x9384ED009284ED00, 0x9584ED009484ED00,
-    0x9784ED009684ED00, 0x9984ED009884ED00, 0x9B84ED009A84ED00,
-    0x9D84ED009C84ED00, 0x9F84ED009E84ED00, 0xA184ED00A084ED00,
-    0xA384ED00A284ED00, 0xA584ED00A484ED00, 0xA784ED00A684ED00,
-    0xA984ED00A884ED00, 0xAB84ED00AA84ED00, 0xAD84ED00AC84ED00,
-    0xAF84ED00AE84ED00, 0xB184ED00B084ED00, 0xB384ED00B284ED00,
-    0xB584ED00B484ED00, 0xB784ED00B684ED00, 0xB984ED00B884ED00,
-    0xBB84ED00BA84ED00, 0xBD84ED00BC84ED00, 0xBF84ED00BE84ED00,
-    0x8185ED008085ED00, 0x8385ED008285ED00, 0x8585ED008485ED00,
-    0x8785ED008685ED00, 0x8985ED008885ED00, 0x8B85ED008A85ED00,
-    0x8D85ED008C85ED00, 0x8F85ED008E85ED00, 0x9185ED009085ED00,
-    0x9385ED009285ED00, 0x9585ED009485ED00, 0x9785ED009685ED00,
-    0x9985ED009885ED00, 0x9B85ED009A85ED00, 0x9D85ED009C85ED00,
-    0x9F85ED009E85ED00, 0xA185ED00A085ED00, 0xA385ED00A285ED00,
-    0xA585ED00A485ED00, 0xA785ED00A685ED00, 0xA985ED00A885ED00,
-    0xAB85ED00AA85ED00, 0xAD85ED00AC85ED00, 0xAF85ED00AE85ED00,
-    0xB185ED00B085ED00, 0xB385ED00B285ED00, 0xB585ED00B485ED00,
-    0xB785ED00B685ED00, 0xB985ED00B885ED00, 0xBB85ED00BA85ED00,
-    0xBD85ED00BC85ED00, 0xBF85ED00BE85ED00, 0x8186ED008086ED00,
-    0x8386ED008286ED00, 0x8586ED008486ED00, 0x8786ED008686ED00,
-    0x8986ED008886ED00, 0x8B86ED008A86ED00, 0x8D86ED008C86ED00,
-    0x8F86ED008E86ED00, 0x9186ED009086ED00, 0x9386ED009286ED00,
-    0x9586ED009486ED00, 0x9786ED009686ED00, 0x9986ED009886ED00,
-    0x9B86ED009A86ED00, 0x9D86ED009C86ED00, 0x9F86ED009E86ED00,
-    0xA186ED00A086ED00, 0xA386ED00A286ED00, 0xA586ED00A486ED00,
-    0xA786ED00A686ED00, 0xA986ED00A886ED00, 0xAB86ED00AA86ED00,
-    0xAD86ED00AC86ED00, 0xAF86ED00AE86ED00, 0xB186ED00B086ED00,
-    0xB386ED00B286ED00, 0xB586ED00B486ED00, 0xB786ED00B686ED00,
-    0xB986ED00B886ED00, 0xBB86ED00BA86ED00, 0xBD86ED00BC86ED00,
-    0xBF86ED00BE86ED00, 0x8187ED008087ED00, 0x8387ED008287ED00,
-    0x8587ED008487ED00, 0x8787ED008687ED00, 0x8987ED008887ED00,
-    0x8B87ED008A87ED00, 0x8D87ED008C87ED00, 0x8F87ED008E87ED00,
-    0x9187ED009087ED00, 0x9387ED009287ED00, 0x9587ED009487ED00,
-    0x9787ED009687ED00, 0x9987ED009887ED00, 0x9B87ED009A87ED00,
-    0x9D87ED009C87ED00, 0x9F87ED009E87ED00, 0xA187ED00A087ED00,
-    0xA387ED00A287ED00, 0xA587ED00A487ED00, 0xA787ED00A687ED00,
-    0xA987ED00A887ED00, 0xAB87ED00AA87ED00, 0xAD87ED00AC87ED00,
-    0xAF87ED00AE87ED00, 0xB187ED00B087ED00, 0xB387ED00B287ED00,
-    0xB587ED00B487ED00, 0xB787ED00B687ED00, 0xB987ED00B887ED00,
-    0xBB87ED00BA87ED00, 0xBD87ED00BC87ED00, 0xBF87ED00BE87ED00,
-    0x8188ED008088ED00, 0x8388ED008288ED00, 0x8588ED008488ED00,
-    0x8788ED008688ED00, 0x8988ED008888ED00, 0x8B88ED008A88ED00,
-    0x8D88ED008C88ED00, 0x8F88ED008E88ED00, 0x9188ED009088ED00,
-    0x9388ED009288ED00, 0x9588ED009488ED00, 0x9788ED009688ED00,
-    0x9988ED009888ED00, 0x9B88ED009A88ED00, 0x9D88ED009C88ED00,
-    0x9F88ED009E88ED00, 0xA188ED00A088ED00, 0xA388ED00A288ED00,
-    0xA588ED00A488ED00, 0xA788ED00A688ED00, 0xA988ED00A888ED00,
-    0xAB88ED00AA88ED00, 0xAD88ED00AC88ED00, 0xAF88ED00AE88ED00,
-    0xB188ED00B088ED00, 0xB388ED00B288ED00, 0xB588ED00B488ED00,
-    0xB788ED00B688ED00, 0xB988ED00B888ED00, 0xBB88ED00BA88ED00,
-    0xBD88ED00BC88ED00, 0xBF88ED00BE88ED00, 0x8189ED008089ED00,
-    0x8389ED008289ED00, 0x8589ED008489ED00, 0x8789ED008689ED00,
-    0x8989ED008889ED00, 0x8B89ED008A89ED00, 0x8D89ED008C89ED00,
-    0x8F89ED008E89ED00, 0x9189ED009089ED00, 0x9389ED009289ED00,
-    0x9589ED009489ED00, 0x9789ED009689ED00, 0x9989ED009889ED00,
-    0x9B89ED009A89ED00, 0x9D89ED009C89ED00, 0x9F89ED009E89ED00,
-    0xA189ED00A089ED00, 0xA389ED00A289ED00, 0xA589ED00A489ED00,
-    0xA789ED00A689ED00, 0xA989ED00A889ED00, 0xAB89ED00AA89ED00,
-    0xAD89ED00AC89ED00, 0xAF89ED00AE89ED00, 0xB189ED00B089ED00,
-    0xB389ED00B289ED00, 0xB589ED00B489ED00, 0xB789ED00B689ED00,
-    0xB989ED00B889ED00, 0xBB89ED00BA89ED00, 0xBD89ED00BC89ED00,
-    0xBF89ED00BE89ED00, 0x818AED00808AED00, 0x838AED00828AED00,
-    0x858AED00848AED00, 0x878AED00868AED00, 0x898AED00888AED00,
-    0x8B8AED008A8AED00, 0x8D8AED008C8AED00, 0x8F8AED008E8AED00,
-    0x918AED00908AED00, 0x938AED00928AED00, 0x958AED00948AED00,
-    0x978AED00968AED00, 0x998AED00988AED00, 0x9B8AED009A8AED00,
-    0x9D8AED009C8AED00, 0x9F8AED009E8AED00, 0xA18AED00A08AED00,
-    0xA38AED00A28AED00, 0xA58AED00A48AED00, 0xA78AED00A68AED00,
-    0xA98AED00A88AED00, 0xAB8AED00AA8AED00, 0xAD8AED00AC8AED00,
-    0xAF8AED00AE8AED00, 0xB18AED00B08AED00, 0xB38AED00B28AED00,
-    0xB58AED00B48AED00, 0xB78AED00B68AED00, 0xB98AED00B88AED00,
-    0xBB8AED00BA8AED00, 0xBD8AED00BC8AED00, 0xBF8AED00BE8AED00,
-    0x818BED00808BED00, 0x838BED00828BED00, 0x858BED00848BED00,
-    0x878BED00868BED00, 0x898BED00888BED00, 0x8B8BED008A8BED00,
-    0x8D8BED008C8BED00, 0x8F8BED008E8BED00, 0x918BED00908BED00,
-    0x938BED00928BED00, 0x958BED00948BED00, 0x978BED00968BED00,
-    0x998BED00988BED00, 0x9B8BED009A8BED00, 0x9D8BED009C8BED00,
-    0x9F8BED009E8BED00, 0xA18BED00A08BED00, 0xA38BED00A28BED00,
-    0xA58BED00A48BED00, 0xA78BED00A68BED00, 0xA98BED00A88BED00,
-    0xAB8BED00AA8BED00, 0xAD8BED00AC8BED00, 0xAF8BED00AE8BED00,
-    0xB18BED00B08BED00, 0xB38BED00B28BED00, 0xB58BED00B48BED00,
-    0xB78BED00B68BED00, 0xB98BED00B88BED00, 0xBB8BED00BA8BED00,
-    0xBD8BED00BC8BED00, 0xBF8BED00BE8BED00, 0x818CED00808CED00,
-    0x838CED00828CED00, 0x858CED00848CED00, 0x878CED00868CED00,
-    0x898CED00888CED00, 0x8B8CED008A8CED00, 0x8D8CED008C8CED00,
-    0x8F8CED008E8CED00, 0x918CED00908CED00, 0x938CED00928CED00,
-    0x958CED00948CED00, 0x978CED00968CED00, 0x998CED00988CED00,
-    0x9B8CED009A8CED00, 0x9D8CED009C8CED00, 0x9F8CED009E8CED00,
-    0xA18CED00A08CED00, 0xA38CED00A28CED00, 0xA58CED00A48CED00,
-    0xA78CED00A68CED00, 0xA98CED00A88CED00, 0xAB8CED00AA8CED00,
-    0xAD8CED00AC8CED00, 0xAF8CED00AE8CED00, 0xB18CED00B08CED00,
-    0xB38CED00B28CED00, 0xB58CED00B48CED00, 0xB78CED00B68CED00,
-    0xB98CED00B88CED00, 0xBB8CED00BA8CED00, 0xBD8CED00BC8CED00,
-    0xBF8CED00BE8CED00, 0x818DED00808DED00, 0x838DED00828DED00,
-    0x858DED00848DED00, 0x878DED00868DED00, 0x898DED00888DED00,
-    0x8B8DED008A8DED00, 0x8D8DED008C8DED00, 0x8F8DED008E8DED00,
-    0x918DED00908DED00, 0x938DED00928DED00, 0x958DED00948DED00,
-    0x978DED00968DED00, 0x998DED00988DED00, 0x9B8DED009A8DED00,
-    0x9D8DED009C8DED00, 0x9F8DED009E8DED00, 0xA18DED00A08DED00,
-    0xA38DED00A28DED00, 0xA58DED00A48DED00, 0xA78DED00A68DED00,
-    0xA98DED00A88DED00, 0xAB8DED00AA8DED00, 0xAD8DED00AC8DED00,
-    0xAF8DED00AE8DED00, 0xB18DED00B08DED00, 0xB38DED00B28DED00,
-    0xB58DED00B48DED00, 0xB78DED00B68DED00, 0xB98DED00B88DED00,
-    0xBB8DED00BA8DED00, 0xBD8DED00BC8DED00, 0xBF8DED00BE8DED00,
-    0x818EED00808EED00, 0x838EED00828EED00, 0x858EED00848EED00,
-    0x878EED00868EED00, 0x898EED00888EED00, 0x8B8EED008A8EED00,
-    0x8D8EED008C8EED00, 0x8F8EED008E8EED00, 0x918EED00908EED00,
-    0x938EED00928EED00, 0x958EED00948EED00, 0x978EED00968EED00,
-    0x998EED00988EED00, 0x9B8EED009A8EED00, 0x9D8EED009C8EED00,
-    0x9F8EED009E8EED00, 0xA18EED00A08EED00, 0xA38EED00A28EED00,
-    0xA58EED00A48EED00, 0xA78EED00A68EED00, 0xA98EED00A88EED00,
-    0xAB8EED00AA8EED00, 0xAD8EED00AC8EED00, 0xAF8EED00AE8EED00,
-    0xB18EED00B08EED00, 0xB38EED00B28EED00, 0xB58EED00B48EED00,
-    0xB78EED00B68EED00, 0xB98EED00B88EED00, 0xBB8EED00BA8EED00,
-    0xBD8EED00BC8EED00, 0xBF8EED00BE8EED00, 0x818FED00808FED00,
-    0x838FED00828FED00, 0x858FED00848FED00, 0x878FED00868FED00,
-    0x898FED00888FED00, 0x8B8FED008A8FED00, 0x8D8FED008C8FED00,
-    0x8F8FED008E8FED00, 0x918FED00908FED00, 0x938FED00928FED00,
-    0x958FED00948FED00, 0x978FED00968FED00, 0x998FED00988FED00,
-    0x9B8FED009A8FED00, 0x9D8FED009C8FED00, 0x9F8FED009E8FED00,
-    0xA18FED00A08FED00, 0xA38FED00A28FED00, 0xA58FED00A48FED00,
-    0xA78FED00A68FED00, 0xA98FED00A88FED00, 0xAB8FED00AA8FED00,
-    0xAD8FED00AC8FED00, 0xAF8FED00AE8FED00, 0xB18FED00B08FED00,
-    0xB38FED00B28FED00, 0xB58FED00B48FED00, 0xB78FED00B68FED00,
-    0xB98FED00B88FED00, 0xBB8FED00BA8FED00, 0xBD8FED00BC8FED00,
-    0xBF8FED00BE8FED00, 0x8190ED008090ED00, 0x8390ED008290ED00,
-    0x8590ED008490ED00, 0x8790ED008690ED00, 0x8990ED008890ED00,
-    0x8B90ED008A90ED00, 0x8D90ED008C90ED00, 0x8F90ED008E90ED00,
-    0x9190ED009090ED00, 0x9390ED009290ED00, 0x9590ED009490ED00,
-    0x9790ED009690ED00, 0x9990ED009890ED00, 0x9B90ED009A90ED00,
-    0x9D90ED009C90ED00, 0x9F90ED009E90ED00, 0xA190ED00A090ED00,
-    0xA390ED00A290ED00, 0xA590ED00A490ED00, 0xA790ED00A690ED00,
-    0xA990ED00A890ED00, 0xAB90ED00AA90ED00, 0xAD90ED00AC90ED00,
-    0xAF90ED00AE90ED00, 0xB190ED00B090ED00, 0xB390ED00B290ED00,
-    0xB590ED00B490ED00, 0xB790ED00B690ED00, 0xB990ED00B890ED00,
-    0xBB90ED00BA90ED00, 0xBD90ED00BC90ED00, 0xBF90ED00BE90ED00,
-    0x8191ED008091ED00, 0x8391ED008291ED00, 0x8591ED008491ED00,
-    0x8791ED008691ED00, 0x8991ED008891ED00, 0x8B91ED008A91ED00,
-    0x8D91ED008C91ED00, 0x8F91ED008E91ED00, 0x9191ED009091ED00,
-    0x9391ED009291ED00, 0x9591ED009491ED00, 0x9791ED009691ED00,
-    0x9991ED009891ED00, 0x9B91ED009A91ED00, 0x9D91ED009C91ED00,
-    0x9F91ED009E91ED00, 0xA191ED00A091ED00, 0xA391ED00A291ED00,
-    0xA591ED00A491ED00, 0xA791ED00A691ED00, 0xA991ED00A891ED00,
-    0xAB91ED00AA91ED00, 0xAD91ED00AC91ED00, 0xAF91ED00AE91ED00,
-    0xB191ED00B091ED00, 0xB391ED00B291ED00, 0xB591ED00B491ED00,
-    0xB791ED00B691ED00, 0xB991ED00B891ED00, 0xBB91ED00BA91ED00,
-    0xBD91ED00BC91ED00, 0xBF91ED00BE91ED00, 0x8192ED008092ED00,
-    0x8392ED008292ED00, 0x8592ED008492ED00, 0x8792ED008692ED00,
-    0x8992ED008892ED00, 0x8B92ED008A92ED00, 0x8D92ED008C92ED00,
-    0x8F92ED008E92ED00, 0x9192ED009092ED00, 0x9392ED009292ED00,
-    0x9592ED009492ED00, 0x9792ED009692ED00, 0x9992ED009892ED00,
-    0x9B92ED009A92ED00, 0x9D92ED009C92ED00, 0x9F92ED009E92ED00,
-    0xA192ED00A092ED00, 0xA392ED00A292ED00, 0xA592ED00A492ED00,
-    0xA792ED00A692ED00, 0xA992ED00A892ED00, 0xAB92ED00AA92ED00,
-    0xAD92ED00AC92ED00, 0xAF92ED00AE92ED00, 0xB192ED00B092ED00,
-    0xB392ED00B292ED00, 0xB592ED00B492ED00, 0xB792ED00B692ED00,
-    0xB992ED00B892ED00, 0xBB92ED00BA92ED00, 0xBD92ED00BC92ED00,
-    0xBF92ED00BE92ED00, 0x8193ED008093ED00, 0x8393ED008293ED00,
-    0x8593ED008493ED00, 0x8793ED008693ED00, 0x8993ED008893ED00,
-    0x8B93ED008A93ED00, 0x8D93ED008C93ED00, 0x8F93ED008E93ED00,
-    0x9193ED009093ED00, 0x9393ED009293ED00, 0x9593ED009493ED00,
-    0x9793ED009693ED00, 0x9993ED009893ED00, 0x9B93ED009A93ED00,
-    0x9D93ED009C93ED00, 0x9F93ED009E93ED00, 0xA193ED00A093ED00,
-    0xA393ED00A293ED00, 0xA593ED00A493ED00, 0xA793ED00A693ED00,
-    0xA993ED00A893ED00, 0xAB93ED00AA93ED00, 0xAD93ED00AC93ED00,
-    0xAF93ED00AE93ED00, 0xB193ED00B093ED00, 0xB393ED00B293ED00,
-    0xB593ED00B493ED00, 0xB793ED00B693ED00, 0xB993ED00B893ED00,
-    0xBB93ED00BA93ED00, 0xBD93ED00BC93ED00, 0xBF93ED00BE93ED00,
-    0x8194ED008094ED00, 0x8394ED008294ED00, 0x8594ED008494ED00,
-    0x8794ED008694ED00, 0x8994ED008894ED00, 0x8B94ED008A94ED00,
-    0x8D94ED008C94ED00, 0x8F94ED008E94ED00, 0x9194ED009094ED00,
-    0x9394ED009294ED00, 0x9594ED009494ED00, 0x9794ED009694ED00,
-    0x9994ED009894ED00, 0x9B94ED009A94ED00, 0x9D94ED009C94ED00,
-    0x9F94ED009E94ED00, 0xA194ED00A094ED00, 0xA394ED00A294ED00,
-    0xA594ED00A494ED00, 0xA794ED00A694ED00, 0xA994ED00A894ED00,
-    0xAB94ED00AA94ED00, 0xAD94ED00AC94ED00, 0xAF94ED00AE94ED00,
-    0xB194ED00B094ED00, 0xB394ED00B294ED00, 0xB594ED00B494ED00,
-    0xB794ED00B694ED00, 0xB994ED00B894ED00, 0xBB94ED00BA94ED00,
-    0xBD94ED00BC94ED00, 0xBF94ED00BE94ED00, 0x8195ED008095ED00,
-    0x8395ED008295ED00, 0x8595ED008495ED00, 0x8795ED008695ED00,
-    0x8995ED008895ED00, 0x8B95ED008A95ED00, 0x8D95ED008C95ED00,
-    0x8F95ED008E95ED00, 0x9195ED009095ED00, 0x9395ED009295ED00,
-    0x9595ED009495ED00, 0x9795ED009695ED00, 0x9995ED009895ED00,
-    0x9B95ED009A95ED00, 0x9D95ED009C95ED00, 0x9F95ED009E95ED00,
-    0xA195ED00A095ED00, 0xA395ED00A295ED00, 0xA595ED00A495ED00,
-    0xA795ED00A695ED00, 0xA995ED00A895ED00, 0xAB95ED00AA95ED00,
-    0xAD95ED00AC95ED00, 0xAF95ED00AE95ED00, 0xB195ED00B095ED00,
-    0xB395ED00B295ED00, 0xB595ED00B495ED00, 0xB795ED00B695ED00,
-    0xB995ED00B895ED00, 0xBB95ED00BA95ED00, 0xBD95ED00BC95ED00,
-    0xBF95ED00BE95ED00, 0x8196ED008096ED00, 0x8396ED008296ED00,
-    0x8596ED008496ED00, 0x8796ED008696ED00, 0x8996ED008896ED00,
-    0x8B96ED008A96ED00, 0x8D96ED008C96ED00, 0x8F96ED008E96ED00,
-    0x9196ED009096ED00, 0x9396ED009296ED00, 0x9596ED009496ED00,
-    0x9796ED009696ED00, 0x9996ED009896ED00, 0x9B96ED009A96ED00,
-    0x9D96ED009C96ED00, 0x9F96ED009E96ED00, 0xA196ED00A096ED00,
-    0xA396ED00A296ED00, 0xA596ED00A496ED00, 0xA796ED00A696ED00,
-    0xA996ED00A896ED00, 0xAB96ED00AA96ED00, 0xAD96ED00AC96ED00,
-    0xAF96ED00AE96ED00, 0xB196ED00B096ED00, 0xB396ED00B296ED00,
-    0xB596ED00B496ED00, 0xB796ED00B696ED00, 0xB996ED00B896ED00,
-    0xBB96ED00BA96ED00, 0xBD96ED00BC96ED00, 0xBF96ED00BE96ED00,
-    0x8197ED008097ED00, 0x8397ED008297ED00, 0x8597ED008497ED00,
-    0x8797ED008697ED00, 0x8997ED008897ED00, 0x8B97ED008A97ED00,
-    0x8D97ED008C97ED00, 0x8F97ED008E97ED00, 0x9197ED009097ED00,
-    0x9397ED009297ED00, 0x9597ED009497ED00, 0x9797ED009697ED00,
-    0x9997ED009897ED00, 0x9B97ED009A97ED00, 0x9D97ED009C97ED00,
-    0x9F97ED009E97ED00, 0xA197ED00A097ED00, 0xA397ED00A297ED00,
-    0xA597ED00A497ED00, 0xA797ED00A697ED00, 0xA997ED00A897ED00,
-    0xAB97ED00AA97ED00, 0xAD97ED00AC97ED00, 0xAF97ED00AE97ED00,
-    0xB197ED00B097ED00, 0xB397ED00B297ED00, 0xB597ED00B497ED00,
-    0xB797ED00B697ED00, 0xB997ED00B897ED00, 0xBB97ED00BA97ED00,
-    0xBD97ED00BC97ED00, 0xBF97ED00BE97ED00, 0x8198ED008098ED00,
-    0x8398ED008298ED00, 0x8598ED008498ED00, 0x8798ED008698ED00,
-    0x8998ED008898ED00, 0x8B98ED008A98ED00, 0x8D98ED008C98ED00,
-    0x8F98ED008E98ED00, 0x9198ED009098ED00, 0x9398ED009298ED00,
-    0x9598ED009498ED00, 0x9798ED009698ED00, 0x9998ED009898ED00,
-    0x9B98ED009A98ED00, 0x9D98ED009C98ED00, 0x9F98ED009E98ED00,
-    0xA198ED00A098ED00, 0xA398ED00A298ED00, 0xA598ED00A498ED00,
-    0xA798ED00A698ED00, 0xA998ED00A898ED00, 0xAB98ED00AA98ED00,
-    0xAD98ED00AC98ED00, 0xAF98ED00AE98ED00, 0xB198ED00B098ED00,
-    0xB398ED00B298ED00, 0xB598ED00B498ED00, 0xB798ED00B698ED00,
-    0xB998ED00B898ED00, 0xBB98ED00BA98ED00, 0xBD98ED00BC98ED00,
-    0xBF98ED00BE98ED00, 0x8199ED008099ED00, 0x8399ED008299ED00,
-    0x8599ED008499ED00, 0x8799ED008699ED00, 0x8999ED008899ED00,
-    0x8B99ED008A99ED00, 0x8D99ED008C99ED00, 0x8F99ED008E99ED00,
-    0x9199ED009099ED00, 0x9399ED009299ED00, 0x9599ED009499ED00,
-    0x9799ED009699ED00, 0x9999ED009899ED00, 0x9B99ED009A99ED00,
-    0x9D99ED009C99ED00, 0x9F99ED009E99ED00, 0xA199ED00A099ED00,
-    0xA399ED00A299ED00, 0xA599ED00A499ED00, 0xA799ED00A699ED00,
-    0xA999ED00A899ED00, 0xAB99ED00AA99ED00, 0xAD99ED00AC99ED00,
-    0xAF99ED00AE99ED00, 0xB199ED00B099ED00, 0xB399ED00B299ED00,
-    0xB599ED00B499ED00, 0xB799ED00B699ED00, 0xB999ED00B899ED00,
-    0xBB99ED00BA99ED00, 0xBD99ED00BC99ED00, 0xBF99ED00BE99ED00,
-    0x819AED00809AED00, 0x839AED00829AED00, 0x859AED00849AED00,
-    0x879AED00869AED00, 0x899AED00889AED00, 0x8B9AED008A9AED00,
-    0x8D9AED008C9AED00, 0x8F9AED008E9AED00, 0x919AED00909AED00,
-    0x939AED00929AED00, 0x959AED00949AED00, 0x979AED00969AED00,
-    0x999AED00989AED00, 0x9B9AED009A9AED00, 0x9D9AED009C9AED00,
-    0x9F9AED009E9AED00, 0xA19AED00A09AED00, 0xA39AED00A29AED00,
-    0xA59AED00A49AED00, 0xA79AED00A69AED00, 0xA99AED00A89AED00,
-    0xAB9AED00AA9AED00, 0xAD9AED00AC9AED00, 0xAF9AED00AE9AED00,
-    0xB19AED00B09AED00, 0xB39AED00B29AED00, 0xB59AED00B49AED00,
-    0xB79AED00B69AED00, 0xB99AED00B89AED00, 0xBB9AED00BA9AED00,
-    0xBD9AED00BC9AED00, 0xBF9AED00BE9AED00, 0x819BED00809BED00,
-    0x839BED00829BED00, 0x859BED00849BED00, 0x879BED00869BED00,
-    0x899BED00889BED00, 0x8B9BED008A9BED00, 0x8D9BED008C9BED00,
-    0x8F9BED008E9BED00, 0x919BED00909BED00, 0x939BED00929BED00,
-    0x959BED00949BED00, 0x979BED00969BED00, 0x999BED00989BED00,
-    0x9B9BED009A9BED00, 0x9D9BED009C9BED00, 0x9F9BED009E9BED00,
-    0xA19BED00A09BED00, 0xA39BED00A29BED00, 0xA59BED00A49BED00,
-    0xA79BED00A69BED00, 0xA99BED00A89BED00, 0xAB9BED00AA9BED00,
-    0xAD9BED00AC9BED00, 0xAF9BED00AE9BED00, 0xB19BED00B09BED00,
-    0xB39BED00B29BED00, 0xB59BED00B49BED00, 0xB79BED00B69BED00,
-    0xB99BED00B89BED00, 0xBB9BED00BA9BED00, 0xBD9BED00BC9BED00,
-    0xBF9BED00BE9BED00, 0x819CED00809CED00, 0x839CED00829CED00,
-    0x859CED00849CED00, 0x879CED00869CED00, 0x899CED00889CED00,
-    0x8B9CED008A9CED00, 0x8D9CED008C9CED00, 0x8F9CED008E9CED00,
-    0x919CED00909CED00, 0x939CED00929CED00, 0x959CED00949CED00,
-    0x979CED00969CED00, 0x999CED00989CED00, 0x9B9CED009A9CED00,
-    0x9D9CED009C9CED00, 0x9F9CED009E9CED00, 0xA19CED00A09CED00,
-    0xA39CED00A29CED00, 0xA59CED00A49CED00, 0xA79CED00A69CED00,
-    0xA99CED00A89CED00, 0xAB9CED00AA9CED00, 0xAD9CED00AC9CED00,
-    0xAF9CED00AE9CED00, 0xB19CED00B09CED00, 0xB39CED00B29CED00,
-    0xB59CED00B49CED00, 0xB79CED00B69CED00, 0xB99CED00B89CED00,
-    0xBB9CED00BA9CED00, 0xBD9CED00BC9CED00, 0xBF9CED00BE9CED00,
-    0x819DED00809DED00, 0x839DED00829DED00, 0x859DED00849DED00,
-    0x879DED00869DED00, 0x899DED00889DED00, 0x8B9DED008A9DED00,
-    0x8D9DED008C9DED00, 0x8F9DED008E9DED00, 0x919DED00909DED00,
-    0x939DED00929DED00, 0x959DED00949DED00, 0x979DED00969DED00,
-    0x999DED00989DED00, 0x9B9DED009A9DED00, 0x9D9DED009C9DED00,
-    0x9F9DED009E9DED00, 0xA19DED00A09DED00, 0xA39DED00A29DED00,
-    0xA59DED00A49DED00, 0xA79DED00A69DED00, 0xA99DED00A89DED00,
-    0xAB9DED00AA9DED00, 0xAD9DED00AC9DED00, 0xAF9DED00AE9DED00,
-    0xB19DED00B09DED00, 0xB39DED00B29DED00, 0xB59DED00B49DED00,
-    0xB79DED00B69DED00, 0xB99DED00B89DED00, 0xBB9DED00BA9DED00,
-    0xBD9DED00BC9DED00, 0xBF9DED00BE9DED00, 0x819EED00809EED00,
-    0x839EED00829EED00, 0x859EED00849EED00, 0x879EED00869EED00,
-    0x899EED00889EED00, 0x8B9EED008A9EED00, 0x8D9EED008C9EED00,
-    0x8F9EED008E9EED00, 0x919EED00909EED00, 0x939EED00929EED00,
-    0x959EED00949EED00, 0x979EED00969EED00, 0x999EED00989EED00,
-    0x9B9EED009A9EED00, 0x9D9EED009C9EED00, 0x9F9EED009E9EED00,
-    0xA19EED00A09EED00, 0xA39EED00A29EED00, 0x90F000A89090F000,
-    0x00AA9090F000A990, 0x9090F000AB9090F0, 0xF000AD9090F000AC,
-    0xAF9090F000AE9090, 0x90F000B09090F000, 0x00B29090F000B190,
-    0x9090F000B39090F0, 0xF000B59090F000B4, 0xB79090F000B69090,
-    0x90F000B89090F000, 0x00BA9090F000B990, 0x9090F000BB9090F0,
-    0xF000BD9090F000BC, 0xBF9090F000BE9090, 0x90F000809190F000,
-    0x00829190F0008191, 0x9190F000839190F0, 0xF000859190F00084,
-    0x879190F000869190, 0x90F000889190F000, 0x008A9190F0008991,
-    0x9190F0008B9190F0, 0xF0008D9190F0008C, 0x8F9190F0008E9190,
-    0x90F000989390F000, 0x009A9390F0009993, 0x9390F0009B9390F0,
-    0xF0009D9390F0009C, 0x9F9390F0009E9390, 0x90F000A09390F000,
-    0x00A29390F000A193, 0x9390F000A39390F0, 0xF000A59390F000A4,
-    0xA79390F000A69390, 0x90F000A89390F000, 0x00AA9390F000A993,
-    0x9390F000AB9390F0, 0xF000AD9390F000AC, 0xAF9390F000AE9390,
-    0x90F000B09390F000, 0x00B29390F000B193, 0x9390F000B39390F0,
-    0xF000B59390F000B4, 0xB79390F000B69390, 0x90F000B89390F000,
-    0x00BA9390F000B993, 0x9690F000BB9390F0, 0xF000989690F00097,
-    0x9A9690F000999690, 0x90F0009B9690F000, 0x009D9690F0009C96,
-    0x9690F0009E9690F0, 0xF000A09690F0009F, 0xA39690F000A19690,
-    0x90F000A49690F000, 0x00A69690F000A596, 0x9690F000A79690F0,
-    0xF000A99690F000A8, 0xAB9690F000AA9690, 0x90F000AC9690F000,
-    0x00AE9690F000AD96, 0x9690F000AF9690F0, 0xF000B19690F000B0,
-    0xB49690F000B39690, 0x90F000B59690F000, 0x00B79690F000B696,
-    0x9690F000B89690F0, 0xF000BB9690F000B9, 0x80B390F000BC9690,
-    0x90F00081B390F000, 0x0083B390F00082B3, 0xB390F00084B390F0,
-    0xF00086B390F00085, 0x88B390F00087B390, 0x90F00089B390F000,
-    0x008BB390F0008AB3, 0xB390F0008CB390F0, 0xF0008EB390F0008D,
-    0x90B390F0008FB390, 0x90F00091B390F000, 0x0093B390F00092B3,
-    0xB390F00094B390F0, 0xF00096B390F00095, 0x98B390F00097B390,
-    0x90F00099B390F000, 0x009BB390F0009AB3, 0xB390F0009CB390F0,
-    0xF0009EB390F0009D, 0xA0B390F0009FB390, 0x90F000A1B390F000,
-    0x00A3B390F000A2B3, 0xB390F000A4B390F0, 0xF000A6B390F000A5,
-    0xA8B390F000A7B390, 0x90F000A9B390F000, 0x00ABB390F000AAB3,
-    0xB390F000ACB390F0, 0xF000AEB390F000AD, 0xB0B390F000AFB390,
-    0x90F000B1B390F000, 0x009A8291F000B2B3, 0x8291F0009C8291F0,
-    0xF000AE8491F000AB, 0x8B8D91F000AF8491, 0x91F0008C8D91F000,
-    0x00BC9291F000BB92, 0x9691F000BE9291F0, 0xF000BB9691F000BA,
-    0x81A391F00080A391, 0x91F00082A391F000, 0x0084A391F00083A3,
-    0xA391F00085A391F0, 0xF00087A391F00086, 0x89A391F00088A391,
-    0x91F0008AA391F000, 0x008CA391F0008BA3, 0xA391F0008DA391F0,
-    0xF0008FA391F0008E, 0x91A391F00090A391, 0x91F00092A391F000,
-    0x0094A391F00093A3, 0xA391F00095A391F0, 0xF00097A391F00096,
-    0x99A391F00098A391, 0x91F0009AA391F000, 0x009CA391F0009BA3,
-    0xA391F0009DA391F0, 0xF0009FA391F0009E, 0xA0B996F000B8A491,
-    0x96F000A1B996F000, 0x00A3B996F000A2B9, 0xB996F000A4B996F0,
-    0xF000A6B996F000A5, 0xA8B996F000A7B996, 0x96F000A9B996F000,
-    0x00ABB996F000AAB9, 0xB996F000ACB996F0, 0xF000AEB996F000AD,
-    0xB0B996F000AFB996, 0x96F000B1B996F000, 0x00B3B996F000B2B9,
-    0xB996F000B4B996F0, 0xF000B6B996F000B5, 0xB8B996F000B7B996,
-    0x96F000B9B996F000, 0x00BBB996F000BAB9, 0xB996F000BCB996F0,
-    0xF000BEB996F000BD, 0x97859DF000BFB996, 0x859DF000A5859DF0,
-    0x9DF000A5859DF098, 0x9DF0A5859DF09885, 0xF098859DF000AE85,
-    0x00AF859DF0A5859D, 0xA5859DF098859DF0, 0x859DF000B0859DF0,
-    0x859DF0A5859DF098, 0x9DF098859DF000B1, 0xF000B2859DF0A585,
-    0x00A5859DF0B9869D, 0xA5859DF0B9869DF0, 0x869DF000AE859DF0,
-    0x859DF0A5859DF0B9, 0x9DF0BA869DF000AF, 0xF0BA869DF000A585,
-    0x00AE859DF0A5859D, 0xA5859DF0BA869DF0, 0xBC9DF000AF859DF0,
-    0xF00085BC9DF00084, 0x88BC9DF00086BC9D, 0x9DF0008ABC9DF000,
-    0x00A2A49EF0009EBC, 0xA49EF000A3A49EF0, 0xF000A5A49EF000A4,
-    0xA7A49EF000A6A49E, 0x9EF000A8A49EF000, 0x00AAA49EF000A9A4,
-    0xA49EF000ABA49EF0, 0xF000ADA49EF000AC, 0xAFA49EF000AEA49E,
-    0x9EF000B0A49EF000, 0x00B2A49EF000B1A4, 0xA49EF000B3A49EF0,
-    0xF000B5A49EF000B4, 0xB7A49EF000B6A49E, 0x9EF000B8A49EF000,
-    0x00BAA49EF000B9A4, 0xA49EF000BBA49EF0, 0xF000BDA49EF000BC,
-    0xBFA49EF000BEA49E, 0x9EF00080A59EF000, 0x0082A59EF00081A5,
-    0x84A0F00083A59EF0, 0xF0009C94A0F000A2, 0x8B95A0F000A594A0,
-    0xA0F000BA98A0F000, 0x009EA3A0F00084A0, 0xADA0F000ACA8A0F0,
-    0xF000A493A1F000A3, 0xAA9BA1F000A89AA1, 0xA1F00088A7A1F000,
-    0x008BB4A1F00098AC, 0xB7A1F000A4B7A1F0, 0xF0008386A2F000A6,
-    0xB18CA2F0009F86A2, 0xA2F000949BA2F000, 0x008AA1A2F00084A1,
-    0xAFA2F0008CACA2F0, 0xF0008A80A3F000B1, 0x9F8DA3F000B88AA3,
-    0xA3F000938EA3F000, 0x00838FA3F0009C8E, 0x91A3F000958FA3F0,
-    0xF000A39AA3F000AD, 0x8DAAA3F000A7A2A3, 0xA3F000BAABA3F000,
-    0x009EB4A3F000BCB2, 0xBDA3F00091BBA3F0, 0xF0008EBEA3F0009E,
-    0xAE8BA4F000A389A4, 0xA4F000AB8EA4F000, 0x00B59CA4F0008898,
-    0xB0A4F00094A0A4F0, 0xF00092B2A4F000B6, 0xB8BEA4F000A1BEA4,
-    0xA5F0008481A5F000, 0x00B383A5F000B283, 0x84A5F0009984A5F0,
-    0xF0008989A5F000B3, 0xA698A5F0009D90A5, 0xA5F0009A9AA5F000,
-    0x00BCA5A5F000859B, 0xAEA5F000A7AAA5F0, 0xF00080B2A5F000AB,
-    0x86BEA5F00090B3A5, 0xA6F0009A87A6F000, 0x008789A6F000A888,
-    0x8CA6F000998BA6F0, 0xF0009A93A6F000BE, 0xA896A6F000A394A6,
-    0xA6F000A79EA6F000, 0x00BCACA6F000B59E, 0xB3A6F000B6B0A6F0,
-    0xF000ABB5A6F00095, 0xB1BEA6F000ACBCA6, 0xA7F0009283A7F000,
-    0x00A799A7F0008A8F, 0xA5A7F000AEA2A7F0, 0xF000A8B2A7F000A6,
-    0xAFBCA7F00093BBA7, 0xA8F0009297A8F000, 0x00AE9CA8F000AD97,
-    0xB5A8F000BAAFA8F0, 0xF0008585A9F000B7, 0x9A88A9F0009F87A9,
-    0xA9F0008A90A9F000, 0x00B696A9F0009692, 0x83AAF000B0ACA9F0,
-    0xF0008584AAF0008E, 0x918AAAF0008E88AA, 0xAAF000928EAAF000,
-    0x0000000000008098, 0x000084000002CC00, 0x8000000180000000,
-    0x000005B80002FCCC, 0x0002E4CC8000215C, 0x8000215000003DB8,
-    0x000005B80002ECCC, 0x800008B280002160, 0x80001E0A800008B5,
-    0x8000066380001E06, 0x800008B880001E12, 0x0000158380001E0E,
-    0x00006581800008BB, 0x0000598000006580, 0x8000090F00005981,
-    0x0003583D80000912, 0x000035890003483C, 0x000085010003683E,
-    0x00008D030001F984, 0x0000950500008902, 0x00009D0700009104,
-    0x0000A90900009906, 0x0000AD0B0000A108, 0x0000B90D0000A50A,
-    0x0000BD0F0000BD0C, 0x0000C5110000B90E, 0x0000CD130000C110,
-    0x0000D5150000C912, 0x0000DD170000D114, 0x0000E5190000D916,
-    0x0000ED1B0000E118, 0x0000F51D0000E91A, 0x0000FD1F0000F11C,
-    0x0009B0610000F91E, 0x0009906380001E1E, 0x0009E06500092462,
-    0x0009CC6700099C64, 0x0009E0690009E066, 0x0009E46B0009FC68,
-    0x000E906D0009F06A, 0x000EA86F000E646C, 0x80001E1A000E946E,
-    0x000EA073000EDC70, 0x000E8075000ECC72, 0x000E8077000E6074,
-    0x000EA879000E6076, 0x80001E26000E4878, 0x80001E22000F047A,
-    0x00017D7F80000A8F, 0x0001C14180000A2F, 0x00058143800008BE,
-    0x0005B54500058942, 0x0005194700058944, 0x00051D490005CD46,
-    0x0004F54B00053148, 0x0004154D0004E94A, 0x0004554F0004354C,
-    0x0006055100047D4E, 0x000995530009FD50, 0x0009395500098D52,
-    0x000C015700097554, 0x000DE95900082556, 0x00001184000DE558,
-    0x0000F580000DF55A, 0x80000A2C0000B581, 0x0002558380001E02,
-    0x00023D8100024D82, 0x0003658700023D80, 0x0000398100027986,
-    0x80000A5000027584, 0x0003D5890004018A, 0x0000098F0003A988,
-    0x80000A5980000A11, 0x80000A5C0000098C, 0x000009910000D0CC,
-    0x0003558080001E16, 0x80001E2A00031581, 0x0000118700033583,
-    0x0000258280001CC2, 0x80001CC680000915, 0x0000218600030589,
-    0x000681A380001CCA, 0x0002B98080001CCE, 0x800009180002B981,
-    0x0000FDA500003D81, 0x8000091E8000091B, 0x80001CD280000921,
-    0x000021810000C1A8, 0x00003DB100014CCC, 0x8000092480001CD6,
-    0x0001D98980001CDA, 0x80001CE680001CE2, 0x800008CA80001CDE,
-    0x00001983800008CD, 0x00001D8180001E3E, 0x00008DA300000980,
-    0x80001E4680001E3A, 0x002904E100019182, 0x025148E3002514E0,
-    0x80001E42023D0CE2, 0x0001FD8000009586, 0x80001E360001E981,
-    0x0000C58780001CEA, 0x00001580026C00EA, 0x0273A8EF00007581,
-    0x80001CEE0000C181, 0x8000092D029244F0, 0x800008D380000930,
-    0x0000A8CD80001E32, 0x8000093600030CCC, 0x80000A6280000A5F,
-    0x00037CCC80001E4A, 0x000124CC0003FCCD, 0x80001CFA00011CCD,
-    0x000E08C380000933, 0x001190C5000FB8C2, 0x001298C7001000C4,
-    0x0015A4C9001398C6, 0x001598CB001400C8, 0x0017B0CD0015E0CA,
-    0x001C00CF80001CF2, 0x001E84D10016ACCE, 0x0020C8D3001D20D0,
-    0x002290D50021D4D2, 0x80001CF600232CD4, 0x002400D9002304D6,
-    0x0025E0DB0022BCD8, 0x0001298080001CFE, 0x0006698400012981,
-    0x80001DFE80000A4A, 0x800006888000067F, 0x800008C70006AD87,
-    0x0006858200068581, 0x0006858C800006B2, 0x800008D0800006D2,
-    0x8000092A80000939, 0x80001E2E80001D02, 0x0006FD8700075591,
-    0x0007298C0007218F, 0x800006DB80000A92, 0x000709888000093C,
-    0x0007298600070989, 0x0000598400072987, 0x000059820006AD8C,
-    0x0006A18000075583, 0x800006F50006A181, 0x80000A388000093F,
-    0x8000094200009DA7, 0x00064DB180000709, 0x80001D0680000945,
-    0x80000A8680001D0E, 0x80001D1280001D0A, 0x000761B080001D16,
-    0x00060DA78000071E, 0x800008D9800008D6, 0x000615A3000779AD,
-    0x00075DA8800008DC, 0x8000094E8000094B, 0x000639AD0000A9A7,
-    0x8000095780000951, 0x800008DF0000CDA3, 0x80001D1E0001FD81,
-    0x80000A1480001E4E, 0x00050CCD80000A65, 0x80000A68000508CC,
-    0x80001E520001958C, 0x0006D58100010982, 0x0001B58780000954,
-    0x80001D1A00016186, 0x0001DD8200010984, 0x0001818180000736,
-    0x0000098200018980, 0x8000095D0001F587, 0x80001D2280000738,
-    0x0001CD8880000A3B, 0x8000096080001D26, 0x0001D18C80001D2A,
-    0x8000076E00001D81, 0x8000096980000963, 0x80001D3280001D2E,
-    0x80001D3A80000966, 0x0000258180001D36, 0x80001D3E80000788,
-    0x000121A780001D42, 0x0000118180001D46, 0x800007DA80000A4D,
-    0x8000096F000169A3, 0x80001D4A800008E5, 0x80000975000165A7,
-    0x8000097280001D4E, 0x80001D5280001D56, 0x000145AE800007F4,
-    0x00004580800008E8, 0x0002DD8300003581, 0x0002C1810002D182,
-    0x0002E1870002C980, 0x800008EB0002F986, 0x80001E620002FD84,
-    0x00032D89800008C1, 0x00033D8F0002D588, 0x80001E6A80001E5E,
-    0x0003E9810003358C, 0x0003519180001E66, 0x80001D5A00032D81,
-    0x0003F5870003DD8C, 0x0003E58100000D81, 0x000110CD80000A98,
-    0x80001D5E000224CC, 0x00031D8C00012CCC, 0x0003A1A380001D62,
-    0x8000097E80001D66, 0x00012CCD0003F9B1, 0x0000B9800002D0CC,
-    0x00033DB10000A981, 0x00000D8480000A9E, 0x0002D4CC000381A8,
-    0x80000A95000110CD, 0x00007984000754CC, 0x00037DA30003A9A3,
-    0x000349AD0003F5B0, 0x000360CD0003A5A7, 0x00006581000C00CC,
-    0x000365A300009980, 0x000398CC80001E5A, 0x00036DA700010CCD,
-    0x000630CC80000981, 0x0007FCCC000634CD, 0x0006E0CD80000A17,
-    0x0007F4CC00016CCC, 0x0007B0CC000604CD, 0x80000A6B0007B4CD,
-    0x000750CC000170CC, 0x000128CC000748CC, 0x00012584000788CD,
-    0x0000198080000A6E, 0x80001E7600001181, 0x0002818C00010581,
-    0x000654CC80001E72, 0x0002D5800002A983, 0x000101880002A581,
-    0x0000E5810002A587, 0x80001E7E000670CC, 0x0000C4CD80001E7A,
-    0x80001E82000380CC, 0x80000A4180001E6E, 0x0000398480001D6A,
-    0x00003D820000ECCC, 0x00009CCC000245B1, 0x00024D8000068CCD,
-    0x80001D6E0001A581, 0x000130CC80000824, 0x0001E581000134CD,
-    0x0001E1830001F580, 0x00002187000239AD, 0x80000987000239A3,
-    0x0001B58980001D72, 0x00006581000215A7, 0x800008EE8000098D,
-    0x800008F180000A71, 0x8000097B000F8989, 0x80000A9B80000A74,
-    0x800009E4800008F4, 0x000F7980000FA583, 0x80001E56000F7181,
-    0x800007F980000A3E, 0x8000082600007984, 0x80001D7680001D7A,
-    0x000061A38000098A, 0x000FB18180001D7E, 0x8000099080000840,
-    0x80001D8A00000D87, 0x80001D8280000993, 0x0000F98780000999,
-    0x80001D8E000049A8, 0x0000058780001D86, 0x80000A7D80001D92,
-    0x000FDD9180000996, 0x800008540000FD81, 0x80001D96000F35A3,
-    0x0000919B8000099F, 0x80000A8080001D9A, 0x0000D9878000099C,
-    0x80001DA2000E4D86, 0x0000D9830000DD84, 0x000EC1810000D582,
-    0x000FC98F000D2980, 0x8000086180001D9E, 0x000E0D8B000FE18C,
-    0x000E1589800008FA, 0x000EE9810000E188, 0x0001698100016982,
-    0x00016187800008FD, 0x000EF58780000900, 0x80001DB200003D81,
-    0x800009A880001DB6, 0x00001588800009AB, 0x00013D8F000EC98C,
-    0x000155910001498C, 0x80000903800009A5, 0x80000A2600002984,
-    0x000039810001FD87, 0x80000A2000000D80, 0x80000A2380000A1D,
-    0x000021800000058C, 0x0001D58C00001581, 0x000105A380001E8A,
-    0x0001C1A3800009AE, 0x000105A7000169B1, 0x0001D5A7000105A6,
-    0x80000A1A800009B1, 0x80000A7A80000A77, 0x80001E8E80001E92,
-    0x800009F680001E9A, 0x000195B180001E96, 0x80001E9E000171A3,
-    0x000169A680001E86, 0x00007580000169A7, 0x80001DA600006D81,
-    0x80001DAE800009B4, 0x0001058080001DAA, 0x800008630000F981,
-    0x00034583000135AD, 0x0003058100031182, 0x80001DBA0002F980,
-    0x000108CC00033986, 0x0003BD8B00033584, 0x000355890003A18A,
-    0x0003958F00036588, 0x80001DC280001DBE, 0x80001DC600039D8C,
-    0x0003E191800009B7, 0x80001DCA80001DCE, 0x0000DD8880001DD2,
-    0x0003D99B000395A3, 0x0000798300000987, 0x000168CD80001DD6,
-    0x000124CC00016CCC, 0x00034DA300010CCD, 0x000258CC80001DDA,
-    0x80001EA2000154CD, 0x00007D81001FE194, 0x0000E18100037DA4,
-    0x8000090600009D80, 0x800009BA000349A8, 0x000351AD80001EAE,
-    0x80001DDE80000AA1, 0x0002ACCC8000090C, 0x00037D80000321B0,
-    0x80001EAA00036581, 0x80001EA600036983, 0x0000DDA3800009BD,
-    0x800009C080001DE2, 0x0000698000034989, 0x000EA0CC00005581,
-    0x000EF4CC000EB4CC, 0x000138CC000E70CD, 0x000B64CC000E5CCD,
-    0x000FC4CC800009C3, 0x000F90CC000FC0CD, 0x000EE4CC000FFCCC,
-    0x000E68CC000FACCD, 0x00010D87000F64CD, 0x0007C581000120CC,
-    0x000128CC80001DE6, 0x0001198100011182, 0x80001DEA00016580,
-    0x0001F0CD80001DEE, 0x000074CC000060CC, 0x000154CC000128CD,
-    0x000B8CCC00012D88, 0x000ED182000B0CCD, 0x0000D8CC80000A47,
-    0x80001DF680001DF2, 0x8000001B80001DFA, 0x000DB8CC8000000B,
-    0x00003D8C000B18CD, 0x000DE4CC000120CC, 0x000DB8CC80000483,
-    0x80000522000A34CD, 0x000185A3000AF8CC, 0x0003D4CC80000003,
-    0x000AD8CD00002CCD, 0x0005A4CC000CC8CC, 0x0004F0CC00026CCC,
-    0x000C80CD000CB8CD, 0x80000882000FB4CC, 0x8000088B00030CCC,
-    0x0006D588000468CC, 0x8000088D0006CD89, 0x00069D84000464CC,
-    0x8000006580000C12, 0x0006A1800006ED87, 0x0006898200069D81,
-    0x0006A58100068983, 0x800003600008D982, 0x0008F1870006A182,
-    0x0006A98780000450, 0x000804CC80000444, 0x80000554000BC0CD,
-    0x800008B80008C988, 0x800008BE0007698C, 0x0007E1B10008D58C,
-    0x0007EDB1800008C1, 0x0008D4CC80000A53, 0x0009F0CC000B28CD,
-    0x00067CCC000B64CD, 0x000954CC00091CCD, 0x0007B4CC0009D0CD,
-    0x80000A320006E8CD, 0x00086DA300065DA3, 0x000785A3800008C7,
-    0x000861A7800008D0, 0x0007D4CC800008DF, 0x0004ECCC000644CD,
-    0x0003A0CC000714CD, 0x800008E2000448CD, 0x00030CCC000841AE,
-    0x0009C580000404CD, 0x0009DD820009CD81, 0x0009F1840009D183,
-    0x0009F5860005BD87, 0x0009D988800008C4, 0x0006BD8800082189,
-    0x0008398C800008EE, 0x0005958C800008F1, 0x000705820008318F,
-    0x0007718000085D91, 0x800008F400070D81, 0x00031CCC00071987,
-    0x0006F98A0005C0CD, 0x000739880001E981, 0x0001C984800008F7,
-    0x000799A300000981, 0x0002C18C80000A56, 0x80000903000531A3,
-    0x000529A60008ADA3, 0x80000909000529A7, 0x00088DA880000912,
-    0x0002C8CC80000927, 0x0002C0CC0004E0CD, 0x8000092A0002C4CD,
-    0x0008F9B0000575AD, 0x80000948000791A3, 0x00070989000529B1,
-    0x8000087700071188, 0x800007270006418A, 0x0007298700075984,
-    0x0007598180000784, 0x00074D8300076580, 0x8000096C00074D82,
-    0x000E0D8F80000BB5, 0x8000097B80000978, 0x8000099080000984,
-    0x800009A280000999, 0x000764CD800009A5, 0x000180CC00014CCC,
-    0x80000ACB800009A8, 0x800009CC800009C9, 0x000E4D9F001FB585,
-    0x80000ACE001FB585, 0x000128CD800009CF, 0x80000AD400012CCC,
-    0x001C6D8580000AD7, 0x80000A29800009D2, 0x80000AE080000ADD,
-    0x001C9185800009D5, 0x00052DAF80000AEC, 0x80000B0780000AF5,
-    0x00156D81000535A8, 0x80000B04000EDDAA, 0x80000B19800009DB,
-    0x80000B2B80000B22, 0x800009E4000EBDA0, 0x00029DBD800009E7,
-    0x800009EA000291BC, 0x000299B90002A5BE, 0x80000B460002FDB8,
-    0x0002C5B5000B31BA, 0x0007B4CC000531B4, 0x800009ED000648CD,
-    0x000531B380000B4F, 0x001D2585000541B2, 0x0016F583800009F3,
-    0x800009F680000B58, 0x000180A680000B5E, 0x800009F980000B61,
-    0x00016CCC800009FC, 0x000748CC0007B0CD, 0x000108CC001D7985,
-    0x80000A440007C4CD, 0x001C1985800009FF, 0x001C098580000A02,
-    0x80000770800006CE, 0x0001E0B8800007E2, 0x80000A3E80000A35,
-    0x001C1185800006CB, 0x80000A05800009D8, 0x000FB18080000A83,
-    0x0003F982000FB181, 0x0003E984000FD983, 0x0002858600031185,
-    0x000D798800035987, 0x0003658A000D7989, 0x000B758C000CFD8B,
-    0x000B6D8E000B7D8D, 0x000331900003698F, 0x000A3D92000A0991,
-    0x0002FD940009D593, 0x000289960002E195, 0x00028D98800009E1,
-    0x0006E59A00071199, 0x0002799C0006F99B, 0x0002619E00055D9D,
-    0x001DED8080000A89, 0x80000B76001E0581, 0x80000A8C0015C583,
-    0x80000A8F0015F585, 0x80000A920015C587, 0x001E258A001E2189,
-    0x001E218C80000A9B, 0x0016B18E800022A2, 0x0016BD9080000AA7,
-    0x001E5D92800009DE, 0x0017B994800022A6, 0x0018159680000AB0,
-    0x00179D98800009C6, 0x001E719A80000B5B, 0x00181D9C80000B64,
-    0x0018ED9E80000AB3, 0x001E9DA080000AB6, 0x001835A280000AB9,
-    0x00183DA480000ABC, 0x001EB9A680000ABF, 0x001E85A880000709,
-    0x001E9DAA80000AFE, 0x00199DAC80000736, 0x00193DAE80000B37,
-    0x00195DB080000B3D, 0x00192DB280000B49, 0x001A95B480000013,
-    0x001A69B680000017, 0x001A25B880000031, 0x80000069001A01B9,
-    0x80000007001A1DBB, 0x80000035001BE9BD, 0x00027D850018A1BF,
-    0x80000B79001599A4, 0x80000BBF001CC5A6, 0x80000BC2001591A0,
-    0x80000BCA00158DA2, 0x80000BC50015C9AC, 0x80000C09001401AE,
-    0x80000BCD0015E9A8, 0x80000BD00015DDAA, 0x80000BB200141DB4,
-    0x80000BD300143DB6, 0x001CB9B380000075, 0x0014E9BD001CBDB2,
-    0x001C85BF8000063E, 0x0014F5B980000C8F, 0x0014EDBB80000025,
-    0x800008AC80000BE2, 0x80000BE50013E984, 0x80000BE800131D86,
-    0x800008770011FD80, 0x80000BEB001C0182, 0x80000C5F0013018C,
-    0x80000C6200136D8E, 0x80000C6500132988, 0x80001F8E00131D8A,
-    0x8000088B00128D94, 0x8000200E00128196, 0x80000BF1001C7190,
-    0x8000200A001C6592, 0x80001EC60015C99C, 0x80001FB20015B99E,
-    0x80001FBA00127198, 0x001C398100129D9A, 0x80001FAA80001FC2,
-    0x80001FB6001C3182, 0x001C4D87001C5584, 0x001C4189001C4D86,
-    0x001C458B80001FBE, 0x0003D980001C498A, 0x001C4D8F00065181,
-    0x001C3D91001C4D8E, 0x001CCD93001C3D90, 0x80001FAE80001FC6,
-    0x001CF997001CD994, 0x0006A582001CCD96, 0x00038180001CC198,
-    0x001CDD9D00060981, 0x001CC99F001CDD9C, 0x8000201600007585,
-    0x80000BF4001C3DA0, 0x80000BF7001C31A2, 0x001C69A7001C5DA4,
-    0x001C55A9001C45A6, 0x001DE0CC80000BFA, 0x80001F92001A08CD,
-    0x001DA9AF001DB1AC, 0x001DD5B1001DB1AE, 0x001DE9B380001EE6,
-    0x001DF5B5001DD1B2, 0x0002D1B780001EEE, 0x80001EEA001E8CCD,
-    0x0006C8CC001DA1B8, 0x80001EF200040CCD, 0x001D0CCC001DB9BC,
-    0x0007D581001A04CD, 0x0001DD810007D580, 0x001C45840001C980,
-    0x001C4D86001C4185, 0x001C7988001C4587, 0x001C758A001C7D89,
-    0x001C6D8C001C718B, 0x001C88CC0011018D, 0x80000C000016358F,
-    0x80001F9600144991, 0x80001FD2001AED93, 0x80001FDA001A3195,
-    0x80001FE2001A3D97, 0x80001FCA001A0D99, 0x00019180001A3D9B,
-    0x0012819E00018581, 0x0012FDA080001FD6, 0x001F31A280001FDE,
-    0x001CF5A480001FE6, 0x001711A6000E9988, 0x0016C5A8001A1CCC,
-    0x001CC1AA80001FCE, 0x001429AC001DE0CC, 0x001CF9AE8000202A,
-    0x0000458580000C06, 0x001CA5B2001CA9B1, 0x00102DB4001CA1B3,
-    0x001CADB680001F9A, 0x00165DB8001CA5B7, 0x0011C9BA8000203E,
-    0x001F7DBC8000203A, 0x001E85BE80000C59, 0x0004558180001F1E,
-    0x0004558300190492, 0x00047D8500078185, 0x0002958500047184,
-    0x0002898500046986, 0x80001EBE00045988, 0x80001F260004A98A,
-    0x80001F2E0004B98C, 0x0004B19C0004A98E, 0x0000B5800004BD9D,
-    0x00049D9800014D81, 0x0004859A00048199, 0x0000BD820004B59B,
-    0x0000B98080001ECE, 0x0005F9810000B981, 0x000161830006C580,
-    0x001CC1B300062D85, 0x00042DB1000429B2, 0x000509B7000411B0,
-    0x000425B5000451B6, 0x0005A9A4000429B4, 0x80001F2A80001F22,
-    0x001A3DA0000565B8, 0x001CE5A2001135A1, 0x80001F320018BDA3,
-    0x80001EC280001ED6, 0x80001ECA80000C0C, 0x80000C0F80001ED2,
-    0x0005218080001EDA, 0x80000BFD00059981, 0x80000C350000C583,
-    0x80000C3800067585, 0x80001F0680001EFE, 0x80001F0E000525A8,
-    0x0003A9850013B5AA, 0x80001EDE001371AC, 0x80001F9E000531AE,
-    0x80001F3E001AF9A0, 0x80001F46001AD1A2, 0x80001F42000525A4,
-    0x00003980000525A6, 0x000575BB0000CD81, 0x000515BD000579BA,
-    0x80001F4A80001EE2, 0x80000C3E000511BE, 0x80000C030013C1B0,
-    0x80001EF6001FD1B2, 0x80001F028000205E, 0x80001EFA80000C44,
-    0x80001F0A00108988, 0x80000C470010958A, 0x80001F120012BD8C,
-    0x001D64CC0012D98E, 0x80000C0900179D80, 0x80001FA20017ED82,
-    0x800020520016D184, 0x8000204E0016D586, 0x80000C5C001D5598,
-    0x80001F16001D759A, 0x80001F5A001AA99C, 0x80001F620011759E,
-    0x80001F6A001DE190, 0x80000C4D001DF992, 0x80000C50001C5194,
-    0x00071180001C5596, 0x80000C5300071D81, 0x0007398400072D83,
-    0x0004A98000072585, 0x0007E18000079581, 0x000724CC00075981,
-    0x0007E184000720CD, 0x0007D986001B70CD, 0x000008CC001B58CD,
-    0x00031D94000238CD, 0x80001FA6000558CD, 0x0002F8CC00034D93,
-    0x000710CD000C00CD, 0x00050185000714CC, 0x0002B99400024593,
-    0x0003798100037D82, 0x0003F98000028180, 0x0003118200030D81,
-    0x80000C4180001FF2, 0x000CB8CC00053185, 0x0005E980000CDCCD,
-    0x80001FFA00073581, 0x0001F0CC0000B4CD, 0x80002002000234CD,
-    0x0007F9B0000768CD, 0x000789B280001FEA, 0x000799B480001FF6,
-    0x000789B680001FFE, 0x0003F4CC80002006, 0x0007BDBA000344CD,
-    0x0003719400039D93, 0x00079DBE00003985, 0x001CD584000791BF,
-    0x0006758600067185, 0x0003218100068D87, 0x000728CC0004E980,
-    0x0006898C000774CD, 0x0006998E80001FEE, 0x0006B9880006958F,
-    0x0006998A0006A589, 0x0005E5948000206E, 0x000409960003AD95,
-    0x001B08CD00022597, 0x0005D19200068D91, 0x001B459C0005CD93,
-    0x0003919E0003959D, 0x0004CD98000F719F, 0x0003B59A0002DD99,
-    0x000211A40003A59B, 0x000279A6000FD1A5, 0x0002F1A0000275A7,
-    0x00014D85000105A1, 0x00045D800002E1A3, 0x0003918200039581,
-    0x0001B9A880002012, 0x00044DAA000FDDA9, 0x00011D85000275AB,
-    0x001A94CD000D8CB5, 0x00007D85000D70B7, 0x0003C0CC000DD8B1,
-    0x001AF4CD0001A8CD, 0x001AC8CD001B70CD, 0x001B8CCD000ECCBF,
-    0x0006FCCC000D04B9, 0x000BF588000194CD, 0x000CAD8080001F1A,
-    0x000CB580000CA581, 0x000CD982000CBD81, 0x000BF58080000C3B,
-    0x000D098000098581, 0x000BD58400098581, 0x000BD58680001F36,
-    0x000B958080001F3A, 0x000BFD850009E581, 0x000D2183000D9D83,
-    0x0009C98180000C4A, 0x80001F52000B1580, 0x80000C5600005193,
-    0x00007594000DBD93, 0x0000459480001F6E, 0x000DA98080001F76,
-    0x000DB180000DB581, 0x000DB180000DD581, 0x000D518800093D81,
-    0x000D61860000A983, 0x000D118480001F7E, 0x000D698080001F86,
-    0x000D0980000D5581, 0x000D198000090D81, 0x000D2182000D2581,
-    0x000D618080001F72, 0x0000A58000096581, 0x0000458100007981,
-    0x0000899400009980, 0x000E358500004182, 0x00000D8080001F7A,
-    0x000F51800001F581, 0x80001F820008BD81, 0x00000582000E2985,
-    0x0001ED8080001F8A, 0x000F718000001581, 0x80001F8E00089D81,
-    0x80000BE200015983, 0x000258CD000E0985, 0x000005850002BCCC,
-    0x8000201E8000201A, 0x00019D9400013993, 0x8000202200000D85,
-    0x0000498080001F96, 0x0000218200005581, 0x0000E585000D4583,
-    0x000D60CC000BF585, 0x0000A980000F08CD, 0x0001C8CD0000B581,
-    0x0008418400022CCC, 0x0008798680002026, 0x0000FD8080000BE8,
-    0x0000058500003981, 0x800020328000202E, 0x00000D8580002036,
-    0x0001E8CC80000C59, 0x80002042000154CD, 0x000D1994000DE593,
-    0x0000118080000BDF, 0x0000298000000981, 0x00001D8200003181,
-    0x000C7194000C9D93, 0x800020468000204A, 0x800020568000205A,
-    0x0003B588000150CD, 0x000038CC0000B985, 0x80000BEE000084CD,
-    0x000FF4CD000FD8CD, 0x0009FD8080002066, 0x0009DD86000B8D81,
-    0x0009DD8480001F56, 0x00031D80000FB0CD, 0x0003F9820003D981,
-    0x80001F5E000F3D83, 0x000278CC0009E185, 0x80001F66001758CD,
-    0x00010CCD000FB593, 0x000F059400039193, 0x00027D9480002062,
-    0x0000458500013CCD, 0x000168CD000128CD, 0x00094580000154CC,
-    0x00036182000B3581, 0x000240CC000F4D83, 0x000114CD0003D4CD,
-    0x0002D1940006A593, 0x0000B9808000206A, 0x0006C1880000A581,
-    0x0000FD8080002072, 0x0000D18000009D81, 0x0000FD820000CD81,
-    0x000F51808000207A, 0x000384CC000B5581, 0x000F498400036CCD,
-    0x000F398680000C68, 0x000180CD00002185, 0x000180CD000184CC,
-    0x00012580000184CC, 0x0001798200014581, 0x0003559380000C6B,
-    0x0001B8CD80000C6E, 0x000120CD000104CC, 0x80000C7100013CCC,
-    0x0001A4CD0003F194, 0x000D44CD00019CCC, 0x0001E0CC80000C74,
-    0x000058CC000194CD, 0x0001B0CD00007CCD, 0x00077D810001DCCC,
-    0x000EC18000077D80, 0x00010D82000AC581, 0x0007918000077583,
-    0x0007918200079181, 0x000DA4CD0003F983, 0x000C9CCD000CA985,
-    0x0000498080000C77, 0x80000C7D80000C7A, 0x80000C8000061588,
-    0x80000C8680000C83, 0x000F20CD00068981, 0x80000C8C80000C89,
-    0x80000C9580000C92, 0x0006B98680000BD6, 0x0006958880000BD9,
-    0x80000CF880000BDC, 0x80000CFE80000CFB, 0x0000E0CD80000D01,
-    0x80000C5C000104CC, 0x80001F4E0001D4B8, 0x80001FA60001D4BA,
-    0x8000207680000C65, 0x80000C410001F0BE, 0x0001D0B30001D4B0,
-    0x0001D0B580000D04, 0x0001D0B780000D07, 0x0002C9A90001C8B6,
-    0x003C45AB0002C5A8, 0x003C41AD0002C1AA, 0x0002EDAF0002C9AC,
-    0x000225A10002EDAE, 0x003C6DA300025DA0, 0x000201A5000225A2,
-    0x003C71A7000211A4, 0x00028199000209A6, 0x0002999B00014998,
-    0x0002B19D003C019A, 0x0002A99F00028D9C, 0x00012191003C9D9E,
-    0x000125930001FD90, 0x0001259500012592, 0x00010D9700012594,
-    0x0001DD8900011D96, 0x0001E58B0001D188, 0x0001ED8D0001E98A,
-    0x00018D8F0001E18C, 0x00070D810001958E, 0x00071D8300070D80,
-    0x0000C58500071D82, 0x0001F5870000D984, 0x0006C9AE0001F986,
-    0x0006CDAC80000D0A, 0x0006A9AA80000D0D, 0x0006ADA880000D10,
-    0x000691A680000D13, 0x000681A480000D16, 0x000695A280000D19,
-    0x0006A5A080000D1C, 0x00069DBE80000D1F, 0x80000D220006A1BF,
-    0x000681BA000699BD, 0x80000C9880000D25, 0x80000DC4000821B9,
-    0x000ED5B40006A9B7, 0x00080DB2000985B5, 0x0009B5B080000DC7,
-    0x80000C9B00015DB1, 0x80000C9E0007718F, 0x80000CA480000CA1,
-    0x80000DCD80000CA7, 0x80000CAA00044489, 0x80000DD680000DAC,
-    0x0008CD8200043C85, 0x80000DD980000CAD, 0x00068D9E0004D481,
-    0x0006F19C00062D80, 0x0006F59A80000CB0, 0x0006C59880000DDF,
-    0x0008819600004586, 0x0008219400060988, 0x0005A59200081D95,
-    0x000F7190000C4593, 0x80000CB3000E4191, 0x80000DE20003A186,
-    0x80000CB900033180, 0x80000CBC00001586, 0x80000CBF0003C984,
-    0x80000DE500031980, 0x80000CC500002588, 0x80000CC8000F80CC,
-    0x0000698800039588, 0x80000CCE80000CCB, 0x00001D8480000DF1,
-    0x0003118680000DF4, 0x0000358880000DF7, 0x0007A8CC80000CD4,
-    0x80000CD70000318B, 0x0000058880000CDA, 0x80000CE080000DFA,
-    0x80000CE3000708CC, 0x00028CCC80000CE6, 0x80000E0000000D88,
-    0x000700CD80000CEC, 0x00000588000160CC, 0x80000CF280000DEE,
-    0x001224CD80000CF5, 0x000700CC001220CC, 0x001254CC0002D8CC,
-    0x000114CC001250CD, 0x000754CC000774CD, 0x0002ECCC0002BCCC,
-    0x000750CC0002C0CD, 0x0002B0CC000754CD, 0x0002DCCD0002ACCC,
-    0x0003B4CC0002F0CC, 0x000380CC80000D28, 0x000730CC000EB8CC,
-    0x80000D2B001210CD, 0x0007B4CD00032096, 0x80000D2E00073CCD,
-    0x000788CC0007C4CD, 0x0007F0CC0007E4CD, 0x80000D3100062CCD,
-    0x0002888380000D34, 0x80000D3A80000D37, 0x00029C8780000D3D,
-    0x80000D4380000D40, 0x0003588B80000D49, 0x0003488D0000098F,
-    0x80000D4C000224CC, 0x80000D5280000D4F, 0x80000D550000D9B0,
-    0x0003D4B50000EDB2, 0x80000D58003F81B4, 0x80000D5B0000F9B6,
-    0x80000D5E0000DDB8, 0x80000D610000D1BA, 0x003E38D9000035BC,
-    0x80000D64000035BE, 0x80000D670000C1A0, 0x80000D6A0000D5A2,
-    0x80000D6D0000FDA4, 0x80000D7000009DA6, 0x80000D730000A1A8,
-    0x80000D760000A1AA, 0x80000D790000A5AC, 0x80000CC20000A9AE,
-    0x80000CDD80000CD1, 0x80000CEF80000CE9, 0x003F818080000D46,
-    0x003F918E80000D7C, 0x003F818C80000D7F, 0x003F918A80000D82,
-    0x80000D8880000D85, 0x003F999680000D8B, 0x003F999480000D8E,
-    0x003F999280000D91, 0x003F999080000D94, 0x003F999E80000D97,
-    0x003F999C80000D9A, 0x003F999A80000D9D, 0x003F999880000DA0,
-    0x0000B5A680000DA3, 0x0000A5A480000DA6, 0x0000B5A280000CB6,
-    0x000085A080000DA9, 0x0000EDAE80000DC1, 0x0000EDAC80000DAF,
-    0x00008DAA80000DB2, 0x000081A880000DB5, 0x0000ADB680000DB8,
-    0x0000ADB480000DBB, 0x00008DB280000DBE, 0x00008DB080000DCA,
-    0x000075BE80000DD0, 0x00006DBC00000988, 0x00008DBA80000DD3,
-    0x00008DB8000230CC, 0x00019D8080000DDC, 0x80000DE8003CA181,
-    0x000009880001A983, 0x80000DEB0001B985, 0x000210CC0001A987,
-    0x80000DFD00019989, 0x80000E030001E98B, 0x80000E060001F98D,
-    0x003C219080000E09, 0x003C3D9280000E0C, 0x0001859480000E0F,
-    0x003CC19680000E12, 0x0001BD9880000E15, 0x0001859A0002A099,
-    0x003CD99C80000E18, 0x003CC19E80000E1B, 0x00017DA080000E1E,
-    0x003C5DA280000E21, 0x003F99A480000E24, 0x003FB1A680000E27,
-    0x0000A9A880000E2A, 0x0000B1AA0003B4A9, 0x003F1DAC80000E2D,
-    0x003FF5AE80000E30, 0x003F95B080000E33, 0x003F95B280000E36,
-    0x003FADB480000E39, 0x0000F1B680000E3C, 0x003F79B880000E3F,
-    0x0000D9BA80000E42, 0x0000C9BC80000E45, 0x0000C5BE80000E48,
-    0x0000BDBE80000E4B, 0x0000A5BC0000C9BF, 0x00008DBA0000B9BD,
-    0x000095B80000B1BB, 0x00005DB6000089B9, 0x000045B40000A1B7,
-    0x00006DB2000059B5, 0x000108CC000051B3, 0x00000DAE000069B1,
-    0x00000DAC80000E4E, 0x0003ADAA80000E51, 0x0003ADA880000E54,
-    0x00038DA680000E57, 0x00038DA480000E5A, 0x0003ADA280000E5D,
-    0x0003ADA080000E60, 0x00034D9E80000E68, 0x00034D9C80000E6B,
-    0x00032D9A80000E6E, 0x00032D9880000E71, 0x00030D9680000E74,
-    0x0003359480000E77, 0x0003259280000E7A, 0x0003359080000E7D,
-    0x0003458E80000E80, 0x0003B58C00010587, 0x0003A58A80000E83,
-    0x0003B58880000E86, 0x0003BD8680000E89, 0x0003BD8480000E8C,
-    0x0003BD8280000E8F, 0x0003BD8080000E92, 0x00009D9080000EA9,
-    0x0000AD920000A591, 0x0000BD940000A593, 0x0000AD960000A595,
-    0x003FA8CC80000EAC, 0x80000EAF003E74CC, 0x003F50CC80000EB2,
-    0x003F60CC003F74CC, 0x00019580003F88CC, 0x0001AD8200019981,
-    0x0001A5840001A183, 0x000045860001A985, 0x00004D8800007587,
-    0x0000ED8A0000E589, 0x0000FD8C0000E58B, 0x0000ED8E0000E58D,
-    0x80000EB50000E58F, 0x80000EBB80000EB8, 0x80000EC180000EBE,
-    0x80000ECC80000EC4, 0x80000ED280000ECF, 0x80000ED880000ED5,
-    0x80000EDE80000EDB, 0x003A00CD80000EE1, 0x80000E63003A04CC,
-    0x80000FBB003F18A7, 0x0000099400000593, 0x80000FBE00000995,
-    0x0000099480000FC4, 0x80000FC1000254D9, 0x00000994000128CC,
-    0x80000FC7000254D9, 0x8000165D80001039, 0x800016BB8000172C,
-    0x000254D900037881, 0x8000172000000D94, 0x0000059400026CD9,
-    0x000254D980001743, 0x8000171A00001D94, 0x000294E000038CA8,
-    0x00033892000288A4, 0x8000177000000DBC, 0x0002C0A400028CE0,
-    0x0003B0B3000398B0, 0x000005BC00031095, 0x000294E080001782,
-    0x000005BC0002ECA4, 0x8000174680001786, 0x000214A48000174D,
-    0x800017540003A4A5, 0x8000175B000394A7, 0x8000176980001762,
-    0x8000177B80001774, 0x000350AE00033CAD, 0x00034CA8000368AF,
-    0x000328E0000350A9, 0x00007DBE8000179F, 0x800017A3000084B5,
-    0x00001997000030B7, 0x00003CB20000C8B1, 0x000030BC0000E8B3,
-    0x0001D4BE0001ECBD, 0x0000A8B88000178A, 0x0000F4BA80001791,
-    0x000109B8000044BB, 0x000330A600023C87, 0x80001798000380A7,
-    0x800017CA800017C3, 0x800017AE800017A7, 0x800017B500003DB6,
-    0x000025B3800017BC, 0x000139B60006C9B5, 0x00013998000139B7,
-    0x0001C19A00013999, 0x0001C59C0001DD9B, 0x0001C59E0001CD9D,
-    0x0002A0E00001FD9F, 0x0001859C0000759F, 0x0002F0AC0001999D,
-    0x000005BE0002BCAD, 0x00009D99800017E3, 0x0000A19B00036487,
-    0x000009960000AD9A, 0x800017DF00003597, 0x800017E70000B19E,
-    0x800017D8800017D1, 0x00039C920002A4E0, 0x800017EB000238AF,
-    0x000540880002E4E0, 0x0005588A000009BE, 0x00006997800017EF,
-    0x0003AC8700002597, 0x0002DCAE0003E086, 0x00004D9C00026CAF,
-    0x800017F70000459D, 0x0002FCAE0002B4E0, 0x800017F30003FDBE,
-    0x0002DCE000038C86, 0x000244B100001996, 0x000384BF800017FB,
-    0x800017FF0002F4E0, 0x000C0182000244B3, 0x0002C0E000039C86,
-    0x0000719500001595, 0x00022CB300021CB3, 0x8000180F0002F4E0,
-    0x8000180780001803, 0x00039C8600003DBE, 0x0002CCB40003A087,
-    0x80001813000248B5, 0x00001D978000181B, 0x0000299600003995,
-    0x0002F0B400022CE0, 0x00000DBE80001817, 0x000210B70002C0E0,
-    0x00039499000025B3, 0x800018278000181F, 0x0000358A8000182B,
-    0x8000182F00000D8A, 0x80001844000230B7, 0x000079B3000EBD8F,
-    0x00000D9C80001836, 0x8000183D00001D9D, 0x800018568000184B,
-    0x0000358C8000185D, 0x8000186B80001864, 0x8000184F80001872,
-    0x8000188080001879, 0x800018C20000599F, 0x006000BF0001DD83,
-    0x800018D3800018B8, 0x0003BCA5800018C9, 0x00026CE180001887,
-    0x8000189C80001895, 0x800018AA800018A3, 0x00002581800018B1,
-    0x0001A5920001E18D, 0x0000619D8000188E, 0x00000DAE0002A480,
-    0x000ED0840001BD97, 0x80002396800018DA, 0x8000239E8000239A,
-    0x00005593800023A2, 0x0001959C000FDC85, 0x000E88BE800023A6,
-    0x000984BE800023AA, 0x000E90840000B9AC, 0x004410BF800023AE,
-    0x000091A7800023B2, 0x800023BA800023B6, 0x800023BE00015DA9,
-    0x000B38BE000099A2, 0x0002A0E1800023C2, 0x001614E3000131B3,
-    0x0000D1B60008C084, 0x00009DB90000D9B7, 0x000FD886000125B5,
-    0x0000F9B8000F5487, 0x800023C60000F9B9, 0x00020CEF800023CA,
-    0x80001823000A3C84, 0x00028083000D8880, 0x800023CE00038C82,
-    0x800023D20003D484, 0x800023DA800023D6, 0x800023E2800023DE,
-    0x800023EA800023E6, 0x021AC08F800023EE, 0x800023F6800023F2,
-    0x800023FE800023FA, 0x8000240680002402, 0x8000240E8000240A,
-    0x8000241680002412, 0x8000241E8000241A, 0x8000242201D45C9A,
-    0x0000DDBC80002426, 0x0000E9828000242A, 0x0000E9800000F583,
-    0x8000242E0000E981, 0x0000ED8400000587, 0x800024320000E185,
-    0x001A64848000190E, 0x021B24AD80003D3E, 0x80003D420219A0AC,
-    0x80003D4600001D8D, 0x80003D4E80003D4A, 0x021E6CB5021C00B2,
-    0x80003D52021E64B4, 0x0212E0B9021FB0B6, 0x0212ECBB021000B8,
-    0x0211D0BD021DACBA, 0x021400BF0213D4BC, 0x000E61A00213FCBE,
-    0x000E61A2000E61A1, 0x000E6DA4000E61A3, 0x000E8DA6000E91A5,
-    0x000E85A8000E81A7, 0x000E99AA000E8DA9, 0x000EC1AC000E91AB,
-    0x000129AE000EC1AD, 0x00015DB0000125AF, 0x00016DB2000165B1,
-    0x00017DB4000165B3, 0x00016DB6000165B5, 0x000121B8000161B7,
-    0x000121BA000121B9, 0x000121BC000121BB, 0x000121BE000121BD,
-    0x01F50090000121BF, 0x023F089201D08091, 0x80003D5A80003D56,
-    0x80003D6280003D5E, 0x80003D6A80003D66, 0x000D24E080003D6E,
-    0x80003D7680003D72, 0x80003D7E80003D7A, 0x000D848080003D82,
-    0x0074A8820069CC81, 0x00BCA08400517883, 0x00E74C8600985485,
-    0x0129E48800CDDC87, 0x0110988A01340489, 0x015B708C017C688B,
-    0x018DFC8E01A1D08D, 0x001E6DAA01E96C8F, 0x001C01A8001001AB,
-    0x001359AE001E09A9, 0x00121DAC001615AF, 0x000401A2001215AD,
-    0x8000180B000605A3, 0x001A35A6000281A1, 0x0006C9A4001A3DA7,
-    0x80003D860007D1A5, 0x80003D8E80003D8A, 0x80003D9680003D92,
-    0x80003D9E80003D9A, 0x006801B280003DA2, 0x00166DB0006A45B3,
-    0x80003DA6001741B1, 0x006A95B480003DAA, 0x000311AE006C01B5,
-    0x000251AC000311AF, 0x000255AA000301AD, 0x000221A800025DAB,
-    0x0003608700025DA9, 0x80003DB200040486, 0x0000ED8080003DB6,
-    0x0000CD820000C581, 0x000095BE80003DBA, 0x000085BC00009DBF,
-    0x000095BA00009DBD, 0x000365B800009DBB, 0x000355B600009DB9,
-    0x000345B400035DB7, 0x000351B200035DB5, 0x000361B0000351B3,
-    0x80003DBE000361B1, 0x80003DC680003DC2, 0x000E68EF80003DCA,
-    0x80003DD280003DCE, 0x000E68E380003DD6, 0x000ED0E180003DDA,
-    0x80003DE280003DDE, 0x80003DEA80003DE6, 0x80003DF280003DEE,
-    0x80003DF6000381B3, 0x0002DDB680003DFA, 0x80003E0280003DFE,
-    0x0002F9BA000385B5, 0x80003E060002F9BB, 0x0002DDBE80003E0A,
-    0x0002D9BC0002C5BF, 0x000FE8E00002D9BD, 0x80003E0E000341A3,
-    0x00021DA680003E12, 0x80003E1A80003E16, 0x000239AA000345A5,
-    0x80003E22000239AB, 0x00021DAE80003E26, 0x000219AC000205AF,
-    0x00004D80000219AD, 0x0000518200005181, 0x80003E2A000171B3,
-    0x0000D1B60000D1B5, 0x0004408680003E2E, 0x0000F1BA00050487,
-    0x0000D5BC0000F1BB, 0x0000C9BE0000D9BD, 0x0006A9A80000C9BF,
-    0x0006B5AA0006A9A9, 0x000715AC00070DAB, 0x000715AE00071DAD,
-    0x80003E1E000709AF, 0x80003E3680003E32, 0x80003E3E80003E3A,
-    0x80003E4680003E42, 0x000701B880003E4A, 0x000705BA00070DB9,
-    0x000709BC000719BB, 0x000735BE00073DBD, 0x000779B00007D5BF,
-    0x00074DB2000779B1, 0x000741B4000741B3, 0x000741B6000741B5,
-    0x00063DA3000741B7, 0x80003E5280003E4E, 0x80003E5A80003E56,
-    0x00079DA500079DA6, 0x0007BDAB80003E5E, 0x80003E620007BDAA,
-    0x000785AF80003E66, 0x000795AD000785AE, 0x80003E6A000799AC,
-    0x80003E7280003E6E, 0x80003E7A80003E76, 0x80003E8280003E7E,
-    0x0000398080003E86, 0x0000058200003981, 0x80003E8E80003E8A,
-    0x0000B88780003E92, 0x0006A9A80000FC86, 0x0006D9AA0006A9A9,
-    0x000135AC0006D1AB, 0x000135AE00013DAD, 0x80003E9600012DAF,
-    0x80003E9E80003E9A, 0x80003EA680003EA2, 0x80003EAE80003EAA,
-    0x0001E9B880003EB2, 0x000189BA0001E9B9, 0x000199BC000189BB,
-    0x000189BE000199BD, 0x000155B0000189BF, 0x000155B200015DB1,
-    0x0001F9B40001E9B3, 0x0001E9B60001F9B5, 0x80003EB60001E1B7,
-    0x80003EBE80003EBA, 0x80003EC680003EC2, 0x80003ECE80003ECA,
-    0x80003ED20008ECEF, 0x80003ED6000518E1, 0x80003EDA00082CE3,
-    0x80003EE280003EDE, 0x000611B380003EE6, 0x80003EEE80003EEA,
-    0x80003EF680003EF2, 0x0001FDB50001FDB6, 0x0001D9BB80003EFA,
-    0x0000A4BE0001D1BA, 0x00017DBF80003EFE, 0x00017DBD00017DBE,
-    0x001C00860001C1BC, 0x80003F0200034487, 0x80003F060005DDA3,
-    0x000231A6000231A5, 0x80003DAE80003F0A, 0x00021DAA80003F0E,
-    0x00020DAC000215AB, 0x0002B1AE0002B1AD, 0x00038D800002B1AF,
-    0x0003958200039581, 0x0003B8EF80003F12, 0x80003F1A80003F16,
-    0x0004D8E380003F1E, 0x000270E180003F22, 0x80003F2A80003F26,
-    0x80003F3280003F2E, 0x80003F3A80003F36, 0x80003F4280003F3E,
-    0x80003F4A80003F46, 0x80003F5280003F4E, 0x000255A880003F56,
-    0x00038DAA000385A9, 0x00039DAC000385AB, 0x00038DAE000385AD,
-    0x0003FDB0000385AF, 0x00034DB2000345B1, 0x00035DB4000345B3,
-    0x00034DB6000345B5, 0x000341B8000345B7, 0x000359BA00034DB9,
-    0x000349BC000359BB, 0x00033DBE000349BD, 0x80003F5A000335BF,
-    0x80003F6280003F5E, 0x001DD8BE000229B3, 0x000229B6000239B5,
-    0x80003F6A80003F66, 0x000215BA80003F72, 0x0002B5BC0002ADBB,
-    0x0002B5BE0002BDBD, 0x000049800002ADBF, 0x0000698200005981,
-    0x001CF08600005CEF, 0x80003F76001D1C87, 0x80003F7E80003F7A,
-    0x80003F8680003F82, 0x80003F8E80003F8A, 0x80003F92000190E1,
-    0x80003F96000A90E3, 0x80003F9E80003F9A, 0x80003FA680003FA2,
-    0x80003FAE80003FAA, 0x001CE8BE000ADCEF, 0x80003FB200065CE1,
-    0x80003FB6000A70E3, 0x80003FBE80003FBA, 0x0002E5A380003FC2,
-    0x80003FCA80003FC6, 0x80003FD280003FCE, 0x0002F5A50002E5A6,
-    0x000261AB80003FD6, 0x80003FDA0002D9AA, 0x000261AF001CA084,
-    0x000271AD000279AE, 0x001E1DB3000279AC, 0x80003FDE80003F6E,
-    0x80003FE680003FE2, 0x001E55B5001E41B6, 0x001E4DBB80003FEA,
-    0x00008886001E41BA, 0x001E4DBF0003E487, 0x001E4DBD001E4DBE,
-    0x001FA982001E4DBC, 0x001F9980001E59A3, 0x001E05A6001F9181,
-    0x80003FF280003FEE, 0x001E05AA001E11A5, 0x80003FF6001E09AB,
-    0x001E09AE80003FFA, 0x001E09AC001E09AF, 0x80003FFE001E09AD,
-    0x8000400680004002, 0x8000400E8000400A, 0x8000401680004012,
-    0x001EA9A88000401A, 0x001EB5AA001EB5A9, 0x001EFDAC001E89AB,
-    0x001EE1AE001EE1AD, 0x001EA5B0001EE1AF, 0x001EB9B2001EA9B1,
-    0x001EA9B4001EB9B3, 0x001E59B6001EA9B5, 0x001E69B8001E59B7,
-    0x001E79BA001E69B9, 0x001E69BC001E79BB, 0x001FF9BE001E69BD,
-    0x001E25A8001FF9BF, 0x001E55AA001E5DA9, 0x001E75AC001E6DAB,
-    0x001E75AE001E7DAD, 0x8000401E001E6DAF, 0x8000402680004022,
-    0x8000402E8000402A, 0x8000403680004032, 0x001EE9B88000403A,
-    0x001EFDBA001EF5B9, 0x001EEDBC001EF5BB, 0x001E91BE001E91BD,
-    0x001E19B0001E91BF, 0x001EEDB2001EE5B1, 0x001EFDB4001EE5B3,
-    0x001EEDB6001EE5B5, 0x001EDDB3001EE5B7, 0x800040428000403E,
-    0x8000404A80004046, 0x001EFDB5001EFDB6, 0x001E21BB0001D084,
-    0x0001E0BE001EDDBA, 0x001E21BF80004052, 0x001E31BD001E39BE,
-    0x00005182001E39BC, 0x00004180001E99A3, 0x001EB9A600004181,
-    0x8000405A80004056, 0x001E99AA001EB9A5, 0x00004886001E65AB,
-    0x001E7DAE00007887, 0x001E7DAC001E65AF, 0x00020DA8001E75AD,
-    0x000215AA000215A9, 0x000249AC000259AB, 0x000275AE00027DAD,
-    0x8000405E00029DAF, 0x8000406680004062, 0x00033C848000406A,
-    0x800040728000406E, 0x000365B880004076, 0x000365BA00036DB9,
-    0x000365BC00037DBB, 0x000365BE00036DBD, 0x0002E5B00003D9BF,
-    0x0002E5B20002EDB1, 0x0002E5B40002FDB3, 0x0002E5B60002EDB5,
-    0x00029DB300035DB7, 0x8000407E8000407A, 0x8000408680004082,
-    0x0002C1B50002FDB6, 0x000229BB8000408A, 0x8000408E000221BA,
-    0x000229BF80004092, 0x000229BD000221BE, 0x80004096000231BC,
-    0x000444BE0002D9A3, 0x0002B9A600050084, 0x8000409E8000409A,
-    0x000265AA000285A5, 0x800040A200026DAB, 0x000265AE800040A6,
-    0x000275AC00026DAF, 0x0002ADA800026DAD, 0x0002D1AA0002D1A9,
-    0x000135AC00012DAB, 0x000135AE00013DAD, 0x800040AA000129AF,
-    0x800040B2800040AE, 0x800040BA800040B6, 0x800040C6800040C2,
-    0x0001EDB8800040CA, 0x00019DBA000185B9, 0x0001B9BC000195BB,
-    0x000079BE0001B9BD, 0x000159B0000079BF, 0x0001EDB2000159B1,
-    0x0001FDB40001E5B3, 0x0001E5B60001E5B5, 0x001EC8E10001D5B7,
-    0x001E74E3000008E3, 0x0000BD810001BCE1, 0x800040CE0000BD80,
-    0x000580860000AD82, 0x800040D20005F487, 0x800040DA800040D6,
-    0x001EE0EF0000E4EF, 0x000219B3800040DE, 0x00757085800040E2,
-    0x800040EA800040E6, 0x0001FDB50001FDB6, 0x0001DDBB800040EE,
-    0x800040F20001DDBA, 0x000161BF800040F6, 0x000171BD000179BE,
-    0x0005D9A3000179BC, 0x800040FA800040BE, 0x80004102800040FE,
-    0x00063DA500063DA6, 0x00061DAB80004106, 0x8000410A00061DAA,
-    0x0006A1AF8000410E, 0x0006B1AD0006B9AE, 0x800041120006B9AC,
-    0x8000411A80004116, 0x00000D808000411E, 0x00003D8200003581,
-    0x8000412680004122, 0x0003A0BE8000412A, 0x0000F0868000412E,
-    0x800041320003E487, 0x0007B5A880004136, 0x0007B5AA0007BDA9,
-    0x000739AC000715AB, 0x000729AE000739AD, 0x8000413A000721AF,
-    0x800041428000413E, 0x8000414A80004146, 0x800041528000414E,
-    0x0000D1B880004156, 0x0000E1BA0000D9B9, 0x000091BC0000E1BB,
-    0x000091BE000091BD, 0x000725B0000091BF, 0x000739B2000729B1,
-    0x000719B4000731B3, 0x0000F1B6000719B5, 0x00063DB30000F1B7,
-    0x8000415E8000415A, 0x8000416680004162, 0x0006B1B50006B9B6,
-    0x0006EDBB8000416A, 0x8000416E0006E1BA, 0x0006DDBF80004172,
-    0x0006E5BD0006DDBE, 0x800041760006EDBC, 0x8000417A000679A3,
-    0x0006FDA68000417E, 0x8000418680004182, 0x0006A5AA0006F5A5,
-    0x8000418A0006A9AB, 0x000699AE8000418E, 0x0006A9AC000699AF,
-    0x800041920006A1AD, 0x800041960005A4E3, 0x0000118000067CE1,
-    0x0000058200001D81, 0x0002AC848000419A, 0x000164870001A4BE,
-    0x800041A200000C86, 0x800041A6000584EF, 0x000161B3800041AA,
-    0x800041B2800041AE, 0x800041BA800041B6, 0x000171B5000179B6,
-    0x00015DBB800041BE, 0x800041C200015DBA, 0x0000E1BF800041C6,
-    0x0000F1BD0000F9BE, 0x800041CA0000F9BC, 0x8000404E800041CE,
-    0x800041D6800041D2, 0x800041DE800041DA, 0x000259A8800041E2,
-    0x000289AA000259A9, 0x000299AC000289AB, 0x000289AE000299AD,
-    0x0002F9B0000289AF, 0x00034DB20002F9B1, 0x00035DB4000345B3,
-    0x000345B6000345B5, 0x00034DB8000375B7, 0x000321BA000325B9,
-    0x000325BC000321BB, 0x000311BE00032DBD, 0x800041E6000311BF,
-    0x800041EE800041EA, 0x000C58BE00022DA3, 0x000235A600023DA5,
-    0x800041F6800041F2, 0x000211AA800041FA, 0x0003B5AC000211AB,
-    0x0003B5AE0003BDAD, 0x800041FE0003ADAF, 0x00001D8100001D82,
-    0x8000420200001D80, 0x8000420A80004206, 0x000C2C8480004212,
-    0x0003288780004216, 0x8000421A000CCC86, 0x800042228000421E,
-    0x0002B1A880004226, 0x0002B5AA0002BDA9, 0x0002D9AC0002C9AB,
-    0x0002C9AE0002D9AD, 0x8000422A0002C1AF, 0x800042328000422E,
-    0x8000423A80004236, 0x800042428000423E, 0x00015DB880004246,
-    0x0001CDBA0001C5B9, 0x0001CDBC0001C5BB, 0x0001FDBE0001F5BD,
-    0x000289B000019DBF, 0x000169B2000289B1, 0x00017DB4000169B3,
-    0x00016DB6000165B5, 0x000638E1000165B7, 0x000754E38000424A,
-    0x00708C858000424E, 0x8000425680004252, 0x8000425E8000425A,
-    0x8000426680004262, 0x000000E98000426A, 0x000128EF000C7884,
-    0x000015808000426E, 0x0000158200001D81, 0x00008CEF000C04BE,
-    0x8000427680004272, 0x000050E38000427A, 0x000180E180004282,
-    0x8000428A80004286, 0x000F78868000428E, 0x000DBDB3000CF087,
-    0x8000427E8000420E, 0x8000429680004292, 0x000DD1B5000DD9B6,
-    0x000DCDBB8000429A, 0x8000429E000DCDBA, 0x000E21BF800042A2,
-    0x000E3DBD000E39BE, 0x800042A6000E3DBC, 0x800042AA000DF9A3,
-    0x000D9DA6800042AE, 0x800042B6800042B2, 0x000D89AA000D95A5,
-    0x800042BA000D89AB, 0x000E7DAE800042BE, 0x000E79AC000E65AF,
-    0x000E5DA8000E79AD, 0x000E79AA000E61A9, 0x000E69AC000E79AB,
-    0x000FB9AE000E69AD, 0x800042C2000FB9AF, 0x800042CA800042C6,
-    0x800042D2800042CE, 0x800042DA800042D6, 0x000F55B8800042DE,
-    0x000F55BA000F5DB9, 0x000F75BC000F6DBB, 0x000F75BE000F7DBD,
-    0x000FC9B0000F61BF, 0x000FD9B2000FC9B1, 0x000FC9B4000FD9B3,
-    0x000F7DB6000FC9B5, 0x000FF1B3000F71B7, 0x0000158100002D82,
-    0x800042E200001D80, 0x000FE1B5000FE9B6, 0x000E0DBB800042E6,
-    0x800042EA000E0DBA, 0x000E0DBF00022484, 0x000E0DBD000E0DBE,
-    0x000140BE000E0DBC, 0x800042EE000FB5A3, 0x000FADA6800042F2,
-    0x800042FA800042F6, 0x000E49AA000FA5A5, 0x0000C886000E49AB,
-    0x000E49AE00017887, 0x000E49AC000E49AF, 0x000E34E1000E49AD,
-    0x000D78E3800042FE, 0x8000430680004302, 0x8000430E8000430A,
-    0x0000AC8480004312, 0x8000431A80004316, 0x800043228000431E,
-    0x000DD0EF80004326, 0x8000432E8000432A, 0x80004332000E35B3,
-    0x0001C1B680004336, 0x8000433E8000433A, 0x0001CDBA0001D1B5,
-    0x800043420001A5BB, 0x0001A5BE80004346, 0x0001B1BC0001ADBF,
-    0x8000434A0001B1BD, 0x8000434E8000419E, 0x0000098180004352,
-    0x8000435600001980, 0x000DC5A800009D82, 0x000DD9AA000DC9A9,
-    0x00023DAC000DD9AB, 0x000225AE000225AD, 0x00026DB0000215AF,
-    0x000201B2000205B1, 0x000205B4000201B3, 0x000231B600020DB5,
-    0x0002C1B8000231B7, 0x0002C1BA0002C1B9, 0x0002C1BC0002C1BB,
-    0x0002C1BE0002C1BD, 0x8000435A0002C1BF, 0x80004362000400BE,
-    0x80004366000D7DA3, 0x000289A6000299A5, 0x8000436E8000436A,
-    0x000285AA80004372, 0x0002F9AC0002EDAB, 0x0002EDAE0002F9AD,
-    0x800043760002E5AF, 0x00044884000174E3, 0x0004F0860001BCE1,
-    0x8000437A00049487, 0x800043828000437E, 0x8000438A80004386,
-    0x800043928000438E, 0x80004396000288EF, 0x00024DA88000439A,
-    0x000251AA000251A9, 0x00029DAC000289AB, 0x00028DAE000285AD,
-    0x8000439E000285AF, 0x800043A6800043A2, 0x800043AE800043AA,
-    0x800043B6800043B2, 0x000289B80004B0BE, 0x00029DBA000289B9,
-    0x0002B9BC000295BB, 0x000159BE0002B9BD, 0x0002FDB0000155BF,
-    0x0002C1B20002C1B1, 0x0002C1B40002C1B3, 0x0002C5B60002C1B5,
-    0x000F74E10002BDB7, 0x000E2CE30000B8E3, 0x800043BA000E6CE1,
-    0x800043C2800043BE, 0x800043CA800043C6, 0x800043D6800043D2,
-    0x000000EA800043DA, 0x000E8CEF00000CEF, 0x000299B3800043DE,
-    0x0004E886800043E2, 0x800043E60004C087, 0x000255B5000245B6,
-    0x000245BB800043EA, 0x800043EE000279BA, 0x000245BF800043F2,
-    0x000255BD000245BE, 0x8000435E000255BC, 0x000551A3800043CE,
-    0x00059DA5800043F6, 0x800043FE800043FA, 0x8000440200058DA6,
-    0x00058DAB80004406, 0x00059DAD0005B1AA, 0x00058DAF00059DAC,
-    0x0007D98000058DAE, 0x0007E9820007D981, 0x8000440A0006D1B3,
-    0x000699B60006F1B5, 0x800044128000440E, 0x000689BA80004416,
-    0x000699BC000689BB, 0x000689BE000699BD, 0x000609A8000685BF,
-    0x000615AA000615A9, 0x000635AC00062DAB, 0x000641AE000655AD,
-    0x8000441A000641AF, 0x800044228000441E, 0x8000442A80004426,
-    0x800044328000442E, 0x0007E5B880004436, 0x0007E5BA0007EDB9,
-    0x0007E5BC0007FDBB, 0x000749BE0007EDBD, 0x000601B0000749BF,
-    0x000601B2000601B1, 0x000601B4000601B3, 0x000605B600060DB5,
-    0x00079DA30007DDB7, 0x0001B58100001582, 0x8000443A0001B580,
-    0x0007BDA50007D5A6, 0x0007C5AB000190BE, 0x0002A0840007C5AA,
-    0x0007C9AF8000443E, 0x0007D5AD0007C5AE, 0x800044420007D5AC,
-    0x0000E8860006E9B3, 0x000681B600012487, 0x8000444A80004446,
-    0x00068DBA000699B5, 0x8000444E000169BB, 0x000151BE80004452,
-    0x00016DBC000151BF, 0x000625A8000165BD, 0x000625AA00062DA9,
-    0x000655AC000631AB, 0x000645AE000645AD, 0x0001A884000675AF,
-    0x8000445A80004456, 0x800044628000445E, 0x8000446A80004466,
-    0x0001A1B88000446E, 0x000169BA0001A1B9, 0x000179BC000169BB,
-    0x000169BE000179BD, 0x00060DB0000169BF, 0x0001FDB20001E9B1,
-    0x0001D1B40001F5B3, 0x0001A9B60001D1B5, 0x0005ADA30001A1B7,
-    0x8000447680004472, 0x8000447E8000447A, 0x0005DDA50005C5A6,
-    0x00022DAB80004482, 0x800044860005C9AA, 0x000215AF8000448A,
-    0x000221AD000215AE, 0x005C7885000229AC, 0x800044928000448E,
-    0x8000449A80004496, 0x800044A28000449E, 0x00000582800044A6,
-    0x00003180800044AA, 0x00033CBE00003181, 0x800044B6800044B2,
-    0x007C2086800044BA, 0x800044BE0003B487, 0x000218EF800044C2,
-    0x800044CA800044C6, 0x000110E3800044CE, 0x0001F8E1800044D2,
-    0x800044DA800044D6, 0x800044E2800044DE, 0x800044EA800044E6,
-    0x800044EE00039DB3, 0x000375B6800044F2, 0x800044FA800044F6,
-    0x000351BA0003B1B5, 0x800044FE000351BB, 0x0000F5BE80004502,
-    0x000341BC0000FDBF, 0x000241A8000341BD, 0x000251AA000251A9,
-    0x00027DAC000265AB, 0x0002B1AE0002B1AD, 0x007D68BE0002B1AF,
-    0x8000450A80004506, 0x00000D808000450E, 0x0000AD820000A581,
-    0x00015DB880004512, 0x000165BA00016DB9, 0x000119BC000109BB,
-    0x000109BE000119BD, 0x0002D1B0000109BF, 0x0002D1B20002D1B1,
-    0x000175B40002D1B3, 0x000175B600017DB5, 0x000650E100016DB7,
-    0x0000B0E300074CE1, 0x80004516000754E3, 0x007DD4848000451A,
-    0x8000452680004522, 0x8000452E8000452A, 0x007F30BE80004532,
-    0x00005CEF80004536, 0x000299A3000678EF, 0x007C68868000453A,
-    0x8000453E007D0487, 0x0002B5A5000271A6, 0x000255AB80004542,
-    0x80004546000255AA, 0x0001F9AF8000454A, 0x000245AD0001F1AE,
-    0x007E75A8000245AC, 0x007E75AA007E7DA9, 0x007E9DAC007E85AB,
-    0x007E81AE007E81AD, 0x8000451E007E81AF, 0x800045528000454E,
-    0x8000455A80004556, 0x800045628000455E, 0x007E8DB880004566,
-    0x007E91BA007E91B9, 0x007EBDBC007EA5BB, 0x007F71BE007F71BD,
-    0x007EC5B0007F71BF, 0x007ED9B2007EC9B1, 0x007EC9B4007ED9B3,
-    0x007EB9B6007EC9B5, 0x007EFDB3007EB5B7, 0x8000456E8000456A,
-    0x8000457680004572, 0x007E15B5007E01B6, 0x007E0DBB8000457A,
-    0x8000457E007E01BA, 0x007E0DBF80004582, 0x007E0DBD007E0DBE,
-    0x00001982007E0DBC, 0x00006580007EB9A3, 0x007E45A600001981,
-    0x8000458A80004586, 0x007E45AA007E51A5, 0x0003A484007E49AB,
-    0x007E49AE8000458E, 0x007E49AC007E49AF, 0x007E9DA8007E49AD,
-    0x007EADAA007EA5A9, 0x007EBDAC007EA5AB, 0x007EDDAE007EA1AD,
-    0x00012086007ECDAF, 0x8000459200016487, 0x8000459A80004596,
-    0x800045A28000459E, 0x000179B8800045A6, 0x0001CDBA000179B9,
-    0x0001DDBC0001C5BB, 0x0001C5BE0001C5BD, 0x007EBDB00001F5BF,
-    0x007E85B2007E8DB1, 0x000159B4000149B3, 0x000149B6000159B5,
-    0x007E25B3000149B7, 0x800045AE800045AA, 0x800045B6800045B2,
-    0x007E35B5007E21B6, 0x007E49BB800045BA, 0x800045BE007E79BA,
-    0x00019DBF800045C2, 0x00019DBD00019DBE, 0x800045C6007E51BC,
-    0x800045CE800045CA, 0x800045D6800045D2, 0x800045DE800045DA,
-    0x800045E2007C04EF, 0x800045E6007D7CE1, 0x800045EA007C64E3,
-    0x800045F2800045EE, 0x800045FA800045F6, 0x0002C8BE0003D884,
-    0x80004602007DA9A3, 0x007DADA6007DB9A5, 0x8000460A80004606,
-    0x007DF5AA8000460E, 0x007DDDAC007DC5AB, 0x000211AE000211AD,
-    0x0003B980000211AF, 0x000049820003B981, 0x8000461200037DA3,
-    0x000365A600036DA5, 0x800044AE80004616, 0x0003E5AA8000461A,
-    0x0003EDAC0003FDAB, 0x0003D5AE0003DDAD, 0x000420860003BDAF,
-    0x8000461E0003F487, 0x8000462680004622, 0x8000462E8000462A,
-    0x000389A880004632, 0x000399AA000389A9, 0x000389AC000399AB,
-    0x00033DAE000389AD, 0x00034DB0000335AF, 0x000355B2000355B1,
-    0x00037DB4000365B3, 0x00031DB6000315B5, 0x00032DB8000315B7,
-    0x00033DBA000335B9, 0x000311BC000335BB, 0x000309BE00031DBD,
-    0x80004636000309BF, 0x000231B5000239B6, 0x000221B38000463A,
-    0x800046428000463E, 0x0002E1BF80004646, 0x000281BD000289BE,
-    0x0002A1BB0002B1BC, 0x000000CF0002B9BA, 0x8000464E8000464A,
-    0x8000465680004652, 0x0003A0EF00024084, 0x8000465E8000465A,
-    0x000374E380004662, 0x000050E180004666, 0x00002D808000466A,
-    0x0000158200001181, 0x00058CBE80004672, 0x80004676007FD0E3,
-    0x8000467A007608E1, 0x800046828000467E, 0x0007EC84007ED4EF,
-    0x80004686007E74E1, 0x8000468A000154E3, 0x000758860066D0EF,
-    0x0001E1B30004C887, 0x800046928000468E, 0x8000469A80004696,
-    0x000189B5000189B6, 0x000171BB8000469E, 0x800046A2000169BA,
-    0x0000F1BF800046A6, 0x000161BD000169BE, 0x000611A8000169BC,
-    0x000629AA000619A9, 0x000641AC000629AB, 0x00067DAE000649AD,
-    0x8000466E000695AF, 0x800046AE800046AA, 0x0004DDA2800046B2,
-    0x000659A0000249A3, 0x000759B800041DA1, 0x000775BA000759B9,
-    0x000715BC00070DBB, 0x00070DBE00071DBD, 0x0006EDB0000705BF,
-    0x0006FDB20006F5B1, 0x0006EDB40006F5B3, 0x000775B6000775B5,
-    0x0006A1A3000769B7, 0x800046B6006C298F, 0x800046BE800046BA,
-    0x0006C9A50006C9A6, 0x000631AB800046C2, 0x800046C6000629AA,
-    0x0007B1AF800046CA, 0x000621AD000629AE, 0x0078C59E000629AC,
-    0x007D819C0079E99F, 0x007CD59A0078CD9D, 0x0070B598007CDD9B,
-    0x0072E996007ED999, 0x0074ED940070E197, 0x006801920074E595,
-    0x006AD99000691993, 0x006DAD820068DD91, 0x800046CE006B2583,
-    0x0068FD86800046D2, 0x0069398400751187, 0x0075298A0069ED85,
-    0x800046D600732D8B, 0x007F358E800046DA, 0x00713D8C007D318F,
-    0x007D49920071518D, 0x000100BE007C0D93, 0x00787D96800046E2,
-    0x0079599400073597, 0x0005419A00795995, 0x0000C8860005899B,
-    0x0000598100004C87, 0x0003059C00007980, 0x00090DA700005182,
-    0x000CE1A5000905A6, 0x000D91A3000D1DA4, 0x0001B9A10003EDA2,
-    0x001C01AF800046E6, 0x001189AD001179AE, 0x00149DAB001001AC,
-    0x001401A9001595AA, 0x0004B8EF00096DA8, 0x001805B500000CEF,
-    0x0019D5B3001991B4, 0x001DE9B1001801B2, 0x800046EA001DFDB0,
-    0x800046F2800046EE, 0x0004ACE3800046F6, 0x000560E10000D0E3,
-    0x00022DB800028CE1, 0x000285BA00028DB9, 0x00028DBC00029DBB,
-    0x0002B5BE0002BDBD, 0x0005E9B0000315BF, 0x0005F9B20005E9B1,
-    0x0005D9B40005F1B3, 0x00021DB60005D9B5, 0x00053DA8000215B7,
-    0x0005D9AA0005CDA9, 0x0005C9AC0005D9AB, 0x0005E9AE0005FDAD,
-    0x800046FA000599AF, 0x80004702800046FE, 0x8000470A80004706,
-    0x800046DE8000470E, 0x000511B3800045FE, 0x8000471680004712,
-    0x8000471E8000471A, 0x000501B500053DB6, 0x00056DBB80004722,
-    0x8000472600056DBA, 0x00055DBF8000472A, 0x000565BD00055DBE,
-    0x8000472E00056DBC, 0x80004732000555A3, 0x000579A680004736,
-    0x8000473E8000473A, 0x000529AA000545A5, 0x80004742000529AB,
-    0x000519AE80004746, 0x000529AC000519AF, 0x00024DA8000521AD,
-    0x0002E9AA000299A9, 0x0002F9AC0002E9AB, 0x000179AE0002F9AD,
-    0x0001C980000179AF, 0x000005820001C981, 0x000C00BE8000474A,
-    0x8000475200023084, 0x0001DDB880004756, 0x0001E5BA0001EDB9,
-    0x000199BC00018DBB, 0x00018DBE000199BD, 0x000109B0000185BF,
-    0x000119B2000109B1, 0x0001F5B4000119B3, 0x0001F5B60001FDB5,
-    0x000D00860001EDB7, 0x8000475A000D0087, 0x800047628000475E,
-    0x8000476A80004766, 0x8000476E00002CEF, 0x80004772000604E1,
-    0x000C3CBE00061CE3, 0x8000477A80004776, 0x800047828000477E,
-    0x000229B380004786, 0x000239B58000478A, 0x800047928000478E,
-    0x80004796000231B6, 0x000185BB8000479A, 0x000185BD000185BA,
-    0x000171BF000185BC, 0x8000474E000179BE, 0x800047A28000479E,
-    0x800047AA800047A6, 0x800047B2800047AE, 0x000F19A8800047B6,
-    0x000FA1AA000F19A9, 0x000FA1AC000FA1AB, 0x000FA1AE000FA1AD,
-    0x000FE9B0000F9DAF, 0x000FF1B2000FF1B1, 0x000F91B4000FF1B3,
-    0x000F95B6000F91B5, 0x000FB9B8000F89B7, 0x000F8DBA000F85B9,
-    0x000079BC000069BB, 0x000069BE000079BD, 0x00009D81000069BF,
-    0x800047BA00009D80, 0x800047BE00000582, 0x00029484800047C2,
-    0x00000086800047C6, 0x800047CA0003F487, 0x800047D2800047CE,
-    0x800047DA800047D6, 0x000F49B3800047DE, 0x800047E6800047E2,
-    0x800047EE800047EA, 0x000F49B5000F41B6, 0x000F55BB800047F2,
-    0x800047F6000F4DBA, 0x000F35BF800047FA, 0x000F45BD000F35BE,
-    0x000E89A3000F45BC, 0x80004802800047FE, 0x8000480A80004806,
-    0x000E89A5000E81A6, 0x000E95AB8000480E, 0x80004812000E8DAA,
-    0x000EF5AF80004816, 0x000E85AD000EF5AE, 0x8000481A000E85AC,
-    0x800048228000481E, 0x0000198080004826, 0x0000058200001981,
-    0x00026C848000482A, 0x8000482E000148BE, 0x0004008680004832,
-    0x800048360001B487, 0x000FF1AC8000483A, 0x000F0DAE000FF1AD,
-    0x8000483E000F05AF, 0x000FF1AA80004842, 0x80004846000FF1AB,
-    0x000C45A6000C4DA5, 0x8000484E8000484A, 0x8000485680004852,
-    0x000169B3000C5DA3, 0x8000485E8000485A, 0x8000486680004862,
-    0x000169B5000161B6, 0x000129BB8000486A, 0x8000486E000121BA,
-    0x000119BF80004872, 0x000129BD000129BE, 0x000295A8000131BC,
-    0x0002A1AA000299A9, 0x0002A1AC0002A1AB, 0x0002A1AE0002A1AD,
-    0x800048760002A1AF, 0x8000487E8000487A, 0x8000488680004882,
-    0x8000488E8000488A, 0x000349B880004892, 0x000359BA000349B9,
-    0x000379BC000351BB, 0x000319BE000379BD, 0x0002A5B0000319BF,
-    0x0002A5B20002ADB1, 0x0002ADB40002BDB3, 0x000295B600029DB5,
-    0x000249AB000379B7, 0x00030484000241AA, 0x000279AF80004896,
-    0x000249AD000249AE, 0x000209A3000251AC, 0x0003F0BE8000489A,
-    0x800048A6800048A2, 0x000209A5000201A6, 0x0000FD80800048AA,
-    0x00006D8200004981, 0x800048B2800048AE, 0x800048B6000284EF,
-    0x800048BE800048BA, 0x800048C200027CE3, 0x800048C6000184E1,
-    0x00045886800048CA, 0x800048CE00036487, 0x800048D6800048D2,
-    0x800048DA000540BE, 0x800048DE0001ACE1, 0x800048E200006CE3,
-    0x800048EA800048E6, 0x000578840000B0EF, 0x800048F2800048EE,
-    0x00023DB3800048F6, 0x800048FE800048FA, 0x000418BE80004902,
-    0x0002CDB50002F5B6, 0x0002D9BB80004906, 0x000000D20002D1BA,
-    0x0002A1BF8000490A, 0x0002B1BD0002B9BE, 0x800049120002B9BC,
-    0x80004916000E1CE1, 0x0000FD82000104E3, 0x0000D5808000491A,
-    0x00056C870000C181, 0x8000491E00040486, 0x8000492680004922,
-    0x8000492E8000492A, 0x000261A6000EF4EF, 0x8000493680004932,
-    0x8000493A000259A5, 0x8000493E0002A9A3, 0x00022DAE80004942,
-    0x00022DAC000235AF, 0x000245AA000225AD, 0x8000494600024DAB,
-    0x0006ADA88000494A, 0x0006B5AA0006B5A9, 0x0006C1AC0006C1AB,
-    0x0006C5AE0006CDAD, 0x8000494E0006FDAF, 0x8000495680004952,
-    0x8000495E8000495A, 0x8000496680004962, 0x000761B88000496A,
-    0x000761BA000761B9, 0x000769BC000779BB, 0x00071DBE000769BD,
-    0x000689B0000715BF, 0x00069DB2000695B1, 0x000775B400076DB3,
-    0x000769B600077DB5, 0x8000489E000765B7, 0x8000496E000639B3,
-    0x000649B68000490E, 0x8000497680004972, 0x000649BA00064DB5,
-    0x8000497A000655BB, 0x0007B5BE0003E084, 0x00064DBC0007BDBF,
-    0x00067DA30007BDBD, 0x0000518100006D82, 0x8000497E00005980,
-    0x000609A500060DA6, 0x000611AB80004982, 0x0000488600060DAA,
-    0x0007F9AF00006887, 0x0007F9AD0007F1AE, 0x000681A8000609AC,
-    0x00069DAA00068DA9, 0x0006BDAC0006A5AB, 0x0006ADAE0006A5AD,
-    0x800049860006A5AF, 0x8000498E8000498A, 0x8000499680004992,
-    0x8000499E8000499A, 0x000165B8800049A2, 0x000165BA00016DB9,
-    0x000165BC00017DBB, 0x000165BE00016DBD, 0x0006A5B00001D5BF,
-    0x0006B9B20006ADB1, 0x0006ADB40006B5B3, 0x000685B600068DB5,
-    0x000615B300015DB7, 0x800049AA800049A6, 0x800049B2800049AE,
-    0x000635B5000625B6, 0x000669BB800049B6, 0x800049BA000661BA,
-    0x000649BF800049BE, 0x000661BD000669BE, 0x800049C200066DBC,
-    0x800049C6000651A3, 0x000661A6800049CA, 0x800049D2800049CE,
-    0x000625AA000671A5, 0x800049D600062DAB, 0x00062DAE800049DA,
-    0x000629AC00060DAF, 0x0002DDA8000625AD, 0x00032DAA000325A9,
-    0x00033DAC000325AB, 0x00032DAE000325AD, 0x800049DE000325AF,
-    0x800049E6800049E2, 0x00027CBE800049EA, 0x800049F6800049F2,
-    0x000381B8800049FA, 0x000385BA00038DB9, 0x000389BC000399BB,
-    0x0003B5BE0003BDBD, 0x00035DB000007DBF, 0x0003E1B20003E1B1,
-    0x0003E1B40003E1B3, 0x0003D1B60003E9B5, 0x0000A9810003D1B7,
-    0x0003ADB30000A180, 0x0003BDB50000A182, 0x80004A02800049FE,
-    0x0002E4840003ADB6, 0x00032DBB80004A06, 0x00033DBD000391BA,
-    0x000325BF00033DBC, 0x0003E5A300032DBE, 0x0003A88780004A0A,
-    0x80004A0E001C6C86, 0x0003F5A50003E5A6, 0x000365AB80004A12,
-    0x80004A160003D9AA, 0x00036DAF80004A1A, 0x000375AD000365AE,
-    0x0001BCE3000375AC, 0x0006B4E10007E4E1, 0x80004A1E00062CE3,
-    0x80004A2680004A22, 0x80004A2E80004A2A, 0x80004A3680004A32,
-    0x000698EF80004A3A, 0x80004A4280004A3E, 0x80004A460006C8EF,
-    0x001D048480004A4A, 0x80004A4E001D04BE, 0x80004A520001A4E1,
-    0x000039800003ECE3, 0x00001D8200001581, 0x80004A560003A4EF,
-    0x001C588780004A5A, 0x0005ADBC001CDC86, 0x0005BDBE0005B5BD,
-    0x00067DB80005B5BF, 0x0005BDBA0005B5B9, 0x00066DB40005B5BB,
-    0x000675B6000675B5, 0x000605B0000645B7, 0x00067DB200060DB1,
-    0x000655AC000675B3, 0x000665AE000665AD, 0x0006BDA800067DAF,
-    0x0006A5AA0006ADA9, 0x80004A6200064DAB, 0x80004A6A80004A66,
-    0x80004A7280004A6E, 0x80004A7A80004A76, 0x001E35A880004A7E,
-    0x001E35AA001E3DA9, 0x001E89AC001E99AB, 0x001EDDAE001EB1AD,
-    0x80004A5E001ECDAF, 0x80004A8680004A82, 0x80004A8E80004A8A,
-    0x80004A9680004A92, 0x001F65B880004A9A, 0x001F65BA001F6DB9,
-    0x001F69BC001F79BB, 0x001F1DBE001F15BD, 0x001EBDB0001F0DBF,
-    0x001E85B2001E8DB1, 0x001F65B4001F7DB3, 0x001F65B6001F6DB5,
-    0x001EBDB3001F5DB7, 0x80004AA280004A9E, 0x80004AAA80004AA6,
-    0x001E01B5001E31B6, 0x001E49BB80004AAE, 0x80004AB2001E79BA,
-    0x001FA9BF80004AB6, 0x001FA9BD001FA1BE, 0x80004ABA001E51BC,
-    0x80004ABE001EF9A3, 0x001E75A680004AC2, 0x80004ACA80004AC6,
-    0x001E3DAA001E45A5, 0x0003B484001E0DAB, 0x001FE5AE0003A0BE,
-    0x001E15AC001FEDAF, 0x001EE1A8001FEDAD, 0x001EE1AA001EE1A9,
-    0x001EE1AC001EE1AB, 0x001EE1AE001EE1AD, 0x0001ED80001EE1AF,
-    0x0001F1820001F181, 0x0000908680004ACE, 0x80004AD200017487,
-    0x0001C9B880004AD6, 0x0001D9BA0001C9B9, 0x0001F9BC0001D1BB,
-    0x000199BE0001F9BD, 0x000145B0000195BF, 0x000145B200014DB1,
-    0x000145B400015DB3, 0x000145B600014DB5, 0x80004ADA0001F9B7,
-    0x001E79B5001E71B6, 0x001E79B380004ADE, 0x80004AE680004AE2,
-    0x000199BF80004AEA, 0x000199BD000191BE, 0x000199BB000181BC,
-    0x800049EE000191BA, 0x001E29A380004AEE, 0x80004AF680004AF2,
-    0x80004AFE80004AFA, 0x001E29A5001E21A6, 0x0001C9AB80004B02,
-    0x80004B060001C1AA, 0x0001C9AF80004B0A, 0x0001C9AD0001C1AE,
-    0x80004B0E0001D1AC, 0x80004B12001FB0EF, 0x80004B1A80004B16,
-    0x80004B1E00029484, 0x80004B2680004B22, 0x80004B2E80004B2A,
-    0x80004B3680004B32, 0x80004B3A001F6CE3, 0x00006D80001EE8E1,
-    0x0000718200007581, 0x80004B420003CDB3, 0x0003C1B60003C1B5,
-    0x0004C086000210BE, 0x0003ADBA0003E087, 0x0003BDBC0003A5BB,
-    0x0003A5BE0003A5BD, 0x80004B46000395BF, 0x80004B4E80004B4A,
-    0x000234EF80004B52, 0x80004B5A80004B56, 0x0002A4E380004B5E,
-    0x000180E180004B62, 0x80004B6A80004B66, 0x80004B7280004B6E,
-    0x000341A380004B76, 0x80004B7E80004B7A, 0x80004B8680004B82,
-    0x00034DA500034DA6, 0x000329AB80004B8A, 0x80004B8E000321AA,
-    0x000319AF80004B92, 0x000329AD000329AE, 0x000239A8000331AC,
-    0x000289AA000239A9, 0x000285AC000281AB, 0x000285AE00028DAD,
-    0x00022C840002A1AF, 0x002C00BF80004B96, 0x80004B9E80004B9A,
-    0x80004BA680004BA2, 0x000149B880004BAA, 0x000159BA000149B9,
-    0x00014DBC000155BB, 0x00017DBE000175BD, 0x0002A5B0000175BF,
-    0x0002A5B20002ADB1, 0x0002ADB40002BDB3, 0x000295B600029DB5,
-    0x0000B98100017DB7, 0x000584BE0000B180, 0x000430860000B182,
-    0x80004BB200055487, 0x0005A084000188EF, 0x80004BB6001EC0E1,
-    0x80004BBA000154E3, 0x80004BC280004BBE, 0x000145B380004BC6,
-    0x80004BCE80004BCA, 0x80004BD680004BD2, 0x000155B5000141B6,
-    0x00014DBB80004BDA, 0x80004BDE000141BA, 0x00014DBF80004BE2,
-    0x00014DBD00014DBE, 0x000651A800014DBC, 0x00068DAA000685A9,
-    0x00069DAC000685AB, 0x000681AE000681AD, 0x80004BAE000681AF,
-    0x80004BEA80004BE6, 0x80004BF280004BEE, 0x80004BFA80004BF6,
-    0x00068DB880004BFE, 0x000695BA000695B9, 0x0006BDBC0006A5BB,
-    0x000775BE000771BD, 0x0006C5B000076DBF, 0x0006D9B20006C9B1,
-    0x0006C9B40006D9B3, 0x0006BDB60006C9B5, 0x000605A30006B5B7,
-    0x80004C0680004C02, 0x80004C0E80004C0A, 0x000615A5000601A6,
-    0x00060DAB80004C12, 0x80004C16000601AA, 0x00060DAF80004C1A,
-    0x00060DAD00060DAE, 0x00005D8000060DAC, 0x0000658200006D81,
-    0x0003DCBE00077DB3, 0x000765B600076DB5, 0x0000E08680004C1E,
-    0x000725BA00031487, 0x00072DBC00073DBB, 0x000715BE00071DBD,
-    0x000699A80000D9BF, 0x0006B9AA0006ADA9, 0x0006A9AC0006B9AB,
-    0x0006DDAE0006A9AD, 0x80004C220006D5AF, 0x80004C2A80004C26,
-    0x80004C3280004C2E, 0x80004C3A80004C36, 0x000171B880004C3E,
-    0x000171BA000171B9, 0x0001D5BC000171BB, 0x0001D5BE0001DDBD,
-    0x0006ADB00001CDBF, 0x0006B5B20006B5B1, 0x00069DB4000685B3,
-    0x000151B6000151B5, 0x000631A3000151B7, 0x80004C4680004C42,
-    0x80004C4E80004C4A, 0x000621A5000629A6, 0x000671AB80004C52,
-    0x80004C56000669AA, 0x000195AF80004B3E, 0x000651AD000659AE,
-    0x80004C5A000661AC, 0x80004C6280004C5E, 0x80004C6A80004C66,
-    0x80004C7280004C6E, 0x80004C7A80004C76, 0x80004C8280004C7E,
-    0x0000198080004C86, 0x0000058200001981, 0x0002808480004C8A,
-    0x00030887000384BE, 0x000760EF000CAC86, 0x80004C9680004C92,
-    0x000710E380004C9A, 0x000650E180004C9E, 0x80004CA680004CA2,
-    0x80004CAE80004CAA, 0x80004CB680004CB2, 0x80004CBE80004CBA,
-    0x80004CC2000365B3, 0x00036DB6000365B5, 0x80004CCA80004CC6,
-    0x0003EDBA80004CCE, 0x0003FDBC0003E5BB, 0x0003E5BE0003E5BD,
-    0x80004CD20003D5BF, 0x80004CDA80004CD6, 0x80004CE280004CDE,
-    0x80004CEA80004CE6, 0x000381A880004CEE, 0x000381AA000381A9,
-    0x000381AC000381AB, 0x000381AE000381AD, 0x000345B0000381AF,
-    0x000345B200034DB1, 0x00034DB400035DB3, 0x000319B600037DB5,
-    0x000329B8000319B7, 0x000339BA000329B9, 0x000329BC000339BB,
-    0x000319BE000329BD, 0x00001180000319BF, 0x0000298200001981,
-    0x80004CF2000225A3, 0x00022DA6000225A5, 0x80004CFA80004CF6,
-    0x0002ADAA80004D02, 0x0002BDAC0002A5AB, 0x0002A5AE0002A5AD,
-    0x00024C84000295AF, 0x000D6487000D48BE, 0x000340E3000CEC86,
-    0x0001A0E180004D06, 0x00039CEF80004D0A, 0x80004D1280004D0E,
-    0x80004D1A80004D16, 0x80004D2280004D1E, 0x80004D2A80004D26,
-    0x80004D3280004D2E, 0x80004D3A80004D36, 0x80004D4280004D3E,
-    0x000DA8840001B4EF, 0x80004D4600062CE1, 0x80004D4A00061CE3,
-    0x80004D5280004D4E, 0x0001DDB380004D56, 0x80004D5E80004D5A,
-    0x80004D6680004D62, 0x0001F5B50001E5B6, 0x0001A1BB80004D6A,
-    0x000CECBE0001D9BA, 0x0001ADBF80004D72, 0x0001B1BD0001A5BE,
-    0x000E15A80001B9BC, 0x000E15AA000E1DA9, 0x000E3DAC000E29AB,
-    0x000E2DAE000E25AD, 0x80004CFE000E25AF, 0x000FF981000FF982,
-    0x80004D6E000FE180, 0x00001C8680004D76, 0x000EC1B80003B887,
-    0x000EC1BA000EC1B9, 0x000EC1BC000EC1BB, 0x000EC1BE000EC1BD,
-    0x000E5DB0000EC1BF, 0x000E2DB2000E25B1, 0x000E21B4000E25B3,
-    0x000E19B6000E29B5, 0x000E9DA3000E19B7, 0x80004D7E80004D7A,
-    0x80004D8680004D82, 0x000EB5A5000EA5A6, 0x000EE1AB80004D8A,
-    0x80004D8E000E99AA, 0x000EEDAF80004D92, 0x000EF1AD000EE5AE,
-    0x80004D96000EF9AC, 0x80004D9A000F99B3, 0x000FBDB680004D9E,
-    0x80004DA680004DA2, 0x000F99BA000FB5B5, 0x80004DAA000F99BB,
-    0x000F79BE80004DAE, 0x000F79BC000F79BF, 0x000E25A8000F79BD,
-    0x000E55AA000E51A9, 0x000E79AC000E69AB, 0x000E69AE000E79AD,
-    0x80004DB2000E69AF, 0x80004DBA80004DB6, 0x80004DC280004DBE,
-    0x80004DCA80004DC6, 0x000ED9B880004DCE, 0x000EF9BA000EEDB9,
-    0x000EE9BC000EF9BB, 0x000E9DBE000EE9BD, 0x000E19B0000E95BF,
-    0x000EE9B2000E19B1, 0x000EF9B4000EE9B3, 0x000EE9B6000EF9B5,
-    0x00001981000EE9B7, 0x000ED5A300001180, 0x000EF9A500002982,
-    0x80004DD680004DD2, 0x80004C8E000EF1A6, 0x000ED5AB80004DDA,
-    0x000E35AD000ED5AA, 0x000E35AF000E35AC, 0x0002C884000E35AE,
-    0x00010CBE000141B3, 0x000141B680004DE2, 0x80004DEA80004DE6,
-    0x00014DBA000151B5, 0x00004086000125BB, 0x000125BE00015087,
-    0x000131BC00012DBF, 0x80004DEE000131BD, 0x0003048480004DF2,
-    0x80004DF6000480BE, 0x80004DFE80004DFA, 0x00CEA8BF000FA0EF,
-    0x80004E02000E30E1, 0x80004E06000F7CE3, 0x80004E0E80004E0A,
-    0x80004E1680004E12, 0x0002CDA380004E1A, 0x0002DDA580004E1E,
-    0x80004E2680004E22, 0x80004E2A0002CDA6, 0x0002A9AB80004E2E,
-    0x0002BDAD0002C1AA, 0x0002A1AF0002BDAC, 0x0002B5A90002A9AE,
-    0x000201AB000D69A8, 0x000201AD000209AA, 0x000231AF000219AC,
-    0x00056C84000201AE, 0x80004E3680004E32, 0x80004E3E80004E3A,
-    0x00019D8100019D82, 0x0003E5B900019D80, 0x0003E5BB0003E5B8,
-    0x0003E5BD0003E5BA, 0x0003E5BF0003E5BC, 0x000221B10003E5BE,
-    0x000225B3000249B0, 0x000229B5000225B2, 0x000215B7000221B4,
-    0x0002D9AA000215B6, 0x0002C5A8000105AB, 0x000105AE0002D9A9,
-    0x00011DAC00010DAF, 0x80004E4200010DAD, 0x80004E4A80004E46,
-    0x0004F8BE80004E52, 0x80004E5A80004E56, 0x000121BA80004E5E,
-    0x00011DB8000121BB, 0x0001FDBE000115B9, 0x0001FDBC0001E5BF,
-    0x000155B20001F5BD, 0x000141B000015DB3, 0x000135B600015DB1,
-    0x000145B4000139B7, 0x0005CC8700013DB5, 0x80004E620005A486,
-    0x80004E660000BCEF, 0x80004E6E80004E6A, 0x00076C84000E8CEF,
-    0x80004E72000EF0E1, 0x80004E76000E1CE3, 0x80004E7A000194E1,
-    0x000219B300006CE3, 0x80004E8280004E7E, 0x0038008480004E86,
-    0x0001DDB50001C5B6, 0x0001D1BB80004E8A, 0x80004E8E0001CDBA,
-    0x0001B1BF80004E92, 0x0001C1BD0001C9BE, 0x0005D9A30001C9BC,
-    0x80004E9680004E4E, 0x80004E9E80004E9A, 0x00061DA5000605A6,
-    0x000611AB80004EA2, 0x80004EA600060DAA, 0x000671AF80004EAA,
-    0x000601AD000609AE, 0x80004EAE000609AC, 0x00001D8100001D82,
-    0x80004EB200001D80, 0x80004EBA80004EB6, 0x0003208680004EBE,
-    0x80004EC20003A487, 0x80004ECA80004EC6, 0x80004ED280004ECE,
-    0x0007ADA880004ED6, 0x0007B1AA0007B1A9, 0x000719AC0007B1AB,
-    0x00070DAE000705AD, 0x80004EDA000705AF, 0x80004EE280004EDE,
-    0x80004EEA80004EE6, 0x80004EF280004EEE, 0x00072DB880004EF6,
-    0x0000C5BA0000CDB9, 0x0000CDBC0000DDBB, 0x0000F5BE0000FDBD,
-    0x000749B000009DBF, 0x000751B2000755B1, 0x000739B4000729B3,
-    0x000725B6000739B5, 0x000639B3000715B7, 0x80004EFE80004EFA,
-    0x80004F0680004F02, 0x000691B5000685B6, 0x000691BB80004F0A,
-    0x80004F0E00068DBA, 0x0006FDBF80004F12, 0x0006FDBD0006FDBE,
-    0x80004F16000685BC, 0x80004F1A00067DA3, 0x0006C1A680004F1E,
-    0x80004F2680004F22, 0x0006C9AA0006D5A5, 0x80004F2A0006D5AB,
-    0x0006B9AE00016CBE, 0x0006C1AC0006B9AF, 0x000129B00006B9AD,
-    0x000139B2000129B1, 0x000129B4000139B3, 0x00009DB6000129B5,
-    0x0000ADB8000095B7, 0x0000B5BA0000B5B9, 0x00009DBC000085BB,
-    0x0000F1BE0000F5BD, 0x000095800000F1BF, 0x0000A18200009D81,
-    0x80004F3280004F2E, 0x0001088780004DDE, 0x000189A800007C86,
-    0x000199AA000189A9, 0x000189AC000199AB, 0x000159AE000189AD,
-    0x000271B3000159AF, 0x80004F3A80004F36, 0x80004F4280004F3E,
-    0x000261B5000269B6, 0x00028DBB80004F46, 0x80004F4A00028DBA,
-    0x00028DBF80004F4E, 0x00028DBD00028DBE, 0x000780E300028DBC,
-    0x00018CE180004F52, 0x0005A0E100000096, 0x000640E380004F56,
-    0x80004F5E80004F5A, 0x00069CEF80004F62, 0x0002F0EF80004F66,
-    0x003CF4BE80004F6A, 0x80004F7280004F6E, 0x80004F7A80004F76,
-    0x80004F7E0002FDA3, 0x0002E5A60002EDA5, 0x80004F8680004F82,
-    0x000201AA80004F8A, 0x000201AC000201AB, 0x000201AE000201AD,
-    0x80004F8E000201AF, 0x80004F9680004F92, 0x0000158080004F9A,
-    0x0000158200001D81, 0x003CA485000F08EF, 0x80004FA2000210E1,
-    0x003E5086003F48E3, 0x80004FA600033487, 0x0002D1A880004FAA,
-    0x000101AA000101A9, 0x000101AC000101AB, 0x000131AE000109AD,
-    0x80004FAE000131AF, 0x80004FB680004FB2, 0x80004FBE80004FBA,
-    0x80004FC680004FC2, 0x0001C1B880004FCA, 0x0001C1BA0001C1B9,
-    0x0001CDBC0001DDBB, 0x0001FDBE0001F5BD, 0x000125B000019DBF,
-    0x000135B2000131B1, 0x000115B400010DB3, 0x00010DB6000119B5,
-    0x00061CE1000101B7, 0x000738E380004FCE, 0x80004FD680004FD2,
-    0x80004FDA003C1884, 0x80004FE280004FDE, 0x80004FEA80004FE6,
-    0x80004FF280004FEE, 0x000670EF80004FF6, 0x00001D8080004FFA,
-    0x0000218200002181, 0x000660E3003FC0BE, 0x0001A0E180004FFE,
-    0x000074EF80005002, 0x8000500A80005006, 0x003E508680005012,
-    0x80005016003C0087, 0x003E09A38000501A, 0x80004F9E0024258F,
-    0x8000500E00000884, 0x003E09A5003E09A6, 0x003E75AB8000501E,
-    0x80005022003E61AA, 0x003E5DAF80005026, 0x003E59AD003E51AE,
-    0x0038619E003E61AC, 0x0034959C0038759F, 0x0035FD9A0039ED9D,
-    0x003079980034759B, 0x002CB19600317199, 0x002CE9940031DD97,
-    0x0028E192002CE195, 0x0026019000291D93, 0x00140DB100287D91,
-    0x001801B3001405B0, 0x001879B5001455B2, 0x8000502A0018B5B4,
-    0x0001E0BE8000502E, 0x8000503680005032, 0x003CEDA28000503A,
-    0x003D6DA0003D79A3, 0x0000C9A1003CB9A1, 0x000401A38000503E,
-    0x00047DA50000DDA2, 0x0008F9A70004F5A4, 0x000C01A9000801A6,
-    0x000CCDAB00086DA8, 0x0010D1AD000CCDAA, 0x0011D5AF001001AC,
-    0x800050420010D9AE, 0x003F15B6003F15B5, 0x8000504A80005046,
-    0x800050528000504E, 0x003F7DBC003F15B3, 0x003F4DBE003F45BD,
-    0x80005056003F41BF, 0x003F7DBA8000505A, 0x003FBDB3003F69BB,
-    0x800050628000505E, 0x8000506A80005066, 0x003E7DB5003E7DB6,
-    0x003E59BB8000506E, 0x80005072003E51BA, 0x003EE1BF80005076,
-    0x003EFDBD003EF9BE, 0x0025A182003E41BC, 0x8000507A0027FD83,
-    0x0028F98680005082, 0x00299984002D1587, 0x002D2D8A00290D85,
-    0x00070886002D758B, 0x00309D8E00008C87, 0x0030018C0034198F,
-    0x0035E19200319D8D, 0x8000508600351993, 0x0039E1968000508A,
-    0x00387994003C0197, 0x003DE59A0039FD95, 0x8000508E003DB99B,
-    0x00009D8180005092, 0x003C359C00009D80, 0x003CE9A800007582,
-    0x003CF5AA003CFDA9, 0x003D35AC003D2DAB, 0x003D35AE003D3DAD,
-    0x80005096003D2DAF, 0x8000509E8000509A, 0x800050A6800050A2,
-    0x000470BE800050AA, 0x003DE1B8800050AE, 0x003DE1BA003DE1B9,
-    0x003DE1BC003DE1BB, 0x003DE5BE003DE1BD, 0x003D55B0003D89BF,
-    0x003D55B2003D5DB1, 0x003DFDB4003DE9B3, 0x003DE1B6003DE5B5,
-    0x003C7DA3003DE1B7, 0x800050B6800050B2, 0x800050BA0005C4BE,
-    0x003DBDA5003DBDA6, 0x003D99AB800050BE, 0x800050C2003D91AA,
-    0x003D21AF800050C6, 0x003D3DAD003D39AE, 0x800050CA003D81AC,
-    0x003D35B3800050CE, 0x0002F9B5800050D2, 0x800050DA800050D6,
-    0x800050DE0002F1B6, 0x000241BB800050E2, 0x000235BD000249BA,
-    0x0003F5BF000259BC, 0x003CEDA80003FDBE, 0x003CF5AA003CFDA9,
-    0x000339AC000311AB, 0x000381AE000339AD, 0x0001C580000381AF,
-    0x0000198200000981, 0x800050EA800050E6, 0x800050F200057884,
-    0x000259B8800050F6, 0x00027DBA000275B9, 0x000219BC000209BB,
-    0x000209BE000219BD, 0x0003C5B0000201BF, 0x0003C5B20003CDB1,
-    0x000271B40003A9B3, 0x000269B6000279B5, 0x800050FA000269B7,
-    0x0004C484003E00E3, 0x800050FE003F64E1, 0x80005102003E7CE1,
-    0x00040086000180E3, 0x8000510600048487, 0x001A14850001F4EF,
-    0x8000510A003E80EF, 0x800051128000510E, 0x8000511A80005116,
-    0x00079084003C31A3, 0x0003F5A60003FDA5, 0x800051228000511E,
-    0x00034DAA80005126, 0x00035DAC000345AB, 0x0002F9AE000331AD,
-    0x00058DA80002F1AF, 0x0005C5AA0005CDA9, 0x0005FDAC0005E5AB,
-    0x00063DAE000635AD, 0x800050EE000635AF, 0x8000512E8000512A,
-    0x8000513680005132, 0x8000513E8000513A, 0x0006E9B880005142,
-    0x000689BA0006E9B9, 0x00069DBC000689BB, 0x000681BE000681BD,
-    0x00064DB00006BDBF, 0x00065DB2000655B1, 0x0006F5B4000651B3,
-    0x0006E9B60006FDB5, 0x00054DB30006E1B7, 0x8000514A80005146,
-    0x800051528000514E, 0x000575B50006B5B6, 0x000691BB80005156,
-    0x8000515A000689BA, 0x000639BF8000515E, 0x000685BD000631BE,
-    0x80005162000689BC, 0x8000507E000509A3, 0x0006F1A680005166,
-    0x8000516E8000516A, 0x0006CDAA000531A5, 0x000048860006D5AB,
-    0x000675AE00006C87, 0x0006CDAC00067DAF, 0x000639A80006C1AD,
-    0x000689AA000639A9, 0x000689AC00069DAB, 0x0006A1AE0006BDAD,
-    0x0001B1800006A1AF, 0x000181820001B981, 0x8000517680005172,
-    0x8000517E8000517A, 0x00016DB880005182, 0x00017DBA000175B9,
-    0x00016DBC000175BB, 0x0001D1BE0001D1BD, 0x0006A1B00001D1BF,
-    0x0006A1B20006A1B1, 0x000691B40006B1B3, 0x00015DB6000691B5,
-    0x000601B3000155B7, 0x8000518A80005186, 0x800051928000518E,
-    0x000601B500063DB6, 0x000665BB80005196, 0x8000519A000665BA,
-    0x000655BF8000519E, 0x000665BD000655BE, 0x800051A2000675BC,
-    0x800051A6000645A3, 0x000679A6800051AA, 0x800051B2800051AE,
-    0x000621AA000645A5, 0x800051B6000621AB, 0x000611AE800051BA,
-    0x000631AC000611AF, 0x0002C1A8000621AD, 0x0002DDAA0002C9A9,
-    0x00033DAC0002D5AB, 0x00032DAE000325AD, 0x800051BE000325AF,
-    0x800051C6800051C2, 0x800051D2800051CA, 0x800051DA800051D6,
-    0x000381B8000270BE, 0x000385BA00038DB9, 0x000389BC000399BB,
-    0x0003B5BE0003BDBD, 0x00035DB000007DBF, 0x0003E1B20003E1B1,
-    0x0003E1B40003E1B3, 0x0003D1B60003E9B5, 0x0000A5800003D1B7,
-    0x0000A5820000AD81, 0x800051DE000259B3, 0x0003A9B60003B9B5,
-    0x000C0086800051E2, 0x000395BA0003F487, 0x000339BC000329BB,
-    0x000329BE000339BD, 0x000215A3000321BF, 0x800051EA800051E6,
-    0x800051F2800051EE, 0x0003F5A50003E5A6, 0x000365AB800051F6,
-    0x800051FA0003D9AA, 0x00036DAF800051FE, 0x000375AD000365AE,
-    0x80005202000375AC, 0x8000520A80005206, 0x0003D4EF8000520E,
-    0x8000521680005212, 0x0006E0E38000521A, 0x0001D8E18000521E,
-    0x8000522680005222, 0x8000522E8000522A, 0x8000523680005232,
-    0x00024484000660E3, 0x00006D800004FCE1, 0x0000058200001581,
-    0x000D6CBE8000523A, 0x000CC88680005242, 0x80005246000D5C87,
-    0x8000524A0007B4EF, 0x800052528000524E, 0x8000525A80005256,
-    0x800052628000525E, 0x8000526A80005266, 0x000DA0840007E8EF,
-    0x8000526E00061CE1, 0x80005272000634E3, 0x8000527A80005276,
-    0x0001F9B38000527E, 0x8000528680005282, 0x8000528E8000528A,
-    0x000145B5000111B6, 0x00013DBB80005292, 0x80005296000135BA,
-    0x0000F5BF8000529A, 0x000115BD0000F5BE, 0x000E31A8000115BC,
-    0x000E31AA000E31A9, 0x000E8DAC000E9DAB, 0x000ED1AE000ED1AD,
-    0x8000523E000ED1AF, 0x800052A28000529E, 0x00001980800052A6,
-    0x0000058200001981, 0x000F55B8800052AA, 0x000F55BA000F5DB9,
-    0x000F7DBC000F69BB, 0x000F6DBE000F65BD, 0x000EB1B0000F65BF,
-    0x000E81B2000EB9B1, 0x000F71B4000E81B3, 0x000F71B6000F71B5,
-    0x000EB9A3000F71B7, 0x800052AE000388BE, 0x800052B6800052B2,
-    0x000E05A5000E51A6, 0x000E7DAB800052BA, 0x00000886000E75AA,
-    0x000FB5AF00036C87, 0x000E55AD000FB5AE, 0x800052BE000E55AC,
-    0x800052C6800052C2, 0x800052CA000FB5B3, 0x000F5DB6000F55B5,
-    0x800052CE800051CE, 0x000F79BA800052D2, 0x000F69BC000F79BB,
-    0x000F49BE000F5DBD, 0x000E65A8000F49BF, 0x000E71AA000E75A9,
-    0x000E91AC000E71AB, 0x000E91AE000E91AD, 0x800052D6000E91AF,
-    0x800052DE800052DA, 0x800052E6800052E2, 0x800052EE800052EA,
-    0x000E85B8800052F2, 0x000E85BA000E8DB9, 0x000E8DBC000E9DBB,
-    0x000EB5BE000EBDBD, 0x000EF1B0000179BF, 0x000EF1B2000EF1B1,
-    0x000EC1B4000EC5B3, 0x000EC1B6000EC1B5, 0x000EF9A3000EC1B7,
-    0x800052FA800052F6, 0x80005302800052FE, 0x000E19A5000E11A6,
-    0x000E35AB80005306, 0x8000530A000E35AA, 0x000E05AF8000530E,
-    0x000E11AD000E05AE, 0x00000D80000E25AC, 0x00001D8200001581,
-    0x8000531680005312, 0x000194848000531A, 0x00074086000194BE,
-    0x800053220000F487, 0x8000532A80005326, 0x800053328000532E,
-    0x00028DA880005336, 0x000295AA000295A9, 0x0002D5AC0002CDAB,
-    0x0002C9AE0002DDAD, 0x8000533A0002FDAF, 0x800053428000533E,
-    0x0000008C80005346, 0x8000534E8000534A, 0x00037DB880005352,
-    0x0003C1BA0003C1B9, 0x0003C1BC0003C1BB, 0x0003F1BE0003C9BD,
-    0x000285B00003F1BF, 0x00034DB2000345B1, 0x00035DB4000345B3,
-    0x00034DB6000345B5, 0x00021DB3000345B7, 0x8000535A80005356,
-    0x800053628000535E, 0x00025DB500025DB6, 0x000381BB80005366,
-    0x8000536A000241BA, 0x000381BF8000536E, 0x000391BD000399BE,
-    0x80005372000399BC, 0x80005376000259A3, 0x000219A68000537A,
-    0x800053828000537E, 0x000205AA000219A5, 0x800053860003C5AB,
-    0x0003DDAE8000538A, 0x0003DDAC0003C5AF, 0x800053920003D5AD,
-    0x000204840000CCE3, 0x0001C9800001BCE1, 0x0001F5820001FD81,
-    0x80005396000510BE, 0x8000539A000428BE, 0x800053A28000539E,
-    0x800053A6000010EF, 0x0004E086800053AA, 0x800053AE0002F487,
-    0x000E08E3800053B2, 0x000FE0E1800053B6, 0x000EF8EF800053BA,
-    0x800053C2800053BE, 0x800053CA800053C6, 0x800053D2800053CE,
-    0x800053DA800053D6, 0x800053E2800053DE, 0x800053EA800053E6,
-    0x800053F2800053EE, 0x001524850001E0EF, 0x800053F6000ED4E1,
-    0x00002980000E1CE3, 0x0000058200001D81, 0x000231B3800053FA,
-    0x0005CC8480005402, 0x8000540A80005406, 0x000221B5000229B6,
-    0x0001CDBB8000540E, 0x800054120001CDBA, 0x000175BF80005416,
-    0x0001C1BD0001C9BE, 0x0005E9A80001C9BC, 0x0005F9AA0005E9A9,
-    0x0005E9AC0005F9AB, 0x000639AE0005E9AD, 0x8000538E000639AF,
-    0x00008886800053FE, 0x8000541A00030087, 0x800054228000541E,
-    0x0006D1B880005426, 0x0006E1BA0006D9B9, 0x000691BC0006E1BB,
-    0x000695BE00069DBD, 0x000649B0000689BF, 0x00065DB2000649B1,
-    0x00064DB4000655B3, 0x0006F1B60006F1B5, 0x000571A30006F1B7,
-    0x8000542E8000542A, 0x8000543680005432, 0x000561A5000569A6,
-    0x00068DAB8000543A, 0x8000543E00068DAA, 0x000635AF80005442,
-    0x000681AD000689AE, 0x80005446000689AC, 0x0007E1B38000544A,
-    0x0007E1B58000544E, 0x8000545680005452, 0x8000531E0007E5B6,
-    0x0007BDBB8000545A, 0x000795BD0007A1BA, 0x000795BF0007A9BC,
-    0x000601A8000795BE, 0x000621AA000625A9, 0x000621AC000621AB,
-    0x000625AE00062DAD, 0x8000545E000655AF, 0x00001D8100001D82,
-    0x8000546200001D80, 0x8000546A80005466, 0x000639B8000130BE,
-    0x0006C9BA000639B9, 0x0006D9BC0006C9BB, 0x0006C9BE0006D9BD,
-    0x00062DB00006C9BF, 0x000635B2000631B1, 0x000619B4000609B3,
-    0x000609B6000619B5, 0x0006A9A3000609B7, 0x001F288600028C84,
-    0x8000547200014487, 0x0006A9A50006ADA6, 0x0006F5AB80005476,
-    0x8000547A0006E9AA, 0x0006DDAF8000547E, 0x0006DDAD0006DDAE,
-    0x800054820006E1AC, 0x80005486000615B3, 0x00063DB68000548A,
-    0x800054928000548E, 0x0001D9BA000635B5, 0x800054960001D9BB,
-    0x00017DBE8000549A, 0x00017DBC000165BF, 0x0005C5A8000175BD,
-    0x0005D9AA0005C9A9, 0x0005F9AC0005D1AB, 0x000229AE0005F9AD,
-    0x8000549E000229AF, 0x800054A6800054A2, 0x0000008C800054AA,
-    0x800054B2800054AE, 0x0002EDB8800054B6, 0x00028DBA000285B9,
-    0x000285BC000281BB, 0x0002BDBE00028DBD, 0x000259B000037DBF,
-    0x0002EDB2000259B1, 0x0002FDB40002E5B3, 0x0002E5B60002E5B5,
-    0x000551A30002D5B7, 0x800054BE800054BA, 0x800054C6800054C2,
-    0x000571A5000579A6, 0x00029DAB800054CA, 0x800054CE00029DAA,
-    0x000221AF800054D2, 0x000231AD000239AE, 0x00001182000239AC,
-    0x00006580800054D6, 0x800054DA00000981, 0x800054E2000398BE,
-    0x00032484800054E6, 0x00036887800054EA, 0x800054EE001C8C86,
-    0x800054F6800054F2, 0x800054FE800054FA, 0x0002A1B380005502,
-    0x0002D1B5001CC084, 0x8000550A80005506, 0x8000550E0002D5B6,
-    0x0002C1BB80005512, 0x000111BD0002F5BA, 0x000111BF0002D9BC,
-    0x80005516000119BE, 0x8000551E8000551A, 0x8000552680005522,
-    0x8000552E8000552A, 0x800055320006B8EF, 0x8000553600049CE1,
-    0x8000553A000674E3, 0x800055428000553E, 0x0000F98080005546,
-    0x000005820000F981, 0x001C58BE8000554A, 0x00005CEF001F5884,
-    0x000190E1000640EF, 0x00003CE30006FCE1, 0x80005552000674E3,
-    0x001C988680005556, 0x000269A3001CF487, 0x8000555A001F0CBE,
-    0x800055628000555E, 0x000219A500021DA6, 0x000209AB80005566,
-    0x8000556A00023DAA, 0x0001D9AF8000556E, 0x0001D9AD0001D1AE,
-    0x001D91A8000211AC, 0x001DA1AA001D99A9, 0x001DD1AC001DA1AB,
-    0x001DD5AE001DDDAD, 0x800054DE001DC9AF, 0x800055728000554E,
-    0x8000557A80005576, 0x800055828000557E, 0x001E79B880005586,
-    0x001ECDBA001E79B9, 0x001EDDBC001EC5BB, 0x001EC5BE001EC5BD,
-    0x001DB9B0001EF5BF, 0x001D85B2001D8DB1, 0x001E55B4001E4DB3,
-    0x001E55B6001E5DB5, 0x001FCDB8001E49B7, 0x001FDDBA001FD5B9,
-    0x001FF1BC001FD5BB, 0x001FE9BE001FFDBD, 0x001FA5B0001FE9BF,
-    0x001FB1B2001FB1B1, 0x001F95B4001F8DB3, 0x001FFDB6001FF5B5,
-    0x001E19A8001FF5B7, 0x001E2DAA001E19A9, 0x001E25AC001E3DAB,
-    0x001E25AE001E2DAD, 0x8000558A001FDDAF, 0x800055928000558E,
-    0x8000559A80005596, 0x8000559E8000546E, 0x001FE5B3800055A2,
-    0x800055AA800055A6, 0x800055B2800055AE, 0x001FEDB5001FEDB6,
-    0x001E35BB800055B6, 0x800055BA001E21BA, 0x001E71BF800055BE,
-    0x001E19BD001E11BE, 0x00006982001E25BC, 0x00005980001FA1A3,
-    0x001FA9A600005181, 0x800055C6800055C2, 0x001E65AA001FA9A5,
-    0x00040086001E71AB, 0x001E55AE0001EC87, 0x001E61AC001E35AF,
-    0x001E31A8001E5DAD, 0x001E41AA001E31A9, 0x001E41AC001E41AB,
-    0x001E71AE001E49AD, 0x800055CA001E71AF, 0x800055D2800055CE,
-    0x800055DA800055D6, 0x800055E2800055DE, 0x000129B8800055E6,
-    0x000139BA000129B9, 0x00012DBC000135BB, 0x0000D1BE0000D1BD,
-    0x0001C9B00000D1BF, 0x0001D9B20001C9B1, 0x0001C9B40001D9B3,
-    0x000119B60001C9B5, 0x001DC9B3000119B7, 0x800055EE800055EA,
-    0x800055F6800055F2, 0x001DC9B5001DC9B6, 0x00020DBB800055FA,
-    0x800055FE00020DBA, 0x00020DBF80005602, 0x00020DBD00020DBE,
-    0x8000560600020DBC, 0x8000560A001D8DA3, 0x001D8DA68000560E,
-    0x8000561680005612, 0x000249AA001D8DA5, 0x8000561A000249AB,
-    0x000249AE8000561E, 0x000249AC000249AF, 0x00000D80000249AD,
-    0x0000118200001181, 0x0002CCEF80005622, 0x8000562A80005626,
-    0x00022CE300029084, 0x0001D8E1000328BE, 0x0004108680005632,
-    0x800056360003D487, 0x000361B38000563A, 0x800056428000563E,
-    0x8000564A80005646, 0x000371B5000379B6, 0x00035DBB8000564E,
-    0x8000565200035DBA, 0x0000E1BF80005656, 0x0000F1BD0000F9BE,
-    0x0002A1A30000F9BC, 0x8000565E8000565A, 0x8000566680005662,
-    0x0002B1A50002B9A6, 0x00029DAB8000566A, 0x8000566E00029DAA,
-    0x000121AF80005672, 0x000131AD000139AE, 0x80005676000139AC,
-    0x001F64EF8000567A, 0x800056828000567E, 0x8000568A80005686,
-    0x00000D808000568E, 0x0000158200001181, 0x001CE0E180005692,
-    0x001F88E380005696, 0x000280848000569A, 0x0005488700058CBE,
-    0x800056A200042C86, 0x001EA4EF800056A6, 0x001E00E1001EF4EF,
-    0x001E2CE3001F50E1, 0x800056AA001EB0E3, 0x800056B2800056AE,
-    0x800056BA800056B6, 0x00048484800056BE, 0x800056C2000171B3,
-    0x000115B600011DB5, 0x800056CA800056C6, 0x000101BA800056CE,
-    0x0000F5BC000101BB, 0x0000F5BE0000FDBD, 0x0006ADA80000EDBF,
-    0x0006B5AA0006BDA9, 0x0006D9AC0006C9AB, 0x0006C9AE0006D9AD,
-    0x800056D20006C1AF, 0x800056DA800056D6, 0x800056E2800056DE,
-    0x800056EA800056E6, 0x00076DB8800056EE, 0x00070DBA000705B9,
-    0x000701BC000701BB, 0x000701BE000701BD, 0x000689B0000701BF,
-    0x00076DB2000689B1, 0x00077DB4000765B3, 0x000765B6000765B5,
-    0x8000562E000755B7, 0x800056F2000639A3, 0x00065DA68000569E,
-    0x800056FA800056F6, 0x000649AA000655A5, 0x800056FE000649AB,
-    0x0007BDAE80005702, 0x0007BDAC0007A5AF, 0x00006D800007B5AD,
-    0x0000198200000981, 0x8000570A80005706, 0x8000570E00039CBE,
-    0x0000408680005712, 0x8000571600006087, 0x8000571E8000571A,
-    0x8000572680005722, 0x000691A88000572A, 0x0006B9AA000691A9,
-    0x0006A9AC0006B9AB, 0x0006D9AE0006A9AD, 0x8000572E0006D9AF,
-    0x8000573680005732, 0x8000573E8000573A, 0x8000574680005742,
-    0x000171B88000574A, 0x000171BA000171B9, 0x0001D9BC000171BB,
-    0x0001C1BE0001C1BD, 0x0006B1B00001F5BF, 0x000689B20006B9B1,
-    0x000151B4000689B3, 0x000151B6000151B5, 0x000631B3000151B7,
-    0x800057528000574E, 0x8000575A80005756, 0x000621B5000629B6,
-    0x00067DBB8000575E, 0x80005762000675BA, 0x000195BF80005766,
-    0x00065DBD000195BE, 0x8000576A00065DBC, 0x8000576E000675A3,
-    0x00066DA680005772, 0x8000577A80005776, 0x000631AA000665A5,
-    0x0001AC84000639AB, 0x0001D1AE0001A0BE, 0x000619AC0001D1AF,
-    0x0002DDA8000619AD, 0x0002E5AA0002EDA9, 0x0002E5AC0002FDAB,
-    0x0002E5AE0002EDAD, 0x8000578200033DAF, 0x8000578A80005786,
-    0x8000578E000C64BE, 0x8000579680005792, 0x0003C9B88000579A,
-    0x0003D9BA0003C9B9, 0x0003F9BC0003D1BB, 0x000399BE0003F9BD,
-    0x000345B0000395BF, 0x000345B200034DB1, 0x000345B400035DB3,
-    0x000345B600034DB5, 0x000355810003F9B7, 0x000265B300034980,
-    0x000265B500035582, 0x800057A28000579E, 0x000C0086000265B6,
-    0x000381BB0003E487, 0x000381BD000389BA, 0x000381BF000399BC,
-    0x00022DA3000389BE, 0x800057AA800057A6, 0x800057B2800057AE,
-    0x00022DA500022DA6, 0x0003C9AB800057B6, 0x800057BA0003C1AA,
-    0x0003C9AF800057BE, 0x0003C9AD0003C1AE, 0x0006D8E30003D1AC,
-    0x00069CE10007ACE1, 0x800057C200069CE3, 0x800057C6000D5884,
-    0x800057CE800057CA, 0x800057D6800057D2, 0x00011CEF800057DA,
-    0x800057E2800057DE, 0x00007980000640EF, 0x0000118200001581,
-    0x800057E6000C0084, 0x800057EA0000ACE3, 0x800057F20001A4E1,
-    0x000CC886000070EF, 0x800057F6000DA487, 0x800057FE800057FA,
-    0x0005C9BA80005802, 0x0005A5B80005C9BB, 0x0005F9BE0005CDB9,
-    0x0005CDBC0005F9BF, 0x0005A5B20005C5BD, 0x000611B00005ADB3,
-    0x0005ADB6000611B1, 0x0005B5B400059DB7, 0x000661AA0005ADB5,
-    0x00069DA8000661AB, 0x000661AE000695A9, 0x000671AC000661AF,
-    0x80005806000671AD, 0x8000580E8000580A, 0x8000581680005812,
-    0x000CB0BE8000581A, 0x000E21A88000581E, 0x000E21AA000E21A9,
-    0x000E25AC000E3DAB, 0x000E25AE000E2DAD, 0x800057EE000E89AF,
-    0x8000582680005822, 0x8000582E8000582A, 0x8000583680005832,
-    0x000F51B88000583A, 0x000F55BA000F5DB9, 0x000F75BC000F6DBB,
-    0x000F75BE000F7DBD, 0x000EF9B0000F69BF, 0x000EA1B2000EA1B1,
-    0x000EA1B4000EA1B3, 0x000E91B6000EA9B5, 0x000EA9B3000E91B7,
-    0x8000577E8000583E, 0x8000584680005842, 0x000EADB5000EADB6,
-    0x000E65BB8000584A, 0x8000584E000E3DBA, 0x000E6DBF80005852,
-    0x000E75BD000E65BE, 0x00001982000E75BC, 0x00006580000EEDA3,
-    0x000EE9A600001981, 0x8000585A80005856, 0x000E79AA000EE9A5,
-    0x8000585E000E21AB, 0x000E21AE80005862, 0x000E31AC000E29AF,
-    0x000E61A8000E31AD, 0x000E7DAA000E75A9, 0x000E6DAC000E75AB,
-    0x000EFDAE000EF5AD, 0x00018086000EF5AF, 0x800058660001A487,
-    0x8000586E8000586A, 0x8000587680005872, 0x000175B88000587A,
-    0x000175BA00017DB9, 0x0001DDBC0001C9BB, 0x0001C5BE0001C5BD,
-    0x000E8DB00001F5BF, 0x000E9DB2000E95B1, 0x000155B4000E91B3,
-    0x000155B600015DB5, 0x000EF5B300014DB7, 0x800058828000587E,
-    0x8000588A80005886, 0x000EE5B5000E75B6, 0x000E51BB8000588E,
-    0x80005892000E49BA, 0x000199BF80005896, 0x000E45BD000191BE,
-    0x8000589A000E49BC, 0x8000589E000EB1A3, 0x000E31A6800058A2,
-    0x800058AA800058A6, 0x000E0DAA000EA1A5, 0x800058AE000E15AB,
-    0x0001D5AE800058B2, 0x000E0DAC0001DDAF, 0x0002EDA8000E01AD,
-    0x000325AA00032DA9, 0x000325AC00033DAB, 0x000325AE00032DAD,
-    0x800058B6000399AF, 0x800058BE800058BA, 0x800058C6800058C2,
-    0x000274BE800058CA, 0x00038DB8800058D2, 0x000391BA000391B9,
-    0x0003BDBC0003A5BB, 0x00007DBE000075BD, 0x0003E9B0000075BF,
-    0x0003F9B20003E9B1, 0x0003D9B40003F1B3, 0x0003B9B60003D9B5,
-    0x0000AD800003B5B7, 0x0000BD820000B581, 0x800058D60003A1B3,
-    0x0003A1B60003A1B5, 0x0002E084800058DA, 0x000321BA800058DE,
-    0x000321BC000321BB, 0x000311BE000329BD, 0x0003EDA3000311BF,
-    0x0004E886800058E2, 0x800058E600034C87, 0x0003EDA50003EDA6,
-    0x00036DAB800058EA, 0x800058EE00036DAA, 0x00035DAF800058F2,
-    0x000365AD00035DAE, 0x800058F600036DAC, 0x0023D4EF000E00E3,
-    0x800058FA000FB4E1, 0x80005902800058FE, 0x0003A9A180005906,
-    0x0003C1A3000FFDA0, 0x001284E100030DA2, 0x00F3E4E3000FE0EF,
-    0x8000590E8000590A, 0x0004048480005912, 0x80005916000404BE,
-    0x8000591A000394EF, 0x800059228000591E, 0x800059260003D4E3,
-    0x8000592A000054E1, 0x800059328000592E, 0x0000F98080005936,
-    0x000005820000F981, 0x800059428000593A, 0x001B918F80005946,
-    0x0007EC84000E9CEF, 0x8000594A000ED4E1, 0x8000594E000EF0E3,
-    0x0007188680005952, 0x0017E5990004E887, 0x000BE99B0017E598,
-    0x000AFD9D000B399A, 0x000F559F000BF19C, 0x001F2991000F5D9E,
-    0x001F4593001BF590, 0x0013D195001FCD92, 0x0017199700132994,
-    0x000005A100137596, 0x00FC01A3800058CE, 0x00FC01A5000055A2,
-    0x00F819A700FC11A4, 0x00F909A900FD45A6, 0x00F409AB00F821A8,
-    0x00F001AD00F401AA, 0x00F0F1AF00F515AC, 0x00EC01B100F01DAE,
-    0x00E801B300EC01B0, 0x00E84DB500EC11B2, 0x8000593E00E851B4,
-    0x8000595A80005956, 0x800059628000595E, 0x8000596A80005966,
-    0x800059728000596E, 0x0006BDA880005976, 0x000615AA000615A9,
-    0x000655AC000629AB, 0x000645AE000645AD, 0x0000E8BE000675AF,
-    0x8000597E8000597A, 0x8000598680005982, 0x8000598E8000598A,
-    0x000785B880005992, 0x000791BA000795B9, 0x0007B1BC000791BB,
-    0x0007A9BE0007BDBD, 0x00060DB00007A9BF, 0x0007F9B20007E5B1,
-    0x0007E9B40007F9B3, 0x0007C5B60007E9B5, 0x0015218A0007BDB7,
-    0x8000599600158D8B, 0x0011358E8000599A, 0x0011F98C000DE18F,
-    0x001D298200113D8D, 0x8000599E001D3583, 0x00199D86800059A2,
-    0x0019F1840015E987, 0x0005C99A00191D85, 0x800059A60005659B,
-    0x00001981800059AA, 0x0005B19C00007980, 0x000D419200001182,
-    0x0000C886000D4193, 0x0009FD960001A487, 0x000C55940009AD97,
-    0x00071DB30009FD95, 0x800059B6800059B2, 0x800059BE800059BA,
-    0x00075DB5000745B6, 0x0006A1BB800059C2, 0x800059C6000749BA,
-    0x0006ADBF800059CA, 0x0006ADBD0006ADBE, 0x800059CE0006ADBC,
-    0x800059D2000759A3, 0x000701A6800059D6, 0x800059DE800059DA,
-    0x00070DAA000719A5, 0x800059E20006E5AB, 0x0006E9AE800059E6,
-    0x0006E9AC0006E9AF, 0x000591A80006E9AD, 0x0005A1AA000591A9,
-    0x0005A5AC0005BDAB, 0x000279AE000249AD, 0x00FCEC84000279AF,
-    0x800059EE800059EA, 0x800059F6800059F2, 0x800059FE800059FA,
-    0x0002EDB880005A02, 0x0002FDBA0002F5B9, 0x0002EDBC0002F5BB,
-    0x000355BE00035DBD, 0x000209B0000375BF, 0x000219B2000209B1,
-    0x000209B4000219B3, 0x0002DDB6000209B5, 0x000529B30002D5B7,
-    0x80005A0600FD4884, 0x80005A0E80005A0A, 0x00052DB50002C9B6,
-    0x0002CDBB80005A12, 0x80005A160002CDBA, 0x0002CDBF80005A1A,
-    0x0002CDBD0002CDBE, 0x00006D820002CDBC, 0x0000558000056DA3,
-    0x00028DA600006581, 0x80005A2200FC1CBE, 0x000289AA000569A5,
-    0x00FCC886000289AB, 0x000289AE00031C87, 0x000289AC000289AF,
-    0x000285A8000289AD, 0x000285AA00028DA9, 0x00028DAC00029DAB,
-    0x0002B5AE0002BDAD, 0x80005A260002DDAF, 0x00FCE0BE80005A2A,
-    0x80005A3280005A2E, 0x80005A3A80005A36, 0x00016DB880005A3E,
-    0x00017DBA000175B9, 0x00016DBC000175BB, 0x0001D5BE0001DDBD,
-    0x0002A9B00001CDBF, 0x0002BDB20002A9B1, 0x000299B40002B5B3,
-    0x00015DB6000299B5, 0x000750E1000155B7, 0x0006D0E380005A42,
-    0x80005A4A80005A46, 0x80005A5280005A4E, 0x80005A5A80005A56,
-    0x80005A6280005A5E, 0x80005A6A80005A66, 0x00060CEF80005A6E,
-    0x80005A7680005A72, 0x80005A7E80005A7A, 0x80005A8680005A82,
-    0x80005A8A000664E3, 0x80005A9200067CE1, 0x00FCA0870001BCEF,
-    0x00003D8000FCCC86, 0x0000218200002581, 0x00FD8DB380005A96,
-    0x800059AE80005A1E, 0x80005A8E80005A9A, 0x00FD85B500FDBDB6,
-    0x00FDEDBB80005A9E, 0x80005AA200FDEDBA, 0x00FDDDBF80005AA6,
-    0x00FDE1BD00FDDDBE, 0x80005AAA00FDE9BC, 0x80005AAE00FDC9A3,
-    0x00FDF9A680005AB2, 0x80005ABA80005AB6, 0x00FDA9AA00FDC1A5,
-    0x80005ABE00FDA9AB, 0x00FD99AE80005AC2, 0x00FDADAC00FD99AF,
-    0x00FE59A800FDA5AD, 0x00FFADAA00FFA5A9, 0x00FFA1AC00FFA1AB,
-    0x00FFA1AE00FFA1AD, 0x80005AC600FFA1AF, 0x80005ACE80005ACA,
-    0x80005AD680005AD2, 0x80005ADE80005ADA, 0x00FF69B880005AE2,
-    0x00FF79BA00FF69B9, 0x00FF6DBC00FF79BB, 0x00FF1DBE00FF15BD,
-    0x00FF65B000FF15BF, 0x00FF79B200FF69B1, 0x00FF69B400FF79B3,
-    0x00FF59B600FF69B5, 0x00FE25B300FF59B7, 0x80005AEA80005AE6,
-    0x80005AF280005AEE, 0x00FE29B500FE21B6, 0x00FEADBB80005AF6,
-    0x80005AFA00FEA1BA, 0x00FEADBF0001E084, 0x00FEADBD00FEADBE,
-    0x0000698200FEADBC, 0x0000598000FE61A3, 0x00FE65A600005181,
-    0x80005AFE00019CBE, 0x00FEE5AA00FE6DA5, 0x0000C88600FEE9AB,
-    0x00FEE9AE0001FC87, 0x00FEE9AC00FEE9AF, 0x00FE61A800FEE9AD,
-    0x000185AA00FE79A9, 0x0001B9AC000191AB, 0x0001D9AE0001B9AD,
-    0x0000A0840001D5AF, 0x80005B0680005B02, 0x80005B0E80005B0A,
-    0x80005B1680005B12, 0x00016DB880005B1A, 0x00017DBA000175B9,
-    0x00016DBC000175BB, 0x0000D1BE0000D1BD, 0x0001ADB00000D1BF,
-    0x0001B1B20001B1B1, 0x00019DB4000185B3, 0x00015DB6000155B5,
-    0x00020DB3000155B7, 0x80005B2280005B1E, 0x80005B2A80005B26,
-    0x000231B5000239B6, 0x000275BB80005B2E, 0x80005B32000269BA,
-    0x000291BF80005B36, 0x000251BD000259BE, 0x80005B3A000265BC,
-    0x80005B3E000249A3, 0x00027DA680005B42, 0x80005B4A80005B46,
-    0x00022DAA000275A5, 0x80005B4E000231AB, 0x00021DAE80005B52,
-    0x000221AC0002D5AF, 0x80005B56000215AD, 0x80005B5E80005B5A,
-    0x00001D8080005B62, 0x0000398200000981, 0x80005B6A80005B66,
-    0x80005B72000400BE, 0x80005B7A80005B76, 0x80005B8280005B7E,
-    0x0004E08680005B86, 0x000544BE00030487, 0x80005B8A00044084,
-    0x80005B8E000164E3, 0x80005B920001ACE1, 0x80005B960002B8EF,
-    0x80005B9E80005B9A, 0x80005BA680005BA2, 0x0002A5A880005BAA,
-    0x0002A5AA0002ADA9, 0x0002A5AC0002BDAB, 0x0002A5AE0002ADAD,
-    0x0004A084000119AF, 0x80005BB280005BAE, 0x80005BBA80005BB6,
-    0x80005BC280005BBE, 0x00010DB80007B0BE, 0x000111BA000111B9,
-    0x00013DBC000125BB, 0x0001DDBE0001D5BD, 0x000169B00001D5BF,
-    0x000179B2000169B1, 0x000159B4000171B3, 0x000139B6000159B5,
-    0x00FF70E1000135B7, 0x00FF1CE300FFB4E3, 0x80005BC600FCFCE1,
-    0x80005BCA00FE60E1, 0x80005BCE0000A0E3, 0x80005BD680005BD2,
-    0x0006FCBE000000EF, 0x00FE68EF00FF98EF, 0x0000658080005BDA,
-    0x0000758200006D81, 0x80005BE2000261B3, 0x000261B6000261B5,
-    0x0004C08680005BE6, 0x0001E1BA0004A487, 0x0001E1BC0001E1BB,
-    0x0001D1BE0001E9BD, 0x80005B6E0001D1BF, 0x000605AB80005BDE,
-    0x00060DAD000605AA, 0x000635AF000605AC, 0x80005BEA000635AE,
-    0x000585A380005BEE, 0x000585A580005BF2, 0x80005BFA80005BF6,
-    0x0006F1B3000585A6, 0x80005C0280005BFE, 0x80005C0A80005C06,
-    0x0006E1B50006E9B6, 0x000689BB80005C0E, 0x80005C12000685BA,
-    0x000689BF80005C16, 0x000689BD000681BE, 0x000629A8000691BC,
-    0x000639AA000629A9, 0x000629AC000639AB, 0x000649AE000629AD,
-    0x80005C1A000645AF, 0x80005C2280005C1E, 0x80005C2A80005C26,
-    0x80005C3280005C2E, 0x0007E9B880005C36, 0x0007F9BA0007E9B9,
-    0x0007E9BC0007F9BB, 0x000741BE000745BD, 0x000601B000077DBF,
-    0x000605B200060DB1, 0x000605B400061DB3, 0x000605B600060DB5,
-    0x0007B1A30007D9B7, 0x80005C3E80005C3A, 0x80005C4680005C42,
-    0x0007A1A50007A9A6, 0x0007C9AB80005C4A, 0x80005C4E0007C5AA,
-    0x0007C9AF80005C52, 0x0007C9AD0007C1AE, 0x00004D800007D1AC,
-    0x0000558200005581, 0x80005C5600016DB3, 0x00016DB6000179B5,
-    0x0000C08680005C5A, 0x000121BA00010487, 0x000121BC000121BB,
-    0x000111BE000129BD, 0x000651A8000111BF, 0x0001A9AA000659A9,
-    0x0001BDAC0001A9AB, 0x0001A1AE0001A1AD, 0x0001AC840001D9AF,
-    0x80005C6280005C5E, 0x80005C6A80005C66, 0x80005C7280005C6E,
-    0x000169B880005C76, 0x000179BA000169B9, 0x000169BC000179BB,
-    0x0000DDBE000169BD, 0x0001A9B00000D5BF, 0x0001B9B20001A9B1,
-    0x000199B40001B1B3, 0x000159B6000199B5, 0x000221A3000159B7,
-    0x80005C7E80005C7A, 0x80005C8680005C82, 0x000235A5000221A6,
-    0x00026DAB80005C8A, 0x80005C8E00026DAA, 0x00025DAF80005C92,
-    0x000265AD00025DAE, 0x80005C9600026DAC, 0x80005C9E80005C9A,
-    0x80005CA680005CA2, 0x80005CAE80005CAA, 0x00003D8080005CB2,
-    0x0000198200000981, 0x80005CBA80005CB6, 0x0003448780005CC2,
-    0x80005CC6000CDC86, 0x80005CCE80005CCA, 0x80005CD2000CC084,
-    0x80005CD60007DCEF, 0x80005CDE80005CDA, 0x80005CE2000114E3,
-    0x80005CE60001D4E1, 0x80005CEE80005CEA, 0x00039DB380005CF2,
-    0x80005CFA80005CF6, 0x80005D0280005CFE, 0x000371B5000379B6,
-    0x00035DBB80005D06, 0x80005D0A00035DBA, 0x0000E1BF80005D0E,
-    0x0000F1BD0000F9BE, 0x00022DA80000F9BC, 0x000255AA000255A9,
-    0x0002B9AC0002A9AB, 0x0002A9AE0002B9AD, 0x000D68BE0002A9AF,
-    0x80005D1680005D12, 0x80005D1E80005D1A, 0x0000AD810000AD82,
-    0x00016DB80000AD80, 0x00010DBA000105B9, 0x00011DBC000105BB,
-    0x00010DBE000105BD, 0x0002D9B0000105BF, 0x00016DB20002D9B1,
-    0x00017DB4000165B3, 0x000165B6000165B5, 0x000678E1000155B7,
-    0x0006B8E300075CE1, 0x80005D22000750E3, 0x80005D2A80005D26,
-    0x80005D3680005D32, 0x80005D3E80005D3A, 0x000CF8BE80005D42,
-    0x00067CEF80005D46, 0x00029DA300078CEF, 0x000C688680005D4A,
-    0x80005D4E000D0487, 0x000271A5000279A6, 0x00025DAB80005D52,
-    0x80005D5600025DAA, 0x0001E1AF80005D5A, 0x0001F1AD0001F9AE,
-    0x000E75A80001F9AC, 0x000E8DAA000E85A9, 0x000E81AC000E85AB,
-    0x000E85AE000E8DAD, 0x80005D2E000EBDAF, 0x80005D6280005D5E,
-    0x80005D6680005CBE, 0x80005D6E80005D6A, 0x000E91B880005D72,
-    0x000EA1BA000E99B9, 0x000F75BC000EA1BB, 0x000F75BE000F7DBD,
-    0x000EC9B0000F6DBF, 0x000EDDB2000ED5B1, 0x000ECDB4000ED5B3,
-    0x000EB1B6000EB1B5, 0x000E05B3000EB1B7, 0x80005D7A80005D76,
-    0x80005D8280005D7E, 0x000E09B5000E0DB6, 0x000E11BB80005D86,
-    0x80005D8A000E0DBA, 0x000E75BF80005D8E, 0x000E01BD000E09BE,
-    0x00002982000E09BC, 0x00001D80000E41A3, 0x000E49A600001581,
-    0x80005D9680005D92, 0x000E49AA000E4DA5, 0x80005D9A000E55AB,
-    0x000E4DAE0003A084, 0x000E4DAC000E31AF, 0x000E9DA8000E45AD,
-    0x000EADAA000EA5A9, 0x000EBDAC000EA5AB, 0x000EDDAE000EA1AD,
-    0x80005D9E000ECDAF, 0x0001888680005DA2, 0x80005DA60001AC87,
-    0x80005DAE80005DAA, 0x00017DB880005DB2, 0x0001C1BA0001C1B9,
-    0x0001C1BC0001C1BB, 0x0001F1BE0001C9BD, 0x000EBDB00001F1BF,
-    0x000E85B2000E8DB1, 0x000159B4000149B3, 0x00014DB6000145B5,
-    0x000E25B3000145B7, 0x80005DBA80005DB6, 0x80005DC280005DBE,
-    0x000E35B5000E21B6, 0x000E49BB80005DC6, 0x80005DCA000E79BA,
-    0x000181BF80005DCE, 0x00019DBD00019DBE, 0x80005DD2000E51BC,
-    0x80005DD6000E61A3, 0x000E65A680005DDA, 0x80005DE280005DDE,
-    0x000E3DAA000E71A5, 0x80005DE6000E0DAB, 0x0001D9AE80005DEA,
-    0x000E15AC0001C5AF, 0x000325A80001D9AD, 0x00033DAA000335A9,
-    0x00032DAC000335AB, 0x000391AE000391AD, 0x80005DEE000391AF,
-    0x80005DF680005DF2, 0x80005DFE80005DFA, 0x80005E0680005E02,
-    0x000399B880005E0A, 0x0003A5BA0003ADB9, 0x000075BC00006DBB,
-    0x000075BE00007DBD, 0x0003F1B000006DBF, 0x0003C1B20003F9B1,
-    0x0003B1B40003C1B3, 0x0003B5B60003BDB5, 0x0000B5800003A9B7,
-    0x0000B5820000BD81, 0x80005E1200030CBE, 0x80005E1A80005E16,
-    0x0002AC840002E8EF, 0x80005E1E000194E1, 0x80005E2200021CE3,
-    0x0003D88780005E26, 0x0003F5B300041C86, 0x001FA48580005E2A,
-    0x80005E3280005E2E, 0x0003E5B5000351B6, 0x00037DBB80005E36,
-    0x80005E3A000375BA, 0x000339BF80005E3E, 0x000355BD000335BE,
-    0x80005E42000355BC, 0x80005E4A80005E46, 0x80005E4E000239A3,
-    0x00029DA6000229A5, 0x80005E5680005E52, 0x0002B9AA80005E5A,
-    0x000299AC0002B1AB, 0x0002F9AE000299AD, 0x80005E5E0002F5AF,
-    0x0002448480005E62, 0x80005E66000540BE, 0x80005E6E80005E6A,
-    0x00003D8080005E72, 0x0000198200000981, 0x80005E7A80005E76,
-    0x0005F48780005E82, 0x80005E860004DC86, 0x80005E8E80005E8A,
-    0x000E48EF80005E92, 0x80005E9A80005E96, 0x000F3CE3000ED4EF,
-    0x000FCCE1000E74E1, 0x80005E9E000150E3, 0x80005EA680005EA2,
-    0x000159B380005EAA, 0x80005EB280005EAE, 0x80005EBA80005EB6,
-    0x000169B5000161B6, 0x00013DBB80005EBE, 0x80005EC2000135BA,
-    0x0000F5BF80005EC6, 0x000115BD0000F5BE, 0x000691A8000115BC,
-    0x000691AA000691A9, 0x0006A1AC0006A5AB, 0x0006DDAE0006A1AD,
-    0x80005E7E0006D1AF, 0x80005ECE80005ECA, 0x80005E0E80005ED2,
-    0x80005EDA80005ED6, 0x00075DB880005EDE, 0x00076DBA000765B9,
-    0x000715BC00070DBB, 0x00070DBE000705BD, 0x0006B1B0000705BF,
-    0x000681B20006B9B1, 0x000771B4000681B3, 0x000771B6000771B5,
-    0x000619A300076DB7, 0x0000158100002D82, 0x80005EE200001D80,
-    0x000629A5000621A6, 0x00067DAB80005EE6, 0x80005EEA000675AA,
-    0x0007B5AF0002A484, 0x000655AD0007B5AE, 0x0003CCBE000655AC,
-    0x80005EEE000741B3, 0x000745B680005EF2, 0x80005EFA80005EF6,
-    0x000755BA000751B5, 0x0000488600075DBB, 0x000739BE00030C87,
-    0x000745BC000739BF, 0x000675A8000739BD, 0x00068DAA000685A9,
-    0x00069DAC000685AB, 0x000681AE000681AD, 0x80005EFE0006ADAF,
-    0x80005F0680005F02, 0x80005F0E80005F0A, 0x80005F1680005F12,
-    0x0006A9B880005F1A, 0x00016DBA0006A9B9, 0x00017DBC000165BB,
-    0x000161BE000161BD, 0x0006D5B0000161BF, 0x0006D5B20006DDB1,
-    0x0006B9B40006A9B3, 0x0006A9B60006B9B5, 0x000605A30006A1B7,
-    0x80005F2280005F1E, 0x80005F2A80005F26, 0x000615A5000601A6,
-    0x000619AB80005F2E, 0x80005F32000611AA, 0x00067DAF80005F36,
-    0x00067DAD00067DAE, 0x80005F3A000601AC, 0x80005F4280005F3E,
-    0x80005F4A80005F46, 0x80005F5280005F4E, 0x0000398080005F56,
-    0x0000058200003981, 0x0001B8BE80005F5A, 0x80005F620001B884,
-    0x0002CDA880005F66, 0x0002D1AA0002D1A9, 0x000335AC00032DAB,
-    0x000335AE00033DAD, 0x80005F6A00032DAF, 0x80005F7280005F6E,
-    0x80005F7A80005F76, 0x001C34BE80005F7E, 0x0003E9B880005F82,
-    0x000389BA0003E9B9, 0x000399BC000389BB, 0x000389BE000399BD,
-    0x000355B0000389BF, 0x000355B200035DB1, 0x0003F9B40003E9B3,
-    0x0003E9B60003F9B5, 0x000259B30003E1B7, 0x0003948780005F86,
-    0x80005F8A001C4C86, 0x0003BDB50003BDB6, 0x000399BB80005F8E,
-    0x80005F92000391BA, 0x00033DBF80005F96, 0x00033DBD00033DBE,
-    0x80005F9A000381BC, 0x80005F9E00021DA3, 0x0003F9A680005FA2,
-    0x80005FAA80005FA6, 0x0003D5AA0003F9A5, 0x80005FAE0003DDAB,
-    0x000379AE80005FB2, 0x0003C5AC000379AF, 0x80005FB6000379AD,
-    0x0002048480005FBA, 0x80005FBE001C00BE, 0x00003D8100003D82,
-    0x80005FC200003D80, 0x80005FCA80005FC6, 0x80005FD680005FD2,
-    0x80005FDE80005FDA, 0x80005FE680005FE2, 0x80005FEA0000A0E3,
-    0x001CF0860001BCE1, 0x80005FEE001D0487, 0x80005FF680005FF2,
-    0x80005FFE80005FFA, 0x8000600680006002, 0x8000600A00007CEF,
-    0x800060128000600E, 0x8000601A80006016, 0x8000601E000638EF,
-    0x8000602680006022, 0x001DA8840001A4EF, 0x8000602A0006C0E1,
-    0x8000602E000110E3, 0x80006032000710E1, 0x0001E1B30006E4E3,
-    0x8000603A80006036, 0x8000604680006042, 0x0001F1B50001F9B6,
-    0x000165BB8000604A, 0x001F48860001D9BA, 0x00015DBF001CD087,
-    0x000165BD00015DBE, 0x001E3DB000016DBC, 0x001E99B2001E81B1,
-    0x001EB9B4001E91B3, 0x001ED9B6001EB9B5, 0x001EEDB8001ED5B7,
-    0x001EF1BA001EF1B9, 0x001EDDBC001EC5BB, 0x001F3DBE001F35BD,
-    0x80005FCE001F2DBF, 0x8000604E80005F5E, 0x001FE98080006052,
-    0x001F9182001F9181, 0x001D99A88000603E, 0x001E49AA001D99A9,
-    0x001E59AC001E49AB, 0x001E49AE001E59AD, 0x001EA1A3001E49AF,
-    0x8000605A80006056, 0x800060628000605E, 0x001EB1A5001EB9A6,
-    0x001E25AB80006066, 0x8000606A001E99AA, 0x001E1DAF8000606E,
-    0x001E25AD001E1DAE, 0x80006072001E2DAC, 0x80006076001FF1B3,
-    0x001F91B68000607A, 0x800060828000607E, 0x001FBDBA001FE1B5,
-    0x80006086001F85BB, 0x001F9DBE8000608A, 0x001F9DBC001F85BF,
-    0x001E2DA8001F95BD, 0x001E3DAA001E35A9, 0x001E2DAC001E35AB,
-    0x001E51AE001E55AD, 0x8000608E001E45AF, 0x8000609680006092,
-    0x8000609E8000609A, 0x0000B1810000B982, 0x001EC1B80000B980,
-    0x001EC1BA001EC1B9, 0x001EC5BC001EDDBB, 0x001EF1BE001ECDBD,
-    0x001E3DB0001EEDBF, 0x001E0DB2001E05B1, 0x001E1DB4001E05B3,
-    0x001E01B6001E05B5, 0x001EB5A3001E01B7, 0x800060A6800060A2,
-    0x00029884800060AA, 0x001EA5A5001ED5A6, 0x001EC1AB000190BE,
-    0x00000886001EF9AA, 0x001EC1AF00017C87, 0x001ED1AD001ED9AE,
-    0x800060B2001ED9AC, 0x800060B6001E19B3, 0x000155B6800060BA,
-    0x800060C2800060BE, 0x000171BA000191B5, 0x800060C6000179BB,
-    0x000131BE800060CA, 0x000151BC00013DBF, 0x001DCDA8000151BD,
-    0x001DD5AA001DD5A9, 0x00025DAC000245AB, 0x000291AE000291AD,
-    0x800060CE000291AF, 0x800060D6800060D2, 0x800060DE800060DA,
-    0x800060E6800060E2, 0x000299B8800060EA, 0x0002A5BA0002ADB9,
-    0x000375BC00036DBB, 0x000375BE00037DBD, 0x0002F1B000036DBF,
-    0x0002C1B20002F9B1, 0x0002B1B40002C1B3, 0x0002B5B60002BDB5,
-    0x001D5DA30002A9B7, 0x800060F2800060EE, 0x800060FA800060F6,
-    0x0002D5A5000211A6, 0x00023DAB800060FE, 0x80006102000235AA,
-    0x000279AF80006106, 0x000215AD000275AE, 0x00044CBE000215AC,
-    0x8000610E8000610A, 0x00001D8080006112, 0x0000398200000981,
-    0x8000611A80006116, 0x0004C8860003A884, 0x800061220004A887,
-    0x8000612A80006126, 0x0002C1A88000612E, 0x0002C5AA0002CDA9,
-    0x0002C9AC0002D9AB, 0x0002F5AE0002FDAD, 0x8000613200013DAF,
-    0x8000613A80006136, 0x800061428000613E, 0x8000614A80006146,
-    0x0001C9B80004F4BE, 0x0001D9BA0001C9B9, 0x0001F9BC0001D1BB,
-    0x00019DBE0001F9BD, 0x000145B0000195BF, 0x000145B200014DB1,
-    0x000145B400015DB3, 0x000145B600014DB5, 0x00024DB30001F9B7,
-    0x800061528000614E, 0x8000615A80006156, 0x000241B500027DB6,
-    0x000189BB8000615E, 0x80006162000181BA, 0x000189BF80006166,
-    0x000189BD000181BE, 0x001F54E1000191BC, 0x001ED0E3001E8CE1,
-    0x000050E3001E7CE3, 0x001E28E18000616A, 0x000084EF8000616E,
-    0x8000617680006172, 0x8000617E8000617A, 0x001F38EF80006182,
-    0x0005CDA3001E08EF, 0x800061868000611E, 0x0003DC848000618A,
-    0x0005C1A50005FDA6, 0x000609AB8000618E, 0x00030886000601AA,
-    0x000609AF00036487, 0x000609AD000601AE, 0x00004981000611AC,
-    0x800060AE0007B180, 0x8000619200004182, 0x8000619A80006196,
-    0x800061A28000619E, 0x800061AA800061A6, 0x800061B2800061AE,
-    0x800061BA800061B6, 0x000671A8800061BE, 0x000671AA000671A9,
-    0x0007B1AC000671AB, 0x0007B1AE0007B1AD, 0x800061C20007B1AF,
-    0x800061CA800061C6, 0x800061D2800061CE, 0x800061DA800061D6,
-    0x000769B8800061DE, 0x00071DBA000769B9, 0x000735BC000711BB,
-    0x000735BE00073DBD, 0x0007D1B000072DBF, 0x0007D1B20007D1B1,
-    0x000779B4000769B3, 0x000769B6000779B5, 0x0007F9B3000761B7,
-    0x800061E6800061E2, 0x800061EE800061EA, 0x000639B5000631B6,
-    0x000615BB800061F2, 0x800061F6000615BA, 0x0006BDBF800061FA,
-    0x0006BDBD0006BDBE, 0x00006D82000605BC, 0x000055800007BDA3,
-    0x000675A600006581, 0x000190BE00019884, 0x000651AA00067DA5,
-    0x000C0085000651AB, 0x0006F9AE000F60BE, 0x000641AC0006F9AF,
-    0x0001D1A80006F9AD, 0x0001D1AA0001D1A9, 0x0001F1AC0001D1AB,
-    0x0001F1AE0001F1AD, 0x000F20860001F1AF, 0x800062020000E487,
-    0x8000620A80006206, 0x800062128000620E, 0x000171B880006216,
-    0x000171BA000171B9, 0x0000D5BC000171BB, 0x0000D5BE0000DDBD,
-    0x000191B00000CDBF, 0x000191B2000191B1, 0x000151B4000191B3,
-    0x000151B6000151B5, 0x000259B3000151B7, 0x8000621E8000621A,
-    0x8000622680006222, 0x000279B5000271B6, 0x000255BB8000622A,
-    0x8000622E000255BA, 0x000295BF80006232, 0x000245BD000295BE,
-    0x80006236000245BC, 0x8000623A00021DA3, 0x000235A68000623E,
-    0x8000624680006242, 0x000211AA00023DA5, 0x8000624A000211AB,
-    0x0002D1AE8000624E, 0x000201AC0002D1AF, 0x00EA8582000201AD,
-    0x8000625200EADD83, 0x0016818600030C84, 0x001691840017CD87,
-    0x0012A18A00169985, 0x800062560012A18B, 0x001E918E8000625A,
-    0x0013918C001F958F, 0x001AF192001E918D, 0x8000625E001AE593,
-    0x00069196002458EF, 0x001A119400068997, 0x0002A99A00069995,
-    0x800062620002E19B, 0x8000626A80006266, 0x0002E59C000188E1,
-    0x001B319B003728E3, 0x001F2999001E259A, 0x0007519F001F2198,
-    0x001BD19D0007519E, 0x00173193001B3D9C, 0x0017219100172192,
-    0x0012199700EB7590, 0x0013399500133196, 0x0001A98000166594,
-    0x0001B9820001A981, 0x00EB558F0002FDB3, 0x0002E5B60002EDB5,
-    0x000CC08680006272, 0x000165BA000DE887, 0x00016DBC00017DBB,
-    0x000155BE00015DBD, 0x000770E100013DBF, 0x000188E380006276,
-    0x0000D4E38000627A, 0x0001A8E1000D9C84, 0x000000EF8000627E,
-    0x8000628680006282, 0x8000628E8000628A, 0x00067CEF80006292,
-    0x8000629A80006296, 0x000CC4BE000730E3, 0x8000629E00047CE1,
-    0x800062A6800062A2, 0x0003FDA3800062AA, 0x0003A1A10003E1A2,
-    0x800062AE0007B1A0, 0x800062B20007CCEF, 0x800062BA800062B6,
-    0x800061FE800062BE, 0x800062C28000626E, 0x800062CA800062C6,
-    0x000D89A8800062CE, 0x000D99AA000D89A9, 0x000D89AC000D99AB,
-    0x000E3DAE000D89AD, 0x000E4DB0000E35AF, 0x000E55B2000E55B1,
-    0x000E7DB4000E65B3, 0x000E11B6000E15B5, 0x000E35B8000E11B7,
-    0x000E01BA000E3DB9, 0x000EF9BC000E01BB, 0x000ED5BE000EDDBD,
-    0x000D31A3000EB5BF, 0x0000158100002D82, 0x800062D200001D80,
-    0x000D21A5000D29A6, 0x000EB1AB800062D6, 0x800062DA000EA9AA,
-    0x000EF1AF800062DE, 0x000E91AD000E99AE, 0x800062E2000EA1AC,
-    0x00006886000F49B3, 0x000F79B600031C87, 0x800062EA800062E6,
-    0x000F55BA000F79B5, 0x800062EE000F59BB, 0x000F25BE800062F2,
-    0x000F39BC000F2DBF, 0x000E5DA8000F2DBD, 0x000E85AA000E81A9,
-    0x000EB5AC000E91AB, 0x000EB5AE000EBDAD, 0x800062F6000EADAF,
-    0x800062FE800062FA, 0x8000630680006302, 0x8000630E8000630A,
-    0x000EB9B880006312, 0x000149BA000EB9B9, 0x000141BC000149BB,
-    0x000141BE000141BD, 0x000ED5B0000141BF, 0x000EA9B2000ED9B1,
-    0x000EA5B4000EBDB3, 0x000EA5B6000EADB5, 0x000E0DA3000E8DB7,
-    0x8000631A80006316, 0x800063228000631E, 0x000E3DA5000E3DA6,
-    0x000E1DAB80006326, 0x8000632A000E11AA, 0x000E69AF8000632E,
-    0x000E69AD000E61AE, 0x80006332000E7DAC, 0x8000633A80006336,
-    0x00000D808000633E, 0x00003D8200003581, 0x8000634680006342,
-    0x0001A4848000634A, 0x0007D0860001A4BE, 0x800063520000F887,
-    0x0002C1A880006356, 0x0002D1AA0002D1A9, 0x0002FDAC0002E5AB,
-    0x00033DAE000335AD, 0x8000635A000335AF, 0x800063628000635E,
-    0x8000636A80006366, 0x800063728000636E, 0x0003D1B880006376,
-    0x0003E1BA0003D9B9, 0x000391BC0003E1BB, 0x000391BE000391BD,
-    0x00034DB0000391BF, 0x00035DB2000355B1, 0x00034DB4000355B3,
-    0x0003F1B60003F1B5, 0x000259B30003F1B7, 0x8000637E8000637A,
-    0x80006382000400BE, 0x000275B50003B5B6, 0x000391BB80006386,
-    0x8000638A000389BA, 0x000335BF8000638E, 0x000381BD000389BE,
-    0x80006392000389BC, 0x8000639600021DA3, 0x0003F1A68000639A,
-    0x800063A28000639E, 0x0003CDAA000231A5, 0x800063A60003D5AB,
-    0x0003CDAE800063AA, 0x0003CDAC000371AF, 0x800063AE0003C5AD,
-    0x0000158100002182, 0x000004EF00001580, 0x800063B200029484,
-    0x0000D8E3800063B6, 0x0001F8E1800063BA, 0x800063C6800063C2,
-    0x800063CE800063CA, 0x0004E086800063D2, 0x800063D600051887,
-    0x800063DA00016DB3, 0x000175B600017DB5, 0x800063E2800063DE,
-    0x000151BA800063E6, 0x0001F5BC000151BB, 0x0001E9BE0001FDBD,
-    0x000E04E30001E1BF, 0x000EF0E1000F18E3, 0x800063EA000C34E1,
-    0x800063F2800063EE, 0x0037ADB2800063F6, 0x0032B9B0003719B3,
-    0x000E80EF003401B1, 0x002801B4000C20EF, 0x003C01AA002BBDB5,
-    0x003BCDA8003FC1AB, 0x0033D5AE003BA1A9, 0x003FC9AC0033DDAF,
-    0x0003EDA2003001AD, 0x0004A0BE000349A3, 0x000781A60003E5A1,
-    0x000401A4003801A7, 0x000621A30007F9A5, 0x800063FA800063BE,
-    0x80006402800063FE, 0x000631A5000639A6, 0x00061DAB80006406,
-    0x8000640A00061DAA, 0x0006ADAF8000640E, 0x0006B1AD0006A5AE,
-    0x800064120006B9AC, 0x80006416000430EF, 0x8000641E8000641A,
-    0x00003D8100003D82, 0x8000642200003D80, 0x8000642A80006426,
-    0x8000642E00032084, 0x80006432000570E3, 0x000579A8000794E1,
-    0x0006A9AA000579A9, 0x0006B9AC0006A9AB, 0x0006A9AE0006B9AD,
-    0x000000860006A9AF, 0x8000643600008487, 0x8000643E8000643A,
-    0x8000644680006442, 0x000601B88000644A, 0x000605BA00060DB9,
-    0x000609BC000619BB, 0x000635BE00063DBD, 0x0006D9B00007DDBF,
-    0x00066DB20006D9B1, 0x000661B4000665B3, 0x000651B6000669B5,
-    0x00062DA8000651B7, 0x00063DAA000635A9, 0x00062DAC000635AB,
-    0x000649AE00065DAD, 0x8000644E000641AF, 0x8000645680006452,
-    0x8000645E8000645A, 0x8000646680006462, 0x000151B88000646A,
-    0x000161BA000159B9, 0x000111BC000161BB, 0x000115BE00011DBD,
-    0x000601B000010DBF, 0x0006F9B2000601B1, 0x0006E9B40006F9B3,
-    0x000171B60006E9B5, 0x0006B5B3000171B7, 0x800064728000646E,
-    0x8000647A80006476, 0x0006A5B50006DDB6, 0x0006C5BB8000647E,
-    0x800064820006C5BA, 0x00062DBF80006486, 0x00062DBD00062DBE,
-    0x8000648A00062DBC, 0x8000648E0006F1A3, 0x000699A680006492,
-    0x8000649A80006496, 0x000681AA0006E1A5, 0x0001B4BE000681AB,
-    0x000669AE0001A884, 0x000669AC000669AF, 0x000279A8000669AD,
-    0x0002A5AA0002ADA9, 0x0002A5AC0002BDAB, 0x0002A5AE0002ADAD,
-    0x0003ED8000033DAF, 0x0003F1820003F181, 0x800064A6800064A2,
-    0x0003AC87800064AA, 0x0003C9B800039C86, 0x0003D9BA0003C9B9,
-    0x0003F9BC0003D1BB, 0x000399BE0003F9BD, 0x000345B0000395BF,
-    0x000345B200034DB1, 0x000345B400035DB3, 0x000345B600034DB5,
-    0x800064AE0003F9B7, 0x000225B3800064B2, 0x000225B5003C4084,
-    0x800064BA800064B6, 0x800064BE000225B6, 0x000381BB800064C2,
-    0x000381BD000389BA, 0x000381BF000399BC, 0x800064C6000389BE,
-    0x00027DA500027DA6, 0x00027DA3800064CA, 0x800064D2800064CE,
-    0x0003D9AF800064D6, 0x0003D9AD0003D1AE, 0x0003D9AB0003C1AC,
-    0x800064DA0003D1AA, 0x00012DA8800064DE, 0x00013DAA000135A9,
-    0x000195AC00018DAB, 0x000181AE000181AD, 0x800064E20001BDAF,
-    0x800064EA800064E6, 0x800064F2800064EE, 0x800064FA800064F6,
-    0x000061B8800064FE, 0x000061BA000061B9, 0x000061BC000061BB,
-    0x000061BE000061BD, 0x0001C5B0000061BF, 0x0001A5B20001A5B1,
-    0x0001A9B40001B9B3, 0x000181B60001A9B5, 0x0006D8E1000181B7,
-    0x0001F4E380006502, 0x00001D8180006506, 0x8000650A00001180,
-    0x003DACBE00000582, 0x003C088680006512, 0x80006516003DFC87,
-    0x0006A8EF8000651A, 0x800065228000651E, 0x8000652A80006526,
-    0x8000652E0000DCEF, 0x8000653680006532, 0x8000653E8000653A,
-    0x8000654680006542, 0x8000654E8000654A, 0x8000634E00014CE1,
-    0x003E35B3000000E3, 0x800065528000650E, 0x8000655A80006556,
-    0x003E9DB5003E81B6, 0x003EE1BB8000655E, 0x80006562003E81BA,
-    0x003EEDBF80006566, 0x003EEDBD003EEDBE, 0x8000656A003EF1BC,
-    0x003D10EF8000656E, 0x00001D8080006572, 0x0000398200000981,
-    0x8000657A80006576, 0x000364BE00026C84, 0x003C54E18000657E,
-    0x003CE4E380006582, 0x003EB5A380006586, 0x000348878000658A,
-    0x8000658E00002C86, 0x003E1DA5003E01A6, 0x003E61AB80006592,
-    0x80006596003E01AA, 0x003E6DAF8000659A, 0x003E6DAD003E6DAE,
-    0x8000659E003E71AC, 0x800065A2003F71B3, 0x003F15B6800065A6,
-    0x800065AE800065AA, 0x003F35BA003F1DB5, 0x800065B2003F3DBB,
-    0x003F15BE800065B6, 0x003F25BC003F1DBF, 0x003EA5A8003F1DBD,
-    0x003EBDAA003EB5A9, 0x003ED1AC003EB1AB, 0x003ED1AE003ED1AD,
-    0x800065BA003ECDAF, 0x800065C2800065BE, 0x800065CA800065C6,
-    0x800065D2800065CE, 0x000149B8800065D6, 0x000159BA000149B9,
-    0x000179BC000151BB, 0x000119BE000179BD, 0x003EB5B0000119BF,
-    0x003EB5B2003EBDB1, 0x003E95B4003E8DB3, 0x003E95B6003E9DB5,
-    0x003E35A3003E8DB7, 0x800065DE800065DA, 0x800065E6800065E2,
-    0x003E59A5003E51A6, 0x003E79AB800065EA, 0x0002A884003E71AA,
-    0x003E59AF0001ACBE, 0x003E59AD003E51AE, 0x00004D80003E61AC,
-    0x0000558200005581, 0x800065F2000149B3, 0x000171B6000149B5,
-    0x00004086800065F6, 0x000121BA00011487, 0x000125BC000121BB,
-    0x000111BE00012DBD, 0x000275A8000111BF, 0x00028DAA000279A9,
-    0x00028DAC00029DAB, 0x0002B5AE0002BDAD, 0x800065FA0002DDAF,
-    0x80006602800065FE, 0x8000660A80006606, 0x800066128000660E,
-    0x000369B880006616, 0x000379BA000369B9, 0x000369BC000379BB,
-    0x0003DDBE000369BD, 0x0002A9B00003D5BF, 0x0002BDB20002A9B1,
-    0x000299B40002B5B3, 0x000359B6000299B5, 0x8000661A000359B7,
-    0x800066228000661E, 0x00045C84000209A3, 0x000231A6000209A5,
-    0x8000662A80006626, 0x000261AA8000662E, 0x000265AC000261AB,
-    0x000251AE00026DAD, 0x80006632000251AF, 0x8000663A80006636,
-    0x8000663E0004C0BE, 0x8000664680006642, 0x8000664E8000664A,
-    0x8000665680006652, 0x000015808000665A, 0x0000158200001D81,
-    0x0002D1A880006662, 0x000129AA0002D9A9, 0x000139AC000129AB,
-    0x000129AE000139AD, 0x80006666000129AF, 0x8000666E8000666A,
-    0x8000667680006672, 0x8000667E8000667A, 0x000181B880006682,
-    0x000181BA000181B9, 0x000185BC00019DBB, 0x000185BE00018DBD,
-    0x000159B00001BDBF, 0x0001E9B2000159B1, 0x0001F9B40001E9B3,
-    0x0001D9B60001E9B5, 0x003E30E10001D9B7, 0x0001FCE380006686,
-    0x8000668E8000668A, 0x0004388680006692, 0x0005A88400026487,
-    0x8000669A80006696, 0x800066A28000669E, 0x000128EF800066A6,
-    0x0013FC85800066AA, 0x800066B2800066AE, 0x000084EF800066B6,
-    0x800066BE800066BA, 0x000058E3800066C2, 0x00017CE1800066C6,
-    0x00001D82800066CA, 0x000025808000649E, 0x000621B300001981,
-    0x000304BE8000665E, 0x800066CE800066D2, 0x000631B5000639B6,
-    0x00061DBB800066D6, 0x0000888600061DBA, 0x0006ADBF00030C87,
-    0x0006BDBD0006BDBE, 0x800066DA0006BDBC, 0x800066DE0007BCE1,
-    0x800066E20004CCE3, 0x800066EA800066E6, 0x800066F2800066EE,
-    0x800066FA800066F6, 0x80006702800066FE, 0x8000670A80006706,
-    0x0006A1A3000448EF, 0x800067128000670E, 0x8000671A80006716,
-    0x0006B1A50006B9A6, 0x00069DAB8000671E, 0x8000672200069DAA,
-    0x00062DAF80006726, 0x00063DAD00063DAE, 0x8000672A00063DAC,
-    0x8000672E0007E1B3, 0x0007E1B680006732, 0x8000673A80006736,
-    0x0007DDBA0007F1B5, 0x0003E084000761BB, 0x000765BE0000E4BE,
-    0x000775BC000751BF, 0x0007FDA800076DBD, 0x00062DAA000625A9,
-    0x00063DAC000625AB, 0x00062DAE000625AD, 0x0001ED80000625AF,
-    0x0001F1820001F181, 0x0000908680006742, 0x800067460001F087,
-    0x0006F1B88000674A, 0x0006F5BA0006F1B9, 0x000699BC000689BB,
-    0x00068DBE000699BD, 0x00065DB0000685BF, 0x0006E5B20006E1B1,
-    0x0006E5B40006FDB3, 0x0006D1B60006D1B5, 0x0006A5A30006D1B7,
-    0x800067528000674E, 0x8000675A80006756, 0x0006B5A50006A5A6,
-    0x000625AB8000675E, 0x80006762000699AA, 0x000615AF80006766,
-    0x000629AD000621AE, 0x8000676A000631AC, 0x8000676E000185B3,
-    0x000185B680006772, 0x8000677A80006776, 0x000181BA00018DB5,
-    0x8000677E000181BB, 0x000181BE80006782, 0x000191BC000181BF,
-    0x00020DA8000191BD, 0x000249AA000219A9, 0x000241AC000245AB,
-    0x000241AE000241AD, 0x80006786000241AF, 0x8000678E8000678A,
-    0x0003388480006792, 0x8000679A80006796, 0x0002C9B88000679E,
-    0x0002D9BA0002C9B9, 0x0002F9BC0002D1BB, 0x000299BE0002F9BD,
-    0x000201B0000295BF, 0x000201B2000201B1, 0x000201B4000201B3,
-    0x000201B6000201B5, 0x800067A20002F9B7, 0x800067A60002C5A3,
-    0x0002C5A6800067AA, 0x800067B6800067B2, 0x0002C1AA0002CDA5,
-    0x000C88860002C1AB, 0x0002C1AE00031887, 0x0002D1AC0002C1AF,
-    0x000049810002D1AD, 0x000C44BE00004580, 0x800067BA00004182,
-    0x800067C2800067BE, 0x800067CA800067C6, 0x800067D2800067CE,
-    0x000D3C84800067D6, 0x000D30BE800067DA, 0x00022DA8800067DE,
-    0x000235AA00023DA9, 0x000299AC000289AB, 0x000289AE000299AD,
-    0x000FEC84000281AF, 0x000FE0BE800067E2, 0x800067EA800067E6,
-    0x800067F2800067EE, 0x0002ADB8800067F6, 0x00016DBA000165B9,
-    0x00017DBC000165BB, 0x000161BE000161BD, 0x0002C9B0000161BF,
-    0x0002A9B20002C9B1, 0x0002BDB40002A5B3, 0x0002A1B60002A1B5,
-    0x000600E1000295B7, 0x000638E3800067FA, 0x80006802800067FE,
-    0x800068060001A4E1, 0x000FA084000164E3, 0x8000680E8000680A,
-    0x00007D810001B8EF, 0x000148EF00007D80, 0x800065EE00006982,
-    0x800068120002F9B3, 0x00025DB680006816, 0x800068228000681A,
-    0x000249BA000255B5, 0x000CC886000249BB, 0x00023DBE000D5087,
-    0x000239BC000221BF, 0x000D3DA3000235BD, 0x8000681E800067AE,
-    0x8000682A80006826, 0x000D91A5000D99A6, 0x000D8DAB8000682E,
-    0x80006832000D8DAA, 0x000DE5AF80006836, 0x000DF1AD000DF9AE,
-    0x8000683A000DFDAC, 0x800068428000683E, 0x000C70E380006846,
-    0x000F70E18000684A, 0x000C50EF8000684E, 0x8000685680006852,
-    0x8000685E8000685A, 0x8000686680006862, 0x000E15A88000686A,
-    0x000E59AA000E1DA9, 0x000E49AC000E59AB, 0x000E79AE000E49AD,
-    0x8000686E000E79AF, 0x8000687680006872, 0x8000687E8000687A,
-    0x0001B1810001B982, 0x000FF9B80001B980, 0x000F4DBA000FF9B9,
-    0x000F41BC000F41BB, 0x000F7DBE000F4DBD, 0x000E09B0000F75BF,
-    0x000E1DB2000E09B1, 0x000FD5B4000E11B3, 0x000FC9B6000FD9B5,
-    0x000FD1B3000FC9B7, 0x8000688680006882, 0x8000688E8000688A,
-    0x000FC1B5000FC9B6, 0x000FCDBB80006892, 0x00008886000FCDBA,
-    0x000E1DBF0001AC87, 0x000FC5BD000E11BE, 0x80006896000FC9BC,
-    0x8000689A000F95A3, 0x000F8DA68000689E, 0x800068A6800068A2,
-    0x000F89AA000F85A5, 0x800068AA000F89AB, 0x000E55AE800068AE,
-    0x000F8DAC000E59AF, 0x000E59A8000F81AD, 0x000199AA000189A9,
-    0x0001B9AC000191AB, 0x0001D9AE0001B9AD, 0x0000FCBE0001D9AF,
-    0x800068B6800068B2, 0x800068BE800068BA, 0x800068C6800068C2,
-    0x0001B5B8800068CA, 0x0001B5BA0001BDB9, 0x000175BC00016DBB,
-    0x000175BE00017DBD, 0x0001A9B000016DBF, 0x0001BDB20001A9B1,
-    0x000195B40001B1B3, 0x000195B600019DB5, 0x000211B300018DB7,
-    0x800068D2800068CE, 0x800068DA800068D6, 0x000231B5000239B6,
-    0x00026DBB800068DE, 0x800068E200026DBA, 0x000259BF800068E6,
-    0x000265BD000251BE, 0x800068EA000269BC, 0x800068EE000255A3,
-    0x00027DA6800068F2, 0x800068FA800068F6, 0x000229AA000275A5,
-    0x80006902000229AB, 0x000215AE80006906, 0x00022DAC00021DAF,
-    0x0000CD80000221AD, 0x0000D1820000D181, 0x8000690A00040084,
-    0x8000690E0001A4E1, 0x00048086000F2CE3, 0x800069120004A487,
-    0x800069160001F4EF, 0x8000691E8000691A, 0x80006922000000AA,
-    0x8000692A80006926, 0x8000692E000385B3, 0x000381B6000381B5,
-    0x8000693680006932, 0x00034DBA8000693A, 0x00035DBC000345BB,
-    0x00034DBE000345BD, 0x00021DA8000345BF, 0x000241AA000245A9,
-    0x000245AC000241AB, 0x000271AE00024DAD, 0x0004ACBE000271AF,
-    0x800069428000693E, 0x0004B88480006946, 0x8000694E8000694A,
-    0x000149B880006952, 0x000159BA000149B9, 0x000179BC000151BB,
-    0x000119BE000179BD, 0x0002C5B0000115BF, 0x0002C5B20002CDB1,
-    0x0002C5B40002DDB3, 0x0002C5B60002CDB5, 0x000F30E1000179B7,
-    0x000F68E3000EB4E3, 0x80006956000E7CE1, 0x8000695E8000695A,
-    0x8000696680006962, 0x800069728000696A, 0x8000697A80006976,
-    0x000FB8EF000E64EF, 0x800069828000697E, 0x00029DA500029DA6,
-    0x000299A3000740BE, 0x8000698A80006986, 0x000259AF8000673E,
-    0x000259AD000251AE, 0x000259AB000241AC, 0x0004D886000251AA,
-    0x00079D8000055087, 0x0007AD820007A581, 0x0005C8EF8000696E,
-    0x800069928000698E, 0x0004BCE380006996, 0x000490E18000699A,
-    0x800069A28000699E, 0x800069AA800069A6, 0x0006D1B3800069AE,
-    0x800069B6800069B2, 0x800069BE800069BA, 0x0006F1B50006E5B6,
-    0x0006A5BB800069C2, 0x800069C60006A5BA, 0x000691BF800069CA,
-    0x0006A9BD0006A9BE, 0x0006E9A80006B5BC, 0x000619AA0006E9A9,
-    0x000639AC000615AB, 0x00062DAE000625AD, 0x800069CE000625AF,
-    0x00001D8100001D82, 0x800069D200001D80, 0x800069DA800069D6,
-    0x0007E9B8800069DE, 0x0007FDBA0007E9B9, 0x0007EDBC0007F5BB,
-    0x00074DBE000745BD, 0x000621B0000745BF, 0x000621B2000621B1,
-    0x00062DB400063DB3, 0x00061DB6000615B5, 0x000791A30007D9B7,
-    0x00012886800069E2, 0x800069E600014C87, 0x0007B1A50007A5A6,
-    0x0007E5AB800069EA, 0x800069EE0007E5AA, 0x0007D1AF800069F2,
-    0x0007E9AD0007E9AE, 0x800069F60007F5AC, 0x800069FA0006F5B3,
-    0x000685B6800069FE, 0x80006A0680006A02, 0x000179BA00069DB5,
-    0x80006A0A000179BB, 0x000141BE80006A0E, 0x00015DBC000141BF,
-    0x000629A8000151BD, 0x000639AA000629A9, 0x000651AC000631AB,
-    0x000641AE000641AD, 0x80006A12000189AF, 0x80006A1A80006A16,
-    0x80006A2280006A1E, 0x80006A2A80006A26, 0x000185B880006A2E,
-    0x000185BA00018DB9, 0x00018DBC00019DBB, 0x0001B5BE0001BDBD,
-    0x0001F9B000017DBF, 0x0001CDB20001F9B1, 0x0001C1B40001C1B3,
-    0x0001C1B60001C1B5, 0x0005B1A30001C1B7, 0x80006A3680006A32,
-    0x80006A3E80006A3A, 0x0005D9A50005C1A6, 0x00023DAB80006A42,
-    0x80006A4600023DAA, 0x000205AF80006A4A, 0x000215AD000205AE,
-    0x00005D80000219AC, 0x0000658200006D81, 0x00039CBE000351B3,
-    0x000319B6000379B5, 0x0002E08480006A52, 0x00033DBA80006A56,
-    0x000319BC000335BB, 0x0003D9BE000319BD, 0x000385A80003D9BF,
-    0x000395AA000395A9, 0x0003BDAC0003A5AB, 0x0003D1AE0003D5AD,
-    0x80006A5A0003D1AF, 0x80006A6280006A5E, 0x00032087001C60BE,
-    0x80006A66001C9486, 0x000371B880006A6A, 0x000371BA000371B9,
-    0x0000D5BC000371BB, 0x0000D5BE0000DDBD, 0x0003B5B00000CDBF,
-    0x000381B20003BDB1, 0x000351B4000381B3, 0x000351B6000351B5,
-    0x0003A8EF000351B7, 0x80006A7280006A6E, 0x00021C8480006A76,
-    0x80006A7E80006A7A, 0x001C6C8480006A82, 0x80006A8A80006A86,
-    0x000340E380006A8E, 0x000028E180006A92, 0x00025DA380006A96,
-    0x80006A9E80006A9A, 0x80006AA680006AA2, 0x000275A5000215A6,
-    0x000239AB80006AAA, 0x80006AAE000231AA, 0x0002D5AF80006AB2,
-    0x000215AD0002D5AE, 0x0006A0E3000215AC, 0x000668E180006AB6,
-    0x0000FD80001CC4BE, 0x0000058200001181, 0x80006AC280006ABA,
-    0x80006ACA80006AC6, 0x000684EF800068FE, 0x80006AD280006ACE,
-    0x000161B380006AD6, 0x001C688680006ADA, 0x80006ADE001D6C87,
-    0x000171B500016DB6, 0x0000EDBB80006AE2, 0x80006AE60000EDBA,
-    0x0000D5BF80006AEA, 0x0000E1BD0000E9BE, 0x80006ABE0000E9BC,
-    0x80006AEE001C50EF, 0x80006AF680006AF2, 0x80006AFE80006AFA,
-    0x80006B0680006B02, 0x80006B0E80006B0A, 0x80006B1680006B12,
-    0x80006B1A001D34E3, 0x001EE1A3001C90E1, 0x80006B2280006B1E,
-    0x80006B2A80006B26, 0x001EF1A5001EEDA6, 0x001F6DAB00001C85,
-    0x000160BF001F6DAA, 0x001F55AF80006B2E, 0x001F61AD001F69AE,
-    0x001EDDA8001F69AC, 0x001E29AA001EE1A9, 0x001E39AC001E29AB,
-    0x001E25AE001E2DAD, 0x0000E980001F99AF, 0x0000F9820000E981,
-    0x0004008680006B32, 0x80006B3600001487, 0x001FF5B880006B3A,
-    0x001F8DBA001FF9B9, 0x001F9DBC001F85BB, 0x001F85BE001F85BD,
-    0x001FE9B0001FB5BF, 0x001FFDB2001FE9B1, 0x001FEDB4001FF5B3,
-    0x001FD1B6001FD1B5, 0x001EA1A8001FD1B7, 0x001EA1AA001EA1A9,
-    0x001EA9AC001EB9AB, 0x001EDDAE001EA9AD, 0x80006B3E001ECDAF,
-    0x80006B4680006B42, 0x80006B4E80006B4A, 0x80006B5680006B52,
-    0x00015DB880006B5A, 0x00016DBA000165B9, 0x00017DBC000165BB,
-    0x000161BE000165BD, 0x001EBDB0000119BF, 0x001E85B2001E8DB1,
-    0x00017DB4000165B3, 0x00016DB6000165B5, 0x001E29B3000165B7,
-    0x80006B6280006B5E, 0x80006B6A80006B66, 0x001E21B5001E29B6,
-    0x001E49BB80006B6E, 0x80006B72001E79BA, 0x0001A1BF80006B76,
-    0x0001B1BD0001B9BE, 0x80006B7A001E51BC, 0x80006B7E001E6DA3,
-    0x001E6DA680006B82, 0x80006B8A80006B86, 0x001E3DAA001E65A5,
-    0x80006B8E001E0DAB, 0x0001FDAE80006B92, 0x001E15AC0001E5AF,
-    0x002A098F0001F5AD, 0x80006B9600004CE1, 0x80006B9A000A58E3,
-    0x0004D0BE0004D884, 0x0005E8BE80006BA2, 0x0003888780006BA6,
-    0x80006BAA00040C86, 0x80006BB280006BAE, 0x001DED9F007B4CEF,
-    0x001A419D001B559E, 0x00198D9B001A499C, 0x00165D990016459A,
-    0x00120D9700159D98, 0x00119D9500126196, 0x002E7593002E1D94,
-    0x002AD591002D8D92, 0x007629B0002A7190, 0x007AA9B2007801B1,
-    0x007C01B4007AB1B3, 0x80006BB6007E91B5, 0x00034D8080006BBA,
-    0x0000B9820000AD81, 0x0002B1A380006BBE, 0x001E05A100010DA2,
-    0x80006BC2001EB1A0, 0x0006CDA20002A9A1, 0x000801A40006C5A3,
-    0x000AD1A6000AD1A5, 0x000EE1A8000BB9A7, 0x000FF9AA000EE1A9,
-    0x0072D1AC007001AB, 0x007401AE00722DAD, 0x002B09820076DDAF,
-    0x80006BC6002B0983, 0x002F158680006BCA, 0x002BB584002F6587,
-    0x0010D58A002CE585, 0x80006BCE0013218B, 0x0017358E80006BD2,
-    0x0013558C0017258F, 0x00189D920014FD8D, 0x00048484001B4193,
-    0x001CFD9600070884, 0x001B5194001F5997, 0x001FDD9A001A4D95,
-    0x80006BD60000DD9B, 0x80006BDE80006BDA, 0x0003799C80006BE2,
-    0x80006BEA80006BE6, 0x80006BEE001E54E1, 0x80006BF2001EBCE3,
-    0x80006BFA80006BF6, 0x00000D8080006BFE, 0x00001D8200001581,
-    0x80006C0680006C02, 0x80006A4E80006C0A, 0x00056C87001EC8EF,
-    0x80006C120004E486, 0x000330EF80006C16, 0x80006C1E80006C1A,
-    0x0003E4E380006C22, 0x000140E180006C26, 0x80006C2E80006C2A,
-    0x80006C3680006C32, 0x0004BDB380006C3A, 0x80006C0E80006B9E,
-    0x80006C4280006C3E, 0x00053DB500053DB6, 0x000519BB80006C46,
-    0x80006C4A000511BA, 0x0005BDBF80006C4E, 0x0005BDBD0005BDBE,
-    0x80006C52000501BC, 0x80006C560004F9A3, 0x000579A680006C5A,
-    0x80006C6280006C5E, 0x000555AA000579A5, 0x80006C6600055DAB,
-    0x0005F9AE80006C6A, 0x000545AC0005F9AF, 0x00069DA80005F9AD,
-    0x0006B9AA0006A1A9, 0x0006A9AC0006B9AB, 0x000605AE00060DAD,
-    0x80006C6E000631AF, 0x0000118100001982, 0x0000ECBE00001980,
-    0x80006C7680006C72, 0x0007B5B880006C7A, 0x000749BA0007B9B9,
-    0x000741BC000745BB, 0x000741BE000741BD, 0x000621B0000741BF,
-    0x0007E5B20007EDB1, 0x0007DDB40007C5B3, 0x000795B600079DB5,
-    0x000731B300078DB7, 0x0000A88680006C82, 0x80006C8600018887,
-    0x000721B500078DB6, 0x000629BB80006C8A, 0x80006C8E0007E5BA,
-    0x000659BF80006C92, 0x000611BD000619BE, 0x80006C96000631BC,
-    0x80006C9A000775A3, 0x0007C9A680006C9E, 0x80006CA680006CA2,
-    0x0007A1AA000765A5, 0x80006CAA00066DAB, 0x00065DAE80006CAE,
-    0x000675AC00061DAF, 0x000649A8000655AD, 0x00065DAA000649A9,
-    0x00064DAC000655AB, 0x000675AE000671AD, 0x0000AC8400066DAF,
-    0x80006CB680006CB2, 0x80006CBE80006CBA, 0x80006CC680006CC2,
-    0x000181B880006CCA, 0x000185BA00018DB9, 0x000189BC000199BB,
-    0x0001B5BE0001BDBD, 0x000615B000017DBF, 0x000615B200061DB1,
-    0x0001E1B40001E1B3, 0x0001D1B60001E9B5, 0x000725AC0001D1B7,
-    0x000725AE00072DAD, 0x0006BDA800073DAF, 0x0006A5AA0006ADA9,
-    0x000CACBE00073DAB, 0x80006CD280006CCE, 0x80006CDA80006CD6,
-    0x80006CE280006CDE, 0x00078DBC80006CE6, 0x000791BE000791BD,
-    0x0007F9B80007A5BF, 0x000799BA0007F9B9, 0x0007E9B4000795BB,
-    0x0007F9B60007E9B5, 0x000745B00007F1B7, 0x000745B200074DB1,
-    0x0000BD810007F9B3, 0x00055DB30000BD80, 0x00054DB500001982,
-    0x80006CF280006CEA, 0x80006CF6000579B6, 0x000551BB80006CFA,
-    0x0002BDBD000559BA, 0x0002A5BF000549BC, 0x80006CFE0002ADBE,
-    0x000515A5000521A6, 0x000505A3000C4C84, 0x000CF88680006D02,
-    0x0002FDAF000C9C87, 0x0002E5AD0002F5AE, 0x000509AB000511AC,
-    0x80006D06000501AA, 0x000205A880006D0A, 0x000215AA000215A9,
-    0x00023DAC000225AB, 0x000251AE000255AD, 0x000CACBE000251AF,
-    0x80006D1280006D0E, 0x80006D1A80006D16, 0x80006D2280006D1E,
-    0x0002F5B880006D26, 0x0002F5BA0002FDB9, 0x000179BC000169BB,
-    0x000169BE000179BD, 0x000235B0000169BF, 0x000201B200023DB1,
-    0x0002D5B4000201B3, 0x0002D5B60002DDB5, 0x0007B4E10002CDB7,
-    0x0006D0E30006D8E3, 0x80006D2A000650E1, 0x80006D3280006D2E,
-    0x80006D3A80006D36, 0x80006D4280006D3E, 0x000000EC80006D46,
-    0x000760EF0001FCEF, 0x0002DDB380006D4A, 0x000CE88680006D4E,
-    0x80006D52000C8C87, 0x0002F5B5000295B6, 0x0002B9BB80006D56,
-    0x80006D5A0002B1BA, 0x000259BF80006D62, 0x000295BD000251BE,
-    0x000D1DA3000295BC, 0x80006D5E80006CEE, 0x80006D6A80006D66,
-    0x000D35A5000D55A6, 0x000D79AB80006D6E, 0x80006D72000D71AA,
-    0x000D99AF80006D76, 0x000D55AD000D91AE, 0x000F9180000D55AC,
-    0x000FA182000F9981, 0x80006D7A000EE9B3, 0x000EA1B6000E81B5,
-    0x80006D8280006D7E, 0x000F41BA80006D86, 0x000F41BC000F41BB,
-    0x000F41BE000F41BD, 0x000E19A8000F41BF, 0x000E2DAA000E25A9,
-    0x000E45AC000E4DAB, 0x000E65AE000E6DAD, 0x80006D8A000FB9AF,
-    0x80006D9280006D8E, 0x80006D9A80006D96, 0x80006DA280006D9E,
-    0x000F59B880006DA6, 0x000F69BA000F59B9, 0x000F79BC000F69BB,
-    0x000F65BE000F79BD, 0x000FC9B0000F19BF, 0x000FD9B2000FC9B1,
-    0x000FC9B4000FD9B3, 0x000F69B6000FC9B5, 0x000FA5A3000F69B7,
-    0x80006DAE80006DAA, 0x80006DB680006DB2, 0x000FCDA5000FEDA6,
-    0x000E0DAB80006DBA, 0x80006DBE000E0DAA, 0x000E0DAF80006DC2,
-    0x000E0DAD000E0DAE, 0x00006D80000E0DAC, 0x0000198200000981,
-    0x80006DCA80006DC6, 0x00019484000E44E1, 0x0001A8BE000EECE3,
-    0x80006DD280006DCE, 0x0000D086000DB4EF, 0x80006DD60001FC87,
-    0x000E6DA880006DDA, 0x000199AA000189A9, 0x0001B9AC000191AB,
-    0x0001C9AE0001B9AD, 0x0000A0840001C1AF, 0x80006DE280006DDE,
-    0x80006DEA80006DE6, 0x80006DF280006DEE, 0x000179B880006DF6,
-    0x0000CDBA000179B9, 0x0000DDBC0000C5BB, 0x0000C5BE0000C5BD,
-    0x000181B00000F5BF, 0x000149B2000181B1, 0x000159B4000149B3,
-    0x000149B6000159B5, 0x000211B3000149B7, 0x80006DFE80006DFA,
-    0x80006E0680006E02, 0x000231B5000239B6, 0x000245BB80006E0A,
-    0x80006E0E000245BA, 0x00029DBF80006E12, 0x00029DBD00029DBE,
-    0x0075FC8500029DBC, 0x80006E16000255A3, 0x00027DA680006E1A,
-    0x80006E2280006E1E, 0x000201AA000275A5, 0x80006E26000201AB,
-    0x0002D9AE80006E2A, 0x0002D9AC0002D9AF, 0x000019820002D9AD,
-    0x0000198080006E2E, 0x80006E3200001981, 0x80006E3A80006E36,
-    0x000400BE80006E42, 0x80006E4A80006E46, 0x80006E5280006E4E,
-    0x80006E5A80006E56, 0x0004E08680006E5E, 0x80006E6200035487,
-    0x80006C7E80006E66, 0x000280EF80006E6A, 0x000188E180006E6E,
-    0x000158E380006E72, 0x80006E7A80006E76, 0x80006E8280006E7E,
-    0x000269A880006E86, 0x000275AA00027DA9, 0x0002B9AC0002A9AB,
-    0x0002A9AE0002B9AD, 0x0005AC840002A9AF, 0x80006E8E80006E8A,
-    0x80006E9680006E92, 0x80006E9E80006E9A, 0x00016DB880006EA2,
-    0x000105BA00010DB9, 0x000109BC000119BB, 0x000135BE00013DBD,
-    0x0002D9B00001D9BF, 0x00016DB20002D9B1, 0x00017DB4000165B3,
-    0x000165B6000165B5, 0x000F40E1000155B7, 0x000F54E30000CCE3,
-    0x0000AD80000E5CE1, 0x0000398200000981, 0x80006EAA80006EA6,
-    0x80006EB2000428BE, 0x00073CBE80006EB6, 0x000E60EF00002CEF,
-    0x0004E08680006EBA, 0x80006EBE00057487, 0x80006EC20002B1B3,
-    0x000275B600027DB5, 0x80006ECA80006EC6, 0x000251BA80006ECE,
-    0x0001F5BC000251BB, 0x0001F5BE0001FDBD, 0x80006E3E0001EDBF,
-    0x0005B1AA80006EAE, 0x000615AC0005B1AB, 0x000615AE00061DAD,
-    0x80006ED200060DAF, 0x80006EDA80006ED6, 0x80006EDE000551A3,
-    0x000595A600059DA5, 0x0006E9B380006EE2, 0x80006EEA80006EE6,
-    0x80006EF280006EEE, 0x0006E9B50006E1B6, 0x0006ADBB80006EF6,
-    0x80006EFA0006ADBA, 0x00069DBF80006EFE, 0x0006A5BD000691BE,
-    0x00061DA80006A9BC, 0x000621AA000621A9, 0x000621AC000621AB,
-    0x000621AE000621AD, 0x80006F02000621AF, 0x80006F0A80006F06,
-    0x00000D8080006F0E, 0x0001B1820001B181, 0x000639B880006F12,
-    0x0007EDBA000639B9, 0x0007FDBC0007E5BB, 0x0007EDBE0007E5BD,
-    0x000629B00007E5BF, 0x00063DB2000629B1, 0x000615B4000631B3,
-    0x000609B6000619B5, 0x0007A9A3000609B7, 0x80006F1A80006F16,
-    0x80006F1E00028484, 0x0007A9A50007A1A6, 0x0007EDAB80006F22,
-    0x000088860007EDAA, 0x0007DDAF00017887, 0x0007E5AD0007D1AE,
-    0x80006F260007E9AC, 0x80006F2A0006F9B3, 0x00068DB680006F2E,
-    0x80006F3680006F32, 0x000141BA00068DB5, 0x80006F3A000141BB,
-    0x00014DBE80006F3E, 0x000151BC000131BF, 0x000629A800014DBD,
-    0x000635AA000629A9, 0x000641AC000645AB, 0x000679AE000649AD,
-    0x0001AC84000679AF, 0x80006F4680006F42, 0x80006F4E80006F4A,
-    0x80006F5680006F52, 0x000191B880006F5A, 0x0001A1BA000199B9,
-    0x000171BC0001A1BB, 0x000171BE000171BD, 0x0001C1B0000171BF,
-    0x0001C1B20001C1B1, 0x0001CDB40001DDB3, 0x0001B1B60001B1B5,
-    0x0005BDA30001B1B7, 0x80006F6280006F5E, 0x80006F6A80006F66,
-    0x0005C9A50005C9A6, 0x000205AB80006F6E, 0x80006F72000205AA,
-    0x000275AF80006F76, 0x000209AD000209AE, 0x80006F7A000215AC,
-    0x80006F8280006F7E, 0x00001D8180006F86, 0x80006F8A00001D80,
-    0x000324BE00000582, 0x007CC88680006F92, 0x80006F9600035C87,
-    0x80006F9E80006F9A, 0x80006FA680006FA2, 0x80006FAE80006FAA,
-    0x80006FB2007CC084, 0x80006FB60001A4E1, 0x007CE8BE00012CE3,
-    0x80006FBE80006FBA, 0x000000AE000234EF, 0x80006FC680006FC2,
-    0x000389B380006FCA, 0x80006FD280006FCE, 0x80006FDA80006FD6,
-    0x00038DB500038DB6, 0x000349BB80006FDE, 0x80006FE2000341BA,
-    0x000349BF80006FE6, 0x000349BD000341BE, 0x00021DA8000351BC,
-    0x000241AA000245A9, 0x000245AC000241AB, 0x000271AE00024DAD,
-    0x007FA0BE000271AF, 0x80006FEE80006FEA, 0x00001D8080006FF2,
-    0x0000A98200000981, 0x000149B880006FF6, 0x000159BA000149B9,
-    0x000179BC000151BB, 0x000119BE000179BD, 0x0002C5B0000119BF,
-    0x0002C5B20002CDB1, 0x0002C5B40002DDB3, 0x0002C5B60002CDB5,
-    0x000658E1000179B7, 0x000094E300075CE1, 0x80006FFA0006BCE3,
-    0x007CB88680007002, 0x80007006007D5487, 0x8000700E8000700A,
-    0x000000EE80007012, 0x0006F0EF80007016, 0x8000701A0006E0EF,
-    0x000281A38000701E, 0x000285A580007022, 0x8000702A80007026,
-    0x8000702E000285A6, 0x000241AB80007032, 0x000241AD000249AA,
-    0x000241AF000259AC, 0x007E65A8000249AE, 0x007E7DAA007E75A9,
-    0x007E6DAC007E75AB, 0x007E9DAE007E95AD, 0x80006FFE007E95AF,
-    0x8000703A80007036, 0x800070428000703E, 0x8000704A80007046,
-    0x007E89B88000704E, 0x007E99BA007E89B9, 0x007EB9BC007E91BB,
-    0x007F7DBE007EB9BD, 0x007EF1B0007F75BF, 0x007ECDB2007EFDB1,
-    0x007EC5B4007EC5B3, 0x007EC5B6007ECDB5, 0x007EF5B3007EB9B7,
-    0x8000705680007052, 0x8000705E8000705A, 0x007EE5B5007E15B6,
-    0x007E39BB80007062, 0x80007066007E35BA, 0x007E09BF8000706A,
-    0x007E11BD007E01BE, 0x00006D82007E19BC, 0x00005580007EB1A3,
-    0x007E51A600006581, 0x8000706E00031884, 0x007E71AA007EA1A5,
-    0x00008886007E7DAB, 0x007E45AE00035C87, 0x007E5DAC007E4DAF,
-    0x007EA5B3007E55AD, 0x8000707680007072, 0x8000707E8000707A,
-    0x007EB5B5007EA1B6, 0x007EF5BB80007082, 0x80007086007EE9BA,
-    0x000115BF8000708A, 0x007ED9BD007ED1BE, 0x8000708E007EE5BC,
-    0x8000709680007092, 0x007D08EF8000709A, 0x800070A28000709E,
-    0x007DFCE3800070A6, 0x007D18E1800070AA, 0x800070B2800070AE,
-    0x800070BA800070B6, 0x007E15A8800070BE, 0x007E2DAA007E25A9,
-    0x007E3DAC007E25AB, 0x007E59AE007E21AD, 0x800070C2007E59AF,
-    0x800070CA800070C6, 0x00000D80800070CE, 0x0003BD820003B581,
-    0x0001F1B8800070D2, 0x0001F1BA0001F1B9, 0x000155BC0001F1BB,
-    0x000155BE00015DBD, 0x007E2DB000014DBF, 0x007E31B2007E31B1,
-    0x007E15B4007E0DB3, 0x0001D1B60001D1B5, 0x800070D60001D1B7,
-    0x00029084800070DA, 0x00039CBE007DA9A3, 0x007DADA6007DB9A5,
-    0x800070E6800070E2, 0x007DE5AA800070EA, 0x007DE9AC007DF9AB,
-    0x007DDDAE007DD5AD, 0x000369B6000219AF, 0x00034884800070EE,
-    0x800070F2000351B5, 0x0004F886000359B3, 0x000339BE00033487,
-    0x000325BC000339BF, 0x000335BA000339BD, 0x800070F600033DBB,
-    0x0003F1B8800070FA, 0x0003F1BA0003F1B9, 0x000331BC0003F1BB,
-    0x000331BE000331BD, 0x0003E5B0000331BF, 0x0003E5B20003EDB1,
-    0x0003E9B40003F9B3, 0x0003DDB60003E9B5, 0x0003C5A80003D1B7,
-    0x0003C5AA0003CDA9, 0x0003CDAC0003DDAB, 0x0003F5AE0003FDAD,
-    0x800070FE00039DAF, 0x8000710680007102, 0x0004FCBE80006F8E,
-    0x8000710E8000710A, 0x00020DA380007112, 0x8000711A80007116,
-    0x800071228000711E, 0x000205A500023DA6, 0x000269AB80007126,
-    0x8000712A000261AA, 0x00026DAF8000712E, 0x00026DAD00026DAE,
-    0x80007132000271AC, 0x8000713A80007136, 0x800071428000713E,
-    0x8000714600034CE3, 0x000011820001ACE1, 0x0000398000038CEF,
-    0x8000714A00003581, 0x0004588680007152, 0x8000715600054087,
-    0x8000715E8000715A, 0x8000716680007162, 0x8000716E8000716A,
-    0x80007172007E44EF, 0x80007176007E24E1, 0x000000EC000168E3,
-    0x8000717E8000717A, 0x8000718680007182, 0x8000718E8000718A,
-    0x800071920001E9B3, 0x0001A1B6000185B5, 0x8000719A80007196,
-    0x00010DBA8000719E, 0x000105BC000105BB, 0x000105BE00010DBD,
-    0x00061DA80000F9BF, 0x00062DAA000625A9, 0x000641AC000649AB,
-    0x000665AE00066DAD, 0x8000714E0006FDAF, 0x800071A6800071A2,
-    0x800071AE800071AA, 0x0000B1810000B982, 0x000751B80000B980,
-    0x000761BA000759B9, 0x000711BC000761BB, 0x000711BE000711BD,
-    0x000685B0000711BF, 0x00069DB200068DB1, 0x00068DB4000695B3,
-    0x000771B6000771B5, 0x0006A5A3000771B7, 0x800071B6800071B2,
-    0x00029884800071BA, 0x0006C9A50006EDA6, 0x000649AB000310BE,
-    0x800071BE000641AA, 0x0007B5AF800071C2, 0x000641AD000649AE,
-    0x800071C6000649AC, 0x00006886000741B3, 0x000741B600036C87,
-    0x800071CE800071CA, 0x000741BA000751B5, 0x800071D200074DBB,
-    0x00074DBE800071D6, 0x00074DBC000735BF, 0x000685A3000745BD,
-    0x800071DE800071DA, 0x800071E6800071E2, 0x000695A5000685A6,
-    0x000689AB800071EA, 0x800071EE000685AA, 0x0006F1AF800071F2,
-    0x000681AD000689AE, 0x800071F6000689AC, 0x800071FA00054CEF,
-    0x80007202800071FE, 0x8000720A80007206, 0x800072128000720E,
-    0x8000721A80007216, 0x800072228000721E, 0x8000722600052CE3,
-    0x0006F5A800053CE1, 0x00060DAA000605A9, 0x00061DAC000605AB,
-    0x00060DAE000605AD, 0x0000D180000605AF, 0x000005820000DD81,
-    0x0001ECBE8000722A, 0x800072320001F884, 0x000611B880007236,
-    0x000629BA000619B9, 0x0001F5BC000629BB, 0x0001F5BE0001FDBD,
-    0x000641B00001EDBF, 0x00065DB200064DB1, 0x00064DB4000655B3,
-    0x00063DB6000635B5, 0x0002ADA8000635B7, 0x0002DDAA0002C5A9,
-    0x0002F9AC0002D5AB, 0x00031DAE0002F9AD, 0x000C8086000315AF,
-    0x8000723A00033C87, 0x800072428000723E, 0x8000724A80007246,
-    0x000311B8000378BE, 0x000311BA000311B9, 0x000331BC000311BB,
-    0x000331BE000331BD, 0x00036DB0000331BF, 0x000375B2000375B1,
-    0x00035DB4000345B3, 0x000331B6000335B5, 0x000255B3000331B7,
-    0x800072528000724E, 0x8000725A80007256, 0x000271B5000279B6,
-    0x0003B1BB8000725E, 0x800072620003A9BA, 0x0003F1BF80007266,
-    0x000391BD000399BE, 0x8000726A0003A1BC, 0x000221A5000229A6,
-    0x000205A38000726E, 0x8000727680007272, 0x0003A1AF8000727A,
-    0x0003C1AD0003C9AE, 0x0003E1AB0003F1AC, 0x800070DE0003F9AA,
-    0x800072828000727E, 0x8000728A80007286, 0x000025828000728E,
-    0x0000318000005CEF, 0x0002AC8400002D81, 0x80007292000080E3,
-    0x800072960001FCE1, 0x000D18868000729A, 0x000169B3000D1087,
-    0x000C44BE800072A2, 0x800072AA800072A6, 0x000169B5000161B6,
-    0x0001EDBB800072AE, 0x800072B20001EDBA, 0x0001DDBF800072B6,
-    0x0001E5BD0001DDBE, 0x800072BA0001EDBC, 0x800072C2800072BE,
-    0x800072C6000C0084, 0x800072CE800072CA, 0x000BEC850001F8EF,
-    0x800072D20006C4E1, 0x800072D600012CE3, 0x800072DE800072DA,
-    0x800072E6800072E2, 0x0001E1A3800072EA, 0x0001E1A5800072EE,
-    0x800072F6800072F2, 0x800072FA0001E9A6, 0x000165AB800072FE,
-    0x00016DAD000165AA, 0x000155AF000165AC, 0x000DB9A8000155AE,
-    0x000E21AA000DB9A9, 0x000E21AC000E21AB, 0x000E21AE000E21AD,
-    0x8000729E000E99AF, 0x00001D8100001D82, 0x8000730200001D80,
-    0x8000730A80007306, 0x000E95B88000730E, 0x000E95BA000E9DB9,
-    0x000F51BC000EB1BB, 0x000F51BE000F51BD, 0x000EE9B0000F51BF,
-    0x000EF9B2000EE9B1, 0x000ED9B4000EF1B3, 0x000EB9B6000ED9B5,
-    0x000FC5A8000EB1B7, 0x000FD9AA000FC9A9, 0x000FF9AC000FD1AB,
-    0x000F19AE000FF9AD, 0x00000086000F19AF, 0x8000731200036487,
-    0x8000731A80007316, 0x800073228000731E, 0x000F7DB880007326,
-    0x000F0DBA000F05B9, 0x000F05BC000F01BB, 0x000F39BE000F0DBD,
-    0x000F69B0000F39BF, 0x000F7DB2000F75B1, 0x000F59B4000F75B3,
-    0x000F4DB6000F45B5, 0x000D38E1000F45B7, 0x000C38E38000732A,
-    0x800073328000732E, 0x8000733A80007336, 0x800073428000733E,
-    0x8000734A80007346, 0x800073528000734E, 0x000DFCEF80007356,
-    0x000E11B38000735A, 0x800073628000735E, 0x8000736A80007366,
-    0x000E31B5000E39B6, 0x000EF1BB8000736E, 0x80007372000EEDBA,
-    0x000EC1BF80007376, 0x000EE1BD000EDDBE, 0x000ED1A3000EE9BC,
-    0x000288848000737A, 0x80007382000180BE, 0x000EF1A5000EF9A6,
-    0x000E31AB80007386, 0x00000886000E2DAA, 0x000E01AF00014C87,
-    0x000E21AD000E1DAE, 0x0000A980000E29AC, 0x00004D820000A981,
-    0x8000738A00017DB3, 0x000165B600016DB5, 0x800073928000738E,
-    0x0000F9BA80007396, 0x0000E9BC0000F9BB, 0x0000D9BE0000E9BD,
-    0x0002A9A80000D9BF, 0x0002B9AA0002A9A9, 0x0002A9AC0002B9AB,
-    0x00031DAE0002A9AD, 0x0004ACBE000311AF, 0x8000739E8000739A,
-    0x800073A20005E0BE, 0x800073AA800073A6, 0x000371B8800073AE,
-    0x000375BA000371B9, 0x000319BC000309BB, 0x000309BE000319BD,
-    0x000371B0000301BF, 0x000371B2000371B1, 0x000351B4000371B3,
-    0x000351B6000351B5, 0x000231A3000351B7, 0x800073B28000722E,
-    0x800073BA800073B6, 0x000221A5000229A6, 0x0003B5AB800073BE,
-    0x800073C20003B5AA, 0x000395AF800073C6, 0x0003A5AD000395AE,
-    0x800073CA0003A5AC, 0x800073D2800073CE, 0x800073DA800073D6,
-    0x800073E2800073DE, 0x00003980800073E6, 0x0000058200003981,
-    0x800073F2800073EA, 0x800073F60003FC84, 0x000291A8800073FA,
-    0x0002D9AA000299A9, 0x0002F1AC0002D1AB, 0x00013DAE0002F1AD,
-    0x00050086000135AF, 0x800073FE00056487, 0x8000740680007402,
-    0x8000740E8000740A, 0x0001E9B880007412, 0x000189BA0001E9B9,
-    0x000199BC000189BB, 0x000189BE000199BD, 0x000151B0000189BF,
-    0x000151B2000151B1, 0x0001F9B4000151B3, 0x0001E9B60001F9B5,
-    0x0000F4EF0001E1B7, 0x8000741A80007416, 0x8000741E00078CBF,
-    0x0005188480007422, 0x80007426000E2CEF, 0x8000742A000E84E1,
-    0x000028E3000E30E3, 0x000100E18000742E, 0x0020FC8580007432,
-    0x80007436000215B3, 0x000235B68000743A, 0x800074428000743E,
-    0x0001D1BA00023DB5, 0x800074460001D1BB, 0x000179BE8000744A,
-    0x0001C1BC000179BF, 0x0005D1A30001C1BD, 0x8000744E800073EE,
-    0x8000745680007452, 0x0005F9A50005F1A6, 0x000615AB8000745A,
-    0x8000745E000615AA, 0x0006BDAF80007462, 0x000605AD0006BDAE,
-    0x00004D80000605AC, 0x0000558200005581, 0x800074660006B1B3,
-    0x000769B6000779B5, 0x0001C0868000746A, 0x000755BA0003E887,
-    0x0007F9BC0007E9BB, 0x0007EDBE0007E5BD, 0x000671A80007DDBF,
-    0x000675AA00067DA9, 0x0007BDAC0007A5AB, 0x0007ADAE0007A5AD,
-    0x8000746E0007A5AF, 0x8000747680007472, 0x8000747E8000747A,
-    0x8000748680007482, 0x00077DB88000748A, 0x000705BA00070DB9,
-    0x000709BC000719BB, 0x000735BE00073DBD, 0x0007DDB00000D5BF,
-    0x000761B2000761B1, 0x000765B400077DB3, 0x000751B6000755B5,
-    0x8000748E000749B7, 0x8000749680007492, 0x0008A4E18000749A,
-    0x0004CCE30006A4E1, 0x8000749E0005F8E3, 0x00187CEF800074A2,
-    0x000135A30005F4EF, 0x000505A10003A5A2, 0x0007BDA300051DA0,
-    0x800074AA800074A6, 0x800074B2800074AE, 0x000675A5000665A6,
-    0x0006E5AB800074B6, 0x800074BA000659AA, 0x0006D1AF800074BE,
-    0x0006E9AD0006E1AE, 0x007DED8F0006F5AC, 0x00001D8100001D82,
-    0x800074C200001D80, 0x800074CA800074C6, 0x0000BC84001C00BE,
-    0x00009887800074D2, 0x800074D6001F2C86, 0x800074DE800074DA,
-    0x0007459F800074E2, 0x0009E99D00096D9E, 0x000D099B000B359C,
-    0x000F3599000DF59A, 0x0071E5970070CD98, 0x00748D950071E596,
-    0x0075F5930075E594, 0x0079F19100772592, 0x800074E60079F990,
-    0x007E6582800074EA, 0x007A1984007A0983, 0x00761D860074C585,
-    0x001C6C8400761587, 0x00774D8A800074EE, 0x0072358C00722D8B,
-    0x000CD58E0073258D, 0x00250485000E1D8F, 0x000E9192800074F2,
-    0x000A71940008B593, 0x0004DD96000AFD95, 0x800074F600067D97,
-    0x0006559A800074FA, 0x0002759C0000FD9B, 0x001C30BE800074FE,
-    0x002AA1B380007502, 0x002FA9B1002801B2, 0x8000737E002E85B0,
-    0x002401B5001DD884, 0x80007506002BA9B4, 0x8000750E8000750A,
-    0x8000751680007512, 0x8000751E8000751A, 0x001EC5A380007522,
-    0x0002E1A1001EC5A2, 0x001A01A780007526, 0x001801A5001AF5A6,
-    0x001001AB001E05A4, 0x0016C9A90016D5AA, 0x002EF5AF001401A8,
-    0x0012A5AD00138DAE, 0x00032DA80012F9AC, 0x000355AA000355A9,
-    0x0003B5AC0003ADAB, 0x0003B5AE0003BDAD, 0x8000752A0003ADAF,
-    0x800075328000752E, 0x0000198080007536, 0x0000058200001981,
-    0x000265B88000753A, 0x000201BA000201B9, 0x000201BC000201BB,
-    0x000231BE000209BD, 0x0003D5B0000231BF, 0x0003D5B20003DDB1,
-    0x000275B400026DB3, 0x000275B600027DB5, 0x0006B8E100025DB7,
-    0x000600E380007542, 0x80007546001CC4BE, 0x001C38860000F0EF,
-    0x8000754A001CBC87, 0x8000754E000014E3, 0x001F10860001D4E1,
-    0x00076CEF001F8487, 0x80007552001FF084, 0x00036DB500036DB6,
-    0x00038DB3001F84BE, 0x8000755A80007556, 0x0002E9BF8000755E,
-    0x0002E9BD0002E1BE, 0x000349BB000351BC, 0x80007562000341BA,
-    0x800074CE80007566, 0x8000756A8000753E, 0x800075728000756E,
-    0x8000757A80007576, 0x001D85A88000757E, 0x001D9DAA001D95A9,
-    0x001D8DAC001D95AB, 0x001E3DAE001E35AD, 0x001E05B0001E19AF,
-    0x001E2DB2001E25B1, 0x001E2DB4001E39B3, 0x001EFDB6001EF5B5,
-    0x001ECDB8001EF5B7, 0x001EDDBA001ED5B9, 0x001ECDBC001ED5BB,
-    0x001F59BE001EB9BD, 0x001D35B3001F59BF, 0x8000758680007582,
-    0x8000758E8000758A, 0x001D25B5001E95B6, 0x001EC1BB80007592,
-    0x80007596001EE1BA, 0x001E11BF8000759A, 0x001ECDBD001EC9BE,
-    0x8000759E001ED9BC, 0x800075A2001D71A3, 0x001ED1A6800075A6,
-    0x800075AE800075AA, 0x001EA5AA001D61A5, 0x800075B2001E85AB,
-    0x001E8DAE800075B6, 0x001E9DAC001E55AF, 0x001E85A8001E89AD,
-    0x001EA9AA001E95A9, 0x001ED1AC001EA5AB, 0x001EC1AE001EC1AD,
-    0x0001E980001EC1AF, 0x0001FD820001E981, 0x800075BE800075BA,
-    0x800075C6800075C2, 0x0001C5B8800075CA, 0x0001C5BA0001CDB9,
-    0x0001CDBC0001DDBB, 0x0001F5BE0001FDBD, 0x001E99B000019DBF,
-    0x000161B2001E99B1, 0x000161B4000161B3, 0x000161B6000161B5,
-    0x001E21B3000161B7, 0x800075D2800075CE, 0x800075DA800075D6,
-    0x001E59B5001E41B6, 0x001E5DBB00009084, 0x800075DE001E5DBA,
-    0x0001A5BF800075E2, 0x0001B5BD0001A5BE, 0x800075E60001B5BC,
-    0x800075EA001E65A3, 0x001E05A6800075EE, 0x800075F6800075F2,
-    0x001E19AA001E1DA5, 0x800075FA001E19AB, 0x0001E1AE800075FE,
-    0x0001F1AC0001E1AF, 0x0002E9A80001F1AD, 0x000325AA0002E9A9,
-    0x000325AC00033DAB, 0x000325AE00032DAD, 0x80007602000399AF,
-    0x8000760A80007606, 0x00000D808000760E, 0x0000B1820000B181,
-    0x000395B880007612, 0x0003A1BA00039DB9, 0x000071BC0003A1BB,
-    0x000071BE000071BD, 0x0003E9B0000071BF, 0x0003F9B20003E9B1,
-    0x0003DDB40003C5B3, 0x0003B1B60003B5B5, 0x0003ADB30003B1B7,
-    0x8000761A80007616, 0x8000762680007622, 0x0003ADB50003ADB6,
-    0x00032DBB8000762A, 0x0004088600032DBA, 0x000371BF0003DC87,
-    0x000311BD000319BE, 0x8000762E00032DBC, 0x800076320003E9A3,
-    0x0003E9A680007636, 0x8000763E8000763A, 0x000369AA0003E9A5,
-    0x80007642000369AB, 0x00035DAE80007646, 0x000369AC000335AF,
-    0x8000764A000355AD, 0x001FADA5001FADA6, 0x001F4DA38000764E,
-    0x8000765680007652, 0x001E29AF8000765A, 0x001E29AD001E21AE,
-    0x001F89AB001F91AC, 0x8000765E001F81AA, 0x8000766680007662,
-    0x000544BE8000766A, 0x8000766E00048084, 0x80007672001EF8E3,
-    0x80007676001EF4E1, 0x8000767A001E50EF, 0x00000D808000767E,
-    0x0000ED820000E581, 0x001E00E180007682, 0x001F0CE3001C30E1,
-    0x80007686001FBCE3, 0x0004B8868000768A, 0x000720BE00055C87,
-    0x8000769680007692, 0x8000769E8000769A, 0x001ECCEF800076A2,
-    0x00015DB3001F3CEF, 0x800076AA800076A6, 0x800076B2800076AE,
-    0x000111B5000119B6, 0x00010DBB800076B6, 0x800076BA00010DBA,
-    0x0000E9BF800076BE, 0x0000F5BD0000E1BE, 0x00068DA80000FDBC,
-    0x000695AA000695A9, 0x0006D9AC0006CDAB, 0x0006CDAE0006D9AD,
-    0x8000768E0006C5AF, 0x800076C6800076C2, 0x800076CE800076CA,
-    0x800076D6800076D2, 0x00077DB8800076DA, 0x00070DBA000705B9,
-    0x000709BC00071DBB, 0x000731BE000735BD, 0x000689B0000731BF,
-    0x000769B2000689B1, 0x000765B4000761B3, 0x000765B600076DB5,
-    0x00061DA3000745B7, 0x800076E2800076DE, 0x800076EA800076E6,
-    0x000651A5000659A6, 0x00064DAB800076EE, 0x800076F200064DAA,
-    0x0007A9AF800076F6, 0x0007B5AD0007A1AE, 0x000069800007BDAC,
-    0x0000058200006981, 0x800076FE800076FA, 0x0003A88780007702,
-    0x8000770600009C86, 0x8000770E8000770A, 0x8000771680007712,
-    0x8000771E8000771A, 0x0006ADB380007722, 0x8000772A80007726,
-    0x800077328000772E, 0x0006ADB50006D1B6, 0x0006FDBB80007736,
-    0x8000773A0006F5BA, 0x000119BF8000773E, 0x0006D5BD000111BE,
-    0x800077420006D5BC, 0x8000774A80007746, 0x800077528000774E,
-    0x800077560006D8E3, 0x8000775A0005ECE1, 0x8000775E0004D4EF,
-    0x00001D8080007762, 0x0000D98200000981, 0x000625A880007766,
-    0x000625AA00062DA9, 0x000625AC00063DAB, 0x000651AE000651AD,
-    0x8000776A000651AF, 0x000024BE00013C84, 0x000F108680007772,
-    0x800077760000C487, 0x0001F5B88000777A, 0x0001F5BA0001FDB9,
-    0x000159BC000149BB, 0x000149BE000159BD, 0x000631B0000141BF,
-    0x000601B2000639B1, 0x0001D5B4000601B3, 0x0001D5B60001DDB5,
-    0x8000777E0001CDB7, 0x8000778680007782, 0x8000778A0005A1A3,
-    0x0005DDA60005A1A5, 0x800077928000778E, 0x0005F9AA80007796,
-    0x0005D9AC0005F1AB, 0x00021DAE0005D9AD, 0x0002B1B6000215AF,
-    0x8000779E8000779A, 0x800077A20002A1B5, 0x800077A60002B1B3,
-    0x000371BE800077AA, 0x000361BC000379BF, 0x00028DBA000361BD,
-    0x800077AE000371BB, 0x00027DB0800077B2, 0x000381B2000381B1,
-    0x000381B4000381B3, 0x0003B1B6000389B5, 0x0003E1B80003B1B7,
-    0x0003E1BA0003E1B9, 0x0003E1BC0003E1BB, 0x0003D9BE0003EDBD,
-    0x000DE0860003D5BF, 0x800077B6000CE887, 0x800077BE800077BA,
-    0x8000761E800077C2, 0x000235A8800077C6, 0x00020DAA000205A9,
-    0x00021DAC000205AB, 0x00020DAE000205AD, 0x0003E5A3000205AF,
-    0x800077CE800077CA, 0x800077D6800077D2, 0x0003F5A50003E5A6,
-    0x000225AB800077DA, 0x000228840003D9AA, 0x00022DAF000D20BE,
-    0x000235AD000225AE, 0x800077E2000235AC, 0x800077EA800077E6,
-    0x000CDCBE800077EE, 0x800077F2000190E1, 0x800077F60000E8E3,
-    0x800077FE800077FA, 0x00007D8000003CEF, 0x0000658200006581,
-    0x8000780680007802, 0x8000780E8000780A, 0x8000781680007812,
-    0x8000781E8000781A, 0x800078220006C4EF, 0x003FA0BF000624E1,
-    0x800078260006F0E3, 0x8000782A000DBC84, 0x800078328000782E,
-    0x8000783A80007836, 0x8000783E000271B3, 0x0001D1B60001D1B5,
-    0x8000784680007842, 0x0001C9BA8000784A, 0x0001B5BC0001C9BB,
-    0x0001A9BE0001BDBD, 0x000DADA80001A5BF, 0x000DB5AA000DBDA9,
-    0x000E15AC000E0DAB, 0x000E09AE000E1DAD, 0x800077DE000E01AF,
-    0x000FC981000FC182, 0x8000784E000FF180, 0x0000988680007852,
-    0x000E3DB800036087, 0x000E85BA000E8DB9, 0x000E85BC000E9DBB,
-    0x000F5DBE000E89BD, 0x000E41B0000F55BF, 0x000E2DB2000E41B1,
-    0x000E21B4000E25B3, 0x000E25B6000E2DB5, 0x000D3DA3000E05B7,
-    0x8000785A80007856, 0x800078628000785E, 0x000E9DA5000E9DA6,
-    0x000E85AB80007866, 0x8000786A000E85AA, 0x000EE9AF8000786E,
-    0x000EF1AD000EE5AE, 0x80007872000EF9AC, 0x80007876000F8DB3,
-    0x000FB1B68000787A, 0x800078828000787E, 0x000F9DBA000F85B5,
-    0x80007886000F65BB, 0x000F7DBE8000788A, 0x000F7DBC000F65BF,
-    0x000E49A3000F75BD, 0x800078928000788E, 0x8000789A80007896,
-    0x000E41A5000E75A6, 0x000EA1AB8000789E, 0x800078A2000E59AA,
-    0x000EA1AF800078A6, 0x000EB1AD000EB9AE, 0x800078AA000EB9AC,
-    0x800078B2800078AE, 0x000C84EF800078B6, 0x0000158100003182,
-    0x000C58E300001580, 0x000EC0E1800078BA, 0x000138BE00012084,
-    0x0000D886800078C2, 0x000EC1A800005C87, 0x000EC1AA000EC1A9,
-    0x000EC9AC000EC1AB, 0x000EFDAE000EF5AD, 0x0000A084000EF5AF,
-    0x800078CA800078C6, 0x800078D2800078CE, 0x800078DA800078D6,
-    0x000E61B8800078DE, 0x000E71BA000E75B9, 0x000E11BC000E71BB,
-    0x000E11BE000E11BD, 0x000E9DB0000E0DBF, 0x000E6DB2000E65B1,
-    0x000E7DB4000E65B3, 0x000E6DB6000E65B5, 0x000235A8000E65B7,
-    0x000281AA000281A9, 0x000281AC000281AB, 0x0002B1AE000289AD,
-    0x01DD28BF0002B1AF, 0x800078E6800078E2, 0x800078EA0005E084,
-    0x800078F2800078EE, 0x000349B8800078F6, 0x000359BA000349B9,
-    0x000379BC000351BB, 0x000319BE000379BD, 0x0002A1B0000319BF,
-    0x0002A5B20002ADB1, 0x0002A9B40002B9B3, 0x000295B600029DB5,
-    0x000209B3000379B7, 0x800078FE800078FA, 0x8000790680007902,
-    0x000209B5000209B6, 0x000269BB8000790A, 0x8000790E000265BA,
-    0x000259BF80007912, 0x00026DBD00026DBE, 0x80007916000271BC,
-    0x8000791A00024DA3, 0x00024DA68000791E, 0x8000792680007922,
-    0x000221AA00024DA5, 0x8000792A00022DAB, 0x000229AE8000776E,
-    0x000235AC00021DAF, 0x000275A8000229AD, 0x0002A1AA0002A1A9,
-    0x0002A1AC0002A1AB, 0x0002A1AE0002A1AD, 0x0001F5800002A1AF,
-    0x0001C1820001FD81, 0x800079320005A0BE, 0x0005748780007936,
-    0x0001CDB800051C86, 0x0001D5BA0001D5B9, 0x0001FDBC0001E5BB,
-    0x000191BE000195BD, 0x000145B0000191BF, 0x000159B200014DB1,
-    0x000149B4000159B3, 0x0001FDB6000149B5, 0x8000793A0001F5B7,
-    0x000484BE8000793E, 0x80007942000004EF, 0x8000794A80007946,
-    0x8000794E000E34EF, 0x80007952000F7CE1, 0x000000EC000E08E3,
-    0x80007956000194E1, 0x8000795A0000C0E3, 0x800079628000795E,
-    0x80007966000265B3, 0x00026DB6000265B5, 0x8000796E8000796A,
-    0x0001CDBA80007972, 0x0001C1BC0001C5BB, 0x0001C1BE0001C1BD,
-    0x0005A9A30001C1BF, 0x800079768000792E, 0x8000797E8000797A,
-    0x0005A9A50005A1A6, 0x000609AB80007982, 0x80007986000601AA,
-    0x00060DAF8000798A, 0x00060DAD00060DAE, 0x8000798E00060DAC,
-    0x00001D8100001D82, 0x8000799200001D80, 0x8000799A80007996,
-    0x0000C0868000799E, 0x800079A200033487, 0x800079AA800079A6,
-    0x800079B2800079AE, 0x0007B1A8800079B6, 0x000709AA0007B1A9,
-    0x000701AC000709AB, 0x000701AE000701AD, 0x0003A88400073DAF,
-    0x800079BE800079BA, 0x800079C6800079C2, 0x800079CE800079CA,
-    0x000725B8800079D2, 0x0000C5BA0000CDB9, 0x0000CDBC0000DDBB,
-    0x0000F5BE0000FDBD, 0x000749B000009DBF, 0x000729B200075DB1,
-    0x000739B4000729B3, 0x000725B6000725B5, 0x000681B300071DB7,
-    0x800079DA800079D6, 0x800079E2800079DE, 0x000689B5000681B6,
-    0x000699BB800079E6, 0x800079EA00068DBA, 0x0006E1BF800079EE,
-    0x0006FDBD0006FDBE, 0x800079F20006FDBC, 0x800079FA800079F6,
-    0x00003180800079FE, 0x0000298200003D81, 0x80007A02000488EF,
-    0x80007A06000604E1, 0x80007A0A0006DCE3, 0x80007A0E000174BE,
-    0x000601A380007A12, 0x0000288680007A16, 0x80007A1A00014487,
-    0x000609A5000601A6, 0x000619AB80007A1E, 0x80007A2200060DAA,
-    0x000661AF80007A26, 0x00067DAD00067DAE, 0x80007A2A00067DAC,
-    0x80007A2E0001B5B3, 0x0001ADB680007A32, 0x80007A3A80007A36,
-    0x00014DBA0001A5B5, 0x80007A3E000155BB, 0x00014DBE80007A42,
-    0x00014DBC0000F1BF, 0x000269A8000145BD, 0x000279AA000269A9,
-    0x000269AC000279AB, 0x0002BDAE000269AD, 0x003C00840002B5AF,
-    0x80007A4A80007A46, 0x80007A5280007A4E, 0x80007A5A80007A56,
-    0x000351B880007A5E, 0x000361BA000359B9, 0x000311BC000361BB,
-    0x000311BE000311BD, 0x0002CDB0000311BF, 0x0002DDB20002D5B1,
-    0x0002CDB40002D5B3, 0x000371B6000371B5, 0x80007A62000371B7,
-    0x80007A6A80007A66, 0x80007A6E0002FDA3, 0x0002E5A60002EDA5,
-    0x80007A7680007A72, 0x000205AA80007A7A, 0x000205AC00021DAB,
-    0x000205AE00020DAD, 0x000065820003B9AF, 0x000055800002E4EF,
-    0x0003148700006D81, 0x003C54BE003D7486, 0x800078BE80007A82,
-    0x80007A8A80007A86, 0x80007A9280007A8E, 0x80007A960002C4E3,
-    0x00027DA800016CE1, 0x000299AA000281A9, 0x0002B9AC000291AB,
-    0x0002D9AE0002B9AD, 0x80007A9A0002D5AF, 0x80007AA280007A9E,
-    0x80007AAA80007AA6, 0x80007AB280007AAE, 0x000171B880007AB6,
-    0x000171BA000171B9, 0x0001D1BC000171BB, 0x0001D1BE0001D1BD,
-    0x0002ADB00001CDBF, 0x0002B1B20002B1B1, 0x00029DB4000285B3,
-    0x000151B6000151B5, 0x0006A4E1000151B7, 0x0007B4E380007ABA,
-    0x003D1884003D80BE, 0x80007AC280007ABE, 0x80007ACA80007AC6,
-    0x80007AD280007ACE, 0x80007ADA80007AD6, 0x0006C8EF80007ADE,
-    0x0000098180007AE2, 0x80007AE600007980, 0x000160E300001982,
-    0x0001A0E180007AEA, 0x00006CEF80007AF2, 0x003CC88680007AF6,
-    0x80007AFA003D0C87, 0x80007B0280007AFE, 0x003D91A380007B06,
-    0x80007AEE80007A7E, 0x80007B0E80007B0A, 0x003DB1A5003DB9A6,
-    0x003DF5AB80007B12, 0x80007B16003DE9AA, 0x003E15AF80007B1A,
-    0x003DD1AD003DD9AE, 0x80007B1E003DE5AC, 0x80007B22003EADB3,
-    0x003EA9B680007B26, 0x80007B2E80007B2A, 0x003F49BA003EADB5,
-    0x80007B32003F49BB, 0x003F49BE80007B36, 0x003F49BC003F49BF,
-    0x003E69A3003F49BD, 0x80007B3E80007B3A, 0x80007B4680007B42,
-    0x003E69A5003E6DA6, 0x003F8DAB80007B4A, 0x80007B4E003F8DAA,
-    0x003F8DAF80007B52, 0x003F8DAD003F8DAE, 0x80007B56003F8DAC,
-    0x003D05BA80007B5A, 0x003D09BC003D19BB, 0x003D35BE003D3DBD,
-    0x00004D80003EF9BF, 0x00002D8200002581, 0x000038BE003D7DB3,
-    0x003D55B6003D5DB5, 0x003FDDA800003084, 0x003FE1AA003FE1A9,
-    0x003FE1AC003FE1AB, 0x003E39AE003FEDAD, 0x80007B5E003E39AF,
-    0x0000888680007B62, 0x80007B660001AC87, 0x80007B6E80007B6A,
-    0x003ED9B880007B72, 0x003EE9BA003ED9B9, 0x003EF9BC003EE9BB,
-    0x003EE9BE003EF9BD, 0x003E49B0003EE9BF, 0x003E59B2003E49B1,
-    0x003E49B4003E59B3, 0x003EFDB6003E49B5, 0x003E5DA8003EE9B7,
-    0x003E79AA003E61A9, 0x003E69AC003E79AB, 0x000195AE003E69AD,
-    0x0001A08400018DAF, 0x80007B7A80007B76, 0x80007B8280007B7E,
-    0x80007B8A80007B86, 0x0001A9B880007B8E, 0x00014DBA0001A9B9,
-    0x00015DBC000145BB, 0x000145BE000145BD, 0x0001FDB0000175BF,
-    0x0001C5B20001CDB1, 0x0001B9B40001ADB3, 0x0001ADB60001B9B5,
-    0x003DF1B30001A5B7, 0x000344BE80007B92, 0x80007B96000444BE,
-    0x003DE1B5003DE9B6, 0x000209BB80007B9A, 0x80007B9E000239BA,
-    0x00027DBF80007BA2, 0x000279BD00027DBE, 0x80007BA6000211BC,
-    0x80007BAA003DB5A3, 0x003DADA680007BAE, 0x80007BB680007BB2,
-    0x00027DAA003DA5A5, 0x80007BBA00024DAB, 0x000239AE80007BBE,
-    0x000255AC000239AF, 0x0000118000023DAD, 0x0000118200001181,
-    0x80007BC200060CBE, 0x80007BCA80007BC6, 0x80007BD200026CEF,
-    0x80007BD6000194E1, 0x80007BDA00011CE3, 0x0004188680007BDE,
-    0x80007BE200037087, 0x80007BEA80007BE6, 0x80007BEE0003F9B3,
-    0x00037DB60003E9B5, 0x0007ACBE80007BF2, 0x000359BA80007BF6,
-    0x000349BC000359BB, 0x0000FDBE000349BD, 0x000225A80000EDBF,
-    0x00023DAA000235A9, 0x00022DAC000235AB, 0x0002B1AE0002B1AD,
-    0x000968BE0002B1AF, 0x80007BFE80007BFA, 0x000AF8BE80007C02,
-    0x80007C0A80007C06, 0x00016DB880007C0E, 0x000101BA000105B9,
-    0x000105BC000101BB, 0x000131BE00010DBD, 0x0002D1B0000131BF,
-    0x0002D1B20002D1B1, 0x000175B40002D1B3, 0x000165B6000165B5,
-    0x003F4CE1000155B7, 0x003E08E3000168E3, 0x80007C12003E7CE1,
-    0x80007C1A80007C16, 0x000D2CBE80007C1E, 0x80007C2680007C22,
-    0x000C3CBE80007C2A, 0x003ECCEF0001BCEF, 0x0002F5A380007C2E,
-    0x0000158100002D82, 0x80007C3200001D80, 0x0002E5A5000271A6,
-    0x000255AB80007C36, 0x80007C3A000255AA, 0x0001E1AF0005A084,
-    0x000245AD0001F1AE, 0x0006A1A8000245AC, 0x0006B9AA0006ADA9,
-    0x0006D1AC0006B9AB, 0x0006D1AE0006D1AD, 0x80007C3E0006CDAF,
-    0x0003088680007C42, 0x80007C4600030887, 0x80007C4E80007C4A,
-    0x0007D1B880007C52, 0x0007D1BA0007D1B9, 0x0007F5BC0007D1BB,
-    0x0007F5BE0007FDBD, 0x0006BDB00007E1BF, 0x00074DB2000681B1,
-    0x00075DB4000745B3, 0x000741B6000741B5, 0x000631B3000741B7,
-    0x80007C5A80007C56, 0x80007C6280007C5E, 0x000659B5000651B6,
-    0x000645BB80007C66, 0x80007C6A000679BA, 0x000785BF80007C6E,
-    0x000791BD000799BE, 0x80007C72000799BC, 0x80007C76000675A3,
-    0x000615A680007C7A, 0x80007C8280007C7E, 0x00063DAA00061DA5,
-    0x80007C86000601AB, 0x0007DDAE80007C8A, 0x0007DDAC0007C1AF,
-    0x0005A0E10007D5AD, 0x0004F4E380007C8E, 0x80007C9680007C92,
-    0x80007C9E80007C9A, 0x0000098180007CA2, 0x80007CA600003980,
-    0x80007CAA00001D82, 0x0005CCEF80007CAE, 0x000175B30002B084,
-    0x0000E88680007CB2, 0x80007CB600012C87, 0x000165B50001D1B6,
-    0x0001FDBB80007CBA, 0x80007CBE0001F5BA, 0x0001B9BF80007CC2,
-    0x0001D5BD0001B5BE, 0x0001A5A80001D5BC, 0x0001BDAA0001B5A9,
-    0x0001ADAC0001B5AB, 0x000111AE000111AD, 0x0001AC84000111AF,
-    0x80007CCA80007CC6, 0x80007CD280007CCE, 0x80007CDA80007CD6,
-    0x000119B880007CDE, 0x000125BA00012DB9, 0x0000DDBC0000C9BB,
-    0x0000CDBE0000C5BD, 0x000171B00000C5BF, 0x000141B2000179B1,
-    0x000131B4000141B3, 0x000135B600013DB5, 0x80007CE2000129B7,
-    0x80007CEA80007CE6, 0x80007CEE000239A3, 0x00029DA6000229A5,
-    0x80007CF680007CF2, 0x0002B9AA80007CFA, 0x000299AC0002B1AB,
-    0x0002F9AE000299AD, 0x80007CFE0002F5AF, 0x000CE88680007D02,
-    0x80007D0600036887, 0x80007D0E80007D0A, 0x00003D8080007D12,
-    0x0000198200000981, 0x80007D1A80007D16, 0x80007D2680007D22,
-    0x80007D2E80007D2A, 0x000690EF80007BCE, 0x80007D3680007D32,
-    0x80007D3E80007D3A, 0x000CEC8480007D42, 0x80007D4A80007D46,
-    0x0001E0E180007D4E, 0x000108E380007D52, 0x00039DB380007D56,
-    0x80007D5E80007D5A, 0x80007D6680007D62, 0x0003B5B5000371B6,
-    0x000355BB80007D6A, 0x80007D6E000355BA, 0x0000F5BF80007D72,
-    0x000345BD0000F5BE, 0x000245A8000345BC, 0x000255AA000255A9,
-    0x00027DAC000265AB, 0x0002B1AE0002B1AD, 0x00001D810002B1AF,
-    0x80007D7600001D80, 0x80007D7A0000AD82, 0x80007D8280007D7E,
-    0x00015DB880007D86, 0x00016DBA000165B9, 0x000115BC00010DBB,
-    0x00010DBE000105BD, 0x0002D1B0000105BF, 0x0002D1B20002D1B1,
-    0x000171B40002D1B3, 0x000171B6000171B5, 0x80007D8A00016DB7,
-    0x80007D92002BD0E1, 0x000F90860017D8E3, 0x80007D96000D5C87,
-    0x80007D9A0007E8EF, 0x80007D9E000194E1, 0x000BB0BE000080E3,
-    0x80007DA680007DA2, 0x00029DA3003F60EF, 0x80007DAE80007DAA,
-    0x80007DB680007DB2, 0x0002B5A5000271A6, 0x000255AB80007DBA,
-    0x80007DBE000255AA, 0x0001F5AF80007DC2, 0x000245AD0001F5AE,
-    0x000E75A8000245AC, 0x000E75AA000E7DA9, 0x000E9DAC000E8DAB,
-    0x000E8DAE000E85AD, 0x80007D8E000E85AF, 0x80007DCA80007DC6,
-    0x80007DD280007DCE, 0x80007DDA80007DD6, 0x000E8DB880007DDE,
-    0x000E95BA000E95B9, 0x000EBDBC000EA5BB, 0x000F69BE000F7DBD,
-    0x000EFDB0000F69BF, 0x000EC1B2000EC1B1, 0x000EC5B4000EC1B3,
-    0x000EBDB6000EC9B5, 0x000EFDB3000EB5B7, 0x80007DE680007DE2,
-    0x80007DEE80007DEA, 0x000E15B5000E05B6, 0x000E05BB80007DF2,
-    0x80007DF6000E39BA, 0x000E0DBF80007DFA, 0x000E15BD000E01BE,
-    0x00006D82000E15BC, 0x00005580000EB9A3, 0x000E41A600006581,
-    0x80007E0200011884, 0x000E7DAA000E51A5, 0x00040086000E41AB,
-    0x000E45AE00007887, 0x000E51AC000E49AF, 0x0025298F000E51AD,
-    0x80007E060003F4E1, 0x80007E0A000154E3, 0x80007E0E000DA4E3,
-    0x80007E12000D4CE1, 0x80007E160008FCEF, 0x80007E1E80007E1A,
-    0x80007E2680007E22, 0x0008619F000864EF, 0x00094D9D00094D9E,
-    0x00350D9B0034359C, 0x0037B9990035519A, 0x0031419700316D98,
-    0x003DB99500338996, 0x003FA993003D4D94, 0x00397D910039AD92,
-    0x0015A5B100397D90, 0x01E989B30015A5B0, 0x01E855B5001441B2,
-    0x80007E2A01E989B4, 0x80007E3280007E2E, 0x80007E3A80007E36,
-    0x000329A380007E3E, 0x000D99A1000C81A2, 0x0001E1A1000F39A0,
-    0x0005CDA380007E42, 0x000481A500034DA2, 0x0019D5A70005DDA4,
-    0x001DE1A90019DDA6, 0x001CE5AB00188DA8, 0x0011D1AD001DE9AA,
-    0x001401AF001001AC, 0x00262982001151AE, 0x80007E46003A0583,
-    0x003CC98680007E4A, 0x003A1D84003E3587, 0x003EF18A003AC585,
-    0x80007E4E0030C98B, 0x0034D18E80007E52, 0x0032198C0036318F,
-    0x00361D9200320D8D, 0x0003448400373D93, 0x000B71960004C484,
-    0x000A5D94000CD197, 0x000E419A000A5595, 0x80007E56000F519B,
-    0x80007E5E80007E5A, 0x0000E19C80007E62, 0x80007DFE80007E66,
-    0x80007E7280007E6A, 0x0000ED8180007E76, 0x80007D1E0000E180,
-    0x000D29A800001982, 0x000D09AA000D3DA9, 0x000D05AC000D05AB,
-    0x000D7DAE000D0DAD, 0x000D0DB0000D75AF, 0x000D15B2000D15B1,
-    0x000D3DB4000D25B3, 0x0002F1B60002F1B5, 0x0002D1B80002F1B7,
-    0x0002D1BA0002D1B9, 0x000275BC0002D1BB, 0x000275BE00027DBD,
-    0x000CA1B300026DBF, 0x0002B48780007E7A, 0x80007E7E00052C86,
-    0x000CADB5000CADB6, 0x000CF1BB80007E82, 0x80007E86000CE9BA,
-    0x000315BF80007E8A, 0x000CD1BD000CD9BE, 0x80007E8E000CE1BC,
-    0x80007E92000CE5A3, 0x000CE9A680007E96, 0x80007E9E80007E9A,
-    0x000CADAA000CE9A5, 0x80007EA2000CB5AB, 0x000C9DAE80007EA6,
-    0x000CA5AC000351AF, 0x80007EAA000C95AD, 0x80007EB280007EAE,
-    0x80007EBA80007EB6, 0x80007EC280007EBE, 0x80007EC6000A54EF,
-    0x80007ECA000E2CE1, 0x80007ECE000168E3, 0x0000718100007982,
-    0x0002D1B300007980, 0x80007ED680007ED2, 0x80007EE280007EDA,
-    0x0002F9B50002FDB6, 0x0002D9BB80007EE6, 0x0004C8860002D1BA,
-    0x000255BF0004AC87, 0x000265BD000255BE, 0x000511A3000265BC,
-    0x80007EDE80007E6E, 0x80007EEE80007EEA, 0x000539A500053DA6,
-    0x000519AB80007EF2, 0x80007EF6000511AA, 0x000595AF80007EFA,
-    0x0005A5AD000595AE, 0x80007EFE0005A5AC, 0x80007F02000741B3,
-    0x000779B680007F06, 0x80007F0E80007F0A, 0x000729BA000741B5,
-    0x80007F12000729BB, 0x000719BE80007F16, 0x00072DBC000719BF,
-    0x000609A8000725BD, 0x00078DAA000609A9, 0x00078DAC00079DAB,
-    0x0007B5AE0007BDAD, 0x80007F1A0007DDAF, 0x80007F2280007F1E,
-    0x80007F2A80007F26, 0x80007F3280007F2E, 0x000749B880007F36,
-    0x000759BA000749B9, 0x000749BC000759BB, 0x000771BE000775BD,
-    0x0007A9B000076DBF, 0x0007BDB20007A9B1, 0x000799B40007B5B3,
-    0x000779B6000799B5, 0x000605A3000779B7, 0x80007F3E80007F3A,
-    0x80007F4680007F42, 0x000605A500063DA6, 0x00066DAB80007F4A,
-    0x0002A88400066DAA, 0x00065DAF80007F52, 0x000661AD00065DAE,
-    0x00004D80000669AC, 0x0000558200005581, 0x80007F4E000141B3,
-    0x000171B6000149B5, 0x0000C08680007F56, 0x000121BA0001A887,
-    0x000125BC000121BB, 0x000111BE00012DBD, 0x000651A8000111BF,
-    0x000185AA000659A9, 0x00018DAC00019DAB, 0x0001B5AE0001BDAD,
-    0x0000A0840001DDAF, 0x80007F5E80007F5A, 0x80007F6680007F62,
-    0x80007F6E80007F6A, 0x000169B880007F72, 0x000179BA000169B9,
-    0x000169BC000179BB, 0x0000DDBE000169BD, 0x0001A9B00000D5BF,
-    0x0001BDB20001A9B1, 0x000199B40001B5B3, 0x000159B6000199B5,
-    0x00040DBA000159B7, 0x0007CDB800042DBB, 0x00047DBE000405B9,
-    0x000435BC000465BF, 0x0007B9B2000475BD, 0x00071DB00007B9B3,
-    0x000799B6000715B1, 0x000799B4000799B7, 0x000751AA000799B5,
-    0x000691A800075DAB, 0x00077DAE000745A9, 0x000779AC000765AF,
-    0x80007F76000775AD, 0x80007F7E80007F7A, 0x80007F8680007F82,
-    0x80007F8E80007F8A, 0x80007F9680007F92, 0x001C00BE80007F9A,
-    0x80007F9E000281A3, 0x0002B1A6000289A5, 0x80007FA680007FA2,
-    0x0002E1AA80007FAA, 0x0002E5AC0002E1AB, 0x0002D1AE0002EDAD,
-    0x80007FAE0002D1AF, 0x0000118100002582, 0x80007FB200001D80,
-    0x80007FB60001ACE1, 0x80007FBA00016CE3, 0x80007FC680007FC2,
-    0x001CB08700029CEF, 0x80007FCA001CD486, 0x00022DA880007FCE,
-    0x0002ADAA000259A9, 0x0002BDAC0002A5AB, 0x0002ADAE0002A5AD,
-    0x001DA0840002A5AF, 0x80007FD680007FD2, 0x80007FDE80007FDA,
-    0x001CB4BE80007FE2, 0x00013DB880007FE6, 0x0001CDBA0001C5B9,
-    0x0001CDBC0001DDBB, 0x0001FDBE0001F5BD, 0x0002DDB000019DBF,
-    0x00010DB2000105B1, 0x00011DB4000105B3, 0x00010DB6000105B5,
-    0x000694E1000105B7, 0x0007D4E30000BCE3, 0x80007FEA00067CE1,
-    0x80007FF280007FEE, 0x80007FFA80007FF6, 0x8000800280007FFE,
-    0x001F78BE80008006, 0x000EF0EF000AFCEF, 0x8000800E8000800A,
-    0x8000801680008012, 0x8000801A000269B3, 0x000269B6000279B5,
-    0x800080228000801E, 0x000255BA80008026, 0x000195BC00018DBB,
-    0x000195BE00019DBD, 0x80007FBE00018DBF, 0x8000802A001DA1A3,
-    0x001DA1A68000802E, 0x8000803680008032, 0x001D9DAA001DB1A5,
-    0x00030886001E45AB, 0x001E5DAE00037C87, 0x001E5DAC001E45AF,
-    0x001FA180001E55AD, 0x00004182001FA181, 0x8000803A001EB5B3,
-    0x001F71B6001F71B5, 0x800080428000803E, 0x001F5DBA80008046,
-    0x001FE1BC001F51BB, 0x001FE1BE001FE1BD, 0x001E65A8001FE1BF,
-    0x001E71AA001E6DA9, 0x001FB5AC001E71AB, 0x001FB5AE001FBDAD,
-    0x8000804A001FADAF, 0x800080528000804E, 0x8000805A80008056,
-    0x800080628000805E, 0x001F69B880008066, 0x001F75BA001F69B9,
-    0x001F15BC001F0DBB, 0x001F09BE001F1DBD, 0x001FD5B0001F01BF,
-    0x001F79B2001FD9B1, 0x001F69B4001F79B3, 0x001F59B6001F69B5,
-    0x001FF9A3001F59B7, 0x8000806E8000806A, 0x8000807680008072,
-    0x001E3DA5001E3DA6, 0x001E1DAB8000807A, 0x8000807E001E11AA,
-    0x001EADAF80008082, 0x001EADAD001EADAE, 0x80008086001EADAC,
-    0x8000808E8000808A, 0x00001D8080008092, 0x0000398200000981,
-    0x8000809A80008096, 0x8000809E0001A884, 0x0000D086800080A2,
-    0x800080A60001FC87, 0x000195A8800080AA, 0x0001D9AA00019DA9,
-    0x0001F1AC0001D1AB, 0x000139AE0001F1AD, 0x0000A084000139AF,
-    0x800080B2800080AE, 0x800080BA800080B6, 0x800080C2800080BE,
-    0x0000CDB8800080C6, 0x0000D5BA0000D5B9, 0x0000FDBC0000E5BB,
-    0x00009DBE000095BD, 0x000149B0000095BF, 0x000159B2000149B1,
-    0x000149B4000159B3, 0x0000FDB6000149B5, 0x000251B30000F5B7,
-    0x800080CE800080CA, 0x800080D6800080D2, 0x000279B5000271B6,
-    0x00028DBB800080DA, 0x800080DE00028DBA, 0x00028DBF800080E2,
-    0x00028DBD00028DBE, 0x01F43C8500028DBC, 0x800080E6000215A3,
-    0x000235A6800080EA, 0x800080F2800080EE, 0x0002C9AA00023DA5,
-    0x800080F60002C9AB, 0x0002C9AE800080FA, 0x0002C9AC0002C9AF,
-    0x000019820002C9AD, 0x00001980800080FE, 0x8000810200001981,
-    0x8000810A80008106, 0x000400BE80008112, 0x8000811A80008116,
-    0x800081228000811E, 0x8000812A80008126, 0x0004E0868000812E,
-    0x8000813200035487, 0x8000813A80008136, 0x8000813E00028CEF,
-    0x8000814680008142, 0x8000814A000154E3, 0x8000814E000184E1,
-    0x8000815680008152, 0x00026DA88000815A, 0x000275AA00027DA9,
-    0x0002B9AC0002A9AB, 0x0002A9AE0002B9AD, 0x0005AC840002A9AF,
-    0x800081628000815E, 0x8000816A80008166, 0x800081728000816E,
-    0x00016DB880008176, 0x000105BA00010DB9, 0x000109BC000119BB,
-    0x000135BE00013DBD, 0x0002D9B00001D9BF, 0x00016DB20002D9B1,
-    0x00017DB4000165B3, 0x000165B6000165B5, 0x0000F4E3000155B7,
-    0x0001B8E1001F68E1, 0x0000A980001EBCE3, 0x000005820000A981,
-    0x00042CBE8000817A, 0x8000818680008182, 0x001EF8EF8000818A,
-    0x800081928000818E, 0x0004E086001E7CEF, 0x8000819600057487,
-    0x8000819A0002B1B3, 0x000275B600027DB5, 0x800081A28000819E,
-    0x000251BA800081A6, 0x0001F5BC000251BB, 0x0001F5BE0001FDBD,
-    0x8000810E0001EDBF, 0x0005B1AA8000817E, 0x000615AC0005B1AB,
-    0x000615AE00061DAD, 0x800081AA00060DAF, 0x800081B2800081AE,
-    0x800081B6000551A3, 0x000595A600059DA5, 0x0006E9B3800081BA,
-    0x800081C2800081BE, 0x800081CA800081C6, 0x0006E9B50006E1B6,
-    0x0006ADBB800081CE, 0x800081D20006ADBA, 0x000691BF800081D6,
-    0x0006A1BD0006A9BE, 0x00061DA80006ADBC, 0x000621AA000621A9,
-    0x000621AC000621AB, 0x000621AE000621AD, 0x800081DA000621AF,
-    0x800081E2800081DE, 0x00001980800081E6, 0x0000058200001981,
-    0x000635B8800081EA, 0x000635BA00063DB9, 0x0007F9BC0007E9BB,
-    0x0007E9BE0007F9BD, 0x000629B00007E9BF, 0x000639B2000629B1,
-    0x00062DB4000635B3, 0x000611B6000615B5, 0x0007A9A300060DB7,
-    0x000184BE00028C84, 0x800081F2800081EE, 0x0007A9A50007A1A6,
-    0x0007EDAB800081F6, 0x000088860007EDAA, 0x0007D1AF00017887,
-    0x0007E1AD0007E9AE, 0x800081FA0007EDAC, 0x800081FE0006E9B3,
-    0x000699B680008202, 0x8000820A80008206, 0x000681BA0006E5B5,
-    0x8000820E000681BB, 0x000159BE80008212, 0x000159BC000145BF,
-    0x000625A8000159BD, 0x000625AA00062DA9, 0x000629AC000639AB,
-    0x000649AE00065DAD, 0x0001AC84000641AF, 0x8000821A80008216,
-    0x800082228000821E, 0x8000822A80008226, 0x000189B88000822E,
-    0x000199BA000189B9, 0x0001B9BC000191BB, 0x000179BE0001B9BD,
-    0x000601B0000179BF, 0x0001C9B2000601B1, 0x0001D9B40001C9B3,
-    0x0001C5B60001C5B5, 0x0005ADA30001B9B7, 0x8000823680008232,
-    0x8000823E8000823A, 0x0005A1A50005DDA6, 0x0005C5AB80008242,
-    0x800082460005C5AA, 0x000201AF8000824A, 0x00021DAD00021DAE,
-    0x8000824E00021DAC, 0x8000825680008252, 0x000019818000825A,
-    0x8000826200001180, 0x000324BE00001182, 0x000CC88680008266,
-    0x8000826A00034C87, 0x800082728000826E, 0x000130E380008276,
-    0x0001B8E18000827A, 0x8000827E000CC084, 0x8000828680008282,
-    0x000CE8BE8000828A, 0x800082928000828E, 0x000228EF80008296,
-    0x8000829E8000829A, 0x000381B3800082A2, 0x800082AA800082A6,
-    0x800082B2800082AE, 0x000395B5000381B6, 0x000345BB800082B6,
-    0x800082BA000389BA, 0x000345BF800082BE, 0x000355BD000345BE,
-    0x000211A8000355BC, 0x000249AA000219A9, 0x00025DAC000245AB,
-    0x000241AE000241AD, 0x00000D80000275AF, 0x0000118200001181,
-    0x000C64BE800082C2, 0x800082CA800082C6, 0x000145B8800082D2,
-    0x000145BA00014DB9, 0x00014DBC00015DBB, 0x000175BE00017DBD,
-    0x00020DB000011DBF, 0x0002C1B20002C1B1, 0x0002C1B40002C1B3,
-    0x0002C1B60002C1B5, 0x000658E10002C1B7, 0x00009CE300075CE1,
-    0x000C30860006BCE3, 0x800082D6000D5C87, 0x800082DE800082DA,
-    0x800082E6800082E2, 0x000000EE800082EA, 0x000048EF800082EE,
-    0x800082F20006E4EF, 0x000289A3800082F6, 0x00029DA5800082FA,
-    0x80008302800082FE, 0x80008306000289A6, 0x00024DAB8000830A,
-    0x00025DAD000281AA, 0x00024DAF00025DAC, 0x000E59A800024DAE,
-    0x000E79AA000E61A9, 0x000E69AC000E79AB, 0x000E9DAE000E69AD,
-    0x800082CE000E8DAF, 0x800083128000830E, 0x8000831A80008316,
-    0x800083228000831E, 0x000EFDB880008326, 0x000E81BA000E81B9,
-    0x000E81BC000E81BB, 0x000EB1BE000E89BD, 0x000EF5B0000EB1BF,
-    0x000EF5B2000EFDB1, 0x000ED9B4000EC9B3, 0x000ECDB6000ED9B5,
-    0x000EF1B3000EC5B7, 0x8000832E8000832A, 0x8000833680008332,
-    0x000EE1B5000EE9B6, 0x000E39BB8000833A, 0x0003E484000E31BA,
-    0x000E1DBF8000833E, 0x000E1DBD000E1DBE, 0x00005D82000E21BC,
-    0x00004580000EB5A3, 0x000EADA600005D81, 0x8000834680008342,
-    0x000E75AA000EA5A5, 0x0000E886000E7DAB, 0x000E59AE0000E087,
-    0x000E65AC000E59AF, 0x000E3DA8000E59AD, 0x000E95AA000E9DA9,
-    0x000EB5AC000EADAB, 0x000EB5AE000EBDAD, 0x8000834A000EA9AF,
-    0x800083528000834E, 0x8000835A80008356, 0x800083628000835E,
-    0x000169B880008366, 0x000179BA000169B9, 0x000169BC000179BB,
-    0x0001DDBE000169BD, 0x000EA1B00001D5BF, 0x000EA5B2000EA1B1,
-    0x000EA9B4000EB9B3, 0x000E9DB6000E95B5, 0x000E1DB3000159B7,
-    0x8000836E8000836A, 0x8000837680008372, 0x000E3DB5000E3DB6,
-    0x000E65BB8000837A, 0x8000837E000E65BA, 0x000E51BF80008382,
-    0x000E6DBD000E6DBE, 0x000E65A5000E71BC, 0x8000838A80008386,
-    0x8000838E000E65A6, 0x000E45A380008392, 0x000E35AD80008396,
-    0x000E09AF000E29AC, 0x8000839A000E35AE, 0x000E3DAB8000825E,
-    0x8000839E000E3DAA, 0x00001D8100001D82, 0x800083A200001D80,
-    0x800083AA800083A6, 0x800083B2000350BE, 0x800083BA800083B6,
-    0x800083C2800083BE, 0x800083CA800083C6, 0x0003C084800083CE,
-    0x000E10EF800083D2, 0x0004F08600040084, 0x800083D600035487,
-    0x800083DE800083DA, 0x800083E6800083E2, 0x000E00E1800083EA,
-    0x000EE4E3800083EE, 0x800083F6800083F2, 0x800083FE800083FA,
-    0x8000840680008402, 0x8000840E8000840A, 0x000275A880008412,
-    0x000389AA00027DA9, 0x000399AC000389AB, 0x000389AE000399AD,
-    0x0003F9B0000389AF, 0x00034DB20003F9B1, 0x00035DB4000345B3,
-    0x000345B6000345B5, 0x00034DB8000375B7, 0x000321BA000321B9,
-    0x000325BC00033DBB, 0x000309BE000329BD, 0x00019D800000FDBF,
-    0x00001D8200000981, 0x0003F4EF80008416, 0x000550BE8000841A,
-    0x000370E380008422, 0x0001D8E180008426, 0x8000842E8000842A,
-    0x8000843680008432, 0x0004E0868000843A, 0x8000843E00028887,
-    0x8000844200026DB3, 0x000275B600027DB5, 0x8000844A80008446,
-    0x000251BA8000844E, 0x0002F5BC000251BB, 0x0002F5BE0002FDBD,
-    0x800084520002EDBF, 0x8000845A80008456, 0x800084628000845E,
-    0x8000846A80008466, 0x0004EC84000E54EF, 0x8000846E000E2CE1,
-    0x80008472000114E3, 0x8000847A80008476, 0x0002E1A38000847E,
-    0x0000158100002D82, 0x8000848200001D80, 0x0002F1A50002F9A6,
-    0x0002DDAB80008486, 0x8000848A0002DDAA, 0x000261AF0004B0BE,
-    0x000271AD000279AE, 0x000621A8000279AC, 0x000621AA000621A9,
-    0x000621AC000621AB, 0x000621AE000621AD, 0x8000841E000621AF,
-    0x000000868000848E, 0x800084920003A087, 0x8000849A80008496,
-    0x00063DB88000849E, 0x0006CDBA0006C5B9, 0x0006DDBC0006C5BB,
-    0x0006CDBE0006C5BD, 0x000625B00006C5BF, 0x000635B200063DB1,
-    0x000615B400060DB3, 0x000615B600061DB5, 0x0006A9B3000609B7,
-    0x800084A6800084A2, 0x800084AE800084AA, 0x0006A9B50006A1B6,
-    0x0006F9BB800084B2, 0x800084B60006E1BA, 0x0006D9BF800084BA,
-    0x0006D9BD0006D1BE, 0x800084BE0006E1BC, 0x800084C20006EDA3,
-    0x0006E5A6800084C6, 0x800084CE800084CA, 0x0006A5AA0006EDA5,
-    0x800084D20006BDAB, 0x000695AE800084D6, 0x0006A5AC00069DAF,
-    0x00061DA800069DAD, 0x000625AA000621A9, 0x000625AC00063DAB,
-    0x000625AE00062DAD, 0x800084DA000655AF, 0x800084E2800084DE,
-    0x800084EA800084E6, 0x800084F2800084EE, 0x0006D1B8800084F6,
-    0x0006D1BA0006D1B9, 0x0006E1BC0006E5BB, 0x0006E1BE0006E1BD,
-    0x000631B00006E1BF, 0x000609B2000639B1, 0x0006F1B4000609B3,
-    0x0006F1B60006F1B5, 0x0006ADB30006F1B7, 0x0000B58100001582,
-    0x800084FA0000B580, 0x0006ADB50006ADB6, 0x0006FDBB000010BE,
-    0x00016C840006F5BA, 0x000635BF80008502, 0x0006DDBD000635BE,
-    0x800085060006DDBC, 0x0006FDA50006FDA6, 0x0006FDA3800084FE,
-    0x01FC00868000850A, 0x000665AF00016087, 0x00068DAD000665AE,
-    0x0006ADAB00068DAC, 0x800083AE0006A5AA, 0x800085128000850E,
-    0x000231B380008516, 0x000295B58000851A, 0x800085228000851E,
-    0x80008526000295B6, 0x000281BB8000852A, 0x0002F1BD0002B1BA,
-    0x0002F5BF000299BC, 0x00021CE30002F5BE, 0x000188E10007A4EF,
-    0x8000852E0003C484, 0x8000853680008532, 0x8000853E8000853A,
-    0x8000854680008542, 0x000228EF8000854A, 0x8000854E000628E3,
-    0x0002B9AA000634E1, 0x0002B9A80002B9AB, 0x00021DAE0002B9A9,
-    0x00021DAC000205AF, 0x0003ECBE000215AD, 0x8000855680008552,
-    0x8000855E8000855A, 0x0001B1810001B982, 0x0003E1BA0001B980,
-    0x000221B80003E1BB, 0x0003E1BE000221B9, 0x0003F1BC0003E1BF,
-    0x00024DB20003F1BD, 0x000275B0000225B3, 0x000225B6000245B1,
-    0x000231B400022DB7, 0x000339A3000231B5, 0x8000856680008562,
-    0x800085728000856A, 0x00039DA500039DA6, 0x000389AB80008576,
-    0x8000857A0003B9AA, 0x0003FDAF8000857E, 0x0003F9AD0003FDAE,
-    0x80008582000391AC, 0x01FDC886000291B3, 0x000155B601FD4087,
-    0x8000858A80008586, 0x000179BA00015DB5, 0x8000858E000145BB,
-    0x000125BE80008592, 0x00015DBC00012DBF, 0x8000859600012DBD,
-    0x8000859E8000859A, 0x800085A201FDC084, 0x0000FCEF800085A6,
-    0x000190E1000688EF, 0x000010E30006FCE1, 0x800085AA00061CE3,
-    0x01FDF0BE800085AE, 0x000215A3800085B2, 0x800085BA800085B6,
-    0x800085C2800085BE, 0x0001D9A50001D1A6, 0x0001C1AB800085C6,
-    0x800085CA0001FDAA, 0x0001A9AF800085CE, 0x0001A9AD0001A1AE,
-    0x8000856E0001D9AC, 0x00001D8100001D82, 0x800085D200001D80,
-    0x800085DA800085D6, 0x01FF31A8800085DE, 0x01FF01AA01FF39A9,
-    0x01FF71AC01FF01AB, 0x01FF61AE01FF61AD, 0x01FF2DB001FF55AF,
-    0x01FFCDB201FFC5B1, 0x01FFC1B401FFC5B3, 0x01FFE9B601FFC1B5,
-    0x01FFDDB801FFE9B7, 0x01FFA1BA01FFA1B9, 0x01FFA1BC01FFA1BB,
-    0x01FF91BE01FFA9BD, 0x01FFA9B301FF91BF, 0x00032886800085E2,
-    0x800085E600031C87, 0x01FFD9B501FFC1B6, 0x01FF21BB800085EA,
-    0x800085EE01FFC9BA, 0x01FF25BF800085F2, 0x01FF31BD01FF25BE,
-    0x800085F601FF39BC, 0x800085FA01FFEDA3, 0x01FF85A6800085FE,
-    0x8000860680008602, 0x01FF8DAA01FF9DA5, 0x8000860A01FF65AB,
-    0x01FF61AE8000860E, 0x01FF7DAC01FF61AF, 0x000185A801FF75AD,
-    0x000195AA000195A9, 0x0001BDAC0001A5AB, 0x0001D1AE0001D5AD,
-    0x800086120001D1AF, 0x8000861A80008616, 0x800086228000861E,
-    0x8000862A80008626, 0x000071B88000862E, 0x000071BA000071B9,
-    0x0000D1BC000071BB, 0x0000D1BE0000D1BD, 0x0001B5B00000D1BF,
-    0x000181B20001BDB1, 0x000051B4000181B3, 0x000051B6000051B5,
-    0x80008632000051B7, 0x8000863A80008636, 0x00000D808000863E,
-    0x00003D8200003581, 0x8000864680008642, 0x0002A4848000864A,
-    0x00019086000120BE, 0x8000865200011487, 0x00015DB380008656,
-    0x8000865A8000864E, 0x800086628000865E, 0x000175B5000115B6,
-    0x000139BB80008666, 0x8000866A000131BA, 0x0000D5BF8000866E,
-    0x000115BD0000D5BE, 0x00029DA3000115BC, 0x8000867680008672,
-    0x8000867A000400BE, 0x0002B5A50002D5A6, 0x0002F9AB8000867E,
-    0x800086820002F1AA, 0x000315AF80008686, 0x0002D5AD000315AE,
-    0x0001E8E10002D5AC, 0x000214E38000868A, 0x8000868E00034084,
-    0x8000869680008692, 0x8000869E8000869A, 0x800086A6800086A2,
-    0x800086AE800086AA, 0x00028CEF800086B2, 0x000339A8800086B6,
-    0x000341AA000339A9, 0x000341AC000341AB, 0x000371AE000349AD,
-    0x0000D580000371AF, 0x000005820000D981, 0x0005B485800086BA,
-    0x800086C6800086C2, 0x000095B8800086CA, 0x000095BA00009DB9,
-    0x0000B5BC0000ADBB, 0x0000B5BE0000BDBD, 0x0000E5B00000ADBF,
-    0x0000F9B20000EDB1, 0x0000D9B40000F1B3, 0x0000BDB60000D9B5,
-    0x01FD5DB60000ADB7, 0x01FD5DB401FD65B7, 0x01FD35B201FD55B5,
-    0x01FD01B001FD45B3, 0x01FD8DBE01FD3DB1, 0x01FD8DBC01FD95BF,
-    0x01FD5DBA01FD85BD, 0x01FD5DB801FD4DBB, 0x0007E08601FD55B9,
-    0x800086CE00051887, 0x800086D6800086D2, 0x800086DA00071484,
-    0x01FD6DAE800086DE, 0x01FD69AC01FD71AF, 0x01FEA5AA01FD69AD,
-    0x01FED5A801FE89AB, 0x000324E101FEBDA9, 0x001FDCE301FED4E1,
-    0x000028E301FE10E3, 0x000108E1800086E2, 0x0000F4EF800086E6,
-    0x0007F4BE800086EA, 0x800086F2800086EE, 0x001F7CEF800086F6,
-    0x000209B301FEFCEF, 0x800086FE800086FA, 0x8000870680008702,
-    0x000209B5000209B6, 0x0001E9BB8000870A, 0x8000870E0001E1BA,
-    0x0001DDBF80008712, 0x0001E5BD0001DDBE, 0x01F8259F0001EDBC,
-    0x01F9F19D01F8259E, 0x01FC319B01FCA19C, 0x01F1159901FDE19A,
-    0x01F0219701F01998, 0x01F4259501F54D96, 0x01E9619301F42D94,
-    0x01E83D9101E83D92, 0x01EC618F01EEBD90, 0x80008716800086BE,
-    0x8000871E8000871A, 0x00003D8100003D82, 0x8000872200003D80,
-    0x8000872A80008726, 0x0001D0868000872E, 0x8000873200030087,
-    0x000549A380008736, 0x8000873E8000873A, 0x8000874680008742,
-    0x000549A5000549A6, 0x0006A9AB8000874A, 0x8000874E0006A1AA,
-    0x00069DAF80008752, 0x0006A5AD00069DAE, 0x800087560006ADAC,
-    0x8000875E8000875A, 0x8000876680008762, 0x8000876E8000876A,
-    0x8000877680008772, 0x8000877E8000877A, 0x000535A380008782,
-    0x0004F9A10004E9A2, 0x01ED218201F931A0, 0x8000878601EF8583,
-    0x01E841868000878A, 0x01E91D8401F51587, 0x01F52D8A01E91585,
-    0x8000878E01F44D8B, 0x01F03D8E80008792, 0x01F1398C01FCC18F,
-    0x01FD619201F1398D, 0x8000879601FDE993, 0x01F975968000879A,
-    0x01F8A19401F89997, 0x0005619A01F97D95, 0x00002C840005799B,
-    0x00003981000C00BE, 0x0004BD9C00005980, 0x0001C5A100003182,
-    0x000D89A3800087A2, 0x000D01A50001E1A2, 0x0009B9A7000D91A4,
-    0x001401A9000801A6, 0x001599AB00090DA8, 0x0011B1AD00158DAA,
-    0x001C01AF001421AC, 0x001DCDB1001115AE, 0x001801B3001DC5B0,
-    0x001939B5001C55B2, 0x000F58860019F5B4, 0x800087A60000FC87,
-    0x800087AE800087AA, 0x800087B6800087B2, 0x800087BE800087BA,
-    0x0004B5A8800087C2, 0x0004B5AA0004BDA9, 0x00055DAC000545AB,
-    0x000541AE000545AD, 0x800087C6000541AF, 0x800087CE800087CA,
-    0x800087D6800087D2, 0x800087DE800087DA, 0x00058DB8800087E2,
-    0x000595BA000595B9, 0x0005BDBC0005A5BB, 0x000215BE00021DBD,
-    0x0005F5B000020DBF, 0x0005CDB20005F9B1, 0x0005C1B40005C1B3,
-    0x0005BDB60005C1B5, 0x800087E60005B5B7, 0x800087EE800087EA,
-    0x800087F2000431B3, 0x0005C9B60005D9B5, 0x800087FA800087F6,
-    0x00053DBA800087FE, 0x000515BC000531BB, 0x000509BE000519BD,
-    0x80008802000509BF, 0x000585A5000595A6, 0x00046DA380008806,
-    0x000C90BE8000880A, 0x000555AF80008812, 0x000545AD000555AE,
-    0x00056DAB000549AC, 0x80008816000561AA, 0x000379A88000881A,
-    0x00038DAA000379A9, 0x00039DAC000385AB, 0x000385AE000385AD,
-    0x0001E1800003B5AF, 0x0001FD820001E981, 0x8000881E000D6084,
-    0x000D248780008822, 0x00039DB8000C9C86, 0x0003ADBA0003A5B9,
-    0x0003BDBC0003A5BB, 0x000255BE00025DBD, 0x0003CDB0000275BF,
-    0x0003ADB20003A5B1, 0x0003BDB40003A5B3, 0x0003ADB60003A5B5,
-    0x800088260003A5B7, 0x00030DB38000882A, 0x00031DB58000882E,
-    0x80008832000CDCBE, 0x8000883600030DB6, 0x000369BB8000883A,
-    0x000379BD000301BA, 0x000369BF000371BC, 0x000698E1000371BE,
-    0x00068CE38000883E, 0x000044EF80008842, 0x8000884A80008846,
-    0x000098E38000884E, 0x000180E180008852, 0x8000885A80008856,
-    0x0006A8EF8000885E, 0x8000879E80008862, 0x800088668000880E,
-    0x8000886E8000886A, 0x0000918100009982, 0x000DA5A800009980,
-    0x000DCDAA000DC5A9, 0x000DDDAC000DC5AB, 0x000DCDAE000DC5AD,
-    0x000DBDB0000DC5AF, 0x000E6DB2000E65B1, 0x000E7DB4000E65B3,
-    0x000E6DB6000E65B5, 0x000E5DB8000E65B7, 0x000EEDBA000EE5B9,
-    0x000EE1BC000EE5BB, 0x000ED9BE000EE9BD, 0x000D65B3000ED9BF,
-    0x8000887680008872, 0x8000887E8000887A, 0x000D75B5000D65B6,
-    0x000E81BB0003D084, 0x00000886000D59BA, 0x000E81BF00008C87,
-    0x000E91BD000E99BE, 0x80008882000E99BC, 0x80008886000D21A3,
-    0x000D21A68000888A, 0x800088928000888E, 0x000D1DAA000D31A5,
-    0x80008896000EC5AB, 0x000EDDAE8000889A, 0x000EDDAC000EC5AF,
-    0x000FDDA8000ED5AD, 0x000FF5AA000FFDA9, 0x000E25AC000E3DAB,
-    0x000E9DAE000E95AD, 0x8000889E000E91AF, 0x800088A6800088A2,
-    0x800088AE800088AA, 0x800088B6800088B2, 0x000EA5B8800088BA,
-    0x000EB9BA000EA9B9, 0x000EADBC000EB9BB, 0x000155BE00015DBD,
-    0x000EF9B000014DBF, 0x000ECDB2000EC5B1, 0x000EBDB4000EA9B3,
-    0x000EA1B6000EA1B5, 0x000F7DB3000EA1B7, 0x800088C2800088BE,
-    0x800088CA800088C6, 0x000EADB5000E15B6, 0x000E01BB800088CE,
-    0x800088D2000E3DBA, 0x000E65BF800088D6, 0x000E7DBD000E79BE,
-    0x800088DA000E19BC, 0x800088DE000F39A3, 0x000E51A6800088E2,
-    0x800088EA800088E6, 0x000E79AA000EE9A5, 0x0001A484000E45AB,
-    0x000E3DAE0001A4BE, 0x000E5DAC000E21AF, 0x0002C1A8000E39AD,
-    0x0002C5AA0002CDA9, 0x0002C9AC0002D9AB, 0x0002F5AE0002FDAD,
-    0x0003ED8000033DAF, 0x0003F1820003F181, 0x00049086800088F2,
-    0x800088F600037487, 0x0003C9B8800088FA, 0x0003D9BA0003C9B9,
-    0x0003F9BC0003D1BB, 0x000399BE0003F9BD, 0x000345B0000395BF,
-    0x000345B200034DB1, 0x000345B400035DB3, 0x000345B600034DB5,
-    0x00024DB30003F9B7, 0x80008902800088FE, 0x8000890600030C84,
-    0x000241B500027DB6, 0x000389BB8000890A, 0x8000890E000381BA,
-    0x000389BF80008912, 0x000389BD000381BE, 0x80008916000391BC,
-    0x8000891A000209A3, 0x000239A68000891E, 0x8000892680008922,
-    0x0003C5AA000205A5, 0x8000892A0003CDAB, 0x0003C5AE8000892E,
-    0x0003D5AC0003CDAF, 0x800089320003CDAD, 0x8000893A80008936,
-    0x800089428000893E, 0x800089460001A4E1, 0x8000894A0000E4E3,
-    0x800089528000894E, 0x80008956000038EF, 0x800089628000895A,
-    0x00006980000530BE, 0x00007D8200007581, 0x80008966000171B3,
-    0x0001C9B60001DDB5, 0x0004C0868000896A, 0x0001C5BA0005E887,
-    0x0001B9BC0001ADBB, 0x0001ADBE0001B9BD, 0x000EA4E10001A5BF,
-    0x000104E3000C44E1, 0x8000896E000EB0E3, 0x8000897680008972,
-    0x0005A8848000897A, 0x800089828000897E, 0x8000898A80008986,
-    0x000ED4EF8000898E, 0x000F01AB000F9CEF, 0x80008992000F69AA,
-    0x000F01AF80008996, 0x000F11AD000F19AE, 0x000F65A3000F19AC,
-    0x8000899E8000899A, 0x800089A6800089A2, 0x000F75A5000F65A6,
-    0x00062DA8800089AA, 0x000635AA00063DA9, 0x000699AC000689AB,
-    0x000685AE000685AD, 0x8000895E0006B5AF, 0x800089B2800089AE,
-    0x800089BA800089B6, 0x800089C2800089BE, 0x000741B8800089C6,
-    0x000741BA000741B9, 0x000741BC000741BB, 0x000779BE00074DBD,
-    0x0006CDB0000779BF, 0x0006A1B20006A5B1, 0x0006A5B40006A1B3,
-    0x000691B60006ADB5, 0x0006BDA3000691B7, 0x0000B58100001582,
-    0x800089CA0000B580, 0x000611A5000605A6, 0x000661AB000390BE,
-    0x800089CE000609AA, 0x000669AF800089D2, 0x000675AD000661AE,
-    0x800089D6000675AC, 0x0000C886000799B3, 0x000749B600031C87,
-    0x800089DE800089DA, 0x000749BA000749B5, 0x800089E2000749BB,
-    0x000739BE800089E6, 0x000749BC000739BF, 0x000639A8000749BD,
-    0x000655AA000655A9, 0x000685AC00069DAB, 0x000685AE00068DAD,
-    0x800089EA0006B9AF, 0x800089F2800089EE, 0x800089FA800089F6,
-    0x80008A02800089FE, 0x0006ADB880008A06, 0x0006B5BA0006B5B9,
-    0x000179BC000169BB, 0x000169BE000179BD, 0x0006C9B0000169BF,
-    0x0006D9B20006C9B1, 0x0006B9B40006D9B3, 0x00069DB60006B9B5,
-    0x0006DDA3000695B7, 0x80008A0E80008A0A, 0x80008A1680008A12,
-    0x00060DA500060DA6, 0x00060DAB800088EE, 0x80008A1A00060DAA,
-    0x00067DAF80008A1E, 0x00060DAD00067DAE, 0x80008A2200060DAC,
-    0x80008A2A80008A26, 0x00000D8080008A2E, 0x00003D8200003581,
-    0x80008A3680008A32, 0x0001648480008A3A, 0x001F5086000034BE,
-    0x80008A4200016887, 0x80008A4A80008A46, 0x0003048480008A4E,
-    0x00068CE3001C00BE, 0x000630E180008A52, 0x0007E4EF80008A56,
-    0x80008A5E80008A5A, 0x80008A6680008A62, 0x80008A6E80008A6A,
-    0x000295B380008A72, 0x80008A7A80008A76, 0x80008A8280008A7E,
-    0x0002B1B50002B9B6, 0x00034DBB80008A86, 0x80008A8A00034DBA,
-    0x00034DBF80008A8E, 0x00034DBD00034DBE, 0x80008A9200034DBC,
-    0x80008A9A80008A96, 0x80008A9E001D88BE, 0x00019D8100019D82,
-    0x00020DA800019D80, 0x00027DAA000265A9, 0x000259AC000275AB,
-    0x000399AE000259AD, 0x0003E9B0000399AF, 0x0003F9B20003E9B1,
-    0x0003E9B40003F9B3, 0x00035DB60003E9B5, 0x00036DB8000355B7,
-    0x000341BA000375B9, 0x000331BC000341BB, 0x000331BE000331BD,
-    0x80008AA2000331BF, 0x80008AAA80008AA6, 0x80008AB2001D0484,
-    0x001CB886000000EF, 0x80008AB6001D3087, 0x80008ABA0000DCE3,
-    0x80008ABE0001FCE1, 0x80008AC680008AC2, 0x0003ADA680008ACA,
-    0x80008AD280008ACE, 0x80008AD60003A5A5, 0x80008ADA000381A3,
-    0x000259AE80008ADE, 0x000259AC000259AF, 0x000259AA000259AD,
-    0x80008AE2000259AB, 0x80008AEA80008AE6, 0x80008AF280008AEE,
-    0x80008AFA80008AF6, 0x80008B0280008AFE, 0x001CAC840001F8EF,
-    0x80008B060006DCE1, 0x80008B0A00061CE3, 0x80008B1280008B0E,
-    0x0001C1B380008B16, 0x001FEC8680008B1A, 0x000000F5001CF087,
-    0x0001CDB50001C9B6, 0x0001D1BB80008B22, 0x001F2CBE0001C9BA,
-    0x0001B5BF80008B26, 0x0001C1BD0001C9BE, 0x001DF9A80001C9BC,
-    0x001E09AA001DF9A9, 0x001E05AC001E09AB, 0x001E3DAE001E09AD,
-    0x80008AAE001E35AF, 0x001F9581001FE182, 0x80008B1E001F9580,
-    0x80008B2E80008B2A, 0x001E11B880008B32, 0x001E21BA001E19B9,
-    0x001ED1BC001E21BB, 0x001ED1BE001ED1BD, 0x001E4DB0001ECDBF,
-    0x001E5DB2001E55B1, 0x001E4DB4001E55B3, 0x001E31B6001E31B5,
-    0x001E81A3001E31B7, 0x80008B3A80008B36, 0x80008B4280008B3E,
-    0x001E8DA5001E89A6, 0x001E91AB80008B46, 0x80008B4A001E89AA,
-    0x001EF5AF80008B4E, 0x001E81AD001E89AE, 0x80008B52001E89AC,
-    0x80008B56001FEDB3, 0x001FE9B680008B5A, 0x80008B6280008B5E,
-    0x001FA9BA001FE9B5, 0x80008B66001FB5BB, 0x001F89BE80008B6A,
-    0x001FA1BC001F89BF, 0x001E11A8001FA1BD, 0x001E21AA001E25A9,
-    0x001E25AC001E3DAB, 0x001E25AE001E2DAD, 0x80008B6E001E59AF,
-    0x80008B7680008B72, 0x80008B7E80008B7A, 0x0000BD810000B982,
-    0x001EC9B80000BD80, 0x001ED9BA001EC9B9, 0x001EC9BC001ED9BB,
-    0x001EE9BE001EFDBD, 0x001E29B0001EE9BF, 0x001E31B2001E35B1,
-    0x001E09B4001E31B3, 0x001EF9B6001E09B5, 0x80008B82001EF9B7,
-    0x001EA1A380008B86, 0x001EA5A580008B8A, 0x80008B9280008B8E,
-    0x80008A3E001EA5A6, 0x001EF9AB0000A884, 0x001EEDAD001EE5AA,
-    0x001EC5AF001EEDAC, 0x0001F5B3001EC5AE, 0x00006C8680008B96,
-    0x80008B9A00017C87, 0x0001E5B5000151B6, 0x00017DBB80008B9E,
-    0x80008BA2000175BA, 0x000139BF80008BA6, 0x000155BD000135BE,
-    0x80008BAA000155BC, 0x000400BE80008BAE, 0x0001C0E100040084,
-    0x0001ACE380008BB2, 0x80008BB6001EE8EF, 0x0002D8EF001DCCE1,
-    0x80008BBA001E34E3, 0x80008BC280008BBE, 0x80008BCA80008BC6,
-    0x00027DA380008BCE, 0x00026DA580008BD2, 0x80008BDA80008BD6,
-    0x80008BDE0002D9A6, 0x0002F5AB80008BE2, 0x0002DDAD0002FDAA,
-    0x0002B1AF0002DDAC, 0x001D6DA90002BDAE, 0x0002B1AB001D59A8,
-    0x000215AD0002B9AA, 0x000215AF0002A9AC, 0x00052CBE000215AE,
-    0x80008BEA80008BE6, 0x00000D8080008BEE, 0x00019D8200019581,
-    0x000229B980008BF2, 0x0003E9BB00021DB8, 0x0003F9BD000221BA,
-    0x0003E9BF0003F1BC, 0x00027DB10003F1BE, 0x000245B3000275B0,
-    0x000239B5000245B2, 0x00022DB7000235B4, 0x0002C5A9000231B6,
-    0x0002D5AB00029DA8, 0x0002FDAD0002DDAA, 0x00011DAF0002FDAC,
-    0x80008BF600011DAE, 0x0005A08780008BFA, 0x80008C020005AC86,
-    0x80008C0A80008C06, 0x000115B9000000CF, 0x000129BB000109B8,
-    0x0001F5BD000115BA, 0x0001F5BF000131BC, 0x000141B10001F5BE,
-    0x00015DB3000171B0, 0x00014DB500015DB2, 0x00013DB700014DB4,
-    0x80008C0E00013DB6, 0x80008C1680008C12, 0x80008C1A001E9CEF,
-    0x80008C2280008C1E, 0x00076C84001FE8EF, 0x80008C26001EA4E1,
-    0x80008C2A001EB0E3, 0x80008C2E001E64E1, 0x000211B3001E0CE3,
-    0x80008C3680008C32, 0x00065C8480008C3A, 0x000231B5000239B6,
-    0x0001C1BB80008C3E, 0x80008C420001F1BA, 0x0001CDBF80008C46,
-    0x0001CDBD0001CDBE, 0x80008BFE0001CDBC, 0x80008C4E80008C4A,
-    0x80008C5680008C52, 0x80008C5E80008C5A, 0x000711A880008C62,
-    0x000711AA000711A9, 0x0007B5AC0007ADAB, 0x0007A9AE0007BDAD,
-    0x0007D9B00007A9AF, 0x0007F9B20007EDB1, 0x000795B40007F5B3,
-    0x000795B600079DB5, 0x000789B8000781B7, 0x00006DBA000789B9,
-    0x00007DBC000065BB, 0x00006DBE000065BD, 0x00001982000065BF,
-    0x00009580000555A3, 0x00057DA600001981, 0x80008C6A80008C66,
-    0x0006B5AA000575A5, 0x80008C6E000685AB, 0x000689AE80008C72,
-    0x000689AC000689AF, 0x0007F9B3000689AD, 0x0003E88680008C76,
-    0x80008C7A0003BC87, 0x00075DB500075DB6, 0x000749BB80008C7E,
-    0x80008C8200077DBA, 0x000739BF80008C86, 0x000741BD000731BE,
-    0x000639A300074DBC, 0x80008C8E80008C8A, 0x80008C9680008C92,
-    0x00069DA500069DA6, 0x000689AB80008C9A, 0x80008C9E0006BDAA,
-    0x0006F9AF80008CA2, 0x000681AD0006F1AE, 0x80008CA600068DAC,
-    0x80008CAE80008CAA, 0x80008CB680008CB2, 0x80008CBE80008CBA,
-    0x80008CC680008CC2, 0x80008CCE80008CCA, 0x00001D8080008CD2,
-    0x0000D98200000981, 0x00012DA880008CD6, 0x000159AA000141A9,
-    0x000179AC000151AB, 0x00009DAE000179AD, 0x80008CDA000095AF,
-    0x0001888600016CBE, 0x80008CE20000AC87, 0x80008CEA80008CE6,
-    0x00008DB880008CEE, 0x00009DBA000095B9, 0x00008DBC000095BB,
-    0x0000BDBE0000B5BD, 0x0000EDB00000B5BF, 0x0000F5B20000F5B1,
-    0x0000DDB40000C5B3, 0x0000BDB60000B5B5, 0x0002D1B30000B5B7,
-    0x000370BE80008CF2, 0x80008CF6000C44BE, 0x0002F1B50002F9B6,
-    0x000331BB80008CFA, 0x80008CFE000329BA, 0x000371BF80008D02,
-    0x000311BD000319BE, 0x80008D06000321BC, 0x80008D0A000295A3,
-    0x0002BDA680008D0E, 0x80008D1680008D12, 0x00036DAA0002B5A5,
-    0x80008D1A000375AB, 0x00035DAE80008D1E, 0x000365AC000335AF,
-    0x000331A8000355AD, 0x000331AA000331A9, 0x000095AC000331AB,
-    0x000085AE000085AD, 0x80008D220000B5AF, 0x80008D2A80008D26,
-    0x80008D3280008D2E, 0x80008D3A80008D36, 0x000061B880008D3E,
-    0x000061BA000061B9, 0x000061BC000061BB, 0x000061BE000061BD,
-    0x0000CDB0000061BF, 0x0000A1B20000A5B1, 0x0000A5B40000A1B3,
-    0x000091B60000ADB5, 0x00000D80000091B7, 0x0000118200001581,
-    0x80008D4680008D42, 0x80008D4A000048EF, 0x000D6C8480008D52,
-    0x80008D56000694E3, 0x80008D5A0001FCE1, 0x000DB88680008D5E,
-    0x0002F9B30002C487, 0x80008D6680008D62, 0x80008D6E80008D6A,
-    0x00015DB5000145B6, 0x000121BB000F98BE, 0x80008D72000149BA,
-    0x000129BF80008D76, 0x000135BD000121BE, 0x80008D7A000135BC,
-    0x80008D7E0006B0E1, 0x80008D82000790E3, 0x80008D8A80008D86,
-    0x80008D9280008D8E, 0x80008D9A80008D96, 0x000000E580008D9E,
-    0x80008DA680008DA2, 0x000195AE0007E4EF, 0x000181AC00019DAF,
-    0x0001FDAA000181AD, 0x80008DAA000195AB, 0x0001F1A680008DAE,
-    0x80008DB680008DB2, 0x000FF5820001E9A5, 0x00001D8000024DA3,
-    0x80008D4E00001581, 0x80008DC280008DBA, 0x0003B08680008DBE,
-    0x80008DC600039487, 0x000F3DA880008DCA, 0x000F19AA000F0DA9,
-    0x000F75AC000F19AB, 0x000F75AE000F7DAD, 0x000F29B0000F6DAF,
-    0x000FC9B2000F29B1, 0x000FC1B4000FC5B3, 0x000FC1B6000FC1B5,
-    0x000FC1B8000FC1B7, 0x000FC1BA000FC1B9, 0x000FC5BC000FC1BB,
-    0x000FA1BE000FCDBD, 0x000FB1B3000FA1BF, 0x80008DD280008DCE,
-    0x80008DDA80008DD6, 0x000FDDB5000FDDB6, 0x000FCDBB80008DDE,
-    0x80008DE2000FCDBA, 0x000F25BF80008DE6, 0x000F31BD000F25BE,
-    0x80008DEA000F3DBC, 0x80008DEE000FF5A3, 0x000F99A680008DF2,
-    0x80008DFA80008DF6, 0x000F89AA000F99A5, 0x80008DFE000F89AB,
-    0x000F61AE80008E02, 0x000F79AC000F61AF, 0x000109A8000F75AD,
-    0x000115AA00011DA9, 0x000199AC000189AB, 0x000189AE000199AD,
-    0x80008E06000181AF, 0x80008E0E80008E0A, 0x80008E1680008E12,
-    0x80008E1E80008E1A, 0x0001ADB880008E22, 0x00006DBA000065B9,
-    0x00007DBC000065BB, 0x00006DBE000065BD, 0x0001C9B0000065BF,
-    0x0001A9B20001C9B1, 0x0001BDB40001A5B3, 0x0001A1B60001A1B5,
-    0x0000BD80000195B7, 0x0000198200000981, 0x80008E2A80008E26,
-    0x80008E2E00029884, 0x0001808680008E32, 0x80008E3600010487,
-    0x80008E3E80008E3A, 0x80008CDE80008E42, 0x0001DDB380008E46,
-    0x80008E4E80008E4A, 0x80008E5680008E52, 0x000151B5000159B6,
-    0x00014DBB80008E5A, 0x80008E5E00014DBA, 0x000125BF80008E62,
-    0x000131BD000139BE, 0x80008E6600013DBC, 0x0003D0BE000EF4EF,
-    0x80008E6E80008E6A, 0x80008E72000248E3, 0x80008E760001A4E1,
-    0x80008E7A0002D4EF, 0x80008E8280008E7E, 0x80008E86000BF4E3,
-    0x80008E8A000D8CE1, 0x000255A380008E8E, 0x0002D9A500038C84,
-    0x80008E9680008E92, 0x80008E9A0002D1A6, 0x0002C5AB80008EA2,
-    0x0002B9AD0002C5AA, 0x0002ADAF0002B5AC, 0x00031DA80002B1AE,
-    0x000325AA00032DA9, 0x000325AC00033DAB, 0x00035DAE000329AD,
-    0x0000F580000355AF, 0x0000C9820000FD81, 0x80008EAA80008EA6,
-    0x00036C8780008EAE, 0x0000D9B800041C86, 0x0000E1BA0000E5B9,
-    0x0000E1BC0000E1BB, 0x000099BE0000EDBD, 0x00032DB0000099BF,
-    0x000335B2000335B1, 0x00031DB4000305B3, 0x0000F5B60000F5B5,
-    0x80008EB20000EDB7, 0x80008EBA80008EB6, 0x80008EBE0002A1B3,
-    0x0002A5B60002A1B5, 0x80008EC680008EC2, 0x0002E5BA80008ECA,
-    0x0002EDBC0002FDBB, 0x0002D5BE0002DDBD, 0x0002FDA600013DBF,
-    0x80008ED280008ECE, 0x80008ED60002F9A5, 0x80008EDA0002F9A3,
-    0x00028DAE80008EDE, 0x0002B5AC000165AF, 0x0002BDAA000285AD,
-    0x80008EE20002A5AB, 0x000004E380008EE6, 0x000180E1000EA4E1,
-    0x80008EEA000EBCE3, 0x80008EF280008EEE, 0x0005E8BE80008EF6,
-    0x80008EFE80008EFA, 0x000E14EF80008F02, 0x0004F0BE80008F06,
-    0x80008F0A000F48EF, 0x80008F1680008F12, 0x0000658180008F1A,
-    0x80008F1E00007980, 0x0005AC8400001D82, 0x0004C88780008F22,
-    0x80008F2600044C86, 0x80008F2E80008F2A, 0x000665A880008F32,
-    0x0006B9AA0006A9A9, 0x000711AC0006B9AB, 0x000711AE000711AD,
-    0x80008F0E00070DAF, 0x80008F3A80008F36, 0x80008F4280008F3E,
-    0x80008F4A80008F46, 0x000739B880008F4E, 0x0007D9BA000739B9,
-    0x0007C9BC0007D9BB, 0x0007FDBE0007C9BD, 0x00077DB00007F5BF,
-    0x00074DB2000745B1, 0x000739B4000729B3, 0x000721B6000725B5,
-    0x000631B3000721B7, 0x80008F5680008F52, 0x80008F5E80008F5A,
-    0x000799B5000791B6, 0x000781BB80008F62, 0x80008F660007B9BA,
-    0x0007E1BF80008F6A, 0x0007FDBD0007FDBE, 0x80008F6E000799BC,
-    0x80008F72000675A3, 0x0007D5A680008F76, 0x0003948480008F7A,
-    0x0007FDAA0007DDA5, 0x80008F7E0007C5AB, 0x0007B9AE80008F82,
-    0x0007DDAC0007A5AF, 0x0006C5A80007B9AD, 0x0006C5AA0006CDA9,
-    0x0006F5AC0006D5AB, 0x0006EDAE0006E5AD, 0x0001E9800006E5AF,
-    0x0001FD820001E981, 0x0001308680008F86, 0x80008F8A00016487,
-    0x0001C5B880008F8E, 0x0001C5BA0001CDB9, 0x0001CDBC0001DDBB,
-    0x0001F5BE0001FDBD, 0x00069DB000019DBF, 0x000141B2000141B1,
-    0x000141B4000141B3, 0x000141B6000141B5, 0x00064DB3000141B7,
-    0x80008F9680008F92, 0x80008F9E80008F9A, 0x00067DB5000665B6,
-    0x000185BB80008FA2, 0x80008FA6000659BA, 0x000185BF80008FAA,
-    0x000195BD000185BE, 0x80008FAE000195BC, 0x80008E9E000609A3,
-    0x000621A680008FB2, 0x80008FBA80008FB6, 0x00061DAA000639A5,
-    0x80008FBE0001C1AB, 0x0001C1AE80008FC2, 0x0001D1AC0001C1AF,
-    0x80008FC60001D1AD, 0x80008FCE80008FCA, 0x80008FD680008FD2,
-    0x80008FDE80008FDA, 0x80008FE680008FE2, 0x000364BE80008FEA,
-    0x80008FF680008FF2, 0x003D30BF80008FFA, 0x003C008580008FFE,
-    0x00038C840007B4E3, 0x00004D800006E8E1, 0x00007D8200007581,
-    0x003C4485002A8CEF, 0x80009002000210E1, 0x80009006002EA8E3,
-    0x003D588600072CEF, 0x8000900A00030C87, 0x800090128000900E,
-    0x8000901A80009016, 0x800090228000901E, 0x00024DA880009026,
-    0x000391AA000389A9, 0x000335AC000391AB, 0x000335AE00033DAD,
-    0x00035DB000032DAF, 0x000365B200036DB1, 0x00031DB400030DB3,
-    0x000301B6000301B5, 0x00030DB8000335B7, 0x0000EDBA0000E5B9,
-    0x0000FDBC0000E5BB, 0x0000EDBE0000E5BD, 0x8000902A0000E5BF,
-    0x000231B38000902E, 0x000295B580009032, 0x8000903A80009036,
-    0x003DBC84000295B6, 0x000281BB8000903E, 0x0002F1BD0002B1BA,
-    0x0002EDBF000299BC, 0x800090420002F1BE, 0x003FC4BE80009046,
-    0x8000904A003FC484, 0x8000904E000384EF, 0x8000905680009052,
-    0x8000905A00033CE3, 0x00002D80000044E1, 0x0000118200001581,
-    0x003FA08680009062, 0x80009066003D5887, 0x8000906E8000906A,
-    0x8000907680009072, 0x003E2CBE0006DCEF, 0x8000907A000650E1,
-    0x8000907E000164E3, 0x8000908680009082, 0x0002F9A38000908A,
-    0x800090928000908E, 0x8000909A80009096, 0x00025DA500025DA6,
-    0x000249AB8000909E, 0x800090A2000279AA, 0x000225AF800090A6,
-    0x000239AD000239AE, 0x00246D90000251AC, 0x00280D9200280591,
-    0x002D8D9400289D93, 0x002CB196002C3595, 0x003001980031BD97,
-    0x0031BD9A00301599, 0x0034999C00342D9B, 0x0038399E00398D9D,
-    0x80008FEE0038319F, 0x003E21B5003E35B6, 0x003E65B38000905E,
-    0x800090AE800090AA, 0x003EFDBF800090B2, 0x003EFDBD003EFDBE,
-    0x003EFDBB003EFDBC, 0x800090B6003EFDBA, 0x000945A90024398F,
-    0x000CA9AB00089DA8, 0x001001AD000C01AA, 0x001035AF000D1DAC,
-    0x0000E1A10010BDAE, 0x000401A3800090BA, 0x000495A50000E1A2,
-    0x0008B9A7000495A4, 0x003959A0000801A6, 0x003CE9A2003D0DA1,
-    0x800090BE003C25A3, 0x000025810001DD82, 0x0014C1B100002580,
-    0x001801B3001401B0, 0x0018D5B50014D5B2, 0x800090C20018D5B4,
-    0x00254582800090C6, 0x800090CA00256983, 0x00298186800090D2,
-    0x00291984002C8987, 0x002D018A00291985, 0x00078886002D8D8B,
-    0x0031018E00018C87, 0x0030998C0030A18F, 0x0035419200311D8D,
-    0x800090D60035ED93, 0x00395D96800090DA, 0x0038E1940038CD97,
-    0x003CA99A00395D95, 0x800090DE003D459B, 0x800090E6800090E2,
-    0x003D919C800090EA, 0x003FADA8800090EE, 0x003FBDAA003FB5A9,
-    0x003FF9AC003FB1AB, 0x003F19AE003FE5AD, 0x800090F2003F19AF,
-    0x800090FA800090F6, 0x80009102800090FE, 0x8000910A80009106,
-    0x003ECDB88000910E, 0x003ED1BA003ED1B9, 0x003E91BC003EE9BB,
-    0x003E89BE003E91BD, 0x003F69B0003E85BF, 0x003F79B2003F69B1,
-    0x003F69B4003F79B3, 0x003EF9B6003F69B5, 0x003C35A3003EF9B7,
-    0x8000911680009112, 0x8000911E8000911A, 0x003C71A5003C65A6,
-    0x003CADAB80009122, 0x80009126003CADAA, 0x003CADAF8000912A,
-    0x003CADAD003CADAE, 0x0003C884003CADAC, 0x8000912E003DC9B3,
-    0x003DC9B680009132, 0x8000913A80009136, 0x003DD5BA003DC9B5,
-    0x00036CBE003DDDBB, 0x003DB9BE80009142, 0x003DC5BC003DB9BF,
-    0x003DFDA8003DB9BD, 0x003D05AA003D0DA9, 0x003D05AC003D1DAB,
-    0x003D31AE003D0DAD, 0x0000ED80003D2DAF, 0x0000F1820000F181,
-    0x0004908680009146, 0x8000914A00036487, 0x003D29B88000914E,
-    0x00029DBA003D29B9, 0x00028DBC000295BB, 0x00027DBE0002ADBD,
-    0x003D55B000020DBF, 0x003D55B2003D5DB1, 0x003D39B4003D29B3,
-    0x003D29B6003D39B5, 0x003C8DA3003D21B7, 0x0004C48480009152,
-    0x8000915A80009156, 0x003C8DA5003C8DA6, 0x003C99AB8000915E,
-    0x80009162003C91AA, 0x003CFDAF80009166, 0x003CFDAD003CFDAE,
-    0x8000916A003C81AC, 0x8000916E000209B3, 0x000221B680009172,
-    0x8000917A80009176, 0x0002C1BA000231B5, 0x8000917E0002C1BB,
-    0x0002C1BE80009182, 0x0002D1BC0002C1BF, 0x800091860002D1BD,
-    0x0002C5A38000918A, 0x0002FDA58000918E, 0x8000919680009192,
-    0x8000919A0002EDA6, 0x00020DAB8000919E, 0x00021DAD00020DAA,
-    0x00020DAF00021DAC, 0x003E9CE300020DAE, 0x003E50E1003FA4E1,
-    0x00000D80003EFCE3, 0x00003D8200003581, 0x800091A6800091A2,
-    0x000420BE800091AA, 0x003EA8EF800091B2, 0x800091BA800091B6,
-    0x00050DA8003E4CEF, 0x000545AA00054DA9, 0x00057DAC000565AB,
-    0x00056DAE000565AD, 0x8000913E000561AF, 0x00030886800091AE,
-    0x800091BE00038C87, 0x800091C6800091C2, 0x000645B8800091CA,
-    0x000645BA00064DB9, 0x00064DBC00065DBB, 0x000675BE00067DBD,
-    0x0005C1B0000615BF, 0x0005C1B20005C1B1, 0x0005C1B40005C1B3,
-    0x0005C1B60005C1B5, 0x0006EDA80005C1B7, 0x000631AA000629A9,
-    0x000795AC000631AB, 0x000789AE000799AD, 0x800091CE000781AF,
-    0x800090CE800091D2, 0x800091DA800091D6, 0x800091E2800091DE,
-    0x0007ADB8800091E6, 0x00074DBA000745B9, 0x000745BC000741BB,
-    0x000779BE00074DBD, 0x0007C9B0000779BF, 0x0007A9B20007C9B1,
-    0x0007BDB40007A5B3, 0x0007A1B60007A1B5, 0x0007B9B3000795B7,
-    0x800091EE800091EA, 0x800091F6800091F2, 0x00061DB5000619B6,
-    0x00060DBB800091FA, 0x800091FE00060DBA, 0x000665BF80009202,
-    0x000671BD000679BE, 0x8000920600067DBC, 0x8000920A0007FDA3,
-    0x00065DA68000920E, 0x8000921680009212, 0x000649AA000659A5,
-    0x8000921A000649AB, 0x00063DAE0001A084, 0x000639AC000621AF,
-    0x00066DA8000635AD, 0x000671AA000671A9, 0x0001A5AC0001BDAB,
-    0x0001A5AE0001ADAD, 0x0000F58000013DAF, 0x0000C1820000FD81,
-    0x00019886000164BE, 0x800092220000AC87, 0x0000CDB880009226,
-    0x0000D5BA0000D5B9, 0x0000FDBC0000E5BB, 0x000091BE000095BD,
-    0x000145B0000091BF, 0x000159B2000149B1, 0x000149B4000159B3,
-    0x0000FDB6000149B5, 0x0005F9B30000F5B7, 0x8000922E8000922A,
-    0x8000923680009232, 0x00022DB500022DB6, 0x00028DBB8000923A,
-    0x8000923E000281BA, 0x00028DBF80009242, 0x00028DBD00028DBE,
-    0x8000924600028DBC, 0x8000924A0005BDA3, 0x000269A68000924E,
-    0x8000925680009252, 0x0002C5AA000269A5, 0x8000925A0002C9AB,
-    0x0002C9AE8000925E, 0x0002C9AC0002C9AF, 0x000085AA0002C9AD,
-    0x000331A800008DAB, 0x00008DAE000331A9, 0x000095AC0000BDAF,
-    0x0003E4BE00008DAD, 0x8000926680009262, 0x8000926E8000926A,
-    0x8000927680009272, 0x00006DBA8000927A, 0x00006DB8000075BB,
-    0x00006DBE000065B9, 0x00006DBC0001D5BF, 0x0000A9B2000065BD,
-    0x0000C5B00000A9B3, 0x000099B60000ADB1, 0x0000ADB4000099B7,
-    0x8000927E0000A5B5, 0x8000928680009282, 0x00001D828000928A,
-    0x00001D80000784EF, 0x0002AC8400001581, 0x80009292000774E3,
-    0x8000929600041CE1, 0x000DB8868000929A, 0x000145B3000D3087,
-    0x800092A28000929E, 0x800092AA800092A6, 0x000155B5000145B6,
-    0x000121BB800092AE, 0x800092B2000149BA, 0x000129BF800092B6,
-    0x000135BD000121BE, 0x800092BA000135BC, 0x800092C2800092BE,
-    0x800092CA800092C6, 0x0000CCEF800092CE, 0x000190E100068CEF,
-    0x0000D4E300070CE1, 0x800092D20006B4E3, 0x800092DA800092D6,
-    0x800092E2800092DE, 0x0001CDA3800092E6, 0x0001DDA5000D9084,
-    0x800092EE800092EA, 0x800092F20001CDA6, 0x0001A9AB800092F6,
-    0x0001BDAD0001C1AA, 0x0001A1AF0001BDAC, 0x8000928E0001A9AE,
-    0x000FED8100001582, 0x800092FA000FED80, 0x800092FE000258BE,
-    0x000F09A880009302, 0x000F1DAA000F15A9, 0x000F79AC000F69AB,
-    0x000F69AE000F79AD, 0x000F29B0000F61AF, 0x000FC9B2000F29B1,
-    0x000FD9B4000FC9B3, 0x000FC9B6000FD9B5, 0x000FD1B8000FC9B7,
-    0x000FD1BA000FD1B9, 0x000FB1BC000FD1BB, 0x000FADBE000FB1BD,
-    0x80009306000FA5BF, 0x000FC1B5000FC9B6, 0x000FA5B38000930A,
-    0x000308878000930E, 0x000F2DBF00033C86, 0x000F2DBD000F2DBE,
-    0x000FDDBB000F2DBC, 0x80009312000FDDBA, 0x000FF5A38000921E,
-    0x8000931A80009316, 0x800093228000931E, 0x000F91A5000F99A6,
-    0x000F8DAB80009326, 0x8000932A000F8DAA, 0x000F7DAF8000932E,
-    0x000F7DAD000F7DAE, 0x00011DA8000F7DAC, 0x000185AA000185A9,
-    0x0001A9AC000199AB, 0x0001D1AE0001D1AD, 0x800093320001D1AF,
-    0x8000933A80009336, 0x800093428000933E, 0x8000934A80009346,
-    0x000071B88000934E, 0x000071BA000071B9, 0x0000D1BC000071BB,
-    0x0000D1BE0000D1BD, 0x0001B1B00000D1BF, 0x000181B20001B9B1,
-    0x000051B4000181B3, 0x000051B6000051B5, 0x80009352000051B7,
-    0x8000935A80009356, 0x00000D808000935E, 0x00003D8200003581,
-    0x8000936680009362, 0x0002A4848000936A, 0x00019086000124BE,
-    0x8000937200011487, 0x00014DB380009376, 0x8000937A8000936E,
-    0x800093828000937E, 0x000161B5000111B6, 0x00013DBB80009386,
-    0x8000938A000135BA, 0x0000D5BF8000938E, 0x000115BD0000D5BE,
-    0x80009392000115BC, 0x80009396000289A3, 0x0002D5A6000340BE,
-    0x8000939E8000939A, 0x0002F1AA0002A5A5, 0x800093A20002F9AB,
-    0x000311AE800093A6, 0x0002D1AC000311AF, 0x0001E8E10002D1AD,
-    0x000214E3800093AA, 0x800093AE0003D484, 0x800093B6800093B2,
-    0x800093BA00041084, 0x800093C2800093BE, 0x800093CA800093C6,
-    0x00028CEF800093CE, 0x000339A8800093D2, 0x000341AA000339A9,
-    0x000341AC000341AB, 0x000371AE000349AD, 0x0000DD80000371AF,
-    0x0000198200000981, 0x800093DA800093D6, 0x00030887800093E2,
-    0x000095B800041C86, 0x0000A9BA00009DB9, 0x0000B9BC0000A9BB,
-    0x0000A9BE0000B9BD, 0x0000E1B00000A9BF, 0x0000FDB20000EDB1,
-    0x0000D9B40000F5B3, 0x0000BDB60000D9B5, 0x000F58EF0000ADB7,
-    0x800093EA800093E6, 0x800093EE000544BE, 0x800093F6800093F2,
-    0x800093FE800093FA, 0x8000940680009402, 0x000E00E38000940A,
-    0x000C90E18000940E, 0x00029DB680009412, 0x8000941A80009416,
-    0x8000941E00029DB5, 0x8000942200029DB3, 0x000149BE80009426,
-    0x00017DBC000149BF, 0x000171BA000175BD, 0x8000942A00017DBB,
-    0x800094328000942E, 0x80009436000EA4E1, 0x0004DCBE000E50E3,
-    0x8000943A00002CE3, 0x0006ECBE000118E1, 0x0019E0BE0000F0EF,
-    0x000075808000943E, 0x0000758200007D81, 0x000209A3000150EF,
-    0x8000944680009442, 0x800094528000944A, 0x000209A5000209A6,
-    0x0001E9AB80009456, 0x000448860001E5AA, 0x0001DDAF0004DC87,
-    0x0001E1AD0001DDAE, 0x0006B5A80001E9AC, 0x000701AA000701A9,
-    0x000705AC00071DAB, 0x000705AE00070DAD, 0x8000944E00073DAF,
-    0x8000945E8000945A, 0x8000946680009462, 0x8000946E8000946A,
-    0x000719B880009472, 0x000731BA000719B9, 0x0007D1BC000731BB,
-    0x0007C1BE0007DDBD, 0x000745B00007F1BF, 0x000745B200074DB1,
-    0x00074DB4000759B3, 0x000731B6000731B5, 0x000789B300072DB7,
-    0x8000947A80009476, 0x8000947E800093DE, 0x00078DB500078DB6,
-    0x000789BB80009482, 0x80009486000781BA, 0x0007F5BF8000948A,
-    0x00078DBD000789BE, 0x8000948E000791BC, 0x800094920007CDA3,
-    0x0007C9A680009496, 0x8000949E8000949A, 0x0007C5AA0007C9A5,
-    0x800094A20007CDAB, 0x0007CDAE800094A6, 0x0007D5AC0007B1AF,
-    0x0006D1A80007C9AD, 0x00012DAA0006D9A9, 0x00013DAC000125AB,
-    0x00012DAE000125AD, 0x800094AA000125AF, 0x800094B2800094AE,
-    0x00001980800094B6, 0x0000058200001981, 0x000189B8800094BA,
-    0x00019DBA000189B9, 0x0001B9BC000195BB, 0x000079BE0001B9BD,
-    0x00015DB0000079BF, 0x0001E1B20001E1B1, 0x0001EDB40001FDB3,
-    0x0001D5B60001DDB5, 0x0002C8840001BDB7, 0x800094BE0001A1B3,
-    0x0001A1B6800094C2, 0x800094CA800094C6, 0x00019DBA0001B1B5,
-    0x0001A886000121BB, 0x00012DBE0001E087, 0x000131BC00011DBF,
-    0x0001E5A300012DBD, 0x800094D2800094CE, 0x800094DA800094D6,
-    0x0001F5A50001E5A6, 0x000165AB800094DE, 0x800094E20001D9AA,
-    0x000159AF800094E6, 0x000169AD000169AE, 0x800094EA000175AC,
-    0x800094EE000274E3, 0x800094F20001BCE1, 0x800094FA800094F6,
-    0x80009502800094FE, 0x8000950A80009506, 0x800095128000950E,
-    0x8000951600076CEF, 0x8000951E8000951A, 0x00001D8100001D82,
-    0x8000952200001D80, 0x8000952A80009526, 0x0003E48480009532,
-    0x00032C8780009536, 0x8000953A001CCC86, 0x800095428000953E,
-    0x8000954A80009546, 0x800095528000954E, 0x8000955A80009556,
-    0x800095628000955E, 0x00032DA880009566, 0x000331AA000331A9,
-    0x000379AC000369AB, 0x000369AE000379AD, 0x000329B0000361AF,
-    0x0000C9B2000329B1, 0x0000C1B40000C5B3, 0x0000C1B60000C1B5,
-    0x0000C1B80000C1B7, 0x0000C1BA0000C1B9, 0x0000C1BC0000C1BB,
-    0x0000C5BE0000CDBD, 0x00074CE30000BDBF, 0x0007ACE18000956A,
-    0x800095728000956E, 0x8000957A80009576, 0x800095828000957E,
-    0x8000958A80009586, 0x0006C4EF8000958E, 0x8000959680009592,
-    0x000079818000959A, 0x0002D1B300007180, 0x000299B500007182,
-    0x001D1484800095A2, 0x001CC086000291B6, 0x000285BB001DE887,
-    0x000179BD000285BA, 0x00016DBF000175BC, 0x800095A600016DBE,
-    0x800095AE800095AA, 0x001CDC84800095B2, 0x800095BA800095B6,
-    0x001CE8BE0006F4EF, 0x800095BE000654E1, 0x001FF08400003CE3,
-    0x800095C6800095C2, 0x000259A3800095CA, 0x800095D2800095CE,
-    0x800095DA800095D6, 0x000211A5000219A6, 0x00020DAB800095DE,
-    0x800095E200020DAA, 0x0001E5AF800095E6, 0x0001F1AD0001E5AE,
-    0x001E5DA80001FDAC, 0x001E79AA001E61A9, 0x001E69AC001E79AB,
-    0x001E99AE001E69AD, 0x8000959E001E99AF, 0x800095EE800095EA,
-    0x800095F6800095F2, 0x800095FE800095FA, 0x001F51B880009602,
-    0x001F51BA001F51B9, 0x001F71BC001F51BB, 0x001F71BE001F71BD,
-    0x001EE9B0001F71BF, 0x001EF9B2001EE9B1, 0x001EE9B4001EF9B3,
-    0x001F7DB6001EE9B5, 0x001EF1B3001F71B7, 0x0000158100002D82,
-    0x8000952E0000B180, 0x001EE1B5001EE9B6, 0x001E2DBB80009606,
-    0x8000960A001E2DBA, 0x001E2DBF8000960E, 0x001E2DBD001E2DBE,
-    0x80009612001E2DBC, 0x00006886001EB5A3, 0x001EADA600004C87,
-    0x8000961A80009616, 0x001E69AA001EA5A5, 0x8000961E001E69AB,
-    0x001E69AE80009622, 0x001E69AC001E69AF, 0x001E55A8001E69AD,
-    0x001E8DAA001E85A9, 0x001E9DAC001E85AB, 0x001E81AE001E81AD,
-    0x80009626001EB5AF, 0x8000962E8000962A, 0x8000963680009632,
-    0x8000963E8000963A, 0x001EA9B880009642, 0x00016DBA001EA9B9,
-    0x00017DBC000165BB, 0x00016DBE000165BD, 0x001ED5B0000165BF,
-    0x001ED5B2001EDDB1, 0x001EB9B4001EA9B3, 0x001EA9B6001EB9B5,
-    0x001E05B3001EA1B7, 0x8000964A80009646, 0x800096528000964E,
-    0x001E15B5001E01B6, 0x001E19BB80009656, 0x8000965A001E11BA,
-    0x001E7DBF8000965E, 0x001E7DBD001E7DBE, 0x80009662001E01BC,
-    0x80009666001E41A3, 0x001E45A68000966A, 0x800096728000966E,
-    0x001E55AA001E51A5, 0x80009676001E5DAB, 0x001E39AE8000967A,
-    0x001E45AC001E39AF, 0x0002C1A8001E39AD, 0x0002D1AA0002D1A9,
-    0x0002FDAC0002E5AB, 0x00033DAE000335AD, 0x80009682000335AF,
-    0x80009686000268BE, 0x8000968E8000968A, 0x8000969680009692,
-    0x0003D1B88000969A, 0x0003E1BA0003D9B9, 0x000391BC0003E1BB,
-    0x000395BE00039DBD, 0x00034DB0000389BF, 0x00035DB2000355B1,
-    0x00034DB4000355B3, 0x0003F1B60003F1B5, 0x000359800003F1B7,
-    0x0003658200036D81, 0x8000969E000255B3, 0x0003B9B6000279B5,
-    0x00040086800096A2, 0x000385BA0003E487, 0x000385BC00039DBB,
-    0x000385BE00038DBD, 0x000219A3000339BF, 0x800096AA800096A6,
-    0x000BAC85800096AE, 0x000235A50003F5A6, 0x0003D1AB800096B2,
-    0x800096B60003C9AA, 0x000375AF800096BA, 0x0003C1AD0003C9AE,
-    0x800096BE0003C9AC, 0x800096C6800096C2, 0x0003CCEF800096CA,
-    0x800096CE001F60E1, 0x001E80E3001E54E3, 0x0001F8E1800096D2,
-    0x800096D6001A7CEF, 0x800096DE800096DA, 0x00007D80800096E2,
-    0x00001D8200001581, 0x800096E600024084, 0x000554BE800096EA,
-    0x0004C086800096F2, 0x800096F600054887, 0x800096FE800096FA,
-    0x8000970680009702, 0x8000970E8000970A, 0x8000971680009712,
-    0x8000971E8000971A, 0x8000972680009722, 0x0005A484001EC4EF,
-    0x8000972A001EDCE1, 0x8000972E001E2CE3, 0x8000973680009732,
-    0x0001EDB38000973A, 0x800097428000973E, 0x8000974A80009746,
-    0x0001EDB50001EDB6, 0x000125BB8000974E, 0x80009752000179BA,
-    0x000119BF80009756, 0x000135BD000129BE, 0x000621A8000135BC,
-    0x000625AA00062DA9, 0x000625AC00063DAB, 0x000625AE00062DAD,
-    0x800096EE00068DAF, 0x0007C58100001582, 0x8000975A0007C580,
-    0x8000975E000338BE, 0x000745B880009762, 0x000745BA00074DB9,
-    0x000775BC000755BB, 0x00076DBE000765BD, 0x0006FDB0000765BF,
-    0x0006A1B20006A1B1, 0x0006ADB40006A1B3, 0x000695B600069DB5,
-    0x0006ADA300077DB7, 0x0000288680009766, 0x8000976A00034C87,
-    0x0006ADA50006ADA6, 0x000665AB8000976E, 0x80009772000639AA,
-    0x000659AF80009776, 0x000675AD000669AE, 0x8000977A000675AC,
-    0x0007A5B38000977E, 0x0007A5B580009782, 0x8000978A80009786,
-    0x8000978E0007A5B6, 0x000761BB8000967E, 0x000761BD000769BA,
-    0x000761BF000779BC, 0x00065DA8000769BE, 0x000665AA00066DA9,
-    0x000665AC00067DAB, 0x000665AE00066DAD, 0x8000979200069DAF,
-    0x8000979A80009796, 0x800097A28000979E, 0x800097AA800097A6,
-    0x0006F5B8800097AE, 0x0006F5BA0006FDB9, 0x000699BC000689BB,
-    0x000689BE000699BD, 0x0006E5B0000681BF, 0x0006E5B20006EDB1,
-    0x0006E5B40006FDB3, 0x0006D1B60006EDB5, 0x0006EDA30006CDB7,
-    0x800097B6800097B2, 0x800097BE800097BA, 0x0006EDA50006EDA6,
-    0x000629AB800097C2, 0x800097C6000621AA, 0x000629AF800097CA,
-    0x000629AD000621AE, 0x00005D80000631AC, 0x0000658200006D81,
-    0x000198BE00018084, 0x0000D887800097D2, 0x800097D6000F7C86,
-    0x800097DE800097DA, 0x800097E6800097E2, 0x800097EE800097EA,
-    0x00023DA8800097F2, 0x000299AA000281A9, 0x0002B9AC000291AB,
-    0x0002D9AE0002B9AD, 0x800097F60002D9AF, 0x800097FE800097FA,
-    0x80009802000EA4BF, 0x8000980A80009806, 0x000371B88000980E,
-    0x000371BA000371B9, 0x0003D5BC000371BB, 0x0003D5BE0003DDBD,
-    0x0002ADB00003CDBF, 0x0002B1B20002B1B1, 0x000351B400028DB3,
-    0x000351B6000351B5, 0x000211B3000351B7, 0x8000981680009812,
-    0x000350848000981A, 0x000231B5000239B6, 0x000275BB8000981E,
-    0x80009822000269BA, 0x000395BF80009826, 0x000259BD000395BE,
-    0x8000982A000265BC, 0x8000982E000255A3, 0x00027DA6000380BE,
-    0x8000983680009832, 0x00022DAA000275A5, 0x8000983A000231AB,
-    0x0003D1AE80009842, 0x000221AC0003D1AF, 0x0001D58000021DAD,
-    0x0001E1820001DD81, 0x800098460002E9B3, 0x0002E9B60002F9B5,
-    0x8000984E8000984A, 0x0002D5BA80009852, 0x000115BC00010DBB,
-    0x000109BE00011DBD, 0x000D2086000109BF, 0x8000985600028887,
-    0x8000985E8000985A, 0x0006F8EF80009862, 0x0007D8E180009866,
-    0x000634E38000986A, 0x800098728000986E, 0x8000987A80009876,
-    0x800098828000987E, 0x800098860000ACEF, 0x8000988E8000988A,
-    0x8000989680009892, 0x000C6C84000150EF, 0x8000989A0006DCE1,
-    0x8000989E00061CE3, 0x800098A2000068E3, 0x000225A300012CE1,
-    0x0000198100002D82, 0x800098A600007580, 0x000235A5000225A6,
-    0x0001C1AB00080085, 0x800098AA000219AA, 0x0001C5AF000C24BE,
-    0x0001D1AD0001C5AE, 0x000DD5A80001D9AC, 0x000DEDAA000DE5A9,
-    0x000DFDAC000DE5AB, 0x000DEDAE000DE5AD, 0x8000983E000DE5AF,
-    0x00030886800098AE, 0x800098B200038C87, 0x800098BA800098B6,
-    0x000EC5B8800098BE, 0x000EC5BA000ECDB9, 0x000EF9BC000ED1BB,
-    0x000E99BE000EF9BD, 0x000D9DB0000E95BF, 0x000E4DB2000E45B1,
-    0x000E41B4000E45B3, 0x000E41B6000E41B5, 0x000F01B8000E41B7,
-    0x000F05BA000F0DB9, 0x000F05BC000F1DBB, 0x000F31BE000F0DBD,
-    0x000FF5B0000F2DBF, 0x000FCDB2000FF9B1, 0x000FC1B4000FC1B3,
-    0x000FC1B6000FC1B5, 0x000E2DA8000FC1B7, 0x000FD5AA000FD5A9,
-    0x000FF5AC000FEDAB, 0x000F95AE000F9DAD, 0x800098C2000F8DAF,
-    0x800097CE800098C6, 0x800098CE800098CA, 0x800098D6800098D2,
-    0x000E1DB3800098DA, 0x800098E2800098DE, 0x800098EA800098E6,
-    0x000E3DB5000E5DB6, 0x000E7DBB800098EE, 0x800098F2000E71BA,
-    0x000E45BF800098F6, 0x000E55BD000E45BE, 0x800098FA000E59BC,
-    0x800098FE000E59A3, 0x000E19A680009902, 0x8000990A80009906,
-    0x000E35AA000E79A5, 0x8000990E000E39AB, 0x000E01AE80009912,
-    0x000E1DAC000E01AF, 0x0001B1A8000E11AD, 0x0001B1AA0001B1A9,
-    0x000171AC0001B1AB, 0x000175AE000171AD, 0x0000CD8000016DAF,
-    0x0000198200000981, 0x8000991A80009916, 0x000170BE00017884,
-    0x0000E9B880009922, 0x000089BA0000E9B9, 0x00009DBC000085BB,
-    0x000081BE000081BD, 0x000115B00000B5BF, 0x000115B200011DB1,
-    0x0000F9B40000E9B3, 0x0000E9B60000F9B5, 0x000239B30000E1B7,
-    0x0004888680009926, 0x8000992A00037887, 0x0002F9B50002F1B6,
-    0x0002D9BB8000992E, 0x800099320002D1BA, 0x00033DBF80009936,
-    0x00033DBD00033DBE, 0x8000993A0002C1BC, 0x8000993E00027DA3,
-    0x0002B5A680009942, 0x8000994A80009946, 0x000295AA0002BDA5,
-    0x8000994E00029DAB, 0x000379AE80009952, 0x000285AC000379AF,
-    0x001AC182000379AD, 0x80009956001AC983, 0x001EB5868000995A,
-    0x001E318400122987, 0x0012A18A001EBD85, 0x8000995E0012BD8B,
-    0x0016918E80009962, 0x00137D8C00173D8F, 0x000A51920016318D,
-    0x80009966000AE593, 0x000EF196006CDCEF, 0x000A7594000EED97,
-    0x0002099A000E3995, 0x8000996A0002C59B, 0x800099728000996E,
-    0x0002DD9C000F3CE1, 0x000F219F0076A8E3, 0x000BA59D000FE19E,
-    0x000B819B000B2D9C, 0x0017399900169D9A, 0x0012E59700179198,
-    0x0013299500133D96, 0x001F159300139194, 0x001F9191001F1592,
-    0x001B018F001B7190, 0x8000997A80009976, 0x800099828000997E,
-    0x80009986000F60EF, 0x000528BE8000998A, 0x00052484000FE4E3,
-    0x00002981000CA8E1, 0x8000999200002580, 0x00050C8700002182,
-    0x800099960004A486, 0x8000999E8000999A, 0x0000D8EF800099A2,
-    0x000190E1000E88EF, 0x00000CE3000EFCE1, 0x800099A6000E1CE3,
-    0x800099AE800099AA, 0x800099B6800099B2, 0x00070484800099BA,
-    0x800099BE000704BE, 0x800099C6800099C2, 0x800099CE800099CA,
-    0x800099D6800099D2, 0x0003D5A3800099DA, 0x000311A10003D5A2,
-    0x8000991E000FC5A0, 0x800099DE8000998E, 0x800099E6800099E2,
-    0x800099EE800099EA, 0x00059DA8800099F2, 0x000639AA000621A9,
-    0x000619AC000631AB, 0x000679AE000619AD, 0x00060DB0000675AF,
-    0x000611B2000611B1, 0x00063DB4000625B3, 0x0006D5B60006DDB5,
-    0x0006CDB80006F5B7, 0x000681BA0006ADB9, 0x000681BC000681BB,
-    0x000731BE000731BD, 0x000691B3000731BF, 0x0000198100001982,
-    0x800099F600009180, 0x0006B1B50006B9B6, 0x0006F5BB800099FA,
-    0x800099FE0006E9BA, 0x000639BF80009A02, 0x0006D1BD0006D9BE,
-    0x80009A060006E5BC, 0x0000E8860006D5A3, 0x0006FDA600008487,
-    0x80009A0E80009A0A, 0x0006ADAA0006F5A5, 0x80009A120006B1AB,
-    0x00069DAE80009A16, 0x0006A1AC00067DAF, 0x80009A1A000695AD,
-    0x006929B580009A1E, 0x006C51B3006801B4, 0x006D19B1006D11B2,
-    0x80009A22007071B0, 0x80009A2A80009A26, 0x80009A3280009A2E,
-    0x80009A3A80009A36, 0x007801A780009A3E, 0x007D69A5007DE9A6,
-    0x0001F9A3007C01A4, 0x000159A1000151A2, 0x007169AF80009A42,
-    0x007001AD007161AE, 0x007551AB007549AC, 0x0079F5A9007401AA,
-    0x0006F9A800795DA8, 0x000605AA00060DA9, 0x000605AC00061DAB,
-    0x00062DAE000625AD, 0x80009A46000625AF, 0x80009A4E80009A4A,
-    0x80009A5680009A52, 0x80009A5E80009A5A, 0x000625B880009A62,
-    0x000625BA00062DB9, 0x0001DDBC0001C5BB, 0x0001CDBE0001C5BD,
-    0x00065DB00001C5BF, 0x00062DB2000625B1, 0x000625B400063DB3,
-    0x000625B600062DB5, 0x00058DB300061DB7, 0x80009A6A80009A66,
-    0x80009A72000340BE, 0x00058DB50005A5B6, 0x0005E1BB80009A76,
-    0x0003A484000599BA, 0x0005E9BF80009A7A, 0x0005E9BD0005E1BE,
-    0x000049820005F9BC, 0x0003F5800005C9A3, 0x0005E1A600004981,
-    0x80009A8280009A7E, 0x0005DDAA0005C9A5, 0x007CC8860005A5AB,
-    0x0005A5AE0003FC87, 0x0005BDAC0005ADAF, 0x00022DA80005ADAD,
-    0x00025DAA000245A9, 0x000279AC000255AB, 0x000399AE000279AD,
-    0x80009A86000399AF, 0x80009A8E80009A8A, 0x80009A9680009A92,
-    0x80009A9E80009A9A, 0x00039DB880009AA2, 0x00036DBA000365B9,
-    0x00037DBC000365BB, 0x00036DBE000365BD, 0x0003C1B0000365BF,
-    0x0003C1B20003C1B1, 0x0003C1B40003C1B3, 0x0003C1B60003C1B5,
-    0x80009AA60003C1B7, 0x80009AAE80009AAA, 0x80009AB680009AB2,
-    0x80009ABA0003E0EF, 0x007D68BE80009ABE, 0x80009AC200031CE3,
-    0x80009AC60001FCE1, 0x80009ACE80009ACA, 0x000395B380009AD2,
-    0x80009ADA80009AD6, 0x007D1C8480009AE2, 0x0003B1B50003B9B6,
-    0x000245BB80009AE6, 0x007CA886000245BA, 0x000245BF007DA487,
-    0x000255BD000245BE, 0x00061CE1000255BC, 0x000650E3000794E1,
-    0x0000AD800006F4E3, 0x0000D1820000D181, 0x80009AEE80009AEA,
-    0x80009AF680009AF2, 0x80009AFE80009AFA, 0x00068CEF80009B02,
-    0x000315A3000624EF, 0x80009B0A80009B06, 0x80009B1280009B0E,
-    0x000331A5000339A6, 0x0002C5AB007C9084, 0x80009B160002C5AA,
-    0x0002C5AF80009B1A, 0x0002D5AD0002C5AE, 0x007E11A80002D5AC,
-    0x007E25AA007E19A9, 0x007E49AC007E59AB, 0x007EBDAE007EB5AD,
-    0x80009ADE007EB5AF, 0x80009B2280009B1E, 0x80009B2A80009B26,
-    0x80009B3280009B2E, 0x007F4DB880009B36, 0x007F51BA007F55B9,
-    0x007F79BC007F51BB, 0x007F6DBE007F65BD, 0x007ECDB0007F65BF,
-    0x007EDDB2007ED5B1, 0x007ECDB4007ED5B3, 0x007F7DB6007F75B5,
-    0x007EADB3007F75B7, 0x80009B3E80009B3A, 0x80009B4680009B42,
-    0x007EC1B5007E35B6, 0x007E11BB80009B4A, 0x0003E0BE007E09BA,
-    0x007FB1BF80009B4E, 0x007E01BD007E09BE, 0x00005D82007E09BC,
-    0x00004580007EE9A3, 0x007E71A600005D81, 0x80009B5680009B52,
-    0x007E4DAA007E85A5, 0x00004886007E55AB, 0x007E4DAE00006C87,
-    0x007E4DAC007FF5AF, 0x007EB5B0007E45AD, 0x007EC1B2007EC1B1,
-    0x007EC1B4007EC1B3, 0x007EF5B6007EFDB5, 0x000109B8000139B7,
-    0x000119BA000109B9, 0x000109BC000119BB, 0x0001BDBE000109BD,
-    0x80009B5A0001B5BF, 0x80009B6280009B5E, 0x80009B6680009A6E,
-    0x80009B6E80009B6A, 0x007E2DA880009B72, 0x007EF5AA007E39A9,
-    0x007ED5AC007ECDAB, 0x007ED5AE007EDDAD, 0x007E1DB3007ECDAF,
-    0x80009B7A80009B76, 0x80009B8280009B7E, 0x007E3DB5007E3DB6,
-    0x007E65BB80009B86, 0x80009B8A007E11BA, 0x007E59BF80009B8E,
-    0x007E75BD007E65BE, 0x80009B92007E75BC, 0x80009B96007E59A3,
-    0x007E79A680009B9A, 0x80009BA280009B9E, 0x007E55AA007E79A5,
-    0x80009BA6007E21AB, 0x007E21AE80009BAA, 0x007E31AC007E1DAF,
-    0x0002D5A8007E31AD, 0x000349AA0002DDA9, 0x000359AC000349AB,
-    0x00034DAE000345AD, 0x80009BAE000345AF, 0x80009BB680009BB2,
-    0x0003F4BE80009BBA, 0x80009BC680009BC2, 0x000381B880009BCA,
-    0x000385BA00038DB9, 0x000389BC000399BB, 0x0003B5BE0003BDBD,
-    0x00033DB000007DBF, 0x0003E1B20003E1B1, 0x0003E1B40003E1B3,
-    0x0003D1B60003E9B5, 0x0000A5800003D1B7, 0x0000A5820000AD81,
-    0x80009BCE0003CDB3, 0x0003C9B60003DDB5, 0x0002E08480009BD2,
-    0x0003F5BA80009BD6, 0x000339BC000329BB, 0x000329BE000339BD,
-    0x007F80E3000321BF, 0x007EB8E180009BDA, 0x0003E08780009BDE,
-    0x80009BE200047486, 0x80009BEA80009BE6, 0x80009BF280009BEE,
-    0x007F28EF80009BF6, 0x80009BFE80009BFA, 0x80009C0680009C02,
-    0x80009C0E80009C0A, 0x0003D8EF80009C12, 0x80009C1A80009C16,
-    0x000390E380009C1E, 0x0001D8E180009C22, 0x0000158080009C26,
-    0x0000058200001981, 0x00054CBE80009C2A, 0x000209A380009C32,
-    0x000219A5000440BE, 0x80009C3A80009C36, 0x00026C8400020DA6,
-    0x0002EDAB80009C3E, 0x0002FDAD000231AA, 0x0002E5AF0002FDAC,
-    0x000440860002EDAE, 0x80009C420005B487, 0x80009C4A80009C46,
-    0x80009C5280009C4E, 0x80009C560001CCEF, 0x80009C5A007ED4E1,
-    0x80009C5E007E60E3, 0x80009C6680009C62, 0x0001FDB380009C6A,
-    0x80009C7280009C6E, 0x80009C7A80009C76, 0x000145B5000171B6,
-    0x00013DBB80009C7E, 0x80009C8200013DBA, 0x00011DBF80009C86,
-    0x00011DBD00011DBE, 0x000635A800011DBC, 0x000635AA00063DA9,
-    0x00068DAC00069DAB, 0x0006DDAE0006B1AD, 0x80009C2E0006D1AF,
-    0x80009C8E80009C8A, 0x00001D8080009C92, 0x0000BD8200000981,
-    0x0006BDB880009C96, 0x00074DBA000745B9, 0x00075DBC000745BB,
-    0x000741BE000741BD, 0x0006B9B0000741BF, 0x000689B20006B9B1,
-    0x000699B4000689B3, 0x00068DB6000699B5, 0x0006BDA3000685B7,
-    0x0002808480009C9A, 0x80009CA280009C9E, 0x000605A5000631A6,
-    0x00067DAB80009CA6, 0x0000088600067DAA, 0x00065DAF00036087,
-    0x00065DAD00065DAE, 0x80009CAA00065DAC, 0x80009CAE00079DB3,
-    0x00075DB680009CB2, 0x80009CBA80009CB6, 0x000779BA000755B5,
-    0x80009CBE000745BB, 0x000745BE80009CC2, 0x000751BC00074DBF,
-    0x000699B000074DBD, 0x0006A1B20006A5B1, 0x0006A5B40006BDB3,
-    0x0006A5B60006ADB5, 0x0006E9B80006D9B7, 0x0006F9BA0006E9B9,
-    0x0006D9BC0006F1BB, 0x00011DBE0006D9BD, 0x80009BBE000115BF,
-    0x80009CCA80009CC6, 0x80009CD280009CCE, 0x80009CDA80009CD6,
-    0x000625A880009CDE, 0x000631AA000631A9, 0x0006F9AC000609AB,
-    0x0006E9AE0006F9AD, 0x0006D9A30006E9AF, 0x80009CE680009CE2,
-    0x80009CEE80009CEA, 0x000611A5000619A6, 0x000601AB80009CF2,
-    0x80009CF600063DAA, 0x000609AF80009CFA, 0x000609AD000601AE,
-    0x80009CFE000615AC, 0x00001D8100001D82, 0x80009D0200001D80,
-    0x80009D0A80009D06, 0x0001AC84000010BE, 0x0001088780009D12,
-    0x80009D16000F4C86, 0x80009D1E80009D1A, 0x000291A880009D22,
-    0x0002C9AA000299A9, 0x0002DDAC0002C5AB, 0x0002C1AE0002C1AD,
-    0x80009D260002F5AF, 0x80009D2E80009D2A, 0x0000008D80009D32,
-    0x80009D3A80009D36, 0x0003C5B880009D3E, 0x0003C5BA0003CDB9,
-    0x0003CDBC0003DDBB, 0x0003F5BE0003FDBD, 0x00028DB000039DBF,
-    0x000321B2000321B1, 0x000321B4000321B3, 0x000321B6000321B5,
-    0x000241B3000321B7, 0x80009D4680009D42, 0x80009D4E80009D4A,
-    0x000255B5000241B6, 0x0003E5BB80009D52, 0x80009D56000249BA,
-    0x0003E5BF80009D5A, 0x0003F5BD0003E5BE, 0x80009D5E0003F5BC,
-    0x80009D62000608E1, 0x80009D660007BCE3, 0x80009D6E80009D6A,
-    0x00003D8080009D72, 0x0000198200000981, 0x80009D7A80009D76,
-    0x000CF4BE0003F884, 0x0000F0E3000744EF, 0x0001B8E180009D82,
-    0x000CF086000D0484, 0x80009D860002A487, 0x80009D8E80009D8A,
-    0x80009D92000C28BE, 0x000014EF80009D96, 0x80009D9E80009D9A,
-    0x000381A380009DA2, 0x80009DAA80009DA6, 0x80009DB280009DAE,
-    0x000395A5000381A6, 0x000225AB80009DB6, 0x80009DBA000389AA,
-    0x000225AF80009DBE, 0x000235AD000225AE, 0x80009DC2000235AC,
-    0x80009DCA80009DC6, 0x80009DD280009DCE, 0x80009DDA80009DD6,
-    0x80009DDE000628EF, 0x80009DE200063CE1, 0x00001980000134E3,
-    0x00007D8200000981, 0x000235B380009DE6, 0x80009DF280009DEA,
-    0x80009DF6000FC0BE, 0x000225B50001F5B6, 0x0001D5BB80009DFA,
-    0x80009DFE0001C9BA, 0x000179BF80009E02, 0x0001C5BD000171BE,
-    0x000DE5A80001C5BC, 0x000DFDAA000DF5A9, 0x000DEDAC000DF5AB,
-    0x000E3DAE000E35AD, 0x80009D7E000E35AF, 0x0000888680009DEE,
-    0x80009E0600034087, 0x80009E0E80009E0A, 0x000EDDB880009E12,
-    0x000EE5BA000EEDB9, 0x000E99BC000E8DBB, 0x000E8DBE000E85BD,
-    0x000E4DB0000E85BF, 0x000E51B2000E51B1, 0x000EF5B4000E51B3,
-    0x000EF5B6000EFDB5, 0x000D75A3000EEDB7, 0x80009E1A80009E16,
-    0x80009E2280009E1E, 0x000D65A5000EB5A6, 0x000E95AB80009E26,
-    0x80009E2A000E89AA, 0x000E39AF80009E2E, 0x000E85AD000E31AE,
-    0x80009E32000E85AC, 0x80009E36000FD9B3, 0x000FFDB680009E3A,
-    0x80009E4280009E3E, 0x000FDDBA000FF5B5, 0x80009E46000FA1BB,
-    0x000FA1BE80009E4A, 0x000FB1BC000FA9BF, 0x000E81B8000FB1BD,
-    0x000E81BA000E81B9, 0x000E81BC000E81BB, 0x000E81BE000E81BD,
-    0x000E1DB0000EBDBF, 0x000E61B2000E61B1, 0x000E61B4000E61B3,
-    0x000E51B6000E69B5, 0x000E45A8000E51B7, 0x000E55AA000E5DA9,
-    0x000E79AC000E6DAB, 0x000E6DAE000E79AD, 0x80009E4E000E65AF,
-    0x0000118100001982, 0x80009E5200001D80, 0x80009E5680009D0E,
-    0x000E9DA380009E5A, 0x0000288600028C84, 0x80009E6200014887,
-    0x000EB1A5000EB9A6, 0x000EE5AB80009E66, 0x80009E6A000E99AA,
-    0x000EEDAF80009E6E, 0x000EF5AD000EE5AE, 0x80009E72000EF5AC,
-    0x80009E76000E19B3, 0x0001E1B680009E7A, 0x80009E8280009E7E,
-    0x0001DDBA000E31B5, 0x80009E86000161BB, 0x000161BE80009E8A,
-    0x000171BC000169BF, 0x000DC9A8000171BD, 0x000DD5AA000DD5A9,
-    0x000DFDAC000DE5AB, 0x00022DAE000225AD, 0x80009E8E000225AF,
-    0x80009E9680009E92, 0x0000008C80009E9A, 0x80009EA280009E9E,
-    0x000281B880009EA6, 0x000285BA00028DB9, 0x000289BC000299BB,
-    0x0002B5BE0002BDBD, 0x00025DB000033DBF, 0x0002E1B20002E1B1,
-    0x0002E1B40002E1B3, 0x0002D1B60002E9B5, 0x000D5DA30002D1B7,
-    0x80009EAE80009EAA, 0x80009EB680009EB2, 0x000D75A50002A5A6,
-    0x000225AB80009EBA, 0x80009EBE000299AA, 0x00022DAF80009EC2,
-    0x000235AD000225AE, 0x00001181000235AC, 0x000EC8EF00006980,
-    0x80009EC600001982, 0x80009ED280009ECA, 0x0003EC8480009ED6,
-    0x0004A887000468BE, 0x000D04E10004CC86, 0x000E1CE380009EDA,
-    0x000291A880009EDE, 0x0002A1AA000299A9, 0x0002D1AC0002A1AB,
-    0x0002D5AE0002DDAD, 0x80009EE20002C9AF, 0x80009EEA80009EE6,
-    0x80009EF280009EEE, 0x80009EFA80009EF6, 0x000179B880009EFE,
-    0x0001CDBA000179B9, 0x0001DDBC0001C5BB, 0x0001C5BE0001C5BD,
-    0x0002B9B00001F5BF, 0x000285B200028DB1, 0x000155B400014DB3,
-    0x000155B600015DB5, 0x80009F02000149B7, 0x80009F0A80009F06,
-    0x0005108480009F0E, 0x000EACEF000418BE, 0x000190E1000E40EF,
-    0x00003CE3000EACE1, 0x80009F12000EB0E3, 0x80009F1A80009F16,
-    0x000269B380009F1E, 0x80009F2680009F22, 0x80009F2E80009F2A,
-    0x000219B500021DB6, 0x000209BB80009F32, 0x80009F3600023DBA,
-    0x0001D9BF80009F3A, 0x0001D9BD0001D1BE, 0x0005A9A3000211BC,
-    0x80009F3E80009ECE, 0x80009F4680009F42, 0x0005D9A50005DDA6,
-    0x0005C9AB80009F4A, 0x000308860005FDAA, 0x000619AF00037C87,
-    0x000619AD000611AE, 0x00079D800005D1AC, 0x0007A5820007AD81,
-    0x80009F4E0006B9B3, 0x00077DB6000775B5, 0x80009F5680009F52,
-    0x000759BA80009F5A, 0x000749BC000759BB, 0x0007F9BE000749BD,
-    0x000659A80007F9BF, 0x00067DAA000659A9, 0x0007B1AC000671AB,
-    0x0007B1AE0007B1AD, 0x80009F5E0007B1AF, 0x80009F6680009F62,
-    0x80009F6E80009F6A, 0x80009F7680009F72, 0x000771B880009F7A,
-    0x000771BA000771B9, 0x000715BC000771BB, 0x000715BE00071DBD,
-    0x0007D1B000070DBF, 0x0007D1B20007D1B1, 0x000771B40007D1B3,
-    0x000771B6000771B5, 0x0007F5A3000771B7, 0x80009F8280009F7E,
-    0x80009F8A80009F86, 0x000639A5000631A6, 0x000615AB80009F8E,
-    0x80009F92000615AA, 0x0006B5AF80009F96, 0x000605AD0006B5AE,
-    0x80009F9A000605AC, 0x80009FA280009F9E, 0x0000198080009FA6,
-    0x0000058200001981, 0x0001AC8480009FAA, 0x80009E5E0001ACBE,
-    0x0000D08680009FAE, 0x80009FB20001E487, 0x00018DA880009FB6,
-    0x000195AA000195A9, 0x0001CDAC0001DDAB, 0x0001FDAE0001F5AD,
-    0x0000A08400013DAF, 0x80009FBE80009FBA, 0x80009FC680009FC2,
-    0x80009FCE80009FCA, 0x0000C9B880009FD2, 0x0000D9BA0000C9B9,
-    0x0000F9BC0000D1BB, 0x000099BE0000F9BD, 0x000145B0000099BF,
-    0x000145B200014DB1, 0x000145B400015DB3, 0x000145B600014DB5,
-    0x00021DB30000F9B7, 0x80009FDA80009FD6, 0x80009FE280009FDE,
-    0x000245B5000275B6, 0x000289BB80009FE6, 0x80009FEA000281BA,
-    0x000289BF80009FEE, 0x000289BD000281BE, 0x001A5C85000291BC,
-    0x80009FF2000259A3, 0x000231A680009FF6, 0x80009FFE80009FFA,
-    0x0002C5AA000201A5, 0x8000A0020002CDAB, 0x0002C5AE8000A006,
-    0x0002D5AC0002CDAF, 0x001C00BE0002CDAD, 0x8000A00E8000A00A,
-    0x00000D80001C0084, 0x00003D8200003581, 0x8000A0168000A012,
-    0x8000A0228000A01A, 0x8000A02A8000A026, 0x8000A0328000A02E,
-    0x001CE0868000A036, 0x8000A03A00030487, 0x0002A8EF8000A03E,
-    0x8000A0420006C8EF, 0x000150E38000A046, 0x000180E1000614E3,
-    0x8000A04A00057CE1, 0x8000A0528000A04E, 0x000279A88000A056,
-    0x0002ADAA000279A9, 0x0002BDAC0002A5AB, 0x0002ADAE0002A5AD,
-    0x001CACBE0002A5AF, 0x8000A05E8000A05A, 0x8000A0668000A062,
-    0x8000A06E8000A06A, 0x000109B88000A072, 0x00011DBA000109B9,
-    0x000139BC000115BB, 0x0001DDBE000139BD, 0x0002DDB00001D5BF,
-    0x000161B2000161B1, 0x000161B4000161B3, 0x000151B6000169B5,
-    0x000774E1000151B7, 0x000608E30000F4E3, 0x8000A0760001BCE1,
-    0x8000A07E8000A07A, 0x8000A082001C1484, 0x8000A08A8000A086,
-    0x000000EC8000A092, 0x0006D4EF000028EF, 0x000061808000A096,
-    0x0000618200006181, 0x8000A09A000269B3, 0x000269B6000279B5,
-    0x001CC0868000A09E, 0x000255BA001CE887, 0x0001F9BC0001E9BB,
-    0x0001E9BE0001F9BD, 0x8000A01E0001E1BF, 0x8000A0A28000A08E,
-    0x8000A0A6001DA9A3, 0x001DA9A6001DB9A5, 0x8000A0AE8000A0AA,
-    0x001D95AA8000A0B2, 0x001E39AC001E29AB, 0x001E29AE001E39AD,
-    0x001EF1B3001E21AF, 0x8000A0BA8000A0B6, 0x8000A0C28000A0BE,
-    0x001EE1B5001EE9B6, 0x001EBDBB8000A0C6, 0x8000A0CA001EBDBA,
-    0x001E8DBF8000A0CE, 0x001E8DBD001E8DBE, 0x001E29A8001E8DBC,
-    0x001E39AA001E29A9, 0x001E29AC001E39AB, 0x001E51AE001E29AD,
-    0x8000A0D2001E51AF, 0x8000A0DA8000A0D6, 0x8000A0E28000A0DE,
-    0x8000A0EA8000A0E6, 0x001FEDB88000A0EE, 0x001FFDBA001FF5B9,
-    0x001FEDBC001FF5BB, 0x001F4DBE001F45BD, 0x001E39B0001F45BF,
-    0x001E19B2001E39B1, 0x001E09B4001E19B3, 0x001FDDB6001E09B5,
-    0x001FB1A3001FD5B7, 0x0000158100002D82, 0x8000A0F200001D80,
-    0x001FA1A5001FA9A6, 0x001FFDAB8000A0F6, 0x8000A0FA001FFDAA,
-    0x001FCDAF0001A0BE, 0x001FCDAD001FCDAE, 0x8000A0FE001FCDAC,
-    0x0000E886001E95B3, 0x000141B600012C87, 0x8000A1068000A102,
-    0x000145BA000155B5, 0x8000A10A000121BB, 0x000121BE8000A10E,
-    0x000131BC000129BF, 0x001E2DA8000131BD, 0x001E59AA001E39A9,
-    0x000199AC001E51AB, 0x000185AE000185AD, 0x0001E0840001BDAF,
-    0x8000A1168000A112, 0x8000A11E8000A11A, 0x8000A1268000A122,
-    0x000161B88000A12A, 0x000161BA000161B9, 0x000161BC000161BB,
-    0x000161BE000161BD, 0x0001C5B0000161BF, 0x0001A1B20001A1B1,
-    0x0001A1B40001A1B3, 0x000191B60001A9B5, 0x001DD1A3000191B7,
-    0x8000A1328000A12E, 0x8000A13A8000A136, 0x000211A5000205A6,
-    0x000265AB8000A13E, 0x8000A142000201AA, 0x00026DAF8000A146,
-    0x000275AD000265AE, 0x8000A14A000275AC, 0x8000A1528000A14E,
-    0x8000A15A8000A156, 0x8000A1628000A15E, 0x000039808000A166,
-    0x0000058200003981, 0x00033CBE8000A16A, 0x000344878000A172,
-    0x0004CC840004DC86, 0x8000A17A8000A176, 0x8000A17E0003E5B3,
-    0x000331B6000331B5, 0x8000A1868000A182, 0x00031DBA8000A18A,
-    0x00030DBC000315BB, 0x0000B1BE0000B1BD, 0x8000A18E0000B1BF,
-    0x8000A192000194E1, 0x8000A19600017CE3, 0x8000A19E8000A19A,
-    0x8000A1A68000A1A2, 0x8000A1AE8000A1AA, 0x8000A1B68000A1B2,
-    0x8000A1BE8000A1BA, 0x000255A8000274EF, 0x000261AA00025DA9,
-    0x0002B5AC000261AB, 0x0002B5AE0002BDAD, 0x000568BE0002ADAF,
-    0x8000A1C68000A1C2, 0x8000A1CE8000A1CA, 0x0000AD810000AD82,
-    0x000169B80000AD80, 0x000109BA000169B9, 0x000119BC000109BB,
-    0x000109BE000119BD, 0x0002D5B0000109BF, 0x0002D5B20002DDB1,
-    0x000179B4000169B3, 0x000169B6000179B5, 0x001E78E1000161B7,
-    0x001E04E3001F5CE1, 0x8000A1D2001F10E3, 0x8000A1DA8000A1D6,
-    0x8000A1E68000A1E2, 0x8000A1EE8000A1EA, 0x0004F8BE8000A1F2,
-    0x00006CEF8000A1F6, 0x0002A9A3001E64EF, 0x000448868000A1FA,
-    0x8000A1FE00050487, 0x00027DA500027DA6, 0x000259AB8000A202,
-    0x8000A206000251AA, 0x0001FDAF8000A20A, 0x0001FDAD0001FDAE,
-    0x000675A8000241AC, 0x000681AA000681A9, 0x00068DAC00069DAB,
-    0x0006BDAE0006B5AD, 0x8000A1DE0006DDAF, 0x8000A2128000A20E,
-    0x8000A21A8000A216, 0x8000A2228000A21E, 0x0006ADB88000A226,
-    0x0006BDBA0006B5B9, 0x000775BC0006B5BB, 0x000775BE00077DBD,
-    0x0006A9B000076DBF, 0x0006BDB20006B5B1, 0x0006ADB40006B5B3,
-    0x00069DB6000695B5, 0x000609B3000695B7, 0x8000A22E8000A22A,
-    0x8000A2368000A232, 0x000605B5000635B6, 0x000671BB8000A23A,
-    0x8000A23E00066DBA, 0x000651BF8000A242, 0x000661BD000669BE,
-    0x00001982000669BC, 0x0000658000064DA3, 0x000671A600001981,
-    0x8000A24A8000A246, 0x000629AA000641A5, 0x0003A484000635AB,
-    0x00062DAE0001E4BF, 0x00062DAC000615AF, 0x00016C85000625AD,
-    0x000008848000A252, 0x000C00868000A16E, 0x8000A25600006087,
-    0x0006BDA88000A25A, 0x00068DAA000685A9, 0x00069DAC000685AB,
-    0x0006FDAE000681AD, 0x00069DB00006EDAF, 0x0006A5B20006ADB1,
-    0x000179B4000169B3, 0x000169B6000179B5, 0x000159B8000169B7,
-    0x0001EDBA000159B9, 0x0001FDBC0001E5BB, 0x0001E5BE0001E5BD,
-    0x000625B30001D5BF, 0x8000A24E000F08BE, 0x8000A2628000A25E,
-    0x000635B5000621B6, 0x000649BB8000A266, 0x8000A26A000679BA,
-    0x00019DBF8000A26E, 0x00019DBD00019DBE, 0x8000A272000651BC,
-    0x8000A276000661A3, 0x000665A68000A27A, 0x8000A2828000A27E,
-    0x00063DAA000671A5, 0x8000A28600060DAB, 0x0001D9AE8000A28A,
-    0x000615AC0001D9AF, 0x000114E10001D9AD, 0x0036D0E300011CE3,
-    0x000C9084000104E1, 0x8000A2928000A28E, 0x006A218F8000A296,
-    0x8000A29E8000A29A, 0x8000A2A68000A2A2, 0x0037FCEF000148EF,
-    0x0012CD978000A2AA, 0x0013099500126596, 0x001671930016E194,
-    0x006AF19100171992, 0x0006619F006A5990, 0x001A619D001B099E,
-    0x001B199B001A699C, 0x001E5999001E519A, 0x000610E3001F0998,
-    0x0006ECE18000A2AE, 0x0002B1A38000A2B2, 0x0007D9A1000399A2,
-    0x003A9DB3000699A0, 0x003801B1003A85B2, 0x000768EF003E01B0,
-    0x0026BDB58000A2B6, 0x003001AB002401B4, 0x003645A9003619AA,
-    0x003EF1AF00364DA8, 0x0033C5AD003C01AE, 0x000E65A30032FDAC,
-    0x000361A1000E65A2, 0x000BE1A78000A2BA, 0x000A79A5000A71A6,
-    0x006B0182000F89A4, 0x000D0886006B0183, 0x00173186000D4C87,
-    0x006BB1840017B987, 0x0012618A00165185, 0x8000A2C20013058B,
-    0x001F118E8000A2C6, 0x0013B58C001F118F, 0x001A2992001E498D,
-    0x000CC4BE001B4193, 0x00075196000F8084, 0x001B519400075197,
-    0x0006219A001A3195, 0x8000A2CA0002699B, 0x0001C5818000A2CE,
-    0x0003599C0001F980, 0x000F48BE00019182, 0x8000A2D68000A2D2,
-    0x8000A2DE8000A2DA, 0x8000A2E68000A2E2, 0x8000A2EA0006D4EF,
-    0x8000A2EE000674E1, 0x8000A2F2000154E3, 0x8000A2FA8000A2F6,
-    0x8000A3028000A2FE, 0x8000A30A8000A306, 0x8000A30E0002D1B3,
-    0x000245B6000245B5, 0x8000A3168000A312, 0x000229BA8000A31A,
-    0x00023DBC000225BB, 0x000221BE000221BD, 0x000D1DA3000215BF,
-    0x8000A3228000A31E, 0x8000A32A8000A326, 0x000D89A5000D89A6,
-    0x000DE9AB8000A332, 0x00030886000DE5AA, 0x000DD9AF00034C87,
-    0x000DEDAD000DEDAE, 0x00004980000DF1AC, 0x0000598200004981,
-    0x8000A336000F25B3, 0x000FD1B6000FD1B5, 0x8000A33E8000A33A,
-    0x000FFDBA8000A342, 0x000FD1BC000FF5BB, 0x000FC9BE000FDDBD,
-    0x8000A2BE000FC9BF, 0x8000A3468000A32E, 0x8000A34E8000A34A,
-    0x8000A3568000A352, 0x000F89A88000A35A, 0x000FC1AA000F89A9,
-    0x000F35AC000FC1AB, 0x000F35AE000F3DAD, 0x000F55B0000F2DAF,
-    0x000F69B2000F5DB1, 0x000F61B4000F65B3, 0x000F61B6000F61B5,
-    0x000F21B8000F61B7, 0x000F21BA000F21B9, 0x000F2DBC000F3DBB,
-    0x000F1DBE000F15BD, 0x000E69A30000F9BF, 0x8000A3628000A35E,
-    0x00009C848000A366, 0x000E9DA5000E9DA6, 0x000EB9AB8000A36A,
-    0x8000A36E000EB1AA, 0x000E85AF8000A372, 0x000E91AD000E85AE,
-    0x8000A376000E9DAC, 0x8000A37E8000A37A, 0x8000A3868000A382,
-    0x8000A38E8000A38A, 0x00003D808000A392, 0x0000198200000981,
-    0x8000A39A8000A396, 0x0001D8878000A39E, 0x00019DA800040086,
-    0x0001D5AA0001D5A9, 0x0001F5AC0001EDAB, 0x000115AE00011DAD,
-    0x8000A3A200010DAF, 0x8000A3AA8000A3A6, 0x8000A3B28000A3AE,
-    0x8000A3BA8000A3B6, 0x0000EDB88000A3BE, 0x00008DBA000085B9,
-    0x00009DBC000085BB, 0x00008DBE000085BD, 0x000175B0000085BF,
-    0x000175B200017DB1, 0x0000F9B40000E9B3, 0x0000E5B60000F9B5,
-    0x00025DB30000D5B7, 0x8000A3C68000A3C2, 0x8000A3CE8000A3CA,
-    0x00027DB500029DB6, 0x0002B9BB8000A3D2, 0x8000A3D60002B1BA,
-    0x00033DBF8000A3DA, 0x00033DBD00033DBE, 0x0004CCBE0002A1BC,
-    0x8000A3DE000219A3, 0x0002D9A68000A3E2, 0x8000A3EA8000A3E6,
-    0x0002F5AA000239A5, 0x8000A3EE0002FDAB, 0x000379AE8000A3F2,
-    0x0002E5AC000379AF, 0x00050CBE000379AD, 0x8000A3FA8000A3F6,
-    0x8000A3FE000610BE, 0x8000A4068000A402, 0x000035808000A40A,
-    0x0000358200003D81, 0x0004B0868000A412, 0x8000A41600033487,
-    0x8000A41E8000A41A, 0x000270EF8000A422, 0x8000A42A8000A426,
-    0x8000A4328000A42E, 0x8000A43A8000A436, 0x8000A4428000A43E,
-    0x000194E18000A446, 0x0001F8E38000A44A, 0x000130E18000A44E,
-    0x000E54E3000E14E1, 0x8000A452000F0CE3, 0x8000A45A8000A456,
-    0x0005A8848000A45E, 0x8000A4668000A462, 0x8000A46E8000A46A,
-    0x000E40EF8000A472, 0x0002E9B3000E8CEF, 0x8000A47A8000A476,
-    0x8000A4868000A482, 0x0002E9B50002E1B6, 0x000169BB8000A48A,
-    0x8000A48E000161BA, 0x000159BF8000A492, 0x000169BD000169BE,
-    0x0000B580000171BC, 0x0000C1820000C181, 0x0004B08600040484,
-    0x8000A49600050887, 0x000E31AC8000A49A, 0x000E01AE000E39AD,
-    0x000EB5A8000E01AF, 0x000EA5AA000EADA9, 0x000E19B4000E31AB,
-    0x000E25B6000E2DB5, 0x000E11B0000DD9B7, 0x000E15B2000E1DB1,
-    0x000DCDBC000E09B3, 0x000DD5BE000DD5BD, 0x000DE9B8000D69BF,
-    0x000DDDBA000DE9B9, 0x0005A9A3000DD5BB, 0x8000A4A28000A49E,
-    0x8000A4AA8000A4A6, 0x0005A9A50005A1A6, 0x000629AB8000A4AE,
-    0x8000A4B2000621AA, 0x000619AF8000A4B6, 0x000629AD000629AE,
-    0x8000A4BA000631AC, 0x8000A4BE0006F9B3, 0x00069DB68000A4C2,
-    0x8000A4CA8000A4C6, 0x0006B9BA000699B5, 0x8000A4CE0006B9BB,
-    0x000699BE8000A4D2, 0x000699BC000699BF, 0x8000A40E000699BD,
-    0x8000A4D68000A47E, 0x8000A4DE8000A4DA, 0x0001918100019982,
-    0x00060DA800019980, 0x000615AA00061DA9, 0x000675AC00066DAB,
-    0x000669AE000679AD, 0x000619B0000669AF, 0x000629B2000619B1,
-    0x000639B4000629B3, 0x00062DB6000639B5, 0x0007F1B8000621B7,
-    0x0007F1BA0007F1B9, 0x0007D5BC0007F1BB, 0x0007D5BE0007DDBD,
-    0x0007BDA30007CDBF, 0x8000A4E68000A4E2, 0x8000A4EE8000A4EA,
-    0x0007DDA50007D9A6, 0x0007FDAB8000A4F2, 0x003C00860007FDAA,
-    0x0007DDAF00017C87, 0x0007DDAD0007DDAE, 0x8000A4F60007DDAC,
-    0x8000A4FA000619B3, 0x0006C9B68000A4FE, 0x8000A5068000A502,
-    0x000169BA000631B5, 0x8000A50A000169BB, 0x000159BE8000A50E,
-    0x000169BC000159BF, 0x0006CDA8000161BD, 0x0006D5AA0006D5A9,
-    0x0006FDAC0006E5AB, 0x000605AE00060DAD, 0x8000A512000199AF,
-    0x8000A51A8000A516, 0x8000A5228000A51E, 0x8000A52A8000A526,
-    0x00018DB88000A52E, 0x000191BA000191B9, 0x0001BDBC0001A5BB,
-    0x000155BE00015DBD, 0x0001E9B000014DBF, 0x0001F9B20001E9B1,
-    0x0001D9B40001F1B3, 0x0001B9B60001D9B5, 0x0003CCBE0001B5B7,
-    0x000565A500059DA6, 0x00054DA38000A532, 0x8000A53A8000A536,
-    0x00020DAF8000A53E, 0x000235AD00020DAE, 0x00023DAB00023DAC,
-    0x8000A54200023DAA, 0x8000A54A8000A546, 0x8000A5528000A54E,
-    0x000019808000A556, 0x0000058200001981, 0x000260848000A55A,
-    0x000348878000A562, 0x8000A566003CCC86, 0x8000A56E8000A56A,
-    0x8000A5768000A572, 0x003CC0848000A57A, 0x8000A57E000000A6,
-    0x8000A5820001A4E1, 0x8000A5860002ECE3, 0x8000A58E8000A58A,
-    0x8000A592000254EF, 0x8000A59A8000A596, 0x8000A5A28000A59E,
-    0x8000A5AA8000A5A6, 0x8000A5AE000385B3, 0x000381B6000381B5,
-    0x8000A5B68000A5B2, 0x000349BA8000A5BA, 0x000359BC000349BB,
-    0x000349BE000359BD, 0x000688E1000349BF, 0x000694E30007ACE1,
-    0x8000A5BE00063CE3, 0x8000A5C68000A5C2, 0x000035808000A5CA,
-    0x0000358200003D81, 0x003C9086003D60BE, 0x00074CEF003DF887,
-    0x000209A3000744EF, 0x8000A5D68000A5D2, 0x8000A5DE8000A5DA,
-    0x00020DA500020DA6, 0x0002C5AB8000A5E2, 0x8000A5E60002C5AA,
-    0x0002C5AF8000A5EA, 0x0002D5AD0002C5AE, 0x000511BF0002D5AC,
-    0x00050DBD00050DBE, 0x0005F9BB0005E1BC, 0x0005CDB90005CDBA,
-    0x0005E9B70005D1B8, 0x000581B5000589B6, 0x0005A1B30005B1B4,
-    0x0005B1B10005B9B2, 0x00067DAF0005B9B0, 0x00067DAD00067DAE,
-    0x00067DAB00067DAC, 0x00067DA900067DAA, 0x8000A5EE00067DA8,
-    0x8000A5F68000A5F2, 0x8000A5FE8000A5FA, 0x8000A6068000A602,
-    0x003E65A88000A60A, 0x003E7DAA003E75A9, 0x003E6DAC003E75AB,
-    0x003E9DAE003E95AD, 0x8000A60E003E8DAF, 0x8000A6168000A612,
-    0x8000A61E8000A61A, 0x8000A6268000A622, 0x003E81B88000A62A,
-    0x003E81BA003E81B9, 0x003E8DBC003E9DBB, 0x003EBDBE003EB5BD,
-    0x003EF9B0003F4DBF, 0x003ECDB2003EC5B1, 0x003EC1B4003EC5B3,
-    0x003EC1B6003EC1B5, 0x8000A55E003EC1B7, 0x8000A62E003EF1B3,
-    0x003E11B68000A5CE, 0x8000A6368000A632, 0x003E39BA003EE1B5,
-    0x8000A63A003E05BB, 0x003E01BE0003E084, 0x003E1DBC003E01BF,
-    0x003EB5A3003E15BD, 0x0000518100006D82, 0x000398BE00005980,
-    0x003EA5A5003E55A6, 0x003E41AB8000A63E, 0x00004886003E7DAA,
-    0x003E45AF00006887, 0x003E51AD003E45AE, 0x003E45A8003E59AC,
-    0x003E45AA003E4DA9, 0x003E4DAC003E5DAB, 0x003E75AE003E7DAD,
-    0x8000A642003EFDAF, 0x00016C848000A646, 0x8000A64E8000A64A,
-    0x8000A6568000A652, 0x000165B88000A65A, 0x000165BA00016DB9,
-    0x000165BC00017DBB, 0x000165BE00016DBD, 0x003E85B00001D5BF,
-    0x003E85B2003E8DB1, 0x003E85B4003E9DB3, 0x003E85B6003E8DB5,
-    0x003ECDB300015DB7, 0x8000A6628000A65E, 0x8000A66A8000A666,
-    0x003EC5B5003EFDB6, 0x003E49BB8000A66E, 0x8000A672003E41BA,
-    0x003E49BF8000A676, 0x003E49BD003E41BE, 0x8000A67A003E51BC,
-    0x8000A67E003E89A3, 0x003EB9A68000A682, 0x8000A68A8000A686,
-    0x003E05AA003E81A5, 0x8000A68E003E0DAB, 0x003E05AE8000A692,
-    0x003E15AC003E0DAF, 0x000321A8003E0DAD, 0x000325AA000321A9,
-    0x000325AC00033DAB, 0x000325AE00032DAD, 0x8000A696000399AF,
-    0x000264BE8000A69A, 0x8000A6A28000A69E, 0x8000A6AA8000A6A6,
-    0x00038DB88000A6B2, 0x000391BA000391B9, 0x0003BDBC0003A5BB,
-    0x00007DBE000075BD, 0x0003E9B0000071BF, 0x0003F9B20003E9B1,
-    0x0003D9B40003F1B3, 0x0003B9B60003D9B5, 0x0000B5810003B5B7,
-    0x0003A5B30000B580, 0x0003A5B50000B582, 0x8000A6BA8000A6B6,
-    0x000400860003A5B6, 0x000325BB0003D487, 0x00032DBD000325BA,
-    0x000315BF000325BC, 0x0003EDA3000315BE, 0x8000A6C28000A6BE,
-    0x8000A6C600116085, 0x0003EDA50003EDA6, 0x00036DAB8000A6CA,
-    0x8000A6CE00036DAA, 0x00035DAF8000A6D2, 0x000365AD00035DAE,
-    0x0001BCE300036DAC, 0x003E20E1003F70E1, 0x8000A6D6003F94E3,
-    0x8000A6DE8000A6DA, 0x8000A6E68000A6E2, 0x8000A6EE8000A6EA,
-    0x000118EF8000A6F2, 0x001338BF8000A6F6, 0x8000A6FA003EE8EF,
-    0x000504848000A6FE, 0x000068E3000504BE, 0x0001A8E18000A702,
-    0x0003B0EF8000A706, 0x8000A70E8000A70A, 0x00001D808000A712,
-    0x0000DD8200000981, 0x003EC9AA8000A716, 0x003E95A8003EC9AB,
-    0x003EF9AE003ECDA9, 0x003ECDAC003EF9AF, 0x00EFA8BF003EC5AD,
-    0x000488868000A71A, 0x8000A7220004AC87, 0x8000A72A8000A726,
-    0x003DD9BA8000A72E, 0x003DC5B8003DD9BB, 0x003D91BE003DD9B9,
-    0x003DE1BC003D91BF, 0x003E49B2003DE1BD, 0x003E49B0003E49B3,
-    0x003E49B6003E49B1, 0x003E49B4003E49B7, 0x00063DA8003E49B5,
-    0x000691AA000691A9, 0x0006D9AC0006C9AB, 0x0006C5AE0006C5AD,
-    0x8000A71E0006F5AF, 0x8000A7368000A732, 0x8000A73E8000A73A,
-    0x8000A7468000A742, 0x00076DB88000A74A, 0x000709BA000771B9,
-    0x000705BC000709BB, 0x00073DBE000709BD, 0x00068DB0000735BF,
-    0x00077DB2000775B1, 0x00076DB4000775B3, 0x00075DB6000755B5,
-    0x000619B3000755B7, 0x8000A7528000A74E, 0x8000A75A8000A756,
-    0x000651B5000645B6, 0x0007B1BB8000A75E, 0x8000A762000649BA,
-    0x000791BF8000A766, 0x0007A1BD0007A9BE, 0x8000A76A0007A9BC,
-    0x8000A76E00065DA3, 0x000601A68000A772, 0x8000A77A8000A776,
-    0x00060DAA000615A5, 0x8000A77E0007F5AB, 0x0007EDAE8000A782,
-    0x0007EDAC0007D5AF, 0x0001ED800007E5AD, 0x0000198200000981,
-    0x8000A78A8000A786, 0x0001C4BE0001D084, 0x000080868000A78E,
-    0x8000A79200012087, 0x8000A79A8000A796, 0x8000A7A28000A79E,
-    0x8000A7AA8000A7A6, 0x8000A7AE0005FCE1, 0x8000A7B2000414E3,
-    0x8000A7BA8000A7B6, 0x8000A7C28000A7BE, 0x8000A7CA8000A7C6,
-    0x8000A7D28000A7CE, 0x8000A7DA8000A7D6, 0x000631A80004C8EF,
-    0x000631AA000631A9, 0x0001B5AC000631AB, 0x0001B5AE0001BDAD,
-    0x0001A8840001ADAF, 0x8000A7E28000A7DE, 0x8000A7EA8000A7E6,
-    0x8000A7F28000A7EE, 0x000169B88000A7F6, 0x000109BA000169B9,
-    0x00011DBC000105BB, 0x000101BE000101BD, 0x0001D5B0000135BF,
-    0x0001D5B20001DDB1, 0x000179B4000169B3, 0x000169B6000179B5,
-    0x0003BD81000161B7, 0x0005B1B30003BD80, 0x000235B500001982,
-    0x000354BE8000A7FA, 0x8000A802000235B6, 0x000211BB8000A806,
-    0x0002B5BD000219BA, 0x0002B5BF000209BC, 0x8000A80A0002B5BE,
-    0x00026DA500026DA6, 0x0005E9A3000C0084, 0x000CF8868000A80E,
-    0x0002EDAF00037C87, 0x0002EDAD0002EDAE, 0x000249AB000251AC,
-    0x8000A812000241AA, 0x8000A81A8000A816, 0x8000A8228000A81E,
-    0x000120E38000A826, 0x0001A0E18000A82A, 0x000240EF8000A82E,
-    0x8000A8368000A832, 0x8000A83E8000A83A, 0x8000A8468000A842,
-    0x8000A6AE8000A84A, 0x8000A84E0003F5B3, 0x0003EDB68000A852,
-    0x8000A85A8000A856, 0x0003B1BA0003E5B5, 0x8000A85E0003B1BB,
-    0x000391BE8000A862, 0x0003A1BC000391BF, 0x000229A80003A1BD,
-    0x000239AA000229A9, 0x000229AC000239AB, 0x000259AE000229AD,
-    0x0001CD80000255AF, 0x0000198200000981, 0x8000A86A8000A866,
-    0x000DB487000D7CBE, 0x000231B8000C1C86, 0x000231BA000231B9,
-    0x0002F9BC0002E9BB, 0x0002E9BE0002F9BD, 0x000231B00002E9BF,
-    0x000231B2000231B1, 0x000211B4000231B3, 0x000211B6000211B5,
-    0x8000A872000211B7, 0x8000A87A8000A876, 0x8000A8828000A87E,
-    0x8000A88A8000A886, 0x8000A88E00007CEF, 0x8000A892000614E1,
-    0x000000EC0000A8E3, 0x8000A89A8000A896, 0x0002F1A38000A89E,
-    0x8000A8A68000A8A2, 0x8000A8AE8000A8AA, 0x0002E1A50002E9A6,
-    0x0002B5AB8000A8B2, 0x8000A8B60002B5AA, 0x000295AF8000A8BA,
-    0x0002A5AD000295AE, 0x000E21A80002A5AC, 0x000E49AA000E21A9,
-    0x000E69AC000E59AB, 0x000E99AE000E69AD, 0x8000A86E000E99AF,
-    0x8000A8C28000A8BE, 0x8000A8CA8000A8C6, 0x8000A8D28000A8CE,
-    0x000EF5B88000A8D6, 0x000EF5BA000EFDB9, 0x000E9DBC000E85BB,
-    0x000E81BE000E81BD, 0x000EE9B0000EB5BF, 0x000EFDB2000EE9B1,
-    0x000ED5B4000EF1B3, 0x000ED5B6000EDDB5, 0x000EC1B3000ECDB7,
-    0x0000B58100001582, 0x8000A8DA0000B580, 0x000EE1B5000EE9B6,
-    0x000E2DBB000010BE, 0x00036484000E2DBA, 0x000E19BF8000A8DE,
-    0x000E25BD000E11BE, 0x8000A8E2000E29BC, 0x00008886000E85A3,
-    0x000EADA600036C87, 0x8000A8EA8000A8E6, 0x000E69AA000EA5A5,
-    0x8000A8EE000E69AB, 0x000E55AE8000A8F2, 0x000E6DAC000E5DAF,
-    0x000E89B3000E61AD, 0x8000A8FA8000A8F6, 0x8000A9028000A8FE,
-    0x000E89B5000E81B6, 0x000E95BB8000A906, 0x8000A90A000E8DBA,
-    0x000EF9BF8000A90E, 0x000E85BD000EF1BE, 0x8000A912000E85BC,
-    0x8000A91A8000A916, 0x000D10E18000A91E, 0x000D04E38000A922,
-    0x8000A92A8000A926, 0x000CA4EF8000A92E, 0x8000A9368000A932,
-    0x8000A93E8000A93A, 0x000E01A88000A942, 0x000E01AA000E01A9,
-    0x000E01AC000E01AB, 0x000E01AE000E01AD, 0x0000DD80000E3DAF,
-    0x0000198200000981, 0x8000A94A8000A946, 0x000074BE00013C84,
-    0x000E1DB88000A952, 0x000E25BA000E2DB9, 0x0001F9BC0001E9BB,
-    0x0001E9BE0001F9BD, 0x000E49B00001E9BF, 0x000E51B2000E51B1,
-    0x000E35B4000E51B3, 0x000E35B6000E3DB5, 0x000D89A3000E2DB7,
-    0x0004A8868000A956, 0x8000A95A00033C87, 0x000D89A5000D81A6,
-    0x000D95AB8000A95E, 0x8000A962000D8DAA, 0x000DF9AF8000A966,
-    0x000D85AD000DF1AE, 0x8000A96A000D85AC, 0x0003488400029DB3,
-    0x000355B6000440BE, 0x8000A9728000A96E, 0x000371BA0002B1B5,
-    0x8000A976000379BB, 0x000331BE8000A97A, 0x000351BC000331BF,
-    0x000391B0000351BD, 0x0003A1B2000399B1, 0x0003D1B40003A1B3,
-    0x0003D1B60003D1B5, 0x0003F5B80003D1B7, 0x0003C9BA0003F9B9,
-    0x0003DDBC0003C9BB, 0x0003CDBE0003C5BD, 0x8000A97E0003C5BF,
-    0x8000A9868000A982, 0x001AEC858000A98A, 0x8000A98E8000A7FE,
-    0x000221A88000A992, 0x000231AA000231A9, 0x00021DAC000205AB,
-    0x0003F1AE0003F1AD, 0x8000A9960003F1AF, 0x8000A99E8000A99A,
-    0x000050EF8000A9A2, 0x8000A9AA8000A9A6, 0x000084E38000A9AE,
-    0x0001F8E18000A9B2, 0x000015808000A9B6, 0x0000058200001981,
-    0x000399A38000A9BA, 0x000468868000A9C2, 0x8000A9C600056087,
-    0x0003B5A5000251A6, 0x00027DAB8000A9CA, 0x8000A9CE000275AA,
-    0x000235AF8000A9D2, 0x000255AD000235AE, 0x8000A9D6000255AC,
-    0x8000A9DE8000A9DA, 0x8000A9E68000A9E2, 0x8000A9EE8000A9EA,
-    0x00046CBE0001F8EF, 0x8000A9F2000E80E1, 0x8000A9F6000114E3,
-    0x8000A9FE8000A9FA, 0x8000AA068000AA02, 0x8000AA0E8000AA0A,
-    0x8000AA120001DDB3, 0x0001F5B60001FDB5, 0x8000AA1A8000AA16,
-    0x0001B1BA8000AA1E, 0x00019DBC000185BB, 0x00013DBE000135BD,
-    0x0005BDA8000135BF, 0x000619AA000601A9, 0x000639AC000611AB,
-    0x000659AE000639AD, 0x8000A9BE000649AF, 0x00001D8100001D82,
-    0x8000AA2200001D80, 0x8000AA2A8000AA26, 0x00069DB88000AA2E,
-    0x00074DBA000745B9, 0x00075DBC000745BB, 0x000741BE000741BD,
-    0x000639B0000741BF, 0x000605B200060DB1, 0x0006B5B40006ADB3,
-    0x0006B5B60006BDB5, 0x000691A30006A5B7, 0x0000288600028C84,
-    0x8000AA320003C087, 0x0006B1A50006B9A6, 0x0006C9AB8000AA36,
-    0x8000AA3A0006FDAA, 0x000679AF8000AA3E, 0x000679AD000671AE,
-    0x8000AA420006D1AC, 0x8000AA46000799B3, 0x000751B68000AA4A,
-    0x8000AA528000AA4E, 0x00076DBA0007B1B5, 0x8000AA56000775BB,
-    0x000759BE8000AA5A, 0x00076DBC000745BF, 0x8000AA5E000765BD,
-    0x8000AA668000AA62, 0x8000AA6E8000AA6A, 0x8000AA720005A0EF,
-    0x8000AA7A8000AA76, 0x8000AA7E00057CE3, 0x8000AA8200053CE1,
-    0x8000AA8A8000AA86, 0x00061DA38000AA8E, 0x0000158100002D82,
-    0x8000AA9200001D80, 0x000635A50006D5A6, 0x0006F1AB8000AA96,
-    0x8000AA9A0006E9AA, 0x0006C1AF00012084, 0x0006E1AD0006DDAE,
-    0x0006C5A80006E9AC, 0x0006D5AA0006D5A9, 0x0006FDAC0006E5AB,
-    0x00061DAE000615AD, 0x0001E8BE000615AF, 0x000008868000AAA2,
-    0x8000AAA600008887, 0x8000AAAE8000AAAA, 0x00067DB88000AAB2,
-    0x00060DBA000605B9, 0x000601BC000605BB, 0x000639BE000609BD,
-    0x00066DB0000639BF, 0x00067DB2000675B1, 0x000659B4000675B3,
-    0x00064DB6000645B5, 0x000291A8000645B7, 0x0002A1AA000299A9,
-    0x0002D1AC0002A1AB, 0x0002D5AE0002DDAD, 0x8000AAB60002C9AF,
-    0x8000AABE8000AABA, 0x8000AAC200C524BF, 0x8000AACA8000AAC6,
-    0x000359B88000AACE, 0x000369BA000359B9, 0x000361BC000365BB,
-    0x000361BE000361BD, 0x0002B9B0000361BF, 0x000285B200028DB1,
-    0x000375B400036DB3, 0x000375B600037DB5, 0x000229B300036DB7,
-    0x8000AAD68000AAD2, 0x000384848000AADA, 0x000259B500025DB6,
-    0x000249BB8000AADE, 0x8000A94E00027DBA, 0x0003B9BF8000AAE2,
-    0x0003B9BD0003B1BE, 0x8000AAE6000251BC, 0x8000AAEA00026DA3,
-    0x000219A68000AAEE, 0x8000AAF68000AAF2, 0x000239AA00021DA5,
-    0x8000AAFA00020DAB, 0x0003F5AE8000AAFE, 0x000215AC0003FDAF,
-    0x0002A5A80003FDAD, 0x0002BDAA0002B5A9, 0x0002ADAC0002B5AB,
-    0x00013DAE000135AD, 0x00001181000135AF, 0x8000AB0200000D80,
-    0x8000AB0600001982, 0x001DB4BE8000AB0A, 0x0001EDB88000AB12,
-    0x000181BA000185B9, 0x000185BC000181BB, 0x0001B1BE00018DBD,
-    0x000151B00001B1BF, 0x000151B2000151B1, 0x0001F5B4000151B3,
-    0x0001F5B60001FDB5, 0x0000D8EF0001E1B7, 0x001DA8868000AB16,
-    0x8000AB1A001DDC87, 0x8000AB228000AB1E, 0x8000AB260006C4EF,
-    0x8000AB2A0006F4E1, 0x0000CCE3000138E3, 0x000128E18000AB2E,
-    0x00027DB58000AB32, 0x8000AB3A8000AB36, 0x8000AB3E0001EDB6,
-    0x00026DB38000AB42, 0x0001DDBD8000AB46, 0x000161BF0001DDBC,
-    0x8000AB4A000169BE, 0x0001CDBB8000AB4E, 0x001DB5A30001CDBA,
-    0x8000AB528000AB0E, 0x8000AB5A8000AB56, 0x001DA5A5001E35A6,
-    0x001E15AB8000AB5E, 0x8000AB62001E15AA, 0x001EB9AF8000AB66,
-    0x001E05AD001EB1AE, 0x8000AB6A001E05AC, 0x8000AB728000AB6E,
-    0x000019808000AB76, 0x0000058200001981, 0x0003EC848000AB7A,
-    0x8000AB7E000368BE, 0x0001D0868000AB82, 0x8000AB860000B487,
-    0x001FBDA88000AB8A, 0x001F19AA001F01A9, 0x001F09AC001F19AB,
-    0x001F31AE001F35AD, 0x8000AB8E001F29AF, 0x8000AB968000AB92,
-    0x8000AB9E8000AB9A, 0x8000ABA68000ABA2, 0x0000C9B88000ABAA,
-    0x0000D9BA0000C9B9, 0x0000F9BC0000D1BB, 0x000099BE0000F9BD,
-    0x001F5DB0000099BF, 0x001F21B2001F21B1, 0x001F25B4001F3DB3,
-    0x001F09B6001F29B5, 0x001CC0EF001F09B7, 0x8000ABAE0019C0E1,
-    0x8000ABB20009D0E3, 0x8000ABBA8000ABB6, 0x000125A38000ABBE,
-    0x001D15A1000399A2, 0x001DB0E3001D0DA0, 0x001E88E18000ABC2,
-    0x001ED1B3001060EF, 0x8000ABCA8000ABC6, 0x8000ABD28000ABCE,
-    0x001EC1B5001EF5B6, 0x001EA5BB8000ABD6, 0x8000ABDA001ED9BA,
-    0x001EADBF8000ABDE, 0x001EA9BD001EA1BE, 0x001E11A3001EB5BC,
-    0x8000ABE68000ABE2, 0x0024198F8000ABEA, 0x001E01A5001E35A6,
-    0x001E65AB000710BE, 0x00040084001E19AA, 0x001E6DAF8000ABF2,
-    0x001E69AD001E61AE, 0x0014DD9B001E75AC, 0x001709990015E59A,
-    0x001F799F00111198, 0x0019F99D0019799E, 0x002D6193001B099C,
-    0x00296D9100289192, 0x00117D9700296D90, 0x002CE99500117596,
-    0x00042086002D7994, 0x0027818200039487, 0x002A9584002A8D83,
-    0x002E9986002B6185, 0x0004A884002E9987, 0x002F398A8000ABF6,
-    0x0012BD8C0012AD8B, 0x0014F98E0012E98D, 0x8000ABFA00161D8F,
-    0x001755928000ABFE, 0x001A719400189993, 0x001CF196001AF995,
-    0x8000AC02001E7D97, 0x001E559A8000AC06, 0x0002759C0000D19B,
-    0x8000AC0E8000AC0A, 0x00F685AF8000AC12, 0x00F2F1AD00F401AE,
-    0x00F001AB00F289AC, 0x000E9DA9000E85AA, 0x000A75A7000E9DA8,
-    0x000801A5000A85A6, 0x0006B9A3000601A4, 0x0002D1A10006B1A2,
-    0x000345808000AA9E, 0x0003458200034D81, 0x8000AC1A8000AC16,
-    0x8000AC228000AC1E, 0x8000AC2A8000AC26, 0x00FC01B58000AC2E,
-    0x00FAF5B300FBC1B4, 0x00F7BDB100FAF5B2, 0x00036DA800F6F5B0,
-    0x000375AA00037DA9, 0x0003B5AC0003ADAB, 0x0003B5AE0003BDAD,
-    0x8000AC320003ADAF, 0x8000AC3A8000AC36, 0x8000AC428000AC3E,
-    0x8000AC4A8000AC46, 0x000275B88000AC4E, 0x000201BA000201B9,
-    0x000201BC000201BB, 0x000231BE000209BD, 0x0003D5B0000231BF,
-    0x0003D5B20003DDB1, 0x000261B4000265B3, 0x000261B6000261B5,
-    0x001E10E1000261B7, 0x0001F8E38000AC52, 0x00001580000400BE,
-    0x0000398200000981, 0x8000AC5A8000AC56, 0x000488868000AC62,
-    0x8000AC6600050C87, 0x000124EF8000AC6A, 0x8000AC728000AC6E,
-    0x8000AC7A8000AC76, 0x8000AC828000AC7E, 0x8000AC860000B0EF,
-    0x8000AC8E8000AC8A, 0x8000AC9200005CE3, 0x8000AC96000184E1,
-    0x8000AC9E8000AC9A, 0x00047DA38000ACA2, 0x8000AC5E8000ABEE,
-    0x8000ACAA8000ACA6, 0x0004BDA50004BDA6, 0x000499AB8000ACAE,
-    0x8000ACB2000491AA, 0x000525AF8000ACB6, 0x000531AD000525AE,
-    0x8000ACBA000481AC, 0x8000ACBE000605B3, 0x000635B68000ACC2,
-    0x8000ACCA8000ACC6, 0x00061DBA00063DB5, 0x0003E0840006A9BB,
-    0x0006A9BE0000E0BE, 0x0006A9BC0006A9BF, 0x00052DA80006A9BD,
-    0x0006C9AA0006A1A9, 0x0006F1AC0006C9AB, 0x0006EDAE0006F1AD,
-    0x0000ED800006E5AF, 0x0000F1820000F181, 0x000090868000ACCE,
-    0x8000ACD20000F087, 0x0007E1B88000ACD6, 0x0007E1BA0007E1B9,
-    0x0007E1BC0007E1BB, 0x0007E1BE0007E1BD, 0x00069DB00007E1BF,
-    0x000639B2000629B1, 0x000629B4000639B3, 0x000619B6000629B5,
-    0x000741A3000619B7, 0x8000ACDE8000ACDA, 0x8000ACE68000ACE2,
-    0x000779A5000771A6, 0x0007EDAB8000ACEA, 0x8000ACEE000759AA,
-    0x0007EDAF8000ACF2, 0x0007EDAD0007EDAE, 0x8000ACF60007EDAC,
-    0x8000ACFA0006F1B3, 0x000695B68000ACFE, 0x8000AD068000AD02,
-    0x0006B1BA00069DB5, 0x8000AD0A0006BDBB, 0x000699BE8000AD0E,
-    0x0006A5BC000685BF, 0x000625A800069DBD, 0x00063DAA000635A9,
-    0x000651AC000631AB, 0x000651AE000651AD, 0x8000AD12000651AF,
-    0x00012C848000AD16, 0x8000AD1E8000AD1A, 0x8000AD268000AD22,
-    0x00063DB88000AD2A, 0x0001D9BA0001C9B9, 0x0001F9BC0001D1BB,
-    0x000199BE0001F9BD, 0x000631B0000195BF, 0x000635B200063DB1,
-    0x000619B400060DB3, 0x00060DB6000605B5, 0x0005B5A3000605B7,
-    0x8000AD328000AD2E, 0x8000AD3A8000AD36, 0x0005D9A50005D1A6,
-    0x0005F9AB000358BE, 0x000C88860005F5AA, 0x0005C1AF0003BC87,
-    0x0005D9AD0005DDAE, 0x00004D810005E1AC, 0x00034DB300004D80,
-    0x00035DB500005D82, 0x8000AD468000AD42, 0x8000AD4A00034DB6,
-    0x000329BB8000AD4E, 0x00033DBD000341BA, 0x000321BF00033DBC,
-    0x000239A8000329BE, 0x00038DAA000239A9, 0x00039DAC000385AB,
-    0x000385AE000385AD, 0x000CEC840003B5AF, 0x000CE0BE8000AD52,
-    0x8000AD5A8000AD56, 0x8000AD628000AD5E, 0x000361B88000AD66,
-    0x000365BA000361B9, 0x000365BC00037DBB, 0x000365BE00036DBD,
-    0x0003CDB00000D9BF, 0x0003A1B20003A5B1, 0x0003A5B40003A1B3,
-    0x000391B60003ADB5, 0x8000AD6A000391B7, 0x8000AD728000AD6E,
-    0x0003F4EF8000AD76, 0x8000AD7E8000AD7A, 0x000098E38000AD82,
-    0x000180E18000AD86, 0x8000AD8E8000AD8A, 0x8000AD968000AD92,
-    0x0007F5BA8000AD9A, 0x8000AD9E0007FDBB, 0x000641BE8000ADA2,
-    0x0007E5BC000641BF, 0x8000ADA6000655BD, 0x8000ADAA000719B3,
-    0x0007D9B68000ADB2, 0x8000ADBA8000ADB6, 0x000798E10007D9B5,
-    0x0007B8E30001FCE3, 0x0000DD8000063CE1, 0x0000F1820000F581,
-    0x8000ADC28000ADBE, 0x8000ADCA8000ADC6, 0x8000ADD28000ADCE,
-    0x0006CCEF000120EF, 0x000CE0868000ADD6, 0x0002CDA3000D0C87,
-    0x0002DDA58000ADDA, 0x8000ADE28000ADDE, 0x8000ADE60002CDA6,
-    0x0002A9AB8000ADEA, 0x0002BDAD0002C1AA, 0x0002A1AF0002BDAC,
-    0x000E55A80002A9AE, 0x000E69AA000E5DA9, 0x000EB5AC000E69AB,
-    0x000EA9AE000EB9AD, 0x8000ADAE000EA9AF, 0x8000ADF28000ADEE,
-    0x8000ADFA8000ADF6, 0x8000AE028000ADFE, 0x000F5DB88000AE06,
-    0x000F65BA000F61B9, 0x000F65BC000F7DBB, 0x000F65BE000F6DBD,
-    0x000ED9B0000F15BF, 0x000F69B2000ED9B1, 0x000F79B4000F69B3,
-    0x000F6DB6000F79B5, 0x000EE1B3000F65B7, 0x8000AE0E8000AE0A,
-    0x8000AE168000AE12, 0x000E3DB5000E39B6, 0x000E1DBB8000AE1A,
-    0x00032484000E1DBA, 0x000FBDBF00002CBE, 0x000FBDBD000FBDBE,
-    0x00005D82000FBDBC, 0x00004580000EA5A3, 0x000E7DA600005D81,
-    0x8000AE228000AE1E, 0x000E59AA000E79A5, 0x00040086000E59AB,
-    0x000FF9AE00035C87, 0x000FF9AC000FF9AF, 0x8000AE26000FF9AD,
-    0x8000AE2E8000AE2A, 0x8000AE368000AE32, 0x8000AE3A000D44EF,
-    0x8000AE428000AE3E, 0x8000AE46000D94E3, 0x8000AE4A000D1CE1,
-    0x8000AE528000AE4E, 0x000E39B38000AE56, 0x8000AE5E8000AE5A,
-    0x8000AE668000AE62, 0x000EF9B5000EF1B6, 0x000EDDBB8000AE6A,
-    0x8000AE6E000ED5BA, 0x000E91BF8000AE72, 0x000EADBD000EADBE,
-    0x000EEDA8000EB1BC, 0x000EF1AA000EF1A9, 0x000E31AC000E11AB,
-    0x000E31AE000E31AD, 0x8000AE76000E31AF, 0x8000AE7E8000AE7A,
-    0x00001D808000AE82, 0x0003B98200000981, 0x0001E9B88000AE86,
-    0x0001F9BA0001E9B9, 0x0001E9BC0001F9BB, 0x00015DBE0001E9BD,
-    0x000E51B0000155BF, 0x000E25B2000E59B1, 0x000E29B4000E39B3,
-    0x000E1DB6000E15B5, 0x8000AE8A0001D9B7, 0x000204BE00029C84,
-    0x8000AE92000D75A3, 0x000DBDA6000DB5A5, 0x8000AE9A8000AE96,
-    0x000D99AA8000AE9E, 0x000DFDAC000D91AB, 0x000DE1AE000DE1AD,
-    0x000351B6000DDDAF, 0x000380848000AEA2, 0x8000AEA6000341B5,
-    0x0004D886000351B3, 0x000359BE0002A887, 0x000359BC000359BF,
-    0x000351BA000359BD, 0x8000AEAA00035DBB, 0x000385A38000AEAE,
-    0x8000AEB68000AEB2, 0x8000AEBE8000AEBA, 0x000395A5000385A6,
-    0x000389AB8000AEC2, 0x8000AD3E000385AA, 0x00038DAF8000AEC6,
-    0x00038DAD00038DAE, 0x8000AECA00038DAC, 0x00F5B4858000AECE,
-    0x8000AED68000AED2, 0x8000AEDE8000AEDA, 0x8000AEE68000AEE2,
-    0x8000AEEE8000AEEA, 0x000009818000AEF2, 0x8000AEF600001980,
-    0x8000AEFA0001FD82, 0x000248BE000125B3, 0x000141B68000AF02,
-    0x8000AF0A8000AF06, 0x0001A5BA000159B5, 0x8000AF0E0001ADBB,
-    0x0001ADBE8000AF12, 0x0001B5BC000199BF, 0x000560860001ADBD,
-    0x8000AF160002F487, 0x0001A0E18000AF1A, 0x00034CE38000AF1E,
-    0x8000AF268000AF22, 0x000398EF8000AF2A, 0x8000AF328000AF2E,
-    0x8000AF3A8000AF36, 0x8000AF428000AF3E, 0x8000AF4A8000AF46,
-    0x8000AF4E00044084, 0x8000AF568000AF52, 0x8000AF5A0001C8EF,
-    0x8000AF5E000E7CE1, 0x00000D80000114E3, 0x00007D8200007581,
-    0x8000AF668000AF62, 0x8000AF728000AF6A, 0x8000AF760001EDA3,
-    0x000189A6000191A5, 0x0004C086000414BE, 0x00016DAA00044487,
-    0x00017DAC000165AB, 0x000165AE000165AD, 0x8000AE8E000151AF,
-    0x8000AF7A8000AEFE, 0x8000AF6E8000AF7E, 0x8000AF868000AF82,
-    0x000631A88000AF8A, 0x000609AA000639A9, 0x000675AC000609AB,
-    0x000641AE000665AD, 0x0006C5B0000641AF, 0x0006C5B20006CDB1,
-    0x0006CDB40006DDB3, 0x000699B60006F9B5, 0x0006ADB8000699B7,
-    0x0006BDBA0006B5B9, 0x0006ADBC0006B5BB, 0x00069DBE000695BD,
-    0x00078DA8000695BF, 0x000791AA000791A9, 0x0007BDAC0007A5AB,
-    0x0007D5AE0007DDAD, 0x8000AF8E0007C9AF, 0x8000AF968000AF92,
-    0x8000AF9E8000AF9A, 0x8000AFA68000AFA2, 0x000741B88000AFAA,
-    0x000741BA000741B9, 0x000745BC00075DBB, 0x000779BE00074DBD,
-    0x0007BDB0000775BF, 0x00078DB2000785B1, 0x000781B4000781B3,
-    0x000781B6000781B5, 0x000619B3000781B7, 0x8000AFB28000AFAE,
-    0x8000AFBA8000AFB6, 0x000635B500065DB6, 0x000641BB8000AFBE,
-    0x8000AFC2000679BA, 0x000645BF8000AFC6, 0x000655BD000645BE,
-    0x8000AFCA000659BC, 0x8000AFCE00065DA3, 0x000619A68000AFD2,
-    0x8000AFDA8000AFD6, 0x00063DAA000671A5, 0x00012484000605AB,
-    0x000601AE000124BE, 0x00061DAC000601AF, 0x000649A8000611AD,
-    0x000181AA000655A9, 0x000185AC000181AB, 0x000185AE00018DAD,
-    0x0000ED8000013DAF, 0x0000F1820000F181, 0x000010868000AFE2,
-    0x8000AFE600009487, 0x0000C9B88000AFEA, 0x0000D9BA0000C9B9,
-    0x0000F9BC0000D1BB, 0x000099BE0000F9BD, 0x000145B0000095BF,
-    0x000145B200014DB1, 0x000145B400015DB3, 0x000145B600014DB5,
-    0x00075DBA0000F9B7, 0x000745B800076DBB, 0x0007B5BE00075DB9,
-    0x000775BC0007BDBF, 0x000711B20007BDBD, 0x000709B0000711B3,
-    0x000775B6000705B1, 0x000701B400077DB7, 0x000745AA000701B5,
-    0x000691A800074DAB, 0x000749AE00074DA9, 0x000755AC000749AF,
-    0x8000AFEE00074DAD, 0x8000AFF68000AFF2, 0x8000AFFE8000AFFA,
-    0x8000B0068000B002, 0x000289B38000B00A, 0x8000B0128000B00E,
-    0x8000B01A8000B016, 0x00028DB500028DB6, 0x000209BB8000B01E,
-    0x8000B022000201BA, 0x000209BF8000B026, 0x000209BD000201BE,
-    0x8000B02A000211BC, 0x8000B02E0002CDA3, 0x0002C9A68000B032,
-    0x8000B03A8000B036, 0x000245AA0002C9A5, 0x8000B03E00024DAB,
-    0x000245AE8000B042, 0x000255AC00024DAF, 0x0002ADA800024DAD,
-    0x0002B5AA0002BDA9, 0x000175AC00016DAB, 0x000175AE00017DAD,
-    0x0001CD8000016DAF, 0x0000198200000981, 0x8000B04A8000B046,
-    0x00FDB4BE00023484, 0x0001E9B88000B052, 0x000189BA0001E9B9,
-    0x00019DBC000185BB, 0x000181BE000181BD, 0x000115B00001B5BF,
-    0x000115B200011DB1, 0x0001F9B40001E9B3, 0x0001E9B60001F9B5,
-    0x0006D8E10001E1B7, 0x000614E30007C4E1, 0x8000B0560007D4E3,
-    0x8000B05A00066CE1, 0x00FC2C840000E4E3, 0x00FCA4BE8000B05E,
-    0x00FD9086000038EF, 0x00069CEF00FDD487, 0x8000B06200064CEF,
-    0x00FCC8848000B066, 0x8000B06A000271B3, 0x0001B1B60001B1B5,
-    0x8000B0728000B06E, 0x00019DBA8000B076, 0x00018DBC000195BB,
-    0x000171BE000171BD, 0x00FDBDA3000171BF, 0x8000B07A8000B04E,
-    0x8000B0828000B07E, 0x00FE7DA500FE7DA6, 0x00FE59AB8000B086,
-    0x8000B08A00FE51AA, 0x00FEBDAF8000B08E, 0x00FEBDAD00FEBDAE,
-    0x8000B09200FE41AC, 0x8000B09A8000B096, 0x8000B0A28000B09E,
-    0x8000B0AA8000B0A6, 0x00000D808000B0AE, 0x00001D8200001581,
-    0x8000B0B68000B0B2, 0x8000B0BE8000B0BA, 0x00FFBDA88000B0C2,
-    0x00FF11AA00FF15A9, 0x00FF51AC00FF29AB, 0x00FF49AE00FF51AD,
-    0x0000008600FF41AF, 0x8000B0C600036487, 0x8000B0CE8000B0CA,
-    0x8000B0D68000B0D2, 0x000081B88000B0DA, 0x000081BA000081B9,
-    0x000081BC000081BB, 0x000081BE000081BD, 0x00FF01B0000081BF,
-    0x0000E1B200FF01B1, 0x0000E1B40000E1B3, 0x0000D1B60000E9B5,
-    0x8000B0DE0000D1B7, 0x8000B0E68000B0E2, 0x8000B0EE8000B0EA,
-    0x8000B0F200FEBCE1, 0x8000B0F600FC38E3, 0x8000B0FE8000B0FA,
-    0x8000B10200FCECEF, 0x8000B10A8000B106, 0x00FED9B38000B10E,
-    0x8000B1168000B112, 0x8000B11E8000B11A, 0x00FE99B500FE91B6,
-    0x00FE85BB8000B122, 0x8000B12600FE85BA, 0x00016DBF8000B12A,
-    0x000175BD000165BE, 0x00FE19A3000175BC, 0x00002886000184BE,
-    0x8000B13900015C87, 0x00FE59A500FE51A6, 0x00FE45AB8000B13D,
-    0x8000B14100FE45AA, 0x0001ADAF8000B145, 0x0001B5AD0001A5AE,
-    0x00009D800001B5AC, 0x0000A5820000AD81, 0x8000B149000161B3,
-    0x000161B6000161B5, 0x8000B1518000B14D, 0x0000EDBA8000B155,
-    0x0000FDBC0000E5BB, 0x0000E1BE0000E5BD, 0x00027DA80000DDBF,
-    0x0002A5AA0002ADA9, 0x0002A5AC0002BDAB, 0x0002A5AE0002ADAD,
-    0x0004EC8400031DAF, 0x8000B15D8000B159, 0x8000B1658000B161,
-    0x8000B16D8000B169, 0x000371B88000B171, 0x000371BA000371B9,
-    0x000315BC000371BB, 0x000315BE00031DBD, 0x000365B000030DBF,
-    0x000365B200036DB1, 0x000369B400037DB3, 0x00035DB6000355B5,
-    0x8000B175000355B7, 0x8000B17D8000B179, 0x8000AFDE000221A3,
-    0x000221A6000221A5, 0x8000B1858000B181, 0x0003ADAA8000B189,
-    0x0003BDAC0003A5AB, 0x0003A1AE0003A5AD, 0x8000B18D00039DAF,
-    0x8000B1958000B191, 0x00001D808000B199, 0x0000398200000981,
-    0x8000B1A18000B19D, 0x8000B1AD8000B1A9, 0x0004D086000424BE,
-    0x8000B1B100032087, 0x000291A88000B1B5, 0x0002A1AA000299A9,
-    0x0002C1AC0002A1AB, 0x0002F9AE0002C9AD, 0x8000B1B90002F9AF,
-    0x8000B1C18000B1BD, 0x8000B1C98000B1C5, 0x8000B1D18000B1CD,
-    0x0001C5B88000B1D5, 0x0001D5BA0001CDB9, 0x0001F5BC0001EDBB,
-    0x00019DBE000195BD, 0x000145B0000195BF, 0x000145B200014DB1,
-    0x000145B400015DB3, 0x000145B600014DB5, 0x00FE28E10001FDB7,
-    0x00FE38E38000B1D9, 0x00041C84000480BE, 0x8000B1E18000B1DD,
-    0x8000B1E98000B1E5, 0x8000B1F18000B1ED, 0x8000B1F98000B1F5,
-    0x00FEE8EF8000B1FD, 0x000000F68000B201, 0x8000B2050000B8EF,
-    0x000009818000B209, 0x8000B20D00001980, 0x8000B21100003D82,
-    0x000748868000B219, 0x8000B21D0004FC87, 0x8000B22100FE2CE3,
-    0x0005A9B3000144E1, 0x8000B2158000B1A5, 0x8000B2298000B225,
-    0x0005C9B50005C9B6, 0x000609BB8000B22D, 0x8000B231000601BA,
-    0x000609BF8000B235, 0x000609BD000601BE, 0x8000B239000611BC,
-    0x8000B23D0005EDA3, 0x00058DA68000B241, 0x8000B2498000B245,
-    0x000645AA00058DA5, 0x8000B24D00064DAB, 0x000645AE8000B251,
-    0x000655AC00064DAF, 0x000665A800064DAD, 0x000671AA00066DA9,
-    0x0007B1AC000671AB, 0x0007B1AE0007B1AD, 0x0003A8840007B1AF,
-    0x8000B2598000B255, 0x8000B2618000B25D, 0x8000B2698000B265,
-    0x000761B88000B26D, 0x000761BA000761B9, 0x000761BC000761BB,
-    0x000765BE000761BD, 0x0007D1B0000715BF, 0x0007D1B20007D1B1,
-    0x000775B40007D1B3, 0x000761B6000761B5, 0x000690E1000761B7,
-    0x000538E38000B271, 0x00000D808000B275, 0x00003D8200003581,
-    0x8000B27D8000B279, 0x000128BE8000B281, 0x000090868000B289,
-    0x00052CEF00010487, 0x0007B9B38000B28D, 0x8000B2918000B285,
-    0x8000B2998000B295, 0x000679B5000671B6, 0x000655BB8000B29D,
-    0x8000B2A1000655BA, 0x0006E5BF8000B2A5, 0x000645BD0006F1BE,
-    0x000779A3000645BC, 0x8000B2AD8000B2A9, 0x8000B2B58000B2B1,
-    0x0006B9A50006B1A6, 0x000695AB8000B2B9, 0x8000B2BD000695AA,
-    0x000625AF8000B2C1, 0x000685AD000631AE, 0x8000B2C5000685AC,
-    0x8000B2C900063DB3, 0x0001CDB68000B2CD, 0x8000B2D58000B2D1,
-    0x0001A1BA0001CDB5, 0x8000B2D90001ADBB, 0x0001A9BE8000B2DD,
-    0x0001B5BC00019DBF, 0x0005E5A80001A9BD, 0x0005F1AA0005F5A9,
-    0x000201AC000201AB, 0x000231AE000209AD, 0x000C00BE000231AF,
-    0x8000B2E58000B2E1, 0x00001D808000B2E9, 0x0000B98200000981,
-    0x0002E9B88000B2ED, 0x0002F9BA0002E9B9, 0x0002E9BC0002F9BB,
-    0x00035DBE0002E9BD, 0x000221B0000351BF, 0x000225B200022DB1,
-    0x000229B4000239B3, 0x000215B600021DB5, 0x8000B2F10002D9B7,
-    0x000380848000B12E, 0x8000B2F9000575A3, 0x000285A6000285A5,
-    0x8000B3018000B2FD, 0x0002E9AA8000B305, 0x0002FDAC0002E5AB,
-    0x0002E1AE0002E1AD, 0x0001E8E10002D5AF, 0x000250E38000B309,
-    0x000C5086000CC084, 0x8000B30D00037087, 0x8000B3158000B311,
-    0x8000B31D8000B319, 0x8000B3258000B321, 0x000220EF8000B329,
-    0x00022DA88000B32D, 0x000241AA000239A9, 0x000241AC000241AB,
-    0x000271AE000249AD, 0x8000B331000271AF, 0x000CA4BE8000B335,
-    0x8000B33D8000B339, 0x8000B3458000B341, 0x000289B88000B349,
-    0x00029DBA000289B9, 0x0002B9BC000295BB, 0x00017DBE0002B9BD,
-    0x0002E5B0000175BF, 0x0002E5B20002EDB1, 0x0002EDB40002FDB3,
-    0x0002D5B60002DDB5, 0x000768E10002BDB7, 0x000608E3000134E3,
-    0x8000B34D00067CE1, 0x8000B3558000B351, 0x000009818000B359,
-    0x8000B35D00003980, 0x8000B36100001D82, 0x0006CCEF000064EF,
-    0x8000B36D8000B369, 0x8000B37100028DB3, 0x00028DB68000B375,
-    0x8000B37D8000B379, 0x000265BA00028DB5, 0x000CC88600026DBB,
-    0x00026DBE000CE087, 0x000275BC00025DBF, 0x8000B2F500026DBD,
-    0x8000B3818000B365, 0x8000B3898000B385, 0x8000B3918000B38D,
-    0x000EB1A88000B395, 0x000E89AA000EB9A9, 0x000EF5AC000E89AB,
-    0x000EE5AE000EEDAD, 0x000EA5B0000EDDAF, 0x000EA5B2000EADB1,
-    0x000EA5B4000EBDB3, 0x000F79B6000EA9B5, 0x000F4DB8000F79B7,
-    0x000F5DBA000F55B9, 0x000F4DBC000F55BB, 0x000FEDBE000FE5BD,
-    0x8000B399000FE5BF, 0x000D0DAA8000B39D, 0x000D1DAC000D05AB,
-    0x000D05AE000D05AD, 0x8000B3A1000D35AF, 0x8000B3A98000B3A5,
-    0x8000B3AD000DE5A3, 0x000DE5A6000DE5A5, 0x000E61B38000B3B1,
-    0x8000B3B98000B3B5, 0x8000B3C18000B3BD, 0x000E1DB5000E0DB6,
-    0x000E09BB8000B3C5, 0x8000B3C9000E01BA, 0x000E0DBF8000B3CD,
-    0x000E09BD000E01BE, 0x000FA1A3000E11BC, 0x0001E58100001582,
-    0x8000B3D10001E580, 0x000FDDA5000FCDA6, 0x000FC9AB000190BE,
-    0x0002A084000FC1AA, 0x000FCDAF8000B3D5, 0x000FC9AD000FC1AE,
-    0x8000B3D9000FD1AC, 0x0000E886000E8DB3, 0x000E89B600012C87,
-    0x8000B3E18000B3DD, 0x000171BA000E89B5, 0x8000B3E5000171BB,
-    0x0001F1BE8000B3E9, 0x000161BC0001F1BF, 0x000E3DA8000161BD,
-    0x000E41AA000E45A9, 0x000E45AC000E41AB, 0x000191AE000E4DAD,
-    0x00018C84000191AF, 0x8000B3F18000B3ED, 0x8000B3F98000B3F5,
-    0x8000B4018000B3FD, 0x000169B88000B405, 0x000175BA000175B9,
-    0x000119BC000109BB, 0x000109BE000119BD, 0x0001F1B0000101BF,
-    0x0001F1B20001F1B1, 0x000171B40001F1B3, 0x000159B6000171B5,
-    0x000DC9A3000159B7, 0x8000B40D8000B409, 0x8000B411000340BE,
-    0x000DCDA5000DCDA6, 0x000235AB8000B415, 0x8000B419000235AA,
-    0x0002B5AF8000B41D, 0x000225AD0002B5AE, 0x8000B421000225AC,
-    0x8000B4298000B425, 0x8000B4318000B42D, 0x8000B4398000B435,
-    0x000039808000B43D, 0x0000058200003981, 0x0004D0868000B441,
-    0x8000B44900034487, 0x0004CC848000B44D, 0x8000B4558000B451,
-    0x000264EF8000B459, 0x8000B4618000B45D, 0x000110E38000B465,
-    0x0001F8E18000B469, 0x8000B4718000B46D, 0x8000B4798000B475,
-    0x000375BB8000B47D, 0x8000B481000375BA, 0x0000D9BF8000B485,
-    0x000365BD0000D1BE, 0x0003B9B3000365BC, 0x8000B48D8000B489,
-    0x8000B4958000B491, 0x000395B5000351B6, 0x000241A88000B499,
-    0x000251AA000251A9, 0x00027DAC000265AB, 0x0002B1AE0002B1AD,
-    0x000568BE0002B1AF, 0x8000B4A18000B49D, 0x00000D808000B4A5,
-    0x0000AD820000A581, 0x000169B88000B4A9, 0x000109BA000169B9,
-    0x000119BC000109BB, 0x000109BE000119BD, 0x0002D1B0000109BF,
-    0x0002D1B20002D1B1, 0x000175B40002D1B3, 0x000169B600017DB5,
-    0x000E70E1000161B7, 0x0000B0E3000F4CE1, 0x8000B4AD000F50E3,
-    0x00E748EF8000B4B1, 0x000228E100049885, 0x001370E38000B4B9,
-    0x000678BE8000B4BD, 0x00005CEF8000B4C1, 0x000299A3000E6CEF,
-    0x0007A8868000B4C5, 0x8000B4C900048087, 0x0002B5A5000271A6,
-    0x000255AB8000B4CD, 0x8000B4D1000255AA, 0x0001F9AF8000B4D5,
-    0x000245AD0001F1AE, 0x000679A8000245AC, 0x00068DAA000679A9,
-    0x00069DAC000685AB, 0x000681AE000681AD, 0x8000B4B5000681AF,
-    0x8000B4DD8000B4D9, 0x8000B4E58000B4E1, 0x8000B4ED8000B4E9,
-    0x0006A5B88000B4F1, 0x0006A5BA0006ADB9, 0x0006ADBC0006BDBB,
-    0x000771BE000771BD, 0x0006C5B0000771BF, 0x0006D9B20006C9B1,
-    0x0006C9B40006D9B3, 0x0006B1B60006C9B5, 0x000605B30006B1B7,
-    0x8000B4F98000B4F5, 0x8000B5018000B4FD, 0x000615B5000601B6,
-    0x00060DBB8000B505, 0x8000B509000601BA, 0x00060DBF8000B50D,
-    0x00060DBD00060DBE, 0x8000B51100060DBC, 0x8000B515000641A3,
-    0x000645A68000B519, 0x8000B5218000B51D, 0x000645AA000651A5,
-    0x00012484000649AB, 0x000649AE8000B529, 0x000649AC000649AF,
-    0x0008459F000649AD, 0x00090D9D00090D9E, 0x000D699B000B899C,
-    0x000F8999000D119A, 0x00F1259700F10D98, 0x00F5FD9500F12596,
-    0x00F7999300F50D94, 0x00F9319100F99992, 0x00FD298F00F93990,
-    0x00019D8100019D82, 0x8000B52D00018180, 0x0001ECEF00048CEF,
-    0x000208E18000B531, 0x000130E30005F4E3, 0x001C0086000504E1,
-    0x8000B53500018887, 0x0001A1A18000B539, 0x001D09A38000B53D,
-    0x001CF9A5000369A2, 0x00191DA7001D11A4, 0x001521A900191DA6,
-    0x001579AB0018B5A8, 0x001191AD001521AA, 0x002C01AF001001AC,
-    0x002D5DB1001089AE, 0x0029E9B3002DE5B0, 0x002801B5002801B2,
-    0x8000B5410029EDB4, 0x8000B5498000B545, 0x8000B5518000B54D,
-    0x000319A38000B555, 0x0005FDA10004D9A2, 0x00FE69820005FDA0,
-    0x8000B55900FA0183, 0x00F4B9868000B55D, 0x00FA118400F63587,
-    0x00F6F58A00FAB185, 0x8000B56100F2258B, 0x000CA18E8000B565,
-    0x00F23D8C000E318F, 0x000E399200F22D8D, 0x000344840008C193,
-    0x000B1596001CC0BE, 0x000A5D9400065597, 0x0006659A000A5595,
-    0x8000B5690007499B, 0x8000B5718000B56D, 0x0000919C8000B575,
-    0x000209B88000B579, 0x000219BA000209B9, 0x000239BC000211BB,
-    0x000259BE000239BD, 0x000285B0000255BF, 0x000285B200028DB1,
-    0x000285B400029DB3, 0x000285B600028DB5, 0x000565A8000239B7,
-    0x00057DAA000569A9, 0x000501AC000509AB, 0x000531AE000509AD,
-    0x8000B57D000531AF, 0x8000B4458000B525, 0x8000B5858000B581,
-    0x0001F5810001E582, 0x0004B5B30001F580, 0x8000B589001D8C84,
-    0x8000B5918000B58D, 0x0004C9B50004C9B6, 0x000309BB8000B599,
-    0x001D2886000301BA, 0x000309BF001D2087, 0x000309BD000301BE,
-    0x8000B59D000311BC, 0x8000B5A10004F1A3, 0x00048DA6001CC0BE,
-    0x8000B5A98000B5A5, 0x000345AA00048DA5, 0x001FA0BE00034DAB,
-    0x000345AE8000B5AD, 0x000355AC00034DAF, 0x8000B5B100034DAD,
-    0x8000B5B98000B5B5, 0x8000B5C18000B5BD, 0x8000B5C98000B5C5,
-    0x8000B5CD0006D4EF, 0x8000B5D1000674E1, 0x8000B5D5000154E3,
-    0x8000B5DD8000B5D9, 0x00029DB38000B5E1, 0x8000B5E98000B5E5,
-    0x8000B5F18000B5ED, 0x0002A5B5000251B6, 0x000205BB8000B5F5,
-    0x8000B5F9000205BA, 0x0001FDBF8000B5FD, 0x000215BD0001FDBE,
-    0x001D5DA3000215BC, 0x001FE58100001582, 0x8000B601001FE580,
-    0x001D65A5001D91A6, 0x001DC5AB8000B609, 0x00030886001DC5AA,
-    0x001E3DAF00037C87, 0x001DD5AD001E3DAE, 0x8000B60D001DD5AC,
-    0x8000B611001F0DB3, 0x001F3DB68000B615, 0x8000B61D8000B619,
-    0x001FE9BA001F05B5, 0x8000B621001FE9BB, 0x001FE9BE8000B625,
-    0x001FE9BC001FD5BF, 0x8000B595001FE9BD, 0x8000B6298000B605,
-    0x8000B6318000B62D, 0x8000B6398000B635, 0x001FBDA88000B63D,
-    0x001FE1AA001FE5A9, 0x001FE9AC001FF9AB, 0x001F39AE001FD9AD,
-    0x001F49B0001F39AF, 0x001F59B2001F49B1, 0x001F49B4001F59B3,
-    0x001F71B6001F75B5, 0x001F55B8001F6DB7, 0x001F55BA001F5DB9,
-    0x001F39BC001F29BB, 0x001F29BE001F39BD, 0x001E49A3001F21BF,
-    0x8000B6458000B641, 0x8000B64D8000B649, 0x001E41A5001E79A6,
-    0x001EADAB8000B651, 0x8000B655001EADAA, 0x001E91AF8000B659,
-    0x001EADAD001EADAE, 0x8000B65D001EADAC, 0x8000B6658000B661,
-    0x00001D808000B669, 0x0000398200000981, 0x8000B6718000B66D,
-    0x8000B6750001A884, 0x000400868000B679, 0x8000B67D0001D487,
-    0x00019DA88000B681, 0x0001C1AA0001C5A9, 0x0001C5AC0001C1AB,
-    0x0001F1AE0001CDAD, 0x8000B6850001F1AF, 0x8000B68D8000B689,
-    0x8000B6958000B691, 0x8000B69D8000B699, 0x0000CDB88000B6A1,
-    0x0000D5BA0000D5B9, 0x0000FDBC0000E5BB, 0x00009DBE000095BD,
-    0x000169B0000095BF, 0x000179B2000169B1, 0x000169B4000179B3,
-    0x0000FDB6000169B5, 0x000249B30000F5B7, 0x8000B6A98000B6A5,
-    0x8000B6AD000480BE, 0x00024DB500024DB6, 0x0002ADBB8000B6B1,
-    0x8000B6B50002ADBA, 0x0002ADBF8000B6B9, 0x0002ADBD0002ADBE,
-    0x8000B6BD0002ADBC, 0x8000B6C100020DA3, 0x000209A68000B6C5,
-    0x8000B6CD8000B6C9, 0x0002E9AA000209A5, 0x8000B6D10002E9AB,
-    0x0002E9AE8000B6D5, 0x0002E9AC0002E9AF, 0x8000B6D90002E9AD,
-    0x000504BE8000B6DD, 0x8000B6E1000614BE, 0x8000B6ED8000B6E9,
-    0x000021808000B6F1, 0x0000398200002181, 0x0004B08600032084,
-    0x8000B6F500035487, 0x8000B6FD8000B6F9, 0x8000B7058000B701,
-    0x8000B70D8000B709, 0x8000B711000220EF, 0x8000B7198000B715,
-    0x8000B71D000258E3, 0x8000B721000184E1, 0x8000B7298000B725,
-    0x000130E18000B72D, 0x00007CE3001E34E1, 0x8000B731001E18E3,
-    0x8000B7398000B735, 0x0005A8848000B73D, 0x8000B7458000B741,
-    0x8000B74D8000B749, 0x001EA8EF8000B751, 0x00023DB3001E88EF,
-    0x000468868000B759, 0x000000D40005DC87, 0x0002FDB50002FDB6,
-    0x0002D9BB8000B75D, 0x0006A4BE0002D1BA, 0x000161BF8000B761,
-    0x00017DBD00017DBE, 0x001DFDBC0002C1BC, 0x001D65BE001D65BD,
-    0x001DF9B8001D7DBF, 0x001DE9BA001DF9B9, 0x001DA9B4001DE9BB,
-    0x001D81B6001DA9B5, 0x001E51B0001D81B7, 0x001E51B2001E51B1,
-    0x001EBDAC001E51B3, 0x001E11AE001E41AD, 0x001F3DA8001E11AF,
-    0x001E85AA001E85A9, 0x0000CD80001EA5AB, 0x0000DD820000D581,
-    0x8000B7698000B765, 0x8000B7718000B76D, 0x00057DA38000B775,
-    0x8000B77D8000B779, 0x8000B7858000B781, 0x0005BDA50005BDA6,
-    0x000599AB8000B789, 0x8000B78D000591AA, 0x000621AF8000B791,
-    0x00063DAD00063DAE, 0x8000B795000581AC, 0x8000B7990006D9B3,
-    0x000691B68000B79D, 0x8000B7A58000B7A1, 0x0006ADBA0006E1B5,
-    0x8000B7A90006B5BB, 0x000691BE8000B7AD, 0x0006A1BC00069DBF,
-    0x8000B6E50006A1BD, 0x8000B7B18000B755, 0x00000D808000B7B5,
-    0x0001918200019181, 0x000635A88000B7B9, 0x000635AA00063DA9,
-    0x00060DAC000615AB, 0x00067DAE000675AD, 0x00060DB0000675AF,
-    0x000611B2000615B1, 0x000631B4000611B3, 0x000629B600063DB5,
-    0x000619B8000629B7, 0x0007F9BA000619B9, 0x0007E9BC0007F9BB,
-    0x0007DDBE0007E9BD, 0x00079DA30007D5BF, 0x8000B7C18000B7BD,
-    0x8000B7C500028484, 0x0007A5A50007D5A6, 0x0007F1AB8000B7C9,
-    0x000C00860007E9AA, 0x0007D9AF00017C87, 0x0007E5AD0007D5AE,
-    0x8000B7CD0007E5AC, 0x8000B7D1000629B3, 0x0006A5B68000B7D5,
-    0x8000B7DD8000B7D9, 0x000165BA0006ADB5, 0x8000B7E100016DBB,
-    0x00016DBE8000B7E5, 0x000175BC00015DBF, 0x0006C5A800016DBD,
-    0x0006E5AA0006D9A9, 0x000661AC000661AB, 0x000661AE000661AD,
-    0x8000B7E9000661AF, 0x8000B7F18000B7ED, 0x8000B7F98000B7F5,
-    0x8000B8018000B7FD, 0x000189B88000B805, 0x00019DBA000189B9,
-    0x0001B9BC000195BB, 0x000151BE0001B9BD, 0x0001E5B0000151BF,
-    0x0001E5B20001EDB1, 0x0001EDB40001FDB3, 0x0001D5B60001DDB5,
-    0x0005F5A60001BDB7, 0x8000B80D8000B809, 0x8000B8110005FDA5,
-    0x8000B815000579A3, 0x00023DAE8000B819, 0x000225AC00020DAF,
-    0x000235AA00023DAD, 0x8000B81D00023DAB, 0x8000B8258000B821,
-    0x8000B82D8000B829, 0x000015808000B831, 0x0000158200001D81,
-    0x000260848000B839, 0x000348878000B83D, 0x8000B841000CCC86,
-    0x8000B8498000B845, 0x8000B8518000B84D, 0x000D28868000B855,
-    0x000000A1000C4087, 0x8000B8590001A4E1, 0x8000B85D0002ECE3,
-    0x8000B8658000B861, 0x8000B86900073CEF, 0x8000B8718000B86D,
-    0x8000B8798000B875, 0x8000B8818000B87D, 0x8000B885000385B3,
-    0x000381B6000381B5, 0x8000B88D8000B889, 0x000349BA8000B891,
-    0x000359BC000349BB, 0x000349BE000359BD, 0x000650E1000349BF,
-    0x000014E30007ACE1, 0x8000B89500063CE3, 0x8000B89D8000B899,
-    0x00003580000D5484, 0x0000058200003D81, 0x000D7CBE8000B8A1,
-    0x00074CEF8000B8A9, 0x000209A300074CEF, 0x8000B8B18000B8AD,
-    0x8000B8B98000B8B5, 0x00020DA500020DA6, 0x0002C5AB8000B8BD,
-    0x8000B8C10002C5AA, 0x0002C5AF8000B8C5, 0x0002D5AD0002C5AE,
-    0x0005D5BC0002D5AC, 0x0005E1BE0005DDBD, 0x0005BDB80005E1BF,
-    0x0005D1BA0005D5B9, 0x0005A9B40005D1BB, 0x0005B5B60005B5B5,
-    0x000605B0000585B7, 0x000605B200060DB1, 0x000665AC0005B9B3,
-    0x000665AE00066DAD, 0x0006BDA800067DAF, 0x0006A5AA0006ADA9,
-    0x8000B8C900067DAB, 0x8000B8D18000B8CD, 0x8000B8D98000B8D5,
-    0x8000B8E18000B8DD, 0x000E5DA88000B8E5, 0x000E79AA000E61A9,
-    0x000E69AC000E79AB, 0x000E99AE000E69AD, 0x8000B8E9000E99AF,
-    0x8000B8F18000B8ED, 0x8000B8F98000B8F5, 0x8000B9018000B8FD,
-    0x000E81B88000B905, 0x000E81BA000E81B9, 0x000E8DBC000E9DBB,
-    0x000EBDBE000EB5BD, 0x000EF9B0000F4DBF, 0x000ECDB2000EC5B1,
-    0x000EC1B4000EC5B3, 0x000EC1B6000EC1B5, 0x8000B835000EC1B7,
-    0x8000B909000EF5B3, 0x000EEDB68000B8A5, 0x8000B9118000B90D,
-    0x000E39BA000EE5B5, 0x0003E484000E05BB, 0x000E01BE0003E4BE,
-    0x000E1DBC000E01BF, 0x000EB1A3000E15BD, 0x00005D8100005D82,
-    0x8000B91500004580, 0x000EA1A5000EA9A6, 0x000E41AB8000B919,
-    0x00004886000E7DAA, 0x000E45AF00006887, 0x000E51AD000E45AE,
-    0x000E35A8000E59AC, 0x000E41AA000E41A9, 0x000E41AC000E41AB,
-    0x000E71AE000E49AD, 0x8000B91D000E71AF, 0x00016C848000B921,
-    0x8000B9298000B925, 0x8000B9318000B92D, 0x000161B88000B935,
-    0x000161BA000161B9, 0x000161BC000161BB, 0x000161BE000161BD,
-    0x000E81B0000161BF, 0x000E81B2000E81B1, 0x000E81B4000E81B3,
-    0x000E81B6000E81B5, 0x000EC9B3000E81B7, 0x8000B93D8000B939,
-    0x8000B9458000B941, 0x000EC9B5000EC9B6, 0x000E45BB8000B949,
-    0x8000B94D000E45BA, 0x000E45BF8000B951, 0x000E55BD000E45BE,
-    0x8000B955000E55BC, 0x8000B959000E8DA3, 0x000E8DA68000B95D,
-    0x8000B9658000B961, 0x000E01AA000E8DA5, 0x8000B969000E01AB,
-    0x000E01AE8000B96D, 0x000E11AC000E01AF, 0x0002DDA8000E11AD,
-    0x00032DAA000325A9, 0x000321AC000321AB, 0x000321AE000321AD,
-    0x8000B971000321AF, 0x8000B9798000B975, 0x8000B9818000B97D,
-    0x8000B989000278BE, 0x000389B88000B98D, 0x00039DBA000389B9,
-    0x0003B9BC000395BB, 0x000079BE0003B9BD, 0x0003E5B0000079BF,
-    0x0003E5B20003EDB1, 0x0003EDB40003FDB3, 0x0003D5B60003DDB5,
-    0x0000AD810003BDB7, 0x0003ADB30000AD80, 0x0003A1B50000BD82,
-    0x8000B9958000B991, 0x000400860003A9B6, 0x000321BB0003D487,
-    0x000321BD000329BA, 0x000311BF000339BC, 0x0003E5A3000321BE,
-    0x8000B99D8000B999, 0x8000B9A10030F085, 0x0003E9A50003E1A6,
-    0x000369AB8000B9A5, 0x8000B9A9000361AA, 0x000359AF8000B9AD,
-    0x000369AD000369AE, 0x0001B8E3000371AC, 0x000F98E1000F70E1,
-    0x8000B9B1000E00E3, 0x8000B9B98000B9B5, 0x8000B9C18000B9BD,
-    0x8000B9C98000B9C5, 0x000164EF8000B9CD, 0x005D30BF8000B9D1,
-    0x8000B9D5000EF4EF, 0x000504848000B9D9, 0x8000B9DD000504BE,
-    0x8000B9E10001A4E1, 0x8000B9E5000064E3, 0x8000B9ED8000B9E9,
-    0x000015800003A0EF, 0x0000058200001D81, 0x000EC9AA8000B9F1,
-    0x000E95A8000EC9AB, 0x000EF9AE000ECDA9, 0x000ECDAC000EF9AF,
-    0x00040086000EC5AD, 0x8000B9F900048487, 0x8000BA018000B9FD,
-    0x8000BA098000BA05, 0x000DD1BA8000BA0D, 0x000DCDB8000DDDBB,
-    0x000DEDBE000DC5B9, 0x000DC5BC000D95BF, 0x000E49B2000DE5BD,
-    0x000E49B0000E49B3, 0x000E49B6000E49B1, 0x000E49B4000E49B7,
-    0x000639A8000E49B5, 0x00069DAA000639A9, 0x0006D5AC000695AB,
-    0x0006C9AE0006DDAD, 0x8000B9F50006C1AF, 0x8000BA158000BA11,
-    0x8000BA1D8000BA19, 0x8000BA258000BA21, 0x000769B88000BA29,
-    0x00077DBA000769B9, 0x000715BC00070DBB, 0x000701BE000705BD,
-    0x000689B000073DBF, 0x000779B2000689B1, 0x000769B4000779B3,
-    0x000759B6000769B5, 0x000615B3000759B7, 0x8000BA318000BA2D,
-    0x8000BA398000BA35, 0x00065DB500065DB6, 0x00064DBB8000BA3D,
-    0x8000BA4100064DBA, 0x0007ADBF8000BA45, 0x0007ADBD0007ADBE,
-    0x8000BA490007ADBC, 0x8000BA4D000651A3, 0x000619A68000BA51,
-    0x8000BA598000BA55, 0x000609AA000619A5, 0x8000BA5D000609AB,
-    0x0007E9AE8000BA61, 0x0007E9AC0007E9AF, 0x0006A9A80007E9AD,
-    0x0006B9AA0006A9A9, 0x0006A9AC0006B9AB, 0x00013DAE0006A9AD,
-    0x0001F980000135AF, 0x0001C5820001CD81, 0x00027CBE00027884,
-    0x000178878000BA65, 0x0001D1B800009C86, 0x0001E1BA0001D9B9,
-    0x000191BC0001E1BB, 0x000195BE00019DBD, 0x00014DB0000189BF,
-    0x00015DB2000155B1, 0x00014DB4000155B3, 0x0001F1B60001F1B5,
-    0x000639B50001F1B7, 0x8000BA6D8000BA69, 0x8000BA71000631B6,
-    0x000629B38000BA75, 0x000199BD8000BA79, 0x00012DBF000191BC,
-    0x8000BA7D000191BE, 0x000189BB8000B985, 0x000671A3000191BA,
-    0x8000BA858000BA81, 0x8000BA8D8000BA89, 0x000661A5000669A6,
-    0x0001D1AB8000BA91, 0x8000BA950001C9AA, 0x000175AF8000BA99,
-    0x0001C1AD0001C9AE, 0x8000BA9D0001C9AC, 0x0007BCEF8000BAA1,
-    0x8000BAA98000BAA5, 0x8000BAB18000BAAD, 0x8000BAB98000BAB5,
-    0x8000BAC18000BABD, 0x000688E18000BAC5, 0x000760E38000BAC9,
-    0x000075808000BACD, 0x0000198200007981, 0x8000BAD100024084,
-    0x000314BE8000BAD9, 0x003CC0868000BADD, 0x8000BAE100034487,
-    0x8000BAE98000BAE5, 0x8000BAF18000BAED, 0x8000BAF98000BAF5,
-    0x000228EF8000BAFD, 0x8000BB058000BB01, 0x8000BB0D8000BB09,
-    0x003CEC848000BB11, 0x8000BB198000BB15, 0x000194E18000BB1D,
-    0x0002E4E38000BB21, 0x000351B38000BB25, 0x8000BB2D8000BB29,
-    0x8000BB358000BB31, 0x000345B5000365B6, 0x000331BB8000BB39,
-    0x8000BB3D000329BA, 0x0000F1BF8000BB41, 0x000311BD000319BE,
-    0x000285A8000321BC, 0x000299AA00028DA9, 0x00028DAC000299AB,
-    0x0002DDAE0002A5AD, 0x0001FD800002D5AF, 0x0001C5820001CD81,
-    0x003C9086003D64BE, 0x8000BB49003DB887, 0x000151B8003C6084,
-    0x000151BA000151B9, 0x000171BC000151BB, 0x000171BE000171BD,
-    0x0002ADB0000171BF, 0x0002B5B20002B5B1, 0x00029DB4000285B3,
-    0x00017DB6000175B5, 0x8000BB4D000171B7, 0x8000BB558000BB51,
-    0x8000BB5D8000BB59, 0x8000BB658000BB61, 0x003CECBE000054EF,
-    0x8000BB6900061CE1, 0x8000BB6D000090E3, 0x8000BB758000BB71,
-    0x000251A38000BB79, 0x003F44848000BB7D, 0x8000BB858000BB81,
-    0x000245A5000265A6, 0x000231AB8000BB89, 0x8000BB8D000229AA,
-    0x0001F1AF8000BB91, 0x000211AD000219AE, 0x003E55A8000221AC,
-    0x003E61AA003E59A9, 0x003E61AC003E61AB, 0x003E61AE003E61AD,
-    0x8000BB45003E61AF, 0x8000BB998000BB95, 0x8000BBA18000BB9D,
-    0x8000BBA98000BBA5, 0x003EE9B88000BBAD, 0x003EFDBA003EE9B9,
-    0x003E95BC003EF1BB, 0x003E95BE003E9DBD, 0x003EE1B0003E8DBF,
-    0x003EE5B2003EE1B1, 0x003EE5B4003EFDB3, 0x003EE5B6003EEDB5,
-    0x003EE9B3003ED9B7, 0x0000B58100001582, 0x8000BBB10000B580,
-    0x003EE9B5003EE1B6, 0x003E25BB8000BBB5, 0x0000A886003E25BA,
-    0x003E29BF0000A487, 0x003E29BD003E21BE, 0x8000BBB9003E31BC,
-    0x8000BBBD003EADA3, 0x003EA5A68000BBC1, 0x8000BBC98000BBC5,
-    0x003E61AA003EADA5, 0x8000BBCD003E61AB, 0x003E65AE8000BBD1,
-    0x003E75AC003E6DAF, 0x003E39A8003E6DAD, 0x003E8DAA003E85A9,
-    0x003E9DAC003E85AB, 0x003E8DAE003E85AD, 0x8000BBD5003E85AF,
-    0x8000BBDD8000BBD9, 0x8000BBE58000BBE1, 0x8000BBED8000BBE9,
-    0x003E9DB88000BBF1, 0x003EA5BA003EADB9, 0x000179BC000169BB,
-    0x000169BE000179BD, 0x003EC9B0000169BF, 0x003EDDB2003ED5B1,
-    0x003ECDB4003ED5B3, 0x003EB5B6003EB1B5, 0x8000BBF5003EADB7,
-    0x003E05B5003E15B6, 0x003E15B38000BBF9, 0x8000BC018000BBFD,
-    0x003E65BF8000BC05, 0x003E11BD003E19BE, 0x003E01BB003E19BC,
-    0x8000BAD5003E1DBA, 0x003E45A38000BC09, 0x8000BC118000BC0D,
-    0x8000BC198000BC15, 0x003E55A5003E45A6, 0x003E51AB8000BC1D,
-    0x8000BC21003E4DAA, 0x003E35AF00002CBE, 0x003E41AD003E49AE,
-    0x0003E982003E49AC, 0x0003D980003FD0EF, 0x000350870003D181,
-    0x8000BC290004B486, 0x8000BC318000BC2D, 0x8000BC398000BC35,
-    0x8000BC418000BC3D, 0x8000BC45003E84E3, 0x00028DB3003E44E1,
-    0x000344848000BC49, 0x8000BC4D000440BE, 0x000281B50002BDB6,
-    0x000349BB8000BC51, 0x8000BC55000341BA, 0x000349BF8000BC59,
-    0x000349BD000341BE, 0x8000BC5D000351BC, 0x8000BC658000BC61,
-    0x8000BC69000580BE, 0x8000BC718000BC6D, 0x000261A88000BC75,
-    0x000265AA00026DA9, 0x000269AC000279AB, 0x000255AE00025DAD,
-    0x0003E5B000039DAF, 0x0003E5B20003EDB1, 0x0003E5B40003FDB3,
-    0x0003E5B60003EDB5, 0x00036DB800035DB7, 0x000371BA000371B9,
-    0x00034DBC000371BB, 0x000331BE000331BD, 0x8000BC79000331BF,
-    0x8000BC818000BC7D, 0x003EC8E38000BC85, 0x0001A8E100051C84,
-    0x000034EF8000BC89, 0x8000BC918000BC8D, 0x8000BC9D8000BC99,
-    0x8000BCA58000BCA1, 0x00005D808000BCA9, 0x0000618200006181,
-    0x8000BCAD000381A3, 0x0003B1A600038DA5, 0x0004C08600079884,
-    0x00024DAA00054887, 0x00025DAC000245AB, 0x00024DAE000245AD,
-    0x8000BCB1000245AF, 0x8000BCB98000BCB5, 0x8000BCC18000BCBD,
-    0x8000BCC98000BCC5, 0x000620BE003EC4EF, 0x8000BCCD003E8CE1,
-    0x8000BCD1003E80E3, 0x8000BCD98000BCD5, 0x0001E1B38000BCDD,
-    0x8000BCE58000BCE1, 0x8000BCED8000BCE9, 0x0001F1B50001F9B6,
-    0x0001B1BB8000BCF1, 0x8000BCF50001A1BA, 0x000171BF8000BCF9,
-    0x000191BD000199BE, 0x000619A80001A1BC, 0x000629AA000619A9,
-    0x000639AC000629AB, 0x000629AE000639AD, 0x8000BC95000625AF,
-    0x0000198100001182, 0x8000BCFD0007C180, 0x8000BD058000BD01,
-    0x0006D1B88000BD09, 0x0006D1BA0006D1B9, 0x000751BC0006D1BB,
-    0x000751BE000751BD, 0x000625B0000751BF, 0x000635B2000635B1,
-    0x00061DB4000605B3, 0x0006FDB60006F5B5, 0x0006A1A30006F5B7,
-    0x0000AC868000BD0D, 0x8000BD1100034C87, 0x0006B1A50006B9A6,
-    0x0006F1AB8000BD15, 0x8000BD190006E1AA, 0x000631AF8000BD1D,
-    0x0006D1AD0006D9AE, 0x8000BD210006E1AC, 0x8000BD250007F1B3,
-    0x000781B68000BD29, 0x8000BD318000BD2D, 0x000769BA0007E1B5,
-    0x8000BD35000769BB, 0x000769BE8000BD39, 0x000769BC00075DBF,
-    0x00062DA8000769BD, 0x00063DAA000635A9, 0x00062DAC000635AB,
-    0x000699AE000645AD, 0x8000BD3D000699AF, 0x8000BD458000BD41,
-    0x8000BD4D8000BD49, 0x8000BD558000BD51, 0x0006F9B88000BD59,
-    0x00068DBA0006F9B9, 0x00069DBC000685BB, 0x000685BE000685BD,
-    0x0006E9B00006B5BF, 0x0006F9B20006E9B1, 0x0006E9B40006F9B3,
-    0x0006C9B60006DDB5, 0x0006B5A30006C9B7, 0x8000BD618000BD5D,
-    0x8000BD698000BD65, 0x0006A5A50006C5A6, 0x00062DAB8000BD6D,
-    0x8000BD7100062DAA, 0x000619AF8000BD79, 0x00062DAD00062DAE,
-    0x0000518000062DAC, 0x0000618200005981, 0x8000BD758000BC25,
-    0x000178878000BD7D, 0x8000BD81000C0086, 0x8000BD898000BD85,
-    0x8000BD918000BD8D, 0x8000BD998000BD95, 0x8000BDA18000BD9D,
-    0x000299B38000BDA5, 0x0002B9B500039484, 0x8000BDAD8000BDA9,
-    0x8000BDB10002B1B6, 0x0002FDBB8000BDB5, 0x0002D9BD0002E1BA,
-    0x000319BF0002EDBC, 0x000618E30002D1BE, 0x000644E18000BDB9,
-    0x8000BDC18000BDBD, 0x8000BDC5000190E1, 0x8000BDC9000240E3,
-    0x8000BDD18000BDCD, 0x0006C4EF000224EF, 0x8000BDD98000BDD5,
-    0x000205A98000BDDD, 0x000215AB0002B9A8, 0x00023DAD00021DAA,
-    0x000251AF00023DAC, 0x0000D98000025DAE, 0x000005820000D981,
-    0x0003FCBE8000BDE1, 0x000CA8878000BDE9, 0x0003F1B9000D9C86,
-    0x0003F1BB0003E9B8, 0x000355BD0003F9BA, 0x000355BF0003E9BC,
-    0x000235B1000355BE, 0x000201B3000229B0, 0x0003D1B5000235B2,
-    0x0003D1B7000219B4, 0x000D44850003D9B6, 0x8000BDF18000BDED,
-    0x8000BDF500039DA3, 0x0003B5A60003BDA5, 0x8000BDFD8000BDF9,
-    0x0003E5AA8000BE01, 0x0003E9AC0003F9AB, 0x0003D5AE0003DDAD,
-    0x0002ADB600021DAF, 0x8000BE098000BE05, 0x8000BE0D0002A5B5,
-    0x8000BE110002B5B3, 0x000175BE8000BE15, 0x000169BC00017DBF,
-    0x000289BA000169BD, 0x8000BE19000289BB, 0x000294E18000BE1D,
-    0x00123CE30006DCEF, 0x8000BE258000BE21, 0x8000BE348000BE30,
-    0x000F6C840007E8EF, 0x8000BE380006DCE1, 0x8000BE3C0006A8E3,
-    0x003E7CEF0006F8E3, 0x000221A300012CE1, 0x8000BE40000000F1,
-    0x8000BE488000BE44, 0x000231A5000239A6, 0x00021DAB8000BE4C,
-    0x8000BE5000021DAA, 0x0001E9AF8000BE54, 0x0001FDAD0001E1AE,
-    0x0024218F0001FDAC, 0x8000BE588000BDE5, 0x000FF9800000C4BE,
-    0x00001182000FF981, 0x000000868000BE5C, 0x8000BE6000038487,
-    0x8000BE688000BE64, 0x8000BE708000BE6C, 0x0008DD9F8000BE74,
-    0x0009F19D0008659E, 0x0034719B0035799C, 0x003089990035E19A,
-    0x0030E1970030D198, 0x003CED95003D7596, 0x00391993003CED94,
-    0x0038F9910038F192, 0x0014A5B100269990, 0x006809B30014A5B0,
-    0x006805B50015E9B2, 0x8000BE78006819B4, 0x8000BE808000BE7C,
-    0x8000BE888000BE84, 0x000D61A38000BE8C, 0x000C99A1000C8DA2,
-    0x0000C5A1000D71A0, 0x0004C9A38000BE90, 0x000439A50000D1A2,
-    0x0018F9A70004D1A4, 0x001C01A9001801A6, 0x001C6DAB001879A8,
-    0x0010D1AD001CCDAA, 0x001401AF001001AC, 0x002521820010D9AE,
-    0x8000BE940027E183, 0x003879868000BE98, 0x00399984003D9587,
-    0x003DA58A00399985, 0x8000BE9C003C358B, 0x0031E18E8000BEA0,
-    0x0031B98C0034818F, 0x003561920031B18D, 0x8000BEA400347193,
-    0x0009CD968000BEA8, 0x0008E194000C9997, 0x000D419A00097D95,
-    0x8000BEAC000D5D9B, 0x8000BEB48000BEB0, 0x000CA99C8000BEB8,
-    0x000F35A88000BEBC, 0x000E81AA000E81A9, 0x000E8DAC000E9DAB,
-    0x000EBDAE000EB5AD, 0x00000D80000ED9AF, 0x00001D8200001581,
-    0x8000BEC48000BEC0, 0x0000BC848000BEC8, 0x000E55B8000400BE,
-    0x000E61BA000E5DB9, 0x000E19BC000E7DBB, 0x000E09BE000E19BD,
-    0x000EA9B0000E01BF, 0x000EB1B2000EB5B1, 0x000E89B4000EB1B3,
-    0x000E75B6000E89B5, 0x000D09B3000E6DB7, 0x000488868000BED0,
-    0x00045C840003C887, 0x000D05B5000D35B6, 0x000D71BB8000BED4,
-    0x8000BED8000D6DBA, 0x000D4DBF8000BEDC, 0x000D65BD000D4DBE,
-    0x8000BEE0000D65BC, 0x8000BEE4000D4DA3, 0x000D71A68000BEE8,
-    0x8000BEF08000BEEC, 0x000D29AA000D41A5, 0x8000BEF4000D35AB,
-    0x000D09AE8000BEF8, 0x000D21AC000D09AF, 0x0002A9A8000D21AD,
-    0x0002B9AA0002A9A9, 0x0002A9AC0002B9AB, 0x000201AE0002A9AD,
-    0x8000BEFC000201AF, 0x8000BF048000BF00, 0x8000BF0C8000BF08,
-    0x8000BF148000BF10, 0x0003EDB88000BF18, 0x000385BA00038DB9,
-    0x0003BDBC0003A5BB, 0x0003ADBE0003A5BD, 0x000241B00003A1BF,
-    0x000241B2000241B1, 0x000235B4000241B3, 0x0003F5B60003FDB5,
-    0x000D25B70003D5B7, 0x000D11B5000D25B6, 0x000D15B3000D19B4,
-    0x000EE5B1000D15B2, 0x000D55BF000EE5B0, 0x000D4DBD000D4DBE,
-    0x000D69BB000D4DBC, 0x000D75B9000D65BA, 0x8000BF1C000D79B8,
-    0x8000BF2000050CBE, 0x8000BF288000BF24, 0x8000BF308000BF2C,
-    0x000E95AF8000BF34, 0x000E85AD000E95AE, 0x000E95AB000E85AC,
-    0x000EADA9000E95AA, 0x000F24E1000ED1A8, 0x000E38E38000BF38,
-    0x000038E38000BF40, 0x0001A0E18000BF44, 0x0000E4EF8000BF48,
-    0x0004E4848000BF4C, 0x000410868000BF50, 0x00014CEF00049487,
-    0x00009D818000BF54, 0x0003F9B300009D80, 0x0003E9B500004982,
-    0x8000BF5C8000BF58, 0x8000BF600003E1B6, 0x00034DBB8000BF64,
-    0x00035DBD00034DBA, 0x0002F1BF00035DBC, 0x8000BECC000339BE,
-    0x8000BF688000BF3C, 0x8000BF708000BF6C, 0x8000BF788000BF74,
-    0x000625A88000BF7C, 0x000601AA000639A9, 0x000671AC000601AB,
-    0x000675AE00067DAD, 0x000619B0000669AF, 0x000625B200062DB1,
-    0x0006F5B40006EDB3, 0x0006F5B60006FDB5, 0x0006D5B80006EDB7,
-    0x0006D5BA0006DDB9, 0x000739BC00072DBB, 0x000729BE000739BD,
-    0x0006A9B3000729BF, 0x8000BF848000BF80, 0x8000BF8C8000BF88,
-    0x0006D9B50006DDB6, 0x0006C9BB8000BF90, 0x8000BF940006FDBA,
-    0x000619BF8000BF98, 0x000619BD000611BE, 0x8000BF9C0006D1BC,
-    0x8000BFA00006EDA3, 0x000699A68000BFA4, 0x000390BE8000BFA8,
-    0x0006B9AA00069DA5, 0x8000BFAC00068DAB, 0x000655AE8000BFB0,
-    0x000695AC00065DAF, 0x00068DA800065DAD, 0x000695AA000691A9,
-    0x0006BDAC0006A9AB, 0x0006CDAE0006A9AD, 0x0001E9800006C5AF,
-    0x0001FD820001E981, 0x000190868000BFB4, 0x8000BFB80001B487,
-    0x0001C5B88000BFBC, 0x0001C5BA0001CDB9, 0x0001CDBC0001DDBB,
-    0x0001F5BE0001FDBD, 0x000681B000019DBF, 0x000161B2000681B1,
-    0x000161B4000161B3, 0x000161B6000161B5, 0x00061DB3000161B7,
-    0x8000BFC48000BFC0, 0x8000BFCC8000BFC8, 0x000635B5000629B6,
-    0x000645BB8000BFD0, 0x8000BFD4000645BA, 0x0001A5BF8000BFD8,
-    0x0001B5BD0001A5BE, 0x8000BFDC0001B5BC, 0x8000BFE0000659A3,
-    0x00066DA68000BFE4, 0x8000BFEC8000BFE8, 0x000601AA000671A5,
-    0x8000BFF0000601AB, 0x0001E1AE8000BFF4, 0x0001F1AC0001E1AF,
-    0x0002EDA80001F1AD, 0x000331AA000329A9, 0x000395AC000331AB,
-    0x000395AE00039DAD, 0x00020C8400038DAF, 0x8000BFFC8000BFF8,
-    0x8000C0048000C000, 0x8000C00C8000C008, 0x0003A9B88000C010,
-    0x00006DBA0003A9B9, 0x00007DBC000065BB, 0x00006DBE000065BD,
-    0x0003FDB0000065BF, 0x0003C5B20003CDB1, 0x0003B9B40003ADB3,
-    0x0003ADB60003B9B5, 0x0000BD800003A5B7, 0x0000198200000981,
-    0x8000C0188000C014, 0x8000C0200003D8BE, 0x001C00868000C024,
-    0x8000C0280003D487, 0x8000C0308000C02C, 0x8000C0388000C034,
-    0x000369BB8000C03C, 0x8000C040000359BA, 0x00031DBF8000C044,
-    0x000319BD00031DBE, 0x0003D9B3000371BC, 0x8000C04C8000C048,
-    0x8000C0548000C050, 0x00037DB500037DB6, 0x000609AF8000C058,
-    0x0007A5AD0007C1AE, 0x0007B5AB0007A5AC, 0x8000C05C0007B5AA,
-    0x001D6C848000C060, 0x000711A5000719A6, 0x000701A38000C064,
-    0x8000C06C8000C068, 0x8000C0748000C070, 0x001D08BE8000C078,
-    0x8000C07C001C0084, 0x8000C080000338EF, 0x8000C0888000C084,
-    0x8000C090000354E3, 0x00002980000044E1, 0x0000198200002981,
-    0x000628E18000C094, 0x000630E3000794E1, 0x8000C09800078CE3,
-    0x001C38868000C09C, 0x001CE8BE001D5C87, 0x8000C0A48000C0A0,
-    0x8000C0AC8000C0A8, 0x000198EF8000C0B0, 0x0002F9A3000600EF,
-    0x8000C0B88000C0B4, 0x8000C0C08000C0BC, 0x00025DA500025DA6,
-    0x000249AB8000C0C4, 0x8000C0C8000279AA, 0x00023DAF8000C0CC,
-    0x000239AD00023DAE, 0x001E99A8000251AC, 0x001EA1AA001E99A9,
-    0x001EA1AC001EA1AB, 0x001EA1AE001EA1AD, 0x8000C08C001EA1AF,
-    0x8000C0D48000C0D0, 0x8000C0DC8000C0D8, 0x8000C0E48000C0E0,
-    0x001F5DB88000C0E8, 0x001F7DBA001F69B9, 0x001F6DBC001F75BB,
-    0x001F1DBE001F15BD, 0x001EA9B0001F0DBF, 0x001E8DB2001EA9B1,
-    0x001E9DB4001E85B3, 0x001E85B6001E85B5, 0x001E29B3001F65B7,
-    0x8000C0F08000C0EC, 0x8000C0F88000C0F4, 0x001E29B5001E21B6,
-    0x001E6DBB8000C0FC, 0x00000886001E6DBA, 0x001E41BF00002C87,
-    0x001E51BD001E59BE, 0x00004980001E59BC, 0x0000598200004981,
-    0x8000C100001E65A3, 0x001E6DA6001E65A5, 0x8000C1048000C01C,
-    0x001E21AA8000C108, 0x001E15AC001E21AB, 0x001E15AE001E1DAD,
-    0x001E69A8001E0DAF, 0x001E99AA001E79A9, 0x001EB5AC001E99AB,
-    0x001EB5AE001EBDAD, 0x8000C10C001EADAF, 0x8000C1148000C110,
-    0x8000C11C8000C118, 0x8000C1248000C120, 0x000165B88000C128,
-    0x000165BA00016DB9, 0x000165BC00017DBB, 0x000165BE00016DBD,
-    0x001ED5B00001D9BF, 0x001EA1B2001EA1B1, 0x001EA1B4001EA1B3,
-    0x001E91B6001EA9B5, 0x001E11B3001E91B7, 0x8000C1308000C12C,
-    0x8000C1388000C134, 0x001E3DB5001E3DB6, 0x001E65BB8000C13C,
-    0x8000C140001E11BA, 0x001E6DBF8000C144, 0x001E75BD001E65BE,
-    0x8000C148001E75BC, 0x8000C14C001E55A3, 0x001E79A68000C150,
-    0x8000C1588000C154, 0x001E55AA001E79A5, 0x8000C15C001E21AB,
-    0x001E21AE8000C160, 0x001E31AC001E29AF, 0x0002D1A8001E31AD,
-    0x0002EDAA0002E5A9, 0x000335AC00032DAB, 0x000335AE00033DAD,
-    0x8000C16400032DAF, 0x8000C1708000C168, 0x8000C174000260BE,
-    0x8000C17C8000C178, 0x0003E9B88000C180, 0x000389BA0003E9B9,
-    0x00039DBC000385BB, 0x000381BE000381BD, 0x000355B00003B5BF,
-    0x000355B200035DB1, 0x0003F9B40003E9B3, 0x0003E9B60003F9B5,
-    0x00036D800003E1B7, 0x0000AD820000A581, 0x8000C184000269B3,
-    0x0003B1B60003B1B5, 0x0002E0848000C188, 0x00039DBA8000C18C,
-    0x00038DBC000395BB, 0x000331BE000331BD, 0x000225A3000331BF,
-    0x8000C1948000C190, 0x8000C19C8000C198, 0x0003FDA50003FDA6,
-    0x0003D9AB8000C1A0, 0x000448860003D1AA, 0x00037DAF0003AC87,
-    0x00037DAD00037DAE, 0x8000C1A40003C1AC, 0x8000C1AC8000C1A8,
-    0x0003A0E38000C1B0, 0x0001A8E18000C1B4, 0x0003F4EF8000C1B8,
-    0x8000C1C08000C1BC, 0x8000C1C88000C1C4, 0x8000C1D08000C1CC,
-    0x000079808000C1D4, 0x00001D8200007981, 0x001E60E300024084,
-    0x001FE0E18000C1D8, 0x001ED0EF000550BE, 0x0004C8868000C1E0,
-    0x8000C1E400054087, 0x8000C1EC8000C1E8, 0x8000C1F48000C1F0,
-    0x8000C1FC8000C1F8, 0x8000C2048000C200, 0x8000C20C8000C208,
-    0x0005A484001EC4EF, 0x8000C210001E24E1, 0x8000C214001EF0E3,
-    0x8000C21C8000C218, 0x0001F5B38000C220, 0x8000C2288000C224,
-    0x8000C2308000C22C, 0x0001E5B5000145B6, 0x000129BB8000C234,
-    0x8000C238000121BA, 0x0000F1BF8000C23C, 0x000129BD000129BE,
-    0x000629A8000131BC, 0x00063DAA000635A9, 0x00062DAC000635AB,
-    0x000685AE000685AD, 0x8000C1DC0006B9AF, 0x00001D8100001D82,
-    0x8000C24000001D80, 0x8000C2488000C244, 0x00074DB8000330BE,
-    0x000755BA000751B9, 0x000779BC000769BB, 0x000769BE000779BD,
-    0x0006A5B0000769BF, 0x0006A5B20006ADB1, 0x0006ADB40006BDB3,
-    0x00077DB6000775B5, 0x0006B5A3000775B7, 0x000028868000C24C,
-    0x8000C25000034C87, 0x0006A5A5000605A6, 0x000669AB8000C254,
-    0x8000C258000661AA, 0x0007B1AF8000C25C, 0x000669AD000669AE,
-    0x8000C260000671AC, 0x0007BDB38000C264, 0x0007ADB58000C268,
-    0x8000C2708000C26C, 0x8000C16C00075DB6, 0x000779BB8000C274,
-    0x000769BD000761BA, 0x000749BF000761BC, 0x000665A8000755BE,
-    0x00067DAA000675A9, 0x00066DAC000675AB, 0x00069DAE000695AD,
-    0x8000C278000695AF, 0x8000C2808000C27C, 0x8000C2888000C284,
-    0x8000C2908000C28C, 0x0006FDB88000C294, 0x000681BA000681B9,
-    0x000681BC000681BB, 0x0006B1BE000689BD, 0x0006EDB00006B1BF,
-    0x0006FDB20006F5B1, 0x0006D9B40006F5B3, 0x0006CDB60006C5B5,
-    0x0006F5A30006C5B7, 0x8000C29C8000C298, 0x8000C2A48000C2A0,
-    0x0006E5A5000615A6, 0x000631AB8000C2A8, 0x8000C2AC000629AA,
-    0x000601AF8000C2B0, 0x000621AD00061DAE, 0x00006D80000629AC,
-    0x0000198200000981, 0x8000C2B88000C2B4, 0x000190BE00019884,
-    0x000F40868000C2C0, 0x8000C2C40000F487, 0x8000C2CC8000C2C8,
-    0x8000C2D48000C2D0, 0x000291A88000C2D8, 0x0002A1AA000299A9,
-    0x0002D1AC0002A1AB, 0x0002D5AE0002D1AD, 0x8000C2DC0002C9AF,
-    0x8000C2E48000C2E0, 0x8000C2E800032084, 0x8000C2F08000C2EC,
-    0x000379B88000C2F4, 0x0003CDBA000379B9, 0x0003DDBC0003C5BB,
-    0x0003C5BE0003C5BD, 0x0002B9B00003F5BF, 0x000349B2000281B1,
-    0x000359B4000349B3, 0x000349B6000359B5, 0x000229B3000349B7,
-    0x8000C2FC8000C2F8, 0x000C9CBE8000C300, 0x000259B5000251B6,
-    0x000245BB8000C304, 0x8000C30800027DBA, 0x00039DBF8000C30C,
-    0x00039DBD00039DBE, 0x8000C31000039DBC, 0x8000C31400026DA3,
-    0x000215A68000C318, 0x8000C3208000C31C, 0x000239AA00021DA5,
-    0x8000C324000201AB, 0x0003D9AE8000C328, 0x0003D9AC0003D9AF,
-    0x000190E10003D9AD, 0x0000CCE38000C330, 0x0001C980000D1884,
-    0x0001F5820001FD81, 0x000C2CBE8000C334, 0x8000C33C8000C338,
-    0x8000C3448000C340, 0x000018EF8000C348, 0x000CE0868000C34C,
-    0x8000C3500002F487, 0x8000C3588000C354, 0x8000C3608000C35C,
-    0x8000C3640006F8EF, 0x8000C368000794E1, 0x8000C36C000630E3,
-    0x8000C3748000C370, 0x8000C37C8000C378, 0x8000C3848000C380,
-    0x8000C38C8000C388, 0x8000C3948000C390, 0x00700085000158EF,
-    0x8000C3980006D4E1, 0x0000258000061CE3, 0x0000118200002D81,
-    0x00022DB38000C3A0, 0x000CE887000000F1, 0x8000C3A4000F6C86,
-    0x00022DB500022DB6, 0x0001C9BB8000C3A8, 0x8000C3AC0001C1BA,
-    0x0001CDBF8000C3B0, 0x0001CDBD0001CDBE, 0x000DDDA80001CDBC,
-    0x000DE5AA000DEDA9, 0x000DE5AC000DFDAB, 0x000DE5AE000DEDAD,
-    0x8000C32C000E3DAF, 0x8000C3B48000C39C, 0x8000C3BC8000C3B8,
-    0x8000C3C48000C3C0, 0x000ECDB88000C3C8, 0x000EE1BA000ED9B9,
-    0x000E91BC000EE1BB, 0x000E95BE000E9DBD, 0x000E45B0000E89BF,
-    0x000E59B2000E4DB1, 0x000E49B4000E59B3, 0x000EFDB6000E49B5,
-    0x000D6DA3000EF5B7, 0x8000C3D08000C3CC, 0x8000C3D88000C3D4,
-    0x000D6DA5000D6DA6, 0x000E89AB8000C3DC, 0x8000C3E0000E81AA,
-    0x000E8DAF8000C3E4, 0x000E8DAD000E8DAE, 0x000FEDB5000E8DAC,
-    0x8000C3EC8000C3E8, 0x8000C3F0000FFDB6, 0x000FFDB38000C3F4,
-    0x000FB5BD8000C3F8, 0x000F89BF000FB5BC, 0x8000C3FC000FB5BE,
-    0x000FB5BB8000C2BC, 0x000E05A8000FB5BA, 0x000E2DAA000E19A9,
-    0x000E3DAC000E25AB, 0x000E2DAE000E25AD, 0x00000D80000E21AF,
-    0x00001D8200001581, 0x8000C4048000C400, 0x000230848000C408,
-    0x000E35B8000130BE, 0x000E35BA000E3DB9, 0x000ED5BC000ECDBB,
-    0x000ED5BE000EDDBD, 0x000E29B0000ECDBF, 0x000E39B2000E29B1,
-    0x000E29B4000E39B3, 0x000E11B6000E15B5, 0x000EA5A3000E0DB7,
-    0x000728868000C410, 0x8000C41400014C87, 0x000EB5A5000EA5A6,
-    0x000EEDAB8000C418, 0x8000C41C000EEDAA, 0x000ED1AF8000C420,
-    0x000EEDAD000EEDAE, 0x8000C424000EEDAC, 0x8000C428000E19B3,
-    0x0001F9B68000C42C, 0x8000C4348000C430, 0x0001D5BA000E31B5,
-    0x8000C4380001DDBB, 0x000179BE8000C43C, 0x0001C5BC000179BF,
-    0x000DBDA8000179BD, 0x000DD5AA000DD5A9, 0x000DFDAC000DE5AB,
-    0x000235AE00023DAD, 0x8000C44000022DAF, 0x8000C4488000C444,
-    0x8000C4508000C44C, 0x8000C4588000C454, 0x0002E9B88000C45C,
-    0x00028DBA0002E9B9, 0x000281BC000285BB, 0x000285BE000281BD,
-    0x000255B00002ADBF, 0x000255B200025DB1, 0x0002F9B40002E9B3,
-    0x0002E9B60002F9B5, 0x000D5DA30002E1B7, 0x8000C4648000C460,
-    0x8000C46C8000C468, 0x000D75A50002BDA6, 0x000299AB8000C470,
-    0x8000C474000291AA, 0x00023DAF8000C478, 0x00023DAD00023DAE,
-    0x00005D80000281AC, 0x0000658200006D81, 0x8000C480000351B3,
-    0x000319B6000379B5, 0x00036CBE0002DC84, 0x00033DBA8000C484,
-    0x000319BC000335BB, 0x0000D9BE000319BD, 0x00029DA30000D9BF,
-    0x000568868000C488, 0x8000C48C0002BC87, 0x0002B5A50002D5A6,
-    0x0002F9AB8000C490, 0x000220840002F1AA, 0x000115AF8000C494,
-    0x0002D5AD000115AE, 0x8000C4980002D5AC, 0x000F20EF8000C49C,
-    0x8000C4A48000C4A0, 0x8000C4AC8000C4A8, 0x8000C4B48000C4B0,
-    0x8000C4BC8000C4B8, 0x000C88E18000C4C0, 0x000F44E38000C4C4,
-    0x8000C4CC8000C4C8, 0x8000C4D48000C4D0, 0x00041CBE8000C4D8,
-    0x00005CEF8000C4DC, 0x000E74E1000E44EF, 0x000080E3000EFCE1,
-    0x00000D80000EB0E3, 0x0000718200007581, 0x8000C4E48000C4E0,
-    0x8000C4E80000C1B3, 0x0000FDB68000C4F0, 0x8000C4F88000C4F4,
-    0x0000ADBA0000C1B5, 0x0004C8860000B5BB, 0x0000ADBE0004A087,
-    0x0000ADBC000095BF, 0x000585A80000A5BD, 0x000595AA000595A9,
-    0x0005BDAC0005A5AB, 0x0005D1AE0005D5AD, 0x8000C47C0005D1AF,
-    0x8000C4FC8000C4EC, 0x8000C5048000C500, 0x8000C50C8000C508,
-    0x000671B88000C510, 0x000671BA000671B9, 0x0006D5BC000671BB,
-    0x0006D5BE0006DDBD, 0x0005B5B00006CDBF, 0x000581B20005BDB1,
-    0x000651B4000581B3, 0x000651B6000651B5, 0x0007EDB8000651B7,
-    0x0007F5BA0007F1B9, 0x000715BC00070DBB, 0x000715BE00071DBD,
-    0x0007EDB0000709BF, 0x0007FDB20007F5B1, 0x0007EDB40007F5B3,
-    0x0007DDB60007D5B5, 0x000611A80007D5B7, 0x0007CDAA000621A9,
-    0x0007CDAC0007DDAB, 0x000799AE0007F9AD, 0x8000C514000799AF,
-    0x8000C51C8000C518, 0x8000C5248000C520, 0x8000C40C8000C528,
-    0x000605A38000C52C, 0x8000C5348000C530, 0x8000C53C8000C538,
-    0x000605A5000639A6, 0x000671AB8000C540, 0x8000C544000669AA,
-    0x000651AF8000C548, 0x000661AD000669AE, 0x8000C54C000669AC,
-    0x00001D8100001D82, 0x8000C55000001D80, 0x8000C5588000C554,
-    0x0001AC84000194BE, 0x0001E8878000C55C, 0x8000C560007C0086,
-    0x8000C5688000C564, 0x00064DA88000C56C, 0x000651AA000651A9,
-    0x0001BDAC0001A5AB, 0x0001ADAE0001A5AD, 0x8000C5700001A5AF,
-    0x8000C5788000C574, 0x8000C5808000C57C, 0x8000C5888000C584,
-    0x0000C5B88000C58C, 0x0000C5BA0000CDB9, 0x0000CDBC0000DDBB,
-    0x0000F5BE0000FDBD, 0x0001DDB000009DBF, 0x00010DB2000105B1,
-    0x00011DB4000105B3, 0x00010DB6000105B5, 0x0005D9B3000101B7,
-    0x8000C5948000C590, 0x8000C59C8000C598, 0x000235B5000225B6,
-    0x0002C1BB8000C5A0, 0x8000C5A4000219BA, 0x0002C1BF8000C5A8,
-    0x0002D1BD0002D9BE, 0x8000C5AC0002D9BC, 0x8000C5B000059DA3,
-    0x000261A68000C5B4, 0x8000C5BC8000C5B8, 0x00025DAA000271A5,
-    0x8000C5C0000285AB, 0x00029DAE8000C5C4, 0x00029DAC000285AF,
-    0x0000E980000295AD, 0x000005820000E981, 0x0003DCBE8000C5C8,
-    0x8000C5D000035884, 0x007C00868000C5D4, 0x8000C5D80003EC87,
-    0x8000C5E08000C5DC, 0x8000C5E88000C5E4, 0x8000C5F08000C5EC,
-    0x8000C5F88000C5F4, 0x8000C6008000C5FC, 0x0002B0EF8000C604,
-    0x000194E18000C608, 0x00026CE38000C60C, 0x8000C6148000C610,
-    0x8000C61C8000C618, 0x000279B38000C620, 0x8000C6288000C624,
-    0x8000C6308000C62C, 0x0002B9B50002B1B6, 0x000295BB8000C634,
-    0x8000C638000295BA, 0x000139BF007D6484, 0x000285BD000131BE,
-    0x00006982000285BC, 0x0000598000075CEF, 0x8000C64000005181,
-    0x007D68878000C644, 0x007D2CBE007CFC86, 0x8000C64C8000C648,
-    0x007C3C848000C650, 0x8000C6540007B8E3, 0x8000C6580004ECE1,
-    0x8000C6608000C65C, 0x8000C664007FC084, 0x00003CEF8000C668,
-    0x000650E10006F4EF, 0x0000D8E30006B4E1, 0x8000C66C000694E3,
-    0x8000C6748000C670, 0x0002B9A38000C678, 0x8000C6808000C67C,
-    0x8000C6888000C684, 0x000279A5000271A6, 0x000255AB8000C68C,
-    0x8000C690000255AA, 0x0001F9AF8000C694, 0x000245AD0001F1AE,
-    0x007E99A8000245AC, 0x007EA1AA007E99A9, 0x007EA1AC007EA1AB,
-    0x007EA1AE007EA1AD, 0x8000C698007EA1AF, 0x8000C6A08000C69C,
-    0x000019808000C6A4, 0x0000058200001981, 0x007EADB88000C6A8,
-    0x007EBDBA007EB5B9, 0x007EADBC007EB5BB, 0x007F51BE007F51BD,
-    0x007EA1B0007F51BF, 0x007EA1B2007EA1B1, 0x007EA1B4007EA1B3,
-    0x007EA5B6007EA1B5, 0x8000C5CC007E99B7, 0x0003C0BE007E2DB3,
-    0x007E25B68000C63C, 0x8000C6B08000C6AC, 0x007E61BA007E2DB5,
-    0x00000886007E61BB, 0x007E61BE00009C87, 0x007E71BC007E61BF,
-    0x007E69A3007E71BD, 0x8000C6B88000C6B4, 0x8000C6C08000C6BC,
-    0x007E69A5007E61A6, 0x007E25AB8000C6C4, 0x8000C6C8007E25AA,
-    0x007E25AF8000C6CC, 0x007E35AD007E25AE, 0x007E51A8007E35AC,
-    0x007E6DAA007E51A9, 0x007E7DAC007E65AB, 0x007E99AE007E6DAD,
-    0x8000C6D0007E99AF, 0x8000C6D88000C6D4, 0x8000C6E08000C6DC,
-    0x8000C6E88000C6E4, 0x007E89B88000C6EC, 0x007E99BA007E89B9,
-    0x007EB9BC007E91BB, 0x000179BE007EB9BD, 0x007EE9B0000179BF,
-    0x007EFDB2007EF5B1, 0x007EC5B4007EDDB3, 0x007EC5B6007ECDB5,
-    0x007EE5B3007EB9B7, 0x8000C6F48000C6F0, 0x8000C6FC8000C6F8,
-    0x007EF5B5007EEDB6, 0x007E31BB8000C700, 0x8000C704007E2DBA,
-    0x007E09BF8000C708, 0x007E09BD007E01BE, 0x8000C70C007E29BC,
-    0x8000C710007EA1A3, 0x007EA9A68000C714, 0x000190BE8000C718,
-    0x007E69AA007EB1A5, 0x0007EC85007E75AB, 0x007E45AE8000C720,
-    0x007E6DAC007E4DAF, 0x000299A8007E4DAD, 0x0002C1AA0002C1A9,
-    0x0002C1AC0002C1AB, 0x0002F1AE0002C9AD, 0x000584860002F1AF,
-    0x8000C72400037887, 0x0003F9808000C728, 0x0003ED820003F981,
-    0x0003C5B88000C72C, 0x0003C5BA0003CDB9, 0x0003CDBC0003DDBB,
-    0x0003F5BE0003FDBD, 0x000341B000039DBF, 0x000341B2000341B1,
-    0x000341B4000341B3, 0x000341B6000341B5, 0x000249B3000341B7,
-    0x8000C7348000C730, 0x8000C73C8000C738, 0x000249B5000249B6,
-    0x000385BB8000C740, 0x8000C744000385BA, 0x000385BF8000C748,
-    0x000395BD000385BE, 0x8000C74C000395BC, 0x8000C75000020DA3,
-    0x00020DA68000C754, 0x8000C75C8000C758, 0x0003C1AA00020DA5,
-    0x8000C7600003C1AB, 0x0003C1AE8000C764, 0x0003D1AC0003C1AF,
-    0x007E60E30003D1AD, 0x007E4CE1007FBCE1, 0x0001A0E1007E30E3,
-    0x007E84E30000E8E1, 0x0007ACBE0017C4E3, 0x000008EF8000C768,
-    0x000140EF001C7CEF, 0x8000C7708000C76C, 0x8000C774007EA4EF,
-    0x000404848000C778, 0x8000C77C0006C0BE, 0x00003D8100003D82,
-    0x8000C78000003D80, 0x8000C7888000C784, 0x000750868000C790,
-    0x8000C79400058487, 0x007DC1BB8000C798, 0x007DA1B9007DA9BA,
-    0x007DC9BF007D91B8, 0x007DD5BD007DC1BE, 0x007E1DB3007DD5BC,
-    0x007E1DB1007E1DB2, 0x007DA1B7007E1DB0, 0x007DB1B5007DB9B6,
-    0x007EBDAB007DB9B4, 0x007EBDA9007EBDAA, 0x007E7DAF007EBDA8,
-    0x007E7DAD007E7DAE, 0x8000C79C007E7DAC, 0x8000C7A48000C7A0,
-    0x8000C7AC8000C7A8, 0x8000C7B48000C7B0, 0x006C05820006B0BE,
-    0x8000C71C006C1983, 0x0068F9868000C78C, 0x006819840074C987,
-    0x0074018A00681985, 0x8000C7B80074B98B, 0x0070B58E8000C7BC,
-    0x0070D98C007CC18F, 0x007C619200701D8D, 0x8000C7C0007C6193,
-    0x007875968000C7C4, 0x0078A19400793197, 0x0004E19A00787D95,
-    0x8000C7C80004E99B, 0x8000C7D08000C7CC, 0x0004ED9C8000C7D4,
-    0x000089A18000C7D8, 0x000C01A38000C7DC, 0x000C15A5000081A2,
-    0x0008B9A7000CB5A4, 0x0009F9A9000801A6, 0x001425AB000881A8,
-    0x001001AD00142DAA, 0x001199AF001421AC, 0x001CE1B100109DAE,
-    0x001801B3001C01B0, 0x0018F5B5001C65B2, 0x8000C7E00018F5B4,
-    0x8000C7E88000C7E4, 0x8000C7F08000C7EC, 0x000019808000C7F4,
-    0x0000058200001981, 0x000729A98000C7F8, 0x000735AB0004E9A8,
-    0x000799AD000721AA, 0x000789AF000791AC, 0x007845A0000791AE,
-    0x0005A5A20005A5A1, 0x0001B8840005B9A3, 0x000034878000C800,
-    0x0007ADB9000F9C86, 0x00067DBB0007ADB8, 0x000611BD000669BA,
-    0x000609BF00066DBC, 0x0007C9B1000619BE, 0x0007A9B30007F9B0,
-    0x0007BDB50007C1B2, 0x0007A1B70007BDB4, 0x0007ADB30007A9B6,
-    0x8000C8088000C804, 0x8000C80C00000C84, 0x00071DB500071DB6,
-    0x000709BB006D998F, 0x8000C810000739BA, 0x00077DBF8000C814,
-    0x000779BD00077DBE, 0x0071E198000711BC, 0x007D3D9A00714199,
-    0x007DF59C007DED9B, 0x0079F99E00792D9D, 0x006DFD900079F99F,
-    0x0069ED920069E591, 0x00752D9400690593, 0x007565960075D595,
-    0x00042DA300711D97, 0x000DC0BE000DC884, 0x8000C81C8000C818,
-    0x00049DA500049DA6, 0x000489AB8000C820, 0x8000C8240004B9AA,
-    0x0004FDAF8000C828, 0x0004F9AD0004FDAE, 0x8000C82C000491AC,
-    0x8000C830000539B3, 0x0005B9B68000C834, 0x8000C83C8000C838,
-    0x000595BA0005B9B5, 0x8000C84000059DBB, 0x000239BE8000C844,
-    0x000251BC000239BF, 0x0005EDA8000251BD, 0x0005F5AA0005FDA9,
-    0x000575AC00056DAB, 0x000575AE00057DAD, 0x8000C84800056DAF,
-    0x8000C8508000C84C, 0x8000C8588000C854, 0x8000C8608000C85C,
-    0x0002B1B88000C864, 0x0002B5BA0002B1B9, 0x000209BC000249BB,
-    0x000239BE000209BD, 0x000515B0000239BF, 0x0002C1B200051DB1,
-    0x0002B9B40002C1B3, 0x000291B60002B9B5, 0x00047DA3000291B7,
-    0x0001B58100001582, 0x8000C8680001B580, 0x0004FDA50004FDA6,
-    0x0004D9AB8000C870, 0x000D28860004D1AA, 0x00037DAF000D1C87,
-    0x000315AD00037DAE, 0x8000C874000315AC, 0x8000C87C8000C878,
-    0x8000C8800002C5B3, 0x0002F5B60002FDB5, 0x8000C8888000C884,
-    0x000175BA8000C88C, 0x00015DBC000145BB, 0x000131BE000135BD,
-    0x0000A0E3000131BF, 0x0006E0E10006DCE3, 0x8000C8900007B4E1,
-    0x8000C8988000C894, 0x000C8086000F90BE, 0x8000C89C000C1887,
-    0x000604EF8000C8A0, 0x8000C8A400066CEF, 0x000249A38000C8A8,
-    0x8000C8B08000C8AC, 0x8000C8B88000C8B4, 0x000271A5000279A6,
-    0x0001C9AB8000C8BC, 0x8000C8C00001F9AA, 0x0001BDAF8000C8C4,
-    0x0001B9AD0001BDAE, 0x000D35A80001D1AC, 0x000D81AA000D81A9,
-    0x000DB9AC000D9DAB, 0x000E15AE000DB9AD, 0x8000C86C000E0DAF,
-    0x8000C8CC8000C8C8, 0x00001D808000C8D0, 0x0000B98200000981,
-    0x000E29B88000C8D4, 0x000EE9BA000E29B9, 0x000EF9BC000EE9BB,
-    0x000EE9BE000EF9BD, 0x000E7DB0000EE9BF, 0x000E45B2000E4DB1,
-    0x000E39B4000E2DB3, 0x000E2DB6000E39B5, 0x8000C8D8000E25B7,
-    0x8000C7FC8000C8DC, 0x8000C8E48000C8E0, 0x8000C8EC8000C8E8,
-    0x000FBDA88000C8F0, 0x000FEDAA000FE5A9, 0x000FE1AC000FE5AB,
-    0x000FD9AE000FE9AD, 0x000F51B0000FD9AF, 0x000F51B2000F51B1,
-    0x000F7DB4000F65B3, 0x000F65B6000F6DB5, 0x000F29B8000F19B7,
-    0x000F39BA000F29B9, 0x000F19BC000F31BB, 0x0000DDBE000F19BD,
-    0x000E45B30000D5BF, 0x8000C8F88000C8F4, 0x8000C9008000C8FC,
-    0x000E49B5000E49B6, 0x000EB5BB8000C904, 0x8000C908000EB5BA,
-    0x000E89BF8000C90C, 0x000E91BD000E99BE, 0x8000C910000EA5BC,
-    0x8000C914000E01A3, 0x000E0DA68000C918, 0x8000C9208000C91C,
-    0x000EF1AA000E0DA5, 0x8000C924000EF1AB, 0x000EDDAE8000C928,
-    0x000EE1AC000ECDAF, 0x00010DA8000ED5AD, 0x00011DAA000115A9,
-    0x00010DAC000115AB, 0x000091AE000091AD, 0x8000C92C000091AF,
-    0x8000C9348000C930, 0x8000C93C8000C938, 0x0003B1810003B982,
-    0x000099B80003B980, 0x0000A5BA0000ADB9, 0x000075BC00006DBB,
-    0x000075BE00007DBD, 0x0000F1B000006DBF, 0x0000C1B20000F9B1,
-    0x0000B1B40000C1B3, 0x0000B5B60000BDB5, 0x0002AC840000A9B7,
-    0x8000C9448000C940, 0x0004B0868000C948, 0x8000C9500002D887,
-    0x8000C9588000C954, 0x8000C9608000C95C, 0x000374BE8000C964,
-    0x8000C96C8000C968, 0x00035DBB8000C970, 0x8000C974000355BA,
-    0x000319BF8000C978, 0x000375BD000315BE, 0x0002F5B3000375BC,
-    0x8000C9808000C97C, 0x8000C9888000C984, 0x0002E5B5000371B6,
-    0x000215A38000C98C, 0x8000C9948000C990, 0x8000C99C8000C998,
-    0x000205A5000391A6, 0x0003BDAB8000C9A0, 0x8000C9A40003B5AA,
-    0x0003F9AF8000C9A8, 0x000395AD0003F5AE, 0x8000C9AC000395AC,
-    0x00001D8100001D82, 0x8000C9B000001D80, 0x8000C9B88000C9B4,
-    0x00036C848000C9C0, 0x000378878000C9C4, 0x8000C9C80004EC86,
-    0x8000C9D08000C9CC, 0x000EB8E18000C9D4, 0x000318E3000C98E3,
-    0x8000C9D8000C28E1, 0x8000C9E08000C9DC, 0x0004EC848000C9E4,
-    0x8000C9EC8000C9E8, 0x8000C9F48000C9F0, 0x000E5CEF000A30EF,
-    0x000161B38000C9F8, 0x8000CA008000C9FC, 0x8000CA088000CA04,
-    0x000171B5000179B6, 0x000129BB8000CA0C, 0x8000CA10000159BA,
-    0x00011DBF8000CA14, 0x00011DBD00011DBE, 0x8000CA18000131BC,
-    0x8000CA208000CA1C, 0x000019808000CA24, 0x0000398200001981,
-    0x8000CA28000F6CEF, 0x000460BE000E94E1, 0x8000CA30000FB8E3,
-    0x8000CA388000CA34, 0x0004E086000000EF, 0x0001E9A300048C87,
-    0x0001F9A500070084, 0x8000CA408000CA3C, 0x8000CA440001F1A6,
-    0x0001A1AB8000CA48, 0x000195AD0001D1AA, 0x000195AF0001B9AC,
-    0x000619A8000195AE, 0x000629AA000619A9, 0x000639AC000629AB,
-    0x00062DAE000639AD, 0x8000C9BC000625AF, 0x8000CA4C8000CA2C,
-    0x8000CA548000CA50, 0x8000CA5C8000CA58, 0x0006C5B88000CA60,
-    0x0006C5BA0006CDB9, 0x0006CDBC0006DDBB, 0x0006F5BE0006FDBD,
-    0x00065DB000069DBF, 0x000625B200062DB1, 0x000619B4000609B3,
-    0x000609B6000619B5, 0x8000CA64000601B7, 0x8000CA6C8000CA68,
-    0x8000CA748000CA70, 0x8000CA7C8000CA78, 0x00067DA88000C94C,
-    0x000659AA000641A9, 0x000649AC000659AB, 0x0007A9AE000649AD,
-    0x0007D9B00007A9AF, 0x0007EDB20007D9B1, 0x0007FDB40007E5B3,
-    0x0007E5B60007E5B5, 0x0007A9B8000799B7, 0x0007B9BA0007A9B9,
-    0x000799BC0007B1BB, 0x000779BE000799BD, 0x0007F1B3000779BF,
-    0x8000CA848000CA80, 0x8000CA8C8000CA88, 0x0007E1B50007E9B6,
-    0x00063DBB8000CA90, 0x8000CA9400063DBA, 0x000601BF8000CA98,
-    0x000611BD000619BE, 0x00006D82000619BC, 0x000055800007B5A3,
-    0x0007ADA600006581, 0x8000CA9C00019CBE, 0x000679AA0007A5A5,
-    0x001C0086000679AB, 0x00065DAE0001CC87, 0x00065DAC000645AF,
-    0x000631A8000655AD, 0x000631AA000631A9, 0x000649AC000659AB,
-    0x00067DAE000675AD, 0x8000CAA00001B5AF, 0x8000CAA88000CAA4,
-    0x8000CAB08000CAAC, 0x8000CAB88000CAB4, 0x00014DB88000CABC,
-    0x00015DBA000155B9, 0x0000D5BC000155BB, 0x0000D5BE0000DDBD,
-    0x0001CDB00000CDBF, 0x0001DDB20001D5B1, 0x0001CDB40001D5B3,
-    0x00017DB6000175B5, 0x0005B9B3000175B7, 0x8000CAC48000CAC0,
-    0x8000CACC8000CAC8, 0x0005C1B50005F5B6, 0x000211BB8000CAD0,
-    0x8000CAD4000209BA, 0x0002B1BF8000CAD8, 0x000201BD000209BE,
-    0x8000CADC000209BC, 0x8000CAE00005FDA3, 0x0005B1A68000CAE4,
-    0x8000CAEC8000CAE8, 0x00024DAA000585A5, 0x8000CAF0000255AB,
-    0x00024DAE8000CAF4, 0x00024DAC0002F5AF, 0x8000CAF8000245AD,
-    0x8000CB008000CAFC, 0x8000CB088000CB04, 0x8000CB100003D8BE,
-    0x000029808000CB14, 0x0000398200002981, 0x001C108600032084,
-    0x8000CB1800038487, 0x8000CB208000CB1C, 0x8000CB288000CB24,
-    0x8000CB308000CB2C, 0x8000CB34000268E3, 0x8000CB380001ACE1,
-    0x8000CB3C000288EF, 0x8000CB448000CB40, 0x8000CB4C8000CB48,
-    0x8000CB548000CB50, 0x8000CB5C8000CB58, 0x0007E0E38000CB60,
-    0x000710E18000CB64, 0x0007E8EF8000CB68, 0x8000CB708000CB6C,
-    0x000019808000CB74, 0x0000058200001981, 0x0002E9B38000CB78,
-    0x001CE886001D0C84, 0x8000CB80001D8087, 0x0002E9B50002E1B6,
-    0x000169BB8000CB84, 0x001D24BE000161BA, 0x000159BF8000CB88,
-    0x000169BD000169BE, 0x8000CB8C000171BC, 0x8000CB948000CB90,
-    0x8000CB9C8000CB98, 0x000034EF8000CBA0, 0x000678E10006CCEF,
-    0x000674E3000750E1, 0x8000CBA400069CE3, 0x8000CBAC8000CBA8,
-    0x000269A38000CBB0, 0x001C84848000CBB4, 0x8000CBB8001C84BE,
-    0x000269A5000261A6, 0x0001E9AB8000CBBC, 0x8000CBC00001E1AA,
-    0x0001D9AF8000CBC4, 0x0001E9AD0001E9AE, 0x001E99A80001F1AC,
-    0x001EB1AA001E99A9, 0x001ED1AC001EB1AB, 0x001ED1AE001ED1AD,
-    0x8000CBC8001ED1AF, 0x8000CBD08000CBCC, 0x8000CBD88000CBD4,
-    0x8000CBE08000CBDC, 0x001EBDB88000CBE4, 0x001F4DBA001F45B9,
-    0x001F5DBC001F45BB, 0x001F45BE001F41BD, 0x001EB1B0001F79BF,
-    0x001E89B2001EBDB1, 0x001E9DB4001E85B3, 0x001E8DB6001E85B5,
-    0x8000CB0C001E85B7, 0x8000CBE8001E3DB3, 0x001E55B68000CB7C,
-    0x8000CBF08000CBEC, 0x001E71BA001E5DB5, 0x00000886001E7DBB,
-    0x001E5DBE00002887, 0x001E59BC001E45BF, 0x001E79A3001E55BD,
-    0x0000F58100004D82, 0x8000CBF40000F580, 0x001E19A5001E11A6,
-    0x001E39AB8000CBF8, 0x8000CBFC001E35AA, 0x001E01AF8000CC00,
-    0x001E11AD001E19AE, 0x001E51A8001E1DAC, 0x001E65AA001E59A9,
-    0x001E95AC001E8DAB, 0x001E95AE001E9DAD, 0x8000CC04001E8DAF,
-    0x8000CC0C8000CC08, 0x8000CC148000CC10, 0x8000CC1C8000CC18,
-    0x001EA9B88000CC20, 0x000169BA001EA9B9, 0x000179BC000169BB,
-    0x000169BE000179BD, 0x001EF5B0000169BF, 0x001EDDB2001ED5B1,
-    0x001ECDB4001ED5B3, 0x001EA9B6001EB1B5, 0x001EEDB3001EA1B7,
-    0x8000CC288000CC24, 0x8000CC308000CC2C, 0x001E1DB5001E1DB6,
-    0x001E11BB8000CC34, 0x8000CC38001E31BA, 0x001E75BF8000CC3C,
-    0x001E01BD001E09BE, 0x8000CC40001E09BC, 0x8000CC44001EA9A3,
-    0x001E59A68000CC48, 0x8000CC508000CC4C, 0x001E75AA001E59A5,
-    0x8000CC54001E55AB, 0x001E4DAE8000CC58, 0x001E4DAC001E31AF,
-    0x000299A8001E45AD, 0x0002D5AA0002D5A9, 0x0002FDAC0002E5AB,
-    0x00033DAE000335AD, 0x00026CBE000335AF, 0x8000CC648000CC60,
-    0x8000CC6C8000CC68, 0x8000CC748000CC70, 0x0003D1B88000CC78,
-    0x0003E1BA0003D9B9, 0x000391BC0003E1BB, 0x000395BE00039DBD,
-    0x00034DB0000389BF, 0x00035DB2000355B1, 0x00034DB4000355B3,
-    0x0003F1B60003F1B5, 0x000369810003F1B7, 0x000255B300035D80,
-    0x00027DB500036182, 0x8000CC808000CC7C, 0x0002E4840003BDB6,
-    0x000399BB8000CC84, 0x000389BD000381BA, 0x00033DBF000381BC,
-    0x00021DA3000381BE, 0x000398878000CC88, 0x8000CC8C00046C86,
-    0x000235A50003F5A6, 0x0003D1AB8000CC90, 0x8000CC940003C9AA,
-    0x000375AF8000CC98, 0x0003C1AD0003C9AE, 0x001ED8E30003C9AC,
-    0x001E58E1001FACE1, 0x0001A0E1001E58E3, 0x0003A8E38000CC9C,
-    0x8000CCA48000CCA0, 0x0003F0EF8000CCA8, 0x001E04EF8000CCAC,
-    0x8000CCB48000CCB0, 0x8000CCB8001EACEF, 0x00051C848000CCBC,
-    0x00000D80000500BE, 0x00003D8200003581, 0x8000CCC48000CCC0,
-    0x0004C8868000CCC8, 0x8000CCD00005A087, 0x8000CCD88000CCD4,
-    0x001DD5BA8000CCDC, 0x001DC1B8001DDDBB, 0x001D11BE001DC1B9,
-    0x001DF1BC001D11BF, 0x001DADB2001DF1BD, 0x001DADB0001DB5B3,
-    0x001D9DB6001DA5B1, 0x001DA5B4001DF5B7, 0x001E69AA001D95B5,
-    0x001E69A8001E69AB, 0x001E69AE001E69A9, 0x001E69AC001E69AF,
-    0x8000CCE0001E69AD, 0x8000CCE88000CCE4, 0x8000CCF08000CCEC,
-    0x8000CCF88000CCF4, 0x000625A88000CCFC, 0x000625AA00062DA9,
-    0x000625AC00063DAB, 0x000625AE00062DAD, 0x8000CCCC000689AF,
-    0x8000CD048000CD00, 0x8000CD0C8000CD08, 0x8000CD148000CD10,
-    0x00075DB88000CD18, 0x00076DBA000765B9, 0x000765BC000761BB,
-    0x000719BE00076DBD, 0x0006F9B0000719BF, 0x0006ADB20006C1B1,
-    0x0006ADB40006BDB3, 0x000695B600069DB5, 0x0006B5B5000769B7,
-    0x8000CD208000CD1C, 0x8000CD240006B5B6, 0x0006B5B38000CD28,
-    0x000671BD8000CD2C, 0x000641BF000661BC, 0x8000CC5C000671BE,
-    0x00061DBB8000CD30, 0x0006EDA3000625BA, 0x0000198100001982,
-    0x8000CD3400007980, 0x0006EDA50006EDA6, 0x000645AB8000CD38,
-    0x0003A48400067DAA, 0x000619AF0003A4BE, 0x000629AD000629AE,
-    0x000671A8000639AC, 0x000671AA000671A9, 0x0006F1AC000671AB,
-    0x0006F1AE0006F1AD, 0x8000CD3C0006F1AF, 0x000188868000CD40,
-    0x8000CD440001AC87, 0x8000CD4C8000CD48, 0x000175B88000CD50,
-    0x000175BA00017DB9, 0x0001D9BC0001C9BB, 0x0001CDBE0001C5BD,
-    0x000691B00001C1BF, 0x000691B2000691B1, 0x000155B4000691B3,
-    0x000155B600015DB5, 0x0006F9B300014DB7, 0x8000CD588000CD54,
-    0x8000CD608000CD5C, 0x000679B5000671B6, 0x000655BB8000CD64,
-    0x8000CD68000655BA, 0x000199BF8000CD6C, 0x000645BD000191BE,
-    0x8000CD70000645BC, 0x8000CD740006BDA3, 0x000635A68000CD78,
-    0x8000CD808000CD7C, 0x000611AA00063DA5, 0x8000CD84000611AB,
-    0x0001D5AE8000CD88, 0x000601AC0001DDAF, 0x000325A8000601AD,
-    0x00033DAA000335A9, 0x00032DAC000335AB, 0x000391AE000391AD,
-    0x8000CD8C000391AF, 0x8000CD948000CD90, 0x8000CD9C8000CD98,
-    0x8000CDA48000CDA0, 0x000399B88000CDA8, 0x0003A5BA0003ADB9,
-    0x000075BC00006DBB, 0x000075BE00007DBD, 0x0003F1B000006DBF,
-    0x0003C1B20003F9B1, 0x0003B1B40003C1B3, 0x0003B5B60003BDB5,
-    0x0000B5800003A9B7, 0x0000B5820000BD81, 0x00031087000348BE,
-    0x8000CDB0000C3486, 0x000CACBE0002E8EF, 0x0002A484000194E1,
-    0x8000CDB4000214E3, 0x8000CDBC8000CDB8, 0x0003F5B38000CDC0,
-    0x8000CDC88000CDC4, 0x8000CDCC000A7485, 0x0003E5B5000351B6,
-    0x00037DBB8000CDD0, 0x8000CDD4000375BA, 0x000339BF8000CDD8,
-    0x000355BD000335BE, 0x8000CDDC000355BC, 0x8000CDE48000CDE0,
-    0x8000CDE80006B8EF, 0x8000CDF08000CDEC, 0x8000CDF88000CDF4,
-    0x8000CE008000CDFC, 0x000000CC8000CE04, 0x8000CE080007E0E1,
-    0x8000CE0C000664E3, 0x00027DA38000CE10, 0x00026DA500024084,
-    0x8000CE188000CE14, 0x000CC0860002D9A6, 0x0002F5AB000DAC87,
-    0x0002DDAD0002FDAA, 0x0002B1AF0002DDAC, 0x0000C5800002BDAE,
-    0x0000D1820000D181, 0x8000CE248000CE20, 0x8000CE2C8000CE28,
-    0x000CEC84000188EF, 0x8000CE300006D4E1, 0x8000CE34000154E3,
-    0x8000CE3C8000CE38, 0x000175B38000CE40, 0x8000CE488000CE44,
-    0x8000CE508000CE4C, 0x000165B5000111B6, 0x000109BB8000CE54,
-    0x8000CE58000139BA, 0x0000F9BF8000CE5C, 0x0000F9BD0000F1BE,
-    0x000E99A8000111BC, 0x000EBDAA000EB5A9, 0x000EADAC000EB5AB,
-    0x000ED1AE000ED1AD, 0x8000CE1C000ECDAF, 0x8000CE648000CE60,
-    0x8000CE6C8000CE68, 0x8000CE748000CE70, 0x000F59B88000CE78,
-    0x000F69BA000F59B9, 0x000F79BC000F69BB, 0x000F61BE000F79BD,
-    0x000EBDB0000F1DBF, 0x000E85B2000E8DB1, 0x000F75B4000F6DB3,
-    0x000F69B6000F7DB5, 0x000E35A3000F69B7, 0x8000CE808000CE7C,
-    0x8000CE888000CE84, 0x000E25A5000E51A6, 0x000E49AB0003D4BE,
-    0x8000CE8C000E79AA, 0x000FB9AF8000CE90, 0x000FB9AD000FB1AE,
-    0x00005182000E51AC, 0x000041808000CDAC, 0x0003188700004181,
-    0x8000CE9400007486, 0x8000CE9C8000CE98, 0x8000CEA48000CEA0,
-    0x8000CEAC8000CEA8, 0x8000CEB48000CEB0, 0x000E8DA88000CEB8,
-    0x000E9DAA000E95A9, 0x000EB9AC000E95AB, 0x000EADAE000EA5AD,
-    0x8000CEBC000EA5AF, 0x8000CEC48000CEC0, 0x8000CECC8000CEC8,
-    0x8000CED48000CED0, 0x000165B88000CED8, 0x000165BA00016DB9,
-    0x000165BC00017DBB, 0x000165BE00016DBD, 0x000EDDB00001D9BF,
-    0x000EA1B2000EA1B1, 0x000EA1B4000EA1B3, 0x000E91B6000EA9B5,
-    0x000E15B3000E91B7, 0x8000CEE08000CEDC, 0x8000CEE88000CEE4,
-    0x000E31B5000E25B6, 0x000E65BB8000CEEC, 0x8000CEF0000E19BA,
-    0x000E6DBF8000CEF4, 0x000E75BD000E65BE, 0x00001582000E75BC,
-    0x00006180000E51A3, 0x000E61A600006181, 0x000190BE8000CEF8,
-    0x000E5DAA000E75A5, 0x0001A484000E21AB, 0x000E21AE8000CF00,
-    0x000E31AC000E29AF, 0x0002ADA8000E31AD, 0x0002D5AA0002DDA9,
-    0x0002F5AC0002EDAB, 0x0002F5AE0002FDAD, 0x8000CF040002EDAF,
-    0x8000CF0C8000CF08, 0x8000CF148000CF10, 0x8000CF1C8000CF18,
-    0x000375B88000CF20, 0x000375BA00037DB9, 0x0003D9BC0003C9BB,
-    0x0003C9BE0003D9BD, 0x000295B00003C1BF, 0x000295B200029DB1,
-    0x000355B400034DB3, 0x000355B600035DB5, 0x00025DB300034DB7,
-    0x0003D8878000CF24, 0x8000CF2800042C86, 0x00027DB500027DB6,
-    0x000259BB8000CF2C, 0x8000CF30000251BA, 0x000399BF8000CF34,
-    0x000399BD000391BE, 0x8000CF38000241BC, 0x00056886000219A3,
-    0x000239A600030087, 0x8000CF408000CF3C, 0x000215AA000239A5,
-    0x8000CF4400021DAB, 0x0003D5AE8000CF48, 0x000205AC0003DDAF,
-    0x8000CF4C0003DDAD, 0x8000CF548000CF50, 0x8000CF5C8000CF58,
-    0x8000CF60000034EF, 0x00052CBE8000CF64, 0x8000CF680000DCE3,
-    0x8000CF700001E0E1, 0x8000CF788000CF74, 0x000FCCE18000CF7C,
-    0x000E14E300024884, 0x0000498000054484, 0x0000718200004981,
-    0x8000CF848000CF80, 0x8000CF8C8000CF88, 0x8000CF948000CF90,
-    0x000F20EF8000CF98, 0x8000CFA08000CF9C, 0x8000CFA88000CFA4,
-    0x8000CFB08000CFAC, 0x8000CFB88000CFB4, 0x0004ACBE0001F8EF,
-    0x8000CFBC000EDCE1, 0x8000CFC0000E1CE3, 0x8000CFC88000CFC4,
-    0x0001E5B38000CFCC, 0x0000158100002D82, 0x8000CFD000001D80,
-    0x0001F5B50001E5B6, 0x000165BB8000CFD4, 0x8000CFD80001D9BA,
-    0x00015DBF8000CFE0, 0x000165BD00015DBE, 0x0005EDA800016DBC,
-    0x00062DAA000625A9, 0x00063DAC000625AB, 0x00062DAE000625AD,
-    0x8000CF6C000625AF, 0x000088868000CFDC, 0x8000CFE40003A087,
-    0x8000CFEC8000CFE8, 0x000699B88000CFF0, 0x0006A5BA0006ADB9,
-    0x000755BC00074DBB, 0x000755BE00075DBD, 0x00065DB000074DBF,
-    0x0006F9B20006E1B1, 0x0006D9B40006F1B3, 0x0006B9B60006D9B5,
-    0x0006A5A30006A9B7, 0x8000CFF88000CFF4, 0x8000D0008000CFFC,
-    0x0006B5A50006A5A6, 0x000625AB8000D004, 0x8000D008000699AA,
-    0x00061DAF8000D00C, 0x000625AD00061DAE, 0x8000D01000062DAC,
-    0x8000CEFC0007FDB3, 0x0007B1B68000D014, 0x8000D01C8000D018,
-    0x00078DBA000781B5, 0x8000D020000795BB, 0x000771BE8000D024,
-    0x00078DBC000771BF, 0x000631A8000781BD, 0x000631AA000631A9,
-    0x00064DAC000631AB, 0x00067DAE000675AD, 0x8000D028000675AF,
-    0x8000D0308000D02C, 0x8000D0388000D034, 0x8000D0408000D03C,
-    0x0006D1B88000D044, 0x0006D1BA0006D1B9, 0x0006E1BC0006E5BB,
-    0x0006E1BE0006E1BD, 0x00060DB00006E1BF, 0x00061DB2000615B1,
-    0x0006F1B4000611B3, 0x0006F1B60006F1B5, 0x0006B9A30006F1B7,
-    0x0000B58100001582, 0x8000D0480000B580, 0x0006C5A50006F5A6,
-    0x0006D1AB000190BE, 0x0002A0840006C9AA, 0x000635AF003C00BF,
-    0x0006C5AD000635AE, 0x000084850006C9AC, 0x003F08860001F1B3,
-    0x000155B600009C87, 0x8000D0548000D050, 0x000171BA0001E1B5,
-    0x8000D058000179BB, 0x000131BE8000D05C, 0x000151BC000139BF,
-    0x0005EDA8000151BD, 0x00023DAA000235A9, 0x00022DAC000235AB,
-    0x000291AE000291AD, 0x8000D060000291AF, 0x8000D0688000D064,
-    0x000211A28000D06C, 0x000691A0000211A3, 0x0002A9B80002F5A1,
-    0x000369BA0002A9B9, 0x000379BC000369BB, 0x000369BE000379BD,
-    0x0002F1B0000369BF, 0x0002C1B20002F9B1, 0x0002B1B40002C1B3,
-    0x0002ADB60002B9B5, 0x0002B5A30002A5B7, 0x8000D070001A8D8F,
-    0x8000D0788000D074, 0x0002A5A5000211A6, 0x00023DAB8000D07C,
-    0x8000D080000235AA, 0x00027DAF8000D084, 0x000215AD000275AE,
-    0x0006059E000215AC, 0x000AE99C0006E59F, 0x000E459A000BED9D,
-    0x000E5998000A659B, 0x0012C596000E5199, 0x001275940013BD97,
-    0x0016D1920012CD95, 0x001AF59000167193, 0x8000D08800167591,
-    0x003D44840000D0E1, 0x8000D08C002410E3, 0x8000D0948000D090,
-    0x8000D09C8000D098, 0x8000D0A48000D0A0, 0x8000D0AC8000D0A8,
-    0x8000D0B48000D0B0, 0x8000D0B80024CCEF, 0x003D00BE000610E3,
-    0x000065800007D8E1, 0x0000198200006D81, 0x003CC0868000D0C0,
-    0x8000D0C4003C4487, 0x8000D0CC8000D0C8, 0x8000D0D00006ACEF,
-    0x002B85B18000D0D4, 0x002AA1B3002801B0, 0x002791B5002B9DB2,
-    0x0000FCEF002401B4, 0x000190E10006C4EF, 0x000010E3000774E1,
-    0x8000D0D80006F0E3, 0x003CF4848000D0DC, 0x003F95A18000D0E0,
-    0x003F21A38000D0E4, 0x003BD5A5003F49A2, 0x003401A7003BD5A4,
-    0x003741A9003B01A6, 0x0033E9AB0037E1A8, 0x002C01AD003001AA,
-    0x002FA1AF0033C1AC, 0x003E89B0002FF9AE, 0x003E99B2003E89B1,
-    0x003EB9B4003E91B3, 0x003F19B6003EB9B5, 0x003F35B8003F0DB7,
-    0x003F35BA003F3DB9, 0x003F15BC003F0DBB, 0x003F15BE003F1DBD,
-    0x8000D0E8003F0DBF, 0x8000D0F08000D0EC, 0x8000D0F88000D0F4,
-    0x8000D0FC8000D04C, 0x003DB5A88000D100, 0x003D81AA003DBDA9,
-    0x003E51AC003D81AB, 0x003E51AE003E51AD, 0x003D69B3003E51AF,
-    0x8000D1048000D0BC, 0x8000D10C8000D108, 0x003EB9B5003EB1B6,
-    0x003E2DBB8000D110, 0x8000D114003E2DBA, 0x003E1DBF8000D118,
-    0x003E25BD003E1DBE, 0x8000D11C003E2DBC, 0x8000D120003D2DA3,
-    0x003EF5A68000D124, 0x000390BE8000D128, 0x003E69AA003EFDA5,
-    0x8000D12C003E69AB, 0x003E59AE8000D130, 0x003E69AC003E59AF,
-    0x003E5DA8003E61AD, 0x003E65AA003E6DA9, 0x003E69AC003E7DAB,
-    0x003EA9AE003E69AD, 0x0001E980003EA9AF, 0x0001FD820001E981,
-    0x000190868000D134, 0x8000D1380001B487, 0x00011DB88000D13C,
-    0x0001C5BA0001CDB9, 0x0001CDBC0001DDBB, 0x0001F5BE0001FDBD,
-    0x003EDDB000019DBF, 0x000129B2003EA1B1, 0x000139B4000129B3,
-    0x000129B6000139B5, 0x003EEDB3000125B7, 0x8000D1448000D140,
-    0x8000D14C8000D148, 0x003EE1B5003EE9B6, 0x003E65BB8000D150,
-    0x8000D154003E19BA, 0x0001FDBF0001A084, 0x0001FDBD0001FDBE,
-    0x8000D1580001FDBC, 0x8000D15C003EA9A3, 0x003EADA68000D160,
-    0x8000D1688000D164, 0x003E5DAA003EA5A5, 0x8000D16C003E21AB,
-    0x0001B9AE8000D170, 0x0001B9AC0001B9AF, 0x0002EDA80001B9AD,
-    0x000331AA000329A9, 0x000395AC000331AB, 0x000395AE00039DAD,
-    0x000400BE00038DAF, 0x8000D1788000D174, 0x8000D1808000D17C,
-    0x8000D1888000D184, 0x0003ADB88000D18C, 0x000061BA000061B9,
-    0x000061BC000061BB, 0x000061BE000061BD, 0x0003FDB0000061BF,
-    0x0003A9B20003CDB1, 0x0003BDB40003A5B3, 0x0003A1B60003A1B5,
-    0x00000D80000395B7, 0x00001D8200001581, 0x8000D1948000D190,
-    0x8000D1A08000D198, 0x000420868000D1A4, 0x8000D1A80003E487,
-    0x8000D1B08000D1AC, 0x8000D1B88000D1B4, 0x0003F9B38000D1BC,
-    0x8000D1C48000D1C0, 0x8000D1CC8000D1C8, 0x00035DB500035DB6,
-    0x000349BB8000D1D0, 0x8000D1D4000379BA, 0x000325BF8000D1D8,
-    0x000331BD000339BE, 0x003E78E100033DBC, 0x003EA0E38000D1DC,
-    0x8000D1E48000D1E0, 0x8000D1EC8000D1E8, 0x00052C848000D1F0,
-    0x0004A4BE8000D1F4, 0x8000D1FC8000D1F8, 0x000118EF8000D200,
-    0x000075808000D204, 0x0000198200007981, 0x8000D2080004C484,
-    0x0005948700038CEF, 0x8000D21000047C86, 0x8000D2140000D8E3,
-    0x8000D218000044E1, 0x8000D2208000D21C, 0x002FA58B8000D224,
-    0x8000D228002FA58A, 0x0033058F8000D22C, 0x0033198D0033198E,
-    0x0027AD83002EC58C, 0x8000D2300027AD82, 0x002F2D878000D234,
-    0x002B9985002B3186, 0x003FE19B002B3584, 0x8000D238003F0D9A,
-    0x003C00E18000D23C, 0x003EC0E38000D240, 0x0037C193003FF99C,
-    0x003E58EF00375592, 0x003B59978000D244, 0x003B3D95003BD596,
-    0x8000D19C0036DD94, 0x8000D2488000D20C, 0x8000D2508000D24C,
-    0x8000D2588000D254, 0x0006B9A88000D25C, 0x000681AA0006B9A9,
-    0x000681AC000681AB, 0x000681AE000681AD, 0x000685B0000681AF,
-    0x000685B200068DB1, 0x00068DB400069DB3, 0x0006B5B60006BDB5,
-    0x00076DB8000759B7, 0x00077DBA000775B9, 0x00076DBC000775BB,
-    0x000755BE000751BD, 0x8000D26000074DBF, 0x000515AA8000D264,
-    0x000551AC000525AB, 0x000555AE00055DAD, 0x8000D268000549AF,
-    0x8000D2708000D26C, 0x8000D274000595A3, 0x000531A6000531A5,
-    0x000669B38000D278, 0x0000798100007982, 0x00005CBE00006580,
-    0x000669B5000661B6, 0x000629BB8000D27C, 0x0000C886000621BA,
-    0x000619BF0000AC87, 0x000629BD000629BE, 0x0007A9A3000631BC,
-    0x8000D2848000D280, 0x8000D28C8000D288, 0x0007A9A50007A1A6,
-    0x0007E9AB8000D290, 0x8000D2940007E1AA, 0x0007D9AF8000D298,
-    0x0007E9AD0007E9AE, 0x8000D29C0007F1AC, 0x8000D2A0000635B3,
-    0x00062DB68000D2A4, 0x8000D2AC8000D2A8, 0x0006C9BA000625B5,
-    0x8000D2B00006C9BB, 0x0006A5BE8000D2B4, 0x0006B1BC000699BF,
-    0x0006E1A80006B1BD, 0x0006F9AA0006EDA9, 0x0006E9AC0006F9AB,
-    0x000611AE0006E9AD, 0x8000D2B8000611AF, 0x8000D2C08000D2BC,
-    0x8000D2C88000D2C4, 0x8000D2D08000D2CC, 0x0001D1B88000D2D4,
-    0x0001D1BA0001D1B9, 0x000151BC0001D1BB, 0x000179BE000151BD,
-    0x000649B0000179BF, 0x000621B2000649B1, 0x000625B4000621B3,
-    0x0001F1B6000619B5, 0x000571A30001F1B7, 0x8000D2DC8000D2D8,
-    0x8000D2E00002CCBE, 0x000561A5000569A6, 0x00058DAB8000D2E4,
-    0x0003E48400058DAA, 0x0005DDAF8000D2E8, 0x0005F5AD0005E1AE,
-    0x000051800005F5AC, 0x0000658200005981, 0x8000D2F0000341A3,
-    0x000371A6000349A5, 0x000C40868000D2F4, 0x000321AA00031487,
-    0x000325AC000321AB, 0x000311AE00032DAD, 0x000385A8000311AF,
-    0x000385AA00038DA9, 0x00038DAC00039DAB, 0x0003B5AE0003BDAD,
-    0x8000D2F80003DDAF, 0x8000D3008000D2FC, 0x8000D3088000D304,
-    0x8000D3108000D30C, 0x000369B88000D314, 0x000379BA000369B9,
-    0x000369BC000379BB, 0x0000DDBE000369BD, 0x0003A9B00000D5BF,
-    0x0003BDB20003A9B1, 0x000399B40003B5B3, 0x000359B6000399B5,
-    0x00020DB3000359B7, 0x8000D31C8000D318, 0x8000D3248000D320,
-    0x000205B500023DB6, 0x00026DBB8000D328, 0x8000D32C00026DBA,
-    0x00025DBF8000D330, 0x000261BD00025DBE, 0x8000D334000269BC,
-    0x8000D33C8000D338, 0x000368E38000D340, 0x0001A8E18000D344,
-    0x000394EF8000D348, 0x8000D3508000D34C, 0x000019808000D354,
-    0x0000058200001981, 0x0006C4E18000D358, 0x0006D4E300073CE1,
-    0x000290840006BCE3, 0x000D14BE8000D360, 0x8000D3688000D364,
-    0x8000D3708000D36C, 0x8000D3788000D374, 0x000190EF8000D37C,
-    0x000C608600066CEF, 0x8000D380000DAC87, 0x8000D38400016DB3,
-    0x000175B600017DB5, 0x8000D38C8000D388, 0x000151BA8000D390,
-    0x0000F5BC000151BB, 0x0000F5BE0000FDBD, 0x000EA1A30000EDBF,
-    0x8000D3948000D35C, 0x8000D39C8000D398, 0x000EB1A5000EB9A6,
-    0x000E9DAB8000D3A0, 0x8000D3A4000E9DAA, 0x000F21AF8000D3A8,
-    0x000F31AD000F39AE, 0x8000D3AC000F39AC, 0x8000D3B48000D3B0,
-    0x8000D3BC8000D3B8, 0x8000D3C0000DD4E3, 0x00001582000C24E1,
-    0x00003180000D04EF, 0x8000D3C400001D81, 0x000330BE8000D3C8,
-    0x000ED1A800033084, 0x000E29AA000ED9A9, 0x000E39AC000E29AB,
-    0x000E29AE000E39AD, 0x8000D3CC000E29AF, 0x00000C860000F884,
-    0x8000D3D000008C87, 0x8000D3D88000D3D4, 0x000FEDB88000D3DC,
-    0x000F8DBA000F85B9, 0x000F9DBC000F85BB, 0x000F85BE000F81BD,
-    0x000E59B0000FB9BF, 0x000FEDB2000E59B1, 0x000FFDB4000FE5B3,
-    0x000FE5B6000FE5B5, 0x000E95A8000FD5B7, 0x000EA9AA000E99A9,
-    0x000EB9AC000EA9AB, 0x000EA9AE000EB9AD, 0x8000D3E0000EA9AF,
-    0x8000D3E88000D3E4, 0x8000D3F08000D3EC, 0x8000D3F88000D3F4,
-    0x00015DB88000D3FC, 0x000165BA00016DB9, 0x000115BC00010DBB,
-    0x000115BE00011DBD, 0x000EA9B000010DBF, 0x000EB5B2000EB5B1,
-    0x000E95B4000E8DB3, 0x000175B600017DB5, 0x000E21B300016DB7,
-    0x8000D4048000D400, 0x8000D40C8000D408, 0x000E31B5000E39B6,
-    0x000E71BB8000D410, 0x8000D414000E6DBA, 0x0001B9BF8000D418,
-    0x000E59BD000E51BE, 0x8000D41C000E61BC, 0x8000D420000E65A3,
-    0x000E7DA68000D424, 0x8000D42C8000D428, 0x000E29AA000E75A5,
-    0x8000D430000E35AB, 0x000E15AE8000D434, 0x000E25AC0001FDAF,
-    0x0002B5A9000E1DAD, 0x0002B5AB0002B5A8, 0x000335AD0002B5AA,
-    0x000335AF000335AC, 0x0003C980000335AE, 0x000005820003C981,
-    0x000264BE8000D438, 0x000378878000D440, 0x0003E9B900041C86,
-    0x000389BB0003D9B8, 0x00039DBD0003E1BA, 0x000381BF00039DBC,
-    0x000355B1000389BE, 0x000355B3000355B0, 0x0003F9B5000355B2,
-    0x0003E9B70003F1B4, 0x000300840003F1B6, 0x0003A9B50003A1B6,
-    0x000229B30005C084, 0x8000D4488000D444, 0x000329BF8000D44C,
-    0x000395BD000321BE, 0x000385BB000395BC, 0x8000D450000385BA,
-    0x0003E1A58000D454, 0x8000D45C8000D458, 0x8000D4600003E9A6,
-    0x000261A38000D464, 0x0003DDAD8000D468, 0x000361AF0003DDAC,
-    0x8000D46C000369AE, 0x0003CDAB8000D470, 0x8000D4740003CDAA,
-    0x8000D47C8000D478, 0x8000D4848000D480, 0x8000D48C8000D488,
-    0x8000D4900003D4EF, 0x8000D494000194E1, 0x8000D498000198E3,
-    0x8000D4A08000D49C, 0x000079808000D4A4, 0x0000198200007981,
-    0x8000D4A80001D5B3, 0x0001E5B60001EDB5, 0x0004C086000214BE,
-    0x0001B9BA00054087, 0x0001ADBC0001B9BB, 0x00018DBE000185BD,
-    0x000E58E100007DBF, 0x0001E4E38000D4B0, 0x8000D4B88000D4B4,
-    0x8000D4C08000D4BC, 0x0004E8BE8000D4C4, 0x8000D4CC8000D4C8,
-    0x8000D4D48000D4D0, 0x000138EF8000D4D8, 0x8000D4E08000D4DC,
-    0x8000D4E400015DA3, 0x00016DA68000D4E8, 0x8000D2EC8000D4EC,
-    0x000131AA000165A5, 0x8000D4F0000131AB, 0x000105AE8000D4F4,
-    0x000125AC0000F5AF, 0x8000D4AC00010DAD, 0x8000D4FC8000D4F8,
-    0x8000D5048000D500, 0x8000D50C8000D508, 0x000009818000D510,
-    0x8000D51400003980, 0x8000D51800001D82, 0x8000D5208000D51C,
-    0x0000E0868000D524, 0x8000D52800037487, 0x0005F4EF8000D52C,
-    0x8000D5348000D530, 0x0004F0E38000D538, 0x00044CE18000D53C,
-    0x8000D5448000D540, 0x8000D54C8000D548, 0x0006B5A88000D550,
-    0x000611AA000611A9, 0x000629AC000611AB, 0x00065DAE000629AD,
-    0x0000E484000655AF, 0x8000D5588000D554, 0x8000D5608000D55C,
-    0x8000D5688000D564, 0x0007CDB88000D56C, 0x0007DDBA0007D5B9,
-    0x0007F1BC0007D5BB, 0x0007E9BE0007FDBD, 0x000631B00007E9BF,
-    0x000635B2000631B1, 0x000605B400061DB3, 0x0007FDB60007F5B5,
-    0x000685A80007F5B7, 0x000699AA000689A9, 0x000689AC000699AB,
-    0x0006BDAE000689AD, 0x8000D5700006B5AF, 0x8000D5788000D574,
-    0x00000D808000D57C, 0x0000B1820000B181, 0x0006A9B88000D580,
-    0x00014DBA0006A9B9, 0x00015DBC000145BB, 0x00014DBE000145BD,
-    0x0006CDB0000145BF, 0x0006DDB20006D5B1, 0x0006B1B40006D1B3,
-    0x0006B5B60006BDB5, 0x000611B30006A9B7, 0x8000D5888000D584,
-    0x8000D5900001CCBE, 0x000601B5000609B6, 0x000611BB8000D594,
-    0x00000886000609BA, 0x000679BF00009C87, 0x000605BD000675BE,
-    0x8000D598000609BC, 0x8000D59C000655A3, 0x00064DA68000D5A0,
-    0x8000D5A88000D5A4, 0x00064DAA000645A5, 0x8000D5AC000655AB,
-    0x000631AE8000D5B0, 0x00064DAC00063DAF, 0x000239A8000641AD,
-    0x000241AA000239A9, 0x000245AC000241AB, 0x000245AE00024DAD,
-    0x001CACBE000279AF, 0x8000D5B4001DECBE, 0x8000D5BC8000D5B8,
-    0x8000D5C48000D5C0, 0x0002F1B88000D5C8, 0x0002F1BA0002F1B9,
-    0x000371BC0002F1BB, 0x000371BE000371BD, 0x00020DB0000371BF,
-    0x00021DB2000215B1, 0x00020DB4000215B3, 0x0002D1B60002D1B5,
-    0x8000D5CC0002D1B7, 0x0002C1B38000D5D0, 0x0002C5B58000D5D4,
-    0x8000D5DC8000D5D8, 0x8000D5E00002C5B6, 0x0002D9BB8000D5E4,
-    0x0002C9BD0002C1BA, 0x00021DBF0002C1BC, 0x8000D5E80002C1BE,
-    0x8000D5F08000D5EC, 0x8000D5F4000285A3, 0x000281A6000281A5,
-    0x0003EC848000D5F8, 0x000285AA8000D600, 0x000285AC00029DAB,
-    0x000285AE00028DAD, 0x000291A8000259AF, 0x0002C9AA000299A9,
-    0x0002DDAC0002C5AB, 0x0002C1AE0002C1AD, 0x001D00860002F5AF,
-    0x8000D604001D6487, 0x0001F9808000D608, 0x0001E5820001E581,
-    0x0001C5B88000D60C, 0x0001DDBA0001C9B9, 0x0001F1BC0001D5BB,
-    0x000199BE0001F1BD, 0x00028DB0000199BF, 0x000145B200014DB1,
-    0x000145B400015DB3, 0x000145B600014DB5, 0x8000D6100001FDB7,
-    0x0000FCEF0006FCE1, 0x8000D614000710E3, 0x8000D61C8000D618,
-    0x001C6C848000D620, 0x8000D6288000D624, 0x0001E8E18000D62C,
-    0x000108E38000D630, 0x000201B30006E4EF, 0x8000D6388000D634,
-    0x8000D6408000D63C, 0x000215B5000201B6, 0x0001C9BB8000D644,
-    0x8000D648000209BA, 0x0001C9BF8000D43C, 0x0001C9BD0001C1BE,
-    0x001DC1A30001D1BC, 0x8000D64C8000D5FC, 0x8000D6548000D650,
-    0x001DD5A5001DC1A6, 0x001E09AB8000D658, 0x8000D65C001DC9AA,
-    0x001E09AF8000D660, 0x001E09AD001E01AE, 0x00006D80001E11AC,
-    0x00001D8200001581, 0x8000D664001C94EF, 0x8000D66C8000D668,
-    0x000340868000D670, 0x8000D6740003C487, 0x8000D67C8000D678,
-    0x8000D680001FC4E1, 0x001E65A8001C8CE3, 0x001FA5AA001FADA9,
-    0x001F11AC001FB1AB, 0x001F0DAE001F05AD, 0x8000D684001F05AF,
-    0x8000D68C8000D688, 0x8000D6948000D690, 0x8000D69C8000D698,
-    0x001F11B88000D6A0, 0x001F21BA001F19B9, 0x0000DDBC0000C5BB,
-    0x0000C5BE0000C5BD, 0x001F7DB00000F5BF, 0x001F41B2001F41B1,
-    0x001F45B4001F41B3, 0x001F31B6001F31B5, 0x001E2DB3001F31B7,
-    0x8000D6A88000D6A4, 0x8000D6B08000D6AC, 0x001E99B5001E85B6,
-    0x001E85BB8000D6B4, 0x8000D6B8001EB9BA, 0x001EF5BF8000D6BC,
-    0x001E95BD001E81BE, 0x8000D6C0001E95BC, 0x8000D6C4001E69A3,
-    0x001EC1A68000D6C8, 0x8000D6D08000D6CC, 0x001EFDAA001EDDA5,
-    0x8000D6D4001EC1AB, 0x001EC5AE8000D6D8, 0x001ED1AC001EB1AF,
-    0x0001DDA8001ED1AD, 0x0001E5AA0001EDA9, 0x0001E9AC0001FDAB,
-    0x00013DAE000135AD, 0x0000F980000135AF, 0x0000C5820000CD81,
-    0x000178BE00016084, 0x0000AC878000D6E0, 0x0000D1B800019C86,
-    0x0000E1BA0000D9B9, 0x000091BC0000E1BB, 0x000095BE00009DBD,
-    0x00014DB0000089BF, 0x00015DB2000155B1, 0x00014DB4000155B3,
-    0x0000F1B60000F1B5, 0x00026DB30000F1B7, 0x8000D6E88000D6E4,
-    0x8000D6F08000D6EC, 0x000261B50002B5B6, 0x000291BB8000D6F4,
-    0x8000D6F8000289BA, 0x000335BF8000D6FC, 0x000281BD000289BE,
-    0x00000090000289BC, 0x8000D700000229A3, 0x0002F1A68000D704,
-    0x8000D70C8000D708, 0x0002CDAA000225A5, 0x8000D7100002D5AB,
-    0x0002CDAE8000D714, 0x0002CDAC000371AF, 0x000335A80002C5AD,
-    0x000081AA000081A9, 0x000081AC000081AB, 0x0000B1AE000089AD,
-    0x8000D7180000B1AF, 0x8000D7208000D71C, 0x8000D7288000D724,
-    0x8000D7308000D72C, 0x000069B88000D734, 0x000079BA000069B9,
-    0x000069BC000079BB, 0x0001D9BE000069BD, 0x0000A9B00001D9BF,
-    0x0000BDB20000A9B1, 0x000099B40000B5B3, 0x000059B6000099B5,
-    0x8000D738000059B7, 0x8000D7408000D73C, 0x00001D818000D744,
-    0x8000D74800001D80, 0x0002AC8400000582, 0x0002C0870005E8BE,
-    0x8000D75000050C86, 0x8000D7588000D754, 0x000149B38000D75C,
-    0x000000D28000D760, 0x8000D7688000D764, 0x000149B5000149B6,
-    0x00012DBB8000D76C, 0x8000D77000012DBA, 0x00011DBF8000D774,
-    0x000121BD00011DBE, 0x001EB8E1000129BC, 0x001E38E38000D778,
-    0x8000D7808000D77C, 0x8000D7888000D784, 0x00046C848000D78C,
-    0x8000D79000046CBE, 0x8000D7988000D794, 0x001F64EF8000D79C,
-    0x8000D7A000077084, 0x0004E8868000D7A4, 0x8000D7A80004C087,
-    0x8000D7AC000068E3, 0x00001D820001A4E1, 0x00001D800000ACEF,
-    0x8000D58C00001581, 0x8000D7B48000D7B0, 0x0007A4E18000D7B8,
-    0x000538E38000D74C, 0x8000D7BC8000D7C0, 0x8000D7C88000D7C4,
-    0x8000D7D08000D7CC, 0x8000D7D88000D7D4, 0x8000D7E08000D7DC,
-    0x0005E4EF8000D7E4, 0x000649A38000D7E8, 0x8000D7F08000D7EC,
-    0x8000D7F88000D7F4, 0x000649A5000649A6, 0x00062DAB8000D7FC,
-    0x8000D80000062DAA, 0x00061DAF8000D804, 0x000621AD00061DAE,
-    0x0007B1A8000629AC, 0x0007B1AA0007B1A9, 0x0007D5AC0007B1AB,
-    0x0007C9AE0007D9AD, 0x8000D8080007C1AF, 0x8000D8108000D80C,
-    0x8000D8188000D814, 0x8000D8208000D81C, 0x000769B88000D824,
-    0x000779BA000769B9, 0x000711BC000779BB, 0x000709BE000711BD,
-    0x000789B0000705BF, 0x00076DB2000789B1, 0x000761B4000761B3,
-    0x000761B6000761B5, 0x000639B3000761B7, 0x0001B58100001582,
-    0x8000D8280001B580, 0x00065DB5000659B6, 0x00064DBB0001D4BE,
-    0x8000D82C00064DBA, 0x0006A5BF8000D830, 0x0006B5BD0006A5BE,
-    0x8000D8340006B9BC, 0x0000E88600067DA3, 0x00061DA60001CC87,
-    0x8000D83C8000D838, 0x000609AA000619A5, 0x8000D840000609AB,
-    0x0006E1AE8000D844, 0x0006FDAC0006E1AF, 0x00018DA80006F1AD,
-    0x000191AA000191A9, 0x0001B1AC000191AB, 0x0001B1AE0001B1AD,
-    0x0001A8840001B1AF, 0x8000D84C8000D848, 0x8000D8548000D850,
-    0x8000D85C8000D858, 0x0001A9B88000D860, 0x000169BA0001A9B9,
-    0x000179BC000169BB, 0x000169BE000179BD, 0x0001D1B0000169BF,
-    0x0001D5B20001D1B1, 0x0001B9B40001A9B3, 0x0001A9B60001B9B5,
-    0x000219B30001A1B7, 0x8000D8688000D864, 0x8000D8708000D86C,
-    0x000239B5000231B6, 0x000215BB8000D874, 0x8000D878000215BA,
-    0x00027DBF8000D87C, 0x00027DBD00027DBE, 0x8000D880000201BC,
-    0x8000D88400025DA3, 0x000275A68000D888, 0x8000D8908000D88C,
-    0x000251AA00027DA5, 0x8000D894000251AB, 0x000239AE8000D898,
-    0x000245AC000239AF, 0x0000D980000239AD, 0x0000E5820000ED81,
-    0x8000D8A00003C0BE, 0x000354870002E0EF, 0x8000D8A4000CBC86,
-    0x8000D8A8000114E3, 0x8000D8AC0001FCE1, 0x8000D8B48000D8B0,
-    0x8000D8BC8000D8B8, 0x8000D8C0000391B3, 0x0003BDB6000C4084,
-    0x8000D8C88000D8C4, 0x00034DBA0003B5B5, 0x8000D8CC000355BB,
-    0x00034DBE8000D8D0, 0x00034DBC0000F1BF, 0x00022DA8000345BD,
-    0x00025DAA000245A9, 0x000279AC000255AB, 0x0002BDAE000279AD,
-    0x000CACBE0002B5AF, 0x8000D8D88000D8D4, 0x8000D8E08000D8DC,
-    0x8000D8E88000D8E4, 0x000151B88000D8EC, 0x000161BA000159B9,
-    0x000111BC000161BB, 0x000115BE00011DBD, 0x0002CDB0000109BF,
-    0x0002DDB20002D5B1, 0x0002CDB40002D5B3, 0x000171B6000171B5,
-    0x0006B4E1000171B7, 0x000638E38000D6DC, 0x8000D8F48000D8F0,
-    0x8000D8F8000644E1, 0x8000D8FC000630E3, 0x8000D9048000D900,
-    0x000F3CBE000048EF, 0x0006E4EF8000D908, 0x000295A38000D910,
-    0x0000618100007182, 0x8000D91400006980, 0x0002B1A50002B9A6,
-    0x000251AB8000D918, 0x000CC886000249AA, 0x0001F5AF000D5C87,
-    0x000241AD000249AE, 0x8000D90C000249AC, 0x8000D9208000D91C,
-    0x000C94E18000D924, 0x000CC4E38000D928, 0x8000D9308000D92C,
-    0x000DF0EF8000D934, 0x8000D93C8000D938, 0x8000D9448000D940,
-    0x000EE9B38000D948, 0x8000D9508000D94C, 0x8000D9588000D954,
-    0x000EEDB5000EEDB6, 0x000EB5BB8000D95C, 0x8000D960000EADBA,
-    0x000F55BF8000D964, 0x000E91BD000E99BE, 0x000E11A8000EA5BC,
-    0x000E21AA000E25A9, 0x000E25AC000E3DAB, 0x000E25AE000E2DAD,
-    0x8000D968000E5DAF, 0x8000D9708000D96C, 0x000019808000D974,
-    0x0000058200001981, 0x000FF1B88000D978, 0x000FF1BA000FF1B9,
-    0x000F5DBC000FF1BB, 0x000F41BE000F45BD, 0x000E29B0000F7DBF,
-    0x000E31B2000E31B1, 0x000E1DB4000E05B3, 0x000FD1B6000FD1B5,
-    0x000FA9A3000FD1B7, 0x0001188500028C84, 0x00017485000004BF,
-    0x000FADA5000FADA6, 0x000FF5AB8000D980, 0x00078886000FEDAA,
-    0x000E15AF00005887, 0x000FD1AD000FD9AE, 0x8000D984000FE5AC,
-    0x8000D988000E99B3, 0x00015DB68000D98C, 0x8000D9948000D990,
-    0x000141BA000155B5, 0x8000D998000141BB, 0x000145BE8000D99C,
-    0x000155BC00014DBF, 0x000E3DA800014DBD, 0x000E55AA000E55A9,
-    0x000199AC000189AB, 0x00018DAE000199AD, 0x8000D9A0000181AF,
-    0x8000D9A88000D9A4, 0x8000D9B08000D9AC, 0x8000D9B88000D9B4,
-    0x000189B88000D9BC, 0x000199BA000189B9, 0x0001B9BC000191BB,
-    0x000179BE0001B9BD, 0x0001C1B0000179BF, 0x0001C5B20001C1B1,
-    0x0001C5B40001DDB3, 0x0001C5B60001CDB5, 0x000DDDA30001B9B7,
-    0x8000D9C48000D9C0, 0x8000D9CC8000D9C8, 0x000211A5000219A6,
-    0x000205AB8000D9D0, 0x8000D9D4000205AA, 0x000209AF001A798F,
-    0x000209AD000201AE, 0x001E9D94000211AC, 0x00124D9600119D95,
-    0x001AD5900012ED97, 0x001E5992001ABD91, 0x000A659C001E5993,
-    0x000DED9E000A3D9D, 0x00159D98000E699F, 0x00164D9A00167599,
-    0x000184E100098D9B, 0x03FEDCE38000D9D8, 0x000E81A000038084,
-    0x00010DA2000FBDA1, 0x03E801B100029DA3, 0x03EA69B303F649B0,
-    0x03EE91B503EA8DB2, 0x03EC60EF03EC01B4, 0x03FED1A98000D9DC,
-    0x03F2CDAB03FEC1A8, 0x03F391AD03FFBDAA, 0x03F6F9AF03F2DDAC,
-    0x000401A103F401AE, 0x000679A38000D9E0, 0x03FAD1A50006CDA2,
-    0x03FC01A703F801A4, 0x001B558203FAC1A6, 0x8000D9E4001B8183,
-    0x001F15868000D9E8, 0x001BF184001FC987, 0x0010F98A001F1D85,
-    0x000524BE0013218B, 0x0017318E8000D9F0, 0x0013998C0017018F,
-    0x0008B1920014D18D, 0x00046886000B4193, 0x000F559600057C87,
-    0x000B0194000F4597, 0x000E499A000CD995, 0x8000D9F40000F19B,
-    0x000079818000D9F8, 0x0003799C00019180, 0x8000D9FC00007182,
-    0x00058084000E74E1, 0x8000DA00000098E3, 0x8000DA048000D89C,
-    0x8000DA0C8000DA08, 0x8000DA148000DA10, 0x8000DA1C8000DA18,
-    0x8000DA248000DA20, 0x000289B3000EF4EF, 0x8000DA2C8000DA28,
-    0x8000DA348000DA30, 0x00028DB500028DB6, 0x000245BB8000DA38,
-    0x8000DA3C000245BA, 0x000245BF8000DA40, 0x000255BD000245BE,
-    0x00051DA8000255BC, 0x000541AA000545A9, 0x000545AC000541AB,
-    0x000571AE00054DAD, 0x8000D9EC000571AF, 0x8000DA488000DA44,
-    0x00000D808000DA4C, 0x0000B1820000B181, 0x000645B88000DA50,
-    0x000645BA00064DB9, 0x00064DBC00065DBB, 0x0006DDBE00067DBD,
-    0x0005C1B00006E1BF, 0x0005C1B20005C1B1, 0x0005C1B40005C1B3,
-    0x0005C1B60005C1B5, 0x0005C9A30005C1B7, 0x8000DA588000DA54,
-    0x8000DA5C000300BE, 0x0005CDA50005CDA6, 0x000505AB8000DA60,
-    0x00000886000505AA, 0x000505AF0003DC87, 0x000515AD000505AE,
-    0x8000DA64000515AC, 0x8000DA680007A9B3, 0x000749B68000DA6C,
-    0x8000DA748000DA70, 0x000745BA00074DB5, 0x8000DA7800074DBB,
-    0x000745BE8000DA7C, 0x000755BC00074DBF, 0x00066DA300074DBD,
-    0x8000DA848000DA80, 0x8000DA8C8000DA88, 0x000689A500068DA6,
-    0x000689AB8000DA90, 0x8000DA94000681AA, 0x000689AF8000DA98,
-    0x000689AD000681AE, 0x8000DA9C000691AC, 0x8000DAA48000DAA0,
-    0x8000DAAC8000DAA8, 0x8000DAB48000DAB0, 0x8000DAB800048CEF,
-    0x8000DABC000654E1, 0x00000D800005F0E3, 0x0000DD820000D581,
-    0x0006ADA88000DAC0, 0x0006E5AA0006EDA9, 0x000601AC00060DAB,
-    0x000605AE00060DAD, 0x8000DAC400063DAF, 0x0001E4848000DAC8,
-    0x03FF10860001E8BE, 0x8000DAD000009487, 0x00060DB88000DAD4,
-    0x000615BA000615B9, 0x00063DBC000625BB, 0x0001F1BE0001F1BD,
-    0x000645B00001F1BF, 0x000645B200064DB1, 0x000645B400065DB3,
-    0x00063DB600064DB5, 0x00029DA8000635B7, 0x0002C1AA0002C5A9,
-    0x0002C5AC0002C1AB, 0x0002F1AE0002CDAD, 0x8000DAD80002F1AF,
-    0x8000DAE08000DADC, 0x01FE39808000DAE4, 0x8000DAEC8000DAE8,
-    0x000305B88000DAF0, 0x000305BA00030DB9, 0x000305BC00031DBB,
-    0x000339BE00030DBD, 0x000365B0000335BF, 0x000365B200036DB1,
-    0x00036DB400037DB3, 0x000355B600035DB5, 0x000249B300033DB7,
-    0x8000DAF88000DAF4, 0x8000DB008000DAFC, 0x00024DB500024DB6,
-    0x0003A9BB8000DB04, 0x8000DB080003A1BA, 0x000399BF8000DB0C,
-    0x0003A9BD0003A9BE, 0x8000DB100003B1BC, 0x000201A38000DB14,
-    0x000205A58000DB18, 0x8000DB208000DB1C, 0x8000D97C000205A6,
-    0x0003E1AB8000DB24, 0x0003E1AD0003E9AA, 0x0003D1AF0003F9AC,
-    0x8000DB280003E1AE, 0x8000DB308000DB2C, 0x8000DB388000DB34,
-    0x000244BE0001A4E1, 0x000029800000A4E3, 0x0000398200002981,
-    0x8000DB40000078EF, 0x8000DB440002B884, 0x000161B38000DB48,
-    0x8000DB508000DB4C, 0x8000DB588000DB54, 0x000171B5000179B6,
-    0x0001E5BB8000DB5C, 0x03FD4886000159BA, 0x0001EDBF03FD4087,
-    0x0001F5BD0001E5BE, 0x8000DB600001F5BC, 0x8000DB688000DB64,
-    0x8000DB6C03FC4084, 0x8000DB748000DB70, 0x8000DB780001E8EF,
-    0x03F3B4AC0006F4E1, 0x8000DB7C000634E3, 0x03FCB0848000DB80,
-    0x8000DB888000DB84, 0x0001E1A50001E9A6, 0x0001F1A38000DB8C,
-    0x8000DB948000DB90, 0x00017DAF8000DB98, 0x000165AD000175AE,
-    0x000175AB000165AC, 0x8000DB9C0001C9AA, 0x03FDB9A88000DBA0,
-    0x03FE29AA03FDB9A9, 0x03FE39AC03FE29AB, 0x03FE2DAE03FE39AD,
-    0x8000DB3C03FE25AF, 0x0000198100001182, 0x8000DBA403FFD180,
-    0x000334BE8000DBA8, 0x03FE81B88000DBAC, 0x03FE85BA03FE81B9,
-    0x03FE89BC03FE99BB, 0x03FF59BE03FEB1BD, 0x03FE5DB003FF59BF,
-    0x03FEE1B203FEE1B1, 0x03FEE1B403FEE1B3, 0x03FED1B603FEE9B5,
-    0x03FFB1A803FED1B7, 0x03FFB5AA03FFBDA9, 0x03FFD9AC03FFC9AB,
-    0x03FFC9AE03FFD9AD, 0x0000008603FFC1AF, 0x8000DBB000032887,
-    0x8000DBB88000DBB4, 0x8000DBC08000DBBC, 0x03FF75B88000DBC4,
-    0x03FF09BA03FF79B9, 0x03FF05BC03FF01BB, 0x03FF05BE03FF0DBD,
-    0x03FF61B003FF25BF, 0x03FF61B203FF61B1, 0x03FF65B403FF7DB3,
-    0x03FF59B603FF6DB5, 0x03FE3DB303FF4DB7, 0x8000DBCC8000DBC8,
-    0x8000DBD48000DBD0, 0x03FE51B503FE59B6, 0x03FEA5BB8000DBD8,
-    0x8000DBDC03FEA5BA, 0x03FEA9BF8000DBE0, 0x03FEA9BD03FEA1BE,
-    0x8000DBE403FEB5BC, 0x03FCA8EF8000DBE8, 0x8000DBF08000DBEC,
-    0x8000DBF88000DBF4, 0x8000DC008000DBFC, 0x8000DC088000DC04,
-    0x03FED8E18000DC0C, 0x03FC5CE38000DC10, 0x03FEBDA38000DC14,
-    0x0000F58100001582, 0x8000DC180000F980, 0x03FED1A503FED9A6,
-    0x03FE25AB000190BE, 0x0000088603FE25AA, 0x03FE29AF00017087,
-    0x03FE29AD03FE21AE, 0x8000DC2003FE35AC, 0x8000DC24000175B3,
-    0x00016DB68000DC28, 0x8000DC308000DC2C, 0x0000EDBA000165B5,
-    0x8000DC340000F5BB, 0x0000DDBE8000DC38, 0x0000E5BC0000B5BF,
-    0x0002A9A80000D5BD, 0x0002B9AA0002A9A9, 0x0002A9AC0002B9AB,
-    0x00031DAE0002A9AD, 0x0004ACBE000315AF, 0x8000DC408000DC3C,
-    0x8000DC440005A0BE, 0x8000DC4C8000DC48, 0x000311B88000DC50,
-    0x000311BA000311B9, 0x000331BC000311BB, 0x000331BE000331BD,
-    0x00036DB0000331BF, 0x000375B2000375B1, 0x00035DB4000345B3,
-    0x00033DB6000335B5, 0x000231A3000331B7, 0x8000DC588000DC54,
-    0x8000DC608000DC5C, 0x000221A5000229A6, 0x0003B1AB8000DC64,
-    0x8000DC680003A9AA, 0x0003F1AF8000DC6C, 0x000391AD000399AE,
-    0x0003C0840003A1AC, 0x8000DC748000DC70, 0x8000DC7C8000DC78,
-    0x00003D8100003D82, 0x8000DC8000003D80, 0x8000DC888000DC84,
-    0x8000DC948000DC90, 0x0004D8868000DACC, 0x0002C1A80004BC87,
-    0x0002C5AA0002CDA9, 0x0002C9AC0002D9AB, 0x0002F5AE0002FDAD,
-    0x8000DC98000139AF, 0x8000DCA08000DC9C, 0x8000DCA88000DCA4,
-    0x8000DCB08000DCAC, 0x0001D1B88000DCB4, 0x0001E1BA0001D9B9,
-    0x000195BC0001E1BB, 0x000195BE00019DBD, 0x000149B000018DBF,
-    0x00015DB2000155B1, 0x00014DB4000155B3, 0x0001F1B60001F1B5,
-    0x0000E4EF0001F1B7, 0x8000DCBC8000DCB8, 0x8000DCC000050C84,
-    0x8000DCC88000DCC4, 0x8000DCCC00015CEF, 0x8000DCD003FED4E1,
-    0x03FE18E303FE30E3, 0x000100E18000DCD4, 0x8000DCDC8000DCD8,
-    0x8000DCE48000DCE0, 0x8000DCE8000201B3, 0x000231B600020DB5,
-    0x8000DCF08000DCEC, 0x0001C1BA8000DCF4, 0x0001C5BC0001DDBB,
-    0x0001C5BE0001CDBD, 0x0005CDA3000179BF, 0x8000DCF88000DC8C,
-    0x8000DCFC8000DD00, 0x0005C1A50005FDA6, 0x000611AB8000DD04,
-    0x0003088600060DAA, 0x0006B5AF00037C87, 0x000601AD000609AE,
-    0x0007A180000609AC, 0x000041820007A181, 0x8000DD0800076DB3,
-    0x000775B600077DB5, 0x8000DD108000DD0C, 0x000751BA8000DD14,
-    0x0007F5BC000751BB, 0x0007E9BE0007FDBD, 0x00066DA80007E9BF,
-    0x0007A9AA000675A9, 0x0007B9AC0007A9AB, 0x0007A9AE0007B9AD,
-    0x8000DD180007A9AF, 0x8000DD208000DD1C, 0x8000DD288000DD24,
-    0x8000DD308000DD2C, 0x000769B88000DD34, 0x000775BA000771B9,
-    0x000715BC00070DBB, 0x000709BE00071DBD, 0x0007D9B0000701BF,
-    0x00076DB20007D9B1, 0x000761B4000765B3, 0x000761B6000761B5,
-    0x000621A300075DB7, 0x8000DD3C8000DD38, 0x8000DD448000DD40,
-    0x000631A5000639A6, 0x00061DAB8000DD48, 0x8000DD4C00061DAA,
-    0x0006A5AF8000DD50, 0x0006B1AD0006A5AE, 0x8000DD540006B9AC,
-    0x8000DD5800059CEF, 0x00000D808000DD5C, 0x00003D8200003581,
-    0x8000DD648000DD60, 0x0001A4848000DD68, 0x8000DD6C0001A4BE,
-    0x8000DD700005A4E3, 0x00018DA80006CCE1, 0x0001D9AA000199A9,
-    0x0001F1AC0001D1AB, 0x000139AE0001F1AD, 0x00000086000139AF,
-    0x8000DD7400016487, 0x8000DD7C8000DD78, 0x8000DD848000DD80,
-    0x0000CDB88000DD88, 0x0000D5BA0000D5B9, 0x0000FDBC0000E5BB,
-    0x000091BE000095BD, 0x000149B0000091BF, 0x000159B2000149B1,
-    0x000149B4000159B3, 0x0000FDB6000149B5, 0x000251B30000F5B7,
-    0x8000DD908000DD8C, 0x8000DD988000DD94, 0x000279B5000271B6,
-    0x00028DBB8000DD9C, 0x8000DDA000028DBA, 0x00028DBF8000DDA4,
-    0x00028DBD00028DBE, 0x00126C8500028DBC, 0x8000DDA8000215A3,
-    0x000235A68000DDAC, 0x8000DDB48000DDB0, 0x0002C9AA00023DA5,
-    0x8000DDB80002C9AB, 0x0002C9AE8000DDBC, 0x0002C9AC0002C9AF,
-    0x000380840002C9AD, 0x8000DDC48000DDC0, 0x8000DDC8000C00BE,
-    0x8000DDD08000DDCC, 0x8000DDD88000DDD4, 0x8000DDE48000DDE0,
-    0x8000DDEC8000DDE8, 0x8000DDF48000DDF0, 0x000059808000DDF8,
-    0x0000698200005981, 0x8000DDFC0002B8EF, 0x000318878000DE00,
-    0x8000DE04000CFC86, 0x8000DE0C8000DE08, 0x8000DE148000DE10,
-    0x8000DC1C00014CE1, 0x000279A800026CE3, 0x0002A5AA0002ADA9,
-    0x0002A5AC0002BDAB, 0x0002A5AE0002ADAD, 0x000DA884000119AF,
-    0x8000DE1C8000DE18, 0x8000DE248000DE20, 0x8000DE2C8000DE28,
-    0x000115B88000DE30, 0x000121BA00011DB9, 0x0001D5BC000121BB,
-    0x0001C9BE0001DDBD, 0x000169B00001C9BF, 0x000179B2000169B1,
-    0x000159B4000171B3, 0x000131B6000159B5, 0x000760E1000131B7,
-    0x000614E38000DE34, 0x8000DE38000C00BE, 0x8000DE3C0001A4E1,
-    0x8000DE400000E4E3, 0x8000DE488000DE44, 0x8000DE50000038EF,
-    0x00076CEF8000DE54, 0x000061808000DE58, 0x0000618200006181,
-    0x8000DE5C000261B3, 0x000261B6000261B5, 0x000CC0868000DE60,
-    0x0001E1BA000CE887, 0x0001E1BC0001E1BB, 0x0001D1BE0001E9BD,
-    0x000DADA30001D1BF, 0x8000DE4C8000DDDC, 0x8000DE688000DE64,
-    0x000DADA5000DADA6, 0x000E2DAB8000DE6C, 0x8000DE70000E2DAA,
-    0x000E1DAF8000DE74, 0x000E25AD000E1DAE, 0x8000DE78000E2DAC,
-    0x8000DE7C000EE9B3, 0x000EE9B68000DE80, 0x8000DE888000DE84,
-    0x000EB5BA000EE9B5, 0x8000DE8C000E85BB, 0x000E9DBE8000DE90,
-    0x000E9DBC000E85BF, 0x000E25A8000E95BD, 0x000E25AA000E2DA9,
-    0x000E25AC000E3DAB, 0x000E25AE000E2DAD, 0x8000DE94000E4DAF,
-    0x8000DE9C8000DE98, 0x8000DEA48000DEA0, 0x8000DEAC8000DEA8,
-    0x000FEDB88000DEB0, 0x000FFDBA000FF5B9, 0x000FEDBC000FF5BB,
-    0x000F49BE000F51BD, 0x000E35B0000F45BF, 0x000E0DB2000E05B1,
-    0x000E1DB4000E05B3, 0x000FDDB6000E05B5, 0x000FADA3000FD5B7,
-    0x0000198100001982, 0x8000DEB40001B180, 0x000FADA5000FADA6,
-    0x000FC1AB8000DEB8, 0x0002A084000FF1AA, 0x000FC1AF0001A0BE,
-    0x000FD1AD000FD9AE, 0x8000DEBC000FD9AC, 0x0000E886000EF9B3,
-    0x00015DB600012C87, 0x8000DEC48000DEC0, 0x000149BA000E81B5,
-    0x8000DEC8000149BB, 0x00013DBE8000DECC, 0x000139BC000121BF,
-    0x000E31A8000135BD, 0x000E35AA000E31A9, 0x000E4DAC000E5DAB,
-    0x000189AE000199AD, 0x0001E484000181AF, 0x8000DED48000DED0,
-    0x8000DEDC8000DED8, 0x8000DEE48000DEE0, 0x0001ADB88000DEE8,
-    0x00016DBA000165B9, 0x00017DBC000165BB, 0x00016DBE000165BD,
-    0x0001C9B0000165BF, 0x0001A9B20001C9B1, 0x0001BDB40001A5B3,
-    0x0001A1B60001A1B5, 0x000DBDA3000195B7, 0x8000DEF08000DEEC,
-    0x8000DEF88000DEF4, 0x000DC5A5000219A6, 0x00020DAB8000DEFC,
-    0x8000DF0000020DAA, 0x000265AF8000DF04, 0x000271AD000279AE,
-    0x8000DF0800027DAC, 0x8000DF108000DF0C, 0x8000DF188000DF14,
-    0x8000DF208000DF1C, 0x000039808000DF24, 0x0000058200003981,
-    0x00033CBE8000DF28, 0x000344878000DF30, 0x8000DF340004DC86,
-    0x0003E1B38000DF38, 0x000335B50004C084, 0x8000DF408000DF3C,
-    0x8000DF44000335B6, 0x000311BB8000DF48, 0x0000B5BD000319BA,
-    0x0000B5BF000309BC, 0x8000DF4C0000B5BE, 0x8000DF50000194E1,
-    0x8000DF5400017CE3, 0x8000DF5C8000DF58, 0x8000DF648000DF60,
-    0x8000DF6C8000DF68, 0x8000DF748000DF70, 0x8000DF7C8000DF78,
-    0x000255A8000274EF, 0x000261AA00025DA9, 0x0002B5AC000261AB,
-    0x0002B5AE0002BDAD, 0x000568BE0002ADAF, 0x8000DF848000DF80,
-    0x8000DF8C8000DF88, 0x0000AD810000AD82, 0x000169B80000AD80,
-    0x000109BA000169B9, 0x000119BC000109BB, 0x000109BE000119BD,
-    0x0002D5B0000109BF, 0x0002D5B20002DDB1, 0x000179B4000169B3,
-    0x000169B6000179B5, 0x000F44E1000161B7, 0x000EB8E30000BCE3,
-    0x8000DF90000E74E1, 0x8000DF988000DF94, 0x8000DFA48000DFA0,
-    0x8000DFAC8000DFA8, 0x000730BE8000DFB0, 0x000ECCEF000060EF,
-    0x0002A9A38000DFB4, 0x0004E8868000DFB8, 0x8000DFBC00050487,
-    0x00027DA500027DA6, 0x000259AB8000DFC0, 0x8000DFC4000251AA,
-    0x0001FDAF8000DFC8, 0x0001FDAD0001FDAE, 0x000675A8000241AC,
-    0x00068DAA000685A9, 0x000689AC000699AB, 0x0006BDAE0006B5AD,
-    0x8000DF9C0006DDAF, 0x8000DFD08000DFCC, 0x8000DFD88000DFD4,
-    0x8000DFE08000DFDC, 0x0006ADB88000DFE4, 0x0006BDBA0006B5B9,
-    0x000775BC0006B5BB, 0x000775BE00077DBD, 0x0006A9B000076DBF,
-    0x0006BDB20006B5B1, 0x0006ADB40006B5B3, 0x00069DB6000695B5,
-    0x000605B3000695B7, 0x8000DFEC8000DFE8, 0x8000DFF48000DFF0,
-    0x000601B5000635B6, 0x000671BB8000DFF8, 0x8000DFFC00066DBA,
-    0x000651BF8000E000, 0x000661BD000669BE, 0x00001982000669BC,
-    0x00006580000641A3, 0x000671A600001981, 0x8000E0088000E004,
-    0x000629AA000645A5, 0x00032484000635AB, 0x00062DAE8000E00C,
-    0x00062DAC000615AF, 0x0006A5B3000625AD, 0x000148868000E010,
-    0x8000E01400010C87, 0x0006B5B50006A9B6, 0x0006C9BB8000E018,
-    0x8000E01C0006F9BA, 0x00011DBF8000E020, 0x00011DBD00011DBE,
-    0x8000E0240006D1BC, 0x8000E02C8000E028, 0x0005CCE38000E030,
-    0x00053CE18000E034, 0x0004B8EF8000E038, 0x8000E0408000E03C,
-    0x8000E0488000E044, 0x8000E0508000E04C, 0x00061DA88000E054,
-    0x00062DAA000625A9, 0x00063DAC000625AB, 0x00065DAE000629AD,
-    0x8000E05800064DAF, 0x8000E0608000E05C, 0x8000E0688000E064,
-    0x8000E0708000E06C, 0x0001F9B88000E074, 0x00014DBA0001F9B9,
-    0x00015DBC000145BB, 0x000145BE000145BD, 0x00063DB0000175BF,
-    0x000605B200060DB1, 0x0001D9B40001C9B3, 0x0001C9B60001D9B5,
-    0x0003B9800001C9B7, 0x000019820003B981, 0x8000E0780005A9A3,
-    0x0005A5A60005B9A5, 0x8000E08000028484, 0x0005F5AA0003A8BE,
-    0x0005DDAC0005C5AB, 0x000211AE000211AD, 0x000301B6000211AF,
-    0x000344848000E084, 0x8000E088000361B5, 0x0003D887000349B3,
-    0x000309BE001CF886, 0x000335BC000309BF, 0x00033DBA000309BD,
-    0x8000E08C000321BB, 0x0003CDB08000E090, 0x0003D5B20003D1B1,
-    0x0003F9B40003E9B3, 0x0003E9B60003F9B5, 0x0003D9B80003E9B7,
-    0x000309BA0003D9B9, 0x000319BC000309BB, 0x000309BE000319BD,
-    0x8000E094000309BF, 0x00210DB58000E098, 0x8000E0A08000E09C,
-    0x8000E0A88000E0A4, 0x0003E5A88000DF2C, 0x0003F5AA0003F5A9,
-    0x0003DDAC0003C5AB, 0x0003BDAE0003B5AD, 0x00021DA30003B5AF,
-    0x8000E0B08000E0AC, 0x001D1CBE8000E0B4, 0x000235A5000255A6,
-    0x000275AB8000E0B8, 0x8000E0BC000269AA, 0x00025DAF8000E0C0,
-    0x00025DAD00025DAE, 0x8000E0C4000261AC, 0x8000E0CC8000E0C8,
-    0x00038CEF8000E0D0, 0x8000E0D88000E0D4, 0x00036CE38000E0DC,
-    0x000078E18000E0E0, 0x8000E0E88000E0E4, 0x001C58868000E0F0,
-    0x0000E580001DF087, 0x0000F9820000E981, 0x8000E0F88000E0F4,
-    0x8000E1008000E0FC, 0x8000E1040006D4EF, 0x8000E108000674E1,
-    0x8000E10C000154E3, 0x8000E1148000E110, 0x000191B38000E118,
-    0x8000E1208000E11C, 0x8000E124001DC484, 0x0001A5B5000115B6,
-    0x000135BB8000E128, 0x8000E12C000135BA, 0x0000FDBF8000E130,
-    0x0000FDBD0000FDBE, 0x001E45A9000125BC, 0x001E71AB001E25A8,
-    0x001ED1AD001E5DAA, 0x001ED5AF001E69AC, 0x8000E0EC001ED9AE,
-    0x8000E1388000E134, 0x8000E1408000E13C, 0x8000E1488000E144,
-    0x001F01B98000E14C, 0x001F01BB001F69B8, 0x001F05BD001F09BA,
-    0x001F05BF001F19BC, 0x001EB5B1001F05BE, 0x001F6DB3001EB5B0,
-    0x001F7DB5001EB5B2, 0x001F51B7001F7DB4, 0x001ED1A3001F61B6,
-    0x0000158100002D82, 0x8000E15000001D80, 0x001EE5A5001E55A6,
-    0x001E75AB8000E154, 0x8000E158001E75AA, 0x001FBDAF000360BE,
-    0x001FBDAD001FBDAE, 0x8000E15C001E65AC, 0x00006886001F55B3,
-    0x001F71B600036C87, 0x8000E1648000E160, 0x001F55BA001F45B5,
-    0x8000E168001F5DBB, 0x001F39BE8000E16C, 0x001F45BC001F25BF,
-    0x8000E170001F39BD, 0x8000E1788000E174, 0x8000E1808000E17C,
-    0x8000E1888000E184, 0x8000E18C001CF4EF, 0x8000E190001D60E1,
-    0x8000E194001D74E3, 0x8000E19C8000E198, 0x001ED1A38000E1A0,
-    0x8000E1A88000E1A4, 0x8000E1B08000E1AC, 0x001EC1A5001EF5A6,
-    0x001ED9AB8000E1B4, 0x8000E1B8001ED1AA, 0x001EA1AF8000E1BC,
-    0x001EBDAD001EBDAE, 0x001EF5A8001EC1AC, 0x001E19AA001E09A9,
-    0x001E09AC001E19AB, 0x001E31AE001E35AD, 0x00000D80001E2DAF,
-    0x00001D8200001581, 0x8000E1C48000E1C0, 0x0001FC848000E1C8,
-    0x001E2DB80001F0BE, 0x0001E1BA0001E1B9, 0x0001E1BC0001E1BB,
-    0x0001E1BE0001E1BD, 0x001E55B00001E1BF, 0x001E55B2001E5DB1,
-    0x001E39B4001E29B3, 0x001E25B6001E25B5, 0x0002D5A8001E15B7,
-    0x0002E1AA0002DDA9, 0x000311AC0002E1AB, 0x000311AE000311AD,
-    0x8000E1D0000311AF, 0x8000E1D88000E1D4, 0x8000E1E08000E1DC,
-    0x8000E1E88000E1E4, 0x00037DB88000E1EC, 0x000301BA000301B9,
-    0x000301BC000301BB, 0x000331BE000309BD, 0x000371B0000331BF,
-    0x000371B2000371B1, 0x000359B400034DB3, 0x00034DB6000345B5,
-    0x00042086000345B7, 0x000261B30003A087, 0x000391B5000480BE,
-    0x8000E1F48000E1F0, 0x8000E07C000399B6, 0x0003BDBB0003B884,
-    0x000391BD0003BDBA, 0x000389BF0003ADBC, 0x000229A3000395BE,
-    0x8000E1FC8000E1F8, 0x000694858000E200, 0x0003D9A50003D1A6,
-    0x0003F5AB8000E204, 0x8000E2080003F5AA, 0x0003C1AF8000E20C,
-    0x0003D9AD0003DDAE, 0x8000E2100003E5AC, 0x8000E2188000E214,
-    0x8000E2208000E21C, 0x8000E2288000E224, 0x000011818000E22C,
-    0x8000E23000000D80, 0x8000E23400001982, 0x8000E2408000E238,
-    0x8000E2488000E244, 0x8000E24C000194E1, 0x8000E2500000E4E3,
-    0x000478868000E254, 0x8000E25800028087, 0x8000E2608000E25C,
-    0x8000E2688000E264, 0x8000E2708000E26C, 0x8000E274000038EF,
-    0x8000E27C8000E278, 0x8000E2848000E280, 0x8000E28C8000E288,
-    0x00046CBE0001C8EF, 0x0005A084001EC4E1, 0x8000E290001E70E3,
-    0x8000E2988000E294, 0x8000E2A08000E29C, 0x0001F9B50001F1B6,
-    0x0001F9B30020F485, 0x0000218100002982, 0x00013DBF00001D80,
-    0x000121BD000129BE, 0x000169BB000129BC, 0x8000E2A4000161BA,
-    0x000621A88000E2A8, 0x000621AA000621A9, 0x000621AC000621AB,
-    0x000621AE000621AD, 0x8000E23C000621AF, 0x000C008600012CBF,
-    0x8000E2AC00036087, 0x8000E2B48000E2B0, 0x000741B88000E2B8,
-    0x000741BA000741B9, 0x000741BC000741BB, 0x000741BE000741BD,
-    0x0006F5B0000741BF, 0x0006ADB20006FDB1, 0x0006BDB40006A5B3,
-    0x000691B60006A9B5, 0x0006A9A3000691B7, 0x8000E2C08000E2BC,
-    0x8000E2C88000E2C4, 0x0006A9A50006A1A6, 0x000639AB8000E2CC,
-    0x8000E2D0000631AA, 0x00066DAF8000E2D4, 0x000671AD000679AE,
-    0x8000E2D8000679AC, 0x8000E2DC000785B3, 0x000785B68000E2E0,
-    0x8000E2E88000E2E4, 0x000789BA00078DB5, 0x8000E2EC000789BB,
-    0x000789BE8000E2F0, 0x000789BC000775BF, 0x8000E2F4000789BD,
-    0x00057CEF0005C0E1, 0x8000E2F80004D4E3, 0x8000E3008000E2FC,
-    0x8000E3088000E304, 0x8000E3108000E30C, 0x0005A8E18000E314,
-    0x00058CE38000E318, 0x000601A3000CF8EF, 0x0000658100007182,
-    0x0001808400006580, 0x000609A5000601A6, 0x00060DAB000010BE,
-    0x000FE88600060DAA, 0x0006F1AF00001887, 0x00060DAD00060DAE,
-    0x0006B9A800060DAC, 0x0006C9AA0006B9A9, 0x0006C1AC0006C5AB,
-    0x0006F9AE0006C1AD, 0x8000E3200006F9AF, 0x8000E3288000E324,
-    0x8000E3308000E32C, 0x8000E3388000E334, 0x00064DB88000E33C,
-    0x00065DBA000655B9, 0x000661BC000655BB, 0x000661BE000661BD,
-    0x000689B0000661BF, 0x000699B2000689B1, 0x000689B4000699B3,
-    0x00067DB6000675B5, 0x8000E340000675B7, 0x000285B600028DB5,
-    0x03E8F18F000C4084, 0x8000E3488000E344, 0x0002E1BC000221B3,
-    0x0002E5BE0002EDBD, 0x0032C8AC0002F9BF, 0x000291BA8000E34C,
-    0x001F019B000291BB, 0x001E4D99001E4D9A, 0x0004F19F00133D98,
-    0x001A799D001AB99E, 0x001641930018819C, 0x03EAF99100149192,
-    0x0012619703EA4D90, 0x0010819500127196, 0x003801B100164194,
-    0x003A81B3003EE1B0, 0x002691B5003A89B2, 0x8000E350003BD1B4,
-    0x000281A38000E354, 0x000669A1000011A2, 0x8000E35800068DA0,
-    0x8000E35C8000E1CC, 0x0002F1A18000E360, 0x000ECDA38000E364,
-    0x000801A5000ECDA2, 0x000BC5A7000F51A4, 0x0036C1A9000AFDA6,
-    0x003001AB003401A8, 0x0032CDAD003641AA, 0x003EFDAF0032D5AC,
-    0x03E9E182003C01AE, 0x8000E36803EB2583, 0x001731868000E36C,
-    0x03EB898400172987, 0x0011D98A0015E985, 0x8000E3700013218B,
-    0x001F358E8000E374, 0x0013298C001F3D8F, 0x001E019200124D8D,
-    0x8000E37800199993, 0x0005F196000C0084, 0x001B799400077597,
-    0x0007A19A001BB195, 0x8000E37C0001D19B, 0x0000F1818000E380,
-    0x0003599C00001980, 0x8000E3840000F982, 0x000DC0BE8000E388,
-    0x000C3086000028EF, 0x8000E390000D5C87, 0x8000E3940006C4EF,
-    0x8000E3980006E4E1, 0x8000E39C000614E3, 0x8000E3A0000184E1,
-    0x0003FDA30000FCE3, 0x8000E3A88000E3A4, 0x8000E3AC000F4C84,
-    0x000351A5000359A6, 0x00034DAB8000E3B0, 0x8000E3B400034DAA,
-    0x000325AF8000E3B8, 0x000331AD000339AE, 0x8000E31C00033DAC,
-    0x8000E3BC8000E38C, 0x8000E3C48000E3C0, 0x8000E3CC8000E3C8,
-    0x000D81A88000E3D0, 0x000D81AA000D81A9, 0x000D81AC000D81AB,
-    0x000D81AE000D81AD, 0x000E41B0000D81AF, 0x000E51B2000E51B1,
-    0x000EC9B4000E79B3, 0x000EE9B6000EC9B5, 0x000EDDB8000EE5B7,
-    0x000F61BA000EA1B9, 0x000F59BC000F61BB, 0x000F31BE000F59BD,
-    0x000D29B3000F15BF, 0x8000E3D88000E3D4, 0x8000E3E08000E3DC,
-    0x000D29B5000D21B6, 0x000EB5BB8000E3E4, 0x8000E3E8000EA5BA,
-    0x000E2DBF8000E3EC, 0x000E8DBD000E2DBE, 0x00002982000EA5BC,
-    0x00001D80000D6DA3, 0x000D65A600001581, 0x8000E3F48000E3F0,
-    0x000EE1AA000D6DA5, 0x8000E3F8000EF1AB, 0x000E69AE0003A4BE,
-    0x000EE1AC000E69AF, 0x000E99A8000EC9AD, 0x000EA1AA000EA5A9,
-    0x000EA9AC000EBDAB, 0x000ED1AE000ED1AD, 0x00008086000ECDAF,
-    0x8000E40000010487, 0x8000E4088000E404, 0x8000E4108000E40C,
-    0x00015DB88000E414, 0x0001C1BA0001C1B9, 0x0001C1BC0001C1BB,
-    0x0001F1BE0001C9BD, 0x000EBDB00001F1BF, 0x000E85B2000E8DB1,
-    0x00017DB4000165B3, 0x00016DB6000165B5, 0x000E29B3000165B7,
-    0x8000E4188000E3FC, 0x8000E4208000E41C, 0x000E21B5000E51B6,
-    0x000E49BB8000E424, 0x8000E428000E79BA, 0x0001A1BF8000E42C,
-    0x0001B1BD0001B9BE, 0x8000E430000E51BC, 0x8000E434000E6DA3,
-    0x000E15A68000E438, 0x8000E4408000E43C, 0x000E3DAA000E65A5,
-    0x8000E444000E0DAB, 0x0001FDAE8000E448, 0x000E15AC0001E5AF,
-    0x000138E10001F5AD, 0x000FD4E38000E44C, 0x8000E4548000E450,
-    0x8000E458000358BE, 0x00000D808000E45C, 0x00001D8200001581,
-    0x8000E4648000E460, 0x0002D4EF8000E468, 0x0004E0868000E470,
-    0x000400BE0003E087, 0x8000E47400037DB3, 0x000365B600036DB5,
-    0x8000E47C8000E478, 0x0003F9BA8000E480, 0x0003D9BC0003F1BB,
-    0x0003B9BE0003D9BD, 0x000EBDAA0003B5BF, 0x000EBDA8000E01AB,
-    0x000E01AE000EB5A9, 0x000E11AC000E09AF, 0x8000E484000E11AD,
-    0x8000E48C8000E488, 0x8000E4948000E490, 0x8000E49C8000E498,
-    0x000FD1BA8000E4A0, 0x000E25B8000FD1BB, 0x000FF1BE000FCDB9,
-    0x000FC1BC000FF1BF, 0x000E21B2000FC1BD, 0x000E41B0000E2DB3,
-    0x000E29B6000E41B1, 0x000E35B4000E1DB7, 0x000229A8000E29B5,
-    0x000239AA000229A9, 0x000229AC000239AB, 0x000291AE000229AD,
-    0x8000E4A4000291AF, 0x8000E4AC8000E4A8, 0x8000E4B48000E4B0,
-    0x002B30BF8000E4B8, 0x000299B88000E4BC, 0x0002A5BA0002ADB9,
-    0x00015DBC000149BB, 0x000141BE000145BD, 0x0002F1B0000141BF,
-    0x0002C1B20002F9B1, 0x0002B1B40002C1B3, 0x0002B5B60002BDB5,
-    0x000E08E10002A9B7, 0x000ED4E3000CF4E1, 0x8000E4C0000F90E3,
-    0x8000E4C88000E4C4, 0x000011818000E4CC, 0x8000E4D000000D80,
-    0x8000E4D400001982, 0x0001ACEF8000E4D8, 0x0002F1A3000ECCEF,
-    0x8000E4E48000E4E0, 0x8000E4E800058084, 0x0002E1A50002E9A6,
-    0x00027DAB8000E4EC, 0x0004C886000275AA, 0x000239AF0004EC87,
-    0x000255AD000235AE, 0x00063DA8000255AC, 0x000655AA000655A9,
-    0x000675AC00066DAB, 0x0006BDAE0006B5AD, 0x8000E4DC0006B5AF,
-    0x8000E4F48000E4F0, 0x8000E4FC8000E4F8, 0x8000E5048000E500,
-    0x00076DB88000E508, 0x00077DBA000775B9, 0x0007C1BC000771BB,
-    0x0007C1BE0007C1BD, 0x0006CDB00007C1BF, 0x0006DDB20006D5B1,
-    0x0006CDB40006D5B3, 0x00075DB6000755B5, 0x0006DDB3000755B7,
-    0x8000E5108000E50C, 0x8000E5188000E514, 0x0006FDB5000635B6,
-    0x000611BB8000E51C, 0x8000E520000609BA, 0x000791BF8000E524,
-    0x000601BD000609BE, 0x8000E528000609BC, 0x8000E52C000699A3,
-    0x000671A68000E530, 0x8000E5388000E534, 0x00064DAA0006B9A5,
-    0x8000E53C000655AB, 0x00064DAE8000E540, 0x00064DAC0007D5AF,
-    0x0001ED80000645AD, 0x0000198200000981, 0x8000E5488000E544,
-    0x0001C4BE0001D084, 0x000080868000E54C, 0x8000E55000012087,
-    0x8000E5588000E554, 0x8000E5608000E55C, 0x8000E5688000E564,
-    0x8000E5708000E56C, 0x8000E5788000E574, 0x8000E57C00050CE1,
-    0x8000E580000480E3, 0x8000E5888000E584, 0x8000E58C0005DCEF,
-    0x8000E5948000E590, 0x000645A88000E598, 0x000659AA00064DA9,
-    0x0001B5AC000651AB, 0x0001B5AE0001BDAD, 0x0001AC840001ADAF,
-    0x8000E5A08000E59C, 0x8000E5A88000E5A4, 0x8000E5B08000E5AC,
-    0x000169B88000E5B4, 0x000109BA000169B9, 0x00011DBC000105BB,
-    0x000101BE000101BD, 0x0001D5B0000135BF, 0x0001D5B20001DDB1,
-    0x000179B4000169B3, 0x000169B6000179B5, 0x0003BD81000161B7,
-    0x0005D9B30003BD80, 0x000235B500001982, 0x000344BE8000E5B8,
-    0x8000E5C0000235B6, 0x000211BB8000E5C4, 0x0002B5BD000219BA,
-    0x0002B5BF000209BC, 0x8000E5C80002B5BE, 0x00026DA500026DA6,
-    0x000581A3003C0084, 0x003CF8868000E5CC, 0x0002EDAF00037C87,
-    0x0002EDAD0002EDAE, 0x000249AB000251AC, 0x8000E5D0000241AA,
-    0x8000E5D88000E5D4, 0x8000E5E08000E5DC, 0x000120E38000E5E4,
-    0x0001A0E18000E5E8, 0x000240EF8000E5EC, 0x8000E5F48000E5F0,
-    0x8000E5FC8000E5F8, 0x8000E6048000E600, 0x0003F1B38000E608,
-    0x8000E6108000E60C, 0x8000E46C8000E614, 0x0003E1B50003E9B6,
-    0x0003ADBB8000E618, 0x8000E61C0003ADBA, 0x000395BF8000E620,
-    0x0003A1BD000395BE, 0x000229A80003ADBC, 0x000239AA000229A9,
-    0x000229AC000239AB, 0x000259AE000229AD, 0x0001CD80000259AF,
-    0x0000198200000981, 0x8000E6288000E624, 0x003DB487003D7CBE,
-    0x000235B8003C1C86, 0x000235BA00023DB9, 0x0002F9BC0002E9BB,
-    0x0002E9BE0002F9BD, 0x000229B00002E9BF, 0x000239B2000229B1,
-    0x000211B4000235B3, 0x000211B6000211B5, 0x8000E63000020DB7,
-    0x8000E6388000E634, 0x8000E6408000E63C, 0x8000E6488000E644,
-    0x8000E64C00007CEF, 0x8000E650000654E1, 0x0032A8AC0000A8E3,
-    0x8000E6588000E654, 0x8000E6608000E65C, 0x0002F1A50002F9A6,
-    0x0002E1A38000E664, 0x8000E66C8000E668, 0x000285AF8000E670,
-    0x0002B1AD000285AE, 0x0002BDAB0002BDAC, 0x8000E6740002BDAA,
-    0x003E21A88000E678, 0x003E49AA003E21A9, 0x003E69AC003E59AB,
-    0x003E99AE003E69AD, 0x8000E62C003E99AF, 0x8000E6808000E67C,
-    0x8000E6888000E684, 0x8000E6908000E68C, 0x003EF5B88000E694,
-    0x003EF5BA003EFDB9, 0x003E99BC003E89BB, 0x003E85BE003E99BD,
-    0x003EE9B0003EB5BF, 0x003EFDB2003EE9B1, 0x003ED5B4003EF1B3,
-    0x003ED5B6003EDDB5, 0x003EC1B3003ECDB7, 0x0000B58100001582,
-    0x8000E6980000B580, 0x003EE1B5003EE9B6, 0x003E2DBB000010BE,
-    0x00036484003E2DBA, 0x003E19BF8000E69C, 0x003E25BD003E11BE,
-    0x8000E6A0003E29BC, 0x0000A886003E85A3, 0x003EADA600036C87,
-    0x8000E6A88000E6A4, 0x003E69AA003EA5A5, 0x8000E6AC003E69AB,
-    0x003E55AE8000E6B0, 0x003E6DAC003E5DAF, 0x003E89B3003E61AD,
-    0x8000E6B88000E6B4, 0x8000E6C08000E6BC, 0x003E89B5003E81B6,
-    0x003E95BB8000E6C4, 0x8000E6C8003E8DBA, 0x003EFDBF8000E6CC,
-    0x003E85BD003EF5BE, 0x8000E6D0003E85BC, 0x8000E6D88000E6D4,
-    0x8000E6E08000E6DC, 0x003C94EF8000E6E4, 0x003D20E18000E6E8,
-    0x003D34E38000E6EC, 0x8000E6F48000E6F0, 0x8000E6FC8000E6F8,
-    0x003E01A88000E700, 0x003E01AA003E01A9, 0x003E01AC003E01AB,
-    0x003E01AE003E01AD, 0x0000DD80003E3DAF, 0x0000198200000981,
-    0x8000E7088000E704, 0x000074BE00013C84, 0x003E1DB88000E710,
-    0x003E35BA003E3DB9, 0x0001F9BC0001E9BB, 0x0001E9BE0001F9BD,
-    0x003E49B00001E9BF, 0x003E51B2003E51B1, 0x003E31B4003E51B3,
-    0x003E2DB6003E39B5, 0x003D89A3003E25B7, 0x0004A8868000E714,
-    0x8000E71800033C87, 0x003D89A5003D81A6, 0x003D95AB8000E71C,
-    0x8000E720003D8DAA, 0x003DFDAF8000E724, 0x003D85AD003DF5AE,
-    0x8000E728003D85AC, 0x0003488400029DB3, 0x000355B6000440BE,
-    0x8000E7308000E72C, 0x000371BA0002B1B5, 0x8000E734000379BB,
-    0x000331BE8000E738, 0x000351BC000331BF, 0x0003D1B8000351BD,
-    0x0003D5BA0003D1B9, 0x0003FDBC0003E9BB, 0x0003EDBE0003E5BD,
-    0x0003B1B00003E5BF, 0x000381B20003B9B1, 0x0003F1B4000381B3,
-    0x0003F1B60003F1B5, 0x000201A80003F1B7, 0x000211AA000211A9,
-    0x00023DAC000225AB, 0x0003D1AE0003D1AD, 0x8000E73C0003D1AF,
-    0x8000E7408000E5BC, 0x8000E7488000E744, 0x8000E7508000E74C,
-    0x8000E7588000E754, 0x8000E7608000E75C, 0x000050EF8000E764,
-    0x8000E76C8000E768, 0x000084E38000E770, 0x0001F8E18000E774,
-    0x00002D808000E778, 0x0000158200001181, 0x000399A38000E780,
-    0x8000E7888000E784, 0x8000E7908000E78C, 0x0003B5A5000251A6,
-    0x00027DAB8000E794, 0x8000E798000275AA, 0x000235AF8000E79C,
-    0x000255AD000235AE, 0x0004A086000255AC, 0x8000E7A00005A887,
-    0x8000E7A88000E7A4, 0x8000E7B08000E7AC, 0x00042CBE0001F8EF,
-    0x8000E7B4003E80E1, 0x8000E7B8000114E3, 0x8000E7C08000E7BC,
-    0x8000E7C88000E7C4, 0x8000E7D08000E7CC, 0x8000E7D40001DDB3,
-    0x0001F5B60001FDB5, 0x8000E7DC8000E7D8, 0x0001B1BA8000E7E0,
-    0x00019DBC000185BB, 0x000111BE000115BD, 0x0005BDA8000111BF,
-    0x000619AA000601A9, 0x000639AC000611AB, 0x000659AE000639AD,
-    0x8000E77C000649AF, 0x0000198100001182, 0x8000E7E40007C980,
-    0x000230848000E7E8, 0x0006B9B88000E7EC, 0x000749BA0006B9B9,
-    0x000759BC000749BB, 0x00074DBE000759BD, 0x000639B0000745BF,
-    0x000605B200060DB1, 0x000699B400068DB3, 0x000689B6000699B5,
-    0x000691A3000689B7, 0x000028868000E7F0, 0x8000E7F40003FC87,
-    0x0006B1A50006B9A6, 0x0006C9AB8000E7F8, 0x8000E7FC0006FDAA,
-    0x00065DAF8000E800, 0x000659AD00065DAE, 0x8000E8040006D1AC,
-    0x8000E80800078DB3, 0x00078DB68000E80C, 0x8000E8148000E810,
-    0x000769BA00078DB5, 0x8000E818000769BB, 0x000769BE8000E81C,
-    0x000769BC00075DBF, 0x000649A3000769BD, 0x8000E8248000E820,
-    0x8000E82C8000E828, 0x000649A5000649A6, 0x0006ADAB8000E830,
-    0x8000E8340006ADAA, 0x000699AF8000E838, 0x0006ADAD0006ADAE,
-    0x8000E83C0006ADAC, 0x8000E8448000E840, 0x8000E84C8000E848,
-    0x8000E8500005E0EF, 0x8000E8588000E854, 0x0001248400053CE3,
-    0x0000298000057CE1, 0x0000198200002981, 0x0006BDA8000134BE,
-    0x0006C1AA0006C1A9, 0x0006C1AC0006C1AB, 0x0006F1AE0006C9AD,
-    0x000000860006F1AF, 0x8000E86000008487, 0x8000E8688000E864,
-    0x8000E8708000E86C, 0x000679B88000E874, 0x000609BA000679B9,
-    0x000619BC000609BB, 0x000605BE000605BD, 0x000669B000063DBF,
-    0x000679B2000669B1, 0x000669B4000679B3, 0x000649B600065DB5,
-    0x000285A8000649B7, 0x000295AA000295A9, 0x0002BDAC0002A5AB,
-    0x0002D1AE0002D5AD, 0x8000E8780002D1AF, 0x000328848000E87C,
-    0x8000E8800009E5B5, 0x8000E8888000E884, 0x00035DB88000E88C,
-    0x000365BA00036DB9, 0x000315BC00030DBB, 0x000315BE00031DBD,
-    0x0002B5B000030DBF, 0x000281B20002BDB1, 0x000375B4000281B3,
-    0x000375B600037DB5, 0x00021DB300036DB7, 0x8000E8948000E890,
-    0x000C1CBE8000E898, 0x000235B5000255B6, 0x000279BB8000E89C,
-    0x8000E8A0000271BA, 0x0003B9BF8000E8A4, 0x000255BD0003B1BE,
-    0x8000E8A8000255BC, 0x8000E8AC000259A3, 0x000211A68000E8B0,
-    0x8000E8B88000E8B4, 0x000235AA000271A5, 0x8000E70C00023DAB,
-    0x0003F5AE8000E8BC, 0x000211AC0003FDAF, 0x0002A9A8000211AD,
-    0x0002B9AA0002A9A9, 0x0002A9AC0002B9AB, 0x000139AE0002A9AD,
-    0x00000D80000139AF, 0x00001D8200001581, 0x8000E8C48000E8C0,
-    0x000D68878000E8C8, 0x0001E9B8000D1C86, 0x000189BA0001E9B9,
-    0x00019DBC000185BB, 0x000181BE000181BD, 0x000151B00001B1BF,
-    0x000151B2000151B1, 0x0001F5B4000151B3, 0x0001F5B60001FDB5,
-    0x000004EF0001E1B7, 0x8000E8D0000C4884, 0x8000E8D4000C80BE,
-    0x8000E8DC8000E8D8, 0x8000E8E00006C4EF, 0x8000E8E400063CE1,
-    0x0000C8E3000184E3, 0x000128E18000E8E8, 0x8000E8F08000E8EC,
-    0x8000E8F88000E8F4, 0x8000E8FC00027DB3, 0x000265B600026DB5,
-    0x8000E9048000E900, 0x0001D9BA8000E908, 0x0001C9BC0001D9BB,
-    0x00017DBE0001C9BD, 0x000DB1A3000175BF, 0x8000E90C8000E8CC,
-    0x8000E9148000E910, 0x000DA1A5000DA9A6, 0x000E15AB8000E918,
-    0x8000E91C000E15AA, 0x000EB9AF8000E920, 0x000E05AD000EB1AE,
-    0x8000E924000E05AC, 0x8000E92C8000E928, 0x00001D808000E930,
-    0x0000398200000981, 0x8000E9388000E934, 0x0000C8868000E93C,
-    0x8000E94000030C87, 0x8000E9488000E944, 0x000F09A88000E94C,
-    0x000F1DAA000F15A9, 0x000F31AC000F15AB, 0x000F2DAE000F3DAD,
-    0x0003A884000F21AF, 0x8000E9548000E950, 0x8000E95C8000E958,
-    0x8000E9648000E960, 0x0000CDB88000E968, 0x0000D5BA0000D5B9,
-    0x0000FDBC0000E5BB, 0x00009DBE000095BD, 0x000F21B0000095BF,
-    0x000F21B2000F21B1, 0x000F2DB4000F3DB3, 0x000F1DB6000F15B5,
-    0x000E95B30000F5B7, 0x8000E9708000E96C, 0x8000E9788000E974,
-    0x000EB9B5000EBDB6, 0x000EE5BB8000E97C, 0x8000E980000EE5BA,
-    0x000ED1BF8000E984, 0x000EE9BD000EE9BE, 0x000D28EF000EF5BC,
-    0x8000E98C8000E988, 0x8000E9948000E990, 0x8000E99C8000E998,
-    0x00003D808000E9A0, 0x00001D8200000981, 0x000D48E38000E9A4,
-    0x000EC0E18000E9A8, 0x000E15A300017CBE, 0x000028868000E9B0,
-    0x8000E9B400014C87, 0x000E39A5000E3DA6, 0x000E65AB8000E9B8,
-    0x8000E9BC000E65AA, 0x000E51AF8000E9C0, 0x000E69AD000E69AE,
-    0x8000E9C4000E75AC, 0x8000E9C80001BDB3, 0x000179B68000E9CC,
-    0x8000E9D48000E9D0, 0x000155BA000179B5, 0x8000E9D800015DBB,
-    0x0000F9BE8000E9DC, 0x000145BC0000F9BF, 0x000271A80000F9BD,
-    0x000271AA000271A9, 0x0002B5AC000271AB, 0x0002B5AE0002BDAD,
-    0x0004EC840002ADAF, 0x8000E9E48000E9E0, 0x8000E9EC8000E9E8,
-    0x8000E9F48000E9F0, 0x000369B88000E9F8, 0x00030DBA000369B9,
-    0x00031DBC000305BB, 0x00030DBE000305BD, 0x0002D5B0000305BF,
-    0x0002D5B20002DDB1, 0x000379B4000369B3, 0x000369B6000379B5,
-    0x8000E9FC000361B7, 0x8000EA048000EA00, 0x8000EA080002F5A3,
-    0x000231A6000231A5, 0x8000EA108000EA0C, 0x00021DAA8000EA14,
-    0x00020DAC000215AB, 0x0003B1AE0003B1AD, 0x000061800003B1AF,
-    0x0000058200006181, 0x0004F0868000EA18, 0x000410BE00036087,
-    0x8000EA248000EA20, 0x8000EA2C8000EA28, 0x8000EA348000EA30,
-    0x8000EA3C8000EA38, 0x000285A88000E85C, 0x000295AA000295A9,
-    0x0002BDAC0002A5AB, 0x0002D1AE0002D5AD, 0x8000EA400002D1AF,
-    0x8000EA488000EA44, 0x8000EA508000EA4C, 0x8000EA588000EA54,
-    0x000175B88000EA5C, 0x000175BA00017DB9, 0x0001D5BC0001CDBB,
-    0x0001C9BE0001DDBD, 0x0002B5B00001C1BF, 0x000281B20002BDB1,
-    0x000155B4000281B3, 0x000155B600015DB5, 0x000EA4E100014DB7,
-    0x000E1CE300040884, 0x00049CBE8000EA60, 0x8000EA688000EA64,
-    0x8000EA708000EA6C, 0x8000EA788000EA74, 0x000075818000EA7C,
-    0x000ECCEF00007580, 0x8000EA8000007582, 0x8000EA888000EA84,
-    0x8000EA900004C484, 0x00005CEF8000EA94, 0x000190E18000EA98,
-    0x000158E38000EA9C, 0x8000EAA48000EAA0, 0x000758868000EAA8,
-    0x00059DB30004EC87, 0x8000EA8C8000EA1C, 0x8000EAB08000EAAC,
-    0x0005B5B50005D5B6, 0x0005F9BB8000EAB4, 0x8000EAB80005F1BA,
-    0x000619BF8000EABC, 0x0005D5BD000611BE, 0x8000EAC00005D5BC,
-    0x8000EAC40005D9A3, 0x000591A68000EAC8, 0x8000EAD08000EACC,
-    0x0005B5AA0005F1A5, 0x8000EAD40005BDAB, 0x000655AE8000EAD8,
-    0x000591AC00065DAF, 0x00065DA8000591AD, 0x00066DAA000665A9,
-    0x00067DAC000665AB, 0x00066DAE000665AD, 0x0000A0BE000665AF,
-    0x8000EAE08000EADC, 0x8000EAE88000EAE4, 0x8000EAF08000EAEC,
-    0x00074DB88000EAF4, 0x000751BA000751B9, 0x000771BC000751BB,
-    0x000771BE000771BD, 0x0007C1B0000771BF, 0x0007C1B20007C1B1,
-    0x0007C1B40007C1B3, 0x0007C1B60007C1B5, 0x8000EAF80007C1B7,
-    0x8000EB00000694E1, 0x0001A98000053CE3, 0x000019820001A981,
-    0x000080868000EB04, 0x8000EB0800010487, 0x8000EB108000EB0C,
-    0x8000EB188000EB14, 0x0007A5B300051CEF, 0x8000EB1C8000EAFC,
-    0x8000EB248000EB20, 0x0007B5B50007A5B6, 0x000645BB8000EB28,
-    0x8000EB2C000645BA, 0x000645BF8000EB30, 0x000655BD000645BE,
-    0x000765A3000655BC, 0x8000EB388000EB34, 0x8000EB408000EB3C,
-    0x000775A5000765A6, 0x000685AB8000EB44, 0x8000EB48000685AA,
-    0x000685AF8000EB4C, 0x000695AD000685AE, 0x8000EB50000695AC,
-    0x8000EB54000619B3, 0x000621B68000EB58, 0x8000EB608000EB5C,
-    0x0001F1BA000631B5, 0x8000EB640001F9BB, 0x0001B1BE8000EB68,
-    0x0001D1BC0001BDBF, 0x0005D1A80001D1BD, 0x0005D5AA0005D1A9,
-    0x0005FDAC0005E9AB, 0x0005EDAE0005E5AD, 0x0003F980000211AF,
-    0x0003C5820003CD81, 0x001C00BE00032084, 0x000304878000EB70,
-    0x000219B8001CBC86, 0x000225BA00022DB9, 0x0002F9BC0002E9BB,
-    0x0002E9BE0002F9BD, 0x000271B00002E9BF, 0x000241B2000279B1,
-    0x000231B4000241B3, 0x000235B600023DB5, 0x8000E9AC000229B7,
-    0x8000EB788000EB74, 0x8000EB7C000551A3, 0x000569A6000579A5,
-    0x8000EB848000EB80, 0x0002B9AA8000EB88, 0x000299AC0002B1AB,
-    0x0002F9AE000299AD, 0x0001F8E10002F5AF, 0x0002F4E38000EB8C,
-    0x8000EB90001CC084, 0x8000EB988000EB94, 0x8000EBA08000EB9C,
-    0x8000EBA88000EBA4, 0x8000EBB08000EBAC, 0x000230EF8000EBB4,
-    0x000209A88000EBB8, 0x000221AA000219A9, 0x000221AC000221AB,
-    0x000221AE000221AD, 0x8000EBBC000255AF, 0x8000EBC48000EBC0,
-    0x8000EBCC8000EBC8, 0x0000AD810000AD82, 0x0002D1B80000AD80,
-    0x0002E1BA0002D9B9, 0x000291BC0002E1BB, 0x000295BE00029DBD,
-    0x00022DB0000289BF, 0x000235B2000235B1, 0x00021DB4000205B3,
-    0x0002F1B60002F1B5, 0x000748E10002F1B7, 0x00061CE3000658E3,
-    0x8000EBD000066CE1, 0x8000EBD88000EBD4, 0x80001C4E001C10BE,
-    0x80001C5680001C52, 0x80001C5E80001C5A, 0x0006D0EF0001ECEF,
-    0x0002E9B380001C62, 0x80001C6A80001C66, 0x80001C6E00001DB5,
-    0x0002E9B50002E1B6, 0x0002B1BB80001C72, 0x001D90AC0002A9BA,
-    0x000275BF80001C76, 0x000291BD000299BE, 0x8000EB6C0002A1BC,
-    0x001FB9BC03CD1CE1, 0x001FB9BA001FB1BD, 0x001FB9B8001FB9BB,
-    0x001E15A8001FB9B9, 0x001EA1AA001EA1A9, 0x001EADAC001EBDAB,
-    0x001E9DAE001E95AD, 0x001E91B0001EF9AF, 0x001E91B2001E91B1,
-    0x001EB1B4001E91B3, 0x001EB1B6001EB1B5, 0x001E91B8001EB1B7,
-    0x001E95BA001E91B9, 0x001F55BC001F4DBB, 0x001F55BE001F5DBD,
-    0x001DDDAC001F4DBF, 0x001DE5AE001DEDAD, 0x002410E1001D09AF,
-    0x001DD5AA002D94E1, 0x001DF8E1001DCDAB, 0x001D9DA6001D95A5,
-    0x0002B4E1001F71B5, 0x001F55B50002D8AC, 0x001E49B3001D95A3,
-    0x000019B50002B8E1, 0x80001C7A0002D0AC, 0x001E45B5001E75B6,
-    0x001E35BB000284E1, 0x0002D0AC001E35BA, 0x001E15BF0001E5B5,
-    0x001E25BD001E15BE, 0x001F89A3001E25BC, 0x80001C7E0025358F,
-    0x0002D8AC0002ACE1, 0x001F85A5001FB5A6, 0x001FF5AB000035B5,
-    0x80001C82001FF5AA, 0x001FD5AF00021CE1, 0x001FE5AD001FD5AE,
-    0x0019259E001FE5AC, 0x001BB59C001FC59F, 0x0015719A0019499D,
-    0x0011F5980015699B, 0x001385960017B599, 0x002D499400114597,
-    0x0028F592002C7195, 0x00297990002FA593, 0x0002D0AC00297991,
-    0x80001C8600000DB5, 0x0002DCAC0002ACE1, 0x00000DB580001C8A,
-    0x0002B8E10003B882, 0x0002D8AC000009B5, 0x001CF5A280001C8E,
-    0x001DBDA00001B5A3, 0x80000CC2001DB5A1, 0x800018DE80003B82,
-    0x800018E6800018E2, 0x800018EE800018EA, 0x800018F6800018F2,
-    0x800018FE800018FA, 0x8000190680001902, 0x800019128000190A,
-    0x0026318280001916, 0x0002E4CC002A0583, 0x002CC186000961B8,
-    0x002A1D84002E3587, 0x002F3D8A002B9985, 0x8000191A0010C18B,
-    0x0014D98E8000191E, 0x0012198C0016158F, 0x00167D920012918D,
-    0x800019220018B993, 0x001CD19680001926, 0x001A5D94001E5197,
-    0x001E699A001A3995, 0x8000192A001F219B, 0x8000192E0002F0CC,
-    0x0002799C80001932, 0x0002E8CC00090DB8, 0x0009FDB880001936,
-    0x8000193E8000193A, 0x8000194680001942, 0x8000194E8000194A,
-    0x0002B0CC80001952, 0x8000195A80001956, 0x800019628000195E,
-    0x8000196A80001966, 0x800019728000196E, 0x8000197A80001976,
-    0x800019828000197E, 0x8000198A80001986, 0x800019928000198E,
-    0x80000AC580000AC2, 0x000981B880001C92, 0x0000808580000AE3,
-    0x00001C8780001C96, 0x0002B0CC80001C9A, 0x000288CC0002ACCC,
-    0x0000148D80001CA2, 0x80001CA6000244CC, 0x0000208980000AC8,
-    0x00001C8B000240CC, 0x00083DB880000AD1, 0x80000AEF80000AE9,
-    0x0000109100020CCC, 0x0008EDB880000AFB, 0x80001CAE80000B0A,
-    0x80000B1380000B88, 0x0008C9B880001CB6, 0x80000B1F80000B8E,
-    0x000841B880000B1C, 0x80000B25001FA0E1, 0x80000B28001F50E3,
-    0x80000B34000845B8, 0x000879B880000B4C, 0x80001C9E800009F0,
-    0x80000B7380000B70, 0x80001D7680001D2E, 0x80001D86001ED8EF,
-    0x000468CC80000675, 0x80001DFE800008AF, 0x80001E5680001E2E,
-    0x80001EB680001EB2, 0x0001E0BF00016CBC, 0x80001EBA000144BE,
-    0x80001EC280001EBE, 0x023441AC0000A4BA, 0x02300DAE023919AD,
-    0x8000071E80000854, 0x0221E9A90221EDAA, 0x0226B1A702283DA8,
-    0x023DD1A50226ADA6, 0x023C71A3023CC5A4, 0x0221C1A1000025A2,
-    0x0231ADBC0221CDA0, 0x023C71BE023BB9BD, 0x0231D1B8023C0DBF,
-    0x023199BA0231F1B9, 0x0230F5B4022A7DB8, 0x02312DB60230D9B5,
-    0x02306DB0023131B7, 0x023A09B2023051B1, 0x0231DD8F0230FDB3,
-    0x0230098D80001EC6, 0x023AE98B0006718C, 0x0230BD89023AE18A,
-    0x0230A58702308188, 0x0006598500065186, 0x0234F18300064184,
-    0x023C518102396D82, 0x0226659F023C1D80, 0x0226419D0226559E,
-    0x023D299B0007E59C, 0x0007C599023A7D9A, 0x0000ED97023C3598,
-    0x0007E595023CDD96, 0x023AA1930007E994, 0x023AF59102311992,
-    0x00062D9B02315990, 0x80001ECE80001ECA, 0x0007ED9F80001ED2,
-    0x0006259D0239E59E, 0x80001ED60231159C, 0x80001EDE80001EDA,
-    0x80001EF680001EE2, 0x80001EFE80001EFA, 0x000E9DB880001F02,
-    0x80001F0A80001F06, 0x80001F1280001F0E, 0x80001F1A80001F16,
-    0x80001F3A80001F36, 0x80001F6E80001F52, 0x80001F7680001F72,
-    0x80001F7E80001F7A, 0x023999BB80001F82, 0x023A95B90225E9BA,
-    0x02262DBF000625B8, 0x000629BD02259DBE, 0x00063DB3000631BC,
-    0x023B19B100063DB2, 0x023AA5B700063DB0, 0x00061DB50225D1B6,
-    0x0006B5AB023B49B4, 0x0006A1A90006A9AA, 0x000659AF0006A9A8,
-    0x0006A5AD023B79AE, 0x0006B5A30006A5AC, 0x0006ADA10006ADA2,
-    0x000691A70231BDA0, 0x023B69A50006A9A6, 0x0004C1B3023B5DA4,
-    0x0004B1B10004B9B2, 0x0004C1B70004B9B0, 0x0004D1B50004D9B6,
-    0x80001F860004D9B4, 0x0004F1B90004F9BA, 0x0004D9BF0004F9B8,
-    0x0004D9BD0004D1BE, 0x000429A380001F8A, 0x000401A1000409A2,
-    0x0004DDA7000461A0, 0x00042DA50004D1A6, 0x0004FDAB00042DA4,
-    0x0004FDA90004FDAA, 0x0004C1AF0004FDA8, 0x0004DDAD0004D9AE,
-    0x000755930004DDAC, 0x0007459100075592, 0x0007759700074590,
-    0x0007459500077596, 0x0007559B00074594, 0x000745990007559A,
-    0x0004999F00074598, 0x0007419D0007499E, 0x02288583021E199C,
-    0x022B598100070582, 0x0229C987022B4580, 0x0228918502285986,
-    0x8000003902289584, 0x8000002B80000071, 0x8000003D00072988,
-    0x8000004980000043, 0x023DED9080000053, 0x023C459280000059,
-    0x023C71948000005F, 0x800006618000001F, 0x023820CC8000004F,
-    0x0004ED9A023068CD, 0x8000207E023F5D9B, 0x0004F59E8000006D,
-    0x023D7980023834CC, 0x023D098280000342, 0x023D018480000344,
-    0x023D158680000347, 0x023D35888000208A, 0x023D058A80002091,
-    0x023DF58C000F8CCC, 0x023DB98E800020A8, 0x023685B0800020AF,
-    0x0236F9B2800000C9, 0x0236F9B48000000F, 0x0235C1B680000652,
-    0x000475B88000064F, 0x0234FDBA800000CC, 0x023491BC8000209B,
-    0x02354DBE8000034D, 0x000411A080000575, 0x023635A2800005B0,
-    0x0237EDA4800005D6, 0x0237E1A6800005ED, 0x0237ADA88000060A,
-    0x0237ADAA80000621, 0x0235BDAC8000033C, 0x0236B9AE80002105,
-    0x023DAD80800000DB, 0x023DBD82023AE180, 0x023DC184023B1D82,
-    0x023DC586023AED84, 0x023DC988023B7186, 0x023DCD8A023D2188,
-    0x023E618C023D0D8A, 0x023E498E000BED8C, 0x023E3D90023ED18E,
-    0x023ECD92023E7190, 0x023FCD94023E6D92, 0x023F9996023E1194,
-    0x023F8D98023E6596, 0x000BD19A023FF198, 0x0230359C023F319A,
-    0x02302D9E023F259C, 0x0230F1A0023FF99E, 0x000B49A2023F05A0,
-    0x0230F9A4023F1DA2, 0x023019A6023159A4, 0x02302DA80231E1A6,
-    0x023001AA023115A8, 0x023005AC023101AA, 0x02305DAE023109AC,
-    0x023009B0023109AE, 0x023129B202310DB0, 0x02313DB402325DB2,
-    0x023131B6023235B4, 0x0231C1B8023205B6, 0x0231CDBA023251B8,
-    0x0232E1BC000B31BA, 0x0232CDBE000B15BC, 0x000AF588000B0DBE,
-    0x000D8D8A000AFD89, 0x000C198C000C098B, 0x000C398E000C198D,
-    0x022B2988000C398F, 0x0228D98A022B7589, 0x0228C98C022B3D8B,
-    0x022B3D8E022B118D, 0x000C6998022B718F, 0x022B8D9A000C6999,
-    0x022B859C022BAD9B, 0x022170CC022BA59D, 0x022A09980221CCCD,
-    0x022A399A022AC199, 0x022A299C022A119B, 0x022A199E022AF19D,
-    0x000CB1A8022AF19F, 0x000CB1AA000CB1A9, 0x000CD5AC000CB1AB,
-    0x000CD5AE000CDDAD, 0x022F65A8000CCDAF, 0x022F25AA022E81A9,
-    0x022F19AC022F61AB, 0x022F01AE022F61AD, 0x000C95B8022F61AF,
-    0x022A35BA000C9DB9, 0x022DE1BC022D69BB, 0x022DE1BE022D7DBD,
-    0x022801B8022D1DBF, 0x022829BA022835B9, 0x022B71BC022891BB,
-    0x000391BE000F15BD, 0x000635B8000F1DBF, 0x80000338023F95B8,
-    0x8000083D023F91BA, 0x8000066A023F8DBC, 0x8000066E023FB5BE,
-    0x800008A4023F45B0, 0x8000068A023F41B2, 0x8000068E023F41B4,
-    0x800008A8000989B6, 0x800007E8023EF5A8, 0x80000845023EF9AA,
-    0x80000856023F6DAC, 0x8000085E023F5DAE, 0x80000C09023E51A0,
-    0x800020BD023ECDA2, 0x80000EE4023ED9A4, 0x80000EFB023ED9A6,
-    0x80000F0802393198, 0x80000F100239559A, 0x800006DD0239159C,
-    0x800021010239719E, 0x8000046202386990, 0x8000046E02387D92,
-    0x8000043D02399194, 0x8000044A0239FD96, 0x8000050702386D88,
-    0x800004560238718A, 0x8000050D0238118C, 0x8000055A0238098E,
-    0x8000059B02388180, 0x8000045C02389D82, 0x800005BB02388984,
-    0x000E5D8802387986, 0x022E218A000E6589, 0x022E218C022E358B,
-    0x80000468022E4D8D, 0x0005F8CC80000560, 0x800005F5800005C1,
-    0x8000072080000438, 0x8000073080000723, 0x8000086880000865,
-    0x8000086C000E2199, 0x80000733022E959B, 0x80000884022EB59D,
-    0x80000887022EB59F, 0x80000355800020C1, 0x800020D9800020D5,
-    0x800020E1800020DD, 0x000EE9A8800020E9, 0x000EF9AA000EE9A9,
-    0x000EE9AC000EF9AB, 0x000DB9AE000EE9AD, 0x000A3890000DA1AF,
-    0x000B8892000AF892, 0x000B9494000A9490, 0x800020F1800020E5,
-    0x800020F9000A0C94, 0x0000A189022A6DB9, 0x800020FD022A7DBB,
-    0x80002109022A6DBD, 0x000085838000210D, 0x80002111022A99B1,
-    0x80002118022A65B3, 0x8000212F022A65B5, 0x00030080022A5DB7,
-    0x007C008200011881, 0x007E108480002136, 0x0005F486007F1485,
-    0x0001BC880005FC87, 0x007DEC8A007D2889, 0x0000248C80000642,
-    0x000005B880002140, 0x0002F0CC80002144, 0x007A489200780091,
-    0x000005B8007B4C93, 0x0000D49680002148, 0x000005B80002DCCC,
-    0x0002E4CC8000214C, 0x8000215800001DB8, 0x000005B80002ECCC,
-    0x0002E4CC80002154, 0x8000216400000DB8, 0x000005B80002FCCC,
-    0x0002E4CC80002168, 0x0001A4A8000035B8, 0x8000216C007C80A9,
-    0x0002E4CC00018CAB, 0x8000217000000DB8, 0x0079C0B000028CCC,
-    0x007400B2007844B1, 0x000009B8007658B3, 0x8000217400012CB5,
-    0x800021780002D8CC, 0x0001CCBA000005B8, 0x0077F0BC000150BB,
-    0x0073F8BE007000BD, 0x023D098000717CBF, 0x023D0182023D0D81,
-    0x023D1984023D0583, 0x023D1186023D1D85, 0x023D2988023D1587,
-    0x023D218A023D2D89, 0x023D398C023D258B, 0x023D318E000075B8,
-    0x8000217C023D358F, 0x00029CCC000A7591, 0x0002FCCC000019B8,
-    0x0002DCCC80002180, 0x800021A4000A6997, 0x0002F8CC000005B8,
-    0x00001DB8800021A8, 0x000005B80002ECCC, 0x000009B880002184,
-    0x800021880002E4CC, 0x000AD5A40002F0CC, 0x000ACDA6000AD9A5,
-    0x023DA9A8000035B8, 0x8000218C023DADA9, 0x000005B80002FCCC,
-    0x000298CC80002190, 0x000005B8023DB5AF, 0x0002FCCC800021AC,
-    0x000AB5B4000A91B3, 0x000AADB6000035B8, 0x800021B0000AA1B7,
-    0x000005B80002ECCC, 0x000A75BC80002194, 0x000A75BE0002E8CC,
-    0x00091DBE0003DDB8, 0x022935BC0002C0CC, 0x022841BA0235D9BD,
-    0x022B39B8022FADBB, 0x80002198022FB5B9, 0x8000219C00000DB8,
-    0x0002FCCC000019B8, 0x0002E4CC800021A0, 0x022F35AE000051B8,
-    0x022B3DAC00095DAF, 0x022BE1AA000959AD, 0x022B99A8022FE5AB,
-    0x0002E4CC022BE5A9, 0x0002F8CC0000C5B8, 0x0000D5B8800021B4,
-    0x0002E4CC02246DA3, 0x0009859E0000D5B8, 0x023E81810009899F,
-    0x022D459A0008719D, 0x022D7598022F359B, 0x800021B8022D7999,
-    0x800021C0800021BC, 0x00001DA980002442, 0x8000244602276193,
-    0x0008358E000005AA, 0x022D7D8C00083D8F, 0x022D998A00083D8D,
-    0x022C6988022F458B, 0x80000368022F5589, 0x8000039A80000381,
-    0x800003C6800003B3, 0x00081D80800003D9, 0x0008958700081581,
-    0x000FF0A5800003EC, 0x000F6CA3800003FF, 0x8000042580000412,
-    0x0004D9AF023DE0CC, 0x0004D9AD8000048B, 0x800000DD0004C5AC,
-    0x00089D89007DD18C, 0x0008ED970008A188, 0x80000113007739AF,
-    0x007C559C8000011C, 0x0077D99F80000120, 0x023F699F000351B0,
-    0x8000012880000124, 0x0077F9B3000B30BC, 0x800001308000012C,
-    0x006F999900029882, 0x80000134006F9D9E, 0x000FB88300024482,
-    0x0002B4EF800000E1, 0x800000EB800000E6, 0x0002DCE3800000F0,
-    0x000FB48B006F8D99, 0x800000FA800000F5, 0x000815B7000FA888,
-    0x00080DB5000809B6, 0x800000FF000871B4, 0x000BC9B1000228BE,
-    0x023A21BF000869B0, 0x000815BD0005EDBE, 0x000805BB0003C1BC,
-    0x00081DB9000819BA, 0x007A1980000821B8, 0x02440182007A2581,
-    0x80000104007A2583, 0x007A2186007A2585, 0x80000109024E6587,
-    0x0244A98A007A6589, 0x0244A58C0244B98B, 0x0250A58E0244A18D,
-    0x0244FD90024DF58F, 0x0245C9920244F991, 0x8000010E0245CD93,
-    0x007A119602459595, 0x8000036580000117, 0x0247AD9A02485D99,
-    0x0248319C02482D9B, 0x800004880248359D, 0x007AD1A080000527,
-    0x007AE9A2007ADDA1, 0x024C89A48000057A, 0x0258F1A6800005B5,
-    0x024CB9A8800005DB, 0x0245F5AA800005F2, 0x0244B5AC024C29AB,
-    0x8000060F0244BDAD, 0x0244E5B0024499AF, 0x007AB9B2024495B1,
-    0x02453DB4024569B3, 0x007A59B6007AADB5, 0x007A71B8007A55B7,
-    0x800006260074F5B9, 0x025591BC007A75BB, 0x025269BE025265BD,
-    0x007BA18002559DBF, 0x800003848000036B, 0x8000039D00112999,
-    0x0245058602450985, 0x0075318802453987, 0x800003B602518989,
-    0x800003DC800003C9, 0x80000402800003EF, 0x007BE99080000415,
-    0x007BF192007BF591, 0x007B9994007B8D93, 0x007B8196007B8595,
-    0x007B8998007BBD97, 0x007B919A007B9599, 0x007A6D9C007A759B,
-    0x007A519E007A519D, 0x02453DA0007A519F, 0x007ABDA2007AB5A1,
-    0x024869A4007AB5A3, 0x007AD9A6007AA9A5, 0x007AF1A8007AD5A7,
-    0x007AF1AA024DA5A9, 0x024431AC007ACDAB, 0x0245A5AE0245BDAD,
-    0x0075D1B0024419AF, 0x007AFDB2007AF5B1, 0x024829B4007AF5B3,
-    0x007A99B6007AE9B5, 0x007AB1B8007A95B7, 0x007AB1BA024DE5B9,
-    0x024471BC007A8DBB, 0x0245E5BE0245FDBD, 0x0249BD82024459BF,
-    0x0075BD800249CD83, 0x0075DD86024E4581, 0x0075A5840075C187,
-    0x0075CD8A0075D985, 0x0075F5880075D18B, 0x0078398E0075C989,
-    0x007E158C80000428, 0x0247AD920075C18D, 0x02420990024E4593,
-    0x02470D9602466991, 0x024C2D940247D597, 0x024A599A0252C195,
-    0x0247CD98024A399B, 0x8000048E0247AD99, 0x0076759C80000138,
-    0x800001408000013C, 0x8000014880000144, 0x800001508000014C,
-    0x8000015880000154, 0x800001608000015C, 0x0078B9A880000164,
-    0x8000016C80000168, 0x8000017480000170, 0x8000017C80000178,
-    0x8000018480000180, 0x8000018C80000188, 0x8000019480000190,
-    0x8000019C80000198, 0x8000063A80002122, 0x00001DB580000644,
-    0x8000064700001DB4, 0x007B1083000005B6, 0x007E58B3800021D8,
-    0x007E58B5007E04B2, 0x007C6885007E4CB4, 0x800021E2800021DE,
-    0x800021EA800021E6, 0x800021EE007C4088, 0x007C608D800021F2,
-    0x800021FA800021F6, 0x80002202800021FE, 0x80002206007D98A2,
-    0x8000220E8000220A, 0x007E54A980002212, 0x007E68AB007DA0A8,
-    0x8000221A80002216, 0x800022228000221E, 0x007D089180002226,
-    0x007CCCA18000222A, 0x8000222E007D7492, 0x007C24A580002232,
-    0x80002236007C2CA4, 0x8000223E8000223A, 0x8000224680002242,
-    0x8000224E8000224A, 0x007C40B380002252, 0x007D7C83007C6CB2,
-    0x007C3CB7007D8882, 0x007D4487007C58B6, 0x007D80BB007D6086,
-    0x80002256007D9CBA, 0x8000225E8000225A, 0x007DA0BD80002262,
-    0x007BA580007D94BC, 0x007BB182007BB181, 0x00040184007BDD83,
-    0x0004498600043585, 0x0004418800044187, 0x0004558A00045D89,
-    0x0004A58C0004A18B, 0x00049D8E0004A98D, 0x0004A9900004858F,
-    0x0005419200054191, 0x0005799400055D93, 0x0005699600057995,
-    0x0005599800056997, 0x00065D9A00055999, 0x0006059C00061D9B,
-    0x0006F99E0006099D, 0x000609A00006F99F, 0x000619A2000609A1,
-    0x000609A4000619A3, 0x000639A6000609A5, 0x00060DA8000639A7,
-    0x00061DAA000615A9, 0x00060DAC000615AB, 0x00067DAE000675AD,
-    0x00060DB0000675AF, 0x00061DB2000615B1, 0x00060DB4000615B3,
-    0x024499B6000635B5, 0x0240E9B80240D9B7, 0x0240CDBA0240F9B9,
-    0x024071BC0240BDBB, 0x024045BE024061BD, 0x02405D800241B9BF,
-    0x0240AD8202408581, 0x0240FD840240CD83, 0x0242DD86024D2D85,
-    0x024D7D88024D5D87, 0x024DC18A024D8589, 0x0248E98C024CD58B,
-    0x0249318E0249358D, 0x024501900249258F, 0x0241419202414191,
-    0x0241759402415193, 0x0241D99602410595, 0x007171980255C197,
-    0x0240359A0255C999, 0x0240D59C0240ED9B, 0x0240959E0240B59D,
-    0x024245A0024D459F, 0x024DD5A2024DC5A1, 0x024D79A4007181A3,
-    0x025401A6024C6DA5, 0x025469A802499DA7, 0x00721DAA0249BDA9,
-    0x8000226A80002266, 0x800022728000226E, 0x8000227A80002276,
-    0x800022828000227E, 0x8000228A80002286, 0x800022928000228E,
-    0x8000229A80002296, 0x80000B0D8000229E, 0x80000B4080001CB2,
-    0x800022AE800022AA, 0x0003E5A0800022B2, 0x0003E1A2800022BA,
-    0x0003E1A40003E1A3, 0x80000AAA800022BE, 0x80000AAD0003E9A7,
-    0x800022C20003DDA9, 0x800022C60003D1AB, 0x007541AE0074C5AD,
-    0x00756DB00074D1AF, 0x00004DB2800022CA, 0x800022D2800022CE,
-    0x800022D6000041B5, 0x800022DE800022DA, 0x800022E6800022E2,
-    0x0256DDBC800022EA, 0x000061BE024F8DBD, 0x007995B100005DBF,
-    0x0079ADB3007989B0, 0x0079D9B5007995B2, 0x0079CDB70079A5B4,
-    0x0079C9B90079D1B6, 0x0079D1BB0079C1B8, 0x007925BD0079D9BA,
-    0x007919BF00793DBC, 0x024D7DA1007921BE, 0x007175A3024A8DA0,
-    0x00711DA5024D05A2, 0x007105A7007111A4, 0x007801A9007109A6,
-    0x007801AB007101A8, 0x007801AD007809AA, 0x0079FDAF007819AC,
-    0x00041D80007809AE, 0x00042D8200042581, 0x0004218400042583,
-    0x0004258600042185, 0x0004658800045D87, 0x0004798A00046D89,
-    0x0004698C0004798B, 0x0004558E0004698D, 0x0004359000044D8F,
-    0x0004359200043D91, 0x0004C1940004C993, 0x0004F9960004C195,
-    0x0004C9980004F997, 0x0004D99A0004C999, 0x0004C99C0004D99B,
-    0x0004919E0004C99D, 0x000471A00004919F, 0x000225A2000471A1,
-    0x000225A400023DA3, 0x000225A600022DA5, 0x000265A800025DA7,
-    0x000265AA00026DA9, 0x000265AC00027DAB, 0x000265AE00026DAD,
-    0x800022EE00025DAF, 0x800022F6800022F2, 0x800022FE800022FA,
-    0x8000230680002302, 0x8000230E8000230A, 0x8000231680002312,
-    0x8000231E8000231A, 0x8000232680002322, 0x8000232E8000232A,
-    0x8000233680002332, 0x8000233E8000233A, 0x8000234200007DA0,
-    0x000055AD000071A2, 0x8000234A80002346, 0x800023528000234E,
-    0x000049AB80002356, 0x8000235E8000235A, 0x8000236680002362,
-    0x8000236E8000236A, 0x8000237680002372, 0x8000237A000025B2,
-    0x800023828000237E, 0x8000238A80002386, 0x800023928000238E,
-    0x800032FA80002436, 0x80003B660000AD84, 0x80002BC60000A186,
-    0x80002BEA00008D80, 0x80002BF60000B982, 0x80002C060000998C,
-    0x80002C0E0000ED8E, 0x80002C1A00008588, 0x80002C220000898A,
-    0x000024BE00008994, 0x80002C2A00009D96, 0x80002C3600009190,
-    0x80002C9100009592, 0x80002CA90000859C, 0x80002CB90000899E,
-    0x80002CC90000A198, 0x000191800000A59A, 0x0001A98200019D81,
-    0x0001A1840001A583, 0x0001D9860001AD85, 0x0001F9880001CD87,
-    0x0001C18A0001C589, 0x0001C98C0001DD8B, 0x0003358E0003318D,
-    0x00035D900003298F, 0x0003659200036191, 0x00036D9400037993,
-    0x0003159600031195, 0x00033D9800030997, 0x0003059A00030199,
-    0x00030D9C0003199B, 0x0003F59E0003F19D, 0x000325A00003E19F,
-    0x00033DA2000329A1, 0x000355A4000331A3, 0x00034DA6000359A5,
-    0x000345A8000341A7, 0x00035DAA000349A9, 0x0002B5AC000351AB,
-    0x0002B5AE0002BDAD, 0x0002D5B00002ADAF, 0x0002D5B20002DDB1,
-    0x0002F5B40002EDB3, 0x0002F5B60002FDB5, 0x0002D5B80002EDB7,
-    0x0002D5BA0002DDB9, 0x0002B5BC0002ADBB, 0x0002B5BE0002BDBD,
-    0x000DF9B20002ADBF, 0x000DFDB080002CD9, 0x000211B680002CE9,
-    0x000DE5B480002D09, 0x000229BA80002D0D, 0x00022DB880002D11,
-    0x000221BE80002D45, 0x000235BC80002D6D, 0x000D99A280002D7D,
-    0x000D9DA080002D85, 0x000DB1A680002D89, 0x000D85A480002D95,
-    0x000D89AA80002DB1, 0x000D8DA880002DB5, 0x000D81AE80002DCD,
-    0x000D95AC80002DD1, 0x000CE99280002DD9, 0x000CED9080002DE9,
-    0x000D519680002E85, 0x000CF594001AB59E, 0x000D699A80002E99,
-    0x000D6D9880002ED9, 0x000D619E80002EED, 0x000D759C80002EF5,
-    0x000CF18280002EF9, 0x000CE18080002F09, 0x000C9D8680002F2C,
-    0x000C9D8480002F58, 0x000CB98A80002F64, 0x000CB98880002F84,
-    0x000C918E80002F90, 0x000CA58C80002F94, 0x0008688280002F98,
-    0x80002FB480002FB0, 0x80002FEC80002FE0, 0x80002FFC80002FF4,
-    0x8000301B8000300C, 0x8000304F80003023, 0x8000305B80003053,
-    0x800030678000305F, 0x8000307F8000306F, 0x800031138000309B,
-    0x8000312B80003127, 0x800031AF800031AB, 0x800031C7800031C3,
-    0x800031D7800031CF, 0x800031EB800031DF, 0x8000323180003221,
-    0x800032C280003245, 0x800032DE800032CE, 0x800032F6800032EA,
-    0x80003302800032FE, 0x8000330A80003306, 0x800033C68000330E,
-    0x8000341A8000340E, 0x800034228000341E, 0x8000342E80003426,
-    0x8000344680003432, 0x8000346A80003462, 0x800034AE800034AA,
-    0x800034BA800034B6, 0x800034C6800034C2, 0x800034EE800034EA,
-    0x8000350E8000350A, 0x8000351680003512, 0x8000354E8000352A,
-    0x8000355680003552, 0x800035BA8000357A, 0x800035DE800035BE,
-    0x800035F6800035EE, 0x8000363A80003616, 0x800036868000367E,
-    0x800036A68000369A, 0x800036B6800036AE, 0x800036BE800036BA,
-    0x800036DA800036D6, 0x800036FE800036F6, 0x8000370A80003706,
-    0x8000371E80003716, 0x8000372680003722, 0x800037328000372E,
-    0x800037E6800037D2, 0x8000381E8000381A, 0x8000385280003826,
-    0x800038628000385A, 0x800038A68000386A, 0x800038B2800038AA,
-    0x800038BA800038B6, 0x800038EA800038E6, 0x80003906800038F2,
-    0x800039268000390A, 0x800039328000392E, 0x8000397680003956,
-    0x8000398E8000398A, 0x800039C280003996, 0x800039C600018CBE,
-    0x00015580800039D6, 0x0001558200015D81, 0x0001758400016D83,
-    0x0001758600017D85, 0x0001558800016D87, 0x0001558A00015D89,
-    0x0001B58C0001AD8B, 0x0001B58E0001BD8D, 0x0001D5900001AD8F,
-    0x0001D5920001DD91, 0x0001F5940001ED93, 0x0001F5960001FD95,
-    0x0001D5980001ED97, 0x0001D59A0001DD99, 0x0001B59C0001AD9B,
-    0x0001B59E0001BD9D, 0x000155A00001AD9F, 0x000155A200015DA1,
-    0x000175A400016DA3, 0x000175A600017DA5, 0x000159A800016DA7,
-    0x0002ADAA000159A9, 0x0002BDAC0002A5AB, 0x0002ADAE0002A5AD,
-    0x0002DDB00002A5AF, 0x0002EDB20002E5B1, 0x0002FDB40002E5B3,
-    0x0002EDB60002E5B5, 0x0002DDB80002E5B7, 0x0002ADBA0002A5B9,
-    0x0002BDBC0002A5BB, 0x0002ADBE0002A5BD, 0x000249BD0002A5BF,
-    0x000259BF000241BC, 0x000229B9000241BE, 0x000259BB000221B8,
-    0x000209B5000221BA, 0x000219B7000201B4, 0x000229B1000201B6,
-    0x000219B3000221B0, 0x000249AD000221B2, 0x000259AF000241AC,
-    0x0005A9A9000241AE, 0x000259AB0005A1A8, 0x000589A50005A1AA,
-    0x000599A7000581A4, 0x0005A9A1000581A6, 0x000599A30005A1A0,
-    0x0005499D0005A1A2, 0x0005599F0005419C, 0x000529990005419E,
-    0x0005599B00052198, 0x000509950005219A, 0x0005199700050194,
-    0x0005299100050196, 0x0005199300052190, 0x0005498D00052192,
-    0x0005598F0005418C, 0x0005A9890005418E, 0x0005598B0005A188,
-    0x000589850005A18A, 0x0005998700058184, 0x0005A98100058186,
-    0x000599830005A180, 0x001F259E0005A182, 0x80003A06800039FA,
-    0x80003A2E80003A16, 0x80003A3A80003A36, 0x80003A4280003A3E,
-    0x80003A5280003A4E, 0x80003A7680003A5A, 0x80003A7E80003A7A,
-    0x80003A9A80003A96, 0x80003AB680003AA2, 0x80003ABE80003ABA,
-    0x80003ACA80003AC6, 0x80003AD280003ACE, 0x80003AE680003AD6,
-    0x80003B0280003AFE, 0x80003B1280003B0E, 0x0000DD8080003B16,
-    0x0000ED820000E581, 0x0000FD840000E583, 0x0000ED860000E585,
-    0x0000DD880000E587, 0x0000758A00007D89, 0x0000558C00004D8B,
-    0x0000558E00005D8D, 0x0000359000004D8F, 0x0000359200003D91,
-    0x001FD194001C0193, 0x80003B1A001FD195, 0x80003B2680003B22,
-    0x80003B3680003B2A, 0x80003B4680003B3E, 0x80003B5280003B4A,
-    0x80003B6280003B56, 0x0250018080003B6A, 0x80002D998000246A,
-    0x8000255280002D9D, 0x800024E0006D98E3, 0x00027C82800024E4,
-    0x00000599001A1CEF, 0x00027CBE800024E8, 0x000248800000099E,
-    0x0001048200035C81, 0x0005908400074C83, 0x0008D886000B5485,
-    0x000C00880000CD99, 0x000D688A000F6489, 0x0038008C00336C8B,
-    0x003B788E003A748D, 0x001988E300397C8F, 0x00020CE30002A4E3,
-    0x800024EC0002B482, 0x0000099E000274BE, 0x00026C82000234E3,
-    0x00000D99800024F0, 0x001EBCEF000290EF, 0x0002D4EF000274EF,
-    0x00021CE30000199E, 0x00000599000200BE, 0x0000DDB6800024F4,
-    0x0000159E00027882, 0x0000F5BA00027CBE, 0x001ABDB800027482,
-    0x0002D4EF0000FDB9, 0x800024F800000D99, 0x006D2486000380BE,
-    0x000278820001FD9E, 0x000210E30003E0E3, 0x800024FC00000D99,
-    0x0002C4EF00024CBE, 0x0000319E00022C8F, 0x00026C820002208D,
-    0x000079990002A48B, 0x80002500000338EF, 0x000260BE0003C497,
-    0x0000119E00020095, 0x000260E300024493, 0x0002748200025891,
-    0x0002D8EF0003789F, 0x0000099900036C9D, 0x800025040003A09B,
-    0x000004A60003E099, 0x000004A40000FD9E, 0x000230E30002E4E3,
-    0x80002508000268BE, 0x0002E8EF0003C4A1, 0x00026082000068AF,
-    0x000208BE00003599, 0x000028A800021882, 0x0000599E0002D0EF,
-    0x00000D998000250C, 0x000004B20000C4B5, 0x00027CBE000268E3,
-    0x00026C820000199E, 0x000005990002C4EF, 0x00001D9E80002510,
-    0x0000E8B80000A0BB, 0x0001E8E300019482, 0x0001FCE300018CBE,
-    0x8000251400001D99, 0x0000099E000274BE, 0x00026082000234E3,
-    0x8000251800000D99, 0x0002CCEF0002E8EF, 0x0002D4EF00001D9E,
-    0x000204E3000270BE, 0x000019990000199A, 0x8000251C00026082,
-    0x00023CE3000248BE, 0x00003D9F80002520, 0x0002ECEF0000359E,
-    0x0002DCEF00028882, 0x8000252880002524, 0x0000F599000294BE,
-    0x00024882000204E3, 0x0000E59E0000F59A, 0x8000252C0000ED9F,
-    0x000218BE00022CE3, 0x0000199A00003D99, 0x800025300002A4EF,
-    0x00020C820002ECEF, 0x0000599E80002534, 0x800025380000759F,
-    0x00006D9900018D99, 0x000204E300024882, 0x00005D9E0000159A,
-    0x000310EF00002D9F, 0x000268BE0001AD9E, 0x000005990001E99A,
-    0x000270BE80002543, 0x0001F59E000218E3, 0x000384BE0000119F,
-    0x000041990003BC82, 0x0000059B80002547, 0x00025482800000BD,
-    0x0000759F800000C2, 0x0003B09D0002DCEF, 0x8000255E0000119C,
-    0x00027CBE8000254B, 0x000254820000159E, 0x80002A9A000274E3,
-    0x00001599000204E3, 0x000298820002D4EF, 0x000270BE0002E4EF,
-    0x00000D9E80002635, 0x000408B3000000D3, 0x0003DCB10002F0EF,
-    0x00045CB7000250BE, 0x000400B500000999, 0x0004ACBB80002698,
-    0x000470B90000399E, 0x0004A0BF00027482, 0x0004BCBD0002D4E3,
-    0x800026DB00007D99, 0x000260BE000260E3, 0x0000119E80003B5A,
-    0x00023CE300030CA6, 0x000330AB00040199, 0x000510EF0004399E,
-    0x0003B8AF00025482, 0x0003C0AD0002C4EF, 0x00059CBE00059082,
-    0x000005990005F0E3, 0x000270BE80002731, 0x0000159E000200E3,
-    0x0000099900025882, 0x800027560002F0EF, 0x0000199E000264BE,
-    0x000260820002DCEF, 0x000270E300007999, 0x00027CBE8000276B,
-    0x800027730000099E, 0x00026082000204E3, 0x000270BE00000D99,
-    0x0000399E0002FCEF, 0x0002C0EF00027082, 0x00000D998000278F,
-    0x0002E4E3000280BE, 0x000214E30000F19E, 0x00001D9980002797,
-    0x0000159E00026082, 0x00026C82000234E3, 0x00001D99000244BE,
-    0x000234EF000290EF, 0x0002E4EF8000279F, 0x000200E30000099E,
-    0x00027C82000200BE, 0x00001D99800027B1, 0x0000159E000274BE,
-    0x800027BD000230E3, 0x0000099900026882, 0x0002DCEF0002E0EF,
-    0x0000199E000274BE, 0x00026C82800027C5, 0x0001F59900024881,
-    0x00024084000250BE, 0x0003888600024082, 0x0003E888000214E3,
-    0x800027DB0001D99E, 0x000019990002D8EF, 0x000200BE00020082,
-    0x0000699E0003D88F, 0x00034C92000258E3, 0x00001D9980002819,
-    0x0002D8EF00030095, 0x0000E8980000399A, 0x000254E38000282A,
-    0x000280820000E89B, 0x000280BE0002B4EF, 0x0000CD9E8000284C,
-    0x0002F0E30000E19F, 0x000240BE80002857, 0x0000159A0000E999,
-    0x800028A700022882, 0x0000D59E800028BB, 0x00022CBE0000219F,
-    0x0002D0EF0001D082, 0x00001CB0000078AF, 0x0000C0B20000C8B1,
-    0x00007199800028DD, 0x00024CE3800028EB, 0x0000459F0000799A,
-    0x0002A0EF0000459E, 0x80002933000184BE, 0x800029410000E4BD,
-    0x0003CD9900032DBF, 0x000218E30001B8E3, 0x0003DD9E0003D99A,
-    0x000258820003C19F, 0x80002A9E000224BE, 0x0003D19A0003D599,
-    0x0000119900027882, 0x0002DCEF000120EF, 0x0003B19E000208BE,
-    0x00026CE30003BD9F, 0x00020CE300006D9E, 0x0000319900026882,
-    0x0002C4EF000240BE, 0x80002AA200000D9E, 0x0002788200001999,
-    0x0002BCEF80002AA6, 0x0002F8E300028CBE, 0x0000FD9E00021CE3,
-    0x0002788200001999, 0x000240BE80002AAA, 0x0000359E000230E3,
-    0x0000159E00027882, 0x000234EF00029CEF, 0x000005990002D4EF,
-    0x000000B080002AB6, 0x000218BE000000B1, 0x80002716000000B3,
-    0x80001ABE000000B5, 0x000000B880001AC2, 0x000000BA000000B9,
-    0x80001AC6000000BB, 0x000000BE000000BD, 0x000000C0000000BF,
-    0x000F09B2000C01B1, 0x000F15B40001D5B3, 0x0001E9B60001C9B5,
-    0x000F35B8000F0DB7, 0x000069BA000F65B9, 0x000069BC000069BB,
-    0x00005DBE000055BD, 0x80001ACA00004DBF, 0x000000D280001ACE,
-    0x000000D4000000D3, 0x80001AD2000000D5, 0x80001AD6000000D7,
-    0x000000DA80001ADA, 0x80001ADE000000DB, 0x000000DE800019EA,
-    0x000000E0800019FE, 0x000000E280001A4A, 0x80001A5280001A4E,
-    0x000000E6000000E5, 0x80001A56000000E7, 0x80001A5E80001A5A,
-    0x80001A6680001A62, 0x80001A6A000000ED, 0x000000F0000000EF,
-    0x80001A7280001A6E, 0x000000F4000000F3, 0x000000F680001A76,
-    0x80001A7A80002707, 0x80001A8280001A7E, 0x000000FC80001A86,
-    0x000000FE000000FD, 0x8000199A80001A8A, 0x80001A8E8000199E,
-    0x800019A280001A92, 0x80001A96800019A6, 0x80001A4680001A9A,
-    0x800019E6800019E2, 0x80001AE680001AE2, 0x80001AEE80001AEA,
-    0x800019AA80001AF2, 0x80001AF6800019AE, 0x800019B280001AFA,
-    0x800019EE800019B6, 0x80001B0280001AFE, 0x800019F6800019F2,
-    0x80001A02800019FA, 0x80001A0A80001A06, 0x027DE0E180001A0E,
-    0x026954E3800019BA, 0x800019BE80001A12, 0x00415CEF800019C2,
-    0x004864E1004E44EF, 0x004970E303A3A0E1, 0x800019C6005CF8E3,
-    0x027D4CEF80001A16, 0x800019CA80001A1A, 0x80001A1E800019CE,
-    0x800019D280001A22, 0x80001A26800019D6, 0x80001A2E80001A2A,
-    0x80001A3680001A32, 0x80001B0A80001B06, 0x80001A3E80001A3A,
-    0x000E758080001A42, 0x0001598200014181, 0x000E6D8400015983,
-    0x0001898600019185, 0x0001918800018987, 0x0001ED8A0001E589,
-    0x0001E58C0001FD8B, 0x00021D8E0002158D, 0x000ECD90000EB98F,
-    0x000EE992000ED591, 0x000EF994000EE993, 0x000EE996000EF995,
-    0x000EA598000EE597, 0x000EBD9A000EAD99, 0x000E959C000E8D9B,
-    0x000E959E000E9D9D, 0x000189A0000E819F, 0x000191A2000189A1,
-    0x0001B1A4000191A3, 0x0001B1A60001B1A5, 0x000191A80001B1A7,
-    0x000191AA000191A9, 0x0001F9AC000191AB, 0x0001E1AE0001F9AD,
-    0x0001A1B00001E1AF, 0x0001A1B20001A1B1, 0x0001A1B40001A1B3,
-    0x0001A1B60001A1B5, 0x00016DB80001A1B7, 0x000155BA000151B9,
-    0x00013DBC000125BB, 0x000125BE00012DBD, 0x80001A9E00011DBF,
-    0x80001AA680001AA2, 0x80001AAE80001AAA, 0x80001AB680001AB2,
-    0x80002BD280001ABA, 0x80002BD680002E89, 0x80002BDA80002BEE,
-    0x80002BE2800034CA, 0x80002F1480002BCA, 0x800001A080002E9D,
-    0x800001AC800001A6, 0x800001B8800001B2, 0x800001C4800001BE,
-    0x800001D0800001CA, 0x800001DC800001D6, 0x800001E8800001E2,
-    0x00026580800001EE, 0x0002658200026D81, 0x0002658400027D83,
-    0x0002658600026D85, 0x0000A58800009D87, 0x0000A58A0000AD89,
-    0x0000A58C0000BD8B, 0x0000A58E0000AD8D, 0x800002D2800002CC,
-    0x00125992800002D8, 0x0000F19400120D93, 0x0000F1960000F195,
-    0x0000D1980000F197, 0x0012599A0000D199, 0x0000CD9C0000D59B,
-    0x0000BD9E0000B59D, 0x034774E10012259F, 0x036BCCE380001996,
-    0x030C04E1033A50EF, 0x033BC0E3036EDCE1, 0x800019DA031748E3,
-    0x033AECEF800019DE, 0x800002DE030D3CEF, 0x036FE8EF0333F0E1,
-    0x03FE54E10323E0E3, 0x03EEFCE303D2CCE1, 0x038484E103C394E3,
-    0x03A044E30398A8E1, 0x800002E4038F84E3, 0x03A098EF800002EA,
-    0x800002F003843CEF, 0x03C2BCEF800002F6, 0x020774E10337ACEF,
-    0x023A2CE3026E00E1, 0x02CCECE10267D8E3, 0x02FA4CE3023220E1,
-    0x8000030E02D340E3, 0x02E93CEF8000031A, 0x80000320020370EF,
-    0x020A6CEF80000326, 0x029E74E1027ED8EF, 0x02921CE302E98CE1,
-    0x035B74E102E398E3, 0x03478CE3029780E1, 0x8000032C02AB88E3,
-    0x03472CEF80000332, 0x80000314028330EF, 0x028EACEF800002FC,
-    0x8000030502928CEF, 0x8000020C800001F4, 0x8000025480000200,
-    0x8000023080000212, 0x8000022A800001FA, 0x8000023C80000206,
-    0x8000028A8000026C, 0x8000027880000284, 0x8000025A800002C6,
-    0x8000027E80000266, 0x8000029C80000272, 0x8000029080000248,
-    0x00000059800002BA, 0x80000236800002A2, 0x8000024E80000218,
-    0x8000029680000260, 0x800002C08000021E, 0x800002A880000242,
-    0x800002AE80000224, 0x80002E3D800002B4, 0x800035FE8000301F,
-    0x8000049E8000081A, 0x800004BC800004AD, 0x800004DA800004CB,
-    0x800004EC800004E3, 0x800004FE800004F5, 0x800005338000052A,
-    0x8000053F8000053C, 0x8000054580000542, 0x80004F9E80003D3E,
-    0x80006B2E800058CE, 0x80007D8E8000745E, 0x0001898080008FEE,
-    0x0001998200018981, 0x0001898400019983, 0x0016458600018985,
-    0x8000A24E0001BD87, 0x8000BDE58000AB7E, 0x8000D04C8000C71C,
-    0x8000E2AC8000D97C, 0x0001E5908000BE29, 0x0001E5920001ED91,
-    0x0001E5940001FD93, 0x0001E5960001ED95, 0x0001A59800019D97,
-    0x0001A59A0001AD99, 0x0001A59C0001BD9B, 0x0001A59E0001AD9D,
-    0x000685A000019D9F, 0x000599A2000581A1, 0x0005D5A40005CDA3,
-    0x00051DA60005C5A5, 0x000511A8000509A7, 0x000579AA000561A9,
-    0x000691AC000579AB, 0x00015DAE000691AD, 0x00012DB0000155AF,
-    0x00013DB2000135B1, 0x00012DB4000135B3, 0x0000FDB60000F5B5,
-    0x0000CDB80000F5B7, 0x0000DDBA0000D5B9, 0x0000CDBC0000D5BB,
-    0x0003E5BE000009BD, 0x0005B1988000B132, 0x0005E99A0005B199,
-    0x0005E19C0005E99B, 0x0002199E0005E19D, 0x0004E5908000A7FE,
-    0x0005499200051591, 0x0005419400054993, 0x0005999600054195,
-    0x0004298800059997, 0x0004398A00042989, 0x0004298C0004398B,
-    0x0004E58E0004298D, 0x0004298000049D8F, 0x0004398200042981,
-    0x0004298400043983, 0x0004198600042985, 0x0002E9B800041987,
-    0x0002F9BA0002E9B9, 0x0002E9BC0002F9BB, 0x000219BE0002E9BD,
-    0x0002E5B0000219BF, 0x0002E5B20002EDB1, 0x0002E5B40002FDB3,
-    0x0002E5B60002EDB5, 0x0002E5A80002D9B7, 0x0002E5AA0002EDA9,
-    0x0002E5AC0002FDAB, 0x0002E5AE0002EDAD, 0x0002E5A000029DAF,
-    0x0002E5A20002EDA1, 0x0002E5A40002FDA3, 0x0002E5A60002EDA5,
-    0x001791800002DDA7, 0x000529820017C581, 0x003C018400052983,
-    0x003F1D86003F1585, 0x003F29880017ED87, 0x0014518A00157189,
-    0x0014958C0014858B, 0x0013D58E0014498D, 0x001431900015A18F,
-    0x003F5992003F4991, 0x003F4994003F5993, 0x003F7996003F4995,
-    0x003F4998003F7997, 0x003F599A003F4999, 0x003F459C0015A19B,
-    0x003F459E003F4D9D, 0x003FC5A0003F3D9F, 0x003FC5A2003FCDA1,
-    0x0005B9A4003FDDA3, 0x0005A9A60005B9A5, 0x003FF9A8003FC9A7,
-    0x003FC9AA003FF9A9, 0x003FD9AC003FC9AB, 0x003FC9AE003FD9AD,
-    0x003FB9B0003FC9AF, 0x003F49B2003FB9B1, 0x003F59B4003F49B3,
-    0x003F49B6003F59B5, 0x003F79B8003F49B7, 0x003F49BA003F79B9,
-    0x003F59BC003F49BB, 0x003F49BE003F59BD, 0x80002CBD003F49BF,
-    0x80002C1280002BCE, 0x8000323580003275, 0x80002E0980003582,
-    0x800038BE8000343A, 0x80002D5180003596, 0x800034D2800035C6,
-    0x80002C8D80003946, 0x8000332E80002DB9, 0x80002C5180003A5E,
-    0x800032D280002CE1, 0x80002DFD80002FF0, 0x80002F6C80002D8D,
-    0x8000352280002F60, 0x800038D280002C4D, 0x80002F0180002DA9,
-    0x8000054B80000548, 0x800005518000054E, 0x800005808000057D,
-    0x8000058680000583, 0x8000058C80000589, 0x800005928000058F,
-    0x8000059880000595, 0x80000479800005B8, 0x8000056B80000518,
-    0x800005CC800005A6, 0x80000600800005E3, 0x8000062E80000617,
-    0x8000038E80000375, 0x8000070E800003A7, 0x800006D8800006D4,
-    0x800025668000077D, 0x0001B9808000259F, 0x0001C9820001B981,
-    0x0001D9840001C983, 0x0001C9860001D985, 0x0001F9880001C987,
-    0x0001C98A0001F989, 0x0001D98C0001C98B, 0x0001C98E0001D98D,
-    0x0001B9900001C98F, 0x000401920001B991, 0x0003559400034D93,
-    0x00055D9600055595, 0x00056D9800055597, 0x00057D9A00057599,
-    0x00056D9C0005759B, 0x00053D9E0005359D, 0x0005CDA00005359F,
-    0x0005DDA20005D5A1, 0x000399A40005D5A3, 0x000389A6000399A5,
-    0x0003B9A8000389A7, 0x000529AA0005F9A9, 0x000539AC000529AB,
-    0x0003A5AE000539AD, 0x0003E5B000039DAF, 0x0003E5B20003EDB1,
-    0x0003E5B40003FDB3, 0x0003E5B60003EDB5, 0x0003A5B800039DB7,
-    0x0003A5BA0003ADB9, 0x000565BC00057DBB, 0x000565BE00056DBD,
-    0x800025D00003ADBF, 0x800027E6800025F5, 0x800027F5800027F1,
-    0x800027FD800027F9, 0x8000299B8000295C, 0x800029C7800029C3,
-    0x800029F4800029EC, 0x80002A2280002A10, 0x80002A2A80002A26,
-    0x80002A5D80002A42, 0x80002C4680002A78, 0x800025778000256A,
-    0x8000259180002584, 0x800025B0800025A3, 0x800025D4800025C2,
-    0x800025F9800025E4, 0x8000261180002603, 0x800026288000261B,
-    0x8000264380002639, 0x800026A38000269C, 0x800026AD80002651,
-    0x800026658000265E, 0x8000268880002675, 0x800026E9800026DF,
-    0x800026CE800026BE, 0x8000271D800026FD, 0x8000273980002727,
-    0x8000275E80002746, 0x800027858000277B, 0x800027C9800027A3,
-    0x800027D4800027DF, 0x80002801800027EA, 0x000000808000280F,
-    0x8000283E8000282E, 0x80002C268000281D, 0x8000286B8000285B,
-    0x8000285080002875, 0x8000289080002880, 0x8000289D800028AB,
-    0x800028EF800028C3, 0x800028D3800028F6, 0x8000290A80002900,
-    0x80002945800028E1, 0x8000291880002937, 0x8000291F80002952,
-    0x8000296080002929, 0x800029778000296D, 0x8000298B80002981,
-    0x800029AC8000299F, 0x800029CB800029B3, 0x800029DF800029D2,
-    0x80002A02800029F8, 0x80002A2E80002A14, 0x80002A4680002A3B,
-    0x80002A6180002A50, 0x80002A8080002A68, 0x8000047E8000035B,
-    0x800005708000051D, 0x800005D1800005AB, 0x80000605800005E8,
-    0x800006338000061C, 0x800003948000037B, 0x800003C0800003AD,
-    0x800003E6800003D3, 0x8000260D800003F9, 0x800026BA8000264D,
-    0x8000270B800026F9, 0x8000275A80002735, 0x8000040C8000276F,
-    0x800004328000041F, 0x800004A780000498, 0x800004C5800004B6,
-    0x80002777800004D4, 0x8000279B80002793, 0x800027B9800027AD,
-    0x800027D0800027C1, 0x800007148000280B, 0x80000672800006B4,
-    0x800007F680000681, 0x800006C080000803, 0x80002848800006C3,
-    0x800028BF8000287C, 0x80002A7C80002914, 0x80002A8E80002A8A,
-    0x800006C780002A92, 0x800030108000072D, 0x80002F0D800031FE,
-    0x80003279800031EF, 0x800007DC80000800, 0x8000079080000C15,
-    0x8000073D8000073A, 0x800006FB80000793, 0x8000074080000692,
-    0x800007DF80000806, 0x80000C1D80000C19, 0x800007458000079C,
-    0x800007488000071B, 0x800006FE8000079F, 0x80000C218000085A,
-    0x800006BD800007A7, 0x800006F28000074F, 0x80000C25800007E5,
-    0x8000069C800007AA, 0x800007AD80000752, 0x8000078A8000069F,
-    0x800025BE80000755, 0x800006A3800007B1, 0x800007598000078D,
-    0x800007D3800007CD, 0x800007B98000075D, 0x8000082880000702,
-    0x800008348000082C, 0x800007EB80000817, 0x800007C080000C29,
-    0x800007EE8000081E, 0x800007C380000C2D, 0x8000082180000764,
-    0x80000C31800007F1, 0x80000767800007C6, 0x800007C98000076A,
-    0x8000068580000665, 0x8000069980000696, 0x800006A7800006AB,
-    0x80000706800006B7, 0x800007118000070B, 0x8000074C8000072A,
-    0x8000077380000761, 0x8000077980000776, 0x800007A380000781,
-    0x80000809800007B5, 0x8000080C800007FB, 0x8000084880000814,
-    0x8000087980000851, 0x8000067980000871, 0x8000051380000474,
-    0x800005A180000566, 0x800005DE800005C7, 0x80000612800005FB,
-    0x8000036F80000629, 0x800003A180000388, 0x800003CD800003BA,
-    0x800003F3800003E0, 0x8000041980000406, 0x800004928000042C,
-    0x800004B0800004A1, 0x800004CE800004BF, 0x800004E6800004DD,
-    0x800004F8800004EF, 0x8000052D80000501, 0x800006F780000536,
-    0x80003B7280003B6E, 0x80003B7A80003B76, 0x80003B8680003B7E,
-    0x80003B8A0037E4E3, 0x80003B9280003B8E, 0x80003B9A80003B96,
-    0x80003BA280003B9E, 0x80003BAA80003BA6, 0x80003BB280003BAE,
-    0x80003BB6003744EF, 0x003310E380003BBA, 0x00309CE30030B8E3,
-    0x003388E3003074E3, 0x003344E300336CE3, 0x003340EF003338E3,
-    0x00338CEF0033A0EF, 0x0033D4EF0033E8EF, 0x003070EF003020EF,
-    0x80003BBE003054EF, 0x80003BC680003BC2, 0x80003BCE80003BCA,
-    0x80003BD680003BD2, 0x80003BDE80003BDA, 0x80003BE680003BE2,
-    0x80003BEE80003BEA, 0x80003BF680003BF2, 0x003664E380003BFA,
-    0x0036E8E3003620E3, 0x0031C0E30036D4E3, 0x00318CE30031A4E3,
-    0x003138EF003668E3, 0x00317CEF003114EF, 0x003148EF003144EF,
-    0x0036CCEF003684EF, 0x003100E3003634EF, 0x003154E300316CE3,
-    0x0030ECE30031A8E3, 0x0030D0E30030C8E3, 0x003060EF003130E3,
-    0x00301CEF003024EF, 0x0031DCEF0031E0EF, 0x0031A4EF0031C8EF,
-    0x000B5D80003178EF, 0x000B6D82000B6581, 0x000B7D84000B6583,
-    0x000B6D86000B6585, 0x000B5D88000B6587, 0x000BAD8A000BA589,
-    0x000BBD8C000BA58B, 0x000BAD8E000BA58D, 0x000BDD90000BA58F,
-    0x000BED92000BE591, 0x000BFD94000BE593, 0x000BED96000BE595,
-    0x000BDD98000BE597, 0x000BAD9A000BA599, 0x000BBD9C000BA59B,
-    0x000BAD9E000BA59D, 0x000B5DA0000BA59F, 0x000B6DA2000B65A1,
-    0x000B7DA4000B65A3, 0x000B6DA6000B65A5, 0x000B5DA8000B65A7,
-    0x000C8DAA000C85A9, 0x000C9DAC000C85AB, 0x000C8DAE000C85AD,
-    0x000CFDB0000C85AF, 0x000CADB2000CA5B1, 0x000CBDB4000CA5B3,
-    0x000CADB6000CA5B5, 0x000C9DB8000CA5B7, 0x000C4DBA000C45B9,
-    0x000C5DBC000C45BB, 0x000C4DBE000C45BD, 0x000DBD80000C45BF,
-    0x000DCD82000DC581, 0x000DDD84000DC583, 0x000DCD86000DC585,
-    0x000DFD88000DC587, 0x000DCD8A000DC589, 0x000DDD8C000DC58B,
-    0x000DCD8E000DC58D, 0x000DBD90000DC58F, 0x000C4D92000C4591,
-    0x000C5D94000C4593, 0x000C4D96000C4595, 0x000C7D98000C4597,
-    0x000C4D9A000C4599, 0x000C5D9C000C459B, 0x000C4D9E000C459D,
-    0x000CBDA0000C459F, 0x000CCDA2000CC5A1, 0x000CC1A4000CC1A3,
-    0x000CC1A6000CC1A5, 0x000CC1A8000CC1A7, 0x000CCDAA000D19A9,
-    0x000CDDAC000CC5AB, 0x000CCDAE000CC5AD, 0x000CBDB0000CC5AF,
-    0x000C4DB2000C45B1, 0x000C5DB4000C45B3, 0x000C4DB6000C45B5,
-    0x000C7DB8000C45B7, 0x000C49BA000C7DB9, 0x000C59BC000C49BB,
-    0x000C49BE000C59BD, 0x0008C59C000C51BF, 0x0008C59E0008CD9D,
-    0x0008A5980008DD9F, 0x0008A59A0008AD99, 0x000885940008DD9B,
-    0x0008859600088D95, 0x0008A59000089D97, 0x0008A5920008AD91,
-    0x0008C58C00089D93, 0x0008C58E0008CD8D, 0x000825880008DD8F,
-    0x0008258A00082D89, 0x000805840008DD8B, 0x0008058600080D85,
-    0x0008258000081D87, 0x0008258200082D81, 0x0009CDBC00081D83,
-    0x0009DDBE0009D5BD, 0x0009ADB80009D5BF, 0x0009DDBA0009D5B9,
-    0x00098DB40009D5BB, 0x00099DB6000995B5, 0x0009ADB0000995B7,
-    0x00099DB2000995B1, 0x0009C9AC000995B3, 0x0009DDAE0009C9AD,
-    0x000825A80009D5AF, 0x000825AA00082DA9, 0x000805A40009D9AB,
-    0x000805A600080DA5, 0x000825A000081DA7, 0x000825A200082DA1,
-    0x000EB58000081DA3, 0x000EB582000EBD81, 0x000ED584000ECD83,
-    0x000ED586000EDD85, 0x000EF588000ECD87, 0x000EF58A000EFD89,
-    0x000ED58C000ECD8B, 0x000E0D8E000EDD8D, 0x000EB190000ED18F,
-    0x000EB192000EB191, 0x000F8594000EB193, 0x000D5D96000D5595,
-    0x000D6D98000D5597, 0x000D7D9A000D7599, 0x000D6D9C000D759B,
-    0x000D5D9E000D559D, 0x000DADA0000D559F, 0x000DBDA2000DB5A1,
-    0x000DADA4000DB5A3, 0x000DDDA6000DD5A5, 0x000DEDA8000DD5A7,
-    0x000DFDAA000DF5A9, 0x000DEDAC000DF5AB, 0x000DDDAE000DD5AD,
-    0x000DADB0000DD5AF, 0x000DBDB2000DB5B1, 0x000DADB4000DB5B3,
-    0x000D5DB6000D55B5, 0x000D6DB8000D55B7, 0x000D7DBA000D75B9,
-    0x000D6DBC000D75BB, 0x000D5DBE000D55BD, 0x000AAD80000D55BF,
-    0x000AB98280003BFE, 0x000AA5840000E5B2, 0x000AA1860000EDB4,
-    0x000A9D880000E1B6, 0x005DED8A000125B8, 0x000AF18C000129BA,
-    0x000AF18E00013DBC, 0x000A8D90000131BE, 0x000A999280003C02,
-    0x000A8594000009A2, 0x000A81960000EDA4, 0x000ABD9800009DA6,
-    0x000A899A0000ADA8, 0x000A959C0000BDAA, 0x000A919E0000ADAC,
-    0x000A6DA000009DAE, 0x000A79A2000BC980, 0x000A69A4000BDD82,
-    0x000A65A6000BC184, 0x000BE1A8000BC586, 0x000BEDAA000BF988,
-    0x000BF9AC000BCD8A, 0x80003C06000BD18C, 0x80003C0A000BD58E,
-    0x80003C0E000BA990, 0x80003C12000BBD92, 0x80003C16000BA194,
-    0x80003C1A000BA596, 0x80003C1E000B9998, 0x02726D9D0000E99A,
-    0x80003C220272799C, 0x80003C2A80003C26, 0x80003C3280003C2E,
-    0x80003C3A80003C36, 0x80003C4280003C3E, 0x80003C4A80003C46,
-    0x80003C5280003C4E, 0x80003C5A80003C56, 0x80003C6280003C5E,
-    0x80003C6A80003C66, 0x000348990003C09A, 0x001C009F80003C6E,
-    0x0001589D0000109E, 0x80003C720003989C, 0x80003C7A80003C76,
-    0x80003C8280003C7E, 0x80003C8A80003C86, 0x80003C9280003C8E,
-    0x80003C9A80003C96, 0x80001CAA80003C9E, 0x001FE8AD001E94AE,
-    0x80003CA680003CA2, 0x80003CAE80003CAA, 0x80003CB680003CB2,
-    0x80003CC280003CBE, 0x80003CCE80003CCA, 0x80003CD680003CD2,
-    0x80003CDE80003CDA, 0x80003CE680003CE2, 0x80003CEE80003CEA,
-    0x80000B8B80000B10, 0x80003D3280000B55, 0x80003CF680003CF2,
-    0x0003118080003CFA, 0x00031D8280003CFE, 0x0003098480003D02,
-    0x0003058680003D06, 0x0003018880003D0A, 0x00030D8A80003D0E,
-    0x0003198C80003CC6, 0x0003158E80001CBA, 0x0003719080003D12,
-    0x00037D9280003D16, 0x0003699480003D1A, 0x0003659680003D1E,
-    0x0003219880003D22, 0x0003259A80003D26, 0x0003119C80003D2A,
-    0x00031D9E80003D2E, 0x0003E9A080000B67, 0x0003E5A2001F6C82,
-    0x0003E1A480001B0E, 0x0003EDA680001B12, 0x0003D9A880001B16,
-    0x0003D5AA80001B1A, 0x000331AC80001B1E, 0x00033DAE80001B22,
-    0x000345B080001B26, 0x80001B2A00024891, 0x80001B3280001B2E,
-    0x80001B3A80001B36, 0x80001B4280001B3E, 0x80001B4600036DB9,
-    0x80001B4A000361BB, 0x000365BE000365BD, 0x80001B5280001B4E,
-    0x80001B560002B190, 0x80001B5A0002B592, 0x80001B6280001B5E,
-    0x80001B660002A196, 0x80001B6A0002E598, 0x80001B6E0002E99A,
-    0x80001B720002FD9C, 0x80001B760002F19E, 0x80001B7A0002E580,
-    0x80001B7E0002E982, 0x80001B820002FD84, 0x80001B860002F186,
-    0x0002D98B80001B8A, 0x005C258D80001B8E, 0x80001B9680001B92,
-    0x00026DB180001B9A, 0x000279B300026DB0, 0x80001B9E005CE9B2,
-    0x80001BA2000261B4, 0x80001BA6000265B6, 0x80001BAA000259B8,
-    0x80001BAE0001A9BA, 0x80001BB20001B9BC, 0x80001BB60001A9BE,
-    0x80001BBA000215A0, 0x80001BBE000219A2, 0x80001BC200020DA4,
-    0x80001BC6000201A6, 0x005D15AB000205A8, 0x000215AD027F49AA,
-    0x80001BCA005CBDAC, 0x001CC19C005CA1AE, 0x0035E19E001E699D,
-    0x80001BCE001E799F, 0x80001BD680001BD2, 0x80001BDE80001BDA,
-    0x80001BE680001BE2, 0x80001BEE80001BEA, 0x80001BF680001BF2,
-    0x80001BFE80001BFA, 0x80001C0680001C02, 0x80001C0E80001C0A,
-    0x80001C1680001C12, 0x80001C1E80001C1A, 0x80001C2680001C22,
-    0x80001C2E80001C2A, 0x80001C3680001C32, 0x001E5DBC80001C3A,
-    0x001E61BE001E65BD, 0x001E5DB8001E7DBF, 0x001E4DBA001E45B9,
-    0x001E0DB4001E45BB, 0x001E15B6001E11B5, 0x001E3DB0001E65B7,
-    0x001E25B2001E21B1, 0x80001C3E001E19B3, 0x80001C4680001C42,
-    0x800038AE80001C4A, 0x80003225001EA9A9, 0x8000338E800038D6,
-    0x80002DED80002BF2, 0x80002E5180002F20, 0x8000310B80002F1C,
-    0x001FFD8080003506, 0x001E0182001E0581, 0x001E0584001E1D83,
-    0x001E3986001E0D85, 0x001E1188001E3587, 0x001E698A001E1D89,
-    0x001E618C001E658B, 0x001E598E001E6D8D, 0x001E2D90001E598F,
-    0x001E3192001E3191, 0x001EC194001E3193, 0x001EF996001ECD95,
-    0x001ED198001EF597, 0x001EA99A001EDD99, 0x001EA19C001EA59B,
-    0x001E999E001EAD9D, 0x000191A0001E859F, 0x000191A2000191A1,
-    0x0001B1A4000191A3, 0x0001B1A60001B1A5, 0x000191A80001B1A7,
-    0x000191AA000191A9, 0x0001F1AC000191AB, 0x0001F1AE0001F1AD,
-    0x000191B00001F1AF, 0x000191B2000191B1, 0x0001B1B4000191B3,
-    0x0001B1B60001B1B5, 0x000191B80001B1B7, 0x000191BA000191B9,
-    0x000131BC000191BB, 0x000131BE000131BD, 0x000400A4000131BF,
-    0x000708A6000504A5, 0x8000369600090CA7, 0x8000380A800037CE,
-    0x00F120AC8000383E, 0x00F328AE00F224AD, 0x000B30A800F560AF,
-    0x000E38AA000C00A9, 0x00FC00B400F000AB, 0x00FE48B600FD04B5,
-    0x00F658B000E000B7, 0x00F958B200F800B1, 0x00E664BC00FA5CB3,
-    0x00E800BE00E764BD, 0x00E130B800E964BF, 0x00E400BA00E278B9,
-    0x8000395200E57CBB, 0x80003336800032AA, 0x8000347A800033E2,
-    0x8000397E80003796, 0x80002C1680003AAA, 0x800032BE80002DC1,
-    0x800037CA80003406, 0x80002FC880003AFA, 0x800037BA800033AE,
-    0x8000313F8000384E, 0x80003816800036F2, 0x800032398000303B,
-    0x800034528000334E, 0x80002C5D8000395E, 0x80002D6180002CF5,
-    0x800032B6800031A7, 0x00EEB9BA80003402, 0x001D5D9680003526,
-    0x001D5D90800037C2, 0x800038FE800037DA, 0x80003ADA80003A26,
-    0x8000356280003AF6, 0x001D6D98800035A2, 0x8000377A8000365A,
-    0x001D6D84800039B2, 0x001D7D860040D985, 0x001D4D80001D7187,
-    0x001D7D8280003882, 0x8000305780002ED5, 0x800036D280003612,
-    0x8000356A80003436, 0x800038CE001D3189, 0x027445B480003A1E,
-    0x80002ED1001DE1B5, 0x800032AE80002FA4, 0x027791B280003362,
-    0x800033960277CDB3, 0x800036728000364A, 0x027E79B8800039DE,
-    0x80002D5D027DA1B9, 0x80002D01800036DE, 0x800035CE80002CFD,
-    0x800037928000365E, 0x8000389E800039E6, 0x8000388E80003143,
-    0x80002F7880002BFE, 0x80003466800030AF, 0x80002D81800034E6,
-    0x0004D18080003572, 0x0038C5820004DD81, 0x0004F1840004D183,
-    0x0004F1860004F185, 0x0024218800241D87, 0x0038C58A0004C989,
-    0x0004C58C0004DD8B, 0x0004C58E0004CD8D, 0x0006559000062D8F,
-    0x0006559200065D91, 0x0006D5940006CD93, 0x0006D5960006DD95,
-    0x0006F5980006CD97, 0x0006F59A0006FD99, 0x0006D59C0006CD9B,
-    0x0006D59E0006DD9D, 0x000635A00006CD9F, 0x000635A200063DA1,
-    0x000655A400064DA3, 0x000655A600065DA5, 0x000675A800064DA7,
-    0x000675AA00067DA9, 0x000655AC00064DAB, 0x000655AE00065DAD,
-    0x000635B000064DAF, 0x000635B200063DB1, 0x003895B40007C9B3,
-    0x0007CDB60007D9B5, 0x0007FDB80007C5B7, 0x0007CDBA0007C5B9,
-    0x0007DDBC0007C5BB, 0x0007C1BE0007C1BD, 0x003A71800007C1BF,
-    0x0006558200065181, 0x0006758400066D83, 0x0006758600067D85,
-    0x0006598800066D87, 0x0006A18A00065989, 0x0006A18C0006A18B,
-    0x0006998E0006A18D, 0x0006E9900006998F, 0x0006F1920006E991,
-    0x000691940006F193, 0x0006919600069195, 0x0006B19800069197,
-    0x0006B19A0006B199, 0x0007F59C0006B19B, 0x00069D9E0006959D,
-    0x00066DA00006959F, 0x00067DA2000675A1, 0x00066DA4000675A3,
-    0x000391A6000391A5, 0x0003B1A8000391A7, 0x0003B1AA0003B1A9,
-    0x000391AC0003B1AB, 0x000391AE000391AD, 0x0003F1B0000391AF,
-    0x0003FDB2003911B1, 0x0003EDB40003F5B3, 0x00039DB6000395B5,
-    0x0003ADB8000395B7, 0x0003BDBA0003B5B9, 0x0003ADBC0003B5BB,
-    0x00039DBE000395BD, 0x001A80F0000395BF, 0x8000309380002C6D,
-    0x8000332280002BDE, 0x80003646800031BF, 0x80002EC180002DD5,
-    0x8000379A80003532, 0x800032EE80003872, 0x800031538000331A,
-    0x8000317380003752, 0x80002C32800034E2, 0x80002CF980002CB5,
-    0x800036328000328A, 0x8000387E8000372A, 0x80002D6980003992,
-    0x8000304B80002E1D, 0x800033B2800031DB, 0x800039CE80003576,
-    0x80003B0680003AB2, 0x8000321D80003B1E, 0x80003922800032DA,
-    0x800030F380003017, 0x8000319F8000310F, 0x800033EE800033A2,
-    0x800035C28000349E, 0x800036CA80003666, 0x800037A280003912,
-    0x800039B68000393A, 0x80002D4980002D1D, 0x800033DE80002E31,
-    0x800030378000382A, 0x8000316B800030A7, 0x8000360E800032E6,
-    0x00016D808000345A, 0x00017982003DC981, 0x0001698400017983,
-    0x0001998600016985, 0x0001A98800019987, 0x003D058A0001A989,
-    0x0001A58C0001BD8B, 0x0001A58E0001AD8D, 0x0001E59000019D8F,
-    0x0001E5920001ED91, 0x0001E5940001FD93, 0x0001E5960001ED95,
-    0x0001A59800019D97, 0x0001A59A0001AD99, 0x0001A59C0001BD9B,
-    0x0001A59E0001AD9D, 0x000165A000019D9F, 0x000161A20001A9A1,
-    0x000161A4000161A3, 0x000161A6000161A5, 0x000C01A8000161A7,
-    0x000441AA000FA5A9, 0x000FB9AC000FA9AB, 0x000FA9AE000FB9AD,
-    0x000FD9B0000FA9AF, 0x000FE9B2000FD9B1, 0x000FF9B4000FE9B3,
-    0x000FE9B6000FF9B5, 0x000FD9B8000FE9B7, 0x000FA9BA000FD9B9,
-    0x000FB9BC000FA9BB, 0x000FA9BE000FB9BD, 0x80002E8D000579BF,
-    0x800030B380002FDC, 0x8000349280003476, 0x800036C2800036A2,
-    0x80003A1A8000399A, 0x80003A6680003A2A, 0x800035B680002C61,
-    0x80003A0280003986, 0x80002C1E800030CB, 0x80002F7C80002C85,
-    0x800031CB80002F9C, 0x80003502800033FA, 0x8000394E800037AE,
-    0x800039DA8000320D, 0x8000325980002D41, 0x8000334680003265,
-    0x800034828000337E, 0x8000355E800034DE, 0x80003A7280003642,
-    0x800039EA8000311F, 0x80002FC080002C75, 0x8000391680003366,
-    0x800030D380003083, 0x800039F28000326D, 0x80002E0D80002D25,
-    0x800031F680002FAC, 0x8000329680003249, 0x8000347E8000332A,
-    0x8000368E800034F2, 0x8000383A8000382E, 0x80002D9180003A0E,
-    0x80002E1180003382, 0x00015580800033FE, 0x0001558200015D81,
-    0x002ADD8400016D83, 0x0001718600017185, 0x0001518800017187,
-    0x0001518A00015189, 0x0001B18C0001518B, 0x0001B18E0001B18D,
-    0x0001D1900001B18F, 0x0001DD9200194591, 0x0001CD940001D593,
-    0x0001FD960001F595, 0x0001CD980001F597, 0x0001DD9A0001D599,
-    0x0001C99C000A9D9B, 0x0001B99E0001C99D, 0x000149A00001B99F,
-    0x000159A2000149A1, 0x000149A4000159A3, 0x000179A6000149A5,
-    0x000149A8000179A7, 0x000145AA00354DA9, 0x000145AC00015DAB,
-    0x000145AE00014DAD, 0x0003C5B00003BDAF, 0x0003C5B20003CDB1,
-    0x0003C5B40003DDB3, 0x0003C5B60003CDB5, 0x0003C1B80035F9B7,
-    0x0003C1BA0003C1B9, 0x0003C1BC0003C1BB, 0x0003C1BE0003C1BD,
-    0x800034A20003C1BF, 0x800039F6800037BE, 0x80003B0A80003AE2,
-    0x8000335E80003261, 0x800035FA800036FA, 0x8000344E8000361E,
-    0x8000389A800033D6, 0x8000376280002C3A, 0x80002D1980002D29,
-    0x8000314780003027, 0x80002F688000362A, 0x800032198000333A,
-    0x800039E28000391E, 0x80002C9580003043, 0x80002EBD80002E69,
-    0x80002D0580003209, 0x8000351A80003456, 0x8000359A8000357E,
-    0x800035B28000359E, 0x8000362280003A32, 0x8000388A800037C6,
-    0x8000396A8000393E, 0x80003A8A80003A86, 0x80003AF280003A8E,
-    0x800039FE80003962, 0x80002C8980002C65, 0x80002D5980002C9D,
-    0x80002DA580002D65, 0x80002E7580002E59, 0x80002EB980002E7D,
-    0x80002FA880002EC5, 0x800030E7800030BF, 0x00014180800030EF,
-    0x0001418200014181, 0x0001418400014183, 0x0001418600014185,
-    0x0037B18800014187, 0x0037B58A00014589, 0x0001598C0001498B,
-    0x800031070001598D, 0x00013190800031B3, 0x00013D92800031E3,
-    0x8000328E80003211, 0x0001259600012D95, 0x0001E5980001DD97,
-    0x0001E59A0001ED99, 0x0001E59C0001FD9B, 0x0037399E0001ED9D,
-    0x00011DA080003352, 0x000129A28000336E, 0x800033F28000339E,
-    0x000131A6000131A5, 0x80002C4280003412, 0x000105AA80003486,
-    0x000105AC00011DAB, 0x000105AE00010DAD, 0x000105B000017DAF,
-    0x000105B200010DB1, 0x000105B400011DB3, 0x000105B600010DB5,
-    0x000105B800013DB7, 0x000105BA00010DB9, 0x0030F5BC00011DBB,
-    0x000101BE000101BD, 0x00008D8D000101BF, 0x00010D8F001C118C,
-    0x00078D8900010D8E, 0x0000918B00078588, 0x00075D850007B98A,
-    0x0007A58700077584, 0x00070D810007A986, 0x0007758300070D80,
-    0x0001759D00077982, 0x0001719F0001759C, 0x000175990001799E,
-    0x0001759B00017598, 0x000179950001759A, 0x000AD59700017194,
-    0x001C559100017196, 0x0001699300017D90, 0x0001EDAD00017192,
-    0x800035660001EDAC, 0x0001CDA98000358A, 0x0001FDAB0001CDA8,
-    0x0001F1A50001FDAA, 0x00060DA70001E9A4, 0x0001F1A10001F9A6,
-    0x0001F1A3000189A0, 0x00015DBD0001F9A2, 0x00016DBF0006B5BC,
-    0x000141B900016DBE, 0x000141BB000609B8, 0x000165B5000149BA,
-    0x000165B7000165B4, 0x000185B1000165B6, 0x000185B3000185B0,
-    0x80003586000185B2, 0x800035928000358E, 0x800035AE800035AA,
-    0x800035E6800035D2, 0x8000366E80003602, 0x8000368A80003676,
-    0x80003702800036B2, 0x8000379E80003736, 0x8000385E80003846,
-    0x8000389680003892, 0x800038DE800038DA, 0x80003A1280003936,
-    0x80003A6E80003A56, 0x8000F2BF800030B7, 0x80002BE68000371A,
-    0x80002CB180002CF1, 0x80002C9980002C59, 0x80002D5580002CC5,
-    0x80002E4580002D71, 0x80002E7180002E55, 0x80002F1880002ECD,
-    0x80002F4080002F24, 0x8000303F80002F50, 0x8000307780003047,
-    0x800030C780003097, 0x800030CF800030DF, 0x80003123800030E3,
-    0x800031878000317B, 0x800031B78000318F, 0x800032518000323D,
-    0x8000338A80003392, 0x8000354A800033BA, 0x0003A98080003416,
-    0x0003B9820003A981, 0x0003A9840003B983, 0x0003D9860003A985,
-    0x0003E9880003D987, 0x0003E58A0004D989, 0x0003E18C0005158B,
-    0x0003E18E0003E18D, 0x0003A1900003E18F, 0x000C919200043191,
-    0x0003B9940003A993, 0x000869960035DD95, 0x00038D9800062597,
-    0x0005E59A00039599, 0x0005E19C0003999B, 0x0003859E00038D9D,
-    0x0004FDA000017D9F, 0x000181A2000181A1, 0x000181A4000181A3,
-    0x000419A6000181A5, 0x0001BDA8000185A7, 0x00018DAA000185A9,
-    0x00019DAC000185AB, 0x000189AE0003A9AD, 0x00087DB0000189AF,
-    0x0003E5B20001FDB1, 0x000191B40001F1B3, 0x000191B6000191B5,
-    0x0003A1B8000191B7, 0x000449BA0001B5B9, 0x0003B5BC0001B9BB,
-    0x0001A1BE000D15BD, 0x8000344A0003B5BF, 0x800034B280003496,
-    0x800034FA800034D6, 0x8000351E800034FE, 0x800035428000352E,
-    0x8000356E8000353E, 0x8000361A800035EA, 0x8000368280003652,
-    0x8000378E80003766, 0x8000384A80003802, 0x8000387680003856,
-    0x800038868000387A, 0x8000391A800038A2, 0x800039828000394A,
-    0x800039EE800039A2, 0x80003A6280003A46, 0x8000F25680003AC2,
-    0x8000F2838000F251, 0x80002B3680002B0E, 0x8000F30580002B3A,
-    0x8000F3AA8000F332, 0x80003B5E80003B4E, 0x800006EC800006E1,
-    0x800006E4800006EF, 0x8000084E800006E8, 0x80000E9580000EA4,
-    0x80000EC780000E9A, 0x80000F3480000E9F, 0x80000F6A80000FB3,
-    0x80000F4380000F18, 0x80000F5880000F50, 0x80000FAB80000F86,
-    0x00FF818F80000FA1, 0x00FF758D00D0058E, 0x00FF698B00FD6D8C,
-    0x00FF6D8900FD498A, 0x00FD798700FA8988, 0x00FF458500FF5586,
-    0x00FF558300FF4584, 0x00FD698100FF5582, 0x80000FA600FF6180,
-    0x80000F9A80000F93, 0x80000EEC80000EE7, 0x00FFF19980000EF1,
-    0x00FFC19700FFF998, 0x00FFF19500FFF996, 0x00FFE19300FFF994,
-    0x00FFF19100FFF992, 0x80000EFE00FFF990, 0x80000F1380000F0B,
-    0x80000F2580000F1B, 0x80000F2F80000F2A, 0x80000F3E80000F39,
-    0x80000F5380000F46, 0x80000F6080000F5B, 0x80000F6D80000F65,
-    0x80000F7C80000F72, 0x80000F8980000F81, 0x80000FAE80000F8E,
-    0x80000F0380000F20, 0x80000F7780000F4B, 0x800016C680000EF6,
-    0x800016D2800016CF, 0x800016CC800016D8, 0x00FEAD80800016D5,
-    0x00FEBD8200FEB581, 0x00FEAD8400FEB583, 0x00FED98600FED585,
-    0x800016FC800016C9, 0x800016DE800016FF, 0x800016E1800016DB,
-    0x800016ED800016E4, 0x800016F0800016EA, 0x800016F6800016E7,
-    0x00FE999400FE8993, 0x00FE899600FE9995, 0x800016F300FE8997,
-    0x8000170880001702, 0x8000170B8000170E, 0x8000171100FEAD9D,
-    0x00FE61A000FEA19F, 0x00BB19A200BBFDA1, 0x00FE79A400FE69A3,
-    0x00FE69A600FE79A5, 0x00FE59A800FE69A7, 0x00FE55AA00B649A9,
-    0x0001F5AC0001EDAB, 0x0001F5AE0001FDAD, 0x000195B00001EDAF,
-    0x00014DB2000145B1, 0x00015DB4000145B3, 0x00014DB6000145B5,
-    0x000179B880001714, 0x000149BA000179B9, 0x000159BC000149BB,
-    0x000155BE80001717, 0x0002B18080001705, 0x800017290002B181,
-    0x0002AD840002B583, 0x0002D98680001726, 0x0002E9880002D987,
-    0x0002F98A0002E989, 0x0002E98C0002F98B, 0x0002D98E0002E98D,
-    0x0002A9900002D98F, 0x0002A5920002AD91, 0x0002BD940002A193,
-    0x0002A9960002B995, 0x000291980002AD97, 0x00029D9A00029599,
-    0x0002859C0002999B, 0x0002819E0002819D, 0x000279A00002859F,
-    0x000275A200027DA1, 0x00026DA4000271A3, 0x000385A6000269A5,
-    0x0003BDA8000381A7, 0x000389AA0003B9A9, 0x000391AC00038DAB,
-    0x00039DAE000395AD, 0x0003E5B0000399AF, 0x0003E1B20003E1B1,
-    0x0003F9B40003E5B3, 0x0003F5B60003FDB5, 0x0003CDB80003F1B7,
-    0x0003D9BA0003C9B9, 0x0003C1BC0003DDBB, 0x0003CDBE0003C5BD,
-    0x000235800003C9BF, 0x0002318200023181, 0x00022D8400023583,
-    0x0002598600022985, 0x0002658800025D87, 0x0002618A00026189,
-    0x0002518C0002658B, 0x00025D8E0002558D, 0x000225900002598F,
-    0x0002219200022191, 0x0002399400022593, 0x0002359600023D95,
-    0x00020D9800023197, 0x0002199A00020999, 0x0002019C00021D9B,
-    0x0002019E0002059D, 0x00025DA00002059F, 0x000255A2000259A1,
-    0x02A7A1A4000251A3, 0x000401A602A7A5A5, 0x000439A8000405A7,
-    0x00025DAA00043DA9, 0x000245AC000259AB, 0x0007B9AE000241AD,
-    0x02A7E5B00007BDAF, 0x8000173102A7E1B1, 0x8000172380001737,
-    0x8000173D80001734, 0x80000FCA80001662, 0x80000FF78000101A,
-    0x800010068000100B, 0x8000101580001010, 0x8000174080000FFC,
-    0x000840918000173A, 0x80000FD480000FCF, 0x8000100180000FE8,
-    0x8000104680001041, 0x8000106880001052, 0x8000107C80001077,
-    0x800010A180001087, 0x800010D8800010B4, 0x8000110F8000110A,
-    0x8000112680001117, 0x8000113A80001135, 0x8000116580001142,
-    0x800011988000118C, 0x800011BD800011B8, 0x800011FC800011C2,
-    0x8000121B8000120F, 0x800012AD80001233, 0x8000131C80001303,
-    0x8000133480001321, 0x8000135780001345, 0x800013838000135C,
-    0x800013A08000138B, 0x800013CC800013C7, 0x8000142B80001426,
-    0x8000143C80001430, 0x8000144D80001448, 0x8000146180001455,
-    0x800014808000147B, 0x8000148D80001488, 0x8000149780001492,
-    0x800014A18000149C, 0x0004FD9302A2B8D9, 0x800014B9800014B4,
-    0x00071997800014D5, 0x0004E5950004E996, 0x0006E99B0004E194,
-    0x00073D990007319A, 0x0006FD9F00072598, 0x02A13D9D0006F99E,
-    0x800014EF0006F59C, 0x8000151580001509, 0x800015328000152D,
-    0x800015608000153F, 0x8000159D80001583, 0x800015AE800015A9,
-    0x800015D7800015B6, 0x80001600800015F1, 0x000671B38000161D,
-    0x000665B1000675B2, 0x000669B7000661B0, 0x00066DB500066DB6,
-    0x00065DBB000669B4, 0x000655B9000659BA, 0x0001B5BF000655B8,
-    0x0001BDBD0001B1BE, 0x000615A30001B9BC, 0x000619A1000611A2,
-    0x000609A700061DA0, 0x000601A500060DA6, 0x000601AB000605A4,
-    0x000635A9000605AA, 0x000619AF000631A8, 0x00061DAD00061DAE,
-    0x00021582000619AC, 0x000205800005F983, 0x0002318600021D81,
-    0x0002018400023187, 0x0002118A00020185, 0x000201880002118B,
-    0x0002718E00020189, 0x0002018C0002718F, 0x000211920002018D,
-    0x0002019000021193, 0x0002319600020191, 0x0002019400023197,
-    0x0002119A00020195, 0x000201980002119B, 0x0002F19E00020199,
-    0x0002019C0002F19F, 0x000211A20002019D, 0x000201A0000211A3,
-    0x000231A6000201A1, 0x000201A4000231A7, 0x000211AA000201A5,
-    0x000201A8000211AB, 0x000271AE000201A9, 0x000201AC000271AF,
-    0x000211B2000201AD, 0x000201B0000211B3, 0x000231B6000201B1,
-    0x000201B4000231B7, 0x000211BA000201B5, 0x000201B8000211BB,
-    0x0003F9BE000201B9, 0x000201BC0003F9BF, 0x800016220003F9BD,
-    0x8000162F8000162A, 0x8000164780001642, 0x800016798000166D,
-    0x8000169480001685, 0x800016B6800016B1, 0x800011E8800011D7,
-    0x8000008180001665, 0x800000958000008B, 0x800000A98000009F,
-    0x800000B3001AF1BA, 0x80000FE380000FDE, 0x8000105E80000FED,
-    0x8000106302A114D9, 0x800010CE8000106D, 0x80001100800010D3,
-    0x800011218000111C, 0x8000153A8000112B, 0x800015FB800015F6,
-    0x8000171D80001613, 0x8000168F8000168A, 0x80000FD9800016A7,
-    0x8000107280000FF2, 0x800012C080001105, 0x8000161880001528,
-    0x800016AC8000164C, 0x8000124D80001130, 0x8000129B80001288,
-    0x80001406800013FA, 0x8000137680001412, 0x800013BA8000137B,
-    0x800013E6800013BF, 0x000F1595800013EB, 0x000F1597000F1594,
-    0x000F1591000F1596, 0x000F1593000F1590, 0x000F159D000F1592,
-    0x000F159F000F159C, 0x000F1599000F159E, 0x000F159B000F1598,
-    0x000D1585000F159A, 0x000D1587000D1584, 0x000D1581000D1586,
-    0x000D1583000D1580, 0x000D158D000D1582, 0x000D158F000D158C,
-    0x000D1589000D158E, 0x000D158B000D1588, 0x000CC9B5000D158A,
-    0x000F09B7000CC1B4, 0x000FE5B1000FF1B6, 0x000FE1B3000FE5B0,
-    0x000C71BD000CDDB2, 0x000C49BF000C69BC, 0x000F35B9000C71BE,
-    0x000CE1BB000CE5B8, 0x000F99A5000CE9BA, 0x000F95A7000F91A4,
-    0x000F95A1000CD9A6, 0x000F89A3000F95A0, 0x000F8DAD000F95A2,
-    0x000CD9AF000CCDAC, 0x000CE9A9000CC1AE, 0x000F91AB000B09A8,
-    0x000DE580000F9DAA, 0x000DF182000DF181, 0x000C1984000C0983,
-    0x000C5586000C5D85, 0x000EA988000C4D87, 0x000EA58A000C4589,
-    0x000C498C000EBD8B, 0x000C458E000EA18D, 0x000E7590000E3D8F,
-    0x000EE992000ED991, 0x000EF594000E6D93, 0x000E7196000E7195,
-    0x000D7198000D5197, 0x000D7D9A000ECD99, 0x000D6D9C000EC19B,
-    0x000D5D9E000D559D, 0x000E31A0000D559F, 0x000DB1A2000DB1A1,
-    0x000DD1A4000DB1A3, 0x000DC9A6000E21A5, 0x000DC1A8000DC1A7,
-    0x000DC1AA000DC1A9, 0x000DC5AC000DC1AB, 0x000DC5AE000DCDAD,
-    0x000D45B0000D3DAF, 0x000E71B2000D4DB1, 0x000D41B4000D41B3,
-    0x000D41B6000D41B5, 0x000D41B8000D41B7, 0x000D45BA000D4DB9,
-    0x000D45BC000D5DBB, 0x000D45BE000D4DBD, 0x000EC580000D3DBF,
-    0x000EDD82000ECD81, 0x000EF984000ED583, 0x000EE986000EF985,
-    0x000ED988000EE987, 0x000F698A000F2989, 0x000F798C000F698B,
-    0x000F6D8E000D898D, 0x000F1D90000F658F, 0x000F2592000F2591,
-    0x000F2594000F3D93, 0x000D9996000F2D95, 0x000D6998000D5997,
-    0x000D7D9A000DA199, 0x000D6D9C000D759B, 0x000DB19E000D559D,
-    0x000D2DA0000E699F, 0x000D39A2000EADA1, 0x000D25A4000ECDA3,
-    0x000D55A6000EC5A5, 0x000D71A8000E19A7, 0x000D71AA000D71A9,
-    0x000E49AC000E59AB, 0x000FC1AE000FFDAD, 0x000DE1B0000D7DAF,
-    0x000D15B2000D0DB1, 0x000D35B4000D2DB3, 0x000D35B6000D3DB5,
-    0x000D15B8000D2DB7, 0x000D15BA000D1DB9, 0x0001F9BC0001E9BB,
-    0x0001E9BE0001F9BD, 0x000D4CF00001E9BF, 0x8000125780001252,
-    0x800012A5800012A0, 0x800011B0800011AB, 0x800011848000117F,
-    0x800011CC800011C7, 0x800013148000130F, 0x8000134F8000134A,
-    0x8000126B8000125F, 0x800012838000127E, 0x800012C58000122E,
-    0x8000103480001340, 0x800010A68000108C, 0x800010B9800010AD,
-    0x800010E4800010DD, 0x8000116A800010EB, 0x8000119D800011A4,
-    0x8000120180001214, 0x8000123F80001208, 0x8000124680001238,
-    0x80001308800012B2, 0x8000126480001270, 0x800012948000128D,
-    0x8000133980001326, 0x8000136880001361, 0x800013908000136F,
-    0x800013AC800013A5, 0x800013DF800013D1, 0x80001435800013D8,
-    0x8000146D80001466, 0x80001502800014F4, 0x800014DA800014FB,
-    0x000755908000150E, 0x000BE592000BED91, 0x000BE594000BFD93,
-    0x000BE596000BED95, 0x0007299800071997, 0x0007399A00072999,
-    0x0007299C0007399B, 0x0007199E0007299D, 0x000759800007199F,
-    0x0007698200075981, 0x0007798400076983, 0x0007698600077985,
-    0x0007598800076987, 0x0007298A00075989, 0x000BA58C0007298B,
-    0x0007358E00073D8D, 0x000B55B000072D8F, 0x000B51B2000B55B1,
-    0x0008E9B4000829B3, 0x000819B60008E9B5, 0x0007E1B80007A1B7,
-    0x00080DBA0007E1B9, 0x0007E9BC000805BB, 0x8000151A0007EDBD,
-    0x0007E9A080001565, 0x0007F9A20007E9A1, 0x0007E9A40007F9A3,
-    0x000799A60007E9A5, 0x000B35A8000799A7, 0x0007A5AA0007ADA9,
-    0x0007A5AC0007BDAB, 0x0007B9AE0007C9AD, 0x00065D900007B9AF,
-    0x0006699200066591, 0x0006799400066993, 0x0006699600067995,
-    0x0006599800066997, 0x0006559A00065D99, 0x0006B59C0006AD9B,
-    0x0006B59E0006BD9D, 0x8000156C0006AD9F, 0x800015448000157C,
-    0x8000158880001552, 0x8000154B8000158F, 0x8000163B80001634,
-    0x800015E3800015DC, 0x800015C9800015C2, 0x800016058000160C,
-    0x000639B080001699, 0x000635B2000639B1, 0x000655B400064DB3,
-    0x000655B600065DB5, 0x000671B8000651B7, 0x000671BA000671B9,
-    0x000651BC000671BB, 0x00065DBE000655BD, 0x000651A0000655BF,
-    0x000651A2000651A1, 0x00064DA4000655A3, 0x000679A6000649A5,
-    0x000645A8000679A7, 0x000645AA00064DA9, 0x000641AC000641AB,
-    0x00064DAE000645AD, 0x0005ED80000645AF, 0x0005FD820005F581,
-    0x0005E9840005F583, 0x0005E5860005E985, 0x000299880002A587,
-    0x0003698A00029989, 0x0003798C0003698B, 0x0003698E0003798D,
-    0x000BE4820003698F, 0x0003119280001057, 0x0003319400031193,
-    0x0003319600033195, 0x00030D9800033197, 0x00031D9A00031599,
-    0x00030D9C0003159B, 0x00003D9E0003099D, 0x000185A000017D9F,
-    0x000185A200018DA1, 0x000185A400019DA3, 0x000185A600018DA5,
-    0x000185A80001BDA7, 0x000185AA00018DA9, 0x000185AC00019DAB,
-    0x000185AE00018DAD, 0x000189B00001F9AF, 0x000199B2000189B1,
-    0x000535B4000199B3, 0x000181B6000539B5, 0x000181B8000181B7,
-    0x000181BA000181B9, 0x000199BC000181BB, 0x000189BE000195BD,
-    0x8000109A000189BF, 0x800010C780001093, 0x800010F9800010C0,
-    0x80001178800010F2, 0x8000117180001147, 0x800012B980001220,
-    0x8000132D80001277, 0x80001521800014E8, 0x001120F08000167E,
-    0x800016A080001672, 0x80001474800015A2, 0x800013B3800015EA,
-    0x800015BB800014AD, 0x800014E180001596, 0x8000114E800014A6,
-    0x8000155980001191, 0x8000104B80001441, 0x800015D080001227,
-    0x8000145A800012FC, 0x800010228000102B, 0x800012CA80001573,
-    0x80001397800011DF, 0x800012D380001654, 0x80001155800012DA,
-    0x8000033E800011ED, 0x8000243E8000243A, 0x800000C780000638,
-    0x800024D88000064D, 0x80002086800024DC, 0x8000065F80002082,
-    0x800008938000088F, 0x800024D48000246E, 0x8000246680002462,
-    0x8000244E8000244A, 0x8000245200A0E0CC, 0x8000245A80002456,
-    0x800006578000245E, 0x800000D18000065B, 0x8000033A800000D7,
-    0x8000064080000340, 0x001F2D838000064B, 0x001F3581001FC582,
-    0x001FD987001F3580, 0x001BF985001FD186, 0x80000659001B9584,
-    0x800000D5800000D3, 0x8000007980000655, 0x8000007D800013F0,
-    0x001F19B380000087, 0x001F19B1001F11B2, 0x001F39B7001F01B0,
-    0x001F39B5001F31B6, 0x001F19BB001F01B4, 0x001F19B9001F11BA,
-    0x80000091001F01B8, 0x8000009B800013F5, 0x80001401001F01BC,
-    0x8000140D800000A5, 0x80001419800000AF, 0x8000141E800000B9,
-    0x8000101F80000FB8, 0x800010818000103E, 0x8000111480001084,
-    0x800011898000113F, 0x001EC190800011B5, 0x001EC192001EC191,
-    0x02807594001EC193, 0x001ECD96001EC595, 0x001EFD98001EC597,
-    0x800011D100A67199, 0x800011DC800011D4, 0x800011F9800011F6,
-    0x800012AA8000125C, 0x8000135480001319, 0x8000138880001380,
-    0x80001423800013C4, 0x8000148580001452, 0x80001537800014BE,
-    0x80001627800015B3, 0x800014C6800014C1, 0x00A6CDB080001651,
-    0x001E6DB2001E65B1, 0x001E79B4001E65B3, 0x00AB91B600A669B5,
-    0x001E51B8001E71B7, 0x001E51BA001E51B9, 0x0001B1BC001E51BB,
-    0x0001B1BE0001B1BD, 0x800014CB00AF11BF, 0x800000CF800014D0,
-    0x8000166A800000D9, 0x8000065D8000034B, 0x800021D080000891,
-    0x80002AAE800021D4, 0x8000259B80002562, 0x800025CC800025BA,
-    0x00021983800025F1, 0x0002198100021182, 0x0002318700ACB580,
-    0x02B3A8D9800029F0, 0x00A5AD8B00020184, 0x00A5A58900A5A98A,
-    0x001DD98F00020988, 0x001DD18D001DDD8E, 0x80002A0C00A5B18C,
-    0x001D8D9100A50192, 0x001D8197001D8D90, 0x001D9595001D8196,
-    0x001D8D9B02833994, 0x00A5959900A86D9A, 0x0002599F001D8D98,
-    0x001D8D9D001D8D9E, 0x0002B5A3001D8D9C, 0x0002B9A100A569A2,
-    0x80002A1E0002A1A0, 0x0002A5A500AEB9A6, 0x0002CDAB0002A5A4,
-    0x0002CDA90002CDAA, 0x800027B50002CDA8, 0x80002AB202B314D9,
-    0x8000255680002A96, 0x0002BDB10002BDB2, 0x000265B70002BDB0,
-    0x8000255A000265B6, 0x000229BB0002A1B4, 0x000229B9000221BA,
-    0x000229BF000255B8, 0x000229BD000221BE, 0x80000895000231BC,
-    0x0291F8CC80000898, 0x800008A10291C4CC, 0x8000089B8000089E,
-    0x800021C4800020B9, 0x800020D1800020C9, 0x800021CC800021C8,
-    0x8000EBE18000EBDC, 0x0002C090800020C5, 0x000E48B200300091,
-    0x8000EBEB8000EBE6, 0x0030BC968000EBF0, 0x8000EBFA8000EBF5,
-    0x8000EC048000EBFF, 0x8000EC0E8000EC09, 0x0021609E0032689D,
-    0x8000EC1300212C9F, 0x8000EC1802B204D9, 0x8000EC228000EC1D,
-    0x8000EC2C8000EC27, 0x8000EC368000EC31, 0x8000EC408000EC3B,
-    0x8000EC4A8000EC45, 0x8000EC548000EC4F, 0x000BC090002A7CAF,
-    0x000BCC928000EC59, 0x8000EC5E000ACC93, 0x000C0096000AD095,
-    0x8000EC688000EC63, 0x8000EC728000EC6D, 0x8000EC7C8000EC77,
-    0x000FDC9E8000EC81, 0x0004F59E8000EC86, 0x0004F99C0004F19F,
-    0x0004E19A0004F99D, 0x0004F5980004E59B, 0x0004CD960004ED99,
-    0x0004C1940004C997, 0x0004D5920004C195, 0x0004DD900004DD93,
-    0x0004A18E0004D991, 0x02B7B18C0004A18F, 0x02B7A98A0004AD8D,
-    0x02B7858802B7AD8B, 0x02B7898602B7A589, 0x02B7B18402B7B987,
-    0x02B7B98202B78585, 0x0004A58002B7AD83, 0x0005C5BE02B7B581,
-    0x0005C9BC0005C1BF, 0x0005D1BA0005C9BD, 0x0005A5B80005D5BB,
-    0x00059DB60005DDB9, 0x000591B4000599B7, 0x000589B2000591B5,
-    0x00058DB000058DB3, 0x0005F1AE000585B1, 0x0005E5AC0005F1AF,
-    0x0005F9AA0005FDAD, 0x000425A80005F9AB, 0x00041DA60005F5A9,
-    0x000411A4000419A7, 0x000409A2000411A5, 0x00040DA000040DA3,
-    0x00044580000405A1, 0x0004418200044D81, 0x0004598400044583,
-    0x0004558600045985, 0x00046D8800045187, 0x0004798A00047589,
-    0x0004618C00047D8B, 0x00046D8E0004618D, 0x000415900004698F,
-    0x0004119200041D91, 0x0004099400041593, 0x0004059600040995,
-    0x00043D9800040197, 0x0004099A00040599, 0x0004119C00040D9B,
-    0x00041D9E0004119D, 0x0004E5A00004199F, 0x0004E1A20004EDA1,
-    0x0004F9A40004E5A3, 0x0004F5A60004F9A5, 0x0004CDA80004F1A7,
-    0x0004D9AA0004D5A9, 0x0004C1AC0004DDAB, 0x0004C5AE0004C9AD,
-    0x00161DB0001661AF, 0x000419B2000415B1, 0x000401B400041DB3,
-    0x0004ADB60004A1B5, 0x000499B80004A5B7, 0x000429BA000425B9,
-    0x000435BC000429BB, 0x8000EC908000EC8B, 0x0019D1810295F5BF,
-    0x000621838000EC95, 0x000679850007C582, 0x0007D58700067184,
-    0x00AC298900063186, 0x00A1C58B00A1D988, 0x0006058D00061D8A,
-    0x0007ED8F0019F58C, 0x029A099100A1798E, 0x029D819300A1F190,
-    0x00A19595029DF192, 0x00A18D9700A1E994, 0x00A1859900A19196,
-    0x02870D9B00A1B998, 0x00AA5D9D0019A19A, 0x0019AD9F00064D9C,
-    0x0296C1A10006419E, 0x029281A30006E9A0, 0x0292B5A5029289A2,
-    0x029219A7029289A4, 0x02921DA90292CDA6, 0x0293F5AB029231A8,
-    0x029315AD0293E9AA, 0x029355AF029335AC, 0x029105B102937DAE,
-    0x029E95B3029EFDB0, 0x029E39B5029E8DB2, 0x029B01B7029E75B4,
-    0x029AE9B9029F25B6, 0x0006C9BB029AE5B8, 0x0006D5BD029AF5BA,
-    0x001911BF000695BC, 0x00A01980000725BE, 0x02930D8202974581,
-    0x02930D8402930583, 0x0293498602933185, 0x0287BD8802939D87,
-    0x0287858A00A33589, 0x0292B18C0292718B, 0x0292F98E0292918D,
-    0x029F79900292D18F, 0x029F099202908191, 0x00A35D94029F1193,
-    0x029EA196029FBD95, 0x029B61980286C597, 0x029B719A0286AD99,
-    0x0006AD9C0018819B, 0x8000EC9A0018859D, 0x000665A000069D9F,
-    0x000799A200184DA1, 0x001855A4000799A3, 0x00F305A600067DA5,
-    0x000651A8000671A7, 0x000651AA000651A9, 0x0005A5AC000651AB,
-    0x000529AE0005F5AD, 0x000579B000050DAF, 0x00F735B200F73DB1,
-    0x00F5D1B400F285B3, 0x00F3D9B600F5D1B5, 0x00F3E9B800F3D9B7,
-    0x00F3F9BA00F3E9B9, 0x00F3E9BC00F3F9BB, 0x00F3B9BE00F3E9BD,
-    0x00FA499A00F3B9BF, 0x00FA499800FA499B, 0x000AD19E00FA4999,
-    0x00FC899C000ACD9F, 0x00FA0992000AA19D, 0x00FA099000FA0993,
-    0x00FA099600FA0991, 0x00FA099400FA0997, 0x00FC1D8A00FA0995,
-    0x00FA5D8800FCC58B, 0x00FCDD8E00FA5589, 0x00FCDD8C00FA498F,
-    0x00FC318200FCD58D, 0x00FC218000FC3183, 0x00FA7D8600FC2181,
-    0x00FC218400FA6587, 0x00C045BA00FA7585, 0x00C039B800C055BB,
-    0x00C3B5BE00C04DB9, 0x00C04DBC800020CD, 0x00C0F1B200C3BDBD,
-    0x00CFDDB000C0F1B3, 0x00C021B600C0E9B1, 0x00CFC5B400C021B7,
-    0x00CF81AA00C039B5, 0x00C0DDA800CF81AB, 0x00CFB5AE00C08DA9,
-    0x00CF81AC00CFA5AF, 0x00C0E1A200CFBDAD, 0x00C0C9A000CE3DA3,
-    0x00CE01A600C3E9A1, 0x00C0FDA400C0E5A7, 0x00CE858200CE21A5,
-    0x8000EC9F00CE8983, 0x00CEB5868000ECA4, 0x00CE898400CEBD87,
-    0x00CE8D8A00CEBD85, 0x8000ECA900CE958B, 0x00CEF98E8000ECAE,
-    0x00CE818C00CEE18F, 0x00CEA19200CEF98D, 0x8000ECB300CEA193,
-    0x00CEAD968000ECB8, 0x00CEB19400C15D97, 0x00C16D9A00CEB195,
-    0x8000ECBD00C17D9B, 0x8000ECC78000ECC2, 0x00C1659C8000ECCC,
-    0x000A91A28000ECD1, 0x000A89A0029221A3, 0x000A8DA6000A89A1,
-    0x000A8DA48000ECD6, 0x000A85AA000A85A5, 0x000A89A8000131AB,
-    0x000A81AE000A91A9, 0x000A91AC8000ECDB, 0x8000ECE0000A91AD,
-    0x8000ECEA8000ECE5, 0x8000ECF48000ECEF, 0x8000ECFE8000ECF9,
-    0x8000ED088000ED03, 0x8000ED128000ED0D, 0x8000ED1C8000ED17,
-    0x8000ED268000ED21, 0x000935800299F5BD, 0x0009458200093D81,
-    0x0009458400095D83, 0x0009798600094985, 0x0009498800097987,
-    0x0009598A00094989, 0x0009B98C0009AD8B, 0x0009A98E0009B98D,
-    0x0009D9900009A98F, 0x0009E9920009D991, 0x0009F9940009E993,
-    0x0009E9960009F995, 0x0009D9980009E997, 0x0009B99A0009A199,
-    0x0009919C0009B19B, 0x0009919E0009919D, 0x000971A00009919F,
-    0x000975A2000971A1, 0x000F69A4000F79A3, 0x000DE1A6000C91A5,
-    0x8000ED2B000395A7, 0x8000ED358000ED30, 0x8000ED3F8000ED3A,
-    0x8000ED498000ED44, 0x0003E9B18000ED4E, 0x000381B30003C9B0,
-    0x0003B9B50003E1B2, 0x0003A9B7000399B4, 0x00039DB90003B1B6,
-    0x00035DBB000391B8, 0x000335BD000379BA, 0x000335BF000335BC,
-    0x0002D180000335BE, 0x0002D1820002D181, 0x0002F1840002D183,
-    0x0002F1860002F185, 0x0002D1880002F187, 0x0002D18A0002D189,
-    0x0001958C00018D8B, 0x0001958E00019D8D, 0x0001F59000018D8F,
-    0x0001F5920001FD91, 0x8000ED5300001D93, 0x8000ED5D8000ED58,
-    0x8000ED678000ED62, 0x8000ED718000ED6C, 0x8000ED7B8000ED76,
-    0x8000ED858000ED80, 0x8000ED8F8000ED8A, 0x8000ED998000ED94,
-    0x8000EDA38000ED9E, 0x8000EDAD8000EDA8, 0x8000EDB78000EDB2,
-    0x8000EDBC00AB38CC, 0x8000EDC68000EDC1, 0x8000EDD08000EDCB,
-    0x000099B18000EDD5, 0x0000A9B3000091B0, 0x0000B9B5000091B2,
-    0x0000A9B70000B1B4, 0x000099B90000B1B6, 0x8000EDDA000091B8,
-    0x000075BD000091BA, 0x000075BF000075BC, 0x8000EDDF000075BE,
-    0x8000EDE98000EDE4, 0x8000EDF38000EDEE, 0x8000EDFD8000EDF8,
-    0x80000BB98000EE02, 0x80000BBC00032C91, 0x80000B9780000B7F,
-    0x80000B9D80003D36, 0x80001CBE80000B9A, 0x80000AE680000ADA,
-    0x80000AF880000BA9, 0x80000B8280000AF2, 0x80000B0180000B85,
-    0x80000BAC80000B52, 0x8000F12080000BAF, 0x80000B1680003CBA,
-    0x80000B6A8000F125, 0x80000B2E8000F12A, 0x80000B3A80000B31,
-    0x80000B438000F12F, 0x80000BA080000BA6, 0x80000BA380003D3A,
-    0x80000B6D800022B6, 0x80000B9480000B91, 0x80000A0880000B7C,
-    0x80000A0E80000A0B, 0x8000F1398000F134, 0x8000EE0C8000EE07,
-    0x8000EE168000EE11, 0x8000EE208000EE1B, 0x8000EE2A8000EE25,
-    0x8000EE348000EE2F, 0x00021C828000EE39, 0x8000EE438000EE3E,
-    0x8000EE4D8000EE48, 0x8000EE578000EE52, 0x8000EE618000EE5C,
-    0x8000EE6B8000EE66, 0x8000EE758000EE70, 0x8000EE7F8000EE7A,
-    0x8000EE898000EE84, 0x8000EE938000EE8E, 0x8000EE9D8000EE98,
-    0x8000EEA78000EEA2, 0x8000EEB18000EEAC, 0x8000EEBB8000EEB6,
-    0x8000EEC58000EEC0, 0x8000EECF8000EECA, 0x8000EED98000EED4,
-    0x000191948000EEDE, 0x0001859200019195, 0x000195908000EEE3,
-    0x8000EEE800018D91, 0x8000EEF28000EEED, 0x8000EEFC8000EEF7,
-    0x8000EF068000EF01, 0x00066D868000EF0B, 0x0006698400067587,
-    0x0006498200066585, 0x0006498000064983, 0x0001E58E00064981,
-    0x0006018C0001ED8F, 0x00064D8A0001ED8D, 0x00064D880003D1BA,
-    0x0002398900064589, 0x0290698B00023188, 0x0002518D0002318A,
-    0x0002518F0290718C, 0x000201810002598E, 0x029335838000EF10,
-    0x0290098500020582, 0x0002098700021984, 0x000229990003B4F0,
-    0x00F3AD9B00022198, 0x0002259D0002219A, 0x0002259F0002259C,
-    0x000231910002259E, 0x0290ED9300022990, 0x0293B99500023992,
-    0x0002199700022D94, 0x000219A900022196, 0x000215AB00DB75A8,
-    0x000235AD029055AA, 0x029195AF000235AC, 0x000225A1000235AE,
-    0x000221A3000225A0, 0x029975A50293C1A2, 0x00022DA7000239A4,
-    0x00020DB900022DA6, 0x8000EF1500020DB8, 0x8000EF1A00021DBA,
-    0x8000EF248000EF1F, 0x8000EF298000EF2E, 0x00025DB3000249B0,
-    0x00024DB500025DB2, 0x00023DB700024DB4, 0x0003D58000023DB6,
-    0x0003D5820003DD81, 0x0003F5840003ED83, 0x0003F5860003FD85,
-    0x0003D5880003ED87, 0x0003D58A0003DD89, 0x0002398C0002298B,
-    0x0002298E0002398D, 0x000259900002298F, 0x0002699200025991,
-    0x0002799400026993, 0x0002699600027995, 0x0002599800026997,
-    0x0002299A00025999, 0x0002399C0002299B, 0x0002299E0002399D,
-    0x0002D9A00002299F, 0x0002E9A20002D9A1, 0x0002F9A40002E9A3,
-    0x0002E9A60002F9A5, 0x0002D9A80002E9A7, 0x000229AA0002D9A9,
-    0x000225AC000229AB, 0x00021DAE000215AD, 0x00026DB0000215AF,
-    0x00027DB2000275B1, 0x003D488400020C91, 0x003D508400021C91,
-    0x8000EF3300020491, 0x003D308C003C3C87, 0x8000EF38003D948D,
-    0x8000EF428000EF3D, 0x00D5C0998000EF47, 0x0028C89B003ECDA7,
-    0x003DECB1003D9CF0, 0x003D80F0003EC5A7, 0x003EF5BE0003D0B2,
-    0x0003C4B9003EFD97, 0x00024C91000258F0, 0x0003ACB90002B492,
-    0x003E9DB00003B8B8, 0x00025C91000248F0, 0x003FA1AF0002A096,
-    0x00025C91000248F0, 0x8000EF4C00028C96, 0x003E89BA8000EF51,
-    0x8000EF56003F99AF, 0x8000EF5B003F95BD, 0x0003A8B58000EF60,
-    0x8000EF6A8000EF65, 0x8000EF748000EF6F, 0x8000EF7E8000EF79,
-    0x8000EF888000EF83, 0x8000EF928000EF8D, 0x8000EF9C8000EF97,
-    0x8000EFA68000EFA1, 0x8000EFB08000EFAB, 0x8000EFBA8000EFB5,
-    0x8000EFC48000EFBF, 0x8000EFCE8000EFC9, 0x8000EFD88000EFD3,
-    0x000240F08000EFDD, 0x002C70A500028491, 0x0002CCA4000009B0,
-    0x0001E8B98000EFE2, 0x8000EFEC8000EFE7, 0x8000EFF10003C092,
-    0x8000EFFB8000EFF6, 0x8000F0000003C096, 0x8000F00A8000F005,
-    0x8000F0148000F00F, 0x8000F01E8000F019, 0x8000F0288000F023,
-    0x8000F0328000F02D, 0x8000F03700032882, 0x8000F03C00038084,
-    0x8000F0468000F041, 0x8000F0508000F04B, 0x003C888D8000F055,
-    0x8000F05F8000F05A, 0x8000F0698000F064, 0x8000F0738000F06E,
-    0x8000F07D8000F078, 0x8000F0878000F082, 0x8000F0998000F090,
-    0x8000F0B38000F0A6, 0x8000F0CD8000F0C0, 0x8000F0FD8000F0DA,
-    0x8000F1068000F0E3, 0x000019808000F0F0, 0x8000F11300014CA2,
-    0x8000095A000380A4, 0x000055BB80000AA4, 0x800020ED800020F5,
-    0x000035BF8000F13E, 0x000045BD000035BE, 0x003DF9A0000045BC,
-    0x000269A2003DF9A1, 0x000261A4000269A3, 0x000265A600026DA5,
-    0x0002A9A8000299A7, 0x0002B9AA0002A9A9, 0x0002A9AC0002B9AB,
-    0x000299AE0002A9AD, 0x0002E9B0000299AF, 0x0002F9B20002E9B1,
-    0x0002E9B40002F9B3, 0x000299B60002E9B5, 0x0002A9B8000299B7,
-    0x0002B9BA0002A9B9, 0x0002A9BC0002B9BB, 0x000299BE0002A9BD,
-    0x0003019A000299BF, 0x0003F1980003019B, 0x0003019E0003F199,
-    0x0003119C0003019F, 0x0003FD920003119D, 0x0003F9900003C593,
-    0x0003DD960003F591, 0x0003DD940003C197, 0x0003F58A0003D595,
-    0x0003E5880003FD8B, 0x0003958E0003FD89, 0x0003E58C00039D8F,
-    0x0003F18200039D8D, 0x0003ED800003F183, 0x0003D5860003E581,
-    0x0003E1840003DD87, 0x0000C8870003DD85, 0x00025C8500000886,
-    0x8000F143000000C3, 0x8000F14D8000F148, 0x8000F1578000F152,
-    0x0002B59F8000F15C, 0x8000F1610002B59E, 0x8000F16B8000F166,
-    0x000A40978000F170, 0x00094895000B4496, 0x0006509300080094,
-    0x0005589100075492, 0x0015609F00040090, 0x00136C9D0014009E,
-    0x000E709B00106C9C, 0x000D7899000F749A, 0x02BE0DB3000C0098,
-    0x02BE09B102BE0DB2, 0x02BE25B702BED5B0, 0x02BE0DB502BE35B6,
-    0x02BDB9BB02BE0DB4, 0x028325B902BDB5BA, 0x02BDF5BF02BE61B8,
-    0x02BD89BD02BD8DBE, 0x8000F1750283BDBC, 0x000000F2000275A4,
-    0x000275A1000000F3, 0x000275A3000275A0, 0x8000F17A000275A2,
-    0x8000F17F000000F9, 0x8000F184000000FB, 0x8000F18E8000F189,
-    0x02A1A5808000F193, 0x02A1E58202A1C581, 0x02A3D58402AC3583,
-    0x02AC458602AC5585, 0x02ACC98802ACBD87, 0x02A9F18A02ADDD89,
-    0x02A8398C02A82D8B, 0x02A4798E02A82D8D, 0x02A0499002A0398F,
-    0x02A06D9202A05991, 0x02A0D19402A01D93, 0x009049968000F198,
-    0x02A13D9802B4F197, 0x02A1CD9A02A1E599, 0x02A19D9C02A1AD9B,
-    0x02A3BD9E02AC4D9D, 0x02ACDDA002AC3D9F, 0x02AC61A2009089A1,
-    0x02B509A402AD75A3, 0x02B551A602A895A5, 0x02A4E1A802A885A7,
-    0x02A0A1AA02A0A1A9, 0x02A095AC02A0B1AB, 0x02A039AE02A0E5AD,
-    0x02A07DB002A029AF, 0x02A195B202A181B1, 0x02A175B402A14DB3,
-    0x02A135B602A115B5, 0x02A325B802ACE5B7, 0x02ACB5BA02ACA5B9,
-    0x02AC19BC02AC4DBB, 0x02A921BE02AD0DBD, 0x02A9019702A8FDBF,
-    0x02ACF19502A8DD96, 0x02ADB19302ADE594, 0x02AD599102AD4992,
-    0x02A1719F02A2D990, 0x02A1359D02A1019E, 0x02A1359B02A1259C,
-    0x02A9219902A5759A, 0x02A1D58702A93598, 0x02A1198502A1C586,
-    0x02A14D8302A16984, 0x02A15D8102A15D82, 0x02AD598F02A51D80,
-    0x02A0A98D02A0898E, 0x02A0F18B02A0C98C, 0x02A03D8902A0298A,
-    0x02A19DB702A1C188, 0x02A18DB502A18DB6, 0x02A989B302A5CDB4,
-    0x02A999B102B45DB2, 0x02A021BF02B405B0, 0x02A0EDBD02A0F9BE,
-    0x02A125BB02A111BC, 0x02A1E9B902A135BA, 0x02A001A702A199B8,
-    0x02A049A502A061A6, 0x02B57DA302A091A4, 0x02B545A10091C5A2,
-    0x02AC19AF02A15DA0, 0x0091E5AD02AD0DAE, 0x02ADB1AB02ADB1AC,
-    0x02ADC1A902A231AA, 0x02AA098002A011A8, 0x02A6498202AA1D81,
-    0x02A2198402A20983, 0x02A23D8602A20985, 0x02A2A18802A24D87,
-    0x0092398A02B6B989, 0x02A36D8C02B6818B, 0x02A39D8E02A3B58D,
-    0x02A3AD9002A3FD8F, 0x02A18D9202AE7D91, 0x02AE0D9402AE0D93,
-    0x02AEB19600925995, 0x02B7F99802AFA597, 0x02B7A19A02AA6599,
-    0x02A6319C02AA759B, 0x02A2719E8000F19D, 0x8000F1A202A2619F,
-    0x02A209A28000F1A7, 0x8000F1B18000F1AC, 0x02A3C5A602A3D1A5,
-    0x8000F1BB8000F1B6, 0x02A345AA02A365A9, 0x02A175AC02AE95AB,
-    0x02AEE5AE8000F1C0, 0x02AE29B002AE1DAF, 0x02AB11B202AF3DB1,
-    0x02AAD9B402AACDB3, 0x02A699B602AACDB5, 0x02A2E9B802A2D9B7,
-    0x8000F1C502A2F9B9, 0x8000F1CA02A2BDBB, 0x0092E9BE02B669BD,
-    0x02A25D8002B651BF, 0x02A2AD8202A28581, 0x8000F1CF02A2CD83,
-    0x02A0DD8602AF2D85, 0x02AF7D8802AF5D87, 0x02AFC18A00932989,
-    0x02B6A98C02AED58B, 0x02B6F18E02AB358D, 0x02A7019002AB258F,
-    0x02A3419202A34191, 0x02A3759402A35193, 0x02A3D99602A30595,
-    0x02A3DD9802A3C997, 0x02A2359A02A22199, 0x02A2D59C02A2ED9B,
-    0x02A2959E02A2B59D, 0x02A045A002AF459F, 0x02AFD5A202AFC5A1,
-    0x02AF79A402AF2DA3, 0x02AA41A602AE6DA5, 0x02ABA9A802AB9DA7,
-    0x02A7E9AA02ABBDA9, 0x02A3B9AC02A3A9AB, 0x02A39DAE02A3A9AD,
-    0x02A341B002A3EDAF, 0x0093D9B202B759B1, 0x02A28DB402B761B3,
-    0x02A27DB602A255B5, 0x02A20DB802A21DB7, 0x02A02DBA02AFDDB9,
-    0x02AFADBC02AFADBB, 0x02AF11BE0093F9BD, 0x02ACBD8002AE05BF,
-    0x02AD558202AD4181, 0x02ADB58402AD8D83, 0x02ADF5868000F1D4,
-    0x8000F1DE8000F1D9, 0x02A0758A8000F1E3, 0x02A0D98C02A08D8B,
-    0x02A5E18E02A1CD8D, 0x02A4499002A43D8F, 0x02A80992800016C0,
-    0x02AC599402AC4993, 0x02AC7D9602AC4995, 0x02ACE19802AC0D97,
-    0x009C799A02B8F999, 0x02AD2D9C02B8C19B, 0x02ADDD9E02ADF59D,
-    0x02AD6DA002ADBD9F, 0x02AF4DA202A0BDA1, 0x02A0CDA402A0CDA3,
-    0x02A071A6009C99A5, 0x02B939A802A165A7, 0x02B961AA02A4A5A9,
-    0x02A8F1AC02A4B5AB, 0x02ACB1AE02ACB1AD, 0x02ACE5B002ACA1AF,
-    0x02AC49B202AC95B1, 0x02AC6DB402AC59B3, 0x02AD85B602AD91B5,
-    0x02AD45B802AD5DB7, 0x02AD05BA02AD25B9, 0x02AF35BC02A0D5BB,
-    0x02A0A5BE02A0B5BD, 0x02A1399302A05DBF, 0x02ACC99102ACE992,
-    0x02A1B19702ACA990, 0x800016F902A14996, 0x02A5219B02AED994,
-    0x02A0D19902A4FD9A, 0x02AD359F02A1C598, 0x800016C302A9759E,
-    0x02A5598302A5359C, 0x02A5498102B88D82, 0x02AD4D8702B8D580,
-    0x02AD5D8580000352, 0x8000034F02A91D84, 0x02AD398902ADE58A,
-    0x02ACF18F02AD4988, 0x02AC3D8D02AC298E, 0x02A079B380000362,
-    0x009D85B102A16DB2, 0x02A589B702A1D1B0, 0x02A599B502B85DB6,
-    0x02ADBDBB02B805B4, 0x02ADADB980000485, 0x8000052402A9EDB8,
-    0x02ADE9BD02AD35BE, 0x02AD91A302AD99BC, 0x02ADD5A102ADE1A2,
-    0x02B97DA702ADC5A0, 0x02B945A5009DC5A6, 0x02AC21AB02AD5DA4,
-    0x02AC69A902AC41AA, 0x02A1B1AF02ACB1A8, 0x02A1C1AD02AE31AE,
-    0x02A2E98002AC11AC, 0x02A7D18202A3FD81, 0x02A6198402A60D83,
-    0x02AA598602A60D85, 0x02AE298802AE1987, 0x02AE0D8A02AE3989,
-    0x02AEB18C02AE7D8B, 0x009E298E02BAA98D, 0x02AF1D9002BA918F,
-    0x02AFED9202AFC591, 0x02AFBD9402AF8D93, 0x02AD9D9602A26D95,
-    0x02A23D9802A21D97, 0x02A2819A009E6999, 0x02BBE99C02A3959B,
-    0x02BBB19E02A6759D, 0x02AAC1A002A6659F, 0x02AE81A202AE81A1,
-    0x02AEB5A402AE91A3, 0x02AE19A602AEC5A5, 0x02AE1DA802AE09A7,
-    0x02AFF5AA02AFE1A9, 0x02AF15AC02AF2DAB, 0x02AF55AE02AF75AD,
-    0x02AD05B002A285AF, 0x02A295B202A285B1, 0x02A239B402A26DB3,
-    0x02A701B602A32DB5, 0x02A6E9B802A6DDB7, 0x02AAA9BA02A6FDB9,
-    0x02AEF9BC02AEE9BB, 0x02AEDDBE02AEE9BD, 0x02AF818002AEADBF,
-    0x009F198202BB9981, 0x02AE4D8402BBA183, 0x02AEBD8602AE9585,
-    0x02AECD8802AEDD87, 0x02ACED8A02A31D89, 0x02A36D8C02A36D8B,
-    0x02A3D18E009F398D, 0x02BAD99002A2C58F, 0x02BA819202A74591,
-    0x02AB119402A75593, 0x02AF519602AF5195, 0x02AF459802AF4197,
-    0x02AFE99A02AF3599, 0x02AFCD9C02AFF99B, 0x02AE259E02AE319D,
-    0x02AE25A002AEFD9F, 0x02AE65A202AE45A1, 0x02AC55A402A3B5A3,
-    0x02A3C5A602A3D5A5, 0x02A349A802A33DA7, 0x02A671AA02A25DA9,
-    0x02A7B9AC02A7ADAB, 0x02ABF9AE02A7ADAD, 0x02AFC9B002AFB9AF,
-    0x02AFEDB202AFD9B1, 0x02AF51B402AF9DB3, 0x009FC9B602BB49B5,
-    0x02AEBDB802BB71B7, 0x02AE4DBA02AE65B9, 0x02AE1DBC02AE2DBB,
-    0x02AC3DBE02A3CDBD, 0x02A8258002A3BDBF, 0x02A8898202A85581,
-    0x02A8AD8402A89983, 0x02A9458602A95185, 0x02A9858802A99D87,
-    0x02A9C58A02A9E589, 0x02ABF58C02A4158B, 0x02A4658E02A4758D,
-    0x02A4A99002A49D8F, 0x02A1919202A5BD91, 0x02A0599402A04D93,
-    0x02AC199602A04D95, 0x02A8699802A85997, 0x02A84D9A02A87999,
-    0x02A8F19C02A83D9B, 0x0098699E02BCE99D, 0x02A9DDA002BCD19F,
-    0x02A92DA202A905A1, 0x02A97DA402A94DA3, 0x02AB5DA602A4ADA5,
-    0x02A4FDA802A4DDA7, 0x02A441AA0098A9A9, 0x02BD29AC02A555AB,
-    0x02BD71AE02A0B5AD, 0x02AC81B002A0A5AF, 0x02A8C1B202A8C1B1,
-    0x02A8F5B402A8D1B3, 0x02A859B602A885B5, 0x02A85DB802A849B7,
-    0x02A9B5BA02A9A1B9, 0x02A955BC02A96DBB, 0x02A915BE02A935BD,
-    0x02A1019F02A4C5BF, 0x02A4F19D02A0DD9E, 0x02A5B19B02A5E59C,
-    0x02A5599902A5499A, 0x02A5199702AAD998, 0x02A8E99502A8C996,
-    0x02A8B19302A88994, 0x02A87D9102A86992, 0x02A9D58F02A98190,
-    0x02A9198D02A9C58E, 0x02A94D8B02A9698C, 0x02A95D8902A95D8A,
-    0x02A1798702AD1D88, 0x02A1698502BCAD86, 0x02A4898302BCF584,
-    0x0099758102A59D82, 0x02A99DBF02A52180, 0x02A98DBD02A98DBE,
-    0x02A189BB02ADCDBC, 0x02A199B902BC5DBA, 0x02A459B702BC05B8,
-    0x0099A5B502A54DB6, 0x02A5F1B302A5F1B4, 0x02A581B102AA71B2,
-    0x02A801AF02A851B0, 0x02A849AD02A861AE, 0x02BD7DAB02A891AC,
-    0x02BD45A90099C5AA, 0x02A9B1A702A95DA8, 0x02A9F5A502A9C1A6,
-    0x02A9F5A302A9E5A4, 0x02A1E1A102ADB5A2, 0x02A9C58002A1F5A0,
-    0x02A6558202A64581, 0x02A6F98402A6AD83, 0x02A3C18602A7ED85,
-    0x02A2298802A21D87, 0x02AE698A02A23D89, 0x02AA398C02AA298B,
-    0x02AA1D8E02AA298D, 0x02AAC19002AA6D8F, 0x009A599202BED991,
-    0x02AB0D9402BEE193, 0x02ABFD9602ABD595, 0x02AB8D9802AB9D97,
-    0x02A9AD9A02A65D99, 0x02A62D9C02A62D9B, 0x02A6919E009A799D,
-    0x02BF19A002A7859F, 0x02BF41A202A285A1, 0x000F49A402A295A3,
-    0x80000577000F49A5, 0x02BF79A8800005B2, 0x02BC39AA02BC2DA9,
-    0x02BA59AC02BC09AB, 0x02BBD1AE02BDF5AD, 0x02BB69B002BD01AF,
-    0x02BA11B202BA09B1, 0x02BA21B402A2E9B3, 0x02B6C5B602BA2DB5,
-    0x02B871B802BBBDB7, 0x02BB95BA02BD59B9, 0x02B6A5BC02BB7DBB,
-    0x02BB01BE02BB05BD, 0x02B7698002B8F5BF, 0x001C0182000EE181,
-    0x02BD818402BD8583, 0x02BBF18602BDB185, 0x001F218802BC5D87,
-    0x0095418A02BC9989, 0x02BBE98C02BBE18B, 0x02BBC98E02A3118D,
-    0x02B75D9002BBC58F, 0x02B3699202BA2591, 0x02BA2D9402BA3193,
-    0x001F519602BAC595, 0x02BA999802BAAD97, 0x001F799A02B96D99,
-    0x02BB999C000E859B, 0x02BBB59E02BCC99D, 0x02B3A1A002BA8D9F,
-    0x02BE51A202BAE5A1, 0x02BD01A402BD05A3, 0x02BB71A602BD31A5,
-    0x02BAC9A802BCDDA7, 0x02BA01AA02BC19A9, 0x02BB69AC02BB61AB,
-    0x02BB49AE02A391AD, 0x02B7DDB002BB45AF, 0x02B959B202BAA5B1,
-    0x02BAADB402BC71B3, 0x02B78DB602BA45B5, 0x02BA19B802BA2DB7,
-    0x02B781BA02B9EDB9, 0x001CF9BC000E09BB, 0x02BD69BE02BD7DBD,
-    0x80000C0302BD59BF, 0x80000C4A00006190, 0x80000C56800005D8,
-    0x00115480800005EF, 0x8000060C0001DC81, 0x8000247280000623,
-    0x800007188000087F, 0x02ABB8CC80000842, 0x000B408402AB5CCD,
-    0x00020C8600094485, 0x800008108000084B, 0x800007968000087C,
-    0x800007BD80000799, 0x8000253C800006BA, 0x000C00888000270F,
-    0x80002F5C0004DC89, 0x80002EFD80002DE1, 0x000488A480003866,
-    0x800031FA80002C2E, 0x80002D35800033E6, 0x80002CCD80003087,
-    0x80002D21800031D3, 0x800038C28000364E, 0x80002E1980002EE1,
-    0x800031378000339A, 0x8000394280003163, 0x8000312F80003157,
-    0x800035E2800035A6, 0x8000337A80002E05, 0x80002D39800034CE,
-    0x000AF0BA0004E8AF, 0x0005F8B880002E65, 0x02A5E980000800B9,
-    0x0001098202A24581, 0x008B798402A2B183, 0x02A5C18602A5D985,
-    0x02A5D18802BD3987, 0x02A9358A02A5DD89, 0x02AD118C02A44D8B,
-    0x02A4458E02A4498D, 0x0001499002A4AD8F, 0x02A4B19202A4B591,
-    0x0001419402A74593, 0x02A5B1960010BD95, 0x02A5AD9802A2E197,
-    0x02AD499A02A49599, 0x02A0A99C02A40D9B, 0x02A3E99E02A3FD9D,
-    0x02A569A002A3D99F, 0x02A4E1A202A2C5A1, 0x02A439A402A231A3,
-    0x02A541A602A559A5, 0x02A551A802BDB9A7, 0x02A9B5AA02A55DA9,
-    0x02A721AC02A4CDAB, 0x02A4C5AE02A209AD, 0x02A995B002A42DAF,
-    0x02A431B202A435B1, 0x02A9B9B402A7C5B3, 0x0002D1B6001031B5,
-    0x02A371B802A355B7, 0x02A501BA02A341B9, 0x0001F1BC02A2ADBB,
-    0x008B91BE02A249BD, 0x02BE258102A531BF, 0x02A6F18302A6DD80,
-    0x02A7718502A6FD82, 0x02A7658702AA0984, 0x02A7B18902AE3D86,
-    0x02A7D98B02A75988, 0x02A4398D0002258A, 0x0013D18F02A7CD8C,
-    0x02A1FD9100022D8E, 0x02A7B99302A6AD90, 0x02A7319502A68192,
-    0x02A0D19702AE7594, 0x02A0C59902A38596, 0x02A1299B02A0F598,
-    0x02A1CD9D02A6859A, 0x02A6B59F02A71D9C, 0x02BEA5A102A7D59E,
-    0x02A671A302A65DA0, 0x02A7F1A502A67DA2, 0x02A125A702AA89A4,
-    0x02A731A902A40DA6, 0x02A759AB02A7D9A8, 0x02A4B9AD02AAF9AA,
-    0x00135DAF02A74DAC, 0x02A049B102AAD5AE, 0x02A06DB30001CDB0,
-    0x02A191B502A05DB2, 0x02A165B702A63DB4, 0x02A62DB90002DDB6,
-    0x02BECDBB00888DB8, 0x02A609BD02A635BA, 0x02A799BF02A605BC,
-    0x000000C602AAE1BE, 0x80000BF18000397A, 0x8000247A800024B6,
-    0x8000249880002484, 0x02AA4CCC800024C0, 0x800024A202AAA4CD,
-    0x02A890CC800024CA, 0x02ABD8CC02AB74CD, 0x02AD25A002AA4CCD,
-    0x02AD3DA202AD35A1, 0x02ADD1A4029345A3, 0x029355A602ADD9A5,
-    0x02AE65A802AC9DA7, 0x02AE3DAA02AF51A9, 0x00D5BDAC029351AB,
-    0x8000248E029321AD, 0x02931D90800024AC, 0x02AE099202AEDD91,
-    0x02AE1D9402AE0593, 0x02AE0D9602AE0595, 0x02AE159802AE3D97,
-    0x0293259A02AE6999, 0x0293B19C02ADBD9B, 0x02ADFD9E02AD859D,
-    0x02AD89800293B99F, 0x02ADAD820293C181, 0x02ADA58402ADBD83,
-    0x02AD59860293DD85, 0x0293F1880293D187, 0x00D4B58A02AD5589,
-    0x02AEF58C0293318B, 0x0292F18E02AEE18D, 0x02A1F9800293B18F,
-    0x0004198202A20D81, 0x02A0F9840015E583, 0x02A0D58602A7A985,
-    0x02A8018802A1ED87, 0x02AF118A02A14589, 0x8000308B02AF158B,
-    0x00838D8E80002DF5, 0x02BFFD9002B8758F, 0x0083F59202BF8D91,
-    0x00839D9400838993, 0x0083859600838195, 0x0083D5980083B997,
-    0x02BFD59A02B82D99, 0x0083CD9C02BFA59B, 0x0083B59E0083B19D,
-    0x00835DA00083A99F, 0x00833DA2008361A1, 0x02BF2DA402B8C5A3,
-    0x008325A602BF5DA5, 0x00836DA8008359A7, 0x008375AA008371A9,
-    0x008305AC008349AB, 0x02BF05AE02B8FDAD, 0x00837DB002BF75AF,
-    0x008305B2008301B1, 0x00830DB4008319B3, 0x00836DB6008331B5,
-    0x02BF5DB802B895B7, 0x008355BA02BF2DB9, 0x00833DBC008329BB,
-    0x008325BE008321BD, 0x02A1CD99008319BF, 0x02B92D9B02A0AD98,
-    0x02A1E99D02A1D59A, 0x02A0799F02A1E59C, 0x02A7A99102AD019E,
-    0x02A78D9302A4FD90, 0x02A6719502A7BD92, 0x02A6859702A1DD94,
-    0x0014A98902A05596, 0x02A6F58B00055588, 0x02A0A18D02A1A58A,
-    0x02A0398F02A1998C, 0x02A01D8102A97D8E, 0x02A0F98302A94580,
-    0x02A0818502A01182, 0x02A3718700057D84, 0x02A0E1B902A08586,
-    0x02A0F5BB02AD99B8, 0x02A001BD02A9ADBA, 0x02A069BF02A0E9BC,
-    0x02A61DB1000595BE, 0x02A165B30005A5B0, 0x02B995B5008FC5B2,
-    0x02A141B702A16DB4, 0x001425A902A14DB6, 0x02A741AB02ADADA8,
-    0x02A775AD0006C5AA, 0x02A699AF02A745AC, 0x02A65DA102A135AE,
-    0x02A079A302A375A0, 0x02A001A502A091A2, 0x02A3F1A702ADA1A4,
-    0x0017ED8002A005A6, 0x00151D8200151581, 0x00150D8400151583,
-    0x00153D8600153585, 0x0015158800152D87, 0x0015158A00151D89,
-    0x0015D18C0015ED8B, 0x0015C98E0015DD8D, 0x0015B9900015C98F,
-    0x00121D920015B991, 0x0010D19400103993, 0x0010D5960010D195,
-    0x0010FD980010CD97, 0x0010DD9A0010C599, 0x0010819C0010859B,
-    0x001E6D9E0011659D, 0x001EA1A0001E619F, 0x0005F588001EA1A1,
-    0x80002BFA80002C02, 0x8000F1E880002C0A, 0x0005B58280002C55,
-    0x0005A5800005BD83, 0x00059D860005BD81, 0x0005A5840005CD87,
-    0x02BA89B100059585, 0x02BD01B3008171B0, 0x008115B502BD71B2,
-    0x00810DB7008169B4, 0x008105B9008111B6, 0x0002D990008139B8,
-    0x80002C690002D991, 0x80002C7980002C71, 0x0027E18080002C7D,
-    0x0027FD820027E181, 0x80002C8100262583, 0x00262D8600266985,
-    0x002669880027E587, 0x0026498A0026F589, 0x0026598C0026498B,
-    0x0026018E0026598D, 0x0026199000266D8F, 0x0026159200260591,
-    0x00266D9400267D93, 0x0027A9960027B995, 0x00264D980027A197,
-    0x0026259A00262199, 0x001F359C0026399B, 0x001E619E00360D9D,
-    0x80002ABA001E459F, 0x00277DA2002761A1, 0x0026E1A48000F1FC,
-    0x80002CA580002CA1, 0x80002CC1002765A7, 0x0026C9AA002675A9,
-    0x0026D9AC0026C9AB, 0x002681AE0026D9AD, 0x002699B00026EDAF,
-    0x002695B2002685B1, 0x0026EDB48000F1ED, 0x002729B6002739B5,
-    0x80002ABE002721B7, 0x80002CAD0026A1B9, 0x8000F1F70026B9BB,
-    0x80002CE580002CDD, 0x0028E18080002C3E, 0x0028FD820028E181,
-    0x0029618400292583, 0x00292D8600296985, 0x002969880028E587,
-    0x80002CED0029F589, 0x0029598C0029498B, 0x0029018E0029598D,
-    0x0029199000296D8F, 0x0029159200290591, 0x00296D9400297D93,
-    0x0028A9960028B995, 0x00294D980028A197, 0x0029259A00292199,
-    0x8000F3D20029399B, 0x80002AC680002D15, 0x80002D3180002D2D,
-    0x00287DA2002861A1, 0x80002D3D0029A5A3, 0x0029ADA60029E9A5,
-    0x0029E9A8002865A7, 0x80002ACA002975A9, 0x0029D9AC0029C9AB,
-    0x002981AE0029D9AD, 0x002999B00029EDAF, 0x002995B2002985B1,
-    0x0029EDB40029FDB3, 0x002829B6002839B5, 0x0029CDB8002821B7,
-    0x0029A5BA0029A1B9, 0x80002D750029B9BB, 0x80002DA180002D79,
-    0x001055800000003F, 0x0010758200106581, 0x0010A1840010A983,
-    0x0017858600178D85, 0x0009358800091187, 0x00090D8A00090589,
-    0x80002DBD80002DAD, 0x80002DC980002DC5, 0x00FB35908000F20B,
-    0x00FBC99200FB3991, 0x00FBD99400FBC993, 0x00FBC99600FBD995,
-    0x00FBF99800FBC997, 0x00FBC99A00FBF999, 0x00FBDD9C00FBC99B,
-    0x00FBCD9E00FBC59D, 0x00FB3DA000FBC59F, 0x00FB4DA200FB45A1,
-    0x00FB5DA400FB45A3, 0x00FB4DA600FB45A5, 0x00FB7DA800FB45A7,
-    0x000981AA00FB45A9, 0x02B1F1AC02BDA5AB, 0x000995AE00DC7DAD,
-    0x02B981B0800033CA, 0x02BDC1B202BDC1B1, 0x02BDF5B402BDD1B3,
-    0x02BD59B602BD85B5, 0x02BD5DB802BD49B7, 0x02BCB5BA02BCA1B9,
-    0x02BC55BC02BC6DBB, 0x02BC15BE02BC35BD, 0x0000008002B1C5BF,
-    0x80002DDD00131DBE, 0x002B45BB001249BC, 0x002B5DB9002B59BA,
-    0x002ADDB780002DE5, 0x002AC5B5002AD5B6, 0x8000F210002B11B4,
-    0x002B79B1002B69B2, 0x002B11AF002B65B0, 0x002B25AD002B7DAE,
-    0x80002DF1002B25AC, 0x002B89A9002B35AA, 0x002A99A7002B15A8,
-    0x80002E0180002DF9, 0x80002E29002B1DA4, 0x002A9DA1002A81A2,
-    0x0013B99F80002E15, 0x003BF19D80002E21, 0x002BC59B80002E25,
-    0x002BDD9980002E2D, 0x002A5D9780002E35, 0x80002E4180002E39,
-    0x80002E49002B9194, 0x002BF991002BE992, 0x002B918F80002E4D,
-    0x002BA58D002BFD8E, 0x002BB58B80002E5D, 0x002B098980002E61,
-    0x002A198780002E6D, 0x80002E9580002E91, 0x80002E8180002E79,
-    0x80002EDD002A0182, 0x02BCC58080002EA5, 0x02B3558202B34581,
-    0x02B3F98402B3AD83, 0x02B6C18602B2ED85, 0x02B7298802B71D87,
-    0x000B198A02B73D89, 0x000B058C00DEB58B, 0x000B698E000B798D,
-    0x000540A0000B698F, 0x000648A2000444A1, 0x007C00A400784CA3,
-    0x007E58A6007D54A5, 0x007260A800705CA7, 0x80002EAD80002EA9,
-    0x80002EB580002EA1, 0x80002EC980002EB1, 0x80002EE58000F215,
-    0x80002EF180002EE9, 0x80002F2880002F05, 0x8000F21F8000F21A,
-    0x80002F3880002F34, 0x000BF1AA80002F3C, 0x000BD1AC000BF1AB,
-    0x80002F4480002F30, 0x80002AD280002ACE, 0x80002F5480002F4C,
-    0x80002F708000F224, 0x80002F8080002F74, 0x80002F888000F229,
-    0x8000306B80002F8C, 0x80002FA080002AD6, 0x80002FB880002FBC,
-    0x000A8D808000F233, 0x000A9D82000A9581, 0x000A8D84000A9583,
-    0x000AE98600C6D185, 0x000AD988000AE987, 0x00C6E18A000AD989,
-    0x000BB18C000B3D8B, 0x00C6F58E000BB98D, 0x000BD990000BA98F,
-    0x000B8592000BE591, 0x000BB994000BAD93, 0x000BA996000CF595,
-    0x000B9998000BA997, 0x0004419A000B9999, 0x0004019C00C1199B,
-    0x00040D9E00F6699D, 0x0004FDA00004059F, 0x000C61A2000405A1,
-    0x000439A400041DA3, 0x00C649A600C11DA5, 0x00C12DA800C649A7,
-    0x00045DAA000455A9, 0x000449AC00CF55AB, 0x000595AE00C66DAD,
-    0x0005F5B000058DAF, 0x0005F1B20005FDB1, 0x0005EDB40005F5B3,
-    0x0006D1B6000569B5, 0x0006D1B80006C1B7, 0x00068DBA0006B5B9,
-    0x000681BC000685BB, 0x00067DBE000689BD, 0x00079580000671BF,
-    0x0007AD8200079981, 0x0007A1840007A583, 0x0007A5860007A985,
-    0x0007ED8800C08987, 0x0007F58A0007F189, 0x00C7D18C0007C98B,
-    0x0007D58E0007DD8D, 0x00C701900007CD8F, 0x0007B5920007BD91,
-    0x0006399400062993, 0x0006299600063995, 0x0006199800062997,
-    0x0006E99A00061999, 0x0006F99C0006E99B, 0x0006F59E000DF59D,
-    0x000615A00006ED9F, 0x000615A200061DA1, 0x000635A400062DA3,
-    0x000621A6000621A5, 0x000661A8000621A7, 0x000661AA000661A9,
-    0x00067DAC000665AB, 0x00066DAE000665AD, 0x000619B000CE55AF,
-    0x000629B2000619B1, 0x000639B4000629B3, 0x000635B600F721B5,
-    0x00F6E5B800062DB7, 0x000611BA000611B9, 0x0003B1BC000611BB,
-    0x0003B1BE0003B1BD, 0x000E999C0003B1BF, 0x000EA99E000E999D,
-    0x000E9D98000EA99F, 0x000E8D9A000E8599, 0x00F6E19400C64D9B,
-    0x000EAD96000EA595, 0x00F5619000E46D97, 0x000E4592000E4D91,
-    0x80002FC400EAC193, 0x80002FD08000F238, 0x80002FD480002FCC,
-    0x80002FE880002FE4, 0x80002FF880002ADA, 0x8000300480003000,
-    0x000E198080003008, 0x00DA6D82000E1981, 0x8000F23D80002ADE,
-    0x8000302B80002AE2, 0x000E75B88000302F, 0x000E6DBA000E7DB9,
-    0x000E7DB4000D99BB, 0x00DEB9B6000E45B5, 0x00F271B000F5D5B7,
-    0x000E6DB2000E65B1, 0x00DE95AC000E65B3, 0x00DE91AE00E4E9AD,
-    0x000E15A8000E01AF, 0x00E4F9AA00F645A9, 0x000E55A4000E15AB,
-    0x000E55A6000E5DA5, 0x000E59A0000E2DA7, 0x00F519A2000E59A1,
-    0x00015180000E4DA3, 0x0001518200015181, 0x0001718400015183,
-    0x0001718600017185, 0x00015D8800017187, 0x0001AD8A0001A589,
-    0x0001BD8C0001A58B, 0x00CDED8E007C018D, 0x00F46D90007F3D8F,
-    0x007F4D92007F4191, 0x007F5D94007F4593, 0x007F4996007F5995,
-    0x007F7998007F4997, 0x007F499A007F7999, 0x007F599C007F499B,
-    0x007F499E007F599D, 0x007FB9A0007F499F, 0x007FC9A2007FB9A1,
-    0x007FD5A400C579A3, 0x007FD5A6007FDDA5, 0x00C265A8007FCDA7,
-    0x007FFDAA007FF5A9, 0x007FE9AC00C551AB, 0x007FD9AE007FE9AD,
-    0x00C4F9B0007FD9AF, 0x007FA1B200CF3DB1, 0x007FA1B4007FA1B3,
-    0x007FA1B6007FA1B5, 0x007F61B8007FA1B7, 0x007F61BA007F61B9,
-    0x007F61BC007F61BB, 0x007F61BE007F61BD, 0x8000F404007F61BF,
-    0x8000370E8000F247, 0x80002AE680003063, 0x8000F33C8000F26A,
-    0x8000307B80003073, 0x8000308F80002AEA, 0x800030A38000309F,
-    0x800030BB800030AB, 0x80002AEE80002AF2, 0x800030C38000F24C,
-    0x800030DB800030D7, 0x800030FF800030EB, 0x800030FB800030F7,
-    0x8000311780003103, 0x800031338000311B, 0x8000314B8000313B,
-    0x8000F25B8000315F, 0x8000314F8000315B, 0x8000316F80003167,
-    0x8000F26080003183, 0x8000317F8000318B, 0x80002AF680003177,
-    0x8000319B80003193, 0x80003197800031A3, 0x800031BB80002AFA,
-    0x800031E78000F265, 0x8000320580003229, 0x80002B0280002B06,
-    0x80002CD180002AFE, 0x8000322D80002CD5, 0x800036E280003215,
-    0x8000324180002B6A, 0x0001A18080003255, 0x0001A1820001A181,
-    0x0001A1840001A183, 0x0001A1860001A185, 0x00BA1D880001A187,
-    0x0001ED8A0001E589, 0x0001FD8C0001E58B, 0x0001ED8E0001E58D,
-    0x0001999000BA718F, 0x0001A99200019991, 0x0001B9940001A993,
-    0x0001A9960001B995, 0x00B3B9980001A997, 0x0001919A00BC8199,
-    0x0003719C0001919B, 0x0003719E0003719D, 0x000391A00003719F,
-    0x00BAD9A2000391A1, 0x00038DA4000395A3, 0x0003BDA60003B5A5,
-    0x000389A800BE11A7, 0x000399AA000389A9, 0x000389AC000399AB,
-    0x0003F9AE000389AD, 0x000389B00003F9AF, 0x000399B2000389B1,
-    0x000389B4000399B3, 0x000385B600B7E5B5, 0x000385B80003BDB7,
-    0x000385BA00038DB9, 0x000385BC00039DBB, 0x000385BE00038DBD,
-    0x8000324D00037DBF, 0x80002B0A8000F27E, 0x8000325D80003269,
-    0x8000F28880003286, 0x8000327180003292, 0x8000329E8000329A,
-    0x800032A6800032A2, 0x8000F28D800032B2, 0x80002B12800032BA,
-    0x8000F292800032C6, 0x80002B16800032CA, 0x800032E2800032D6,
-    0x8000F297800032F2, 0x8000F29C8000F22E, 0x8000F2A180003312,
-    0x800033268000331E, 0x8000333280003316, 0x8000334A80003342,
-    0x8000335A80003356, 0x8000333E8000F2A6, 0x8000337680003372,
-    0x8000338680002B1A, 0x8000336A8000F2AB, 0x8000F2B0800033A6,
-    0x800033AA8000F2B5, 0x800033B6800033BE, 0x800033C280002B1E,
-    0x800033CE800033D2, 0x8000F1F2800033DA, 0x8000F2BA800033EA,
-    0x8000F2C4800033F6, 0x8000342A8000340A, 0x000000808000F2C9,
-    0x800034428000343E, 0x8000F2D38000F2CE, 0x8000346E8000345E,
-    0x8000347280002B22, 0x8000348A80002B26, 0x8000349A8000348E,
-    0x800034A680003033, 0x800034BE80002B2A, 0x800034DA8000F2D8,
-    0x8000F2428000F2DD, 0x8000F2E2800034F6, 0x8000F2EC8000F2E7,
-    0x80002B3280002B2E, 0x8000F2F18000F2F6, 0x8000F3008000F2FB,
-    0x8000353A80003536, 0x80002B3E80003546, 0x8000F30A80002B42,
-    0x80002B468000355A, 0x8000F3148000F30F, 0x800035CA8000F319,
-    0x800035D680002B4A, 0x8000F31E800035DA, 0x800035F28000F323,
-    0x8000F32880002B4E, 0x8000360A80003606, 0x8000F32D80002B52,
-    0x80002B5680003626, 0x8000362E80003636, 0x8000F3378000363E,
-    0x80002B5A80003656, 0x00BB1D8080003662, 0x0007798200076981,
-    0x0007698400077983, 0x0007998600076985, 0x00BB018800079D87,
-    0x0007AD8A00BBB189, 0x0007BD8C0007A58B, 0x0007AD8E0007A58D,
-    0x00BB699000B7158F, 0x0007D5920007DD91, 0x0007F19400B94593,
-    0x00BEC9960007F195, 0x0007CD980007F597, 0x0007D99A00B97199,
-    0x0007C99C0007D99B, 0x0007C59E0007C99D, 0x000745A00007BD9F,
-    0x000745A200074DA1, 0x000745A400075DA3, 0x000745A600074DA5,
-    0x000745A800077DA7, 0x000745AA00074DA9, 0x000745AC00075DAB,
-    0x000745AE00074DAD, 0x0002C9B00002B9AF, 0x0002D9B20002C9B1,
-    0x0002C9B40002D9B3, 0x0002F9B60002C9B5, 0x0002C9B80002F9B7,
-    0x00B9C5BA0002C9B9, 0x0002C5BC0002DDBB, 0x0002C5BE0002CDBD,
-    0x00B8F5940002BDBF, 0x0007C5960007CD95, 0x0007A9900007DD97,
-    0x0007D9920007A991, 0x0007C59C0007D993, 0x0007C59E0007CD9D,
-    0x0007E5980007DD9F, 0x0007E59A0007ED99, 0x0007CD840007DD9B,
-    0x0007DD860007D585, 0x0007D5800007D587, 0x00BD218200BE2981,
-    0x0007C98C0007D583, 0x0007D98E0007C98D, 0x0007ED880007D98F,
-    0x0007DD8A0007D589, 0x0006C1B400B88D8B, 0x0006C1B60006C1B5,
-    0x00BA95B00006C1B7, 0x0006A1B20006A5B1, 0x0006FDBC0006A1B3,
-    0x0006CDBE0006C5BD, 0x00B645B80006C5BF, 0x0006EDBA0006E5B9,
-    0x0006C5A40006E5BB, 0x0006C5A60006CDA5, 0x000725A00006DDA7,
-    0x000721A200B829A1, 0x0006C5AC000721A3, 0x0006C1AE0006C1AD,
-    0x0006E5A80006C1AF, 0x0006E5AA0006EDA9, 0x000000800006DDAB,
-    0x8000367A8000366A, 0x8000F34180002B5E, 0x80002B628000F346,
-    0x800036928000F34B, 0x8000369E8000F350, 0x8000F355800036AA,
-    0x800036C68000F35A, 0x800036CE8000F35F, 0x80002B668000F26F,
-    0x800036EA800036E6, 0x800036EE80002B6E, 0x8000F36480002F48,
-    0x8000F2748000F369, 0x800037128000F279, 0x80002B728000392A,
-    0x8000373A8000373E, 0x80002D4D80003742, 0x8000374A80003746,
-    0x800037568000374E, 0x8000375A8000F36E, 0x800037728000376E,
-    0x800037768000375E, 0x8000376A8000378A, 0x800037828000377E,
-    0x8000F37380003786, 0x8000F3788000F37D, 0x800037A680002B76,
-    0x800037B2800037AA, 0x800037B68000F391, 0x80002B7A8000F382,
-    0x8000F38780002B7E, 0x000145808000F38C, 0x0001458200014D81,
-    0x0001458400015D83, 0x0001458600014D85, 0x0001458800017D87,
-    0x0001458A00014D89, 0x0001418C0007258B, 0x0001418E0001418D,
-    0x0001C1900001418F, 0x0001C1920001C191, 0x0001C1940001C193,
-    0x0001C1960001C195, 0x00B15D980001C197, 0x0001CD9A0001C599,
-    0x0001DD9C0001C59B, 0x0001CD9E0001C59D, 0x000139A000BA459F,
-    0x000149A2000139A1, 0x000159A4000149A3, 0x000149A6000159A5,
-    0x000179A8000149A7, 0x000149AA000179A9, 0x000159AC000149AB,
-    0x000149AE000159AD, 0x000139B0000149AF, 0x000C01B2000139B1,
-    0x00B4C5B4000FCDB3, 0x000FD1B6000FD1B5, 0x000FF1B8000FD1B7,
-    0x000FF1BA000FF1B9, 0x000FEDBC00B8A9BB, 0x000FDDBE000FD5BD,
-    0x800037D6000FD5BF, 0x800037E2800037DE, 0x800037EA800037EE,
-    0x800037F2800037F6, 0x800037FE800037FA, 0x80002B8680003806,
-    0x800038128000380E, 0x8000382280002B8A, 0x800038328000F396,
-    0x80002B8E80003836, 0x80002AC280003842, 0x8000F3A08000F39B,
-    0x80002B9680002B92, 0x8000F3A58000386E, 0x800038CA800038C6,
-    0x800038EE800038E2, 0x8000F2018000F3AF, 0x800038F6800038FA,
-    0x8000F20680003902, 0x8000F3B48000390E, 0x8000395A8000F3B9,
-    0x8000396E80003966, 0x800039728000F3BE, 0x800039AA8000399E,
-    0x800039A6800039AE, 0x800039BE800039BA, 0x800039CA8000F3C3,
-    0x800039D280002B9A, 0x80002B9E8000F3C8, 0x80002FD880003A0A,
-    0x8000F3CD80003A22, 0x80002BA28000F3D7, 0x00012D8080002BA6,
-    0x00013D8200013581, 0x0085858400013583, 0x0001598600012985,
-    0x0001698800015987, 0x0001798A00016989, 0x0001698C0001798B,
-    0x0001598E0001698D, 0x00B619900001598F, 0x0085F99200B61991,
-    0x00013D9400012593, 0x00012D9600012595, 0x00011D9800012597,
-    0x0001ED9A0001E599, 0x0001FD9C0001E59B, 0x0001ED9E0001E59D,
-    0x000119A000B61D9F, 0x000129A2000119A1, 0x000139A4000129A3,
-    0x000129A6000139A5, 0x000119A8000129A7, 0x000169AA000119A9,
-    0x000179AC000169AB, 0x000169AE000179AD, 0x000119B0000169AF,
-    0x000129B2000119B1, 0x000139B4000129B3, 0x000129B6000139B5,
-    0x000119B8000129B7, 0x0003E9BA000119B9, 0x0003F9BC0003E9BB,
-    0x00B681BE0003F9BD, 0x80003A4A00B685BF, 0x80002BAA8000F3DC,
-    0x80003A6A8000F3E1, 0x80003A828000F3E6, 0x80003A9280002BAE,
-    0x80003AA680003A9E, 0x80002BB280003AAE, 0x80003ADE8000F3EB,
-    0x80002BB680003AEA, 0x80003AEE80002BBA, 0x80002BBE8000F3F0,
-    0x8000F3FA8000F3F5, 0x80002BC28000F3FF, 0x80003B3280003B2E,
-    0x80003B4280003B3A, 0x000000DF8000F409, 0x000000E1000000DE,
-    0x000000E3000000E0, 0x000000E5000000E2, 0x000000E7000000E4,
-    0x000000E9000000E6, 0x000000EB000000E8, 0x000000ED000000EA,
-    0x000000EF000000EC, 0x000000F1000000EE, 0x000000F3000000F0,
-    0x000000F5000000F2, 0x000000F7000000F4, 0x000000F9000000F6,
-    0x000000FB000000F8, 0x000000FD000000FA, 0x000000FF000000FC,
-    0x00011180000000FE, 0x0001118200011181, 0x0001318400011183,
-    0x0001318600013185, 0x0001118800013187, 0x00B4558A00011189,
-    0x00010D8C0001158B, 0x00017D8E0001758D, 0x00010D900001758F,
-    0x00011D9200011591, 0x00010D9400011593, 0x0001399600842D95,
-    0x000105980084BD97, 0x0001059A00010D99, 0x0084859C00011D9B,
-    0x0000009F0001019D, 0x000000A10000009E, 0x000000A3000000A0,
-    0x000000A5000000A2, 0x000000A7000000A4, 0x000000A9000000A6,
-    0x000000AB000000A8, 0x000000AD000000AA, 0x000000AF000000AC,
-    0x000000B1000000AE, 0x80002B82000000B0, 0x000000B5000000B2,
-    0x000000B7000000B4, 0x000000B9000000B6, 0x000000BB000000B8,
-    0x000000BD000000BA, 0x000000BF000000BC, 0x20002000000000BE,
-    0x200083CC200081CC, 0x200085CC200084CC, 0x200087CC200086CC,
-    0x80CC88CC200088CC, 0x200081CC88CC2000, 0x8ACC200082CD88CC,
-    0x93CC20008BCC2000, 0x200080CC93CC2000, 0x93CC200081CC93CC,
-    0x200094CC200082CD, 0x94CC200080CC94CC, 0x82CD94CC200081CC,
-    0xA8CC2000A7CC2000, 0x82CD2000B3CC2000, 0x8BD92000B9CE2000,
-    0x8CD920008CD92000, 0x20008DD9200091D9, 0x8ED9200091D98DD9,
-    0x200091D98ED92000, 0x91D98FD920008FD9, 0x90D9200090D92000,
-    0x200091D9200091D9, 0x92D92000B0D991D9, 0xE320009982E32000,
-    0x0021210021009A82, 0x2400230022003F21, 0x2800270026002500,
-    0x3031280029312800, 0x2800293131280029, 0x2933312800293231,
-    0x3128002934312800, 0x0029363128002935, 0x3831280029373128,
-    0x2800293931280029, 0x0029303228002932, 0x0029342800293328,
-    0x0029362800293528, 0x0029382800293728, 0x0029612800293928,
-    0x0029632800296228, 0x0029652800296428, 0x0029672800296628,
-    0x0029692800296828, 0x00296B2800296A28, 0x00296D2800296C28,
-    0x00296F2800296E28, 0x0029712800297028, 0x0029732800297228,
-    0x0029752800297428, 0x0029772800297628, 0x0029792800297828,
-    0x8084E12800297A28, 0x00298284E1280029, 0xE12800298384E128,
-    0x8684E12800298584, 0x00298784E1280029, 0xE12800298984E128,
-    0x8C84E12800298B84, 0x00298E84E1280029, 0xE12800298F84E128,
-    0x9184E12800299084, 0x00299284E1280029, 0xE428002980B8E428,
-    0x89B8E428002983B8, 0x00299DB9E4280029, 0xE42800298CBAE428,
-    0xA3BBE428002994BA, 0x002981BCE4280029, 0xE528002991BCE428,
-    0xAD85E5280029AB85, 0x0029B48AE5280029, 0xE5280029818DE528,
-    0x8D90E5280029948D, 0x0029BC91E5280029, 0xE52800299B9BE528,
-    0xA6ADE52800299F9C, 0x0029A597E6280029, 0xE6280029889CE628,
-    0xA89CE6280029899C, 0x0029AAA0E6280029, 0xE7280029B4B0E628,
-    0xB989E7280029AB81, 0x0029A39BE7280029, 0xE7280029BEA4E728,
-    0xADA5E72800299DA5, 0x0029AA87E8280029, 0xE8280029B387E828,
-    0x87B3E8280029A1B2, 0x00299187E9280029, 0xEB28002980B0EA28,
-    0xA48BEB2800299882, 0x0029BC9DEB280029, 0xEB28002988A7EB28,
-    0xAC82EC28002994B0, 0x00298495EC280029, 0x2984A0ECA498EC28,
-    0x849BEDA498EC2800, 0x0029909EEC280029, 0xEC280029BCA3EC28,
-    0xB4B9EC280029A8B0, 0x00298083ED280029, 0xED2800298C8CED28,
-    0x002A002900299895, 0x002E002D002C002B, 0x2F002E2E2E002E2E,
-    0x2E30002C30003000, 0x3000338481E23000, 0x2C31003100B982E7,
-    0x31003031002E3100, 0xA597E63031002E30, 0x3100889CE6303100,
-    0x00313100B982E730, 0x97E63131002E3131, 0x00889CE6313100A5,
-    0x323100B982E73131, 0xE63231002E323100, 0x889CE6323100A597,
-    0x3100B982E7323100, 0x3331002E33310033, 0x82E7333100A597E6,
-    0x2E343100343100B9, 0x3100A597E6343100, 0x00353100B982E734,
-    0x97E63531002E3531, 0x00B982E7353100A5, 0x31002E3631003631,
-    0xE7363100A597E636, 0x373100373100B982, 0x00A597E63731002E,
-    0x383100B982E73731, 0xE63831002E383100, 0xB982E7383100A597,
-    0x002E393100393100, 0x393100A597E63931, 0x8481E23100B982E7,
-    0x0030318481E23100, 0xE23100328481E231, 0x8481E23100338481,
-    0x00358481E2310034, 0xE23100368481E231, 0x8481E23100378481,
-    0x00398481E2310038, 0x9CE63100A597E631, 0x3200B982E7310088,
-    0x32002E32002C3200, 0x3032002E30320030, 0x82E7303200A597E6,
-    0xE6313200313200B9, 0xB982E7313200A597, 0x97E6323200323200,
-    0x00B982E7323200A5, 0xA597E63332003332, 0x3200B982E7333200,
-    0x00A597E634320034, 0x353200B982E73432, 0x3200A597E6353200,
-    0x00A597E636320036, 0xA597E63732003732, 0x97E6383200383200,
-    0xE6393200393200A5, 0x338481E23200A597, 0x3200358481E23200,
-    0x889CE63200A597E6, 0x003300B982E73200, 0x3033002E33002C33,
-    0x3300A597E6303300, 0x00A597E631330031, 0x3433003333003233,
-    0x3300363300353300, 0x0039330038330037, 0xE23300348481E233,
-    0x8481E23300358481, 0x3300A597E6330038, 0xB982E73300889CE6,
-    0x2E34002C34003400, 0x3400313400303400, 0x0034340033340032,
-    0x3734003634003534, 0x3400393400383400, 0x97E63400358481E2,
-    0x3400889CE63400A5, 0x2C35003500B982E7, 0x35003035002E3500,
-    0x81E23500368481E2, 0x00A597E635003884, 0x82E73500889CE635,
-    0x36002C36003600B9, 0x3600A597E636002E, 0xB982E73600889CE6,
-    0x2E37002C37003700, 0x3700388481E23700, 0x889CE63700A597E6,
-    0x003800B982E73700, 0xE638002E38002C38, 0x00889CE63800A597,
-    0x39003900B982E738, 0x97E639002E39002C, 0x3900889CE63900A5,
-    0x3A3A003A00B982E7, 0x003D003C003B003D, 0x3E003D3D3D003D3D,
-    0x3F3F00213F003F00, 0x5D005C005B004000, 0x610060005F005E00,
-    0x2F61002E6D2E6100, 0x756100732F610063, 0x88E26100BECA6100,
-    0x7261620062006D95, 0x2F63006300716200, 0x616300752F63006F,
-    0x006463006363006C, 0x6300326D63006D63, 0x63002E6F6300336D,
-    0x006400676B9588E2, 0x6A64006264006164, 0x64006D64006C6400,
-    0x6400336D6400326D, 0x006500BEC564007A, 0x6600766500677265,
-    0x0066660078616600, 0x006C666600696666, 0x6D66006C66006966,
-    0x67006C6167006700, 0x7067007A68670062, 0x6800680079670061,
-    0x0070680067680061, 0x6800766800617068, 0x690069690069007A,
-    0x6E69006A69006969, 0x6900766900756900, 0x616B006B006A0078,
-    0x6C61636B00626B00, 0x007A686B00676B00, 0x6D6B006C6B006B6B,
-    0x336D6B00326D6B00, 0x00746B0061706B00, 0xCF6B00776B00766B,
-    0x6C006A6C006C0089, 0x676F6C006E6C006D, 0x00786C0064746C00,
-    0x326D006D00B7C26C, 0x6D00616D00336D00, 0x00646D00636D0062,
-    0x6D007A686D00676D, 0x6D6D006C6D006C69, 0x336D6D00326D6D00,
-    0x61706D006C6F6D00, 0x6D00736D00726D00, 0x89CF6D00776D0076,
-    0x6D00739588E26D00, 0x006E0032739588E2, 0x6A6E00666E00616E,
-    0x6E006F6E006D6E00, 0x00776E00766E0073, 0x70007000766F006F,
-    0x70006170002E6D2E, 0x0068700066700063, 0x00767070006D7070,
-    0x7470007370007270, 0x0077700076700065, 0x0064617200720071,
-    0x00739588E2646172, 0x32739588E2646172, 0x6473007300737200,
-    0x73007273006D7300, 0x0076730074730073, 0x6874006C65740074,
-    0x760075006D74007A, 0x0069697600697600, 0x88E2760069696976,
-    0x0062770077006D95, 0x0078007A77006377, 0x7900696978006978,
-    0x7D007C007B007A00, 0xC200A3C200A2C200, 0x00ACC200A6C200A5,
-    0x0066B0C20063B0C2, 0xA0C3009FC300B7C2, 0xC300A2C300A1C300,
-    0x00A5C300A4C300A3, 0xA8C300A7C300A6C3, 0xC300AAC300A9C300,
-    0x00ADC300ACC300AB, 0xB0C300AFC300AEC3, 0xC300B2C300B1C300,
-    0x00B5C300B4C300B3, 0xB9C300B8C300B6C3, 0xC300BBC300BAC300,
-    0x00BEC300BDC300BC, 0x83C40081C400BFC3, 0xC40087C40085C400,
-    0x008DC4008BC40089, 0x93C40091C4008FC4, 0xC40097C40095C400,
-    0x009DC4009BC40099, 0xA3C400A1C4009FC4, 0xC400A7C400A5C400,
-    0x00ADC400ABC400A9, 0xB1C400B0C400AFC4, 0xC400B7C400B5C400,
-    0x00BEC400BCC400BA, 0x86C50084C50082C5, 0xC5008BC50088C500,
-    0x0091C5008FC5008D, 0x97C50095C50093C5, 0xC5009BC50099C500,
-    0x00A1C5009FC5009D, 0xA7C500A5C500A3C5, 0xC500ABC500A9C500,
-    0x00B1C500AFC500AD, 0xB7C500B5C500B3C5, 0xC500BCC500BAC500,
-    0x0083C60080C600BE, 0x8CC60088C60085C6, 0xC60095C60092C600,
-    0x009EC6009AC60099, 0xA5C600A3C600A1C6, 0xC600ABC600A8C600,
-    0x00B4C600B0C600AD, 0xBDC600B9C600B6C6, 0xC70080C700BFC600,
-    0x008EC70082C70081, 0x94C70092C70090C7, 0xC70098C70096C700,
-    0x009DC7009CC7009A, 0xA3C700A1C7009FC7, 0xC700A7C700A5C700,
-    0x00ADC700ABC700A9, 0xB5C700B0C700AFC7, 0xC700BBC700B9C700,
-    0x0081C800BFC700BD, 0x87C80085C80083C8, 0xC8008BC80089C800,
-    0x0091C8008FC8008D, 0x97C80095C80093C8, 0xC8009BC80099C800,
-    0x00A3C8009FC8009D, 0xA9C800A7C800A5C8, 0xC800ADC800ABC800,
-    0x00B3C800B1C800AF, 0xBFC800BCC800B7C8, 0xC90082C90080C900,
-    0x008BC90089C90087, 0x90C9008FC9008DC9, 0xC90092C90091C900,
-    0x0095C90094C90093, 0x98C90097C90096C9, 0xC9009BC90099C900,
-    0x009FC9009EC9009C, 0xA2C900A1C900A0C9, 0xC900A4C900A3C900,
-    0x00A7C900A6C900A5, 0xAAC900A9C900A8C9, 0xC900ACC900ABC900,
-    0x00AFC900AEC900AD, 0xB2C900B1C900B0C9, 0xC900B4C900B3C900,
-    0x00B7C900B6C900B5, 0xBAC900B9C900B8C9, 0xC900BDC900BBC900,
-    0x0081CA0080CA00BE, 0x84CA0083CA0082CA, 0xCA0088CA0087CA00,
-    0x008BCA008ACA0089, 0x8ECA008DCA008CCA, 0xCA0090CA008FCA00,
-    0x0095CA0092CA0091, 0x9BCA0099CA0098CA, 0xCA009DCA009CCA00,
-    0x00A1CA009FCA009E, 0xA4CA00A3CA00A2CA, 0xCA00A6CA00A5CA00,
-    0x00A9CA00A8CA00A7, 0xB9CA00ABCA00AACA, 0x0090CB006EBCCA00,
-    0x81CC0080CC0091CB, 0x93CC0081CC88CC00, 0xCD00B3CD00B1CD00,
-    0x00BCCD00BBCD00B7, 0xACCE0090CE00BDCD, 0xCE00AECE00ADCE00,
-    0x00B1CE00B0CE00AF, 0xB4CE00B3CE00B2CE, 0xCE00B6CE00B5CE00,
-    0x00B9CE00B8CE00B7, 0xBCCE00BBCE00BACE, 0x66BCCE0061BCCE00,
-    0x6CBCCE0067BCCE00, 0x73BCCE006DBCCE00, 0x77BCCE0076BCCE00,
-    0xCE00BECE00BDCE00, 0x0081CF0080CF00BF, 0x85CF0084CF0083CF,
-    0xCF0087CF0086CF00, 0x008ACF0089CF0088, 0x8DCF008CCF008BCF,
-    0xCF0097CF008ECF00, 0x009DCF009BCF0099, 0xA3CF00A1CF009FCF,
-    0xCF00A7CF00A5CF00, 0x00ADCF00ABCF00A9, 0xB8CF00B3CF00AFCF,
-    0xD000B0D000BBCF00, 0x00B3D000B2D000B1, 0xB6D000B5D000B4D0,
-    0xD000B8D000B7D000, 0x00BBD000BAD000B9, 0xBED000BDD000BCD0,
-    0xD10080D100BFD000, 0x0083D10082D10081, 0x86D10085D10084D1,
-    0xD10088D10087D100, 0x008BD1008AD10089, 0x8ED1008DD1008CD1,
-    0xD10090D1008FD100, 0x0093D10092D10091, 0x96D10095D10094D1,
-    0xD10098D10097D100, 0x009BD1009AD10099, 0x9ED1009DD1009CD1,
-    0xD100A1D1009FD100, 0x00A7D100A5D100A3, 0xADD100ABD100A9D1,
-    0xD100B1D100AFD100, 0x00B7D100B5D100B3, 0xBDD100BBD100B9D1,
-    0xD20081D200BFD100, 0x008FD2008DD2008B, 0x95D20093D20091D2,
-    0xD20099D20097D200, 0x009FD2009DD2009B, 0xA5D200A3D200A1D2,
-    0xD200A9D200A7D200, 0x00AFD200ADD200AB, 0xB5D200B3D200B1D2,
-    0xD200B9D200B7D200, 0x00BFD200BDD200BB, 0x86D30084D30082D3,
-    0xD3008AD30088D300, 0x008FD3008ED3008C, 0x95D30093D30091D3,
-    0xD30099D30097D300, 0x009FD3009DD3009B, 0xA5D300A3D300A1D3,
-    0xD300A9D300A7D300, 0x00AFD300ADD300AB, 0xB5D300B3D300B1D3,
-    0xD300B9D300B7D300, 0x00BFD300BDD300BB, 0x85D40083D40081D4,
-    0xD40089D40087D400, 0x008FD4008DD4008B, 0x95D40093D40091D4,
-    0xD40099D40097D400, 0x009FD4009DD4009B, 0xA5D400A3D400A1D4,
-    0xD400A9D400A7D400, 0x00AFD400ADD400AB, 0xA3D500A2D500A1D5,
-    0xD500A5D500A4D500, 0xD500A6D50082D6A5, 0x00A9D500A8D500A7,
-    0xACD500ABD500AAD5, 0xD500AED500ADD500, 0x00B1D500B0D500AF,
-    0xB4D500B3D500B2D5, 0xB4D500A5D5B4D500, 0x00ADD5B4D500ABD5,
-    0x00B5D500B6D5B4D5, 0xB8D500B7D500B6D5, 0xD500BAD500B9D500,
-    0x00BDD500BCD500BB, 0x00B6D5BED500BED5, 0x81D60080D600BFD5,
-    0xD60083D60082D600, 0x0086D60085D60084, 0x00B7D690D70090D7,
-    0xD690D700B8D690D7, 0xD7009CD790D700BC, 0xD700BCD691D70091,
-    0xD70092D700BFD691, 0xD70093D700BCD692, 0xD70094D700BCD693,
-    0xB9D695D700BCD694, 0x96D700BCD695D700, 0x00BCD698D700BCD6,
-    0xD699D700B4D699D7, 0xD700BCD69AD700BC, 0xD700BCD69BD7009B,
-    0xD7009CD700BFD69B, 0xD7009DD700BCD69C, 0xBCD6A0D700BCD69E,
-    0xA2D700BCD6A1D700, 0xA4D700BCD6A3D700, 0x00BFD6A4D700BCD6,
-    0xD6A7D700BCD6A6D7, 0xD6A8D700A8D700BC, 0xD700BCD6A9D700BC,
-    0xA9D70081D7BCD6A9, 0xD7A9D70082D7BCD6, 0xD70082D7A9D70081,
-    0xD700BCD6AAD700AA, 0xD800A1D800B7D6B2, 0x00A4D800A3D800A2,
-    0xA6D800A6D800A5D8, 0x00ACD8A6D800A7D8, 0xD8A6D800ADD8A6D8,
-    0xD800B1D8A6D800AE, 0x85D9A6D800B2D8A6, 0xA6D80086D9A6D800,
-    0x0088D9A6D80087D9, 0xD9A6D80089D9A6D8, 0xD80086DBA6D8008A,
-    0x88DBA6D80087DBA6, 0xA6D80090DBA6D800, 0xA7D800A7D80095DB,
-    0xD800B1D8A8D883D9, 0x0087D984D984D9A7, 0xD9A7D8008BD9A7D8,
-    0xD8A8D800A8D800B4, 0xD800ADD8A8D800AC, 0xA8D8008AD9ADD8A8,
-    0xD9AED8A8D800AED8, 0xD800B1D8A8D8008A, 0x85D9A8D800B2D8A8,
-    0xA8D80086D9A8D800, 0x0089D9A8D80087D9, 0x00A9D8008AD9A8D8,
-    0x00ACD8AAD800AAD8, 0xD80085D9ACD8AAD8, 0xAAD80089D9ACD8AA,
-    0xD8AAD8008AD9ACD8, 0xACD8ADD8AAD800AD, 0x0085D9ADD8AAD800,
-    0xD8AAD800AED8AAD8, 0xAED8AAD80085D9AE, 0xD9AED8AAD80089D9,
-    0xD800B1D8AAD8008A, 0x85D9AAD800B2D8AA, 0x00ACD885D9AAD800,
-    0xD800ADD885D9AAD8, 0xAAD800AED885D9AA, 0xD9AAD80089D985D9,
-    0x86D9AAD8008AD985, 0xAAD80087D9AAD800, 0x008AD9AAD80089D9,
-    0x00ACD8ABD800ABD8, 0xD8ABD800B1D8ABD8, 0xD80085D9ABD800B2,
-    0x87D9ABD80086D9AB, 0xABD80089D9ABD800, 0xACD800ACD8008AD9,
-    0xD9ADD8ACD800ADD8, 0x8AD9ADD8ACD80089, 0xACD82084D9ACD800,
-    0x87D984D9A7D884D9, 0xACD80085D9ACD800, 0xD9ACD800ADD885D9,
-    0x85D9ACD80089D985, 0x0089D9ACD8008AD9, 0x00ADD8008AD9ACD8,
-    0xD8ADD800ACD8ADD8, 0x85D9ADD8008AD9AC, 0x0089D985D9ADD800,
-    0xD8008AD985D9ADD8, 0x8AD9ADD80089D9AD, 0xACD8AED800AED800,
-    0xAED800ADD8AED800, 0x0089D9AED80085D9, 0x00AFD8008AD9AED8,
-    0x00B0D9B0D800B0D8, 0xD9B3D8B1D800B1D8, 0xB0D9B1D80084D988,
-    0xD9A7D88CDBB1D800, 0x00B3D800B2D80084, 0xD8B3D800ACD8B3D8,
-    0xACD8B3D800ADD8AC, 0x00ADD8B3D80089D9, 0xD800ACD8ADD8B3D8,
-    0xAED8B3D800AED8B3, 0xD9AED8B3D80089D9, 0xD800B1D8B3D8008A,
-    0x85D9B3D80085D9B3, 0xD885D9B3D800ACD8, 0x85D985D9B3D800AD,
-    0xB3D80087D9B3D800, 0x008AD9B3D80089D9, 0x00ACD8B4D800B4D8,
-    0xD8008AD9ACD8B4D8, 0xADD8B4D800ADD8B4, 0xD9ADD8B4D80085D9,
-    0xD800AED8B4D8008A, 0x85D9B4D800B1D8B4, 0x00AED885D9B4D800,
-    0xD80085D985D9B4D8, 0x89D9B4D80087D9B4, 0xB5D8008AD9B4D800,
-    0xB5D800ADD8B5D800, 0xD8B5D800ADD8ADD8, 0xAED8B5D8008AD9AD,
-    0xB5D800B1D8B5D800, 0xD80085D9B9D884D9, 0xB5D80089D984D9B5,
-    0xD9A7D82089D984D9, 0xB9D82087D984D984, 0xD92087D98AD984D9,
-    0x0085D984D9B3D888, 0xD80092DB84D9B5D8, 0x85D9B5D80085D9B5,
-    0x0089D9B5D80085D9, 0x00B6D8008AD9B5D8, 0xD8B6D800ACD8B6D8,
-    0x89D9ADD8B6D800AD, 0x008AD9ADD8B6D800, 0xD8B6D800AED8B6D8,
-    0xB1D8B6D80085D9AE, 0xB6D80085D9B6D800, 0x008AD9B6D80089D9,
-    0x00ADD8B7D800B7D8, 0xD9B7D80085D9B7D8, 0x85D9B7D800ADD885,
-    0xD985D9B7D80085D9, 0xD80089D9B7D8008A, 0xD800B8D8008AD9B7,
-    0xD800B9D80085D9B8, 0xACD8B9D800ACD8B9, 0xD984D9B9D80085D9,
-    0x85D9B9D80087D98A, 0x0085D985D9B9D800, 0xD80089D985D9B9D8,
-    0xB9D8008AD985D9B9, 0x008AD9B9D80089D9, 0x00ACD8BAD800BAD8,
-    0xD9BAD80085D9BAD8, 0x85D9BAD80085D985, 0xD985D9BAD80089D9,
-    0xD80089D9BAD8008A, 0x8BD980D9008AD9BA, 0x80D9008ED980D900,
-    0xD980D90091D98ED9, 0x91D98FD980D9008F, 0x80D90090D980D900,
-    0xD980D90091D990D9, 0xD90092D980D90091, 0xD900ACD881D90081,
-    0xAED881D900ADD881, 0x0085D9AED881D900, 0xD981D90085D981D9,
-    0x89D981D9008AD985, 0x82D9008AD981D900, 0x82D900ADD882D900,
-    0xD982D90092DB84D9, 0xADD885D982D90085, 0x0085D985D982D900,
-    0xD9008AD985D982D9, 0x8AD982D90089D982, 0xA7D883D90083D900,
-    0x83D900ACD883D900, 0x00AED883D900ADD8, 0xD983D90084D983D9,
-    0x85D985D983D90085, 0x008AD985D983D900, 0xD983D90089D983D9,
-    0xD884D90084D9008A, 0xD900A3D884D900A2, 0xA7D884D900A5D884,
-    0x84D900ACD884D900, 0xD884D900ACD8ACD8, 0xACD884D90085D9AC,
-    0x00ADD884D9008AD9, 0xD90085D9ADD884D9, 0x84D90089D9ADD884,
-    0xD884D9008AD9ADD8, 0x85D9AED884D900AE, 0x84D90085D984D900,
-    0xD984D900ADD885D9, 0x87D984D9008AD985, 0x84D90089D984D900,
-    0x85D90085D9008AD9, 0x00ACD885D900A7D8, 0xD900ADD8ACD885D9,
-    0x85D900AED8ACD885, 0xD885D90085D9ACD8, 0xADD885D9008AD9AC,
-    0x00ACD8ADD885D900, 0xD90085D9ADD885D9, 0x00AFD885D9ADD885,
-    0xD9008AD9ADD885D9, 0xAED885D900AED885, 0xD9AED885D900ACD8,
-    0x8AD9AED885D90085, 0x85D90085D985D900, 0xD985D9008AD985D9,
-    0xD9008AD985D90089, 0xD900ACD886D90086, 0x86D900ADD8ACD886,
-    0xD886D90085D9ACD8, 0xACD886D90089D9AC, 0x00ADD886D9008AD9,
-    0xD90085D9ADD886D9, 0x86D90089D9ADD886, 0xD886D9008AD9ADD8,
-    0xD900B1D886D900AE, 0x85D986D900B2D886, 0x0089D985D986D900,
-    0xD9008AD985D986D9, 0x87D986D90086D986, 0x86D90089D986D900,
-    0x87D90087D9008AD9, 0x0085D987D900ACD8, 0xD900ACD885D987D9,
-    0x87D90085D985D987, 0x008AD987D90089D9, 0x0088D900B0D987D9,
-    0x85D984D9B3D888D9, 0x89D900B4D988D900, 0x8AD900B0D989D900,
-    0x8AD900ACD88AD900, 0xD88AD9008AD9ACD8, 0x8AD9ADD88AD900AD,
-    0x8AD900AED88AD900, 0x00B2D88AD900B1D8, 0xD98AD90085D98AD9,
-    0x85D98AD90085D985, 0x0086D98AD9008AD9, 0xD98AD90087D98AD9,
-    0xD9008AD98AD90089, 0xD900AED900B4D98A, 0x00B9D900B1D900AF,
-    0xBED900BBD900BAD9, 0xDA0080DA00BFD900, 0x0086DA0084DA0083,
-    0x8CDA0088DA0087DA, 0xDA008EDA008DDA00, 0x00A1DA0098DA0091,
-    0xA9DA00A6DA00A4DA, 0xDA00AFDA00ADDA00, 0x00BADA00B3DA00B1,
-    0x80DB00BEDA00BBDA, 0xDB0082DB0081DB00, 0x0087DB0086DB0085,
-    0x0088DB00B4D987DB, 0x8CDB008BDB0089DB, 0xDB0092DB0090DB00,
-    0xBCA4E095A4E00093, 0x00BCA4E096A4E000, 0xE000BCA4E097A4E0,
-    0xA4E000BCA4E09CA4, 0xA2A4E000BCA4E0A1, 0x00A9A4E000BCA4E0,
-    0xE000BCA4E0ABA4E0, 0xA4E000BCA4E0AFA4, 0xA6E000B4A4E000B1,
-    0xA2A6E000BCA6E0A1, 0xE0AFA6E000BCA6E0, 0xE0008BA7E000BCA6,
-    0xA8E096A8E0008CA7, 0xBCA8E097A8E000BC, 0x00BCA8E09CA8E000,
-    0xE000BCA8E0ABA8E0, 0xA8E000BCA8E0B2A8, 0xA1ACE000BCA8E0B8,
-    0xE0A2ACE000BCACE0, 0xE00088ADE000BCAC, 0xE0008CADE0008BAD,
-    0xE0008AAFE00094AE, 0xE0008CAFE0008BAF, 0xE00080B3E00088B1,
-    0xE00088B3E00087B3, 0xE0008BB3E0008AB3, 0xE0008BB5E0008AB5,
-    0xE0009AB7E0008CB5, 0xE0009DB7E0009CB7, 0xB8E08DB9E0009EB7,
-    0x99BAE0ABBAE000B2, 0x00A1BAE0ABBAE000, 0xE000B2BAE08DBBE0,
-    0xBEE080BDE0008BBC, 0xB7BEE082BDE000B5, 0x00B7BEE08CBDE000,
-    0xE000B7BEE091BDE0, 0xBDE000B7BEE096BD, 0xB1BDE000B7BEE09B,
-    0xE0B1BDE000B2BDE0, 0xBEE0B1BDE000B4BD, 0xB5BEE090BEE00080,
-    0x00B7BEE092BEE000, 0xE000B7BEE09CBEE0, 0xBEE000B7BEE0A1BE,
-    0xABBEE000B7BEE0A6, 0xE0B2BEE000B7BEE0, 0xBEE00080BEE0B1BD,
-    0xB3BEE00080BEE0B2, 0xE00080BEE0B1BDE0, 0x80E10080BEE0B3BE,
-    0x83E1009083E100A6, 0x83E1009283E10091, 0x83E1009483E10093,
-    0x83E1009683E10095, 0x83E1009883E10097, 0x83E1009A83E10099,
-    0x83E1009C83E1009B, 0x83E1009E83E1009D, 0x83E100A083E1009F,
-    0x83E100A283E100A1, 0x83E100A483E100A3, 0x83E100A683E100A5,
-    0x83E100A883E100A7, 0x83E100AA83E100A9, 0x83E100AC83E100AB,
-    0x83E100AE83E100AD, 0x83E100B083E100AF, 0x83E100B283E100B1,
-    0x83E100B483E100B3, 0x83E100B683E100B5, 0x83E100B883E100B7,
-    0x83E100BA83E100B9, 0x83E100BE83E100BD, 0x84E1008084E100BF,
-    0x84E1008284E10081, 0x84E1008484E10083, 0x84E1008684E10085,
-    0x84E1008884E10087, 0x84E1008A84E10089, 0x84E1008C84E1008B,
-    0x84E1008E84E1008D, 0x84E1009084E1008F, 0x84E1009284E10091,
-    0x84E1009584E10094, 0x84E1009C84E1009A, 0x84E1009E84E1009D,
-    0x84E100A184E100A0, 0x84E100A384E100A2, 0x84E100A984E100A7,
-    0x84E100AC84E100AB, 0x84E100AE84E100AD, 0x84E100B284E100AF,
-    0x85E1008085E100B6, 0x85E1008C85E10087, 0x85E1009885E10097,
-    0x85E100A085E10099, 0x85E100A285E100A1, 0x85E100A485E100A3,
-    0x85E100A685E100A5, 0x85E100A885E100A7, 0x85E100AA85E100A9,
-    0x85E100AC85E100AB, 0x85E100AE85E100AD, 0x85E100B085E100AF,
-    0x85E100B285E100B1, 0x85E100B485E100B3, 0x86E1008486E100B5,
-    0x86E1008886E10085, 0x86E1009286E10091, 0x86E1009E86E10094,
-    0x86E100AA86E100A1, 0x86E100AD86E100AC, 0x86E100B186E100B0,
-    0x86E100B386E100B2, 0x87E100B586E100B4, 0x87E1008887E10087,
-    0x87E1008E87E1008C, 0x87E1009787E10093, 0x87E1009D87E10099,
-    0x87E100B187E1009F, 0x8EE100A08EE100B2, 0x8EE100A28EE100A1,
-    0x8EE100A48EE100A3, 0x8EE100A68EE100A5, 0x8EE100A88EE100A7,
-    0x8EE100AA8EE100A9, 0x8EE100AC8EE100AB, 0x8EE100AE8EE100AD,
-    0x8EE100B08EE100AF, 0x8EE100B28EE100B1, 0x8EE100B48EE100B3,
-    0x8EE100B68EE100B5, 0x8EE100B88EE100B7, 0x8EE100BA8EE100B9,
-    0x8EE100BC8EE100BB, 0x8EE100BE8EE100BD, 0x8FE100808FE100BF,
-    0x8FE100828FE10081, 0x8FE100848FE10083, 0x8FE100868FE10085,
-    0x8FE100888FE10087, 0x8FE1008A8FE10089, 0x8FE1008C8FE1008B,
-    0x8FE1008E8FE1008D, 0x8FE100908FE1008F, 0x8FE100928FE10091,
-    0x8FE100948FE10093, 0x8FE100968FE10095, 0x8FE100988FE10097,
-    0x8FE1009A8FE10099, 0x8FE1009C8FE1009B, 0x8FE1009E8FE1009D,
-    0x8FE100A08FE1009F, 0x8FE100A28FE100A1, 0x8FE100A48FE100A3,
-    0x8FE100A68FE100A5, 0x8FE100A88FE100A7, 0x8FE100AA8FE100A9,
-    0x8FE100AC8FE100AB, 0x8FE100AE8FE100AD, 0x8FE100B08FE100AF,
-    0x8FE100B28FE100B1, 0x8FE100B48FE100B3, 0xACE10086ACE100B5,
-    0xACE1008AACE10088, 0xACE1008EACE1008C, 0xACE100BBACE10092,
-    0xADE10080ADE100BD, 0xB4E10083ADE10081, 0xB4E10096B4E10082,
-    0xB4E1009CB4E10097, 0xB5E100A5B4E1009D, 0xB5E100BBB5E100B9,
-    0xB6E10085B6E100BD, 0xB8E10091B6E1008E, 0xB8E10083B8E10081,
-    0xB8E10087B8E10085, 0xB8E1008BB8E10089, 0xB8E1008FB8E1008D,
-    0xB8E10093B8E10091, 0xB8E10097B8E10095, 0xB8E1009BB8E10099,
-    0xB8E1009FB8E1009D, 0xB8E100A3B8E100A1, 0xB8E100A7B8E100A5,
-    0xB8E100ABB8E100A9, 0xB8E100AFB8E100AD, 0xB8E100B3B8E100B1,
-    0xB8E100B7B8E100B5, 0xB8E100BBB8E100B9, 0xB9E100BFB8E100BD,
-    0xB9E10083B9E10081, 0xB9E10087B9E10085, 0xB9E1008BB9E10089,
-    0xB9E1008FB9E1008D, 0xB9E10093B9E10091, 0xB9E10097B9E10095,
-    0xB9E1009BB9E10099, 0xB9E1009FB9E1009D, 0xB9E100A3B9E100A1,
-    0xB9E100A7B9E100A5, 0xB9E100ABB9E100A9, 0xB9E100AFB9E100AD,
-    0xB9E100B3B9E100B1, 0xB9E100B7B9E100B5, 0xB9E100BBB9E100B9,
-    0xBAE100BFB9E100BD, 0xBAE10083BAE10081, 0xBAE10087BAE10085,
-    0xBAE1008BBAE10089, 0xBAE1008FBAE1008D, 0xBAE10093BAE10091,
-    0xBAE10096BAE10095, 0xBAE10098BAE10097, 0xBAE100A1BAE10099,
-    0xBAE100A5BAE100A3, 0xBAE100A9BAE100A7, 0xBAE100ADBAE100AB,
-    0xBAE100B1BAE100AF, 0xBAE100B5BAE100B3, 0xBAE100B9BAE100B7,
-    0xBAE100BDBAE100BB, 0xBBE10081BBE100BF, 0xBBE10085BBE10083,
-    0xBBE10089BBE10087, 0xBBE1008DBBE1008B, 0xBBE10091BBE1008F,
-    0xBBE10095BBE10093, 0xBBE10099BBE10097, 0xBBE1009DBBE1009B,
-    0xBBE100A1BBE1009F, 0xBBE100A5BBE100A3, 0xBBE100A9BBE100A7,
-    0xBBE100ADBBE100AB, 0xBBE100B1BBE100AF, 0xBBE100B5BBE100B3,
-    0xBBE100B9BBE100B7, 0xBBE100BDBBE100BB, 0xBCE10080BCE100BF,
-    0xBCE10082BCE10081, 0xBCE10084BCE10083, 0xBCE10086BCE10085,
-    0xBCE10090BCE10087, 0xBCE10092BCE10091, 0xBCE10094BCE10093,
-    0xBCE100A0BCE10095, 0xBCE100A2BCE100A1, 0xBCE100A4BCE100A3,
-    0xBCE100A6BCE100A5, 0xBCE100B0BCE100A7, 0xBCE100B2BCE100B1,
-    0xBCE100B4BCE100B3, 0xBCE100B6BCE100B5, 0xBDE10080BDE100B7,
-    0xBDE10082BDE10081, 0xBDE10084BDE10083, 0xBDE10090BDE10085,
-    0xBDE10092BDE10091, 0xBDE10094BDE10093, 0xBDE10096BDE10095,
-    0xBDE100A0BDE10097, 0xBDE100A2BDE100A1, 0xBDE100A4BDE100A3,
-    0xBDE100A6BDE100A5, 0xBDE100B0BDE100A7, 0xBDE100B4BDE100B2,
-    0xBDE100B8BDE100B6, 0xBEE100BCBDE100BA, 0xBEE10081BEE10080,
-    0xBEE10083BEE10082, 0xBEE10085BEE10084, 0xBEE10087BEE10086,
-    0xBEE10091BEE10090, 0xBEE10093BEE10092, 0xBEE10095BEE10094,
-    0xBEE10097BEE10096, 0xBEE100A1BEE100A0, 0xBEE100A3BEE100A2,
-    0xBEE100A5BEE100A4, 0xBEE100A7BEE100A6, 0xBEE100B1BEE100B0,
-    0xBEE100B3BEE100B2, 0xBEE100B6BEE100B4, 0xBFE10082BFE100B7,
-    0xBFE10084BFE10083, 0xBFE10087BFE10086, 0xBFE10091BFE10090,
-    0xBFE10096BFE10092, 0xBFE100A0BFE10097, 0xBFE100A2BFE100A1,
-    0xBFE100A5BFE100A4, 0xBFE100A7BFE100A6, 0xBFE100B3BFE100B2,
-    0xBFE100B6BFE100B4, 0x80E2009080E200B7, 0x80E2009480E20093,
-    0xB280E200B280E2B2, 0xE200B280E2B280E2, 0xB280E2B280E2B280,
-    0xE2B580E200B280E2, 0x80E2B580E200B580, 0xA982E200B580E2B5,
-    0x8486E2008E85E200, 0x9186E2009086E200, 0x9386E2009286E200,
-    0x9B86E2009A86E200, 0x8D87E200AE86E200, 0x8F87E2008E87E200,
-    0x8488E2008288E200, 0x8988E2008788E200, 0x9188E2008C88E200,
-    0xA488E2009288E200, 0xAB88E200A688E200, 0xE2AB88E200AB88E2,
-    0x88E200AB88E2AB88, 0xE2AB88E2AB88E2AB, 0x88E2AE88E200AB88,
-    0xAE88E2AE88E200AE, 0x008189E200AE88E2, 0x008789E2008489E2,
-    0x00A089E2008989E2, 0x00AD89E200A289E2, 0x00AF89E200AE89E2,
-    0x00B189E200B089E2, 0x00B589E200B489E2, 0x00B989E200B889E2,
-    0x00818AE200808AE2, 0x00858AE200848AE2, 0x00898AE200888AE2,
-    0x00AD8AE200AC8AE2, 0x00AF8AE200AE8AE2, 0x00A18BE200A08BE2,
-    0x00A38BE200A28BE2, 0x00AB8BE200AA8BE2, 0x00AD8BE200AC8BE2,
-    0x00A096E2008294E2, 0x0085A6E2008B97E2, 0xCC9DABE20086A6E2,
-    0xB0E200B0B0E200B8, 0xB0E200B2B0E200B1, 0xB0E200B4B0E200B3,
-    0xB0E200B6B0E200B5, 0xB0E200B8B0E200B7, 0xB0E200BAB0E200B9,
-    0xB0E200BCB0E200BB, 0xB0E200BEB0E200BD, 0xB1E20080B1E200BF,
-    0xB1E20082B1E20081, 0xB1E20084B1E20083, 0xB1E20086B1E20085,
-    0xB1E20088B1E20087, 0xB1E2008AB1E20089, 0xB1E2008CB1E2008B,
-    0xB1E2008EB1E2008D, 0xB1E20090B1E2008F, 0xB1E20092B1E20091,
-    0xB1E20094B1E20093, 0xB1E20096B1E20095, 0xB1E20098B1E20097,
-    0xB1E2009AB1E20099, 0xB1E2009CB1E2009B, 0xB1E2009EB1E2009D,
-    0xB1E200A1B1E2009F, 0xB1E200A6B1E200A5, 0xB1E200AAB1E200A8,
-    0xB1E200B1B1E200AC, 0xB2E200B6B1E200B3, 0xB2E20083B2E20081,
-    0xB2E20087B2E20085, 0xB2E2008BB2E20089, 0xB2E2008FB2E2008D,
-    0xB2E20093B2E20091, 0xB2E20097B2E20095, 0xB2E2009BB2E20099,
-    0xB2E2009FB2E2009D, 0xB2E200A3B2E200A1, 0xB2E200A7B2E200A5,
-    0xB2E200ABB2E200A9, 0xB2E200AFB2E200AD, 0xB2E200B3B2E200B1,
-    0xB2E200B7B2E200B5, 0xB2E200BBB2E200B9, 0xB3E200BFB2E200BD,
-    0xB3E20083B3E20081, 0xB3E20087B3E20085, 0xB3E2008BB3E20089,
-    0xB3E2008FB3E2008D, 0xB3E20093B3E20091, 0xB3E20097B3E20095,
-    0xB3E2009BB3E20099, 0xB3E2009FB3E2009D, 0xB3E200A3B3E200A1,
-    0xB3E200AEB3E200AC, 0xB4E20080B4E200B3, 0xB4E20082B4E20081,
-    0xB4E20084B4E20083, 0xB4E20086B4E20085, 0xB4E20088B4E20087,
-    0xB4E2008AB4E20089, 0xB4E2008CB4E2008B, 0xB4E2008EB4E2008D,
-    0xB4E20090B4E2008F, 0xB4E20092B4E20091, 0xB4E20094B4E20093,
-    0xB4E20096B4E20095, 0xB4E20098B4E20097, 0xB4E2009AB4E20099,
-    0xB4E2009CB4E2009B, 0xB4E2009EB4E2009D, 0xB4E200A0B4E2009F,
-    0xB4E200A2B4E200A1, 0xB4E200A4B4E200A3, 0xB4E200A7B4E200A5,
-    0x80E300A1B5E200AD, 0x80E3008280E30081, 0x80E3008980E30088,
-    0x80E3008B80E3008A, 0x80E3008D80E3008C, 0x80E3008F80E3008E,
-    0x80E3009180E30090, 0x80E3009480E30092, 0x80E3009580E37394,
-    0x009580E389B8E494, 0x80E38CBAE49480E3, 0x9D8BE59480E30095,
-    0xE59480E3009580E3, 0x80E3009580E389AE, 0x009580E39389E694,
-    0x80E39795E69480E3, 0xAC9CE69480E30095, 0xE79480E3009580E3,
-    0x80E3009580E3B982, 0x009580E3979BE794, 0x009680E3009580E3,
-    0x008C81E3009780E3, 0x009081E3008E81E3, 0x009481E3009281E3,
-    0x009881E3009681E3, 0x009C81E3009A81E3, 0x00A081E3009E81E3,
-    0x00A581E300A281E3, 0x00A981E300A781E3, 0x00B181E300B081E3,
-    0x00B481E300B381E3, 0x00B781E300B681E3, 0x00BA81E300B981E3,
-    0xE3008B81E3BB81E3, 0xE300BD81E300BC81, 0x82E3008A82E38882,
-    0x82E3009982E30094, 0x82E3009E82E3009A, 0x82E300A282E300A1,
-    0xE3BC83E39183E3A2, 0x83E3A282E3008883, 0x00A182E39583E3AB,
-    0x83E3B383E3A282E3, 0xA282E300A282E39A, 0xE300AB83E3BC83E3,
-    0xE300A482E300A382, 0xB383E38B83E3A482, 0xE3A482E300B082E3,
-    0x82E3008183E3B383, 0x82E300A682E300A5, 0x00B383E3A982E3A6,
-    0x00A882E300A782E3, 0x82E3B982E3A882E3, 0x008983E3BC83E3AF,
-    0x82E3BC83E3A882E3, 0xA982E300BC83E3AB, 0xAA82E300AA82E300,
-    0xE300B982E3B383E3, 0xA083E3BC83E3AA82, 0xAB82E300AB82E300,
-    0xE300AA83E3A482E3, 0x8383E3A983E3AB82, 0xE3AB82E3008883E3,
-    0xBC83E3AA83E3AD83, 0xAC82E300AC82E300, 0xE300B383E3AD83E3,
-    0x9E83E3B383E3AC82, 0xAD82E300AD82E300, 0x83E3AA83E3A583E3,
-    0xAD83E3AD82E300BC, 0xE3AD83E3AD82E300, 0xA083E3A983E3B082,
-    0xE3AD83E3AD82E300, 0x8883E3BC83E3A183, 0xE3AD82E300AB83E3,
-    0x8383E3AF83E3AD83, 0x00AE82E3008883E3, 0xE300AC82E3AE82E3,
-    0xBC83E38B83E3AE82, 0xE3AB83E3AE82E300, 0x82E300BC83E38083,
-    0xAB83E3AF82E300AF, 0x83E3A482E3BC82E3, 0xAD83E3AF82E300AD,
-    0xE3008D83E3BC83E3, 0x83E3B082E300B082, 0xB082E300A083E3A9,
-    0x83E3A083E3A983E3, 0xB182E300B383E388, 0xE3BC83E3B182E300,
-    0xE300B282E300B982, 0x82E3B382E300B382, 0x8883E3B382E300B3,
-    0xE3AB83E3B382E300, 0x83E3B382E3008A83, 0xB482E3009D83E3BC,
-    0xB582E300B582E300, 0x83E3AF82E3A482E3, 0xB383E3B582E300AB,
-    0x83E3BC83E38183E3, 0x82E300B682E300A0, 0xAA83E3B782E300B7,
-    0xE300B082E3B383E3, 0xE300B982E300B882, 0xE300BB82E300BA82,
-    0x8183E3B383E3BB82, 0xE3B383E3BB82E300, 0xE300BC82E3008883,
-    0xE300BE82E300BD82, 0xE3008083E300BF82, 0xB982E3BC83E38083,
-    0x8283E3008183E300, 0x8483E3008383E300, 0x8683E3008583E300,
-    0x8783E3008783E300, 0x008883E300B782E3, 0xE300B383E38883E3,
-    0x83E38983E3008983, 0x83E3008A83E300AB, 0x8B83E3008E83E38A,
-    0x8D83E3008C83E300, 0x8E83E3008E83E300, 0xE3008883E38383E3,
-    0x82E38F83E3008F83, 0x9083E3008483E3A4, 0xE3BC83E39083E300,
-    0x83E300AB83E3AC83, 0xBC83E39183E30091, 0x83E3B383E3BB82E3,
-    0xBC83E39183E30088, 0x009283E3008483E3, 0xE39383E3009383E3,
-    0xE3009483E300AB83, 0xB982E3A282E39483, 0xE300AB83E38883E3,
-    0xAB83E3AF82E39483, 0x00B382E39483E300, 0xE39583E3009583E3,
-    0x8383E3A983E3A182, 0xE39583E3008983E3, 0x8883E3BC83E3A382,
-    0xE3A983E39583E300, 0xE3009683E300B383, 0xB782E38383E39683,
-    0xE300AB83E3A782E3, 0xE3009883E3009783, 0xBF82E3AF82E39883,
-    0xE300AB83E3BC83E3, 0x8483E3AB83E39883, 0x9983E3009983E300,
-    0xE300BF82E3BC83E3, 0x82E39A83E3009A83, 0x8B83E39A83E300BD,
-    0xE39A83E3009283E3, 0x83E300B982E3B383, 0x00B882E3BC83E39A,
-    0xE39B83E3009B83E3, 0x83E39B83E300B383, 0x9B83E300AB83E3BC,
-    0xE300B383E3BC83E3, 0x83E39C83E3009C83, 0x9D83E3008883E3AB,
-    0xE3A482E39D83E300, 0x83E3008883E3B383, 0x008983E3B383E39D,
-    0xE39E83E3009E83E3, 0xAD83E3AF82E3A482, 0xE3A482E39E83E300,
-    0x83E39E83E300AB83, 0x9E83E3008F83E383, 0xE300AF82E3AB83E3,
-    0xB782E3B383E39E83, 0xE300B383E3A783E3, 0x82E39F83E3009F83,
-    0x00B383E3AD83E3AF, 0xE300AA83E39F83E3, 0x9083E3AA83E39F83,
-    0xE300AB83E3BC83E3, 0xE300A183E300A083, 0x83E300AC82E3A183,
-    0xE38883E3AC82E3A1, 0x83E3A183E300B383, 0x00AB83E38883E3BC,
-    0x00A383E300A283E3, 0xE3A483E300A483E3, 0x83E3008983E3BC83,
-    0x00AB83E3BC83E3A4, 0x00A683E300A583E3, 0x83E3A282E3A683E3,
-    0x83E300A783E300B3, 0x83E300A983E300A8, 0x8383E3AA83E300AA,
-    0xE300AB83E38883E3, 0x83E300A983E3AA83, 0x9483E3AB83E300AB,
-    0xE3AB83E300BC83E3, 0xAB83E39683E3BC83, 0xAC83E300AC83E300,
-    0xE3AC83E300A083E3, 0xB282E38883E3B383, 0x00AD83E300B383E3,
-    0xE3AF83E300AF83E3, 0x83E3008883E38383, 0x83E300B183E300B0,
-    0x83E300B383E300B2, 0x83E300B783E300B4, 0x83E300B983E300B8,
-    0x83E300BB83E300BA, 0x92E300BE83E300BC, 0x92E300B992E3009E,
-    0x94E3009F93E300BB, 0x9BE300AE9BE30095, 0xA0E300819EE300BC,
-    0xA1E300A2A1E300AF, 0xA3E30087A3E300BC, 0xA4E3009CA4E300A3,
-    0xA9E300AEA8E300BA, 0xACE300A4ABE300AC, 0xADE30099ACE30088,
-    0xB0E3009DAEE30089, 0xB4E3008EB1E30098, 0xBAE30096B6E300B3,
-    0xBCE300B8BAE300AC, 0x80E400BCBFE3009B, 0x80E4009880E40088,
-    0x82E4008681E400B9, 0x84E400A383E40096, 0x88E4008288E400AF,
-    0x8CE400A08AE400A7, 0x8DE400B48CE40081, 0x8FE400958FE40099,
-    0x91E4008B90E40099, 0x95E400AB94E400AB, 0x95E400A195E4009D,
-    0x97E4009797E400AB, 0x9AE400B598E400B9, 0xA6E400879BE400BE,
-    0xA9E400A6A7E40095, 0xAAE400B6A9E400AE, 0xAFE400B3ACE400B2,
-    0xB3E4008EB3E4008E, 0xB5E400B8B3E400AD, 0xB8E40080B8E40096,
-    0xB8E40083B8E40081, 0xB8E4008AB8E40089, 0xB8E4008DB8E4008B,
-    0xB8E400A6B8E40099, 0xB8E400ADB8E400A8, 0xB8E400B6B8E400B2,
-    0xB8E400B9B8E400B8, 0xB9E400BFB8E400BD, 0xB9E40099B9E40081,
-    0xBAE40082BAE4009D, 0xBAE40086BAE40085, 0xBAE40094BAE4008C,
-    0xBAE400A4BAE400A0, 0xBBE400BABAE400AE, 0xBBE4008CBBE40080,
-    0x8C92E5A4BBE400A4, 0x91BCE40081BCE400, 0x80BEE400A0BDE400,
-    0x8BBEE40086BEE400, 0xBBBEE400AEBEE400, 0x8280E500BFBEE400,
-    0xBA81E500AB80E500, 0x8F83E5009982E500, 0xA783E5009A83E500,
-    0xBF84E500AA84E500, 0x8585E5008085E500, 0x9485E5008D85E500,
-    0xA585E500A485E500, 0xA885E500A785E500, 0xAB85E500A985E500,
-    0xB785E500AD85E500, 0x8286E5008086E500, 0x9286E5008D86E500,
-    0x9686E5009586E500, 0x9986E5009786E500, 0xAB86E500A486E500,
-    0xB586E500AC86E500, 0x8987E500B786E500, 0x9C87E5008C87E500,
-    0xA087E5009E87E500, 0x8088E500B587E500, 0x8788E5008388E500,
-    0x9D88E5009788E500, 0xBA88E500A988E500, 0x8689E500BB88E500,
-    0xB289E5008D89E500, 0x898AE500B789E500, 0xA38AE5009B8AE500,
-    0xB48AE500B38AE500, 0x898BE500878BE500, 0x9E8BE500928BE500,
-    0xB58BE500A48BE500, 0xBA8BE500B98BE500, 0x868CE500858CE500,
-    0x978CE500958CE500, 0xB88CE5009A8CE500, 0xBF8CE500BB8CE500,
-    0x848DE500818DE500, 0x898DE500858DE500, 0x948DE500918DE500,
-    0x9C8DE5009A8DE500, 0xB08DE500A98DE500, 0xB58DE500B38DE500,
-    0xBF8DE500BD8DE500, 0xB68EE500828EE500, 0x888FE500838FE500,
-    0x8C8FE5008A8FE500, 0xA38FE5009F8FE500, 0xAB8FE500A58FE500,
-    0xB18FE500AF8FE500, 0x8690E500B38FE500, 0x8D90E5008890E500,
-    0x9D90E5008F90E500, 0xB990E500B890E500, 0x8891E5008291E500,
-    0x9E92E500A891E500, 0xBD92E500A292E500, 0x9094E500B693E500,
-    0x9395E5008F95E500, 0xA395E5009595E500, 0x8796E5008496E500,
-    0x9D96E5009996E500, 0xB396E500AB96E500, 0x8097E500B696E500,
-    0xA297E5008297E500, 0x9199E5008698E500, 0xB499E500A899E500,
-    0x9B9BE500979BE500, 0x969CE500B99BE500, 0x9F9CE500979CE500,
-    0x8B9EE500B09CE500, 0xB49FE5008E9FE500, 0xB1A0E5008DA0E500,
-    0x80A1E500B2A0E500, 0x9EA1E5009AA1E500, 0xACA2E500A8A2E500,
-    0x98A3E500B3A2E500, 0xABA3E5009FA3E500, 0xB0A3E500AEA3E500,
-    0xB7A3E500B2A3E500, 0x86A4E50082A4E500, 0x95A4E5008AA4E500,
-    0x9CA4E5009AA4E500, 0xA7A4E500A2A4E500, 0x00A3ADE6A7A4E500,
-    0x0084A5E500A9A4E5, 0x0091A5E50088A5E5, 0x00A2A5E50094A5E5,
-    0x0098A7E500B3A5E5, 0x009BA8E500ACA7E5, 0x00A2A9E500A7A8E5,
-    0x00B5AAE500A6A9E5, 0x00A8ACE50088ACE5, 0x0090ADE500BEACE5,
-    0x00A6ADE50097ADE5, 0x0085AEE50080AEE5, 0x0083AFE50097AEE5,
-    0x00A7AFE50098AFE5, 0x00B3AFE500AEAFE5, 0x00BFAFE500B8AFE5,
-    0x008FB0E50086B0E5, 0x00B8B0E500A2B0E5, 0x00A0B1E500BFB0E5,
-    0x00A4B1E500A2B1E5, 0x00AEB1E500A5B1E5, 0x008DB2E500B1B1E5,
-    0x0099B4E50080B3E5, 0x0090B5E50083B5E5, 0x00AEB5E500ABB5E5,
-    0x00B2B6E500BCB5E5, 0x009BB7E500BAB6E5, 0x00A2B7E500A1B7E5,
-    0x00A6B7E500A5B7E5, 0x00BDB7E500B1B7E5, 0x00A8B8E500BEB7E5,
-    0x00A9B9E500BDB8E5, 0xE6B3B9E500B2B9E5, 0xE500B4B9E5009088,
-    0xE500BCB9E500BAB9, 0xE500A6BAE500BFB9, 0xE500B3BAE500B0BA,
-    0xE50089BBE500B6BA, 0xE50092BBE5008ABB, 0xE50099BBE50093BB,
-    0xE500B4BBE500ACBB, 0xE50084BCE500BEBB, 0xE50093BCE5008BBC,
-    0xE50090BDE500A2BC, 0xE500A1BDE50093BD, 0xE500A9BDE500A2BD,
-    0xE500B3BDE500ABBD, 0xE5008CBEE5008BBE, 0xE5009ABEE50097BE,
-    0xE500ADBEE500A9BE, 0xE5008DBFE50083BF, 0xE500B5BFE50097BF,
-    0xE6009280E600B9BF, 0xE600B581E6009C80, 0xE6009482E6008182,
-    0xE6009883E6008783, 0xE6008884E600A183, 0xE6008885E6008485,
-    0xE6008E85E6008C85, 0xE600A885E600A085, 0xE6008E86E600BA85,
-    0xE600A486E6009086, 0xE600B286E600AF86, 0xE600B287E6009E87,
-    0xE6008088E600B687, 0xE6009088E6008888, 0xE600AE88E6009B88,
-    0xE600B688E600B488, 0xE6009389E6008B89, 0xE600958AE6009D89,
-    0xE600898BE600B18A, 0xE600938BE6008F8B, 0xE600BC8BE600948B,
-    0xE600878CE600BE8B, 0xE600908DE600BD8C, 0xE600A88DE600958D,
-    0xE600838EE600BB8D, 0xE600A98EE600A08E, 0xE600858FE600848F,
-    0xE6009C90E600A48F, 0xE6009291E600A290, 0xE600B791E600A991,
-    0xE6009A92E600BE91, 0xE6008493E6009D92, 0xE600B494E600AF94,
-    0xE6009695E6008F95, 0xE600B895E600AC95, 0xE6009796E6008796,
-    0xE600A496E6009996, 0xE600B996E600B096, 0xE600A097E6008597,
-    0xE600A397E600A297, 0xB2E68E98E600A597, 0x98E6009398E600BB,
-    0x8C92E5AD98E600A0, 0xB499E6008999E600, 0x919AE600889AE600,
-    0xB49AE6009C9AE600, 0xB09BE600869BE600, 0xB89BE600B49BE600,
-    0x889CE600809CE600, 0x979CE600899CE600, 0xA19CE6009B9CE600,
-    0x8E9DE600A89CE600, 0x969DE600939DE600, 0xBB9DE6009E9DE600,
-    0x979EE600859EE600, 0xBA9FE600B39FE600, 0x9FA0E60097A0E600,
-    0xAAA0E600AAA0E600, 0xA4E79ABCE48FBCE5, 0xA2E60092A1E600BE,
-    0xA2E60085A2E60081, 0xA4E600A8A2E6008E, 0xA6E60082A5E60094,
-    0xA8E600AAA7E600A3, 0xAAE60093A8E60082, 0xABE60093ABE600A8,
-    0xACE60084ACE6009B, 0xADE600A1ACE600A0, 0xADE600A2ADE60094,
-    0xADE600B2ADE600A3, 0xAEE600B9ADE600B7, 0xAEE600AEAEE6009F,
-    0xAEE600BAAEE600B3, 0xAFE6008BAFE600BB, 0xAFE60094AFE6008D,
-    0xB0E6008FB0E6009B, 0xB1E600B4B0E60094, 0xB2E600A7B1E6008E,
-    0xB3E600BFB2E60088, 0xB3E6008DB3E6008C, 0xB4E600A8B3E600A5,
-    0xB4E6009BB4E60096, 0xB4E600B4B4E6009E, 0xB5E60081B5E600BE,
-    0xB5E600AAB5E600A9, 0xB6E600B8B5E600B7, 0xB7E6008BB7E60085,
-    0xB7E600AAB7E6009A, 0xB8E6009AB8E600B9, 0xBAE600AEB9E600AF,
-    0xBAE6009CBAE60080, 0xBBE60087BBE600BA, 0xBBE60091BBE6008B,
-    0xBCE6008FBCE6009B, 0xBCE600A2BCE60094, 0xBFE600AEBDE600A3,
-    0xBFE600ABBFE60086, 0x80E7009B80E700BE, 0x81E700B980E7009E,
-    0x81E700AB81E7008A, 0x81E700B781E700B0, 0x82E7009982E700BD,
-    0x83E7008883E700AD, 0x85E700A184E70099, 0x85E7008985E70085,
-    0x87E7009C86E700AE, 0x88E7009087E7008E, 0x88E7009B88E70090,
-    0x88E700AA88E700A8, 0x88E700B588E700AB, 0x88E700BB88E700B6,
-    0x89E7008789E700BF, 0x89E7009989E70090, 0x89E700A289E7009B,
-    0x8AE700808AE700B9, 0x8AE700AC8AE70095, 0x8BE700808BE700AF,
-    0x8DE700AA8CE700BC, 0x8EE700BA8DE700B5, 0x8EE700878EE70084,
-    0x8EE7008B8EE70089, 0x8FE700B28EE700A5, 0x90E7008690E7009E,
-    0x91E700A290E70089, 0x91E7009C91E70087, 0x92E700B191E700A9,
-    0x92E7008992E70085, 0x93E7008A93E70098, 0x94E700A693E7009C,
-    0x94E7009894E70086, 0x94E700A494E7009F, 0x94E700B094E700A8,
-    0x94E700B394E700B2, 0x94E700BB94E700B7, 0x95E7009995E700BE,
-    0x96E700B095E700A5, 0x97E7009296E7008B, 0x98E7009098E700A2,
-    0x99E7009F98E7009D, 0x99E700A999E70082, 0x9AE700BD99E700B6,
-    0x9BE700BF9AE700AE, 0x9BE7009B9BE7008A, 0x9BE700A79BE700A3,
-    0x9CE700B49BE700AE, 0x9CE7009E9CE70081, 0x9DE700809DE7009F,
-    0x9EE7008B9EE7008A, 0x9FE7009B9FE700A7, 0xA1E700B39FE700A2,
-    0xA2E700ABA1E7008E, 0xA3E70091A2E7008C, 0xA3E7008CA3E7008A,
-    0xA4E700AAA4E700BB, 0xA4E700BCA4E700BA, 0xA5E70088A5E700BE,
-    0xA5E70090A5E70089, 0xA5E7009DA5E70096, 0xA5E700A5A5E7009E,
-    0xA6E70081A6E700BF, 0xA6E7008EA6E7008D, 0xA6E700AEA6E7008F,
-    0xA7E700BEA6E700B8, 0xA7E70098A7E7008A, 0xA9E7009CA8E700AB,
-    0xA9E7008AA9E70080, 0xA9E700B4A9E7008F, 0xAAE70081AAE700BA,
-    0xABE7008BABE700B1, 0xACE700B9ABE700AE, 0xAFE7008FAEE700A0,
-    0xAFE70086AFE70080, 0xB1E700BEB0E70089, 0xB1E700B3B1E700A0,
-    0xB2E70092B2E700BB, 0xB3E70092B3E700BE, 0xB3E700A3B3E70096,
-    0xB3E700A8B3E700A7, 0xB4E70080B4E700B8, 0xB4E700A2B4E70090,
-    0xB5E70082B5E700AF, 0xB6E700A3B5E7009B, 0xB7E700BEB6E700A0,
-    0xB8E700B4B7E70087, 0xB8E70089B8E70082, 0xB9E70081B9E700B7,
-    0xBCE700B6BCE70085, 0xBDE70091BDE700BE, 0xBDE700B9BDE700B2,
-    0xBEE70085BEE700BA, 0xBEE70095BEE7008A, 0xBFE700BDBEE7009A,
-    0x80E8008180E800BA, 0x80E8008C80E80085, 0x81E800B380E80092,
-    0x81E800A081E80086, 0x81E800B081E800AF, 0x82E800BF81E800BE,
-    0x82E8008B82E80089, 0x84E800B282E800AD, 0x87E800BE84E80083,
-    0x87E800A387E80098, 0x87E800AA87E800A8, 0x87E800B387E800AD,
-    0x88E8008188E800BC, 0x88E8008C88E80084, 0x88E8009B88E80098,
-    0x89E800AE89E8009F, 0x89E800B289E800AF, 0x8AE800B989E800B8,
-    0x8AE800918AE8008B, 0x8AE800B18AE8009D, 0x8BE800BD8AE800B3,
-    0x8CE800A68BE800A5, 0x8CE800A38CE8009D, 0x8DE800928DE800B6,
-    0x8EE800A38DE80093, 0x8FE800BD8EE800AD, 0x8FE8008A8FE80089,
-    0x8FE8009C8FE8008C, 0x8FE800AF8FE800A7, 0x91E800BD90E800B1,
-    0x93E8009791E80089, 0x93E800B193E800AE, 0x94E800BC93E800B3,
-    0x97E800A495E80096, 0x98E800BA97E8008D, 0x98E8009298E80086,
-    0x99E800BF98E800AD, 0x99E8009099E8008D, 0x99E800A799E8009C,
-    0x9AE800AB99E800A9, 0x9BE800A99AE80088, 0x9CE8008E9CE800A2,
-    0x9DE800AB9DE800A8, 0x9EE800869EE800B9, 0xA0E800A19FE800BA,
-    0xA1E8009FA0E80081, 0xA1E8008CA1E80080, 0xA3E800A3A1E800A0,
-    0xA3E8008FA3E80082, 0xA3E8009EA3E80097, 0xA3E800B8A3E800A1,
-    0xA5E80090A4E800BA, 0xA5E800A4A5E80081, 0xA6E80086A6E800BE,
-    0xA7E80096A6E8008B, 0xA8E800A3A7E80092, 0xAAE800A0AAE80080,
-    0xABE800BFAAE800AA, 0xABE80092ABE8008B, 0xABE800ADABE80096,
-    0xACE800BEABE800B8, 0xADE800B9ACE80081, 0xAEE80080AEE80098,
-    0xB1E800B7B0E8008A, 0xB1E80088B1E80086, 0xB2E800B8B1E80095,
-    0xB2E800A1B2E8009D, 0xB3E800ABB2E800A9, 0xB3E80082B3E80081,
-    0xB3E80088B3E80087, 0xB4E80088B4E80093, 0xB5E800A4B5E8009B,
-    0xB6E800B7B5E800B0, 0xB7E800BCB6E800B3, 0xB7E800AFB7E8008B,
-    0xBBE800ABBAE800B0, 0xBCE80094BBE8008A, 0xBCE800AABCE800A6,
-    0xBDE800BBBCE800B8, 0xBEE8009BBEE800A2, 0xBEE800B0BEE8009E,
-    0x80E900B6BEE800B5, 0x81E900B880E900A3, 0x81E900A981E9008A,
-    0x82E900BC81E900B2, 0x82E9009182E9008F, 0x83E9008E83E90094,
-    0x83E900B183E9009E, 0x84E9009184E900BD, 0x85E9008985E9009B,
-    0x86E900AA85E9008D, 0x87E900B486E90099, 0x87E9008C87E90086,
-    0x88E9009187E9008F, 0x89E900B888E900B4, 0x8BE900BC89E900B6,
-    0x8CE900988BE90097, 0x8FE9008A8DE90084, 0x95E9009590E900B9,
-    0x96E9008096E900B7, 0x96E900AD96E9008B, 0x98E9009C98E900B7,
-    0x99E9008B99E900AE, 0x99E900B599E9008D, 0x9AE900BC99E900B8,
-    0x9AE900A39AE90086, 0x9AE900B79AE900B6, 0x9BE900B99AE900B8,
-    0x9BE900A29BE90083, 0x9BE900A89BE900A3, 0x9CE900B79BE900B6,
-    0x9DE900B29CE900A3, 0x9DE900919DE90088, 0x9DE9009E9DE90096,
-    0x9FE900A99DE900A2, 0x9FE9009B9FE9008B, 0x9FE900AD9FE900A0,
-    0xA0E900BF9FE900B3, 0xA0E90085A0E90081, 0xA0E90098A0E9008B,
-    0xA1E900BBA0E900A9, 0xA3E900A8A2E9009E, 0xA3E9009FA3E9009B,
-    0xA3E900AFA3E900A2, 0xA4E900A8A4E900BC, 0xA6E90096A6E900A9,
-    0xA6E900A7A6E90099, 0xA7E90082A7E900AC, 0xA9E900BEA7E900B1,
-    0xABE900A8AAE900AA, 0xACE9009FABE90098, 0xACE900A5ACE90092,
-    0xACE900B2ACE900AF, 0xADE9009AADE900BC, 0xB1E90080B1E900AF,
-    0xB3E900A5B3E90097, 0xB6E900A7B5E900BD, 0xB8E900BAB7E900B4,
-    0xB9E900B5B9E9009E, 0xBAE90097BAE900BF, 0xBAE900A5BAE9009F,
-    0xBBE90083BBE900BB, 0xBBE9008EBBE9008D, 0xBBE900B9BBE90091,
-    0xBCE900BEBBE900BD, 0xBCE9008EBCE90085, 0xBCE90093BCE9008F,
-    0xBCE900A0BCE90096, 0xBDE90083BDE900BB, 0xBEE90092BDE9008A,
-    0xBEE9008EBEE9008D, 0xBEE9009FBEE9009C, 0x99EA008199EA00A0,
-    0x99EA008599EA0083, 0x99EA008999EA0087, 0x99EA008D99EA008B,
-    0x99EA009199EA008F, 0x99EA009599EA0093, 0x99EA009999EA0097,
-    0x99EA009D99EA009B, 0x99EA00A199EA009F, 0x99EA00A599EA00A3,
-    0x99EA00A999EA00A7, 0x9AEA00AD99EA00AB, 0x9AEA00839AEA0081,
-    0x9AEA00879AEA0085, 0x9AEA008B9AEA0089, 0x9AEA008F9AEA008D,
-    0x9AEA00939AEA0091, 0x9AEA00979AEA0095, 0x9CEA009B9AEA0099,
-    0x9CEA00A59CEA00A3, 0x9CEA00A99CEA00A7, 0x9CEA00AD9CEA00AB,
-    0x9CEA00B39CEA00AF, 0x9CEA00B79CEA00B5, 0x9CEA00BB9CEA00B9,
-    0x9DEA00BF9CEA00BD, 0x9DEA00839DEA0081, 0x9DEA00879DEA0085,
-    0x9DEA008B9DEA0089, 0x9DEA008F9DEA008D, 0x9DEA00939DEA0091,
-    0x9DEA00979DEA0095, 0x9DEA009B9DEA0099, 0x9DEA009F9DEA009D,
-    0x9DEA00A39DEA00A1, 0x9DEA00A79DEA00A5, 0x9DEA00AB9DEA00A9,
-    0x9DEA00AF9DEA00AD, 0x9DEA00BC9DEA00BA, 0x9EEA00819EEA00BF,
-    0x9EEA00859EEA0083, 0x9EEA008C9EEA0087, 0x9EEA00919EEA008E,
-    0x9EEA00949EEA0093, 0x9EEA00999EEA0097, 0x9EEA009D9EEA009B,
-    0x9EEA00A19EEA009F, 0x9EEA00A59EEA00A3, 0x9EEA00A99EEA00A7,
-    0x9EEA00B79EEA00B5, 0x9EEA00BB9EEA00B9, 0x9FEA00BF9EEA00BD,
-    0x9FEA00839FEA0081, 0x9FEA008A9FEA0088, 0x9FEA00979FEA0091,
-    0xACEA00B69FEA0099, 0xADEA0092ADEA00B7, 0xADEA00A6ADEA0093,
-    0xB0EA0080B0EA00A7, 0xB0EA0082B0EA0081, 0xB0EA0084B0EA0083,
-    0xB0EA0086B0EA0085, 0xB0EA0088B0EA0087, 0xB0EA008AB0EA0089,
-    0xB0EA008CB0EA008B, 0xB0EA008EB0EA008D, 0xB0EA0090B0EA008F,
-    0xB0EA0092B0EA0091, 0xB0EA0094B0EA0093, 0xB0EA0096B0EA0095,
-    0xB0EA0098B0EA0097, 0xB0EA009AB0EA0099, 0xB0EA009CB0EA009B,
-    0xB0EA009EB0EA009D, 0xB0EA00A0B0EA009F, 0xB0EA00A2B0EA00A1,
-    0xB0EA00A4B0EA00A3, 0xB0EA00A6B0EA00A5, 0xB0EA00A8B0EA00A7,
-    0xB0EA00AAB0EA00A9, 0xB0EA00ACB0EA00AB, 0xB0EA00AEB0EA00AD,
-    0xB0EA00B0B0EA00AF, 0xB0EA00B2B0EA00B1, 0xB0EA00B4B0EA00B3,
-    0xB0EA00B6B0EA00B5, 0xB0EA00B8B0EA00B7, 0xB0EA00BAB0EA00B9,
-    0xB0EA00BCB0EA00BB, 0xB0EA00BEB0EA00BD, 0xB1EA0080B1EA00BF,
-    0xB1EA0082B1EA0081, 0xB1EA0084B1EA0083, 0xB1EA0086B1EA0085,
-    0xB1EA0088B1EA0087, 0xB1EA008AB1EA0089, 0xB1EA008CB1EA008B,
-    0xB1EA008EB1EA008D, 0xB1EA0090B1EA008F, 0xB1EA0092B1EA0091,
-    0xB1EA0094B1EA0093, 0xB1EA0096B1EA0095, 0xB1EA0098B1EA0097,
-    0xB1EA009AB1EA0099, 0xB1EA009CB1EA009B, 0xB1EA009EB1EA009D,
-    0xB1EA00A0B1EA009F, 0xB1EA00A2B1EA00A1, 0xB1EA00A4B1EA00A3,
-    0xB1EA00A6B1EA00A5, 0xB1EA00A8B1EA00A7, 0xB1EA00AAB1EA00A9,
-    0xB1EA00ACB1EA00AB, 0xB1EA00AEB1EA00AD, 0xB1EA00B0B1EA00AF,
-    0xB1EA00B2B1EA00B1, 0xB1EA00B4B1EA00B3, 0xB1EA00B6B1EA00B5,
-    0xB1EA00B8B1EA00B7, 0xB1EA00BAB1EA00B9, 0xB1EA00BCB1EA00BB,
-    0xB1EA00BEB1EA00BD, 0xB2EA0080B2EA00BF, 0xB2EA0082B2EA0081,
-    0xB2EA0084B2EA0083, 0xB2EA0086B2EA0085, 0xB2EA0088B2EA0087,
-    0xB2EA008AB2EA0089, 0xB2EA008CB2EA008B, 0xB2EA008EB2EA008D,
-    0xB2EA0090B2EA008F, 0xB2EA0092B2EA0091, 0xB2EA0094B2EA0093,
-    0xB2EA0096B2EA0095, 0xB2EA0098B2EA0097, 0xB2EA009AB2EA0099,
-    0xB2EA009CB2EA009B, 0xB2EA009EB2EA009D, 0xB2EA00A0B2EA009F,
-    0xB2EA00A2B2EA00A1, 0xB2EA00A4B2EA00A3, 0xB2EA00A6B2EA00A5,
-    0xB2EA00A8B2EA00A7, 0xB2EA00AAB2EA00A9, 0xB2EA00ACB2EA00AB,
-    0xB2EA00AEB2EA00AD, 0xB2EA00B0B2EA00AF, 0xB2EA00B2B2EA00B1,
-    0xB2EA00B4B2EA00B3, 0xB2EA00B6B2EA00B5, 0xB2EA00B8B2EA00B7,
-    0xB2EA00BAB2EA00B9, 0xB2EA00BCB2EA00BB, 0xB2EA00BEB2EA00BD,
-    0xB3EA0080B3EA00BF, 0xB3EA0082B3EA0081, 0xB3EA0084B3EA0083,
-    0xB3EA0086B3EA0085, 0xB3EA0088B3EA0087, 0xB3EA008AB3EA0089,
-    0xB3EA008CB3EA008B, 0xB3EA008EB3EA008D, 0xB3EA0090B3EA008F,
-    0xB3EA0092B3EA0091, 0xB3EA0094B3EA0093, 0xB3EA0096B3EA0095,
-    0xB3EA0098B3EA0097, 0xB3EA009AB3EA0099, 0xB3EA009CB3EA009B,
-    0xB3EA009EB3EA009D, 0xB3EA00A0B3EA009F, 0xB3EA00A2B3EA00A1,
-    0xB3EA00A4B3EA00A3, 0xB3EA00A6B3EA00A5, 0xB3EA00A8B3EA00A7,
-    0xB3EA00AAB3EA00A9, 0xB3EA00ACB3EA00AB, 0xB3EA00AEB3EA00AD,
-    0xB3EA00B0B3EA00AF, 0xB3EA00B2B3EA00B1, 0xB3EA00B4B3EA00B3,
-    0xB3EA00B6B3EA00B5, 0xB3EA00B8B3EA00B7, 0xB3EA00BAB3EA00B9,
-    0xB3EA00BCB3EA00BB, 0xB3EA00BEB3EA00BD, 0xB4EA0080B4EA00BF,
-    0xB4EA0082B4EA0081, 0xB4EA0084B4EA0083, 0xB4EA0086B4EA0085,
-    0xB4EA0088B4EA0087, 0xB4EA008AB4EA0089, 0xB4EA008CB4EA008B,
-    0xB4EA008EB4EA008D, 0xB4EA0090B4EA008F, 0xB4EA0092B4EA0091,
-    0xB4EA0094B4EA0093, 0xB4EA0096B4EA0095, 0xB4EA0098B4EA0097,
-    0xB4EA009AB4EA0099, 0xB4EA009CB4EA009B, 0xB4EA009EB4EA009D,
-    0xB4EA00A0B4EA009F, 0xB4EA00A2B4EA00A1, 0xB4EA00A4B4EA00A3,
-    0xB4EA00A6B4EA00A5, 0xB4EA00A8B4EA00A7, 0xB4EA00AAB4EA00A9,
-    0xB4EA00ACB4EA00AB, 0xB4EA00AEB4EA00AD, 0xB4EA00B0B4EA00AF,
-    0xB4EA00B2B4EA00B1, 0xB4EA00B4B4EA00B3, 0xB4EA00B6B4EA00B5,
-    0xB4EA00B8B4EA00B7, 0xB4EA00BAB4EA00B9, 0xB4EA00BCB4EA00BB,
-    0xB4EA00BEB4EA00BD, 0xB5EA0080B5EA00BF, 0xB5EA0082B5EA0081,
-    0xB5EA0084B5EA0083, 0xB5EA0086B5EA0085, 0xB5EA0088B5EA0087,
-    0xB5EA008AB5EA0089, 0xB5EA008CB5EA008B, 0xB5EA008EB5EA008D,
-    0xB5EA0090B5EA008F, 0xB5EA0092B5EA0091, 0xB5EA0094B5EA0093,
-    0xB5EA0096B5EA0095, 0xB5EA0098B5EA0097, 0xB5EA009AB5EA0099,
-    0xB5EA009CB5EA009B, 0xB5EA009EB5EA009D, 0xB5EA00A0B5EA009F,
-    0xB5EA00A2B5EA00A1, 0xB5EA00A4B5EA00A3, 0xB5EA00A6B5EA00A5,
-    0xB5EA00A8B5EA00A7, 0xB5EA00AAB5EA00A9, 0xB5EA00ACB5EA00AB,
-    0xB5EA00AEB5EA00AD, 0xB5EA00B0B5EA00AF, 0xB5EA00B2B5EA00B1,
-    0xB5EA00B4B5EA00B3, 0xB5EA00B6B5EA00B5, 0xB5EA00B8B5EA00B7,
-    0xB5EA00BAB5EA00B9, 0xB5EA00BCB5EA00BB, 0xB5EA00BEB5EA00BD,
-    0xB6EA0080B6EA00BF, 0xB6EA0082B6EA0081, 0xB6EA0084B6EA0083,
-    0xB6EA0086B6EA0085, 0xB6EA0088B6EA0087, 0xB6EA008AB6EA0089,
-    0xB6EA008CB6EA008B, 0xB6EA008EB6EA008D, 0xB6EA0090B6EA008F,
-    0xB6EA0092B6EA0091, 0xB6EA0094B6EA0093, 0xB6EA0096B6EA0095,
-    0xB6EA0098B6EA0097, 0xB6EA009AB6EA0099, 0xB6EA009CB6EA009B,
-    0xB6EA009EB6EA009D, 0xB6EA00A0B6EA009F, 0xB6EA00A2B6EA00A1,
-    0xB6EA00A4B6EA00A3, 0xB6EA00A6B6EA00A5, 0xB6EA00A8B6EA00A7,
-    0xB6EA00AAB6EA00A9, 0xB6EA00ACB6EA00AB, 0xB6EA00AEB6EA00AD,
-    0xB6EA00B0B6EA00AF, 0xB6EA00B2B6EA00B1, 0xB6EA00B4B6EA00B3,
-    0xB6EA00B6B6EA00B5, 0xB6EA00B8B6EA00B7, 0xB6EA00BAB6EA00B9,
-    0xB6EA00BCB6EA00BB, 0xB6EA00BEB6EA00BD, 0xB7EA0080B7EA00BF,
-    0xB7EA0082B7EA0081, 0xB7EA0084B7EA0083, 0xB7EA0086B7EA0085,
-    0xB7EA0088B7EA0087, 0xB7EA008AB7EA0089, 0xB7EA008CB7EA008B,
-    0xB7EA008EB7EA008D, 0xB7EA0090B7EA008F, 0xB7EA0092B7EA0091,
-    0xB7EA0094B7EA0093, 0xB7EA0096B7EA0095, 0xB7EA0098B7EA0097,
-    0xB7EA009AB7EA0099, 0xB7EA009CB7EA009B, 0xB7EA009EB7EA009D,
-    0xB7EA00A0B7EA009F, 0xB7EA00A2B7EA00A1, 0xB7EA00A4B7EA00A3,
-    0xB7EA00A6B7EA00A5, 0xB7EA00A8B7EA00A7, 0xB7EA00AAB7EA00A9,
-    0xB7EA00ACB7EA00AB, 0xB7EA00AEB7EA00AD, 0xB7EA00B0B7EA00AF,
-    0xB7EA00B2B7EA00B1, 0xB7EA00B4B7EA00B3, 0xB7EA00B6B7EA00B5,
-    0xB7EA00B8B7EA00B7, 0xB7EA00BAB7EA00B9, 0xB7EA00BCB7EA00BB,
-    0xB7EA00BEB7EA00BD, 0xB8EA0080B8EA00BF, 0xB8EA0082B8EA0081,
-    0xB8EA0084B8EA0083, 0xB8EA0086B8EA0085, 0xB8EA0088B8EA0087,
-    0xB8EA008AB8EA0089, 0xB8EA008CB8EA008B, 0xB8EA008EB8EA008D,
-    0xB8EA0090B8EA008F, 0xB8EA0092B8EA0091, 0xB8EA0094B8EA0093,
-    0xB8EA0096B8EA0095, 0xB8EA0098B8EA0097, 0xB8EA009AB8EA0099,
-    0xB8EA009CB8EA009B, 0xB8EA009EB8EA009D, 0xB8EA00A0B8EA009F,
-    0xB8EA00A2B8EA00A1, 0xB8EA00A4B8EA00A3, 0xB8EA00A6B8EA00A5,
-    0xB8EA00A8B8EA00A7, 0xB8EA00AAB8EA00A9, 0xB8EA00ACB8EA00AB,
-    0xB8EA00AEB8EA00AD, 0xB8EA00B0B8EA00AF, 0xB8EA00B2B8EA00B1,
-    0xB8EA00B4B8EA00B3, 0xB8EA00B6B8EA00B5, 0xB8EA00B8B8EA00B7,
-    0xB8EA00BAB8EA00B9, 0xB8EA00BCB8EA00BB, 0xB8EA00BEB8EA00BD,
-    0xB9EA0080B9EA00BF, 0xB9EA0082B9EA0081, 0xB9EA0084B9EA0083,
-    0xB9EA0086B9EA0085, 0xB9EA0088B9EA0087, 0xB9EA008AB9EA0089,
-    0xB9EA008CB9EA008B, 0xB9EA008EB9EA008D, 0xB9EA0090B9EA008F,
-    0xB9EA0092B9EA0091, 0xB9EA0094B9EA0093, 0xB9EA0096B9EA0095,
-    0xB9EA0098B9EA0097, 0xB9EA009AB9EA0099, 0xB9EA009CB9EA009B,
-    0xB9EA009EB9EA009D, 0xB9EA00A0B9EA009F, 0xB9EA00A2B9EA00A1,
-    0xB9EA00A4B9EA00A3, 0xB9EA00A6B9EA00A5, 0xB9EA00A8B9EA00A7,
-    0xB9EA00AAB9EA00A9, 0xB9EA00ACB9EA00AB, 0xB9EA00AEB9EA00AD,
-    0xB9EA00B0B9EA00AF, 0xB9EA00B2B9EA00B1, 0xB9EA00B4B9EA00B3,
-    0xB9EA00B6B9EA00B5, 0xB9EA00B8B9EA00B7, 0xB9EA00BAB9EA00B9,
-    0xB9EA00BCB9EA00BB, 0xB9EA00BEB9EA00BD, 0xBAEA0080BAEA00BF,
-    0xBAEA0082BAEA0081, 0xBAEA0084BAEA0083, 0xBAEA0086BAEA0085,
-    0xBAEA0088BAEA0087, 0xBAEA008ABAEA0089, 0xBAEA008CBAEA008B,
-    0xBAEA008EBAEA008D, 0xBAEA0090BAEA008F, 0xBAEA0092BAEA0091,
-    0xBAEA0094BAEA0093, 0xBAEA0096BAEA0095, 0xBAEA0098BAEA0097,
-    0xBAEA009ABAEA0099, 0xBAEA009CBAEA009B, 0xBAEA009EBAEA009D,
-    0xBAEA00A0BAEA009F, 0xBAEA00A2BAEA00A1, 0xBAEA00A4BAEA00A3,
-    0xBAEA00A6BAEA00A5, 0xBAEA00A8BAEA00A7, 0xBAEA00AABAEA00A9,
-    0xBAEA00ACBAEA00AB, 0xBAEA00AEBAEA00AD, 0xBAEA00B0BAEA00AF,
-    0xBAEA00B2BAEA00B1, 0xBAEA00B4BAEA00B3, 0xBAEA00B6BAEA00B5,
-    0xBAEA00B8BAEA00B7, 0xBAEA00BABAEA00B9, 0xBAEA00BCBAEA00BB,
-    0xBAEA00BEBAEA00BD, 0xBBEA0080BBEA00BF, 0xBBEA0082BBEA0081,
-    0xBBEA0084BBEA0083, 0xBBEA0086BBEA0085, 0xBBEA0088BBEA0087,
-    0xBBEA008ABBEA0089, 0xBBEA008CBBEA008B, 0xBBEA008EBBEA008D,
-    0xBBEA0090BBEA008F, 0xBBEA0092BBEA0091, 0xBBEA0094BBEA0093,
-    0xBBEA0096BBEA0095, 0xBBEA0098BBEA0097, 0xBBEA009ABBEA0099,
-    0xBBEA009CBBEA009B, 0xBBEA009EBBEA009D, 0xBBEA00A0BBEA009F,
-    0xBBEA00A2BBEA00A1, 0xBBEA00A4BBEA00A3, 0xBBEA00A6BBEA00A5,
-    0xBBEA00A8BBEA00A7, 0xBBEA00AABBEA00A9, 0xBBEA00ACBBEA00AB,
-    0xBBEA00AEBBEA00AD, 0xBBEA00B0BBEA00AF, 0xBBEA00B2BBEA00B1,
-    0xBBEA00B4BBEA00B3, 0xBBEA00B6BBEA00B5, 0xBBEA00B8BBEA00B7,
-    0xBBEA00BABBEA00B9, 0xBBEA00BCBBEA00BB, 0xBBEA00BEBBEA00BD,
-    0xBCEA0080BCEA00BF, 0xBCEA0082BCEA0081, 0xBCEA0084BCEA0083,
-    0xBCEA0086BCEA0085, 0xBCEA0088BCEA0087, 0xBCEA008ABCEA0089,
-    0xBCEA008CBCEA008B, 0xBCEA008EBCEA008D, 0xBCEA0090BCEA008F,
-    0xBCEA0092BCEA0091, 0xBCEA0094BCEA0093, 0xBCEA0096BCEA0095,
-    0xBCEA0098BCEA0097, 0xBCEA009ABCEA0099, 0xBCEA009CBCEA009B,
-    0xBCEA009EBCEA009D, 0xBCEA00A0BCEA009F, 0xBCEA00A2BCEA00A1,
-    0xBCEA00A4BCEA00A3, 0xBCEA00A6BCEA00A5, 0xBCEA00A8BCEA00A7,
-    0xBCEA00AABCEA00A9, 0xBCEA00ACBCEA00AB, 0xBCEA00AEBCEA00AD,
-    0xBCEA00B0BCEA00AF, 0xBCEA00B2BCEA00B1, 0xBCEA00B4BCEA00B3,
-    0xBCEA00B6BCEA00B5, 0xBCEA00B8BCEA00B7, 0xBCEA00BABCEA00B9,
-    0xBCEA00BCBCEA00BB, 0xBCEA00BEBCEA00BD, 0xBDEA0080BDEA00BF,
-    0xBDEA0082BDEA0081, 0xBDEA0084BDEA0083, 0xBDEA0086BDEA0085,
-    0xBDEA0088BDEA0087, 0xBDEA008ABDEA0089, 0xBDEA008CBDEA008B,
-    0xBDEA008EBDEA008D, 0xBDEA0090BDEA008F, 0xBDEA0092BDEA0091,
-    0xBDEA0094BDEA0093, 0xBDEA0096BDEA0095, 0xBDEA0098BDEA0097,
-    0xBDEA009ABDEA0099, 0xBDEA009CBDEA009B, 0xBDEA009EBDEA009D,
-    0xBDEA00A0BDEA009F, 0xBDEA00A2BDEA00A1, 0xBDEA00A4BDEA00A3,
-    0xBDEA00A6BDEA00A5, 0xBDEA00A8BDEA00A7, 0xBDEA00AABDEA00A9,
-    0xBDEA00ACBDEA00AB, 0xBDEA00AEBDEA00AD, 0xBDEA00B0BDEA00AF,
-    0xBDEA00B2BDEA00B1, 0xBDEA00B4BDEA00B3, 0xBDEA00B6BDEA00B5,
-    0xBDEA00B8BDEA00B7, 0xBDEA00BABDEA00B9, 0xBDEA00BCBDEA00BB,
-    0xBDEA00BEBDEA00BD, 0xBEEA0080BEEA00BF, 0xBEEA0082BEEA0081,
-    0xBEEA0084BEEA0083, 0xBEEA0086BEEA0085, 0xBEEA0088BEEA0087,
-    0xBEEA008ABEEA0089, 0xBEEA008CBEEA008B, 0xBEEA008EBEEA008D,
-    0xBEEA0090BEEA008F, 0xBEEA0092BEEA0091, 0xBEEA0094BEEA0093,
-    0xBEEA0096BEEA0095, 0xBEEA0098BEEA0097, 0xBEEA009ABEEA0099,
-    0xBEEA009CBEEA009B, 0xBEEA009EBEEA009D, 0xBEEA00A0BEEA009F,
-    0xBEEA00A2BEEA00A1, 0xBEEA00A4BEEA00A3, 0xBEEA00A6BEEA00A5,
-    0xBEEA00A8BEEA00A7, 0xBEEA00AABEEA00A9, 0xBEEA00ACBEEA00AB,
-    0xBEEA00AEBEEA00AD, 0xBEEA00B0BEEA00AF, 0xBEEA00B2BEEA00B1,
-    0xBEEA00B4BEEA00B3, 0xBEEA00B6BEEA00B5, 0xBEEA00B8BEEA00B7,
-    0xBEEA00BABEEA00B9, 0xBEEA00BCBEEA00BB, 0xBEEA00BEBEEA00BD,
-    0xBFEA0080BFEA00BF, 0xBFEA0082BFEA0081, 0xBFEA0084BFEA0083,
-    0xBFEA0086BFEA0085, 0xBFEA0088BFEA0087, 0xBFEA008ABFEA0089,
-    0xBFEA008CBFEA008B, 0xBFEA008EBFEA008D, 0xBFEA0090BFEA008F,
-    0xBFEA0092BFEA0091, 0xBFEA0094BFEA0093, 0xBFEA0096BFEA0095,
-    0xBFEA0098BFEA0097, 0xBFEA009ABFEA0099, 0xBFEA009CBFEA009B,
-    0xBFEA009EBFEA009D, 0xBFEA00A0BFEA009F, 0xBFEA00A2BFEA00A1,
-    0xBFEA00A4BFEA00A3, 0xBFEA00A6BFEA00A5, 0xBFEA00A8BFEA00A7,
-    0xBFEA00AABFEA00A9, 0xBFEA00ACBFEA00AB, 0xBFEA00AEBFEA00AD,
-    0xBFEA00B0BFEA00AF, 0xBFEA00B2BFEA00B1, 0xBFEA00B4BFEA00B3,
-    0xBFEA00B6BFEA00B5, 0xBFEA00B8BFEA00B7, 0xBFEA00BABFEA00B9,
-    0xBFEA00BCBFEA00BB, 0xBFEA00BEBFEA00BD, 0x80EB008080EB00BF,
-    0x80EB008280EB0081, 0x80EB008480EB0083, 0x80EB008680EB0085,
-    0x80EB008880EB0087, 0x80EB008A80EB0089, 0x80EB008C80EB008B,
-    0x80EB008E80EB008D, 0x80EB009080EB008F, 0x80EB009280EB0091,
-    0x80EB009480EB0093, 0x80EB009680EB0095, 0x80EB009880EB0097,
-    0x80EB009A80EB0099, 0x80EB009C80EB009B, 0x80EB009E80EB009D,
-    0x80EB00A080EB009F, 0x80EB00A280EB00A1, 0x80EB00A480EB00A3,
-    0x80EB00A680EB00A5, 0x80EB00A880EB00A7, 0x80EB00AA80EB00A9,
-    0x80EB00AC80EB00AB, 0x80EB00AE80EB00AD, 0x80EB00B080EB00AF,
-    0x80EB00B280EB00B1, 0x80EB00B480EB00B3, 0x80EB00B680EB00B5,
-    0x80EB00B880EB00B7, 0x80EB00BA80EB00B9, 0x80EB00BC80EB00BB,
-    0x80EB00BE80EB00BD, 0x81EB008081EB00BF, 0x81EB008281EB0081,
-    0x81EB008481EB0083, 0x81EB008681EB0085, 0x81EB008881EB0087,
-    0x81EB008A81EB0089, 0x81EB008C81EB008B, 0x81EB008E81EB008D,
-    0x81EB009081EB008F, 0x81EB009281EB0091, 0x81EB009481EB0093,
-    0x81EB009681EB0095, 0x81EB009881EB0097, 0x81EB009A81EB0099,
-    0x81EB009C81EB009B, 0x81EB009E81EB009D, 0x81EB00A081EB009F,
-    0x81EB00A281EB00A1, 0x81EB00A481EB00A3, 0x81EB00A681EB00A5,
-    0x81EB00A881EB00A7, 0x81EB00AA81EB00A9, 0x81EB00AC81EB00AB,
-    0x81EB00AE81EB00AD, 0x81EB00B081EB00AF, 0x81EB00B281EB00B1,
-    0x81EB00B481EB00B3, 0x81EB00B681EB00B5, 0x81EB00B881EB00B7,
-    0x81EB00BA81EB00B9, 0x81EB00BC81EB00BB, 0x81EB00BE81EB00BD,
-    0x82EB008082EB00BF, 0x82EB008282EB0081, 0x82EB008482EB0083,
-    0x82EB008682EB0085, 0x82EB008882EB0087, 0x82EB008A82EB0089,
-    0x82EB008C82EB008B, 0x82EB008E82EB008D, 0x82EB009082EB008F,
-    0x82EB009282EB0091, 0x82EB009482EB0093, 0x82EB009682EB0095,
-    0x82EB009882EB0097, 0x82EB009A82EB0099, 0x82EB009C82EB009B,
-    0x82EB009E82EB009D, 0x82EB00A082EB009F, 0x82EB00A282EB00A1,
-    0x82EB00A482EB00A3, 0x82EB00A682EB00A5, 0x82EB00A882EB00A7,
-    0x82EB00AA82EB00A9, 0x82EB00AC82EB00AB, 0x82EB00AE82EB00AD,
-    0x82EB00B082EB00AF, 0x82EB00B282EB00B1, 0x82EB00B482EB00B3,
-    0x82EB00B682EB00B5, 0x82EB00B882EB00B7, 0x82EB00BA82EB00B9,
-    0x82EB00BC82EB00BB, 0x82EB00BE82EB00BD, 0x83EB008083EB00BF,
-    0x83EB008283EB0081, 0x83EB008483EB0083, 0x83EB008683EB0085,
-    0x83EB008883EB0087, 0x83EB008A83EB0089, 0x83EB008C83EB008B,
-    0x83EB008E83EB008D, 0x83EB009083EB008F, 0x83EB009283EB0091,
-    0x83EB009483EB0093, 0x83EB009683EB0095, 0x83EB009883EB0097,
-    0x83EB009A83EB0099, 0x83EB009C83EB009B, 0x83EB009E83EB009D,
-    0x83EB00A083EB009F, 0x83EB00A283EB00A1, 0x83EB00A483EB00A3,
-    0x83EB00A683EB00A5, 0x83EB00A883EB00A7, 0x83EB00AA83EB00A9,
-    0x83EB00AC83EB00AB, 0x83EB00AE83EB00AD, 0x83EB00B083EB00AF,
-    0x83EB00B283EB00B1, 0x83EB00B483EB00B3, 0x83EB00B683EB00B5,
-    0x83EB00B883EB00B7, 0x83EB00BA83EB00B9, 0x83EB00BC83EB00BB,
-    0x83EB00BE83EB00BD, 0x84EB008084EB00BF, 0x84EB008284EB0081,
-    0x84EB008484EB0083, 0x84EB008684EB0085, 0x84EB008884EB0087,
-    0x84EB008A84EB0089, 0x84EB008C84EB008B, 0x84EB008E84EB008D,
-    0x84EB009084EB008F, 0x84EB009284EB0091, 0x84EB009484EB0093,
-    0x84EB009684EB0095, 0x84EB009884EB0097, 0x84EB009A84EB0099,
-    0x84EB009C84EB009B, 0x84EB009E84EB009D, 0x84EB00A084EB009F,
-    0x84EB00A284EB00A1, 0x84EB00A484EB00A3, 0x84EB00A684EB00A5,
-    0x84EB00A884EB00A7, 0x84EB00AA84EB00A9, 0x84EB00AC84EB00AB,
-    0x84EB00AE84EB00AD, 0x84EB00B084EB00AF, 0x84EB00B284EB00B1,
-    0x84EB00B484EB00B3, 0x84EB00B684EB00B5, 0x84EB00B884EB00B7,
-    0x84EB00BA84EB00B9, 0x84EB00BC84EB00BB, 0x84EB00BE84EB00BD,
-    0x85EB008085EB00BF, 0x85EB008285EB0081, 0x85EB008485EB0083,
-    0x85EB008685EB0085, 0x85EB008885EB0087, 0x85EB008A85EB0089,
-    0x85EB008C85EB008B, 0x85EB008E85EB008D, 0x85EB009085EB008F,
-    0x85EB009285EB0091, 0x85EB009485EB0093, 0x85EB009685EB0095,
-    0x85EB009885EB0097, 0x85EB009A85EB0099, 0x85EB009C85EB009B,
-    0x85EB009E85EB009D, 0x85EB00A085EB009F, 0x85EB00A285EB00A1,
-    0x85EB00A485EB00A3, 0x85EB00A685EB00A5, 0x85EB00A885EB00A7,
-    0x85EB00AA85EB00A9, 0x85EB00AC85EB00AB, 0x85EB00AE85EB00AD,
-    0x85EB00B085EB00AF, 0x85EB00B285EB00B1, 0x85EB00B485EB00B3,
-    0x85EB00B685EB00B5, 0x85EB00B885EB00B7, 0x85EB00BA85EB00B9,
-    0x85EB00BC85EB00BB, 0x85EB00BE85EB00BD, 0x86EB008086EB00BF,
-    0x86EB008286EB0081, 0x86EB008486EB0083, 0x86EB008686EB0085,
-    0x86EB008886EB0087, 0x86EB008A86EB0089, 0x86EB008C86EB008B,
-    0x86EB008E86EB008D, 0x86EB009086EB008F, 0x86EB009286EB0091,
-    0x86EB009486EB0093, 0x86EB009686EB0095, 0x86EB009886EB0097,
-    0x86EB009A86EB0099, 0x86EB009C86EB009B, 0x86EB009E86EB009D,
-    0x86EB00A086EB009F, 0x86EB00A286EB00A1, 0x86EB00A486EB00A3,
-    0x86EB00A686EB00A5, 0x86EB00A886EB00A7, 0x86EB00AA86EB00A9,
-    0x86EB00AC86EB00AB, 0x86EB00AE86EB00AD, 0x86EB00B086EB00AF,
-    0x86EB00B286EB00B1, 0x86EB00B486EB00B3, 0x86EB00B686EB00B5,
-    0x86EB00B886EB00B7, 0x86EB00BA86EB00B9, 0x86EB00BC86EB00BB,
-    0x86EB00BE86EB00BD, 0x87EB008087EB00BF, 0x87EB008287EB0081,
-    0x87EB008487EB0083, 0x87EB008687EB0085, 0x87EB008887EB0087,
-    0x87EB008A87EB0089, 0x87EB008C87EB008B, 0x87EB008E87EB008D,
-    0x87EB009087EB008F, 0x87EB009287EB0091, 0x87EB009487EB0093,
-    0x87EB009687EB0095, 0x87EB009887EB0097, 0x87EB009A87EB0099,
-    0x87EB009C87EB009B, 0x87EB009E87EB009D, 0x87EB00A087EB009F,
-    0x87EB00A287EB00A1, 0x87EB00A487EB00A3, 0x87EB00A687EB00A5,
-    0x87EB00A887EB00A7, 0x87EB00AA87EB00A9, 0x87EB00AC87EB00AB,
-    0x87EB00AE87EB00AD, 0x87EB00B087EB00AF, 0x87EB00B287EB00B1,
-    0x87EB00B487EB00B3, 0x87EB00B687EB00B5, 0x87EB00B887EB00B7,
-    0x87EB00BA87EB00B9, 0x87EB00BC87EB00BB, 0x87EB00BE87EB00BD,
-    0x88EB008088EB00BF, 0x88EB008288EB0081, 0x88EB008488EB0083,
-    0x88EB008688EB0085, 0x88EB008888EB0087, 0x88EB008A88EB0089,
-    0x88EB008C88EB008B, 0x88EB008E88EB008D, 0x88EB009088EB008F,
-    0x88EB009288EB0091, 0x88EB009488EB0093, 0x88EB009688EB0095,
-    0x88EB009888EB0097, 0x88EB009A88EB0099, 0x88EB009C88EB009B,
-    0x88EB009E88EB009D, 0x88EB00A088EB009F, 0x88EB00A288EB00A1,
-    0x88EB00A488EB00A3, 0x88EB00A688EB00A5, 0x88EB00A888EB00A7,
-    0x88EB00AA88EB00A9, 0x88EB00AC88EB00AB, 0x88EB00AE88EB00AD,
-    0x88EB00B088EB00AF, 0x88EB00B288EB00B1, 0x88EB00B488EB00B3,
-    0x88EB00B688EB00B5, 0x88EB00B888EB00B7, 0x88EB00BA88EB00B9,
-    0x88EB00BC88EB00BB, 0x88EB00BE88EB00BD, 0x89EB008089EB00BF,
-    0x89EB008289EB0081, 0x89EB008489EB0083, 0x89EB008689EB0085,
-    0x89EB008889EB0087, 0x89EB008A89EB0089, 0x89EB008C89EB008B,
-    0x89EB008E89EB008D, 0x89EB009089EB008F, 0x89EB009289EB0091,
-    0x89EB009489EB0093, 0x89EB009689EB0095, 0x89EB009889EB0097,
-    0x89EB009A89EB0099, 0x89EB009C89EB009B, 0x89EB009E89EB009D,
-    0x89EB00A089EB009F, 0x89EB00A289EB00A1, 0x89EB00A489EB00A3,
-    0x89EB00A689EB00A5, 0x89EB00A889EB00A7, 0x89EB00AA89EB00A9,
-    0x89EB00AC89EB00AB, 0x89EB00AE89EB00AD, 0x89EB00B089EB00AF,
-    0x89EB00B289EB00B1, 0x89EB00B489EB00B3, 0x89EB00B689EB00B5,
-    0x89EB00B889EB00B7, 0x89EB00BA89EB00B9, 0x89EB00BC89EB00BB,
-    0x89EB00BE89EB00BD, 0x8AEB00808AEB00BF, 0x8AEB00828AEB0081,
-    0x8AEB00848AEB0083, 0x8AEB00868AEB0085, 0x8AEB00888AEB0087,
-    0x8AEB008A8AEB0089, 0x8AEB008C8AEB008B, 0x8AEB008E8AEB008D,
-    0x8AEB00908AEB008F, 0x8AEB00928AEB0091, 0x8AEB00948AEB0093,
-    0x8AEB00968AEB0095, 0x8AEB00988AEB0097, 0x8AEB009A8AEB0099,
-    0x8AEB009C8AEB009B, 0x8AEB009E8AEB009D, 0x8AEB00A08AEB009F,
-    0x8AEB00A28AEB00A1, 0x8AEB00A48AEB00A3, 0x8AEB00A68AEB00A5,
-    0x8AEB00A88AEB00A7, 0x8AEB00AA8AEB00A9, 0x8AEB00AC8AEB00AB,
-    0x8AEB00AE8AEB00AD, 0x8AEB00B08AEB00AF, 0x8AEB00B28AEB00B1,
-    0x8AEB00B48AEB00B3, 0x8AEB00B68AEB00B5, 0x8AEB00B88AEB00B7,
-    0x8AEB00BA8AEB00B9, 0x8AEB00BC8AEB00BB, 0x8AEB00BE8AEB00BD,
-    0x8BEB00808BEB00BF, 0x8BEB00828BEB0081, 0x8BEB00848BEB0083,
-    0x8BEB00868BEB0085, 0x8BEB00888BEB0087, 0x8BEB008A8BEB0089,
-    0x8BEB008C8BEB008B, 0x8BEB008E8BEB008D, 0x8BEB00908BEB008F,
-    0x8BEB00928BEB0091, 0x8BEB00948BEB0093, 0x8BEB00968BEB0095,
-    0x8BEB00988BEB0097, 0x8BEB009A8BEB0099, 0x8BEB009C8BEB009B,
-    0x8BEB009E8BEB009D, 0x8BEB00A08BEB009F, 0x8BEB00A28BEB00A1,
-    0x8BEB00A48BEB00A3, 0x8BEB00A68BEB00A5, 0x8BEB00A88BEB00A7,
-    0x8BEB00AA8BEB00A9, 0x8BEB00AC8BEB00AB, 0x8BEB00AE8BEB00AD,
-    0x8BEB00B08BEB00AF, 0x8BEB00B28BEB00B1, 0x8BEB00B48BEB00B3,
-    0x8BEB00B68BEB00B5, 0x8BEB00B88BEB00B7, 0x8BEB00BA8BEB00B9,
-    0x8BEB00BC8BEB00BB, 0x8BEB00BE8BEB00BD, 0x8CEB00808CEB00BF,
-    0x8CEB00828CEB0081, 0x8CEB00848CEB0083, 0x8CEB00868CEB0085,
-    0x8CEB00888CEB0087, 0x8CEB008A8CEB0089, 0x8CEB008C8CEB008B,
-    0x8CEB008E8CEB008D, 0x8CEB00908CEB008F, 0x8CEB00928CEB0091,
-    0x8CEB00948CEB0093, 0x8CEB00968CEB0095, 0x8CEB00988CEB0097,
-    0x8CEB009A8CEB0099, 0x8CEB009C8CEB009B, 0x8CEB009E8CEB009D,
-    0x8CEB00A08CEB009F, 0x8CEB00A28CEB00A1, 0x8CEB00A48CEB00A3,
-    0x8CEB00A68CEB00A5, 0x8CEB00A88CEB00A7, 0x8CEB00AA8CEB00A9,
-    0x8CEB00AC8CEB00AB, 0x8CEB00AE8CEB00AD, 0x8CEB00B08CEB00AF,
-    0x8CEB00B28CEB00B1, 0x8CEB00B48CEB00B3, 0x8CEB00B68CEB00B5,
-    0x8CEB00B88CEB00B7, 0x8CEB00BA8CEB00B9, 0x8CEB00BC8CEB00BB,
-    0x8CEB00BE8CEB00BD, 0x8DEB00808DEB00BF, 0x8DEB00828DEB0081,
-    0x8DEB00848DEB0083, 0x8DEB00868DEB0085, 0x8DEB00888DEB0087,
-    0x8DEB008A8DEB0089, 0x8DEB008C8DEB008B, 0x8DEB008E8DEB008D,
-    0x8DEB00908DEB008F, 0x8DEB00928DEB0091, 0x8DEB00948DEB0093,
-    0x8DEB00968DEB0095, 0x8DEB00988DEB0097, 0x8DEB009A8DEB0099,
-    0x8DEB009C8DEB009B, 0x8DEB009E8DEB009D, 0x8DEB00A08DEB009F,
-    0x8DEB00A28DEB00A1, 0x8DEB00A48DEB00A3, 0x8DEB00A68DEB00A5,
-    0x8DEB00A88DEB00A7, 0x8DEB00AA8DEB00A9, 0x8DEB00AC8DEB00AB,
-    0x8DEB00AE8DEB00AD, 0x8DEB00B08DEB00AF, 0x8DEB00B28DEB00B1,
-    0x8DEB00B48DEB00B3, 0x8DEB00B68DEB00B5, 0x8DEB00B88DEB00B7,
-    0x8DEB00BA8DEB00B9, 0x8DEB00BC8DEB00BB, 0x8DEB00BE8DEB00BD,
-    0x8EEB00808EEB00BF, 0x8EEB00828EEB0081, 0x8EEB00848EEB0083,
-    0x8EEB00868EEB0085, 0x8EEB00888EEB0087, 0x8EEB008A8EEB0089,
-    0x8EEB008C8EEB008B, 0x8EEB008E8EEB008D, 0x8EEB00908EEB008F,
-    0x8EEB00928EEB0091, 0x8EEB00948EEB0093, 0x8EEB00968EEB0095,
-    0x8EEB00988EEB0097, 0x8EEB009A8EEB0099, 0x8EEB009C8EEB009B,
-    0x8EEB009E8EEB009D, 0x8EEB00A08EEB009F, 0x8EEB00A28EEB00A1,
-    0x8EEB00A48EEB00A3, 0x8EEB00A68EEB00A5, 0x8EEB00A88EEB00A7,
-    0x8EEB00AA8EEB00A9, 0x8EEB00AC8EEB00AB, 0x8EEB00AE8EEB00AD,
-    0x8EEB00B08EEB00AF, 0x8EEB00B28EEB00B1, 0x8EEB00B48EEB00B3,
-    0x8EEB00B68EEB00B5, 0x8EEB00B88EEB00B7, 0x8EEB00BA8EEB00B9,
-    0x8EEB00BC8EEB00BB, 0x8EEB00BE8EEB00BD, 0x8FEB00808FEB00BF,
-    0x8FEB00828FEB0081, 0x8FEB00848FEB0083, 0x8FEB00868FEB0085,
-    0x8FEB00888FEB0087, 0x8FEB008A8FEB0089, 0x8FEB008C8FEB008B,
-    0x8FEB008E8FEB008D, 0x8FEB00908FEB008F, 0x8FEB00928FEB0091,
-    0x8FEB00948FEB0093, 0x8FEB00968FEB0095, 0x8FEB00988FEB0097,
-    0x8FEB009A8FEB0099, 0x8FEB009C8FEB009B, 0x8FEB009E8FEB009D,
-    0x8FEB00A08FEB009F, 0x8FEB00A28FEB00A1, 0x8FEB00A48FEB00A3,
-    0x8FEB00A68FEB00A5, 0x8FEB00A88FEB00A7, 0x8FEB00AA8FEB00A9,
-    0x8FEB00AC8FEB00AB, 0x8FEB00AE8FEB00AD, 0x8FEB00B08FEB00AF,
-    0x8FEB00B28FEB00B1, 0x8FEB00B48FEB00B3, 0x8FEB00B68FEB00B5,
-    0x8FEB00B88FEB00B7, 0x8FEB00BA8FEB00B9, 0x8FEB00BC8FEB00BB,
-    0x8FEB00BE8FEB00BD, 0x90EB008090EB00BF, 0x90EB008290EB0081,
-    0x90EB008490EB0083, 0x90EB008690EB0085, 0x90EB008890EB0087,
-    0x90EB008A90EB0089, 0x90EB008C90EB008B, 0x90EB008E90EB008D,
-    0x90EB009090EB008F, 0x90EB009290EB0091, 0x90EB009490EB0093,
-    0x90EB009690EB0095, 0x90EB009890EB0097, 0x90EB009A90EB0099,
-    0x90EB009C90EB009B, 0x90EB009E90EB009D, 0x90EB00A090EB009F,
-    0x90EB00A290EB00A1, 0x90EB00A490EB00A3, 0x90EB00A690EB00A5,
-    0x90EB00A890EB00A7, 0x90EB00AA90EB00A9, 0x90EB00AC90EB00AB,
-    0x90EB00AE90EB00AD, 0x90EB00B090EB00AF, 0x90EB00B290EB00B1,
-    0x90EB00B490EB00B3, 0x90EB00B690EB00B5, 0x90EB00B890EB00B7,
-    0x90EB00BA90EB00B9, 0x90EB00BC90EB00BB, 0x90EB00BE90EB00BD,
-    0x91EB008091EB00BF, 0x91EB008291EB0081, 0x91EB008491EB0083,
-    0x91EB008691EB0085, 0x91EB008891EB0087, 0x91EB008A91EB0089,
-    0x91EB008C91EB008B, 0x91EB008E91EB008D, 0x91EB009091EB008F,
-    0x91EB009291EB0091, 0x91EB009491EB0093, 0x91EB009691EB0095,
-    0x91EB009891EB0097, 0x91EB009A91EB0099, 0x91EB009C91EB009B,
-    0x91EB009E91EB009D, 0x91EB00A091EB009F, 0x91EB00A291EB00A1,
-    0x91EB00A491EB00A3, 0x91EB00A691EB00A5, 0x91EB00A891EB00A7,
-    0x91EB00AA91EB00A9, 0x91EB00AC91EB00AB, 0x91EB00AE91EB00AD,
-    0x91EB00B091EB00AF, 0x91EB00B291EB00B1, 0x91EB00B491EB00B3,
-    0x91EB00B691EB00B5, 0x91EB00B891EB00B7, 0x91EB00BA91EB00B9,
-    0x91EB00BC91EB00BB, 0x91EB00BE91EB00BD, 0x92EB008092EB00BF,
-    0x92EB008292EB0081, 0x92EB008492EB0083, 0x92EB008692EB0085,
-    0x92EB008892EB0087, 0x92EB008A92EB0089, 0x92EB008C92EB008B,
-    0x92EB008E92EB008D, 0x92EB009092EB008F, 0x92EB009292EB0091,
-    0x92EB009492EB0093, 0x92EB009692EB0095, 0x92EB009892EB0097,
-    0x92EB009A92EB0099, 0x92EB009C92EB009B, 0x92EB009E92EB009D,
-    0x92EB00A092EB009F, 0x92EB00A292EB00A1, 0x92EB00A492EB00A3,
-    0x92EB00A692EB00A5, 0x92EB00A892EB00A7, 0x92EB00AA92EB00A9,
-    0x92EB00AC92EB00AB, 0x92EB00AE92EB00AD, 0x92EB00B092EB00AF,
-    0x92EB00B292EB00B1, 0x92EB00B492EB00B3, 0x92EB00B692EB00B5,
-    0x92EB00B892EB00B7, 0x92EB00BA92EB00B9, 0x92EB00BC92EB00BB,
-    0x92EB00BE92EB00BD, 0x93EB008093EB00BF, 0x93EB008293EB0081,
-    0x93EB008493EB0083, 0x93EB008693EB0085, 0x93EB008893EB0087,
-    0x93EB008A93EB0089, 0x93EB008C93EB008B, 0x93EB008E93EB008D,
-    0x93EB009093EB008F, 0x93EB009293EB0091, 0x93EB009493EB0093,
-    0x93EB009693EB0095, 0x93EB009893EB0097, 0x93EB009A93EB0099,
-    0x93EB009C93EB009B, 0x93EB009E93EB009D, 0x93EB00A093EB009F,
-    0x93EB00A293EB00A1, 0x93EB00A493EB00A3, 0x93EB00A693EB00A5,
-    0x93EB00A893EB00A7, 0x93EB00AA93EB00A9, 0x93EB00AC93EB00AB,
-    0x93EB00AE93EB00AD, 0x93EB00B093EB00AF, 0x93EB00B293EB00B1,
-    0x93EB00B493EB00B3, 0x93EB00B693EB00B5, 0x93EB00B893EB00B7,
-    0x93EB00BA93EB00B9, 0x93EB00BC93EB00BB, 0x93EB00BE93EB00BD,
-    0x94EB008094EB00BF, 0x94EB008294EB0081, 0x94EB008494EB0083,
-    0x94EB008694EB0085, 0x94EB008894EB0087, 0x94EB008A94EB0089,
-    0x94EB008C94EB008B, 0x94EB008E94EB008D, 0x94EB009094EB008F,
-    0x94EB009294EB0091, 0x94EB009494EB0093, 0x94EB009694EB0095,
-    0x94EB009894EB0097, 0x94EB009A94EB0099, 0x94EB009C94EB009B,
-    0x94EB009E94EB009D, 0x94EB00A094EB009F, 0x94EB00A294EB00A1,
-    0x94EB00A494EB00A3, 0x94EB00A694EB00A5, 0x94EB00A894EB00A7,
-    0x94EB00AA94EB00A9, 0x94EB00AC94EB00AB, 0x94EB00AE94EB00AD,
-    0x94EB00B094EB00AF, 0x94EB00B294EB00B1, 0x94EB00B494EB00B3,
-    0x94EB00B694EB00B5, 0x94EB00B894EB00B7, 0x94EB00BA94EB00B9,
-    0x94EB00BC94EB00BB, 0x94EB00BE94EB00BD, 0x95EB008095EB00BF,
-    0x95EB008295EB0081, 0x95EB008495EB0083, 0x95EB008695EB0085,
-    0x95EB008895EB0087, 0x95EB008A95EB0089, 0x95EB008C95EB008B,
-    0x95EB008E95EB008D, 0x95EB009095EB008F, 0x95EB009295EB0091,
-    0x95EB009495EB0093, 0x95EB009695EB0095, 0x95EB009895EB0097,
-    0x95EB009A95EB0099, 0x95EB009C95EB009B, 0x95EB009E95EB009D,
-    0x95EB00A095EB009F, 0x95EB00A295EB00A1, 0x95EB00A495EB00A3,
-    0x95EB00A695EB00A5, 0x95EB00A895EB00A7, 0x95EB00AA95EB00A9,
-    0x95EB00AC95EB00AB, 0x95EB00AE95EB00AD, 0x95EB00B095EB00AF,
-    0x95EB00B295EB00B1, 0x95EB00B495EB00B3, 0x95EB00B695EB00B5,
-    0x95EB00B895EB00B7, 0x95EB00BA95EB00B9, 0x95EB00BC95EB00BB,
-    0x95EB00BE95EB00BD, 0x96EB008096EB00BF, 0x96EB008296EB0081,
-    0x96EB008496EB0083, 0x96EB008696EB0085, 0x96EB008896EB0087,
-    0x96EB008A96EB0089, 0x96EB008C96EB008B, 0x96EB008E96EB008D,
-    0x96EB009096EB008F, 0x96EB009296EB0091, 0x96EB009496EB0093,
-    0x96EB009696EB0095, 0x96EB009896EB0097, 0x96EB009A96EB0099,
-    0x96EB009C96EB009B, 0x96EB009E96EB009D, 0x96EB00A096EB009F,
-    0x96EB00A296EB00A1, 0x96EB00A496EB00A3, 0x96EB00A696EB00A5,
-    0x96EB00A896EB00A7, 0x96EB00AA96EB00A9, 0x96EB00AC96EB00AB,
-    0x96EB00AE96EB00AD, 0x96EB00B096EB00AF, 0x96EB00B296EB00B1,
-    0x96EB00B496EB00B3, 0x96EB00B696EB00B5, 0x96EB00B896EB00B7,
-    0x96EB00BA96EB00B9, 0x96EB00BC96EB00BB, 0x96EB00BE96EB00BD,
-    0x97EB008097EB00BF, 0x97EB008297EB0081, 0x97EB008497EB0083,
-    0x97EB008697EB0085, 0x97EB008897EB0087, 0x97EB008A97EB0089,
-    0x97EB008C97EB008B, 0x97EB008E97EB008D, 0x97EB009097EB008F,
-    0x97EB009297EB0091, 0x97EB009497EB0093, 0x97EB009697EB0095,
-    0x97EB009897EB0097, 0x97EB009A97EB0099, 0x97EB009C97EB009B,
-    0x97EB009E97EB009D, 0x97EB00A097EB009F, 0x97EB00A297EB00A1,
-    0x97EB00A497EB00A3, 0x97EB00A697EB00A5, 0x97EB00A897EB00A7,
-    0x97EB00AA97EB00A9, 0x97EB00AC97EB00AB, 0x97EB00AE97EB00AD,
-    0x97EB00B097EB00AF, 0x97EB00B297EB00B1, 0x97EB00B497EB00B3,
-    0x97EB00B697EB00B5, 0x97EB00B897EB00B7, 0x97EB00BA97EB00B9,
-    0x97EB00BC97EB00BB, 0x97EB00BE97EB00BD, 0x98EB008098EB00BF,
-    0x98EB008298EB0081, 0x98EB008498EB0083, 0x98EB008698EB0085,
-    0x98EB008898EB0087, 0x98EB008A98EB0089, 0x98EB008C98EB008B,
-    0x98EB008E98EB008D, 0x98EB009098EB008F, 0x98EB009298EB0091,
-    0x98EB009498EB0093, 0x98EB009698EB0095, 0x98EB009898EB0097,
-    0x98EB009A98EB0099, 0x98EB009C98EB009B, 0x98EB009E98EB009D,
-    0x98EB00A098EB009F, 0x98EB00A298EB00A1, 0x98EB00A498EB00A3,
-    0x98EB00A698EB00A5, 0x98EB00A898EB00A7, 0x98EB00AA98EB00A9,
-    0x98EB00AC98EB00AB, 0x98EB00AE98EB00AD, 0x98EB00B098EB00AF,
-    0x98EB00B298EB00B1, 0x98EB00B498EB00B3, 0x98EB00B698EB00B5,
-    0x98EB00B898EB00B7, 0x98EB00BA98EB00B9, 0x98EB00BC98EB00BB,
-    0x98EB00BE98EB00BD, 0x99EB008099EB00BF, 0x99EB008299EB0081,
-    0x99EB008499EB0083, 0x99EB008699EB0085, 0x99EB008899EB0087,
-    0x99EB008A99EB0089, 0x99EB008C99EB008B, 0x99EB008E99EB008D,
-    0x99EB009099EB008F, 0x99EB009299EB0091, 0x99EB009499EB0093,
-    0x99EB009699EB0095, 0x99EB009899EB0097, 0x99EB009A99EB0099,
-    0x99EB009C99EB009B, 0x99EB009E99EB009D, 0x99EB00A099EB009F,
-    0x99EB00A299EB00A1, 0x99EB00A499EB00A3, 0x99EB00A699EB00A5,
-    0x99EB00A899EB00A7, 0x99EB00AA99EB00A9, 0x99EB00AC99EB00AB,
-    0x99EB00AE99EB00AD, 0x99EB00B099EB00AF, 0x99EB00B299EB00B1,
-    0x99EB00B499EB00B3, 0x99EB00B699EB00B5, 0x99EB00B899EB00B7,
-    0x99EB00BA99EB00B9, 0x99EB00BC99EB00BB, 0x99EB00BE99EB00BD,
-    0x9AEB00809AEB00BF, 0x9AEB00829AEB0081, 0x9AEB00849AEB0083,
-    0x9AEB00869AEB0085, 0x9AEB00889AEB0087, 0x9AEB008A9AEB0089,
-    0x9AEB008C9AEB008B, 0x9AEB008E9AEB008D, 0x9AEB00909AEB008F,
-    0x9AEB00929AEB0091, 0x9AEB00949AEB0093, 0x9AEB00969AEB0095,
-    0x9AEB00989AEB0097, 0x9AEB009A9AEB0099, 0x9AEB009C9AEB009B,
-    0x9AEB009E9AEB009D, 0x9AEB00A09AEB009F, 0x9AEB00A29AEB00A1,
-    0x9AEB00A49AEB00A3, 0x9AEB00A69AEB00A5, 0x9AEB00A89AEB00A7,
-    0x9AEB00AA9AEB00A9, 0x9AEB00AC9AEB00AB, 0x9AEB00AE9AEB00AD,
-    0x9AEB00B09AEB00AF, 0x9AEB00B29AEB00B1, 0x9AEB00B49AEB00B3,
-    0x9AEB00B69AEB00B5, 0x9AEB00B89AEB00B7, 0x9AEB00BA9AEB00B9,
-    0x9AEB00BC9AEB00BB, 0x9AEB00BE9AEB00BD, 0x9BEB00809BEB00BF,
-    0x9BEB00829BEB0081, 0x9BEB00849BEB0083, 0x9BEB00869BEB0085,
-    0x9BEB00889BEB0087, 0x9BEB008A9BEB0089, 0x9BEB008C9BEB008B,
-    0x9BEB008E9BEB008D, 0x9BEB00909BEB008F, 0x9BEB00929BEB0091,
-    0x9BEB00949BEB0093, 0x9BEB00969BEB0095, 0x9BEB00989BEB0097,
-    0x9BEB009A9BEB0099, 0x9BEB009C9BEB009B, 0x9BEB009E9BEB009D,
-    0x9BEB00A09BEB009F, 0x9BEB00A29BEB00A1, 0x9BEB00A49BEB00A3,
-    0x9BEB00A69BEB00A5, 0x9BEB00A89BEB00A7, 0x9BEB00AA9BEB00A9,
-    0x9BEB00AC9BEB00AB, 0x9BEB00AE9BEB00AD, 0x9BEB00B09BEB00AF,
-    0x9BEB00B29BEB00B1, 0x9BEB00B49BEB00B3, 0x9BEB00B69BEB00B5,
-    0x9BEB00B89BEB00B7, 0x9BEB00BA9BEB00B9, 0x9BEB00BC9BEB00BB,
-    0x9BEB00BE9BEB00BD, 0x9CEB00809CEB00BF, 0x9CEB00829CEB0081,
-    0x9CEB00849CEB0083, 0x9CEB00869CEB0085, 0x9CEB00889CEB0087,
-    0x9CEB008A9CEB0089, 0x9CEB008C9CEB008B, 0x9CEB008E9CEB008D,
-    0x9CEB00909CEB008F, 0x9CEB00929CEB0091, 0x9CEB00949CEB0093,
-    0x9CEB00969CEB0095, 0x9CEB00989CEB0097, 0x9CEB009A9CEB0099,
-    0x9CEB009C9CEB009B, 0x9CEB009E9CEB009D, 0x9CEB00A09CEB009F,
-    0x9CEB00A29CEB00A1, 0x9CEB00A49CEB00A3, 0x9CEB00A69CEB00A5,
-    0x9CEB00A89CEB00A7, 0x9CEB00AA9CEB00A9, 0x9CEB00AC9CEB00AB,
-    0x9CEB00AE9CEB00AD, 0x9CEB00B09CEB00AF, 0x9CEB00B29CEB00B1,
-    0x9CEB00B49CEB00B3, 0x9CEB00B69CEB00B5, 0x9CEB00B89CEB00B7,
-    0x9CEB00BA9CEB00B9, 0x9CEB00BC9CEB00BB, 0x9CEB00BE9CEB00BD,
-    0x9DEB00809DEB00BF, 0x9DEB00829DEB0081, 0x9DEB00849DEB0083,
-    0x9DEB00869DEB0085, 0x9DEB00889DEB0087, 0x9DEB008A9DEB0089,
-    0x9DEB008C9DEB008B, 0x9DEB008E9DEB008D, 0x9DEB00909DEB008F,
-    0x9DEB00929DEB0091, 0x9DEB00949DEB0093, 0x9DEB00969DEB0095,
-    0x9DEB00989DEB0097, 0x9DEB009A9DEB0099, 0x9DEB009C9DEB009B,
-    0x9DEB009E9DEB009D, 0x9DEB00A09DEB009F, 0x9DEB00A29DEB00A1,
-    0x9DEB00A49DEB00A3, 0x9DEB00A69DEB00A5, 0x9DEB00A89DEB00A7,
-    0x9DEB00AA9DEB00A9, 0x9DEB00AC9DEB00AB, 0x9DEB00AE9DEB00AD,
-    0x9DEB00B09DEB00AF, 0x9DEB00B29DEB00B1, 0x9DEB00B49DEB00B3,
-    0x9DEB00B69DEB00B5, 0x9DEB00B89DEB00B7, 0x9DEB00BA9DEB00B9,
-    0x9DEB00BC9DEB00BB, 0x9DEB00BE9DEB00BD, 0x9EEB00809EEB00BF,
-    0x9EEB00829EEB0081, 0x9EEB00849EEB0083, 0x9EEB00869EEB0085,
-    0x9EEB00889EEB0087, 0x9EEB008A9EEB0089, 0x9EEB008C9EEB008B,
-    0x9EEB008E9EEB008D, 0x9EEB00909EEB008F, 0x9EEB00929EEB0091,
-    0x9EEB00949EEB0093, 0x9EEB00969EEB0095, 0x9EEB00989EEB0097,
-    0x9EEB009A9EEB0099, 0x9EEB009C9EEB009B, 0x9EEB009E9EEB009D,
-    0x9EEB00A09EEB009F, 0x9EEB00A29EEB00A1, 0x9EEB00A49EEB00A3,
-    0x9EEB00A69EEB00A5, 0x9EEB00A89EEB00A7, 0x9EEB00AA9EEB00A9,
-    0x9EEB00AC9EEB00AB, 0x9EEB00AE9EEB00AD, 0x9EEB00B09EEB00AF,
-    0x9EEB00B29EEB00B1, 0x9EEB00B49EEB00B3, 0x9EEB00B69EEB00B5,
-    0x9EEB00B89EEB00B7, 0x9EEB00BA9EEB00B9, 0x9EEB00BC9EEB00BB,
-    0x9EEB00BE9EEB00BD, 0x9FEB00809FEB00BF, 0x9FEB00829FEB0081,
-    0x9FEB00849FEB0083, 0x9FEB00869FEB0085, 0x9FEB00889FEB0087,
-    0x9FEB008A9FEB0089, 0x9FEB008C9FEB008B, 0x9FEB008E9FEB008D,
-    0x9FEB00909FEB008F, 0x9FEB00929FEB0091, 0x9FEB00949FEB0093,
-    0x9FEB00969FEB0095, 0x9FEB00989FEB0097, 0x9FEB009A9FEB0099,
-    0x9FEB009C9FEB009B, 0x9FEB009E9FEB009D, 0x9FEB00A09FEB009F,
-    0x9FEB00A29FEB00A1, 0x9FEB00A49FEB00A3, 0x9FEB00A69FEB00A5,
-    0x9FEB00A89FEB00A7, 0x9FEB00AA9FEB00A9, 0x9FEB00AC9FEB00AB,
-    0x9FEB00AE9FEB00AD, 0x9FEB00B09FEB00AF, 0x9FEB00B29FEB00B1,
-    0x9FEB00B49FEB00B3, 0x9FEB00B69FEB00B5, 0x9FEB00B89FEB00B7,
-    0x9FEB00BA9FEB00B9, 0x9FEB00BC9FEB00BB, 0x9FEB00BE9FEB00BD,
-    0xA0EB0080A0EB00BF, 0xA0EB0082A0EB0081, 0xA0EB0084A0EB0083,
-    0xA0EB0086A0EB0085, 0xA0EB0088A0EB0087, 0xA0EB008AA0EB0089,
-    0xA0EB008CA0EB008B, 0xA0EB008EA0EB008D, 0xA0EB0090A0EB008F,
-    0xA0EB0092A0EB0091, 0xA0EB0094A0EB0093, 0xA0EB0096A0EB0095,
-    0xA0EB0098A0EB0097, 0xA0EB009AA0EB0099, 0xA0EB009CA0EB009B,
-    0xA0EB009EA0EB009D, 0xA0EB00A0A0EB009F, 0xA0EB00A2A0EB00A1,
-    0xA0EB00A4A0EB00A3, 0xA0EB00A6A0EB00A5, 0xA0EB00A8A0EB00A7,
-    0xA0EB00AAA0EB00A9, 0xA0EB00ACA0EB00AB, 0xA0EB00AEA0EB00AD,
-    0xA0EB00B0A0EB00AF, 0xA0EB00B2A0EB00B1, 0xA0EB00B4A0EB00B3,
-    0xA0EB00B6A0EB00B5, 0xA0EB00B8A0EB00B7, 0xA0EB00BAA0EB00B9,
-    0xA0EB00BCA0EB00BB, 0xA0EB00BEA0EB00BD, 0xA1EB0080A1EB00BF,
-    0xA1EB0082A1EB0081, 0xA1EB0084A1EB0083, 0xA1EB0086A1EB0085,
-    0xA1EB0088A1EB0087, 0xA1EB008AA1EB0089, 0xA1EB008CA1EB008B,
-    0xA1EB008EA1EB008D, 0xA1EB0090A1EB008F, 0xA1EB0092A1EB0091,
-    0xA1EB0094A1EB0093, 0xA1EB0096A1EB0095, 0xA1EB0098A1EB0097,
-    0xA1EB009AA1EB0099, 0xA1EB009CA1EB009B, 0xA1EB009EA1EB009D,
-    0xA1EB00A0A1EB009F, 0xA1EB00A2A1EB00A1, 0xA1EB00A4A1EB00A3,
-    0xA1EB00A6A1EB00A5, 0xA1EB00A8A1EB00A7, 0xA1EB00AAA1EB00A9,
-    0xA1EB00ACA1EB00AB, 0xA1EB00AEA1EB00AD, 0xA1EB00B0A1EB00AF,
-    0xA1EB00B2A1EB00B1, 0xA1EB00B4A1EB00B3, 0xA1EB00B6A1EB00B5,
-    0xA1EB00B8A1EB00B7, 0xA1EB00BAA1EB00B9, 0xA1EB00BCA1EB00BB,
-    0xA1EB00BEA1EB00BD, 0xA2EB0080A2EB00BF, 0xA2EB0082A2EB0081,
-    0xA2EB0084A2EB0083, 0xA2EB0086A2EB0085, 0xA2EB0088A2EB0087,
-    0xA2EB008AA2EB0089, 0xA2EB008CA2EB008B, 0xA2EB008EA2EB008D,
-    0xA2EB0090A2EB008F, 0xA2EB0092A2EB0091, 0xA2EB0094A2EB0093,
-    0xA2EB0096A2EB0095, 0xA2EB0098A2EB0097, 0xA2EB009AA2EB0099,
-    0xA2EB009CA2EB009B, 0xA2EB009EA2EB009D, 0xA2EB00A0A2EB009F,
-    0xA2EB00A2A2EB00A1, 0xA2EB00A4A2EB00A3, 0xA2EB00A6A2EB00A5,
-    0xA2EB00A8A2EB00A7, 0xA2EB00AAA2EB00A9, 0xA2EB00ACA2EB00AB,
-    0xA2EB00AEA2EB00AD, 0xA2EB00B0A2EB00AF, 0xA2EB00B2A2EB00B1,
-    0xA2EB00B4A2EB00B3, 0xA2EB00B6A2EB00B5, 0xA2EB00B8A2EB00B7,
-    0xA2EB00BAA2EB00B9, 0xA2EB00BCA2EB00BB, 0xA2EB00BEA2EB00BD,
-    0xA3EB0080A3EB00BF, 0xA3EB0082A3EB0081, 0xA3EB0084A3EB0083,
-    0xA3EB0086A3EB0085, 0xA3EB0088A3EB0087, 0xA3EB008AA3EB0089,
-    0xA3EB008CA3EB008B, 0xA3EB008EA3EB008D, 0xA3EB0090A3EB008F,
-    0xA3EB0092A3EB0091, 0xA3EB0094A3EB0093, 0xA3EB0096A3EB0095,
-    0xA3EB0098A3EB0097, 0xA3EB009AA3EB0099, 0xA3EB009CA3EB009B,
-    0xA3EB009EA3EB009D, 0xA3EB00A0A3EB009F, 0xA3EB00A2A3EB00A1,
-    0xA3EB00A4A3EB00A3, 0xA3EB00A6A3EB00A5, 0xA3EB00A8A3EB00A7,
-    0xA3EB00AAA3EB00A9, 0xA3EB00ACA3EB00AB, 0xA3EB00AEA3EB00AD,
-    0xA3EB00B0A3EB00AF, 0xA3EB00B2A3EB00B1, 0xA3EB00B4A3EB00B3,
-    0xA3EB00B6A3EB00B5, 0xA3EB00B8A3EB00B7, 0xA3EB00BAA3EB00B9,
-    0xA3EB00BCA3EB00BB, 0xA3EB00BEA3EB00BD, 0xA4EB0080A4EB00BF,
-    0xA4EB0082A4EB0081, 0xA4EB0084A4EB0083, 0xA4EB0086A4EB0085,
-    0xA4EB0088A4EB0087, 0xA4EB008AA4EB0089, 0xA4EB008CA4EB008B,
-    0xA4EB008EA4EB008D, 0xA4EB0090A4EB008F, 0xA4EB0092A4EB0091,
-    0xA4EB0094A4EB0093, 0xA4EB0096A4EB0095, 0xA4EB0098A4EB0097,
-    0xA4EB009AA4EB0099, 0xA4EB009CA4EB009B, 0xA4EB009EA4EB009D,
-    0xA4EB00A0A4EB009F, 0xA4EB00A2A4EB00A1, 0xA4EB00A4A4EB00A3,
-    0xA4EB00A6A4EB00A5, 0xA4EB00A8A4EB00A7, 0xA4EB00AAA4EB00A9,
-    0xA4EB00ACA4EB00AB, 0xA4EB00AEA4EB00AD, 0xA4EB00B0A4EB00AF,
-    0xA4EB00B2A4EB00B1, 0xA4EB00B4A4EB00B3, 0xA4EB00B6A4EB00B5,
-    0xA4EB00B8A4EB00B7, 0xA4EB00BAA4EB00B9, 0xA4EB00BCA4EB00BB,
-    0xA4EB00BEA4EB00BD, 0xA5EB0080A5EB00BF, 0xA5EB0082A5EB0081,
-    0xA5EB0084A5EB0083, 0xA5EB0086A5EB0085, 0xA5EB0088A5EB0087,
-    0xA5EB008AA5EB0089, 0xA5EB008CA5EB008B, 0xA5EB008EA5EB008D,
-    0xA5EB0090A5EB008F, 0xA5EB0092A5EB0091, 0xA5EB0094A5EB0093,
-    0xA5EB0096A5EB0095, 0xA5EB0098A5EB0097, 0xA5EB009AA5EB0099,
-    0xA5EB009CA5EB009B, 0xA5EB009EA5EB009D, 0xA5EB00A0A5EB009F,
-    0xA5EB00A2A5EB00A1, 0xA5EB00A4A5EB00A3, 0xA5EB00A6A5EB00A5,
-    0xA5EB00A8A5EB00A7, 0xA5EB00AAA5EB00A9, 0xA5EB00ACA5EB00AB,
-    0xA5EB00AEA5EB00AD, 0xA5EB00B0A5EB00AF, 0xA5EB00B2A5EB00B1,
-    0xA5EB00B4A5EB00B3, 0xA5EB00B6A5EB00B5, 0xA5EB00B8A5EB00B7,
-    0xA5EB00BAA5EB00B9, 0xA5EB00BCA5EB00BB, 0xA5EB00BEA5EB00BD,
-    0xA6EB0080A6EB00BF, 0xA6EB0082A6EB0081, 0xA6EB0084A6EB0083,
-    0xA6EB0086A6EB0085, 0xA6EB0088A6EB0087, 0xA6EB008AA6EB0089,
-    0xA6EB008CA6EB008B, 0xA6EB008EA6EB008D, 0xA6EB0090A6EB008F,
-    0xA6EB0092A6EB0091, 0xA6EB0094A6EB0093, 0xA6EB0096A6EB0095,
-    0xA6EB0098A6EB0097, 0xA6EB009AA6EB0099, 0xA6EB009CA6EB009B,
-    0xA6EB009EA6EB009D, 0xA6EB00A0A6EB009F, 0xA6EB00A2A6EB00A1,
-    0xA6EB00A4A6EB00A3, 0xA6EB00A6A6EB00A5, 0xA6EB00A8A6EB00A7,
-    0xA6EB00AAA6EB00A9, 0xA6EB00ACA6EB00AB, 0xA6EB00AEA6EB00AD,
-    0xA6EB00B0A6EB00AF, 0xA6EB00B2A6EB00B1, 0xA6EB00B4A6EB00B3,
-    0xA6EB00B6A6EB00B5, 0xA6EB00B8A6EB00B7, 0xA6EB00BAA6EB00B9,
-    0xA6EB00BCA6EB00BB, 0xA6EB00BEA6EB00BD, 0xA7EB0080A7EB00BF,
-    0xA7EB0082A7EB0081, 0xA7EB0084A7EB0083, 0xA7EB0086A7EB0085,
-    0xA7EB0088A7EB0087, 0xA7EB008AA7EB0089, 0xA7EB008CA7EB008B,
-    0xA7EB008EA7EB008D, 0xA7EB0090A7EB008F, 0xA7EB0092A7EB0091,
-    0xA7EB0094A7EB0093, 0xA7EB0096A7EB0095, 0xA7EB0098A7EB0097,
-    0xA7EB009AA7EB0099, 0xA7EB009CA7EB009B, 0xA7EB009EA7EB009D,
-    0xA7EB00A0A7EB009F, 0xA7EB00A2A7EB00A1, 0xA7EB00A4A7EB00A3,
-    0xA7EB00A6A7EB00A5, 0xA7EB00A8A7EB00A7, 0xA7EB00AAA7EB00A9,
-    0xA7EB00ACA7EB00AB, 0xA7EB00AEA7EB00AD, 0xA7EB00B0A7EB00AF,
-    0xA7EB00B2A7EB00B1, 0xA7EB00B4A7EB00B3, 0xA7EB00B6A7EB00B5,
-    0xA7EB00B8A7EB00B7, 0xA7EB00BAA7EB00B9, 0xA7EB00BCA7EB00BB,
-    0xA7EB00BEA7EB00BD, 0xA8EB0080A8EB00BF, 0xA8EB0082A8EB0081,
-    0xA8EB0084A8EB0083, 0xA8EB0086A8EB0085, 0xA8EB0088A8EB0087,
-    0xA8EB008AA8EB0089, 0xA8EB008CA8EB008B, 0xA8EB008EA8EB008D,
-    0xA8EB0090A8EB008F, 0xA8EB0092A8EB0091, 0xA8EB0094A8EB0093,
-    0xA8EB0096A8EB0095, 0xA8EB0098A8EB0097, 0xA8EB009AA8EB0099,
-    0xA8EB009CA8EB009B, 0xA8EB009EA8EB009D, 0xA8EB00A0A8EB009F,
-    0xA8EB00A2A8EB00A1, 0xA8EB00A4A8EB00A3, 0xA8EB00A6A8EB00A5,
-    0xA8EB00A8A8EB00A7, 0xA8EB00AAA8EB00A9, 0xA8EB00ACA8EB00AB,
-    0xA8EB00AEA8EB00AD, 0xA8EB00B0A8EB00AF, 0xA8EB00B2A8EB00B1,
-    0xA8EB00B4A8EB00B3, 0xA8EB00B6A8EB00B5, 0xA8EB00B8A8EB00B7,
-    0xA8EB00BAA8EB00B9, 0xA8EB00BCA8EB00BB, 0xA8EB00BEA8EB00BD,
-    0xA9EB0080A9EB00BF, 0xA9EB0082A9EB0081, 0xA9EB0084A9EB0083,
-    0xA9EB0086A9EB0085, 0xA9EB0088A9EB0087, 0xA9EB008AA9EB0089,
-    0xA9EB008CA9EB008B, 0xA9EB008EA9EB008D, 0xA9EB0090A9EB008F,
-    0xA9EB0092A9EB0091, 0xA9EB0094A9EB0093, 0xA9EB0096A9EB0095,
-    0xA9EB0098A9EB0097, 0xA9EB009AA9EB0099, 0xA9EB009CA9EB009B,
-    0xA9EB009EA9EB009D, 0xA9EB00A0A9EB009F, 0xA9EB00A2A9EB00A1,
-    0xA9EB00A4A9EB00A3, 0xA9EB00A6A9EB00A5, 0xA9EB00A8A9EB00A7,
-    0xA9EB00AAA9EB00A9, 0xA9EB00ACA9EB00AB, 0xA9EB00AEA9EB00AD,
-    0xA9EB00B0A9EB00AF, 0xA9EB00B2A9EB00B1, 0xA9EB00B4A9EB00B3,
-    0xA9EB00B6A9EB00B5, 0xA9EB00B8A9EB00B7, 0xA9EB00BAA9EB00B9,
-    0xA9EB00BCA9EB00BB, 0xA9EB00BEA9EB00BD, 0xAAEB0080AAEB00BF,
-    0xAAEB0082AAEB0081, 0xAAEB0084AAEB0083, 0xAAEB0086AAEB0085,
-    0xAAEB0088AAEB0087, 0xAAEB008AAAEB0089, 0xAAEB008CAAEB008B,
-    0xAAEB008EAAEB008D, 0xAAEB0090AAEB008F, 0xAAEB0092AAEB0091,
-    0xAAEB0094AAEB0093, 0xAAEB0096AAEB0095, 0xAAEB0098AAEB0097,
-    0xAAEB009AAAEB0099, 0xAAEB009CAAEB009B, 0xAAEB009EAAEB009D,
-    0xAAEB00A0AAEB009F, 0xAAEB00A2AAEB00A1, 0xAAEB00A4AAEB00A3,
-    0xAAEB00A6AAEB00A5, 0xAAEB00A8AAEB00A7, 0xAAEB00AAAAEB00A9,
-    0xAAEB00ACAAEB00AB, 0xAAEB00AEAAEB00AD, 0xAAEB00B0AAEB00AF,
-    0xAAEB00B2AAEB00B1, 0xAAEB00B4AAEB00B3, 0xAAEB00B6AAEB00B5,
-    0xAAEB00B8AAEB00B7, 0xAAEB00BAAAEB00B9, 0xAAEB00BCAAEB00BB,
-    0xAAEB00BEAAEB00BD, 0xABEB0080ABEB00BF, 0xABEB0082ABEB0081,
-    0xABEB0084ABEB0083, 0xABEB0086ABEB0085, 0xABEB0088ABEB0087,
-    0xABEB008AABEB0089, 0xABEB008CABEB008B, 0xABEB008EABEB008D,
-    0xABEB0090ABEB008F, 0xABEB0092ABEB0091, 0xABEB0094ABEB0093,
-    0xABEB0096ABEB0095, 0xABEB0098ABEB0097, 0xABEB009AABEB0099,
-    0xABEB009CABEB009B, 0xABEB009EABEB009D, 0xABEB00A0ABEB009F,
-    0xABEB00A2ABEB00A1, 0xABEB00A4ABEB00A3, 0xABEB00A6ABEB00A5,
-    0xABEB00A8ABEB00A7, 0xABEB00AAABEB00A9, 0xABEB00ACABEB00AB,
-    0xABEB00AEABEB00AD, 0xABEB00B0ABEB00AF, 0xABEB00B2ABEB00B1,
-    0xABEB00B4ABEB00B3, 0xABEB00B6ABEB00B5, 0xABEB00B8ABEB00B7,
-    0xABEB00BAABEB00B9, 0xABEB00BCABEB00BB, 0xABEB00BEABEB00BD,
-    0xACEB0080ACEB00BF, 0xACEB0082ACEB0081, 0xACEB0084ACEB0083,
-    0xACEB0086ACEB0085, 0xACEB0088ACEB0087, 0xACEB008AACEB0089,
-    0xACEB008CACEB008B, 0xACEB008EACEB008D, 0xACEB0090ACEB008F,
-    0xACEB0092ACEB0091, 0xACEB0094ACEB0093, 0xACEB0096ACEB0095,
-    0xACEB0098ACEB0097, 0xACEB009AACEB0099, 0xACEB009CACEB009B,
-    0xACEB009EACEB009D, 0xACEB00A0ACEB009F, 0xACEB00A2ACEB00A1,
-    0xACEB00A4ACEB00A3, 0xACEB00A6ACEB00A5, 0xACEB00A8ACEB00A7,
-    0xACEB00AAACEB00A9, 0xACEB00ACACEB00AB, 0xACEB00AEACEB00AD,
-    0xACEB00B0ACEB00AF, 0xACEB00B2ACEB00B1, 0xACEB00B4ACEB00B3,
-    0xACEB00B6ACEB00B5, 0xACEB00B8ACEB00B7, 0xACEB00BAACEB00B9,
-    0xACEB00BCACEB00BB, 0xACEB00BEACEB00BD, 0xADEB0080ADEB00BF,
-    0xADEB0082ADEB0081, 0xADEB0084ADEB0083, 0xADEB0086ADEB0085,
-    0xADEB0088ADEB0087, 0xADEB008AADEB0089, 0xADEB008CADEB008B,
-    0xADEB008EADEB008D, 0xADEB0090ADEB008F, 0xADEB0092ADEB0091,
-    0xADEB0094ADEB0093, 0xADEB0096ADEB0095, 0xADEB0098ADEB0097,
-    0xADEB009AADEB0099, 0xADEB009CADEB009B, 0xADEB009EADEB009D,
-    0xADEB00A0ADEB009F, 0xADEB00A2ADEB00A1, 0xADEB00A4ADEB00A3,
-    0xADEB00A6ADEB00A5, 0xADEB00A8ADEB00A7, 0xADEB00AAADEB00A9,
-    0xADEB00ACADEB00AB, 0xADEB00AEADEB00AD, 0xADEB00B0ADEB00AF,
-    0xADEB00B2ADEB00B1, 0xADEB00B4ADEB00B3, 0xADEB00B6ADEB00B5,
-    0xADEB00B8ADEB00B7, 0xADEB00BAADEB00B9, 0xADEB00BCADEB00BB,
-    0xADEB00BEADEB00BD, 0xAEEB0080AEEB00BF, 0xAEEB0082AEEB0081,
-    0xAEEB0084AEEB0083, 0xAEEB0086AEEB0085, 0xAEEB0088AEEB0087,
-    0xAEEB008AAEEB0089, 0xAEEB008CAEEB008B, 0xAEEB008EAEEB008D,
-    0xAEEB0090AEEB008F, 0xAEEB0092AEEB0091, 0xAEEB0094AEEB0093,
-    0xAEEB0096AEEB0095, 0xAEEB0098AEEB0097, 0xAEEB009AAEEB0099,
-    0xAEEB009CAEEB009B, 0xAEEB009EAEEB009D, 0xAEEB00A0AEEB009F,
-    0xAEEB00A2AEEB00A1, 0xAEEB00A4AEEB00A3, 0xAEEB00A6AEEB00A5,
-    0xAEEB00A8AEEB00A7, 0xAEEB00AAAEEB00A9, 0xAEEB00ACAEEB00AB,
-    0xAEEB00AEAEEB00AD, 0xAEEB00B0AEEB00AF, 0xAEEB00B2AEEB00B1,
-    0xAEEB00B4AEEB00B3, 0xAEEB00B6AEEB00B5, 0xAEEB00B8AEEB00B7,
-    0xAEEB00BAAEEB00B9, 0xAEEB00BCAEEB00BB, 0xAEEB00BEAEEB00BD,
-    0xAFEB0080AFEB00BF, 0xAFEB0082AFEB0081, 0xAFEB0084AFEB0083,
-    0xAFEB0086AFEB0085, 0xAFEB0088AFEB0087, 0xAFEB008AAFEB0089,
-    0xAFEB008CAFEB008B, 0xAFEB008EAFEB008D, 0xAFEB0090AFEB008F,
-    0xAFEB0092AFEB0091, 0xAFEB0094AFEB0093, 0xAFEB0096AFEB0095,
-    0xAFEB0098AFEB0097, 0xAFEB009AAFEB0099, 0xAFEB009CAFEB009B,
-    0xAFEB009EAFEB009D, 0xAFEB00A0AFEB009F, 0xAFEB00A2AFEB00A1,
-    0xAFEB00A4AFEB00A3, 0xAFEB00A6AFEB00A5, 0xAFEB00A8AFEB00A7,
-    0xAFEB00AAAFEB00A9, 0xAFEB00ACAFEB00AB, 0xAFEB00AEAFEB00AD,
-    0xAFEB00B0AFEB00AF, 0xAFEB00B2AFEB00B1, 0xAFEB00B4AFEB00B3,
-    0xAFEB00B6AFEB00B5, 0xAFEB00B8AFEB00B7, 0xAFEB00BAAFEB00B9,
-    0xAFEB00BCAFEB00BB, 0xAFEB00BEAFEB00BD, 0xB0EB0080B0EB00BF,
-    0xB0EB0082B0EB0081, 0xB0EB0084B0EB0083, 0xB0EB0086B0EB0085,
-    0xB0EB0088B0EB0087, 0xB0EB008AB0EB0089, 0xB0EB008CB0EB008B,
-    0xB0EB008EB0EB008D, 0xB0EB0090B0EB008F, 0xB0EB0092B0EB0091,
-    0xB0EB0094B0EB0093, 0xB0EB0096B0EB0095, 0xB0EB0098B0EB0097,
-    0xB0EB009AB0EB0099, 0xB0EB009CB0EB009B, 0xB0EB009EB0EB009D,
-    0xB0EB00A0B0EB009F, 0xB0EB00A2B0EB00A1, 0xB0EB00A4B0EB00A3,
-    0xB0EB00A6B0EB00A5, 0xB0EB00A8B0EB00A7, 0xB0EB00AAB0EB00A9,
-    0xB0EB00ACB0EB00AB, 0xB0EB00AEB0EB00AD, 0xB0EB00B0B0EB00AF,
-    0xB0EB00B2B0EB00B1, 0xB0EB00B4B0EB00B3, 0xB0EB00B6B0EB00B5,
-    0xB0EB00B8B0EB00B7, 0xB0EB00BAB0EB00B9, 0xB0EB00BCB0EB00BB,
-    0xB0EB00BEB0EB00BD, 0xB1EB0080B1EB00BF, 0xB1EB0082B1EB0081,
-    0xB1EB0084B1EB0083, 0xB1EB0086B1EB0085, 0xB1EB0088B1EB0087,
-    0xB1EB008AB1EB0089, 0xB1EB008CB1EB008B, 0xB1EB008EB1EB008D,
-    0xB1EB0090B1EB008F, 0xB1EB0092B1EB0091, 0xB1EB0094B1EB0093,
-    0xB1EB0096B1EB0095, 0xB1EB0098B1EB0097, 0xB1EB009AB1EB0099,
-    0xB1EB009CB1EB009B, 0xB1EB009EB1EB009D, 0xB1EB00A0B1EB009F,
-    0xB1EB00A2B1EB00A1, 0xB1EB00A4B1EB00A3, 0xB1EB00A6B1EB00A5,
-    0xB1EB00A8B1EB00A7, 0xB1EB00AAB1EB00A9, 0xB1EB00ACB1EB00AB,
-    0xB1EB00AEB1EB00AD, 0xB1EB00B0B1EB00AF, 0xB1EB00B2B1EB00B1,
-    0xB1EB00B4B1EB00B3, 0xB1EB00B6B1EB00B5, 0xB1EB00B8B1EB00B7,
-    0xB1EB00BAB1EB00B9, 0xB1EB00BCB1EB00BB, 0xB1EB00BEB1EB00BD,
-    0xB2EB0080B2EB00BF, 0xB2EB0082B2EB0081, 0xB2EB0084B2EB0083,
-    0xB2EB0086B2EB0085, 0xB2EB0088B2EB0087, 0xB2EB008AB2EB0089,
-    0xB2EB008CB2EB008B, 0xB2EB008EB2EB008D, 0xB2EB0090B2EB008F,
-    0xB2EB0092B2EB0091, 0xB2EB0094B2EB0093, 0xB2EB0096B2EB0095,
-    0xB2EB0098B2EB0097, 0xB2EB009AB2EB0099, 0xB2EB009CB2EB009B,
-    0xB2EB009EB2EB009D, 0xB2EB00A0B2EB009F, 0xB2EB00A2B2EB00A1,
-    0xB2EB00A4B2EB00A3, 0xB2EB00A6B2EB00A5, 0xB2EB00A8B2EB00A7,
-    0xB2EB00AAB2EB00A9, 0xB2EB00ACB2EB00AB, 0xB2EB00AEB2EB00AD,
-    0xB2EB00B0B2EB00AF, 0xB2EB00B2B2EB00B1, 0xB2EB00B4B2EB00B3,
-    0xB2EB00B6B2EB00B5, 0xB2EB00B8B2EB00B7, 0xB2EB00BAB2EB00B9,
-    0xB2EB00BCB2EB00BB, 0xB2EB00BEB2EB00BD, 0xB3EB0080B3EB00BF,
-    0xB3EB0082B3EB0081, 0xB3EB0084B3EB0083, 0xB3EB0086B3EB0085,
-    0xB3EB0088B3EB0087, 0xB3EB008AB3EB0089, 0xB3EB008CB3EB008B,
-    0xB3EB008EB3EB008D, 0xB3EB0090B3EB008F, 0xB3EB0092B3EB0091,
-    0xB3EB0094B3EB0093, 0xB3EB0096B3EB0095, 0xB3EB0098B3EB0097,
-    0xB3EB009AB3EB0099, 0xB3EB009CB3EB009B, 0xB3EB009EB3EB009D,
-    0xB3EB00A0B3EB009F, 0xB3EB00A2B3EB00A1, 0xB3EB00A4B3EB00A3,
-    0xB3EB00A6B3EB00A5, 0xB3EB00A8B3EB00A7, 0xB3EB00AAB3EB00A9,
-    0xB3EB00ACB3EB00AB, 0xB3EB00AEB3EB00AD, 0xB3EB00B0B3EB00AF,
-    0xB3EB00B2B3EB00B1, 0xB3EB00B4B3EB00B3, 0xB3EB00B6B3EB00B5,
-    0xB3EB00B8B3EB00B7, 0xB3EB00BAB3EB00B9, 0xB3EB00BCB3EB00BB,
-    0xB3EB00BEB3EB00BD, 0xB4EB0080B4EB00BF, 0xB4EB0082B4EB0081,
-    0xB4EB0084B4EB0083, 0xB4EB0086B4EB0085, 0xB4EB0088B4EB0087,
-    0xB4EB008AB4EB0089, 0xB4EB008CB4EB008B, 0xB4EB008EB4EB008D,
-    0xB4EB0090B4EB008F, 0xB4EB0092B4EB0091, 0xB4EB0094B4EB0093,
-    0xB4EB0096B4EB0095, 0xB4EB0098B4EB0097, 0xB4EB009AB4EB0099,
-    0xB4EB009CB4EB009B, 0xB4EB009EB4EB009D, 0xB4EB00A0B4EB009F,
-    0xB4EB00A2B4EB00A1, 0xB4EB00A4B4EB00A3, 0xB4EB00A6B4EB00A5,
-    0xB4EB00A8B4EB00A7, 0xB4EB00AAB4EB00A9, 0xB4EB00ACB4EB00AB,
-    0xB4EB00AEB4EB00AD, 0xB4EB00B0B4EB00AF, 0xB4EB00B2B4EB00B1,
-    0xB4EB00B4B4EB00B3, 0xB4EB00B6B4EB00B5, 0xB4EB00B8B4EB00B7,
-    0xB4EB00BAB4EB00B9, 0xB4EB00BCB4EB00BB, 0xB4EB00BEB4EB00BD,
-    0xB5EB0080B5EB00BF, 0xB5EB0082B5EB0081, 0xB5EB0084B5EB0083,
-    0xB5EB0086B5EB0085, 0xB5EB0088B5EB0087, 0xB5EB008AB5EB0089,
-    0xB5EB008CB5EB008B, 0xB5EB008EB5EB008D, 0xB5EB0090B5EB008F,
-    0xB5EB0092B5EB0091, 0xB5EB0094B5EB0093, 0xB5EB0096B5EB0095,
-    0xB5EB0098B5EB0097, 0xB5EB009AB5EB0099, 0xB5EB009CB5EB009B,
-    0xB5EB009EB5EB009D, 0xB5EB00A0B5EB009F, 0xB5EB00A2B5EB00A1,
-    0xB5EB00A4B5EB00A3, 0xB5EB00A6B5EB00A5, 0xB5EB00A8B5EB00A7,
-    0xB5EB00AAB5EB00A9, 0xB5EB00ACB5EB00AB, 0xB5EB00AEB5EB00AD,
-    0xB5EB00B0B5EB00AF, 0xB5EB00B2B5EB00B1, 0xB5EB00B4B5EB00B3,
-    0xB5EB00B6B5EB00B5, 0xB5EB00B8B5EB00B7, 0xB5EB00BAB5EB00B9,
-    0xB5EB00BCB5EB00BB, 0xB5EB00BEB5EB00BD, 0xB6EB0080B6EB00BF,
-    0xB6EB0082B6EB0081, 0xB6EB0084B6EB0083, 0xB6EB0086B6EB0085,
-    0xB6EB0088B6EB0087, 0xB6EB008AB6EB0089, 0xB6EB008CB6EB008B,
-    0xB6EB008EB6EB008D, 0xB6EB0090B6EB008F, 0xB6EB0092B6EB0091,
-    0xB6EB0094B6EB0093, 0xB6EB0096B6EB0095, 0xB6EB0098B6EB0097,
-    0xB6EB009AB6EB0099, 0xB6EB009CB6EB009B, 0xB6EB009EB6EB009D,
-    0xB6EB00A0B6EB009F, 0xB6EB00A2B6EB00A1, 0xB6EB00A4B6EB00A3,
-    0xB6EB00A6B6EB00A5, 0xB6EB00A8B6EB00A7, 0xB6EB00AAB6EB00A9,
-    0xB6EB00ACB6EB00AB, 0xB6EB00AEB6EB00AD, 0xB6EB00B0B6EB00AF,
-    0xB6EB00B2B6EB00B1, 0xB6EB00B4B6EB00B3, 0xB6EB00B6B6EB00B5,
-    0xB6EB00B8B6EB00B7, 0xB6EB00BAB6EB00B9, 0xB6EB00BCB6EB00BB,
-    0xB6EB00BEB6EB00BD, 0xB7EB0080B7EB00BF, 0xB7EB0082B7EB0081,
-    0xB7EB0084B7EB0083, 0xB7EB0086B7EB0085, 0xB7EB0088B7EB0087,
-    0xB7EB008AB7EB0089, 0xB7EB008CB7EB008B, 0xB7EB008EB7EB008D,
-    0xB7EB0090B7EB008F, 0xB7EB0092B7EB0091, 0xB7EB0094B7EB0093,
-    0xB7EB0096B7EB0095, 0xB7EB0098B7EB0097, 0xB7EB009AB7EB0099,
-    0xB7EB009CB7EB009B, 0xB7EB009EB7EB009D, 0xB7EB00A0B7EB009F,
-    0xB7EB00A2B7EB00A1, 0xB7EB00A4B7EB00A3, 0xB7EB00A6B7EB00A5,
-    0xB7EB00A8B7EB00A7, 0xB7EB00AAB7EB00A9, 0xB7EB00ACB7EB00AB,
-    0xB7EB00AEB7EB00AD, 0xB7EB00B0B7EB00AF, 0xB7EB00B2B7EB00B1,
-    0xB7EB00B4B7EB00B3, 0xB7EB00B6B7EB00B5, 0xB7EB00B8B7EB00B7,
-    0xB7EB00BAB7EB00B9, 0xB7EB00BCB7EB00BB, 0xB7EB00BEB7EB00BD,
-    0xB8EB0080B8EB00BF, 0xB8EB0082B8EB0081, 0xB8EB0084B8EB0083,
-    0xB8EB0086B8EB0085, 0xB8EB0088B8EB0087, 0xB8EB008AB8EB0089,
-    0xB8EB008CB8EB008B, 0xB8EB008EB8EB008D, 0xB8EB0090B8EB008F,
-    0xB8EB0092B8EB0091, 0xB8EB0094B8EB0093, 0xB8EB0096B8EB0095,
-    0xB8EB0098B8EB0097, 0xB8EB009AB8EB0099, 0xB8EB009CB8EB009B,
-    0xB8EB009EB8EB009D, 0xB8EB00A0B8EB009F, 0xB8EB00A2B8EB00A1,
-    0xB8EB00A4B8EB00A3, 0xB8EB00A6B8EB00A5, 0xB8EB00A8B8EB00A7,
-    0xB8EB00AAB8EB00A9, 0xB8EB00ACB8EB00AB, 0xB8EB00AEB8EB00AD,
-    0xB8EB00B0B8EB00AF, 0xB8EB00B2B8EB00B1, 0xB8EB00B4B8EB00B3,
-    0xB8EB00B6B8EB00B5, 0xB8EB00B8B8EB00B7, 0xB8EB00BAB8EB00B9,
-    0xB8EB00BCB8EB00BB, 0xB8EB00BEB8EB00BD, 0xB9EB0080B9EB00BF,
-    0xB9EB0082B9EB0081, 0xB9EB0084B9EB0083, 0xB9EB0086B9EB0085,
-    0xB9EB0088B9EB0087, 0xB9EB008AB9EB0089, 0xB9EB008CB9EB008B,
-    0xB9EB008EB9EB008D, 0xB9EB0090B9EB008F, 0xB9EB0092B9EB0091,
-    0xB9EB0094B9EB0093, 0xB9EB0096B9EB0095, 0xB9EB0098B9EB0097,
-    0xB9EB009AB9EB0099, 0xB9EB009CB9EB009B, 0xB9EB009EB9EB009D,
-    0xB9EB00A0B9EB009F, 0xB9EB00A2B9EB00A1, 0xB9EB00A4B9EB00A3,
-    0xB9EB00A6B9EB00A5, 0xB9EB00A8B9EB00A7, 0xB9EB00AAB9EB00A9,
-    0xB9EB00ACB9EB00AB, 0xB9EB00AEB9EB00AD, 0xB9EB00B0B9EB00AF,
-    0xB9EB00B2B9EB00B1, 0xB9EB00B4B9EB00B3, 0xB9EB00B6B9EB00B5,
-    0xB9EB00B8B9EB00B7, 0xB9EB00BAB9EB00B9, 0xB9EB00BCB9EB00BB,
-    0xB9EB00BEB9EB00BD, 0xBAEB0080BAEB00BF, 0xBAEB0082BAEB0081,
-    0xBAEB0084BAEB0083, 0xBAEB0086BAEB0085, 0xBAEB0088BAEB0087,
-    0xBAEB008ABAEB0089, 0xBAEB008CBAEB008B, 0xBAEB008EBAEB008D,
-    0xBAEB0090BAEB008F, 0xBAEB0092BAEB0091, 0xBAEB0094BAEB0093,
-    0xBAEB0096BAEB0095, 0xBAEB0098BAEB0097, 0xBAEB009ABAEB0099,
-    0xBAEB009CBAEB009B, 0xBAEB009EBAEB009D, 0xBAEB00A0BAEB009F,
-    0xBAEB00A2BAEB00A1, 0xBAEB00A4BAEB00A3, 0xBAEB00A6BAEB00A5,
-    0xBAEB00A8BAEB00A7, 0xBAEB00AABAEB00A9, 0xBAEB00ACBAEB00AB,
-    0xBAEB00AEBAEB00AD, 0xBAEB00B0BAEB00AF, 0xBAEB00B2BAEB00B1,
-    0xBAEB00B4BAEB00B3, 0xBAEB00B6BAEB00B5, 0xBAEB00B8BAEB00B7,
-    0xBAEB00BABAEB00B9, 0xBAEB00BCBAEB00BB, 0xBAEB00BEBAEB00BD,
-    0xBBEB0080BBEB00BF, 0xBBEB0082BBEB0081, 0xBBEB0084BBEB0083,
-    0xBBEB0086BBEB0085, 0xBBEB0088BBEB0087, 0xBBEB008ABBEB0089,
-    0xBBEB008CBBEB008B, 0xBBEB008EBBEB008D, 0xBBEB0090BBEB008F,
-    0xBBEB0092BBEB0091, 0xBBEB0094BBEB0093, 0xBBEB0096BBEB0095,
-    0xBBEB0098BBEB0097, 0xBBEB009ABBEB0099, 0xBBEB009CBBEB009B,
-    0xBBEB009EBBEB009D, 0xBBEB00A0BBEB009F, 0xBBEB00A2BBEB00A1,
-    0xBBEB00A4BBEB00A3, 0xBBEB00A6BBEB00A5, 0xBBEB00A8BBEB00A7,
-    0xBBEB00AABBEB00A9, 0xBBEB00ACBBEB00AB, 0xBBEB00AEBBEB00AD,
-    0xBBEB00B0BBEB00AF, 0xBBEB00B2BBEB00B1, 0xBBEB00B4BBEB00B3,
-    0xBBEB00B6BBEB00B5, 0xBBEB00B8BBEB00B7, 0xBBEB00BABBEB00B9,
-    0xBBEB00BCBBEB00BB, 0xBBEB00BEBBEB00BD, 0xBCEB0080BCEB00BF,
-    0xBCEB0082BCEB0081, 0xBCEB0084BCEB0083, 0xBCEB0086BCEB0085,
-    0xBCEB0088BCEB0087, 0xBCEB008ABCEB0089, 0xBCEB008CBCEB008B,
-    0xBCEB008EBCEB008D, 0xBCEB0090BCEB008F, 0xBCEB0092BCEB0091,
-    0xBCEB0094BCEB0093, 0xBCEB0096BCEB0095, 0xBCEB0098BCEB0097,
-    0xBCEB009ABCEB0099, 0xBCEB009CBCEB009B, 0xBCEB009EBCEB009D,
-    0xBCEB00A0BCEB009F, 0xBCEB00A2BCEB00A1, 0xBCEB00A4BCEB00A3,
-    0xBCEB00A6BCEB00A5, 0xBCEB00A8BCEB00A7, 0xBCEB00AABCEB00A9,
-    0xBCEB00ACBCEB00AB, 0xBCEB00AEBCEB00AD, 0xBCEB00B0BCEB00AF,
-    0xBCEB00B2BCEB00B1, 0xBCEB00B4BCEB00B3, 0xBCEB00B6BCEB00B5,
-    0xBCEB00B8BCEB00B7, 0xBCEB00BABCEB00B9, 0xBCEB00BCBCEB00BB,
-    0xBCEB00BEBCEB00BD, 0xBDEB0080BDEB00BF, 0xBDEB0082BDEB0081,
-    0xBDEB0084BDEB0083, 0xBDEB0086BDEB0085, 0xBDEB0088BDEB0087,
-    0xBDEB008ABDEB0089, 0xBDEB008CBDEB008B, 0xBDEB008EBDEB008D,
-    0xBDEB0090BDEB008F, 0xBDEB0092BDEB0091, 0xBDEB0094BDEB0093,
-    0xBDEB0096BDEB0095, 0xBDEB0098BDEB0097, 0xBDEB009ABDEB0099,
-    0xBDEB009CBDEB009B, 0xBDEB009EBDEB009D, 0xBDEB00A0BDEB009F,
-    0xBDEB00A2BDEB00A1, 0xBDEB00A4BDEB00A3, 0xBDEB00A6BDEB00A5,
-    0xBDEB00A8BDEB00A7, 0xBDEB00AABDEB00A9, 0xBDEB00ACBDEB00AB,
-    0xBDEB00AEBDEB00AD, 0xBDEB00B0BDEB00AF, 0xBDEB00B2BDEB00B1,
-    0xBDEB00B4BDEB00B3, 0xBDEB00B6BDEB00B5, 0xBDEB00B8BDEB00B7,
-    0xBDEB00BABDEB00B9, 0xBDEB00BCBDEB00BB, 0xBDEB00BEBDEB00BD,
-    0xBEEB0080BEEB00BF, 0xBEEB0082BEEB0081, 0xBEEB0084BEEB0083,
-    0xBEEB0086BEEB0085, 0xBEEB0088BEEB0087, 0xBEEB008ABEEB0089,
-    0xBEEB008CBEEB008B, 0xBEEB008EBEEB008D, 0xBEEB0090BEEB008F,
-    0xBEEB0092BEEB0091, 0xBEEB0094BEEB0093, 0xBEEB0096BEEB0095,
-    0xBEEB0098BEEB0097, 0xBEEB009ABEEB0099, 0xBEEB009CBEEB009B,
-    0xBEEB009EBEEB009D, 0xBEEB00A0BEEB009F, 0xBEEB00A2BEEB00A1,
-    0xBEEB00A4BEEB00A3, 0xBEEB00A6BEEB00A5, 0xBEEB00A8BEEB00A7,
-    0xBEEB00AABEEB00A9, 0xBEEB00ACBEEB00AB, 0xBEEB00AEBEEB00AD,
-    0xBEEB00B0BEEB00AF, 0xBEEB00B2BEEB00B1, 0xBEEB00B4BEEB00B3,
-    0xBEEB00B6BEEB00B5, 0xBEEB00B8BEEB00B7, 0xBEEB00BABEEB00B9,
-    0xBEEB00BCBEEB00BB, 0xBEEB00BEBEEB00BD, 0xBFEB0080BFEB00BF,
-    0xBFEB0082BFEB0081, 0xBFEB0084BFEB0083, 0xBFEB0086BFEB0085,
-    0xBFEB0088BFEB0087, 0xBFEB008ABFEB0089, 0xBFEB008CBFEB008B,
-    0xBFEB008EBFEB008D, 0xBFEB0090BFEB008F, 0xBFEB0092BFEB0091,
-    0xBFEB0094BFEB0093, 0xBFEB0096BFEB0095, 0xBFEB0098BFEB0097,
-    0xBFEB009ABFEB0099, 0xBFEB009CBFEB009B, 0xBFEB009EBFEB009D,
-    0xBFEB00A0BFEB009F, 0xBFEB00A2BFEB00A1, 0xBFEB00A4BFEB00A3,
-    0xBFEB00A6BFEB00A5, 0xBFEB00A8BFEB00A7, 0xBFEB00AABFEB00A9,
-    0xBFEB00ACBFEB00AB, 0xBFEB00AEBFEB00AD, 0xBFEB00B0BFEB00AF,
-    0xBFEB00B2BFEB00B1, 0xBFEB00B4BFEB00B3, 0xBFEB00B6BFEB00B5,
-    0xBFEB00B8BFEB00B7, 0xBFEB00BABFEB00B9, 0xBFEB00BCBFEB00BB,
-    0xBFEB00BEBFEB00BD, 0x80EC008080EC00BF, 0x80EC008280EC0081,
-    0x80EC008480EC0083, 0x80EC008680EC0085, 0x80EC008880EC0087,
-    0x80EC008A80EC0089, 0x80EC008C80EC008B, 0x80EC008E80EC008D,
-    0x80EC009080EC008F, 0x80EC009280EC0091, 0x80EC009480EC0093,
-    0x80EC009680EC0095, 0x80EC009880EC0097, 0x80EC009A80EC0099,
-    0x80EC009C80EC009B, 0x80EC009E80EC009D, 0x80EC00A080EC009F,
-    0x80EC00A280EC00A1, 0x80EC00A480EC00A3, 0x80EC00A680EC00A5,
-    0x80EC00A880EC00A7, 0x80EC00AA80EC00A9, 0x80EC00AC80EC00AB,
-    0x80EC00AE80EC00AD, 0x80EC00B080EC00AF, 0x80EC00B280EC00B1,
-    0x80EC00B480EC00B3, 0x80EC00B680EC00B5, 0x80EC00B880EC00B7,
-    0x80EC00BA80EC00B9, 0x80EC00BC80EC00BB, 0x80EC00BE80EC00BD,
-    0x81EC008081EC00BF, 0x81EC008281EC0081, 0x81EC008481EC0083,
-    0x81EC008681EC0085, 0x81EC008881EC0087, 0x81EC008A81EC0089,
-    0x81EC008C81EC008B, 0x81EC008E81EC008D, 0x81EC009081EC008F,
-    0x81EC009281EC0091, 0x81EC009481EC0093, 0x81EC009681EC0095,
-    0x81EC009881EC0097, 0x81EC009A81EC0099, 0x81EC009C81EC009B,
-    0x81EC009E81EC009D, 0x81EC00A081EC009F, 0x81EC00A281EC00A1,
-    0x81EC00A481EC00A3, 0x81EC00A681EC00A5, 0x81EC00A881EC00A7,
-    0x81EC00AA81EC00A9, 0x81EC00AC81EC00AB, 0x81EC00AE81EC00AD,
-    0x81EC00B081EC00AF, 0x81EC00B281EC00B1, 0x81EC00B481EC00B3,
-    0x81EC00B681EC00B5, 0x81EC00B881EC00B7, 0x81EC00BA81EC00B9,
-    0x81EC00BC81EC00BB, 0x81EC00BE81EC00BD, 0x82EC008082EC00BF,
-    0x82EC008282EC0081, 0x82EC008482EC0083, 0x82EC008682EC0085,
-    0x82EC008882EC0087, 0x82EC008A82EC0089, 0x82EC008C82EC008B,
-    0x82EC008E82EC008D, 0x82EC009082EC008F, 0x82EC009282EC0091,
-    0x82EC009482EC0093, 0x82EC009682EC0095, 0x82EC009882EC0097,
-    0x82EC009A82EC0099, 0x82EC009C82EC009B, 0x82EC009E82EC009D,
-    0x82EC00A082EC009F, 0x82EC00A282EC00A1, 0x82EC00A482EC00A3,
-    0x82EC00A682EC00A5, 0x82EC00A882EC00A7, 0x82EC00AA82EC00A9,
-    0x82EC00AC82EC00AB, 0x82EC00AE82EC00AD, 0x82EC00B082EC00AF,
-    0x82EC00B282EC00B1, 0x82EC00B482EC00B3, 0x82EC00B682EC00B5,
-    0x82EC00B882EC00B7, 0x82EC00BA82EC00B9, 0x82EC00BC82EC00BB,
-    0x82EC00BE82EC00BD, 0x83EC008083EC00BF, 0x83EC008283EC0081,
-    0x83EC008483EC0083, 0x83EC008683EC0085, 0x83EC008883EC0087,
-    0x83EC008A83EC0089, 0x83EC008C83EC008B, 0x83EC008E83EC008D,
-    0x83EC009083EC008F, 0x83EC009283EC0091, 0x83EC009483EC0093,
-    0x83EC009683EC0095, 0x83EC009883EC0097, 0x83EC009A83EC0099,
-    0x83EC009C83EC009B, 0x83EC009E83EC009D, 0x83EC00A083EC009F,
-    0x83EC00A283EC00A1, 0x83EC00A483EC00A3, 0x83EC00A683EC00A5,
-    0x83EC00A883EC00A7, 0x83EC00AA83EC00A9, 0x83EC00AC83EC00AB,
-    0x83EC00AE83EC00AD, 0x83EC00B083EC00AF, 0x83EC00B283EC00B1,
-    0x83EC00B483EC00B3, 0x83EC00B683EC00B5, 0x83EC00B883EC00B7,
-    0x83EC00BA83EC00B9, 0x83EC00BC83EC00BB, 0x83EC00BE83EC00BD,
-    0x84EC008084EC00BF, 0x84EC008284EC0081, 0x84EC008484EC0083,
-    0x84EC008684EC0085, 0x84EC008884EC0087, 0x84EC008A84EC0089,
-    0x84EC008C84EC008B, 0x84EC008E84EC008D, 0x84EC009084EC008F,
-    0x84EC009284EC0091, 0x84EC009484EC0093, 0x84EC009684EC0095,
-    0x84EC009884EC0097, 0x84EC009A84EC0099, 0x84EC009C84EC009B,
-    0x84EC009E84EC009D, 0x84EC00A084EC009F, 0x84EC00A284EC00A1,
-    0x84EC00A484EC00A3, 0x84EC00A684EC00A5, 0x84EC00A884EC00A7,
-    0x84EC00AA84EC00A9, 0x84EC00AC84EC00AB, 0x84EC00AE84EC00AD,
-    0x84EC00B084EC00AF, 0x84EC00B284EC00B1, 0x84EC00B484EC00B3,
-    0x84EC00B684EC00B5, 0x84EC00B884EC00B7, 0x84EC00BA84EC00B9,
-    0x84EC00BC84EC00BB, 0x84EC00BE84EC00BD, 0x85EC008085EC00BF,
-    0x85EC008285EC0081, 0x85EC008485EC0083, 0x85EC008685EC0085,
-    0x85EC008885EC0087, 0x85EC008A85EC0089, 0x85EC008C85EC008B,
-    0x85EC008E85EC008D, 0x85EC009085EC008F, 0x85EC009285EC0091,
-    0x85EC009485EC0093, 0x85EC009685EC0095, 0x85EC009885EC0097,
-    0x85EC009A85EC0099, 0x85EC009C85EC009B, 0x85EC009E85EC009D,
-    0x85EC00A085EC009F, 0x85EC00A285EC00A1, 0x85EC00A485EC00A3,
-    0x85EC00A685EC00A5, 0x85EC00A885EC00A7, 0x85EC00AA85EC00A9,
-    0x85EC00AC85EC00AB, 0x85EC00AE85EC00AD, 0x85EC00B085EC00AF,
-    0x85EC00B285EC00B1, 0x85EC00B485EC00B3, 0x85EC00B685EC00B5,
-    0x85EC00B885EC00B7, 0x85EC00BA85EC00B9, 0x85EC00BC85EC00BB,
-    0x85EC00BE85EC00BD, 0x86EC008086EC00BF, 0x86EC008286EC0081,
-    0x86EC008486EC0083, 0x86EC008686EC0085, 0x86EC008886EC0087,
-    0x86EC008A86EC0089, 0x86EC008C86EC008B, 0x86EC008E86EC008D,
-    0x86EC009086EC008F, 0x86EC009286EC0091, 0x86EC009486EC0093,
-    0x86EC009686EC0095, 0x86EC009886EC0097, 0x86EC009A86EC0099,
-    0x86EC009C86EC009B, 0x86EC009E86EC009D, 0x86EC00A086EC009F,
-    0x86EC00A286EC00A1, 0x86EC00A486EC00A3, 0x86EC00A686EC00A5,
-    0x86EC00A886EC00A7, 0x86EC00AA86EC00A9, 0x86EC00AC86EC00AB,
-    0x86EC00AE86EC00AD, 0x86EC00B086EC00AF, 0x86EC00B286EC00B1,
-    0x86EC00B486EC00B3, 0x86EC00B686EC00B5, 0x86EC00B886EC00B7,
-    0x86EC00BA86EC00B9, 0x86EC00BC86EC00BB, 0x86EC00BE86EC00BD,
-    0x87EC008087EC00BF, 0x87EC008287EC0081, 0x87EC008487EC0083,
-    0x87EC008687EC0085, 0x87EC008887EC0087, 0x87EC008A87EC0089,
-    0x87EC008C87EC008B, 0x87EC008E87EC008D, 0x87EC009087EC008F,
-    0x87EC009287EC0091, 0x87EC009487EC0093, 0x87EC009687EC0095,
-    0x87EC009887EC0097, 0x87EC009A87EC0099, 0x87EC009C87EC009B,
-    0x87EC009E87EC009D, 0x87EC00A087EC009F, 0x87EC00A287EC00A1,
-    0x87EC00A487EC00A3, 0x87EC00A687EC00A5, 0x87EC00A887EC00A7,
-    0x87EC00AA87EC00A9, 0x87EC00AC87EC00AB, 0x87EC00AE87EC00AD,
-    0x87EC00B087EC00AF, 0x87EC00B287EC00B1, 0x87EC00B487EC00B3,
-    0x87EC00B687EC00B5, 0x87EC00B887EC00B7, 0x87EC00BA87EC00B9,
-    0x87EC00BC87EC00BB, 0x87EC00BE87EC00BD, 0x88EC008088EC00BF,
-    0x88EC008288EC0081, 0x88EC008488EC0083, 0x88EC008688EC0085,
-    0x88EC008888EC0087, 0x88EC008A88EC0089, 0x88EC008C88EC008B,
-    0x88EC008E88EC008D, 0x88EC009088EC008F, 0x88EC009288EC0091,
-    0x88EC009488EC0093, 0x88EC009688EC0095, 0x88EC009888EC0097,
-    0x88EC009A88EC0099, 0x88EC009C88EC009B, 0x88EC009E88EC009D,
-    0x88EC00A088EC009F, 0x88EC00A288EC00A1, 0x88EC00A488EC00A3,
-    0x88EC00A688EC00A5, 0x88EC00A888EC00A7, 0x88EC00AA88EC00A9,
-    0x88EC00AC88EC00AB, 0x88EC00AE88EC00AD, 0x88EC00B088EC00AF,
-    0x88EC00B288EC00B1, 0x88EC00B488EC00B3, 0x88EC00B688EC00B5,
-    0x88EC00B888EC00B7, 0x88EC00BA88EC00B9, 0x88EC00BC88EC00BB,
-    0x88EC00BE88EC00BD, 0x89EC008089EC00BF, 0x89EC008289EC0081,
-    0x89EC008489EC0083, 0x89EC008689EC0085, 0x89EC008889EC0087,
-    0x89EC008A89EC0089, 0x89EC008C89EC008B, 0x89EC008E89EC008D,
-    0x89EC009089EC008F, 0x89EC009289EC0091, 0x89EC009489EC0093,
-    0x89EC009689EC0095, 0x89EC009889EC0097, 0x89EC009A89EC0099,
-    0x89EC009C89EC009B, 0x89EC009E89EC009D, 0x89EC00A089EC009F,
-    0x89EC00A289EC00A1, 0x89EC00A489EC00A3, 0x89EC00A689EC00A5,
-    0x89EC00A889EC00A7, 0x89EC00AA89EC00A9, 0x89EC00AC89EC00AB,
-    0x89EC00AE89EC00AD, 0x89EC00B089EC00AF, 0x89EC00B289EC00B1,
-    0x89EC00B489EC00B3, 0x89EC00B689EC00B5, 0x89EC00B889EC00B7,
-    0x89EC00BA89EC00B9, 0x89EC00BC89EC00BB, 0x89EC00BE89EC00BD,
-    0x8AEC00808AEC00BF, 0x8AEC00828AEC0081, 0x8AEC00848AEC0083,
-    0x8AEC00868AEC0085, 0x8AEC00888AEC0087, 0x8AEC008A8AEC0089,
-    0x8AEC008C8AEC008B, 0x8AEC008E8AEC008D, 0x8AEC00908AEC008F,
-    0x8AEC00928AEC0091, 0x8AEC00948AEC0093, 0x8AEC00968AEC0095,
-    0x8AEC00988AEC0097, 0x8AEC009A8AEC0099, 0x8AEC009C8AEC009B,
-    0x8AEC009E8AEC009D, 0x8AEC00A08AEC009F, 0x8AEC00A28AEC00A1,
-    0x8AEC00A48AEC00A3, 0x8AEC00A68AEC00A5, 0x8AEC00A88AEC00A7,
-    0x8AEC00AA8AEC00A9, 0x8AEC00AC8AEC00AB, 0x8AEC00AE8AEC00AD,
-    0x8AEC00B08AEC00AF, 0x8AEC00B28AEC00B1, 0x8AEC00B48AEC00B3,
-    0x8AEC00B68AEC00B5, 0x8AEC00B88AEC00B7, 0x8AEC00BA8AEC00B9,
-    0x8AEC00BC8AEC00BB, 0x8AEC00BE8AEC00BD, 0x8BEC00808BEC00BF,
-    0x8BEC00828BEC0081, 0x8BEC00848BEC0083, 0x8BEC00868BEC0085,
-    0x8BEC00888BEC0087, 0x8BEC008A8BEC0089, 0x8BEC008C8BEC008B,
-    0x8BEC008E8BEC008D, 0x8BEC00908BEC008F, 0x8BEC00928BEC0091,
-    0x8BEC00948BEC0093, 0x8BEC00968BEC0095, 0x8BEC00988BEC0097,
-    0x8BEC009A8BEC0099, 0x8BEC009C8BEC009B, 0x8BEC009E8BEC009D,
-    0x8BEC00A08BEC009F, 0x8BEC00A28BEC00A1, 0x8BEC00A48BEC00A3,
-    0x8BEC00A68BEC00A5, 0x8BEC00A88BEC00A7, 0x8BEC00AA8BEC00A9,
-    0x8BEC00AC8BEC00AB, 0x8BEC00AE8BEC00AD, 0x8BEC00B08BEC00AF,
-    0x8BEC00B28BEC00B1, 0x8BEC00B48BEC00B3, 0x8BEC00B68BEC00B5,
-    0x8BEC00B88BEC00B7, 0x8BEC00BA8BEC00B9, 0x8BEC00BC8BEC00BB,
-    0x8BEC00BE8BEC00BD, 0x8CEC00808CEC00BF, 0x8CEC00828CEC0081,
-    0x8CEC00848CEC0083, 0x8CEC00868CEC0085, 0x8CEC00888CEC0087,
-    0x8CEC008A8CEC0089, 0x8CEC008C8CEC008B, 0x8CEC008E8CEC008D,
-    0x8CEC00908CEC008F, 0x8CEC00928CEC0091, 0x8CEC00948CEC0093,
-    0x8CEC00968CEC0095, 0x8CEC00988CEC0097, 0x8CEC009A8CEC0099,
-    0x8CEC009C8CEC009B, 0x8CEC009E8CEC009D, 0x8CEC00A08CEC009F,
-    0x8CEC00A28CEC00A1, 0x8CEC00A48CEC00A3, 0x8CEC00A68CEC00A5,
-    0x8CEC00A88CEC00A7, 0x8CEC00AA8CEC00A9, 0x8CEC00AC8CEC00AB,
-    0x8CEC00AE8CEC00AD, 0x8CEC00B08CEC00AF, 0x8CEC00B28CEC00B1,
-    0x8CEC00B48CEC00B3, 0x8CEC00B68CEC00B5, 0x8CEC00B88CEC00B7,
-    0x8CEC00BA8CEC00B9, 0x8CEC00BC8CEC00BB, 0x8CEC00BE8CEC00BD,
-    0x8DEC00808DEC00BF, 0x8DEC00828DEC0081, 0x8DEC00848DEC0083,
-    0x8DEC00868DEC0085, 0x8DEC00888DEC0087, 0x8DEC008A8DEC0089,
-    0x8DEC008C8DEC008B, 0x8DEC008E8DEC008D, 0x8DEC00908DEC008F,
-    0x8DEC00928DEC0091, 0x8DEC00948DEC0093, 0x8DEC00968DEC0095,
-    0x8DEC00988DEC0097, 0x8DEC009A8DEC0099, 0x8DEC009C8DEC009B,
-    0x8DEC009E8DEC009D, 0x8DEC00A08DEC009F, 0x8DEC00A28DEC00A1,
-    0x8DEC00A48DEC00A3, 0x8DEC00A68DEC00A5, 0x8DEC00A88DEC00A7,
-    0x8DEC00AA8DEC00A9, 0x8DEC00AC8DEC00AB, 0x8DEC00AE8DEC00AD,
-    0x8DEC00B08DEC00AF, 0x8DEC00B28DEC00B1, 0x8DEC00B48DEC00B3,
-    0x8DEC00B68DEC00B5, 0x8DEC00B88DEC00B7, 0x8DEC00BA8DEC00B9,
-    0x8DEC00BC8DEC00BB, 0x8DEC00BE8DEC00BD, 0x8EEC00808EEC00BF,
-    0x8EEC00828EEC0081, 0x8EEC00848EEC0083, 0x8EEC00868EEC0085,
-    0x8EEC00888EEC0087, 0x8EEC008A8EEC0089, 0x8EEC008C8EEC008B,
-    0x8EEC008E8EEC008D, 0x8EEC00908EEC008F, 0x8EEC00928EEC0091,
-    0x8EEC00948EEC0093, 0x8EEC00968EEC0095, 0x8EEC00988EEC0097,
-    0x8EEC009A8EEC0099, 0x8EEC009C8EEC009B, 0x8EEC009E8EEC009D,
-    0x8EEC00A08EEC009F, 0x8EEC00A28EEC00A1, 0x8EEC00A48EEC00A3,
-    0x8EEC00A68EEC00A5, 0x8EEC00A88EEC00A7, 0x8EEC00AA8EEC00A9,
-    0x8EEC00AC8EEC00AB, 0x8EEC00AE8EEC00AD, 0x8EEC00B08EEC00AF,
-    0x8EEC00B28EEC00B1, 0x8EEC00B48EEC00B3, 0x8EEC00B68EEC00B5,
-    0x8EEC00B88EEC00B7, 0x8EEC00BA8EEC00B9, 0x8EEC00BC8EEC00BB,
-    0x8EEC00BE8EEC00BD, 0x8FEC00808FEC00BF, 0x8FEC00828FEC0081,
-    0x8FEC00848FEC0083, 0x8FEC00868FEC0085, 0x8FEC00888FEC0087,
-    0x8FEC008A8FEC0089, 0x8FEC008C8FEC008B, 0x8FEC008E8FEC008D,
-    0x8FEC00908FEC008F, 0x8FEC00928FEC0091, 0x8FEC00948FEC0093,
-    0x8FEC00968FEC0095, 0x8FEC00988FEC0097, 0x8FEC009A8FEC0099,
-    0x8FEC009C8FEC009B, 0x8FEC009E8FEC009D, 0x8FEC00A08FEC009F,
-    0x8FEC00A28FEC00A1, 0x8FEC00A48FEC00A3, 0x8FEC00A68FEC00A5,
-    0x8FEC00A88FEC00A7, 0x8FEC00AA8FEC00A9, 0x8FEC00AC8FEC00AB,
-    0x8FEC00AE8FEC00AD, 0x8FEC00B08FEC00AF, 0x8FEC00B28FEC00B1,
-    0x8FEC00B48FEC00B3, 0x8FEC00B68FEC00B5, 0x8FEC00B88FEC00B7,
-    0x8FEC00BA8FEC00B9, 0x8FEC00BC8FEC00BB, 0x8FEC00BE8FEC00BD,
-    0x90EC008090EC00BF, 0x90EC008290EC0081, 0x90EC008490EC0083,
-    0x90EC008690EC0085, 0x90EC008890EC0087, 0x90EC008A90EC0089,
-    0x90EC008C90EC008B, 0x90EC008E90EC008D, 0x90EC009090EC008F,
-    0x90EC009290EC0091, 0x90EC009490EC0093, 0x90EC009690EC0095,
-    0x90EC009890EC0097, 0x90EC009A90EC0099, 0x90EC009C90EC009B,
-    0x90EC009E90EC009D, 0x90EC00A090EC009F, 0x90EC00A290EC00A1,
-    0x90EC00A490EC00A3, 0x90EC00A690EC00A5, 0x90EC00A890EC00A7,
-    0x90EC00AA90EC00A9, 0x90EC00AC90EC00AB, 0x90EC00AE90EC00AD,
-    0x90EC00B090EC00AF, 0x90EC00B290EC00B1, 0x90EC00B490EC00B3,
-    0x90EC00B690EC00B5, 0x90EC00B890EC00B7, 0x90EC00BA90EC00B9,
-    0x90EC00BC90EC00BB, 0x90EC00BE90EC00BD, 0x91EC008091EC00BF,
-    0x91EC008291EC0081, 0x91EC008491EC0083, 0x91EC008691EC0085,
-    0x91EC008891EC0087, 0x91EC008A91EC0089, 0x91EC008C91EC008B,
-    0x91EC008E91EC008D, 0x91EC009091EC008F, 0x91EC009291EC0091,
-    0x91EC009491EC0093, 0x91EC009691EC0095, 0x91EC009891EC0097,
-    0x91EC009A91EC0099, 0x91EC009C91EC009B, 0x91EC009E91EC009D,
-    0x91EC00A091EC009F, 0x91EC00A291EC00A1, 0x91EC00A491EC00A3,
-    0x91EC00A691EC00A5, 0x91EC00A891EC00A7, 0x91EC00AA91EC00A9,
-    0x91EC00AC91EC00AB, 0x91EC00AE91EC00AD, 0x91EC00B091EC00AF,
-    0x91EC00B291EC00B1, 0x91EC00B491EC00B3, 0x91EC00B691EC00B5,
-    0x91EC00B891EC00B7, 0x91EC00BA91EC00B9, 0x91EC00BC91EC00BB,
-    0x91EC00BE91EC00BD, 0x92EC008092EC00BF, 0x92EC008292EC0081,
-    0x92EC008492EC0083, 0x92EC008692EC0085, 0x92EC008892EC0087,
-    0x92EC008A92EC0089, 0x92EC008C92EC008B, 0x92EC008E92EC008D,
-    0x92EC009092EC008F, 0x92EC009292EC0091, 0x92EC009492EC0093,
-    0x92EC009692EC0095, 0x92EC009892EC0097, 0x92EC009A92EC0099,
-    0x92EC009C92EC009B, 0x92EC009E92EC009D, 0x92EC00A092EC009F,
-    0x92EC00A292EC00A1, 0x92EC00A492EC00A3, 0x92EC00A692EC00A5,
-    0x92EC00A892EC00A7, 0x92EC00AA92EC00A9, 0x92EC00AC92EC00AB,
-    0x92EC00AE92EC00AD, 0x92EC00B092EC00AF, 0x92EC00B292EC00B1,
-    0x92EC00B492EC00B3, 0x92EC00B692EC00B5, 0x92EC00B892EC00B7,
-    0x92EC00BA92EC00B9, 0x92EC00BC92EC00BB, 0x92EC00BE92EC00BD,
-    0x93EC008093EC00BF, 0x93EC008293EC0081, 0x93EC008493EC0083,
-    0x93EC008693EC0085, 0x93EC008893EC0087, 0x93EC008A93EC0089,
-    0x93EC008C93EC008B, 0x93EC008E93EC008D, 0x93EC009093EC008F,
-    0x93EC009293EC0091, 0x93EC009493EC0093, 0x93EC009693EC0095,
-    0x93EC009893EC0097, 0x93EC009A93EC0099, 0x93EC009C93EC009B,
-    0x93EC009E93EC009D, 0x93EC00A093EC009F, 0x93EC00A293EC00A1,
-    0x93EC00A493EC00A3, 0x93EC00A693EC00A5, 0x93EC00A893EC00A7,
-    0x93EC00AA93EC00A9, 0x93EC00AC93EC00AB, 0x93EC00AE93EC00AD,
-    0x93EC00B093EC00AF, 0x93EC00B293EC00B1, 0x93EC00B493EC00B3,
-    0x93EC00B693EC00B5, 0x93EC00B893EC00B7, 0x93EC00BA93EC00B9,
-    0x93EC00BC93EC00BB, 0x93EC00BE93EC00BD, 0x94EC008094EC00BF,
-    0x94EC008294EC0081, 0x94EC008494EC0083, 0x94EC008694EC0085,
-    0x94EC008894EC0087, 0x94EC008A94EC0089, 0x94EC008C94EC008B,
-    0x94EC008E94EC008D, 0x94EC009094EC008F, 0x94EC009294EC0091,
-    0x94EC009494EC0093, 0x94EC009694EC0095, 0x94EC009894EC0097,
-    0x94EC009A94EC0099, 0x94EC009C94EC009B, 0x94EC009E94EC009D,
-    0x94EC00A094EC009F, 0x94EC00A294EC00A1, 0x94EC00A494EC00A3,
-    0x94EC00A694EC00A5, 0x94EC00A894EC00A7, 0x94EC00AA94EC00A9,
-    0x94EC00AC94EC00AB, 0x94EC00AE94EC00AD, 0x94EC00B094EC00AF,
-    0x94EC00B294EC00B1, 0x94EC00B494EC00B3, 0x94EC00B694EC00B5,
-    0x94EC00B894EC00B7, 0x94EC00BA94EC00B9, 0x94EC00BC94EC00BB,
-    0x94EC00BE94EC00BD, 0x95EC008095EC00BF, 0x95EC008295EC0081,
-    0x95EC008495EC0083, 0x95EC008695EC0085, 0x95EC008895EC0087,
-    0x95EC008A95EC0089, 0x95EC008C95EC008B, 0x95EC008E95EC008D,
-    0x95EC009095EC008F, 0x95EC009295EC0091, 0x95EC009495EC0093,
-    0x95EC009695EC0095, 0x95EC009895EC0097, 0x95EC009A95EC0099,
-    0x95EC009C95EC009B, 0x95EC009E95EC009D, 0x95EC00A095EC009F,
-    0x95EC00A295EC00A1, 0x95EC00A495EC00A3, 0x95EC00A695EC00A5,
-    0x95EC00A895EC00A7, 0x95EC00AA95EC00A9, 0x95EC00AC95EC00AB,
-    0x95EC00AE95EC00AD, 0x95EC00B095EC00AF, 0x95EC00B295EC00B1,
-    0x95EC00B495EC00B3, 0x95EC00B695EC00B5, 0x95EC00B895EC00B7,
-    0x95EC00BA95EC00B9, 0x95EC00BC95EC00BB, 0x95EC00BE95EC00BD,
-    0x96EC008096EC00BF, 0x96EC008296EC0081, 0x96EC008496EC0083,
-    0x96EC008696EC0085, 0x96EC008896EC0087, 0x96EC008A96EC0089,
-    0x96EC008C96EC008B, 0x96EC008E96EC008D, 0x96EC009096EC008F,
-    0x96EC009296EC0091, 0x96EC009496EC0093, 0x96EC009696EC0095,
-    0x96EC009896EC0097, 0x96EC009A96EC0099, 0x96EC009C96EC009B,
-    0x96EC009E96EC009D, 0x96EC00A096EC009F, 0x96EC00A296EC00A1,
-    0x96EC00A496EC00A3, 0x96EC00A696EC00A5, 0x96EC00A896EC00A7,
-    0x96EC00AA96EC00A9, 0x96EC00AC96EC00AB, 0x96EC00AE96EC00AD,
-    0x96EC00B096EC00AF, 0x96EC00B296EC00B1, 0x96EC00B496EC00B3,
-    0x96EC00B696EC00B5, 0x96EC00B896EC00B7, 0x96EC00BA96EC00B9,
-    0x96EC00BC96EC00BB, 0x96EC00BE96EC00BD, 0x97EC008097EC00BF,
-    0x97EC008297EC0081, 0x97EC008497EC0083, 0x97EC008697EC0085,
-    0x97EC008897EC0087, 0x97EC008A97EC0089, 0x97EC008C97EC008B,
-    0x97EC008E97EC008D, 0x97EC009097EC008F, 0x97EC009297EC0091,
-    0x97EC009497EC0093, 0x97EC009697EC0095, 0x97EC009897EC0097,
-    0x97EC009A97EC0099, 0x97EC009C97EC009B, 0x97EC009E97EC009D,
-    0x97EC00A097EC009F, 0x97EC00A297EC00A1, 0x97EC00A497EC00A3,
-    0x97EC00A697EC00A5, 0x97EC00A897EC00A7, 0x97EC00AA97EC00A9,
-    0x97EC00AC97EC00AB, 0x97EC00AE97EC00AD, 0x97EC00B097EC00AF,
-    0x97EC00B297EC00B1, 0x97EC00B497EC00B3, 0x97EC00B697EC00B5,
-    0x97EC00B897EC00B7, 0x97EC00BA97EC00B9, 0x97EC00BC97EC00BB,
-    0x97EC00BE97EC00BD, 0x98EC008098EC00BF, 0x98EC008298EC0081,
-    0x98EC008498EC0083, 0x98EC008698EC0085, 0x98EC008898EC0087,
-    0x98EC008A98EC0089, 0x98EC008C98EC008B, 0x98EC008E98EC008D,
-    0x98EC009098EC008F, 0x98EC009298EC0091, 0x98EC009498EC0093,
-    0x98EC009698EC0095, 0x98EC009898EC0097, 0x98EC009A98EC0099,
-    0x98EC009C98EC009B, 0x98EC009E98EC009D, 0x98EC00A098EC009F,
-    0x98EC00A298EC00A1, 0x98EC00A498EC00A3, 0x98EC00A698EC00A5,
-    0x98EC00A898EC00A7, 0x98EC00AA98EC00A9, 0x98EC00AC98EC00AB,
-    0x98EC00AE98EC00AD, 0x98EC00B098EC00AF, 0x98EC00B298EC00B1,
-    0x98EC00B498EC00B3, 0x98EC00B698EC00B5, 0x98EC00B898EC00B7,
-    0x98EC00BA98EC00B9, 0x98EC00BC98EC00BB, 0x98EC00BE98EC00BD,
-    0x99EC008099EC00BF, 0x99EC008299EC0081, 0x99EC008499EC0083,
-    0x99EC008699EC0085, 0x99EC008899EC0087, 0x99EC008A99EC0089,
-    0x99EC008C99EC008B, 0x99EC008E99EC008D, 0x99EC009099EC008F,
-    0x99EC009299EC0091, 0x99EC009499EC0093, 0x99EC009699EC0095,
-    0x99EC009899EC0097, 0x99EC009A99EC0099, 0x99EC009C99EC009B,
-    0x99EC009E99EC009D, 0x99EC00A099EC009F, 0x99EC00A299EC00A1,
-    0x99EC00A499EC00A3, 0x99EC00A699EC00A5, 0x99EC00A899EC00A7,
-    0x99EC00AA99EC00A9, 0x99EC00AC99EC00AB, 0x99EC00AE99EC00AD,
-    0x99EC00B099EC00AF, 0x99EC00B299EC00B1, 0x99EC00B499EC00B3,
-    0x99EC00B699EC00B5, 0x99EC00B899EC00B7, 0x99EC00BA99EC00B9,
-    0x99EC00BC99EC00BB, 0x99EC00BE99EC00BD, 0x9AEC00809AEC00BF,
-    0x9AEC00829AEC0081, 0x9AEC00849AEC0083, 0x9AEC00869AEC0085,
-    0x9AEC00889AEC0087, 0x9AEC008A9AEC0089, 0x9AEC008C9AEC008B,
-    0x9AEC008E9AEC008D, 0x9AEC00909AEC008F, 0x9AEC00929AEC0091,
-    0x9AEC00949AEC0093, 0x9AEC00969AEC0095, 0x9AEC00989AEC0097,
-    0x9AEC009A9AEC0099, 0x9AEC009C9AEC009B, 0x9AEC009E9AEC009D,
-    0x9AEC00A09AEC009F, 0x9AEC00A29AEC00A1, 0x9AEC00A49AEC00A3,
-    0x9AEC00A69AEC00A5, 0x9AEC00A89AEC00A7, 0x9AEC00AA9AEC00A9,
-    0x9AEC00AC9AEC00AB, 0x9AEC00AE9AEC00AD, 0x9AEC00B09AEC00AF,
-    0x9AEC00B29AEC00B1, 0x9AEC00B49AEC00B3, 0x9AEC00B69AEC00B5,
-    0x9AEC00B89AEC00B7, 0x9AEC00BA9AEC00B9, 0x9AEC00BC9AEC00BB,
-    0x9AEC00BE9AEC00BD, 0x9BEC00809BEC00BF, 0x9BEC00829BEC0081,
-    0x9BEC00849BEC0083, 0x9BEC00869BEC0085, 0x9BEC00889BEC0087,
-    0x9BEC008A9BEC0089, 0x9BEC008C9BEC008B, 0x9BEC008E9BEC008D,
-    0x9BEC00909BEC008F, 0x9BEC00929BEC0091, 0x9BEC00949BEC0093,
-    0x9BEC00969BEC0095, 0x9BEC00989BEC0097, 0x9BEC009A9BEC0099,
-    0x9BEC009C9BEC009B, 0x9BEC009E9BEC009D, 0x9BEC00A09BEC009F,
-    0x9BEC00A29BEC00A1, 0x9BEC00A49BEC00A3, 0x9BEC00A69BEC00A5,
-    0x9BEC00A89BEC00A7, 0x9BEC00AA9BEC00A9, 0x9BEC00AC9BEC00AB,
-    0x9BEC00AE9BEC00AD, 0x9BEC00B09BEC00AF, 0x9BEC00B29BEC00B1,
-    0x9BEC00B49BEC00B3, 0x9BEC00B69BEC00B5, 0x9BEC00B89BEC00B7,
-    0x9BEC00BA9BEC00B9, 0x9BEC00BC9BEC00BB, 0x9BEC00BE9BEC00BD,
-    0x9CEC00809CEC00BF, 0x9CEC00829CEC0081, 0x9CEC00849CEC0083,
-    0x9CEC00869CEC0085, 0x9CEC00889CEC0087, 0x9CEC008A9CEC0089,
-    0x9CEC008C9CEC008B, 0x9CEC008E9CEC008D, 0x9CEC00909CEC008F,
-    0x9CEC00929CEC0091, 0x9CEC00949CEC0093, 0x9CEC00969CEC0095,
-    0x9CEC00989CEC0097, 0x9CEC009A9CEC0099, 0x9CEC009C9CEC009B,
-    0x9CEC009E9CEC009D, 0x9CEC00A09CEC009F, 0x9CEC00A29CEC00A1,
-    0x9CEC00A49CEC00A3, 0x9CEC00A69CEC00A5, 0x9CEC00A89CEC00A7,
-    0x9CEC00AA9CEC00A9, 0x9CEC00AC9CEC00AB, 0x9CEC00AE9CEC00AD,
-    0x9CEC00B09CEC00AF, 0x9CEC00B29CEC00B1, 0x9CEC00B49CEC00B3,
-    0x9CEC00B69CEC00B5, 0x9CEC00B89CEC00B7, 0x9CEC00BA9CEC00B9,
-    0x9CEC00BC9CEC00BB, 0x9CEC00BE9CEC00BD, 0x9DEC00809DEC00BF,
-    0x9DEC00829DEC0081, 0x9DEC00849DEC0083, 0x9DEC00869DEC0085,
-    0x9DEC00889DEC0087, 0x9DEC008A9DEC0089, 0x9DEC008C9DEC008B,
-    0x9DEC008E9DEC008D, 0x9DEC00909DEC008F, 0x9DEC00929DEC0091,
-    0x9DEC00949DEC0093, 0x9DEC00969DEC0095, 0x9DEC00989DEC0097,
-    0x9DEC009A9DEC0099, 0x9DEC009C9DEC009B, 0x9DEC009E9DEC009D,
-    0x9DEC00A09DEC009F, 0x9DEC00A29DEC00A1, 0x9DEC00A49DEC00A3,
-    0x9DEC00A69DEC00A5, 0x9DEC00A89DEC00A7, 0x9DEC00AA9DEC00A9,
-    0x9DEC00AC9DEC00AB, 0x9DEC00AE9DEC00AD, 0x9DEC00B09DEC00AF,
-    0x9DEC00B29DEC00B1, 0x9DEC00B49DEC00B3, 0x9DEC00B69DEC00B5,
-    0x9DEC00B89DEC00B7, 0x9DEC00BA9DEC00B9, 0x9DEC00BC9DEC00BB,
-    0x9DEC00BE9DEC00BD, 0x9EEC00809EEC00BF, 0x9EEC00829EEC0081,
-    0x9EEC00849EEC0083, 0x9EEC00869EEC0085, 0x9EEC00889EEC0087,
-    0x9EEC008A9EEC0089, 0x9EEC008C9EEC008B, 0x9EEC008E9EEC008D,
-    0x9EEC00909EEC008F, 0x9EEC00929EEC0091, 0x9EEC00949EEC0093,
-    0x9EEC00969EEC0095, 0x9EEC00989EEC0097, 0x9EEC009A9EEC0099,
-    0x9EEC009C9EEC009B, 0x9EEC009E9EEC009D, 0x9EEC00A09EEC009F,
-    0x9EEC00A29EEC00A1, 0x9EEC00A49EEC00A3, 0x9EEC00A69EEC00A5,
-    0x9EEC00A89EEC00A7, 0x9EEC00AA9EEC00A9, 0x9EEC00AC9EEC00AB,
-    0x9EEC00AE9EEC00AD, 0x9EEC00B09EEC00AF, 0x9EEC00B29EEC00B1,
-    0x9EEC00B49EEC00B3, 0x9EEC00B69EEC00B5, 0x9EEC00B89EEC00B7,
-    0x9EEC00BA9EEC00B9, 0x9EEC00BC9EEC00BB, 0x9EEC00BE9EEC00BD,
-    0x9FEC00809FEC00BF, 0x9FEC00829FEC0081, 0x9FEC00849FEC0083,
-    0x9FEC00869FEC0085, 0x9FEC00889FEC0087, 0x9FEC008A9FEC0089,
-    0x9FEC008C9FEC008B, 0x9FEC008E9FEC008D, 0x9FEC00909FEC008F,
-    0x9FEC00929FEC0091, 0x9FEC00949FEC0093, 0x9FEC00969FEC0095,
-    0x9FEC00989FEC0097, 0x9FEC009A9FEC0099, 0x9FEC009C9FEC009B,
-    0x9FEC009E9FEC009D, 0x9FEC00A09FEC009F, 0x9FEC00A29FEC00A1,
-    0x9FEC00A49FEC00A3, 0x9FEC00A69FEC00A5, 0x9FEC00A89FEC00A7,
-    0x9FEC00AA9FEC00A9, 0x9FEC00AC9FEC00AB, 0x9FEC00AE9FEC00AD,
-    0x9FEC00B09FEC00AF, 0x9FEC00B29FEC00B1, 0x9FEC00B49FEC00B3,
-    0x9FEC00B69FEC00B5, 0x9FEC00B89FEC00B7, 0x9FEC00BA9FEC00B9,
-    0x9FEC00BC9FEC00BB, 0x9FEC00BE9FEC00BD, 0xA0EC0080A0EC00BF,
-    0xA0EC0082A0EC0081, 0xA0EC0084A0EC0083, 0xA0EC0086A0EC0085,
-    0xA0EC0088A0EC0087, 0xA0EC008AA0EC0089, 0xA0EC008CA0EC008B,
-    0xA0EC008EA0EC008D, 0xA0EC0090A0EC008F, 0xA0EC0092A0EC0091,
-    0xA0EC0094A0EC0093, 0xA0EC0096A0EC0095, 0xA0EC0098A0EC0097,
-    0xA0EC009AA0EC0099, 0xA0EC009CA0EC009B, 0xA0EC009EA0EC009D,
-    0xA0EC00A0A0EC009F, 0xA0EC00A2A0EC00A1, 0xA0EC00A4A0EC00A3,
-    0xA0EC00A6A0EC00A5, 0xA0EC00A8A0EC00A7, 0xA0EC00AAA0EC00A9,
-    0xA0EC00ACA0EC00AB, 0xA0EC00AEA0EC00AD, 0xA0EC00B0A0EC00AF,
-    0xA0EC00B2A0EC00B1, 0xA0EC00B4A0EC00B3, 0xA0EC00B6A0EC00B5,
-    0xA0EC00B8A0EC00B7, 0xA0EC00BAA0EC00B9, 0xA0EC00BCA0EC00BB,
-    0xA0EC00BEA0EC00BD, 0xA1EC0080A1EC00BF, 0xA1EC0082A1EC0081,
-    0xA1EC0084A1EC0083, 0xA1EC0086A1EC0085, 0xA1EC0088A1EC0087,
-    0xA1EC008AA1EC0089, 0xA1EC008CA1EC008B, 0xA1EC008EA1EC008D,
-    0xA1EC0090A1EC008F, 0xA1EC0092A1EC0091, 0xA1EC0094A1EC0093,
-    0xA1EC0096A1EC0095, 0xA1EC0098A1EC0097, 0xA1EC009AA1EC0099,
-    0xA1EC009CA1EC009B, 0xA1EC009EA1EC009D, 0xA1EC00A0A1EC009F,
-    0xA1EC00A2A1EC00A1, 0xA1EC00A4A1EC00A3, 0xA1EC00A6A1EC00A5,
-    0xA1EC00A8A1EC00A7, 0xA1EC00AAA1EC00A9, 0xA1EC00ACA1EC00AB,
-    0xA1EC00AEA1EC00AD, 0xA1EC00B0A1EC00AF, 0xA1EC00B2A1EC00B1,
-    0xA1EC00B4A1EC00B3, 0xA1EC00B6A1EC00B5, 0xA1EC00B8A1EC00B7,
-    0xA1EC00BAA1EC00B9, 0xA1EC00BCA1EC00BB, 0xA1EC00BEA1EC00BD,
-    0xA2EC0080A2EC00BF, 0xA2EC0082A2EC0081, 0xA2EC0084A2EC0083,
-    0xA2EC0086A2EC0085, 0xA2EC0088A2EC0087, 0xA2EC008AA2EC0089,
-    0xA2EC008CA2EC008B, 0xA2EC008EA2EC008D, 0xA2EC0090A2EC008F,
-    0xA2EC0092A2EC0091, 0xA2EC0094A2EC0093, 0xA2EC0096A2EC0095,
-    0xA2EC0098A2EC0097, 0xA2EC009AA2EC0099, 0xA2EC009CA2EC009B,
-    0xA2EC009EA2EC009D, 0xA2EC00A0A2EC009F, 0xA2EC00A2A2EC00A1,
-    0xA2EC00A4A2EC00A3, 0xA2EC00A6A2EC00A5, 0xA2EC00A8A2EC00A7,
-    0xA2EC00AAA2EC00A9, 0xA2EC00ACA2EC00AB, 0xA2EC00AEA2EC00AD,
-    0xA2EC00B0A2EC00AF, 0xA2EC00B2A2EC00B1, 0xA2EC00B4A2EC00B3,
-    0xA2EC00B6A2EC00B5, 0xA2EC00B8A2EC00B7, 0xA2EC00BAA2EC00B9,
-    0xA2EC00BCA2EC00BB, 0xA2EC00BEA2EC00BD, 0xA3EC0080A3EC00BF,
-    0xA3EC0082A3EC0081, 0xA3EC0084A3EC0083, 0xA3EC0086A3EC0085,
-    0xA3EC0088A3EC0087, 0xA3EC008AA3EC0089, 0xA3EC008CA3EC008B,
-    0xA3EC008EA3EC008D, 0xA3EC0090A3EC008F, 0xA3EC0092A3EC0091,
-    0xA3EC0094A3EC0093, 0xA3EC0096A3EC0095, 0xA3EC0098A3EC0097,
-    0xA3EC009AA3EC0099, 0xA3EC009CA3EC009B, 0xA3EC009EA3EC009D,
-    0xA3EC00A0A3EC009F, 0xA3EC00A2A3EC00A1, 0xA3EC00A4A3EC00A3,
-    0xA3EC00A6A3EC00A5, 0xA3EC00A8A3EC00A7, 0xA3EC00AAA3EC00A9,
-    0xA3EC00ACA3EC00AB, 0xA3EC00AEA3EC00AD, 0xA3EC00B0A3EC00AF,
-    0xA3EC00B2A3EC00B1, 0xA3EC00B4A3EC00B3, 0xA3EC00B6A3EC00B5,
-    0xA3EC00B8A3EC00B7, 0xA3EC00BAA3EC00B9, 0xA3EC00BCA3EC00BB,
-    0xBDA3EC00989DECBC, 0xBFA3EC00BEA3EC00, 0x81A4EC0080A4EC00,
-    0x83A4EC0082A4EC00, 0x85A4EC0084A4EC00, 0x87A4EC0086A4EC00,
-    0x89A4EC0088A4EC00, 0x8BA4EC008AA4EC00, 0x8DA4EC008CA4EC00,
-    0x8FA4EC008EA4EC00, 0x91A4EC0090A4EC00, 0x93A4EC0092A4EC00,
-    0x95A4EC0094A4EC00, 0x97A4EC0096A4EC00, 0x99A4EC0098A4EC00,
-    0x9BA4EC009AA4EC00, 0x9DA4EC009CA4EC00, 0x9FA4EC009EA4EC00,
-    0xA1A4EC00A0A4EC00, 0xA3A4EC00A2A4EC00, 0xA5A4EC00A4A4EC00,
-    0xA7A4EC00A6A4EC00, 0xA9A4EC00A8A4EC00, 0xABA4EC00AAA4EC00,
-    0xADA4EC00ACA4EC00, 0xAFA4EC00AEA4EC00, 0xB1A4EC00B0A4EC00,
-    0xB3A4EC00B2A4EC00, 0xB5A4EC00B4A4EC00, 0xB7A4EC00B6A4EC00,
-    0xB9A4EC00B8A4EC00, 0xBBA4EC00BAA4EC00, 0xBDA4EC00BCA4EC00,
-    0xBFA4EC00BEA4EC00, 0x81A5EC0080A5EC00, 0x83A5EC0082A5EC00,
-    0x85A5EC0084A5EC00, 0x87A5EC0086A5EC00, 0x89A5EC0088A5EC00,
-    0x8BA5EC008AA5EC00, 0x8DA5EC008CA5EC00, 0x8FA5EC008EA5EC00,
-    0x91A5EC0090A5EC00, 0x93A5EC0092A5EC00, 0x95A5EC0094A5EC00,
-    0x97A5EC0096A5EC00, 0x99A5EC0098A5EC00, 0x9BA5EC009AA5EC00,
-    0x9DA5EC009CA5EC00, 0x9FA5EC009EA5EC00, 0xA1A5EC00A0A5EC00,
-    0xA3A5EC00A2A5EC00, 0xA5A5EC00A4A5EC00, 0xA7A5EC00A6A5EC00,
-    0xA9A5EC00A8A5EC00, 0xABA5EC00AAA5EC00, 0xADA5EC00ACA5EC00,
-    0xAFA5EC00AEA5EC00, 0xB1A5EC00B0A5EC00, 0xB3A5EC00B2A5EC00,
-    0xB5A5EC00B4A5EC00, 0xB7A5EC00B6A5EC00, 0xB9A5EC00B8A5EC00,
-    0xBBA5EC00BAA5EC00, 0xBDA5EC00BCA5EC00, 0xBFA5EC00BEA5EC00,
-    0x81A6EC0080A6EC00, 0x83A6EC0082A6EC00, 0x85A6EC0084A6EC00,
-    0x87A6EC0086A6EC00, 0x89A6EC0088A6EC00, 0x8BA6EC008AA6EC00,
-    0x8DA6EC008CA6EC00, 0x8FA6EC008EA6EC00, 0x91A6EC0090A6EC00,
-    0x93A6EC0092A6EC00, 0x95A6EC0094A6EC00, 0x97A6EC0096A6EC00,
-    0x99A6EC0098A6EC00, 0x9BA6EC009AA6EC00, 0x9DA6EC009CA6EC00,
-    0x9FA6EC009EA6EC00, 0xA1A6EC00A0A6EC00, 0xA3A6EC00A2A6EC00,
-    0xA5A6EC00A4A6EC00, 0xA7A6EC00A6A6EC00, 0xA9A6EC00A8A6EC00,
-    0xABA6EC00AAA6EC00, 0xADA6EC00ACA6EC00, 0xAFA6EC00AEA6EC00,
-    0xB1A6EC00B0A6EC00, 0xB3A6EC00B2A6EC00, 0xB5A6EC00B4A6EC00,
-    0xB7A6EC00B6A6EC00, 0xB9A6EC00B8A6EC00, 0xBBA6EC00BAA6EC00,
-    0xBDA6EC00BCA6EC00, 0xBFA6EC00BEA6EC00, 0x81A7EC0080A7EC00,
-    0x83A7EC0082A7EC00, 0x85A7EC0084A7EC00, 0x87A7EC0086A7EC00,
-    0x89A7EC0088A7EC00, 0x8BA7EC008AA7EC00, 0x8DA7EC008CA7EC00,
-    0x8FA7EC008EA7EC00, 0x91A7EC0090A7EC00, 0x93A7EC0092A7EC00,
-    0x95A7EC0094A7EC00, 0x97A7EC0096A7EC00, 0x99A7EC0098A7EC00,
-    0x9BA7EC009AA7EC00, 0x9DA7EC009CA7EC00, 0x9FA7EC009EA7EC00,
-    0xA1A7EC00A0A7EC00, 0xA3A7EC00A2A7EC00, 0xA5A7EC00A4A7EC00,
-    0xA7A7EC00A6A7EC00, 0xA9A7EC00A8A7EC00, 0xABA7EC00AAA7EC00,
-    0xADA7EC00ACA7EC00, 0xAFA7EC00AEA7EC00, 0xB1A7EC00B0A7EC00,
-    0xB3A7EC00B2A7EC00, 0xB5A7EC00B4A7EC00, 0xB7A7EC00B6A7EC00,
-    0xB9A7EC00B8A7EC00, 0xBBA7EC00BAA7EC00, 0xBDA7EC00BCA7EC00,
-    0xBFA7EC00BEA7EC00, 0x81A8EC0080A8EC00, 0x83A8EC0082A8EC00,
-    0x85A8EC0084A8EC00, 0x87A8EC0086A8EC00, 0x89A8EC0088A8EC00,
-    0x8BA8EC008AA8EC00, 0x8DA8EC008CA8EC00, 0x8FA8EC008EA8EC00,
-    0x91A8EC0090A8EC00, 0x93A8EC0092A8EC00, 0x95A8EC0094A8EC00,
-    0x97A8EC0096A8EC00, 0x99A8EC0098A8EC00, 0x9BA8EC009AA8EC00,
-    0x9DA8EC009CA8EC00, 0x9FA8EC009EA8EC00, 0xA1A8EC00A0A8EC00,
-    0xA3A8EC00A2A8EC00, 0xA5A8EC00A4A8EC00, 0xA7A8EC00A6A8EC00,
-    0xA9A8EC00A8A8EC00, 0xABA8EC00AAA8EC00, 0xADA8EC00ACA8EC00,
-    0xAFA8EC00AEA8EC00, 0xB1A8EC00B0A8EC00, 0xB3A8EC00B2A8EC00,
-    0xB5A8EC00B4A8EC00, 0xB7A8EC00B6A8EC00, 0xB9A8EC00B8A8EC00,
-    0xBBA8EC00BAA8EC00, 0xBDA8EC00BCA8EC00, 0xBFA8EC00BEA8EC00,
-    0x81A9EC0080A9EC00, 0x83A9EC0082A9EC00, 0x85A9EC0084A9EC00,
-    0x87A9EC0086A9EC00, 0x89A9EC0088A9EC00, 0x8BA9EC008AA9EC00,
-    0x8DA9EC008CA9EC00, 0x8FA9EC008EA9EC00, 0x91A9EC0090A9EC00,
-    0x93A9EC0092A9EC00, 0x95A9EC0094A9EC00, 0x97A9EC0096A9EC00,
-    0x99A9EC0098A9EC00, 0x9BA9EC009AA9EC00, 0x9DA9EC009CA9EC00,
-    0x9FA9EC009EA9EC00, 0xA1A9EC00A0A9EC00, 0xA3A9EC00A2A9EC00,
-    0xA5A9EC00A4A9EC00, 0xA7A9EC00A6A9EC00, 0xA9A9EC00A8A9EC00,
-    0xABA9EC00AAA9EC00, 0xADA9EC00ACA9EC00, 0xAFA9EC00AEA9EC00,
-    0xB1A9EC00B0A9EC00, 0xB3A9EC00B2A9EC00, 0xB5A9EC00B4A9EC00,
-    0xB7A9EC00B6A9EC00, 0xB9A9EC00B8A9EC00, 0xBBA9EC00BAA9EC00,
-    0xBDA9EC00BCA9EC00, 0xBFA9EC00BEA9EC00, 0x81AAEC0080AAEC00,
-    0x83AAEC0082AAEC00, 0x85AAEC0084AAEC00, 0x87AAEC0086AAEC00,
-    0x89AAEC0088AAEC00, 0x8BAAEC008AAAEC00, 0x8DAAEC008CAAEC00,
-    0x8FAAEC008EAAEC00, 0x91AAEC0090AAEC00, 0x93AAEC0092AAEC00,
-    0x95AAEC0094AAEC00, 0x97AAEC0096AAEC00, 0x99AAEC0098AAEC00,
-    0x9BAAEC009AAAEC00, 0x9DAAEC009CAAEC00, 0x9FAAEC009EAAEC00,
-    0xA1AAEC00A0AAEC00, 0xA3AAEC00A2AAEC00, 0xA5AAEC00A4AAEC00,
-    0xA7AAEC00A6AAEC00, 0xA9AAEC00A8AAEC00, 0xABAAEC00AAAAEC00,
-    0xADAAEC00ACAAEC00, 0xAFAAEC00AEAAEC00, 0xB1AAEC00B0AAEC00,
-    0xB3AAEC00B2AAEC00, 0xB5AAEC00B4AAEC00, 0xB7AAEC00B6AAEC00,
-    0xB9AAEC00B8AAEC00, 0xBBAAEC00BAAAEC00, 0xBDAAEC00BCAAEC00,
-    0xBFAAEC00BEAAEC00, 0x81ABEC0080ABEC00, 0x83ABEC0082ABEC00,
-    0x85ABEC0084ABEC00, 0x87ABEC0086ABEC00, 0x89ABEC0088ABEC00,
-    0x8BABEC008AABEC00, 0x8DABEC008CABEC00, 0x8FABEC008EABEC00,
-    0x91ABEC0090ABEC00, 0x93ABEC0092ABEC00, 0x95ABEC0094ABEC00,
-    0x97ABEC0096ABEC00, 0x99ABEC0098ABEC00, 0x9BABEC009AABEC00,
-    0x9DABEC009CABEC00, 0x9FABEC009EABEC00, 0xA1ABEC00A0ABEC00,
-    0xA3ABEC00A2ABEC00, 0xA5ABEC00A4ABEC00, 0xA7ABEC00A6ABEC00,
-    0xA9ABEC00A8ABEC00, 0xABABEC00AAABEC00, 0xADABEC00ACABEC00,
-    0xAFABEC00AEABEC00, 0xB1ABEC00B0ABEC00, 0xB3ABEC00B2ABEC00,
-    0xB5ABEC00B4ABEC00, 0xB7ABEC00B6ABEC00, 0xB9ABEC00B8ABEC00,
-    0xBBABEC00BAABEC00, 0xBDABEC00BCABEC00, 0xBFABEC00BEABEC00,
-    0x81ACEC0080ACEC00, 0x83ACEC0082ACEC00, 0x85ACEC0084ACEC00,
-    0x87ACEC0086ACEC00, 0x89ACEC0088ACEC00, 0x8BACEC008AACEC00,
-    0x8DACEC008CACEC00, 0x8FACEC008EACEC00, 0x91ACEC0090ACEC00,
-    0x93ACEC0092ACEC00, 0x95ACEC0094ACEC00, 0x97ACEC0096ACEC00,
-    0x99ACEC0098ACEC00, 0x9BACEC009AACEC00, 0x9DACEC009CACEC00,
-    0x9FACEC009EACEC00, 0xA1ACEC00A0ACEC00, 0xA3ACEC00A2ACEC00,
-    0xA5ACEC00A4ACEC00, 0xA7ACEC00A6ACEC00, 0xA9ACEC00A8ACEC00,
-    0xABACEC00AAACEC00, 0xADACEC00ACACEC00, 0xAFACEC00AEACEC00,
-    0xB1ACEC00B0ACEC00, 0xB3ACEC00B2ACEC00, 0xB5ACEC00B4ACEC00,
-    0xB7ACEC00B6ACEC00, 0xB9ACEC00B8ACEC00, 0xBBACEC00BAACEC00,
-    0xBDACEC00BCACEC00, 0xBFACEC00BEACEC00, 0x81ADEC0080ADEC00,
-    0x83ADEC0082ADEC00, 0x85ADEC0084ADEC00, 0x87ADEC0086ADEC00,
-    0x89ADEC0088ADEC00, 0x8BADEC008AADEC00, 0x8DADEC008CADEC00,
-    0x8FADEC008EADEC00, 0x91ADEC0090ADEC00, 0x93ADEC0092ADEC00,
-    0x95ADEC0094ADEC00, 0x97ADEC0096ADEC00, 0x99ADEC0098ADEC00,
-    0x9BADEC009AADEC00, 0x9DADEC009CADEC00, 0x9FADEC009EADEC00,
-    0xA1ADEC00A0ADEC00, 0xA3ADEC00A2ADEC00, 0xA5ADEC00A4ADEC00,
-    0xA7ADEC00A6ADEC00, 0xA9ADEC00A8ADEC00, 0xABADEC00AAADEC00,
-    0xADADEC00ACADEC00, 0xAFADEC00AEADEC00, 0xB1ADEC00B0ADEC00,
-    0xB3ADEC00B2ADEC00, 0xB5ADEC00B4ADEC00, 0xB7ADEC00B6ADEC00,
-    0xB9ADEC00B8ADEC00, 0xBBADEC00BAADEC00, 0xBDADEC00BCADEC00,
-    0xBFADEC00BEADEC00, 0x81AEEC0080AEEC00, 0x83AEEC0082AEEC00,
-    0x85AEEC0084AEEC00, 0x87AEEC0086AEEC00, 0x89AEEC0088AEEC00,
-    0x8BAEEC008AAEEC00, 0x8DAEEC008CAEEC00, 0x8FAEEC008EAEEC00,
-    0x91AEEC0090AEEC00, 0x93AEEC0092AEEC00, 0x95AEEC0094AEEC00,
-    0x97AEEC0096AEEC00, 0x99AEEC0098AEEC00, 0x9BAEEC009AAEEC00,
-    0x9DAEEC009CAEEC00, 0x9FAEEC009EAEEC00, 0xA1AEEC00A0AEEC00,
-    0xA3AEEC00A2AEEC00, 0xA5AEEC00A4AEEC00, 0xA7AEEC00A6AEEC00,
-    0xA9AEEC00A8AEEC00, 0xABAEEC00AAAEEC00, 0xADAEEC00ACAEEC00,
-    0xAFAEEC00AEAEEC00, 0xB1AEEC00B0AEEC00, 0xB3AEEC00B2AEEC00,
-    0xB5AEEC00B4AEEC00, 0xB7AEEC00B6AEEC00, 0xB9AEEC00B8AEEC00,
-    0xBBAEEC00BAAEEC00, 0xBDAEEC00BCAEEC00, 0xBFAEEC00BEAEEC00,
-    0x81AFEC0080AFEC00, 0x83AFEC0082AFEC00, 0x85AFEC0084AFEC00,
-    0x87AFEC0086AFEC00, 0x89AFEC0088AFEC00, 0x8BAFEC008AAFEC00,
-    0x8DAFEC008CAFEC00, 0x8FAFEC008EAFEC00, 0x91AFEC0090AFEC00,
-    0x93AFEC0092AFEC00, 0x95AFEC0094AFEC00, 0x97AFEC0096AFEC00,
-    0x99AFEC0098AFEC00, 0x9BAFEC009AAFEC00, 0x9DAFEC009CAFEC00,
-    0x9FAFEC009EAFEC00, 0xA1AFEC00A0AFEC00, 0xA3AFEC00A2AFEC00,
-    0xA5AFEC00A4AFEC00, 0xA7AFEC00A6AFEC00, 0xA9AFEC00A8AFEC00,
-    0xABAFEC00AAAFEC00, 0xADAFEC00ACAFEC00, 0xAFAFEC00AEAFEC00,
-    0xB1AFEC00B0AFEC00, 0xB3AFEC00B2AFEC00, 0xB5AFEC00B4AFEC00,
-    0xB7AFEC00B6AFEC00, 0xB9AFEC00B8AFEC00, 0xBBAFEC00BAAFEC00,
-    0xBDAFEC00BCAFEC00, 0xBFAFEC00BEAFEC00, 0x81B0EC0080B0EC00,
-    0x83B0EC0082B0EC00, 0x85B0EC0084B0EC00, 0x87B0EC0086B0EC00,
-    0x89B0EC0088B0EC00, 0x8BB0EC008AB0EC00, 0x8DB0EC008CB0EC00,
-    0x8FB0EC008EB0EC00, 0x91B0EC0090B0EC00, 0x93B0EC0092B0EC00,
-    0x95B0EC0094B0EC00, 0x97B0EC0096B0EC00, 0x99B0EC0098B0EC00,
-    0x9BB0EC009AB0EC00, 0x9DB0EC009CB0EC00, 0x9FB0EC009EB0EC00,
-    0xA1B0EC00A0B0EC00, 0xA3B0EC00A2B0EC00, 0xA5B0EC00A4B0EC00,
-    0xA7B0EC00A6B0EC00, 0xA9B0EC00A8B0EC00, 0xABB0EC00AAB0EC00,
-    0xADB0EC00ACB0EC00, 0xAFB0EC00AEB0EC00, 0xB1B0EC00B0B0EC00,
-    0xB3B0EC00B2B0EC00, 0xB5B0EC00B4B0EC00, 0xB7B0EC00B6B0EC00,
-    0xB8B0EC00B8B0EC00, 0x00B9B0EC00A0B3EA, 0x00BBB0EC00BAB0EC,
-    0x00BDB0EC00BCB0EC, 0x00BFB0EC00BEB0EC, 0x0081B1EC0080B1EC,
-    0x0083B1EC0082B1EC, 0x0085B1EC0084B1EC, 0x0087B1EC0086B1EC,
-    0x0089B1EC0088B1EC, 0x008BB1EC008AB1EC, 0x008DB1EC008CB1EC,
-    0x008FB1EC008EB1EC, 0x0091B1EC0090B1EC, 0x0093B1EC0092B1EC,
-    0x0095B1EC0094B1EC, 0x0097B1EC0096B1EC, 0x0099B1EC0098B1EC,
-    0x009BB1EC009AB1EC, 0x009DB1EC009CB1EC, 0x009FB1EC009EB1EC,
-    0x00A1B1EC00A0B1EC, 0x00A3B1EC00A2B1EC, 0x00A5B1EC00A4B1EC,
-    0x00A7B1EC00A6B1EC, 0x00A9B1EC00A8B1EC, 0x00ABB1EC00AAB1EC,
-    0x00ADB1EC00ACB1EC, 0x00AFB1EC00AEB1EC, 0x00B1B1EC00B0B1EC,
-    0x00B3B1EC00B2B1EC, 0x00B5B1EC00B4B1EC, 0x00B7B1EC00B6B1EC,
-    0x00B9B1EC00B8B1EC, 0x00BBB1EC00BAB1EC, 0x00BDB1EC00BCB1EC,
-    0x00BFB1EC00BEB1EC, 0x0081B2EC0080B2EC, 0x0083B2EC0082B2EC,
-    0x0085B2EC0084B2EC, 0x0087B2EC0086B2EC, 0x0089B2EC0088B2EC,
-    0x008BB2EC008AB2EC, 0x008DB2EC008CB2EC, 0x008FB2EC008EB2EC,
-    0x0091B2EC0090B2EC, 0x0093B2EC0092B2EC, 0x0095B2EC0094B2EC,
-    0x0097B2EC0096B2EC, 0x0099B2EC0098B2EC, 0x009BB2EC009AB2EC,
-    0x009DB2EC009CB2EC, 0x009FB2EC009EB2EC, 0x00A1B2EC00A0B2EC,
-    0x00A3B2EC00A2B2EC, 0x00A5B2EC00A4B2EC, 0x00A7B2EC00A6B2EC,
-    0x00A9B2EC00A8B2EC, 0x00ABB2EC00AAB2EC, 0x00ADB2EC00ACB2EC,
-    0x00AFB2EC00AEB2EC, 0x00B1B2EC00B0B2EC, 0x00B3B2EC00B2B2EC,
-    0x00B5B2EC00B4B2EC, 0x00B7B2EC00B6B2EC, 0x00B9B2EC00B8B2EC,
-    0x00BBB2EC00BAB2EC, 0x00BDB2EC00BCB2EC, 0x00BFB2EC00BEB2EC,
-    0x0081B3EC0080B3EC, 0x0083B3EC0082B3EC, 0x0085B3EC0084B3EC,
-    0x0087B3EC0086B3EC, 0x0089B3EC0088B3EC, 0x008BB3EC008AB3EC,
-    0x008DB3EC008CB3EC, 0x008FB3EC008EB3EC, 0x0091B3EC0090B3EC,
-    0x0093B3EC0092B3EC, 0x0095B3EC0094B3EC, 0x0097B3EC0096B3EC,
-    0x0099B3EC0098B3EC, 0x009BB3EC009AB3EC, 0x009DB3EC009CB3EC,
-    0x009FB3EC009EB3EC, 0x00A1B3EC00A0B3EC, 0x00A3B3EC00A2B3EC,
-    0x00A5B3EC00A4B3EC, 0x00A7B3EC00A6B3EC, 0x00A9B3EC00A8B3EC,
-    0x00ABB3EC00AAB3EC, 0x00ADB3EC00ACB3EC, 0x00AFB3EC00AEB3EC,
-    0x00B1B3EC00B0B3EC, 0x00B3B3EC00B2B3EC, 0x00B5B3EC00B4B3EC,
-    0x00B7B3EC00B6B3EC, 0x00B9B3EC00B8B3EC, 0x00BBB3EC00BAB3EC,
-    0x00BDB3EC00BCB3EC, 0x00BFB3EC00BEB3EC, 0x0081B4EC0080B4EC,
-    0x0083B4EC0082B4EC, 0x0085B4EC0084B4EC, 0x0087B4EC0086B4EC,
-    0x0089B4EC0088B4EC, 0x008BB4EC008AB4EC, 0x008DB4EC008CB4EC,
-    0x008FB4EC008EB4EC, 0x0091B4EC0090B4EC, 0x0093B4EC0092B4EC,
-    0x0095B4EC0094B4EC, 0x0097B4EC0096B4EC, 0x0099B4EC0098B4EC,
-    0x009BB4EC009AB4EC, 0x009DB4EC009CB4EC, 0x009FB4EC009EB4EC,
-    0x00A1B4EC00A0B4EC, 0x00A3B4EC00A2B4EC, 0x00A5B4EC00A4B4EC,
-    0x00A7B4EC00A6B4EC, 0x00A9B4EC00A8B4EC, 0x00ABB4EC00AAB4EC,
-    0x00ADB4EC00ACB4EC, 0x00AFB4EC00AEB4EC, 0x00B1B4EC00B0B4EC,
-    0x00B3B4EC00B2B4EC, 0x00B5B4EC00B4B4EC, 0x00B7B4EC00B6B4EC,
-    0x00B9B4EC00B8B4EC, 0x00BBB4EC00BAB4EC, 0x00BDB4EC00BCB4EC,
-    0x00BFB4EC00BEB4EC, 0x0081B5EC0080B5EC, 0x0083B5EC0082B5EC,
-    0x0085B5EC0084B5EC, 0x0087B5EC0086B5EC, 0x0089B5EC0088B5EC,
-    0x008BB5EC008AB5EC, 0x008DB5EC008CB5EC, 0x008FB5EC008EB5EC,
-    0x0091B5EC0090B5EC, 0x0093B5EC0092B5EC, 0x0095B5EC0094B5EC,
-    0x0097B5EC0096B5EC, 0x0099B5EC0098B5EC, 0x009BB5EC009AB5EC,
-    0x009DB5EC009CB5EC, 0x009FB5EC009EB5EC, 0x00A1B5EC00A0B5EC,
-    0x00A3B5EC00A2B5EC, 0x00A5B5EC00A4B5EC, 0x00A7B5EC00A6B5EC,
-    0x00A9B5EC00A8B5EC, 0x00ABB5EC00AAB5EC, 0x00ADB5EC00ACB5EC,
-    0x00AFB5EC00AEB5EC, 0x00B1B5EC00B0B5EC, 0x00B3B5EC00B2B5EC,
-    0x00B5B5EC00B4B5EC, 0x00B7B5EC00B6B5EC, 0x00B9B5EC00B8B5EC,
-    0x00BBB5EC00BAB5EC, 0x00BDB5EC00BCB5EC, 0x00BFB5EC00BEB5EC,
-    0x0081B6EC0080B6EC, 0x0083B6EC0082B6EC, 0x0085B6EC0084B6EC,
-    0x0087B6EC0086B6EC, 0x0089B6EC0088B6EC, 0x008BB6EC008AB6EC,
-    0x008DB6EC008CB6EC, 0x008FB6EC008EB6EC, 0x0091B6EC0090B6EC,
-    0x0093B6EC0092B6EC, 0x0095B6EC0094B6EC, 0x0097B6EC0096B6EC,
-    0x0099B6EC0098B6EC, 0x009BB6EC009AB6EC, 0x009DB6EC009CB6EC,
-    0x009FB6EC009EB6EC, 0x00A1B6EC00A0B6EC, 0x00A3B6EC00A2B6EC,
-    0x00A5B6EC00A4B6EC, 0x00A7B6EC00A6B6EC, 0x00A9B6EC00A8B6EC,
-    0x00ABB6EC00AAB6EC, 0x00ADB6EC00ACB6EC, 0x00AFB6EC00AEB6EC,
-    0x00B1B6EC00B0B6EC, 0x00B3B6EC00B2B6EC, 0x00B5B6EC00B4B6EC,
-    0x00B7B6EC00B6B6EC, 0x00B9B6EC00B8B6EC, 0x00BBB6EC00BAB6EC,
-    0x00BDB6EC00BCB6EC, 0x00BFB6EC00BEB6EC, 0x0081B7EC0080B7EC,
-    0x0083B7EC0082B7EC, 0x0085B7EC0084B7EC, 0x0087B7EC0086B7EC,
-    0x0089B7EC0088B7EC, 0x008BB7EC008AB7EC, 0x008DB7EC008CB7EC,
-    0x008FB7EC008EB7EC, 0x0091B7EC0090B7EC, 0x0093B7EC0092B7EC,
-    0x0095B7EC0094B7EC, 0x0097B7EC0096B7EC, 0x0099B7EC0098B7EC,
-    0x009BB7EC009AB7EC, 0x009DB7EC009CB7EC, 0x009FB7EC009EB7EC,
-    0x00A1B7EC00A0B7EC, 0x00A3B7EC00A2B7EC, 0x00A5B7EC00A4B7EC,
-    0x00A7B7EC00A6B7EC, 0x00A9B7EC00A8B7EC, 0x00ABB7EC00AAB7EC,
-    0x00ADB7EC00ACB7EC, 0x00AFB7EC00AEB7EC, 0x00B1B7EC00B0B7EC,
-    0x00B3B7EC00B2B7EC, 0x00B5B7EC00B4B7EC, 0x00B7B7EC00B6B7EC,
-    0x00B9B7EC00B8B7EC, 0x00BBB7EC00BAB7EC, 0x00BDB7EC00BCB7EC,
-    0x00BFB7EC00BEB7EC, 0x0081B8EC0080B8EC, 0x0083B8EC0082B8EC,
-    0x0085B8EC0084B8EC, 0x0087B8EC0086B8EC, 0x0089B8EC0088B8EC,
-    0x008BB8EC008AB8EC, 0x008DB8EC008CB8EC, 0x008FB8EC008EB8EC,
-    0x0091B8EC0090B8EC, 0x0093B8EC0092B8EC, 0x0095B8EC0094B8EC,
-    0x0097B8EC0096B8EC, 0x0099B8EC0098B8EC, 0x009BB8EC009AB8EC,
-    0x009DB8EC009CB8EC, 0x009FB8EC009EB8EC, 0x00A1B8EC00A0B8EC,
-    0x00A3B8EC00A2B8EC, 0x00A5B8EC00A4B8EC, 0x00A7B8EC00A6B8EC,
-    0x00A9B8EC00A8B8EC, 0x00ABB8EC00AAB8EC, 0x00ADB8EC00ACB8EC,
-    0x00AFB8EC00AEB8EC, 0x00B1B8EC00B0B8EC, 0x00B3B8EC00B2B8EC,
-    0x00B5B8EC00B4B8EC, 0x00B7B8EC00B6B8EC, 0x00B9B8EC00B8B8EC,
-    0x00BBB8EC00BAB8EC, 0x00BDB8EC00BCB8EC, 0x00BFB8EC00BEB8EC,
-    0x0081B9EC0080B9EC, 0x0083B9EC0082B9EC, 0x0085B9EC0084B9EC,
-    0x0087B9EC0086B9EC, 0x0089B9EC0088B9EC, 0x008BB9EC008AB9EC,
-    0x008DB9EC008CB9EC, 0x008FB9EC008EB9EC, 0x0091B9EC0090B9EC,
-    0x0093B9EC0092B9EC, 0x0095B9EC0094B9EC, 0x0097B9EC0096B9EC,
-    0x0099B9EC0098B9EC, 0x009BB9EC009AB9EC, 0x009DB9EC009CB9EC,
-    0x009FB9EC009EB9EC, 0x00A1B9EC00A0B9EC, 0x00A3B9EC00A2B9EC,
-    0x00A5B9EC00A4B9EC, 0x00A7B9EC00A6B9EC, 0x00A9B9EC00A8B9EC,
-    0x00ABB9EC00AAB9EC, 0x00ADB9EC00ACB9EC, 0x00AFB9EC00AEB9EC,
-    0x00B1B9EC00B0B9EC, 0x00B3B9EC00B2B9EC, 0x00B5B9EC00B4B9EC,
-    0x00B7B9EC00B6B9EC, 0x00B9B9EC00B8B9EC, 0x00BBB9EC00BAB9EC,
-    0x00BDB9EC00BCB9EC, 0x00BFB9EC00BEB9EC, 0x0081BAEC0080BAEC,
-    0x0083BAEC0082BAEC, 0x0085BAEC0084BAEC, 0x0087BAEC0086BAEC,
-    0x0089BAEC0088BAEC, 0x008BBAEC008ABAEC, 0x008DBAEC008CBAEC,
-    0x008FBAEC008EBAEC, 0x0091BAEC0090BAEC, 0x0093BAEC0092BAEC,
-    0x0095BAEC0094BAEC, 0x0097BAEC0096BAEC, 0x0099BAEC0098BAEC,
-    0x009BBAEC009ABAEC, 0x009DBAEC009CBAEC, 0x009FBAEC009EBAEC,
-    0x00A1BAEC00A0BAEC, 0x00A3BAEC00A2BAEC, 0x00A5BAEC00A4BAEC,
-    0x00A7BAEC00A6BAEC, 0x00A9BAEC00A8BAEC, 0x00ABBAEC00AABAEC,
-    0x00ADBAEC00ACBAEC, 0x00AFBAEC00AEBAEC, 0x00B1BAEC00B0BAEC,
-    0x00B3BAEC00B2BAEC, 0x00B5BAEC00B4BAEC, 0x00B7BAEC00B6BAEC,
-    0x00B9BAEC00B8BAEC, 0x00BBBAEC00BABAEC, 0x00BDBAEC00BCBAEC,
-    0x00BFBAEC00BEBAEC, 0x0081BBEC0080BBEC, 0x0083BBEC0082BBEC,
-    0x0085BBEC0084BBEC, 0x0087BBEC0086BBEC, 0x0089BBEC0088BBEC,
-    0x008BBBEC008ABBEC, 0x008DBBEC008CBBEC, 0x008FBBEC008EBBEC,
-    0x0091BBEC0090BBEC, 0x0093BBEC0092BBEC, 0x0095BBEC0094BBEC,
-    0x0097BBEC0096BBEC, 0x0099BBEC0098BBEC, 0x009BBBEC009ABBEC,
-    0x009DBBEC009CBBEC, 0x009FBBEC009EBBEC, 0x00A1BBEC00A0BBEC,
-    0x00A3BBEC00A2BBEC, 0x00A5BBEC00A4BBEC, 0x00A7BBEC00A6BBEC,
-    0x00A9BBEC00A8BBEC, 0x00ABBBEC00AABBEC, 0x00ADBBEC00ACBBEC,
-    0x00AFBBEC00AEBBEC, 0x00B1BBEC00B0BBEC, 0x00B3BBEC00B2BBEC,
-    0x00B5BBEC00B4BBEC, 0x00B7BBEC00B6BBEC, 0x00B9BBEC00B8BBEC,
-    0x00BBBBEC00BABBEC, 0x00BDBBEC00BCBBEC, 0x00BFBBEC00BEBBEC,
-    0x0081BCEC0080BCEC, 0x0083BCEC0082BCEC, 0x0085BCEC0084BCEC,
-    0x0087BCEC0086BCEC, 0x0089BCEC0088BCEC, 0x008BBCEC008ABCEC,
-    0x008DBCEC008CBCEC, 0x008FBCEC008EBCEC, 0x0091BCEC0090BCEC,
-    0x0093BCEC0092BCEC, 0x0095BCEC0094BCEC, 0x0097BCEC0096BCEC,
-    0x0099BCEC0098BCEC, 0x009BBCEC009ABCEC, 0x009DBCEC009CBCEC,
-    0x009FBCEC009EBCEC, 0x00A1BCEC00A0BCEC, 0x00A3BCEC00A2BCEC,
-    0x00A5BCEC00A4BCEC, 0x00A7BCEC00A6BCEC, 0x00A9BCEC00A8BCEC,
-    0x00ABBCEC00AABCEC, 0x00ADBCEC00ACBCEC, 0x00AFBCEC00AEBCEC,
-    0x00B1BCEC00B0BCEC, 0x00B3BCEC00B2BCEC, 0x00B5BCEC00B4BCEC,
-    0x00B7BCEC00B6BCEC, 0x00B9BCEC00B8BCEC, 0x00BBBCEC00BABCEC,
-    0x00BDBCEC00BCBCEC, 0x00BFBCEC00BEBCEC, 0x0081BDEC0080BDEC,
-    0x0083BDEC0082BDEC, 0x0085BDEC0084BDEC, 0x0087BDEC0086BDEC,
-    0x0089BDEC0088BDEC, 0x008BBDEC008ABDEC, 0x008DBDEC008CBDEC,
-    0x008FBDEC008EBDEC, 0x0091BDEC0090BDEC, 0x0093BDEC0092BDEC,
-    0x0095BDEC0094BDEC, 0x0097BDEC0096BDEC, 0x0099BDEC0098BDEC,
-    0x009BBDEC009ABDEC, 0x009DBDEC009CBDEC, 0x009FBDEC009EBDEC,
-    0x00A1BDEC00A0BDEC, 0x00A3BDEC00A2BDEC, 0x00A5BDEC00A4BDEC,
-    0x00A7BDEC00A6BDEC, 0x00A9BDEC00A8BDEC, 0x00ABBDEC00AABDEC,
-    0x00ADBDEC00ACBDEC, 0x00AFBDEC00AEBDEC, 0x00B1BDEC00B0BDEC,
-    0x00B3BDEC00B2BDEC, 0x00B5BDEC00B4BDEC, 0x00B7BDEC00B6BDEC,
-    0x00B9BDEC00B8BDEC, 0x00BBBDEC00BABDEC, 0x00BDBDEC00BCBDEC,
-    0x00BFBDEC00BEBDEC, 0x0081BEEC0080BEEC, 0x0083BEEC0082BEEC,
-    0x0085BEEC0084BEEC, 0x0087BEEC0086BEEC, 0x0089BEEC0088BEEC,
-    0x008BBEEC008ABEEC, 0x008DBEEC008CBEEC, 0x008FBEEC008EBEEC,
-    0x0091BEEC0090BEEC, 0x0093BEEC0092BEEC, 0x0095BEEC0094BEEC,
-    0x0097BEEC0096BEEC, 0x0099BEEC0098BEEC, 0x009BBEEC009ABEEC,
-    0x009DBEEC009CBEEC, 0x009FBEEC009EBEEC, 0x00A1BEEC00A0BEEC,
-    0x00A3BEEC00A2BEEC, 0x00A5BEEC00A4BEEC, 0x00A7BEEC00A6BEEC,
-    0x00A9BEEC00A8BEEC, 0x00ABBEEC00AABEEC, 0x00ADBEEC00ACBEEC,
-    0x00AFBEEC00AEBEEC, 0x00B1BEEC00B0BEEC, 0x00B3BEEC00B2BEEC,
-    0x00B5BEEC00B4BEEC, 0x00B7BEEC00B6BEEC, 0x00B9BEEC00B8BEEC,
-    0x00BBBEEC00BABEEC, 0x00BDBEEC00BCBEEC, 0x00BFBEEC00BEBEEC,
-    0x0081BFEC0080BFEC, 0x0083BFEC0082BFEC, 0x0085BFEC0084BFEC,
-    0x0087BFEC0086BFEC, 0x0089BFEC0088BFEC, 0x008BBFEC008ABFEC,
-    0x008DBFEC008CBFEC, 0x008FBFEC008EBFEC, 0x0091BFEC0090BFEC,
-    0x0093BFEC0092BFEC, 0x0095BFEC0094BFEC, 0x0097BFEC0096BFEC,
-    0x0099BFEC0098BFEC, 0x009BBFEC009ABFEC, 0x009DBFEC009CBFEC,
-    0x009FBFEC009EBFEC, 0x00A1BFEC00A0BFEC, 0x00A3BFEC00A2BFEC,
-    0x00A5BFEC00A4BFEC, 0x00A7BFEC00A6BFEC, 0x00A9BFEC00A8BFEC,
-    0x00ABBFEC00AABFEC, 0x00ADBFEC00ACBFEC, 0x00AFBFEC00AEBFEC,
-    0x00B1BFEC00B0BFEC, 0x00B3BFEC00B2BFEC, 0x00B5BFEC00B4BFEC,
-    0x00B7BFEC00B6BFEC, 0x00B9BFEC00B8BFEC, 0x00BBBFEC00BABFEC,
-    0x00BDBFEC00BCBFEC, 0x00BFBFEC00BEBFEC, 0x008180ED008080ED,
-    0x008380ED008280ED, 0x008580ED008480ED, 0x008780ED008680ED,
-    0x008980ED008880ED, 0x008B80ED008A80ED, 0x008D80ED008C80ED,
-    0x008F80ED008E80ED, 0x009180ED009080ED, 0x009380ED009280ED,
-    0x009580ED009480ED, 0x009780ED009680ED, 0x009980ED009880ED,
-    0x009B80ED009A80ED, 0x009D80ED009C80ED, 0x009F80ED009E80ED,
-    0x00A180ED00A080ED, 0x00A380ED00A280ED, 0x00A580ED00A480ED,
-    0x00A780ED00A680ED, 0x00A980ED00A880ED, 0x00AB80ED00AA80ED,
-    0x00AD80ED00AC80ED, 0x00AF80ED00AE80ED, 0x00B180ED00B080ED,
-    0x00B380ED00B280ED, 0x00B580ED00B480ED, 0x00B780ED00B680ED,
-    0x00B980ED00B880ED, 0x00BB80ED00BA80ED, 0x00BD80ED00BC80ED,
-    0x00BF80ED00BE80ED, 0x008181ED008081ED, 0x008381ED008281ED,
-    0x008581ED008481ED, 0x008781ED008681ED, 0x008981ED008881ED,
-    0x008B81ED008A81ED, 0x008D81ED008C81ED, 0x008F81ED008E81ED,
-    0x009181ED009081ED, 0x009381ED009281ED, 0x009581ED009481ED,
-    0x009781ED009681ED, 0x009981ED009881ED, 0x009B81ED009A81ED,
-    0x009D81ED009C81ED, 0x009F81ED009E81ED, 0x00A181ED00A081ED,
-    0x00A381ED00A281ED, 0x00A581ED00A481ED, 0x00A781ED00A681ED,
-    0x00A981ED00A881ED, 0x00AB81ED00AA81ED, 0x00AD81ED00AC81ED,
-    0x00AF81ED00AE81ED, 0x00B181ED00B081ED, 0x00B381ED00B281ED,
-    0x00B581ED00B481ED, 0x00B781ED00B681ED, 0x00B981ED00B881ED,
-    0x00BB81ED00BA81ED, 0x00BD81ED00BC81ED, 0x00BF81ED00BE81ED,
-    0x008182ED008082ED, 0x008382ED008282ED, 0x008582ED008482ED,
-    0x008782ED008682ED, 0x008982ED008882ED, 0x008B82ED008A82ED,
-    0x008D82ED008C82ED, 0x008F82ED008E82ED, 0x009182ED009082ED,
-    0x009382ED009282ED, 0x009582ED009482ED, 0x009782ED009682ED,
-    0x009982ED009882ED, 0x009B82ED009A82ED, 0x009D82ED009C82ED,
-    0x009F82ED009E82ED, 0x00A182ED00A082ED, 0x00A382ED00A282ED,
-    0x00A582ED00A482ED, 0x00A782ED00A682ED, 0x00A982ED00A882ED,
-    0x00AB82ED00AA82ED, 0x00AD82ED00AC82ED, 0x00AF82ED00AE82ED,
-    0x00B182ED00B082ED, 0x00B382ED00B282ED, 0x00B582ED00B482ED,
-    0x00B782ED00B682ED, 0x00B982ED00B882ED, 0x00BB82ED00BA82ED,
-    0x00BD82ED00BC82ED, 0x00BF82ED00BE82ED, 0x008183ED008083ED,
-    0x008383ED008283ED, 0x008583ED008483ED, 0x008783ED008683ED,
-    0x008983ED008883ED, 0x008B83ED008A83ED, 0x008D83ED008C83ED,
-    0x008F83ED008E83ED, 0x009183ED009083ED, 0x009383ED009283ED,
-    0x009583ED009483ED, 0x009783ED009683ED, 0x009983ED009883ED,
-    0x009B83ED009A83ED, 0x009D83ED009C83ED, 0x009F83ED009E83ED,
-    0x00A183ED00A083ED, 0x00A383ED00A283ED, 0x00A583ED00A483ED,
-    0x00A783ED00A683ED, 0x00A983ED00A883ED, 0x00AB83ED00AA83ED,
-    0x00AD83ED00AC83ED, 0x00AF83ED00AE83ED, 0x00B183ED00B083ED,
-    0x00B383ED00B283ED, 0x00B583ED00B483ED, 0x00B783ED00B683ED,
-    0x00B983ED00B883ED, 0x00BB83ED00BA83ED, 0x00BD83ED00BC83ED,
-    0x00BF83ED00BE83ED, 0x008184ED008084ED, 0x008384ED008284ED,
-    0x008584ED008484ED, 0x008784ED008684ED, 0x008984ED008884ED,
-    0x008B84ED008A84ED, 0x008D84ED008C84ED, 0x008F84ED008E84ED,
-    0x009184ED009084ED, 0x009384ED009284ED, 0x009584ED009484ED,
-    0x009784ED009684ED, 0x009984ED009884ED, 0x009B84ED009A84ED,
-    0x009D84ED009C84ED, 0x009F84ED009E84ED, 0x00A184ED00A084ED,
-    0x00A384ED00A284ED, 0x00A584ED00A484ED, 0x00A784ED00A684ED,
-    0x00A984ED00A884ED, 0x00AB84ED00AA84ED, 0x00AD84ED00AC84ED,
-    0x00AF84ED00AE84ED, 0x00B184ED00B084ED, 0x00B384ED00B284ED,
-    0x00B584ED00B484ED, 0x00B784ED00B684ED, 0x00B984ED00B884ED,
-    0x00BB84ED00BA84ED, 0x00BD84ED00BC84ED, 0x00BF84ED00BE84ED,
-    0x008185ED008085ED, 0x008385ED008285ED, 0x008585ED008485ED,
-    0x008785ED008685ED, 0x008985ED008885ED, 0x008B85ED008A85ED,
-    0x008D85ED008C85ED, 0x008F85ED008E85ED, 0x009185ED009085ED,
-    0x009385ED009285ED, 0x009585ED009485ED, 0x009785ED009685ED,
-    0x009985ED009885ED, 0x009B85ED009A85ED, 0x009D85ED009C85ED,
-    0x009F85ED009E85ED, 0x00A185ED00A085ED, 0x00A385ED00A285ED,
-    0x00A585ED00A485ED, 0x00A785ED00A685ED, 0x00A985ED00A885ED,
-    0x00AB85ED00AA85ED, 0x00AD85ED00AC85ED, 0x00AF85ED00AE85ED,
-    0x00B185ED00B085ED, 0x00B385ED00B285ED, 0x00B585ED00B485ED,
-    0x00B785ED00B685ED, 0x00B985ED00B885ED, 0x00BB85ED00BA85ED,
-    0x00BD85ED00BC85ED, 0x00BF85ED00BE85ED, 0x008186ED008086ED,
-    0x008386ED008286ED, 0x008586ED008486ED, 0x008786ED008686ED,
-    0x008986ED008886ED, 0x008B86ED008A86ED, 0x008D86ED008C86ED,
-    0x008F86ED008E86ED, 0x009186ED009086ED, 0x009386ED009286ED,
-    0x009586ED009486ED, 0x009786ED009686ED, 0x009986ED009886ED,
-    0x009B86ED009A86ED, 0x009D86ED009C86ED, 0x009F86ED009E86ED,
-    0x00A186ED00A086ED, 0x00A386ED00A286ED, 0x00A586ED00A486ED,
-    0x00A786ED00A686ED, 0x00A986ED00A886ED, 0x00AB86ED00AA86ED,
-    0x00AD86ED00AC86ED, 0x00AF86ED00AE86ED, 0x00B186ED00B086ED,
-    0x00B386ED00B286ED, 0x00B586ED00B486ED, 0x00B786ED00B686ED,
-    0x00B986ED00B886ED, 0x00BB86ED00BA86ED, 0x00BD86ED00BC86ED,
-    0x00BF86ED00BE86ED, 0x008187ED008087ED, 0x008387ED008287ED,
-    0x008587ED008487ED, 0x008787ED008687ED, 0x008987ED008887ED,
-    0x008B87ED008A87ED, 0x008D87ED008C87ED, 0x008F87ED008E87ED,
-    0x009187ED009087ED, 0x009387ED009287ED, 0x009587ED009487ED,
-    0x009787ED009687ED, 0x009987ED009887ED, 0x009B87ED009A87ED,
-    0x009D87ED009C87ED, 0x009F87ED009E87ED, 0x00A187ED00A087ED,
-    0x00A387ED00A287ED, 0x00A587ED00A487ED, 0x00A787ED00A687ED,
-    0x00A987ED00A887ED, 0x00AB87ED00AA87ED, 0x00AD87ED00AC87ED,
-    0x00AF87ED00AE87ED, 0x00B187ED00B087ED, 0x00B387ED00B287ED,
-    0x00B587ED00B487ED, 0x00B787ED00B687ED, 0x00B987ED00B887ED,
-    0x00BB87ED00BA87ED, 0x00BD87ED00BC87ED, 0x00BF87ED00BE87ED,
-    0x008188ED008088ED, 0x008388ED008288ED, 0x008588ED008488ED,
-    0x008788ED008688ED, 0x008988ED008888ED, 0x008B88ED008A88ED,
-    0x008D88ED008C88ED, 0x008F88ED008E88ED, 0x009188ED009088ED,
-    0x009388ED009288ED, 0x009588ED009488ED, 0x009788ED009688ED,
-    0x009988ED009888ED, 0x009B88ED009A88ED, 0x009D88ED009C88ED,
-    0x009F88ED009E88ED, 0x00A188ED00A088ED, 0x00A388ED00A288ED,
-    0x00A588ED00A488ED, 0x00A788ED00A688ED, 0x00A988ED00A888ED,
-    0x00AB88ED00AA88ED, 0x00AD88ED00AC88ED, 0x00AF88ED00AE88ED,
-    0x00B188ED00B088ED, 0x00B388ED00B288ED, 0x00B588ED00B488ED,
-    0x00B788ED00B688ED, 0x00B988ED00B888ED, 0x00BB88ED00BA88ED,
-    0x00BD88ED00BC88ED, 0x00BF88ED00BE88ED, 0x008189ED008089ED,
-    0x008389ED008289ED, 0x008589ED008489ED, 0x008789ED008689ED,
-    0x008989ED008889ED, 0x008B89ED008A89ED, 0x008D89ED008C89ED,
-    0x008F89ED008E89ED, 0x009189ED009089ED, 0x009389ED009289ED,
-    0x009589ED009489ED, 0x009789ED009689ED, 0x009989ED009889ED,
-    0x009B89ED009A89ED, 0x009D89ED009C89ED, 0x009F89ED009E89ED,
-    0x00A189ED00A089ED, 0x00A389ED00A289ED, 0x00A589ED00A489ED,
-    0x00A789ED00A689ED, 0x00A989ED00A889ED, 0x00AB89ED00AA89ED,
-    0x00AD89ED00AC89ED, 0x00AF89ED00AE89ED, 0x00B189ED00B089ED,
-    0x00B389ED00B289ED, 0x00B589ED00B489ED, 0x00B789ED00B689ED,
-    0x00B989ED00B889ED, 0x00BB89ED00BA89ED, 0x00BD89ED00BC89ED,
-    0x00BF89ED00BE89ED, 0x00818AED00808AED, 0x00838AED00828AED,
-    0x00858AED00848AED, 0x00878AED00868AED, 0x00898AED00888AED,
-    0x008B8AED008A8AED, 0x008D8AED008C8AED, 0x008F8AED008E8AED,
-    0x00918AED00908AED, 0x00938AED00928AED, 0x00958AED00948AED,
-    0x00978AED00968AED, 0x00998AED00988AED, 0x009B8AED009A8AED,
-    0x009D8AED009C8AED, 0x009F8AED009E8AED, 0x00A18AED00A08AED,
-    0x00A38AED00A28AED, 0x00A58AED00A48AED, 0x00A78AED00A68AED,
-    0x00A98AED00A88AED, 0x00AB8AED00AA8AED, 0x00AD8AED00AC8AED,
-    0x00AF8AED00AE8AED, 0x00B18AED00B08AED, 0x00B38AED00B28AED,
-    0x00B58AED00B48AED, 0x00B78AED00B68AED, 0x00B98AED00B88AED,
-    0x00BB8AED00BA8AED, 0x00BD8AED00BC8AED, 0x00BF8AED00BE8AED,
-    0x00818BED00808BED, 0x00838BED00828BED, 0x00858BED00848BED,
-    0x00878BED00868BED, 0x00898BED00888BED, 0x008B8BED008A8BED,
-    0x008D8BED008C8BED, 0x008F8BED008E8BED, 0x00918BED00908BED,
-    0x00938BED00928BED, 0x00958BED00948BED, 0x00978BED00968BED,
-    0x00998BED00988BED, 0x009B8BED009A8BED, 0x009D8BED009C8BED,
-    0x009F8BED009E8BED, 0x00A18BED00A08BED, 0x00A38BED00A28BED,
-    0x00A58BED00A48BED, 0x00A78BED00A68BED, 0x00A98BED00A88BED,
-    0x00AB8BED00AA8BED, 0x00AD8BED00AC8BED, 0x00AF8BED00AE8BED,
-    0x00B18BED00B08BED, 0x00B38BED00B28BED, 0x00B58BED00B48BED,
-    0x00B78BED00B68BED, 0x00B98BED00B88BED, 0x00BB8BED00BA8BED,
-    0x00BD8BED00BC8BED, 0x00BF8BED00BE8BED, 0x00818CED00808CED,
-    0x00838CED00828CED, 0x00858CED00848CED, 0x00878CED00868CED,
-    0x00898CED00888CED, 0x008B8CED008A8CED, 0x008D8CED008C8CED,
-    0x008F8CED008E8CED, 0x00918CED00908CED, 0x00938CED00928CED,
-    0x00958CED00948CED, 0x00978CED00968CED, 0x00998CED00988CED,
-    0x009B8CED009A8CED, 0x009D8CED009C8CED, 0x009F8CED009E8CED,
-    0x00A18CED00A08CED, 0x00A38CED00A28CED, 0x00A58CED00A48CED,
-    0x00A78CED00A68CED, 0x00A98CED00A88CED, 0x00AB8CED00AA8CED,
-    0x00AD8CED00AC8CED, 0x00AF8CED00AE8CED, 0x00B18CED00B08CED,
-    0x00B38CED00B28CED, 0x00B58CED00B48CED, 0x00B78CED00B68CED,
-    0x00B98CED00B88CED, 0x00BB8CED00BA8CED, 0x00BD8CED00BC8CED,
-    0x00BF8CED00BE8CED, 0x00818DED00808DED, 0x00838DED00828DED,
-    0x00858DED00848DED, 0x00878DED00868DED, 0x00898DED00888DED,
-    0x008B8DED008A8DED, 0x008D8DED008C8DED, 0x008F8DED008E8DED,
-    0x00918DED00908DED, 0x00938DED00928DED, 0x00958DED00948DED,
-    0x00978DED00968DED, 0x00998DED00988DED, 0x009B8DED009A8DED,
-    0x009D8DED009C8DED, 0x009F8DED009E8DED, 0x00A18DED00A08DED,
-    0x00A38DED00A28DED, 0x00A58DED00A48DED, 0x00A78DED00A68DED,
-    0x00A98DED00A88DED, 0x00AB8DED00AA8DED, 0x00AD8DED00AC8DED,
-    0x00AF8DED00AE8DED, 0x00B18DED00B08DED, 0x00B38DED00B28DED,
-    0x00B58DED00B48DED, 0x00B78DED00B68DED, 0x00B98DED00B88DED,
-    0x00BB8DED00BA8DED, 0x00BD8DED00BC8DED, 0x00BF8DED00BE8DED,
-    0x00818EED00808EED, 0x00838EED00828EED, 0x00858EED00848EED,
-    0x00878EED00868EED, 0x00898EED00888EED, 0x008B8EED008A8EED,
-    0x008D8EED008C8EED, 0x008F8EED008E8EED, 0x00918EED00908EED,
-    0x00938EED00928EED, 0x00958EED00948EED, 0x00978EED00968EED,
-    0x00998EED00988EED, 0x009B8EED009A8EED, 0x009D8EED009C8EED,
-    0x009F8EED009E8EED, 0x00A18EED00A08EED, 0x00A38EED00A28EED,
-    0x00A58EED00A48EED, 0x00A78EED00A68EED, 0x00A98EED00A88EED,
-    0x00AB8EED00AA8EED, 0x00AD8EED00AC8EED, 0x00AF8EED00AE8EED,
-    0x00B18EED00B08EED, 0x00B38EED00B28EED, 0x00B58EED00B48EED,
-    0x00B78EED00B68EED, 0x00B98EED00B88EED, 0x00BB8EED00BA8EED,
-    0x00BD8EED00BC8EED, 0x00BF8EED00BE8EED, 0x00818FED00808FED,
-    0x00838FED00828FED, 0x00858FED00848FED, 0x00878FED00868FED,
-    0x00898FED00888FED, 0x008B8FED008A8FED, 0x008D8FED008C8FED,
-    0x008F8FED008E8FED, 0x00918FED00908FED, 0x00938FED00928FED,
-    0x00958FED00948FED, 0x00978FED00968FED, 0x00998FED00988FED,
-    0x009B8FED009A8FED, 0x009D8FED009C8FED, 0x009F8FED009E8FED,
-    0x00A18FED00A08FED, 0x00A38FED00A28FED, 0x00A58FED00A48FED,
-    0x00A78FED00A68FED, 0x00A98FED00A88FED, 0x00AB8FED00AA8FED,
-    0x00AD8FED00AC8FED, 0x00AF8FED00AE8FED, 0x00B18FED00B08FED,
-    0x00B38FED00B28FED, 0x00B58FED00B48FED, 0x00B78FED00B68FED,
-    0x00B98FED00B88FED, 0x00BB8FED00BA8FED, 0x00BD8FED00BC8FED,
-    0x00BF8FED00BE8FED, 0x008190ED008090ED, 0x008390ED008290ED,
-    0x008590ED008490ED, 0x008790ED008690ED, 0x008990ED008890ED,
-    0x008B90ED008A90ED, 0x008D90ED008C90ED, 0x008F90ED008E90ED,
-    0x009190ED009090ED, 0x009390ED009290ED, 0x009590ED009490ED,
-    0x009790ED009690ED, 0x009990ED009890ED, 0x009B90ED009A90ED,
-    0x009D90ED009C90ED, 0x009F90ED009E90ED, 0x00A190ED00A090ED,
-    0x00A390ED00A290ED, 0x00A590ED00A490ED, 0x00A790ED00A690ED,
-    0x00A990ED00A890ED, 0x00AB90ED00AA90ED, 0x00AD90ED00AC90ED,
-    0x00AF90ED00AE90ED, 0x00B190ED00B090ED, 0x00B390ED00B290ED,
-    0x00B590ED00B490ED, 0x00B790ED00B690ED, 0x00B990ED00B890ED,
-    0x00BB90ED00BA90ED, 0x00BD90ED00BC90ED, 0x00BF90ED00BE90ED,
-    0x008191ED008091ED, 0x008391ED008291ED, 0x008591ED008491ED,
-    0x008791ED008691ED, 0x008991ED008891ED, 0x008B91ED008A91ED,
-    0x008D91ED008C91ED, 0x008F91ED008E91ED, 0x009191ED009091ED,
-    0x009391ED009291ED, 0x009591ED009491ED, 0x009791ED009691ED,
-    0x009991ED009891ED, 0x009B91ED009A91ED, 0x009D91ED009C91ED,
-    0x009F91ED009E91ED, 0x00A191ED00A091ED, 0x00A391ED00A291ED,
-    0x00A591ED00A491ED, 0x00A791ED00A691ED, 0x00A991ED00A891ED,
-    0x00AB91ED00AA91ED, 0x00AD91ED00AC91ED, 0x00AF91ED00AE91ED,
-    0x00B191ED00B091ED, 0x00B391ED00B291ED, 0x00B591ED00B491ED,
-    0x00B791ED00B691ED, 0x00B991ED00B891ED, 0x00BB91ED00BA91ED,
-    0x00BD91ED00BC91ED, 0x00BF91ED00BE91ED, 0x008192ED008092ED,
-    0x008392ED008292ED, 0x008592ED008492ED, 0x008792ED008692ED,
-    0x008992ED008892ED, 0x008B92ED008A92ED, 0x008D92ED008C92ED,
-    0x008F92ED008E92ED, 0x009192ED009092ED, 0x009392ED009292ED,
-    0x009592ED009492ED, 0x009792ED009692ED, 0x009992ED009892ED,
-    0x009B92ED009A92ED, 0x009D92ED009C92ED, 0x009F92ED009E92ED,
-    0x00A192ED00A092ED, 0x00A392ED00A292ED, 0x00A592ED00A492ED,
-    0x00A792ED00A692ED, 0x00A992ED00A892ED, 0x00AB92ED00AA92ED,
-    0x00AD92ED00AC92ED, 0x00AF92ED00AE92ED, 0x00B192ED00B092ED,
-    0x00B392ED00B292ED, 0x00B592ED00B492ED, 0x00B792ED00B692ED,
-    0x00B992ED00B892ED, 0x00BB92ED00BA92ED, 0x00BD92ED00BC92ED,
-    0x00BF92ED00BE92ED, 0x008193ED008093ED, 0x008393ED008293ED,
-    0x008593ED008493ED, 0x008793ED008693ED, 0x008993ED008893ED,
-    0x008B93ED008A93ED, 0x008D93ED008C93ED, 0x008F93ED008E93ED,
-    0x009193ED009093ED, 0x009393ED009293ED, 0x009593ED009493ED,
-    0x009793ED009693ED, 0x009993ED009893ED, 0x009B93ED009A93ED,
-    0x009D93ED009C93ED, 0x009F93ED009E93ED, 0x00A193ED00A093ED,
-    0x00A393ED00A293ED, 0x00A593ED00A493ED, 0x00A793ED00A693ED,
-    0x00A993ED00A893ED, 0x00AB93ED00AA93ED, 0x00AD93ED00AC93ED,
-    0x00AF93ED00AE93ED, 0x00B193ED00B093ED, 0x00B393ED00B293ED,
-    0x00B593ED00B493ED, 0x00B793ED00B693ED, 0x00B993ED00B893ED,
-    0x00BB93ED00BA93ED, 0x00BD93ED00BC93ED, 0x00BF93ED00BE93ED,
-    0x008194ED008094ED, 0x008394ED008294ED, 0x008594ED008494ED,
-    0x008794ED008694ED, 0x008994ED008894ED, 0x008B94ED008A94ED,
-    0x008D94ED008C94ED, 0x008F94ED008E94ED, 0x009194ED009094ED,
-    0x009394ED009294ED, 0x009594ED009494ED, 0x009794ED009694ED,
-    0x009994ED009894ED, 0x009B94ED009A94ED, 0x009D94ED009C94ED,
-    0x009F94ED009E94ED, 0x00A194ED00A094ED, 0x00A394ED00A294ED,
-    0x00A594ED00A494ED, 0x00A794ED00A694ED, 0x00A994ED00A894ED,
-    0x00AB94ED00AA94ED, 0x00AD94ED00AC94ED, 0x00AF94ED00AE94ED,
-    0x00B194ED00B094ED, 0x00B394ED00B294ED, 0x00B594ED00B494ED,
-    0x00B794ED00B694ED, 0x00B994ED00B894ED, 0x00BB94ED00BA94ED,
-    0x00BD94ED00BC94ED, 0x00BF94ED00BE94ED, 0x008195ED008095ED,
-    0x008395ED008295ED, 0x008595ED008495ED, 0x008795ED008695ED,
-    0x008995ED008895ED, 0x008B95ED008A95ED, 0x008D95ED008C95ED,
-    0x008F95ED008E95ED, 0x009195ED009095ED, 0x009395ED009295ED,
-    0x009595ED009495ED, 0x009795ED009695ED, 0x009995ED009895ED,
-    0x009B95ED009A95ED, 0x009D95ED009C95ED, 0x009F95ED009E95ED,
-    0x00A195ED00A095ED, 0x00A395ED00A295ED, 0x00A595ED00A495ED,
-    0x00A795ED00A695ED, 0x00A995ED00A895ED, 0x00AB95ED00AA95ED,
-    0x00AD95ED00AC95ED, 0x00AF95ED00AE95ED, 0x00B195ED00B095ED,
-    0x00B395ED00B295ED, 0x00B595ED00B495ED, 0x00B795ED00B695ED,
-    0x00B995ED00B895ED, 0x00BB95ED00BA95ED, 0x00BD95ED00BC95ED,
-    0x00BF95ED00BE95ED, 0x008196ED008096ED, 0x008396ED008296ED,
-    0x008596ED008496ED, 0x008796ED008696ED, 0x008996ED008896ED,
-    0x008B96ED008A96ED, 0x008D96ED008C96ED, 0x008F96ED008E96ED,
-    0x009196ED009096ED, 0x009396ED009296ED, 0x009596ED009496ED,
-    0x009796ED009696ED, 0x009996ED009896ED, 0x009B96ED009A96ED,
-    0x009D96ED009C96ED, 0x009F96ED009E96ED, 0x00A196ED00A096ED,
-    0x00A396ED00A296ED, 0x00A596ED00A496ED, 0x00A796ED00A696ED,
-    0x00A996ED00A896ED, 0x00AB96ED00AA96ED, 0x00AD96ED00AC96ED,
-    0x00AF96ED00AE96ED, 0x00B196ED00B096ED, 0x00B396ED00B296ED,
-    0x00B596ED00B496ED, 0x00B796ED00B696ED, 0x00B996ED00B896ED,
-    0x00BB96ED00BA96ED, 0x00BD96ED00BC96ED, 0x00BF96ED00BE96ED,
-    0x008197ED008097ED, 0x008397ED008297ED, 0x008597ED008497ED,
-    0x008797ED008697ED, 0x008997ED008897ED, 0x008B97ED008A97ED,
-    0x008D97ED008C97ED, 0x008F97ED008E97ED, 0x009197ED009097ED,
-    0x009397ED009297ED, 0x009597ED009497ED, 0x009797ED009697ED,
-    0x009997ED009897ED, 0x009B97ED009A97ED, 0x009D97ED009C97ED,
-    0x009F97ED009E97ED, 0x00A197ED00A097ED, 0x00A397ED00A297ED,
-    0x00A597ED00A497ED, 0x00A797ED00A697ED, 0x00A997ED00A897ED,
-    0x00AB97ED00AA97ED, 0x00AD97ED00AC97ED, 0x00AF97ED00AE97ED,
-    0x00B197ED00B097ED, 0x00B397ED00B297ED, 0x00B597ED00B497ED,
-    0x00B797ED00B697ED, 0x00B997ED00B897ED, 0x00BB97ED00BA97ED,
-    0x00BD97ED00BC97ED, 0x00BF97ED00BE97ED, 0x008198ED008098ED,
-    0x008398ED008298ED, 0x008598ED008498ED, 0x008798ED008698ED,
-    0x008998ED008898ED, 0x008B98ED008A98ED, 0x008D98ED008C98ED,
-    0x008F98ED008E98ED, 0x009198ED009098ED, 0x009398ED009298ED,
-    0x009598ED009498ED, 0x009798ED009698ED, 0x009998ED009898ED,
-    0x009B98ED009A98ED, 0x009D98ED009C98ED, 0x009F98ED009E98ED,
-    0x00A198ED00A098ED, 0x00A398ED00A298ED, 0x00A598ED00A498ED,
-    0x00A798ED00A698ED, 0x00A998ED00A898ED, 0x00AB98ED00AA98ED,
-    0x00AD98ED00AC98ED, 0x00AF98ED00AE98ED, 0x00B198ED00B098ED,
-    0x00B398ED00B298ED, 0x00B598ED00B498ED, 0x00B798ED00B698ED,
-    0x00B998ED00B898ED, 0x00BB98ED00BA98ED, 0x00BD98ED00BC98ED,
-    0x00BF98ED00BE98ED, 0x008199ED008099ED, 0x008399ED008299ED,
-    0x008599ED008499ED, 0x008799ED008699ED, 0x008999ED008899ED,
-    0x008B99ED008A99ED, 0x008D99ED008C99ED, 0x008F99ED008E99ED,
-    0x009199ED009099ED, 0x009399ED009299ED, 0x009599ED009499ED,
-    0x009799ED009699ED, 0x009999ED009899ED, 0x009B99ED009A99ED,
-    0x009D99ED009C99ED, 0x009F99ED009E99ED, 0x00A199ED00A099ED,
-    0x00A399ED00A299ED, 0x00A599ED00A499ED, 0x00A799ED00A699ED,
-    0x00A999ED00A899ED, 0x00AB99ED00AA99ED, 0x00AD99ED00AC99ED,
-    0x00AF99ED00AE99ED, 0x00B199ED00B099ED, 0x00B399ED00B299ED,
-    0x00B599ED00B499ED, 0x00B799ED00B699ED, 0x00B999ED00B899ED,
-    0x00BB99ED00BA99ED, 0x00BD99ED00BC99ED, 0x00BF99ED00BE99ED,
-    0x00819AED00809AED, 0x00839AED00829AED, 0x00859AED00849AED,
-    0x00879AED00869AED, 0x00899AED00889AED, 0x008B9AED008A9AED,
-    0x008D9AED008C9AED, 0x008F9AED008E9AED, 0x00919AED00909AED,
-    0x00939AED00929AED, 0x00959AED00949AED, 0x00979AED00969AED,
-    0x00999AED00989AED, 0x009B9AED009A9AED, 0x009D9AED009C9AED,
-    0x009F9AED009E9AED, 0x00A19AED00A09AED, 0x00A39AED00A29AED,
-    0x00A59AED00A49AED, 0x00A79AED00A69AED, 0x00A99AED00A89AED,
-    0x00AB9AED00AA9AED, 0x00AD9AED00AC9AED, 0x00AF9AED00AE9AED,
-    0x00B19AED00B09AED, 0x00B39AED00B29AED, 0x00B59AED00B49AED,
-    0x00B79AED00B69AED, 0x00B99AED00B89AED, 0x00BB9AED00BA9AED,
-    0x00BD9AED00BC9AED, 0x00BF9AED00BE9AED, 0x00819BED00809BED,
-    0x00839BED00829BED, 0x00859BED00849BED, 0x00879BED00869BED,
-    0x00899BED00889BED, 0x008B9BED008A9BED, 0x008D9BED008C9BED,
-    0x008F9BED008E9BED, 0x00919BED00909BED, 0x00939BED00929BED,
-    0x00959BED00949BED, 0x00979BED00969BED, 0x00999BED00989BED,
-    0x009B9BED009A9BED, 0x009D9BED009C9BED, 0x009F9BED009E9BED,
-    0x00A19BED00A09BED, 0x00A39BED00A29BED, 0x00A59BED00A49BED,
-    0x00A79BED00A69BED, 0x00A99BED00A89BED, 0x00AB9BED00AA9BED,
-    0x00AD9BED00AC9BED, 0x00AF9BED00AE9BED, 0x00B19BED00B09BED,
-    0x00B39BED00B29BED, 0x00B59BED00B49BED, 0x00B79BED00B69BED,
-    0x00B99BED00B89BED, 0x00BB9BED00BA9BED, 0x00BD9BED00BC9BED,
-    0x00BF9BED00BE9BED, 0x00819CED00809CED, 0x00839CED00829CED,
-    0x00859CED00849CED, 0x00879CED00869CED, 0x00899CED00889CED,
-    0x008B9CED008A9CED, 0x008D9CED008C9CED, 0x008F9CED008E9CED,
-    0x00919CED00909CED, 0x00939CED00929CED, 0x00959CED00949CED,
-    0x00979CED00969CED, 0x00999CED00989CED, 0x009B9CED009A9CED,
-    0x009D9CED009C9CED, 0x009F9CED009E9CED, 0x00A19CED00A09CED,
-    0x00A39CED00A29CED, 0x00A59CED00A49CED, 0x00A79CED00A69CED,
-    0x00A99CED00A89CED, 0x00AB9CED00AA9CED, 0x00AD9CED00AC9CED,
-    0x00AF9CED00AE9CED, 0x00B19CED00B09CED, 0x00B39CED00B29CED,
-    0x00B59CED00B49CED, 0x00B79CED00B69CED, 0x00B99CED00B89CED,
-    0x00BB9CED00BA9CED, 0x00BD9CED00BC9CED, 0x00BF9CED00BE9CED,
-    0x00819DED00809DED, 0x00839DED00829DED, 0x00859DED00849DED,
-    0x00879DED00869DED, 0x00899DED00889DED, 0x008B9DED008A9DED,
-    0x008D9DED008C9DED, 0x008F9DED008E9DED, 0x00919DED00909DED,
-    0x00939DED00929DED, 0x00959DED00949DED, 0x00979DED00969DED,
-    0x00999DED00989DED, 0x009B9DED009A9DED, 0x009D9DED009C9DED,
-    0x009F9DED009E9DED, 0x00A19DED00A09DED, 0x00A39DED00A29DED,
-    0x00A59DED00A49DED, 0x00A79DED00A69DED, 0x00A99DED00A89DED,
-    0x00AB9DED00AA9DED, 0x00AD9DED00AC9DED, 0x00AF9DED00AE9DED,
-    0x00B19DED00B09DED, 0x00B39DED00B29DED, 0x00B59DED00B49DED,
-    0x00B79DED00B69DED, 0x00B99DED00B89DED, 0x00BB9DED00BA9DED,
-    0x00BD9DED00BC9DED, 0x00BF9DED00BE9DED, 0x00819EED00809EED,
-    0x00839EED00829EED, 0x00859EED00849EED, 0x00879EED00869EED,
-    0x00899EED00889EED, 0x008B9EED008A9EED, 0x008D9EED008C9EED,
-    0x008F9EED008E9EED, 0x00919EED00909EED, 0x00939EED00929EED,
-    0x00959EED00949EED, 0x00979EED00969EED, 0x00999EED00989EED,
-    0x009B9EED009A9EED, 0x009D9EED009C9EED, 0x009F9EED009E9EED,
-    0x00A19EED00A09EED, 0x00A39EED00A29EED, 0x9090F000A89090F0,
-    0xF000AA9090F000A9, 0xAC9090F000AB9090, 0x90F000AD9090F000,
-    0x00AF9090F000AE90, 0x9090F000B09090F0, 0xF000B29090F000B1,
-    0xB49090F000B39090, 0x90F000B59090F000, 0x00B79090F000B690,
-    0x9090F000B89090F0, 0xF000BA9090F000B9, 0xBC9090F000BB9090,
-    0x90F000BD9090F000, 0x00BF9090F000BE90, 0x9190F000809190F0,
-    0xF000829190F00081, 0x849190F000839190, 0x90F000859190F000,
-    0x00879190F0008691, 0x9190F000889190F0, 0xF0008A9190F00089,
-    0x8C9190F0008B9190, 0x90F0008D9190F000, 0x008F9190F0008E91,
-    0x9390F000989390F0, 0xF0009A9390F00099, 0x9C9390F0009B9390,
-    0x90F0009D9390F000, 0x009F9390F0009E93, 0x9390F000A09390F0,
-    0xF000A29390F000A1, 0xA49390F000A39390, 0x90F000A59390F000,
-    0x00A79390F000A693, 0x9390F000A89390F0, 0xF000AA9390F000A9,
-    0xAC9390F000AB9390, 0x90F000AD9390F000, 0x00AF9390F000AE93,
-    0x9390F000B09390F0, 0xF000B29390F000B1, 0xB49390F000B39390,
-    0x90F000B59390F000, 0x00B79390F000B693, 0x9390F000B89390F0,
-    0xF000BA9390F000B9, 0x979690F000BB9390, 0x90F000989690F000,
-    0x009A9690F0009996, 0x9690F0009B9690F0, 0xF0009D9690F0009C,
-    0x9F9690F0009E9690, 0x90F000A09690F000, 0x00A39690F000A196,
-    0x9690F000A49690F0, 0xF000A69690F000A5, 0xA89690F000A79690,
-    0x90F000A99690F000, 0x00AB9690F000AA96, 0x9690F000AC9690F0,
-    0xF000AE9690F000AD, 0xB09690F000AF9690, 0x90F000B19690F000,
-    0x00B49690F000B396, 0x9690F000B59690F0, 0xF000B79690F000B6,
-    0xB99690F000B89690, 0x90F000BB9690F000, 0x0080B390F000BC96,
-    0xB390F00081B390F0, 0xF00083B390F00082, 0x85B390F00084B390,
-    0x90F00086B390F000, 0x0088B390F00087B3, 0xB390F00089B390F0,
-    0xF0008BB390F0008A, 0x8DB390F0008CB390, 0x90F0008EB390F000,
-    0x0090B390F0008FB3, 0xB390F00091B390F0, 0xF00093B390F00092,
-    0x95B390F00094B390, 0x90F00096B390F000, 0x0098B390F00097B3,
-    0xB390F00099B390F0, 0xF0009BB390F0009A, 0x9DB390F0009CB390,
-    0x90F0009EB390F000, 0x00A0B390F0009FB3, 0xB390F000A1B390F0,
-    0xF000A3B390F000A2, 0xA5B390F000A4B390, 0x90F000A6B390F000,
-    0x00A8B390F000A7B3, 0xB390F000A9B390F0, 0xF000ABB390F000AA,
-    0xADB390F000ACB390, 0x90F000AEB390F000, 0x00B0B390F000AFB3,
-    0xB390F000B1B390F0, 0xF0009A8291F000B2, 0xAB8291F0009C8291,
-    0x91F000AE8491F000, 0x008B8D91F000AF84, 0x9291F0008C8D91F0,
-    0xF000BC9291F000BB, 0xBA9691F000BE9291, 0x91F000BB9691F000,
-    0x0081A391F00080A3, 0xA391F00082A391F0, 0xF00084A391F00083,
-    0x86A391F00085A391, 0x91F00087A391F000, 0x0089A391F00088A3,
-    0xA391F0008AA391F0, 0xF0008CA391F0008B, 0x8EA391F0008DA391,
-    0x91F0008FA391F000, 0x0091A391F00090A3, 0xA391F00092A391F0,
-    0xF00094A391F00093, 0x96A391F00095A391, 0x91F00097A391F000,
-    0x0099A391F00098A3, 0xA391F0009AA391F0, 0xF0009CA391F0009B,
-    0x9EA391F0009DA391, 0x91F0009FA391F000, 0x00A0B996F000B8A4,
-    0xB996F000A1B996F0, 0xF000A3B996F000A2, 0xA5B996F000A4B996,
-    0x96F000A6B996F000, 0x00A8B996F000A7B9, 0xB996F000A9B996F0,
-    0xF000ABB996F000AA, 0xADB996F000ACB996, 0x96F000AEB996F000,
-    0x00B0B996F000AFB9, 0xB996F000B1B996F0, 0xF000B3B996F000B2,
-    0xB5B996F000B4B996, 0x96F000B6B996F000, 0x00B8B996F000B7B9,
-    0xB996F000B9B996F0, 0xF000BBB996F000BA, 0xBDB996F000BCB996,
-    0x96F000BEB996F000, 0xF097859DF000BFB9, 0x98859DF000A5859D,
-    0x859DF000A5859DF0, 0x859DF0A5859DF098, 0x9DF098859DF000AE,
-    0xF000AF859DF0A585, 0xF0A5859DF098859D, 0x98859DF000B0859D,
-    0xB1859DF0A5859DF0, 0x859DF098859DF000, 0x9DF000B2859DF0A5,
-    0xF000A5859DF0B986, 0xF0A5859DF0B9869D, 0xB9869DF000AE859D,
-    0xAF859DF0A5859DF0, 0x859DF0BA869DF000, 0x9DF0BA869DF000A5,
-    0xF000AE859DF0A585, 0xF0A5859DF0BA869D, 0x84BC9DF000AF859D,
-    0x9DF00085BC9DF000, 0x0088BC9DF00086BC, 0xBC9DF0008ABC9DF0,
-    0xF000A2A49EF0009E, 0xA4A49EF000A3A49E, 0x9EF000A5A49EF000,
-    0x00A7A49EF000A6A4, 0xA49EF000A8A49EF0, 0xF000AAA49EF000A9,
-    0xACA49EF000ABA49E, 0x9EF000ADA49EF000, 0x00AFA49EF000AEA4,
-    0xA49EF000B0A49EF0, 0xF000B2A49EF000B1, 0xB4A49EF000B3A49E,
-    0x9EF000B5A49EF000, 0x00B7A49EF000B6A4, 0xA49EF000B8A49EF0,
-    0xF000BAA49EF000B9, 0xBCA49EF000BBA49E, 0x9EF000BDA49EF000,
-    0x00BFA49EF000BEA4, 0xA59EF00080A59EF0, 0xF00082A59EF00081,
-    0xA284A0F00083A59E, 0xA0F0009C94A0F000, 0x008B95A0F000A594,
-    0xA0A0F000BA98A0F0, 0xF0009EA3A0F00084, 0xA3ADA0F000ACA8A0,
-    0xA1F000A493A1F000, 0x00AA9BA1F000A89A, 0xACA1F00088A7A1F0,
-    0xF0008BB4A1F00098, 0xA6B7A1F000A4B7A1, 0xA2F0008386A2F000,
-    0x00B18CA2F0009F86, 0xA1A2F000949BA2F0, 0xF0008AA1A2F00084,
-    0xB1AFA2F0008CACA2, 0xA3F0008A80A3F000, 0x009F8DA3F000B88A,
-    0x8EA3F000938EA3F0, 0xF000838FA3F0009C, 0xAD91A3F000958FA3,
-    0xA3F000A39AA3F000, 0x008DAAA3F000A7A2, 0xB2A3F000BAABA3F0,
-    0xF0009EB4A3F000BC, 0x9EBDA3F00091BBA3, 0xA4F0008EBEA3F000,
-    0x00AE8BA4F000A389, 0x98A4F000AB8EA4F0, 0xF000B59CA4F00088,
-    0xB6B0A4F00094A0A4, 0xA4F00092B2A4F000, 0x00B8BEA4F000A1BE,
-    0x83A5F0008481A5F0, 0xF000B383A5F000B2, 0xB384A5F0009984A5,
-    0xA5F0008989A5F000, 0x00A698A5F0009D90, 0x9BA5F0009A9AA5F0,
-    0xF000BCA5A5F00085, 0xABAEA5F000A7AAA5, 0xA5F00080B2A5F000,
-    0x0086BEA5F00090B3, 0x88A6F0009A87A6F0, 0xF0008789A6F000A8,
-    0xBE8CA6F000998BA6, 0xA6F0009A93A6F000, 0x00A896A6F000A394,
-    0x9EA6F000A79EA6F0, 0xF000BCACA6F000B5, 0x95B3A6F000B6B0A6,
-    0xA6F000ABB5A6F000, 0x00B1BEA6F000ACBC, 0x8FA7F0009283A7F0,
-    0xF000A799A7F0008A, 0xA6A5A7F000AEA2A7, 0xA7F000A8B2A7F000,
-    0x00AFBCA7F00093BB, 0x97A8F0009297A8F0, 0xF000AE9CA8F000AD,
-    0xB7B5A8F000BAAFA8, 0xA9F0008585A9F000, 0x009A88A9F0009F87,
-    0x92A9F0008A90A9F0, 0xF000B696A9F00096, 0x8E83AAF000B0ACA9,
-    0xAAF0008584AAF000, 0x00918AAAF0008E88, 0x98AAF000928EAAF0,
-    0x0000000000000080,
-};
-
-const BinaryBlob kNormalizationRules_blob[] = {
-    {"nfkc", 238884,
-     reinterpret_cast<const char *>(kNormalizationRules_blob_uint64 + 0)},
-    {"nmt_nfkc", 238883,
-     reinterpret_cast<const char *>(kNormalizationRules_blob_uint64 + 29861)},
-    {"nfkc_cf", 245779,
-     reinterpret_cast<const char *>(kNormalizationRules_blob_uint64 + 59722)},
-    {"nmt_nfkc_cf", 245778,
-     reinterpret_cast<const char *>(kNormalizationRules_blob_uint64 + 90445)},
-};
-#else
-constexpr BinaryBlob kNormalizationRules_blob[] = {
-    {"nfkc", 238884,
-     "\x00\xB8\x02\x00\x00\xFC\x00\x00\x3E\x1C\x03\x00\x3D\x24\x03\x00\x3C\x2C"
-     "\x03\x00"
-     "\xCC\xE4\x02\x00\xB8\x0D\x00\x00\x82\x23\x00\x80\xCC\xDC\x02\x00\xB8\x05"
-     "\x00\x00"
-     "\x76\x23\x00\x80\xCC\xE4\x02\x00\xB8\x1D\x00\x00\x86\x23\x00\x80\x91\x0D"
-     "\x01\x00"
-     "\xCC\x48\x02\x00\xCD\x38\x00\x00\x8C\x4D\x01\x00\x9F\x09\x00\x80\xA2\x09"
-     "\x00\x80"
-     "\x8F\x5D\x01\x00\x88\xBD\x00\x00\x89\xAD\x00\x00\x8A\x59\x01\x00\xA5\x09"
-     "\x00\x80"
-     "\x84\xC9\x00\x00\x38\x1E\x00\x80\x86\xD9\x00\x00\x87\xA1\x00\x00\x80\x35"
-     "\x00\x00"
-     "\x81\x3D\x00\x00\x82\x25\x00\x00\x83\xDD\x00\x00\x80\xE5\x00\x00\x81\x0D"
-     "\x00\x00"
-     "\x30\x1E\x00\x80\x83\x1D\x00\x00\x48\x1E\x00\x80\x40\x1E\x00\x80\x80\xFD"
-     "\x00\x00"
-     "\x81\x15\x00\x00\xA8\x09\x00\x80\x89\x31\x00\x00\x4A\x0A\x00\x80\x60\x1E"
-     "\x00\x80"
-     "\x50\x0A\x00\x80\x58\x1E\x00\x80\x70\x1E\x00\x80\x68\x1E\x00\x80\x81\x75"
-     "\x00\x00"
-     "\x80\x69\x00\x00\x83\x71\x00\x00\x99\x0C\x00\x80\xA8\x69\x01\x00\xF1\x0B"
-     "\x00\x80"
-     "\x80\x75\x00\x00\x81\x69\x00\x00\x89\x5D\x00\x00\xA5\x69\x01\x00\x84\x3D"
-     "\x00\x00"
-     "\xAB\x09\x00\x80\x84\x05\x00\x00\xEB\x0B\x00\x80\x28\x1E\x00\x80\xA3\x81"
-     "\x01\x00"
-     "\xAE\x09\x00\x80\x81\x0D\x00\x00\x2A\x0C\x00\x80\xBB\x0B\x00\x80\x3C\x0C"
-     "\x00\x80"
-     "\x7A\xD0\x11\x00\x79\x84\x11\x00\x78\x04\x10\x00\x77\x40\x11\x00\x76\x30"
-     "\x10\x00"
-     "\x75\xF8\x13\x00\x74\xA4\x13\x00\x73\x58\x13\x00\x72\x1C\x13\x00\x42\x0C"
-     "\x00\x80"
-     "\x70\xEC\x12\x00\x6F\xF0\x0D\x00\x6E\x8C\x0C\x00\x6D\x8C\x0C\x00\x6C\xDC"
-     "\x0C\x00"
-     "\x6B\x00\x0C\x00\x6A\x64\x0C\x00\x69\xA0\x0F\x00\x68\xF8\x0F\x00\x67\x04"
-     "\x0F\x00"
-     "\x66\x64\x0F\x00\x65\x68\x0E\x00\x64\x94\x09\x00\x63\xEC\x09\x00\x62\xD0"
-     "\x09\x00"
-     "\x61\xD0\x08\x00\x20\x1E\x00\x80\x82\x05\x00\x00\x50\x1E\x00\x80\x78\x1E"
-     "\x00\x80"
-     "\xB8\x1B\x00\x80\x86\x19\x00\x00\x5A\xEC\x08\x00\x59\x48\x0B\x00\x58\x18"
-     "\x0B\x00"
-     "\x57\x00\x0B\x00\x56\x4C\x05\x00\x55\xF8\x0A\x00\x54\x34\x04\x00\x53\x30"
-     "\x05\x00"
-     "\x52\x00\x05\x00\x56\x0A\x00\x80\x50\xDC\x05\x00\x4F\x00\x07\x00\x4E\x6C"
-     "\x07\x00"
-     "\x4D\x90\x07\x00\x4C\xD4\x07\x00\x4B\x08\x06\x00\x4A\xC8\x01\x00\x49\x40"
-     "\x01\x00"
-     "\x48\xE4\x01\x00\x47\xF0\x01\x00\x46\x34\x00\x00\x45\xA0\x00\x00\x44\x9C"
-     "\x00\x00"
-     "\x43\xD8\x00\x00\x42\xC8\x00\x00\x41\xF0\x02\x00\xC0\x1B\x00\x80\x81\x49"
-     "\x02\x00"
-     "\x80\x41\x02\x00\x87\xF5\x03\x00\xCC\x18\x02\x00\xC8\x1B\x00\x80\xD0\x1B"
-     "\x00\x80"
-     "\xCC\x58\x02\x00\xCD\x48\x02\x00\x5C\x0A\x00\x80\x62\x0A\x00\x80\x68\x0A"
-     "\x00\x80"
-     "\x6E\x0A\x00\x80\xCC\x30\x01\x00\xCD\x34\x01\x00\x81\x05\x00\x00\xD8\x1B"
-     "\x00\x80"
-     "\xCC\x5C\x04\x00\x81\x65\x00\x00\x82\x6D\x00\x00\xCC\xC4\x01\x00\x81\x71"
-     "\x00\x00"
-     "\xE0\x1B\x00\x80\x74\x0A\x00\x80\x87\x75\x00\x00\xE8\x1B\x00\x80\xF8\x1B"
-     "\x00\x80"
-     "\x00\x1C\x00\x80\xF0\x1B\x00\x80\x8C\x5D\x00\x00\xB7\x09\x00\x80\xCC\x6C"
-     "\x04\x00"
-     "\xCD\x64\x03\x00\x83\x1D\x00\x00\xBA\x09\x00\x80\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xA0\x1E\x00\x80\x98\x1E\x00\x80\xA3\x89\x00\x00\xB0\x1E\x00\x80\xA8\x1E"
-     "\x00\x80"
-     "\x90\x1E\x00\x80\x89\x09\x00\x00\x08\x1C\x00\x80\x10\x1C\x00\x80\x80\x0A"
-     "\x00\x80"
-     "\x86\x0A\x00\x80\xC0\x09\x00\x80\x80\x51\x00\x00\x81\x51\x00\x00\x80\x65"
-     "\x00\x00"
-     "\x81\x7D\x00\x00\xB1\xC5\x00\x00\x88\x1E\x00\x80\x82\x29\x00\x00\xA7\x01"
-     "\x04\x00"
-     "\xCC\x08\x02\x00\x92\x0A\x00\x80\x48\x0C\x00\x80\x4E\x0C\x00\x80\xB8\x1E"
-     "\x00\x80"
-     "\x87\xE1\x01\x00\x86\x05\x00\x00\x28\x1C\x00\x80\x81\x29\x00\x00\x8C\x0A"
-     "\x00\x80"
-     "\xCC\x70\x05\x00\x18\x1C\x00\x80\x20\x1C\x00\x80\x30\x1C\x00\x80\x80\x6D"
-     "\x01\x00"
-     "\x81\x99\x01\x00\xCC\x44\x05\x00\xCD\x20\x02\x00\x1E\x0C\x00\x80\xCC\xAC"
-     "\x02\x00"
-     "\x98\x0A\x00\x80\x38\x1C\x00\x80\x9E\x0A\x00\x80\xF0\x7C\x8A\x02\xEF\x20"
-     "\x65\x02"
-     "\xA4\x0A\x00\x80\xFD\x0B\x00\x80\xAA\x0A\x00\x80\xB0\x0A\x00\x80\xEA\x00"
-     "\x64\x02"
-     "\x40\x1C\x00\x80\x50\x1C\x00\x80\x90\x0C\x00\x80\x48\x1C\x00\x80\xCD\xF0"
-     "\x01\x00"
-     "\xCC\xEC\x01\x00\xE3\x70\x48\x02\xE2\x74\x38\x02\xE1\x58\x28\x00\xE0\xEC"
-     "\x27\x00"
-     "\xCC\x44\x01\x00\xCD\x5C\x01\x00\x82\x35\x07\x00\x58\x1C\x00\x80\xDB\x30"
-     "\x24\x00"
-     "\x60\x1C\x00\x80\xD9\x94\x24\x00\xD8\x0C\x24\x00\xC3\x09\x00\x80\xD6\x40"
-     "\x25\x00"
-     "\xCC\xA8\x04\x00\xCD\x64\x05\x00\xD3\xD0\x24\x00\xC6\x09\x00\x80\xD1\x34"
-     "\x25\x00"
-     "\xD0\x34\x23\x00\xCF\xC0\x21\x00\xCE\xDC\x19\x00\xCD\xC8\x16\x00\xC9\x09"
-     "\x00\x80"
-     "\xCB\x14\x17\x00\xCA\x88\x17\x00\xBC\x0A\x00\x80\xCC\x3C\x01\x00\xC7\xFC"
-     "\x17\x00"
-     "\xC6\x84\x12\x00\xC5\xBC\x14\x00\xC4\x64\x14\x00\xC3\xE8\x14\x00\xC2\xE4"
-     "\x15\x00"
-     "\xC2\x0A\x00\x80\xCC\x04\x01\x00\x87\x55\x06\x00\x84\x3D\x00\x00\x87\xB1"
-     "\x06\x00"
-     "\x86\xB9\x06\x00\x81\x95\x06\x00\x80\x61\x06\x00\x83\xBD\x06\x00\x82\x3D"
-     "\x00\x00"
-     "\xBD\x09\x00\x80\x8C\xC1\x06\x00\x8F\xC1\x06\x00\x8C\x75\x06\x00\x89\xE5"
-     "\x06\x00"
-     "\x88\x89\x06\x00\x7A\x0A\x00\x80\xC8\x0A\x00\x80\x80\x1E\x00\x80\x9F\x0C"
-     "\x00\x80"
-     "\x8C\x29\x07\x00\xD4\x0A\x00\x80\x91\xBD\x06\x00\x8C\x1D\x07\x00\x88\x05"
-     "\x07\x00"
-     "\xCC\x09\x00\x80\x81\x49\x07\x00\x87\x3D\x07\x00\x81\xED\x00\x00\x82\x5D"
-     "\x07\x00"
-     "\x82\x21\x07\x00\x84\x41\x07\x00\x87\x3D\x07\x00\x86\x45\x07\x00\xA7\xE5"
-     "\x06\x00"
-     "\x70\x1C\x00\x80\xA7\xCD\x00\x00\xC0\x1E\x00\x80\xA3\xF1\x06\x00\xC1\x0B"
-     "\x00\x80"
-     "\xA3\xD9\x00\x00\x54\x0C\x00\x80\xAD\xAD\x07\x00\x5A\x0C\x00\x80\xAD\xC1"
-     "\x06\x00"
-     "\xC8\x1E\x00\x80\xCE\x0A\x00\x80\xA8\xB1\x07\x00\x68\x1C\x00\x80\xDA\x0A"
-     "\x00\x80"
-     "\xAE\x55\x07\x00\x78\x1C\x00\x80\x80\x69\x07\x00\x81\x79\x07\x00\x03\x0C"
-     "\x00\x80"
-     "\xB0\xC5\x07\x00\xB1\xB5\x06\x00\x80\x1C\x00\x80\xE0\x0A\x00\x80\xA7\x69"
-     "\x07\x00"
-     "\xCC\xAC\x03\x00\xCD\x14\x02\x00\x88\x1C\x00\x80\xA3\x91\x07\x00\xA7\xB5"
-     "\x07\x00"
-     "\x81\x39\x00\x00\x80\xA1\x06\x00\x81\xB1\x06\x00\x82\xC5\x06\x00\x83\xD5"
-     "\x06\x00"
-     "\x84\xE9\x06\x00\xE6\x0A\x00\x80\x86\x25\x01\x00\x87\x51\x01\x00\x88\x7D"
-     "\x01\x00"
-     "\x89\xA9\x01\x00\xCC\x78\x02\x00\xCD\x4C\x02\x00\x8C\xA5\x01\x00\xF2\x0A"
-     "\x00\x80"
-     "\x90\x1C\x00\x80\x8F\xA1\x01\x00\x81\x85\x01\x00\x91\xE1\x01\x00\x84\x05"
-     "\x00\x00"
-     "\x98\x1C\x00\x80\xEC\x0A\x00\x80\x81\x41\x00\x00\xA8\x1C\x00\x80\xA0\x1C"
-     "\x00\x80"
-     "\x8C\x55\x00\x00\x81\x71\x00\x00\xCC\x94\x02\x00\xCD\xE0\x02\x00\xA3\xF5"
-     "\x00\x00"
-     "\x8C\xA5\x01\x00\xB0\x1C\x00\x80\xB8\x1C\x00\x80\xB1\x71\x01\x00\xC0\x1C"
-     "\x00\x80"
-     "\x81\xF1\x00\x00\xA3\x21\x01\x00\x24\x0C\x00\x80\xB1\xC9\x00\x00\xCC\x98"
-     "\x01\x00"
-     "\xCD\x4C\x02\x00\xA8\x11\x01\x00\xF8\x0A\x00\x80\xCF\x09\x00\x80\xC8\x1C"
-     "\x00\x80"
-     "\x04\x0B\x00\x80\xD0\x1C\x00\x80\xFE\x0A\x00\x80\xE0\x1C\x00\x80\xB0\x79"
-     "\x01\x00"
-     "\xD8\x1C\x00\x80\xA3\x15\x01\x00\xA7\x9D\x00\x00\x80\x41\x00\x00\x81\x71"
-     "\x00\x00"
-     "\xA7\x39\x01\x00\xA3\xE5\x00\x00\x87\x9D\x00\x00\xAD\xBD\x00\x00\xCD\xD4"
-     "\x05\x00"
-     "\xCC\xD0\x05\x00\xCC\xA4\x01\x00\xCD\xBC\x00\x00\x81\x81\x00\x00\xD2\x09"
-     "\x00\x80"
-     "\x80\x91\x03\x00\x81\xA1\x03\x00\xCC\x50\x01\x00\x83\xA5\x03\x00\x91\xFD"
-     "\x00\x00"
-     "\x81\x29\x00\x00\x84\x81\x00\x00\x87\xB1\x03\x00\xD5\x09\x00\x80\xE8\x1E"
-     "\x00\x80"
-     "\xE0\x1E\x00\x80\xF8\x1E\x00\x80\x8C\x81\x03\x00\xF0\x1E\x00\x80\x9B\x21"
-     "\x0E\x00"
-     "\xE8\x1C\x00\x80\xAB\x0C\x00\x80\x84\x4D\x0E\x00\x87\x59\x0E\x00\x86\xC5"
-     "\x00\x00"
-     "\x81\x71\x00\x00\x80\xA9\x03\x00\x83\x59\x0E\x00\x82\x01\x0C\x00\xF0\x1C"
-     "\x00\x80"
-     "\x8C\x85\x00\x00\x8F\x8D\x00\x00\x84\xDD\x00\x00\x89\xC9\x00\x00\x88\x61"
-     "\x0E\x00"
-     "\x8B\xC5\x00\x00\xF8\x1C\x00\x80\x00\x1D\x00\x80\x89\xB1\x00\x00\x13\x0B"
-     "\x00\x80"
-     "\xA3\x39\x03\x00\x80\xB5\x00\x00\x81\xBD\x00\x00\xB7\x0C\x00\x80\xA7\x25"
-     "\x03\x00"
-     "\x80\x85\x00\x00\x81\x8D\x00\x00\xCC\xE4\x0C\x00\x83\x81\x00\x00\xA5\x0C"
-     "\x00\x80"
-     "\xAD\x09\x03\x00\xD8\x1E\x00\x80\x19\x0B\x00\x80\x82\xF5\x01\x00\xB1\x01"
-     "\x03\x00"
-     "\x84\x89\x01\x00\x88\xAD\x00\x00\x81\x51\x00\x00\x80\xA9\x00\x00\xA3\xF9"
-     "\x0E\x00"
-     "\xB4\x09\x00\x80\xC7\x0B\x00\x80\x60\x0C\x00\x80\x81\xD5\x00\x00\x66\x0C"
-     "\x00\x80"
-     "\x10\x1F\x00\x80\xA8\xD1\x0E\x00\x81\x9D\x01\x00\x84\xBD\x00\x00\x84\x9D"
-     "\x00\x00"
-     "\x89\x2D\x00\x00\xCC\xC0\x07\x00\xCD\xC4\x07\x00\xCC\x9C\x07\x00\xCD\xD4"
-     "\x07\x00"
-     "\x08\x1F\x00\x80\x20\x1F\x00\x80\x80\xD1\x01\x00\x81\x3D\x00\x00\x18\x1F"
-     "\x00\x80"
-     "\x83\x31\x00\x00\x28\x1F\x00\x80\x00\x1F\x00\x80\x81\x21\x00\x00\x80\xCD"
-     "\x01\x00"
-     "\x0F\x0C\x00\x80\x81\x21\x00\x00\x80\xB5\x02\x00\x81\x6D\x01\x00\xCC\x10"
-     "\x02\x00"
-     "\xCD\xA8\x03\x00\x87\xF1\x00\x00\x87\x25\x00\x00\x81\x1D\x00\x00\x08\x1D"
-     "\x00\x80"
-     "\xCD\x04\x01\x00\xCC\x00\x01\x00\xCC\xA0\x07\x00\xCC\x94\x07\x00\x10\x1D"
-     "\x00\x80"
-     "\x22\x0B\x00\x80\xCD\xE4\x02\x00\xCC\x84\x0D\x00\x18\x1D\x00\x80\x81\xA1"
-     "\x0F\x00"
-     "\x2E\x0B\x00\x80\x6C\x0C\x00\x80\x72\x0C\x00\x80\x28\x1D\x00\x80\x28\x0B"
-     "\x00\x80"
-     "\x30\x1D\x00\x80\x48\x1D\x00\x80\xA3\x9D\x00\x00\xCC\xF4\x0D\x00\xCD\x3C"
-     "\x02\x00"
-     "\x3A\x0B\x00\x80\x38\x1D\x00\x80\x87\x05\x00\x00\x50\x1D\x00\x80\x87\x05"
-     "\x00\x00"
-     "\x58\x1D\x00\x80\x88\xA1\x0F\x00\x84\x0C\x00\x80\xCC\x20\x0C\x00\x40\x0B"
-     "\x00\x80"
-     "\x60\x1D\x00\x80\x52\x0B\x00\x80\x68\x1D\x00\x80\x8A\x0C\x00\x80\x4C\x0B"
-     "\x00\x80"
-     "\x78\x1D\x00\x80\x70\x1D\x00\x80\x81\x95\x0F\x00\xDB\x09\x00\x80\x81\x79"
-     "\x0F\x00"
-     "\x0D\x0B\x00\x80\xE1\x09\x00\x80\xB1\x0C\x00\x80\xDE\x09\x00\x80\xA3\x0B"
-     "\x00\x80"
-     "\x87\x95\x0F\x00\xD0\x1E\x00\x80\x09\x0C\x00\x80\x20\x1D\x00\x80\x34\x0B"
-     "\x00\x80"
-     "\x8C\xA1\x0F\x00\x46\x0B\x00\x80\x40\x1D\x00\x80\x8F\xA9\x0F\x00\x81\x6D"
-     "\x00\x00"
-     "\x91\xDD\x0F\x00\xE4\x09\x00\x80\x82\xF5\x0F\x00\x87\xB1\x0F\x00\xE7\x09"
-     "\x00\x80"
-     "\x87\xE5\x0F\x00\x58\x0B\x00\x80\xCC\x68\x03\x00\xCD\x1C\x03\x00\x98\x1D"
-     "\x00\x80"
-     "\x5E\x0B\x00\x80\xA0\x1D\x00\x80\x8C\x41\x00\x00\x64\x0B\x00\x80\x8C\x99"
-     "\x0F\x00"
-     "\xEA\x09\x00\x80\x84\x2D\x00\x00\xB1\x71\x0F\x00\xA3\xA5\x00\x00\x81\x31"
-     "\x00\x00"
-     "\x80\x0D\x00\x00\xE5\x0B\x00\x80\xA7\x3D\x0F\x00\xD9\x0B\x00\x80\x8C\x09"
-     "\x00\x00"
-     "\xD3\x0B\x00\x80\xDF\x0B\x00\x80\x80\x29\x00\x00\x81\x15\x00\x00\x38\x1F"
-     "\x00\x80"
-     "\x6A\x0B\x00\x80\xA3\x29\x0F\x00\xB1\x61\x0F\x00\xA3\xF1\x00\x00\x70\x0B"
-     "\x00\x80"
-     "\xA7\x21\x0F\x00\xA6\x39\x0F\x00\xA7\x05\x0F\x00\xA6\x09\x0F\x00\xCD\x0B"
-     "\x00\x80"
-     "\x78\x0C\x00\x80\x7E\x0C\x00\x80\xAC\x0B\x00\x80\x48\x1F\x00\x80\x40\x1F"
-     "\x00\x80"
-     "\xAD\x0D\x0F\x00\x58\x1F\x00\x80\x82\x55\x01\x00\x83\x59\x01\x00\x80\x05"
-     "\x01\x00"
-     "\x81\x45\x01\x00\x86\x69\x01\x00\x50\x1F\x00\x80\x84\x75\x01\x00\x60\x1F"
-     "\x00\x80"
-     "\x8A\x9D\x01\x00\x8B\xE9\x01\x00\x88\xA9\x01\x00\x89\x95\x01\x00\x30\x1F"
-     "\x00\x80"
-     "\x8F\xD1\x01\x00\x8C\xD9\x01\x00\x80\x1D\x00\x80\x80\xB1\x01\x00\x81\xB1"
-     "\x01\x00"
-     "\x76\x0B\x00\x80\x91\xA5\x01\x00\x81\xA5\x01\x00\x80\xA5\x01\x00\x83\xA5"
-     "\x01\x00"
-     "\x90\x1D\x00\x80\x88\x1D\x00\x80\x9B\x89\x01\x00\xCD\xF8\x0D\x00\xCC\xCC"
-     "\x0D\x00"
-     "\x89\x65\x00\x00\xA3\xFD\x00\x00\x80\x75\x01\x00\x81\x35\x01\x00\xA8\x1D"
-     "\x00\x80"
-     "\xA3\xB5\x00\x00\xB0\x1D\x00\x80\xB8\x1D\x00\x80\xCC\x94\x01\x00\xCD\x54"
-     "\x02\x00"
-     "\xA4\xA5\x00\x00\xC0\x1D\x00\x80\x7C\x0B\x00\x80\xD0\x1D\x00\x80\xA8\xE1"
-     "\x00\x00"
-     "\xC8\x1D\x00\x80\xD8\x1D\x00\x80\xCC\x64\x02\x00\xE0\x1D\x00\x80\xAD\xE1"
-     "\x00\x00"
-     "\x80\x4D\x00\x00\x81\x59\x00\x00\xB0\xA9\x00\x00\x87\x75\x00\x00\xE8\x1D"
-     "\x00\x80"
-     "\x68\x1F\x00\x80\xA3\xC5\x00\x00\xED\x09\x00\x80\xCC\xD0\x01\x00\xCD\xA4"
-     "\x01\x00"
-     "\x82\x0B\x00\x80\x80\x1F\x00\x80\x88\x21\x00\x00\x83\x75\x00\x00\xBD\x0C"
-     "\x00\x80"
-     "\xF0\x1D\x00\x80\x81\x9D\x03\x00\x80\x89\x03\x00\x88\x0B\x00\x80\x82\xAD"
-     "\x03\x00"
-     "\x88\x19\x00\x00\x89\x09\x00\x00\x87\xBD\x03\x00\x78\x1F\x00\x80\x84\xD9"
-     "\x03\x00"
-     "\x88\x89\x03\x00\x70\x1F\x00\x80\x87\xD1\x03\x00\x80\xE5\x03\x00\x81\xE9"
-     "\x03\x00"
-     "\x82\xD1\x03\x00\x83\xD1\x03\x00\x80\x95\x03\x00\x81\x99\x03\x00\xCC\x2C"
-     "\x02\x00"
-     "\xCD\xC8\x02\x00\x8B\x0B\x00\x80\x8C\x3D\x00\x00\xF8\x1D\x00\x80\xD8\x09"
-     "\x00\x80"
-     "\x81\x31\x00\x00\x91\x0B\x00\x80\x97\x0B\x00\x80\x82\x35\x00\x00\x00\x1E"
-     "\x00\x80"
-     "\x08\x1E\x00\x80\x87\x1D\x00\x00\xF0\x09\x00\x80\x81\xD1\x00\x00\x84\x71"
-     "\x00\x00"
-     "\xA3\x39\x03\x00\xF3\x09\x00\x80\x89\x51\x00\x00\x81\xDD\x01\x00\xCC\x00"
-     "\x04\x00"
-     "\xCD\xCC\x03\x00\xB1\xD5\x00\x00\x3C\x1E\x00\x80\x34\x1E\x00\x80\x4C\x1E"
-     "\x00\x80"
-     "\x81\x19\x00\x00\x80\x11\x00\x00\x83\x15\x00\x00\xA3\x95\x00\x00\x44\x1E"
-     "\x00\x80"
-     "\xF9\x09\x00\x80\x80\x6D\x00\x00\x81\x65\x00\x00\x83\x3D\x00\x00\x4D\x0A"
-     "\x00\x80"
-     "\x81\x3D\x00\x00\x80\x3D\x00\x00\x64\x1E\x00\x80\x5C\x1E\x00\x80\xA3\x99"
-     "\x00\x00"
-     "\x74\x1E\x00\x80\x6C\x1E\x00\x80\xF4\x0B\x00\x80\x89\x09\x00\x00\xEE\x0B"
-     "\x00\x80"
-     "\x84\xFD\x01\x00\xCC\xD8\x0E\x00\xCC\xDC\x0E\x00\x2C\x1E\x00\x80\xCC\xD8"
-     "\x0F\x00"
-     "\xCC\xCC\x0F\x00\xCC\x08\x0E\x00\xCC\x58\x0E\x00\xCC\x28\x0F\x00\xCC\x58"
-     "\x0F\x00"
-     "\xCC\x08\x0E\x00\xCD\x10\x0E\x00\xCC\xDC\x0F\x00\xCD\x24\x0F\x00\xCC\xEC"
-     "\x0F\x00"
-     "\xCD\xE8\x0F\x00\x80\xA1\x01\x00\x81\xA1\x01\x00\x2D\x0C\x00\x80\xBE\x0B"
-     "\x00\x80"
-     "\x80\x2D\x01\x00\x81\xD9\x01\x00\x3F\x0C\x00\x80\xCC\xAC\x0E\x00\x45\x0C"
-     "\x00\x80"
-     "\x82\x15\x00\x00\xCD\x90\x0E\x00\xCC\x94\x0E\x00\x54\x1E\x00\x80\x86\x09"
-     "\x00\x00"
-     "\x87\xF1\x00\x00\x7C\x1E\x00\x80\xBC\x1B\x00\x80\x59\x0A\x00\x80\xC4\x1B"
-     "\x00\x80"
-     "\xCC\x1B\x00\x80\xD4\x1B\x00\x80\xCC\xF0\x02\x00\x5F\x0A\x00\x80\x65\x0A"
-     "\x00\x80"
-     "\xB1\x31\x00\x00\x6B\x0A\x00\x80\xCD\x68\x02\x00\xCC\xE8\x05\x00\xCC\x24"
-     "\x01\x00"
-     "\xCD\x04\x01\x00\x71\x0A\x00\x80\xDC\x1B\x00\x80\x81\x7D\x00\x00\x81\x5D"
-     "\x00\x00"
-     "\xCC\x80\x05\x00\xE4\x1B\x00\x80\x82\x21\x07\x00\x77\x0A\x00\x80\xCD\x98"
-     "\x00\x00"
-     "\xCC\x88\x00\x00\xEC\x1B\x00\x80\xFC\x1B\x00\x80\xA3\x65\x00\x00\x04\x1C"
-     "\x00\x80"
-     "\xF4\x1B\x00\x80\x08\x0A\x00\x80\x86\x01\x07\x00\x0B\x0A\x00\x80\xF6\x09"
-     "\x00\x80"
-     "\xA4\x1E\x00\x80\x53\x0A\x00\x80\xA5\x8D\x07\x00\x9C\x0C\x00\x80\xA3\x09"
-     "\x00\x00"
-     "\xFC\x09\x00\x80\x24\x1E\x00\x80\xFF\x09\x00\x80\x9C\x1E\x00\x80\x05\x0A"
-     "\x00\x80"
-     "\xB4\x1E\x00\x80\xCC\x34\x03\x00\xCD\x7C\x03\x00\xA8\xBD\x07\x00\x0E\x0A"
-     "\x00\x80"
-     "\x81\xD9\x07\x00\x89\x0D\x00\x00\xAC\x1E\x00\x80\x82\xD1\x07\x00\x94\x1E"
-     "\x00\x80"
-     "\x87\x99\x07\x00\x87\xFD\x07\x00\x7D\x0A\x00\x80\x80\x65\x00\x00\x81\x41"
-     "\x00\x00"
-     "\x0C\x1C\x00\x80\x83\x41\x00\x00\x14\x1C\x00\x80\x8C\xCD\x07\x00\x8C\xAD"
-     "\x07\x00"
-     "\x83\x0A\x00\x80\x86\x89\x00\x00\x87\x95\x00\x00\x84\x5D\x06\x00\xB1\x7D"
-     "\x07\x00"
-     "\x82\x91\x00\x00\x83\x51\x06\x00\x80\xE5\x06\x00\x81\x11\x06\x00\x89\x0A"
-     "\x00\x80"
-     "\x8F\xFD\x07\x00\x8C\xE5\x07\x00\x11\x0A\x00\x80\x8A\x91\x00\x00\x8C\x1E"
-     "\x00\x80"
-     "\x88\xA1\x00\x00\x89\xB1\x07\x00\x95\x0A\x00\x80\xA3\x25\x07\x00\x80\xCD"
-     "\x00\x00"
-     "\x81\xE9\x00\x00\x4B\x0C\x00\x80\xA7\x31\x07\x00\xA7\xF1\x00\x00\x91\xBD"
-     "\x07\x00"
-     "\x80\x89\x00\x00\x81\x95\x00\x00\x51\x0C\x00\x80\x84\x1E\x00\x80\xBC\x1E"
-     "\x00\x80"
-     "\xA2\x0C\x00\x80\x2C\x1C\x00\x80\xAD\x11\x07\x00\x81\xFD\x06\x00\x80\xF1"
-     "\x06\x00"
-     "\x83\x59\x01\x00\x82\x31\x01\x00\x8F\x0A\x00\x80\x84\x49\x01\x00\x87\xB9"
-     "\x01\x00"
-     "\x86\x61\x01\x00\x89\x95\x01\x00\x88\x89\x01\x00\x1C\x1C\x00\x80\x24\x1C"
-     "\x00\x80"
-     "\xCC\x10\x02\x00\x8C\xF5\x01\x00\x8F\xE9\x01\x00\x87\x79\x00\x00\x91\xA9"
-     "\x01\x00"
-     "\x34\x1C\x00\x80\x80\xBD\x06\x00\x81\xB1\x06\x00\xCD\xC4\x02\x00\xCC\x38"
-     "\x02\x00"
-     "\x21\x0C\x00\x80\x8C\xD9\x00\x00\x9B\x0A\x00\x80\x82\x05\x00\x00\x81\x31"
-     "\x00\x00"
-     "\x3C\x1C\x00\x80\x87\xF1\x00\x00\x86\x0D\x00\x00\xA1\x0A\x00\x80\x84\x11"
-     "\x00\x00"
-     "\xA7\x0A\x00\x80\x00\x0C\x00\x80\xA3\x65\x01\x00\xAD\x0A\x00\x80\x81\xC9"
-     "\x00\x00"
-     "\xCC\x98\x01\x00\xA7\x6D\x01\x00\xB3\x0A\x00\x80\x44\x1C\x00\x80\xA8\xB5"
-     "\x00\x00"
-     "\x54\x1C\x00\x80\x93\x0C\x00\x80\xAD\x99\x00\x00\x4C\x1C\x00\x80\x5C\x1C"
-     "\x00\x80"
-     "\x64\x1C\x00\x80\x10\x1E\x00\x80\xB0\xE9\x00\x00\xCC\x84\x01\x00\xCD\xFC"
-     "\x01\x00"
-     "\x14\x0A\x00\x80\x17\x0A\x00\x80\x1A\x0A\x00\x80\xBF\x0A\x00\x80\xC5\x0A"
-     "\x00\x80"
-     "\xCB\x0A\x00\x80\x1D\x0A\x00\x80\x81\x19\x00\x00\xA7\x7D\x00\x00\x74\x1C"
-     "\x00\x80"
-     "\xC4\x1E\x00\x80\xC4\x0B\x00\x80\xCC\x8C\x02\x00\x82\x99\x03\x00\x91\x1D"
-     "\x00\x00"
-     "\x57\x0C\x00\x80\x87\xB1\x03\x00\x5D\x0C\x00\x80\xCC\x1E\x00\x80\xD1\x0A"
-     "\x00\x80"
-     "\x6C\x1C\x00\x80\xDD\x0A\x00\x80\x1B\x0C\x00\x80\x88\x85\x03\x00\x80\xE1"
-     "\x03\x00"
-     "\x81\xE1\x03\x00\x7C\x1C\x00\x80\x8C\x91\x03\x00\x83\x9D\x03\x00\x82\x9D"
-     "\x03\x00"
-     "\x81\x9D\x03\x00\x80\x9D\x03\x00\x06\x0C\x00\x80\x86\xB1\x03\x00\x84\x1C"
-     "\x00\x80"
-     "\x84\xBD\x03\x00\xCC\x2C\x03\x00\xCD\xF4\x02\x00\x89\x91\x03\x00\x88\x85"
-     "\x03\x00"
-     "\x8F\x7D\x00\x00\xE3\x0A\x00\x80\x8C\x1C\x00\x80\x8C\x81\x03\x00\xA3\x35"
-     "\x03\x00"
-     "\x82\xA1\x00\x00\xE9\x0A\x00\x80\xB0\xAD\x00\x00\xA7\x29\x03\x00\x81\xAD"
-     "\x00\x00"
-     "\x84\x31\x00\x00\xF5\x0A\x00\x80\xCC\xF4\x03\x00\xCD\x6C\x02\x00\x9C\x1C"
-     "\x00\x80"
-     "\xEF\x0A\x00\x80\xAC\x1C\x00\x80\xAE\x09\x03\x00\xA4\x1C\x00\x80\x8C\x95"
-     "\x00\x00"
-     "\xA3\xD9\x00\x00\xB4\x1C\x00\x80\xB1\x09\x03\x00\x81\x45\x00\x00\xBC\x1C"
-     "\x00\x80"
-     "\xC4\x1C\x00\x80\xCC\x30\x01\x00\xCD\x34\x01\x00\x81\x25\x00\x00\x27\x0C"
-     "\x00\x80"
-     "\xFB\x0A\x00\x80\xA8\xF1\x00\x00\xCD\x14\x02\x00\xCC\xB4\x03\x00\xCC\xF8"
-     "\x1C\x00"
-     "\xCD\xFC\x03\x00\x80\xE5\x01\x00\x81\xED\x01\x00\xCD\xC0\x01\x00\xCC\x6C"
-     "\x01\x00"
-     "\xCC\x80\x07\x00\xCD\x84\x07\x00\xCC\xDC\x06\x00\xCD\xD8\x06\x00\xCC\xB4"
-     "\x07\x00"
-     "\x23\x0A\x00\x80\xCC\x38\x06\x00\xCC\x40\x06\x00\xCC\x84\x07\x00\xCD\xFC"
-     "\x07\x00"
-     "\xCC\xE4\x07\x00\xCD\x7C\x06\x00\xCC\x1C\x00\x80\x81\x81\x01\x00\x81\x71"
-     "\x01\x00"
-     "\x07\x0B\x00\x80\x81\xD9\x01\x00\xD4\x1C\x00\x80\x01\x0B\x00\x80\x87\x8D"
-     "\x01\x00"
-     "\xE4\x1C\x00\x80\x8C\xF9\x01\x00\xCD\x88\x01\x00\xCC\x8C\x01\x00\xDC\x1C"
-     "\x00\x80"
-     "\x26\x0A\x00\x80\x29\x0A\x00\x80\x8C\x2D\x01\x00\xCC\xCC\x1D\x00\xCD\x5C"
-     "\x1D\x00"
-     "\xB1\xF1\x01\x00\xEC\x1E\x00\x80\xB1\x29\x01\x00\x83\x31\x00\x00\x80\x15"
-     "\x00\x00"
-     "\x81\x3D\x00\x00\xE4\x1E\x00\x80\xFC\x1E\x00\x80\xF4\x1E\x00\x80\xEC\x1C"
-     "\x00\x80"
-     "\x80\x3D\x00\x00\x81\x15\x00\x00\xAE\x0C\x00\x80\x89\x15\x00\x00\xA3\x99"
-     "\x01\x00"
-     "\xCC\x24\x06\x00\xA7\x65\x01\x00\xA3\x19\x01\x00\xA7\xA5\x01\x00\xF4\x1C"
-     "\x00\x80"
-     "\xA3\x01\x1C\x00\xCC\x04\x06\x00\xAD\x51\x01\x00\xFC\x1C\x00\x80\x81\x45"
-     "\x00\x00"
-     "\x04\x1D\x00\x80\x80\x15\x00\x00\x81\x19\x00\x00\x16\x0B\x00\x80\xBA\x0C"
-     "\x00\x80"
-     "\x80\xE5\x1E\x00\x81\xED\x1E\x00\x91\x71\x00\x00\x83\x29\x1F\x00\x2C\x0A"
-     "\x00\x80"
-     "\x84\xE9\x1F\x00\x2F\x0A\x00\x80\x87\x5D\x1F\x00\x9B\xA5\x00\x00\x10\x0B"
-     "\x00\x80"
-     "\xB4\x0C\x00\x80\x32\x0A\x00\x80\x8C\x7D\x1F\x00\x84\x0D\x00\x00\xA8\x0C"
-     "\x00\x80"
-     "\xDC\x1E\x00\x80\x83\x59\x00\x00\x82\x55\x00\x00\x81\x31\x1F\x00\x80\x39"
-     "\x1F\x00"
-     "\x87\x79\x00\x00\x86\xAD\x1F\x00\x1C\x0B\x00\x80\x84\x79\x00\x00\x8B\x39"
-     "\x00\x00"
-     "\xCA\x0B\x00\x80\x89\x55\x00\x00\x88\x41\x00\x00\x8F\x05\x00\x00\x63\x0C"
-     "\x00\x80"
-     "\x69\x0C\x00\x80\x8C\x19\x00\x00\x81\x2D\x1F\x00\xA6\x0B\x00\x80\x14\x1F"
-     "\x00\x80"
-     "\xA3\xD9\x1F\x00\x0C\x1F\x00\x80\x84\x2D\x1F\x00\x24\x1F\x00\x80\xA7\xC5"
-     "\x1F\x00"
-     "\x1C\x1F\x00\x80\x88\x71\x1F\x00\x2C\x1F\x00\x80\xD4\x1E\x00\x80\x04\x1F"
-     "\x00\x80"
-     "\xAD\xD5\x1F\x00\x80\x31\x00\x00\x81\x2D\x00\x00\xA3\x6D\x00\x00\xB1\xB5"
-     "\x1F\x00"
-     "\x82\x79\x00\x00\x0C\x0C\x00\x80\x12\x0C\x00\x80\x84\x05\x00\x00\x80\xAD"
-     "\x1F\x00"
-     "\x81\xA5\x1F\x00\xCC\xEC\x03\x00\xCD\x18\x02\x00\xCC\x60\x1C\x00\xA8\x21"
-     "\x00\x00"
-     "\x0C\x1D\x00\x80\x14\x1D\x00\x80\x81\x09\x00\x00\x25\x0B\x00\x80\x80\x89"
-     "\x01\x00"
-     "\x81\x95\x01\x00\x81\xF9\x01\x00\x83\x95\x01\x00\x87\xE5\x01\x00\x91\x25"
-     "\x00\x00"
-     "\xCC\x48\x02\x00\xCD\x50\x02\x00\x1C\x1D\x00\x80\x89\x85\x01\x00\x31\x0B"
-     "\x00\x80"
-     "\x6F\x0C\x00\x80\x75\x0C\x00\x80\x24\x1D\x00\x80\x84\x05\x00\x00\x2C\x1D"
-     "\x00\x80"
-     "\x2B\x0B\x00\x80\x34\x1D\x00\x80\x81\xB5\x01\x00\x87\x6D\x00\x00\x8C\xD9"
-     "\x00\x00"
-     "\x81\xA9\x01\x00\xCC\x38\x02\x00\xCD\xE8\x02\x00\x8C\x49\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\xE9\x00\x00\x8F\x41\x00\x00\x37\x0B\x00\x80\x87\x0D\x00\x00\x4C\x1D"
-     "\x00\x80"
-     "\x87\xF9\x00\x00\x3D\x0B\x00\x80\x3C\x1D\x00\x80\x49\x0B\x00\x80\xA3\x25"
-     "\x01\x00"
-     "\x87\x09\x00\x00\xB1\xD1\x00\x00\x54\x1D\x00\x80\x44\x1D\x00\x80\x87\x05"
-     "\x00\x00"
-     "\x5C\x1D\x00\x80\x87\x0C\x00\x80\x43\x0B\x00\x80\x81\xC1\x00\x00\x64\x1D"
-     "\x00\x80"
-     "\xCC\xA4\x01\x00\x14\x1E\x00\x80\x55\x0B\x00\x80\x6C\x1D\x00\x80\x8D\x0C"
-     "\x00\x80"
-     "\xA7\x8D\x00\x00\x4F\x0B\x00\x80\x7C\x1D\x00\x80\x74\x1D\x00\x80\xA3\xE9"
-     "\x00\x00"
-     "\xCC\xBC\x01\x00\xCD\xDC\x00\x00\x38\x0A\x00\x80\xA3\x71\x00\x00\x3B\x0A"
-     "\x00\x80"
-     "\x3E\x0A\x00\x80\xA6\x51\x00\x00\xA7\x51\x00\x00\x87\xB5\x03\x00\xCC\x2C"
-     "\x01\x00"
-     "\x81\x05\x00\x00\x9C\x1D\x00\x80\x88\x05\x00\x00\xA4\x1D\x00\x80\x91\xCD"
-     "\x00\x00"
-     "\x67\x0B\x00\x80\xE8\x0B\x00\x80\xCC\x28\x03\x00\xCC\x90\x03\x00\x8C\xED"
-     "\x03\x00"
-     "\x9B\x21\x06\x00\xDC\x0B\x00\x80\xD6\x0B\x00\x80\x88\x81\x03\x00\xCC\x28"
-     "\x01\x00"
-     "\x86\x59\x00\x00\xE2\x0B\x00\x80\x84\x4D\x06\x00\x83\x01\x04\x00\x82\xA1"
-     "\x03\x00"
-     "\x81\xA9\x03\x00\x80\xB5\x03\x00\x8F\xA9\x00\x00\x3C\x1F\x00\x80\x81\x91"
-     "\x01\x00"
-     "\x8C\xD1\x00\x00\x8B\xC1\x00\x00\x8A\xED\x00\x00\x89\x1D\x00\x00\x88\x75"
-     "\x06\x00"
-     "\xA7\x51\x03\x00\xA6\x4D\x03\x00\x81\xBD\x00\x00\x80\xAD\x00\x00\xA3\x55"
-     "\x03\x00"
-     "\x8C\xDD\x00\x00\x6D\x0B\x00\x80\xB0\xAD\x01\x00\x81\x95\x00\x00\x80\x85"
-     "\x00\x00"
-     "\xAD\x7D\x03\x00\xCC\x10\x01\x00\x73\x0B\x00\x80\x84\x8D\x00\x00\xCC\x34"
-     "\x01\x00"
-     "\xD0\x0B\x00\x80\x80\x1D\x00\x00\x81\x25\x00\x00\x7B\x0C\x00\x80\xA4\x35"
-     "\x00\x00"
-     "\xA3\x25\x00\x00\x81\x0C\x00\x80\xB1\x01\x03\x00\x4C\x1F\x00\x80\x44\x1F"
-     "\x00\x80"
-     "\x5C\x1F\x00\x80\xAD\xE1\x01\x00\x54\x1F\x00\x80\x64\x1F\x00\x80\x34\x1F"
-     "\x00\x80"
-     "\x84\x1D\x00\x80\xA8\xF9\x01\x00\x83\x35\x00\x00\x79\x0B\x00\x80\x94\x1D"
-     "\x00\x80"
-     "\x81\xC5\x07\x00\xCD\xE0\x1E\x00\xCC\xE4\x1E\x00\xCC\x1C\x01\x00\xCC\x38"
-     "\x01\x00"
-     "\xCC\xB8\x1E\x00\xCD\x94\x1E\x00\xCC\xAC\x01\x00\xCD\xC0\x1E\x00\x8C\x1D"
-     "\x00\x80"
-     "\xAC\x1D\x00\x80\x80\xD1\x02\x00\x81\xCD\x02\x00\xB4\x1D\x00\x80\x83\xA1"
-     "\x01\x00"
-     "\x80\x95\x01\x00\x81\xAD\x01\x00\xCC\x74\x05\x00\xCD\xD8\x02\x00\xBC\x1D"
-     "\x00\x80"
-     "\xC4\x1D\x00\x80\x81\x9D\x07\x00\x7F\x0B\x00\x80\x81\x7D\x06\x00\x89\x81"
-     "\x01\x00"
-     "\xD4\x1D\x00\x80\xCC\x1D\x00\x80\x18\x1E\x00\x80\xDC\x1D\x00\x80\xA3\xC1"
-     "\x00\x00"
-     "\xE4\x1D\x00\x80\x81\xD5\x00\x00\x80\xD5\x00\x00\xEC\x1D\x00\x80\x6C\x1F"
-     "\x00\x80"
-     "\x88\x09\x00\x00\xCC\x5C\x01\x00\x44\x0A\x00\x80\x87\x21\x00\x00\xCD\x44"
-     "\x02\x00"
-     "\xCC\xEC\x05\x00\xCD\x08\x01\x00\xCC\xB4\x00\x00\x85\x0B\x00\x80\xB7\x5C"
-     "\x03\x00"
-     "\x84\x1F\x00\x80\xA3\x35\x01\x00\xC0\x0C\x00\x80\xF4\x1D\x00\x80\xCC\x88"
-     "\x04\x00"
-     "\x47\x0A\x00\x80\x94\x1C\x00\x80\x7C\x1F\x00\x80\x1C\x1E\x00\x80\x74\x1F"
-     "\x00\x80"
-     "\x81\x49\x00\x00\x80\x79\x00\x00\xCD\x9C\x03\x00\xCC\x64\x05\x00\x8E\x0B"
-     "\x00\x80"
-     "\xCC\x38\x01\x00\x61\x0B\x00\x80\x88\x71\x07\x00\x41\x0A\x00\x80\x8A\x75"
-     "\x07\x00"
-     "\xAF\x0B\x00\x80\xFC\x1D\x00\x80\x94\x0B\x00\x80\xA3\x2D\x00\x00\x81\x7D"
-     "\x07\x00"
-     "\x80\x7D\x07\x00\x9A\x0B\x00\x80\x82\x49\x07\x00\x04\x1E\x00\x80\x0C\x1E"
-     "\x00\x80"
-     "\x87\x49\x07\x00\x00\x00\x00\x80\x80\xD5\x07\x00\x81\xE5\x07\x00\x82\xF1"
-     "\x07\x00"
-     "\x83\x8D\x07\x00\x84\x99\x07\x00\xB1\x61\x00\x00\x1A\x00\x00\x80\x87\x91"
-     "\x07\x00"
-     "\x88\xB5\x07\x00\x89\xB9\x07\x00\x8A\x89\x07\x00\xE9\x07\x00\x80\x0A\x00"
-     "\x00\x80"
-     "\xEA\x04\x00\x80\x89\x05\x00\x80\x02\x00\x00\x80\x61\x0E\x00\x80\x64\x00"
-     "\x00\x80"
-     "\xC7\x03\x00\x80\x87\x95\x00\x00\xA0\xAD\x00\x00\x81\x6D\x07\x00\x82\x8D"
-     "\x00\x00"
-     "\x1D\x09\x00\x80\x8C\x89\x00\x00\xB7\x04\x00\x80\xA3\xD5\x07\x00\xAF\xDD"
-     "\x00\x00"
-     "\xA8\xE9\x00\x00\xAB\x04\x00\x80\xAA\xD5\x00\x00\xBB\x05\x00\x80\xB4\xBD"
-     "\x00\x00"
-     "\xB5\x45\x00\x00\x30\x0C\x00\x80\xA3\x29\x07\x00\xCC\xD0\x03\x00\xCD\xDC"
-     "\x05\x00"
-     "\xB2\xAD\x00\x00\xB3\xA5\x00\x00\xBC\x45\x00\x00\xBD\x51\x00\x00\xBE\x55"
-     "\x00\x00"
-     "\xF7\x0B\x00\x80\xB8\x75\x00\x00\xB9\x7D\x00\x00\xBA\x65\x00\x00\x36\x0C"
-     "\x00\x80"
-     "\x86\xE0\x02\x00\x81\xCD\x01\x00\xCC\x6C\x04\x00\xCD\x68\x04\x00\x84\xFD"
-     "\x01\x00"
-     "\x33\x0C\x00\x80\xCC\x3C\x02\x00\xCD\x70\x04\x00\x81\x35\x00\x00\xFA\x0B"
-     "\x00\x80"
-     "\xCC\x1C\x02\x00\xCD\x18\x02\x00\x81\x09\x00\x00\x84\x11\x00\x00\x39\x0C"
-     "\x00\x80"
-     "\x20\x07\x00\x80\xB2\x7D\x00\x00\xB3\x0D\x00\x00\x75\x08\x00\x80\x45\x07"
-     "\x00\x80"
-     "\xC5\x08\x00\x80\x80\x05\x00\x00\xB3\x0D\x00\x80\x54\x09\x00\x80\xCC\x6C"
-     "\x05\x00"
-     "\xCD\x68\x05\x00\x8C\x05\x00\x00\x15\x0C\x00\x80\x89\x29\x00\x00\xBF\x39"
-     "\x00\x00"
-     "\x98\x40\x03\x00\xE9\x06\x00\x80\xA6\xA8\x03\x00\xED\x06\x00\x80\x3A\x08"
-     "\x00\x80"
-     "\x3B\x07\x00\x80\x42\x07\x00\x80\xB5\x08\x00\x80\x6B\x07\x00\x80\x6E\x07"
-     "\x00\x80"
-     "\x14\x09\x00\x80\xE3\x06\x00\x80\xBF\xF5\x00\x00\xE6\x06\x00\x80\x37\x08"
-     "\x00\x80"
-     "\xCC\x38\x02\x00\x18\x0C\x00\x80\x8C\x05\x00\x00\x63\x08\x00\x80\x02\x0D"
-     "\x00\x80"
-     "\x81\x08\x00\x80\x3B\x0D\x00\x80\x3D\x09\x00\x80\x41\x0D\x00\x80\x4D\x0D"
-     "\x00\x80"
-     "\x6E\x09\x00\x80\x12\x00\x00\x80\x16\x00\x00\x80\x30\x00\x00\x80\x68\x00"
-     "\x00\x80"
-     "\x79\x09\x00\x80\x06\x00\x00\x80\xB8\xE0\x03\x00\x34\x00\x00\x80\x87\x8D"
-     "\x03\x00"
-     "\x86\x8D\x03\x00\x85\x8D\x03\x00\x84\x71\x03\x00\x85\x05\x01\x00\x92\x90"
-     "\x00\x00"
-     "\x85\x35\x01\x00\xF9\x0C\x00\x80\x7A\x0D\x00\x80\x81\xD5\x37\x00\x83\x51"
-     "\x09\x00"
-     "\x8C\x8D\x03\x00\x8B\xAD\x03\x00\x8A\xAD\x03\x00\x89\xAD\x03\x00\x88\xAD"
-     "\x03\x00"
-     "\xA1\x01\x0C\x00\xA0\x59\x00\x00\xA3\x49\x0E\x00\xA2\x11\x03\x00\x5B\x0B"
-     "\x00\x80"
-     "\xA4\x75\x00\x00\xBD\x0D\x00\x80\xC0\x0D\x00\x80\xC8\x0D\x00\x80\xC3\x0D"
-     "\x00\x80"
-     "\xCD\xDC\x05\x00\xCC\x0C\x05\x00\xB0\x0D\x00\x80\x74\x00\x00\x80\xA5\x06"
-     "\x00\x80"
-     "\x24\x00\x00\x80\xB7\x5D\x03\x00\xB6\x5D\x03\x00\xB5\x5D\x03\x00\xB4\x41"
-     "\x03\x00"
-     "\xB3\x41\x03\x00\xB2\x5D\x03\x00\xB1\x5D\x03\x00\xB0\x5D\x03\x00\x99\x45"
-     "\x03\x00"
-     "\x98\x45\x03\x00\x9B\x45\x03\x00\x9A\x45\x03\x00\x9D\x4D\x03\x00\x9C\x41"
-     "\x03\x00"
-     "\x9C\x09\x00\x80\xB8\x4D\x03\x00\xCC\x80\x00\x00\xCD\x28\x07\x00\xCC\xAC"
-     "\x03\x00"
-     "\xCD\x98\x0C\x00\xB3\x61\x03\x00\xB2\x79\x03\x00\xB1\x7D\x03\x00\x18\x22"
-     "\x00\x80"
-     "\xCC\xB0\x05\x00\xCD\xA8\x05\x00\xCD\x64\x07\x00\xCC\xC8\x00\x00\xCC\xFC"
-     "\x1A\x00"
-     "\xCB\x0D\x00\x80\xCC\x74\x1A\x00\x14\x22\x00\x80\x80\x59\x01\x00\x81\x59"
-     "\x01\x00"
-     "\x10\x22\x00\x80\x83\x6D\x01\x00\x84\x75\x01\x00\x5C\x21\x00\x80\x80\x01"
-     "\x0F\x00"
-     "\x81\x01\x0F\x00\xCC\x68\x06\x00\xCD\xE4\x07\x00\x93\x39\x0F\x00\x64\x21"
-     "\x00\x80"
-     "\xCC\x00\x07\x00\x94\x1D\x0F\x00\xCC\x9C\x07\x00\xCD\xB0\x07\x00\x80\x59"
-     "\x0F\x00"
-     "\x81\x59\x0F\x00\x82\x69\x0F\x00\x6C\x21\x00\x80\x85\x1D\x00\x00\x85\x3D"
-     "\x00\x00"
-     "\xCD\x08\x1E\x00\xCC\xC4\x19\x00\x81\x95\x01\x00\x80\xB9\x01\x00\x54\x21"
-     "\x00\x80"
-     "\x85\x1D\x00\x00\x60\x21\x00\x80\x84\x89\x01\x00\x68\x21\x00\x80\x86\x75"
-     "\x00\x00"
-     "\x80\xA5\x0F\x00\x81\xB5\x0F\x00\xCC\xB0\x0D\x00\xCD\x38\x0D\x00\x70\x21"
-     "\x00\x80"
-     "\x58\x21\x00\x80\x1C\x22\x00\x80\x34\x22\x00\x80\x85\xD9\x00\x00\xCE\x0D"
-     "\x00\x80"
-     "\xCD\x70\x05\x00\xCC\xE0\x05\x00\xE8\x1F\x00\x80\xF0\x1F\x00\x80\xEC\x1F"
-     "\x00\x80"
-     "\x85\x2D\x00\x00\xF4\x1F\x00\x80\x38\x22\x00\x80\xCC\xC8\x0C\x00\xCD\xA4"
-     "\x0C\x00"
-     "\xB4\xA1\x01\x00\xD1\x0D\x00\x80\x9C\x21\x00\x80\xA4\x21\x00\x80\xCC\xAC"
-     "\x0F\x00"
-     "\xAC\x21\x00\x80\xBA\x9D\x01\x00\x94\x21\x00\x80\xA0\x21\x00\x80\xA8\x21"
-     "\x00\x80"
-     "\xBE\x81\x01\x00\xB0\x21\x00\x80\x70\x09\x00\x80\x87\xBD\x0E\x00\x84\x75"
-     "\x0C\x00"
-     "\x85\x71\x0E\x00\xA8\x1F\x00\x80\x85\xCD\x0F\x00\xB0\x1F\x00\x80\xB8\x1F"
-     "\x00\x80"
-     "\xC0\x1F\x00\x80\xB4\x1F\x00\x80\xAC\x1F\x00\x80\x98\x21\x00\x80\xBC\x1F"
-     "\x00\x80"
-     "\xC4\x1F\x00\x80\x80\xA5\x0F\x00\x81\x99\x0F\x00\x85\xD9\x0F\x00\x97\xBC"
-     "\x0C\x00"
-     "\xE0\x1F\x00\x80\x95\xF4\x0C\x00\x81\xA5\x0E\x00\x80\x89\x0E\x00\x83\x49"
-     "\x00\x00"
-     "\x91\xA4\x0D\x00\xE4\x1F\x00\x80\x9F\xBC\x02\x00\x94\x09\x00\x00\x18\x20"
-     "\x00\x80"
-     "\x20\x20\x00\x80\x93\x3D\x00\x00\x28\x20\x00\x80\x99\x74\x02\x00\x80\xA5"
-     "\x00\x00"
-     "\x81\xB5\x00\x00\x82\xBD\x00\x00\xA5\x94\x01\x00\x85\x4D\x0F\x00\x85\x01"
-     "\x0F\x00"
-     "\x30\x20\x00\x80\xA1\x4C\x0C\x00\x80\x11\x0F\x00\x81\x11\x0F\x00\x80\x11"
-     "\x0F\x00"
-     "\x81\x6D\x0F\x00\x81\x15\x0F\x00\x80\x29\x0F\x00\x83\xF1\x00\x00\xA9\x2C"
-     "\x01\x00"
-     "\x85\x25\x0F\x00\xB7\x0C\x07\x00\x1C\x20\x00\x80\xB5\x3C\x07\x00\x80\xA1"
-     "\x00\x00"
-     "\x81\xAD\x00\x00\x24\x20\x00\x80\xB1\x4C\x00\x00\x85\x11\x0F\x00\xBF\x3C"
-     "\x05\x00"
-     "\x80\x31\x00\x00\x81\x1D\x00\x00\x2C\x20\x00\x80\x85\x01\x0F\x00\x34\x20"
-     "\x00\x80"
-     "\xB9\x34\x06\x00\x80\xC5\x0E\x00\x81\xD1\x0E\x00\x85\xF5\x0E\x00\x85\xE1"
-     "\x0E\x00"
-     "\x82\x89\x01\x00\x3C\x22\x00\x80\x80\x69\x01\x00\x81\x65\x01\x00\x80\xF9"
-     "\x01\x00"
-     "\x81\xD5\x01\x00\x82\xD1\x01\x00\x5C\x22\x00\x80\xD4\x0D\x00\x80\x85\x19"
-     "\x01\x00"
-     "\xCC\xB8\x02\x00\xCD\xE4\x02\x00\x58\x22\x00\x80\x54\x22\x00\x80\x2B\x0E"
-     "\x00\x80"
-     "\x93\x21\x01\x00\x94\x99\x01\x00\x58\x20\x00\x80\x80\x39\x00\x00\x81\x25"
-     "\x00\x00"
-     "\x60\x20\x00\x80\x83\x09\x01\x00\x80\xAD\x0E\x00\x81\xB9\x0E\x00\xCC\x24"
-     "\x14\x00"
-     "\xCD\xBC\x15\x00\x68\x20\x00\x80\x85\x69\x00\x00\x80\xAD\x00\x00\x81\xB5"
-     "\x00\x00"
-     "\x70\x20\x00\x80\x5C\x20\x00\x80\x84\xD1\x00\x00\x64\x20\x00\x80\x86\xDD"
-     "\x00\x00"
-     "\x6C\x20\x00\x80\x88\xE9\x00\x00\x82\x2D\x00\x00\x81\xD1\x00\x00\x80\xCD"
-     "\x00\x00"
-     "\x80\x25\x00\x00\x81\x29\x00\x00\x74\x20\x00\x80\xA0\x22\x00\x80\x80\x55"
-     "\x00\x00"
-     "\x81\x59\x00\x00\x82\x71\x00\x00\x93\x99\x00\x00\x94\x5D\x00\x00\x83\x81"
-     "\x00\x00"
-     "\x80\xF5\x00\x00\x81\xED\x00\x00\xD7\x0D\x00\x80\x90\x20\x00\x80\xCC\xEC"
-     "\x01\x00"
-     "\xCD\xD4\x01\x00\x80\x09\x00\x00\x81\x09\x00\x00\x98\x20\x00\x80\xA0\x20"
-     "\x00\x80"
-     "\x81\xE1\x03\x00\x80\xB9\x03\x00\x80\x19\x00\x00\x81\x19\x00\x00\x9C\x20"
-     "\x00\x80"
-     "\xA4\x20\x00\x80\x85\xB5\x00\x00\x94\x3D\x00\x00\x88\x22\x00\x80\x84\x22"
-     "\x00\x80"
-     "\xCC\xA0\x02\x00\xCD\xAC\x02\x00\xDA\x0D\x00\x80\x83\xD1\x03\x00\x80\x85"
-     "\x03\x00"
-     "\x81\xDD\x03\x00\xB3\x08\x00\x80\x80\x22\x00\x80\x93\xAD\x03\x00\x7C\x22"
-     "\x00\x80"
-     "\x2E\x0E\x00\x80\x94\x01\x04\x00\x80\x31\x00\x00\x81\x31\x00\x00\xCD\xCC"
-     "\x01\x00"
-     "\xCC\x3C\x01\x00\xCC\x20\x00\x80\xD0\x20\x00\x80\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xD4\x20\x00\x80\x82\x05\x00\x00\x80\xA5\x00\x00\x81\xB5\x00\x00\x80\xAD"
-     "\x00\x00"
-     "\x81\xBD\x00\x00\x86\xDD\x00\x00\xA4\x22\x00\x80\x84\xDD\x00\x00\xDD\x0D"
-     "\x00\x80"
-     "\xCC\x10\x02\x00\xCD\x3C\x03\x00\x88\xF9\x00\x00\xDC\x21\x00\x80\x80\x25"
-     "\x00\x00"
-     "\x81\x29\x00\x00\xCD\x78\x01\x00\xCC\x0C\x01\x00\x80\xD1\x06\x00\x81\xD1"
-     "\x06\x00"
-     "\x85\x05\x00\x00\xE4\x21\x00\x80\xCD\x20\x01\x00\xCC\x24\x01\x00\x94\xD9"
-     "\x06\x00"
-     "\xEC\x21\x00\x80\x94\xC1\x06\x00\x82\xFD\x06\x00\x81\xFD\x06\x00\x80\xFD"
-     "\x06\x00"
-     "\x85\x2D\x00\x00\x85\xF9\x01\x00\x85\xF1\x01\x00\x93\xF5\x06\x00\xCD\xA8"
-     "\x0E\x00"
-     "\xCD\xB0\x0E\x00\xD4\x21\x00\x80\xE0\x21\x00\x80\xCD\x68\x0E\x00\xCD\x50"
-     "\x0E\x00"
-     "\xCC\x98\x0E\x00\xCD\x9C\x01\x00\xCC\x38\x0E\x00\xCD\x64\x0E\x00\xCD\x24"
-     "\x0F\x00"
-     "\xCD\x28\x0F\x00\xCC\x88\x0A\x00\x85\x75\x00\x00\x80\x25\x07\x00\x81\x15"
-     "\x07\x00"
-     "\xE8\x21\x00\x80\xCD\x44\x01\x00\x85\x3D\x00\x00\xCD\xC8\x01\x00\xCC\xF8"
-     "\x01\x00"
-     "\xCD\xFC\x01\x00\xCD\x40\x00\x00\xCC\x8C\x01\x00\x81\x49\x07\x00\x80\x79"
-     "\x07\x00"
-     "\x85\xF9\x00\x00\x82\x41\x07\x00\x85\xF5\x00\x00\xF0\x21\x00\x80\xCC\xD8"
-     "\x01\x00"
-     "\xCD\xDC\x01\x00\x88\x51\x3A\x00\xD8\x21\x00\x80\xA8\x22\x00\x80\x85\x05"
-     "\x00\x00"
-     "\xFC\x21\x00\x80\x83\x9D\x07\x00\x80\x0D\x01\x00\x81\x01\x01\x00\xCC\xC8"
-     "\x05\x00"
-     "\xCD\x04\x05\x00\x04\x22\x00\x80\xCD\x78\x01\x00\x85\x19\x00\x00\xF8\x21"
-     "\x00\x80"
-     "\xF4\x21\x00\x80\x85\x75\x00\x00\x85\x1D\x00\x00\xCD\x24\x01\x00\x3C\x21"
-     "\x00\x80"
-     "\xCD\xD0\x00\x00\x80\x91\x07\x00\x81\x91\x07\x00\x85\x39\x00\x00\x44\x21"
-     "\x00\x80"
-     "\x4C\x21\x00\x80\x34\x21\x00\x80\x40\x21\x00\x80\xCD\x18\x01\x00\x80\x91"
-     "\x07\x00"
-     "\x81\x91\x07\x00\xCC\x10\x01\x00\xCD\xDC\x00\x00\xC8\x20\x00\x80\x85\xED"
-     "\x07\x00"
-     "\xF8\x20\x00\x80\x85\x79\x00\x00\x00\x21\x00\x80\x08\x21\x00\x80\x10\x21"
-     "\x00\x80"
-     "\x04\x21\x00\x80\xFC\x20\x00\x80\x93\x19\x00\x00\x0C\x21\x00\x80\x14\x21"
-     "\x00\x80"
-     "\x18\x21\x00\x80\x31\x0E\x00\x80\x94\x6D\x00\x00\x88\x1F\x00\x80\x90\x1F"
-     "\x00\x80"
-     "\x98\x1F\x00\x80\xA0\x1F\x00\x80\x85\xA9\x07\x00\x82\x19\x00\x00\x8C\x1F"
-     "\x00\x80"
-     "\x80\x19\x00\x00\x81\x19\x00\x00\x94\x1F\x00\x80\x9C\x1F\x00\x80\x80\x59"
-     "\x00\x00"
-     "\x81\x59\x00\x00\x86\xC9\x07\x00\x48\x21\x00\x80\x84\xCD\x07\x00\xA4\x1F"
-     "\x00\x80"
-     "\x80\xE1\x00\x00\x81\xE1\x00\x00\x82\xF5\x00\x00\x85\x1D\x00\x00\x50\x21"
-     "\x00\x80"
-     "\x85\x0D\x00\x00\x38\x21\x00\x80\x08\x22\x00\x80\x80\x91\x00\x00\x81\x91"
-     "\x00\x00"
-     "\x82\x35\x00\x00\x83\x91\x00\x00\x85\x09\x00\x00\x85\x09\x00\x00\x0C\x22"
-     "\x00\x80"
-     "\x00\x22\x00\x80\xCC\xC0\x03\x00\xCD\xD4\x03\x00\x1C\x21\x00\x80\x34\x0E"
-     "\x00\x80"
-     "\xC8\x1F\x00\x80\xD0\x1F\x00\x80\x80\x0D\x00\x00\x81\x3D\x00\x00\xD8\x1F"
-     "\x00\x80"
-     "\xCC\x1F\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\xD4\x1F\x00\x80\xDC\x1F"
-     "\x00\x80"
-     "\xCC\xF0\x03\x00\xCD\x1C\x03\x00\x80\xC9\x01\x00\x81\xC9\x01\x00\x80\x11"
-     "\x00\x00"
-     "\x81\x25\x00\x00\x80\xD9\x01\x00\x81\xD9\x01\x00\x20\x21\x00\x80\x83\xCD"
-     "\x01\x00"
-     "\x85\x05\x00\x00\x20\x22\x00\x80\x37\x0E\x00\x80\x85\x1D\x00\x00\x28\x22"
-     "\x00\x80"
-     "\xF8\x1F\x00\x80\x81\x65\x00\x00\x80\x7D\x00\x00\x00\x20\x00\x80\x93\x71"
-     "\x00\x00"
-     "\x85\x1D\x00\x00\x93\x9D\x01\x00\x94\xB5\x01\x00\x7C\x21\x00\x80\x94\xD5"
-     "\x00\x00"
-     "\x08\x20\x00\x80\xCD\x68\x05\x00\xCC\x6C\x05\x00\x85\x05\x00\x00\x84\x21"
-     "\x00\x80"
-     "\x10\x20\x00\x80\x85\x0D\x00\x00\x8C\x21\x00\x80\x74\x21\x00\x80\x80\xC1"
-     "\x00\x00"
-     "\x81\xD9\x00\x00\x82\xF9\x00\x00\xFC\x1F\x00\x80\x04\x20\x00\x80\x85\xE9"
-     "\x00\x00"
-     "\x80\x09\x00\x00\x81\x35\x00\x00\x85\x05\x00\x00\x80\x21\x00\x80\x0C\x20"
-     "\x00\x80"
-     "\x85\x1D\x00\x00\x88\x21\x00\x80\x14\x20\x00\x80\x85\x05\x00\x00\x90\x21"
-     "\x00\x80"
-     "\x80\x51\x00\x00\x81\x6D\x00\x00\x82\x7D\x00\x00\x78\x21\x00\x80\x2C\x22"
-     "\x00\x80"
-     "\x85\x19\x00\x00\x85\x05\x00\x00\x30\x22\x00\x80\x80\xF9\x00\x00\x81\xCD"
-     "\x00\x00"
-     "\x82\x39\x00\x00\x83\xD9\x00\x00\x24\x22\x00\x80\x85\x05\x00\x00\xCC\xA0"
-     "\x01\x00"
-     "\xCD\x24\x01\x00\xCD\x80\x01\x00\xCC\x38\x01\x00\x94\x01\x3E\x00\x24\x21"
-     "\x00\x80"
-     "\x85\xC9\x3E\x00\x93\x15\x3E\x00\x85\xF9\x3E\x00\xCD\x24\x01\x00\x3A\x0E"
-     "\x00\x80"
-     "\x44\x22\x00\x80\x40\x22\x00\x80\xCD\x18\x01\x00\x48\x22\x00\x80\xE0\x0D"
-     "\x00\x80"
-     "\x80\x09\x00\x00\x81\x09\x00\x00\x86\x69\x00\x00\xCD\x24\x01\x00\x84\x6D"
-     "\x00\x00"
-     "\x50\x22\x00\x80\x40\x20\x00\x80\x94\x20\x00\x80\x80\x55\x00\x00\x81\x7D"
-     "\x00\x00"
-     "\x48\x20\x00\x80\x50\x20\x00\x80\xCD\x14\x01\x00\x44\x20\x00\x80\x4C\x20"
-     "\x00\x80"
-     "\xCD\x28\x01\x00\x88\x01\x3C\x00\x54\x20\x00\x80\xCD\x2C\x01\x00\x82\xC9"
-     "\x00\x00"
-     "\x81\xBD\x00\x00\x80\xBD\x00\x00\x81\xF1\x00\x00\x80\xC5\x00\x00\x8F\x31"
-     "\x3A\x00"
-     "\xCD\x24\x01\x00\x82\xC5\x00\x00\xCD\x10\x01\x00\x80\xF9\x00\x00\x81\xCD"
-     "\x00\x00"
-     "\x80\xDD\x00\x00\x81\xC5\x00\x00\xCD\x78\x01\x00\xCC\x24\x01\x00\x80\xAD"
-     "\x00\x00"
-     "\x81\xB5\x00\x00\x82\xB5\x00\x00\x4C\x22\x00\x80\x82\x1D\x00\x00\x83\xD5"
-     "\x3E\x00"
-     "\x80\xD5\x00\x00\x81\xFD\x00\x00\x28\x21\x00\x80\xCD\x28\x01\x00\xCC\x60"
-     "\x03\x00"
-     "\xCD\x78\x03\x00\xB1\x0E\x00\x80\x80\x20\x00\x80\x80\x0D\x00\x00\x81\xBD"
-     "\x01\x00"
-     "\xCD\xA4\x06\x00\xCD\x78\x06\x00\xCD\x50\x01\x00\xCC\xDC\x06\x00\xCC\x28"
-     "\x06\x00"
-     "\xCD\x70\x06\x00\xCD\x44\x06\x00\xCD\x3C\x06\x00\xCD\xFC\x07\x00\xCD\x28"
-     "\x07\x00"
-     "\xCD\xC8\x07\x00\xCD\xC8\x07\x00\xCC\x68\x07\x00\xCD\x40\x07\x00\xCC\xF8"
-     "\x07\x00"
-     "\xCD\xD0\x07\x00\x88\x20\x00\x80\x84\x20\x00\x80\xCD\xD0\x01\x00\x8C\x20"
-     "\x00\x80"
-     "\xCD\x48\x01\x00\xCD\x44\x01\x00\x80\x1D\x00\x00\x81\x11\x00\x00\xCD\xE0"
-     "\x01\x00"
-     "\xCC\x94\x06\x00\xCD\x4C\x0B\x00\xCC\x80\x02\x00\xCD\xA4\x06\x00\xCD\xDC"
-     "\x06\x00"
-     "\x6C\x22\x00\x80\xCD\xB0\x06\x00\x80\x75\x3E\x00\x70\x22\x00\x80\x80\x69"
-     "\x01\x00"
-     "\x81\x7D\x01\x00\x81\x61\x01\x00\x80\x75\x01\x00\x83\x89\x3F\x00\x2C\x21"
-     "\x00\x80"
-     "\x93\xD9\x00\x00\x93\xF9\x3F\x00\xCD\x0C\x01\x00\x85\x41\x3F\x00\xCC\x58"
-     "\x02\x00"
-     "\xCD\x04\x3D\x00\x94\x39\x3F\x00\x94\x39\x00\x00\x85\xC1\x3E\x00\x93\xD5"
-     "\x3F\x00"
-     "\x88\xE9\x3F\x00\xCC\x6C\x06\x00\xCD\x38\x01\x00\xCC\x3C\x01\x00\x94\xC1"
-     "\x3F\x00"
-     "\xB4\x0E\x00\x80\xCC\x38\x01\x00\xCD\x3C\x01\x00\x80\x75\x00\x00\x81\x31"
-     "\x00\x00"
-     "\x86\x0D\x00\x00\x64\x22\x00\x80\x84\x0D\x00\x00\x60\x22\x00\x80\x38\x20"
-     "\x00\x80"
-     "\x81\x58\x3C\x00\x3C\x20\x00\x80\x68\x22\x00\x80\x78\x20\x00\x80\x85\x94"
-     "\x3C\x00"
-     "\x7C\x20\x00\x80\x3D\x0E\x00\x80\xAE\x0E\x00\x80\x89\xBC\x03\x00\x80\x25"
-     "\x00\x00"
-     "\x81\x31\x00\x00\x80\x3D\x00\x00\x81\x29\x00\x00\x82\x05\x00\x00\x78\x22"
-     "\x00\x80"
-     "\x90\xE5\x01\x00\x91\xED\x01\x00\x92\xE5\x01\x00\x93\x7D\x3A\x00\x94\x01"
-     "\x3A\x00"
-     "\x95\xE5\x01\x00\x96\xED\x01\x00\xA8\x20\x00\x80\x80\x09\x00\x00\x81\x09"
-     "\x00\x00"
-     "\xB0\x20\x00\x80\xB8\x20\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xC0\x20\x00\x80\xAC\x20\x00\x80\xB4\x20\x00\x80\x80\x0D\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\xBC\x20\x00\x80\x82\x35\x00\x00\x81\x09\x00\x00\x80\x19\x00\x00\xC4\x20"
-     "\x00\x80"
-     "\x74\x22\x00\x80\xCC\xA0\x03\x00\xCD\x2C\x03\x00\x80\x2D\x3F\x00\x81\x69"
-     "\x3F\x00"
-     "\x82\x1D\x00\x00\x83\xE1\x00\x00\xB0\x79\x01\x00\xB1\x79\x01\x00\xB2\x09"
-     "\x01\x00"
-     "\x30\x21\x00\x80\xB4\x15\x01\x00\xB5\x1D\x01\x00\x8C\x22\x00\x80\xB7\x0E"
-     "\x00\x80"
-     "\x94\x22\x00\x80\xB9\x29\x01\x00\xD8\x20\x00\x80\xBC\x21\x00\x80\x80\x09"
-     "\x00\x00"
-     "\x81\x09\x00\x00\xE0\x20\x00\x80\xE8\x20\x00\x80\xC4\x21\x00\x80\x85\x2D"
-     "\x00\x00"
-     "\x80\xC5\x01\x00\x81\xD1\x01\x00\x80\xE9\x01\x00\x81\xE9\x01\x00\x82\x05"
-     "\x00\x00"
-     "\xF0\x20\x00\x80\x85\x05\x00\x00\xCC\x21\x00\x80\xB4\x21\x00\x80\xDC\x20"
-     "\x00\x80"
-     "\x80\x09\x00\x00\x81\x65\x00\x00\xE4\x20\x00\x80\x85\x7D\x00\x00\xC0\x21"
-     "\x00\x80"
-     "\x93\xAD\x01\x00\xCC\x70\x3D\x00\xCD\xC0\x3D\x00\xEC\x20\x00\x80\x85\x09"
-     "\x00\x00"
-     "\x94\x75\x00\x00\xC8\x21\x00\x80\x80\x59\x00\x00\x81\x35\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xF4\x20\x00\x80\x85\x09\x00\x00\x85\x09\x00\x00\xD0\x21\x00\x80\xB8\x21"
-     "\x00\x80"
-     "\x80\x4D\x01\x00\x81\x59\x01\x00\x82\x19\x00\x00\x83\x65\x01\x00\x98\x22"
-     "\x00\x80"
-     "\x85\x35\x00\x00\x85\x05\x00\x00\x9C\x22\x00\x80\x90\x22\x00\x80\x43\x0E"
-     "\x00\x80"
-     "\x55\x0E\x00\x80\x1C\x0E\x00\x80\x9F\x0E\x00\x80\x8D\x0E\x00\x80\x5B\x0E"
-     "\x00\x80"
-     "\x90\x0E\x00\x80\x93\x0E\x00\x80\xF8\x0D\x00\x80\x4C\x0E\x00\x80\x16\x0E"
-     "\x00\x80"
-     "\xC9\x0E\x00\x80\xCC\x14\x02\x00\x86\x05\x00\x00\x6E\x0F\x00\x80\x88\x05"
-     "\x00\x00"
-     "\x74\x0F\x00\x80\xCC\xC8\x35\x00\xCD\xCC\x35\x00\xC6\x0E\x00\x80\xC0\x0E"
-     "\x00\x80"
-     "\xCC\x4C\x01\x00\xCD\x7C\x3C\x00\xCC\x54\x02\x00\xA3\xF8\x03\x00\x7A\x0F"
-     "\x00\x80"
-     "\xCD\x24\x01\x00\xC3\x0E\x00\x80\xA7\xFC\x03\x00\xCD\x78\x01\x00\xCC\x2C"
-     "\x01\x00"
-     "\x65\x0F\x00\x80\xAB\x1C\x02\x00\xCD\x1C\x01\x00\x89\x0F\x00\x80\x80\x5D"
-     "\x00\x00"
-     "\xCC\x54\x02\x00\x8F\x0F\x00\x80\xAD\x6C\x02\x00\x88\x79\x00\x00\xB3\xEC"
-     "\x02\x00"
-     "\xB0\x04\x02\x00\x86\x6D\x00\x00\xB6\xA0\x02\x00\xB7\x84\x02\x00\xCD\x3C"
-     "\x01\x00"
-     "\xB5\xC4\x02\x00\xBA\x78\x05\x00\xCD\x24\x01\x00\xB8\x50\x05\x00\xCF\x0E"
-     "\x00\x80"
-     "\xBE\x74\x05\x00\x88\x59\x00\x00\xAB\x0E\x00\x80\x95\x0F\x00\x80\xD5\x0E"
-     "\x00\x80"
-     "\x9B\x0F\x00\x80\xCD\xAC\x36\x00\xCC\x60\x3F\x00\x86\x6C\x3E\x00\x88\x85"
-     "\x00\x00"
-     "\xCD\x18\x3A\x00\xCC\x14\x3A\x00\xCC\x0E\x00\x80\xCD\x28\x01\x00\x81\x09"
-     "\x00\x00"
-     "\x86\xA5\x00\x00\xCC\x74\x02\x00\x80\x41\x03\x00\xCC\xFC\x03\x00\xCD\x8C"
-     "\x02\x00"
-     "\xA1\x0F\x00\x80\x93\x1C\x00\x00\x90\x2C\x00\x00\xCC\x94\x03\x00\x96\xE0"
-     "\x03\x00"
-     "\x97\xD0\x03\x00\xCC\xE4\x03\x00\x95\x14\x00\x00\x9A\xD4\x02\x00\x88\x25"
-     "\x00\x00"
-     "\x98\x60\x03\x00\xB6\x0F\x00\x80\x9E\x70\x03\x00\xD2\x0E\x00\x80\xCC\xF8"
-     "\x3D\x00"
-     "\xCD\xEC\x3D\x00\xCD\xAC\x03\x00\xCC\xA8\x03\x00\xCC\x78\x02\x00\xBC\x0F"
-     "\x00\x80"
-     "\xCC\x38\x01\x00\xCD\x24\x01\x00\xC2\x0F\x00\x80\x88\x51\x00\x00\xCC\xD8"
-     "\x3F\x00"
-     "\xCD\xDC\x3F\x00\xCC\x70\x3E\x00\xCD\x74\x3E\x00\xCC\xA4\x3F\x00\xCD\x9C"
-     "\x3F\x00"
-     "\xCC\x88\x3F\x00\xCD\xFC\x3F\x00\x86\xB5\x01\x00\x80\x29\x01\x00\x84\xA5"
-     "\x01\x00"
-     "\xC8\x0F\x00\x80\x88\x09\x00\x00\x84\x29\x01\x00\xCE\x0F\x00\x80\x86\xDD"
-     "\x01\x00"
-     "\xCC\xE4\x02\x00\x88\xE1\x01\x00\xCD\x18\x01\x00\xCC\x0C\x01\x00\xB0\x0F"
-     "\x00\x80"
-     "\x8B\x6D\x00\x00\x88\x75\x00\x00\xCC\xE4\x02\x00\x86\x05\x00\x00\x71\x0F"
-     "\x00\x80"
-     "\x77\x0F\x00\x80\x80\x2D\x00\x00\x81\x05\x00\x00\x38\x0F\x00\x80\xCC\x0C"
-     "\x02\x00"
-     "\xCD\x08\x02\x00\x32\x0F\x00\x80\x88\xD5\x00\x00\xCC\x00\x04\x00\xCD\x20"
-     "\x02\x00"
-     "\xCC\x38\x01\x00\xCD\x28\x01\x00\x88\x31\x00\x00\x7D\x0F\x00\x80\xCC\xD8"
-     "\x05\x00"
-     "\x35\x0F\x00\x80\xCD\x1C\x3E\x00\xCD\x74\x3E\x00\x68\x0F\x00\x80\x8C\x0F"
-     "\x00\x80"
-     "\xCC\x28\x01\x00\xCD\xCC\x00\x00\xCC\x34\x02\x00\x92\x0F\x00\x80\xCD\xE4"
-     "\x3F\x00"
-     "\x4A\x0F\x00\x80\x98\x0F\x00\x80\x88\x11\x00\x00\xCD\xFC\x3F\x00\xF3\x0E"
-     "\x00\x80"
-     "\x86\xD1\x07\x00\x80\xE9\x07\x00\xCC\x0C\x02\x00\xCD\x2C\x02\x00\x9E\x0F"
-     "\x00\x80"
-     "\x84\xE5\x07\x00\x81\x29\x00\x00\x86\xE1\x07\x00\x80\xCD\x07\x00\x88\x35"
-     "\x00\x00"
-     "\xCC\x34\x02\x00\xCD\x30\x02\x00\x47\x0F\x00\x80\xCC\x54\x02\x00\x88\xED"
-     "\x07\x00"
-     "\xA4\x0F\x00\x80\x88\x7D\x00\x00\xCC\xD8\x02\x00\xB9\x0F\x00\x80\x4D\x0F"
-     "\x00\x80"
-     "\xBF\x0F\x00\x80\xC5\x0F\x00\x80\xB5\xC4\x03\x00\xB4\xC8\x03\x00\xCC\x24"
-     "\x02\x00"
-     "\x88\x0D\x00\x00\xCB\x0F\x00\x80\xCC\x3C\x02\x00\x88\x05\x00\x00\xD1\x0F"
-     "\x00\x80"
-     "\xCC\x24\x02\x00\x88\xF5\x00\x00\x83\xF4\x03\x00\x86\xC9\x00\x00\xB3\x0F"
-     "\x00\x80"
-     "\x84\xC5\x00\x00\x87\xC0\x03\x00\xCC\x2C\x02\x00\x88\x05\x00\x00\x41\x0F"
-     "\x00\x80"
-     "\x8B\xFC\x03\x00\xCC\x00\x38\x00\x53\x0F\x00\x80\xCC\x24\x02\x00\x8B\xE5"
-     "\x00\x00"
-     "\x8F\x75\x00\x00\x8D\xF0\x03\x00\x88\xED\x00\x00\x56\x0F\x00\x80\xCC\x2C"
-     "\x02\x00"
-     "\x88\x05\x00\x00\x83\x0F\x00\x80\xCC\x28\x02\x00\x96\x70\x03\x00\x88\x05"
-     "\x00\x00"
-     "\x86\x0F\x00\x80\xCC\x24\x02\x00\x88\x0D\x00\x00\xAA\x0F\x00\x80\xCC\x3C"
-     "\x02\x00"
-     "\x88\x05\x00\x00\xAD\x0F\x00\x80\x87\x05\x00\x00\xD4\x0F\x00\x80\x98\xF4"
-     "\x02\x00"
-     "\x99\xE4\x02\x00\xD9\x64\x02\x00\xA7\x60\x03\x00\xC9\x10\x00\x80\xCC\x10"
-     "\x00\x80"
-     "\xD2\x10\x00\x80\xD9\x64\x02\x00\x93\x31\x00\x00\x8A\x00\x03\x00\x94\x19"
-     "\x00\x00"
-     "\x88\x54\x03\x00\xCF\x10\x00\x80\x94\x71\x00\x00\x95\x21\x00\x00\x94\x29"
-     "\x00\x00"
-     "\xD9\x24\x02\x00\xD5\x10\x00\x80\x47\x11\x00\x80\x6B\x17\x00\x80\x3A\x18"
-     "\x00\x80"
-     "\xC9\x17\x00\x80\xD9\x68\x02\x00\x2E\x18\x00\x80\x94\x3D\x00\x00\xD9\x58"
-     "\x02\x00"
-     "\x51\x18\x00\x80\x94\x05\x00\x00\xD9\x54\x02\x00\x94\x0D\x00\x00\x28\x18"
-     "\x00\x80"
-     "\xE0\x68\x02\x00\xB0\xAC\x03\x00\xA4\xEC\x02\x00\x88\x4D\x06\x00\xB3\xC4"
-     "\x03\x00"
-     "\x7E\x18\x00\x80\x90\x18\x00\x80\xBC\x09\x00\x00\x94\x18\x00\x80\x54\x18"
-     "\x00\x80"
-     "\x81\x98\x03\x00\x80\x09\x06\x00\xE0\x84\x02\x00\x86\x0D\x06\x00\xBC\x21"
-     "\x00\x00"
-     "\xA4\xFC\x02\x00\xA4\xBC\x02\x00\xA8\x10\x02\x00\xA9\x18\x02\x00\xE0\xE4"
-     "\x02\x00"
-     "\x5B\x18\x00\x80\xB7\x81\x00\x00\xB6\x99\x00\x00\xB5\x91\x00\x00\x62\x18"
-     "\x00\x80"
-     "\xA8\x1C\x03\x00\x69\x18\x00\x80\x92\xE8\x03\x00\xB8\xB9\x00\x00\xBC\x6D"
-     "\x00\x00"
-     "\x95\xE0\x03\x00\x70\x18\x00\x80\x77\x18\x00\x80\xA4\x40\x01\x00\xA5\xE4"
-     "\x02\x00"
-     "\x82\x18\x00\x80\xA7\xB0\x02\x00\x89\x18\x00\x80\xBE\x55\x00\x00\xE0\xB8"
-     "\x02\x00"
-     "\x97\x69\x00\x00\xBE\x29\x01\x00\xAD\x38\x03\x00\xAE\x30\x02\x00\xAF\x18"
-     "\x03\x00"
-     "\xA8\x38\x02\x00\xA9\xB0\x02\x00\xA6\xD4\x02\x00\xA7\x7C\x02\x00\xAD\x18"
-     "\x00\x80"
-     "\xB5\x9C\x03\x00\xE0\x0C\x02\x00\xB7\xE4\x03\x00\x92\xBC\x03\x00\xB1\x68"
-     "\x03\x00"
-     "\xB2\xDC\x03\x00\xB3\xE0\x03\x00\xBC\x64\x0C\x00\xBD\xFC\x0C\x00\xBE\x6C"
-     "\x0D\x00"
-     "\xE0\x68\x02\x00\xB8\xA8\x0C\x00\xB1\x18\x00\x80\xBA\xAC\x0C\x00\xBB\x2C"
-     "\x0C\x00"
-     "\x98\x21\x03\x00\x99\x49\x03\x00\x9A\x55\x03\x00\x9B\x69\x03\x00\x9C\x69"
-     "\x03\x00"
-     "\x9D\x69\x03\x00\x9E\x91\x00\x00\x9F\x8D\x00\x00\x87\x38\x03\x00\x98\x18"
-     "\x00\x80"
-     "\x9F\x18\x00\x80\xA6\x18\x00\x80\xD1\x18\x00\x80\xD8\x18\x00\x80\xAF\x00"
-     "\x02\x00"
-     "\xB5\x18\x00\x80\x9F\x6D\x00\x00\xB3\x75\x00\x00\x9D\x61\x00\x00\x9C\x69"
-     "\x00\x00"
-     "\xB6\x65\x00\x00\xBC\x18\x00\x80\xC3\x18\x00\x80\xCA\x18\x00\x80\x99\x5D"
-     "\x00\x00"
-     "\x97\xB1\x01\x00\x9B\x31\x00\x00\x9A\x39\x00\x00\xAD\xB0\x03\x00\xAC\x2C"
-     "\x02\x00"
-     "\xBE\xA1\x01\x00\x9E\x21\x00\x00\x87\xC8\x02\x00\x97\x75\x00\x00\xF1\x18"
-     "\x00\x80"
-     "\xE0\x8C\x02\x00\xED\x18\x00\x80\xE0\x9C\x02\x00\x96\x09\x00\x00\x97\x3D"
-     "\x00\x00"
-     "\xF5\x18\x00\x80\x86\xE0\x03\x00\x87\xBC\x03\x00\x86\xA0\x03\x00\xAE\x88"
-     "\x02\x00"
-     "\xDF\x18\x00\x80\xAE\x38\x03\x00\xAF\x64\x02\x00\xBE\x3D\x00\x00\xE0\xC8"
-     "\x02\x00"
-     "\xB1\x40\x02\x00\xE6\x18\x00\x80\x96\xD1\x00\x00\xF9\x18\x00\x80\xFD\x18"
-     "\x00\x80"
-     "\x8A\xB9\x00\x00\xB3\x40\x02\x00\xE0\xC8\x02\x00\x9D\x25\x00\x00\x9C\x59"
-     "\x00\x00"
-     "\x05\x19\x00\x80\x95\xF1\x00\x00\xBF\x9C\x03\x00\x01\x19\x00\x80\x09\x19"
-     "\x00\x80"
-     "\x0D\x19\x00\x80\x1D\x19\x00\x80\xE0\xE0\x02\x00\x95\x19\x00\x00\xE0\x94"
-     "\x02\x00"
-     "\x11\x19\x00\x80\x97\x25\x00\x00\xB3\x04\x02\x00\x86\xA8\x03\x00\xB3\x6C"
-     "\x02\x00"
-     "\x82\x01\x0C\x00\x15\x19\x00\x80\x21\x19\x00\x80\x29\x19\x00\x80\x25\x19"
-     "\x00\x80"
-     "\x86\xD4\x03\x00\x87\x9C\x03\x00\x2D\x19\x00\x80\x99\xAC\x03\x00\xB4\xF4"
-     "\x02\x00"
-     "\xB5\x14\x02\x00\xE0\xE0\x02\x00\xBE\x69\x00\x00\xE0\x3C\x0D\x00\x35\x19"
-     "\x00\x80"
-     "\xB4\xE4\x02\x00\x39\x19\x00\x80\x95\x69\x00\x00\xBE\x49\x00\x00\x3D\x19"
-     "\x00\x80"
-     "\x96\x4D\x00\x00\xB7\x2C\x02\x00\x8A\xCD\x0F\x00\x31\x19\x00\x80\xB7\x78"
-     "\x0D\x00"
-     "\x8F\x19\x00\x00\xB3\xED\x0F\x00\xB3\x05\x00\x00\x52\x19\x00\x80\x9C\x09"
-     "\x00\x00"
-     "\x9D\x15\x00\x00\x44\x19\x00\x80\xE0\x48\x0D\x00\x4B\x19\x00\x80\x8C\x0D"
-     "\x00\x00"
-     "\x59\x19\x00\x80\xA9\x55\x00\x00\xB6\x35\x00\x00\xB7\xC5\x00\x00\x64\x19"
-     "\x00\x80"
-     "\xB5\x3D\x00\x00\x9F\xBD\x0F\x00\xB3\x39\x00\x00\x6B\x19\x00\x80\x72\x19"
-     "\x00\x80"
-     "\x79\x19\x00\x80\x80\x19\x00\x80\x5D\x19\x00\x80\x87\x19\x00\x80\x8E\x19"
-     "\x00\x80"
-     "\xD0\x19\x00\x80\xB8\xFD\x00\x00\xB9\xF5\x00\x00\xC6\x19\x00\x80\xE1\x19"
-     "\x00\x80"
-     "\xD7\x19\x00\x80\x95\x19\x00\x80\xA3\x19\x00\x80\x83\xDD\x00\x00\xAA\x19"
-     "\x00\x80"
-     "\xB1\x19\x00\x80\xB8\x19\x00\x80\xBF\x19\x00\x80\x9C\x19\x00\x80\x8D\xF5"
-     "\x00\x00"
-     "\xA5\x80\x03\x00\xE1\x0C\x02\x00\x80\xC0\x02\x00\xE8\x19\x00\x80\xAE\x7D"
-     "\x00\x00"
-     "\x97\xA5\x00\x00\xBC\x05\x00\x00\xEC\x19\x00\x80\x92\x8D\x00\x00\xE1\x18"
-     "\x02\x00"
-     "\x85\xDC\x00\x00\xE3\x14\x01\x00\x64\x3B\x00\x80\x68\x3B\x00\x80\x9C\x8D"
-     "\x00\x00"
-     "\xEF\x14\x01\x00\x86\x00\x04\x00\x87\x88\x03\x00\x6C\x3B\x00\x80\x70\x3B"
-     "\x00\x80"
-     "\x81\x8D\x01\x00\x74\x3B\x00\x80\x78\x3B\x00\x80\x7C\x3B\x00\x80\x80\x3B"
-     "\x00\x80"
-     "\x84\x3B\x00\x80\x88\x3B\x00\x80\x8C\x3B\x00\x80\x80\x04\x03\x00\x90\x3B"
-     "\x00\x80"
-     "\x94\x3B\x00\x80\x83\x14\x03\x00\x84\x30\x01\x00\x98\x3B\x00\x80\x9C\x3B"
-     "\x00\x80"
-     "\xA0\x3B\x00\x80\xA4\x3B\x00\x80\xA8\x3B\x00\x80\x93\xD9\x01\x00\xAC\x3B"
-     "\x00\x80"
-     "\xB0\x3B\x00\x80\xB4\x3B\x00\x80\xB8\x3B\x00\x80\xBC\x3B\x00\x80\xC0\x3B"
-     "\x00\x80"
-     "\xC4\x3B\x00\x80\xC8\x3B\x00\x80\xCC\x3B\x00\x80\x9D\xE9\x01\x00\x82\xF5"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x80\x15\x00\x00\xD0\x3B\x00\x80\x84\x48\x00\x00\xBE\x48"
-     "\x01\x00"
-     "\xA2\x11\x01\x00\xAC\x84\x18\x02\xAD\x94\x1A\x02\xA7\x39\x01\x00\xD8\x3B"
-     "\x00\x80"
-     "\x86\x20\x07\x00\x87\xEC\x00\x00\xDC\x3B\x00\x80\xE0\x3B\x00\x80\xE4\x3B"
-     "\x00\x80"
-     "\xAC\x11\x01\x00\xE8\x3B\x00\x80\xEC\x3B\x00\x80\xF0\x3B\x00\x80\xF4\x3B"
-     "\x00\x80"
-     "\xBA\xD8\x18\x02\x84\xCC\x00\x00\xF8\x3B\x00\x80\xBD\xB8\x18\x02\xBE\xFC"
-     "\x18\x02"
-     "\xBF\xDC\x13\x02\xB9\x49\x01\x00\xFC\x3B\x00\x80\x00\x3C\x00\x80\x04\x3C"
-     "\x00\x80"
-     "\xB4\xC0\x1A\x02\xB5\x30\x1F\x02\xB6\x00\x10\x02\x08\x3C\x00\x80\x80\x50"
-     "\x01\x00"
-     "\x81\x00\x6A\x00\x82\x68\x77\x00\x83\xB8\x52\x00\x84\x60\xBF\x00\x85\x94"
-     "\x9B\x00"
-     "\x86\x8C\xE4\x00\x87\x1C\xCE\x00\x88\x24\x2A\x01\x89\xC4\x37\x01\x8A\x58"
-     "\x13\x01"
-     "\x8B\xA8\x7F\x01\x8C\xB0\x58\x01\x8D\x10\xA2\x01\x8E\x3C\x8E\x01\x8F\xAC"
-     "\xEA\x01"
-     "\x90\xC0\xF6\x01\x91\x40\xD3\x01\x92\xC8\x3C\x02\xA3\xB1\x02\x00\x84\x9C"
-     "\x05\x00"
-     "\xA5\x5D\x03\x00\xA6\x55\x03\x00\x0C\x3C\x00\x80\x10\x3C\x00\x80\x14\x3C"
-     "\x00\x80"
-     "\xAA\x71\x03\x00\xAB\x71\x03\x00\xAC\x55\x03\x00\xAD\x5D\x03\x00\xAE\x4D"
-     "\x03\x00"
-     "\xAF\x45\x03\x00\xA1\x61\x02\x00\x18\x3C\x00\x80\xA3\x8D\x06\x00\xA2\x8D"
-     "\x06\x00"
-     "\xA5\x01\x18\x00\xA4\x21\x06\x00\xA7\x1D\x1A\x00\xA6\xBD\x1A\x00\xA9\x81"
-     "\x1E\x00"
-     "\xA8\x01\x1C\x00\xAB\xC1\x1F\x00\xAA\x9D\x1E\x00\xAD\x91\x12\x00\xAC\xB5"
-     "\x12\x00"
-     "\xAF\xB9\x16\x00\xAE\xD1\x13\x00\xB1\xDD\x17\x00\xB0\xC1\x16\x00\xB3\xE9"
-     "\x6A\x00"
-     "\xB2\x01\x68\x00\xB5\x01\x6C\x00\xB4\x0D\x6A\x00\xBE\xF4\x03\x00\xBE\x14"
-     "\x05\x00"
-     "\x1C\x3C\x00\x80\x20\x3C\x00\x80\xBE\xA4\x06\x00\x24\x3C\x00\x80\x28\x3C"
-     "\x00\x80"
-     "\x2C\x3C\x00\x80\x30\x3C\x00\x80\x34\x3C\x00\x80\xB3\xFD\x03\x00\xBF\x2C"
-     "\x6C\x00"
-     "\x38\x3C\x00\x80\x3C\x3C\x00\x80\x40\x3C\x00\x80\xB6\x19\x02\x00\xB5\x11"
-     "\x02\x00"
-     "\x48\x3C\x00\x80\xBB\x3D\x02\x00\xBA\x3D\x02\x00\x86\x08\x05\x00\x87\x8C"
-     "\x05\x00"
-     "\xBF\x09\x02\x00\xBE\x01\x02\x00\xBD\x11\x02\x00\xBC\x19\x02\x00\x82\x51"
-     "\x00\x00"
-     "\x4C\x3C\x00\x80\x80\x41\x00\x00\x81\x41\x00\x00\xD7\x00\x00\x00\x50\x3C"
-     "\x00\x80"
-     "\x54\x3C\x00\x80\xEF\x2C\x00\x00\x58\x3C\x00\x80\xE1\x8C\x01\x00\x5C\x3C"
-     "\x00\x80"
-     "\xE3\xF8\x00\x00\xBE\x7C\x19\x00\x60\x3C\x00\x80\x64\x3C\x00\x80\xBE\xB0"
-     "\x18\x00"
-     "\xAB\x65\x02\x00\xAA\x55\x02\x00\xE0\x58\x01\x00\x68\x3C\x00\x80\xAF\xCD"
-     "\x01\x00"
-     "\xAE\x69\x02\x00\xAD\x75\x02\x00\xAC\x75\x02\x00\xA3\x0D\x02\x00\x6C\x3C"
-     "\x00\x80"
-     "\x70\x3C\x00\x80\x19\x19\x00\x80\x74\x3C\x00\x80\xA6\x09\x02\x00\xA5\x01"
-     "\x02\x00"
-     "\x78\x3C\x00\x80\x7C\x3C\x00\x80\x80\x3C\x00\x80\x84\x3C\x00\x80\xB3\x61"
-     "\x02\x00"
-     "\x88\x3C\x00\x80\xB5\x6D\x02\x00\xB6\x65\x02\x00\x85\x64\x6D\x00\x8C\x3C"
-     "\x00\x80"
-     "\x90\x3C\x00\x80\xBA\x39\x02\x00\xBB\x09\x02\x00\xBC\x19\x02\x00\xBD\x19"
-     "\x02\x00"
-     "\xBE\x05\x02\x00\xBF\xA1\x01\x00\xD4\x3B\x00\x80\x44\x3C\x00\x80\x94\x3C"
-     "\x00\x80"
-     "\x98\x3C\x00\x80\x9C\x3C\x00\x80\xA0\x3C\x00\x80\xA4\x3C\x00\x80\xA8\x3C"
-     "\x00\x80"
-     "\xA8\xBD\x05\x00\xA9\x8D\x05\x00\xAA\x85\x05\x00\xAB\x9D\x05\x00\xAC\x89"
-     "\x05\x00"
-     "\xAD\x89\x05\x00\xAE\xF9\x05\x00\xAF\xF9\x05\x00\xB0\x91\x05\x00\xB1\xA1"
-     "\x05\x00"
-     "\xB2\xA1\x05\x00\xB3\xA1\x05\x00\xB4\xAD\x05\x00\xB5\x09\x06\x00\xB6\x39"
-     "\x06\x00"
-     "\xB7\x39\x06\x00\xB8\xE5\x06\x00\xB9\x85\x06\x00\xBA\x8D\x06\x00\xBB\x81"
-     "\x06\x00"
-     "\xBC\x81\x06\x00\xBD\x81\x06\x00\xBE\x81\x06\x00\xBF\x81\x06\x00\xA8\x9D"
-     "\x07\x00"
-     "\xA9\xC1\x07\x00\xAA\xDD\x07\x00\xAB\xD5\x07\x00\xAC\xF1\x07\x00\xAD\xFD"
-     "\x07\x00"
-     "\xAE\xED\x07\x00\xAF\xE5\x07\x00\x80\xBD\x00\x00\x81\x09\x00\x00\x82\x19"
-     "\x00\x00"
-     "\xAC\x3C\x00\x80\xB0\x3C\x00\x80\xB4\x3C\x00\x80\x87\xC8\x00\x00\x86\x9C"
-     "\x03\x00"
-     "\xB8\x11\x07\x00\xB9\x25\x07\x00\xBA\x2D\x07\x00\xBB\xC5\x00\x00\xBC\xDD"
-     "\x00\x00"
-     "\xBD\xC5\x00\x00\xBE\xCD\x00\x00\xBF\xC5\x00\x00\xB0\x6D\x07\x00\xB1\x41"
-     "\x07\x00"
-     "\xB2\x41\x07\x00\xB3\x5D\x07\x00\xB4\x4D\x07\x00\xB5\x35\x07\x00\xB6\x3D"
-     "\x07\x00"
-     "\xB7\x35\x07\x00\xB3\x55\x06\x00\xB8\x3C\x00\x80\xBC\x3C\x00\x80\xC0\x3C"
-     "\x00\x80"
-     "\xC4\x3C\x00\x80\xB6\x7D\x06\x00\xB5\x79\x06\x00\xC8\x3C\x00\x80\xBB\x85"
-     "\x06\x00"
-     "\xBA\xA9\x06\x00\xCC\x3C\x00\x80\xD0\x3C\x00\x80\xBF\xF1\x06\x00\xBE\x89"
-     "\x06\x00"
-     "\xBD\x89\x06\x00\xBC\x95\x06\x00\xD4\x3C\x00\x80\xA3\x11\x06\x00\xD8\x3C"
-     "\x00\x80"
-     "\xDC\x3C\x00\x80\xA6\x39\x06\x00\xE0\x3C\x00\x80\xE4\x3C\x00\x80\xA5\x3D"
-     "\x06\x00"
-     "\xAA\xED\x06\x00\xAB\xC1\x06\x00\xE8\x3C\x00\x80\xEC\x3C\x00\x80\xAE\xCD"
-     "\x06\x00"
-     "\xAF\xB5\x06\x00\xAC\xD1\x06\x00\xAD\xCD\x06\x00\xA8\x85\x00\x00\xA9\x8D"
-     "\x00\x00"
-     "\xAA\x85\x00\x00\xAB\x9D\x00\x00\xAC\x8D\x00\x00\xAD\xBD\x00\x00\xAE\xB5"
-     "\x00\x00"
-     "\xAF\xDD\x00\x00\xF0\x3C\x00\x80\xF4\x3C\x00\x80\xF8\x3C\x00\x80\xFC\x3C"
-     "\x00\x80"
-     "\x00\x3D\x00\x80\x04\x3D\x00\x80\x08\x3D\x00\x80\x0C\x3D\x00\x80\xB8\x69"
-     "\x00\x00"
-     "\xB9\x69\x00\x00\xBA\x79\x00\x00\xBB\x79\x00\x00\xBC\x69\x00\x00\xBD\x69"
-     "\x00\x00"
-     "\xBE\xB9\x03\x00\xBF\xB9\x03\x00\xB0\xA9\x00\x00\xB1\xA9\x00\x00\xB2\xBD"
-     "\x00\x00"
-     "\xB3\xB5\x00\x00\xB4\x99\x00\x00\xB5\x99\x00\x00\xB6\x59\x00\x00\xB7\x59"
-     "\x00\x00"
-     "\xBC\x59\x07\x00\xBD\x59\x07\x00\xBE\x69\x07\x00\xBF\x69\x07\x00\xB8\x55"
-     "\x07\x00"
-     "\xB9\x59\x07\x00\xBA\x49\x07\x00\xBB\x49\x07\x00\xB4\x15\x07\x00\xB5\x1D"
-     "\x07\x00"
-     "\xB6\x15\x07\x00\xB7\x6D\x07\x00\xB0\x29\x07\x00\xB1\x29\x07\x00\xB2\x1D"
-     "\x07\x00"
-     "\xB3\x15\x07\x00\xAC\x49\x07\x00\xAD\x49\x07\x00\xAE\x59\x07\x00\xAF\x59"
-     "\x07\x00"
-     "\xA8\xB5\x06\x00\xA9\xAD\x06\x00\xAA\xA5\x06\x00\xAB\x59\x07\x00\x10\x3D"
-     "\x00\x80"
-     "\x14\x3D\x00\x80\x18\x3D\x00\x80\x1C\x3D\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x20\x3D\x00\x80\x84\xC8\x02\x00\xB3\x89\x03\x00\x87\xD8"
-     "\x03\x00"
-     "\x86\x2C\x1C\x00\xB6\xB9\x03\x00\x28\x3D\x00\x80\x2C\x3D\x00\x80\xB5\x81"
-     "\x03\x00"
-     "\xBA\xE9\x03\x00\xBB\xE9\x03\x00\x30\x3D\x00\x80\x34\x3D\x00\x80\xBE\xD9"
-     "\x03\x00"
-     "\xBF\xD9\x03\x00\xBC\xED\x03\x00\xBD\xE5\x03\x00\xA3\xCD\x03\x00\x38\x3D"
-     "\x00\x80"
-     "\x3C\x3D\x00\x80\x40\x3D\x00\x80\x44\x3D\x00\x80\xA6\xFD\x03\x00\xA5\xC5"
-     "\x03\x00"
-     "\x48\x3D\x00\x80\xAB\xAD\x03\x00\xAA\xAD\x03\x00\x4C\x3D\x00\x80\x50\x3D"
-     "\x00\x80"
-     "\xAF\x9D\x03\x00\xAE\x9D\x03\x00\xAD\xA1\x03\x00\xAC\xA9\x03\x00\x54\x3D"
-     "\x00\x80"
-     "\x58\x3D\x00\x80\xEF\x18\x07\x00\x5C\x3D\x00\x80\x60\x3D\x00\x80\x64\x3D"
-     "\x00\x80"
-     "\x68\x3D\x00\x80\x6C\x3D\x00\x80\x84\x20\x02\x00\x70\x3D\x00\x80\x74\x3D"
-     "\x00\x80"
-     "\x78\x3D\x00\x80\xE1\x94\x01\x00\x7C\x3D\x00\x80\xE3\xA8\x03\x00\x80\x3D"
-     "\x00\x80"
-     "\x84\x3D\x00\x80\x88\x3D\x00\x80\x8C\x3D\x00\x80\x90\x3D\x00\x80\x80\x15"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x82\x15\x00\x00\x98\x3D\x00\x80\x86\x40\x1C\x00\x87\x54"
-     "\x1D\x00"
-     "\x9C\x3D\x00\x80\xA0\x3D\x00\x80\xA4\x3D\x00\x80\xA8\x3D\x00\x80\xAC\x3D"
-     "\x00\x80"
-     "\xB0\x3D\x00\x80\xE1\xB0\x06\x00\xE1\x3C\x04\x00\xE3\xC8\x07\x00\xE3\x50"
-     "\x04\x00"
-     "\xB4\x3D\x00\x80\xB8\x3D\x00\x80\xBC\x3D\x00\x80\xC0\x3D\x00\x80\xC4\x3D"
-     "\x00\x80"
-     "\xC8\x3D\x00\x80\xCC\x3D\x00\x80\xD0\x3D\x00\x80\xD4\x3D\x00\x80\xD8\x3D"
-     "\x00\x80"
-     "\xEF\x9C\x07\x00\xEF\x9C\x07\x00\xDC\x3D\x00\x80\xE0\x3D\x00\x80\x84\xC0"
-     "\x1D\x00"
-     "\xB3\x65\x01\x00\xE4\x3D\x00\x80\xB5\x11\x01\x00\xB6\x11\x01\x00\xE8\x3D"
-     "\x00\x80"
-     "\xEC\x3D\x00\x80\xF0\x3D\x00\x80\xBA\x35\x01\x00\xBB\x05\x01\x00\xBC\x1D"
-     "\x01\x00"
-     "\xBD\xF5\x00\x00\xBE\xFD\x00\x00\xBF\xF1\x00\x00\xA8\x95\x1E\x00\xA9\x9D"
-     "\x1E\x00"
-     "\xAA\xA1\x1E\x00\xAB\xA1\x1E\x00\xAC\xD5\x1E\x00\xAD\xDD\x1E\x00\xAE\xD5"
-     "\x1E\x00"
-     "\xAF\xCD\x1E\x00\x94\x3D\x00\x80\xF4\x3D\x00\x80\xF8\x3D\x00\x80\xFC\x3D"
-     "\x00\x80"
-     "\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x00\x3E\x00\x80\xB8\x59"
-     "\x1F\x00"
-     "\xB9\x59\x1F\x00\xBA\x6D\x1F\x00\xBB\x65\x1F\x00\xBC\x7D\x1F\x00\xBD\x61"
-     "\x1F\x00"
-     "\xBE\x65\x1F\x00\xBF\x1D\x1F\x00\xB0\xBD\x1E\x00\xB1\x8D\x1E\x00\xB2\x85"
-     "\x1E\x00"
-     "\xB3\x6D\x1F\x00\xB4\x75\x1F\x00\xB5\x79\x1F\x00\xB6\x69\x1F\x00\xB7\x69"
-     "\x1F\x00"
-     "\xA3\x29\x1E\x00\x04\x3E\x00\x80\x87\xC4\x03\x00\x86\x2C\x00\x00\x08\x3E"
-     "\x00\x80"
-     "\xA6\x5D\x1E\x00\xA5\x5D\x1E\x00\x0C\x3E\x00\x80\xAB\x49\x1E\x00\xAA\x79"
-     "\x1E\x00"
-     "\x10\x3E\x00\x80\x14\x3E\x00\x80\xAF\xBD\x1F\x00\xAE\xB1\x1F\x00\xAD\xB9"
-     "\x1F\x00"
-     "\xAC\x51\x1E\x00\x18\x3E\x00\x80\xB3\xA9\x1F\x00\x1C\x3E\x00\x80\x20\x3E"
-     "\x00\x80"
-     "\xB6\xA9\x1F\x00\x24\x3E\x00\x80\x28\x3E\x00\x80\xB5\xA9\x1F\x00\xBA\x75"
-     "\x1F\x00"
-     "\xBB\x7D\x1F\x00\x2C\x3E\x00\x80\x30\x3E\x00\x80\xBE\x55\x1F\x00\xBF\x5D"
-     "\x1F\x00"
-     "\xBC\x65\x1F\x00\xBD\x5D\x1F\x00\x34\x3E\x00\x80\x38\x3E\x00\x80\x3C\x3E"
-     "\x00\x80"
-     "\x40\x3E\x00\x80\x44\x3E\x00\x80\xE1\x80\x1D\x00\x48\x3E\x00\x80\xE3\x2C"
-     "\x1C\x00"
-     "\x4C\x3E\x00\x80\x50\x3E\x00\x80\x54\x3E\x00\x80\xEF\xA4\x18\x00\x58\x3E"
-     "\x00\x80"
-     "\x5C\x3E\x00\x80\x60\x3E\x00\x80\x64\x3E\x00\x80\xA3\x2D\x1E\x00\x82\x2D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x80\x1D\x00\x00\x68\x3E\x00\x80\xA6\x2D\x1E\x00\xA5\x2D"
-     "\x1E\x00"
-     "\x6C\x3E\x00\x80\xAB\xF9\x1E\x00\xAA\xF1\x1E\x00\x70\x3E\x00\x80\x84\x20"
-     "\x01\x00"
-     "\xAF\xD9\x1E\x00\xAE\xD1\x1E\x00\xAD\xD9\x1E\x00\xAC\xE1\x1E\x00\xA8\xDD"
-     "\x1E\x00"
-     "\xA9\xED\x1E\x00\xAA\xE5\x1E\x00\xAB\xFD\x1E\x00\xAC\xE5\x1E\x00\xAD\xED"
-     "\x1E\x00"
-     "\xAE\xE5\x1E\x00\xAF\x0D\x1E\x00\xBE\xE8\x01\x00\x78\x3E\x00\x80\x86\x08"
-     "\x07\x00"
-     "\x87\xB8\x00\x00\x7C\x3E\x00\x80\x80\x3E\x00\x80\x84\x3E\x00\x80\x88\x3E"
-     "\x00\x80"
-     "\xB8\x75\x1E\x00\xB9\x05\x1E\x00\xBA\x0D\x1E\x00\xBB\x1D\x1E\x00\xBC\x0D"
-     "\x1E\x00"
-     "\xBD\x35\x1E\x00\xBE\x3D\x1E\x00\xBF\xF9\x01\x00\xB0\x75\x1E\x00\xB1\x7D"
-     "\x1E\x00"
-     "\xB2\x75\x1E\x00\xB3\x4D\x1E\x00\xB4\x55\x1E\x00\xB5\x5D\x1E\x00\xB6\x55"
-     "\x1E\x00"
-     "\xB7\x4D\x1E\x00\xA8\x91\x02\x00\xA9\x99\x02\x00\xAA\xA1\x02\x00\xAB\xA1"
-     "\x02\x00"
-     "\xAC\xD1\x02\x00\xAD\xDD\x02\x00\xAE\xD5\x02\x00\xAF\xC9\x02\x00\x8C\x3E"
-     "\x00\x80"
-     "\x90\x3E\x00\x80\x94\x3E\x00\x80\x98\x3E\x00\x80\x8C\x00\x00\x00\x9C\x3E"
-     "\x00\x80"
-     "\xA0\x3E\x00\x80\xA4\x3E\x00\x80\xB8\x59\x03\x00\xB9\x65\x03\x00\xBA\x6D"
-     "\x03\x00"
-     "\xBB\x65\x03\x00\xBC\x61\x03\x00\xBD\x61\x03\x00\xBE\x61\x03\x00\xBF\x61"
-     "\x03\x00"
-     "\xB0\xB9\x02\x00\xB1\x8D\x02\x00\xB2\x85\x02\x00\xB3\x6D\x03\x00\xB4\x75"
-     "\x03\x00"
-     "\xB5\x7D\x03\x00\xB6\x75\x03\x00\xB7\x6D\x03\x00\xB3\x29\x02\x00\xA8\x3E"
-     "\x00\x80"
-     "\xAC\x3E\x00\x80\xB0\x3E\x00\x80\x24\x3D\x00\x80\xB6\x5D\x02\x00\xB5\x59"
-     "\x02\x00"
-     "\xB4\x3E\x00\x80\xBB\x49\x02\x00\xBA\x7D\x02\x00\xB8\x3E\x00\x80\xBC\x3E"
-     "\x00\x80"
-     "\xBF\xB9\x03\x00\xBE\xB1\x03\x00\xBD\xB9\x03\x00\xBC\x51\x02\x00\xC0\x3E"
-     "\x00\x80"
-     "\xA3\x6D\x02\x00\xC4\x3E\x00\x80\xC8\x3E\x00\x80\xA6\x19\x02\x00\xCC\x3E"
-     "\x00\x80"
-     "\xD0\x3E\x00\x80\xA5\x1D\x02\x00\xAA\x39\x02\x00\xAB\x0D\x02\x00\xD4\x3E"
-     "\x00\x80"
-     "\xD8\x3E\x00\x80\xAE\xF5\x03\x00\xAF\xFD\x03\x00\xAC\x15\x02\x00\xAD\xFD"
-     "\x03\x00"
-     "\xE1\x94\x01\x00\xDC\x3E\x00\x80\xE3\x74\x00\x00\xE0\x3E\x00\x80\x81\x15"
-     "\x00\x00"
-     "\x80\xF9\x01\x00\xBE\x54\x02\x00\x82\x11\x00\x00\xBE\x2C\x04\x00\xE8\x3E"
-     "\x00\x80"
-     "\xEC\x3E\x00\x80\xF0\x3E\x00\x80\xF4\x3E\x00\x80\xF8\x3E\x00\x80\xEF\x60"
-     "\x00\x00"
-     "\x84\xB0\x02\x00\xB3\xF5\x02\x00\xFC\x3E\x00\x80\x87\x64\x05\x00\x86\x6C"
-     "\x05\x00"
-     "\x00\x3F\x00\x80\xB6\x75\x01\x00\xB5\xE5\x02\x00\x04\x3F\x00\x80\xBB\x55"
-     "\x01\x00"
-     "\xBA\x49\x01\x00\x08\x3F\x00\x80\x0C\x3F\x00\x80\xBF\xF9\x01\x00\xBE\xF1"
-     "\x01\x00"
-     "\xBD\x45\x01\x00\xBC\x45\x01\x00\x10\x3F\x00\x80\x14\x3F\x00\x80\x18\x3F"
-     "\x00\x80"
-     "\x1C\x3F\x00\x80\x20\x3F\x00\x80\x24\x3F\x00\x80\x28\x3F\x00\x80\xEF\x18"
-     "\x1A\x00"
-     "\x2C\x3F\x00\x80\xE1\xDC\x1E\x00\x30\x3F\x00\x80\xE3\x74\x1E\x00\x84\x3C"
-     "\x04\x00"
-     "\x34\x3F\x00\x80\x38\x3F\x00\x80\x3C\x3F\x00\x80\xA3\x75\x02\x00\x40\x3F"
-     "\x00\x80"
-     "\x44\x3F\x00\x80\x84\xCC\x04\x00\x48\x3F\x00\x80\xA6\xF5\x01\x00\xA5\x65"
-     "\x02\x00"
-     "\x4C\x3F\x00\x80\xAB\xD5\x01\x00\xAA\xC9\x01\x00\x50\x3F\x00\x80\x58\x3F"
-     "\x00\x80"
-     "\xAF\x79\x01\x00\xAE\x71\x01\x00\xAD\xC5\x01\x00\xAC\xC5\x01\x00\xA8\xA5"
-     "\x05\x00"
-     "\xA9\xB5\x05\x00\xAA\xBD\x05\x00\xAB\xB5\x05\x00\xAC\xAD\x05\x00\xAD\x35"
-     "\x06\x00"
-     "\xAE\x3D\x06\x00\xAF\x35\x06\x00\xE4\x3E\x00\x80\x82\xCD\x07\x00\x81\xF5"
-     "\x07\x00"
-     "\x80\xFD\x07\x00\x54\x3F\x00\x80\x5C\x3F\x00\x80\x87\xF8\x03\x00\x86\x9C"
-     "\x00\x00"
-     "\xB8\xE9\x06\x00\xB9\xE9\x06\x00\xBA\x89\x06\x00\xBB\x85\x06\x00\xBC\x9D"
-     "\x06\x00"
-     "\xBD\x81\x06\x00\xBE\x81\x06\x00\xBF\xB5\x06\x00\xB0\x4D\x06\x00\xB1\x51"
-     "\x06\x00"
-     "\xB2\x51\x06\x00\xB3\x51\x06\x00\xB4\xF5\x06\x00\xB5\xFD\x06\x00\xB6\xF5"
-     "\x06\x00"
-     "\xB7\xE1\x06\x00\xA8\xA5\x07\x00\xA9\xAD\x07\x00\xAA\xA5\x07\x00\xAB\xBD"
-     "\x07\x00"
-     "\xAC\xA5\x07\x00\xAD\xD1\x07\x00\xAE\xC9\x07\x00\xAF\xC1\x07\x00\xBE\xA4"
-     "\x00\x00"
-     "\x60\x3F\x00\x80\x64\x3F\x00\x80\x68\x3F\x00\x80\x6C\x3F\x00\x80\x70\x3F"
-     "\x00\x80"
-     "\x74\x3F\x00\x80\x78\x3F\x00\x80\xB8\x59\x07\x00\xB9\x65\x07\x00\xBA\x61"
-     "\x07\x00"
-     "\xBB\x79\x07\x00\xBC\x69\x07\x00\xBD\x69\x07\x00\xBE\x1D\x07\x00\xBF\x15"
-     "\x07\x00"
-     "\xB0\x89\x07\x00\xB1\x89\x07\x00\xB2\x69\x07\x00\xB3\x69\x07\x00\xB4\x79"
-     "\x07\x00"
-     "\xB5\x79\x07\x00\xB6\x69\x07\x00\xB7\x69\x07\x00\x7C\x3F\x00\x80\x80\x3F"
-     "\x00\x80"
-     "\x84\x3F\x00\x80\x88\x3F\x00\x80\x8C\x3F\x00\x80\x90\x3F\x00\x80\xEF\xF8"
-     "\x04\x00"
-     "\x94\x3F\x00\x80\xE1\x90\x06\x00\x98\x3F\x00\x80\xE3\x04\x04\x00\x9C\x3F"
-     "\x00\x80"
-     "\xA0\x3F\x00\x80\xA4\x3F\x00\x80\xA8\x3F\x00\x80\xAC\x3F\x00\x80\xB3\x6D"
-     "\x06\x00"
-     "\xB0\x3F\x00\x80\xB4\x3F\x00\x80\xB8\x3F\x00\x80\xBC\x3F\x00\x80\xB6\x11"
-     "\x06\x00"
-     "\xB5\x6D\x06\x00\xC0\x3F\x00\x80\xBB\x0D\x06\x00\xBA\x0D\x06\x00\xC4\x3F"
-     "\x00\x80"
-     "\xC8\x3F\x00\x80\xBF\xFD\x06\x00\xBE\xFD\x06\x00\xBD\xFD\x06\x00\xBC\xFD"
-     "\x06\x00"
-     "\xA3\xAD\x06\x00\xCC\x3F\x00\x80\x86\x28\x00\x00\x87\xA8\x01\x00\xD0\x3F"
-     "\x00\x80"
-     "\xA6\xD1\x06\x00\xA5\xAD\x06\x00\xD4\x3F\x00\x80\xAB\xCD\x06\x00\xAA\xCD"
-     "\x06\x00"
-     "\xD8\x3F\x00\x80\xDC\x3F\x00\x80\xAF\x3D\x06\x00\xAE\x3D\x06\x00\xAD\x3D"
-     "\x06\x00"
-     "\xAC\x3D\x06\x00\x80\x8D\x00\x00\x81\x95\x00\x00\x82\x95\x00\x00\xB3\x69"
-     "\x01\x00"
-     "\xE0\x3F\x00\x80\xB5\x79\x01\x00\xB6\x69\x01\x00\xE4\x3F\x00\x80\xE8\x3F"
-     "\x00\x80"
-     "\xEC\x3F\x00\x80\xBA\x55\x01\x00\xBB\xED\x00\x00\xBC\xF5\x00\x00\xBD\xFD"
-     "\x00\x00"
-     "\xBE\xF5\x00\x00\xBF\xED\x00\x00\xA8\xED\x05\x00\xA9\x31\x02\x00\xAA\xAD"
-     "\x02\x00"
-     "\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xA5\x02\x00\xAE\xAD\x02\x00\xAF\xA5"
-     "\x02\x00"
-     "\x81\xF9\x00\x00\x80\xF5\x00\x00\xF0\x3F\x00\x80\x82\xF1\x00\x00\xF4\x3F"
-     "\x00\x80"
-     "\xF8\x3F\x00\x80\xFC\x3F\x00\x80\x00\x40\x00\x80\xB8\x65\x03\x00\xB9\x69"
-     "\x03\x00"
-     "\xBA\x79\x03\x00\xBB\x79\x03\x00\xBC\x69\x03\x00\xBD\x69\x03\x00\xBE\x1D"
-     "\x03\x00"
-     "\xBF\x15\x03\x00\xB0\xDD\x02\x00\xB1\x65\x03\x00\xB2\x6D\x03\x00\xB3\x65"
-     "\x03\x00"
-     "\xB4\x7D\x03\x00\xB5\x65\x03\x00\xB6\x6D\x03\x00\xB7\x61\x03\x00\x04\x40"
-     "\x00\x80"
-     "\x08\x40\x00\x80\xA3\x2D\x02\x00\x0C\x40\x00\x80\xA5\x3D\x02\x00\x10\x40"
-     "\x00\x80"
-     "\x14\x40\x00\x80\xA6\x2D\x02\x00\xBE\x00\x0C\x00\x74\x3E\x00\x80\xAB\xA9"
-     "\x03\x00"
-     "\xAA\x11\x02\x00\xAD\xB9\x03\x00\xAC\xB1\x03\x00\xAF\xA9\x03\x00\xAE\xB1"
-     "\x03\x00"
-     "\x18\x40\x00\x80\x1C\x40\x00\x80\x20\x40\x00\x80\x24\x40\x00\x80\x28\x40"
-     "\x00\x80"
-     "\xEF\x8C\x02\x00\x2C\x40\x00\x80\x30\x40\x00\x80\x84\xF8\x03\x00\xE3\x14"
-     "\x02\x00"
-     "\x38\x40\x00\x80\xE1\x7C\x01\x00\x86\xD0\x0C\x00\x87\xB4\x0C\x00\x3C\x40"
-     "\x00\x80"
-     "\x40\x40\x00\x80\xA8\x6D\x02\x00\xA9\x81\x02\x00\xAA\x99\x02\x00\xAB\x91"
-     "\x02\x00"
-     "\xAC\xB9\x02\x00\xAD\xB9\x02\x00\xAE\xC9\x02\x00\xAF\xC1\x02\x00\x44\x40"
-     "\x00\x80"
-     "\x48\x40\x00\x80\x4C\x40\x00\x80\x50\x40\x00\x80\x54\x40\x00\x80\x58\x40"
-     "\x00\x80"
-     "\x5C\x40\x00\x80\x60\x40\x00\x80\xB8\xC5\x01\x00\xB9\xC9\x01\x00\xBA\xDD"
-     "\x01\x00"
-     "\xBB\xD5\x01\x00\xBC\x91\x01\x00\xBD\x91\x01\x00\xBE\x91\x01\x00\xBF\x91"
-     "\x01\x00"
-     "\xB0\x81\x02\x00\xB1\x81\x02\x00\xB2\x41\x01\x00\xB3\x5D\x01\x00\xB4\x45"
-     "\x01\x00"
-     "\xB5\x4D\x01\x00\xB6\x45\x01\x00\xB7\xFD\x01\x00\x64\x40\x00\x80\x68\x40"
-     "\x00\x80"
-     "\x6C\x40\x00\x80\xEF\xD4\x01\x00\x84\x10\x0D\x00\x70\x40\x00\x80\x74\x40"
-     "\x00\x80"
-     "\xEF\x44\x06\x00\x78\x40\x00\x80\xE1\xF0\x06\x00\x7C\x40\x00\x80\xE3\x0C"
-     "\x06\x00"
-     "\x80\x40\x00\x80\xE1\x90\x01\x00\x84\x40\x00\x80\xE3\x60\x06\x00\xB6\x6D"
-     "\x02\x00"
-     "\x88\x40\x00\x80\x8C\x40\x00\x80\xB5\x65\x02\x00\x90\x40\x00\x80\xB3\x45"
-     "\x02\x00"
-     "\x94\x40\x00\x80\x98\x40\x00\x80\xBE\xD5\x01\x00\xBF\xDD\x01\x00\xBC\xC1"
-     "\x01\x00"
-     "\xBD\xDD\x01\x00\xBA\x11\x02\x00\xBB\x11\x02\x00\x9C\x40\x00\x80\xA0\x40"
-     "\x00\x80"
-     "\x34\x40\x00\x80\xBE\x88\x03\x00\xAA\xE9\x0D\x00\xAB\xE9\x0D\x00\xAC\x39"
-     "\x0E\x00"
-     "\xAD\x25\x0E\x00\xAE\x2D\x0E\x00\xAF\x25\x0E\x00\xA4\x40\x00\x80\xA8\x40"
-     "\x00\x80"
-     "\xAC\x40\x00\x80\xA3\xBD\x0D\x00\xB0\x40\x00\x80\xA5\x9D\x0D\x00\xA6\x95"
-     "\x0D\x00"
-     "\xB4\x40\x00\x80\x80\x9D\x0F\x00\x81\xBD\x0F\x00\x82\xB5\x0F\x00\xB3\xB1"
-     "\x0E\x00"
-     "\xB8\x40\x00\x80\xB5\x71\x0F\x00\xB6\x71\x0F\x00\xBC\x40\x00\x80\x86\x40"
-     "\x03\x00"
-     "\x87\x34\x03\x00\xBA\x5D\x0F\x00\xBB\x55\x0F\x00\xBC\x4D\x0F\x00\xBD\xF5"
-     "\x0F\x00"
-     "\xBE\xFD\x0F\x00\xBF\xF5\x0F\x00\xA8\x69\x0E\x00\xA9\x7D\x0E\x00\xAA\x75"
-     "\x0E\x00"
-     "\xAB\xAD\x0F\x00\xAC\xB5\x0F\x00\xAD\xBD\x0F\x00\xAE\xB5\x0F\x00\xAF\xAD"
-     "\x0F\x00"
-     "\xC0\x40\x00\x80\xC4\x40\x00\x80\xC8\x40\x00\x80\xCC\x40\x00\x80\xD0\x40"
-     "\x00\x80"
-     "\xD4\x40\x00\x80\xD8\x40\x00\x80\xDC\x40\x00\x80\xB8\x61\x0F\x00\xB9\x6D"
-     "\x0F\x00"
-     "\xBA\x79\x0F\x00\xBB\x75\x0F\x00\xBC\x15\x0F\x00\xBD\x1D\x0F\x00\xBE\x09"
-     "\x0F\x00"
-     "\xBF\x05\x0F\x00\xB0\xD5\x0F\x00\xB1\xDD\x0F\x00\xB2\xD5\x0F\x00\xB3\x6D"
-     "\x0F\x00"
-     "\xB4\x75\x0F\x00\xB5\x7D\x0F\x00\xB6\x6D\x0F\x00\xB7\x65\x0F\x00\xA3\xFD"
-     "\x0F\x00"
-     "\xE0\x40\x00\x80\xE4\x40\x00\x80\xE8\x40\x00\x80\xEC\x40\x00\x80\xA6\x3D"
-     "\x0E\x00"
-     "\xA5\x3D\x0E\x00\xF0\x40\x00\x80\xAB\x19\x0E\x00\xAA\x11\x0E\x00\xF4\x40"
-     "\x00\x80"
-     "\xF8\x40\x00\x80\xAF\xB9\x0E\x00\xAE\xB1\x0E\x00\xAD\xB9\x0E\x00\xAC\x01"
-     "\x0E\x00"
-     "\xFC\x40\x00\x80\x00\x41\x00\x80\x04\x41\x00\x80\x08\x41\x00\x80\x80\x1D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x39\x00\x00\x0C\x41\x00\x80\x10\x41\x00\x80\x84\xA8"
-     "\x01\x00"
-     "\xBE\xA0\x01\x00\x14\x41\x00\x80\x86\xD0\x00\x00\x87\xE4\x01\x00\x18\x41"
-     "\x00\x80"
-     "\x1C\x41\x00\x80\xA8\xA5\x01\x00\xA9\xAD\x01\x00\xAA\xB9\x01\x00\xAB\xB1"
-     "\x01\x00"
-     "\xAC\xC1\x01\x00\xAD\xC9\x01\x00\xAE\x39\x01\x00\xAF\x39\x01\x00\x84\xA0"
-     "\x00\x00"
-     "\x20\x41\x00\x80\x24\x41\x00\x80\x28\x41\x00\x80\x2C\x41\x00\x80\x30\x41"
-     "\x00\x80"
-     "\x34\x41\x00\x80\x38\x41\x00\x80\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5"
-     "\x00\x00"
-     "\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x9D\x00\x00\xBF\x95"
-     "\x00\x00"
-     "\xB0\x49\x01\x00\xB1\x49\x01\x00\xB2\x59\x01\x00\xB3\x59\x01\x00\xB4\x49"
-     "\x01\x00"
-     "\xB5\x49\x01\x00\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xB3\x31\x02\x00\x3C\x41"
-     "\x00\x80"
-     "\x40\x41\x00\x80\x44\x41\x00\x80\x48\x41\x00\x80\xB6\x49\x02\x00\xB5\x49"
-     "\x02\x00"
-     "\x4C\x41\x00\x80\xBB\x8D\x02\x00\xBA\x8D\x02\x00\x50\x41\x00\x80\x54\x41"
-     "\x00\x80"
-     "\xBF\x8D\x02\x00\xBE\x8D\x02\x00\xBD\x8D\x02\x00\xBC\x8D\x02\x00\x85\x44"
-     "\x7D\x00"
-     "\xA3\x75\x02\x00\x58\x41\x00\x80\x5C\x41\x00\x80\xA6\x0D\x02\x00\x60\x41"
-     "\x00\x80"
-     "\x64\x41\x00\x80\xA5\x0D\x02\x00\xAA\xC9\x02\x00\xAB\xC9\x02\x00\x68\x41"
-     "\x00\x80"
-     "\x6C\x41\x00\x80\xAE\xC9\x02\x00\xAF\xC9\x02\x00\xAC\xC9\x02\x00\xAD\xC9"
-     "\x02\x00"
-     "\xBE\x00\x04\x00\x70\x41\x00\x80\x74\x41\x00\x80\x78\x41\x00\x80\x80\x1D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x3D\x00\x00\x7C\x41\x00\x80\xBE\x2C\x05\x00\x80\x41"
-     "\x00\x80"
-     "\x88\x41\x00\x80\x8C\x41\x00\x80\x90\x41\x00\x80\x94\x41\x00\x80\x98\x41"
-     "\x00\x80"
-     "\x9C\x41\x00\x80\x86\xE0\x04\x00\x87\x54\x03\x00\xA0\x41\x00\x80\xA4\x41"
-     "\x00\x80"
-     "\xEF\xA8\x02\x00\xA8\x41\x00\x80\xAC\x41\x00\x80\xB0\x41\x00\x80\xE3\x50"
-     "\x01\x00"
-     "\xB4\x41\x00\x80\xE1\x80\x01\x00\xB8\x41\x00\x80\xBC\x41\x00\x80\xC0\x41"
-     "\x00\x80"
-     "\xC4\x41\x00\x80\x84\x70\x04\x00\xA8\x79\x02\x00\xA9\x79\x02\x00\xAA\xAD"
-     "\x02\x00"
-     "\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xA5\x02\x00\xAE\xAD\x02\x00\xAF\xA5"
-     "\x02\x00"
-     "\x84\xA8\x04\x00\xC8\x41\x00\x80\xCC\x41\x00\x80\xD0\x41\x00\x80\xD4\x41"
-     "\x00\x80"
-     "\xD8\x41\x00\x80\xDC\x41\x00\x80\xE0\x41\x00\x80\xB8\x09\x01\x00\xB9\x09"
-     "\x01\x00"
-     "\xBA\x1D\x01\x00\xBB\x15\x01\x00\xBC\x39\x01\x00\xBD\x39\x01\x00\xBE\xDD"
-     "\x01\x00"
-     "\xBF\xD5\x01\x00\xB0\xDD\x02\x00\xB1\x61\x01\x00\xB2\x61\x01\x00\xB3\x61"
-     "\x01\x00"
-     "\xB4\x61\x01\x00\xB5\x69\x01\x00\xB6\x51\x01\x00\xB7\x51\x01\x00\xE4\x41"
-     "\x00\x80"
-     "\xE1\x6C\x0F\x00\xE8\x41\x00\x80\xE3\x0C\x0E\x00\x80\x19\x00\x00\x81\x35"
-     "\x00\x00"
-     "\x82\x31\x00\x00\xEF\x0C\x00\x00\xEC\x41\x00\x80\xE1\xDC\x01\x00\xF0\x41"
-     "\x00\x80"
-     "\xE3\xD8\x00\x00\xED\x00\x00\x00\xF8\x41\x00\x80\xFC\x41\x00\x80\xEF\x40"
-     "\x01\x00"
-     "\x86\xE0\x04\x00\x87\x78\x05\x00\x00\x42\x00\x80\xB3\x69\x02\x00\x04\x42"
-     "\x00\x80"
-     "\xB5\x79\x02\x00\xB6\x69\x02\x00\x08\x42\x00\x80\x0C\x42\x00\x80\x10\x42"
-     "\x00\x80"
-     "\xBA\x55\x02\x00\xBB\xE9\x01\x00\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\xE9"
-     "\x01\x00"
-     "\xBF\xE1\x01\x00\x84\x41\x00\x80\xF4\x41\x00\x80\x14\x42\x00\x80\xA3\xA9"
-     "\x05\x00"
-     "\x18\x42\x00\x80\xA5\xB9\x05\x00\xA6\xA9\x05\x00\x1C\x42\x00\x80\x20\x42"
-     "\x00\x80"
-     "\x24\x42\x00\x80\xAA\x95\x05\x00\xAB\x29\x06\x00\xAC\x39\x06\x00\xAD\x39"
-     "\x06\x00"
-     "\xAE\x29\x06\x00\xAF\x21\x06\x00\xB3\xF1\x06\x00\x28\x42\x00\x80\x2C\x42"
-     "\x00\x80"
-     "\x30\x42\x00\x80\x34\x42\x00\x80\xB6\xE9\x06\x00\xB5\xE1\x06\x00\x38\x42"
-     "\x00\x80"
-     "\xBB\xB9\x06\x00\xBA\xB1\x06\x00\x3C\x42\x00\x80\x40\x42\x00\x80\xBF\x81"
-     "\x06\x00"
-     "\xBE\x99\x06\x00\xBD\x9D\x06\x00\xBC\xA1\x06\x00\xA8\x29\x06\x00\xA9\x29"
-     "\x06\x00"
-     "\xAA\x39\x06\x00\xAB\x39\x06\x00\xAC\x29\x06\x00\xAD\x29\x06\x00\xAE\x51"
-     "\x06\x00"
-     "\xAF\x51\x06\x00\x44\x42\x00\x80\x48\x42\x00\x80\x4C\x42\x00\x80\x50\x42"
-     "\x00\x80"
-     "\x54\x42\x00\x80\x82\xB9\x01\x00\x81\xB1\x01\x00\x80\xB9\x01\x00\xB8\xE5"
-     "\x07\x00"
-     "\xB9\xED\x07\x00\xBA\xE5\x07\x00\xBB\xFD\x07\x00\xBC\xE5\x07\x00\xBD\xED"
-     "\x07\x00"
-     "\xBE\xE5\x07\x00\xBF\x5D\x07\x00\xB0\x35\x06\x00\xB1\x3D\x06\x00\xB2\x35"
-     "\x06\x00"
-     "\xB3\x09\x06\x00\xB4\x1D\x06\x00\xB5\x05\x06\x00\xB6\x0D\x06\x00\xB7\x05"
-     "\x06\x00"
-     "\xA3\xB1\x07\x00\x58\x42\x00\x80\x5C\x42\x00\x80\x60\x42\x00\x80\xBE\x9C"
-     "\x01\x00"
-     "\xA6\xA9\x07\x00\xA5\xA1\x07\x00\x64\x42\x00\x80\xAB\xF9\x07\x00\xAA\xF1"
-     "\x07\x00"
-     "\x86\x88\x00\x00\x87\x78\x01\x00\xAF\xC1\x07\x00\xAE\xD9\x07\x00\xAD\xDD"
-     "\x07\x00"
-     "\xAC\xE1\x07\x00\x68\x42\x00\x80\xB3\x8D\x06\x00\x6C\x42\x00\x80\x70\x42"
-     "\x00\x80"
-     "\xB6\xB1\x06\x00\x74\x42\x00\x80\x78\x42\x00\x80\xB5\x81\x06\x00\xBA\x71"
-     "\x01\x00"
-     "\xBB\x79\x01\x00\x7C\x42\x00\x80\x80\x42\x00\x80\xBE\x35\x01\x00\xBF\x3D"
-     "\x01\x00"
-     "\xBC\x59\x01\x00\xBD\x59\x01\x00\xA8\x35\x06\x00\xA9\x3D\x06\x00\xAA\x41"
-     "\x06\x00"
-     "\xAB\x5D\x06\x00\xAC\x4D\x06\x00\xAD\x75\x06\x00\xAE\x7D\x06\x00\xAF\x95"
-     "\x01\x00"
-     "\x84\xAC\x01\x00\x84\x42\x00\x80\x88\x42\x00\x80\x8C\x42\x00\x80\x90\x42"
-     "\x00\x80"
-     "\x94\x42\x00\x80\x98\x42\x00\x80\x9C\x42\x00\x80\xB8\x9D\x01\x00\xB9\xAD"
-     "\x01\x00"
-     "\xBA\xA5\x01\x00\xBB\x6D\x01\x00\xBC\x75\x01\x00\xBD\x7D\x01\x00\xBE\x75"
-     "\x01\x00"
-     "\xBF\x6D\x01\x00\xB0\xF1\x01\x00\xB1\xF9\x01\x00\xB2\xC9\x01\x00\xB3\xC9"
-     "\x01\x00"
-     "\xB4\xB5\x01\x00\xB5\xBD\x01\x00\xB6\xB5\x01\x00\xB7\xAD\x01\x00\xA3\xC9"
-     "\x05\x00"
-     "\xA0\x42\x00\x80\xA4\x42\x00\x80\xA8\x42\x00\x80\xAC\x42\x00\x80\xA6\xF5"
-     "\x05\x00"
-     "\xA5\xC5\x05\x00\xB0\x42\x00\x80\xAB\x3D\x02\x00\xAA\x35\x02\x00\xB4\x42"
-     "\x00\x80"
-     "\xB8\x42\x00\x80\xAF\x79\x02\x00\xAE\x71\x02\x00\xAD\x1D\x02\x00\xAC\x1D"
-     "\x02\x00"
-     "\xBC\x42\x00\x80\xC0\x42\x00\x80\xC4\x42\x00\x80\xC8\x42\x00\x80\x81\x09"
-     "\x00\x00"
-     "\x80\x19\x00\x00\xCC\x42\x00\x80\x82\x3D\x00\x00\xD0\x42\x00\x80\xBE\x20"
-     "\x03\x00"
-     "\x86\xC8\x7C\x00\x87\x5C\x03\x00\xD8\x42\x00\x80\xDC\x42\x00\x80\xE0\x42"
-     "\x00\x80"
-     "\xE4\x42\x00\x80\x84\xCC\x7C\x00\xE8\x42\x00\x80\xEC\x42\x00\x80\xBE\xC4"
-     "\x7C\x00"
-     "\xE3\x20\x01\x00\xF0\x42\x00\x80\xE1\xA8\x01\x00\xF4\x42\x00\x80\xEF\x3C"
-     "\x02\x00"
-     "\xF8\x42\x00\x80\xFC\x42\x00\x80\x00\x43\x00\x80\x04\x43\x00\x80\x08\x43"
-     "\x00\x80"
-     "\x0C\x43\x00\x80\xBF\x34\x5D\x00\xB3\x8D\x03\x00\x10\x43\x00\x80\x14\x43"
-     "\x00\x80"
-     "\x18\x43\x00\x80\x1C\x43\x00\x80\xB6\xBD\x03\x00\xB5\x81\x03\x00\x20\x43"
-     "\x00\x80"
-     "\xBB\x4D\x03\x00\xBA\x4D\x03\x00\x24\x43\x00\x80\x28\x43\x00\x80\xBF\x4D"
-     "\x03\x00"
-     "\xBE\x4D\x03\x00\xBD\x4D\x03\x00\xBC\x4D\x03\x00\xA8\x41\x02\x00\xA9\x4D"
-     "\x02\x00"
-     "\xAA\x45\x02\x00\xAB\x59\x02\x00\xAC\x49\x02\x00\xAD\x7D\x02\x00\xAE\x75"
-     "\x02\x00"
-     "\xAF\xB9\x02\x00\xBE\xA0\x7F\x00\x2C\x43\x00\x80\x30\x43\x00\x80\x34\x43"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\xA5\x00\x00\x82\xAD\x00\x00\x38\x43\x00\x80\xB8\x4D"
-     "\x01\x00"
-     "\xB9\x55\x01\x00\xBA\x55\x01\x00\xBB\x65\x01\x00\xBC\x7D\x01\x00\xBD\x15"
-     "\x01\x00"
-     "\xBE\x1D\x01\x00\xBF\x15\x01\x00\xB0\xC9\x02\x00\xB1\xC9\x02\x00\xB2\xD9"
-     "\x02\x00"
-     "\xB3\xD9\x02\x00\xB4\xC9\x02\x00\xB5\xC9\x02\x00\xB6\x7D\x01\x00\xB7\x75"
-     "\x01\x00"
-     "\xE1\x58\x06\x00\xE1\x5C\x07\x00\xE3\xB8\x06\x00\xE3\x60\x07\x00\x3C\x43"
-     "\x00\x80"
-     "\x40\x43\x00\x80\x86\xB8\x7C\x00\x87\x54\x7D\x00\x48\x43\x00\x80\x4C\x43"
-     "\x00\x80"
-     "\x50\x43\x00\x80\x54\x43\x00\x80\xE9\x00\x00\x00\x58\x43\x00\x80\xEF\x50"
-     "\x00\x00"
-     "\xEF\xE4\x07\x00\x5C\x43\x00\x80\x60\x43\x00\x80\xA3\x85\x02\x00\x64\x43"
-     "\x00\x80"
-     "\xA5\x89\x02\x00\x68\x43\x00\x80\x6C\x43\x00\x80\xA6\xB5\x02\x00\x70\x43"
-     "\x00\x80"
-     "\x74\x43\x00\x80\xAB\x45\x02\x00\xAA\x45\x02\x00\xAD\x45\x02\x00\xAC\x45"
-     "\x02\x00"
-     "\xAF\x45\x02\x00\xAE\x45\x02\x00\xA8\x71\x7E\x00\xA9\x71\x7E\x00\xAA\x71"
-     "\x7E\x00"
-     "\xAB\x71\x7E\x00\xAC\x95\x7E\x00\xAD\x9D\x7E\x00\xAE\x8D\x7E\x00\xAF\x81"
-     "\x7E\x00"
-     "\x44\x43\x00\x80\x78\x43\x00\x80\x7C\x43\x00\x80\x80\x43\x00\x80\x84\x43"
-     "\x00\x80"
-     "\x88\x43\x00\x80\x8C\x43\x00\x80\x90\x43\x00\x80\xB8\x8D\x7E\x00\xB9\x95"
-     "\x7E\x00"
-     "\xBA\x95\x7E\x00\xBB\xA5\x7E\x00\xBC\xBD\x7E\x00\xBD\x71\x7F\x00\xBE\x71"
-     "\x7F\x00"
-     "\xBF\x71\x7F\x00\xB0\xC5\x7E\x00\xB1\xCD\x7E\x00\xB2\xC5\x7E\x00\xB3\xDD"
-     "\x7E\x00"
-     "\xB4\xC5\x7E\x00\xB5\xCD\x7E\x00\xB6\xBD\x7E\x00\xB7\xB5\x7E\x00\xB3\xF9"
-     "\x7E\x00"
-     "\x94\x43\x00\x80\x98\x43\x00\x80\x9C\x43\x00\x80\xA0\x43\x00\x80\xB6\x1D"
-     "\x7E\x00"
-     "\xB5\x1D\x7E\x00\xA4\x43\x00\x80\xBB\x09\x7E\x00\xBA\x01\x7E\x00\xA8\x43"
-     "\x00\x80"
-     "\xAC\x43\x00\x80\xBF\x09\x7E\x00\xBE\x01\x7E\x00\xBD\x09\x7E\x00\xBC\x11"
-     "\x7E\x00"
-     "\x82\x15\x00\x00\xA3\xBD\x7E\x00\x80\x61\x00\x00\x81\x61\x00\x00\xA6\x59"
-     "\x7E\x00"
-     "\xB0\x43\x00\x80\xB4\x43\x00\x80\xA5\x59\x7E\x00\xAA\x45\x7E\x00\xAB\x4D"
-     "\x7E\x00"
-     "\x84\x24\x03\x00\xB8\x43\x00\x80\xAE\x45\x7E\x00\xAF\x4D\x7E\x00\xAC\x55"
-     "\x7E\x00"
-     "\xAD\x4D\x7E\x00\xB3\xA9\x7E\x00\xBC\x43\x00\x80\x86\x48\x01\x00\x87\xB0"
-     "\x02\x00"
-     "\xC0\x43\x00\x80\xB6\xD5\x7E\x00\xB5\xAD\x7E\x00\xC4\x43\x00\x80\xBB\xF9"
-     "\x7E\x00"
-     "\xBA\xF1\x7E\x00\xC8\x43\x00\x80\xCC\x43\x00\x80\xBF\x19\x01\x00\xBE\x11"
-     "\x01\x00"
-     "\xBD\xDD\x7E\x00\xBC\xDD\x7E\x00\xE3\xD0\x7D\x00\xD0\x43\x00\x80\xE1\x38"
-     "\x7D\x00"
-     "\xD4\x43\x00\x80\xD8\x43\x00\x80\xDC\x43\x00\x80\xE0\x43\x00\x80\xE4\x43"
-     "\x00\x80"
-     "\xE8\x43\x00\x80\xEC\x43\x00\x80\xF0\x43\x00\x80\xF4\x43\x00\x80\xEF\x04"
-     "\x7D\x00"
-     "\xF8\x43\x00\x80\xFC\x43\x00\x80\x00\x44\x00\x80\xA8\x21\x7E\x00\xA9\x21"
-     "\x7E\x00"
-     "\xAA\x21\x7E\x00\xAB\x21\x7E\x00\xAC\x25\x7E\x00\xAD\x55\x7E\x00\xAE\x5D"
-     "\x7E\x00"
-     "\xAF\x51\x7E\x00\x04\x44\x00\x80\x08\x44\x00\x80\x0C\x44\x00\x80\x10\x44"
-     "\x00\x80"
-     "\x14\x44\x00\x80\x82\xBD\x03\x00\x81\xBD\x03\x00\x80\xBD\x03\x00\xB8\xF5"
-     "\x01\x00"
-     "\xB9\xFD\x01\x00\xBA\xF5\x01\x00\xBB\x49\x01\x00\xBC\x59\x01\x00\xBD\x59"
-     "\x01\x00"
-     "\xBE\x49\x01\x00\xBF\x41\x01\x00\xB0\x35\x7E\x00\xB1\x3D\x7E\x00\xB2\x09"
-     "\x7E\x00"
-     "\xB3\x09\x7E\x00\xB4\xD5\x01\x00\xB5\xDD\x01\x00\xB6\xD5\x01\x00\xB7\xCD"
-     "\x01\x00"
-     "\x18\x44\x00\x80\x1C\x44\x00\x80\x20\x44\x00\x80\xA3\xA5\x7D\x00\x84\x88"
-     "\x02\x00"
-     "\xA5\xA1\x7D\x00\xA6\xD9\x7D\x00\xBE\x90\x03\x00\x28\x44\x00\x80\x2C\x44"
-     "\x00\x80"
-     "\xAA\xFD\x7D\x00\xAB\xF5\x7D\x00\xAC\xD1\x7D\x00\xAD\xD1\x7D\x00\xAE\x1D"
-     "\x02\x00"
-     "\xAF\x15\x02\x00\xB6\x6D\x03\x00\x30\x44\x00\x80\x84\x44\x03\x00\xB5\x65"
-     "\x03\x00"
-     "\x34\x44\x00\x80\xB3\x45\x03\x00\x87\x28\x03\x00\x86\xFC\x04\x00\xBE\x01"
-     "\x03\x00"
-     "\xBF\x01\x03\x00\xBC\x31\x03\x00\xBD\x31\x03\x00\xBA\x21\x03\x00\xBB\x21"
-     "\x03\x00"
-     "\x38\x44\x00\x80\x3C\x44\x00\x80\xB0\xD1\x03\x00\xB1\xD1\x03\x00\xB2\xD1"
-     "\x03\x00"
-     "\xB3\xD1\x03\x00\xB4\xF1\x03\x00\xB5\xF1\x03\x00\xB6\xF1\x03\x00\xB7\xF1"
-     "\x03\x00"
-     "\xB8\xD5\x03\x00\xB9\xDD\x03\x00\xBA\xD5\x03\x00\xBB\x0D\x03\x00\xBC\x15"
-     "\x03\x00"
-     "\xBD\x1D\x03\x00\xBE\x15\x03\x00\xBF\x0D\x03\x00\x40\x44\x00\x80\x44\x44"
-     "\x00\x80"
-     "\x85\xF0\x20\x00\xD4\x42\x00\x80\x48\x44\x00\x80\x4C\x44\x00\x80\x50\x44"
-     "\x00\x80"
-     "\x54\x44\x00\x80\xA8\x0D\x02\x00\xA9\xE1\x03\x00\xAA\xF9\x03\x00\xAB\xF1"
-     "\x03\x00"
-     "\xAC\xD9\x03\x00\xAD\xD9\x03\x00\xAE\xBD\x03\x00\xAF\xB1\x03\x00\xA3\x11"
-     "\x02\x00"
-     "\x58\x44\x00\x80\x5C\x44\x00\x80\x60\x44\x00\x80\xBE\x1C\x05\x00\xA6\x39"
-     "\x02\x00"
-     "\xA5\x31\x02\x00\x64\x44\x00\x80\xAB\x75\x02\x00\xAA\x75\x02\x00\x68\x44"
-     "\x00\x80"
-     "\x6C\x44\x00\x80\xAF\x55\x02\x00\xAE\x55\x02\x00\xAD\x65\x02\x00\xAC\x65"
-     "\x02\x00"
-     "\x70\x44\x00\x80\x74\x44\x00\x80\x78\x44\x00\x80\x7C\x44\x00\x80\x80\x44"
-     "\x00\x80"
-     "\xE3\x54\x03\x00\x84\x44\x00\x80\xE1\x4C\x00\x00\x88\x44\x00\x80\xEF\x80"
-     "\x03\x00"
-     "\x8C\x44\x00\x80\x90\x44\x00\x80\x80\x11\x00\x00\x81\x1D\x00\x00\x82\x15"
-     "\x00\x00"
-     "\x98\x44\x00\x80\xE1\xD4\x7B\x00\x9C\x44\x00\x80\xE3\x54\x03\x00\xA0\x44"
-     "\x00\x80"
-     "\x86\x30\x07\x00\x87\x74\x04\x00\xA4\x44\x00\x80\xEF\x44\x7E\x00\xA8\x44"
-     "\x00\x80"
-     "\xE1\x24\x7E\x00\xAC\x44\x00\x80\xE3\x54\x01\x00\xB0\x44\x00\x80\xB4\x44"
-     "\x00\x80"
-     "\xEF\xF8\x7F\x00\xBF\x00\x3C\x00\xB3\x85\x01\x00\xB8\x44\x00\x80\xBC\x44"
-     "\x00\x80"
-     "\x84\x80\x07\x00\xC0\x44\x00\x80\xB6\xA9\x01\x00\xB5\xA1\x01\x00\xC4\x44"
-     "\x00\x80"
-     "\xBB\x35\x01\x00\xBA\x35\x01\x00\xC8\x44\x00\x80\xCC\x44\x00\x80\xBF\xF9"
-     "\x00\x00"
-     "\xBE\xF1\x00\x00\xBD\x25\x01\x00\xBC\x25\x01\x00\xA8\x21\x06\x00\xA9\x21"
-     "\x06\x00"
-     "\xAA\x4D\x06\x00\xAB\x5D\x06\x00\xAC\x69\x06\x00\xAD\x69\x06\x00\xAE\xD9"
-     "\x06\x00"
-     "\xAF\xD9\x06\x00\x94\x44\x00\x80\xD0\x44\x00\x80\xD4\x44\x00\x80\xD8\x44"
-     "\x00\x80"
-     "\xA2\xE9\x04\x00\xA3\xE5\x02\x00\xA0\xF5\x06\x00\xA1\x1D\x04\x00\xB8\x69"
-     "\x07\x00"
-     "\xB9\x69\x07\x00\xBA\x09\x07\x00\xBB\x09\x07\x00\xBC\x19\x07\x00\xBD\x19"
-     "\x07\x00"
-     "\xBE\x09\x07\x00\xBF\x05\x07\x00\xB0\xB1\x06\x00\xB1\xB1\x06\x00\xB2\xB1"
-     "\x06\x00"
-     "\xB3\xB1\x06\x00\xB4\x75\x07\x00\xB5\x7D\x07\x00\xB6\x75\x07\x00\xB7\x61"
-     "\x07\x00"
-     "\xA3\xC5\x06\x00\x8F\x65\x6C\x00\xDC\x44\x00\x80\xE0\x44\x00\x80\xE4\x44"
-     "\x00\x80"
-     "\xA6\xE9\x06\x00\xA5\xE1\x06\x00\xE8\x44\x00\x80\xAB\x75\x06\x00\xAA\x75"
-     "\x06\x00"
-     "\xEC\x44\x00\x80\xF0\x44\x00\x80\xAF\xB9\x07\x00\xAE\xB1\x07\x00\xAD\x65"
-     "\x06\x00"
-     "\xAC\x65\x06\x00\x9E\xC5\x78\x00\x9F\xD9\x79\x00\x9C\xA5\x7D\x00\x9D\xCD"
-     "\x78\x00"
-     "\x9A\xD1\x7C\x00\x9B\xF5\x7C\x00\x98\xE9\x70\x00\x99\xB5\x71\x00\x96\x45"
-     "\x72\x00"
-     "\x97\xC5\x70\x00\x94\xC9\x74\x00\x95\x69\x74\x00\x92\x45\x68\x00\x93\x65"
-     "\x76\x00"
-     "\x90\xF9\x68\x00\x91\xF9\x68\x00\x8A\x85\x75\x00\x8B\xA1\x73\x00\xF4\x44"
-     "\x00\x80"
-     "\xF8\x44\x00\x80\x8E\x79\x70\x00\x8F\x15\x7D\x00\x8C\x39\x71\x00\x8D\x2D"
-     "\x71\x00"
-     "\x82\xC1\x6D\x00\x83\x25\x69\x00\xFC\x44\x00\x80\x00\x45\x00\x80\x86\xA1"
-     "\x77\x00"
-     "\x87\x15\x75\x00\x84\x3D\x69\x00\x85\x89\x69\x00\x9A\x61\x05\x00\x9B\x9D"
-     "\x05\x00"
-     "\xBE\x08\x01\x00\x08\x45\x00\x80\x0C\x45\x00\x80\xE1\x10\x05\x00\x9C\x89"
-     "\x03\x00"
-     "\xE3\x04\x05\x00\x92\x41\x7D\x00\x93\x09\x7C\x00\x10\x45\x00\x80\xEF\x84"
-     "\x04\x00"
-     "\x96\x6D\x78\x00\x97\xB9\x07\x00\x94\x79\x79\x00\x95\x79\x79\x00\xA1\xA1"
-     "\x01\x00"
-     "\x14\x45\x00\x80\xA3\x8D\x0D\x00\xA2\x01\x0C\x00\xA5\xED\x0C\x00\xA4\xC5"
-     "\x0D\x00"
-     "\xA7\x91\x09\x00\xA6\x99\x09\x00\xA9\x01\x14\x00\xA8\x65\x09\x00\xAB\x85"
-     "\x14\x00"
-     "\xAA\x8D\x15\x00\xAD\x91\x11\x00\xAC\x01\x10\x00\xAF\x01\x1C\x00\xAE\x61"
-     "\x11\x00"
-     "\xB1\xF1\x1D\x00\xB0\xE5\x1D\x00\xB3\xCD\x19\x00\xB2\x01\x18\x00\xB5\x01"
-     "\x18\x00"
-     "\xB4\xB5\x19\x00\x86\x78\x00\x00\x87\x1C\x01\x00\x80\xED\x00\x00\x81\xCD"
-     "\x00\x00"
-     "\x82\xC5\x00\x00\x18\x45\x00\x80\x1C\x45\x00\x80\x20\x45\x00\x80\x24\x45"
-     "\x00\x80"
-     "\x28\x45\x00\x80\xA8\x7D\x05\x00\xA9\x81\x05\x00\xAA\x85\x05\x00\xAB\x95"
-     "\x05\x00"
-     "\xAC\xB1\x05\x00\xAD\xBD\x05\x00\xAE\xA9\x05\x00\xAF\xA9\x05\x00\x2C\x45"
-     "\x00\x80"
-     "\x30\x45\x00\x80\x34\x45\x00\x80\x38\x45\x00\x80\x3C\x45\x00\x80\x40\x45"
-     "\x00\x80"
-     "\x44\x45\x00\x80\x48\x45\x00\x80\xB8\x6D\x02\x00\xB9\xCD\x02\x00\xBA\xC5"
-     "\x02\x00"
-     "\xBB\xDD\x02\x00\xBC\xCD\x02\x00\xBD\xFD\x02\x00\xBE\xF5\x02\x00\xBF\x55"
-     "\x03\x00"
-     "\xB0\xA5\x05\x00\xB1\xAD\x05\x00\xB2\xB9\x05\x00\xB3\xB1\x05\x00\xB4\x99"
-     "\x05\x00"
-     "\xB5\x99\x05\x00\xB6\x5D\x02\x00\xB7\x55\x02\x00\xB3\x0D\x05\x00\x4C\x45"
-     "\x00\x80"
-     "\x50\x45\x00\x80\x54\x45\x00\x80\x58\x45\x00\x80\xB6\x3D\x05\x00\xB5\x39"
-     "\x05\x00"
-     "\x5C\x45\x00\x80\xBB\x69\x05\x00\xBA\x61\x05\x00\x60\x45\x00\x80\x64\x45"
-     "\x00\x80"
-     "\xBF\x5D\x05\x00\xBE\x5D\x05\x00\xBD\x65\x05\x00\xBC\x6D\x05\x00\x80\x1D"
-     "\x00\x00"
-     "\x81\x25\x00\x00\x82\x2D\x00\x00\xA3\x41\x05\x00\x24\x44\x00\x80\xA5\x75"
-     "\x05\x00"
-     "\xA6\x71\x05\x00\x68\x45\x00\x80\x6C\x45\x00\x80\x70\x45\x00\x80\xAA\x2D"
-     "\x05\x00"
-     "\xAB\x25\x05\x00\xAC\x21\x05\x00\xAD\x29\x05\x00\xAE\x11\x05\x00\xAF\x11"
-     "\x05\x00"
-     "\xA8\x4D\x02\x00\xA9\x55\x02\x00\xAA\x61\x02\x00\xAB\x89\x02\x00\xAC\xFD"
-     "\x02\x00"
-     "\xAD\xE5\x02\x00\xAE\xED\x02\x00\xAF\xE5\x02\x00\x84\x24\x02\x00\xBE\xA8"
-     "\x0D\x00"
-     "\x87\x68\x0D\x00\x86\x0C\x0D\x00\x78\x45\x00\x80\x7C\x45\x00\x80\x80\x45"
-     "\x00\x80"
-     "\x84\x45\x00\x80\xB8\x7D\x01\x00\xB9\xC1\x01\x00\xBA\xC1\x01\x00\xBB\xC1"
-     "\x01\x00"
-     "\xBC\xC1\x01\x00\xBD\xC9\x01\x00\xBE\xF1\x01\x00\xBF\xF1\x01\x00\xB0\x9D"
-     "\x02\x00"
-     "\xB1\x45\x01\x00\xB2\x4D\x01\x00\xB3\x45\x01\x00\xB4\x5D\x01\x00\xB5\x45"
-     "\x01\x00"
-     "\xB6\x4D\x01\x00\xB7\x45\x01\x00\x88\x45\x00\x80\x8C\x45\x00\x80\x90\x45"
-     "\x00\x80"
-     "\x94\x45\x00\x80\x98\x45\x00\x80\x9C\x45\x00\x80\xA0\x45\x00\x80\xEF\x2C"
-     "\x00\x00"
-     "\xA4\x45\x00\x80\xE1\x2C\x06\x00\xA8\x45\x00\x80\xE3\x70\x06\x00\xBE\x3C"
-     "\x0C\x00"
-     "\xAC\x45\x00\x80\xB0\x45\x00\x80\xB4\x45\x00\x80\x84\x40\x0D\x00\xB8\x45"
-     "\x00\x80"
-     "\xB3\xA1\x02\x00\xBC\x45\x00\x80\xB5\x3D\x02\x00\xC0\x45\x00\x80\xC4\x45"
-     "\x00\x80"
-     "\xB6\x2D\x02\x00\xBE\xE0\x0F\x00\xC8\x45\x00\x80\xBB\xC9\x01\x00\xBA\x11"
-     "\x02\x00"
-     "\xBD\xD9\x01\x00\xBC\xD1\x01\x00\xBF\xC9\x01\x00\xBE\xD1\x01\x00\xA3\x69"
-     "\x0D\x00"
-     "\x74\x45\x00\x80\xCC\x45\x00\x80\xD0\x45\x00\x80\xD4\x45\x00\x80\xA6\xE5"
-     "\x0D\x00"
-     "\xA5\xF5\x0D\x00\xD8\x45\x00\x80\xAB\x01\x0E\x00\xAA\xD9\x0D\x00\xDC\x45"
-     "\x00\x80"
-     "\xE0\x45\x00\x80\xAF\x01\x0E\x00\xAE\x19\x0E\x00\xAD\x11\x0E\x00\xAC\x19"
-     "\x0E\x00"
-     "\x80\x5D\x00\x00\x81\x6D\x00\x00\x82\x65\x00\x00\xB3\x75\x0F\x00\xE4\x45"
-     "\x00\x80"
-     "\xB5\xD5\x0F\x00\xB6\xDD\x0F\x00\xE8\x45\x00\x80\x86\x60\x03\x00\x87\x04"
-     "\x03\x00"
-     "\xBA\xF9\x0F\x00\xBB\xF9\x0F\x00\xBC\xE9\x0F\x00\xBD\xD5\x0F\x00\xBE\xD1"
-     "\x0F\x00"
-     "\xBF\xC9\x0F\x00\xA8\xB1\x0F\x00\xA9\xB1\x0F\x00\xAA\xB1\x0F\x00\xAB\xB1"
-     "\x0F\x00"
-     "\xAC\x11\x0F\x00\xAD\x11\x0F\x00\xAE\x11\x0F\x00\xAF\x11\x0F\x00\xEC\x45"
-     "\x00\x80"
-     "\xF0\x45\x00\x80\xF4\x45\x00\x80\xF8\x45\x00\x80\xFC\x45\x00\x80\x00\x46"
-     "\x00\x80"
-     "\x04\x46\x00\x80\x08\x46\x00\x80\xB8\x05\x0F\x00\xB9\x09\x0F\x00\xBA\x1D"
-     "\x0F\x00"
-     "\xBB\x15\x0F\x00\xBC\x35\x0F\x00\xBD\xDD\x00\x00\xBE\xD5\x00\x00\xBF\xCD"
-     "\x00\x00"
-     "\xB0\x71\x0F\x00\xB1\x71\x0F\x00\xB2\x71\x0F\x00\xB3\x4D\x0F\x00\xB4\x59"
-     "\x0F\x00"
-     "\xB5\x41\x0F\x00\xB6\x41\x0F\x00\xB7\x41\x0F\x00\xA3\x39\x0E\x00\x0C\x46"
-     "\x00\x80"
-     "\x10\x46\x00\x80\x14\x46\x00\x80\x18\x46\x00\x80\xA6\x91\x0E\x00\xA5\x99"
-     "\x0E\x00"
-     "\x1C\x46\x00\x80\xAB\xB5\x0E\x00\xAA\xB5\x0E\x00\x20\x46\x00\x80\x24\x46"
-     "\x00\x80"
-     "\xAF\x85\x0E\x00\xAE\x9D\x0E\x00\xAD\x99\x0E\x00\xAC\xA5\x0E\x00\x28\x46"
-     "\x00\x80"
-     "\x2C\x46\x00\x80\x30\x46\x00\x80\x34\x46\x00\x80\x38\x46\x00\x80\x3C\x46"
-     "\x00\x80"
-     "\x40\x46\x00\x80\x44\x46\x00\x80\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\x48\x46\x00\x80\x4C\x46\x00\x80\x50\x46\x00\x80\x84\xF4\x01\x00\x54\x46"
-     "\x00\x80"
-     "\x86\x20\x00\x00\x87\x04\x01\x00\x58\x46\x00\x80\x04\x45\x00\x80\x5C\x46"
-     "\x00\x80"
-     "\x60\x46\x00\x80\x64\x46\x00\x80\x68\x46\x00\x80\xA8\xFD\x01\x00\xA9\x05"
-     "\x01\x00"
-     "\xAA\x0D\x01\x00\xAB\x05\x01\x00\xAC\x1D\x01\x00\xAD\x05\x01\x00\xAE\x0D"
-     "\x01\x00"
-     "\xAF\x05\x01\x00\xB0\x7D\x01\x00\xB1\xC1\x00\x00\xB2\xC1\x00\x00\xB3\xC1"
-     "\x00\x00"
-     "\xB4\xC1\x00\x00\xB5\xC9\x00\x00\xB6\xF1\x00\x00\xB7\xF1\x00\x00\xB8\xA1"
-     "\x00\x00"
-     "\xB9\xA1\x00\x00\xBA\xA1\x00\x00\xBB\xA1\x00\x00\xBC\xA1\x00\x00\xBD\xA1"
-     "\x00\x00"
-     "\xBE\xA1\x00\x00\xBF\xA1\x00\x00\xE3\x30\x0F\x00\x6C\x46\x00\x80\xE1\xC8"
-     "\x0D\x00"
-     "\x70\x46\x00\x80\x74\x46\x00\x80\x78\x46\x00\x80\x7C\x46\x00\x80\x80\x46"
-     "\x00\x80"
-     "\x84\x00\x04\x00\x84\x46\x00\x80\x88\x46\x00\x80\x8C\x46\x00\x80\xEF\xEC"
-     "\x0F\x00"
-     "\x90\x46\x00\x80\x94\x46\x00\x80\x98\x46\x00\x80\xB3\xE5\x02\x00\x9C\x46"
-     "\x00\x80"
-     "\xA0\x46\x00\x80\xA4\x46\x00\x80\xA8\x46\x00\x80\xB6\xE5\x02\x00\xB5\xF5"
-     "\x02\x00"
-     "\xAC\x46\x00\x80\xBB\x65\x03\x00\xBA\xD9\x02\x00\xB0\x46\x00\x80\xB4\x46"
-     "\x00\x80"
-     "\xBF\x6D\x03\x00\xBE\x65\x03\x00\xBD\x75\x03\x00\xBC\x75\x03\x00\x80\x11"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x82\x29\x00\x00\xA3\x29\x02\x00\xB8\x46\x00\x80\xA5\x39"
-     "\x02\x00"
-     "\xA6\x29\x02\x00\xBC\x46\x00\x80\xC0\x46\x00\x80\xC8\x46\x00\x80\xAA\x15"
-     "\x02\x00"
-     "\xAB\xA9\x03\x00\xAC\xB9\x03\x00\xAD\xB9\x03\x00\xAE\xA9\x03\x00\xAF\xA1"
-     "\x03\x00"
-     "\xEF\xFC\x02\x00\xBE\x48\x04\x00\x86\x68\x05\x00\x87\x60\x03\x00\xCC\x46"
-     "\x00\x80"
-     "\xD0\x46\x00\x80\xD4\x46\x00\x80\xD8\x46\x00\x80\xDC\x46\x00\x80\xE0\x46"
-     "\x00\x80"
-     "\xE4\x46\x00\x80\xE8\x46\x00\x80\xE3\xC0\x02\x00\xEC\x46\x00\x80\xE1\x60"
-     "\x01\x00"
-     "\xF0\x46\x00\x80\xA8\x89\x02\x00\xA9\x95\x02\x00\xAA\x9D\x02\x00\xAB\x95"
-     "\x02\x00"
-     "\xAC\x8D\x02\x00\xAD\x11\x01\x00\xAE\x15\x01\x00\xAF\x0D\x01\x00\xF4\x46"
-     "\x00\x80"
-     "\xF8\x46\x00\x80\xFC\x46\x00\x80\x00\x47\x00\x80\x04\x47\x00\x80\x08\x47"
-     "\x00\x80"
-     "\x0C\x47\x00\x80\x10\x47\x00\x80\xB8\x7D\x01\x00\xB9\x01\x01\x00\xBA\x01"
-     "\x01\x00"
-     "\xBB\x01\x01\x00\xBC\x01\x01\x00\xBD\x09\x01\x00\xBE\x31\x01\x00\xBF\x31"
-     "\x01\x00"
-     "\xB0\x75\x01\x00\xB1\x7D\x01\x00\xB2\x75\x01\x00\xB3\x49\x01\x00\xB4\x59"
-     "\x01\x00"
-     "\xB5\x59\x01\x00\xB6\x4D\x01\x00\xB7\x45\x01\x00\xE1\xB4\x0E\x00\x14\x47"
-     "\x00\x80"
-     "\xE3\x58\x0E\x00\xE6\x00\x00\x00\x18\x47\x00\x80\x1C\x47\x00\x80\xBE\x94"
-     "\x05\x00"
-     "\x20\x47\x00\x80\x24\x47\x00\x80\x82\x19\x00\x00\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\x28\x47\x00\x80\x2C\x47\x00\x80\xEF\xF0\x0E\x00\x30\x47\x00\x80\x84\x48"
-     "\x05\x00"
-     "\x38\x47\x00\x80\x87\x28\x05\x00\x86\xEC\x04\x00\x3C\x47\x00\x80\xE1\xAC"
-     "\x01\x00"
-     "\x40\x47\x00\x80\xE3\xB4\x00\x00\x44\x47\x00\x80\x48\x47\x00\x80\x4C\x47"
-     "\x00\x80"
-     "\xEF\xF4\x0E\x00\x50\x47\x00\x80\x54\x47\x00\x80\x58\x47\x00\x80\x5C\x47"
-     "\x00\x80"
-     "\xA3\x95\x05\x00\xC4\x46\x00\x80\x34\x47\x00\x80\x60\x47\x00\x80\x64\x47"
-     "\x00\x80"
-     "\xA6\x11\x06\x00\xA5\x85\x05\x00\x68\x47\x00\x80\xAB\x39\x06\x00\xAA\x31"
-     "\x06\x00"
-     "\x6C\x47\x00\x80\x70\x47\x00\x80\xAF\x1D\x06\x00\xAE\x1D\x06\x00\xAD\x1D"
-     "\x06\x00"
-     "\xAC\x21\x06\x00\x74\x47\x00\x80\xB3\x91\x06\x00\x78\x47\x00\x80\x7C\x47"
-     "\x00\x80"
-     "\xB6\xB1\x06\x00\x80\x47\x00\x80\x84\x47\x00\x80\xB5\x81\x06\x00\xBA\x8D"
-     "\x06\x00"
-     "\xBB\x91\x06\x00\x88\x47\x00\x80\x8C\x47\x00\x80\xBE\x55\x07\x00\xBF\x5D"
-     "\x07\x00"
-     "\xBC\x85\x06\x00\xBD\x5D\x07\x00\xA3\x55\x06\x00\x90\x47\x00\x80\x94\x47"
-     "\x00\x80"
-     "\x98\x47\x00\x80\x9C\x47\x00\x80\xA6\x75\x06\x00\xA5\x45\x06\x00\xA0\x47"
-     "\x00\x80"
-     "\xAB\x55\x06\x00\xAA\x49\x06\x00\x84\x64\x00\x00\xBE\x64\x00\x00\xAF\x99"
-     "\x07\x00"
-     "\xAE\x91\x07\x00\xAD\x99\x07\x00\xAC\x41\x06\x00\xA4\x47\x00\x80\xA8\x47"
-     "\x00\x80"
-     "\xBA\xDD\x06\x00\xBB\xD5\x06\x00\xBC\xCD\x06\x00\xBD\xF1\x06\x00\xBE\xF1"
-     "\x06\x00"
-     "\xBF\xF1\x06\x00\x80\x6D\x00\x00\x81\x75\x00\x00\x82\x75\x00\x00\xB3\x79"
-     "\x05\x00"
-     "\xAC\x47\x00\x80\xB5\x69\x05\x00\xB6\xFD\x06\x00\xB0\x47\x00\x80\xA8\xA5"
-     "\x07\x00"
-     "\xA9\xD5\x07\x00\xAA\xDD\x07\x00\xAB\xD5\x07\x00\xAC\xCD\x07\x00\xAD\xF5"
-     "\x07\x00"
-     "\xAE\xFD\x07\x00\xAF\xF5\x07\x00\x86\x00\x01\x00\x87\x84\x01\x00\xB4\x47"
-     "\x00\x80"
-     "\xB8\x47\x00\x80\xBC\x47\x00\x80\xC0\x47\x00\x80\xC4\x47\x00\x80\xC8\x47"
-     "\x00\x80"
-     "\xB8\x75\x06\x00\xB9\x7D\x06\x00\xBA\x75\x06\x00\xBB\xC9\x06\x00\xBC\xC1"
-     "\x06\x00"
-     "\xBD\xC1\x06\x00\xBE\xC5\x06\x00\xBF\xFD\x06\x00\xB0\x8D\x07\x00\xB1\x91"
-     "\x07\x00"
-     "\xB2\x95\x07\x00\xB3\x4D\x06\x00\xB4\x55\x06\x00\xB5\x5D\x06\x00\xB6\x55"
-     "\x06\x00"
-     "\xB7\x4D\x06\x00\xA8\x45\x06\x00\xA9\x55\x06\x00\xAA\x51\x06\x00\xAB\x6D"
-     "\x06\x00"
-     "\xAC\x75\x06\x00\xAD\x7D\x06\x00\xAE\x69\x06\x00\xAF\x69\x06\x00\xCC\x47"
-     "\x00\x80"
-     "\xD0\x47\x00\x80\xD4\x47\x00\x80\xD8\x47\x00\x80\xDC\x47\x00\x80\xE0\x47"
-     "\x00\x80"
-     "\xE4\x47\x00\x80\xE8\x47\x00\x80\xB8\x89\x01\x00\xB9\x89\x01\x00\xBA\x9D"
-     "\x01\x00"
-     "\xBB\x95\x01\x00\xBC\xB9\x01\x00\xBD\xB9\x01\x00\xBE\x5D\x01\x00\xBF\x55"
-     "\x01\x00"
-     "\xB0\x19\x06\x00\xB1\x19\x06\x00\xB2\xE5\x01\x00\xB3\xFD\x01\x00\xB4\xED"
-     "\x01\x00"
-     "\xB5\xDD\x01\x00\xB6\xD5\x01\x00\xB7\xBD\x01\x00\xB3\xD9\x05\x00\xEC\x47"
-     "\x00\x80"
-     "\xF0\x47\x00\x80\xBE\x40\x03\x00\xF4\x47\x00\x80\xB6\xFD\x05\x00\xB5\xFD"
-     "\x05\x00"
-     "\xF8\x47\x00\x80\xBB\xDD\x05\x00\xBA\xDD\x05\x00\xFC\x47\x00\x80\x00\x48"
-     "\x00\x80"
-     "\xBF\x19\x02\x00\xBE\x29\x02\x00\xBD\x29\x02\x00\xBC\x31\x02\x00\x04\x48"
-     "\x00\x80"
-     "\xA3\x9D\x05\x00\x08\x48\x00\x80\x0C\x48\x00\x80\xA6\xB9\x05\x00\x84\x98"
-     "\x03\x00"
-     "\x10\x48\x00\x80\xA5\xB9\x05\x00\xAA\x99\x05\x00\xAB\x99\x05\x00\x86\xC8"
-     "\x1C\x00"
-     "\x87\xFC\x03\x00\xAE\x6D\x02\x00\xAF\x5D\x02\x00\xAC\x75\x02\x00\xAD\x6D"
-     "\x02\x00"
-     "\xA8\x45\x02\x00\xA9\x75\x02\x00\xAA\x75\x02\x00\xAB\xA5\x03\x00\xAC\xBD"
-     "\x03\x00"
-     "\xAD\xA5\x03\x00\xAE\xAD\x03\x00\xAF\xA5\x03\x00\x80\xE9\x00\x00\x81\xE9"
-     "\x00\x00"
-     "\x82\xF1\x00\x00\x18\x48\x00\x80\x1C\x48\x00\x80\x20\x48\x00\x80\x84\xF4"
-     "\x1C\x00"
-     "\x24\x48\x00\x80\xB8\x7D\x03\x00\xB9\xC1\x00\x00\xBA\xC1\x00\x00\xBB\xDD"
-     "\x00\x00"
-     "\xBC\xCD\x00\x00\xBD\xFD\x00\x00\xBE\xF5\x00\x00\xBF\x9D\x00\x00\xB0\xDD"
-     "\x03\x00"
-     "\xB1\x45\x03\x00\xB2\x4D\x03\x00\xB3\x45\x03\x00\xB4\x5D\x03\x00\xB5\x45"
-     "\x03\x00"
-     "\xB6\x4D\x03\x00\xB7\x45\x03\x00\xB3\xFD\x03\x00\x28\x48\x00\x80\x2C\x48"
-     "\x00\x80"
-     "\x30\x48\x00\x80\x34\x48\x00\x80\xB6\x25\x02\x00\xB5\x35\x02\x00\x38\x48"
-     "\x00\x80"
-     "\xBB\x81\x02\x00\xBA\x19\x02\x00\x3C\x48\x00\x80\x40\x48\x00\x80\xBF\x81"
-     "\x02\x00"
-     "\xBE\x99\x02\x00\xBD\x91\x02\x00\xBC\x99\x02\x00\x44\x48\x00\x80\xE1\x7C"
-     "\x00\x00"
-     "\x48\x48\x00\x80\xE3\x08\x03\x00\x4C\x48\x00\x80\x50\x48\x00\x80\x54\x48"
-     "\x00\x80"
-     "\x58\x48\x00\x80\x5C\x48\x00\x80\x60\x48\x00\x80\x64\x48\x00\x80\x68\x48"
-     "\x00\x80"
-     "\x6C\x48\x00\x80\x70\x48\x00\x80\x74\x48\x00\x80\xEF\x88\x03\x00\xE1\x50"
-     "\x07\x00"
-     "\xE3\x3C\x01\x00\xE3\x58\x06\x00\xE1\x58\x06\x00\x80\x0D\x00\x00\x81\x31"
-     "\x00\x00"
-     "\x82\x31\x00\x00\x78\x48\x00\x80\xBE\xE8\x1D\x00\x7C\x48\x00\x80\x80\x48"
-     "\x00\x80"
-     "\x88\x48\x00\x80\x8C\x48\x00\x80\xEF\xE0\x01\x00\xEF\xD8\x06\x00\x90\x48"
-     "\x00\x80"
-     "\xA3\x3D\x03\x00\x94\x48\x00\x80\x86\x68\x1C\x00\x87\xEC\x1C\x00\x98\x48"
-     "\x00\x80"
-     "\xA6\xE5\x02\x00\xA5\xF5\x02\x00\x9C\x48\x00\x80\xAB\x41\x02\x00\xAA\xD9"
-     "\x02\x00"
-     "\xA0\x48\x00\x80\xA4\x48\x00\x80\xAF\x41\x02\x00\xAE\x59\x02\x00\xAD\x51"
-     "\x02\x00"
-     "\xAC\x59\x02\x00\xA8\xAD\x1E\x00\xA9\xB5\x1E\x00\xAA\xBD\x1E\x00\xAB\xB1"
-     "\x1E\x00"
-     "\xAC\x15\x1F\x00\xAD\x1D\x1F\x00\xAE\x15\x1F\x00\xAF\x0D\x1F\x00\x84\x48"
-     "\x00\x80"
-     "\xA8\x48\x00\x80\xAC\x48\x00\x80\xB0\x48\x00\x80\xB4\x48\x00\x80\xB8\x48"
-     "\x00\x80"
-     "\xBC\x48\x00\x80\xC0\x48\x00\x80\xB8\x1D\x1F\x00\xB9\x25\x1F\x00\xBA\x2D"
-     "\x1F\x00"
-     "\xBB\x25\x1F\x00\xBC\x3D\x1F\x00\xBD\x25\x1F\x00\xBE\x2D\x1F\x00\xBF\x25"
-     "\x1F\x00"
-     "\xB0\x7D\x1F\x00\xB1\x4D\x1F\x00\xB2\x45\x1F\x00\xB3\x2D\x1F\x00\xB4\x39"
-     "\x1F\x00"
-     "\xB5\x25\x1F\x00\xB6\x2D\x1F\x00\xB7\x25\x1F\x00\xB3\x35\x1E\x00\xC4\x48"
-     "\x00\x80"
-     "\xC8\x48\x00\x80\xCC\x48\x00\x80\xD0\x48\x00\x80\xB6\x9D\x1F\x00\xB5\x9D"
-     "\x1F\x00"
-     "\xD4\x48\x00\x80\xBB\x89\x1F\x00\xBA\xB9\x1F\x00\xD8\x48\x00\x80\xDC\x48"
-     "\x00\x80"
-     "\xBF\xE1\x1F\x00\xBE\xFD\x1F\x00\xBD\xF9\x1F\x00\xBC\x91\x1F\x00\xE0\x48"
-     "\x00\x80"
-     "\xA3\x71\x1E\x00\xE4\x48\x00\x80\xE8\x48\x00\x80\xA6\xD9\x1F\x00\xEC\x48"
-     "\x00\x80"
-     "\xF0\x48\x00\x80\xA5\xD9\x1F\x00\xAA\xFD\x1F\x00\xAB\xCD\x1F\x00\x84\xAC"
-     "\x03\x00"
-     "\xBE\xA0\x03\x00\xAE\xB9\x1F\x00\xAF\xA5\x1F\x00\xAC\xD5\x1F\x00\xAD\xBD"
-     "\x1F\x00"
-     "\xA8\xAD\x1E\x00\xA9\xB5\x1E\x00\xAA\xB5\x1E\x00\xAB\xC9\x1E\x00\xAC\xC1"
-     "\x1E\x00"
-     "\xAD\xC9\x1E\x00\xAE\xF1\x1E\x00\xAF\xF1\x1E\x00\x80\xED\x01\x00\x81\xF1"
-     "\x01\x00"
-     "\x82\xF1\x01\x00\xF4\x48\x00\x80\x86\x94\x00\x00\x87\x74\x01\x00\xF8\x48"
-     "\x00\x80"
-     "\xFC\x48\x00\x80\xB8\xC9\x01\x00\xB9\xC9\x01\x00\xBA\xD9\x01\x00\xBB\xD1"
-     "\x01\x00"
-     "\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\x99\x01\x00\xBF\x95\x01\x00\xB0\x45"
-     "\x01\x00"
-     "\xB1\x4D\x01\x00\xB2\x45\x01\x00\xB3\x5D\x01\x00\xB4\x45\x01\x00\xB5\x4D"
-     "\x01\x00"
-     "\xB6\x45\x01\x00\xB7\xF9\x01\x00\x00\x49\x00\x80\x04\x49\x00\x80\xB3\x35"
-     "\x1E\x00"
-     "\x08\x49\x00\x80\xB5\x41\x1E\x00\x0C\x49\x00\x80\x10\x49\x00\x80\xB6\x41"
-     "\x1E\x00"
-     "\x14\x48\x00\x80\x84\x24\x01\x00\xBB\x81\x01\x00\xBA\x89\x01\x00\xBD\x81"
-     "\x01\x00"
-     "\xBC\x99\x01\x00\xBF\x81\x01\x00\xBE\x89\x01\x00\xA3\x7D\x1E\x00\x14\x49"
-     "\x00\x80"
-     "\x18\x49\x00\x80\x1C\x49\x00\x80\x20\x49\x00\x80\xA6\x09\x1E\x00\xA5\x09"
-     "\x1E\x00"
-     "\x24\x49\x00\x80\xAB\xC9\x01\x00\xAA\xC1\x01\x00\x28\x49\x00\x80\x2C\x49"
-     "\x00\x80"
-     "\xAF\xC9\x01\x00\xAE\xC1\x01\x00\xAD\xC9\x01\x00\xAC\xD1\x01\x00\x30\x49"
-     "\x00\x80"
-     "\x34\x49\x00\x80\x38\x49\x00\x80\x3C\x49\x00\x80\x40\x49\x00\x80\xBE\x58"
-     "\x02\x00"
-     "\x44\x49\x00\x80\xEF\x44\x1F\x00\x48\x49\x00\x80\xE1\x54\x1E\x00\x4C\x49"
-     "\x00\x80"
-     "\xE3\x84\x1E\x00\x50\x49\x00\x80\x54\x49\x00\x80\x58\x49\x00\x80\x5C\x49"
-     "\x00\x80"
-     "\x80\x79\x00\x00\x81\x79\x00\x00\x82\x19\x00\x00\xB3\xCD\x03\x00\x60\x49"
-     "\x00\x80"
-     "\xB5\xD9\x03\x00\xB6\xCD\x03\x00\x68\x49\x00\x80\x86\xC0\x04\x00\x87\xEC"
-     "\x03\x00"
-     "\xBA\xA1\x03\x00\xBB\xA1\x03\x00\xBC\xA1\x03\x00\xBD\xA9\x03\x00\xBE\x91"
-     "\x03\x00"
-     "\xBF\x91\x03\x00\x6C\x49\x00\x80\x70\x49\x00\x80\x74\x49\x00\x80\x78\x49"
-     "\x00\x80"
-     "\xEF\x7C\x02\x00\x7C\x49\x00\x80\x80\x49\x00\x80\x84\x49\x00\x80\xE3\x14"
-     "\x01\x00"
-     "\x88\x49\x00\x80\xE1\x80\x01\x00\x8C\x49\x00\x80\x90\x49\x00\x80\x94\x49"
-     "\x00\x80"
-     "\x98\x49\x00\x80\x9C\x49\x00\x80\xA3\x41\x03\x00\xA0\x49\x00\x80\xA4\x49"
-     "\x00\x80"
-     "\xA8\x49\x00\x80\xAC\x49\x00\x80\xA6\x41\x03\x00\xA5\x55\x03\x00\xB0\x49"
-     "\x00\x80"
-     "\xAB\x2D\x03\x00\xAA\x2D\x03\x00\xB4\x49\x00\x80\xB8\x49\x00\x80\xAF\x1D"
-     "\x03\x00"
-     "\xAE\x1D\x03\x00\xAD\x25\x03\x00\xAC\x2D\x03\x00\xA9\x3D\x02\x00\xA8\x3D"
-     "\x02\x00"
-     "\xAB\x8D\x02\x00\xAA\x8D\x02\x00\xAD\x85\x02\x00\xAC\x99\x02\x00\xAF\xDD"
-     "\x02\x00"
-     "\xAE\x85\x02\x00\x84\x2C\x02\x00\xBC\x49\x00\x80\xCA\x00\x00\x00\xC0\x49"
-     "\x00\x80"
-     "\xC4\x49\x00\x80\xC8\x49\x00\x80\xCC\x49\x00\x80\xD0\x49\x00\x80\xB9\x4D"
-     "\x01\x00"
-     "\xB8\x4D\x01\x00\xBB\x51\x01\x00\xBA\x5D\x01\x00\xBD\x79\x01\x00\xBC\x75"
-     "\x01\x00"
-     "\xBF\x6D\x01\x00\xBE\x6D\x01\x00\xB1\xAD\x02\x00\xB0\xAD\x02\x00\xB3\xB5"
-     "\x02\x00"
-     "\xB2\xBD\x02\x00\xB5\x9D\x02\x00\xB4\x9D\x02\x00\xB7\x7D\x01\x00\xB6\x7D"
-     "\x01\x00"
-     "\x82\xBD\x00\x00\xBE\x88\x05\x00\x80\xBD\x00\x00\x81\xB5\x00\x00\xD8\x49"
-     "\x00\x80"
-     "\xDC\x49\x00\x80\xE0\x49\x00\x80\xEF\x88\x01\x00\x84\xA0\x05\x00\xE1\x74"
-     "\x1E\x00"
-     "\xE4\x49\x00\x80\xE3\x54\x01\x00\x86\x10\x04\x00\x87\x74\x05\x00\xE8\x49"
-     "\x00\x80"
-     "\xEC\x49\x00\x80\xB3\x45\x01\x00\xF0\x49\x00\x80\xF4\x49\x00\x80\xF8\x49"
-     "\x00\x80"
-     "\xFC\x49\x00\x80\xB6\x41\x01\x00\xB5\x55\x01\x00\x00\x4A\x00\x80\xBB\x4D"
-     "\x01\x00"
-     "\xBA\x41\x01\x00\x04\x4A\x00\x80\x08\x4A\x00\x80\xBF\x4D\x01\x00\xBE\x4D"
-     "\x01\x00"
-     "\xBD\x4D\x01\x00\xBC\x4D\x01\x00\xA8\x5D\x06\x00\xA9\x85\x06\x00\xAA\x8D"
-     "\x06\x00"
-     "\xAB\x85\x06\x00\xAC\x9D\x06\x00\xAD\x81\x06\x00\xAE\x81\x06\x00\xAF\x81"
-     "\x06\x00"
-     "\xD4\x49\x00\x80\x0C\x4A\x00\x80\x10\x4A\x00\x80\x14\x4A\x00\x80\x18\x4A"
-     "\x00\x80"
-     "\x1C\x4A\x00\x80\x20\x4A\x00\x80\x24\x4A\x00\x80\xB8\x8D\x06\x00\xB9\x95"
-     "\x06\x00"
-     "\xBA\x95\x06\x00\xBB\xA5\x06\x00\xBC\xBD\x06\x00\xBD\x71\x07\x00\xBE\x75"
-     "\x07\x00"
-     "\xBF\x6D\x07\x00\xB0\xC5\x06\x00\xB1\xC9\x06\x00\xB2\xD9\x06\x00\xB3\xD9"
-     "\x06\x00"
-     "\xB4\xC9\x06\x00\xB5\xC9\x06\x00\xB6\xBD\x06\x00\xB7\xB5\x06\x00\xA3\x05"
-     "\x06\x00"
-     "\x28\x4A\x00\x80\x2C\x4A\x00\x80\x30\x4A\x00\x80\x34\x4A\x00\x80\xA6\x01"
-     "\x06\x00"
-     "\xA5\x15\x06\x00\x38\x4A\x00\x80\xAB\x0D\x06\x00\xAA\x01\x06\x00\x3C\x4A"
-     "\x00\x80"
-     "\x40\x4A\x00\x80\xAF\x0D\x06\x00\xAE\x0D\x06\x00\xAD\x0D\x06\x00\xAC\x0D"
-     "\x06\x00"
-     "\x80\x5D\x00\x00\x81\x6D\x00\x00\x82\x65\x00\x00\xB3\x55\x07\x00\xBE\xC4"
-     "\x03\x00"
-     "\xB5\x6D\x07\x00\xB6\x65\x07\x00\x44\x4A\x00\x80\x86\xE0\x00\x00\x87\x14"
-     "\x03\x00"
-     "\xBA\x25\x07\x00\xBB\x3D\x07\x00\xBC\x2D\x07\x00\xBD\x1D\x07\x00\xBE\x15"
-     "\x07\x00"
-     "\xBF\xD9\x00\x00\xA8\x91\x06\x00\xA9\x91\x06\x00\xAA\x91\x06\x00\xAB\xA5"
-     "\x06\x00"
-     "\xAC\xA9\x06\x00\xAD\xA9\x06\x00\xAE\xDD\x06\x00\xAF\xD5\x06\x00\x48\x4A"
-     "\x00\x80"
-     "\x4C\x4A\x00\x80\x50\x4A\x00\x80\x54\x4A\x00\x80\x58\x4A\x00\x80\x5C\x4A"
-     "\x00\x80"
-     "\x60\x4A\x00\x80\x64\x4A\x00\x80\xB8\x71\x01\x00\xB9\x71\x01\x00\xBA\x71"
-     "\x01\x00"
-     "\xBB\x71\x01\x00\xBC\xD5\x01\x00\xBD\xDD\x01\x00\xBE\xD5\x01\x00\xBF\xCD"
-     "\x01\x00"
-     "\xB0\xAD\x06\x00\xB1\xB5\x06\x00\xB2\xB5\x06\x00\xB3\x85\x06\x00\xB4\x9D"
-     "\x06\x00"
-     "\xB5\x51\x01\x00\xB6\x51\x01\x00\xB7\x51\x01\x00\xA5\x39\x06\x00\x68\x4A"
-     "\x00\x80"
-     "\x6C\x4A\x00\x80\xA6\x31\x06\x00\x70\x4A\x00\x80\x74\x4A\x00\x80\xA3\x01"
-     "\x06\x00"
-     "\x78\x4A\x00\x80\xAD\x49\x06\x00\xAC\x79\x06\x00\xAF\x8D\x01\x00\xAE\x41"
-     "\x06\x00"
-     "\x7C\x4A\x00\x80\x64\x49\x00\x80\xAB\x69\x06\x00\xAA\x71\x06\x00\x80\x4A"
-     "\x00\x80"
-     "\x84\x4A\x00\x80\x88\x4A\x00\x80\x8C\x4A\x00\x80\x90\x4A\x00\x80\x94\x4A"
-     "\x00\x80"
-     "\x98\x4A\x00\x80\x9C\x4A\x00\x80\xA0\x4A\x00\x80\xA4\x4A\x00\x80\xA8\x4A"
-     "\x00\x80"
-     "\xAC\x4A\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\xB0\x4A"
-     "\x00\x80"
-     "\x84\x80\x02\x00\xBE\x88\x03\x00\xEF\x74\x07\x00\x84\x00\x03\x00\x86\xB0"
-     "\x0C\x00"
-     "\x87\x54\x03\x00\xB8\x4A\x00\x80\xBC\x4A\x00\x80\xC0\x4A\x00\x80\xC4\x4A"
-     "\x00\x80"
-     "\xBE\x24\x0C\x00\xC8\x4A\x00\x80\xE1\x48\x06\x00\xCC\x4A\x00\x80\xE3\x08"
-     "\x07\x00"
-     "\xD0\x4A\x00\x80\xD4\x4A\x00\x80\xD8\x4A\x00\x80\xDC\x4A\x00\x80\xB3\x65"
-     "\x03\x00"
-     "\xE0\x4A\x00\x80\xB5\x65\x03\x00\xB6\x6D\x03\x00\xE4\x4A\x00\x80\xE8\x4A"
-     "\x00\x80"
-     "\xEC\x4A\x00\x80\xBA\xED\x03\x00\xBB\xE5\x03\x00\xBC\xE1\x03\x00\xBD\xE1"
-     "\x03\x00"
-     "\xBE\xE5\x03\x00\xBF\xCD\x03\x00\xF0\x4A\x00\x80\xF4\x4A\x00\x80\xF8\x4A"
-     "\x00\x80"
-     "\xFC\x4A\x00\x80\x00\x4B\x00\x80\x04\x4B\x00\x80\x08\x4B\x00\x80\x0C\x4B"
-     "\x00\x80"
-     "\xA8\x81\x03\x00\xA9\x81\x03\x00\xAA\x81\x03\x00\xAB\x81\x03\x00\xAC\x81"
-     "\x03\x00"
-     "\xAD\x81\x03\x00\xAE\x81\x03\x00\xAF\x81\x03\x00\xB0\x45\x03\x00\xB1\x4D"
-     "\x03\x00"
-     "\xB2\x59\x03\x00\xB3\x59\x03\x00\xB4\x4D\x03\x00\xB5\x65\x03\x00\xB6\x19"
-     "\x03\x00"
-     "\xB7\x19\x03\x00\xB8\x29\x03\x00\xB9\x29\x03\x00\xBA\x39\x03\x00\xBB\x39"
-     "\x03\x00"
-     "\xBC\x29\x03\x00\xBD\x29\x03\x00\xBE\x19\x03\x00\xBF\x19\x03\x00\xA3\x29"
-     "\x02\x00"
-     "\x10\x4B\x00\x80\x14\x4B\x00\x80\x18\x4B\x00\x80\x1C\x4B\x00\x80\xA6\x21"
-     "\x02\x00"
-     "\xA5\x29\x02\x00\x20\x4B\x00\x80\xAB\xA9\x02\x00\xAA\xA1\x02\x00\x28\x4B"
-     "\x00\x80"
-     "\x2C\x4B\x00\x80\xAF\x81\x02\x00\xAE\xA9\x02\x00\xAD\xAD\x02\x00\xAC\xAD"
-     "\x02\x00"
-     "\x80\x4D\x00\x00\x81\x55\x00\x00\x82\x55\x00\x00\x30\x4B\x00\x80\xE3\x40"
-     "\x03\x00"
-     "\x34\x4B\x00\x80\xE1\xA0\x01\x00\x38\x4B\x00\x80\xEF\x9C\x03\x00\x3C\x4B"
-     "\x00\x80"
-     "\x86\xC8\x0C\x00\x87\x40\x0D\x00\x40\x4B\x00\x80\x44\x4B\x00\x80\x48\x4B"
-     "\x00\x80"
-     "\x4C\x4B\x00\x80\x50\x4B\x00\x80\x54\x4B\x00\x80\x58\x4B\x00\x80\x5C\x4B"
-     "\x00\x80"
-     "\x60\x4B\x00\x80\x64\x4B\x00\x80\x68\x4B\x00\x80\xEF\xB4\x01\x00\xBE\xBC"
-     "\x0C\x00"
-     "\xE1\x0C\x06\x00\x6C\x4B\x00\x80\xE3\x1C\x06\x00\x70\x4B\x00\x80\x74\x4B"
-     "\x00\x80"
-     "\x78\x4B\x00\x80\x7C\x4B\x00\x80\xB3\xE1\x01\x00\x80\x4B\x00\x80\x84\x4B"
-     "\x00\x80"
-     "\x88\x4B\x00\x80\x8C\x4B\x00\x80\xB6\xE1\x01\x00\xB5\xF5\x01\x00\x90\x4B"
-     "\x00\x80"
-     "\xBB\xA9\x01\x00\xBA\xA1\x01\x00\x84\xE0\x0D\x00\x98\x4B\x00\x80\xBF\x95"
-     "\x01\x00"
-     "\xBE\x95\x01\x00\xBD\xA5\x01\x00\xBC\xAD\x01\x00\xA8\x19\x0E\x00\xA9\x19"
-     "\x0E\x00"
-     "\xAA\x29\x0E\x00\xAB\x29\x0E\x00\xAC\x3D\x0E\x00\xAD\x21\x0E\x00\xAE\x25"
-     "\x0E\x00"
-     "\xAF\x5D\x0E\x00\x24\x4B\x00\x80\x82\xF9\x0F\x00\x81\xF9\x0F\x00\x80\xE1"
-     "\x0F\x00"
-     "\x94\x4B\x00\x80\x9C\x4B\x00\x80\x86\x1C\x00\x00\x87\xA0\x03\x00\xB8\xC9"
-     "\x0E\x00"
-     "\xB9\xC9\x0E\x00\xBA\xD9\x0E\x00\xBB\xD9\x0E\x00\xBC\xC9\x0E\x00\xBD\xC9"
-     "\x0E\x00"
-     "\xBE\x5D\x0F\x00\xBF\x55\x0F\x00\xB0\x25\x0E\x00\xB1\x2D\x0E\x00\xB2\x39"
-     "\x0E\x00"
-     "\xB3\x31\x0E\x00\xB4\x11\x0E\x00\xB5\x11\x0E\x00\xB6\xF9\x0E\x00\xB7\xF9"
-     "\x0E\x00"
-     "\xA3\xA1\x0E\x00\xA0\x4B\x00\x80\xA4\x4B\x00\x80\xA8\x4B\x00\x80\xAC\x4B"
-     "\x00\x80"
-     "\xA6\xA1\x0E\x00\xA5\xB5\x0E\x00\xB0\x4B\x00\x80\xAB\xE9\x0E\x00\xAA\xE1"
-     "\x0E\x00"
-     "\xB4\x4B\x00\x80\xB8\x4B\x00\x80\xAF\xD5\x0E\x00\xAE\xD5\x0E\x00\xAD\xE5"
-     "\x0E\x00"
-     "\xAC\xED\x0E\x00\xBC\x4B\x00\x80\xB3\x89\x0F\x00\xC0\x4B\x00\x80\xC4\x4B"
-     "\x00\x80"
-     "\xB6\xB5\x0F\x00\xC8\x4B\x00\x80\xCC\x4B\x00\x80\xB5\x81\x0F\x00\xBA\x99"
-     "\x0F\x00"
-     "\xBB\x99\x0F\x00\xD0\x4B\x00\x80\xD4\x4B\x00\x80\xBE\x79\x0F\x00\xBF\x79"
-     "\x0F\x00"
-     "\xBC\x79\x0F\x00\xBD\x79\x0F\x00\xA8\x29\x0E\x00\xA9\x29\x0E\x00\xAA\x45"
-     "\x0E\x00"
-     "\xAB\x59\x0E\x00\xAC\x4D\x0E\x00\xAD\x71\x0E\x00\xAE\x69\x0E\x00\xAF\x69"
-     "\x0E\x00"
-     "\xD8\x4B\x00\x80\xDC\x4B\x00\x80\xE0\x4B\x00\x80\xE4\x4B\x00\x80\xE8\x4B"
-     "\x00\x80"
-     "\xEC\x4B\x00\x80\xF0\x4B\x00\x80\xF4\x4B\x00\x80\xB8\xD9\x0E\x00\xB9\xED"
-     "\x0E\x00"
-     "\xBA\xF9\x0E\x00\xBB\xF9\x0E\x00\xBC\xE9\x0E\x00\xBD\xE9\x0E\x00\xBE\x9D"
-     "\x0E\x00"
-     "\xBF\x95\x0E\x00\xB0\x19\x0E\x00\xB1\x19\x0E\x00\xB2\xE9\x0E\x00\xB3\xE9"
-     "\x0E\x00"
-     "\xB4\xF9\x0E\x00\xB5\xF9\x0E\x00\xB6\xE9\x0E\x00\xB7\xE9\x0E\x00\xF8\x4B"
-     "\x00\x80"
-     "\xA6\xE1\x0E\x00\xA5\xD5\x0E\x00\xFC\x4B\x00\x80\xA3\xDD\x0E\x00\x82\x29"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x80\x1D\x00\x00\xAF\x2D\x0E\x00\xAE\x2D\x0E\x00\xAD\x2D"
-     "\x0E\x00"
-     "\xAC\x2D\x0E\x00\xAB\xCD\x0E\x00\xAA\xCD\x0E\x00\xB4\x4A\x00\x80\x00\x4C"
-     "\x00\x80"
-     "\x84\xC8\x02\x00\xB3\x41\x01\x00\xBE\x0C\x01\x00\x08\x4C\x00\x80\xB6\x41"
-     "\x01\x00"
-     "\x0C\x4C\x00\x80\x10\x4C\x00\x80\xB5\x51\x01\x00\xBA\x4D\x01\x00\xBB\x25"
-     "\x01\x00"
-     "\x86\x48\x00\x00\x87\x7C\x00\x00\xBE\x25\x01\x00\xBF\x2D\x01\x00\xBC\x31"
-     "\x01\x00"
-     "\xBD\x31\x01\x00\x14\x4C\x00\x80\xEF\xD4\x0E\x00\x84\x04\x03\x00\xBE\x80"
-     "\x04\x00"
-     "\x18\x4C\x00\x80\x1C\x4C\x00\x80\x20\x4C\x00\x80\x24\x4C\x00\x80\xBF\xA8"
-     "\xCE\x00"
-     "\x28\x4C\x00\x80\x2C\x4C\x00\x80\x84\xE0\x05\x00\x30\x4C\x00\x80\xE3\x28"
-     "\x0E\x00"
-     "\x34\x4C\x00\x80\xE1\x34\x0E\x00\x38\x4C\x00\x80\x3C\x4C\x00\x80\xA3\xCD"
-     "\x02\x00"
-     "\x40\x4C\x00\x80\xA5\xDD\x02\x00\x44\x4C\x00\x80\x48\x4C\x00\x80\xA6\xCD"
-     "\x02\x00"
-     "\x4C\x4C\x00\x80\x50\x4C\x00\x80\xAB\xA9\x02\x00\xAA\xC1\x02\x00\xAD\xBD"
-     "\x02\x00"
-     "\xAC\xBD\x02\x00\xAF\xA1\x02\x00\xAE\xA9\x02\x00\xA9\xB5\x02\x00\xA8\x69"
-     "\x0D\x00"
-     "\xAB\x01\x02\x00\xAA\x09\x02\x00\xAD\x01\x02\x00\xAC\x19\x02\x00\xAF\x31"
-     "\x02\x00"
-     "\xAE\x01\x02\x00\xBE\x68\x05\x00\x54\x4C\x00\x80\x58\x4C\x00\x80\x5C\x4C"
-     "\x00\x80"
-     "\x60\x4C\x00\x80\x64\x4C\x00\x80\x68\x4C\x00\x80\x6C\x4C\x00\x80\xB9\xE5"
-     "\x03\x00"
-     "\xB8\xE5\x03\x00\xBB\xE5\x03\x00\xBA\xE5\x03\x00\xBD\xE5\x03\x00\xBC\xE5"
-     "\x03\x00"
-     "\xBF\xE5\x03\x00\xBE\xE5\x03\x00\xB1\x21\x02\x00\xB0\x49\x02\x00\xB3\x25"
-     "\x02\x00"
-     "\xB2\x25\x02\x00\xB5\x29\x02\x00\xB4\x21\x02\x00\xB7\x15\x02\x00\xB6\x15"
-     "\x02\x00"
-     "\xA8\xCD\x02\x00\xA9\xD1\x02\x00\xAA\xD1\x02\x00\xAB\x0D\x01\x00\xAC\x01"
-     "\x01\x00"
-     "\xAD\x01\x01\x00\xAE\x01\x01\x00\xAF\x01\x01\x00\x70\x4C\x00\x80\x78\x4C"
-     "\x00\x80"
-     "\x7C\x4C\x00\x80\x80\x4C\x00\x80\xBE\x30\x07\x00\x84\x4C\x00\x80\x88\x4C"
-     "\x00\x80"
-     "\x8C\x4C\x00\x80\xB8\x19\x01\x00\xB9\x25\x01\x00\xBA\x2D\x01\x00\xBB\xE9"
-     "\x01\x00"
-     "\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\xE9\x01\x00\xBF\xE9\x01\x00\xB0\x45"
-     "\x01\x00"
-     "\xB1\x4D\x01\x00\xB2\x59\x01\x00\xB3\x55\x01\x00\xB4\x31\x01\x00\xB5\x31"
-     "\x01\x00"
-     "\xB6\x35\x01\x00\xB7\x29\x01\x00\x86\xA0\x05\x00\x87\xD4\x05\x00\x90\x4C"
-     "\x00\x80"
-     "\xEF\xBC\x00\x00\x80\x29\x01\x00\x81\x29\x01\x00\x82\xB9\x00\x00\xEF\xC4"
-     "\x0E\x00"
-     "\x84\x6C\x07\x00\xE1\xD4\x0E\x00\x94\x4C\x00\x80\xE3\x54\x0E\x00\x98\x4C"
-     "\x00\x80"
-     "\xE1\x94\x01\x00\x9C\x4C\x00\x80\xE3\x6C\x00\x00\xB3\x19\x02\x00\xA0\x4C"
-     "\x00\x80"
-     "\xA4\x4C\x00\x80\xA8\x4C\x00\x80\x84\x00\x38\x00\xB6\xC1\x01\x00\xB5\xC9"
-     "\x01\x00"
-     "\xAC\x4C\x00\x80\xBB\xC9\x01\x00\xBA\xC1\x01\x00\xB0\x4C\x00\x80\xB4\x4C"
-     "\x00\x80"
-     "\xBF\xB5\x01\x00\xBE\xB5\x01\x00\xBD\xC1\x01\x00\xBC\xCD\x01\x00\xA3\xD9"
-     "\x05\x00"
-     "\x74\x4C\x00\x80\xB8\x4C\x00\x80\xBC\x4C\x00\x80\xC0\x4C\x00\x80\xA6\x01"
-     "\x06\x00"
-     "\xA5\x09\x06\x00\xC4\x4C\x00\x80\xAB\x09\x06\x00\xAA\x01\x06\x00\xC8\x4C"
-     "\x00\x80"
-     "\xCC\x4C\x00\x80\xAF\x75\x06\x00\xAE\x75\x06\x00\xAD\x01\x06\x00\xAC\x0D"
-     "\x06\x00"
-     "\xD0\x4C\x00\x80\xD4\x4C\x00\x80\xD8\x4C\x00\x80\xDC\x4C\x00\x80\x80\x19"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x82\x05\x00\x00\xE0\x4C\x00\x80\xE4\x4C\x00\x80\xE8\x4C"
-     "\x00\x80"
-     "\x86\x28\x03\x00\x87\xAC\x03\x00\xEC\x4C\x00\x80\xF0\x4C\x00\x80\xF4\x4C"
-     "\x00\x80"
-     "\xF8\x4C\x00\x80\xA8\xA5\x07\x00\xA9\xAD\x07\x00\xAA\xA5\x07\x00\xAB\xB9"
-     "\x07\x00"
-     "\xAC\xA9\x07\x00\xAD\xA9\x07\x00\xAE\x11\x07\x00\xAF\x35\x07\x00\xFC\x4C"
-     "\x00\x80"
-     "\x00\x4D\x00\x80\x04\x4D\x00\x80\x08\x4D\x00\x80\x0C\x4D\x00\x80\x10\x4D"
-     "\x00\x80"
-     "\x14\x4D\x00\x80\x18\x4D\x00\x80\xB8\xC9\x00\x00\xB9\xC9\x00\x00\xBA\xD9"
-     "\x00\x00"
-     "\xBB\xD1\x00\x00\xBC\xF9\x00\x00\xBD\xF9\x00\x00\xBE\x99\x00\x00\xBF\x99"
-     "\x00\x00"
-     "\xB0\x51\x07\x00\xB1\x5D\x07\x00\xB2\x2D\x07\x00\xB3\x25\x07\x00\xB4\x3D"
-     "\x07\x00"
-     "\xB5\x29\x07\x00\xB6\x11\x07\x00\xB7\x11\x07\x00\xB3\x2D\x06\x00\x1C\x4D"
-     "\x00\x80"
-     "\x20\x4D\x00\x80\x24\x4D\x00\x80\x28\x4D\x00\x80\xB6\x29\x06\x00\xB5\x21"
-     "\x06\x00"
-     "\x2C\x4D\x00\x80\xBB\x99\x06\x00\xBA\x95\x06\x00\x30\x4D\x00\x80\x34\x4D"
-     "\x00\x80"
-     "\xBF\xED\x06\x00\xBE\xF9\x06\x00\xBD\xF1\x06\x00\xBC\xF9\x06\x00\x38\x4D"
-     "\x00\x80"
-     "\xA3\x69\x06\x00\x3C\x4D\x00\x80\x40\x4D\x00\x80\xA6\x6D\x06\x00\x44\x4D"
-     "\x00\x80"
-     "\x48\x4D\x00\x80\xA5\x65\x06\x00\xAA\xD1\x06\x00\xAB\xDD\x06\x00\x4C\x4D"
-     "\x00\x80"
-     "\x50\x4D\x00\x80\xAE\xBD\x06\x00\xAF\xA9\x06\x00\xAC\xBD\x06\x00\xAD\xB5"
-     "\x06\x00"
-     "\xB0\x2D\x01\x00\xB1\x35\x01\x00\xB2\x3D\x01\x00\xB3\x35\x01\x00\xB4\x2D"
-     "\x01\x00"
-     "\xB5\x91\x00\x00\xB6\x91\x00\x00\xB7\x91\x00\x00\xB8\xB1\x00\x00\xB9\xB9"
-     "\x00\x00"
-     "\xBA\x81\x00\x00\xBB\x81\x00\x00\xBC\xF1\x00\x00\xBD\xFD\x00\x00\xBE\xF5"
-     "\x00\x00"
-     "\xBF\xE9\x00\x00\x80\x99\x00\x00\x81\xAD\x00\x00\x82\xA5\x00\x00\x04\x4C"
-     "\x00\x80"
-     "\x54\x4D\x00\x80\x58\x4D\x00\x80\x87\x18\x01\x00\x86\x7C\x00\x00\xA8\x85"
-     "\x01\x00"
-     "\xA9\x95\x01\x00\xAA\x9D\x01\x00\xAB\x95\x01\x00\xAC\x8D\x01\x00\xAD\x55"
-     "\x01\x00"
-     "\xAE\x5D\x01\x00\xAF\x55\x01\x00\xB3\x75\x02\x00\x5C\x4D\x00\x80\x60\x4D"
-     "\x00\x80"
-     "\x64\x4D\x00\x80\x68\x4D\x00\x80\xB6\xB5\x02\x00\xB5\x65\x02\x00\x6C\x4D"
-     "\x00\x80"
-     "\xBB\x91\x02\x00\xBA\x89\x02\x00\x70\x4D\x00\x80\x74\x4D\x00\x80\xBF\x35"
-     "\x03\x00"
-     "\xBE\x89\x02\x00\xBD\x81\x02\x00\xBC\x89\x02\x00\xE3\x58\x06\x00\x78\x4D"
-     "\x00\x80"
-     "\xE1\xB8\x05\x00\x7C\x4D\x00\x80\x91\x00\x00\x00\xE1\x90\x01\x00\x80\x4D"
-     "\x00\x80"
-     "\xE3\x9C\x07\x00\x84\x4D\x00\x80\x88\x4D\x00\x80\x8C\x4D\x00\x80\xEF\x4C"
-     "\x02\x00"
-     "\xEF\x84\x06\x00\x90\x4D\x00\x80\x94\x4D\x00\x80\x98\x4D\x00\x80\x9C\x4D"
-     "\x00\x80"
-     "\xA0\x4D\x00\x80\xA4\x4D\x00\x80\xA3\xF9\x02\x00\xA8\x4D\x00\x80\xA5\xE9"
-     "\x02\x00"
-     "\xA6\x39\x02\x00\xAC\x4D\x00\x80\xB0\x4D\x00\x80\xB4\x4D\x00\x80\xAA\x05"
-     "\x02\x00"
-     "\xAB\x1D\x02\x00\xAC\x05\x02\x00\xAD\x0D\x02\x00\xAE\x05\x02\x00\xAF\xB9"
-     "\x03\x00"
-     "\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\xB8\x4D\x00\x80\xBC\x4D"
-     "\x00\x80"
-     "\xC0\x4D\x00\x80\xBE\x54\x3C\x00\xEF\x9C\x1F\x00\x85\xA4\x3C\x00\xE1\x10"
-     "\x02\x00"
-     "\xC8\x4D\x00\x80\xE3\x48\x3F\x00\x86\x50\x3E\x00\x87\x34\x03\x00\xCC\x4D"
-     "\x00\x80"
-     "\xD0\x4D\x00\x80\xA8\xD5\x02\x00\xA9\x01\x01\x00\xAA\x01\x01\x00\xAB\x01"
-     "\x01\x00"
-     "\xAC\x01\x01\x00\xAD\x09\x01\x00\xAE\x31\x01\x00\xAF\x31\x01\x00\xD4\x4D"
-     "\x00\x80"
-     "\xD8\x4D\x00\x80\xDC\x4D\x00\x80\xE0\x4D\x00\x80\xE4\x4D\x00\x80\xE8\x4D"
-     "\x00\x80"
-     "\xEC\x4D\x00\x80\xF0\x4D\x00\x80\xB8\x3D\x01\x00\xB9\xC5\x01\x00\xBA\xCD"
-     "\x01\x00"
-     "\xBB\xC5\x01\x00\xBC\xC1\x01\x00\xBD\xC9\x01\x00\xBE\xF9\x01\x00\xBF\xF9"
-     "\x01\x00"
-     "\xB0\x25\x01\x00\xB1\x29\x01\x00\xB2\x31\x01\x00\xB3\x0D\x01\x00\xB4\x15"
-     "\x01\x00"
-     "\xB5\x1D\x01\x00\xB6\x0D\x01\x00\xB7\x05\x01\x00\xE1\x18\x06\x00\xF4\x4D"
-     "\x00\x80"
-     "\xE3\xB4\x07\x00\xF8\x4D\x00\x80\xFC\x4D\x00\x80\x00\x4E\x00\x80\x04\x4E"
-     "\x00\x80"
-     "\x08\x4E\x00\x80\x0C\x4E\x00\x80\x10\x4E\x00\x80\x14\x4E\x00\x80\x18\x4E"
-     "\x00\x80"
-     "\x1C\x4E\x00\x80\x20\x4E\x00\x80\xEF\xE0\x07\x00\x24\x4E\x00\x80\x80\x11"
-     "\x00\x00"
-     "\x81\x11\x00\x00\x82\x11\x00\x00\x84\x40\x3C\x00\x28\x4E\x00\x80\x2C\x4E"
-     "\x00\x80"
-     "\x30\x4E\x00\x80\xEF\x14\x06\x00\xBE\xEC\x3F\x00\xE1\x94\x01\x00\x38\x4E"
-     "\x00\x80"
-     "\xE3\x5C\x06\x00\x86\x50\x3E\x00\x87\x00\x3C\x00\x3C\x4E\x00\x80\x40\x4E"
-     "\x00\x80"
-     "\xA3\x09\x3E\x00\x8F\x25\x24\x00\xC4\x4D\x00\x80\x84\x08\x00\x00\x34\x4E"
-     "\x00\x80"
-     "\xA6\x09\x3E\x00\xA5\x09\x3E\x00\x44\x4E\x00\x80\xAB\x6D\x3E\x00\xAA\x61"
-     "\x3E\x00"
-     "\x48\x4E\x00\x80\x4C\x4E\x00\x80\xAF\x59\x3E\x00\xAE\x51\x3E\x00\xAD\x59"
-     "\x3E\x00"
-     "\xAC\x65\x3E\x00\x9E\x61\x38\x00\x9F\x75\x38\x00\x9C\x95\x34\x00\x9D\xF5"
-     "\x39\x00"
-     "\x9A\xE5\x35\x00\x9B\x75\x34\x00\x98\x79\x30\x00\x99\x71\x31\x00\x96\xA1"
-     "\x2C\x00"
-     "\x97\xC5\x31\x00\x94\x6D\x2C\x00\x95\x65\x2C\x00\x92\x21\x28\x00\x93\x09"
-     "\x29\x00"
-     "\x90\xF5\x29\x00\x91\x79\x28\x00\xB1\x0D\x14\x00\xB0\x05\x14\x00\xB3\x01"
-     "\x18\x00"
-     "\xB2\x55\x14\x00\xB5\x79\x18\x00\xB4\xB5\x18\x00\x50\x4E\x00\x80\x54\x4E"
-     "\x00\x80"
-     "\xBE\xE0\x01\x00\x58\x4E\x00\x80\x5C\x4E\x00\x80\x60\x4E\x00\x80\xA2\xD1"
-     "\x3C\x00"
-     "\xA3\x65\x3D\x00\xA0\x75\x3D\x00\xA1\xB9\x3C\x00\xA1\xC9\x00\x00\x64\x4E"
-     "\x00\x80"
-     "\xA3\x01\x04\x00\xA2\xDD\x00\x00\xA5\x7D\x04\x00\xA4\xF5\x04\x00\xA7\xF9"
-     "\x08\x00"
-     "\xA6\x01\x08\x00\xA9\x01\x0C\x00\xA8\x6D\x08\x00\xAB\xCD\x0C\x00\xAA\xCD"
-     "\x0C\x00"
-     "\xAD\xD1\x10\x00\xAC\x01\x10\x00\xAF\xD5\x11\x00\xAE\xD9\x10\x00\x68\x4E"
-     "\x00\x80"
-     "\xB5\x15\x3F\x00\xB6\x15\x3F\x00\x6C\x4E\x00\x80\x70\x4E\x00\x80\x74\x4E"
-     "\x00\x80"
-     "\x78\x4E\x00\x80\xB3\x15\x3F\x00\xBC\x79\x3F\x00\xBD\x45\x3F\x00\xBE\x4D"
-     "\x3F\x00"
-     "\xBF\x45\x3F\x00\x7C\x4E\x00\x80\x80\x4E\x00\x80\xBA\x7D\x3F\x00\xBB\x71"
-     "\x3F\x00"
-     "\xB3\xBD\x3F\x00\x84\x4E\x00\x80\x88\x4E\x00\x80\x8C\x4E\x00\x80\x90\x4E"
-     "\x00\x80"
-     "\xB6\x7D\x3E\x00\xB5\x7D\x3E\x00\x94\x4E\x00\x80\xBB\x59\x3E\x00\xBA\x51"
-     "\x3E\x00"
-     "\x98\x4E\x00\x80\x9C\x4E\x00\x80\xBF\xE1\x3E\x00\xBE\xF9\x3E\x00\xBD\xFD"
-     "\x3E\x00"
-     "\xBC\x41\x3E\x00\x82\xA1\x25\x00\x83\x09\x28\x00\xA0\x4E\x00\x80\xA8\x4E"
-     "\x00\x80"
-     "\x86\xED\x28\x00\x87\x91\x2D\x00\x84\x9D\x29\x00\x85\xCD\x29\x00\x8A\xA9"
-     "\x2D\x00"
-     "\x8B\x65\x2D\x00\x86\x08\x07\x00\x87\x8C\x00\x00\x8E\x9D\x30\x00\x8F\x01"
-     "\x34\x00"
-     "\x8C\x19\x30\x00\x8D\x9D\x31\x00\x92\xE1\x35\x00\x93\x19\x35\x00\xAC\x4E"
-     "\x00\x80"
-     "\xB0\x4E\x00\x80\x96\xE1\x39\x00\x97\x19\x3C\x00\x94\x61\x38\x00\x95\xFD"
-     "\x39\x00"
-     "\x9A\xE5\x3D\x00\x9B\x85\x3D\x00\xB4\x4E\x00\x80\xB8\x4E\x00\x80\x81\x9D"
-     "\x00\x00"
-     "\x80\x9D\x00\x00\x9C\x29\x3C\x00\x82\x75\x00\x00\xA8\xE9\x3C\x00\xA9\xFD"
-     "\x3C\x00"
-     "\xAA\xF5\x3C\x00\xAB\x2D\x3D\x00\xAC\x35\x3D\x00\xAD\x3D\x3D\x00\xAE\x35"
-     "\x3D\x00"
-     "\xAF\x2D\x3D\x00\xBC\x4E\x00\x80\xC0\x4E\x00\x80\xC4\x4E\x00\x80\xC8\x4E"
-     "\x00\x80"
-     "\xCC\x4E\x00\x80\xD0\x4E\x00\x80\xBE\x70\x04\x00\xD4\x4E\x00\x80\xB8\xDD"
-     "\x3D\x00"
-     "\xB9\xE5\x3D\x00\xBA\xED\x3D\x00\xBB\xE5\x3D\x00\xBC\xFD\x3D\x00\xBD\xE5"
-     "\x3D\x00"
-     "\xBE\xED\x3D\x00\xBF\xE1\x3D\x00\xB0\x55\x3D\x00\xB1\x5D\x3D\x00\xB2\x55"
-     "\x3D\x00"
-     "\xB3\xE9\x3D\x00\xB4\xFD\x3D\x00\xB5\xE5\x3D\x00\xB6\xED\x3D\x00\xB7\xE5"
-     "\x3D\x00"
-     "\xA3\x7D\x3C\x00\xD8\x4E\x00\x80\xDC\x4E\x00\x80\xBE\xC4\x05\x00\xE0\x4E"
-     "\x00\x80"
-     "\xA6\xBD\x3D\x00\xA5\xBD\x3D\x00\xE4\x4E\x00\x80\xAB\x99\x3D\x00\xAA\x91"
-     "\x3D\x00"
-     "\xE8\x4E\x00\x80\xEC\x4E\x00\x80\xAF\x21\x3D\x00\xAE\x39\x3D\x00\xAD\x3D"
-     "\x3D\x00"
-     "\xAC\x81\x3D\x00\xF0\x4E\x00\x80\xF4\x4E\x00\x80\xB3\x35\x3D\x00\xF8\x4E"
-     "\x00\x80"
-     "\xB5\xF9\x02\x00\xFC\x4E\x00\x80\x00\x4F\x00\x80\xB6\xF1\x02\x00\x04\x4F"
-     "\x00\x80"
-     "\x08\x4F\x00\x80\xBB\x41\x02\x00\xBA\x49\x02\x00\xBD\x35\x02\x00\xBC\x59"
-     "\x02\x00"
-     "\xBF\xF5\x03\x00\xBE\xFD\x03\x00\xA8\xE9\x3C\x00\xA9\xFD\x3C\x00\xAA\xF5"
-     "\x3C\x00"
-     "\xAB\x11\x03\x00\xAC\x39\x03\x00\xAD\x39\x03\x00\xAE\x81\x03\x00\xAF\x81"
-     "\x03\x00"
-     "\x80\xC5\x01\x00\x81\x09\x00\x00\x82\x19\x00\x00\x0C\x4F\x00\x80\x10\x4F"
-     "\x00\x80"
-     "\x84\x78\x05\x00\x18\x4F\x00\x80\x1C\x4F\x00\x80\xB8\x59\x02\x00\xB9\x75"
-     "\x02\x00"
-     "\xBA\x7D\x02\x00\xBB\x09\x02\x00\xBC\x19\x02\x00\xBD\x19\x02\x00\xBE\x09"
-     "\x02\x00"
-     "\xBF\x01\x02\x00\xB0\xC5\x03\x00\xB1\xCD\x03\x00\xB2\xC5\x03\x00\xB3\xA9"
-     "\x03\x00"
-     "\xB4\x71\x02\x00\xB5\x79\x02\x00\xB6\x69\x02\x00\xB7\x69\x02\x00\x20\x4F"
-     "\x00\x80"
-     "\xE3\x58\x3E\x00\x84\xC4\x04\x00\xE1\x64\x3F\x00\x24\x4F\x00\x80\xE1\x7C"
-     "\x3E\x00"
-     "\x28\x4F\x00\x80\xE3\x80\x01\x00\x86\x00\x04\x00\x87\x84\x04\x00\x2C\x4F"
-     "\x00\x80"
-     "\xEF\xF4\x01\x00\x85\x14\x1A\x00\xEF\xAC\x3E\x00\x30\x4F\x00\x80\x34\x4F"
-     "\x00\x80"
-     "\x38\x4F\x00\x80\x3C\x4F\x00\x80\x40\x4F\x00\x80\xA3\x31\x3C\x00\x84\x90"
-     "\x07\x00"
-     "\xA5\xFD\x03\x00\xA6\xF5\x03\x00\x44\x4F\x00\x80\x48\x4F\x00\x80\x4C\x4F"
-     "\x00\x80"
-     "\xAA\x4D\x03\x00\xAB\x45\x03\x00\xAC\x5D\x03\x00\xAD\x31\x03\x00\xAE\xF9"
-     "\x02\x00"
-     "\xAF\xF1\x02\x00\xA8\x8D\x05\x00\xA9\xCD\x05\x00\xAA\xC5\x05\x00\xAB\xE5"
-     "\x05\x00"
-     "\xAC\xFD\x05\x00\xAD\x35\x06\x00\xAE\x3D\x06\x00\xAF\x35\x06\x00\x14\x4F"
-     "\x00\x80"
-     "\x50\x4F\x00\x80\x54\x4F\x00\x80\x58\x4F\x00\x80\x5C\x4F\x00\x80\x60\x4F"
-     "\x00\x80"
-     "\x64\x4F\x00\x80\x68\x4F\x00\x80\xB8\xE9\x06\x00\xB9\xE9\x06\x00\xBA\x89"
-     "\x06\x00"
-     "\xBB\x89\x06\x00\xBC\x9D\x06\x00\xBD\x81\x06\x00\xBE\x81\x06\x00\xBF\xBD"
-     "\x06\x00"
-     "\xB0\x4D\x06\x00\xB1\x55\x06\x00\xB2\x5D\x06\x00\xB3\x51\x06\x00\xB4\xF5"
-     "\x06\x00"
-     "\xB5\xFD\x06\x00\xB6\xE9\x06\x00\xB7\xE1\x06\x00\xB3\x4D\x05\x00\x6C\x4F"
-     "\x00\x80"
-     "\x70\x4F\x00\x80\x74\x4F\x00\x80\x78\x4F\x00\x80\xB6\xB5\x06\x00\xB5\x75"
-     "\x05\x00"
-     "\x7C\x4F\x00\x80\xBB\x91\x06\x00\xBA\x89\x06\x00\x80\x4F\x00\x80\x84\x4F"
-     "\x00\x80"
-     "\xBF\x39\x06\x00\xBE\x31\x06\x00\xBD\x85\x06\x00\xBC\x89\x06\x00\x88\x4F"
-     "\x00\x80"
-     "\xA3\x09\x05\x00\xA4\x4E\x00\x80\x8C\x4F\x00\x80\xA6\xF1\x06\x00\x90\x4F"
-     "\x00\x80"
-     "\x94\x4F\x00\x80\xA5\x31\x05\x00\xAA\xCD\x06\x00\xAB\xD5\x06\x00\x86\x48"
-     "\x00\x00"
-     "\x87\x6C\x00\x00\xAE\x75\x06\x00\xAF\x7D\x06\x00\xAC\xCD\x06\x00\xAD\xC1"
-     "\x06\x00"
-     "\xA8\x39\x06\x00\xA9\x39\x06\x00\xAA\x89\x06\x00\xAB\x9D\x06\x00\xAC\x89"
-     "\x06\x00"
-     "\xAD\xBD\x06\x00\xAE\xA1\x06\x00\xAF\xA1\x06\x00\x80\xB1\x01\x00\x81\xB9"
-     "\x01\x00"
-     "\x82\x81\x01\x00\x98\x4F\x00\x80\x9C\x4F\x00\x80\xA0\x4F\x00\x80\xA4\x4F"
-     "\x00\x80"
-     "\xA8\x4F\x00\x80\xB8\x6D\x01\x00\xB9\x75\x01\x00\xBA\x7D\x01\x00\xBB\x75"
-     "\x01\x00"
-     "\xBC\x6D\x01\x00\xBD\xD1\x01\x00\xBE\xD1\x01\x00\xBF\xD1\x01\x00\xB0\xA1"
-     "\x06\x00"
-     "\xB1\xA1\x06\x00\xB2\xA1\x06\x00\xB3\xB1\x06\x00\xB4\x91\x06\x00\xB5\x91"
-     "\x06\x00"
-     "\xB6\x5D\x01\x00\xB7\x55\x01\x00\xB3\x01\x06\x00\xAC\x4F\x00\x80\xB0\x4F"
-     "\x00\x80"
-     "\xB4\x4F\x00\x80\xB8\x4F\x00\x80\xB6\x3D\x06\x00\xB5\x01\x06\x00\xBC\x4F"
-     "\x00\x80"
-     "\xBB\x65\x06\x00\xBA\x65\x06\x00\xC0\x4F\x00\x80\xC4\x4F\x00\x80\xBF\x55"
-     "\x06\x00"
-     "\xBE\x55\x06\x00\xBD\x65\x06\x00\xBC\x75\x06\x00\xC8\x4F\x00\x80\xA3\x45"
-     "\x06\x00"
-     "\xCC\x4F\x00\x80\xD0\x4F\x00\x80\xA6\x79\x06\x00\xD4\x4F\x00\x80\xD8\x4F"
-     "\x00\x80"
-     "\xA5\x45\x06\x00\xAA\x21\x06\x00\xAB\x21\x06\x00\xDC\x4F\x00\x80\xE0\x4F"
-     "\x00\x80"
-     "\xAE\x11\x06\x00\xAF\x11\x06\x00\xAC\x31\x06\x00\xAD\x21\x06\x00\xA8\xC1"
-     "\x02\x00"
-     "\xA9\xC9\x02\x00\xAA\xDD\x02\x00\xAB\xD5\x02\x00\xAC\x3D\x03\x00\xAD\x25"
-     "\x03\x00"
-     "\xAE\x2D\x03\x00\xAF\x25\x03\x00\xE4\x4F\x00\x80\xE8\x4F\x00\x80\xEC\x4F"
-     "\x00\x80"
-     "\xF0\x4F\x00\x80\xF8\x4F\x00\x80\xFC\x4F\x00\x80\x00\x50\x00\x80\xBE\x70"
-     "\x02\x00"
-     "\xB8\x81\x03\x00\xB9\x8D\x03\x00\xBA\x85\x03\x00\xBB\x99\x03\x00\xBC\x89"
-     "\x03\x00"
-     "\xBD\xBD\x03\x00\xBE\xB5\x03\x00\xBF\x7D\x00\x00\xB0\x5D\x03\x00\xB1\xE1"
-     "\x03\x00"
-     "\xB2\xE1\x03\x00\xB3\xE1\x03\x00\xB4\xE1\x03\x00\xB5\xE9\x03\x00\xB6\xD1"
-     "\x03\x00"
-     "\xB7\xD1\x03\x00\x80\xA5\x00\x00\x81\xAD\x00\x00\x82\xA5\x00\x00\xB3\x59"
-     "\x02\x00"
-     "\x04\x50\x00\x80\xB5\xB9\x03\x00\xB6\xA9\x03\x00\x08\x50\x00\x80\x86\x00"
-     "\x0C\x00"
-     "\x87\xF4\x03\x00\xBA\x95\x03\x00\xBB\x29\x03\x00\xBC\x39\x03\x00\xBD\x39"
-     "\x03\x00"
-     "\xBE\x29\x03\x00\xBF\x21\x03\x00\xA3\x15\x02\x00\x0C\x50\x00\x80\x10\x50"
-     "\x00\x80"
-     "\x14\x50\x00\x80\x18\x50\x00\x80\xA6\xE5\x03\x00\xA5\xF5\x03\x00\x1C\x50"
-     "\x00\x80"
-     "\xAB\x65\x03\x00\xAA\xD9\x03\x00\x20\x50\x00\x80\x24\x50\x00\x80\xAF\x6D"
-     "\x03\x00"
-     "\xAE\x65\x03\x00\xAD\x75\x03\x00\xAC\x75\x03\x00\x28\x50\x00\x80\x2C\x50"
-     "\x00\x80"
-     "\x30\x50\x00\x80\x34\x50\x00\x80\xEF\xD4\x03\x00\x38\x50\x00\x80\x3C\x50"
-     "\x00\x80"
-     "\x40\x50\x00\x80\xE3\xE0\x06\x00\x44\x50\x00\x80\xE1\xD8\x01\x00\x48\x50"
-     "\x00\x80"
-     "\x4C\x50\x00\x80\x50\x50\x00\x80\x54\x50\x00\x80\x58\x50\x00\x80\x5C\x50"
-     "\x00\x80"
-     "\xE3\x60\x06\x00\x84\x44\x02\x00\xE1\xFC\x04\x00\x80\x6D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x60\x50\x00\x80\xBE\x6C\x0D\x00\x68\x50\x00\x80\x86\xC8"
-     "\x0C\x00"
-     "\x87\x5C\x0D\x00\x6C\x50\x00\x80\xEF\xB4\x07\x00\x70\x50\x00\x80\x74\x50"
-     "\x00\x80"
-     "\x78\x50\x00\x80\x7C\x50\x00\x80\x80\x50\x00\x80\x84\x50\x00\x80\x88\x50"
-     "\x00\x80"
-     "\x8C\x50\x00\x80\x90\x50\x00\x80\xEF\xE8\x07\x00\x84\xA0\x0D\x00\xE1\x1C"
-     "\x06\x00"
-     "\x94\x50\x00\x80\xE3\x34\x06\x00\x98\x50\x00\x80\x9C\x50\x00\x80\xA0\x50"
-     "\x00\x80"
-     "\xA4\x50\x00\x80\xB3\xF9\x01\x00\xA8\x50\x00\x80\xAC\x50\x00\x80\xB0\x50"
-     "\x00\x80"
-     "\xB4\x50\x00\x80\xB6\x11\x01\x00\xB5\x45\x01\x00\xB8\x50\x00\x80\xBB\x3D"
-     "\x01\x00"
-     "\xBA\x35\x01\x00\xBC\x50\x00\x80\xC0\x50\x00\x80\xBF\xF5\x00\x00\xBE\xF5"
-     "\x00\x00"
-     "\xBD\x15\x01\x00\xBC\x15\x01\x00\xA8\x31\x0E\x00\xA9\x31\x0E\x00\xAA\x31"
-     "\x0E\x00"
-     "\xAB\x9D\x0E\x00\xAC\x8D\x0E\x00\xAD\xD1\x0E\x00\xAE\xD1\x0E\x00\xAF\xD1"
-     "\x0E\x00"
-     "\x64\x50\x00\x80\xC4\x50\x00\x80\xC8\x50\x00\x80\xCC\x50\x00\x80\x80\x19"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x82\x05\x00\x00\xD0\x50\x00\x80\xB8\x55\x0F\x00\xB9\x5D"
-     "\x0F\x00"
-     "\xBA\x55\x0F\x00\xBB\x69\x0F\x00\xBC\x7D\x0F\x00\xBD\x65\x0F\x00\xBE\x6D"
-     "\x0F\x00"
-     "\xBF\x65\x0F\x00\xB0\xB1\x0E\x00\xB1\xB9\x0E\x00\xB2\x81\x0E\x00\xB3\x81"
-     "\x0E\x00"
-     "\xB4\x71\x0F\x00\xB5\x71\x0F\x00\xB6\x71\x0F\x00\xB7\x71\x0F\x00\xA3\xB9"
-     "\x0E\x00"
-     "\xBE\x88\x03\x00\xD4\x50\x00\x80\xD8\x50\x00\x80\xDC\x50\x00\x80\xA6\x51"
-     "\x0E\x00"
-     "\xA5\x05\x0E\x00\xE0\x50\x00\x80\xAB\x7D\x0E\x00\xAA\x75\x0E\x00\x86\x08"
-     "\x00\x00"
-     "\x87\x6C\x03\x00\xAF\xB5\x0F\x00\xAE\xB5\x0F\x00\xAD\x55\x0E\x00\xAC\x55"
-     "\x0E\x00"
-     "\xE4\x50\x00\x80\xE8\x50\x00\x80\xEC\x50\x00\x80\xB3\xB5\x0F\x00\xF0\x50"
-     "\x00\x80"
-     "\xB5\x55\x0F\x00\xB6\x5D\x0F\x00\xF4\x4F\x00\x80\xF4\x50\x00\x80\xF8\x50"
-     "\x00\x80"
-     "\xBA\x79\x0F\x00\xBB\x79\x0F\x00\xBC\x69\x0F\x00\xBD\x5D\x0F\x00\xBE\x49"
-     "\x0F\x00"
-     "\xBF\x49\x0F\x00\xA8\x65\x0E\x00\xA9\x75\x0E\x00\xAA\x71\x0E\x00\xAB\x71"
-     "\x0E\x00"
-     "\xAC\x91\x0E\x00\xAD\x91\x0E\x00\xAE\x91\x0E\x00\xAF\x91\x0E\x00\xFC\x50"
-     "\x00\x80"
-     "\x00\x51\x00\x80\x04\x51\x00\x80\x08\x51\x00\x80\x0C\x51\x00\x80\x10\x51"
-     "\x00\x80"
-     "\x14\x51\x00\x80\x18\x51\x00\x80\xB8\x85\x0E\x00\xB9\x8D\x0E\x00\xBA\x85"
-     "\x0E\x00"
-     "\xBB\x9D\x0E\x00\xBC\x8D\x0E\x00\xBD\xBD\x0E\x00\xBE\xB5\x0E\x00\xBF\x79"
-     "\x01\x00"
-     "\xB0\xF1\x0E\x00\xB1\xF1\x0E\x00\xB2\xF1\x0E\x00\xB3\xC5\x0E\x00\xB4\xC1"
-     "\x0E\x00"
-     "\xB5\xC1\x0E\x00\xB6\xC1\x0E\x00\xB7\xC1\x0E\x00\xA3\xF9\x0E\x00\x1C\x51"
-     "\x00\x80"
-     "\x20\x51\x00\x80\x24\x51\x00\x80\x28\x51\x00\x80\xA6\x11\x0E\x00\xA5\x19"
-     "\x0E\x00"
-     "\x2C\x51\x00\x80\xAB\x35\x0E\x00\xAA\x35\x0E\x00\x30\x51\x00\x80\x34\x51"
-     "\x00\x80"
-     "\xAF\x05\x0E\x00\xAE\x05\x0E\x00\xAD\x11\x0E\x00\xAC\x25\x0E\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x1D\x00\x00\x38\x51\x00\x80\x3C\x51\x00\x80\x40\x51"
-     "\x00\x80"
-     "\x84\x94\x01\x00\xBE\x94\x01\x00\x86\x40\x07\x00\x87\xF4\x00\x00\x48\x51"
-     "\x00\x80"
-     "\x4C\x51\x00\x80\x50\x51\x00\x80\x54\x51\x00\x80\x58\x51\x00\x80\x5C\x51"
-     "\x00\x80"
-     "\xA8\x8D\x02\x00\xA9\x95\x02\x00\xAA\x95\x02\x00\xAB\xCD\x02\x00\xAC\xD5"
-     "\x02\x00"
-     "\xAD\xDD\x02\x00\xAE\xC9\x02\x00\xAF\xFD\x02\x00\x60\x51\x00\x80\x64\x51"
-     "\x00\x80"
-     "\x68\x51\x00\x80\x6C\x51\x00\x80\x8C\x00\x00\x00\x70\x51\x00\x80\x74\x51"
-     "\x00\x80"
-     "\x78\x51\x00\x80\xB8\x7D\x03\x00\xB9\xC1\x03\x00\xBA\xC1\x03\x00\xBB\xC1"
-     "\x03\x00"
-     "\xBC\xC1\x03\x00\xBD\xC9\x03\x00\xBE\xF1\x03\x00\xBF\xF1\x03\x00\xB0\x85"
-     "\x02\x00"
-     "\xB1\x45\x03\x00\xB2\x4D\x03\x00\xB3\x45\x03\x00\xB4\x5D\x03\x00\xB5\x45"
-     "\x03\x00"
-     "\xB6\x4D\x03\x00\xB7\x45\x03\x00\xB3\x1D\x02\x00\x7C\x51\x00\x80\x80\x51"
-     "\x00\x80"
-     "\x84\x51\x00\x80\x88\x51\x00\x80\xB6\x5D\x02\x00\xB5\x5D\x02\x00\x8C\x51"
-     "\x00\x80"
-     "\xBB\x81\x03\x00\xBA\x41\x02\x00\x90\x51\x00\x80\x94\x51\x00\x80\xBF\x81"
-     "\x03\x00"
-     "\xBE\x99\x03\x00\xBD\x91\x03\x00\xBC\x99\x03\x00\x98\x51\x00\x80\xA3\x59"
-     "\x02\x00"
-     "\x9C\x51\x00\x80\xA0\x51\x00\x80\xA6\x19\x02\x00\xA4\x51\x00\x80\xA8\x51"
-     "\x00\x80"
-     "\xA5\x19\x02\x00\xAA\x05\x02\x00\xAB\xC5\x03\x00\xAC\x51\x00\x80\xB0\x51"
-     "\x00\x80"
-     "\xAE\xDD\x03\x00\xAF\xC5\x03\x00\xAC\xDD\x03\x00\xAD\xD5\x03\x00\xB8\x51"
-     "\x00\x80"
-     "\xE3\xCC\x00\x00\x84\x04\x02\x00\xE1\xBC\x01\x00\x80\xC9\x01\x00\x81\xFD"
-     "\x01\x00"
-     "\x82\xF5\x01\x00\xBE\x10\x05\x00\xBC\x51\x00\x80\xBE\x28\x04\x00\xC0\x51"
-     "\x00\x80"
-     "\xC4\x51\x00\x80\xC8\x51\x00\x80\xEF\x10\x00\x00\xCC\x51\x00\x80\xD0\x51"
-     "\x00\x80"
-     "\x86\xE0\x04\x00\x87\xF4\x02\x00\xD4\x51\x00\x80\xD8\x51\x00\x80\xE3\x08"
-     "\x0E\x00"
-     "\xDC\x51\x00\x80\xE1\xE0\x0F\x00\xE0\x51\x00\x80\xEF\xF8\x0E\x00\xE4\x51"
-     "\x00\x80"
-     "\xE8\x51\x00\x80\xEC\x51\x00\x80\xF0\x51\x00\x80\xF4\x51\x00\x80\xF8\x51"
-     "\x00\x80"
-     "\xFC\x51\x00\x80\x00\x52\x00\x80\x04\x52\x00\x80\x08\x52\x00\x80\x0C\x52"
-     "\x00\x80"
-     "\x10\x52\x00\x80\x14\x52\x00\x80\x18\x52\x00\x80\xEF\xE0\x01\x00\x85\x24"
-     "\x15\x00"
-     "\xE1\xD4\x0E\x00\x1C\x52\x00\x80\xE3\x1C\x0E\x00\x80\x29\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x20\x52\x00\x80\xB3\x31\x02\x00\x28\x52\x00\x80\x84\xCC"
-     "\x05\x00"
-     "\x2C\x52\x00\x80\x30\x52\x00\x80\xB6\x29\x02\x00\xB5\x21\x02\x00\x34\x52"
-     "\x00\x80"
-     "\xBB\xCD\x01\x00\xBA\xCD\x01\x00\x38\x52\x00\x80\x3C\x52\x00\x80\xBF\x75"
-     "\x01\x00"
-     "\xBE\xC9\x01\x00\xBD\xC1\x01\x00\xBC\xC9\x01\x00\xA8\xE9\x05\x00\xA9\xE9"
-     "\x05\x00"
-     "\xAA\xF9\x05\x00\xAB\xF9\x05\x00\xAC\xE9\x05\x00\xAD\xE9\x05\x00\xAE\x39"
-     "\x06\x00"
-     "\xAF\x39\x06\x00\xB4\x51\x00\x80\x24\x52\x00\x80\x86\x88\x00\x00\x87\x00"
-     "\x03\x00"
-     "\x40\x52\x00\x80\x44\x52\x00\x80\x48\x52\x00\x80\x4C\x52\x00\x80\xB8\xD1"
-     "\x06\x00"
-     "\xB9\xD9\x06\x00\xBA\xE1\x06\x00\xBB\xE1\x06\x00\xBC\x91\x06\x00\xBD\x9D"
-     "\x06\x00"
-     "\xBE\x95\x06\x00\xBF\x89\x06\x00\xB0\x49\x06\x00\xB1\x49\x06\x00\xB2\x5D"
-     "\x06\x00"
-     "\xB3\x55\x06\x00\xB4\x4D\x06\x00\xB5\xF1\x06\x00\xB6\xF1\x06\x00\xB7\xF1"
-     "\x06\x00"
-     "\xA3\x71\x05\x00\x50\x52\x00\x80\x54\x52\x00\x80\x58\x52\x00\x80\x5C\x52"
-     "\x00\x80"
-     "\xA6\x69\x05\x00\xA5\x61\x05\x00\x60\x52\x00\x80\xAB\x8D\x06\x00\xAA\x8D"
-     "\x06\x00"
-     "\x64\x52\x00\x80\x68\x52\x00\x80\xAF\x35\x06\x00\xAE\x89\x06\x00\xAD\x81"
-     "\x06\x00"
-     "\xAC\x89\x06\x00\x6C\x52\x00\x80\x70\x52\x00\x80\xB3\xE1\x07\x00\x74\x52"
-     "\x00\x80"
-     "\xB5\xE1\x07\x00\x78\x52\x00\x80\x7C\x52\x00\x80\xB6\xE5\x07\x00\x44\x51"
-     "\x00\x80"
-     "\x80\x52\x00\x80\xBB\xBD\x07\x00\xBA\xA1\x07\x00\xBD\x95\x07\x00\xBC\xA9"
-     "\x07\x00"
-     "\xBF\x95\x07\x00\xBE\x95\x07\x00\xA8\x01\x06\x00\xA9\x25\x06\x00\xAA\x21"
-     "\x06\x00"
-     "\xAB\x21\x06\x00\xAC\x21\x06\x00\xAD\x2D\x06\x00\xAE\x25\x06\x00\xAF\x55"
-     "\x06\x00"
-     "\x84\x52\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\x88\x52"
-     "\x00\x80"
-     "\x8C\x52\x00\x80\x90\x52\x00\x80\xBE\x30\x01\x00\xB8\x39\x06\x00\xB9\x39"
-     "\x06\x00"
-     "\xBA\xC9\x06\x00\xBB\xC9\x06\x00\xBC\xD9\x06\x00\xBD\xD9\x06\x00\xBE\xC9"
-     "\x06\x00"
-     "\xBF\xC9\x06\x00\xB0\x2D\x06\x00\xB1\x31\x06\x00\xB2\x35\x06\x00\xB3\x09"
-     "\x06\x00"
-     "\xB4\x19\x06\x00\xB5\x19\x06\x00\xB6\x09\x06\x00\xB7\x09\x06\x00\xA3\xA9"
-     "\x06\x00"
-     "\x84\x8C\x02\x00\x86\x28\x1F\x00\x87\x44\x01\x00\x98\x52\x00\x80\xA6\xAD"
-     "\x06\x00"
-     "\xA5\xA9\x06\x00\x9C\x52\x00\x80\xAB\xF5\x06\x00\xAA\xE9\x06\x00\xA0\x52"
-     "\x00\x80"
-     "\xA4\x52\x00\x80\xAF\xDD\x06\x00\xAE\xDD\x06\x00\xAD\xDD\x06\x00\xAC\xE1"
-     "\x06\x00"
-     "\xA8\x52\x00\x80\xB3\x15\x06\x00\xAC\x52\x00\x80\xB0\x52\x00\x80\xB6\x3D"
-     "\x06\x00"
-     "\xB4\x52\x00\x80\xB8\x52\x00\x80\xB5\x35\x06\x00\xBA\xD9\x01\x00\xBB\xD9"
-     "\x01\x00"
-     "\xBC\x52\x00\x80\xC0\x52\x00\x80\xBE\x7D\x01\x00\xBF\x65\x01\x00\xBC\x7D"
-     "\x01\x00"
-     "\xBD\x75\x01\x00\xA8\xC5\x05\x00\xA9\xC9\x05\x00\xAA\xD9\x05\x00\xAB\xD1"
-     "\x05\x00"
-     "\xAC\xF9\x05\x00\xAD\xF9\x05\x00\xAE\x29\x02\x00\xAF\x29\x02\x00\xC4\x52"
-     "\x00\x80"
-     "\xC8\x52\x00\x80\xCC\x52\x00\x80\xD0\x52\x00\x80\x8C\x00\x00\x00\xD4\x52"
-     "\x00\x80"
-     "\xD8\x52\x00\x80\xDC\x52\x00\x80\xB8\xED\x02\x00\xB9\x85\x02\x00\xBA\x8D"
-     "\x02\x00"
-     "\xBB\x81\x02\x00\xBC\x85\x02\x00\xBD\x8D\x02\x00\xBE\xBD\x02\x00\xBF\x7D"
-     "\x03\x00"
-     "\xB0\x59\x02\x00\xB1\x59\x02\x00\xB2\xED\x02\x00\xB3\xE5\x02\x00\xB4\xFD"
-     "\x02\x00"
-     "\xB5\xE5\x02\x00\xB6\xE5\x02\x00\xB7\xD5\x02\x00\xA3\x51\x05\x00\xE0\x52"
-     "\x00\x80"
-     "\xE4\x52\x00\x80\xE8\x52\x00\x80\xEC\x52\x00\x80\xA6\x79\x05\x00\xA5\x71"
-     "\x05\x00"
-     "\xF0\x52\x00\x80\xAB\x9D\x02\x00\xAA\x9D\x02\x00\xF4\x52\x00\x80\xF8\x52"
-     "\x00\x80"
-     "\xAF\x21\x02\x00\xAE\x39\x02\x00\xAD\x31\x02\x00\xAC\x39\x02\x00\x82\x11"
-     "\x00\x00"
-     "\xFC\x52\x00\x80\x80\x65\x00\x00\x81\x09\x00\x00\x00\x53\x00\x80\xBE\x98"
-     "\x03\x00"
-     "\x08\x53\x00\x80\x0C\x53\x00\x80\x84\x24\x03\x00\x10\x53\x00\x80\x87\x68"
-     "\x03\x00"
-     "\x86\x8C\x1C\x00\x14\x53\x00\x80\x18\x53\x00\x80\x1C\x53\x00\x80\x20\x53"
-     "\x00\x80"
-     "\x24\x53\x00\x80\x28\x53\x00\x80\xB3\xA1\x02\x00\x84\xC0\x1C\x00\xB5\xD1"
-     "\x02\x00"
-     "\x2C\x53\x00\x80\x30\x53\x00\x80\xB6\xD5\x02\x00\x34\x53\x00\x80\x38\x53"
-     "\x00\x80"
-     "\xBB\xC1\x02\x00\xBA\xF5\x02\x00\xBD\x11\x01\x00\xBC\xD9\x02\x00\xBF\x11"
-     "\x01\x00"
-     "\xBE\x19\x01\x00\x3C\x53\x00\x80\x40\x53\x00\x80\x44\x53\x00\x80\x48\x53"
-     "\x00\x80"
-     "\x4C\x53\x00\x80\x50\x53\x00\x80\x54\x53\x00\x80\xEF\xB8\x06\x00\x58\x53"
-     "\x00\x80"
-     "\xE1\x9C\x04\x00\x5C\x53\x00\x80\xE3\x74\x06\x00\x60\x53\x00\x80\x64\x53"
-     "\x00\x80"
-     "\x68\x53\x00\x80\x6C\x53\x00\x80\x80\xF9\x00\x00\x81\xF9\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x70\x53\x00\x80\xBE\x58\x1C\x00\x84\x58\x1F\x00\xEF\x5C\x00\x00\xEF\x40"
-     "\x06\x00"
-     "\xE1\x90\x01\x00\xE1\xFC\x06\x00\xE3\x3C\x00\x00\xE3\x74\x06\x00\x78\x53"
-     "\x00\x80"
-     "\x7C\x53\x00\x80\x86\x98\x1C\x00\x87\xF4\x1C\x00\xA3\x69\x02\x00\xBE\x0C"
-     "\x1F\x00"
-     "\x80\x53\x00\x80\x84\x53\x00\x80\x88\x53\x00\x80\xA6\x1D\x02\x00\xA5\x19"
-     "\x02\x00"
-     "\x8C\x53\x00\x80\xAB\x09\x02\x00\xAA\x3D\x02\x00\x90\x53\x00\x80\x94\x53"
-     "\x00\x80"
-     "\xAF\xD9\x01\x00\xAE\xD1\x01\x00\xAD\xD9\x01\x00\xAC\x11\x02\x00\xA8\x91"
-     "\x1D\x00"
-     "\xA9\x99\x1D\x00\xAA\xA1\x1D\x00\xAB\xA1\x1D\x00\xAC\xD1\x1D\x00\xAD\xDD"
-     "\x1D\x00"
-     "\xAE\xD5\x1D\x00\xAF\xC9\x1D\x00\x04\x53\x00\x80\x74\x53\x00\x80\x98\x53"
-     "\x00\x80"
-     "\x9C\x53\x00\x80\xA0\x53\x00\x80\xA4\x53\x00\x80\xA8\x53\x00\x80\xAC\x53"
-     "\x00\x80"
-     "\xB8\x79\x1E\x00\xB9\x79\x1E\x00\xBA\xCD\x1E\x00\xBB\xC5\x1E\x00\xBC\xDD"
-     "\x1E\x00"
-     "\xBD\xC5\x1E\x00\xBE\xC5\x1E\x00\xBF\xF5\x1E\x00\xB0\xB9\x1D\x00\xB1\x8D"
-     "\x1D\x00"
-     "\xB2\x85\x1D\x00\xB3\x4D\x1E\x00\xB4\x55\x1E\x00\xB5\x5D\x1E\x00\xB6\x55"
-     "\x1E\x00"
-     "\xB7\x49\x1E\x00\xB8\xCD\x1F\x00\xB9\xD5\x1F\x00\xBA\xDD\x1F\x00\xBB\xD5"
-     "\x1F\x00"
-     "\xBC\xF1\x1F\x00\xBD\xFD\x1F\x00\xBE\xE9\x1F\x00\xBF\xE9\x1F\x00\xB0\xA5"
-     "\x1F\x00"
-     "\xB1\xB1\x1F\x00\xB2\xB1\x1F\x00\xB3\x8D\x1F\x00\xB4\x95\x1F\x00\xB5\xF5"
-     "\x1F\x00"
-     "\xB6\xFD\x1F\x00\xB7\xF5\x1F\x00\xA8\x19\x1E\x00\xA9\x19\x1E\x00\xAA\x2D"
-     "\x1E\x00"
-     "\xAB\x3D\x1E\x00\xAC\x25\x1E\x00\xAD\x2D\x1E\x00\xAE\x25\x1E\x00\xAF\xDD"
-     "\x1F\x00"
-     "\xB0\x53\x00\x80\xB4\x53\x00\x80\xB8\x53\x00\x80\xBC\x53\x00\x80\xC0\x53"
-     "\x00\x80"
-     "\x94\x52\x00\x80\xC4\x53\x00\x80\xC8\x53\x00\x80\xB3\xE5\x1F\x00\xCC\x53"
-     "\x00\x80"
-     "\xD0\x53\x00\x80\xD4\x53\x00\x80\xD8\x53\x00\x80\xB6\xED\x1F\x00\xB5\xED"
-     "\x1F\x00"
-     "\xDC\x53\x00\x80\xBB\x35\x1E\x00\xBA\x21\x1E\x00\xE0\x53\x00\x80\xE4\x53"
-     "\x00\x80"
-     "\xBF\x71\x1E\x00\xBE\x11\x1E\x00\xBD\x19\x1E\x00\xBC\x25\x1E\x00\x82\x69"
-     "\x00\x00"
-     "\xA3\xA1\x1F\x00\x80\x59\x00\x00\x81\x51\x00\x00\xA6\xA9\x1F\x00\xE8\x53"
-     "\x00\x80"
-     "\xEC\x53\x00\x80\xA5\xA9\x1F\x00\xAA\x65\x1E\x00\xAB\x71\x1E\x00\x86\x00"
-     "\x04\x00"
-     "\x87\xEC\x01\x00\xAE\x55\x1E\x00\xAF\x35\x1E\x00\xAC\x61\x1E\x00\xAD\x5D"
-     "\x1E\x00"
-     "\xA8\x31\x1E\x00\xA9\x31\x1E\x00\xAA\x41\x1E\x00\xAB\x41\x1E\x00\xAC\x41"
-     "\x1E\x00"
-     "\xAD\x49\x1E\x00\xAE\x71\x1E\x00\xAF\x71\x1E\x00\xF0\x53\x00\x80\xF4\x53"
-     "\x00\x80"
-     "\xF8\x53\x00\x80\xFC\x53\x00\x80\x00\x54\x00\x80\x04\x54\x00\x80\x08\x54"
-     "\x00\x80"
-     "\x0C\x54\x00\x80\xB8\x29\x01\x00\xB9\x29\x01\x00\xBA\x39\x01\x00\xBB\x35"
-     "\x01\x00"
-     "\xBC\x2D\x01\x00\xBD\xD1\x00\x00\xBE\xD1\x00\x00\xBF\xD1\x00\x00\xB0\xC9"
-     "\x01\x00"
-     "\xB1\xC9\x01\x00\xB2\xD9\x01\x00\xB3\xD9\x01\x00\xB4\xC9\x01\x00\xB5\xC9"
-     "\x01\x00"
-     "\xB6\x19\x01\x00\xB7\x19\x01\x00\xB3\xC9\x1D\x00\x10\x54\x00\x80\x14\x54"
-     "\x00\x80"
-     "\x18\x54\x00\x80\x1C\x54\x00\x80\xB6\xC9\x1D\x00\xB5\xC9\x1D\x00\x20\x54"
-     "\x00\x80"
-     "\xBB\x0D\x02\x00\xBA\x0D\x02\x00\x24\x54\x00\x80\x28\x54\x00\x80\xBF\x0D"
-     "\x02\x00"
-     "\xBE\x0D\x02\x00\xBD\x0D\x02\x00\xBC\x0D\x02\x00\x2C\x54\x00\x80\xA3\x8D"
-     "\x1D\x00"
-     "\x30\x54\x00\x80\x34\x54\x00\x80\xA6\x8D\x1D\x00\x38\x54\x00\x80\x3C\x54"
-     "\x00\x80"
-     "\xA5\x8D\x1D\x00\xAA\x49\x02\x00\xAB\x49\x02\x00\x40\x54\x00\x80\x44\x54"
-     "\x00\x80"
-     "\xAE\x49\x02\x00\xAF\x49\x02\x00\xAC\x49\x02\x00\xAD\x49\x02\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x11\x00\x00\x82\x11\x00\x00\x48\x54\x00\x80\xEF\xCC\x02\x00\x4C\x54"
-     "\x00\x80"
-     "\x50\x54\x00\x80\x84\x90\x02\x00\xE3\x2C\x02\x00\xBE\x28\x03\x00\xE1\xD8"
-     "\x01\x00"
-     "\x58\x54\x00\x80\x86\x10\x04\x00\x87\xD4\x03\x00\x5C\x54\x00\x80\x60\x54"
-     "\x00\x80"
-     "\xB3\x61\x03\x00\x64\x54\x00\x80\x68\x54\x00\x80\x6C\x54\x00\x80\x70\x54"
-     "\x00\x80"
-     "\xB6\x79\x03\x00\xB5\x71\x03\x00\x74\x54\x00\x80\xBB\x5D\x03\x00\xBA\x5D"
-     "\x03\x00"
-     "\x78\x54\x00\x80\x7C\x54\x00\x80\xBF\xE1\x00\x00\xBE\xF9\x00\x00\xBD\xF1"
-     "\x00\x00"
-     "\xBC\xF9\x00\x00\xA3\xA1\x02\x00\x80\x54\x00\x80\x84\x54\x00\x80\x88\x54"
-     "\x00\x80"
-     "\x8C\x54\x00\x80\xA6\xB9\x02\x00\xA5\xB1\x02\x00\x90\x54\x00\x80\xAB\x9D"
-     "\x02\x00"
-     "\xAA\x9D\x02\x00\x94\x54\x00\x80\x98\x54\x00\x80\xAF\x21\x01\x00\xAE\x39"
-     "\x01\x00"
-     "\xAD\x31\x01\x00\xAC\x39\x01\x00\x9C\x54\x00\x80\xA0\x54\x00\x80\xEF\x64"
-     "\x1F\x00"
-     "\xA4\x54\x00\x80\xA8\x54\x00\x80\xAC\x54\x00\x80\xB0\x54\x00\x80\xB4\x54"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\x11\x00\x00\x82\x15\x00\x00\xB8\x54\x00\x80\xE1\xE0"
-     "\x1C\x00"
-     "\xBC\x54\x00\x80\xE3\x88\x1F\x00\xC0\x54\x00\x80\x84\x80\x02\x00\xBE\x8C"
-     "\x05\x00"
-     "\x87\x48\x05\x00\x86\x2C\x04\x00\xC8\x54\x00\x80\xCC\x54\x00\x80\xEF\xA4"
-     "\x1E\x00"
-     "\xEF\xF4\x1E\x00\xE1\x00\x1E\x00\xE1\x50\x1F\x00\xE3\x2C\x1E\x00\xE3\xB0"
-     "\x1E\x00"
-     "\xD0\x54\x00\x80\xD4\x54\x00\x80\xD8\x54\x00\x80\xDC\x54\x00\x80\xE0\x54"
-     "\x00\x80"
-     "\xE4\x54\x00\x80\x84\x84\x04\x00\xB3\x71\x01\x00\xE8\x54\x00\x80\xB5\x1D"
-     "\x01\x00"
-     "\xB6\x15\x01\x00\xEC\x54\x00\x80\xF0\x54\x00\x80\xF4\x54\x00\x80\xBA\x01"
-     "\x01\x00"
-     "\xBB\x01\x01\x00\xBC\xF5\x00\x00\xBD\xFD\x00\x00\xBE\xF5\x00\x00\xBF\xED"
-     "\x00\x00"
-     "\xA8\xAD\x06\x00\xA9\xBD\x06\x00\xAA\xB5\x06\x00\xAB\xC9\x06\x00\xAC\xD9"
-     "\x06\x00"
-     "\xAD\xD9\x06\x00\xAE\xC9\x06\x00\xAF\xC1\x06\x00\xF8\x54\x00\x80\xFC\x54"
-     "\x00\x80"
-     "\x00\x55\x00\x80\x04\x55\x00\x80\x08\x55\x00\x80\x0C\x55\x00\x80\x10\x55"
-     "\x00\x80"
-     "\x14\x55\x00\x80\xB8\x6D\x07\x00\xB9\x05\x07\x00\xBA\x0D\x07\x00\xBB\x01"
-     "\x07\x00"
-     "\xBC\x01\x07\x00\xBD\x01\x07\x00\xBE\x01\x07\x00\xBF\x01\x07\x00\xB0\x89"
-     "\x06\x00"
-     "\xB1\x89\x06\x00\xB2\x6D\x07\x00\xB3\x65\x07\x00\xB4\x7D\x07\x00\xB5\x65"
-     "\x07\x00"
-     "\xB6\x65\x07\x00\xB7\x55\x07\x00\x54\x54\x00\x80\xA3\x39\x06\x00\x18\x55"
-     "\x00\x80"
-     "\xC4\x54\x00\x80\xA6\x5D\x06\x00\x1C\x55\x00\x80\x20\x55\x00\x80\xA5\x55"
-     "\x06\x00"
-     "\xAA\x49\x06\x00\xAB\x49\x06\x00\x24\x55\x00\x80\x28\x55\x00\x80\xAE\xBD"
-     "\x07\x00"
-     "\xAF\xA5\x07\x00\xAC\xBD\x07\x00\xAD\xB5\x07\x00\x80\x6D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x2C\x55\x00\x80\x30\x55\x00\x80\xBE\x9C\x03\x00\x34\x55"
-     "\x00\x80"
-     "\x38\x55\x00\x80\x86\x40\x00\x00\x87\x60\x00\x00\x3C\x55\x00\x80\x40\x55"
-     "\x00\x80"
-     "\x44\x55\x00\x80\x48\x55\x00\x80\x4C\x55\x00\x80\x50\x55\x00\x80\xA8\x91"
-     "\x06\x00"
-     "\xA9\x91\x06\x00\xAA\xB9\x06\x00\xAB\xB9\x06\x00\xAC\xA9\x06\x00\xAD\xA9"
-     "\x06\x00"
-     "\xAE\xD9\x06\x00\xAF\xD9\x06\x00\x54\x55\x00\x80\x58\x55\x00\x80\x5C\x55"
-     "\x00\x80"
-     "\x60\x55\x00\x80\x64\x55\x00\x80\x68\x55\x00\x80\x6C\x55\x00\x80\x70\x55"
-     "\x00\x80"
-     "\xB8\x71\x01\x00\xB9\x71\x01\x00\xBA\x71\x01\x00\xBB\x71\x01\x00\xBC\xD9"
-     "\x01\x00"
-     "\xBD\xC1\x01\x00\xBE\xC1\x01\x00\xBF\xF5\x01\x00\xB0\xB1\x06\x00\xB1\xB9"
-     "\x06\x00"
-     "\xB2\x89\x06\x00\xB3\x89\x06\x00\xB4\x51\x01\x00\xB5\x51\x01\x00\xB6\x51"
-     "\x01\x00"
-     "\xB7\x51\x01\x00\xB3\x31\x06\x00\x74\x55\x00\x80\x78\x55\x00\x80\x7C\x55"
-     "\x00\x80"
-     "\x80\x55\x00\x80\xB6\x29\x06\x00\xB5\x21\x06\x00\x84\x55\x00\x80\xBB\x7D"
-     "\x06\x00"
-     "\xBA\x75\x06\x00\x88\x55\x00\x80\x8C\x55\x00\x80\xBF\x95\x01\x00\xBE\x95"
-     "\x01\x00"
-     "\xBD\x5D\x06\x00\xBC\x5D\x06\x00\x90\x55\x00\x80\xA3\x75\x06\x00\x94\x55"
-     "\x00\x80"
-     "\x98\x55\x00\x80\xA6\x6D\x06\x00\x9C\x55\x00\x80\xA0\x55\x00\x80\xA5\x65"
-     "\x06\x00"
-     "\xAA\x31\x06\x00\xAB\x39\x06\x00\x84\xAC\x01\x00\xBE\xA0\x01\x00\xAE\xD1"
-     "\x01\x00"
-     "\xAF\xD1\x01\x00\xAC\x19\x06\x00\xAD\x19\x06\x00\xA8\xDD\x02\x00\xA9\xED"
-     "\x02\x00"
-     "\xAA\xE5\x02\x00\xAB\xFD\x02\x00\xAC\xE5\x02\x00\xAD\xED\x02\x00\xAE\xE5"
-     "\x02\x00"
-     "\xAF\x3D\x03\x00\xA8\x55\x00\x80\xAC\x55\x00\x80\xB0\x55\x00\x80\xBE\x64"
-     "\x0C\x00"
-     "\xB4\x55\x00\x80\xB8\x55\x00\x80\xBC\x55\x00\x80\xC0\x55\x00\x80\xB8\xC9"
-     "\x03\x00"
-     "\xB9\xC9\x03\x00\xBA\xD9\x03\x00\xBB\xD1\x03\x00\xBC\xF9\x03\x00\xBD\xF9"
-     "\x03\x00"
-     "\xBE\x99\x03\x00\xBF\x95\x03\x00\xB0\x45\x03\x00\xB1\x4D\x03\x00\xB2\x45"
-     "\x03\x00"
-     "\xB3\x5D\x03\x00\xB4\x45\x03\x00\xB5\x4D\x03\x00\xB6\x45\x03\x00\xB7\xF9"
-     "\x03\x00"
-     "\x81\x55\x03\x00\x80\x49\x03\x00\xB3\x65\x02\x00\x82\x55\x03\x00\xB5\x65"
-     "\x02\x00"
-     "\xC4\x55\x00\x80\xC8\x55\x00\x80\xB6\x65\x02\x00\x86\x00\x0C\x00\x87\xE4"
-     "\x03\x00"
-     "\xBB\x81\x03\x00\xBA\x89\x03\x00\xBD\x81\x03\x00\xBC\x99\x03\x00\xBF\x81"
-     "\x03\x00"
-     "\xBE\x89\x03\x00\xA3\x2D\x02\x00\xCC\x55\x00\x80\xD0\x55\x00\x80\xD4\x55"
-     "\x00\x80"
-     "\xD8\x55\x00\x80\xA6\x2D\x02\x00\xA5\x2D\x02\x00\xDC\x55\x00\x80\xAB\xC9"
-     "\x03\x00"
-     "\xAA\xC1\x03\x00\xE0\x55\x00\x80\xE4\x55\x00\x80\xAF\xC9\x03\x00\xAE\xC1"
-     "\x03\x00"
-     "\xAD\xC9\x03\x00\xAC\xD1\x03\x00\xE3\xD8\x06\x00\xE1\xAC\x07\x00\xE1\x9C"
-     "\x06\x00"
-     "\xE3\x9C\x06\x00\xE8\x55\x00\x80\x84\x58\x0D\x00\xEC\x55\x00\x80\xF0\x55"
-     "\x00\x80"
-     "\xF4\x55\x00\x80\xF8\x55\x00\x80\xFC\x55\x00\x80\x00\x56\x00\x80\xEF\x1C"
-     "\x01\x00"
-     "\x04\x56\x00\x80\x08\x56\x00\x80\xEF\x40\x06\x00\x80\x79\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x11\x00\x00\x84\x00\x0C\x00\x0C\x56\x00\x80\xE3\xAC\x00\x00\x10\x56"
-     "\x00\x80"
-     "\xE1\xA4\x01\x00\x18\x56\x00\x80\xEF\x70\x00\x00\x86\xC8\x0C\x00\x87\xA4"
-     "\x0D\x00"
-     "\x1C\x56\x00\x80\x20\x56\x00\x80\x24\x56\x00\x80\x28\x56\x00\x80\xBA\xC9"
-     "\x05\x00"
-     "\xBB\xC9\x05\x00\xB8\xA5\x05\x00\xB9\xCD\x05\x00\xBE\xF9\x05\x00\xBF\xF9"
-     "\x05\x00"
-     "\xBC\xCD\x05\x00\xBD\xC5\x05\x00\xB2\xA5\x05\x00\xB3\xAD\x05\x00\xB0\x11"
-     "\x06\x00"
-     "\xB1\x11\x06\x00\xB6\xAD\x05\x00\xB7\x9D\x05\x00\xB4\xB5\x05\x00\xB5\xAD"
-     "\x05\x00"
-     "\xAA\x61\x06\x00\xAB\x61\x06\x00\xA8\x9D\x06\x00\xA9\x95\x06\x00\xAE\x61"
-     "\x06\x00"
-     "\xAF\x61\x06\x00\xAC\x71\x06\x00\xAD\x71\x06\x00\x2C\x56\x00\x80\x30\x56"
-     "\x00\x80"
-     "\x34\x56\x00\x80\x38\x56\x00\x80\x3C\x56\x00\x80\x40\x56\x00\x80\xBE\xB0"
-     "\x0C\x00"
-     "\x44\x56\x00\x80\xA8\x21\x0E\x00\xA9\x21\x0E\x00\xAA\x21\x0E\x00\xAB\x3D"
-     "\x0E\x00"
-     "\xAC\x25\x0E\x00\xAD\x2D\x0E\x00\xAE\x25\x0E\x00\xAF\x89\x0E\x00\x14\x56"
-     "\x00\x80"
-     "\x48\x56\x00\x80\x4C\x56\x00\x80\x50\x56\x00\x80\x54\x56\x00\x80\x58\x56"
-     "\x00\x80"
-     "\x5C\x56\x00\x80\x60\x56\x00\x80\xB8\x51\x0F\x00\xB9\x5D\x0F\x00\xBA\x55"
-     "\x0F\x00"
-     "\xBB\x6D\x0F\x00\xBC\x75\x0F\x00\xBD\x7D\x0F\x00\xBE\x75\x0F\x00\xBF\x69"
-     "\x0F\x00"
-     "\xB0\xF9\x0E\x00\xB1\xA1\x0E\x00\xB2\xA1\x0E\x00\xB3\xA1\x0E\x00\xB4\xA1"
-     "\x0E\x00"
-     "\xB5\xA9\x0E\x00\xB6\x91\x0E\x00\xB7\x91\x0E\x00\xB3\xA9\x0E\x00\x64\x56"
-     "\x00\x80"
-     "\xA4\x55\x00\x80\x68\x56\x00\x80\x6C\x56\x00\x80\xB6\xAD\x0E\x00\xB5\xAD"
-     "\x0E\x00"
-     "\x70\x56\x00\x80\xBB\x65\x0E\x00\xBA\x3D\x0E\x00\x74\x56\x00\x80\x78\x56"
-     "\x00\x80"
-     "\xBF\x6D\x0E\x00\xBE\x65\x0E\x00\xBD\x75\x0E\x00\xBC\x75\x0E\x00\x82\x19"
-     "\x00\x00"
-     "\xA3\xED\x0E\x00\x80\x65\x00\x00\x81\x19\x00\x00\xA6\xE9\x0E\x00\x7C\x56"
-     "\x00\x80"
-     "\x80\x56\x00\x80\xA5\xE9\x0E\x00\xAA\x79\x0E\x00\xAB\x21\x0E\x00\x84\x56"
-     "\x00\x80"
-     "\x88\x56\x00\x80\xAE\x21\x0E\x00\xAF\x29\x0E\x00\xAC\x31\x0E\x00\xAD\x31"
-     "\x0E\x00"
-     "\xA8\x61\x0E\x00\xA9\x75\x0E\x00\xAA\x7D\x0E\x00\xAB\x75\x0E\x00\xAC\x6D"
-     "\x0E\x00"
-     "\xAD\xF5\x0E\x00\xAE\xFD\x0E\x00\xAF\xF5\x0E\x00\x86\x80\x01\x00\x87\xA4"
-     "\x01\x00"
-     "\x8C\x56\x00\x80\x90\x56\x00\x80\x94\x56\x00\x80\x98\x56\x00\x80\x9C\x56"
-     "\x00\x80"
-     "\xA0\x56\x00\x80\xB8\x75\x01\x00\xB9\x7D\x01\x00\xBA\x75\x01\x00\xBB\xC9"
-     "\x01\x00"
-     "\xBC\xDD\x01\x00\xBD\xC5\x01\x00\xBE\xC5\x01\x00\xBF\xF5\x01\x00\xB0\x8D"
-     "\x0E\x00"
-     "\xB1\x95\x0E\x00\xB2\x9D\x0E\x00\xB3\x91\x0E\x00\xB4\x55\x01\x00\xB5\x5D"
-     "\x01\x00"
-     "\xB6\x55\x01\x00\xB7\x4D\x01\x00\xB3\xF5\x0E\x00\xA4\x56\x00\x80\xA8\x56"
-     "\x00\x80"
-     "\xAC\x56\x00\x80\xB0\x56\x00\x80\xB6\x75\x0E\x00\xB5\xE5\x0E\x00\xB4\x56"
-     "\x00\x80"
-     "\xBB\x51\x0E\x00\xBA\x49\x0E\x00\xB8\x56\x00\x80\xBC\x56\x00\x80\xBF\x99"
-     "\x01\x00"
-     "\xBE\x91\x01\x00\xBD\x45\x0E\x00\xBC\x49\x0E\x00\xC0\x56\x00\x80\xA3\xB1"
-     "\x0E\x00"
-     "\xC4\x56\x00\x80\xC8\x56\x00\x80\xA6\x31\x0E\x00\xCC\x56\x00\x80\xD0\x56"
-     "\x00\x80"
-     "\xA5\xA1\x0E\x00\xAA\x0D\x0E\x00\xAB\x15\x0E\x00\xD4\x56\x00\x80\xD8\x56"
-     "\x00\x80"
-     "\xAE\xD5\x01\x00\xAF\xDD\x01\x00\xAC\x0D\x0E\x00\xAD\x01\x0E\x00\xA8\xED"
-     "\x02\x00"
-     "\xA9\x2D\x03\x00\xAA\x25\x03\x00\xAB\x3D\x03\x00\xAC\x25\x03\x00\xAD\x2D"
-     "\x03\x00"
-     "\xAE\x25\x03\x00\xAF\x99\x03\x00\xDC\x56\x00\x80\xE0\x56\x00\x80\xE4\x56"
-     "\x00\x80"
-     "\xE8\x56\x00\x80\xEC\x56\x00\x80\xF0\x56\x00\x80\xBE\x74\x02\x00\xF8\x56"
-     "\x00\x80"
-     "\xB8\x8D\x03\x00\xB9\x91\x03\x00\xBA\x91\x03\x00\xBB\xA5\x03\x00\xBC\xBD"
-     "\x03\x00"
-     "\xBD\x75\x00\x00\xBE\x7D\x00\x00\xBF\x75\x00\x00\xB0\xE9\x03\x00\xB1\xE9"
-     "\x03\x00"
-     "\xB2\xF9\x03\x00\xB3\xF1\x03\x00\xB4\xD9\x03\x00\xB5\xD9\x03\x00\xB6\xB9"
-     "\x03\x00"
-     "\xB7\xB5\x03\x00\x80\xAD\x00\x00\x81\xB5\x00\x00\x82\xBD\x00\x00\xB3\xA1"
-     "\x03\x00"
-     "\xFC\x56\x00\x80\xB5\xA1\x03\x00\xB6\xA1\x03\x00\x00\x57\x00\x80\x84\xE0"
-     "\x02\x00"
-     "\x04\x57\x00\x80\xBA\x21\x03\x00\xBB\x21\x03\x00\xBC\x21\x03\x00\xBD\x29"
-     "\x03\x00"
-     "\xBE\x11\x03\x00\xBF\x11\x03\x00\xA3\xED\x03\x00\x08\x57\x00\x80\x86\xE8"
-     "\x04\x00"
-     "\x87\x4C\x03\x00\x0C\x57\x00\x80\xA6\xED\x03\x00\xA5\xED\x03\x00\x10\x57"
-     "\x00\x80"
-     "\xAB\x6D\x03\x00\xAA\x6D\x03\x00\x14\x57\x00\x80\x18\x57\x00\x80\xAF\x5D"
-     "\x03\x00"
-     "\xAE\x5D\x03\x00\xAD\x65\x03\x00\xAC\x6D\x03\x00\x1C\x57\x00\x80\xE3\x00"
-     "\x0E\x00"
-     "\xEF\xD4\x23\x00\xE1\xB4\x0F\x00\x20\x57\x00\x80\x24\x57\x00\x80\x28\x57"
-     "\x00\x80"
-     "\x2C\x57\x00\x80\xA1\xA9\x03\x00\xA0\xFD\x0F\x00\xA3\xC1\x03\x00\xA2\x0D"
-     "\x03\x00"
-     "\xE1\x84\x12\x00\xEF\xE0\x0F\x00\xE3\xE4\xF3\x00\x30\x57\x00\x80\x34\x57"
-     "\x00\x80"
-     "\x38\x57\x00\x80\x84\x84\x04\x00\xBE\x84\x04\x00\x3C\x57\x00\x80\xEF\x94"
-     "\x03\x00"
-     "\x40\x57\x00\x80\x44\x57\x00\x80\x48\x57\x00\x80\xE3\xD4\x03\x00\x4C\x57"
-     "\x00\x80"
-     "\xE1\x54\x00\x00\x50\x57\x00\x80\x54\x57\x00\x80\x58\x57\x00\x80\x5C\x57"
-     "\x00\x80"
-     "\x80\xF9\x00\x00\x81\xF9\x00\x00\x82\x05\x00\x00\x60\x57\x00\x80\x68\x57"
-     "\x00\x80"
-     "\x6C\x57\x00\x80\x8F\x91\x1B\x00\xEF\x9C\x0E\x00\x84\xEC\x07\x00\xE1\xD4"
-     "\x0E\x00"
-     "\x70\x57\x00\x80\xE3\xF0\x0E\x00\x74\x57\x00\x80\x78\x57\x00\x80\x86\x18"
-     "\x07\x00"
-     "\x87\xF4\x04\x00\x99\xE5\x17\x00\x98\xE5\x17\x00\x9B\xE9\x0B\x00\x9A\x39"
-     "\x0B\x00"
-     "\x9D\xFD\x0A\x00\x9C\xF1\x0B\x00\x9F\x55\x0F\x00\x9E\x5D\x0F\x00\x91\x29"
-     "\x1F\x00"
-     "\x90\xF5\x1B\x00\x93\x45\x1F\x00\x92\xCD\x1F\x00\x95\xD1\x13\x00\x94\x29"
-     "\x13\x00"
-     "\x97\x19\x17\x00\x96\x75\x13\x00\xA1\x05\x00\x00\xF4\x56\x00\x80\xA3\x01"
-     "\xFC\x00"
-     "\xA2\x55\x00\x00\xA5\x01\xFC\x00\xA4\x11\xFC\x00\xA7\x19\xF8\x00\xA6\x45"
-     "\xFD\x00"
-     "\xA9\x09\xF9\x00\xA8\x21\xF8\x00\xAB\x09\xF4\x00\xAA\x01\xF4\x00\xAD\x01"
-     "\xF0\x00"
-     "\xAC\x15\xF5\x00\xAF\xF1\xF0\x00\xAE\x1D\xF0\x00\xB1\x01\xEC\x00\xB0\x01"
-     "\xEC\x00"
-     "\xB3\x01\xE8\x00\xB2\x11\xEC\x00\xB5\x4D\xE8\x00\xB4\x51\xE8\x00\x64\x57"
-     "\x00\x80"
-     "\x7C\x57\x00\x80\x80\x57\x00\x80\x84\x57\x00\x80\x88\x57\x00\x80\x8C\x57"
-     "\x00\x80"
-     "\x90\x57\x00\x80\x94\x57\x00\x80\x98\x57\x00\x80\x9C\x57\x00\x80\xA8\xBD"
-     "\x06\x00"
-     "\xA9\x15\x06\x00\xAA\x15\x06\x00\xAB\x29\x06\x00\xAC\x55\x06\x00\xAD\x45"
-     "\x06\x00"
-     "\xAE\x45\x06\x00\xAF\x75\x06\x00\xBE\xE8\x00\x00\xA0\x57\x00\x80\xA4\x57"
-     "\x00\x80"
-     "\xA8\x57\x00\x80\xAC\x57\x00\x80\xB0\x57\x00\x80\xB4\x57\x00\x80\xB8\x57"
-     "\x00\x80"
-     "\xB8\x85\x07\x00\xB9\x95\x07\x00\xBA\x91\x07\x00\xBB\x91\x07\x00\xBC\xB1"
-     "\x07\x00"
-     "\xBD\xBD\x07\x00\xBE\xA9\x07\x00\xBF\xA9\x07\x00\xB0\x0D\x06\x00\xB1\xE5"
-     "\x07\x00"
-     "\xB2\xF9\x07\x00\xB3\xF9\x07\x00\xB4\xE9\x07\x00\xB5\xE9\x07\x00\xB6\xC5"
-     "\x07\x00"
-     "\xB7\xBD\x07\x00\xB3\x9D\x07\x00\xBC\x57\x00\x80\xC0\x57\x00\x80\xC4\x57"
-     "\x00\x80"
-     "\xC8\x57\x00\x80\xB6\xC5\x07\x00\xB5\xDD\x07\x00\xCC\x57\x00\x80\xBB\x21"
-     "\x06\x00"
-     "\xBA\xC9\x07\x00\xD0\x57\x00\x80\xD8\x57\x00\x80\xBF\x2D\x06\x00\xBE\x2D"
-     "\x06\x00"
-     "\xBD\x2D\x06\x00\xBC\x2D\x06\x00\x82\x69\x00\x00\xA3\xD9\x07\x00\x80\x59"
-     "\x00\x00"
-     "\x81\x51\x00\x00\xA6\x81\x07\x00\xDC\x57\x00\x80\xE0\x57\x00\x80\xA5\x99"
-     "\x07\x00"
-     "\xAA\x8D\x07\x00\xAB\x65\x06\x00\x86\xC8\x00\x00\x87\xA4\x01\x00\xAE\x69"
-     "\x06\x00"
-     "\xAF\x69\x06\x00\xAC\x69\x06\x00\xAD\x69\x06\x00\x8F\x75\x0D\x00\x8E\xA1"
-     "\x11\x00"
-     "\x8D\xA9\x11\x00\x8C\x6D\x11\x00\x8B\x19\x15\x00\x8A\xB5\x15\x00\xE4\x57"
-     "\x00\x80"
-     "\xE8\x57\x00\x80\x87\x7D\x15\x00\x86\x09\x19\x00\x85\x89\x19\x00\x84\x65"
-     "\x19\x00"
-     "\x83\xA1\x1D\x00\x82\xBD\x1D\x00\xEC\x57\x00\x80\xF0\x57\x00\x80\xF4\x57"
-     "\x00\x80"
-     "\xF8\x57\x00\x80\xFC\x57\x00\x80\x9C\x25\x05\x00\x9B\xF1\x05\x00\x9A\x5D"
-     "\x05\x00"
-     "\x00\x58\x00\x80\x04\x58\x00\x80\x97\x39\x09\x00\x96\x69\x09\x00\x95\x69"
-     "\x09\x00"
-     "\x94\xC1\x0C\x00\x93\xD5\x0D\x00\x92\xD5\x0D\x00\x08\x58\x00\x80\x0C\x58"
-     "\x00\x80"
-     "\xA8\x91\x05\x00\xA9\x91\x05\x00\xAA\xA1\x05\x00\xAB\xBD\x05\x00\xAC\xA5"
-     "\x05\x00"
-     "\xAD\x49\x02\x00\xAE\x79\x02\x00\xAF\x79\x02\x00\x84\xEC\xFC\x00\x10\x58"
-     "\x00\x80"
-     "\x14\x58\x00\x80\x18\x58\x00\x80\x1C\x58\x00\x80\x20\x58\x00\x80\x24\x58"
-     "\x00\x80"
-     "\x28\x58\x00\x80\xB8\xED\x02\x00\xB9\xF5\x02\x00\xBA\xFD\x02\x00\xBB\xF5"
-     "\x02\x00"
-     "\xBC\xED\x02\x00\xBD\x51\x03\x00\xBE\x51\x03\x00\xBF\x51\x03\x00\xB0\x09"
-     "\x02\x00"
-     "\xB1\x09\x02\x00\xB2\x19\x02\x00\xB3\x19\x02\x00\xB4\x09\x02\x00\xB5\x09"
-     "\x02\x00"
-     "\xB6\xDD\x02\x00\xB7\xD5\x02\x00\xB3\x29\x05\x00\x84\x48\xFD\x00\x2C\x58"
-     "\x00\x80"
-     "\x30\x58\x00\x80\x34\x58\x00\x80\xB6\xC9\x02\x00\xB5\x2D\x05\x00\x38\x58"
-     "\x00\x80"
-     "\xBB\xCD\x02\x00\xBA\xCD\x02\x00\x3C\x58\x00\x80\x40\x58\x00\x80\xBF\xCD"
-     "\x02\x00"
-     "\xBE\xCD\x02\x00\xBD\xCD\x02\x00\xBC\xCD\x02\x00\x82\x6D\x00\x00\xA3\x6D"
-     "\x05\x00"
-     "\x80\x55\x00\x00\x81\x65\x00\x00\xA6\x8D\x02\x00\xBE\x1C\xFC\x00\x48\x58"
-     "\x00\x80"
-     "\xA5\x69\x05\x00\xAA\x89\x02\x00\xAB\x89\x02\x00\x86\xC8\xFC\x00\x87\x04"
-     "\x03\x00"
-     "\xAE\x89\x02\x00\xAF\x89\x02\x00\xAC\x89\x02\x00\xAD\x89\x02\x00\xA8\x2D"
-     "\x02\x00"
-     "\xA9\x51\x02\x00\xAA\x51\x02\x00\xAB\x95\x02\x00\xAC\x8D\x02\x00\xAD\xAD"
-     "\x02\x00"
-     "\xAE\xA5\x02\x00\xAF\xDD\x02\x00\x4C\x58\x00\x80\x50\x58\x00\x80\x54\x58"
-     "\x00\x80"
-     "\x58\x58\x00\x80\x5C\x58\x00\x80\x60\x58\x00\x80\x64\x58\x00\x80\x68\x58"
-     "\x00\x80"
-     "\xB8\x6D\x01\x00\xB9\x75\x01\x00\xBA\x7D\x01\x00\xBB\x75\x01\x00\xBC\x6D"
-     "\x01\x00"
-     "\xBD\xDD\x01\x00\xBE\xD5\x01\x00\xBF\xCD\x01\x00\xB0\xA5\x02\x00\xB1\xAD"
-     "\x02\x00"
-     "\xB2\xB1\x02\x00\xB3\xB1\x02\x00\xB4\x89\x02\x00\xB5\x89\x02\x00\xB6\x5D"
-     "\x01\x00"
-     "\xB7\x55\x01\x00\xE1\x50\x07\x00\x6C\x58\x00\x80\xE3\xD0\x06\x00\x70\x58"
-     "\x00\x80"
-     "\x74\x58\x00\x80\x78\x58\x00\x80\x7C\x58\x00\x80\x80\x58\x00\x80\x84\x58"
-     "\x00\x80"
-     "\x88\x58\x00\x80\x8C\x58\x00\x80\x90\x58\x00\x80\x94\x58\x00\x80\x98\x58"
-     "\x00\x80"
-     "\xEF\x0C\x06\x00\xBE\x74\xFC\x00\x9C\x58\x00\x80\xA0\x58\x00\x80\x86\xE8"
-     "\xFC\x00"
-     "\x87\x84\xFC\x00\xA4\x58\x00\x80\xE3\x64\x06\x00\xA8\x58\x00\x80\xE1\x7C"
-     "\x06\x00"
-     "\x82\x15\x00\x00\xEF\xBC\x01\x00\x80\x31\x00\x00\x81\x1D\x00\x00\xAC\x58"
-     "\x00\x80"
-     "\xB0\x58\x00\x80\xB8\x58\x00\x80\xBC\x58\x00\x80\xB3\x59\xFD\x00\x44\x58"
-     "\x00\x80"
-     "\xD4\x57\x00\x80\xC0\x58\x00\x80\xB4\x58\x00\x80\xB6\xAD\xFD\x00\xB5\x85"
-     "\xFD\x00"
-     "\xC4\x58\x00\x80\xBB\xE9\xFD\x00\xBA\xE1\xFD\x00\xC8\x58\x00\x80\xCC\x58"
-     "\x00\x80"
-     "\xBF\xCD\xFD\x00\xBE\xCD\xFD\x00\xBD\xE5\xFD\x00\xBC\xE5\xFD\x00\xD0\x58"
-     "\x00\x80"
-     "\xA3\x1D\xFD\x00\xD4\x58\x00\x80\xD8\x58\x00\x80\xA6\xE9\xFD\x00\xDC\x58"
-     "\x00\x80"
-     "\xE0\x58\x00\x80\xA5\xC1\xFD\x00\xAA\xA5\xFD\x00\xAB\xAD\xFD\x00\xE4\x58"
-     "\x00\x80"
-     "\xE8\x58\x00\x80\xAE\x89\xFD\x00\xAF\x89\xFD\x00\xAC\xA1\xFD\x00\xAD\xA1"
-     "\xFD\x00"
-     "\xA8\x59\xFE\x00\xA9\xA5\xFF\x00\xAA\xAD\xFF\x00\xAB\xA5\xFF\x00\xAC\xBD"
-     "\xFF\x00"
-     "\xAD\xA5\xFF\x00\xAE\xAD\xFF\x00\xAF\xA5\xFF\x00\xEC\x58\x00\x80\xF0\x58"
-     "\x00\x80"
-     "\xF4\x58\x00\x80\xF8\x58\x00\x80\xFC\x58\x00\x80\x00\x59\x00\x80\x04\x59"
-     "\x00\x80"
-     "\x08\x59\x00\x80\xB8\x65\xFF\x00\xB9\x6D\xFF\x00\xBA\x65\xFF\x00\xBB\x7D"
-     "\xFF\x00"
-     "\xBC\x65\xFF\x00\xBD\x15\xFF\x00\xBE\x1D\xFF\x00\xBF\x0D\xFF\x00\xB0\xDD"
-     "\xFF\x00"
-     "\xB1\x61\xFF\x00\xB2\x65\xFF\x00\xB3\x7D\xFF\x00\xB4\x65\xFF\x00\xB5\x6D"
-     "\xFF\x00"
-     "\xB6\x65\xFF\x00\xB7\x5D\xFF\x00\xB3\x25\xFE\x00\x0C\x59\x00\x80\x10\x59"
-     "\x00\x80"
-     "\x14\x59\x00\x80\x18\x59\x00\x80\xB6\x25\xFE\x00\xB5\x35\xFE\x00\x1C\x59"
-     "\x00\x80"
-     "\xBB\xA5\xFE\x00\xBA\x19\xFE\x00\x20\x59\x00\x80\x84\xE0\x01\x00\xBF\xA9"
-     "\xFE\x00"
-     "\xBE\xA1\xFE\x00\xBD\xA9\xFE\x00\xBC\xB1\xFE\x00\x82\x69\x00\x00\xA3\x61"
-     "\xFE\x00"
-     "\x80\x59\x00\x00\x81\x51\x00\x00\xA6\x61\xFE\x00\xBE\x9C\x01\x00\x24\x59"
-     "\x00\x80"
-     "\xA5\x71\xFE\x00\xAA\x5D\xFE\x00\xAB\xE1\xFE\x00\x86\xC8\x00\x00\x87\xFC"
-     "\x01\x00"
-     "\xAE\xE5\xFE\x00\xAF\xED\xFE\x00\xAC\xF5\xFE\x00\xAD\xED\xFE\x00\xA8\x79"
-     "\xFE\x00"
-     "\xA9\x79\xFE\x00\xAA\x85\x01\x00\xAB\x91\x01\x00\xAC\xB9\x01\x00\xAD\xB9"
-     "\x01\x00"
-     "\xAE\xD9\x01\x00\xAF\xD5\x01\x00\x84\xA0\x00\x00\x28\x59\x00\x80\x2C\x59"
-     "\x00\x80"
-     "\x30\x59\x00\x80\x34\x59\x00\x80\x38\x59\x00\x80\x3C\x59\x00\x80\x40\x59"
-     "\x00\x80"
-     "\xB8\x6D\x01\x00\xB9\x75\x01\x00\xBA\x7D\x01\x00\xBB\x75\x01\x00\xBC\x6D"
-     "\x01\x00"
-     "\xBD\xD1\x00\x00\xBE\xD1\x00\x00\xBF\xD1\x00\x00\xB0\xAD\x01\x00\xB1\xB1"
-     "\x01\x00"
-     "\xB2\xB1\x01\x00\xB3\x85\x01\x00\xB4\x9D\x01\x00\xB5\x55\x01\x00\xB6\x5D"
-     "\x01\x00"
-     "\xB7\x55\x01\x00\xB3\x0D\x02\x00\x44\x59\x00\x80\x48\x59\x00\x80\x4C\x59"
-     "\x00\x80"
-     "\x50\x59\x00\x80\xB6\x39\x02\x00\xB5\x31\x02\x00\x54\x59\x00\x80\xBB\x75"
-     "\x02\x00"
-     "\xBA\x69\x02\x00\x58\x59\x00\x80\x5C\x59\x00\x80\xBF\x91\x02\x00\xBE\x59"
-     "\x02\x00"
-     "\xBD\x51\x02\x00\xBC\x65\x02\x00\x60\x59\x00\x80\xA3\x49\x02\x00\x64\x59"
-     "\x00\x80"
-     "\x68\x59\x00\x80\xA6\x7D\x02\x00\x6C\x59\x00\x80\x70\x59\x00\x80\xA5\x75"
-     "\x02\x00"
-     "\xAA\x2D\x02\x00\xAB\x31\x02\x00\x74\x59\x00\x80\x78\x59\x00\x80\xAE\x1D"
-     "\x02\x00"
-     "\xAF\xD5\x02\x00\xAC\x21\x02\x00\xAD\x15\x02\x00\x7C\x59\x00\x80\x80\x59"
-     "\x00\x80"
-     "\x84\x59\x00\x80\x88\x59\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39"
-     "\x00\x00"
-     "\x8C\x59\x00\x80\x90\x59\x00\x80\xBE\x00\x04\x00\x98\x59\x00\x80\x9C\x59"
-     "\x00\x80"
-     "\xA0\x59\x00\x80\xA4\x59\x00\x80\xA8\x59\x00\x80\xAC\x59\x00\x80\x86\xE0"
-     "\x04\x00"
-     "\x87\x04\x03\x00\xBE\x44\x05\x00\x84\x40\x04\x00\xB0\x59\x00\x80\xE3\x64"
-     "\x01\x00"
-     "\xB4\x59\x00\x80\xE1\xAC\x01\x00\xB8\x59\x00\x80\xEF\xB8\x02\x00\xBC\x59"
-     "\x00\x80"
-     "\xC0\x59\x00\x80\xC4\x59\x00\x80\xC8\x59\x00\x80\xCC\x59\x00\x80\xD0\x59"
-     "\x00\x80"
-     "\xA8\xA5\x02\x00\xA9\xAD\x02\x00\xAA\xA5\x02\x00\xAB\xBD\x02\x00\xAC\xA5"
-     "\x02\x00"
-     "\xAD\xAD\x02\x00\xAE\xA5\x02\x00\xAF\x19\x01\x00\x84\xA0\x04\x00\xD4\x59"
-     "\x00\x80"
-     "\xD8\x59\x00\x80\xDC\x59\x00\x80\xE0\x59\x00\x80\xE4\x59\x00\x80\xE8\x59"
-     "\x00\x80"
-     "\xBE\xB0\x07\x00\xB8\x0D\x01\x00\xB9\x11\x01\x00\xBA\x11\x01\x00\xBB\x25"
-     "\x01\x00"
-     "\xBC\x3D\x01\x00\xBD\xD5\x01\x00\xBE\xDD\x01\x00\xBF\xD5\x01\x00\xB0\x69"
-     "\x01\x00"
-     "\xB1\x69\x01\x00\xB2\x79\x01\x00\xB3\x71\x01\x00\xB4\x59\x01\x00\xB5\x59"
-     "\x01\x00"
-     "\xB6\x39\x01\x00\xB7\x35\x01\x00\xE1\x70\xFF\x00\xE3\xB4\xFF\x00\xE3\x1C"
-     "\xFF\x00"
-     "\xE1\xFC\xFC\x00\xEC\x59\x00\x80\xE1\x44\xFE\x00\xF0\x59\x00\x80\xE3\xA0"
-     "\x00\x00"
-     "\xF4\x59\x00\x80\xF8\x59\x00\x80\xFC\x59\x00\x80\xEF\x00\x00\x00\xBE\xFC"
-     "\x06\x00"
-     "\xEF\x98\xFF\x00\xEF\xFC\xFE\x00\x00\x5A\x00\x80\x80\x65\x00\x00\x81\x6D"
-     "\x00\x00"
-     "\x82\x75\x00\x00\xB3\x61\x02\x00\x08\x5A\x00\x80\xB5\x61\x02\x00\xB6\x61"
-     "\x02\x00"
-     "\x0C\x5A\x00\x80\x86\xC0\x04\x00\x87\xA4\x04\x00\xBA\xE1\x01\x00\xBB\xE1"
-     "\x01\x00"
-     "\xBC\xE1\x01\x00\xBD\xE9\x01\x00\xBE\xD1\x01\x00\xBF\xD1\x01\x00\x94\x59"
-     "\x00\x80"
-     "\x04\x5A\x00\x80\xAB\x05\x06\x00\xAA\x05\x06\x00\xAD\x0D\x06\x00\xAC\x05"
-     "\x06\x00"
-     "\xAF\x35\x06\x00\xAE\x35\x06\x00\x10\x5A\x00\x80\x14\x5A\x00\x80\xA3\x85"
-     "\x05\x00"
-     "\x18\x5A\x00\x80\xA5\x85\x05\x00\x1C\x5A\x00\x80\x20\x5A\x00\x80\xA6\x85"
-     "\x05\x00"
-     "\xB3\xF1\x06\x00\x24\x5A\x00\x80\x28\x5A\x00\x80\x2C\x5A\x00\x80\x30\x5A"
-     "\x00\x80"
-     "\xB6\xE9\x06\x00\xB5\xE1\x06\x00\x34\x5A\x00\x80\xBB\x89\x06\x00\xBA\x85"
-     "\x06\x00"
-     "\x38\x5A\x00\x80\x3C\x5A\x00\x80\xBF\x89\x06\x00\xBE\x81\x06\x00\xBD\x89"
-     "\x06\x00"
-     "\xBC\x91\x06\x00\xA8\x29\x06\x00\xA9\x29\x06\x00\xAA\x39\x06\x00\xAB\x39"
-     "\x06\x00"
-     "\xAC\x29\x06\x00\xAD\x29\x06\x00\xAE\x49\x06\x00\xAF\x45\x06\x00\x40\x5A"
-     "\x00\x80"
-     "\x44\x5A\x00\x80\x48\x5A\x00\x80\x4C\x5A\x00\x80\x50\x5A\x00\x80\x54\x5A"
-     "\x00\x80"
-     "\x58\x5A\x00\x80\x5C\x5A\x00\x80\xB8\xE9\x07\x00\xB9\xE9\x07\x00\xBA\xF9"
-     "\x07\x00"
-     "\xBB\xF9\x07\x00\xBC\xE9\x07\x00\xBD\x45\x07\x00\xBE\x41\x07\x00\xBF\x7D"
-     "\x07\x00"
-     "\xB0\x01\x06\x00\xB1\x0D\x06\x00\xB2\x05\x06\x00\xB3\x1D\x06\x00\xB4\x05"
-     "\x06\x00"
-     "\xB5\x0D\x06\x00\xB6\x05\x06\x00\xB7\xD9\x07\x00\xA3\xB1\x07\x00\x60\x5A"
-     "\x00\x80"
-     "\x64\x5A\x00\x80\x68\x5A\x00\x80\x6C\x5A\x00\x80\xA6\xA9\x07\x00\xA5\xA1"
-     "\x07\x00"
-     "\x70\x5A\x00\x80\xAB\xC9\x07\x00\xAA\xC5\x07\x00\x74\x5A\x00\x80\x78\x5A"
-     "\x00\x80"
-     "\xAF\xC9\x07\x00\xAE\xC1\x07\x00\xAD\xC9\x07\x00\xAC\xD1\x07\x00\x80\x4D"
-     "\x00\x00"
-     "\x81\x55\x00\x00\x82\x55\x00\x00\xB3\x6D\x01\x00\x7C\x5A\x00\x80\xB5\x79"
-     "\x01\x00"
-     "\xB6\x6D\x01\x00\x80\x5A\x00\x80\x86\xC0\x00\x00\x87\x04\x01\x00\xBA\x21"
-     "\x01\x00"
-     "\xBB\x21\x01\x00\xBC\x21\x01\x00\xBD\x29\x01\x00\xBE\x11\x01\x00\xBF\x11"
-     "\x01\x00"
-     "\xA8\x51\x06\x00\xA9\x59\x06\x00\xAA\xA9\x01\x00\xAB\xA9\x01\x00\xAC\xBD"
-     "\x01\x00"
-     "\xAD\xA1\x01\x00\xAE\xA1\x01\x00\xAF\xD9\x01\x00\x84\xAC\x01\x00\x84\x5A"
-     "\x00\x80"
-     "\x88\x5A\x00\x80\x8C\x5A\x00\x80\x90\x5A\x00\x80\x94\x5A\x00\x80\x98\x5A"
-     "\x00\x80"
-     "\x9C\x5A\x00\x80\xB8\x69\x01\x00\xB9\x69\x01\x00\xBA\x79\x01\x00\xBB\x79"
-     "\x01\x00"
-     "\xBC\x69\x01\x00\xBD\x69\x01\x00\xBE\xDD\x00\x00\xBF\xD5\x00\x00\xB0\xA9"
-     "\x01\x00"
-     "\xB1\xA9\x01\x00\xB2\xB9\x01\x00\xB3\xB1\x01\x00\xB4\x99\x01\x00\xB5\x99"
-     "\x01\x00"
-     "\xB6\x59\x01\x00\xB7\x59\x01\x00\xA3\x21\x02\x00\xA0\x5A\x00\x80\xA4\x5A"
-     "\x00\x80"
-     "\xA8\x5A\x00\x80\xAC\x5A\x00\x80\xA6\x21\x02\x00\xA5\x35\x02\x00\xB0\x5A"
-     "\x00\x80"
-     "\xAB\x6D\x02\x00\xAA\x6D\x02\x00\xB4\x5A\x00\x80\xB8\x5A\x00\x80\xAF\x5D"
-     "\x02\x00"
-     "\xAE\x5D\x02\x00\xAD\x65\x02\x00\xAC\x6D\x02\x00\xBC\x5A\x00\x80\xC0\x5A"
-     "\x00\x80"
-     "\xC4\x5A\x00\x80\xC8\x5A\x00\x80\xCC\x5A\x00\x80\xD0\x5A\x00\x80\xD4\x5A"
-     "\x00\x80"
-     "\xD8\x5A\x00\x80\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\xDC\x5A"
-     "\x00\x80"
-     "\xE0\x5A\x00\x80\xE8\x5A\x00\x80\x87\x44\x03\x00\x86\xDC\x0C\x00\xEC\x5A"
-     "\x00\x80"
-     "\xF0\x5A\x00\x80\xF4\x5A\x00\x80\x84\xC0\x0C\x00\xF8\x5A\x00\x80\xEF\xDC"
-     "\x07\x00"
-     "\xFC\x5A\x00\x80\x00\x5B\x00\x80\x04\x5B\x00\x80\xE3\x14\x01\x00\x08\x5B"
-     "\x00\x80"
-     "\xE1\xD4\x01\x00\x0C\x5B\x00\x80\x10\x5B\x00\x80\x14\x5B\x00\x80\x18\x5B"
-     "\x00\x80"
-     "\xB3\x9D\x03\x00\x1C\x5B\x00\x80\x20\x5B\x00\x80\x24\x5B\x00\x80\x28\x5B"
-     "\x00\x80"
-     "\xB6\x79\x03\x00\xB5\x71\x03\x00\x2C\x5B\x00\x80\xBB\x5D\x03\x00\xBA\x5D"
-     "\x03\x00"
-     "\x30\x5B\x00\x80\x34\x5B\x00\x80\xBF\xE1\x00\x00\xBE\xF9\x00\x00\xBD\xF1"
-     "\x00\x00"
-     "\xBC\xF9\x00\x00\xA8\x2D\x02\x00\xA9\x55\x02\x00\xAA\x55\x02\x00\xAB\xA9"
-     "\x02\x00"
-     "\xAC\xB9\x02\x00\xAD\xB9\x02\x00\xAE\xA9\x02\x00\xAF\xA9\x02\x00\xBE\x68"
-     "\x0D\x00"
-     "\x38\x5B\x00\x80\x3C\x5B\x00\x80\x40\x5B\x00\x80\x44\x5B\x00\x80\x82\xAD"
-     "\x00\x00"
-     "\x81\xAD\x00\x00\x80\xAD\x00\x00\xB8\x6D\x01\x00\xB9\x05\x01\x00\xBA\x0D"
-     "\x01\x00"
-     "\xBB\x05\x01\x00\xBC\x1D\x01\x00\xBD\x05\x01\x00\xBE\x0D\x01\x00\xBF\x05"
-     "\x01\x00"
-     "\xB0\xD9\x02\x00\xB1\xD9\x02\x00\xB2\x6D\x01\x00\xB3\x65\x01\x00\xB4\x7D"
-     "\x01\x00"
-     "\xB5\x65\x01\x00\xB6\x65\x01\x00\xB7\x55\x01\x00\xE1\x78\x06\x00\xE1\x5C"
-     "\x07\x00"
-     "\xE3\xB8\x06\x00\xE3\x50\x07\x00\x48\x5B\x00\x80\x4C\x5B\x00\x80\x50\x5B"
-     "\x00\x80"
-     "\x58\x5B\x00\x80\x5C\x5B\x00\x80\x60\x5B\x00\x80\x64\x5B\x00\x80\x68\x5B"
-     "\x00\x80"
-     "\xBE\xF8\x0C\x00\x6C\x5B\x00\x80\xEF\x7C\x06\x00\xEF\x8C\x07\x00\xA3\x9D"
-     "\x02\x00"
-     "\x70\x5B\x00\x80\x86\x68\x0C\x00\x87\x04\x0D\x00\x74\x5B\x00\x80\xA6\x79"
-     "\x02\x00"
-     "\xA5\x71\x02\x00\x78\x5B\x00\x80\xAB\x5D\x02\x00\xAA\x5D\x02\x00\x7C\x5B"
-     "\x00\x80"
-     "\x80\x5B\x00\x80\xAF\xE1\x01\x00\xAE\xF9\x01\x00\xAD\xF1\x01\x00\xAC\xF9"
-     "\x01\x00"
-     "\xA8\x75\x0E\x00\xA9\x85\x0E\x00\xAA\x8D\x0E\x00\xAB\x85\x0E\x00\xAC\x81"
-     "\x0E\x00"
-     "\xAD\x8D\x0E\x00\xAE\x85\x0E\x00\xAF\xBD\x0E\x00\x54\x5B\x00\x80\x84\x5B"
-     "\x00\x80"
-     "\x88\x5B\x00\x80\xE4\x5A\x00\x80\x8C\x5B\x00\x80\x90\x5B\x00\x80\x94\x5B"
-     "\x00\x80"
-     "\x98\x5B\x00\x80\xB8\x91\x0E\x00\xB9\x99\x0E\x00\xBA\xA1\x0E\x00\xBB\xA1"
-     "\x0E\x00"
-     "\xBC\x75\x0F\x00\xBD\x7D\x0F\x00\xBE\x75\x0F\x00\xBF\x6D\x0F\x00\xB0\xC9"
-     "\x0E\x00"
-     "\xB1\xD5\x0E\x00\xB2\xDD\x0E\x00\xB3\xD5\x0E\x00\xB4\xCD\x0E\x00\xB5\xB1"
-     "\x0E\x00"
-     "\xB6\xB1\x0E\x00\xB7\xB1\x0E\x00\xB3\x05\x0E\x00\x9C\x5B\x00\x80\xA0\x5B"
-     "\x00\x80"
-     "\xA4\x5B\x00\x80\xA8\x5B\x00\x80\xB6\x0D\x0E\x00\xB5\x09\x0E\x00\xAC\x5B"
-     "\x00\x80"
-     "\xBB\x11\x0E\x00\xBA\x0D\x0E\x00\xB0\x5B\x00\x80\xB4\x5B\x00\x80\xBF\x75"
-     "\x0E\x00"
-     "\xBE\x09\x0E\x00\xBD\x01\x0E\x00\xBC\x09\x0E\x00\x82\x29\x00\x00\xA3\x41"
-     "\x0E\x00"
-     "\x80\x1D\x00\x00\x81\x15\x00\x00\xA6\x49\x0E\x00\xB8\x5B\x00\x80\xBC\x5B"
-     "\x00\x80"
-     "\xA5\x4D\x0E\x00\xAA\x49\x0E\x00\xAB\x55\x0E\x00\xC0\x5B\x00\x80\x84\xA0"
-     "\x03\x00"
-     "\xAE\x4D\x0E\x00\xAF\x31\x0E\x00\xAC\x4D\x0E\x00\xAD\x45\x0E\x00\xA8\x9D"
-     "\x0E\x00"
-     "\xA9\xA5\x0E\x00\xAA\xAD\x0E\x00\xAB\xA5\x0E\x00\xAC\xBD\x0E\x00\xAD\xA1"
-     "\x0E\x00"
-     "\xAE\xDD\x0E\x00\xAF\xCD\x0E\x00\xC4\x5B\x00\x80\xC8\x5B\x00\x80\x86\x88"
-     "\x01\x00"
-     "\x87\xAC\x01\x00\xCC\x5B\x00\x80\xD0\x5B\x00\x80\xD4\x5B\x00\x80\xD8\x5B"
-     "\x00\x80"
-     "\xB8\x7D\x01\x00\xB9\xC1\x01\x00\xBA\xC1\x01\x00\xBB\xC1\x01\x00\xBC\xC1"
-     "\x01\x00"
-     "\xBD\xC9\x01\x00\xBE\xF1\x01\x00\xBF\xF1\x01\x00\xB0\xBD\x0E\x00\xB1\x8D"
-     "\x0E\x00"
-     "\xB2\x85\x0E\x00\xB3\x49\x01\x00\xB4\x59\x01\x00\xB5\x45\x01\x00\xB6\x4D"
-     "\x01\x00"
-     "\xB7\x45\x01\x00\xB3\x25\x0E\x00\xDC\x5B\x00\x80\xE0\x5B\x00\x80\xE4\x5B"
-     "\x00\x80"
-     "\xE8\x5B\x00\x80\xB6\x21\x0E\x00\xB5\x35\x0E\x00\xEC\x5B\x00\x80\xBB\x49"
-     "\x0E\x00"
-     "\xBA\x79\x0E\x00\xF0\x5B\x00\x80\xF4\x5B\x00\x80\xBF\x81\x01\x00\xBE\x9D"
-     "\x01\x00"
-     "\xBD\x9D\x01\x00\xBC\x51\x0E\x00\xF8\x5B\x00\x80\xA3\x61\x0E\x00\xFC\x5B"
-     "\x00\x80"
-     "\x00\x5C\x00\x80\xA6\x65\x0E\x00\x04\x5C\x00\x80\x08\x5C\x00\x80\xA5\x71"
-     "\x0E\x00"
-     "\xAA\x3D\x0E\x00\xAB\x0D\x0E\x00\x0C\x5C\x00\x80\x10\x5C\x00\x80\xAE\xD9"
-     "\x01\x00"
-     "\xAF\xC5\x01\x00\xAC\x15\x0E\x00\xAD\xD9\x01\x00\xA8\x25\x03\x00\xA9\x35"
-     "\x03\x00"
-     "\xAA\x3D\x03\x00\xAB\x35\x03\x00\xAC\x2D\x03\x00\xAD\x91\x03\x00\xAE\x91"
-     "\x03\x00"
-     "\xAF\x91\x03\x00\x14\x5C\x00\x80\x18\x5C\x00\x80\x1C\x5C\x00\x80\x20\x5C"
-     "\x00\x80"
-     "\x24\x5C\x00\x80\x28\x5C\x00\x80\x2C\x5C\x00\x80\x30\x5C\x00\x80\xB8\x99"
-     "\x03\x00"
-     "\xB9\xAD\x03\x00\xBA\xA5\x03\x00\xBB\x6D\x00\x00\xBC\x75\x00\x00\xBD\x7D"
-     "\x00\x00"
-     "\xBE\x75\x00\x00\xBF\x6D\x00\x00\xB0\xF1\x03\x00\xB1\xF9\x03\x00\xB2\xC1"
-     "\x03\x00"
-     "\xB3\xC1\x03\x00\xB4\xB1\x03\x00\xB5\xBD\x03\x00\xB6\xB5\x03\x00\xB7\xA9"
-     "\x03\x00"
-     "\x80\xB5\x00\x00\x81\xBD\x00\x00\x82\xB5\x00\x00\xBE\x0C\x03\x00\x38\x5C"
-     "\x00\x80"
-     "\x3C\x5C\x00\x80\x40\x5C\x00\x80\xEF\xE8\x02\x00\x84\xAC\x02\x00\xE1\x94"
-     "\x01\x00"
-     "\x44\x5C\x00\x80\xE3\x1C\x02\x00\x48\x5C\x00\x80\x4C\x5C\x00\x80\x87\xD8"
-     "\x03\x00"
-     "\x86\x1C\x04\x00\xB3\xF5\x03\x00\x50\x5C\x00\x80\x85\xA4\x1F\x00\x54\x5C"
-     "\x00\x80"
-     "\x58\x5C\x00\x80\xB6\x51\x03\x00\xB5\xE5\x03\x00\x5C\x5C\x00\x80\xBB\x7D"
-     "\x03\x00"
-     "\xBA\x75\x03\x00\x60\x5C\x00\x80\x64\x5C\x00\x80\xBF\x39\x03\x00\xBE\x35"
-     "\x03\x00"
-     "\xBD\x55\x03\x00\xBC\x55\x03\x00\x68\x5C\x00\x80\x6C\x5C\x00\x80\x70\x5C"
-     "\x00\x80"
-     "\xA3\x39\x02\x00\x74\x5C\x00\x80\xA5\x29\x02\x00\xA6\x9D\x02\x00\x78\x5C"
-     "\x00\x80"
-     "\x7C\x5C\x00\x80\x80\x5C\x00\x80\xAA\xB9\x02\x00\xAB\xB1\x02\x00\xAC\x99"
-     "\x02\x00"
-     "\xAD\x99\x02\x00\xAE\xF9\x02\x00\xAF\xF5\x02\x00\x84\x5C\x00\x80\x88\x5C"
-     "\x00\x80"
-     "\x84\x44\x02\x00\xBE\x40\x05\x00\x8C\x5C\x00\x80\x90\x5C\x00\x80\x94\x5C"
-     "\x00\x80"
-     "\x98\x5C\x00\x80\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x9C\x5C"
-     "\x00\x80"
-     "\xA0\x5C\x00\x80\xA8\x5C\x00\x80\x87\xF4\x05\x00\x86\xDC\x04\x00\xAC\x5C"
-     "\x00\x80"
-     "\xB0\x5C\x00\x80\xB4\x5C\x00\x80\xB8\x5C\x00\x80\xEF\x48\x0E\x00\xBC\x5C"
-     "\x00\x80"
-     "\xC0\x5C\x00\x80\xEF\xD4\x0E\x00\xE3\x3C\x0F\x00\xE1\x74\x0E\x00\xE1\xCC"
-     "\x0F\x00"
-     "\xE3\x50\x01\x00\xC4\x5C\x00\x80\xC8\x5C\x00\x80\xCC\x5C\x00\x80\xD0\x5C"
-     "\x00\x80"
-     "\xB3\x59\x01\x00\xD4\x5C\x00\x80\xD8\x5C\x00\x80\xDC\x5C\x00\x80\xE0\x5C"
-     "\x00\x80"
-     "\xB6\x61\x01\x00\xB5\x69\x01\x00\xE4\x5C\x00\x80\xBB\x3D\x01\x00\xBA\x35"
-     "\x01\x00"
-     "\xE8\x5C\x00\x80\xEC\x5C\x00\x80\xBF\xF5\x00\x00\xBE\xF5\x00\x00\xBD\x15"
-     "\x01\x00"
-     "\xBC\x15\x01\x00\xA8\x91\x06\x00\xA9\x91\x06\x00\xAA\x91\x06\x00\xAB\xA5"
-     "\x06\x00"
-     "\xAC\xA1\x06\x00\xAD\xA1\x06\x00\xAE\xDD\x06\x00\xAF\xD1\x06\x00\xA4\x5C"
-     "\x00\x80"
-     "\xF0\x5C\x00\x80\xF4\x5C\x00\x80\xF8\x5C\x00\x80\x34\x5C\x00\x80\xFC\x5C"
-     "\x00\x80"
-     "\x00\x5D\x00\x80\x04\x5D\x00\x80\xB8\x5D\x07\x00\xB9\x65\x07\x00\xBA\x6D"
-     "\x07\x00"
-     "\xBB\x0D\x07\x00\xBC\x15\x07\x00\xBD\x05\x07\x00\xBE\x0D\x07\x00\xBF\x05"
-     "\x07\x00"
-     "\xB0\xB1\x06\x00\xB1\xB9\x06\x00\xB2\x81\x06\x00\xB3\x81\x06\x00\xB4\x71"
-     "\x07\x00"
-     "\xB5\x71\x07\x00\xB6\x71\x07\x00\xB7\x6D\x07\x00\xA3\x19\x06\x00\x82\x2D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x80\x1D\x00\x00\x08\x5D\x00\x80\xA6\x21\x06\x00\xA5\x29"
-     "\x06\x00"
-     "\x0C\x5D\x00\x80\xAB\x7D\x06\x00\xAA\x75\x06\x00\x10\x5D\x00\x80\x84\xA4"
-     "\x02\x00"
-     "\xAF\xB5\x07\x00\xAE\xB5\x07\x00\xAD\x55\x06\x00\xAC\x55\x06\x00\xBE\xCC"
-     "\x03\x00"
-     "\xB3\x41\x07\x00\x14\x5D\x00\x80\x18\x5D\x00\x80\xB6\x45\x07\x00\x1C\x5D"
-     "\x00\x80"
-     "\x20\x5D\x00\x80\xB5\x51\x07\x00\xBA\x55\x07\x00\xBB\x5D\x07\x00\x86\x48"
-     "\x00\x00"
-     "\x87\x0C\x03\x00\xBE\x39\x07\x00\xBF\x39\x07\x00\xBC\x45\x07\x00\xBD\x39"
-     "\x07\x00"
-     "\xA8\x75\x06\x00\xA9\x85\x06\x00\xAA\x8D\x06\x00\xAB\x85\x06\x00\xAC\x9D"
-     "\x06\x00"
-     "\xAD\x81\x06\x00\xAE\x81\x06\x00\xAF\xAD\x06\x00\x24\x5D\x00\x80\x28\x5D"
-     "\x00\x80"
-     "\x2C\x5D\x00\x80\x30\x5D\x00\x80\x34\x5D\x00\x80\x38\x5D\x00\x80\x3C\x5D"
-     "\x00\x80"
-     "\x40\x5D\x00\x80\xB8\xA9\x06\x00\xB9\xA9\x06\x00\xBA\x6D\x01\x00\xBB\x65"
-     "\x01\x00"
-     "\xBC\x7D\x01\x00\xBD\x61\x01\x00\xBE\x61\x01\x00\xBF\x61\x01\x00\xB0\xD5"
-     "\x06\x00"
-     "\xB1\xDD\x06\x00\xB2\xD5\x06\x00\xB3\xA9\x06\x00\xB4\xB9\x06\x00\xB5\xB9"
-     "\x06\x00"
-     "\xB6\xA9\x06\x00\xB7\xA1\x06\x00\xA3\x05\x06\x00\x44\x5D\x00\x80\x48\x5D"
-     "\x00\x80"
-     "\x4C\x5D\x00\x80\x50\x5D\x00\x80\xA6\x01\x06\x00\xA5\x15\x06\x00\x54\x5D"
-     "\x00\x80"
-     "\xAB\x19\x06\x00\xAA\x11\x06\x00\x58\x5D\x00\x80\x5C\x5D\x00\x80\xAF\x7D"
-     "\x06\x00"
-     "\xAE\x7D\x06\x00\xAD\x7D\x06\x00\xAC\x01\x06\x00\x60\x5D\x00\x80\x64\x5D"
-     "\x00\x80"
-     "\x68\x5D\x00\x80\x6C\x5D\x00\x80\x70\x5D\x00\x80\x74\x5D\x00\x80\x78\x5D"
-     "\x00\x80"
-     "\x7C\x5D\x00\x80\x80\x39\x00\x00\x81\x39\x00\x00\x82\x05\x00\x00\x80\x5D"
-     "\x00\x80"
-     "\xBE\xB8\x01\x00\x84\xB8\x01\x00\x88\x5D\x00\x80\x8C\x5D\x00\x80\xA8\xCD"
-     "\x02\x00"
-     "\xA9\xD1\x02\x00\xAA\xD1\x02\x00\xAB\x2D\x03\x00\xAC\x35\x03\x00\xAD\x3D"
-     "\x03\x00"
-     "\xAE\x35\x03\x00\xAF\x2D\x03\x00\x90\x5D\x00\x80\x94\x5D\x00\x80\x98\x5D"
-     "\x00\x80"
-     "\x9C\x5D\x00\x80\xA0\x5D\x00\x80\xA4\x5D\x00\x80\xBE\x34\x1C\x00\xA8\x5D"
-     "\x00\x80"
-     "\xB8\xE9\x03\x00\xB9\xE9\x03\x00\xBA\x89\x03\x00\xBB\x89\x03\x00\xBC\x99"
-     "\x03\x00"
-     "\xBD\x99\x03\x00\xBE\x89\x03\x00\xBF\x89\x03\x00\xB0\x55\x03\x00\xB1\x5D"
-     "\x03\x00"
-     "\xB2\x55\x03\x00\xB3\xE9\x03\x00\xB4\xF9\x03\x00\xB5\xF9\x03\x00\xB6\xE9"
-     "\x03\x00"
-     "\xB7\xE1\x03\x00\xB3\x59\x02\x00\xAC\x5D\x00\x80\x87\x94\x03\x00\x86\x4C"
-     "\x1C\x00"
-     "\xB0\x5D\x00\x80\xB6\xBD\x03\x00\xB5\xBD\x03\x00\xB4\x5D\x00\x80\xBB\x99"
-     "\x03\x00"
-     "\xBA\x91\x03\x00\xB8\x5D\x00\x80\xBC\x5D\x00\x80\xBF\x3D\x03\x00\xBE\x3D"
-     "\x03\x00"
-     "\xBD\x3D\x03\x00\xBC\x81\x03\x00\xC0\x5D\x00\x80\xA3\x1D\x02\x00\xC4\x5D"
-     "\x00\x80"
-     "\xC8\x5D\x00\x80\xA6\xF9\x03\x00\xCC\x5D\x00\x80\xD0\x5D\x00\x80\xA5\xF9"
-     "\x03\x00"
-     "\xAA\xD5\x03\x00\xAB\xDD\x03\x00\xD4\x5D\x00\x80\xD8\x5D\x00\x80\xAE\x79"
-     "\x03\x00"
-     "\xAF\x79\x03\x00\xAC\xC5\x03\x00\xAD\x79\x03\x00\xDC\x5D\x00\x80\xE0\x5D"
-     "\x00\x80"
-     "\x84\x04\x02\x00\xBE\x00\x1C\x00\xE4\x5D\x00\x80\x82\x3D\x00\x00\x81\x3D"
-     "\x00\x00"
-     "\x80\x3D\x00\x00\xE8\x5D\x00\x80\xEC\x5D\x00\x80\xF0\x5D\x00\x80\xF8\x5D"
-     "\x00\x80"
-     "\xFC\x5D\x00\x80\x00\x5E\x00\x80\x04\x5E\x00\x80\x08\x5E\x00\x80\x0C\x5E"
-     "\x00\x80"
-     "\xE3\xA0\x00\x00\x10\x5E\x00\x80\xE1\xBC\x01\x00\x86\xF0\x1C\x00\x87\x04"
-     "\x1D\x00"
-     "\x14\x5E\x00\x80\x18\x5E\x00\x80\x1C\x5E\x00\x80\x20\x5E\x00\x80\x24\x5E"
-     "\x00\x80"
-     "\x28\x5E\x00\x80\x2C\x5E\x00\x80\xEF\x7C\x00\x00\x30\x5E\x00\x80\x34\x5E"
-     "\x00\x80"
-     "\x38\x5E\x00\x80\x3C\x5E\x00\x80\x40\x5E\x00\x80\xEF\x38\x06\x00\x44\x5E"
-     "\x00\x80"
-     "\x48\x5E\x00\x80\x4C\x5E\x00\x80\xEF\xA4\x01\x00\x84\xA8\x1D\x00\xE1\xC0"
-     "\x06\x00"
-     "\x50\x5E\x00\x80\xE3\x10\x01\x00\x54\x5E\x00\x80\xE1\x10\x07\x00\x58\x5E"
-     "\x00\x80"
-     "\xE3\xE4\x06\x00\xB3\xE1\x01\x00\x5C\x5E\x00\x80\x60\x5E\x00\x80\x68\x5E"
-     "\x00\x80"
-     "\x6C\x5E\x00\x80\xB6\xF9\x01\x00\xB5\xF1\x01\x00\x70\x5E\x00\x80\xBB\x65"
-     "\x01\x00"
-     "\xBA\xD9\x01\x00\x86\x48\x1F\x00\x87\xD0\x1C\x00\xBF\x5D\x01\x00\xBE\x5D"
-     "\x01\x00"
-     "\xBD\x65\x01\x00\xBC\x6D\x01\x00\xB0\x3D\x1E\x00\xB1\x81\x1E\x00\xB2\x99"
-     "\x1E\x00"
-     "\xB3\x91\x1E\x00\xB4\xB9\x1E\x00\xB5\xB9\x1E\x00\xB6\xD9\x1E\x00\xB7\xD5"
-     "\x1E\x00"
-     "\xB8\xED\x1E\x00\xB9\xF1\x1E\x00\xBA\xF1\x1E\x00\xBB\xC5\x1E\x00\xBC\xDD"
-     "\x1E\x00"
-     "\xBD\x35\x1F\x00\xBE\x3D\x1F\x00\xBF\x2D\x1F\x00\xF4\x5D\x00\x80\x84\x5D"
-     "\x00\x80"
-     "\x74\x5E\x00\x80\x78\x5E\x00\x80\x80\xE9\x1F\x00\x81\x91\x1F\x00\x82\x91"
-     "\x1F\x00"
-     "\x64\x5E\x00\x80\xA8\x99\x1D\x00\xA9\x99\x1D\x00\xAA\x49\x1E\x00\xAB\x49"
-     "\x1E\x00"
-     "\xAC\x59\x1E\x00\xAD\x59\x1E\x00\xAE\x49\x1E\x00\xAF\x49\x1E\x00\xA3\xA1"
-     "\x1E\x00"
-     "\x7C\x5E\x00\x80\x80\x5E\x00\x80\x84\x5E\x00\x80\x88\x5E\x00\x80\xA6\xB9"
-     "\x1E\x00"
-     "\xA5\xB1\x1E\x00\x8C\x5E\x00\x80\xAB\x25\x1E\x00\xAA\x99\x1E\x00\x90\x5E"
-     "\x00\x80"
-     "\x94\x5E\x00\x80\xAF\x1D\x1E\x00\xAE\x1D\x1E\x00\xAD\x25\x1E\x00\xAC\x2D"
-     "\x1E\x00"
-     "\x98\x5E\x00\x80\xB3\xF1\x1F\x00\x9C\x5E\x00\x80\xA0\x5E\x00\x80\xB6\x91"
-     "\x1F\x00"
-     "\xA4\x5E\x00\x80\xA8\x5E\x00\x80\xB5\xE1\x1F\x00\xBA\xBD\x1F\x00\xBB\x85"
-     "\x1F\x00"
-     "\xAC\x5E\x00\x80\xB0\x5E\x00\x80\xBE\x9D\x1F\x00\xBF\x85\x1F\x00\xBC\x9D"
-     "\x1F\x00"
-     "\xBD\x95\x1F\x00\xA8\x2D\x1E\x00\xA9\x35\x1E\x00\xAA\x3D\x1E\x00\xAB\x35"
-     "\x1E\x00"
-     "\xAC\x2D\x1E\x00\xAD\x55\x1E\x00\xAE\x51\x1E\x00\xAF\x45\x1E\x00\xB4\x5E"
-     "\x00\x80"
-     "\xB8\x5E\x00\x80\xBC\x5E\x00\x80\xC0\x5E\x00\x80\xC4\x5E\x00\x80\x82\xB9"
-     "\x00\x00"
-     "\x81\xB1\x00\x00\x80\xB9\x00\x00\xB8\xC1\x1E\x00\xB9\xC1\x1E\x00\xBA\xC1"
-     "\x1E\x00"
-     "\xBB\xDD\x1E\x00\xBC\xC5\x1E\x00\xBD\xCD\x1E\x00\xBE\xF1\x1E\x00\xBF\xED"
-     "\x1E\x00"
-     "\xB0\x3D\x1E\x00\xB1\x05\x1E\x00\xB2\x0D\x1E\x00\xB3\x05\x1E\x00\xB4\x1D"
-     "\x1E\x00"
-     "\xB5\x05\x1E\x00\xB6\x01\x1E\x00\xB7\x01\x1E\x00\xA3\xB5\x1E\x00\xC8\x5E"
-     "\x00\x80"
-     "\xCC\x5E\x00\x80\xD0\x5E\x00\x80\x84\x98\x02\x00\xA6\xD5\x1E\x00\xA5\xA5"
-     "\x1E\x00"
-     "\xBE\x90\x01\x00\xAB\xC1\x1E\x00\xAA\xF9\x1E\x00\x86\x08\x00\x00\x87\x7C"
-     "\x01\x00"
-     "\xAF\xC1\x1E\x00\xAE\xD9\x1E\x00\xAD\xD1\x1E\x00\xAC\xD9\x1E\x00\xD8\x5E"
-     "\x00\x80"
-     "\xB3\x19\x1E\x00\xDC\x5E\x00\x80\xE0\x5E\x00\x80\xB6\x55\x01\x00\xE4\x5E"
-     "\x00\x80"
-     "\xE8\x5E\x00\x80\xB5\x91\x01\x00\xBA\x71\x01\x00\xBB\x79\x01\x00\xEC\x5E"
-     "\x00\x80"
-     "\xF0\x5E\x00\x80\xBE\x31\x01\x00\xBF\x3D\x01\x00\xBC\x51\x01\x00\xBD\x51"
-     "\x01\x00"
-     "\xA8\xCD\x1D\x00\xA9\xD5\x1D\x00\xAA\xD5\x1D\x00\xAB\x45\x02\x00\xAC\x5D"
-     "\x02\x00"
-     "\xAD\x91\x02\x00\xAE\x91\x02\x00\xAF\x91\x02\x00\xF4\x5E\x00\x80\xF8\x5E"
-     "\x00\x80"
-     "\xFC\x5E\x00\x80\x00\x5F\x00\x80\x04\x5F\x00\x80\x08\x5F\x00\x80\x0C\x5F"
-     "\x00\x80"
-     "\x10\x5F\x00\x80\xB8\x99\x02\x00\xB9\xAD\x02\x00\xBA\xA5\x02\x00\xBB\x6D"
-     "\x03\x00"
-     "\xBC\x75\x03\x00\xBD\x7D\x03\x00\xBE\x75\x03\x00\xBF\x6D\x03\x00\xB0\xF1"
-     "\x02\x00"
-     "\xB1\xF9\x02\x00\xB2\xC1\x02\x00\xB3\xC1\x02\x00\xB4\xB1\x02\x00\xB5\xBD"
-     "\x02\x00"
-     "\xB6\xB5\x02\x00\xB7\xA9\x02\x00\xA3\x5D\x1D\x00\x14\x5F\x00\x80\x18\x5F"
-     "\x00\x80"
-     "\x1C\x5F\x00\x80\x20\x5F\x00\x80\xA6\x11\x02\x00\xA5\xD5\x02\x00\x24\x5F"
-     "\x00\x80"
-     "\xAB\x3D\x02\x00\xAA\x35\x02\x00\x28\x5F\x00\x80\x2C\x5F\x00\x80\xAF\x79"
-     "\x02\x00"
-     "\xAE\x75\x02\x00\xAD\x15\x02\x00\xAC\x15\x02\x00\xBE\x4C\x04\x00\x30\x5F"
-     "\x00\x80"
-     "\x34\x5F\x00\x80\x38\x5F\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39"
-     "\x00\x00"
-     "\x3C\x5F\x00\x80\x40\x5F\x00\x80\x84\xA8\x03\x00\x86\xC8\x04\x00\x87\xA8"
-     "\x04\x00"
-     "\x48\x5F\x00\x80\x4C\x5F\x00\x80\x50\x5F\x00\x80\x54\x5F\x00\x80\xA8\xC1"
-     "\x02\x00"
-     "\xA9\xCD\x02\x00\xAA\xC5\x02\x00\xAB\xD9\x02\x00\xAC\xC9\x02\x00\xAD\xFD"
-     "\x02\x00"
-     "\xAE\xF5\x02\x00\xAF\x3D\x01\x00\x58\x5F\x00\x80\x5C\x5F\x00\x80\x60\x5F"
-     "\x00\x80"
-     "\x64\x5F\x00\x80\x68\x5F\x00\x80\x6C\x5F\x00\x80\x70\x5F\x00\x80\xBE\xF4"
-     "\x04\x00"
-     "\xB8\xC9\x01\x00\xB9\xC9\x01\x00\xBA\xD9\x01\x00\xBB\xD1\x01\x00\xBC\xF9"
-     "\x01\x00"
-     "\xBD\xF9\x01\x00\xBE\x9D\x01\x00\xBF\x95\x01\x00\xB0\x45\x01\x00\xB1\x4D"
-     "\x01\x00"
-     "\xB2\x45\x01\x00\xB3\x5D\x01\x00\xB4\x45\x01\x00\xB5\x4D\x01\x00\xB6\x45"
-     "\x01\x00"
-     "\xB7\xF9\x01\x00\xB3\x4D\x02\x00\x74\x5F\x00\x80\x78\x5F\x00\x80\x7C\x5F"
-     "\x00\x80"
-     "\x80\x5F\x00\x80\xB6\x7D\x02\x00\xB5\x41\x02\x00\x84\x5F\x00\x80\xBB\x89"
-     "\x01\x00"
-     "\xBA\x81\x01\x00\x88\x5F\x00\x80\x8C\x5F\x00\x80\xBF\x89\x01\x00\xBE\x81"
-     "\x01\x00"
-     "\xBD\x89\x01\x00\xBC\x91\x01\x00\xE1\x54\x1F\x00\xE1\x8C\x1E\x00\xE3\xD0"
-     "\x1E\x00"
-     "\xE3\x7C\x1E\x00\xE3\x50\x00\x00\x90\x5F\x00\x80\xE1\x28\x1E\x00\x94\x5F"
-     "\x00\x80"
-     "\xEF\x84\x00\x00\x98\x5F\x00\x80\x9C\x5F\x00\x80\xA0\x5F\x00\x80\xA4\x5F"
-     "\x00\x80"
-     "\xA8\x5F\x00\x80\xEF\x38\x1F\x00\xEF\x08\x1E\x00\xA3\xCD\x05\x00\x44\x5F"
-     "\x00\x80"
-     "\xAC\x5F\x00\x80\xB0\x5F\x00\x80\x84\xDC\x03\x00\xA6\xFD\x05\x00\xA5\xC1"
-     "\x05\x00"
-     "\xB4\x5F\x00\x80\xAB\x09\x06\x00\xAA\x01\x06\x00\x86\x08\x03\x00\x87\x64"
-     "\x03\x00"
-     "\xAF\x09\x06\x00\xAE\x01\x06\x00\xAD\x09\x06\x00\xAC\x11\x06\x00\x81\x49"
-     "\x00\x00"
-     "\x80\xB1\x07\x00\xD4\x5E\x00\x80\x82\x41\x00\x00\xB8\x5F\x00\x80\xBC\x5F"
-     "\x00\x80"
-     "\xC0\x5F\x00\x80\xC4\x5F\x00\x80\xC8\x5F\x00\x80\xCC\x5F\x00\x80\xD0\x5F"
-     "\x00\x80"
-     "\xD4\x5F\x00\x80\xD8\x5F\x00\x80\xDC\x5F\x00\x80\xE0\x5F\x00\x80\xE4\x5F"
-     "\x00\x80"
-     "\xA8\x71\x06\x00\xA9\x71\x06\x00\xAA\x71\x06\x00\xAB\x71\x06\x00\xAC\xB1"
-     "\x07\x00"
-     "\xAD\xB1\x07\x00\xAE\xB1\x07\x00\xAF\xB1\x07\x00\xE8\x5F\x00\x80\xEC\x5F"
-     "\x00\x80"
-     "\xF0\x5F\x00\x80\xF4\x5F\x00\x80\xF8\x5F\x00\x80\xFC\x5F\x00\x80\x00\x60"
-     "\x00\x80"
-     "\x04\x60\x00\x80\xB8\x69\x07\x00\xB9\x69\x07\x00\xBA\x1D\x07\x00\xBB\x11"
-     "\x07\x00"
-     "\xBC\x35\x07\x00\xBD\x3D\x07\x00\xBE\x35\x07\x00\xBF\x2D\x07\x00\xB0\xD1"
-     "\x07\x00"
-     "\xB1\xD1\x07\x00\xB2\xD1\x07\x00\xB3\x69\x07\x00\xB4\x79\x07\x00\xB5\x79"
-     "\x07\x00"
-     "\xB6\x69\x07\x00\xB7\x61\x07\x00\xB3\xF9\x07\x00\x08\x60\x00\x80\x0C\x60"
-     "\x00\x80"
-     "\x10\x60\x00\x80\x14\x60\x00\x80\xB6\x31\x06\x00\xB5\x39\x06\x00\x18\x60"
-     "\x00\x80"
-     "\xBB\x15\x06\x00\xBA\x15\x06\x00\x1C\x60\x00\x80\x20\x60\x00\x80\xBF\xBD"
-     "\x06\x00"
-     "\xBE\xBD\x06\x00\xBD\xBD\x06\x00\xBC\x05\x06\x00\x82\x6D\x00\x00\xA3\xBD"
-     "\x07\x00"
-     "\x80\x55\x00\x00\x81\x65\x00\x00\xA6\x75\x06\x00\x84\x98\x01\x00\xBE\x90"
-     "\x01\x00"
-     "\xA5\x7D\x06\x00\xAA\x51\x06\x00\xAB\x51\x06\x00\x85\x00\x0C\x00\xBE\x60"
-     "\x0F\x00"
-     "\xAE\xF9\x06\x00\xAF\xF9\x06\x00\xAC\x41\x06\x00\xAD\xF9\x06\x00\xA8\xD1"
-     "\x01\x00"
-     "\xA9\xD1\x01\x00\xAA\xD1\x01\x00\xAB\xD1\x01\x00\xAC\xF1\x01\x00\xAD\xF1"
-     "\x01\x00"
-     "\xAE\xF1\x01\x00\xAF\xF1\x01\x00\x86\x20\x0F\x00\x87\xE4\x00\x00\x28\x60"
-     "\x00\x80"
-     "\x2C\x60\x00\x80\x30\x60\x00\x80\x34\x60\x00\x80\x38\x60\x00\x80\x3C\x60"
-     "\x00\x80"
-     "\xB8\x71\x01\x00\xB9\x71\x01\x00\xBA\x71\x01\x00\xBB\x71\x01\x00\xBC\xD5"
-     "\x00\x00"
-     "\xBD\xDD\x00\x00\xBE\xD5\x00\x00\xBF\xCD\x00\x00\xB0\x91\x01\x00\xB1\x91"
-     "\x01\x00"
-     "\xB2\x91\x01\x00\xB3\x91\x01\x00\xB4\x51\x01\x00\xB5\x51\x01\x00\xB6\x51"
-     "\x01\x00"
-     "\xB7\x51\x01\x00\xB3\x59\x02\x00\x40\x60\x00\x80\x44\x60\x00\x80\x48\x60"
-     "\x00\x80"
-     "\x4C\x60\x00\x80\xB6\x71\x02\x00\xB5\x79\x02\x00\x50\x60\x00\x80\xBB\x55"
-     "\x02\x00"
-     "\xBA\x55\x02\x00\x54\x60\x00\x80\x58\x60\x00\x80\xBF\x95\x02\x00\xBE\x95"
-     "\x02\x00"
-     "\xBD\x45\x02\x00\xBC\x45\x02\x00\x5C\x60\x00\x80\xA3\x1D\x02\x00\x60\x60"
-     "\x00\x80"
-     "\x64\x60\x00\x80\xA6\x35\x02\x00\x68\x60\x00\x80\x6C\x60\x00\x80\xA5\x3D"
-     "\x02\x00"
-     "\xAA\x11\x02\x00\xAB\x11\x02\x00\x70\x60\x00\x80\x74\x60\x00\x80\xAE\xD1"
-     "\x02\x00"
-     "\xAF\xD1\x02\x00\xAC\x01\x02\x00\xAD\x01\x02\x00\x82\x85\xEA\x00\x83\xDD"
-     "\xEA\x00"
-     "\x78\x60\x00\x80\x84\x0C\x03\x00\x86\x81\x16\x00\x87\xCD\x17\x00\x84\x91"
-     "\x16\x00"
-     "\x85\x99\x16\x00\x8A\xA1\x12\x00\x8B\xA1\x12\x00\x7C\x60\x00\x80\x80\x60"
-     "\x00\x80"
-     "\x8E\x91\x1E\x00\x8F\x95\x1F\x00\x8C\x91\x13\x00\x8D\x91\x1E\x00\x92\xF1"
-     "\x1A\x00"
-     "\x93\xE5\x1A\x00\x84\x60\x00\x80\xEF\x58\x24\x00\x96\x91\x06\x00\x97\x89"
-     "\x06\x00"
-     "\x94\x11\x1A\x00\x95\x99\x06\x00\x9A\xA9\x02\x00\x9B\xE1\x02\x00\x88\x60"
-     "\x00\x80"
-     "\x8C\x60\x00\x80\x90\x60\x00\x80\xE1\x88\x01\x00\x9C\xE5\x02\x00\xE3\x28"
-     "\x37\x00"
-     "\x9B\x31\x1B\x00\x9A\x25\x1E\x00\x99\x29\x1F\x00\x98\x21\x1F\x00\x9F\x51"
-     "\x07\x00"
-     "\x9E\x51\x07\x00\x9D\xD1\x1B\x00\x9C\x3D\x1B\x00\x93\x31\x17\x00\x92\x21"
-     "\x17\x00"
-     "\x91\x21\x17\x00\x90\x75\xEB\x00\x97\x19\x12\x00\x96\x31\x13\x00\x95\x39"
-     "\x13\x00"
-     "\x94\x65\x16\x00\x80\xA9\x01\x00\x81\xA9\x01\x00\x82\xB9\x01\x00\xB3\xFD"
-     "\x02\x00"
-     "\x8F\x55\xEB\x00\xB5\xED\x02\x00\xB6\xE5\x02\x00\x98\x60\x00\x80\x86\xC0"
-     "\x0C\x00"
-     "\x87\xE8\x0D\x00\xBA\x65\x01\x00\xBB\x7D\x01\x00\xBC\x6D\x01\x00\xBD\x5D"
-     "\x01\x00"
-     "\xBE\x55\x01\x00\xBF\x3D\x01\x00\xE1\x70\x07\x00\x9C\x60\x00\x80\xE3\x88"
-     "\x01\x00"
-     "\xA0\x60\x00\x80\xE3\xD4\x00\x00\x84\x9C\x0D\x00\xE1\xA8\x01\x00\xA4\x60"
-     "\x00\x80"
-     "\xEF\x00\x00\x00\xA8\x60\x00\x80\xAC\x60\x00\x80\xB0\x60\x00\x80\xB4\x60"
-     "\x00\x80"
-     "\xB8\x60\x00\x80\xEF\x7C\x06\x00\xBC\x60\x00\x80\xC0\x60\x00\x80\xE3\x30"
-     "\x07\x00"
-     "\xBE\xC4\x0C\x00\xE1\x7C\x04\x00\xC4\x60\x00\x80\xC8\x60\x00\x80\xCC\x60"
-     "\x00\x80"
-     "\xD0\x60\x00\x80\xA3\xFD\x03\x00\xA2\xE1\x03\x00\xA1\xA1\x03\x00\xA0\xB1"
-     "\x07\x00"
-     "\xD4\x60\x00\x80\xEF\xCC\x07\x00\xD8\x60\x00\x80\xDC\x60\x00\x80\xE0\x60"
-     "\x00\x80"
-     "\xE4\x60\x00\x80\x24\x60\x00\x80\x94\x60\x00\x80\xE8\x60\x00\x80\xEC\x60"
-     "\x00\x80"
-     "\xF0\x60\x00\x80\xF4\x60\x00\x80\xA8\x89\x0D\x00\xA9\x89\x0D\x00\xAA\x99"
-     "\x0D\x00"
-     "\xAB\x99\x0D\x00\xAC\x89\x0D\x00\xAD\x89\x0D\x00\xAE\x3D\x0E\x00\xAF\x35"
-     "\x0E\x00"
-     "\xB0\x4D\x0E\x00\xB1\x55\x0E\x00\xB2\x55\x0E\x00\xB3\x65\x0E\x00\xB4\x7D"
-     "\x0E\x00"
-     "\xB5\x15\x0E\x00\xB6\x11\x0E\x00\xB7\x11\x0E\x00\xB8\x35\x0E\x00\xB9\x3D"
-     "\x0E\x00"
-     "\xBA\x01\x0E\x00\xBB\x01\x0E\x00\xBC\xF9\x0E\x00\xBD\xDD\x0E\x00\xBE\xD5"
-     "\x0E\x00"
-     "\xBF\xB5\x0E\x00\xA3\x31\x0D\x00\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xF8\x60\x00\x80\xA6\x29\x0D\x00\xA5\x21\x0D\x00\xFC\x60\x00\x80\xAB\xB1"
-     "\x0E\x00"
-     "\xAA\xA9\x0E\x00\x00\x61\x00\x80\x04\x61\x00\x80\xAF\xF1\x0E\x00\xAE\x99"
-     "\x0E\x00"
-     "\xAD\x91\x0E\x00\xAC\xA1\x0E\x00\x08\x61\x00\x80\xB3\x49\x0F\x00\x86\x68"
-     "\x00\x00"
-     "\x87\x1C\x03\x00\xB6\x79\x0F\x00\x0C\x61\x00\x80\x10\x61\x00\x80\xB5\x79"
-     "\x0F\x00"
-     "\xBA\x55\x0F\x00\xBB\x59\x0F\x00\x14\x61\x00\x80\x18\x61\x00\x80\xBE\x25"
-     "\x0F\x00"
-     "\xBF\x2D\x0F\x00\xBC\x39\x0F\x00\xBD\x2D\x0F\x00\xA8\x5D\x0E\x00\xA9\x81"
-     "\x0E\x00"
-     "\xAA\x85\x0E\x00\xAB\x91\x0E\x00\xAC\xB5\x0E\x00\xAD\xBD\x0E\x00\xAE\xB5"
-     "\x0E\x00"
-     "\xAF\xAD\x0E\x00\x1C\x61\x00\x80\x20\x61\x00\x80\x24\x61\x00\x80\x28\x61"
-     "\x00\x80"
-     "\x2C\x61\x00\x80\x30\x61\x00\x80\x34\x61\x00\x80\x38\x61\x00\x80\xB8\xB9"
-     "\x0E\x00"
-     "\xB9\xB9\x0E\x00\xBA\x49\x01\x00\xBB\x49\x01\x00\xBC\x41\x01\x00\xBD\x41"
-     "\x01\x00"
-     "\xBE\x41\x01\x00\xBF\x41\x01\x00\xB0\xD5\x0E\x00\xB1\xD9\x0E\x00\xB2\xA9"
-     "\x0E\x00"
-     "\xB3\xBD\x0E\x00\xB4\xA5\x0E\x00\xB5\xAD\x0E\x00\xB6\xA5\x0E\x00\xB7\x8D"
-     "\x0E\x00"
-     "\xA3\x0D\x0E\x00\x3C\x61\x00\x80\x40\x61\x00\x80\x44\x61\x00\x80\x48\x61"
-     "\x00\x80"
-     "\xA6\x3D\x0E\x00\xA5\x3D\x0E\x00\x4C\x61\x00\x80\xAB\x1D\x0E\x00\xAA\x11"
-     "\x0E\x00"
-     "\x50\x61\x00\x80\x54\x61\x00\x80\xAF\x69\x0E\x00\xAE\x61\x0E\x00\xAD\x69"
-     "\x0E\x00"
-     "\xAC\x7D\x0E\x00\x58\x61\x00\x80\x5C\x61\x00\x80\x60\x61\x00\x80\x64\x61"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\x68\x61\x00\x80\x6C\x61"
-     "\x00\x80"
-     "\x70\x61\x00\x80\x84\xA4\x01\x00\xBE\xA4\x01\x00\x86\xD0\x07\x00\x87\xF8"
-     "\x00\x00"
-     "\x78\x61\x00\x80\x7C\x61\x00\x80\xA8\xC1\x02\x00\xA9\xD1\x02\x00\xAA\xD1"
-     "\x02\x00"
-     "\xAB\xE5\x02\x00\xAC\xFD\x02\x00\xAD\x35\x03\x00\xAE\x3D\x03\x00\xAF\x35"
-     "\x03\x00"
-     "\x80\x61\x00\x80\x84\x61\x00\x80\x88\x61\x00\x80\x8C\x61\x00\x80\x90\x61"
-     "\x00\x80"
-     "\x94\x61\x00\x80\x98\x61\x00\x80\x9C\x61\x00\x80\xB8\xD1\x03\x00\xB9\xD9"
-     "\x03\x00"
-     "\xBA\xE1\x03\x00\xBB\xE1\x03\x00\xBC\x91\x03\x00\xBD\x91\x03\x00\xBE\x91"
-     "\x03\x00"
-     "\xBF\x91\x03\x00\xB0\x4D\x03\x00\xB1\x55\x03\x00\xB2\x5D\x03\x00\xB3\x55"
-     "\x03\x00"
-     "\xB4\x4D\x03\x00\xB5\xF1\x03\x00\xB6\xF1\x03\x00\xB7\xF1\x03\x00\xB3\x59"
-     "\x02\x00"
-     "\xA0\x61\x00\x80\xA4\x61\x00\x80\xBE\x00\x04\x00\xA8\x61\x00\x80\xB6\xB5"
-     "\x03\x00"
-     "\xB5\x75\x02\x00\xAC\x61\x00\x80\xBB\x91\x03\x00\xBA\x89\x03\x00\xB0\x61"
-     "\x00\x80"
-     "\xB4\x61\x00\x80\xBF\x35\x03\x00\xBE\x89\x03\x00\xBD\x81\x03\x00\xBC\x89"
-     "\x03\x00"
-     "\xB8\x61\x00\x80\xA3\x1D\x02\x00\xBC\x61\x00\x80\xC0\x61\x00\x80\xA6\xF1"
-     "\x03\x00"
-     "\xC4\x61\x00\x80\xC8\x61\x00\x80\xA5\x31\x02\x00\xAA\xCD\x03\x00\xAB\xD5"
-     "\x03\x00"
-     "\xCC\x61\x00\x80\xD0\x61\x00\x80\xAE\xCD\x03\x00\xAF\x71\x03\x00\xAC\xCD"
-     "\x03\x00"
-     "\xAD\xC5\x03\x00\xD4\x61\x00\x80\x82\x21\x00\x00\x81\x15\x00\x00\x80\x15"
-     "\x00\x00"
-     "\xEF\x04\x00\x00\x84\x94\x02\x00\xD8\x61\x00\x80\xDC\x61\x00\x80\xE3\xD8"
-     "\x00\x00"
-     "\xE0\x61\x00\x80\xE1\xF8\x01\x00\xE8\x61\x00\x80\xEC\x61\x00\x80\xF0\x61"
-     "\x00\x80"
-     "\xF4\x61\x00\x80\xF8\x61\x00\x80\x86\xE0\x04\x00\x87\x18\x05\x00\xFC\x61"
-     "\x00\x80"
-     "\xB3\x6D\x01\x00\x00\x62\x00\x80\xB5\x7D\x01\x00\xB6\x75\x01\x00\x04\x62"
-     "\x00\x80"
-     "\x08\x62\x00\x80\x0C\x62\x00\x80\xBA\x51\x01\x00\xBB\x51\x01\x00\xBC\xF5"
-     "\x01\x00"
-     "\xBD\xFD\x01\x00\xBE\xE9\x01\x00\xBF\xE1\x01\x00\xE3\x04\x0E\x00\xE3\x18"
-     "\x0F\x00"
-     "\xE1\xF0\x0E\x00\xE1\x34\x0C\x00\x10\x62\x00\x80\x14\x62\x00\x80\x18\x62"
-     "\x00\x80"
-     "\x1C\x62\x00\x80\xB2\xAD\x37\x00\xB3\x19\x37\x00\xB0\xB9\x32\x00\xB1\x01"
-     "\x34\x00"
-     "\xEF\x80\x0E\x00\xEF\x20\x0C\x00\xB4\x01\x28\x00\xB5\xBD\x2B\x00\xAA\x01"
-     "\x3C\x00"
-     "\xAB\xC1\x3F\x00\xA8\xCD\x3B\x00\xA9\xA1\x3B\x00\xAE\xD5\x33\x00\xAF\xDD"
-     "\x33\x00"
-     "\xAC\xC9\x3F\x00\xAD\x01\x30\x00\xA2\xED\x03\x00\xA3\x49\x03\x00\xBE\xA0"
-     "\x04\x00"
-     "\xA1\xE5\x03\x00\xA6\x81\x07\x00\xA7\x01\x38\x00\xA4\x01\x04\x00\xA5\xF9"
-     "\x07\x00"
-     "\xA3\x21\x06\x00\xE4\x61\x00\x80\x20\x62\x00\x80\x24\x62\x00\x80\x28\x62"
-     "\x00\x80"
-     "\xA6\x39\x06\x00\xA5\x31\x06\x00\x2C\x62\x00\x80\xAB\x1D\x06\x00\xAA\x1D"
-     "\x06\x00"
-     "\x30\x62\x00\x80\x34\x62\x00\x80\xAF\xAD\x06\x00\xAE\xA5\x06\x00\xAD\xB1"
-     "\x06\x00"
-     "\xAC\xB9\x06\x00\x38\x62\x00\x80\xEF\x30\x04\x00\x3C\x62\x00\x80\x40\x62"
-     "\x00\x80"
-     "\x44\x62\x00\x80\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D\x00\x00\x48\x62"
-     "\x00\x80"
-     "\x4C\x62\x00\x80\x50\x62\x00\x80\x84\x20\x03\x00\x54\x62\x00\x80\xE3\x70"
-     "\x05\x00"
-     "\x58\x62\x00\x80\xE1\x94\x07\x00\xA8\x79\x05\x00\xA9\x79\x05\x00\xAA\xA9"
-     "\x06\x00"
-     "\xAB\xA9\x06\x00\xAC\xB9\x06\x00\xAD\xB9\x06\x00\xAE\xA9\x06\x00\xAF\xA9"
-     "\x06\x00"
-     "\x86\x00\x00\x00\x87\x84\x00\x00\x5C\x62\x00\x80\x60\x62\x00\x80\x64\x62"
-     "\x00\x80"
-     "\x68\x62\x00\x80\x6C\x62\x00\x80\x70\x62\x00\x80\xB8\x01\x06\x00\xB9\x0D"
-     "\x06\x00"
-     "\xBA\x05\x06\x00\xBB\x19\x06\x00\xBC\x09\x06\x00\xBD\x3D\x06\x00\xBE\x35"
-     "\x06\x00"
-     "\xBF\xDD\x07\x00\xB0\xD9\x06\x00\xB1\xD9\x06\x00\xB2\x6D\x06\x00\xB3\x65"
-     "\x06\x00"
-     "\xB4\x61\x06\x00\xB5\x69\x06\x00\xB6\x51\x06\x00\xB7\x51\x06\x00\xA8\x2D"
-     "\x06\x00"
-     "\xA9\x35\x06\x00\xAA\x3D\x06\x00\xAB\x35\x06\x00\xAC\x2D\x06\x00\xAD\x5D"
-     "\x06\x00"
-     "\xAE\x49\x06\x00\xAF\x41\x06\x00\x74\x62\x00\x80\x78\x62\x00\x80\x7C\x62"
-     "\x00\x80"
-     "\x80\x62\x00\x80\x84\x62\x00\x80\x88\x62\x00\x80\x8C\x62\x00\x80\x90\x62"
-     "\x00\x80"
-     "\xB8\x51\x01\x00\xB9\x59\x01\x00\xBA\x61\x01\x00\xBB\x61\x01\x00\xBC\x11"
-     "\x01\x00"
-     "\xBD\x1D\x01\x00\xBE\x15\x01\x00\xBF\x0D\x01\x00\xB0\x01\x06\x00\xB1\x01"
-     "\x06\x00"
-     "\xB2\xF9\x06\x00\xB3\xF9\x06\x00\xB4\xE9\x06\x00\xB5\xE9\x06\x00\xB6\x71"
-     "\x01\x00"
-     "\xB7\x71\x01\x00\xB3\xB5\x06\x00\x94\x62\x00\x80\x98\x62\x00\x80\x9C\x62"
-     "\x00\x80"
-     "\xA0\x62\x00\x80\xB6\xDD\x06\x00\xB5\xA5\x06\x00\xA4\x62\x00\x80\xBB\xC5"
-     "\x06\x00"
-     "\xBA\xC5\x06\x00\xA8\x62\x00\x80\xAC\x62\x00\x80\xBF\x2D\x06\x00\xBE\x2D"
-     "\x06\x00"
-     "\xBD\x2D\x06\x00\xBC\x2D\x06\x00\xB0\x62\x00\x80\xA3\xF1\x06\x00\xB4\x62"
-     "\x00\x80"
-     "\xB8\x62\x00\x80\xA6\x99\x06\x00\xBC\x62\x00\x80\xC0\x62\x00\x80\xA5\xE1"
-     "\x06\x00"
-     "\xAA\x81\x06\x00\xAB\x81\x06\x00\xBE\xB4\x01\x00\x84\xA8\x01\x00\xAE\x69"
-     "\x06\x00"
-     "\xAF\x69\x06\x00\xAC\x69\x06\x00\xAD\x69\x06\x00\xA8\x79\x02\x00\xA9\xAD"
-     "\x02\x00"
-     "\xAA\xA5\x02\x00\xAB\xBD\x02\x00\xAC\xA5\x02\x00\xAD\xAD\x02\x00\xAE\xA5"
-     "\x02\x00"
-     "\xAF\x3D\x03\x00\x80\xED\x03\x00\x81\xF1\x03\x00\x82\xF1\x03\x00\xC8\x62"
-     "\x00\x80"
-     "\xCC\x62\x00\x80\xD0\x62\x00\x80\x87\xAC\x03\x00\x86\x9C\x03\x00\xB8\xC9"
-     "\x03\x00"
-     "\xB9\xC9\x03\x00\xBA\xD9\x03\x00\xBB\xD1\x03\x00\xBC\xF9\x03\x00\xBD\xF9"
-     "\x03\x00"
-     "\xBE\x99\x03\x00\xBF\x95\x03\x00\xB0\x45\x03\x00\xB1\x4D\x03\x00\xB2\x45"
-     "\x03\x00"
-     "\xB3\x5D\x03\x00\xB4\x45\x03\x00\xB5\x4D\x03\x00\xB6\x45\x03\x00\xB7\xF9"
-     "\x03\x00"
-     "\xD4\x62\x00\x80\xD8\x62\x00\x80\xB3\x25\x02\x00\x84\x40\x3C\x00\xB5\x25"
-     "\x02\x00"
-     "\xDC\x62\x00\x80\xE0\x62\x00\x80\xB6\x25\x02\x00\xE4\x62\x00\x80\xE8\x62"
-     "\x00\x80"
-     "\xBB\x81\x03\x00\xBA\x89\x03\x00\xBD\x81\x03\x00\xBC\x99\x03\x00\xBF\x81"
-     "\x03\x00"
-     "\xBE\x89\x03\x00\xEC\x62\x00\x80\xA6\x7D\x02\x00\xA5\x7D\x02\x00\xF0\x62"
-     "\x00\x80"
-     "\xA3\x7D\x02\x00\xF4\x62\x00\x80\xF8\x62\x00\x80\xFC\x62\x00\x80\xAF\xD9"
-     "\x03\x00"
-     "\xAE\xD1\x03\x00\xAD\xD9\x03\x00\xAC\xC1\x03\x00\xAB\xD9\x03\x00\xAA\xD1"
-     "\x03\x00"
-     "\x00\x63\x00\x80\x04\x63\x00\x80\xA8\x2D\x01\x00\xA9\x35\x01\x00\xAA\x3D"
-     "\x01\x00"
-     "\xAB\x8D\x01\x00\xAC\x95\x01\x00\xAD\x81\x01\x00\xAE\x81\x01\x00\xAF\xBD"
-     "\x01\x00"
-     "\x08\x63\x00\x80\x0C\x63\x00\x80\x10\x63\x00\x80\x14\x63\x00\x80\x18\x63"
-     "\x00\x80"
-     "\x1C\x63\x00\x80\x20\x63\x00\x80\x24\x63\x00\x80\xB8\x61\x00\x00\xB9\x61"
-     "\x00\x00"
-     "\xBA\x61\x00\x00\xBB\x61\x00\x00\xBC\x61\x00\x00\xBD\x61\x00\x00\xBE\x61"
-     "\x00\x00"
-     "\xBF\x61\x00\x00\xB0\xC5\x01\x00\xB1\xA5\x01\x00\xB2\xA5\x01\x00\xB3\xB9"
-     "\x01\x00"
-     "\xB4\xA9\x01\x00\xB5\xA9\x01\x00\xB6\x81\x01\x00\xB7\x81\x01\x00\xE1\xD8"
-     "\x06\x00"
-     "\x28\x63\x00\x80\xE3\xF4\x01\x00\x2C\x63\x00\x80\x81\x1D\x00\x00\x80\x11"
-     "\x00\x00"
-     "\x30\x63\x00\x80\x82\x05\x00\x00\xBE\xAC\x3D\x00\x38\x63\x00\x80\x86\x08"
-     "\x3C\x00"
-     "\x87\xFC\x3D\x00\x3C\x63\x00\x80\x40\x63\x00\x80\xEF\xA8\x06\x00\x44\x63"
-     "\x00\x80"
-     "\x48\x63\x00\x80\x4C\x63\x00\x80\x50\x63\x00\x80\xEF\xDC\x00\x00\x54\x63"
-     "\x00\x80"
-     "\x58\x63\x00\x80\x5C\x63\x00\x80\x60\x63\x00\x80\x64\x63\x00\x80\x68\x63"
-     "\x00\x80"
-     "\x6C\x63\x00\x80\x70\x63\x00\x80\x74\x63\x00\x80\xE1\x4C\x01\x00\x74\x61"
-     "\x00\x80"
-     "\xE3\x00\x00\x00\xB3\x35\x3E\x00\x34\x63\x00\x80\x78\x63\x00\x80\x7C\x63"
-     "\x00\x80"
-     "\x80\x63\x00\x80\xB6\x81\x3E\x00\xB5\x9D\x3E\x00\x84\x63\x00\x80\xBB\xE1"
-     "\x3E\x00"
-     "\xBA\x81\x3E\x00\x88\x63\x00\x80\x8C\x63\x00\x80\xBF\xED\x3E\x00\xBE\xED"
-     "\x3E\x00"
-     "\xBD\xED\x3E\x00\xBC\xF1\x3E\x00\x90\x63\x00\x80\x94\x63\x00\x80\xEF\x10"
-     "\x3D\x00"
-     "\x98\x63\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39\x00\x00\x9C\x63"
-     "\x00\x80"
-     "\xA0\x63\x00\x80\x84\x6C\x02\x00\xBE\x64\x03\x00\xA4\x63\x00\x80\xE1\x54"
-     "\x3C\x00"
-     "\xA8\x63\x00\x80\xE3\xE4\x3C\x00\xAC\x63\x00\x80\xA3\xB5\x3E\x00\xB0\x63"
-     "\x00\x80"
-     "\x87\x48\x03\x00\x86\x2C\x00\x00\xB4\x63\x00\x80\xA6\x01\x3E\x00\xA5\x1D"
-     "\x3E\x00"
-     "\xB8\x63\x00\x80\xAB\x61\x3E\x00\xAA\x01\x3E\x00\xBC\x63\x00\x80\xC0\x63"
-     "\x00\x80"
-     "\xAF\x6D\x3E\x00\xAE\x6D\x3E\x00\xAD\x6D\x3E\x00\xAC\x71\x3E\x00\xC4\x63"
-     "\x00\x80"
-     "\xB3\x71\x3F\x00\xC8\x63\x00\x80\xCC\x63\x00\x80\xB6\x15\x3F\x00\xD0\x63"
-     "\x00\x80"
-     "\xD4\x63\x00\x80\xB5\x1D\x3F\x00\xBA\x35\x3F\x00\xBB\x3D\x3F\x00\xD8\x63"
-     "\x00\x80"
-     "\xDC\x63\x00\x80\xBE\x15\x3F\x00\xBF\x1D\x3F\x00\xBC\x25\x3F\x00\xBD\x1D"
-     "\x3F\x00"
-     "\xA8\xA5\x3E\x00\xA9\xB5\x3E\x00\xAA\xBD\x3E\x00\xAB\xB1\x3E\x00\xAC\xD1"
-     "\x3E\x00"
-     "\xAD\xD1\x3E\x00\xAE\xD1\x3E\x00\xAF\xCD\x3E\x00\xE0\x63\x00\x80\xE4\x63"
-     "\x00\x80"
-     "\xE8\x63\x00\x80\xEC\x63\x00\x80\xF0\x63\x00\x80\xF4\x63\x00\x80\xF8\x63"
-     "\x00\x80"
-     "\xFC\x63\x00\x80\xB8\x49\x01\x00\xB9\x49\x01\x00\xBA\x59\x01\x00\xBB\x51"
-     "\x01\x00"
-     "\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x19\x01\x00\xBF\x19\x01\x00\xB0\xB5"
-     "\x3E\x00"
-     "\xB1\xBD\x3E\x00\xB2\xB5\x3E\x00\xB3\x8D\x3E\x00\xB4\x95\x3E\x00\xB5\x9D"
-     "\x3E\x00"
-     "\xB6\x95\x3E\x00\xB7\x8D\x3E\x00\xA3\x35\x3E\x00\x00\x64\x00\x80\x04\x64"
-     "\x00\x80"
-     "\x08\x64\x00\x80\x0C\x64\x00\x80\xA6\x51\x3E\x00\xA5\x59\x3E\x00\x10\x64"
-     "\x00\x80"
-     "\xAB\x79\x3E\x00\xAA\x71\x3E\x00\x84\xA8\x02\x00\xBE\xAC\x01\x00\xAF\x59"
-     "\x3E\x00"
-     "\xAE\x51\x3E\x00\xAD\x59\x3E\x00\xAC\x61\x3E\x00\x80\x4D\x00\x00\x81\x55"
-     "\x00\x00"
-     "\x82\x55\x00\x00\xB3\x49\x01\x00\x18\x64\x00\x80\xB5\x49\x01\x00\xB6\x71"
-     "\x01\x00"
-     "\x1C\x64\x00\x80\x86\x40\x00\x00\x87\x14\x01\x00\xBA\x21\x01\x00\xBB\x21"
-     "\x01\x00"
-     "\xBC\x25\x01\x00\xBD\x2D\x01\x00\xBE\x11\x01\x00\xBF\x11\x01\x00\xA8\x75"
-     "\x02\x00"
-     "\xA9\x79\x02\x00\xAA\x8D\x02\x00\xAB\x9D\x02\x00\xAC\x8D\x02\x00\xAD\xBD"
-     "\x02\x00"
-     "\xAE\xB5\x02\x00\xAF\xDD\x02\x00\x20\x64\x00\x80\x24\x64\x00\x80\x28\x64"
-     "\x00\x80"
-     "\x2C\x64\x00\x80\x30\x64\x00\x80\x34\x64\x00\x80\x38\x64\x00\x80\x3C\x64"
-     "\x00\x80"
-     "\xB8\x69\x03\x00\xB9\x69\x03\x00\xBA\x79\x03\x00\xBB\x79\x03\x00\xBC\x69"
-     "\x03\x00"
-     "\xBD\x69\x03\x00\xBE\xDD\x03\x00\xBF\xD5\x03\x00\xB0\xA9\x02\x00\xB1\xA9"
-     "\x02\x00"
-     "\xB2\xBD\x02\x00\xB3\xB5\x02\x00\xB4\x99\x02\x00\xB5\x99\x02\x00\xB6\x59"
-     "\x03\x00"
-     "\xB7\x59\x03\x00\x40\x64\x00\x80\x44\x64\x00\x80\x48\x64\x00\x80\xA3\x09"
-     "\x02\x00"
-     "\x84\x5C\x04\x00\xA5\x09\x02\x00\xA6\x31\x02\x00\x4C\x64\x00\x80\x50\x64"
-     "\x00\x80"
-     "\x54\x64\x00\x80\xAA\x61\x02\x00\xAB\x61\x02\x00\xAC\x65\x02\x00\xAD\x6D"
-     "\x02\x00"
-     "\xAE\x51\x02\x00\xAF\x51\x02\x00\x58\x64\x00\x80\x5C\x64\x00\x80\x60\x64"
-     "\x00\x80"
-     "\xBE\xC0\x04\x00\x64\x64\x00\x80\x68\x64\x00\x80\x6C\x64\x00\x80\x70\x64"
-     "\x00\x80"
-     "\x74\x64\x00\x80\x78\x64\x00\x80\x7C\x64\x00\x80\x80\x64\x00\x80\x80\x15"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x82\x15\x00\x00\x88\x64\x00\x80\xA8\xD1\x02\x00\xA9\xD9"
-     "\x02\x00"
-     "\xAA\x29\x01\x00\xAB\x29\x01\x00\xAC\x39\x01\x00\xAD\x39\x01\x00\xAE\x29"
-     "\x01\x00"
-     "\xAF\x29\x01\x00\x8C\x64\x00\x80\x90\x64\x00\x80\x94\x64\x00\x80\x98\x64"
-     "\x00\x80"
-     "\x9C\x64\x00\x80\xA0\x64\x00\x80\xA4\x64\x00\x80\xA8\x64\x00\x80\xB8\x81"
-     "\x01\x00"
-     "\xB9\x81\x01\x00\xBA\x81\x01\x00\xBB\x9D\x01\x00\xBC\x85\x01\x00\xBD\x8D"
-     "\x01\x00"
-     "\xBE\x85\x01\x00\xBF\xBD\x01\x00\xB0\x59\x01\x00\xB1\x59\x01\x00\xB2\xE9"
-     "\x01\x00"
-     "\xB3\xE9\x01\x00\xB4\xF9\x01\x00\xB5\xE9\x01\x00\xB6\xD9\x01\x00\xB7\xD9"
-     "\x01\x00"
-     "\xE1\x30\x3E\x00\xAC\x64\x00\x80\xE3\xFC\x01\x00\xB0\x64\x00\x80\xB4\x64"
-     "\x00\x80"
-     "\xB8\x64\x00\x80\x86\x38\x04\x00\x87\x64\x02\x00\x84\xA8\x05\x00\xBC\x64"
-     "\x00\x80"
-     "\xC0\x64\x00\x80\xC4\x64\x00\x80\xC8\x64\x00\x80\xCC\x64\x00\x80\xEF\x28"
-     "\x01\x00"
-     "\xD0\x64\x00\x80\x85\xFC\x13\x00\xD4\x64\x00\x80\xD8\x64\x00\x80\xDC\x64"
-     "\x00\x80"
-     "\xEF\x84\x00\x00\xE0\x64\x00\x80\xE4\x64\x00\x80\xE8\x64\x00\x80\xE3\x58"
-     "\x00\x00"
-     "\xEC\x64\x00\x80\xE1\x7C\x01\x00\xF0\x64\x00\x80\x82\x1D\x00\x00\xC4\x62"
-     "\x00\x80"
-     "\x80\x25\x00\x00\x81\x19\x00\x00\xB3\x21\x06\x00\x84\x64\x00\x80\xBE\x04"
-     "\x03\x00"
-     "\xF8\x64\x00\x80\xF4\x64\x00\x80\xB6\x39\x06\x00\xB5\x31\x06\x00\xFC\x64"
-     "\x00\x80"
-     "\xBB\x1D\x06\x00\xBA\x1D\x06\x00\x86\x88\x00\x00\x87\x0C\x03\x00\xBF\xAD"
-     "\x06\x00"
-     "\xBE\xBD\x06\x00\xBD\xBD\x06\x00\xBC\xBD\x06\x00\x00\x65\x00\x80\xE1\xBC"
-     "\x07\x00"
-     "\x04\x65\x00\x80\xE3\xCC\x04\x00\x08\x65\x00\x80\x0C\x65\x00\x80\x10\x65"
-     "\x00\x80"
-     "\x14\x65\x00\x80\x18\x65\x00\x80\x1C\x65\x00\x80\x20\x65\x00\x80\x24\x65"
-     "\x00\x80"
-     "\x28\x65\x00\x80\x2C\x65\x00\x80\x30\x65\x00\x80\xEF\x48\x04\x00\xA3\xA1"
-     "\x06\x00"
-     "\x34\x65\x00\x80\x38\x65\x00\x80\x3C\x65\x00\x80\x40\x65\x00\x80\xA6\xB9"
-     "\x06\x00"
-     "\xA5\xB1\x06\x00\x44\x65\x00\x80\xAB\x9D\x06\x00\xAA\x9D\x06\x00\x48\x65"
-     "\x00\x80"
-     "\x4C\x65\x00\x80\xAF\x2D\x06\x00\xAE\x3D\x06\x00\xAD\x3D\x06\x00\xAC\x3D"
-     "\x06\x00"
-     "\x50\x65\x00\x80\xB3\xE1\x07\x00\x54\x65\x00\x80\x58\x65\x00\x80\xB6\xE1"
-     "\x07\x00"
-     "\x5C\x65\x00\x80\x60\x65\x00\x80\xB5\xF1\x07\x00\xBA\xDD\x07\x00\xBB\x61"
-     "\x07\x00"
-     "\x84\xE0\x03\x00\xBE\xE4\x00\x00\xBE\x65\x07\x00\xBF\x51\x07\x00\xBC\x75"
-     "\x07\x00"
-     "\xBD\x6D\x07\x00\xA8\xFD\x07\x00\xA9\x25\x06\x00\xAA\x2D\x06\x00\xAB\x25"
-     "\x06\x00"
-     "\xAC\x3D\x06\x00\xAD\x25\x06\x00\xAE\x2D\x06\x00\xAF\x25\x06\x00\x80\xED"
-     "\x01\x00"
-     "\x81\xF1\x01\x00\x82\xF1\x01\x00\x68\x65\x00\x80\x86\x90\x00\x00\x87\xF0"
-     "\x01\x00"
-     "\x6C\x65\x00\x80\x70\x65\x00\x80\xB8\xF1\x06\x00\xB9\xF1\x06\x00\xBA\xF5"
-     "\x06\x00"
-     "\xBB\x89\x06\x00\xBC\x99\x06\x00\xBD\x99\x06\x00\xBE\x8D\x06\x00\xBF\x85"
-     "\x06\x00"
-     "\xB0\x5D\x06\x00\xB1\xE1\x06\x00\xB2\xE5\x06\x00\xB3\xFD\x06\x00\xB4\xE5"
-     "\x06\x00"
-     "\xB5\xD1\x06\x00\xB6\xD1\x06\x00\xB7\xD1\x06\x00\xA3\xA5\x06\x00\x74\x65"
-     "\x00\x80"
-     "\x78\x65\x00\x80\x7C\x65\x00\x80\x80\x65\x00\x80\xA6\xA5\x06\x00\xA5\xB5"
-     "\x06\x00"
-     "\x84\x65\x00\x80\xAB\x25\x06\x00\xAA\x99\x06\x00\x88\x65\x00\x80\x8C\x65"
-     "\x00\x80"
-     "\xAF\x15\x06\x00\xAE\x21\x06\x00\xAD\x29\x06\x00\xAC\x31\x06\x00\x90\x65"
-     "\x00\x80"
-     "\xB3\x85\x01\x00\x94\x65\x00\x80\x98\x65\x00\x80\xB6\x85\x01\x00\x9C\x65"
-     "\x00\x80"
-     "\xA0\x65\x00\x80\xB5\x8D\x01\x00\xBA\x81\x01\x00\xBB\x81\x01\x00\xA4\x65"
-     "\x00\x80"
-     "\xA8\x65\x00\x80\xBE\x81\x01\x00\xBF\x81\x01\x00\xBC\x91\x01\x00\xBD\x91"
-     "\x01\x00"
-     "\xA8\x0D\x02\x00\xA9\x19\x02\x00\xAA\x49\x02\x00\xAB\x45\x02\x00\xAC\x41"
-     "\x02\x00"
-     "\xAD\x41\x02\x00\xAE\x41\x02\x00\xAF\x41\x02\x00\xAC\x65\x00\x80\xB0\x65"
-     "\x00\x80"
-     "\xB4\x65\x00\x80\xB8\x65\x00\x80\x84\x38\x03\x00\xBC\x65\x00\x80\xC0\x65"
-     "\x00\x80"
-     "\xC4\x65\x00\x80\xB8\xC9\x02\x00\xB9\xC9\x02\x00\xBA\xD9\x02\x00\xBB\xD1"
-     "\x02\x00"
-     "\xBC\xF9\x02\x00\xBD\xF9\x02\x00\xBE\x99\x02\x00\xBF\x95\x02\x00\xB0\x01"
-     "\x02\x00"
-     "\xB1\x01\x02\x00\xB2\x01\x02\x00\xB3\x01\x02\x00\xB4\x01\x02\x00\xB5\x01"
-     "\x02\x00"
-     "\xB6\x01\x02\x00\xB7\xF9\x02\x00\xC8\x65\x00\x80\xA3\xC5\x02\x00\xCC\x65"
-     "\x00\x80"
-     "\xD0\x65\x00\x80\xA6\xC5\x02\x00\xD8\x65\x00\x80\xDC\x65\x00\x80\xA5\xCD"
-     "\x02\x00"
-     "\xAA\xC1\x02\x00\xAB\xC1\x02\x00\x86\x88\x0C\x00\x87\x18\x03\x00\xAE\xC1"
-     "\x02\x00"
-     "\xAF\xC1\x02\x00\xAC\xD1\x02\x00\xAD\xD1\x02\x00\x81\x49\x00\x00\x80\x45"
-     "\x00\x00"
-     "\xBE\x44\x0C\x00\x82\x41\x00\x00\xE0\x65\x00\x80\xE4\x65\x00\x80\xE8\x65"
-     "\x00\x80"
-     "\xEC\x65\x00\x80\xF0\x65\x00\x80\xF4\x65\x00\x80\xF8\x65\x00\x80\xFC\x65"
-     "\x00\x80"
-     "\x84\x3C\x0D\x00\x00\x66\x00\x80\xBE\x30\x0D\x00\x04\x66\x00\x80\xA8\x2D"
-     "\x02\x00"
-     "\xA9\x3D\x02\x00\xAA\x35\x02\x00\xAB\x89\x02\x00\xAC\x99\x02\x00\xAD\x99"
-     "\x02\x00"
-     "\xAE\x89\x02\x00\xAF\x81\x02\x00\x84\xEC\x0F\x00\x08\x66\x00\x80\xBE\xE0"
-     "\x0F\x00"
-     "\x0C\x66\x00\x80\x10\x66\x00\x80\x14\x66\x00\x80\x18\x66\x00\x80\x1C\x66"
-     "\x00\x80"
-     "\xB8\xAD\x02\x00\xB9\x65\x01\x00\xBA\x6D\x01\x00\xBB\x65\x01\x00\xBC\x7D"
-     "\x01\x00"
-     "\xBD\x61\x01\x00\xBE\x61\x01\x00\xBF\x61\x01\x00\xB0\xC9\x02\x00\xB1\xC9"
-     "\x02\x00"
-     "\xB2\xA9\x02\x00\xB3\xA5\x02\x00\xB4\xBD\x02\x00\xB5\xA1\x02\x00\xB6\xA1"
-     "\x02\x00"
-     "\xB7\x95\x02\x00\xE1\x00\x06\x00\x20\x66\x00\x80\xE3\x38\x06\x00\x24\x66"
-     "\x00\x80"
-     "\x28\x66\x00\x80\xE1\xA4\x01\x00\x2C\x66\x00\x80\xE3\x64\x01\x00\x84\xA0"
-     "\x0F\x00"
-     "\x30\x66\x00\x80\x34\x66\x00\x80\xEF\xB8\x01\x00\x81\x7D\x00\x00\x80\x7D"
-     "\x00\x00"
-     "\xEF\x48\x01\x00\x82\x69\x00\x00\x14\x64\x00\x80\xB3\xF9\x02\x00\x38\x66"
-     "\x00\x80"
-     "\x3C\x66\x00\x80\xB6\x5D\x02\x00\x40\x66\x00\x80\x48\x66\x00\x80\xB5\x55"
-     "\x02\x00"
-     "\xBA\x49\x02\x00\xBB\x49\x02\x00\x86\xC8\x0C\x00\x87\x50\x0D\x00\xBE\x3D"
-     "\x02\x00"
-     "\xBF\x21\x02\x00\xBC\x39\x02\x00\xBD\x35\x02\x00\xA3\x3D\x0D\x00\xD4\x65"
-     "\x00\x80"
-     "\x44\x66\x00\x80\x4C\x66\x00\x80\x50\x66\x00\x80\xA6\x99\x0D\x00\xA5\x91"
-     "\x0D\x00"
-     "\x54\x66\x00\x80\xAB\x8D\x0D\x00\xAA\x8D\x0D\x00\x58\x66\x00\x80\x5C\x66"
-     "\x00\x80"
-     "\xAF\xE5\x0D\x00\xAE\xF9\x0D\x00\xAD\xF1\x0D\x00\xAC\xFD\x0D\x00\x60\x66"
-     "\x00\x80"
-     "\x64\x66\x00\x80\x68\x66\x00\x80\x6C\x66\x00\x80\xE3\x70\x0C\x00\x70\x66"
-     "\x00\x80"
-     "\xE1\x70\x0F\x00\x74\x66\x00\x80\xEF\x50\x0C\x00\x78\x66\x00\x80\x7C\x66"
-     "\x00\x80"
-     "\x80\x66\x00\x80\x84\x66\x00\x80\x88\x66\x00\x80\x8C\x66\x00\x80\x90\x66"
-     "\x00\x80"
-     "\xA8\x15\x0E\x00\xA9\x1D\x0E\x00\xAA\x59\x0E\x00\xAB\x59\x0E\x00\xAC\x49"
-     "\x0E\x00"
-     "\xAD\x49\x0E\x00\xAE\x79\x0E\x00\xAF\x79\x0E\x00\x94\x66\x00\x80\x98\x66"
-     "\x00\x80"
-     "\x9C\x66\x00\x80\xA0\x66\x00\x80\xA4\x66\x00\x80\x82\xB9\x01\x00\x81\xB1"
-     "\x01\x00"
-     "\x80\xB9\x01\x00\xB8\xF9\x0F\x00\xB9\xF9\x0F\x00\xBA\x4D\x0F\x00\xBB\x41"
-     "\x0F\x00"
-     "\xBC\x41\x0F\x00\xBD\x4D\x0F\x00\xBE\x7D\x0F\x00\xBF\x75\x0F\x00\xB0\x09"
-     "\x0E\x00"
-     "\xB1\x09\x0E\x00\xB2\x1D\x0E\x00\xB3\x11\x0E\x00\xB4\xD5\x0F\x00\xB5\xD9"
-     "\x0F\x00"
-     "\xB6\xC9\x0F\x00\xB7\xC9\x0F\x00\xB3\xD1\x0F\x00\xA8\x66\x00\x80\xAC\x66"
-     "\x00\x80"
-     "\xB0\x66\x00\x80\xB4\x66\x00\x80\xB6\xC9\x0F\x00\xB5\xC1\x0F\x00\xB8\x66"
-     "\x00\x80"
-     "\xBB\xCD\x0F\x00\xBA\xCD\x0F\x00\x86\x88\x00\x00\x87\xAC\x01\x00\xBF\x1D"
-     "\x0E\x00"
-     "\xBE\x11\x0E\x00\xBD\xC5\x0F\x00\xBC\xC9\x0F\x00\xBC\x66\x00\x80\xA3\x95"
-     "\x0F\x00"
-     "\xC0\x66\x00\x80\xC4\x66\x00\x80\xA6\x8D\x0F\x00\xC8\x66\x00\x80\xCC\x66"
-     "\x00\x80"
-     "\xA5\x85\x0F\x00\xAA\x89\x0F\x00\xAB\x89\x0F\x00\xD0\x66\x00\x80\xD4\x66"
-     "\x00\x80"
-     "\xAE\x55\x0E\x00\xAF\x59\x0E\x00\xAC\x8D\x0F\x00\xAD\x81\x0F\x00\xA8\x59"
-     "\x0E\x00"
-     "\xA9\x89\x01\x00\xAA\x99\x01\x00\xAB\x91\x01\x00\xAC\xB9\x01\x00\xAD\xB9"
-     "\x01\x00"
-     "\xAE\xD9\x01\x00\xAF\xD9\x01\x00\xBE\xFC\x00\x00\xD8\x66\x00\x80\xDC\x66"
-     "\x00\x80"
-     "\xE0\x66\x00\x80\xE4\x66\x00\x80\xE8\x66\x00\x80\xEC\x66\x00\x80\xF0\x66"
-     "\x00\x80"
-     "\xB8\xB5\x01\x00\xB9\xBD\x01\x00\xBA\xB5\x01\x00\xBB\x6D\x01\x00\xBC\x75"
-     "\x01\x00"
-     "\xBD\x7D\x01\x00\xBE\x75\x01\x00\xBF\x6D\x01\x00\xB0\xA9\x01\x00\xB1\xA9"
-     "\x01\x00"
-     "\xB2\xBD\x01\x00\xB3\xB1\x01\x00\xB4\x95\x01\x00\xB5\x9D\x01\x00\xB6\x95"
-     "\x01\x00"
-     "\xB7\x8D\x01\x00\xB3\x11\x02\x00\xF4\x66\x00\x80\xF8\x66\x00\x80\xFC\x66"
-     "\x00\x80"
-     "\x00\x67\x00\x80\xB6\x39\x02\x00\xB5\x31\x02\x00\x04\x67\x00\x80\xBB\x6D"
-     "\x02\x00"
-     "\xBA\x6D\x02\x00\x08\x67\x00\x80\x0C\x67\x00\x80\xBF\x59\x02\x00\xBE\x51"
-     "\x02\x00"
-     "\xBD\x65\x02\x00\xBC\x69\x02\x00\x10\x67\x00\x80\xA3\x55\x02\x00\x14\x67"
-     "\x00\x80"
-     "\x18\x67\x00\x80\xA6\x7D\x02\x00\x1C\x67\x00\x80\x20\x67\x00\x80\xA5\x75"
-     "\x02\x00"
-     "\xAA\x29\x02\x00\xAB\x29\x02\x00\x28\x67\x00\x80\x2C\x67\x00\x80\xAE\x15"
-     "\x02\x00"
-     "\xAF\x1D\x02\x00\xAC\x2D\x02\x00\xAD\x21\x02\x00\x80\xCD\x00\x00\x81\xD1"
-     "\x00\x00"
-     "\x82\xD1\x00\x00\x84\x00\x04\x00\x30\x67\x00\x80\xE1\xA4\x01\x00\x34\x67"
-     "\x00\x80"
-     "\xE3\x2C\x0F\x00\x86\x80\x04\x00\x87\xA4\x04\x00\x38\x67\x00\x80\xEF\xF4"
-     "\x01\x00"
-     "\x3C\x67\x00\x80\x40\x67\x00\x80\x44\x67\x00\x80\xAA\x00\x00\x00\x48\x67"
-     "\x00\x80"
-     "\x4C\x67\x00\x80\x50\x67\x00\x80\xB3\x85\x03\x00\x54\x67\x00\x80\xB5\x81"
-     "\x03\x00"
-     "\xB6\x81\x03\x00\x58\x67\x00\x80\x5C\x67\x00\x80\x60\x67\x00\x80\xBA\x4D"
-     "\x03\x00"
-     "\xBB\x45\x03\x00\xBC\x5D\x03\x00\xBD\x45\x03\x00\xBE\x4D\x03\x00\xBF\x45"
-     "\x03\x00"
-     "\xA8\x1D\x02\x00\xA9\x45\x02\x00\xAA\x41\x02\x00\xAB\x41\x02\x00\xAC\x45"
-     "\x02\x00"
-     "\xAD\x4D\x02\x00\xAE\x71\x02\x00\xAF\x71\x02\x00\xBE\xAC\x04\x00\x64\x67"
-     "\x00\x80"
-     "\x68\x67\x00\x80\x6C\x67\x00\x80\x84\xB8\x04\x00\x70\x67\x00\x80\x74\x67"
-     "\x00\x80"
-     "\x78\x67\x00\x80\xB8\x49\x01\x00\xB9\x49\x01\x00\xBA\x59\x01\x00\xBB\x51"
-     "\x01\x00"
-     "\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x19\x01\x00\xBF\x15\x01\x00\xB0\xC5"
-     "\x02\x00"
-     "\xB1\xCD\x02\x00\xB2\xC5\x02\x00\xB3\xDD\x02\x00\xB4\xC5\x02\x00\xB5\xCD"
-     "\x02\x00"
-     "\xB6\xC5\x02\x00\xB7\x79\x01\x00\xE1\x30\x0F\x00\xE3\xB4\x0E\x00\xE3\x68"
-     "\x0F\x00"
-     "\xE1\x7C\x0E\x00\x7C\x67\x00\x80\x80\x67\x00\x80\x84\x67\x00\x80\x88\x67"
-     "\x00\x80"
-     "\x8C\x67\x00\x80\x90\x67\x00\x80\x98\x67\x00\x80\x9C\x67\x00\x80\xA0\x67"
-     "\x00\x80"
-     "\xEF\x64\x0E\x00\xEF\xB8\x0F\x00\xA4\x67\x00\x80\xA8\x67\x00\x80\xA6\x9D"
-     "\x02\x00"
-     "\xA5\x9D\x02\x00\xBE\x40\x07\x00\xA3\x99\x02\x00\xAC\x67\x00\x80\xB0\x67"
-     "\x00\x80"
-     "\x64\x65\x00\x80\xAF\x59\x02\x00\xAE\x51\x02\x00\xAD\x59\x02\x00\xAC\x41"
-     "\x02\x00"
-     "\xAB\x59\x02\x00\xAA\x51\x02\x00\x86\xD8\x04\x00\x87\x50\x05\x00\x80\x9D"
-     "\x07\x00"
-     "\x81\xA5\x07\x00\x82\xAD\x07\x00\x94\x67\x00\x80\xEF\xC8\x05\x00\xB4\x67"
-     "\x00\x80"
-     "\xB8\x67\x00\x80\xBC\x67\x00\x80\xE3\xBC\x04\x00\xC0\x67\x00\x80\xE1\x90"
-     "\x04\x00"
-     "\xC4\x67\x00\x80\xC8\x67\x00\x80\xCC\x67\x00\x80\xD0\x67\x00\x80\xD4\x67"
-     "\x00\x80"
-     "\xB3\xD1\x06\x00\xD8\x67\x00\x80\xDC\x67\x00\x80\xE0\x67\x00\x80\xE4\x67"
-     "\x00\x80"
-     "\xB6\xE5\x06\x00\xB5\xF1\x06\x00\xE8\x67\x00\x80\xBB\xA5\x06\x00\xBA\xA5"
-     "\x06\x00"
-     "\xEC\x67\x00\x80\xF0\x67\x00\x80\xBF\x91\x06\x00\xBE\xA9\x06\x00\xBD\xA9"
-     "\x06\x00"
-     "\xBC\xB5\x06\x00\xA8\xE9\x06\x00\xA9\xE9\x06\x00\xAA\x19\x06\x00\xAB\x15"
-     "\x06\x00"
-     "\xAC\x39\x06\x00\xAD\x25\x06\x00\xAE\x2D\x06\x00\xAF\x25\x06\x00\xF4\x67"
-     "\x00\x80"
-     "\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\xF8\x67\x00\x80\xFC\x67"
-     "\x00\x80"
-     "\x00\x68\x00\x80\x04\x68\x00\x80\xB8\xE9\x07\x00\xB9\xE9\x07\x00\xBA\xFD"
-     "\x07\x00"
-     "\xBB\xF5\x07\x00\xBC\xED\x07\x00\xBD\x45\x07\x00\xBE\x4D\x07\x00\xBF\x45"
-     "\x07\x00"
-     "\xB0\x21\x06\x00\xB1\x21\x06\x00\xB2\x21\x06\x00\xB3\x3D\x06\x00\xB4\x2D"
-     "\x06\x00"
-     "\xB5\x15\x06\x00\xB6\x1D\x06\x00\xB7\xD9\x07\x00\xA3\x91\x07\x00\x08\x68"
-     "\x00\x80"
-     "\x86\x28\x01\x00\x87\x4C\x01\x00\x0C\x68\x00\x80\xA6\xA5\x07\x00\xA5\xB1"
-     "\x07\x00"
-     "\x10\x68\x00\x80\xAB\xE5\x07\x00\xAA\xE5\x07\x00\x14\x68\x00\x80\x18\x68"
-     "\x00\x80"
-     "\xAF\xD1\x07\x00\xAE\xE9\x07\x00\xAD\xE9\x07\x00\xAC\xF5\x07\x00\x1C\x68"
-     "\x00\x80"
-     "\xB3\xF5\x06\x00\x20\x68\x00\x80\x24\x68\x00\x80\xB6\x85\x06\x00\x28\x68"
-     "\x00\x80"
-     "\x2C\x68\x00\x80\xB5\x9D\x06\x00\xBA\x79\x01\x00\xBB\x79\x01\x00\x30\x68"
-     "\x00\x80"
-     "\x34\x68\x00\x80\xBE\x41\x01\x00\xBF\x41\x01\x00\xBC\x5D\x01\x00\xBD\x51"
-     "\x01\x00"
-     "\xA8\x29\x06\x00\xA9\x29\x06\x00\xAA\x39\x06\x00\xAB\x31\x06\x00\xAC\x51"
-     "\x06\x00"
-     "\xAD\x41\x06\x00\xAE\x41\x06\x00\xAF\x89\x01\x00\x38\x68\x00\x80\x3C\x68"
-     "\x00\x80"
-     "\x40\x68\x00\x80\x44\x68\x00\x80\x48\x68\x00\x80\x4C\x68\x00\x80\x50\x68"
-     "\x00\x80"
-     "\x54\x68\x00\x80\xB8\x85\x01\x00\xB9\x8D\x01\x00\xBA\x85\x01\x00\xBB\x9D"
-     "\x01\x00"
-     "\xBC\x8D\x01\x00\xBD\xBD\x01\x00\xBE\xB5\x01\x00\xBF\x7D\x01\x00\xB0\xF9"
-     "\x01\x00"
-     "\xB1\xF9\x01\x00\xB2\xCD\x01\x00\xB3\xC1\x01\x00\xB4\xC1\x01\x00\xB5\xC1"
-     "\x01\x00"
-     "\xB6\xC1\x01\x00\xB7\xC1\x01\x00\xA3\xB1\x05\x00\x58\x68\x00\x80\x5C\x68"
-     "\x00\x80"
-     "\x60\x68\x00\x80\x64\x68\x00\x80\xA6\xC1\x05\x00\xA5\xD9\x05\x00\x68\x68"
-     "\x00\x80"
-     "\xAB\x3D\x02\x00\xAA\x3D\x02\x00\x6C\x68\x00\x80\x70\x68\x00\x80\xAF\x05"
-     "\x02\x00"
-     "\xAE\x05\x02\x00\xAD\x15\x02\x00\xAC\x19\x02\x00\x80\x5D\x00\x00\x81\x6D"
-     "\x00\x00"
-     "\x82\x65\x00\x00\xB3\x51\x03\x00\xBE\x9C\x03\x00\xB5\x79\x03\x00\xB6\x19"
-     "\x03\x00"
-     "\x78\x68\x00\x80\x84\xE0\x02\x00\x7C\x68\x00\x80\xBA\x3D\x03\x00\xBB\x35"
-     "\x03\x00"
-     "\xBC\x19\x03\x00\xBD\x19\x03\x00\xBE\xD9\x03\x00\xBF\xD9\x03\x00\xA8\x85"
-     "\x03\x00"
-     "\xA9\x95\x03\x00\xAA\x95\x03\x00\xAB\xA5\x03\x00\xAC\xBD\x03\x00\xAD\xD5"
-     "\x03\x00"
-     "\xAE\xD1\x03\x00\xAF\xD1\x03\x00\x80\x68\x00\x80\x84\x68\x00\x80\x88\x68"
-     "\x00\x80"
-     "\xBE\x60\x1C\x00\x87\x20\x03\x00\x86\x94\x1C\x00\x8C\x68\x00\x80\x90\x68"
-     "\x00\x80"
-     "\xB8\x71\x03\x00\xB9\x71\x03\x00\xBA\x71\x03\x00\xBB\x71\x03\x00\xBC\xD5"
-     "\x00\x00"
-     "\xBD\xDD\x00\x00\xBE\xD5\x00\x00\xBF\xCD\x00\x00\xB0\xB5\x03\x00\xB1\xBD"
-     "\x03\x00"
-     "\xB2\x81\x03\x00\xB3\x81\x03\x00\xB4\x51\x03\x00\xB5\x51\x03\x00\xB6\x51"
-     "\x03\x00"
-     "\xB7\x51\x03\x00\xEF\xA8\x03\x00\x94\x68\x00\x80\x98\x68\x00\x80\x9C\x68"
-     "\x00\x80"
-     "\x84\x1C\x02\x00\xA0\x68\x00\x80\xA4\x68\x00\x80\xA8\x68\x00\x80\x84\x6C"
-     "\x1C\x00"
-     "\xAC\x68\x00\x80\xB0\x68\x00\x80\xB4\x68\x00\x80\xE3\x40\x03\x00\xB8\x68"
-     "\x00\x80"
-     "\xE1\x28\x00\x00\xBC\x68\x00\x80\xA3\x5D\x02\x00\xC0\x68\x00\x80\xC4\x68"
-     "\x00\x80"
-     "\xC8\x68\x00\x80\xCC\x68\x00\x80\xA6\x15\x02\x00\xA5\x75\x02\x00\xD0\x68"
-     "\x00\x80"
-     "\xAB\x39\x02\x00\xAA\x31\x02\x00\xD4\x68\x00\x80\xD8\x68\x00\x80\xAF\xD5"
-     "\x02\x00"
-     "\xAE\xD5\x02\x00\xAD\x15\x02\x00\xAC\x15\x02\x00\xE3\xA0\x06\x00\xDC\x68"
-     "\x00\x80"
-     "\xE1\x68\x06\x00\xBE\xC4\x1C\x00\x80\xFD\x00\x00\x81\x11\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xE0\x68\x00\x80\xE8\x68\x00\x80\xEC\x68\x00\x80\xF0\x68\x00\x80\x24\x67"
-     "\x00\x80"
-     "\xEF\x84\x06\x00\xF4\x68\x00\x80\xF8\x68\x00\x80\xFC\x68\x00\x80\xB3\x61"
-     "\x01\x00"
-     "\x00\x69\x00\x80\x86\x68\x1C\x00\x87\x6C\x1D\x00\x04\x69\x00\x80\xB6\x6D"
-     "\x01\x00"
-     "\xB5\x71\x01\x00\x08\x69\x00\x80\xBB\xED\x00\x00\xBA\xED\x00\x00\x0C\x69"
-     "\x00\x80"
-     "\x10\x69\x00\x80\xBF\xD5\x00\x00\xBE\xE9\x00\x00\xBD\xE1\x00\x00\xBC\xE9"
-     "\x00\x00"
-     "\xE4\x68\x00\x80\xEF\x50\x1C\x00\x14\x69\x00\x80\x18\x69\x00\x80\x1C\x69"
-     "\x00\x80"
-     "\x20\x69\x00\x80\x24\x69\x00\x80\x28\x69\x00\x80\x2C\x69\x00\x80\x30\x69"
-     "\x00\x80"
-     "\x34\x69\x00\x80\x38\x69\x00\x80\x3C\x69\x00\x80\xE3\x34\x1D\x00\x40\x69"
-     "\x00\x80"
-     "\xE1\x90\x1C\x00\xA3\xE1\x1E\x00\x44\x69\x00\x80\x48\x69\x00\x80\x4C\x69"
-     "\x00\x80"
-     "\x50\x69\x00\x80\xA6\xED\x1E\x00\xA5\xF1\x1E\x00\x85\x1C\x00\x00\xAB\x6D"
-     "\x1F\x00"
-     "\xAA\x6D\x1F\x00\xBF\x60\x01\x00\x54\x69\x00\x80\xAF\x55\x1F\x00\xAE\x69"
-     "\x1F\x00"
-     "\xAD\x61\x1F\x00\xAC\x69\x1F\x00\xA8\xDD\x1E\x00\xA9\xE1\x1E\x00\xAA\x29"
-     "\x1E\x00"
-     "\xAB\x29\x1E\x00\xAC\x39\x1E\x00\xAD\x2D\x1E\x00\xAE\x25\x1E\x00\xAF\x99"
-     "\x1F\x00"
-     "\x80\xE9\x00\x00\x81\xE9\x00\x00\x82\xF9\x00\x00\x58\x69\x00\x80\x86\x00"
-     "\x04\x00"
-     "\x87\x14\x00\x00\x5C\x69\x00\x80\x60\x69\x00\x80\xB8\xF5\x1F\x00\xB9\xF9"
-     "\x1F\x00"
-     "\xBA\x8D\x1F\x00\xBB\x85\x1F\x00\xBC\x9D\x1F\x00\xBD\x85\x1F\x00\xBE\x85"
-     "\x1F\x00"
-     "\xBF\xB5\x1F\x00\xB0\xE9\x1F\x00\xB1\xE9\x1F\x00\xB2\xFD\x1F\x00\xB3\xF5"
-     "\x1F\x00"
-     "\xB4\xED\x1F\x00\xB5\xD1\x1F\x00\xB6\xD1\x1F\x00\xB7\xD1\x1F\x00\xA8\xA1"
-     "\x1E\x00"
-     "\xA9\xA1\x1E\x00\xAA\xA1\x1E\x00\xAB\xB9\x1E\x00\xAC\xA9\x1E\x00\xAD\xA9"
-     "\x1E\x00"
-     "\xAE\xDD\x1E\x00\xAF\xCD\x1E\x00\x64\x69\x00\x80\x68\x69\x00\x80\x6C\x69"
-     "\x00\x80"
-     "\x70\x69\x00\x80\x74\x69\x00\x80\x78\x69\x00\x80\x7C\x69\x00\x80\x80\x69"
-     "\x00\x80"
-     "\xB8\x5D\x01\x00\xB9\x65\x01\x00\xBA\x6D\x01\x00\xBB\x65\x01\x00\xBC\x7D"
-     "\x01\x00"
-     "\xBD\x65\x01\x00\xBE\x61\x01\x00\xBF\x19\x01\x00\xB0\xBD\x1E\x00\xB1\x8D"
-     "\x1E\x00"
-     "\xB2\x85\x1E\x00\xB3\x65\x01\x00\xB4\x7D\x01\x00\xB5\x65\x01\x00\xB6\x6D"
-     "\x01\x00"
-     "\xB7\x65\x01\x00\xB3\x29\x1E\x00\x84\x69\x00\x80\x88\x69\x00\x80\x8C\x69"
-     "\x00\x80"
-     "\x90\x69\x00\x80\xB6\x29\x1E\x00\xB5\x21\x1E\x00\x94\x69\x00\x80\xBB\x49"
-     "\x1E\x00"
-     "\xBA\x79\x1E\x00\x98\x69\x00\x80\x9C\x69\x00\x80\xBF\xA1\x01\x00\xBE\xB9"
-     "\x01\x00"
-     "\xBD\xB1\x01\x00\xBC\x51\x1E\x00\xA0\x69\x00\x80\xA3\x6D\x1E\x00\xA4\x69"
-     "\x00\x80"
-     "\xA8\x69\x00\x80\xA6\x6D\x1E\x00\xAC\x69\x00\x80\xB0\x69\x00\x80\xA5\x65"
-     "\x1E\x00"
-     "\xAA\x3D\x1E\x00\xAB\x0D\x1E\x00\xB4\x69\x00\x80\xB8\x69\x00\x80\xAE\xFD"
-     "\x01\x00"
-     "\xAF\xE5\x01\x00\xAC\x15\x1E\x00\xAD\xF5\x01\x00\x8F\x09\x2A\x00\xE1\x4C"
-     "\x00\x00"
-     "\xBC\x69\x00\x80\xE3\x58\x0A\x00\xC0\x69\x00\x80\x84\xD8\x04\x00\xBE\xD0"
-     "\x04\x00"
-     "\xC8\x69\x00\x80\xBE\xE8\x05\x00\xCC\x69\x00\x80\x87\x88\x03\x00\x86\x0C"
-     "\x04\x00"
-     "\xD0\x69\x00\x80\xD4\x69\x00\x80\xD8\x69\x00\x80\xEF\x4C\x7B\x00\x9F\xED"
-     "\x1D\x00"
-     "\x9E\x55\x1B\x00\x9D\x41\x1A\x00\x9C\x49\x1A\x00\x9B\x8D\x19\x00\x9A\x45"
-     "\x16\x00"
-     "\x99\x5D\x16\x00\x98\x9D\x15\x00\x97\x0D\x12\x00\x96\x61\x12\x00\x95\x9D"
-     "\x11\x00"
-     "\x94\x1D\x2E\x00\x93\x75\x2E\x00\x92\x8D\x2D\x00\x91\xD5\x2A\x00\x90\x71"
-     "\x2A\x00"
-     "\xB0\x29\x76\x00\xB1\x01\x78\x00\xB2\xA9\x7A\x00\xB3\xB1\x7A\x00\xB4\x01"
-     "\x7C\x00"
-     "\xB5\x91\x7E\x00\xDC\x69\x00\x80\xE0\x69\x00\x80\x80\x4D\x03\x00\x81\xAD"
-     "\x00\x00"
-     "\x82\xB9\x00\x00\xE4\x69\x00\x80\xA3\xB1\x02\x00\xA2\x0D\x01\x00\xA1\x05"
-     "\x1E\x00"
-     "\xA0\xB1\x1E\x00\xE8\x69\x00\x80\xA1\xA9\x02\x00\xA2\xCD\x06\x00\xA3\xC5"
-     "\x06\x00"
-     "\xA4\x01\x08\x00\xA5\xD1\x0A\x00\xA6\xD1\x0A\x00\xA7\xB9\x0B\x00\xA8\xE1"
-     "\x0E\x00"
-     "\xA9\xE1\x0E\x00\xAA\xF9\x0F\x00\xAB\x01\x70\x00\xAC\xD1\x72\x00\xAD\x2D"
-     "\x72\x00"
-     "\xAE\x01\x74\x00\xAF\xDD\x76\x00\x82\x09\x2B\x00\x83\x09\x2B\x00\xEC\x69"
-     "\x00\x80"
-     "\xF0\x69\x00\x80\x86\x15\x2F\x00\x87\x65\x2F\x00\x84\xB5\x2B\x00\x85\xE5"
-     "\x2C\x00"
-     "\x8A\xD5\x10\x00\x8B\x21\x13\x00\xF4\x69\x00\x80\xF8\x69\x00\x80\x8E\x35"
-     "\x17\x00"
-     "\x8F\x25\x17\x00\x8C\x55\x13\x00\x8D\xFD\x14\x00\x92\x9D\x18\x00\x93\x41"
-     "\x1B\x00"
-     "\x84\x84\x04\x00\x84\x08\x07\x00\x96\xFD\x1C\x00\x97\x59\x1F\x00\x94\x51"
-     "\x1B\x00"
-     "\x95\x4D\x1A\x00\x9A\xDD\x1F\x00\x9B\xDD\x00\x00\xFC\x69\x00\x80\x00\x6A"
-     "\x00\x80"
-     "\x04\x6A\x00\x80\x08\x6A\x00\x80\x9C\x79\x03\x00\x0C\x6A\x00\x80\x10\x6A"
-     "\x00\x80"
-     "\xE1\x54\x1E\x00\x14\x6A\x00\x80\xE3\xBC\x1E\x00\x18\x6A\x00\x80\x1C\x6A"
-     "\x00\x80"
-     "\x20\x6A\x00\x80\x24\x6A\x00\x80\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\x28\x6A\x00\x80\x2C\x6A\x00\x80\x30\x6A\x00\x80\x74\x68\x00\x80\xEF\xC8"
-     "\x1E\x00"
-     "\x87\x6C\x05\x00\x86\xE4\x04\x00\x38\x6A\x00\x80\x3C\x6A\x00\x80\xEF\x30"
-     "\x03\x00"
-     "\x40\x6A\x00\x80\x44\x6A\x00\x80\x48\x6A\x00\x80\xE3\xE4\x03\x00\x4C\x6A"
-     "\x00\x80"
-     "\xE1\x40\x01\x00\x50\x6A\x00\x80\x54\x6A\x00\x80\x58\x6A\x00\x80\x5C\x6A"
-     "\x00\x80"
-     "\x60\x6A\x00\x80\xB3\xBD\x04\x00\xC4\x69\x00\x80\x34\x6A\x00\x80\x64\x6A"
-     "\x00\x80"
-     "\x68\x6A\x00\x80\xB6\x3D\x05\x00\xB5\x3D\x05\x00\x6C\x6A\x00\x80\xBB\x19"
-     "\x05\x00"
-     "\xBA\x11\x05\x00\x70\x6A\x00\x80\x74\x6A\x00\x80\xBF\xBD\x05\x00\xBE\xBD"
-     "\x05\x00"
-     "\xBD\xBD\x05\x00\xBC\x01\x05\x00\x78\x6A\x00\x80\xA3\xF9\x04\x00\x7C\x6A"
-     "\x00\x80"
-     "\x80\x6A\x00\x80\xA6\x79\x05\x00\x84\x6A\x00\x80\x88\x6A\x00\x80\xA5\x79"
-     "\x05\x00"
-     "\xAA\x55\x05\x00\xAB\x5D\x05\x00\x8C\x6A\x00\x80\x90\x6A\x00\x80\xAE\xF9"
-     "\x05\x00"
-     "\xAF\xF9\x05\x00\xAC\x45\x05\x00\xAD\xF9\x05\x00\xA8\x9D\x06\x00\xA9\xA1"
-     "\x06\x00"
-     "\xAA\xB9\x06\x00\xAB\xB9\x06\x00\xAC\xA9\x06\x00\xAD\x0D\x06\x00\xAE\x05"
-     "\x06\x00"
-     "\xAF\x31\x06\x00\x94\x6A\x00\x80\x82\x19\x00\x00\x81\x11\x00\x00\x80\x19"
-     "\x00\x00"
-     "\xBE\xEC\x00\x00\x98\x6A\x00\x80\x9C\x6A\x00\x80\xA0\x6A\x00\x80\xB8\xB5"
-     "\x07\x00"
-     "\xB9\xB9\x07\x00\xBA\x49\x07\x00\xBB\x45\x07\x00\xBC\x41\x07\x00\xBD\x41"
-     "\x07\x00"
-     "\xBE\x41\x07\x00\xBF\x41\x07\x00\xB0\x21\x06\x00\xB1\xED\x07\x00\xB2\xE5"
-     "\x07\x00"
-     "\xB3\xC5\x07\x00\xB4\xDD\x07\x00\xB5\x9D\x07\x00\xB6\x95\x07\x00\xB7\x8D"
-     "\x07\x00"
-     "\xB3\x31\x07\x00\xA8\x6A\x00\x80\x86\xA8\x00\x00\x87\x88\x01\x00\xAC\x6A"
-     "\x00\x80"
-     "\xB6\x8D\x07\x00\xB5\x21\x07\x00\xB0\x6A\x00\x80\xBB\x29\x06\x00\xBA\xE5"
-     "\x07\x00"
-     "\xB4\x6A\x00\x80\xB8\x6A\x00\x80\xBF\x59\x06\x00\xBE\x19\x06\x00\xBD\x11"
-     "\x06\x00"
-     "\xBC\x31\x06\x00\xBC\x6A\x00\x80\xA3\x75\x07\x00\xC0\x6A\x00\x80\xC4\x6A"
-     "\x00\x80"
-     "\xA6\xC9\x07\x00\xC8\x6A\x00\x80\xCC\x6A\x00\x80\xA5\x65\x07\x00\xAA\xA1"
-     "\x07\x00"
-     "\xAB\x6D\x06\x00\xD0\x6A\x00\x80\xD4\x6A\x00\x80\xAE\x5D\x06\x00\xAF\x1D"
-     "\x06\x00"
-     "\xAC\x75\x06\x00\xAD\x55\x06\x00\xA8\x49\x06\x00\xA9\x49\x06\x00\xAA\x5D"
-     "\x06\x00"
-     "\xAB\x55\x06\x00\xAC\x4D\x06\x00\xAD\x71\x06\x00\xAE\x75\x06\x00\xAF\x6D"
-     "\x06\x00"
-     "\x84\xAC\x00\x00\xD8\x6A\x00\x80\xDC\x6A\x00\x80\xE0\x6A\x00\x80\xE4\x6A"
-     "\x00\x80"
-     "\xE8\x6A\x00\x80\xEC\x6A\x00\x80\xF0\x6A\x00\x80\xB8\x81\x01\x00\xB9\x8D"
-     "\x01\x00"
-     "\xBA\x85\x01\x00\xBB\x99\x01\x00\xBC\x89\x01\x00\xBD\xBD\x01\x00\xBE\xB5"
-     "\x01\x00"
-     "\xBF\x7D\x01\x00\xB0\x15\x06\x00\xB1\x1D\x06\x00\xB2\x15\x06\x00\xB3\xE1"
-     "\x01\x00"
-     "\xB4\xE1\x01\x00\xB5\xE9\x01\x00\xB6\xD1\x01\x00\xB7\xD1\x01\x00\xAC\x25"
-     "\x07\x00"
-     "\xAD\x2D\x07\x00\xAE\x25\x07\x00\xAF\x3D\x07\x00\xA8\xBD\x06\x00\xA9\xAD"
-     "\x06\x00"
-     "\xAA\xA5\x06\x00\xAB\x3D\x07\x00\xBE\xAC\x0C\x00\xF4\x6A\x00\x80\xF8\x6A"
-     "\x00\x80"
-     "\xFC\x6A\x00\x80\x00\x6B\x00\x80\x04\x6B\x00\x80\x08\x6B\x00\x80\x0C\x6B"
-     "\x00\x80"
-     "\xBC\x8D\x07\x00\xBD\x91\x07\x00\xBE\x91\x07\x00\xBF\xA5\x07\x00\xB8\xF9"
-     "\x07\x00"
-     "\xB9\xF9\x07\x00\xBA\x99\x07\x00\xBB\x95\x07\x00\xB4\xE9\x07\x00\xB5\xE9"
-     "\x07\x00"
-     "\xB6\xF9\x07\x00\xB7\xF1\x07\x00\xB0\x45\x07\x00\xB1\x4D\x07\x00\xB2\x45"
-     "\x07\x00"
-     "\xB3\xF9\x07\x00\x81\xBD\x00\x00\x80\xBD\x00\x00\xB3\x5D\x05\x00\x82\x19"
-     "\x00\x00"
-     "\xB5\x4D\x05\x00\x10\x6B\x00\x80\x18\x6B\x00\x80\xB6\x79\x05\x00\x1C\x6B"
-     "\x00\x80"
-     "\x20\x6B\x00\x80\xBB\x51\x05\x00\xBA\x59\x05\x00\xBD\xBD\x02\x00\xBC\x49"
-     "\x05\x00"
-     "\xBF\xA5\x02\x00\xBE\xAD\x02\x00\x24\x6B\x00\x80\xA6\x21\x05\x00\xA5\x15"
-     "\x05\x00"
-     "\x84\x4C\x0C\x00\xA3\x05\x05\x00\x28\x6B\x00\x80\x86\xF8\x0C\x00\x87\x9C"
-     "\x0C\x00"
-     "\xAF\xFD\x02\x00\xAE\xF5\x02\x00\xAD\xE5\x02\x00\xAC\x11\x05\x00\xAB\x09"
-     "\x05\x00"
-     "\xAA\x01\x05\x00\x2C\x6B\x00\x80\x30\x6B\x00\x80\xA8\x05\x02\x00\xA9\x15"
-     "\x02\x00"
-     "\xAA\x15\x02\x00\xAB\x25\x02\x00\xAC\x3D\x02\x00\xAD\x55\x02\x00\xAE\x51"
-     "\x02\x00"
-     "\xAF\x51\x02\x00\xBE\xAC\x0C\x00\x34\x6B\x00\x80\x38\x6B\x00\x80\x3C\x6B"
-     "\x00\x80"
-     "\x40\x6B\x00\x80\x44\x6B\x00\x80\x48\x6B\x00\x80\x4C\x6B\x00\x80\xB8\xF5"
-     "\x02\x00"
-     "\xB9\xFD\x02\x00\xBA\xF5\x02\x00\xBB\x69\x01\x00\xBC\x79\x01\x00\xBD\x79"
-     "\x01\x00"
-     "\xBE\x69\x01\x00\xBF\x69\x01\x00\xB0\x35\x02\x00\xB1\x3D\x02\x00\xB2\x01"
-     "\x02\x00"
-     "\xB3\x01\x02\x00\xB4\xD5\x02\x00\xB5\xDD\x02\x00\xB6\xD5\x02\x00\xB7\xCD"
-     "\x02\x00"
-     "\xE1\xB4\x07\x00\xE3\xD8\x06\x00\xE3\xD0\x06\x00\xE1\x50\x06\x00\x50\x6B"
-     "\x00\x80"
-     "\x54\x6B\x00\x80\x58\x6B\x00\x80\x5C\x6B\x00\x80\x60\x6B\x00\x80\x64\x6B"
-     "\x00\x80"
-     "\x68\x6B\x00\x80\x6C\x6B\x00\x80\xEC\x00\x00\x00\xEF\xFC\x01\x00\xEF\x60"
-     "\x07\x00"
-     "\x70\x6B\x00\x80\xB3\xDD\x02\x00\x74\x6B\x00\x80\x86\xE8\x0C\x00\x87\x8C"
-     "\x0C\x00"
-     "\x78\x6B\x00\x80\xB6\x95\x02\x00\xB5\xF5\x02\x00\x7C\x6B\x00\x80\xBB\xB9"
-     "\x02\x00"
-     "\xBA\xB1\x02\x00\x80\x6B\x00\x80\x88\x6B\x00\x80\xBF\x59\x02\x00\xBE\x51"
-     "\x02\x00"
-     "\xBD\x95\x02\x00\xBC\x95\x02\x00\xA3\x1D\x0D\x00\x14\x6B\x00\x80\x84\x6B"
-     "\x00\x80"
-     "\x8C\x6B\x00\x80\x90\x6B\x00\x80\xA6\x55\x0D\x00\xA5\x35\x0D\x00\x94\x6B"
-     "\x00\x80"
-     "\xAB\x79\x0D\x00\xAA\x71\x0D\x00\x98\x6B\x00\x80\x9C\x6B\x00\x80\xAF\x99"
-     "\x0D\x00"
-     "\xAE\x91\x0D\x00\xAD\x55\x0D\x00\xAC\x55\x0D\x00\x80\x91\x0F\x00\x81\x99"
-     "\x0F\x00"
-     "\x82\xA1\x0F\x00\xB3\xE9\x0E\x00\xA0\x6B\x00\x80\xB5\x81\x0E\x00\xB6\xA1"
-     "\x0E\x00"
-     "\xA4\x6B\x00\x80\xA8\x6B\x00\x80\xAC\x6B\x00\x80\xBA\x41\x0F\x00\xBB\x41"
-     "\x0F\x00"
-     "\xBC\x41\x0F\x00\xBD\x41\x0F\x00\xBE\x41\x0F\x00\xBF\x41\x0F\x00\xA8\x19"
-     "\x0E\x00"
-     "\xA9\x25\x0E\x00\xAA\x2D\x0E\x00\xAB\x4D\x0E\x00\xAC\x45\x0E\x00\xAD\x6D"
-     "\x0E\x00"
-     "\xAE\x65\x0E\x00\xAF\xB9\x0F\x00\xB0\x6B\x00\x80\xB4\x6B\x00\x80\xB8\x6B"
-     "\x00\x80"
-     "\xBC\x6B\x00\x80\xC0\x6B\x00\x80\xC4\x6B\x00\x80\xC8\x6B\x00\x80\xCC\x6B"
-     "\x00\x80"
-     "\xB8\x59\x0F\x00\xB9\x59\x0F\x00\xBA\x69\x0F\x00\xBB\x69\x0F\x00\xBC\x79"
-     "\x0F\x00"
-     "\xBD\x79\x0F\x00\xBE\x65\x0F\x00\xBF\x19\x0F\x00\xB0\xC9\x0F\x00\xB1\xC9"
-     "\x0F\x00"
-     "\xB2\xD9\x0F\x00\xB3\xD9\x0F\x00\xB4\xC9\x0F\x00\xB5\xC9\x0F\x00\xB6\x69"
-     "\x0F\x00"
-     "\xB7\x69\x0F\x00\xA3\xA5\x0F\x00\xD0\x6B\x00\x80\xD4\x6B\x00\x80\xD8\x6B"
-     "\x00\x80"
-     "\xDC\x6B\x00\x80\xA6\xED\x0F\x00\xA5\xCD\x0F\x00\xE0\x6B\x00\x80\xAB\x0D"
-     "\x0E\x00"
-     "\xAA\x0D\x0E\x00\xE4\x6B\x00\x80\xE8\x6B\x00\x80\xAF\x0D\x0E\x00\xAE\x0D"
-     "\x0E\x00"
-     "\xAD\x0D\x0E\x00\xAC\x0D\x0E\x00\x80\x6D\x00\x00\x81\x09\x00\x00\x82\x19"
-     "\x00\x00"
-     "\xEC\x6B\x00\x80\xF0\x6B\x00\x80\xE1\x44\x0E\x00\x84\x94\x01\x00\xE3\xEC"
-     "\x0E\x00"
-     "\xBE\xA8\x01\x00\xF4\x6B\x00\x80\xF8\x6B\x00\x80\xEF\xB4\x0D\x00\x86\xD0"
-     "\x00\x00"
-     "\x87\xFC\x01\x00\xFC\x6B\x00\x80\x00\x6C\x00\x80\xA8\x6D\x0E\x00\xA9\x89"
-     "\x01\x00"
-     "\xAA\x99\x01\x00\xAB\x91\x01\x00\xAC\xB9\x01\x00\xAD\xB9\x01\x00\xAE\xC9"
-     "\x01\x00"
-     "\xAF\xC1\x01\x00\x84\xA0\x00\x00\x04\x6C\x00\x80\x08\x6C\x00\x80\x0C\x6C"
-     "\x00\x80"
-     "\x10\x6C\x00\x80\x14\x6C\x00\x80\x18\x6C\x00\x80\x1C\x6C\x00\x80\xB8\x79"
-     "\x01\x00"
-     "\xB9\x79\x01\x00\xBA\xCD\x00\x00\xBB\xC5\x00\x00\xBC\xDD\x00\x00\xBD\xC5"
-     "\x00\x00"
-     "\xBE\xC5\x00\x00\xBF\xF5\x00\x00\xB0\x81\x01\x00\xB1\x81\x01\x00\xB2\x49"
-     "\x01\x00"
-     "\xB3\x49\x01\x00\xB4\x59\x01\x00\xB5\x59\x01\x00\xB6\x49\x01\x00\xB7\x49"
-     "\x01\x00"
-     "\xB3\x11\x02\x00\x20\x6C\x00\x80\x24\x6C\x00\x80\x28\x6C\x00\x80\x2C\x6C"
-     "\x00\x80"
-     "\xB6\x39\x02\x00\xB5\x31\x02\x00\x30\x6C\x00\x80\xBB\x45\x02\x00\xBA\x45"
-     "\x02\x00"
-     "\x34\x6C\x00\x80\x38\x6C\x00\x80\xBF\x9D\x02\x00\xBE\x9D\x02\x00\xBD\x9D"
-     "\x02\x00"
-     "\xBC\x9D\x02\x00\x85\xFC\x75\x00\xA3\x55\x02\x00\x3C\x6C\x00\x80\x40\x6C"
-     "\x00\x80"
-     "\xA6\x7D\x02\x00\x44\x6C\x00\x80\x48\x6C\x00\x80\xA5\x75\x02\x00\xAA\x01"
-     "\x02\x00"
-     "\xAB\x01\x02\x00\x4C\x6C\x00\x80\x50\x6C\x00\x80\xAE\xD9\x02\x00\xAF\xD9"
-     "\x02\x00"
-     "\xAC\xD9\x02\x00\xAD\xD9\x02\x00\x82\x19\x00\x00\x54\x6C\x00\x80\x80\x19"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x58\x6C\x00\x80\x5C\x6C\x00\x80\x60\x6C\x00\x80\x68\x6C"
-     "\x00\x80"
-     "\xBE\x00\x04\x00\x6C\x6C\x00\x80\x70\x6C\x00\x80\x74\x6C\x00\x80\x78\x6C"
-     "\x00\x80"
-     "\x7C\x6C\x00\x80\x80\x6C\x00\x80\x84\x6C\x00\x80\x86\xE0\x04\x00\x87\x54"
-     "\x03\x00"
-     "\x88\x6C\x00\x80\x8C\x6C\x00\x80\xA4\x6A\x00\x80\x90\x6C\x00\x80\xEF\x80"
-     "\x02\x00"
-     "\x94\x6C\x00\x80\xE1\x88\x01\x00\x98\x6C\x00\x80\xE3\x58\x01\x00\x9C\x6C"
-     "\x00\x80"
-     "\xA0\x6C\x00\x80\xA4\x6C\x00\x80\xA8\x6C\x00\x80\xAC\x6C\x00\x80\xA8\x69"
-     "\x02\x00"
-     "\xA9\x7D\x02\x00\xAA\x75\x02\x00\xAB\xA9\x02\x00\xAC\xB9\x02\x00\xAD\xB9"
-     "\x02\x00"
-     "\xAE\xA9\x02\x00\xAF\xA9\x02\x00\x84\xAC\x05\x00\xB0\x6C\x00\x80\xB4\x6C"
-     "\x00\x80"
-     "\xB8\x6C\x00\x80\xBC\x6C\x00\x80\xC0\x6C\x00\x80\xC4\x6C\x00\x80\xC8\x6C"
-     "\x00\x80"
-     "\xB8\x6D\x01\x00\xB9\x0D\x01\x00\xBA\x05\x01\x00\xBB\x19\x01\x00\xBC\x09"
-     "\x01\x00"
-     "\xBD\x3D\x01\x00\xBE\x35\x01\x00\xBF\xD9\x01\x00\xB0\xD9\x02\x00\xB1\xD9"
-     "\x02\x00"
-     "\xB2\x6D\x01\x00\xB3\x65\x01\x00\xB4\x7D\x01\x00\xB5\x65\x01\x00\xB6\x65"
-     "\x01\x00"
-     "\xB7\x55\x01\x00\xE1\x40\x0F\x00\xE3\xCC\x00\x00\xE3\x54\x0F\x00\xE1\x5C"
-     "\x0E\x00"
-     "\x80\xAD\x00\x00\x81\x09\x00\x00\x82\x39\x00\x00\xCC\x6C\x00\x80\xD0\x6C"
-     "\x00\x80"
-     "\xBE\x28\x04\x00\xD8\x6C\x00\x80\xDC\x6C\x00\x80\xBE\x3C\x07\x00\xEF\x2C"
-     "\x00\x00"
-     "\xEF\x60\x0E\x00\xE0\x6C\x00\x80\x86\xE0\x04\x00\x87\x74\x05\x00\xE4\x6C"
-     "\x00\x80"
-     "\xB3\xB1\x02\x00\xE8\x6C\x00\x80\xB5\x7D\x02\x00\xB6\x75\x02\x00\xEC\x6C"
-     "\x00\x80"
-     "\xF0\x6C\x00\x80\xF4\x6C\x00\x80\xBA\x51\x02\x00\xBB\x51\x02\x00\xBC\xF5"
-     "\x01\x00"
-     "\xBD\xFD\x01\x00\xBE\xF5\x01\x00\xBF\xED\x01\x00\x64\x6C\x00\x80\xD4\x6C"
-     "\x00\x80"
-     "\xAA\xB1\x05\x00\xAB\xB1\x05\x00\xAC\x15\x06\x00\xAD\x1D\x06\x00\xAE\x15"
-     "\x06\x00"
-     "\xAF\x0D\x06\x00\xF8\x6C\x00\x80\xFC\x6C\x00\x80\x00\x6D\x00\x80\xA3\x51"
-     "\x05\x00"
-     "\x04\x6D\x00\x80\xA5\x9D\x05\x00\xA6\x95\x05\x00\x08\x6D\x00\x80\xB3\xE9"
-     "\x06\x00"
-     "\x0C\x6D\x00\x80\x10\x6D\x00\x80\x14\x6D\x00\x80\x18\x6D\x00\x80\xB6\xE1"
-     "\x06\x00"
-     "\xB5\xE9\x06\x00\x1C\x6D\x00\x80\xBB\xAD\x06\x00\xBA\xAD\x06\x00\x20\x6D"
-     "\x00\x80"
-     "\x24\x6D\x00\x80\xBF\x9D\x06\x00\xBE\x91\x06\x00\xBD\xA5\x06\x00\xBC\xA9"
-     "\x06\x00"
-     "\xA8\x1D\x06\x00\xA9\x21\x06\x00\xAA\x21\x06\x00\xAB\x21\x06\x00\xAC\x21"
-     "\x06\x00"
-     "\xAD\x21\x06\x00\xAE\x21\x06\x00\xAF\x21\x06\x00\x28\x6D\x00\x80\x2C\x6D"
-     "\x00\x80"
-     "\x30\x6D\x00\x80\x34\x6D\x00\x80\x80\x0D\x00\x00\x81\xB1\x01\x00\x82\xB1"
-     "\x01\x00"
-     "\x38\x6D\x00\x80\xB8\x39\x06\x00\xB9\x39\x06\x00\xBA\xED\x07\x00\xBB\xE5"
-     "\x07\x00"
-     "\xBC\xFD\x07\x00\xBD\xE5\x07\x00\xBE\xED\x07\x00\xBF\xE5\x07\x00\xB0\x29"
-     "\x06\x00"
-     "\xB1\x29\x06\x00\xB2\x3D\x06\x00\xB3\x31\x06\x00\xB4\x15\x06\x00\xB5\x19"
-     "\x06\x00"
-     "\xB6\x09\x06\x00\xB7\x09\x06\x00\xA3\xA9\x07\x00\x3C\x6D\x00\x80\x40\x6D"
-     "\x00\x80"
-     "\x84\x84\x02\x00\x44\x6D\x00\x80\xA6\xA1\x07\x00\xA5\xA9\x07\x00\x48\x6D"
-     "\x00\x80"
-     "\xAB\xED\x07\x00\xAA\xED\x07\x00\x86\x88\x00\x00\x87\x78\x01\x00\xAF\xDD"
-     "\x07\x00"
-     "\xAE\xD1\x07\x00\xAD\xE5\x07\x00\xAC\xE9\x07\x00\x4C\x6D\x00\x80\xB3\xF9"
-     "\x06\x00"
-     "\x50\x6D\x00\x80\x54\x6D\x00\x80\xB6\x8D\x06\x00\x58\x6D\x00\x80\x5C\x6D"
-     "\x00\x80"
-     "\xB5\x8D\x06\x00\xBA\x41\x01\x00\xBB\x41\x01\x00\x60\x6D\x00\x80\x64\x6D"
-     "\x00\x80"
-     "\xBE\x4D\x01\x00\xBF\x31\x01\x00\xBC\x51\x01\x00\xBD\x4D\x01\x00\xA8\x29"
-     "\x06\x00"
-     "\xA9\x29\x06\x00\xAA\x35\x06\x00\xAB\x45\x06\x00\xAC\x41\x06\x00\xAD\x49"
-     "\x06\x00"
-     "\xAE\x79\x06\x00\xAF\x79\x06\x00\x84\xAC\x01\x00\x68\x6D\x00\x80\x6C\x6D"
-     "\x00\x80"
-     "\x70\x6D\x00\x80\x74\x6D\x00\x80\x78\x6D\x00\x80\x7C\x6D\x00\x80\x80\x6D"
-     "\x00\x80"
-     "\xB8\x91\x01\x00\xB9\x99\x01\x00\xBA\xA1\x01\x00\xBB\xA1\x01\x00\xBC\x71"
-     "\x01\x00"
-     "\xBD\x71\x01\x00\xBE\x71\x01\x00\xBF\x71\x01\x00\xB0\xC1\x01\x00\xB1\xC1"
-     "\x01\x00"
-     "\xB2\xC1\x01\x00\xB3\xDD\x01\x00\xB4\xCD\x01\x00\xB5\xB1\x01\x00\xB6\xB1"
-     "\x01\x00"
-     "\xB7\xB1\x01\x00\xA3\xBD\x05\x00\x84\x6D\x00\x80\x88\x6D\x00\x80\x8C\x6D"
-     "\x00\x80"
-     "\x90\x6D\x00\x80\xA6\xC9\x05\x00\xA5\xC9\x05\x00\x94\x6D\x00\x80\xAB\x05"
-     "\x02\x00"
-     "\xAA\x05\x02\x00\x98\x6D\x00\x80\x9C\x6D\x00\x80\xAF\x75\x02\x00\xAE\x09"
-     "\x02\x00"
-     "\xAD\x09\x02\x00\xAC\x15\x02\x00\xA0\x6D\x00\x80\xA4\x6D\x00\x80\xA8\x6D"
-     "\x00\x80"
-     "\xAC\x6D\x00\x80\x81\x1D\x00\x00\x80\x1D\x00\x00\xB0\x6D\x00\x80\x82\x05"
-     "\x00\x00"
-     "\xBE\x24\x03\x00\xB8\x6D\x00\x80\x86\xC8\x7C\x00\x87\x5C\x03\x00\xBC\x6D"
-     "\x00\x80"
-     "\xC0\x6D\x00\x80\xC4\x6D\x00\x80\xC8\x6D\x00\x80\xCC\x6D\x00\x80\xD0\x6D"
-     "\x00\x80"
-     "\xD4\x6D\x00\x80\x84\xC0\x7C\x00\xD8\x6D\x00\x80\xE1\xA4\x01\x00\xDC\x6D"
-     "\x00\x80"
-     "\xE3\x2C\x01\x00\xBE\xE8\x7C\x00\xE0\x6D\x00\x80\xE4\x6D\x00\x80\xEF\x34"
-     "\x02\x00"
-     "\xAE\x00\x00\x00\xE8\x6D\x00\x80\xEC\x6D\x00\x80\xF0\x6D\x00\x80\xB3\x89"
-     "\x03\x00"
-     "\xF4\x6D\x00\x80\xF8\x6D\x00\x80\xFC\x6D\x00\x80\x00\x6E\x00\x80\xB6\x8D"
-     "\x03\x00"
-     "\xB5\x8D\x03\x00\x04\x6E\x00\x80\xBB\x49\x03\x00\xBA\x41\x03\x00\x08\x6E"
-     "\x00\x80"
-     "\x0C\x6E\x00\x80\xBF\x49\x03\x00\xBE\x41\x03\x00\xBD\x49\x03\x00\xBC\x51"
-     "\x03\x00"
-     "\xA8\x1D\x02\x00\xA9\x45\x02\x00\xAA\x41\x02\x00\xAB\x41\x02\x00\xAC\x45"
-     "\x02\x00"
-     "\xAD\x4D\x02\x00\xAE\x71\x02\x00\xAF\x71\x02\x00\xBE\xA0\x7F\x00\x10\x6E"
-     "\x00\x80"
-     "\x14\x6E\x00\x80\x18\x6E\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xA9"
-     "\x00\x00"
-     "\x1C\x6E\x00\x80\xB8\x49\x01\x00\xB9\x49\x01\x00\xBA\x59\x01\x00\xBB\x51"
-     "\x01\x00"
-     "\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x19\x01\x00\xBF\x19\x01\x00\xB0\xC5"
-     "\x02\x00"
-     "\xB1\xCD\x02\x00\xB2\xC5\x02\x00\xB3\xDD\x02\x00\xB4\xC5\x02\x00\xB5\xCD"
-     "\x02\x00"
-     "\xB6\xC5\x02\x00\xB7\x79\x01\x00\xE1\x58\x06\x00\xE1\x5C\x07\x00\xE3\x94"
-     "\x00\x00"
-     "\xE3\xBC\x06\x00\x20\x6E\x00\x80\x28\x6E\x00\x80\x86\xB8\x7C\x00\x87\x54"
-     "\x7D\x00"
-     "\x2C\x6E\x00\x80\x30\x6E\x00\x80\x34\x6E\x00\x80\x38\x6E\x00\x80\xEE\x00"
-     "\x00\x00"
-     "\x3C\x6E\x00\x80\xEF\xF0\x06\x00\xEF\xE0\x06\x00\x40\x6E\x00\x80\x44\x6E"
-     "\x00\x80"
-     "\xA3\x81\x02\x00\x48\x6E\x00\x80\xA5\x85\x02\x00\x4C\x6E\x00\x80\x50\x6E"
-     "\x00\x80"
-     "\xA6\x85\x02\x00\x54\x6E\x00\x80\x58\x6E\x00\x80\xAB\x41\x02\x00\xAA\x49"
-     "\x02\x00"
-     "\xAD\x41\x02\x00\xAC\x59\x02\x00\xAF\x41\x02\x00\xAE\x49\x02\x00\xA8\x65"
-     "\x7E\x00"
-     "\xA9\x75\x7E\x00\xAA\x7D\x7E\x00\xAB\x75\x7E\x00\xAC\x6D\x7E\x00\xAD\x95"
-     "\x7E\x00"
-     "\xAE\x9D\x7E\x00\xAF\x95\x7E\x00\x24\x6E\x00\x80\x5C\x6E\x00\x80\x60\x6E"
-     "\x00\x80"
-     "\x64\x6E\x00\x80\x68\x6E\x00\x80\x6C\x6E\x00\x80\x70\x6E\x00\x80\x74\x6E"
-     "\x00\x80"
-     "\xB8\x89\x7E\x00\xB9\x89\x7E\x00\xBA\x99\x7E\x00\xBB\x91\x7E\x00\xBC\xB9"
-     "\x7E\x00"
-     "\xBD\xB9\x7E\x00\xBE\x7D\x7F\x00\xBF\x75\x7F\x00\xB0\xF1\x7E\x00\xB1\xFD"
-     "\x7E\x00"
-     "\xB2\xCD\x7E\x00\xB3\xC5\x7E\x00\xB4\xC5\x7E\x00\xB5\xCD\x7E\x00\xB6\xC5"
-     "\x7E\x00"
-     "\xB7\xB9\x7E\x00\xB3\xF5\x7E\x00\x78\x6E\x00\x80\x7C\x6E\x00\x80\x80\x6E"
-     "\x00\x80"
-     "\x84\x6E\x00\x80\xB6\x15\x7E\x00\xB5\xE5\x7E\x00\x88\x6E\x00\x80\xBB\x39"
-     "\x7E\x00"
-     "\xBA\x35\x7E\x00\x8C\x6E\x00\x80\x90\x6E\x00\x80\xBF\x09\x7E\x00\xBE\x01"
-     "\x7E\x00"
-     "\xBD\x11\x7E\x00\xBC\x19\x7E\x00\x82\x6D\x00\x00\xA3\xB1\x7E\x00\x80\x55"
-     "\x00\x00"
-     "\x81\x65\x00\x00\xA6\x51\x7E\x00\x84\x18\x03\x00\x94\x6E\x00\x80\xA5\xA1"
-     "\x7E\x00"
-     "\xAA\x71\x7E\x00\xAB\x7D\x7E\x00\x86\x88\x00\x00\x87\x5C\x03\x00\xAE\x45"
-     "\x7E\x00"
-     "\xAF\x4D\x7E\x00\xAC\x5D\x7E\x00\xAD\x55\x7E\x00\xB3\xA5\x7E\x00\x98\x6E"
-     "\x00\x80"
-     "\x9C\x6E\x00\x80\xA0\x6E\x00\x80\xA4\x6E\x00\x80\xB6\xA1\x7E\x00\xB5\xB5"
-     "\x7E\x00"
-     "\xA8\x6E\x00\x80\xBB\xF5\x7E\x00\xBA\xE9\x7E\x00\xAC\x6E\x00\x80\xB0\x6E"
-     "\x00\x80"
-     "\xBF\x15\x01\x00\xBE\xD1\x7E\x00\xBD\xD9\x7E\x00\xBC\xE5\x7E\x00\xB4\x6E"
-     "\x00\x80"
-     "\xB8\x6E\x00\x80\xBC\x6E\x00\x80\xC0\x6E\x00\x80\xEF\x08\x7D\x00\xC4\x6E"
-     "\x00\x80"
-     "\xC8\x6E\x00\x80\xCC\x6E\x00\x80\xE3\xFC\x7D\x00\xD0\x6E\x00\x80\xE1\x18"
-     "\x7D\x00"
-     "\xD4\x6E\x00\x80\xD8\x6E\x00\x80\xDC\x6E\x00\x80\xE0\x6E\x00\x80\xE4\x6E"
-     "\x00\x80"
-     "\xA8\x15\x7E\x00\xA9\x25\x7E\x00\xAA\x2D\x7E\x00\xAB\x25\x7E\x00\xAC\x3D"
-     "\x7E\x00"
-     "\xAD\x21\x7E\x00\xAE\x59\x7E\x00\xAF\x59\x7E\x00\xE8\x6E\x00\x80\xEC\x6E"
-     "\x00\x80"
-     "\xF0\x6E\x00\x80\xF4\x6E\x00\x80\x80\x0D\x00\x00\x81\xB5\x03\x00\x82\xBD"
-     "\x03\x00"
-     "\xF8\x6E\x00\x80\xB8\xF1\x01\x00\xB9\xF1\x01\x00\xBA\xF1\x01\x00\xBB\xF1"
-     "\x01\x00"
-     "\xBC\x55\x01\x00\xBD\x5D\x01\x00\xBE\x55\x01\x00\xBF\x4D\x01\x00\xB0\x2D"
-     "\x7E\x00"
-     "\xB1\x31\x7E\x00\xB2\x31\x7E\x00\xB3\x0D\x7E\x00\xB4\x15\x7E\x00\xB5\xD1"
-     "\x01\x00"
-     "\xB6\xD1\x01\x00\xB7\xD1\x01\x00\xFC\x6E\x00\x80\x00\x6F\x00\x80\x84\x90"
-     "\x02\x00"
-     "\xA3\xA9\x7D\x00\xBE\x9C\x03\x00\xA5\xB9\x7D\x00\xA6\xAD\x7D\x00\x08\x6F"
-     "\x00\x80"
-     "\x0C\x6F\x00\x80\x10\x6F\x00\x80\xAA\xE5\x7D\x00\xAB\xF9\x7D\x00\xAC\xE9"
-     "\x7D\x00"
-     "\xAD\xD5\x7D\x00\xAE\xDD\x7D\x00\xAF\x19\x02\x00\xB6\x69\x03\x00\x14\x6F"
-     "\x00\x80"
-     "\x84\x48\x03\x00\xB5\x51\x03\x00\x18\x6F\x00\x80\xB3\x59\x03\x00\x86\xF8"
-     "\x04\x00"
-     "\x87\x34\x03\x00\xBE\x39\x03\x00\xBF\x39\x03\x00\xBC\x25\x03\x00\xBD\x39"
-     "\x03\x00"
-     "\xBA\x35\x03\x00\xBB\x3D\x03\x00\x1C\x6F\x00\x80\x20\x6F\x00\x80\xB8\xF1"
-     "\x03\x00"
-     "\xB9\xF1\x03\x00\xBA\xF1\x03\x00\xBB\xF1\x03\x00\xBC\x31\x03\x00\xBD\x31"
-     "\x03\x00"
-     "\xBE\x31\x03\x00\xBF\x31\x03\x00\xB0\xE5\x03\x00\xB1\xED\x03\x00\xB2\xE5"
-     "\x03\x00"
-     "\xB3\xF9\x03\x00\xB4\xE9\x03\x00\xB5\xE9\x03\x00\xB6\xDD\x03\x00\xB7\xD1"
-     "\x03\x00"
-     "\xA8\xC5\x03\x00\xA9\xCD\x03\x00\xAA\xC5\x03\x00\xAB\xDD\x03\x00\xAC\xCD"
-     "\x03\x00"
-     "\xAD\xFD\x03\x00\xAE\xF5\x03\x00\xAF\x9D\x03\x00\x24\x6F\x00\x80\x28\x6F"
-     "\x00\x80"
-     "\x2C\x6F\x00\x80\xB4\x6D\x00\x80\xBE\xFC\x04\x00\x30\x6F\x00\x80\x34\x6F"
-     "\x00\x80"
-     "\x38\x6F\x00\x80\xA3\x0D\x02\x00\x3C\x6F\x00\x80\x40\x6F\x00\x80\x44\x6F"
-     "\x00\x80"
-     "\x48\x6F\x00\x80\xA6\x3D\x02\x00\xA5\x05\x02\x00\x4C\x6F\x00\x80\xAB\x69"
-     "\x02\x00"
-     "\xAA\x61\x02\x00\x50\x6F\x00\x80\x54\x6F\x00\x80\xAF\x6D\x02\x00\xAE\x6D"
-     "\x02\x00"
-     "\xAD\x6D\x02\x00\xAC\x71\x02\x00\x58\x6F\x00\x80\x5C\x6F\x00\x80\x60\x6F"
-     "\x00\x80"
-     "\x64\x6F\x00\x80\x68\x6F\x00\x80\xE3\x4C\x03\x00\x6C\x6F\x00\x80\xE1\xAC"
-     "\x01\x00"
-     "\x82\x11\x00\x00\xEF\x8C\x03\x00\x80\x39\x00\x00\x81\x35\x00\x00\x70\x6F"
-     "\x00\x80"
-     "\x78\x6F\x00\x80\x86\x58\x04\x00\x87\x40\x05\x00\x7C\x6F\x00\x80\x80\x6F"
-     "\x00\x80"
-     "\x84\x6F\x00\x80\x88\x6F\x00\x80\x8C\x6F\x00\x80\x90\x6F\x00\x80\x94\x6F"
-     "\x00\x80"
-     "\xEF\x44\x7E\x00\x98\x6F\x00\x80\xE1\x24\x7E\x00\x9C\x6F\x00\x80\xE3\x68"
-     "\x01\x00"
-     "\xEC\x00\x00\x00\xA0\x6F\x00\x80\xA4\x6F\x00\x80\xA8\x6F\x00\x80\xAC\x6F"
-     "\x00\x80"
-     "\xB0\x6F\x00\x80\xB4\x6F\x00\x80\xB3\xE9\x01\x00\xB8\x6F\x00\x80\xB5\x85"
-     "\x01\x00"
-     "\xB6\xA1\x01\x00\xBC\x6F\x00\x80\xC0\x6F\x00\x80\xC4\x6F\x00\x80\xBA\x0D"
-     "\x01\x00"
-     "\xBB\x05\x01\x00\xBC\x05\x01\x00\xBD\x0D\x01\x00\xBE\x05\x01\x00\xBF\xF9"
-     "\x00\x00"
-     "\xA8\x1D\x06\x00\xA9\x25\x06\x00\xAA\x2D\x06\x00\xAB\x49\x06\x00\xAC\x41"
-     "\x06\x00"
-     "\xAD\x6D\x06\x00\xAE\x65\x06\x00\xAF\xFD\x06\x00\x74\x6F\x00\x80\xC8\x6F"
-     "\x00\x80"
-     "\xCC\x6F\x00\x80\xD0\x6F\x00\x80\xD4\x6F\x00\x80\x82\xB9\x00\x00\x81\xB1"
-     "\x00\x00"
-     "\x80\xB9\x00\x00\xB8\x51\x07\x00\xB9\x59\x07\x00\xBA\x61\x07\x00\xBB\x61"
-     "\x07\x00"
-     "\xBC\x11\x07\x00\xBD\x11\x07\x00\xBE\x11\x07\x00\xBF\x11\x07\x00\xB0\x85"
-     "\x06\x00"
-     "\xB1\x8D\x06\x00\xB2\x9D\x06\x00\xB3\x95\x06\x00\xB4\x8D\x06\x00\xB5\x71"
-     "\x07\x00"
-     "\xB6\x71\x07\x00\xB7\x71\x07\x00\xA3\xA5\x06\x00\xD8\x6F\x00\x80\xDC\x6F"
-     "\x00\x80"
-     "\xE0\x6F\x00\x80\x84\x98\x02\x00\xA6\xED\x06\x00\xA5\xC9\x06\x00\xBE\x10"
-     "\x03\x00"
-     "\xAB\x49\x06\x00\xAA\x41\x06\x00\xE4\x6F\x00\x80\xE8\x6F\x00\x80\xAF\xB5"
-     "\x07\x00"
-     "\xAE\x49\x06\x00\xAD\x41\x06\x00\xAC\x49\x06\x00\xEC\x6F\x00\x80\xB3\x41"
-     "\x07\x00"
-     "\x86\x68\x00\x00\x87\x6C\x03\x00\xB6\x41\x07\x00\xF0\x6F\x00\x80\xF4\x6F"
-     "\x00\x80"
-     "\xB5\x51\x07\x00\xBA\x41\x07\x00\xBB\x4D\x07\x00\xF8\x6F\x00\x80\xFC\x6F"
-     "\x00\x80"
-     "\xBE\x4D\x07\x00\xBF\x35\x07\x00\xBC\x4D\x07\x00\xBD\x45\x07\x00\xA3\x85"
-     "\x06\x00"
-     "\x00\x70\x00\x80\x04\x70\x00\x80\x08\x70\x00\x80\x0C\x70\x00\x80\xA6\x85"
-     "\x06\x00"
-     "\xA5\x95\x06\x00\x10\x70\x00\x80\xAB\x89\x06\x00\xAA\x85\x06\x00\x14\x70"
-     "\x00\x80"
-     "\x18\x70\x00\x80\xAF\xF1\x06\x00\xAE\x89\x06\x00\xAD\x81\x06\x00\xAC\x89"
-     "\x06\x00"
-     "\x1C\x70\x00\x80\xEF\x4C\x05\x00\x20\x70\x00\x80\x24\x70\x00\x80\x28\x70"
-     "\x00\x80"
-     "\x2C\x70\x00\x80\x30\x70\x00\x80\x34\x70\x00\x80\x38\x70\x00\x80\x3C\x70"
-     "\x00\x80"
-     "\x40\x70\x00\x80\x44\x70\x00\x80\x48\x70\x00\x80\xE3\x2C\x05\x00\x4C\x70"
-     "\x00\x80"
-     "\xE1\x3C\x05\x00\xA8\xF5\x06\x00\xA9\x05\x06\x00\xAA\x0D\x06\x00\xAB\x05"
-     "\x06\x00"
-     "\xAC\x1D\x06\x00\xAD\x05\x06\x00\xAE\x0D\x06\x00\xAF\x05\x06\x00\x80\xD1"
-     "\x00\x00"
-     "\x81\xDD\x00\x00\x82\x05\x00\x00\x50\x70\x00\x80\xBE\xEC\x01\x00\x84\xF8"
-     "\x01\x00"
-     "\x58\x70\x00\x80\x5C\x70\x00\x80\xB8\x11\x06\x00\xB9\x19\x06\x00\xBA\x29"
-     "\x06\x00"
-     "\xBB\x29\x06\x00\xBC\xF5\x01\x00\xBD\xFD\x01\x00\xBE\xF5\x01\x00\xBF\xED"
-     "\x01\x00"
-     "\xB0\x41\x06\x00\xB1\x4D\x06\x00\xB2\x5D\x06\x00\xB3\x55\x06\x00\xB4\x4D"
-     "\x06\x00"
-     "\xB5\x35\x06\x00\xB6\x3D\x06\x00\xB7\x35\x06\x00\xA8\xAD\x02\x00\xA9\xC5"
-     "\x02\x00"
-     "\xAA\xDD\x02\x00\xAB\xD5\x02\x00\xAC\xF9\x02\x00\xAD\xF9\x02\x00\xAE\x1D"
-     "\x03\x00"
-     "\xAF\x15\x03\x00\x86\x80\x0C\x00\x87\x3C\x03\x00\x60\x70\x00\x80\x64\x70"
-     "\x00\x80"
-     "\x68\x70\x00\x80\x6C\x70\x00\x80\x70\x70\x00\x80\xBE\x78\x03\x00\xB8\x11"
-     "\x03\x00"
-     "\xB9\x11\x03\x00\xBA\x11\x03\x00\xBB\x11\x03\x00\xBC\x31\x03\x00\xBD\x31"
-     "\x03\x00"
-     "\xBE\x31\x03\x00\xBF\x31\x03\x00\xB0\x6D\x03\x00\xB1\x75\x03\x00\xB2\x75"
-     "\x03\x00"
-     "\xB3\x45\x03\x00\xB4\x5D\x03\x00\xB5\x35\x03\x00\xB6\x31\x03\x00\xB7\x31"
-     "\x03\x00"
-     "\xB3\x55\x02\x00\x74\x70\x00\x80\x78\x70\x00\x80\x7C\x70\x00\x80\x80\x70"
-     "\x00\x80"
-     "\xB6\x79\x02\x00\xB5\x71\x02\x00\x84\x70\x00\x80\xBB\xB1\x03\x00\xBA\xA9"
-     "\x03\x00"
-     "\x88\x70\x00\x80\x8C\x70\x00\x80\xBF\xF1\x03\x00\xBE\x99\x03\x00\xBD\x91"
-     "\x03\x00"
-     "\xBC\xA1\x03\x00\x90\x70\x00\x80\xA6\x29\x02\x00\xA5\x21\x02\x00\x94\x70"
-     "\x00\x80"
-     "\xA3\x05\x02\x00\x98\x70\x00\x80\x9C\x70\x00\x80\xA0\x70\x00\x80\xAF\xA1"
-     "\x03\x00"
-     "\xAE\xC9\x03\x00\xAD\xC1\x03\x00\xAC\xF1\x03\x00\xAB\xE1\x03\x00\xAA\xF9"
-     "\x03\x00"
-     "\x04\x6F\x00\x80\xA4\x70\x00\x80\xA8\x70\x00\x80\xAC\x70\x00\x80\xB0\x70"
-     "\x00\x80"
-     "\xB4\x70\x00\x80\x82\x25\x00\x00\xEF\x5C\x00\x00\x80\x31\x00\x00\x81\x2D"
-     "\x00\x00"
-     "\x84\xAC\x02\x00\xE3\x80\x00\x00\xB8\x70\x00\x80\xE1\xFC\x01\x00\xBC\x70"
-     "\x00\x80"
-     "\xC0\x70\x00\x80\x86\x18\x0D\x00\x87\x10\x0D\x00\xB3\x69\x01\x00\xC8\x70"
-     "\x00\x80"
-     "\xBE\x44\x0C\x00\xCC\x70\x00\x80\xD0\x70\x00\x80\xB6\x61\x01\x00\xB5\x69"
-     "\x01\x00"
-     "\xD4\x70\x00\x80\xBB\xED\x01\x00\xBA\xED\x01\x00\xD8\x70\x00\x80\xDC\x70"
-     "\x00\x80"
-     "\xBF\xDD\x01\x00\xBE\xDD\x01\x00\xBD\xE5\x01\x00\xBC\xED\x01\x00\xE0\x70"
-     "\x00\x80"
-     "\xE4\x70\x00\x80\xE8\x70\x00\x80\x84\x00\x0C\x00\xEC\x70\x00\x80\xF0\x70"
-     "\x00\x80"
-     "\xF4\x70\x00\x80\xEF\xF8\x01\x00\x85\xEC\x0B\x00\xE1\xC4\x06\x00\xF8\x70"
-     "\x00\x80"
-     "\xE3\x2C\x01\x00\xFC\x70\x00\x80\x00\x71\x00\x80\x04\x71\x00\x80\x08\x71"
-     "\x00\x80"
-     "\x0C\x71\x00\x80\x10\x71\x00\x80\xA3\xE1\x01\x00\x14\x71\x00\x80\xA5\xE1"
-     "\x01\x00"
-     "\x18\x71\x00\x80\x1C\x71\x00\x80\xA6\xE9\x01\x00\x20\x71\x00\x80\x24\x71"
-     "\x00\x80"
-     "\xAB\x65\x01\x00\xAA\x65\x01\x00\xAD\x6D\x01\x00\xAC\x65\x01\x00\xAF\x55"
-     "\x01\x00"
-     "\xAE\x55\x01\x00\xA8\xB9\x0D\x00\xA9\xB9\x0D\x00\xAA\x21\x0E\x00\xAB\x21"
-     "\x0E\x00"
-     "\xAC\x21\x0E\x00\xAD\x21\x0E\x00\xAE\x21\x0E\x00\xAF\x99\x0E\x00\xC4\x70"
-     "\x00\x80"
-     "\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\x28\x71\x00\x80\x2C\x71"
-     "\x00\x80"
-     "\x30\x71\x00\x80\x34\x71\x00\x80\xB8\x95\x0E\x00\xB9\x9D\x0E\x00\xBA\x95"
-     "\x0E\x00"
-     "\xBB\xB1\x0E\x00\xBC\x51\x0F\x00\xBD\x51\x0F\x00\xBE\x51\x0F\x00\xBF\x51"
-     "\x0F\x00"
-     "\xB0\xE9\x0E\x00\xB1\xE9\x0E\x00\xB2\xF9\x0E\x00\xB3\xF1\x0E\x00\xB4\xD9"
-     "\x0E\x00"
-     "\xB5\xD9\x0E\x00\xB6\xB9\x0E\x00\xB7\xB1\x0E\x00\xA8\xC5\x0F\x00\xA9\xC9"
-     "\x0F\x00"
-     "\xAA\xD9\x0F\x00\xAB\xD1\x0F\x00\xAC\xF9\x0F\x00\xAD\xF9\x0F\x00\xAE\x19"
-     "\x0F\x00"
-     "\xAF\x19\x0F\x00\x86\x00\x00\x00\x87\x64\x03\x00\x38\x71\x00\x80\x3C\x71"
-     "\x00\x80"
-     "\x40\x71\x00\x80\x44\x71\x00\x80\x48\x71\x00\x80\x4C\x71\x00\x80\xB8\x7D"
-     "\x0F\x00"
-     "\xB9\x05\x0F\x00\xBA\x0D\x0F\x00\xBB\x01\x0F\x00\xBC\x05\x0F\x00\xBD\x0D"
-     "\x0F\x00"
-     "\xBE\x39\x0F\x00\xBF\x39\x0F\x00\xB0\x69\x0F\x00\xB1\x75\x0F\x00\xB2\x7D"
-     "\x0F\x00"
-     "\xB3\x75\x0F\x00\xB4\x59\x0F\x00\xB5\x45\x0F\x00\xB6\x4D\x0F\x00\xB7\x45"
-     "\x0F\x00"
-     "\xE1\x38\x0D\x00\x50\x71\x00\x80\xE3\x38\x0C\x00\x54\x71\x00\x80\x58\x71"
-     "\x00\x80"
-     "\x5C\x71\x00\x80\x60\x71\x00\x80\x64\x71\x00\x80\x68\x71\x00\x80\x6C\x71"
-     "\x00\x80"
-     "\x70\x71\x00\x80\x74\x71\x00\x80\x78\x71\x00\x80\x7C\x71\x00\x80\xEF\xFC"
-     "\x0D\x00"
-     "\x80\x71\x00\x80\xB3\x11\x0E\x00\x84\x71\x00\x80\x88\x71\x00\x80\x8C\x71"
-     "\x00\x80"
-     "\x90\x71\x00\x80\xB6\x39\x0E\x00\xB5\x31\x0E\x00\x94\x71\x00\x80\xBB\xF1"
-     "\x0E\x00"
-     "\xBA\xED\x0E\x00\x98\x71\x00\x80\x9C\x71\x00\x80\xBF\xC1\x0E\x00\xBE\xDD"
-     "\x0E\x00"
-     "\xBD\xE1\x0E\x00\xBC\xE9\x0E\x00\xA3\xD1\x0E\x00\xA0\x71\x00\x80\x84\x88"
-     "\x02\x00"
-     "\xBE\x80\x01\x00\xA8\x71\x00\x80\xA6\xF9\x0E\x00\xA5\xF1\x0E\x00\xAC\x71"
-     "\x00\x80"
-     "\xAB\x31\x0E\x00\xAA\x2D\x0E\x00\x86\x08\x00\x00\x87\x4C\x01\x00\xAF\x01"
-     "\x0E\x00"
-     "\xAE\x1D\x0E\x00\xAD\x21\x0E\x00\xAC\x29\x0E\x00\x80\xA9\x00\x00\x81\xA9"
-     "\x00\x00"
-     "\x82\x4D\x00\x00\xB3\x7D\x01\x00\xB0\x71\x00\x80\xB5\x6D\x01\x00\xB6\x65"
-     "\x01\x00"
-     "\xB4\x71\x00\x80\xB8\x71\x00\x80\xBC\x71\x00\x80\xBA\xF9\x00\x00\xBB\xF9"
-     "\x00\x00"
-     "\xBC\xE9\x00\x00\xBD\xE9\x00\x00\xBE\xD9\x00\x00\xBF\xD9\x00\x00\xA8\xA9"
-     "\x02\x00"
-     "\xA9\xA9\x02\x00\xAA\xB9\x02\x00\xAB\xB9\x02\x00\xAC\xA9\x02\x00\xAD\xA9"
-     "\x02\x00"
-     "\xAE\x1D\x03\x00\xAF\x11\x03\x00\xBE\xAC\x04\x00\xC0\x71\x00\x80\xC4\x71"
-     "\x00\x80"
-     "\xBE\xE0\x05\x00\xC8\x71\x00\x80\xCC\x71\x00\x80\xD0\x71\x00\x80\xD4\x71"
-     "\x00\x80"
-     "\xB8\x71\x03\x00\xB9\x71\x03\x00\xBA\x75\x03\x00\xBB\x09\x03\x00\xBC\x19"
-     "\x03\x00"
-     "\xBD\x19\x03\x00\xBE\x09\x03\x00\xBF\x01\x03\x00\xB0\x71\x03\x00\xB1\x71"
-     "\x03\x00"
-     "\xB2\x71\x03\x00\xB3\x71\x03\x00\xB4\x51\x03\x00\xB5\x51\x03\x00\xB6\x51"
-     "\x03\x00"
-     "\xB7\x51\x03\x00\xA3\x31\x02\x00\x54\x70\x00\x80\xD8\x71\x00\x80\xDC\x71"
-     "\x00\x80"
-     "\xE0\x71\x00\x80\xA6\x29\x02\x00\xA5\x21\x02\x00\xE4\x71\x00\x80\xAB\xB5"
-     "\x03\x00"
-     "\xAA\xB5\x03\x00\xE8\x71\x00\x80\xEC\x71\x00\x80\xAF\x95\x03\x00\xAE\x95"
-     "\x03\x00"
-     "\xAD\xA5\x03\x00\xAC\xA5\x03\x00\xF0\x71\x00\x80\xF4\x71\x00\x80\xF8\x71"
-     "\x00\x80"
-     "\xFC\x71\x00\x80\x00\x72\x00\x80\x04\x72\x00\x80\x08\x72\x00\x80\x0C\x72"
-     "\x00\x80"
-     "\x80\x39\x00\x00\x81\x39\x00\x00\x82\x05\x00\x00\x10\x72\x00\x80\x18\x72"
-     "\x00\x80"
-     "\x84\xFC\x03\x00\x1C\x72\x00\x80\x20\x72\x00\x80\xA8\x91\x02\x00\xA9\x99"
-     "\x02\x00"
-     "\xAA\xD9\x02\x00\xAB\xD1\x02\x00\xAC\xF1\x02\x00\xAD\xF1\x02\x00\xAE\x3D"
-     "\x01\x00"
-     "\xAF\x35\x01\x00\x86\x00\x05\x00\x87\x64\x05\x00\x24\x72\x00\x80\x28\x72"
-     "\x00\x80"
-     "\x2C\x72\x00\x80\x30\x72\x00\x80\x34\x72\x00\x80\x38\x72\x00\x80\xB8\xE9"
-     "\x01\x00"
-     "\xB9\xE9\x01\x00\xBA\x89\x01\x00\xBB\x89\x01\x00\xBC\x99\x01\x00\xBD\x99"
-     "\x01\x00"
-     "\xBE\x89\x01\x00\xBF\x89\x01\x00\xB0\x51\x01\x00\xB1\x51\x01\x00\xB2\x51"
-     "\x01\x00"
-     "\xB3\x51\x01\x00\xB4\xF9\x01\x00\xB5\xF9\x01\x00\xB6\xE9\x01\x00\xB7\xE1"
-     "\x01\x00"
-     "\xEF\xF4\x00\x00\x3C\x72\x00\x80\x40\x72\x00\x80\xBF\x8C\x07\x00\x44\x72"
-     "\x00\x80"
-     "\x48\x72\x00\x80\x84\x18\x05\x00\xEF\x2C\x0E\x00\x4C\x72\x00\x80\xE1\x84"
-     "\x0E\x00"
-     "\x50\x72\x00\x80\xE3\x30\x0E\x00\xE3\x28\x00\x00\x54\x72\x00\x80\xE1\x00"
-     "\x01\x00"
-     "\x58\x72\x00\x80\x85\xFC\x20\x00\xB3\x15\x02\x00\x5C\x72\x00\x80\x60\x72"
-     "\x00\x80"
-     "\xB6\x35\x02\x00\x64\x72\x00\x80\x68\x72\x00\x80\xB5\x3D\x02\x00\xBA\xD1"
-     "\x01\x00"
-     "\xBB\xD1\x01\x00\x6C\x72\x00\x80\x70\x72\x00\x80\xBE\x79\x01\x00\xBF\x79"
-     "\x01\x00"
-     "\xBC\xC1\x01\x00\xBD\xC1\x01\x00\xA3\xD1\x05\x00\x14\x72\x00\x80\x74\x72"
-     "\x00\x80"
-     "\x78\x72\x00\x80\x7C\x72\x00\x80\xA6\xF1\x05\x00\xA5\xF9\x05\x00\x80\x72"
-     "\x00\x80"
-     "\xAB\x15\x06\x00\xAA\x15\x06\x00\x84\x72\x00\x80\x88\x72\x00\x80\xAF\xBD"
-     "\x06\x00"
-     "\xAE\xBD\x06\x00\xAD\x05\x06\x00\xAC\x05\x06\x00\x80\x4D\x00\x00\x81\x55"
-     "\x00\x00"
-     "\x82\x55\x00\x00\xB3\xB1\x06\x00\x8C\x72\x00\x80\xB5\x79\x07\x00\xB6\x69"
-     "\x07\x00"
-     "\x90\x72\x00\x80\x86\xC0\x01\x00\x87\xE8\x03\x00\xBA\x55\x07\x00\xBB\xE9"
-     "\x07\x00"
-     "\xBC\xF9\x07\x00\xBD\xE5\x07\x00\xBE\xED\x07\x00\xBF\xDD\x07\x00\xA8\x71"
-     "\x06\x00"
-     "\xA9\x7D\x06\x00\xAA\x75\x06\x00\xAB\xA5\x07\x00\xAC\xBD\x07\x00\xAD\xA5"
-     "\x07\x00"
-     "\xAE\xAD\x07\x00\xAF\xA5\x07\x00\x94\x72\x00\x80\x98\x72\x00\x80\x9C\x72"
-     "\x00\x80"
-     "\xA0\x72\x00\x80\xA4\x72\x00\x80\xA8\x72\x00\x80\xAC\x72\x00\x80\xB0\x72"
-     "\x00\x80"
-     "\xB8\x7D\x07\x00\xB9\x0D\x07\x00\xBA\x05\x07\x00\xBB\x19\x07\x00\xBC\x09"
-     "\x07\x00"
-     "\xBD\x3D\x07\x00\xBE\x35\x07\x00\xBF\xD5\x00\x00\xB0\xDD\x07\x00\xB1\x61"
-     "\x07\x00"
-     "\xB2\x61\x07\x00\xB3\x7D\x07\x00\xB4\x65\x07\x00\xB5\x55\x07\x00\xB6\x51"
-     "\x07\x00"
-     "\xB7\x49\x07\x00\xB4\x72\x00\x80\xB8\x72\x00\x80\xBC\x72\x00\x80\xC0\x72"
-     "\x00\x80"
-     "\xE1\xA4\x08\x00\xE1\xA4\x06\x00\xE3\xCC\x04\x00\xE3\xF8\x05\x00\xC4\x72"
-     "\x00\x80"
-     "\xC8\x72\x00\x80\xEF\x7C\x18\x00\xEF\xF4\x05\x00\xA3\x35\x01\x00\xA2\xA5"
-     "\x03\x00"
-     "\xA1\x05\x05\x00\xA0\x1D\x05\x00\xA3\xBD\x07\x00\xCC\x72\x00\x80\xD0\x72"
-     "\x00\x80"
-     "\xD4\x72\x00\x80\xD8\x72\x00\x80\xA6\x65\x06\x00\xA5\x75\x06\x00\xDC\x72"
-     "\x00\x80"
-     "\xAB\xE5\x06\x00\xAA\x59\x06\x00\xE0\x72\x00\x80\xE4\x72\x00\x80\xAF\xD1"
-     "\x06\x00"
-     "\xAE\xE1\x06\x00\xAD\xE9\x06\x00\xAC\xF5\x06\x00\x8F\xED\x7D\x00\x82\x1D"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x80\x1D\x00\x00\xE8\x72\x00\x80\xEC\x72\x00\x80\xF0\x72"
-     "\x00\x80"
-     "\xBE\x00\x1C\x00\x84\xBC\x00\x00\xF8\x72\x00\x80\x87\x98\x00\x00\x86\x2C"
-     "\x1F\x00"
-     "\xFC\x72\x00\x80\x00\x73\x00\x80\x04\x73\x00\x80\x08\x73\x00\x80\x9F\x45"
-     "\x07\x00"
-     "\x9E\x6D\x09\x00\x9D\xE9\x09\x00\x9C\x35\x0B\x00\x9B\x09\x0D\x00\x9A\xF5"
-     "\x0D\x00"
-     "\x99\x35\x0F\x00\x98\xCD\x70\x00\x97\xE5\x71\x00\x96\xE5\x71\x00\x95\x8D"
-     "\x74\x00"
-     "\x94\xE5\x75\x00\x93\xF5\x75\x00\x92\x25\x77\x00\x91\xF1\x79\x00\x90\xF9"
-     "\x79\x00"
-     "\x0C\x73\x00\x80\x10\x73\x00\x80\x82\x65\x7E\x00\x83\x09\x7A\x00\x84\x19"
-     "\x7A\x00"
-     "\x85\xC5\x74\x00\x86\x1D\x76\x00\x87\x15\x76\x00\x84\x6C\x1C\x00\x14\x73"
-     "\x00\x80"
-     "\x8A\x4D\x77\x00\x8B\x2D\x72\x00\x8C\x35\x72\x00\x8D\x25\x73\x00\x8E\xD5"
-     "\x0C\x00"
-     "\x8F\x1D\x0E\x00\x85\x04\x25\x00\x18\x73\x00\x80\x92\x91\x0E\x00\x93\xB5"
-     "\x08\x00"
-     "\x94\x71\x0A\x00\x95\xFD\x0A\x00\x96\xDD\x04\x00\x97\x7D\x06\x00\x1C\x73"
-     "\x00\x80"
-     "\x20\x73\x00\x80\x9A\x55\x06\x00\x9B\xFD\x00\x00\x9C\x75\x02\x00\x24\x73"
-     "\x00\x80"
-     "\xBE\x30\x1C\x00\x28\x73\x00\x80\xB3\xA1\x2A\x00\xB2\x01\x28\x00\xB1\xA9"
-     "\x2F\x00"
-     "\xB0\x85\x2E\x00\xA4\x71\x00\x80\x84\xD8\x1D\x00\xB5\x01\x24\x00\xB4\xA9"
-     "\x2B\x00"
-     "\x2C\x73\x00\x80\x30\x73\x00\x80\x34\x73\x00\x80\x38\x73\x00\x80\x3C\x73"
-     "\x00\x80"
-     "\x40\x73\x00\x80\x44\x73\x00\x80\x48\x73\x00\x80\xA3\xC5\x1E\x00\xA2\xC5"
-     "\x1E\x00"
-     "\xA1\xE1\x02\x00\x4C\x73\x00\x80\xA7\x01\x1A\x00\xA6\xF5\x1A\x00\xA5\x01"
-     "\x18\x00"
-     "\xA4\x05\x1E\x00\xAB\x01\x10\x00\xAA\xD5\x16\x00\xA9\xC9\x16\x00\xA8\x01"
-     "\x14\x00"
-     "\xAF\xF5\x2E\x00\xAE\x8D\x13\x00\xAD\xA5\x12\x00\xAC\xF9\x12\x00\xA8\x2D"
-     "\x03\x00"
-     "\xA9\x55\x03\x00\xAA\x55\x03\x00\xAB\xAD\x03\x00\xAC\xB5\x03\x00\xAD\xBD"
-     "\x03\x00"
-     "\xAE\xB5\x03\x00\xAF\xAD\x03\x00\x50\x73\x00\x80\x54\x73\x00\x80\x58\x73"
-     "\x00\x80"
-     "\x5C\x73\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x60\x73"
-     "\x00\x80"
-     "\xB8\x65\x02\x00\xB9\x01\x02\x00\xBA\x01\x02\x00\xBB\x01\x02\x00\xBC\x01"
-     "\x02\x00"
-     "\xBD\x09\x02\x00\xBE\x31\x02\x00\xBF\x31\x02\x00\xB0\xD5\x03\x00\xB1\xDD"
-     "\x03\x00"
-     "\xB2\xD5\x03\x00\xB3\x6D\x02\x00\xB4\x75\x02\x00\xB5\x7D\x02\x00\xB6\x75"
-     "\x02\x00"
-     "\xB7\x5D\x02\x00\xE1\xB8\x06\x00\x68\x73\x00\x80\xE3\x00\x06\x00\xBE\xC4"
-     "\x1C\x00"
-     "\x6C\x73\x00\x80\xEF\xF0\x00\x00\x86\x38\x1C\x00\x87\xBC\x1C\x00\x70\x73"
-     "\x00\x80"
-     "\xE3\x14\x00\x00\x74\x73\x00\x80\xE1\xD4\x01\x00\x86\x10\x1F\x00\x87\x84"
-     "\x1F\x00"
-     "\xEF\x6C\x07\x00\x84\xF0\x1F\x00\x78\x73\x00\x80\xB6\x6D\x03\x00\xB5\x6D"
-     "\x03\x00"
-     "\xBE\x84\x1F\x00\xB3\x8D\x03\x00\x7C\x73\x00\x80\x80\x73\x00\x80\x84\x73"
-     "\x00\x80"
-     "\xBF\xE9\x02\x00\xBE\xE1\x02\x00\xBD\xE9\x02\x00\xBC\x51\x03\x00\xBB\x49"
-     "\x03\x00"
-     "\xBA\x41\x03\x00\x88\x73\x00\x80\x8C\x73\x00\x80\xF4\x72\x00\x80\x64\x73"
-     "\x00\x80"
-     "\x90\x73\x00\x80\x94\x73\x00\x80\x98\x73\x00\x80\x9C\x73\x00\x80\xA0\x73"
-     "\x00\x80"
-     "\xA4\x73\x00\x80\xA8\x85\x1D\x00\xA9\x95\x1D\x00\xAA\x9D\x1D\x00\xAB\x95"
-     "\x1D\x00"
-     "\xAC\x8D\x1D\x00\xAD\x35\x1E\x00\xAE\x3D\x1E\x00\xAF\x19\x1E\x00\xB0\x05"
-     "\x1E\x00"
-     "\xB1\x25\x1E\x00\xB2\x2D\x1E\x00\xB3\x39\x1E\x00\xB4\x2D\x1E\x00\xB5\xF5"
-     "\x1E\x00"
-     "\xB6\xFD\x1E\x00\xB7\xF5\x1E\x00\xB8\xCD\x1E\x00\xB9\xD5\x1E\x00\xBA\xDD"
-     "\x1E\x00"
-     "\xBB\xD5\x1E\x00\xBC\xCD\x1E\x00\xBD\xB9\x1E\x00\xBE\x59\x1F\x00\xBF\x59"
-     "\x1F\x00"
-     "\xB3\x35\x1D\x00\xA8\x73\x00\x80\xAC\x73\x00\x80\xB0\x73\x00\x80\xB4\x73"
-     "\x00\x80"
-     "\xB6\x95\x1E\x00\xB5\x25\x1D\x00\xB8\x73\x00\x80\xBB\xC1\x1E\x00\xBA\xE1"
-     "\x1E\x00"
-     "\xBC\x73\x00\x80\xC0\x73\x00\x80\xBF\x11\x1E\x00\xBE\xC9\x1E\x00\xBD\xCD"
-     "\x1E\x00"
-     "\xBC\xD9\x1E\x00\xC4\x73\x00\x80\xA3\x71\x1D\x00\xC8\x73\x00\x80\xCC\x73"
-     "\x00\x80"
-     "\xA6\xD1\x1E\x00\xD0\x73\x00\x80\xD4\x73\x00\x80\xA5\x61\x1D\x00\xAA\xA5"
-     "\x1E\x00"
-     "\xAB\x85\x1E\x00\xD8\x73\x00\x80\xDC\x73\x00\x80\xAE\x8D\x1E\x00\xAF\x55"
-     "\x1E\x00"
-     "\xAC\x9D\x1E\x00\xAD\x89\x1E\x00\xA8\x85\x1E\x00\xA9\x95\x1E\x00\xAA\xA9"
-     "\x1E\x00"
-     "\xAB\xA5\x1E\x00\xAC\xD1\x1E\x00\xAD\xC1\x1E\x00\xAE\xC1\x1E\x00\xAF\xC1"
-     "\x1E\x00"
-     "\x80\xE9\x01\x00\x81\xE9\x01\x00\x82\xFD\x01\x00\xE0\x73\x00\x80\xE4\x73"
-     "\x00\x80"
-     "\xE8\x73\x00\x80\xEC\x73\x00\x80\xF0\x73\x00\x80\xB8\xC5\x01\x00\xB9\xCD"
-     "\x01\x00"
-     "\xBA\xC5\x01\x00\xBB\xDD\x01\x00\xBC\xCD\x01\x00\xBD\xFD\x01\x00\xBE\xF5"
-     "\x01\x00"
-     "\xBF\x9D\x01\x00\xB0\x99\x1E\x00\xB1\x99\x1E\x00\xB2\x61\x01\x00\xB3\x61"
-     "\x01\x00"
-     "\xB4\x61\x01\x00\xB5\x61\x01\x00\xB6\x61\x01\x00\xB7\x61\x01\x00\xB3\x21"
-     "\x1E\x00"
-     "\xF4\x73\x00\x80\xF8\x73\x00\x80\xFC\x73\x00\x80\x00\x74\x00\x80\xB6\x41"
-     "\x1E\x00"
-     "\xB5\x59\x1E\x00\x84\x90\x00\x00\xBB\x5D\x1E\x00\xBA\x5D\x1E\x00\x04\x74"
-     "\x00\x80"
-     "\x08\x74\x00\x80\xBF\xA5\x01\x00\xBE\xA5\x01\x00\xBD\xB5\x01\x00\xBC\xB5"
-     "\x01\x00"
-     "\x0C\x74\x00\x80\xA3\x65\x1E\x00\x10\x74\x00\x80\x14\x74\x00\x80\xA6\x05"
-     "\x1E\x00"
-     "\x18\x74\x00\x80\x1C\x74\x00\x80\xA5\x1D\x1E\x00\xAA\x19\x1E\x00\xAB\x19"
-     "\x1E\x00"
-     "\x20\x74\x00\x80\x24\x74\x00\x80\xAE\xE1\x01\x00\xAF\xE1\x01\x00\xAC\xF1"
-     "\x01\x00"
-     "\xAD\xF1\x01\x00\xA8\xE9\x02\x00\xA9\xE9\x02\x00\xAA\x25\x03\x00\xAB\x3D"
-     "\x03\x00"
-     "\xAC\x25\x03\x00\xAD\x2D\x03\x00\xAE\x25\x03\x00\xAF\x99\x03\x00\x28\x74"
-     "\x00\x80"
-     "\x2C\x74\x00\x80\x30\x74\x00\x80\x34\x74\x00\x80\x80\x0D\x00\x00\x81\xB1"
-     "\x00\x00"
-     "\x82\xB1\x00\x00\x38\x74\x00\x80\xB8\x95\x03\x00\xB9\x9D\x03\x00\xBA\xA1"
-     "\x03\x00"
-     "\xBB\xA1\x03\x00\xBC\x71\x00\x00\xBD\x71\x00\x00\xBE\x71\x00\x00\xBF\x71"
-     "\x00\x00"
-     "\xB0\xE9\x03\x00\xB1\xE9\x03\x00\xB2\xF9\x03\x00\xB3\xC5\x03\x00\xB4\xDD"
-     "\x03\x00"
-     "\xB5\xB5\x03\x00\xB6\xB1\x03\x00\xB7\xB1\x03\x00\xB3\xAD\x03\x00\x3C\x74"
-     "\x00\x80"
-     "\x40\x74\x00\x80\x48\x74\x00\x80\x4C\x74\x00\x80\xB6\xAD\x03\x00\xB5\xAD"
-     "\x03\x00"
-     "\x50\x74\x00\x80\xBB\x2D\x03\x00\xBA\x2D\x03\x00\x86\x08\x04\x00\x87\xDC"
-     "\x03\x00"
-     "\xBF\x71\x03\x00\xBE\x19\x03\x00\xBD\x11\x03\x00\xBC\x2D\x03\x00\x54\x74"
-     "\x00\x80"
-     "\xA3\xE9\x03\x00\x58\x74\x00\x80\x5C\x74\x00\x80\xA6\xE9\x03\x00\x60\x74"
-     "\x00\x80"
-     "\x64\x74\x00\x80\xA5\xE9\x03\x00\xAA\x69\x03\x00\xAB\x69\x03\x00\x68\x74"
-     "\x00\x80"
-     "\x6C\x74\x00\x80\xAE\x5D\x03\x00\xAF\x35\x03\x00\xAC\x69\x03\x00\xAD\x55"
-     "\x03\x00"
-     "\x70\x74\x00\x80\xA6\xAD\x1F\x00\xA5\xAD\x1F\x00\x74\x74\x00\x80\xA3\x4D"
-     "\x1F\x00"
-     "\x78\x74\x00\x80\x7C\x74\x00\x80\x80\x74\x00\x80\xAF\x29\x1E\x00\xAE\x21"
-     "\x1E\x00"
-     "\xAD\x29\x1E\x00\xAC\x91\x1F\x00\xAB\x89\x1F\x00\xAA\x81\x1F\x00\x84\x74"
-     "\x00\x80"
-     "\x88\x74\x00\x80\x8C\x74\x00\x80\x90\x74\x00\x80\xBE\x44\x05\x00\x84\x80"
-     "\x04\x00"
-     "\x94\x74\x00\x80\xE3\xF8\x1E\x00\x98\x74\x00\x80\xE1\xF4\x1E\x00\x9C\x74"
-     "\x00\x80"
-     "\xEF\x50\x1E\x00\xA0\x74\x00\x80\xA4\x74\x00\x80\x80\x0D\x00\x00\x81\xE5"
-     "\x00\x00"
-     "\x82\xED\x00\x00\xA8\x74\x00\x80\xE1\x00\x1E\x00\xE1\x30\x1C\x00\xE3\x0C"
-     "\x1F\x00"
-     "\xE3\xBC\x1F\x00\xAC\x74\x00\x80\xB0\x74\x00\x80\x86\xB8\x04\x00\x87\x5C"
-     "\x05\x00"
-     "\xBE\x20\x07\x00\xB8\x74\x00\x80\xBC\x74\x00\x80\xC0\x74\x00\x80\xC4\x74"
-     "\x00\x80"
-     "\xC8\x74\x00\x80\xEF\xCC\x1E\x00\xEF\x3C\x1F\x00\xB3\x5D\x01\x00\xCC\x74"
-     "\x00\x80"
-     "\xD0\x74\x00\x80\xD4\x74\x00\x80\xD8\x74\x00\x80\xB6\x19\x01\x00\xB5\x11"
-     "\x01\x00"
-     "\xDC\x74\x00\x80\xBB\x0D\x01\x00\xBA\x0D\x01\x00\xE0\x74\x00\x80\xE4\x74"
-     "\x00\x80"
-     "\xBF\xE9\x00\x00\xBE\xE1\x00\x00\xBD\xF5\x00\x00\xBC\xFD\x00\x00\xA8\x8D"
-     "\x06\x00"
-     "\xA9\x95\x06\x00\xAA\x95\x06\x00\xAB\xCD\x06\x00\xAC\xD9\x06\x00\xAD\xD9"
-     "\x06\x00"
-     "\xAE\xCD\x06\x00\xAF\xC5\x06\x00\xB4\x74\x00\x80\xE8\x74\x00\x80\xEC\x74"
-     "\x00\x80"
-     "\xF0\x74\x00\x80\xF4\x74\x00\x80\xF8\x74\x00\x80\xFC\x74\x00\x80\x00\x75"
-     "\x00\x80"
-     "\xB8\x7D\x07\x00\xB9\x05\x07\x00\xBA\x0D\x07\x00\xBB\x1D\x07\x00\xBC\x09"
-     "\x07\x00"
-     "\xBD\x35\x07\x00\xBE\x31\x07\x00\xBF\x31\x07\x00\xB0\x89\x06\x00\xB1\x89"
-     "\x06\x00"
-     "\xB2\x69\x07\x00\xB3\x61\x07\x00\xB4\x65\x07\x00\xB5\x6D\x07\x00\xB6\x65"
-     "\x07\x00"
-     "\xB7\x45\x07\x00\xA3\x1D\x06\x00\x04\x75\x00\x80\x08\x75\x00\x80\x0C\x75"
-     "\x00\x80"
-     "\x10\x75\x00\x80\xA6\x59\x06\x00\xA5\x51\x06\x00\x14\x75\x00\x80\xAB\x4D"
-     "\x06\x00"
-     "\xAA\x4D\x06\x00\x18\x75\x00\x80\x1C\x75\x00\x80\xAF\xA9\x07\x00\xAE\xA1"
-     "\x07\x00"
-     "\xAD\xB5\x07\x00\xAC\xBD\x07\x00\x80\x69\x00\x00\x81\x69\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x20\x75\x00\x80\x24\x75\x00\x80\x28\x75\x00\x80\x87\xA8\x03\x00\x86\x9C"
-     "\x00\x00"
-     "\x2C\x75\x00\x80\x30\x75\x00\x80\x34\x75\x00\x80\x38\x75\x00\x80\x3C\x75"
-     "\x00\x80"
-     "\x40\x75\x00\x80\x44\x75\x00\x80\x48\x75\x00\x80\xB3\xAD\x06\x00\x4C\x75"
-     "\x00\x80"
-     "\x50\x75\x00\x80\x54\x75\x00\x80\x58\x75\x00\x80\xB6\xD1\x06\x00\xB5\xAD"
-     "\x06\x00"
-     "\x5C\x75\x00\x80\xBB\xFD\x06\x00\xBA\xF5\x06\x00\x60\x75\x00\x80\x64\x75"
-     "\x00\x80"
-     "\xBF\x19\x01\x00\xBE\x11\x01\x00\xBD\xD5\x06\x00\xBC\xD5\x06\x00\x68\x75"
-     "\x00\x80"
-     "\x6C\x75\x00\x80\x70\x75\x00\x80\x74\x75\x00\x80\x78\x75\x00\x80\xE3\xD8"
-     "\x06\x00"
-     "\x7C\x75\x00\x80\xE1\xEC\x05\x00\x80\x75\x00\x80\xEF\xD4\x04\x00\x84\x75"
-     "\x00\x80"
-     "\x88\x75\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xD9\x00\x00\x8C\x75"
-     "\x00\x80"
-     "\xA8\x25\x06\x00\xA9\x2D\x06\x00\xAA\x25\x06\x00\xAB\x3D\x06\x00\xAC\x25"
-     "\x06\x00"
-     "\xAD\x51\x06\x00\xAE\x51\x06\x00\xAF\x51\x06\x00\x90\x75\x00\x80\x84\x3C"
-     "\x01\x00"
-     "\xBE\x24\x00\x00\x98\x75\x00\x80\x86\x10\x0F\x00\x87\xC4\x00\x00\x9C\x75"
-     "\x00\x80"
-     "\xA0\x75\x00\x80\xB8\xF5\x01\x00\xB9\xFD\x01\x00\xBA\xF5\x01\x00\xBB\x49"
-     "\x01\x00"
-     "\xBC\x59\x01\x00\xBD\x59\x01\x00\xBE\x49\x01\x00\xBF\x41\x01\x00\xB0\x31"
-     "\x06\x00"
-     "\xB1\x39\x06\x00\xB2\x01\x06\x00\xB3\x01\x06\x00\xB4\xD5\x01\x00\xB5\xDD"
-     "\x01\x00"
-     "\xB6\xD5\x01\x00\xB7\xCD\x01\x00\xA4\x75\x00\x80\xA8\x75\x00\x80\xAC\x75"
-     "\x00\x80"
-     "\xA3\xA1\x05\x00\xB0\x75\x00\x80\xA5\xA1\x05\x00\xA6\xDD\x05\x00\xB4\x75"
-     "\x00\x80"
-     "\xB8\x75\x00\x80\xBC\x75\x00\x80\xAA\xF9\x05\x00\xAB\xF1\x05\x00\xAC\xD9"
-     "\x05\x00"
-     "\xAD\xD9\x05\x00\xAE\x1D\x02\x00\xAF\x15\x02\x00\xB6\xB1\x02\x00\xC0\x75"
-     "\x00\x80"
-     "\xC4\x75\x00\x80\xB5\xA1\x02\x00\xC8\x75\x00\x80\xB3\xB1\x02\x00\xCC\x75"
-     "\x00\x80"
-     "\xD0\x75\x00\x80\xBE\x71\x03\x00\xBF\x79\x03\x00\xBC\x61\x03\x00\xBD\x61"
-     "\x03\x00"
-     "\xBA\x8D\x02\x00\xBB\x71\x03\x00\xD4\x75\x00\x80\xD8\x75\x00\x80\xB0\x7D"
-     "\x02\x00"
-     "\xB1\x81\x03\x00\xB2\x81\x03\x00\xB3\x81\x03\x00\xB4\x81\x03\x00\xB5\x89"
-     "\x03\x00"
-     "\xB6\xB1\x03\x00\xB7\xB1\x03\x00\xB8\xE1\x03\x00\xB9\xE1\x03\x00\xBA\xE1"
-     "\x03\x00"
-     "\xBB\xE1\x03\x00\xBC\xE1\x03\x00\xBD\xED\x03\x00\xBE\xD9\x03\x00\xBF\xD5"
-     "\x03\x00"
-     "\x86\xE0\x0D\x00\x87\xE8\x0C\x00\xDC\x75\x00\x80\xE0\x75\x00\x80\xE4\x75"
-     "\x00\x80"
-     "\xE8\x75\x00\x80\x44\x74\x00\x80\xEC\x75\x00\x80\xA8\x35\x02\x00\xA9\x05"
-     "\x02\x00"
-     "\xAA\x0D\x02\x00\xAB\x05\x02\x00\xAC\x1D\x02\x00\xAD\x05\x02\x00\xAE\x0D"
-     "\x02\x00"
-     "\xAF\x05\x02\x00\xA3\xE5\x03\x00\xF0\x75\x00\x80\xF4\x75\x00\x80\xF8\x75"
-     "\x00\x80"
-     "\xFC\x75\x00\x80\xA6\xE5\x03\x00\xA5\xF5\x03\x00\x00\x76\x00\x80\xAB\x25"
-     "\x02\x00"
-     "\xAA\xD9\x03\x00\x84\x28\x02\x00\xBE\x20\x0D\x00\xAF\x2D\x02\x00\xAE\x25"
-     "\x02\x00"
-     "\xAD\x35\x02\x00\xAC\x35\x02\x00\x08\x76\x00\x80\x0C\x76\x00\x80\x10\x76"
-     "\x00\x80"
-     "\x14\x76\x00\x80\xBE\xDC\x0C\x00\xE1\x90\x01\x00\x18\x76\x00\x80\xE3\xE8"
-     "\x00\x00"
-     "\x1C\x76\x00\x80\x20\x76\x00\x80\x24\x76\x00\x80\xEF\x3C\x00\x00\x80\x7D"
-     "\x00\x00"
-     "\x81\x65\x00\x00\x82\x65\x00\x00\x28\x76\x00\x80\x2C\x76\x00\x80\x30\x76"
-     "\x00\x80"
-     "\x34\x76\x00\x80\x38\x76\x00\x80\x3C\x76\x00\x80\x40\x76\x00\x80\x44\x76"
-     "\x00\x80"
-     "\xEF\xC4\x06\x00\x48\x76\x00\x80\xE1\x24\x06\x00\xBF\xA0\x3F\x00\xE3\xF0"
-     "\x06\x00"
-     "\x4C\x76\x00\x80\x84\xBC\x0D\x00\x50\x76\x00\x80\x54\x76\x00\x80\x58\x76"
-     "\x00\x80"
-     "\x5C\x76\x00\x80\x60\x76\x00\x80\xB3\x71\x02\x00\x64\x76\x00\x80\xB5\xD1"
-     "\x01\x00"
-     "\xB6\xD1\x01\x00\x68\x76\x00\x80\x6C\x76\x00\x80\x70\x76\x00\x80\xBA\xC9"
-     "\x01\x00"
-     "\xBB\xC9\x01\x00\xBC\xB5\x01\x00\xBD\xBD\x01\x00\xBE\xA9\x01\x00\xBF\xA5"
-     "\x01\x00"
-     "\xA8\xAD\x0D\x00\xA9\xBD\x0D\x00\xAA\xB5\x0D\x00\xAB\x0D\x0E\x00\xAC\x15"
-     "\x0E\x00"
-     "\xAD\x1D\x0E\x00\xAE\x09\x0E\x00\xAF\x01\x0E\x00\x04\x76\x00\x80\x82\xC1"
-     "\x0F\x00"
-     "\x81\xC9\x0F\x00\x80\xF1\x0F\x00\x74\x76\x00\x80\x78\x76\x00\x80\x86\x98"
-     "\x00\x00"
-     "\x87\x60\x03\x00\xB8\x3D\x0E\x00\xB9\x8D\x0E\x00\xBA\x85\x0E\x00\xBB\x9D"
-     "\x0E\x00"
-     "\xBC\x85\x0E\x00\xBD\x89\x0E\x00\xBE\x5D\x0F\x00\xBF\x55\x0F\x00\xB0\x41"
-     "\x0E\x00"
-     "\xB1\x41\x0E\x00\xB2\x2D\x0E\x00\xB3\x25\x0E\x00\xB4\x21\x0E\x00\xB5\x2D"
-     "\x0E\x00"
-     "\xB6\x25\x0E\x00\xB7\x05\x0E\x00\xA3\x3D\x0D\x00\x7C\x76\x00\x80\x80\x76"
-     "\x00\x80"
-     "\x84\x76\x00\x80\x88\x76\x00\x80\xA6\x9D\x0E\x00\xA5\x9D\x0E\x00\x8C\x76"
-     "\x00\x80"
-     "\xAB\x85\x0E\x00\xAA\x85\x0E\x00\x90\x76\x00\x80\x94\x76\x00\x80\xAF\xE9"
-     "\x0E\x00"
-     "\xAE\xE5\x0E\x00\xAD\xF1\x0E\x00\xAC\xF9\x0E\x00\x98\x76\x00\x80\xB3\x8D"
-     "\x0F\x00"
-     "\x9C\x76\x00\x80\xA0\x76\x00\x80\xB6\xB1\x0F\x00\xA4\x76\x00\x80\xA8\x76"
-     "\x00\x80"
-     "\xB5\x85\x0F\x00\xBA\x9D\x0F\x00\xBB\x65\x0F\x00\xAC\x76\x00\x80\xB0\x76"
-     "\x00\x80"
-     "\xBE\x7D\x0F\x00\xBF\x65\x0F\x00\xBC\x7D\x0F\x00\xBD\x75\x0F\x00\xA3\x49"
-     "\x0E\x00"
-     "\xB4\x76\x00\x80\xB8\x76\x00\x80\xBC\x76\x00\x80\xC0\x76\x00\x80\xA6\x75"
-     "\x0E\x00"
-     "\xA5\x41\x0E\x00\xC4\x76\x00\x80\xAB\xA1\x0E\x00\xAA\x59\x0E\x00\xC8\x76"
-     "\x00\x80"
-     "\xCC\x76\x00\x80\xAF\xA1\x0E\x00\xAE\xB9\x0E\x00\xAD\xB1\x0E\x00\xAC\xB9"
-     "\x0E\x00"
-     "\xD0\x76\x00\x80\xD4\x76\x00\x80\xD8\x76\x00\x80\xDC\x76\x00\x80\xEF\x84"
-     "\x0C\x00"
-     "\x82\x31\x00\x00\x81\x15\x00\x00\x80\x15\x00\x00\xE3\x58\x0C\x00\xE0\x76"
-     "\x00\x80"
-     "\xE1\xC0\x0E\x00\x84\x20\x01\x00\xBE\x38\x01\x00\xE8\x76\x00\x80\x86\xD8"
-     "\x00\x00"
-     "\x87\x5C\x00\x00\xA8\xC1\x0E\x00\xA9\xC1\x0E\x00\xAA\xC1\x0E\x00\xAB\xC1"
-     "\x0E\x00"
-     "\xAC\xC9\x0E\x00\xAD\xF5\x0E\x00\xAE\xFD\x0E\x00\xAF\xF5\x0E\x00\x84\xA0"
-     "\x00\x00"
-     "\xEC\x76\x00\x80\xF0\x76\x00\x80\xF4\x76\x00\x80\xF8\x76\x00\x80\xFC\x76"
-     "\x00\x80"
-     "\x00\x77\x00\x80\x04\x77\x00\x80\xB8\x61\x0E\x00\xB9\x75\x0E\x00\xBA\x71"
-     "\x0E\x00"
-     "\xBB\x71\x0E\x00\xBC\x11\x0E\x00\xBD\x11\x0E\x00\xBE\x11\x0E\x00\xBF\x0D"
-     "\x0E\x00"
-     "\xB0\x9D\x0E\x00\xB1\x65\x0E\x00\xB2\x6D\x0E\x00\xB3\x65\x0E\x00\xB4\x7D"
-     "\x0E\x00"
-     "\xB5\x65\x0E\x00\xB6\x6D\x0E\x00\xB7\x65\x0E\x00\xA8\x35\x02\x00\xA9\x81"
-     "\x02\x00"
-     "\xAA\x81\x02\x00\xAB\x81\x02\x00\xAC\x81\x02\x00\xAD\x89\x02\x00\xAE\xB1"
-     "\x02\x00"
-     "\xAF\xB1\x02\x00\xBF\x28\xDD\x01\x08\x77\x00\x80\x0C\x77\x00\x80\x84\xE0"
-     "\x05\x00"
-     "\x10\x77\x00\x80\x14\x77\x00\x80\x18\x77\x00\x80\x1C\x77\x00\x80\xB8\x49"
-     "\x03\x00"
-     "\xB9\x49\x03\x00\xBA\x59\x03\x00\xBB\x51\x03\x00\xBC\x79\x03\x00\xBD\x79"
-     "\x03\x00"
-     "\xBE\x19\x03\x00\xBF\x19\x03\x00\xB0\xA1\x02\x00\xB1\xAD\x02\x00\xB2\xA5"
-     "\x02\x00"
-     "\xB3\xB9\x02\x00\xB4\xA9\x02\x00\xB5\x9D\x02\x00\xB6\x95\x02\x00\xB7\x79"
-     "\x03\x00"
-     "\xB3\x09\x02\x00\x20\x77\x00\x80\x24\x77\x00\x80\x28\x77\x00\x80\x2C\x77"
-     "\x00\x80"
-     "\xB6\x09\x02\x00\xB5\x09\x02\x00\x30\x77\x00\x80\xBB\x69\x02\x00\xBA\x65"
-     "\x02\x00"
-     "\x34\x77\x00\x80\x38\x77\x00\x80\xBF\x59\x02\x00\xBE\x6D\x02\x00\xBD\x6D"
-     "\x02\x00"
-     "\xBC\x71\x02\x00\x3C\x77\x00\x80\xA3\x4D\x02\x00\x40\x77\x00\x80\x44\x77"
-     "\x00\x80"
-     "\xA6\x4D\x02\x00\x48\x77\x00\x80\x4C\x77\x00\x80\xA5\x4D\x02\x00\xAA\x21"
-     "\x02\x00"
-     "\xAB\x2D\x02\x00\x50\x77\x00\x80\x94\x75\x00\x80\xAE\x29\x02\x00\xAF\x1D"
-     "\x02\x00"
-     "\xAC\x35\x02\x00\xAD\x29\x02\x00\xA8\x75\x02\x00\xA9\xA1\x02\x00\xAA\xA1"
-     "\x02\x00"
-     "\xAB\xA1\x02\x00\xAC\xA1\x02\x00\xAD\xA1\x02\x00\xAE\xA1\x02\x00\xAF\xA1"
-     "\x02\x00"
-     "\x80\xF5\x01\x00\x81\xFD\x01\x00\x82\xC1\x01\x00\xBE\xA0\x05\x00\x58\x77"
-     "\x00\x80"
-     "\x5C\x77\x00\x80\x87\x74\x05\x00\x86\x1C\x05\x00\xB8\xCD\x01\x00\xB9\xD5"
-     "\x01\x00"
-     "\xBA\xD5\x01\x00\xBB\xE5\x01\x00\xBC\xFD\x01\x00\xBD\x95\x01\x00\xBE\x91"
-     "\x01\x00"
-     "\xBF\x91\x01\x00\xB0\x45\x01\x00\xB1\x4D\x01\x00\xB2\x59\x01\x00\xB3\x59"
-     "\x01\x00"
-     "\xB4\x49\x01\x00\xB5\x49\x01\x00\xB6\xFD\x01\x00\xB7\xF5\x01\x00\x60\x77"
-     "\x00\x80"
-     "\x64\x77\x00\x80\xBE\x84\x04\x00\xEF\x04\x00\x00\x68\x77\x00\x80\x6C\x77"
-     "\x00\x80"
-     "\x70\x77\x00\x80\xEF\x34\x0E\x00\x74\x77\x00\x80\xE1\x7C\x0F\x00\x78\x77"
-     "\x00\x80"
-     "\xE3\x08\x0E\x00\xEC\x00\x00\x00\xE1\x94\x01\x00\x7C\x77\x00\x80\xE3\xC0"
-     "\x00\x00"
-     "\x80\x77\x00\x80\x84\x77\x00\x80\x88\x77\x00\x80\xB3\x65\x02\x00\x8C\x77"
-     "\x00\x80"
-     "\xB5\x65\x02\x00\xB6\x6D\x02\x00\x90\x77\x00\x80\x94\x77\x00\x80\x98\x77"
-     "\x00\x80"
-     "\xBA\xCD\x01\x00\xBB\xC5\x01\x00\xBC\xC1\x01\x00\xBD\xC1\x01\x00\xBE\xC1"
-     "\x01\x00"
-     "\xBF\xC1\x01\x00\xA3\xA9\x05\x00\x54\x77\x00\x80\x9C\x77\x00\x80\xA0\x77"
-     "\x00\x80"
-     "\xA4\x77\x00\x80\xA6\xA1\x05\x00\xA5\xA9\x05\x00\xA8\x77\x00\x80\xAB\x09"
-     "\x06\x00"
-     "\xAA\x01\x06\x00\xAC\x77\x00\x80\xB0\x77\x00\x80\xAF\x0D\x06\x00\xAE\x0D"
-     "\x06\x00"
-     "\xAD\x0D\x06\x00\xAC\x0D\x06\x00\xB4\x77\x00\x80\x82\x1D\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\xB8\x77\x00\x80\xBC\x77\x00\x80\xC0\x77\x00\x80\xC4\x77"
-     "\x00\x80"
-     "\x86\xC0\x00\x00\x87\x34\x03\x00\xC8\x77\x00\x80\xCC\x77\x00\x80\xD0\x77"
-     "\x00\x80"
-     "\xD4\x77\x00\x80\xD8\x77\x00\x80\xDC\x77\x00\x80\xA8\xB1\x07\x00\xA9\xB1"
-     "\x07\x00"
-     "\xAA\x09\x07\x00\xAB\x09\x07\x00\xAC\x01\x07\x00\xAD\x01\x07\x00\xAE\x01"
-     "\x07\x00"
-     "\xAF\x3D\x07\x00\x84\xA8\x03\x00\xE0\x77\x00\x80\xE4\x77\x00\x80\xE8\x77"
-     "\x00\x80"
-     "\xEC\x77\x00\x80\xF0\x77\x00\x80\xF4\x77\x00\x80\xF8\x77\x00\x80\xB8\x25"
-     "\x07\x00"
-     "\xB9\xCD\x00\x00\xBA\xC5\x00\x00\xBB\xDD\x00\x00\xBC\xCD\x00\x00\xBD\xFD"
-     "\x00\x00"
-     "\xBE\xF5\x00\x00\xBF\x9D\x00\x00\xB0\x49\x07\x00\xB1\x5D\x07\x00\xB2\x29"
-     "\x07\x00"
-     "\xB3\x29\x07\x00\xB4\x39\x07\x00\xB5\x25\x07\x00\xB6\x25\x07\x00\xB7\x1D"
-     "\x07\x00"
-     "\xB3\x81\x06\x00\xFC\x77\x00\x80\x00\x78\x00\x80\x04\x78\x00\x80\x08\x78"
-     "\x00\x80"
-     "\xB6\x81\x06\x00\xB5\x89\x06\x00\x0C\x78\x00\x80\xBB\x99\x06\x00\xBA\x8D"
-     "\x06\x00"
-     "\x10\x78\x00\x80\x14\x78\x00\x80\xBF\xE1\x06\x00\xBE\xFD\x06\x00\xBD\xFD"
-     "\x06\x00"
-     "\xBC\xFD\x06\x00\x18\x78\x00\x80\x1C\x78\x00\x80\x20\x78\x00\x80\x24\x78"
-     "\x00\x80"
-     "\x80\x31\x00\x00\x81\x3D\x00\x00\x82\x29\x00\x00\xEF\x88\x04\x00\x28\x78"
-     "\x00\x80"
-     "\xE1\x04\x06\x00\x2C\x78\x00\x80\xE3\xDC\x06\x00\x30\x78\x00\x80\xBE\x74"
-     "\x01\x00"
-     "\x34\x78\x00\x80\x38\x78\x00\x80\xA3\x01\x06\x00\x3C\x78\x00\x80\x86\x28"
-     "\x00\x00"
-     "\x87\x44\x01\x00\x40\x78\x00\x80\xA6\x01\x06\x00\xA5\x09\x06\x00\x44\x78"
-     "\x00\x80"
-     "\xAB\x19\x06\x00\xAA\x0D\x06\x00\x48\x78\x00\x80\x4C\x78\x00\x80\xAF\x61"
-     "\x06\x00"
-     "\xAE\x7D\x06\x00\xAD\x7D\x06\x00\xAC\x7D\x06\x00\x50\x78\x00\x80\xB3\xB5"
-     "\x01\x00"
-     "\x54\x78\x00\x80\x58\x78\x00\x80\xB6\xAD\x01\x00\x5C\x78\x00\x80\x60\x78"
-     "\x00\x80"
-     "\xB5\xA5\x01\x00\xBA\x4D\x01\x00\xBB\x55\x01\x00\x64\x78\x00\x80\x68\x78"
-     "\x00\x80"
-     "\xBE\x4D\x01\x00\xBF\xF1\x00\x00\xBC\x4D\x01\x00\xBD\x45\x01\x00\xA8\x69"
-     "\x02\x00"
-     "\xA9\x69\x02\x00\xAA\x79\x02\x00\xAB\x79\x02\x00\xAC\x69\x02\x00\xAD\x69"
-     "\x02\x00"
-     "\xAE\xBD\x02\x00\xAF\xB5\x02\x00\x84\x00\x3C\x00\x6C\x78\x00\x80\x70\x78"
-     "\x00\x80"
-     "\x74\x78\x00\x80\x78\x78\x00\x80\x7C\x78\x00\x80\x80\x78\x00\x80\x84\x78"
-     "\x00\x80"
-     "\xB8\x51\x03\x00\xB9\x59\x03\x00\xBA\x61\x03\x00\xBB\x61\x03\x00\xBC\x11"
-     "\x03\x00"
-     "\xBD\x11\x03\x00\xBE\x11\x03\x00\xBF\x11\x03\x00\xB0\xCD\x02\x00\xB1\xD5"
-     "\x02\x00"
-     "\xB2\xDD\x02\x00\xB3\xD5\x02\x00\xB4\xCD\x02\x00\xB5\x71\x03\x00\xB6\x71"
-     "\x03\x00"
-     "\xB7\x71\x03\x00\x88\x78\x00\x80\x8C\x78\x00\x80\x90\x78\x00\x80\xA3\xFD"
-     "\x02\x00"
-     "\x94\x78\x00\x80\xA5\xED\x02\x00\xA6\xE5\x02\x00\x98\x78\x00\x80\x9C\x78"
-     "\x00\x80"
-     "\xA0\x78\x00\x80\xAA\x05\x02\x00\xAB\x1D\x02\x00\xAC\x05\x02\x00\xAD\x0D"
-     "\x02\x00"
-     "\xAE\x05\x02\x00\xAF\xB9\x03\x00\x82\x65\x00\x00\xEF\xE4\x02\x00\x80\x55"
-     "\x00\x00"
-     "\x81\x6D\x00\x00\x87\x14\x03\x00\x86\x74\x3D\x00\xBE\x54\x3C\x00\xA8\x78"
-     "\x00\x80"
-     "\xE4\x76\x00\x80\xAC\x78\x00\x80\xB0\x78\x00\x80\xB4\x78\x00\x80\xB8\x78"
-     "\x00\x80"
-     "\xE3\xC4\x02\x00\xBC\x78\x00\x80\xE1\x6C\x01\x00\xA8\x7D\x02\x00\xA9\x81"
-     "\x02\x00"
-     "\xAA\x99\x02\x00\xAB\x91\x02\x00\xAC\xB9\x02\x00\xAD\xB9\x02\x00\xAE\xD9"
-     "\x02\x00"
-     "\xAF\xD5\x02\x00\xC0\x78\x00\x80\xC4\x78\x00\x80\xC8\x78\x00\x80\xCC\x78"
-     "\x00\x80"
-     "\xD0\x78\x00\x80\xD4\x78\x00\x80\xD8\x78\x00\x80\xDC\x78\x00\x80\xB8\x71"
-     "\x01\x00"
-     "\xB9\x71\x01\x00\xBA\x71\x01\x00\xBB\x71\x01\x00\xBC\xD1\x01\x00\xBD\xD1"
-     "\x01\x00"
-     "\xBE\xD1\x01\x00\xBF\xCD\x01\x00\xB0\xAD\x02\x00\xB1\xB1\x02\x00\xB2\xB1"
-     "\x02\x00"
-     "\xB3\x85\x02\x00\xB4\x9D\x02\x00\xB5\x51\x01\x00\xB6\x51\x01\x00\xB7\x51"
-     "\x01\x00"
-     "\xE1\xA4\x06\x00\xE0\x78\x00\x80\xE3\xB4\x07\x00\xBE\x80\x3D\x00\x84\x18"
-     "\x3D\x00"
-     "\xE4\x78\x00\x80\xE8\x78\x00\x80\xEC\x78\x00\x80\xF0\x78\x00\x80\xF4\x78"
-     "\x00\x80"
-     "\xF8\x78\x00\x80\xFC\x78\x00\x80\x00\x79\x00\x80\x04\x79\x00\x80\xEF\xC8"
-     "\x06\x00"
-     "\x08\x79\x00\x80\x81\x09\x00\x00\x80\x79\x00\x00\x0C\x79\x00\x80\x82\x19"
-     "\x00\x00"
-     "\xE3\x60\x01\x00\x10\x79\x00\x80\xE1\xA0\x01\x00\x18\x79\x00\x80\xEF\x6C"
-     "\x00\x00"
-     "\x1C\x79\x00\x80\x86\xC8\x3C\x00\x87\x0C\x3D\x00\x20\x79\x00\x80\x24\x79"
-     "\x00\x80"
-     "\x28\x79\x00\x80\x2C\x79\x00\x80\xA3\x91\x3D\x00\xA4\x78\x00\x80\x14\x79"
-     "\x00\x80"
-     "\x30\x79\x00\x80\x34\x79\x00\x80\xA6\xB9\x3D\x00\xA5\xB1\x3D\x00\x38\x79"
-     "\x00\x80"
-     "\xAB\xF5\x3D\x00\xAA\xE9\x3D\x00\x3C\x79\x00\x80\x40\x79\x00\x80\xAF\x15"
-     "\x3E\x00"
-     "\xAE\xD9\x3D\x00\xAD\xD1\x3D\x00\xAC\xE5\x3D\x00\x44\x79\x00\x80\xB3\xAD"
-     "\x3E\x00"
-     "\x48\x79\x00\x80\x4C\x79\x00\x80\xB6\xA9\x3E\x00\x50\x79\x00\x80\x54\x79"
-     "\x00\x80"
-     "\xB5\xAD\x3E\x00\xBA\x49\x3F\x00\xBB\x49\x3F\x00\x58\x79\x00\x80\x5C\x79"
-     "\x00\x80"
-     "\xBE\x49\x3F\x00\xBF\x49\x3F\x00\xBC\x49\x3F\x00\xBD\x49\x3F\x00\xA3\x69"
-     "\x3E\x00"
-     "\x60\x79\x00\x80\x64\x79\x00\x80\x68\x79\x00\x80\x6C\x79\x00\x80\xA6\x6D"
-     "\x3E\x00"
-     "\xA5\x69\x3E\x00\x70\x79\x00\x80\xAB\x8D\x3F\x00\xAA\x8D\x3F\x00\x74\x79"
-     "\x00\x80"
-     "\x78\x79\x00\x80\xAF\x8D\x3F\x00\xAE\x8D\x3F\x00\xAD\x8D\x3F\x00\xAC\x8D"
-     "\x3F\x00"
-     "\x7C\x79\x00\x80\x80\x79\x00\x80\xBA\x05\x3D\x00\xBB\x19\x3D\x00\xBC\x09"
-     "\x3D\x00"
-     "\xBD\x3D\x3D\x00\xBE\x35\x3D\x00\xBF\xF9\x3E\x00\x80\x4D\x00\x00\x81\x25"
-     "\x00\x00"
-     "\x82\x2D\x00\x00\xB3\x7D\x3D\x00\xBE\x38\x00\x00\xB5\x5D\x3D\x00\xB6\x55"
-     "\x3D\x00"
-     "\x84\x30\x00\x00\xA8\xDD\x3F\x00\xA9\xE1\x3F\x00\xAA\xE1\x3F\x00\xAB\xE1"
-     "\x3F\x00"
-     "\xAC\xE1\x3F\x00\xAD\xED\x3F\x00\xAE\x39\x3E\x00\xAF\x39\x3E\x00\x84\x79"
-     "\x00\x80"
-     "\x88\x79\x00\x80\x86\x88\x00\x00\x87\xAC\x01\x00\x8C\x79\x00\x80\x90\x79"
-     "\x00\x80"
-     "\x94\x79\x00\x80\x98\x79\x00\x80\xB8\xD9\x3E\x00\xB9\xD9\x3E\x00\xBA\xE9"
-     "\x3E\x00"
-     "\xBB\xE9\x3E\x00\xBC\xF9\x3E\x00\xBD\xF9\x3E\x00\xBE\xE9\x3E\x00\xBF\xE9"
-     "\x3E\x00"
-     "\xB0\x49\x3E\x00\xB1\x49\x3E\x00\xB2\x59\x3E\x00\xB3\x59\x3E\x00\xB4\x49"
-     "\x3E\x00"
-     "\xB5\x49\x3E\x00\xB6\xFD\x3E\x00\xB7\xE9\x3E\x00\xA8\x5D\x3E\x00\xA9\x61"
-     "\x3E\x00"
-     "\xAA\x79\x3E\x00\xAB\x79\x3E\x00\xAC\x69\x3E\x00\xAD\x69\x3E\x00\xAE\x95"
-     "\x01\x00"
-     "\xAF\x8D\x01\x00\x84\xA0\x01\x00\x9C\x79\x00\x80\xA0\x79\x00\x80\xA4\x79"
-     "\x00\x80"
-     "\xA8\x79\x00\x80\xAC\x79\x00\x80\xB0\x79\x00\x80\xB4\x79\x00\x80\xB8\xA9"
-     "\x01\x00"
-     "\xB9\xA9\x01\x00\xBA\x4D\x01\x00\xBB\x45\x01\x00\xBC\x5D\x01\x00\xBD\x45"
-     "\x01\x00"
-     "\xBE\x45\x01\x00\xBF\x75\x01\x00\xB0\xFD\x01\x00\xB1\xCD\x01\x00\xB2\xC5"
-     "\x01\x00"
-     "\xB3\xAD\x01\x00\xB4\xB9\x01\x00\xB5\xB9\x01\x00\xB6\xAD\x01\x00\xB7\xA5"
-     "\x01\x00"
-     "\xB3\xF1\x3D\x00\xB8\x79\x00\x80\xBE\x44\x03\x00\xBE\x44\x04\x00\xBC\x79"
-     "\x00\x80"
-     "\xB6\xE9\x3D\x00\xB5\xE1\x3D\x00\xC0\x79\x00\x80\xBB\x09\x02\x00\xBA\x39"
-     "\x02\x00"
-     "\xC4\x79\x00\x80\xC8\x79\x00\x80\xBF\x7D\x02\x00\xBE\x7D\x02\x00\xBD\x79"
-     "\x02\x00"
-     "\xBC\x11\x02\x00\xCC\x79\x00\x80\xA3\xB5\x3D\x00\xD0\x79\x00\x80\xD4\x79"
-     "\x00\x80"
-     "\xA6\xAD\x3D\x00\xD8\x79\x00\x80\xDC\x79\x00\x80\xA5\xA5\x3D\x00\xAA\x7D"
-     "\x02\x00"
-     "\xAB\x4D\x02\x00\xE0\x79\x00\x80\xE4\x79\x00\x80\xAE\x39\x02\x00\xAF\x39"
-     "\x02\x00"
-     "\xAC\x55\x02\x00\xAD\x3D\x02\x00\x80\x11\x00\x00\x81\x11\x00\x00\x82\x11"
-     "\x00\x00"
-     "\xBE\x0C\x06\x00\xE8\x79\x00\x80\xEC\x79\x00\x80\xF0\x79\x00\x80\xEF\x6C"
-     "\x02\x00"
-     "\xF8\x79\x00\x80\xE1\x94\x01\x00\xFC\x79\x00\x80\xE3\x1C\x01\x00\x00\x7A"
-     "\x00\x80"
-     "\x04\x7A\x00\x80\x86\x18\x04\x00\x87\x70\x03\x00\x08\x7A\x00\x80\x0C\x7A"
-     "\x00\x80"
-     "\x10\x7A\x00\x80\xB3\xF9\x03\x00\x14\x7A\x00\x80\xB5\xE9\x03\x00\xB6\x7D"
-     "\x03\x00"
-     "\x18\x7A\x00\x80\xBE\xAC\x07\x00\x1C\x7A\x00\x80\xBA\x59\x03\x00\xBB\x59"
-     "\x03\x00"
-     "\xBC\x49\x03\x00\xBD\x49\x03\x00\xBE\xFD\x00\x00\xBF\xED\x00\x00\xA8\x25"
-     "\x02\x00"
-     "\xA9\x35\x02\x00\xAA\x3D\x02\x00\xAB\x35\x02\x00\xAC\x2D\x02\x00\xAD\xB1"
-     "\x02\x00"
-     "\xAE\xB1\x02\x00\xAF\xB1\x02\x00\xBE\x68\x09\x00\x20\x7A\x00\x80\x24\x7A"
-     "\x00\x80"
-     "\x28\x7A\x00\x80\xBE\xF8\x0A\x00\x2C\x7A\x00\x80\x30\x7A\x00\x80\x34\x7A"
-     "\x00\x80"
-     "\xB8\x6D\x01\x00\xB9\x05\x01\x00\xBA\x01\x01\x00\xBB\x01\x01\x00\xBC\x05"
-     "\x01\x00"
-     "\xBD\x0D\x01\x00\xBE\x31\x01\x00\xBF\x31\x01\x00\xB0\xD1\x02\x00\xB1\xD1"
-     "\x02\x00"
-     "\xB2\xD1\x02\x00\xB3\xD1\x02\x00\xB4\x75\x01\x00\xB5\x65\x01\x00\xB6\x65"
-     "\x01\x00"
-     "\xB7\x55\x01\x00\xE1\x4C\x3F\x00\xE3\x68\x01\x00\xE3\x08\x3E\x00\xE1\x7C"
-     "\x3E\x00"
-     "\x38\x7A\x00\x80\x3C\x7A\x00\x80\x40\x7A\x00\x80\x44\x7A\x00\x80\xBE\x2C"
-     "\x0D\x00"
-     "\x48\x7A\x00\x80\x4C\x7A\x00\x80\x50\x7A\x00\x80\xBE\x3C\x0C\x00\xEF\xBC"
-     "\x01\x00"
-     "\xEF\xCC\x3E\x00\x54\x7A\x00\x80\xA3\xF5\x02\x00\x82\x2D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x58\x7A\x00\x80\xA6\x71\x02\x00\xA5\xE5\x02\x00\x5C\x7A"
-     "\x00\x80"
-     "\xAB\x55\x02\x00\xAA\x55\x02\x00\x60\x7A\x00\x80\x84\xA0\x05\x00\xAF\xE1"
-     "\x01\x00"
-     "\xAE\xF1\x01\x00\xAD\x45\x02\x00\xAC\x45\x02\x00\xA8\xA1\x06\x00\xA9\xAD"
-     "\x06\x00"
-     "\xAA\xB9\x06\x00\xAB\xB9\x06\x00\xAC\xD1\x06\x00\xAD\xD1\x06\x00\xAE\xD1"
-     "\x06\x00"
-     "\xAF\xCD\x06\x00\x64\x7A\x00\x80\x68\x7A\x00\x80\x86\x08\x03\x00\x87\x08"
-     "\x03\x00"
-     "\x6C\x7A\x00\x80\x70\x7A\x00\x80\x74\x7A\x00\x80\x78\x7A\x00\x80\xB8\xD1"
-     "\x07\x00"
-     "\xB9\xD1\x07\x00\xBA\xD1\x07\x00\xBB\xD1\x07\x00\xBC\xF5\x07\x00\xBD\xFD"
-     "\x07\x00"
-     "\xBE\xF5\x07\x00\xBF\xE1\x07\x00\xB0\xBD\x06\x00\xB1\x81\x06\x00\xB2\x4D"
-     "\x07\x00"
-     "\xB3\x45\x07\x00\xB4\x5D\x07\x00\xB5\x41\x07\x00\xB6\x41\x07\x00\xB7\x41"
-     "\x07\x00"
-     "\xB3\x31\x06\x00\x7C\x7A\x00\x80\x80\x7A\x00\x80\x84\x7A\x00\x80\x88\x7A"
-     "\x00\x80"
-     "\xB6\x51\x06\x00\xB5\x59\x06\x00\x8C\x7A\x00\x80\xBB\x45\x06\x00\xBA\x79"
-     "\x06\x00"
-     "\x90\x7A\x00\x80\x94\x7A\x00\x80\xBF\x85\x07\x00\xBE\x99\x07\x00\xBD\x91"
-     "\x07\x00"
-     "\xBC\x99\x07\x00\x98\x7A\x00\x80\xA3\x75\x06\x00\x9C\x7A\x00\x80\xA0\x7A"
-     "\x00\x80"
-     "\xA6\x15\x06\x00\xA4\x7A\x00\x80\xA8\x7A\x00\x80\xA5\x1D\x06\x00\xAA\x3D"
-     "\x06\x00"
-     "\xAB\x01\x06\x00\xAC\x7A\x00\x80\xB0\x7A\x00\x80\xAE\xDD\x07\x00\xAF\xC1"
-     "\x07\x00"
-     "\xAC\xDD\x07\x00\xAD\xD5\x07\x00\xE1\xA0\x05\x00\xB4\x7A\x00\x80\xE3\xF4"
-     "\x04\x00"
-     "\xB8\x7A\x00\x80\xBC\x7A\x00\x80\xC0\x7A\x00\x80\xC4\x7A\x00\x80\xC8\x7A"
-     "\x00\x80"
-     "\x81\x09\x00\x00\x80\x39\x00\x00\xCC\x7A\x00\x80\x82\x1D\x00\x00\xD0\x7A"
-     "\x00\x80"
-     "\xD4\x7A\x00\x80\xEF\xCC\x05\x00\x84\xB0\x02\x00\xB3\x75\x01\x00\xD8\x7A"
-     "\x00\x80"
-     "\x86\xE8\x00\x00\x87\x2C\x01\x00\xDC\x7A\x00\x80\xB6\xD1\x01\x00\xB5\x65"
-     "\x01\x00"
-     "\xE0\x7A\x00\x80\xBB\xFD\x01\x00\xBA\xF5\x01\x00\xE4\x7A\x00\x80\xE8\x7A"
-     "\x00\x80"
-     "\xBF\xB9\x01\x00\xBE\xB5\x01\x00\xBD\xD5\x01\x00\xBC\xD5\x01\x00\xA8\xA5"
-     "\x01\x00"
-     "\xA9\xB5\x01\x00\xAA\xBD\x01\x00\xAB\xB5\x01\x00\xAC\xAD\x01\x00\xAD\x11"
-     "\x01\x00"
-     "\xAE\x11\x01\x00\xAF\x11\x01\x00\x84\xAC\x01\x00\xEC\x7A\x00\x80\xF0\x7A"
-     "\x00\x80"
-     "\xF4\x7A\x00\x80\xF8\x7A\x00\x80\xFC\x7A\x00\x80\x00\x7B\x00\x80\x04\x7B"
-     "\x00\x80"
-     "\xB8\x19\x01\x00\xB9\x2D\x01\x00\xBA\x25\x01\x00\xBB\xC9\x00\x00\xBC\xDD"
-     "\x00\x00"
-     "\xBD\xC5\x00\x00\xBE\xCD\x00\x00\xBF\xC5\x00\x00\xB0\x71\x01\x00\xB1\x79"
-     "\x01\x00"
-     "\xB2\x41\x01\x00\xB3\x41\x01\x00\xB4\x31\x01\x00\xB5\x3D\x01\x00\xB6\x35"
-     "\x01\x00"
-     "\xB7\x29\x01\x00\x08\x7B\x00\x80\x0C\x7B\x00\x80\x10\x7B\x00\x80\xA3\x39"
-     "\x02\x00"
-     "\x14\x7B\x00\x80\xA5\x29\x02\x00\xA6\x9D\x02\x00\x18\x7B\x00\x80\x1C\x7B"
-     "\x00\x80"
-     "\x20\x7B\x00\x80\xAA\xB9\x02\x00\xAB\xB1\x02\x00\xAC\x99\x02\x00\xAD\x99"
-     "\x02\x00"
-     "\xAE\xF9\x02\x00\xAF\xF5\x02\x00\x24\x7B\x00\x80\x28\x7B\x00\x80\x86\xE8"
-     "\x0C\x00"
-     "\x87\x68\x03\x00\x2C\x7B\x00\x80\x30\x7B\x00\x80\x34\x7B\x00\x80\x38\x7B"
-     "\x00\x80"
-     "\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x3C\x7B\x00\x80\x40\x7B"
-     "\x00\x80"
-     "\x48\x7B\x00\x80\x4C\x7B\x00\x80\x50\x7B\x00\x80\x54\x7B\x00\x80\xF4\x79"
-     "\x00\x80"
-     "\xEF\x90\x06\x00\x58\x7B\x00\x80\x5C\x7B\x00\x80\x60\x7B\x00\x80\x64\x7B"
-     "\x00\x80"
-     "\x68\x7B\x00\x80\x84\xEC\x0C\x00\x6C\x7B\x00\x80\x70\x7B\x00\x80\x74\x7B"
-     "\x00\x80"
-     "\xE1\xE0\x01\x00\x78\x7B\x00\x80\xE3\x08\x01\x00\x7C\x7B\x00\x80\xB3\x9D"
-     "\x03\x00"
-     "\x80\x7B\x00\x80\x84\x7B\x00\x80\x88\x7B\x00\x80\x8C\x7B\x00\x80\xB6\x71"
-     "\x03\x00"
-     "\xB5\xB5\x03\x00\x90\x7B\x00\x80\xBB\x55\x03\x00\xBA\x55\x03\x00\x94\x7B"
-     "\x00\x80"
-     "\x98\x7B\x00\x80\xBF\xF5\x00\x00\xBE\xF5\x00\x00\xBD\x45\x03\x00\xBC\x45"
-     "\x03\x00"
-     "\xA8\x45\x02\x00\xA9\x55\x02\x00\xAA\x55\x02\x00\xAB\x65\x02\x00\xAC\x7D"
-     "\x02\x00"
-     "\xAD\xB1\x02\x00\xAE\xB1\x02\x00\xAF\xB1\x02\x00\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\x9C\x7B\x00\x80\x82\xAD\x00\x00\xA0\x7B\x00\x80\xA4\x7B\x00\x80\xA8\x7B"
-     "\x00\x80"
-     "\xAC\x7B\x00\x80\xB8\x5D\x01\x00\xB9\x65\x01\x00\xBA\x6D\x01\x00\xBB\x0D"
-     "\x01\x00"
-     "\xBC\x15\x01\x00\xBD\x05\x01\x00\xBE\x0D\x01\x00\xBF\x05\x01\x00\xB0\xD1"
-     "\x02\x00"
-     "\xB1\xD1\x02\x00\xB2\xD1\x02\x00\xB3\xD1\x02\x00\xB4\x71\x01\x00\xB5\x71"
-     "\x01\x00"
-     "\xB6\x71\x01\x00\xB7\x6D\x01\x00\xB0\x7B\x00\x80\xE1\xD0\x2B\x00\xB8\x7B"
-     "\x00\x80"
-     "\xE3\xD8\x17\x00\x86\x90\x0F\x00\x87\x5C\x0D\x00\xBC\x7B\x00\x80\xEF\xE8"
-     "\x07\x00"
-     "\xC0\x7B\x00\x80\xE1\x94\x01\x00\xC4\x7B\x00\x80\xE3\x80\x00\x00\xBE\xB0"
-     "\x0B\x00"
-     "\xC8\x7B\x00\x80\xCC\x7B\x00\x80\xEF\x60\x3F\x00\xA3\x9D\x02\x00\xD0\x7B"
-     "\x00\x80"
-     "\xD4\x7B\x00\x80\xD8\x7B\x00\x80\xDC\x7B\x00\x80\xA6\x71\x02\x00\xA5\xB5"
-     "\x02\x00"
-     "\xE0\x7B\x00\x80\xAB\x55\x02\x00\xAA\x55\x02\x00\xE4\x7B\x00\x80\xE8\x7B"
-     "\x00\x80"
-     "\xAF\xF5\x01\x00\xAE\xF5\x01\x00\xAD\x45\x02\x00\xAC\x45\x02\x00\xA8\x75"
-     "\x0E\x00"
-     "\xA9\x7D\x0E\x00\xAA\x75\x0E\x00\xAB\x8D\x0E\x00\xAC\x9D\x0E\x00\xAD\x85"
-     "\x0E\x00"
-     "\xAE\x8D\x0E\x00\xAF\x85\x0E\x00\xB4\x7B\x00\x80\xEC\x7B\x00\x80\xF0\x7B"
-     "\x00\x80"
-     "\xF4\x7B\x00\x80\xF8\x7B\x00\x80\xFC\x7B\x00\x80\x00\x7C\x00\x80\x04\x7C"
-     "\x00\x80"
-     "\xB8\x8D\x0E\x00\xB9\x95\x0E\x00\xBA\x95\x0E\x00\xBB\xA5\x0E\x00\xBC\xBD"
-     "\x0E\x00"
-     "\xBD\x7D\x0F\x00\xBE\x69\x0F\x00\xBF\x69\x0F\x00\xB0\xFD\x0E\x00\xB1\xC1"
-     "\x0E\x00"
-     "\xB2\xC1\x0E\x00\xB3\xC1\x0E\x00\xB4\xC5\x0E\x00\xB5\xC9\x0E\x00\xB6\xBD"
-     "\x0E\x00"
-     "\xB7\xB5\x0E\x00\xB3\xFD\x0E\x00\x08\x7C\x00\x80\x0C\x7C\x00\x80\x10\x7C"
-     "\x00\x80"
-     "\x14\x7C\x00\x80\xB6\x05\x0E\x00\xB5\x15\x0E\x00\x18\x7C\x00\x80\xBB\x05"
-     "\x0E\x00"
-     "\xBA\x39\x0E\x00\x1C\x7C\x00\x80\x20\x7C\x00\x80\xBF\x0D\x0E\x00\xBE\x01"
-     "\x0E\x00"
-     "\xBD\x15\x0E\x00\xBC\x15\x0E\x00\x82\x6D\x00\x00\xA3\xB9\x0E\x00\x80\x55"
-     "\x00\x00"
-     "\x81\x65\x00\x00\xA6\x41\x0E\x00\x84\x18\x01\x00\x28\x7C\x00\x80\xA5\x51"
-     "\x0E\x00"
-     "\xAA\x7D\x0E\x00\xAB\x41\x0E\x00\x86\x00\x04\x00\x87\x78\x00\x00\xAE\x45"
-     "\x0E\x00"
-     "\xAF\x49\x0E\x00\xAC\x51\x0E\x00\xAD\x51\x0E\x00\x8F\x29\x25\x00\xE1\xF4"
-     "\x03\x00"
-     "\x2C\x7C\x00\x80\xE3\x54\x01\x00\x30\x7C\x00\x80\xE3\xA4\x0D\x00\x34\x7C"
-     "\x00\x80"
-     "\xE1\x4C\x0D\x00\x38\x7C\x00\x80\xEF\xFC\x08\x00\x3C\x7C\x00\x80\x40\x7C"
-     "\x00\x80"
-     "\x44\x7C\x00\x80\x48\x7C\x00\x80\x4C\x7C\x00\x80\xEF\x64\x08\x00\x9F\x61"
-     "\x08\x00"
-     "\x9E\x4D\x09\x00\x9D\x4D\x09\x00\x9C\x35\x34\x00\x9B\x0D\x35\x00\x9A\x51"
-     "\x35\x00"
-     "\x99\xB9\x37\x00\x98\x6D\x31\x00\x97\x41\x31\x00\x96\x89\x33\x00\x95\xB9"
-     "\x3D\x00"
-     "\x94\x4D\x3D\x00\x93\xA9\x3F\x00\x92\xAD\x39\x00\x91\x7D\x39\x00\x90\x7D"
-     "\x39\x00"
-     "\xB1\xA5\x15\x00\xB0\xA5\x15\x00\xB3\x89\xE9\x01\xB2\x41\x14\x00\xB5\x55"
-     "\xE8\x01"
-     "\xB4\x89\xE9\x01\x50\x7C\x00\x80\x54\x7C\x00\x80\x58\x7C\x00\x80\x5C\x7C"
-     "\x00\x80"
-     "\x60\x7C\x00\x80\x64\x7C\x00\x80\xA3\x29\x03\x00\xA2\x81\x0C\x00\xA1\x99"
-     "\x0D\x00"
-     "\xA0\x39\x0F\x00\xA1\xE1\x01\x00\x68\x7C\x00\x80\xA3\xCD\x05\x00\xA2\x4D"
-     "\x03\x00"
-     "\xA5\x81\x04\x00\xA4\xDD\x05\x00\xA7\xD5\x19\x00\xA6\xDD\x19\x00\xA9\xE1"
-     "\x1D\x00"
-     "\xA8\x8D\x18\x00\xAB\xE5\x1C\x00\xAA\xE9\x1D\x00\xAD\xD1\x11\x00\xAC\x01"
-     "\x10\x00"
-     "\xAF\x01\x14\x00\xAE\x51\x11\x00\x82\x29\x26\x00\x83\x05\x3A\x00\x6C\x7C"
-     "\x00\x80"
-     "\x70\x7C\x00\x80\x86\xC9\x3C\x00\x87\x35\x3E\x00\x84\x1D\x3A\x00\x85\xC5"
-     "\x3A\x00"
-     "\x8A\xF1\x3E\x00\x8B\xC9\x30\x00\x74\x7C\x00\x80\x78\x7C\x00\x80\x8E\xD1"
-     "\x34\x00"
-     "\x8F\x31\x36\x00\x8C\x19\x32\x00\x8D\x0D\x32\x00\x92\x1D\x36\x00\x93\x3D"
-     "\x37\x00"
-     "\x84\x44\x03\x00\x84\xC4\x04\x00\x96\x71\x0B\x00\x97\xD1\x0C\x00\x94\x5D"
-     "\x0A\x00"
-     "\x95\x55\x0A\x00\x9A\x41\x0E\x00\x9B\x51\x0F\x00\x7C\x7C\x00\x80\x80\x7C"
-     "\x00\x80"
-     "\x84\x7C\x00\x80\x88\x7C\x00\x80\x9C\xE1\x00\x00\x8C\x7C\x00\x80\x24\x7C"
-     "\x00\x80"
-     "\x90\x7C\x00\x80\x98\x7C\x00\x80\x9C\x7C\x00\x80\x81\xED\x00\x00\x80\xE1"
-     "\x00\x00"
-     "\x44\x7B\x00\x80\x82\x19\x00\x00\xA8\x29\x0D\x00\xA9\x3D\x0D\x00\xAA\x09"
-     "\x0D\x00"
-     "\xAB\x05\x0D\x00\xAC\x05\x0D\x00\xAD\x0D\x0D\x00\xAE\x7D\x0D\x00\xAF\x75"
-     "\x0D\x00"
-     "\xB0\x0D\x0D\x00\xB1\x15\x0D\x00\xB2\x15\x0D\x00\xB3\x25\x0D\x00\xB4\x3D"
-     "\x0D\x00"
-     "\xB5\xF1\x02\x00\xB6\xF1\x02\x00\xB7\xF1\x02\x00\xB8\xD1\x02\x00\xB9\xD1"
-     "\x02\x00"
-     "\xBA\xD1\x02\x00\xBB\xD1\x02\x00\xBC\x75\x02\x00\xBD\x7D\x02\x00\xBE\x75"
-     "\x02\x00"
-     "\xBF\x6D\x02\x00\xB3\xA1\x0C\x00\xA0\x7C\x00\x80\x87\xB4\x02\x00\x86\x2C"
-     "\x05\x00"
-     "\xA4\x7C\x00\x80\xB6\xAD\x0C\x00\xB5\xAD\x0C\x00\xA8\x7C\x00\x80\xBB\xF1"
-     "\x0C\x00"
-     "\xBA\xE9\x0C\x00\xAC\x7C\x00\x80\xB0\x7C\x00\x80\xBF\x15\x03\x00\xBE\xD9"
-     "\x0C\x00"
-     "\xBD\xD1\x0C\x00\xBC\xE1\x0C\x00\xB4\x7C\x00\x80\xA3\xE5\x0C\x00\xB8\x7C"
-     "\x00\x80"
-     "\xBC\x7C\x00\x80\xA6\xE9\x0C\x00\xC0\x7C\x00\x80\xC4\x7C\x00\x80\xA5\xE9"
-     "\x0C\x00"
-     "\xAA\xAD\x0C\x00\xAB\xB5\x0C\x00\xC8\x7C\x00\x80\xCC\x7C\x00\x80\xAE\x9D"
-     "\x0C\x00"
-     "\xAF\x51\x03\x00\xAC\xA5\x0C\x00\xAD\x95\x0C\x00\xD0\x7C\x00\x80\xD4\x7C"
-     "\x00\x80"
-     "\xD8\x7C\x00\x80\xDC\x7C\x00\x80\xE0\x7C\x00\x80\xE4\x7C\x00\x80\xE8\x7C"
-     "\x00\x80"
-     "\xEF\x54\x0A\x00\xEC\x7C\x00\x80\xE1\x2C\x0E\x00\xF0\x7C\x00\x80\xE3\x68"
-     "\x01\x00"
-     "\xF4\x7C\x00\x80\x82\x79\x00\x00\x81\x71\x00\x00\x80\x79\x00\x00\xB3\xD1"
-     "\x02\x00"
-     "\xF8\x7C\x00\x80\xFC\x7C\x00\x80\x00\x7D\x00\x80\x08\x7D\x00\x80\xB6\xFD"
-     "\x02\x00"
-     "\xB5\xF9\x02\x00\x0C\x7D\x00\x80\xBB\xD9\x02\x00\xBA\xD1\x02\x00\x86\xC8"
-     "\x04\x00"
-     "\x87\xAC\x04\x00\xBF\x55\x02\x00\xBE\x55\x02\x00\xBD\x65\x02\x00\xBC\x65"
-     "\x02\x00"
-     "\xA3\x11\x05\x00\x94\x7C\x00\x80\x04\x7D\x00\x80\x10\x7D\x00\x80\x14\x7D"
-     "\x00\x80"
-     "\xA6\x3D\x05\x00\xA5\x39\x05\x00\x18\x7D\x00\x80\xAB\x19\x05\x00\xAA\x11"
-     "\x05\x00"
-     "\x1C\x7D\x00\x80\x20\x7D\x00\x80\xAF\x95\x05\x00\xAE\x95\x05\x00\xAD\xA5"
-     "\x05\x00"
-     "\xAC\xA5\x05\x00\x24\x7D\x00\x80\xB3\x41\x07\x00\x28\x7D\x00\x80\x2C\x7D"
-     "\x00\x80"
-     "\xB6\x79\x07\x00\x30\x7D\x00\x80\x34\x7D\x00\x80\xB5\x41\x07\x00\xBA\x29"
-     "\x07\x00"
-     "\xBB\x29\x07\x00\x38\x7D\x00\x80\x3C\x7D\x00\x80\xBE\x19\x07\x00\xBF\x19"
-     "\x07\x00"
-     "\xBC\x2D\x07\x00\xBD\x25\x07\x00\xA8\x09\x06\x00\xA9\x09\x06\x00\xAA\x8D"
-     "\x07\x00"
-     "\xAB\x9D\x07\x00\xAC\x8D\x07\x00\xAD\xBD\x07\x00\xAE\xB5\x07\x00\xAF\xDD"
-     "\x07\x00"
-     "\x40\x7D\x00\x80\x44\x7D\x00\x80\x48\x7D\x00\x80\x4C\x7D\x00\x80\x50\x7D"
-     "\x00\x80"
-     "\x54\x7D\x00\x80\x58\x7D\x00\x80\x5C\x7D\x00\x80\xB8\x49\x07\x00\xB9\x49"
-     "\x07\x00"
-     "\xBA\x59\x07\x00\xBB\x59\x07\x00\xBC\x49\x07\x00\xBD\x75\x07\x00\xBE\x71"
-     "\x07\x00"
-     "\xBF\x6D\x07\x00\xB0\xA9\x07\x00\xB1\xA9\x07\x00\xB2\xBD\x07\x00\xB3\xB5"
-     "\x07\x00"
-     "\xB4\x99\x07\x00\xB5\x99\x07\x00\xB6\x79\x07\x00\xB7\x79\x07\x00\xA3\x05"
-     "\x06\x00"
-     "\x60\x7D\x00\x80\x64\x7D\x00\x80\x68\x7D\x00\x80\x6C\x7D\x00\x80\xA6\x3D"
-     "\x06\x00"
-     "\xA5\x05\x06\x00\x70\x7D\x00\x80\xAB\x6D\x06\x00\xAA\x6D\x06\x00\x84\xA8"
-     "\x02\x00"
-     "\x78\x7D\x00\x80\xAF\x5D\x06\x00\xAE\x5D\x06\x00\xAD\x61\x06\x00\xAC\x69"
-     "\x06\x00"
-     "\x80\x4D\x00\x00\x81\x55\x00\x00\x82\x55\x00\x00\xB3\x41\x01\x00\x74\x7D"
-     "\x00\x80"
-     "\xB5\x49\x01\x00\xB6\x71\x01\x00\x7C\x7D\x00\x80\x86\xC0\x00\x00\x87\xA8"
-     "\x01\x00"
-     "\xBA\x21\x01\x00\xBB\x21\x01\x00\xBC\x25\x01\x00\xBD\x2D\x01\x00\xBE\x11"
-     "\x01\x00"
-     "\xBF\x11\x01\x00\xA8\x51\x06\x00\xA9\x59\x06\x00\xAA\x85\x01\x00\xAB\x9D"
-     "\x01\x00"
-     "\xAC\x8D\x01\x00\xAD\xBD\x01\x00\xAE\xB5\x01\x00\xAF\xDD\x01\x00\x84\xA0"
-     "\x00\x00"
-     "\x80\x7D\x00\x80\x84\x7D\x00\x80\x88\x7D\x00\x80\x8C\x7D\x00\x80\x90\x7D"
-     "\x00\x80"
-     "\x94\x7D\x00\x80\x98\x7D\x00\x80\xB8\x69\x01\x00\xB9\x69\x01\x00\xBA\x79"
-     "\x01\x00"
-     "\xBB\x79\x01\x00\xBC\x69\x01\x00\xBD\x69\x01\x00\xBE\xDD\x00\x00\xBF\xD5"
-     "\x00\x00"
-     "\xB0\xA9\x01\x00\xB1\xA9\x01\x00\xB2\xBD\x01\x00\xB3\xB5\x01\x00\xB4\x99"
-     "\x01\x00"
-     "\xB5\x99\x01\x00\xB6\x59\x01\x00\xB7\x59\x01\x00\xBA\x0D\x04\x00\xBB\x2D"
-     "\x04\x00"
-     "\xB8\xCD\x07\x00\xB9\x05\x04\x00\xBE\x7D\x04\x00\xBF\x65\x04\x00\xBC\x35"
-     "\x04\x00"
-     "\xBD\x75\x04\x00\xB2\xB9\x07\x00\xB3\xB9\x07\x00\xB0\x1D\x07\x00\xB1\x15"
-     "\x07\x00"
-     "\xB6\x99\x07\x00\xB7\x99\x07\x00\xB4\x99\x07\x00\xB5\x99\x07\x00\xAA\x51"
-     "\x07\x00"
-     "\xAB\x5D\x07\x00\xA8\x91\x06\x00\xA9\x45\x07\x00\xAE\x7D\x07\x00\xAF\x65"
-     "\x07\x00"
-     "\xAC\x79\x07\x00\xAD\x75\x07\x00\x9C\x7D\x00\x80\xA0\x7D\x00\x80\xA4\x7D"
-     "\x00\x80"
-     "\xA8\x7D\x00\x80\xAC\x7D\x00\x80\xB0\x7D\x00\x80\xB4\x7D\x00\x80\xB8\x7D"
-     "\x00\x80"
-     "\xBC\x7D\x00\x80\xC0\x7D\x00\x80\xBE\x00\x1C\x00\xA3\x81\x02\x00\xC4\x7D"
-     "\x00\x80"
-     "\xA5\x89\x02\x00\xA6\xB1\x02\x00\xC8\x7D\x00\x80\xCC\x7D\x00\x80\xD0\x7D"
-     "\x00\x80"
-     "\xAA\xE1\x02\x00\xAB\xE1\x02\x00\xAC\xE5\x02\x00\xAD\xED\x02\x00\xAE\xD1"
-     "\x02\x00"
-     "\xAF\xD1\x02\x00\xD4\x7D\x00\x80\x82\x25\x00\x00\x81\x11\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xD8\x7D\x00\x80\xE1\xAC\x01\x00\xDC\x7D\x00\x80\xE3\x6C\x01\x00\xE0\x7D"
-     "\x00\x80"
-     "\xE8\x7D\x00\x80\xEC\x7D\x00\x80\xEF\x9C\x02\x00\x87\xB0\x1C\x00\x86\xD4"
-     "\x1C\x00"
-     "\xF0\x7D\x00\x80\xF4\x7D\x00\x80\xA8\x2D\x02\x00\xA9\x59\x02\x00\xAA\xAD"
-     "\x02\x00"
-     "\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xA5\x02\x00\xAE\xAD\x02\x00\xAF\xA5"
-     "\x02\x00"
-     "\x84\xA0\x1D\x00\xF8\x7D\x00\x80\xFC\x7D\x00\x80\x00\x7E\x00\x80\x04\x7E"
-     "\x00\x80"
-     "\x08\x7E\x00\x80\xBE\xB4\x1C\x00\x0C\x7E\x00\x80\xB8\x3D\x01\x00\xB9\xC5"
-     "\x01\x00"
-     "\xBA\xCD\x01\x00\xBB\xDD\x01\x00\xBC\xCD\x01\x00\xBD\xF5\x01\x00\xBE\xFD"
-     "\x01\x00"
-     "\xBF\x9D\x01\x00\xB0\xDD\x02\x00\xB1\x05\x01\x00\xB2\x0D\x01\x00\xB3\x05"
-     "\x01\x00"
-     "\xB4\x1D\x01\x00\xB5\x05\x01\x00\xB6\x0D\x01\x00\xB7\x05\x01\x00\xE1\x94"
-     "\x06\x00"
-     "\xE3\xBC\x00\x00\xE3\xD4\x07\x00\xE1\x7C\x06\x00\x10\x7E\x00\x80\x14\x7E"
-     "\x00\x80"
-     "\x18\x7E\x00\x80\x1C\x7E\x00\x80\x20\x7E\x00\x80\x24\x7E\x00\x80\x28\x7E"
-     "\x00\x80"
-     "\x2C\x7E\x00\x80\xBE\x78\x1F\x00\xEF\xFC\x0A\x00\xEF\xF0\x0E\x00\x30\x7E"
-     "\x00\x80"
-     "\x34\x7E\x00\x80\x38\x7E\x00\x80\x3C\x7E\x00\x80\xB3\x69\x02\x00\x40\x7E"
-     "\x00\x80"
-     "\xB5\x79\x02\x00\xB6\x69\x02\x00\x44\x7E\x00\x80\x48\x7E\x00\x80\x4C\x7E"
-     "\x00\x80"
-     "\xBA\x55\x02\x00\xBB\x8D\x01\x00\xBC\x95\x01\x00\xBD\x9D\x01\x00\xBE\x95"
-     "\x01\x00"
-     "\xBF\x8D\x01\x00\xE4\x7D\x00\x80\xA3\xA1\x1D\x00\x50\x7E\x00\x80\x54\x7E"
-     "\x00\x80"
-     "\xA6\xA1\x1D\x00\x58\x7E\x00\x80\x5C\x7E\x00\x80\xA5\xB1\x1D\x00\xAA\x9D"
-     "\x1D\x00"
-     "\xAB\x45\x1E\x00\x86\x08\x03\x00\x87\x7C\x03\x00\xAE\x5D\x1E\x00\xAF\x45"
-     "\x1E\x00"
-     "\xAC\x5D\x1E\x00\xAD\x55\x1E\x00\x80\xA1\x1F\x00\x81\xA1\x1F\x00\x82\x41"
-     "\x00\x00"
-     "\xB3\xB5\x1E\x00\x60\x7E\x00\x80\xB5\x71\x1F\x00\xB6\x71\x1F\x00\x64\x7E"
-     "\x00\x80"
-     "\x68\x7E\x00\x80\x6C\x7E\x00\x80\xBA\x5D\x1F\x00\xBB\x51\x1F\x00\xBC\xE1"
-     "\x1F\x00"
-     "\xBD\xE1\x1F\x00\xBE\xE1\x1F\x00\xBF\xE1\x1F\x00\xA8\x65\x1E\x00\xA9\x6D"
-     "\x1E\x00"
-     "\xAA\x71\x1E\x00\xAB\x71\x1E\x00\xAC\xB5\x1F\x00\xAD\xBD\x1F\x00\xAE\xB5"
-     "\x1F\x00"
-     "\xAF\xAD\x1F\x00\x70\x7E\x00\x80\x74\x7E\x00\x80\x78\x7E\x00\x80\x7C\x7E"
-     "\x00\x80"
-     "\x80\x7E\x00\x80\x84\x7E\x00\x80\x88\x7E\x00\x80\x8C\x7E\x00\x80\xB8\x69"
-     "\x1F\x00"
-     "\xB9\x69\x1F\x00\xBA\x75\x1F\x00\xBB\x0D\x1F\x00\xBC\x15\x1F\x00\xBD\x1D"
-     "\x1F\x00"
-     "\xBE\x09\x1F\x00\xBF\x01\x1F\x00\xB0\xD5\x1F\x00\xB1\xD9\x1F\x00\xB2\x79"
-     "\x1F\x00"
-     "\xB3\x79\x1F\x00\xB4\x69\x1F\x00\xB5\x69\x1F\x00\xB6\x59\x1F\x00\xB7\x59"
-     "\x1F\x00"
-     "\xA3\xF9\x1F\x00\x90\x7E\x00\x80\x94\x7E\x00\x80\x98\x7E\x00\x80\x9C\x7E"
-     "\x00\x80"
-     "\xA6\x3D\x1E\x00\xA5\x3D\x1E\x00\xA0\x7E\x00\x80\xAB\x1D\x1E\x00\xAA\x11"
-     "\x1E\x00"
-     "\xA4\x7E\x00\x80\xA8\x7E\x00\x80\xAF\xAD\x1E\x00\xAE\xAD\x1E\x00\xAD\xAD"
-     "\x1E\x00"
-     "\xAC\xAD\x1E\x00\xAC\x7E\x00\x80\xB0\x7E\x00\x80\xB4\x7E\x00\x80\xB8\x7E"
-     "\x00\x80"
-     "\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39\x00\x00\xBC\x7E\x00\x80\xC0\x7E"
-     "\x00\x80"
-     "\x84\xA8\x01\x00\xC4\x7E\x00\x80\xC8\x7E\x00\x80\x86\xD0\x00\x00\x87\xFC"
-     "\x01\x00"
-     "\xCC\x7E\x00\x80\xD0\x7E\x00\x80\xA8\x95\x01\x00\xA9\x9D\x01\x00\xAA\xD9"
-     "\x01\x00"
-     "\xAB\xD1\x01\x00\xAC\xF1\x01\x00\xAD\xF1\x01\x00\xAE\x39\x01\x00\xAF\x39"
-     "\x01\x00"
-     "\x84\xA0\x00\x00\xD4\x7E\x00\x80\xD8\x7E\x00\x80\xDC\x7E\x00\x80\xE0\x7E"
-     "\x00\x80"
-     "\xE4\x7E\x00\x80\xE8\x7E\x00\x80\xEC\x7E\x00\x80\xB8\xCD\x00\x00\xB9\xD5"
-     "\x00\x00"
-     "\xBA\xD5\x00\x00\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x9D"
-     "\x00\x00"
-     "\xBF\x95\x00\x00\xB0\x49\x01\x00\xB1\x49\x01\x00\xB2\x59\x01\x00\xB3\x59"
-     "\x01\x00"
-     "\xB4\x49\x01\x00\xB5\x49\x01\x00\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xB3\x51"
-     "\x02\x00"
-     "\xF0\x7E\x00\x80\xF4\x7E\x00\x80\xF8\x7E\x00\x80\xFC\x7E\x00\x80\xB6\x71"
-     "\x02\x00"
-     "\xB5\x79\x02\x00\x00\x7F\x00\x80\xBB\x8D\x02\x00\xBA\x8D\x02\x00\x04\x7F"
-     "\x00\x80"
-     "\x08\x7F\x00\x80\xBF\x8D\x02\x00\xBE\x8D\x02\x00\xBD\x8D\x02\x00\xBC\x8D"
-     "\x02\x00"
-     "\x85\x3C\xF4\x01\xA3\x15\x02\x00\x0C\x7F\x00\x80\x10\x7F\x00\x80\xA6\x35"
-     "\x02\x00"
-     "\x14\x7F\x00\x80\x18\x7F\x00\x80\xA5\x3D\x02\x00\xAA\xC9\x02\x00\xAB\xC9"
-     "\x02\x00"
-     "\x1C\x7F\x00\x80\x20\x7F\x00\x80\xAE\xC9\x02\x00\xAF\xC9\x02\x00\xAC\xC9"
-     "\x02\x00"
-     "\xAD\xC9\x02\x00\x82\x19\x00\x00\x24\x7F\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x28\x7F\x00\x80\x2C\x7F\x00\x80\x30\x7F\x00\x80\x38\x7F\x00\x80\xBE\x00"
-     "\x04\x00"
-     "\x3C\x7F\x00\x80\x40\x7F\x00\x80\x44\x7F\x00\x80\x48\x7F\x00\x80\x4C\x7F"
-     "\x00\x80"
-     "\x50\x7F\x00\x80\x54\x7F\x00\x80\x86\xE0\x04\x00\x87\x54\x03\x00\x58\x7F"
-     "\x00\x80"
-     "\x5C\x7F\x00\x80\x60\x7F\x00\x80\xEF\x8C\x02\x00\x64\x7F\x00\x80\x68\x7F"
-     "\x00\x80"
-     "\x6C\x7F\x00\x80\xE3\x54\x01\x00\x70\x7F\x00\x80\xE1\x84\x01\x00\x74\x7F"
-     "\x00\x80"
-     "\x78\x7F\x00\x80\x7C\x7F\x00\x80\x80\x7F\x00\x80\xA8\x6D\x02\x00\xA9\x7D"
-     "\x02\x00"
-     "\xAA\x75\x02\x00\xAB\xA9\x02\x00\xAC\xB9\x02\x00\xAD\xB9\x02\x00\xAE\xA9"
-     "\x02\x00"
-     "\xAF\xA9\x02\x00\x84\xAC\x05\x00\x84\x7F\x00\x80\x88\x7F\x00\x80\x8C\x7F"
-     "\x00\x80"
-     "\x90\x7F\x00\x80\x94\x7F\x00\x80\x98\x7F\x00\x80\x9C\x7F\x00\x80\xB8\x6D"
-     "\x01\x00"
-     "\xB9\x0D\x01\x00\xBA\x05\x01\x00\xBB\x19\x01\x00\xBC\x09\x01\x00\xBD\x3D"
-     "\x01\x00"
-     "\xBE\x35\x01\x00\xBF\xD9\x01\x00\xB0\xD9\x02\x00\xB1\xD9\x02\x00\xB2\x6D"
-     "\x01\x00"
-     "\xB3\x65\x01\x00\xB4\x7D\x01\x00\xB5\x65\x01\x00\xB6\x65\x01\x00\xB7\x55"
-     "\x01\x00"
-     "\xE3\xF4\x00\x00\xE1\x68\x1F\x00\xE1\xB8\x01\x00\xE3\xBC\x1E\x00\x80\xA9"
-     "\x00\x00"
-     "\x81\xA9\x00\x00\x82\x05\x00\x00\xA0\x7F\x00\x80\xBE\x2C\x04\x00\xA8\x7F"
-     "\x00\x80"
-     "\xAC\x7F\x00\x80\xB0\x7F\x00\x80\xEF\xF8\x1E\x00\xB4\x7F\x00\x80\xB8\x7F"
-     "\x00\x80"
-     "\xEF\x7C\x1E\x00\x86\xE0\x04\x00\x87\x74\x05\x00\xBC\x7F\x00\x80\xB3\xB1"
-     "\x02\x00"
-     "\xC0\x7F\x00\x80\xB5\x7D\x02\x00\xB6\x75\x02\x00\xC4\x7F\x00\x80\xC8\x7F"
-     "\x00\x80"
-     "\xCC\x7F\x00\x80\xBA\x51\x02\x00\xBB\x51\x02\x00\xBC\xF5\x01\x00\xBD\xFD"
-     "\x01\x00"
-     "\xBE\xF5\x01\x00\xBF\xED\x01\x00\x34\x7F\x00\x80\xA4\x7F\x00\x80\xAA\xB1"
-     "\x05\x00"
-     "\xAB\xB1\x05\x00\xAC\x15\x06\x00\xAD\x1D\x06\x00\xAE\x15\x06\x00\xAF\x0D"
-     "\x06\x00"
-     "\xD0\x7F\x00\x80\xD4\x7F\x00\x80\xD8\x7F\x00\x80\xA3\x51\x05\x00\xDC\x7F"
-     "\x00\x80"
-     "\xA5\x9D\x05\x00\xA6\x95\x05\x00\xE0\x7F\x00\x80\xB3\xE9\x06\x00\xE4\x7F"
-     "\x00\x80"
-     "\xE8\x7F\x00\x80\xEC\x7F\x00\x80\xF0\x7F\x00\x80\xB6\xE1\x06\x00\xB5\xE9"
-     "\x06\x00"
-     "\xF4\x7F\x00\x80\xBB\xAD\x06\x00\xBA\xAD\x06\x00\xF8\x7F\x00\x80\xFC\x7F"
-     "\x00\x80"
-     "\xBF\x91\x06\x00\xBE\xA9\x06\x00\xBD\xA1\x06\x00\xBC\xAD\x06\x00\xA8\x1D"
-     "\x06\x00"
-     "\xA9\x21\x06\x00\xAA\x21\x06\x00\xAB\x21\x06\x00\xAC\x21\x06\x00\xAD\x21"
-     "\x06\x00"
-     "\xAE\x21\x06\x00\xAF\x21\x06\x00\x00\x80\x00\x80\x04\x80\x00\x80\x08\x80"
-     "\x00\x80"
-     "\x0C\x80\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x10\x80"
-     "\x00\x80"
-     "\xB8\x35\x06\x00\xB9\x3D\x06\x00\xBA\x35\x06\x00\xBB\xE9\x07\x00\xBC\xF9"
-     "\x07\x00"
-     "\xBD\xF9\x07\x00\xBE\xE9\x07\x00\xBF\xE9\x07\x00\xB0\x29\x06\x00\xB1\x29"
-     "\x06\x00"
-     "\xB2\x39\x06\x00\xB3\x35\x06\x00\xB4\x2D\x06\x00\xB5\x15\x06\x00\xB6\x11"
-     "\x06\x00"
-     "\xB7\x0D\x06\x00\xA3\xA9\x07\x00\x84\x8C\x02\x00\xBE\x84\x01\x00\x14\x80"
-     "\x00\x80"
-     "\x18\x80\x00\x80\xA6\xA1\x07\x00\xA5\xA9\x07\x00\x1C\x80\x00\x80\xAB\xED"
-     "\x07\x00"
-     "\xAA\xED\x07\x00\x86\x88\x00\x00\x87\x78\x01\x00\xAF\xD1\x07\x00\xAE\xE9"
-     "\x07\x00"
-     "\xAD\xE1\x07\x00\xAC\xED\x07\x00\x20\x80\x00\x80\xB3\xE9\x06\x00\x24\x80"
-     "\x00\x80"
-     "\x28\x80\x00\x80\xB6\x99\x06\x00\x2C\x80\x00\x80\x30\x80\x00\x80\xB5\xE5"
-     "\x06\x00"
-     "\xBA\x81\x06\x00\xBB\x81\x06\x00\x34\x80\x00\x80\x38\x80\x00\x80\xBE\x59"
-     "\x01\x00"
-     "\xBF\x45\x01\x00\xBC\x59\x01\x00\xBD\x59\x01\x00\xA8\x25\x06\x00\xA9\x2D"
-     "\x06\x00"
-     "\xAA\x25\x06\x00\xAB\x39\x06\x00\xAC\x29\x06\x00\xAD\x5D\x06\x00\xAE\x49"
-     "\x06\x00"
-     "\xAF\x41\x06\x00\x84\xAC\x01\x00\x3C\x80\x00\x80\x40\x80\x00\x80\x44\x80"
-     "\x00\x80"
-     "\x48\x80\x00\x80\x4C\x80\x00\x80\x50\x80\x00\x80\x54\x80\x00\x80\xB8\x89"
-     "\x01\x00"
-     "\xB9\x89\x01\x00\xBA\x99\x01\x00\xBB\x91\x01\x00\xBC\xB9\x01\x00\xBD\xB9"
-     "\x01\x00"
-     "\xBE\x79\x01\x00\xBF\x79\x01\x00\xB0\x01\x06\x00\xB1\x01\x06\x00\xB2\xC9"
-     "\x01\x00"
-     "\xB3\xC9\x01\x00\xB4\xD9\x01\x00\xB5\xC5\x01\x00\xB6\xC5\x01\x00\xB7\xB9"
-     "\x01\x00"
-     "\xA3\xAD\x05\x00\x58\x80\x00\x80\x5C\x80\x00\x80\x60\x80\x00\x80\x64\x80"
-     "\x00\x80"
-     "\xA6\xDD\x05\x00\xA5\xA1\x05\x00\x68\x80\x00\x80\xAB\xC5\x05\x00\xAA\xC5"
-     "\x05\x00"
-     "\x6C\x80\x00\x80\x70\x80\x00\x80\xAF\x01\x02\x00\xAE\x1D\x02\x00\xAD\x1D"
-     "\x02\x00"
-     "\xAC\x1D\x02\x00\x74\x80\x00\x80\x78\x80\x00\x80\x7C\x80\x00\x80\x80\x80"
-     "\x00\x80"
-     "\x81\x19\x00\x00\x80\x11\x00\x00\x88\x80\x00\x80\x82\x11\x00\x00\xBE\x24"
-     "\x03\x00"
-     "\x8C\x80\x00\x80\x86\xC8\x0C\x00\x87\x4C\x03\x00\x90\x80\x00\x80\x94\x80"
-     "\x00\x80"
-     "\x98\x80\x00\x80\x9C\x80\x00\x80\xE3\x30\x01\x00\xA0\x80\x00\x80\xE1\xB8"
-     "\x01\x00"
-     "\x84\xC0\x0C\x00\xA4\x80\x00\x80\xA8\x80\x00\x80\xAC\x80\x00\x80\xB0\x80"
-     "\x00\x80"
-     "\xBE\xE8\x0C\x00\xB4\x80\x00\x80\xB8\x80\x00\x80\xBC\x80\x00\x80\xEF\x28"
-     "\x02\x00"
-     "\xC0\x80\x00\x80\xC4\x80\x00\x80\xC8\x80\x00\x80\xB3\x81\x03\x00\xCC\x80"
-     "\x00\x80"
-     "\xD0\x80\x00\x80\xD4\x80\x00\x80\xD8\x80\x00\x80\xB6\x81\x03\x00\xB5\x95"
-     "\x03\x00"
-     "\xDC\x80\x00\x80\xBB\x45\x03\x00\xBA\x89\x03\x00\xE0\x80\x00\x80\xE4\x80"
-     "\x00\x80"
-     "\xBF\x45\x03\x00\xBE\x45\x03\x00\xBD\x55\x03\x00\xBC\x55\x03\x00\xA8\x11"
-     "\x02\x00"
-     "\xA9\x19\x02\x00\xAA\x49\x02\x00\xAB\x45\x02\x00\xAC\x5D\x02\x00\xAD\x41"
-     "\x02\x00"
-     "\xAE\x41\x02\x00\xAF\x75\x02\x00\x80\x0D\x00\x00\x81\x11\x00\x00\x82\x11"
-     "\x00\x00"
-     "\xE8\x80\x00\x80\xBE\x64\x0C\x00\xEC\x80\x00\x80\xF0\x80\x00\x80\xF8\x80"
-     "\x00\x80"
-     "\xB8\x45\x01\x00\xB9\x4D\x01\x00\xBA\x45\x01\x00\xBB\x5D\x01\x00\xBC\x4D"
-     "\x01\x00"
-     "\xBD\x7D\x01\x00\xBE\x75\x01\x00\xBF\x1D\x01\x00\xB0\x0D\x02\x00\xB1\xC1"
-     "\x02\x00"
-     "\xB2\xC1\x02\x00\xB3\xC1\x02\x00\xB4\xC1\x02\x00\xB5\xC1\x02\x00\xB6\xC1"
-     "\x02\x00"
-     "\xB7\xC1\x02\x00\xE1\x58\x06\x00\xE1\x5C\x07\x00\xE3\x9C\x00\x00\xE3\xBC"
-     "\x06\x00"
-     "\x86\x30\x0C\x00\x87\x5C\x0D\x00\xFC\x80\x00\x80\x00\x81\x00\x80\x04\x81"
-     "\x00\x80"
-     "\x08\x81\x00\x80\x0C\x81\x00\x80\x10\x81\x00\x80\xEE\x00\x00\x00\x14\x81"
-     "\x00\x80"
-     "\xEF\x48\x00\x00\xEF\xE4\x06\x00\x18\x81\x00\x80\x1C\x81\x00\x80\xA3\x89"
-     "\x02\x00"
-     "\x20\x81\x00\x80\xA5\x9D\x02\x00\x24\x81\x00\x80\x28\x81\x00\x80\xA6\x89"
-     "\x02\x00"
-     "\x2C\x81\x00\x80\x30\x81\x00\x80\xAB\x4D\x02\x00\xAA\x81\x02\x00\xAD\x5D"
-     "\x02\x00"
-     "\xAC\x5D\x02\x00\xAF\x4D\x02\x00\xAE\x4D\x02\x00\xA8\x59\x0E\x00\xA9\x61"
-     "\x0E\x00"
-     "\xAA\x79\x0E\x00\xAB\x79\x0E\x00\xAC\x69\x0E\x00\xAD\x69\x0E\x00\xAE\x9D"
-     "\x0E\x00"
-     "\xAF\x8D\x0E\x00\xF4\x80\x00\x80\x34\x81\x00\x80\x38\x81\x00\x80\x3C\x81"
-     "\x00\x80"
-     "\x40\x81\x00\x80\x44\x81\x00\x80\x48\x81\x00\x80\x4C\x81\x00\x80\xB8\xFD"
-     "\x0E\x00"
-     "\xB9\x81\x0E\x00\xBA\x81\x0E\x00\xBB\x81\x0E\x00\xBC\x81\x0E\x00\xBD\x89"
-     "\x0E\x00"
-     "\xBE\xB1\x0E\x00\xBF\xB1\x0E\x00\xB0\xF5\x0E\x00\xB1\xFD\x0E\x00\xB2\xF5"
-     "\x0E\x00"
-     "\xB3\xC9\x0E\x00\xB4\xD9\x0E\x00\xB5\xD9\x0E\x00\xB6\xCD\x0E\x00\xB7\xC5"
-     "\x0E\x00"
-     "\xB3\xF1\x0E\x00\x50\x81\x00\x80\x54\x81\x00\x80\x58\x81\x00\x80\x5C\x81"
-     "\x00\x80"
-     "\xB6\xE9\x0E\x00\xB5\xE1\x0E\x00\x60\x81\x00\x80\xBB\x39\x0E\x00\xBA\x31"
-     "\x0E\x00"
-     "\x84\xE4\x03\x00\x64\x81\x00\x80\xBF\x1D\x0E\x00\xBE\x1D\x0E\x00\xBD\x1D"
-     "\x0E\x00"
-     "\xBC\x21\x0E\x00\x82\x5D\x00\x00\xA3\xB5\x0E\x00\x80\x45\x00\x00\x81\x5D"
-     "\x00\x00"
-     "\xA6\xAD\x0E\x00\x68\x81\x00\x80\x6C\x81\x00\x80\xA5\xA5\x0E\x00\xAA\x75"
-     "\x0E\x00"
-     "\xAB\x7D\x0E\x00\x86\xE8\x00\x00\x87\xE0\x00\x00\xAE\x59\x0E\x00\xAF\x59"
-     "\x0E\x00"
-     "\xAC\x65\x0E\x00\xAD\x59\x0E\x00\xA8\x3D\x0E\x00\xA9\x9D\x0E\x00\xAA\x95"
-     "\x0E\x00"
-     "\xAB\xAD\x0E\x00\xAC\xB5\x0E\x00\xAD\xBD\x0E\x00\xAE\xB5\x0E\x00\xAF\xA9"
-     "\x0E\x00"
-     "\x70\x81\x00\x80\x74\x81\x00\x80\x78\x81\x00\x80\x7C\x81\x00\x80\x80\x81"
-     "\x00\x80"
-     "\x84\x81\x00\x80\x88\x81\x00\x80\x8C\x81\x00\x80\xB8\x69\x01\x00\xB9\x69"
-     "\x01\x00"
-     "\xBA\x79\x01\x00\xBB\x79\x01\x00\xBC\x69\x01\x00\xBD\x69\x01\x00\xBE\xDD"
-     "\x01\x00"
-     "\xBF\xD5\x01\x00\xB0\xA1\x0E\x00\xB1\xA1\x0E\x00\xB2\xA5\x0E\x00\xB3\xB9"
-     "\x0E\x00"
-     "\xB4\xA9\x0E\x00\xB5\x95\x0E\x00\xB6\x9D\x0E\x00\xB7\x59\x01\x00\xB3\x1D"
-     "\x0E\x00"
-     "\x90\x81\x00\x80\x94\x81\x00\x80\x98\x81\x00\x80\x9C\x81\x00\x80\xB6\x3D"
-     "\x0E\x00"
-     "\xB5\x3D\x0E\x00\xA0\x81\x00\x80\xBB\x65\x0E\x00\xBA\x65\x0E\x00\xA4\x81"
-     "\x00\x80"
-     "\xA8\x81\x00\x80\xBF\x51\x0E\x00\xBE\x6D\x0E\x00\xBD\x6D\x0E\x00\xBC\x71"
-     "\x0E\x00"
-     "\xA5\x65\x0E\x00\xAC\x81\x00\x80\xB0\x81\x00\x80\xA6\x65\x0E\x00\xB4\x81"
-     "\x00\x80"
-     "\xB8\x81\x00\x80\xA3\x45\x0E\x00\xBC\x81\x00\x80\xAD\x35\x0E\x00\xAC\x29"
-     "\x0E\x00"
-     "\xAF\x09\x0E\x00\xAE\x35\x0E\x00\xC0\x81\x00\x80\x84\x80\x00\x80\xAB\x3D"
-     "\x0E\x00"
-     "\xAA\x3D\x0E\x00\xC4\x81\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xC8\x81\x00\x80\xCC\x81\x00\x80\xD0\x81\x00\x80\xBE\x50\x03\x00\xD8\x81"
-     "\x00\x80"
-     "\xDC\x81\x00\x80\xE0\x81\x00\x80\xE4\x81\x00\x80\xE8\x81\x00\x80\xEC\x81"
-     "\x00\x80"
-     "\xF0\x81\x00\x80\xF4\x81\x00\x80\x84\xC0\x03\x00\xF8\x81\x00\x80\xEF\x10"
-     "\x0E\x00"
-     "\x84\x00\x04\x00\x86\xF0\x04\x00\x87\x54\x03\x00\xFC\x81\x00\x80\x00\x82"
-     "\x00\x80"
-     "\x04\x82\x00\x80\x08\x82\x00\x80\x0C\x82\x00\x80\x10\x82\x00\x80\xE1\x00"
-     "\x0E\x00"
-     "\x14\x82\x00\x80\xE3\xE4\x0E\x00\x18\x82\x00\x80\x1C\x82\x00\x80\x20\x82"
-     "\x00\x80"
-     "\x24\x82\x00\x80\x28\x82\x00\x80\x2C\x82\x00\x80\x30\x82\x00\x80\x34\x82"
-     "\x00\x80"
-     "\x38\x82\x00\x80\xA8\x75\x02\x00\xA9\x7D\x02\x00\xAA\x89\x03\x00\xAB\x89"
-     "\x03\x00"
-     "\xAC\x99\x03\x00\xAD\x99\x03\x00\xAE\x89\x03\x00\xAF\x89\x03\x00\xB0\xF9"
-     "\x03\x00"
-     "\xB1\xF9\x03\x00\xB2\x4D\x03\x00\xB3\x45\x03\x00\xB4\x5D\x03\x00\xB5\x45"
-     "\x03\x00"
-     "\xB6\x45\x03\x00\xB7\x75\x03\x00\xB8\x4D\x03\x00\xB9\x21\x03\x00\xBA\x21"
-     "\x03\x00"
-     "\xBB\x3D\x03\x00\xBC\x25\x03\x00\xBD\x29\x03\x00\xBE\x09\x03\x00\xBF\xFD"
-     "\x00\x00"
-     "\x80\x9D\x01\x00\x81\x09\x00\x00\x82\x1D\x00\x00\x3C\x82\x00\x80\xEF\xF4"
-     "\x03\x00"
-     "\x40\x82\x00\x80\xBE\x50\x05\x00\x48\x82\x00\x80\xE3\x70\x03\x00\x4C\x82"
-     "\x00\x80"
-     "\xE1\xD8\x01\x00\x50\x82\x00\x80\x54\x82\x00\x80\x58\x82\x00\x80\x5C\x82"
-     "\x00\x80"
-     "\x60\x82\x00\x80\x86\xE0\x04\x00\x87\x88\x02\x00\x64\x82\x00\x80\xB3\x6D"
-     "\x02\x00"
-     "\x68\x82\x00\x80\xB5\x7D\x02\x00\xB6\x75\x02\x00\x6C\x82\x00\x80\x70\x82"
-     "\x00\x80"
-     "\x74\x82\x00\x80\xBA\x51\x02\x00\xBB\x51\x02\x00\xBC\xF5\x02\x00\xBD\xFD"
-     "\x02\x00"
-     "\xBE\xF5\x02\x00\xBF\xED\x02\x00\x78\x82\x00\x80\x7C\x82\x00\x80\x80\x82"
-     "\x00\x80"
-     "\x84\x82\x00\x80\x88\x82\x00\x80\x8C\x82\x00\x80\x90\x82\x00\x80\xEF\x54"
-     "\x0E\x00"
-     "\x84\xEC\x04\x00\xE1\x2C\x0E\x00\x94\x82\x00\x80\xE3\x14\x01\x00\x98\x82"
-     "\x00\x80"
-     "\x9C\x82\x00\x80\xA0\x82\x00\x80\xA4\x82\x00\x80\xA3\xE1\x02\x00\x82\x2D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x80\x1D\x00\x00\xA8\x82\x00\x80\xA6\xF9\x02\x00\xA5\xF1"
-     "\x02\x00"
-     "\xAC\x82\x00\x80\xAB\xDD\x02\x00\xAA\xDD\x02\x00\xB0\x82\x00\x80\xBE\xB0"
-     "\x04\x00"
-     "\xAF\x61\x02\x00\xAE\x79\x02\x00\xAD\x71\x02\x00\xAC\x79\x02\x00\xA8\x21"
-     "\x06\x00"
-     "\xA9\x21\x06\x00\xAA\x21\x06\x00\xAB\x21\x06\x00\xAC\x21\x06\x00\xAD\x21"
-     "\x06\x00"
-     "\xAE\x21\x06\x00\xAF\x21\x06\x00\x44\x82\x00\x80\xB4\x82\x00\x80\x86\x00"
-     "\x00\x00"
-     "\x87\xA0\x03\x00\xB8\x82\x00\x80\xBC\x82\x00\x80\xC0\x82\x00\x80\xC4\x82"
-     "\x00\x80"
-     "\xB8\x3D\x06\x00\xB9\xC5\x06\x00\xBA\xCD\x06\x00\xBB\xC5\x06\x00\xBC\xDD"
-     "\x06\x00"
-     "\xBD\xC5\x06\x00\xBE\xCD\x06\x00\xBF\xC5\x06\x00\xB0\x25\x06\x00\xB1\x3D"
-     "\x06\x00"
-     "\xB2\x35\x06\x00\xB3\x0D\x06\x00\xB4\x15\x06\x00\xB5\x1D\x06\x00\xB6\x15"
-     "\x06\x00"
-     "\xB7\x09\x06\x00\xB3\xA9\x06\x00\xC8\x82\x00\x80\xCC\x82\x00\x80\xD0\x82"
-     "\x00\x80"
-     "\xD4\x82\x00\x80\xB6\xA1\x06\x00\xB5\xA9\x06\x00\xD8\x82\x00\x80\xBB\xF9"
-     "\x06\x00"
-     "\xBA\xE1\x06\x00\xDC\x82\x00\x80\xE0\x82\x00\x80\xBF\xD9\x06\x00\xBE\xD1"
-     "\x06\x00"
-     "\xBD\xD9\x06\x00\xBC\xE1\x06\x00\xE4\x82\x00\x80\xA3\xED\x06\x00\xE8\x82"
-     "\x00\x80"
-     "\xEC\x82\x00\x80\xA6\xE5\x06\x00\xF0\x82\x00\x80\xF4\x82\x00\x80\xA5\xED"
-     "\x06\x00"
-     "\xAA\xA5\x06\x00\xAB\xBD\x06\x00\xF8\x82\x00\x80\xFC\x82\x00\x80\xAE\x95"
-     "\x06\x00"
-     "\xAF\x9D\x06\x00\xAC\xA5\x06\x00\xAD\x9D\x06\x00\xA8\x1D\x06\x00\xA9\x21"
-     "\x06\x00"
-     "\xAA\x25\x06\x00\xAB\x3D\x06\x00\xAC\x25\x06\x00\xAD\x2D\x06\x00\xAE\x25"
-     "\x06\x00"
-     "\xAF\x55\x06\x00\x00\x83\x00\x80\x04\x83\x00\x80\x08\x83\x00\x80\x0C\x83"
-     "\x00\x80"
-     "\x10\x83\x00\x80\x14\x83\x00\x80\x18\x83\x00\x80\x1C\x83\x00\x80\xB8\xD1"
-     "\x06\x00"
-     "\xB9\xD1\x06\x00\xBA\xD1\x06\x00\xBB\xE5\x06\x00\xBC\xE1\x06\x00\xBD\xE1"
-     "\x06\x00"
-     "\xBE\xE1\x06\x00\xBF\xE1\x06\x00\xB0\x31\x06\x00\xB1\x39\x06\x00\xB2\x09"
-     "\x06\x00"
-     "\xB3\x09\x06\x00\xB4\xF1\x06\x00\xB5\xF1\x06\x00\xB6\xF1\x06\x00\xB7\xF1"
-     "\x06\x00"
-     "\xB3\xAD\x06\x00\x82\x15\x00\x00\x81\xB5\x00\x00\x80\xB5\x00\x00\x20\x83"
-     "\x00\x80"
-     "\xB6\xAD\x06\x00\xB5\xAD\x06\x00\xBE\x10\x00\x00\xBB\xFD\x06\x00\xBA\xF5"
-     "\x06\x00"
-     "\x84\x6C\x01\x00\x28\x83\x00\x80\xBF\x35\x06\x00\xBE\x35\x06\x00\xBD\xDD"
-     "\x06\x00"
-     "\xBC\xDD\x06\x00\x2C\x83\x00\x80\xA6\xFD\x06\x00\xA5\xFD\x06\x00\x24\x83"
-     "\x00\x80"
-     "\xA3\xFD\x06\x00\x30\x83\x00\x80\x86\x00\xFC\x01\x87\x60\x01\x00\xAF\x65"
-     "\x06\x00"
-     "\xAE\x65\x06\x00\xAD\x8D\x06\x00\xAC\x8D\x06\x00\xAB\xAD\x06\x00\xAA\xA5"
-     "\x06\x00"
-     "\xD4\x81\x00\x80\x34\x83\x00\x80\x38\x83\x00\x80\x3C\x83\x00\x80\xB3\x31"
-     "\x02\x00"
-     "\x40\x83\x00\x80\xB5\x95\x02\x00\x44\x83\x00\x80\x48\x83\x00\x80\xB6\x95"
-     "\x02\x00"
-     "\x4C\x83\x00\x80\x50\x83\x00\x80\xBB\x81\x02\x00\xBA\xB1\x02\x00\xBD\xF1"
-     "\x02\x00"
-     "\xBC\x99\x02\x00\xBF\xF5\x02\x00\xBE\xF5\x02\x00\xE3\x1C\x02\x00\xEF\xA4"
-     "\x07\x00"
-     "\xE1\x88\x01\x00\x84\xC4\x03\x00\x54\x83\x00\x80\x58\x83\x00\x80\x5C\x83"
-     "\x00\x80"
-     "\x60\x83\x00\x80\x64\x83\x00\x80\x68\x83\x00\x80\x6C\x83\x00\x80\x70\x83"
-     "\x00\x80"
-     "\xEF\x28\x02\x00\xE3\x28\x06\x00\x74\x83\x00\x80\xE1\x34\x06\x00\xAA\xB9"
-     "\x02\x00"
-     "\xAB\xB9\x02\x00\xA8\xB9\x02\x00\xA9\xB9\x02\x00\xAE\x1D\x02\x00\xAF\x05"
-     "\x02\x00"
-     "\xAC\x1D\x02\x00\xAD\x15\x02\x00\xBE\xEC\x03\x00\x78\x83\x00\x80\x7C\x83"
-     "\x00\x80"
-     "\x80\x83\x00\x80\x84\x83\x00\x80\x82\xB9\x01\x00\x81\xB1\x01\x00\x80\xB9"
-     "\x01\x00"
-     "\xBA\xE1\x03\x00\xBB\xE1\x03\x00\xB8\x21\x02\x00\xB9\x21\x02\x00\xBE\xE1"
-     "\x03\x00"
-     "\xBF\xE1\x03\x00\xBC\xF1\x03\x00\xBD\xF1\x03\x00\xB2\x4D\x02\x00\xB3\x25"
-     "\x02\x00"
-     "\xB0\x75\x02\x00\xB1\x45\x02\x00\xB6\x25\x02\x00\xB7\x2D\x02\x00\xB4\x31"
-     "\x02\x00"
-     "\xB5\x31\x02\x00\xA3\x39\x03\x00\x88\x83\x00\x80\x8C\x83\x00\x80\x90\x83"
-     "\x00\x80"
-     "\x98\x83\x00\x80\xA6\x9D\x03\x00\xA5\x9D\x03\x00\x9C\x83\x00\x80\xAB\x89"
-     "\x03\x00"
-     "\xAA\xB9\x03\x00\xA0\x83\x00\x80\xA4\x83\x00\x80\xAF\xFD\x03\x00\xAE\xFD"
-     "\x03\x00"
-     "\xAD\xF9\x03\x00\xAC\x91\x03\x00\xA8\x83\x00\x80\xB3\x91\x02\x00\x86\xC8"
-     "\xFD\x01"
-     "\x87\x40\xFD\x01\xB6\x55\x01\x00\xAC\x83\x00\x80\xB0\x83\x00\x80\xB5\x5D"
-     "\x01\x00"
-     "\xBA\x79\x01\x00\xBB\x45\x01\x00\xB4\x83\x00\x80\xB8\x83\x00\x80\xBE\x25"
-     "\x01\x00"
-     "\xBF\x2D\x01\x00\xBC\x5D\x01\x00\xBD\x2D\x01\x00\xBC\x83\x00\x80\xC0\x83"
-     "\x00\x80"
-     "\xC4\x83\x00\x80\x84\xC0\xFD\x01\xC8\x83\x00\x80\xCC\x83\x00\x80\xEF\xFC"
-     "\x00\x00"
-     "\xEF\x88\x06\x00\xE1\x90\x01\x00\xE1\xFC\x06\x00\xE3\x10\x00\x00\xE3\x1C"
-     "\x06\x00"
-     "\xD0\x83\x00\x80\xD4\x83\x00\x80\xBE\xF0\xFD\x01\xD8\x83\x00\x80\xA3\x15"
-     "\x02\x00"
-     "\xDC\x83\x00\x80\xE0\x83\x00\x80\xE4\x83\x00\x80\xE8\x83\x00\x80\xA6\xD1"
-     "\x01\x00"
-     "\xA5\xD9\x01\x00\xEC\x83\x00\x80\xAB\xC1\x01\x00\xAA\xFD\x01\x00\xF0\x83"
-     "\x00\x80"
-     "\xF4\x83\x00\x80\xAF\xA9\x01\x00\xAE\xA1\x01\x00\xAD\xA9\x01\x00\xAC\xD9"
-     "\x01\x00"
-     "\x94\x83\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\xF8\x83"
-     "\x00\x80"
-     "\xFC\x83\x00\x80\x00\x84\x00\x80\x04\x84\x00\x80\xA8\x31\xFF\x01\xA9\x39"
-     "\xFF\x01"
-     "\xAA\x01\xFF\x01\xAB\x01\xFF\x01\xAC\x71\xFF\x01\xAD\x61\xFF\x01\xAE\x61"
-     "\xFF\x01"
-     "\xAF\x55\xFF\x01\xB0\x2D\xFF\x01\xB1\xC5\xFF\x01\xB2\xCD\xFF\x01\xB3\xC5"
-     "\xFF\x01"
-     "\xB4\xC1\xFF\x01\xB5\xC1\xFF\x01\xB6\xE9\xFF\x01\xB7\xE9\xFF\x01\xB8\xDD"
-     "\xFF\x01"
-     "\xB9\xA1\xFF\x01\xBA\xA1\xFF\x01\xBB\xA1\xFF\x01\xBC\xA1\xFF\x01\xBD\xA9"
-     "\xFF\x01"
-     "\xBE\x91\xFF\x01\xBF\x91\xFF\x01\xB3\xA9\xFF\x01\x08\x84\x00\x80\x86\x28"
-     "\x03\x00"
-     "\x87\x1C\x03\x00\x0C\x84\x00\x80\xB6\xC1\xFF\x01\xB5\xD9\xFF\x01\x10\x84"
-     "\x00\x80"
-     "\xBB\x21\xFF\x01\xBA\xC9\xFF\x01\x14\x84\x00\x80\x18\x84\x00\x80\xBF\x25"
-     "\xFF\x01"
-     "\xBE\x25\xFF\x01\xBD\x31\xFF\x01\xBC\x39\xFF\x01\x1C\x84\x00\x80\xA3\xED"
-     "\xFF\x01"
-     "\x20\x84\x00\x80\x24\x84\x00\x80\xA6\x85\xFF\x01\x28\x84\x00\x80\x2C\x84"
-     "\x00\x80"
-     "\xA5\x9D\xFF\x01\xAA\x8D\xFF\x01\xAB\x65\xFF\x01\x30\x84\x00\x80\x34\x84"
-     "\x00\x80"
-     "\xAE\x61\xFF\x01\xAF\x61\xFF\x01\xAC\x7D\xFF\x01\xAD\x75\xFF\x01\xA8\x85"
-     "\x01\x00"
-     "\xA9\x95\x01\x00\xAA\x95\x01\x00\xAB\xA5\x01\x00\xAC\xBD\x01\x00\xAD\xD5"
-     "\x01\x00"
-     "\xAE\xD1\x01\x00\xAF\xD1\x01\x00\x38\x84\x00\x80\x3C\x84\x00\x80\x40\x84"
-     "\x00\x80"
-     "\x44\x84\x00\x80\x48\x84\x00\x80\x4C\x84\x00\x80\x50\x84\x00\x80\x54\x84"
-     "\x00\x80"
-     "\xB8\x71\x00\x00\xB9\x71\x00\x00\xBA\x71\x00\x00\xBB\x71\x00\x00\xBC\xD1"
-     "\x00\x00"
-     "\xBD\xD1\x00\x00\xBE\xD1\x00\x00\xBF\xD1\x00\x00\xB0\xB5\x01\x00\xB1\xBD"
-     "\x01\x00"
-     "\xB2\x81\x01\x00\xB3\x81\x01\x00\xB4\x51\x00\x00\xB5\x51\x00\x00\xB6\x51"
-     "\x00\x00"
-     "\xB7\x51\x00\x00\x58\x84\x00\x80\x5C\x84\x00\x80\x60\x84\x00\x80\x64\x84"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\x68\x84\x00\x80\x6C\x84"
-     "\x00\x80"
-     "\x70\x84\x00\x80\x84\xA4\x02\x00\xBE\x20\x01\x00\x86\x90\x01\x00\x87\x14"
-     "\x01\x00"
-     "\x78\x84\x00\x80\x7C\x84\x00\x80\xB3\x5D\x01\x00\x74\x84\x00\x80\x80\x84"
-     "\x00\x80"
-     "\x84\x84\x00\x80\x88\x84\x00\x80\xB6\x15\x01\x00\xB5\x75\x01\x00\x8C\x84"
-     "\x00\x80"
-     "\xBB\x39\x01\x00\xBA\x31\x01\x00\x90\x84\x00\x80\x94\x84\x00\x80\xBF\xD5"
-     "\x00\x00"
-     "\xBE\xD5\x00\x00\xBD\x15\x01\x00\xBC\x15\x01\x00\xA3\x9D\x02\x00\x98\x84"
-     "\x00\x80"
-     "\x9C\x84\x00\x80\xBE\x00\x04\x00\xA0\x84\x00\x80\xA6\xD5\x02\x00\xA5\xB5"
-     "\x02\x00"
-     "\xA4\x84\x00\x80\xAB\xF9\x02\x00\xAA\xF1\x02\x00\xA8\x84\x00\x80\xAC\x84"
-     "\x00\x80"
-     "\xAF\x15\x03\x00\xAE\x15\x03\x00\xAD\xD5\x02\x00\xAC\xD5\x02\x00\xE1\xE8"
-     "\x01\x00"
-     "\xB0\x84\x00\x80\xE3\x14\x02\x00\x84\x40\x03\x00\xB4\x84\x00\x80\xB8\x84"
-     "\x00\x80"
-     "\xBC\x84\x00\x80\xC0\x84\x00\x80\xC4\x84\x00\x80\xC8\x84\x00\x80\xCC\x84"
-     "\x00\x80"
-     "\xD0\x84\x00\x80\xD4\x84\x00\x80\xD8\x84\x00\x80\xEF\x8C\x02\x00\xDC\x84"
-     "\x00\x80"
-     "\xA8\x39\x03\x00\xA9\x39\x03\x00\xAA\x41\x03\x00\xAB\x41\x03\x00\xAC\x41"
-     "\x03\x00"
-     "\xAD\x49\x03\x00\xAE\x71\x03\x00\xAF\x71\x03\x00\x80\xD5\x00\x00\x81\xD9"
-     "\x00\x00"
-     "\x82\x05\x00\x00\xE0\x84\x00\x80\x85\xB4\x05\x00\xE8\x84\x00\x80\xEC\x84"
-     "\x00\x80"
-     "\xF0\x84\x00\x80\xB8\x95\x00\x00\xB9\x9D\x00\x00\xBA\x95\x00\x00\xBB\xAD"
-     "\x00\x00"
-     "\xBC\xB5\x00\x00\xBD\xBD\x00\x00\xBE\xB5\x00\x00\xBF\xAD\x00\x00\xB0\xE5"
-     "\x00\x00"
-     "\xB1\xED\x00\x00\xB2\xF9\x00\x00\xB3\xF1\x00\x00\xB4\xD9\x00\x00\xB5\xD9"
-     "\x00\x00"
-     "\xB6\xBD\x00\x00\xB7\xAD\x00\x00\xB6\x5D\xFD\x01\xB7\x65\xFD\x01\xB4\x5D"
-     "\xFD\x01"
-     "\xB5\x55\xFD\x01\xB2\x35\xFD\x01\xB3\x45\xFD\x01\xB0\x01\xFD\x01\xB1\x3D"
-     "\xFD\x01"
-     "\xBE\x8D\xFD\x01\xBF\x95\xFD\x01\xBC\x8D\xFD\x01\xBD\x85\xFD\x01\xBA\x5D"
-     "\xFD\x01"
-     "\xBB\x4D\xFD\x01\xB8\x5D\xFD\x01\xB9\x55\xFD\x01\x86\xE0\x07\x00\x87\x18"
-     "\x05\x00"
-     "\xF4\x84\x00\x80\xF8\x84\x00\x80\xFC\x84\x00\x80\x84\x14\x07\x00\x00\x85"
-     "\x00\x80"
-     "\x04\x85\x00\x80\xAE\x6D\xFD\x01\xAF\x71\xFD\x01\xAC\x69\xFD\x01\xAD\x69"
-     "\xFD\x01"
-     "\xAA\xA5\xFE\x01\xAB\x89\xFE\x01\xA8\xD5\xFE\x01\xA9\xBD\xFE\x01\xE1\x24"
-     "\x03\x00"
-     "\xE1\xD4\xFE\x01\xE3\xDC\x1F\x00\xE3\x10\xFE\x01\xE3\x28\x00\x00\x08\x85"
-     "\x00\x80"
-     "\xE1\x08\x01\x00\x0C\x85\x00\x80\xEF\xF4\x00\x00\x10\x85\x00\x80\xBE\xF4"
-     "\x07\x00"
-     "\x14\x85\x00\x80\x18\x85\x00\x80\x1C\x85\x00\x80\xEF\x7C\x1F\x00\xEF\xFC"
-     "\xFE\x01"
-     "\xB3\x09\x02\x00\x20\x85\x00\x80\x24\x85\x00\x80\x28\x85\x00\x80\x2C\x85"
-     "\x00\x80"
-     "\xB6\x09\x02\x00\xB5\x09\x02\x00\x30\x85\x00\x80\xBB\xE9\x01\x00\xBA\xE1"
-     "\x01\x00"
-     "\x34\x85\x00\x80\x38\x85\x00\x80\xBF\xDD\x01\x00\xBE\xDD\x01\x00\xBD\xE5"
-     "\x01\x00"
-     "\xBC\xED\x01\x00\x9F\x25\xF8\x01\x9E\x25\xF8\x01\x9D\xF1\xF9\x01\x9C\xA1"
-     "\xFC\x01"
-     "\x9B\x31\xFC\x01\x9A\xE1\xFD\x01\x99\x15\xF1\x01\x98\x19\xF0\x01\x97\x21"
-     "\xF0\x01"
-     "\x96\x4D\xF5\x01\x95\x25\xF4\x01\x94\x2D\xF4\x01\x93\x61\xE9\x01\x92\x3D"
-     "\xE8\x01"
-     "\x91\x3D\xE8\x01\x90\xBD\xEE\x01\x8F\x61\xEC\x01\xE4\x84\x00\x80\x3C\x85"
-     "\x00\x80"
-     "\x40\x85\x00\x80\x44\x85\x00\x80\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D"
-     "\x00\x00"
-     "\x48\x85\x00\x80\x4C\x85\x00\x80\x50\x85\x00\x80\x54\x85\x00\x80\x86\xD0"
-     "\x01\x00"
-     "\x87\x00\x03\x00\x58\x85\x00\x80\x5C\x85\x00\x80\xA3\x49\x05\x00\x60\x85"
-     "\x00\x80"
-     "\x64\x85\x00\x80\x68\x85\x00\x80\x6C\x85\x00\x80\xA6\x49\x05\x00\xA5\x49"
-     "\x05\x00"
-     "\x70\x85\x00\x80\xAB\xA9\x06\x00\xAA\xA1\x06\x00\x74\x85\x00\x80\x78\x85"
-     "\x00\x80"
-     "\xAF\x9D\x06\x00\xAE\x9D\x06\x00\xAD\xA5\x06\x00\xAC\xAD\x06\x00\x7C\x85"
-     "\x00\x80"
-     "\x80\x85\x00\x80\x84\x85\x00\x80\x88\x85\x00\x80\x8C\x85\x00\x80\x90\x85"
-     "\x00\x80"
-     "\x94\x85\x00\x80\x98\x85\x00\x80\x9C\x85\x00\x80\xA0\x85\x00\x80\xA4\x85"
-     "\x00\x80"
-     "\xA8\x85\x00\x80\xA3\x35\x05\x00\xA2\xE9\x04\x00\xA1\xF9\x04\x00\xA0\x31"
-     "\xF9\x01"
-     "\x82\x21\xED\x01\x83\x85\xEF\x01\xAC\x85\x00\x80\xB0\x85\x00\x80\x86\x41"
-     "\xE8\x01"
-     "\x87\x15\xF5\x01\x84\x1D\xE9\x01\x85\x15\xE9\x01\x8A\x2D\xF5\x01\x8B\x4D"
-     "\xF4\x01"
-     "\xB4\x85\x00\x80\xB8\x85\x00\x80\x8E\x3D\xF0\x01\x8F\xC1\xFC\x01\x8C\x39"
-     "\xF1\x01"
-     "\x8D\x39\xF1\x01\x92\x61\xFD\x01\x93\xE9\xFD\x01\xBC\x85\x00\x80\xC0\x85"
-     "\x00\x80"
-     "\x96\x75\xF9\x01\x97\x99\xF8\x01\x94\xA1\xF8\x01\x95\x7D\xF9\x01\x9A\x61"
-     "\x05\x00"
-     "\x9B\x79\x05\x00\x84\x2C\x00\x00\xBE\x00\x0C\x00\x81\x39\x00\x00\x80\x59"
-     "\x00\x00"
-     "\x9C\xBD\x04\x00\x82\x31\x00\x00\xA1\xC5\x01\x00\xC8\x85\x00\x80\xA3\x89"
-     "\x0D\x00"
-     "\xA2\xE1\x01\x00\xA5\x01\x0D\x00\xA4\x91\x0D\x00\xA7\xB9\x09\x00\xA6\x01"
-     "\x08\x00"
-     "\xA9\x01\x14\x00\xA8\x0D\x09\x00\xAB\x99\x15\x00\xAA\x8D\x15\x00\xAD\xB1"
-     "\x11\x00"
-     "\xAC\x21\x14\x00\xAF\x01\x1C\x00\xAE\x15\x11\x00\xB1\xCD\x1D\x00\xB0\xC5"
-     "\x1D\x00"
-     "\xB3\x01\x18\x00\xB2\x55\x1C\x00\xB5\x39\x19\x00\xB4\xF5\x19\x00\x86\x58"
-     "\x0F\x00"
-     "\x87\xFC\x00\x00\xCC\x85\x00\x80\xD0\x85\x00\x80\xD4\x85\x00\x80\xD8\x85"
-     "\x00\x80"
-     "\xDC\x85\x00\x80\xE0\x85\x00\x80\xE4\x85\x00\x80\xE8\x85\x00\x80\xA8\xB5"
-     "\x04\x00"
-     "\xA9\xBD\x04\x00\xAA\xB5\x04\x00\xAB\x45\x05\x00\xAC\x5D\x05\x00\xAD\x45"
-     "\x05\x00"
-     "\xAE\x41\x05\x00\xAF\x41\x05\x00\xEC\x85\x00\x80\xF0\x85\x00\x80\xF4\x85"
-     "\x00\x80"
-     "\xF8\x85\x00\x80\xFC\x85\x00\x80\x00\x86\x00\x80\x04\x86\x00\x80\x08\x86"
-     "\x00\x80"
-     "\xB8\x8D\x05\x00\xB9\x95\x05\x00\xBA\x95\x05\x00\xBB\xA5\x05\x00\xBC\xBD"
-     "\x05\x00"
-     "\xBD\x1D\x02\x00\xBE\x15\x02\x00\xBF\x0D\x02\x00\xB0\xF5\x05\x00\xB1\xF9"
-     "\x05\x00"
-     "\xB2\xCD\x05\x00\xB3\xC1\x05\x00\xB4\xC1\x05\x00\xB5\xC1\x05\x00\xB6\xBD"
-     "\x05\x00"
-     "\xB7\xB5\x05\x00\x0C\x86\x00\x80\x10\x86\x00\x80\x14\x86\x00\x80\xB3\x31"
-     "\x04\x00"
-     "\x18\x86\x00\x80\xB5\xD9\x05\x00\xB6\xC9\x05\x00\x1C\x86\x00\x80\x20\x86"
-     "\x00\x80"
-     "\x24\x86\x00\x80\xBA\x3D\x05\x00\xBB\x31\x05\x00\xBC\x15\x05\x00\xBD\x19"
-     "\x05\x00"
-     "\xBE\x09\x05\x00\xBF\x09\x05\x00\x28\x86\x00\x80\xA6\x95\x05\x00\xA5\x85"
-     "\x05\x00"
-     "\x2C\x86\x00\x80\xA3\x6D\x04\x00\x30\x86\x00\x80\xBE\x90\x0C\x00\x38\x86"
-     "\x00\x80"
-     "\xAF\x55\x05\x00\xAE\x55\x05\x00\xAD\x45\x05\x00\xAC\x49\x05\x00\xAB\x6D"
-     "\x05\x00"
-     "\xAA\x61\x05\x00\x3C\x86\x00\x80\x40\x86\x00\x80\xA8\x79\x03\x00\xA9\x79"
-     "\x03\x00"
-     "\xAA\x8D\x03\x00\xAB\x85\x03\x00\xAC\x9D\x03\x00\xAD\x85\x03\x00\xAE\x85"
-     "\x03\x00"
-     "\xAF\xB5\x03\x00\x80\xE1\x01\x00\x81\xE9\x01\x00\x82\xFD\x01\x00\x84\x60"
-     "\x0D\x00"
-     "\x44\x86\x00\x80\x48\x86\x00\x80\x87\x24\x0D\x00\x86\x9C\x0C\x00\xB8\x9D"
-     "\x03\x00"
-     "\xB9\xA5\x03\x00\xBA\xAD\x03\x00\xBB\xA5\x03\x00\xBC\xBD\x03\x00\xBD\x5D"
-     "\x02\x00"
-     "\xBE\x55\x02\x00\xBF\x75\x02\x00\xB0\xCD\x03\x00\xB1\xA5\x03\x00\xB2\xAD"
-     "\x03\x00"
-     "\xB3\xA5\x03\x00\xB4\xBD\x03\x00\xB5\xA5\x03\x00\xB6\xAD\x03\x00\xB7\xA5"
-     "\x03\x00"
-     "\x4C\x86\x00\x80\x50\x86\x00\x80\xB3\x0D\x03\x00\x54\x86\x00\x80\xB5\x1D"
-     "\x03\x00"
-     "\xBE\xDC\x0C\x00\x58\x86\x00\x80\xB6\x0D\x03\x00\x5C\x86\x00\x80\x60\x86"
-     "\x00\x80"
-     "\xBB\x69\x03\x00\xBA\x01\x03\x00\xBD\x79\x03\x00\xBC\x71\x03\x00\xBF\x69"
-     "\x03\x00"
-     "\xBE\x71\x03\x00\xE1\x98\x06\x00\x64\x86\x00\x80\xE3\x8C\x06\x00\x68\x86"
-     "\x00\x80"
-     "\xEF\x44\x00\x00\x6C\x86\x00\x80\x70\x86\x00\x80\x74\x86\x00\x80\xE3\x98"
-     "\x00\x00"
-     "\x78\x86\x00\x80\xE1\x80\x01\x00\x7C\x86\x00\x80\x80\x86\x00\x80\x84\x86"
-     "\x00\x80"
-     "\xEF\xA8\x06\x00\x88\x86\x00\x80\xC4\x85\x00\x80\x34\x86\x00\x80\x8C\x86"
-     "\x00\x80"
-     "\x90\x86\x00\x80\x94\x86\x00\x80\x82\x99\x00\x00\x81\x91\x00\x00\x80\x99"
-     "\x00\x00"
-     "\xA8\xA5\x0D\x00\xA9\xC5\x0D\x00\xAA\xCD\x0D\x00\xAB\xC5\x0D\x00\xAC\xDD"
-     "\x0D\x00"
-     "\xAD\xC5\x0D\x00\xAE\xCD\x0D\x00\xAF\xC5\x0D\x00\xB0\xBD\x0D\x00\xB1\x65"
-     "\x0E\x00"
-     "\xB2\x6D\x0E\x00\xB3\x65\x0E\x00\xB4\x7D\x0E\x00\xB5\x65\x0E\x00\xB6\x6D"
-     "\x0E\x00"
-     "\xB7\x65\x0E\x00\xB8\x5D\x0E\x00\xB9\xE5\x0E\x00\xBA\xED\x0E\x00\xBB\xE5"
-     "\x0E\x00"
-     "\xBC\xE1\x0E\x00\xBD\xE9\x0E\x00\xBE\xD9\x0E\x00\xBF\xD9\x0E\x00\xB3\x65"
-     "\x0D\x00"
-     "\x98\x86\x00\x80\x9C\x86\x00\x80\xA0\x86\x00\x80\xA4\x86\x00\x80\xB6\x65"
-     "\x0D\x00"
-     "\xB5\x75\x0D\x00\x84\xD0\x03\x00\xBB\x81\x0E\x00\xBA\x59\x0D\x00\x86\x08"
-     "\x00\x00"
-     "\x87\x8C\x00\x00\xBF\x81\x0E\x00\xBE\x99\x0E\x00\xBD\x91\x0E\x00\xBC\x99"
-     "\x0E\x00"
-     "\xA8\x86\x00\x80\xA3\x21\x0D\x00\xAC\x86\x00\x80\xB0\x86\x00\x80\xA6\x21"
-     "\x0D\x00"
-     "\xB4\x86\x00\x80\xB8\x86\x00\x80\xA5\x31\x0D\x00\xAA\x1D\x0D\x00\xAB\xC5"
-     "\x0E\x00"
-     "\xBC\x86\x00\x80\xC0\x86\x00\x80\xAE\xDD\x0E\x00\xAF\xC5\x0E\x00\xAC\xDD"
-     "\x0E\x00"
-     "\xAD\xD5\x0E\x00\xA8\xDD\x0F\x00\xA9\xFD\x0F\x00\xAA\xF5\x0F\x00\xAB\x3D"
-     "\x0E\x00"
-     "\xAC\x25\x0E\x00\xAD\x95\x0E\x00\xAE\x9D\x0E\x00\xAF\x91\x0E\x00\xC4\x86"
-     "\x00\x80"
-     "\xC8\x86\x00\x80\xCC\x86\x00\x80\xD0\x86\x00\x80\xD4\x86\x00\x80\xD8\x86"
-     "\x00\x80"
-     "\xDC\x86\x00\x80\xE0\x86\x00\x80\xB8\xA5\x0E\x00\xB9\xA9\x0E\x00\xBA\xB9"
-     "\x0E\x00"
-     "\xBB\xB9\x0E\x00\xBC\xAD\x0E\x00\xBD\x5D\x01\x00\xBE\x55\x01\x00\xBF\x4D"
-     "\x01\x00"
-     "\xB0\xF9\x0E\x00\xB1\xC5\x0E\x00\xB2\xCD\x0E\x00\xB3\xA9\x0E\x00\xB4\xBD"
-     "\x0E\x00"
-     "\xB5\xA1\x0E\x00\xB6\xA1\x0E\x00\xB7\xA1\x0E\x00\xB3\x7D\x0F\x00\xE4\x86"
-     "\x00\x80"
-     "\xE8\x86\x00\x80\xEC\x86\x00\x80\xF0\x86\x00\x80\xB6\x15\x0E\x00\xB5\xAD"
-     "\x0E\x00"
-     "\xF4\x86\x00\x80\xBB\x01\x0E\x00\xBA\x3D\x0E\x00\xF8\x86\x00\x80\xFC\x86"
-     "\x00\x80"
-     "\xBF\x65\x0E\x00\xBE\x79\x0E\x00\xBD\x7D\x0E\x00\xBC\x19\x0E\x00\x00\x87"
-     "\x00\x80"
-     "\xA3\x39\x0F\x00\x04\x87\x00\x80\x08\x87\x00\x80\xA6\x51\x0E\x00\x0C\x87"
-     "\x00\x80"
-     "\x10\x87\x00\x80\xA5\xE9\x0E\x00\xAA\x79\x0E\x00\xAB\x45\x0E\x00\x84\xA4"
-     "\x01\x00"
-     "\xBE\xA4\x01\x00\xAE\x3D\x0E\x00\xAF\x21\x0E\x00\xAC\x5D\x0E\x00\xAD\x39"
-     "\x0E\x00"
-     "\xA8\xC1\x02\x00\xA9\xCD\x02\x00\xAA\xC5\x02\x00\xAB\xD9\x02\x00\xAC\xC9"
-     "\x02\x00"
-     "\xAD\xFD\x02\x00\xAE\xF5\x02\x00\xAF\x3D\x03\x00\x80\xED\x03\x00\x81\xF1"
-     "\x03\x00"
-     "\x82\xF1\x03\x00\x18\x87\x00\x80\x86\x90\x04\x00\x87\x74\x03\x00\x1C\x87"
-     "\x00\x80"
-     "\x20\x87\x00\x80\xB8\xC9\x03\x00\xB9\xC9\x03\x00\xBA\xD9\x03\x00\xBB\xD1"
-     "\x03\x00"
-     "\xBC\xF9\x03\x00\xBD\xF9\x03\x00\xBE\x99\x03\x00\xBF\x95\x03\x00\xB0\x45"
-     "\x03\x00"
-     "\xB1\x4D\x03\x00\xB2\x45\x03\x00\xB3\x5D\x03\x00\xB4\x45\x03\x00\xB5\x4D"
-     "\x03\x00"
-     "\xB6\x45\x03\x00\xB7\xF9\x03\x00\xB3\x4D\x02\x00\x24\x87\x00\x80\x28\x87"
-     "\x00\x80"
-     "\x84\x0C\x03\x00\x2C\x87\x00\x80\xB6\x7D\x02\x00\xB5\x41\x02\x00\x30\x87"
-     "\x00\x80"
-     "\xBB\x89\x03\x00\xBA\x81\x03\x00\x34\x87\x00\x80\x38\x87\x00\x80\xBF\x89"
-     "\x03\x00"
-     "\xBE\x81\x03\x00\xBD\x89\x03\x00\xBC\x91\x03\x00\x3C\x87\x00\x80\xA3\x09"
-     "\x02\x00"
-     "\x40\x87\x00\x80\x44\x87\x00\x80\xA6\x39\x02\x00\x48\x87\x00\x80\x4C\x87"
-     "\x00\x80"
-     "\xA5\x05\x02\x00\xAA\xC5\x03\x00\xAB\xCD\x03\x00\x50\x87\x00\x80\x54\x87"
-     "\x00\x80"
-     "\xAE\xC5\x03\x00\xAF\xCD\x03\x00\xAC\xD5\x03\x00\xAD\xCD\x03\x00\x58\x87"
-     "\x00\x80"
-     "\x5C\x87\x00\x80\x60\x87\x00\x80\x64\x87\x00\x80\x68\x87\x00\x80\xE1\xA4"
-     "\x01\x00"
-     "\x6C\x87\x00\x80\xE3\xE4\x00\x00\x70\x87\x00\x80\x74\x87\x00\x80\x78\x87"
-     "\x00\x80"
-     "\xEF\x38\x00\x00\x7C\x87\x00\x80\x80\x87\x00\x80\x88\x87\x00\x80\xBE\x30"
-     "\x05\x00"
-     "\x80\x69\x00\x00\x81\x75\x00\x00\x82\x7D\x00\x00\xB3\x71\x01\x00\x8C\x87"
-     "\x00\x80"
-     "\xB5\xDD\x01\x00\xB6\xC9\x01\x00\x90\x87\x00\x80\x86\xC0\x04\x00\x87\xE8"
-     "\x05\x00"
-     "\xBA\xC5\x01\x00\xBB\xAD\x01\x00\xBC\xB9\x01\x00\xBD\xB9\x01\x00\xBE\xAD"
-     "\x01\x00"
-     "\xBF\xA5\x01\x00\xE1\xA4\x0E\x00\xE1\x44\x0C\x00\xE3\x04\x01\x00\xE3\xB0"
-     "\x0E\x00"
-     "\x94\x87\x00\x80\x98\x87\x00\x80\x9C\x87\x00\x80\xA0\x87\x00\x80\x84\xA8"
-     "\x05\x00"
-     "\xA4\x87\x00\x80\xA8\x87\x00\x80\xAC\x87\x00\x80\xB0\x87\x00\x80\xB4\x87"
-     "\x00\x80"
-     "\xEF\xD4\x0E\x00\xEF\x9C\x0F\x00\xAB\x01\x0F\x00\xAA\x69\x0F\x00\xB8\x87"
-     "\x00\x80"
-     "\xBC\x87\x00\x80\xAF\x01\x0F\x00\xAE\x19\x0F\x00\xAD\x11\x0F\x00\xAC\x19"
-     "\x0F\x00"
-     "\xA3\x65\x0F\x00\xC0\x87\x00\x80\xC4\x87\x00\x80\xC8\x87\x00\x80\xCC\x87"
-     "\x00\x80"
-     "\xA6\x65\x0F\x00\xA5\x75\x0F\x00\xD0\x87\x00\x80\xA8\x2D\x06\x00\xA9\x3D"
-     "\x06\x00"
-     "\xAA\x35\x06\x00\xAB\x89\x06\x00\xAC\x99\x06\x00\xAD\x85\x06\x00\xAE\x85"
-     "\x06\x00"
-     "\xAF\xB5\x06\x00\x84\x87\x00\x80\xD4\x87\x00\x80\xD8\x87\x00\x80\xDC\x87"
-     "\x00\x80"
-     "\xE0\x87\x00\x80\xE4\x87\x00\x80\xE8\x87\x00\x80\xEC\x87\x00\x80\xB8\x41"
-     "\x07\x00"
-     "\xB9\x41\x07\x00\xBA\x41\x07\x00\xBB\x41\x07\x00\xBC\x41\x07\x00\xBD\x4D"
-     "\x07\x00"
-     "\xBE\x79\x07\x00\xBF\x79\x07\x00\xB0\xCD\x06\x00\xB1\xA5\x06\x00\xB2\xA1"
-     "\x06\x00"
-     "\xB3\xA1\x06\x00\xB4\xA5\x06\x00\xB5\xAD\x06\x00\xB6\x91\x06\x00\xB7\x91"
-     "\x06\x00"
-     "\xA3\xBD\x06\x00\x82\x15\x00\x00\x81\xB5\x00\x00\x80\xB5\x00\x00\xF0\x87"
-     "\x00\x80"
-     "\xA6\x05\x06\x00\xA5\x11\x06\x00\xBE\x90\x03\x00\xAB\x61\x06\x00\xAA\x09"
-     "\x06\x00"
-     "\xF4\x87\x00\x80\xF8\x87\x00\x80\xAF\x69\x06\x00\xAE\x61\x06\x00\xAD\x75"
-     "\x06\x00"
-     "\xAC\x75\x06\x00\xFC\x87\x00\x80\xB3\x99\x07\x00\x86\xC8\x00\x00\x87\x1C"
-     "\x03\x00"
-     "\xB6\x49\x07\x00\x00\x88\x00\x80\x04\x88\x00\x80\xB5\x49\x07\x00\xBA\x49"
-     "\x07\x00"
-     "\xBB\x49\x07\x00\x08\x88\x00\x80\x0C\x88\x00\x80\xBE\x39\x07\x00\xBF\x39"
-     "\x07\x00"
-     "\xBC\x49\x07\x00\xBD\x49\x07\x00\xA8\x39\x06\x00\xA9\x55\x06\x00\xAA\x55"
-     "\x06\x00"
-     "\xAB\x9D\x06\x00\xAC\x85\x06\x00\xAD\x8D\x06\x00\xAE\x85\x06\x00\xAF\xB9"
-     "\x06\x00"
-     "\x10\x88\x00\x80\x14\x88\x00\x80\x18\x88\x00\x80\x1C\x88\x00\x80\x20\x88"
-     "\x00\x80"
-     "\x24\x88\x00\x80\x28\x88\x00\x80\x2C\x88\x00\x80\xB8\xAD\x06\x00\xB9\xB5"
-     "\x06\x00"
-     "\xBA\xB5\x06\x00\xBB\x69\x01\x00\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x69"
-     "\x01\x00"
-     "\xBF\x69\x01\x00\xB0\xC9\x06\x00\xB1\xC9\x06\x00\xB2\xD9\x06\x00\xB3\xD9"
-     "\x06\x00"
-     "\xB4\xB9\x06\x00\xB5\xB9\x06\x00\xB6\x9D\x06\x00\xB7\x95\x06\x00\xA3\xDD"
-     "\x06\x00"
-     "\x30\x88\x00\x80\x34\x88\x00\x80\x38\x88\x00\x80\x3C\x88\x00\x80\xA6\x0D"
-     "\x06\x00"
-     "\xA5\x0D\x06\x00\x14\x87\x00\x80\xAB\x0D\x06\x00\xAA\x0D\x06\x00\x40\x88"
-     "\x00\x80"
-     "\x44\x88\x00\x80\xAF\x7D\x06\x00\xAE\x7D\x06\x00\xAD\x0D\x06\x00\xAC\x0D"
-     "\x06\x00"
-     "\x48\x88\x00\x80\x4C\x88\x00\x80\x50\x88\x00\x80\x54\x88\x00\x80\x80\x0D"
-     "\x00\x00"
-     "\x81\x35\x00\x00\x82\x3D\x00\x00\x58\x88\x00\x80\x5C\x88\x00\x80\x60\x88"
-     "\x00\x80"
-     "\x84\x64\x01\x00\xBE\x34\x00\x00\x86\x50\x1F\x00\x87\x68\x01\x00\x68\x88"
-     "\x00\x80"
-     "\x6C\x88\x00\x80\x70\x88\x00\x80\x74\x88\x00\x80\x84\x04\x03\x00\xBE\x00"
-     "\x1C\x00"
-     "\xE3\x8C\x06\x00\x78\x88\x00\x80\xE1\x30\x06\x00\x7C\x88\x00\x80\xEF\xE4"
-     "\x07\x00"
-     "\x80\x88\x00\x80\x84\x88\x00\x80\x88\x88\x00\x80\x8C\x88\x00\x80\x90\x88"
-     "\x00\x80"
-     "\x94\x88\x00\x80\x98\x88\x00\x80\xB3\x95\x02\x00\x9C\x88\x00\x80\xA0\x88"
-     "\x00\x80"
-     "\xA4\x88\x00\x80\xA8\x88\x00\x80\xB6\xB9\x02\x00\xB5\xB1\x02\x00\xAC\x88"
-     "\x00\x80"
-     "\xBB\x4D\x03\x00\xBA\x4D\x03\x00\xB0\x88\x00\x80\xB4\x88\x00\x80\xBF\x4D"
-     "\x03\x00"
-     "\xBE\x4D\x03\x00\xBD\x4D\x03\x00\xBC\x4D\x03\x00\xB8\x88\x00\x80\xBC\x88"
-     "\x00\x80"
-     "\xC0\x88\x00\x80\xBE\x88\x1D\x00\xC4\x88\x00\x80\x82\x9D\x01\x00\x81\x9D"
-     "\x01\x00"
-     "\x80\x9D\x01\x00\xA8\x0D\x02\x00\xA9\x65\x02\x00\xAA\x7D\x02\x00\xAB\x75"
-     "\x02\x00"
-     "\xAC\x59\x02\x00\xAD\x59\x02\x00\xAE\x99\x03\x00\xAF\x99\x03\x00\xB0\xE9"
-     "\x03\x00"
-     "\xB1\xE9\x03\x00\xB2\xF9\x03\x00\xB3\xF9\x03\x00\xB4\xE9\x03\x00\xB5\xE9"
-     "\x03\x00"
-     "\xB6\x5D\x03\x00\xB7\x55\x03\x00\xB8\x6D\x03\x00\xB9\x75\x03\x00\xBA\x41"
-     "\x03\x00"
-     "\xBB\x41\x03\x00\xBC\x31\x03\x00\xBD\x31\x03\x00\xBE\x31\x03\x00\xBF\x31"
-     "\x03\x00"
-     "\xC8\x88\x00\x80\xCC\x88\x00\x80\xD0\x88\x00\x80\x84\x04\x1D\x00\xD8\x88"
-     "\x00\x80"
-     "\xEF\x00\x00\x00\x86\xB8\x1C\x00\x87\x30\x1D\x00\xDC\x88\x00\x80\xE3\xDC"
-     "\x00\x00"
-     "\xE0\x88\x00\x80\xE1\xFC\x01\x00\xE4\x88\x00\x80\xE8\x88\x00\x80\xEC\x88"
-     "\x00\x80"
-     "\xF0\x88\x00\x80\xA6\xAD\x03\x00\xF4\x88\x00\x80\xF8\x88\x00\x80\xA5\xA5"
-     "\x03\x00"
-     "\xFC\x88\x00\x80\xA3\x81\x03\x00\x00\x89\x00\x80\x04\x89\x00\x80\xAE\x59"
-     "\x02\x00"
-     "\xAF\x59\x02\x00\xAC\x59\x02\x00\xAD\x59\x02\x00\xAA\x59\x02\x00\xAB\x59"
-     "\x02\x00"
-     "\x08\x89\x00\x80\x0C\x89\x00\x80\x10\x89\x00\x80\x14\x89\x00\x80\x18\x89"
-     "\x00\x80"
-     "\x1C\x89\x00\x80\x20\x89\x00\x80\x24\x89\x00\x80\x28\x89\x00\x80\xEF\xF8"
-     "\x01\x00"
-     "\x84\xAC\x1C\x00\xE1\xDC\x06\x00\x2C\x89\x00\x80\xE3\x1C\x06\x00\x30\x89"
-     "\x00\x80"
-     "\x34\x89\x00\x80\x38\x89\x00\x80\x3C\x89\x00\x80\xB3\xC1\x01\x00\x40\x89"
-     "\x00\x80"
-     "\x86\xEC\x1F\x00\x87\xF0\x1C\x00\xF5\x00\x00\x00\xB6\xC9\x01\x00\xB5\xCD"
-     "\x01\x00"
-     "\x48\x89\x00\x80\xBB\xD1\x01\x00\xBA\xC9\x01\x00\xBE\x2C\x1F\x00\x4C\x89"
-     "\x00\x80"
-     "\xBF\xB5\x01\x00\xBE\xC9\x01\x00\xBD\xC1\x01\x00\xBC\xC9\x01\x00\xA8\xF9"
-     "\x1D\x00"
-     "\xA9\xF9\x1D\x00\xAA\x09\x1E\x00\xAB\x09\x1E\x00\xAC\x05\x1E\x00\xAD\x09"
-     "\x1E\x00"
-     "\xAE\x3D\x1E\x00\xAF\x35\x1E\x00\xD4\x88\x00\x80\x82\xE1\x1F\x00\x81\x95"
-     "\x1F\x00"
-     "\x80\x95\x1F\x00\x44\x89\x00\x80\x50\x89\x00\x80\x54\x89\x00\x80\x58\x89"
-     "\x00\x80"
-     "\xB8\x11\x1E\x00\xB9\x19\x1E\x00\xBA\x21\x1E\x00\xBB\x21\x1E\x00\xBC\xD1"
-     "\x1E\x00"
-     "\xBD\xD1\x1E\x00\xBE\xD1\x1E\x00\xBF\xCD\x1E\x00\xB0\x4D\x1E\x00\xB1\x55"
-     "\x1E\x00"
-     "\xB2\x5D\x1E\x00\xB3\x55\x1E\x00\xB4\x4D\x1E\x00\xB5\x31\x1E\x00\xB6\x31"
-     "\x1E\x00"
-     "\xB7\x31\x1E\x00\xA3\x81\x1E\x00\x5C\x89\x00\x80\x60\x89\x00\x80\x64\x89"
-     "\x00\x80"
-     "\x68\x89\x00\x80\xA6\x89\x1E\x00\xA5\x8D\x1E\x00\x6C\x89\x00\x80\xAB\x91"
-     "\x1E\x00"
-     "\xAA\x89\x1E\x00\x70\x89\x00\x80\x74\x89\x00\x80\xAF\xF5\x1E\x00\xAE\x89"
-     "\x1E\x00"
-     "\xAD\x81\x1E\x00\xAC\x89\x1E\x00\x78\x89\x00\x80\xB3\xED\x1F\x00\x7C\x89"
-     "\x00\x80"
-     "\x80\x89\x00\x80\xB6\xE9\x1F\x00\x84\x89\x00\x80\x88\x89\x00\x80\xB5\xE9"
-     "\x1F\x00"
-     "\xBA\xA9\x1F\x00\xBB\xB5\x1F\x00\x8C\x89\x00\x80\x90\x89\x00\x80\xBE\x89"
-     "\x1F\x00"
-     "\xBF\x89\x1F\x00\xBC\xA1\x1F\x00\xBD\xA1\x1F\x00\xA8\x11\x1E\x00\xA9\x25"
-     "\x1E\x00"
-     "\xAA\x21\x1E\x00\xAB\x3D\x1E\x00\xAC\x25\x1E\x00\xAD\x2D\x1E\x00\xAE\x25"
-     "\x1E\x00"
-     "\xAF\x59\x1E\x00\x94\x89\x00\x80\x98\x89\x00\x80\x9C\x89\x00\x80\xA0\x89"
-     "\x00\x80"
-     "\xA4\x89\x00\x80\x82\xB9\x00\x00\x81\xBD\x00\x00\x80\xBD\x00\x00\xB8\xC9"
-     "\x1E\x00"
-     "\xB9\xC9\x1E\x00\xBA\xD9\x1E\x00\xBB\xD9\x1E\x00\xBC\xC9\x1E\x00\xBD\xFD"
-     "\x1E\x00"
-     "\xBE\xE9\x1E\x00\xBF\xE9\x1E\x00\xB0\x29\x1E\x00\xB1\x35\x1E\x00\xB2\x31"
-     "\x1E\x00"
-     "\xB3\x31\x1E\x00\xB4\x09\x1E\x00\xB5\x09\x1E\x00\xB6\xF9\x1E\x00\xB7\xF9"
-     "\x1E\x00"
-     "\xA8\x89\x00\x80\xAC\x89\x00\x80\xA3\xA1\x1E\x00\xB0\x89\x00\x80\xA5\xA5"
-     "\x1E\x00"
-     "\xB4\x89\x00\x80\xB8\x89\x00\x80\xA6\xA5\x1E\x00\x64\x88\x00\x80\x84\xA8"
-     "\x00\x00"
-     "\xAB\xF9\x1E\x00\xAA\xE5\x1E\x00\xAD\xED\x1E\x00\xAC\xED\x1E\x00\xAF\xC5"
-     "\x1E\x00"
-     "\xAE\xC5\x1E\x00\xB3\xF5\x01\x00\xBC\x89\x00\x80\x86\x6C\x00\x00\x87\x7C"
-     "\x01\x00"
-     "\xC0\x89\x00\x80\xB6\x51\x01\x00\xB5\xE5\x01\x00\xC4\x89\x00\x80\xBB\x7D"
-     "\x01\x00"
-     "\xBA\x75\x01\x00\xC8\x89\x00\x80\xCC\x89\x00\x80\xBF\x39\x01\x00\xBE\x35"
-     "\x01\x00"
-     "\xBD\x55\x01\x00\xBC\x55\x01\x00\xD0\x89\x00\x80\xD4\x89\x00\x80\xBE\x00"
-     "\x04\x00"
-     "\x84\x00\x04\x00\xE1\xC0\x01\x00\xD8\x89\x00\x80\xE3\xAC\x01\x00\xEF\xE8"
-     "\x1E\x00"
-     "\xDC\x89\x00\x80\xE1\xCC\x1D\x00\xEF\xD8\x02\x00\xE3\x34\x1E\x00\xE0\x89"
-     "\x00\x80"
-     "\xE4\x89\x00\x80\xE8\x89\x00\x80\xEC\x89\x00\x80\xF0\x89\x00\x80\xF4\x89"
-     "\x00\x80"
-     "\xA3\x7D\x02\x00\xF8\x89\x00\x80\xA5\x6D\x02\x00\xFC\x89\x00\x80\x00\x8A"
-     "\x00\x80"
-     "\xA6\xD9\x02\x00\x04\x8A\x00\x80\x08\x8A\x00\x80\xAB\xF5\x02\x00\xAA\xFD"
-     "\x02\x00"
-     "\xAD\xDD\x02\x00\xAC\xDD\x02\x00\xAF\xB1\x02\x00\xAE\xBD\x02\x00\xA9\x6D"
-     "\x1D\x00"
-     "\xA8\x59\x1D\x00\xAB\xB1\x02\x00\xAA\xB9\x02\x00\xAD\x15\x02\x00\xAC\xA9"
-     "\x02\x00"
-     "\xAF\x15\x02\x00\xAE\x15\x02\x00\xBE\x2C\x05\x00\x0C\x8A\x00\x80\x10\x8A"
-     "\x00\x80"
-     "\x14\x8A\x00\x80\x80\x0D\x00\x00\x81\x95\x01\x00\x82\x9D\x01\x00\x18\x8A"
-     "\x00\x80"
-     "\xB9\x29\x02\x00\xB8\x1D\x02\x00\xBB\xE9\x03\x00\xBA\x21\x02\x00\xBD\xF9"
-     "\x03\x00"
-     "\xBC\xF1\x03\x00\xBF\xE9\x03\x00\xBE\xF1\x03\x00\xB1\x7D\x02\x00\xB0\x75"
-     "\x02\x00"
-     "\xB3\x45\x02\x00\xB2\x45\x02\x00\xB5\x39\x02\x00\xB4\x35\x02\x00\xB7\x2D"
-     "\x02\x00"
-     "\xB6\x31\x02\x00\xA9\xC5\x02\x00\xA8\x9D\x02\x00\xAB\xD5\x02\x00\xAA\xDD"
-     "\x02\x00"
-     "\xAD\xFD\x02\x00\xAC\xFD\x02\x00\xAF\x1D\x01\x00\xAE\x1D\x01\x00\x1C\x8A"
-     "\x00\x80"
-     "\x20\x8A\x00\x80\x87\xA0\x05\x00\x86\xAC\x05\x00\x28\x8A\x00\x80\x2C\x8A"
-     "\x00\x80"
-     "\x30\x8A\x00\x80\xCF\x00\x00\x00\xB9\x15\x01\x00\xB8\x09\x01\x00\xBB\x29"
-     "\x01\x00"
-     "\xBA\x15\x01\x00\xBD\xF5\x01\x00\xBC\x31\x01\x00\xBF\xF5\x01\x00\xBE\xF5"
-     "\x01\x00"
-     "\xB1\x41\x01\x00\xB0\x71\x01\x00\xB3\x5D\x01\x00\xB2\x5D\x01\x00\xB5\x4D"
-     "\x01\x00"
-     "\xB4\x4D\x01\x00\xB7\x3D\x01\x00\xB6\x3D\x01\x00\x34\x8A\x00\x80\x38\x8A"
-     "\x00\x80"
-     "\x3C\x8A\x00\x80\xEF\x9C\x1E\x00\x40\x8A\x00\x80\x44\x8A\x00\x80\x48\x8A"
-     "\x00\x80"
-     "\xEF\xE8\x1F\x00\x84\x6C\x07\x00\xE1\xA4\x1E\x00\x4C\x8A\x00\x80\xE3\xB0"
-     "\x1E\x00"
-     "\x50\x8A\x00\x80\xE1\x64\x1E\x00\x54\x8A\x00\x80\xE3\x0C\x1E\x00\xB3\x11"
-     "\x02\x00"
-     "\x58\x8A\x00\x80\x5C\x8A\x00\x80\x60\x8A\x00\x80\x84\x5C\x06\x00\xB6\x39"
-     "\x02\x00"
-     "\xB5\x31\x02\x00\x64\x8A\x00\x80\xBB\xC1\x01\x00\xBA\xF1\x01\x00\x68\x8A"
-     "\x00\x80"
-     "\x6C\x8A\x00\x80\xBF\xCD\x01\x00\xBE\xCD\x01\x00\xBD\xCD\x01\x00\xBC\xCD"
-     "\x01\x00"
-     "\x24\x8A\x00\x80\x70\x8A\x00\x80\x74\x8A\x00\x80\x78\x8A\x00\x80\x7C\x8A"
-     "\x00\x80"
-     "\x80\x8A\x00\x80\x84\x8A\x00\x80\x88\x8A\x00\x80\xA8\x11\x07\x00\xA9\x11"
-     "\x07\x00"
-     "\xAA\x11\x07\x00\xAB\xAD\x07\x00\xAC\xB5\x07\x00\xAD\xBD\x07\x00\xAE\xA9"
-     "\x07\x00"
-     "\xAF\xA9\x07\x00\xB0\xD9\x07\x00\xB1\xED\x07\x00\xB2\xF9\x07\x00\xB3\xF5"
-     "\x07\x00"
-     "\xB4\x95\x07\x00\xB5\x9D\x07\x00\xB6\x95\x07\x00\xB7\x81\x07\x00\xB8\x89"
-     "\x07\x00"
-     "\xB9\x89\x07\x00\xBA\x6D\x00\x00\xBB\x65\x00\x00\xBC\x7D\x00\x00\xBD\x65"
-     "\x00\x00"
-     "\xBE\x6D\x00\x00\xBF\x65\x00\x00\x82\x19\x00\x00\xA3\x55\x05\x00\x80\x95"
-     "\x00\x00"
-     "\x81\x19\x00\x00\xA6\x7D\x05\x00\x8C\x8A\x00\x80\x90\x8A\x00\x80\xA5\x75"
-     "\x05\x00"
-     "\xAA\xB5\x06\x00\xAB\x85\x06\x00\x94\x8A\x00\x80\x98\x8A\x00\x80\xAE\x89"
-     "\x06\x00"
-     "\xAF\x89\x06\x00\xAC\x89\x06\x00\xAD\x89\x06\x00\xB3\xF9\x07\x00\x9C\x8A"
-     "\x00\x80"
-     "\x86\xE8\x03\x00\x87\xBC\x03\x00\xA0\x8A\x00\x80\xB6\x5D\x07\x00\xB5\x5D"
-     "\x07\x00"
-     "\xA4\x8A\x00\x80\xBB\x49\x07\x00\xBA\x7D\x07\x00\xA8\x8A\x00\x80\xAC\x8A"
-     "\x00\x80"
-     "\xBF\x39\x07\x00\xBE\x31\x07\x00\xBD\x41\x07\x00\xBC\x4D\x07\x00\xA3\x39"
-     "\x06\x00"
-     "\xB0\x8A\x00\x80\xB4\x8A\x00\x80\xB8\x8A\x00\x80\xBC\x8A\x00\x80\xA6\x9D"
-     "\x06\x00"
-     "\xA5\x9D\x06\x00\xC0\x8A\x00\x80\xAB\x89\x06\x00\xAA\xBD\x06\x00\xC4\x8A"
-     "\x00\x80"
-     "\xC8\x8A\x00\x80\xAF\xF9\x06\x00\xAE\xF1\x06\x00\xAD\x81\x06\x00\xAC\x8D"
-     "\x06\x00"
-     "\xCC\x8A\x00\x80\xD0\x8A\x00\x80\xD4\x8A\x00\x80\xD8\x8A\x00\x80\xDC\x8A"
-     "\x00\x80"
-     "\xE0\x8A\x00\x80\xE4\x8A\x00\x80\xE8\x8A\x00\x80\xEC\x8A\x00\x80\xF0\x8A"
-     "\x00\x80"
-     "\xF4\x8A\x00\x80\xF8\x8A\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xD9"
-     "\x00\x00"
-     "\xFC\x8A\x00\x80\xA8\x2D\x01\x00\xA9\x41\x01\x00\xAA\x59\x01\x00\xAB\x51"
-     "\x01\x00"
-     "\xAC\x79\x01\x00\xAD\x79\x01\x00\xAE\x9D\x00\x00\xAF\x95\x00\x00\x00\x8B"
-     "\x00\x80"
-     "\xBE\x6C\x01\x00\x86\x88\x01\x00\x87\xAC\x00\x00\x08\x8B\x00\x80\x0C\x8B"
-     "\x00\x80"
-     "\x10\x8B\x00\x80\x14\x8B\x00\x80\xB8\x8D\x00\x00\xB9\x95\x00\x00\xBA\x9D"
-     "\x00\x00"
-     "\xBB\x95\x00\x00\xBC\x8D\x00\x00\xBD\xB5\x00\x00\xBE\xBD\x00\x00\xBF\xB5"
-     "\x00\x00"
-     "\xB0\xED\x00\x00\xB1\xF5\x00\x00\xB2\xF5\x00\x00\xB3\xC5\x00\x00\xB4\xDD"
-     "\x00\x00"
-     "\xB5\xB5\x00\x00\xB6\xBD\x00\x00\xB7\xB5\x00\x00\xB3\xD1\x02\x00\x18\x8B"
-     "\x00\x80"
-     "\xBE\x70\x03\x00\xBE\x44\x0C\x00\x1C\x8B\x00\x80\xB6\xF9\x02\x00\xB5\xF1"
-     "\x02\x00"
-     "\x20\x8B\x00\x80\xBB\x31\x03\x00\xBA\x29\x03\x00\x24\x8B\x00\x80\x28\x8B"
-     "\x00\x80"
-     "\xBF\x71\x03\x00\xBE\x19\x03\x00\xBD\x11\x03\x00\xBC\x21\x03\x00\x2C\x8B"
-     "\x00\x80"
-     "\xA3\x95\x02\x00\x30\x8B\x00\x80\x34\x8B\x00\x80\xA6\xBD\x02\x00\x38\x8B"
-     "\x00\x80"
-     "\x3C\x8B\x00\x80\xA5\xB5\x02\x00\xAA\x6D\x03\x00\xAB\x75\x03\x00\x40\x8B"
-     "\x00\x80"
-     "\x44\x8B\x00\x80\xAE\x5D\x03\x00\xAF\x35\x03\x00\xAC\x65\x03\x00\xAD\x55"
-     "\x03\x00"
-     "\xA8\x31\x03\x00\xA9\x31\x03\x00\xAA\x31\x03\x00\xAB\x31\x03\x00\xAC\x95"
-     "\x00\x00"
-     "\xAD\x85\x00\x00\xAE\x85\x00\x00\xAF\xB5\x00\x00\x48\x8B\x00\x80\x4C\x8B"
-     "\x00\x80"
-     "\x50\x8B\x00\x80\x54\x8B\x00\x80\x58\x8B\x00\x80\x5C\x8B\x00\x80\x60\x8B"
-     "\x00\x80"
-     "\x64\x8B\x00\x80\xB8\x61\x00\x00\xB9\x61\x00\x00\xBA\x61\x00\x00\xBB\x61"
-     "\x00\x00"
-     "\xBC\x61\x00\x00\xBD\x61\x00\x00\xBE\x61\x00\x00\xBF\x61\x00\x00\xB0\xCD"
-     "\x00\x00"
-     "\xB1\xA5\x00\x00\xB2\xA1\x00\x00\xB3\xA1\x00\x00\xB4\xA5\x00\x00\xB5\xAD"
-     "\x00\x00"
-     "\xB6\x91\x00\x00\xB7\x91\x00\x00\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x11"
-     "\x00\x00"
-     "\x68\x8B\x00\x80\x6C\x8B\x00\x80\xEF\x48\x00\x00\x70\x8B\x00\x80\x78\x8B"
-     "\x00\x80"
-     "\x84\x6C\x0D\x00\xE3\x94\x06\x00\x7C\x8B\x00\x80\xE1\xFC\x01\x00\x80\x8B"
-     "\x00\x80"
-     "\x84\x8B\x00\x80\x86\xB8\x0D\x00\x87\xC4\x02\x00\xB3\xF9\x02\x00\x88\x8B"
-     "\x00\x80"
-     "\x8C\x8B\x00\x80\x90\x8B\x00\x80\x94\x8B\x00\x80\xB6\x45\x01\x00\xB5\x5D"
-     "\x01\x00"
-     "\xBE\x98\x0F\x00\xBB\x21\x01\x00\xBA\x49\x01\x00\x98\x8B\x00\x80\x9C\x8B"
-     "\x00\x80"
-     "\xBF\x29\x01\x00\xBE\x21\x01\x00\xBD\x35\x01\x00\xBC\x35\x01\x00\xA0\x8B"
-     "\x00\x80"
-     "\xE1\xB0\x06\x00\xA4\x8B\x00\x80\xE3\x90\x07\x00\xA8\x8B\x00\x80\xAC\x8B"
-     "\x00\x80"
-     "\xB0\x8B\x00\x80\xB4\x8B\x00\x80\xB8\x8B\x00\x80\xBC\x8B\x00\x80\xC0\x8B"
-     "\x00\x80"
-     "\xC4\x8B\x00\x80\xE5\x00\x00\x00\xC8\x8B\x00\x80\xCC\x8B\x00\x80\xEF\xE4"
-     "\x07\x00"
-     "\xAE\x95\x01\x00\xAF\x9D\x01\x00\xAC\x81\x01\x00\xAD\x81\x01\x00\xAA\xFD"
-     "\x01\x00"
-     "\xAB\x95\x01\x00\xD0\x8B\x00\x80\xD4\x8B\x00\x80\xA6\xF1\x01\x00\xD8\x8B"
-     "\x00\x80"
-     "\xDC\x8B\x00\x80\xA5\xE9\x01\x00\x82\xF5\x0F\x00\xA3\x4D\x02\x00\x80\x1D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x74\x8B\x00\x80\xE0\x8B\x00\x80\xE8\x8B\x00\x80\xE4\x8B"
-     "\x00\x80"
-     "\x86\xB0\x03\x00\x87\x94\x03\x00\xEC\x8B\x00\x80\xF0\x8B\x00\x80\xA8\x3D"
-     "\x0F\x00"
-     "\xA9\x0D\x0F\x00\xAA\x19\x0F\x00\xAB\x19\x0F\x00\xAC\x75\x0F\x00\xAD\x7D"
-     "\x0F\x00"
-     "\xAE\x75\x0F\x00\xAF\x6D\x0F\x00\xB0\x29\x0F\x00\xB1\x29\x0F\x00\xB2\xC9"
-     "\x0F\x00"
-     "\xB3\xC5\x0F\x00\xB4\xC1\x0F\x00\xB5\xC1\x0F\x00\xB6\xC1\x0F\x00\xB7\xC1"
-     "\x0F\x00"
-     "\xB8\xC1\x0F\x00\xB9\xC1\x0F\x00\xBA\xC1\x0F\x00\xBB\xC1\x0F\x00\xBC\xC5"
-     "\x0F\x00"
-     "\xBD\xCD\x0F\x00\xBE\xA1\x0F\x00\xBF\xA1\x0F\x00\xB3\xB1\x0F\x00\xF4\x8B"
-     "\x00\x80"
-     "\xF8\x8B\x00\x80\xFC\x8B\x00\x80\x00\x8C\x00\x80\xB6\xDD\x0F\x00\xB5\xDD"
-     "\x0F\x00"
-     "\x04\x8C\x00\x80\xBB\xCD\x0F\x00\xBA\xCD\x0F\x00\x08\x8C\x00\x80\x0C\x8C"
-     "\x00\x80"
-     "\xBF\x25\x0F\x00\xBE\x25\x0F\x00\xBD\x31\x0F\x00\xBC\x3D\x0F\x00\x10\x8C"
-     "\x00\x80"
-     "\xA3\xF5\x0F\x00\x14\x8C\x00\x80\x18\x8C\x00\x80\xA6\x99\x0F\x00\x1C\x8C"
-     "\x00\x80"
-     "\x20\x8C\x00\x80\xA5\x99\x0F\x00\xAA\x89\x0F\x00\xAB\x89\x0F\x00\x24\x8C"
-     "\x00\x80"
-     "\x28\x8C\x00\x80\xAE\x61\x0F\x00\xAF\x61\x0F\x00\xAC\x79\x0F\x00\xAD\x75"
-     "\x0F\x00"
-     "\xA8\x09\x01\x00\xA9\x1D\x01\x00\xAA\x15\x01\x00\xAB\x89\x01\x00\xAC\x99"
-     "\x01\x00"
-     "\xAD\x99\x01\x00\xAE\x89\x01\x00\xAF\x81\x01\x00\x2C\x8C\x00\x80\x30\x8C"
-     "\x00\x80"
-     "\x34\x8C\x00\x80\x38\x8C\x00\x80\x3C\x8C\x00\x80\x40\x8C\x00\x80\x44\x8C"
-     "\x00\x80"
-     "\x48\x8C\x00\x80\xB8\xAD\x01\x00\xB9\x65\x00\x00\xBA\x6D\x00\x00\xBB\x65"
-     "\x00\x00"
-     "\xBC\x7D\x00\x00\xBD\x65\x00\x00\xBE\x6D\x00\x00\xBF\x65\x00\x00\xB0\xC9"
-     "\x01\x00"
-     "\xB1\xC9\x01\x00\xB2\xA9\x01\x00\xB3\xA5\x01\x00\xB4\xBD\x01\x00\xB5\xA1"
-     "\x01\x00"
-     "\xB6\xA1\x01\x00\xB7\x95\x01\x00\x80\xBD\x00\x00\x81\x09\x00\x00\x82\x19"
-     "\x00\x00"
-     "\x4C\x8C\x00\x80\x50\x8C\x00\x80\x84\x98\x02\x00\x54\x8C\x00\x80\x58\x8C"
-     "\x00\x80"
-     "\x86\x80\x01\x00\x87\x04\x01\x00\x5C\x8C\x00\x80\x60\x8C\x00\x80\x64\x8C"
-     "\x00\x80"
-     "\x68\x8C\x00\x80\x04\x8B\x00\x80\x6C\x8C\x00\x80\xB3\xDD\x01\x00\x70\x8C"
-     "\x00\x80"
-     "\x74\x8C\x00\x80\x78\x8C\x00\x80\x7C\x8C\x00\x80\xB6\x59\x01\x00\xB5\x51"
-     "\x01\x00"
-     "\x80\x8C\x00\x80\xBB\x4D\x01\x00\xBA\x4D\x01\x00\x84\x8C\x00\x80\x88\x8C"
-     "\x00\x80"
-     "\xBF\x25\x01\x00\xBE\x39\x01\x00\xBD\x31\x01\x00\xBC\x3D\x01\x00\x8C\x8C"
-     "\x00\x80"
-     "\xEF\xF4\x0E\x00\xBE\xD0\x03\x00\x90\x8C\x00\x80\x94\x8C\x00\x80\xE3\x48"
-     "\x02\x00"
-     "\x98\x8C\x00\x80\xE1\xA4\x01\x00\x9C\x8C\x00\x80\xEF\xD4\x02\x00\xA0\x8C"
-     "\x00\x80"
-     "\xA4\x8C\x00\x80\xA8\x8C\x00\x80\xE3\xF4\x0B\x00\xAC\x8C\x00\x80\xE1\x8C"
-     "\x0D\x00"
-     "\xB0\x8C\x00\x80\xB4\x8C\x00\x80\xA3\x55\x02\x00\x84\x8C\x03\x00\xA5\xD9"
-     "\x02\x00"
-     "\xB8\x8C\x00\x80\xBC\x8C\x00\x80\xA6\xD1\x02\x00\xC0\x8C\x00\x80\xC8\x8C"
-     "\x00\x80"
-     "\xAB\xC5\x02\x00\xAA\xC5\x02\x00\xAD\xB9\x02\x00\xAC\xB5\x02\x00\xAF\xAD"
-     "\x02\x00"
-     "\xAE\xB1\x02\x00\xA8\x1D\x03\x00\xA9\x2D\x03\x00\xAA\x25\x03\x00\xAB\x3D"
-     "\x03\x00"
-     "\xAC\x25\x03\x00\xAD\x29\x03\x00\xAE\x5D\x03\x00\xAF\x55\x03\x00\x80\xF5"
-     "\x00\x00"
-     "\x81\xFD\x00\x00\x82\xC9\x00\x00\xCC\x8C\x00\x80\xD0\x8C\x00\x80\xD4\x8C"
-     "\x00\x80"
-     "\x87\x6C\x03\x00\x86\x1C\x04\x00\xB8\xD9\x00\x00\xB9\xE5\x00\x00\xBA\xE1"
-     "\x00\x00"
-     "\xBB\xE1\x00\x00\xBC\xE1\x00\x00\xBD\xED\x00\x00\xBE\x99\x00\x00\xBF\x99"
-     "\x00\x00"
-     "\xB0\x2D\x03\x00\xB1\x35\x03\x00\xB2\x35\x03\x00\xB3\x05\x03\x00\xB4\x1D"
-     "\x03\x00"
-     "\xB5\xF5\x00\x00\xB6\xF5\x00\x00\xB7\xED\x00\x00\xD8\x8C\x00\x80\xDC\x8C"
-     "\x00\x80"
-     "\xE0\x8C\x00\x80\xB3\xA1\x02\x00\xE4\x8C\x00\x80\xB5\xA1\x02\x00\xB6\xA5"
-     "\x02\x00"
-     "\xE8\x8C\x00\x80\xEC\x8C\x00\x80\xF0\x8C\x00\x80\xBA\xE5\x02\x00\xBB\xFD"
-     "\x02\x00"
-     "\xBC\xED\x02\x00\xBD\xDD\x02\x00\xBE\xD5\x02\x00\xBF\x3D\x01\x00\xA6\xFD"
-     "\x02\x00"
-     "\xF4\x8C\x00\x80\xF8\x8C\x00\x80\xA5\xF9\x02\x00\xFC\x8C\x00\x80\xA3\xF9"
-     "\x02\x00"
-     "\x00\x8D\x00\x80\x04\x8D\x00\x80\xAE\x8D\x02\x00\xAF\x65\x01\x00\xAC\xB5"
-     "\x02\x00"
-     "\xAD\x85\x02\x00\xAA\xBD\x02\x00\xAB\xA5\x02\x00\x08\x8D\x00\x80\x0C\x8D"
-     "\x00\x80"
-     "\xE3\x04\x00\x00\xE1\xA4\x0E\x00\xE1\x80\x01\x00\xE3\xBC\x0E\x00\x10\x8D"
-     "\x00\x80"
-     "\x14\x8D\x00\x80\x18\x8D\x00\x80\x1C\x8D\x00\x80\xBE\xE8\x05\x00\x20\x8D"
-     "\x00\x80"
-     "\x24\x8D\x00\x80\x28\x8D\x00\x80\xEF\x14\x0E\x00\x2C\x8D\x00\x80\xBE\xF0"
-     "\x04\x00"
-     "\xEF\x48\x0F\x00\x30\x8D\x00\x80\x38\x8D\x00\x80\x3C\x8D\x00\x80\x40\x8D"
-     "\x00\x80"
-     "\x81\x65\x00\x00\x80\x79\x00\x00\x44\x8D\x00\x80\x82\x1D\x00\x00\x84\xAC"
-     "\x05\x00"
-     "\x48\x8D\x00\x80\x87\xC8\x04\x00\x86\x4C\x04\x00\x4C\x8D\x00\x80\x50\x8D"
-     "\x00\x80"
-     "\x54\x8D\x00\x80\x58\x8D\x00\x80\xA8\x65\x06\x00\xA9\xA9\x06\x00\xAA\xB9"
-     "\x06\x00"
-     "\xAB\xB9\x06\x00\xAC\x11\x07\x00\xAD\x11\x07\x00\xAE\x11\x07\x00\xAF\x0D"
-     "\x07\x00"
-     "\x34\x8D\x00\x80\x5C\x8D\x00\x80\x60\x8D\x00\x80\x64\x8D\x00\x80\x68\x8D"
-     "\x00\x80"
-     "\x6C\x8D\x00\x80\x70\x8D\x00\x80\x74\x8D\x00\x80\xB8\x39\x07\x00\xB9\x39"
-     "\x07\x00"
-     "\xBA\xD9\x07\x00\xBB\xD9\x07\x00\xBC\xC9\x07\x00\xBD\xC9\x07\x00\xBE\xFD"
-     "\x07\x00"
-     "\xBF\xF5\x07\x00\xB0\x7D\x07\x00\xB1\x45\x07\x00\xB2\x4D\x07\x00\xB3\x29"
-     "\x07\x00"
-     "\xB4\x39\x07\x00\xB5\x25\x07\x00\xB6\x21\x07\x00\xB7\x21\x07\x00\xB3\x31"
-     "\x06\x00"
-     "\x78\x8D\x00\x80\x7C\x8D\x00\x80\x80\x8D\x00\x80\x84\x8D\x00\x80\xB6\x91"
-     "\x07\x00"
-     "\xB5\x99\x07\x00\x88\x8D\x00\x80\xBB\x81\x07\x00\xBA\xB9\x07\x00\x8C\x8D"
-     "\x00\x80"
-     "\x90\x8D\x00\x80\xBF\xE1\x07\x00\xBE\xFD\x07\x00\xBD\xFD\x07\x00\xBC\x99"
-     "\x07\x00"
-     "\x94\x8D\x00\x80\xA3\x75\x06\x00\x98\x8D\x00\x80\x9C\x8D\x00\x80\xA6\xD5"
-     "\x07\x00"
-     "\xA0\x8D\x00\x80\x84\x94\x03\x00\xA5\xDD\x07\x00\xAA\xFD\x07\x00\xAB\xC5"
-     "\x07\x00"
-     "\xA4\x8D\x00\x80\xA8\x8D\x00\x80\xAE\xB9\x07\x00\xAF\xA5\x07\x00\xAC\xDD"
-     "\x07\x00"
-     "\xAD\xB9\x07\x00\xA8\xC5\x06\x00\xA9\xCD\x06\x00\xAA\xC5\x06\x00\xAB\xD5"
-     "\x06\x00"
-     "\xAC\xF5\x06\x00\xAD\xE5\x06\x00\xAE\xED\x06\x00\xAF\xE5\x06\x00\x80\xE9"
-     "\x01\x00"
-     "\x81\xE9\x01\x00\x82\xFD\x01\x00\xAC\x8D\x00\x80\x86\x30\x01\x00\x87\x64"
-     "\x01\x00"
-     "\xB0\x8D\x00\x80\xB4\x8D\x00\x80\xB8\xC5\x01\x00\xB9\xCD\x01\x00\xBA\xC5"
-     "\x01\x00"
-     "\xBB\xDD\x01\x00\xBC\xCD\x01\x00\xBD\xFD\x01\x00\xBE\xF5\x01\x00\xBF\x9D"
-     "\x01\x00"
-     "\xB0\x9D\x06\x00\xB1\x41\x01\x00\xB2\x41\x01\x00\xB3\x41\x01\x00\xB4\x41"
-     "\x01\x00"
-     "\xB5\x41\x01\x00\xB6\x41\x01\x00\xB7\x41\x01\x00\xB3\x4D\x06\x00\xB8\x8D"
-     "\x00\x80"
-     "\xBC\x8D\x00\x80\xC0\x8D\x00\x80\xC4\x8D\x00\x80\xB6\x65\x06\x00\xB5\x7D"
-     "\x06\x00"
-     "\xC8\x8D\x00\x80\xBB\x85\x01\x00\xBA\x59\x06\x00\xCC\x8D\x00\x80\xD0\x8D"
-     "\x00\x80"
-     "\xBF\x85\x01\x00\xBE\x85\x01\x00\xBD\x95\x01\x00\xBC\x95\x01\x00\xD4\x8D"
-     "\x00\x80"
-     "\xA3\x09\x06\x00\xC4\x8C\x00\x80\xD8\x8D\x00\x80\xA6\x21\x06\x00\xDC\x8D"
-     "\x00\x80"
-     "\xE0\x8D\x00\x80\xA5\x39\x06\x00\xAA\x1D\x06\x00\xAB\xC1\x01\x00\xE4\x8D"
-     "\x00\x80"
-     "\xE8\x8D\x00\x80\xAE\xC1\x01\x00\xAF\xC1\x01\x00\xAC\xD1\x01\x00\xAD\xD1"
-     "\x01\x00"
-     "\xEC\x8D\x00\x80\xF0\x8D\x00\x80\xF4\x8D\x00\x80\xF8\x8D\x00\x80\xFC\x8D"
-     "\x00\x80"
-     "\x00\x8E\x00\x80\x04\x8E\x00\x80\x08\x8E\x00\x80\x0C\x8E\x00\x80\x10\x8E"
-     "\x00\x80"
-     "\xBE\x64\x03\x00\x18\x8E\x00\x80\x1C\x8E\x00\x80\x20\x8E\x00\x80\xBF\x30"
-     "\x3D\x00"
-     "\x24\x8E\x00\x80\x85\x00\x3C\x00\xE3\xB4\x07\x00\x84\x8C\x03\x00\xE1\xE8"
-     "\x06\x00"
-     "\x80\x4D\x00\x00\x81\x75\x00\x00\x82\x7D\x00\x00\xEF\x8C\x2A\x00\x85\x44"
-     "\x3C\x00"
-     "\xE1\x10\x02\x00\x28\x8E\x00\x80\xE3\xA8\x2E\x00\x2C\x8E\x00\x80\xEF\x2C"
-     "\x07\x00"
-     "\x86\x58\x3D\x00\x87\x0C\x03\x00\x30\x8E\x00\x80\x34\x8E\x00\x80\x38\x8E"
-     "\x00\x80"
-     "\x3C\x8E\x00\x80\x40\x8E\x00\x80\x44\x8E\x00\x80\x48\x8E\x00\x80\x4C\x8E"
-     "\x00\x80"
-     "\xA8\x4D\x02\x00\xA9\x89\x03\x00\xAA\x91\x03\x00\xAB\x91\x03\x00\xAC\x35"
-     "\x03\x00"
-     "\xAD\x3D\x03\x00\xAE\x35\x03\x00\xAF\x2D\x03\x00\xB0\x5D\x03\x00\xB1\x6D"
-     "\x03\x00"
-     "\xB2\x65\x03\x00\xB3\x0D\x03\x00\xB4\x1D\x03\x00\xB5\x01\x03\x00\xB6\x01"
-     "\x03\x00"
-     "\xB7\x35\x03\x00\xB8\x0D\x03\x00\xB9\xE5\x00\x00\xBA\xED\x00\x00\xBB\xE5"
-     "\x00\x00"
-     "\xBC\xFD\x00\x00\xBD\xE5\x00\x00\xBE\xED\x00\x00\xBF\xE5\x00\x00\x50\x8E"
-     "\x00\x80"
-     "\x54\x8E\x00\x80\xB3\x31\x02\x00\x58\x8E\x00\x80\xB5\x95\x02\x00\x5C\x8E"
-     "\x00\x80"
-     "\x60\x8E\x00\x80\xB6\x95\x02\x00\x84\xBC\x3D\x00\x64\x8E\x00\x80\xBB\x81"
-     "\x02\x00"
-     "\xBA\xB1\x02\x00\xBD\xF1\x02\x00\xBC\x99\x02\x00\xBF\xED\x02\x00\xBE\xF1"
-     "\x02\x00"
-     "\x68\x8E\x00\x80\x6C\x8E\x00\x80\xBE\xC4\x3F\x00\x84\xC4\x3F\x00\x70\x8E"
-     "\x00\x80"
-     "\xEF\x84\x03\x00\x74\x8E\x00\x80\x78\x8E\x00\x80\x7C\x8E\x00\x80\xE3\x3C"
-     "\x03\x00"
-     "\x80\x8E\x00\x80\xE1\x44\x00\x00\x80\x2D\x00\x00\x81\x15\x00\x00\x82\x11"
-     "\x00\x00"
-     "\x88\x8E\x00\x80\x86\xA0\x3F\x00\x87\x58\x3D\x00\x8C\x8E\x00\x80\x90\x8E"
-     "\x00\x80"
-     "\x94\x8E\x00\x80\x98\x8E\x00\x80\x9C\x8E\x00\x80\xEF\xDC\x06\x00\xBE\x2C"
-     "\x3E\x00"
-     "\xE1\x50\x06\x00\xA0\x8E\x00\x80\xE3\x64\x01\x00\xA4\x8E\x00\x80\xA8\x8E"
-     "\x00\x80"
-     "\xAC\x8E\x00\x80\xB0\x8E\x00\x80\xA3\xF9\x02\x00\xB4\x8E\x00\x80\xB8\x8E"
-     "\x00\x80"
-     "\xBC\x8E\x00\x80\xC0\x8E\x00\x80\xA6\x5D\x02\x00\xA5\x5D\x02\x00\xC4\x8E"
-     "\x00\x80"
-     "\xAB\x49\x02\x00\xAA\x79\x02\x00\xC8\x8E\x00\x80\xCC\x8E\x00\x80\xAF\x25"
-     "\x02\x00"
-     "\xAE\x39\x02\x00\xAD\x39\x02\x00\xAC\x51\x02\x00\x90\x6D\x24\x00\x91\x05"
-     "\x28\x00"
-     "\x92\x0D\x28\x00\x93\x9D\x28\x00\x94\x8D\x2D\x00\x95\x35\x2C\x00\x96\xB1"
-     "\x2C\x00"
-     "\x97\xBD\x31\x00\x98\x01\x30\x00\x99\x15\x30\x00\x9A\xBD\x31\x00\x9B\x2D"
-     "\x34\x00"
-     "\x9C\x99\x34\x00\x9D\x8D\x39\x00\x9E\x39\x38\x00\x9F\x31\x38\x00\x14\x8E"
-     "\x00\x80"
-     "\xB6\x35\x3E\x00\xB5\x21\x3E\x00\x84\x8E\x00\x80\xB3\x65\x3E\x00\xD0\x8E"
-     "\x00\x80"
-     "\xD4\x8E\x00\x80\xD8\x8E\x00\x80\xBF\xFD\x3E\x00\xBE\xFD\x3E\x00\xBD\xFD"
-     "\x3E\x00"
-     "\xBC\xFD\x3E\x00\xBB\xFD\x3E\x00\xBA\xFD\x3E\x00\xDC\x8E\x00\x80\x8F\x39"
-     "\x24\x00"
-     "\xA9\x45\x09\x00\xA8\x9D\x08\x00\xAB\xA9\x0C\x00\xAA\x01\x0C\x00\xAD\x01"
-     "\x10\x00"
-     "\xAC\x1D\x0D\x00\xAF\x35\x10\x00\xAE\xBD\x10\x00\xA1\xE1\x00\x00\xE0\x8E"
-     "\x00\x80"
-     "\xA3\x01\x04\x00\xA2\xE1\x00\x00\xA5\x95\x04\x00\xA4\x95\x04\x00\xA7\xB9"
-     "\x08\x00"
-     "\xA6\x01\x08\x00\xA0\x59\x39\x00\xA1\x0D\x3D\x00\xA2\xE9\x3C\x00\xA3\x25"
-     "\x3C\x00"
-     "\xE4\x8E\x00\x80\x82\xDD\x01\x00\x81\x25\x00\x00\x80\x25\x00\x00\xB1\xC1"
-     "\x14\x00"
-     "\xB0\x01\x14\x00\xB3\x01\x18\x00\xB2\xD5\x14\x00\xB5\xD5\x18\x00\xB4\xD5"
-     "\x18\x00"
-     "\xE8\x8E\x00\x80\xEC\x8E\x00\x80\x82\x45\x25\x00\x83\x69\x25\x00\xF0\x8E"
-     "\x00\x80"
-     "\xF8\x8E\x00\x80\x86\x81\x29\x00\x87\x89\x2C\x00\x84\x19\x29\x00\x85\x19"
-     "\x29\x00"
-     "\x8A\x01\x2D\x00\x8B\x8D\x2D\x00\x86\x88\x07\x00\x87\x8C\x01\x00\x8E\x01"
-     "\x31\x00"
-     "\x8F\xA1\x30\x00\x8C\x99\x30\x00\x8D\x1D\x31\x00\x92\x41\x35\x00\x93\xED"
-     "\x35\x00"
-     "\xFC\x8E\x00\x80\x00\x8F\x00\x80\x96\x5D\x39\x00\x97\xCD\x38\x00\x94\xE1"
-     "\x38\x00"
-     "\x95\x5D\x39\x00\x9A\xA9\x3C\x00\x9B\x45\x3D\x00\x04\x8F\x00\x80\x08\x8F"
-     "\x00\x80"
-     "\x0C\x8F\x00\x80\x10\x8F\x00\x80\x9C\x91\x3D\x00\x14\x8F\x00\x80\xA8\xAD"
-     "\x3F\x00"
-     "\xA9\xB5\x3F\x00\xAA\xBD\x3F\x00\xAB\xB1\x3F\x00\xAC\xF9\x3F\x00\xAD\xE5"
-     "\x3F\x00"
-     "\xAE\x19\x3F\x00\xAF\x19\x3F\x00\x18\x8F\x00\x80\x1C\x8F\x00\x80\x20\x8F"
-     "\x00\x80"
-     "\x24\x8F\x00\x80\x28\x8F\x00\x80\x2C\x8F\x00\x80\x30\x8F\x00\x80\x34\x8F"
-     "\x00\x80"
-     "\xB8\xCD\x3E\x00\xB9\xD1\x3E\x00\xBA\xD1\x3E\x00\xBB\xE9\x3E\x00\xBC\x91"
-     "\x3E\x00"
-     "\xBD\x91\x3E\x00\xBE\x89\x3E\x00\xBF\x85\x3E\x00\xB0\x69\x3F\x00\xB1\x69"
-     "\x3F\x00"
-     "\xB2\x79\x3F\x00\xB3\x79\x3F\x00\xB4\x69\x3F\x00\xB5\x69\x3F\x00\xB6\xF9"
-     "\x3E\x00"
-     "\xB7\xF9\x3E\x00\xA3\x35\x3C\x00\x38\x8F\x00\x80\x3C\x8F\x00\x80\x40\x8F"
-     "\x00\x80"
-     "\x44\x8F\x00\x80\xA6\x65\x3C\x00\xA5\x71\x3C\x00\x48\x8F\x00\x80\xAB\xAD"
-     "\x3C\x00"
-     "\xAA\xAD\x3C\x00\x4C\x8F\x00\x80\x50\x8F\x00\x80\xAF\xAD\x3C\x00\xAE\xAD"
-     "\x3C\x00"
-     "\xAD\xAD\x3C\x00\xAC\xAD\x3C\x00\x84\xC8\x03\x00\xB3\xC9\x3D\x00\x54\x8F"
-     "\x00\x80"
-     "\x58\x8F\x00\x80\xB6\xC9\x3D\x00\x5C\x8F\x00\x80\x60\x8F\x00\x80\xB5\xC9"
-     "\x3D\x00"
-     "\xBA\xD5\x3D\x00\xBB\xDD\x3D\x00\xBE\x6C\x03\x00\x68\x8F\x00\x80\xBE\xB9"
-     "\x3D\x00"
-     "\xBF\xB9\x3D\x00\xBC\xC5\x3D\x00\xBD\xB9\x3D\x00\xA8\xFD\x3D\x00\xA9\x0D"
-     "\x3D\x00"
-     "\xAA\x05\x3D\x00\xAB\x1D\x3D\x00\xAC\x05\x3D\x00\xAD\x0D\x3D\x00\xAE\x31"
-     "\x3D\x00"
-     "\xAF\x2D\x3D\x00\x80\xED\x00\x00\x81\xF1\x00\x00\x82\xF1\x00\x00\x6C\x8F"
-     "\x00\x80"
-     "\x86\x90\x04\x00\x87\x64\x03\x00\x70\x8F\x00\x80\x74\x8F\x00\x80\xB8\x29"
-     "\x3D\x00"
-     "\xB9\x29\x3D\x00\xBA\x9D\x02\x00\xBB\x95\x02\x00\xBC\x8D\x02\x00\xBD\xAD"
-     "\x02\x00"
-     "\xBE\x7D\x02\x00\xBF\x0D\x02\x00\xB0\x55\x3D\x00\xB1\x5D\x3D\x00\xB2\x55"
-     "\x3D\x00"
-     "\xB3\x29\x3D\x00\xB4\x39\x3D\x00\xB5\x39\x3D\x00\xB6\x29\x3D\x00\xB7\x21"
-     "\x3D\x00"
-     "\xA3\x8D\x3C\x00\x78\x8F\x00\x80\x84\xC4\x04\x00\x7C\x8F\x00\x80\x80\x8F"
-     "\x00\x80"
-     "\xA6\x8D\x3C\x00\xA5\x8D\x3C\x00\x84\x8F\x00\x80\xAB\x99\x3C\x00\xAA\x91"
-     "\x3C\x00"
-     "\x88\x8F\x00\x80\x8C\x8F\x00\x80\xAF\xFD\x3C\x00\xAE\xFD\x3C\x00\xAD\xFD"
-     "\x3C\x00"
-     "\xAC\x81\x3C\x00\x90\x8F\x00\x80\xB3\x09\x02\x00\x94\x8F\x00\x80\x98\x8F"
-     "\x00\x80"
-     "\xB6\x21\x02\x00\x9C\x8F\x00\x80\xA0\x8F\x00\x80\xB5\x31\x02\x00\xBA\xC1"
-     "\x02\x00"
-     "\xBB\xC1\x02\x00\xA4\x8F\x00\x80\xA8\x8F\x00\x80\xBE\xC1\x02\x00\xBF\xC1"
-     "\x02\x00"
-     "\xBC\xD1\x02\x00\xBD\xD1\x02\x00\xAC\x8F\x00\x80\xB0\x8F\x00\x80\xA3\xC5"
-     "\x02\x00"
-     "\xB4\x8F\x00\x80\xA5\xFD\x02\x00\xB8\x8F\x00\x80\xBC\x8F\x00\x80\xA6\xED"
-     "\x02\x00"
-     "\xC0\x8F\x00\x80\xC4\x8F\x00\x80\xAB\x0D\x02\x00\xAA\x0D\x02\x00\xAD\x1D"
-     "\x02\x00"
-     "\xAC\x1D\x02\x00\xAF\x0D\x02\x00\xAE\x0D\x02\x00\xE3\x9C\x3E\x00\xE1\xA4"
-     "\x3F\x00"
-     "\xE1\x50\x3E\x00\xE3\xFC\x3E\x00\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D"
-     "\x00\x00"
-     "\xC8\x8F\x00\x80\xCC\x8F\x00\x80\xD0\x8F\x00\x80\xBE\x20\x04\x00\xD8\x8F"
-     "\x00\x80"
-     "\xEF\xA8\x3E\x00\xDC\x8F\x00\x80\xE0\x8F\x00\x80\xEF\x4C\x3E\x00\xA8\x0D"
-     "\x05\x00"
-     "\xA9\x4D\x05\x00\xAA\x45\x05\x00\xAB\x65\x05\x00\xAC\x7D\x05\x00\xAD\x65"
-     "\x05\x00"
-     "\xAE\x6D\x05\x00\xAF\x61\x05\x00\x64\x8F\x00\x80\xD4\x8F\x00\x80\x86\x08"
-     "\x03\x00"
-     "\x87\x8C\x03\x00\xE4\x8F\x00\x80\xE8\x8F\x00\x80\xEC\x8F\x00\x80\xF0\x8F"
-     "\x00\x80"
-     "\xB8\x45\x06\x00\xB9\x4D\x06\x00\xBA\x45\x06\x00\xBB\x5D\x06\x00\xBC\x4D"
-     "\x06\x00"
-     "\xBD\x7D\x06\x00\xBE\x75\x06\x00\xBF\x15\x06\x00\xB0\xC1\x05\x00\xB1\xC1"
-     "\x05\x00"
-     "\xB2\xC1\x05\x00\xB3\xC1\x05\x00\xB4\xC1\x05\x00\xB5\xC1\x05\x00\xB6\xC1"
-     "\x05\x00"
-     "\xB7\xC1\x05\x00\xA8\xED\x06\x00\xA9\x29\x06\x00\xAA\x31\x06\x00\xAB\x31"
-     "\x06\x00"
-     "\xAC\x95\x07\x00\xAD\x99\x07\x00\xAE\x89\x07\x00\xAF\x81\x07\x00\xF4\x8F"
-     "\x00\x80"
-     "\xF8\x8F\x00\x80\xF4\x8E\x00\x80\xFC\x8F\x00\x80\x00\x90\x00\x80\x04\x90"
-     "\x00\x80"
-     "\x08\x90\x00\x80\x0C\x90\x00\x80\xB8\xAD\x07\x00\xB9\x45\x07\x00\xBA\x4D"
-     "\x07\x00"
-     "\xBB\x41\x07\x00\xBC\x45\x07\x00\xBD\x4D\x07\x00\xBE\x79\x07\x00\xBF\x79"
-     "\x07\x00"
-     "\xB0\xC9\x07\x00\xB1\xC9\x07\x00\xB2\xA9\x07\x00\xB3\xA5\x07\x00\xB4\xBD"
-     "\x07\x00"
-     "\xB5\xA1\x07\x00\xB6\xA1\x07\x00\xB7\x95\x07\x00\xB3\xB9\x07\x00\x10\x90"
-     "\x00\x80"
-     "\x14\x90\x00\x80\x18\x90\x00\x80\x1C\x90\x00\x80\xB6\x19\x06\x00\xB5\x1D"
-     "\x06\x00"
-     "\x20\x90\x00\x80\xBB\x0D\x06\x00\xBA\x0D\x06\x00\x24\x90\x00\x80\x28\x90"
-     "\x00\x80"
-     "\xBF\x65\x06\x00\xBE\x79\x06\x00\xBD\x71\x06\x00\xBC\x7D\x06\x00\x2C\x90"
-     "\x00\x80"
-     "\xA3\xFD\x07\x00\x30\x90\x00\x80\x34\x90\x00\x80\xA6\x5D\x06\x00\x38\x90"
-     "\x00\x80"
-     "\x3C\x90\x00\x80\xA5\x59\x06\x00\xAA\x49\x06\x00\xAB\x49\x06\x00\x40\x90"
-     "\x00\x80"
-     "\x84\xA0\x01\x00\xAE\x3D\x06\x00\xAF\x21\x06\x00\xAC\x39\x06\x00\xAD\x35"
-     "\x06\x00"
-     "\xA8\x6D\x06\x00\xA9\x71\x06\x00\xAA\x71\x06\x00\xAB\xBD\x01\x00\xAC\xA5"
-     "\x01\x00"
-     "\xAD\xAD\x01\x00\xAE\xA5\x01\x00\xAF\x3D\x01\x00\x80\xF5\x00\x00\x81\xFD"
-     "\x00\x00"
-     "\x82\xC1\x00\x00\xBE\x64\x01\x00\x86\x98\x01\x00\x87\xAC\x00\x00\x48\x90"
-     "\x00\x80"
-     "\x4C\x90\x00\x80\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5\x00\x00\xBB\xE5"
-     "\x00\x00"
-     "\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x91\x00\x00\xBF\x91\x00\x00\xB0\x45"
-     "\x01\x00"
-     "\xB1\x49\x01\x00\xB2\x59\x01\x00\xB3\x59\x01\x00\xB4\x49\x01\x00\xB5\x49"
-     "\x01\x00"
-     "\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xB3\xF9\x05\x00\x50\x90\x00\x80\x54\x90"
-     "\x00\x80"
-     "\x58\x90\x00\x80\x5C\x90\x00\x80\xB6\x2D\x02\x00\xB5\x2D\x02\x00\x60\x90"
-     "\x00\x80"
-     "\xBB\x8D\x02\x00\xBA\x81\x02\x00\x64\x90\x00\x80\x68\x90\x00\x80\xBF\x8D"
-     "\x02\x00"
-     "\xBE\x8D\x02\x00\xBD\x8D\x02\x00\xBC\x8D\x02\x00\x6C\x90\x00\x80\xA3\xBD"
-     "\x05\x00"
-     "\x70\x90\x00\x80\x74\x90\x00\x80\xA6\x69\x02\x00\x78\x90\x00\x80\x7C\x90"
-     "\x00\x80"
-     "\xA5\x69\x02\x00\xAA\xC5\x02\x00\xAB\xC9\x02\x00\x80\x90\x00\x80\x84\x90"
-     "\x00\x80"
-     "\xAE\xC9\x02\x00\xAF\xC9\x02\x00\xAC\xC9\x02\x00\xAD\xC9\x02\x00\xAA\x85"
-     "\x00\x00"
-     "\xAB\x8D\x00\x00\xA8\x31\x03\x00\xA9\x31\x03\x00\xAE\x8D\x00\x00\xAF\xBD"
-     "\x00\x00"
-     "\xAC\x95\x00\x00\xAD\x8D\x00\x00\xBE\xE4\x03\x00\x88\x90\x00\x80\x8C\x90"
-     "\x00\x80"
-     "\x90\x90\x00\x80\x94\x90\x00\x80\x98\x90\x00\x80\x9C\x90\x00\x80\xA0\x90"
-     "\x00\x80"
-     "\xBA\x6D\x00\x00\xBB\x75\x00\x00\xB8\x6D\x00\x00\xB9\x65\x00\x00\xBE\x6D"
-     "\x00\x00"
-     "\xBF\xD5\x01\x00\xBC\x6D\x00\x00\xBD\x65\x00\x00\xB2\xA9\x00\x00\xB3\xA9"
-     "\x00\x00"
-     "\xB0\xC5\x00\x00\xB1\xAD\x00\x00\xB6\x99\x00\x00\xB7\x99\x00\x00\xB4\xAD"
-     "\x00\x00"
-     "\xB5\xA5\x00\x00\xA4\x90\x00\x80\xA8\x90\x00\x80\xAC\x90\x00\x80\xB0\x90"
-     "\x00\x80"
-     "\x82\x1D\x00\x00\xEF\x84\x07\x00\x80\x1D\x00\x00\x81\x15\x00\x00\x84\xAC"
-     "\x02\x00"
-     "\xE3\x74\x07\x00\xB8\x90\x00\x80\xE1\x1C\x04\x00\xBC\x90\x00\x80\xC0\x90"
-     "\x00\x80"
-     "\x86\xB8\x0D\x00\x87\x30\x0D\x00\xB3\x45\x01\x00\xC4\x90\x00\x80\xC8\x90"
-     "\x00\x80"
-     "\xCC\x90\x00\x80\xD0\x90\x00\x80\xB6\x45\x01\x00\xB5\x55\x01\x00\xD4\x90"
-     "\x00\x80"
-     "\xBB\x21\x01\x00\xBA\x49\x01\x00\xD8\x90\x00\x80\xDC\x90\x00\x80\xBF\x29"
-     "\x01\x00"
-     "\xBE\x21\x01\x00\xBD\x35\x01\x00\xBC\x35\x01\x00\xE0\x90\x00\x80\xE4\x90"
-     "\x00\x80"
-     "\xE8\x90\x00\x80\xEC\x90\x00\x80\xF0\x90\x00\x80\xF4\x90\x00\x80\xEF\xCC"
-     "\x00\x00"
-     "\xEF\x8C\x06\x00\xE1\x90\x01\x00\xE1\x0C\x07\x00\xE3\xD4\x00\x00\xE3\xB4"
-     "\x06\x00"
-     "\xF8\x90\x00\x80\xFC\x90\x00\x80\x00\x91\x00\x80\x04\x91\x00\x80\x08\x91"
-     "\x00\x80"
-     "\x0C\x91\x00\x80\xA3\xCD\x01\x00\x84\x90\x0D\x00\xA5\xDD\x01\x00\x10\x91"
-     "\x00\x80"
-     "\x14\x91\x00\x80\xA6\xCD\x01\x00\x18\x91\x00\x80\x1C\x91\x00\x80\xAB\xA9"
-     "\x01\x00"
-     "\xAA\xC1\x01\x00\xAD\xBD\x01\x00\xAC\xBD\x01\x00\xAF\xA1\x01\x00\xAE\xA9"
-     "\x01\x00"
-     "\xB4\x90\x00\x80\x82\x15\x00\x00\x81\xED\x0F\x00\x80\xED\x0F\x00\x20\x91"
-     "\x00\x80"
-     "\xBE\x58\x02\x00\x24\x91\x00\x80\x28\x91\x00\x80\xA8\x09\x0F\x00\xA9\x15"
-     "\x0F\x00"
-     "\xAA\x1D\x0F\x00\xAB\x69\x0F\x00\xAC\x79\x0F\x00\xAD\x79\x0F\x00\xAE\x69"
-     "\x0F\x00"
-     "\xAF\x61\x0F\x00\xB0\x29\x0F\x00\xB1\x29\x0F\x00\xB2\xC9\x0F\x00\xB3\xC9"
-     "\x0F\x00"
-     "\xB4\xD9\x0F\x00\xB5\xD9\x0F\x00\xB6\xC9\x0F\x00\xB7\xC9\x0F\x00\xB8\xD1"
-     "\x0F\x00"
-     "\xB9\xD1\x0F\x00\xBA\xD1\x0F\x00\xBB\xD1\x0F\x00\xBC\xB1\x0F\x00\xBD\xB1"
-     "\x0F\x00"
-     "\xBE\xAD\x0F\x00\xBF\xA5\x0F\x00\x2C\x91\x00\x80\xB6\xC9\x0F\x00\xB5\xC1"
-     "\x0F\x00"
-     "\x30\x91\x00\x80\xB3\xA5\x0F\x00\x34\x91\x00\x80\x87\x08\x03\x00\x86\x3C"
-     "\x03\x00"
-     "\xBF\x2D\x0F\x00\xBE\x2D\x0F\x00\xBD\x2D\x0F\x00\xBC\x2D\x0F\x00\xBB\xDD"
-     "\x0F\x00"
-     "\xBA\xDD\x0F\x00\x38\x91\x00\x80\x44\x90\x00\x80\xA3\xF5\x0F\x00\x3C\x91"
-     "\x00\x80"
-     "\x40\x91\x00\x80\x44\x91\x00\x80\x48\x91\x00\x80\xA6\x99\x0F\x00\xA5\x91"
-     "\x0F\x00"
-     "\x4C\x91\x00\x80\xAB\x8D\x0F\x00\xAA\x8D\x0F\x00\x50\x91\x00\x80\x54\x91"
-     "\x00\x80"
-     "\xAF\x7D\x0F\x00\xAE\x7D\x0F\x00\xAD\x7D\x0F\x00\xAC\x7D\x0F\x00\xA8\x1D"
-     "\x01\x00"
-     "\xA9\x85\x01\x00\xAA\x85\x01\x00\xAB\x99\x01\x00\xAC\xA9\x01\x00\xAD\xD1"
-     "\x01\x00"
-     "\xAE\xD1\x01\x00\xAF\xD1\x01\x00\x58\x91\x00\x80\x5C\x91\x00\x80\x60\x91"
-     "\x00\x80"
-     "\x64\x91\x00\x80\x68\x91\x00\x80\x6C\x91\x00\x80\x70\x91\x00\x80\x74\x91"
-     "\x00\x80"
-     "\xB8\x71\x00\x00\xB9\x71\x00\x00\xBA\x71\x00\x00\xBB\x71\x00\x00\xBC\xD1"
-     "\x00\x00"
-     "\xBD\xD1\x00\x00\xBE\xD1\x00\x00\xBF\xD1\x00\x00\xB0\xB1\x01\x00\xB1\xB9"
-     "\x01\x00"
-     "\xB2\x81\x01\x00\xB3\x81\x01\x00\xB4\x51\x00\x00\xB5\x51\x00\x00\xB6\x51"
-     "\x00\x00"
-     "\xB7\x51\x00\x00\x78\x91\x00\x80\x7C\x91\x00\x80\x80\x91\x00\x80\x84\x91"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\x88\x91\x00\x80\x8C\x91"
-     "\x00\x80"
-     "\x90\x91\x00\x80\x84\xA4\x02\x00\xBE\x24\x01\x00\x86\x90\x01\x00\x87\x14"
-     "\x01\x00"
-     "\x98\x91\x00\x80\x9C\x91\x00\x80\xB3\x4D\x01\x00\x94\x91\x00\x80\xA0\x91"
-     "\x00\x80"
-     "\xA4\x91\x00\x80\xA8\x91\x00\x80\xB6\x11\x01\x00\xB5\x61\x01\x00\xAC\x91"
-     "\x00\x80"
-     "\xBB\x3D\x01\x00\xBA\x35\x01\x00\xB0\x91\x00\x80\xB4\x91\x00\x80\xBF\xD5"
-     "\x00\x00"
-     "\xBE\xD5\x00\x00\xBD\x15\x01\x00\xBC\x15\x01\x00\xB8\x91\x00\x80\xA3\x89"
-     "\x02\x00"
-     "\xBC\x91\x00\x80\xBE\x40\x03\x00\xA6\xD5\x02\x00\xC0\x91\x00\x80\xC4\x91"
-     "\x00\x80"
-     "\xA5\xA5\x02\x00\xAA\xF1\x02\x00\xAB\xF9\x02\x00\xC8\x91\x00\x80\xCC\x91"
-     "\x00\x80"
-     "\xAE\x11\x03\x00\xAF\x11\x03\x00\xAC\xD1\x02\x00\xAD\xD1\x02\x00\xE1\xE8"
-     "\x01\x00"
-     "\xD0\x91\x00\x80\xE3\x14\x02\x00\x84\xD4\x03\x00\xD4\x91\x00\x80\xD8\x91"
-     "\x00\x80"
-     "\xDC\x91\x00\x80\x84\x10\x04\x00\xE0\x91\x00\x80\xE4\x91\x00\x80\xE8\x91"
-     "\x00\x80"
-     "\xEC\x91\x00\x80\xF0\x91\x00\x80\xF4\x91\x00\x80\xEF\x8C\x02\x00\xF8\x91"
-     "\x00\x80"
-     "\xA8\x39\x03\x00\xA9\x39\x03\x00\xAA\x41\x03\x00\xAB\x41\x03\x00\xAC\x41"
-     "\x03\x00"
-     "\xAD\x49\x03\x00\xAE\x71\x03\x00\xAF\x71\x03\x00\x80\xDD\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\xFC\x91\x00\x80\x00\x92\x00\x80\x08\x92\x00\x80\x87\x08"
-     "\x03\x00"
-     "\x86\x1C\x04\x00\xB8\x95\x00\x00\xB9\x9D\x00\x00\xBA\xA9\x00\x00\xBB\xA9"
-     "\x00\x00"
-     "\xBC\xB9\x00\x00\xBD\xB9\x00\x00\xBE\xA9\x00\x00\xBF\xA9\x00\x00\xB0\xE1"
-     "\x00\x00"
-     "\xB1\xED\x00\x00\xB2\xFD\x00\x00\xB3\xF5\x00\x00\xB4\xD9\x00\x00\xB5\xD9"
-     "\x00\x00"
-     "\xB6\xBD\x00\x00\xB7\xAD\x00\x00\xEF\x58\x0F\x00\x0C\x92\x00\x80\x10\x92"
-     "\x00\x80"
-     "\xBE\x44\x05\x00\x14\x92\x00\x80\x18\x92\x00\x80\x1C\x92\x00\x80\x20\x92"
-     "\x00\x80"
-     "\x24\x92\x00\x80\x28\x92\x00\x80\x2C\x92\x00\x80\x30\x92\x00\x80\xE3\x00"
-     "\x0E\x00"
-     "\x34\x92\x00\x80\xE1\x90\x0C\x00\x38\x92\x00\x80\xB6\x9D\x02\x00\x3C\x92"
-     "\x00\x80"
-     "\x40\x92\x00\x80\xB5\x9D\x02\x00\x44\x92\x00\x80\xB3\x9D\x02\x00\x48\x92"
-     "\x00\x80"
-     "\x4C\x92\x00\x80\xBE\x49\x01\x00\xBF\x49\x01\x00\xBC\x7D\x01\x00\xBD\x75"
-     "\x01\x00"
-     "\xBA\x71\x01\x00\xBB\x7D\x01\x00\x50\x92\x00\x80\x54\x92\x00\x80\x58\x92"
-     "\x00\x80"
-     "\xE1\xA4\x0E\x00\x5C\x92\x00\x80\xE3\x50\x0E\x00\xBE\xDC\x04\x00\xE3\x2C"
-     "\x00\x00"
-     "\x60\x92\x00\x80\xE1\x18\x01\x00\xBE\xEC\x06\x00\xEF\xF0\x00\x00\xBE\xE0"
-     "\x19\x00"
-     "\x64\x92\x00\x80\x80\x75\x00\x00\x81\x7D\x00\x00\x82\x75\x00\x00\xEF\x50"
-     "\x01\x00"
-     "\xA3\x09\x02\x00\x68\x92\x00\x80\x6C\x92\x00\x80\x70\x92\x00\x80\x78\x92"
-     "\x00\x80"
-     "\xA6\x09\x02\x00\xA5\x09\x02\x00\x7C\x92\x00\x80\xAB\xE9\x01\x00\xAA\xE5"
-     "\x01\x00"
-     "\x86\x48\x04\x00\x87\xDC\x04\x00\xAF\xDD\x01\x00\xAE\xDD\x01\x00\xAD\xE1"
-     "\x01\x00"
-     "\xAC\xE9\x01\x00\xA8\xB5\x06\x00\xA9\x01\x07\x00\xAA\x01\x07\x00\xAB\x1D"
-     "\x07\x00"
-     "\xAC\x05\x07\x00\xAD\x0D\x07\x00\xAE\x05\x07\x00\xAF\x3D\x07\x00\x74\x92"
-     "\x00\x80"
-     "\x80\x92\x00\x80\x84\x92\x00\x80\x88\x92\x00\x80\x8C\x92\x00\x80\x90\x92"
-     "\x00\x80"
-     "\x94\x92\x00\x80\x98\x92\x00\x80\xB8\x19\x07\x00\xB9\x19\x07\x00\xBA\x31"
-     "\x07\x00"
-     "\xBB\x31\x07\x00\xBC\xD1\x07\x00\xBD\xDD\x07\x00\xBE\xC1\x07\x00\xBF\xF1"
-     "\x07\x00"
-     "\xB0\x45\x07\x00\xB1\x4D\x07\x00\xB2\x45\x07\x00\xB3\x59\x07\x00\xB4\x4D"
-     "\x07\x00"
-     "\xB5\x31\x07\x00\xB6\x31\x07\x00\xB7\x2D\x07\x00\xB3\x89\x07\x00\x9C\x92"
-     "\x00\x80"
-     "\xA0\x92\x00\x80\x04\x92\x00\x80\xA4\x92\x00\x80\xB6\x8D\x07\x00\xB5\x8D"
-     "\x07\x00"
-     "\xA8\x92\x00\x80\xBB\x89\x07\x00\xBA\x81\x07\x00\xAC\x92\x00\x80\xB0\x92"
-     "\x00\x80"
-     "\xBF\xF5\x07\x00\xBE\x89\x07\x00\xBD\x8D\x07\x00\xBC\x91\x07\x00\xB4\x92"
-     "\x00\x80"
-     "\xA3\xCD\x07\x00\xB8\x92\x00\x80\xBC\x92\x00\x80\xA6\xC9\x07\x00\xC0\x92"
-     "\x00\x80"
-     "\xC4\x92\x00\x80\xA5\xC9\x07\x00\xAA\xC5\x07\x00\xAB\xCD\x07\x00\xC8\x92"
-     "\x00\x80"
-     "\xCC\x92\x00\x80\xAE\xCD\x07\x00\xAF\xB1\x07\x00\xAC\xD5\x07\x00\xAD\xC9"
-     "\x07\x00"
-     "\xA8\xD1\x06\x00\xA9\xD9\x06\x00\xAA\x2D\x01\x00\xAB\x25\x01\x00\xAC\x3D"
-     "\x01\x00"
-     "\xAD\x25\x01\x00\xAE\x2D\x01\x00\xAF\x25\x01\x00\xD0\x92\x00\x80\xD4\x92"
-     "\x00\x80"
-     "\xD8\x92\x00\x80\xDC\x92\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xE0\x92\x00\x80\xB8\x89\x01\x00\xB9\x89\x01\x00\xBA\x9D\x01\x00\xBB\x95"
-     "\x01\x00"
-     "\xBC\xB9\x01\x00\xBD\xB9\x01\x00\xBE\x79\x00\x00\xBF\x79\x00\x00\xB0\x5D"
-     "\x01\x00"
-     "\xB1\xE1\x01\x00\xB2\xE1\x01\x00\xB3\xFD\x01\x00\xB4\xED\x01\x00\xB5\xDD"
-     "\x01\x00"
-     "\xB6\xD5\x01\x00\xB7\xBD\x01\x00\x84\xC8\x02\x00\xB3\xA1\x01\x00\xE4\x92"
-     "\x00\x80"
-     "\xE8\x92\x00\x80\xB6\xA1\x01\x00\xEC\x92\x00\x80\xF0\x92\x00\x80\xB5\xB1"
-     "\x01\x00"
-     "\xBA\x9D\x01\x00\xBB\x21\x01\x00\x86\xA8\x01\x00\x87\xE0\x01\x00\xBE\x2D"
-     "\x01\x00"
-     "\xBF\x1D\x01\x00\xBC\x31\x01\x00\xBD\x2D\x01\x00\xA3\xE5\x01\x00\xF4\x92"
-     "\x00\x80"
-     "\xF8\x92\x00\x80\xFC\x92\x00\x80\x00\x93\x00\x80\xA6\xE5\x01\x00\xA5\xF5"
-     "\x01\x00"
-     "\x04\x93\x00\x80\xAB\x65\x01\x00\xAA\xD9\x01\x00\x08\x93\x00\x80\x0C\x93"
-     "\x00\x80"
-     "\xAF\x59\x01\x00\xAE\x69\x01\x00\xAD\x69\x01\x00\xAC\x75\x01\x00\x10\x93"
-     "\x00\x80"
-     "\xE3\x74\x02\x00\x14\x93\x00\x80\xE1\xBC\x01\x00\x18\x93\x00\x80\x1C\x93"
-     "\x00\x80"
-     "\x20\x93\x00\x80\x24\x93\x00\x80\x28\x93\x00\x80\x2C\x93\x00\x80\x30\x93"
-     "\x00\x80"
-     "\x34\x93\x00\x80\x38\x93\x00\x80\xEF\x6C\x07\x00\x3C\x93\x00\x80\x40\x93"
-     "\x00\x80"
-     "\x44\x93\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\x48\x93"
-     "\x00\x80"
-     "\x4C\x93\x00\x80\x50\x93\x00\x80\x58\x93\x00\x80\x84\xE4\x03\x00\x5C\x93"
-     "\x00\x80"
-     "\x87\x2C\x03\x00\x86\xCC\x1C\x00\x60\x93\x00\x80\x64\x93\x00\x80\x68\x93"
-     "\x00\x80"
-     "\x6C\x93\x00\x80\x70\x93\x00\x80\x74\x93\x00\x80\x78\x93\x00\x80\x7C\x93"
-     "\x00\x80"
-     "\x80\x93\x00\x80\x84\x93\x00\x80\x88\x93\x00\x80\x8C\x93\x00\x80\xA8\x2D"
-     "\x03\x00"
-     "\xA9\x31\x03\x00\xAA\x31\x03\x00\xAB\x69\x03\x00\xAC\x79\x03\x00\xAD\x79"
-     "\x03\x00"
-     "\xAE\x69\x03\x00\xAF\x61\x03\x00\xB0\x29\x03\x00\xB1\x29\x03\x00\xB2\xC9"
-     "\x00\x00"
-     "\xB3\xC5\x00\x00\xB4\xC1\x00\x00\xB5\xC1\x00\x00\xB6\xC1\x00\x00\xB7\xC1"
-     "\x00\x00"
-     "\xB8\xC1\x00\x00\xB9\xC1\x00\x00\xBA\xC1\x00\x00\xBB\xC1\x00\x00\xBC\xC1"
-     "\x00\x00"
-     "\xBD\xCD\x00\x00\xBE\xC5\x00\x00\xBF\xBD\x00\x00\xE3\x4C\x07\x00\x90\x93"
-     "\x00\x80"
-     "\xE1\xAC\x07\x00\x94\x93\x00\x80\x98\x93\x00\x80\x9C\x93\x00\x80\xA0\x93"
-     "\x00\x80"
-     "\xA4\x93\x00\x80\xA8\x93\x00\x80\xAC\x93\x00\x80\xB0\x93\x00\x80\xB4\x93"
-     "\x00\x80"
-     "\xEF\xC4\x06\x00\xB8\x93\x00\x80\xBC\x93\x00\x80\xC0\x93\x00\x80\x81\x79"
-     "\x00\x00"
-     "\x80\x71\x00\x00\xB3\xD1\x02\x00\x82\x71\x00\x00\xB5\x99\x02\x00\xC8\x93"
-     "\x00\x80"
-     "\x84\x14\x1D\x00\xB6\x91\x02\x00\x86\xC0\x1C\x00\x87\xE8\x1D\x00\xBB\x85"
-     "\x02\x00"
-     "\xBA\x85\x02\x00\xBD\x79\x01\x00\xBC\x75\x01\x00\xBF\x6D\x01\x00\xBE\x6D"
-     "\x01\x00"
-     "\xCC\x93\x00\x80\xD0\x93\x00\x80\xD4\x93\x00\x80\xD8\x93\x00\x80\x84\xDC"
-     "\x1C\x00"
-     "\xDC\x93\x00\x80\xE0\x93\x00\x80\xEF\xF4\x06\x00\xBE\xE8\x1C\x00\xE1\x54"
-     "\x06\x00"
-     "\xE4\x93\x00\x80\xE3\x3C\x00\x00\x84\xF0\x1F\x00\xE8\x93\x00\x80\xEC\x93"
-     "\x00\x80"
-     "\xF0\x93\x00\x80\xA3\x59\x02\x00\xF4\x93\x00\x80\xF8\x93\x00\x80\xFC\x93"
-     "\x00\x80"
-     "\x00\x94\x00\x80\xA6\x19\x02\x00\xA5\x11\x02\x00\x04\x94\x00\x80\xAB\x0D"
-     "\x02\x00"
-     "\xAA\x0D\x02\x00\x08\x94\x00\x80\x0C\x94\x00\x80\xAF\xE5\x01\x00\xAE\xE5"
-     "\x01\x00"
-     "\xAD\xF1\x01\x00\xAC\xFD\x01\x00\xA8\x5D\x1E\x00\xA9\x61\x1E\x00\xAA\x79"
-     "\x1E\x00"
-     "\xAB\x79\x1E\x00\xAC\x69\x1E\x00\xAD\x69\x1E\x00\xAE\x99\x1E\x00\xAF\x99"
-     "\x1E\x00"
-     "\xC4\x93\x00\x80\x10\x94\x00\x80\x14\x94\x00\x80\x18\x94\x00\x80\x1C\x94"
-     "\x00\x80"
-     "\x20\x94\x00\x80\x24\x94\x00\x80\x28\x94\x00\x80\xB8\x51\x1F\x00\xB9\x51"
-     "\x1F\x00"
-     "\xBA\x51\x1F\x00\xBB\x51\x1F\x00\xBC\x71\x1F\x00\xBD\x71\x1F\x00\xBE\x71"
-     "\x1F\x00"
-     "\xBF\x71\x1F\x00\xB0\xE9\x1E\x00\xB1\xE9\x1E\x00\xB2\xF9\x1E\x00\xB3\xF9"
-     "\x1E\x00"
-     "\xB4\xE9\x1E\x00\xB5\xE9\x1E\x00\xB6\x7D\x1F\x00\xB7\x71\x1F\x00\xB3\xF1"
-     "\x1E\x00"
-     "\x82\x2D\x00\x00\x81\x15\x00\x00\x80\xB1\x00\x00\x54\x93\x00\x80\xB6\xE9"
-     "\x1E\x00"
-     "\xB5\xE1\x1E\x00\x2C\x94\x00\x80\xBB\x2D\x1E\x00\xBA\x2D\x1E\x00\x30\x94"
-     "\x00\x80"
-     "\x34\x94\x00\x80\xBF\x2D\x1E\x00\xBE\x2D\x1E\x00\xBD\x2D\x1E\x00\xBC\x2D"
-     "\x1E\x00"
-     "\x38\x94\x00\x80\xA3\xB5\x1E\x00\x86\x68\x00\x00\x87\x4C\x00\x00\xA6\xAD"
-     "\x1E\x00"
-     "\x3C\x94\x00\x80\x40\x94\x00\x80\xA5\xA5\x1E\x00\xAA\x69\x1E\x00\xAB\x69"
-     "\x1E\x00"
-     "\x44\x94\x00\x80\x48\x94\x00\x80\xAE\x69\x1E\x00\xAF\x69\x1E\x00\xAC\x69"
-     "\x1E\x00"
-     "\xAD\x69\x1E\x00\xA8\x55\x1E\x00\xA9\x85\x1E\x00\xAA\x8D\x1E\x00\xAB\x85"
-     "\x1E\x00"
-     "\xAC\x9D\x1E\x00\xAD\x81\x1E\x00\xAE\x81\x1E\x00\xAF\xB5\x1E\x00\x4C\x94"
-     "\x00\x80"
-     "\x50\x94\x00\x80\x54\x94\x00\x80\x58\x94\x00\x80\x5C\x94\x00\x80\x60\x94"
-     "\x00\x80"
-     "\x64\x94\x00\x80\x68\x94\x00\x80\xB8\xA9\x1E\x00\xB9\xA9\x1E\x00\xBA\x6D"
-     "\x01\x00"
-     "\xBB\x65\x01\x00\xBC\x7D\x01\x00\xBD\x65\x01\x00\xBE\x6D\x01\x00\xBF\x65"
-     "\x01\x00"
-     "\xB0\xD5\x1E\x00\xB1\xDD\x1E\x00\xB2\xD5\x1E\x00\xB3\xA9\x1E\x00\xB4\xB9"
-     "\x1E\x00"
-     "\xB5\xB9\x1E\x00\xB6\xA9\x1E\x00\xB7\xA1\x1E\x00\xB3\x05\x1E\x00\x6C\x94"
-     "\x00\x80"
-     "\x70\x94\x00\x80\x74\x94\x00\x80\x78\x94\x00\x80\xB6\x01\x1E\x00\xB5\x15"
-     "\x1E\x00"
-     "\x7C\x94\x00\x80\xBB\x19\x1E\x00\xBA\x11\x1E\x00\x80\x94\x00\x80\x84\x94"
-     "\x00\x80"
-     "\xBF\x7D\x1E\x00\xBE\x7D\x1E\x00\xBD\x7D\x1E\x00\xBC\x01\x1E\x00\x88\x94"
-     "\x00\x80"
-     "\xA3\x41\x1E\x00\x8C\x94\x00\x80\x90\x94\x00\x80\xA6\x45\x1E\x00\x94\x94"
-     "\x00\x80"
-     "\x98\x94\x00\x80\xA5\x51\x1E\x00\xAA\x55\x1E\x00\xAB\x5D\x1E\x00\x9C\x94"
-     "\x00\x80"
-     "\xA0\x94\x00\x80\xAE\x39\x1E\x00\xAF\x39\x1E\x00\xAC\x45\x1E\x00\xAD\x39"
-     "\x1E\x00"
-     "\xA8\xC1\x02\x00\xA9\xD1\x02\x00\xAA\xD1\x02\x00\xAB\xE5\x02\x00\xAC\xFD"
-     "\x02\x00"
-     "\xAD\x35\x03\x00\xAE\x3D\x03\x00\xAF\x35\x03\x00\xA8\x94\x00\x80\xBE\x68"
-     "\x02\x00"
-     "\xAC\x94\x00\x80\xB0\x94\x00\x80\xB4\x94\x00\x80\xB8\x94\x00\x80\xBC\x94"
-     "\x00\x80"
-     "\xC0\x94\x00\x80\xB8\xD1\x03\x00\xB9\xD9\x03\x00\xBA\xE1\x03\x00\xBB\xE1"
-     "\x03\x00"
-     "\xBC\x91\x03\x00\xBD\x9D\x03\x00\xBE\x95\x03\x00\xBF\x89\x03\x00\xB0\x4D"
-     "\x03\x00"
-     "\xB1\x55\x03\x00\xB2\x5D\x03\x00\xB3\x55\x03\x00\xB4\x4D\x03\x00\xB5\xF1"
-     "\x03\x00"
-     "\xB6\xF1\x03\x00\xB7\xF1\x03\x00\x80\x59\x03\x00\x81\x6D\x03\x00\x82\x65"
-     "\x03\x00"
-     "\xB3\x55\x02\x00\xC4\x94\x00\x80\xB5\x79\x02\x00\xB6\xB9\x03\x00\xC8\x94"
-     "\x00\x80"
-     "\x86\x00\x04\x00\x87\xE4\x03\x00\xBA\x85\x03\x00\xBB\x9D\x03\x00\xBC\x85"
-     "\x03\x00"
-     "\xBD\x8D\x03\x00\xBE\x85\x03\x00\xBF\x39\x03\x00\xA3\x19\x02\x00\xCC\x94"
-     "\x00\x80"
-     "\xD0\x94\x00\x80\xD4\x94\x00\x80\x85\xAC\x0B\x00\xA6\xF5\x03\x00\xA5\x35"
-     "\x02\x00"
-     "\xD8\x94\x00\x80\xAB\xD1\x03\x00\xAA\xC9\x03\x00\xDC\x94\x00\x80\xE0\x94"
-     "\x00\x80"
-     "\xAF\x75\x03\x00\xAE\xC9\x03\x00\xAD\xC1\x03\x00\xAC\xC9\x03\x00\xE4\x94"
-     "\x00\x80"
-     "\xE8\x94\x00\x80\xEC\x94\x00\x80\xF0\x94\x00\x80\xEF\xCC\x03\x00\xE1\x60"
-     "\x1F\x00"
-     "\xF4\x94\x00\x80\xE3\x54\x1E\x00\xE3\x80\x1E\x00\xF8\x94\x00\x80\xE1\xF8"
-     "\x01\x00"
-     "\xEF\x7C\x1A\x00\xFC\x94\x00\x80\x00\x95\x00\x80\x04\x95\x00\x80\x08\x95"
-     "\x00\x80"
-     "\x80\x7D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\x84\x40\x02\x00\x0C\x95"
-     "\x00\x80"
-     "\x10\x95\x00\x80\xBE\x54\x05\x00\x18\x95\x00\x80\x86\xC0\x04\x00\x87\x48"
-     "\x05\x00"
-     "\x1C\x95\x00\x80\x20\x95\x00\x80\x24\x95\x00\x80\x28\x95\x00\x80\x2C\x95"
-     "\x00\x80"
-     "\x30\x95\x00\x80\x34\x95\x00\x80\x38\x95\x00\x80\x3C\x95\x00\x80\x40\x95"
-     "\x00\x80"
-     "\x44\x95\x00\x80\x48\x95\x00\x80\x4C\x95\x00\x80\xEF\xC4\x1E\x00\x84\xA4"
-     "\x05\x00"
-     "\xE1\xDC\x1E\x00\x50\x95\x00\x80\xE3\x2C\x1E\x00\x54\x95\x00\x80\x58\x95"
-     "\x00\x80"
-     "\x5C\x95\x00\x80\x60\x95\x00\x80\xB3\xED\x01\x00\x64\x95\x00\x80\x68\x95"
-     "\x00\x80"
-     "\x6C\x95\x00\x80\x70\x95\x00\x80\xB6\xED\x01\x00\xB5\xED\x01\x00\x74\x95"
-     "\x00\x80"
-     "\xBB\x25\x01\x00\xBA\x79\x01\x00\x78\x95\x00\x80\x7C\x95\x00\x80\xBF\x19"
-     "\x01\x00"
-     "\xBE\x29\x01\x00\xBD\x35\x01\x00\xBC\x35\x01\x00\xA8\x21\x06\x00\xA9\x2D"
-     "\x06\x00"
-     "\xAA\x25\x06\x00\xAB\x3D\x06\x00\xAC\x25\x06\x00\xAD\x2D\x06\x00\xAE\x25"
-     "\x06\x00"
-     "\xAF\x8D\x06\x00\x14\x95\x00\x80\x82\x15\x00\x00\x81\xC5\x07\x00\x80\xC5"
-     "\x07\x00"
-     "\x80\x95\x00\x80\xBE\x38\x03\x00\x84\x95\x00\x80\x88\x95\x00\x80\xB8\x45"
-     "\x07\x00"
-     "\xB9\x4D\x07\x00\xBA\x45\x07\x00\xBB\x55\x07\x00\xBC\x75\x07\x00\xBD\x65"
-     "\x07\x00"
-     "\xBE\x6D\x07\x00\xBF\x65\x07\x00\xB0\xFD\x06\x00\xB1\xA1\x06\x00\xB2\xA1"
-     "\x06\x00"
-     "\xB3\xA1\x06\x00\xB4\xAD\x06\x00\xB5\x9D\x06\x00\xB6\x95\x06\x00\xB7\x7D"
-     "\x07\x00"
-     "\xA3\xAD\x06\x00\x8C\x95\x00\x80\x86\x28\x00\x00\x87\x4C\x03\x00\x90\x95"
-     "\x00\x80"
-     "\xA6\xAD\x06\x00\xA5\xAD\x06\x00\x94\x95\x00\x80\xAB\x65\x06\x00\xAA\x39"
-     "\x06\x00"
-     "\x98\x95\x00\x80\x9C\x95\x00\x80\xAF\x59\x06\x00\xAE\x69\x06\x00\xAD\x75"
-     "\x06\x00"
-     "\xAC\x75\x06\x00\xA0\x95\x00\x80\xA4\x95\x00\x80\xB3\xA5\x07\x00\xA8\x95"
-     "\x00\x80"
-     "\xB5\xA5\x07\x00\xAC\x95\x00\x80\xB0\x95\x00\x80\xB6\xA5\x07\x00\xB4\x95"
-     "\x00\x80"
-     "\xA4\x94\x00\x80\xBB\x61\x07\x00\xBA\x69\x07\x00\xBD\x61\x07\x00\xBC\x79"
-     "\x07\x00"
-     "\xBF\x61\x07\x00\xBE\x69\x07\x00\xA8\x5D\x06\x00\xA9\x6D\x06\x00\xAA\x65"
-     "\x06\x00"
-     "\xAB\x7D\x06\x00\xAC\x65\x06\x00\xAD\x6D\x06\x00\xAE\x65\x06\x00\xAF\x9D"
-     "\x06\x00"
-     "\xB8\x95\x00\x80\xBC\x95\x00\x80\xC0\x95\x00\x80\xC4\x95\x00\x80\xC8\x95"
-     "\x00\x80"
-     "\xCC\x95\x00\x80\xD0\x95\x00\x80\xD4\x95\x00\x80\xB8\xF5\x06\x00\xB9\xFD"
-     "\x06\x00"
-     "\xBA\xF5\x06\x00\xBB\x89\x06\x00\xBC\x99\x06\x00\xBD\x99\x06\x00\xBE\x89"
-     "\x06\x00"
-     "\xBF\x81\x06\x00\xB0\xE5\x06\x00\xB1\xED\x06\x00\xB2\xE5\x06\x00\xB3\xFD"
-     "\x06\x00"
-     "\xB4\xE5\x06\x00\xB5\xED\x06\x00\xB6\xD1\x06\x00\xB7\xCD\x06\x00\xA3\xED"
-     "\x06\x00"
-     "\xD8\x95\x00\x80\xDC\x95\x00\x80\xE0\x95\x00\x80\xE4\x95\x00\x80\xA6\xED"
-     "\x06\x00"
-     "\xA5\xED\x06\x00\xE8\x95\x00\x80\xAB\x29\x06\x00\xAA\x21\x06\x00\xEC\x95"
-     "\x00\x80"
-     "\xF0\x95\x00\x80\xAF\x29\x06\x00\xAE\x21\x06\x00\xAD\x29\x06\x00\xAC\x31"
-     "\x06\x00"
-     "\x80\x5D\x00\x00\x81\x6D\x00\x00\x82\x65\x00\x00\x84\x80\x01\x00\xBE\x98"
-     "\x01\x00"
-     "\xF8\x95\x00\x80\x87\xD8\x00\x00\x86\x7C\x0F\x00\xFC\x95\x00\x80\x00\x96"
-     "\x00\x80"
-     "\x04\x96\x00\x80\x08\x96\x00\x80\x0C\x96\x00\x80\x10\x96\x00\x80\x14\x96"
-     "\x00\x80"
-     "\x18\x96\x00\x80\xA8\x3D\x02\x00\xA9\x81\x02\x00\xAA\x99\x02\x00\xAB\x91"
-     "\x02\x00"
-     "\xAC\xB9\x02\x00\xAD\xB9\x02\x00\xAE\xD9\x02\x00\xAF\xD9\x02\x00\x1C\x96"
-     "\x00\x80"
-     "\x20\x96\x00\x80\x24\x96\x00\x80\xBF\xA4\x0E\x00\x28\x96\x00\x80\x2C\x96"
-     "\x00\x80"
-     "\x30\x96\x00\x80\x34\x96\x00\x80\xB8\x71\x03\x00\xB9\x71\x03\x00\xBA\x71"
-     "\x03\x00"
-     "\xBB\x71\x03\x00\xBC\xD5\x03\x00\xBD\xDD\x03\x00\xBE\xD5\x03\x00\xBF\xCD"
-     "\x03\x00"
-     "\xB0\xAD\x02\x00\xB1\xB1\x02\x00\xB2\xB1\x02\x00\xB3\x8D\x02\x00\xB4\x51"
-     "\x03\x00"
-     "\xB5\x51\x03\x00\xB6\x51\x03\x00\xB7\x51\x03\x00\xB3\x11\x02\x00\x38\x96"
-     "\x00\x80"
-     "\x3C\x96\x00\x80\x40\x96\x00\x80\x84\x50\x03\x00\xB6\x39\x02\x00\xB5\x31"
-     "\x02\x00"
-     "\x44\x96\x00\x80\xBB\x75\x02\x00\xBA\x69\x02\x00\x48\x96\x00\x80\x4C\x96"
-     "\x00\x80"
-     "\xBF\x95\x03\x00\xBE\x95\x03\x00\xBD\x59\x02\x00\xBC\x65\x02\x00\x50\x96"
-     "\x00\x80"
-     "\xA3\x55\x02\x00\x54\x96\x00\x80\xBE\x80\x03\x00\xA6\x7D\x02\x00\x58\x96"
-     "\x00\x80"
-     "\x5C\x96\x00\x80\xA5\x75\x02\x00\xAA\x2D\x02\x00\xAB\x31\x02\x00\x60\x96"
-     "\x00\x80"
-     "\x68\x96\x00\x80\xAE\xD1\x03\x00\xAF\xD1\x03\x00\xAC\x21\x02\x00\xAD\x1D"
-     "\x02\x00"
-     "\x80\xD5\x01\x00\x81\xDD\x01\x00\x82\xE1\x01\x00\xB3\xE9\x02\x00\x6C\x96"
-     "\x00\x80"
-     "\xB5\xF9\x02\x00\xB6\xE9\x02\x00\x70\x96\x00\x80\x74\x96\x00\x80\x78\x96"
-     "\x00\x80"
-     "\xBA\xD5\x02\x00\xBB\x0D\x01\x00\xBC\x15\x01\x00\xBD\x1D\x01\x00\xBE\x09"
-     "\x01\x00"
-     "\xBF\x09\x01\x00\x86\x20\x0D\x00\x87\x88\x02\x00\x7C\x96\x00\x80\x80\x96"
-     "\x00\x80"
-     "\x84\x96\x00\x80\x88\x96\x00\x80\xEF\xF8\x06\x00\x8C\x96\x00\x80\xE1\xD8"
-     "\x07\x00"
-     "\x90\x96\x00\x80\xE3\x34\x06\x00\x94\x96\x00\x80\x98\x96\x00\x80\x9C\x96"
-     "\x00\x80"
-     "\xA0\x96\x00\x80\xA4\x96\x00\x80\xA8\x96\x00\x80\xEF\xAC\x00\x00\xAC\x96"
-     "\x00\x80"
-     "\xB0\x96\x00\x80\xB4\x96\x00\x80\xB8\x96\x00\x80\xBC\x96\x00\x80\xEF\x50"
-     "\x01\x00"
-     "\x84\x6C\x0C\x00\xE1\xDC\x06\x00\xC0\x96\x00\x80\xE3\x1C\x06\x00\xC4\x96"
-     "\x00\x80"
-     "\xE3\x68\x00\x00\xC8\x96\x00\x80\xE1\x2C\x01\x00\xA3\x25\x02\x00\x82\x2D"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x80\x75\x00\x00\xCC\x96\x00\x80\xA6\x25\x02\x00\xA5\x35"
-     "\x02\x00"
-     "\x85\x00\x08\x00\xAB\xC1\x01\x00\xAA\x19\x02\x00\xD0\x96\x00\x80\xBE\x24"
-     "\x0C\x00"
-     "\xAF\xC5\x01\x00\xAE\xC5\x01\x00\xAD\xD1\x01\x00\xAC\xD9\x01\x00\xA8\xD5"
-     "\x0D\x00"
-     "\xA9\xE5\x0D\x00\xAA\xED\x0D\x00\xAB\xE5\x0D\x00\xAC\xFD\x0D\x00\xAD\xE5"
-     "\x0D\x00"
-     "\xAE\xED\x0D\x00\xAF\xE5\x0D\x00\x64\x96\x00\x80\xD4\x96\x00\x80\x86\x08"
-     "\x03\x00"
-     "\x87\x8C\x03\x00\xD8\x96\x00\x80\xDC\x96\x00\x80\xE0\x96\x00\x80\xE4\x96"
-     "\x00\x80"
-     "\xB8\xC5\x0E\x00\xB9\xCD\x0E\x00\xBA\xC5\x0E\x00\xBB\xD1\x0E\x00\xBC\xF9"
-     "\x0E\x00"
-     "\xBD\xF9\x0E\x00\xBE\x99\x0E\x00\xBF\x95\x0E\x00\xB0\x9D\x0D\x00\xB1\x45"
-     "\x0E\x00"
-     "\xB2\x4D\x0E\x00\xB3\x45\x0E\x00\xB4\x41\x0E\x00\xB5\x41\x0E\x00\xB6\x41"
-     "\x0E\x00"
-     "\xB7\x41\x0E\x00\xB8\x01\x0F\x00\xB9\x0D\x0F\x00\xBA\x05\x0F\x00\xBB\x1D"
-     "\x0F\x00"
-     "\xBC\x05\x0F\x00\xBD\x0D\x0F\x00\xBE\x31\x0F\x00\xBF\x2D\x0F\x00\xB0\xF5"
-     "\x0F\x00"
-     "\xB1\xF9\x0F\x00\xB2\xCD\x0F\x00\xB3\xC1\x0F\x00\xB4\xC1\x0F\x00\xB5\xC1"
-     "\x0F\x00"
-     "\xB6\xC1\x0F\x00\xB7\xC1\x0F\x00\xA8\x2D\x0E\x00\xA9\xD5\x0F\x00\xAA\xD5"
-     "\x0F\x00"
-     "\xAB\xED\x0F\x00\xAC\xF5\x0F\x00\xAD\x9D\x0F\x00\xAE\x95\x0F\x00\xAF\x8D"
-     "\x0F\x00"
-     "\xE8\x96\x00\x80\xEC\x96\x00\x80\xF4\x95\x00\x80\xF0\x96\x00\x80\xF4\x96"
-     "\x00\x80"
-     "\xF8\x96\x00\x80\xFC\x96\x00\x80\x00\x97\x00\x80\xB3\x1D\x0E\x00\x04\x97"
-     "\x00\x80"
-     "\x08\x97\x00\x80\x0C\x97\x00\x80\x10\x97\x00\x80\xB6\x5D\x0E\x00\xB5\x3D"
-     "\x0E\x00"
-     "\x14\x97\x00\x80\xBB\x7D\x0E\x00\xBA\x71\x0E\x00\x18\x97\x00\x80\x1C\x97"
-     "\x00\x80"
-     "\xBF\x45\x0E\x00\xBE\x45\x0E\x00\xBD\x55\x0E\x00\xBC\x59\x0E\x00\x20\x97"
-     "\x00\x80"
-     "\xA3\x59\x0E\x00\x24\x97\x00\x80\x28\x97\x00\x80\xA6\x19\x0E\x00\x2C\x97"
-     "\x00\x80"
-     "\x30\x97\x00\x80\xA5\x79\x0E\x00\xAA\x35\x0E\x00\xAB\x39\x0E\x00\x34\x97"
-     "\x00\x80"
-     "\x38\x97\x00\x80\xAE\x01\x0E\x00\xAF\x01\x0E\x00\xAC\x1D\x0E\x00\xAD\x11"
-     "\x0E\x00"
-     "\xA8\xB1\x01\x00\xA9\xB1\x01\x00\xAA\xB1\x01\x00\xAB\xB1\x01\x00\xAC\x71"
-     "\x01\x00"
-     "\xAD\x71\x01\x00\xAE\x75\x01\x00\xAF\x6D\x01\x00\x80\xCD\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x3C\x97\x00\x80\x40\x97\x00\x80\x84\x78\x01\x00\xBE\x70"
-     "\x01\x00"
-     "\x48\x97\x00\x80\xB8\xE9\x00\x00\xB9\xE9\x00\x00\xBA\x89\x00\x00\xBB\x85"
-     "\x00\x00"
-     "\xBC\x9D\x00\x00\xBD\x81\x00\x00\xBE\x81\x00\x00\xBF\xB5\x00\x00\xB0\x15"
-     "\x01\x00"
-     "\xB1\x1D\x01\x00\xB2\x15\x01\x00\xB3\xE9\x00\x00\xB4\xF9\x00\x00\xB5\xF9"
-     "\x00\x00"
-     "\xB6\xE9\x00\x00\xB7\xE1\x00\x00\xB3\x39\x02\x00\x4C\x97\x00\x80\x86\x88"
-     "\x04\x00"
-     "\x87\x78\x03\x00\x50\x97\x00\x80\xB6\xF1\x02\x00\xB5\xF9\x02\x00\x54\x97"
-     "\x00\x80"
-     "\xBB\xD9\x02\x00\xBA\xD1\x02\x00\x58\x97\x00\x80\x5C\x97\x00\x80\xBF\x3D"
-     "\x03\x00"
-     "\xBE\x3D\x03\x00\xBD\x3D\x03\x00\xBC\xC1\x02\x00\x60\x97\x00\x80\xA3\x7D"
-     "\x02\x00"
-     "\x64\x97\x00\x80\x68\x97\x00\x80\xA6\xB5\x02\x00\x6C\x97\x00\x80\x70\x97"
-     "\x00\x80"
-     "\xA5\xBD\x02\x00\xAA\x95\x02\x00\xAB\x9D\x02\x00\x74\x97\x00\x80\x78\x97"
-     "\x00\x80"
-     "\xAE\x79\x03\x00\xAF\x79\x03\x00\xAC\x85\x02\x00\xAD\x79\x03\x00\x82\xC1"
-     "\x1A\x00"
-     "\x83\xC9\x1A\x00\x7C\x97\x00\x80\x80\x97\x00\x80\x86\xB5\x1E\x00\x87\x29"
-     "\x12\x00"
-     "\x84\x31\x1E\x00\x85\xBD\x1E\x00\x8A\xA1\x12\x00\x8B\xBD\x12\x00\x84\x97"
-     "\x00\x80"
-     "\x88\x97\x00\x80\x8E\x91\x16\x00\x8F\x3D\x17\x00\x8C\x7D\x13\x00\x8D\x31"
-     "\x16\x00"
-     "\x92\x51\x0A\x00\x93\xE5\x0A\x00\x8C\x97\x00\x80\xEF\xDC\x6C\x00\x96\xF1"
-     "\x0E\x00"
-     "\x97\xED\x0E\x00\x94\x75\x0A\x00\x95\x39\x0E\x00\x9A\x09\x02\x00\x9B\xC5"
-     "\x02\x00"
-     "\x90\x97\x00\x80\x94\x97\x00\x80\x98\x97\x00\x80\xE1\x3C\x0F\x00\x9C\xDD"
-     "\x02\x00"
-     "\xE3\xA8\x76\x00\x9F\x21\x0F\x00\x9E\xE1\x0F\x00\x9D\xA5\x0B\x00\x9C\x2D"
-     "\x0B\x00"
-     "\x9B\x81\x0B\x00\x9A\x9D\x16\x00\x99\x39\x17\x00\x98\x91\x17\x00\x97\xE5"
-     "\x12\x00"
-     "\x96\x3D\x13\x00\x95\x29\x13\x00\x94\x91\x13\x00\x93\x15\x1F\x00\x92\x15"
-     "\x1F\x00"
-     "\x91\x91\x1F\x00\x90\x71\x1B\x00\x8F\x01\x1B\x00\x9C\x97\x00\x80\xA0\x97"
-     "\x00\x80"
-     "\xA4\x97\x00\x80\xA8\x97\x00\x80\xEF\x60\x0F\x00\xAC\x97\x00\x80\xB0\x97"
-     "\x00\x80"
-     "\xBE\x28\x05\x00\xE3\xE4\x0F\x00\x84\x24\x05\x00\xE1\xA8\x0C\x00\x81\x29"
-     "\x00\x00"
-     "\x80\x25\x00\x00\xB8\x97\x00\x80\x82\x21\x00\x00\x87\x0C\x05\x00\x86\xA4"
-     "\x04\x00"
-     "\xBC\x97\x00\x80\xC0\x97\x00\x80\xC4\x97\x00\x80\xC8\x97\x00\x80\xEF\xD8"
-     "\x00\x00"
-     "\xEF\x88\x0E\x00\xE1\x90\x01\x00\xE1\xFC\x0E\x00\xE3\x0C\x00\x00\xE3\x1C"
-     "\x0E\x00"
-     "\xCC\x97\x00\x80\xD0\x97\x00\x80\xD4\x97\x00\x80\xD8\x97\x00\x80\xDC\x97"
-     "\x00\x80"
-     "\xE0\x97\x00\x80\x84\x04\x07\x00\xBE\x04\x07\x00\xE4\x97\x00\x80\xE8\x97"
-     "\x00\x80"
-     "\xEC\x97\x00\x80\xF0\x97\x00\x80\xF4\x97\x00\x80\xF8\x97\x00\x80\xFC\x97"
-     "\x00\x80"
-     "\x00\x98\x00\x80\xA3\xD5\x03\x00\xA2\xD5\x03\x00\xA1\x11\x03\x00\xA0\xC5"
-     "\x0F\x00"
-     "\x44\x97\x00\x80\xB4\x97\x00\x80\x04\x98\x00\x80\x08\x98\x00\x80\x0C\x98"
-     "\x00\x80"
-     "\x10\x98\x00\x80\x14\x98\x00\x80\x18\x98\x00\x80\xA8\x9D\x05\x00\xA9\x21"
-     "\x06\x00"
-     "\xAA\x39\x06\x00\xAB\x31\x06\x00\xAC\x19\x06\x00\xAD\x19\x06\x00\xAE\x79"
-     "\x06\x00"
-     "\xAF\x75\x06\x00\xB0\x0D\x06\x00\xB1\x11\x06\x00\xB2\x11\x06\x00\xB3\x25"
-     "\x06\x00"
-     "\xB4\x3D\x06\x00\xB5\xDD\x06\x00\xB6\xD5\x06\x00\xB7\xF5\x06\x00\xB8\xCD"
-     "\x06\x00"
-     "\xB9\xAD\x06\x00\xBA\x81\x06\x00\xBB\x81\x06\x00\xBC\x81\x06\x00\xBD\x31"
-     "\x07\x00"
-     "\xBE\x31\x07\x00\xBF\x31\x07\x00\xB3\x91\x06\x00\x82\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x80\x91\x00\x00\x1C\x98\x00\x80\xB6\xB9\x06\x00\xB5\xB1\x06\x00\x20\x98"
-     "\x00\x80"
-     "\xBB\xF5\x06\x00\xBA\xE9\x06\x00\x24\x98\x00\x80\x28\x98\x00\x80\xBF\x39"
-     "\x06\x00"
-     "\xBE\xD9\x06\x00\xBD\xD1\x06\x00\xBC\xE5\x06\x00\x2C\x98\x00\x80\xA3\xD5"
-     "\x06\x00"
-     "\x86\xE8\x00\x00\x87\x84\x00\x00\xA6\xFD\x06\x00\x30\x98\x00\x80\x34\x98"
-     "\x00\x80"
-     "\xA5\xF5\x06\x00\xAA\xAD\x06\x00\xAB\xB1\x06\x00\x38\x98\x00\x80\x3C\x98"
-     "\x00\x80"
-     "\xAE\x9D\x06\x00\xAF\x7D\x06\x00\xAC\xA1\x06\x00\xAD\x95\x06\x00\x40\x98"
-     "\x00\x80"
-     "\x44\x98\x00\x80\xB5\x29\x69\x00\xB4\x01\x68\x00\xB3\x51\x6C\x00\xB2\x11"
-     "\x6D\x00"
-     "\xB1\x19\x6D\x00\xB0\x71\x70\x00\x48\x98\x00\x80\x4C\x98\x00\x80\x50\x98"
-     "\x00\x80"
-     "\x54\x98\x00\x80\x58\x98\x00\x80\x5C\x98\x00\x80\x60\x98\x00\x80\x64\x98"
-     "\x00\x80"
-     "\xA7\x01\x78\x00\xA6\xE9\x7D\x00\xA5\x69\x7D\x00\xA4\x01\x7C\x00\xA3\xF9"
-     "\x01\x00"
-     "\xA2\x51\x01\x00\xA1\x59\x01\x00\x68\x98\x00\x80\xAF\x69\x71\x00\xAE\x61"
-     "\x71\x00"
-     "\xAD\x01\x70\x00\xAC\x49\x75\x00\xAB\x51\x75\x00\xAA\x01\x74\x00\xA9\xF5"
-     "\x79\x00"
-     "\xA8\x5D\x79\x00\xA8\xF9\x06\x00\xA9\x0D\x06\x00\xAA\x05\x06\x00\xAB\x1D"
-     "\x06\x00"
-     "\xAC\x05\x06\x00\xAD\x25\x06\x00\xAE\x2D\x06\x00\xAF\x25\x06\x00\x6C\x98"
-     "\x00\x80"
-     "\x70\x98\x00\x80\x74\x98\x00\x80\x78\x98\x00\x80\x7C\x98\x00\x80\x80\x98"
-     "\x00\x80"
-     "\x84\x98\x00\x80\x88\x98\x00\x80\xB8\x25\x06\x00\xB9\x2D\x06\x00\xBA\x25"
-     "\x06\x00"
-     "\xBB\xC5\x01\x00\xBC\xDD\x01\x00\xBD\xC5\x01\x00\xBE\xCD\x01\x00\xBF\xC5"
-     "\x01\x00"
-     "\xB0\x5D\x06\x00\xB1\x25\x06\x00\xB2\x2D\x06\x00\xB3\x3D\x06\x00\xB4\x25"
-     "\x06\x00"
-     "\xB5\x2D\x06\x00\xB6\x25\x06\x00\xB7\x1D\x06\x00\xB3\x8D\x05\x00\x8C\x98"
-     "\x00\x80"
-     "\x90\x98\x00\x80\xBE\x40\x03\x00\x98\x98\x00\x80\xB6\xA5\x05\x00\xB5\x8D"
-     "\x05\x00"
-     "\x9C\x98\x00\x80\xBB\xE1\x05\x00\xBA\x99\x05\x00\x84\xA4\x03\x00\xA0\x98"
-     "\x00\x80"
-     "\xBF\xE9\x05\x00\xBE\xE1\x05\x00\xBD\xE9\x05\x00\xBC\xF9\x05\x00\x82\x49"
-     "\x00\x00"
-     "\xA3\xC9\x05\x00\x80\xF5\x03\x00\x81\x49\x00\x00\xA6\xE1\x05\x00\xA4\x98"
-     "\x00\x80"
-     "\xA8\x98\x00\x80\xA5\xC9\x05\x00\xAA\xDD\x05\x00\xAB\xA5\x05\x00\x86\xC8"
-     "\x7C\x00"
-     "\x87\xFC\x03\x00\xAE\xA5\x05\x00\xAF\xAD\x05\x00\xAC\xBD\x05\x00\xAD\xAD"
-     "\x05\x00"
-     "\xA8\x2D\x02\x00\xA9\x45\x02\x00\xAA\x5D\x02\x00\xAB\x55\x02\x00\xAC\x79"
-     "\x02\x00"
-     "\xAD\x79\x02\x00\xAE\x99\x03\x00\xAF\x99\x03\x00\xAC\x98\x00\x80\xB0\x98"
-     "\x00\x80"
-     "\xB4\x98\x00\x80\xB8\x98\x00\x80\xBC\x98\x00\x80\xC0\x98\x00\x80\xC4\x98"
-     "\x00\x80"
-     "\xC8\x98\x00\x80\xB8\x9D\x03\x00\xB9\x65\x03\x00\xBA\x6D\x03\x00\xBB\x65"
-     "\x03\x00"
-     "\xBC\x7D\x03\x00\xBD\x65\x03\x00\xBE\x6D\x03\x00\xBF\x65\x03\x00\xB0\xC1"
-     "\x03\x00"
-     "\xB1\xC1\x03\x00\xB2\xC1\x03\x00\xB3\xC1\x03\x00\xB4\xC1\x03\x00\xB5\xC1"
-     "\x03\x00"
-     "\xB6\xC1\x03\x00\xB7\xC1\x03\x00\xCC\x98\x00\x80\xD0\x98\x00\x80\xD4\x98"
-     "\x00\x80"
-     "\xD8\x98\x00\x80\xDC\x98\x00\x80\xEF\xE0\x03\x00\xE0\x98\x00\x80\xE4\x98"
-     "\x00\x80"
-     "\xBE\x68\x7D\x00\xE3\x1C\x03\x00\xE8\x98\x00\x80\xE1\xFC\x01\x00\xEC\x98"
-     "\x00\x80"
-     "\xF0\x98\x00\x80\xF4\x98\x00\x80\xF8\x98\x00\x80\xB3\x95\x03\x00\xFC\x98"
-     "\x00\x80"
-     "\x00\x99\x00\x80\x08\x99\x00\x80\x84\x1C\x7D\x00\xB6\xB9\x03\x00\xB5\xB1"
-     "\x03\x00"
-     "\x0C\x99\x00\x80\xBB\x45\x02\x00\xBA\x45\x02\x00\x86\xA8\x7C\x00\x87\xA4"
-     "\x7D\x00"
-     "\xBF\x45\x02\x00\xBE\x45\x02\x00\xBD\x55\x02\x00\xBC\x55\x02\x00\xE1\x1C"
-     "\x06\x00"
-     "\xE1\x94\x07\x00\xE3\x50\x06\x00\xE3\xF4\x06\x00\x80\xAD\x00\x00\x81\xD1"
-     "\x00\x00"
-     "\x82\xD1\x00\x00\x10\x99\x00\x80\x14\x99\x00\x80\x18\x99\x00\x80\x1C\x99"
-     "\x00\x80"
-     "\x20\x99\x00\x80\x24\x99\x00\x80\x28\x99\x00\x80\xEF\x8C\x06\x00\xEF\x24"
-     "\x06\x00"
-     "\xA3\x15\x03\x00\x2C\x99\x00\x80\x30\x99\x00\x80\x34\x99\x00\x80\x38\x99"
-     "\x00\x80"
-     "\xA6\x39\x03\x00\xA5\x31\x03\x00\x84\x90\x7C\x00\xAB\xC5\x02\x00\xAA\xC5"
-     "\x02\x00"
-     "\x3C\x99\x00\x80\x40\x99\x00\x80\xAF\xC5\x02\x00\xAE\xC5\x02\x00\xAD\xD5"
-     "\x02\x00"
-     "\xAC\xD5\x02\x00\xA8\x11\x7E\x00\xA9\x19\x7E\x00\xAA\x25\x7E\x00\xAB\x59"
-     "\x7E\x00"
-     "\xAC\x49\x7E\x00\xAD\xB5\x7E\x00\xAE\xBD\x7E\x00\xAF\xB5\x7E\x00\x04\x99"
-     "\x00\x80"
-     "\x44\x99\x00\x80\x48\x99\x00\x80\x4C\x99\x00\x80\x50\x99\x00\x80\x54\x99"
-     "\x00\x80"
-     "\x58\x99\x00\x80\x5C\x99\x00\x80\xB8\x4D\x7F\x00\xB9\x55\x7F\x00\xBA\x51"
-     "\x7F\x00"
-     "\xBB\x51\x7F\x00\xBC\x79\x7F\x00\xBD\x65\x7F\x00\xBE\x6D\x7F\x00\xBF\x65"
-     "\x7F\x00"
-     "\xB0\xCD\x7E\x00\xB1\xD5\x7E\x00\xB2\xDD\x7E\x00\xB3\xD5\x7E\x00\xB4\xCD"
-     "\x7E\x00"
-     "\xB5\x75\x7F\x00\xB6\x7D\x7F\x00\xB7\x75\x7F\x00\xB3\xAD\x7E\x00\x60\x99"
-     "\x00\x80"
-     "\x64\x99\x00\x80\x68\x99\x00\x80\x6C\x99\x00\x80\xB6\x35\x7E\x00\xB5\xC1"
-     "\x7E\x00"
-     "\x70\x99\x00\x80\xBB\x11\x7E\x00\xBA\x09\x7E\x00\xBE\xE0\x03\x00\x74\x99"
-     "\x00\x80"
-     "\xBF\xB1\x7F\x00\xBE\x09\x7E\x00\xBD\x01\x7E\x00\xBC\x09\x7E\x00\x82\x5D"
-     "\x00\x00"
-     "\xA3\xE9\x7E\x00\x80\x45\x00\x00\x81\x5D\x00\x00\xA6\x71\x7E\x00\x78\x99"
-     "\x00\x80"
-     "\x7C\x99\x00\x80\xA5\x85\x7E\x00\xAA\x4D\x7E\x00\xAB\x55\x7E\x00\x86\x48"
-     "\x00\x00"
-     "\x87\x6C\x00\x00\xAE\x4D\x7E\x00\xAF\xF5\x7F\x00\xAC\x4D\x7E\x00\xAD\x45"
-     "\x7E\x00"
-     "\xB0\xB5\x7E\x00\xB1\xC1\x7E\x00\xB2\xC1\x7E\x00\xB3\xC1\x7E\x00\xB4\xC1"
-     "\x7E\x00"
-     "\xB5\xFD\x7E\x00\xB6\xF5\x7E\x00\xB7\x39\x01\x00\xB8\x09\x01\x00\xB9\x09"
-     "\x01\x00"
-     "\xBA\x19\x01\x00\xBB\x19\x01\x00\xBC\x09\x01\x00\xBD\x09\x01\x00\xBE\xBD"
-     "\x01\x00"
-     "\xBF\xB5\x01\x00\x80\x99\x00\x80\x84\x99\x00\x80\x88\x99\x00\x80\x94\x98"
-     "\x00\x80"
-     "\x8C\x99\x00\x80\x90\x99\x00\x80\x94\x99\x00\x80\x98\x99\x00\x80\xA8\x2D"
-     "\x7E\x00"
-     "\xA9\x39\x7E\x00\xAA\xF5\x7E\x00\xAB\xCD\x7E\x00\xAC\xD5\x7E\x00\xAD\xDD"
-     "\x7E\x00"
-     "\xAE\xD5\x7E\x00\xAF\xCD\x7E\x00\xB3\x1D\x7E\x00\x9C\x99\x00\x80\xA0\x99"
-     "\x00\x80"
-     "\xA4\x99\x00\x80\xA8\x99\x00\x80\xB6\x3D\x7E\x00\xB5\x3D\x7E\x00\xAC\x99"
-     "\x00\x80"
-     "\xBB\x65\x7E\x00\xBA\x11\x7E\x00\xB0\x99\x00\x80\xB4\x99\x00\x80\xBF\x59"
-     "\x7E\x00"
-     "\xBE\x65\x7E\x00\xBD\x75\x7E\x00\xBC\x75\x7E\x00\xB8\x99\x00\x80\xA3\x59"
-     "\x7E\x00"
-     "\xBC\x99\x00\x80\xC0\x99\x00\x80\xA6\x79\x7E\x00\xC4\x99\x00\x80\xC8\x99"
-     "\x00\x80"
-     "\xA5\x79\x7E\x00\xAA\x55\x7E\x00\xAB\x21\x7E\x00\xCC\x99\x00\x80\xD0\x99"
-     "\x00\x80"
-     "\xAE\x21\x7E\x00\xAF\x1D\x7E\x00\xAC\x31\x7E\x00\xAD\x31\x7E\x00\xA8\xD5"
-     "\x02\x00"
-     "\xA9\xDD\x02\x00\xAA\x49\x03\x00\xAB\x49\x03\x00\xAC\x59\x03\x00\xAD\x45"
-     "\x03\x00"
-     "\xAE\x4D\x03\x00\xAF\x45\x03\x00\xD4\x99\x00\x80\xD8\x99\x00\x80\xDC\x99"
-     "\x00\x80"
-     "\xE0\x99\x00\x80\xBE\xF4\x03\x00\xE8\x99\x00\x80\xEC\x99\x00\x80\xF0\x99"
-     "\x00\x80"
-     "\xB8\x81\x03\x00\xB9\x8D\x03\x00\xBA\x85\x03\x00\xBB\x99\x03\x00\xBC\x89"
-     "\x03\x00"
-     "\xBD\xBD\x03\x00\xBE\xB5\x03\x00\xBF\x7D\x00\x00\xB0\x3D\x03\x00\xB1\xE1"
-     "\x03\x00"
-     "\xB2\xE1\x03\x00\xB3\xE1\x03\x00\xB4\xE1\x03\x00\xB5\xE9\x03\x00\xB6\xD1"
-     "\x03\x00"
-     "\xB7\xD1\x03\x00\x80\xA5\x00\x00\x81\xAD\x00\x00\x82\xA5\x00\x00\xB3\xCD"
-     "\x03\x00"
-     "\xF4\x99\x00\x80\xB5\xDD\x03\x00\xB6\xC9\x03\x00\xF8\x99\x00\x80\x84\xE0"
-     "\x02\x00"
-     "\xFC\x99\x00\x80\xBA\xF5\x03\x00\xBB\x29\x03\x00\xBC\x39\x03\x00\xBD\x39"
-     "\x03\x00"
-     "\xBE\x29\x03\x00\xBF\x21\x03\x00\xE3\x80\x7F\x00\x00\x9A\x00\x80\xE1\xB8"
-     "\x7E\x00"
-     "\x04\x9A\x00\x80\x87\xE0\x03\x00\x86\x74\x04\x00\x08\x9A\x00\x80\x0C\x9A"
-     "\x00\x80"
-     "\x10\x9A\x00\x80\x14\x9A\x00\x80\x18\x9A\x00\x80\x1C\x9A\x00\x80\xEF\x28"
-     "\x7F\x00"
-     "\x20\x9A\x00\x80\x24\x9A\x00\x80\x28\x9A\x00\x80\x2C\x9A\x00\x80\x30\x9A"
-     "\x00\x80"
-     "\x34\x9A\x00\x80\x38\x9A\x00\x80\xEF\xD8\x03\x00\x3C\x9A\x00\x80\x40\x9A"
-     "\x00\x80"
-     "\x44\x9A\x00\x80\xE3\x90\x03\x00\x48\x9A\x00\x80\xE1\xD8\x01\x00\x4C\x9A"
-     "\x00\x80"
-     "\x80\x15\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x50\x9A\x00\x80\xBE\x4C"
-     "\x05\x00"
-     "\x58\x9A\x00\x80\xA3\x09\x02\x00\xBE\x40\x04\x00\xA5\x19\x02\x00\x5C\x9A"
-     "\x00\x80"
-     "\x60\x9A\x00\x80\xA6\x0D\x02\x00\x84\x6C\x02\x00\x64\x9A\x00\x80\xAB\xED"
-     "\x02\x00"
-     "\xAA\x31\x02\x00\xAD\xFD\x02\x00\xAC\xFD\x02\x00\xAF\xE5\x02\x00\xAE\xED"
-     "\x02\x00"
-     "\x86\x40\x04\x00\x87\xB4\x05\x00\x68\x9A\x00\x80\x6C\x9A\x00\x80\x70\x9A"
-     "\x00\x80"
-     "\x74\x9A\x00\x80\x78\x9A\x00\x80\xEF\xCC\x01\x00\x7C\x9A\x00\x80\xE1\xD4"
-     "\x7E\x00"
-     "\x80\x9A\x00\x80\xE3\x60\x7E\x00\x84\x9A\x00\x80\x88\x9A\x00\x80\x8C\x9A"
-     "\x00\x80"
-     "\x90\x9A\x00\x80\xB3\xFD\x01\x00\x94\x9A\x00\x80\x98\x9A\x00\x80\x9C\x9A"
-     "\x00\x80"
-     "\xA0\x9A\x00\x80\xB6\x71\x01\x00\xB5\x45\x01\x00\xA4\x9A\x00\x80\xBB\x3D"
-     "\x01\x00"
-     "\xBA\x3D\x01\x00\xA8\x9A\x00\x80\xAC\x9A\x00\x80\xBF\x1D\x01\x00\xBE\x1D"
-     "\x01\x00"
-     "\xBD\x1D\x01\x00\xBC\x1D\x01\x00\xA8\x35\x06\x00\xA9\x3D\x06\x00\xAA\x35"
-     "\x06\x00"
-     "\xAB\x9D\x06\x00\xAC\x8D\x06\x00\xAD\xB1\x06\x00\xAE\xDD\x06\x00\xAF\xD1"
-     "\x06\x00"
-     "\x54\x9A\x00\x80\xB0\x9A\x00\x80\xB4\x9A\x00\x80\xB8\x9A\x00\x80\x80\x1D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\xBD\x00\x00\xBC\x9A\x00\x80\xB8\xBD\x06\x00\xB9\x45"
-     "\x07\x00"
-     "\xBA\x4D\x07\x00\xBB\x45\x07\x00\xBC\x5D\x07\x00\xBD\x41\x07\x00\xBE\x41"
-     "\x07\x00"
-     "\xBF\x41\x07\x00\xB0\xB9\x06\x00\xB1\xB9\x06\x00\xB2\x89\x06\x00\xB3\x89"
-     "\x06\x00"
-     "\xB4\x99\x06\x00\xB5\x99\x06\x00\xB6\x8D\x06\x00\xB7\x85\x06\x00\xA3\xBD"
-     "\x06\x00"
-     "\xC0\x9A\x00\x80\x84\x80\x02\x00\xC4\x9A\x00\x80\xC8\x9A\x00\x80\xA6\x31"
-     "\x06\x00"
-     "\xA5\x05\x06\x00\xCC\x9A\x00\x80\xAB\x7D\x06\x00\xAA\x7D\x06\x00\x86\x08"
-     "\x00\x00"
-     "\x87\x60\x03\x00\xAF\x5D\x06\x00\xAE\x5D\x06\x00\xAD\x5D\x06\x00\xAC\x5D"
-     "\x06\x00"
-     "\xD0\x9A\x00\x80\xB3\x9D\x07\x00\xD4\x9A\x00\x80\xD8\x9A\x00\x80\xB6\x5D"
-     "\x07\x00"
-     "\xDC\x9A\x00\x80\xE0\x9A\x00\x80\xB5\x55\x07\x00\xBA\x79\x07\x00\xBB\x45"
-     "\x07\x00"
-     "\xE4\x9A\x00\x80\xE8\x9A\x00\x80\xBE\x45\x07\x00\xBF\x4D\x07\x00\xBC\x51"
-     "\x07\x00"
-     "\xBD\x4D\x07\x00\xB0\x99\x06\x00\xB1\xA5\x06\x00\xB2\xA1\x06\x00\xB3\xBD"
-     "\x06\x00"
-     "\xB4\xA5\x06\x00\xB5\xAD\x06\x00\xB6\xA5\x06\x00\xB7\xD9\x06\x00\xB8\xE9"
-     "\x06\x00"
-     "\xB9\xE9\x06\x00\xBA\xF9\x06\x00\xBB\xF1\x06\x00\xBC\xD9\x06\x00\xBD\xD9"
-     "\x06\x00"
-     "\xBE\x1D\x01\x00\xBF\x15\x01\x00\xE4\x99\x00\x80\xEC\x9A\x00\x80\xF0\x9A"
-     "\x00\x80"
-     "\xF4\x9A\x00\x80\xF8\x9A\x00\x80\xFC\x9A\x00\x80\x00\x9B\x00\x80\x04\x9B"
-     "\x00\x80"
-     "\xA8\x25\x06\x00\xA9\x31\x06\x00\xAA\x31\x06\x00\xAB\x09\x06\x00\xAC\xF9"
-     "\x06\x00"
-     "\xAD\xF9\x06\x00\xAE\xE9\x06\x00\xAF\xE9\x06\x00\xA3\xD9\x06\x00\x08\x9B"
-     "\x00\x80"
-     "\x0C\x9B\x00\x80\x10\x9B\x00\x80\x14\x9B\x00\x80\xA6\x19\x06\x00\xA5\x11"
-     "\x06\x00"
-     "\x18\x9B\x00\x80\xAB\x01\x06\x00\xAA\x3D\x06\x00\x1C\x9B\x00\x80\x20\x9B"
-     "\x00\x80"
-     "\xAF\x09\x06\x00\xAE\x01\x06\x00\xAD\x09\x06\x00\xAC\x15\x06\x00\x24\x9B"
-     "\x00\x80"
-     "\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\x28\x9B\x00\x80\x2C\x9B"
-     "\x00\x80"
-     "\x30\x9B\x00\x80\xBE\x10\x00\x00\x84\xAC\x01\x00\x38\x9B\x00\x80\x87\x08"
-     "\x01\x00"
-     "\x86\x4C\x0F\x00\x3C\x9B\x00\x80\x40\x9B\x00\x80\x44\x9B\x00\x80\x48\x9B"
-     "\x00\x80"
-     "\xA8\x91\x02\x00\xA9\x99\x02\x00\xAA\xC9\x02\x00\xAB\xC5\x02\x00\xAC\xDD"
-     "\x02\x00"
-     "\xAD\xC1\x02\x00\xAE\xC1\x02\x00\xAF\xF5\x02\x00\x4C\x9B\x00\x80\x50\x9B"
-     "\x00\x80"
-     "\x54\x9B\x00\x80\x58\x9B\x00\x80\x8D\x00\x00\x00\x5C\x9B\x00\x80\x60\x9B"
-     "\x00\x80"
-     "\x64\x9B\x00\x80\xB8\xC5\x03\x00\xB9\xCD\x03\x00\xBA\xC5\x03\x00\xBB\xDD"
-     "\x03\x00"
-     "\xBC\xCD\x03\x00\xBD\xFD\x03\x00\xBE\xF5\x03\x00\xBF\x9D\x03\x00\xB0\x8D"
-     "\x02\x00"
-     "\xB1\x21\x03\x00\xB2\x21\x03\x00\xB3\x21\x03\x00\xB4\x21\x03\x00\xB5\x21"
-     "\x03\x00"
-     "\xB6\x21\x03\x00\xB7\x21\x03\x00\xB3\x41\x02\x00\x68\x9B\x00\x80\x6C\x9B"
-     "\x00\x80"
-     "\x70\x9B\x00\x80\x74\x9B\x00\x80\xB6\x41\x02\x00\xB5\x55\x02\x00\x78\x9B"
-     "\x00\x80"
-     "\xBB\xE5\x03\x00\xBA\x49\x02\x00\x7C\x9B\x00\x80\x80\x9B\x00\x80\xBF\xE5"
-     "\x03\x00"
-     "\xBE\xE5\x03\x00\xBD\xF5\x03\x00\xBC\xF5\x03\x00\x84\x9B\x00\x80\xE1\x08"
-     "\x06\x00"
-     "\x88\x9B\x00\x80\xE3\xBC\x07\x00\x8C\x9B\x00\x80\x90\x9B\x00\x80\x94\x9B"
-     "\x00\x80"
-     "\x98\x9B\x00\x80\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x9C\x9B"
-     "\x00\x80"
-     "\xA0\x9B\x00\x80\x84\xF8\x03\x00\xBE\xF4\x0C\x00\xEF\x44\x07\x00\xE3\xF0"
-     "\x00\x00"
-     "\xA8\x9B\x00\x80\xE1\xB8\x01\x00\x84\x04\x0D\x00\x86\xF0\x0C\x00\x87\xA4"
-     "\x02\x00"
-     "\xAC\x9B\x00\x80\xB0\x9B\x00\x80\xB4\x9B\x00\x80\xBE\x28\x0C\x00\xB8\x9B"
-     "\x00\x80"
-     "\xBC\x9B\x00\x80\xEF\x14\x00\x00\xC0\x9B\x00\x80\xC4\x9B\x00\x80\xC8\x9B"
-     "\x00\x80"
-     "\xA3\x81\x03\x00\xCC\x9B\x00\x80\xD0\x9B\x00\x80\xD4\x9B\x00\x80\xD8\x9B"
-     "\x00\x80"
-     "\xA6\x81\x03\x00\xA5\x95\x03\x00\xDC\x9B\x00\x80\xAB\x25\x02\x00\xAA\x89"
-     "\x03\x00"
-     "\xE0\x9B\x00\x80\xE4\x9B\x00\x80\xAF\x25\x02\x00\xAE\x25\x02\x00\xAD\x35"
-     "\x02\x00"
-     "\xAC\x35\x02\x00\xE8\x9B\x00\x80\xEC\x9B\x00\x80\xF0\x9B\x00\x80\xF4\x9B"
-     "\x00\x80"
-     "\xF8\x9B\x00\x80\xFC\x9B\x00\x80\x00\x9C\x00\x80\xEF\x28\x06\x00\x04\x9C"
-     "\x00\x80"
-     "\xE1\x3C\x06\x00\x08\x9C\x00\x80\xE3\x34\x01\x00\x80\x19\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x7D\x00\x00\x0C\x9C\x00\x80\xB3\x35\x02\x00\x10\x9C\x00\x80\x18\x9C"
-     "\x00\x80"
-     "\xBE\xC0\x0F\x00\x1C\x9C\x00\x80\xB6\xF5\x01\x00\xB5\x25\x02\x00\x20\x9C"
-     "\x00\x80"
-     "\xBB\xD5\x01\x00\xBA\xC9\x01\x00\x24\x9C\x00\x80\x28\x9C\x00\x80\xBF\x79"
-     "\x01\x00"
-     "\xBE\x71\x01\x00\xBD\xC5\x01\x00\xBC\xC5\x01\x00\xA8\xE5\x0D\x00\xA9\xF5"
-     "\x0D\x00"
-     "\xAA\xFD\x0D\x00\xAB\xF5\x0D\x00\xAC\xED\x0D\x00\xAD\x35\x0E\x00\xAE\x3D"
-     "\x0E\x00"
-     "\xAF\x35\x0E\x00\xA4\x9B\x00\x80\x14\x9C\x00\x80\x86\x88\x00\x00\x87\x40"
-     "\x03\x00"
-     "\x2C\x9C\x00\x80\x30\x9C\x00\x80\x34\x9C\x00\x80\x38\x9C\x00\x80\xB8\xDD"
-     "\x0E\x00"
-     "\xB9\xED\x0E\x00\xBA\xE5\x0E\x00\xBB\x8D\x0E\x00\xBC\x99\x0E\x00\xBD\x85"
-     "\x0E\x00"
-     "\xBE\x8D\x0E\x00\xBF\x85\x0E\x00\xB0\x4D\x0E\x00\xB1\x51\x0E\x00\xB2\x51"
-     "\x0E\x00"
-     "\xB3\x51\x0E\x00\xB4\xF5\x0E\x00\xB5\xFD\x0E\x00\xB6\xF5\x0E\x00\xB7\xED"
-     "\x0E\x00"
-     "\xA3\x75\x0D\x00\x3C\x9C\x00\x80\x40\x9C\x00\x80\x44\x9C\x00\x80\x48\x9C"
-     "\x00\x80"
-     "\xA6\xB5\x0E\x00\xA5\x65\x0D\x00\x4C\x9C\x00\x80\xAB\x95\x0E\x00\xAA\x89"
-     "\x0E\x00"
-     "\x50\x9C\x00\x80\x54\x9C\x00\x80\xAF\x39\x0E\x00\xAE\x31\x0E\x00\xAD\x85"
-     "\x0E\x00"
-     "\xAC\x85\x0E\x00\x58\x9C\x00\x80\xB3\xD9\x0F\x00\x5C\x9C\x00\x80\x60\x9C"
-     "\x00\x80"
-     "\xB6\xFD\x0F\x00\x64\x9C\x00\x80\x68\x9C\x00\x80\xB5\xF5\x0F\x00\xBA\xDD"
-     "\x0F\x00"
-     "\xBB\xA1\x0F\x00\x6C\x9C\x00\x80\x70\x9C\x00\x80\xBE\xA1\x0F\x00\xBF\xA9"
-     "\x0F\x00"
-     "\xBC\xB1\x0F\x00\xBD\xB1\x0F\x00\xB8\x81\x0E\x00\xB9\x81\x0E\x00\xBA\x81"
-     "\x0E\x00"
-     "\xBB\x81\x0E\x00\xBC\x81\x0E\x00\xBD\x81\x0E\x00\xBE\x81\x0E\x00\xBF\xBD"
-     "\x0E\x00"
-     "\xB0\x1D\x0E\x00\xB1\x61\x0E\x00\xB2\x61\x0E\x00\xB3\x61\x0E\x00\xB4\x61"
-     "\x0E\x00"
-     "\xB5\x69\x0E\x00\xB6\x51\x0E\x00\xB7\x51\x0E\x00\xA8\x45\x0E\x00\xA9\x5D"
-     "\x0E\x00"
-     "\xAA\x55\x0E\x00\xAB\x6D\x0E\x00\xAC\x79\x0E\x00\xAD\x79\x0E\x00\xAE\x6D"
-     "\x0E\x00"
-     "\xAF\x65\x0E\x00\x74\x9C\x00\x80\x82\x19\x00\x00\x81\x11\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\x78\x9C\x00\x80\x34\x9B\x00\x80\x7C\x9C\x00\x80\x80\x9C\x00\x80\xA3\x9D"
-     "\x0E\x00"
-     "\x84\x8C\x02\x00\x86\x28\x00\x00\x87\x48\x01\x00\x88\x9C\x00\x80\xA6\xB9"
-     "\x0E\x00"
-     "\xA5\xB1\x0E\x00\x8C\x9C\x00\x80\xAB\xE5\x0E\x00\xAA\x99\x0E\x00\x90\x9C"
-     "\x00\x80"
-     "\x94\x9C\x00\x80\xAF\xED\x0E\x00\xAE\xE5\x0E\x00\xAD\xF5\x0E\x00\xAC\xF5"
-     "\x0E\x00"
-     "\x98\x9C\x00\x80\xB3\x19\x0E\x00\x9C\x9C\x00\x80\xA0\x9C\x00\x80\xB6\xE1"
-     "\x01\x00"
-     "\xA4\x9C\x00\x80\xA8\x9C\x00\x80\xB5\x31\x0E\x00\xBA\xDD\x01\x00\xBB\x61"
-     "\x01\x00"
-     "\xAC\x9C\x00\x80\xB0\x9C\x00\x80\xBE\x61\x01\x00\xBF\x69\x01\x00\xBC\x71"
-     "\x01\x00"
-     "\xBD\x71\x01\x00\xA8\xC9\x0D\x00\xA9\xD5\x0D\x00\xAA\xD5\x0D\x00\xAB\xE5"
-     "\x0D\x00"
-     "\xAC\xFD\x0D\x00\xAD\x25\x02\x00\xAE\x2D\x02\x00\xAF\x25\x02\x00\xB4\x9C"
-     "\x00\x80"
-     "\xB8\x9C\x00\x80\xBC\x9C\x00\x80\xC0\x9C\x00\x80\x8C\x00\x00\x00\xC4\x9C"
-     "\x00\x80"
-     "\xC8\x9C\x00\x80\xCC\x9C\x00\x80\xB8\x81\x02\x00\xB9\x8D\x02\x00\xBA\x85"
-     "\x02\x00"
-     "\xBB\x99\x02\x00\xBC\x89\x02\x00\xBD\xBD\x02\x00\xBE\xB5\x02\x00\xBF\x3D"
-     "\x03\x00"
-     "\xB0\x5D\x02\x00\xB1\xE1\x02\x00\xB2\xE1\x02\x00\xB3\xE1\x02\x00\xB4\xE1"
-     "\x02\x00"
-     "\xB5\xE9\x02\x00\xB6\xD1\x02\x00\xB7\xD1\x02\x00\xA3\x5D\x0D\x00\xD0\x9C"
-     "\x00\x80"
-     "\xD4\x9C\x00\x80\xD8\x9C\x00\x80\xDC\x9C\x00\x80\xA6\xA5\x02\x00\xA5\x75"
-     "\x0D\x00"
-     "\xE0\x9C\x00\x80\xAB\x25\x02\x00\xAA\x99\x02\x00\xE4\x9C\x00\x80\xE8\x9C"
-     "\x00\x80"
-     "\xAF\x2D\x02\x00\xAE\x25\x02\x00\xAD\x35\x02\x00\xAC\x35\x02\x00\x81\x11"
-     "\x00\x00"
-     "\x80\x69\x00\x00\xEF\xC8\x0E\x00\x82\x19\x00\x00\xEC\x9C\x00\x80\xF0\x9C"
-     "\x00\x80"
-     "\xF8\x9C\x00\x80\xFC\x9C\x00\x80\x84\xEC\x03\x00\xBE\x68\x04\x00\x87\xA8"
-     "\x04\x00"
-     "\x86\xCC\x04\x00\xE1\x04\x0D\x00\x00\x9D\x00\x80\xE3\x1C\x0E\x00\x04\x9D"
-     "\x00\x80"
-     "\xA8\x91\x02\x00\xA9\x99\x02\x00\xAA\xA1\x02\x00\xAB\xA1\x02\x00\xAC\xD1"
-     "\x02\x00"
-     "\xAD\xDD\x02\x00\xAE\xD5\x02\x00\xAF\xC9\x02\x00\x08\x9D\x00\x80\x0C\x9D"
-     "\x00\x80"
-     "\x10\x9D\x00\x80\x14\x9D\x00\x80\x18\x9D\x00\x80\x1C\x9D\x00\x80\x20\x9D"
-     "\x00\x80"
-     "\x24\x9D\x00\x80\xB8\x79\x01\x00\xB9\x79\x01\x00\xBA\xCD\x01\x00\xBB\xC5"
-     "\x01\x00"
-     "\xBC\xDD\x01\x00\xBD\xC5\x01\x00\xBE\xC5\x01\x00\xBF\xF5\x01\x00\xB0\xB9"
-     "\x02\x00"
-     "\xB1\x8D\x02\x00\xB2\x85\x02\x00\xB3\x4D\x01\x00\xB4\x55\x01\x00\xB5\x5D"
-     "\x01\x00"
-     "\xB6\x55\x01\x00\xB7\x49\x01\x00\x28\x9D\x00\x80\x2C\x9D\x00\x80\x30\x9D"
-     "\x00\x80"
-     "\x34\x9D\x00\x80\x84\x10\x05\x00\xBE\x18\x04\x00\xEF\xAC\x0E\x00\xEF\x40"
-     "\x0E\x00"
-     "\xE1\x90\x01\x00\xE1\xAC\x0E\x00\xE3\x3C\x00\x00\xE3\xB0\x0E\x00\x38\x9D"
-     "\x00\x80"
-     "\x3C\x9D\x00\x80\x40\x9D\x00\x80\x44\x9D\x00\x80\xB3\x69\x02\x00\x48\x9D"
-     "\x00\x80"
-     "\x4C\x9D\x00\x80\x50\x9D\x00\x80\x54\x9D\x00\x80\xB6\x1D\x02\x00\xB5\x19"
-     "\x02\x00"
-     "\x58\x9D\x00\x80\xBB\x09\x02\x00\xBA\x3D\x02\x00\x5C\x9D\x00\x80\x60\x9D"
-     "\x00\x80"
-     "\xBF\xD9\x01\x00\xBE\xD1\x01\x00\xBD\xD9\x01\x00\xBC\x11\x02\x00\xA3\xA9"
-     "\x05\x00"
-     "\xF4\x9C\x00\x80\x64\x9D\x00\x80\x68\x9D\x00\x80\x6C\x9D\x00\x80\xA6\xDD"
-     "\x05\x00"
-     "\xA5\xD9\x05\x00\x70\x9D\x00\x80\xAB\xC9\x05\x00\xAA\xFD\x05\x00\x86\x08"
-     "\x03\x00"
-     "\x87\x7C\x03\x00\xAF\x19\x06\x00\xAE\x11\x06\x00\xAD\x19\x06\x00\xAC\xD1"
-     "\x05\x00"
-     "\x80\x9D\x07\x00\x81\xAD\x07\x00\x82\xA5\x07\x00\xB3\xB9\x06\x00\x74\x9D"
-     "\x00\x80"
-     "\xB5\x75\x07\x00\xB6\x7D\x07\x00\x78\x9D\x00\x80\x7C\x9D\x00\x80\x80\x9D"
-     "\x00\x80"
-     "\xBA\x59\x07\x00\xBB\x59\x07\x00\xBC\x49\x07\x00\xBD\x49\x07\x00\xBE\xF9"
-     "\x07\x00"
-     "\xBF\xF9\x07\x00\xA8\x59\x06\x00\xA9\x59\x06\x00\xAA\x7D\x06\x00\xAB\x71"
-     "\x06\x00"
-     "\xAC\xB1\x07\x00\xAD\xB1\x07\x00\xAE\xB1\x07\x00\xAF\xB1\x07\x00\x84\x9D"
-     "\x00\x80"
-     "\x88\x9D\x00\x80\x8C\x9D\x00\x80\x90\x9D\x00\x80\x94\x9D\x00\x80\x98\x9D"
-     "\x00\x80"
-     "\x9C\x9D\x00\x80\xA0\x9D\x00\x80\xB8\x71\x07\x00\xB9\x71\x07\x00\xBA\x71"
-     "\x07\x00"
-     "\xBB\x71\x07\x00\xBC\x15\x07\x00\xBD\x1D\x07\x00\xBE\x15\x07\x00\xBF\x0D"
-     "\x07\x00"
-     "\xB0\xD1\x07\x00\xB1\xD1\x07\x00\xB2\xD1\x07\x00\xB3\xD1\x07\x00\xB4\x71"
-     "\x07\x00"
-     "\xB5\x71\x07\x00\xB6\x71\x07\x00\xB7\x71\x07\x00\xA3\xF5\x07\x00\xA4\x9D"
-     "\x00\x80"
-     "\xA8\x9D\x00\x80\xAC\x9D\x00\x80\xB0\x9D\x00\x80\xA6\x31\x06\x00\xA5\x39"
-     "\x06\x00"
-     "\xB4\x9D\x00\x80\xAB\x15\x06\x00\xAA\x15\x06\x00\xB8\x9D\x00\x80\xBC\x9D"
-     "\x00\x80"
-     "\xAF\xB5\x06\x00\xAE\xB5\x06\x00\xAD\x05\x06\x00\xAC\x05\x06\x00\xC0\x9D"
-     "\x00\x80"
-     "\xC4\x9D\x00\x80\xC8\x9D\x00\x80\xCC\x9D\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x05\x00\x00\xD0\x9D\x00\x80\x84\xAC\x01\x00\xBE\xAC\x01\x00\x84\x9C"
-     "\x00\x80"
-     "\xD4\x9D\x00\x80\x86\xD0\x00\x00\x87\xE4\x01\x00\xD8\x9D\x00\x80\xDC\x9D"
-     "\x00\x80"
-     "\xA8\x8D\x01\x00\xA9\x95\x01\x00\xAA\x95\x01\x00\xAB\xDD\x01\x00\xAC\xCD"
-     "\x01\x00"
-     "\xAD\xF5\x01\x00\xAE\xFD\x01\x00\xAF\x3D\x01\x00\x84\xA0\x00\x00\xE0\x9D"
-     "\x00\x80"
-     "\xE4\x9D\x00\x80\xE8\x9D\x00\x80\xEC\x9D\x00\x80\xF0\x9D\x00\x80\xF4\x9D"
-     "\x00\x80"
-     "\xF8\x9D\x00\x80\xB8\xC9\x00\x00\xB9\xC9\x00\x00\xBA\xD9\x00\x00\xBB\xD1"
-     "\x00\x00"
-     "\xBC\xF9\x00\x00\xBD\xF9\x00\x00\xBE\x99\x00\x00\xBF\x99\x00\x00\xB0\x45"
-     "\x01\x00"
-     "\xB1\x4D\x01\x00\xB2\x45\x01\x00\xB3\x5D\x01\x00\xB4\x45\x01\x00\xB5\x4D"
-     "\x01\x00"
-     "\xB6\x45\x01\x00\xB7\xF9\x00\x00\xB3\x1D\x02\x00\xFC\x9D\x00\x80\x00\x9E"
-     "\x00\x80"
-     "\x04\x9E\x00\x80\x08\x9E\x00\x80\xB6\x75\x02\x00\xB5\x45\x02\x00\x0C\x9E"
-     "\x00\x80"
-     "\xBB\x89\x02\x00\xBA\x81\x02\x00\x10\x9E\x00\x80\x14\x9E\x00\x80\xBF\x89"
-     "\x02\x00"
-     "\xBE\x81\x02\x00\xBD\x89\x02\x00\xBC\x91\x02\x00\x85\x5C\x1A\x00\xA3\x59"
-     "\x02\x00"
-     "\x18\x9E\x00\x80\x1C\x9E\x00\x80\xA6\x31\x02\x00\x20\x9E\x00\x80\x24\x9E"
-     "\x00\x80"
-     "\xA5\x01\x02\x00\xAA\xC5\x02\x00\xAB\xCD\x02\x00\x28\x9E\x00\x80\x2C\x9E"
-     "\x00\x80"
-     "\xAE\xC5\x02\x00\xAF\xCD\x02\x00\xAC\xD5\x02\x00\xAD\xCD\x02\x00\xBE\x00"
-     "\x1C\x00"
-     "\x30\x9E\x00\x80\x34\x9E\x00\x80\x84\x00\x1C\x00\x80\x0D\x00\x00\x81\x35"
-     "\x00\x00"
-     "\x82\x3D\x00\x00\x38\x9E\x00\x80\x3C\x9E\x00\x80\x40\x9E\x00\x80\x48\x9E"
-     "\x00\x80"
-     "\x4C\x9E\x00\x80\x50\x9E\x00\x80\x54\x9E\x00\x80\x58\x9E\x00\x80\x5C\x9E"
-     "\x00\x80"
-     "\x86\xE0\x1C\x00\x87\x04\x03\x00\x60\x9E\x00\x80\x64\x9E\x00\x80\xEF\xA8"
-     "\x02\x00"
-     "\xEF\xC8\x06\x00\x68\x9E\x00\x80\x6C\x9E\x00\x80\xE3\x50\x01\x00\xE3\x14"
-     "\x06\x00"
-     "\xE1\x80\x01\x00\xE1\x7C\x05\x00\x70\x9E\x00\x80\x74\x9E\x00\x80\x78\x9E"
-     "\x00\x80"
-     "\x7C\x9E\x00\x80\xA8\x79\x02\x00\xA9\x79\x02\x00\xAA\xAD\x02\x00\xAB\xA5"
-     "\x02\x00"
-     "\xAC\xBD\x02\x00\xAD\xA5\x02\x00\xAE\xAD\x02\x00\xAF\xA5\x02\x00\xBE\xAC"
-     "\x1C\x00"
-     "\x80\x9E\x00\x80\x84\x9E\x00\x80\x88\x9E\x00\x80\x8C\x9E\x00\x80\x90\x9E"
-     "\x00\x80"
-     "\x94\x9E\x00\x80\x98\x9E\x00\x80\xB8\x09\x01\x00\xB9\x09\x01\x00\xBA\x1D"
-     "\x01\x00"
-     "\xBB\x15\x01\x00\xBC\x39\x01\x00\xBD\x39\x01\x00\xBE\xDD\x01\x00\xBF\xD5"
-     "\x01\x00"
-     "\xB0\xDD\x02\x00\xB1\x61\x01\x00\xB2\x61\x01\x00\xB3\x61\x01\x00\xB4\x61"
-     "\x01\x00"
-     "\xB5\x69\x01\x00\xB6\x51\x01\x00\xB7\x51\x01\x00\xE1\x74\x07\x00\xE3\xF4"
-     "\x00\x00"
-     "\xE3\x08\x06\x00\xE1\xBC\x01\x00\x9C\x9E\x00\x80\xA0\x9E\x00\x80\xA4\x9E"
-     "\x00\x80"
-     "\x84\x14\x1C\x00\xA8\x9E\x00\x80\xAC\x9E\x00\x80\xB0\x9E\x00\x80\xB8\x9E"
-     "\x00\x80"
-     "\xEC\x00\x00\x00\xEF\x28\x00\x00\xEF\xD4\x06\x00\xBC\x9E\x00\x80\x80\x61"
-     "\x00\x00"
-     "\x81\x61\x00\x00\x82\x61\x00\x00\xB3\x69\x02\x00\xC0\x9E\x00\x80\xB5\x79"
-     "\x02\x00"
-     "\xB6\x69\x02\x00\xC4\x9E\x00\x80\x86\xC0\x1C\x00\x87\xE8\x1C\x00\xBA\x55"
-     "\x02\x00"
-     "\xBB\xE9\x01\x00\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\xE9\x01\x00\xBF\xE1"
-     "\x01\x00"
-     "\x44\x9E\x00\x80\xB4\x9E\x00\x80\xC8\x9E\x00\x80\xA3\xA9\x1D\x00\xCC\x9E"
-     "\x00\x80"
-     "\xA5\xB9\x1D\x00\xA6\xA9\x1D\x00\xD0\x9E\x00\x80\xD4\x9E\x00\x80\xD8\x9E"
-     "\x00\x80"
-     "\xAA\x95\x1D\x00\xAB\x29\x1E\x00\xAC\x39\x1E\x00\xAD\x39\x1E\x00\xAE\x29"
-     "\x1E\x00"
-     "\xAF\x21\x1E\x00\xB3\xF1\x1E\x00\xDC\x9E\x00\x80\xE0\x9E\x00\x80\xE4\x9E"
-     "\x00\x80"
-     "\xE8\x9E\x00\x80\xB6\xE9\x1E\x00\xB5\xE1\x1E\x00\xEC\x9E\x00\x80\xBB\xBD"
-     "\x1E\x00"
-     "\xBA\xBD\x1E\x00\xF0\x9E\x00\x80\xF4\x9E\x00\x80\xBF\x8D\x1E\x00\xBE\x8D"
-     "\x1E\x00"
-     "\xBD\x8D\x1E\x00\xBC\x8D\x1E\x00\xA8\x29\x1E\x00\xA9\x29\x1E\x00\xAA\x39"
-     "\x1E\x00"
-     "\xAB\x39\x1E\x00\xAC\x29\x1E\x00\xAD\x29\x1E\x00\xAE\x51\x1E\x00\xAF\x51"
-     "\x1E\x00"
-     "\xF8\x9E\x00\x80\xFC\x9E\x00\x80\x00\x9F\x00\x80\x04\x9F\x00\x80\x08\x9F"
-     "\x00\x80"
-     "\x0C\x9F\x00\x80\x10\x9F\x00\x80\x14\x9F\x00\x80\xB8\xED\x1F\x00\xB9\xF5"
-     "\x1F\x00"
-     "\xBA\xFD\x1F\x00\xBB\xF5\x1F\x00\xBC\xED\x1F\x00\xBD\x45\x1F\x00\xBE\x4D"
-     "\x1F\x00"
-     "\xBF\x45\x1F\x00\xB0\x39\x1E\x00\xB1\x39\x1E\x00\xB2\x19\x1E\x00\xB3\x19"
-     "\x1E\x00"
-     "\xB4\x09\x1E\x00\xB5\x09\x1E\x00\xB6\xDD\x1F\x00\xB7\xD5\x1F\x00\xA3\xB1"
-     "\x1F\x00"
-     "\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D\x00\x00\x18\x9F\x00\x80\xA6\xA9"
-     "\x1F\x00"
-     "\xA5\xA1\x1F\x00\x1C\x9F\x00\x80\xAB\xFD\x1F\x00\xAA\xFD\x1F\x00\x20\x9F"
-     "\x00\x80"
-     "\xBE\xA0\x01\x00\xAF\xCD\x1F\x00\xAE\xCD\x1F\x00\xAD\xCD\x1F\x00\xAC\xCD"
-     "\x1F\x00"
-     "\x24\x9F\x00\x80\xB3\x95\x1E\x00\x86\xE8\x00\x00\x87\x2C\x01\x00\xB6\x41"
-     "\x01\x00"
-     "\x28\x9F\x00\x80\x2C\x9F\x00\x80\xB5\x55\x01\x00\xBA\x45\x01\x00\xBB\x21"
-     "\x01\x00"
-     "\x30\x9F\x00\x80\x34\x9F\x00\x80\xBE\x21\x01\x00\xBF\x29\x01\x00\xBC\x31"
-     "\x01\x00"
-     "\xBD\x31\x01\x00\xA8\x2D\x1E\x00\xA9\x39\x1E\x00\xAA\x59\x1E\x00\xAB\x51"
-     "\x1E\x00"
-     "\xAC\x99\x01\x00\xAD\x85\x01\x00\xAE\x85\x01\x00\xAF\xBD\x01\x00\x84\xE0"
-     "\x01\x00"
-     "\x38\x9F\x00\x80\x3C\x9F\x00\x80\x40\x9F\x00\x80\x44\x9F\x00\x80\x48\x9F"
-     "\x00\x80"
-     "\x4C\x9F\x00\x80\x50\x9F\x00\x80\xB8\x61\x01\x00\xB9\x61\x01\x00\xBA\x61"
-     "\x01\x00"
-     "\xBB\x61\x01\x00\xBC\x61\x01\x00\xBD\x61\x01\x00\xBE\x61\x01\x00\xBF\x61"
-     "\x01\x00"
-     "\xB0\xC5\x01\x00\xB1\xA1\x01\x00\xB2\xA1\x01\x00\xB3\xA1\x01\x00\xB4\xA1"
-     "\x01\x00"
-     "\xB5\xA9\x01\x00\xB6\x91\x01\x00\xB7\x91\x01\x00\xA3\xD1\x1D\x00\x54\x9F"
-     "\x00\x80"
-     "\x58\x9F\x00\x80\x5C\x9F\x00\x80\x60\x9F\x00\x80\xA6\x05\x02\x00\xA5\x11"
-     "\x02\x00"
-     "\x64\x9F\x00\x80\xAB\x65\x02\x00\xAA\x01\x02\x00\x68\x9F\x00\x80\x6C\x9F"
-     "\x00\x80"
-     "\xAF\x6D\x02\x00\xAE\x65\x02\x00\xAD\x75\x02\x00\xAC\x75\x02\x00\x70\x9F"
-     "\x00\x80"
-     "\x74\x9F\x00\x80\x78\x9F\x00\x80\x7C\x9F\x00\x80\x80\x9F\x00\x80\x84\x9F"
-     "\x00\x80"
-     "\x88\x9F\x00\x80\x8C\x9F\x00\x80\x80\x39\x00\x00\x81\x39\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x90\x9F\x00\x80\xBE\x3C\x03\x00\x98\x9F\x00\x80\x87\x44\x03\x00\x86\xDC"
-     "\x04\x00"
-     "\x84\xCC\x04\x00\x9C\x9F\x00\x80\xA0\x9F\x00\x80\xB3\xE5\x03\x00\xA4\x9F"
-     "\x00\x80"
-     "\xB5\x31\x03\x00\xB6\x31\x03\x00\xA8\x9F\x00\x80\xAC\x9F\x00\x80\xB0\x9F"
-     "\x00\x80"
-     "\xBA\x1D\x03\x00\xBB\x15\x03\x00\xBC\x0D\x03\x00\xBD\xB1\x00\x00\xBE\xB1"
-     "\x00\x00"
-     "\xBF\xB1\x00\x00\xB4\x9F\x00\x80\xE1\x94\x01\x00\xB8\x9F\x00\x80\xE3\x7C"
-     "\x01\x00"
-     "\xBC\x9F\x00\x80\xC0\x9F\x00\x80\xC4\x9F\x00\x80\xC8\x9F\x00\x80\xCC\x9F"
-     "\x00\x80"
-     "\xD0\x9F\x00\x80\xD4\x9F\x00\x80\xD8\x9F\x00\x80\xDC\x9F\x00\x80\xE0\x9F"
-     "\x00\x80"
-     "\xE4\x9F\x00\x80\xEF\x74\x02\x00\xA8\x55\x02\x00\xA9\x5D\x02\x00\xAA\x61"
-     "\x02\x00"
-     "\xAB\x61\x02\x00\xAC\xB5\x02\x00\xAD\xBD\x02\x00\xAE\xB5\x02\x00\xAF\xAD"
-     "\x02\x00"
-     "\xBE\x68\x05\x00\xE8\x9F\x00\x80\xEC\x9F\x00\x80\xF0\x9F\x00\x80\xF4\x9F"
-     "\x00\x80"
-     "\x82\xAD\x00\x00\x81\xAD\x00\x00\x80\xAD\x00\x00\xB8\x69\x01\x00\xB9\x69"
-     "\x01\x00"
-     "\xBA\x09\x01\x00\xBB\x09\x01\x00\xBC\x19\x01\x00\xBD\x19\x01\x00\xBE\x09"
-     "\x01\x00"
-     "\xBF\x09\x01\x00\xB0\xD5\x02\x00\xB1\xDD\x02\x00\xB2\xD5\x02\x00\xB3\x69"
-     "\x01\x00"
-     "\xB4\x79\x01\x00\xB5\x79\x01\x00\xB6\x69\x01\x00\xB7\x61\x01\x00\xE1\x78"
-     "\x1E\x00"
-     "\xE1\x5C\x1F\x00\xE3\x04\x1E\x00\xE3\x10\x1F\x00\xF8\x9F\x00\x80\xFC\x9F"
-     "\x00\x80"
-     "\x00\xA0\x00\x80\x08\xA0\x00\x80\x0C\xA0\x00\x80\x10\xA0\x00\x80\x14\xA0"
-     "\x00\x80"
-     "\x18\xA0\x00\x80\xBE\xF8\x04\x00\x1C\xA0\x00\x80\xEF\x6C\x00\x00\xEF\x64"
-     "\x1E\x00"
-     "\xA3\xA9\x02\x00\x20\xA0\x00\x80\x86\x48\x04\x00\x87\x04\x05\x00\x24\xA0"
-     "\x00\x80"
-     "\xA6\x7D\x02\x00\xA5\x7D\x02\x00\x28\xA0\x00\x80\xAB\x59\x02\x00\xAA\x51"
-     "\x02\x00"
-     "\x2C\xA0\x00\x80\x30\xA0\x00\x80\xAF\xFD\x01\x00\xAE\xFD\x01\x00\xAD\xFD"
-     "\x01\x00"
-     "\xAC\x41\x02\x00\xA8\x75\x06\x00\xA9\x81\x06\x00\xAA\x81\x06\x00\xAB\x9D"
-     "\x06\x00"
-     "\xAC\x8D\x06\x00\xAD\xB5\x06\x00\xAE\xBD\x06\x00\xAF\xDD\x06\x00\x04\xA0"
-     "\x00\x80"
-     "\x34\xA0\x00\x80\x38\xA0\x00\x80\x3C\xA0\x00\x80\x40\xA0\x00\x80\x44\xA0"
-     "\x00\x80"
-     "\x48\xA0\x00\x80\x4C\xA0\x00\x80\xB8\xAD\x06\x00\xB9\xB5\x06\x00\xBA\xBD"
-     "\x06\x00"
-     "\xBB\xB5\x06\x00\xBC\x75\x07\x00\xBD\x7D\x07\x00\xBE\x75\x07\x00\xBF\x6D"
-     "\x07\x00"
-     "\xB0\xA9\x06\x00\xB1\xB5\x06\x00\xB2\xBD\x06\x00\xB3\xB5\x06\x00\xB4\xAD"
-     "\x06\x00"
-     "\xB5\x95\x06\x00\xB6\x9D\x06\x00\xB7\x95\x06\x00\xB3\x09\x06\x00\x50\xA0"
-     "\x00\x80"
-     "\x54\xA0\x00\x80\x58\xA0\x00\x80\x5C\xA0\x00\x80\xB6\x35\x06\x00\xB5\x05"
-     "\x06\x00"
-     "\x60\xA0\x00\x80\xBB\x71\x06\x00\xBA\x6D\x06\x00\x64\xA0\x00\x80\x68\xA0"
-     "\x00\x80"
-     "\xBF\x51\x06\x00\xBE\x69\x06\x00\xBD\x61\x06\x00\xBC\x69\x06\x00\x82\x19"
-     "\x00\x00"
-     "\xA3\x4D\x06\x00\x80\x65\x00\x00\x81\x19\x00\x00\xA6\x71\x06\x00\x6C\xA0"
-     "\x00\x80"
-     "\x70\xA0\x00\x80\xA5\x41\x06\x00\xAA\x29\x06\x00\xAB\x35\x06\x00\x84\xA4"
-     "\x03\x00"
-     "\xBF\xE4\x01\x00\xAE\x2D\x06\x00\xAF\x15\x06\x00\xAC\x2D\x06\x00\xAD\x25"
-     "\x06\x00"
-     "\x85\x6C\x01\x00\x78\xA0\x00\x80\x84\x08\x00\x00\x94\x9F\x00\x80\x86\x00"
-     "\x0C\x00"
-     "\x87\x60\x00\x00\x7C\xA0\x00\x80\x80\xA0\x00\x80\xA8\xBD\x06\x00\xA9\x85"
-     "\x06\x00"
-     "\xAA\x8D\x06\x00\xAB\x85\x06\x00\xAC\x9D\x06\x00\xAD\x81\x06\x00\xAE\xFD"
-     "\x06\x00"
-     "\xAF\xED\x06\x00\xB0\x9D\x06\x00\xB1\xAD\x06\x00\xB2\xA5\x06\x00\xB3\x69"
-     "\x01\x00"
-     "\xB4\x79\x01\x00\xB5\x79\x01\x00\xB6\x69\x01\x00\xB7\x69\x01\x00\xB8\x59"
-     "\x01\x00"
-     "\xB9\x59\x01\x00\xBA\xED\x01\x00\xBB\xE5\x01\x00\xBC\xFD\x01\x00\xBD\xE5"
-     "\x01\x00"
-     "\xBE\xE5\x01\x00\xBF\xD5\x01\x00\xB3\x25\x06\x00\xBE\x08\x0F\x00\x74\xA0"
-     "\x00\x80"
-     "\x84\xA0\x00\x80\x88\xA0\x00\x80\xB6\x21\x06\x00\xB5\x35\x06\x00\x8C\xA0"
-     "\x00\x80"
-     "\xBB\x49\x06\x00\xBA\x79\x06\x00\x90\xA0\x00\x80\x94\xA0\x00\x80\xBF\x9D"
-     "\x01\x00"
-     "\xBE\x9D\x01\x00\xBD\x9D\x01\x00\xBC\x51\x06\x00\x98\xA0\x00\x80\xA3\x61"
-     "\x06\x00"
-     "\x9C\xA0\x00\x80\xA0\xA0\x00\x80\xA6\x65\x06\x00\xA4\xA0\x00\x80\xA8\xA0"
-     "\x00\x80"
-     "\xA5\x71\x06\x00\xAA\x3D\x06\x00\xAB\x0D\x06\x00\xAC\xA0\x00\x80\xB0\xA0"
-     "\x00\x80"
-     "\xAE\xD9\x01\x00\xAF\xD9\x01\x00\xAC\x15\x06\x00\xAD\xD9\x01\x00\xE1\x14"
-     "\x01\x00"
-     "\xE3\x1C\x01\x00\xE3\xD0\x36\x00\xE1\x04\x01\x00\x84\x90\x0C\x00\xB4\xA0"
-     "\x00\x80"
-     "\xB8\xA0\x00\x80\xBC\xA0\x00\x80\x8F\x21\x6A\x00\xC0\xA0\x00\x80\xC4\xA0"
-     "\x00\x80"
-     "\xC8\xA0\x00\x80\xCC\xA0\x00\x80\xEF\x48\x01\x00\xEF\xFC\x37\x00\xD0\xA0"
-     "\x00\x80"
-     "\x97\xCD\x12\x00\x96\x65\x12\x00\x95\x09\x13\x00\x94\xE1\x16\x00\x93\x71"
-     "\x16\x00"
-     "\x92\x19\x17\x00\x91\xF1\x6A\x00\x90\x59\x6A\x00\x9F\x61\x06\x00\x9E\x09"
-     "\x1B\x00"
-     "\x9D\x61\x1A\x00\x9C\x69\x1A\x00\x9B\x19\x1B\x00\x9A\x51\x1E\x00\x99\x59"
-     "\x1E\x00"
-     "\x98\x09\x1F\x00\xE3\x10\x06\x00\xD4\xA0\x00\x80\xE1\xEC\x06\x00\xD8\xA0"
-     "\x00\x80"
-     "\xA3\xB1\x02\x00\xA2\x99\x03\x00\xA1\xD9\x07\x00\xA0\x99\x06\x00\xB3\x9D"
-     "\x3A\x00"
-     "\xB2\x85\x3A\x00\xB1\x01\x38\x00\xB0\x01\x3E\x00\xEF\x68\x07\x00\xDC\xA0"
-     "\x00\x80"
-     "\xB5\xBD\x26\x00\xB4\x01\x24\x00\xAB\x01\x30\x00\xAA\x19\x36\x00\xA9\x45"
-     "\x36\x00"
-     "\xA8\x4D\x36\x00\xAF\xF1\x3E\x00\xAE\x01\x3C\x00\xAD\xC5\x33\x00\xAC\xFD"
-     "\x32\x00"
-     "\xA3\x65\x0E\x00\xA2\x65\x0E\x00\xA1\x61\x03\x00\xE0\xA0\x00\x80\xA7\xE1"
-     "\x0B\x00"
-     "\xA6\x71\x0A\x00\xA5\x79\x0A\x00\xA4\x89\x0F\x00\x82\x01\x6B\x00\x83\x01"
-     "\x6B\x00"
-     "\x86\x08\x0D\x00\x87\x4C\x0D\x00\x86\x31\x17\x00\x87\xB9\x17\x00\x84\xB1"
-     "\x6B\x00"
-     "\x85\x51\x16\x00\x8A\x61\x12\x00\x8B\x05\x13\x00\xE8\xA0\x00\x80\xEC\xA0"
-     "\x00\x80"
-     "\x8E\x11\x1F\x00\x8F\x11\x1F\x00\x8C\xB5\x13\x00\x8D\x49\x1E\x00\x92\x29"
-     "\x1A\x00"
-     "\x93\x41\x1B\x00\xBE\xC4\x0C\x00\x84\x80\x0F\x00\x96\x51\x07\x00\x97\x51"
-     "\x07\x00"
-     "\x94\x51\x1B\x00\x95\x31\x1A\x00\x9A\x21\x06\x00\x9B\x69\x02\x00\xF0\xA0"
-     "\x00\x80"
-     "\xF4\xA0\x00\x80\x81\xC5\x01\x00\x80\xF9\x01\x00\x9C\x59\x03\x00\x82\x91"
-     "\x01\x00"
-     "\xBE\x48\x0F\x00\xF8\xA0\x00\x80\xFC\xA0\x00\x80\x00\xA1\x00\x80\x04\xA1"
-     "\x00\x80"
-     "\x08\xA1\x00\x80\x0C\xA1\x00\x80\xEF\xD4\x06\x00\x10\xA1\x00\x80\xE1\x74"
-     "\x06\x00"
-     "\x14\xA1\x00\x80\xE3\x54\x01\x00\x18\xA1\x00\x80\x1C\xA1\x00\x80\x20\xA1"
-     "\x00\x80"
-     "\x24\xA1\x00\x80\x28\xA1\x00\x80\x2C\xA1\x00\x80\x30\xA1\x00\x80\xB3\xD1"
-     "\x02\x00"
-     "\x34\xA1\x00\x80\xB5\x45\x02\x00\xB6\x45\x02\x00\x38\xA1\x00\x80\x3C\xA1"
-     "\x00\x80"
-     "\x40\xA1\x00\x80\xBA\x29\x02\x00\xBB\x25\x02\x00\xBC\x3D\x02\x00\xBD\x21"
-     "\x02\x00"
-     "\xBE\x21\x02\x00\xBF\x15\x02\x00\xA3\x1D\x0D\x00\x44\xA1\x00\x80\x48\xA1"
-     "\x00\x80"
-     "\x4C\xA1\x00\x80\x50\xA1\x00\x80\xA6\x89\x0D\x00\xA5\x89\x0D\x00\x58\xA1"
-     "\x00\x80"
-     "\xAB\xE9\x0D\x00\xAA\xE5\x0D\x00\x86\x08\x03\x00\x87\x4C\x03\x00\xAF\xD9"
-     "\x0D\x00"
-     "\xAE\xED\x0D\x00\xAD\xED\x0D\x00\xAC\xF1\x0D\x00\x80\x49\x00\x00\x81\x49"
-     "\x00\x00"
-     "\x82\x59\x00\x00\xB3\x25\x0F\x00\x5C\xA1\x00\x80\xB5\xD1\x0F\x00\xB6\xD1"
-     "\x0F\x00"
-     "\x60\xA1\x00\x80\x64\xA1\x00\x80\x68\xA1\x00\x80\xBA\xFD\x0F\x00\xBB\xF5"
-     "\x0F\x00"
-     "\xBC\xD1\x0F\x00\xBD\xDD\x0F\x00\xBE\xC9\x0F\x00\xBF\xC9\x0F\x00\xE4\xA0"
-     "\x00\x80"
-     "\x54\xA1\x00\x80\x6C\xA1\x00\x80\x70\xA1\x00\x80\x74\xA1\x00\x80\x78\xA1"
-     "\x00\x80"
-     "\x7C\xA1\x00\x80\x80\xA1\x00\x80\xA8\x89\x0F\x00\xA9\x89\x0F\x00\xAA\xC1"
-     "\x0F\x00"
-     "\xAB\xC1\x0F\x00\xAC\x35\x0F\x00\xAD\x3D\x0F\x00\xAE\x35\x0F\x00\xAF\x2D"
-     "\x0F\x00"
-     "\xB0\x55\x0F\x00\xB1\x5D\x0F\x00\xB2\x69\x0F\x00\xB3\x65\x0F\x00\xB4\x61"
-     "\x0F\x00"
-     "\xB5\x61\x0F\x00\xB6\x61\x0F\x00\xB7\x61\x0F\x00\xB8\x21\x0F\x00\xB9\x21"
-     "\x0F\x00"
-     "\xBA\x21\x0F\x00\xBB\x3D\x0F\x00\xBC\x2D\x0F\x00\xBD\x15\x0F\x00\xBE\x1D"
-     "\x0F\x00"
-     "\xBF\xF9\x00\x00\xA3\x69\x0E\x00\x84\xA1\x00\x80\x88\xA1\x00\x80\x8C\xA1"
-     "\x00\x80"
-     "\x84\x9C\x00\x00\xA6\x9D\x0E\x00\xA5\x9D\x0E\x00\x90\xA1\x00\x80\xAB\xB9"
-     "\x0E\x00"
-     "\xAA\xB1\x0E\x00\x94\xA1\x00\x80\x98\xA1\x00\x80\xAF\x85\x0E\x00\xAE\x85"
-     "\x0E\x00"
-     "\xAD\x91\x0E\x00\xAC\x9D\x0E\x00\x9C\xA1\x00\x80\xA0\xA1\x00\x80\xA4\xA1"
-     "\x00\x80"
-     "\xA8\xA1\x00\x80\xAC\xA1\x00\x80\xB0\xA1\x00\x80\xB4\xA1\x00\x80\xB8\xA1"
-     "\x00\x80"
-     "\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\xBC\xA1\x00\x80\xC0\xA1"
-     "\x00\x80"
-     "\xC4\xA1\x00\x80\x87\xD8\x01\x00\x86\x00\x04\x00\xA8\x9D\x01\x00\xA9\xD5"
-     "\x01\x00"
-     "\xAA\xD5\x01\x00\xAB\xED\x01\x00\xAC\xF5\x01\x00\xAD\x1D\x01\x00\xAE\x15"
-     "\x01\x00"
-     "\xAF\x0D\x01\x00\xC8\xA1\x00\x80\xCC\xA1\x00\x80\xD0\xA1\x00\x80\xD4\xA1"
-     "\x00\x80"
-     "\xD8\xA1\x00\x80\xDC\xA1\x00\x80\xE0\xA1\x00\x80\xE4\xA1\x00\x80\xB8\xED"
-     "\x00\x00"
-     "\xB9\x85\x00\x00\xBA\x8D\x00\x00\xBB\x85\x00\x00\xBC\x9D\x00\x00\xBD\x85"
-     "\x00\x00"
-     "\xBE\x8D\x00\x00\xBF\x85\x00\x00\xB0\x75\x01\x00\xB1\x7D\x01\x00\xB2\x75"
-     "\x01\x00"
-     "\xB3\xE9\x00\x00\xB4\xF9\x00\x00\xB5\xF9\x00\x00\xB6\xE5\x00\x00\xB7\xD5"
-     "\x00\x00"
-     "\xB3\x5D\x02\x00\xE8\xA1\x00\x80\xEC\xA1\x00\x80\xF0\xA1\x00\x80\xF4\xA1"
-     "\x00\x80"
-     "\xB6\x9D\x02\x00\xB5\x7D\x02\x00\xF8\xA1\x00\x80\xBB\xB9\x02\x00\xBA\xB1"
-     "\x02\x00"
-     "\xFC\xA1\x00\x80\x00\xA2\x00\x80\xBF\x3D\x03\x00\xBE\x3D\x03\x00\xBD\x3D"
-     "\x03\x00"
-     "\xBC\xA1\x02\x00\xBE\xCC\x04\x00\xA3\x19\x02\x00\x04\xA2\x00\x80\x08\xA2"
-     "\x00\x80"
-     "\xA6\xD9\x02\x00\x0C\xA2\x00\x80\x10\xA2\x00\x80\xA5\x39\x02\x00\xAA\xF5"
-     "\x02\x00"
-     "\xAB\xFD\x02\x00\x14\xA2\x00\x80\x18\xA2\x00\x80\xAE\x79\x03\x00\xAF\x79"
-     "\x03\x00"
-     "\xAC\xE5\x02\x00\xAD\x79\x03\x00\xBE\x0C\x05\x00\x1C\xA2\x00\x80\x20\xA2"
-     "\x00\x80"
-     "\xBE\x10\x06\x00\x24\xA2\x00\x80\x28\xA2\x00\x80\x2C\xA2\x00\x80\x30\xA2"
-     "\x00\x80"
-     "\x80\x35\x00\x00\x81\x3D\x00\x00\x82\x35\x00\x00\x38\xA2\x00\x80\x86\xB0"
-     "\x04\x00"
-     "\x87\x34\x03\x00\x3C\xA2\x00\x80\x40\xA2\x00\x80\x44\xA2\x00\x80\x48\xA2"
-     "\x00\x80"
-     "\xEF\x70\x02\x00\x4C\xA2\x00\x80\x50\xA2\x00\x80\x54\xA2\x00\x80\x58\xA2"
-     "\x00\x80"
-     "\x5C\xA2\x00\x80\x60\xA2\x00\x80\x64\xA2\x00\x80\x68\xA2\x00\x80\x6C\xA2"
-     "\x00\x80"
-     "\xE1\x94\x01\x00\x70\xA2\x00\x80\xE3\xF8\x01\x00\x74\xA2\x00\x80\xE1\x30"
-     "\x01\x00"
-     "\xE1\x14\x0E\x00\xE3\x54\x0E\x00\xE3\x0C\x0F\x00\x78\xA2\x00\x80\x7C\xA2"
-     "\x00\x80"
-     "\x80\xA2\x00\x80\x84\xA2\x00\x80\x84\xA8\x05\x00\x88\xA2\x00\x80\x8C\xA2"
-     "\x00\x80"
-     "\x90\xA2\x00\x80\x94\xA2\x00\x80\x98\xA2\x00\x80\xEF\x40\x0E\x00\xEF\x8C"
-     "\x0E\x00"
-     "\xB3\xE9\x02\x00\x9C\xA2\x00\x80\xA0\xA2\x00\x80\xA8\xA2\x00\x80\xAC\xA2"
-     "\x00\x80"
-     "\xB6\xE1\x02\x00\xB5\xE9\x02\x00\xB0\xA2\x00\x80\xBB\x69\x01\x00\xBA\x61"
-     "\x01\x00"
-     "\xB4\xA2\x00\x80\xB8\xA2\x00\x80\xBF\x59\x01\x00\xBE\x69\x01\x00\xBD\x69"
-     "\x01\x00"
-     "\xBC\x71\x01\x00\x80\xB5\x00\x00\x81\xC1\x00\x00\x82\xC1\x00\x00\x84\x04"
-     "\x04\x00"
-     "\x86\xB0\x04\x00\x87\x08\x05\x00\xBC\xA2\x00\x80\xC0\xA2\x00\x80\xAC\x31"
-     "\x0E\x00"
-     "\xAD\x39\x0E\x00\xAE\x01\x0E\x00\xAF\x01\x0E\x00\xA8\xB5\x0E\x00\xA9\xAD"
-     "\x0E\x00"
-     "\xAA\xA5\x0E\x00\xAB\x31\x0E\x00\xB4\x19\x0E\x00\xB5\x2D\x0E\x00\xB6\x25"
-     "\x0E\x00"
-     "\xB7\xD9\x0D\x00\xB0\x11\x0E\x00\xB1\x1D\x0E\x00\xB2\x15\x0E\x00\xB3\x09"
-     "\x0E\x00"
-     "\xBC\xCD\x0D\x00\xBD\xD5\x0D\x00\xBE\xD5\x0D\x00\xBF\x69\x0D\x00\xB8\xE9"
-     "\x0D\x00"
-     "\xB9\xE9\x0D\x00\xBA\xDD\x0D\x00\xBB\xD5\x0D\x00\xA3\xA9\x05\x00\xC4\xA2"
-     "\x00\x80"
-     "\xC8\xA2\x00\x80\xCC\xA2\x00\x80\xD0\xA2\x00\x80\xA6\xA1\x05\x00\xA5\xA9"
-     "\x05\x00"
-     "\xD4\xA2\x00\x80\xAB\x29\x06\x00\xAA\x21\x06\x00\xD8\xA2\x00\x80\xDC\xA2"
-     "\x00\x80"
-     "\xAF\x19\x06\x00\xAE\x29\x06\x00\xAD\x29\x06\x00\xAC\x31\x06\x00\xE0\xA2"
-     "\x00\x80"
-     "\xB3\xF9\x06\x00\xE4\xA2\x00\x80\xE8\xA2\x00\x80\xB6\x9D\x06\x00\xEC\xA2"
-     "\x00\x80"
-     "\xF0\xA2\x00\x80\xB5\x99\x06\x00\xBA\xB9\x06\x00\xBB\xB9\x06\x00\xF4\xA2"
-     "\x00\x80"
-     "\xF8\xA2\x00\x80\xBE\x99\x06\x00\xBF\x99\x06\x00\xBC\x99\x06\x00\xBD\x99"
-     "\x06\x00"
-     "\x34\xA2\x00\x80\xA4\xA2\x00\x80\xFC\xA2\x00\x80\x00\xA3\x00\x80\x04\xA3"
-     "\x00\x80"
-     "\x82\x99\x01\x00\x81\x91\x01\x00\x80\x99\x01\x00\xA8\x0D\x06\x00\xA9\x1D"
-     "\x06\x00"
-     "\xAA\x15\x06\x00\xAB\x6D\x06\x00\xAC\x75\x06\x00\xAD\x79\x06\x00\xAE\x69"
-     "\x06\x00"
-     "\xAF\x69\x06\x00\xB0\x19\x06\x00\xB1\x19\x06\x00\xB2\x29\x06\x00\xB3\x29"
-     "\x06\x00"
-     "\xB4\x39\x06\x00\xB5\x39\x06\x00\xB6\x2D\x06\x00\xB7\x21\x06\x00\xB8\xF1"
-     "\x07\x00"
-     "\xB9\xF1\x07\x00\xBA\xF1\x07\x00\xBB\xF1\x07\x00\xBC\xD5\x07\x00\xBD\xDD"
-     "\x07\x00"
-     "\xBE\xD5\x07\x00\xBF\xCD\x07\x00\xA3\xBD\x07\x00\x08\xA3\x00\x80\x0C\xA3"
-     "\x00\x80"
-     "\x10\xA3\x00\x80\x14\xA3\x00\x80\xA6\xD9\x07\x00\xA5\xDD\x07\x00\x18\xA3"
-     "\x00\x80"
-     "\xAB\xFD\x07\x00\xAA\xFD\x07\x00\x86\x00\x3C\x00\x87\x7C\x01\x00\xAF\xDD"
-     "\x07\x00"
-     "\xAE\xDD\x07\x00\xAD\xDD\x07\x00\xAC\xDD\x07\x00\x1C\xA3\x00\x80\xB3\x19"
-     "\x06\x00"
-     "\x20\xA3\x00\x80\x24\xA3\x00\x80\xB6\xC9\x06\x00\x28\xA3\x00\x80\x2C\xA3"
-     "\x00\x80"
-     "\xB5\x31\x06\x00\xBA\x69\x01\x00\xBB\x69\x01\x00\x30\xA3\x00\x80\x34\xA3"
-     "\x00\x80"
-     "\xBE\x59\x01\x00\xBF\x59\x01\x00\xBC\x69\x01\x00\xBD\x61\x01\x00\xA8\xCD"
-     "\x06\x00"
-     "\xA9\xD5\x06\x00\xAA\xD5\x06\x00\xAB\xE5\x06\x00\xAC\xFD\x06\x00\xAD\x0D"
-     "\x06\x00"
-     "\xAE\x05\x06\x00\xAF\x99\x01\x00\x38\xA3\x00\x80\x3C\xA3\x00\x80\x40\xA3"
-     "\x00\x80"
-     "\x44\xA3\x00\x80\x48\xA3\x00\x80\x4C\xA3\x00\x80\x50\xA3\x00\x80\x54\xA3"
-     "\x00\x80"
-     "\xB8\x8D\x01\x00\xB9\x91\x01\x00\xBA\x91\x01\x00\xBB\xA5\x01\x00\xBC\xBD"
-     "\x01\x00"
-     "\xBD\x5D\x01\x00\xBE\x55\x01\x00\xBF\x4D\x01\x00\xB0\xE9\x01\x00\xB1\xE9"
-     "\x01\x00"
-     "\xB2\xF9\x01\x00\xB3\xF1\x01\x00\xB4\xD9\x01\x00\xB5\xD9\x01\x00\xB6\xB9"
-     "\x01\x00"
-     "\xB7\xB5\x01\x00\xBE\xCC\x03\x00\xA6\x9D\x05\x00\xA5\x65\x05\x00\x58\xA3"
-     "\x00\x80"
-     "\xA3\x4D\x05\x00\x5C\xA3\x00\x80\x60\xA3\x00\x80\x64\xA3\x00\x80\xAF\x0D"
-     "\x02\x00"
-     "\xAE\x0D\x02\x00\xAD\x35\x02\x00\xAC\x3D\x02\x00\xAB\x3D\x02\x00\xAA\x3D"
-     "\x02\x00"
-     "\x68\xA3\x00\x80\x6C\xA3\x00\x80\x70\xA3\x00\x80\x74\xA3\x00\x80\x78\xA3"
-     "\x00\x80"
-     "\x7C\xA3\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x80\xA3"
-     "\x00\x80"
-     "\x84\x60\x02\x00\x88\xA3\x00\x80\x87\x48\x03\x00\x86\xCC\x3C\x00\x8C\xA3"
-     "\x00\x80"
-     "\x90\xA3\x00\x80\x94\xA3\x00\x80\x98\xA3\x00\x80\x9C\xA3\x00\x80\xA0\xA3"
-     "\x00\x80"
-     "\x84\xC0\x3C\x00\xA6\x00\x00\x00\xA4\xA3\x00\x80\xE1\xA4\x01\x00\xA8\xA3"
-     "\x00\x80"
-     "\xE3\xEC\x02\x00\xAC\xA3\x00\x80\xB0\xA3\x00\x80\xB4\xA3\x00\x80\xEF\x54"
-     "\x02\x00"
-     "\xB8\xA3\x00\x80\xBC\xA3\x00\x80\xC0\xA3\x00\x80\xC4\xA3\x00\x80\xC8\xA3"
-     "\x00\x80"
-     "\xCC\xA3\x00\x80\xD0\xA3\x00\x80\xB3\x85\x03\x00\xD4\xA3\x00\x80\xB5\x81"
-     "\x03\x00"
-     "\xB6\x81\x03\x00\xD8\xA3\x00\x80\xDC\xA3\x00\x80\xE0\xA3\x00\x80\xBA\x49"
-     "\x03\x00"
-     "\xBB\x49\x03\x00\xBC\x59\x03\x00\xBD\x59\x03\x00\xBE\x49\x03\x00\xBF\x49"
-     "\x03\x00"
-     "\xE1\x88\x06\x00\xE1\xAC\x07\x00\xE3\x94\x06\x00\xE3\x3C\x06\x00\xE4\xA3"
-     "\x00\x80"
-     "\xE8\xA3\x00\x80\xEC\xA3\x00\x80\xF0\xA3\x00\x80\x80\x35\x00\x00\x81\x3D"
-     "\x00\x00"
-     "\x82\x35\x00\x00\xBE\x60\x3D\x00\x86\x90\x3C\x00\x87\xF8\x3D\x00\xEF\x4C"
-     "\x07\x00"
-     "\xEF\x44\x07\x00\xA3\x09\x02\x00\xF8\xA3\x00\x80\xFC\xA3\x00\x80\x00\xA4"
-     "\x00\x80"
-     "\x04\xA4\x00\x80\xA6\x0D\x02\x00\xA5\x0D\x02\x00\x08\xA4\x00\x80\xAB\xC5"
-     "\x02\x00"
-     "\xAA\xC5\x02\x00\x0C\xA4\x00\x80\x10\xA4\x00\x80\xAF\xC5\x02\x00\xAE\xC5"
-     "\x02\x00"
-     "\xAD\xD5\x02\x00\xAC\xD5\x02\x00\xBF\x11\x05\x00\xBE\x0D\x05\x00\xBD\x0D"
-     "\x05\x00"
-     "\xBC\xE1\x05\x00\xBB\xF9\x05\x00\xBA\xCD\x05\x00\xB9\xCD\x05\x00\xB8\xD1"
-     "\x05\x00"
-     "\xB7\xE9\x05\x00\xB6\x89\x05\x00\xB5\x81\x05\x00\xB4\xB1\x05\x00\xB3\xA1"
-     "\x05\x00"
-     "\xB2\xB9\x05\x00\xB1\xB1\x05\x00\xB0\xB9\x05\x00\xAF\x7D\x06\x00\xAE\x7D"
-     "\x06\x00"
-     "\xAD\x7D\x06\x00\xAC\x7D\x06\x00\xAB\x7D\x06\x00\xAA\x7D\x06\x00\xA9\x7D"
-     "\x06\x00"
-     "\xA8\x7D\x06\x00\x14\xA4\x00\x80\x18\xA4\x00\x80\x1C\xA4\x00\x80\x20\xA4"
-     "\x00\x80"
-     "\x24\xA4\x00\x80\x28\xA4\x00\x80\x2C\xA4\x00\x80\x30\xA4\x00\x80\xA8\x65"
-     "\x3E\x00"
-     "\xA9\x75\x3E\x00\xAA\x7D\x3E\x00\xAB\x75\x3E\x00\xAC\x6D\x3E\x00\xAD\x95"
-     "\x3E\x00"
-     "\xAE\x9D\x3E\x00\xAF\x8D\x3E\x00\x34\xA4\x00\x80\x38\xA4\x00\x80\x3C\xA4"
-     "\x00\x80"
-     "\x40\xA4\x00\x80\x44\xA4\x00\x80\x48\xA4\x00\x80\x4C\xA4\x00\x80\x50\xA4"
-     "\x00\x80"
-     "\xB8\x81\x3E\x00\xB9\x81\x3E\x00\xBA\x81\x3E\x00\xBB\x9D\x3E\x00\xBC\x8D"
-     "\x3E\x00"
-     "\xBD\xB5\x3E\x00\xBE\xBD\x3E\x00\xBF\x4D\x3F\x00\xB0\xF9\x3E\x00\xB1\xC5"
-     "\x3E\x00"
-     "\xB2\xCD\x3E\x00\xB3\xC5\x3E\x00\xB4\xC1\x3E\x00\xB5\xC1\x3E\x00\xB6\xC1"
-     "\x3E\x00"
-     "\xB7\xC1\x3E\x00\x84\xA3\x00\x80\xB3\xF1\x3E\x00\x54\xA4\x00\x80\xF4\xA3"
-     "\x00\x80"
-     "\xB6\x11\x3E\x00\x58\xA4\x00\x80\x5C\xA4\x00\x80\xB5\xE1\x3E\x00\xBA\x39"
-     "\x3E\x00"
-     "\xBB\x05\x3E\x00\x60\xA4\x00\x80\x84\xE0\x03\x00\xBE\x01\x3E\x00\xBF\x01"
-     "\x3E\x00"
-     "\xBC\x1D\x3E\x00\xBD\x15\x3E\x00\xA3\xB5\x3E\x00\x82\x6D\x00\x00\x81\x51"
-     "\x00\x00"
-     "\x80\x59\x00\x00\xBE\x98\x03\x00\xA6\x55\x3E\x00\xA5\xA5\x3E\x00\x64\xA4"
-     "\x00\x80"
-     "\xAB\x41\x3E\x00\xAA\x7D\x3E\x00\x86\x48\x00\x00\x87\x68\x00\x00\xAF\x45"
-     "\x3E\x00"
-     "\xAE\x45\x3E\x00\xAD\x51\x3E\x00\xAC\x59\x3E\x00\xA8\x45\x3E\x00\xA9\x4D"
-     "\x3E\x00"
-     "\xAA\x45\x3E\x00\xAB\x5D\x3E\x00\xAC\x4D\x3E\x00\xAD\x7D\x3E\x00\xAE\x75"
-     "\x3E\x00"
-     "\xAF\xFD\x3E\x00\x68\xA4\x00\x80\x6C\xA4\x00\x80\x84\x6C\x01\x00\x70\xA4"
-     "\x00\x80"
-     "\x74\xA4\x00\x80\x78\xA4\x00\x80\x7C\xA4\x00\x80\x80\xA4\x00\x80\xB8\x65"
-     "\x01\x00"
-     "\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x7D\x01\x00\xBC\x65\x01\x00\xBD\x6D"
-     "\x01\x00"
-     "\xBE\x65\x01\x00\xBF\xD5\x01\x00\xB0\x85\x3E\x00\xB1\x8D\x3E\x00\xB2\x85"
-     "\x3E\x00"
-     "\xB3\x9D\x3E\x00\xB4\x85\x3E\x00\xB5\x8D\x3E\x00\xB6\x85\x3E\x00\xB7\x5D"
-     "\x01\x00"
-     "\xB3\xCD\x3E\x00\x84\xA4\x00\x80\x88\xA4\x00\x80\x8C\xA4\x00\x80\x90\xA4"
-     "\x00\x80"
-     "\xB6\xFD\x3E\x00\xB5\xC5\x3E\x00\x94\xA4\x00\x80\xBB\x49\x3E\x00\xBA\x41"
-     "\x3E\x00"
-     "\x98\xA4\x00\x80\x9C\xA4\x00\x80\xBF\x49\x3E\x00\xBE\x41\x3E\x00\xBD\x49"
-     "\x3E\x00"
-     "\xBC\x51\x3E\x00\xA0\xA4\x00\x80\xA3\x89\x3E\x00\xA4\xA4\x00\x80\xA8\xA4"
-     "\x00\x80"
-     "\xA6\xB9\x3E\x00\xAC\xA4\x00\x80\xB0\xA4\x00\x80\xA5\x81\x3E\x00\xAA\x05"
-     "\x3E\x00"
-     "\xAB\x0D\x3E\x00\xB4\xA4\x00\x80\xB8\xA4\x00\x80\xAE\x05\x3E\x00\xAF\x0D"
-     "\x3E\x00"
-     "\xAC\x15\x3E\x00\xAD\x0D\x3E\x00\xA8\x21\x03\x00\xA9\x21\x03\x00\xAA\x25"
-     "\x03\x00"
-     "\xAB\x3D\x03\x00\xAC\x25\x03\x00\xAD\x2D\x03\x00\xAE\x25\x03\x00\xAF\x99"
-     "\x03\x00"
-     "\xBC\xA4\x00\x80\xC0\xA4\x00\x80\xBE\x64\x02\x00\xC4\xA4\x00\x80\xC8\xA4"
-     "\x00\x80"
-     "\xCC\xA4\x00\x80\xD0\xA4\x00\x80\xD8\xA4\x00\x80\xB8\x8D\x03\x00\xB9\x91"
-     "\x03\x00"
-     "\xBA\x91\x03\x00\xBB\xA5\x03\x00\xBC\xBD\x03\x00\xBD\x75\x00\x00\xBE\x7D"
-     "\x00\x00"
-     "\xBF\x71\x00\x00\xB0\xE9\x03\x00\xB1\xE9\x03\x00\xB2\xF9\x03\x00\xB3\xF1"
-     "\x03\x00"
-     "\xB4\xD9\x03\x00\xB5\xD9\x03\x00\xB6\xB9\x03\x00\xB7\xB5\x03\x00\x81\xB5"
-     "\x00\x00"
-     "\x80\xB5\x00\x00\xB3\xA5\x03\x00\x82\xB5\x00\x00\xB5\xA5\x03\x00\xDC\xA4"
-     "\x00\x80"
-     "\xE0\xA4\x00\x80\xB6\xA5\x03\x00\x86\x00\x04\x00\x87\xD4\x03\x00\xBB\x25"
-     "\x03\x00"
-     "\xBA\x25\x03\x00\xBD\x2D\x03\x00\xBC\x25\x03\x00\xBF\x15\x03\x00\xBE\x15"
-     "\x03\x00"
-     "\xA3\xED\x03\x00\xE4\xA4\x00\x80\xE8\xA4\x00\x80\x85\x60\x11\x00\xEC\xA4"
-     "\x00\x80"
-     "\xA6\xED\x03\x00\xA5\xED\x03\x00\xF0\xA4\x00\x80\xAB\x6D\x03\x00\xAA\x6D"
-     "\x03\x00"
-     "\xF4\xA4\x00\x80\xF8\xA4\x00\x80\xAF\x5D\x03\x00\xAE\x5D\x03\x00\xAD\x65"
-     "\x03\x00"
-     "\xAC\x6D\x03\x00\xE3\xBC\x01\x00\xE1\x70\x3F\x00\xE1\x20\x3E\x00\xE3\x94"
-     "\x3F\x00"
-     "\xFC\xA4\x00\x80\x00\xA5\x00\x80\x04\xA5\x00\x80\x08\xA5\x00\x80\x0C\xA5"
-     "\x00\x80"
-     "\x10\xA5\x00\x80\x14\xA5\x00\x80\x18\xA5\x00\x80\xEF\x18\x01\x00\x1C\xA5"
-     "\x00\x80"
-     "\xBF\x38\x13\x00\xEF\xE8\x3E\x00\x20\xA5\x00\x80\x24\xA5\x00\x80\x84\x04"
-     "\x05\x00"
-     "\xBE\x04\x05\x00\xE3\x68\x00\x00\x28\xA5\x00\x80\xE1\xA8\x01\x00\x2C\xA5"
-     "\x00\x80"
-     "\xEF\xB0\x03\x00\x30\xA5\x00\x80\x34\xA5\x00\x80\x38\xA5\x00\x80\x80\x1D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\xDD\x00\x00\x3C\xA5\x00\x80\xAA\xC9\x3E\x00\xAB\xC9"
-     "\x3E\x00"
-     "\xA8\x95\x3E\x00\xA9\xCD\x3E\x00\xAE\xF9\x3E\x00\xAF\xF9\x3E\x00\xAC\xCD"
-     "\x3E\x00"
-     "\xAD\xC5\x3E\x00\xBF\xA8\xEF\x00\x40\xA5\x00\x80\x86\x88\x04\x00\x87\xAC"
-     "\x04\x00"
-     "\x48\xA5\x00\x80\x4C\xA5\x00\x80\x50\xA5\x00\x80\x54\xA5\x00\x80\xBA\xD9"
-     "\x3D\x00"
-     "\xBB\xD9\x3D\x00\xB8\xC5\x3D\x00\xB9\xD9\x3D\x00\xBE\x91\x3D\x00\xBF\x91"
-     "\x3D\x00"
-     "\xBC\xE1\x3D\x00\xBD\xE1\x3D\x00\xB2\x49\x3E\x00\xB3\x49\x3E\x00\xB0\x49"
-     "\x3E\x00"
-     "\xB1\x49\x3E\x00\xB6\x49\x3E\x00\xB7\x49\x3E\x00\xB4\x49\x3E\x00\xB5\x49"
-     "\x3E\x00"
-     "\xA8\x3D\x06\x00\xA9\x91\x06\x00\xAA\x91\x06\x00\xAB\xC9\x06\x00\xAC\xD9"
-     "\x06\x00"
-     "\xAD\xC5\x06\x00\xAE\xC5\x06\x00\xAF\xF5\x06\x00\x44\xA5\x00\x80\x58\xA5"
-     "\x00\x80"
-     "\x5C\xA5\x00\x80\x60\xA5\x00\x80\x64\xA5\x00\x80\x68\xA5\x00\x80\x6C\xA5"
-     "\x00\x80"
-     "\x70\xA5\x00\x80\xB8\x6D\x07\x00\xB9\x71\x07\x00\xBA\x09\x07\x00\xBB\x09"
-     "\x07\x00"
-     "\xBC\x05\x07\x00\xBD\x09\x07\x00\xBE\x3D\x07\x00\xBF\x35\x07\x00\xB0\x8D"
-     "\x06\x00"
-     "\xB1\x75\x07\x00\xB2\x7D\x07\x00\xB3\x75\x07\x00\xB4\x6D\x07\x00\xB5\x55"
-     "\x07\x00"
-     "\xB6\x5D\x07\x00\xB7\x55\x07\x00\xB3\x19\x06\x00\x74\xA5\x00\x80\x78\xA5"
-     "\x00\x80"
-     "\x7C\xA5\x00\x80\x80\xA5\x00\x80\xB6\x45\x06\x00\xB5\x51\x06\x00\x84\xA5"
-     "\x00\x80"
-     "\xBB\xB1\x07\x00\xBA\x49\x06\x00\x88\xA5\x00\x80\x8C\xA5\x00\x80\xBF\x91"
-     "\x07\x00"
-     "\xBE\xA9\x07\x00\xBD\xA1\x07\x00\xBC\xA9\x07\x00\x90\xA5\x00\x80\xA3\x5D"
-     "\x06\x00"
-     "\x94\xA5\x00\x80\x98\xA5\x00\x80\xA6\x01\x06\x00\x9C\xA5\x00\x80\xA0\xA5"
-     "\x00\x80"
-     "\xA5\x15\x06\x00\xAA\x0D\x06\x00\xAB\xF5\x07\x00\xA4\xA5\x00\x80\xA8\xA5"
-     "\x00\x80"
-     "\xAE\xED\x07\x00\xAF\xD5\x07\x00\xAC\xED\x07\x00\xAD\xE5\x07\x00\x80\xED"
-     "\x01\x00"
-     "\x81\x09\x00\x00\x82\x19\x00\x00\xAC\xA5\x00\x80\xB0\xA5\x00\x80\x84\xD0"
-     "\x01\x00"
-     "\xBE\xC4\x01\x00\xB4\xA5\x00\x80\x86\x80\x00\x00\x87\x20\x01\x00\xB8\xA5"
-     "\x00\x80"
-     "\xBC\xA5\x00\x80\xC0\xA5\x00\x80\xC4\xA5\x00\x80\xC8\xA5\x00\x80\xCC\xA5"
-     "\x00\x80"
-     "\xD0\xA5\x00\x80\xE1\xFC\x05\x00\xD4\xA5\x00\x80\xE3\x14\x04\x00\xD8\xA5"
-     "\x00\x80"
-     "\xDC\xA5\x00\x80\xE0\xA5\x00\x80\xE4\xA5\x00\x80\xE8\xA5\x00\x80\xEC\xA5"
-     "\x00\x80"
-     "\xF0\xA5\x00\x80\xF4\xA5\x00\x80\xF8\xA5\x00\x80\xFC\xA5\x00\x80\x00\xA6"
-     "\x00\x80"
-     "\xEF\xC8\x04\x00\xA8\x31\x06\x00\xA9\x31\x06\x00\xAA\x31\x06\x00\xAB\x31"
-     "\x06\x00"
-     "\xAC\xB5\x01\x00\xAD\xBD\x01\x00\xAE\xB5\x01\x00\xAF\xAD\x01\x00\x84\xA8"
-     "\x01\x00"
-     "\x04\xA6\x00\x80\x08\xA6\x00\x80\x0C\xA6\x00\x80\x10\xA6\x00\x80\x14\xA6"
-     "\x00\x80"
-     "\x18\xA6\x00\x80\x1C\xA6\x00\x80\xB8\x69\x01\x00\xB9\x69\x01\x00\xBA\x09"
-     "\x01\x00"
-     "\xBB\x05\x01\x00\xBC\x1D\x01\x00\xBD\x01\x01\x00\xBE\x01\x01\x00\xBF\x35"
-     "\x01\x00"
-     "\xB0\xD5\x01\x00\xB1\xDD\x01\x00\xB2\xD5\x01\x00\xB3\x69\x01\x00\xB4\x79"
-     "\x01\x00"
-     "\xB5\x79\x01\x00\xB6\x69\x01\x00\xB7\x61\x01\x00\x81\xBD\x03\x00\x80\xBD"
-     "\x03\x00"
-     "\xB3\xB1\x05\x00\x82\x19\x00\x00\xB5\x35\x02\x00\x20\xA6\x00\x80\xBE\x54"
-     "\x03\x00"
-     "\xB6\x35\x02\x00\x28\xA6\x00\x80\x2C\xA6\x00\x80\xBB\x11\x02\x00\xBA\x19"
-     "\x02\x00"
-     "\xBD\xB5\x02\x00\xBC\x09\x02\x00\xBF\xB5\x02\x00\xBE\xB5\x02\x00\x30\xA6"
-     "\x00\x80"
-     "\xA6\x6D\x02\x00\xA5\x6D\x02\x00\x84\x00\x0C\x00\xA3\xE9\x05\x00\x34\xA6"
-     "\x00\x80"
-     "\x86\xF8\x0C\x00\x87\x7C\x03\x00\xAF\xED\x02\x00\xAE\xED\x02\x00\xAD\xED"
-     "\x02\x00"
-     "\xAC\x51\x02\x00\xAB\x49\x02\x00\xAA\x41\x02\x00\x38\xA6\x00\x80\x3C\xA6"
-     "\x00\x80"
-     "\x40\xA6\x00\x80\x44\xA6\x00\x80\x48\xA6\x00\x80\x4C\xA6\x00\x80\xE3\x20"
-     "\x01\x00"
-     "\x50\xA6\x00\x80\xE1\xA0\x01\x00\x54\xA6\x00\x80\xEF\x40\x02\x00\x58\xA6"
-     "\x00\x80"
-     "\x5C\xA6\x00\x80\x60\xA6\x00\x80\x64\xA6\x00\x80\x68\xA6\x00\x80\x6C\xA6"
-     "\x00\x80"
-     "\x70\xA6\x00\x80\xD4\xA4\x00\x80\xB3\xF5\x03\x00\x74\xA6\x00\x80\x78\xA6"
-     "\x00\x80"
-     "\xB6\xED\x03\x00\x7C\xA6\x00\x80\x80\xA6\x00\x80\xB5\xE5\x03\x00\xBA\xB1"
-     "\x03\x00"
-     "\xBB\xB1\x03\x00\x84\xA6\x00\x80\x88\xA6\x00\x80\xBE\x91\x03\x00\xBF\x91"
-     "\x03\x00"
-     "\xBC\xA1\x03\x00\xBD\xA1\x03\x00\xA8\x29\x02\x00\xA9\x29\x02\x00\xAA\x39"
-     "\x02\x00"
-     "\xAB\x39\x02\x00\xAC\x29\x02\x00\xAD\x29\x02\x00\xAE\x59\x02\x00\xAF\x55"
-     "\x02\x00"
-     "\x80\xCD\x01\x00\x81\x09\x00\x00\x82\x19\x00\x00\x8C\xA6\x00\x80\x90\xA6"
-     "\x00\x80"
-     "\xBE\x7C\x0D\x00\x87\xB4\x0D\x00\x86\x1C\x0C\x00\xB8\x31\x02\x00\xB9\x31"
-     "\x02\x00"
-     "\xBA\x31\x02\x00\xBB\xE9\x02\x00\xBC\xF9\x02\x00\xBD\xF9\x02\x00\xBE\xE9"
-     "\x02\x00"
-     "\xBF\xE9\x02\x00\xB0\x31\x02\x00\xB1\x31\x02\x00\xB2\x31\x02\x00\xB3\x31"
-     "\x02\x00"
-     "\xB4\x11\x02\x00\xB5\x11\x02\x00\xB6\x11\x02\x00\xB7\x11\x02\x00\x98\xA6"
-     "\x00\x80"
-     "\x9C\xA6\x00\x80\xA0\xA6\x00\x80\xA4\xA6\x00\x80\xA8\xA6\x00\x80\xAC\xA6"
-     "\x00\x80"
-     "\xB0\xA6\x00\x80\xEF\x7C\x00\x00\xB4\xA6\x00\x80\xE1\x14\x06\x00\xB8\xA6"
-     "\x00\x80"
-     "\xE3\xA8\x00\x00\xEC\x00\x00\x00\xBC\xA6\x00\x80\xC0\xA6\x00\x80\xC4\xA6"
-     "\x00\x80"
-     "\xA3\xF1\x02\x00\xC8\xA6\x00\x80\xCC\xA6\x00\x80\xD0\xA6\x00\x80\xD4\xA6"
-     "\x00\x80"
-     "\xA6\xE9\x02\x00\xA5\xE1\x02\x00\xD8\xA6\x00\x80\xAB\xB5\x02\x00\xAA\xB5"
-     "\x02\x00"
-     "\xDC\xA6\x00\x80\xE0\xA6\x00\x80\xAF\x95\x02\x00\xAE\x95\x02\x00\xAD\xA5"
-     "\x02\x00"
-     "\xAC\xA5\x02\x00\xA8\x21\x0E\x00\xA9\x21\x0E\x00\xAA\x49\x0E\x00\xAB\x59"
-     "\x0E\x00"
-     "\xAC\x69\x0E\x00\xAD\x69\x0E\x00\xAE\x99\x0E\x00\xAF\x99\x0E\x00\x94\xA6"
-     "\x00\x80"
-     "\xE4\xA6\x00\x80\xE8\xA6\x00\x80\xEC\xA6\x00\x80\xF0\xA6\x00\x80\xF4\xA6"
-     "\x00\x80"
-     "\xF8\xA6\x00\x80\xFC\xA6\x00\x80\xB8\xF5\x0E\x00\xB9\xFD\x0E\x00\xBA\xF5"
-     "\x0E\x00"
-     "\xBB\x85\x0E\x00\xBC\x9D\x0E\x00\xBD\x81\x0E\x00\xBE\x81\x0E\x00\xBF\xB5"
-     "\x0E\x00"
-     "\xB0\xE9\x0E\x00\xB1\xE9\x0E\x00\xB2\xFD\x0E\x00\xB3\xF1\x0E\x00\xB4\xD5"
-     "\x0E\x00"
-     "\xB5\xDD\x0E\x00\xB6\xD5\x0E\x00\xB7\xCD\x0E\x00\xB3\xC1\x0E\x00\x82\x15"
-     "\x00\x00"
-     "\x81\xB5\x00\x00\x80\xB5\x00\x00\x00\xA7\x00\x80\xB6\xE9\x0E\x00\xB5\xE1"
-     "\x0E\x00"
-     "\xBE\x10\x00\x00\xBB\x2D\x0E\x00\xBA\x2D\x0E\x00\x84\x64\x03\x00\x04\xA7"
-     "\x00\x80"
-     "\xBF\x19\x0E\x00\xBE\x11\x0E\x00\xBD\x25\x0E\x00\xBC\x29\x0E\x00\x08\xA7"
-     "\x00\x80"
-     "\xA3\x85\x0E\x00\x86\x88\x00\x00\x87\x6C\x03\x00\xA6\xAD\x0E\x00\x0C\xA7"
-     "\x00\x80"
-     "\x10\xA7\x00\x80\xA5\xA5\x0E\x00\xAA\x69\x0E\x00\xAB\x69\x0E\x00\x14\xA7"
-     "\x00\x80"
-     "\x18\xA7\x00\x80\xAE\x55\x0E\x00\xAF\x5D\x0E\x00\xAC\x6D\x0E\x00\xAD\x61"
-     "\x0E\x00"
-     "\xB3\x89\x0E\x00\x1C\xA7\x00\x80\x20\xA7\x00\x80\x24\xA7\x00\x80\x28\xA7"
-     "\x00\x80"
-     "\xB6\x81\x0E\x00\xB5\x89\x0E\x00\x2C\xA7\x00\x80\xBB\x95\x0E\x00\xBA\x8D"
-     "\x0E\x00"
-     "\x30\xA7\x00\x80\x34\xA7\x00\x80\xBF\xF9\x0E\x00\xBE\xF1\x0E\x00\xBD\x85"
-     "\x0E\x00"
-     "\xBC\x85\x0E\x00\x38\xA7\x00\x80\x3C\xA7\x00\x80\x40\xA7\x00\x80\x44\xA7"
-     "\x00\x80"
-     "\xE1\x10\x0D\x00\x48\xA7\x00\x80\xE3\x04\x0D\x00\x4C\xA7\x00\x80\x50\xA7"
-     "\x00\x80"
-     "\x54\xA7\x00\x80\xEF\xA4\x0C\x00\x58\xA7\x00\x80\x5C\xA7\x00\x80\x60\xA7"
-     "\x00\x80"
-     "\x64\xA7\x00\x80\x68\xA7\x00\x80\xA8\x01\x0E\x00\xA9\x01\x0E\x00\xAA\x01"
-     "\x0E\x00"
-     "\xAB\x01\x0E\x00\xAC\x01\x0E\x00\xAD\x01\x0E\x00\xAE\x01\x0E\x00\xAF\x3D"
-     "\x0E\x00"
-     "\x80\xDD\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x6C\xA7\x00\x80\x70\xA7"
-     "\x00\x80"
-     "\x84\x3C\x01\x00\xBE\x74\x00\x00\x78\xA7\x00\x80\xB8\x1D\x0E\x00\xB9\x2D"
-     "\x0E\x00"
-     "\xBA\x25\x0E\x00\xBB\xE9\x01\x00\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\xE9"
-     "\x01\x00"
-     "\xBF\xE9\x01\x00\xB0\x49\x0E\x00\xB1\x51\x0E\x00\xB2\x51\x0E\x00\xB3\x51"
-     "\x0E\x00"
-     "\xB4\x35\x0E\x00\xB5\x3D\x0E\x00\xB6\x35\x0E\x00\xB7\x2D\x0E\x00\xA3\x89"
-     "\x0D\x00"
-     "\x7C\xA7\x00\x80\x86\xA8\x04\x00\x87\x3C\x03\x00\x80\xA7\x00\x80\xA6\x81"
-     "\x0D\x00"
-     "\xA5\x89\x0D\x00\x84\xA7\x00\x80\xAB\x95\x0D\x00\xAA\x8D\x0D\x00\x88\xA7"
-     "\x00\x80"
-     "\x8C\xA7\x00\x80\xAF\xF9\x0D\x00\xAE\xF1\x0D\x00\xAD\x85\x0D\x00\xAC\x85"
-     "\x0D\x00"
-     "\x90\xA7\x00\x80\xB3\x9D\x02\x00\x84\x48\x03\x00\xBE\x40\x04\x00\xB6\x55"
-     "\x03\x00"
-     "\x94\xA7\x00\x80\x98\xA7\x00\x80\xB5\xB1\x02\x00\xBA\x71\x03\x00\xBB\x79"
-     "\x03\x00"
-     "\x9C\xA7\x00\x80\xA0\xA7\x00\x80\xBE\x31\x03\x00\xBF\x31\x03\x00\xBC\x51"
-     "\x03\x00"
-     "\xBD\x51\x03\x00\xB0\x91\x03\x00\xB1\x99\x03\x00\xB2\xA1\x03\x00\xB3\xA1"
-     "\x03\x00"
-     "\xB4\xD1\x03\x00\xB5\xD1\x03\x00\xB6\xD1\x03\x00\xB7\xD1\x03\x00\xB8\xF5"
-     "\x03\x00"
-     "\xB9\xF9\x03\x00\xBA\xC9\x03\x00\xBB\xC9\x03\x00\xBC\xDD\x03\x00\xBD\xC5"
-     "\x03\x00"
-     "\xBE\xCD\x03\x00\xBF\xC5\x03\x00\xA4\xA7\x00\x80\xA8\xA7\x00\x80\xAC\xA7"
-     "\x00\x80"
-     "\xB0\xA7\x00\x80\x85\xEC\x1A\x00\x24\xA6\x00\x80\xB4\xA7\x00\x80\xB8\xA7"
-     "\x00\x80"
-     "\xA8\x21\x02\x00\xA9\x31\x02\x00\xAA\x31\x02\x00\xAB\x05\x02\x00\xAC\x1D"
-     "\x02\x00"
-     "\xAD\xF1\x03\x00\xAE\xF1\x03\x00\xAF\xF1\x03\x00\xBC\xA7\x00\x80\xC0\xA7"
-     "\x00\x80"
-     "\xC4\xA7\x00\x80\xC8\xA7\x00\x80\xEF\x50\x00\x00\xCC\xA7\x00\x80\xD0\xA7"
-     "\x00\x80"
-     "\xD4\xA7\x00\x80\xE3\x84\x00\x00\xD8\xA7\x00\x80\xE1\xF8\x01\x00\xDC\xA7"
-     "\x00\x80"
-     "\x80\x15\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\xE0\xA7\x00\x80\xA3\x99"
-     "\x03\x00"
-     "\xE8\xA7\x00\x80\x86\x68\x04\x00\x87\x60\x05\x00\xEC\xA7\x00\x80\xA6\x51"
-     "\x02\x00"
-     "\xA5\xB5\x03\x00\xF0\xA7\x00\x80\xAB\x7D\x02\x00\xAA\x75\x02\x00\xF4\xA7"
-     "\x00\x80"
-     "\xF8\xA7\x00\x80\xAF\x35\x02\x00\xAE\x35\x02\x00\xAD\x55\x02\x00\xAC\x55"
-     "\x02\x00"
-     "\xFC\xA7\x00\x80\x00\xA8\x00\x80\x04\xA8\x00\x80\x08\xA8\x00\x80\x0C\xA8"
-     "\x00\x80"
-     "\x10\xA8\x00\x80\x14\xA8\x00\x80\xEF\xF8\x01\x00\xBE\x6C\x04\x00\xE1\x80"
-     "\x0E\x00"
-     "\x18\xA8\x00\x80\xE3\x14\x01\x00\x1C\xA8\x00\x80\x20\xA8\x00\x80\x24\xA8"
-     "\x00\x80"
-     "\x28\xA8\x00\x80\x2C\xA8\x00\x80\x30\xA8\x00\x80\x34\xA8\x00\x80\xB3\xDD"
-     "\x01\x00"
-     "\x38\xA8\x00\x80\xB5\xFD\x01\x00\xB6\xF5\x01\x00\x3C\xA8\x00\x80\x40\xA8"
-     "\x00\x80"
-     "\x44\xA8\x00\x80\xBA\xB1\x01\x00\xBB\x85\x01\x00\xBC\x9D\x01\x00\xBD\x35"
-     "\x01\x00"
-     "\xBE\x3D\x01\x00\xBF\x35\x01\x00\xA8\xBD\x05\x00\xA9\x01\x06\x00\xAA\x19"
-     "\x06\x00"
-     "\xAB\x11\x06\x00\xAC\x39\x06\x00\xAD\x39\x06\x00\xAE\x59\x06\x00\xAF\x49"
-     "\x06\x00"
-     "\xE4\xA7\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\x48\xA8"
-     "\x00\x80"
-     "\x4C\xA8\x00\x80\x50\xA8\x00\x80\x54\xA8\x00\x80\xB8\x9D\x06\x00\xB9\x45"
-     "\x07\x00"
-     "\xBA\x4D\x07\x00\xBB\x45\x07\x00\xBC\x5D\x07\x00\xBD\x41\x07\x00\xBE\x41"
-     "\x07\x00"
-     "\xBF\x41\x07\x00\xB0\x39\x06\x00\xB1\x0D\x06\x00\xB2\x05\x06\x00\xB3\xAD"
-     "\x06\x00"
-     "\xB4\xB5\x06\x00\xB5\xBD\x06\x00\xB6\xB5\x06\x00\xB7\xA5\x06\x00\xA3\x91"
-     "\x06\x00"
-     "\x84\x8C\x02\x00\x86\x28\x00\x00\x87\xC0\x03\x00\x58\xA8\x00\x80\xA6\xB9"
-     "\x06\x00"
-     "\xA5\xB1\x06\x00\x5C\xA8\x00\x80\xAB\xC9\x06\x00\xAA\xFD\x06\x00\x60\xA8"
-     "\x00\x80"
-     "\x64\xA8\x00\x80\xAF\x79\x06\x00\xAE\x71\x06\x00\xAD\x79\x06\x00\xAC\xD1"
-     "\x06\x00"
-     "\x68\xA8\x00\x80\xB3\x99\x07\x00\x6C\xA8\x00\x80\x70\xA8\x00\x80\xB6\x51"
-     "\x07\x00"
-     "\x74\xA8\x00\x80\x78\xA8\x00\x80\xB5\xB1\x07\x00\xBA\x6D\x07\x00\xBB\x75"
-     "\x07\x00"
-     "\x7C\xA8\x00\x80\x80\xA8\x00\x80\xBE\x59\x07\x00\xBF\x45\x07\x00\xBC\x6D"
-     "\x07\x00"
-     "\xBD\x65\x07\x00\x84\xA8\x00\x80\x88\xA8\x00\x80\x8C\xA8\x00\x80\x90\xA8"
-     "\x00\x80"
-     "\x94\xA8\x00\x80\xEF\xA0\x05\x00\x98\xA8\x00\x80\x9C\xA8\x00\x80\xA0\xA8"
-     "\x00\x80"
-     "\xE3\x7C\x05\x00\xA4\xA8\x00\x80\xE1\x3C\x05\x00\xA8\xA8\x00\x80\xAC\xA8"
-     "\x00\x80"
-     "\xB0\xA8\x00\x80\xB4\xA8\x00\x80\xA3\x1D\x06\x00\x82\x2D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\xB8\xA8\x00\x80\xA6\xD5\x06\x00\xA5\x35\x06\x00\xBC\xA8"
-     "\x00\x80"
-     "\xAB\xF1\x06\x00\xAA\xE9\x06\x00\xC0\xA8\x00\x80\x84\x20\x01\x00\xAF\xC1"
-     "\x06\x00"
-     "\xAE\xDD\x06\x00\xAD\xE1\x06\x00\xAC\xE9\x06\x00\xA8\xC5\x06\x00\xA9\xD5"
-     "\x06\x00"
-     "\xAA\xD5\x06\x00\xAB\xE5\x06\x00\xAC\xFD\x06\x00\xAD\x15\x06\x00\xAE\x1D"
-     "\x06\x00"
-     "\xAF\x15\x06\x00\xBE\xE8\x01\x00\xC8\xA8\x00\x80\x86\x08\x00\x00\x87\x88"
-     "\x00\x00"
-     "\xCC\xA8\x00\x80\xD0\xA8\x00\x80\xD4\xA8\x00\x80\xD8\xA8\x00\x80\xB8\x7D"
-     "\x06\x00"
-     "\xB9\x05\x06\x00\xBA\x0D\x06\x00\xBB\x05\x06\x00\xBC\x01\x06\x00\xBD\x09"
-     "\x06\x00"
-     "\xBE\x39\x06\x00\xBF\x39\x06\x00\xB0\x6D\x06\x00\xB1\x75\x06\x00\xB2\x7D"
-     "\x06\x00"
-     "\xB3\x75\x06\x00\xB4\x59\x06\x00\xB5\x45\x06\x00\xB6\x4D\x06\x00\xB7\x45"
-     "\x06\x00"
-     "\xA8\x91\x02\x00\xA9\x99\x02\x00\xAA\xA1\x02\x00\xAB\xA1\x02\x00\xAC\xD1"
-     "\x02\x00"
-     "\xAD\xDD\x02\x00\xAE\xD5\x02\x00\xAF\xC9\x02\x00\xDC\xA8\x00\x80\xE0\xA8"
-     "\x00\x80"
-     "\xE4\xA8\x00\x80\xBF\x24\xC5\x00\xE8\xA8\x00\x80\xEC\xA8\x00\x80\xF0\xA8"
-     "\x00\x80"
-     "\xF4\xA8\x00\x80\xB8\x59\x03\x00\xB9\x59\x03\x00\xBA\x69\x03\x00\xBB\x65"
-     "\x03\x00"
-     "\xBC\x61\x03\x00\xBD\x61\x03\x00\xBE\x61\x03\x00\xBF\x61\x03\x00\xB0\xB9"
-     "\x02\x00"
-     "\xB1\x8D\x02\x00\xB2\x85\x02\x00\xB3\x6D\x03\x00\xB4\x75\x03\x00\xB5\x7D"
-     "\x03\x00"
-     "\xB6\x75\x03\x00\xB7\x6D\x03\x00\xB3\x29\x02\x00\xF8\xA8\x00\x80\xFC\xA8"
-     "\x00\x80"
-     "\x00\xA9\x00\x80\x84\x84\x03\x00\xB6\x5D\x02\x00\xB5\x59\x02\x00\x04\xA9"
-     "\x00\x80"
-     "\xBB\x49\x02\x00\xBA\x7D\x02\x00\x74\xA7\x00\x80\x08\xA9\x00\x80\xBF\xB9"
-     "\x03\x00"
-     "\xBE\xB1\x03\x00\xBD\xB9\x03\x00\xBC\x51\x02\x00\x0C\xA9\x00\x80\xA3\x6D"
-     "\x02\x00"
-     "\x10\xA9\x00\x80\x14\xA9\x00\x80\xA6\x19\x02\x00\x18\xA9\x00\x80\x1C\xA9"
-     "\x00\x80"
-     "\xA5\x1D\x02\x00\xAA\x39\x02\x00\xAB\x0D\x02\x00\x20\xA9\x00\x80\x24\xA9"
-     "\x00\x80"
-     "\xAE\xF5\x03\x00\xAF\xFD\x03\x00\xAC\x15\x02\x00\xAD\xFD\x03\x00\xA8\xA5"
-     "\x02\x00"
-     "\xA9\xB5\x02\x00\xAA\xBD\x02\x00\xAB\xB5\x02\x00\xAC\xAD\x02\x00\xAD\x35"
-     "\x01\x00"
-     "\xAE\x3D\x01\x00\xAF\x35\x01\x00\x81\x11\x00\x00\x80\x0D\x00\x00\x28\xA9"
-     "\x00\x80"
-     "\x82\x19\x00\x00\x2C\xA9\x00\x80\x30\xA9\x00\x80\xBE\xB4\x1D\x00\x38\xA9"
-     "\x00\x80"
-     "\xB8\xED\x01\x00\xB9\x85\x01\x00\xBA\x81\x01\x00\xBB\x81\x01\x00\xBC\x85"
-     "\x01\x00"
-     "\xBD\x8D\x01\x00\xBE\xB1\x01\x00\xBF\xB1\x01\x00\xB0\x51\x01\x00\xB1\x51"
-     "\x01\x00"
-     "\xB2\x51\x01\x00\xB3\x51\x01\x00\xB4\xF5\x01\x00\xB5\xFD\x01\x00\xB6\xF5"
-     "\x01\x00"
-     "\xB7\xE1\x01\x00\xEF\xD8\x00\x00\x3C\xA9\x00\x80\x86\xA8\x1D\x00\x87\xDC"
-     "\x1D\x00"
-     "\x40\xA9\x00\x80\x44\xA9\x00\x80\x48\xA9\x00\x80\xEF\xC4\x06\x00\x4C\xA9"
-     "\x00\x80"
-     "\xE1\xF4\x06\x00\x50\xA9\x00\x80\xE3\x38\x01\x00\xE3\xCC\x00\x00\x54\xA9"
-     "\x00\x80"
-     "\xE1\x28\x01\x00\x58\xA9\x00\x80\xB5\x7D\x02\x00\x5C\xA9\x00\x80\x60\xA9"
-     "\x00\x80"
-     "\xB6\xED\x01\x00\x64\xA9\x00\x80\x68\xA9\x00\x80\xB3\x6D\x02\x00\x6C\xA9"
-     "\x00\x80"
-     "\xBD\xDD\x01\x00\xBC\xDD\x01\x00\xBF\x61\x01\x00\xBE\x69\x01\x00\x70\xA9"
-     "\x00\x80"
-     "\x74\xA9\x00\x80\xBB\xCD\x01\x00\xBA\xCD\x01\x00\xA3\xB5\x1D\x00\x34\xA9"
-     "\x00\x80"
-     "\x78\xA9\x00\x80\x7C\xA9\x00\x80\x80\xA9\x00\x80\xA6\x35\x1E\x00\xA5\xA5"
-     "\x1D\x00"
-     "\x84\xA9\x00\x80\xAB\x15\x1E\x00\xAA\x15\x1E\x00\x88\xA9\x00\x80\x8C\xA9"
-     "\x00\x80"
-     "\xAF\xB9\x1E\x00\xAE\xB1\x1E\x00\xAD\x05\x1E\x00\xAC\x05\x1E\x00\x90\xA9"
-     "\x00\x80"
-     "\x94\xA9\x00\x80\x98\xA9\x00\x80\x9C\xA9\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x05\x00\x00\xA0\xA9\x00\x80\x84\xEC\x03\x00\xBE\x68\x03\x00\xA4\xA9"
-     "\x00\x80"
-     "\xA8\xA9\x00\x80\x86\xD0\x01\x00\x87\xB4\x00\x00\xAC\xA9\x00\x80\xB0\xA9"
-     "\x00\x80"
-     "\xA8\xBD\x1F\x00\xA9\x01\x1F\x00\xAA\x19\x1F\x00\xAB\x19\x1F\x00\xAC\x09"
-     "\x1F\x00"
-     "\xAD\x35\x1F\x00\xAE\x31\x1F\x00\xAF\x29\x1F\x00\xB4\xA9\x00\x80\xB8\xA9"
-     "\x00\x80"
-     "\xBC\xA9\x00\x80\xC0\xA9\x00\x80\xC4\xA9\x00\x80\xC8\xA9\x00\x80\xCC\xA9"
-     "\x00\x80"
-     "\xD0\xA9\x00\x80\xB8\xC9\x00\x00\xB9\xC9\x00\x00\xBA\xD9\x00\x00\xBB\xD1"
-     "\x00\x00"
-     "\xBC\xF9\x00\x00\xBD\xF9\x00\x00\xBE\x99\x00\x00\xBF\x99\x00\x00\xB0\x5D"
-     "\x1F\x00"
-     "\xB1\x21\x1F\x00\xB2\x21\x1F\x00\xB3\x3D\x1F\x00\xB4\x25\x1F\x00\xB5\x29"
-     "\x1F\x00"
-     "\xB6\x09\x1F\x00\xB7\x09\x1F\x00\xEF\xC0\x1C\x00\xE1\xC0\x19\x00\xD4\xA9"
-     "\x00\x80"
-     "\xE3\xD0\x09\x00\xD8\xA9\x00\x80\xDC\xA9\x00\x80\xE0\xA9\x00\x80\xE4\xA9"
-     "\x00\x80"
-     "\xA3\x25\x01\x00\xA2\x99\x03\x00\xA1\x15\x1D\x00\xA0\x0D\x1D\x00\xE3\xB0"
-     "\x1D\x00"
-     "\xE8\xA9\x00\x80\xE1\x88\x1E\x00\xEF\x60\x10\x00\xB3\xD1\x1E\x00\xEC\xA9"
-     "\x00\x80"
-     "\xF0\xA9\x00\x80\xF4\xA9\x00\x80\xF8\xA9\x00\x80\xB6\xF5\x1E\x00\xB5\xC1"
-     "\x1E\x00"
-     "\xFC\xA9\x00\x80\xBB\xA5\x1E\x00\xBA\xD9\x1E\x00\x00\xAA\x00\x80\x04\xAA"
-     "\x00\x80"
-     "\xBF\xAD\x1E\x00\xBE\xA1\x1E\x00\xBD\xA9\x1E\x00\xBC\xB5\x1E\x00\xA3\x11"
-     "\x1E\x00"
-     "\x08\xAA\x00\x80\x0C\xAA\x00\x80\x10\xAA\x00\x80\x8F\x19\x24\x00\xA6\x35"
-     "\x1E\x00"
-     "\xA5\x01\x1E\x00\xBE\x10\x07\x00\xAB\x65\x1E\x00\xAA\x19\x1E\x00\x84\x00"
-     "\x04\x00"
-     "\x18\xAA\x00\x80\xAF\x6D\x1E\x00\xAE\x61\x1E\x00\xAD\x69\x1E\x00\xAC\x75"
-     "\x1E\x00"
-     "\x9B\xDD\x14\x00\x9A\xE5\x15\x00\x99\x09\x17\x00\x98\x11\x11\x00\x9F\x79"
-     "\x1F\x00"
-     "\x9E\x79\x19\x00\x9D\xF9\x19\x00\x9C\x09\x1B\x00\x93\x61\x2D\x00\x92\x91"
-     "\x28\x00"
-     "\x91\x6D\x29\x00\x90\x6D\x29\x00\x97\x7D\x11\x00\x96\x75\x11\x00\x95\xE9"
-     "\x2C\x00"
-     "\x94\x79\x2D\x00\x86\x20\x04\x00\x87\x94\x03\x00\x82\x81\x27\x00\x83\x8D"
-     "\x2A\x00"
-     "\x84\x95\x2A\x00\x85\x61\x2B\x00\x86\x99\x2E\x00\x87\x99\x2E\x00\x84\xA8"
-     "\x04\x00"
-     "\x1C\xAA\x00\x80\x8A\x39\x2F\x00\x8B\xAD\x12\x00\x8C\xBD\x12\x00\x8D\xE9"
-     "\x12\x00"
-     "\x8E\xF9\x14\x00\x8F\x1D\x16\x00\x20\xAA\x00\x80\x24\xAA\x00\x80\x92\x55"
-     "\x17\x00"
-     "\x93\x99\x18\x00\x94\x71\x1A\x00\x95\xF9\x1A\x00\x96\xF1\x1C\x00\x97\x7D"
-     "\x1E\x00"
-     "\x28\xAA\x00\x80\x2C\xAA\x00\x80\x9A\x55\x1E\x00\x9B\xD1\x00\x00\x9C\x75"
-     "\x02\x00"
-     "\x30\xAA\x00\x80\x34\xAA\x00\x80\x38\xAA\x00\x80\xAF\x85\xF6\x00\xAE\x01"
-     "\xF4\x00"
-     "\xAD\xF1\xF2\x00\xAC\x89\xF2\x00\xAB\x01\xF0\x00\xAA\x85\x0E\x00\xA9\x9D"
-     "\x0E\x00"
-     "\xA8\x9D\x0E\x00\xA7\x75\x0A\x00\xA6\x85\x0A\x00\xA5\x01\x08\x00\xA4\x01"
-     "\x06\x00"
-     "\xA3\xB9\x06\x00\xA2\xB1\x06\x00\xA1\xD1\x02\x00\xC4\xA8\x00\x80\x80\x45"
-     "\x03\x00"
-     "\x81\x4D\x03\x00\x82\x45\x03\x00\x3C\xAA\x00\x80\x40\xAA\x00\x80\x44\xAA"
-     "\x00\x80"
-     "\x48\xAA\x00\x80\x4C\xAA\x00\x80\x50\xAA\x00\x80\x54\xAA\x00\x80\xB5\x01"
-     "\xFC\x00"
-     "\xB4\xC1\xFB\x00\xB3\xF5\xFA\x00\xB2\xF5\xFA\x00\xB1\xBD\xF7\x00\xB0\xF5"
-     "\xF6\x00"
-     "\xA8\x6D\x03\x00\xA9\x7D\x03\x00\xAA\x75\x03\x00\xAB\xAD\x03\x00\xAC\xB5"
-     "\x03\x00"
-     "\xAD\xBD\x03\x00\xAE\xB5\x03\x00\xAF\xAD\x03\x00\x58\xAA\x00\x80\x5C\xAA"
-     "\x00\x80"
-     "\x60\xAA\x00\x80\x64\xAA\x00\x80\x68\xAA\x00\x80\x6C\xAA\x00\x80\x70\xAA"
-     "\x00\x80"
-     "\x74\xAA\x00\x80\xB8\x75\x02\x00\xB9\x01\x02\x00\xBA\x01\x02\x00\xBB\x01"
-     "\x02\x00"
-     "\xBC\x01\x02\x00\xBD\x09\x02\x00\xBE\x31\x02\x00\xBF\x31\x02\x00\xB0\xD5"
-     "\x03\x00"
-     "\xB1\xDD\x03\x00\xB2\xD5\x03\x00\xB3\x65\x02\x00\xB4\x61\x02\x00\xB5\x61"
-     "\x02\x00"
-     "\xB6\x61\x02\x00\xB7\x61\x02\x00\xE1\x10\x1E\x00\x78\xAA\x00\x80\xE3\xF8"
-     "\x01\x00"
-     "\xBE\x00\x04\x00\x80\x15\x00\x00\x81\x09\x00\x00\x82\x39\x00\x00\x7C\xAA"
-     "\x00\x80"
-     "\x80\xAA\x00\x80\x88\xAA\x00\x80\x86\x88\x04\x00\x87\x0C\x05\x00\x8C\xAA"
-     "\x00\x80"
-     "\x90\xAA\x00\x80\xEF\x24\x01\x00\x94\xAA\x00\x80\x98\xAA\x00\x80\x9C\xAA"
-     "\x00\x80"
-     "\xA0\xAA\x00\x80\xA4\xAA\x00\x80\xA8\xAA\x00\x80\xEF\xB0\x00\x00\xAC\xAA"
-     "\x00\x80"
-     "\xB0\xAA\x00\x80\xB4\xAA\x00\x80\xE3\x5C\x00\x00\xB8\xAA\x00\x80\xE1\x84"
-     "\x01\x00"
-     "\xBC\xAA\x00\x80\xC0\xAA\x00\x80\xC4\xAA\x00\x80\xC8\xAA\x00\x80\xA3\x7D"
-     "\x04\x00"
-     "\x14\xAA\x00\x80\x84\xAA\x00\x80\xCC\xAA\x00\x80\xD0\xAA\x00\x80\xA6\xBD"
-     "\x04\x00"
-     "\xA5\xBD\x04\x00\xD4\xAA\x00\x80\xAB\x99\x04\x00\xAA\x91\x04\x00\xD8\xAA"
-     "\x00\x80"
-     "\xDC\xAA\x00\x80\xAF\x25\x05\x00\xAE\x25\x05\x00\xAD\x31\x05\x00\xAC\x81"
-     "\x04\x00"
-     "\xE0\xAA\x00\x80\xB3\x05\x06\x00\xE4\xAA\x00\x80\xE8\xAA\x00\x80\xB6\x35"
-     "\x06\x00"
-     "\xEC\xAA\x00\x80\xF0\xAA\x00\x80\xB5\x3D\x06\x00\xBA\x1D\x06\x00\xBB\xA9"
-     "\x06\x00"
-     "\x84\xE0\x03\x00\xBE\xE0\x00\x00\xBE\xA9\x06\x00\xBF\xA9\x06\x00\xBC\xA9"
-     "\x06\x00"
-     "\xBD\xA9\x06\x00\xA8\x2D\x05\x00\xA9\xA1\x06\x00\xAA\xC9\x06\x00\xAB\xC9"
-     "\x06\x00"
-     "\xAC\xF1\x06\x00\xAD\xF1\x06\x00\xAE\xED\x06\x00\xAF\xE5\x06\x00\x80\xED"
-     "\x00\x00"
-     "\x81\xF1\x00\x00\x82\xF1\x00\x00\xF4\xAA\x00\x80\x86\x90\x00\x00\x87\xF0"
-     "\x00\x00"
-     "\xF8\xAA\x00\x80\xFC\xAA\x00\x80\xB8\xE1\x07\x00\xB9\xE1\x07\x00\xBA\xE1"
-     "\x07\x00"
-     "\xBB\xE1\x07\x00\xBC\xE1\x07\x00\xBD\xE1\x07\x00\xBE\xE1\x07\x00\xBF\xE1"
-     "\x07\x00"
-     "\xB0\x9D\x06\x00\xB1\x29\x06\x00\xB2\x39\x06\x00\xB3\x39\x06\x00\xB4\x29"
-     "\x06\x00"
-     "\xB5\x29\x06\x00\xB6\x19\x06\x00\xB7\x19\x06\x00\xA3\x41\x07\x00\x00\xAB"
-     "\x00\x80"
-     "\x04\xAB\x00\x80\x08\xAB\x00\x80\x0C\xAB\x00\x80\xA6\x71\x07\x00\xA5\x79"
-     "\x07\x00"
-     "\x10\xAB\x00\x80\xAB\xED\x07\x00\xAA\x59\x07\x00\x14\xAB\x00\x80\x18\xAB"
-     "\x00\x80"
-     "\xAF\xED\x07\x00\xAE\xED\x07\x00\xAD\xED\x07\x00\xAC\xED\x07\x00\x1C\xAB"
-     "\x00\x80"
-     "\xB3\xF1\x06\x00\x20\xAB\x00\x80\x24\xAB\x00\x80\xB6\x95\x06\x00\x28\xAB"
-     "\x00\x80"
-     "\x2C\xAB\x00\x80\xB5\x9D\x06\x00\xBA\xB1\x06\x00\xBB\xBD\x06\x00\x30\xAB"
-     "\x00\x80"
-     "\x34\xAB\x00\x80\xBE\x99\x06\x00\xBF\x85\x06\x00\xBC\xA5\x06\x00\xBD\x9D"
-     "\x06\x00"
-     "\xA8\x25\x06\x00\xA9\x35\x06\x00\xAA\x3D\x06\x00\xAB\x31\x06\x00\xAC\x51"
-     "\x06\x00"
-     "\xAD\x51\x06\x00\xAE\x51\x06\x00\xAF\x51\x06\x00\x38\xAB\x00\x80\x3C\xAB"
-     "\x00\x80"
-     "\x84\x2C\x01\x00\x40\xAB\x00\x80\x44\xAB\x00\x80\x48\xAB\x00\x80\x4C\xAB"
-     "\x00\x80"
-     "\x50\xAB\x00\x80\xB8\x3D\x06\x00\xB9\xC9\x01\x00\xBA\xD9\x01\x00\xBB\xD1"
-     "\x01\x00"
-     "\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\x99\x01\x00\xBF\x95\x01\x00\xB0\x31"
-     "\x06\x00"
-     "\xB1\x3D\x06\x00\xB2\x35\x06\x00\xB3\x0D\x06\x00\xB4\x19\x06\x00\xB5\x05"
-     "\x06\x00"
-     "\xB6\x0D\x06\x00\xB7\x05\x06\x00\xA3\xB5\x05\x00\x54\xAB\x00\x80\x58\xAB"
-     "\x00\x80"
-     "\x5C\xAB\x00\x80\x60\xAB\x00\x80\xA6\xD1\x05\x00\xA5\xD9\x05\x00\xBE\x58"
-     "\x03\x00"
-     "\xAB\xF9\x05\x00\xAA\xF5\x05\x00\x86\x88\x0C\x00\x87\xBC\x03\x00\xAF\xC1"
-     "\x05\x00"
-     "\xAE\xDD\x05\x00\xAD\xD9\x05\x00\xAC\xE1\x05\x00\x81\x4D\x00\x00\x80\x4D"
-     "\x00\x00"
-     "\xB3\x4D\x03\x00\x82\x5D\x00\x00\xB5\x5D\x03\x00\x68\xAB\x00\x80\x6C\xAB"
-     "\x00\x80"
-     "\xB6\x4D\x03\x00\x70\xAB\x00\x80\x74\xAB\x00\x80\xBB\x29\x03\x00\xBA\x41"
-     "\x03\x00"
-     "\xBD\x3D\x03\x00\xBC\x3D\x03\x00\xBF\x21\x03\x00\xBE\x29\x03\x00\xA8\x39"
-     "\x02\x00"
-     "\xA9\x39\x02\x00\xAA\x8D\x03\x00\xAB\x85\x03\x00\xAC\x9D\x03\x00\xAD\x85"
-     "\x03\x00"
-     "\xAE\x85\x03\x00\xAF\xB5\x03\x00\x84\xEC\x0C\x00\x78\xAB\x00\x80\xBE\xE0"
-     "\x0C\x00"
-     "\x7C\xAB\x00\x80\x80\xAB\x00\x80\x84\xAB\x00\x80\x88\xAB\x00\x80\x8C\xAB"
-     "\x00\x80"
-     "\xB8\x61\x03\x00\xB9\x61\x03\x00\xBA\x65\x03\x00\xBB\x7D\x03\x00\xBC\x65"
-     "\x03\x00"
-     "\xBD\x6D\x03\x00\xBE\x65\x03\x00\xBF\xD9\x00\x00\xB0\xCD\x03\x00\xB1\xA5"
-     "\x03\x00"
-     "\xB2\xA1\x03\x00\xB3\xA1\x03\x00\xB4\xA5\x03\x00\xB5\xAD\x03\x00\xB6\x91"
-     "\x03\x00"
-     "\xB7\x91\x03\x00\x90\xAB\x00\x80\x94\xAB\x00\x80\x98\xAB\x00\x80\x9C\xAB"
-     "\x00\x80"
-     "\xEF\xF4\x03\x00\xA0\xAB\x00\x80\xA4\xAB\x00\x80\xA8\xAB\x00\x80\xE3\x98"
-     "\x00\x00"
-     "\xAC\xAB\x00\x80\xE1\x80\x01\x00\xB0\xAB\x00\x80\xB4\xAB\x00\x80\xB8\xAB"
-     "\x00\x80"
-     "\xBC\xAB\x00\x80\xC0\xAB\x00\x80\xBA\xF5\x07\x00\xBB\xFD\x07\x00\xC4\xAB"
-     "\x00\x80"
-     "\xC8\xAB\x00\x80\xBE\x41\x06\x00\xBF\x41\x06\x00\xBC\xE5\x07\x00\xBD\x55"
-     "\x06\x00"
-     "\xCC\xAB\x00\x80\xB3\x19\x07\x00\xD0\xAB\x00\x80\xD8\xAB\x00\x80\xB6\xD9"
-     "\x07\x00"
-     "\xDC\xAB\x00\x80\xE0\xAB\x00\x80\xB5\xD9\x07\x00\xE1\x98\x07\x00\xE3\xFC"
-     "\x01\x00"
-     "\xE3\xB8\x07\x00\xE1\x3C\x06\x00\x80\xDD\x00\x00\x81\xF5\x00\x00\x82\xF1"
-     "\x00\x00"
-     "\xE4\xAB\x00\x80\xE8\xAB\x00\x80\xEC\xAB\x00\x80\xF0\xAB\x00\x80\xF4\xAB"
-     "\x00\x80"
-     "\xF8\xAB\x00\x80\xEF\x20\x01\x00\xEF\xCC\x06\x00\xFC\xAB\x00\x80\x86\xE0"
-     "\x0C\x00"
-     "\x87\x0C\x0D\x00\xA3\xCD\x02\x00\x00\xAC\x00\x80\xA5\xDD\x02\x00\x04\xAC"
-     "\x00\x80"
-     "\x08\xAC\x00\x80\xA6\xCD\x02\x00\x0C\xAC\x00\x80\x10\xAC\x00\x80\xAB\xA9"
-     "\x02\x00"
-     "\xAA\xC1\x02\x00\xAD\xBD\x02\x00\xAC\xBD\x02\x00\xAF\xA1\x02\x00\xAE\xA9"
-     "\x02\x00"
-     "\xA8\x55\x0E\x00\xA9\x5D\x0E\x00\xAA\x69\x0E\x00\xAB\x69\x0E\x00\xAC\xB5"
-     "\x0E\x00"
-     "\xAD\xB9\x0E\x00\xAE\xA9\x0E\x00\xAF\xA9\x0E\x00\xD4\xAB\x00\x80\x14\xAC"
-     "\x00\x80"
-     "\x18\xAC\x00\x80\x1C\xAC\x00\x80\x20\xAC\x00\x80\x24\xAC\x00\x80\x28\xAC"
-     "\x00\x80"
-     "\x2C\xAC\x00\x80\xB8\x5D\x0F\x00\xB9\x61\x0F\x00\xBA\x65\x0F\x00\xBB\x7D"
-     "\x0F\x00"
-     "\xBC\x65\x0F\x00\xBD\x6D\x0F\x00\xBE\x65\x0F\x00\xBF\x15\x0F\x00\xB0\xD9"
-     "\x0E\x00"
-     "\xB1\xD9\x0E\x00\xB2\x69\x0F\x00\xB3\x69\x0F\x00\xB4\x79\x0F\x00\xB5\x79"
-     "\x0F\x00"
-     "\xB6\x6D\x0F\x00\xB7\x65\x0F\x00\xB3\xE1\x0E\x00\x30\xAC\x00\x80\x34\xAC"
-     "\x00\x80"
-     "\x38\xAC\x00\x80\x3C\xAC\x00\x80\xB6\x39\x0E\x00\xB5\x3D\x0E\x00\x40\xAC"
-     "\x00\x80"
-     "\xBB\x1D\x0E\x00\xBA\x1D\x0E\x00\x84\x24\x03\x00\xBE\x2C\x00\x00\xBF\xBD"
-     "\x0F\x00"
-     "\xBE\xBD\x0F\x00\xBD\xBD\x0F\x00\xBC\xBD\x0F\x00\x82\x5D\x00\x00\xA3\xA5"
-     "\x0E\x00"
-     "\x80\x45\x00\x00\x81\x5D\x00\x00\xA6\x7D\x0E\x00\x44\xAC\x00\x80\x48\xAC"
-     "\x00\x80"
-     "\xA5\x79\x0E\x00\xAA\x59\x0E\x00\xAB\x59\x0E\x00\x86\x00\x04\x00\x87\x5C"
-     "\x03\x00"
-     "\xAE\xF9\x0F\x00\xAF\xF9\x0F\x00\xAC\xF9\x0F\x00\xAD\xF9\x0F\x00\x4C\xAC"
-     "\x00\x80"
-     "\x50\xAC\x00\x80\x54\xAC\x00\x80\x58\xAC\x00\x80\x5C\xAC\x00\x80\xEF\x44"
-     "\x0D\x00"
-     "\x60\xAC\x00\x80\x64\xAC\x00\x80\x68\xAC\x00\x80\xE3\x94\x0D\x00\x6C\xAC"
-     "\x00\x80"
-     "\xE1\x1C\x0D\x00\x70\xAC\x00\x80\x74\xAC\x00\x80\x78\xAC\x00\x80\x7C\xAC"
-     "\x00\x80"
-     "\xB3\x39\x0E\x00\x80\xAC\x00\x80\x84\xAC\x00\x80\x88\xAC\x00\x80\x8C\xAC"
-     "\x00\x80"
-     "\xB6\xF1\x0E\x00\xB5\xF9\x0E\x00\x90\xAC\x00\x80\xBB\xDD\x0E\x00\xBA\xD5"
-     "\x0E\x00"
-     "\x94\xAC\x00\x80\x98\xAC\x00\x80\xBF\x91\x0E\x00\xBE\xAD\x0E\x00\xBD\xAD"
-     "\x0E\x00"
-     "\xBC\xB1\x0E\x00\xA8\xED\x0E\x00\xA9\xF1\x0E\x00\xAA\xF1\x0E\x00\xAB\x11"
-     "\x0E\x00"
-     "\xAC\x31\x0E\x00\xAD\x31\x0E\x00\xAE\x31\x0E\x00\xAF\x31\x0E\x00\x9C\xAC"
-     "\x00\x80"
-     "\xA0\xAC\x00\x80\xA4\xAC\x00\x80\xA8\xAC\x00\x80\x80\x1D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\xB9\x03\x00\xAC\xAC\x00\x80\xB8\xE9\x01\x00\xB9\xE9\x01\x00\xBA\xF9"
-     "\x01\x00"
-     "\xBB\xF9\x01\x00\xBC\xE9\x01\x00\xBD\xE9\x01\x00\xBE\x5D\x01\x00\xBF\x55"
-     "\x01\x00"
-     "\xB0\x51\x0E\x00\xB1\x59\x0E\x00\xB2\x25\x0E\x00\xB3\x39\x0E\x00\xB4\x29"
-     "\x0E\x00"
-     "\xB5\x15\x0E\x00\xB6\x1D\x0E\x00\xB7\xD9\x01\x00\xB0\xAC\x00\x80\x84\x9C"
-     "\x02\x00"
-     "\xBE\x04\x02\x00\xA3\x75\x0D\x00\xB8\xAC\x00\x80\xA5\xB5\x0D\x00\xA6\xBD"
-     "\x0D\x00"
-     "\xBC\xAC\x00\x80\xC0\xAC\x00\x80\xC4\xAC\x00\x80\xAA\x99\x0D\x00\xAB\x91"
-     "\x0D\x00"
-     "\xAC\xFD\x0D\x00\xAD\xE1\x0D\x00\xAE\xE1\x0D\x00\xAF\xDD\x0D\x00\xB6\x51"
-     "\x03\x00"
-     "\xC8\xAC\x00\x80\x84\x80\x03\x00\xB5\x41\x03\x00\xCC\xAC\x00\x80\xB3\x51"
-     "\x03\x00"
-     "\x86\xD8\x04\x00\x87\xA8\x02\x00\xBE\x59\x03\x00\xBF\x59\x03\x00\xBC\x59"
-     "\x03\x00"
-     "\xBD\x59\x03\x00\xBA\x51\x03\x00\xBB\x5D\x03\x00\xD0\xAC\x00\x80\xD4\xAC"
-     "\x00\x80"
-     "\xA3\x85\x03\x00\xD8\xAC\x00\x80\xDC\xAC\x00\x80\xE0\xAC\x00\x80\xE4\xAC"
-     "\x00\x80"
-     "\xA6\x85\x03\x00\xA5\x95\x03\x00\xE8\xAC\x00\x80\xAB\x89\x03\x00\xAA\x85"
-     "\x03\x00"
-     "\x64\xAB\x00\x80\xEC\xAC\x00\x80\xAF\x8D\x03\x00\xAE\x8D\x03\x00\xAD\x8D"
-     "\x03\x00"
-     "\xAC\x8D\x03\x00\xF0\xAC\x00\x80\xF4\xAC\x00\x80\x85\xB4\xF5\x00\xF8\xAC"
-     "\x00\x80"
-     "\xFC\xAC\x00\x80\x00\xAD\x00\x80\x04\xAD\x00\x80\x08\xAD\x00\x80\x0C\xAD"
-     "\x00\x80"
-     "\x10\xAD\x00\x80\x14\xAD\x00\x80\x18\xAD\x00\x80\x81\x09\x00\x00\x80\x19"
-     "\x00\x00"
-     "\x1C\xAD\x00\x80\x82\xFD\x01\x00\x20\xAD\x00\x80\xB3\x25\x01\x00\xBE\x48"
-     "\x02\x00"
-     "\x28\xAD\x00\x80\xB6\x41\x01\x00\x2C\xAD\x00\x80\x30\xAD\x00\x80\xB5\x59"
-     "\x01\x00"
-     "\xBA\xA5\x01\x00\xBB\xAD\x01\x00\x34\xAD\x00\x80\x38\xAD\x00\x80\xBE\xAD"
-     "\x01\x00"
-     "\xBF\x99\x01\x00\xBC\xB5\x01\x00\xBD\xAD\x01\x00\x86\x60\x05\x00\x87\xF4"
-     "\x02\x00"
-     "\x3C\xAD\x00\x80\x40\xAD\x00\x80\xE1\xA0\x01\x00\x44\xAD\x00\x80\xE3\x4C"
-     "\x03\x00"
-     "\x48\xAD\x00\x80\x4C\xAD\x00\x80\x50\xAD\x00\x80\xEF\x98\x03\x00\x54\xAD"
-     "\x00\x80"
-     "\x58\xAD\x00\x80\x5C\xAD\x00\x80\x60\xAD\x00\x80\x64\xAD\x00\x80\x68\xAD"
-     "\x00\x80"
-     "\x6C\xAD\x00\x80\x70\xAD\x00\x80\x84\x40\x04\x00\x74\xAD\x00\x80\x78\xAD"
-     "\x00\x80"
-     "\x7C\xAD\x00\x80\xEF\xC8\x01\x00\x80\xAD\x00\x80\xE1\x7C\x0E\x00\x84\xAD"
-     "\x00\x80"
-     "\xE3\x14\x01\x00\x80\x0D\x00\x00\x81\x75\x00\x00\x82\x7D\x00\x00\x88\xAD"
-     "\x00\x80"
-     "\x8C\xAD\x00\x80\x90\xAD\x00\x80\x98\xAD\x00\x80\xA3\xED\x01\x00\x9C\xAD"
-     "\x00\x80"
-     "\xA5\x91\x01\x00\xA6\x89\x01\x00\xBE\x14\x04\x00\x86\xC0\x04\x00\x87\x44"
-     "\x04\x00"
-     "\xAA\x6D\x01\x00\xAB\x65\x01\x00\xAC\x7D\x01\x00\xAD\x65\x01\x00\xAE\x65"
-     "\x01\x00"
-     "\xAF\x51\x01\x00\xB4\xAC\x00\x80\x24\xAD\x00\x80\xA0\xAD\x00\x80\xA4\xAD"
-     "\x00\x80"
-     "\x94\xAD\x00\x80\xA8\xAD\x00\x80\xAC\xAD\x00\x80\xB0\xAD\x00\x80\xA8\x31"
-     "\x06\x00"
-     "\xA9\x39\x06\x00\xAA\x09\x06\x00\xAB\x09\x06\x00\xAC\x75\x06\x00\xAD\x65"
-     "\x06\x00"
-     "\xAE\x41\x06\x00\xAF\x41\x06\x00\xB0\xC5\x06\x00\xB1\xCD\x06\x00\xB2\xC5"
-     "\x06\x00"
-     "\xB3\xDD\x06\x00\xB4\xCD\x06\x00\xB5\xF9\x06\x00\xB6\x99\x06\x00\xB7\x99"
-     "\x06\x00"
-     "\xB8\xAD\x06\x00\xB9\xB5\x06\x00\xBA\xBD\x06\x00\xBB\xB5\x06\x00\xBC\xAD"
-     "\x06\x00"
-     "\xBD\x95\x06\x00\xBE\x9D\x06\x00\xBF\x95\x06\x00\xA8\x8D\x07\x00\xA9\x91"
-     "\x07\x00"
-     "\xAA\x91\x07\x00\xAB\xA5\x07\x00\xAC\xBD\x07\x00\xAD\xDD\x07\x00\xAE\xD5"
-     "\x07\x00"
-     "\xAF\xC9\x07\x00\xB4\xAD\x00\x80\xB8\xAD\x00\x80\xBC\xAD\x00\x80\xC0\xAD"
-     "\x00\x80"
-     "\xC4\xAD\x00\x80\xC8\xAD\x00\x80\xCC\xAD\x00\x80\xD0\xAD\x00\x80\xB8\x41"
-     "\x07\x00"
-     "\xB9\x41\x07\x00\xBA\x41\x07\x00\xBB\x5D\x07\x00\xBC\x45\x07\x00\xBD\x4D"
-     "\x07\x00"
-     "\xBE\x79\x07\x00\xBF\x75\x07\x00\xB0\xBD\x07\x00\xB1\x85\x07\x00\xB2\x8D"
-     "\x07\x00"
-     "\xB3\x81\x07\x00\xB4\x81\x07\x00\xB5\x81\x07\x00\xB6\x81\x07\x00\xB7\x81"
-     "\x07\x00"
-     "\xB3\x19\x06\x00\xD4\xAD\x00\x80\xD8\xAD\x00\x80\xDC\xAD\x00\x80\xE0\xAD"
-     "\x00\x80"
-     "\xB6\x5D\x06\x00\xB5\x35\x06\x00\xE4\xAD\x00\x80\xBB\x41\x06\x00\xBA\x79"
-     "\x06\x00"
-     "\xE8\xAD\x00\x80\xEC\xAD\x00\x80\xBF\x45\x06\x00\xBE\x45\x06\x00\xBD\x55"
-     "\x06\x00"
-     "\xBC\x59\x06\x00\xF0\xAD\x00\x80\xA3\x5D\x06\x00\xF4\xAD\x00\x80\xF8\xAD"
-     "\x00\x80"
-     "\xA6\x19\x06\x00\xFC\xAD\x00\x80\x00\xAE\x00\x80\xA5\x71\x06\x00\xAA\x3D"
-     "\x06\x00"
-     "\xAB\x05\x06\x00\x84\x24\x01\x00\xBE\x24\x01\x00\xAE\x01\x06\x00\xAF\x01"
-     "\x06\x00"
-     "\xAC\x1D\x06\x00\xAD\x11\x06\x00\xA8\x49\x06\x00\xA9\x55\x06\x00\xAA\x81"
-     "\x01\x00"
-     "\xAB\x81\x01\x00\xAC\x85\x01\x00\xAD\x8D\x01\x00\xAE\x85\x01\x00\xAF\x3D"
-     "\x01\x00"
-     "\x80\xED\x00\x00\x81\xF1\x00\x00\x82\xF1\x00\x00\x08\xAE\x00\x80\x86\x10"
-     "\x00\x00"
-     "\x87\x94\x00\x00\x0C\xAE\x00\x80\x10\xAE\x00\x80\xB8\xC9\x00\x00\xB9\xC9"
-     "\x00\x00"
-     "\xBA\xD9\x00\x00\xBB\xD1\x00\x00\xBC\xF9\x00\x00\xBD\xF9\x00\x00\xBE\x99"
-     "\x00\x00"
-     "\xBF\x95\x00\x00\xB0\x45\x01\x00\xB1\x4D\x01\x00\xB2\x45\x01\x00\xB3\x5D"
-     "\x01\x00"
-     "\xB4\x45\x01\x00\xB5\x4D\x01\x00\xB6\x45\x01\x00\xB7\xF9\x00\x00\xBA\x5D"
-     "\x07\x00"
-     "\xBB\x6D\x07\x00\xB8\x45\x07\x00\xB9\x5D\x07\x00\xBE\xB5\x07\x00\xBF\xBD"
-     "\x07\x00"
-     "\xBC\x75\x07\x00\xBD\xBD\x07\x00\xB2\x11\x07\x00\xB3\x11\x07\x00\xB0\x09"
-     "\x07\x00"
-     "\xB1\x05\x07\x00\xB6\x75\x07\x00\xB7\x7D\x07\x00\xB4\x01\x07\x00\xB5\x01"
-     "\x07\x00"
-     "\xAA\x45\x07\x00\xAB\x4D\x07\x00\xA8\x91\x06\x00\xA9\x4D\x07\x00\xAE\x49"
-     "\x07\x00"
-     "\xAF\x49\x07\x00\xAC\x55\x07\x00\xAD\x4D\x07\x00\x14\xAE\x00\x80\x18\xAE"
-     "\x00\x80"
-     "\x1C\xAE\x00\x80\x20\xAE\x00\x80\x24\xAE\x00\x80\x28\xAE\x00\x80\x2C\xAE"
-     "\x00\x80"
-     "\x30\xAE\x00\x80\xB3\x89\x02\x00\x34\xAE\x00\x80\x38\xAE\x00\x80\x3C\xAE"
-     "\x00\x80"
-     "\x40\xAE\x00\x80\xB6\x8D\x02\x00\xB5\x8D\x02\x00\x44\xAE\x00\x80\xBB\x09"
-     "\x02\x00"
-     "\xBA\x01\x02\x00\x48\xAE\x00\x80\x4C\xAE\x00\x80\xBF\x09\x02\x00\xBE\x01"
-     "\x02\x00"
-     "\xBD\x09\x02\x00\xBC\x11\x02\x00\x50\xAE\x00\x80\xA3\xCD\x02\x00\x54\xAE"
-     "\x00\x80"
-     "\x58\xAE\x00\x80\xA6\xC9\x02\x00\x5C\xAE\x00\x80\x60\xAE\x00\x80\xA5\xC9"
-     "\x02\x00"
-     "\xAA\x45\x02\x00\xAB\x4D\x02\x00\x64\xAE\x00\x80\x68\xAE\x00\x80\xAE\x45"
-     "\x02\x00"
-     "\xAF\x4D\x02\x00\xAC\x55\x02\x00\xAD\x4D\x02\x00\xA8\xAD\x02\x00\xA9\xBD"
-     "\x02\x00"
-     "\xAA\xB5\x02\x00\xAB\x6D\x01\x00\xAC\x75\x01\x00\xAD\x7D\x01\x00\xAE\x75"
-     "\x01\x00"
-     "\xAF\x6D\x01\x00\x80\xCD\x01\x00\x81\x09\x00\x00\x82\x19\x00\x00\x6C\xAE"
-     "\x00\x80"
-     "\x70\xAE\x00\x80\x84\x34\x02\x00\xBE\xB4\xFD\x00\x78\xAE\x00\x80\xB8\xE9"
-     "\x01\x00"
-     "\xB9\xE9\x01\x00\xBA\x89\x01\x00\xBB\x85\x01\x00\xBC\x9D\x01\x00\xBD\x81"
-     "\x01\x00"
-     "\xBE\x81\x01\x00\xBF\xB5\x01\x00\xB0\x15\x01\x00\xB1\x1D\x01\x00\xB2\x15"
-     "\x01\x00"
-     "\xB3\xE9\x01\x00\xB4\xF9\x01\x00\xB5\xF9\x01\x00\xB6\xE9\x01\x00\xB7\xE1"
-     "\x01\x00"
-     "\xE1\xD8\x06\x00\xE1\xC4\x07\x00\xE3\x14\x06\x00\xE3\xD4\x07\x00\x7C\xAE"
-     "\x00\x80"
-     "\xE1\x6C\x06\x00\x80\xAE\x00\x80\xE3\xE4\x00\x00\x84\x2C\xFC\x00\x84\xAE"
-     "\x00\x80"
-     "\xBE\xA4\xFC\x00\xEF\x38\x00\x00\x86\x90\xFD\x00\x87\xD4\xFD\x00\xEF\x9C"
-     "\x06\x00"
-     "\xEF\x4C\x06\x00\x88\xAE\x00\x80\x8C\xAE\x00\x80\x84\xC8\xFC\x00\xB3\x71"
-     "\x02\x00"
-     "\x90\xAE\x00\x80\xB5\xB1\x01\x00\xB6\xB1\x01\x00\x94\xAE\x00\x80\x98\xAE"
-     "\x00\x80"
-     "\x9C\xAE\x00\x80\xBA\x9D\x01\x00\xBB\x95\x01\x00\xBC\x8D\x01\x00\xBD\x71"
-     "\x01\x00"
-     "\xBE\x71\x01\x00\xBF\x71\x01\x00\xA3\xBD\xFD\x00\x74\xAE\x00\x80\xA0\xAE"
-     "\x00\x80"
-     "\xA4\xAE\x00\x80\xA8\xAE\x00\x80\xA6\x7D\xFE\x00\xA5\x7D\xFE\x00\xAC\xAE"
-     "\x00\x80"
-     "\xAB\x59\xFE\x00\xAA\x51\xFE\x00\xB0\xAE\x00\x80\xB4\xAE\x00\x80\xAF\xBD"
-     "\xFE\x00"
-     "\xAE\xBD\xFE\x00\xAD\xBD\xFE\x00\xAC\x41\xFE\x00\xB8\xAE\x00\x80\xBC\xAE"
-     "\x00\x80"
-     "\xC0\xAE\x00\x80\xC4\xAE\x00\x80\xC8\xAE\x00\x80\xCC\xAE\x00\x80\xD0\xAE"
-     "\x00\x80"
-     "\xD4\xAE\x00\x80\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\xD8\xAE"
-     "\x00\x80"
-     "\xDC\xAE\x00\x80\xE0\xAE\x00\x80\xE4\xAE\x00\x80\xE8\xAE\x00\x80\xA8\xBD"
-     "\xFF\x00"
-     "\xA9\x15\xFF\x00\xAA\x11\xFF\x00\xAB\x29\xFF\x00\xAC\x51\xFF\x00\xAD\x51"
-     "\xFF\x00"
-     "\xAE\x49\xFF\x00\xAF\x41\xFF\x00\x86\x00\x00\x00\x87\x64\x03\x00\xEC\xAE"
-     "\x00\x80"
-     "\xF0\xAE\x00\x80\xF4\xAE\x00\x80\xF8\xAE\x00\x80\xFC\xAE\x00\x80\x00\xAF"
-     "\x00\x80"
-     "\xB8\x81\x00\x00\xB9\x81\x00\x00\xBA\x81\x00\x00\xBB\x81\x00\x00\xBC\x81"
-     "\x00\x00"
-     "\xBD\x81\x00\x00\xBE\x81\x00\x00\xBF\x81\x00\x00\xB0\x01\xFF\x00\xB1\x01"
-     "\xFF\x00"
-     "\xB2\xE1\x00\x00\xB3\xE1\x00\x00\xB4\xE1\x00\x00\xB5\xE9\x00\x00\xB6\xD1"
-     "\x00\x00"
-     "\xB7\xD1\x00\x00\x04\xAF\x00\x80\x08\xAF\x00\x80\x0C\xAF\x00\x80\x10\xAF"
-     "\x00\x80"
-     "\x14\xAF\x00\x80\xE1\xBC\xFE\x00\x18\xAF\x00\x80\xE3\x38\xFC\x00\x1C\xAF"
-     "\x00\x80"
-     "\x20\xAF\x00\x80\x24\xAF\x00\x80\xEF\xEC\xFC\x00\x28\xAF\x00\x80\x2C\xAF"
-     "\x00\x80"
-     "\x30\xAF\x00\x80\x34\xAF\x00\x80\xB3\xD9\xFE\x00\x38\xAF\x00\x80\x3C\xAF"
-     "\x00\x80"
-     "\x40\xAF\x00\x80\x44\xAF\x00\x80\xB6\x91\xFE\x00\xB5\x99\xFE\x00\x48\xAF"
-     "\x00\x80"
-     "\xBB\x85\xFE\x00\xBA\x85\xFE\x00\x4C\xAF\x00\x80\x50\xAF\x00\x80\xBF\x6D"
-     "\x01\x00"
-     "\xBE\x65\x01\x00\xBD\x75\x01\x00\xBC\x75\x01\x00\xA3\x19\xFE\x00\xBE\x84"
-     "\x01\x00"
-     "\x86\x28\x00\x00\x87\x5C\x01\x00\x5F\xAF\x00\x80\xA6\x51\xFE\x00\xA5\x59"
-     "\xFE\x00"
-     "\x63\xAF\x00\x80\xAB\x45\xFE\x00\xAA\x45\xFE\x00\x67\xAF\x00\x80\x6B\xAF"
-     "\x00\x80"
-     "\xAF\xAD\x01\x00\xAE\xA5\x01\x00\xAD\xB5\x01\x00\xAC\xB5\x01\x00\x80\x9D"
-     "\x00\x00"
-     "\x81\xAD\x00\x00\x82\xA5\x00\x00\xB3\x61\x01\x00\x6F\xAF\x00\x80\xB5\x61"
-     "\x01\x00"
-     "\xB6\x61\x01\x00\x73\xAF\x00\x80\x77\xAF\x00\x80\x7B\xAF\x00\x80\xBA\xED"
-     "\x00\x00"
-     "\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\xE5\x00\x00\xBE\xE1\x00\x00\xBF\xDD"
-     "\x00\x00"
-     "\xA8\x7D\x02\x00\xA9\xAD\x02\x00\xAA\xA5\x02\x00\xAB\xBD\x02\x00\xAC\xA5"
-     "\x02\x00"
-     "\xAD\xAD\x02\x00\xAE\xA5\x02\x00\xAF\x1D\x03\x00\x84\xEC\x04\x00\x7F\xAF"
-     "\x00\x80"
-     "\x83\xAF\x00\x80\x87\xAF\x00\x80\x8B\xAF\x00\x80\x8F\xAF\x00\x80\x93\xAF"
-     "\x00\x80"
-     "\x97\xAF\x00\x80\xB8\x71\x03\x00\xB9\x71\x03\x00\xBA\x71\x03\x00\xBB\x71"
-     "\x03\x00"
-     "\xBC\x15\x03\x00\xBD\x1D\x03\x00\xBE\x15\x03\x00\xBF\x0D\x03\x00\xB0\x65"
-     "\x03\x00"
-     "\xB1\x6D\x03\x00\xB2\x65\x03\x00\xB3\x7D\x03\x00\xB4\x69\x03\x00\xB5\x55"
-     "\x03\x00"
-     "\xB6\x5D\x03\x00\xB7\x55\x03\x00\x9B\xAF\x00\x80\x9F\xAF\x00\x80\xA3\xAF"
-     "\x00\x80"
-     "\xA3\x21\x02\x00\x04\xAE\x00\x80\xA5\x21\x02\x00\xA6\x21\x02\x00\xA7\xAF"
-     "\x00\x80"
-     "\xAB\xAF\x00\x80\xAF\xAF\x00\x80\xAA\xAD\x03\x00\xAB\xA5\x03\x00\xAC\xBD"
-     "\x03\x00"
-     "\xAD\xA5\x03\x00\xAE\xA1\x03\x00\xAF\x9D\x03\x00\xB3\xAF\x00\x80\xB7\xAF"
-     "\x00\x80"
-     "\xBB\xAF\x00\x80\xBF\xAF\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39"
-     "\x00\x00"
-     "\xC3\xAF\x00\x80\xC7\xAF\x00\x80\xCF\xAF\x00\x80\xD3\xAF\x00\x80\xBE\x24"
-     "\x04\x00"
-     "\x86\xD0\x04\x00\x87\x20\x03\x00\xD7\xAF\x00\x80\xDB\xAF\x00\x80\xA8\x91"
-     "\x02\x00"
-     "\xA9\x99\x02\x00\xAA\xA1\x02\x00\xAB\xA1\x02\x00\xAC\xC1\x02\x00\xAD\xC9"
-     "\x02\x00"
-     "\xAE\xF9\x02\x00\xAF\xF9\x02\x00\xDF\xAF\x00\x80\xE3\xAF\x00\x80\xE7\xAF"
-     "\x00\x80"
-     "\xEB\xAF\x00\x80\xEF\xAF\x00\x80\xF3\xAF\x00\x80\xF7\xAF\x00\x80\xFB\xAF"
-     "\x00\x80"
-     "\xB8\xC5\x01\x00\xB9\xCD\x01\x00\xBA\xD5\x01\x00\xBB\xED\x01\x00\xBC\xF5"
-     "\x01\x00"
-     "\xBD\x95\x01\x00\xBE\x9D\x01\x00\xBF\x95\x01\x00\xB0\x45\x01\x00\xB1\x4D"
-     "\x01\x00"
-     "\xB2\x45\x01\x00\xB3\x5D\x01\x00\xB4\x45\x01\x00\xB5\x4D\x01\x00\xB6\x45"
-     "\x01\x00"
-     "\xB7\xFD\x01\x00\xE1\x28\xFE\x00\xFF\xAF\x00\x80\xE3\x38\xFE\x00\xBE\x80"
-     "\x04\x00"
-     "\x84\x1C\x04\x00\x03\xB0\x00\x80\x07\xB0\x00\x80\x0B\xB0\x00\x80\x0F\xB0"
-     "\x00\x80"
-     "\x13\xB0\x00\x80\x17\xB0\x00\x80\x1B\xB0\x00\x80\x1F\xB0\x00\x80\x23\xB0"
-     "\x00\x80"
-     "\xEF\xE8\xFE\x00\x27\xB0\x00\x80\xF6\x00\x00\x00\xEF\xB8\x00\x00\x2B\xB0"
-     "\x00\x80"
-     "\x2F\xB0\x00\x80\x81\x09\x00\x00\x80\x19\x00\x00\x33\xB0\x00\x80\x82\x3D"
-     "\x00\x00"
-     "\x37\xB0\x00\x80\x3F\xB0\x00\x80\x86\x48\x07\x00\x87\xFC\x04\x00\x43\xB0"
-     "\x00\x80"
-     "\xE3\x2C\xFE\x00\x47\xB0\x00\x80\xE1\x44\x01\x00\xB3\xA9\x05\x00\xCB\xAF"
-     "\x00\x80"
-     "\x3B\xB0\x00\x80\x4B\xB0\x00\x80\x4F\xB0\x00\x80\xB6\xC9\x05\x00\xB5\xC9"
-     "\x05\x00"
-     "\x53\xB0\x00\x80\xBB\x09\x06\x00\xBA\x01\x06\x00\x57\xB0\x00\x80\x5B\xB0"
-     "\x00\x80"
-     "\xBF\x09\x06\x00\xBE\x01\x06\x00\xBD\x09\x06\x00\xBC\x11\x06\x00\x5F\xB0"
-     "\x00\x80"
-     "\xA3\xED\x05\x00\x63\xB0\x00\x80\x67\xB0\x00\x80\xA6\x8D\x05\x00\x6B\xB0"
-     "\x00\x80"
-     "\x6F\xB0\x00\x80\xA5\x8D\x05\x00\xAA\x45\x06\x00\xAB\x4D\x06\x00\x73\xB0"
-     "\x00\x80"
-     "\x77\xB0\x00\x80\xAE\x45\x06\x00\xAF\x4D\x06\x00\xAC\x55\x06\x00\xAD\x4D"
-     "\x06\x00"
-     "\xA8\x65\x06\x00\xA9\x6D\x06\x00\xAA\x71\x06\x00\xAB\x71\x06\x00\xAC\xB1"
-     "\x07\x00"
-     "\xAD\xB1\x07\x00\xAE\xB1\x07\x00\xAF\xB1\x07\x00\x84\xA8\x03\x00\x7B\xB0"
-     "\x00\x80"
-     "\x7F\xB0\x00\x80\x83\xB0\x00\x80\x87\xB0\x00\x80\x8B\xB0\x00\x80\x8F\xB0"
-     "\x00\x80"
-     "\x93\xB0\x00\x80\xB8\x61\x07\x00\xB9\x61\x07\x00\xBA\x61\x07\x00\xBB\x61"
-     "\x07\x00"
-     "\xBC\x61\x07\x00\xBD\x61\x07\x00\xBE\x65\x07\x00\xBF\x15\x07\x00\xB0\xD1"
-     "\x07\x00"
-     "\xB1\xD1\x07\x00\xB2\xD1\x07\x00\xB3\xD1\x07\x00\xB4\x75\x07\x00\xB5\x61"
-     "\x07\x00"
-     "\xB6\x61\x07\x00\xB7\x61\x07\x00\xE1\x90\x06\x00\x97\xB0\x00\x80\xE3\x38"
-     "\x05\x00"
-     "\x9B\xB0\x00\x80\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\x9F\xB0"
-     "\x00\x80"
-     "\xA3\xB0\x00\x80\xA7\xB0\x00\x80\xBE\x28\x01\x00\xAF\xB0\x00\x80\x86\x90"
-     "\x00\x00"
-     "\x87\x04\x01\x00\xEF\x2C\x05\x00\xB3\xB0\x00\x80\xB3\xB9\x07\x00\xAB\xB0"
-     "\x00\x80"
-     "\xB7\xB0\x00\x80\xBB\xB0\x00\x80\xBF\xB0\x00\x80\xB6\x71\x06\x00\xB5\x79"
-     "\x06\x00"
-     "\xC3\xB0\x00\x80\xBB\x55\x06\x00\xBA\x55\x06\x00\xC7\xB0\x00\x80\xCB\xB0"
-     "\x00\x80"
-     "\xBF\xE5\x06\x00\xBE\xF1\x06\x00\xBD\x45\x06\x00\xBC\x45\x06\x00\xA3\x79"
-     "\x07\x00"
-     "\xCF\xB0\x00\x80\xD3\xB0\x00\x80\xD7\xB0\x00\x80\xDB\xB0\x00\x80\xA6\xB1"
-     "\x06\x00"
-     "\xA5\xB9\x06\x00\xDF\xB0\x00\x80\xAB\x95\x06\x00\xAA\x95\x06\x00\xE3\xB0"
-     "\x00\x80"
-     "\xE7\xB0\x00\x80\xAF\x25\x06\x00\xAE\x31\x06\x00\xAD\x85\x06\x00\xAC\x85"
-     "\x06\x00"
-     "\xEB\xB0\x00\x80\xB3\x3D\x06\x00\xEF\xB0\x00\x80\xF3\xB0\x00\x80\xB6\xCD"
-     "\x01\x00"
-     "\xF7\xB0\x00\x80\xFB\xB0\x00\x80\xB5\xCD\x01\x00\xBA\xA1\x01\x00\xBB\xAD"
-     "\x01\x00"
-     "\xFF\xB0\x00\x80\x03\xB1\x00\x80\xBE\xA9\x01\x00\xBF\x9D\x01\x00\xBC\xB5"
-     "\x01\x00"
-     "\xBD\xA9\x01\x00\xA8\xE5\x05\x00\xA9\xF5\x05\x00\xAA\xF1\x05\x00\xAB\x01"
-     "\x02\x00"
-     "\xAC\x01\x02\x00\xAD\x09\x02\x00\xAE\x31\x02\x00\xAF\x31\x02\x00\xBE\x00"
-     "\x0C\x00"
-     "\x07\xB1\x00\x80\x0B\xB1\x00\x80\x0F\xB1\x00\x80\x80\x1D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\xB9\x00\x00\x13\xB1\x00\x80\xB8\xE9\x02\x00\xB9\xE9\x02\x00\xBA\xF9"
-     "\x02\x00"
-     "\xBB\xF9\x02\x00\xBC\xE9\x02\x00\xBD\xE9\x02\x00\xBE\x5D\x03\x00\xBF\x51"
-     "\x03\x00"
-     "\xB0\x21\x02\x00\xB1\x2D\x02\x00\xB2\x25\x02\x00\xB3\x39\x02\x00\xB4\x29"
-     "\x02\x00"
-     "\xB5\x1D\x02\x00\xB6\x15\x02\x00\xB7\xD9\x02\x00\x17\xB1\x00\x80\x54\xAF"
-     "\x00\x80"
-     "\x84\x80\x03\x00\xA3\x75\x05\x00\x1F\xB1\x00\x80\xA5\x85\x02\x00\xA6\x85"
-     "\x02\x00"
-     "\x23\xB1\x00\x80\x27\xB1\x00\x80\x2B\xB1\x00\x80\xAA\xE9\x02\x00\xAB\xE5"
-     "\x02\x00"
-     "\xAC\xFD\x02\x00\xAD\xE1\x02\x00\xAE\xE1\x02\x00\xAF\xD5\x02\x00\xE1\xE8"
-     "\x01\x00"
-     "\x2F\xB1\x00\x80\xE3\x50\x02\x00\x84\xC0\x0C\x00\x86\x50\x0C\x00\x87\x70"
-     "\x03\x00"
-     "\x33\xB1\x00\x80\x37\xB1\x00\x80\x3B\xB1\x00\x80\x3F\xB1\x00\x80\x43\xB1"
-     "\x00\x80"
-     "\x47\xB1\x00\x80\x4B\xB1\x00\x80\x4F\xB1\x00\x80\xEF\x20\x02\x00\x53\xB1"
-     "\x00\x80"
-     "\xA8\x2D\x02\x00\xA9\x39\x02\x00\xAA\x41\x02\x00\xAB\x41\x02\x00\xAC\x41"
-     "\x02\x00"
-     "\xAD\x49\x02\x00\xAE\x71\x02\x00\xAF\x71\x02\x00\x57\xB1\x00\x80\x5B\xB1"
-     "\x00\x80"
-     "\xBE\xA4\x0C\x00\x5F\xB1\x00\x80\x63\xB1\x00\x80\x67\xB1\x00\x80\x6B\xB1"
-     "\x00\x80"
-     "\x6F\xB1\x00\x80\xB8\x89\x02\x00\xB9\x89\x02\x00\xBA\x9D\x02\x00\xBB\x95"
-     "\x02\x00"
-     "\xBC\xB9\x02\x00\xBD\xB9\x02\x00\xBE\x7D\x01\x00\xBF\x75\x01\x00\xB0\xE5"
-     "\x02\x00"
-     "\xB1\xED\x02\x00\xB2\xE5\x02\x00\xB3\xFD\x02\x00\xB4\xED\x02\x00\xB5\xDD"
-     "\x02\x00"
-     "\xB6\xD5\x02\x00\xB7\xBD\x02\x00\xE1\x68\x07\x00\xE3\x34\x01\x00\xE3\x08"
-     "\x06\x00"
-     "\xE1\x7C\x06\x00\x73\xB1\x00\x80\x77\xB1\x00\x80\x7B\xB1\x00\x80\x7F\xB1"
-     "\x00\x80"
-     "\x81\x09\x00\x00\x80\x39\x00\x00\x83\xB1\x00\x80\x82\x1D\x00\x00\x87\xB1"
-     "\x00\x80"
-     "\xEF\x64\x00\x00\xEF\xCC\x06\x00\x8F\xB1\x00\x80\x93\xB1\x00\x80\xB3\x8D"
-     "\x02\x00"
-     "\x97\xB1\x00\x80\x9B\xB1\x00\x80\xB6\x8D\x02\x00\x9F\xB1\x00\x80\xA3\xB1"
-     "\x00\x80"
-     "\xB5\x8D\x02\x00\xBA\x65\x02\x00\xBB\x6D\x02\x00\x86\xC8\x0C\x00\x87\xE0"
-     "\x0C\x00"
-     "\xBE\x6D\x02\x00\xBF\x5D\x02\x00\xBC\x75\x02\x00\xBD\x6D\x02\x00\x1B\xB1"
-     "\x00\x80"
-     "\x8B\xB1\x00\x80\xA7\xB1\x00\x80\xAB\xB1\x00\x80\xAF\xB1\x00\x80\xB3\xB1"
-     "\x00\x80"
-     "\xB7\xB1\x00\x80\xBB\xB1\x00\x80\xA8\xB1\x0E\x00\xA9\xB9\x0E\x00\xAA\x89"
-     "\x0E\x00"
-     "\xAB\x89\x0E\x00\xAC\xF5\x0E\x00\xAD\xED\x0E\x00\xAE\xE5\x0E\x00\xAF\xDD"
-     "\x0E\x00"
-     "\xB0\xA5\x0E\x00\xB1\xAD\x0E\x00\xB2\xA5\x0E\x00\xB3\xBD\x0E\x00\xB4\xA5"
-     "\x0E\x00"
-     "\xB5\xA9\x0E\x00\xB6\x79\x0F\x00\xB7\x79\x0F\x00\xB8\x4D\x0F\x00\xB9\x55"
-     "\x0F\x00"
-     "\xBA\x5D\x0F\x00\xBB\x55\x0F\x00\xBC\x4D\x0F\x00\xBD\xE5\x0F\x00\xBE\xED"
-     "\x0F\x00"
-     "\xBF\xE5\x0F\x00\xBF\xB1\x00\x80\xC3\xB1\x00\x80\xAA\x0D\x0D\x00\xAB\x05"
-     "\x0D\x00"
-     "\xAC\x1D\x0D\x00\xAD\x05\x0D\x00\xAE\x05\x0D\x00\xAF\x35\x0D\x00\xC7\xB1"
-     "\x00\x80"
-     "\xCB\xB1\x00\x80\xCF\xB1\x00\x80\xA3\xE5\x0D\x00\xD3\xB1\x00\x80\xA5\xE5"
-     "\x0D\x00"
-     "\xA6\xE5\x0D\x00\xD7\xB1\x00\x80\xB3\x61\x0E\x00\xDB\xB1\x00\x80\xDF\xB1"
-     "\x00\x80"
-     "\xE3\xB1\x00\x80\xE7\xB1\x00\x80\xB6\x0D\x0E\x00\xB5\x1D\x0E\x00\xEB\xB1"
-     "\x00\x80"
-     "\xBB\x09\x0E\x00\xBA\x01\x0E\x00\xEF\xB1\x00\x80\xF3\xB1\x00\x80\xBF\x0D"
-     "\x0E\x00"
-     "\xBE\x01\x0E\x00\xBD\x09\x0E\x00\xBC\x11\x0E\x00\xA3\xA1\x0F\x00\x82\x15"
-     "\x00\x00"
-     "\x81\xE5\x01\x00\x80\xE5\x01\x00\xF7\xB1\x00\x80\xA6\xCD\x0F\x00\xA5\xDD"
-     "\x0F\x00"
-     "\xBE\x90\x01\x00\xAB\xC9\x0F\x00\xAA\xC1\x0F\x00\x84\xA0\x02\x00\xFB\xB1"
-     "\x00\x80"
-     "\xAF\xCD\x0F\x00\xAE\xC1\x0F\x00\xAD\xC9\x0F\x00\xAC\xD1\x0F\x00\xFF\xB1"
-     "\x00\x80"
-     "\xB3\x8D\x0E\x00\x86\xE8\x00\x00\x87\x2C\x01\x00\xB6\x89\x0E\x00\x03\xB2"
-     "\x00\x80"
-     "\x07\xB2\x00\x80\xB5\x89\x0E\x00\xBA\x71\x01\x00\xBB\x71\x01\x00\x0B\xB2"
-     "\x00\x80"
-     "\x0F\xB2\x00\x80\xBE\xF1\x01\x00\xBF\xF1\x01\x00\xBC\x61\x01\x00\xBD\x61"
-     "\x01\x00"
-     "\xA8\x3D\x0E\x00\xA9\x45\x0E\x00\xAA\x41\x0E\x00\xAB\x41\x0E\x00\xAC\x45"
-     "\x0E\x00"
-     "\xAD\x4D\x0E\x00\xAE\x91\x01\x00\xAF\x91\x01\x00\x84\x8C\x01\x00\x13\xB2"
-     "\x00\x80"
-     "\x17\xB2\x00\x80\x1B\xB2\x00\x80\x1F\xB2\x00\x80\x23\xB2\x00\x80\x27\xB2"
-     "\x00\x80"
-     "\x2B\xB2\x00\x80\xB8\x69\x01\x00\xB9\x75\x01\x00\xBA\x75\x01\x00\xBB\x09"
-     "\x01\x00"
-     "\xBC\x19\x01\x00\xBD\x19\x01\x00\xBE\x09\x01\x00\xBF\x01\x01\x00\xB0\xF1"
-     "\x01\x00"
-     "\xB1\xF1\x01\x00\xB2\xF1\x01\x00\xB3\xF1\x01\x00\xB4\x71\x01\x00\xB5\x71"
-     "\x01\x00"
-     "\xB6\x59\x01\x00\xB7\x59\x01\x00\xA3\xC9\x0D\x00\x2F\xB2\x00\x80\x33\xB2"
-     "\x00\x80"
-     "\xBE\x40\x03\x00\x37\xB2\x00\x80\xA6\xCD\x0D\x00\xA5\xCD\x0D\x00\x3B\xB2"
-     "\x00\x80"
-     "\xAB\x35\x02\x00\xAA\x35\x02\x00\x3F\xB2\x00\x80\x43\xB2\x00\x80\xAF\xB5"
-     "\x02\x00"
-     "\xAE\xB5\x02\x00\xAD\x25\x02\x00\xAC\x25\x02\x00\x47\xB2\x00\x80\x4B\xB2"
-     "\x00\x80"
-     "\x4F\xB2\x00\x80\x53\xB2\x00\x80\x57\xB2\x00\x80\x5B\xB2\x00\x80\x5F\xB2"
-     "\x00\x80"
-     "\x63\xB2\x00\x80\x80\x39\x00\x00\x81\x39\x00\x00\x82\x05\x00\x00\x67\xB2"
-     "\x00\x80"
-     "\x86\xD0\x04\x00\x87\x44\x03\x00\x6F\xB2\x00\x80\x73\xB2\x00\x80\x84\xCC"
-     "\x04\x00"
-     "\x77\xB2\x00\x80\x7B\xB2\x00\x80\x7F\xB2\x00\x80\xEF\x64\x02\x00\x83\xB2"
-     "\x00\x80"
-     "\x87\xB2\x00\x80\x8B\xB2\x00\x80\xE3\x10\x01\x00\x8F\xB2\x00\x80\xE1\xF8"
-     "\x01\x00"
-     "\x93\xB2\x00\x80\x97\xB2\x00\x80\x9B\xB2\x00\x80\x9F\xB2\x00\x80\xA3\xB2"
-     "\x00\x80"
-     "\xBB\x75\x03\x00\xBA\x75\x03\x00\xA7\xB2\x00\x80\xAB\xB2\x00\x80\xBF\xD9"
-     "\x00\x00"
-     "\xBE\xD1\x00\x00\xBD\x65\x03\x00\xBC\x65\x03\x00\xB3\xB9\x03\x00\xAF\xB2"
-     "\x00\x80"
-     "\xB3\xB2\x00\x80\xB7\xB2\x00\x80\xBB\xB2\x00\x80\xB6\x51\x03\x00\xB5\x95"
-     "\x03\x00"
-     "\xBF\xB2\x00\x80\xA8\x41\x02\x00\xA9\x51\x02\x00\xAA\x51\x02\x00\xAB\x65"
-     "\x02\x00"
-     "\xAC\x7D\x02\x00\xAD\xB1\x02\x00\xAE\xB1\x02\x00\xAF\xB1\x02\x00\xBE\x68"
-     "\x05\x00"
-     "\xC3\xB2\x00\x80\xC7\xB2\x00\x80\xCB\xB2\x00\x80\x80\x0D\x00\x00\x81\xA5"
-     "\x00\x00"
-     "\x82\xAD\x00\x00\xCF\xB2\x00\x80\xB8\x69\x01\x00\xB9\x69\x01\x00\xBA\x09"
-     "\x01\x00"
-     "\xBB\x09\x01\x00\xBC\x19\x01\x00\xBD\x19\x01\x00\xBE\x09\x01\x00\xBF\x09"
-     "\x01\x00"
-     "\xB0\xD1\x02\x00\xB1\xD1\x02\x00\xB2\xD1\x02\x00\xB3\xD1\x02\x00\xB4\x75"
-     "\x01\x00"
-     "\xB5\x7D\x01\x00\xB6\x69\x01\x00\xB7\x61\x01\x00\xE1\x70\x0E\x00\xE1\x4C"
-     "\x0F\x00"
-     "\xE3\xB0\x00\x00\xE3\x50\x0F\x00\xD3\xB2\x00\x80\xD7\xB2\x00\x80\xEF\x48"
-     "\xE7\x00"
-     "\x85\x98\x04\x00\xE1\x28\x02\x00\xDF\xB2\x00\x80\xE3\x70\x13\x00\xE3\xB2"
-     "\x00\x80"
-     "\xBE\x78\x06\x00\xE7\xB2\x00\x80\xEF\x5C\x00\x00\xEF\x6C\x0E\x00\xA3\x99"
-     "\x02\x00"
-     "\xEB\xB2\x00\x80\x86\xA8\x07\x00\x87\x80\x04\x00\xEF\xB2\x00\x80\xA6\x71"
-     "\x02\x00"
-     "\xA5\xB5\x02\x00\xF3\xB2\x00\x80\xAB\x55\x02\x00\xAA\x55\x02\x00\xF7\xB2"
-     "\x00\x80"
-     "\xFB\xB2\x00\x80\xAF\xF9\x01\x00\xAE\xF1\x01\x00\xAD\x45\x02\x00\xAC\x45"
-     "\x02\x00"
-     "\xA8\x79\x06\x00\xA9\x79\x06\x00\xAA\x8D\x06\x00\xAB\x85\x06\x00\xAC\x9D"
-     "\x06\x00"
-     "\xAD\x81\x06\x00\xAE\x81\x06\x00\xAF\x81\x06\x00\xDB\xB2\x00\x80\xFF\xB2"
-     "\x00\x80"
-     "\x03\xB3\x00\x80\x07\xB3\x00\x80\x0B\xB3\x00\x80\x0F\xB3\x00\x80\x13\xB3"
-     "\x00\x80"
-     "\x17\xB3\x00\x80\xB8\xA5\x06\x00\xB9\xAD\x06\x00\xBA\xA5\x06\x00\xBB\xBD"
-     "\x06\x00"
-     "\xBC\xAD\x06\x00\xBD\x71\x07\x00\xBE\x71\x07\x00\xBF\x71\x07\x00\xB0\xC5"
-     "\x06\x00"
-     "\xB1\xC9\x06\x00\xB2\xD9\x06\x00\xB3\xD9\x06\x00\xB4\xC9\x06\x00\xB5\xC9"
-     "\x06\x00"
-     "\xB6\xB1\x06\x00\xB7\xB1\x06\x00\xB3\x05\x06\x00\x1B\xB3\x00\x80\x1F\xB3"
-     "\x00\x80"
-     "\x23\xB3\x00\x80\x27\xB3\x00\x80\xB6\x01\x06\x00\xB5\x15\x06\x00\x2B\xB3"
-     "\x00\x80"
-     "\xBB\x0D\x06\x00\xBA\x01\x06\x00\x2F\xB3\x00\x80\x33\xB3\x00\x80\xBF\x0D"
-     "\x06\x00"
-     "\xBE\x0D\x06\x00\xBD\x0D\x06\x00\xBC\x0D\x06\x00\x37\xB3\x00\x80\xA3\x41"
-     "\x06\x00"
-     "\x3B\xB3\x00\x80\x3F\xB3\x00\x80\xA6\x45\x06\x00\x43\xB3\x00\x80\x47\xB3"
-     "\x00\x80"
-     "\xA5\x51\x06\x00\xAA\x45\x06\x00\xAB\x49\x06\x00\x84\x24\x01\x00\x4F\xB3"
-     "\x00\x80"
-     "\xAE\x49\x06\x00\xAF\x49\x06\x00\xAC\x49\x06\x00\xAD\x49\x06\x00\x9F\x45"
-     "\x08\x00"
-     "\x9E\x0D\x09\x00\x9D\x0D\x09\x00\x9C\x89\x0B\x00\x9B\x69\x0D\x00\x9A\x11"
-     "\x0D\x00"
-     "\x99\x89\x0F\x00\x98\x0D\xF1\x00\x97\x25\xF1\x00\x96\x25\xF1\x00\x95\xFD"
-     "\xF5\x00"
-     "\x94\x0D\xF5\x00\x93\x99\xF7\x00\x92\x99\xF9\x00\x91\x31\xF9\x00\x90\x39"
-     "\xF9\x00"
-     "\x8F\x29\xFD\x00\x82\x9D\x01\x00\x81\x9D\x01\x00\x80\x81\x01\x00\x53\xB3"
-     "\x00\x80"
-     "\xEF\x8C\x04\x00\xEF\xEC\x01\x00\x57\xB3\x00\x80\xE1\x08\x02\x00\xE3\xF4"
-     "\x05\x00"
-     "\xE3\x30\x01\x00\xE1\x04\x05\x00\x86\x00\x1C\x00\x87\x88\x01\x00\x5B\xB3"
-     "\x00\x80"
-     "\x5F\xB3\x00\x80\xA1\xA1\x01\x00\x63\xB3\x00\x80\xA3\x09\x1D\x00\xA2\x69"
-     "\x03\x00"
-     "\xA5\xF9\x1C\x00\xA4\x11\x1D\x00\xA7\x1D\x19\x00\xA6\x1D\x19\x00\xA9\x21"
-     "\x15\x00"
-     "\xA8\xB5\x18\x00\xAB\x79\x15\x00\xAA\x21\x15\x00\xAD\x91\x11\x00\xAC\x01"
-     "\x10\x00"
-     "\xAF\x01\x2C\x00\xAE\x89\x10\x00\xB1\x5D\x2D\x00\xB0\xE5\x2D\x00\xB3\xE9"
-     "\x29\x00"
-     "\xB2\x01\x28\x00\xB5\x01\x28\x00\xB4\xED\x29\x00\x67\xB3\x00\x80\x6B\xB3"
-     "\x00\x80"
-     "\x6F\xB3\x00\x80\x73\xB3\x00\x80\x77\xB3\x00\x80\x7B\xB3\x00\x80\xA3\x19"
-     "\x03\x00"
-     "\xA2\xD9\x04\x00\xA1\xFD\x05\x00\xA0\xFD\x05\x00\x82\x69\xFE\x00\x83\x01"
-     "\xFA\x00"
-     "\x7F\xB3\x00\x80\x83\xB3\x00\x80\x86\xB9\xF4\x00\x87\x35\xF6\x00\x84\x11"
-     "\xFA\x00"
-     "\x85\xB1\xFA\x00\x8A\xF5\xF6\x00\x8B\x25\xF2\x00\x87\xB3\x00\x80\x8B\xB3"
-     "\x00\x80"
-     "\x8E\xA1\x0C\x00\x8F\x31\x0E\x00\x8C\x3D\xF2\x00\x8D\x2D\xF2\x00\x92\x39"
-     "\x0E\x00"
-     "\x93\xC1\x08\x00\x84\x44\x03\x00\xBE\xC0\x1C\x00\x96\x15\x0B\x00\x97\x55"
-     "\x06\x00"
-     "\x94\x5D\x0A\x00\x95\x55\x0A\x00\x9A\x65\x06\x00\x9B\x49\x07\x00\x8F\xB3"
-     "\x00\x80"
-     "\x93\xB3\x00\x80\x97\xB3\x00\x80\x9B\xB3\x00\x80\x9C\x91\x00\x00\x9F\xB3"
-     "\x00\x80"
-     "\xB8\x09\x02\x00\xB9\x09\x02\x00\xBA\x19\x02\x00\xBB\x11\x02\x00\xBC\x39"
-     "\x02\x00"
-     "\xBD\x39\x02\x00\xBE\x59\x02\x00\xBF\x55\x02\x00\xB0\x85\x02\x00\xB1\x8D"
-     "\x02\x00"
-     "\xB2\x85\x02\x00\xB3\x9D\x02\x00\xB4\x85\x02\x00\xB5\x8D\x02\x00\xB6\x85"
-     "\x02\x00"
-     "\xB7\x39\x02\x00\xA8\x65\x05\x00\xA9\x69\x05\x00\xAA\x7D\x05\x00\xAB\x09"
-     "\x05\x00"
-     "\xAC\x01\x05\x00\xAD\x09\x05\x00\xAE\x31\x05\x00\xAF\x31\x05\x00\xA3\xB3"
-     "\x00\x80"
-     "\x4B\xB3\x00\x80\x6B\xB2\x00\x80\xA7\xB3\x00\x80\xAB\xB3\x00\x80\x82\xE5"
-     "\x01\x00"
-     "\x81\xF5\x01\x00\x80\xF5\x01\x00\xB3\xB5\x04\x00\x84\x8C\x1D\x00\xAF\xB3"
-     "\x00\x80"
-     "\xB3\xB3\x00\x80\xB7\xB3\x00\x80\xB6\xC9\x04\x00\xB5\xC9\x04\x00\xBF\xB3"
-     "\x00\x80"
-     "\xBB\x09\x03\x00\xBA\x01\x03\x00\x86\x28\x1D\x00\x87\x20\x1D\x00\xBF\x09"
-     "\x03\x00"
-     "\xBE\x01\x03\x00\xBD\x09\x03\x00\xBC\x11\x03\x00\xC3\xB3\x00\x80\xA3\xF1"
-     "\x04\x00"
-     "\xC7\xB3\x00\x80\xBE\xC0\x1C\x00\xA6\x8D\x04\x00\xCB\xB3\x00\x80\xCF\xB3"
-     "\x00\x80"
-     "\xA5\x8D\x04\x00\xAA\x45\x03\x00\xAB\x4D\x03\x00\xBE\xA0\x1F\x00\xD3\xB3"
-     "\x00\x80"
-     "\xAE\x45\x03\x00\xAF\x4D\x03\x00\xAC\x55\x03\x00\xAD\x4D\x03\x00\xD7\xB3"
-     "\x00\x80"
-     "\xDB\xB3\x00\x80\xDF\xB3\x00\x80\xE3\xB3\x00\x80\xE7\xB3\x00\x80\xEB\xB3"
-     "\x00\x80"
-     "\xEF\xB3\x00\x80\xEF\xD4\x06\x00\xF3\xB3\x00\x80\xE1\x74\x06\x00\xF7\xB3"
-     "\x00\x80"
-     "\xE3\x54\x01\x00\xFB\xB3\x00\x80\xFF\xB3\x00\x80\x03\xB4\x00\x80\x07\xB4"
-     "\x00\x80"
-     "\xB3\x9D\x02\x00\x0B\xB4\x00\x80\x0F\xB4\x00\x80\x13\xB4\x00\x80\x17\xB4"
-     "\x00\x80"
-     "\xB6\x51\x02\x00\xB5\xA5\x02\x00\x1B\xB4\x00\x80\xBB\x05\x02\x00\xBA\x05"
-     "\x02\x00"
-     "\x1F\xB4\x00\x80\x23\xB4\x00\x80\xBF\xFD\x01\x00\xBE\xFD\x01\x00\xBD\x15"
-     "\x02\x00"
-     "\xBC\x15\x02\x00\xA3\x5D\x1D\x00\x82\x15\x00\x00\x81\xE5\x1F\x00\x80\xE5"
-     "\x1F\x00"
-     "\x27\xB4\x00\x80\xA6\x91\x1D\x00\xA5\x65\x1D\x00\x2F\xB4\x00\x80\xAB\xC5"
-     "\x1D\x00"
-     "\xAA\xC5\x1D\x00\x86\x08\x03\x00\x87\x7C\x03\x00\xAF\x3D\x1E\x00\xAE\x3D"
-     "\x1E\x00"
-     "\xAD\xD5\x1D\x00\xAC\xD5\x1D\x00\x33\xB4\x00\x80\xB3\x0D\x1F\x00\x37\xB4"
-     "\x00\x80"
-     "\x3B\xB4\x00\x80\xB6\x3D\x1F\x00\x3F\xB4\x00\x80\x43\xB4\x00\x80\xB5\x05"
-     "\x1F\x00"
-     "\xBA\xE9\x1F\x00\xBB\xE9\x1F\x00\x47\xB4\x00\x80\x4B\xB4\x00\x80\xBE\xE9"
-     "\x1F\x00"
-     "\xBF\xD5\x1F\x00\xBC\xE9\x1F\x00\xBD\xE9\x1F\x00\xBB\xB3\x00\x80\x2B\xB4"
-     "\x00\x80"
-     "\x4F\xB4\x00\x80\x53\xB4\x00\x80\x57\xB4\x00\x80\x5B\xB4\x00\x80\x5F\xB4"
-     "\x00\x80"
-     "\x63\xB4\x00\x80\xA8\xBD\x1F\x00\xA9\xE5\x1F\x00\xAA\xE1\x1F\x00\xAB\xF9"
-     "\x1F\x00"
-     "\xAC\xE9\x1F\x00\xAD\xD9\x1F\x00\xAE\x39\x1F\x00\xAF\x39\x1F\x00\xB0\x49"
-     "\x1F\x00"
-     "\xB1\x49\x1F\x00\xB2\x59\x1F\x00\xB3\x59\x1F\x00\xB4\x49\x1F\x00\xB5\x75"
-     "\x1F\x00"
-     "\xB6\x71\x1F\x00\xB7\x6D\x1F\x00\xB8\x55\x1F\x00\xB9\x5D\x1F\x00\xBA\x55"
-     "\x1F\x00"
-     "\xBB\x29\x1F\x00\xBC\x39\x1F\x00\xBD\x39\x1F\x00\xBE\x29\x1F\x00\xBF\x21"
-     "\x1F\x00"
-     "\xA3\x49\x1E\x00\x67\xB4\x00\x80\x6B\xB4\x00\x80\x6F\xB4\x00\x80\x73\xB4"
-     "\x00\x80"
-     "\xA6\x79\x1E\x00\xA5\x41\x1E\x00\x77\xB4\x00\x80\xAB\xAD\x1E\x00\xAA\xAD"
-     "\x1E\x00"
-     "\x7B\xB4\x00\x80\x7F\xB4\x00\x80\xAF\x91\x1E\x00\xAE\xAD\x1E\x00\xAD\xAD"
-     "\x1E\x00"
-     "\xAC\xAD\x1E\x00\x83\xB4\x00\x80\x87\xB4\x00\x80\x8B\xB4\x00\x80\x8F\xB4"
-     "\x00\x80"
-     "\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39\x00\x00\x93\xB4\x00\x80\x97\xB4"
-     "\x00\x80"
-     "\x84\xA8\x01\x00\x9B\xB4\x00\x80\x9F\xB4\x00\x80\x86\x00\x04\x00\x87\xD4"
-     "\x01\x00"
-     "\xA3\xB4\x00\x80\xA7\xB4\x00\x80\xA8\x9D\x01\x00\xA9\xC5\x01\x00\xAA\xC1"
-     "\x01\x00"
-     "\xAB\xC1\x01\x00\xAC\xC5\x01\x00\xAD\xCD\x01\x00\xAE\xF1\x01\x00\xAF\xF1"
-     "\x01\x00"
-     "\xAB\xB4\x00\x80\xAF\xB4\x00\x80\xB3\xB4\x00\x80\xB7\xB4\x00\x80\xBB\xB4"
-     "\x00\x80"
-     "\xBF\xB4\x00\x80\xC3\xB4\x00\x80\xC7\xB4\x00\x80\xB8\xCD\x00\x00\xB9\xD5"
-     "\x00\x00"
-     "\xBA\xD5\x00\x00\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x9D"
-     "\x00\x00"
-     "\xBF\x95\x00\x00\xB0\x69\x01\x00\xB1\x69\x01\x00\xB2\x79\x01\x00\xB3\x79"
-     "\x01\x00"
-     "\xB4\x69\x01\x00\xB5\x69\x01\x00\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xB3\x49"
-     "\x02\x00"
-     "\xCB\xB4\x00\x80\xCF\xB4\x00\x80\xBE\x80\x04\x00\xD3\xB4\x00\x80\xB6\x4D"
-     "\x02\x00"
-     "\xB5\x4D\x02\x00\xD7\xB4\x00\x80\xBB\xAD\x02\x00\xBA\xAD\x02\x00\xDB\xB4"
-     "\x00\x80"
-     "\xDF\xB4\x00\x80\xBF\xAD\x02\x00\xBE\xAD\x02\x00\xBD\xAD\x02\x00\xBC\xAD"
-     "\x02\x00"
-     "\xE3\xB4\x00\x80\xA3\x0D\x02\x00\xE7\xB4\x00\x80\xEB\xB4\x00\x80\xA6\x09"
-     "\x02\x00"
-     "\xEF\xB4\x00\x80\xF3\xB4\x00\x80\xA5\x09\x02\x00\xAA\xE9\x02\x00\xAB\xE9"
-     "\x02\x00"
-     "\xF7\xB4\x00\x80\xFB\xB4\x00\x80\xAE\xE9\x02\x00\xAF\xE9\x02\x00\xAC\xE9"
-     "\x02\x00"
-     "\xAD\xE9\x02\x00\xFF\xB4\x00\x80\x03\xB5\x00\x80\xBE\x04\x05\x00\xBE\x14"
-     "\x06\x00"
-     "\x07\xB5\x00\x80\x0F\xB5\x00\x80\x13\xB5\x00\x80\x17\xB5\x00\x80\x80\x21"
-     "\x00\x00"
-     "\x81\x21\x00\x00\x82\x39\x00\x00\x84\x20\x03\x00\x86\xB0\x04\x00\x87\x54"
-     "\x03\x00"
-     "\x1B\xB5\x00\x80\x1F\xB5\x00\x80\x23\xB5\x00\x80\x27\xB5\x00\x80\x2B\xB5"
-     "\x00\x80"
-     "\x2F\xB5\x00\x80\x33\xB5\x00\x80\xEF\x20\x02\x00\x37\xB5\x00\x80\x3B\xB5"
-     "\x00\x80"
-     "\x3F\xB5\x00\x80\xE3\x58\x02\x00\x43\xB5\x00\x80\xE1\x84\x01\x00\x47\xB5"
-     "\x00\x80"
-     "\x4B\xB5\x00\x80\x4F\xB5\x00\x80\x53\xB5\x00\x80\xE1\x30\x01\x00\xE1\x34"
-     "\x1E\x00"
-     "\xE3\x7C\x00\x00\xE3\x18\x1E\x00\x57\xB5\x00\x80\x5B\xB5\x00\x80\x5F\xB5"
-     "\x00\x80"
-     "\x63\xB5\x00\x80\x84\xA8\x05\x00\x67\xB5\x00\x80\x6B\xB5\x00\x80\x6F\xB5"
-     "\x00\x80"
-     "\x73\xB5\x00\x80\x77\xB5\x00\x80\xEF\xA8\x1E\x00\xEF\x88\x1E\x00\xB3\x3D"
-     "\x02\x00"
-     "\x7F\xB5\x00\x80\x86\x68\x04\x00\x87\xDC\x05\x00\xD4\x00\x00\x00\xB6\xFD"
-     "\x02\x00"
-     "\xB5\xFD\x02\x00\x83\xB5\x00\x80\xBB\xD9\x02\x00\xBA\xD1\x02\x00\xBE\xA4"
-     "\x06\x00"
-     "\x87\xB5\x00\x80\xBF\x61\x01\x00\xBE\x7D\x01\x00\xBD\x7D\x01\x00\xBC\xC1"
-     "\x02\x00"
-     "\xBC\xFD\x1D\x00\xBD\x65\x1D\x00\xBE\x65\x1D\x00\xBF\x7D\x1D\x00\xB8\xF9"
-     "\x1D\x00"
-     "\xB9\xF9\x1D\x00\xBA\xE9\x1D\x00\xBB\xE9\x1D\x00\xB4\xA9\x1D\x00\xB5\xA9"
-     "\x1D\x00"
-     "\xB6\x81\x1D\x00\xB7\x81\x1D\x00\xB0\x51\x1E\x00\xB1\x51\x1E\x00\xB2\x51"
-     "\x1E\x00"
-     "\xB3\x51\x1E\x00\xAC\xBD\x1E\x00\xAD\x41\x1E\x00\xAE\x11\x1E\x00\xAF\x11"
-     "\x1E\x00"
-     "\xA8\x3D\x1F\x00\xA9\x85\x1E\x00\xAA\x85\x1E\x00\xAB\xA5\x1E\x00\x80\xCD"
-     "\x00\x00"
-     "\x81\xD5\x00\x00\x82\xDD\x00\x00\x8B\xB5\x00\x80\x8F\xB5\x00\x80\x93\xB5"
-     "\x00\x80"
-     "\x97\xB5\x00\x80\x9B\xB5\x00\x80\xA3\x7D\x05\x00\x9F\xB5\x00\x80\xA3\xB5"
-     "\x00\x80"
-     "\xA7\xB5\x00\x80\xAB\xB5\x00\x80\xA6\xBD\x05\x00\xA5\xBD\x05\x00\xAF\xB5"
-     "\x00\x80"
-     "\xAB\x99\x05\x00\xAA\x91\x05\x00\xB3\xB5\x00\x80\xB7\xB5\x00\x80\xAF\x21"
-     "\x06\x00"
-     "\xAE\x3D\x06\x00\xAD\x3D\x06\x00\xAC\x81\x05\x00\xBB\xB5\x00\x80\xB3\xD9"
-     "\x06\x00"
-     "\xBF\xB5\x00\x80\xC3\xB5\x00\x80\xB6\x91\x06\x00\xC7\xB5\x00\x80\xCB\xB5"
-     "\x00\x80"
-     "\xB5\xE1\x06\x00\xBA\xAD\x06\x00\xBB\xB5\x06\x00\xCF\xB5\x00\x80\xD3\xB5"
-     "\x00\x80"
-     "\xBE\x91\x06\x00\xBF\x9D\x06\x00\xBC\xA1\x06\x00\xBD\xA1\x06\x00\x0B\xB5"
-     "\x00\x80"
-     "\x7B\xB5\x00\x80\xD7\xB5\x00\x80\xDB\xB5\x00\x80\x80\x0D\x00\x00\x81\x91"
-     "\x01\x00"
-     "\x82\x91\x01\x00\xDF\xB5\x00\x80\xA8\x35\x06\x00\xA9\x3D\x06\x00\xAA\x35"
-     "\x06\x00"
-     "\xAB\x15\x06\x00\xAC\x0D\x06\x00\xAD\x75\x06\x00\xAE\x7D\x06\x00\xAF\x75"
-     "\x06\x00"
-     "\xB0\x0D\x06\x00\xB1\x15\x06\x00\xB2\x11\x06\x00\xB3\x11\x06\x00\xB4\x31"
-     "\x06\x00"
-     "\xB5\x3D\x06\x00\xB6\x29\x06\x00\xB7\x29\x06\x00\xB8\x19\x06\x00\xB9\x19"
-     "\x06\x00"
-     "\xBA\xF9\x07\x00\xBB\xF9\x07\x00\xBC\xE9\x07\x00\xBD\xE9\x07\x00\xBE\xDD"
-     "\x07\x00"
-     "\xBF\xD5\x07\x00\xA3\x9D\x07\x00\xE3\xB5\x00\x80\xE7\xB5\x00\x80\x84\x84"
-     "\x02\x00"
-     "\xEB\xB5\x00\x80\xA6\xD5\x07\x00\xA5\xA5\x07\x00\xEF\xB5\x00\x80\xAB\xF1"
-     "\x07\x00"
-     "\xAA\xE9\x07\x00\x86\x00\x0C\x00\x87\x7C\x01\x00\xAF\xD9\x07\x00\xAE\xD5"
-     "\x07\x00"
-     "\xAD\xE5\x07\x00\xAC\xE5\x07\x00\xF3\xB5\x00\x80\xB3\x29\x06\x00\xF7\xB5"
-     "\x00\x80"
-     "\xFB\xB5\x00\x80\xB6\xA5\x06\x00\xFF\xB5\x00\x80\x03\xB6\x00\x80\xB5\xAD"
-     "\x06\x00"
-     "\xBA\x65\x01\x00\xBB\x6D\x01\x00\x07\xB6\x00\x80\x0B\xB6\x00\x80\xBE\x6D"
-     "\x01\x00"
-     "\xBF\x5D\x01\x00\xBC\x75\x01\x00\xBD\x6D\x01\x00\xA8\xC5\x06\x00\xA9\xD9"
-     "\x06\x00"
-     "\xAA\xE5\x06\x00\xAB\x61\x06\x00\xAC\x61\x06\x00\xAD\x61\x06\x00\xAE\x61"
-     "\x06\x00"
-     "\xAF\x61\x06\x00\x0F\xB6\x00\x80\x13\xB6\x00\x80\x17\xB6\x00\x80\x1B\xB6"
-     "\x00\x80"
-     "\x1F\xB6\x00\x80\x23\xB6\x00\x80\x27\xB6\x00\x80\x2B\xB6\x00\x80\xB8\x89"
-     "\x01\x00"
-     "\xB9\x89\x01\x00\xBA\x9D\x01\x00\xBB\x95\x01\x00\xBC\xB9\x01\x00\xBD\xB9"
-     "\x01\x00"
-     "\xBE\x51\x01\x00\xBF\x51\x01\x00\xB0\xE5\x01\x00\xB1\xED\x01\x00\xB2\xE5"
-     "\x01\x00"
-     "\xB3\xFD\x01\x00\xB4\xED\x01\x00\xB5\xDD\x01\x00\xB6\xD5\x01\x00\xB7\xBD"
-     "\x01\x00"
-     "\xA6\xF5\x05\x00\x2F\xB6\x00\x80\x33\xB6\x00\x80\xA5\xFD\x05\x00\x37\xB6"
-     "\x00\x80"
-     "\xA3\x79\x05\x00\x3B\xB6\x00\x80\x3F\xB6\x00\x80\xAE\x3D\x02\x00\xAF\x0D"
-     "\x02\x00"
-     "\xAC\x25\x02\x00\xAD\x3D\x02\x00\xAA\x35\x02\x00\xAB\x3D\x02\x00\x43\xB6"
-     "\x00\x80"
-     "\x47\xB6\x00\x80\x4B\xB6\x00\x80\x4F\xB6\x00\x80\x53\xB6\x00\x80\x57\xB6"
-     "\x00\x80"
-     "\x80\x15\x00\x00\x81\x1D\x00\x00\x82\x15\x00\x00\x5F\xB6\x00\x80\x84\x60"
-     "\x02\x00"
-     "\x63\xB6\x00\x80\x87\x48\x03\x00\x86\xCC\x0C\x00\x67\xB6\x00\x80\x6B\xB6"
-     "\x00\x80"
-     "\x6F\xB6\x00\x80\x73\xB6\x00\x80\x77\xB6\x00\x80\x7B\xB6\x00\x80\x86\x28"
-     "\x0D\x00"
-     "\x87\x40\x0C\x00\xA1\x00\x00\x00\xE1\xA4\x01\x00\x7F\xB6\x00\x80\xE3\xEC"
-     "\x02\x00"
-     "\x83\xB6\x00\x80\x87\xB6\x00\x80\x8B\xB6\x00\x80\xEF\x3C\x07\x00\x8F\xB6"
-     "\x00\x80"
-     "\x93\xB6\x00\x80\x97\xB6\x00\x80\x9B\xB6\x00\x80\x9F\xB6\x00\x80\xA3\xB6"
-     "\x00\x80"
-     "\xA7\xB6\x00\x80\xB3\x85\x03\x00\xAB\xB6\x00\x80\xB5\x81\x03\x00\xB6\x81"
-     "\x03\x00"
-     "\xAF\xB6\x00\x80\xB3\xB6\x00\x80\xB7\xB6\x00\x80\xBA\x49\x03\x00\xBB\x49"
-     "\x03\x00"
-     "\xBC\x59\x03\x00\xBD\x59\x03\x00\xBE\x49\x03\x00\xBF\x49\x03\x00\xE1\x50"
-     "\x06\x00"
-     "\xE1\xAC\x07\x00\xE3\x14\x00\x00\xE3\x3C\x06\x00\xBB\xB6\x00\x80\xBF\xB6"
-     "\x00\x80"
-     "\xC3\xB6\x00\x80\x84\x54\x0D\x00\x80\x35\x00\x00\x81\x3D\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xC7\xB6\x00\x80\xBE\x7C\x0D\x00\xCF\xB6\x00\x80\xEF\x4C\x07\x00\xEF\x4C"
-     "\x07\x00"
-     "\xA3\x09\x02\x00\xD3\xB6\x00\x80\xD7\xB6\x00\x80\xDB\xB6\x00\x80\xDF\xB6"
-     "\x00\x80"
-     "\xA6\x0D\x02\x00\xA5\x0D\x02\x00\xE3\xB6\x00\x80\xAB\xC5\x02\x00\xAA\xC5"
-     "\x02\x00"
-     "\xE7\xB6\x00\x80\xEB\xB6\x00\x80\xAF\xC5\x02\x00\xAE\xC5\x02\x00\xAD\xD5"
-     "\x02\x00"
-     "\xAC\xD5\x02\x00\xBC\xD5\x05\x00\xBD\xDD\x05\x00\xBE\xE1\x05\x00\xBF\xE1"
-     "\x05\x00"
-     "\xB8\xBD\x05\x00\xB9\xD5\x05\x00\xBA\xD1\x05\x00\xBB\xD1\x05\x00\xB4\xA9"
-     "\x05\x00"
-     "\xB5\xB5\x05\x00\xB6\xB5\x05\x00\xB7\x85\x05\x00\xB0\x05\x06\x00\xB1\x0D"
-     "\x06\x00"
-     "\xB2\x05\x06\x00\xB3\xB9\x05\x00\xAC\x65\x06\x00\xAD\x6D\x06\x00\xAE\x65"
-     "\x06\x00"
-     "\xAF\x7D\x06\x00\xA8\xBD\x06\x00\xA9\xAD\x06\x00\xAA\xA5\x06\x00\xAB\x7D"
-     "\x06\x00"
-     "\xEF\xB6\x00\x80\xF3\xB6\x00\x80\xF7\xB6\x00\x80\xFB\xB6\x00\x80\xFF\xB6"
-     "\x00\x80"
-     "\x03\xB7\x00\x80\x07\xB7\x00\x80\x0B\xB7\x00\x80\xA8\x5D\x0E\x00\xA9\x61"
-     "\x0E\x00"
-     "\xAA\x79\x0E\x00\xAB\x79\x0E\x00\xAC\x69\x0E\x00\xAD\x69\x0E\x00\xAE\x99"
-     "\x0E\x00"
-     "\xAF\x99\x0E\x00\x0F\xB7\x00\x80\x13\xB7\x00\x80\x17\xB7\x00\x80\x1B\xB7"
-     "\x00\x80"
-     "\x1F\xB7\x00\x80\x23\xB7\x00\x80\x27\xB7\x00\x80\x2B\xB7\x00\x80\xB8\x81"
-     "\x0E\x00"
-     "\xB9\x81\x0E\x00\xBA\x81\x0E\x00\xBB\x9D\x0E\x00\xBC\x8D\x0E\x00\xBD\xB5"
-     "\x0E\x00"
-     "\xBE\xBD\x0E\x00\xBF\x4D\x0F\x00\xB0\xF9\x0E\x00\xB1\xC5\x0E\x00\xB2\xCD"
-     "\x0E\x00"
-     "\xB3\xC5\x0E\x00\xB4\xC1\x0E\x00\xB5\xC1\x0E\x00\xB6\xC1\x0E\x00\xB7\xC1"
-     "\x0E\x00"
-     "\x5B\xB6\x00\x80\xB3\xF5\x0E\x00\x2F\xB7\x00\x80\xCB\xB6\x00\x80\xB6\xED"
-     "\x0E\x00"
-     "\x33\xB7\x00\x80\x37\xB7\x00\x80\xB5\xE5\x0E\x00\xBA\x39\x0E\x00\xBB\x05"
-     "\x0E\x00"
-     "\x84\xE4\x03\x00\xBE\xE4\x03\x00\xBE\x01\x0E\x00\xBF\x01\x0E\x00\xBC\x1D"
-     "\x0E\x00"
-     "\xBD\x15\x0E\x00\xA3\xB1\x0E\x00\x82\x5D\x00\x00\x81\x5D\x00\x00\x80\x45"
-     "\x00\x00"
-     "\x3B\xB7\x00\x80\xA6\xA9\x0E\x00\xA5\xA1\x0E\x00\x3F\xB7\x00\x80\xAB\x41"
-     "\x0E\x00"
-     "\xAA\x7D\x0E\x00\x86\x48\x00\x00\x87\x68\x00\x00\xAF\x45\x0E\x00\xAE\x45"
-     "\x0E\x00"
-     "\xAD\x51\x0E\x00\xAC\x59\x0E\x00\xA8\x35\x0E\x00\xA9\x41\x0E\x00\xAA\x41"
-     "\x0E\x00"
-     "\xAB\x41\x0E\x00\xAC\x41\x0E\x00\xAD\x49\x0E\x00\xAE\x71\x0E\x00\xAF\x71"
-     "\x0E\x00"
-     "\x43\xB7\x00\x80\x47\xB7\x00\x80\x84\x6C\x01\x00\x4B\xB7\x00\x80\x4F\xB7"
-     "\x00\x80"
-     "\x53\xB7\x00\x80\x57\xB7\x00\x80\x5B\xB7\x00\x80\xB8\x61\x01\x00\xB9\x61"
-     "\x01\x00"
-     "\xBA\x61\x01\x00\xBB\x61\x01\x00\xBC\x61\x01\x00\xBD\x61\x01\x00\xBE\x61"
-     "\x01\x00"
-     "\xBF\x61\x01\x00\xB0\x81\x0E\x00\xB1\x81\x0E\x00\xB2\x81\x0E\x00\xB3\x81"
-     "\x0E\x00"
-     "\xB4\x81\x0E\x00\xB5\x81\x0E\x00\xB6\x81\x0E\x00\xB7\x81\x0E\x00\xB3\xC9"
-     "\x0E\x00"
-     "\x5F\xB7\x00\x80\x63\xB7\x00\x80\x67\xB7\x00\x80\x6B\xB7\x00\x80\xB6\xC9"
-     "\x0E\x00"
-     "\xB5\xC9\x0E\x00\x6F\xB7\x00\x80\xBB\x45\x0E\x00\xBA\x45\x0E\x00\x73\xB7"
-     "\x00\x80"
-     "\x77\xB7\x00\x80\xBF\x45\x0E\x00\xBE\x45\x0E\x00\xBD\x55\x0E\x00\xBC\x55"
-     "\x0E\x00"
-     "\x7B\xB7\x00\x80\xA3\x8D\x0E\x00\x7F\xB7\x00\x80\x83\xB7\x00\x80\xA6\x8D"
-     "\x0E\x00"
-     "\x87\xB7\x00\x80\x8B\xB7\x00\x80\xA5\x8D\x0E\x00\xAA\x01\x0E\x00\xAB\x01"
-     "\x0E\x00"
-     "\x8F\xB7\x00\x80\x93\xB7\x00\x80\xAE\x01\x0E\x00\xAF\x01\x0E\x00\xAC\x11"
-     "\x0E\x00"
-     "\xAD\x11\x0E\x00\xA8\xDD\x02\x00\xA9\x25\x03\x00\xAA\x2D\x03\x00\xAB\x21"
-     "\x03\x00"
-     "\xAC\x21\x03\x00\xAD\x21\x03\x00\xAE\x21\x03\x00\xAF\x21\x03\x00\x97\xB7"
-     "\x00\x80"
-     "\x9B\xB7\x00\x80\x9F\xB7\x00\x80\xA3\xB7\x00\x80\xA7\xB7\x00\x80\xBE\x78"
-     "\x02\x00"
-     "\xAF\xB7\x00\x80\xB3\xB7\x00\x80\xB8\x89\x03\x00\xB9\x89\x03\x00\xBA\x9D"
-     "\x03\x00"
-     "\xBB\x95\x03\x00\xBC\xB9\x03\x00\xBD\xB9\x03\x00\xBE\x79\x00\x00\xBF\x79"
-     "\x00\x00"
-     "\xB0\xE5\x03\x00\xB1\xED\x03\x00\xB2\xE5\x03\x00\xB3\xFD\x03\x00\xB4\xED"
-     "\x03\x00"
-     "\xB5\xDD\x03\x00\xB6\xD5\x03\x00\xB7\xBD\x03\x00\x81\xAD\x00\x00\x80\xAD"
-     "\x00\x00"
-     "\xB3\xAD\x03\x00\x82\xBD\x00\x00\xB5\xA1\x03\x00\xB7\xB7\x00\x80\xBB\xB7"
-     "\x00\x80"
-     "\xB6\xA9\x03\x00\x86\x00\x04\x00\x87\xD4\x03\x00\xBB\x21\x03\x00\xBA\x29"
-     "\x03\x00"
-     "\xBD\x21\x03\x00\xBC\x39\x03\x00\xBF\x11\x03\x00\xBE\x21\x03\x00\xA3\xE5"
-     "\x03\x00"
-     "\xBF\xB7\x00\x80\xC3\xB7\x00\x80\x85\xF0\x30\x00\xC7\xB7\x00\x80\xA6\xE1"
-     "\x03\x00"
-     "\xA5\xE9\x03\x00\xCB\xB7\x00\x80\xAB\x69\x03\x00\xAA\x61\x03\x00\xCF\xB7"
-     "\x00\x80"
-     "\xD3\xB7\x00\x80\xAF\x59\x03\x00\xAE\x69\x03\x00\xAD\x69\x03\x00\xAC\x71"
-     "\x03\x00"
-     "\xE3\xB8\x01\x00\xE1\x70\x0F\x00\xE1\x98\x0F\x00\xE3\x00\x0E\x00\xD7\xB7"
-     "\x00\x80"
-     "\xDB\xB7\x00\x80\xDF\xB7\x00\x80\xE3\xB7\x00\x80\xE7\xB7\x00\x80\xEB\xB7"
-     "\x00\x80"
-     "\xEF\xB7\x00\x80\xF3\xB7\x00\x80\xEF\x64\x01\x00\xF7\xB7\x00\x80\xBF\x30"
-     "\x5D\x00"
-     "\xEF\xF4\x0E\x00\xFB\xB7\x00\x80\xFF\xB7\x00\x80\x84\x04\x05\x00\xBE\x04"
-     "\x05\x00"
-     "\x03\xB8\x00\x80\xE1\xA4\x01\x00\x07\xB8\x00\x80\xE3\x64\x00\x00\x0B\xB8"
-     "\x00\x80"
-     "\x0F\xB8\x00\x80\x13\xB8\x00\x80\xEF\xA0\x03\x00\x80\x15\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x17\xB8\x00\x80\xAA\xC9\x0E\x00\xAB\xC9\x0E\x00\xA8\x95"
-     "\x0E\x00"
-     "\xA9\xCD\x0E\x00\xAE\xF9\x0E\x00\xAF\xF9\x0E\x00\xAC\xCD\x0E\x00\xAD\xC5"
-     "\x0E\x00"
-     "\x86\x00\x04\x00\x87\x84\x04\x00\x1F\xB8\x00\x80\x23\xB8\x00\x80\x27\xB8"
-     "\x00\x80"
-     "\x2B\xB8\x00\x80\x2F\xB8\x00\x80\x33\xB8\x00\x80\xBA\xD1\x0D\x00\xBB\xDD"
-     "\x0D\x00"
-     "\xB8\xCD\x0D\x00\xB9\xC5\x0D\x00\xBE\xED\x0D\x00\xBF\x95\x0D\x00\xBC\xC5"
-     "\x0D\x00"
-     "\xBD\xE5\x0D\x00\xB2\x49\x0E\x00\xB3\x49\x0E\x00\xB0\x49\x0E\x00\xB1\x49"
-     "\x0E\x00"
-     "\xB6\x49\x0E\x00\xB7\x49\x0E\x00\xB4\x49\x0E\x00\xB5\x49\x0E\x00\xA8\x39"
-     "\x06\x00"
-     "\xA9\x39\x06\x00\xAA\x9D\x06\x00\xAB\x95\x06\x00\xAC\xD5\x06\x00\xAD\xDD"
-     "\x06\x00"
-     "\xAE\xC9\x06\x00\xAF\xC1\x06\x00\x1B\xB8\x00\x80\x37\xB8\x00\x80\x3B\xB8"
-     "\x00\x80"
-     "\x3F\xB8\x00\x80\x43\xB8\x00\x80\x47\xB8\x00\x80\x4B\xB8\x00\x80\x4F\xB8"
-     "\x00\x80"
-     "\xB8\x69\x07\x00\xB9\x69\x07\x00\xBA\x7D\x07\x00\xBB\x0D\x07\x00\xBC\x15"
-     "\x07\x00"
-     "\xBD\x05\x07\x00\xBE\x01\x07\x00\xBF\x3D\x07\x00\xB0\x89\x06\x00\xB1\x89"
-     "\x06\x00"
-     "\xB2\x79\x07\x00\xB3\x79\x07\x00\xB4\x69\x07\x00\xB5\x69\x07\x00\xB6\x59"
-     "\x07\x00"
-     "\xB7\x59\x07\x00\xB3\x15\x06\x00\x53\xB8\x00\x80\x57\xB8\x00\x80\x5B\xB8"
-     "\x00\x80"
-     "\x5F\xB8\x00\x80\xB6\x5D\x06\x00\xB5\x5D\x06\x00\x63\xB8\x00\x80\xBB\x4D"
-     "\x06\x00"
-     "\xBA\x4D\x06\x00\x67\xB8\x00\x80\x6B\xB8\x00\x80\xBF\xAD\x07\x00\xBE\xAD"
-     "\x07\x00"
-     "\xBD\xAD\x07\x00\xBC\xAD\x07\x00\x6F\xB8\x00\x80\xA3\x51\x06\x00\x73\xB8"
-     "\x00\x80"
-     "\x77\xB8\x00\x80\xA6\x19\x06\x00\x7B\xB8\x00\x80\x7F\xB8\x00\x80\xA5\x19"
-     "\x06\x00"
-     "\xAA\x09\x06\x00\xAB\x09\x06\x00\x83\xB8\x00\x80\x87\xB8\x00\x80\xAE\xE9"
-     "\x07\x00"
-     "\xAF\xE9\x07\x00\xAC\xE9\x07\x00\xAD\xE9\x07\x00\xA8\xA9\x06\x00\xA9\xA9"
-     "\x06\x00"
-     "\xAA\xB9\x06\x00\xAB\xB9\x06\x00\xAC\xA9\x06\x00\xAD\xA9\x06\x00\xAE\x3D"
-     "\x01\x00"
-     "\xAF\x35\x01\x00\x80\xF9\x01\x00\x81\xCD\x01\x00\x82\xC5\x01\x00\x84\x78"
-     "\x02\x00"
-     "\xBE\x7C\x02\x00\x8B\xB8\x00\x80\x87\x78\x01\x00\x86\x9C\x00\x00\xB8\xD1"
-     "\x01\x00"
-     "\xB9\xD9\x01\x00\xBA\xE1\x01\x00\xBB\xE1\x01\x00\xBC\x91\x01\x00\xBD\x9D"
-     "\x01\x00"
-     "\xBE\x95\x01\x00\xBF\x89\x01\x00\xB0\x4D\x01\x00\xB1\x55\x01\x00\xB2\x5D"
-     "\x01\x00"
-     "\xB3\x55\x01\x00\xB4\x4D\x01\x00\xB5\xF1\x01\x00\xB6\xF1\x01\x00\xB7\xF1"
-     "\x01\x00"
-     "\xB5\x39\x06\x00\x8F\xB8\x00\x80\x93\xB8\x00\x80\xB6\x31\x06\x00\x97\xB8"
-     "\x00\x80"
-     "\x9B\xB8\x00\x80\xB3\x29\x06\x00\x9F\xB8\x00\x80\xBD\x99\x01\x00\xBC\x91"
-     "\x01\x00"
-     "\xBF\x2D\x01\x00\xBE\x91\x01\x00\xA3\xB8\x00\x80\xAB\xB7\x00\x80\xBB\x89"
-     "\x01\x00"
-     "\xBA\x91\x01\x00\xA3\x71\x06\x00\xA7\xB8\x00\x80\xAB\xB8\x00\x80\xAF\xB8"
-     "\x00\x80"
-     "\xB3\xB8\x00\x80\xA6\x69\x06\x00\xA5\x61\x06\x00\xB7\xB8\x00\x80\xAB\xD1"
-     "\x01\x00"
-     "\xAA\xC9\x01\x00\xBB\xB8\x00\x80\xBF\xB8\x00\x80\xAF\x75\x01\x00\xAE\xC9"
-     "\x01\x00"
-     "\xAD\xC1\x01\x00\xAC\xC9\x01\x00\xC3\xB8\x00\x80\xC7\xB8\x00\x80\xEF\xBC"
-     "\x07\x00"
-     "\xCB\xB8\x00\x80\xCF\xB8\x00\x80\xD3\xB8\x00\x80\xD7\xB8\x00\x80\xDB\xB8"
-     "\x00\x80"
-     "\xDF\xB8\x00\x80\xE3\xB8\x00\x80\xE7\xB8\x00\x80\xEB\xB8\x00\x80\xE1\x88"
-     "\x06\x00"
-     "\xEF\xB8\x00\x80\xE3\x60\x07\x00\xF3\xB8\x00\x80\x80\x75\x00\x00\x81\x79"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x84\x40\x02\x00\xF7\xB8\x00\x80\xFF\xB8\x00\x80\xBE\x14"
-     "\x03\x00"
-     "\x03\xB9\x00\x80\x86\xC0\x3C\x00\x87\x44\x03\x00\x07\xB9\x00\x80\x0B\xB9"
-     "\x00\x80"
-     "\x0F\xB9\x00\x80\x13\xB9\x00\x80\x17\xB9\x00\x80\x1B\xB9\x00\x80\x1F\xB9"
-     "\x00\x80"
-     "\x23\xB9\x00\x80\xEF\x28\x02\x00\x27\xB9\x00\x80\x2B\xB9\x00\x80\x2F\xB9"
-     "\x00\x80"
-     "\x33\xB9\x00\x80\x37\xB9\x00\x80\x84\xEC\x3C\x00\x3B\xB9\x00\x80\x3F\xB9"
-     "\x00\x80"
-     "\x43\xB9\x00\x80\xE1\x94\x01\x00\x47\xB9\x00\x80\xE3\xE4\x02\x00\x4B\xB9"
-     "\x00\x80"
-     "\xB3\x51\x03\x00\x4F\xB9\x00\x80\x53\xB9\x00\x80\x57\xB9\x00\x80\x5B\xB9"
-     "\x00\x80"
-     "\xB6\x65\x03\x00\xB5\x45\x03\x00\x5F\xB9\x00\x80\xBB\x31\x03\x00\xBA\x29"
-     "\x03\x00"
-     "\x63\xB9\x00\x80\x67\xB9\x00\x80\xBF\xF1\x00\x00\xBE\x19\x03\x00\xBD\x11"
-     "\x03\x00"
-     "\xBC\x21\x03\x00\xA8\x85\x02\x00\xA9\x8D\x02\x00\xAA\x99\x02\x00\xAB\x99"
-     "\x02\x00"
-     "\xAC\x8D\x02\x00\xAD\xA5\x02\x00\xAE\xDD\x02\x00\xAF\xD5\x02\x00\x80\xFD"
-     "\x01\x00"
-     "\x81\xCD\x01\x00\x82\xC5\x01\x00\xBE\x64\x3D\x00\x86\x90\x3C\x00\x87\xB8"
-     "\x3D\x00"
-     "\x6F\xB9\x00\x80\x84\x60\x3C\x00\xB8\x51\x01\x00\xB9\x51\x01\x00\xBA\x51"
-     "\x01\x00"
-     "\xBB\x51\x01\x00\xBC\x71\x01\x00\xBD\x71\x01\x00\xBE\x71\x01\x00\xBF\x71"
-     "\x01\x00"
-     "\xB0\xAD\x02\x00\xB1\xB5\x02\x00\xB2\xB5\x02\x00\xB3\x85\x02\x00\xB4\x9D"
-     "\x02\x00"
-     "\xB5\x75\x01\x00\xB6\x7D\x01\x00\xB7\x71\x01\x00\x73\xB9\x00\x80\x77\xB9"
-     "\x00\x80"
-     "\x7B\xB9\x00\x80\x7F\xB9\x00\x80\x83\xB9\x00\x80\x87\xB9\x00\x80\x8B\xB9"
-     "\x00\x80"
-     "\xEF\x54\x00\x00\xBE\xEC\x3C\x00\xE1\x1C\x06\x00\x8F\xB9\x00\x80\xE3\x90"
-     "\x00\x00"
-     "\x93\xB9\x00\x80\x97\xB9\x00\x80\x9B\xB9\x00\x80\x9F\xB9\x00\x80\xA3\x51"
-     "\x02\x00"
-     "\xA3\xB9\x00\x80\x84\x44\x3F\x00\xA7\xB9\x00\x80\xAB\xB9\x00\x80\xA6\x65"
-     "\x02\x00"
-     "\xA5\x45\x02\x00\xAF\xB9\x00\x80\xAB\x31\x02\x00\xAA\x29\x02\x00\xB3\xB9"
-     "\x00\x80"
-     "\xB7\xB9\x00\x80\xAF\xF1\x01\x00\xAE\x19\x02\x00\xAD\x11\x02\x00\xAC\x21"
-     "\x02\x00"
-     "\xA8\x55\x3E\x00\xA9\x59\x3E\x00\xAA\x61\x3E\x00\xAB\x61\x3E\x00\xAC\x61"
-     "\x3E\x00"
-     "\xAD\x61\x3E\x00\xAE\x61\x3E\x00\xAF\x61\x3E\x00\x6B\xB9\x00\x80\xBB\xB9"
-     "\x00\x80"
-     "\xBF\xB9\x00\x80\xC3\xB9\x00\x80\xC7\xB9\x00\x80\xCB\xB9\x00\x80\xCF\xB9"
-     "\x00\x80"
-     "\xD3\xB9\x00\x80\xB8\xE9\x3E\x00\xB9\xE9\x3E\x00\xBA\xFD\x3E\x00\xBB\xF1"
-     "\x3E\x00"
-     "\xBC\x95\x3E\x00\xBD\x9D\x3E\x00\xBE\x95\x3E\x00\xBF\x8D\x3E\x00\xB0\xE1"
-     "\x3E\x00"
-     "\xB1\xE1\x3E\x00\xB2\xE5\x3E\x00\xB3\xFD\x3E\x00\xB4\xE5\x3E\x00\xB5\xED"
-     "\x3E\x00"
-     "\xB6\xE5\x3E\x00\xB7\xD9\x3E\x00\xB3\xE9\x3E\x00\x82\x15\x00\x00\x81\xB5"
-     "\x00\x00"
-     "\x80\xB5\x00\x00\xD7\xB9\x00\x80\xB6\xE1\x3E\x00\xB5\xE9\x3E\x00\xDB\xB9"
-     "\x00\x80"
-     "\xBB\x25\x3E\x00\xBA\x25\x3E\x00\x86\xA8\x00\x00\x87\xA4\x00\x00\xBF\x29"
-     "\x3E\x00"
-     "\xBE\x21\x3E\x00\xBD\x29\x3E\x00\xBC\x31\x3E\x00\xDF\xB9\x00\x80\xA3\xAD"
-     "\x3E\x00"
-     "\xE3\xB9\x00\x80\xE7\xB9\x00\x80\xA6\xA5\x3E\x00\xEB\xB9\x00\x80\xEF\xB9"
-     "\x00\x80"
-     "\xA5\xAD\x3E\x00\xAA\x61\x3E\x00\xAB\x61\x3E\x00\xF3\xB9\x00\x80\xF7\xB9"
-     "\x00\x80"
-     "\xAE\x65\x3E\x00\xAF\x6D\x3E\x00\xAC\x75\x3E\x00\xAD\x6D\x3E\x00\xA8\x39"
-     "\x3E\x00"
-     "\xA9\x85\x3E\x00\xAA\x8D\x3E\x00\xAB\x85\x3E\x00\xAC\x9D\x3E\x00\xAD\x85"
-     "\x3E\x00"
-     "\xAE\x8D\x3E\x00\xAF\x85\x3E\x00\xFB\xB9\x00\x80\xFF\xB9\x00\x80\x03\xBA"
-     "\x00\x80"
-     "\x07\xBA\x00\x80\x0B\xBA\x00\x80\x0F\xBA\x00\x80\x13\xBA\x00\x80\x17\xBA"
-     "\x00\x80"
-     "\xB8\x9D\x3E\x00\xB9\xAD\x3E\x00\xBA\xA5\x3E\x00\xBB\x69\x01\x00\xBC\x79"
-     "\x01\x00"
-     "\xBD\x79\x01\x00\xBE\x69\x01\x00\xBF\x69\x01\x00\xB0\xC9\x3E\x00\xB1\xD5"
-     "\x3E\x00"
-     "\xB2\xDD\x3E\x00\xB3\xD5\x3E\x00\xB4\xCD\x3E\x00\xB5\xB1\x3E\x00\xB6\xB5"
-     "\x3E\x00"
-     "\xB7\xAD\x3E\x00\x1B\xBA\x00\x80\xB6\x15\x3E\x00\xB5\x05\x3E\x00\x1F\xBA"
-     "\x00\x80"
-     "\xB3\x15\x3E\x00\x23\xBA\x00\x80\x27\xBA\x00\x80\x2B\xBA\x00\x80\xBF\x65"
-     "\x3E\x00"
-     "\xBE\x19\x3E\x00\xBD\x11\x3E\x00\xBC\x19\x3E\x00\xBB\x01\x3E\x00\xBA\x1D"
-     "\x3E\x00"
-     "\xFB\xB8\x00\x80\x2F\xBA\x00\x80\xA3\x45\x3E\x00\x33\xBA\x00\x80\x37\xBA"
-     "\x00\x80"
-     "\x3B\xBA\x00\x80\x3F\xBA\x00\x80\xA6\x45\x3E\x00\xA5\x55\x3E\x00\x43\xBA"
-     "\x00\x80"
-     "\xAB\x51\x3E\x00\xAA\x4D\x3E\x00\x47\xBA\x00\x80\xBE\x2C\x00\x00\xAF\x35"
-     "\x3E\x00"
-     "\xAE\x49\x3E\x00\xAD\x41\x3E\x00\xAC\x49\x3E\x00\x82\xE9\x03\x00\xEF\xD0"
-     "\x3F\x00"
-     "\x80\xD9\x03\x00\x81\xD1\x03\x00\x87\x50\x03\x00\x86\xB4\x04\x00\x4F\xBA"
-     "\x00\x80"
-     "\x53\xBA\x00\x80\x57\xBA\x00\x80\x5B\xBA\x00\x80\x5F\xBA\x00\x80\x63\xBA"
-     "\x00\x80"
-     "\x67\xBA\x00\x80\xE3\x84\x3E\x00\x6B\xBA\x00\x80\xE1\x44\x3E\x00\xB3\x8D"
-     "\x02\x00"
-     "\x6F\xBA\x00\x80\x84\x44\x03\x00\xBE\x40\x04\x00\x73\xBA\x00\x80\xB6\xBD"
-     "\x02\x00"
-     "\xB5\x81\x02\x00\x77\xBA\x00\x80\xBB\x49\x03\x00\xBA\x41\x03\x00\x7B\xBA"
-     "\x00\x80"
-     "\x7F\xBA\x00\x80\xBF\x49\x03\x00\xBE\x41\x03\x00\xBD\x49\x03\x00\xBC\x51"
-     "\x03\x00"
-     "\x83\xBA\x00\x80\x87\xBA\x00\x80\x8B\xBA\x00\x80\xBE\x80\x05\x00\x8F\xBA"
-     "\x00\x80"
-     "\x93\xBA\x00\x80\x97\xBA\x00\x80\x9B\xBA\x00\x80\xA8\x61\x02\x00\xA9\x6D"
-     "\x02\x00"
-     "\xAA\x65\x02\x00\xAB\x79\x02\x00\xAC\x69\x02\x00\xAD\x5D\x02\x00\xAE\x55"
-     "\x02\x00"
-     "\xAF\x9D\x03\x00\xB0\xE5\x03\x00\xB1\xED\x03\x00\xB2\xE5\x03\x00\xB3\xFD"
-     "\x03\x00"
-     "\xB4\xE5\x03\x00\xB5\xED\x03\x00\xB6\xE5\x03\x00\xB7\x5D\x03\x00\xB8\x6D"
-     "\x03\x00"
-     "\xB9\x71\x03\x00\xBA\x71\x03\x00\xBB\x71\x03\x00\xBC\x4D\x03\x00\xBD\x31"
-     "\x03\x00"
-     "\xBE\x31\x03\x00\xBF\x31\x03\x00\x9F\xBA\x00\x80\xA3\xBA\x00\x80\xA7\xBA"
-     "\x00\x80"
-     "\xAB\xBA\x00\x80\xE3\xC8\x3E\x00\x84\x1C\x05\x00\xE1\xA8\x01\x00\xAF\xBA"
-     "\x00\x80"
-     "\xEF\x34\x00\x00\xB3\xBA\x00\x80\xB7\xBA\x00\x80\xBF\xBA\x00\x80\xC3\xBA"
-     "\x00\x80"
-     "\xC7\xBA\x00\x80\xCB\xBA\x00\x80\xCF\xBA\x00\x80\x80\x5D\x00\x00\x81\x61"
-     "\x00\x00"
-     "\x82\x61\x00\x00\xA3\x81\x03\x00\xD3\xBA\x00\x80\xA5\x8D\x03\x00\xA6\xB1"
-     "\x03\x00"
-     "\x84\x98\x07\x00\x86\xC0\x04\x00\x87\x48\x05\x00\xAA\x4D\x02\x00\xAB\x45"
-     "\x02\x00"
-     "\xAC\x5D\x02\x00\xAD\x45\x02\x00\xAE\x4D\x02\x00\xAF\x45\x02\x00\xD7\xBA"
-     "\x00\x80"
-     "\xDB\xBA\x00\x80\xDF\xBA\x00\x80\xE3\xBA\x00\x80\xE7\xBA\x00\x80\xEB\xBA"
-     "\x00\x80"
-     "\xEF\xBA\x00\x80\xEF\xC4\x3E\x00\xBE\x20\x06\x00\xE1\x8C\x3E\x00\xF3\xBA"
-     "\x00\x80"
-     "\xE3\x80\x3E\x00\xF7\xBA\x00\x80\xFB\xBA\x00\x80\xFF\xBA\x00\x80\x03\xBB"
-     "\x00\x80"
-     "\xB3\xE1\x01\x00\x07\xBB\x00\x80\x0B\xBB\x00\x80\x0F\xBB\x00\x80\x13\xBB"
-     "\x00\x80"
-     "\xB6\xF9\x01\x00\xB5\xF1\x01\x00\x17\xBB\x00\x80\xBB\xB1\x01\x00\xBA\xA1"
-     "\x01\x00"
-     "\x1B\xBB\x00\x80\x1F\xBB\x00\x80\xBF\x71\x01\x00\xBE\x99\x01\x00\xBD\x91"
-     "\x01\x00"
-     "\xBC\xA1\x01\x00\xA8\x19\x06\x00\xA9\x19\x06\x00\xAA\x29\x06\x00\xAB\x29"
-     "\x06\x00"
-     "\xAC\x39\x06\x00\xAD\x39\x06\x00\xAE\x29\x06\x00\xAF\x25\x06\x00\xBB\xBA"
-     "\x00\x80"
-     "\x82\x11\x00\x00\x81\x19\x00\x00\x80\xC1\x07\x00\x23\xBB\x00\x80\x27\xBB"
-     "\x00\x80"
-     "\x2B\xBB\x00\x80\x2F\xBB\x00\x80\xB8\xD1\x06\x00\xB9\xD1\x06\x00\xBA\xD1"
-     "\x06\x00"
-     "\xBB\xD1\x06\x00\xBC\x51\x07\x00\xBD\x51\x07\x00\xBE\x51\x07\x00\xBF\x51"
-     "\x07\x00"
-     "\xB0\x25\x06\x00\xB1\x35\x06\x00\xB2\x35\x06\x00\xB3\x05\x06\x00\xB4\x1D"
-     "\x06\x00"
-     "\xB5\xF5\x06\x00\xB6\xFD\x06\x00\xB7\xF5\x06\x00\xA3\xA1\x06\x00\x33\xBB"
-     "\x00\x80"
-     "\x86\xAC\x00\x00\x87\x4C\x03\x00\x37\xBB\x00\x80\xA6\xB9\x06\x00\xA5\xB1"
-     "\x06\x00"
-     "\x3B\xBB\x00\x80\xAB\xF1\x06\x00\xAA\xE1\x06\x00\x3F\xBB\x00\x80\x43\xBB"
-     "\x00\x80"
-     "\xAF\x31\x06\x00\xAE\xD9\x06\x00\xAD\xD1\x06\x00\xAC\xE1\x06\x00\x47\xBB"
-     "\x00\x80"
-     "\xB3\xF1\x07\x00\x4B\xBB\x00\x80\x4F\xBB\x00\x80\xB6\x81\x07\x00\x53\xBB"
-     "\x00\x80"
-     "\x57\xBB\x00\x80\xB5\xE1\x07\x00\xBA\x69\x07\x00\xBB\x69\x07\x00\x5B\xBB"
-     "\x00\x80"
-     "\x5F\xBB\x00\x80\xBE\x69\x07\x00\xBF\x5D\x07\x00\xBC\x69\x07\x00\xBD\x69"
-     "\x07\x00"
-     "\xA8\x2D\x06\x00\xA9\x35\x06\x00\xAA\x3D\x06\x00\xAB\x35\x06\x00\xAC\x2D"
-     "\x06\x00"
-     "\xAD\x45\x06\x00\xAE\x99\x06\x00\xAF\x99\x06\x00\x63\xBB\x00\x80\x67\xBB"
-     "\x00\x80"
-     "\x6B\xBB\x00\x80\x6F\xBB\x00\x80\x73\xBB\x00\x80\x77\xBB\x00\x80\x7B\xBB"
-     "\x00\x80"
-     "\x7F\xBB\x00\x80\xB8\xF9\x06\x00\xB9\xF9\x06\x00\xBA\x8D\x06\x00\xBB\x85"
-     "\x06\x00"
-     "\xBC\x9D\x06\x00\xBD\x85\x06\x00\xBE\x85\x06\x00\xBF\xB5\x06\x00\xB0\xE9"
-     "\x06\x00"
-     "\xB1\xE9\x06\x00\xB2\xF9\x06\x00\xB3\xF9\x06\x00\xB4\xE9\x06\x00\xB5\xDD"
-     "\x06\x00"
-     "\xB6\xC9\x06\x00\xB7\xC9\x06\x00\xA3\xB5\x06\x00\x83\xBB\x00\x80\x87\xBB"
-     "\x00\x80"
-     "\x8B\xBB\x00\x80\x8F\xBB\x00\x80\xA6\xC5\x06\x00\xA5\xA5\x06\x00\x93\xBB"
-     "\x00\x80"
-     "\xAB\x2D\x06\x00\xAA\x2D\x06\x00\x97\xBB\x00\x80\x9F\xBB\x00\x80\xAF\x19"
-     "\x06\x00"
-     "\xAE\x2D\x06\x00\xAD\x2D\x06\x00\xAC\x2D\x06\x00\x80\x51\x00\x00\x81\x59"
-     "\x00\x00"
-     "\x82\x61\x00\x00\x4B\xBA\x00\x80\x9B\xBB\x00\x80\xA3\xBB\x00\x80\x87\x78"
-     "\x01\x00"
-     "\x86\x00\x0C\x00\xA7\xBB\x00\x80\xAB\xBB\x00\x80\xAF\xBB\x00\x80\xB3\xBB"
-     "\x00\x80"
-     "\xB7\xBB\x00\x80\xBB\xBB\x00\x80\xBF\xBB\x00\x80\xC3\xBB\x00\x80\xC7\xBB"
-     "\x00\x80"
-     "\xCB\xBB\x00\x80\xB3\x99\x02\x00\x84\x94\x03\x00\xB5\xB9\x02\x00\xCF\xBB"
-     "\x00\x80"
-     "\xD3\xBB\x00\x80\xB6\xB1\x02\x00\xD7\xBB\x00\x80\xDB\xBB\x00\x80\xBB\xFD"
-     "\x02\x00"
-     "\xBA\xE1\x02\x00\xBD\xD9\x02\x00\xBC\xED\x02\x00\xBF\x19\x03\x00\xBE\xD1"
-     "\x02\x00"
-     "\xE3\x18\x06\x00\xDF\xBB\x00\x80\xE1\x44\x06\x00\xE3\xBB\x00\x80\xE7\xBB"
-     "\x00\x80"
-     "\xE1\x90\x01\x00\xEB\xBB\x00\x80\xE3\x40\x02\x00\xEF\xBB\x00\x80\xF3\xBB"
-     "\x00\x80"
-     "\xF7\xBB\x00\x80\xEF\x24\x02\x00\xEF\xC4\x06\x00\xFB\xBB\x00\x80\xFF\xBB"
-     "\x00\x80"
-     "\x03\xBC\x00\x80\xA9\x05\x02\x00\xA8\xB9\x02\x00\xAB\x15\x02\x00\xAA\x1D"
-     "\x02\x00"
-     "\xAD\x3D\x02\x00\xAC\x3D\x02\x00\xAF\x51\x02\x00\xAE\x5D\x02\x00\x80\xD9"
-     "\x00\x00"
-     "\x81\xD9\x00\x00\x82\x05\x00\x00\x07\xBC\x00\x80\xBE\xFC\x03\x00\x0F\xBC"
-     "\x00\x80"
-     "\x87\xA8\x0C\x00\x86\x9C\x0D\x00\xB9\xF1\x03\x00\xB8\xE9\x03\x00\xBB\xF1"
-     "\x03\x00"
-     "\xBA\xF9\x03\x00\xBD\x55\x03\x00\xBC\xE9\x03\x00\xBF\x55\x03\x00\xBE\x55"
-     "\x03\x00"
-     "\xB1\x35\x02\x00\xB0\x29\x02\x00\xB3\x01\x02\x00\xB2\x35\x02\x00\xB5\xD1"
-     "\x03\x00"
-     "\xB4\x19\x02\x00\xB7\xD1\x03\x00\xB6\xD9\x03\x00\x85\x44\x0D\x00\x13\xBC"
-     "\x00\x80"
-     "\x17\xBC\x00\x80\xA3\x9D\x03\x00\x1B\xBC\x00\x80\xA5\xBD\x03\x00\xA6\xB5"
-     "\x03\x00"
-     "\x1F\xBC\x00\x80\x23\xBC\x00\x80\x27\xBC\x00\x80\xAA\xE5\x03\x00\xAB\xF9"
-     "\x03\x00"
-     "\xAC\xE9\x03\x00\xAD\xDD\x03\x00\xAE\xD5\x03\x00\xAF\x1D\x02\x00\xB6\xAD"
-     "\x02\x00"
-     "\x2B\xBC\x00\x80\x2F\xBC\x00\x80\xB5\xA5\x02\x00\x33\xBC\x00\x80\xB3\xB5"
-     "\x02\x00"
-     "\x37\xBC\x00\x80\x3B\xBC\x00\x80\xBE\x75\x01\x00\xBF\x7D\x01\x00\xBC\x69"
-     "\x01\x00"
-     "\xBD\x69\x01\x00\xBA\x89\x02\x00\xBB\x89\x02\x00\x3F\xBC\x00\x80\x43\xBC"
-     "\x00\x80"
-     "\xE1\x94\x02\x00\xEF\xDC\x06\x00\xE3\x3C\x12\x00\x47\xBC\x00\x80\x4B\xBC"
-     "\x00\x80"
-     "\x56\xBC\x00\x80\x5A\xBC\x00\x80\xEF\xE8\x07\x00\x84\x6C\x0F\x00\xE1\xDC"
-     "\x06\x00"
-     "\x5E\xBC\x00\x80\xE3\xA8\x06\x00\x62\xBC\x00\x80\xE3\xF8\x06\x00\xEF\x7C"
-     "\x3E\x00"
-     "\xE1\x2C\x01\x00\xA3\x21\x02\x00\xF1\x00\x00\x00\x66\xBC\x00\x80\x6A\xBC"
-     "\x00\x80"
-     "\x6E\xBC\x00\x80\xA6\x39\x02\x00\xA5\x31\x02\x00\x72\xBC\x00\x80\xAB\x1D"
-     "\x02\x00"
-     "\xAA\x1D\x02\x00\x76\xBC\x00\x80\x7A\xBC\x00\x80\xAF\xE9\x01\x00\xAE\xE1"
-     "\x01\x00"
-     "\xAD\xFD\x01\x00\xAC\xFD\x01\x00\x8F\x21\x24\x00\x0B\xBC\x00\x80\x7E\xBC"
-     "\x00\x80"
-     "\xBE\xC4\x00\x00\x80\xF9\x0F\x00\x81\xF9\x0F\x00\x82\x11\x00\x00\x82\xBC"
-     "\x00\x80"
-     "\x86\x00\x00\x00\x87\x84\x03\x00\x86\xBC\x00\x80\x8A\xBC\x00\x80\x8E\xBC"
-     "\x00\x80"
-     "\x92\xBC\x00\x80\x96\xBC\x00\x80\x9A\xBC\x00\x80\x9F\xDD\x08\x00\x9E\x65"
-     "\x08\x00"
-     "\x9D\xF1\x09\x00\x9C\x79\x35\x00\x9B\x71\x34\x00\x9A\xE1\x35\x00\x99\x89"
-     "\x30\x00"
-     "\x98\xD1\x30\x00\x97\xE1\x30\x00\x96\x75\x3D\x00\x95\xED\x3C\x00\x94\xED"
-     "\x3C\x00"
-     "\x93\x19\x39\x00\x92\xF1\x38\x00\x91\xF9\x38\x00\x90\x99\x26\x00\xB1\xA5"
-     "\x14\x00"
-     "\xB0\xA5\x14\x00\xB3\x09\x68\x00\xB2\xE9\x15\x00\xB5\x05\x68\x00\xB4\x19"
-     "\x68\x00"
-     "\x9E\xBC\x00\x80\xA2\xBC\x00\x80\xA6\xBC\x00\x80\xAA\xBC\x00\x80\xAE\xBC"
-     "\x00\x80"
-     "\xB2\xBC\x00\x80\xA3\x61\x0D\x00\xA2\x8D\x0C\x00\xA1\x99\x0C\x00\xA0\x71"
-     "\x0D\x00"
-     "\xA1\xC5\x00\x00\xB6\xBC\x00\x80\xA3\xC9\x04\x00\xA2\xD1\x00\x00\xA5\x39"
-     "\x04\x00"
-     "\xA4\xD1\x04\x00\xA7\xF9\x18\x00\xA6\x01\x18\x00\xA9\x01\x1C\x00\xA8\x79"
-     "\x18\x00"
-     "\xAB\x6D\x1C\x00\xAA\xCD\x1C\x00\xAD\xD1\x10\x00\xAC\x01\x10\x00\xAF\x01"
-     "\x14\x00"
-     "\xAE\xD9\x10\x00\x82\x21\x25\x00\x83\xE1\x27\x00\xBA\xBC\x00\x80\xBE\xBC"
-     "\x00\x80"
-     "\x86\x79\x38\x00\x87\x95\x3D\x00\x84\x99\x39\x00\x85\x99\x39\x00\x8A\xA5"
-     "\x3D\x00"
-     "\x8B\x35\x3C\x00\xC2\xBC\x00\x80\xC6\xBC\x00\x80\x8E\xE1\x31\x00\x8F\x81"
-     "\x34\x00"
-     "\x8C\xB9\x31\x00\x8D\xB1\x31\x00\x92\x61\x35\x00\x93\x71\x34\x00\xCA\xBC"
-     "\x00\x80"
-     "\xCE\xBC\x00\x80\x96\xCD\x09\x00\x97\x99\x0C\x00\x94\xE1\x08\x00\x95\x7D"
-     "\x09\x00"
-     "\x9A\x41\x0D\x00\x9B\x5D\x0D\x00\xD2\xBC\x00\x80\xD6\xBC\x00\x80\xDA\xBC"
-     "\x00\x80"
-     "\xDE\xBC\x00\x80\x9C\xA9\x0C\x00\xE2\xBC\x00\x80\xA8\x35\x0F\x00\xA9\x81"
-     "\x0E\x00"
-     "\xAA\x81\x0E\x00\xAB\x9D\x0E\x00\xAC\x8D\x0E\x00\xAD\xB5\x0E\x00\xAE\xBD"
-     "\x0E\x00"
-     "\xAF\xD9\x0E\x00\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\xE6\xBC"
-     "\x00\x80"
-     "\xEA\xBC\x00\x80\xEE\xBC\x00\x80\x84\xBC\x00\x00\xBE\x00\x04\x00\xB8\x55"
-     "\x0E\x00"
-     "\xB9\x5D\x0E\x00\xBA\x61\x0E\x00\xBB\x7D\x0E\x00\xBC\x19\x0E\x00\xBD\x19"
-     "\x0E\x00"
-     "\xBE\x09\x0E\x00\xBF\x01\x0E\x00\xB0\xA9\x0E\x00\xB1\xB5\x0E\x00\xB2\xB1"
-     "\x0E\x00"
-     "\xB3\xB1\x0E\x00\xB4\x89\x0E\x00\xB5\x89\x0E\x00\xB6\x75\x0E\x00\xB7\x6D"
-     "\x0E\x00"
-     "\xB3\x09\x0D\x00\xF6\xBC\x00\x80\x86\x88\x04\x00\x87\xC8\x03\x00\x84\x5C"
-     "\x04\x00"
-     "\xB6\x35\x0D\x00\xB5\x05\x0D\x00\xFA\xBC\x00\x80\xBB\x71\x0D\x00\xBA\x6D"
-     "\x0D\x00"
-     "\xFE\xBC\x00\x80\x02\xBD\x00\x80\xBF\x4D\x0D\x00\xBE\x4D\x0D\x00\xBD\x65"
-     "\x0D\x00"
-     "\xBC\x65\x0D\x00\x06\xBD\x00\x80\xA3\x4D\x0D\x00\x0A\xBD\x00\x80\x0E\xBD"
-     "\x00\x80"
-     "\xA6\x71\x0D\x00\x12\xBD\x00\x80\x16\xBD\x00\x80\xA5\x41\x0D\x00\xAA\x29"
-     "\x0D\x00"
-     "\xAB\x35\x0D\x00\x1A\xBD\x00\x80\x1E\xBD\x00\x80\xAE\x09\x0D\x00\xAF\x09"
-     "\x0D\x00"
-     "\xAC\x21\x0D\x00\xAD\x21\x0D\x00\xA8\xA9\x02\x00\xA9\xA9\x02\x00\xAA\xB9"
-     "\x02\x00"
-     "\xAB\xB9\x02\x00\xAC\xA9\x02\x00\xAD\xA9\x02\x00\xAE\x01\x02\x00\xAF\x01"
-     "\x02\x00"
-     "\x22\xBD\x00\x80\x26\xBD\x00\x80\x2A\xBD\x00\x80\x2E\xBD\x00\x80\x32\xBD"
-     "\x00\x80"
-     "\x36\xBD\x00\x80\x3A\xBD\x00\x80\x3E\xBD\x00\x80\xB8\xED\x03\x00\xB9\x8D"
-     "\x03\x00"
-     "\xBA\x85\x03\x00\xBB\xA5\x03\x00\xBC\xBD\x03\x00\xBD\xA5\x03\x00\xBE\xAD"
-     "\x03\x00"
-     "\xBF\xA1\x03\x00\xB0\x41\x02\x00\xB1\x41\x02\x00\xB2\x41\x02\x00\xB3\x41"
-     "\x02\x00"
-     "\xB4\x35\x02\x00\xB5\xFD\x03\x00\xB6\xF5\x03\x00\xB7\xD5\x03\x00\xB7\x25"
-     "\x0D\x00"
-     "\xB6\x25\x0D\x00\xB5\x11\x0D\x00\xB4\x19\x0D\x00\xB3\x15\x0D\x00\xB2\x15"
-     "\x0D\x00"
-     "\xB1\xE5\x0E\x00\xB0\xE5\x0E\x00\xBF\x55\x0D\x00\xBE\x4D\x0D\x00\xBD\x4D"
-     "\x0D\x00"
-     "\xBC\x4D\x0D\x00\xBB\x69\x0D\x00\xBA\x65\x0D\x00\xB9\x75\x0D\x00\xB8\x79"
-     "\x0D\x00"
-     "\x42\xBD\x00\x80\xBE\x0C\x05\x00\x46\xBD\x00\x80\x4A\xBD\x00\x80\x4E\xBD"
-     "\x00\x80"
-     "\x52\xBD\x00\x80\x56\xBD\x00\x80\x5A\xBD\x00\x80\xAF\x95\x0E\x00\xAE\x95"
-     "\x0E\x00"
-     "\xAD\x85\x0E\x00\xAC\x85\x0E\x00\xAB\x95\x0E\x00\xAA\x95\x0E\x00\xA9\xAD"
-     "\x0E\x00"
-     "\xA8\xD1\x0E\x00\xE1\x24\x0F\x00\x5E\xBD\x00\x80\xE3\x38\x0E\x00\x66\xBD"
-     "\x00\x80"
-     "\xE3\x38\x00\x00\x6A\xBD\x00\x80\xE1\xA0\x01\x00\x6E\xBD\x00\x80\xEF\xE4"
-     "\x00\x00"
-     "\x72\xBD\x00\x80\x84\xE4\x04\x00\x76\xBD\x00\x80\x86\x10\x04\x00\x87\x94"
-     "\x04\x00"
-     "\xEF\x4C\x01\x00\x7A\xBD\x00\x80\x81\x9D\x00\x00\x80\x9D\x00\x00\xB3\xF9"
-     "\x03\x00"
-     "\x82\x49\x00\x00\xB5\xE9\x03\x00\x7E\xBD\x00\x80\x82\xBD\x00\x80\xB6\xE1"
-     "\x03\x00"
-     "\x86\xBD\x00\x80\x8A\xBD\x00\x80\xBB\x4D\x03\x00\xBA\x4D\x03\x00\xBD\x5D"
-     "\x03\x00"
-     "\xBC\x5D\x03\x00\xBF\xF1\x02\x00\xBE\x39\x03\x00\xF2\xBC\x00\x80\x62\xBD"
-     "\x00\x80"
-     "\x8E\xBD\x00\x80\x92\xBD\x00\x80\x96\xBD\x00\x80\x9A\xBD\x00\x80\x9E\xBD"
-     "\x00\x80"
-     "\xA2\xBD\x00\x80\xA8\x25\x06\x00\xA9\x39\x06\x00\xAA\x01\x06\x00\xAB\x01"
-     "\x06\x00"
-     "\xAC\x71\x06\x00\xAD\x7D\x06\x00\xAE\x75\x06\x00\xAF\x69\x06\x00\xB0\x19"
-     "\x06\x00"
-     "\xB1\x2D\x06\x00\xB2\x25\x06\x00\xB3\xED\x06\x00\xB4\xF5\x06\x00\xB5\xFD"
-     "\x06\x00"
-     "\xB6\xF5\x06\x00\xB7\xED\x06\x00\xB8\xD5\x06\x00\xB9\xDD\x06\x00\xBA\xD5"
-     "\x06\x00"
-     "\xBB\x2D\x07\x00\xBC\x39\x07\x00\xBD\x39\x07\x00\xBE\x29\x07\x00\xBF\x29"
-     "\x07\x00"
-     "\xB3\xA9\x06\x00\xA6\xBD\x00\x80\xAA\xBD\x00\x80\xAE\xBD\x00\x80\xB2\xBD"
-     "\x00\x80"
-     "\xB6\xDD\x06\x00\xB5\xD9\x06\x00\xB6\xBD\x00\x80\xBB\xC9\x06\x00\xBA\xFD"
-     "\x06\x00"
-     "\xBA\xBD\x00\x80\xBE\xBD\x00\x80\xBF\x19\x06\x00\xBE\x11\x06\x00\xBD\x19"
-     "\x06\x00"
-     "\xBC\xD1\x06\x00\xC2\xBD\x00\x80\xA3\xED\x06\x00\xC6\xBD\x00\x80\xCA\xBD"
-     "\x00\x80"
-     "\xA6\x99\x06\x00\xCE\xBD\x00\x80\xBE\x90\x03\x00\xA5\x9D\x06\x00\xAA\xB9"
-     "\x06\x00"
-     "\xAB\x8D\x06\x00\xD2\xBD\x00\x80\xD6\xBD\x00\x80\xAE\x55\x06\x00\xAF\x5D"
-     "\x06\x00"
-     "\xAC\x95\x06\x00\xAD\x5D\x06\x00\xA8\x8D\x06\x00\xA9\x91\x06\x00\xAA\x95"
-     "\x06\x00"
-     "\xAB\xA9\x06\x00\xAC\xBD\x06\x00\xAD\xA9\x06\x00\xAE\xCD\x06\x00\xAF\xC5"
-     "\x06\x00"
-     "\x80\xE9\x01\x00\x81\xE9\x01\x00\x82\xFD\x01\x00\xDA\xBD\x00\x80\x86\x90"
-     "\x01\x00"
-     "\x87\xB4\x01\x00\xDE\xBD\x00\x80\xE2\xBD\x00\x80\xB8\xC5\x01\x00\xB9\xCD"
-     "\x01\x00"
-     "\xBA\xC5\x01\x00\xBB\xDD\x01\x00\xBC\xCD\x01\x00\xBD\xFD\x01\x00\xBE\xF5"
-     "\x01\x00"
-     "\xBF\x9D\x01\x00\xB0\x81\x06\x00\xB1\x81\x06\x00\xB2\x61\x01\x00\xB3\x61"
-     "\x01\x00"
-     "\xB4\x61\x01\x00\xB5\x61\x01\x00\xB6\x61\x01\x00\xB7\x61\x01\x00\xB3\x1D"
-     "\x06\x00"
-     "\xE6\xBD\x00\x80\xEA\xBD\x00\x80\xEE\xBD\x00\x80\xF2\xBD\x00\x80\xB6\x29"
-     "\x06\x00"
-     "\xB5\x35\x06\x00\xF6\xBD\x00\x80\xBB\x45\x06\x00\xBA\x45\x06\x00\xFA\xBD"
-     "\x00\x80"
-     "\xFE\xBD\x00\x80\xBF\xA5\x01\x00\xBE\xA5\x01\x00\xBD\xB5\x01\x00\xBC\xB5"
-     "\x01\x00"
-     "\x02\xBE\x00\x80\xA3\x59\x06\x00\x06\xBE\x00\x80\x0A\xBE\x00\x80\xA6\x6D"
-     "\x06\x00"
-     "\x0E\xBE\x00\x80\x12\xBE\x00\x80\xA5\x71\x06\x00\xAA\x01\x06\x00\xAB\x01"
-     "\x06\x00"
-     "\x16\xBE\x00\x80\x1A\xBE\x00\x80\xAE\xE1\x01\x00\xAF\xE1\x01\x00\xAC\xF1"
-     "\x01\x00"
-     "\xAD\xF1\x01\x00\xA8\xED\x02\x00\xA9\x29\x03\x00\xAA\x31\x03\x00\xAB\x31"
-     "\x03\x00"
-     "\xAC\x95\x03\x00\xAD\x9D\x03\x00\xAE\x95\x03\x00\xAF\x8D\x03\x00\x84\x0C"
-     "\x02\x00"
-     "\x1E\xBE\x00\x80\x22\xBE\x00\x80\x26\xBE\x00\x80\x2A\xBE\x00\x80\x2E\xBE"
-     "\x00\x80"
-     "\x32\xBE\x00\x80\x36\xBE\x00\x80\xB8\xA9\x03\x00\xB9\xA9\x03\x00\xBA\x6D"
-     "\x00\x00"
-     "\xBB\x65\x00\x00\xBC\x7D\x00\x00\xBD\x65\x00\x00\xBE\x6D\x00\x00\xBF\x65"
-     "\x00\x00"
-     "\xB0\xFD\x03\x00\xB1\xCD\x03\x00\xB2\xC5\x03\x00\xB3\xAD\x03\x00\xB4\xB9"
-     "\x03\x00"
-     "\xB5\xB9\x03\x00\xB6\xAD\x03\x00\xB7\xA5\x03\x00\x80\xBD\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x3A\xBE\x00\x80\x3E\xBE\x00\x80\xBE\xD8\x03\x00\x46\xBE"
-     "\x00\x80"
-     "\x4A\xBE\x00\x80\x86\x00\x1C\x00\x87\xD4\x03\x00\x4E\xBE\x00\x80\x52\xBE"
-     "\x00\x80"
-     "\x56\xBE\x00\x80\x5A\xBE\x00\x80\x5E\xBE\x00\x80\x62\xBE\x00\x80\xBB\x69"
-     "\x03\x00"
-     "\xBA\x59\x03\x00\x66\xBE\x00\x80\x6A\xBE\x00\x80\xBF\x1D\x03\x00\xBE\x1D"
-     "\x03\x00"
-     "\xBD\x19\x03\x00\xBC\x71\x03\x00\xB3\xD9\x03\x00\x6E\xBE\x00\x80\x72\xBE"
-     "\x00\x80"
-     "\x76\xBE\x00\x80\x7A\xBE\x00\x80\xB6\x7D\x03\x00\xB5\x7D\x03\x00\x7E\xBE"
-     "\x00\x80"
-     "\xAF\x09\x06\x00\xAE\xC1\x07\x00\xAD\xA5\x07\x00\xAC\xA5\x07\x00\xAB\xB5"
-     "\x07\x00"
-     "\xAA\xB5\x07\x00\x82\xBE\x00\x80\x86\xBE\x00\x80\x84\x6C\x1D\x00\xA6\x19"
-     "\x07\x00"
-     "\xA5\x11\x07\x00\x8A\xBE\x00\x80\xA3\x01\x07\x00\x8E\xBE\x00\x80\x92\xBE"
-     "\x00\x80"
-     "\x96\xBE\x00\x80\x9A\xBE\x00\x80\x9E\xBE\x00\x80\xBE\x08\x1D\x00\x84\x00"
-     "\x1C\x00"
-     "\xA2\xBE\x00\x80\xEF\x38\x03\x00\xA6\xBE\x00\x80\xAA\xBE\x00\x80\xAE\xBE"
-     "\x00\x80"
-     "\xE3\x54\x03\x00\xB6\xBE\x00\x80\xE1\x44\x00\x00\x80\x29\x00\x00\x81\x29"
-     "\x00\x00"
-     "\x82\x19\x00\x00\xBA\xBE\x00\x80\xE1\x28\x06\x00\xE1\x94\x07\x00\xE3\x30"
-     "\x06\x00"
-     "\xE3\x8C\x07\x00\xBE\xBE\x00\x80\xC2\xBE\x00\x80\x86\x38\x1C\x00\x87\x5C"
-     "\x1D\x00"
-     "\xBE\xE8\x1C\x00\xC6\xBE\x00\x80\xCA\xBE\x00\x80\xCE\xBE\x00\x80\xD2\xBE"
-     "\x00\x80"
-     "\xD6\xBE\x00\x80\xEF\x98\x01\x00\xEF\x00\x06\x00\xA3\xF9\x02\x00\xDA\xBE"
-     "\x00\x80"
-     "\xDE\xBE\x00\x80\xE2\xBE\x00\x80\xE6\xBE\x00\x80\xA6\x5D\x02\x00\xA5\x5D"
-     "\x02\x00"
-     "\xEA\xBE\x00\x80\xAB\x49\x02\x00\xAA\x79\x02\x00\xEE\xBE\x00\x80\xF2\xBE"
-     "\x00\x80"
-     "\xAF\x3D\x02\x00\xAE\x3D\x02\x00\xAD\x39\x02\x00\xAC\x51\x02\x00\xA8\x99"
-     "\x1E\x00"
-     "\xA9\x99\x1E\x00\xAA\xA1\x1E\x00\xAB\xA1\x1E\x00\xAC\xA1\x1E\x00\xAD\xA1"
-     "\x1E\x00"
-     "\xAE\xA1\x1E\x00\xAF\xA1\x1E\x00\xB2\xBE\x00\x80\xF6\xBE\x00\x80\xFA\xBE"
-     "\x00\x80"
-     "\xFE\xBE\x00\x80\x02\xBF\x00\x80\x06\xBF\x00\x80\x0A\xBF\x00\x80\x0E\xBF"
-     "\x00\x80"
-     "\xB8\x5D\x1F\x00\xB9\x69\x1F\x00\xBA\x7D\x1F\x00\xBB\x75\x1F\x00\xBC\x6D"
-     "\x1F\x00"
-     "\xBD\x15\x1F\x00\xBE\x1D\x1F\x00\xBF\x0D\x1F\x00\xB0\xA9\x1E\x00\xB1\xA9"
-     "\x1E\x00"
-     "\xB2\x8D\x1E\x00\xB3\x85\x1E\x00\xB4\x9D\x1E\x00\xB5\x85\x1E\x00\xB6\x85"
-     "\x1E\x00"
-     "\xB7\x65\x1F\x00\xB3\x29\x1E\x00\x12\xBF\x00\x80\x16\xBF\x00\x80\x1A\xBF"
-     "\x00\x80"
-     "\x1E\xBF\x00\x80\xB6\x21\x1E\x00\xB5\x29\x1E\x00\x22\xBF\x00\x80\xBB\x6D"
-     "\x1E\x00"
-     "\xBA\x6D\x1E\x00\x86\x08\x00\x00\x87\x2C\x00\x00\xBF\x41\x1E\x00\xBE\x59"
-     "\x1E\x00"
-     "\xBD\x51\x1E\x00\xBC\x59\x1E\x00\x80\x49\x00\x00\x81\x49\x00\x00\x82\x59"
-     "\x00\x00"
-     "\xA3\x65\x1E\x00\x26\xBF\x00\x80\xA5\x65\x1E\x00\xA6\x6D\x1E\x00\x42\xBE"
-     "\x00\x80"
-     "\x2A\xBF\x00\x80\x2E\xBF\x00\x80\xAA\x21\x1E\x00\xAB\x21\x1E\x00\xAC\x15"
-     "\x1E\x00"
-     "\xAD\x1D\x1E\x00\xAE\x15\x1E\x00\xAF\x0D\x1E\x00\xA8\x69\x1E\x00\xA9\x79"
-     "\x1E\x00"
-     "\xAA\x99\x1E\x00\xAB\x99\x1E\x00\xAC\xB5\x1E\x00\xAD\xBD\x1E\x00\xAE\xB5"
-     "\x1E\x00"
-     "\xAF\xAD\x1E\x00\x32\xBF\x00\x80\x36\xBF\x00\x80\x3A\xBF\x00\x80\x3E\xBF"
-     "\x00\x80"
-     "\x42\xBF\x00\x80\x46\xBF\x00\x80\x4A\xBF\x00\x80\x4E\xBF\x00\x80\xB8\x65"
-     "\x01\x00"
-     "\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x7D\x01\x00\xBC\x65\x01\x00\xBD\x6D"
-     "\x01\x00"
-     "\xBE\x65\x01\x00\xBF\xD9\x01\x00\xB0\xD5\x1E\x00\xB1\xA1\x1E\x00\xB2\xA1"
-     "\x1E\x00"
-     "\xB3\xA1\x1E\x00\xB4\xA1\x1E\x00\xB5\xA9\x1E\x00\xB6\x91\x1E\x00\xB7\x91"
-     "\x1E\x00"
-     "\xB3\x11\x1E\x00\x52\xBF\x00\x80\x56\xBF\x00\x80\x5A\xBF\x00\x80\x5E\xBF"
-     "\x00\x80"
-     "\xB6\x3D\x1E\x00\xB5\x3D\x1E\x00\x62\xBF\x00\x80\xBB\x65\x1E\x00\xBA\x11"
-     "\x1E\x00"
-     "\x66\xBF\x00\x80\x6A\xBF\x00\x80\xBF\x6D\x1E\x00\xBE\x65\x1E\x00\xBD\x75"
-     "\x1E\x00"
-     "\xBC\x75\x1E\x00\x6E\xBF\x00\x80\xA3\x55\x1E\x00\x72\xBF\x00\x80\x76\xBF"
-     "\x00\x80"
-     "\xA6\x79\x1E\x00\x7A\xBF\x00\x80\x7E\xBF\x00\x80\xA5\x79\x1E\x00\xAA\x55"
-     "\x1E\x00"
-     "\xAB\x21\x1E\x00\x82\xBF\x00\x80\x86\xBF\x00\x80\xAE\x21\x1E\x00\xAF\x29"
-     "\x1E\x00"
-     "\xAC\x31\x1E\x00\xAD\x31\x1E\x00\xA8\xD1\x02\x00\xA9\xE5\x02\x00\xAA\xED"
-     "\x02\x00"
-     "\xAB\x2D\x03\x00\xAC\x35\x03\x00\xAD\x3D\x03\x00\xAE\x35\x03\x00\xAF\x2D"
-     "\x03\x00"
-     "\x8A\xBF\x00\x80\x8E\xBF\x00\x80\x96\xBF\x00\x80\xBE\x60\x02\x00\x9A\xBF"
-     "\x00\x80"
-     "\x9E\xBF\x00\x80\xA2\xBF\x00\x80\xA6\xBF\x00\x80\xB8\xE9\x03\x00\xB9\xE9"
-     "\x03\x00"
-     "\xBA\x89\x03\x00\xBB\x85\x03\x00\xBC\x9D\x03\x00\xBD\x81\x03\x00\xBE\x81"
-     "\x03\x00"
-     "\xBF\xB5\x03\x00\xB0\x55\x03\x00\xB1\x5D\x03\x00\xB2\x55\x03\x00\xB3\xE9"
-     "\x03\x00"
-     "\xB4\xF9\x03\x00\xB5\xF9\x03\x00\xB6\xE9\x03\x00\xB7\xE1\x03\x00\x80\x6D"
-     "\x03\x00"
-     "\x81\xA5\x00\x00\x82\xAD\x00\x00\xB3\x69\x02\x00\xAA\xBF\x00\x80\xB5\xB1"
-     "\x03\x00"
-     "\xB6\xB1\x03\x00\xAE\xBF\x00\x80\x84\xE0\x02\x00\xB2\xBF\x00\x80\xBA\x9D"
-     "\x03\x00"
-     "\xBB\x95\x03\x00\xBC\x8D\x03\x00\xBD\x31\x03\x00\xBE\x31\x03\x00\xBF\x31"
-     "\x03\x00"
-     "\xA3\x25\x02\x00\xB6\xBF\x00\x80\xBA\xBF\x00\x80\xBE\xBF\x00\x80\xC2\xBF"
-     "\x00\x80"
-     "\xA6\xFD\x03\x00\xA5\xFD\x03\x00\xC6\xBF\x00\x80\xAB\xD9\x03\x00\xAA\xD1"
-     "\x03\x00"
-     "\x86\x48\x04\x00\x87\xAC\x03\x00\xAF\x7D\x03\x00\xAE\x7D\x03\x00\xAD\x7D"
-     "\x03\x00"
-     "\xAC\xC1\x03\x00\xCA\xBF\x00\x80\xCE\xBF\x00\x80\xD2\xBF\x00\x80\xD6\xBF"
-     "\x00\x80"
-     "\xE3\xA0\x03\x00\xDA\xBF\x00\x80\xE1\xA8\x01\x00\xDE\xBF\x00\x80\xEF\xF4"
-     "\x03\x00"
-     "\xE2\xBF\x00\x80\xE6\xBF\x00\x80\xEA\xBF\x00\x80\xEE\xBF\x00\x80\xF2\xBF"
-     "\x00\x80"
-     "\xF6\xBF\x00\x80\xFA\xBF\x00\x80\x80\x79\x00\x00\x81\x79\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\x84\x40\x02\x00\xE3\x60\x1E\x00\xFE\xBF\x00\x80\xE1\xE0\x1F\x00\xBE\x50"
-     "\x05\x00"
-     "\xEF\xD0\x1E\x00\x06\xC0\x00\x80\x86\xC8\x04\x00\x87\x40\x05\x00\x0A\xC0"
-     "\x00\x80"
-     "\x0E\xC0\x00\x80\x12\xC0\x00\x80\x16\xC0\x00\x80\x1A\xC0\x00\x80\x1E\xC0"
-     "\x00\x80"
-     "\x22\xC0\x00\x80\x26\xC0\x00\x80\x2A\xC0\x00\x80\x2E\xC0\x00\x80\x32\xC0"
-     "\x00\x80"
-     "\xEF\xC4\x1E\x00\x84\xA4\x05\x00\xE1\x24\x1E\x00\x36\xC0\x00\x80\xE3\xF0"
-     "\x1E\x00"
-     "\x3A\xC0\x00\x80\x3E\xC0\x00\x80\x42\xC0\x00\x80\x46\xC0\x00\x80\xB3\xF5"
-     "\x01\x00"
-     "\x4A\xC0\x00\x80\x4E\xC0\x00\x80\x52\xC0\x00\x80\x56\xC0\x00\x80\xB6\x45"
-     "\x01\x00"
-     "\xB5\xE5\x01\x00\x5A\xC0\x00\x80\xBB\x29\x01\x00\xBA\x21\x01\x00\x5E\xC0"
-     "\x00\x80"
-     "\x62\xC0\x00\x80\xBF\xF1\x00\x00\xBE\x29\x01\x00\xBD\x29\x01\x00\xBC\x31"
-     "\x01\x00"
-     "\xA8\x29\x06\x00\xA9\x35\x06\x00\xAA\x3D\x06\x00\xAB\x35\x06\x00\xAC\x2D"
-     "\x06\x00"
-     "\xAD\x85\x06\x00\xAE\x85\x06\x00\xAF\xB9\x06\x00\x02\xC0\x00\x80\x82\x1D"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x80\x1D\x00\x00\x66\xC0\x00\x80\x6A\xC0\x00\x80\x6E\xC0"
-     "\x00\x80"
-     "\xBE\x30\x03\x00\xB8\x4D\x07\x00\xB9\x51\x07\x00\xBA\x55\x07\x00\xBB\x69"
-     "\x07\x00"
-     "\xBC\x79\x07\x00\xBD\x79\x07\x00\xBE\x69\x07\x00\xBF\x69\x07\x00\xB0\xA5"
-     "\x06\x00"
-     "\xB1\xAD\x06\x00\xB2\xA5\x06\x00\xB3\xBD\x06\x00\xB4\xAD\x06\x00\xB5\x75"
-     "\x07\x00"
-     "\xB6\x7D\x07\x00\xB7\x75\x07\x00\xA3\xB5\x06\x00\x72\xC0\x00\x80\x86\x28"
-     "\x00\x00"
-     "\x87\x4C\x03\x00\x76\xC0\x00\x80\xA6\x05\x06\x00\xA5\xA5\x06\x00\x7A\xC0"
-     "\x00\x80"
-     "\xAB\x69\x06\x00\xAA\x61\x06\x00\x7E\xC0\x00\x80\x82\xC0\x00\x80\xAF\xB1"
-     "\x07\x00"
-     "\xAE\x69\x06\x00\xAD\x69\x06\x00\xAC\x71\x06\x00\x86\xC0\x00\x80\x8A\xC0"
-     "\x00\x80"
-     "\xB3\xBD\x07\x00\x8E\xC0\x00\x80\xB5\xAD\x07\x00\x92\xC0\x00\x80\x96\xC0"
-     "\x00\x80"
-     "\xB6\x5D\x07\x00\x92\xBF\x00\x80\x9A\xC0\x00\x80\xBB\x79\x07\x00\xBA\x61"
-     "\x07\x00"
-     "\xBD\x69\x07\x00\xBC\x61\x07\x00\xBF\x49\x07\x00\xBE\x55\x07\x00\xA8\x65"
-     "\x06\x00"
-     "\xA9\x75\x06\x00\xAA\x7D\x06\x00\xAB\x75\x06\x00\xAC\x6D\x06\x00\xAD\x95"
-     "\x06\x00"
-     "\xAE\x9D\x06\x00\xAF\x95\x06\x00\x9E\xC0\x00\x80\xA2\xC0\x00\x80\xA6\xC0"
-     "\x00\x80"
-     "\xAA\xC0\x00\x80\xAE\xC0\x00\x80\xB2\xC0\x00\x80\xB6\xC0\x00\x80\xBA\xC0"
-     "\x00\x80"
-     "\xB8\xFD\x06\x00\xB9\x81\x06\x00\xBA\x81\x06\x00\xBB\x81\x06\x00\xBC\x81"
-     "\x06\x00"
-     "\xBD\x89\x06\x00\xBE\xB1\x06\x00\xBF\xB1\x06\x00\xB0\xED\x06\x00\xB1\xF5"
-     "\x06\x00"
-     "\xB2\xFD\x06\x00\xB3\xF5\x06\x00\xB4\xD9\x06\x00\xB5\xC5\x06\x00\xB6\xCD"
-     "\x06\x00"
-     "\xB7\xC5\x06\x00\xA3\xF5\x06\x00\xBE\xC0\x00\x80\xC2\xC0\x00\x80\xC6\xC0"
-     "\x00\x80"
-     "\xCA\xC0\x00\x80\xA6\x15\x06\x00\xA5\xE5\x06\x00\xCE\xC0\x00\x80\xAB\x31"
-     "\x06\x00"
-     "\xAA\x29\x06\x00\xD2\xC0\x00\x80\xD6\xC0\x00\x80\xAF\x01\x06\x00\xAE\x1D"
-     "\x06\x00"
-     "\xAD\x21\x06\x00\xAC\x29\x06\x00\x80\x6D\x00\x00\x81\x09\x00\x00\x82\x19"
-     "\x00\x00"
-     "\xDA\xC0\x00\x80\xDE\xC0\x00\x80\x84\x98\x01\x00\xBE\x90\x01\x00\xE6\xC0"
-     "\x00\x80"
-     "\x86\x40\x0F\x00\x87\xF4\x00\x00\xEA\xC0\x00\x80\xEE\xC0\x00\x80\xF2\xC0"
-     "\x00\x80"
-     "\xF6\xC0\x00\x80\xFA\xC0\x00\x80\xFE\xC0\x00\x80\xA8\x91\x02\x00\xA9\x99"
-     "\x02\x00"
-     "\xAA\xA1\x02\x00\xAB\xA1\x02\x00\xAC\xD1\x02\x00\xAD\xD1\x02\x00\xAE\xD5"
-     "\x02\x00"
-     "\xAF\xC9\x02\x00\x02\xC1\x00\x80\x06\xC1\x00\x80\x0A\xC1\x00\x80\x84\x20"
-     "\x03\x00"
-     "\x0E\xC1\x00\x80\x12\xC1\x00\x80\x16\xC1\x00\x80\x1A\xC1\x00\x80\xB8\x79"
-     "\x03\x00"
-     "\xB9\x79\x03\x00\xBA\xCD\x03\x00\xBB\xC5\x03\x00\xBC\xDD\x03\x00\xBD\xC5"
-     "\x03\x00"
-     "\xBE\xC5\x03\x00\xBF\xF5\x03\x00\xB0\xB9\x02\x00\xB1\x81\x02\x00\xB2\x49"
-     "\x03\x00"
-     "\xB3\x49\x03\x00\xB4\x59\x03\x00\xB5\x59\x03\x00\xB6\x49\x03\x00\xB7\x49"
-     "\x03\x00"
-     "\xB3\x29\x02\x00\x1E\xC1\x00\x80\x22\xC1\x00\x80\x26\xC1\x00\x80\xBE\x9C"
-     "\x0C\x00"
-     "\xB6\x51\x02\x00\xB5\x59\x02\x00\x2A\xC1\x00\x80\xBB\x45\x02\x00\xBA\x7D"
-     "\x02\x00"
-     "\x2E\xC1\x00\x80\x32\xC1\x00\x80\xBF\x9D\x03\x00\xBE\x9D\x03\x00\xBD\x9D"
-     "\x03\x00"
-     "\xBC\x9D\x03\x00\x36\xC1\x00\x80\xA3\x6D\x02\x00\x3A\xC1\x00\x80\x3E\xC1"
-     "\x00\x80"
-     "\xA6\x15\x02\x00\x42\xC1\x00\x80\x46\xC1\x00\x80\xA5\x1D\x02\x00\xAA\x39"
-     "\x02\x00"
-     "\xAB\x01\x02\x00\x4A\xC1\x00\x80\x4E\xC1\x00\x80\xAE\xD9\x03\x00\xAF\xD9"
-     "\x03\x00"
-     "\xAC\xD9\x03\x00\xAD\xD9\x03\x00\xE1\x90\x01\x00\x56\xC1\x00\x80\xE3\xCC"
-     "\x00\x00"
-     "\x84\x18\x0D\x00\x80\xC9\x01\x00\x81\xFD\x01\x00\x82\xF5\x01\x00\x5A\xC1"
-     "\x00\x80"
-     "\xBE\x2C\x0C\x00\x5E\xC1\x00\x80\x62\xC1\x00\x80\x66\xC1\x00\x80\x6A\xC1"
-     "\x00\x80"
-     "\x6E\xC1\x00\x80\xEF\x18\x00\x00\x72\xC1\x00\x80\x86\xE0\x0C\x00\x87\xF4"
-     "\x02\x00"
-     "\x76\xC1\x00\x80\x7A\xC1\x00\x80\x7E\xC1\x00\x80\x82\xC1\x00\x80\x86\xC1"
-     "\x00\x80"
-     "\xEF\xF8\x06\x00\x8A\xC1\x00\x80\xE1\x94\x07\x00\x8E\xC1\x00\x80\xE3\x30"
-     "\x06\x00"
-     "\x92\xC1\x00\x80\x96\xC1\x00\x80\x9A\xC1\x00\x80\x9E\xC1\x00\x80\xA2\xC1"
-     "\x00\x80"
-     "\xA6\xC1\x00\x80\xAA\xC1\x00\x80\xAE\xC1\x00\x80\xB2\xC1\x00\x80\xB6\xC1"
-     "\x00\x80"
-     "\xBA\xC1\x00\x80\xEF\x58\x01\x00\x85\x00\x70\x00\xE1\xD4\x06\x00\xBE\xC1"
-     "\x00\x80"
-     "\xE3\x1C\x06\x00\x80\x25\x00\x00\x81\x2D\x00\x00\x82\x11\x00\x00\xC6\xC1"
-     "\x00\x80"
-     "\xB3\x2D\x02\x00\xF1\x00\x00\x00\x87\xE8\x0C\x00\x86\x6C\x0F\x00\xCA\xC1"
-     "\x00\x80"
-     "\xB6\x2D\x02\x00\xB5\x2D\x02\x00\xCE\xC1\x00\x80\xBB\xC9\x01\x00\xBA\xC1"
-     "\x01\x00"
-     "\xD2\xC1\x00\x80\xD6\xC1\x00\x80\xBF\xCD\x01\x00\xBE\xCD\x01\x00\xBD\xCD"
-     "\x01\x00"
-     "\xBC\xCD\x01\x00\xA8\xDD\x0D\x00\xA9\xED\x0D\x00\xAA\xE5\x0D\x00\xAB\xFD"
-     "\x0D\x00"
-     "\xAC\xE5\x0D\x00\xAD\xED\x0D\x00\xAE\xE5\x0D\x00\xAF\x3D\x0E\x00\x52\xC1"
-     "\x00\x80"
-     "\xC2\xC1\x00\x80\xDA\xC1\x00\x80\xDE\xC1\x00\x80\xE2\xC1\x00\x80\xE6\xC1"
-     "\x00\x80"
-     "\xEA\xC1\x00\x80\xEE\xC1\x00\x80\xB8\xCD\x0E\x00\xB9\xD9\x0E\x00\xBA\xE1"
-     "\x0E\x00"
-     "\xBB\xE1\x0E\x00\xBC\x91\x0E\x00\xBD\x9D\x0E\x00\xBE\x95\x0E\x00\xBF\x89"
-     "\x0E\x00"
-     "\xB0\x45\x0E\x00\xB1\x4D\x0E\x00\xB2\x59\x0E\x00\xB3\x59\x0E\x00\xB4\x49"
-     "\x0E\x00"
-     "\xB5\x49\x0E\x00\xB6\xFD\x0E\x00\xB7\xF5\x0E\x00\xA3\x6D\x0D\x00\xF2\xC1"
-     "\x00\x80"
-     "\xF6\xC1\x00\x80\xFA\xC1\x00\x80\xFE\xC1\x00\x80\xA6\x6D\x0D\x00\xA5\x6D"
-     "\x0D\x00"
-     "\x02\xC2\x00\x80\xAB\x89\x0E\x00\xAA\x81\x0E\x00\x06\xC2\x00\x80\x0A\xC2"
-     "\x00\x80"
-     "\xAF\x8D\x0E\x00\xAE\x8D\x0E\x00\xAD\x8D\x0E\x00\xAC\x8D\x0E\x00\xB5\xED"
-     "\x0F\x00"
-     "\x0E\xC2\x00\x80\x12\xC2\x00\x80\xB6\xFD\x0F\x00\x16\xC2\x00\x80\x1A\xC2"
-     "\x00\x80"
-     "\xB3\xFD\x0F\x00\x1E\xC2\x00\x80\xBD\xB5\x0F\x00\xBC\xB5\x0F\x00\xBF\x89"
-     "\x0F\x00"
-     "\xBE\xB5\x0F\x00\x22\xC2\x00\x80\xE2\xC0\x00\x80\xBB\xB5\x0F\x00\xBA\xB5"
-     "\x0F\x00"
-     "\xA8\x05\x0E\x00\xA9\x19\x0E\x00\xAA\x2D\x0E\x00\xAB\x25\x0E\x00\xAC\x3D"
-     "\x0E\x00"
-     "\xAD\x25\x0E\x00\xAE\x2D\x0E\x00\xAF\x21\x0E\x00\x80\x0D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x1D\x00\x00\x26\xC2\x00\x80\x2A\xC2\x00\x80\x2E\xC2\x00\x80\x84\x30"
-     "\x02\x00"
-     "\xBE\x30\x01\x00\xB8\x35\x0E\x00\xB9\x3D\x0E\x00\xBA\x35\x0E\x00\xBB\xCD"
-     "\x0E\x00"
-     "\xBC\xD5\x0E\x00\xBD\xDD\x0E\x00\xBE\xD5\x0E\x00\xBF\xCD\x0E\x00\xB0\x29"
-     "\x0E\x00"
-     "\xB1\x29\x0E\x00\xB2\x39\x0E\x00\xB3\x39\x0E\x00\xB4\x29\x0E\x00\xB5\x15"
-     "\x0E\x00"
-     "\xB6\x11\x0E\x00\xB7\x0D\x0E\x00\xA3\xA5\x0E\x00\x36\xC2\x00\x80\x86\x28"
-     "\x07\x00"
-     "\x87\x4C\x01\x00\x3A\xC2\x00\x80\xA6\xA5\x0E\x00\xA5\xB5\x0E\x00\x3E\xC2"
-     "\x00\x80"
-     "\xAB\xED\x0E\x00\xAA\xED\x0E\x00\x42\xC2\x00\x80\x46\xC2\x00\x80\xAF\xD1"
-     "\x0E\x00"
-     "\xAE\xED\x0E\x00\xAD\xED\x0E\x00\xAC\xED\x0E\x00\x4A\xC2\x00\x80\xB3\x19"
-     "\x0E\x00"
-     "\x4E\xC2\x00\x80\x52\xC2\x00\x80\xB6\xF9\x01\x00\x56\xC2\x00\x80\x5A\xC2"
-     "\x00\x80"
-     "\xB5\x31\x0E\x00\xBA\xD5\x01\x00\xBB\xDD\x01\x00\x5E\xC2\x00\x80\x62\xC2"
-     "\x00\x80"
-     "\xBE\x79\x01\x00\xBF\x79\x01\x00\xBC\xC5\x01\x00\xBD\x79\x01\x00\xA8\xBD"
-     "\x0D\x00"
-     "\xA9\xD5\x0D\x00\xAA\xD5\x0D\x00\xAB\xE5\x0D\x00\xAC\xFD\x0D\x00\xAD\x3D"
-     "\x02\x00"
-     "\xAE\x35\x02\x00\xAF\x2D\x02\x00\x66\xC2\x00\x80\x6A\xC2\x00\x80\x6E\xC2"
-     "\x00\x80"
-     "\x72\xC2\x00\x80\x76\xC2\x00\x80\x7A\xC2\x00\x80\x7E\xC2\x00\x80\x82\xC2"
-     "\x00\x80"
-     "\xB8\xE9\x02\x00\xB9\xE9\x02\x00\xBA\x8D\x02\x00\xBB\x85\x02\x00\xBC\x81"
-     "\x02\x00"
-     "\xBD\x81\x02\x00\xBE\x85\x02\x00\xBF\xAD\x02\x00\xB0\x55\x02\x00\xB1\x5D"
-     "\x02\x00"
-     "\xB2\x55\x02\x00\xB3\xE9\x02\x00\xB4\xF9\x02\x00\xB5\xF9\x02\x00\xB6\xE9"
-     "\x02\x00"
-     "\xB7\xE1\x02\x00\xA3\x5D\x0D\x00\x86\xC2\x00\x80\x8A\xC2\x00\x80\x8E\xC2"
-     "\x00\x80"
-     "\x92\xC2\x00\x80\xA6\xBD\x02\x00\xA5\x75\x0D\x00\x96\xC2\x00\x80\xAB\x99"
-     "\x02\x00"
-     "\xAA\x91\x02\x00\x9A\xC2\x00\x80\x9E\xC2\x00\x80\xAF\x3D\x02\x00\xAE\x3D"
-     "\x02\x00"
-     "\xAD\x3D\x02\x00\xAC\x81\x02\x00\x80\x5D\x00\x00\x81\x6D\x00\x00\x82\x65"
-     "\x00\x00"
-     "\xB3\x51\x03\x00\xA6\xC2\x00\x80\xB5\x79\x03\x00\xB6\x19\x03\x00\x84\xDC"
-     "\x02\x00"
-     "\xBE\x6C\x03\x00\xAA\xC2\x00\x80\xBA\x3D\x03\x00\xBB\x35\x03\x00\xBC\x19"
-     "\x03\x00"
-     "\xBD\x19\x03\x00\xBE\xD9\x00\x00\xBF\xD9\x00\x00\xA3\x9D\x02\x00\xAE\xC2"
-     "\x00\x80"
-     "\x86\x68\x05\x00\x87\xBC\x02\x00\xB2\xC2\x00\x80\xA6\xD5\x02\x00\xA5\xB5"
-     "\x02\x00"
-     "\xB6\xC2\x00\x80\xAB\xF9\x02\x00\xAA\xF1\x02\x00\x84\x20\x02\x00\xBA\xC2"
-     "\x00\x80"
-     "\xAF\x15\x01\x00\xAE\x15\x01\x00\xAD\xD5\x02\x00\xAC\xD5\x02\x00\xBE\xC2"
-     "\x00\x80"
-     "\xC2\xC2\x00\x80\xEF\x20\x0F\x00\xC6\xC2\x00\x80\xCA\xC2\x00\x80\xCE\xC2"
-     "\x00\x80"
-     "\xD2\xC2\x00\x80\xD6\xC2\x00\x80\xDA\xC2\x00\x80\xDE\xC2\x00\x80\xE2\xC2"
-     "\x00\x80"
-     "\xE6\xC2\x00\x80\xE1\x88\x0C\x00\xEA\xC2\x00\x80\xE3\x44\x0F\x00\xEE\xC2"
-     "\x00\x80"
-     "\xF2\xC2\x00\x80\xF6\xC2\x00\x80\xFA\xC2\x00\x80\xFE\xC2\x00\x80\xBE\x1C"
-     "\x04\x00"
-     "\x02\xC3\x00\x80\xEF\x5C\x00\x00\xEF\x44\x0E\x00\xE1\x74\x0E\x00\xE1\xFC"
-     "\x0E\x00"
-     "\xE3\x80\x00\x00\xE3\xB0\x0E\x00\x80\x0D\x00\x00\x81\x75\x00\x00\x82\x71"
-     "\x00\x00"
-     "\x06\xC3\x00\x80\x0A\xC3\x00\x80\xB3\xC1\x00\x00\x0E\xC3\x00\x80\x16\xC3"
-     "\x00\x80"
-     "\xB6\xFD\x00\x00\x1A\xC3\x00\x80\x1E\xC3\x00\x80\xB5\xC1\x00\x00\xBA\xAD"
-     "\x00\x00"
-     "\xBB\xB5\x00\x00\x86\xC8\x04\x00\x87\xA0\x04\x00\xBE\xAD\x00\x00\xBF\x95"
-     "\x00\x00"
-     "\xBC\xAD\x00\x00\xBD\xA5\x00\x00\xA8\x85\x05\x00\xA9\x95\x05\x00\xAA\x95"
-     "\x05\x00"
-     "\xAB\xA5\x05\x00\xAC\xBD\x05\x00\xAD\xD5\x05\x00\xAE\xD1\x05\x00\xAF\xD1"
-     "\x05\x00"
-     "\xA2\xC2\x00\x80\x12\xC3\x00\x80\x22\xC3\x00\x80\x26\xC3\x00\x80\x2A\xC3"
-     "\x00\x80"
-     "\x2E\xC3\x00\x80\x32\xC3\x00\x80\x36\xC3\x00\x80\xB8\x71\x06\x00\xB9\x71"
-     "\x06\x00"
-     "\xBA\x71\x06\x00\xBB\x71\x06\x00\xBC\xD5\x06\x00\xBD\xDD\x06\x00\xBE\xD5"
-     "\x06\x00"
-     "\xBF\xCD\x06\x00\xB0\xB5\x05\x00\xB1\xBD\x05\x00\xB2\x81\x05\x00\xB3\x81"
-     "\x05\x00"
-     "\xB4\x51\x06\x00\xB5\x51\x06\x00\xB6\x51\x06\x00\xB7\x51\x06\x00\xB8\xED"
-     "\x07\x00"
-     "\xB9\xF1\x07\x00\xBA\xF5\x07\x00\xBB\x0D\x07\x00\xBC\x15\x07\x00\xBD\x1D"
-     "\x07\x00"
-     "\xBE\x15\x07\x00\xBF\x09\x07\x00\xB0\xED\x07\x00\xB1\xF5\x07\x00\xB2\xFD"
-     "\x07\x00"
-     "\xB3\xF5\x07\x00\xB4\xED\x07\x00\xB5\xD5\x07\x00\xB6\xDD\x07\x00\xB7\xD5"
-     "\x07\x00"
-     "\xA8\x11\x06\x00\xA9\x21\x06\x00\xAA\xCD\x07\x00\xAB\xDD\x07\x00\xAC\xCD"
-     "\x07\x00"
-     "\xAD\xF9\x07\x00\xAE\x99\x07\x00\xAF\x99\x07\x00\x3A\xC3\x00\x80\x3E\xC3"
-     "\x00\x80"
-     "\x42\xC3\x00\x80\x46\xC3\x00\x80\x4A\xC3\x00\x80\x4E\xC3\x00\x80\x32\xC2"
-     "\x00\x80"
-     "\x52\xC3\x00\x80\xA3\x05\x06\x00\x56\xC3\x00\x80\x5A\xC3\x00\x80\x5E\xC3"
-     "\x00\x80"
-     "\x62\xC3\x00\x80\xA6\x39\x06\x00\xA5\x05\x06\x00\x66\xC3\x00\x80\xAB\x71"
-     "\x06\x00"
-     "\xAA\x69\x06\x00\x6A\xC3\x00\x80\x6E\xC3\x00\x80\xAF\x51\x06\x00\xAE\x69"
-     "\x06\x00"
-     "\xAD\x61\x06\x00\xAC\x69\x06\x00\x72\xC3\x00\x80\x82\x1D\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x76\xC3\x00\x80\x7A\xC3\x00\x80\x7E\xC3\x00\x80\xBE\x94"
-     "\x01\x00"
-     "\x84\xAC\x01\x00\x82\xC3\x00\x80\x87\xE8\x01\x00\x86\x00\x7C\x00\x86\xC3"
-     "\x00\x80"
-     "\x8A\xC3\x00\x80\x8E\xC3\x00\x80\x92\xC3\x00\x80\xA8\x4D\x06\x00\xA9\x51"
-     "\x06\x00"
-     "\xAA\x51\x06\x00\xAB\xA5\x01\x00\xAC\xBD\x01\x00\xAD\xA5\x01\x00\xAE\xAD"
-     "\x01\x00"
-     "\xAF\xA5\x01\x00\x96\xC3\x00\x80\x9A\xC3\x00\x80\x9E\xC3\x00\x80\xA2\xC3"
-     "\x00\x80"
-     "\xA6\xC3\x00\x80\xAA\xC3\x00\x80\xAE\xC3\x00\x80\xB2\xC3\x00\x80\xB8\xC5"
-     "\x00\x00"
-     "\xB9\xCD\x00\x00\xBA\xC5\x00\x00\xBB\xDD\x00\x00\xBC\xCD\x00\x00\xBD\xFD"
-     "\x00\x00"
-     "\xBE\xF5\x00\x00\xBF\x9D\x00\x00\xB0\xDD\x01\x00\xB1\x05\x01\x00\xB2\x0D"
-     "\x01\x00"
-     "\xB3\x05\x01\x00\xB4\x1D\x01\x00\xB5\x05\x01\x00\xB6\x0D\x01\x00\xB7\x01"
-     "\x01\x00"
-     "\xB3\xD9\x05\x00\xB6\xC3\x00\x80\xBA\xC3\x00\x80\xBE\xC3\x00\x80\xC2\xC3"
-     "\x00\x80"
-     "\xB6\x25\x02\x00\xB5\x35\x02\x00\xC6\xC3\x00\x80\xBB\xC1\x02\x00\xBA\x19"
-     "\x02\x00"
-     "\xCA\xC3\x00\x80\xCE\xC3\x00\x80\xBF\xC1\x02\x00\xBE\xD9\x02\x00\xBD\xD1"
-     "\x02\x00"
-     "\xBC\xD9\x02\x00\xD2\xC3\x00\x80\xA3\x9D\x05\x00\xD6\xC3\x00\x80\xDA\xC3"
-     "\x00\x80"
-     "\xA6\x61\x02\x00\xDE\xC3\x00\x80\xE2\xC3\x00\x80\xA5\x71\x02\x00\xAA\x5D"
-     "\x02\x00"
-     "\xAB\x85\x02\x00\xE6\xC3\x00\x80\xEA\xC3\x00\x80\xAE\x9D\x02\x00\xAF\x85"
-     "\x02\x00"
-     "\xAC\x9D\x02\x00\xAD\x95\x02\x00\x80\xE9\x00\x00\x81\xE9\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xEE\xC3\x00\x80\xBE\xDC\x03\x00\x84\x58\x03\x00\xF6\xC3\x00\x80\xFA\xC3"
-     "\x00\x80"
-     "\x86\x00\x7C\x00\x87\xEC\x03\x00\xFE\xC3\x00\x80\x02\xC4\x00\x80\x06\xC4"
-     "\x00\x80"
-     "\x0A\xC4\x00\x80\x0E\xC4\x00\x80\x12\xC4\x00\x80\x16\xC4\x00\x80\x1A\xC4"
-     "\x00\x80"
-     "\x1E\xC4\x00\x80\x22\xC4\x00\x80\x26\xC4\x00\x80\x2A\xC4\x00\x80\xEF\xB0"
-     "\x02\x00"
-     "\x2E\xC4\x00\x80\xE1\x94\x01\x00\x32\xC4\x00\x80\xE3\x6C\x02\x00\x36\xC4"
-     "\x00\x80"
-     "\x3A\xC4\x00\x80\x3E\xC4\x00\x80\x42\xC4\x00\x80\x46\xC4\x00\x80\xB3\x79"
-     "\x02\x00"
-     "\x4A\xC4\x00\x80\x4E\xC4\x00\x80\x52\xC4\x00\x80\x56\xC4\x00\x80\xB6\xB1"
-     "\x02\x00"
-     "\xB5\xB9\x02\x00\x5A\xC4\x00\x80\xBB\x95\x02\x00\xBA\x95\x02\x00\x5E\xC4"
-     "\x00\x80"
-     "\x84\x64\x7D\x00\xBF\x39\x01\x00\xBE\x31\x01\x00\xBD\x85\x02\x00\xBC\x85"
-     "\x02\x00"
-     "\x82\x69\x00\x00\xEF\x5C\x07\x00\x80\x59\x00\x00\x81\x51\x00\x00\x66\xC4"
-     "\x00\x80"
-     "\x6A\xC4\x00\x80\x87\x68\x7D\x00\x86\xFC\x7C\x00\xBE\x2C\x7D\x00\x6E\xC4"
-     "\x00\x80"
-     "\x72\xC4\x00\x80\x76\xC4\x00\x80\x84\x3C\x7C\x00\xE3\xB8\x07\x00\x7A\xC4"
-     "\x00\x80"
-     "\xE1\xEC\x04\x00\x7E\xC4\x00\x80\x82\xC4\x00\x80\x86\xC4\x00\x80\x84\xC0"
-     "\x7F\x00"
-     "\x8A\xC4\x00\x80\x8E\xC4\x00\x80\xEF\x3C\x00\x00\xEF\xF4\x06\x00\xE1\x50"
-     "\x06\x00"
-     "\xE1\xB4\x06\x00\xE3\xD8\x00\x00\xE3\x94\x06\x00\x92\xC4\x00\x80\x96\xC4"
-     "\x00\x80"
-     "\x9A\xC4\x00\x80\x9E\xC4\x00\x80\xA3\xB9\x02\x00\xA2\xC4\x00\x80\xA6\xC4"
-     "\x00\x80"
-     "\xAA\xC4\x00\x80\xAE\xC4\x00\x80\xA6\x71\x02\x00\xA5\x79\x02\x00\xB2\xC4"
-     "\x00\x80"
-     "\xAB\x55\x02\x00\xAA\x55\x02\x00\xB6\xC4\x00\x80\xBA\xC4\x00\x80\xAF\xF9"
-     "\x01\x00"
-     "\xAE\xF1\x01\x00\xAD\x45\x02\x00\xAC\x45\x02\x00\xA8\x99\x7E\x00\xA9\x99"
-     "\x7E\x00"
-     "\xAA\xA1\x7E\x00\xAB\xA1\x7E\x00\xAC\xA1\x7E\x00\xAD\xA1\x7E\x00\xAE\xA1"
-     "\x7E\x00"
-     "\xAF\xA1\x7E\x00\xBE\xC4\x00\x80\xC2\xC4\x00\x80\xC6\xC4\x00\x80\xCA\xC4"
-     "\x00\x80"
-     "\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\xCE\xC4\x00\x80\xB8\xAD"
-     "\x7E\x00"
-     "\xB9\xB5\x7E\x00\xBA\xBD\x7E\x00\xBB\xB5\x7E\x00\xBC\xAD\x7E\x00\xBD\x51"
-     "\x7F\x00"
-     "\xBE\x51\x7F\x00\xBF\x51\x7F\x00\xB0\xA1\x7E\x00\xB1\xA1\x7E\x00\xB2\xA1"
-     "\x7E\x00"
-     "\xB3\xA1\x7E\x00\xB4\xA1\x7E\x00\xB5\xA1\x7E\x00\xB6\xA5\x7E\x00\xB7\x99"
-     "\x7E\x00"
-     "\xF2\xC3\x00\x80\xB3\x2D\x7E\x00\xBE\xC0\x03\x00\x62\xC4\x00\x80\xB6\x25"
-     "\x7E\x00"
-     "\xD2\xC4\x00\x80\xD6\xC4\x00\x80\xB5\x2D\x7E\x00\xBA\x61\x7E\x00\xBB\x61"
-     "\x7E\x00"
-     "\x86\x08\x00\x00\x87\x9C\x00\x00\xBE\x61\x7E\x00\xBF\x61\x7E\x00\xBC\x71"
-     "\x7E\x00"
-     "\xBD\x71\x7E\x00\xA3\x69\x7E\x00\xDA\xC4\x00\x80\xDE\xC4\x00\x80\xE2\xC4"
-     "\x00\x80"
-     "\xE6\xC4\x00\x80\xA6\x61\x7E\x00\xA5\x69\x7E\x00\xEA\xC4\x00\x80\xAB\x25"
-     "\x7E\x00"
-     "\xAA\x25\x7E\x00\xEE\xC4\x00\x80\xF2\xC4\x00\x80\xAF\x25\x7E\x00\xAE\x25"
-     "\x7E\x00"
-     "\xAD\x35\x7E\x00\xAC\x35\x7E\x00\xA8\x51\x7E\x00\xA9\x51\x7E\x00\xAA\x6D"
-     "\x7E\x00"
-     "\xAB\x65\x7E\x00\xAC\x7D\x7E\x00\xAD\x6D\x7E\x00\xAE\x99\x7E\x00\xAF\x99"
-     "\x7E\x00"
-     "\xF6\xC4\x00\x80\xFA\xC4\x00\x80\xFE\xC4\x00\x80\x02\xC5\x00\x80\x06\xC5"
-     "\x00\x80"
-     "\x0A\xC5\x00\x80\x0E\xC5\x00\x80\x12\xC5\x00\x80\xB8\x89\x7E\x00\xB9\x89"
-     "\x7E\x00"
-     "\xBA\x99\x7E\x00\xBB\x91\x7E\x00\xBC\xB9\x7E\x00\xBD\xB9\x7E\x00\xBE\x79"
-     "\x01\x00"
-     "\xBF\x79\x01\x00\xB0\xE9\x7E\x00\xB1\xF5\x7E\x00\xB2\xFD\x7E\x00\xB3\xDD"
-     "\x7E\x00"
-     "\xB4\xC5\x7E\x00\xB5\xCD\x7E\x00\xB6\xC5\x7E\x00\xB7\xB9\x7E\x00\xB3\xE5"
-     "\x7E\x00"
-     "\x16\xC5\x00\x80\x1A\xC5\x00\x80\x1E\xC5\x00\x80\x22\xC5\x00\x80\xB6\xED"
-     "\x7E\x00"
-     "\xB5\xF5\x7E\x00\x26\xC5\x00\x80\xBB\x31\x7E\x00\xBA\x2D\x7E\x00\x2A\xC5"
-     "\x00\x80"
-     "\x2E\xC5\x00\x80\xBF\x09\x7E\x00\xBE\x01\x7E\x00\xBD\x09\x7E\x00\xBC\x29"
-     "\x7E\x00"
-     "\x32\xC5\x00\x80\xA3\xA1\x7E\x00\x36\xC5\x00\x80\x3A\xC5\x00\x80\xA6\xA9"
-     "\x7E\x00"
-     "\x3E\xC5\x00\x80\xBE\x90\x01\x00\xA5\xB1\x7E\x00\xAA\x69\x7E\x00\xAB\x75"
-     "\x7E\x00"
-     "\x85\xEC\x07\x00\x46\xC5\x00\x80\xAE\x45\x7E\x00\xAF\x4D\x7E\x00\xAC\x6D"
-     "\x7E\x00"
-     "\xAD\x4D\x7E\x00\xA8\x99\x02\x00\xA9\xC1\x02\x00\xAA\xC1\x02\x00\xAB\xC1"
-     "\x02\x00"
-     "\xAC\xC1\x02\x00\xAD\xC9\x02\x00\xAE\xF1\x02\x00\xAF\xF1\x02\x00\x86\x84"
-     "\x05\x00"
-     "\x87\x78\x03\x00\x4A\xC5\x00\x80\x4E\xC5\x00\x80\x80\xF9\x03\x00\x81\xF9"
-     "\x03\x00"
-     "\x82\xED\x03\x00\x52\xC5\x00\x80\xB8\xC5\x03\x00\xB9\xCD\x03\x00\xBA\xC5"
-     "\x03\x00"
-     "\xBB\xDD\x03\x00\xBC\xCD\x03\x00\xBD\xFD\x03\x00\xBE\xF5\x03\x00\xBF\x9D"
-     "\x03\x00"
-     "\xB0\x41\x03\x00\xB1\x41\x03\x00\xB2\x41\x03\x00\xB3\x41\x03\x00\xB4\x41"
-     "\x03\x00"
-     "\xB5\x41\x03\x00\xB6\x41\x03\x00\xB7\x41\x03\x00\xB3\x49\x02\x00\x56\xC5"
-     "\x00\x80"
-     "\x5A\xC5\x00\x80\x5E\xC5\x00\x80\x62\xC5\x00\x80\xB6\x49\x02\x00\xB5\x49"
-     "\x02\x00"
-     "\x66\xC5\x00\x80\xBB\x85\x03\x00\xBA\x85\x03\x00\x6A\xC5\x00\x80\x6E\xC5"
-     "\x00\x80"
-     "\xBF\x85\x03\x00\xBE\x85\x03\x00\xBD\x95\x03\x00\xBC\x95\x03\x00\x72\xC5"
-     "\x00\x80"
-     "\xA3\x0D\x02\x00\x76\xC5\x00\x80\x7A\xC5\x00\x80\xA6\x0D\x02\x00\x7E\xC5"
-     "\x00\x80"
-     "\x82\xC5\x00\x80\xA5\x0D\x02\x00\xAA\xC1\x03\x00\xAB\xC1\x03\x00\x86\xC5"
-     "\x00\x80"
-     "\x8A\xC5\x00\x80\xAE\xC1\x03\x00\xAF\xC1\x03\x00\xAC\xD1\x03\x00\xAD\xD1"
-     "\x03\x00"
-     "\xE3\x60\x7E\x00\xE1\xBC\x7F\x00\xE1\x4C\x7E\x00\xE3\x30\x7E\x00\xE1\xA0"
-     "\x01\x00"
-     "\xE1\xE8\x00\x00\xE3\x84\x7E\x00\xE3\xC4\x17\x00\xBE\xAC\x07\x00\x8E\xC5"
-     "\x00\x80"
-     "\xEF\x08\x00\x00\xEF\x7C\x1C\x00\xEF\x40\x01\x00\x92\xC5\x00\x80\x96\xC5"
-     "\x00\x80"
-     "\xEF\xA4\x7E\x00\x9A\xC5\x00\x80\x9E\xC5\x00\x80\x84\x04\x04\x00\xBE\xC0"
-     "\x06\x00"
-     "\xA2\xC5\x00\x80\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D\x00\x00\xA6\xC5"
-     "\x00\x80"
-     "\xAA\xC5\x00\x80\xAE\xC5\x00\x80\xB6\xC5\x00\x80\x86\x50\x07\x00\x87\x84"
-     "\x05\x00"
-     "\xBA\xC5\x00\x80\xBE\xC5\x00\x80\xBB\xC1\x7D\x00\xBA\xA9\x7D\x00\xB9\xA1"
-     "\x7D\x00"
-     "\xB8\x91\x7D\x00\xBF\xC9\x7D\x00\xBE\xC1\x7D\x00\xBD\xD5\x7D\x00\xBC\xD5"
-     "\x7D\x00"
-     "\xB3\x1D\x7E\x00\xB2\x1D\x7E\x00\xB1\x1D\x7E\x00\xB0\x1D\x7E\x00\xB7\xA1"
-     "\x7D\x00"
-     "\xB6\xB9\x7D\x00\xB5\xB1\x7D\x00\xB4\xB9\x7D\x00\xAB\xBD\x7E\x00\xAA\xBD"
-     "\x7E\x00"
-     "\xA9\xBD\x7E\x00\xA8\xBD\x7E\x00\xAF\x7D\x7E\x00\xAE\x7D\x7E\x00\xAD\x7D"
-     "\x7E\x00"
-     "\xAC\x7D\x7E\x00\xC2\xC5\x00\x80\xC6\xC5\x00\x80\xCA\xC5\x00\x80\xCE\xC5"
-     "\x00\x80"
-     "\xD2\xC5\x00\x80\xD6\xC5\x00\x80\xDA\xC5\x00\x80\xBE\xB0\x06\x00\x82\x05"
-     "\x6C\x00"
-     "\x83\x19\x6C\x00\x42\xC5\x00\x80\xB2\xC5\x00\x80\x86\xF9\x68\x00\x87\xC9"
-     "\x74\x00"
-     "\x84\x19\x68\x00\x85\x19\x68\x00\x8A\x01\x74\x00\x8B\xB9\x74\x00\xDE\xC5"
-     "\x00\x80"
-     "\xE2\xC5\x00\x80\x8E\xB5\x70\x00\x8F\xC1\x7C\x00\x8C\xD9\x70\x00\x8D\x1D"
-     "\x70\x00"
-     "\x92\x61\x7C\x00\x93\x61\x7C\x00\xE6\xC5\x00\x80\xEA\xC5\x00\x80\x96\x75"
-     "\x78\x00"
-     "\x97\x31\x79\x00\x94\xA1\x78\x00\x95\x7D\x78\x00\x9A\xE1\x04\x00\x9B\xE9"
-     "\x04\x00"
-     "\xEE\xC5\x00\x80\xF2\xC5\x00\x80\xF6\xC5\x00\x80\xFA\xC5\x00\x80\x9C\xED"
-     "\x04\x00"
-     "\xFE\xC5\x00\x80\xA1\x89\x00\x00\x02\xC6\x00\x80\xA3\x01\x0C\x00\xA2\x81"
-     "\x00\x00"
-     "\xA5\x15\x0C\x00\xA4\xB5\x0C\x00\xA7\xB9\x08\x00\xA6\x01\x08\x00\xA9\xF9"
-     "\x09\x00"
-     "\xA8\x81\x08\x00\xAB\x25\x14\x00\xAA\x2D\x14\x00\xAD\x01\x10\x00\xAC\x21"
-     "\x14\x00"
-     "\xAF\x99\x11\x00\xAE\x9D\x10\x00\xB1\xE1\x1C\x00\xB0\x01\x1C\x00\xB3\x01"
-     "\x18\x00"
-     "\xB2\x65\x1C\x00\xB5\xF5\x18\x00\xB4\xF5\x18\x00\x06\xC6\x00\x80\x0A\xC6"
-     "\x00\x80"
-     "\x0E\xC6\x00\x80\x12\xC6\x00\x80\x16\xC6\x00\x80\x1A\xC6\x00\x80\x80\x19"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x82\x05\x00\x00\x1E\xC6\x00\x80\xA9\x29\x07\x00\xA8\xE9"
-     "\x04\x00"
-     "\xAB\x35\x07\x00\xAA\x21\x07\x00\xAD\x99\x07\x00\xAC\x91\x07\x00\xAF\x89"
-     "\x07\x00"
-     "\xAE\x91\x07\x00\xA0\x45\x78\x00\xA1\xA5\x05\x00\xA2\xA5\x05\x00\xA3\xB9"
-     "\x05\x00"
-     "\x84\xB8\x01\x00\x26\xC6\x00\x80\x87\x34\x00\x00\x86\x9C\x0F\x00\xB9\xAD"
-     "\x07\x00"
-     "\xB8\xAD\x07\x00\xBB\x7D\x06\x00\xBA\x69\x06\x00\xBD\x11\x06\x00\xBC\x6D"
-     "\x06\x00"
-     "\xBF\x09\x06\x00\xBE\x19\x06\x00\xB1\xC9\x07\x00\xB0\xF9\x07\x00\xB3\xA9"
-     "\x07\x00"
-     "\xB2\xC1\x07\x00\xB5\xBD\x07\x00\xB4\xBD\x07\x00\xB7\xA1\x07\x00\xB6\xA9"
-     "\x07\x00"
-     "\xB3\xAD\x07\x00\x2A\xC6\x00\x80\x2E\xC6\x00\x80\x84\x0C\x00\x00\x32\xC6"
-     "\x00\x80"
-     "\xB6\x1D\x07\x00\xB5\x1D\x07\x00\x8F\x99\x6D\x00\xBB\x09\x07\x00\xBA\x39"
-     "\x07\x00"
-     "\x36\xC6\x00\x80\x3A\xC6\x00\x80\xBF\x7D\x07\x00\xBE\x7D\x07\x00\xBD\x79"
-     "\x07\x00"
-     "\xBC\x11\x07\x00\x98\xE1\x71\x00\x99\x41\x71\x00\x9A\x3D\x7D\x00\x9B\xED"
-     "\x7D\x00"
-     "\x9C\xF5\x7D\x00\x9D\x2D\x79\x00\x9E\xF9\x79\x00\x9F\xF9\x79\x00\x90\xFD"
-     "\x6D\x00"
-     "\x91\xE5\x69\x00\x92\xED\x69\x00\x93\x05\x69\x00\x94\x2D\x75\x00\x95\xD5"
-     "\x75\x00"
-     "\x96\x65\x75\x00\x97\x1D\x71\x00\xA3\x2D\x04\x00\x84\xC8\x0D\x00\xBE\xC0"
-     "\x0D\x00"
-     "\x3E\xC6\x00\x80\x42\xC6\x00\x80\xA6\x9D\x04\x00\xA5\x9D\x04\x00\x46\xC6"
-     "\x00\x80"
-     "\xAB\x89\x04\x00\xAA\xB9\x04\x00\x4A\xC6\x00\x80\x4E\xC6\x00\x80\xAF\xFD"
-     "\x04\x00"
-     "\xAE\xFD\x04\x00\xAD\xF9\x04\x00\xAC\x91\x04\x00\x52\xC6\x00\x80\xB3\x39"
-     "\x05\x00"
-     "\x56\xC6\x00\x80\x5A\xC6\x00\x80\xB6\xB9\x05\x00\x5E\xC6\x00\x80\x62\xC6"
-     "\x00\x80"
-     "\xB5\xB9\x05\x00\xBA\x95\x05\x00\xBB\x9D\x05\x00\x66\xC6\x00\x80\x6A\xC6"
-     "\x00\x80"
-     "\xBE\x39\x02\x00\xBF\x39\x02\x00\xBC\x51\x02\x00\xBD\x51\x02\x00\xA8\xED"
-     "\x05\x00"
-     "\xA9\xFD\x05\x00\xAA\xF5\x05\x00\xAB\x6D\x05\x00\xAC\x75\x05\x00\xAD\x7D"
-     "\x05\x00"
-     "\xAE\x75\x05\x00\xAF\x6D\x05\x00\x6E\xC6\x00\x80\x72\xC6\x00\x80\x76\xC6"
-     "\x00\x80"
-     "\x7A\xC6\x00\x80\x7E\xC6\x00\x80\x82\xC6\x00\x80\x86\xC6\x00\x80\x8A\xC6"
-     "\x00\x80"
-     "\xB8\xB1\x02\x00\xB9\xB1\x02\x00\xBA\xB5\x02\x00\xBB\x49\x02\x00\xBC\x09"
-     "\x02\x00"
-     "\xBD\x09\x02\x00\xBE\x39\x02\x00\xBF\x39\x02\x00\xB0\x15\x05\x00\xB1\x1D"
-     "\x05\x00"
-     "\xB2\xC1\x02\x00\xB3\xC1\x02\x00\xB4\xB9\x02\x00\xB5\xB9\x02\x00\xB6\x91"
-     "\x02\x00"
-     "\xB7\x91\x02\x00\xA3\x7D\x04\x00\x82\x15\x00\x00\x81\xB5\x01\x00\x80\xB5"
-     "\x01\x00"
-     "\x8E\xC6\x00\x80\xA6\xFD\x04\x00\xA5\xFD\x04\x00\x96\xC6\x00\x80\xAB\xD9"
-     "\x04\x00"
-     "\xAA\xD1\x04\x00\x86\x28\x0D\x00\x87\x1C\x0D\x00\xAF\x7D\x03\x00\xAE\x7D"
-     "\x03\x00"
-     "\xAD\x15\x03\x00\xAC\x15\x03\x00\x9A\xC6\x00\x80\x9E\xC6\x00\x80\xA2\xC6"
-     "\x00\x80"
-     "\xB3\xC5\x02\x00\xA6\xC6\x00\x80\xB5\xFD\x02\x00\xB6\xF5\x02\x00\xAA\xC6"
-     "\x00\x80"
-     "\xAE\xC6\x00\x80\xB2\xC6\x00\x80\xBA\x75\x01\x00\xBB\x45\x01\x00\xBC\x5D"
-     "\x01\x00"
-     "\xBD\x35\x01\x00\xBE\x31\x01\x00\xBF\x31\x01\x00\xE3\xA0\x00\x00\xE3\xDC"
-     "\x06\x00"
-     "\xE1\xE0\x06\x00\xE1\xB4\x07\x00\xB6\xC6\x00\x80\xBA\xC6\x00\x80\xBE\xC6"
-     "\x00\x80"
-     "\xBE\x90\x0F\x00\x86\x80\x0C\x00\x87\x18\x0C\x00\xC2\xC6\x00\x80\xC6\xC6"
-     "\x00\x80"
-     "\xEF\x04\x06\x00\xEF\x6C\x06\x00\xCA\xC6\x00\x80\xCE\xC6\x00\x80\xA3\x49"
-     "\x02\x00"
-     "\xD2\xC6\x00\x80\xD6\xC6\x00\x80\xDA\xC6\x00\x80\xDE\xC6\x00\x80\xA6\x79"
-     "\x02\x00"
-     "\xA5\x71\x02\x00\xE2\xC6\x00\x80\xAB\xC9\x01\x00\xAA\xF9\x01\x00\xE6\xC6"
-     "\x00\x80"
-     "\xEA\xC6\x00\x80\xAF\xBD\x01\x00\xAE\xBD\x01\x00\xAD\xB9\x01\x00\xAC\xD1"
-     "\x01\x00"
-     "\xA8\x35\x0D\x00\xA9\x81\x0D\x00\xAA\x81\x0D\x00\xAB\x9D\x0D\x00\xAC\xB9"
-     "\x0D\x00"
-     "\xAD\xB9\x0D\x00\xAE\x15\x0E\x00\xAF\x0D\x0E\x00\x92\xC6\x00\x80\xEE\xC6"
-     "\x00\x80"
-     "\xF2\xC6\x00\x80\xF6\xC6\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xB9"
-     "\x00\x00"
-     "\xFA\xC6\x00\x80\xB8\x29\x0E\x00\xB9\x29\x0E\x00\xBA\xE9\x0E\x00\xBB\xE9"
-     "\x0E\x00"
-     "\xBC\xF9\x0E\x00\xBD\xF9\x0E\x00\xBE\xE9\x0E\x00\xBF\xE9\x0E\x00\xB0\x7D"
-     "\x0E\x00"
-     "\xB1\x4D\x0E\x00\xB2\x45\x0E\x00\xB3\x2D\x0E\x00\xB4\x39\x0E\x00\xB5\x39"
-     "\x0E\x00"
-     "\xB6\x2D\x0E\x00\xB7\x25\x0E\x00\xFE\xC6\x00\x80\x02\xC7\x00\x80\x22\xC6"
-     "\x00\x80"
-     "\x06\xC7\x00\x80\x0A\xC7\x00\x80\x0E\xC7\x00\x80\x12\xC7\x00\x80\x16\xC7"
-     "\x00\x80"
-     "\xA8\xBD\x0F\x00\xA9\xE5\x0F\x00\xAA\xED\x0F\x00\xAB\xE5\x0F\x00\xAC\xE1"
-     "\x0F\x00"
-     "\xAD\xE9\x0F\x00\xAE\xD9\x0F\x00\xAF\xD9\x0F\x00\xB0\x51\x0F\x00\xB1\x51"
-     "\x0F\x00"
-     "\xB2\x51\x0F\x00\xB3\x65\x0F\x00\xB4\x7D\x0F\x00\xB5\x6D\x0F\x00\xB6\x65"
-     "\x0F\x00"
-     "\xB7\x19\x0F\x00\xB8\x29\x0F\x00\xB9\x29\x0F\x00\xBA\x39\x0F\x00\xBB\x31"
-     "\x0F\x00"
-     "\xBC\x19\x0F\x00\xBD\x19\x0F\x00\xBE\xDD\x00\x00\xBF\xD5\x00\x00\xB3\x45"
-     "\x0E\x00"
-     "\x1A\xC7\x00\x80\x1E\xC7\x00\x80\x22\xC7\x00\x80\x26\xC7\x00\x80\xB6\x49"
-     "\x0E\x00"
-     "\xB5\x49\x0E\x00\x2A\xC7\x00\x80\xBB\xB5\x0E\x00\xBA\xB5\x0E\x00\x2E\xC7"
-     "\x00\x80"
-     "\x32\xC7\x00\x80\xBF\x89\x0E\x00\xBE\x99\x0E\x00\xBD\x91\x0E\x00\xBC\xA5"
-     "\x0E\x00"
-     "\x36\xC7\x00\x80\xA3\x01\x0E\x00\x3A\xC7\x00\x80\x3E\xC7\x00\x80\xA6\x0D"
-     "\x0E\x00"
-     "\x42\xC7\x00\x80\x46\xC7\x00\x80\xA5\x0D\x0E\x00\xAA\xF1\x0E\x00\xAB\xF1"
-     "\x0E\x00"
-     "\x4A\xC7\x00\x80\x4E\xC7\x00\x80\xAE\xDD\x0E\x00\xAF\xCD\x0E\x00\xAC\xE1"
-     "\x0E\x00"
-     "\xAD\xD5\x0E\x00\xA8\x0D\x01\x00\xA9\x15\x01\x00\xAA\x1D\x01\x00\xAB\x15"
-     "\x01\x00"
-     "\xAC\x0D\x01\x00\xAD\x91\x00\x00\xAE\x91\x00\x00\xAF\x91\x00\x00\x52\xC7"
-     "\x00\x80"
-     "\x56\xC7\x00\x80\x5A\xC7\x00\x80\x5E\xC7\x00\x80\x62\xC7\x00\x80\x82\xB9"
-     "\x03\x00"
-     "\x81\xB1\x03\x00\x80\xB9\x03\x00\xB8\x99\x00\x00\xB9\xAD\x00\x00\xBA\xA5"
-     "\x00\x00"
-     "\xBB\x6D\x00\x00\xBC\x75\x00\x00\xBD\x7D\x00\x00\xBE\x75\x00\x00\xBF\x6D"
-     "\x00\x00"
-     "\xB0\xF1\x00\x00\xB1\xF9\x00\x00\xB2\xC1\x00\x00\xB3\xC1\x00\x00\xB4\xB1"
-     "\x00\x00"
-     "\xB5\xBD\x00\x00\xB6\xB5\x00\x00\xB7\xA9\x00\x00\x84\xAC\x02\x00\x66\xC7"
-     "\x00\x80"
-     "\x6A\xC7\x00\x80\x6E\xC7\x00\x80\x86\xB0\x04\x00\x87\xD8\x02\x00\x76\xC7"
-     "\x00\x80"
-     "\x7A\xC7\x00\x80\x7E\xC7\x00\x80\x82\xC7\x00\x80\x86\xC7\x00\x80\x8A\xC7"
-     "\x00\x80"
-     "\xBE\x74\x03\x00\x8E\xC7\x00\x80\x92\xC7\x00\x80\x96\xC7\x00\x80\xBB\x5D"
-     "\x03\x00"
-     "\xBA\x55\x03\x00\x9A\xC7\x00\x80\x9E\xC7\x00\x80\xBF\x19\x03\x00\xBE\x15"
-     "\x03\x00"
-     "\xBD\x75\x03\x00\xBC\x75\x03\x00\xB3\xF5\x02\x00\xA2\xC7\x00\x80\xA6\xC7"
-     "\x00\x80"
-     "\xAA\xC7\x00\x80\xAE\xC7\x00\x80\xB6\x71\x03\x00\xB5\xE5\x02\x00\xB2\xC7"
-     "\x00\x80"
-     "\xA3\x15\x02\x00\xB6\xC7\x00\x80\xBA\xC7\x00\x80\xBE\xC7\x00\x80\xC2\xC7"
-     "\x00\x80"
-     "\xA6\x91\x03\x00\xA5\x05\x02\x00\xC6\xC7\x00\x80\xAB\xBD\x03\x00\xAA\xB5"
-     "\x03\x00"
-     "\xCA\xC7\x00\x80\xCE\xC7\x00\x80\xAF\xF9\x03\x00\xAE\xF5\x03\x00\xAD\x95"
-     "\x03\x00"
-     "\xAC\x95\x03\x00\xD2\xC7\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xD6\xC7\x00\x80\xDA\xC7\x00\x80\xDE\xC7\x00\x80\xE6\xC7\x00\x80\x84\x6C"
-     "\x03\x00"
-     "\xEA\xC7\x00\x80\x87\x78\x03\x00\x86\xEC\x04\x00\xEE\xC7\x00\x80\xF2\xC7"
-     "\x00\x80"
-     "\xF6\xC7\x00\x80\xFA\xC7\x00\x80\xE1\xB8\x0E\x00\xE3\x98\x0C\x00\xE3\x18"
-     "\x03\x00"
-     "\xE1\x28\x0C\x00\xFE\xC7\x00\x80\x02\xC8\x00\x80\x06\xC8\x00\x80\x0A\xC8"
-     "\x00\x80"
-     "\x84\xEC\x04\x00\x0E\xC8\x00\x80\x12\xC8\x00\x80\x16\xC8\x00\x80\x1A\xC8"
-     "\x00\x80"
-     "\xEF\x30\x0A\x00\xEF\x5C\x0E\x00\x1E\xC8\x00\x80\xB3\x61\x01\x00\x22\xC8"
-     "\x00\x80"
-     "\x26\xC8\x00\x80\x2A\xC8\x00\x80\x2E\xC8\x00\x80\xB6\x79\x01\x00\xB5\x71"
-     "\x01\x00"
-     "\x32\xC8\x00\x80\xBB\x29\x01\x00\xBA\x59\x01\x00\x36\xC8\x00\x80\x3A\xC8"
-     "\x00\x80"
-     "\xBF\x1D\x01\x00\xBE\x1D\x01\x00\xBD\x1D\x01\x00\xBC\x31\x01\x00\x3E\xC8"
-     "\x00\x80"
-     "\x42\xC8\x00\x80\x46\xC8\x00\x80\x4A\xC8\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x39\x00\x00\xEF\x6C\x0F\x00\x4E\xC8\x00\x80\xE1\x94\x0E\x00\xBE\x60"
-     "\x04\x00"
-     "\xE3\xB8\x0F\x00\x56\xC8\x00\x80\x5A\xC8\x00\x80\x5E\xC8\x00\x80\xEF\x00"
-     "\x00\x00"
-     "\x86\xE0\x04\x00\x87\x8C\x04\x00\xA3\xE9\x01\x00\x84\x00\x07\x00\xA5\xF9"
-     "\x01\x00"
-     "\x62\xC8\x00\x80\x66\xC8\x00\x80\xA6\xF1\x01\x00\x6A\xC8\x00\x80\x6E\xC8"
-     "\x00\x80"
-     "\xAB\xA1\x01\x00\xAA\xD1\x01\x00\xAD\x95\x01\x00\xAC\xB9\x01\x00\xAF\x95"
-     "\x01\x00"
-     "\xAE\x95\x01\x00\xA8\x19\x06\x00\xA9\x19\x06\x00\xAA\x29\x06\x00\xAB\x29"
-     "\x06\x00"
-     "\xAC\x39\x06\x00\xAD\x39\x06\x00\xAE\x2D\x06\x00\xAF\x25\x06\x00\xE2\xC7"
-     "\x00\x80"
-     "\x52\xC8\x00\x80\x72\xC8\x00\x80\x76\xC8\x00\x80\x7A\xC8\x00\x80\x7E\xC8"
-     "\x00\x80"
-     "\x82\xC8\x00\x80\x86\xC8\x00\x80\xB8\xC5\x06\x00\xB9\xCD\x06\x00\xBA\xC5"
-     "\x06\x00"
-     "\xBB\xDD\x06\x00\xBC\xCD\x06\x00\xBD\xFD\x06\x00\xBE\xF5\x06\x00\xBF\x9D"
-     "\x06\x00"
-     "\xB0\x5D\x06\x00\xB1\x2D\x06\x00\xB2\x25\x06\x00\xB3\x09\x06\x00\xB4\x19"
-     "\x06\x00"
-     "\xB5\x19\x06\x00\xB6\x09\x06\x00\xB7\x01\x06\x00\x8A\xC8\x00\x80\x8E\xC8"
-     "\x00\x80"
-     "\x92\xC8\x00\x80\x96\xC8\x00\x80\x9A\xC8\x00\x80\x9E\xC8\x00\x80\xA2\xC8"
-     "\x00\x80"
-     "\x72\xC7\x00\x80\xA8\x7D\x06\x00\xA9\x41\x06\x00\xAA\x59\x06\x00\xAB\x59"
-     "\x06\x00"
-     "\xAC\x49\x06\x00\xAD\x49\x06\x00\xAE\xA9\x07\x00\xAF\xA9\x07\x00\xB0\xD9"
-     "\x07\x00"
-     "\xB1\xD9\x07\x00\xB2\xED\x07\x00\xB3\xE5\x07\x00\xB4\xFD\x07\x00\xB5\xE5"
-     "\x07\x00"
-     "\xB6\xE5\x07\x00\xB7\x99\x07\x00\xB8\xA9\x07\x00\xB9\xA9\x07\x00\xBA\xB9"
-     "\x07\x00"
-     "\xBB\xB1\x07\x00\xBC\x99\x07\x00\xBD\x99\x07\x00\xBE\x79\x07\x00\xBF\x79"
-     "\x07\x00"
-     "\xB3\xF1\x07\x00\xA6\xC8\x00\x80\xAA\xC8\x00\x80\xAE\xC8\x00\x80\xB2\xC8"
-     "\x00\x80"
-     "\xB6\xE9\x07\x00\xB5\xE1\x07\x00\xB6\xC8\x00\x80\xBB\x3D\x06\x00\xBA\x3D"
-     "\x06\x00"
-     "\xBA\xC8\x00\x80\xBE\xC8\x00\x80\xBF\x01\x06\x00\xBE\x19\x06\x00\xBD\x11"
-     "\x06\x00"
-     "\xBC\x19\x06\x00\x82\x6D\x00\x00\xA3\xB5\x07\x00\x80\x55\x00\x00\x81\x65"
-     "\x00\x00"
-     "\xA6\xAD\x07\x00\xBE\x9C\x01\x00\xC2\xC8\x00\x80\xA5\xA5\x07\x00\xAA\x79"
-     "\x06\x00"
-     "\xAB\x79\x06\x00\x86\x00\x1C\x00\x87\xCC\x01\x00\xAE\x5D\x06\x00\xAF\x45"
-     "\x06\x00"
-     "\xAC\x5D\x06\x00\xAD\x55\x06\x00\xA8\x31\x06\x00\xA9\x31\x06\x00\xAA\x31"
-     "\x06\x00"
-     "\xAB\x59\x06\x00\xAC\x49\x06\x00\xAD\x75\x06\x00\xAE\x7D\x06\x00\xAF\xB5"
-     "\x01\x00"
-     "\xC6\xC8\x00\x80\xCA\xC8\x00\x80\xCE\xC8\x00\x80\xD2\xC8\x00\x80\xD6\xC8"
-     "\x00\x80"
-     "\xDA\xC8\x00\x80\xDE\xC8\x00\x80\xE2\xC8\x00\x80\xB8\x4D\x01\x00\xB9\x55"
-     "\x01\x00"
-     "\xBA\x5D\x01\x00\xBB\x55\x01\x00\xBC\xD5\x00\x00\xBD\xDD\x00\x00\xBE\xD5"
-     "\x00\x00"
-     "\xBF\xCD\x00\x00\xB0\xCD\x01\x00\xB1\xD5\x01\x00\xB2\xDD\x01\x00\xB3\xD5"
-     "\x01\x00"
-     "\xB4\xCD\x01\x00\xB5\x75\x01\x00\xB6\x7D\x01\x00\xB7\x75\x01\x00\xB3\xB9"
-     "\x05\x00"
-     "\xE6\xC8\x00\x80\xEA\xC8\x00\x80\xEE\xC8\x00\x80\xF2\xC8\x00\x80\xB6\xF5"
-     "\x05\x00"
-     "\xB5\xC1\x05\x00\xF6\xC8\x00\x80\xBB\x11\x02\x00\xBA\x09\x02\x00\xFA\xC8"
-     "\x00\x80"
-     "\xFE\xC8\x00\x80\xBF\xB1\x02\x00\xBE\x09\x02\x00\xBD\x01\x02\x00\xBC\x09"
-     "\x02\x00"
-     "\x02\xC9\x00\x80\xA3\xFD\x05\x00\x06\xC9\x00\x80\x0A\xC9\x00\x80\xA6\xB1"
-     "\x05\x00"
-     "\x0E\xC9\x00\x80\x12\xC9\x00\x80\xA5\x85\x05\x00\xAA\x4D\x02\x00\xAB\x55"
-     "\x02\x00"
-     "\x16\xC9\x00\x80\x1A\xC9\x00\x80\xAE\x4D\x02\x00\xAF\xF5\x02\x00\xAC\x4D"
-     "\x02\x00"
-     "\xAD\x45\x02\x00\x1E\xC9\x00\x80\x22\xC9\x00\x80\x26\xC9\x00\x80\x2A\xC9"
-     "\x00\x80"
-     "\x2E\xC9\x00\x80\xBE\xD8\x03\x00\x36\xC9\x00\x80\x3A\xC9\x00\x80\x80\x29"
-     "\x00\x00"
-     "\x81\x29\x00\x00\x82\x39\x00\x00\x84\x20\x03\x00\x86\x10\x1C\x00\x87\x84"
-     "\x03\x00"
-     "\x3E\xC9\x00\x80\x42\xC9\x00\x80\x46\xC9\x00\x80\x4A\xC9\x00\x80\x4E\xC9"
-     "\x00\x80"
-     "\x52\xC9\x00\x80\x56\xC9\x00\x80\xE3\x68\x02\x00\x5A\xC9\x00\x80\xE1\xAC"
-     "\x01\x00"
-     "\x5E\xC9\x00\x80\xEF\x88\x02\x00\x62\xC9\x00\x80\x66\xC9\x00\x80\x6A\xC9"
-     "\x00\x80"
-     "\x6E\xC9\x00\x80\x72\xC9\x00\x80\x76\xC9\x00\x80\x7A\xC9\x00\x80\x7E\xC9"
-     "\x00\x80"
-     "\x82\xC9\x00\x80\x86\xC9\x00\x80\xE3\xE0\x07\x00\x8A\xC9\x00\x80\xE1\x10"
-     "\x07\x00"
-     "\x8E\xC9\x00\x80\xEF\xE8\x07\x00\x92\xC9\x00\x80\x96\xC9\x00\x80\x9A\xC9"
-     "\x00\x80"
-     "\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x9E\xC9\x00\x80\xB3\xE9"
-     "\x02\x00"
-     "\x84\x0C\x1D\x00\x86\xE8\x1C\x00\x87\x80\x1D\x00\xA6\xC9\x00\x80\xB6\xE1"
-     "\x02\x00"
-     "\xB5\xE9\x02\x00\xAA\xC9\x00\x80\xBB\x69\x01\x00\xBA\x61\x01\x00\xBE\x24"
-     "\x1D\x00"
-     "\xAE\xC9\x00\x80\xBF\x59\x01\x00\xBE\x69\x01\x00\xBD\x69\x01\x00\xBC\x71"
-     "\x01\x00"
-     "\xB2\xC9\x00\x80\xB6\xC9\x00\x80\xBA\xC9\x00\x80\xBE\xC9\x00\x80\xC2\xC9"
-     "\x00\x80"
-     "\xC6\xC9\x00\x80\xEF\x34\x00\x00\xEF\xCC\x06\x00\xE1\x78\x06\x00\xE1\x50"
-     "\x07\x00"
-     "\xE3\x74\x06\x00\xE3\x9C\x06\x00\xCA\xC9\x00\x80\xCE\xC9\x00\x80\xD2\xC9"
-     "\x00\x80"
-     "\xD6\xC9\x00\x80\xA3\x69\x02\x00\xDA\xC9\x00\x80\x84\x84\x1C\x00\xBE\x84"
-     "\x1C\x00"
-     "\xDE\xC9\x00\x80\xA6\x61\x02\x00\xA5\x69\x02\x00\xE2\xC9\x00\x80\xAB\xE9"
-     "\x01\x00"
-     "\xAA\xE1\x01\x00\xE6\xC9\x00\x80\xEA\xC9\x00\x80\xAF\xD9\x01\x00\xAE\xE9"
-     "\x01\x00"
-     "\xAD\xE9\x01\x00\xAC\xF1\x01\x00\xA8\x99\x1E\x00\xA9\x99\x1E\x00\xAA\xB1"
-     "\x1E\x00"
-     "\xAB\xB1\x1E\x00\xAC\xD1\x1E\x00\xAD\xD1\x1E\x00\xAE\xD1\x1E\x00\xAF\xD1"
-     "\x1E\x00"
-     "\xEE\xC9\x00\x80\xF2\xC9\x00\x80\xF6\xC9\x00\x80\xFA\xC9\x00\x80\xFE\xC9"
-     "\x00\x80"
-     "\x02\xCA\x00\x80\x06\xCA\x00\x80\x0A\xCA\x00\x80\xB8\xBD\x1E\x00\xB9\x45"
-     "\x1F\x00"
-     "\xBA\x4D\x1F\x00\xBB\x45\x1F\x00\xBC\x5D\x1F\x00\xBD\x41\x1F\x00\xBE\x45"
-     "\x1F\x00"
-     "\xBF\x79\x1F\x00\xB0\xB1\x1E\x00\xB1\xBD\x1E\x00\xB2\x89\x1E\x00\xB3\x85"
-     "\x1E\x00"
-     "\xB4\x9D\x1E\x00\xB5\x85\x1E\x00\xB6\x8D\x1E\x00\xB7\x85\x1E\x00\x32\xC9"
-     "\x00\x80"
-     "\xB3\x3D\x1E\x00\x0E\xCA\x00\x80\xA2\xC9\x00\x80\xB6\x55\x1E\x00\x12\xCA"
-     "\x00\x80"
-     "\x16\xCA\x00\x80\xB5\x5D\x1E\x00\xBA\x71\x1E\x00\xBB\x7D\x1E\x00\x86\x08"
-     "\x00\x00"
-     "\x87\x28\x00\x00\xBE\x5D\x1E\x00\xBF\x45\x1E\x00\xBC\x59\x1E\x00\xBD\x55"
-     "\x1E\x00"
-     "\xA3\x79\x1E\x00\x82\x4D\x00\x00\x81\xF5\x00\x00\x80\xF5\x00\x00\x1A\xCA"
-     "\x00\x80"
-     "\xA6\x11\x1E\x00\xA5\x19\x1E\x00\x1E\xCA\x00\x80\xAB\x39\x1E\x00\xAA\x35"
-     "\x1E\x00"
-     "\x22\xCA\x00\x80\x26\xCA\x00\x80\xAF\x01\x1E\x00\xAE\x19\x1E\x00\xAD\x11"
-     "\x1E\x00"
-     "\xAC\x1D\x1E\x00\xA8\x51\x1E\x00\xA9\x59\x1E\x00\xAA\x65\x1E\x00\xAB\x8D"
-     "\x1E\x00"
-     "\xAC\x95\x1E\x00\xAD\x9D\x1E\x00\xAE\x95\x1E\x00\xAF\x8D\x1E\x00\x2A\xCA"
-     "\x00\x80"
-     "\x2E\xCA\x00\x80\x32\xCA\x00\x80\x36\xCA\x00\x80\x3A\xCA\x00\x80\x3E\xCA"
-     "\x00\x80"
-     "\x42\xCA\x00\x80\x46\xCA\x00\x80\xB8\xA9\x1E\x00\xB9\xA9\x1E\x00\xBA\x69"
-     "\x01\x00"
-     "\xBB\x69\x01\x00\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x69\x01\x00\xBF\x69"
-     "\x01\x00"
-     "\xB0\xF5\x1E\x00\xB1\xD5\x1E\x00\xB2\xDD\x1E\x00\xB3\xD5\x1E\x00\xB4\xCD"
-     "\x1E\x00"
-     "\xB5\xB1\x1E\x00\xB6\xA9\x1E\x00\xB7\xA1\x1E\x00\xB3\xED\x1E\x00\x4A\xCA"
-     "\x00\x80"
-     "\x4E\xCA\x00\x80\x52\xCA\x00\x80\x56\xCA\x00\x80\xB6\x1D\x1E\x00\xB5\x1D"
-     "\x1E\x00"
-     "\x5A\xCA\x00\x80\xBB\x11\x1E\x00\xBA\x31\x1E\x00\x5E\xCA\x00\x80\x62\xCA"
-     "\x00\x80"
-     "\xBF\x75\x1E\x00\xBE\x09\x1E\x00\xBD\x01\x1E\x00\xBC\x09\x1E\x00\x66\xCA"
-     "\x00\x80"
-     "\xA3\xA9\x1E\x00\x6A\xCA\x00\x80\x6E\xCA\x00\x80\xA6\x59\x1E\x00\x72\xCA"
-     "\x00\x80"
-     "\x76\xCA\x00\x80\xA5\x59\x1E\x00\xAA\x75\x1E\x00\xAB\x55\x1E\x00\x7A\xCA"
-     "\x00\x80"
-     "\x7E\xCA\x00\x80\xAE\x4D\x1E\x00\xAF\x31\x1E\x00\xAC\x4D\x1E\x00\xAD\x45"
-     "\x1E\x00"
-     "\xA8\x99\x02\x00\xA9\xD5\x02\x00\xAA\xD5\x02\x00\xAB\xE5\x02\x00\xAC\xFD"
-     "\x02\x00"
-     "\xAD\x35\x03\x00\xAE\x3D\x03\x00\xAF\x35\x03\x00\xBE\x6C\x02\x00\x86\xCA"
-     "\x00\x80"
-     "\x8A\xCA\x00\x80\x8E\xCA\x00\x80\x92\xCA\x00\x80\x96\xCA\x00\x80\x9A\xCA"
-     "\x00\x80"
-     "\x9E\xCA\x00\x80\xB8\xD1\x03\x00\xB9\xD9\x03\x00\xBA\xE1\x03\x00\xBB\xE1"
-     "\x03\x00"
-     "\xBC\x91\x03\x00\xBD\x9D\x03\x00\xBE\x95\x03\x00\xBF\x89\x03\x00\xB0\x4D"
-     "\x03\x00"
-     "\xB1\x55\x03\x00\xB2\x5D\x03\x00\xB3\x55\x03\x00\xB4\x4D\x03\x00\xB5\xF1"
-     "\x03\x00"
-     "\xB6\xF1\x03\x00\xB7\xF1\x03\x00\x81\x69\x03\x00\x80\x5D\x03\x00\xB3\x55"
-     "\x02\x00"
-     "\x82\x61\x03\x00\xB5\x7D\x02\x00\xA2\xCA\x00\x80\xA6\xCA\x00\x80\xB6\xBD"
-     "\x03\x00"
-     "\x84\xE4\x02\x00\xAA\xCA\x00\x80\xBB\x99\x03\x00\xBA\x81\x03\x00\xBD\x89"
-     "\x03\x00"
-     "\xBC\x81\x03\x00\xBF\x3D\x03\x00\xBE\x81\x03\x00\xA3\x1D\x02\x00\xAE\xCA"
-     "\x00\x80"
-     "\x87\x98\x03\x00\x86\x6C\x04\x00\xB2\xCA\x00\x80\xA6\xF5\x03\x00\xA5\x35"
-     "\x02\x00"
-     "\xB6\xCA\x00\x80\xAB\xD1\x03\x00\xAA\xC9\x03\x00\xBA\xCA\x00\x80\xBE\xCA"
-     "\x00\x80"
-     "\xAF\x75\x03\x00\xAE\xC9\x03\x00\xAD\xC1\x03\x00\xAC\xC9\x03\x00\xE3\xD8"
-     "\x1E\x00"
-     "\xE1\xAC\x1F\x00\xE1\x58\x1E\x00\xE3\x58\x1E\x00\xE1\xA0\x01\x00\xC2\xCA"
-     "\x00\x80"
-     "\xE3\xA8\x03\x00\xC6\xCA\x00\x80\xCA\xCA\x00\x80\xCE\xCA\x00\x80\xEF\xF0"
-     "\x03\x00"
-     "\xD2\xCA\x00\x80\xEF\x04\x1E\x00\xD6\xCA\x00\x80\xDA\xCA\x00\x80\xEF\xAC"
-     "\x1E\x00"
-     "\xDE\xCA\x00\x80\xE2\xCA\x00\x80\x84\x1C\x05\x00\xBE\x00\x05\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x35\x00\x00\x82\x3D\x00\x00\xE6\xCA\x00\x80\xEA\xCA\x00\x80\xEE\xCA"
-     "\x00\x80"
-     "\x86\xC8\x04\x00\x87\xA0\x05\x00\xF6\xCA\x00\x80\xFA\xCA\x00\x80\xFE\xCA"
-     "\x00\x80"
-     "\x02\xCB\x00\x80\xBA\xD5\x1D\x00\xBB\xDD\x1D\x00\xB8\xC1\x1D\x00\xB9\xC1"
-     "\x1D\x00"
-     "\xBE\x11\x1D\x00\xBF\x11\x1D\x00\xBC\xF1\x1D\x00\xBD\xF1\x1D\x00\xB2\xAD"
-     "\x1D\x00"
-     "\xB3\xB5\x1D\x00\xB0\xAD\x1D\x00\xB1\xA5\x1D\x00\xB6\x9D\x1D\x00\xB7\xF5"
-     "\x1D\x00"
-     "\xB4\xA5\x1D\x00\xB5\x95\x1D\x00\xAA\x69\x1E\x00\xAB\x69\x1E\x00\xA8\x69"
-     "\x1E\x00"
-     "\xA9\x69\x1E\x00\xAE\x69\x1E\x00\xAF\x69\x1E\x00\xAC\x69\x1E\x00\xAD\x69"
-     "\x1E\x00"
-     "\x06\xCB\x00\x80\x0A\xCB\x00\x80\x0E\xCB\x00\x80\x12\xCB\x00\x80\x16\xCB"
-     "\x00\x80"
-     "\x1A\xCB\x00\x80\x1E\xCB\x00\x80\x22\xCB\x00\x80\xA8\x25\x06\x00\xA9\x2D"
-     "\x06\x00"
-     "\xAA\x25\x06\x00\xAB\x3D\x06\x00\xAC\x25\x06\x00\xAD\x2D\x06\x00\xAE\x25"
-     "\x06\x00"
-     "\xAF\x89\x06\x00\xF2\xCA\x00\x80\x26\xCB\x00\x80\x2A\xCB\x00\x80\x2E\xCB"
-     "\x00\x80"
-     "\x32\xCB\x00\x80\x36\xCB\x00\x80\x3A\xCB\x00\x80\x3E\xCB\x00\x80\xB8\x5D"
-     "\x07\x00"
-     "\xB9\x65\x07\x00\xBA\x6D\x07\x00\xBB\x61\x07\x00\xBC\x65\x07\x00\xBD\x6D"
-     "\x07\x00"
-     "\xBE\x19\x07\x00\xBF\x19\x07\x00\xB0\xF9\x06\x00\xB1\xC1\x06\x00\xB2\xAD"
-     "\x06\x00"
-     "\xB3\xBD\x06\x00\xB4\xAD\x06\x00\xB5\x9D\x06\x00\xB6\x95\x06\x00\xB7\x69"
-     "\x07\x00"
-     "\xB5\xB5\x06\x00\x42\xCB\x00\x80\x46\xCB\x00\x80\xB6\xB5\x06\x00\x4A\xCB"
-     "\x00\x80"
-     "\x4E\xCB\x00\x80\xB3\xB5\x06\x00\x52\xCB\x00\x80\xBD\x71\x06\x00\xBC\x61"
-     "\x06\x00"
-     "\xBF\x41\x06\x00\xBE\x71\x06\x00\x82\xCA\x00\x80\x56\xCB\x00\x80\xBB\x1D"
-     "\x06\x00"
-     "\xBA\x25\x06\x00\xA3\xED\x06\x00\x82\x19\x00\x00\x81\x19\x00\x00\x80\x79"
-     "\x00\x00"
-     "\x5A\xCB\x00\x80\xA6\xED\x06\x00\xA5\xED\x06\x00\x5E\xCB\x00\x80\xAB\x45"
-     "\x06\x00"
-     "\xAA\x7D\x06\x00\x84\xA4\x03\x00\xBE\xA4\x03\x00\xAF\x19\x06\x00\xAE\x29"
-     "\x06\x00"
-     "\xAD\x29\x06\x00\xAC\x39\x06\x00\xA8\x71\x06\x00\xA9\x71\x06\x00\xAA\x71"
-     "\x06\x00"
-     "\xAB\x71\x06\x00\xAC\xF1\x06\x00\xAD\xF1\x06\x00\xAE\xF1\x06\x00\xAF\xF1"
-     "\x06\x00"
-     "\x62\xCB\x00\x80\x66\xCB\x00\x80\x86\x88\x01\x00\x87\xAC\x01\x00\x6A\xCB"
-     "\x00\x80"
-     "\x6E\xCB\x00\x80\x72\xCB\x00\x80\x76\xCB\x00\x80\xB8\x75\x01\x00\xB9\x7D"
-     "\x01\x00"
-     "\xBA\x75\x01\x00\xBB\xC9\x01\x00\xBC\xD9\x01\x00\xBD\xC5\x01\x00\xBE\xCD"
-     "\x01\x00"
-     "\xBF\xC1\x01\x00\xB0\x91\x06\x00\xB1\x91\x06\x00\xB2\x91\x06\x00\xB3\x91"
-     "\x06\x00"
-     "\xB4\x55\x01\x00\xB5\x5D\x01\x00\xB6\x55\x01\x00\xB7\x4D\x01\x00\xB3\xF9"
-     "\x06\x00"
-     "\x7A\xCB\x00\x80\x7E\xCB\x00\x80\x82\xCB\x00\x80\x86\xCB\x00\x80\xB6\x71"
-     "\x06\x00"
-     "\xB5\x79\x06\x00\x8A\xCB\x00\x80\xBB\x55\x06\x00\xBA\x55\x06\x00\x8E\xCB"
-     "\x00\x80"
-     "\x92\xCB\x00\x80\xBF\x99\x01\x00\xBE\x91\x01\x00\xBD\x45\x06\x00\xBC\x45"
-     "\x06\x00"
-     "\x96\xCB\x00\x80\xA3\xBD\x06\x00\x9A\xCB\x00\x80\x9E\xCB\x00\x80\xA6\x35"
-     "\x06\x00"
-     "\xA2\xCB\x00\x80\xA6\xCB\x00\x80\xA5\x3D\x06\x00\xAA\x11\x06\x00\xAB\x11"
-     "\x06\x00"
-     "\xAA\xCB\x00\x80\xAE\xCB\x00\x80\xAE\xD5\x01\x00\xAF\xDD\x01\x00\xAC\x01"
-     "\x06\x00"
-     "\xAD\x01\x06\x00\xA8\x25\x03\x00\xA9\x35\x03\x00\xAA\x3D\x03\x00\xAB\x35"
-     "\x03\x00"
-     "\xAC\x2D\x03\x00\xAD\x91\x03\x00\xAE\x91\x03\x00\xAF\x91\x03\x00\xB2\xCB"
-     "\x00\x80"
-     "\xB6\xCB\x00\x80\xBA\xCB\x00\x80\xBE\xCB\x00\x80\xC2\xCB\x00\x80\xC6\xCB"
-     "\x00\x80"
-     "\xCA\xCB\x00\x80\xCE\xCB\x00\x80\xB8\x99\x03\x00\xB9\xAD\x03\x00\xBA\xA5"
-     "\x03\x00"
-     "\xBB\x6D\x00\x00\xBC\x75\x00\x00\xBD\x7D\x00\x00\xBE\x75\x00\x00\xBF\x6D"
-     "\x00\x00"
-     "\xB0\xF1\x03\x00\xB1\xF9\x03\x00\xB2\xC1\x03\x00\xB3\xC1\x03\x00\xB4\xB1"
-     "\x03\x00"
-     "\xB5\xBD\x03\x00\xB6\xB5\x03\x00\xB7\xA9\x03\x00\x80\xB5\x00\x00\x81\xBD"
-     "\x00\x00"
-     "\x82\xB5\x00\x00\xBE\x48\x03\x00\x87\x10\x03\x00\x86\x34\x0C\x00\xD6\xCB"
-     "\x00\x80"
-     "\xEF\xE8\x02\x00\xBE\xAC\x0C\x00\xE1\x94\x01\x00\x84\xA4\x02\x00\xE3\x14"
-     "\x02\x00"
-     "\xDA\xCB\x00\x80\xDE\xCB\x00\x80\xE2\xCB\x00\x80\xE6\xCB\x00\x80\xB3\xF5"
-     "\x03\x00"
-     "\xEA\xCB\x00\x80\xEE\xCB\x00\x80\x85\x74\x0A\x00\xF2\xCB\x00\x80\xB6\x51"
-     "\x03\x00"
-     "\xB5\xE5\x03\x00\xF6\xCB\x00\x80\xBB\x7D\x03\x00\xBA\x75\x03\x00\xFA\xCB"
-     "\x00\x80"
-     "\xFE\xCB\x00\x80\xBF\x39\x03\x00\xBE\x35\x03\x00\xBD\x55\x03\x00\xBC\x55"
-     "\x03\x00"
-     "\x02\xCC\x00\x80\x06\xCC\x00\x80\x0A\xCC\x00\x80\xEF\xB8\x06\x00\x0E\xCC"
-     "\x00\x80"
-     "\x12\xCC\x00\x80\x16\xCC\x00\x80\x1A\xCC\x00\x80\x1E\xCC\x00\x80\x22\xCC"
-     "\x00\x80"
-     "\x26\xCC\x00\x80\x2A\xCC\x00\x80\xCC\x00\x00\x00\xE1\xE0\x07\x00\x2E\xCC"
-     "\x00\x80"
-     "\xE3\x64\x06\x00\x32\xCC\x00\x80\x36\xCC\x00\x80\xA3\x7D\x02\x00\x84\x40"
-     "\x02\x00"
-     "\xA5\x6D\x02\x00\x3A\xCC\x00\x80\x3E\xCC\x00\x80\xA6\xD9\x02\x00\x86\xC0"
-     "\x0C\x00"
-     "\x87\xAC\x0D\x00\xAB\xF5\x02\x00\xAA\xFD\x02\x00\xAD\xDD\x02\x00\xAC\xDD"
-     "\x02\x00"
-     "\xAF\xB1\x02\x00\xAE\xBD\x02\x00\x80\xC5\x00\x00\x81\xD1\x00\x00\x82\xD1"
-     "\x00\x00"
-     "\x46\xCC\x00\x80\x4A\xCC\x00\x80\x4E\xCC\x00\x80\x52\xCC\x00\x80\xEF\x88"
-     "\x01\x00"
-     "\x84\xEC\x0C\x00\xE1\xD4\x06\x00\x56\xCC\x00\x80\xE3\x54\x01\x00\x5A\xCC"
-     "\x00\x80"
-     "\x5E\xCC\x00\x80\x62\xCC\x00\x80\x66\xCC\x00\x80\xB3\x75\x01\x00\x6A\xCC"
-     "\x00\x80"
-     "\x6E\xCC\x00\x80\x72\xCC\x00\x80\x76\xCC\x00\x80\xB6\x11\x01\x00\xB5\x65"
-     "\x01\x00"
-     "\x7A\xCC\x00\x80\xBB\x09\x01\x00\xBA\x39\x01\x00\x7E\xCC\x00\x80\x82\xCC"
-     "\x00\x80"
-     "\xBF\xF9\x00\x00\xBE\xF1\x00\x00\xBD\xF9\x00\x00\xBC\x11\x01\x00\xA8\x99"
-     "\x0E\x00"
-     "\xA9\xB5\x0E\x00\xAA\xBD\x0E\x00\xAB\xB5\x0E\x00\xAC\xAD\x0E\x00\xAD\xD1"
-     "\x0E\x00"
-     "\xAE\xD1\x0E\x00\xAF\xCD\x0E\x00\x42\xCC\x00\x80\x86\xCC\x00\x80\x8A\xCC"
-     "\x00\x80"
-     "\x8E\xCC\x00\x80\x92\xCC\x00\x80\x96\xCC\x00\x80\x9A\xCC\x00\x80\x9E\xCC"
-     "\x00\x80"
-     "\xB8\x59\x0F\x00\xB9\x59\x0F\x00\xBA\x69\x0F\x00\xBB\x69\x0F\x00\xBC\x79"
-     "\x0F\x00"
-     "\xBD\x79\x0F\x00\xBE\x61\x0F\x00\xBF\x1D\x0F\x00\xB0\xBD\x0E\x00\xB1\x8D"
-     "\x0E\x00"
-     "\xB2\x85\x0E\x00\xB3\x6D\x0F\x00\xB4\x75\x0F\x00\xB5\x7D\x0F\x00\xB6\x69"
-     "\x0F\x00"
-     "\xB7\x69\x0F\x00\xA3\x35\x0E\x00\xA2\xCC\x00\x80\xA6\xCC\x00\x80\xAA\xCC"
-     "\x00\x80"
-     "\xAE\xCC\x00\x80\xA6\x51\x0E\x00\xA5\x25\x0E\x00\xBE\xD4\x03\x00\xAB\x49"
-     "\x0E\x00"
-     "\xAA\x79\x0E\x00\xB2\xCC\x00\x80\xB6\xCC\x00\x80\xAF\xB9\x0F\x00\xAE\xB1"
-     "\x0F\x00"
-     "\xAD\xB9\x0F\x00\xAC\x51\x0E\x00\x82\x51\x00\x00\xD2\xCB\x00\x80\x80\x41"
-     "\x00\x00"
-     "\x81\x41\x00\x00\x87\x18\x03\x00\x86\x74\x00\x00\xBA\xCC\x00\x80\xBE\xCC"
-     "\x00\x80"
-     "\xC2\xCC\x00\x80\xC6\xCC\x00\x80\xCA\xCC\x00\x80\xCE\xCC\x00\x80\xD2\xCC"
-     "\x00\x80"
-     "\xD6\xCC\x00\x80\xDA\xCC\x00\x80\xDE\xCC\x00\x80\xA8\x8D\x0E\x00\xA9\x95"
-     "\x0E\x00"
-     "\xAA\x9D\x0E\x00\xAB\x95\x0E\x00\xAC\xB9\x0E\x00\xAD\xA5\x0E\x00\xAE\xAD"
-     "\x0E\x00"
-     "\xAF\xA5\x0E\x00\xE2\xCC\x00\x80\xE6\xCC\x00\x80\xEA\xCC\x00\x80\xEE\xCC"
-     "\x00\x80"
-     "\xF2\xCC\x00\x80\xF6\xCC\x00\x80\xFA\xCC\x00\x80\xFE\xCC\x00\x80\xB8\x65"
-     "\x01\x00"
-     "\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x7D\x01\x00\xBC\x65\x01\x00\xBD\x6D"
-     "\x01\x00"
-     "\xBE\x65\x01\x00\xBF\xD9\x01\x00\xB0\xDD\x0E\x00\xB1\xA1\x0E\x00\xB2\xA1"
-     "\x0E\x00"
-     "\xB3\xA1\x0E\x00\xB4\xA1\x0E\x00\xB5\xA9\x0E\x00\xB6\x91\x0E\x00\xB7\x91"
-     "\x0E\x00"
-     "\xB3\x15\x0E\x00\x02\xCD\x00\x80\x06\xCD\x00\x80\x0A\xCD\x00\x80\x0E\xCD"
-     "\x00\x80"
-     "\xB6\x25\x0E\x00\xB5\x31\x0E\x00\x12\xCD\x00\x80\xBB\x65\x0E\x00\xBA\x19"
-     "\x0E\x00"
-     "\x16\xCD\x00\x80\x1A\xCD\x00\x80\xBF\x6D\x0E\x00\xBE\x65\x0E\x00\xBD\x75"
-     "\x0E\x00"
-     "\xBC\x75\x0E\x00\x82\x15\x00\x00\xA3\x51\x0E\x00\x80\x61\x00\x00\x81\x61"
-     "\x00\x00"
-     "\xA6\x61\x0E\x00\x1E\xCD\x00\x80\xBE\x90\x01\x00\xA5\x75\x0E\x00\xAA\x5D"
-     "\x0E\x00"
-     "\xAB\x21\x0E\x00\x84\xA4\x01\x00\x26\xCD\x00\x80\xAE\x21\x0E\x00\xAF\x29"
-     "\x0E\x00"
-     "\xAC\x31\x0E\x00\xAD\x31\x0E\x00\xA8\xAD\x02\x00\xA9\xDD\x02\x00\xAA\xD5"
-     "\x02\x00"
-     "\xAB\xED\x02\x00\xAC\xF5\x02\x00\xAD\xFD\x02\x00\xAE\xF5\x02\x00\xAF\xED"
-     "\x02\x00"
-     "\x2A\xCD\x00\x80\x2E\xCD\x00\x80\x32\xCD\x00\x80\x36\xCD\x00\x80\x3A\xCD"
-     "\x00\x80"
-     "\x3E\xCD\x00\x80\x42\xCD\x00\x80\x46\xCD\x00\x80\xB8\x75\x03\x00\xB9\x7D"
-     "\x03\x00"
-     "\xBA\x75\x03\x00\xBB\xC9\x03\x00\xBC\xD9\x03\x00\xBD\xD9\x03\x00\xBE\xC9"
-     "\x03\x00"
-     "\xBF\xC1\x03\x00\xB0\x95\x02\x00\xB1\x9D\x02\x00\xB2\x95\x02\x00\xB3\x4D"
-     "\x03\x00"
-     "\xB4\x55\x03\x00\xB5\x5D\x03\x00\xB6\x55\x03\x00\xB7\x4D\x03\x00\xB3\x5D"
-     "\x02\x00"
-     "\x4A\xCD\x00\x80\x87\xD8\x03\x00\x86\x2C\x04\x00\x4E\xCD\x00\x80\xB6\x7D"
-     "\x02\x00"
-     "\xB5\x7D\x02\x00\x52\xCD\x00\x80\xBB\x59\x02\x00\xBA\x51\x02\x00\x56\xCD"
-     "\x00\x80"
-     "\x5A\xCD\x00\x80\xBF\x99\x03\x00\xBE\x91\x03\x00\xBD\x99\x03\x00\xBC\x41"
-     "\x02\x00"
-     "\x5E\xCD\x00\x80\xA3\x19\x02\x00\x86\x68\x05\x00\x87\x00\x03\x00\xA6\x39"
-     "\x02\x00"
-     "\x62\xCD\x00\x80\x66\xCD\x00\x80\xA5\x39\x02\x00\xAA\x15\x02\x00\xAB\x1D"
-     "\x02\x00"
-     "\x6A\xCD\x00\x80\x6E\xCD\x00\x80\xAE\xD5\x03\x00\xAF\xDD\x03\x00\xAC\x05"
-     "\x02\x00"
-     "\xAD\xDD\x03\x00\x72\xCD\x00\x80\x76\xCD\x00\x80\x7A\xCD\x00\x80\x7E\xCD"
-     "\x00\x80"
-     "\x82\xCD\x00\x80\xEF\x34\x00\x00\x86\xCD\x00\x80\x8A\xCD\x00\x80\xBE\x2C"
-     "\x05\x00"
-     "\xE3\xDC\x00\x00\x8E\xCD\x00\x80\xE1\xE0\x01\x00\x96\xCD\x00\x80\x9A\xCD"
-     "\x00\x80"
-     "\x9E\xCD\x00\x80\xA2\xCD\x00\x80\xE1\xCC\x0F\x00\x84\x48\x02\x00\xE3\x14"
-     "\x0E\x00"
-     "\x84\x44\x05\x00\x80\x49\x00\x00\x81\x49\x00\x00\x82\x71\x00\x00\xA6\xCD"
-     "\x00\x80"
-     "\xAA\xCD\x00\x80\xAE\xCD\x00\x80\xB2\xCD\x00\x80\xB6\xCD\x00\x80\xBA\xCD"
-     "\x00\x80"
-     "\xBE\xCD\x00\x80\xEF\x20\x0F\x00\xC2\xCD\x00\x80\xC6\xCD\x00\x80\xCA\xCD"
-     "\x00\x80"
-     "\xCE\xCD\x00\x80\xD2\xCD\x00\x80\xD6\xCD\x00\x80\xDA\xCD\x00\x80\xDE\xCD"
-     "\x00\x80"
-     "\xEF\xF8\x01\x00\xBE\xAC\x04\x00\xE1\xDC\x0E\x00\xE2\xCD\x00\x80\xE3\x1C"
-     "\x0E\x00"
-     "\xE6\xCD\x00\x80\xEA\xCD\x00\x80\xEE\xCD\x00\x80\xF2\xCD\x00\x80\xB3\xE5"
-     "\x01\x00"
-     "\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D\x00\x00\xF6\xCD\x00\x80\xB6\xE5"
-     "\x01\x00"
-     "\xB5\xF5\x01\x00\xFA\xCD\x00\x80\xBB\x65\x01\x00\xBA\xD9\x01\x00\xFE\xCD"
-     "\x00\x80"
-     "\x06\xCE\x00\x80\xBF\x5D\x01\x00\xBE\x5D\x01\x00\xBD\x65\x01\x00\xBC\x6D"
-     "\x01\x00"
-     "\xA8\xED\x05\x00\xA9\x25\x06\x00\xAA\x2D\x06\x00\xAB\x25\x06\x00\xAC\x3D"
-     "\x06\x00"
-     "\xAD\x25\x06\x00\xAE\x2D\x06\x00\xAF\x25\x06\x00\x92\xCD\x00\x80\x02\xCE"
-     "\x00\x80"
-     "\x86\x88\x00\x00\x87\xA0\x03\x00\x0A\xCE\x00\x80\x0E\xCE\x00\x80\x12\xCE"
-     "\x00\x80"
-     "\x16\xCE\x00\x80\xB8\x99\x06\x00\xB9\xAD\x06\x00\xBA\xA5\x06\x00\xBB\x4D"
-     "\x07\x00"
-     "\xBC\x55\x07\x00\xBD\x5D\x07\x00\xBE\x55\x07\x00\xBF\x4D\x07\x00\xB0\x5D"
-     "\x06\x00"
-     "\xB1\xE1\x06\x00\xB2\xF9\x06\x00\xB3\xF1\x06\x00\xB4\xD9\x06\x00\xB5\xD9"
-     "\x06\x00"
-     "\xB6\xB9\x06\x00\xB7\xA9\x06\x00\xA3\xA5\x06\x00\x1A\xCE\x00\x80\x1E\xCE"
-     "\x00\x80"
-     "\x22\xCE\x00\x80\x26\xCE\x00\x80\xA6\xA5\x06\x00\xA5\xB5\x06\x00\x2A\xCE"
-     "\x00\x80"
-     "\xAB\x25\x06\x00\xAA\x99\x06\x00\x2E\xCE\x00\x80\x32\xCE\x00\x80\xAF\x1D"
-     "\x06\x00"
-     "\xAE\x1D\x06\x00\xAD\x25\x06\x00\xAC\x2D\x06\x00\x36\xCE\x00\x80\xB3\xFD"
-     "\x07\x00"
-     "\x22\xCD\x00\x80\x3A\xCE\x00\x80\xB6\xB1\x07\x00\x3E\xCE\x00\x80\x42\xCE"
-     "\x00\x80"
-     "\xB5\x81\x07\x00\xBA\x8D\x07\x00\xBB\x95\x07\x00\x46\xCE\x00\x80\x4A\xCE"
-     "\x00\x80"
-     "\xBE\x71\x07\x00\xBF\x71\x07\x00\xBC\x8D\x07\x00\xBD\x81\x07\x00\xA8\x31"
-     "\x06\x00"
-     "\xA9\x31\x06\x00\xAA\x31\x06\x00\xAB\x31\x06\x00\xAC\x4D\x06\x00\xAD\x75"
-     "\x06\x00"
-     "\xAE\x7D\x06\x00\xAF\x75\x06\x00\x4E\xCE\x00\x80\x52\xCE\x00\x80\x56\xCE"
-     "\x00\x80"
-     "\x5A\xCE\x00\x80\x5E\xCE\x00\x80\x62\xCE\x00\x80\x66\xCE\x00\x80\x6A\xCE"
-     "\x00\x80"
-     "\xB8\xD1\x06\x00\xB9\xD1\x06\x00\xBA\xD1\x06\x00\xBB\xE5\x06\x00\xBC\xE1"
-     "\x06\x00"
-     "\xBD\xE1\x06\x00\xBE\xE1\x06\x00\xBF\xE1\x06\x00\xB0\x0D\x06\x00\xB1\x15"
-     "\x06\x00"
-     "\xB2\x1D\x06\x00\xB3\x11\x06\x00\xB4\xF1\x06\x00\xB5\xF1\x06\x00\xB6\xF1"
-     "\x06\x00"
-     "\xB7\xF1\x06\x00\xA3\xB9\x06\x00\x82\x15\x00\x00\x81\xB5\x00\x00\x80\xB5"
-     "\x00\x00"
-     "\x6E\xCE\x00\x80\xA6\xF5\x06\x00\xA5\xC5\x06\x00\xBE\x90\x01\x00\xAB\xD1"
-     "\x06\x00"
-     "\xAA\xC9\x06\x00\x84\xA0\x02\x00\xBF\x00\x3C\x00\xAF\x35\x06\x00\xAE\x35"
-     "\x06\x00"
-     "\xAD\xC5\x06\x00\xAC\xC9\x06\x00\x85\x84\x00\x00\xB3\xF1\x01\x00\x86\x08"
-     "\x3F\x00"
-     "\x87\x9C\x00\x00\xB6\x55\x01\x00\x76\xCE\x00\x80\x7A\xCE\x00\x80\xB5\xE1"
-     "\x01\x00"
-     "\xBA\x71\x01\x00\xBB\x79\x01\x00\x7E\xCE\x00\x80\x82\xCE\x00\x80\xBE\x31"
-     "\x01\x00"
-     "\xBF\x39\x01\x00\xBC\x51\x01\x00\xBD\x51\x01\x00\xA8\xED\x05\x00\xA9\x35"
-     "\x02\x00"
-     "\xAA\x3D\x02\x00\xAB\x35\x02\x00\xAC\x2D\x02\x00\xAD\x91\x02\x00\xAE\x91"
-     "\x02\x00"
-     "\xAF\x91\x02\x00\x86\xCE\x00\x80\x8A\xCE\x00\x80\x8E\xCE\x00\x80\x92\xCE"
-     "\x00\x80"
-     "\xA2\x11\x02\x00\xA3\x11\x02\x00\xA0\x91\x06\x00\xA1\xF5\x02\x00\xB8\xA9"
-     "\x02\x00"
-     "\xB9\xA9\x02\x00\xBA\x69\x03\x00\xBB\x69\x03\x00\xBC\x79\x03\x00\xBD\x79"
-     "\x03\x00"
-     "\xBE\x69\x03\x00\xBF\x69\x03\x00\xB0\xF1\x02\x00\xB1\xF9\x02\x00\xB2\xC1"
-     "\x02\x00"
-     "\xB3\xC1\x02\x00\xB4\xB1\x02\x00\xB5\xB9\x02\x00\xB6\xAD\x02\x00\xB7\xA5"
-     "\x02\x00"
-     "\xA3\xB5\x02\x00\x8F\x8D\x1A\x00\x96\xCE\x00\x80\x9A\xCE\x00\x80\x9E\xCE"
-     "\x00\x80"
-     "\xA6\x11\x02\x00\xA5\xA5\x02\x00\xA2\xCE\x00\x80\xAB\x3D\x02\x00\xAA\x35"
-     "\x02\x00"
-     "\xA6\xCE\x00\x80\xAA\xCE\x00\x80\xAF\x7D\x02\x00\xAE\x75\x02\x00\xAD\x15"
-     "\x02\x00"
-     "\xAC\x15\x02\x00\x9E\x05\x06\x00\x9F\xE5\x06\x00\x9C\xE9\x0A\x00\x9D\xED"
-     "\x0B\x00"
-     "\x9A\x45\x0E\x00\x9B\x65\x0A\x00\x98\x59\x0E\x00\x99\x51\x0E\x00\x96\xC5"
-     "\x12\x00"
-     "\x97\xBD\x13\x00\x94\x75\x12\x00\x95\xCD\x12\x00\x92\xD1\x16\x00\x93\x71"
-     "\x16\x00"
-     "\x90\xF5\x1A\x00\x91\x75\x16\x00\xAE\xCE\x00\x80\xE1\xD0\x00\x00\x84\x44"
-     "\x3D\x00"
-     "\xE3\x10\x24\x00\xB2\xCE\x00\x80\xB6\xCE\x00\x80\xBA\xCE\x00\x80\xBE\xCE"
-     "\x00\x80"
-     "\xC2\xCE\x00\x80\xC6\xCE\x00\x80\xCA\xCE\x00\x80\xCE\xCE\x00\x80\xD2\xCE"
-     "\x00\x80"
-     "\xD6\xCE\x00\x80\xDA\xCE\x00\x80\xEF\xCC\x24\x00\xDE\xCE\x00\x80\xE3\x10"
-     "\x06\x00"
-     "\xBE\x00\x3D\x00\xE1\xD8\x07\x00\x80\x65\x00\x00\x81\x6D\x00\x00\x82\x19"
-     "\x00\x00"
-     "\xE6\xCE\x00\x80\x86\xC0\x3C\x00\x87\x44\x3C\x00\xEA\xCE\x00\x80\xEE\xCE"
-     "\x00\x80"
-     "\xF2\xCE\x00\x80\xEF\xAC\x06\x00\xF6\xCE\x00\x80\xFA\xCE\x00\x80\xB1\x85"
-     "\x2B\x00"
-     "\xB0\x01\x28\x00\xB3\xA1\x2A\x00\xB2\x9D\x2B\x00\xB5\x91\x27\x00\xB4\x01"
-     "\x24\x00"
-     "\xEF\xFC\x00\x00\xEF\xC4\x06\x00\xE1\x90\x01\x00\xE1\x74\x07\x00\xE3\x10"
-     "\x00\x00"
-     "\xE3\xF0\x06\x00\xFE\xCE\x00\x80\x02\xCF\x00\x80\x84\xF4\x3C\x00\x06\xCF"
-     "\x00\x80"
-     "\xA1\x95\x3F\x00\x0A\xCF\x00\x80\xA3\x21\x3F\x00\xA2\x49\x3F\x00\xA5\xD5"
-     "\x3B\x00"
-     "\xA4\xD5\x3B\x00\xA7\x01\x34\x00\xA6\x01\x3B\x00\xA9\x41\x37\x00\xA8\xE1"
-     "\x37\x00"
-     "\xAB\xE9\x33\x00\xAA\x01\x30\x00\xAD\x01\x2C\x00\xAC\xC1\x33\x00\xAF\xA1"
-     "\x2F\x00"
-     "\xAE\xF9\x2F\x00\xB0\x89\x3E\x00\xB1\x89\x3E\x00\xB2\x99\x3E\x00\xB3\x91"
-     "\x3E\x00"
-     "\xB4\xB9\x3E\x00\xB5\xB9\x3E\x00\xB6\x19\x3F\x00\xB7\x0D\x3F\x00\xB8\x35"
-     "\x3F\x00"
-     "\xB9\x3D\x3F\x00\xBA\x35\x3F\x00\xBB\x0D\x3F\x00\xBC\x15\x3F\x00\xBD\x1D"
-     "\x3F\x00"
-     "\xBE\x15\x3F\x00\xBF\x0D\x3F\x00\x0E\xCF\x00\x80\x12\xCF\x00\x80\x16\xCF"
-     "\x00\x80"
-     "\x1A\xCF\x00\x80\x1E\xCF\x00\x80\x72\xCE\x00\x80\x22\xCF\x00\x80\x26\xCF"
-     "\x00\x80"
-     "\xA8\xB5\x3D\x00\xA9\xBD\x3D\x00\xAA\x81\x3D\x00\xAB\x81\x3D\x00\xAC\x51"
-     "\x3E\x00"
-     "\xAD\x51\x3E\x00\xAE\x51\x3E\x00\xAF\x51\x3E\x00\xB3\x69\x3D\x00\xE2\xCE"
-     "\x00\x80"
-     "\x2A\xCF\x00\x80\x2E\xCF\x00\x80\x32\xCF\x00\x80\xB6\xB1\x3E\x00\xB5\xB9"
-     "\x3E\x00"
-     "\x36\xCF\x00\x80\xBB\x2D\x3E\x00\xBA\x2D\x3E\x00\x3A\xCF\x00\x80\x3E\xCF"
-     "\x00\x80"
-     "\xBF\x1D\x3E\x00\xBE\x1D\x3E\x00\xBD\x25\x3E\x00\xBC\x2D\x3E\x00\x42\xCF"
-     "\x00\x80"
-     "\xA3\x2D\x3D\x00\x46\xCF\x00\x80\x4A\xCF\x00\x80\xA6\xF5\x3E\x00\x4E\xCF"
-     "\x00\x80"
-     "\xBE\x90\x03\x00\xA5\xFD\x3E\x00\xAA\x69\x3E\x00\xAB\x69\x3E\x00\x52\xCF"
-     "\x00\x80"
-     "\x56\xCF\x00\x80\xAE\x59\x3E\x00\xAF\x59\x3E\x00\xAC\x69\x3E\x00\xAD\x61"
-     "\x3E\x00"
-     "\xA8\x5D\x3E\x00\xA9\x6D\x3E\x00\xAA\x65\x3E\x00\xAB\x7D\x3E\x00\xAC\x69"
-     "\x3E\x00"
-     "\xAD\x69\x3E\x00\xAE\xA9\x3E\x00\xAF\xA9\x3E\x00\x80\xE9\x01\x00\x81\xE9"
-     "\x01\x00"
-     "\x82\xFD\x01\x00\x5A\xCF\x00\x80\x86\x90\x01\x00\x87\xB4\x01\x00\x5E\xCF"
-     "\x00\x80"
-     "\x62\xCF\x00\x80\xB8\x1D\x01\x00\xB9\xCD\x01\x00\xBA\xC5\x01\x00\xBB\xDD"
-     "\x01\x00"
-     "\xBC\xCD\x01\x00\xBD\xFD\x01\x00\xBE\xF5\x01\x00\xBF\x9D\x01\x00\xB0\xDD"
-     "\x3E\x00"
-     "\xB1\xA1\x3E\x00\xB2\x29\x01\x00\xB3\x29\x01\x00\xB4\x39\x01\x00\xB5\x39"
-     "\x01\x00"
-     "\xB6\x29\x01\x00\xB7\x25\x01\x00\xB3\xED\x3E\x00\x66\xCF\x00\x80\x6A\xCF"
-     "\x00\x80"
-     "\x6E\xCF\x00\x80\x72\xCF\x00\x80\xB6\xE9\x3E\x00\xB5\xE1\x3E\x00\x76\xCF"
-     "\x00\x80"
-     "\xBB\x65\x3E\x00\xBA\x19\x3E\x00\x7A\xCF\x00\x80\x84\xA0\x01\x00\xBF\xFD"
-     "\x01\x00"
-     "\xBE\xFD\x01\x00\xBD\xFD\x01\x00\xBC\xFD\x01\x00\x7E\xCF\x00\x80\xA3\xA9"
-     "\x3E\x00"
-     "\x82\xCF\x00\x80\x86\xCF\x00\x80\xA6\xAD\x3E\x00\x8A\xCF\x00\x80\x8E\xCF"
-     "\x00\x80"
-     "\xA5\xA5\x3E\x00\xAA\x5D\x3E\x00\xAB\x21\x3E\x00\x92\xCF\x00\x80\x96\xCF"
-     "\x00\x80"
-     "\xAE\xB9\x01\x00\xAF\xB9\x01\x00\xAC\xB9\x01\x00\xAD\xB9\x01\x00\xA8\xED"
-     "\x02\x00"
-     "\xA9\x29\x03\x00\xAA\x31\x03\x00\xAB\x31\x03\x00\xAC\x95\x03\x00\xAD\x9D"
-     "\x03\x00"
-     "\xAE\x95\x03\x00\xAF\x8D\x03\x00\xBE\x00\x04\x00\x9A\xCF\x00\x80\x9E\xCF"
-     "\x00\x80"
-     "\xA2\xCF\x00\x80\xA6\xCF\x00\x80\xAA\xCF\x00\x80\xAE\xCF\x00\x80\xB2\xCF"
-     "\x00\x80"
-     "\xB8\xAD\x03\x00\xB9\x61\x00\x00\xBA\x61\x00\x00\xBB\x61\x00\x00\xBC\x61"
-     "\x00\x00"
-     "\xBD\x61\x00\x00\xBE\x61\x00\x00\xBF\x61\x00\x00\xB0\xFD\x03\x00\xB1\xCD"
-     "\x03\x00"
-     "\xB2\xA9\x03\x00\xB3\xA5\x03\x00\xB4\xBD\x03\x00\xB5\xA1\x03\x00\xB6\xA1"
-     "\x03\x00"
-     "\xB7\x95\x03\x00\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\xB6\xCF"
-     "\x00\x80"
-     "\xBA\xCF\x00\x80\xBE\xCF\x00\x80\xC6\xCF\x00\x80\xCA\xCF\x00\x80\x86\x20"
-     "\x04\x00"
-     "\x87\xE4\x03\x00\xCE\xCF\x00\x80\xD2\xCF\x00\x80\xD6\xCF\x00\x80\xDA\xCF"
-     "\x00\x80"
-     "\xDE\xCF\x00\x80\xE2\xCF\x00\x80\xB3\xF9\x03\x00\xE6\xCF\x00\x80\xEA\xCF"
-     "\x00\x80"
-     "\xEE\xCF\x00\x80\xF2\xCF\x00\x80\xB6\x5D\x03\x00\xB5\x5D\x03\x00\xF6\xCF"
-     "\x00\x80"
-     "\xBB\x49\x03\x00\xBA\x79\x03\x00\xFA\xCF\x00\x80\xFE\xCF\x00\x80\xBF\x25"
-     "\x03\x00"
-     "\xBE\x39\x03\x00\xBD\x31\x03\x00\xBC\x3D\x03\x00\xE1\x78\x3E\x00\x02\xD0"
-     "\x00\x80"
-     "\xE3\xA0\x3E\x00\x06\xD0\x00\x80\x0A\xD0\x00\x80\x0E\xD0\x00\x80\x12\xD0"
-     "\x00\x80"
-     "\x16\xD0\x00\x80\x84\x2C\x05\x00\x1A\xD0\x00\x80\xBE\xA4\x04\x00\x1E\xD0"
-     "\x00\x80"
-     "\x22\xD0\x00\x80\x26\xD0\x00\x80\xEF\x18\x01\x00\x2A\xD0\x00\x80\x80\x75"
-     "\x00\x00"
-     "\x81\x79\x00\x00\x82\x19\x00\x00\x84\xC4\x04\x00\x2E\xD0\x00\x80\xEF\x8C"
-     "\x03\x00"
-     "\x87\x94\x05\x00\x86\x7C\x04\x00\x36\xD0\x00\x80\xE3\xD8\x00\x00\x3A\xD0"
-     "\x00\x80"
-     "\xE1\x44\x00\x00\x3E\xD0\x00\x80\x42\xD0\x00\x80\x46\xD0\x00\x80\x4A\xD0"
-     "\x00\x80"
-     "\x8B\xA5\x2F\x00\x8A\xA5\x2F\x00\x4E\xD0\x00\x80\x52\xD0\x00\x80\x8F\x05"
-     "\x33\x00"
-     "\x8E\x19\x33\x00\x8D\x19\x33\x00\x8C\xC5\x2E\x00\x83\xAD\x27\x00\x82\xAD"
-     "\x27\x00"
-     "\x56\xD0\x00\x80\x5A\xD0\x00\x80\x87\x2D\x2F\x00\x86\x31\x2B\x00\x85\x99"
-     "\x2B\x00"
-     "\x84\x35\x2B\x00\x9B\xE1\x3F\x00\x9A\x0D\x3F\x00\x5E\xD0\x00\x80\x62\xD0"
-     "\x00\x80"
-     "\xE1\x00\x3C\x00\x66\xD0\x00\x80\xE3\xC0\x3E\x00\x9C\xF9\x3F\x00\x93\xC1"
-     "\x37\x00"
-     "\x92\x55\x37\x00\xEF\x58\x3E\x00\x6A\xD0\x00\x80\x97\x59\x3B\x00\x96\xD5"
-     "\x3B\x00"
-     "\x95\x3D\x3B\x00\x94\xDD\x36\x00\xC2\xCF\x00\x80\x32\xD0\x00\x80\x6E\xD0"
-     "\x00\x80"
-     "\x72\xD0\x00\x80\x76\xD0\x00\x80\x7A\xD0\x00\x80\x7E\xD0\x00\x80\x82\xD0"
-     "\x00\x80"
-     "\xA8\xB9\x06\x00\xA9\xB9\x06\x00\xAA\x81\x06\x00\xAB\x81\x06\x00\xAC\x81"
-     "\x06\x00"
-     "\xAD\x81\x06\x00\xAE\x81\x06\x00\xAF\x81\x06\x00\xB0\x85\x06\x00\xB1\x8D"
-     "\x06\x00"
-     "\xB2\x85\x06\x00\xB3\x9D\x06\x00\xB4\x8D\x06\x00\xB5\xBD\x06\x00\xB6\xB5"
-     "\x06\x00"
-     "\xB7\x59\x07\x00\xB8\x6D\x07\x00\xB9\x75\x07\x00\xBA\x7D\x07\x00\xBB\x75"
-     "\x07\x00"
-     "\xBC\x6D\x07\x00\xBD\x51\x07\x00\xBE\x55\x07\x00\xBF\x4D\x07\x00\x86\xD0"
-     "\x00\x80"
-     "\x8A\xD0\x00\x80\xAA\x15\x05\x00\xAB\x25\x05\x00\xAC\x51\x05\x00\xAD\x5D"
-     "\x05\x00"
-     "\xAE\x55\x05\x00\xAF\x49\x05\x00\x8E\xD0\x00\x80\x92\xD0\x00\x80\x96\xD0"
-     "\x00\x80"
-     "\xA3\x95\x05\x00\x9A\xD0\x00\x80\xA5\x31\x05\x00\xA6\x31\x05\x00\x9E\xD0"
-     "\x00\x80"
-     "\xB3\x69\x06\x00\x82\x79\x00\x00\x81\x79\x00\x00\x80\x65\x00\x00\xBE\x5C"
-     "\x00\x00"
-     "\xB6\x61\x06\x00\xB5\x69\x06\x00\xA2\xD0\x00\x80\xBB\x29\x06\x00\xBA\x21"
-     "\x06\x00"
-     "\x86\xC8\x00\x00\x87\xAC\x00\x00\xBF\x19\x06\x00\xBE\x29\x06\x00\xBD\x29"
-     "\x06\x00"
-     "\xBC\x31\x06\x00\xA3\xA9\x07\x00\xA6\xD0\x00\x80\xAA\xD0\x00\x80\xAE\xD0"
-     "\x00\x80"
-     "\xB2\xD0\x00\x80\xA6\xA1\x07\x00\xA5\xA9\x07\x00\xB6\xD0\x00\x80\xAB\xE9"
-     "\x07\x00"
-     "\xAA\xE1\x07\x00\xBA\xD0\x00\x80\xBE\xD0\x00\x80\xAF\xD9\x07\x00\xAE\xE9"
-     "\x07\x00"
-     "\xAD\xE9\x07\x00\xAC\xF1\x07\x00\xC2\xD0\x00\x80\xB3\x35\x06\x00\xC6\xD0"
-     "\x00\x80"
-     "\xCA\xD0\x00\x80\xB6\x2D\x06\x00\xCE\xD0\x00\x80\xD2\xD0\x00\x80\xB5\x25"
-     "\x06\x00"
-     "\xBA\xC9\x06\x00\xBB\xC9\x06\x00\xD6\xD0\x00\x80\xDA\xD0\x00\x80\xBE\xA5"
-     "\x06\x00"
-     "\xBF\x99\x06\x00\xBC\xB1\x06\x00\xBD\xB1\x06\x00\xA8\xE1\x06\x00\xA9\xED"
-     "\x06\x00"
-     "\xAA\xF9\x06\x00\xAB\xF9\x06\x00\xAC\xE9\x06\x00\xAD\xE9\x06\x00\xAE\x11"
-     "\x06\x00"
-     "\xAF\x11\x06\x00\xDE\xD0\x00\x80\xE2\xD0\x00\x80\xE6\xD0\x00\x80\xEA\xD0"
-     "\x00\x80"
-     "\xEE\xD0\x00\x80\xF2\xD0\x00\x80\xF6\xD0\x00\x80\xFA\xD0\x00\x80\xB8\xD1"
-     "\x01\x00"
-     "\xB9\xD1\x01\x00\xBA\xD1\x01\x00\xBB\xD1\x01\x00\xBC\x51\x01\x00\xBD\x51"
-     "\x01\x00"
-     "\xBE\x79\x01\x00\xBF\x79\x01\x00\xB0\x49\x06\x00\xB1\x49\x06\x00\xB2\x21"
-     "\x06\x00"
-     "\xB3\x21\x06\x00\xB4\x25\x06\x00\xB5\x19\x06\x00\xB6\xF1\x01\x00\xB7\xF1"
-     "\x01\x00"
-     "\xA3\x71\x05\x00\xFE\xD0\x00\x80\x02\xD1\x00\x80\xBE\xCC\x02\x00\x06\xD1"
-     "\x00\x80"
-     "\xA6\x69\x05\x00\xA5\x61\x05\x00\x0A\xD1\x00\x80\xAB\x8D\x05\x00\xAA\x8D"
-     "\x05\x00"
-     "\x84\xE4\x03\x00\x0E\xD1\x00\x80\xAF\xDD\x05\x00\xAE\xE1\x05\x00\xAD\xF5"
-     "\x05\x00"
-     "\xAC\xF5\x05\x00\x80\x51\x00\x00\x81\x59\x00\x00\x82\x65\x00\x00\xA3\x41"
-     "\x03\x00"
-     "\x16\xD1\x00\x80\xA5\x49\x03\x00\xA6\x71\x03\x00\x1A\xD1\x00\x80\x86\x40"
-     "\x0C\x00"
-     "\x87\x14\x03\x00\xAA\x21\x03\x00\xAB\x21\x03\x00\xAC\x25\x03\x00\xAD\x2D"
-     "\x03\x00"
-     "\xAE\x11\x03\x00\xAF\x11\x03\x00\xA8\x85\x03\x00\xA9\x8D\x03\x00\xAA\x85"
-     "\x03\x00"
-     "\xAB\x9D\x03\x00\xAC\x8D\x03\x00\xAD\xBD\x03\x00\xAE\xB5\x03\x00\xAF\xDD"
-     "\x03\x00"
-     "\x1E\xD1\x00\x80\x22\xD1\x00\x80\x26\xD1\x00\x80\x2A\xD1\x00\x80\x2E\xD1"
-     "\x00\x80"
-     "\x32\xD1\x00\x80\x36\xD1\x00\x80\x3A\xD1\x00\x80\xB8\x69\x03\x00\xB9\x69"
-     "\x03\x00"
-     "\xBA\x79\x03\x00\xBB\x79\x03\x00\xBC\x69\x03\x00\xBD\x69\x03\x00\xBE\xDD"
-     "\x00\x00"
-     "\xBF\xD5\x00\x00\xB0\xA9\x03\x00\xB1\xA9\x03\x00\xB2\xBD\x03\x00\xB3\xB5"
-     "\x03\x00"
-     "\xB4\x99\x03\x00\xB5\x99\x03\x00\xB6\x59\x03\x00\xB7\x59\x03\x00\xB3\x0D"
-     "\x02\x00"
-     "\x3E\xD1\x00\x80\x42\xD1\x00\x80\x46\xD1\x00\x80\x4A\xD1\x00\x80\xB6\x3D"
-     "\x02\x00"
-     "\xB5\x05\x02\x00\x4E\xD1\x00\x80\xBB\x6D\x02\x00\xBA\x6D\x02\x00\x52\xD1"
-     "\x00\x80"
-     "\x56\xD1\x00\x80\xBF\x5D\x02\x00\xBE\x5D\x02\x00\xBD\x61\x02\x00\xBC\x69"
-     "\x02\x00"
-     "\x5A\xD1\x00\x80\x5E\xD1\x00\x80\x62\xD1\x00\x80\x66\xD1\x00\x80\xE3\x68"
-     "\x03\x00"
-     "\x6A\xD1\x00\x80\xE1\xA8\x01\x00\x6E\xD1\x00\x80\xEF\x94\x03\x00\x72\xD1"
-     "\x00\x80"
-     "\x76\xD1\x00\x80\x7A\xD1\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x7E\xD1\x00\x80\xE1\xC4\x06\x00\xE1\x3C\x07\x00\xE3\xD4\x06\x00\xE3\xBC"
-     "\x06\x00"
-     "\x84\x90\x02\x00\x86\xD1\x00\x80\xBE\x14\x0D\x00\x8A\xD1\x00\x80\x8E\xD1"
-     "\x00\x80"
-     "\x92\xD1\x00\x80\x96\xD1\x00\x80\x9A\xD1\x00\x80\x9E\xD1\x00\x80\xA2\xD1"
-     "\x00\x80"
-     "\xEF\x90\x01\x00\xEF\x6C\x06\x00\x86\x60\x0C\x00\x87\xAC\x0D\x00\xA6\xD1"
-     "\x00\x80"
-     "\xB3\x6D\x01\x00\xAA\xD1\x00\x80\xB5\x7D\x01\x00\xB6\x75\x01\x00\xAE\xD1"
-     "\x00\x80"
-     "\xB2\xD1\x00\x80\xB6\xD1\x00\x80\xBA\x51\x01\x00\xBB\x51\x01\x00\xBC\xF5"
-     "\x00\x00"
-     "\xBD\xFD\x00\x00\xBE\xF5\x00\x00\xBF\xED\x00\x00\xA3\xA1\x0E\x00\x82\xD1"
-     "\x00\x80"
-     "\xBA\xD1\x00\x80\xBE\xD1\x00\x80\xC2\xD1\x00\x80\xA6\xB9\x0E\x00\xA5\xB1"
-     "\x0E\x00"
-     "\xC6\xD1\x00\x80\xAB\x9D\x0E\x00\xAA\x9D\x0E\x00\xCA\xD1\x00\x80\xCE\xD1"
-     "\x00\x80"
-     "\xAF\x21\x0F\x00\xAE\x39\x0F\x00\xAD\x31\x0F\x00\xAC\x39\x0F\x00\xD2\xD1"
-     "\x00\x80"
-     "\xD6\xD1\x00\x80\xDA\xD1\x00\x80\xDE\xD1\x00\x80\xE2\xD1\x00\x80\xE3\xD4"
-     "\x0D\x00"
-     "\xE6\xD1\x00\x80\xE1\x24\x0C\x00\x82\x15\x00\x00\xEF\x04\x0D\x00\x80\x31"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\xEA\xD1\x00\x80\xEE\xD1\x00\x80\xBE\x30\x03\x00\x84\x30"
-     "\x03\x00"
-     "\xA8\xD1\x0E\x00\xA9\xD9\x0E\x00\xAA\x29\x0E\x00\xAB\x29\x0E\x00\xAC\x39"
-     "\x0E\x00"
-     "\xAD\x39\x0E\x00\xAE\x29\x0E\x00\xAF\x29\x0E\x00\xF2\xD1\x00\x80\x84\xF8"
-     "\x00\x00"
-     "\x86\x0C\x00\x00\x87\x8C\x00\x00\xF6\xD1\x00\x80\xFA\xD1\x00\x80\xFE\xD1"
-     "\x00\x80"
-     "\x02\xD2\x00\x80\xB8\xED\x0F\x00\xB9\x85\x0F\x00\xBA\x8D\x0F\x00\xBB\x85"
-     "\x0F\x00"
-     "\xBC\x9D\x0F\x00\xBD\x81\x0F\x00\xBE\x85\x0F\x00\xBF\xB9\x0F\x00\xB0\x59"
-     "\x0E\x00"
-     "\xB1\x59\x0E\x00\xB2\xED\x0F\x00\xB3\xE5\x0F\x00\xB4\xFD\x0F\x00\xB5\xE5"
-     "\x0F\x00"
-     "\xB6\xE5\x0F\x00\xB7\xD5\x0F\x00\xA8\x95\x0E\x00\xA9\x99\x0E\x00\xAA\xA9"
-     "\x0E\x00"
-     "\xAB\xA9\x0E\x00\xAC\xB9\x0E\x00\xAD\xB9\x0E\x00\xAE\xA9\x0E\x00\xAF\xA9"
-     "\x0E\x00"
-     "\x06\xD2\x00\x80\x0A\xD2\x00\x80\x0E\xD2\x00\x80\x12\xD2\x00\x80\x16\xD2"
-     "\x00\x80"
-     "\x1A\xD2\x00\x80\x1E\xD2\x00\x80\x22\xD2\x00\x80\xB8\x5D\x01\x00\xB9\x6D"
-     "\x01\x00"
-     "\xBA\x65\x01\x00\xBB\x0D\x01\x00\xBC\x15\x01\x00\xBD\x1D\x01\x00\xBE\x15"
-     "\x01\x00"
-     "\xBF\x0D\x01\x00\xB0\xA9\x0E\x00\xB1\xB5\x0E\x00\xB2\xB5\x0E\x00\xB3\x8D"
-     "\x0E\x00"
-     "\xB4\x95\x0E\x00\xB5\x7D\x01\x00\xB6\x75\x01\x00\xB7\x6D\x01\x00\xB3\x21"
-     "\x0E\x00"
-     "\x26\xD2\x00\x80\x2A\xD2\x00\x80\x2E\xD2\x00\x80\x32\xD2\x00\x80\xB6\x39"
-     "\x0E\x00"
-     "\xB5\x31\x0E\x00\x36\xD2\x00\x80\xBB\x71\x0E\x00\xBA\x6D\x0E\x00\x3A\xD2"
-     "\x00\x80"
-     "\x3E\xD2\x00\x80\xBF\xB9\x01\x00\xBE\x51\x0E\x00\xBD\x59\x0E\x00\xBC\x61"
-     "\x0E\x00"
-     "\x42\xD2\x00\x80\xA3\x65\x0E\x00\x46\xD2\x00\x80\x4A\xD2\x00\x80\xA6\x7D"
-     "\x0E\x00"
-     "\x4E\xD2\x00\x80\x52\xD2\x00\x80\xA5\x75\x0E\x00\xAA\x29\x0E\x00\xAB\x35"
-     "\x0E\x00"
-     "\x56\xD2\x00\x80\x5A\xD2\x00\x80\xAE\x15\x0E\x00\xAF\xFD\x01\x00\xAC\x25"
-     "\x0E\x00"
-     "\xAD\x1D\x0E\x00\xA9\xB5\x02\x00\xA8\xB5\x02\x00\xAB\xB5\x02\x00\xAA\xB5"
-     "\x02\x00"
-     "\xAD\x35\x03\x00\xAC\x35\x03\x00\xAF\x35\x03\x00\xAE\x35\x03\x00\x80\xC9"
-     "\x03\x00"
-     "\x81\xC9\x03\x00\x82\x05\x00\x00\x5E\xD2\x00\x80\xBE\x64\x02\x00\x66\xD2"
-     "\x00\x80"
-     "\x87\x78\x03\x00\x86\x1C\x04\x00\xB9\xE9\x03\x00\xB8\xD9\x03\x00\xBB\x89"
-     "\x03\x00"
-     "\xBA\xE1\x03\x00\xBD\x9D\x03\x00\xBC\x9D\x03\x00\xBF\x81\x03\x00\xBE\x89"
-     "\x03\x00"
-     "\xB1\x55\x03\x00\xB0\x55\x03\x00\xB3\x55\x03\x00\xB2\x55\x03\x00\xB5\xF9"
-     "\x03\x00"
-     "\xB4\xF1\x03\x00\xB7\xE9\x03\x00\xB6\xF1\x03\x00\x84\x00\x03\x00\xB6\xA1"
-     "\x03\x00"
-     "\xB5\xA9\x03\x00\x84\xC0\x05\x00\xB3\x29\x02\x00\x6A\xD2\x00\x80\x6E\xD2"
-     "\x00\x80"
-     "\x72\xD2\x00\x80\xBF\x29\x03\x00\xBE\x21\x03\x00\xBD\x95\x03\x00\xBC\x95"
-     "\x03\x00"
-     "\xBB\x85\x03\x00\xBA\x85\x03\x00\x76\xD2\x00\x80\x7A\xD2\x00\x80\xA5\xE1"
-     "\x03\x00"
-     "\x7E\xD2\x00\x80\x82\xD2\x00\x80\xA6\xE9\x03\x00\x86\xD2\x00\x80\x8A\xD2"
-     "\x00\x80"
-     "\xA3\x61\x02\x00\x8E\xD2\x00\x80\xAD\xDD\x03\x00\xAC\xDD\x03\x00\xAF\x61"
-     "\x03\x00"
-     "\xAE\x69\x03\x00\x92\xD2\x00\x80\x96\xD2\x00\x80\xAB\xCD\x03\x00\xAA\xCD"
-     "\x03\x00"
-     "\x9A\xD2\x00\x80\x9E\xD2\x00\x80\xA2\xD2\x00\x80\xA6\xD2\x00\x80\xAA\xD2"
-     "\x00\x80"
-     "\xAE\xD2\x00\x80\xB2\xD2\x00\x80\xEF\xD4\x03\x00\xB6\xD2\x00\x80\xE1\x94"
-     "\x01\x00"
-     "\xBA\xD2\x00\x80\xE3\x98\x01\x00\xBE\xD2\x00\x80\xC2\xD2\x00\x80\xC6\xD2"
-     "\x00\x80"
-     "\xCA\xD2\x00\x80\x80\x79\x00\x00\x81\x79\x00\x00\x82\x19\x00\x00\xB3\xD5"
-     "\x01\x00"
-     "\xCE\xD2\x00\x80\xB5\xED\x01\x00\xB6\xE5\x01\x00\xBE\x14\x02\x00\x86\xC0"
-     "\x04\x00"
-     "\x87\x40\x05\x00\xBA\xB9\x01\x00\xBB\xB9\x01\x00\xBC\xAD\x01\x00\xBD\x85"
-     "\x01\x00"
-     "\xBE\x8D\x01\x00\xBF\x7D\x00\x00\xE1\x58\x0E\x00\xD6\xD2\x00\x80\xE3\xE4"
-     "\x01\x00"
-     "\xDA\xD2\x00\x80\xDE\xD2\x00\x80\xE2\xD2\x00\x80\xE6\xD2\x00\x80\xEA\xD2"
-     "\x00\x80"
-     "\xBE\xE8\x04\x00\xEE\xD2\x00\x80\xF2\xD2\x00\x80\xF6\xD2\x00\x80\xFA\xD2"
-     "\x00\x80"
-     "\xFE\xD2\x00\x80\xEF\x38\x01\x00\x02\xD3\x00\x80\x06\xD3\x00\x80\xA3\x5D"
-     "\x01\x00"
-     "\x0A\xD3\x00\x80\x0E\xD3\x00\x80\xA6\x6D\x01\x00\x12\xD3\x00\x80\x12\xD1"
-     "\x00\x80"
-     "\xA5\x65\x01\x00\xAA\x31\x01\x00\xAB\x31\x01\x00\x16\xD3\x00\x80\x1A\xD3"
-     "\x00\x80"
-     "\xAE\x05\x01\x00\xAF\xF5\x00\x00\xAC\x25\x01\x00\xAD\x0D\x01\x00\xD2\xD2"
-     "\x00\x80"
-     "\x1E\xD3\x00\x80\x22\xD3\x00\x80\x26\xD3\x00\x80\x2A\xD3\x00\x80\x2E\xD3"
-     "\x00\x80"
-     "\x32\xD3\x00\x80\x36\xD3\x00\x80\x81\x09\x00\x00\x80\x39\x00\x00\x3A\xD3"
-     "\x00\x80"
-     "\x82\x1D\x00\x00\x3E\xD3\x00\x80\x42\xD3\x00\x80\x46\xD3\x00\x80\x4A\xD3"
-     "\x00\x80"
-     "\x86\xE0\x00\x00\x87\x74\x03\x00\x4E\xD3\x00\x80\x52\xD3\x00\x80\xEF\xF4"
-     "\x05\x00"
-     "\x56\xD3\x00\x80\x5A\xD3\x00\x80\x5E\xD3\x00\x80\xE3\xF0\x04\x00\x62\xD3"
-     "\x00\x80"
-     "\xE1\x4C\x04\x00\x66\xD3\x00\x80\x6A\xD3\x00\x80\x6E\xD3\x00\x80\x72\xD3"
-     "\x00\x80"
-     "\x76\xD3\x00\x80\xA8\xB5\x06\x00\xA9\x11\x06\x00\xAA\x11\x06\x00\xAB\x11"
-     "\x06\x00"
-     "\xAC\x29\x06\x00\xAD\x29\x06\x00\xAE\x5D\x06\x00\xAF\x55\x06\x00\x84\xE4"
-     "\x00\x00"
-     "\x7A\xD3\x00\x80\x7E\xD3\x00\x80\x82\xD3\x00\x80\x86\xD3\x00\x80\x8A\xD3"
-     "\x00\x80"
-     "\x8E\xD3\x00\x80\x92\xD3\x00\x80\xB8\xCD\x07\x00\xB9\xD5\x07\x00\xBA\xDD"
-     "\x07\x00"
-     "\xBB\xD5\x07\x00\xBC\xF1\x07\x00\xBD\xFD\x07\x00\xBE\xE9\x07\x00\xBF\xE9"
-     "\x07\x00"
-     "\xB0\x31\x06\x00\xB1\x31\x06\x00\xB2\x35\x06\x00\xB3\x1D\x06\x00\xB4\x05"
-     "\x06\x00"
-     "\xB5\xF5\x07\x00\xB6\xFD\x07\x00\xB7\xF5\x07\x00\xA8\x85\x06\x00\xA9\x89"
-     "\x06\x00"
-     "\xAA\x99\x06\x00\xAB\x99\x06\x00\xAC\x89\x06\x00\xAD\x89\x06\x00\xAE\xBD"
-     "\x06\x00"
-     "\xAF\xB5\x06\x00\x96\xD3\x00\x80\x9A\xD3\x00\x80\x9E\xD3\x00\x80\xA2\xD3"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\xB1\x00\x00\x82\xB1\x00\x00\xA6\xD3\x00\x80\xB8\xA9"
-     "\x06\x00"
-     "\xB9\xA9\x06\x00\xBA\x4D\x01\x00\xBB\x45\x01\x00\xBC\x5D\x01\x00\xBD\x45"
-     "\x01\x00"
-     "\xBE\x4D\x01\x00\xBF\x45\x01\x00\xB0\xCD\x06\x00\xB1\xD5\x06\x00\xB2\xDD"
-     "\x06\x00"
-     "\xB3\xD1\x06\x00\xB4\xB1\x06\x00\xB5\xBD\x06\x00\xB6\xB5\x06\x00\xB7\xA9"
-     "\x06\x00"
-     "\xB3\x11\x06\x00\xAA\xD3\x00\x80\xAE\xD3\x00\x80\xBE\xCC\x01\x00\xB6\xD3"
-     "\x00\x80"
-     "\xB6\x09\x06\x00\xB5\x01\x06\x00\xBA\xD3\x00\x80\xBB\x11\x06\x00\xBA\x09"
-     "\x06\x00"
-     "\x86\x08\x00\x00\x87\x9C\x00\x00\xBF\x79\x06\x00\xBE\x75\x06\x00\xBD\x05"
-     "\x06\x00"
-     "\xBC\x09\x06\x00\xBE\xD3\x00\x80\xA3\x55\x06\x00\xC2\xD3\x00\x80\xC6\xD3"
-     "\x00\x80"
-     "\xA6\x4D\x06\x00\xCA\xD3\x00\x80\xCE\xD3\x00\x80\xA5\x45\x06\x00\xAA\x4D"
-     "\x06\x00"
-     "\xAB\x55\x06\x00\xD2\xD3\x00\x80\xD6\xD3\x00\x80\xAE\x31\x06\x00\xAF\x3D"
-     "\x06\x00"
-     "\xAC\x4D\x06\x00\xAD\x41\x06\x00\xA8\x39\x02\x00\xA9\x39\x02\x00\xAA\x41"
-     "\x02\x00"
-     "\xAB\x41\x02\x00\xAC\x45\x02\x00\xAD\x4D\x02\x00\xAE\x45\x02\x00\xAF\x79"
-     "\x02\x00"
-     "\xBE\xAC\x1C\x00\xBE\xEC\x1D\x00\xDA\xD3\x00\x80\xDE\xD3\x00\x80\xE2\xD3"
-     "\x00\x80"
-     "\xE6\xD3\x00\x80\xEA\xD3\x00\x80\xEE\xD3\x00\x80\xB8\xF1\x02\x00\xB9\xF1"
-     "\x02\x00"
-     "\xBA\xF1\x02\x00\xBB\xF1\x02\x00\xBC\x71\x03\x00\xBD\x71\x03\x00\xBE\x71"
-     "\x03\x00"
-     "\xBF\x71\x03\x00\xB0\x0D\x02\x00\xB1\x15\x02\x00\xB2\x1D\x02\x00\xB3\x15"
-     "\x02\x00"
-     "\xB4\x0D\x02\x00\xB5\xD1\x02\x00\xB6\xD1\x02\x00\xB7\xD1\x02\x00\xF2\xD3"
-     "\x00\x80"
-     "\xF6\xD3\x00\x80\xB3\xC1\x02\x00\xFA\xD3\x00\x80\xB5\xC5\x02\x00\xFE\xD3"
-     "\x00\x80"
-     "\x02\xD4\x00\x80\xB6\xC5\x02\x00\x06\xD4\x00\x80\x0A\xD4\x00\x80\xBB\xD9"
-     "\x02\x00"
-     "\xBA\xC1\x02\x00\xBD\xC9\x02\x00\xBC\xC1\x02\x00\xBF\x1D\x02\x00\xBE\xC1"
-     "\x02\x00"
-     "\x0E\xD4\x00\x80\x12\xD4\x00\x80\x16\xD4\x00\x80\xA3\x85\x02\x00\x1A\xD4"
-     "\x00\x80"
-     "\xA5\x81\x02\x00\xA6\x81\x02\x00\x1E\xD4\x00\x80\x84\xEC\x03\x00\x26\xD4"
-     "\x00\x80"
-     "\xAA\x85\x02\x00\xAB\x9D\x02\x00\xAC\x85\x02\x00\xAD\x8D\x02\x00\xAE\x85"
-     "\x02\x00"
-     "\xAF\x59\x02\x00\xA8\x91\x02\x00\xA9\x99\x02\x00\xAA\xC9\x02\x00\xAB\xC5"
-     "\x02\x00"
-     "\xAC\xDD\x02\x00\xAD\xC1\x02\x00\xAE\xC1\x02\x00\xAF\xF5\x02\x00\x86\x00"
-     "\x1D\x00"
-     "\x87\x64\x1D\x00\x2A\xD4\x00\x80\x2E\xD4\x00\x80\x80\xF9\x01\x00\x81\xE5"
-     "\x01\x00"
-     "\x82\xE5\x01\x00\x32\xD4\x00\x80\xB8\xC5\x01\x00\xB9\xC9\x01\x00\xBA\xDD"
-     "\x01\x00"
-     "\xBB\xD5\x01\x00\xBC\xF1\x01\x00\xBD\xF1\x01\x00\xBE\x99\x01\x00\xBF\x99"
-     "\x01\x00"
-     "\xB0\x8D\x02\x00\xB1\x4D\x01\x00\xB2\x45\x01\x00\xB3\x5D\x01\x00\xB4\x45"
-     "\x01\x00"
-     "\xB5\x4D\x01\x00\xB6\x45\x01\x00\xB7\xFD\x01\x00\x36\xD4\x00\x80\xE1\xFC"
-     "\x06\x00"
-     "\xEF\xFC\x00\x00\xE3\x10\x07\x00\x3A\xD4\x00\x80\x3E\xD4\x00\x80\x42\xD4"
-     "\x00\x80"
-     "\x46\xD4\x00\x80\x84\x6C\x1C\x00\x4A\xD4\x00\x80\x4E\xD4\x00\x80\x52\xD4"
-     "\x00\x80"
-     "\xE1\xE8\x01\x00\x56\xD4\x00\x80\xE3\x08\x01\x00\xEF\xE4\x06\x00\xB3\x01"
-     "\x02\x00"
-     "\x5A\xD4\x00\x80\x5E\xD4\x00\x80\x62\xD4\x00\x80\x66\xD4\x00\x80\xB6\x01"
-     "\x02\x00"
-     "\xB5\x15\x02\x00\x6A\xD4\x00\x80\xBB\xC9\x01\x00\xBA\x09\x02\x00\x6E\xD4"
-     "\x00\x80"
-     "\x62\xD2\x00\x80\xBF\xC9\x01\x00\xBE\xC1\x01\x00\xBD\xC9\x01\x00\xBC\xD1"
-     "\x01\x00"
-     "\xA3\xC1\x1D\x00\x22\xD4\x00\x80\x72\xD4\x00\x80\x76\xD4\x00\x80\x7A\xD4"
-     "\x00\x80"
-     "\xA6\xC1\x1D\x00\xA5\xD5\x1D\x00\x7E\xD4\x00\x80\xAB\x09\x1E\x00\xAA\xC9"
-     "\x1D\x00"
-     "\x82\xD4\x00\x80\x86\xD4\x00\x80\xAF\x09\x1E\x00\xAE\x01\x1E\x00\xAD\x09"
-     "\x1E\x00"
-     "\xAC\x11\x1E\x00\x80\x6D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\xEF\x94"
-     "\x1C\x00"
-     "\x8A\xD4\x00\x80\x8E\xD4\x00\x80\x92\xD4\x00\x80\x96\xD4\x00\x80\x86\x40"
-     "\x03\x00"
-     "\x87\xC4\x03\x00\x9A\xD4\x00\x80\x9E\xD4\x00\x80\xA2\xD4\x00\x80\xE1\xC4"
-     "\x1F\x00"
-     "\xA6\xD4\x00\x80\xE3\x8C\x1C\x00\xA8\x65\x1E\x00\xA9\xAD\x1F\x00\xAA\xA5"
-     "\x1F\x00"
-     "\xAB\xB1\x1F\x00\xAC\x11\x1F\x00\xAD\x05\x1F\x00\xAE\x0D\x1F\x00\xAF\x05"
-     "\x1F\x00"
-     "\xAA\xD4\x00\x80\xAE\xD4\x00\x80\xB2\xD4\x00\x80\xB6\xD4\x00\x80\xBA\xD4"
-     "\x00\x80"
-     "\xBE\xD4\x00\x80\xC2\xD4\x00\x80\xC6\xD4\x00\x80\xB8\x11\x1F\x00\xB9\x19"
-     "\x1F\x00"
-     "\xBA\x21\x1F\x00\xBB\xC5\x00\x00\xBC\xDD\x00\x00\xBD\xC5\x00\x00\xBE\xC5"
-     "\x00\x00"
-     "\xBF\xF5\x00\x00\xB0\x7D\x1F\x00\xB1\x41\x1F\x00\xB2\x41\x1F\x00\xB3\x41"
-     "\x1F\x00"
-     "\xB4\x45\x1F\x00\xB5\x31\x1F\x00\xB6\x31\x1F\x00\xB7\x31\x1F\x00\xB3\x2D"
-     "\x1E\x00"
-     "\xCA\xD4\x00\x80\xCE\xD4\x00\x80\xD2\xD4\x00\x80\xD6\xD4\x00\x80\xB6\x85"
-     "\x1E\x00"
-     "\xB5\x99\x1E\x00\xDA\xD4\x00\x80\xBB\x85\x1E\x00\xBA\xB9\x1E\x00\xDE\xD4"
-     "\x00\x80"
-     "\xE2\xD4\x00\x80\xBF\xF5\x1E\x00\xBE\x81\x1E\x00\xBD\x95\x1E\x00\xBC\x95"
-     "\x1E\x00"
-     "\xE6\xD4\x00\x80\xA3\x69\x1E\x00\xEA\xD4\x00\x80\xEE\xD4\x00\x80\xA6\xC1"
-     "\x1E\x00"
-     "\xF2\xD4\x00\x80\xF6\xD4\x00\x80\xA5\xDD\x1E\x00\xAA\xFD\x1E\x00\xAB\xC1"
-     "\x1E\x00"
-     "\xFA\xD4\x00\x80\xFE\xD4\x00\x80\xAE\xC5\x1E\x00\xAF\xB1\x1E\x00\xAC\xD1"
-     "\x1E\x00"
-     "\xAD\xD1\x1E\x00\xA8\xDD\x01\x00\xA9\xED\x01\x00\xAA\xE5\x01\x00\xAB\xFD"
-     "\x01\x00"
-     "\xAC\xE9\x01\x00\xAD\x35\x01\x00\xAE\x3D\x01\x00\xAF\x35\x01\x00\x80\xF9"
-     "\x00\x00"
-     "\x81\xCD\x00\x00\x82\xC5\x00\x00\x84\x60\x01\x00\xBE\x78\x01\x00\x06\xD5"
-     "\x00\x80"
-     "\x87\xAC\x00\x00\x86\x9C\x01\x00\xB8\xD1\x00\x00\xB9\xD9\x00\x00\xBA\xE1"
-     "\x00\x00"
-     "\xBB\xE1\x00\x00\xBC\x91\x00\x00\xBD\x9D\x00\x00\xBE\x95\x00\x00\xBF\x89"
-     "\x00\x00"
-     "\xB0\x4D\x01\x00\xB1\x55\x01\x00\xB2\x5D\x01\x00\xB3\x55\x01\x00\xB4\x4D"
-     "\x01\x00"
-     "\xB5\xF1\x00\x00\xB6\xF1\x00\x00\xB7\xF1\x00\x00\xB3\x6D\x02\x00\x0A\xD5"
-     "\x00\x80"
-     "\x0E\xD5\x00\x80\x12\xD5\x00\x80\x16\xD5\x00\x80\xB6\xB5\x02\x00\xB5\x61"
-     "\x02\x00"
-     "\x1A\xD5\x00\x80\xBB\x91\x02\x00\xBA\x89\x02\x00\x1E\xD5\x00\x80\x22\xD5"
-     "\x00\x80"
-     "\xBF\x35\x03\x00\xBE\x89\x02\x00\xBD\x81\x02\x00\xBC\x89\x02\x00\x90\x00"
-     "\x00\x00"
-     "\xA3\x29\x02\x00\x26\xD5\x00\x80\x2A\xD5\x00\x80\xA6\xF1\x02\x00\x2E\xD5"
-     "\x00\x80"
-     "\x32\xD5\x00\x80\xA5\x25\x02\x00\xAA\xCD\x02\x00\xAB\xD5\x02\x00\x36\xD5"
-     "\x00\x80"
-     "\x3A\xD5\x00\x80\xAE\xCD\x02\x00\xAF\x71\x03\x00\xAC\xCD\x02\x00\xAD\xC5"
-     "\x02\x00"
-     "\xA8\x35\x03\x00\xA9\x81\x00\x00\xAA\x81\x00\x00\xAB\x81\x00\x00\xAC\x81"
-     "\x00\x00"
-     "\xAD\x89\x00\x00\xAE\xB1\x00\x00\xAF\xB1\x00\x00\x3E\xD5\x00\x80\x42\xD5"
-     "\x00\x80"
-     "\x46\xD5\x00\x80\x4A\xD5\x00\x80\x4E\xD5\x00\x80\x52\xD5\x00\x80\x56\xD5"
-     "\x00\x80"
-     "\x5A\xD5\x00\x80\xB8\x69\x00\x00\xB9\x69\x00\x00\xBA\x79\x00\x00\xBB\x79"
-     "\x00\x00"
-     "\xBC\x69\x00\x00\xBD\x69\x00\x00\xBE\xD9\x01\x00\xBF\xD9\x01\x00\xB0\xA9"
-     "\x00\x00"
-     "\xB1\xA9\x00\x00\xB2\xBD\x00\x00\xB3\xB5\x00\x00\xB4\x99\x00\x00\xB5\x99"
-     "\x00\x00"
-     "\xB6\x59\x00\x00\xB7\x59\x00\x00\x5E\xD5\x00\x80\x62\xD5\x00\x80\x66\xD5"
-     "\x00\x80"
-     "\x6A\xD5\x00\x80\x81\x1D\x00\x00\x80\x1D\x00\x00\x6E\xD5\x00\x80\x82\x05"
-     "\x00\x00"
-     "\x84\xAC\x02\x00\xBE\xE8\x05\x00\x87\xC0\x02\x00\x86\x0C\x05\x00\x76\xD5"
-     "\x00\x80"
-     "\x7A\xD5\x00\x80\x7E\xD5\x00\x80\x82\xD5\x00\x80\xB3\x49\x01\x00\x86\xD5"
-     "\x00\x80"
-     "\xD2\x00\x00\x00\x8A\xD5\x00\x80\x8E\xD5\x00\x80\xB6\x49\x01\x00\xB5\x49"
-     "\x01\x00"
-     "\x92\xD5\x00\x80\xBB\x2D\x01\x00\xBA\x2D\x01\x00\x96\xD5\x00\x80\x9A\xD5"
-     "\x00\x80"
-     "\xBF\x1D\x01\x00\xBE\x1D\x01\x00\xBD\x21\x01\x00\xBC\x29\x01\x00\xE1\xB8"
-     "\x1E\x00"
-     "\x9E\xD5\x00\x80\xE3\x38\x1E\x00\xA2\xD5\x00\x80\xA6\xD5\x00\x80\xAA\xD5"
-     "\x00\x80"
-     "\xAE\xD5\x00\x80\xB2\xD5\x00\x80\x84\x6C\x04\x00\xBE\x6C\x04\x00\xB6\xD5"
-     "\x00\x80"
-     "\xBA\xD5\x00\x80\xBE\xD5\x00\x80\xC2\xD5\x00\x80\xEF\x64\x1F\x00\x84\x70"
-     "\x07\x00"
-     "\xC6\xD5\x00\x80\xCA\xD5\x00\x80\x86\xE8\x04\x00\x87\xC0\x04\x00\xCE\xD5"
-     "\x00\x80"
-     "\xE3\x68\x00\x00\xD2\xD5\x00\x80\xE1\xA4\x01\x00\x82\x1D\x00\x00\xEF\xAC"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x81\x15\x00\x00\xB2\xD3\x00\x80\xD6\xD5\x00\x80\xDA\xD5"
-     "\x00\x80"
-     "\xDE\xD5\x00\x80\xE1\xA4\x07\x00\x72\xD5\x00\x80\xE3\x38\x05\x00\xE6\xD5"
-     "\x00\x80"
-     "\xE2\xD5\x00\x80\xEA\xD5\x00\x80\xEE\xD5\x00\x80\xF2\xD5\x00\x80\xF6\xD5"
-     "\x00\x80"
-     "\xFA\xD5\x00\x80\xFE\xD5\x00\x80\x02\xD6\x00\x80\x06\xD6\x00\x80\x0A\xD6"
-     "\x00\x80"
-     "\xEF\xE4\x05\x00\x0E\xD6\x00\x80\xA3\x49\x06\x00\x12\xD6\x00\x80\x16\xD6"
-     "\x00\x80"
-     "\x1A\xD6\x00\x80\x1E\xD6\x00\x80\xA6\x49\x06\x00\xA5\x49\x06\x00\x22\xD6"
-     "\x00\x80"
-     "\xAB\x2D\x06\x00\xAA\x2D\x06\x00\x26\xD6\x00\x80\x2A\xD6\x00\x80\xAF\x1D"
-     "\x06\x00"
-     "\xAE\x1D\x06\x00\xAD\x21\x06\x00\xAC\x29\x06\x00\xA8\xB1\x07\x00\xA9\xB1"
-     "\x07\x00"
-     "\xAA\xB1\x07\x00\xAB\xB1\x07\x00\xAC\xD5\x07\x00\xAD\xD9\x07\x00\xAE\xC9"
-     "\x07\x00"
-     "\xAF\xC1\x07\x00\x2E\xD6\x00\x80\x32\xD6\x00\x80\x36\xD6\x00\x80\x3A\xD6"
-     "\x00\x80"
-     "\x3E\xD6\x00\x80\x42\xD6\x00\x80\x46\xD6\x00\x80\x4A\xD6\x00\x80\xB8\x69"
-     "\x07\x00"
-     "\xB9\x69\x07\x00\xBA\x79\x07\x00\xBB\x79\x07\x00\xBC\x11\x07\x00\xBD\x11"
-     "\x07\x00"
-     "\xBE\x09\x07\x00\xBF\x05\x07\x00\xB0\x89\x07\x00\xB1\x89\x07\x00\xB2\x6D"
-     "\x07\x00"
-     "\xB3\x61\x07\x00\xB4\x61\x07\x00\xB5\x61\x07\x00\xB6\x61\x07\x00\xB7\x61"
-     "\x07\x00"
-     "\xB3\x39\x06\x00\x82\x15\x00\x00\x81\xB5\x01\x00\x80\xB5\x01\x00\x4E\xD6"
-     "\x00\x80"
-     "\xB6\x59\x06\x00\xB5\x5D\x06\x00\xBE\xD4\x01\x00\xBB\x4D\x06\x00\xBA\x4D"
-     "\x06\x00"
-     "\x52\xD6\x00\x80\x56\xD6\x00\x80\xBF\xA5\x06\x00\xBE\xA5\x06\x00\xBD\xB5"
-     "\x06\x00"
-     "\xBC\xB9\x06\x00\x5A\xD6\x00\x80\xA3\x7D\x06\x00\x86\xE8\x00\x00\x87\xCC"
-     "\x01\x00"
-     "\xA6\x1D\x06\x00\x5E\xD6\x00\x80\x62\xD6\x00\x80\xA5\x19\x06\x00\xAA\x09"
-     "\x06\x00"
-     "\xAB\x09\x06\x00\x66\xD6\x00\x80\x6A\xD6\x00\x80\xAE\xE1\x06\x00\xAF\xE1"
-     "\x06\x00"
-     "\xAC\xFD\x06\x00\xAD\xF1\x06\x00\xA8\x8D\x01\x00\xA9\x91\x01\x00\xAA\x91"
-     "\x01\x00"
-     "\xAB\x91\x01\x00\xAC\xB1\x01\x00\xAD\xB1\x01\x00\xAE\xB1\x01\x00\xAF\xB1"
-     "\x01\x00"
-     "\x84\xA8\x01\x00\x6E\xD6\x00\x80\x72\xD6\x00\x80\x76\xD6\x00\x80\x7A\xD6"
-     "\x00\x80"
-     "\x7E\xD6\x00\x80\x82\xD6\x00\x80\x86\xD6\x00\x80\xB8\xA9\x01\x00\xB9\xA9"
-     "\x01\x00"
-     "\xBA\x69\x01\x00\xBB\x69\x01\x00\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x69"
-     "\x01\x00"
-     "\xBF\x69\x01\x00\xB0\xD1\x01\x00\xB1\xD1\x01\x00\xB2\xD5\x01\x00\xB3\xA9"
-     "\x01\x00"
-     "\xB4\xB9\x01\x00\xB5\xB9\x01\x00\xB6\xA9\x01\x00\xB7\xA1\x01\x00\xB3\x19"
-     "\x02\x00"
-     "\x8A\xD6\x00\x80\x8E\xD6\x00\x80\x92\xD6\x00\x80\x96\xD6\x00\x80\xB6\x31"
-     "\x02\x00"
-     "\xB5\x39\x02\x00\x9A\xD6\x00\x80\xBB\x15\x02\x00\xBA\x15\x02\x00\x9E\xD6"
-     "\x00\x80"
-     "\xA2\xD6\x00\x80\xBF\x7D\x02\x00\xBE\x7D\x02\x00\xBD\x7D\x02\x00\xBC\x01"
-     "\x02\x00"
-     "\xA6\xD6\x00\x80\xA3\x5D\x02\x00\xAA\xD6\x00\x80\xAE\xD6\x00\x80\xA6\x75"
-     "\x02\x00"
-     "\xB2\xD6\x00\x80\xB6\xD6\x00\x80\xA5\x7D\x02\x00\xAA\x51\x02\x00\xAB\x51"
-     "\x02\x00"
-     "\xBA\xD6\x00\x80\xBE\xD6\x00\x80\xAE\x39\x02\x00\xAF\x39\x02\x00\xAC\x45"
-     "\x02\x00"
-     "\xAD\x39\x02\x00\x80\xD9\x00\x00\x81\xED\x00\x00\x82\xE5\x00\x00\xBE\xC0"
-     "\x03\x00"
-     "\xC6\xD6\x00\x80\xEF\xE0\x02\x00\x87\x54\x03\x00\x86\xBC\x0C\x00\xCA\xD6"
-     "\x00\x80"
-     "\xE3\x14\x01\x00\xCE\xD6\x00\x80\xE1\xFC\x01\x00\xD2\xD6\x00\x80\xD6\xD6"
-     "\x00\x80"
-     "\xDA\xD6\x00\x80\xDE\xD6\x00\x80\xE2\xD6\x00\x80\xB3\x91\x03\x00\xE6\xD6"
-     "\x00\x80"
-     "\x84\x40\x0C\x00\xB6\xBD\x03\x00\xEA\xD6\x00\x80\xEE\xD6\x00\x80\xB5\xB5"
-     "\x03\x00"
-     "\xBA\x4D\x03\x00\xBB\x55\x03\x00\xF2\xD6\x00\x80\xF6\xD6\x00\x80\xBE\x4D"
-     "\x03\x00"
-     "\xBF\xF1\x00\x00\xBC\x4D\x03\x00\xBD\x45\x03\x00\xA8\x2D\x02\x00\xA9\x45"
-     "\x02\x00"
-     "\xAA\x5D\x02\x00\xAB\x55\x02\x00\xAC\x79\x02\x00\xAD\x79\x02\x00\xAE\xBD"
-     "\x02\x00"
-     "\xAF\xB5\x02\x00\xBE\xAC\x0C\x00\xFA\xD6\x00\x80\xFE\xD6\x00\x80\x02\xD7"
-     "\x00\x80"
-     "\x06\xD7\x00\x80\x0A\xD7\x00\x80\x0E\xD7\x00\x80\x12\xD7\x00\x80\xB8\x51"
-     "\x01\x00"
-     "\xB9\x59\x01\x00\xBA\x61\x01\x00\xBB\x61\x01\x00\xBC\x11\x01\x00\xBD\x1D"
-     "\x01\x00"
-     "\xBE\x15\x01\x00\xBF\x09\x01\x00\xB0\xCD\x02\x00\xB1\xD5\x02\x00\xB2\xDD"
-     "\x02\x00"
-     "\xB3\xD5\x02\x00\xB4\xCD\x02\x00\xB5\x71\x01\x00\xB6\x71\x01\x00\xB7\x71"
-     "\x01\x00"
-     "\xE1\xB4\x06\x00\x02\xD5\x00\x80\xE3\x38\x06\x00\x16\xD7\x00\x80\x1A\xD7"
-     "\x00\x80"
-     "\xE1\x44\x06\x00\x1E\xD7\x00\x80\xE3\x30\x06\x00\x22\xD7\x00\x80\x26\xD7"
-     "\x00\x80"
-     "\x2A\xD7\x00\x80\xEF\x48\x00\x00\xBE\x3C\x0F\x00\x2E\xD7\x00\x80\xEF\xE4"
-     "\x06\x00"
-     "\x36\xD7\x00\x80\xA3\x95\x02\x00\x82\x71\x00\x00\x81\x61\x00\x00\x80\x69"
-     "\x00\x00"
-     "\x3A\xD7\x00\x80\xA6\xB9\x02\x00\xA5\xB1\x02\x00\x3E\xD7\x00\x80\xAB\x51"
-     "\x02\x00"
-     "\xAA\x49\x02\x00\x86\xC8\x0C\x00\x87\x5C\x0D\x00\xAF\xF5\x01\x00\xAE\x49"
-     "\x02\x00"
-     "\xAD\x41\x02\x00\xAC\x49\x02\x00\x32\xD7\x00\x80\x42\xD7\x00\x80\x46\xD7"
-     "\x00\x80"
-     "\x4A\xD7\x00\x80\xE1\x94\x0C\x00\x4E\xD7\x00\x80\xE3\xC4\x0C\x00\x52\xD7"
-     "\x00\x80"
-     "\x56\xD7\x00\x80\x5A\xD7\x00\x80\xEF\xF0\x0D\x00\x5E\xD7\x00\x80\x62\xD7"
-     "\x00\x80"
-     "\x66\xD7\x00\x80\x6A\xD7\x00\x80\x6E\xD7\x00\x80\xB3\xE9\x0E\x00\x72\xD7"
-     "\x00\x80"
-     "\x76\xD7\x00\x80\x7A\xD7\x00\x80\x7E\xD7\x00\x80\xB6\xED\x0E\x00\xB5\xED"
-     "\x0E\x00"
-     "\x82\xD7\x00\x80\xBB\xB5\x0E\x00\xBA\xAD\x0E\x00\x86\xD7\x00\x80\x8A\xD7"
-     "\x00\x80"
-     "\xBF\x55\x0F\x00\xBE\x99\x0E\x00\xBD\x91\x0E\x00\xBC\xA5\x0E\x00\xA8\x11"
-     "\x0E\x00"
-     "\xA9\x25\x0E\x00\xAA\x21\x0E\x00\xAB\x3D\x0E\x00\xAC\x25\x0E\x00\xAD\x2D"
-     "\x0E\x00"
-     "\xAE\x25\x0E\x00\xAF\x5D\x0E\x00\x8E\xD7\x00\x80\x92\xD7\x00\x80\x96\xD7"
-     "\x00\x80"
-     "\x9A\xD7\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x9E\xD7"
-     "\x00\x80"
-     "\xB8\xF1\x0F\x00\xB9\xF1\x0F\x00\xBA\xF1\x0F\x00\xBB\xF1\x0F\x00\xBC\x5D"
-     "\x0F\x00"
-     "\xBD\x45\x0F\x00\xBE\x41\x0F\x00\xBF\x7D\x0F\x00\xB0\x29\x0E\x00\xB1\x31"
-     "\x0E\x00"
-     "\xB2\x31\x0E\x00\xB3\x05\x0E\x00\xB4\x1D\x0E\x00\xB5\xD1\x0F\x00\xB6\xD1"
-     "\x0F\x00"
-     "\xB7\xD1\x0F\x00\xA3\xA9\x0F\x00\x84\x8C\x02\x00\x85\x18\x01\x00\xBF\x04"
-     "\x00\x00"
-     "\x85\x74\x01\x00\xA6\xAD\x0F\x00\xA5\xAD\x0F\x00\xA6\xD7\x00\x80\xAB\xF5"
-     "\x0F\x00"
-     "\xAA\xED\x0F\x00\x86\x88\x07\x00\x87\x58\x00\x00\xAF\x15\x0E\x00\xAE\xD9"
-     "\x0F\x00"
-     "\xAD\xD1\x0F\x00\xAC\xE5\x0F\x00\xAA\xD7\x00\x80\xB3\x99\x0E\x00\xAE\xD7"
-     "\x00\x80"
-     "\xB2\xD7\x00\x80\xB6\x5D\x01\x00\xB6\xD7\x00\x80\xBA\xD7\x00\x80\xB5\x55"
-     "\x01\x00"
-     "\xBA\x41\x01\x00\xBB\x41\x01\x00\xBE\xD7\x00\x80\xC2\xD7\x00\x80\xBE\x45"
-     "\x01\x00"
-     "\xBF\x4D\x01\x00\xBC\x55\x01\x00\xBD\x4D\x01\x00\xA8\x3D\x0E\x00\xA9\x55"
-     "\x0E\x00"
-     "\xAA\x55\x0E\x00\xAB\x89\x01\x00\xAC\x99\x01\x00\xAD\x99\x01\x00\xAE\x8D"
-     "\x01\x00"
-     "\xAF\x81\x01\x00\xC6\xD7\x00\x80\xCA\xD7\x00\x80\xCE\xD7\x00\x80\xD2\xD7"
-     "\x00\x80"
-     "\xD6\xD7\x00\x80\xDA\xD7\x00\x80\xDE\xD7\x00\x80\xE2\xD7\x00\x80\xB8\x89"
-     "\x01\x00"
-     "\xB9\x89\x01\x00\xBA\x99\x01\x00\xBB\x91\x01\x00\xBC\xB9\x01\x00\xBD\xB9"
-     "\x01\x00"
-     "\xBE\x79\x01\x00\xBF\x79\x01\x00\xB0\xC1\x01\x00\xB1\xC1\x01\x00\xB2\xC5"
-     "\x01\x00"
-     "\xB3\xDD\x01\x00\xB4\xC5\x01\x00\xB5\xCD\x01\x00\xB6\xC5\x01\x00\xB7\xB9"
-     "\x01\x00"
-     "\xA3\xDD\x0D\x00\xE6\xD7\x00\x80\xEA\xD7\x00\x80\xEE\xD7\x00\x80\xF2\xD7"
-     "\x00\x80"
-     "\xA6\x19\x02\x00\xA5\x11\x02\x00\xF6\xD7\x00\x80\xAB\x05\x02\x00\xAA\x05"
-     "\x02\x00"
-     "\xFA\xD7\x00\x80\x8F\x79\x1A\x00\xAF\x09\x02\x00\xAE\x01\x02\x00\xAD\x09"
-     "\x02\x00"
-     "\xAC\x11\x02\x00\x94\x9D\x1E\x00\x95\x9D\x11\x00\x96\x4D\x12\x00\x97\xED"
-     "\x12\x00"
-     "\x90\xD5\x1A\x00\x91\xBD\x1A\x00\x92\x59\x1E\x00\x93\x59\x1E\x00\x9C\x65"
-     "\x0A\x00"
-     "\x9D\x3D\x0A\x00\x9E\xED\x0D\x00\x9F\x69\x0E\x00\x98\x9D\x15\x00\x99\x75"
-     "\x16\x00"
-     "\x9A\x4D\x16\x00\x9B\x8D\x09\x00\xE1\x84\x01\x00\xFE\xD7\x00\x80\xE3\xDC"
-     "\xFE\x03"
-     "\x84\x80\x03\x00\xA0\x81\x0E\x00\xA1\xBD\x0F\x00\xA2\x0D\x01\x00\xA3\x9D"
-     "\x02\x00"
-     "\xB1\x01\xE8\x03\xB0\x49\xF6\x03\xB3\x69\xEA\x03\xB2\x8D\xEA\x03\xB5\x91"
-     "\xEE\x03"
-     "\xB4\x01\xEC\x03\xEF\x60\xEC\x03\x02\xD8\x00\x80\xA9\xD1\xFE\x03\xA8\xC1"
-     "\xFE\x03"
-     "\xAB\xCD\xF2\x03\xAA\xBD\xFF\x03\xAD\x91\xF3\x03\xAC\xDD\xF2\x03\xAF\xF9"
-     "\xF6\x03"
-     "\xAE\x01\xF4\x03\xA1\x01\x04\x00\x06\xD8\x00\x80\xA3\x79\x06\x00\xA2\xCD"
-     "\x06\x00"
-     "\xA5\xD1\xFA\x03\xA4\x01\xF8\x03\xA7\x01\xFC\x03\xA6\xC1\xFA\x03\x82\x55"
-     "\x1B\x00"
-     "\x83\x81\x1B\x00\x0A\xD8\x00\x80\x0E\xD8\x00\x80\x86\x15\x1F\x00\x87\xC9"
-     "\x1F\x00"
-     "\x84\xF1\x1B\x00\x85\x1D\x1F\x00\x8A\xF9\x10\x00\x8B\x21\x13\x00\xBE\x24"
-     "\x05\x00"
-     "\x16\xD8\x00\x80\x8E\x31\x17\x00\x8F\x01\x17\x00\x8C\x99\x13\x00\x8D\xD1"
-     "\x14\x00"
-     "\x92\xB1\x08\x00\x93\x41\x0B\x00\x86\x68\x04\x00\x87\x7C\x05\x00\x96\x55"
-     "\x0F\x00"
-     "\x97\x45\x0F\x00\x94\x01\x0B\x00\x95\xD9\x0C\x00\x9A\x49\x0E\x00\x9B\xF1"
-     "\x00\x00"
-     "\x1A\xD8\x00\x80\x1E\xD8\x00\x80\x81\x79\x00\x00\x80\x91\x01\x00\x9C\x79"
-     "\x03\x00"
-     "\x82\x71\x00\x00\x22\xD8\x00\x80\xE1\x74\x0E\x00\x84\x80\x05\x00\xE3\x98"
-     "\x00\x00"
-     "\x26\xD8\x00\x80\xC2\xD6\x00\x80\x2A\xD8\x00\x80\x2E\xD8\x00\x80\x32\xD8"
-     "\x00\x80"
-     "\x36\xD8\x00\x80\x3A\xD8\x00\x80\x3E\xD8\x00\x80\x42\xD8\x00\x80\x46\xD8"
-     "\x00\x80"
-     "\x4A\xD8\x00\x80\xEF\xF4\x0E\x00\xB3\x89\x02\x00\x4E\xD8\x00\x80\x52\xD8"
-     "\x00\x80"
-     "\x56\xD8\x00\x80\x5A\xD8\x00\x80\xB6\x8D\x02\x00\xB5\x8D\x02\x00\x5E\xD8"
-     "\x00\x80"
-     "\xBB\x45\x02\x00\xBA\x45\x02\x00\x62\xD8\x00\x80\x66\xD8\x00\x80\xBF\x45"
-     "\x02\x00"
-     "\xBE\x45\x02\x00\xBD\x55\x02\x00\xBC\x55\x02\x00\xA8\x1D\x05\x00\xA9\x45"
-     "\x05\x00"
-     "\xAA\x41\x05\x00\xAB\x41\x05\x00\xAC\x45\x05\x00\xAD\x4D\x05\x00\xAE\x71"
-     "\x05\x00"
-     "\xAF\x71\x05\x00\x12\xD8\x00\x80\x6A\xD8\x00\x80\x6E\xD8\x00\x80\x72\xD8"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\xB1\x00\x00\x82\xB1\x00\x00\x76\xD8\x00\x80\xB8\x45"
-     "\x06\x00"
-     "\xB9\x4D\x06\x00\xBA\x45\x06\x00\xBB\x5D\x06\x00\xBC\x4D\x06\x00\xBD\x7D"
-     "\x06\x00"
-     "\xBE\xDD\x06\x00\xBF\xE1\x06\x00\xB0\xC1\x05\x00\xB1\xC1\x05\x00\xB2\xC1"
-     "\x05\x00"
-     "\xB3\xC1\x05\x00\xB4\xC1\x05\x00\xB5\xC1\x05\x00\xB6\xC1\x05\x00\xB7\xC1"
-     "\x05\x00"
-     "\xA3\xC9\x05\x00\x7A\xD8\x00\x80\x7E\xD8\x00\x80\xBE\x00\x03\x00\x82\xD8"
-     "\x00\x80"
-     "\xA6\xCD\x05\x00\xA5\xCD\x05\x00\x86\xD8\x00\x80\xAB\x05\x05\x00\xAA\x05"
-     "\x05\x00"
-     "\x86\x08\x00\x00\x87\xDC\x03\x00\xAF\x05\x05\x00\xAE\x05\x05\x00\xAD\x15"
-     "\x05\x00"
-     "\xAC\x15\x05\x00\x8A\xD8\x00\x80\xB3\xA9\x07\x00\x8E\xD8\x00\x80\x92\xD8"
-     "\x00\x80"
-     "\xB6\x49\x07\x00\x96\xD8\x00\x80\x9A\xD8\x00\x80\xB5\x4D\x07\x00\xBA\x45"
-     "\x07\x00"
-     "\xBB\x4D\x07\x00\x9E\xD8\x00\x80\xA2\xD8\x00\x80\xBE\x45\x07\x00\xBF\x4D"
-     "\x07\x00"
-     "\xBC\x55\x07\x00\xBD\x4D\x07\x00\xA3\x6D\x06\x00\xA6\xD8\x00\x80\xAA\xD8"
-     "\x00\x80"
-     "\xAE\xD8\x00\x80\xB2\xD8\x00\x80\xA6\x8D\x06\x00\xA5\x89\x06\x00\xB6\xD8"
-     "\x00\x80"
-     "\xAB\x89\x06\x00\xAA\x81\x06\x00\xBA\xD8\x00\x80\xBE\xD8\x00\x80\xAF\x89"
-     "\x06\x00"
-     "\xAE\x81\x06\x00\xAD\x89\x06\x00\xAC\x91\x06\x00\xC2\xD8\x00\x80\xC6\xD8"
-     "\x00\x80"
-     "\xCA\xD8\x00\x80\xCE\xD8\x00\x80\xD2\xD8\x00\x80\xD6\xD8\x00\x80\xDA\xD8"
-     "\x00\x80"
-     "\xEF\x8C\x04\x00\xDE\xD8\x00\x80\xE1\x54\x06\x00\xE2\xD8\x00\x80\xE3\xF0"
-     "\x05\x00"
-     "\x80\x0D\x00\x00\x81\xD5\x00\x00\x82\xDD\x00\x00\xE6\xD8\x00\x80\xA8\xAD"
-     "\x06\x00"
-     "\xA9\xED\x06\x00\xAA\xE5\x06\x00\xAB\x0D\x06\x00\xAC\x01\x06\x00\xAD\x0D"
-     "\x06\x00"
-     "\xAE\x05\x06\x00\xAF\x3D\x06\x00\xEA\xD8\x00\x80\xEE\xD8\x00\x80\x84\xE4"
-     "\x01\x00"
-     "\xBE\xE8\x01\x00\x86\x10\xFF\x03\x87\x94\x00\x00\xF6\xD8\x00\x80\xFA\xD8"
-     "\x00\x80"
-     "\xB8\x0D\x06\x00\xB9\x15\x06\x00\xBA\x15\x06\x00\xBB\x25\x06\x00\xBC\x3D"
-     "\x06\x00"
-     "\xBD\xF1\x01\x00\xBE\xF1\x01\x00\xBF\xF1\x01\x00\xB0\x45\x06\x00\xB1\x4D"
-     "\x06\x00"
-     "\xB2\x45\x06\x00\xB3\x5D\x06\x00\xB4\x45\x06\x00\xB5\x4D\x06\x00\xB6\x3D"
-     "\x06\x00"
-     "\xB7\x35\x06\x00\xA8\x9D\x02\x00\xA9\xC5\x02\x00\xAA\xC1\x02\x00\xAB\xC1"
-     "\x02\x00"
-     "\xAC\xC5\x02\x00\xAD\xCD\x02\x00\xAE\xF1\x02\x00\xAF\xF1\x02\x00\xFE\xD8"
-     "\x00\x80"
-     "\x02\xD9\x00\x80\x06\xD9\x00\x80\x0A\xD9\x00\x80\xE1\x28\x03\x00\x0E\xD9"
-     "\x00\x80"
-     "\x12\xD9\x00\x80\x16\xD9\x00\x80\xB8\x05\x03\x00\xB9\x0D\x03\x00\xBA\x05"
-     "\x03\x00"
-     "\xBB\x1D\x03\x00\xBC\x05\x03\x00\xBD\x0D\x03\x00\xBE\x39\x03\x00\xBF\x35"
-     "\x03\x00"
-     "\xB0\x65\x03\x00\xB1\x6D\x03\x00\xB2\x65\x03\x00\xB3\x7D\x03\x00\xB4\x6D"
-     "\x03\x00"
-     "\xB5\x5D\x03\x00\xB6\x55\x03\x00\xB7\x3D\x03\x00\xB3\x49\x02\x00\x1A\xD9"
-     "\x00\x80"
-     "\x1E\xD9\x00\x80\x22\xD9\x00\x80\x26\xD9\x00\x80\xB6\x4D\x02\x00\xB5\x4D"
-     "\x02\x00"
-     "\x2A\xD9\x00\x80\xBB\xA9\x03\x00\xBA\xA1\x03\x00\x2E\xD9\x00\x80\x32\xD9"
-     "\x00\x80"
-     "\xBF\x99\x03\x00\xBE\xA9\x03\x00\xBD\xA9\x03\x00\xBC\xB1\x03\x00\x36\xD9"
-     "\x00\x80"
-     "\x3A\xD9\x00\x80\xA3\x01\x02\x00\x3E\xD9\x00\x80\xA5\x05\x02\x00\x42\xD9"
-     "\x00\x80"
-     "\x46\xD9\x00\x80\xA6\x05\x02\x00\xA2\xD7\x00\x80\x4A\xD9\x00\x80\xAB\xE1"
-     "\x03\x00"
-     "\xAA\xE9\x03\x00\xAD\xE1\x03\x00\xAC\xF9\x03\x00\xAF\xD1\x03\x00\xAE\xE1"
-     "\x03\x00"
-     "\x4E\xD9\x00\x80\x52\xD9\x00\x80\x56\xD9\x00\x80\x5A\xD9\x00\x80\x5E\xD9"
-     "\x00\x80"
-     "\xE1\xA4\x01\x00\xBE\x44\x02\x00\xE3\xA4\x00\x00\x80\x29\x00\x00\x81\x29"
-     "\x00\x00"
-     "\x82\x39\x00\x00\xEF\x78\x00\x00\x66\xD9\x00\x80\x84\xB8\x02\x00\x6A\xD9"
-     "\x00\x80"
-     "\x6E\xD9\x00\x80\xB3\x61\x01\x00\x72\xD9\x00\x80\x76\xD9\x00\x80\x7A\xD9"
-     "\x00\x80"
-     "\x7E\xD9\x00\x80\xB6\x79\x01\x00\xB5\x71\x01\x00\x82\xD9\x00\x80\xBB\xE5"
-     "\x01\x00"
-     "\xBA\x59\x01\x00\x86\x48\xFD\x03\x87\x40\xFD\x03\xBF\xED\x01\x00\xBE\xE5"
-     "\x01\x00"
-     "\xBD\xF5\x01\x00\xBC\xF5\x01\x00\x86\xD9\x00\x80\x8A\xD9\x00\x80\x8E\xD9"
-     "\x00\x80"
-     "\x84\x40\xFC\x03\x92\xD9\x00\x80\x96\xD9\x00\x80\x9A\xD9\x00\x80\xEF\xE8"
-     "\x01\x00"
-     "\x9E\xD9\x00\x80\xE1\xF4\x06\x00\xAC\xB4\xF3\x03\xE3\x34\x06\x00\xA2\xD9"
-     "\x00\x80"
-     "\xA6\xD9\x00\x80\x84\xB0\xFC\x03\xAA\xD9\x00\x80\xAE\xD9\x00\x80\xA6\xE9"
-     "\x01\x00"
-     "\xA5\xE1\x01\x00\xB2\xD9\x00\x80\xA3\xF1\x01\x00\xB6\xD9\x00\x80\xBA\xD9"
-     "\x00\x80"
-     "\xBE\xD9\x00\x80\xAF\x7D\x01\x00\xAE\x75\x01\x00\xAD\x65\x01\x00\xAC\x65"
-     "\x01\x00"
-     "\xAB\x75\x01\x00\xAA\xC9\x01\x00\xC2\xD9\x00\x80\xC6\xD9\x00\x80\xA8\xB9"
-     "\xFD\x03"
-     "\xA9\xB9\xFD\x03\xAA\x29\xFE\x03\xAB\x29\xFE\x03\xAC\x39\xFE\x03\xAD\x39"
-     "\xFE\x03"
-     "\xAE\x2D\xFE\x03\xAF\x25\xFE\x03\x62\xD9\x00\x80\x82\x11\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x80\xD1\xFF\x03\xCA\xD9\x00\x80\xCE\xD9\x00\x80\xBE\x34\x03\x00\xD2\xD9"
-     "\x00\x80"
-     "\xB8\x81\xFE\x03\xB9\x81\xFE\x03\xBA\x85\xFE\x03\xBB\x99\xFE\x03\xBC\x89"
-     "\xFE\x03"
-     "\xBD\xB1\xFE\x03\xBE\x59\xFF\x03\xBF\x59\xFF\x03\xB0\x5D\xFE\x03\xB1\xE1"
-     "\xFE\x03"
-     "\xB2\xE1\xFE\x03\xB3\xE1\xFE\x03\xB4\xE1\xFE\x03\xB5\xE9\xFE\x03\xB6\xD1"
-     "\xFE\x03"
-     "\xB7\xD1\xFE\x03\xA8\xB1\xFF\x03\xA9\xBD\xFF\x03\xAA\xB5\xFF\x03\xAB\xC9"
-     "\xFF\x03"
-     "\xAC\xD9\xFF\x03\xAD\xD9\xFF\x03\xAE\xC9\xFF\x03\xAF\xC1\xFF\x03\x86\x00"
-     "\x00\x00"
-     "\x87\x28\x03\x00\xD6\xD9\x00\x80\xDA\xD9\x00\x80\xDE\xD9\x00\x80\xE2\xD9"
-     "\x00\x80"
-     "\xE6\xD9\x00\x80\xEA\xD9\x00\x80\xB8\x75\xFF\x03\xB9\x79\xFF\x03\xBA\x09"
-     "\xFF\x03"
-     "\xBB\x01\xFF\x03\xBC\x05\xFF\x03\xBD\x0D\xFF\x03\xBE\x05\xFF\x03\xBF\x25"
-     "\xFF\x03"
-     "\xB0\x61\xFF\x03\xB1\x61\xFF\x03\xB2\x61\xFF\x03\xB3\x7D\xFF\x03\xB4\x65"
-     "\xFF\x03"
-     "\xB5\x6D\xFF\x03\xB6\x59\xFF\x03\xB7\x4D\xFF\x03\xB3\x3D\xFE\x03\xEE\xD9"
-     "\x00\x80"
-     "\xF2\xD9\x00\x80\xF6\xD9\x00\x80\xFA\xD9\x00\x80\xB6\x59\xFE\x03\xB5\x51"
-     "\xFE\x03"
-     "\xFE\xD9\x00\x80\xBB\xA5\xFE\x03\xBA\xA5\xFE\x03\x02\xDA\x00\x80\x06\xDA"
-     "\x00\x80"
-     "\xBF\xA9\xFE\x03\xBE\xA1\xFE\x03\xBD\xA9\xFE\x03\xBC\xB5\xFE\x03\x0A\xDA"
-     "\x00\x80"
-     "\x0E\xDA\x00\x80\xEF\xA8\xFC\x03\x12\xDA\x00\x80\x16\xDA\x00\x80\x1A\xDA"
-     "\x00\x80"
-     "\x1E\xDA\x00\x80\x22\xDA\x00\x80\x26\xDA\x00\x80\x2A\xDA\x00\x80\x2E\xDA"
-     "\x00\x80"
-     "\x32\xDA\x00\x80\xE1\xD8\xFE\x03\x36\xDA\x00\x80\xE3\x5C\xFC\x03\x3A\xDA"
-     "\x00\x80"
-     "\xA3\xBD\xFE\x03\x82\x15\x00\x00\x81\xF5\x00\x00\x80\xF9\x00\x00\x3E\xDA"
-     "\x00\x80"
-     "\xA6\xD9\xFE\x03\xA5\xD1\xFE\x03\xBE\x90\x01\x00\xAB\x25\xFE\x03\xAA\x25"
-     "\xFE\x03"
-     "\x86\x08\x00\x00\x87\x70\x01\x00\xAF\x29\xFE\x03\xAE\x21\xFE\x03\xAD\x29"
-     "\xFE\x03"
-     "\xAC\x35\xFE\x03\x46\xDA\x00\x80\xB3\x75\x01\x00\x4A\xDA\x00\x80\x4E\xDA"
-     "\x00\x80"
-     "\xB6\x6D\x01\x00\x52\xDA\x00\x80\x56\xDA\x00\x80\xB5\x65\x01\x00\xBA\xED"
-     "\x00\x00"
-     "\xBB\xF5\x00\x00\x5A\xDA\x00\x80\x5E\xDA\x00\x80\xBE\xDD\x00\x00\xBF\xB5"
-     "\x00\x00"
-     "\xBC\xE5\x00\x00\xBD\xD5\x00\x00\xA8\xA9\x02\x00\xA9\xA9\x02\x00\xAA\xB9"
-     "\x02\x00"
-     "\xAB\xB9\x02\x00\xAC\xA9\x02\x00\xAD\xA9\x02\x00\xAE\x1D\x03\x00\xAF\x15"
-     "\x03\x00"
-     "\xBE\xAC\x04\x00\x62\xDA\x00\x80\x66\xDA\x00\x80\xBE\xA0\x05\x00\x6A\xDA"
-     "\x00\x80"
-     "\x6E\xDA\x00\x80\x72\xDA\x00\x80\x76\xDA\x00\x80\xB8\x11\x03\x00\xB9\x11"
-     "\x03\x00"
-     "\xBA\x11\x03\x00\xBB\x11\x03\x00\xBC\x31\x03\x00\xBD\x31\x03\x00\xBE\x31"
-     "\x03\x00"
-     "\xBF\x31\x03\x00\xB0\x6D\x03\x00\xB1\x75\x03\x00\xB2\x75\x03\x00\xB3\x45"
-     "\x03\x00"
-     "\xB4\x5D\x03\x00\xB5\x35\x03\x00\xB6\x3D\x03\x00\xB7\x31\x03\x00\xA3\x31"
-     "\x02\x00"
-     "\x7A\xDA\x00\x80\x7E\xDA\x00\x80\x82\xDA\x00\x80\x86\xDA\x00\x80\xA6\x29"
-     "\x02\x00"
-     "\xA5\x21\x02\x00\x8A\xDA\x00\x80\xAB\xB1\x03\x00\xAA\xA9\x03\x00\x8E\xDA"
-     "\x00\x80"
-     "\x92\xDA\x00\x80\xAF\xF1\x03\x00\xAE\x99\x03\x00\xAD\x91\x03\x00\xAC\xA1"
-     "\x03\x00"
-     "\x84\xC0\x03\x00\x96\xDA\x00\x80\x9A\xDA\x00\x80\x9E\xDA\x00\x80\xA2\xDA"
-     "\x00\x80"
-     "\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D\x00\x00\xA6\xDA\x00\x80\xAA\xDA"
-     "\x00\x80"
-     "\xAE\xDA\x00\x80\xB6\xDA\x00\x80\xBA\xDA\x00\x80\xF2\xD8\x00\x80\x86\xD8"
-     "\x04\x00"
-     "\x87\xBC\x04\x00\xA8\xC1\x02\x00\xA9\xCD\x02\x00\xAA\xC5\x02\x00\xAB\xD9"
-     "\x02\x00"
-     "\xAC\xC9\x02\x00\xAD\xFD\x02\x00\xAE\xF5\x02\x00\xAF\x39\x01\x00\xBE\xDA"
-     "\x00\x80"
-     "\xC2\xDA\x00\x80\xC6\xDA\x00\x80\xCA\xDA\x00\x80\xCE\xDA\x00\x80\xD2\xDA"
-     "\x00\x80"
-     "\xD6\xDA\x00\x80\xDA\xDA\x00\x80\xB8\xD1\x01\x00\xB9\xD9\x01\x00\xBA\xE1"
-     "\x01\x00"
-     "\xBB\xE1\x01\x00\xBC\x95\x01\x00\xBD\x9D\x01\x00\xBE\x95\x01\x00\xBF\x8D"
-     "\x01\x00"
-     "\xB0\x49\x01\x00\xB1\x55\x01\x00\xB2\x5D\x01\x00\xB3\x55\x01\x00\xB4\x4D"
-     "\x01\x00"
-     "\xB5\xF1\x01\x00\xB6\xF1\x01\x00\xB7\xF1\x01\x00\xEF\xE4\x00\x00\xDE\xDA"
-     "\x00\x80"
-     "\xE2\xDA\x00\x80\x84\x0C\x05\x00\xE6\xDA\x00\x80\xEA\xDA\x00\x80\xEE\xDA"
-     "\x00\x80"
-     "\xEF\x5C\x01\x00\xF2\xDA\x00\x80\xE1\xD4\xFE\x03\xF6\xDA\x00\x80\xE3\x30"
-     "\xFE\x03"
-     "\xE3\x18\xFE\x03\xFA\xDA\x00\x80\xE1\x00\x01\x00\xFE\xDA\x00\x80\x02\xDB"
-     "\x00\x80"
-     "\x06\xDB\x00\x80\x0A\xDB\x00\x80\xB3\x01\x02\x00\x0E\xDB\x00\x80\xB5\x0D"
-     "\x02\x00"
-     "\xB6\x31\x02\x00\x12\xDB\x00\x80\x16\xDB\x00\x80\x1A\xDB\x00\x80\xBA\xC1"
-     "\x01\x00"
-     "\xBB\xDD\x01\x00\xBC\xC5\x01\x00\xBD\xCD\x01\x00\xBE\xC5\x01\x00\xBF\x79"
-     "\x01\x00"
-     "\xA3\xCD\x05\x00\xB2\xDA\x00\x80\x1E\xDB\x00\x80\x26\xDB\x00\x80\x22\xDB"
-     "\x00\x80"
-     "\xA6\xFD\x05\x00\xA5\xC1\x05\x00\x2A\xDB\x00\x80\xAB\x11\x06\x00\xAA\x0D"
-     "\x06\x00"
-     "\x86\x08\x03\x00\x87\x7C\x03\x00\xAF\xB5\x06\x00\xAE\x09\x06\x00\xAD\x01"
-     "\x06\x00"
-     "\xAC\x09\x06\x00\x80\xA1\x07\x00\x81\xA1\x07\x00\x82\x41\x00\x00\xB3\x6D"
-     "\x07\x00"
-     "\x2E\xDB\x00\x80\xB5\x7D\x07\x00\xB6\x75\x07\x00\x32\xDB\x00\x80\x36\xDB"
-     "\x00\x80"
-     "\x3A\xDB\x00\x80\xBA\x51\x07\x00\xBB\x51\x07\x00\xBC\xF5\x07\x00\xBD\xFD"
-     "\x07\x00"
-     "\xBE\xE9\x07\x00\xBF\xE9\x07\x00\xA8\x6D\x06\x00\xA9\x75\x06\x00\xAA\xA9"
-     "\x07\x00"
-     "\xAB\xA9\x07\x00\xAC\xB9\x07\x00\xAD\xB9\x07\x00\xAE\xA9\x07\x00\xAF\xA9"
-     "\x07\x00"
-     "\x3E\xDB\x00\x80\x42\xDB\x00\x80\x46\xDB\x00\x80\x4A\xDB\x00\x80\x4E\xDB"
-     "\x00\x80"
-     "\x52\xDB\x00\x80\x56\xDB\x00\x80\x5A\xDB\x00\x80\xB8\x69\x07\x00\xB9\x71"
-     "\x07\x00"
-     "\xBA\x75\x07\x00\xBB\x0D\x07\x00\xBC\x15\x07\x00\xBD\x1D\x07\x00\xBE\x09"
-     "\x07\x00"
-     "\xBF\x01\x07\x00\xB0\xD9\x07\x00\xB1\xD9\x07\x00\xB2\x6D\x07\x00\xB3\x65"
-     "\x07\x00"
-     "\xB4\x61\x07\x00\xB5\x61\x07\x00\xB6\x61\x07\x00\xB7\x5D\x07\x00\xA3\x21"
-     "\x06\x00"
-     "\x5E\xDB\x00\x80\x62\xDB\x00\x80\x66\xDB\x00\x80\x6A\xDB\x00\x80\xA6\x39"
-     "\x06\x00"
-     "\xA5\x31\x06\x00\x6E\xDB\x00\x80\xAB\x1D\x06\x00\xAA\x1D\x06\x00\x72\xDB"
-     "\x00\x80"
-     "\x76\xDB\x00\x80\xAF\xA5\x06\x00\xAE\xA5\x06\x00\xAD\xB1\x06\x00\xAC\xB9"
-     "\x06\x00"
-     "\x7A\xDB\x00\x80\xEF\x9C\x05\x00\x7E\xDB\x00\x80\x82\xDB\x00\x80\x80\x0D"
-     "\x00\x00"
-     "\x81\x35\x00\x00\x82\x3D\x00\x00\x86\xDB\x00\x80\x8A\xDB\x00\x80\x8E\xDB"
-     "\x00\x80"
-     "\x84\xA4\x01\x00\xBE\xA4\x01\x00\x92\xDB\x00\x80\xE3\xA4\x05\x00\x96\xDB"
-     "\x00\x80"
-     "\xE1\xCC\x06\x00\xA8\x8D\x01\x00\xA9\x99\x01\x00\xAA\xD9\x01\x00\xAB\xD1"
-     "\x01\x00"
-     "\xAC\xF1\x01\x00\xAD\xF1\x01\x00\xAE\x39\x01\x00\xAF\x39\x01\x00\x86\x00"
-     "\x00\x00"
-     "\x87\x64\x01\x00\x9A\xDB\x00\x80\x9E\xDB\x00\x80\xA2\xDB\x00\x80\xA6\xDB"
-     "\x00\x80"
-     "\xAA\xDB\x00\x80\xAE\xDB\x00\x80\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5"
-     "\x00\x00"
-     "\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x91\x00\x00\xBF\x91"
-     "\x00\x00"
-     "\xB0\x49\x01\x00\xB1\x49\x01\x00\xB2\x59\x01\x00\xB3\x59\x01\x00\xB4\x49"
-     "\x01\x00"
-     "\xB5\x49\x01\x00\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xB3\x51\x02\x00\xB2\xDB"
-     "\x00\x80"
-     "\xB6\xDB\x00\x80\xBA\xDB\x00\x80\xBE\xDB\x00\x80\xB6\x71\x02\x00\xB5\x79"
-     "\x02\x00"
-     "\xC2\xDB\x00\x80\xBB\x8D\x02\x00\xBA\x8D\x02\x00\xC6\xDB\x00\x80\xCA\xDB"
-     "\x00\x80"
-     "\xBF\x8D\x02\x00\xBE\x8D\x02\x00\xBD\x8D\x02\x00\xBC\x8D\x02\x00\x85\x6C"
-     "\x12\x00"
-     "\xA3\x15\x02\x00\xCE\xDB\x00\x80\xD2\xDB\x00\x80\xA6\x35\x02\x00\xD6\xDB"
-     "\x00\x80"
-     "\xDA\xDB\x00\x80\xA5\x3D\x02\x00\xAA\xC9\x02\x00\xAB\xC9\x02\x00\xDE\xDB"
-     "\x00\x80"
-     "\xE2\xDB\x00\x80\xAE\xC9\x02\x00\xAF\xC9\x02\x00\xAC\xC9\x02\x00\xAD\xC9"
-     "\x02\x00"
-     "\x84\x80\x03\x00\xE6\xDB\x00\x80\xEA\xDB\x00\x80\xBE\x00\x0C\x00\xEE\xDB"
-     "\x00\x80"
-     "\xF2\xDB\x00\x80\xF6\xDB\x00\x80\xFA\xDB\x00\x80\xFE\xDB\x00\x80\x06\xDC"
-     "\x00\x80"
-     "\x0A\xDC\x00\x80\x0E\xDC\x00\x80\x12\xDC\x00\x80\x16\xDC\x00\x80\x1A\xDC"
-     "\x00\x80"
-     "\x1E\xDC\x00\x80\x80\x59\x00\x00\x81\x59\x00\x00\x82\x69\x00\x00\xEF\xB8"
-     "\x02\x00"
-     "\x22\xDC\x00\x80\x26\xDC\x00\x80\x87\x18\x03\x00\x86\xFC\x0C\x00\x2A\xDC"
-     "\x00\x80"
-     "\x2E\xDC\x00\x80\x32\xDC\x00\x80\x36\xDC\x00\x80\x3A\xDC\x00\x80\xE1\x4C"
-     "\x01\x00"
-     "\x42\xDA\x00\x80\xE3\x6C\x02\x00\xA8\x79\x02\x00\xA9\xAD\x02\x00\xAA\xA5"
-     "\x02\x00"
-     "\xAB\xBD\x02\x00\xAC\xA5\x02\x00\xAD\xAD\x02\x00\xAE\xA5\x02\x00\xAF\x19"
-     "\x01\x00"
-     "\x84\xA8\x0D\x00\x3E\xDC\x00\x80\x42\xDC\x00\x80\x46\xDC\x00\x80\x4A\xDC"
-     "\x00\x80"
-     "\x4E\xDC\x00\x80\x52\xDC\x00\x80\x56\xDC\x00\x80\xB8\x15\x01\x00\xB9\x1D"
-     "\x01\x00"
-     "\xBA\x21\x01\x00\xBB\x21\x01\x00\xBC\xD5\x01\x00\xBD\xDD\x01\x00\xBE\xC9"
-     "\x01\x00"
-     "\xBF\xC9\x01\x00\xB0\x69\x01\x00\xB1\x69\x01\x00\xB2\x79\x01\x00\xB3\x71"
-     "\x01\x00"
-     "\xB4\x59\x01\x00\xB5\x59\x01\x00\xB6\x31\x01\x00\xB7\x31\x01\x00\xE1\x60"
-     "\x07\x00"
-     "\x5A\xDC\x00\x80\xE3\x14\x06\x00\xBE\x00\x0C\x00\x5E\xDC\x00\x80\xE1\xA4"
-     "\x01\x00"
-     "\x62\xDC\x00\x80\xE3\xE4\x00\x00\x66\xDC\x00\x80\x6A\xDC\x00\x80\x6E\xDC"
-     "\x00\x80"
-     "\xEF\x38\x00\x00\x76\xDC\x00\x80\x7A\xDC\x00\x80\xEF\x6C\x07\x00\x7E\xDC"
-     "\x00\x80"
-     "\x80\x61\x00\x00\x81\x61\x00\x00\x82\x61\x00\x00\xB3\x61\x02\x00\x82\xDC"
-     "\x00\x80"
-     "\xB5\x61\x02\x00\xB6\x61\x02\x00\x86\xDC\x00\x80\x86\xC0\x0C\x00\x87\xE8"
-     "\x0C\x00"
-     "\xBA\xE1\x01\x00\xBB\xE1\x01\x00\xBC\xE1\x01\x00\xBD\xE9\x01\x00\xBE\xD1"
-     "\x01\x00"
-     "\xBF\xD1\x01\x00\xA3\xAD\x0D\x00\x02\xDC\x00\x80\x72\xDC\x00\x80\x8A\xDC"
-     "\x00\x80"
-     "\x8E\xDC\x00\x80\xA6\xAD\x0D\x00\xA5\xAD\x0D\x00\x92\xDC\x00\x80\xAB\x2D"
-     "\x0E\x00"
-     "\xAA\x2D\x0E\x00\x96\xDC\x00\x80\x9A\xDC\x00\x80\xAF\x1D\x0E\x00\xAE\x1D"
-     "\x0E\x00"
-     "\xAD\x25\x0E\x00\xAC\x2D\x0E\x00\x9E\xDC\x00\x80\xB3\xE9\x0E\x00\xA2\xDC"
-     "\x00\x80"
-     "\xA6\xDC\x00\x80\xB6\xE9\x0E\x00\xAA\xDC\x00\x80\xAE\xDC\x00\x80\xB5\xE9"
-     "\x0E\x00"
-     "\xBA\xB5\x0E\x00\xBB\x85\x0E\x00\xB2\xDC\x00\x80\xB6\xDC\x00\x80\xBE\x9D"
-     "\x0E\x00"
-     "\xBF\x85\x0E\x00\xBC\x9D\x0E\x00\xBD\x95\x0E\x00\xA8\x25\x0E\x00\xA9\x2D"
-     "\x0E\x00"
-     "\xAA\x25\x0E\x00\xAB\x3D\x0E\x00\xAC\x25\x0E\x00\xAD\x2D\x0E\x00\xAE\x25"
-     "\x0E\x00"
-     "\xAF\x4D\x0E\x00\xBA\xDC\x00\x80\xBE\xDC\x00\x80\xC2\xDC\x00\x80\xC6\xDC"
-     "\x00\x80"
-     "\xCA\xDC\x00\x80\xCE\xDC\x00\x80\xD2\xDC\x00\x80\xD6\xDC\x00\x80\xB8\xED"
-     "\x0F\x00"
-     "\xB9\xF5\x0F\x00\xBA\xFD\x0F\x00\xBB\xF5\x0F\x00\xBC\xED\x0F\x00\xBD\x51"
-     "\x0F\x00"
-     "\xBE\x49\x0F\x00\xBF\x45\x0F\x00\xB0\x35\x0E\x00\xB1\x05\x0E\x00\xB2\x0D"
-     "\x0E\x00"
-     "\xB3\x05\x0E\x00\xB4\x1D\x0E\x00\xB5\x05\x0E\x00\xB6\xDD\x0F\x00\xB7\xD5"
-     "\x0F\x00"
-     "\xA3\xAD\x0F\x00\x82\x19\x00\x00\x81\x19\x00\x00\x80\xB1\x01\x00\xDA\xDC"
-     "\x00\x80"
-     "\xA6\xAD\x0F\x00\xA5\xAD\x0F\x00\xDE\xDC\x00\x80\xAB\xC1\x0F\x00\xAA\xF1"
-     "\x0F\x00"
-     "\x84\xA0\x02\x00\xBE\xA0\x01\x00\xAF\xC1\x0F\x00\xAE\xD9\x0F\x00\xAD\xD1"
-     "\x0F\x00"
-     "\xAC\xD9\x0F\x00\xE2\xDC\x00\x80\xB3\xF9\x0E\x00\x86\xE8\x00\x00\x87\x2C"
-     "\x01\x00"
-     "\xB6\x5D\x01\x00\xE6\xDC\x00\x80\xEA\xDC\x00\x80\xB5\x81\x0E\x00\xBA\x49"
-     "\x01\x00"
-     "\xBB\x49\x01\x00\xEE\xDC\x00\x80\xF2\xDC\x00\x80\xBE\x3D\x01\x00\xBF\x21"
-     "\x01\x00"
-     "\xBC\x39\x01\x00\xBD\x35\x01\x00\xA8\x31\x0E\x00\xA9\x31\x0E\x00\xAA\x35"
-     "\x0E\x00"
-     "\xAB\x5D\x0E\x00\xAC\x4D\x0E\x00\xAD\x99\x01\x00\xAE\x89\x01\x00\xAF\x81"
-     "\x01\x00"
-     "\x84\xE4\x01\x00\xF6\xDC\x00\x80\xFA\xDC\x00\x80\xFE\xDC\x00\x80\x02\xDD"
-     "\x00\x80"
-     "\x06\xDD\x00\x80\x0A\xDD\x00\x80\x0E\xDD\x00\x80\xB8\xAD\x01\x00\xB9\x65"
-     "\x01\x00"
-     "\xBA\x6D\x01\x00\xBB\x65\x01\x00\xBC\x7D\x01\x00\xBD\x65\x01\x00\xBE\x6D"
-     "\x01\x00"
-     "\xBF\x65\x01\x00\xB0\xC9\x01\x00\xB1\xC9\x01\x00\xB2\xA9\x01\x00\xB3\xA5"
-     "\x01\x00"
-     "\xB4\xBD\x01\x00\xB5\xA1\x01\x00\xB6\xA1\x01\x00\xB7\x95\x01\x00\xA3\xBD"
-     "\x0D\x00"
-     "\x12\xDD\x00\x80\x16\xDD\x00\x80\x1A\xDD\x00\x80\x1E\xDD\x00\x80\xA6\x19"
-     "\x02\x00"
-     "\xA5\xC5\x0D\x00\x22\xDD\x00\x80\xAB\x0D\x02\x00\xAA\x0D\x02\x00\x26\xDD"
-     "\x00\x80"
-     "\x2A\xDD\x00\x80\xAF\x65\x02\x00\xAE\x79\x02\x00\xAD\x71\x02\x00\xAC\x7D"
-     "\x02\x00"
-     "\x2E\xDD\x00\x80\x32\xDD\x00\x80\x36\xDD\x00\x80\x3A\xDD\x00\x80\x3E\xDD"
-     "\x00\x80"
-     "\x42\xDD\x00\x80\x46\xDD\x00\x80\x4A\xDD\x00\x80\x80\x39\x00\x00\x81\x39"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x4E\xDD\x00\x80\xBE\x3C\x03\x00\x56\xDD\x00\x80\x87\x44"
-     "\x03\x00"
-     "\x86\xDC\x04\x00\x5A\xDD\x00\x80\x5E\xDD\x00\x80\xB3\xE1\x03\x00\x84\xC0"
-     "\x04\x00"
-     "\xB5\x35\x03\x00\x62\xDD\x00\x80\x66\xDD\x00\x80\xB6\x35\x03\x00\x6A\xDD"
-     "\x00\x80"
-     "\x6E\xDD\x00\x80\xBB\x11\x03\x00\xBA\x19\x03\x00\xBD\xB5\x00\x00\xBC\x09"
-     "\x03\x00"
-     "\xBF\xB5\x00\x00\xBE\xB5\x00\x00\x72\xDD\x00\x80\xE1\x94\x01\x00\x76\xDD"
-     "\x00\x80"
-     "\xE3\x7C\x01\x00\x7A\xDD\x00\x80\x7E\xDD\x00\x80\x82\xDD\x00\x80\x86\xDD"
-     "\x00\x80"
-     "\x8A\xDD\x00\x80\x8E\xDD\x00\x80\x92\xDD\x00\x80\x96\xDD\x00\x80\x9A\xDD"
-     "\x00\x80"
-     "\x9E\xDD\x00\x80\xA2\xDD\x00\x80\xEF\x74\x02\x00\xA8\x55\x02\x00\xA9\x5D"
-     "\x02\x00"
-     "\xAA\x61\x02\x00\xAB\x61\x02\x00\xAC\xB5\x02\x00\xAD\xBD\x02\x00\xAE\xB5"
-     "\x02\x00"
-     "\xAF\xAD\x02\x00\xBE\x68\x05\x00\xA6\xDD\x00\x80\xAA\xDD\x00\x80\xAE\xDD"
-     "\x00\x80"
-     "\xB2\xDD\x00\x80\x82\xAD\x00\x00\x81\xAD\x00\x00\x80\xAD\x00\x00\xB8\x69"
-     "\x01\x00"
-     "\xB9\x69\x01\x00\xBA\x09\x01\x00\xBB\x09\x01\x00\xBC\x19\x01\x00\xBD\x19"
-     "\x01\x00"
-     "\xBE\x09\x01\x00\xBF\x09\x01\x00\xB0\xD5\x02\x00\xB1\xDD\x02\x00\xB2\xD5"
-     "\x02\x00"
-     "\xB3\x69\x01\x00\xB4\x79\x01\x00\xB5\x79\x01\x00\xB6\x69\x01\x00\xB7\x61"
-     "\x01\x00"
-     "\xE1\x44\x0F\x00\xE3\xBC\x00\x00\xE3\xB8\x0E\x00\xE1\x74\x0E\x00\xB6\xDD"
-     "\x00\x80"
-     "\xBA\xDD\x00\x80\xBE\xDD\x00\x80\xC6\xDD\x00\x80\xCA\xDD\x00\x80\xCE\xDD"
-     "\x00\x80"
-     "\xD2\xDD\x00\x80\xD6\xDD\x00\x80\xBE\x30\x07\x00\xEF\x60\x00\x00\xEF\xCC"
-     "\x0E\x00"
-     "\xDA\xDD\x00\x80\xA3\xA9\x02\x00\xDE\xDD\x00\x80\x86\xE8\x04\x00\x87\x04"
-     "\x05\x00"
-     "\xE2\xDD\x00\x80\xA6\x7D\x02\x00\xA5\x7D\x02\x00\xE6\xDD\x00\x80\xAB\x59"
-     "\x02\x00"
-     "\xAA\x51\x02\x00\xEA\xDD\x00\x80\xEE\xDD\x00\x80\xAF\xFD\x01\x00\xAE\xFD"
-     "\x01\x00"
-     "\xAD\xFD\x01\x00\xAC\x41\x02\x00\xA8\x75\x06\x00\xA9\x85\x06\x00\xAA\x8D"
-     "\x06\x00"
-     "\xAB\x99\x06\x00\xAC\x89\x06\x00\xAD\xB5\x06\x00\xAE\xBD\x06\x00\xAF\xDD"
-     "\x06\x00"
-     "\xC2\xDD\x00\x80\xF2\xDD\x00\x80\xF6\xDD\x00\x80\xFA\xDD\x00\x80\xFE\xDD"
-     "\x00\x80"
-     "\x02\xDE\x00\x80\x06\xDE\x00\x80\x0A\xDE\x00\x80\xB8\xAD\x06\x00\xB9\xB5"
-     "\x06\x00"
-     "\xBA\xBD\x06\x00\xBB\xB5\x06\x00\xBC\x75\x07\x00\xBD\x7D\x07\x00\xBE\x75"
-     "\x07\x00"
-     "\xBF\x6D\x07\x00\xB0\xA9\x06\x00\xB1\xB5\x06\x00\xB2\xBD\x06\x00\xB3\xB5"
-     "\x06\x00"
-     "\xB4\xAD\x06\x00\xB5\x95\x06\x00\xB6\x9D\x06\x00\xB7\x95\x06\x00\xB3\x05"
-     "\x06\x00"
-     "\x0E\xDE\x00\x80\x12\xDE\x00\x80\x16\xDE\x00\x80\x1A\xDE\x00\x80\xB6\x35"
-     "\x06\x00"
-     "\xB5\x01\x06\x00\x1E\xDE\x00\x80\xBB\x71\x06\x00\xBA\x6D\x06\x00\x22\xDE"
-     "\x00\x80"
-     "\x26\xDE\x00\x80\xBF\x51\x06\x00\xBE\x69\x06\x00\xBD\x61\x06\x00\xBC\x69"
-     "\x06\x00"
-     "\x82\x19\x00\x00\xA3\x41\x06\x00\x80\x65\x00\x00\x81\x19\x00\x00\xA6\x71"
-     "\x06\x00"
-     "\x2A\xDE\x00\x80\x2E\xDE\x00\x80\xA5\x45\x06\x00\xAA\x29\x06\x00\xAB\x35"
-     "\x06\x00"
-     "\x84\x24\x03\x00\x32\xDE\x00\x80\xAE\x2D\x06\x00\xAF\x15\x06\x00\xAC\x2D"
-     "\x06\x00"
-     "\xAD\x25\x06\x00\xB3\xA5\x06\x00\x36\xDE\x00\x80\x86\x48\x01\x00\x87\x0C"
-     "\x01\x00"
-     "\x3A\xDE\x00\x80\xB6\xA9\x06\x00\xB5\xB5\x06\x00\x3E\xDE\x00\x80\xBB\xC9"
-     "\x06\x00"
-     "\xBA\xF9\x06\x00\x42\xDE\x00\x80\x46\xDE\x00\x80\xBF\x1D\x01\x00\xBE\x1D"
-     "\x01\x00"
-     "\xBD\x1D\x01\x00\xBC\xD1\x06\x00\x4A\xDE\x00\x80\x4E\xDE\x00\x80\x52\xDE"
-     "\x00\x80"
-     "\x56\xDE\x00\x80\xE3\xCC\x05\x00\x5A\xDE\x00\x80\xE1\x3C\x05\x00\x5E\xDE"
-     "\x00\x80"
-     "\xEF\xB8\x04\x00\x62\xDE\x00\x80\x66\xDE\x00\x80\x6A\xDE\x00\x80\x6E\xDE"
-     "\x00\x80"
-     "\x72\xDE\x00\x80\x76\xDE\x00\x80\x7A\xDE\x00\x80\xA8\x1D\x06\x00\xA9\x25"
-     "\x06\x00"
-     "\xAA\x2D\x06\x00\xAB\x25\x06\x00\xAC\x3D\x06\x00\xAD\x29\x06\x00\xAE\x5D"
-     "\x06\x00"
-     "\xAF\x4D\x06\x00\x7E\xDE\x00\x80\x82\xDE\x00\x80\x86\xDE\x00\x80\x8A\xDE"
-     "\x00\x80"
-     "\x8E\xDE\x00\x80\x92\xDE\x00\x80\x96\xDE\x00\x80\x9A\xDE\x00\x80\xB8\xF9"
-     "\x01\x00"
-     "\xB9\xF9\x01\x00\xBA\x4D\x01\x00\xBB\x45\x01\x00\xBC\x5D\x01\x00\xBD\x45"
-     "\x01\x00"
-     "\xBE\x45\x01\x00\xBF\x75\x01\x00\xB0\x3D\x06\x00\xB1\x0D\x06\x00\xB2\x05"
-     "\x06\x00"
-     "\xB3\xC9\x01\x00\xB4\xD9\x01\x00\xB5\xD9\x01\x00\xB6\xC9\x01\x00\xB7\xC9"
-     "\x01\x00"
-     "\x80\xB9\x03\x00\x81\xB9\x03\x00\x82\x19\x00\x00\xA3\xA9\x05\x00\x9E\xDE"
-     "\x00\x80"
-     "\xA5\xB9\x05\x00\xA6\xA5\x05\x00\x84\x84\x02\x00\xA6\xDE\x00\x80\xBE\xA8"
-     "\x03\x00"
-     "\xAA\xF5\x05\x00\xAB\xC5\x05\x00\xAC\xDD\x05\x00\xAD\x11\x02\x00\xAE\x11"
-     "\x02\x00"
-     "\xAF\x11\x02\x00\xB6\x01\x03\x00\xAA\xDE\x00\x80\x84\x44\x03\x00\xB5\x61"
-     "\x03\x00"
-     "\xAE\xDE\x00\x80\xB3\x49\x03\x00\x87\xD8\x03\x00\x86\xF8\x1C\x00\xBE\x09"
-     "\x03\x00"
-     "\xBF\x09\x03\x00\xBC\x35\x03\x00\xBD\x09\x03\x00\xBA\x3D\x03\x00\xBB\x21"
-     "\x03\x00"
-     "\xB2\xDE\x00\x80\xB6\xDE\x00\x80\xB0\xCD\x03\x00\xB1\xD1\x03\x00\xB2\xD5"
-     "\x03\x00"
-     "\xB3\xE9\x03\x00\xB4\xF9\x03\x00\xB5\xF9\x03\x00\xB6\xE9\x03\x00\xB7\xE9"
-     "\x03\x00"
-     "\xB8\xD9\x03\x00\xB9\xD9\x03\x00\xBA\x09\x03\x00\xBB\x09\x03\x00\xBC\x19"
-     "\x03\x00"
-     "\xBD\x19\x03\x00\xBE\x09\x03\x00\xBF\x09\x03\x00\xBA\xDE\x00\x80\xBE\xDE"
-     "\x00\x80"
-     "\xB5\x69\x21\x00\xC2\xDE\x00\x80\xC6\xDE\x00\x80\xCA\xDE\x00\x80\xCE\xDE"
-     "\x00\x80"
-     "\x52\xDD\x00\x80\xA8\xE5\x03\x00\xA9\xF5\x03\x00\xAA\xF5\x03\x00\xAB\xC5"
-     "\x03\x00"
-     "\xAC\xDD\x03\x00\xAD\xB5\x03\x00\xAE\xBD\x03\x00\xAF\xB5\x03\x00\xA3\x1D"
-     "\x02\x00"
-     "\xD2\xDE\x00\x80\xD6\xDE\x00\x80\xDA\xDE\x00\x80\xBE\x1C\x1D\x00\xA6\x55"
-     "\x02\x00"
-     "\xA5\x35\x02\x00\xDE\xDE\x00\x80\xAB\x75\x02\x00\xAA\x69\x02\x00\xE2\xDE"
-     "\x00\x80"
-     "\xE6\xDE\x00\x80\xAF\x5D\x02\x00\xAE\x5D\x02\x00\xAD\x5D\x02\x00\xAC\x61"
-     "\x02\x00"
-     "\xEA\xDE\x00\x80\xEE\xDE\x00\x80\xF2\xDE\x00\x80\xF6\xDE\x00\x80\xEF\x8C"
-     "\x03\x00"
-     "\xFA\xDE\x00\x80\xFE\xDE\x00\x80\x02\xDF\x00\x80\xE3\x6C\x03\x00\x06\xDF"
-     "\x00\x80"
-     "\xE1\x78\x00\x00\x0A\xDF\x00\x80\x0E\xDF\x00\x80\x16\xDF\x00\x80\x86\x58"
-     "\x1C\x00"
-     "\x87\xF0\x1D\x00\x80\xE5\x00\x00\x81\xE9\x00\x00\x82\xF9\x00\x00\x1A\xDF"
-     "\x00\x80"
-     "\x1E\xDF\x00\x80\x22\xDF\x00\x80\x26\xDF\x00\x80\xEF\xD4\x06\x00\x2A\xDF"
-     "\x00\x80"
-     "\xE1\x74\x06\x00\x2E\xDF\x00\x80\xE3\x54\x01\x00\x32\xDF\x00\x80\x36\xDF"
-     "\x00\x80"
-     "\x3A\xDF\x00\x80\x3E\xDF\x00\x80\xB3\x91\x01\x00\x42\xDF\x00\x80\x46\xDF"
-     "\x00\x80"
-     "\x84\xC4\x1D\x00\x4A\xDF\x00\x80\xB6\x15\x01\x00\xB5\xA5\x01\x00\x4E\xDF"
-     "\x00\x80"
-     "\xBB\x35\x01\x00\xBA\x35\x01\x00\x52\xDF\x00\x80\x56\xDF\x00\x80\xBF\xFD"
-     "\x00\x00"
-     "\xBE\xFD\x00\x00\xBD\xFD\x00\x00\xBC\x25\x01\x00\xA9\x45\x1E\x00\xA8\x25"
-     "\x1E\x00"
-     "\xAB\x71\x1E\x00\xAA\x5D\x1E\x00\xAD\xD1\x1E\x00\xAC\x69\x1E\x00\xAF\xD5"
-     "\x1E\x00"
-     "\xAE\xD9\x1E\x00\x12\xDF\x00\x80\x5A\xDF\x00\x80\x5E\xDF\x00\x80\x62\xDF"
-     "\x00\x80"
-     "\x66\xDF\x00\x80\x6A\xDF\x00\x80\x6E\xDF\x00\x80\x72\xDF\x00\x80\xB9\x01"
-     "\x1F\x00"
-     "\xB8\x69\x1F\x00\xBB\x01\x1F\x00\xBA\x09\x1F\x00\xBD\x05\x1F\x00\xBC\x19"
-     "\x1F\x00"
-     "\xBF\x05\x1F\x00\xBE\x05\x1F\x00\xB1\xB5\x1E\x00\xB0\xB5\x1E\x00\xB3\x6D"
-     "\x1F\x00"
-     "\xB2\xB5\x1E\x00\xB5\x7D\x1F\x00\xB4\x7D\x1F\x00\xB7\x51\x1F\x00\xB6\x61"
-     "\x1F\x00"
-     "\xA3\xD1\x1E\x00\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D\x00\x00\x76\xDF"
-     "\x00\x80"
-     "\xA6\x55\x1E\x00\xA5\xE5\x1E\x00\x7A\xDF\x00\x80\xAB\x75\x1E\x00\xAA\x75"
-     "\x1E\x00"
-     "\x7E\xDF\x00\x80\xBE\x60\x03\x00\xAF\xBD\x1F\x00\xAE\xBD\x1F\x00\xAD\xBD"
-     "\x1F\x00"
-     "\xAC\x65\x1E\x00\x82\xDF\x00\x80\xB3\x55\x1F\x00\x86\x68\x00\x00\x87\x6C"
-     "\x03\x00"
-     "\xB6\x71\x1F\x00\x86\xDF\x00\x80\x8A\xDF\x00\x80\xB5\x45\x1F\x00\xBA\x55"
-     "\x1F\x00"
-     "\xBB\x5D\x1F\x00\x8E\xDF\x00\x80\x92\xDF\x00\x80\xBE\x39\x1F\x00\xBF\x25"
-     "\x1F\x00"
-     "\xBC\x45\x1F\x00\xBD\x39\x1F\x00\x96\xDF\x00\x80\x9A\xDF\x00\x80\x9E\xDF"
-     "\x00\x80"
-     "\xA2\xDF\x00\x80\xA6\xDF\x00\x80\xAA\xDF\x00\x80\xAE\xDF\x00\x80\xEF\xF4"
-     "\x1C\x00"
-     "\xB2\xDF\x00\x80\xE1\x60\x1D\x00\xB6\xDF\x00\x80\xE3\x74\x1D\x00\xBA\xDF"
-     "\x00\x80"
-     "\xBE\xDF\x00\x80\xC2\xDF\x00\x80\xC6\xDF\x00\x80\xA3\xD1\x1E\x00\xCA\xDF"
-     "\x00\x80"
-     "\xCE\xDF\x00\x80\xD2\xDF\x00\x80\xD6\xDF\x00\x80\xA6\xF5\x1E\x00\xA5\xC1"
-     "\x1E\x00"
-     "\xDA\xDF\x00\x80\xAB\xD9\x1E\x00\xAA\xD1\x1E\x00\xDE\xDF\x00\x80\xE2\xDF"
-     "\x00\x80"
-     "\xAF\xA1\x1E\x00\xAE\xBD\x1E\x00\xAD\xBD\x1E\x00\xAC\xC1\x1E\x00\xA8\xF5"
-     "\x1E\x00"
-     "\xA9\x09\x1E\x00\xAA\x19\x1E\x00\xAB\x19\x1E\x00\xAC\x09\x1E\x00\xAD\x35"
-     "\x1E\x00"
-     "\xAE\x31\x1E\x00\xAF\x2D\x1E\x00\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\xE6\xDF\x00\x80\xEA\xDF\x00\x80\xEE\xDF\x00\x80\x84\xFC\x01\x00\xBE\xF0"
-     "\x01\x00"
-     "\xB8\x2D\x1E\x00\xB9\xE1\x01\x00\xBA\xE1\x01\x00\xBB\xE1\x01\x00\xBC\xE1"
-     "\x01\x00"
-     "\xBD\xE1\x01\x00\xBE\xE1\x01\x00\xBF\xE1\x01\x00\xB0\x55\x1E\x00\xB1\x5D"
-     "\x1E\x00"
-     "\xB2\x55\x1E\x00\xB3\x29\x1E\x00\xB4\x39\x1E\x00\xB5\x25\x1E\x00\xB6\x25"
-     "\x1E\x00"
-     "\xB7\x15\x1E\x00\xA8\xD5\x02\x00\xA9\xDD\x02\x00\xAA\xE1\x02\x00\xAB\xE1"
-     "\x02\x00"
-     "\xAC\x11\x03\x00\xAD\x11\x03\x00\xAE\x11\x03\x00\xAF\x11\x03\x00\xF6\xDF"
-     "\x00\x80"
-     "\xFA\xDF\x00\x80\xFE\xDF\x00\x80\x02\xE0\x00\x80\x06\xE0\x00\x80\x0A\xE0"
-     "\x00\x80"
-     "\x0E\xE0\x00\x80\x12\xE0\x00\x80\xB8\x7D\x03\x00\xB9\x01\x03\x00\xBA\x01"
-     "\x03\x00"
-     "\xBB\x01\x03\x00\xBC\x01\x03\x00\xBD\x09\x03\x00\xBE\x31\x03\x00\xBF\x31"
-     "\x03\x00"
-     "\xB0\x71\x03\x00\xB1\x71\x03\x00\xB2\x71\x03\x00\xB3\x4D\x03\x00\xB4\x59"
-     "\x03\x00"
-     "\xB5\x45\x03\x00\xB6\x4D\x03\x00\xB7\x45\x03\x00\x86\x20\x04\x00\x87\xA0"
-     "\x03\x00"
-     "\xB3\x61\x02\x00\xBE\x80\x04\x00\xB5\x91\x03\x00\x16\xE0\x00\x80\x1A\xE0"
-     "\x00\x80"
-     "\xB6\x99\x03\x00\xA2\xDE\x00\x80\x84\xB8\x03\x00\xBB\xBD\x03\x00\xBA\xBD"
-     "\x03\x00"
-     "\xBD\x91\x03\x00\xBC\xAD\x03\x00\xBF\x89\x03\x00\xBE\x95\x03\x00\xA3\x29"
-     "\x02\x00"
-     "\x1E\xE0\x00\x80\x22\xE0\x00\x80\x26\xE0\x00\x80\x85\x94\x06\x00\xA6\xD1"
-     "\x03\x00"
-     "\xA5\xD9\x03\x00\x2A\xE0\x00\x80\xAB\xF5\x03\x00\xAA\xF5\x03\x00\x2E\xE0"
-     "\x00\x80"
-     "\x32\xE0\x00\x80\xAF\xC1\x03\x00\xAE\xDD\x03\x00\xAD\xD9\x03\x00\xAC\xE5"
-     "\x03\x00"
-     "\x36\xE0\x00\x80\x3A\xE0\x00\x80\x3E\xE0\x00\x80\x42\xE0\x00\x80\x46\xE0"
-     "\x00\x80"
-     "\x4A\xE0\x00\x80\x4E\xE0\x00\x80\x52\xE0\x00\x80\x81\x11\x00\x00\x80\x0D"
-     "\x00\x00"
-     "\x56\xE0\x00\x80\x82\x19\x00\x00\x5A\xE0\x00\x80\x5E\xE0\x00\x80\x66\xE0"
-     "\x00\x80"
-     "\x6A\xE0\x00\x80\x6E\xE0\x00\x80\xE1\x94\x01\x00\x72\xE0\x00\x80\xE3\xE4"
-     "\x00\x00"
-     "\x76\xE0\x00\x80\x7A\xE0\x00\x80\x86\x78\x04\x00\x87\x80\x02\x00\x7E\xE0"
-     "\x00\x80"
-     "\x82\xE0\x00\x80\x86\xE0\x00\x80\x8A\xE0\x00\x80\x8E\xE0\x00\x80\x92\xE0"
-     "\x00\x80"
-     "\x96\xE0\x00\x80\xEF\x38\x00\x00\x9A\xE0\x00\x80\x9E\xE0\x00\x80\xA2\xE0"
-     "\x00\x80"
-     "\xA6\xE0\x00\x80\xAA\xE0\x00\x80\xAE\xE0\x00\x80\xB2\xE0\x00\x80\xEF\xC8"
-     "\x01\x00"
-     "\xBE\x6C\x04\x00\xE1\xC4\x1E\x00\x84\xA0\x05\x00\xE3\x70\x1E\x00\xB6\xE0"
-     "\x00\x80"
-     "\xBA\xE0\x00\x80\xBE\xE0\x00\x80\xC2\xE0\x00\x80\xC6\xE0\x00\x80\xB6\xF1"
-     "\x01\x00"
-     "\xB5\xF9\x01\x00\x85\xF4\x20\x00\xB3\xF9\x01\x00\x82\x29\x00\x00\x81\x21"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\xBF\x3D\x01\x00\xBE\x29\x01\x00\xBD\x21\x01\x00\xBC\x29"
-     "\x01\x00"
-     "\xBB\x69\x01\x00\xBA\x61\x01\x00\xCA\xE0\x00\x80\xCE\xE0\x00\x80\xA8\x21"
-     "\x06\x00"
-     "\xA9\x21\x06\x00\xAA\x21\x06\x00\xAB\x21\x06\x00\xAC\x21\x06\x00\xAD\x21"
-     "\x06\x00"
-     "\xAE\x21\x06\x00\xAF\x21\x06\x00\x62\xE0\x00\x80\xBF\x2C\x01\x00\x86\x00"
-     "\x0C\x00"
-     "\x87\x60\x03\x00\xD2\xE0\x00\x80\xD6\xE0\x00\x80\xDA\xE0\x00\x80\xDE\xE0"
-     "\x00\x80"
-     "\xB8\x41\x07\x00\xB9\x41\x07\x00\xBA\x41\x07\x00\xBB\x41\x07\x00\xBC\x41"
-     "\x07\x00"
-     "\xBD\x41\x07\x00\xBE\x41\x07\x00\xBF\x41\x07\x00\xB0\xF5\x06\x00\xB1\xFD"
-     "\x06\x00"
-     "\xB2\xAD\x06\x00\xB3\xA5\x06\x00\xB4\xBD\x06\x00\xB5\xA9\x06\x00\xB6\x91"
-     "\x06\x00"
-     "\xB7\x91\x06\x00\xA3\xA9\x06\x00\xE2\xE0\x00\x80\xE6\xE0\x00\x80\xEA\xE0"
-     "\x00\x80"
-     "\xEE\xE0\x00\x80\xA6\xA1\x06\x00\xA5\xA9\x06\x00\xF2\xE0\x00\x80\xAB\x39"
-     "\x06\x00"
-     "\xAA\x31\x06\x00\xF6\xE0\x00\x80\xFA\xE0\x00\x80\xAF\x6D\x06\x00\xAE\x79"
-     "\x06\x00"
-     "\xAD\x71\x06\x00\xAC\x79\x06\x00\xFE\xE0\x00\x80\xB3\x85\x07\x00\x02\xE1"
-     "\x00\x80"
-     "\x06\xE1\x00\x80\xB6\x85\x07\x00\x0A\xE1\x00\x80\x0E\xE1\x00\x80\xB5\x8D"
-     "\x07\x00"
-     "\xBA\x89\x07\x00\xBB\x89\x07\x00\x12\xE1\x00\x80\x16\xE1\x00\x80\xBE\x89"
-     "\x07\x00"
-     "\xBF\x75\x07\x00\xBC\x89\x07\x00\xBD\x89\x07\x00\x1A\xE1\x00\x80\xE1\xC0"
-     "\x05\x00"
-     "\xEF\x7C\x05\x00\xE3\xD4\x04\x00\x1E\xE1\x00\x80\x22\xE1\x00\x80\x26\xE1"
-     "\x00\x80"
-     "\x2A\xE1\x00\x80\x2E\xE1\x00\x80\x32\xE1\x00\x80\x36\xE1\x00\x80\x3A\xE1"
-     "\x00\x80"
-     "\xE1\xA8\x05\x00\x3E\xE1\x00\x80\xE3\x8C\x05\x00\xEF\xF8\x0C\x00\xA3\x01"
-     "\x06\x00"
-     "\x82\x71\x00\x00\x81\x65\x00\x00\x80\x65\x00\x00\x84\x80\x01\x00\xA6\x01"
-     "\x06\x00"
-     "\xA5\x09\x06\x00\xBE\x10\x00\x00\xAB\x0D\x06\x00\xAA\x0D\x06\x00\x86\xE8"
-     "\x0F\x00"
-     "\x87\x18\x00\x00\xAF\xF1\x06\x00\xAE\x0D\x06\x00\xAD\x0D\x06\x00\xAC\x0D"
-     "\x06\x00"
-     "\xA8\xB9\x06\x00\xA9\xB9\x06\x00\xAA\xC9\x06\x00\xAB\xC5\x06\x00\xAC\xC1"
-     "\x06\x00"
-     "\xAD\xC1\x06\x00\xAE\xF9\x06\x00\xAF\xF9\x06\x00\x46\xE1\x00\x80\x4A\xE1"
-     "\x00\x80"
-     "\x4E\xE1\x00\x80\x52\xE1\x00\x80\x56\xE1\x00\x80\x5A\xE1\x00\x80\x5E\xE1"
-     "\x00\x80"
-     "\x62\xE1\x00\x80\xB8\x4D\x06\x00\xB9\x55\x06\x00\xBA\x5D\x06\x00\xBB\x55"
-     "\x06\x00"
-     "\xBC\x61\x06\x00\xBD\x61\x06\x00\xBE\x61\x06\x00\xBF\x61\x06\x00\xB0\x89"
-     "\x06\x00"
-     "\xB1\x89\x06\x00\xB2\x99\x06\x00\xB3\x99\x06\x00\xB4\x89\x06\x00\xB5\x75"
-     "\x06\x00"
-     "\xB6\x7D\x06\x00\xB7\x75\x06\x00\x66\xE1\x00\x80\xB5\x8D\x02\x00\xB6\x85"
-     "\x02\x00"
-     "\x84\x40\x0C\x00\x8F\xF1\xE8\x03\x6A\xE1\x00\x80\x6E\xE1\x00\x80\xB3\x21"
-     "\x02\x00"
-     "\xBC\xE1\x02\x00\xBD\xED\x02\x00\xBE\xE5\x02\x00\xBF\xF9\x02\x00\xAC\xC8"
-     "\x32\x00"
-     "\x72\xE1\x00\x80\xBA\x91\x02\x00\xBB\x91\x02\x00\x9B\x01\x1F\x00\x9A\x4D"
-     "\x1E\x00"
-     "\x99\x4D\x1E\x00\x98\x3D\x13\x00\x9F\xF1\x04\x00\x9E\xB9\x1A\x00\x9D\x79"
-     "\x1A\x00"
-     "\x9C\x81\x18\x00\x93\x41\x16\x00\x92\x91\x14\x00\x91\xF9\xEA\x03\x90\x4D"
-     "\xEA\x03"
-     "\x97\x61\x12\x00\x96\x71\x12\x00\x95\x81\x10\x00\x94\x41\x16\x00\xB1\x01"
-     "\x38\x00"
-     "\xB0\xE1\x3E\x00\xB3\x81\x3A\x00\xB2\x89\x3A\x00\xB5\x91\x26\x00\xB4\xD1"
-     "\x3B\x00"
-     "\x76\xE1\x00\x80\x7A\xE1\x00\x80\xA3\x81\x02\x00\xA2\x11\x00\x00\xA1\x69"
-     "\x06\x00"
-     "\xA0\x8D\x06\x00\x7E\xE1\x00\x80\xF2\xDF\x00\x80\x82\xE1\x00\x80\x86\xE1"
-     "\x00\x80"
-     "\xA1\xF1\x02\x00\x8A\xE1\x00\x80\xA3\xCD\x0E\x00\xA2\xCD\x0E\x00\xA5\x01"
-     "\x08\x00"
-     "\xA4\x51\x0F\x00\xA7\xC5\x0B\x00\xA6\xFD\x0A\x00\xA9\xC1\x36\x00\xA8\x01"
-     "\x34\x00"
-     "\xAB\x01\x30\x00\xAA\x41\x36\x00\xAD\xCD\x32\x00\xAC\xD5\x32\x00\xAF\xFD"
-     "\x3E\x00"
-     "\xAE\x01\x3C\x00\x82\xE1\xE9\x03\x83\x25\xEB\x03\x8E\xE1\x00\x80\x92\xE1"
-     "\x00\x80"
-     "\x86\x31\x17\x00\x87\x29\x17\x00\x84\x89\xEB\x03\x85\xE9\x15\x00\x8A\xD9"
-     "\x11\x00"
-     "\x8B\x21\x13\x00\x96\xE1\x00\x80\x9A\xE1\x00\x80\x8E\x35\x1F\x00\x8F\x3D"
-     "\x1F\x00"
-     "\x8C\x29\x13\x00\x8D\x4D\x12\x00\x92\x01\x1E\x00\x93\x99\x19\x00\x9E\xE1"
-     "\x00\x80"
-     "\x84\x00\x0C\x00\x96\xF1\x05\x00\x97\x75\x07\x00\x94\x79\x1B\x00\x95\xB1"
-     "\x1B\x00"
-     "\x9A\xA1\x07\x00\x9B\xD1\x01\x00\xA2\xE1\x00\x80\xA6\xE1\x00\x80\x81\xF1"
-     "\x00\x00"
-     "\x80\x19\x00\x00\x9C\x59\x03\x00\x82\xF9\x00\x00\xAA\xE1\x00\x80\xAE\xE1"
-     "\x00\x80"
-     "\xBE\xC0\x0D\x00\xEF\x28\x00\x00\x86\x30\x0C\x00\x87\x5C\x0D\x00\xB6\xE1"
-     "\x00\x80"
-     "\xEF\xC4\x06\x00\xBA\xE1\x00\x80\xE1\xE4\x06\x00\xBE\xE1\x00\x80\xE3\x14"
-     "\x06\x00"
-     "\xC2\xE1\x00\x80\xE1\x84\x01\x00\xC6\xE1\x00\x80\xE3\xFC\x00\x00\xA3\xFD"
-     "\x03\x00"
-     "\xCA\xE1\x00\x80\xCE\xE1\x00\x80\x84\x4C\x0F\x00\xD2\xE1\x00\x80\xA6\x59"
-     "\x03\x00"
-     "\xA5\x51\x03\x00\xD6\xE1\x00\x80\xAB\x4D\x03\x00\xAA\x4D\x03\x00\xDA\xE1"
-     "\x00\x80"
-     "\xDE\xE1\x00\x80\xAF\x25\x03\x00\xAE\x39\x03\x00\xAD\x31\x03\x00\xAC\x3D"
-     "\x03\x00"
-     "\x42\xE1\x00\x80\xB2\xE1\x00\x80\xE2\xE1\x00\x80\xE6\xE1\x00\x80\xEA\xE1"
-     "\x00\x80"
-     "\xEE\xE1\x00\x80\xF2\xE1\x00\x80\xF6\xE1\x00\x80\xA8\x81\x0D\x00\xA9\x81"
-     "\x0D\x00"
-     "\xAA\x81\x0D\x00\xAB\x81\x0D\x00\xAC\x81\x0D\x00\xAD\x81\x0D\x00\xAE\x81"
-     "\x0D\x00"
-     "\xAF\x81\x0D\x00\xB0\x41\x0E\x00\xB1\x51\x0E\x00\xB2\x51\x0E\x00\xB3\x79"
-     "\x0E\x00"
-     "\xB4\xC9\x0E\x00\xB5\xC9\x0E\x00\xB6\xE9\x0E\x00\xB7\xE5\x0E\x00\xB8\xDD"
-     "\x0E\x00"
-     "\xB9\xA1\x0E\x00\xBA\x61\x0F\x00\xBB\x61\x0F\x00\xBC\x59\x0F\x00\xBD\x59"
-     "\x0F\x00"
-     "\xBE\x31\x0F\x00\xBF\x15\x0F\x00\xB3\x29\x0D\x00\xFA\xE1\x00\x80\xFE\xE1"
-     "\x00\x80"
-     "\x02\xE2\x00\x80\x06\xE2\x00\x80\xB6\x21\x0D\x00\xB5\x29\x0D\x00\x0A\xE2"
-     "\x00\x80"
-     "\xBB\xB5\x0E\x00\xBA\xA5\x0E\x00\x0E\xE2\x00\x80\x12\xE2\x00\x80\xBF\x2D"
-     "\x0E\x00"
-     "\xBE\x2D\x0E\x00\xBD\x8D\x0E\x00\xBC\xA5\x0E\x00\x82\x29\x00\x00\xA3\x6D"
-     "\x0D\x00"
-     "\x80\x1D\x00\x00\x81\x15\x00\x00\xA6\x65\x0D\x00\x16\xE2\x00\x80\x1A\xE2"
-     "\x00\x80"
-     "\xA5\x6D\x0D\x00\xAA\xE1\x0E\x00\xAB\xF1\x0E\x00\x1E\xE2\x00\x80\xBE\xA4"
-     "\x03\x00"
-     "\xAE\x69\x0E\x00\xAF\x69\x0E\x00\xAC\xE1\x0E\x00\xAD\xC9\x0E\x00\xA8\x99"
-     "\x0E\x00"
-     "\xA9\xA5\x0E\x00\xAA\xA1\x0E\x00\xAB\xBD\x0E\x00\xAC\xA9\x0E\x00\xAD\xD1"
-     "\x0E\x00"
-     "\xAE\xD1\x0E\x00\xAF\xCD\x0E\x00\x86\x80\x00\x00\x87\x04\x01\x00\x26\xE2"
-     "\x00\x80"
-     "\x2A\xE2\x00\x80\x2E\xE2\x00\x80\x32\xE2\x00\x80\x36\xE2\x00\x80\x3A\xE2"
-     "\x00\x80"
-     "\xB8\x5D\x01\x00\xB9\xC1\x01\x00\xBA\xC1\x01\x00\xBB\xC1\x01\x00\xBC\xC1"
-     "\x01\x00"
-     "\xBD\xC9\x01\x00\xBE\xF1\x01\x00\xBF\xF1\x01\x00\xB0\xBD\x0E\x00\xB1\x8D"
-     "\x0E\x00"
-     "\xB2\x85\x0E\x00\xB3\x65\x01\x00\xB4\x7D\x01\x00\xB5\x65\x01\x00\xB6\x6D"
-     "\x01\x00"
-     "\xB7\x65\x01\x00\xB3\x29\x0E\x00\x22\xE2\x00\x80\x3E\xE2\x00\x80\x42\xE2"
-     "\x00\x80"
-     "\x46\xE2\x00\x80\xB6\x51\x0E\x00\xB5\x21\x0E\x00\x4A\xE2\x00\x80\xBB\x49"
-     "\x0E\x00"
-     "\xBA\x79\x0E\x00\x4E\xE2\x00\x80\x52\xE2\x00\x80\xBF\xA1\x01\x00\xBE\xB9"
-     "\x01\x00"
-     "\xBD\xB1\x01\x00\xBC\x51\x0E\x00\x56\xE2\x00\x80\xA3\x6D\x0E\x00\x5A\xE2"
-     "\x00\x80"
-     "\x5E\xE2\x00\x80\xA6\x15\x0E\x00\x62\xE2\x00\x80\x66\xE2\x00\x80\xA5\x65"
-     "\x0E\x00"
-     "\xAA\x3D\x0E\x00\xAB\x0D\x0E\x00\x6A\xE2\x00\x80\x6E\xE2\x00\x80\xAE\xFD"
-     "\x01\x00"
-     "\xAF\xE5\x01\x00\xAC\x15\x0E\x00\xAD\xF5\x01\x00\xE1\x38\x01\x00\x72\xE2"
-     "\x00\x80"
-     "\xE3\xD4\x0F\x00\x76\xE2\x00\x80\x7A\xE2\x00\x80\xBE\x58\x03\x00\x7E\xE2"
-     "\x00\x80"
-     "\x82\xE2\x00\x80\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\x86\xE2"
-     "\x00\x80"
-     "\x8A\xE2\x00\x80\x8E\xE2\x00\x80\xEF\xD4\x02\x00\x96\xE2\x00\x80\x86\xE0"
-     "\x04\x00"
-     "\x87\xE0\x03\x00\xBE\x00\x04\x00\xB3\x7D\x03\x00\x9A\xE2\x00\x80\xB5\x6D"
-     "\x03\x00"
-     "\xB6\x65\x03\x00\x9E\xE2\x00\x80\xA2\xE2\x00\x80\xA6\xE2\x00\x80\xBA\xF9"
-     "\x03\x00"
-     "\xBB\xF1\x03\x00\xBC\xD9\x03\x00\xBD\xD9\x03\x00\xBE\xB9\x03\x00\xBF\xB5"
-     "\x03\x00"
-     "\xAA\xBD\x0E\x00\xAB\x01\x0E\x00\xA8\xBD\x0E\x00\xA9\xB5\x0E\x00\xAE\x01"
-     "\x0E\x00"
-     "\xAF\x09\x0E\x00\xAC\x11\x0E\x00\xAD\x11\x0E\x00\xAA\xE2\x00\x80\xAE\xE2"
-     "\x00\x80"
-     "\xB2\xE2\x00\x80\xB6\xE2\x00\x80\xBA\xE2\x00\x80\xBE\xE2\x00\x80\xC2\xE2"
-     "\x00\x80"
-     "\xC6\xE2\x00\x80\xBA\xD1\x0F\x00\xBB\xD1\x0F\x00\xB8\x25\x0E\x00\xB9\xCD"
-     "\x0F\x00"
-     "\xBE\xF1\x0F\x00\xBF\xF1\x0F\x00\xBC\xC1\x0F\x00\xBD\xC1\x0F\x00\xB2\x21"
-     "\x0E\x00"
-     "\xB3\x2D\x0E\x00\xB0\x41\x0E\x00\xB1\x41\x0E\x00\xB6\x29\x0E\x00\xB7\x1D"
-     "\x0E\x00"
-     "\xB4\x35\x0E\x00\xB5\x29\x0E\x00\xA8\x29\x02\x00\xA9\x29\x02\x00\xAA\x39"
-     "\x02\x00"
-     "\xAB\x39\x02\x00\xAC\x29\x02\x00\xAD\x29\x02\x00\xAE\x91\x02\x00\xAF\x91"
-     "\x02\x00"
-     "\xCA\xE2\x00\x80\xCE\xE2\x00\x80\xD2\xE2\x00\x80\xD6\xE2\x00\x80\xDA\xE2"
-     "\x00\x80"
-     "\xDE\xE2\x00\x80\xBF\x30\x2B\x00\xE2\xE2\x00\x80\xB8\x99\x02\x00\xB9\xAD"
-     "\x02\x00"
-     "\xBA\xA5\x02\x00\xBB\x49\x01\x00\xBC\x5D\x01\x00\xBD\x45\x01\x00\xBE\x41"
-     "\x01\x00"
-     "\xBF\x41\x01\x00\xB0\xF1\x02\x00\xB1\xF9\x02\x00\xB2\xC1\x02\x00\xB3\xC1"
-     "\x02\x00"
-     "\xB4\xB1\x02\x00\xB5\xBD\x02\x00\xB6\xB5\x02\x00\xB7\xA9\x02\x00\xE1\x08"
-     "\x0E\x00"
-     "\xE1\xF4\x0C\x00\xE3\xD4\x0E\x00\xE3\x90\x0F\x00\xE6\xE2\x00\x80\xEA\xE2"
-     "\x00\x80"
-     "\xEE\xE2\x00\x80\xF2\xE2\x00\x80\x81\x11\x00\x00\x80\x0D\x00\x00\xF6\xE2"
-     "\x00\x80"
-     "\x82\x19\x00\x00\xFA\xE2\x00\x80\xFE\xE2\x00\x80\xEF\xAC\x01\x00\xEF\xCC"
-     "\x0E\x00"
-     "\xA3\xF1\x02\x00\x06\xE3\x00\x80\x0A\xE3\x00\x80\x84\x80\x05\x00\x0E\xE3"
-     "\x00\x80"
-     "\xA6\xE9\x02\x00\xA5\xE1\x02\x00\x12\xE3\x00\x80\xAB\x7D\x02\x00\xAA\x75"
-     "\x02\x00"
-     "\x86\xC8\x04\x00\x87\xEC\x04\x00\xAF\x39\x02\x00\xAE\x35\x02\x00\xAD\x55"
-     "\x02\x00"
-     "\xAC\x55\x02\x00\xA8\x3D\x06\x00\xA9\x55\x06\x00\xAA\x55\x06\x00\xAB\x6D"
-     "\x06\x00"
-     "\xAC\x75\x06\x00\xAD\xB5\x06\x00\xAE\xBD\x06\x00\xAF\xB5\x06\x00\x02\xE3"
-     "\x00\x80"
-     "\x16\xE3\x00\x80\x1A\xE3\x00\x80\x1E\xE3\x00\x80\x22\xE3\x00\x80\x26\xE3"
-     "\x00\x80"
-     "\x2A\xE3\x00\x80\x2E\xE3\x00\x80\xB8\x6D\x07\x00\xB9\x75\x07\x00\xBA\x7D"
-     "\x07\x00"
-     "\xBB\x71\x07\x00\xBC\xC1\x07\x00\xBD\xC1\x07\x00\xBE\xC1\x07\x00\xBF\xC1"
-     "\x07\x00"
-     "\xB0\xCD\x06\x00\xB1\xD5\x06\x00\xB2\xDD\x06\x00\xB3\xD5\x06\x00\xB4\xCD"
-     "\x06\x00"
-     "\xB5\x55\x07\x00\xB6\x5D\x07\x00\xB7\x55\x07\x00\xB3\xDD\x06\x00\x32\xE3"
-     "\x00\x80"
-     "\x36\xE3\x00\x80\x3A\xE3\x00\x80\x3E\xE3\x00\x80\xB6\x35\x06\x00\xB5\xFD"
-     "\x06\x00"
-     "\x42\xE3\x00\x80\xBB\x11\x06\x00\xBA\x09\x06\x00\x46\xE3\x00\x80\x4A\xE3"
-     "\x00\x80"
-     "\xBF\x91\x07\x00\xBE\x09\x06\x00\xBD\x01\x06\x00\xBC\x09\x06\x00\x4E\xE3"
-     "\x00\x80"
-     "\xA3\x99\x06\x00\x52\xE3\x00\x80\x56\xE3\x00\x80\xA6\x71\x06\x00\x5A\xE3"
-     "\x00\x80"
-     "\x5E\xE3\x00\x80\xA5\xB9\x06\x00\xAA\x4D\x06\x00\xAB\x55\x06\x00\x62\xE3"
-     "\x00\x80"
-     "\x66\xE3\x00\x80\xAE\x4D\x06\x00\xAF\xD5\x07\x00\xAC\x4D\x06\x00\xAD\x45"
-     "\x06\x00"
-     "\x80\xED\x01\x00\x81\x09\x00\x00\x82\x19\x00\x00\x6A\xE3\x00\x80\x6E\xE3"
-     "\x00\x80"
-     "\x84\xD0\x01\x00\xBE\xC4\x01\x00\x72\xE3\x00\x80\x86\x80\x00\x00\x87\x20"
-     "\x01\x00"
-     "\x76\xE3\x00\x80\x7A\xE3\x00\x80\x7E\xE3\x00\x80\x82\xE3\x00\x80\x86\xE3"
-     "\x00\x80"
-     "\x8A\xE3\x00\x80\x8E\xE3\x00\x80\x92\xE3\x00\x80\x96\xE3\x00\x80\x9A\xE3"
-     "\x00\x80"
-     "\x9E\xE3\x00\x80\xE1\x0C\x05\x00\xA2\xE3\x00\x80\xE3\x80\x04\x00\xA6\xE3"
-     "\x00\x80"
-     "\xAA\xE3\x00\x80\xAE\xE3\x00\x80\xEF\xDC\x05\x00\xB2\xE3\x00\x80\xB6\xE3"
-     "\x00\x80"
-     "\xBA\xE3\x00\x80\xBE\xE3\x00\x80\xA8\x45\x06\x00\xA9\x4D\x06\x00\xAA\x59"
-     "\x06\x00"
-     "\xAB\x51\x06\x00\xAC\xB5\x01\x00\xAD\xBD\x01\x00\xAE\xB5\x01\x00\xAF\xAD"
-     "\x01\x00"
-     "\x84\xAC\x01\x00\xC2\xE3\x00\x80\xC6\xE3\x00\x80\xCA\xE3\x00\x80\xCE\xE3"
-     "\x00\x80"
-     "\xD2\xE3\x00\x80\xD6\xE3\x00\x80\xDA\xE3\x00\x80\xB8\x69\x01\x00\xB9\x69"
-     "\x01\x00"
-     "\xBA\x09\x01\x00\xBB\x05\x01\x00\xBC\x1D\x01\x00\xBD\x01\x01\x00\xBE\x01"
-     "\x01\x00"
-     "\xBF\x35\x01\x00\xB0\xD5\x01\x00\xB1\xDD\x01\x00\xB2\xD5\x01\x00\xB3\x69"
-     "\x01\x00"
-     "\xB4\x79\x01\x00\xB5\x79\x01\x00\xB6\x69\x01\x00\xB7\x61\x01\x00\x81\xBD"
-     "\x03\x00"
-     "\x80\xBD\x03\x00\xB3\xD9\x05\x00\x82\x19\x00\x00\xB5\x35\x02\x00\xDE\xE3"
-     "\x00\x80"
-     "\xBE\x44\x03\x00\xB6\x35\x02\x00\xE6\xE3\x00\x80\xEA\xE3\x00\x80\xBB\x11"
-     "\x02\x00"
-     "\xBA\x19\x02\x00\xBD\xB5\x02\x00\xBC\x09\x02\x00\xBF\xB5\x02\x00\xBE\xB5"
-     "\x02\x00"
-     "\xEE\xE3\x00\x80\xA6\x6D\x02\x00\xA5\x6D\x02\x00\x84\x00\x3C\x00\xA3\x81"
-     "\x05\x00"
-     "\xF2\xE3\x00\x80\x86\xF8\x3C\x00\x87\x7C\x03\x00\xAF\xED\x02\x00\xAE\xED"
-     "\x02\x00"
-     "\xAD\xED\x02\x00\xAC\x51\x02\x00\xAB\x49\x02\x00\xAA\x41\x02\x00\xF6\xE3"
-     "\x00\x80"
-     "\xFA\xE3\x00\x80\xFE\xE3\x00\x80\x02\xE4\x00\x80\x06\xE4\x00\x80\x0A\xE4"
-     "\x00\x80"
-     "\xE3\x20\x01\x00\x0E\xE4\x00\x80\xE1\xA0\x01\x00\x12\xE4\x00\x80\xEF\x40"
-     "\x02\x00"
-     "\x16\xE4\x00\x80\x1A\xE4\x00\x80\x1E\xE4\x00\x80\x22\xE4\x00\x80\x26\xE4"
-     "\x00\x80"
-     "\x2A\xE4\x00\x80\x2E\xE4\x00\x80\xB3\xF1\x03\x00\x32\xE4\x00\x80\x36\xE4"
-     "\x00\x80"
-     "\x3A\xE4\x00\x80\x92\xE2\x00\x80\xB6\xE9\x03\x00\xB5\xE1\x03\x00\x3E\xE4"
-     "\x00\x80"
-     "\xBB\xAD\x03\x00\xBA\xAD\x03\x00\x42\xE4\x00\x80\x46\xE4\x00\x80\xBF\x95"
-     "\x03\x00"
-     "\xBE\x95\x03\x00\xBD\xA1\x03\x00\xBC\xAD\x03\x00\xA8\x29\x02\x00\xA9\x29"
-     "\x02\x00"
-     "\xAA\x39\x02\x00\xAB\x39\x02\x00\xAC\x29\x02\x00\xAD\x29\x02\x00\xAE\x59"
-     "\x02\x00"
-     "\xAF\x59\x02\x00\x80\xCD\x01\x00\x81\x09\x00\x00\x82\x19\x00\x00\x4A\xE4"
-     "\x00\x80"
-     "\x4E\xE4\x00\x80\xBE\x7C\x3D\x00\x87\xB4\x3D\x00\x86\x1C\x3C\x00\xB8\x35"
-     "\x02\x00"
-     "\xB9\x3D\x02\x00\xBA\x35\x02\x00\xBB\xE9\x02\x00\xBC\xF9\x02\x00\xBD\xF9"
-     "\x02\x00"
-     "\xBE\xE9\x02\x00\xBF\xE9\x02\x00\xB0\x29\x02\x00\xB1\x29\x02\x00\xB2\x39"
-     "\x02\x00"
-     "\xB3\x35\x02\x00\xB4\x11\x02\x00\xB5\x11\x02\x00\xB6\x11\x02\x00\xB7\x0D"
-     "\x02\x00"
-     "\x56\xE4\x00\x80\x5A\xE4\x00\x80\x5E\xE4\x00\x80\x62\xE4\x00\x80\x66\xE4"
-     "\x00\x80"
-     "\x6A\xE4\x00\x80\x6E\xE4\x00\x80\xEF\x7C\x00\x00\x72\xE4\x00\x80\xE1\x54"
-     "\x06\x00"
-     "\x76\xE4\x00\x80\xE3\xA8\x00\x00\xAC\xCC\x32\x00\x7A\xE4\x00\x80\x7E\xE4"
-     "\x00\x80"
-     "\x82\xE4\x00\x80\x86\xE4\x00\x80\xA6\xF9\x02\x00\xA5\xF1\x02\x00\x8A\xE4"
-     "\x00\x80"
-     "\xA3\xE1\x02\x00\x8E\xE4\x00\x80\x92\xE4\x00\x80\x96\xE4\x00\x80\xAF\x85"
-     "\x02\x00"
-     "\xAE\x85\x02\x00\xAD\xB1\x02\x00\xAC\xBD\x02\x00\xAB\xBD\x02\x00\xAA\xBD"
-     "\x02\x00"
-     "\x9A\xE4\x00\x80\x9E\xE4\x00\x80\xA8\x21\x3E\x00\xA9\x21\x3E\x00\xAA\x49"
-     "\x3E\x00"
-     "\xAB\x59\x3E\x00\xAC\x69\x3E\x00\xAD\x69\x3E\x00\xAE\x99\x3E\x00\xAF\x99"
-     "\x3E\x00"
-     "\x52\xE4\x00\x80\xA2\xE4\x00\x80\xA6\xE4\x00\x80\xAA\xE4\x00\x80\xAE\xE4"
-     "\x00\x80"
-     "\xB2\xE4\x00\x80\xB6\xE4\x00\x80\xBA\xE4\x00\x80\xB8\xF5\x3E\x00\xB9\xFD"
-     "\x3E\x00"
-     "\xBA\xF5\x3E\x00\xBB\x89\x3E\x00\xBC\x99\x3E\x00\xBD\x99\x3E\x00\xBE\x85"
-     "\x3E\x00"
-     "\xBF\xB5\x3E\x00\xB0\xE9\x3E\x00\xB1\xE9\x3E\x00\xB2\xFD\x3E\x00\xB3\xF1"
-     "\x3E\x00"
-     "\xB4\xD5\x3E\x00\xB5\xDD\x3E\x00\xB6\xD5\x3E\x00\xB7\xCD\x3E\x00\xB3\xC1"
-     "\x3E\x00"
-     "\x82\x15\x00\x00\x81\xB5\x00\x00\x80\xB5\x00\x00\xBE\xE4\x00\x80\xB6\xE9"
-     "\x3E\x00"
-     "\xB5\xE1\x3E\x00\xBE\x10\x00\x00\xBB\x2D\x3E\x00\xBA\x2D\x3E\x00\x84\x64"
-     "\x03\x00"
-     "\xC2\xE4\x00\x80\xBF\x19\x3E\x00\xBE\x11\x3E\x00\xBD\x25\x3E\x00\xBC\x29"
-     "\x3E\x00"
-     "\xC6\xE4\x00\x80\xA3\x85\x3E\x00\x86\xA8\x00\x00\x87\x6C\x03\x00\xA6\xAD"
-     "\x3E\x00"
-     "\xCA\xE4\x00\x80\xCE\xE4\x00\x80\xA5\xA5\x3E\x00\xAA\x69\x3E\x00\xAB\x69"
-     "\x3E\x00"
-     "\xD2\xE4\x00\x80\xD6\xE4\x00\x80\xAE\x55\x3E\x00\xAF\x5D\x3E\x00\xAC\x6D"
-     "\x3E\x00"
-     "\xAD\x61\x3E\x00\xB3\x89\x3E\x00\xDA\xE4\x00\x80\xDE\xE4\x00\x80\xE2\xE4"
-     "\x00\x80"
-     "\xE6\xE4\x00\x80\xB6\x81\x3E\x00\xB5\x89\x3E\x00\xEA\xE4\x00\x80\xBB\x95"
-     "\x3E\x00"
-     "\xBA\x8D\x3E\x00\xEE\xE4\x00\x80\xF2\xE4\x00\x80\xBF\xFD\x3E\x00\xBE\xF5"
-     "\x3E\x00"
-     "\xBD\x85\x3E\x00\xBC\x85\x3E\x00\xF6\xE4\x00\x80\xFA\xE4\x00\x80\xFE\xE4"
-     "\x00\x80"
-     "\x02\xE5\x00\x80\x06\xE5\x00\x80\x0A\xE5\x00\x80\xEF\x94\x3C\x00\x0E\xE5"
-     "\x00\x80"
-     "\xE1\x20\x3D\x00\x12\xE5\x00\x80\xE3\x34\x3D\x00\x16\xE5\x00\x80\x1A\xE5"
-     "\x00\x80"
-     "\x1E\xE5\x00\x80\x22\xE5\x00\x80\x26\xE5\x00\x80\xA8\x01\x3E\x00\xA9\x01"
-     "\x3E\x00"
-     "\xAA\x01\x3E\x00\xAB\x01\x3E\x00\xAC\x01\x3E\x00\xAD\x01\x3E\x00\xAE\x01"
-     "\x3E\x00"
-     "\xAF\x3D\x3E\x00\x80\xDD\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x2A\xE5"
-     "\x00\x80"
-     "\x2E\xE5\x00\x80\x84\x3C\x01\x00\xBE\x74\x00\x00\x36\xE5\x00\x80\xB8\x1D"
-     "\x3E\x00"
-     "\xB9\x3D\x3E\x00\xBA\x35\x3E\x00\xBB\xE9\x01\x00\xBC\xF9\x01\x00\xBD\xF9"
-     "\x01\x00"
-     "\xBE\xE9\x01\x00\xBF\xE9\x01\x00\xB0\x49\x3E\x00\xB1\x51\x3E\x00\xB2\x51"
-     "\x3E\x00"
-     "\xB3\x51\x3E\x00\xB4\x31\x3E\x00\xB5\x39\x3E\x00\xB6\x2D\x3E\x00\xB7\x25"
-     "\x3E\x00"
-     "\xA3\x89\x3D\x00\x3A\xE5\x00\x80\x86\xA8\x04\x00\x87\x3C\x03\x00\x3E\xE5"
-     "\x00\x80"
-     "\xA6\x81\x3D\x00\xA5\x89\x3D\x00\x42\xE5\x00\x80\xAB\x95\x3D\x00\xAA\x8D"
-     "\x3D\x00"
-     "\x46\xE5\x00\x80\x4A\xE5\x00\x80\xAF\xFD\x3D\x00\xAE\xF5\x3D\x00\xAD\x85"
-     "\x3D\x00"
-     "\xAC\x85\x3D\x00\x4E\xE5\x00\x80\xB3\x9D\x02\x00\x84\x48\x03\x00\xBE\x40"
-     "\x04\x00"
-     "\xB6\x55\x03\x00\x52\xE5\x00\x80\x56\xE5\x00\x80\xB5\xB1\x02\x00\xBA\x71"
-     "\x03\x00"
-     "\xBB\x79\x03\x00\x5A\xE5\x00\x80\x5E\xE5\x00\x80\xBE\x31\x03\x00\xBF\x31"
-     "\x03\x00"
-     "\xBC\x51\x03\x00\xBD\x51\x03\x00\xB8\xD1\x03\x00\xB9\xD1\x03\x00\xBA\xD5"
-     "\x03\x00"
-     "\xBB\xE9\x03\x00\xBC\xFD\x03\x00\xBD\xE5\x03\x00\xBE\xED\x03\x00\xBF\xE5"
-     "\x03\x00"
-     "\xB0\xB1\x03\x00\xB1\xB9\x03\x00\xB2\x81\x03\x00\xB3\x81\x03\x00\xB4\xF1"
-     "\x03\x00"
-     "\xB5\xF1\x03\x00\xB6\xF1\x03\x00\xB7\xF1\x03\x00\xA8\x01\x02\x00\xA9\x11"
-     "\x02\x00"
-     "\xAA\x11\x02\x00\xAB\x25\x02\x00\xAC\x3D\x02\x00\xAD\xD1\x03\x00\xAE\xD1"
-     "\x03\x00"
-     "\xAF\xD1\x03\x00\x62\xE5\x00\x80\xE2\xE3\x00\x80\x66\xE5\x00\x80\x6A\xE5"
-     "\x00\x80"
-     "\x6E\xE5\x00\x80\x72\xE5\x00\x80\x76\xE5\x00\x80\x7A\xE5\x00\x80\x7E\xE5"
-     "\x00\x80"
-     "\x82\xE5\x00\x80\x86\xE5\x00\x80\x8A\xE5\x00\x80\xEF\x50\x00\x00\x8E\xE5"
-     "\x00\x80"
-     "\x92\xE5\x00\x80\x96\xE5\x00\x80\xE3\x84\x00\x00\x9A\xE5\x00\x80\xE1\xF8"
-     "\x01\x00"
-     "\x9E\xE5\x00\x80\x80\x2D\x00\x00\x81\x11\x00\x00\x82\x15\x00\x00\xA6\xE5"
-     "\x00\x80"
-     "\xA3\x99\x03\x00\xAA\xE5\x00\x80\xAE\xE5\x00\x80\xB2\xE5\x00\x80\xB6\xE5"
-     "\x00\x80"
-     "\xA6\x51\x02\x00\xA5\xB5\x03\x00\xBA\xE5\x00\x80\xAB\x7D\x02\x00\xAA\x75"
-     "\x02\x00"
-     "\xBE\xE5\x00\x80\xC2\xE5\x00\x80\xAF\x35\x02\x00\xAE\x35\x02\x00\xAD\x55"
-     "\x02\x00"
-     "\xAC\x55\x02\x00\x86\xA0\x04\x00\x87\xA8\x05\x00\xC6\xE5\x00\x80\xCA\xE5"
-     "\x00\x80"
-     "\xCE\xE5\x00\x80\xD2\xE5\x00\x80\xD6\xE5\x00\x80\xEF\xF8\x01\x00\xBE\x2C"
-     "\x04\x00"
-     "\xE1\x80\x3E\x00\xDA\xE5\x00\x80\xE3\x14\x01\x00\xDE\xE5\x00\x80\xE2\xE5"
-     "\x00\x80"
-     "\xE6\xE5\x00\x80\xEA\xE5\x00\x80\xEE\xE5\x00\x80\xF2\xE5\x00\x80\xF6\xE5"
-     "\x00\x80"
-     "\xB3\xDD\x01\x00\xFA\xE5\x00\x80\xB5\xFD\x01\x00\xB6\xF5\x01\x00\xFE\xE5"
-     "\x00\x80"
-     "\x02\xE6\x00\x80\x06\xE6\x00\x80\xBA\xB1\x01\x00\xBB\x85\x01\x00\xBC\x9D"
-     "\x01\x00"
-     "\xBD\x15\x01\x00\xBE\x11\x01\x00\xBF\x11\x01\x00\xA8\xBD\x05\x00\xA9\x01"
-     "\x06\x00"
-     "\xAA\x19\x06\x00\xAB\x11\x06\x00\xAC\x39\x06\x00\xAD\x39\x06\x00\xAE\x59"
-     "\x06\x00"
-     "\xAF\x49\x06\x00\xA2\xE5\x00\x80\x82\x11\x00\x00\x81\x19\x00\x00\x80\xC9"
-     "\x07\x00"
-     "\x0A\xE6\x00\x80\x0E\xE6\x00\x80\x84\x30\x02\x00\x12\xE6\x00\x80\xB8\xB9"
-     "\x06\x00"
-     "\xB9\xB9\x06\x00\xBA\x49\x07\x00\xBB\x49\x07\x00\xBC\x59\x07\x00\xBD\x59"
-     "\x07\x00"
-     "\xBE\x4D\x07\x00\xBF\x45\x07\x00\xB0\x39\x06\x00\xB1\x0D\x06\x00\xB2\x05"
-     "\x06\x00"
-     "\xB3\x8D\x06\x00\xB4\x99\x06\x00\xB5\x99\x06\x00\xB6\x89\x06\x00\xB7\x89"
-     "\x06\x00"
-     "\xA3\x91\x06\x00\x16\xE6\x00\x80\x86\x28\x00\x00\x87\xFC\x03\x00\x1A\xE6"
-     "\x00\x80"
-     "\xA6\xB9\x06\x00\xA5\xB1\x06\x00\x1E\xE6\x00\x80\xAB\xC9\x06\x00\xAA\xFD"
-     "\x06\x00"
-     "\x22\xE6\x00\x80\x26\xE6\x00\x80\xAF\x5D\x06\x00\xAE\x5D\x06\x00\xAD\x59"
-     "\x06\x00"
-     "\xAC\xD1\x06\x00\x2A\xE6\x00\x80\xB3\x8D\x07\x00\x2E\xE6\x00\x80\x32\xE6"
-     "\x00\x80"
-     "\xB6\x8D\x07\x00\x36\xE6\x00\x80\x3A\xE6\x00\x80\xB5\x8D\x07\x00\xBA\x69"
-     "\x07\x00"
-     "\xBB\x69\x07\x00\x3E\xE6\x00\x80\x42\xE6\x00\x80\xBE\x69\x07\x00\xBF\x5D"
-     "\x07\x00"
-     "\xBC\x69\x07\x00\xBD\x69\x07\x00\xA3\x49\x06\x00\x46\xE6\x00\x80\x4A\xE6"
-     "\x00\x80"
-     "\x4E\xE6\x00\x80\x52\xE6\x00\x80\xA6\x49\x06\x00\xA5\x49\x06\x00\x56\xE6"
-     "\x00\x80"
-     "\xAB\xAD\x06\x00\xAA\xAD\x06\x00\x5A\xE6\x00\x80\x5E\xE6\x00\x80\xAF\x99"
-     "\x06\x00"
-     "\xAE\xAD\x06\x00\xAD\xAD\x06\x00\xAC\xAD\x06\x00\x62\xE6\x00\x80\x66\xE6"
-     "\x00\x80"
-     "\x6A\xE6\x00\x80\x6E\xE6\x00\x80\x72\xE6\x00\x80\xEF\xE0\x05\x00\x76\xE6"
-     "\x00\x80"
-     "\x7A\xE6\x00\x80\x7E\xE6\x00\x80\xE3\x3C\x05\x00\x84\x24\x01\x00\xE1\x7C"
-     "\x05\x00"
-     "\x80\x29\x00\x00\x81\x29\x00\x00\x82\x19\x00\x00\xBE\x34\x01\x00\xA8\xBD"
-     "\x06\x00"
-     "\xA9\xC1\x06\x00\xAA\xC1\x06\x00\xAB\xC1\x06\x00\xAC\xC1\x06\x00\xAD\xC9"
-     "\x06\x00"
-     "\xAE\xF1\x06\x00\xAF\xF1\x06\x00\x86\x00\x00\x00\x87\x84\x00\x00\x86\xE6"
-     "\x00\x80"
-     "\x8A\xE6\x00\x80\x8E\xE6\x00\x80\x92\xE6\x00\x80\x96\xE6\x00\x80\x9A\xE6"
-     "\x00\x80"
-     "\xB8\x79\x06\x00\xB9\x79\x06\x00\xBA\x09\x06\x00\xBB\x09\x06\x00\xBC\x19"
-     "\x06\x00"
-     "\xBD\x05\x06\x00\xBE\x05\x06\x00\xBF\x3D\x06\x00\xB0\x69\x06\x00\xB1\x69"
-     "\x06\x00"
-     "\xB2\x79\x06\x00\xB3\x79\x06\x00\xB4\x69\x06\x00\xB5\x5D\x06\x00\xB6\x49"
-     "\x06\x00"
-     "\xB7\x49\x06\x00\xA8\x85\x02\x00\xA9\x95\x02\x00\xAA\x95\x02\x00\xAB\xA5"
-     "\x02\x00"
-     "\xAC\xBD\x02\x00\xAD\xD5\x02\x00\xAE\xD1\x02\x00\xAF\xD1\x02\x00\x9E\xE6"
-     "\x00\x80"
-     "\xA2\xE6\x00\x80\x84\x28\x03\x00\xB5\x89\x09\x00\xA6\xE6\x00\x80\xAA\xE6"
-     "\x00\x80"
-     "\xAE\xE6\x00\x80\xB2\xE6\x00\x80\xB8\x5D\x03\x00\xB9\x6D\x03\x00\xBA\x65"
-     "\x03\x00"
-     "\xBB\x0D\x03\x00\xBC\x15\x03\x00\xBD\x1D\x03\x00\xBE\x15\x03\x00\xBF\x0D"
-     "\x03\x00"
-     "\xB0\xB5\x02\x00\xB1\xBD\x02\x00\xB2\x81\x02\x00\xB3\x81\x02\x00\xB4\x75"
-     "\x03\x00"
-     "\xB5\x7D\x03\x00\xB6\x75\x03\x00\xB7\x6D\x03\x00\xB3\x1D\x02\x00\xB6\xE6"
-     "\x00\x80"
-     "\xBA\xE6\x00\x80\xBE\xE6\x00\x80\xBE\x1C\x0C\x00\xB6\x55\x02\x00\xB5\x35"
-     "\x02\x00"
-     "\xC2\xE6\x00\x80\xBB\x79\x02\x00\xBA\x71\x02\x00\xC6\xE6\x00\x80\xCA\xE6"
-     "\x00\x80"
-     "\xBF\xB9\x03\x00\xBE\xB1\x03\x00\xBD\x55\x02\x00\xBC\x55\x02\x00\xCE\xE6"
-     "\x00\x80"
-     "\xA3\x59\x02\x00\xD2\xE6\x00\x80\xD6\xE6\x00\x80\xA6\x11\x02\x00\xDA\xE6"
-     "\x00\x80"
-     "\xDE\xE6\x00\x80\xA5\x71\x02\x00\xAA\x35\x02\x00\xAB\x3D\x02\x00\x32\xE5"
-     "\x00\x80"
-     "\xE2\xE6\x00\x80\xAE\xF5\x03\x00\xAF\xFD\x03\x00\xAC\x11\x02\x00\xAD\x11"
-     "\x02\x00"
-     "\xA8\xA9\x02\x00\xA9\xA9\x02\x00\xAA\xB9\x02\x00\xAB\xB9\x02\x00\xAC\xA9"
-     "\x02\x00"
-     "\xAD\xA9\x02\x00\xAE\x39\x01\x00\xAF\x39\x01\x00\x80\x0D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x1D\x00\x00\xE6\xE6\x00\x80\xEA\xE6\x00\x80\xEE\xE6\x00\x80\x87\x68"
-     "\x0D\x00"
-     "\x86\x1C\x0D\x00\xB8\xE9\x01\x00\xB9\xE9\x01\x00\xBA\x89\x01\x00\xBB\x85"
-     "\x01\x00"
-     "\xBC\x9D\x01\x00\xBD\x81\x01\x00\xBE\x81\x01\x00\xBF\xB1\x01\x00\xB0\x51"
-     "\x01\x00"
-     "\xB1\x51\x01\x00\xB2\x51\x01\x00\xB3\x51\x01\x00\xB4\xF5\x01\x00\xB5\xFD"
-     "\x01\x00"
-     "\xB6\xF5\x01\x00\xB7\xE1\x01\x00\xEF\x04\x00\x00\x84\x48\x0C\x00\xF6\xE6"
-     "\x00\x80"
-     "\xBE\x80\x0C\x00\xFA\xE6\x00\x80\xFE\xE6\x00\x80\x02\xE7\x00\x80\xEF\xC4"
-     "\x06\x00"
-     "\x06\xE7\x00\x80\xE1\x3C\x06\x00\x0A\xE7\x00\x80\xE3\x84\x01\x00\xE3\xC8"
-     "\x00\x00"
-     "\x0E\xE7\x00\x80\xE1\x28\x01\x00\x12\xE7\x00\x80\x16\xE7\x00\x80\x1A\xE7"
-     "\x00\x80"
-     "\x1E\xE7\x00\x80\xB3\x7D\x02\x00\x22\xE7\x00\x80\xB5\x6D\x02\x00\xB6\x65"
-     "\x02\x00"
-     "\x26\xE7\x00\x80\x2A\xE7\x00\x80\x2E\xE7\x00\x80\xBA\xD9\x01\x00\xBB\xD9"
-     "\x01\x00"
-     "\xBC\xC9\x01\x00\xBD\xC9\x01\x00\xBE\x7D\x01\x00\xBF\x75\x01\x00\xA3\xB1"
-     "\x0D\x00"
-     "\xF2\xE6\x00\x80\x32\xE7\x00\x80\x36\xE7\x00\x80\x3A\xE7\x00\x80\xA6\xA9"
-     "\x0D\x00"
-     "\xA5\xA1\x0D\x00\x3E\xE7\x00\x80\xAB\x15\x0E\x00\xAA\x15\x0E\x00\x42\xE7"
-     "\x00\x80"
-     "\x46\xE7\x00\x80\xAF\xB9\x0E\x00\xAE\xB1\x0E\x00\xAD\x05\x0E\x00\xAC\x05"
-     "\x0E\x00"
-     "\x4A\xE7\x00\x80\x4E\xE7\x00\x80\x52\xE7\x00\x80\x56\xE7\x00\x80\x80\x1D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x39\x00\x00\x5A\xE7\x00\x80\x5E\xE7\x00\x80\x62\xE7"
-     "\x00\x80"
-     "\x86\xC8\x00\x00\x87\x0C\x03\x00\x66\xE7\x00\x80\x6A\xE7\x00\x80\x6E\xE7"
-     "\x00\x80"
-     "\x72\xE7\x00\x80\xA8\x09\x0F\x00\xA9\x15\x0F\x00\xAA\x1D\x0F\x00\xAB\x15"
-     "\x0F\x00"
-     "\xAC\x31\x0F\x00\xAD\x3D\x0F\x00\xAE\x2D\x0F\x00\xAF\x21\x0F\x00\x84\xA8"
-     "\x03\x00"
-     "\x76\xE7\x00\x80\x7A\xE7\x00\x80\x7E\xE7\x00\x80\x82\xE7\x00\x80\x86\xE7"
-     "\x00\x80"
-     "\x8A\xE7\x00\x80\x8E\xE7\x00\x80\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5"
-     "\x00\x00"
-     "\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x9D\x00\x00\xBF\x95"
-     "\x00\x00"
-     "\xB0\x21\x0F\x00\xB1\x21\x0F\x00\xB2\x21\x0F\x00\xB3\x3D\x0F\x00\xB4\x2D"
-     "\x0F\x00"
-     "\xB5\x15\x0F\x00\xB6\x1D\x0F\x00\xB7\xF5\x00\x00\xB3\x95\x0E\x00\x92\xE7"
-     "\x00\x80"
-     "\x96\xE7\x00\x80\x9A\xE7\x00\x80\x9E\xE7\x00\x80\xB6\xBD\x0E\x00\xB5\xB9"
-     "\x0E\x00"
-     "\xA2\xE7\x00\x80\xBB\xE5\x0E\x00\xBA\xE5\x0E\x00\xA6\xE7\x00\x80\xAA\xE7"
-     "\x00\x80"
-     "\xBF\xD1\x0E\x00\xBE\xE9\x0E\x00\xBD\xE9\x0E\x00\xBC\xF5\x0E\x00\xEF\x28"
-     "\x0D\x00"
-     "\xAE\xE7\x00\x80\xB2\xE7\x00\x80\xB6\xE7\x00\x80\xBA\xE7\x00\x80\xBE\xE7"
-     "\x00\x80"
-     "\xC2\xE7\x00\x80\xC6\xE7\x00\x80\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\xCA\xE7\x00\x80\xE3\x48\x0D\x00\xCE\xE7\x00\x80\xE1\xC0\x0E\x00\xBE\x7C"
-     "\x01\x00"
-     "\xA3\x15\x0E\x00\xD6\xE7\x00\x80\x86\x28\x00\x00\x87\x4C\x01\x00\xDA\xE7"
-     "\x00\x80"
-     "\xA6\x3D\x0E\x00\xA5\x39\x0E\x00\xDE\xE7\x00\x80\xAB\x65\x0E\x00\xAA\x65"
-     "\x0E\x00"
-     "\xE2\xE7\x00\x80\xE6\xE7\x00\x80\xAF\x51\x0E\x00\xAE\x69\x0E\x00\xAD\x69"
-     "\x0E\x00"
-     "\xAC\x75\x0E\x00\xEA\xE7\x00\x80\xB3\xBD\x01\x00\xEE\xE7\x00\x80\xF2\xE7"
-     "\x00\x80"
-     "\xB6\x79\x01\x00\xF6\xE7\x00\x80\xFA\xE7\x00\x80\xB5\x79\x01\x00\xBA\x55"
-     "\x01\x00"
-     "\xBB\x5D\x01\x00\xFE\xE7\x00\x80\x02\xE8\x00\x80\xBE\xF9\x00\x00\xBF\xF9"
-     "\x00\x00"
-     "\xBC\x45\x01\x00\xBD\xF9\x00\x00\xA8\x71\x02\x00\xA9\x71\x02\x00\xAA\x71"
-     "\x02\x00"
-     "\xAB\x71\x02\x00\xAC\xB5\x02\x00\xAD\xBD\x02\x00\xAE\xB5\x02\x00\xAF\xAD"
-     "\x02\x00"
-     "\x84\xEC\x04\x00\x06\xE8\x00\x80\x0A\xE8\x00\x80\x0E\xE8\x00\x80\x12\xE8"
-     "\x00\x80"
-     "\x16\xE8\x00\x80\x1A\xE8\x00\x80\x1E\xE8\x00\x80\xB8\x69\x03\x00\xB9\x69"
-     "\x03\x00"
-     "\xBA\x0D\x03\x00\xBB\x05\x03\x00\xBC\x1D\x03\x00\xBD\x05\x03\x00\xBE\x0D"
-     "\x03\x00"
-     "\xBF\x05\x03\x00\xB0\xD5\x02\x00\xB1\xDD\x02\x00\xB2\xD5\x02\x00\xB3\x69"
-     "\x03\x00"
-     "\xB4\x79\x03\x00\xB5\x79\x03\x00\xB6\x69\x03\x00\xB7\x61\x03\x00\x22\xE8"
-     "\x00\x80"
-     "\x26\xE8\x00\x80\x2A\xE8\x00\x80\xA3\xF5\x02\x00\x2E\xE8\x00\x80\xA5\x31"
-     "\x02\x00"
-     "\xA6\x31\x02\x00\x32\xE8\x00\x80\x36\xE8\x00\x80\x3A\xE8\x00\x80\xAA\x1D"
-     "\x02\x00"
-     "\xAB\x15\x02\x00\xAC\x0D\x02\x00\xAD\xB1\x03\x00\xAE\xB1\x03\x00\xAF\xB1"
-     "\x03\x00"
-     "\x80\x61\x00\x00\x81\x61\x00\x00\x82\x05\x00\x00\x3E\xE8\x00\x80\x86\xF0"
-     "\x04\x00"
-     "\x87\x60\x03\x00\xBE\x10\x04\x00\x46\xE8\x00\x80\x4A\xE8\x00\x80\x4E\xE8"
-     "\x00\x80"
-     "\x52\xE8\x00\x80\x56\xE8\x00\x80\x5A\xE8\x00\x80\x5E\xE8\x00\x80\x62\xE8"
-     "\x00\x80"
-     "\x82\xE6\x00\x80\xA8\x85\x02\x00\xA9\x95\x02\x00\xAA\x95\x02\x00\xAB\xA5"
-     "\x02\x00"
-     "\xAC\xBD\x02\x00\xAD\xD5\x02\x00\xAE\xD1\x02\x00\xAF\xD1\x02\x00\x66\xE8"
-     "\x00\x80"
-     "\x6A\xE8\x00\x80\x6E\xE8\x00\x80\x72\xE8\x00\x80\x76\xE8\x00\x80\x7A\xE8"
-     "\x00\x80"
-     "\x7E\xE8\x00\x80\x82\xE8\x00\x80\xB8\x75\x01\x00\xB9\x7D\x01\x00\xBA\x75"
-     "\x01\x00"
-     "\xBB\xCD\x01\x00\xBC\xD5\x01\x00\xBD\xDD\x01\x00\xBE\xC9\x01\x00\xBF\xC1"
-     "\x01\x00"
-     "\xB0\xB5\x02\x00\xB1\xBD\x02\x00\xB2\x81\x02\x00\xB3\x81\x02\x00\xB4\x55"
-     "\x01\x00"
-     "\xB5\x5D\x01\x00\xB6\x55\x01\x00\xB7\x4D\x01\x00\xE1\xA4\x0E\x00\x84\x08"
-     "\x04\x00"
-     "\xE3\x1C\x0E\x00\x86\xE8\x00\x80\xBE\x9C\x04\x00\x8A\xE8\x00\x80\x8E\xE8"
-     "\x00\x80"
-     "\x92\xE8\x00\x80\x96\xE8\x00\x80\x9A\xE8\x00\x80\x9E\xE8\x00\x80\xA2\xE8"
-     "\x00\x80"
-     "\x81\x75\x00\x00\x80\x75\x00\x00\xEF\xCC\x0E\x00\x82\x75\x00\x00\xA6\xE8"
-     "\x00\x80"
-     "\xAA\xE8\x00\x80\xAE\xE8\x00\x80\x84\xC4\x04\x00\xB6\xE8\x00\x80\xBA\xE8"
-     "\x00\x80"
-     "\xEF\x5C\x00\x00\xBE\xE8\x00\x80\xE1\x90\x01\x00\xC2\xE8\x00\x80\xE3\x58"
-     "\x01\x00"
-     "\xC6\xE8\x00\x80\xCA\xE8\x00\x80\xCE\xE8\x00\x80\x86\x58\x07\x00\x87\xEC"
-     "\x04\x00"
-     "\xB3\x9D\x05\x00\x42\xE8\x00\x80\xB2\xE8\x00\x80\xD2\xE8\x00\x80\xD6\xE8"
-     "\x00\x80"
-     "\xB6\xD5\x05\x00\xB5\xB5\x05\x00\xDA\xE8\x00\x80\xBB\xF9\x05\x00\xBA\xF1"
-     "\x05\x00"
-     "\xDE\xE8\x00\x80\xE2\xE8\x00\x80\xBF\x19\x06\x00\xBE\x11\x06\x00\xBD\xD5"
-     "\x05\x00"
-     "\xBC\xD5\x05\x00\xE6\xE8\x00\x80\xA3\xD9\x05\x00\xEA\xE8\x00\x80\xEE\xE8"
-     "\x00\x80"
-     "\xA6\x91\x05\x00\xF2\xE8\x00\x80\xF6\xE8\x00\x80\xA5\xF1\x05\x00\xAA\xB5"
-     "\x05\x00"
-     "\xAB\xBD\x05\x00\xFA\xE8\x00\x80\xFE\xE8\x00\x80\xAE\x55\x06\x00\xAF\x5D"
-     "\x06\x00"
-     "\xAC\x91\x05\x00\xAD\x91\x05\x00\xA8\x5D\x06\x00\xA9\x65\x06\x00\xAA\x6D"
-     "\x06\x00"
-     "\xAB\x65\x06\x00\xAC\x7D\x06\x00\xAD\x65\x06\x00\xAE\x6D\x06\x00\xAF\x65"
-     "\x06\x00"
-     "\xBE\xA0\x00\x00\x02\xE9\x00\x80\x06\xE9\x00\x80\x0A\xE9\x00\x80\x0E\xE9"
-     "\x00\x80"
-     "\x12\xE9\x00\x80\x16\xE9\x00\x80\x1A\xE9\x00\x80\xB8\x4D\x07\x00\xB9\x51"
-     "\x07\x00"
-     "\xBA\x51\x07\x00\xBB\x51\x07\x00\xBC\x71\x07\x00\xBD\x71\x07\x00\xBE\x71"
-     "\x07\x00"
-     "\xBF\x71\x07\x00\xB0\xC1\x07\x00\xB1\xC1\x07\x00\xB2\xC1\x07\x00\xB3\xC1"
-     "\x07\x00"
-     "\xB4\xC1\x07\x00\xB5\xC1\x07\x00\xB6\xC1\x07\x00\xB7\xC1\x07\x00\x1E\xE9"
-     "\x00\x80"
-     "\xE1\x94\x06\x00\x26\xE9\x00\x80\xE3\x3C\x05\x00\x80\xA9\x01\x00\x81\xA9"
-     "\x01\x00"
-     "\x82\x19\x00\x00\x2A\xE9\x00\x80\x86\x80\x00\x00\x87\x04\x01\x00\x2E\xE9"
-     "\x00\x80"
-     "\x32\xE9\x00\x80\x36\xE9\x00\x80\x3A\xE9\x00\x80\x3E\xE9\x00\x80\xEF\x1C"
-     "\x05\x00"
-     "\xB3\xA5\x07\x00\x22\xE9\x00\x80\x42\xE9\x00\x80\x46\xE9\x00\x80\x4A\xE9"
-     "\x00\x80"
-     "\xB6\xA5\x07\x00\xB5\xB5\x07\x00\x4E\xE9\x00\x80\xBB\x45\x06\x00\xBA\x45"
-     "\x06\x00"
-     "\x52\xE9\x00\x80\x56\xE9\x00\x80\xBF\x45\x06\x00\xBE\x45\x06\x00\xBD\x55"
-     "\x06\x00"
-     "\xBC\x55\x06\x00\xA3\x65\x07\x00\x5A\xE9\x00\x80\x5E\xE9\x00\x80\x62\xE9"
-     "\x00\x80"
-     "\x66\xE9\x00\x80\xA6\x65\x07\x00\xA5\x75\x07\x00\x6A\xE9\x00\x80\xAB\x85"
-     "\x06\x00"
-     "\xAA\x85\x06\x00\x6E\xE9\x00\x80\x72\xE9\x00\x80\xAF\x85\x06\x00\xAE\x85"
-     "\x06\x00"
-     "\xAD\x95\x06\x00\xAC\x95\x06\x00\x76\xE9\x00\x80\xB3\x19\x06\x00\x7A\xE9"
-     "\x00\x80"
-     "\x7E\xE9\x00\x80\xB6\x21\x06\x00\x82\xE9\x00\x80\x86\xE9\x00\x80\xB5\x31"
-     "\x06\x00"
-     "\xBA\xF1\x01\x00\xBB\xF9\x01\x00\x8A\xE9\x00\x80\x8E\xE9\x00\x80\xBE\xB1"
-     "\x01\x00"
-     "\xBF\xBD\x01\x00\xBC\xD1\x01\x00\xBD\xD1\x01\x00\xA8\xD1\x05\x00\xA9\xD1"
-     "\x05\x00"
-     "\xAA\xD5\x05\x00\xAB\xE9\x05\x00\xAC\xFD\x05\x00\xAD\xE5\x05\x00\xAE\xED"
-     "\x05\x00"
-     "\xAF\x11\x02\x00\x80\xF9\x03\x00\x81\xCD\x03\x00\x82\xC5\x03\x00\x84\x20"
-     "\x03\x00"
-     "\xBE\x00\x1C\x00\x96\xE9\x00\x80\x87\x04\x03\x00\x86\xBC\x1C\x00\xB8\x19"
-     "\x02\x00"
-     "\xB9\x2D\x02\x00\xBA\x25\x02\x00\xBB\xE9\x02\x00\xBC\xF9\x02\x00\xBD\xF9"
-     "\x02\x00"
-     "\xBE\xE9\x02\x00\xBF\xE9\x02\x00\xB0\x71\x02\x00\xB1\x79\x02\x00\xB2\x41"
-     "\x02\x00"
-     "\xB3\x41\x02\x00\xB4\x31\x02\x00\xB5\x3D\x02\x00\xB6\x35\x02\x00\xB7\x29"
-     "\x02\x00"
-     "\xD2\xE7\x00\x80\x9A\xE9\x00\x80\x9E\xE9\x00\x80\xA3\x51\x05\x00\xA2\xE9"
-     "\x00\x80"
-     "\xA5\x79\x05\x00\xA6\x69\x05\x00\xA6\xE9\x00\x80\xAA\xE9\x00\x80\xAE\xE9"
-     "\x00\x80"
-     "\xAA\xB9\x02\x00\xAB\xB1\x02\x00\xAC\x99\x02\x00\xAD\x99\x02\x00\xAE\xF9"
-     "\x02\x00"
-     "\xAF\xF5\x02\x00\xE1\xF8\x01\x00\xB2\xE9\x00\x80\xE3\xF4\x02\x00\x84\xC0"
-     "\x1C\x00"
-     "\xB6\xE9\x00\x80\xBA\xE9\x00\x80\xBE\xE9\x00\x80\xC2\xE9\x00\x80\xC6\xE9"
-     "\x00\x80"
-     "\xCA\xE9\x00\x80\xCE\xE9\x00\x80\xD2\xE9\x00\x80\xD6\xE9\x00\x80\xDA\xE9"
-     "\x00\x80"
-     "\xEF\x30\x02\x00\xDE\xE9\x00\x80\xA8\x09\x02\x00\xA9\x19\x02\x00\xAA\x21"
-     "\x02\x00"
-     "\xAB\x21\x02\x00\xAC\x21\x02\x00\xAD\x21\x02\x00\xAE\x21\x02\x00\xAF\x55"
-     "\x02\x00"
-     "\xE2\xE9\x00\x80\xE6\xE9\x00\x80\xEA\xE9\x00\x80\xEE\xE9\x00\x80\xF2\xE9"
-     "\x00\x80"
-     "\x82\xAD\x00\x00\x81\xAD\x00\x00\x80\xAD\x00\x00\xB8\xD1\x02\x00\xB9\xD9"
-     "\x02\x00"
-     "\xBA\xE1\x02\x00\xBB\xE1\x02\x00\xBC\x91\x02\x00\xBD\x9D\x02\x00\xBE\x95"
-     "\x02\x00"
-     "\xBF\x89\x02\x00\xB0\x2D\x02\x00\xB1\x35\x02\x00\xB2\x35\x02\x00\xB3\x05"
-     "\x02\x00"
-     "\xB4\x1D\x02\x00\xB5\xF1\x02\x00\xB6\xF1\x02\x00\xB7\xF1\x02\x00\xE1\x48"
-     "\x07\x00"
-     "\xE3\x58\x06\x00\xE3\x1C\x06\x00\xE1\x6C\x06\x00\xF6\xE9\x00\x80\xFA\xE9"
-     "\x00\x80"
-     "\xFE\xE9\x00\x80\xBE\x10\x1C\x00\x68\x1B\x00\x80\x6C\x1B\x00\x80\xB5\x05"
-     "\x00\x00"
-     "\x70\x1B\x00\x80\xE1\x30\x3B\x00\xEF\xEC\x01\x00\xEF\xD0\x06\x00\xE1\xBC"
-     "\x32\x00"
-     "\xB3\xE9\x02\x00\xAC\xDC\x02\x00\x74\x1B\x00\x80\xB5\x05\x00\x00\xAC\x0C"
-     "\x1D\x00"
-     "\xB6\xE1\x02\x00\xB5\xE9\x02\x00\x78\x1B\x00\x80\xBB\xB1\x02\x00\xBA\xA9"
-     "\x02\x00"
-     "\xAC\x28\x1D\x00\x7C\x1B\x00\x80\xBF\x75\x02\x00\xBE\x99\x02\x00\xBD\x91"
-     "\x02\x00"
-     "\xBC\xA1\x02\x00\x92\xE9\x00\x80\xE1\x70\x1D\x00\xB5\xD5\x1F\x00\xE1\x6C"
-     "\x1D\x00"
-     "\xE1\x48\x1D\x00\xB5\xF9\x1F\x00\xE1\xB4\x02\x00\xAC\x58\x02\x00\xA8\x15"
-     "\x1E\x00"
-     "\xA9\xA1\x1E\x00\xAA\xA1\x1E\x00\xAB\xBD\x1E\x00\xAC\xAD\x1E\x00\xAD\x95"
-     "\x1E\x00"
-     "\xAE\x9D\x1E\x00\xAF\xF9\x1E\x00\xB0\x91\x1E\x00\xB1\x91\x1E\x00\xB2\x91"
-     "\x1E\x00"
-     "\xB3\x91\x1E\x00\xB4\xB1\x1E\x00\xB5\xB1\x1E\x00\xB6\xB1\x1E\x00\xB7\xB1"
-     "\x1E\x00"
-     "\xB8\x91\x1E\x00\xB9\x91\x1E\x00\xBA\x95\x1E\x00\xBB\x4D\x1F\x00\xBC\x55"
-     "\x1F\x00"
-     "\xBD\x5D\x1F\x00\xBE\x55\x1F\x00\xBF\x4D\x1F\x00\xAC\xDD\x1D\x00\xAD\xED"
-     "\x1D\x00"
-     "\xAE\xE5\x1D\x00\xAF\x09\x1D\x00\xB5\x05\x00\x00\x80\x1B\x00\x80\xAA\xD5"
-     "\x1D\x00"
-     "\xAB\xCD\x1D\x00\xE1\xBC\x02\x00\xA5\x95\x1D\x00\xA6\x9D\x1D\x00\xAC\xC8"
-     "\x02\x00"
-     "\xB5\x05\x00\x00\x84\x1B\x00\x80\xE1\xCC\x02\x00\xA3\x95\x1D\x00\xB3\x49"
-     "\x1E\x00"
-     "\xAC\xDC\x02\x00\x88\x1B\x00\x80\xB5\x05\x00\x00\xE1\xBC\x02\x00\xB6\x75"
-     "\x1E\x00"
-     "\xB5\x45\x1E\x00\xAC\xE0\x02\x00\xBB\x35\x1E\x00\xBA\x35\x1E\x00\xB5\x05"
-     "\x00\x00"
-     "\x8C\x1B\x00\x80\xBF\x15\x1E\x00\xBE\x15\x1E\x00\xBD\x25\x1E\x00\xBC\x25"
-     "\x1E\x00"
-     "\xA3\x89\x1F\x00\x8F\x35\x25\x00\x82\x80\x03\x00\xE1\xAC\x02\x00\xAC\xD8"
-     "\x02\x00"
-     "\xA6\xB5\x1F\x00\xA5\x85\x1F\x00\xB5\x35\x00\x00\xAB\xF5\x1F\x00\xAA\xF5"
-     "\x1F\x00"
-     "\x90\x1B\x00\x80\x9D\x0B\x00\x80\xAF\xD5\x1F\x00\xAE\xD5\x1F\x00\xAD\xE5"
-     "\x1F\x00"
-     "\xAC\xE5\x1F\x00\x9E\x25\x19\x00\x9F\xC5\x1F\x00\x9C\xB5\x1B\x00\x9D\x49"
-     "\x19\x00"
-     "\x9A\x71\x15\x00\x9B\x69\x15\x00\x98\xF5\x11\x00\x99\xB5\x17\x00\x96\x85"
-     "\x13\x00"
-     "\x97\x45\x11\x00\x94\x49\x2D\x00\x95\x71\x2C\x00\x92\xF5\x28\x00\x93\xA5"
-     "\x2F\x00"
-     "\x90\x79\x29\x00\x91\x79\x29\x00\xAC\x0D\x00\x00\xAD\x15\x00\x00\xAE\x11"
-     "\x00\x00"
-     "\xBE\x06\x00\x80\xB1\x09\x00\x80\x96\x0C\x00\x80\xC9\x06\x00\x80\xDE\x06"
-     "\x00\x80"
-     "\xF1\x06\x00\x80\xC6\x0C\x00\x80\xF9\x06\x00\x80\xC9\x0C\x00\x80\xA2\xF5"
-     "\x1C\x00"
-     "\xA3\xB5\x01\x00\xA0\xBD\x1D\x00\xA1\xB5\x1D\x00\xBC\xE9\x05\x00\xBD\x51"
-     "\x00\x00"
-     "\xBE\xF9\x05\x00\xBF\xF5\x05\x00\xB8\xC1\x05\x00\xB9\xCD\x05\x00\xBA\xF9"
-     "\x05\x00"
-     "\x09\x07\x00\x80\xB4\x3D\x00\x00\xB5\x01\x04\x00\xB6\xE9\x05\x00\xB7\xE5"
-     "\x05\x00"
-     "\xB0\x6D\x00\x00\xB1\x55\x00\x00\xB2\xD5\x00\x00\xB3\x55\x00\x00\x82\x31"
-     "\x26\x00"
-     "\x83\x05\x2A\x00\x94\x1B\x00\x80\xCC\x80\x02\x00\x86\xC1\x2C\x00\x87\x35"
-     "\x2E\x00"
-     "\x84\x1D\x2A\x00\x85\x99\x2B\x00\x8A\x3D\x2F\x00\x8B\xC1\x10\x00\xE1\x0C"
-     "\x00\x80"
-     "\xE4\x0C\x00\x80\x8E\xD9\x14\x00\x8F\x15\x16\x00\x8C\x19\x12\x00\x8D\x91"
-     "\x12\x00"
-     "\x92\x7D\x16\x00\x93\xB9\x18\x00\xE7\x0C\x00\x80\x0A\x0B\x00\x80\x96\xD1"
-     "\x1C\x00"
-     "\x97\x51\x1E\x00\x94\x5D\x1A\x00\x95\x39\x1A\x00\x9A\x69\x1E\x00\x9B\x21"
-     "\x1F\x00"
-     "\xD2\x0C\x00\x80\xB8\x15\x05\x00\x98\x1B\x00\x80\xCC\x14\x02\x00\x9C\x79"
-     "\x02\x00"
-     "\x9C\x1B\x00\x80\xB8\xFD\x05\x00\xA4\x1B\x00\x80\xCC\xA8\x02\x00\x1D\x0D"
-     "\x00\x80"
-     "\xCC\x84\x34\x02\xCD\x8C\x33\x02\xCC\xFC\x34\x02\xCD\x98\x33\x02\xCC\xA0"
-     "\x25\x02"
-     "\xCD\xAC\x23\x02\xCC\xBC\x32\x02\xCC\xB0\x34\x02\xA8\x1B\x00\x80\x46\x0E"
-     "\x00\x80"
-     "\xCD\xBC\x32\x02\xCC\xCC\x30\x02\xB8\x41\x05\x00\x49\x0E\x00\x80\xA2\x0E"
-     "\x00\x80"
-     "\xFF\x0E\x00\x80\xCC\x0C\x00\x80\xD5\x0C\x00\x80\x20\x0A\x00\x80\xCC\x18"
-     "\x03\x00"
-     "\xED\x0C\x00\x80\xF3\x0C\x00\x80\xFF\x0C\x00\x80\xCC\x94\x34\x02\x08\x0D"
-     "\x00\x80"
-     "\x0B\x0D\x00\x80\x0E\x0D\x00\x80\xAC\x1B\x00\x80\x85\xB4\xCE\x03\x89\x0D"
-     "\x00\x80"
-     "\x87\x3C\x1F\x00\x17\x0D\x00\x80\xB0\x1B\x00\x80\x8C\x0D\x00\x80\x23\x0D"
-     "\x00\x80"
-     "\x20\x0D\x00\x80\x8D\xA8\x00\x00\xB8\xA1\x04\x00\x26\x0D\x00\x80\x29\x0D"
-     "\x00\x80"
-     "\x89\x08\x1F\x00\x2C\x0D\x00\x80\x8B\xB8\x00\x00\xCC\x88\x02\x00\x2F\x0D"
-     "\x00\x80"
-     "\x38\x0D\x00\x80\x50\x0D\x00\x80\x53\x0D\x00\x80\x91\xC4\x00\x00\xB8\xDD"
-     "\x04\x00"
-     "\xA9\x0B\x00\x80\x5C\x0D\x00\x80\x5F\x0D\x00\x80\xA0\x1B\x00\x80\x62\x0D"
-     "\x00\x80"
-     "\x65\x0D\x00\x80\x71\x0D\x00\x80\x74\x0D\x00\x80\x9A\xE9\x00\x00\x9B\xF9"
-     "\x22\x02"
-     "\xBB\x5D\x2C\x02\xE1\xA0\x1F\x00\xB9\x79\x2D\x02\xE3\x50\x1F\x00\xF8\x07"
-     "\x00\x80"
-     "\xCC\xE8\x02\x00\xBD\x2D\x00\x00\xB8\x11\x04\x00\xB3\x6D\x2F\x02\x31\x0E"
-     "\x00\x80"
-     "\xB1\x0D\x00\x00\x37\x0E\x00\x80\xB7\x91\x2D\x02\xB7\x0E\x00\x80\xB5\x15"
-     "\x00\x00"
-     "\xEF\xD8\x1E\x00\x38\x00\x00\x80\x70\x00\x00\x80\xBB\x3D\x25\x02\x2A\x00"
-     "\x00\x80"
-     "\xBD\x11\x00\x00\x3C\x00\x00\x80\xBF\x19\x00\x00\xBE\x8D\x06\x00\x42\x00"
-     "\x00\x80"
-     "\xBC\xC0\x00\x00\xBF\xBC\x00\x00\xBE\xD0\x00\x00\x48\x00\x00\x80\x52\x00"
-     "\x00\x80"
-     "\x58\x00\x00\x80\xBA\x60\x00\x00\x2C\x07\x00\x80\xA2\x07\x00\x80\x2E\x07"
-     "\x00\x80"
-     "\x5E\x00\x00\x80\x39\x07\x00\x80\x1E\x00\x00\x80\x49\x07\x00\x80\xE7\x07"
-     "\x00\x80"
-     "\x69\x07\x00\x80\xB8\xE9\x06\x00\x74\x07\x00\x80\x4E\x00\x00\x80\x76\x07"
-     "\x00\x80"
-     "\xAC\x22\x00\x80\x8F\x07\x00\x80\x6C\x00\x00\x80\xA1\x89\x06\x00\xA0\xF5"
-     "\x15\x02"
-     "\xA3\x99\x21\x02\xA2\x7D\x01\x00\xA5\x75\x01\x00\xA4\x7D\x01\x00\xA7\xED"
-     "\x06\x00"
-     "\xA6\xF9\x15\x02\xA9\xD1\x15\x02\xA8\xD9\x15\x02\xAF\x07\x00\x80\xAA\xA5"
-     "\x06\x00"
-     "\xC5\x07\x00\x80\xA9\x03\x00\x80\xFC\x07\x00\x80\x21\x08\x00\x80\x91\xCD"
-     "\x06\x00"
-     "\x90\x95\x01\x00\x93\xE1\x06\x00\x92\x11\x20\x02\x95\xED\x06\x00\x94\xE5"
-     "\x06\x00"
-     "\x97\xB9\x01\x00\x96\x85\x01\x00\x99\x25\x06\x00\x98\x8D\x01\x00\x9B\x8D"
-     "\x01\x00"
-     "\x9A\x21\x06\x00\x9D\x15\x15\x02\x9C\x05\x06\x00\x9F\x7D\x06\x00\x9E\x09"
-     "\x06\x00"
-     "\x81\xA9\x00\x00\x80\xC1\x00\x00\x83\xA5\x20\x02\x82\xBD\x00\x00\x85\x7D"
-     "\x05\x00"
-     "\x84\x71\x05\x00\x87\xA1\x00\x00\x86\xD5\x06\x00\x89\x1D\x00\x00\x88\x99"
-     "\x00\x00"
-     "\x8B\xC5\x06\x00\x8A\xC5\x06\x00\x8D\xF1\x01\x00\x8C\xBD\x06\x00\x8F\xF1"
-     "\x01\x00"
-     "\x3E\x08\x00\x80\x5D\x08\x00\x80\xAB\x03\x00\x80\x83\x08\x00\x80\xAE\x03"
-     "\x00\x80"
-     "\xC9\x08\x00\x80\xB8\x22\x00\x80\x22\x09\x00\x80\xBF\x22\x00\x80\x5C\x09"
-     "\x00\x80"
-     "\x60\x09\x00\x80\x5E\x09\x00\x80\xD6\x22\x00\x80\x6C\x08\x00\x80\xDD\x22"
-     "\x00\x80"
-     "\x02\x08\x00\x80\xC8\x00\x00\x80\x47\x08\x00\x80\x77\x0D\x00\x80\x7B\x09"
-     "\x00\x80"
-     "\x0E\x00\x00\x80\x58\x0E\x00\x80\xB9\x06\x00\x80\xB6\x06\x00\x80\xCB\x00"
-     "\x00\x80"
-     "\xC9\x22\x00\x80\xB4\x03\x00\x80\xDC\x05\x00\x80\x17\x06\x00\x80\x3D\x06"
-     "\x00\x80"
-     "\x54\x06\x00\x80\x71\x06\x00\x80\x88\x06\x00\x80\xA3\x03\x00\x80\x2B\x23"
-     "\x00\x80"
-     "\xA9\x06\x00\x80\xDA\x00\x00\x80\x9F\x03\x00\x80\x06\x09\x00\x80\x91\x07"
-     "\x00\x80"
-     "\xF0\x07\x00\x80\xF4\x07\x00\x80\xCE\x06\x00\x80\x94\x09\x00\x80\x04\x08"
-     "\x00\x80"
-     "\x08\x08\x00\x80\xA0\x0B\x00\x80\x98\x09\x00\x80\xB9\x0A\x00\x80\x71\x07"
-     "\x00\x80"
-     "\x8D\x07\x00\x80\x99\x07\x00\x80\xA4\x07\x00\x80\xAC\x07\x00\x80\xDB\x07"
-     "\x00\x80"
-     "\x28\x0E\x00\x80\xAE\x09\x00\x80\xF2\x0F\x00\x80\x17\x07\x00\x80\xF3\x06"
-     "\x00\x80"
-     "\x09\x10\x00\x80\x16\x10\x00\x80\x1E\x10\x00\x80\xF5\x06\x00\x80\xE9\x0D"
-     "\x00\x80"
-     "\xCC\x4C\x3A\x02\xCD\x34\x3A\x02\xCC\x4C\x38\x02\xCD\xCC\x3F\x02\xCC\x0C"
-     "\x26\x02"
-     "\xCC\x18\x27\x02\xCC\x44\x26\x02\xCD\x24\x27\x02\xCC\x60\x3B\x02\xCD\xC0"
-     "\x3A\x02"
-     "\xCC\x24\x25\x02\xCD\xD0\x24\x02\xCC\x5C\x39\x02\xCD\xA0\x39\x02\xCC\x74"
-     "\x3B\x02"
-     "\xCD\xAC\x3A\x02\x10\x0E\x00\x80\x27\x23\x00\x80\xC9\x04\x00\x80\xD5\x04"
-     "\x00\x80"
-     "\xA4\x04\x00\x80\xB1\x04\x00\x80\x6E\x05\x00\x80\xBD\x04\x00\x80\xCC\x38"
-     "\x2A\x02"
-     "\xCD\xBC\x2A\x02\x74\x05\x00\x80\xC1\x05\x00\x80\xCC\x50\x39\x02\xCC\xF8"
-     "\x21\x02"
-     "\xCC\xE8\x20\x02\xCD\x10\x26\x02\x02\x06\x00\x80\xC3\x04\x00\x80\x22\x06"
-     "\x00\x80"
-     "\xCF\x04\x00\x80\xCD\xC4\x36\x02\xCC\xD4\x36\x02\xC7\x05\x00\x80\x28\x06"
-     "\x00\x80"
-     "\xCD\xA0\x3B\x02\xCC\xA4\x24\x02\x5C\x06\x00\x80\x9F\x04\x00\x80\x19\x07"
-     "\x00\x80"
-     "\x1C\x07\x00\x80\xCC\x78\x34\x02\x26\x07\x00\x80\xB3\x07\x00\x80\xB8\x51"
-     "\x01\x00"
-     "\xB6\x07\x00\x80\xBA\x07\x00\x80\xCD\x18\x30\x02\xCC\xD4\x31\x02\xCC\x30"
-     "\x36\x02"
-     "\x29\x07\x00\x80\xCD\x64\x3E\x02\xCC\x80\x3E\x02\xCC\x58\x2E\x02\xCD\x74"
-     "\x2F\x02"
-     "\xD2\x07\x00\x80\xD5\x07\x00\x80\x6E\x08\x00\x80\x71\x08\x00\x80\xCC\x40"
-     "\x32\x02"
-     "\xCD\xF8\x33\x02\xCC\x08\x3D\x02\xCD\x68\x32\x02\xCC\x10\x33\x02\xCD\xC4"
-     "\x30\x02"
-     "\xCC\x4C\x33\x02\xCD\x54\x32\x02\xCC\x70\x30\x02\xCD\x44\x30\x02\xCC\xD4"
-     "\x30\x02"
-     "\xCD\xA0\x30\x02\xCC\xB0\x36\x02\xCC\x78\x3D\x02\xCC\x7C\x30\x02\xCD\xC4"
-     "\x31\x02"
-     "\xCC\x44\x3F\x02\xCD\x84\x3C\x02\xCC\x2C\x3E\x02\xCC\x04\x39\x02\x7B\x08"
-     "\x00\x80"
-     "\x62\x09\x00\x80\xCC\x50\x3C\x02\xCD\x24\x3C\x02\x65\x09\x00\x80\x69\x09"
-     "\x00\x80"
-     "\x7E\x08\x00\x80\x72\x09\x00\x80\x75\x09\x00\x80\xBC\x03\x00\x80\xFB\x22"
-     "\x00\x80"
-     "\xFF\x22\x00\x80\x03\x23\x00\x80\x07\x23\x00\x80\x0F\x23\x00\x80\x0B\x23"
-     "\x00\x80"
-     "\xCC\xB4\x03\x00\x17\x23\x00\x80\xCC\x74\x0D\x00\x1F\x23\x00\x80\xCC\x54"
-     "\x0D\x00"
-     "\x23\x23\x00\x80\xCC\x54\x0D\x00\x2F\x23\x00\x80\xCC\x70\x0D\x00\x33\x23"
-     "\x00\x80"
-     "\x37\x23\x00\x80\x3E\x23\x00\x80\x55\x23\x00\x80\x5C\x23\x00\x80\xCD\x80"
-     "\x38\x02"
-     "\xCC\xC8\x38\x02\xCC\x18\x0D\x00\x66\x23\x00\x80\xCC\x78\x0D\x00\x6A\x23"
-     "\x00\x80"
-     "\xCC\x48\x0D\x00\x6E\x23\x00\x80\xCC\x58\x0D\x00\x72\x23\x00\x80\xCC\x48"
-     "\x0D\x00"
-     "\x7E\x23\x00\x80\xCC\x78\x0D\x00\x7A\x23\x00\x80\x80\xC5\x0B\x00\x81\xC9"
-     "\x0B\x00"
-     "\x94\xAC\x08\x00\xB8\x91\x0F\x00\x92\x1C\x08\x00\xB8\xB1\x0F\x00\x90\xF4"
-     "\x09\x00"
-     "\xB8\xB1\x0F\x00\xB8\x95\x0F\x00\x89\x81\x2F\x02\xB8\xFD\x0F\x00\x8B\xF9"
-     "\x2F\x02"
-     "\x89\x45\x0F\x00\x8D\xE1\x0B\x00\x8E\xD9\x0B\x00\x8F\xCD\x0B\x00\xB8\x9D"
-     "\x0F\x00"
-     "\x94\xFC\x09\x00\xB8\xAD\x0F\x00\x93\x79\x26\x02\xB8\xBD\x0F\x00\x90\x68"
-     "\x09\x00"
-     "\xB8\xAD\x0F\x00\x92\x18\x09\x00\xB8\x9D\x0F\x00\xCC\xEC\x02\x00\xB8\x19"
-     "\x00\x00"
-     "\x9B\x5D\x21\x02\x8A\x23\x00\x80\x9D\x81\x0B\x00\x9E\x81\x0B\x00\x9F\x71"
-     "\x0C\x00"
-     "\xB1\x07\x00\x80\xD0\x07\x00\x80\xCC\xF8\x02\x00\xA3\x91\x18\x02\xB8\x05"
-     "\x00\x00"
-     "\x8E\x23\x00\x80\xCC\xE4\x02\x00\xB8\x3D\x00\x00\x92\x23\x00\x80\xCC\xEC"
-     "\x02\x00"
-     "\xB8\x19\x00\x00\xAB\x9D\x22\x02\x96\x23\x00\x80\xAD\xA1\x0C\x00\xAE\xC5"
-     "\x2E\x02"
-     "\xAF\xA1\x0C\x00\xCC\xE4\x02\x00\xB8\x0D\x00\x00\x9A\x23\x00\x80\xCC\xFC"
-     "\x02\x00"
-     "\xB8\x05\x00\x00\x9E\x23\x00\x80\xCC\xE4\x02\x00\xB8\x3D\x00\x00\xA2\x23"
-     "\x00\x80"
-     "\xB9\x05\x21\x02\xCC\xF8\x02\x00\xBB\x71\x22\x02\xB8\x0D\x00\x00\xBD\x89"
-     "\x2C\x02"
-     "\xBE\xD5\x0C\x00\xA6\x23\x00\x80\x80\x00\x01\x00\x81\x00\x04\x00\x82\x08"
-     "\x07\x00"
-     "\xB8\x39\x00\x00\x84\x10\x06\x00\x85\x14\x05\x00\x86\x0C\x03\x00\x87\x44"
-     "\x03\x00"
-     "\x88\xA4\x04\x00\x89\x00\x78\x00\x8A\x20\x7B\x00\xCC\xC0\x02\x00\x8C\xC0"
-     "\x01\x00"
-     "\xCA\x23\x00\x80\xCC\xE4\x02\x00\xB8\x7D\x00\x00\xCE\x23\x00\x80\x91\xC4"
-     "\x7A\x00"
-     "\x92\x48\x79\x00\x93\x00\x7C\x00\xCC\xE4\x02\x00\xB8\x0D\x00\x00\xAA\x23"
-     "\x00\x80"
-     "\xCC\xDC\x02\x00\xB8\x05\x00\x00\xAE\x23\x00\x80\xB8\x19\x00\x00\xCC\xE4"
-     "\x02\x00"
-     "\xB2\x23\x00\x80\xCC\xEC\x02\x00\xB8\x05\x00\x00\xB6\x23\x00\x80\xB8\x09"
-     "\x00\x00"
-     "\xCC\xE4\x02\x00\xD2\x23\x00\x80\xCC\xFC\x02\x00\xB8\x05\x00\x00\xD6\x23"
-     "\x00\x80"
-     "\xB8\x31\x00\x00\xCC\xE4\x02\x00\xA8\xF0\x01\x00\xA9\xF0\x01\x00\xBA\x23"
-     "\x00\x80"
-     "\xAB\xB0\x01\x00\xB8\x09\x00\x00\xCC\xE4\x02\x00\xBE\x23\x00\x80\xCC\x9C"
-     "\x02\x00"
-     "\xB8\x09\x00\x00\xB1\x9C\x01\x00\xC2\x23\x00\x80\xB8\x15\x00\x00\xCC\xFC"
-     "\x02\x00"
-     "\xB5\x18\x01\x00\xC6\x23\x00\x80\xB8\x39\x00\x00\xCC\xDC\x02\x00\xDA\x23"
-     "\x00\x80"
-     "\xBA\x1C\x01\x00\xBB\x80\x01\x00\xBC\xF0\x7F\x00\xBD\xF4\x7E\x00\xBE\xF8"
-     "\x7D\x00"
-     "\xBF\x7C\x73\x00\x80\x3D\x2E\x02\x81\x39\x2E\x02\x82\x35\x2E\x02\x83\x31"
-     "\x2E\x02"
-     "\x84\x2D\x2E\x02\x85\x29\x2E\x02\x86\x25\x2E\x02\x87\x21\x2E\x02\x88\x1D"
-     "\x2E\x02"
-     "\x89\x19\x2E\x02\x8A\x15\x2E\x02\xCC\xFC\x02\x00\xB8\x05\x00\x00\xDE\x23"
-     "\x00\x80"
-     "\xCC\xE4\x02\x00\xB8\x7D\x00\x00\xE2\x23\x00\x80\x91\x71\x0E\x00\xCC\xE4"
-     "\x02\x00"
-     "\xB8\x1D\x00\x00\xE6\x23\x00\x80\xA9\x31\x00\x00\xAA\x31\x00\x00\x97\x61"
-     "\x0E\x00"
-     "\x8C\x4D\x04\x00\x14\x24\x00\x80\x18\x24\x00\x80\x9C\x51\x04\x00\xAF\x75"
-     "\x05\x00"
-     "\xCF\x03\x00\x80\xE8\x03\x00\x80\x82\xD8\x02\x00\xB5\xB5\x04\x00\xB4\xB5"
-     "\x04\x00"
-     "\x9F\x85\x05\x00\xB6\xB5\x04\x00\xA4\xE5\x0E\x00\xA5\x91\x0F\x00\xA6\x95"
-     "\x0F\x00"
-     "\x01\x04\x00\x80\xB3\xB5\x05\x00\x1A\x04\x00\x80\xBC\x61\x2F\x02\xBD\x2D"
-     "\x02\x00"
-     "\xBE\x70\x02\x00\x82\x3C\x02\x00\x9E\x41\x72\x00\xAF\x81\x2E\x02\x99\x39"
-     "\x72\x00"
-     "\x2D\x04\x00\x80\xBE\x50\x02\x00\xB3\xD9\x0F\x00\xB4\xCD\x0F\x00\x40\x04"
-     "\x00\x80"
-     "\xB6\xF5\x0F\x00\xB7\xE9\x0F\x00\x9E\x1D\x72\x00\x53\x04\x00\x80\x66\x04"
-     "\x00\x80"
-     "\x99\x11\x72\x00\xBC\xCD\x0F\x00\x99\xB5\x73\x00\xBE\xB5\x0F\x00\xBE\x78"
-     "\x03\x00"
-     "\xBE\x45\x0E\x00\xBF\xAD\x1A\x02\xBC\x79\x08\x00\xBD\x79\x08\x00\xBA\x7D"
-     "\x08\x00"
-     "\xBB\x5D\x0E\x00\xB8\x7D\x08\x00\xB9\x75\x08\x00\xB6\x7D\x08\x00\xB7\x45"
-     "\x08\x00"
-     "\xB4\x65\x08\x00\xB5\x75\x08\x00\xB2\x75\x08\x00\xB3\x7D\x08\x00\xB0\x0D"
-     "\x08\x00"
-     "\xB1\x7D\x08\x00\xAE\x69\x08\x00\xAF\x69\x08\x00\xAC\x51\x08\x00\xAD\x51"
-     "\x08\x00"
-     "\xAA\x41\x08\x00\xAB\x41\x08\x00\xA8\x5D\x08\x00\xA9\x51\x08\x00\xA6\x99"
-     "\x09\x00"
-     "\xA7\x99\x09\x00\xA4\x99\x09\x00\xA5\x99\x09\x00\xA2\x95\x09\x00\xA3\x9D"
-     "\x09\x00"
-     "\xA0\x39\x0E\x00\xA1\x9D\x09\x00\x9E\x59\x09\x00\x9F\xCD\x09\x00\x9C\xB1"
-     "\x0E\x00"
-     "\x9D\x59\x09\x00\x79\x04\x00\x80\x9B\x45\x09\x00\xE3\xEC\x03\x00\x8C\x04"
-     "\x00\x80"
-     "\xF2\x04\x00\x80\xDC\x00\x00\x80\xEF\xE0\x03\x00\x12\x01\x00\x80\xCC\x48"
-     "\x30\x02"
-     "\xCC\x64\x30\x02\x82\x28\x03\x00\x1B\x01\x00\x80\x1F\x01\x00\x80\x23\x01"
-     "\x00\x80"
-     "\x27\x01\x00\x80\x2B\x01\x00\x80\x2F\x01\x00\x80\x33\x01\x00\x80\xEF\xE8"
-     "\x03\x00"
-     "\xEF\xD8\x03\x00\xE0\x00\x00\x80\xE5\x00\x00\x80\xE3\x54\x03\x00\xE3\x5C"
-     "\x02\x00"
-     "\xB0\xF1\x5F\x00\xEA\x00\x00\x80\xEF\x00\x00\x80\x9E\xBD\x72\x00\x80\x65"
-     "\x09\x00"
-     "\x81\x8D\x28\x02\x82\x7D\x28\x02\x83\x75\x28\x02\x84\x79\x09\x00\x85\x79"
-     "\x09\x00"
-     "\x86\x69\x09\x00\x87\x69\x09\x00\x88\x59\x09\x00\x89\x59\x09\x00\x8A\xA9"
-     "\x09\x00"
-     "\x8B\xA9\x09\x00\x8C\xB9\x09\x00\x8D\xB9\x09\x00\x8E\xA9\x09\x00\xF4\x00"
-     "\x00\x80"
-     "\x90\x2D\x28\x02\x91\xB9\x08\x00\x92\xD5\x28\x02\x93\x4D\x24\x02\x94\x79"
-     "\x0E\x00"
-     "\x95\x95\x29\x02\x96\x51\x09\x00\x97\x1D\x26\x02\x98\x71\x09\x00\x99\xF1"
-     "\x09\x00"
-     "\x9A\x71\x09\x00\x9B\x31\x29\x02\x9C\x51\x09\x00\xF9\x00\x00\x80\xFE\x00"
-     "\x00\x80"
-     "\x03\x01\x00\x80\x08\x01\x00\x80\x0D\x01\x00\x80\x16\x01\x00\x80\xCC\x03"
-     "\x00\x80"
-     "\xEF\x04\x00\x80\x8E\x05\x00\x80\xE1\x05\x00\x80\x1C\x06\x00\x80\xA8\x39"
-     "\x09\x00"
-     "\x42\x06\x00\x80\x59\x06\x00\x80\x76\x06\x00\x80\x8D\x06\x00\x80\xD2\x03"
-     "\x00\x80"
-     "\xEB\x03\x00\x80\x04\x04\x00\x80\x1D\x04\x00\x80\x30\x04\x00\x80\x43\x04"
-     "\x00\x80"
-     "\x56\x04\x00\x80\x69\x04\x00\x80\x7C\x04\x00\x80\x8F\x04\x00\x80\xF5\x04"
-     "\x00\x80"
-     "\x9F\x01\x00\x80\xA3\x01\x00\x80\xA7\x01\x00\x80\xAB\x01\x00\x80\xAF\x01"
-     "\x00\x80"
-     "\xB3\x01\x00\x80\xB7\x01\x00\x80\xBB\x01\x00\x80\x80\x9D\x08\x00\x81\xA5"
-     "\x08\x00"
-     "\x82\xAD\x08\x00\x83\xA5\x08\x00\xBF\x01\x00\x80\x85\xB9\x08\x00\x86\xA9"
-     "\x08\x00"
-     "\x87\xA9\x08\x00\xC3\x01\x00\x80\x89\x9D\x08\x00\x8A\x29\x08\x00\x8B\xF1"
-     "\x0C\x00"
-     "\x8C\xED\x0C\x00\x8D\xE9\x0C\x00\x8E\xF9\x28\x02\x8F\x81\x08\x00\x90\xA5"
-     "\x08\x00"
-     "\x91\xA1\x08\x00\x92\x59\x09\x00\x93\x0D\x27\x02\xC7\x01\x00\x80\x95\x75"
-     "\x09\x00"
-     "\x96\x99\x08\x00\xCB\x01\x00\x80\xCF\x01\x00\x80\x99\x55\x09\x00\x9A\xAD"
-     "\x08\x00"
-     "\x9B\x55\x09\x00\x9C\x49\x09\x00\x9D\x4D\x09\x00\xD3\x01\x00\x80\xD7\x01"
-     "\x00\x80"
-     "\xA0\x49\x08\x00\xA1\x49\x08\x00\xA2\x59\x08\x00\xDB\x01\x00\x80\xA4\x45"
-     "\x08\x00"
-     "\xDF\x01\x00\x80\xA6\x41\x08\x00\xE3\x01\x00\x80\xA8\x75\x08\x00\xE7\x01"
-     "\x00\x80"
-     "\xAA\xA1\x09\x00\xAB\x71\x08\x00\xAC\x29\x0C\x00\xAD\x11\x08\x00\xEB\x01"
-     "\x00\x80"
-     "\xAF\x41\x09\x00\xB0\x61\x0C\x00\xB1\x2D\x08\x00\xEF\x01\x00\x80\xB3\xD5"
-     "\x09\x00"
-     "\xB4\x4D\x29\x02\xB5\x35\x08\x00\xB6\x35\x08\x00\xB7\x2D\x08\x00\xB8\x15"
-     "\x08\x00"
-     "\xB9\xD5\x08\x00\xF3\x01\x00\x80\xBB\x15\x08\x00\xBC\x45\x1C\x02\xBD\x41"
-     "\x0F\x00"
-     "\xBE\x05\x08\x00\xBF\xBD\x0B\x00\x80\x45\x08\x00\xF7\x01\x00\x80\xFB\x01"
-     "\x00\x80"
-     "\xFF\x01\x00\x80\x03\x02\x00\x80\x85\x89\x0F\x00\x86\x65\x0A\x00\x87\xE1"
-     "\x0A\x00"
-     "\x88\x11\x0B\x00\x89\xED\x2B\x02\x3B\x09\x00\x80\x48\x23\x00\x80\xA1\x06"
-     "\x00\x80"
-     "\xAB\x06\x00\x80\xAE\x06\x00\x80\xFE\x23\x00\x80\x90\x09\x08\x00\x91\x09"
-     "\x08\x00"
-     "\x92\x19\x08\x00\x93\x19\x08\x00\x94\x09\x08\x00\x95\x09\x08\x00\x96\x31"
-     "\x08\x00"
-     "\x97\x31\x08\x00\x98\xE1\x08\x00\x99\xE1\x08\x00\x9A\xE1\x08\x00\x9B\xE1"
-     "\x08\x00"
-     "\x9C\xE9\x08\x00\x9D\xE9\x08\x00\x9E\xD1\x08\x00\x9F\xD1\x08\x00\xA0\x65"
-     "\x0B\x00"
-     "\xA1\x35\x08\x00\xA2\x3D\x08\x00\xA3\x31\x08\x00\xA4\x11\x06\x00\xA5\x55"
-     "\x08\x00"
-     "\xA6\x51\x08\x00\xA7\x51\x08\x00\xA8\x7D\x08\x00\xA9\x21\x06\x00\xAA\x59"
-     "\x08\x00"
-     "\xAB\x59\x08\x00\xAC\xA1\x0A\x00\xAD\x11\x0B\x00\xAE\x25\x0F\x00\xAF\xA5"
-     "\x0A\x00"
-     "\xB0\xF1\x0B\x00\xB1\x3D\x08\x00\xB2\x35\x08\x00\xB3\x9D\x09\x00\xB4\xF5"
-     "\x0B\x00"
-     "\xB5\x81\x09\x00\xB6\xB9\x09\x00\xB7\xB9\x09\x00\xB8\x89\x09\x00\xB9\xE5"
-     "\x26\x02"
-     "\xBA\x85\x09\x00\xBB\x9D\x09\x00\xBC\xB1\x24\x02\xBD\xCD\x0B\x00\xBE\x85"
-     "\x0A\x00"
-     "\xBF\xED\x0B\x00\x82\x2C\x74\x00\x08\x24\x00\x80\xA3\x90\x0B\x00\xCC\x32"
-     "\x00\x80"
-     "\xA5\x90\x0B\x00\x38\x3B\x00\x80\x87\x4D\x0A\x00\x98\x2B\x00\x80\x89\x7D"
-     "\x0A\x00"
-     "\x88\x7D\x0A\x00\xBC\x2B\x00\x80\xC8\x2B\x00\x80\xAD\x8D\x08\x00\xAC\x8D"
-     "\x08\x00"
-     "\xAF\xF9\x08\x00\xD8\x2B\x00\x80\xE0\x2B\x00\x80\xB0\x81\x08\x00\xEC\x2B"
-     "\x00\x80"
-     "\xF4\x2B\x00\x80\xFC\x2B\x00\x80\x08\x2C\x00\x80\x97\x39\x0A\x00\x63\x2C"
-     "\x00\x80"
-     "\x7B\x2C\x00\x80\x8B\x2C\x00\x80\x9B\x2C\x00\x80\xAB\x2C\x00\x80\xBB\x2C"
-     "\x00\x80"
-     "\xBC\x94\x0B\x00\x9F\x45\x2B\x02\xDB\x2C\x00\x80\xDF\x2C\x00\x80\xE3\x2C"
-     "\x00\x80"
-     "\x83\x28\x0B\x00\x17\x2D\x00\x80\x3F\x2D\x00\x80\x4F\x2D\x00\x80\x57\x2D"
-     "\x00\x80"
-     "\x5B\x2D\x00\x80\x67\x2D\x00\x80\x88\x0C\x0B\x00\x8B\x38\x0B\x00\x83\x2D"
-     "\x00\x80"
-     "\x87\x2D\x00\x80\x9F\x2D\x00\x80\xA3\x2D\x00\x80\xAB\x2D\x00\x80\xB1\xF1"
-     "\x0A\x00"
-     "\xB0\xA1\x0A\x00\xBB\x2D\x00\x80\x57\x2E\x00\x80\xB5\xBD\x0A\x00\xB4\xBD"
-     "\x0A\x00"
-     "\xB7\xAD\x0A\x00\xB6\xAD\x0A\x00\xB9\x9D\x0A\x00\xB8\x9D\x0A\x00\xBB\x6D"
-     "\x0A\x00"
-     "\xBA\x6D\x0A\x00\xBD\x7D\x0A\x00\xBC\x7D\x0A\x00\xBF\x6D\x0A\x00\xBE\x6D"
-     "\x0A\x00"
-     "\x82\x60\x7A\x00\x83\x68\x7A\x00\x94\x07\x00\x80\x6B\x2E\x00\x80\x86\x4C"
-     "\x7A\x00"
-     "\x87\x50\x7A\x00\xAB\x2E\x00\x80\xBF\x2E\x00\x80\xC7\x2E\x00\x80\xCB\x2E"
-     "\x00\x80"
-     "\xD9\x07\x00\x80\xDB\x2E\x00\x80\xFE\x2E\x00\x80\x2A\x2F\x00\x80\x36\x2F"
-     "\x00\x80"
-     "\x56\x2F\x00\x80\x92\xFC\x7A\x00\x62\x2F\x00\x80\x66\x2F\x00\x80\x91\xF8"
-     "\x7A\x00"
-     "\x6A\x2F\x00\x80\x82\x2F\x00\x80\x86\x2F\x00\x80\xB2\x2F\x00\x80\xBE\x2F"
-     "\x00\x80"
-     "\xC6\x2F\x00\x80\xCE\x2F\x00\x80\xDE\x2F\x00\x80\xED\x2F\x00\x80\xF5\x2F"
-     "\x00\x80"
-     "\x21\x30\x00\x80\x25\x30\x00\x80\xA2\x2C\x7A\x00\x2D\x30\x00\x80\x31\x30"
-     "\x00\x80"
-     "\x39\x30\x00\x80\x41\x30\x00\x80\x51\x30\x00\x80\x6D\x30\x00\x80\xE5\x30"
-     "\x00\x80"
-     "\xF9\x30\x00\x80\xAB\x44\x7A\x00\xA8\x2C\x7A\x00\xA9\x20\x7A\x00\xFD\x30"
-     "\x00\x80"
-     "\x7D\x31\x00\x80\x81\x31\x00\x80\x95\x31\x00\x80\xB2\x10\x7A\x00\xB3\xD8"
-     "\x79\x00"
-     "\x99\x31\x00\x80\xA1\x31\x00\x80\xA9\x31\x00\x80\xB1\x31\x00\x80\xB4\xD0"
-     "\x79\x00"
-     "\xB5\xA4\x79\x00\xBD\x31\x00\x80\xF3\x31\x00\x80\x03\x32\x00\x80\x17\x32"
-     "\x00\x80"
-     "\x94\x32\x00\x80\xA0\x32\x00\x80\xB0\x32\x00\x80\xBC\x32\x00\x80\xA0\x89"
-     "\x55\x02"
-     "\xA1\x71\x55\x02\xA2\x71\x55\x02\xA3\x65\x74\x00\xA4\x7D\x74\x00\xA5\x65"
-     "\x74\x00"
-     "\xA6\x6D\x74\x00\xA7\x65\x74\x00\xA8\x5D\x74\x00\xA9\x51\x7B\x00\xAA\x51"
-     "\x7B\x00"
-     "\xAB\xB1\x7B\x00\xAC\x95\x7B\x00\xAD\xF1\x7B\x00\xAE\xED\x7B\x00\xAF\xD9"
-     "\x7B\x00"
-     "\xB0\xA9\x7B\x00\xB1\xD5\x7A\x00\xB2\xD5\x7A\x00\xB3\xED\x7A\x00\xB4\xF5"
-     "\x7A\x00"
-     "\xB5\xF9\x7A\x00\xB6\xE5\x7A\x00\xB7\x9D\x7A\x00\xB8\xA5\x7A\x00\xB9\xAD"
-     "\x7A\x00"
-     "\xBA\xA5\x7A\x00\xBB\xBD\x7A\x00\xBC\xA5\x7A\x00\xBD\x95\x7A\x00\xBE\x9D"
-     "\x7A\x00"
-     "\xBF\x89\x7A\x00\xC8\x32\x00\x80\xD0\x32\x00\x80\xD4\x32\x00\x80\xD8\x32"
-     "\x00\x80"
-     "\xDC\x32\x00\x80\xE0\x32\x00\x80\x98\x33\x00\x80\xE0\x33\x00\x80\xEC\x33"
-     "\x00\x80"
-     "\xF0\x33\x00\x80\xF4\x33\x00\x80\xF8\x33\x00\x80\x00\x34\x00\x80\x04\x34"
-     "\x00\x80"
-     "\x18\x34\x00\x80\x34\x34\x00\x80\x3C\x34\x00\x80\x7C\x34\x00\x80\x80\x34"
-     "\x00\x80"
-     "\x88\x34\x00\x80\x8C\x34\x00\x80\x94\x34\x00\x80\x98\x34\x00\x80\xBC\x34"
-     "\x00\x80"
-     "\xC0\x34\x00\x80\xDC\x34\x00\x80\xE0\x34\x00\x80\xE4\x34\x00\x80\xE8\x34"
-     "\x00\x80"
-     "\xFC\x34\x00\x80\x20\x35\x00\x80\x24\x35\x00\x80\x80\xF9\x79\x00\x81\x39"
-     "\x7E\x00"
-     "\x82\x7D\x7E\x00\x83\x75\x7E\x00\x84\x6D\x7E\x00\x85\x95\x7E\x00\x86\x9D"
-     "\x7E\x00"
-     "\x87\x95\x7E\x00\x88\xAD\x7E\x00\x89\xB5\x7E\x00\x8A\xBD\x7E\x00\x8B\xB5"
-     "\x7E\x00"
-     "\x8C\xAD\x7E\x00\x8D\x91\x7E\x00\x8E\x91\x7E\x00\x8F\x91\x7E\x00\x90\xF1"
-     "\x7E\x00"
-     "\x91\xF1\x7E\x00\x92\xF1\x7E\x00\x93\xF1\x7E\x00\x94\x91\x7E\x00\x95\x91"
-     "\x7E\x00"
-     "\x96\x91\x7E\x00\x97\x91\x7E\x00\x98\xB1\x7E\x00\x99\xB1\x7E\x00\x9A\xB1"
-     "\x7E\x00"
-     "\x9B\xB1\x7E\x00\x9C\x91\x7E\x00\x9D\x91\x7E\x00\x9E\x91\x7E\x00\x9F\x91"
-     "\x7E\x00"
-     "\xA0\x71\x7E\x00\xA1\x71\x7E\x00\xA2\x71\x7E\x00\xA3\x71\x7E\x00\xA4\x81"
-     "\x7F\x00"
-     "\xA5\xB5\x7F\x00\xA6\xC9\x7F\x00\xA7\xC1\x7F\x00\xA8\xC1\x7F\x00\xA9\xDD"
-     "\x7F\x00"
-     "\xAA\xD5\x7F\x00\xAB\x21\x7F\x00\xAC\x25\x7F\x00\xAD\x29\x7F\x00\xAE\x1D"
-     "\x7F\x00"
-     "\xAF\x05\x7F\x00\xB0\x09\x7F\x00\xB1\x2D\x7F\x00\xB2\xCD\x7C\x00\xB3\xC5"
-     "\x7C\x00"
-     "\xB4\xDD\x7C\x00\xB5\xC5\x7C\x00\xB6\x55\x73\x00\xB7\x45\x73\x00\xB8\x45"
-     "\x77\x00"
-     "\xB9\x79\x73\x00\xBA\x49\x73\x00\xBB\x05\x77\x00\xBC\x59\x73\x00\xBD\x29"
-     "\x73\x00"
-     "\xBE\x1D\x77\x00\xBF\xFD\x76\x00\x28\x35\x00\x80\x88\xDC\x77\x00\x4C\x35"
-     "\x00\x80"
-     "\x8C\x35\x00\x80\x8D\xD0\x77\x00\x90\x35\x00\x80\xCC\x28\x4B\x02\xCD\x5C"
-     "\x4B\x02"
-     "\xB0\x35\x00\x80\xC0\x35\x00\x80\x83\xC0\x77\x00\xC8\x35\x00\x80\x85\xE0"
-     "\x77\x00"
-     "\xE8\x35\x00\x80\xCC\x7C\x49\x02\xCD\xB4\x46\x02\x0C\x36\x00\x80\x50\x36"
-     "\x00\x80"
-     "\x58\x36\x00\x80\x6C\x36\x00\x80\x78\x36\x00\x80\x80\x36\x00\x80\x88\x36"
-     "\x00\x80"
-     "\x8C\x36\x00\x80\x90\x36\x00\x80\xA8\x36\x00\x80\xAC\x36\x00\x80\xC8\x36"
-     "\x00\x80"
-     "\xD0\x36\x00\x80\xD8\x36\x00\x80\xDC\x36\x00\x80\xE8\x36\x00\x80\xBE\xA8"
-     "\x08\x00"
-     "\xF0\x36\x00\x80\xF4\x36\x00\x80\xF8\x36\x00\x80\x00\x37\x00\x80\x04\x37"
-     "\x00\x80"
-     "\xA4\x37\x00\x80\xB8\x37\x00\x80\xA1\x68\x77\x00\xEC\x37\x00\x80\xF0\x37"
-     "\x00\x80"
-     "\xF8\x37\x00\x80\xA5\x08\x78\x00\xA4\xE0\x78\x00\x24\x38\x00\x80\x2C\x38"
-     "\x00\x80"
-     "\x34\x38\x00\x80\x3C\x38\x00\x80\xBB\x10\x78\x00\xBA\x24\x78\x00\xBD\xD8"
-     "\x79\x00"
-     "\xBC\xC8\x79\x00\x78\x38\x00\x80\x7C\x38\x00\x80\x84\x38\x00\x80\x88\x38"
-     "\x00\x80"
-     "\xB3\x7C\x78\x00\xB2\x44\x78\x00\x8C\x38\x00\x80\xB8\x38\x00\x80\xB7\x04"
-     "\x78\x00"
-     "\xB6\x0C\x78\x00\x80\x1D\x7B\x00\x81\x2D\x7B\x00\x82\x25\x7B\x00\x83\x31"
-     "\x7B\x00"
-     "\x84\x51\x7B\x00\x85\x5D\x7B\x00\x86\x55\x7B\x00\x87\x4D\x7B\x00\x88\x75"
-     "\x7B\x00"
-     "\x89\x79\x7B\x00\x8A\x49\x7B\x00\x8B\x49\x7B\x00\x8C\x59\x7B\x00\x8D\x59"
-     "\x7B\x00"
-     "\x8E\x49\x7B\x00\x8F\x41\x7B\x00\x90\xC1\x7B\x00\x91\xC1\x7B\x00\x92\xC5"
-     "\x7B\x00"
-     "\x93\xDD\x7B\x00\x94\xC5\x7B\x00\x95\xCD\x7B\x00\x96\xC5\x7B\x00\x97\xFD"
-     "\x7B\x00"
-     "\x98\xCD\x7B\x00\x99\xD5\x7B\x00\x9A\xDD\x7B\x00\x9B\xD5\x7B\x00\x9C\xCD"
-     "\x7B\x00"
-     "\x9D\xBD\x7B\x00\x9E\xB5\x7B\x00\x9F\x71\x04\x00\xA0\x99\x04\x00\xA1\x99"
-     "\x04\x00"
-     "\xA2\xA9\x04\x00\xA3\xA9\x04\x00\xA4\xBD\x04\x00\xA5\xA5\x04\x00\xA6\xAD"
-     "\x04\x00"
-     "\xA7\xA5\x04\x00\xA8\x9D\x04\x00\xA9\xE1\x04\x00\xAA\xE1\x04\x00\xAB\xFD"
-     "\x04\x00"
-     "\xAC\xE5\x04\x00\xAD\xED\x04\x00\xAE\xE5\x04\x00\xAF\xDD\x04\x00\xB0\xA5"
-     "\x04\x00"
-     "\xB1\xAD\x04\x00\xB2\xA5\x04\x00\xB3\xBD\x04\x00\xB4\xA5\x04\x00\xB5\xAD"
-     "\x04\x00"
-     "\xB6\xA5\x04\x00\xB7\x59\x04\x00\xB8\x69\x04\x00\xB9\x69\x04\x00\xBA\x79"
-     "\x04\x00"
-     "\xBB\x79\x04\x00\xBC\x69\x04\x00\xBD\x69\x04\x00\xBE\x59\x04\x00\xBF\x4D"
-     "\x04\x00"
-     "\x80\xB5\x05\x00\x81\xBD\x05\x00\x82\xB5\x05\x00\x83\xC5\x05\x00\x84\xDD"
-     "\x05\x00"
-     "\x85\xC5\x05\x00\x86\xCD\x05\x00\x87\xFD\x05\x00\x88\xC5\x05\x00\x89\xCD"
-     "\x05\x00"
-     "\x8A\xC5\x05\x00\x8B\xDD\x05\x00\x8C\xC5\x05\x00\x8D\xCD\x05\x00\x8E\xC5"
-     "\x05\x00"
-     "\x8F\xBD\x04\x00\x90\xC5\x04\x00\x91\xCD\x04\x00\x92\xC5\x04\x00\x93\xDD"
-     "\x04\x00"
-     "\x94\xC5\x04\x00\x95\xCD\x04\x00\x96\xC5\x04\x00\x97\xFD\x04\x00\x98\xC5"
-     "\x04\x00"
-     "\x99\xCD\x04\x00\x9A\xC5\x04\x00\x9B\xDD\x04\x00\x9C\xC5\x04\x00\x9D\xCD"
-     "\x04\x00"
-     "\x9E\xC5\x04\x00\x9F\xBD\x04\x00\xA0\x45\x04\x00\xA1\x4D\x04\x00\xA2\x45"
-     "\x04\x00"
-     "\xA3\x5D\x04\x00\xA4\x45\x04\x00\xA5\x4D\x04\x00\xA6\x45\x04\x00\xA7\x7D"
-     "\x04\x00"
-     "\xA8\x45\x04\x00\xA9\x4D\x04\x00\xAA\x45\x04\x00\xAB\x5D\x04\x00\xAC\x45"
-     "\x04\x00"
-     "\xAD\x4D\x04\x00\xAE\x45\x04\x00\xAF\xBD\x06\x00\xB0\xC9\x06\x00\xB1\xC9"
-     "\x06\x00"
-     "\xB2\xDD\x06\x00\xB3\xED\x06\x00\xB4\xF5\x06\x00\xB5\xF9\x06\x00\xB6\xED"
-     "\x06\x00"
-     "\xB7\x9D\x06\x00\xB8\xA5\x06\x00\xB9\xAD\x06\x00\xBA\xA5\x06\x00\xBB\xBD"
-     "\x06\x00"
-     "\xBC\xA5\x06\x00\xBD\xAD\x06\x00\xBE\xA5\x06\x00\xBF\x9D\x06\x00\x80\x65"
-     "\x05\x00"
-     "\x81\x6D\x05\x00\x82\x65\x05\x00\x83\x7D\x05\x00\x84\x65\x05\x00\x85\x6D"
-     "\x05\x00"
-     "\x86\x65\x05\x00\x87\x99\x05\x00\x88\xA9\x05\x00\x89\xA9\x05\x00\x8A\xB9"
-     "\x05\x00"
-     "\x8B\xB9\x05\x00\x8C\xA9\x05\x00\x8D\xA9\x05\x00\x8E\x9D\x05\x00\x8F\x95"
-     "\x05\x00"
-     "\x90\xED\x05\x00\x91\xFD\x05\x00\x92\xF5\x05\x00\x93\x8D\x05\x00\x94\x95"
-     "\x05\x00"
-     "\x95\x8D\x05\x00\x96\x85\x05\x00\x97\xBD\x05\x00\x98\x85\x05\x00\x99\x95"
-     "\x05\x00"
-     "\x9A\x9D\x05\x00\x9B\xED\x01\x00\x9C\xF5\x01\x00\x9D\xFD\x01\x00\x9E\xF5"
-     "\x01\x00"
-     "\x9F\xED\x01\x00\xA0\x15\x01\x00\xA1\x1D\x01\x00\xA2\x15\x01\x00\xA3\x2D"
-     "\x01\x00"
-     "\xA4\x35\x01\x00\xA5\x3D\x01\x00\xA6\x35\x01\x00\xA7\x2D\x01\x00\xA8\x15"
-     "\x01\x00"
-     "\xA9\x1D\x01\x00\xAA\x15\x01\x00\xAB\x4D\x01\x00\xAC\x41\x01\x00\xAD\x41"
-     "\x01\x00"
-     "\xAE\x41\x01\x00\xAF\x41\x01\x00\xB0\x01\x0C\x00\xB1\xC5\x0E\x00\xB2\xCD"
-     "\x0E\x00"
-     "\xB3\xC5\x0E\x00\xB4\xDD\x0E\x00\xB5\xC5\x0E\x00\xB6\xCD\x0E\x00\xB7\xC5"
-     "\x0E\x00"
-     "\xB8\xFD\x0E\x00\xB9\xC5\x0E\x00\xBA\xCD\x0E\x00\xBB\xC5\x0E\x00\xBC\xDD"
-     "\x0E\x00"
-     "\xBD\xC5\x0E\x00\xBE\xCD\x0E\x00\xBF\xC5\x0E\x00\x93\x7D\x73\x00\x92\x3D"
-     "\x72\x00"
-     "\x91\x71\x73\x00\x90\x61\x53\x02\x97\xD1\x52\x02\x96\x15\x72\x00\x95\x59"
-     "\x72\x00"
-     "\x94\xE1\x73\x00\x9B\x61\x5D\x02\x9A\x2D\x72\x00\x99\xE1\x52\x02\x98\x1D"
-     "\x72\x00"
-     "\x9F\x29\x72\x00\x9E\xA9\x53\x02\x9D\xAD\x72\x00\x9C\x2D\x72\x00\x83\x61"
-     "\x73\x00"
-     "\x82\x5D\x73\x00\x81\x59\x73\x00\x80\x45\x73\x00\x87\x69\x73\x00\x86\x91"
-     "\x72\x00"
-     "\x85\x69\x73\x00\x84\x75\x73\x00\x8B\xE1\x7F\x00\x8A\x0D\x73\x00\x89\x6D"
-     "\x73\x00"
-     "\x88\x65\x07\x00\x8F\xA5\x72\x00\x8E\x5D\x07\x00\x8D\xFD\x7F\x00\x8C\x0D"
-     "\x73\x00"
-     "\xBC\x38\x00\x80\xC4\x38\x00\x80\xD8\x38\x00\x80\xDC\x38\x00\x80\xF8\x38"
-     "\x00\x80"
-     "\x00\x39\x00\x80\x04\x39\x00\x80\x28\x39\x00\x80\x48\x39\x00\x80\x5C\x39"
-     "\x00\x80"
-     "\x60\x39\x00\x80\x68\x39\x00\x80\x94\x39\x00\x80\x98\x39\x00\x80\xA8\x39"
-     "\x00\x80"
-     "\xCC\x39\x00\x80\xA3\xE1\x72\x00\xA2\x99\x52\x02\xA1\x19\x52\x02\xA0\xE5"
-     "\x79\x00"
-     "\xA7\xD1\x5F\x02\xA6\x09\x52\x02\xA5\xFD\x72\x00\xA4\xF5\x72\x00\xD8\x39"
-     "\x00\x80"
-     "\xAA\x81\x72\x00\xA9\xA1\x72\x00\xA8\x1D\x52\x02\xE8\x39\x00\x80\x00\x3A"
-     "\x00\x80"
-     "\x08\x3A\x00\x80\x0C\x3A\x00\x80\x14\x3A\x00\x80\x20\x3A\x00\x80\x24\x3A"
-     "\x00\x80"
-     "\x2C\x3A\x00\x80\x48\x3A\x00\x80\x4C\x3A\x00\x80\x50\x3A\x00\x80\x68\x3A"
-     "\x00\x80"
-     "\x6C\x3A\x00\x80\x74\x3A\x00\x80\x88\x3A\x00\x80\x8C\x3A\x00\x80\x90\x3A"
-     "\x00\x80"
-     "\x98\x3A\x00\x80\x9C\x3A\x00\x80\xA0\x3A\x00\x80\xA4\x3A\x00\x80\xA8\x3A"
-     "\x00\x80"
-     "\xB8\x3A\x00\x80\xD0\x3A\x00\x80\xD4\x3A\x00\x80\xE0\x3A\x00\x80\xE4\x3A"
-     "\x00\x80"
-     "\xE8\x3A\x00\x80\xEC\x3A\x00\x80\xF4\x3A\x00\x80\xF8\x3A\x00\x80\xFC\x3A"
-     "\x00\x80"
-     "\x08\x3B\x00\x80\x10\x3B\x00\x80\x18\x3B\x00\x80\x1C\x3B\x00\x80\x80\xBD"
-     "\x00\x00"
-     "\x81\xC5\x00\x00\x82\xCD\x00\x00\x83\xC5\x00\x00\x84\xDD\x00\x00\x85\xC5"
-     "\x00\x00"
-     "\x86\xCD\x00\x00\x87\xC5\x00\x00\x88\xFD\x00\x00\x89\xC5\x00\x00\x8A\xCD"
-     "\x00\x00"
-     "\x8B\xC5\x00\x00\x8C\xDD\x00\x00\x8D\xC5\x00\x00\x8E\xCD\x00\x00\x8F\xC5"
-     "\x00\x00"
-     "\x90\xBD\x00\x00\x91\x1D\x00\x00\x92\x15\x00\x00\x93\x2D\x00\x00\x94\x35"
-     "\x00\x00"
-     "\x95\x3D\x00\x00\x24\x3B\x00\x80\x28\x3B\x00\x80\x2C\x3B\x00\x80\x34\x3B"
-     "\x00\x80"
-     "\x3C\x3B\x00\x80\x3C\x24\x00\x80\x6B\x2D\x00\x80\x6F\x2D\x00\x80\x24\x25"
-     "\x00\x80"
-     "\xB2\x24\x00\x80\x80\x1C\x02\x00\x81\x9C\x01\x00\x82\x24\x07\x00\x83\x08"
-     "\x05\x00"
-     "\x84\x1C\x04\x00\x85\x14\x1A\x00\x86\x80\x19\x00\x80\x21\x5D\x02\x88\x00"
-     "\x1C\x00"
-     "\x89\x24\x1E\x00\x8A\x00\x10\x00\x8B\x2C\x12\x00\x8C\xF4\x11\x00\x8D\x00"
-     "\x28\x00"
-     "\x8E\x38\x2B\x00\x8F\x3C\x2A\x00\xB6\x24\x00\x80\xE3\x4C\x74\x00\x99\x05"
-     "\x00\x00"
-     "\xBA\x24\x00\x80\x9E\x1D\x00\x00\xE3\x04\x02\x00\x82\x60\x02\x00\x99\x3D"
-     "\x00\x00"
-     "\xBE\x24\x00\x80\xEF\xF4\x02\x00\xBE\x7C\x02\x00\x9E\x0D\x00\x00\x82\x68"
-     "\x02\x00"
-     "\xEF\x0C\x08\x00\xC2\x24\x00\x80\x99\x05\x00\x00\xE3\xF8\x02\x00\xBE\x74"
-     "\x02\x00"
-     "\x9E\xF1\x00\x00\xE3\x14\x02\x00\x82\x6C\x02\x00\xC6\x24\x00\x80\x99\x0D"
-     "\x00\x00"
-     "\xBE\x44\x02\x00\x9E\x35\x00\x00\x82\x68\x02\x00\x99\x05\x00\x00\xCA\x24"
-     "\x00\x80"
-     "\xEF\xCC\x02\x00\xBE\x74\x02\x00\x9E\x15\x00\x00\xEF\xD8\x02\x00\xE3\x6C"
-     "\x02\x00"
-     "\xB6\x29\x01\x00\xCE\x24\x00\x80\x82\x7C\x02\x00\xE3\x14\x02\x00\x99\x1D"
-     "\x00\x00"
-     "\xBE\x7C\x02\x00\x9E\x15\x00\x00\xEF\xC0\x02\x00\x82\x68\x02\x00\x99\x05"
-     "\x00\x00"
-     "\xD2\x24\x00\x80\xEF\xBC\x02\x00\xBA\x01\x01\x00\xB9\x09\x01\x00\xB8\x5D"
-     "\x76\x00"
-     "\xA6\xAC\x02\x00\xE3\xE8\x01\x00\xA4\xD8\x03\x00\xE3\x80\x02\x00\xD6\x24"
-     "\x00\x80"
-     "\xDA\x24\x00\x80\xDE\x24\x00\x80\xA1\xC4\x03\x00\xE2\x24\x00\x80\xAF\xA8"
-     "\x05\x00"
-     "\xBE\xE0\x02\x00\xBE\xEC\x02\x00\xE6\x24\x00\x80\xEF\xE4\x02\x00\xA8\x00"
-     "\x04\x00"
-     "\xEF\xE8\x02\x00\xE3\xC0\x02\x00\xBE\xF4\x02\x00\xEA\x24\x00\x80\xB5\xC8"
-     "\x05\x00"
-     "\xB2\xD8\x05\x00\xE3\xD0\x02\x00\xBE\xC8\x02\x00\xE3\xC0\x02\x00\xEE\x24"
-     "\x00\x80"
-     "\xEF\xC4\x02\x00\xF2\x24\x00\x80\xEF\x94\x03\x00\xEF\xCC\x02\x00\xBB\xE8"
-     "\x05\x00"
-     "\xB8\xF8\x05\x00\xF6\x24\x00\x80\x86\x94\x72\x00\x82\x40\x02\x00\x82\x44"
-     "\x02\x00"
-     "\x82\x38\x02\x00\xFA\x24\x00\x80\x82\x00\x02\x00\xFE\x24\x00\x80\x02\x25"
-     "\x00\x80"
-     "\x99\xB1\x00\x00\x8F\x6C\x00\x00\x99\xBD\x00\x00\x8D\xCC\x72\x00\x9E\x5D"
-     "\x03\x00"
-     "\x8B\xD0\x72\x00\x9E\xA9\x00\x00\x06\x25\x00\x80\x10\x3A\x00\x80\x97\x88"
-     "\x00\x00"
-     "\x9E\xDD\x00\x00\x95\xCC\x00\x00\x99\xC9\x00\x00\x93\xD8\x00\x00\x94\x00"
-     "\x00\x00"
-     "\x91\x04\x00\x00\x0A\x25\x00\x80\x9F\x28\x03\x00\x9E\xF1\x00\x00\x9D\x6C"
-     "\x03\x00"
-     "\x99\xD1\x00\x00\x9B\x7C\x03\x00\x15\x25\x00\x80\x99\xA0\x00\x00\xBE\x7C"
-     "\x02\x00"
-     "\x9E\x25\x01\x00\x82\x60\x02\x00\x99\x3D\x01\x00\xEF\xF0\x02\x00\x9E\x25"
-     "\x01\x00"
-     "\xBE\x74\x02\x00\x82\x54\x02\x00\xE3\x20\x02\x00\xBE\x74\x02\x00\x9E\x61"
-     "\x01\x00"
-     "\x99\x65\x01\x00\x82\x6C\x02\x00\x9A\x5D\x01\x00\x99\x59\x01\x00\xBE\x00"
-     "\x02\x00"
-     "\x9F\x29\x01\x00\x9E\x25\x01\x00\xBE\x50\x02\x00\x82\x78\x02\x00\x99\x2D"
-     "\x01\x00"
-     "\x99\xCD\x01\x00\x9A\xC5\x01\x00\x9A\xCD\x01\x00\x9E\x1D\x01\x00\x9F\xF5"
-     "\x01\x00"
-     "\xBE\x60\x02\x00\x82\x5C\x02\x00\x9E\xE9\x01\x00\x9F\xE9\x01\x00\x82\x9C"
-     "\x02\x00"
-     "\xBE\x94\x02\x00\x99\x3D\x01\x00\x82\x44\x02\x00\x19\x25\x00\x80\x9A\x6D"
-     "\x01\x00"
-     "\x9E\x2D\x01\x00\x9F\x75\x01\x00\x99\x21\x00\x00\xBC\x00\x00\x80\xBE\x18"
-     "\x02\x00"
-     "\x99\x5D\x01\x00\x9A\x21\x00\x00\x82\x50\x02\x00\xCE\x00\x00\x00\xC1\x00"
-     "\x00\x80"
-     "\x30\x25\x00\x80\x1D\x25\x00\x80\x9E\x39\x01\x00\x9F\x4D\x00\x00\x6C\x2A"
-     "\x00\x80"
-     "\x9E\x75\x00\x00\xBE\x64\x02\x00\xBE\x4C\x02\x00\x99\x11\x00\x00\x82\x60"
-     "\x02\x00"
-     "\x9E\x29\x00\x00\x07\x26\x00\x80\xBE\xD8\x05\x00\xD9\x00\x00\x00\xDE\x00"
-     "\x00\x00"
-     "\x6A\x26\x00\x80\x82\xD8\x05\x00\x82\xF8\x05\x00\xE3\x38\x05\x00\xE3\x40"
-     "\x05\x00"
-     "\xE3\x6C\x05\x00\xE3\x7C\x05\x00\xE3\x9C\x05\x00\xE3\xB0\x05\x00\xE3\xCC"
-     "\x05\x00"
-     "\xE3\xEC\x05\x00\xEF\x78\x05\x00\xEF\x0C\x05\x00\xEF\x04\x05\x00\xEF\x3C"
-     "\x05\x00"
-     "\xEF\xF4\x05\x00\xEF\x84\x05\x00\xEF\xA8\x05\x00\xEF\xB8\x05\x00\x9B\xDD"
-     "\x07\x00"
-     "\x99\xA1\x07\x00\x9E\xAD\x07\x00\xE3\xB8\x05\x00\x9F\xED\x07\x00\xBE\x48"
-     "\x02\x00"
-     "\x9D\xC0\x03\x00\x9C\xE9\x07\x00\x99\x95\x07\x00\x9E\x91\x07\x00\x82\x60"
-     "\x02\x00"
-     "\x99\x1D\x00\x00\xAD\x26\x00\x80\xBE\x74\x02\x00\x9E\x09\x00\x00\xEF\xD0"
-     "\x02\x00"
-     "\xAB\x10\x03\x00\x99\x19\x00\x00\xE3\xE8\x02\x00\xE3\x10\x02\x00\xAF\x94"
-     "\x03\x00"
-     "\x82\x74\x02\x00\xAD\x58\x03\x00\xD9\x26\x00\x80\xBE\x7C\x02\x00\x9E\x39"
-     "\x00\x00"
-     "\x82\x60\x02\x00\x99\x1D\x00\x00\x03\x27\x00\x80\xA6\x20\x03\x00\xEF\x34"
-     "\x02\x00"
-     "\xEF\xE4\x02\x00\xBB\x04\x02\x00\xE3\x64\x02\x00\xB9\x04\x02\x00\xBE\x68"
-     "\x02\x00"
-     "\xBF\xC8\x02\x00\xE3\x30\x02\x00\xBD\x0C\x02\x00\x9E\x6D\x00\x00\xB3\xA8"
-     "\x03\x00"
-     "\xEF\x18\x03\x00\xB1\xE8\x03\x00\x82\x74\x02\x00\xB7\x70\x02\x00\xEF\xC0"
-     "\x02\x00"
-     "\xB5\xA0\x03\x00\x99\xFD\x01\x00\x28\x27\x00\x80\xBE\x74\x02\x00\x9E\x09"
-     "\x00\x00"
-     "\xE3\x14\x02\x00\x82\x6C\x02\x00\x3D\x27\x00\x80\x99\x0D\x00\x00\xBE\x40"
-     "\x02\x00"
-     "\x82\x6C\x02\x00\x9E\x31\x00\x00\x99\x05\x00\x00\x45\x27\x00\x80\xBE\x70"
-     "\x02\x00"
-     "\x82\x14\x02\x00\x9E\x15\x00\x00\xEF\xD8\x02\x00\xE3\x68\x02\x00\x99\x09"
-     "\x00\x00"
-     "\xE3\x74\x02\x00\x61\x27\x00\x80\xBE\x7C\x02\x00\x9E\x19\x00\x00\xE3\x0C"
-     "\x02\x00"
-     "\x82\x54\x02\x00\x69\x27\x00\x80\x9E\x05\x00\x00\xEF\x10\x02\x00\x99\x0D"
-     "\x00\x00"
-     "\xEF\xB8\x02\x00\x99\xFD\x00\x00\xEF\xD0\x02\x00\x82\x6C\x02\x00\xBE\x9C"
-     "\x02\x00"
-     "\xEF\x58\x04\x00\x71\x27\x00\x80\x9E\x05\x00\x00\xBE\x64\x02\x00\xE3\xE0"
-     "\x02\x00"
-     "\xE3\x3C\x02\x00\x99\x3D\x00\x00\x83\x27\x00\x80\xEF\xCC\x02\x00\xEF\xEC"
-     "\x02\x00"
-     "\x82\x54\x02\x00\x99\x79\x00\x00\xE3\xA0\x04\x00\x82\x6C\x02\x00\xBE\x04"
-     "\x02\x00"
-     "\x9E\x61\x00\x00\xE3\x74\x02\x00\x8F\x27\x00\x80\xBE\x60\x02\x00\xE3\x00"
-     "\x02\x00"
-     "\x9E\x1D\x00\x00\x82\xB8\x01\x00\x97\x27\x00\x80\xEF\xF4\x02\x00\x82\x88"
-     "\x01\x00"
-     "\xAD\x27\x00\x80\xBE\x98\x01\x00\xEB\x27\x00\x80\xEF\xC0\x02\x00\x82\xAC"
-     "\x01\x00"
-     "\xFC\x27\x00\x80\x81\x14\x00\x00\x99\xD9\x03\x00\x99\xE1\x03\x00\x9E\xD1"
-     "\x03\x00"
-     "\xE3\xFC\x01\x00\x84\x5C\x00\x00\xBE\x48\x02\x00\x86\x40\x00\x00\xEF\xC0"
-     "\x02\x00"
-     "\x88\xB8\x03\x00\x9E\xC1\x03\x00\xBE\x0C\x02\x00\x99\xC1\x03\x00\x82\x3C"
-     "\x02\x00"
-     "\x8F\xF4\x03\x00\x9A\xC1\x03\x00\xE3\xB0\x01\x00\x1E\x28\x00\x80\xE3\x74"
-     "\x02\x00"
-     "\x92\x88\x03\x00\x95\x4C\x03\x00\x29\x28\x00\x80\x9E\xA9\x03\x00\x9F\xA1"
-     "\x03\x00"
-     "\x9A\x35\x00\x00\x98\x54\x03\x00\x9B\xE0\x02\x00\x99\x29\x00\x00\xEF\xA4"
-     "\x02\x00"
-     "\xBE\x8C\x02\x00\xEF\xF4\x02\x00\x79\x28\x00\x80\x9E\xC5\x00\x00\x9F\xD1"
-     "\x00\x00"
-     "\x82\x78\x02\x00\xEF\xBC\x02\x00\xE3\x10\x02\x00\x99\xE9\x00\x00\x9A\x31"
-     "\x00\x00"
-     "\xBE\x54\x02\x00\xEF\xC4\x02\x00\x82\x10\x02\x00\x8D\x28\x00\x80\x82\xC8"
-     "\x03\x00"
-     "\x9E\xCD\x00\x00\x9F\x1D\x00\x00\xAF\x3C\x02\x00\xE3\x10\x02\x00\xB1\x04"
-     "\x02\x00"
-     "\xB0\x08\x02\x00\xBE\x50\x02\x00\xB2\x28\x02\x00\x99\x09\x00\x00\xBE\xBC"
-     "\x03\x00"
-     "\xAF\x28\x00\x80\x9A\x35\x00\x00\x9E\x39\x00\x00\x9F\x0D\x00\x00\xBD\x28"
-     "\x00\x80"
-     "\x05\x29\x00\x80\xBD\xFC\x02\x00\x13\x29\x00\x80\xBF\x45\x01\x00\x82\x80"
-     "\x03\x00"
-     "\x99\xED\x01\x00\xE3\xA0\x03\x00\xE3\xFC\x03\x00\x9A\xF9\x01\x00\x9E\xFD"
-     "\x01\x00"
-     "\x9F\xE1\x01\x00\x99\x3D\x00\x00\xBE\x44\x02\x00\x9E\x05\x00\x00\x70\x2A"
-     "\x00\x80"
-     "\x82\x60\x02\x00\x99\x1D\x00\x00\x74\x2A\x00\x80\xEF\x18\x03\x00\xEF\xDC"
-     "\x02\x00"
-     "\x9E\x0D\x00\x00\xE3\x60\x02\x00\xBE\x00\x02\x00\xE3\x14\x02\x00\x99\x1D"
-     "\x00\x00"
-     "\x78\x2A\x00\x80\x82\x7C\x02\x00\xEF\xD4\x02\x00\xBE\x44\x02\x00\x9E\x31"
-     "\x00\x00"
-     "\x82\x6C\x02\x00\xE3\x04\x02\x00\x99\xCD\x00\x00\xEF\xCC\x02\x00\xBE\x70"
-     "\x02\x00"
-     "\xEF\xDC\x02\x00\x9E\xDD\x00\x00\xE3\x00\x02\x00\x99\x31\x00\x00\x82\x68"
-     "\x02\x00"
-     "\xBE\x60\x02\x00\xE4\x00\x00\x00\x9E\x21\x00\x00\xE6\x00\x00\x00\xE7\x00"
-     "\x00\x00"
-     "\x7C\x2A\x00\x80\xE9\x00\x00\x00\x9C\x51\x2D\x00\x9D\x51\x2D\x00\xEF\xC4"
-     "\x02\x00"
-     "\x88\x2A\x00\x80\xEE\x00\x00\x00\xE8\x26\x00\x80\xB3\x0D\x00\x00\xB2\xC5"
-     "\x1F\x00"
-     "\xB1\x01\x1C\x00\x18\x1B\x00\x80\xB7\xC1\x1F\x00\xB6\xD9\x1F\x00\xB5\xD1"
-     "\x1F\x00"
-     "\xB4\xD9\x1F\x00\xBB\xFD\x1F\x00\xBA\xFD\x1F\x00\xB9\xA9\x1F\x00\xB8\xF9"
-     "\x1F\x00"
-     "\xBF\xDD\x1F\x00\xBE\xDD\x1F\x00\xBD\xDD\x1F\x00\xBC\xDD\x1F\x00\xF4\x19"
-     "\x00\x80"
-     "\xF8\x19\x00\x80\x1C\x1B\x00\x80\x20\x1B\x00\x80\xFC\x19\x00\x80\x00\x1A"
-     "\x00\x80"
-     "\x24\x1B\x00\x80\x28\x1B\x00\x80\x2C\x1B\x00\x80\x30\x1B\x00\x80\x34\x1B"
-     "\x00\x80"
-     "\x38\x1B\x00\x80\x44\x1A\x00\x80\x58\x1A\x00\x80\xA4\x1A\x00\x80\xA8\x1A"
-     "\x00\x80"
-     "\x04\x1A\x00\x80\xE3\x8C\x79\x02\xAC\x1A\x00\x80\xE1\x80\x7A\x02\x08\x1A"
-     "\x00\x80"
-     "\x0C\x1A\x00\x80\xB0\x1A\x00\x80\xB4\x1A\x00\x80\xB8\x1A\x00\x80\xBC\x1A"
-     "\x00\x80"
-     "\xC0\x1A\x00\x80\xC4\x1A\x00\x80\xC8\x1A\x00\x80\xEF\x8C\x79\x02\xCC\x1A"
-     "\x00\x80"
-     "\x10\x1A\x00\x80\x14\x1A\x00\x80\x18\x1A\x00\x80\xD0\x1A\x00\x80\xD4\x1A"
-     "\x00\x80"
-     "\x1C\x1A\x00\x80\x20\x1A\x00\x80\xEF\x5C\x49\x00\xEF\x44\x46\x00\xE1\x64"
-     "\x40\x00"
-     "\xE1\xA0\xAB\x03\xE3\x70\x41\x00\xE3\xF8\x54\x00\xD8\x1A\x00\x80\xDC\x1A"
-     "\x00\x80"
-     "\xE0\x1A\x00\x80\xE4\x1A\x00\x80\x24\x1A\x00\x80\x28\x1A\x00\x80\xE8\x1A"
-     "\x00\x80"
-     "\xEC\x1A\x00\x80\x2C\x1A\x00\x80\x30\x1A\x00\x80\xF0\x1A\x00\x80\xF4\x1A"
-     "\x00\x80"
-     "\xA0\x1A\x00\x80\x3C\x1A\x00\x80\x40\x1A\x00\x80\x3C\x1B\x00\x80\x40\x1B"
-     "\x00\x80"
-     "\x44\x1B\x00\x80\x48\x1B\x00\x80\x4C\x1B\x00\x80\x80\x31\x01\x00\x81\x55"
-     "\x01\x00"
-     "\x82\x5D\x01\x00\x83\x71\x01\x00\x84\x25\x01\x00\x85\x95\x01\x00\x86\x9D"
-     "\x01\x00"
-     "\x87\x8D\x01\x00\x88\xB5\x01\x00\x89\xE5\x01\x00\x8A\xED\x01\x00\x8B\xFD"
-     "\x01\x00"
-     "\x8C\xE5\x01\x00\x8D\x15\x02\x00\x8E\x1D\x02\x00\x8F\x05\x01\x00\x90\x7D"
-     "\x01\x00"
-     "\x91\x0D\x01\x00\x92\x11\x01\x00\x93\x11\x01\x00\x94\x31\x01\x00\x95\x31"
-     "\x01\x00"
-     "\x96\x31\x01\x00\x97\x31\x01\x00\x98\x11\x01\x00\x99\x1D\x01\x00\x9A\xE1"
-     "\x01\x00"
-     "\x9B\xE1\x01\x00\x9C\xC1\x01\x00\x9D\xC1\x01\x00\x9E\xC1\x01\x00\x9F\xC1"
-     "\x01\x00"
-     "\xA0\x49\x01\x00\xA1\x49\x01\x00\xA2\x51\x01\x00\xA3\x51\x01\x00\xA4\x71"
-     "\x01\x00"
-     "\xA5\x71\x01\x00\xA6\x71\x01\x00\xA7\x71\x01\x00\xA8\x51\x01\x00\xA9\x51"
-     "\x01\x00"
-     "\xAA\x51\x01\x00\xAB\x51\x01\x00\xAC\xB1\x03\x00\xAD\xB1\x03\x00\xAE\xB1"
-     "\x03\x00"
-     "\xAF\xB1\x03\x00\xB0\xD1\x03\x00\xB1\xD1\x03\x00\xB2\xD1\x03\x00\xB3\xD1"
-     "\x03\x00"
-     "\xB4\xF1\x03\x00\xB5\xF1\x03\x00\xB6\xF1\x03\x00\xB7\xF1\x03\x00\xB8\xD1"
-     "\x03\x00"
-     "\xB9\xD1\x03\x00\xBA\xD1\x03\x00\xBB\xD1\x03\x00\xBC\xB1\x03\x00\xBD\xB1"
-     "\x03\x00"
-     "\xBE\xB1\x03\x00\xBF\xB1\x03\x00\x50\x1B\x00\x80\x54\x1B\x00\x80\x48\x1A"
-     "\x00\x80"
-     "\x58\x1B\x00\x80\x5C\x1B\x00\x80\x4C\x1A\x00\x80\x50\x1A\x00\x80\x54\x1A"
-     "\x00\x80"
-     "\x5C\x1A\x00\x80\x60\x1A\x00\x80\x64\x1A\x00\x80\x68\x1A\x00\x80\x6C\x1A"
-     "\x00\x80"
-     "\x70\x1A\x00\x80\x74\x1A\x00\x80\x78\x1A\x00\x80\x7C\x1A\x00\x80\x80\x1A"
-     "\x00\x80"
-     "\x84\x1A\x00\x80\x88\x1A\x00\x80\x8C\x1A\x00\x80\x90\x1A\x00\x80\x60\x1B"
-     "\x00\x80"
-     "\x64\x1B\x00\x80\x94\x1A\x00\x80\x98\x1A\x00\x80\x9C\x1A\x00\x80\xF8\x1A"
-     "\x00\x80"
-     "\xFC\x1A\x00\x80\x00\x1B\x00\x80\x04\x1B\x00\x80\x08\x1B\x00\x80\x86\xE9"
-     "\x00\x00"
-     "\x87\xE9\x00\x00\x84\xE9\x00\x00\x85\xE9\x00\x00\x82\xC9\x00\x00\x83\xC9"
-     "\x00\x00"
-     "\x80\xC9\x00\x00\x81\xC9\x00\x00\x8E\x7D\x00\x00\x0C\x1B\x00\x80\x8C\x0D"
-     "\x00\x00"
-     "\x8D\x75\x00\x00\x8A\xC9\x00\x00\x8B\xC9\x00\x00\x88\xC9\x00\x00\x89\xC9"
-     "\x00\x00"
-     "\x96\xA1\x01\x00\x97\xA1\x01\x00\x94\xF9\x01\x00\x95\xF9\x01\x00\x92\xCD"
-     "\x6D\x00"
-     "\x93\x99\x6D\x00\x10\x1B\x00\x80\x14\x1B\x00\x80\x9E\x55\x01\x00\x9F\xB1"
-     "\x6D\x00"
-     "\x9C\x4D\x01\x00\x9D\x5D\x01\x00\x9A\xB1\x6D\x00\x9B\x55\x01\x00\x98\x89"
-     "\x01\x00"
-     "\x99\x89\x01\x00\xE1\xC8\x66\x03\xE1\x70\x4F\x03\xE3\x5C\x1F\x03\xE3\xC8"
-     "\x63\x03"
-     "\xE1\xD4\x3B\x03\xE1\x00\x04\x03\xE3\xC4\x2B\x03\xE3\xC4\x33\x03\x34\x1A"
-     "\x00\x80"
-     "\x38\x1A\x00\x80\xEF\x74\x32\x03\xEF\xE8\x32\x03\xA4\x2B\x00\x80\x5B\x2E"
-     "\x00\x80"
-     "\xEF\x28\x05\x03\xEF\xEC\x67\x03\xE1\x54\xF6\x03\xE1\xCC\xDA\x03\xE3\xFC"
-     "\xE6\x03"
-     "\xE3\x94\xCB\x03\xE1\x84\x8C\x03\xE1\xA8\x90\x03\xE3\x44\xA8\x03\xE3\x84"
-     "\x87\x03"
-     "\xA8\x2B\x00\x80\xC0\x2B\x00\x80\xEF\x98\xA8\x03\xEF\x3C\x8C\x03\xAC\x2B"
-     "\x00\x80"
-     "\x9C\x34\x00\x80\xEF\xBC\xCA\x03\xEF\xAC\x3F\x03\xE1\x74\x0F\x02\xE1\xD0"
-     "\x6B\x02"
-     "\xE3\x2C\x32\x02\xE3\x50\x13\x02\xE1\xEC\xC4\x02\xE1\x20\x3A\x02\xE3\x4C"
-     "\xF2\x02"
-     "\xE3\x40\xDB\x02\xB4\x2B\x00\x80\x9C\x2B\x00\x80\xEF\x3C\xE1\x02\xEF\x70"
-     "\x0B\x02"
-     "\xE6\x2E\x00\x80\x6F\x2E\x00\x80\xEF\xF8\x12\x02\xEF\xFC\x6F\x02\xEF\x04"
-     "\x4F\x03"
-     "\xE1\x8C\xE1\x02\xF0\x19\x00\x80\xE3\x98\xEB\x02\xE1\x84\x9F\x02\xE1\x60"
-     "\x96\x02"
-     "\xE3\x8C\xA3\x02\xE3\x08\x9A\x02\x07\x02\x00\x80\x0D\x02\x00\x80\xEF\x34"
-     "\x8B\x02"
-     "\xEF\xB8\x86\x02\xE3\xA4\x4F\x03\x13\x02\x00\x80\xE1\x5C\x53\x03\xEF\x8C"
-     "\x9A\x02"
-     "\x19\x02\x00\x80\x1F\x02\x00\x80\x25\x02\x00\x80\x2B\x02\x00\x80\x31\x02"
-     "\x00\x80"
-     "\x37\x02\x00\x80\x3D\x02\x00\x80\x43\x02\x00\x80\x49\x02\x00\x80\x4F\x02"
-     "\x00\x80"
-     "\x55\x02\x00\x80\x33\x03\x00\x80\x39\x03\x00\x80\x3F\x03\x00\x80\x45\x03"
-     "\x00\x80"
-     "\x4B\x03\x00\x80\x51\x03\x00\x80\x57\x03\x00\x80\x5D\x03\x00\x80\x75\x03"
-     "\x00\x80"
-     "\x81\x03\x00\x80\x87\x03\x00\x80\x8D\x03\x00\x80\x93\x03\x00\x80\x99\x03"
-     "\x00\x80"
-     "\x7B\x03\x00\x80\x63\x03\x00\x80\x6C\x03\x00\x80\x5B\x02\x00\x80\x73\x02"
-     "\x00\x80"
-     "\x67\x02\x00\x80\xBB\x02\x00\x80\x79\x02\x00\x80\x97\x02\x00\x80\x61\x02"
-     "\x00\x80"
-     "\x91\x02\x00\x80\x6D\x02\x00\x80\xA3\x02\x00\x80\xD3\x02\x00\x80\xF1\x02"
-     "\x00\x80"
-     "\xEB\x02\x00\x80\xDF\x02\x00\x80\x2D\x03\x00\x80\xC1\x02\x00\x80\xCD\x02"
-     "\x00\x80"
-     "\xE5\x02\x00\x80\xD9\x02\x00\x80\x03\x03\x00\x80\xAF\x02\x00\x80\xF7\x02"
-     "\x00\x80"
-     "\x21\x03\x00\x80\x09\x03\x00\x80\x9D\x02\x00\x80\x7F\x02\x00\x80\xB5\x02"
-     "\x00\x80"
-     "\xC7\x02\x00\x80\xFD\x02\x00\x80\x85\x02\x00\x80\x27\x03\x00\x80\xA9\x02"
-     "\x00\x80"
-     "\x0F\x03\x00\x80\x8B\x02\x00\x80\x15\x03\x00\x80\x1B\x03\x00\x80\x80\xF1"
-     "\x01\x00"
-     "\x81\xF1\x01\x00\x82\xF1\x01\x00\x83\xF1\x01\x00\x84\x35\x00\x00\x85\x3D"
-     "\x00\x00"
-     "\x86\xA5\x15\x00\x87\x31\x00\x00\x9E\x24\xDA\x00\x0F\x2E\x00\x80\xF1\x2F"
-     "\x00\x80"
-     "\xD0\x35\x00\x80\x86\x07\x00\x80\x05\x05\x00\x80\x14\x05\x00\x80\x23\x05"
-     "\x00\x80"
-     "\x90\x71\x00\x00\x91\x71\x00\x00\x92\x71\x00\x00\x93\x71\x00\x00\x94\xAD"
-     "\x00\x00"
-     "\x95\x55\x03\x00\x96\x5D\x03\x00\x97\x55\x03\x00\x98\x6D\x03\x00\x99\x75"
-     "\x03\x00"
-     "\x9A\x7D\x03\x00\x9B\x75\x03\x00\x9C\x6D\x03\x00\x9D\x55\x03\x00\x9E\x5D"
-     "\x03\x00"
-     "\x9F\x55\x03\x00\xA0\x49\x06\x00\xA1\x81\x05\x00\xA2\x99\x05\x00\xA3\xCD"
-     "\x05\x00"
-     "\xA4\xC1\x05\x00\xA5\xC1\x05\x00\xA6\x19\x05\x00\xA7\x0D\x05\x00\xA8\x35"
-     "\x05\x00"
-     "\xA9\x61\x05\x00\xAA\x79\x05\x00\xAB\x79\x05\x00\xAC\x91\x06\x00\xAD\x91"
-     "\x06\x00"
-     "\xAE\x95\x03\x00\xAF\x8D\x03\x00\xB0\xF5\x03\x00\xB1\xFD\x03\x00\xB2\xF5"
-     "\x03\x00"
-     "\xB3\x8D\x03\x00\xB4\x95\x03\x00\xB5\x9D\x03\x00\xB6\x95\x03\x00\xB7\x8D"
-     "\x03\x00"
-     "\xB8\xB5\x03\x00\xB9\xBD\x03\x00\xBA\xB5\x03\x00\xBB\x8D\x03\x00\xBC\x95"
-     "\x03\x00"
-     "\xBD\x9D\x03\x00\xBE\x95\x03\x00\x32\x05\x00\x80\x41\x05\x00\x80\x4A\x05"
-     "\x00\x80"
-     "\x53\x05\x00\x80\x5C\x05\x00\x80\x65\x05\x00\x80\x91\x05\x00\x80\x9A\x05"
-     "\x00\x80"
-     "\xA3\x05\x00\x80\xA6\x05\x00\x80\xA9\x05\x00\x80\xAC\x05\x00\x80\x64\x3B"
-     "\x00\x80"
-     "\xC4\x4D\x00\x80\xF4\x56\x00\x80\x54\x69\x00\x80\x84\x72\x00\x80\xB4\x7B"
-     "\x00\x80"
-     "\x14\x8E\x00\x80\x74\xA0\x00\x80\xA4\xA9\x00\x80\x0B\xBC\x00\x80\x42\xC5"
-     "\x00\x80"
-     "\x72\xCE\x00\x80\xA2\xD7\x00\x80\xD2\xE0\x00\x80\x4F\xBC\x00\x80\x58\xAF"
-     "\x00\x80"
-     "\x24\xA6\x00\x80\xF8\x2B\x00\x80\x8F\x2C\x00\x80\xA0\x2B\x00\x80\xE4\x2B"
-     "\x00\x80"
-     "\x47\x32\x00\x80\x07\x32\x00\x80\x54\x35\x00\x80\xDB\x2D\x00\x80\x0C\x34"
-     "\x00\x80"
-     "\x90\x38\x00\x80\x68\x35\x00\x80\x23\x2D\x00\x80\x98\x35\x00\x80\xA4\x34"
-     "\x00\x80"
-     "\x18\x39\x00\x80\x5F\x2C\x00\x80\x8B\x2D\x00\x80\x00\x33\x00\x80\x30\x3A"
-     "\x00\x80"
-     "\x23\x2C\x00\x80\xB3\x2C\x00\x80\xA4\x32\x00\x80\xC2\x2F\x00\x80\xCF\x2D"
-     "\x00\x80"
-     "\x5F\x2D\x00\x80\x3E\x2F\x00\x80\x32\x2F\x00\x80\xF4\x34\x00\x80\x1F\x2C"
-     "\x00\x80"
-     "\xA4\x38\x00\x80\x7B\x2D\x00\x80\xD3\x2E\x00\x80\xAF\x05\x00\x80\xB2\x05"
-     "\x00\x80"
-     "\xB5\x05\x00\x80\xB8\x05\x00\x80\x88\x15\x03\x00\x89\x1D\x03\x00\x8A\x15"
-     "\x03\x00"
-     "\x8B\x2D\x03\x00\x8C\x35\x03\x00\x8D\x3D\x03\x00\x8E\x35\x03\x00\x8F\x2D"
-     "\x03\x00"
-     "\x80\xC1\x04\x00\x81\xC1\x04\x00\x82\xDD\x04\x00\x83\x2D\x03\x00\x84\x35"
-     "\x03\x00"
-     "\x85\x3D\x03\x00\x86\x35\x03\x00\x87\x2D\x03\x00\x98\x15\x03\x00\x99\x1D"
-     "\x03\x00"
-     "\x9A\x15\x03\x00\x9B\x2D\x03\x00\x9C\x35\x03\x00\x9D\x3D\x03\x00\x9E\x35"
-     "\x03\x00"
-     "\xE4\x05\x00\x80\x90\x55\x03\x00\x91\x5D\x03\x00\x92\x55\x03\x00\x93\x2D"
-     "\x03\x00"
-     "\x94\x35\x03\x00\x95\x3D\x03\x00\x96\x35\x03\x00\x97\x2D\x03\x00\xA8\x11"
-     "\x03\x00"
-     "\xA9\x11\x03\x00\xAA\x11\x03\x00\xAB\x11\x03\x00\xAC\x31\x03\x00\xAD\x31"
-     "\x03\x00"
-     "\xAE\x31\x03\x00\xAF\x31\x03\x00\xA0\xD1\x03\x00\xA1\xD1\x03\x00\xA2\xD1"
-     "\x03\x00"
-     "\xA3\xD1\x03\x00\xA4\x31\x03\x00\xA5\x31\x03\x00\xA6\x31\x03\x00\xA7\x31"
-     "\x03\x00"
-     "\xB8\x11\x03\x00\xB9\x11\x03\x00\xBA\x11\x03\x00\xBB\x11\x03\x00\xBC\x31"
-     "\x03\x00"
-     "\xBD\x31\x03\x00\xBE\x31\x03\x00\xBF\x31\x03\x00\xB0\x51\x03\x00\xB1\x51"
-     "\x03\x00"
-     "\xB2\x51\x03\x00\xB3\x51\x03\x00\xB4\x31\x03\x00\xB5\x31\x03\x00\xB6\x31"
-     "\x03\x00"
-     "\xB7\x31\x03\x00\xE7\x05\x00\x80\xEA\x05\x00\x80\xED\x05\x00\x80\xF0\x05"
-     "\x00\x80"
-     "\xF3\x05\x00\x80\xF6\x05\x00\x80\xF9\x05\x00\x80\xFC\x05\x00\x80\xFF\x05"
-     "\x00\x80"
-     "\x1F\x06\x00\x80\xE0\x04\x00\x80\x7F\x05\x00\x80\xD2\x05\x00\x80\x0D\x06"
-     "\x00\x80"
-     "\x33\x06\x00\x80\x4A\x06\x00\x80\x67\x06\x00\x80\x7E\x06\x00\x80\x95\x06"
-     "\x00\x80"
-     "\xDC\x03\x00\x80\xF5\x03\x00\x80\x0E\x04\x00\x80\x11\x07\x00\x80\x43\x08"
-     "\x00\x80"
-     "\x40\x08\x00\x80\x3E\x07\x00\x80\x38\x25\x00\x80\x71\x25\x00\x80\xA2\x25"
-     "\x00\x80"
-     "\xC7\x25\x00\x80\xB8\x27\x00\x80\xC3\x27\x00\x80\xC7\x27\x00\x80\xCB\x27"
-     "\x00\x80"
-     "\xCF\x27\x00\x80\x2E\x29\x00\x80\x6D\x29\x00\x80\x95\x29\x00\x80\x99\x29"
-     "\x00\x80"
-     "\xBE\x29\x00\x80\xC6\x29\x00\x80\xE2\x29\x00\x80\xF4\x29\x00\x80\xF8\x29"
-     "\x00\x80"
-     "\xFC\x29\x00\x80\x14\x2A\x00\x80\x2F\x2A\x00\x80\x4A\x2A\x00\x80\x18\x2C"
-     "\x00\x80"
-     "\xF0\x80\x38\x00\x3C\x25\x00\x80\x49\x25\x00\x80\x56\x25\x00\x80\x63\x25"
-     "\x00\x80"
-     "\x75\x25\x00\x80\x82\x25\x00\x80\x94\x25\x00\x80\xA6\x25\x00\x80\xB6\x25"
-     "\x00\x80"
-     "\xCB\x25\x00\x80\xD5\x25\x00\x80\xE3\x25\x00\x80\xED\x25\x00\x80\xFA\x25"
-     "\x00\x80"
-     "\x80\x29\x01\x00\x81\x29\x01\x00\x82\x39\x01\x00\x83\x39\x01\x00\x84\x29"
-     "\x01\x00"
-     "\x85\x29\x01\x00\x86\x59\x01\x00\x87\x59\x01\x00\x88\x69\x01\x00\x89\x69"
-     "\x01\x00"
-     "\x8A\x79\x01\x00\x8B\x79\x01\x00\x8C\x69\x01\x00\x8D\x69\x01\x00\x8E\x59"
-     "\x01\x00"
-     "\x8F\x59\x01\x00\x90\x29\x01\x00\x91\x29\x01\x00\x92\x01\x04\x00\x93\x3D"
-     "\x01\x00"
-     "\x94\x25\x01\x00\x95\x55\x05\x00\x96\x5D\x05\x00\x97\x55\x05\x00\x98\x6D"
-     "\x05\x00"
-     "\x99\x75\x05\x00\x9A\x7D\x05\x00\x9B\x75\x05\x00\x9C\x6D\x05\x00\x9D\x35"
-     "\x05\x00"
-     "\x9E\x3D\x05\x00\x9F\x35\x05\x00\xA0\xCD\x05\x00\xA1\xD5\x05\x00\xA2\xDD"
-     "\x05\x00"
-     "\xA3\xD5\x05\x00\xA4\xE9\x01\x00\xA5\xE9\x01\x00\xA6\x19\x01\x00\xA7\x19"
-     "\x01\x00"
-     "\xA8\x29\x01\x00\xA9\xF9\x05\x00\xAA\x29\x05\x00\xAB\x29\x05\x00\xAC\x39"
-     "\x05\x00"
-     "\xAD\x39\x05\x00\xAE\x35\x01\x00\xAF\x2D\x01\x00\xB0\x55\x01\x00\xB1\x5D"
-     "\x01\x00"
-     "\xB2\x55\x01\x00\xB3\x6D\x01\x00\xB4\x75\x01\x00\xB5\x7D\x01\x00\xB6\x75"
-     "\x01\x00"
-     "\xB7\x6D\x01\x00\xB8\x55\x01\x00\xB9\x5D\x01\x00\xBA\x55\x01\x00\xBB\x7D"
-     "\x05\x00"
-     "\xBC\x65\x05\x00\xBD\x6D\x05\x00\xBE\x65\x05\x00\xBF\x3D\x01\x00\xB1\xED"
-     "\x06\x00"
-     "\xB0\xED\x06\x00\xB3\xFD\x06\x00\xB2\xFD\x06\x00\xB5\xED\x06\x00\xB4\xED"
-     "\x06\x00"
-     "\xB7\x1D\x07\x00\xB6\x1D\x07\x00\xB9\x2D\x07\x00\xB8\x2D\x07\x00\xBB\x3D"
-     "\x07\x00"
-     "\xBA\x3D\x07\x00\xBD\x2D\x07\x00\xBC\x2D\x07\x00\xBF\x1D\x07\x00\xBE\x1D"
-     "\x07\x00"
-     "\xA1\xED\x06\x00\xA0\xED\x06\x00\xA3\xFD\x06\x00\xA2\xFD\x06\x00\xA5\xED"
-     "\x06\x00"
-     "\xA4\xED\x06\x00\xA7\x9D\x06\x00\xA6\x9D\x06\x00\xA9\xAD\x06\x00\xA8\xAD"
-     "\x06\x00"
-     "\xAB\xBD\x06\x00\xAA\xBD\x06\x00\xAD\xAD\x06\x00\xAC\xAD\x06\x00\xAF\x9D"
-     "\x06\x00"
-     "\xAE\x9D\x06\x00\x91\xAD\x06\x00\x90\xAD\x06\x00\x93\xBD\x06\x00\x92\xBD"
-     "\x06\x00"
-     "\x95\xAD\x06\x00\x94\xAD\x06\x00\x97\xFD\x06\x00\x96\xFD\x06\x00\x99\xCD"
-     "\x06\x00"
-     "\x98\xCD\x06\x00\x9B\xDD\x06\x00\x9A\xDD\x06\x00\x9D\xCD\x06\x00\x9C\xCD"
-     "\x06\x00"
-     "\x9F\x1D\x06\x00\x9E\x1D\x06\x00\x81\x0D\x02\x00\x80\x0D\x02\x00\x83\x1D"
-     "\x02\x00"
-     "\x82\x1D\x02\x00\x85\x0D\x02\x00\x84\x0D\x02\x00\x87\x3D\x02\x00\x86\x3D"
-     "\x02\x00"
-     "\x89\x0D\x02\x00\x88\x0D\x02\x00\x8B\x1D\x02\x00\x8A\x1D\x02\x00\x8D\x0D"
-     "\x02\x00"
-     "\x8C\x0D\x02\x00\x8F\xDD\x06\x00\x8E\xDD\x06\x00\x8A\xE9\x1B\x00\x8B\x9D"
-     "\x1A\x00"
-     "\x88\x75\x0D\x00\x89\xAD\x64\x00\x8E\xBD\x1C\x00\x8F\x19\x1B\x00\x8C\x8D"
-     "\x1A\x00"
-     "\x8D\xF1\x1B\x00\x82\x11\x08\x00\x83\x11\x08\x00\x80\x35\x64\x00\x81\x61"
-     "\x64\x00"
-     "\x86\x49\x0D\x00\x87\x51\x64\x00\x84\x49\x0D\x00\x85\x49\x0D\x00\x9A\xE5"
-     "\x0D\x00"
-     "\x9B\x75\x1B\x00\x98\xD5\x0D\x00\x99\xED\x0D\x00\x9E\xE1\x0D\x00\x9F\xE1"
-     "\x0D\x00"
-     "\x9C\xF1\x0D\x00\x9D\xF1\x0D\x00\x92\xE5\x0D\x00\x93\xED\x0D\x00\x90\x89"
-     "\x1B\x00"
-     "\x91\xED\x0D\x00\x96\xE5\x0D\x00\x97\xED\x0D\x00\x94\xF5\x0D\x00\x95\xED"
-     "\x0D\x00"
-     "\xAA\x55\x0D\x00\xAB\x5D\x0D\x00\xA8\x45\x0D\x00\xA9\x5D\x0D\x00\xAE\x75"
-     "\x0D\x00"
-     "\xAF\x7D\x0D\x00\xAC\x45\x0D\x00\xAD\x7D\x0D\x00\xA2\x61\x0D\x00\xA3\x61"
-     "\x0D\x00"
-     "\xA0\x11\x0D\x00\xA1\x11\x0D\x00\xA6\xC1\x08\x00\xA7\x7D\x0D\x00\xA4\xD9"
-     "\x08\x00"
-     "\xA5\xD9\x08\x00\xBA\xD1\x03\x00\xBB\xD1\x03\x00\xB8\xC1\x03\x00\xB9\xC1"
-     "\x03\x00"
-     "\xBE\xF1\x03\x00\xBF\xF1\x03\x00\xBC\xC1\x03\x00\xBD\xC1\x03\x00\xB2\x15"
-     "\x0D\x00"
-     "\xB3\x1D\x0D\x00\xB0\x05\x0D\x00\xB1\x1D\x0D\x00\xB6\xF1\x03\x00\xB7\xF1"
-     "\x03\x00"
-     "\xB4\x05\x0D\x00\xB5\xA1\x0C\x00\xDF\x25\x00\x80\x1F\x26\x00\x80\x8C\x26"
-     "\x00\x80"
-     "\xCB\x26\x00\x80\xDD\x26\x00\x80\x07\x27\x00\x80\x2C\x27\x00\x80\x41\x27"
-     "\x00\x80"
-     "\x0B\x26\x00\x80\x15\x26\x00\x80\x6E\x26\x00\x80\x75\x26\x00\x80\x23\x26"
-     "\x00\x80"
-     "\x7F\x26\x00\x80\x30\x26\x00\x80\x37\x26\x00\x80\x49\x27\x00\x80\x65\x27"
-     "\x00\x80"
-     "\x6D\x27\x00\x80\x7F\x27\x00\x80\x8B\x27\x00\x80\x93\x27\x00\x80\xA2\x27"
-     "\x00\x80"
-     "\xDD\x27\x00\x80\x47\x26\x00\x80\x5A\x26\x00\x80\xB1\x26\x00\x80\xBB\x26"
-     "\x00\x80"
-     "\x90\x26\x00\x80\xA0\x26\x00\x80\xCF\x26\x00\x80\xEF\x26\x00\x80\x1A\x28"
-     "\x00\x80"
-     "\x4E\x28\x00\x80\x91\x28\x00\x80\xE6\x28\x00\x80\x4E\x2A\x00\x80\x5C\x2A"
-     "\x00\x80"
-     "\x60\x2A\x00\x80\x64\x2A\x00\x80\xF9\x26\x00\x80\x0B\x27\x00\x80\x18\x27"
-     "\x00\x80"
-     "\x30\x27\x00\x80\x4D\x27\x00\x80\x57\x27\x00\x80\x75\x27\x00\x80\x9B\x27"
-     "\x00\x80"
-     "\xB1\x27\x00\x80\xA6\x27\x00\x80\xBC\x27\x00\x80\xD3\x27\x00\x80\xE1\x27"
-     "\x00\x80"
-     "\x00\x28\x00\x80\x10\x28\x00\x80\xEF\x27\x00\x80\x2D\x28\x00\x80\x3D\x28"
-     "\x00\x80"
-     "\x47\x28\x00\x80\x22\x28\x00\x80\x52\x28\x00\x80\x62\x28\x00\x80\x7D\x28"
-     "\x00\x80"
-     "\x6F\x28\x00\x80\x95\x28\x00\x80\xC1\x28\x00\x80\xC8\x28\x00\x80\xA5\x28"
-     "\x00\x80"
-     "\xD2\x28\x00\x80\xDC\x28\x00\x80\xB3\x28\x00\x80\x17\x29\x00\x80\x09\x29"
-     "\x00\x80"
-     "\xEA\x28\x00\x80\x24\x29\x00\x80\xF1\x28\x00\x80\xFB\x28\x00\x80\x32\x29"
-     "\x00\x80"
-     "\x3F\x29\x00\x80\x49\x29\x00\x80\x53\x29\x00\x80\x5D\x29\x00\x80\x90\x25"
-     "\x00\x80"
-     "\x71\x29\x00\x80\x7E\x29\x00\x80\x85\x29\x00\x80\x9D\x29\x00\x80\xA4\x29"
-     "\x00\x80"
-     "\xB1\x29\x00\x80\xCA\x29\x00\x80\xD4\x29\x00\x80\xE6\x29\x00\x80\x00\x2A"
-     "\x00\x80"
-     "\x0D\x2A\x00\x80\x18\x2A\x00\x80\x22\x2A\x00\x80\x33\x2A\x00\x80\x3A\x2A"
-     "\x00\x80"
-     "\x52\x2A\x00\x80\xC2\x03\x00\x80\xE5\x04\x00\x80\x84\x05\x00\x80\xD7\x05"
-     "\x00\x80"
-     "\x12\x06\x00\x80\x38\x06\x00\x80\x4F\x06\x00\x80\x6C\x06\x00\x80\x83\x06"
-     "\x00\x80"
-     "\x9A\x06\x00\x80\xE2\x03\x00\x80\xFB\x03\x00\x80\x14\x04\x00\x80\x27\x04"
-     "\x00\x80"
-     "\x3A\x04\x00\x80\x4D\x04\x00\x80\x60\x04\x00\x80\x73\x04\x00\x80\x86\x04"
-     "\x00\x80"
-     "\x99\x04\x00\x80\xFF\x04\x00\x80\x0E\x05\x00\x80\x1D\x05\x00\x80\x2C\x05"
-     "\x00\x80"
-     "\x3B\x05\x00\x80\x65\x08\x00\x80\x26\x08\x00\x80\xC0\x06\x00\x80\xFE\x07"
-     "\x00\x80"
-     "\x1F\x09\x00\x80\x35\x09\x00\x80\x2C\x08\x00\x80\x2F\x08\x00\x80\x33\x08"
-     "\x00\x80"
-     "\x23\x07\x00\x80\xE2\x2F\x00\x80\xD0\x31\x00\x80\xDF\x2E\x00\x80\xC1\x31"
-     "\x00\x80"
-     "\x4B\x32\x00\x80\x29\x09\x00\x80\x08\x09\x00\x80\x64\x0E\x00\x80\xD1\x08"
-     "\x00\x80"
-     "\x85\x08\x00\x80\x30\x07\x00\x80\x4B\x07\x00\x80\xFB\x06\x00\x80\x0C\x08"
-     "\x00\x80"
-     "\x96\x08\x00\x80\x2C\x09\x00\x80\x0B\x09\x00\x80\x68\x0E\x00\x80\x74\x0E"
-     "\x00\x80"
-     "\xDD\x08\x00\x80\x9B\x08\x00\x80\x14\x07\x00\x80\x88\x08\x00\x80\x54\x07"
-     "\x00\x80"
-     "\xFE\x06\x00\x80\xA8\x07\x00\x80\x78\x0E\x00\x80\xE4\x08\x00\x80\x29\x08"
-     "\x00\x80"
-     "\x9E\x08\x00\x80\x5A\x08\x00\x80\x17\x09\x00\x80\x7C\x0E\x00\x80\xE7\x08"
-     "\x00\x80"
-     "\x16\x08\x00\x80\xA1\x08\x00\x80\xEA\x08\x00\x80\x19\x08\x00\x80\xCB\x08"
-     "\x00\x80"
-     "\xA4\x08\x00\x80\xEE\x08\x00\x80\x1D\x08\x00\x80\xCE\x08\x00\x80\xE3\x10"
-     "\x17\x00"
-     "\xA8\x08\x00\x80\xF9\x08\x00\x80\xFF\x08\x00\x80\x8A\x07\x00\x80\x8C\x08"
-     "\x00\x80"
-     "\x58\x07\x00\x80\x02\x07\x00\x80\x3F\x09\x00\x80\x43\x09\x00\x80\x4B\x09"
-     "\x00\x80"
-     "\x38\x09\x00\x80\xEF\xD8\x17\x00\x1A\x09\x00\x80\x80\x0E\x00\x80\xE3\xA4"
-     "\x13\x00"
-     "\xE3\xA8\x13\x00\xE3\x88\x13\x00\xE3\x30\x13\x00\xE3\x7C\x12\x00\xE3\x48"
-     "\x12\x00"
-     "\xE3\x38\x12\x00\xE3\xE8\x13\x00\xEF\xAC\x12\x00\xEF\x80\x12\x00\xEF\x94"
-     "\x12\x00"
-     "\xEF\xC8\x12\x00\xEF\x1C\x13\x00\xEF\x04\x13\x00\xEF\x5C\x13\x00\xEF\xB0"
-     "\x13\x00"
-     "\xF6\x08\x00\x80\x2F\x09\x00\x80\x0E\x09\x00\x80\x6C\x0E\x00\x80\xD4\x08"
-     "\x00\x80"
-     "\x90\x08\x00\x80\x5F\x07\x00\x80\x32\x09\x00\x80\x11\x09\x00\x80\x70\x0E"
-     "\x00\x80"
-     "\xD7\x08\x00\x80\x93\x08\x00\x80\x62\x07\x00\x80\xAF\x08\x00\x80\x65\x07"
-     "\x00\x80"
-     "\xEB\x07\x00\x80\xE3\x1C\x11\x00\xE3\xE4\x11\x00\xE3\x7C\x16\x00\xE3\x7C"
-     "\x16\x00"
-     "\xE3\xB8\x10\x00\xE3\x8C\x10\x00\xE3\x7C\x11\x00\xE3\x3C\x11\x00\xEF\x68"
-     "\x10\x00"
-     "\xEF\x40\x10\x00\xEF\x54\x10\x00\xEF\x98\x11\x00\xEF\xE8\x11\x00\xEF\x38"
-     "\x11\x00"
-     "\xEF\x14\x11\x00\xEF\xAC\x16\x00\xE3\x64\x10\x00\xE3\xB0\x10\x00\xE3\x9C"
-     "\x10\x00"
-     "\xE3\xF0\x10\x00\xE3\x20\x10\x00\xE3\x38\x10\x00\xE3\x08\x10\x00\xE3\x6C"
-     "\x10\x00"
-     "\xEF\x34\x13\x00\xEF\x04\x13\x00\xEF\xFC\x10\x00\xEF\xD8\x10\x00\xEF\xB4"
-     "\x10\x00"
-     "\xEF\x9C\x10\x00\xEF\x80\x10\x00\xEF\x54\x10\x00\x80\x2D\x3E\x00\x81\x35"
-     "\x3E\x00"
-     "\x82\x3D\x3E\x00\x83\x35\x3E\x00\x84\x2D\x3E\x00\x85\x55\x3E\x00\x86\x5D"
-     "\x3E\x00"
-     "\x87\x55\x3E\x00\x88\x6D\x3E\x00\x89\x75\x3E\x00\x8A\x7D\x3E\x00\x8B\x75"
-     "\x3E\x00"
-     "\x8C\x6D\x3E\x00\x8D\x55\x3E\x00\x8E\x5D\x3E\x00\x8F\x55\x3E\x00\x90\x2D"
-     "\x3E\x00"
-     "\x91\x35\x3E\x00\x92\x3D\x3E\x00\x93\x35\x3E\x00\x94\x2D\x3E\x00\x95\xD5"
-     "\x3E\x00"
-     "\x96\xDD\x3E\x00\x97\xD5\x3E\x00\x98\xED\x3E\x00\x99\xF5\x3E\x00\x9A\xFD"
-     "\x3E\x00"
-     "\x9B\xF5\x3E\x00\x9C\xED\x3E\x00\x9D\xD5\x3E\x00\x9E\xDD\x3E\x00\x9F\xD5"
-     "\x3E\x00"
-     "\xA0\x2D\x3E\x00\xA1\x35\x3E\x00\xA2\x3D\x3E\x00\xA3\x35\x3E\x00\xA4\x2D"
-     "\x3E\x00"
-     "\xA5\x55\x3E\x00\xA6\x51\x3E\x00\xA7\x51\x3E\x00\xA8\x71\x3E\x00\xA9\x71"
-     "\x3E\x00"
-     "\xAA\x71\x3E\x00\xAB\x71\x3E\x00\xAC\x51\x3E\x00\xAD\x51\x3E\x00\xAE\x51"
-     "\x3E\x00"
-     "\xAF\x51\x3E\x00\xB0\x31\x3E\x00\xB1\x3D\x3E\x00\xB2\x35\x3E\x00\xB3\x8D"
-     "\x3F\x00"
-     "\xB4\x95\x3F\x00\xB5\x9D\x3F\x00\xB6\x95\x3F\x00\xB7\x8D\x3F\x00\xB8\xB5"
-     "\x3F\x00"
-     "\xB9\xBD\x3F\x00\xBA\xB5\x3F\x00\xBB\x8D\x3F\x00\xBC\x95\x3F\x00\xBD\x9D"
-     "\x3F\x00"
-     "\xBE\x95\x3F\x00\xBF\x8D\x3F\x00\x80\x75\x3E\x00\x81\x7D\x3E\x00\x82\x75"
-     "\x3E\x00"
-     "\x83\x09\x03\x00\x84\x19\x03\x00\x85\x19\x03\x00\x86\x09\x03\x00\x87\x09"
-     "\x03\x00"
-     "\x88\x39\x03\x00\x89\x39\x03\x00\x8A\x09\x03\x00\x8B\x09\x03\x00\x8C\x19"
-     "\x03\x00"
-     "\x8D\x19\x03\x00\x8E\x09\x03\x00\x8F\x09\x03\x00\x90\x79\x03\x00\x91\x79"
-     "\x03\x00"
-     "\x92\x09\x03\x00\x93\x09\x03\x00\x94\x19\x03\x00\x95\x19\x03\x00\x96\x09"
-     "\x03\x00"
-     "\x97\x09\x03\x00\x98\x39\x03\x00\x99\x39\x03\x00\x9A\x09\x03\x00\x9B\x09"
-     "\x03\x00"
-     "\x9C\x19\x03\x00\x9D\x19\x03\x00\x9E\x09\x03\x00\x9F\x09\x03\x00\xA0\xF9"
-     "\x03\x00"
-     "\xA1\xF9\x03\x00\xA2\x09\x03\x00\xA3\x09\x03\x00\xA4\x19\x03\x00\xA5\x19"
-     "\x03\x00"
-     "\xA6\x09\x03\x00\xA7\x09\x03\x00\xA8\x39\x03\x00\xA9\x39\x03\x00\xAA\x09"
-     "\x03\x00"
-     "\xAB\x09\x03\x00\xAC\x19\x03\x00\xAD\x19\x03\x00\xAE\x09\x03\x00\xAF\x09"
-     "\x03\x00"
-     "\xB0\x79\x03\x00\xB1\x79\x03\x00\xB2\x09\x03\x00\xB3\x09\x03\x00\xB4\x19"
-     "\x03\x00"
-     "\xB5\x19\x03\x00\xB6\x09\x03\x00\xB7\x09\x03\x00\xB8\x39\x03\x00\xB9\x39"
-     "\x03\x00"
-     "\xBA\x09\x03\x00\xBB\x09\x03\x00\xBC\x19\x03\x00\xBD\x19\x03\x00\xBE\x09"
-     "\x03\x00"
-     "\xBF\x09\x03\x00\x91\x08\x0B\x00\xCB\x06\x00\x80\x10\x08\x00\x80\x13\x08"
-     "\x00\x80"
-     "\xD7\x06\x00\x80\xD3\x06\x00\x80\x23\x08\x00\x80\x06\x07\x00\x80\x69\x08"
-     "\x00\x80"
-     "\x0B\x07\x00\x80\x78\x08\x00\x80\x33\x07\x00\x80\x36\x07\x00\x80\xAC\x08"
-     "\x00\x80"
-     "\xB8\x08\x00\x80\xBB\x08\x00\x80\xBE\x08\x00\x80\xC2\x08\x00\x80\xDA\x08"
-     "\x00\x80"
-     "\xE0\x08\x00\x80\xF2\x08\x00\x80\x78\x07\x00\x80\x24\x09\x00\x80\x7B\x07"
-     "\x00\x80"
-     "\x83\x07\x00\x80\x56\x09\x00\x80\x9F\x07\x00\x80\xCD\x07\x00\x80\xBF\x07"
-     "\x00\x80"
-     "\xC3\x06\x00\x80\xDB\x04\x00\x80\x7A\x05\x00\x80\xCD\x05\x00\x80\x08\x06"
-     "\x00\x80"
-     "\x2E\x06\x00\x80\x45\x06\x00\x80\x62\x06\x00\x80\x79\x06\x00\x80\x90\x06"
-     "\x00\x80"
-     "\xD6\x03\x00\x80\xEF\x03\x00\x80\x08\x04\x00\x80\x21\x04\x00\x80\x34\x04"
-     "\x00\x80"
-     "\x47\x04\x00\x80\x5A\x04\x00\x80\x6D\x04\x00\x80\x80\x04\x00\x80\x93\x04"
-     "\x00\x80"
-     "\xF9\x04\x00\x80\x08\x05\x00\x80\x17\x05\x00\x80\x26\x05\x00\x80\x35\x05"
-     "\x00\x80"
-     "\x44\x05\x00\x80\x4D\x05\x00\x80\x56\x05\x00\x80\x5F\x05\x00\x80\x68\x05"
-     "\x00\x80"
-     "\x94\x05\x00\x80\x9D\x05\x00\x80\x5F\x08\x00\x80\x23\x0F\x00\x80\x29\x0F"
-     "\x00\x80"
-     "\x8D\x61\x3E\x00\x8C\x59\x3E\x00\x8F\x61\x3E\x00\x8E\x69\x3E\x00\x89\x41"
-     "\x3E\x00"
-     "\x88\x59\x3E\x00\x8B\x41\x3E\x00\x8A\x49\x3E\x00\x85\x1D\x3E\x00\x84\x1D"
-     "\x3E\x00"
-     "\x87\x6D\x3E\x00\x86\x6D\x3E\x00\x81\x1D\x3E\x00\x80\x1D\x3E\x00\x83\x0D"
-     "\x3E\x00"
-     "\x82\x0D\x3E\x00\x9D\xE1\x3E\x00\x9C\xD9\x3E\x00\x9F\xE1\x3E\x00\x9E\xE9"
-     "\x3E\x00"
-     "\x99\xC1\x3E\x00\x98\xD9\x3E\x00\x9B\xC1\x3E\x00\x9A\xC9\x3E\x00\x95\xE1"
-     "\x3E\x00"
-     "\x94\x19\x3E\x00\x97\xE1\x3E\x00\x96\xE9\x3E\x00\x91\x01\x3E\x00\x90\x19"
-     "\x3E\x00"
-     "\x93\x01\x3E\x00\x92\x09\x3E\x00\xAD\x61\x3E\x00\xAC\x59\x3E\x00\xAF\x61"
-     "\x3E\x00"
-     "\xAE\x69\x3E\x00\xA9\x41\x3E\x00\xA8\x59\x3E\x00\xAB\x41\x3E\x00\xAA\x49"
-     "\x3E\x00"
-     "\xA5\x61\x3E\x00\xA4\x19\x3E\x00\xA7\x61\x3E\x00\xA6\x69\x3E\x00\xA1\x01"
-     "\x3E\x00"
-     "\xA0\x19\x3E\x00\xA3\x01\x3E\x00\xA2\x09\x3E\x00\xBD\xE1\x3D\x00\xBC\xD9"
-     "\x3D\x00"
-     "\xBF\xE1\x3D\x00\xBE\xE9\x3D\x00\xB9\xC1\x3D\x00\xB8\xD9\x3D\x00\xBB\xC1"
-     "\x3D\x00"
-     "\xBA\xC9\x3D\x00\xB5\xE1\x3D\x00\xB4\x19\x3E\x00\xB7\xE1\x3D\x00\xB6\xE9"
-     "\x3D\x00"
-     "\xB1\x01\x3E\x00\xB0\x19\x3E\x00\xB3\x01\x3E\x00\xB2\x09\x3E\x00\x4C\x3B"
-     "\x00\x80"
-     "\xB6\x0A\x00\x80\xB2\xAD\x51\x00\xB3\xE5\x51\x00\xB4\xD5\x51\x00\x1F\x0B"
-     "\x00\x80"
-     "\x48\x3B\x00\x80\x54\x3B\x00\x80\xB8\x25\x00\x00\xB9\x31\x00\x00\x11\x0D"
-     "\x00\x80"
-     "\x58\x3B\x00\x80\x9C\x29\x00\x00\x9D\x29\x00\x00\x9E\x11\x00\x00\x9F\x11"
-     "\x00\x00"
-     "\x68\x0D\x00\x80\x80\x38\x00\x80\xF7\x31\x00\x80\xA8\x38\x00\x80\x60\x33"
-     "\x00\x80"
-     "\xC4\x2B\x00\x80\xBF\x2D\x00\x80\xF2\x2E\x00\x80\x23\x2E\x00\x80\xEE\x2E"
-     "\x00\x80"
-     "\xDD\x30\x00\x80\xD8\x34\x00\x80\x68\x36\x00\x80\xA0\x37\x00\x80\xDC\x37"
-     "\x00\x80"
-     "\x10\x38\x00\x80\xB8\xC0\xE5\x00\xB9\x00\xE0\x00\xBA\x88\xE3\x00\xBB\x74"
-     "\xE2\x00"
-     "\xBC\xAC\xEE\x00\xBD\x94\xEF\x00\xBE\x98\xEC\x00\xBF\x48\xE1\x00\xB0\x00"
-     "\x1C\x00"
-     "\xB1\xA4\x1F\x00\xB2\xA8\x1D\x00\xB3\xAC\x1A\x00\xB4\xB0\x18\x00\xB5\x74"
-     "\x19\x00"
-     "\xB6\x00\xE4\x00\xB7\xBC\xE7\x00\xA8\xC0\x0A\x00\xA9\xD0\x09\x00\xAA\x00"
-     "\x14\x00"
-     "\xAB\xCC\x17\x00\xAC\xD0\x15\x00\xAD\x00\x10\x00\xAE\x08\x12\x00\xAF\x90"
-     "\x11\x00"
-     "\x24\x39\x00\x80\xA9\xA5\x00\x00\x7C\x32\x00\x80\x08\x33\x00\x80\xA4\xF8"
-     "\x00\x00"
-     "\xA5\xD4\x01\x00\xA6\xF8\x0F\x00\xA7\x00\x08\x00\xB4\x33\x00\x80\x4C\x34"
-     "\x00\x80"
-     "\x68\x37\x00\x80\x50\x39\x00\x80\x7C\x3A\x00\x80\xE8\x2B\x00\x80\x93\x2D"
-     "\x00\x80"
-     "\x90\x32\x00\x80\x9D\x10\x5C\x00\xD8\x33\x00\x80\x9F\xE8\x03\x00\x9C\x37"
-     "\x00\x80"
-     "\xCC\x3A\x00\x80\x9A\x2F\x00\x80\x80\x33\x00\x80\x9A\xE4\x28\x00\x8C\x37"
-     "\x00\x80"
-     "\x20\x38\x00\x80\x11\x31\x00\x80\xC4\x36\x00\x80\xE8\x37\x00\x80\x0D\x30"
-     "\x00\x80"
-     "\x0B\x32\x00\x80\x20\x33\x00\x80\x24\x34\x00\x80\x30\x39\x00\x80\x2F\x2C"
-     "\x00\x80"
-     "\xC7\x2C\x00\x80\x33\x2D\x00\x80\x79\x31\x00\x80\x88\x32\x00\x80\xD4\x33"
-     "\x00\x80"
-     "\xF8\x34\x00\x80\x94\x37\x00\x80\xAC\x37\x00\x80\xD0\x38\x00\x80\xF8\x39"
-     "\x00\x80"
-     "\xAC\x3A\x00\x80\xC8\x3A\x00\x80\x34\x35\x00\x80\x74\x35\x00\x80\x2C\x36"
-     "\x00\x80"
-     "\x4C\x37\x00\x80\x84\x39\x00\x80\x54\x38\x00\x80\xA7\x2E\x00\x80\x29\x30"
-     "\x00\x80"
-     "\xE4\x35\x00\x80\xA4\x36\x00\x80\x08\x34\x00\x80\x3C\x35\x00\x80\xA0\x38"
-     "\x00\x80"
-     "\xF0\x39\x00\x80\xA3\x2E\x00\x80\x76\x2F\x00\x80\x80\x32\x00\x80\xAD\xA0"
-     "\x03\x00"
-     "\x34\x33\x00\x80\x68\x33\x00\x80\x1C\x36\x00\x80\x44\x36\x00\x80\xB0\x39"
-     "\x00\x80"
-     "\x2F\x2D\x00\x80\xB0\x36\x00\x80\x82\x8D\x21\x00\x83\x49\x02\x00\x80\x4D"
-     "\x02\x00"
-     "\x81\x45\x02\x00\x86\x49\x02\x00\x87\xF1\x2E\x00\x84\x49\x02\x00\x85\x49"
-     "\x02\x00"
-     "\x8A\x8D\x21\x00\x8B\x45\x02\x00\x88\xCD\x2E\x00\x89\x71\x02\x00\x8E\x5D"
-     "\x02\x00"
-     "\x8F\x45\x02\x00\x8C\x5D\x02\x00\x8D\x55\x02\x00\x92\x3D\x02\x00\x93\xA5"
-     "\x02\x00"
-     "\x90\x3D\x02\x00\x91\x35\x02\x00\x96\x51\x03\x00\x97\x51\x03\x00\x94\xB1"
-     "\x02\x00"
-     "\x95\xB1\x02\x00\x9A\x71\x03\x00\x9B\x71\x03\x00\x98\x61\x03\x00\x99\x61"
-     "\x03\x00"
-     "\x9E\x55\x03\x00\x9F\x59\x03\x00\x9C\x61\x03\x00\x9D\x61\x03\x00\xA2\xB9"
-     "\x03\x00"
-     "\xA3\xC5\x03\x00\xA0\xB9\x03\x00\xA1\xB9\x03\x00\xA6\xDD\x03\x00\xA7\xC5"
-     "\x03\x00"
-     "\xA4\xDD\x03\x00\xA5\xD5\x03\x00\xAA\xFD\x03\x00\xAB\xC5\x03\x00\xA8\xFD"
-     "\x03\x00"
-     "\xA9\xF5\x03\x00\xAE\xDD\x03\x00\xAF\xC5\x03\x00\xAC\xDD\x03\x00\xAD\xD5"
-     "\x03\x00"
-     "\xB2\xBD\x03\x00\xB3\x45\x03\x00\xB0\xBD\x03\x00\xB1\xB5\x03\x00\xB6\x59"
-     "\x03\x00"
-     "\xB7\x59\x03\x00\xB4\x8D\x25\x00\xB5\x59\x03\x00\xBA\x79\x03\x00\xBB\x79"
-     "\x03\x00"
-     "\xB8\x79\x03\x00\xB9\x79\x03\x00\xBE\x59\x03\x00\xBF\x59\x03\x00\xBC\x59"
-     "\x03\x00"
-     "\xBD\x59\x03\x00\x88\xCD\x02\x00\x89\xD5\x02\x00\x8A\xDD\x02\x00\x8B\xD5"
-     "\x02\x00"
-     "\x8C\xCD\x02\x00\x8D\xF1\x02\x00\x8E\xF1\x02\x00\x8F\xF1\x02\x00\x80\x71"
-     "\x2F\x00"
-     "\x81\x95\x02\x00\x82\x9D\x02\x00\x83\x95\x02\x00\x84\x8D\x02\x00\x85\xF5"
-     "\x02\x00"
-     "\x86\xFD\x02\x00\x87\xF5\x02\x00\x98\x4D\x0F\x00\x99\x55\x0F\x00\x9A\x5D"
-     "\x0F\x00"
-     "\x9B\x55\x0F\x00\x9C\xB9\x03\x00\x9D\x49\x0F\x00\x9E\x79\x0F\x00\x9F\x79"
-     "\x0F\x00"
-     "\x90\x91\x02\x00\x91\x91\x02\x00\x92\x91\x02\x00\x93\x91\x02\x00\x94\x01"
-     "\x0C\x00"
-     "\x95\x75\x0F\x00\x96\x7D\x0F\x00\x97\x75\x0F\x00\xA8\xC9\x0F\x00\xA9\xC9"
-     "\x0F\x00"
-     "\xAA\xD9\x0F\x00\xAB\xD9\x0F\x00\xAC\xC9\x0F\x00\xAD\xC9\x0F\x00\xAE\xF9"
-     "\x0F\x00"
-     "\xAF\xF9\x0F\x00\xA0\x89\x0F\x00\xA1\x89\x0F\x00\xA2\x99\x0F\x00\xA3\x99"
-     "\x0F\x00"
-     "\xA4\x89\x0F\x00\xA5\x89\x0F\x00\xA6\xF9\x0F\x00\xA7\xF9\x0F\x00\xB8\x45"
-     "\x0F\x00"
-     "\xB9\x4D\x0F\x00\xBA\x45\x0F\x00\xBB\x5D\x0F\x00\xBC\x45\x0F\x00\xBD\x4D"
-     "\x0F\x00"
-     "\xBE\x45\x0F\x00\xBF\x7D\x0F\x00\xB0\x89\x0F\x00\xB1\x71\x20\x00\xB2\x85"
-     "\x0F\x00"
-     "\xB3\x9D\x0F\x00\xB4\x85\x0F\x00\xB5\x8D\x0F\x00\xB6\x85\x0F\x00\xB7\x7D"
-     "\x0F\x00"
-     "\xCF\x2C\x00\x80\xA0\x35\x00\x80\x30\x36\x00\x80\x64\x37\x00\x80\xB8\x39"
-     "\x00\x80"
-     "\x70\x38\x00\x80\x15\x31\x00\x80\x60\x38\x00\x80\xD0\x2B\x00\x80\x4A\x2F"
-     "\x00\x80"
-     "\x81\x30\x00\x80\x38\x34\x00\x80\xB8\x34\x00\x80\x53\x2D\x00\x80\x44\x35"
-     "\x00\x80"
-     "\x3F\x2C\x00\x80\x65\x30\x00\x80\xB0\x2B\x00\x80\xF4\x32\x00\x80\x91\x31"
-     "\x00\x80"
-     "\x18\x36\x00\x80\xA7\x2D\x00\x80\x93\x2E\x00\x80\x04\x35\x00\x80\x6C\x37"
-     "\x00\x80"
-     "\x44\x38\x00\x80\xC0\x32\x00\x80\xEC\x32\x00\x80\x25\x31\x00\x80\x24\x37"
-     "\x00\x80"
-     "\x45\x31\x00\x80\xB4\x34\x00\x80\x04\x2C\x00\x80\x87\x2C\x00\x80\xCB\x2C"
-     "\x00\x80"
-     "\x5C\x32\x00\x80\x04\x36\x00\x80\xFC\x36\x00\x80\x50\x38\x00\x80\x64\x39"
-     "\x00\x80"
-     "\x3B\x2D\x00\x80\xEF\x2D\x00\x80\x1D\x30\x00\x80\xAD\x31\x00\x80\x84\x33"
-     "\x00\x80"
-     "\x48\x35\x00\x80\xA0\x39\x00\x80\x84\x3A\x00\x80\xD8\x3A\x00\x80\xF0\x3A"
-     "\x00\x80"
-     "\xEF\x31\x00\x80\xAC\x32\x00\x80\xF4\x38\x00\x80\xE9\x2F\x00\x80\xC5\x30"
-     "\x00\x80"
-     "\xE1\x30\x00\x80\x71\x31\x00\x80\x74\x33\x00\x80\xC0\x33\x00\x80\x70\x34"
-     "\x00\x80"
-     "\x94\x35\x00\x80\x38\x36\x00\x80\x9C\x36\x00\x80\xE4\x38\x00\x80\x80\xA5"
-     "\x01\x00"
-     "\x81\x35\x29\x00\x82\xA1\x01\x00\x83\xA1\x01\x00\x84\xA1\x01\x00\x85\xA1"
-     "\x01\x00"
-     "\x86\xA1\x01\x00\x87\xA1\x01\x00\x88\xE1\x01\x00\x89\xE1\x01\x00\x8A\xA5"
-     "\x56\x00"
-     "\x8B\xE5\x01\x00\x8C\xFD\x01\x00\x8D\xE5\x01\x00\x8E\xED\x01\x00\x8F\xE5"
-     "\x01\x00"
-     "\x90\x9D\x01\x00\x91\xA5\x01\x00\x92\xAD\x01\x00\x93\xA5\x01\x00\x94\xBD"
-     "\x01\x00"
-     "\x95\xA5\x01\x00\x96\xAD\x01\x00\x97\xA5\x01\x00\x98\x9D\x01\x00\x99\x65"
-     "\x03\x00"
-     "\x9A\x6D\x03\x00\x9B\x65\x03\x00\x9C\x7D\x03\x00\x9D\x65\x03\x00\x9E\x6D"
-     "\x03\x00"
-     "\x9F\x65\x03\x00\xA0\x9D\x03\x00\xA1\xE1\x01\x00\xA2\xA9\x03\x00\xA3\xA9"
-     "\x03\x00"
-     "\xA4\xB9\x03\x00\xA5\xB9\x03\x00\xA6\xA9\x03\x00\xA7\xA9\x03\x00\xA8\x99"
-     "\x03\x00"
-     "\xA9\x99\x03\x00\xAA\x8D\x01\x00\xAB\xED\x03\x00\xAC\xF5\x03\x00\xAD\xFD"
-     "\x03\x00"
-     "\xAE\xF5\x03\x00\xAF\xED\x03\x00\xB0\x95\x03\x00\xB1\x9D\x03\x00\xB2\x95"
-     "\x03\x00"
-     "\xB3\xAD\x03\x00\xB4\xB5\x03\x00\xB5\xBD\x03\x00\xB6\xB5\x03\x00\xB7\xAD"
-     "\x03\x00"
-     "\xB8\x95\x03\x00\xB9\x9D\x03\x00\xBA\x95\x03\x00\xBB\x6D\x03\x00\xBC\x75"
-     "\x03\x00"
-     "\xBD\x7D\x03\x00\xBE\x75\x03\x00\xBF\x15\x0D\x00\x74\x37\x00\x80\x0C\x39"
-     "\x00\x80"
-     "\x88\x39\x00\x80\xEF\x2C\x00\x80\x1B\x2D\x00\x80\x03\x2E\x00\x80\xB0\x33"
-     "\x00\x80"
-     "\xFC\x37\x00\x80\x09\x30\x00\x80\x79\x30\x00\x80\x3D\x31\x00\x80\xB8\x32"
-     "\x00\x80"
-     "\xE0\x35\x00\x80\x2C\x34\x00\x80\x14\x2C\x00\x80\x5F\x2E\x00\x80\xAE\x2F"
-     "\x00\x80"
-     "\x85\x30\x00\x80\x48\x34\x00\x80\x64\x34\x00\x80\x74\x36\x00\x80\x94\x36"
-     "\x00\x80"
-     "\x6C\x39\x00\x80\xEC\x39\x00\x80\xFC\x39\x00\x80\x38\x3A\x00\x80\x33\x2C"
-     "\x00\x80"
-     "\x88\x35\x00\x80\x58\x39\x00\x80\xD4\x39\x00\x80\x9D\x30\x00\x80\xF0\x2B"
-     "\x00\x80"
-     "\x57\x2C\x00\x80\x4E\x2F\x00\x80\x6E\x2F\x00\x80\x9D\x31\x00\x80\xCC\x33"
-     "\x00\x80"
-     "\xD4\x34\x00\x80\x80\x37\x00\x80\x20\x39\x00\x80\xDF\x31\x00\x80\xAC\x39"
-     "\x00\x80"
-     "\x13\x2D\x00\x80\x2B\x32\x00\x80\x37\x32\x00\x80\x18\x33\x00\x80\x50\x33"
-     "\x00\x80"
-     "\x54\x34\x00\x80\xB0\x34\x00\x80\x30\x35\x00\x80\x14\x36\x00\x80\x44\x3A"
-     "\x00\x80"
-     "\xF1\x30\x00\x80\xBC\x39\x00\x80\x47\x2C\x00\x80\x92\x2F\x00\x80\x38\x33"
-     "\x00\x80"
-     "\xE8\x38\x00\x80\x55\x30\x00\x80\xA5\x30\x00\x80\x3F\x32\x00\x80\xC4\x39"
-     "\x00\x80"
-     "\xF7\x2C\x00\x80\xDF\x2D\x00\x80\x82\xFC\x04\x00\x7E\x2F\x00\x80\xC8\x31"
-     "\x00\x80"
-     "\x1B\x32\x00\x80\x68\x32\x00\x80\xFC\x32\x00\x80\x50\x34\x00\x80\xC4\x34"
-     "\x00\x80"
-     "\x60\x36\x00\x80\x00\x38\x00\x80\x0C\x38\x00\x80\xE0\x39\x00\x80\x63\x2D"
-     "\x00\x80"
-     "\x54\x33\x00\x80\xE3\x2D\x00\x80\xD0\x33\x00\x80\x74\x34\x00\x80\x90\x37"
-     "\x00\x80"
-     "\xC8\x39\x00\x80\xB4\x3A\x00\x80\xDC\x3A\x00\x80\x33\x32\x00\x80\x30\x33"
-     "\x00\x80"
-     "\xCC\x36\x00\x80\xCC\x35\x00\x80\xF0\x35\x00\x80\x20\x34\x00\x80\xA8\x33"
-     "\x00\x80"
-     "\xD3\x2C\x00\x80\x6C\x38\x00\x80\x0C\x2C\x00\x80\x34\x37\x00\x80\xFB\x2C"
-     "\x00\x80"
-     "\xEB\x2C\x00\x80\xF9\x2F\x00\x80\x19\x31\x00\x80\xFC\x35\x00\x80\x3A\x2F"
-     "\x00\x80"
-     "\x0C\x33\x00\x80\xEB\x31\x00\x80\xF0\x38\x00\x80\xB4\x39\x00\x80\x15\x30"
-     "\x00\x80"
-     "\x67\x2C\x00\x80\x3B\x2E\x00\x80\x8F\x2E\x00\x80\xDB\x31\x00\x80\xD7\x2C"
-     "\x00\x80"
-     "\x28\x34\x00\x80\xEC\x34\x00\x80\x50\x35\x00\x80\x6C\x35\x00\x80\x70\x35"
-     "\x00\x80"
-     "\x84\x35\x00\x80\x04\x3A\x00\x80\xF4\x35\x00\x80\x98\x37\x00\x80\x5C\x38"
-     "\x00\x80"
-     "\x10\x39\x00\x80\x3C\x39\x00\x80\x58\x3A\x00\x80\x5C\x3A\x00\x80\x60\x3A"
-     "\x00\x80"
-     "\xC4\x3A\x00\x80\x80\x85\x07\x00\x81\x8D\x07\x00\x82\x85\x07\x00\x83\x9D"
-     "\x07\x00"
-     "\x84\x85\x07\x00\x85\x8D\x07\x00\x86\x85\x07\x00\x87\xBD\x07\x00\x88\x89"
-     "\x2F\x00"
-     "\x89\x81\x07\x00\x8A\x95\x2F\x00\x8B\x85\x07\x00\x8C\x9D\x07\x00\x8D\x85"
-     "\x07\x00"
-     "\x34\x39\x00\x80\xD0\x39\x00\x80\x90\xF5\x07\x00\x37\x2C\x00\x80\x92\xF1"
-     "\x07\x00"
-     "\x5B\x2C\x00\x80\x6F\x2C\x00\x80\x95\xE9\x07\x00\x96\x99\x07\x00\x97\x99"
-     "\x07\x00"
-     "\x98\xA9\x07\x00\x99\xA9\x07\x00\x9A\xB9\x07\x00\x9B\xB9\x07\x00\x9C\xA9"
-     "\x07\x00"
-     "\x9D\xA9\x07\x00\x9E\x29\x2F\x00\x2B\x2D\x00\x80\xA0\x61\x07\x00\x37\x2D"
-     "\x00\x80"
-     "\xA2\x6D\x07\x00\x77\x2D\x00\x80\x2B\x2E\x00\x80\xA5\x7D\x07\x00\xA6\x75"
-     "\x07\x00"
-     "\x47\x2E\x00\x80\x4F\x2E\x00\x80\x8B\x2E\x00\x80\xAA\x59\x07\x00\xAB\x59"
-     "\x07\x00"
-     "\xAC\x49\x07\x00\xAD\x49\x07\x00\xAE\x81\x00\x00\xAF\x81\x00\x00\xB0\x85"
-     "\x00\x00"
-     "\xB1\x89\x00\x00\xB2\x99\x00\x00\xB3\xB1\x00\x00\xB4\x55\x00\x00\xB5\x59"
-     "\x00\x00"
-     "\xB6\x49\x00\x00\xB7\x41\x00\x00\xB8\x41\x00\x00\xB9\x41\x00\x00\xBA\xE9"
-     "\x01\x00"
-     "\xBB\xE9\x01\x00\xBC\xA5\x2C\x00\xBD\xFD\x01\x00\xBE\xF5\x01\x00\xBF\xED"
-     "\x01\x00"
-     "\x97\x2E\x00\x80\x7A\x2F\x00\x80\x91\x30\x00\x80\xB9\x30\x00\x80\xC1\x30"
-     "\x00\x80"
-     "\xBA\xC9\xE6\x00\xD9\x30\x00\x80\x85\x31\x00\x80\xB5\x31\x00\x80\xE3\x31"
-     "\x00\x80"
-     "\x60\x32\x00\x80\x24\x33\x00\x80\x40\x33\x00\x80\x70\x33\x00\x80\xC4\x33"
-     "\x00\x80"
-     "\xE4\x33\x00\x80\x58\x34\x00\x80\x38\x35\x00\x80\x5C\x35\x00\x80\x58\x35"
-     "\x00\x80"
-     "\x60\x35\x00\x80\x64\x35\x00\x80\x7C\x35\x00\x80\x80\x35\x00\x80\xA4\x35"
-     "\x00\x80"
-     "\xB8\x35\x00\x80\xD4\x35\x00\x80\x40\x36\x00\x80\x48\x36\x00\x80\x5C\x36"
-     "\x00\x80"
-     "\x84\x36\x00\x80\xD4\x36\x00\x80\x08\x37\x00\x80\x70\x37\x00\x80\x18\x38"
-     "\x00\x80"
-     "\x30\x38\x00\x80\x64\x38\x00\x80\x68\x38\x00\x80\xAC\x38\x00\x80\xB0\x38"
-     "\x00\x80"
-     "\x08\x39\x00\x80\xE4\x39\x00\x80\x28\x3A\x00\x80\x40\x3A\x00\x80\x89\x30"
-     "\x00\x80"
-     "\xD1\xEB\x00\x80\xEC\x36\x00\x80\xB8\x2B\x00\x80\xC3\x2C\x00\x80\x83\x2C"
-     "\x00\x80"
-     "\x2B\x2C\x00\x80\x6B\x2C\x00\x80\x97\x2C\x00\x80\x27\x2D\x00\x80\x43\x2D"
-     "\x00\x80"
-     "\x17\x2E\x00\x80\x27\x2E\x00\x80\x43\x2E\x00\x80\x9F\x2E\x00\x80\xEA\x2E"
-     "\x00\x80"
-     "\xF6\x2E\x00\x80\x12\x2F\x00\x80\x22\x2F\x00\x80\x11\x30\x00\x80\xBF\x81"
-     "\x05\x00"
-     "\xBE\x79\x05\x00\xBD\x71\x05\x00\xBC\x79\x05\x00\xBB\x7D\x05\x00\xBA\x7D"
-     "\x05\x00"
-     "\xB9\x7D\x05\x00\xB8\x7D\x05\x00\xB7\x05\x2D\x00\xB6\x79\x05\x00\xB5\x71"
-     "\x05\x00"
-     "\xB4\x79\x05\x00\xB3\x61\x05\x00\xB2\x79\x05\x00\xB1\x71\x05\x00\xB0\x79"
-     "\x05\x00"
-     "\xAF\x01\x05\x00\xAE\x79\x05\x00\xAD\x71\x05\x00\xAC\x79\x05\x00\xAB\x61"
-     "\x05\x00"
-     "\xAA\x79\x05\x00\xA9\xF1\x29\x00\xA8\x75\x05\x00\xA7\x45\x05\x00\xA6\x45"
-     "\x05\x00"
-     "\xA5\x75\x05\x00\xA4\x75\x05\x00\xA3\x65\x05\x00\xA2\x65\x05\x00\xA1\x75"
-     "\x05\x00"
-     "\xA0\x75\x05\x00\x9F\x85\x05\x00\x9E\x79\x04\x00\x9D\x71\x04\x00\x9C\x79"
-     "\x04\x00"
-     "\x9B\xBD\x06\x00\x9A\x65\x04\x00\x99\x75\x04\x00\x98\x75\x04\x00\x97\x45"
-     "\x04\x00"
-     "\x96\x45\x04\x00\x95\x75\x04\x00\x94\x75\x04\x00\x93\x65\x04\x00\x92\x65"
-     "\x04\x00"
-     "\x91\xCD\x38\x00\x90\x71\x04\x00\x8F\x09\x04\x00\x8E\x01\x04\x00\x8D\x09"
-     "\x04\x00"
-     "\x8C\x71\x04\x00\x8B\x69\x04\x00\x8A\x61\x04\x00\x89\x69\x04\x00\x88\x71"
-     "\x04\x00"
-     "\x87\x49\x04\x00\x86\x41\x04\x00\x85\x49\x04\x00\x84\x0D\x26\x00\x83\x6D"
-     "\x04\x00"
-     "\x82\x6D\x04\x00\x81\x6D\x04\x00\x80\x6D\x04\x00\x85\x2D\x02\x00\x84\x2D"
-     "\x02\x00"
-     "\x87\x3D\x02\x00\x86\x3D\x02\x00\x81\x0D\x02\x00\x80\x0D\x02\x00\x83\x3D"
-     "\x02\x00"
-     "\x82\x3D\x02\x00\x8D\x69\x02\x00\x8C\xAD\x39\x00\x8F\x79\x02\x00\x8E\x61"
-     "\x02\x00"
-     "\x89\x0D\x02\x00\x88\x0D\x02\x00\x8B\x7D\x02\x00\x8A\x7D\x02\x00\x95\x25"
-     "\x02\x00"
-     "\x94\x25\x02\x00\x97\xBD\x07\x00\x96\x25\x02\x00\x91\xC9\x39\x00\x90\x01"
-     "\x02\x00"
-     "\x93\x05\x02\x00\x92\x05\x02\x00\x9D\xE1\x02\x00\x9C\x19\x02\x00\x9F\xED"
-     "\x02\x00"
-     "\x9E\xED\x02\x00\x99\x01\x02\x00\x98\x19\x02\x00\x9B\x01\x02\x00\x9A\x09"
-     "\x02\x00"
-     "\xA5\x1D\x02\x00\xA4\x1D\x02\x00\xA7\x61\x04\x00\xA6\x2D\x02\x00\xA1\x1D"
-     "\x02\x00"
-     "\xA0\x1D\x02\x00\xA3\x0D\x02\x00\xA2\x0D\x02\x00\xAD\x19\x02\x00\xAC\x11"
-     "\x02\x00"
-     "\x19\x30\x00\x80\x49\x30\x00\x80\xA9\x19\x02\x00\xA8\x11\x02\x00\xAB\x09"
-     "\x02\x00"
-     "\xAA\x11\x02\x00\xB5\x11\x02\x00\xB4\x09\x02\x00\xB7\x11\x02\x00\xB6\x19"
-     "\x02\x00"
-     "\xB1\x11\x02\x00\xB0\x69\x02\x00\xB3\x11\x02\x00\xB2\x19\x02\x00\xBD\x09"
-     "\x02\x00"
-     "\xBC\x51\x04\x00\xBF\x19\x02\x00\xBE\x01\x02\x00\xB9\x2D\x02\x00\xB8\x65"
-     "\x03\x00"
-     "\xBB\x1D\x02\x00\xBA\x1D\x02\x00\x69\x30\x00\x80\x99\x30\x00\x80\xB1\x30"
-     "\x00\x80"
-     "\xA1\x30\x00\x80\xB5\x30\x00\x80\xF5\x30\x00\x80\x4D\x31\x00\x80\x59\x31"
-     "\x00\x80"
-     "\x61\x31\x00\x80\x89\x31\x00\x80\x0F\x32\x00\x80\x23\x32\x00\x80\x64\x33"
-     "\x00\x80"
-     "\x5C\x33\x00\x80\x8C\x33\x00\x80\x1C\x35\x00\x80\xE8\x33\x00\x80\x1C\x34"
-     "\x00\x80"
-     "\x68\x34\x00\x80\x84\x34\x00\x80\xA8\x34\x00\x80\xCC\x34\x00\x80\xD0\x34"
-     "\x00\x80"
-     "\xF0\x34\x00\x80\x00\x35\x00\x80\x14\x35\x00\x80\x10\x35\x00\x80\x40\x35"
-     "\x00\x80"
-     "\xBC\x35\x00\x80\xEC\x35\x00\x80\x24\x36\x00\x80\x54\x36\x00\x80\x38\x37"
-     "\x00\x80"
-     "\x60\x37\x00\x80\xD4\x37\x00\x80\x1C\x38\x00\x80\x28\x38\x00\x80\x48\x38"
-     "\x00\x80"
-     "\x4C\x38\x00\x80\x58\x38\x00\x80\x74\x38\x00\x80\xEC\x38\x00\x80\x1C\x39"
-     "\x00\x80"
-     "\x54\x39\x00\x80\x74\x39\x00\x80\xC0\x39\x00\x80\x18\x3A\x00\x80\x34\x3A"
-     "\x00\x80"
-     "\x94\x3A\x00\x80\x68\xEB\x00\x80\xF0\x2C\x0A\x00\x63\xEB\x00\x80\x95\xEB"
-     "\x00\x80"
-     "\xE0\x2A\x00\x80\x08\x2B\x00\x80\x0C\x2B\x00\x80\x17\xEC\x00\x80\x44\xEC"
-     "\x00\x80"
-     "\xBC\xEC\x00\x80\x20\x3B\x00\x80\x30\x3B\x00\x80\x49\x08\x00\x80\x54\x08"
-     "\x00\x80"
-     "\x57\x08\x00\x80\x80\xA1\x01\x00\x81\x99\x1C\x00\x82\xAD\x01\x00\x83\xA5"
-     "\x01\x00"
-     "\x84\xBD\x01\x00\x85\xA5\x01\x00\x86\xAD\x01\x00\x87\xB9\x1C\x00\x88\xF9"
-     "\x1A\x00"
-     "\x89\x9D\x01\x00\x8A\x81\x1C\x00\x8B\x91\x01\x00\x8C\x9D\x1C\x00\x8D\xF5"
-     "\x01\x00"
-     "\x8E\xDD\x39\x00\x8F\xF9\x01\x00\x90\x8D\x01\x00\x91\x95\x01\x00\x92\x9D"
-     "\x01\x00"
-     "\x93\x95\x01\x00\x94\x8D\x01\x00\x95\xB5\x01\x00\x96\xBD\x01\x00\x97\xB5"
-     "\x01\x00"
-     "\x98\x8D\x01\x00\x99\x95\x01\x00\x4C\x08\x00\x80\x50\x08\x00\x80\x59\x09"
-     "\x00\x80"
-     "\xE8\x0F\x00\x80\xD9\x0F\x00\x80\xDE\x0F\x00\x80\xED\x0F\x00\x80\xE3\x0F"
-     "\x00\x80"
-     "\x42\x10\x00\x80\xC1\x10\x00\x80\x78\x10\x00\x80\x26\x10\x00\x80\x51\x10"
-     "\x00\x80"
-     "\x5E\x10\x00\x80\x66\x10\x00\x80\x94\x10\x00\x80\xB9\x10\x00\x80\xAF\x10"
-     "\x00\x80"
-     "\xB4\x10\x00\x80\xA1\x10\x00\x80\xA8\x10\x00\x80\xF5\x0F\x00\x80\xFA\x0F"
-     "\x00\x80"
-     "\xFF\x0F\x00\x80\x0C\x10\x00\x80\x19\x10\x00\x80\x21\x10\x00\x80\x29\x10"
-     "\x00\x80"
-     "\x33\x10\x00\x80\x38\x10\x00\x80\x3D\x10\x00\x80\x47\x10\x00\x80\x4C\x10"
-     "\x00\x80"
-     "\x54\x10\x00\x80\x61\x10\x00\x80\x69\x10\x00\x80\x6E\x10\x00\x80\x73\x10"
-     "\x00\x80"
-     "\xB2\x79\x1F\x00\xB3\x85\x02\x00\xB0\xFD\x1A\x00\xB1\x71\x02\x00\xB6\x9D"
-     "\x02\x00"
-     "\xB7\x85\x02\x00\xB4\x9D\x02\x00\xB5\x95\x02\x00\xBA\xC5\x19\x00\xBB\xBD"
-     "\x02\x00"
-     "\xB8\xA5\x1F\x00\xB9\xB9\x02\x00\xBE\xAD\x02\x00\xBF\xA1\x1F\x00\xBC\xA9"
-     "\x1F\x00"
-     "\xBD\x21\x1A\x00\xA2\x0D\x02\x00\xA3\x15\x02\x00\xA0\x89\x19\x00\xA1\x05"
-     "\x02\x00"
-     "\xA6\x95\x19\x00\xA7\x09\x02\x00\xA4\x0D\x02\x00\xA5\x05\x02\x00\xAA\x09"
-     "\x02\x00"
-     "\xAB\x09\x02\x00\xA8\x09\x02\x00\xA9\x09\x02\x00\xAE\x05\x02\x00\xAF\x0D"
-     "\x02\x00"
-     "\xAC\x09\x02\x00\xAD\x15\x1F\x00\x92\xD9\x17\x00\x93\xAD\x02\x00\x90\xAD"
-     "\x02\x00"
-     "\x91\x35\x19\x00\x96\xF9\x1A\x00\x97\x21\x18\x00\x94\xB5\x02\x00\x95\x65"
-     "\x32\x00"
-     "\x9A\x95\x1F\x00\x9B\x9D\x02\x00\x98\x99\x02\x00\x99\x99\x02\x00\x9E\xF1"
-     "\x02\x00"
-     "\x9F\xF1\x02\x00\x9C\x81\x1F\x00\x9D\x81\x02\x00\x82\x3D\x1F\x00\x83\x6D"
-     "\x1D\x00"
-     "\x80\x3D\x1F\x00\x81\x35\x1F\x00\x86\xD5\x02\x00\x87\xDD\x02\x00\x84\x75"
-     "\x1D\x00"
-     "\x85\xDD\x02\x00\x8A\xF1\x02\x00\x8B\xFD\x1F\x00\x88\xE5\x02\x00\x89\xFD"
-     "\x1F\x00"
-     "\x8E\xED\x02\x00\x8F\xD5\x02\x00\x8C\xED\x02\x00\x8D\xE5\x02\x00\x80\xF5"
-     "\x03\x00"
-     "\x81\xFD\x03\x00\x82\xF5\x03\x00\x83\x65\x02\x00\x84\x7D\x02\x00\x85\x65"
-     "\x02\x00"
-     "\x86\x69\x02\x00\x7B\x10\x00\x80\x80\x10\x00\x80\x8A\x10\x00\x80\x8F\x10"
-     "\x00\x80"
-     "\x97\x10\x00\x80\x9C\x10\x00\x80\xBC\x10\x00\x80\x2E\x10\x00\x80\x11\x10"
-     "\x00\x80"
-     "\x59\x10\x00\x80\x85\x10\x00\x80\x04\x10\x00\x80\x93\x39\x02\x00\x94\x29"
-     "\x02\x00"
-     "\x95\x29\x02\x00\x96\xD9\x02\x00\x97\xD9\x02\x00\xD4\x17\x00\x80\xDD\x17"
-     "\x00\x80"
-     "\xE0\x17\x00\x80\xE6\x17\x00\x80\xDA\x17\x00\x80\x9D\xFD\x02\x00\xE3\x17"
-     "\x00\x80"
-     "\x9F\xF1\x02\x00\xA0\x11\x02\x00\xA1\x65\x46\x00\xA2\x7D\x46\x00\xA3\x19"
-     "\x02\x00"
-     "\xA4\x09\x02\x00\xA5\x09\x02\x00\xA6\x39\x02\x00\xA7\x39\x02\x00\xA8\x09"
-     "\x02\x00"
-     "\xA9\x25\x46\x00\xAA\x05\x02\x00\xAB\x1D\x02\x00\xAC\x05\x02\x00\xAD\x0D"
-     "\x02\x00"
-     "\xAE\x05\x02\x00\xAF\x7D\x02\x00\xB0\x05\x02\x00\xB1\x0D\x02\x00\xB2\x05"
-     "\x02\x00"
-     "\xB3\x1D\x02\x00\xB4\x05\x02\x00\xB5\x0D\x02\x00\xB6\x05\x02\x00\xD7\x17"
-     "\x00\x80"
-     "\xB8\x01\x02\x00\xB9\x01\x02\x00\xBA\x01\x02\x00\xBB\x01\x02\x00\xBC\x01"
-     "\x02\x00"
-     "\x0A\x18\x00\x80\xBE\x0D\x02\x00\x0D\x18\x00\x80\xEC\x17\x00\x80\xE9\x17"
-     "\x00\x80"
-     "\xEF\x17\x00\x80\xF2\x17\x00\x80\xB0\x79\x51\x02\xB1\x7D\x51\x02\xFB\x17"
-     "\x00\x80"
-     "\xF8\x17\x00\x80\xFE\x17\x00\x80\xF5\x17\x00\x80\x04\x18\x00\x80\x01\x18"
-     "\x00\x80"
-     "\x10\x18\x00\x80\x16\x18\x00\x80\x1C\x18\x00\x80\x19\x18\x00\x80\xA4\x39"
-     "\x51\x02"
-     "\xA5\x3D\x51\x02\xA6\xED\x01\x00\xA7\xE9\x01\x00\xA0\x55\x01\x00\xA1\x51"
-     "\x01\x00"
-     "\xA2\x5D\x01\x00\xA3\x59\x01\x00\xAC\x69\x01\x00\xAD\x6D\x01\x00\xAE\xD1"
-     "\x01\x00"
-     "\xAF\xD5\x01\x00\xA8\xD5\x01\x00\xA9\xD1\x01\x00\xAA\x71\x01\x00\xAB\x75"
-     "\x01\x00"
-     "\x94\xB5\x00\x00\x95\xB1\x00\x00\x96\xB1\x00\x00\x97\xB5\x00\x00\x90\xA1"
-     "\x00\x00"
-     "\x91\xA5\x00\x00\x92\xAD\x00\x00\x93\xA9\x00\x00\x9C\x9D\x00\x00\x9D\x99"
-     "\x00\x00"
-     "\x9E\xA9\x01\x00\x9F\xAD\x01\x00\x98\x89\x00\x00\x99\x8D\x00\x00\x9A\x85"
-     "\x00\x00"
-     "\x9B\x81\x00\x00\x84\xDD\x00\x00\x85\xD9\x00\x00\x86\xE9\x00\x00\x87\xED"
-     "\x00\x00"
-     "\x80\xDD\x00\x00\x81\xD9\x00\x00\x82\xC1\x00\x00\x83\xC5\x00\x00\x8C\xCD"
-     "\x00\x00"
-     "\x8D\xC9\x00\x00\x8E\xD9\x00\x00\x8F\xDD\x00\x00\x88\xD5\x00\x00\x89\xD1"
-     "\x00\x00"
-     "\x8A\xD1\x00\x00\x8B\xD5\x00\x00\x1F\x18\x00\x80\x22\x18\x00\x80\x25\x18"
-     "\x00\x80"
-     "\x13\x18\x00\x80\x37\x18\x00\x80\x34\x18\x00\x80\x3F\x18\x00\x80\x45\x18"
-     "\x00\x80"
-     "\x31\x18\x00\x80\x42\x18\x00\x80\x4B\x18\x00\x80\x70\x17\x00\x80\xD8\x10"
-     "\x00\x80"
-     "\x28\x11\x00\x80\x05\x11\x00\x80\x19\x11\x00\x80\x14\x11\x00\x80\x1E\x11"
-     "\x00\x80"
-     "\x23\x11\x00\x80\xBA\xAD\xFE\x00\x0A\x11\x00\x80\x48\x18\x00\x80\xDD\x10"
-     "\x00\x80"
-     "\xE2\x10\x00\x80\xF6\x10\x00\x80\x0F\x11\x00\x80\x4F\x11\x00\x80\x54\x11"
-     "\x00\x80"
-     "\x60\x11\x00\x80\x76\x11\x00\x80\x85\x11\x00\x80\x8A\x11\x00\x80\x95\x11"
-     "\x00\x80"
-     "\xAF\x11\x00\x80\xC2\x11\x00\x80\xE6\x11\x00\x80\x18\x12\x00\x80\x1D\x12"
-     "\x00\x80"
-     "\x25\x12\x00\x80\x34\x12\x00\x80\x43\x12\x00\x80\x2B\x18\x00\x80\x48\x12"
-     "\x00\x80"
-     "\x50\x12\x00\x80\x73\x12\x00\x80\x9A\x12\x00\x80\x4E\x18\x00\x80\xA6\x12"
-     "\x00\x80"
-     "\xC6\x12\x00\x80\xCB\x12\x00\x80\xD0\x12\x00\x80\x0A\x13\x00\x80\x1D\x13"
-     "\x00\x80"
-     "\x29\x13\x00\x80\x41\x13\x00\x80\xBB\x13\x00\x80\x11\x14\x00\x80\x2A\x14"
-     "\x00\x80"
-     "\x2F\x14\x00\x80\x42\x14\x00\x80\x53\x14\x00\x80\x65\x14\x00\x80\x6A\x14"
-     "\x00\x80"
-     "\x91\x14\x00\x80\xB5\x01\x02\x00\xB4\x05\x02\x00\xB7\x0D\x02\x00\xB6\x09"
-     "\x02\x00"
-     "\xB1\xED\x05\x00\xB0\xE9\x05\x00\xB3\x19\x02\x00\xB2\x1D\x02\x00\xBD\x29"
-     "\x02\x00"
-     "\xBC\x2D\x02\x00\xBF\x25\x02\x00\xBE\x21\x02\x00\xB9\x35\x02\x00\xB8\x31"
-     "\x02\x00"
-     "\xBB\x31\x02\x00\xBA\x35\x02\x00\xA5\x39\x05\x00\xA4\x3D\x05\x00\xA7\x95"
-     "\x05\x00"
-     "\xA6\x91\x05\x00\xA1\x21\x05\x00\xA0\x25\x05\x00\xA3\x21\x05\x00\xA2\x25"
-     "\x05\x00"
-     "\xAD\x95\x05\x00\xAC\x91\x05\x00\xAF\x95\x05\x00\xAE\x91\x05\x00\xA9\xAD"
-     "\x05\x00"
-     "\xA8\xA9\x05\x00\xAB\x8D\x05\x00\xAA\x89\x05\x00\x95\xE5\x05\x00\x94\xE1"
-     "\x05\x00"
-     "\x97\xE1\x05\x00\x96\xE5\x05\x00\x91\xF1\x05\x00\x90\xF5\x05\x00\x93\xFD"
-     "\x05\x00"
-     "\x92\xF9\x05\x00\x9D\xCD\x05\x00\x9C\xC9\x05\x00\x9F\xD9\x05\x00\x9E\xDD"
-     "\x05\x00"
-     "\x99\xD9\x05\x00\x98\xDD\x05\x00\x9B\xD5\x05\x00\x9A\xD1\x05\x00\xD9\xF0"
-     "\x51\x02"
-     "\x84\xE5\x05\x00\x87\xE1\x05\x00\x86\xE9\x05\x00\x81\x2D\x07\x00\x80\x2D"
-     "\x07\x00"
-     "\x83\xC5\x05\x00\xD9\xE0\x51\x02\x8D\xA1\x05\x00\x8C\xD9\x05\x00\x8F\xA1"
-     "\x05\x00"
-     "\x8E\xA9\x05\x00\x89\xC1\x05\x00\x88\xD9\x05\x00\x8B\xC1\x05\x00\x8A\xC9"
-     "\x05\x00"
-     "\x93\x0D\x02\x00\x99\x14\x00\x80\xAE\x14\x00\x80\xD5\x14\x00\x80\x97\x01"
-     "\x02\x00"
-     "\x96\x19\x02\x00\x95\x15\x02\x00\x94\x11\x02\x00\x9B\x39\x02\x00\x9A\x31"
-     "\x02\x00"
-     "\x99\x3D\x02\x00\x98\x3D\x02\x00\x9F\x2D\x02\x00\x9E\x29\x02\x00\x9D\x69"
-     "\x52\x02"
-     "\x9C\x25\x02\x00\xDA\x14\x00\x80\x34\x15\x00\x80\x39\x15\x00\x80\x3E\x15"
-     "\x00\x80"
-     "\x4A\x15\x00\x80\x56\x15\x00\x80\x5B\x15\x00\x80\x63\x15\x00\x80\x6F\x15"
-     "\x00\x80"
-     "\x89\x15\x00\x80\x8E\x15\x00\x80\x96\x15\x00\x80\x9B\x15\x00\x80\xA0\x15"
-     "\x00\x80"
-     "\xA5\x15\x00\x80\xAA\x15\x00\x80\xB3\xC1\x02\x00\xB2\xC5\x02\x00\xB1\xB5"
-     "\x02\x00"
-     "\xB0\xB1\x02\x00\xB7\xD9\x02\x00\xB6\xDD\x02\x00\xB5\xDD\x02\x00\xB4\xD9"
-     "\x02\x00"
-     "\xBB\xF1\x02\x00\xBA\xF5\x02\x00\xB9\xF9\x02\x00\xB8\xE5\x02\x00\xBF\xE5"
-     "\x02\x00"
-     "\xBE\xE1\x02\x00\xBD\xED\x02\x00\xBC\xE9\x02\x00\xA3\xE5\x02\x00\xA2\xE1"
-     "\x02\x00"
-     "\xA1\xE9\x02\x00\xA0\xED\x02\x00\xA7\xF9\x02\x00\xA6\xFD\x02\x00\xA5\xF1"
-     "\x02\x00"
-     "\xA4\xF5\x02\x00\xAB\xD1\x02\x00\xAA\xD5\x02\x00\xA9\xC5\x02\x00\xA8\xC1"
-     "\x02\x00"
-     "\xAF\xC9\x02\x00\xAE\xCD\x02\x00\xAD\xCD\x02\x00\xAC\xC9\x02\x00\xAF\x15"
-     "\x00\x80"
-     "\xC2\x15\x00\x80\xC7\x15\x00\x80\xE3\x15\x00\x80\xFD\x15\x00\x80\x17\x16"
-     "\x00\x80"
-     "\x23\x16\x00\x80\x3B\x16\x00\x80\x40\x16\x00\x80\x4D\x16\x00\x80\x6E\x16"
-     "\x00\x80"
-     "\x91\x16\x00\x80\xAB\x16\x00\x80\xB7\x16\x00\x80\xBC\x16\x00\x80\xC4\x16"
-     "\x00\x80"
-     "\xE5\x16\x00\x80\xFF\x16\x00\x80\x0E\x17\x00\x80\x2B\x17\x00\x80\x30\x17"
-     "\x00\x80"
-     "\x38\x17\x00\x80\x3D\x17\x00\x80\x50\x17\x00\x80\x55\x17\x00\x80\x7B\x17"
-     "\x00\x80"
-     "\x87\x17\x00\x80\x93\x17\x00\x80\xA2\x17\x00\x80\xBF\x17\x00\x80\xC4\x17"
-     "\x00\x80"
-     "\xE5\x12\x00\x80\xF6\x12\x00\x80\x73\x17\x00\x80\x80\x00\x00\x80\x8A\x00"
-     "\x00\x80"
-     "\x94\x00\x00\x80\x9E\x00\x00\x80\xA8\x00\x00\x80\xB2\x00\x00\x80\x91\xA8"
-     "\xFD\x00"
-     "\xEC\x10\x00\x80\xF1\x10\x00\x80\xFB\x10\x00\x80\x6C\x11\x00\x80\x71\x11"
-     "\x00\x80"
-     "\x7B\x11\x00\x80\xDC\x11\x00\x80\xE1\x11\x00\x80\x0E\x12\x00\x80\x2A\x12"
-     "\x00\x80"
-     "\x2F\x12\x00\x80\x39\x12\x00\x80\x48\x16\x00\x80\x04\x17\x00\x80\x09\x17"
-     "\x00\x80"
-     "\x21\x17\x00\x80\x98\x17\x00\x80\x9D\x17\x00\x80\xB5\x17\x00\x80\xE7\x10"
-     "\x00\x80"
-     "\x00\x11\x00\x80\x80\x11\x00\x80\x13\x12\x00\x80\x80\x11\x01\x00\x81\x11"
-     "\x01\x00"
-     "\x82\x11\x01\x00\x83\x11\x01\x00\x84\x31\x01\x00\x85\x31\x01\x00\x86\x31"
-     "\x01\x00"
-     "\x87\x31\x01\x00\x88\x11\x01\x00\x89\x11\x01\x00\x8A\x11\x01\x00\x8B\x11"
-     "\x01\x00"
-     "\x8C\x71\x01\x00\x8D\x71\x01\x00\x8E\x71\x01\x00\x8F\x71\x01\x00\x90\x11"
-     "\x01\x00"
-     "\x91\x11\x01\x00\x92\x11\x01\x00\x93\x11\x01\x00\x94\x31\x01\x00\x95\x31"
-     "\x01\x00"
-     "\x96\x31\x01\x00\x97\x31\x01\x00\x98\x11\x01\x00\x99\x11\x01\x00\x9A\x11"
-     "\x01\x00"
-     "\x9B\x11\x01\x00\x9C\xF1\x01\x00\x9D\xF1\x01\x00\x9E\xF1\x01\x00\x9F\xF1"
-     "\x01\x00"
-     "\xA0\x11\x01\x00\xA1\x11\x01\x00\xA2\x11\x01\x00\xA3\x11\x01\x00\xA4\x35"
-     "\x01\x00"
-     "\xA5\x3D\x01\x00\xA6\xF9\x0C\x00\xA7\x31\x01\x00\xA8\xF1\x0C\x00\xA9\xC1"
-     "\x0C\x00"
-     "\xAA\x19\x01\x00\xAB\x19\x01\x00\xAC\xC5\x0C\x00\xAD\x0D\x01\x00\xAE\xC1"
-     "\x0C\x00"
-     "\xAF\xC1\x0C\x00\xB0\x7D\x01\x00\xB1\x05\x01\x00\xB2\x45\x0C\x00\xB3\x09"
-     "\x01\x00"
-     "\xB4\x41\x0C\x00\xB5\x41\x0C\x00\xB6\x11\x01\x00\xB7\x11\x01\x00\xB8\x7D"
-     "\x0C\x00"
-     "\xB9\x35\x01\x00\xBA\x49\x0C\x00\xBB\x45\x0C\x00\xBC\xA5\x0E\x00\xBD\xAD"
-     "\x0E\x00"
-     "\xBE\x9D\x0E\x00\xBF\x95\x0E\x00\xA8\xFD\x0F\x00\xA9\x01\x0D\x00\xAA\x01"
-     "\x0D\x00"
-     "\xAB\x01\x0D\x00\xAC\x51\x0D\x00\xAD\x51\x0D\x00\xAE\x51\x0D\x00\xAF\x51"
-     "\x0D\x00"
-     "\xA0\xFD\x0F\x00\xA1\xC5\x0F\x00\xA2\xCD\x0F\x00\xA3\xC5\x0F\x00\xA4\xDD"
-     "\x0F\x00"
-     "\xA5\xC5\x0F\x00\xA6\xCD\x0F\x00\xA7\xC5\x0F\x00\xB8\x31\x0D\x00\xB9\x31"
-     "\x0D\x00"
-     "\xBA\x31\x0D\x00\xBB\x31\x0D\x00\xBC\x51\x02\x00\xBD\x51\x02\x00\xBE\x51"
-     "\x02\x00"
-     "\xBF\x51\x02\x00\xB0\x31\x0D\x00\xB1\x31\x0D\x00\xB2\x31\x0D\x00\xB3\x31"
-     "\x0D\x00"
-     "\xB4\x11\x0D\x00\xB5\x11\x0D\x00\xB6\x11\x0D\x00\xB7\x11\x0D\x00\x88\xF5"
-     "\x0F\x00"
-     "\x89\xFD\x0F\x00\x8A\xF5\x0F\x00\x8B\x0D\x0F\x00\x8C\x15\x0F\x00\x8D\x1D"
-     "\x0F\x00"
-     "\x8E\x15\x0F\x00\x8F\x0D\x0F\x00\x80\xF9\x0F\x00\x81\xF9\x0F\x00\x82\xC9"
-     "\x0F\x00"
-     "\x83\xFD\x0F\x00\x84\xD5\x0F\x00\x85\xDD\x0F\x00\x86\xD5\x0F\x00\x87\xCD"
-     "\x0F\x00"
-     "\x98\x7D\x0F\x00\x99\x05\x0F\x00\x9A\x0D\x0F\x00\x9B\x05\x0F\x00\x9C\x1D"
-     "\x0F\x00"
-     "\x9D\x05\x0F\x00\x9E\x0D\x0F\x00\x9F\x05\x0F\x00\x90\x75\x0F\x00\x91\x7D"
-     "\x0F\x00"
-     "\x92\x75\x0F\x00\x93\x4D\x0F\x00\x94\x59\x0F\x00\x95\x59\x0F\x00\x96\x49"
-     "\x0F\x00"
-     "\x97\x49\x0F\x00\x80\x6D\x0C\x00\x81\x79\x0C\x00\x82\x09\x03\x00\x83\x09"
-     "\x03\x00"
-     "\x84\x19\x03\x00\x85\x0D\x03\x00\x86\x05\x03\x00\x87\x3D\x03\x00\x88\xF5"
-     "\x03\x00"
-     "\x89\x15\x03\x00\x8A\xF1\x03\x00\x8B\xF1\x03\x00\x8C\x79\x03\x00\x8D\xD5"
-     "\x03\x00"
-     "\x8E\x75\x03\x00\x8F\x6D\x03\x00\x90\xC5\x03\x00\x91\xA1\x03\x00\x92\xA1"
-     "\x03\x00"
-     "\x93\x3D\x03\x00\x94\xBD\x03\x00\x95\x21\x03\x00\x96\x21\x03\x00\x97\x05"
-     "\x0E\x00"
-     "\x98\x3D\x0E\x00\x99\x95\x03\x00\x9A\x09\x0E\x00\x9B\x99\x03\x00\x9C\x19"
-     "\x0E\x00"
-     "\x9D\x19\x0E\x00\x9E\x09\x0E\x00\x9F\x09\x0E\x00\xA0\x79\x03\x00\xA1\x05"
-     "\x0E\x00"
-     "\xA2\x0D\x0E\x00\xA3\x05\x0E\x00\xA4\x1D\x0E\x00\xA5\x69\x03\x00\xA6\x05"
-     "\x0E\x00"
-     "\xA7\x31\x0E\x00\xA8\x11\x0E\x00\xA9\x11\x0E\x00\xAA\x15\x0E\x00\xAB\x6D"
-     "\x0E\x00"
-     "\xAC\x79\x0E\x00\xAD\x79\x0E\x00\xAE\x69\x0E\x00\xAF\x69\x0E\x00\xB0\x19"
-     "\x0E\x00"
-     "\xB1\x19\x0E\x00\xB2\x01\x04\x00\xB3\x2D\x0E\x00\xB4\x35\x0E\x00\xB5\x3D"
-     "\x0E\x00"
-     "\xB6\x35\x0E\x00\xB7\x2D\x0E\x00\xB8\x15\x0E\x00\xB9\x19\x0E\x00\xBA\xED"
-     "\x0C\x00"
-     "\xBB\xE5\x0C\x00\xBC\xFD\x0C\x00\xBD\xB5\x0C\x00\xBE\xBD\x0C\x00\xBF\xB5"
-     "\x0C\x00"
-     "\x80\x4D\x0B\x00\x81\x55\x0B\x00\x82\x55\x0B\x00\x83\x6D\x0B\x00\x84\x61"
-     "\x0B\x00"
-     "\x85\x61\x0B\x00\x86\x61\x0B\x00\x87\x61\x0B\x00\x88\x21\x04\x00\x89\x29"
-     "\x04\x00"
-     "\x8A\x39\x04\x00\x8B\x39\x04\x00\x8C\x29\x04\x00\x8D\x35\x01\x00\x8E\x1D"
-     "\x04\x00"
-     "\x8F\x15\x04\x00\x90\x6D\x04\x00\x91\x75\x04\x00\x92\x75\x04\x00\x93\x0D"
-     "\x04\x00"
-     "\x94\x15\x04\x00\x95\x1D\x04\x00\x96\x5D\x01\x00\x97\x09\x04\x00\x98\x39"
-     "\x04\x00"
-     "\x99\x6D\x01\x00\x9A\x0D\x04\x00\x9B\x05\x04\x00\x9C\x1D\x04\x00\x9D\x05"
-     "\x04\x00"
-     "\x9E\x75\x01\x00\x9F\x1D\x09\x00\xA0\x75\x04\x00\xA1\xF1\x09\x00\xA2\x71"
-     "\x04\x00"
-     "\xA3\x01\x09\x00\xA4\x6D\x04\x00\xA5\x09\x09\x00\xA6\x89\x01\x00\xA7\x45"
-     "\x09\x00"
-     "\xA8\xB5\x01\x00\xA9\xBD\x01\x00\xAA\xB5\x01\x00\xAB\xD1\x0B\x00\xAC\xB1"
-     "\x04\x00"
-     "\xAD\xAD\x04\x00\xAE\xF1\x04\x00\xAF\xB9\x01\x00\xB0\xB1\x04\x00\xB1\xC9"
-     "\x01\x00"
-     "\xB2\xE9\x01\x00\xB3\xE9\x01\x00\xB4\xF9\x01\x00\xB5\xF9\x01\x00\xB6\xE9"
-     "\x01\x00"
-     "\xB7\xE9\x01\x00\xB8\xD9\x01\x00\xB9\xD9\x01\x00\xBA\xA9\x01\x00\xBB\xA9"
-     "\x01\x00"
-     "\xBC\xB9\x01\x00\xBD\xB9\x01\x00\xBE\xA9\x01\x00\xBF\xA9\x01\x00\x36\x16"
-     "\x00\x80"
-     "\x26\x17\x00\x80\x5A\x17\x00\x80\xBA\x17\x00\x80\x3E\x12\x00\x80\x5B\x13"
-     "\x00\x80"
-     "\x96\x13\x00\x80\xA9\x13\x00\x80\x08\x15\x00\x80\x14\x15\x00\x80\x20\x15"
-     "\x00\x80"
-     "\x84\x14\x00\x80\x89\x14\x00\x80\xC8\x14\x00\x80\xCD\x14\x00\x80\xF4\x14"
-     "\x00\x80"
-     "\xF9\x14\x00\x80\x60\x13\x00\x80\x65\x13\x00\x80\xAE\x13\x00\x80\xB3\x13"
-     "\x00\x80"
-     "\xB9\x12\x00\x80\xBE\x12\x00\x80\x8D\x12\x00\x80\x92\x12\x00\x80\xD5\x12"
-     "\x00\x80"
-     "\xDA\x12\x00\x80\x1D\x14\x00\x80\x22\x14\x00\x80\x58\x14\x00\x80\x5D\x14"
-     "\x00\x80"
-     "\x6D\x13\x00\x80\x79\x13\x00\x80\x8C\x13\x00\x80\x91\x13\x00\x80\x3C\x13"
-     "\x00\x80"
-     "\xD3\x13\x00\x80\x4E\x14\x00\x80\x42\x11\x00\x80\x9A\x11\x00\x80\xB4\x11"
-     "\x00\x80"
-     "\xBB\x11\x00\x80\xC7\x11\x00\x80\xEB\x11\x00\x80\xF2\x11\x00\x80\xF9\x11"
-     "\x00\x80"
-     "\x78\x12\x00\x80\xB2\x12\x00\x80\x91\x2C\xF4\x00\xAB\x12\x00\x80\x22\x13"
-     "\x00\x80"
-     "\x0F\x13\x00\x80\x16\x13\x00\x80\x4D\x13\x00\x80\x46\x13\x00\x80\x54\x13"
-     "\x00\x80"
-     "\xC0\x13\x00\x80\x16\x14\x00\x80\x7E\x13\x00\x80\x72\x13\x00\x80\x9B\x13"
-     "\x00\x80"
-     "\xA2\x13\x00\x80\x34\x14\x00\x80\x47\x14\x00\x80\x80\x59\x03\x00\x81\x59"
-     "\x03\x00"
-     "\x82\x69\x03\x00\x83\x69\x03\x00\x84\x79\x03\x00\x85\x79\x03\x00\x86\x69"
-     "\x03\x00"
-     "\x87\x69\x03\x00\x88\x59\x03\x00\x89\x59\x03\x00\x8A\xA9\x03\x00\x8B\xA9"
-     "\x03\x00"
-     "\x8C\x29\x03\x00\x8D\xBD\x03\x00\x8E\xB5\x03\x00\x8F\xAD\x03\x00\x90\xD5"
-     "\x03\x00"
-     "\x91\x69\x03\x00\x92\x79\x03\x00\x93\x79\x03\x00\x94\x69\x03\x00\x95\x69"
-     "\x03\x00"
-     "\x96\x19\x03\x00\x97\x19\x03\x00\x98\x29\x03\x00\x99\x29\x03\x00\x9A\x39"
-     "\x03\x00"
-     "\x9B\x39\x03\x00\x9C\x29\x03\x00\x9D\x29\x03\x00\x9E\x19\x03\x00\x9F\x19"
-     "\x03\x00"
-     "\xA0\xE9\x03\x00\xA1\xE9\x03\x00\xA2\xF9\x03\x00\xA3\xF9\x03\x00\xA4\xE9"
-     "\x03\x00"
-     "\xA5\xE9\x03\x00\xA6\x19\x03\x00\xA7\x19\x03\x00\xA8\xB9\x03\x00\xA9\x2D"
-     "\x03\x00"
-     "\xAA\x25\x03\x00\xAB\x3D\x03\x00\xAC\x25\x03\x00\xAD\xC9\x03\x00\xAE\x39"
-     "\x03\x00"
-     "\xAF\x39\x03\x00\xB0\xD9\x03\x00\xB1\xD1\x03\x00\xB2\xD5\x03\x00\xB3\x35"
-     "\x0B\x00"
-     "\xB4\x1D\x0B\x00\xB5\x05\x0B\x00\xB6\x0D\x0B\x00\xB7\xA1\x03\x00\xB8\x61"
-     "\x03\x00"
-     "\xB9\x61\x03\x00\xBA\x11\x0B\x00\xBB\x11\x0B\x00\xBC\x69\x03\x00\xBD\x6D"
-     "\x03\x00"
-     "\x6F\x14\x00\x80\x76\x14\x00\x80\xB0\xFD\x02\x00\xB1\xFD\x01\x00\xB2\xF1"
-     "\x01\x00"
-     "\xB3\xF1\x01\x00\xB4\xD1\x00\x00\xB5\xD1\x00\x00\xB6\xD5\x00\x00\xB7\xD1"
-     "\x00\x00"
-     "\xB8\xF1\x00\x00\xB9\xF1\x00\x00\xBA\xF1\x00\x00\xBB\xF1\x00\x00\xBC\xD1"
-     "\x00\x00"
-     "\xBD\xD5\x00\x00\xBE\xDD\x00\x00\xBF\xD5\x00\x00\xA0\x95\x02\x00\xA1\x9D"
-     "\x02\x00"
-     "\xA2\x95\x02\x00\xA3\x91\x02\x00\xA4\xB1\x02\x00\xA5\xB5\x02\x00\xA6\xBD"
-     "\x02\x00"
-     "\xA7\xB5\x02\x00\xA8\x89\x02\x00\xA9\x89\x02\x00\xAA\x99\x02\x00\xAB\x9D"
-     "\x02\x00"
-     "\xAC\x85\x02\x00\xAD\x81\x02\x00\xAE\x81\x02\x00\xAF\x81\x02\x00\x90\x1D"
-     "\x02\x00"
-     "\x91\x25\x02\x00\x92\x29\x02\x00\x93\x29\x02\x00\x94\x39\x02\x00\x95\x39"
-     "\x02\x00"
-     "\x96\x29\x02\x00\x97\x29\x02\x00\x98\x19\x02\x00\x99\x1D\x02\x00\x9A\x15"
-     "\x02\x00"
-     "\x9B\x69\x02\x00\x9C\x79\x02\x00\x9D\x79\x02\x00\x9E\x69\x02\x00\x9F\x69"
-     "\x02\x00"
-     "\x7D\x14\x00\x80\x9E\x14\x00\x80\xCE\x13\x00\x80\xB3\x14\x00\x80\xBA\x14"
-     "\x00\x80"
-     "\xDF\x14\x00\x80\xED\x14\x00\x80\xE6\x14\x00\x80\x43\x15\x00\x80\x74\x15"
-     "\x00\x80"
-     "\x7B\x15\x00\x80\x02\x16\x00\x80\x10\x16\x00\x80\x09\x16\x00\x80\xE8\x15"
-     "\x00\x80"
-     "\x1C\x16\x00\x80\x28\x16\x00\x80\x73\x16\x00\x80\x7A\x16\x00\x80\x8A\x16"
-     "\x00\x80"
-     "\x52\x16\x00\x80\x60\x16\x00\x80\x96\x16\x00\x80\x9D\x16\x00\x80\x59\x16"
-     "\x00\x80"
-     "\x42\x17\x00\x80\x49\x17\x00\x80\xEA\x16\x00\x80\xF1\x16\x00\x80\xD0\x16"
-     "\x00\x80"
-     "\xD7\x16\x00\x80\x1A\x17\x00\x80\x13\x17\x00\x80\xA7\x17\x00\x80\x65\x11"
-     "\x00\x80"
-     "\xA8\x11\x00\x80\xA1\x11\x00\x80\xD5\x11\x00\x80\xCE\x11\x00\x80\x07\x12"
-     "\x00\x80"
-     "\x00\x12\x00\x80\x86\x12\x00\x80\x55\x12\x00\x80\x7F\x12\x00\x80\x2E\x13"
-     "\x00\x80"
-     "\xC7\x13\x00\x80\x85\x13\x00\x80\x3B\x14\x00\x80\xF6\x15\x00\x80\x2F\x16"
-     "\x00\x80"
-     "\x8C\x17\x00\x80\x80\x17\x00\x80\xAE\x17\x00\x80\xB0\x16\x00\x80\x82\x15"
-     "\x00\x80"
-     "\xF8\x16\x00\x80\xC1\x14\x00\x80\xBB\x15\x00\x80\xC9\x16\x00\x80\xA4\x16"
-     "\x00\x80"
-     "\xEF\x15\x00\x80\xB4\x15\x00\x80\xD9\xD4\xA7\x02\x5C\x12\x00\x80\x9F\x12"
-     "\x00\x80"
-     "\x67\x16\x00\x80\x4F\x15\x00\x80\x59\x11\x00\x80\x35\x13\x00\x80\xDE\x16"
-     "\x00\x80"
-     "\x0A\x14\x00\x80\x68\x15\x00\x80\x39\x11\x00\x80\x30\x11\x00\x80\x81\x16"
-     "\x00\x80"
-     "\xD8\x13\x00\x80\xED\x12\x00\x80\xA5\x14\x00\x80\x62\x17\x00\x80\xE1\x13"
-     "\x00\x80"
-     "\x80\xC5\x01\x00\x81\xCD\x01\x00\x82\xC5\x01\x00\x83\xB9\x01\x00\x84\xD5"
-     "\xFE\x00"
-     "\x85\xF1\xFC\x00\x86\xC9\x01\x00\x87\xC9\x01\x00\xE8\x13\x00\x80\x63\x12"
-     "\x00\x80"
-     "\xFB\x12\x00\x80\xA5\x03\x00\x80\x0C\x24\x00\x80\x10\x24\x00\x80\x9F\x06"
-     "\x00\x80"
-     "\xC6\x00\x00\x80\xF0\xC4\xFE\x00\xB4\x06\x00\x80\xAA\x24\x00\x80\xAE\x24"
-     "\x00\x80"
-     "\xB4\x22\x00\x80\xB0\x22\x00\x80\xE5\x07\x00\x80\x7D\x09\x00\x80\x81\x09"
-     "\x00\x80"
-     "\x40\x24\x00\x80\xA6\x24\x00\x80\x34\x24\x00\x80\x38\x24\x00\x80\x1C\x24"
-     "\x00\x80"
-     "\x20\x24\x00\x80\x24\x24\x00\x80\x28\x24\x00\x80\x2C\x24\x00\x80\x30\x24"
-     "\x00\x80"
-     "\xDD\x07\x00\x80\xE1\x07\x00\x80\xD0\x00\x00\x80\xD6\x00\x00\x80\xA1\x03"
-     "\x00\x80"
-     "\xA7\x03\x00\x80\xDF\x07\x00\x80\xD2\x00\x00\x80\xD4\x00\x00\x80\xBC\x06"
-     "\x00\x80"
-     "\x78\x00\x00\x80\xFE\x14\x00\x80\x7C\x00\x00\x80\xB0\x19\x01\x00\xB1\x19"
-     "\x01\x00"
-     "\xB2\x29\x01\x00\xB3\x29\x01\x00\xB4\x39\x01\x00\xB5\x39\x01\x00\xB6\x29"
-     "\x01\x00"
-     "\xB7\x29\x01\x00\xB8\x19\x01\x00\xB9\x19\x01\x00\xBA\xC9\x00\x00\xBB\xC9"
-     "\x00\x00"
-     "\xBC\xD9\x00\x00\x86\x00\x00\x80\x90\x00\x00\x80\x03\x15\x00\x80\xAE\x89"
-     "\x02\x00"
-     "\xAF\x89\x02\x00\xAC\x89\x02\x00\xAD\x89\x02\x00\xAA\x69\x02\x00\xAB\x69"
-     "\x02\x00"
-     "\xA8\x69\x02\x00\xA9\x69\x02\x00\xA6\x49\x02\x00\xA7\x49\x02\x00\xA4\x49"
-     "\x02\x00"
-     "\xA5\x49\x02\x00\xA2\x69\x02\x00\xA3\x69\x02\x00\xA0\x69\x02\x00\xA1\x69"
-     "\x02\x00"
-     "\xBE\xFD\x02\x00\xBF\x85\x02\x00\xBC\xF9\x02\x00\xBD\xE5\x02\x00\xBA\xE1"
-     "\x02\x00"
-     "\xBB\xE1\x02\x00\xB8\xF1\x02\x00\xB9\xF1\x02\x00\xB6\xC1\x02\x00\xB7\xC1"
-     "\x02\x00"
-     "\xB4\x8D\xFD\x00\xB5\x81\xFD\x00\xB2\xE9\x02\x00\xB3\xE9\x02\x00\xB0\xE9"
-     "\x02\x00"
-     "\xB1\xE9\x02\x00\x8E\x99\x02\x00\x8F\x99\x02\x00\x8C\x99\x02\x00\x8D\x99"
-     "\x02\x00"
-     "\x8A\x99\x02\x00\x8B\x99\x02\x00\x88\x99\x02\x00\x89\x99\x02\x00\x86\x5D"
-     "\xFD\x00"
-     "\x87\xA5\x02\x00\x84\x51\xFD\x00\x85\x51\xFD\x00\x82\x45\xFD\x00\x83\x4D"
-     "\xFD\x00"
-     "\x80\x55\xFD\x00\x81\x4D\xFD\x00\x9E\x89\x02\x00\x9F\x89\x02\x00\x9C\x8D"
-     "\x02\x00"
-     "\x9D\x89\x02\x00\x9A\xED\x02\x00\x9B\x95\x02\x00\x98\xED\x02\x00\x99\xE5"
-     "\x02\x00"
-     "\x96\xD1\x02\x00\x97\xD1\x02\x00\x94\xC1\x02\x00\x95\xC1\x02\x00\x92\xD1"
-     "\x02\x00"
-     "\x93\xD1\x02\x00\x9A\x00\x00\x80\x0F\x15\x00\x80\x90\x2D\x02\x00\x91\x35"
-     "\x02\x00"
-     "\x92\x3D\x02\x00\x93\x35\x02\x00\x94\x69\x95\x02\x95\x29\x02\x00\x96\x5D"
-     "\x02\x00"
-     "\x97\x55\x02\x00\x98\x6D\x02\x00\x99\x29\xB6\x00\xA4\x00\x00\x80\x1B\x15"
-     "\x00\x80"
-     "\xAE\x00\x00\x80\x27\x15\x00\x80\xB8\x00\x00\x80\x2C\x15\x00\x80\xC6\x10"
-     "\x00\x80"
-     "\x4C\x11\x00\x80\x8F\x11\x00\x80\x92\x11\x00\x80\x22\x12\x00\x80\x4D\x12"
-     "\x00\x80"
-     "\x97\x12\x00\x80\xC3\x12\x00\x80\xDF\x12\x00\x80\xE2\x12\x00\x80\xEA\x12"
-     "\x00\x80"
-     "\x04\x13\x00\x80\x07\x13\x00\x80\x6A\x13\x00\x80\xB8\x13\x00\x80\x27\x14"
-     "\x00\x80"
-     "\xB0\x85\xB6\x00\xB1\xD5\x02\x00\xB2\xDD\x02\x00\xB3\xD5\x02\x00\xB4\xC9"
-     "\x02\x00"
-     "\xB5\x19\xB6\x00\xB6\x09\xB6\x00\xB7\xC1\x02\x00\xB8\xC1\x02\x00\xB9\xC1"
-     "\x02\x00"
-     "\xBA\xC1\x02\x00\xBB\xC1\x02\x00\xBC\xC1\x02\x00\xBD\xC1\x02\x00\xBE\xC1"
-     "\x02\x00"
-     "\xBF\xD9\xB9\x00\x62\x14\x00\x80\x8E\x14\x00\x80\x96\x14\x00\x80\xD2\x14"
-     "\x00\x80"
-     "\x31\x15\x00\x80\x60\x15\x00\x80\x93\x15\x00\x80\x2D\x11\x00\x80\xCC\x15"
-     "\x00\x80"
-     "\x45\x16\x00\x80\xC1\x16\x00\x80\x35\x17\x00\x80\xCF\x15\x00\x80\xD4\x15"
-     "\x00\x80"
-     "\xD9\x15\x00\x80\xDE\x15\x00\x80\xA7\x06\x00\x80\xB2\x06\x00\x80\xBC\xF5"
-     "\x07\x00"
-     "\x5F\x17\x00\x80\xBA\xE9\x07\x00\xBB\xE9\x07\x00\xB8\xED\x07\x00\xB9\xE5"
-     "\x07\x00"
-     "\xB6\xD1\x07\x00\xB7\xD1\x07\x00\xB4\x15\x0C\x00\xB5\xDD\x07\x00\xB2\x0D"
-     "\x0C\x00"
-     "\xB3\x09\x0C\x00\xB0\x15\xF1\x00\xB1\x01\x0C\x00\xAE\x4D\x00\x00\xAF\x69"
-     "\xF1\x00"
-     "\xAC\xA5\x07\x00\xAD\x41\x00\x00\xAA\xBD\x07\x00\xAB\xB9\x07\x00\xA8\xB1"
-     "\x07\x00"
-     "\xA9\xB1\x07\x00\xA6\x89\x07\x00\xA7\x8D\x07\x00\xA4\x8D\x07\x00\xA5\x85"
-     "\x07\x00"
-     "\xA2\x95\x07\x00\xA3\x91\x07\x00\xA0\x99\x07\x00\xA1\x99\x07\x00\x9E\x61"
-     "\x07\x00"
-     "\x9F\x65\x07\x00\x9C\x51\x07\x00\x9D\x6D\x07\x00\x9A\x49\x07\x00\x9B\x4D"
-     "\x07\x00"
-     "\x98\x4D\x07\x00\x99\x45\x07\x00\x96\x75\x07\x00\x97\x71\x07\x00\x94\x79"
-     "\x07\x00"
-     "\x95\x79\x07\x00\x92\x61\x07\x00\x93\x65\x07\x00\x90\x75\x07\x00\x91\x6D"
-     "\x07\x00"
-     "\x8E\x0D\x07\x00\x8F\x09\x07\x00\x8C\x01\x07\x00\x8D\x01\x07\x00\x8A\x19"
-     "\x07\x00"
-     "\x8B\x1D\x07\x00\x88\x1D\x07\x00\x89\x15\x07\x00\x86\x25\x07\x00\x87\x21"
-     "\x07\x00"
-     "\x84\x29\x07\x00\x85\x29\x07\x00\x82\x31\x07\x00\x83\x35\x07\x00\x80\x85"
-     "\x07\x00"
-     "\x81\x3D\x07\x00\x80\x41\x06\x00\x81\x15\xA0\x02\x82\x19\xA0\x02\x83\x19"
-     "\xA0\x02"
-     "\x84\x05\xA0\x02\x85\x25\xA0\x02\x86\x29\xA0\x02\x87\x05\xA0\x02\x88\x39"
-     "\xA0\x02"
-     "\x89\x61\xA0\x02\x8A\x6D\xA0\x02\x8B\x69\xA0\x02\x8C\x75\xA0\x02\x8D\xE9"
-     "\x06\x00"
-     "\x8E\xE5\x06\x00\x8F\x79\x06\x00\x90\x05\x06\x00\x91\x01\x06\x00\x92\x0D"
-     "\x06\x00"
-     "\x93\x05\x06\x00\x94\x19\x06\x00\x95\x19\x06\x00\x96\x15\x06\x00\x97\x11"
-     "\x06\x00"
-     "\x98\x2D\x06\x00\x99\x35\x06\x00\x9A\x39\x06\x00\x9B\x3D\x06\x00\x9C\x21"
-     "\x06\x00"
-     "\x9D\x21\x06\x00\x9E\x2D\x06\x00\x9F\x29\x06\x00\xA0\xD5\x06\x00\xA1\xDD"
-     "\x06\x00"
-     "\xA2\xD1\x06\x00\xA3\xD5\x06\x00\xA4\xC9\x06\x00\xA5\xC9\x06\x00\xA6\xC5"
-     "\x06\x00"
-     "\xA7\xC1\x06\x00\xA8\xFD\x06\x00\xA9\xC5\x06\x00\xAA\xC9\x06\x00\xAB\xC9"
-     "\x06\x00"
-     "\xAC\xD5\x06\x00\xAD\xDD\x06\x00\xAE\xD1\x06\x00\xAF\xD1\x06\x00\xB0\xAD"
-     "\x06\x00"
-     "\xB1\xB5\x06\x00\xB2\xB9\x06\x00\xB3\xBD\x06\x00\xB4\xA1\x06\x00\xB5\xA1"
-     "\x06\x00"
-     "\xB6\xAD\x06\x00\xB7\xA9\x06\x00\xB8\x95\x06\x00\xB9\x9D\x06\x00\xBA\x91"
-     "\x06\x00"
-     "\xBB\x95\x06\x00\xBC\x89\x06\x00\xBD\x89\x06\x00\xBE\x85\x06\x00\xBF\x81"
-     "\x06\x00"
-     "\xA1\x9D\x07\x00\xA0\x9D\x07\x00\xA3\xA9\x07\x00\xA2\x8D\xB3\x00\xA5\x15"
-     "\x02\x00"
-     "\xA4\x15\x02\x00\xCE\x00\x00\x80\xA6\x95\xB3\x00\xA9\x89\x07\x00\xA8\x81"
-     "\x07\x00"
-     "\xAB\x99\x07\x00\xAA\x81\x07\x00\xD8\x00\x00\x80\xB2\x03\x00\x80\xE3\x07"
-     "\x00\x80"
-     "\x7F\x09\x00\x80\xB1\xF9\x07\x00\xB0\xF1\x07\x00\x83\x09\x00\x80\xB2\xF1"
-     "\x07\x00"
-     "\xF6\x23\x00\x80\xB4\xA1\x07\x00\xB7\xA5\x07\x00\xB6\xA5\x07\x00\xB9\x9D"
-     "\x04\x00"
-     "\xB8\x9D\x04\x00\xBB\x45\x05\x00\xBA\x45\x05\x00\xBD\x45\x05\x00\xBC\x45"
-     "\x05\x00"
-     "\xBF\x45\x05\x00\xBE\x45\x05\x00\x81\xFD\x07\x00\x80\xED\xBC\x00\x83\x0D"
-     "\x07\x00"
-     "\x82\x0D\x07\x00\xFA\x23\x00\x80\x84\x1D\x07\x00\x87\x15\x07\x00\x80\x2A"
-     "\x00\x80"
-     "\x89\xED\xB3\x00\x88\x35\x07\x00\x8B\xE5\xB3\x00\x8A\xE1\xB3\x00\x8D\xE9"
-     "\x07\x00"
-     "\x8C\xF9\xB3\x00\x8F\xE1\x07\x00\x8E\xE5\x07\x00\x91\xF1\x07\x00\x90\xE9"
-     "\x07\x00"
-     "\x34\x25\x00\x80\x92\xB9\xB2\x00\x95\xE9\x07\x00\x94\xAD\x90\x02\x97\xE5"
-     "\x07\x00"
-     "\x96\x99\x07\x00\x99\x6D\xB3\x00\x98\xB5\x07\x00\x9B\xB5\x07\x00\x9A\x7D"
-     "\xB3\x00"
-     "\x9D\x95\x07\x00\x9C\x95\x07\x00\x9F\x6D\x07\x00\x9E\x95\x07\x00\x82\xA4"
-     "\xF0\x00"
-     "\x6D\x25\x00\x80\x8C\x25\x00\x80\x9E\x25\x00\x80\xC3\x25\x00\x80\xC2\x29"
-     "\x00\x80"
-     "\xDE\x29\x00\x80\xF0\x29\x00\x80\x87\x27\x00\x80\x84\x2A\x00\x80\x68\x2A"
-     "\x00\x80"
-     "\x28\x25\x00\x80\xCC\xC4\x85\x02\xCC\xF8\x85\x02\x2C\x25\x00\x80\x85\x09"
-     "\x00\x80"
-     "\xA5\xE1\x00\x00\xA4\xD1\x00\x00\x88\x09\x00\x80\xA6\xE5\x00\x00\xA1\x19"
-     "\x00\x00"
-     "\xA0\x69\x00\x00\xA3\x29\x93\x02\xA2\xCD\x00\x00\xAD\xB1\x00\x00\xAC\xD1"
-     "\x00\x00"
-     "\xD9\x10\xA0\x02\xAE\xB9\x00\x00\xA9\x01\x08\x00\xA8\xD9\x00\x00\xAB\x79"
-     "\x0B\x00"
-     "\xAA\xCD\x00\x00\x95\x39\xD7\x00\x94\x31\xD7\x00\x97\x41\xD7\x00\x96\x31"
-     "\xD7\x00"
-     "\x91\x09\x00\x80\x8E\x09\x00\x80\x93\x29\xD7\x00\x92\x11\xD7\x00\x8B\x09"
-     "\x00\x80"
-     "\x9C\x79\xD7\x00\xE7\x22\x00\x80\xEA\x23\x00\x80\xEF\x22\x00\x80\xF7\x22"
-     "\x00\x80"
-     "\x9B\x61\xD7\x00\x9A\x71\xD7\x00\x85\x99\xD7\x00\x84\x8D\xD7\x00\x87\xA9"
-     "\xD7\x00"
-     "\x86\x91\xD7\x00\xEE\x23\x00\x80\xF2\x23\x00\x80\x83\xE5\xD7\x00\x82\xE5"
-     "\xD7\x00"
-     "\x8D\x91\xD7\x00\x8C\x89\xD7\x00\x8F\x61\xD7\x00\x8E\x95\xD7\x00\xB7\x0D"
-     "\x00\x80"
-     "\xBA\x0D\x00\x80\x8B\x91\xD7\x00\x8A\x99\xD7\x00\x8F\x35\x0E\x00\x8E\x71"
-     "\xBC\x00"
-     "\x8D\xA9\x09\x00\x8C\x21\x09\x00\x8B\x91\xBD\x00\x8A\x89\x09\x00\x89\x89"
-     "\xBD\x00"
-     "\x88\x91\xBD\x00\x87\x11\x0E\x00\x86\xBD\x09\x00\x85\x85\x09\x00\x84\x85"
-     "\x09\x00"
-     "\x83\xA5\x09\x00\x82\x2D\x0E\x00\x81\x05\x09\x00\x78\x17\x00\x80\x9F\x05"
-     "\x09\x00"
-     "\x9E\x41\x0C\x00\x9D\xC1\xBD\x00\x9C\x4D\x0C\x00\x9B\x29\x9E\x02\x9A\x6D"
-     "\x09\x00"
-     "\x99\x25\xBD\x00\x98\x25\xBD\x00\x97\x15\xBD\x00\x96\x15\xBD\x00\x95\x05"
-     "\xBD\x00"
-     "\x94\x05\xBD\x00\x93\x09\x9C\x02\x92\x01\x9C\x02\x91\xF1\x9C\x02\x90\x19"
-     "\xBD\x00"
-     "\xAF\xA9\xBC\x00\xAE\xA5\xBC\x00\xAD\x91\xBC\x00\xAC\x9D\xBC\x00\xAB\x99"
-     "\xBC\x00"
-     "\xAA\x95\xBC\x00\xA9\x7D\xBD\x00\xA8\x41\xB9\x00\xA7\x09\xB9\x00\xA6\x4D"
-     "\xBD\x00"
-     "\xA5\x09\xB9\x00\xA4\x31\xB9\x00\xA3\x15\xBD\x00\xA2\x2D\xB9\x00\xA1\x35"
-     "\xB9\x00"
-     "\xA0\x0D\x09\x00\xBF\x99\x09\x00\xBE\xFD\x0E\x00\xBD\x59\x09\x00\xBC\x69"
-     "\x09\x00"
-     "\xBB\x5D\x09\x00\xBA\x01\xBD\x00\xB9\xF1\xC8\x00\xB8\x59\xB0\x00\xB7\x91"
-     "\xBC\x00"
-     "\xB6\x65\xB0\x00\xB5\x81\xBC\x00\xB4\xE9\xBC\x00\xB3\xF9\xC8\x00\xB2\xCD"
-     "\xBC\x00"
-     "\xB1\x3D\xBD\x00\xB0\xCD\xBC\x00\x80\x1D\xBD\x00\x81\x69\x9D\x02\x82\x71"
-     "\xBD\x00"
-     "\x83\x35\xBC\x00\x84\x6D\xBD\x00\x85\xE9\xBD\x00\x86\x59\xBC\x00\x87\x1D"
-     "\xBC\x00"
-     "\x88\xE1\x9C\x02\x89\x15\xBC\x00\x8A\xE1\x9C\x02\x8B\x25\xBC\x00\x8C\x71"
-     "\x93\x02"
-     "\x8D\x25\xBC\x00\x8E\xAD\xBC\x00\x8F\xA1\x9D\x02\x90\x59\xBC\x00\x91\x6D"
-     "\xB7\x00"
-     "\x92\x99\x9C\x02\x93\x11\x9C\x02\x94\x71\xBC\x00\x95\x7D\xBC\x00\x96\x7D"
-     "\xBC\x00"
-     "\x97\x81\x9C\x02\x98\x61\x91\x02\x99\x95\x9C\x02\x9A\x21\xBC\x00\x9B\x31"
-     "\x08\x00"
-     "\x9C\x4D\x0F\x00\x9D\x15\x08\x00\x9E\x31\x0F\x00\x9F\x2D\x0F\x00\xA0\x11"
-     "\x0F\x00"
-     "\xA1\xB1\x08\x00\xA2\xF5\x08\x00\xA3\x0D\x08\x00\xA4\xA1\x08\x00\xA5\x09"
-     "\x0F\x00"
-     "\xA6\x05\xD5\x00\xA7\x55\x0F\x00\xA8\xA5\x0E\x00\xA9\xAD\x0E\x00\xAA\xA5"
-     "\x0E\x00"
-     "\xAB\xBD\x0E\x00\xAC\xA5\x0E\x00\xAD\xAD\x0E\x00\xAE\xA5\x0E\x00\xAF\x9D"
-     "\x0E\x00"
-     "\xB0\xE5\x0E\x00\xB1\xAD\xD1\x00\xB2\xA5\xD1\x00\xB3\x85\xD4\x00\xB4\xA1"
-     "\xD1\x00"
-     "\xB5\xA1\xD1\x00\xB6\xD9\xD5\x00\xB7\xD9\xD5\x00\xB8\xE9\xD5\x00\xB9\xE9"
-     "\xD5\x00"
-     "\xBA\xF9\xD5\x00\xBB\xF9\xD5\x00\xBC\xE9\xD5\x00\xBD\xE9\xD5\x00\xBE\xB9"
-     "\xD5\x00"
-     "\xBF\xB9\xD5\x00\x80\x49\xD6\x00\x81\x49\xD6\x00\x82\x59\xD6\x00\x83\x59"
-     "\xD6\x00"
-     "\x84\x49\xD6\x00\x85\x6D\xD2\x00\x86\x65\xD2\x00\x87\x9D\xD2\x00\x88\xA5"
-     "\xD2\x00"
-     "\x89\xAD\xD2\x00\x8A\x75\xD6\x00\x8B\xAD\xD6\x00\x8C\xB5\xD6\x00\x8D\xBD"
-     "\xD6\x00"
-     "\x8E\xB5\xD6\x00\x8F\xB1\xD2\x00\x90\xD1\xD2\x00\x91\xD1\xD2\x00\x92\xD1"
-     "\xD2\x00"
-     "\x93\xD1\xD2\x00\x94\xF1\xD2\x00\x95\xF1\xD2\x00\x96\xF1\xD2\x00\x97\xF1"
-     "\xD2\x00"
-     "\x98\xD1\xD2\x00\x99\xD1\xD2\x00\x9A\xD1\xD2\x00\x9B\xD1\xD2\x00\x9C\xE1"
-     "\xD6\x00"
-     "\x9D\x5D\x0D\x00\x9E\x55\x0D\x00\x9F\x45\x0D\x00\xA0\x61\xDA\x00\xA1\x4D"
-     "\xD9\x00"
-     "\xA2\x89\xDA\x00\xA3\x41\xC5\x00\xA4\x95\xDA\x00\xA5\x9D\xDA\x00\xA6\x95"
-     "\xDA\x00"
-     "\xA7\x8D\xDA\x00\xA8\xB5\xDA\x00\xA9\xE5\xDA\x00\xAA\xB1\xDA\x00\xAB\xB1"
-     "\xDA\x00"
-     "\xAC\x91\xDA\x00\xAD\x91\xDA\x00\xAE\x91\xDA\x00\xAF\x91\xDA\x00\xB0\xF1"
-     "\xDA\x00"
-     "\xB1\x95\xDA\x00\xB2\x9D\xDA\x00\xB3\xB1\xDA\x00\xB4\xE5\xDA\x00\xB5\x55"
-     "\xDA\x00"
-     "\xB6\x5D\xDA\x00\xB7\x4D\xDA\x00\xB8\x75\xDA\x00\xB9\x25\xDA\x00\xBA\x2D"
-     "\xDA\x00"
-     "\xBB\x3D\xDA\x00\xBC\x25\xDA\x00\xBD\xD5\xD9\x00\xBE\xDD\xD9\x00\x80\x0D"
-     "\x00\x80"
-     "\x90\x58\xDC\x00\x91\x08\x05\x00\xCF\x0C\x00\x80\x95\x0D\x00\x80\x5C\x3B"
-     "\x00\x80"
-     "\x9B\x0D\x00\x80\x98\x0D\x00\x80\xD8\x0C\x00\x80\xDB\x0C\x00\x80\xB4\x1B"
-     "\x00\x80"
-     "\xDE\x0C\x00\x80\xEA\x0C\x00\x80\xA7\x0D\x00\x80\x9D\x20\x07\x00\x9E\x10"
-     "\x06\x00"
-     "\x9F\x84\x15\x00\xFC\x0C\x00\x80\xF6\x0C\x00\x80\xF0\x0C\x00\x80\x83\x0D"
-     "\x00\x80"
-     "\x86\x0D\x00\x80\x05\x0D\x00\x80\xD9\x3C\xAC\x02\x56\x0D\x00\x80\xAA\x0D"
-     "\x00\x80"
-     "\xAD\x0D\x00\x80\x14\x0D\x00\x80\xDC\xEA\x00\x80\x50\x3B\x00\x80\x1A\x0D"
-     "\x00\x80"
-     "\xE1\xEA\x00\x80\x6B\x0D\x00\x80\xE6\xEA\x00\x80\x32\x0D\x00\x80\x35\x0D"
-     "\x00\x80"
-     "\x3E\x0D\x00\x80\xEB\xEA\x00\x80\x44\x0D\x00\x80\x47\x0D\x00\x80\x4A\x0D"
-     "\x00\x80"
-     "\xA4\x0D\x00\x80\x9E\x0D\x00\x80\x60\x3B\x00\x80\xA1\x0D\x00\x80\x59\x0D"
-     "\x00\x80"
-     "\x04\x24\x00\x80\x6E\x0D\x00\x80\x8F\x0D\x00\x80\x92\x0D\x00\x80\x7D\x0D"
-     "\x00\x80"
-     "\xB2\x0B\x00\x80\xB5\x0B\x00\x80\xB8\x0B\x00\x80\xF0\xEA\x00\x80\xF5\xEA"
-     "\x00\x80"
-     "\x02\xEA\x00\x80\x07\xEA\x00\x80\x82\xE4\x02\x00\xBA\x05\x00\x00\x0C\xEA"
-     "\x00\x80"
-     "\xF0\xB8\x05\x00\xB1\xC4\x03\x00\xB2\x14\x04\x00\xAF\xF8\x2E\x00\xF3\x22"
-     "\x00\x80"
-     "\x82\x40\x02\x00\x02\x0A\x00\x80\x91\x08\x02\x00\x35\x0A\x00\x80\x84\x90"
-     "\x02\x00"
-     "\xA7\x3D\x00\x00\x84\x2C\x05\x00\x99\x24\xD7\x00\x11\xEA\x00\x80\x9B\x20"
-     "\xFF\x00"
-     "\xF0\x58\x02\x00\x91\x14\x02\x00\x84\x94\x02\x00\x8D\x68\x02\x00\xA7\x75"
-     "\x00\x00"
-     "\xCC\xCC\xBC\x00\x92\xD4\x02\x00\x16\xEA\x00\x80\x87\xDC\x03\x00\xF0\x4C"
-     "\x02\x00"
-     "\x96\x74\x02\x00\x91\x0C\x02\x00\xBE\x35\x00\x00\x8D\x50\x02\x00\x8C\xC0"
-     "\x02\x00"
-     "\x1B\xEA\x00\x80\x97\x1D\x00\x00\x20\xEA\x00\x80\xB9\xCC\x03\x00\xF0\x40"
-     "\x02\x00"
-     "\x91\xB4\x02\x00\x92\xCC\x02\x00\x2A\xEA\x00\x80\x25\xEA\x00\x80\xB0\x09"
-     "\x00\x00"
-     "\x2F\xEA\x00\x80\xF0\x48\x02\x00\x91\x60\x02\x00\xA4\x9C\x02\x00\x96\xB4"
-     "\x02\x00"
-     "\xBA\x2D\x00\x00\xAF\x05\x00\x00\x34\xEA\x00\x80\x91\x54\x02\x00\xF0\x40"
-     "\x02\x00"
-     "\xBD\x29\x00\x00\x96\xCC\x02\x00\xB8\x94\x03\x00\xB9\xB0\x03\x00\x39\xEA"
-     "\x00\x80"
-     "\xAF\x05\x00\x00\xA5\x50\x0E\x00\xB5\xCC\x03\x00\xF0\x40\x02\x00\x91\xA4"
-     "\x02\x00"
-     "\x3E\xEA\x00\x80\x43\xEA\x00\x80\xA4\xC4\x02\x00\xB0\x0D\x00\x00\x4C\xEA"
-     "\x00\x80"
-     "\x55\xEA\x00\x80\x62\xEA\x00\x80\x6F\xEA\x00\x80\x81\xF1\x0A\x00\x7C\xEA"
-     "\x00\x80"
-     "\x83\x01\x01\x00\x82\xF9\x0A\x00\x85\x19\x06\x00\x84\xE9\x07\x00\x87\x15"
-     "\x06\x00"
-     "\x89\xEA\x00\x80\x89\x35\x06\x00\x88\x35\x06\x00\x8B\x35\x06\x00\x8A\x35"
-     "\x06\x00"
-     "\x8D\x15\x06\x00\x8C\x15\x06\x00\x8F\x15\x06\x00\x8E\x15\x06\x00\x91\x01"
-     "\x06\x00"
-     "\x90\x75\x06\x00\x93\x01\x06\x00\x92\x09\x06\x00\x95\xED\xB8\x00\x94\x19"
-     "\x06\x00"
-     "\x97\x0D\x06\x00\x96\x0D\x06\x00\x99\x01\x06\x00\x98\x39\x06\x00\x9B\x01"
-     "\x06\x00"
-     "\x9A\x09\x06\x00\x9D\x01\x06\x00\x9C\x19\x06\x00\x9F\x01\x06\x00\x9E\x09"
-     "\x06\x00"
-     "\xA1\x01\x06\x00\xA0\xF9\x06\x00\xA3\x0D\x06\x00\xA2\x9D\x01\x00\xA5\xB9"
-     "\xB3\x00"
-     "\xA4\x1D\x06\x00\xA7\x09\x06\x00\xA6\x11\x06\x00\xA9\x39\x06\x00\xA8\x31"
-     "\x06\x00"
-     "\xAB\x09\x06\x00\xAA\x31\x06\x00\xAD\x19\x06\x00\xAC\x11\x06\x00\xAF\x09"
-     "\x06\x00"
-     "\xAE\x11\x06\x00\x96\xEA\x00\x80\xB0\x71\x06\x00\xB3\x75\x06\x00\xB2\x75"
-     "\x06\x00"
-     "\xB5\x15\x06\x00\xB4\x15\x06\x00\xB7\x15\x06\x00\xB6\x15\x06\x00\xB9\x35"
-     "\x06\x00"
-     "\xB8\x35\x06\x00\xB9\xEA\x00\x80\xBA\x35\x06\x00\x9F\xEA\x00\x80\xC2\xEA"
-     "\x00\x80"
-     "\xAC\xEA\x00\x80\x80\xE5\x03\x00\x85\x6C\x02\x00\x90\xE5\x12\x00\x87\xF4"
-     "\x01\x00"
-     "\x86\xC0\x02\x00\x9F\xE1\x02\x00\x9E\xF1\x02\x00\xCF\xEA\x00\x80\x86\x00"
-     "\x00\x00"
-     "\xBB\xE1\x03\x00\x88\x00\x00\x00\xD7\x0A\x00\x80\x8A\x00\x00\x00\xBF\xC9"
-     "\x03\x00"
-     "\xBE\xC1\x03\x00\xBD\xC9\x03\x00\xBC\xD5\x03\x00\x95\x40\x19\x00\x94\x00"
-     "\x18\x00"
-     "\x97\x48\x1B\x00\x96\x4C\x1A\x00\x91\x50\x1D\x00\x90\x00\x1C\x00\x93\x58"
-     "\x1F\x00"
-     "\x92\x5C\x1E\x00\x9D\x60\x2D\x00\x9C\x00\x2C\x00\x9F\xC8\x11\x00\x9E\x6C"
-     "\x2E\x00"
-     "\x99\x70\x15\x00\x98\x00\x14\x00\x9B\x78\x17\x00\x9A\x7C\x16\x00\xC3\x0C"
-     "\x00\x80"
-     "\xA0\x00\x00\x00\xE6\x0D\x00\x80\xEC\x0D\x00\x80\xA5\x00\x00\x00\xF2\x0D"
-     "\x00\x80"
-     "\xA7\x00\x00\x00\xFE\x0D\x00\x80\xA9\x00\x00\x00\xA8\x00\x00\x00\x01\x0E"
-     "\x00\x80"
-     "\x04\x0E\x00\x80\x07\x0E\x00\x80\x0A\x0E\x00\x80\x19\x0E\x00\x80\x1F\x0E"
-     "\x00\x80"
-     "\x22\x0E\x00\x80\x25\x0E\x00\x80\xB3\x00\x00\x00\x1B\x23\x00\x80\x4F\x0E"
-     "\x00\x80"
-     "\x5E\x0E\x00\x80\x84\x0E\x00\x80\xB6\x00\x00\x00\xA3\xE5\x03\x00\xA2\x19"
-     "\x02\x00"
-     "\xA1\x11\x02\x00\xA0\x19\x02\x00\x87\x0E\x00\x80\xBC\x00\x00\x00\xBF\x00"
-     "\x00\x00"
-     "\xA4\xE1\x03\x00\x96\x0E\x00\x80\xC0\x00\x00\x00\x99\x0E\x00\x80\xA5\x0E"
-     "\x00\x80"
-     "\x13\x23\x00\x80\xBA\x0E\x00\x80\xC7\x00\x00\x00\xBD\x0E\x00\x80\xCC\x14"
-     "\x98\x02"
-     "\xCD\x10\x98\x02\x80\x68\x02\x00\x81\x30\x13\x00\xCC\x50\xBF\x00\xDB\x0E"
-     "\x00\x80"
-     "\xCC\x1C\x98\x02\xCD\x50\x9E\x02\xDE\x0E\x00\x80\xE4\x0E\x00\x80\xF9\x0E"
-     "\x00\x80"
-     "\xFC\x0E\x00\x80\x05\x0F\x00\x80\xD4\x00\x00\x00\x08\x0F\x00\x80\x0B\x0F"
-     "\x00\x80"
-     "\x0E\x0F\x00\x80\x14\x0F\x00\x80\x17\x0F\x00\x80\x1A\x0F\x00\x80\xEB\x22"
-     "\x00\x80"
-     "\xDC\x00\x00\x00\x20\x0F\x00\x80\x2F\x0F\x00\x80\xB0\xA5\x13\x00\xB1\xB1"
-     "\x00\x00"
-     "\xB2\xC9\x00\x00\xB3\xDD\x13\x00\xB4\xD5\x00\x00\xB5\xCD\x13\x00\xB6\xF9"
-     "\x13\x00"
-     "\xB7\xF9\x13\x00\xB8\xC9\x13\x00\xB9\xD5\x13\x00\xBA\xE1\x00\x00\xBB\xE1"
-     "\x00\x00"
-     "\xBC\xC9\x13\x00\xBD\xE5\x00\x00\xBE\xE1\x00\x00\xBF\xE1\x00\x00\xBA\xC0"
-     "\x29\x00"
-     "\x44\x3B\x00\x80\xB8\xC8\x2E\x00\xB9\x00\x28\x00\x44\x0F\x00\x80\x5F\x0F"
-     "\x00\x80"
-     "\x6B\x0F\x00\x80\x59\x0F\x00\x80\x3B\x0F\x00\x80\x50\x0F\x00\x80\x5C\x0F"
-     "\x00\x80"
-     "\xFA\x00\x00\x00\x40\x3B\x00\x80\x62\x0F\x00\x80\xCE\x17\x00\x80\xFE\x00"
-     "\x00\x00"
-     "\x95\x81\xA8\x00\x94\x6D\xA4\x00\x97\x8D\xA8\x00\x96\x7D\xA4\x00\x91\x29"
-     "\xA4\x00"
-     "\x90\xD1\xA5\x00\x93\x1D\xA4\x00\x92\x15\xD0\x00\x9D\x5D\xA4\x00\x9C\x1D"
-     "\xA5\x00"
-     "\x9F\x69\xA5\x00\x9E\xD1\xA4\x00\x99\xE5\xA5\x00\x98\x1D\xD0\x00\x9B\x41"
-     "\xA4\x00"
-     "\x9A\x51\x84\x02\x85\xA9\xA5\x00\x84\xE5\xA1\x00\x87\x95\xA1\x00\x86\xE5"
-     "\xA1\x00"
-     "\x81\xC9\xA1\x00\x80\xD9\xA1\x00\x83\xC5\xA1\x00\x82\xF9\xA5\x00\x8D\x41"
-     "\xA4\x00"
-     "\x8C\x7D\xA4\x00\x8F\x59\xA4\x00\x8E\x45\xA4\x00\x89\x71\xA4\x00\x88\x91"
-     "\xA5\x00"
-     "\x8B\x69\xA4\x00\x8A\x75\xA4\x00\xB5\x19\xA1\x00\xB4\x09\xA1\x00\xB7\x15"
-     "\xA1\x00"
-     "\xB6\x29\xA5\x00\xB1\x91\x88\x02\xB0\x49\x85\x02\xB3\xD1\xA5\x00\xB2\x6D"
-     "\x85\x02"
-     "\xBD\xA1\xA4\x00\xBC\x41\xA5\x00\xBF\xB9\xA4\x00\xBE\xA5\xA4\x00\xB9\x59"
-     "\xA5\x00"
-     "\xB8\x15\xA1\x00\xBB\x65\xA1\x00\xBA\x15\xA1\x00\xA5\x81\x8A\x02\xA4\xCD"
-     "\xA5\x00"
-     "\xA7\x5D\xA5\x00\xA6\xDD\xA5\x00\xA1\x11\x85\x02\xA0\xD5\xA5\x00\xA3\x11"
-     "\x85\x02"
-     "\xA2\xED\xA5\x00\xAD\xC1\xA5\x00\xAC\xB9\x85\x02\xAF\xCD\xA5\x00\xAE\xC5"
-     "\xA5\x00"
-     "\xA9\xE9\xA5\x00\xA8\x69\x84\x02\xAB\x29\x85\x02\xAA\xD5\xAE\x00\x80\x19"
-     "\xA5\x00"
-     "\x81\x25\xA5\x00\x82\x21\xA5\x00\x83\x3D\xA5\x00\x84\xD5\xA4\x00\x85\x2D"
-     "\xA5\x00"
-     "\x86\x11\xD1\x00\x87\x19\xA5\x00\x88\x49\xA5\x00\x89\xA5\xA9\x00\x8A\x59"
-     "\xA5\x00"
-     "\x8B\xA9\xA9\x00\x8C\x19\xD1\x00\x8D\xE1\xA4\x00\x8E\x55\x85\x02\x8F\x45"
-     "\xA5\x00"
-     "\x90\x79\xA4\x00\x91\x39\xA5\x00\x92\xB5\xA5\x00\x93\x0D\xA4\x00\x94\x51"
-     "\xA4\x00"
-     "\x07\x18\x00\x80\x96\x69\xA4\x00\x97\x95\x84\x02\x98\x69\xA4\x00\x99\x25"
-     "\x8B\x02"
-     "\x9A\x79\xA4\x00\x9B\xF9\xA4\x00\x9C\xED\x85\x02\x9D\x6D\xA4\x00\x9E\x51"
-     "\xAF\x00"
-     "\x9F\xAD\x84\x02\xA0\xDD\x84\x02\xA1\xA5\xA4\x00\xA2\xA1\xA4\x00\xA3\xA9"
-     "\xA4\x00"
-     "\xA4\x4D\x84\x02\xA5\x95\x89\x02\xA6\x69\x84\x02\xA7\xD5\xA4\x00\xA8\x2D"
-     "\xA0\x00"
-     "\xA9\x3D\xA0\x00\xAA\x0D\xA4\x00\xAB\x31\xA0\x00\xAC\x11\xA0\x00\xAD\x5D"
-     "\xA4\x00"
-     "\xAE\x11\xA0\x00\xAF\x61\xA0\x00\xB0\x25\xA4\x00\xB1\xC5\xA5\x00\xB2\xC1"
-     "\xA5\x00"
-     "\xB3\xDD\xA5\x00\xB4\xC9\xA5\x00\xB5\xF5\xA5\x00\xB6\xF1\xA5\x00\xB7\xED"
-     "\xA5\x00"
-     "\xB8\x25\xA4\x00\xB9\xDD\xA5\x00\xBA\xE1\xD1\x00\xBB\xE9\xA5\x00\xBC\x99"
-     "\xA5\x00"
-     "\xBD\x75\xA9\x00\xBE\x89\xA5\x00\xBF\x79\xA9\x00\x80\x29\xD2\x00\x81\xD1"
-     "\xA7\x00"
-     "\x82\x65\x86\x02\x83\x75\xA6\x00\x84\x29\xA7\x00\x85\x69\xA6\x00\x86\xE5"
-     "\xA6\x00"
-     "\x87\x5D\xA7\x00\x88\x21\xA7\x00\x89\xE5\x87\x02\x8A\x19\xA7\x00\x8B\xE5"
-     "\x87\x02"
-     "\x8C\x39\xA7\x00\x8D\x75\x88\x02\x8E\x29\xA7\x00\x8F\xA9\xA7\x00\x90\xDD"
-     "\x86\x02"
-     "\x91\x5D\xA7\x00\x92\x61\xAC\x00\x93\x9D\x87\x02\x94\x0D\x87\x02\x95\x75"
-     "\xA7\x00"
-     "\x96\x71\xA7\x00\x97\x79\xA7\x00\x98\xBD\x87\x02\x99\x65\x8A\x02\x9A\x99"
-     "\x87\x02"
-     "\x9B\x25\xA7\x00\x9C\xFD\xA3\x00\xD1\x17\x00\x80\x9E\xDD\xA7\x00\x9F\xE1"
-     "\xA3\x00"
-     "\xB9\x03\x00\x80\xB6\x03\x00\x80\xA2\x21\xA3\x00\xC9\x03\x00\x80\xEC\x04"
-     "\x00\x80"
-     "\xA5\x95\xA6\x00\xA6\x91\xA6\x00\x8B\x05\x00\x80\xDE\x05\x00\x80\xA9\x85"
-     "\xA6\x00"
-     "\xAA\x81\xA6\x00\xAB\x9D\xA6\x00\xAC\x75\xA7\x00\x19\x06\x00\x80\xAE\xB1"
-     "\xD2\x00"
-     "\xAF\xB9\xA6\x00\xB0\xA9\xA6\x00\xB1\x45\xAA\x00\xB2\xB9\xA6\x00\xB3\x49"
-     "\xAA\x00"
-     "\xB4\xF9\xD2\x00\xB5\x01\xA7\x00\xB6\xB5\x86\x02\xB7\xA5\xA6\x00\xB8\xD9"
-     "\xA7\x00"
-     "\xB9\x99\xA6\x00\x3F\x06\x00\x80\xBB\xAD\xA7\x00\x56\x06\x00\x80\xBD\x35"
-     "\x87\x02"
-     "\xBE\xC9\xA7\x00\xBF\x35\x87\x02\x80\x09\xA6\x00\x81\x45\x89\x02\x82\x19"
-     "\xA6\x00"
-     "\x83\x99\xA6\x00\x73\x06\x00\x80\x85\x0D\xA6\x00\x86\x31\xAD\x00\x87\xCD"
-     "\x86\x02"
-     "\x88\x7D\x86\x02\x89\x05\xA6\x00\x8A\x01\xA6\x00\x8B\x09\xA6\x00\x8C\xED"
-     "\x86\x02"
-     "\x8D\x35\x8B\x02\x8E\xC9\x86\x02\x8F\x75\xA6\x00\x90\xCD\xA2\x00\x91\xDD"
-     "\xA2\x00"
-     "\x92\xED\xA6\x00\x93\xD1\xA2\x00\x94\xF1\xA2\x00\x95\xBD\xA6\x00\x96\xF1"
-     "\xA2\x00"
-     "\x97\x81\xA2\x00\x98\x85\xA6\x00\x99\x65\xA7\x00\x9A\x61\xA7\x00\x9B\x7D"
-     "\xA7\x00"
-     "\x9C\x69\xA7\x00\x9D\x55\xA7\x00\x9E\x51\xA7\x00\x9F\x4D\xA7\x00\xA0\x45"
-     "\xA6\x00"
-     "\xA1\xBD\xA7\x00\xA2\x81\xD3\x00\xA3\x89\xA7\x00\xA4\xF9\xA7\x00\xA5\x15"
-     "\xAB\x00"
-     "\xA6\xE9\xA7\x00\xA7\x19\xAB\x00\xA8\x89\xD3\x00\xA9\x71\xA6\x00\xAA\xC5"
-     "\x87\x02"
-     "\xAB\xD5\xA7\x00\xAC\x89\xA6\x00\xAD\xC9\xA7\x00\xAE\x45\xA7\x00\xAF\xFD"
-     "\xA6\x00"
-     "\xB0\xC1\xA6\x00\xB1\x05\x86\x02\xB2\xF9\xA6\x00\xB3\x05\x86\x02\xB4\xD9"
-     "\xA6\x00"
-     "\xB5\x95\x89\x02\xB6\xC9\xA6\x00\xB7\x49\xA6\x00\xB8\x7D\x87\x02\xB9\xFD"
-     "\xA6\x00"
-     "\xBA\xC1\xAD\x00\xBB\x3D\x86\x02\xBC\xAD\x86\x02\xBD\xD5\xA6\x00\xBE\xD1"
-     "\xA6\x00"
-     "\xBF\xD9\xA6\x00\x91\x21\xA0\x00\x90\x1D\xA0\x00\x93\x39\xA0\x00\x92\x25"
-     "\xA0\x00"
-     "\x8A\x06\x00\x80\x94\xD1\xA1\x00\x97\x1D\xA0\x00\x96\x15\xD4\x00\x99\xA1"
-     "\xAC\x00"
-     "\x98\x4D\xA0\x00\x9B\xAD\xAC\x00\x9A\x5D\xA0\x00\x37\x01\x00\x80\x9C\x1D"
-     "\xD4\x00"
-     "\x9F\x41\xA0\x00\x9E\x51\x80\x02\x81\x41\x8C\x02\x80\x99\x81\x02\x83\x01"
-     "\xA1\x00"
-     "\x82\xBD\x81\x02\x85\xC9\xA5\x00\x84\xD9\xA5\x00\x87\xC5\xA5\x00\x3B\x01"
-     "\x00\x80"
-     "\x89\x89\xA1\x00\x88\xC5\xA5\x00\x3F\x01\x00\x80\x8A\xC5\xA5\x00\x8D\x71"
-     "\xA0\x00"
-     "\x43\x01\x00\x80\x8F\x69\xA0\x00\x8E\x75\xA0\x00\xB1\xA1\xA1\x00\xB0\xD9"
-     "\x81\x02"
-     "\xB3\xAD\xA1\x00\xB2\xA5\xA1\x00\xB5\x91\x8C\x02\xB4\x49\x81\x02\xB7\xD1"
-     "\xA1\x00"
-     "\xB6\x6D\x81\x02\xB9\x39\xA5\x00\xB8\x29\xA5\x00\xBB\x35\xA5\x00\x47\x01"
-     "\x00\x80"
-     "\xBD\x59\xA1\x00\xBC\x15\xA5\x00\x4B\x01\x00\x80\xBE\x15\xA5\x00\xA1\xBD"
-     "\xA0\x00"
-     "\xA0\xFD\xA1\x00\xA3\x89\xA1\x00\xA2\x31\xA0\x00\xA5\x11\x81\x02\xA4\xD5"
-     "\xA1\x00"
-     "\xA7\x11\x81\x02\xA6\xED\xA1\x00\xA9\xA1\x8E\x02\xA8\xED\xA1\x00\xAB\x7D"
-     "\xA1\x00"
-     "\xAA\xFD\xA1\x00\xAD\xE9\xA1\x00\xAC\x69\x80\x02\xAF\x29\x81\x02\xAE\xD5"
-     "\xAA\x00"
-     "\x80\xE5\xA0\x00\x81\x05\xA1\x00\x82\x01\xA1\x00\x83\x1D\xA1\x00\x84\x09"
-     "\xA1\x00"
-     "\x4F\x01\x00\x80\x86\x31\xA1\x00\x53\x01\x00\x80\x57\x01\x00\x80\x5B\x01"
-     "\x00\x80"
-     "\x8A\x21\xD5\x00\x8B\x29\xA1\x00\x8C\x59\xA1\x00\x8D\xB5\xAD\x00\x8E\x49"
-     "\xA1\x00"
-     "\x8F\xB9\xAD\x00\x90\x69\xD5\x00\x5F\x01\x00\x80\x92\x25\x81\x02\x93\x35"
-     "\xA1\x00"
-     "\x94\x69\xA0\x00\x95\x29\xA1\x00\x96\xA5\xA1\x00\x97\x1D\xA0\x00\x98\x61"
-     "\xA0\x00"
-     "\x99\xA5\x80\x02\x9A\x59\xA0\x00\x9B\xA5\x80\x02\x9C\x79\xA0\x00\x9D\x35"
-     "\x8F\x02"
-     "\x9E\x69\xA0\x00\x9F\xE9\xA0\x00\xA0\x1D\x81\x02\xA1\x9D\xA0\x00\xA2\xA1"
-     "\xAB\x00"
-     "\xA3\x5D\x80\x02\xA4\xCD\x80\x02\xA5\xB5\xA0\x00\xA6\xB1\xA0\x00\xA7\xB9"
-     "\xA0\x00"
-     "\xA8\x7D\x80\x02\xA9\xA5\x8D\x02\xAA\x59\x80\x02\xAB\xE5\xA0\x00\xAC\x3D"
-     "\xA4\x00"
-     "\xAD\x2D\xA4\x00\xAE\x1D\xA0\x00\xAF\x21\xA4\x00\xB0\x61\xA4\x00\xB1\x2D"
-     "\xA0\x00"
-     "\xB2\x61\xA4\x00\xB3\x11\xA4\x00\xB4\x35\xA0\x00\xB5\xD5\xA1\x00\xB6\xD1"
-     "\xA1\x00"
-     "\xB7\xCD\xA1\x00\xB8\xF9\xA1\x00\xB9\xC5\xA1\x00\xBA\xC1\xA1\x00\xBB\xDD"
-     "\xA1\x00"
-     "\xBC\x35\xA0\x00\xBD\xCD\xA1\x00\xBE\xF1\xD5\x00\xBF\xF9\xA1\x00\x80\x69"
-     "\xA2\x00"
-     "\x81\x85\xAE\x00\x82\x79\xA2\x00\x83\x89\xAE\x00\x84\x39\xD6\x00\x85\xC1"
-     "\xA3\x00"
-     "\x86\x75\x82\x02\x87\x65\xA2\x00\x88\x19\xA3\x00\x89\x59\xA2\x00\x8A\xD5"
-     "\xA2\x00"
-     "\x8B\x6D\xA3\x00\x8C\x31\xA3\x00\x8D\xF5\x83\x02\x8E\x09\xA3\x00\x8F\xF5"
-     "\x83\x02"
-     "\x90\x49\xA3\x00\x91\x05\x8C\x02\x92\x59\xA3\x00\x93\xD9\xA3\x00\x94\xCD"
-     "\x82\x02"
-     "\x95\x4D\xA3\x00\x96\x71\xA8\x00\x97\x8D\x83\x02\x98\x3D\x83\x02\x99\x45"
-     "\xA3\x00"
-     "\x9A\x41\xA3\x00\x9B\x49\xA3\x00\x9C\xAD\x83\x02\x9D\x75\x8E\x02\x9E\x89"
-     "\x83\x02"
-     "\x9F\x35\xA3\x00\xA0\x0D\xA7\x00\xA1\x1D\xA7\x00\xA2\x2D\xA3\x00\xA3\x11"
-     "\xA7\x00"
-     "\xA4\x31\xA7\x00\xA5\x7D\xA3\x00\xA6\x31\xA7\x00\xA7\x41\xA7\x00\xA8\x45"
-     "\xA3\x00"
-     "\xA9\xA5\xA2\x00\xAA\xA1\xA2\x00\xAB\xBD\xA2\x00\xAC\xA9\xA2\x00\xAD\x95"
-     "\xA2\x00"
-     "\xAE\x91\xA2\x00\xAF\x8D\xA2\x00\xB0\x05\xA3\x00\xB1\xFD\xA2\x00\xB2\xC1"
-     "\xD6\x00"
-     "\xB3\xC9\xA2\x00\xB4\xB9\xA2\x00\xB5\x55\xAE\x00\xB6\xA9\xA2\x00\xB7\x59"
-     "\xAE\x00"
-     "\xB8\xC9\xD6\x00\xB9\x31\xA3\x00\xBA\x85\x82\x02\xBB\x95\xA2\x00\xBC\xC9"
-     "\xA3\x00"
-     "\xBD\x89\xA2\x00\xBE\x05\xA2\x00\xBF\xBD\xA3\x00\x80\x01\xA2\x00\x81\xC5"
-     "\x82\x02"
-     "\x82\x39\xA2\x00\x83\xC5\x82\x02\x84\x19\xA2\x00\x85\x55\x8D\x02\x86\x09"
-     "\xA2\x00"
-     "\x87\x89\xA2\x00\x88\xBD\x83\x02\x89\x3D\xA2\x00\x8A\x01\xA9\x00\x8B\xFD"
-     "\x82\x02"
-     "\x8C\x6D\x82\x02\x8D\x15\xA2\x00\x8E\x11\xA2\x00\x8F\x19\xA2\x00\x90\x9D"
-     "\x82\x02"
-     "\x91\x45\x8F\x02\x92\xB9\x82\x02\x93\x05\xA2\x00\x94\xDD\xA6\x00\x95\xCD"
-     "\xA6\x00"
-     "\x96\xFD\xA2\x00\x97\xC1\xA6\x00\x98\xC1\xA6\x00\x99\x8D\xA2\x00\x9A\xC1"
-     "\xA6\x00"
-     "\x9B\xB1\xA6\x00\x9C\x95\xA2\x00\x9D\x75\xA3\x00\x9E\x71\xA3\x00\x9F\x6D"
-     "\xA3\x00"
-     "\xA0\x99\xA3\x00\xA1\xA5\xA3\x00\xA2\xA1\xA3\x00\xA3\xBD\xA3\x00\xA4\x55"
-     "\xA2\x00"
-     "\xA5\xAD\xA3\x00\xA6\x91\xD7\x00\xA7\x99\xA3\x00\xA8\xC9\xA3\x00\xA9\x25"
-     "\xAF\x00"
-     "\xAA\xD9\xA3\x00\xAB\x29\xAF\x00\xAC\x99\xD7\x00\xAD\x61\xA2\x00\xAE\xD5"
-     "\x83\x02"
-     "\xAF\xC5\xA3\x00\xB0\xF9\xA2\x00\xB1\xB9\xA3\x00\xB2\x35\xA3\x00\xB3\x8D"
-     "\xA2\x00"
-     "\xB4\xD1\xA2\x00\xB5\x15\x82\x02\xB6\xE9\xA2\x00\xB7\x15\x82\x02\xB8\xE9"
-     "\xA2\x00"
-     "\xB9\xA5\x8D\x02\xBA\xF9\xA2\x00\xBB\x79\xA2\x00\xBC\x6D\x83\x02\xBD\xED"
-     "\xA2\x00"
-     "\xBE\xD1\xA9\x00\xBF\x2D\x82\x02\x9D\x81\xA0\x00\x9C\x6D\xAC\x00\x9F\x8D"
-     "\xA0\x00"
-     "\x9E\x7D\xAC\x00\x99\x29\xAC\x00\x98\xD1\xAD\x00\x9B\x1D\xAC\x00\x9A\x15"
-     "\xD8\x00"
-     "\x95\x01\xAC\x00\x94\x3D\xAC\x00\x97\x19\xAC\x00\x96\x05\xAC\x00\x91\x31"
-     "\xAC\x00"
-     "\x90\xD1\xAD\x00\x93\x29\xAC\x00\x92\x35\xAC\x00\x8D\xA9\xAD\x00\x8C\xE5"
-     "\xA9\x00"
-     "\x8F\x95\xA9\x00\x8E\xE5\xA9\x00\x89\xC9\xA9\x00\x88\xD9\xA9\x00\x8B\xC5"
-     "\xA9\x00"
-     "\x8A\xF9\xAD\x00\x85\x61\x80\x02\x84\xB9\x8D\x02\x87\x21\xAD\x00\x86\x9D"
-     "\x8D\x02"
-     "\x81\x51\xAD\x00\x80\x29\x8D\x02\x83\x5D\xAD\x00\x82\x55\xAD\x00\xBD\x19"
-     "\xA9\x00"
-     "\xBC\x09\xA9\x00\xBF\x15\xA9\x00\xBE\x29\xAD\x00\xB9\x91\x80\x02\xB8\x49"
-     "\x8D\x02"
-     "\xBB\xD1\xAD\x00\xBA\x6D\x8D\x02\xB5\x81\xAD\x00\xB4\xF9\x8D\x02\xB7\x8D"
-     "\xAD\x00"
-     "\xB6\x85\xAD\x00\xB1\xA9\xAD\x00\xB0\x29\x8C\x02\xB3\x69\x8D\x02\xB2\x95"
-     "\xA6\x00"
-     "\xAD\x81\x82\x02\xAC\xCD\xAD\x00\xAF\x5D\xAD\x00\xAE\xDD\xAD\x00\xA9\x11"
-     "\x8D\x02"
-     "\xA8\xD5\xAD\x00\xAB\x11\x8D\x02\xAA\xED\xAD\x00\xA5\x9D\xAC\x00\xA4\xDD"
-     "\xAD\x00"
-     "\xA7\xA9\xAD\x00\xA6\x11\xAC\x00\xA1\x25\xAD\x00\xA0\xDD\xD8\x00\xA3\x81"
-     "\xAC\x00"
-     "\xA2\x91\x8C\x02\x80\xA1\xA8\x00\x81\xED\xAC\x00\x82\xA1\xA8\x00\x83\xD1"
-     "\xA8\x00"
-     "\x84\xF5\xAC\x00\x85\x15\xAD\x00\x86\x11\xAD\x00\x87\x0D\xAD\x00\x88\x39"
-     "\xAD\x00"
-     "\x89\x05\xAD\x00\x8A\x01\xAD\x00\x8B\x1D\xAD\x00\x8C\xF5\xAC\x00\x8D\x0D"
-     "\xAD\x00"
-     "\x8E\x31\xD9\x00\x8F\x39\xAD\x00\x90\x29\xAD\x00\x91\xC5\xA1\x00\x92\x39"
-     "\xAD\x00"
-     "\x93\xC9\xA1\x00\x94\x79\xD9\x00\x95\x81\xAC\x00\x96\x35\x8D\x02\x97\x25"
-     "\xAD\x00"
-     "\x98\x59\xAC\x00\x99\x19\xAD\x00\x9A\x95\xAD\x00\x9B\x2D\xAC\x00\x9C\x71"
-     "\xAC\x00"
-     "\x9D\xB5\x8C\x02\x9E\x49\xAC\x00\x9F\xB5\x8C\x02\xA0\x89\xAC\x00\xA1\xC5"
-     "\x83\x02"
-     "\xA2\x99\xAC\x00\xA3\x19\xAC\x00\xA4\x0D\x8D\x02\xA5\x8D\xAC\x00\xA6\xB1"
-     "\xA7\x00"
-     "\xA7\x4D\x8C\x02\xA8\xFD\x8C\x02\xA9\x85\xAC\x00\xAA\x81\xAC\x00\xAB\x89"
-     "\xAC\x00"
-     "\xAC\x6D\x8C\x02\xAD\xB5\x81\x02\xAE\x49\x8C\x02\xAF\xF5\xAC\x00\xB0\x4D"
-     "\xA8\x00"
-     "\xB1\x5D\xA8\x00\xB2\x6D\xAC\x00\xB3\x51\xA8\x00\xB4\x71\xA8\x00\xB5\x3D"
-     "\xAC\x00"
-     "\xB6\x71\xA8\x00\xB7\x01\xA8\x00\xB8\x05\xAC\x00\xB9\xE5\xAD\x00\xBA\xE1"
-     "\xAD\x00"
-     "\xBB\xFD\xAD\x00\xBC\xE9\xAD\x00\xBD\xD5\xAD\x00\xBE\xD1\xAD\x00\xBF\xCD"
-     "\xAD\x00"
-     "\x80\xC5\xAF\x00\x81\x3D\xAE\x00\x82\x01\xDA\x00\x83\x09\xAE\x00\x84\x79"
-     "\xAE\x00"
-     "\x85\x95\xA2\x00\x86\x69\xAE\x00\x87\x99\xA2\x00\x88\x09\xDA\x00\x89\xF1"
-     "\xAF\x00"
-     "\x8A\x45\x8E\x02\x8B\x55\xAE\x00\x8C\x09\xAF\x00\x8D\x49\xAE\x00\x8E\xC5"
-     "\xAE\x00"
-     "\x8F\x7D\xAF\x00\x90\x41\xAF\x00\x91\x85\x8F\x02\x92\x79\xAF\x00\x93\x85"
-     "\x8F\x02"
-     "\x94\x59\xAF\x00\x95\x15\x80\x02\x96\x49\xAF\x00\x97\xC9\xAF\x00\x98\xFD"
-     "\x8E\x02"
-     "\x99\x7D\xAF\x00\x9A\x41\xA4\x00\x9B\xBD\x8F\x02\x9C\x2D\x8F\x02\x9D\x55"
-     "\xAF\x00"
-     "\x9E\x51\xAF\x00\x9F\x59\xAF\x00\xA0\x5D\x8F\x02\xA1\x85\x82\x02\xA2\x79"
-     "\x8F\x02"
-     "\xA3\xC5\xAF\x00\xA4\xB9\x14\x00\xA5\x15\x14\x00\x63\x01\x00\x80\x67\x01"
-     "\x00\x80"
-     "\xA8\x01\x04\x00\xA9\x2D\x14\x00\xAA\x55\xAF\x00\xAB\x21\x14\x00\xAC\xB1"
-     "\x06\x00"
-     "\xAD\x21\x14\x00\xAE\xB1\x06\x00\xAF\x79\xBB\x02\xB0\xD1\x06\x00\xB1\x59"
-     "\x14\x00"
-     "\xB2\x61\x14\x00\xB3\x61\x14\x00\xB4\x61\x14\x00\xB5\x61\x14\x00\xB6\xC1"
-     "\x06\x00"
-     "\xB7\x9D\xAC\x00\xB8\xC1\x06\x00\xB9\x21\xBB\x02\xBA\x25\xBB\x02\xBB\x55"
-     "\x14\x00"
-     "\xBC\x5D\xBB\x02\xBD\x49\x14\x00\xBE\x39\x14\x00\xBF\x39\x14\x00\x80\xD9"
-     "\xAE\x00"
-     "\x81\xC9\x15\x00\x82\x21\x07\x00\x83\xA9\xBA\x02\x84\xA5\xA9\x00\x85\xD1"
-     "\xA9\x00"
-     "\x86\xD1\xBA\x02\x87\xCD\x15\x00\x88\x11\x07\x00\x89\x8D\xBA\x02\x8A\x79"
-     "\xAE\x00"
-     "\x8B\x95\xBA\x02\x8C\xE5\x15\x00\x8D\xB5\x8F\x02\x8E\xE1\x15\x00\x8F\xCD"
-     "\x15\x00"
-     "\x90\x79\x07\x00\x91\xF1\xBA\x02\x92\xF5\xBA\x02\x93\x8D\xBA\x02\x94\x51"
-     "\x14\x00"
-     "\x95\x5D\x14\x00\x96\x19\x07\x00\x97\xED\xBA\x02\x98\xA9\xA9\x00\x99\x69"
-     "\x14\x00"
-     "\x9A\x21\x07\x00\x9B\x7D\x14\x00\x9C\xB9\xBA\x02\x9D\xDD\xBA\x02\x9E\xC1"
-     "\xBA\x02"
-     "\x9F\xCD\xBA\x02\xA0\x3D\xBA\x02\xA1\x31\xBA\x02\xA2\x29\x05\x00\xA3\x05"
-     "\x15\x00"
-     "\xA4\x6D\xAE\x00\xA5\x19\x15\x00\xA6\x99\x07\x00\xA7\x09\x15\x00\xA8\xA9"
-     "\x07\x00"
-     "\xA9\x61\xBA\x02\xAA\xB9\x07\x00\xAB\x31\x15\x00\xAC\x19\x15\x00\xAD\x19"
-     "\x15\x00"
-     "\xAE\x09\x15\x00\xAF\x09\x15\x00\xB0\xD9\x07\x00\xB1\x85\xAD\x00\xB2\xE9"
-     "\x07\x00"
-     "\xB3\x09\xBA\x02\xB4\x1D\xBA\x02\xB5\x6D\x15\x00\xB6\x75\xBA\x02\xB7\x61"
-     "\x15\x00"
-     "\xB8\x21\x15\x00\xB9\x21\x15\x00\xBA\x31\xAE\x00\xBB\x21\x15\x00\xBC\xD9"
-     "\x07\x00"
-     "\xBD\x51\xBA\x02\xBE\x4D\xA9\x00\xBF\x39\xA9\x00\xEF\x0D\x00\x80\xAF\x84"
-     "\x39\x00"
-     "\xF5\x0D\x00\x80\x6B\x01\x00\x80\xFB\x0D\x00\x80\x6F\x01\x00\x80\x0D\x0E"
-     "\x00\x80"
-     "\x06\x00\x00\x00\x13\x0E\x00\x80\xD8\x0E\x00\x80\x40\x0E\x00\x80\x73\x01"
-     "\x00\x80"
-     "\x52\x0E\x00\x80\x77\x01\x00\x80\x8A\x0E\x00\x80\x7B\x01\x00\x80\x9C\x0E"
-     "\x00\x80"
-     "\x7F\x01\x00\x80\xA8\x0E\x00\x80\x83\x01\x00\x80\xE1\x0E\x00\x80\x87\x01"
-     "\x00\x80"
-     "\xE7\x0E\x00\x80\x8B\x01\x00\x80\xEA\x0E\x00\x80\xED\x0E\x00\x80\xF0\x0E"
-     "\x00\x80"
-     "\x8F\x01\x00\x80\xF6\x0E\x00\x80\x93\x01\x00\x80\x02\x0F\x00\x80\x11\x0F"
-     "\x00\x80"
-     "\x1D\x0F\x00\x80\x26\x0F\x00\x80\x2C\x0F\x00\x80\x80\x0F\x00\x80\x3E\x0F"
-     "\x00\x80"
-     "\xA7\x0F\x00\x80\x88\x00\x34\x00\x89\x74\x01\x00\x86\xE4\x10\x00\x97\x01"
-     "\x00\x80"
-     "\x84\xA8\x38\x00\x85\xAC\x39\x00\x9B\x01\x00\x80\x44\x24\x00\x80\xD0\x06"
-     "\x00\x80"
-     "\xCA\x07\x00\x80\x0E\x07\x00\x80\x96\x07\x00\x80\x9C\x07\x00\x80\x7F\x07"
-     "\x00\x80"
-     "\xC7\x07\x00\x80\x4E\x07\x00\x80\x51\x07\x00\x80\x5C\x07\x00\x80\xE0\x06"
-     "\x00\x80"
-     "\x0E\x25\x00\x80\xE1\x26\x00\x80\x2E\x2F\x00\x80\xB3\x2D\x00\x80\xCF\x2E"
-     "\x00\x80"
-     "\x38\x38\x00\x80\x00\x2C\x00\x80\xA8\x71\xA8\x00\xA9\x71\xA8\x00\xAA\x71"
-     "\xA8\x00"
-     "\xAB\x71\xA8\x00\xAC\x75\xA8\x00\xAD\x9D\x89\x02\xAE\x6D\x89\x02\xAF\x65"
-     "\x89\x02"
-     "\xA0\x59\xA8\x00\xA1\x59\xA8\x00\xA2\x4D\xA8\x00\xA3\xA5\x89\x02\xA4\x45"
-     "\x89\x02"
-     "\xA5\x4D\x89\x02\xA6\x51\xA8\x00\xA7\x51\xA8\x00\xB8\x35\x89\x02\xB9\x3D"
-     "\x89\x02"
-     "\xBA\x21\xA8\x00\xBB\x21\xA8\x00\xBC\x21\xA8\x00\xBD\x21\xA8\x00\xBE\x21"
-     "\xA8\x00"
-     "\xBF\x21\xA8\x00\xB0\x09\xA8\x00\xB1\x09\xA8\x00\xB2\x19\xA8\x00\xB3\x19"
-     "\xA8\x00"
-     "\xB4\x09\xA8\x00\xB5\x09\xA8\x00\xB6\x1D\xA8\x00\xB7\xF5\x89\x02\x88\x3D"
-     "\xBC\x02"
-     "\x89\x31\xBC\x02\x8A\x9D\x12\x00\x8B\x91\x12\x00\xCC\xF4\x88\x02\xCD\xC0"
-     "\x87\x02"
-     "\x8E\xFD\xA8\x00\x8F\x15\x89\x02\x80\x69\xAF\x00\x81\xA9\x12\x00\x82\xE1"
-     "\x01\x00"
-     "\x83\xBD\x12\x00\x84\x79\xBC\x02\x85\x1D\xBC\x02\x86\x01\xBC\x02\x87\x0D"
-     "\xBC\x02"
-     "\x98\xA5\xA8\x00\x99\x4D\x89\x02\x9A\xBD\x89\x02\x9B\xB5\x89\x02\x9C\xB9"
-     "\xA8\x00"
-     "\x9D\xB9\xA8\x00\x9E\xA9\xA8\x00\x9F\xA9\xA8\x00\x90\x95\x89\x02\x91\x9D"
-     "\x89\x02"
-     "\x92\x81\xA8\x00\x93\x81\xA8\x00\x94\x81\xA8\x00\x95\x81\xA8\x00\x96\x81"
-     "\xA8\x00"
-     "\x97\x81\xA8\x00\x87\x85\x02\x00\x86\x7D\x11\x00\x85\x89\x02\x00\x84\x61"
-     "\x11\x00"
-     "\x83\x79\x11\x00\x82\x71\x11\x00\x81\x65\x02\x00\x80\x7D\x11\x00\x8F\xF5"
-     "\x11\x00"
-     "\x8E\xB1\x02\x00\x8D\xF9\x11\x00\x8C\xBD\x02\x00\x8B\xBD\x02\x00\x8A\xF1"
-     "\x11\x00"
-     "\x89\x45\x11\x00\x88\xB5\x02\x00\x97\x21\x02\x00\x96\x1D\x02\x00\x95\x0D"
-     "\x11\x00"
-     "\x94\x1D\x02\x00\x93\x11\x11\x00\x92\x61\x11\x00\x91\x7D\x02\x00\x90\x85"
-     "\x11\x00"
-     "\x9F\x1D\x02\x00\x9E\x39\x11\x00\x9D\x39\x11\x00\x9C\x15\x02\x00\x9B\x39"
-     "\x11\x00"
-     "\x9A\x05\x02\x00\x99\x11\x02\x00\x98\x19\x02\x00\xA7\x5D\x11\x00\xA6\x01"
-     "\x02\x00"
-     "\xA5\x71\xFC\x00\xA4\xF5\x11\x00\xA3\x11\x02\x00\xA2\xF9\x11\x00\xA1\xF1"
-     "\x11\x00"
-     "\xA0\xF9\x11\x00\xE3\x0D\x00\x80\xCC\x31\x00\x80\xAD\x5D\x11\x00\xAC\x5D"
-     "\x11\x00"
-     "\xAB\x59\x11\x00\xAA\x51\x11\x00\xA9\x59\x11\x00\xA8\x2D\x02\x00\xB8\x33"
-     "\x00\x80"
-     "\x07\x2D\x00\x80\x59\x30\x00\x80\x9F\x2C\x00\x80\xA5\x31\x00\x80\xF3\x2C"
-     "\x00\x80"
-     "\x20\x36\x00\x80\x94\x38\x00\x80\xB3\x2E\x00\x80\xEB\x2D\x00\x80\x6C\x33"
-     "\x00\x80"
-     "\x09\x31\x00\x80\x35\x31\x00\x80\x14\x39\x00\x80\x29\x31\x00\x80\x01\x31"
-     "\x00\x80"
-     "\xCC\x08\x80\x02\xCD\xE4\x80\x02\xCC\x80\x82\x02\xCD\x98\x82\x02\xCC\x5C"
-     "\x87\x02"
-     "\xCC\x8C\xBE\x02\xCC\x70\x81\x02\xCD\xBC\x81\x02\xCC\xA0\x84\x02\xCD\xFC"
-     "\x84\x02"
-     "\xCC\x2C\x87\x02\xCD\x14\x87\x02\xCC\x08\x85\x02\xCD\x80\x85\x02\xCC\x38"
-     "\x84\x02"
-     "\xCD\x54\x84\x02\xCC\x5C\xB9\x02\xCD\x58\xB9\x02\xCC\x28\xB9\x02\xCC\xD0"
-     "\xB9\x02"
-     "\xCC\x48\xB9\x02\xCC\x80\xBE\x02\xCC\x5C\xB9\x02\xCD\x7C\xB9\x02\xCC\xF4"
-     "\xBE\x02"
-     "\xCD\xF0\xBE\x02\xCC\xC0\xB8\x02\xCD\xE0\xBE\x02\xCC\x98\x82\x02\xCD\xC4"
-     "\xBD\x02"
-     "\xCC\x70\xBC\x02\xCD\xFC\xBC\x02\x78\x35\x00\x80\xB4\x35\x00\x80\xD7\x2D"
-     "\x00\x80"
-     "\x4C\x33\x00\x80\xA0\x34\x00\x80\x0B\x2D\x00\x80\x37\x2E\x00\x80\x4C\x39"
-     "\x00\x80"
-     "\xCC\x18\xB9\x02\xCC\x74\xB9\x02\x90\x25\x3C\x00\x91\x39\x3C\x00\xCC\xF4"
-     "\xB9\x02"
-     "\xCC\xE4\xB9\x02\xCC\xE4\xB9\x02\xCC\x44\xB9\x02\x88\x24\x00\x80\x4C\x24"
-     "\x00\x80"
-     "\x88\x49\x3C\x00\x56\x24\x00\x80\x6A\x24\x00\x80\x92\x24\x00\x80\x74\x24"
-     "\x00\x80"
-     "\x9C\x24\x00\x80\x82\x2D\x00\x00\x83\x35\x00\x00\x80\x29\x00\x00\x81\x29"
-     "\x00\x00"
-     "\x86\x2D\x00\x00\x87\xE5\x3C\x00\x84\x2D\x00\x00\x85\x25\x00\x00\x99\xB1"
-     "\x81\x02"
-     "\x98\xBD\x81\x02\x9B\xBD\x81\x02\x9A\xB1\x81\x02\x9D\x99\x2E\x00\x9C\xB5"
-     "\x3E\x00"
-     "\x9F\x95\x2E\x00\x9E\xE1\x95\x00\x91\x91\x81\x02\x90\x65\x3C\x00\x93\x25"
-     "\x2F\x00"
-     "\x92\xE5\x92\x00\x95\x21\x2F\x00\x94\x7D\x3C\x00\x97\x91\x81\x02\x96\xF5"
-     "\x81\x02"
-     "\x89\xB1\x2E\x00\x88\x9D\x2E\x00\x8B\xFD\x81\x02\x8A\x75\x3C\x00\x8D\x91"
-     "\x81\x02"
-     "\x8C\xE9\x81\x02\x8F\x51\x2F\x00\x8E\x5D\x2F\x00\x81\xB1\x2E\x00\x80\xAD"
-     "\x81\x02"
-     "\x83\xC1\x81\x02\x82\x5D\x3C\x00\x85\xC9\x81\x02\x84\x25\x95\x00\x87\xF9"
-     "\xB4\x02"
-     "\x86\xA9\x2E\x00\xB9\x45\x92\x00\xB8\x31\x92\x00\xBB\x59\x2E\x00\xBA\x45"
-     "\x81\x02"
-     "\xBD\x39\x81\x02\xBC\xA5\x3C\x00\xBF\x21\x81\x02\xBE\xCD\x95\x00\xB1\x1D"
-     "\x2E\x00"
-     "\xB0\x09\x81\x02\xB3\x6D\x2E\x00\xB2\x6D\x2E\x00\xB5\x7D\x2E\x00\xB4\x6D"
-     "\x95\x00"
-     "\xB7\x1D\x81\x02\xB6\x95\x3C\x00\xA9\x75\x2E\x00\xA8\x75\x2E\x00\xAB\x89"
-     "\x96\x00"
-     "\xAA\xD5\x3C\x00\xAD\x15\x81\x02\xAC\xF5\x3C\x00\xAF\x61\x2E\x00\xAE\x11"
-     "\x81\x02"
-     "\xA1\x75\x2E\x00\xA0\xE5\x3C\x00\xA3\x2D\x81\x02\xA2\xE5\x3C\x00\xA5\x6D"
-     "\x2E\x00"
-     "\xA4\xE5\x3C\x00\xA7\x55\x2E\x00\xA6\x55\x2E\x00\x80\xD5\x2F\x00\x81\x85"
-     "\xB5\x02"
-     "\x82\xD1\x2F\x00\x83\xFD\x2F\x00\x84\x29\x3D\x00\x85\xA1\x80\x02\x86\xA5"
-     "\x80\x02"
-     "\x87\xDD\x80\x02\x88\x21\x2E\x00\x89\x2D\x2E\x00\x8A\x69\x3D\x00\x8B\x9D"
-     "\x80\x02"
-     "\x8C\xF9\x93\x00\x8D\x39\x2E\x00\x8E\x71\x3D\x00\x8F\x2D\x2E\x00\x90\x89"
-     "\x80\x02"
-     "\x91\xED\x80\x02\x92\xF1\x80\x02\x93\xFD\x80\x02\x94\xED\x80\x02\x95\xE1"
-     "\x80\x02"
-     "\x96\xF9\x3F\x00\x97\xD5\x2F\x00\x98\x9D\x94\x00\x99\xE9\x2F\x00\x9A\x69"
-     "\x3D\x00"
-     "\x9B\xF9\x2F\x00\x9C\x79\x3D\x00\x9D\xB1\x80\x02\x9E\x69\x3D\x00\x9F\xE1"
-     "\x2F\x00"
-     "\xA0\x29\x2F\x00\xA1\x29\x2F\x00\xA2\x39\x2F\x00\xA3\x39\x2F\x00\xA4\x89"
-     "\x3D\x00"
-     "\xA5\xD5\x97\x00\xA6\xB9\x3D\x00\xA7\x59\x80\x02\xA8\x6D\x80\x02\xA9\x1D"
-     "\x2F\x00"
-     "\xAA\x05\x80\x02\xAB\x11\x2F\x00\xAC\x71\x2F\x00\xAD\x71\x2F\x00\xAE\x61"
-     "\x94\x00"
-     "\xAF\x71\x2F\x00\xB0\xE9\x3D\x00\xB1\x61\x80\x02\xB2\x7D\x93\x00\xB3\x09"
-     "\x93\x00"
-     "\xB4\x19\x80\x02\xB5\x05\x2F\x00\xB6\xE9\x3D\x00\xB7\x75\x80\x02\xB8\xB1"
-     "\x94\x00"
-     "\xB9\x5D\x80\x02\xBA\x3D\x2F\x00\xBB\x6D\xB5\x02\xBC\x29\x2F\x00\xBD\x05"
-     "\x2F\x00"
-     "\xBE\xC1\x3D\x00\xBF\x49\x80\x02\x80\xBD\x83\x02\x81\xC5\x83\x02\x82\x09"
-     "\x2D\x00"
-     "\x83\x05\x2D\x00\x84\x51\x3E\x00\x85\xA5\x83\x02\x86\xD1\x90\x00\x87\x11"
-     "\x2D\x00"
-     "\x88\x69\x3E\x00\x89\x35\x2D\x00\x8A\xE1\x83\x02\x8B\x85\x83\x02\x8C\x89"
-     "\x83\x02"
-     "\x8D\x85\x83\x02\x8E\x85\x83\x02\x8F\x89\x83\x02\x90\xE1\x3C\x00\x91\xCD"
-     "\x2C\x00"
-     "\x92\xB5\x97\x00\x93\xC1\x2C\x00\x94\x51\x3E\x00\x95\xC1\x2C\x00\x96\x51"
-     "\x3E\x00"
-     "\x97\x99\x83\x02\x98\x71\x3E\x00\x99\xF9\x2C\x00\x9A\xC1\x2C\x00\x9B\xC1"
-     "\x2C\x00"
-     "\x9C\xC1\x2C\x00\x9D\xC1\x2C\x00\x9E\x61\x3E\x00\x9F\x3D\x94\x00\xA0\xA1"
-     "\x3E\x00"
-     "\xA1\x41\x83\x02\xA2\x45\x83\x02\xA3\x35\x2C\x00\xA4\x3D\x83\x02\xA5\x29"
-     "\x2C\x00"
-     "\xA6\x59\x2C\x00\xA7\x59\x2C\x00\xA8\x79\x97\x00\xA9\x69\x2C\x00\xAA\x81"
-     "\x3E\x00"
-     "\xAB\x09\x83\x02\xAC\x05\x90\x00\xAD\x71\x90\x00\xAE\x71\x83\x02\xAF\x6D"
-     "\x2C\x00"
-     "\xB0\xF1\x3E\x00\xB1\x6D\x83\x02\xB2\x99\x97\x00\xB3\x75\x83\x02\xB4\x05"
-     "\x2C\x00"
-     "\xB5\x55\xB6\x02\xB6\x01\x2C\x00\xB7\x2D\x2C\x00\xB8\xD9\x3E\x00\xB9\x51"
-     "\x83\x02"
-     "\xBA\x55\x83\x02\xBB\x2D\x83\x02\xBC\xF1\x2D\x00\xBD\xFD\x2D\x00\xBE\xB9"
-     "\x3E\x00"
-     "\xBF\x4D\x83\x02\x80\xDD\x20\x00\x81\x41\x20\x00\x82\x5D\x20\x00\x83\x6D"
-     "\x20\x00"
-     "\x60\x24\x00\x80\x85\x19\x27\x00\x86\x75\x20\x00\x87\x45\x20\x00\x88\xE1"
-     "\x20\x00"
-     "\x89\x19\x2B\x00\x8A\xF1\x20\x00\x8B\xCD\x20\x00\x8C\xD5\x20\x00\x8D\xDD"
-     "\x20\x00"
-     "\x8E\x49\x20\x00\x8F\xF5\x20\x00\x90\x91\x20\x00\x91\x3D\x20\x00\x92\x9D"
-     "\x20\x00"
-     "\x93\x25\x20\x00\x94\x25\x20\x00\x95\x19\x20\x00\x96\x09\x20\x00\x97\x01"
-     "\x20\x00"
-     "\x98\x05\x20\x00\x99\x19\x20\x00\x9A\xAD\x20\x00\x9B\xA1\x20\x00\x9C\x89"
-     "\x2C\x00"
-     "\x9D\x75\xD6\x00\x9E\x2D\x32\x00\x9F\x09\x31\x00\x7E\x24\x00\x80\xA1\xC1"
-     "\x20\x00"
-     "\xA2\xDD\x20\x00\x5D\x30\x00\x80\xA4\x7D\x20\x00\xC7\x2D\x00\x80\xD4\x2B"
-     "\x00\x80"
-     "\xA7\xC5\x20\x00\xCC\x2B\x00\x80\xA9\x99\x2B\x00\xAA\x71\x20\x00\xAB\x4D"
-     "\x20\x00"
-     "\xAC\x55\x20\x00\xAD\x5D\x20\x00\xAE\xC9\x20\x00\xAF\x75\x20\x00\xB0\x11"
-     "\x20\x00"
-     "\xB1\xBD\x20\x00\xB2\x1D\x20\x00\xDC\x2B\x00\x80\xB4\xA5\x20\x00\xB5\x99"
-     "\x20\x00"
-     "\xB6\x89\x20\x00\xB7\x81\x20\x00\xFA\xEA\x00\x80\xB9\x99\x20\x00\x27\x2C"
-     "\x00\x80"
-     "\xBB\x21\x20\x00\x3B\x2C\x00\x80\x43\x2C\x00\x80\x4B\x2C\x00\x80\x4F\x2C"
-     "\x00\x80"
-     "\x80\xDD\x27\x00\x81\x41\x27\x00\x82\x5D\x27\x00\x83\x6D\x27\x00\x84\xFD"
-     "\x27\x00"
-     "\x85\x19\x20\x00\x86\x75\x27\x00\x87\x45\x27\x00\x88\xE1\x27\x00\x89\x19"
-     "\x2C\x00"
-     "\x53\x2C\x00\x80\x8B\xCD\x27\x00\x8C\xD5\x27\x00\x8D\xDD\x27\x00\x8E\x49"
-     "\x27\x00"
-     "\x8F\xF5\x27\x00\x90\x91\x27\x00\x91\x3D\x27\x00\x92\x9D\x27\x00\x93\x25"
-     "\x27\x00"
-     "\x94\x25\x27\x00\x95\x19\x27\x00\x96\x09\x27\x00\x97\x01\x27\x00\x98\x05"
-     "\x27\x00"
-     "\x99\x19\x27\x00\x9A\xAD\x27\x00\x9B\xA1\x27\x00\x8C\x2A\x00\x80\x0E\xEB"
-     "\x00\x80"
-     "\x73\x2C\x00\x80\x77\x2C\x00\x80\x93\x2C\x00\x80\xA1\xC1\x27\x00\xA2\xDD"
-     "\x27\x00"
-     "\xA3\xED\x27\x00\xFF\xEA\x00\x80\xA5\x99\x20\x00\xA6\xF5\x27\x00\xA7\xC5"
-     "\x27\x00"
-     "\xA8\x61\x27\x00\xA9\x99\x2C\x00\x90\x2A\x00\x80\xAB\x4D\x27\x00\xAC\x55"
-     "\x27\x00"
-     "\xAD\x5D\x27\x00\xAE\xC9\x27\x00\xAF\x75\x27\x00\xB0\x11\x27\x00\xB1\xBD"
-     "\x27\x00"
-     "\xB2\x1D\x27\x00\xB3\xA5\x27\x00\xB4\xA5\x27\x00\xB5\x99\x27\x00\xB6\x89"
-     "\x27\x00"
-     "\xB7\x81\x27\x00\xB8\x85\x27\x00\xB9\x99\x27\x00\xBA\x2D\x27\x00\xBB\x21"
-     "\x27\x00"
-     "\x7F\x2C\x00\x80\x09\xEB\x00\x80\xAF\x2C\x00\x80\xB7\x2C\x00\x80\x41\x00"
-     "\x00\x00"
-     "\xB2\xD1\x26\x00\xB1\x71\x26\x00\xB0\xDD\x26\x00\xB7\x4D\x26\x00\xB6\x45"
-     "\x26\x00"
-     "\xB5\x55\x26\x00\xB4\x69\x26\x00\xBB\xED\x26\x00\xBA\xE1\x26\x00\xB9\x55"
-     "\x26\x00"
-     "\x10\x2C\x00\x80\xBF\x2C\x00\x80\xBE\x61\x34\x00\xE4\xEC\x00\x80\xBC\xC5"
-     "\x2A\x00"
-     "\xE7\x2C\x00\x80\xA2\x11\x26\x00\xA1\x0D\x26\x00\x98\x2A\x00\x80\xA7\x09"
-     "\x26\x00"
-     "\xFF\x2C\x00\x80\x03\x2D\x00\x80\xA4\xB1\x26\x00\x0F\x2D\x00\x80\xAA\xBD"
-     "\x26\x00"
-     "\xA9\x55\x2D\x00\xA8\xAD\x26\x00\xAF\xB9\x26\x00\xAE\x05\x26\x00\xAD\x91"
-     "\x26\x00"
-     "\xAC\x99\x26\x00\x61\x00\x00\x00\x92\x51\x26\x00\x91\xF1\x26\x00\x9C\x2A"
-     "\x00\x80"
-     "\x97\xCD\x26\x00\x47\x2D\x00\x80\x4B\x2D\x00\x80\x94\xE9\x26\x00\x9B\x6D"
-     "\x26\x00"
-     "\x73\x2D\x00\x80\x99\xD5\x26\x00\x7F\x2D\x00\x80\x9F\xC5\x37\x00\x8F\x2D"
-     "\x00\x80"
-     "\x9D\xB9\xD0\x00\x97\x2D\x00\x80\x9B\x2D\x00\x80\x82\x91\x26\x00\x1D\xEB"
-     "\x00\x80"
-     "\x9C\x33\x00\x80\x87\x89\x26\x00\xAF\x2D\x00\x80\xB7\x2D\x00\x80\x22\xEB"
-     "\x00\x80"
-     "\x8B\x01\x26\x00\xC3\x2D\x00\x80\x89\xD5\x2D\x00\xCB\x2D\x00\x80\x8F\x39"
-     "\x26\x00"
-     "\x8E\x85\x26\x00\x8D\x11\x26\x00\xD3\x2D\x00\x80\x80\x81\x34\x00\x81\x81"
-     "\x34\x00"
-     "\x82\x85\x34\x00\x83\x9D\x34\x00\x84\x8D\x34\x00\x85\xB5\x34\x00\x86\xAD"
-     "\x34\x00"
-     "\x87\xF5\x34\x00\x88\xD1\x34\x00\x89\x35\x35\x00\x8A\x39\x32\x00\xFB\x2D"
-     "\x00\x80"
-     "\xE7\x2D\x00\x80\xF3\x2D\x00\x80\xF7\x2D\x00\x80\xFF\x2D\x00\x80\x90\x71"
-     "\x32\x00"
-     "\x91\x19\x32\x00\x92\x21\x32\x00\x93\x39\x32\x00\x94\xFD\x32\x00\x95\xED"
-     "\x32\x00"
-     "\x96\x4D\x33\x00\x97\x41\x33\x00\x98\x41\x33\x00\x99\x41\x33\x00\x9A\x2D"
-     "\x33\x00"
-     "\x9B\xF5\x3C\x00\x9C\xED\x3C\x00\x9D\x79\x3A\x00\x9E\x6D\x3A\x00\x9F\x51"
-     "\x3A\x00"
-     "\xA0\xB5\x3A\x00\xA1\xB9\x3A\x00\xA2\xCD\x3A\x00\xA3\xC1\x3A\x00\xA4\xC5"
-     "\x3A\x00"
-     "\xA5\xC9\x3A\x00\xA6\xF5\x3A\x00\xA7\xE9\x3A\x00\xA8\x05\x3A\x00\xA9\x15"
-     "\x3A\x00"
-     "\xAA\x1D\x3A\x00\xAB\x09\x93\x00\xAC\x89\x92\x00\xAD\x0D\x3A\x00\xAE\x05"
-     "\x3A\x00"
-     "\x07\x2E\x00\x80\xB0\x4D\x97\x00\xB1\x5D\x97\x00\xB2\x6D\x93\x00\xB3\x51"
-     "\x97\x00"
-     "\xB4\x71\x97\x00\xB5\x3D\x93\x00\xB6\x71\x97\x00\xB7\x01\x97\x00\xB8\x05"
-     "\x93\x00"
-     "\xB9\xE5\x92\x00\xBA\xE1\x92\x00\xBB\xFD\x92\x00\xBC\xE9\x92\x00\xBD\xD5"
-     "\x92\x00"
-     "\xBE\xD1\x92\x00\xBF\xCD\x92\x00\x80\xC5\x92\x00\x81\x3D\x93\x00\x82\x01"
-     "\xE7\x00"
-     "\x83\x09\x93\x00\x84\x79\x93\x00\x85\x95\x9F\x00\x86\x69\x93\x00\x87\x99"
-     "\x9F\x00"
-     "\x88\x09\xE7\x00\x89\xF1\x92\x00\x8A\xE9\x3B\x00\x8B\x75\xF8\x00\x8C\xF5"
-     "\x3B\x00"
-     "\x8D\xFD\x3B\x00\x8E\xF5\x3B\x00\x8F\xED\x3B\x00\xB0\x25\x92\x00\xB1\xCD"
-     "\xB3\x02"
-     "\xB2\x3D\xB3\x02\xB3\x35\xB3\x02\xB4\x39\x92\x00\xB5\x39\x92\x00\xB6\x29"
-     "\x92\x00"
-     "\xB7\x29\x92\x00\xB8\x19\x92\x00\xB9\x19\x92\x00\x0B\x2E\x00\x80\x13\x2E"
-     "\x00\x80"
-     "\x1B\x2E\x00\x80\x1F\x2E\x00\x80\x2F\x2E\x00\x80\x33\x2E\x00\x80\xA1\x80"
-     "\x0C\x00"
-     "\xA0\x84\x0D\x00\xA3\x80\x0B\x00\xA2\x00\x08\x00\xA5\x90\x14\x00\xA4\x94"
-     "\x0A\x00"
-     "\xA7\x98\x11\x00\xA6\x00\x10\x00\x3F\x2E\x00\x80\xA8\x24\x13\x00\xAA\x7D"
-     "\x3B\x00"
-     "\xAB\x75\x3B\x00\xAC\x6D\x3B\x00\x63\x2E\x00\x80\x67\x2E\x00\x80\x4B\x2E"
-     "\x00\x80"
-     "\x53\x2E\x00\x80\xAF\x2E\x00\x80\x77\x2E\x00\x80\xF2\x00\x00\x00\x7B\x2E"
-     "\x00\x80"
-     "\x7F\x2E\x00\x80\x73\x2E\x00\x80\x87\x2E\x00\x80\x83\x2E\x00\x80\x9B\x2E"
-     "\x00\x80"
-     "\x27\xEB\x00\x80\xB7\x2E\x00\x80\xBB\x2E\x00\x80\xC3\x2E\x00\x80\xD7\x2E"
-     "\x00\x80"
-     "\xFA\x2E\x00\x80\x80\x99\x0B\x00\x81\xA5\x0B\x00\x82\xC5\x0B\x00\x83\xED"
-     "\x0B\x00"
-     "\x84\xF9\x0B\x00\x85\x51\xC4\x00\x86\xE9\x0B\x00\x87\xE9\x0B\x00\x88\xD9"
-     "\x0B\x00"
-     "\x89\xD9\x0B\x00\x8A\x65\xC4\x00\x8B\x05\x0C\x00\x8C\x41\x0C\x00\x8D\x41"
-     "\x0C\x00"
-     "\x8E\x69\xC4\x00\x8F\x45\x0C\x00\x90\x3D\x0C\x00\x91\xC5\x0C\x00\x92\xD9"
-     "\x0C\x00"
-     "\x93\xE5\x0C\x00\x94\xA1\x0C\x00\x95\x99\x36\x00\x96\xAD\x0C\x00\x97\xA5"
-     "\x0C\x00"
-     "\x98\x9D\x0C\x00\x99\x49\x0D\x00\x9A\x59\x0D\x00\x9B\xAD\xC5\x00\x9C\x49"
-     "\x0D\x00"
-     "\x9D\xF5\x97\x00\x9E\x39\x0D\x00\x9F\x31\x0D\x00\xA0\xD5\x0D\x00\xA1\xDD"
-     "\x0D\x00"
-     "\xA2\x1D\x37\x00\xA3\xED\x0D\x00\xA4\x1D\x0D\x00\xA5\x41\xC5\x00\xA6\xE5"
-     "\xC4\x00"
-     "\xA7\x19\xC4\x00\xA8\x79\xC5\x00\xA9\x31\x0D\x00\xAA\x31\x0D\x00\xAB\x99"
-     "\xC0\x00"
-     "\xAC\x15\x0D\x00\xAD\x39\xC4\x00\xAE\x15\x0D\x00\xAF\x09\x0D\x00\xB0\x7D"
-     "\x0D\x00"
-     "\xB1\x05\x0D\x00\xB2\x09\x0D\x00\xB3\x0D\x0D\x00\xB4\x19\x0D\x00\xB5\x19"
-     "\x0D\x00"
-     "\xB6\x0D\x0D\x00\xB7\x05\x0D\x00\xB8\x3D\x0D\x00\xB9\x01\x0D\x00\xBA\x05"
-     "\x0D\x00"
-     "\xBB\x11\x0D\x00\xBC\xDD\x0E\x00\xBD\xC5\x0E\x00\xBE\xCD\x0E\x00\xBF\xC5"
-     "\x0E\x00"
-     "\x80\x3D\x0F\x00\x81\xB9\x0F\x00\x82\x61\x0E\x00\x83\x61\x0E\x00\x84\x61"
-     "\x0E\x00"
-     "\x85\x61\x0E\x00\x86\x6D\x0E\x00\x87\xFD\xC4\x00\x88\x59\x0E\x00\x89\x59"
-     "\x0E\x00"
-     "\x8A\x89\x0E\x00\x8B\x99\x0E\x00\x8C\xAD\xC5\x00\x8D\x8D\x0E\x00\x8E\x85"
-     "\x0E\x00"
-     "\x8F\xFD\x0E\x00\x90\xC5\xC2\x00\x91\x81\x0E\x00\x92\x81\x0E\x00\x93\x9D"
-     "\x0E\x00"
-     "\x94\x85\x0E\x00\x95\x8D\x0E\x00\x96\x85\x0E\x00\x97\xBD\x0E\x00\x98\x85"
-     "\x0E\x00"
-     "\x99\x8D\x0E\x00\x9A\x85\x0E\x00\x9B\x9D\x0E\x00\x9C\x85\x0E\x00\x9D\x81"
-     "\x35\x00"
-     "\x9E\x81\x0E\x00\x9F\x81\x0E\x00\xA0\xE1\x03\x00\xA1\xE1\x03\x00\xA2\xE1"
-     "\x03\x00"
-     "\xA3\xE1\x03\x00\xA4\x11\x03\x00\xA5\x11\x03\x00\xA6\x11\x03\x00\xA7\x11"
-     "\x03\x00"
-     "\xA8\x35\x03\x00\xA9\x05\x03\x00\xAA\x0D\x03\x00\xAB\x09\x03\x00\xAC\x19"
-     "\x03\x00"
-     "\xAD\x19\x03\x00\xAE\x09\x03\x00\xAF\x99\xC1\x00\xB0\x75\x03\x00\xB1\x7D"
-     "\x03\x00"
-     "\xB2\x75\x03\x00\xB3\xC9\x02\x00\xB4\xD9\x02\x00\xB5\x9D\xE9\x00\xB6\xD5"
-     "\x02\x00"
-     "\xB7\xCD\x02\x00\xB8\xB5\xE9\x00\xB9\xF1\x02\x00\xBA\xF1\x02\x00\xBB\xF1"
-     "\x02\x00"
-     "\xBC\xD1\x02\x00\xBD\xD1\x02\x00\xBE\xD1\x02\x00\xBF\xD1\x02\x00\xA6\xE9"
-     "\x34\x00"
-     "\xA7\xE9\x34\x00\xA4\xE9\x34\x00\xA5\xE9\x34\x00\xA2\xC1\xEB\x00\xA3\xC9"
-     "\x34\x00"
-     "\xA0\xCD\x34\x00\xA1\xC5\x34\x00\xAE\xED\xFC\x00\xAF\xDD\x34\x00\xAC\xE1"
-     "\xFC\x00"
-     "\xAD\x49\xF9\x00\xAA\x01\xF9\x00\xAB\xC1\x34\x00\xA8\xC9\x34\x00\xA9\x21"
-     "\x95\x00"
-     "\xB6\xC5\xFC\x00\xB7\xAD\xEA\x00\xB4\xC1\x35\x00\xB5\xC1\x35\x00\xB2\xD1"
-     "\x35\x00"
-     "\xB3\xD1\x35\x00\xB0\xD9\xED\x00\xB1\xA1\x34\x00\x2C\xEB\x00\x80\x31\xEB"
-     "\x00\x80"
-     "\x06\x2F\x00\x80\x0A\x2F\x00\x80\xBA\xE9\x35\x00\xBB\xE9\x35\x00\xB8\xE9"
-     "\x35\x00"
-     "\xB9\xE9\x35\x00\x0E\x2F\x00\x80\x02\x2F\x00\x80\x16\x2F\x00\x80\xA0\x2A"
-     "\x00\x80"
-     "\x82\x85\xF6\x00\xA4\x2A\x00\x80\x80\x7D\x36\x00\x81\x75\x36\x00\x1E\x2F"
-     "\x00\x80"
-     "\x26\x2F\x00\x80\x36\xEB\x00\x80\x42\x2F\x00\x80\x46\x2F\x00\x80\x52\x2F"
-     "\x00\x80"
-     "\x3B\xEB\x00\x80\x5A\x2F\x00\x80\x96\x39\x36\x00\x97\x75\xF9\x00\x94\x85"
-     "\x95\x00"
-     "\x95\x39\x36\x00\x92\x21\x36\x00\x93\x09\xE7\x00\x90\x69\xEA\x00\x91\x31"
-     "\x36\x00"
-     "\x9E\x2D\x34\x00\x9F\x35\x34\x00\x9C\x2D\x34\x00\x9D\x25\x34\x00\x9A\x31"
-     "\x34\x00"
-     "\x9B\x79\xC0\x00\x98\x05\x36\x00\x99\x59\x34\x00\xC0\x00\x00\x00\x5E\x2F"
-     "\x00\x80"
-     "\x3D\x30\x00\x80\xA8\x2A\x00\x80\x72\x2F\x00\x80\x8E\x2F\x00\x80\x8A\x2F"
-     "\x00\x80"
-     "\x45\xEB\x00\x80\x96\x2F\x00\x80\x4A\xEB\x00\x80\xA2\x2F\x00\x80\x9E\x2F"
-     "\x00\x80"
-     "\xA6\x2F\x00\x80\xB6\x2F\x00\x80\xBA\x2F\x00\x80\xAC\x2A\x00\x80\xCA\x2F"
-     "\x00\x80"
-     "\xD2\x2F\x00\x80\xD6\x2F\x00\x80\xDA\x2F\x00\x80\xB0\x2A\x00\x80\x4F\xEB"
-     "\x00\x80"
-     "\xB4\x2A\x00\x80\xFD\x2F\x00\x80\x01\x30\x00\x80\x05\x30\x00\x80\x16\xED"
-     "\x00\x80"
-     "\x59\xEB\x00\x80\xE0\x36\x00\x80\x35\x30\x00\x80\xB8\x2A\x00\x80\x7C\xEB"
-     "\x00\x80"
-     "\x4E\xEC\x00\x80\x45\x30\x00\x80\x4D\x30\x00\x80\xBC\x2A\x00\x80\x61\x30"
-     "\x00\x80"
-     "\x71\x30\x00\x80\x75\x30\x00\x80\x7D\x30\x00\x80\x8D\x30\x00\x80\xC4\x2A"
-     "\x00\x80"
-     "\xC0\x2A\x00\x80\x5E\xEB\x00\x80\x95\x30\x00\x80\xA9\x30\x00\x80\xAD\x30"
-     "\x00\x80"
-     "\xBD\x30\x00\x80\xD1\x30\x00\x80\xC9\x30\x00\x80\xCD\x30\x00\x80\xD5\x30"
-     "\x00\x80"
-     "\xE9\x30\x00\x80\xED\x30\x00\x80\x05\x31\x00\x80\x0D\x31\x00\x80\x1D\x31"
-     "\x00\x80"
-     "\x31\x31\x00\x80\x6D\xEB\x00\x80\x2D\x31\x00\x80\x21\x31\x00\x80\x39\x31"
-     "\x00\x80"
-     "\x41\x31\x00\x80\x55\x31\x00\x80\x72\xEB\x00\x80\x5D\x31\x00\x80\x43\x00"
-     "\x00\x00"
-     "\x51\x31\x00\x80\x49\x31\x00\x80\xC8\x2A\x00\x80\x65\x31\x00\x80\x6D\x31"
-     "\x00\x80"
-     "\x75\x31\x00\x80\x69\x31\x00\x80\xCC\x2A\x00\x80\x8D\x31\x00\x80\x77\xEB"
-     "\x00\x80"
-     "\xB9\x31\x00\x80\xFB\x31\x00\x80\xD7\x31\x00\x80\xD8\x2A\x00\x80\xD4\x2A"
-     "\x00\x80"
-     "\xD0\x2A\x00\x80\xA3\x2C\x00\x80\xA7\x2C\x00\x80\xFF\x31\x00\x80\xE7\x31"
-     "\x00\x80"
-     "\xB4\x36\x00\x80\x3C\x2B\x00\x80\x13\x32\x00\x80\x27\x32\x00\x80\x1F\x32"
-     "\x00\x80"
-     "\x90\xEB\x00\x80\xDC\x2A\x00\x80\x3B\x32\x00\x80\x2F\x32\x00\x80\x58\x32"
-     "\x00\x80"
-     "\x9A\xEB\x00\x80\x64\x32\x00\x80\x43\x32\x00\x80\x6C\x32\x00\x80\x70\x32"
-     "\x00\x80"
-     "\x74\x32\x00\x80\x78\x32\x00\x80\x84\x32\x00\x80\x9F\xEB\x00\x80\x8C\x32"
-     "\x00\x80"
-     "\xE4\x2A\x00\x80\x98\x32\x00\x80\xA4\xEB\x00\x80\x9C\x32\x00\x80\xE8\x2A"
-     "\x00\x80"
-     "\xA8\x32\x00\x80\xB4\x32\x00\x80\xC4\x32\x00\x80\xA9\xEB\x00\x80\x40\xEB"
-     "\x00\x80"
-     "\xAE\xEB\x00\x80\xE4\x32\x00\x80\xB3\xEB\x00\x80\xF0\x32\x00\x80\xF8\x32"
-     "\x00\x80"
-     "\xE8\x32\x00\x80\x04\x33\x00\x80\x14\x33\x00\x80\x1C\x33\x00\x80\x28\x33"
-     "\x00\x80"
-     "\x2C\x33\x00\x80\xA3\x09\xC0\x00\xA2\xAD\x06\x00\xA1\xAD\x06\x00\xA0\xAD"
-     "\x06\x00"
-     "\xA7\xA9\x06\x00\xA6\xA1\x06\x00\xA5\xA9\x06\x00\xA4\xB1\x06\x00\xAB\x31"
-     "\xC0\x00"
-     "\xAA\x99\x06\x00\xA9\x91\x06\x00\xA8\x25\xDD\x00\xAF\xFD\x06\x00\xAE\xFD"
-     "\x06\x00"
-     "\xAD\xFD\x06\x00\xAC\xFD\x06\x00\xB3\x95\x06\x00\xB2\x95\x06\x00\xB1\x21"
-     "\xC9\x00"
-     "\xB0\x55\xC0\x00\xB7\xB5\x06\x00\xB6\xB5\x06\x00\xB5\x85\x06\x00\xB4\x85"
-     "\x06\x00"
-     "\xBB\x95\x06\x00\xBA\x95\x06\x00\xB9\x85\x06\x00\xB8\x85\x06\x00\xBF\x75"
-     "\x01\x00"
-     "\xBE\x75\x01\x00\xBD\x85\x06\x00\xBC\x85\x06\x00\x83\xBD\x06\x00\x82\xBD"
-     "\x06\x00"
-     "\x81\xBD\x06\x00\x80\xBD\x06\x00\x87\xDD\x06\x00\x86\xDD\x06\x00\x85\xDD"
-     "\x06\x00"
-     "\x84\xDD\x06\x00\x8B\xFD\x06\x00\x8A\xFD\x06\x00\x89\xFD\x06\x00\x88\xFD"
-     "\x06\x00"
-     "\x8F\xD9\x06\x00\x8E\xE1\xC8\x00\x8D\xDD\x06\x00\x8C\xDD\x06\x00\x93\xB1"
-     "\x06\x00"
-     "\x92\xA9\x06\x00\x91\xA5\x06\x00\x90\xB1\xED\x00\x97\xAD\x06\x00\x96\xAD"
-     "\x06\x00"
-     "\x95\xAD\x06\x00\x94\xA9\x06\x00\x9B\x6D\x06\x00\x9A\x6D\x06\x00\x99\x6D"
-     "\x06\x00"
-     "\x98\x6D\x06\x00\x9F\x6D\x06\x00\x9E\x6D\x06\x00\x9D\x6D\x06\x00\x9C\x6D"
-     "\x06\x00"
-     "\x82\x15\x02\x00\x83\x1D\x02\x00\x80\x05\x02\x00\x81\x1D\x02\x00\x86\x35"
-     "\x02\x00"
-     "\x87\x3D\x02\x00\x84\x05\x02\x00\x85\x3D\x02\x00\x8A\x11\x02\x00\x8B\x11"
-     "\x02\x00"
-     "\x88\x35\xC3\x00\x89\x01\x02\x00\x8E\x71\x02\x00\x8F\x11\xC3\x00\x8C\x01"
-     "\x02\x00"
-     "\x8D\x01\x02\x00\x92\x0D\x02\x00\x93\x15\x02\x00\x90\x0D\x02\x00\x91\x05"
-     "\x02\x00"
-     "\x96\x0D\x02\x00\x97\x35\x02\x00\x94\x0D\x02\x00\x95\x05\x02\x00\x9A\x05"
-     "\x02\x00"
-     "\x9B\x0D\x02\x00\x98\x31\xCB\x00\x99\x45\xDE\x00\x9E\x05\x02\x00\x9F\x0D"
-     "\x02\x00"
-     "\x9C\x15\x02\x00\x9D\x0D\x02\x00\xA2\xA9\xC3\x00\xA3\x01\x02\x00\xA0\xF5"
-     "\x02\x00"
-     "\xA1\x0D\x02\x00\xA6\x01\x02\x00\xA7\x41\xDE\x00\xA4\x11\x02\x00\xA5\x11"
-     "\x02\x00"
-     "\xAA\x3D\x02\x00\xAB\x05\x02\x00\xA8\x3D\x02\x00\xA9\x35\x02\x00\xAE\x1D"
-     "\x02\x00"
-     "\xAF\x05\x02\x00\xAC\x1D\x02\x00\xAD\x15\x02\x00\xB2\x7D\x02\x00\xB3\x05"
-     "\x02\x00"
-     "\xB0\x7D\x02\x00\xB1\x75\x02\x00\xB6\x19\x02\x00\xB7\x19\x02\x00\xB4\x1D"
-     "\x02\x00"
-     "\xB5\xB5\xC6\x00\xBA\x39\x02\x00\xBB\x39\x02\x00\xB8\x39\x02\x00\xB9\x39"
-     "\x02\x00"
-     "\xBE\x19\x02\x00\xBF\x19\x02\x00\xBC\x19\x02\x00\xBD\x19\x02\x00\x80\xF1"
-     "\x03\x00"
-     "\x81\x29\xC2\x00\x82\xBD\xC5\x00\x83\xF9\x03\x00\x84\xE9\x03\x00\x85\xE9"
-     "\x03\x00"
-     "\x86\x01\x1C\x00\x87\x1D\x1F\x00\x88\x25\x1F\x00\x89\x2D\x1F\x00\x8A\x25"
-     "\x1F\x00"
-     "\x8B\x19\xDF\x00\x8C\x21\x1F\x00\x8D\x21\x1F\x00\x8E\x21\x1F\x00\x8F\x21"
-     "\x1F\x00"
-     "\x90\x61\x1F\x00\x91\x61\x1F\x00\x92\x61\x1F\x00\x93\x61\x1F\x00\x94\x69"
-     "\xDF\x00"
-     "\x95\x65\x1F\x00\x96\x6D\x1F\x00\x97\x65\x1F\x00\x98\x5D\x1F\x00\x99\x25"
-     "\x1F\x00"
-     "\x9A\x2D\x1F\x00\x9B\x25\x1F\x00\x9C\x3D\x1F\x00\x9D\x25\x1F\x00\x9E\x2D"
-     "\x1F\x00"
-     "\x9F\x25\x1F\x00\xA0\xDD\x1F\x00\xA1\xC5\xDF\x00\xA2\xE9\x1F\x00\xA3\xE9"
-     "\x1F\x00"
-     "\xA4\xF9\x1F\x00\xA5\xF9\x1F\x00\xA6\xE9\x1F\x00\xA7\xE9\x1F\x00\xA8\xD9"
-     "\x1F\x00"
-     "\xA9\xD9\x1F\x00\xAA\x29\x1F\x00\xAB\x29\x1F\x00\xAC\x39\x1F\x00\xAD\x3D"
-     "\x1F\x00"
-     "\xAE\x35\x1F\x00\xAF\x2D\x1F\x00\xB0\x89\xDF\x00\xB1\x51\x1F\x00\xB2\x51"
-     "\x1F\x00"
-     "\xB3\x51\x1F\x00\xB4\x71\x1F\x00\xB5\x71\x1F\x00\xB6\x71\x1F\x00\xB7\x71"
-     "\x1F\x00"
-     "\xB8\xD1\xC7\x00\xB9\x55\x1F\x00\xBA\x5D\x1F\x00\xBB\x55\x1F\x00\xBC\x4D"
-     "\x1F\x00"
-     "\xBD\x35\x1F\x00\xBE\x3D\x1F\x00\xBF\x35\x1F\x00\x10\x33\x00\x80\x44\x33"
-     "\x00\x80"
-     "\x48\x33\x00\x80\xEC\x2A\x00\x80\x58\x33\x00\x80\xBD\xEB\x00\x80\x3C\x33"
-     "\x00\x80"
-     "\x78\x33\x00\x80\xC2\xEB\x00\x80\xC7\xEB\x00\x80\x7C\x33\x00\x80\x90\x33"
-     "\x00\x80"
-     "\x88\x33\x00\x80\xF0\x2A\x00\x80\x94\x33\x00\x80\xA4\x33\x00\x80\xA0\x33"
-     "\x00\x80"
-     "\xAC\x33\x00\x80\x04\xEB\x00\x80\xBC\x33\x00\x80\xCC\xEB\x00\x80\xC8\x33"
-     "\x00\x80"
-     "\xD6\xEB\x00\x80\xDC\x33\x00\x80\xFC\x33\x00\x80\xDB\xEB\x00\x80\x10\x34"
-     "\x00\x80"
-     "\x14\x34\x00\x80\xE0\xEB\x00\x80\xE5\xEB\x00\x80\x30\x34\x00\x80\x40\x34"
-     "\x00\x80"
-     "\xF4\x2A\x00\x80\x44\x34\x00\x80\xF8\x2A\x00\x80\x5C\x34\x00\x80\x60\x34"
-     "\x00\x80"
-     "\x6C\x34\x00\x80\x78\x34\x00\x80\xFC\x2A\x00\x80\x90\x34\x00\x80\xEA\xEB"
-     "\x00\x80"
-     "\xAC\x34\x00\x80\xEF\xEB\x00\x80\x54\xEB\x00\x80\xC8\x34\x00\x80\xF4\xEB"
-     "\x00\x80"
-     "\xF9\xEB\x00\x80\xFE\xEB\x00\x80\x00\x2B\x00\x80\x04\x2B\x00\x80\x08\xEC"
-     "\x00\x80"
-     "\x03\xEC\x00\x80\x0D\xEC\x00\x80\x12\xEC\x00\x80\x08\x35\x00\x80\x0C\x35"
-     "\x00\x80"
-     "\x18\x35\x00\x80\x10\x2B\x00\x80\x14\x2B\x00\x80\x1C\xEC\x00\x80\x2C\x35"
-     "\x00\x80"
-     "\x18\x2B\x00\x80\x21\xEC\x00\x80\x80\x61\xC1\x00\x81\xC9\x01\x00\x82\xD9"
-     "\x01\x00"
-     "\x83\xD9\x01\x00\x84\xC9\x01\x00\x85\xC9\x01\x00\x86\xF9\x01\x00\x87\xFD"
-     "\x01\x00"
-     "\x88\x45\xC1\x00\x89\xF9\xC1\x00\x8A\xCD\x01\x00\x8B\xC5\x01\x00\x8C\xDD"
-     "\x01\x00"
-     "\x8D\xC5\x01\x00\x8E\xCD\x01\x00\x8F\xA1\xD4\x00\x90\x2D\xC1\x00\x91\xBD"
-     "\x01\x00"
-     "\x92\xB5\x01\x00\x93\x19\xDC\x00\x94\x51\x03\x00\x95\x51\x03\x00\x96\x8D"
-     "\xDA\x00"
-     "\x97\x55\x03\x00\x98\x6D\x03\x00\x99\x05\xDC\x00\x9A\x79\x03\x00\x9B\x79"
-     "\x03\x00"
-     "\x9C\x69\x03\x00\x9D\x69\x03\x00\x9E\x65\x03\x00\x9F\x5D\x03\x00\xA0\xA5"
-     "\x03\x00"
-     "\xA1\xAD\x03\x00\xA2\xA5\x03\x00\xA3\xBD\x03\x00\xA4\xA5\x03\x00\xA5\xAD"
-     "\x03\x00"
-     "\xA6\xA5\x03\x00\xA7\xDD\x03\x00\xA8\xE5\x03\x00\xA9\xED\x03\x00\xAA\xE5"
-     "\x03\x00"
-     "\xAB\xFD\x03\x00\xAC\xE5\x03\x00\xAD\xED\x03\x00\xAE\xE5\x03\x00\xAF\xDD"
-     "\x03\x00"
-     "\xB0\xA5\x03\x00\xB1\xAD\x03\x00\xB2\xA5\x03\x00\xB3\xBD\x03\x00\xB4\xA5"
-     "\x03\x00"
-     "\xB5\xAD\x03\x00\xB6\xA5\x03\x00\xB7\x5D\x03\x00\xB8\x65\x03\x00\xB9\x6D"
-     "\x03\x00"
-     "\xBA\x91\xDC\x00\xBB\x61\x03\x00\xBC\x61\x03\x00\xBD\x61\x03\x00\xBE\x61"
-     "\x03\x00"
-     "\xBF\x61\x03\x00\x26\xEC\x00\x80\x2B\xEC\x00\x80\x9C\x35\x00\x80\x1C\x2B"
-     "\x00\x80"
-     "\xA8\x35\x00\x80\xAC\x35\x00\x80\x30\xEC\x00\x80\x35\xEC\x00\x80\xC4\x35"
-     "\x00\x80"
-     "\x20\x2B\x00\x80\x3A\xEC\x00\x80\xD8\x35\x00\x80\xDC\x35\x00\x80\x24\x2B"
-     "\x00\x80"
-     "\x3F\xEC\x00\x80\xF8\x35\x00\x80\x28\x2B\x00\x80\x08\x36\x00\x80\x00\x36"
-     "\x00\x80"
-     "\x10\x36\x00\x80\x49\xEC\x00\x80\x28\x36\x00\x80\x2C\x2B\x00\x80\x34\x36"
-     "\x00\x80"
-     "\x3C\x36\x00\x80\x4C\x36\x00\x80\x30\x2B\x00\x80\x53\xEC\x00\x80\x58\xEC"
-     "\x00\x80"
-     "\x34\x2B\x00\x80\x5D\xEC\x00\x80\x64\x36\x00\x80\x62\xEC\x00\x80\x70\x36"
-     "\x00\x80"
-     "\x7C\x36\x00\x80\x67\xEC\x00\x80\x6C\xEC\x00\x80\x98\x36\x00\x80\x71\xEC"
-     "\x00\x80"
-     "\xA0\x36\x00\x80\x81\xEB\x00\x80\x38\x2B\x00\x80\xB8\x36\x00\x80\xBC\x36"
-     "\x00\x80"
-     "\x40\x2B\x00\x80\xC0\x36\x00\x80\x1A\x2F\x00\x80\x76\xEC\x00\x80\x7B\xEC"
-     "\x00\x80"
-     "\x86\xEB\x00\x80\x8B\xEB\x00\x80\xE4\x36\x00\x80\xFC\x38\x00\x80\x44\x2B"
-     "\x00\x80"
-     "\x10\x37\x00\x80\x0C\x37\x00\x80\x14\x37\x00\x80\x1F\x2D\x00\x80\x18\x37"
-     "\x00\x80"
-     "\x1C\x37\x00\x80\x20\x37\x00\x80\x28\x37\x00\x80\x80\xEC\x00\x80\x2C\x37"
-     "\x00\x80"
-     "\x80\x00\x00\x00\x40\x37\x00\x80\x44\x37\x00\x80\x30\x37\x00\x80\x48\x37"
-     "\x00\x80"
-     "\x5C\x37\x00\x80\xB8\xEB\x00\x80\x3C\x37\x00\x80\x50\x37\x00\x80\x54\x37"
-     "\x00\x80"
-     "\x58\x37\x00\x80\x85\xEC\x00\x80\x8F\xEC\x00\x80\x8A\xEC\x00\x80\x48\x2B"
-     "\x00\x80"
-     "\x78\x37\x00\x80\x7C\x37\x00\x80\x84\x37\x00\x80\xA3\xEC\x00\x80\x88\x37"
-     "\x00\x80"
-     "\x94\xEC\x00\x80\x4C\x2B\x00\x80\x50\x2B\x00\x80\x99\xEC\x00\x80\x9E\xEC"
-     "\x00\x80"
-     "\x54\x2B\x00\x80\xA8\x37\x00\x80\xB0\x37\x00\x80\xB4\x37\x00\x80\xC0\x37"
-     "\x00\x80"
-     "\xBC\x37\x00\x80\xC8\x37\x00\x80\xC4\x37\x00\x80\xCC\x37\x00\x80\xD0\x37"
-     "\x00\x80"
-     "\xD8\x37\x00\x80\x58\x2B\x00\x80\xE0\x37\x00\x80\xE4\x37\x00\x80\x5C\x2B"
-     "\x00\x80"
-     "\xF4\x37\x00\x80\xA8\xEC\x00\x80\x04\x38\x00\x80\x08\x38\x00\x80\x60\x2B"
-     "\x00\x80"
-     "\x14\x38\x00\x80\x94\x2A\x00\x80\xAD\xEC\x00\x80\xB2\xEC\x00\x80\x64\x2B"
-     "\x00\x80"
-     "\x68\x2B\x00\x80\x40\x38\x00\x80\xB7\xEC\x00\x80\x98\x38\x00\x80\x9C\x38"
-     "\x00\x80"
-     "\xB4\x38\x00\x80\xC0\x38\x00\x80\xC1\xEC\x00\x80\x13\xEB\x00\x80\xCC\x38"
-     "\x00\x80"
-     "\xC8\x38\x00\x80\xD4\x38\x00\x80\x18\xEB\x00\x80\xE0\x38\x00\x80\x80\x95"
-     "\x07\x00"
-     "\x81\x9D\x07\x00\x82\x95\x07\x00\x83\xAD\x07\x00\x84\xBD\xF7\x00\x85\xB1"
-     "\x07\x00"
-     "\x86\xB1\x07\x00\x87\xB1\x07\x00\x88\x91\x07\x00\x89\x91\x07\x00\x8A\x91"
-     "\x07\x00"
-     "\x8B\x91\x07\x00\x8C\xF1\x07\x00\x8D\xF1\x07\x00\x8E\xF1\x07\x00\x8F\xF1"
-     "\x07\x00"
-     "\x90\xDD\xC4\x00\x91\xE5\xC4\x00\x92\xA9\xF7\x00\x93\x9D\x07\x00\x94\x85"
-     "\x07\x00"
-     "\x95\x8D\x07\x00\x96\x85\x07\x00\x97\xBD\x07\x00\x98\x85\x07\x00\x99\x8D"
-     "\x07\x00"
-     "\x9A\x85\x07\x00\x9B\x9D\x07\x00\x9C\x85\x07\x00\x9D\x8D\x07\x00\x9E\x85"
-     "\x07\x00"
-     "\x9F\xD9\xC4\x00\xA0\x85\x02\x00\xA1\x8D\x02\x00\xA2\x85\x02\x00\xA3\x9D"
-     "\x02\x00"
-     "\xA4\x85\x02\x00\xA5\x8D\x02\x00\xA6\x85\x02\x00\xA7\xBD\x02\x00\xA8\x85"
-     "\x02\x00"
-     "\xA9\x8D\x02\x00\xAA\x85\x02\x00\xAB\x9D\x02\x00\xAC\x85\x02\x00\xAD\x8D"
-     "\x02\x00"
-     "\xAE\x85\x02\x00\xAF\xFD\x02\x00\xB0\x85\x02\x00\xB1\x8D\x02\x00\xB2\x85"
-     "\x02\x00"
-     "\xB3\x9D\x02\x00\xB4\x85\x02\x00\xB5\x8D\x02\x00\xB6\x85\x02\x00\xB7\xBD"
-     "\x02\x00"
-     "\xB8\x85\x02\x00\xB9\x8D\x02\x00\xBA\x85\x02\x00\xBB\x9D\x02\x00\xBC\x85"
-     "\x02\x00"
-     "\xBD\x8D\x02\x00\xBE\x45\xC4\x00\xBF\x41\xC4\x00\xA7\x39\x06\x00\xA6\x31"
-     "\x06\x00"
-     "\xA5\x39\x06\x00\xA4\x21\x06\x00\xA3\x39\x06\x00\xA2\x31\x06\x00\xA1\x39"
-     "\x06\x00"
-     "\xA0\x01\x06\x00\xAF\x79\x06\x00\xAE\x71\x06\x00\xAD\x79\x06\x00\xAC\x61"
-     "\x06\x00"
-     "\xAB\x79\x06\x00\xAA\x71\x06\x00\xA9\x79\x06\x00\xA8\x01\x06\x00\xB7\x35"
-     "\x06\x00"
-     "\xB6\x35\x06\x00\xB5\x25\x06\x00\xB4\xC5\xD0\x00\xB3\x39\x06\x00\xB2\x31"
-     "\x06\x00"
-     "\xB1\x39\x06\x00\xB0\x01\x06\x00\xBF\xF1\x06\x00\xBE\xE9\x06\x00\xBD\xE1"
-     "\x06\x00"
-     "\xBC\xE9\x06\x00\xBB\xF9\xC5\x00\xBA\xF5\x06\x00\xB9\x05\x06\x00\xB8\x05"
-     "\x06\x00"
-     "\x87\x3D\x07\x00\x86\x3D\x07\x00\x85\x3D\x07\x00\x84\x3D\x07\x00\x83\x3D"
-     "\x07\x00"
-     "\x82\x3D\x07\x00\x81\x3D\x07\x00\x80\x3D\x07\x00\x8F\x79\x07\x00\x8E\x71"
-     "\x07\x00"
-     "\x8D\x79\x07\x00\x8C\x61\x07\x00\x8B\xC1\x1E\x00\x8A\x7D\x07\x00\x89\x7D"
-     "\x07\x00"
-     "\x88\x7D\x07\x00\x97\x39\x07\x00\x96\x31\x07\x00\x95\x39\x07\x00\x94\x21"
-     "\x07\x00"
-     "\x93\x39\x07\x00\x92\x31\x07\x00\x91\x39\x07\x00\x90\x01\x07\x00\x9F\x65"
-     "\xD8\x00"
-     "\x9E\xF1\x06\x00\x9D\xF9\x06\x00\x9C\xE1\x06\x00\x9B\xF9\x06\x00\x9A\xF1"
-     "\x06\x00"
-     "\x99\x05\x07\x00\x98\x89\xDD\x00\x80\x00\x00\x00\xC6\xEC\x00\x80\xCB\xEC"
-     "\x00\x80"
-     "\x2C\x39\x00\x80\x38\x39\x00\x80\x40\x39\x00\x80\xD0\xEC\x00\x80\x44\x39"
-     "\x00\x80"
-     "\x70\x39\x00\x80\x7C\x39\x00\x80\x80\x39\x00\x80\x78\x39\x00\x80\x8C\x39"
-     "\x00\x80"
-     "\x90\x39\x00\x80\xD5\xEC\x00\x80\x9C\x39\x00\x80\x6C\x2B\x00\x80\xA4\x39"
-     "\x00\x80"
-     "\xDA\xEC\x00\x80\x70\x2B\x00\x80\xDC\x39\x00\x80\xAA\x2F\x00\x80\xF4\x39"
-     "\x00\x80"
-     "\xDF\xEC\x00\x80\xE9\xEC\x00\x80\x74\x2B\x00\x80\x78\x2B\x00\x80\x1C\x3A"
-     "\x00\x80"
-     "\xEE\xEC\x00\x80\x7C\x2B\x00\x80\xF3\xEC\x00\x80\x3C\x3A\x00\x80\x80\xFD"
-     "\x00\x00"
-     "\x81\x7D\x00\x00\x82\x75\x00\x00\x83\x8D\x01\x00\x84\x95\x01\x00\x85\x9D"
-     "\x01\x00"
-     "\x86\x95\x01\x00\x87\x8D\x01\x00\x88\xB5\x01\x00\x89\xBD\x01\x00\x8A\x69"
-     "\xC6\x00"
-     "\x8B\xB1\x01\x00\x8C\x91\x01\x00\x8D\x91\x01\x00\x8E\x91\x01\x00\x8F\x91"
-     "\x01\x00"
-     "\x90\xF1\x01\x00\x91\xF1\x01\x00\x92\xF1\x01\x00\x93\xF1\x01\x00\x94\x91"
-     "\x01\x00"
-     "\x95\x8D\xFE\x00\x96\x9D\x01\x00\x97\xB5\xFE\x00\x98\xA9\x01\x00\x99\xA9"
-     "\x01\x00"
-     "\x9A\xB9\x01\x00\x9B\xB9\x01\x00\x9C\x8D\xFE\x00\x9D\xAD\x01\x00\xF8\xEC"
-     "\x00\x80"
-     "\x54\x3A\x00\x80\x80\x2B\x00\x80\x64\x3A\x00\x80\x70\x3A\x00\x80\x78\x3A"
-     "\x00\x80"
-     "\x80\x3A\x00\x80\x84\x2B\x00\x80\xFD\xEC\x00\x80\xB0\x3A\x00\x80\xBC\x3A"
-     "\x00\x80"
-     "\x88\x2B\x00\x80\x8C\x2B\x00\x80\xC0\x3A\x00\x80\x02\xED\x00\x80\x90\x2B"
-     "\x00\x80"
-     "\x07\xED\x00\x80\x0C\xED\x00\x80\x11\xED\x00\x80\x94\x2B\x00\x80\x00\x3B"
-     "\x00\x80"
-     "\x04\x3B\x00\x80\x0C\x3B\x00\x80\x14\x3B\x00\x80\x1B\xED\x00\x80\xD7\x00"
-     "\x00\x00"
-     "\xD8\x00\x00\x00\xD9\x00\x00\x00\xDA\x00\x00\x00\xDB\x00\x00\x00\xDC\x00"
-     "\x00\x00"
-     "\xDD\x00\x00\x00\xDE\x00\x00\x00\xDF\x00\x00\x00\xE0\x00\x00\x00\xE1\x00"
-     "\x00\x00"
-     "\xE2\x00\x00\x00\xE3\x00\x00\x00\xE4\x00\x00\x00\xE5\x00\x00\x00\xE6\x00"
-     "\x00\x00"
-     "\xE7\x00\x00\x00\xE8\x00\x00\x00\xE9\x00\x00\x00\xEA\x00\x00\x00\xEB\x00"
-     "\x00\x00"
-     "\xEC\x00\x00\x00\xED\x00\x00\x00\xEE\x00\x00\x00\xEF\x00\x00\x00\xF0\x00"
-     "\x00\x00"
-     "\xF1\x00\x00\x00\xF2\x00\x00\x00\xF3\x00\x00\x00\xF4\x00\x00\x00\xF5\x00"
-     "\x00\x00"
-     "\xF6\x00\x00\x00\xF7\x00\x00\x00\xF8\x00\x00\x00\xF9\x00\x00\x00\xFA\x00"
-     "\x00\x00"
-     "\xFB\x00\x00\x00\xFC\x00\x00\x00\xFD\x00\x00\x00\xFE\x00\x00\x00\xFF\x00"
-     "\x00\x00"
-     "\x20\x00\x20\xCC\x81\x00\x20\xCC\x83\x00\x20\xCC\x84\x00\x20\xCC\x85\x00"
-     "\x20\xCC"
-     "\x86\x00\x20\xCC\x87\x00\x20\xCC\x88\x00\x20\xCC\x88\xCC\x80\x00\x20\xCC"
-     "\x88\xCC"
-     "\x81\x00\x20\xCC\x88\xCD\x82\x00\x20\xCC\x8A\x00\x20\xCC\x8B\x00\x20\xCC"
-     "\x93\x00"
-     "\x20\xCC\x93\xCC\x80\x00\x20\xCC\x93\xCC\x81\x00\x20\xCC\x93\xCD\x82\x00"
-     "\x20\xCC"
-     "\x94\x00\x20\xCC\x94\xCC\x80\x00\x20\xCC\x94\xCC\x81\x00\x20\xCC\x94\xCD"
-     "\x82\x00"
-     "\x20\xCC\xA7\x00\x20\xCC\xA8\x00\x20\xCC\xB3\x00\x20\xCD\x82\x00\x20\xCD"
-     "\x85\x00"
-     "\x20\xD9\x8B\x00\x20\xD9\x8C\x00\x20\xD9\x8C\xD9\x91\x00\x20\xD9\x8D\x00"
-     "\x20\xD9"
-     "\x8D\xD9\x91\x00\x20\xD9\x8E\x00\x20\xD9\x8E\xD9\x91\x00\x20\xD9\x8F\x00"
-     "\x20\xD9"
-     "\x8F\xD9\x91\x00\x20\xD9\x90\x00\x20\xD9\x90\xD9\x91\x00\x20\xD9\x91\x00"
-     "\x20\xD9"
-     "\x91\xD9\xB0\x00\x20\xD9\x92\x00\x20\xE3\x82\x99\x00\x20\xE3\x82\x9A\x00"
-     "\x21\x00"
-     "\x21\x21\x00\x21\x3F\x00\x22\x00\x23\x00\x24\x00\x25\x00\x26\x00\x27\x00"
-     "\x28\x00"
-     "\x28\x31\x29\x00\x28\x31\x30\x29\x00\x28\x31\x31\x29\x00\x28\x31\x32\x29"
-     "\x00\x28"
-     "\x31\x33\x29\x00\x28\x31\x34\x29\x00\x28\x31\x35\x29\x00\x28\x31\x36\x29"
-     "\x00\x28"
-     "\x31\x37\x29\x00\x28\x31\x38\x29\x00\x28\x31\x39\x29\x00\x28\x32\x29\x00"
-     "\x28\x32"
-     "\x30\x29\x00\x28\x33\x29\x00\x28\x34\x29\x00\x28\x35\x29\x00\x28\x36\x29"
-     "\x00\x28"
-     "\x37\x29\x00\x28\x38\x29\x00\x28\x39\x29\x00\x28\x41\x29\x00\x28\x42\x29"
-     "\x00\x28"
-     "\x43\x29\x00\x28\x44\x29\x00\x28\x45\x29\x00\x28\x46\x29\x00\x28\x47\x29"
-     "\x00\x28"
-     "\x48\x29\x00\x28\x49\x29\x00\x28\x4A\x29\x00\x28\x4B\x29\x00\x28\x4C\x29"
-     "\x00\x28"
-     "\x4D\x29\x00\x28\x4E\x29\x00\x28\x4F\x29\x00\x28\x50\x29\x00\x28\x51\x29"
-     "\x00\x28"
-     "\x52\x29\x00\x28\x53\x29\x00\x28\x54\x29\x00\x28\x55\x29\x00\x28\x56\x29"
-     "\x00\x28"
-     "\x57\x29\x00\x28\x58\x29\x00\x28\x59\x29\x00\x28\x5A\x29\x00\x28\x61\x29"
-     "\x00\x28"
-     "\x62\x29\x00\x28\x63\x29\x00\x28\x64\x29\x00\x28\x65\x29\x00\x28\x66\x29"
-     "\x00\x28"
-     "\x67\x29\x00\x28\x68\x29\x00\x28\x69\x29\x00\x28\x6A\x29\x00\x28\x6B\x29"
-     "\x00\x28"
-     "\x6C\x29\x00\x28\x6D\x29\x00\x28\x6E\x29\x00\x28\x6F\x29\x00\x28\x70\x29"
-     "\x00\x28"
-     "\x71\x29\x00\x28\x72\x29\x00\x28\x73\x29\x00\x28\x74\x29\x00\x28\x75\x29"
-     "\x00\x28"
-     "\x76\x29\x00\x28\x77\x29\x00\x28\x78\x29\x00\x28\x79\x29\x00\x28\x7A\x29"
-     "\x00\x28"
-     "\xE1\x84\x80\x29\x00\x28\xE1\x84\x82\x29\x00\x28\xE1\x84\x83\x29\x00\x28"
-     "\xE1\x84"
-     "\x85\x29\x00\x28\xE1\x84\x86\x29\x00\x28\xE1\x84\x87\x29\x00\x28\xE1\x84"
-     "\x89\x29"
-     "\x00\x28\xE1\x84\x8B\x29\x00\x28\xE1\x84\x8C\x29\x00\x28\xE1\x84\x8E\x29"
-     "\x00\x28"
-     "\xE1\x84\x8F\x29\x00\x28\xE1\x84\x90\x29\x00\x28\xE1\x84\x91\x29\x00\x28"
-     "\xE1\x84"
-     "\x92\x29\x00\x28\xE4\xB8\x80\x29\x00\x28\xE4\xB8\x83\x29\x00\x28\xE4\xB8"
-     "\x89\x29"
-     "\x00\x28\xE4\xB9\x9D\x29\x00\x28\xE4\xBA\x8C\x29\x00\x28\xE4\xBA\x94\x29"
-     "\x00\x28"
-     "\xE4\xBB\xA3\x29\x00\x28\xE4\xBC\x81\x29\x00\x28\xE4\xBC\x91\x29\x00\x28"
-     "\xE5\x85"
-     "\xAB\x29\x00\x28\xE5\x85\xAD\x29\x00\x28\xE5\x8A\xB4\x29\x00\x28\xE5\x8D"
-     "\x81\x29"
-     "\x00\x28\xE5\x8D\x94\x29\x00\x28\xE5\x90\x8D\x29\x00\x28\xE5\x91\xBC\x29"
-     "\x00\x28"
-     "\xE5\x9B\x9B\x29\x00\x28\xE5\x9C\x9F\x29\x00\x28\xE5\xAD\xA6\x29\x00\x28"
-     "\xE6\x97"
-     "\xA5\x29\x00\x28\xE6\x9C\x88\x29\x00\x28\xE6\x9C\x89\x29\x00\x28\xE6\x9C"
-     "\xA8\x29"
-     "\x00\x28\xE6\xA0\xAA\x29\x00\x28\xE6\xB0\xB4\x29\x00\x28\xE7\x81\xAB\x29"
-     "\x00\x28"
-     "\xE7\x89\xB9\x29\x00\x28\xE7\x9B\xA3\x29\x00\x28\xE7\xA4\xBE\x29\x00\x28"
-     "\xE7\xA5"
-     "\x9D\x29\x00\x28\xE7\xA5\xAD\x29\x00\x28\xE8\x87\xAA\x29\x00\x28\xE8\x87"
-     "\xB3\x29"
-     "\x00\x28\xE8\xB2\xA1\x29\x00\x28\xE8\xB3\x87\x29\x00\x28\xE9\x87\x91\x29"
-     "\x00\x28"
-     "\xEA\xB0\x80\x29\x00\x28\xEB\x82\x98\x29\x00\x28\xEB\x8B\xA4\x29\x00\x28"
-     "\xEB\x9D"
-     "\xBC\x29\x00\x28\xEB\xA7\x88\x29\x00\x28\xEB\xB0\x94\x29\x00\x28\xEC\x82"
-     "\xAC\x29"
-     "\x00\x28\xEC\x95\x84\x29\x00\x28\xEC\x98\xA4\xEC\xA0\x84\x29\x00\x28\xEC"
-     "\x98\xA4"
-     "\xED\x9B\x84\x29\x00\x28\xEC\x9E\x90\x29\x00\x28\xEC\xA3\xBC\x29\x00\x28"
-     "\xEC\xB0"
-     "\xA8\x29\x00\x28\xEC\xB9\xB4\x29\x00\x28\xED\x83\x80\x29\x00\x28\xED\x8C"
-     "\x8C\x29"
-     "\x00\x28\xED\x95\x98\x29\x00\x29\x00\x2A\x00\x2B\x00\x2C\x00\x2D\x00\x2E"
-     "\x00\x2E"
-     "\x2E\x00\x2E\x2E\x2E\x00\x2F\x00\x30\x00\x30\x2C\x00\x30\x2E\x00\x30\xE2"
-     "\x81\x84"
-     "\x33\x00\x30\xE7\x82\xB9\x00\x31\x00\x31\x2C\x00\x31\x2E\x00\x31\x30\x00"
-     "\x31\x30"
-     "\x2E\x00\x31\x30\xE6\x97\xA5\x00\x31\x30\xE6\x9C\x88\x00\x31\x30\xE7\x82"
-     "\xB9\x00"
-     "\x31\x31\x00\x31\x31\x2E\x00\x31\x31\xE6\x97\xA5\x00\x31\x31\xE6\x9C\x88"
-     "\x00\x31"
-     "\x31\xE7\x82\xB9\x00\x31\x32\x00\x31\x32\x2E\x00\x31\x32\xE6\x97\xA5\x00"
-     "\x31\x32"
-     "\xE6\x9C\x88\x00\x31\x32\xE7\x82\xB9\x00\x31\x33\x00\x31\x33\x2E\x00\x31"
-     "\x33\xE6"
-     "\x97\xA5\x00\x31\x33\xE7\x82\xB9\x00\x31\x34\x00\x31\x34\x2E\x00\x31\x34"
-     "\xE6\x97"
-     "\xA5\x00\x31\x34\xE7\x82\xB9\x00\x31\x35\x00\x31\x35\x2E\x00\x31\x35\xE6"
-     "\x97\xA5"
-     "\x00\x31\x35\xE7\x82\xB9\x00\x31\x36\x00\x31\x36\x2E\x00\x31\x36\xE6\x97"
-     "\xA5\x00"
-     "\x31\x36\xE7\x82\xB9\x00\x31\x37\x00\x31\x37\x2E\x00\x31\x37\xE6\x97\xA5"
-     "\x00\x31"
-     "\x37\xE7\x82\xB9\x00\x31\x38\x00\x31\x38\x2E\x00\x31\x38\xE6\x97\xA5\x00"
-     "\x31\x38"
-     "\xE7\x82\xB9\x00\x31\x39\x00\x31\x39\x2E\x00\x31\x39\xE6\x97\xA5\x00\x31"
-     "\x39\xE7"
-     "\x82\xB9\x00\x31\xE2\x81\x84\x00\x31\xE2\x81\x84\x31\x30\x00\x31\xE2\x81"
-     "\x84\x32"
-     "\x00\x31\xE2\x81\x84\x33\x00\x31\xE2\x81\x84\x34\x00\x31\xE2\x81\x84\x35"
-     "\x00\x31"
-     "\xE2\x81\x84\x36\x00\x31\xE2\x81\x84\x37\x00\x31\xE2\x81\x84\x38\x00\x31"
-     "\xE2\x81"
-     "\x84\x39\x00\x31\xE6\x97\xA5\x00\x31\xE6\x9C\x88\x00\x31\xE7\x82\xB9\x00"
-     "\x32\x00"
-     "\x32\x2C\x00\x32\x2E\x00\x32\x30\x00\x32\x30\x2E\x00\x32\x30\xE6\x97\xA5"
-     "\x00\x32"
-     "\x30\xE7\x82\xB9\x00\x32\x31\x00\x32\x31\xE6\x97\xA5\x00\x32\x31\xE7\x82"
-     "\xB9\x00"
-     "\x32\x32\x00\x32\x32\xE6\x97\xA5\x00\x32\x32\xE7\x82\xB9\x00\x32\x33\x00"
-     "\x32\x33"
-     "\xE6\x97\xA5\x00\x32\x33\xE7\x82\xB9\x00\x32\x34\x00\x32\x34\xE6\x97\xA5"
-     "\x00\x32"
-     "\x34\xE7\x82\xB9\x00\x32\x35\x00\x32\x35\xE6\x97\xA5\x00\x32\x36\x00\x32"
-     "\x36\xE6"
-     "\x97\xA5\x00\x32\x37\x00\x32\x37\xE6\x97\xA5\x00\x32\x38\x00\x32\x38\xE6"
-     "\x97\xA5"
-     "\x00\x32\x39\x00\x32\x39\xE6\x97\xA5\x00\x32\xE2\x81\x84\x33\x00\x32\xE2"
-     "\x81\x84"
-     "\x35\x00\x32\xE6\x97\xA5\x00\x32\xE6\x9C\x88\x00\x32\xE7\x82\xB9\x00\x33"
-     "\x00\x33"
-     "\x2C\x00\x33\x2E\x00\x33\x30\x00\x33\x30\xE6\x97\xA5\x00\x33\x31\x00\x33"
-     "\x31\xE6"
-     "\x97\xA5\x00\x33\x32\x00\x33\x33\x00\x33\x34\x00\x33\x35\x00\x33\x36\x00"
-     "\x33\x37"
-     "\x00\x33\x38\x00\x33\x39\x00\x33\xE2\x81\x84\x34\x00\x33\xE2\x81\x84\x35"
-     "\x00\x33"
-     "\xE2\x81\x84\x38\x00\x33\xE6\x97\xA5\x00\x33\xE6\x9C\x88\x00\x33\xE7\x82"
-     "\xB9\x00"
-     "\x34\x00\x34\x2C\x00\x34\x2E\x00\x34\x30\x00\x34\x31\x00\x34\x32\x00\x34"
-     "\x33\x00"
-     "\x34\x34\x00\x34\x35\x00\x34\x36\x00\x34\x37\x00\x34\x38\x00\x34\x39\x00"
-     "\x34\xE2"
-     "\x81\x84\x35\x00\x34\xE6\x97\xA5\x00\x34\xE6\x9C\x88\x00\x34\xE7\x82\xB9"
-     "\x00\x35"
-     "\x00\x35\x2C\x00\x35\x2E\x00\x35\x30\x00\x35\xE2\x81\x84\x36\x00\x35\xE2"
-     "\x81\x84"
-     "\x38\x00\x35\xE6\x97\xA5\x00\x35\xE6\x9C\x88\x00\x35\xE7\x82\xB9\x00\x36"
-     "\x00\x36"
-     "\x2C\x00\x36\x2E\x00\x36\xE6\x97\xA5\x00\x36\xE6\x9C\x88\x00\x36\xE7\x82"
-     "\xB9\x00"
-     "\x37\x00\x37\x2C\x00\x37\x2E\x00\x37\xE2\x81\x84\x38\x00\x37\xE6\x97\xA5"
-     "\x00\x37"
-     "\xE6\x9C\x88\x00\x37\xE7\x82\xB9\x00\x38\x00\x38\x2C\x00\x38\x2E\x00\x38"
-     "\xE6\x97"
-     "\xA5\x00\x38\xE6\x9C\x88\x00\x38\xE7\x82\xB9\x00\x39\x00\x39\x2C\x00\x39"
-     "\x2E\x00"
-     "\x39\xE6\x97\xA5\x00\x39\xE6\x9C\x88\x00\x39\xE7\x82\xB9\x00\x3A\x00\x3A"
-     "\x3A\x3D"
-     "\x00\x3B\x00\x3C\x00\x3D\x00\x3D\x3D\x00\x3D\x3D\x3D\x00\x3E\x00\x3F\x00"
-     "\x3F\x21"
-     "\x00\x3F\x3F\x00\x40\x00\x41\x00\x41\x55\x00\x41\xE2\x88\x95\x6D\x00\x42"
-     "\x00\x42"
-     "\x71\x00\x43\x00\x43\x44\x00\x43\x6F\x2E\x00\x43\xE2\x88\x95\x6B\x67\x00"
-     "\x44\x00"
-     "\x44\x4A\x00\x44\x5A\x00\x44\x7A\x00\x44\xC5\xBD\x00\x44\xC5\xBE\x00\x45"
-     "\x00\x46"
-     "\x00\x46\x41\x58\x00\x47\x00\x47\x42\x00\x47\x48\x7A\x00\x47\x50\x61\x00"
-     "\x47\x79"
-     "\x00\x48\x00\x48\x50\x00\x48\x56\x00\x48\x67\x00\x48\x7A\x00\x49\x00\x49"
-     "\x49\x00"
-     "\x49\x49\x49\x00\x49\x4A\x00\x49\x55\x00\x49\x56\x00\x49\x58\x00\x4A\x00"
-     "\x4B\x00"
-     "\x4B\x42\x00\x4B\x4B\x00\x4B\x4D\x00\x4C\x00\x4C\x4A\x00\x4C\x54\x44\x00"
-     "\x4C\x6A"
-     "\x00\x4C\xC2\xB7\x00\x4D\x00\x4D\x42\x00\x4D\x43\x00\x4D\x44\x00\x4D\x48"
-     "\x7A\x00"
-     "\x4D\x50\x61\x00\x4D\x52\x00\x4D\x56\x00\x4D\x57\x00\x4D\xCE\xA9\x00\x4E"
-     "\x00\x4E"
-     "\x4A\x00\x4E\x6A\x00\x4E\x6F\x00\x4F\x00\x50\x00\x50\x48\x00\x50\x50\x4D"
-     "\x00\x50"
-     "\x50\x56\x00\x50\x52\x00\x50\x54\x45\x00\x50\x61\x00\x51\x00\x52\x00\x52"
-     "\x73\x00"
-     "\x53\x00\x53\x44\x00\x53\x4D\x00\x53\x53\x00\x53\x76\x00\x54\x00\x54\x45"
-     "\x4C\x00"
-     "\x54\x48\x7A\x00\x54\x4D\x00\x55\x00\x56\x00\x56\x49\x00\x56\x49\x49\x00"
-     "\x56\x49"
-     "\x49\x49\x00\x56\xE2\x88\x95\x6D\x00\x57\x00\x57\x43\x00\x57\x5A\x00\x57"
-     "\x62\x00"
-     "\x58\x00\x58\x49\x00\x58\x49\x49\x00\x59\x00\x5A\x00\x5B\x00\x5C\x00\x5D"
-     "\x00\x5E"
-     "\x00\x5F\x00\x60\x00\x61\x00\x61\x2E\x6D\x2E\x00\x61\x2F\x63\x00\x61\x2F"
-     "\x73\x00"
-     "\x61\xCA\xBE\x00\x62\x00\x62\x61\x72\x00\x63\x00\x63\x2F\x6F\x00\x63\x2F"
-     "\x75\x00"
-     "\x63\x61\x6C\x00\x63\x63\x00\x63\x64\x00\x63\x6D\x00\x63\x6D\x32\x00\x63"
-     "\x6D\x33"
-     "\x00\x64\x00\x64\x42\x00\x64\x61\x00\x64\x6C\x00\x64\x6D\x00\x64\x6D\x32"
-     "\x00\x64"
-     "\x6D\x33\x00\x64\x7A\x00\x64\xC5\xBE\x00\x65\x00\x65\x56\x00\x65\x72\x67"
-     "\x00\x66"
-     "\x00\x66\x66\x00\x66\x66\x69\x00\x66\x66\x6C\x00\x66\x69\x00\x66\x6C\x00"
-     "\x66\x6D"
-     "\x00\x67\x00\x67\x61\x6C\x00\x68\x00\x68\x50\x61\x00\x68\x61\x00\x69\x00"
-     "\x69\x69"
-     "\x00\x69\x69\x69\x00\x69\x6A\x00\x69\x6E\x00\x69\x76\x00\x69\x78\x00\x6A"
-     "\x00\x6B"
-     "\x00\x6B\x41\x00\x6B\x48\x7A\x00\x6B\x50\x61\x00\x6B\x56\x00\x6B\x57\x00"
-     "\x6B\x63"
-     "\x61\x6C\x00\x6B\x67\x00\x6B\x6C\x00\x6B\x6D\x00\x6B\x6D\x32\x00\x6B\x6D"
-     "\x33\x00"
-     "\x6B\x74\x00\x6B\xCE\xA9\x00\x6C\x00\x6C\x6A\x00\x6C\x6D\x00\x6C\x6E\x00"
-     "\x6C\x6F"
-     "\x67\x00\x6C\x78\x00\x6C\xC2\xB7\x00\x6D\x00\x6D\x32\x00\x6D\x33\x00\x6D"
-     "\x41\x00"
-     "\x6D\x56\x00\x6D\x57\x00\x6D\x62\x00\x6D\x67\x00\x6D\x69\x6C\x00\x6D\x6C"
-     "\x00\x6D"
-     "\x6D\x00\x6D\x6D\x32\x00\x6D\x6D\x33\x00\x6D\x6F\x6C\x00\x6D\x73\x00\x6D"
-     "\xE2\x88"
-     "\x95\x73\x00\x6D\xE2\x88\x95\x73\x32\x00\x6E\x00\x6E\x41\x00\x6E\x46\x00"
-     "\x6E\x56"
-     "\x00\x6E\x57\x00\x6E\x6A\x00\x6E\x6D\x00\x6E\x73\x00\x6F\x00\x6F\x56\x00"
-     "\x70\x00"
-     "\x70\x2E\x6D\x2E\x00\x70\x41\x00\x70\x46\x00\x70\x56\x00\x70\x57\x00\x70"
-     "\x63\x00"
-     "\x70\x73\x00\x71\x00\x72\x00\x72\x61\x64\x00\x72\x61\x64\xE2\x88\x95\x73"
-     "\x00\x72"
-     "\x61\x64\xE2\x88\x95\x73\x32\x00\x73\x00\x73\x72\x00\x73\x74\x00\x74\x00"
-     "\x75\x00"
-     "\x76\x00\x76\x69\x00\x76\x69\x69\x00\x76\x69\x69\x69\x00\x77\x00\x78\x00"
-     "\x78\x69"
-     "\x00\x78\x69\x69\x00\x79\x00\x7A\x00\x7B\x00\x7C\x00\x7D\x00\x7E\x00\xC2"
-     "\xA2\x00"
-     "\xC2\xA3\x00\xC2\xA5\x00\xC2\xA6\x00\xC2\xAC\x00\xC2\xB0\x43\x00\xC2\xB0"
-     "\x46\x00"
-     "\xC2\xB7\x00\xC3\x80\x00\xC3\x81\x00\xC3\x82\x00\xC3\x83\x00\xC3\x84\x00"
-     "\xC3\x85"
-     "\x00\xC3\x86\x00\xC3\x87\x00\xC3\x88\x00\xC3\x89\x00\xC3\x8A\x00\xC3\x8B"
-     "\x00\xC3"
-     "\x8C\x00\xC3\x8D\x00\xC3\x8E\x00\xC3\x8F\x00\xC3\x91\x00\xC3\x92\x00\xC3"
-     "\x93\x00"
-     "\xC3\x94\x00\xC3\x95\x00\xC3\x96\x00\xC3\x99\x00\xC3\x9A\x00\xC3\x9B\x00"
-     "\xC3\x9C"
-     "\x00\xC3\x9D\x00\xC3\xA0\x00\xC3\xA1\x00\xC3\xA2\x00\xC3\xA3\x00\xC3\xA4"
-     "\x00\xC3"
-     "\xA5\x00\xC3\xA6\x00\xC3\xA7\x00\xC3\xA8\x00\xC3\xA9\x00\xC3\xAA\x00\xC3"
-     "\xAB\x00"
-     "\xC3\xAC\x00\xC3\xAD\x00\xC3\xAE\x00\xC3\xAF\x00\xC3\xB0\x00\xC3\xB1\x00"
-     "\xC3\xB2"
-     "\x00\xC3\xB3\x00\xC3\xB4\x00\xC3\xB5\x00\xC3\xB6\x00\xC3\xB8\x00\xC3\xB9"
-     "\x00\xC3"
-     "\xBA\x00\xC3\xBB\x00\xC3\xBC\x00\xC3\xBD\x00\xC3\xBF\x00\xC4\x80\x00\xC4"
-     "\x81\x00"
-     "\xC4\x82\x00\xC4\x83\x00\xC4\x84\x00\xC4\x85\x00\xC4\x86\x00\xC4\x87\x00"
-     "\xC4\x88"
-     "\x00\xC4\x89\x00\xC4\x8A\x00\xC4\x8B\x00\xC4\x8C\x00\xC4\x8D\x00\xC4\x8E"
-     "\x00\xC4"
-     "\x8F\x00\xC4\x92\x00\xC4\x93\x00\xC4\x94\x00\xC4\x95\x00\xC4\x96\x00\xC4"
-     "\x97\x00"
-     "\xC4\x98\x00\xC4\x99\x00\xC4\x9A\x00\xC4\x9B\x00\xC4\x9C\x00\xC4\x9D\x00"
-     "\xC4\x9E"
-     "\x00\xC4\x9F\x00\xC4\xA0\x00\xC4\xA1\x00\xC4\xA2\x00\xC4\xA3\x00\xC4\xA4"
-     "\x00\xC4"
-     "\xA5\x00\xC4\xA6\x00\xC4\xA7\x00\xC4\xA8\x00\xC4\xA9\x00\xC4\xAA\x00\xC4"
-     "\xAB\x00"
-     "\xC4\xAC\x00\xC4\xAD\x00\xC4\xAE\x00\xC4\xAF\x00\xC4\xB0\x00\xC4\xB1\x00"
-     "\xC4\xB4"
-     "\x00\xC4\xB5\x00\xC4\xB6\x00\xC4\xB7\x00\xC4\xB9\x00\xC4\xBA\x00\xC4\xBB"
-     "\x00\xC4"
-     "\xBC\x00\xC4\xBD\x00\xC4\xBE\x00\xC5\x83\x00\xC5\x84\x00\xC5\x85\x00\xC5"
-     "\x86\x00"
-     "\xC5\x87\x00\xC5\x88\x00\xC5\x8B\x00\xC5\x8C\x00\xC5\x8D\x00\xC5\x8E\x00"
-     "\xC5\x8F"
-     "\x00\xC5\x90\x00\xC5\x91\x00\xC5\x93\x00\xC5\x94\x00\xC5\x95\x00\xC5\x96"
-     "\x00\xC5"
-     "\x97\x00\xC5\x98\x00\xC5\x99\x00\xC5\x9A\x00\xC5\x9B\x00\xC5\x9C\x00\xC5"
-     "\x9D\x00"
-     "\xC5\x9E\x00\xC5\x9F\x00\xC5\xA0\x00\xC5\xA1\x00\xC5\xA2\x00\xC5\xA3\x00"
-     "\xC5\xA4"
-     "\x00\xC5\xA5\x00\xC5\xA8\x00\xC5\xA9\x00\xC5\xAA\x00\xC5\xAB\x00\xC5\xAC"
-     "\x00\xC5"
-     "\xAD\x00\xC5\xAE\x00\xC5\xAF\x00\xC5\xB0\x00\xC5\xB1\x00\xC5\xB2\x00\xC5"
-     "\xB3\x00"
-     "\xC5\xB4\x00\xC5\xB5\x00\xC5\xB6\x00\xC5\xB7\x00\xC5\xB8\x00\xC5\xB9\x00"
-     "\xC5\xBA"
-     "\x00\xC5\xBB\x00\xC5\xBC\x00\xC5\xBD\x00\xC5\xBE\x00\xC6\x8E\x00\xC6\x90"
-     "\x00\xC6"
-     "\xA0\x00\xC6\xA1\x00\xC6\xAB\x00\xC6\xAF\x00\xC6\xB0\x00\xC7\x80\x00\xC7"
-     "\x81\x00"
-     "\xC7\x82\x00\xC7\x8D\x00\xC7\x8E\x00\xC7\x8F\x00\xC7\x90\x00\xC7\x91\x00"
-     "\xC7\x92"
-     "\x00\xC7\x93\x00\xC7\x94\x00\xC7\x95\x00\xC7\x96\x00\xC7\x97\x00\xC7\x98"
-     "\x00\xC7"
-     "\x99\x00\xC7\x9A\x00\xC7\x9B\x00\xC7\x9C\x00\xC7\x9E\x00\xC7\x9F\x00\xC7"
-     "\xA0\x00"
-     "\xC7\xA1\x00\xC7\xA2\x00\xC7\xA3\x00\xC7\xA6\x00\xC7\xA7\x00\xC7\xA8\x00"
-     "\xC7\xA9"
-     "\x00\xC7\xAA\x00\xC7\xAB\x00\xC7\xAC\x00\xC7\xAD\x00\xC7\xAE\x00\xC7\xAF"
-     "\x00\xC7"
-     "\xB0\x00\xC7\xB4\x00\xC7\xB5\x00\xC7\xB8\x00\xC7\xB9\x00\xC7\xBA\x00\xC7"
-     "\xBB\x00"
-     "\xC7\xBC\x00\xC7\xBD\x00\xC7\xBE\x00\xC7\xBF\x00\xC8\x80\x00\xC8\x81\x00"
-     "\xC8\x82"
-     "\x00\xC8\x83\x00\xC8\x84\x00\xC8\x85\x00\xC8\x86\x00\xC8\x87\x00\xC8\x88"
-     "\x00\xC8"
-     "\x89\x00\xC8\x8A\x00\xC8\x8B\x00\xC8\x8C\x00\xC8\x8D\x00\xC8\x8E\x00\xC8"
-     "\x8F\x00"
-     "\xC8\x90\x00\xC8\x91\x00\xC8\x92\x00\xC8\x93\x00\xC8\x94\x00\xC8\x95\x00"
-     "\xC8\x96"
-     "\x00\xC8\x97\x00\xC8\x98\x00\xC8\x99\x00\xC8\x9A\x00\xC8\x9B\x00\xC8\x9E"
-     "\x00\xC8"
-     "\x9F\x00\xC8\xA2\x00\xC8\xA6\x00\xC8\xA7\x00\xC8\xA8\x00\xC8\xA9\x00\xC8"
-     "\xAA\x00"
-     "\xC8\xAB\x00\xC8\xAC\x00\xC8\xAD\x00\xC8\xAE\x00\xC8\xAF\x00\xC8\xB0\x00"
-     "\xC8\xB1"
-     "\x00\xC8\xB2\x00\xC8\xB3\x00\xC8\xB7\x00\xC9\x90\x00\xC9\x91\x00\xC9\x92"
-     "\x00\xC9"
-     "\x93\x00\xC9\x94\x00\xC9\x95\x00\xC9\x96\x00\xC9\x97\x00\xC9\x98\x00\xC9"
-     "\x99\x00"
-     "\xC9\x9B\x00\xC9\x9C\x00\xC9\x9E\x00\xC9\x9F\x00\xC9\xA0\x00\xC9\xA1\x00"
-     "\xC9\xA2"
-     "\x00\xC9\xA3\x00\xC9\xA4\x00\xC9\xA5\x00\xC9\xA6\x00\xC9\xA7\x00\xC9\xA8"
-     "\x00\xC9"
-     "\xA9\x00\xC9\xAA\x00\xC9\xAB\x00\xC9\xAC\x00\xC9\xAD\x00\xC9\xAE\x00\xC9"
-     "\xAF\x00"
-     "\xC9\xB0\x00\xC9\xB1\x00\xC9\xB2\x00\xC9\xB3\x00\xC9\xB4\x00\xC9\xB5\x00"
-     "\xC9\xB6"
-     "\x00\xC9\xB7\x00\xC9\xB8\x00\xC9\xB9\x00\xC9\xBA\x00\xC9\xBB\x00\xC9\xBD"
-     "\x00\xC9"
-     "\xBE\x00\xCA\x80\x00\xCA\x81\x00\xCA\x82\x00\xCA\x83\x00\xCA\x84\x00\xCA"
-     "\x88\x00"
-     "\xCA\x89\x00\xCA\x8A\x00\xCA\x8B\x00\xCA\x8C\x00\xCA\x8D\x00\xCA\x8E\x00"
-     "\xCA\x8F"
-     "\x00\xCA\x90\x00\xCA\x91\x00\xCA\x92\x00\xCA\x95\x00\xCA\x98\x00\xCA\x99"
-     "\x00\xCA"
-     "\x9B\x00\xCA\x9C\x00\xCA\x9D\x00\xCA\x9F\x00\xCA\xA1\x00\xCA\xA2\x00\xCA"
-     "\xA3\x00"
-     "\xCA\xA4\x00\xCA\xA5\x00\xCA\xA6\x00\xCA\xA7\x00\xCA\xA8\x00\xCA\xA9\x00"
-     "\xCA\xAA"
-     "\x00\xCA\xAB\x00\xCA\xB9\x00\xCA\xBC\x6E\x00\xCB\x90\x00\xCB\x91\x00\xCC"
-     "\x80\x00"
-     "\xCC\x81\x00\xCC\x88\xCC\x81\x00\xCC\x93\x00\xCE\x86\x00\xCE\x88\x00\xCE"
-     "\x89\x00"
-     "\xCE\x8A\x00\xCE\x8C\x00\xCE\x8E\x00\xCE\x8F\x00\xCE\x90\x00\xCE\x91\x00"
-     "\xCE\x92"
-     "\x00\xCE\x93\x00\xCE\x94\x00\xCE\x95\x00\xCE\x96\x00\xCE\x97\x00\xCE\x98"
-     "\x00\xCE"
-     "\x99\x00\xCE\x9A\x00\xCE\x9B\x00\xCE\x9C\x00\xCE\x9D\x00\xCE\x9E\x00\xCE"
-     "\x9F\x00"
-     "\xCE\xA0\x00\xCE\xA1\x00\xCE\xA3\x00\xCE\xA4\x00\xCE\xA5\x00\xCE\xA6\x00"
-     "\xCE\xA7"
-     "\x00\xCE\xA8\x00\xCE\xA9\x00\xCE\xAA\x00\xCE\xAB\x00\xCE\xAC\x00\xCE\xAD"
-     "\x00\xCE"
-     "\xAE\x00\xCE\xAF\x00\xCE\xB0\x00\xCE\xB1\x00\xCE\xB2\x00\xCE\xB3\x00\xCE"
-     "\xB4\x00"
-     "\xCE\xB5\x00\xCE\xB6\x00\xCE\xB7\x00\xCE\xB8\x00\xCE\xB9\x00\xCE\xBA\x00"
-     "\xCE\xBB"
-     "\x00\xCE\xBC\x00\xCE\xBC\x41\x00\xCE\xBC\x46\x00\xCE\xBC\x56\x00\xCE\xBC"
-     "\x57\x00"
-     "\xCE\xBC\x67\x00\xCE\xBC\x6C\x00\xCE\xBC\x6D\x00\xCE\xBC\x73\x00\xCE\xBD"
-     "\x00\xCE"
-     "\xBE\x00\xCE\xBF\x00\xCF\x80\x00\xCF\x81\x00\xCF\x82\x00\xCF\x83\x00\xCF"
-     "\x84\x00"
-     "\xCF\x85\x00\xCF\x86\x00\xCF\x87\x00\xCF\x88\x00\xCF\x89\x00\xCF\x8A\x00"
-     "\xCF\x8B"
-     "\x00\xCF\x8C\x00\xCF\x8D\x00\xCF\x8E\x00\xCF\x9C\x00\xCF\x9D\x00\xD0\x80"
-     "\x00\xD0"
-     "\x81\x00\xD0\x83\x00\xD0\x87\x00\xD0\x8C\x00\xD0\x8D\x00\xD0\x8E\x00\xD0"
-     "\x99\x00"
-     "\xD0\xB0\x00\xD0\xB1\x00\xD0\xB2\x00\xD0\xB3\x00\xD0\xB4\x00\xD0\xB5\x00"
-     "\xD0\xB6"
-     "\x00\xD0\xB7\x00\xD0\xB8\x00\xD0\xB9\x00\xD0\xBA\x00\xD0\xBB\x00\xD0\xBC"
-     "\x00\xD0"
-     "\xBD\x00\xD0\xBE\x00\xD0\xBF\x00\xD1\x80\x00\xD1\x81\x00\xD1\x82\x00\xD1"
-     "\x83\x00"
-     "\xD1\x84\x00\xD1\x85\x00\xD1\x86\x00\xD1\x87\x00\xD1\x88\x00\xD1\x8A\x00"
-     "\xD1\x8B"
-     "\x00\xD1\x8C\x00\xD1\x8D\x00\xD1\x8E\x00\xD1\x90\x00\xD1\x91\x00\xD1\x93"
-     "\x00\xD1"
-     "\x95\x00\xD1\x96\x00\xD1\x97\x00\xD1\x98\x00\xD1\x9C\x00\xD1\x9D\x00\xD1"
-     "\x9E\x00"
-     "\xD1\x9F\x00\xD1\xB6\x00\xD1\xB7\x00\xD2\x91\x00\xD2\xAB\x00\xD2\xAF\x00"
-     "\xD2\xB1"
-     "\x00\xD3\x81\x00\xD3\x82\x00\xD3\x8F\x00\xD3\x90\x00\xD3\x91\x00\xD3\x92"
-     "\x00\xD3"
-     "\x93\x00\xD3\x96\x00\xD3\x97\x00\xD3\x99\x00\xD3\x9A\x00\xD3\x9B\x00\xD3"
-     "\x9C\x00"
-     "\xD3\x9D\x00\xD3\x9E\x00\xD3\x9F\x00\xD3\xA2\x00\xD3\xA3\x00\xD3\xA4\x00"
-     "\xD3\xA5"
-     "\x00\xD3\xA6\x00\xD3\xA7\x00\xD3\xA9\x00\xD3\xAA\x00\xD3\xAB\x00\xD3\xAC"
-     "\x00\xD3"
-     "\xAD\x00\xD3\xAE\x00\xD3\xAF\x00\xD3\xB0\x00\xD3\xB1\x00\xD3\xB2\x00\xD3"
-     "\xB3\x00"
-     "\xD3\xB4\x00\xD3\xB5\x00\xD3\xB8\x00\xD3\xB9\x00\xD5\xA5\xD6\x82\x00\xD5"
-     "\xB4\xD5"
-     "\xA5\x00\xD5\xB4\xD5\xAB\x00\xD5\xB4\xD5\xAD\x00\xD5\xB4\xD5\xB6\x00\xD5"
-     "\xBE\xD5"
-     "\xB6\x00\xD7\x90\x00\xD7\x90\xD6\xB7\x00\xD7\x90\xD6\xB8\x00\xD7\x90\xD6"
-     "\xBC\x00"
-     "\xD7\x90\xD7\x9C\x00\xD7\x91\x00\xD7\x91\xD6\xBC\x00\xD7\x91\xD6\xBF\x00"
-     "\xD7\x92"
-     "\x00\xD7\x92\xD6\xBC\x00\xD7\x93\x00\xD7\x93\xD6\xBC\x00\xD7\x94\x00\xD7"
-     "\x94\xD6"
-     "\xBC\x00\xD7\x95\xD6\xB9\x00\xD7\x95\xD6\xBC\x00\xD7\x96\xD6\xBC\x00\xD7"
-     "\x98\xD6"
-     "\xBC\x00\xD7\x99\xD6\xB4\x00\xD7\x99\xD6\xBC\x00\xD7\x9A\xD6\xBC\x00\xD7"
-     "\x9B\x00"
-     "\xD7\x9B\xD6\xBC\x00\xD7\x9B\xD6\xBF\x00\xD7\x9C\x00\xD7\x9C\xD6\xBC\x00"
-     "\xD7\x9D"
-     "\x00\xD7\x9E\xD6\xBC\x00\xD7\xA0\xD6\xBC\x00\xD7\xA1\xD6\xBC\x00\xD7\xA2"
-     "\x00\xD7"
-     "\xA3\xD6\xBC\x00\xD7\xA4\xD6\xBC\x00\xD7\xA4\xD6\xBF\x00\xD7\xA6\xD6\xBC"
-     "\x00\xD7"
-     "\xA7\xD6\xBC\x00\xD7\xA8\x00\xD7\xA8\xD6\xBC\x00\xD7\xA9\xD6\xBC\x00\xD7"
-     "\xA9\xD6"
-     "\xBC\xD7\x81\x00\xD7\xA9\xD6\xBC\xD7\x82\x00\xD7\xA9\xD7\x81\x00\xD7\xA9"
-     "\xD7\x82"
-     "\x00\xD7\xAA\x00\xD7\xAA\xD6\xBC\x00\xD7\xB2\xD6\xB7\x00\xD8\xA1\x00\xD8"
-     "\xA2\x00"
-     "\xD8\xA3\x00\xD8\xA4\x00\xD8\xA5\x00\xD8\xA6\x00\xD8\xA6\xD8\xA7\x00\xD8"
-     "\xA6\xD8"
-     "\xAC\x00\xD8\xA6\xD8\xAD\x00\xD8\xA6\xD8\xAE\x00\xD8\xA6\xD8\xB1\x00\xD8"
-     "\xA6\xD8"
-     "\xB2\x00\xD8\xA6\xD9\x85\x00\xD8\xA6\xD9\x86\x00\xD8\xA6\xD9\x87\x00\xD8"
-     "\xA6\xD9"
-     "\x88\x00\xD8\xA6\xD9\x89\x00\xD8\xA6\xD9\x8A\x00\xD8\xA6\xDB\x86\x00\xD8"
-     "\xA6\xDB"
-     "\x87\x00\xD8\xA6\xDB\x88\x00\xD8\xA6\xDB\x90\x00\xD8\xA6\xDB\x95\x00\xD8"
-     "\xA7\x00"
-     "\xD8\xA7\xD9\x83\xD8\xA8\xD8\xB1\x00\xD8\xA7\xD9\x84\xD9\x84\xD9\x87\x00"
-     "\xD8\xA7"
-     "\xD9\x8B\x00\xD8\xA7\xD9\xB4\x00\xD8\xA8\x00\xD8\xA8\xD8\xAC\x00\xD8\xA8"
-     "\xD8\xAD"
-     "\x00\xD8\xA8\xD8\xAD\xD9\x8A\x00\xD8\xA8\xD8\xAE\x00\xD8\xA8\xD8\xAE\xD9"
-     "\x8A\x00"
-     "\xD8\xA8\xD8\xB1\x00\xD8\xA8\xD8\xB2\x00\xD8\xA8\xD9\x85\x00\xD8\xA8\xD9"
-     "\x86\x00"
-     "\xD8\xA8\xD9\x87\x00\xD8\xA8\xD9\x89\x00\xD8\xA8\xD9\x8A\x00\xD8\xA9\x00"
-     "\xD8\xAA"
-     "\x00\xD8\xAA\xD8\xAC\x00\xD8\xAA\xD8\xAC\xD9\x85\x00\xD8\xAA\xD8\xAC\xD9"
-     "\x89\x00"
-     "\xD8\xAA\xD8\xAC\xD9\x8A\x00\xD8\xAA\xD8\xAD\x00\xD8\xAA\xD8\xAD\xD8\xAC"
-     "\x00\xD8"
-     "\xAA\xD8\xAD\xD9\x85\x00\xD8\xAA\xD8\xAE\x00\xD8\xAA\xD8\xAE\xD9\x85\x00"
-     "\xD8\xAA"
-     "\xD8\xAE\xD9\x89\x00\xD8\xAA\xD8\xAE\xD9\x8A\x00\xD8\xAA\xD8\xB1\x00\xD8"
-     "\xAA\xD8"
-     "\xB2\x00\xD8\xAA\xD9\x85\x00\xD8\xAA\xD9\x85\xD8\xAC\x00\xD8\xAA\xD9\x85"
-     "\xD8\xAD"
-     "\x00\xD8\xAA\xD9\x85\xD8\xAE\x00\xD8\xAA\xD9\x85\xD9\x89\x00\xD8\xAA\xD9"
-     "\x85\xD9"
-     "\x8A\x00\xD8\xAA\xD9\x86\x00\xD8\xAA\xD9\x87\x00\xD8\xAA\xD9\x89\x00\xD8"
-     "\xAA\xD9"
-     "\x8A\x00\xD8\xAB\x00\xD8\xAB\xD8\xAC\x00\xD8\xAB\xD8\xB1\x00\xD8\xAB\xD8"
-     "\xB2\x00"
-     "\xD8\xAB\xD9\x85\x00\xD8\xAB\xD9\x86\x00\xD8\xAB\xD9\x87\x00\xD8\xAB\xD9"
-     "\x89\x00"
-     "\xD8\xAB\xD9\x8A\x00\xD8\xAC\x00\xD8\xAC\xD8\xAD\x00\xD8\xAC\xD8\xAD\xD9"
-     "\x89\x00"
-     "\xD8\xAC\xD8\xAD\xD9\x8A\x00\xD8\xAC\xD9\x84\x20\xD8\xAC\xD9\x84\xD8\xA7"
-     "\xD9\x84"
-     "\xD9\x87\x00\xD8\xAC\xD9\x85\x00\xD8\xAC\xD9\x85\xD8\xAD\x00\xD8\xAC\xD9"
-     "\x85\xD9"
-     "\x89\x00\xD8\xAC\xD9\x85\xD9\x8A\x00\xD8\xAC\xD9\x89\x00\xD8\xAC\xD9\x8A"
-     "\x00\xD8"
-     "\xAD\x00\xD8\xAD\xD8\xAC\x00\xD8\xAD\xD8\xAC\xD9\x8A\x00\xD8\xAD\xD9\x85"
-     "\x00\xD8"
-     "\xAD\xD9\x85\xD9\x89\x00\xD8\xAD\xD9\x85\xD9\x8A\x00\xD8\xAD\xD9\x89\x00"
-     "\xD8\xAD"
-     "\xD9\x8A\x00\xD8\xAE\x00\xD8\xAE\xD8\xAC\x00\xD8\xAE\xD8\xAD\x00\xD8\xAE"
-     "\xD9\x85"
-     "\x00\xD8\xAE\xD9\x89\x00\xD8\xAE\xD9\x8A\x00\xD8\xAF\x00\xD8\xB0\x00\xD8"
-     "\xB0\xD9"
-     "\xB0\x00\xD8\xB1\x00\xD8\xB1\xD8\xB3\xD9\x88\xD9\x84\x00\xD8\xB1\xD9\xB0"
-     "\x00\xD8"
-     "\xB1\xDB\x8C\xD8\xA7\xD9\x84\x00\xD8\xB2\x00\xD8\xB3\x00\xD8\xB3\xD8\xAC"
-     "\x00\xD8"
-     "\xB3\xD8\xAC\xD8\xAD\x00\xD8\xB3\xD8\xAC\xD9\x89\x00\xD8\xB3\xD8\xAD\x00"
-     "\xD8\xB3"
-     "\xD8\xAD\xD8\xAC\x00\xD8\xB3\xD8\xAE\x00\xD8\xB3\xD8\xAE\xD9\x89\x00\xD8"
-     "\xB3\xD8"
-     "\xAE\xD9\x8A\x00\xD8\xB3\xD8\xB1\x00\xD8\xB3\xD9\x85\x00\xD8\xB3\xD9\x85"
-     "\xD8\xAC"
-     "\x00\xD8\xB3\xD9\x85\xD8\xAD\x00\xD8\xB3\xD9\x85\xD9\x85\x00\xD8\xB3\xD9"
-     "\x87\x00"
-     "\xD8\xB3\xD9\x89\x00\xD8\xB3\xD9\x8A\x00\xD8\xB4\x00\xD8\xB4\xD8\xAC\x00"
-     "\xD8\xB4"
-     "\xD8\xAC\xD9\x8A\x00\xD8\xB4\xD8\xAD\x00\xD8\xB4\xD8\xAD\xD9\x85\x00\xD8"
-     "\xB4\xD8"
-     "\xAD\xD9\x8A\x00\xD8\xB4\xD8\xAE\x00\xD8\xB4\xD8\xB1\x00\xD8\xB4\xD9\x85"
-     "\x00\xD8"
-     "\xB4\xD9\x85\xD8\xAE\x00\xD8\xB4\xD9\x85\xD9\x85\x00\xD8\xB4\xD9\x87\x00"
-     "\xD8\xB4"
-     "\xD9\x89\x00\xD8\xB4\xD9\x8A\x00\xD8\xB5\x00\xD8\xB5\xD8\xAD\x00\xD8\xB5"
-     "\xD8\xAD"
-     "\xD8\xAD\x00\xD8\xB5\xD8\xAD\xD9\x8A\x00\xD8\xB5\xD8\xAE\x00\xD8\xB5\xD8"
-     "\xB1\x00"
-     "\xD8\xB5\xD9\x84\xD8\xB9\xD9\x85\x00\xD8\xB5\xD9\x84\xD9\x89\x00\xD8\xB5"
-     "\xD9\x84"
-     "\xD9\x89\x20\xD8\xA7\xD9\x84\xD9\x84\xD9\x87\x20\xD8\xB9\xD9\x84\xD9\x8A"
-     "\xD9\x87"
-     "\x20\xD9\x88\xD8\xB3\xD9\x84\xD9\x85\x00\xD8\xB5\xD9\x84\xDB\x92\x00\xD8"
-     "\xB5\xD9"
-     "\x85\x00\xD8\xB5\xD9\x85\xD9\x85\x00\xD8\xB5\xD9\x89\x00\xD8\xB5\xD9\x8A"
-     "\x00\xD8"
-     "\xB6\x00\xD8\xB6\xD8\xAC\x00\xD8\xB6\xD8\xAD\x00\xD8\xB6\xD8\xAD\xD9\x89"
-     "\x00\xD8"
-     "\xB6\xD8\xAD\xD9\x8A\x00\xD8\xB6\xD8\xAE\x00\xD8\xB6\xD8\xAE\xD9\x85\x00"
-     "\xD8\xB6"
-     "\xD8\xB1\x00\xD8\xB6\xD9\x85\x00\xD8\xB6\xD9\x89\x00\xD8\xB6\xD9\x8A\x00"
-     "\xD8\xB7"
-     "\x00\xD8\xB7\xD8\xAD\x00\xD8\xB7\xD9\x85\x00\xD8\xB7\xD9\x85\xD8\xAD\x00"
-     "\xD8\xB7"
-     "\xD9\x85\xD9\x85\x00\xD8\xB7\xD9\x85\xD9\x8A\x00\xD8\xB7\xD9\x89\x00\xD8"
-     "\xB7\xD9"
-     "\x8A\x00\xD8\xB8\x00\xD8\xB8\xD9\x85\x00\xD8\xB9\x00\xD8\xB9\xD8\xAC\x00"
-     "\xD8\xB9"
-     "\xD8\xAC\xD9\x85\x00\xD8\xB9\xD9\x84\xD9\x8A\xD9\x87\x00\xD8\xB9\xD9\x85"
-     "\x00\xD8"
-     "\xB9\xD9\x85\xD9\x85\x00\xD8\xB9\xD9\x85\xD9\x89\x00\xD8\xB9\xD9\x85\xD9"
-     "\x8A\x00"
-     "\xD8\xB9\xD9\x89\x00\xD8\xB9\xD9\x8A\x00\xD8\xBA\x00\xD8\xBA\xD8\xAC\x00"
-     "\xD8\xBA"
-     "\xD9\x85\x00\xD8\xBA\xD9\x85\xD9\x85\x00\xD8\xBA\xD9\x85\xD9\x89\x00\xD8"
-     "\xBA\xD9"
-     "\x85\xD9\x8A\x00\xD8\xBA\xD9\x89\x00\xD8\xBA\xD9\x8A\x00\xD9\x80\xD9\x8B"
-     "\x00\xD9"
-     "\x80\xD9\x8E\x00\xD9\x80\xD9\x8E\xD9\x91\x00\xD9\x80\xD9\x8F\x00\xD9\x80"
-     "\xD9\x8F"
-     "\xD9\x91\x00\xD9\x80\xD9\x90\x00\xD9\x80\xD9\x90\xD9\x91\x00\xD9\x80\xD9"
-     "\x91\x00"
-     "\xD9\x80\xD9\x92\x00\xD9\x81\x00\xD9\x81\xD8\xAC\x00\xD9\x81\xD8\xAD\x00"
-     "\xD9\x81"
-     "\xD8\xAE\x00\xD9\x81\xD8\xAE\xD9\x85\x00\xD9\x81\xD9\x85\x00\xD9\x81\xD9"
-     "\x85\xD9"
-     "\x8A\x00\xD9\x81\xD9\x89\x00\xD9\x81\xD9\x8A\x00\xD9\x82\x00\xD9\x82\xD8"
-     "\xAD\x00"
-     "\xD9\x82\xD9\x84\xDB\x92\x00\xD9\x82\xD9\x85\x00\xD9\x82\xD9\x85\xD8\xAD"
-     "\x00\xD9"
-     "\x82\xD9\x85\xD9\x85\x00\xD9\x82\xD9\x85\xD9\x8A\x00\xD9\x82\xD9\x89\x00"
-     "\xD9\x82"
-     "\xD9\x8A\x00\xD9\x83\x00\xD9\x83\xD8\xA7\x00\xD9\x83\xD8\xAC\x00\xD9\x83"
-     "\xD8\xAD"
-     "\x00\xD9\x83\xD8\xAE\x00\xD9\x83\xD9\x84\x00\xD9\x83\xD9\x85\x00\xD9\x83"
-     "\xD9\x85"
-     "\xD9\x85\x00\xD9\x83\xD9\x85\xD9\x8A\x00\xD9\x83\xD9\x89\x00\xD9\x83\xD9"
-     "\x8A\x00"
-     "\xD9\x84\x00\xD9\x84\xD8\xA2\x00\xD9\x84\xD8\xA3\x00\xD9\x84\xD8\xA5\x00"
-     "\xD9\x84"
-     "\xD8\xA7\x00\xD9\x84\xD8\xAC\x00\xD9\x84\xD8\xAC\xD8\xAC\x00\xD9\x84\xD8"
-     "\xAC\xD9"
-     "\x85\x00\xD9\x84\xD8\xAC\xD9\x8A\x00\xD9\x84\xD8\xAD\x00\xD9\x84\xD8\xAD"
-     "\xD9\x85"
-     "\x00\xD9\x84\xD8\xAD\xD9\x89\x00\xD9\x84\xD8\xAD\xD9\x8A\x00\xD9\x84\xD8"
-     "\xAE\x00"
-     "\xD9\x84\xD8\xAE\xD9\x85\x00\xD9\x84\xD9\x85\x00\xD9\x84\xD9\x85\xD8\xAD"
-     "\x00\xD9"
-     "\x84\xD9\x85\xD9\x8A\x00\xD9\x84\xD9\x87\x00\xD9\x84\xD9\x89\x00\xD9\x84"
-     "\xD9\x8A"
-     "\x00\xD9\x85\x00\xD9\x85\xD8\xA7\x00\xD9\x85\xD8\xAC\x00\xD9\x85\xD8\xAC"
-     "\xD8\xAD"
-     "\x00\xD9\x85\xD8\xAC\xD8\xAE\x00\xD9\x85\xD8\xAC\xD9\x85\x00\xD9\x85\xD8"
-     "\xAC\xD9"
-     "\x8A\x00\xD9\x85\xD8\xAD\x00\xD9\x85\xD8\xAD\xD8\xAC\x00\xD9\x85\xD8\xAD"
-     "\xD9\x85"
-     "\x00\xD9\x85\xD8\xAD\xD9\x85\xD8\xAF\x00\xD9\x85\xD8\xAD\xD9\x8A\x00\xD9"
-     "\x85\xD8"
-     "\xAE\x00\xD9\x85\xD8\xAE\xD8\xAC\x00\xD9\x85\xD8\xAE\xD9\x85\x00\xD9\x85"
-     "\xD8\xAE"
-     "\xD9\x8A\x00\xD9\x85\xD9\x85\x00\xD9\x85\xD9\x85\xD9\x8A\x00\xD9\x85\xD9"
-     "\x89\x00"
-     "\xD9\x85\xD9\x8A\x00\xD9\x86\x00\xD9\x86\xD8\xAC\x00\xD9\x86\xD8\xAC\xD8"
-     "\xAD\x00"
-     "\xD9\x86\xD8\xAC\xD9\x85\x00\xD9\x86\xD8\xAC\xD9\x89\x00\xD9\x86\xD8\xAC"
-     "\xD9\x8A"
-     "\x00\xD9\x86\xD8\xAD\x00\xD9\x86\xD8\xAD\xD9\x85\x00\xD9\x86\xD8\xAD\xD9"
-     "\x89\x00"
-     "\xD9\x86\xD8\xAD\xD9\x8A\x00\xD9\x86\xD8\xAE\x00\xD9\x86\xD8\xB1\x00\xD9"
-     "\x86\xD8"
-     "\xB2\x00\xD9\x86\xD9\x85\x00\xD9\x86\xD9\x85\xD9\x89\x00\xD9\x86\xD9\x85"
-     "\xD9\x8A"
-     "\x00\xD9\x86\xD9\x86\x00\xD9\x86\xD9\x87\x00\xD9\x86\xD9\x89\x00\xD9\x86"
-     "\xD9\x8A"
-     "\x00\xD9\x87\x00\xD9\x87\xD8\xAC\x00\xD9\x87\xD9\x85\x00\xD9\x87\xD9\x85"
-     "\xD8\xAC"
-     "\x00\xD9\x87\xD9\x85\xD9\x85\x00\xD9\x87\xD9\x89\x00\xD9\x87\xD9\x8A\x00"
-     "\xD9\x87"
-     "\xD9\xB0\x00\xD9\x88\x00\xD9\x88\xD8\xB3\xD9\x84\xD9\x85\x00\xD9\x88\xD9"
-     "\xB4\x00"
-     "\xD9\x89\x00\xD9\x89\xD9\xB0\x00\xD9\x8A\x00\xD9\x8A\xD8\xAC\x00\xD9\x8A"
-     "\xD8\xAC"
-     "\xD9\x8A\x00\xD9\x8A\xD8\xAD\x00\xD9\x8A\xD8\xAD\xD9\x8A\x00\xD9\x8A\xD8"
-     "\xAE\x00"
-     "\xD9\x8A\xD8\xB1\x00\xD9\x8A\xD8\xB2\x00\xD9\x8A\xD9\x85\x00\xD9\x8A\xD9"
-     "\x85\xD9"
-     "\x85\x00\xD9\x8A\xD9\x85\xD9\x8A\x00\xD9\x8A\xD9\x86\x00\xD9\x8A\xD9\x87"
-     "\x00\xD9"
-     "\x8A\xD9\x89\x00\xD9\x8A\xD9\x8A\x00\xD9\x8A\xD9\xB4\x00\xD9\xAE\x00\xD9"
-     "\xAF\x00"
-     "\xD9\xB1\x00\xD9\xB9\x00\xD9\xBA\x00\xD9\xBB\x00\xD9\xBE\x00\xD9\xBF\x00"
-     "\xDA\x80"
-     "\x00\xDA\x83\x00\xDA\x84\x00\xDA\x86\x00\xDA\x87\x00\xDA\x88\x00\xDA\x8C"
-     "\x00\xDA"
-     "\x8D\x00\xDA\x8E\x00\xDA\x91\x00\xDA\x98\x00\xDA\xA1\x00\xDA\xA4\x00\xDA"
-     "\xA6\x00"
-     "\xDA\xA9\x00\xDA\xAD\x00\xDA\xAF\x00\xDA\xB1\x00\xDA\xB3\x00\xDA\xBA\x00"
-     "\xDA\xBB"
-     "\x00\xDA\xBE\x00\xDB\x80\x00\xDB\x81\x00\xDB\x82\x00\xDB\x85\x00\xDB\x86"
-     "\x00\xDB"
-     "\x87\x00\xDB\x87\xD9\xB4\x00\xDB\x88\x00\xDB\x89\x00\xDB\x8B\x00\xDB\x8C"
-     "\x00\xDB"
-     "\x90\x00\xDB\x92\x00\xDB\x93\x00\xE0\xA4\x95\xE0\xA4\xBC\x00\xE0\xA4\x96"
-     "\xE0\xA4"
-     "\xBC\x00\xE0\xA4\x97\xE0\xA4\xBC\x00\xE0\xA4\x9C\xE0\xA4\xBC\x00\xE0\xA4"
-     "\xA1\xE0"
-     "\xA4\xBC\x00\xE0\xA4\xA2\xE0\xA4\xBC\x00\xE0\xA4\xA9\x00\xE0\xA4\xAB\xE0"
-     "\xA4\xBC"
-     "\x00\xE0\xA4\xAF\xE0\xA4\xBC\x00\xE0\xA4\xB1\x00\xE0\xA4\xB4\x00\xE0\xA6"
-     "\xA1\xE0"
-     "\xA6\xBC\x00\xE0\xA6\xA2\xE0\xA6\xBC\x00\xE0\xA6\xAF\xE0\xA6\xBC\x00\xE0"
-     "\xA7\x8B"
-     "\x00\xE0\xA7\x8C\x00\xE0\xA8\x96\xE0\xA8\xBC\x00\xE0\xA8\x97\xE0\xA8\xBC"
-     "\x00\xE0"
-     "\xA8\x9C\xE0\xA8\xBC\x00\xE0\xA8\xAB\xE0\xA8\xBC\x00\xE0\xA8\xB2\xE0\xA8"
-     "\xBC\x00"
-     "\xE0\xA8\xB8\xE0\xA8\xBC\x00\xE0\xAC\xA1\xE0\xAC\xBC\x00\xE0\xAC\xA2\xE0"
-     "\xAC\xBC"
-     "\x00\xE0\xAD\x88\x00\xE0\xAD\x8B\x00\xE0\xAD\x8C\x00\xE0\xAE\x94\x00\xE0"
-     "\xAF\x8A"
-     "\x00\xE0\xAF\x8B\x00\xE0\xAF\x8C\x00\xE0\xB1\x88\x00\xE0\xB3\x80\x00\xE0"
-     "\xB3\x87"
-     "\x00\xE0\xB3\x88\x00\xE0\xB3\x8A\x00\xE0\xB3\x8B\x00\xE0\xB5\x8A\x00\xE0"
-     "\xB5\x8B"
-     "\x00\xE0\xB5\x8C\x00\xE0\xB7\x9A\x00\xE0\xB7\x9C\x00\xE0\xB7\x9D\x00\xE0"
-     "\xB7\x9E"
-     "\x00\xE0\xB9\x8D\xE0\xB8\xB2\x00\xE0\xBA\xAB\xE0\xBA\x99\x00\xE0\xBA\xAB"
-     "\xE0\xBA"
-     "\xA1\x00\xE0\xBB\x8D\xE0\xBA\xB2\x00\xE0\xBC\x8B\x00\xE0\xBD\x80\xE0\xBE"
-     "\xB5\x00"
-     "\xE0\xBD\x82\xE0\xBE\xB7\x00\xE0\xBD\x8C\xE0\xBE\xB7\x00\xE0\xBD\x91\xE0"
-     "\xBE\xB7"
-     "\x00\xE0\xBD\x96\xE0\xBE\xB7\x00\xE0\xBD\x9B\xE0\xBE\xB7\x00\xE0\xBD\xB1"
-     "\xE0\xBD"
-     "\xB2\x00\xE0\xBD\xB1\xE0\xBD\xB4\x00\xE0\xBD\xB1\xE0\xBE\x80\x00\xE0\xBE"
-     "\x90\xE0"
-     "\xBE\xB5\x00\xE0\xBE\x92\xE0\xBE\xB7\x00\xE0\xBE\x9C\xE0\xBE\xB7\x00\xE0"
-     "\xBE\xA1"
-     "\xE0\xBE\xB7\x00\xE0\xBE\xA6\xE0\xBE\xB7\x00\xE0\xBE\xAB\xE0\xBE\xB7\x00"
-     "\xE0\xBE"
-     "\xB2\xE0\xBD\xB1\xE0\xBE\x80\x00\xE0\xBE\xB2\xE0\xBE\x80\x00\xE0\xBE\xB3"
-     "\xE0\xBD"
-     "\xB1\xE0\xBE\x80\x00\xE0\xBE\xB3\xE0\xBE\x80\x00\xE1\x80\xA6\x00\xE1\x83"
-     "\x9C\x00"
-     "\xE1\x84\x80\x00\xE1\x84\x81\x00\xE1\x84\x82\x00\xE1\x84\x83\x00\xE1\x84"
-     "\x84\x00"
-     "\xE1\x84\x85\x00\xE1\x84\x86\x00\xE1\x84\x87\x00\xE1\x84\x88\x00\xE1\x84"
-     "\x89\x00"
-     "\xE1\x84\x8A\x00\xE1\x84\x8B\x00\xE1\x84\x8C\x00\xE1\x84\x8D\x00\xE1\x84"
-     "\x8E\x00"
-     "\xE1\x84\x8F\x00\xE1\x84\x90\x00\xE1\x84\x91\x00\xE1\x84\x92\x00\xE1\x84"
-     "\x94\x00"
-     "\xE1\x84\x95\x00\xE1\x84\x9A\x00\xE1\x84\x9C\x00\xE1\x84\x9D\x00\xE1\x84"
-     "\x9E\x00"
-     "\xE1\x84\xA0\x00\xE1\x84\xA1\x00\xE1\x84\xA2\x00\xE1\x84\xA3\x00\xE1\x84"
-     "\xA7\x00"
-     "\xE1\x84\xA9\x00\xE1\x84\xAB\x00\xE1\x84\xAC\x00\xE1\x84\xAD\x00\xE1\x84"
-     "\xAE\x00"
-     "\xE1\x84\xAF\x00\xE1\x84\xB2\x00\xE1\x84\xB6\x00\xE1\x85\x80\x00\xE1\x85"
-     "\x87\x00"
-     "\xE1\x85\x8C\x00\xE1\x85\x97\x00\xE1\x85\x98\x00\xE1\x85\x99\x00\xE1\x85"
-     "\xA0\x00"
-     "\xE1\x85\xA1\x00\xE1\x85\xA2\x00\xE1\x85\xA3\x00\xE1\x85\xA4\x00\xE1\x85"
-     "\xA5\x00"
-     "\xE1\x85\xA6\x00\xE1\x85\xA7\x00\xE1\x85\xA8\x00\xE1\x85\xA9\x00\xE1\x85"
-     "\xAA\x00"
-     "\xE1\x85\xAB\x00\xE1\x85\xAC\x00\xE1\x85\xAD\x00\xE1\x85\xAE\x00\xE1\x85"
-     "\xAF\x00"
-     "\xE1\x85\xB0\x00\xE1\x85\xB1\x00\xE1\x85\xB2\x00\xE1\x85\xB3\x00\xE1\x85"
-     "\xB4\x00"
-     "\xE1\x85\xB5\x00\xE1\x86\x84\x00\xE1\x86\x85\x00\xE1\x86\x88\x00\xE1\x86"
-     "\x91\x00"
-     "\xE1\x86\x92\x00\xE1\x86\x94\x00\xE1\x86\x9E\x00\xE1\x86\xA1\x00\xE1\x86"
-     "\xAA\x00"
-     "\xE1\x86\xAC\x00\xE1\x86\xAD\x00\xE1\x86\xB0\x00\xE1\x86\xB1\x00\xE1\x86"
-     "\xB2\x00"
-     "\xE1\x86\xB3\x00\xE1\x86\xB4\x00\xE1\x86\xB5\x00\xE1\x87\x87\x00\xE1\x87"
-     "\x88\x00"
-     "\xE1\x87\x8C\x00\xE1\x87\x8E\x00\xE1\x87\x93\x00\xE1\x87\x97\x00\xE1\x87"
-     "\x99\x00"
-     "\xE1\x87\x9D\x00\xE1\x87\x9F\x00\xE1\x87\xB1\x00\xE1\x87\xB2\x00\xE1\xAC"
-     "\x86\x00"
-     "\xE1\xAC\x88\x00\xE1\xAC\x8A\x00\xE1\xAC\x8C\x00\xE1\xAC\x8E\x00\xE1\xAC"
-     "\x92\x00"
-     "\xE1\xAC\xBB\x00\xE1\xAC\xBD\x00\xE1\xAD\x80\x00\xE1\xAD\x81\x00\xE1\xAD"
-     "\x83\x00"
-     "\xE1\xB4\x82\x00\xE1\xB4\x96\x00\xE1\xB4\x97\x00\xE1\xB4\x9C\x00\xE1\xB4"
-     "\x9D\x00"
-     "\xE1\xB4\xA5\x00\xE1\xB5\xBB\x00\xE1\xB6\x85\x00\xE1\xB6\x91\x00\xE1\xB8"
-     "\x80\x00"
-     "\xE1\xB8\x81\x00\xE1\xB8\x82\x00\xE1\xB8\x83\x00\xE1\xB8\x84\x00\xE1\xB8"
-     "\x85\x00"
-     "\xE1\xB8\x86\x00\xE1\xB8\x87\x00\xE1\xB8\x88\x00\xE1\xB8\x89\x00\xE1\xB8"
-     "\x8A\x00"
-     "\xE1\xB8\x8B\x00\xE1\xB8\x8C\x00\xE1\xB8\x8D\x00\xE1\xB8\x8E\x00\xE1\xB8"
-     "\x8F\x00"
-     "\xE1\xB8\x90\x00\xE1\xB8\x91\x00\xE1\xB8\x92\x00\xE1\xB8\x93\x00\xE1\xB8"
-     "\x94\x00"
-     "\xE1\xB8\x95\x00\xE1\xB8\x96\x00\xE1\xB8\x97\x00\xE1\xB8\x98\x00\xE1\xB8"
-     "\x99\x00"
-     "\xE1\xB8\x9A\x00\xE1\xB8\x9B\x00\xE1\xB8\x9C\x00\xE1\xB8\x9D\x00\xE1\xB8"
-     "\x9E\x00"
-     "\xE1\xB8\x9F\x00\xE1\xB8\xA0\x00\xE1\xB8\xA1\x00\xE1\xB8\xA2\x00\xE1\xB8"
-     "\xA3\x00"
-     "\xE1\xB8\xA4\x00\xE1\xB8\xA5\x00\xE1\xB8\xA6\x00\xE1\xB8\xA7\x00\xE1\xB8"
-     "\xA8\x00"
-     "\xE1\xB8\xA9\x00\xE1\xB8\xAA\x00\xE1\xB8\xAB\x00\xE1\xB8\xAC\x00\xE1\xB8"
-     "\xAD\x00"
-     "\xE1\xB8\xAE\x00\xE1\xB8\xAF\x00\xE1\xB8\xB0\x00\xE1\xB8\xB1\x00\xE1\xB8"
-     "\xB2\x00"
-     "\xE1\xB8\xB3\x00\xE1\xB8\xB4\x00\xE1\xB8\xB5\x00\xE1\xB8\xB6\x00\xE1\xB8"
-     "\xB7\x00"
-     "\xE1\xB8\xB8\x00\xE1\xB8\xB9\x00\xE1\xB8\xBA\x00\xE1\xB8\xBB\x00\xE1\xB8"
-     "\xBC\x00"
-     "\xE1\xB8\xBD\x00\xE1\xB8\xBE\x00\xE1\xB8\xBF\x00\xE1\xB9\x80\x00\xE1\xB9"
-     "\x81\x00"
-     "\xE1\xB9\x82\x00\xE1\xB9\x83\x00\xE1\xB9\x84\x00\xE1\xB9\x85\x00\xE1\xB9"
-     "\x86\x00"
-     "\xE1\xB9\x87\x00\xE1\xB9\x88\x00\xE1\xB9\x89\x00\xE1\xB9\x8A\x00\xE1\xB9"
-     "\x8B\x00"
-     "\xE1\xB9\x8C\x00\xE1\xB9\x8D\x00\xE1\xB9\x8E\x00\xE1\xB9\x8F\x00\xE1\xB9"
-     "\x90\x00"
-     "\xE1\xB9\x91\x00\xE1\xB9\x92\x00\xE1\xB9\x93\x00\xE1\xB9\x94\x00\xE1\xB9"
-     "\x95\x00"
-     "\xE1\xB9\x96\x00\xE1\xB9\x97\x00\xE1\xB9\x98\x00\xE1\xB9\x99\x00\xE1\xB9"
-     "\x9A\x00"
-     "\xE1\xB9\x9B\x00\xE1\xB9\x9C\x00\xE1\xB9\x9D\x00\xE1\xB9\x9E\x00\xE1\xB9"
-     "\x9F\x00"
-     "\xE1\xB9\xA0\x00\xE1\xB9\xA1\x00\xE1\xB9\xA2\x00\xE1\xB9\xA3\x00\xE1\xB9"
-     "\xA4\x00"
-     "\xE1\xB9\xA5\x00\xE1\xB9\xA6\x00\xE1\xB9\xA7\x00\xE1\xB9\xA8\x00\xE1\xB9"
-     "\xA9\x00"
-     "\xE1\xB9\xAA\x00\xE1\xB9\xAB\x00\xE1\xB9\xAC\x00\xE1\xB9\xAD\x00\xE1\xB9"
-     "\xAE\x00"
-     "\xE1\xB9\xAF\x00\xE1\xB9\xB0\x00\xE1\xB9\xB1\x00\xE1\xB9\xB2\x00\xE1\xB9"
-     "\xB3\x00"
-     "\xE1\xB9\xB4\x00\xE1\xB9\xB5\x00\xE1\xB9\xB6\x00\xE1\xB9\xB7\x00\xE1\xB9"
-     "\xB8\x00"
-     "\xE1\xB9\xB9\x00\xE1\xB9\xBA\x00\xE1\xB9\xBB\x00\xE1\xB9\xBC\x00\xE1\xB9"
-     "\xBD\x00"
-     "\xE1\xB9\xBE\x00\xE1\xB9\xBF\x00\xE1\xBA\x80\x00\xE1\xBA\x81\x00\xE1\xBA"
-     "\x82\x00"
-     "\xE1\xBA\x83\x00\xE1\xBA\x84\x00\xE1\xBA\x85\x00\xE1\xBA\x86\x00\xE1\xBA"
-     "\x87\x00"
-     "\xE1\xBA\x88\x00\xE1\xBA\x89\x00\xE1\xBA\x8A\x00\xE1\xBA\x8B\x00\xE1\xBA"
-     "\x8C\x00"
-     "\xE1\xBA\x8D\x00\xE1\xBA\x8E\x00\xE1\xBA\x8F\x00\xE1\xBA\x90\x00\xE1\xBA"
-     "\x91\x00"
-     "\xE1\xBA\x92\x00\xE1\xBA\x93\x00\xE1\xBA\x94\x00\xE1\xBA\x95\x00\xE1\xBA"
-     "\x96\x00"
-     "\xE1\xBA\x97\x00\xE1\xBA\x98\x00\xE1\xBA\x99\x00\xE1\xBA\xA0\x00\xE1\xBA"
-     "\xA1\x00"
-     "\xE1\xBA\xA2\x00\xE1\xBA\xA3\x00\xE1\xBA\xA4\x00\xE1\xBA\xA5\x00\xE1\xBA"
-     "\xA6\x00"
-     "\xE1\xBA\xA7\x00\xE1\xBA\xA8\x00\xE1\xBA\xA9\x00\xE1\xBA\xAA\x00\xE1\xBA"
-     "\xAB\x00"
-     "\xE1\xBA\xAC\x00\xE1\xBA\xAD\x00\xE1\xBA\xAE\x00\xE1\xBA\xAF\x00\xE1\xBA"
-     "\xB0\x00"
-     "\xE1\xBA\xB1\x00\xE1\xBA\xB2\x00\xE1\xBA\xB3\x00\xE1\xBA\xB4\x00\xE1\xBA"
-     "\xB5\x00"
-     "\xE1\xBA\xB6\x00\xE1\xBA\xB7\x00\xE1\xBA\xB8\x00\xE1\xBA\xB9\x00\xE1\xBA"
-     "\xBA\x00"
-     "\xE1\xBA\xBB\x00\xE1\xBA\xBC\x00\xE1\xBA\xBD\x00\xE1\xBA\xBE\x00\xE1\xBA"
-     "\xBF\x00"
-     "\xE1\xBB\x80\x00\xE1\xBB\x81\x00\xE1\xBB\x82\x00\xE1\xBB\x83\x00\xE1\xBB"
-     "\x84\x00"
-     "\xE1\xBB\x85\x00\xE1\xBB\x86\x00\xE1\xBB\x87\x00\xE1\xBB\x88\x00\xE1\xBB"
-     "\x89\x00"
-     "\xE1\xBB\x8A\x00\xE1\xBB\x8B\x00\xE1\xBB\x8C\x00\xE1\xBB\x8D\x00\xE1\xBB"
-     "\x8E\x00"
-     "\xE1\xBB\x8F\x00\xE1\xBB\x90\x00\xE1\xBB\x91\x00\xE1\xBB\x92\x00\xE1\xBB"
-     "\x93\x00"
-     "\xE1\xBB\x94\x00\xE1\xBB\x95\x00\xE1\xBB\x96\x00\xE1\xBB\x97\x00\xE1\xBB"
-     "\x98\x00"
-     "\xE1\xBB\x99\x00\xE1\xBB\x9A\x00\xE1\xBB\x9B\x00\xE1\xBB\x9C\x00\xE1\xBB"
-     "\x9D\x00"
-     "\xE1\xBB\x9E\x00\xE1\xBB\x9F\x00\xE1\xBB\xA0\x00\xE1\xBB\xA1\x00\xE1\xBB"
-     "\xA2\x00"
-     "\xE1\xBB\xA3\x00\xE1\xBB\xA4\x00\xE1\xBB\xA5\x00\xE1\xBB\xA6\x00\xE1\xBB"
-     "\xA7\x00"
-     "\xE1\xBB\xA8\x00\xE1\xBB\xA9\x00\xE1\xBB\xAA\x00\xE1\xBB\xAB\x00\xE1\xBB"
-     "\xAC\x00"
-     "\xE1\xBB\xAD\x00\xE1\xBB\xAE\x00\xE1\xBB\xAF\x00\xE1\xBB\xB0\x00\xE1\xBB"
-     "\xB1\x00"
-     "\xE1\xBB\xB2\x00\xE1\xBB\xB3\x00\xE1\xBB\xB4\x00\xE1\xBB\xB5\x00\xE1\xBB"
-     "\xB6\x00"
-     "\xE1\xBB\xB7\x00\xE1\xBB\xB8\x00\xE1\xBB\xB9\x00\xE1\xBC\x80\x00\xE1\xBC"
-     "\x81\x00"
-     "\xE1\xBC\x82\x00\xE1\xBC\x83\x00\xE1\xBC\x84\x00\xE1\xBC\x85\x00\xE1\xBC"
-     "\x86\x00"
-     "\xE1\xBC\x87\x00\xE1\xBC\x88\x00\xE1\xBC\x89\x00\xE1\xBC\x8A\x00\xE1\xBC"
-     "\x8B\x00"
-     "\xE1\xBC\x8C\x00\xE1\xBC\x8D\x00\xE1\xBC\x8E\x00\xE1\xBC\x8F\x00\xE1\xBC"
-     "\x90\x00"
-     "\xE1\xBC\x91\x00\xE1\xBC\x92\x00\xE1\xBC\x93\x00\xE1\xBC\x94\x00\xE1\xBC"
-     "\x95\x00"
-     "\xE1\xBC\x98\x00\xE1\xBC\x99\x00\xE1\xBC\x9A\x00\xE1\xBC\x9B\x00\xE1\xBC"
-     "\x9C\x00"
-     "\xE1\xBC\x9D\x00\xE1\xBC\xA0\x00\xE1\xBC\xA1\x00\xE1\xBC\xA2\x00\xE1\xBC"
-     "\xA3\x00"
-     "\xE1\xBC\xA4\x00\xE1\xBC\xA5\x00\xE1\xBC\xA6\x00\xE1\xBC\xA7\x00\xE1\xBC"
-     "\xA8\x00"
-     "\xE1\xBC\xA9\x00\xE1\xBC\xAA\x00\xE1\xBC\xAB\x00\xE1\xBC\xAC\x00\xE1\xBC"
-     "\xAD\x00"
-     "\xE1\xBC\xAE\x00\xE1\xBC\xAF\x00\xE1\xBC\xB0\x00\xE1\xBC\xB1\x00\xE1\xBC"
-     "\xB2\x00"
-     "\xE1\xBC\xB3\x00\xE1\xBC\xB4\x00\xE1\xBC\xB5\x00\xE1\xBC\xB6\x00\xE1\xBC"
-     "\xB7\x00"
-     "\xE1\xBC\xB8\x00\xE1\xBC\xB9\x00\xE1\xBC\xBA\x00\xE1\xBC\xBB\x00\xE1\xBC"
-     "\xBC\x00"
-     "\xE1\xBC\xBD\x00\xE1\xBC\xBE\x00\xE1\xBC\xBF\x00\xE1\xBD\x80\x00\xE1\xBD"
-     "\x81\x00"
-     "\xE1\xBD\x82\x00\xE1\xBD\x83\x00\xE1\xBD\x84\x00\xE1\xBD\x85\x00\xE1\xBD"
-     "\x88\x00"
-     "\xE1\xBD\x89\x00\xE1\xBD\x8A\x00\xE1\xBD\x8B\x00\xE1\xBD\x8C\x00\xE1\xBD"
-     "\x8D\x00"
-     "\xE1\xBD\x90\x00\xE1\xBD\x91\x00\xE1\xBD\x92\x00\xE1\xBD\x93\x00\xE1\xBD"
-     "\x94\x00"
-     "\xE1\xBD\x95\x00\xE1\xBD\x96\x00\xE1\xBD\x97\x00\xE1\xBD\x99\x00\xE1\xBD"
-     "\x9B\x00"
-     "\xE1\xBD\x9D\x00\xE1\xBD\x9F\x00\xE1\xBD\xA0\x00\xE1\xBD\xA1\x00\xE1\xBD"
-     "\xA2\x00"
-     "\xE1\xBD\xA3\x00\xE1\xBD\xA4\x00\xE1\xBD\xA5\x00\xE1\xBD\xA6\x00\xE1\xBD"
-     "\xA7\x00"
-     "\xE1\xBD\xA8\x00\xE1\xBD\xA9\x00\xE1\xBD\xAA\x00\xE1\xBD\xAB\x00\xE1\xBD"
-     "\xAC\x00"
-     "\xE1\xBD\xAD\x00\xE1\xBD\xAE\x00\xE1\xBD\xAF\x00\xE1\xBD\xB0\x00\xE1\xBD"
-     "\xB2\x00"
-     "\xE1\xBD\xB4\x00\xE1\xBD\xB6\x00\xE1\xBD\xB8\x00\xE1\xBD\xBA\x00\xE1\xBD"
-     "\xBC\x00"
-     "\xE1\xBE\x80\x00\xE1\xBE\x81\x00\xE1\xBE\x82\x00\xE1\xBE\x83\x00\xE1\xBE"
-     "\x84\x00"
-     "\xE1\xBE\x85\x00\xE1\xBE\x86\x00\xE1\xBE\x87\x00\xE1\xBE\x88\x00\xE1\xBE"
-     "\x89\x00"
-     "\xE1\xBE\x8A\x00\xE1\xBE\x8B\x00\xE1\xBE\x8C\x00\xE1\xBE\x8D\x00\xE1\xBE"
-     "\x8E\x00"
-     "\xE1\xBE\x8F\x00\xE1\xBE\x90\x00\xE1\xBE\x91\x00\xE1\xBE\x92\x00\xE1\xBE"
-     "\x93\x00"
-     "\xE1\xBE\x94\x00\xE1\xBE\x95\x00\xE1\xBE\x96\x00\xE1\xBE\x97\x00\xE1\xBE"
-     "\x98\x00"
-     "\xE1\xBE\x99\x00\xE1\xBE\x9A\x00\xE1\xBE\x9B\x00\xE1\xBE\x9C\x00\xE1\xBE"
-     "\x9D\x00"
-     "\xE1\xBE\x9E\x00\xE1\xBE\x9F\x00\xE1\xBE\xA0\x00\xE1\xBE\xA1\x00\xE1\xBE"
-     "\xA2\x00"
-     "\xE1\xBE\xA3\x00\xE1\xBE\xA4\x00\xE1\xBE\xA5\x00\xE1\xBE\xA6\x00\xE1\xBE"
-     "\xA7\x00"
-     "\xE1\xBE\xA8\x00\xE1\xBE\xA9\x00\xE1\xBE\xAA\x00\xE1\xBE\xAB\x00\xE1\xBE"
-     "\xAC\x00"
-     "\xE1\xBE\xAD\x00\xE1\xBE\xAE\x00\xE1\xBE\xAF\x00\xE1\xBE\xB0\x00\xE1\xBE"
-     "\xB1\x00"
-     "\xE1\xBE\xB2\x00\xE1\xBE\xB3\x00\xE1\xBE\xB4\x00\xE1\xBE\xB6\x00\xE1\xBE"
-     "\xB7\x00"
-     "\xE1\xBE\xB8\x00\xE1\xBE\xB9\x00\xE1\xBE\xBA\x00\xE1\xBE\xBC\x00\xE1\xBF"
-     "\x82\x00"
-     "\xE1\xBF\x83\x00\xE1\xBF\x84\x00\xE1\xBF\x86\x00\xE1\xBF\x87\x00\xE1\xBF"
-     "\x88\x00"
-     "\xE1\xBF\x8A\x00\xE1\xBF\x8C\x00\xE1\xBF\x90\x00\xE1\xBF\x91\x00\xE1\xBF"
-     "\x92\x00"
-     "\xE1\xBF\x96\x00\xE1\xBF\x97\x00\xE1\xBF\x98\x00\xE1\xBF\x99\x00\xE1\xBF"
-     "\x9A\x00"
-     "\xE1\xBF\xA0\x00\xE1\xBF\xA1\x00\xE1\xBF\xA2\x00\xE1\xBF\xA4\x00\xE1\xBF"
-     "\xA5\x00"
-     "\xE1\xBF\xA6\x00\xE1\xBF\xA7\x00\xE1\xBF\xA8\x00\xE1\xBF\xA9\x00\xE1\xBF"
-     "\xAA\x00"
-     "\xE1\xBF\xAC\x00\xE1\xBF\xB2\x00\xE1\xBF\xB3\x00\xE1\xBF\xB4\x00\xE1\xBF"
-     "\xB6\x00"
-     "\xE1\xBF\xB7\x00\xE1\xBF\xB8\x00\xE1\xBF\xBA\x00\xE1\xBF\xBC\x00\xE2\x80"
-     "\x90\x00"
-     "\xE2\x80\x93\x00\xE2\x80\x94\x00\xE2\x80\xB2\xE2\x80\xB2\x00\xE2\x80\xB2"
-     "\xE2\x80"
-     "\xB2\xE2\x80\xB2\x00\xE2\x80\xB2\xE2\x80\xB2\xE2\x80\xB2\xE2\x80\xB2\x00"
-     "\xE2\x80"
-     "\xB5\xE2\x80\xB5\x00\xE2\x80\xB5\xE2\x80\xB5\xE2\x80\xB5\x00\xE2\x82\xA9"
-     "\x00\xE2"
-     "\x86\x90\x00\xE2\x86\x91\x00\xE2\x86\x92\x00\xE2\x86\x93\x00\xE2\x86\x9A"
-     "\x00\xE2"
-     "\x86\x9B\x00\xE2\x86\xAE\x00\xE2\x87\x8D\x00\xE2\x87\x8E\x00\xE2\x87\x8F"
-     "\x00\xE2"
-     "\x88\x82\x00\xE2\x88\x84\x00\xE2\x88\x87\x00\xE2\x88\x89\x00\xE2\x88\x8C"
-     "\x00\xE2"
-     "\x88\x91\x00\xE2\x88\x92\x00\xE2\x88\xA4\x00\xE2\x88\xA6\x00\xE2\x88\xAB"
-     "\xE2\x88"
-     "\xAB\x00\xE2\x88\xAB\xE2\x88\xAB\xE2\x88\xAB\x00\xE2\x88\xAB\xE2\x88\xAB"
-     "\xE2\x88"
-     "\xAB\xE2\x88\xAB\x00\xE2\x88\xAE\xE2\x88\xAE\x00\xE2\x88\xAE\xE2\x88\xAE"
-     "\xE2\x88"
-     "\xAE\x00\xE2\x89\x81\x00\xE2\x89\x84\x00\xE2\x89\x87\x00\xE2\x89\x89\x00"
-     "\xE2\x89"
-     "\xA0\x00\xE2\x89\xA2\x00\xE2\x89\xAD\x00\xE2\x89\xAE\x00\xE2\x89\xAF\x00"
-     "\xE2\x89"
-     "\xB0\x00\xE2\x89\xB1\x00\xE2\x89\xB4\x00\xE2\x89\xB5\x00\xE2\x89\xB8\x00"
-     "\xE2\x89"
-     "\xB9\x00\xE2\x8A\x80\x00\xE2\x8A\x81\x00\xE2\x8A\x84\x00\xE2\x8A\x85\x00"
-     "\xE2\x8A"
-     "\x88\x00\xE2\x8A\x89\x00\xE2\x8A\xAC\x00\xE2\x8A\xAD\x00\xE2\x8A\xAE\x00"
-     "\xE2\x8A"
-     "\xAF\x00\xE2\x8B\xA0\x00\xE2\x8B\xA1\x00\xE2\x8B\xA2\x00\xE2\x8B\xA3\x00"
-     "\xE2\x8B"
-     "\xAA\x00\xE2\x8B\xAB\x00\xE2\x8B\xAC\x00\xE2\x8B\xAD\x00\xE2\x94\x82\x00"
-     "\xE2\x96"
-     "\xA0\x00\xE2\x97\x8B\x00\xE2\xA6\x85\x00\xE2\xA6\x86\x00\xE2\xAB\x9D\xCC"
-     "\xB8\x00"
-     "\xE2\xB1\xB1\x00\xE2\xB5\xA1\x00\xE3\x80\x81\x00\xE3\x80\x82\x00\xE3\x80"
-     "\x88\x00"
-     "\xE3\x80\x89\x00\xE3\x80\x8A\x00\xE3\x80\x8B\x00\xE3\x80\x8C\x00\xE3\x80"
-     "\x8D\x00"
-     "\xE3\x80\x8E\x00\xE3\x80\x8F\x00\xE3\x80\x90\x00\xE3\x80\x91\x00\xE3\x80"
-     "\x92\x00"
-     "\xE3\x80\x94\x00\xE3\x80\x94\x53\xE3\x80\x95\x00\xE3\x80\x94\xE4\xB8\x89"
-     "\xE3\x80"
-     "\x95\x00\xE3\x80\x94\xE4\xBA\x8C\xE3\x80\x95\x00\xE3\x80\x94\xE5\x8B\x9D"
-     "\xE3\x80"
-     "\x95\x00\xE3\x80\x94\xE5\xAE\x89\xE3\x80\x95\x00\xE3\x80\x94\xE6\x89\x93"
-     "\xE3\x80"
-     "\x95\x00\xE3\x80\x94\xE6\x95\x97\xE3\x80\x95\x00\xE3\x80\x94\xE6\x9C\xAC"
-     "\xE3\x80"
-     "\x95\x00\xE3\x80\x94\xE7\x82\xB9\xE3\x80\x95\x00\xE3\x80\x94\xE7\x9B\x97"
-     "\xE3\x80"
-     "\x95\x00\xE3\x80\x95\x00\xE3\x80\x96\x00\xE3\x80\x97\x00\xE3\x81\x8C\x00"
-     "\xE3\x81"
-     "\x8E\x00\xE3\x81\x90\x00\xE3\x81\x92\x00\xE3\x81\x94\x00\xE3\x81\x96\x00"
-     "\xE3\x81"
-     "\x98\x00\xE3\x81\x9A\x00\xE3\x81\x9C\x00\xE3\x81\x9E\x00\xE3\x81\xA0\x00"
-     "\xE3\x81"
-     "\xA2\x00\xE3\x81\xA5\x00\xE3\x81\xA7\x00\xE3\x81\xA9\x00\xE3\x81\xB0\x00"
-     "\xE3\x81"
-     "\xB1\x00\xE3\x81\xB3\x00\xE3\x81\xB4\x00\xE3\x81\xB6\x00\xE3\x81\xB7\x00"
-     "\xE3\x81"
-     "\xB9\x00\xE3\x81\xBA\x00\xE3\x81\xBB\xE3\x81\x8B\x00\xE3\x81\xBC\x00\xE3"
-     "\x81\xBD"
-     "\x00\xE3\x82\x88\xE3\x82\x8A\x00\xE3\x82\x94\x00\xE3\x82\x99\x00\xE3\x82"
-     "\x9A\x00"
-     "\xE3\x82\x9E\x00\xE3\x82\xA1\x00\xE3\x82\xA2\x00\xE3\x82\xA2\xE3\x83\x91"
-     "\xE3\x83"
-     "\xBC\xE3\x83\x88\x00\xE3\x82\xA2\xE3\x83\xAB\xE3\x83\x95\xE3\x82\xA1\x00"
-     "\xE3\x82"
-     "\xA2\xE3\x83\xB3\xE3\x83\x9A\xE3\x82\xA2\x00\xE3\x82\xA2\xE3\x83\xBC\xE3"
-     "\x83\xAB"
-     "\x00\xE3\x82\xA3\x00\xE3\x82\xA4\x00\xE3\x82\xA4\xE3\x83\x8B\xE3\x83\xB3"
-     "\xE3\x82"
-     "\xB0\x00\xE3\x82\xA4\xE3\x83\xB3\xE3\x83\x81\x00\xE3\x82\xA5\x00\xE3\x82"
-     "\xA6\x00"
-     "\xE3\x82\xA6\xE3\x82\xA9\xE3\x83\xB3\x00\xE3\x82\xA7\x00\xE3\x82\xA8\x00"
-     "\xE3\x82"
-     "\xA8\xE3\x82\xB9\xE3\x82\xAF\xE3\x83\xBC\xE3\x83\x89\x00\xE3\x82\xA8\xE3"
-     "\x83\xBC"
-     "\xE3\x82\xAB\xE3\x83\xBC\x00\xE3\x82\xA9\x00\xE3\x82\xAA\x00\xE3\x82\xAA"
-     "\xE3\x83"
-     "\xB3\xE3\x82\xB9\x00\xE3\x82\xAA\xE3\x83\xBC\xE3\x83\xA0\x00\xE3\x82\xAB"
-     "\x00\xE3"
-     "\x82\xAB\xE3\x82\xA4\xE3\x83\xAA\x00\xE3\x82\xAB\xE3\x83\xA9\xE3\x83\x83"
-     "\xE3\x83"
-     "\x88\x00\xE3\x82\xAB\xE3\x83\xAD\xE3\x83\xAA\xE3\x83\xBC\x00\xE3\x82\xAC"
-     "\x00\xE3"
-     "\x82\xAC\xE3\x83\xAD\xE3\x83\xB3\x00\xE3\x82\xAC\xE3\x83\xB3\xE3\x83\x9E"
-     "\x00\xE3"
-     "\x82\xAD\x00\xE3\x82\xAD\xE3\x83\xA5\xE3\x83\xAA\xE3\x83\xBC\x00\xE3\x82"
-     "\xAD\xE3"
-     "\x83\xAD\x00\xE3\x82\xAD\xE3\x83\xAD\xE3\x82\xB0\xE3\x83\xA9\xE3\x83\xA0"
-     "\x00\xE3"
-     "\x82\xAD\xE3\x83\xAD\xE3\x83\xA1\xE3\x83\xBC\xE3\x83\x88\xE3\x83\xAB\x00"
-     "\xE3\x82"
-     "\xAD\xE3\x83\xAD\xE3\x83\xAF\xE3\x83\x83\xE3\x83\x88\x00\xE3\x82\xAE\x00"
-     "\xE3\x82"
-     "\xAE\xE3\x82\xAC\x00\xE3\x82\xAE\xE3\x83\x8B\xE3\x83\xBC\x00\xE3\x82\xAE"
-     "\xE3\x83"
-     "\xAB\xE3\x83\x80\xE3\x83\xBC\x00\xE3\x82\xAF\x00\xE3\x82\xAF\xE3\x83\xAB"
-     "\xE3\x82"
-     "\xBC\xE3\x82\xA4\xE3\x83\xAD\x00\xE3\x82\xAF\xE3\x83\xAD\xE3\x83\xBC\xE3"
-     "\x83\x8D"
-     "\x00\xE3\x82\xB0\x00\xE3\x82\xB0\xE3\x83\xA9\xE3\x83\xA0\x00\xE3\x82\xB0"
-     "\xE3\x83"
-     "\xA9\xE3\x83\xA0\xE3\x83\x88\xE3\x83\xB3\x00\xE3\x82\xB1\x00\xE3\x82\xB1"
-     "\xE3\x83"
-     "\xBC\xE3\x82\xB9\x00\xE3\x82\xB2\x00\xE3\x82\xB3\x00\xE3\x82\xB3\xE3\x82"
-     "\xB3\x00"
-     "\xE3\x82\xB3\xE3\x83\x88\x00\xE3\x82\xB3\xE3\x83\xAB\xE3\x83\x8A\x00\xE3"
-     "\x82\xB3"
-     "\xE3\x83\xBC\xE3\x83\x9D\x00\xE3\x82\xB4\x00\xE3\x82\xB5\x00\xE3\x82\xB5"
-     "\xE3\x82"
-     "\xA4\xE3\x82\xAF\xE3\x83\xAB\x00\xE3\x82\xB5\xE3\x83\xB3\xE3\x83\x81\xE3"
-     "\x83\xBC"
-     "\xE3\x83\xA0\x00\xE3\x82\xB6\x00\xE3\x82\xB7\x00\xE3\x82\xB7\xE3\x83\xAA"
-     "\xE3\x83"
-     "\xB3\xE3\x82\xB0\x00\xE3\x82\xB8\x00\xE3\x82\xB9\x00\xE3\x82\xBA\x00\xE3"
-     "\x82\xBB"
-     "\x00\xE3\x82\xBB\xE3\x83\xB3\xE3\x83\x81\x00\xE3\x82\xBB\xE3\x83\xB3\xE3"
-     "\x83\x88"
-     "\x00\xE3\x82\xBC\x00\xE3\x82\xBD\x00\xE3\x82\xBE\x00\xE3\x82\xBF\x00\xE3"
-     "\x83\x80"
-     "\x00\xE3\x83\x80\xE3\x83\xBC\xE3\x82\xB9\x00\xE3\x83\x81\x00\xE3\x83\x82"
-     "\x00\xE3"
-     "\x83\x83\x00\xE3\x83\x84\x00\xE3\x83\x85\x00\xE3\x83\x86\x00\xE3\x83\x87"
-     "\x00\xE3"
-     "\x83\x87\xE3\x82\xB7\x00\xE3\x83\x88\x00\xE3\x83\x88\xE3\x83\xB3\x00\xE3"
-     "\x83\x89"
-     "\x00\xE3\x83\x89\xE3\x83\xAB\x00\xE3\x83\x8A\x00\xE3\x83\x8A\xE3\x83\x8E"
-     "\x00\xE3"
-     "\x83\x8B\x00\xE3\x83\x8C\x00\xE3\x83\x8D\x00\xE3\x83\x8E\x00\xE3\x83\x8E"
-     "\xE3\x83"
-     "\x83\xE3\x83\x88\x00\xE3\x83\x8F\x00\xE3\x83\x8F\xE3\x82\xA4\xE3\x83\x84"
-     "\x00\xE3"
-     "\x83\x90\x00\xE3\x83\x90\xE3\x83\xBC\xE3\x83\xAC\xE3\x83\xAB\x00\xE3\x83"
-     "\x91\x00"
-     "\xE3\x83\x91\xE3\x83\xBC\xE3\x82\xBB\xE3\x83\xB3\xE3\x83\x88\x00\xE3\x83"
-     "\x91\xE3"
-     "\x83\xBC\xE3\x83\x84\x00\xE3\x83\x92\x00\xE3\x83\x93\x00\xE3\x83\x93\xE3"
-     "\x83\xAB"
-     "\x00\xE3\x83\x94\x00\xE3\x83\x94\xE3\x82\xA2\xE3\x82\xB9\xE3\x83\x88\xE3"
-     "\x83\xAB"
-     "\x00\xE3\x83\x94\xE3\x82\xAF\xE3\x83\xAB\x00\xE3\x83\x94\xE3\x82\xB3\x00"
-     "\xE3\x83"
-     "\x95\x00\xE3\x83\x95\xE3\x82\xA1\xE3\x83\xA9\xE3\x83\x83\xE3\x83\x89\x00"
-     "\xE3\x83"
-     "\x95\xE3\x82\xA3\xE3\x83\xBC\xE3\x83\x88\x00\xE3\x83\x95\xE3\x83\xA9\xE3"
-     "\x83\xB3"
-     "\x00\xE3\x83\x96\x00\xE3\x83\x96\xE3\x83\x83\xE3\x82\xB7\xE3\x82\xA7\xE3"
-     "\x83\xAB"
-     "\x00\xE3\x83\x97\x00\xE3\x83\x98\x00\xE3\x83\x98\xE3\x82\xAF\xE3\x82\xBF"
-     "\xE3\x83"
-     "\xBC\xE3\x83\xAB\x00\xE3\x83\x98\xE3\x83\xAB\xE3\x83\x84\x00\xE3\x83\x99"
-     "\x00\xE3"
-     "\x83\x99\xE3\x83\xBC\xE3\x82\xBF\x00\xE3\x83\x9A\x00\xE3\x83\x9A\xE3\x82"
-     "\xBD\x00"
-     "\xE3\x83\x9A\xE3\x83\x8B\xE3\x83\x92\x00\xE3\x83\x9A\xE3\x83\xB3\xE3\x82"
-     "\xB9\x00"
-     "\xE3\x83\x9A\xE3\x83\xBC\xE3\x82\xB8\x00\xE3\x83\x9B\x00\xE3\x83\x9B\xE3"
-     "\x83\xB3"
-     "\x00\xE3\x83\x9B\xE3\x83\xBC\xE3\x83\xAB\x00\xE3\x83\x9B\xE3\x83\xBC\xE3"
-     "\x83\xB3"
-     "\x00\xE3\x83\x9C\x00\xE3\x83\x9C\xE3\x83\xAB\xE3\x83\x88\x00\xE3\x83\x9D"
-     "\x00\xE3"
-     "\x83\x9D\xE3\x82\xA4\xE3\x83\xB3\xE3\x83\x88\x00\xE3\x83\x9D\xE3\x83\xB3"
-     "\xE3\x83"
-     "\x89\x00\xE3\x83\x9E\x00\xE3\x83\x9E\xE3\x82\xA4\xE3\x82\xAF\xE3\x83\xAD"
-     "\x00\xE3"
-     "\x83\x9E\xE3\x82\xA4\xE3\x83\xAB\x00\xE3\x83\x9E\xE3\x83\x83\xE3\x83\x8F"
-     "\x00\xE3"
-     "\x83\x9E\xE3\x83\xAB\xE3\x82\xAF\x00\xE3\x83\x9E\xE3\x83\xB3\xE3\x82\xB7"
-     "\xE3\x83"
-     "\xA7\xE3\x83\xB3\x00\xE3\x83\x9F\x00\xE3\x83\x9F\xE3\x82\xAF\xE3\x83\xAD"
-     "\xE3\x83"
-     "\xB3\x00\xE3\x83\x9F\xE3\x83\xAA\x00\xE3\x83\x9F\xE3\x83\xAA\xE3\x83\x90"
-     "\xE3\x83"
-     "\xBC\xE3\x83\xAB\x00\xE3\x83\xA0\x00\xE3\x83\xA1\x00\xE3\x83\xA1\xE3\x82"
-     "\xAC\x00"
-     "\xE3\x83\xA1\xE3\x82\xAC\xE3\x83\x88\xE3\x83\xB3\x00\xE3\x83\xA1\xE3\x83"
-     "\xBC\xE3"
-     "\x83\x88\xE3\x83\xAB\x00\xE3\x83\xA2\x00\xE3\x83\xA3\x00\xE3\x83\xA4\x00"
-     "\xE3\x83"
-     "\xA4\xE3\x83\xBC\xE3\x83\x89\x00\xE3\x83\xA4\xE3\x83\xBC\xE3\x83\xAB\x00"
-     "\xE3\x83"
-     "\xA5\x00\xE3\x83\xA6\x00\xE3\x83\xA6\xE3\x82\xA2\xE3\x83\xB3\x00\xE3\x83"
-     "\xA7\x00"
-     "\xE3\x83\xA8\x00\xE3\x83\xA9\x00\xE3\x83\xAA\x00\xE3\x83\xAA\xE3\x83\x83"
-     "\xE3\x83"
-     "\x88\xE3\x83\xAB\x00\xE3\x83\xAA\xE3\x83\xA9\x00\xE3\x83\xAB\x00\xE3\x83"
-     "\xAB\xE3"
-     "\x83\x94\xE3\x83\xBC\x00\xE3\x83\xAB\xE3\x83\xBC\xE3\x83\x96\xE3\x83\xAB"
-     "\x00\xE3"
-     "\x83\xAC\x00\xE3\x83\xAC\xE3\x83\xA0\x00\xE3\x83\xAC\xE3\x83\xB3\xE3\x83"
-     "\x88\xE3"
-     "\x82\xB2\xE3\x83\xB3\x00\xE3\x83\xAD\x00\xE3\x83\xAF\x00\xE3\x83\xAF\xE3"
-     "\x83\x83"
-     "\xE3\x83\x88\x00\xE3\x83\xB0\x00\xE3\x83\xB1\x00\xE3\x83\xB2\x00\xE3\x83"
-     "\xB3\x00"
-     "\xE3\x83\xB4\x00\xE3\x83\xB7\x00\xE3\x83\xB8\x00\xE3\x83\xB9\x00\xE3\x83"
-     "\xBA\x00"
-     "\xE3\x83\xBB\x00\xE3\x83\xBC\x00\xE3\x83\xBE\x00\xE3\x92\x9E\x00\xE3\x92"
-     "\xB9\x00"
-     "\xE3\x92\xBB\x00\xE3\x93\x9F\x00\xE3\x94\x95\x00\xE3\x9B\xAE\x00\xE3\x9B"
-     "\xBC\x00"
-     "\xE3\x9E\x81\x00\xE3\xA0\xAF\x00\xE3\xA1\xA2\x00\xE3\xA1\xBC\x00\xE3\xA3"
-     "\x87\x00"
-     "\xE3\xA3\xA3\x00\xE3\xA4\x9C\x00\xE3\xA4\xBA\x00\xE3\xA8\xAE\x00\xE3\xA9"
-     "\xAC\x00"
-     "\xE3\xAB\xA4\x00\xE3\xAC\x88\x00\xE3\xAC\x99\x00\xE3\xAD\x89\x00\xE3\xAE"
-     "\x9D\x00"
-     "\xE3\xB0\x98\x00\xE3\xB1\x8E\x00\xE3\xB4\xB3\x00\xE3\xB6\x96\x00\xE3\xBA"
-     "\xAC\x00"
-     "\xE3\xBA\xB8\x00\xE3\xBC\x9B\x00\xE3\xBF\xBC\x00\xE4\x80\x88\x00\xE4\x80"
-     "\x98\x00"
-     "\xE4\x80\xB9\x00\xE4\x81\x86\x00\xE4\x82\x96\x00\xE4\x83\xA3\x00\xE4\x84"
-     "\xAF\x00"
-     "\xE4\x88\x82\x00\xE4\x88\xA7\x00\xE4\x8A\xA0\x00\xE4\x8C\x81\x00\xE4\x8C"
-     "\xB4\x00"
-     "\xE4\x8D\x99\x00\xE4\x8F\x95\x00\xE4\x8F\x99\x00\xE4\x90\x8B\x00\xE4\x91"
-     "\xAB\x00"
-     "\xE4\x94\xAB\x00\xE4\x95\x9D\x00\xE4\x95\xA1\x00\xE4\x95\xAB\x00\xE4\x97"
-     "\x97\x00"
-     "\xE4\x97\xB9\x00\xE4\x98\xB5\x00\xE4\x9A\xBE\x00\xE4\x9B\x87\x00\xE4\xA6"
-     "\x95\x00"
-     "\xE4\xA7\xA6\x00\xE4\xA9\xAE\x00\xE4\xA9\xB6\x00\xE4\xAA\xB2\x00\xE4\xAC"
-     "\xB3\x00"
-     "\xE4\xAF\x8E\x00\xE4\xB3\x8E\x00\xE4\xB3\xAD\x00\xE4\xB3\xB8\x00\xE4\xB5"
-     "\x96\x00"
-     "\xE4\xB8\x80\x00\xE4\xB8\x81\x00\xE4\xB8\x83\x00\xE4\xB8\x89\x00\xE4\xB8"
-     "\x8A\x00"
-     "\xE4\xB8\x8B\x00\xE4\xB8\x8D\x00\xE4\xB8\x99\x00\xE4\xB8\xA6\x00\xE4\xB8"
-     "\xA8\x00"
-     "\xE4\xB8\xAD\x00\xE4\xB8\xB2\x00\xE4\xB8\xB6\x00\xE4\xB8\xB8\x00\xE4\xB8"
-     "\xB9\x00"
-     "\xE4\xB8\xBD\x00\xE4\xB8\xBF\x00\xE4\xB9\x81\x00\xE4\xB9\x99\x00\xE4\xB9"
-     "\x9D\x00"
-     "\xE4\xBA\x82\x00\xE4\xBA\x85\x00\xE4\xBA\x86\x00\xE4\xBA\x8C\x00\xE4\xBA"
-     "\x94\x00"
-     "\xE4\xBA\xA0\x00\xE4\xBA\xA4\x00\xE4\xBA\xAE\x00\xE4\xBA\xBA\x00\xE4\xBB"
-     "\x80\x00"
-     "\xE4\xBB\x8C\x00\xE4\xBB\xA4\x00\xE4\xBB\xA4\xE5\x92\x8C\x00\xE4\xBC\x81"
-     "\x00\xE4"
-     "\xBC\x91\x00\xE4\xBD\xA0\x00\xE4\xBE\x80\x00\xE4\xBE\x86\x00\xE4\xBE\x8B"
-     "\x00\xE4"
-     "\xBE\xAE\x00\xE4\xBE\xBB\x00\xE4\xBE\xBF\x00\xE5\x80\x82\x00\xE5\x80\xAB"
-     "\x00\xE5"
-     "\x81\xBA\x00\xE5\x82\x99\x00\xE5\x83\x8F\x00\xE5\x83\x9A\x00\xE5\x83\xA7"
-     "\x00\xE5"
-     "\x84\xAA\x00\xE5\x84\xBF\x00\xE5\x85\x80\x00\xE5\x85\x85\x00\xE5\x85\x8D"
-     "\x00\xE5"
-     "\x85\x94\x00\xE5\x85\xA4\x00\xE5\x85\xA5\x00\xE5\x85\xA7\x00\xE5\x85\xA8"
-     "\x00\xE5"
-     "\x85\xA9\x00\xE5\x85\xAB\x00\xE5\x85\xAD\x00\xE5\x85\xB7\x00\xE5\x86\x80"
-     "\x00\xE5"
-     "\x86\x82\x00\xE5\x86\x8D\x00\xE5\x86\x92\x00\xE5\x86\x95\x00\xE5\x86\x96"
-     "\x00\xE5"
-     "\x86\x97\x00\xE5\x86\x99\x00\xE5\x86\xA4\x00\xE5\x86\xAB\x00\xE5\x86\xAC"
-     "\x00\xE5"
-     "\x86\xB5\x00\xE5\x86\xB7\x00\xE5\x87\x89\x00\xE5\x87\x8C\x00\xE5\x87\x9C"
-     "\x00\xE5"
-     "\x87\x9E\x00\xE5\x87\xA0\x00\xE5\x87\xB5\x00\xE5\x88\x80\x00\xE5\x88\x83"
-     "\x00\xE5"
-     "\x88\x87\x00\xE5\x88\x97\x00\xE5\x88\x9D\x00\xE5\x88\xA9\x00\xE5\x88\xBA"
-     "\x00\xE5"
-     "\x88\xBB\x00\xE5\x89\x86\x00\xE5\x89\x8D\x00\xE5\x89\xB2\x00\xE5\x89\xB7"
-     "\x00\xE5"
-     "\x8A\x89\x00\xE5\x8A\x9B\x00\xE5\x8A\xA3\x00\xE5\x8A\xB3\x00\xE5\x8A\xB4"
-     "\x00\xE5"
-     "\x8B\x87\x00\xE5\x8B\x89\x00\xE5\x8B\x92\x00\xE5\x8B\x9E\x00\xE5\x8B\xA4"
-     "\x00\xE5"
-     "\x8B\xB5\x00\xE5\x8B\xB9\x00\xE5\x8B\xBA\x00\xE5\x8C\x85\x00\xE5\x8C\x86"
-     "\x00\xE5"
-     "\x8C\x95\x00\xE5\x8C\x97\x00\xE5\x8C\x9A\x00\xE5\x8C\xB8\x00\xE5\x8C\xBB"
-     "\x00\xE5"
-     "\x8C\xBF\x00\xE5\x8D\x81\x00\xE5\x8D\x84\x00\xE5\x8D\x85\x00\xE5\x8D\x89"
-     "\x00\xE5"
-     "\x8D\x91\x00\xE5\x8D\x94\x00\xE5\x8D\x9A\x00\xE5\x8D\x9C\x00\xE5\x8D\xA9"
-     "\x00\xE5"
-     "\x8D\xB0\x00\xE5\x8D\xB3\x00\xE5\x8D\xB5\x00\xE5\x8D\xBD\x00\xE5\x8D\xBF"
-     "\x00\xE5"
-     "\x8E\x82\x00\xE5\x8E\xB6\x00\xE5\x8F\x83\x00\xE5\x8F\x88\x00\xE5\x8F\x8A"
-     "\x00\xE5"
-     "\x8F\x8C\x00\xE5\x8F\x9F\x00\xE5\x8F\xA3\x00\xE5\x8F\xA5\x00\xE5\x8F\xAB"
-     "\x00\xE5"
-     "\x8F\xAF\x00\xE5\x8F\xB1\x00\xE5\x8F\xB3\x00\xE5\x90\x86\x00\xE5\x90\x88"
-     "\x00\xE5"
-     "\x90\x8D\x00\xE5\x90\x8F\x00\xE5\x90\x9D\x00\xE5\x90\xB8\x00\xE5\x90\xB9"
-     "\x00\xE5"
-     "\x91\x82\x00\xE5\x91\x88\x00\xE5\x91\xA8\x00\xE5\x92\x9E\x00\xE5\x92\xA2"
-     "\x00\xE5"
-     "\x92\xBD\x00\xE5\x93\xB6\x00\xE5\x94\x90\x00\xE5\x95\x8F\x00\xE5\x95\x93"
-     "\x00\xE5"
-     "\x95\x95\x00\xE5\x95\xA3\x00\xE5\x96\x84\x00\xE5\x96\x87\x00\xE5\x96\x99"
-     "\x00\xE5"
-     "\x96\x9D\x00\xE5\x96\xAB\x00\xE5\x96\xB3\x00\xE5\x96\xB6\x00\xE5\x97\x80"
-     "\x00\xE5"
-     "\x97\x82\x00\xE5\x97\xA2\x00\xE5\x98\x86\x00\xE5\x99\x91\x00\xE5\x99\xA8"
-     "\x00\xE5"
-     "\x99\xB4\x00\xE5\x9B\x97\x00\xE5\x9B\x9B\x00\xE5\x9B\xB9\x00\xE5\x9C\x96"
-     "\x00\xE5"
-     "\x9C\x97\x00\xE5\x9C\x9F\x00\xE5\x9C\xB0\x00\xE5\x9E\x8B\x00\xE5\x9F\x8E"
-     "\x00\xE5"
-     "\x9F\xB4\x00\xE5\xA0\x8D\x00\xE5\xA0\xB1\x00\xE5\xA0\xB2\x00\xE5\xA1\x80"
-     "\x00\xE5"
-     "\xA1\x9A\x00\xE5\xA1\x9E\x00\xE5\xA2\xA8\x00\xE5\xA2\xAC\x00\xE5\xA2\xB3"
-     "\x00\xE5"
-     "\xA3\x98\x00\xE5\xA3\x9F\x00\xE5\xA3\xAB\x00\xE5\xA3\xAE\x00\xE5\xA3\xB0"
-     "\x00\xE5"
-     "\xA3\xB2\x00\xE5\xA3\xB7\x00\xE5\xA4\x82\x00\xE5\xA4\x86\x00\xE5\xA4\x8A"
-     "\x00\xE5"
-     "\xA4\x95\x00\xE5\xA4\x9A\x00\xE5\xA4\x9C\x00\xE5\xA4\xA2\x00\xE5\xA4\xA7"
-     "\x00\xE5"
-     "\xA4\xA7\xE6\xAD\xA3\x00\xE5\xA4\xA9\x00\xE5\xA5\x84\x00\xE5\xA5\x88\x00"
-     "\xE5\xA5"
-     "\x91\x00\xE5\xA5\x94\x00\xE5\xA5\xA2\x00\xE5\xA5\xB3\x00\xE5\xA7\x98\x00"
-     "\xE5\xA7"
-     "\xAC\x00\xE5\xA8\x9B\x00\xE5\xA8\xA7\x00\xE5\xA9\xA2\x00\xE5\xA9\xA6\x00"
-     "\xE5\xAA"
-     "\xB5\x00\xE5\xAC\x88\x00\xE5\xAC\xA8\x00\xE5\xAC\xBE\x00\xE5\xAD\x90\x00"
-     "\xE5\xAD"
-     "\x97\x00\xE5\xAD\xA6\x00\xE5\xAE\x80\x00\xE5\xAE\x85\x00\xE5\xAE\x97\x00"
-     "\xE5\xAF"
-     "\x83\x00\xE5\xAF\x98\x00\xE5\xAF\xA7\x00\xE5\xAF\xAE\x00\xE5\xAF\xB3\x00"
-     "\xE5\xAF"
-     "\xB8\x00\xE5\xAF\xBF\x00\xE5\xB0\x86\x00\xE5\xB0\x8F\x00\xE5\xB0\xA2\x00"
-     "\xE5\xB0"
-     "\xB8\x00\xE5\xB0\xBF\x00\xE5\xB1\xA0\x00\xE5\xB1\xA2\x00\xE5\xB1\xA4\x00"
-     "\xE5\xB1"
-     "\xA5\x00\xE5\xB1\xAE\x00\xE5\xB1\xB1\x00\xE5\xB2\x8D\x00\xE5\xB3\x80\x00"
-     "\xE5\xB4"
-     "\x99\x00\xE5\xB5\x83\x00\xE5\xB5\x90\x00\xE5\xB5\xAB\x00\xE5\xB5\xAE\x00"
-     "\xE5\xB5"
-     "\xBC\x00\xE5\xB6\xB2\x00\xE5\xB6\xBA\x00\xE5\xB7\x9B\x00\xE5\xB7\xA1\x00"
-     "\xE5\xB7"
-     "\xA2\x00\xE5\xB7\xA5\x00\xE5\xB7\xA6\x00\xE5\xB7\xB1\x00\xE5\xB7\xBD\x00"
-     "\xE5\xB7"
-     "\xBE\x00\xE5\xB8\xA8\x00\xE5\xB8\xBD\x00\xE5\xB9\xA9\x00\xE5\xB9\xB2\x00"
-     "\xE5\xB9"
-     "\xB3\xE6\x88\x90\x00\xE5\xB9\xB4\x00\xE5\xB9\xBA\x00\xE5\xB9\xBC\x00\xE5"
-     "\xB9\xBF"
-     "\x00\xE5\xBA\xA6\x00\xE5\xBA\xB0\x00\xE5\xBA\xB3\x00\xE5\xBA\xB6\x00\xE5"
-     "\xBB\x89"
-     "\x00\xE5\xBB\x8A\x00\xE5\xBB\x92\x00\xE5\xBB\x93\x00\xE5\xBB\x99\x00\xE5"
-     "\xBB\xAC"
-     "\x00\xE5\xBB\xB4\x00\xE5\xBB\xBE\x00\xE5\xBC\x84\x00\xE5\xBC\x8B\x00\xE5"
-     "\xBC\x93"
-     "\x00\xE5\xBC\xA2\x00\xE5\xBD\x90\x00\xE5\xBD\x93\x00\xE5\xBD\xA1\x00\xE5"
-     "\xBD\xA2"
-     "\x00\xE5\xBD\xA9\x00\xE5\xBD\xAB\x00\xE5\xBD\xB3\x00\xE5\xBE\x8B\x00\xE5"
-     "\xBE\x8C"
-     "\x00\xE5\xBE\x97\x00\xE5\xBE\x9A\x00\xE5\xBE\xA9\x00\xE5\xBE\xAD\x00\xE5"
-     "\xBF\x83"
-     "\x00\xE5\xBF\x8D\x00\xE5\xBF\x97\x00\xE5\xBF\xB5\x00\xE5\xBF\xB9\x00\xE6"
-     "\x80\x92"
-     "\x00\xE6\x80\x9C\x00\xE6\x81\xB5\x00\xE6\x82\x81\x00\xE6\x82\x94\x00\xE6"
-     "\x83\x87"
-     "\x00\xE6\x83\x98\x00\xE6\x83\xA1\x00\xE6\x84\x88\x00\xE6\x85\x84\x00\xE6"
-     "\x85\x88"
-     "\x00\xE6\x85\x8C\x00\xE6\x85\x8E\x00\xE6\x85\xA0\x00\xE6\x85\xA8\x00\xE6"
-     "\x85\xBA"
-     "\x00\xE6\x86\x8E\x00\xE6\x86\x90\x00\xE6\x86\xA4\x00\xE6\x86\xAF\x00\xE6"
-     "\x86\xB2"
-     "\x00\xE6\x87\x9E\x00\xE6\x87\xB2\x00\xE6\x87\xB6\x00\xE6\x88\x80\x00\xE6"
-     "\x88\x88"
-     "\x00\xE6\x88\x90\x00\xE6\x88\x9B\x00\xE6\x88\xAE\x00\xE6\x88\xB4\x00\xE6"
-     "\x88\xB6"
-     "\x00\xE6\x89\x8B\x00\xE6\x89\x93\x00\xE6\x89\x9D\x00\xE6\x8A\x95\x00\xE6"
-     "\x8A\xB1"
-     "\x00\xE6\x8B\x89\x00\xE6\x8B\x8F\x00\xE6\x8B\x93\x00\xE6\x8B\x94\x00\xE6"
-     "\x8B\xBC"
-     "\x00\xE6\x8B\xBE\x00\xE6\x8C\x87\x00\xE6\x8C\xBD\x00\xE6\x8D\x90\x00\xE6"
-     "\x8D\x95"
-     "\x00\xE6\x8D\xA8\x00\xE6\x8D\xBB\x00\xE6\x8E\x83\x00\xE6\x8E\xA0\x00\xE6"
-     "\x8E\xA9"
-     "\x00\xE6\x8F\x84\x00\xE6\x8F\x85\x00\xE6\x8F\xA4\x00\xE6\x90\x9C\x00\xE6"
-     "\x90\xA2"
-     "\x00\xE6\x91\x92\x00\xE6\x91\xA9\x00\xE6\x91\xB7\x00\xE6\x91\xBE\x00\xE6"
-     "\x92\x9A"
-     "\x00\xE6\x92\x9D\x00\xE6\x93\x84\x00\xE6\x94\xAF\x00\xE6\x94\xB4\x00\xE6"
-     "\x95\x8F"
-     "\x00\xE6\x95\x96\x00\xE6\x95\xAC\x00\xE6\x95\xB8\x00\xE6\x96\x87\x00\xE6"
-     "\x96\x97"
-     "\x00\xE6\x96\x99\x00\xE6\x96\xA4\x00\xE6\x96\xB0\x00\xE6\x96\xB9\x00\xE6"
-     "\x97\x85"
-     "\x00\xE6\x97\xA0\x00\xE6\x97\xA2\x00\xE6\x97\xA3\x00\xE6\x97\xA5\x00\xE6"
-     "\x98\x8E"
-     "\xE6\xB2\xBB\x00\xE6\x98\x93\x00\xE6\x98\xA0\x00\xE6\x98\xAD\xE5\x92\x8C"
-     "\x00\xE6"
-     "\x99\x89\x00\xE6\x99\xB4\x00\xE6\x9A\x88\x00\xE6\x9A\x91\x00\xE6\x9A\x9C"
-     "\x00\xE6"
-     "\x9A\xB4\x00\xE6\x9B\x86\x00\xE6\x9B\xB0\x00\xE6\x9B\xB4\x00\xE6\x9B\xB8"
-     "\x00\xE6"
-     "\x9C\x80\x00\xE6\x9C\x88\x00\xE6\x9C\x89\x00\xE6\x9C\x97\x00\xE6\x9C\x9B"
-     "\x00\xE6"
-     "\x9C\xA1\x00\xE6\x9C\xA8\x00\xE6\x9D\x8E\x00\xE6\x9D\x93\x00\xE6\x9D\x96"
-     "\x00\xE6"
-     "\x9D\x9E\x00\xE6\x9D\xBB\x00\xE6\x9E\x85\x00\xE6\x9E\x97\x00\xE6\x9F\xB3"
-     "\x00\xE6"
-     "\x9F\xBA\x00\xE6\xA0\x97\x00\xE6\xA0\x9F\x00\xE6\xA0\xAA\x00\xE6\xA0\xAA"
-     "\xE5\xBC"
-     "\x8F\xE4\xBC\x9A\xE7\xA4\xBE\x00\xE6\xA1\x92\x00\xE6\xA2\x81\x00\xE6\xA2"
-     "\x85\x00"
-     "\xE6\xA2\x8E\x00\xE6\xA2\xA8\x00\xE6\xA4\x94\x00\xE6\xA5\x82\x00\xE6\xA6"
-     "\xA3\x00"
-     "\xE6\xA7\xAA\x00\xE6\xA8\x82\x00\xE6\xA8\x93\x00\xE6\xAA\xA8\x00\xE6\xAB"
-     "\x93\x00"
-     "\xE6\xAB\x9B\x00\xE6\xAC\x84\x00\xE6\xAC\xA0\x00\xE6\xAC\xA1\x00\xE6\xAD"
-     "\x94\x00"
-     "\xE6\xAD\xA2\x00\xE6\xAD\xA3\x00\xE6\xAD\xB2\x00\xE6\xAD\xB7\x00\xE6\xAD"
-     "\xB9\x00"
-     "\xE6\xAE\x9F\x00\xE6\xAE\xAE\x00\xE6\xAE\xB3\x00\xE6\xAE\xBA\x00\xE6\xAE"
-     "\xBB\x00"
-     "\xE6\xAF\x8B\x00\xE6\xAF\x8D\x00\xE6\xAF\x94\x00\xE6\xAF\x9B\x00\xE6\xB0"
-     "\x8F\x00"
-     "\xE6\xB0\x94\x00\xE6\xB0\xB4\x00\xE6\xB1\x8E\x00\xE6\xB1\xA7\x00\xE6\xB2"
-     "\x88\x00"
-     "\xE6\xB2\xBF\x00\xE6\xB3\x8C\x00\xE6\xB3\x8D\x00\xE6\xB3\xA5\x00\xE6\xB3"
-     "\xA8\x00"
-     "\xE6\xB4\x96\x00\xE6\xB4\x9B\x00\xE6\xB4\x9E\x00\xE6\xB4\xB4\x00\xE6\xB4"
-     "\xBE\x00"
-     "\xE6\xB5\x81\x00\xE6\xB5\xA9\x00\xE6\xB5\xAA\x00\xE6\xB5\xB7\x00\xE6\xB5"
-     "\xB8\x00"
-     "\xE6\xB6\x85\x00\xE6\xB7\x8B\x00\xE6\xB7\x9A\x00\xE6\xB7\xAA\x00\xE6\xB7"
-     "\xB9\x00"
-     "\xE6\xB8\x9A\x00\xE6\xB8\xAF\x00\xE6\xB9\xAE\x00\xE6\xBA\x80\x00\xE6\xBA"
-     "\x9C\x00"
-     "\xE6\xBA\xBA\x00\xE6\xBB\x87\x00\xE6\xBB\x8B\x00\xE6\xBB\x91\x00\xE6\xBB"
-     "\x9B\x00"
-     "\xE6\xBC\x8F\x00\xE6\xBC\x94\x00\xE6\xBC\xA2\x00\xE6\xBC\xA3\x00\xE6\xBD"
-     "\xAE\x00"
-     "\xE6\xBF\x86\x00\xE6\xBF\xAB\x00\xE6\xBF\xBE\x00\xE7\x80\x9B\x00\xE7\x80"
-     "\x9E\x00"
-     "\xE7\x80\xB9\x00\xE7\x81\x8A\x00\xE7\x81\xAB\x00\xE7\x81\xB0\x00\xE7\x81"
-     "\xB7\x00"
-     "\xE7\x81\xBD\x00\xE7\x82\x99\x00\xE7\x82\xAD\x00\xE7\x83\x88\x00\xE7\x83"
-     "\x99\x00"
-     "\xE7\x84\xA1\x00\xE7\x85\x85\x00\xE7\x85\x89\x00\xE7\x85\xAE\x00\xE7\x86"
-     "\x9C\x00"
-     "\xE7\x87\x8E\x00\xE7\x87\x90\x00\xE7\x88\x90\x00\xE7\x88\x9B\x00\xE7\x88"
-     "\xA8\x00"
-     "\xE7\x88\xAA\x00\xE7\x88\xAB\x00\xE7\x88\xB5\x00\xE7\x88\xB6\x00\xE7\x88"
-     "\xBB\x00"
-     "\xE7\x88\xBF\x00\xE7\x89\x87\x00\xE7\x89\x90\x00\xE7\x89\x99\x00\xE7\x89"
-     "\x9B\x00"
-     "\xE7\x89\xA2\x00\xE7\x89\xB9\x00\xE7\x8A\x80\x00\xE7\x8A\x95\x00\xE7\x8A"
-     "\xAC\x00"
-     "\xE7\x8A\xAF\x00\xE7\x8B\x80\x00\xE7\x8B\xBC\x00\xE7\x8C\xAA\x00\xE7\x8D"
-     "\xB5\x00"
-     "\xE7\x8D\xBA\x00\xE7\x8E\x84\x00\xE7\x8E\x87\x00\xE7\x8E\x89\x00\xE7\x8E"
-     "\x8B\x00"
-     "\xE7\x8E\xA5\x00\xE7\x8E\xB2\x00\xE7\x8F\x9E\x00\xE7\x90\x86\x00\xE7\x90"
-     "\x89\x00"
-     "\xE7\x90\xA2\x00\xE7\x91\x87\x00\xE7\x91\x9C\x00\xE7\x91\xA9\x00\xE7\x91"
-     "\xB1\x00"
-     "\xE7\x92\x85\x00\xE7\x92\x89\x00\xE7\x92\x98\x00\xE7\x93\x8A\x00\xE7\x93"
-     "\x9C\x00"
-     "\xE7\x93\xA6\x00\xE7\x94\x86\x00\xE7\x94\x98\x00\xE7\x94\x9F\x00\xE7\x94"
-     "\xA4\x00"
-     "\xE7\x94\xA8\x00\xE7\x94\xB0\x00\xE7\x94\xB2\x00\xE7\x94\xB3\x00\xE7\x94"
-     "\xB7\x00"
-     "\xE7\x94\xBB\x00\xE7\x94\xBE\x00\xE7\x95\x99\x00\xE7\x95\xA5\x00\xE7\x95"
-     "\xB0\x00"
-     "\xE7\x96\x8B\x00\xE7\x96\x92\x00\xE7\x97\xA2\x00\xE7\x98\x90\x00\xE7\x98"
-     "\x9D\x00"
-     "\xE7\x98\x9F\x00\xE7\x99\x82\x00\xE7\x99\xA9\x00\xE7\x99\xB6\x00\xE7\x99"
-     "\xBD\x00"
-     "\xE7\x9A\xAE\x00\xE7\x9A\xBF\x00\xE7\x9B\x8A\x00\xE7\x9B\x9B\x00\xE7\x9B"
-     "\xA3\x00"
-     "\xE7\x9B\xA7\x00\xE7\x9B\xAE\x00\xE7\x9B\xB4\x00\xE7\x9C\x81\x00\xE7\x9C"
-     "\x9E\x00"
-     "\xE7\x9C\x9F\x00\xE7\x9D\x80\x00\xE7\x9D\x8A\x00\xE7\x9E\x8B\x00\xE7\x9E"
-     "\xA7\x00"
-     "\xE7\x9F\x9B\x00\xE7\x9F\xA2\x00\xE7\x9F\xB3\x00\xE7\xA1\x8E\x00\xE7\xA1"
-     "\xAB\x00"
-     "\xE7\xA2\x8C\x00\xE7\xA2\x91\x00\xE7\xA3\x8A\x00\xE7\xA3\x8C\x00\xE7\xA3"
-     "\xBB\x00"
-     "\xE7\xA4\xAA\x00\xE7\xA4\xBA\x00\xE7\xA4\xBC\x00\xE7\xA4\xBE\x00\xE7\xA5"
-     "\x88\x00"
-     "\xE7\xA5\x89\x00\xE7\xA5\x90\x00\xE7\xA5\x96\x00\xE7\xA5\x9D\x00\xE7\xA5"
-     "\x9E\x00"
-     "\xE7\xA5\xA5\x00\xE7\xA5\xBF\x00\xE7\xA6\x81\x00\xE7\xA6\x8D\x00\xE7\xA6"
-     "\x8E\x00"
-     "\xE7\xA6\x8F\x00\xE7\xA6\xAE\x00\xE7\xA6\xB8\x00\xE7\xA6\xBE\x00\xE7\xA7"
-     "\x8A\x00"
-     "\xE7\xA7\x98\x00\xE7\xA7\xAB\x00\xE7\xA8\x9C\x00\xE7\xA9\x80\x00\xE7\xA9"
-     "\x8A\x00"
-     "\xE7\xA9\x8F\x00\xE7\xA9\xB4\x00\xE7\xA9\xBA\x00\xE7\xAA\x81\x00\xE7\xAA"
-     "\xB1\x00"
-     "\xE7\xAB\x8B\x00\xE7\xAB\xAE\x00\xE7\xAB\xB9\x00\xE7\xAC\xA0\x00\xE7\xAE"
-     "\x8F\x00"
-     "\xE7\xAF\x80\x00\xE7\xAF\x86\x00\xE7\xAF\x89\x00\xE7\xB0\xBE\x00\xE7\xB1"
-     "\xA0\x00"
-     "\xE7\xB1\xB3\x00\xE7\xB1\xBB\x00\xE7\xB2\x92\x00\xE7\xB2\xBE\x00\xE7\xB3"
-     "\x92\x00"
-     "\xE7\xB3\x96\x00\xE7\xB3\xA3\x00\xE7\xB3\xA7\x00\xE7\xB3\xA8\x00\xE7\xB3"
-     "\xB8\x00"
-     "\xE7\xB4\x80\x00\xE7\xB4\x90\x00\xE7\xB4\xA2\x00\xE7\xB4\xAF\x00\xE7\xB5"
-     "\x82\x00"
-     "\xE7\xB5\x9B\x00\xE7\xB5\xA3\x00\xE7\xB6\xA0\x00\xE7\xB6\xBE\x00\xE7\xB7"
-     "\x87\x00"
-     "\xE7\xB7\xB4\x00\xE7\xB8\x82\x00\xE7\xB8\x89\x00\xE7\xB8\xB7\x00\xE7\xB9"
-     "\x81\x00"
-     "\xE7\xB9\x85\x00\xE7\xBC\xB6\x00\xE7\xBC\xBE\x00\xE7\xBD\x91\x00\xE7\xBD"
-     "\xB2\x00"
-     "\xE7\xBD\xB9\x00\xE7\xBD\xBA\x00\xE7\xBE\x85\x00\xE7\xBE\x8A\x00\xE7\xBE"
-     "\x95\x00"
-     "\xE7\xBE\x9A\x00\xE7\xBE\xBD\x00\xE7\xBF\xBA\x00\xE8\x80\x81\x00\xE8\x80"
-     "\x85\x00"
-     "\xE8\x80\x8C\x00\xE8\x80\x92\x00\xE8\x80\xB3\x00\xE8\x81\x86\x00\xE8\x81"
-     "\xA0\x00"
-     "\xE8\x81\xAF\x00\xE8\x81\xB0\x00\xE8\x81\xBE\x00\xE8\x81\xBF\x00\xE8\x82"
-     "\x89\x00"
-     "\xE8\x82\x8B\x00\xE8\x82\xAD\x00\xE8\x82\xB2\x00\xE8\x84\x83\x00\xE8\x84"
-     "\xBE\x00"
-     "\xE8\x87\x98\x00\xE8\x87\xA3\x00\xE8\x87\xA8\x00\xE8\x87\xAA\x00\xE8\x87"
-     "\xAD\x00"
-     "\xE8\x87\xB3\x00\xE8\x87\xBC\x00\xE8\x88\x81\x00\xE8\x88\x84\x00\xE8\x88"
-     "\x8C\x00"
-     "\xE8\x88\x98\x00\xE8\x88\x9B\x00\xE8\x88\x9F\x00\xE8\x89\xAE\x00\xE8\x89"
-     "\xAF\x00"
-     "\xE8\x89\xB2\x00\xE8\x89\xB8\x00\xE8\x89\xB9\x00\xE8\x8A\x8B\x00\xE8\x8A"
-     "\x91\x00"
-     "\xE8\x8A\x9D\x00\xE8\x8A\xB1\x00\xE8\x8A\xB3\x00\xE8\x8A\xBD\x00\xE8\x8B"
-     "\xA5\x00"
-     "\xE8\x8B\xA6\x00\xE8\x8C\x9D\x00\xE8\x8C\xA3\x00\xE8\x8C\xB6\x00\xE8\x8D"
-     "\x92\x00"
-     "\xE8\x8D\x93\x00\xE8\x8D\xA3\x00\xE8\x8E\xAD\x00\xE8\x8E\xBD\x00\xE8\x8F"
-     "\x89\x00"
-     "\xE8\x8F\x8A\x00\xE8\x8F\x8C\x00\xE8\x8F\x9C\x00\xE8\x8F\xA7\x00\xE8\x8F"
-     "\xAF\x00"
-     "\xE8\x8F\xB1\x00\xE8\x90\xBD\x00\xE8\x91\x89\x00\xE8\x91\x97\x00\xE8\x93"
-     "\xAE\x00"
-     "\xE8\x93\xB1\x00\xE8\x93\xB3\x00\xE8\x93\xBC\x00\xE8\x94\x96\x00\xE8\x95"
-     "\xA4\x00"
-     "\xE8\x97\x8D\x00\xE8\x97\xBA\x00\xE8\x98\x86\x00\xE8\x98\x92\x00\xE8\x98"
-     "\xAD\x00"
-     "\xE8\x98\xBF\x00\xE8\x99\x8D\x00\xE8\x99\x90\x00\xE8\x99\x9C\x00\xE8\x99"
-     "\xA7\x00"
-     "\xE8\x99\xA9\x00\xE8\x99\xAB\x00\xE8\x9A\x88\x00\xE8\x9A\xA9\x00\xE8\x9B"
-     "\xA2\x00"
-     "\xE8\x9C\x8E\x00\xE8\x9C\xA8\x00\xE8\x9D\xAB\x00\xE8\x9D\xB9\x00\xE8\x9E"
-     "\x86\x00"
-     "\xE8\x9E\xBA\x00\xE8\x9F\xA1\x00\xE8\xA0\x81\x00\xE8\xA0\x9F\x00\xE8\xA1"
-     "\x80\x00"
-     "\xE8\xA1\x8C\x00\xE8\xA1\xA0\x00\xE8\xA1\xA3\x00\xE8\xA3\x82\x00\xE8\xA3"
-     "\x8F\x00"
-     "\xE8\xA3\x97\x00\xE8\xA3\x9E\x00\xE8\xA3\xA1\x00\xE8\xA3\xB8\x00\xE8\xA3"
-     "\xBA\x00"
-     "\xE8\xA4\x90\x00\xE8\xA5\x81\x00\xE8\xA5\xA4\x00\xE8\xA5\xBE\x00\xE8\xA6"
-     "\x86\x00"
-     "\xE8\xA6\x8B\x00\xE8\xA6\x96\x00\xE8\xA7\x92\x00\xE8\xA7\xA3\x00\xE8\xA8"
-     "\x80\x00"
-     "\xE8\xAA\xA0\x00\xE8\xAA\xAA\x00\xE8\xAA\xBF\x00\xE8\xAB\x8B\x00\xE8\xAB"
-     "\x92\x00"
-     "\xE8\xAB\x96\x00\xE8\xAB\xAD\x00\xE8\xAB\xB8\x00\xE8\xAB\xBE\x00\xE8\xAC"
-     "\x81\x00"
-     "\xE8\xAC\xB9\x00\xE8\xAD\x98\x00\xE8\xAE\x80\x00\xE8\xAE\x8A\x00\xE8\xB0"
-     "\xB7\x00"
-     "\xE8\xB1\x86\x00\xE8\xB1\x88\x00\xE8\xB1\x95\x00\xE8\xB1\xB8\x00\xE8\xB2"
-     "\x9D\x00"
-     "\xE8\xB2\xA1\x00\xE8\xB2\xA9\x00\xE8\xB2\xAB\x00\xE8\xB3\x81\x00\xE8\xB3"
-     "\x82\x00"
-     "\xE8\xB3\x87\x00\xE8\xB3\x88\x00\xE8\xB3\x93\x00\xE8\xB4\x88\x00\xE8\xB4"
-     "\x9B\x00"
-     "\xE8\xB5\xA4\x00\xE8\xB5\xB0\x00\xE8\xB5\xB7\x00\xE8\xB6\xB3\x00\xE8\xB6"
-     "\xBC\x00"
-     "\xE8\xB7\x8B\x00\xE8\xB7\xAF\x00\xE8\xB7\xB0\x00\xE8\xBA\xAB\x00\xE8\xBB"
-     "\x8A\x00"
-     "\xE8\xBB\x94\x00\xE8\xBC\xA6\x00\xE8\xBC\xAA\x00\xE8\xBC\xB8\x00\xE8\xBC"
-     "\xBB\x00"
-     "\xE8\xBD\xA2\x00\xE8\xBE\x9B\x00\xE8\xBE\x9E\x00\xE8\xBE\xB0\x00\xE8\xBE"
-     "\xB5\x00"
-     "\xE8\xBE\xB6\x00\xE9\x80\xA3\x00\xE9\x80\xB8\x00\xE9\x81\x8A\x00\xE9\x81"
-     "\xA9\x00"
-     "\xE9\x81\xB2\x00\xE9\x81\xBC\x00\xE9\x82\x8F\x00\xE9\x82\x91\x00\xE9\x82"
-     "\x94\x00"
-     "\xE9\x83\x8E\x00\xE9\x83\x9E\x00\xE9\x83\xB1\x00\xE9\x83\xBD\x00\xE9\x84"
-     "\x91\x00"
-     "\xE9\x84\x9B\x00\xE9\x85\x89\x00\xE9\x85\x8D\x00\xE9\x85\xAA\x00\xE9\x86"
-     "\x99\x00"
-     "\xE9\x86\xB4\x00\xE9\x87\x86\x00\xE9\x87\x8C\x00\xE9\x87\x8F\x00\xE9\x87"
-     "\x91\x00"
-     "\xE9\x88\xB4\x00\xE9\x88\xB8\x00\xE9\x89\xB6\x00\xE9\x89\xBC\x00\xE9\x8B"
-     "\x97\x00"
-     "\xE9\x8B\x98\x00\xE9\x8C\x84\x00\xE9\x8D\x8A\x00\xE9\x8F\xB9\x00\xE9\x90"
-     "\x95\x00"
-     "\xE9\x95\xB7\x00\xE9\x96\x80\x00\xE9\x96\x8B\x00\xE9\x96\xAD\x00\xE9\x96"
-     "\xB7\x00"
-     "\xE9\x98\x9C\x00\xE9\x98\xAE\x00\xE9\x99\x8B\x00\xE9\x99\x8D\x00\xE9\x99"
-     "\xB5\x00"
-     "\xE9\x99\xB8\x00\xE9\x99\xBC\x00\xE9\x9A\x86\x00\xE9\x9A\xA3\x00\xE9\x9A"
-     "\xB6\x00"
-     "\xE9\x9A\xB7\x00\xE9\x9A\xB8\x00\xE9\x9A\xB9\x00\xE9\x9B\x83\x00\xE9\x9B"
-     "\xA2\x00"
-     "\xE9\x9B\xA3\x00\xE9\x9B\xA8\x00\xE9\x9B\xB6\x00\xE9\x9B\xB7\x00\xE9\x9C"
-     "\xA3\x00"
-     "\xE9\x9C\xB2\x00\xE9\x9D\x88\x00\xE9\x9D\x91\x00\xE9\x9D\x96\x00\xE9\x9D"
-     "\x9E\x00"
-     "\xE9\x9D\xA2\x00\xE9\x9D\xA9\x00\xE9\x9F\x8B\x00\xE9\x9F\x9B\x00\xE9\x9F"
-     "\xA0\x00"
-     "\xE9\x9F\xAD\x00\xE9\x9F\xB3\x00\xE9\x9F\xBF\x00\xE9\xA0\x81\x00\xE9\xA0"
-     "\x85\x00"
-     "\xE9\xA0\x8B\x00\xE9\xA0\x98\x00\xE9\xA0\xA9\x00\xE9\xA0\xBB\x00\xE9\xA1"
-     "\x9E\x00"
-     "\xE9\xA2\xA8\x00\xE9\xA3\x9B\x00\xE9\xA3\x9F\x00\xE9\xA3\xA2\x00\xE9\xA3"
-     "\xAF\x00"
-     "\xE9\xA3\xBC\x00\xE9\xA4\xA8\x00\xE9\xA4\xA9\x00\xE9\xA6\x96\x00\xE9\xA6"
-     "\x99\x00"
-     "\xE9\xA6\xA7\x00\xE9\xA6\xAC\x00\xE9\xA7\x82\x00\xE9\xA7\xB1\x00\xE9\xA7"
-     "\xBE\x00"
-     "\xE9\xA9\xAA\x00\xE9\xAA\xA8\x00\xE9\xAB\x98\x00\xE9\xAB\x9F\x00\xE9\xAC"
-     "\x92\x00"
-     "\xE9\xAC\xA5\x00\xE9\xAC\xAF\x00\xE9\xAC\xB2\x00\xE9\xAC\xBC\x00\xE9\xAD"
-     "\x9A\x00"
-     "\xE9\xAD\xAF\x00\xE9\xB1\x80\x00\xE9\xB1\x97\x00\xE9\xB3\xA5\x00\xE9\xB3"
-     "\xBD\x00"
-     "\xE9\xB5\xA7\x00\xE9\xB6\xB4\x00\xE9\xB7\xBA\x00\xE9\xB8\x9E\x00\xE9\xB9"
-     "\xB5\x00"
-     "\xE9\xB9\xBF\x00\xE9\xBA\x97\x00\xE9\xBA\x9F\x00\xE9\xBA\xA5\x00\xE9\xBA"
-     "\xBB\x00"
-     "\xE9\xBB\x83\x00\xE9\xBB\x8D\x00\xE9\xBB\x8E\x00\xE9\xBB\x91\x00\xE9\xBB"
-     "\xB9\x00"
-     "\xE9\xBB\xBD\x00\xE9\xBB\xBE\x00\xE9\xBC\x85\x00\xE9\xBC\x8E\x00\xE9\xBC"
-     "\x8F\x00"
-     "\xE9\xBC\x93\x00\xE9\xBC\x96\x00\xE9\xBC\xA0\x00\xE9\xBC\xBB\x00\xE9\xBD"
-     "\x83\x00"
-     "\xE9\xBD\x8A\x00\xE9\xBD\x92\x00\xE9\xBE\x8D\x00\xE9\xBE\x8E\x00\xE9\xBE"
-     "\x9C\x00"
-     "\xE9\xBE\x9F\x00\xE9\xBE\xA0\x00\xEA\x99\x91\x00\xEA\x9A\x89\x00\xEA\x9C"
-     "\xA7\x00"
-     "\xEA\x9D\xAF\x00\xEA\x9E\x8E\x00\xEA\xAC\xB7\x00\xEA\xAD\x92\x00\xEA\xAD"
-     "\xA6\x00"
-     "\xEA\xAD\xA7\x00\xEA\xB0\x80\x00\xEA\xB0\x81\x00\xEA\xB0\x82\x00\xEA\xB0"
-     "\x83\x00"
-     "\xEA\xB0\x84\x00\xEA\xB0\x85\x00\xEA\xB0\x86\x00\xEA\xB0\x87\x00\xEA\xB0"
-     "\x88\x00"
-     "\xEA\xB0\x89\x00\xEA\xB0\x8A\x00\xEA\xB0\x8B\x00\xEA\xB0\x8C\x00\xEA\xB0"
-     "\x8D\x00"
-     "\xEA\xB0\x8E\x00\xEA\xB0\x8F\x00\xEA\xB0\x90\x00\xEA\xB0\x91\x00\xEA\xB0"
-     "\x92\x00"
-     "\xEA\xB0\x93\x00\xEA\xB0\x94\x00\xEA\xB0\x95\x00\xEA\xB0\x96\x00\xEA\xB0"
-     "\x97\x00"
-     "\xEA\xB0\x98\x00\xEA\xB0\x99\x00\xEA\xB0\x9A\x00\xEA\xB0\x9B\x00\xEA\xB0"
-     "\x9C\x00"
-     "\xEA\xB0\x9D\x00\xEA\xB0\x9E\x00\xEA\xB0\x9F\x00\xEA\xB0\xA0\x00\xEA\xB0"
-     "\xA1\x00"
-     "\xEA\xB0\xA2\x00\xEA\xB0\xA3\x00\xEA\xB0\xA4\x00\xEA\xB0\xA5\x00\xEA\xB0"
-     "\xA6\x00"
-     "\xEA\xB0\xA7\x00\xEA\xB0\xA8\x00\xEA\xB0\xA9\x00\xEA\xB0\xAA\x00\xEA\xB0"
-     "\xAB\x00"
-     "\xEA\xB0\xAC\x00\xEA\xB0\xAD\x00\xEA\xB0\xAE\x00\xEA\xB0\xAF\x00\xEA\xB0"
-     "\xB0\x00"
-     "\xEA\xB0\xB1\x00\xEA\xB0\xB2\x00\xEA\xB0\xB3\x00\xEA\xB0\xB4\x00\xEA\xB0"
-     "\xB5\x00"
-     "\xEA\xB0\xB6\x00\xEA\xB0\xB7\x00\xEA\xB0\xB8\x00\xEA\xB0\xB9\x00\xEA\xB0"
-     "\xBA\x00"
-     "\xEA\xB0\xBB\x00\xEA\xB0\xBC\x00\xEA\xB0\xBD\x00\xEA\xB0\xBE\x00\xEA\xB0"
-     "\xBF\x00"
-     "\xEA\xB1\x80\x00\xEA\xB1\x81\x00\xEA\xB1\x82\x00\xEA\xB1\x83\x00\xEA\xB1"
-     "\x84\x00"
-     "\xEA\xB1\x85\x00\xEA\xB1\x86\x00\xEA\xB1\x87\x00\xEA\xB1\x88\x00\xEA\xB1"
-     "\x89\x00"
-     "\xEA\xB1\x8A\x00\xEA\xB1\x8B\x00\xEA\xB1\x8C\x00\xEA\xB1\x8D\x00\xEA\xB1"
-     "\x8E\x00"
-     "\xEA\xB1\x8F\x00\xEA\xB1\x90\x00\xEA\xB1\x91\x00\xEA\xB1\x92\x00\xEA\xB1"
-     "\x93\x00"
-     "\xEA\xB1\x94\x00\xEA\xB1\x95\x00\xEA\xB1\x96\x00\xEA\xB1\x97\x00\xEA\xB1"
-     "\x98\x00"
-     "\xEA\xB1\x99\x00\xEA\xB1\x9A\x00\xEA\xB1\x9B\x00\xEA\xB1\x9C\x00\xEA\xB1"
-     "\x9D\x00"
-     "\xEA\xB1\x9E\x00\xEA\xB1\x9F\x00\xEA\xB1\xA0\x00\xEA\xB1\xA1\x00\xEA\xB1"
-     "\xA2\x00"
-     "\xEA\xB1\xA3\x00\xEA\xB1\xA4\x00\xEA\xB1\xA5\x00\xEA\xB1\xA6\x00\xEA\xB1"
-     "\xA7\x00"
-     "\xEA\xB1\xA8\x00\xEA\xB1\xA9\x00\xEA\xB1\xAA\x00\xEA\xB1\xAB\x00\xEA\xB1"
-     "\xAC\x00"
-     "\xEA\xB1\xAD\x00\xEA\xB1\xAE\x00\xEA\xB1\xAF\x00\xEA\xB1\xB0\x00\xEA\xB1"
-     "\xB1\x00"
-     "\xEA\xB1\xB2\x00\xEA\xB1\xB3\x00\xEA\xB1\xB4\x00\xEA\xB1\xB5\x00\xEA\xB1"
-     "\xB6\x00"
-     "\xEA\xB1\xB7\x00\xEA\xB1\xB8\x00\xEA\xB1\xB9\x00\xEA\xB1\xBA\x00\xEA\xB1"
-     "\xBB\x00"
-     "\xEA\xB1\xBC\x00\xEA\xB1\xBD\x00\xEA\xB1\xBE\x00\xEA\xB1\xBF\x00\xEA\xB2"
-     "\x80\x00"
-     "\xEA\xB2\x81\x00\xEA\xB2\x82\x00\xEA\xB2\x83\x00\xEA\xB2\x84\x00\xEA\xB2"
-     "\x85\x00"
-     "\xEA\xB2\x86\x00\xEA\xB2\x87\x00\xEA\xB2\x88\x00\xEA\xB2\x89\x00\xEA\xB2"
-     "\x8A\x00"
-     "\xEA\xB2\x8B\x00\xEA\xB2\x8C\x00\xEA\xB2\x8D\x00\xEA\xB2\x8E\x00\xEA\xB2"
-     "\x8F\x00"
-     "\xEA\xB2\x90\x00\xEA\xB2\x91\x00\xEA\xB2\x92\x00\xEA\xB2\x93\x00\xEA\xB2"
-     "\x94\x00"
-     "\xEA\xB2\x95\x00\xEA\xB2\x96\x00\xEA\xB2\x97\x00\xEA\xB2\x98\x00\xEA\xB2"
-     "\x99\x00"
-     "\xEA\xB2\x9A\x00\xEA\xB2\x9B\x00\xEA\xB2\x9C\x00\xEA\xB2\x9D\x00\xEA\xB2"
-     "\x9E\x00"
-     "\xEA\xB2\x9F\x00\xEA\xB2\xA0\x00\xEA\xB2\xA1\x00\xEA\xB2\xA2\x00\xEA\xB2"
-     "\xA3\x00"
-     "\xEA\xB2\xA4\x00\xEA\xB2\xA5\x00\xEA\xB2\xA6\x00\xEA\xB2\xA7\x00\xEA\xB2"
-     "\xA8\x00"
-     "\xEA\xB2\xA9\x00\xEA\xB2\xAA\x00\xEA\xB2\xAB\x00\xEA\xB2\xAC\x00\xEA\xB2"
-     "\xAD\x00"
-     "\xEA\xB2\xAE\x00\xEA\xB2\xAF\x00\xEA\xB2\xB0\x00\xEA\xB2\xB1\x00\xEA\xB2"
-     "\xB2\x00"
-     "\xEA\xB2\xB3\x00\xEA\xB2\xB4\x00\xEA\xB2\xB5\x00\xEA\xB2\xB6\x00\xEA\xB2"
-     "\xB7\x00"
-     "\xEA\xB2\xB8\x00\xEA\xB2\xB9\x00\xEA\xB2\xBA\x00\xEA\xB2\xBB\x00\xEA\xB2"
-     "\xBC\x00"
-     "\xEA\xB2\xBD\x00\xEA\xB2\xBE\x00\xEA\xB2\xBF\x00\xEA\xB3\x80\x00\xEA\xB3"
-     "\x81\x00"
-     "\xEA\xB3\x82\x00\xEA\xB3\x83\x00\xEA\xB3\x84\x00\xEA\xB3\x85\x00\xEA\xB3"
-     "\x86\x00"
-     "\xEA\xB3\x87\x00\xEA\xB3\x88\x00\xEA\xB3\x89\x00\xEA\xB3\x8A\x00\xEA\xB3"
-     "\x8B\x00"
-     "\xEA\xB3\x8C\x00\xEA\xB3\x8D\x00\xEA\xB3\x8E\x00\xEA\xB3\x8F\x00\xEA\xB3"
-     "\x90\x00"
-     "\xEA\xB3\x91\x00\xEA\xB3\x92\x00\xEA\xB3\x93\x00\xEA\xB3\x94\x00\xEA\xB3"
-     "\x95\x00"
-     "\xEA\xB3\x96\x00\xEA\xB3\x97\x00\xEA\xB3\x98\x00\xEA\xB3\x99\x00\xEA\xB3"
-     "\x9A\x00"
-     "\xEA\xB3\x9B\x00\xEA\xB3\x9C\x00\xEA\xB3\x9D\x00\xEA\xB3\x9E\x00\xEA\xB3"
-     "\x9F\x00"
-     "\xEA\xB3\xA0\x00\xEA\xB3\xA1\x00\xEA\xB3\xA2\x00\xEA\xB3\xA3\x00\xEA\xB3"
-     "\xA4\x00"
-     "\xEA\xB3\xA5\x00\xEA\xB3\xA6\x00\xEA\xB3\xA7\x00\xEA\xB3\xA8\x00\xEA\xB3"
-     "\xA9\x00"
-     "\xEA\xB3\xAA\x00\xEA\xB3\xAB\x00\xEA\xB3\xAC\x00\xEA\xB3\xAD\x00\xEA\xB3"
-     "\xAE\x00"
-     "\xEA\xB3\xAF\x00\xEA\xB3\xB0\x00\xEA\xB3\xB1\x00\xEA\xB3\xB2\x00\xEA\xB3"
-     "\xB3\x00"
-     "\xEA\xB3\xB4\x00\xEA\xB3\xB5\x00\xEA\xB3\xB6\x00\xEA\xB3\xB7\x00\xEA\xB3"
-     "\xB8\x00"
-     "\xEA\xB3\xB9\x00\xEA\xB3\xBA\x00\xEA\xB3\xBB\x00\xEA\xB3\xBC\x00\xEA\xB3"
-     "\xBD\x00"
-     "\xEA\xB3\xBE\x00\xEA\xB3\xBF\x00\xEA\xB4\x80\x00\xEA\xB4\x81\x00\xEA\xB4"
-     "\x82\x00"
-     "\xEA\xB4\x83\x00\xEA\xB4\x84\x00\xEA\xB4\x85\x00\xEA\xB4\x86\x00\xEA\xB4"
-     "\x87\x00"
-     "\xEA\xB4\x88\x00\xEA\xB4\x89\x00\xEA\xB4\x8A\x00\xEA\xB4\x8B\x00\xEA\xB4"
-     "\x8C\x00"
-     "\xEA\xB4\x8D\x00\xEA\xB4\x8E\x00\xEA\xB4\x8F\x00\xEA\xB4\x90\x00\xEA\xB4"
-     "\x91\x00"
-     "\xEA\xB4\x92\x00\xEA\xB4\x93\x00\xEA\xB4\x94\x00\xEA\xB4\x95\x00\xEA\xB4"
-     "\x96\x00"
-     "\xEA\xB4\x97\x00\xEA\xB4\x98\x00\xEA\xB4\x99\x00\xEA\xB4\x9A\x00\xEA\xB4"
-     "\x9B\x00"
-     "\xEA\xB4\x9C\x00\xEA\xB4\x9D\x00\xEA\xB4\x9E\x00\xEA\xB4\x9F\x00\xEA\xB4"
-     "\xA0\x00"
-     "\xEA\xB4\xA1\x00\xEA\xB4\xA2\x00\xEA\xB4\xA3\x00\xEA\xB4\xA4\x00\xEA\xB4"
-     "\xA5\x00"
-     "\xEA\xB4\xA6\x00\xEA\xB4\xA7\x00\xEA\xB4\xA8\x00\xEA\xB4\xA9\x00\xEA\xB4"
-     "\xAA\x00"
-     "\xEA\xB4\xAB\x00\xEA\xB4\xAC\x00\xEA\xB4\xAD\x00\xEA\xB4\xAE\x00\xEA\xB4"
-     "\xAF\x00"
-     "\xEA\xB4\xB0\x00\xEA\xB4\xB1\x00\xEA\xB4\xB2\x00\xEA\xB4\xB3\x00\xEA\xB4"
-     "\xB4\x00"
-     "\xEA\xB4\xB5\x00\xEA\xB4\xB6\x00\xEA\xB4\xB7\x00\xEA\xB4\xB8\x00\xEA\xB4"
-     "\xB9\x00"
-     "\xEA\xB4\xBA\x00\xEA\xB4\xBB\x00\xEA\xB4\xBC\x00\xEA\xB4\xBD\x00\xEA\xB4"
-     "\xBE\x00"
-     "\xEA\xB4\xBF\x00\xEA\xB5\x80\x00\xEA\xB5\x81\x00\xEA\xB5\x82\x00\xEA\xB5"
-     "\x83\x00"
-     "\xEA\xB5\x84\x00\xEA\xB5\x85\x00\xEA\xB5\x86\x00\xEA\xB5\x87\x00\xEA\xB5"
-     "\x88\x00"
-     "\xEA\xB5\x89\x00\xEA\xB5\x8A\x00\xEA\xB5\x8B\x00\xEA\xB5\x8C\x00\xEA\xB5"
-     "\x8D\x00"
-     "\xEA\xB5\x8E\x00\xEA\xB5\x8F\x00\xEA\xB5\x90\x00\xEA\xB5\x91\x00\xEA\xB5"
-     "\x92\x00"
-     "\xEA\xB5\x93\x00\xEA\xB5\x94\x00\xEA\xB5\x95\x00\xEA\xB5\x96\x00\xEA\xB5"
-     "\x97\x00"
-     "\xEA\xB5\x98\x00\xEA\xB5\x99\x00\xEA\xB5\x9A\x00\xEA\xB5\x9B\x00\xEA\xB5"
-     "\x9C\x00"
-     "\xEA\xB5\x9D\x00\xEA\xB5\x9E\x00\xEA\xB5\x9F\x00\xEA\xB5\xA0\x00\xEA\xB5"
-     "\xA1\x00"
-     "\xEA\xB5\xA2\x00\xEA\xB5\xA3\x00\xEA\xB5\xA4\x00\xEA\xB5\xA5\x00\xEA\xB5"
-     "\xA6\x00"
-     "\xEA\xB5\xA7\x00\xEA\xB5\xA8\x00\xEA\xB5\xA9\x00\xEA\xB5\xAA\x00\xEA\xB5"
-     "\xAB\x00"
-     "\xEA\xB5\xAC\x00\xEA\xB5\xAD\x00\xEA\xB5\xAE\x00\xEA\xB5\xAF\x00\xEA\xB5"
-     "\xB0\x00"
-     "\xEA\xB5\xB1\x00\xEA\xB5\xB2\x00\xEA\xB5\xB3\x00\xEA\xB5\xB4\x00\xEA\xB5"
-     "\xB5\x00"
-     "\xEA\xB5\xB6\x00\xEA\xB5\xB7\x00\xEA\xB5\xB8\x00\xEA\xB5\xB9\x00\xEA\xB5"
-     "\xBA\x00"
-     "\xEA\xB5\xBB\x00\xEA\xB5\xBC\x00\xEA\xB5\xBD\x00\xEA\xB5\xBE\x00\xEA\xB5"
-     "\xBF\x00"
-     "\xEA\xB6\x80\x00\xEA\xB6\x81\x00\xEA\xB6\x82\x00\xEA\xB6\x83\x00\xEA\xB6"
-     "\x84\x00"
-     "\xEA\xB6\x85\x00\xEA\xB6\x86\x00\xEA\xB6\x87\x00\xEA\xB6\x88\x00\xEA\xB6"
-     "\x89\x00"
-     "\xEA\xB6\x8A\x00\xEA\xB6\x8B\x00\xEA\xB6\x8C\x00\xEA\xB6\x8D\x00\xEA\xB6"
-     "\x8E\x00"
-     "\xEA\xB6\x8F\x00\xEA\xB6\x90\x00\xEA\xB6\x91\x00\xEA\xB6\x92\x00\xEA\xB6"
-     "\x93\x00"
-     "\xEA\xB6\x94\x00\xEA\xB6\x95\x00\xEA\xB6\x96\x00\xEA\xB6\x97\x00\xEA\xB6"
-     "\x98\x00"
-     "\xEA\xB6\x99\x00\xEA\xB6\x9A\x00\xEA\xB6\x9B\x00\xEA\xB6\x9C\x00\xEA\xB6"
-     "\x9D\x00"
-     "\xEA\xB6\x9E\x00\xEA\xB6\x9F\x00\xEA\xB6\xA0\x00\xEA\xB6\xA1\x00\xEA\xB6"
-     "\xA2\x00"
-     "\xEA\xB6\xA3\x00\xEA\xB6\xA4\x00\xEA\xB6\xA5\x00\xEA\xB6\xA6\x00\xEA\xB6"
-     "\xA7\x00"
-     "\xEA\xB6\xA8\x00\xEA\xB6\xA9\x00\xEA\xB6\xAA\x00\xEA\xB6\xAB\x00\xEA\xB6"
-     "\xAC\x00"
-     "\xEA\xB6\xAD\x00\xEA\xB6\xAE\x00\xEA\xB6\xAF\x00\xEA\xB6\xB0\x00\xEA\xB6"
-     "\xB1\x00"
-     "\xEA\xB6\xB2\x00\xEA\xB6\xB3\x00\xEA\xB6\xB4\x00\xEA\xB6\xB5\x00\xEA\xB6"
-     "\xB6\x00"
-     "\xEA\xB6\xB7\x00\xEA\xB6\xB8\x00\xEA\xB6\xB9\x00\xEA\xB6\xBA\x00\xEA\xB6"
-     "\xBB\x00"
-     "\xEA\xB6\xBC\x00\xEA\xB6\xBD\x00\xEA\xB6\xBE\x00\xEA\xB6\xBF\x00\xEA\xB7"
-     "\x80\x00"
-     "\xEA\xB7\x81\x00\xEA\xB7\x82\x00\xEA\xB7\x83\x00\xEA\xB7\x84\x00\xEA\xB7"
-     "\x85\x00"
-     "\xEA\xB7\x86\x00\xEA\xB7\x87\x00\xEA\xB7\x88\x00\xEA\xB7\x89\x00\xEA\xB7"
-     "\x8A\x00"
-     "\xEA\xB7\x8B\x00\xEA\xB7\x8C\x00\xEA\xB7\x8D\x00\xEA\xB7\x8E\x00\xEA\xB7"
-     "\x8F\x00"
-     "\xEA\xB7\x90\x00\xEA\xB7\x91\x00\xEA\xB7\x92\x00\xEA\xB7\x93\x00\xEA\xB7"
-     "\x94\x00"
-     "\xEA\xB7\x95\x00\xEA\xB7\x96\x00\xEA\xB7\x97\x00\xEA\xB7\x98\x00\xEA\xB7"
-     "\x99\x00"
-     "\xEA\xB7\x9A\x00\xEA\xB7\x9B\x00\xEA\xB7\x9C\x00\xEA\xB7\x9D\x00\xEA\xB7"
-     "\x9E\x00"
-     "\xEA\xB7\x9F\x00\xEA\xB7\xA0\x00\xEA\xB7\xA1\x00\xEA\xB7\xA2\x00\xEA\xB7"
-     "\xA3\x00"
-     "\xEA\xB7\xA4\x00\xEA\xB7\xA5\x00\xEA\xB7\xA6\x00\xEA\xB7\xA7\x00\xEA\xB7"
-     "\xA8\x00"
-     "\xEA\xB7\xA9\x00\xEA\xB7\xAA\x00\xEA\xB7\xAB\x00\xEA\xB7\xAC\x00\xEA\xB7"
-     "\xAD\x00"
-     "\xEA\xB7\xAE\x00\xEA\xB7\xAF\x00\xEA\xB7\xB0\x00\xEA\xB7\xB1\x00\xEA\xB7"
-     "\xB2\x00"
-     "\xEA\xB7\xB3\x00\xEA\xB7\xB4\x00\xEA\xB7\xB5\x00\xEA\xB7\xB6\x00\xEA\xB7"
-     "\xB7\x00"
-     "\xEA\xB7\xB8\x00\xEA\xB7\xB9\x00\xEA\xB7\xBA\x00\xEA\xB7\xBB\x00\xEA\xB7"
-     "\xBC\x00"
-     "\xEA\xB7\xBD\x00\xEA\xB7\xBE\x00\xEA\xB7\xBF\x00\xEA\xB8\x80\x00\xEA\xB8"
-     "\x81\x00"
-     "\xEA\xB8\x82\x00\xEA\xB8\x83\x00\xEA\xB8\x84\x00\xEA\xB8\x85\x00\xEA\xB8"
-     "\x86\x00"
-     "\xEA\xB8\x87\x00\xEA\xB8\x88\x00\xEA\xB8\x89\x00\xEA\xB8\x8A\x00\xEA\xB8"
-     "\x8B\x00"
-     "\xEA\xB8\x8C\x00\xEA\xB8\x8D\x00\xEA\xB8\x8E\x00\xEA\xB8\x8F\x00\xEA\xB8"
-     "\x90\x00"
-     "\xEA\xB8\x91\x00\xEA\xB8\x92\x00\xEA\xB8\x93\x00\xEA\xB8\x94\x00\xEA\xB8"
-     "\x95\x00"
-     "\xEA\xB8\x96\x00\xEA\xB8\x97\x00\xEA\xB8\x98\x00\xEA\xB8\x99\x00\xEA\xB8"
-     "\x9A\x00"
-     "\xEA\xB8\x9B\x00\xEA\xB8\x9C\x00\xEA\xB8\x9D\x00\xEA\xB8\x9E\x00\xEA\xB8"
-     "\x9F\x00"
-     "\xEA\xB8\xA0\x00\xEA\xB8\xA1\x00\xEA\xB8\xA2\x00\xEA\xB8\xA3\x00\xEA\xB8"
-     "\xA4\x00"
-     "\xEA\xB8\xA5\x00\xEA\xB8\xA6\x00\xEA\xB8\xA7\x00\xEA\xB8\xA8\x00\xEA\xB8"
-     "\xA9\x00"
-     "\xEA\xB8\xAA\x00\xEA\xB8\xAB\x00\xEA\xB8\xAC\x00\xEA\xB8\xAD\x00\xEA\xB8"
-     "\xAE\x00"
-     "\xEA\xB8\xAF\x00\xEA\xB8\xB0\x00\xEA\xB8\xB1\x00\xEA\xB8\xB2\x00\xEA\xB8"
-     "\xB3\x00"
-     "\xEA\xB8\xB4\x00\xEA\xB8\xB5\x00\xEA\xB8\xB6\x00\xEA\xB8\xB7\x00\xEA\xB8"
-     "\xB8\x00"
-     "\xEA\xB8\xB9\x00\xEA\xB8\xBA\x00\xEA\xB8\xBB\x00\xEA\xB8\xBC\x00\xEA\xB8"
-     "\xBD\x00"
-     "\xEA\xB8\xBE\x00\xEA\xB8\xBF\x00\xEA\xB9\x80\x00\xEA\xB9\x81\x00\xEA\xB9"
-     "\x82\x00"
-     "\xEA\xB9\x83\x00\xEA\xB9\x84\x00\xEA\xB9\x85\x00\xEA\xB9\x86\x00\xEA\xB9"
-     "\x87\x00"
-     "\xEA\xB9\x88\x00\xEA\xB9\x89\x00\xEA\xB9\x8A\x00\xEA\xB9\x8B\x00\xEA\xB9"
-     "\x8C\x00"
-     "\xEA\xB9\x8D\x00\xEA\xB9\x8E\x00\xEA\xB9\x8F\x00\xEA\xB9\x90\x00\xEA\xB9"
-     "\x91\x00"
-     "\xEA\xB9\x92\x00\xEA\xB9\x93\x00\xEA\xB9\x94\x00\xEA\xB9\x95\x00\xEA\xB9"
-     "\x96\x00"
-     "\xEA\xB9\x97\x00\xEA\xB9\x98\x00\xEA\xB9\x99\x00\xEA\xB9\x9A\x00\xEA\xB9"
-     "\x9B\x00"
-     "\xEA\xB9\x9C\x00\xEA\xB9\x9D\x00\xEA\xB9\x9E\x00\xEA\xB9\x9F\x00\xEA\xB9"
-     "\xA0\x00"
-     "\xEA\xB9\xA1\x00\xEA\xB9\xA2\x00\xEA\xB9\xA3\x00\xEA\xB9\xA4\x00\xEA\xB9"
-     "\xA5\x00"
-     "\xEA\xB9\xA6\x00\xEA\xB9\xA7\x00\xEA\xB9\xA8\x00\xEA\xB9\xA9\x00\xEA\xB9"
-     "\xAA\x00"
-     "\xEA\xB9\xAB\x00\xEA\xB9\xAC\x00\xEA\xB9\xAD\x00\xEA\xB9\xAE\x00\xEA\xB9"
-     "\xAF\x00"
-     "\xEA\xB9\xB0\x00\xEA\xB9\xB1\x00\xEA\xB9\xB2\x00\xEA\xB9\xB3\x00\xEA\xB9"
-     "\xB4\x00"
-     "\xEA\xB9\xB5\x00\xEA\xB9\xB6\x00\xEA\xB9\xB7\x00\xEA\xB9\xB8\x00\xEA\xB9"
-     "\xB9\x00"
-     "\xEA\xB9\xBA\x00\xEA\xB9\xBB\x00\xEA\xB9\xBC\x00\xEA\xB9\xBD\x00\xEA\xB9"
-     "\xBE\x00"
-     "\xEA\xB9\xBF\x00\xEA\xBA\x80\x00\xEA\xBA\x81\x00\xEA\xBA\x82\x00\xEA\xBA"
-     "\x83\x00"
-     "\xEA\xBA\x84\x00\xEA\xBA\x85\x00\xEA\xBA\x86\x00\xEA\xBA\x87\x00\xEA\xBA"
-     "\x88\x00"
-     "\xEA\xBA\x89\x00\xEA\xBA\x8A\x00\xEA\xBA\x8B\x00\xEA\xBA\x8C\x00\xEA\xBA"
-     "\x8D\x00"
-     "\xEA\xBA\x8E\x00\xEA\xBA\x8F\x00\xEA\xBA\x90\x00\xEA\xBA\x91\x00\xEA\xBA"
-     "\x92\x00"
-     "\xEA\xBA\x93\x00\xEA\xBA\x94\x00\xEA\xBA\x95\x00\xEA\xBA\x96\x00\xEA\xBA"
-     "\x97\x00"
-     "\xEA\xBA\x98\x00\xEA\xBA\x99\x00\xEA\xBA\x9A\x00\xEA\xBA\x9B\x00\xEA\xBA"
-     "\x9C\x00"
-     "\xEA\xBA\x9D\x00\xEA\xBA\x9E\x00\xEA\xBA\x9F\x00\xEA\xBA\xA0\x00\xEA\xBA"
-     "\xA1\x00"
-     "\xEA\xBA\xA2\x00\xEA\xBA\xA3\x00\xEA\xBA\xA4\x00\xEA\xBA\xA5\x00\xEA\xBA"
-     "\xA6\x00"
-     "\xEA\xBA\xA7\x00\xEA\xBA\xA8\x00\xEA\xBA\xA9\x00\xEA\xBA\xAA\x00\xEA\xBA"
-     "\xAB\x00"
-     "\xEA\xBA\xAC\x00\xEA\xBA\xAD\x00\xEA\xBA\xAE\x00\xEA\xBA\xAF\x00\xEA\xBA"
-     "\xB0\x00"
-     "\xEA\xBA\xB1\x00\xEA\xBA\xB2\x00\xEA\xBA\xB3\x00\xEA\xBA\xB4\x00\xEA\xBA"
-     "\xB5\x00"
-     "\xEA\xBA\xB6\x00\xEA\xBA\xB7\x00\xEA\xBA\xB8\x00\xEA\xBA\xB9\x00\xEA\xBA"
-     "\xBA\x00"
-     "\xEA\xBA\xBB\x00\xEA\xBA\xBC\x00\xEA\xBA\xBD\x00\xEA\xBA\xBE\x00\xEA\xBA"
-     "\xBF\x00"
-     "\xEA\xBB\x80\x00\xEA\xBB\x81\x00\xEA\xBB\x82\x00\xEA\xBB\x83\x00\xEA\xBB"
-     "\x84\x00"
-     "\xEA\xBB\x85\x00\xEA\xBB\x86\x00\xEA\xBB\x87\x00\xEA\xBB\x88\x00\xEA\xBB"
-     "\x89\x00"
-     "\xEA\xBB\x8A\x00\xEA\xBB\x8B\x00\xEA\xBB\x8C\x00\xEA\xBB\x8D\x00\xEA\xBB"
-     "\x8E\x00"
-     "\xEA\xBB\x8F\x00\xEA\xBB\x90\x00\xEA\xBB\x91\x00\xEA\xBB\x92\x00\xEA\xBB"
-     "\x93\x00"
-     "\xEA\xBB\x94\x00\xEA\xBB\x95\x00\xEA\xBB\x96\x00\xEA\xBB\x97\x00\xEA\xBB"
-     "\x98\x00"
-     "\xEA\xBB\x99\x00\xEA\xBB\x9A\x00\xEA\xBB\x9B\x00\xEA\xBB\x9C\x00\xEA\xBB"
-     "\x9D\x00"
-     "\xEA\xBB\x9E\x00\xEA\xBB\x9F\x00\xEA\xBB\xA0\x00\xEA\xBB\xA1\x00\xEA\xBB"
-     "\xA2\x00"
-     "\xEA\xBB\xA3\x00\xEA\xBB\xA4\x00\xEA\xBB\xA5\x00\xEA\xBB\xA6\x00\xEA\xBB"
-     "\xA7\x00"
-     "\xEA\xBB\xA8\x00\xEA\xBB\xA9\x00\xEA\xBB\xAA\x00\xEA\xBB\xAB\x00\xEA\xBB"
-     "\xAC\x00"
-     "\xEA\xBB\xAD\x00\xEA\xBB\xAE\x00\xEA\xBB\xAF\x00\xEA\xBB\xB0\x00\xEA\xBB"
-     "\xB1\x00"
-     "\xEA\xBB\xB2\x00\xEA\xBB\xB3\x00\xEA\xBB\xB4\x00\xEA\xBB\xB5\x00\xEA\xBB"
-     "\xB6\x00"
-     "\xEA\xBB\xB7\x00\xEA\xBB\xB8\x00\xEA\xBB\xB9\x00\xEA\xBB\xBA\x00\xEA\xBB"
-     "\xBB\x00"
-     "\xEA\xBB\xBC\x00\xEA\xBB\xBD\x00\xEA\xBB\xBE\x00\xEA\xBB\xBF\x00\xEA\xBC"
-     "\x80\x00"
-     "\xEA\xBC\x81\x00\xEA\xBC\x82\x00\xEA\xBC\x83\x00\xEA\xBC\x84\x00\xEA\xBC"
-     "\x85\x00"
-     "\xEA\xBC\x86\x00\xEA\xBC\x87\x00\xEA\xBC\x88\x00\xEA\xBC\x89\x00\xEA\xBC"
-     "\x8A\x00"
-     "\xEA\xBC\x8B\x00\xEA\xBC\x8C\x00\xEA\xBC\x8D\x00\xEA\xBC\x8E\x00\xEA\xBC"
-     "\x8F\x00"
-     "\xEA\xBC\x90\x00\xEA\xBC\x91\x00\xEA\xBC\x92\x00\xEA\xBC\x93\x00\xEA\xBC"
-     "\x94\x00"
-     "\xEA\xBC\x95\x00\xEA\xBC\x96\x00\xEA\xBC\x97\x00\xEA\xBC\x98\x00\xEA\xBC"
-     "\x99\x00"
-     "\xEA\xBC\x9A\x00\xEA\xBC\x9B\x00\xEA\xBC\x9C\x00\xEA\xBC\x9D\x00\xEA\xBC"
-     "\x9E\x00"
-     "\xEA\xBC\x9F\x00\xEA\xBC\xA0\x00\xEA\xBC\xA1\x00\xEA\xBC\xA2\x00\xEA\xBC"
-     "\xA3\x00"
-     "\xEA\xBC\xA4\x00\xEA\xBC\xA5\x00\xEA\xBC\xA6\x00\xEA\xBC\xA7\x00\xEA\xBC"
-     "\xA8\x00"
-     "\xEA\xBC\xA9\x00\xEA\xBC\xAA\x00\xEA\xBC\xAB\x00\xEA\xBC\xAC\x00\xEA\xBC"
-     "\xAD\x00"
-     "\xEA\xBC\xAE\x00\xEA\xBC\xAF\x00\xEA\xBC\xB0\x00\xEA\xBC\xB1\x00\xEA\xBC"
-     "\xB2\x00"
-     "\xEA\xBC\xB3\x00\xEA\xBC\xB4\x00\xEA\xBC\xB5\x00\xEA\xBC\xB6\x00\xEA\xBC"
-     "\xB7\x00"
-     "\xEA\xBC\xB8\x00\xEA\xBC\xB9\x00\xEA\xBC\xBA\x00\xEA\xBC\xBB\x00\xEA\xBC"
-     "\xBC\x00"
-     "\xEA\xBC\xBD\x00\xEA\xBC\xBE\x00\xEA\xBC\xBF\x00\xEA\xBD\x80\x00\xEA\xBD"
-     "\x81\x00"
-     "\xEA\xBD\x82\x00\xEA\xBD\x83\x00\xEA\xBD\x84\x00\xEA\xBD\x85\x00\xEA\xBD"
-     "\x86\x00"
-     "\xEA\xBD\x87\x00\xEA\xBD\x88\x00\xEA\xBD\x89\x00\xEA\xBD\x8A\x00\xEA\xBD"
-     "\x8B\x00"
-     "\xEA\xBD\x8C\x00\xEA\xBD\x8D\x00\xEA\xBD\x8E\x00\xEA\xBD\x8F\x00\xEA\xBD"
-     "\x90\x00"
-     "\xEA\xBD\x91\x00\xEA\xBD\x92\x00\xEA\xBD\x93\x00\xEA\xBD\x94\x00\xEA\xBD"
-     "\x95\x00"
-     "\xEA\xBD\x96\x00\xEA\xBD\x97\x00\xEA\xBD\x98\x00\xEA\xBD\x99\x00\xEA\xBD"
-     "\x9A\x00"
-     "\xEA\xBD\x9B\x00\xEA\xBD\x9C\x00\xEA\xBD\x9D\x00\xEA\xBD\x9E\x00\xEA\xBD"
-     "\x9F\x00"
-     "\xEA\xBD\xA0\x00\xEA\xBD\xA1\x00\xEA\xBD\xA2\x00\xEA\xBD\xA3\x00\xEA\xBD"
-     "\xA4\x00"
-     "\xEA\xBD\xA5\x00\xEA\xBD\xA6\x00\xEA\xBD\xA7\x00\xEA\xBD\xA8\x00\xEA\xBD"
-     "\xA9\x00"
-     "\xEA\xBD\xAA\x00\xEA\xBD\xAB\x00\xEA\xBD\xAC\x00\xEA\xBD\xAD\x00\xEA\xBD"
-     "\xAE\x00"
-     "\xEA\xBD\xAF\x00\xEA\xBD\xB0\x00\xEA\xBD\xB1\x00\xEA\xBD\xB2\x00\xEA\xBD"
-     "\xB3\x00"
-     "\xEA\xBD\xB4\x00\xEA\xBD\xB5\x00\xEA\xBD\xB6\x00\xEA\xBD\xB7\x00\xEA\xBD"
-     "\xB8\x00"
-     "\xEA\xBD\xB9\x00\xEA\xBD\xBA\x00\xEA\xBD\xBB\x00\xEA\xBD\xBC\x00\xEA\xBD"
-     "\xBD\x00"
-     "\xEA\xBD\xBE\x00\xEA\xBD\xBF\x00\xEA\xBE\x80\x00\xEA\xBE\x81\x00\xEA\xBE"
-     "\x82\x00"
-     "\xEA\xBE\x83\x00\xEA\xBE\x84\x00\xEA\xBE\x85\x00\xEA\xBE\x86\x00\xEA\xBE"
-     "\x87\x00"
-     "\xEA\xBE\x88\x00\xEA\xBE\x89\x00\xEA\xBE\x8A\x00\xEA\xBE\x8B\x00\xEA\xBE"
-     "\x8C\x00"
-     "\xEA\xBE\x8D\x00\xEA\xBE\x8E\x00\xEA\xBE\x8F\x00\xEA\xBE\x90\x00\xEA\xBE"
-     "\x91\x00"
-     "\xEA\xBE\x92\x00\xEA\xBE\x93\x00\xEA\xBE\x94\x00\xEA\xBE\x95\x00\xEA\xBE"
-     "\x96\x00"
-     "\xEA\xBE\x97\x00\xEA\xBE\x98\x00\xEA\xBE\x99\x00\xEA\xBE\x9A\x00\xEA\xBE"
-     "\x9B\x00"
-     "\xEA\xBE\x9C\x00\xEA\xBE\x9D\x00\xEA\xBE\x9E\x00\xEA\xBE\x9F\x00\xEA\xBE"
-     "\xA0\x00"
-     "\xEA\xBE\xA1\x00\xEA\xBE\xA2\x00\xEA\xBE\xA3\x00\xEA\xBE\xA4\x00\xEA\xBE"
-     "\xA5\x00"
-     "\xEA\xBE\xA6\x00\xEA\xBE\xA7\x00\xEA\xBE\xA8\x00\xEA\xBE\xA9\x00\xEA\xBE"
-     "\xAA\x00"
-     "\xEA\xBE\xAB\x00\xEA\xBE\xAC\x00\xEA\xBE\xAD\x00\xEA\xBE\xAE\x00\xEA\xBE"
-     "\xAF\x00"
-     "\xEA\xBE\xB0\x00\xEA\xBE\xB1\x00\xEA\xBE\xB2\x00\xEA\xBE\xB3\x00\xEA\xBE"
-     "\xB4\x00"
-     "\xEA\xBE\xB5\x00\xEA\xBE\xB6\x00\xEA\xBE\xB7\x00\xEA\xBE\xB8\x00\xEA\xBE"
-     "\xB9\x00"
-     "\xEA\xBE\xBA\x00\xEA\xBE\xBB\x00\xEA\xBE\xBC\x00\xEA\xBE\xBD\x00\xEA\xBE"
-     "\xBE\x00"
-     "\xEA\xBE\xBF\x00\xEA\xBF\x80\x00\xEA\xBF\x81\x00\xEA\xBF\x82\x00\xEA\xBF"
-     "\x83\x00"
-     "\xEA\xBF\x84\x00\xEA\xBF\x85\x00\xEA\xBF\x86\x00\xEA\xBF\x87\x00\xEA\xBF"
-     "\x88\x00"
-     "\xEA\xBF\x89\x00\xEA\xBF\x8A\x00\xEA\xBF\x8B\x00\xEA\xBF\x8C\x00\xEA\xBF"
-     "\x8D\x00"
-     "\xEA\xBF\x8E\x00\xEA\xBF\x8F\x00\xEA\xBF\x90\x00\xEA\xBF\x91\x00\xEA\xBF"
-     "\x92\x00"
-     "\xEA\xBF\x93\x00\xEA\xBF\x94\x00\xEA\xBF\x95\x00\xEA\xBF\x96\x00\xEA\xBF"
-     "\x97\x00"
-     "\xEA\xBF\x98\x00\xEA\xBF\x99\x00\xEA\xBF\x9A\x00\xEA\xBF\x9B\x00\xEA\xBF"
-     "\x9C\x00"
-     "\xEA\xBF\x9D\x00\xEA\xBF\x9E\x00\xEA\xBF\x9F\x00\xEA\xBF\xA0\x00\xEA\xBF"
-     "\xA1\x00"
-     "\xEA\xBF\xA2\x00\xEA\xBF\xA3\x00\xEA\xBF\xA4\x00\xEA\xBF\xA5\x00\xEA\xBF"
-     "\xA6\x00"
-     "\xEA\xBF\xA7\x00\xEA\xBF\xA8\x00\xEA\xBF\xA9\x00\xEA\xBF\xAA\x00\xEA\xBF"
-     "\xAB\x00"
-     "\xEA\xBF\xAC\x00\xEA\xBF\xAD\x00\xEA\xBF\xAE\x00\xEA\xBF\xAF\x00\xEA\xBF"
-     "\xB0\x00"
-     "\xEA\xBF\xB1\x00\xEA\xBF\xB2\x00\xEA\xBF\xB3\x00\xEA\xBF\xB4\x00\xEA\xBF"
-     "\xB5\x00"
-     "\xEA\xBF\xB6\x00\xEA\xBF\xB7\x00\xEA\xBF\xB8\x00\xEA\xBF\xB9\x00\xEA\xBF"
-     "\xBA\x00"
-     "\xEA\xBF\xBB\x00\xEA\xBF\xBC\x00\xEA\xBF\xBD\x00\xEA\xBF\xBE\x00\xEA\xBF"
-     "\xBF\x00"
-     "\xEB\x80\x80\x00\xEB\x80\x81\x00\xEB\x80\x82\x00\xEB\x80\x83\x00\xEB\x80"
-     "\x84\x00"
-     "\xEB\x80\x85\x00\xEB\x80\x86\x00\xEB\x80\x87\x00\xEB\x80\x88\x00\xEB\x80"
-     "\x89\x00"
-     "\xEB\x80\x8A\x00\xEB\x80\x8B\x00\xEB\x80\x8C\x00\xEB\x80\x8D\x00\xEB\x80"
-     "\x8E\x00"
-     "\xEB\x80\x8F\x00\xEB\x80\x90\x00\xEB\x80\x91\x00\xEB\x80\x92\x00\xEB\x80"
-     "\x93\x00"
-     "\xEB\x80\x94\x00\xEB\x80\x95\x00\xEB\x80\x96\x00\xEB\x80\x97\x00\xEB\x80"
-     "\x98\x00"
-     "\xEB\x80\x99\x00\xEB\x80\x9A\x00\xEB\x80\x9B\x00\xEB\x80\x9C\x00\xEB\x80"
-     "\x9D\x00"
-     "\xEB\x80\x9E\x00\xEB\x80\x9F\x00\xEB\x80\xA0\x00\xEB\x80\xA1\x00\xEB\x80"
-     "\xA2\x00"
-     "\xEB\x80\xA3\x00\xEB\x80\xA4\x00\xEB\x80\xA5\x00\xEB\x80\xA6\x00\xEB\x80"
-     "\xA7\x00"
-     "\xEB\x80\xA8\x00\xEB\x80\xA9\x00\xEB\x80\xAA\x00\xEB\x80\xAB\x00\xEB\x80"
-     "\xAC\x00"
-     "\xEB\x80\xAD\x00\xEB\x80\xAE\x00\xEB\x80\xAF\x00\xEB\x80\xB0\x00\xEB\x80"
-     "\xB1\x00"
-     "\xEB\x80\xB2\x00\xEB\x80\xB3\x00\xEB\x80\xB4\x00\xEB\x80\xB5\x00\xEB\x80"
-     "\xB6\x00"
-     "\xEB\x80\xB7\x00\xEB\x80\xB8\x00\xEB\x80\xB9\x00\xEB\x80\xBA\x00\xEB\x80"
-     "\xBB\x00"
-     "\xEB\x80\xBC\x00\xEB\x80\xBD\x00\xEB\x80\xBE\x00\xEB\x80\xBF\x00\xEB\x81"
-     "\x80\x00"
-     "\xEB\x81\x81\x00\xEB\x81\x82\x00\xEB\x81\x83\x00\xEB\x81\x84\x00\xEB\x81"
-     "\x85\x00"
-     "\xEB\x81\x86\x00\xEB\x81\x87\x00\xEB\x81\x88\x00\xEB\x81\x89\x00\xEB\x81"
-     "\x8A\x00"
-     "\xEB\x81\x8B\x00\xEB\x81\x8C\x00\xEB\x81\x8D\x00\xEB\x81\x8E\x00\xEB\x81"
-     "\x8F\x00"
-     "\xEB\x81\x90\x00\xEB\x81\x91\x00\xEB\x81\x92\x00\xEB\x81\x93\x00\xEB\x81"
-     "\x94\x00"
-     "\xEB\x81\x95\x00\xEB\x81\x96\x00\xEB\x81\x97\x00\xEB\x81\x98\x00\xEB\x81"
-     "\x99\x00"
-     "\xEB\x81\x9A\x00\xEB\x81\x9B\x00\xEB\x81\x9C\x00\xEB\x81\x9D\x00\xEB\x81"
-     "\x9E\x00"
-     "\xEB\x81\x9F\x00\xEB\x81\xA0\x00\xEB\x81\xA1\x00\xEB\x81\xA2\x00\xEB\x81"
-     "\xA3\x00"
-     "\xEB\x81\xA4\x00\xEB\x81\xA5\x00\xEB\x81\xA6\x00\xEB\x81\xA7\x00\xEB\x81"
-     "\xA8\x00"
-     "\xEB\x81\xA9\x00\xEB\x81\xAA\x00\xEB\x81\xAB\x00\xEB\x81\xAC\x00\xEB\x81"
-     "\xAD\x00"
-     "\xEB\x81\xAE\x00\xEB\x81\xAF\x00\xEB\x81\xB0\x00\xEB\x81\xB1\x00\xEB\x81"
-     "\xB2\x00"
-     "\xEB\x81\xB3\x00\xEB\x81\xB4\x00\xEB\x81\xB5\x00\xEB\x81\xB6\x00\xEB\x81"
-     "\xB7\x00"
-     "\xEB\x81\xB8\x00\xEB\x81\xB9\x00\xEB\x81\xBA\x00\xEB\x81\xBB\x00\xEB\x81"
-     "\xBC\x00"
-     "\xEB\x81\xBD\x00\xEB\x81\xBE\x00\xEB\x81\xBF\x00\xEB\x82\x80\x00\xEB\x82"
-     "\x81\x00"
-     "\xEB\x82\x82\x00\xEB\x82\x83\x00\xEB\x82\x84\x00\xEB\x82\x85\x00\xEB\x82"
-     "\x86\x00"
-     "\xEB\x82\x87\x00\xEB\x82\x88\x00\xEB\x82\x89\x00\xEB\x82\x8A\x00\xEB\x82"
-     "\x8B\x00"
-     "\xEB\x82\x8C\x00\xEB\x82\x8D\x00\xEB\x82\x8E\x00\xEB\x82\x8F\x00\xEB\x82"
-     "\x90\x00"
-     "\xEB\x82\x91\x00\xEB\x82\x92\x00\xEB\x82\x93\x00\xEB\x82\x94\x00\xEB\x82"
-     "\x95\x00"
-     "\xEB\x82\x96\x00\xEB\x82\x97\x00\xEB\x82\x98\x00\xEB\x82\x99\x00\xEB\x82"
-     "\x9A\x00"
-     "\xEB\x82\x9B\x00\xEB\x82\x9C\x00\xEB\x82\x9D\x00\xEB\x82\x9E\x00\xEB\x82"
-     "\x9F\x00"
-     "\xEB\x82\xA0\x00\xEB\x82\xA1\x00\xEB\x82\xA2\x00\xEB\x82\xA3\x00\xEB\x82"
-     "\xA4\x00"
-     "\xEB\x82\xA5\x00\xEB\x82\xA6\x00\xEB\x82\xA7\x00\xEB\x82\xA8\x00\xEB\x82"
-     "\xA9\x00"
-     "\xEB\x82\xAA\x00\xEB\x82\xAB\x00\xEB\x82\xAC\x00\xEB\x82\xAD\x00\xEB\x82"
-     "\xAE\x00"
-     "\xEB\x82\xAF\x00\xEB\x82\xB0\x00\xEB\x82\xB1\x00\xEB\x82\xB2\x00\xEB\x82"
-     "\xB3\x00"
-     "\xEB\x82\xB4\x00\xEB\x82\xB5\x00\xEB\x82\xB6\x00\xEB\x82\xB7\x00\xEB\x82"
-     "\xB8\x00"
-     "\xEB\x82\xB9\x00\xEB\x82\xBA\x00\xEB\x82\xBB\x00\xEB\x82\xBC\x00\xEB\x82"
-     "\xBD\x00"
-     "\xEB\x82\xBE\x00\xEB\x82\xBF\x00\xEB\x83\x80\x00\xEB\x83\x81\x00\xEB\x83"
-     "\x82\x00"
-     "\xEB\x83\x83\x00\xEB\x83\x84\x00\xEB\x83\x85\x00\xEB\x83\x86\x00\xEB\x83"
-     "\x87\x00"
-     "\xEB\x83\x88\x00\xEB\x83\x89\x00\xEB\x83\x8A\x00\xEB\x83\x8B\x00\xEB\x83"
-     "\x8C\x00"
-     "\xEB\x83\x8D\x00\xEB\x83\x8E\x00\xEB\x83\x8F\x00\xEB\x83\x90\x00\xEB\x83"
-     "\x91\x00"
-     "\xEB\x83\x92\x00\xEB\x83\x93\x00\xEB\x83\x94\x00\xEB\x83\x95\x00\xEB\x83"
-     "\x96\x00"
-     "\xEB\x83\x97\x00\xEB\x83\x98\x00\xEB\x83\x99\x00\xEB\x83\x9A\x00\xEB\x83"
-     "\x9B\x00"
-     "\xEB\x83\x9C\x00\xEB\x83\x9D\x00\xEB\x83\x9E\x00\xEB\x83\x9F\x00\xEB\x83"
-     "\xA0\x00"
-     "\xEB\x83\xA1\x00\xEB\x83\xA2\x00\xEB\x83\xA3\x00\xEB\x83\xA4\x00\xEB\x83"
-     "\xA5\x00"
-     "\xEB\x83\xA6\x00\xEB\x83\xA7\x00\xEB\x83\xA8\x00\xEB\x83\xA9\x00\xEB\x83"
-     "\xAA\x00"
-     "\xEB\x83\xAB\x00\xEB\x83\xAC\x00\xEB\x83\xAD\x00\xEB\x83\xAE\x00\xEB\x83"
-     "\xAF\x00"
-     "\xEB\x83\xB0\x00\xEB\x83\xB1\x00\xEB\x83\xB2\x00\xEB\x83\xB3\x00\xEB\x83"
-     "\xB4\x00"
-     "\xEB\x83\xB5\x00\xEB\x83\xB6\x00\xEB\x83\xB7\x00\xEB\x83\xB8\x00\xEB\x83"
-     "\xB9\x00"
-     "\xEB\x83\xBA\x00\xEB\x83\xBB\x00\xEB\x83\xBC\x00\xEB\x83\xBD\x00\xEB\x83"
-     "\xBE\x00"
-     "\xEB\x83\xBF\x00\xEB\x84\x80\x00\xEB\x84\x81\x00\xEB\x84\x82\x00\xEB\x84"
-     "\x83\x00"
-     "\xEB\x84\x84\x00\xEB\x84\x85\x00\xEB\x84\x86\x00\xEB\x84\x87\x00\xEB\x84"
-     "\x88\x00"
-     "\xEB\x84\x89\x00\xEB\x84\x8A\x00\xEB\x84\x8B\x00\xEB\x84\x8C\x00\xEB\x84"
-     "\x8D\x00"
-     "\xEB\x84\x8E\x00\xEB\x84\x8F\x00\xEB\x84\x90\x00\xEB\x84\x91\x00\xEB\x84"
-     "\x92\x00"
-     "\xEB\x84\x93\x00\xEB\x84\x94\x00\xEB\x84\x95\x00\xEB\x84\x96\x00\xEB\x84"
-     "\x97\x00"
-     "\xEB\x84\x98\x00\xEB\x84\x99\x00\xEB\x84\x9A\x00\xEB\x84\x9B\x00\xEB\x84"
-     "\x9C\x00"
-     "\xEB\x84\x9D\x00\xEB\x84\x9E\x00\xEB\x84\x9F\x00\xEB\x84\xA0\x00\xEB\x84"
-     "\xA1\x00"
-     "\xEB\x84\xA2\x00\xEB\x84\xA3\x00\xEB\x84\xA4\x00\xEB\x84\xA5\x00\xEB\x84"
-     "\xA6\x00"
-     "\xEB\x84\xA7\x00\xEB\x84\xA8\x00\xEB\x84\xA9\x00\xEB\x84\xAA\x00\xEB\x84"
-     "\xAB\x00"
-     "\xEB\x84\xAC\x00\xEB\x84\xAD\x00\xEB\x84\xAE\x00\xEB\x84\xAF\x00\xEB\x84"
-     "\xB0\x00"
-     "\xEB\x84\xB1\x00\xEB\x84\xB2\x00\xEB\x84\xB3\x00\xEB\x84\xB4\x00\xEB\x84"
-     "\xB5\x00"
-     "\xEB\x84\xB6\x00\xEB\x84\xB7\x00\xEB\x84\xB8\x00\xEB\x84\xB9\x00\xEB\x84"
-     "\xBA\x00"
-     "\xEB\x84\xBB\x00\xEB\x84\xBC\x00\xEB\x84\xBD\x00\xEB\x84\xBE\x00\xEB\x84"
-     "\xBF\x00"
-     "\xEB\x85\x80\x00\xEB\x85\x81\x00\xEB\x85\x82\x00\xEB\x85\x83\x00\xEB\x85"
-     "\x84\x00"
-     "\xEB\x85\x85\x00\xEB\x85\x86\x00\xEB\x85\x87\x00\xEB\x85\x88\x00\xEB\x85"
-     "\x89\x00"
-     "\xEB\x85\x8A\x00\xEB\x85\x8B\x00\xEB\x85\x8C\x00\xEB\x85\x8D\x00\xEB\x85"
-     "\x8E\x00"
-     "\xEB\x85\x8F\x00\xEB\x85\x90\x00\xEB\x85\x91\x00\xEB\x85\x92\x00\xEB\x85"
-     "\x93\x00"
-     "\xEB\x85\x94\x00\xEB\x85\x95\x00\xEB\x85\x96\x00\xEB\x85\x97\x00\xEB\x85"
-     "\x98\x00"
-     "\xEB\x85\x99\x00\xEB\x85\x9A\x00\xEB\x85\x9B\x00\xEB\x85\x9C\x00\xEB\x85"
-     "\x9D\x00"
-     "\xEB\x85\x9E\x00\xEB\x85\x9F\x00\xEB\x85\xA0\x00\xEB\x85\xA1\x00\xEB\x85"
-     "\xA2\x00"
-     "\xEB\x85\xA3\x00\xEB\x85\xA4\x00\xEB\x85\xA5\x00\xEB\x85\xA6\x00\xEB\x85"
-     "\xA7\x00"
-     "\xEB\x85\xA8\x00\xEB\x85\xA9\x00\xEB\x85\xAA\x00\xEB\x85\xAB\x00\xEB\x85"
-     "\xAC\x00"
-     "\xEB\x85\xAD\x00\xEB\x85\xAE\x00\xEB\x85\xAF\x00\xEB\x85\xB0\x00\xEB\x85"
-     "\xB1\x00"
-     "\xEB\x85\xB2\x00\xEB\x85\xB3\x00\xEB\x85\xB4\x00\xEB\x85\xB5\x00\xEB\x85"
-     "\xB6\x00"
-     "\xEB\x85\xB7\x00\xEB\x85\xB8\x00\xEB\x85\xB9\x00\xEB\x85\xBA\x00\xEB\x85"
-     "\xBB\x00"
-     "\xEB\x85\xBC\x00\xEB\x85\xBD\x00\xEB\x85\xBE\x00\xEB\x85\xBF\x00\xEB\x86"
-     "\x80\x00"
-     "\xEB\x86\x81\x00\xEB\x86\x82\x00\xEB\x86\x83\x00\xEB\x86\x84\x00\xEB\x86"
-     "\x85\x00"
-     "\xEB\x86\x86\x00\xEB\x86\x87\x00\xEB\x86\x88\x00\xEB\x86\x89\x00\xEB\x86"
-     "\x8A\x00"
-     "\xEB\x86\x8B\x00\xEB\x86\x8C\x00\xEB\x86\x8D\x00\xEB\x86\x8E\x00\xEB\x86"
-     "\x8F\x00"
-     "\xEB\x86\x90\x00\xEB\x86\x91\x00\xEB\x86\x92\x00\xEB\x86\x93\x00\xEB\x86"
-     "\x94\x00"
-     "\xEB\x86\x95\x00\xEB\x86\x96\x00\xEB\x86\x97\x00\xEB\x86\x98\x00\xEB\x86"
-     "\x99\x00"
-     "\xEB\x86\x9A\x00\xEB\x86\x9B\x00\xEB\x86\x9C\x00\xEB\x86\x9D\x00\xEB\x86"
-     "\x9E\x00"
-     "\xEB\x86\x9F\x00\xEB\x86\xA0\x00\xEB\x86\xA1\x00\xEB\x86\xA2\x00\xEB\x86"
-     "\xA3\x00"
-     "\xEB\x86\xA4\x00\xEB\x86\xA5\x00\xEB\x86\xA6\x00\xEB\x86\xA7\x00\xEB\x86"
-     "\xA8\x00"
-     "\xEB\x86\xA9\x00\xEB\x86\xAA\x00\xEB\x86\xAB\x00\xEB\x86\xAC\x00\xEB\x86"
-     "\xAD\x00"
-     "\xEB\x86\xAE\x00\xEB\x86\xAF\x00\xEB\x86\xB0\x00\xEB\x86\xB1\x00\xEB\x86"
-     "\xB2\x00"
-     "\xEB\x86\xB3\x00\xEB\x86\xB4\x00\xEB\x86\xB5\x00\xEB\x86\xB6\x00\xEB\x86"
-     "\xB7\x00"
-     "\xEB\x86\xB8\x00\xEB\x86\xB9\x00\xEB\x86\xBA\x00\xEB\x86\xBB\x00\xEB\x86"
-     "\xBC\x00"
-     "\xEB\x86\xBD\x00\xEB\x86\xBE\x00\xEB\x86\xBF\x00\xEB\x87\x80\x00\xEB\x87"
-     "\x81\x00"
-     "\xEB\x87\x82\x00\xEB\x87\x83\x00\xEB\x87\x84\x00\xEB\x87\x85\x00\xEB\x87"
-     "\x86\x00"
-     "\xEB\x87\x87\x00\xEB\x87\x88\x00\xEB\x87\x89\x00\xEB\x87\x8A\x00\xEB\x87"
-     "\x8B\x00"
-     "\xEB\x87\x8C\x00\xEB\x87\x8D\x00\xEB\x87\x8E\x00\xEB\x87\x8F\x00\xEB\x87"
-     "\x90\x00"
-     "\xEB\x87\x91\x00\xEB\x87\x92\x00\xEB\x87\x93\x00\xEB\x87\x94\x00\xEB\x87"
-     "\x95\x00"
-     "\xEB\x87\x96\x00\xEB\x87\x97\x00\xEB\x87\x98\x00\xEB\x87\x99\x00\xEB\x87"
-     "\x9A\x00"
-     "\xEB\x87\x9B\x00\xEB\x87\x9C\x00\xEB\x87\x9D\x00\xEB\x87\x9E\x00\xEB\x87"
-     "\x9F\x00"
-     "\xEB\x87\xA0\x00\xEB\x87\xA1\x00\xEB\x87\xA2\x00\xEB\x87\xA3\x00\xEB\x87"
-     "\xA4\x00"
-     "\xEB\x87\xA5\x00\xEB\x87\xA6\x00\xEB\x87\xA7\x00\xEB\x87\xA8\x00\xEB\x87"
-     "\xA9\x00"
-     "\xEB\x87\xAA\x00\xEB\x87\xAB\x00\xEB\x87\xAC\x00\xEB\x87\xAD\x00\xEB\x87"
-     "\xAE\x00"
-     "\xEB\x87\xAF\x00\xEB\x87\xB0\x00\xEB\x87\xB1\x00\xEB\x87\xB2\x00\xEB\x87"
-     "\xB3\x00"
-     "\xEB\x87\xB4\x00\xEB\x87\xB5\x00\xEB\x87\xB6\x00\xEB\x87\xB7\x00\xEB\x87"
-     "\xB8\x00"
-     "\xEB\x87\xB9\x00\xEB\x87\xBA\x00\xEB\x87\xBB\x00\xEB\x87\xBC\x00\xEB\x87"
-     "\xBD\x00"
-     "\xEB\x87\xBE\x00\xEB\x87\xBF\x00\xEB\x88\x80\x00\xEB\x88\x81\x00\xEB\x88"
-     "\x82\x00"
-     "\xEB\x88\x83\x00\xEB\x88\x84\x00\xEB\x88\x85\x00\xEB\x88\x86\x00\xEB\x88"
-     "\x87\x00"
-     "\xEB\x88\x88\x00\xEB\x88\x89\x00\xEB\x88\x8A\x00\xEB\x88\x8B\x00\xEB\x88"
-     "\x8C\x00"
-     "\xEB\x88\x8D\x00\xEB\x88\x8E\x00\xEB\x88\x8F\x00\xEB\x88\x90\x00\xEB\x88"
-     "\x91\x00"
-     "\xEB\x88\x92\x00\xEB\x88\x93\x00\xEB\x88\x94\x00\xEB\x88\x95\x00\xEB\x88"
-     "\x96\x00"
-     "\xEB\x88\x97\x00\xEB\x88\x98\x00\xEB\x88\x99\x00\xEB\x88\x9A\x00\xEB\x88"
-     "\x9B\x00"
-     "\xEB\x88\x9C\x00\xEB\x88\x9D\x00\xEB\x88\x9E\x00\xEB\x88\x9F\x00\xEB\x88"
-     "\xA0\x00"
-     "\xEB\x88\xA1\x00\xEB\x88\xA2\x00\xEB\x88\xA3\x00\xEB\x88\xA4\x00\xEB\x88"
-     "\xA5\x00"
-     "\xEB\x88\xA6\x00\xEB\x88\xA7\x00\xEB\x88\xA8\x00\xEB\x88\xA9\x00\xEB\x88"
-     "\xAA\x00"
-     "\xEB\x88\xAB\x00\xEB\x88\xAC\x00\xEB\x88\xAD\x00\xEB\x88\xAE\x00\xEB\x88"
-     "\xAF\x00"
-     "\xEB\x88\xB0\x00\xEB\x88\xB1\x00\xEB\x88\xB2\x00\xEB\x88\xB3\x00\xEB\x88"
-     "\xB4\x00"
-     "\xEB\x88\xB5\x00\xEB\x88\xB6\x00\xEB\x88\xB7\x00\xEB\x88\xB8\x00\xEB\x88"
-     "\xB9\x00"
-     "\xEB\x88\xBA\x00\xEB\x88\xBB\x00\xEB\x88\xBC\x00\xEB\x88\xBD\x00\xEB\x88"
-     "\xBE\x00"
-     "\xEB\x88\xBF\x00\xEB\x89\x80\x00\xEB\x89\x81\x00\xEB\x89\x82\x00\xEB\x89"
-     "\x83\x00"
-     "\xEB\x89\x84\x00\xEB\x89\x85\x00\xEB\x89\x86\x00\xEB\x89\x87\x00\xEB\x89"
-     "\x88\x00"
-     "\xEB\x89\x89\x00\xEB\x89\x8A\x00\xEB\x89\x8B\x00\xEB\x89\x8C\x00\xEB\x89"
-     "\x8D\x00"
-     "\xEB\x89\x8E\x00\xEB\x89\x8F\x00\xEB\x89\x90\x00\xEB\x89\x91\x00\xEB\x89"
-     "\x92\x00"
-     "\xEB\x89\x93\x00\xEB\x89\x94\x00\xEB\x89\x95\x00\xEB\x89\x96\x00\xEB\x89"
-     "\x97\x00"
-     "\xEB\x89\x98\x00\xEB\x89\x99\x00\xEB\x89\x9A\x00\xEB\x89\x9B\x00\xEB\x89"
-     "\x9C\x00"
-     "\xEB\x89\x9D\x00\xEB\x89\x9E\x00\xEB\x89\x9F\x00\xEB\x89\xA0\x00\xEB\x89"
-     "\xA1\x00"
-     "\xEB\x89\xA2\x00\xEB\x89\xA3\x00\xEB\x89\xA4\x00\xEB\x89\xA5\x00\xEB\x89"
-     "\xA6\x00"
-     "\xEB\x89\xA7\x00\xEB\x89\xA8\x00\xEB\x89\xA9\x00\xEB\x89\xAA\x00\xEB\x89"
-     "\xAB\x00"
-     "\xEB\x89\xAC\x00\xEB\x89\xAD\x00\xEB\x89\xAE\x00\xEB\x89\xAF\x00\xEB\x89"
-     "\xB0\x00"
-     "\xEB\x89\xB1\x00\xEB\x89\xB2\x00\xEB\x89\xB3\x00\xEB\x89\xB4\x00\xEB\x89"
-     "\xB5\x00"
-     "\xEB\x89\xB6\x00\xEB\x89\xB7\x00\xEB\x89\xB8\x00\xEB\x89\xB9\x00\xEB\x89"
-     "\xBA\x00"
-     "\xEB\x89\xBB\x00\xEB\x89\xBC\x00\xEB\x89\xBD\x00\xEB\x89\xBE\x00\xEB\x89"
-     "\xBF\x00"
-     "\xEB\x8A\x80\x00\xEB\x8A\x81\x00\xEB\x8A\x82\x00\xEB\x8A\x83\x00\xEB\x8A"
-     "\x84\x00"
-     "\xEB\x8A\x85\x00\xEB\x8A\x86\x00\xEB\x8A\x87\x00\xEB\x8A\x88\x00\xEB\x8A"
-     "\x89\x00"
-     "\xEB\x8A\x8A\x00\xEB\x8A\x8B\x00\xEB\x8A\x8C\x00\xEB\x8A\x8D\x00\xEB\x8A"
-     "\x8E\x00"
-     "\xEB\x8A\x8F\x00\xEB\x8A\x90\x00\xEB\x8A\x91\x00\xEB\x8A\x92\x00\xEB\x8A"
-     "\x93\x00"
-     "\xEB\x8A\x94\x00\xEB\x8A\x95\x00\xEB\x8A\x96\x00\xEB\x8A\x97\x00\xEB\x8A"
-     "\x98\x00"
-     "\xEB\x8A\x99\x00\xEB\x8A\x9A\x00\xEB\x8A\x9B\x00\xEB\x8A\x9C\x00\xEB\x8A"
-     "\x9D\x00"
-     "\xEB\x8A\x9E\x00\xEB\x8A\x9F\x00\xEB\x8A\xA0\x00\xEB\x8A\xA1\x00\xEB\x8A"
-     "\xA2\x00"
-     "\xEB\x8A\xA3\x00\xEB\x8A\xA4\x00\xEB\x8A\xA5\x00\xEB\x8A\xA6\x00\xEB\x8A"
-     "\xA7\x00"
-     "\xEB\x8A\xA8\x00\xEB\x8A\xA9\x00\xEB\x8A\xAA\x00\xEB\x8A\xAB\x00\xEB\x8A"
-     "\xAC\x00"
-     "\xEB\x8A\xAD\x00\xEB\x8A\xAE\x00\xEB\x8A\xAF\x00\xEB\x8A\xB0\x00\xEB\x8A"
-     "\xB1\x00"
-     "\xEB\x8A\xB2\x00\xEB\x8A\xB3\x00\xEB\x8A\xB4\x00\xEB\x8A\xB5\x00\xEB\x8A"
-     "\xB6\x00"
-     "\xEB\x8A\xB7\x00\xEB\x8A\xB8\x00\xEB\x8A\xB9\x00\xEB\x8A\xBA\x00\xEB\x8A"
-     "\xBB\x00"
-     "\xEB\x8A\xBC\x00\xEB\x8A\xBD\x00\xEB\x8A\xBE\x00\xEB\x8A\xBF\x00\xEB\x8B"
-     "\x80\x00"
-     "\xEB\x8B\x81\x00\xEB\x8B\x82\x00\xEB\x8B\x83\x00\xEB\x8B\x84\x00\xEB\x8B"
-     "\x85\x00"
-     "\xEB\x8B\x86\x00\xEB\x8B\x87\x00\xEB\x8B\x88\x00\xEB\x8B\x89\x00\xEB\x8B"
-     "\x8A\x00"
-     "\xEB\x8B\x8B\x00\xEB\x8B\x8C\x00\xEB\x8B\x8D\x00\xEB\x8B\x8E\x00\xEB\x8B"
-     "\x8F\x00"
-     "\xEB\x8B\x90\x00\xEB\x8B\x91\x00\xEB\x8B\x92\x00\xEB\x8B\x93\x00\xEB\x8B"
-     "\x94\x00"
-     "\xEB\x8B\x95\x00\xEB\x8B\x96\x00\xEB\x8B\x97\x00\xEB\x8B\x98\x00\xEB\x8B"
-     "\x99\x00"
-     "\xEB\x8B\x9A\x00\xEB\x8B\x9B\x00\xEB\x8B\x9C\x00\xEB\x8B\x9D\x00\xEB\x8B"
-     "\x9E\x00"
-     "\xEB\x8B\x9F\x00\xEB\x8B\xA0\x00\xEB\x8B\xA1\x00\xEB\x8B\xA2\x00\xEB\x8B"
-     "\xA3\x00"
-     "\xEB\x8B\xA4\x00\xEB\x8B\xA5\x00\xEB\x8B\xA6\x00\xEB\x8B\xA7\x00\xEB\x8B"
-     "\xA8\x00"
-     "\xEB\x8B\xA9\x00\xEB\x8B\xAA\x00\xEB\x8B\xAB\x00\xEB\x8B\xAC\x00\xEB\x8B"
-     "\xAD\x00"
-     "\xEB\x8B\xAE\x00\xEB\x8B\xAF\x00\xEB\x8B\xB0\x00\xEB\x8B\xB1\x00\xEB\x8B"
-     "\xB2\x00"
-     "\xEB\x8B\xB3\x00\xEB\x8B\xB4\x00\xEB\x8B\xB5\x00\xEB\x8B\xB6\x00\xEB\x8B"
-     "\xB7\x00"
-     "\xEB\x8B\xB8\x00\xEB\x8B\xB9\x00\xEB\x8B\xBA\x00\xEB\x8B\xBB\x00\xEB\x8B"
-     "\xBC\x00"
-     "\xEB\x8B\xBD\x00\xEB\x8B\xBE\x00\xEB\x8B\xBF\x00\xEB\x8C\x80\x00\xEB\x8C"
-     "\x81\x00"
-     "\xEB\x8C\x82\x00\xEB\x8C\x83\x00\xEB\x8C\x84\x00\xEB\x8C\x85\x00\xEB\x8C"
-     "\x86\x00"
-     "\xEB\x8C\x87\x00\xEB\x8C\x88\x00\xEB\x8C\x89\x00\xEB\x8C\x8A\x00\xEB\x8C"
-     "\x8B\x00"
-     "\xEB\x8C\x8C\x00\xEB\x8C\x8D\x00\xEB\x8C\x8E\x00\xEB\x8C\x8F\x00\xEB\x8C"
-     "\x90\x00"
-     "\xEB\x8C\x91\x00\xEB\x8C\x92\x00\xEB\x8C\x93\x00\xEB\x8C\x94\x00\xEB\x8C"
-     "\x95\x00"
-     "\xEB\x8C\x96\x00\xEB\x8C\x97\x00\xEB\x8C\x98\x00\xEB\x8C\x99\x00\xEB\x8C"
-     "\x9A\x00"
-     "\xEB\x8C\x9B\x00\xEB\x8C\x9C\x00\xEB\x8C\x9D\x00\xEB\x8C\x9E\x00\xEB\x8C"
-     "\x9F\x00"
-     "\xEB\x8C\xA0\x00\xEB\x8C\xA1\x00\xEB\x8C\xA2\x00\xEB\x8C\xA3\x00\xEB\x8C"
-     "\xA4\x00"
-     "\xEB\x8C\xA5\x00\xEB\x8C\xA6\x00\xEB\x8C\xA7\x00\xEB\x8C\xA8\x00\xEB\x8C"
-     "\xA9\x00"
-     "\xEB\x8C\xAA\x00\xEB\x8C\xAB\x00\xEB\x8C\xAC\x00\xEB\x8C\xAD\x00\xEB\x8C"
-     "\xAE\x00"
-     "\xEB\x8C\xAF\x00\xEB\x8C\xB0\x00\xEB\x8C\xB1\x00\xEB\x8C\xB2\x00\xEB\x8C"
-     "\xB3\x00"
-     "\xEB\x8C\xB4\x00\xEB\x8C\xB5\x00\xEB\x8C\xB6\x00\xEB\x8C\xB7\x00\xEB\x8C"
-     "\xB8\x00"
-     "\xEB\x8C\xB9\x00\xEB\x8C\xBA\x00\xEB\x8C\xBB\x00\xEB\x8C\xBC\x00\xEB\x8C"
-     "\xBD\x00"
-     "\xEB\x8C\xBE\x00\xEB\x8C\xBF\x00\xEB\x8D\x80\x00\xEB\x8D\x81\x00\xEB\x8D"
-     "\x82\x00"
-     "\xEB\x8D\x83\x00\xEB\x8D\x84\x00\xEB\x8D\x85\x00\xEB\x8D\x86\x00\xEB\x8D"
-     "\x87\x00"
-     "\xEB\x8D\x88\x00\xEB\x8D\x89\x00\xEB\x8D\x8A\x00\xEB\x8D\x8B\x00\xEB\x8D"
-     "\x8C\x00"
-     "\xEB\x8D\x8D\x00\xEB\x8D\x8E\x00\xEB\x8D\x8F\x00\xEB\x8D\x90\x00\xEB\x8D"
-     "\x91\x00"
-     "\xEB\x8D\x92\x00\xEB\x8D\x93\x00\xEB\x8D\x94\x00\xEB\x8D\x95\x00\xEB\x8D"
-     "\x96\x00"
-     "\xEB\x8D\x97\x00\xEB\x8D\x98\x00\xEB\x8D\x99\x00\xEB\x8D\x9A\x00\xEB\x8D"
-     "\x9B\x00"
-     "\xEB\x8D\x9C\x00\xEB\x8D\x9D\x00\xEB\x8D\x9E\x00\xEB\x8D\x9F\x00\xEB\x8D"
-     "\xA0\x00"
-     "\xEB\x8D\xA1\x00\xEB\x8D\xA2\x00\xEB\x8D\xA3\x00\xEB\x8D\xA4\x00\xEB\x8D"
-     "\xA5\x00"
-     "\xEB\x8D\xA6\x00\xEB\x8D\xA7\x00\xEB\x8D\xA8\x00\xEB\x8D\xA9\x00\xEB\x8D"
-     "\xAA\x00"
-     "\xEB\x8D\xAB\x00\xEB\x8D\xAC\x00\xEB\x8D\xAD\x00\xEB\x8D\xAE\x00\xEB\x8D"
-     "\xAF\x00"
-     "\xEB\x8D\xB0\x00\xEB\x8D\xB1\x00\xEB\x8D\xB2\x00\xEB\x8D\xB3\x00\xEB\x8D"
-     "\xB4\x00"
-     "\xEB\x8D\xB5\x00\xEB\x8D\xB6\x00\xEB\x8D\xB7\x00\xEB\x8D\xB8\x00\xEB\x8D"
-     "\xB9\x00"
-     "\xEB\x8D\xBA\x00\xEB\x8D\xBB\x00\xEB\x8D\xBC\x00\xEB\x8D\xBD\x00\xEB\x8D"
-     "\xBE\x00"
-     "\xEB\x8D\xBF\x00\xEB\x8E\x80\x00\xEB\x8E\x81\x00\xEB\x8E\x82\x00\xEB\x8E"
-     "\x83\x00"
-     "\xEB\x8E\x84\x00\xEB\x8E\x85\x00\xEB\x8E\x86\x00\xEB\x8E\x87\x00\xEB\x8E"
-     "\x88\x00"
-     "\xEB\x8E\x89\x00\xEB\x8E\x8A\x00\xEB\x8E\x8B\x00\xEB\x8E\x8C\x00\xEB\x8E"
-     "\x8D\x00"
-     "\xEB\x8E\x8E\x00\xEB\x8E\x8F\x00\xEB\x8E\x90\x00\xEB\x8E\x91\x00\xEB\x8E"
-     "\x92\x00"
-     "\xEB\x8E\x93\x00\xEB\x8E\x94\x00\xEB\x8E\x95\x00\xEB\x8E\x96\x00\xEB\x8E"
-     "\x97\x00"
-     "\xEB\x8E\x98\x00\xEB\x8E\x99\x00\xEB\x8E\x9A\x00\xEB\x8E\x9B\x00\xEB\x8E"
-     "\x9C\x00"
-     "\xEB\x8E\x9D\x00\xEB\x8E\x9E\x00\xEB\x8E\x9F\x00\xEB\x8E\xA0\x00\xEB\x8E"
-     "\xA1\x00"
-     "\xEB\x8E\xA2\x00\xEB\x8E\xA3\x00\xEB\x8E\xA4\x00\xEB\x8E\xA5\x00\xEB\x8E"
-     "\xA6\x00"
-     "\xEB\x8E\xA7\x00\xEB\x8E\xA8\x00\xEB\x8E\xA9\x00\xEB\x8E\xAA\x00\xEB\x8E"
-     "\xAB\x00"
-     "\xEB\x8E\xAC\x00\xEB\x8E\xAD\x00\xEB\x8E\xAE\x00\xEB\x8E\xAF\x00\xEB\x8E"
-     "\xB0\x00"
-     "\xEB\x8E\xB1\x00\xEB\x8E\xB2\x00\xEB\x8E\xB3\x00\xEB\x8E\xB4\x00\xEB\x8E"
-     "\xB5\x00"
-     "\xEB\x8E\xB6\x00\xEB\x8E\xB7\x00\xEB\x8E\xB8\x00\xEB\x8E\xB9\x00\xEB\x8E"
-     "\xBA\x00"
-     "\xEB\x8E\xBB\x00\xEB\x8E\xBC\x00\xEB\x8E\xBD\x00\xEB\x8E\xBE\x00\xEB\x8E"
-     "\xBF\x00"
-     "\xEB\x8F\x80\x00\xEB\x8F\x81\x00\xEB\x8F\x82\x00\xEB\x8F\x83\x00\xEB\x8F"
-     "\x84\x00"
-     "\xEB\x8F\x85\x00\xEB\x8F\x86\x00\xEB\x8F\x87\x00\xEB\x8F\x88\x00\xEB\x8F"
-     "\x89\x00"
-     "\xEB\x8F\x8A\x00\xEB\x8F\x8B\x00\xEB\x8F\x8C\x00\xEB\x8F\x8D\x00\xEB\x8F"
-     "\x8E\x00"
-     "\xEB\x8F\x8F\x00\xEB\x8F\x90\x00\xEB\x8F\x91\x00\xEB\x8F\x92\x00\xEB\x8F"
-     "\x93\x00"
-     "\xEB\x8F\x94\x00\xEB\x8F\x95\x00\xEB\x8F\x96\x00\xEB\x8F\x97\x00\xEB\x8F"
-     "\x98\x00"
-     "\xEB\x8F\x99\x00\xEB\x8F\x9A\x00\xEB\x8F\x9B\x00\xEB\x8F\x9C\x00\xEB\x8F"
-     "\x9D\x00"
-     "\xEB\x8F\x9E\x00\xEB\x8F\x9F\x00\xEB\x8F\xA0\x00\xEB\x8F\xA1\x00\xEB\x8F"
-     "\xA2\x00"
-     "\xEB\x8F\xA3\x00\xEB\x8F\xA4\x00\xEB\x8F\xA5\x00\xEB\x8F\xA6\x00\xEB\x8F"
-     "\xA7\x00"
-     "\xEB\x8F\xA8\x00\xEB\x8F\xA9\x00\xEB\x8F\xAA\x00\xEB\x8F\xAB\x00\xEB\x8F"
-     "\xAC\x00"
-     "\xEB\x8F\xAD\x00\xEB\x8F\xAE\x00\xEB\x8F\xAF\x00\xEB\x8F\xB0\x00\xEB\x8F"
-     "\xB1\x00"
-     "\xEB\x8F\xB2\x00\xEB\x8F\xB3\x00\xEB\x8F\xB4\x00\xEB\x8F\xB5\x00\xEB\x8F"
-     "\xB6\x00"
-     "\xEB\x8F\xB7\x00\xEB\x8F\xB8\x00\xEB\x8F\xB9\x00\xEB\x8F\xBA\x00\xEB\x8F"
-     "\xBB\x00"
-     "\xEB\x8F\xBC\x00\xEB\x8F\xBD\x00\xEB\x8F\xBE\x00\xEB\x8F\xBF\x00\xEB\x90"
-     "\x80\x00"
-     "\xEB\x90\x81\x00\xEB\x90\x82\x00\xEB\x90\x83\x00\xEB\x90\x84\x00\xEB\x90"
-     "\x85\x00"
-     "\xEB\x90\x86\x00\xEB\x90\x87\x00\xEB\x90\x88\x00\xEB\x90\x89\x00\xEB\x90"
-     "\x8A\x00"
-     "\xEB\x90\x8B\x00\xEB\x90\x8C\x00\xEB\x90\x8D\x00\xEB\x90\x8E\x00\xEB\x90"
-     "\x8F\x00"
-     "\xEB\x90\x90\x00\xEB\x90\x91\x00\xEB\x90\x92\x00\xEB\x90\x93\x00\xEB\x90"
-     "\x94\x00"
-     "\xEB\x90\x95\x00\xEB\x90\x96\x00\xEB\x90\x97\x00\xEB\x90\x98\x00\xEB\x90"
-     "\x99\x00"
-     "\xEB\x90\x9A\x00\xEB\x90\x9B\x00\xEB\x90\x9C\x00\xEB\x90\x9D\x00\xEB\x90"
-     "\x9E\x00"
-     "\xEB\x90\x9F\x00\xEB\x90\xA0\x00\xEB\x90\xA1\x00\xEB\x90\xA2\x00\xEB\x90"
-     "\xA3\x00"
-     "\xEB\x90\xA4\x00\xEB\x90\xA5\x00\xEB\x90\xA6\x00\xEB\x90\xA7\x00\xEB\x90"
-     "\xA8\x00"
-     "\xEB\x90\xA9\x00\xEB\x90\xAA\x00\xEB\x90\xAB\x00\xEB\x90\xAC\x00\xEB\x90"
-     "\xAD\x00"
-     "\xEB\x90\xAE\x00\xEB\x90\xAF\x00\xEB\x90\xB0\x00\xEB\x90\xB1\x00\xEB\x90"
-     "\xB2\x00"
-     "\xEB\x90\xB3\x00\xEB\x90\xB4\x00\xEB\x90\xB5\x00\xEB\x90\xB6\x00\xEB\x90"
-     "\xB7\x00"
-     "\xEB\x90\xB8\x00\xEB\x90\xB9\x00\xEB\x90\xBA\x00\xEB\x90\xBB\x00\xEB\x90"
-     "\xBC\x00"
-     "\xEB\x90\xBD\x00\xEB\x90\xBE\x00\xEB\x90\xBF\x00\xEB\x91\x80\x00\xEB\x91"
-     "\x81\x00"
-     "\xEB\x91\x82\x00\xEB\x91\x83\x00\xEB\x91\x84\x00\xEB\x91\x85\x00\xEB\x91"
-     "\x86\x00"
-     "\xEB\x91\x87\x00\xEB\x91\x88\x00\xEB\x91\x89\x00\xEB\x91\x8A\x00\xEB\x91"
-     "\x8B\x00"
-     "\xEB\x91\x8C\x00\xEB\x91\x8D\x00\xEB\x91\x8E\x00\xEB\x91\x8F\x00\xEB\x91"
-     "\x90\x00"
-     "\xEB\x91\x91\x00\xEB\x91\x92\x00\xEB\x91\x93\x00\xEB\x91\x94\x00\xEB\x91"
-     "\x95\x00"
-     "\xEB\x91\x96\x00\xEB\x91\x97\x00\xEB\x91\x98\x00\xEB\x91\x99\x00\xEB\x91"
-     "\x9A\x00"
-     "\xEB\x91\x9B\x00\xEB\x91\x9C\x00\xEB\x91\x9D\x00\xEB\x91\x9E\x00\xEB\x91"
-     "\x9F\x00"
-     "\xEB\x91\xA0\x00\xEB\x91\xA1\x00\xEB\x91\xA2\x00\xEB\x91\xA3\x00\xEB\x91"
-     "\xA4\x00"
-     "\xEB\x91\xA5\x00\xEB\x91\xA6\x00\xEB\x91\xA7\x00\xEB\x91\xA8\x00\xEB\x91"
-     "\xA9\x00"
-     "\xEB\x91\xAA\x00\xEB\x91\xAB\x00\xEB\x91\xAC\x00\xEB\x91\xAD\x00\xEB\x91"
-     "\xAE\x00"
-     "\xEB\x91\xAF\x00\xEB\x91\xB0\x00\xEB\x91\xB1\x00\xEB\x91\xB2\x00\xEB\x91"
-     "\xB3\x00"
-     "\xEB\x91\xB4\x00\xEB\x91\xB5\x00\xEB\x91\xB6\x00\xEB\x91\xB7\x00\xEB\x91"
-     "\xB8\x00"
-     "\xEB\x91\xB9\x00\xEB\x91\xBA\x00\xEB\x91\xBB\x00\xEB\x91\xBC\x00\xEB\x91"
-     "\xBD\x00"
-     "\xEB\x91\xBE\x00\xEB\x91\xBF\x00\xEB\x92\x80\x00\xEB\x92\x81\x00\xEB\x92"
-     "\x82\x00"
-     "\xEB\x92\x83\x00\xEB\x92\x84\x00\xEB\x92\x85\x00\xEB\x92\x86\x00\xEB\x92"
-     "\x87\x00"
-     "\xEB\x92\x88\x00\xEB\x92\x89\x00\xEB\x92\x8A\x00\xEB\x92\x8B\x00\xEB\x92"
-     "\x8C\x00"
-     "\xEB\x92\x8D\x00\xEB\x92\x8E\x00\xEB\x92\x8F\x00\xEB\x92\x90\x00\xEB\x92"
-     "\x91\x00"
-     "\xEB\x92\x92\x00\xEB\x92\x93\x00\xEB\x92\x94\x00\xEB\x92\x95\x00\xEB\x92"
-     "\x96\x00"
-     "\xEB\x92\x97\x00\xEB\x92\x98\x00\xEB\x92\x99\x00\xEB\x92\x9A\x00\xEB\x92"
-     "\x9B\x00"
-     "\xEB\x92\x9C\x00\xEB\x92\x9D\x00\xEB\x92\x9E\x00\xEB\x92\x9F\x00\xEB\x92"
-     "\xA0\x00"
-     "\xEB\x92\xA1\x00\xEB\x92\xA2\x00\xEB\x92\xA3\x00\xEB\x92\xA4\x00\xEB\x92"
-     "\xA5\x00"
-     "\xEB\x92\xA6\x00\xEB\x92\xA7\x00\xEB\x92\xA8\x00\xEB\x92\xA9\x00\xEB\x92"
-     "\xAA\x00"
-     "\xEB\x92\xAB\x00\xEB\x92\xAC\x00\xEB\x92\xAD\x00\xEB\x92\xAE\x00\xEB\x92"
-     "\xAF\x00"
-     "\xEB\x92\xB0\x00\xEB\x92\xB1\x00\xEB\x92\xB2\x00\xEB\x92\xB3\x00\xEB\x92"
-     "\xB4\x00"
-     "\xEB\x92\xB5\x00\xEB\x92\xB6\x00\xEB\x92\xB7\x00\xEB\x92\xB8\x00\xEB\x92"
-     "\xB9\x00"
-     "\xEB\x92\xBA\x00\xEB\x92\xBB\x00\xEB\x92\xBC\x00\xEB\x92\xBD\x00\xEB\x92"
-     "\xBE\x00"
-     "\xEB\x92\xBF\x00\xEB\x93\x80\x00\xEB\x93\x81\x00\xEB\x93\x82\x00\xEB\x93"
-     "\x83\x00"
-     "\xEB\x93\x84\x00\xEB\x93\x85\x00\xEB\x93\x86\x00\xEB\x93\x87\x00\xEB\x93"
-     "\x88\x00"
-     "\xEB\x93\x89\x00\xEB\x93\x8A\x00\xEB\x93\x8B\x00\xEB\x93\x8C\x00\xEB\x93"
-     "\x8D\x00"
-     "\xEB\x93\x8E\x00\xEB\x93\x8F\x00\xEB\x93\x90\x00\xEB\x93\x91\x00\xEB\x93"
-     "\x92\x00"
-     "\xEB\x93\x93\x00\xEB\x93\x94\x00\xEB\x93\x95\x00\xEB\x93\x96\x00\xEB\x93"
-     "\x97\x00"
-     "\xEB\x93\x98\x00\xEB\x93\x99\x00\xEB\x93\x9A\x00\xEB\x93\x9B\x00\xEB\x93"
-     "\x9C\x00"
-     "\xEB\x93\x9D\x00\xEB\x93\x9E\x00\xEB\x93\x9F\x00\xEB\x93\xA0\x00\xEB\x93"
-     "\xA1\x00"
-     "\xEB\x93\xA2\x00\xEB\x93\xA3\x00\xEB\x93\xA4\x00\xEB\x93\xA5\x00\xEB\x93"
-     "\xA6\x00"
-     "\xEB\x93\xA7\x00\xEB\x93\xA8\x00\xEB\x93\xA9\x00\xEB\x93\xAA\x00\xEB\x93"
-     "\xAB\x00"
-     "\xEB\x93\xAC\x00\xEB\x93\xAD\x00\xEB\x93\xAE\x00\xEB\x93\xAF\x00\xEB\x93"
-     "\xB0\x00"
-     "\xEB\x93\xB1\x00\xEB\x93\xB2\x00\xEB\x93\xB3\x00\xEB\x93\xB4\x00\xEB\x93"
-     "\xB5\x00"
-     "\xEB\x93\xB6\x00\xEB\x93\xB7\x00\xEB\x93\xB8\x00\xEB\x93\xB9\x00\xEB\x93"
-     "\xBA\x00"
-     "\xEB\x93\xBB\x00\xEB\x93\xBC\x00\xEB\x93\xBD\x00\xEB\x93\xBE\x00\xEB\x93"
-     "\xBF\x00"
-     "\xEB\x94\x80\x00\xEB\x94\x81\x00\xEB\x94\x82\x00\xEB\x94\x83\x00\xEB\x94"
-     "\x84\x00"
-     "\xEB\x94\x85\x00\xEB\x94\x86\x00\xEB\x94\x87\x00\xEB\x94\x88\x00\xEB\x94"
-     "\x89\x00"
-     "\xEB\x94\x8A\x00\xEB\x94\x8B\x00\xEB\x94\x8C\x00\xEB\x94\x8D\x00\xEB\x94"
-     "\x8E\x00"
-     "\xEB\x94\x8F\x00\xEB\x94\x90\x00\xEB\x94\x91\x00\xEB\x94\x92\x00\xEB\x94"
-     "\x93\x00"
-     "\xEB\x94\x94\x00\xEB\x94\x95\x00\xEB\x94\x96\x00\xEB\x94\x97\x00\xEB\x94"
-     "\x98\x00"
-     "\xEB\x94\x99\x00\xEB\x94\x9A\x00\xEB\x94\x9B\x00\xEB\x94\x9C\x00\xEB\x94"
-     "\x9D\x00"
-     "\xEB\x94\x9E\x00\xEB\x94\x9F\x00\xEB\x94\xA0\x00\xEB\x94\xA1\x00\xEB\x94"
-     "\xA2\x00"
-     "\xEB\x94\xA3\x00\xEB\x94\xA4\x00\xEB\x94\xA5\x00\xEB\x94\xA6\x00\xEB\x94"
-     "\xA7\x00"
-     "\xEB\x94\xA8\x00\xEB\x94\xA9\x00\xEB\x94\xAA\x00\xEB\x94\xAB\x00\xEB\x94"
-     "\xAC\x00"
-     "\xEB\x94\xAD\x00\xEB\x94\xAE\x00\xEB\x94\xAF\x00\xEB\x94\xB0\x00\xEB\x94"
-     "\xB1\x00"
-     "\xEB\x94\xB2\x00\xEB\x94\xB3\x00\xEB\x94\xB4\x00\xEB\x94\xB5\x00\xEB\x94"
-     "\xB6\x00"
-     "\xEB\x94\xB7\x00\xEB\x94\xB8\x00\xEB\x94\xB9\x00\xEB\x94\xBA\x00\xEB\x94"
-     "\xBB\x00"
-     "\xEB\x94\xBC\x00\xEB\x94\xBD\x00\xEB\x94\xBE\x00\xEB\x94\xBF\x00\xEB\x95"
-     "\x80\x00"
-     "\xEB\x95\x81\x00\xEB\x95\x82\x00\xEB\x95\x83\x00\xEB\x95\x84\x00\xEB\x95"
-     "\x85\x00"
-     "\xEB\x95\x86\x00\xEB\x95\x87\x00\xEB\x95\x88\x00\xEB\x95\x89\x00\xEB\x95"
-     "\x8A\x00"
-     "\xEB\x95\x8B\x00\xEB\x95\x8C\x00\xEB\x95\x8D\x00\xEB\x95\x8E\x00\xEB\x95"
-     "\x8F\x00"
-     "\xEB\x95\x90\x00\xEB\x95\x91\x00\xEB\x95\x92\x00\xEB\x95\x93\x00\xEB\x95"
-     "\x94\x00"
-     "\xEB\x95\x95\x00\xEB\x95\x96\x00\xEB\x95\x97\x00\xEB\x95\x98\x00\xEB\x95"
-     "\x99\x00"
-     "\xEB\x95\x9A\x00\xEB\x95\x9B\x00\xEB\x95\x9C\x00\xEB\x95\x9D\x00\xEB\x95"
-     "\x9E\x00"
-     "\xEB\x95\x9F\x00\xEB\x95\xA0\x00\xEB\x95\xA1\x00\xEB\x95\xA2\x00\xEB\x95"
-     "\xA3\x00"
-     "\xEB\x95\xA4\x00\xEB\x95\xA5\x00\xEB\x95\xA6\x00\xEB\x95\xA7\x00\xEB\x95"
-     "\xA8\x00"
-     "\xEB\x95\xA9\x00\xEB\x95\xAA\x00\xEB\x95\xAB\x00\xEB\x95\xAC\x00\xEB\x95"
-     "\xAD\x00"
-     "\xEB\x95\xAE\x00\xEB\x95\xAF\x00\xEB\x95\xB0\x00\xEB\x95\xB1\x00\xEB\x95"
-     "\xB2\x00"
-     "\xEB\x95\xB3\x00\xEB\x95\xB4\x00\xEB\x95\xB5\x00\xEB\x95\xB6\x00\xEB\x95"
-     "\xB7\x00"
-     "\xEB\x95\xB8\x00\xEB\x95\xB9\x00\xEB\x95\xBA\x00\xEB\x95\xBB\x00\xEB\x95"
-     "\xBC\x00"
-     "\xEB\x95\xBD\x00\xEB\x95\xBE\x00\xEB\x95\xBF\x00\xEB\x96\x80\x00\xEB\x96"
-     "\x81\x00"
-     "\xEB\x96\x82\x00\xEB\x96\x83\x00\xEB\x96\x84\x00\xEB\x96\x85\x00\xEB\x96"
-     "\x86\x00"
-     "\xEB\x96\x87\x00\xEB\x96\x88\x00\xEB\x96\x89\x00\xEB\x96\x8A\x00\xEB\x96"
-     "\x8B\x00"
-     "\xEB\x96\x8C\x00\xEB\x96\x8D\x00\xEB\x96\x8E\x00\xEB\x96\x8F\x00\xEB\x96"
-     "\x90\x00"
-     "\xEB\x96\x91\x00\xEB\x96\x92\x00\xEB\x96\x93\x00\xEB\x96\x94\x00\xEB\x96"
-     "\x95\x00"
-     "\xEB\x96\x96\x00\xEB\x96\x97\x00\xEB\x96\x98\x00\xEB\x96\x99\x00\xEB\x96"
-     "\x9A\x00"
-     "\xEB\x96\x9B\x00\xEB\x96\x9C\x00\xEB\x96\x9D\x00\xEB\x96\x9E\x00\xEB\x96"
-     "\x9F\x00"
-     "\xEB\x96\xA0\x00\xEB\x96\xA1\x00\xEB\x96\xA2\x00\xEB\x96\xA3\x00\xEB\x96"
-     "\xA4\x00"
-     "\xEB\x96\xA5\x00\xEB\x96\xA6\x00\xEB\x96\xA7\x00\xEB\x96\xA8\x00\xEB\x96"
-     "\xA9\x00"
-     "\xEB\x96\xAA\x00\xEB\x96\xAB\x00\xEB\x96\xAC\x00\xEB\x96\xAD\x00\xEB\x96"
-     "\xAE\x00"
-     "\xEB\x96\xAF\x00\xEB\x96\xB0\x00\xEB\x96\xB1\x00\xEB\x96\xB2\x00\xEB\x96"
-     "\xB3\x00"
-     "\xEB\x96\xB4\x00\xEB\x96\xB5\x00\xEB\x96\xB6\x00\xEB\x96\xB7\x00\xEB\x96"
-     "\xB8\x00"
-     "\xEB\x96\xB9\x00\xEB\x96\xBA\x00\xEB\x96\xBB\x00\xEB\x96\xBC\x00\xEB\x96"
-     "\xBD\x00"
-     "\xEB\x96\xBE\x00\xEB\x96\xBF\x00\xEB\x97\x80\x00\xEB\x97\x81\x00\xEB\x97"
-     "\x82\x00"
-     "\xEB\x97\x83\x00\xEB\x97\x84\x00\xEB\x97\x85\x00\xEB\x97\x86\x00\xEB\x97"
-     "\x87\x00"
-     "\xEB\x97\x88\x00\xEB\x97\x89\x00\xEB\x97\x8A\x00\xEB\x97\x8B\x00\xEB\x97"
-     "\x8C\x00"
-     "\xEB\x97\x8D\x00\xEB\x97\x8E\x00\xEB\x97\x8F\x00\xEB\x97\x90\x00\xEB\x97"
-     "\x91\x00"
-     "\xEB\x97\x92\x00\xEB\x97\x93\x00\xEB\x97\x94\x00\xEB\x97\x95\x00\xEB\x97"
-     "\x96\x00"
-     "\xEB\x97\x97\x00\xEB\x97\x98\x00\xEB\x97\x99\x00\xEB\x97\x9A\x00\xEB\x97"
-     "\x9B\x00"
-     "\xEB\x97\x9C\x00\xEB\x97\x9D\x00\xEB\x97\x9E\x00\xEB\x97\x9F\x00\xEB\x97"
-     "\xA0\x00"
-     "\xEB\x97\xA1\x00\xEB\x97\xA2\x00\xEB\x97\xA3\x00\xEB\x97\xA4\x00\xEB\x97"
-     "\xA5\x00"
-     "\xEB\x97\xA6\x00\xEB\x97\xA7\x00\xEB\x97\xA8\x00\xEB\x97\xA9\x00\xEB\x97"
-     "\xAA\x00"
-     "\xEB\x97\xAB\x00\xEB\x97\xAC\x00\xEB\x97\xAD\x00\xEB\x97\xAE\x00\xEB\x97"
-     "\xAF\x00"
-     "\xEB\x97\xB0\x00\xEB\x97\xB1\x00\xEB\x97\xB2\x00\xEB\x97\xB3\x00\xEB\x97"
-     "\xB4\x00"
-     "\xEB\x97\xB5\x00\xEB\x97\xB6\x00\xEB\x97\xB7\x00\xEB\x97\xB8\x00\xEB\x97"
-     "\xB9\x00"
-     "\xEB\x97\xBA\x00\xEB\x97\xBB\x00\xEB\x97\xBC\x00\xEB\x97\xBD\x00\xEB\x97"
-     "\xBE\x00"
-     "\xEB\x97\xBF\x00\xEB\x98\x80\x00\xEB\x98\x81\x00\xEB\x98\x82\x00\xEB\x98"
-     "\x83\x00"
-     "\xEB\x98\x84\x00\xEB\x98\x85\x00\xEB\x98\x86\x00\xEB\x98\x87\x00\xEB\x98"
-     "\x88\x00"
-     "\xEB\x98\x89\x00\xEB\x98\x8A\x00\xEB\x98\x8B\x00\xEB\x98\x8C\x00\xEB\x98"
-     "\x8D\x00"
-     "\xEB\x98\x8E\x00\xEB\x98\x8F\x00\xEB\x98\x90\x00\xEB\x98\x91\x00\xEB\x98"
-     "\x92\x00"
-     "\xEB\x98\x93\x00\xEB\x98\x94\x00\xEB\x98\x95\x00\xEB\x98\x96\x00\xEB\x98"
-     "\x97\x00"
-     "\xEB\x98\x98\x00\xEB\x98\x99\x00\xEB\x98\x9A\x00\xEB\x98\x9B\x00\xEB\x98"
-     "\x9C\x00"
-     "\xEB\x98\x9D\x00\xEB\x98\x9E\x00\xEB\x98\x9F\x00\xEB\x98\xA0\x00\xEB\x98"
-     "\xA1\x00"
-     "\xEB\x98\xA2\x00\xEB\x98\xA3\x00\xEB\x98\xA4\x00\xEB\x98\xA5\x00\xEB\x98"
-     "\xA6\x00"
-     "\xEB\x98\xA7\x00\xEB\x98\xA8\x00\xEB\x98\xA9\x00\xEB\x98\xAA\x00\xEB\x98"
-     "\xAB\x00"
-     "\xEB\x98\xAC\x00\xEB\x98\xAD\x00\xEB\x98\xAE\x00\xEB\x98\xAF\x00\xEB\x98"
-     "\xB0\x00"
-     "\xEB\x98\xB1\x00\xEB\x98\xB2\x00\xEB\x98\xB3\x00\xEB\x98\xB4\x00\xEB\x98"
-     "\xB5\x00"
-     "\xEB\x98\xB6\x00\xEB\x98\xB7\x00\xEB\x98\xB8\x00\xEB\x98\xB9\x00\xEB\x98"
-     "\xBA\x00"
-     "\xEB\x98\xBB\x00\xEB\x98\xBC\x00\xEB\x98\xBD\x00\xEB\x98\xBE\x00\xEB\x98"
-     "\xBF\x00"
-     "\xEB\x99\x80\x00\xEB\x99\x81\x00\xEB\x99\x82\x00\xEB\x99\x83\x00\xEB\x99"
-     "\x84\x00"
-     "\xEB\x99\x85\x00\xEB\x99\x86\x00\xEB\x99\x87\x00\xEB\x99\x88\x00\xEB\x99"
-     "\x89\x00"
-     "\xEB\x99\x8A\x00\xEB\x99\x8B\x00\xEB\x99\x8C\x00\xEB\x99\x8D\x00\xEB\x99"
-     "\x8E\x00"
-     "\xEB\x99\x8F\x00\xEB\x99\x90\x00\xEB\x99\x91\x00\xEB\x99\x92\x00\xEB\x99"
-     "\x93\x00"
-     "\xEB\x99\x94\x00\xEB\x99\x95\x00\xEB\x99\x96\x00\xEB\x99\x97\x00\xEB\x99"
-     "\x98\x00"
-     "\xEB\x99\x99\x00\xEB\x99\x9A\x00\xEB\x99\x9B\x00\xEB\x99\x9C\x00\xEB\x99"
-     "\x9D\x00"
-     "\xEB\x99\x9E\x00\xEB\x99\x9F\x00\xEB\x99\xA0\x00\xEB\x99\xA1\x00\xEB\x99"
-     "\xA2\x00"
-     "\xEB\x99\xA3\x00\xEB\x99\xA4\x00\xEB\x99\xA5\x00\xEB\x99\xA6\x00\xEB\x99"
-     "\xA7\x00"
-     "\xEB\x99\xA8\x00\xEB\x99\xA9\x00\xEB\x99\xAA\x00\xEB\x99\xAB\x00\xEB\x99"
-     "\xAC\x00"
-     "\xEB\x99\xAD\x00\xEB\x99\xAE\x00\xEB\x99\xAF\x00\xEB\x99\xB0\x00\xEB\x99"
-     "\xB1\x00"
-     "\xEB\x99\xB2\x00\xEB\x99\xB3\x00\xEB\x99\xB4\x00\xEB\x99\xB5\x00\xEB\x99"
-     "\xB6\x00"
-     "\xEB\x99\xB7\x00\xEB\x99\xB8\x00\xEB\x99\xB9\x00\xEB\x99\xBA\x00\xEB\x99"
-     "\xBB\x00"
-     "\xEB\x99\xBC\x00\xEB\x99\xBD\x00\xEB\x99\xBE\x00\xEB\x99\xBF\x00\xEB\x9A"
-     "\x80\x00"
-     "\xEB\x9A\x81\x00\xEB\x9A\x82\x00\xEB\x9A\x83\x00\xEB\x9A\x84\x00\xEB\x9A"
-     "\x85\x00"
-     "\xEB\x9A\x86\x00\xEB\x9A\x87\x00\xEB\x9A\x88\x00\xEB\x9A\x89\x00\xEB\x9A"
-     "\x8A\x00"
-     "\xEB\x9A\x8B\x00\xEB\x9A\x8C\x00\xEB\x9A\x8D\x00\xEB\x9A\x8E\x00\xEB\x9A"
-     "\x8F\x00"
-     "\xEB\x9A\x90\x00\xEB\x9A\x91\x00\xEB\x9A\x92\x00\xEB\x9A\x93\x00\xEB\x9A"
-     "\x94\x00"
-     "\xEB\x9A\x95\x00\xEB\x9A\x96\x00\xEB\x9A\x97\x00\xEB\x9A\x98\x00\xEB\x9A"
-     "\x99\x00"
-     "\xEB\x9A\x9A\x00\xEB\x9A\x9B\x00\xEB\x9A\x9C\x00\xEB\x9A\x9D\x00\xEB\x9A"
-     "\x9E\x00"
-     "\xEB\x9A\x9F\x00\xEB\x9A\xA0\x00\xEB\x9A\xA1\x00\xEB\x9A\xA2\x00\xEB\x9A"
-     "\xA3\x00"
-     "\xEB\x9A\xA4\x00\xEB\x9A\xA5\x00\xEB\x9A\xA6\x00\xEB\x9A\xA7\x00\xEB\x9A"
-     "\xA8\x00"
-     "\xEB\x9A\xA9\x00\xEB\x9A\xAA\x00\xEB\x9A\xAB\x00\xEB\x9A\xAC\x00\xEB\x9A"
-     "\xAD\x00"
-     "\xEB\x9A\xAE\x00\xEB\x9A\xAF\x00\xEB\x9A\xB0\x00\xEB\x9A\xB1\x00\xEB\x9A"
-     "\xB2\x00"
-     "\xEB\x9A\xB3\x00\xEB\x9A\xB4\x00\xEB\x9A\xB5\x00\xEB\x9A\xB6\x00\xEB\x9A"
-     "\xB7\x00"
-     "\xEB\x9A\xB8\x00\xEB\x9A\xB9\x00\xEB\x9A\xBA\x00\xEB\x9A\xBB\x00\xEB\x9A"
-     "\xBC\x00"
-     "\xEB\x9A\xBD\x00\xEB\x9A\xBE\x00\xEB\x9A\xBF\x00\xEB\x9B\x80\x00\xEB\x9B"
-     "\x81\x00"
-     "\xEB\x9B\x82\x00\xEB\x9B\x83\x00\xEB\x9B\x84\x00\xEB\x9B\x85\x00\xEB\x9B"
-     "\x86\x00"
-     "\xEB\x9B\x87\x00\xEB\x9B\x88\x00\xEB\x9B\x89\x00\xEB\x9B\x8A\x00\xEB\x9B"
-     "\x8B\x00"
-     "\xEB\x9B\x8C\x00\xEB\x9B\x8D\x00\xEB\x9B\x8E\x00\xEB\x9B\x8F\x00\xEB\x9B"
-     "\x90\x00"
-     "\xEB\x9B\x91\x00\xEB\x9B\x92\x00\xEB\x9B\x93\x00\xEB\x9B\x94\x00\xEB\x9B"
-     "\x95\x00"
-     "\xEB\x9B\x96\x00\xEB\x9B\x97\x00\xEB\x9B\x98\x00\xEB\x9B\x99\x00\xEB\x9B"
-     "\x9A\x00"
-     "\xEB\x9B\x9B\x00\xEB\x9B\x9C\x00\xEB\x9B\x9D\x00\xEB\x9B\x9E\x00\xEB\x9B"
-     "\x9F\x00"
-     "\xEB\x9B\xA0\x00\xEB\x9B\xA1\x00\xEB\x9B\xA2\x00\xEB\x9B\xA3\x00\xEB\x9B"
-     "\xA4\x00"
-     "\xEB\x9B\xA5\x00\xEB\x9B\xA6\x00\xEB\x9B\xA7\x00\xEB\x9B\xA8\x00\xEB\x9B"
-     "\xA9\x00"
-     "\xEB\x9B\xAA\x00\xEB\x9B\xAB\x00\xEB\x9B\xAC\x00\xEB\x9B\xAD\x00\xEB\x9B"
-     "\xAE\x00"
-     "\xEB\x9B\xAF\x00\xEB\x9B\xB0\x00\xEB\x9B\xB1\x00\xEB\x9B\xB2\x00\xEB\x9B"
-     "\xB3\x00"
-     "\xEB\x9B\xB4\x00\xEB\x9B\xB5\x00\xEB\x9B\xB6\x00\xEB\x9B\xB7\x00\xEB\x9B"
-     "\xB8\x00"
-     "\xEB\x9B\xB9\x00\xEB\x9B\xBA\x00\xEB\x9B\xBB\x00\xEB\x9B\xBC\x00\xEB\x9B"
-     "\xBD\x00"
-     "\xEB\x9B\xBE\x00\xEB\x9B\xBF\x00\xEB\x9C\x80\x00\xEB\x9C\x81\x00\xEB\x9C"
-     "\x82\x00"
-     "\xEB\x9C\x83\x00\xEB\x9C\x84\x00\xEB\x9C\x85\x00\xEB\x9C\x86\x00\xEB\x9C"
-     "\x87\x00"
-     "\xEB\x9C\x88\x00\xEB\x9C\x89\x00\xEB\x9C\x8A\x00\xEB\x9C\x8B\x00\xEB\x9C"
-     "\x8C\x00"
-     "\xEB\x9C\x8D\x00\xEB\x9C\x8E\x00\xEB\x9C\x8F\x00\xEB\x9C\x90\x00\xEB\x9C"
-     "\x91\x00"
-     "\xEB\x9C\x92\x00\xEB\x9C\x93\x00\xEB\x9C\x94\x00\xEB\x9C\x95\x00\xEB\x9C"
-     "\x96\x00"
-     "\xEB\x9C\x97\x00\xEB\x9C\x98\x00\xEB\x9C\x99\x00\xEB\x9C\x9A\x00\xEB\x9C"
-     "\x9B\x00"
-     "\xEB\x9C\x9C\x00\xEB\x9C\x9D\x00\xEB\x9C\x9E\x00\xEB\x9C\x9F\x00\xEB\x9C"
-     "\xA0\x00"
-     "\xEB\x9C\xA1\x00\xEB\x9C\xA2\x00\xEB\x9C\xA3\x00\xEB\x9C\xA4\x00\xEB\x9C"
-     "\xA5\x00"
-     "\xEB\x9C\xA6\x00\xEB\x9C\xA7\x00\xEB\x9C\xA8\x00\xEB\x9C\xA9\x00\xEB\x9C"
-     "\xAA\x00"
-     "\xEB\x9C\xAB\x00\xEB\x9C\xAC\x00\xEB\x9C\xAD\x00\xEB\x9C\xAE\x00\xEB\x9C"
-     "\xAF\x00"
-     "\xEB\x9C\xB0\x00\xEB\x9C\xB1\x00\xEB\x9C\xB2\x00\xEB\x9C\xB3\x00\xEB\x9C"
-     "\xB4\x00"
-     "\xEB\x9C\xB5\x00\xEB\x9C\xB6\x00\xEB\x9C\xB7\x00\xEB\x9C\xB8\x00\xEB\x9C"
-     "\xB9\x00"
-     "\xEB\x9C\xBA\x00\xEB\x9C\xBB\x00\xEB\x9C\xBC\x00\xEB\x9C\xBD\x00\xEB\x9C"
-     "\xBE\x00"
-     "\xEB\x9C\xBF\x00\xEB\x9D\x80\x00\xEB\x9D\x81\x00\xEB\x9D\x82\x00\xEB\x9D"
-     "\x83\x00"
-     "\xEB\x9D\x84\x00\xEB\x9D\x85\x00\xEB\x9D\x86\x00\xEB\x9D\x87\x00\xEB\x9D"
-     "\x88\x00"
-     "\xEB\x9D\x89\x00\xEB\x9D\x8A\x00\xEB\x9D\x8B\x00\xEB\x9D\x8C\x00\xEB\x9D"
-     "\x8D\x00"
-     "\xEB\x9D\x8E\x00\xEB\x9D\x8F\x00\xEB\x9D\x90\x00\xEB\x9D\x91\x00\xEB\x9D"
-     "\x92\x00"
-     "\xEB\x9D\x93\x00\xEB\x9D\x94\x00\xEB\x9D\x95\x00\xEB\x9D\x96\x00\xEB\x9D"
-     "\x97\x00"
-     "\xEB\x9D\x98\x00\xEB\x9D\x99\x00\xEB\x9D\x9A\x00\xEB\x9D\x9B\x00\xEB\x9D"
-     "\x9C\x00"
-     "\xEB\x9D\x9D\x00\xEB\x9D\x9E\x00\xEB\x9D\x9F\x00\xEB\x9D\xA0\x00\xEB\x9D"
-     "\xA1\x00"
-     "\xEB\x9D\xA2\x00\xEB\x9D\xA3\x00\xEB\x9D\xA4\x00\xEB\x9D\xA5\x00\xEB\x9D"
-     "\xA6\x00"
-     "\xEB\x9D\xA7\x00\xEB\x9D\xA8\x00\xEB\x9D\xA9\x00\xEB\x9D\xAA\x00\xEB\x9D"
-     "\xAB\x00"
-     "\xEB\x9D\xAC\x00\xEB\x9D\xAD\x00\xEB\x9D\xAE\x00\xEB\x9D\xAF\x00\xEB\x9D"
-     "\xB0\x00"
-     "\xEB\x9D\xB1\x00\xEB\x9D\xB2\x00\xEB\x9D\xB3\x00\xEB\x9D\xB4\x00\xEB\x9D"
-     "\xB5\x00"
-     "\xEB\x9D\xB6\x00\xEB\x9D\xB7\x00\xEB\x9D\xB8\x00\xEB\x9D\xB9\x00\xEB\x9D"
-     "\xBA\x00"
-     "\xEB\x9D\xBB\x00\xEB\x9D\xBC\x00\xEB\x9D\xBD\x00\xEB\x9D\xBE\x00\xEB\x9D"
-     "\xBF\x00"
-     "\xEB\x9E\x80\x00\xEB\x9E\x81\x00\xEB\x9E\x82\x00\xEB\x9E\x83\x00\xEB\x9E"
-     "\x84\x00"
-     "\xEB\x9E\x85\x00\xEB\x9E\x86\x00\xEB\x9E\x87\x00\xEB\x9E\x88\x00\xEB\x9E"
-     "\x89\x00"
-     "\xEB\x9E\x8A\x00\xEB\x9E\x8B\x00\xEB\x9E\x8C\x00\xEB\x9E\x8D\x00\xEB\x9E"
-     "\x8E\x00"
-     "\xEB\x9E\x8F\x00\xEB\x9E\x90\x00\xEB\x9E\x91\x00\xEB\x9E\x92\x00\xEB\x9E"
-     "\x93\x00"
-     "\xEB\x9E\x94\x00\xEB\x9E\x95\x00\xEB\x9E\x96\x00\xEB\x9E\x97\x00\xEB\x9E"
-     "\x98\x00"
-     "\xEB\x9E\x99\x00\xEB\x9E\x9A\x00\xEB\x9E\x9B\x00\xEB\x9E\x9C\x00\xEB\x9E"
-     "\x9D\x00"
-     "\xEB\x9E\x9E\x00\xEB\x9E\x9F\x00\xEB\x9E\xA0\x00\xEB\x9E\xA1\x00\xEB\x9E"
-     "\xA2\x00"
-     "\xEB\x9E\xA3\x00\xEB\x9E\xA4\x00\xEB\x9E\xA5\x00\xEB\x9E\xA6\x00\xEB\x9E"
-     "\xA7\x00"
-     "\xEB\x9E\xA8\x00\xEB\x9E\xA9\x00\xEB\x9E\xAA\x00\xEB\x9E\xAB\x00\xEB\x9E"
-     "\xAC\x00"
-     "\xEB\x9E\xAD\x00\xEB\x9E\xAE\x00\xEB\x9E\xAF\x00\xEB\x9E\xB0\x00\xEB\x9E"
-     "\xB1\x00"
-     "\xEB\x9E\xB2\x00\xEB\x9E\xB3\x00\xEB\x9E\xB4\x00\xEB\x9E\xB5\x00\xEB\x9E"
-     "\xB6\x00"
-     "\xEB\x9E\xB7\x00\xEB\x9E\xB8\x00\xEB\x9E\xB9\x00\xEB\x9E\xBA\x00\xEB\x9E"
-     "\xBB\x00"
-     "\xEB\x9E\xBC\x00\xEB\x9E\xBD\x00\xEB\x9E\xBE\x00\xEB\x9E\xBF\x00\xEB\x9F"
-     "\x80\x00"
-     "\xEB\x9F\x81\x00\xEB\x9F\x82\x00\xEB\x9F\x83\x00\xEB\x9F\x84\x00\xEB\x9F"
-     "\x85\x00"
-     "\xEB\x9F\x86\x00\xEB\x9F\x87\x00\xEB\x9F\x88\x00\xEB\x9F\x89\x00\xEB\x9F"
-     "\x8A\x00"
-     "\xEB\x9F\x8B\x00\xEB\x9F\x8C\x00\xEB\x9F\x8D\x00\xEB\x9F\x8E\x00\xEB\x9F"
-     "\x8F\x00"
-     "\xEB\x9F\x90\x00\xEB\x9F\x91\x00\xEB\x9F\x92\x00\xEB\x9F\x93\x00\xEB\x9F"
-     "\x94\x00"
-     "\xEB\x9F\x95\x00\xEB\x9F\x96\x00\xEB\x9F\x97\x00\xEB\x9F\x98\x00\xEB\x9F"
-     "\x99\x00"
-     "\xEB\x9F\x9A\x00\xEB\x9F\x9B\x00\xEB\x9F\x9C\x00\xEB\x9F\x9D\x00\xEB\x9F"
-     "\x9E\x00"
-     "\xEB\x9F\x9F\x00\xEB\x9F\xA0\x00\xEB\x9F\xA1\x00\xEB\x9F\xA2\x00\xEB\x9F"
-     "\xA3\x00"
-     "\xEB\x9F\xA4\x00\xEB\x9F\xA5\x00\xEB\x9F\xA6\x00\xEB\x9F\xA7\x00\xEB\x9F"
-     "\xA8\x00"
-     "\xEB\x9F\xA9\x00\xEB\x9F\xAA\x00\xEB\x9F\xAB\x00\xEB\x9F\xAC\x00\xEB\x9F"
-     "\xAD\x00"
-     "\xEB\x9F\xAE\x00\xEB\x9F\xAF\x00\xEB\x9F\xB0\x00\xEB\x9F\xB1\x00\xEB\x9F"
-     "\xB2\x00"
-     "\xEB\x9F\xB3\x00\xEB\x9F\xB4\x00\xEB\x9F\xB5\x00\xEB\x9F\xB6\x00\xEB\x9F"
-     "\xB7\x00"
-     "\xEB\x9F\xB8\x00\xEB\x9F\xB9\x00\xEB\x9F\xBA\x00\xEB\x9F\xBB\x00\xEB\x9F"
-     "\xBC\x00"
-     "\xEB\x9F\xBD\x00\xEB\x9F\xBE\x00\xEB\x9F\xBF\x00\xEB\xA0\x80\x00\xEB\xA0"
-     "\x81\x00"
-     "\xEB\xA0\x82\x00\xEB\xA0\x83\x00\xEB\xA0\x84\x00\xEB\xA0\x85\x00\xEB\xA0"
-     "\x86\x00"
-     "\xEB\xA0\x87\x00\xEB\xA0\x88\x00\xEB\xA0\x89\x00\xEB\xA0\x8A\x00\xEB\xA0"
-     "\x8B\x00"
-     "\xEB\xA0\x8C\x00\xEB\xA0\x8D\x00\xEB\xA0\x8E\x00\xEB\xA0\x8F\x00\xEB\xA0"
-     "\x90\x00"
-     "\xEB\xA0\x91\x00\xEB\xA0\x92\x00\xEB\xA0\x93\x00\xEB\xA0\x94\x00\xEB\xA0"
-     "\x95\x00"
-     "\xEB\xA0\x96\x00\xEB\xA0\x97\x00\xEB\xA0\x98\x00\xEB\xA0\x99\x00\xEB\xA0"
-     "\x9A\x00"
-     "\xEB\xA0\x9B\x00\xEB\xA0\x9C\x00\xEB\xA0\x9D\x00\xEB\xA0\x9E\x00\xEB\xA0"
-     "\x9F\x00"
-     "\xEB\xA0\xA0\x00\xEB\xA0\xA1\x00\xEB\xA0\xA2\x00\xEB\xA0\xA3\x00\xEB\xA0"
-     "\xA4\x00"
-     "\xEB\xA0\xA5\x00\xEB\xA0\xA6\x00\xEB\xA0\xA7\x00\xEB\xA0\xA8\x00\xEB\xA0"
-     "\xA9\x00"
-     "\xEB\xA0\xAA\x00\xEB\xA0\xAB\x00\xEB\xA0\xAC\x00\xEB\xA0\xAD\x00\xEB\xA0"
-     "\xAE\x00"
-     "\xEB\xA0\xAF\x00\xEB\xA0\xB0\x00\xEB\xA0\xB1\x00\xEB\xA0\xB2\x00\xEB\xA0"
-     "\xB3\x00"
-     "\xEB\xA0\xB4\x00\xEB\xA0\xB5\x00\xEB\xA0\xB6\x00\xEB\xA0\xB7\x00\xEB\xA0"
-     "\xB8\x00"
-     "\xEB\xA0\xB9\x00\xEB\xA0\xBA\x00\xEB\xA0\xBB\x00\xEB\xA0\xBC\x00\xEB\xA0"
-     "\xBD\x00"
-     "\xEB\xA0\xBE\x00\xEB\xA0\xBF\x00\xEB\xA1\x80\x00\xEB\xA1\x81\x00\xEB\xA1"
-     "\x82\x00"
-     "\xEB\xA1\x83\x00\xEB\xA1\x84\x00\xEB\xA1\x85\x00\xEB\xA1\x86\x00\xEB\xA1"
-     "\x87\x00"
-     "\xEB\xA1\x88\x00\xEB\xA1\x89\x00\xEB\xA1\x8A\x00\xEB\xA1\x8B\x00\xEB\xA1"
-     "\x8C\x00"
-     "\xEB\xA1\x8D\x00\xEB\xA1\x8E\x00\xEB\xA1\x8F\x00\xEB\xA1\x90\x00\xEB\xA1"
-     "\x91\x00"
-     "\xEB\xA1\x92\x00\xEB\xA1\x93\x00\xEB\xA1\x94\x00\xEB\xA1\x95\x00\xEB\xA1"
-     "\x96\x00"
-     "\xEB\xA1\x97\x00\xEB\xA1\x98\x00\xEB\xA1\x99\x00\xEB\xA1\x9A\x00\xEB\xA1"
-     "\x9B\x00"
-     "\xEB\xA1\x9C\x00\xEB\xA1\x9D\x00\xEB\xA1\x9E\x00\xEB\xA1\x9F\x00\xEB\xA1"
-     "\xA0\x00"
-     "\xEB\xA1\xA1\x00\xEB\xA1\xA2\x00\xEB\xA1\xA3\x00\xEB\xA1\xA4\x00\xEB\xA1"
-     "\xA5\x00"
-     "\xEB\xA1\xA6\x00\xEB\xA1\xA7\x00\xEB\xA1\xA8\x00\xEB\xA1\xA9\x00\xEB\xA1"
-     "\xAA\x00"
-     "\xEB\xA1\xAB\x00\xEB\xA1\xAC\x00\xEB\xA1\xAD\x00\xEB\xA1\xAE\x00\xEB\xA1"
-     "\xAF\x00"
-     "\xEB\xA1\xB0\x00\xEB\xA1\xB1\x00\xEB\xA1\xB2\x00\xEB\xA1\xB3\x00\xEB\xA1"
-     "\xB4\x00"
-     "\xEB\xA1\xB5\x00\xEB\xA1\xB6\x00\xEB\xA1\xB7\x00\xEB\xA1\xB8\x00\xEB\xA1"
-     "\xB9\x00"
-     "\xEB\xA1\xBA\x00\xEB\xA1\xBB\x00\xEB\xA1\xBC\x00\xEB\xA1\xBD\x00\xEB\xA1"
-     "\xBE\x00"
-     "\xEB\xA1\xBF\x00\xEB\xA2\x80\x00\xEB\xA2\x81\x00\xEB\xA2\x82\x00\xEB\xA2"
-     "\x83\x00"
-     "\xEB\xA2\x84\x00\xEB\xA2\x85\x00\xEB\xA2\x86\x00\xEB\xA2\x87\x00\xEB\xA2"
-     "\x88\x00"
-     "\xEB\xA2\x89\x00\xEB\xA2\x8A\x00\xEB\xA2\x8B\x00\xEB\xA2\x8C\x00\xEB\xA2"
-     "\x8D\x00"
-     "\xEB\xA2\x8E\x00\xEB\xA2\x8F\x00\xEB\xA2\x90\x00\xEB\xA2\x91\x00\xEB\xA2"
-     "\x92\x00"
-     "\xEB\xA2\x93\x00\xEB\xA2\x94\x00\xEB\xA2\x95\x00\xEB\xA2\x96\x00\xEB\xA2"
-     "\x97\x00"
-     "\xEB\xA2\x98\x00\xEB\xA2\x99\x00\xEB\xA2\x9A\x00\xEB\xA2\x9B\x00\xEB\xA2"
-     "\x9C\x00"
-     "\xEB\xA2\x9D\x00\xEB\xA2\x9E\x00\xEB\xA2\x9F\x00\xEB\xA2\xA0\x00\xEB\xA2"
-     "\xA1\x00"
-     "\xEB\xA2\xA2\x00\xEB\xA2\xA3\x00\xEB\xA2\xA4\x00\xEB\xA2\xA5\x00\xEB\xA2"
-     "\xA6\x00"
-     "\xEB\xA2\xA7\x00\xEB\xA2\xA8\x00\xEB\xA2\xA9\x00\xEB\xA2\xAA\x00\xEB\xA2"
-     "\xAB\x00"
-     "\xEB\xA2\xAC\x00\xEB\xA2\xAD\x00\xEB\xA2\xAE\x00\xEB\xA2\xAF\x00\xEB\xA2"
-     "\xB0\x00"
-     "\xEB\xA2\xB1\x00\xEB\xA2\xB2\x00\xEB\xA2\xB3\x00\xEB\xA2\xB4\x00\xEB\xA2"
-     "\xB5\x00"
-     "\xEB\xA2\xB6\x00\xEB\xA2\xB7\x00\xEB\xA2\xB8\x00\xEB\xA2\xB9\x00\xEB\xA2"
-     "\xBA\x00"
-     "\xEB\xA2\xBB\x00\xEB\xA2\xBC\x00\xEB\xA2\xBD\x00\xEB\xA2\xBE\x00\xEB\xA2"
-     "\xBF\x00"
-     "\xEB\xA3\x80\x00\xEB\xA3\x81\x00\xEB\xA3\x82\x00\xEB\xA3\x83\x00\xEB\xA3"
-     "\x84\x00"
-     "\xEB\xA3\x85\x00\xEB\xA3\x86\x00\xEB\xA3\x87\x00\xEB\xA3\x88\x00\xEB\xA3"
-     "\x89\x00"
-     "\xEB\xA3\x8A\x00\xEB\xA3\x8B\x00\xEB\xA3\x8C\x00\xEB\xA3\x8D\x00\xEB\xA3"
-     "\x8E\x00"
-     "\xEB\xA3\x8F\x00\xEB\xA3\x90\x00\xEB\xA3\x91\x00\xEB\xA3\x92\x00\xEB\xA3"
-     "\x93\x00"
-     "\xEB\xA3\x94\x00\xEB\xA3\x95\x00\xEB\xA3\x96\x00\xEB\xA3\x97\x00\xEB\xA3"
-     "\x98\x00"
-     "\xEB\xA3\x99\x00\xEB\xA3\x9A\x00\xEB\xA3\x9B\x00\xEB\xA3\x9C\x00\xEB\xA3"
-     "\x9D\x00"
-     "\xEB\xA3\x9E\x00\xEB\xA3\x9F\x00\xEB\xA3\xA0\x00\xEB\xA3\xA1\x00\xEB\xA3"
-     "\xA2\x00"
-     "\xEB\xA3\xA3\x00\xEB\xA3\xA4\x00\xEB\xA3\xA5\x00\xEB\xA3\xA6\x00\xEB\xA3"
-     "\xA7\x00"
-     "\xEB\xA3\xA8\x00\xEB\xA3\xA9\x00\xEB\xA3\xAA\x00\xEB\xA3\xAB\x00\xEB\xA3"
-     "\xAC\x00"
-     "\xEB\xA3\xAD\x00\xEB\xA3\xAE\x00\xEB\xA3\xAF\x00\xEB\xA3\xB0\x00\xEB\xA3"
-     "\xB1\x00"
-     "\xEB\xA3\xB2\x00\xEB\xA3\xB3\x00\xEB\xA3\xB4\x00\xEB\xA3\xB5\x00\xEB\xA3"
-     "\xB6\x00"
-     "\xEB\xA3\xB7\x00\xEB\xA3\xB8\x00\xEB\xA3\xB9\x00\xEB\xA3\xBA\x00\xEB\xA3"
-     "\xBB\x00"
-     "\xEB\xA3\xBC\x00\xEB\xA3\xBD\x00\xEB\xA3\xBE\x00\xEB\xA3\xBF\x00\xEB\xA4"
-     "\x80\x00"
-     "\xEB\xA4\x81\x00\xEB\xA4\x82\x00\xEB\xA4\x83\x00\xEB\xA4\x84\x00\xEB\xA4"
-     "\x85\x00"
-     "\xEB\xA4\x86\x00\xEB\xA4\x87\x00\xEB\xA4\x88\x00\xEB\xA4\x89\x00\xEB\xA4"
-     "\x8A\x00"
-     "\xEB\xA4\x8B\x00\xEB\xA4\x8C\x00\xEB\xA4\x8D\x00\xEB\xA4\x8E\x00\xEB\xA4"
-     "\x8F\x00"
-     "\xEB\xA4\x90\x00\xEB\xA4\x91\x00\xEB\xA4\x92\x00\xEB\xA4\x93\x00\xEB\xA4"
-     "\x94\x00"
-     "\xEB\xA4\x95\x00\xEB\xA4\x96\x00\xEB\xA4\x97\x00\xEB\xA4\x98\x00\xEB\xA4"
-     "\x99\x00"
-     "\xEB\xA4\x9A\x00\xEB\xA4\x9B\x00\xEB\xA4\x9C\x00\xEB\xA4\x9D\x00\xEB\xA4"
-     "\x9E\x00"
-     "\xEB\xA4\x9F\x00\xEB\xA4\xA0\x00\xEB\xA4\xA1\x00\xEB\xA4\xA2\x00\xEB\xA4"
-     "\xA3\x00"
-     "\xEB\xA4\xA4\x00\xEB\xA4\xA5\x00\xEB\xA4\xA6\x00\xEB\xA4\xA7\x00\xEB\xA4"
-     "\xA8\x00"
-     "\xEB\xA4\xA9\x00\xEB\xA4\xAA\x00\xEB\xA4\xAB\x00\xEB\xA4\xAC\x00\xEB\xA4"
-     "\xAD\x00"
-     "\xEB\xA4\xAE\x00\xEB\xA4\xAF\x00\xEB\xA4\xB0\x00\xEB\xA4\xB1\x00\xEB\xA4"
-     "\xB2\x00"
-     "\xEB\xA4\xB3\x00\xEB\xA4\xB4\x00\xEB\xA4\xB5\x00\xEB\xA4\xB6\x00\xEB\xA4"
-     "\xB7\x00"
-     "\xEB\xA4\xB8\x00\xEB\xA4\xB9\x00\xEB\xA4\xBA\x00\xEB\xA4\xBB\x00\xEB\xA4"
-     "\xBC\x00"
-     "\xEB\xA4\xBD\x00\xEB\xA4\xBE\x00\xEB\xA4\xBF\x00\xEB\xA5\x80\x00\xEB\xA5"
-     "\x81\x00"
-     "\xEB\xA5\x82\x00\xEB\xA5\x83\x00\xEB\xA5\x84\x00\xEB\xA5\x85\x00\xEB\xA5"
-     "\x86\x00"
-     "\xEB\xA5\x87\x00\xEB\xA5\x88\x00\xEB\xA5\x89\x00\xEB\xA5\x8A\x00\xEB\xA5"
-     "\x8B\x00"
-     "\xEB\xA5\x8C\x00\xEB\xA5\x8D\x00\xEB\xA5\x8E\x00\xEB\xA5\x8F\x00\xEB\xA5"
-     "\x90\x00"
-     "\xEB\xA5\x91\x00\xEB\xA5\x92\x00\xEB\xA5\x93\x00\xEB\xA5\x94\x00\xEB\xA5"
-     "\x95\x00"
-     "\xEB\xA5\x96\x00\xEB\xA5\x97\x00\xEB\xA5\x98\x00\xEB\xA5\x99\x00\xEB\xA5"
-     "\x9A\x00"
-     "\xEB\xA5\x9B\x00\xEB\xA5\x9C\x00\xEB\xA5\x9D\x00\xEB\xA5\x9E\x00\xEB\xA5"
-     "\x9F\x00"
-     "\xEB\xA5\xA0\x00\xEB\xA5\xA1\x00\xEB\xA5\xA2\x00\xEB\xA5\xA3\x00\xEB\xA5"
-     "\xA4\x00"
-     "\xEB\xA5\xA5\x00\xEB\xA5\xA6\x00\xEB\xA5\xA7\x00\xEB\xA5\xA8\x00\xEB\xA5"
-     "\xA9\x00"
-     "\xEB\xA5\xAA\x00\xEB\xA5\xAB\x00\xEB\xA5\xAC\x00\xEB\xA5\xAD\x00\xEB\xA5"
-     "\xAE\x00"
-     "\xEB\xA5\xAF\x00\xEB\xA5\xB0\x00\xEB\xA5\xB1\x00\xEB\xA5\xB2\x00\xEB\xA5"
-     "\xB3\x00"
-     "\xEB\xA5\xB4\x00\xEB\xA5\xB5\x00\xEB\xA5\xB6\x00\xEB\xA5\xB7\x00\xEB\xA5"
-     "\xB8\x00"
-     "\xEB\xA5\xB9\x00\xEB\xA5\xBA\x00\xEB\xA5\xBB\x00\xEB\xA5\xBC\x00\xEB\xA5"
-     "\xBD\x00"
-     "\xEB\xA5\xBE\x00\xEB\xA5\xBF\x00\xEB\xA6\x80\x00\xEB\xA6\x81\x00\xEB\xA6"
-     "\x82\x00"
-     "\xEB\xA6\x83\x00\xEB\xA6\x84\x00\xEB\xA6\x85\x00\xEB\xA6\x86\x00\xEB\xA6"
-     "\x87\x00"
-     "\xEB\xA6\x88\x00\xEB\xA6\x89\x00\xEB\xA6\x8A\x00\xEB\xA6\x8B\x00\xEB\xA6"
-     "\x8C\x00"
-     "\xEB\xA6\x8D\x00\xEB\xA6\x8E\x00\xEB\xA6\x8F\x00\xEB\xA6\x90\x00\xEB\xA6"
-     "\x91\x00"
-     "\xEB\xA6\x92\x00\xEB\xA6\x93\x00\xEB\xA6\x94\x00\xEB\xA6\x95\x00\xEB\xA6"
-     "\x96\x00"
-     "\xEB\xA6\x97\x00\xEB\xA6\x98\x00\xEB\xA6\x99\x00\xEB\xA6\x9A\x00\xEB\xA6"
-     "\x9B\x00"
-     "\xEB\xA6\x9C\x00\xEB\xA6\x9D\x00\xEB\xA6\x9E\x00\xEB\xA6\x9F\x00\xEB\xA6"
-     "\xA0\x00"
-     "\xEB\xA6\xA1\x00\xEB\xA6\xA2\x00\xEB\xA6\xA3\x00\xEB\xA6\xA4\x00\xEB\xA6"
-     "\xA5\x00"
-     "\xEB\xA6\xA6\x00\xEB\xA6\xA7\x00\xEB\xA6\xA8\x00\xEB\xA6\xA9\x00\xEB\xA6"
-     "\xAA\x00"
-     "\xEB\xA6\xAB\x00\xEB\xA6\xAC\x00\xEB\xA6\xAD\x00\xEB\xA6\xAE\x00\xEB\xA6"
-     "\xAF\x00"
-     "\xEB\xA6\xB0\x00\xEB\xA6\xB1\x00\xEB\xA6\xB2\x00\xEB\xA6\xB3\x00\xEB\xA6"
-     "\xB4\x00"
-     "\xEB\xA6\xB5\x00\xEB\xA6\xB6\x00\xEB\xA6\xB7\x00\xEB\xA6\xB8\x00\xEB\xA6"
-     "\xB9\x00"
-     "\xEB\xA6\xBA\x00\xEB\xA6\xBB\x00\xEB\xA6\xBC\x00\xEB\xA6\xBD\x00\xEB\xA6"
-     "\xBE\x00"
-     "\xEB\xA6\xBF\x00\xEB\xA7\x80\x00\xEB\xA7\x81\x00\xEB\xA7\x82\x00\xEB\xA7"
-     "\x83\x00"
-     "\xEB\xA7\x84\x00\xEB\xA7\x85\x00\xEB\xA7\x86\x00\xEB\xA7\x87\x00\xEB\xA7"
-     "\x88\x00"
-     "\xEB\xA7\x89\x00\xEB\xA7\x8A\x00\xEB\xA7\x8B\x00\xEB\xA7\x8C\x00\xEB\xA7"
-     "\x8D\x00"
-     "\xEB\xA7\x8E\x00\xEB\xA7\x8F\x00\xEB\xA7\x90\x00\xEB\xA7\x91\x00\xEB\xA7"
-     "\x92\x00"
-     "\xEB\xA7\x93\x00\xEB\xA7\x94\x00\xEB\xA7\x95\x00\xEB\xA7\x96\x00\xEB\xA7"
-     "\x97\x00"
-     "\xEB\xA7\x98\x00\xEB\xA7\x99\x00\xEB\xA7\x9A\x00\xEB\xA7\x9B\x00\xEB\xA7"
-     "\x9C\x00"
-     "\xEB\xA7\x9D\x00\xEB\xA7\x9E\x00\xEB\xA7\x9F\x00\xEB\xA7\xA0\x00\xEB\xA7"
-     "\xA1\x00"
-     "\xEB\xA7\xA2\x00\xEB\xA7\xA3\x00\xEB\xA7\xA4\x00\xEB\xA7\xA5\x00\xEB\xA7"
-     "\xA6\x00"
-     "\xEB\xA7\xA7\x00\xEB\xA7\xA8\x00\xEB\xA7\xA9\x00\xEB\xA7\xAA\x00\xEB\xA7"
-     "\xAB\x00"
-     "\xEB\xA7\xAC\x00\xEB\xA7\xAD\x00\xEB\xA7\xAE\x00\xEB\xA7\xAF\x00\xEB\xA7"
-     "\xB0\x00"
-     "\xEB\xA7\xB1\x00\xEB\xA7\xB2\x00\xEB\xA7\xB3\x00\xEB\xA7\xB4\x00\xEB\xA7"
-     "\xB5\x00"
-     "\xEB\xA7\xB6\x00\xEB\xA7\xB7\x00\xEB\xA7\xB8\x00\xEB\xA7\xB9\x00\xEB\xA7"
-     "\xBA\x00"
-     "\xEB\xA7\xBB\x00\xEB\xA7\xBC\x00\xEB\xA7\xBD\x00\xEB\xA7\xBE\x00\xEB\xA7"
-     "\xBF\x00"
-     "\xEB\xA8\x80\x00\xEB\xA8\x81\x00\xEB\xA8\x82\x00\xEB\xA8\x83\x00\xEB\xA8"
-     "\x84\x00"
-     "\xEB\xA8\x85\x00\xEB\xA8\x86\x00\xEB\xA8\x87\x00\xEB\xA8\x88\x00\xEB\xA8"
-     "\x89\x00"
-     "\xEB\xA8\x8A\x00\xEB\xA8\x8B\x00\xEB\xA8\x8C\x00\xEB\xA8\x8D\x00\xEB\xA8"
-     "\x8E\x00"
-     "\xEB\xA8\x8F\x00\xEB\xA8\x90\x00\xEB\xA8\x91\x00\xEB\xA8\x92\x00\xEB\xA8"
-     "\x93\x00"
-     "\xEB\xA8\x94\x00\xEB\xA8\x95\x00\xEB\xA8\x96\x00\xEB\xA8\x97\x00\xEB\xA8"
-     "\x98\x00"
-     "\xEB\xA8\x99\x00\xEB\xA8\x9A\x00\xEB\xA8\x9B\x00\xEB\xA8\x9C\x00\xEB\xA8"
-     "\x9D\x00"
-     "\xEB\xA8\x9E\x00\xEB\xA8\x9F\x00\xEB\xA8\xA0\x00\xEB\xA8\xA1\x00\xEB\xA8"
-     "\xA2\x00"
-     "\xEB\xA8\xA3\x00\xEB\xA8\xA4\x00\xEB\xA8\xA5\x00\xEB\xA8\xA6\x00\xEB\xA8"
-     "\xA7\x00"
-     "\xEB\xA8\xA8\x00\xEB\xA8\xA9\x00\xEB\xA8\xAA\x00\xEB\xA8\xAB\x00\xEB\xA8"
-     "\xAC\x00"
-     "\xEB\xA8\xAD\x00\xEB\xA8\xAE\x00\xEB\xA8\xAF\x00\xEB\xA8\xB0\x00\xEB\xA8"
-     "\xB1\x00"
-     "\xEB\xA8\xB2\x00\xEB\xA8\xB3\x00\xEB\xA8\xB4\x00\xEB\xA8\xB5\x00\xEB\xA8"
-     "\xB6\x00"
-     "\xEB\xA8\xB7\x00\xEB\xA8\xB8\x00\xEB\xA8\xB9\x00\xEB\xA8\xBA\x00\xEB\xA8"
-     "\xBB\x00"
-     "\xEB\xA8\xBC\x00\xEB\xA8\xBD\x00\xEB\xA8\xBE\x00\xEB\xA8\xBF\x00\xEB\xA9"
-     "\x80\x00"
-     "\xEB\xA9\x81\x00\xEB\xA9\x82\x00\xEB\xA9\x83\x00\xEB\xA9\x84\x00\xEB\xA9"
-     "\x85\x00"
-     "\xEB\xA9\x86\x00\xEB\xA9\x87\x00\xEB\xA9\x88\x00\xEB\xA9\x89\x00\xEB\xA9"
-     "\x8A\x00"
-     "\xEB\xA9\x8B\x00\xEB\xA9\x8C\x00\xEB\xA9\x8D\x00\xEB\xA9\x8E\x00\xEB\xA9"
-     "\x8F\x00"
-     "\xEB\xA9\x90\x00\xEB\xA9\x91\x00\xEB\xA9\x92\x00\xEB\xA9\x93\x00\xEB\xA9"
-     "\x94\x00"
-     "\xEB\xA9\x95\x00\xEB\xA9\x96\x00\xEB\xA9\x97\x00\xEB\xA9\x98\x00\xEB\xA9"
-     "\x99\x00"
-     "\xEB\xA9\x9A\x00\xEB\xA9\x9B\x00\xEB\xA9\x9C\x00\xEB\xA9\x9D\x00\xEB\xA9"
-     "\x9E\x00"
-     "\xEB\xA9\x9F\x00\xEB\xA9\xA0\x00\xEB\xA9\xA1\x00\xEB\xA9\xA2\x00\xEB\xA9"
-     "\xA3\x00"
-     "\xEB\xA9\xA4\x00\xEB\xA9\xA5\x00\xEB\xA9\xA6\x00\xEB\xA9\xA7\x00\xEB\xA9"
-     "\xA8\x00"
-     "\xEB\xA9\xA9\x00\xEB\xA9\xAA\x00\xEB\xA9\xAB\x00\xEB\xA9\xAC\x00\xEB\xA9"
-     "\xAD\x00"
-     "\xEB\xA9\xAE\x00\xEB\xA9\xAF\x00\xEB\xA9\xB0\x00\xEB\xA9\xB1\x00\xEB\xA9"
-     "\xB2\x00"
-     "\xEB\xA9\xB3\x00\xEB\xA9\xB4\x00\xEB\xA9\xB5\x00\xEB\xA9\xB6\x00\xEB\xA9"
-     "\xB7\x00"
-     "\xEB\xA9\xB8\x00\xEB\xA9\xB9\x00\xEB\xA9\xBA\x00\xEB\xA9\xBB\x00\xEB\xA9"
-     "\xBC\x00"
-     "\xEB\xA9\xBD\x00\xEB\xA9\xBE\x00\xEB\xA9\xBF\x00\xEB\xAA\x80\x00\xEB\xAA"
-     "\x81\x00"
-     "\xEB\xAA\x82\x00\xEB\xAA\x83\x00\xEB\xAA\x84\x00\xEB\xAA\x85\x00\xEB\xAA"
-     "\x86\x00"
-     "\xEB\xAA\x87\x00\xEB\xAA\x88\x00\xEB\xAA\x89\x00\xEB\xAA\x8A\x00\xEB\xAA"
-     "\x8B\x00"
-     "\xEB\xAA\x8C\x00\xEB\xAA\x8D\x00\xEB\xAA\x8E\x00\xEB\xAA\x8F\x00\xEB\xAA"
-     "\x90\x00"
-     "\xEB\xAA\x91\x00\xEB\xAA\x92\x00\xEB\xAA\x93\x00\xEB\xAA\x94\x00\xEB\xAA"
-     "\x95\x00"
-     "\xEB\xAA\x96\x00\xEB\xAA\x97\x00\xEB\xAA\x98\x00\xEB\xAA\x99\x00\xEB\xAA"
-     "\x9A\x00"
-     "\xEB\xAA\x9B\x00\xEB\xAA\x9C\x00\xEB\xAA\x9D\x00\xEB\xAA\x9E\x00\xEB\xAA"
-     "\x9F\x00"
-     "\xEB\xAA\xA0\x00\xEB\xAA\xA1\x00\xEB\xAA\xA2\x00\xEB\xAA\xA3\x00\xEB\xAA"
-     "\xA4\x00"
-     "\xEB\xAA\xA5\x00\xEB\xAA\xA6\x00\xEB\xAA\xA7\x00\xEB\xAA\xA8\x00\xEB\xAA"
-     "\xA9\x00"
-     "\xEB\xAA\xAA\x00\xEB\xAA\xAB\x00\xEB\xAA\xAC\x00\xEB\xAA\xAD\x00\xEB\xAA"
-     "\xAE\x00"
-     "\xEB\xAA\xAF\x00\xEB\xAA\xB0\x00\xEB\xAA\xB1\x00\xEB\xAA\xB2\x00\xEB\xAA"
-     "\xB3\x00"
-     "\xEB\xAA\xB4\x00\xEB\xAA\xB5\x00\xEB\xAA\xB6\x00\xEB\xAA\xB7\x00\xEB\xAA"
-     "\xB8\x00"
-     "\xEB\xAA\xB9\x00\xEB\xAA\xBA\x00\xEB\xAA\xBB\x00\xEB\xAA\xBC\x00\xEB\xAA"
-     "\xBD\x00"
-     "\xEB\xAA\xBE\x00\xEB\xAA\xBF\x00\xEB\xAB\x80\x00\xEB\xAB\x81\x00\xEB\xAB"
-     "\x82\x00"
-     "\xEB\xAB\x83\x00\xEB\xAB\x84\x00\xEB\xAB\x85\x00\xEB\xAB\x86\x00\xEB\xAB"
-     "\x87\x00"
-     "\xEB\xAB\x88\x00\xEB\xAB\x89\x00\xEB\xAB\x8A\x00\xEB\xAB\x8B\x00\xEB\xAB"
-     "\x8C\x00"
-     "\xEB\xAB\x8D\x00\xEB\xAB\x8E\x00\xEB\xAB\x8F\x00\xEB\xAB\x90\x00\xEB\xAB"
-     "\x91\x00"
-     "\xEB\xAB\x92\x00\xEB\xAB\x93\x00\xEB\xAB\x94\x00\xEB\xAB\x95\x00\xEB\xAB"
-     "\x96\x00"
-     "\xEB\xAB\x97\x00\xEB\xAB\x98\x00\xEB\xAB\x99\x00\xEB\xAB\x9A\x00\xEB\xAB"
-     "\x9B\x00"
-     "\xEB\xAB\x9C\x00\xEB\xAB\x9D\x00\xEB\xAB\x9E\x00\xEB\xAB\x9F\x00\xEB\xAB"
-     "\xA0\x00"
-     "\xEB\xAB\xA1\x00\xEB\xAB\xA2\x00\xEB\xAB\xA3\x00\xEB\xAB\xA4\x00\xEB\xAB"
-     "\xA5\x00"
-     "\xEB\xAB\xA6\x00\xEB\xAB\xA7\x00\xEB\xAB\xA8\x00\xEB\xAB\xA9\x00\xEB\xAB"
-     "\xAA\x00"
-     "\xEB\xAB\xAB\x00\xEB\xAB\xAC\x00\xEB\xAB\xAD\x00\xEB\xAB\xAE\x00\xEB\xAB"
-     "\xAF\x00"
-     "\xEB\xAB\xB0\x00\xEB\xAB\xB1\x00\xEB\xAB\xB2\x00\xEB\xAB\xB3\x00\xEB\xAB"
-     "\xB4\x00"
-     "\xEB\xAB\xB5\x00\xEB\xAB\xB6\x00\xEB\xAB\xB7\x00\xEB\xAB\xB8\x00\xEB\xAB"
-     "\xB9\x00"
-     "\xEB\xAB\xBA\x00\xEB\xAB\xBB\x00\xEB\xAB\xBC\x00\xEB\xAB\xBD\x00\xEB\xAB"
-     "\xBE\x00"
-     "\xEB\xAB\xBF\x00\xEB\xAC\x80\x00\xEB\xAC\x81\x00\xEB\xAC\x82\x00\xEB\xAC"
-     "\x83\x00"
-     "\xEB\xAC\x84\x00\xEB\xAC\x85\x00\xEB\xAC\x86\x00\xEB\xAC\x87\x00\xEB\xAC"
-     "\x88\x00"
-     "\xEB\xAC\x89\x00\xEB\xAC\x8A\x00\xEB\xAC\x8B\x00\xEB\xAC\x8C\x00\xEB\xAC"
-     "\x8D\x00"
-     "\xEB\xAC\x8E\x00\xEB\xAC\x8F\x00\xEB\xAC\x90\x00\xEB\xAC\x91\x00\xEB\xAC"
-     "\x92\x00"
-     "\xEB\xAC\x93\x00\xEB\xAC\x94\x00\xEB\xAC\x95\x00\xEB\xAC\x96\x00\xEB\xAC"
-     "\x97\x00"
-     "\xEB\xAC\x98\x00\xEB\xAC\x99\x00\xEB\xAC\x9A\x00\xEB\xAC\x9B\x00\xEB\xAC"
-     "\x9C\x00"
-     "\xEB\xAC\x9D\x00\xEB\xAC\x9E\x00\xEB\xAC\x9F\x00\xEB\xAC\xA0\x00\xEB\xAC"
-     "\xA1\x00"
-     "\xEB\xAC\xA2\x00\xEB\xAC\xA3\x00\xEB\xAC\xA4\x00\xEB\xAC\xA5\x00\xEB\xAC"
-     "\xA6\x00"
-     "\xEB\xAC\xA7\x00\xEB\xAC\xA8\x00\xEB\xAC\xA9\x00\xEB\xAC\xAA\x00\xEB\xAC"
-     "\xAB\x00"
-     "\xEB\xAC\xAC\x00\xEB\xAC\xAD\x00\xEB\xAC\xAE\x00\xEB\xAC\xAF\x00\xEB\xAC"
-     "\xB0\x00"
-     "\xEB\xAC\xB1\x00\xEB\xAC\xB2\x00\xEB\xAC\xB3\x00\xEB\xAC\xB4\x00\xEB\xAC"
-     "\xB5\x00"
-     "\xEB\xAC\xB6\x00\xEB\xAC\xB7\x00\xEB\xAC\xB8\x00\xEB\xAC\xB9\x00\xEB\xAC"
-     "\xBA\x00"
-     "\xEB\xAC\xBB\x00\xEB\xAC\xBC\x00\xEB\xAC\xBD\x00\xEB\xAC\xBE\x00\xEB\xAC"
-     "\xBF\x00"
-     "\xEB\xAD\x80\x00\xEB\xAD\x81\x00\xEB\xAD\x82\x00\xEB\xAD\x83\x00\xEB\xAD"
-     "\x84\x00"
-     "\xEB\xAD\x85\x00\xEB\xAD\x86\x00\xEB\xAD\x87\x00\xEB\xAD\x88\x00\xEB\xAD"
-     "\x89\x00"
-     "\xEB\xAD\x8A\x00\xEB\xAD\x8B\x00\xEB\xAD\x8C\x00\xEB\xAD\x8D\x00\xEB\xAD"
-     "\x8E\x00"
-     "\xEB\xAD\x8F\x00\xEB\xAD\x90\x00\xEB\xAD\x91\x00\xEB\xAD\x92\x00\xEB\xAD"
-     "\x93\x00"
-     "\xEB\xAD\x94\x00\xEB\xAD\x95\x00\xEB\xAD\x96\x00\xEB\xAD\x97\x00\xEB\xAD"
-     "\x98\x00"
-     "\xEB\xAD\x99\x00\xEB\xAD\x9A\x00\xEB\xAD\x9B\x00\xEB\xAD\x9C\x00\xEB\xAD"
-     "\x9D\x00"
-     "\xEB\xAD\x9E\x00\xEB\xAD\x9F\x00\xEB\xAD\xA0\x00\xEB\xAD\xA1\x00\xEB\xAD"
-     "\xA2\x00"
-     "\xEB\xAD\xA3\x00\xEB\xAD\xA4\x00\xEB\xAD\xA5\x00\xEB\xAD\xA6\x00\xEB\xAD"
-     "\xA7\x00"
-     "\xEB\xAD\xA8\x00\xEB\xAD\xA9\x00\xEB\xAD\xAA\x00\xEB\xAD\xAB\x00\xEB\xAD"
-     "\xAC\x00"
-     "\xEB\xAD\xAD\x00\xEB\xAD\xAE\x00\xEB\xAD\xAF\x00\xEB\xAD\xB0\x00\xEB\xAD"
-     "\xB1\x00"
-     "\xEB\xAD\xB2\x00\xEB\xAD\xB3\x00\xEB\xAD\xB4\x00\xEB\xAD\xB5\x00\xEB\xAD"
-     "\xB6\x00"
-     "\xEB\xAD\xB7\x00\xEB\xAD\xB8\x00\xEB\xAD\xB9\x00\xEB\xAD\xBA\x00\xEB\xAD"
-     "\xBB\x00"
-     "\xEB\xAD\xBC\x00\xEB\xAD\xBD\x00\xEB\xAD\xBE\x00\xEB\xAD\xBF\x00\xEB\xAE"
-     "\x80\x00"
-     "\xEB\xAE\x81\x00\xEB\xAE\x82\x00\xEB\xAE\x83\x00\xEB\xAE\x84\x00\xEB\xAE"
-     "\x85\x00"
-     "\xEB\xAE\x86\x00\xEB\xAE\x87\x00\xEB\xAE\x88\x00\xEB\xAE\x89\x00\xEB\xAE"
-     "\x8A\x00"
-     "\xEB\xAE\x8B\x00\xEB\xAE\x8C\x00\xEB\xAE\x8D\x00\xEB\xAE\x8E\x00\xEB\xAE"
-     "\x8F\x00"
-     "\xEB\xAE\x90\x00\xEB\xAE\x91\x00\xEB\xAE\x92\x00\xEB\xAE\x93\x00\xEB\xAE"
-     "\x94\x00"
-     "\xEB\xAE\x95\x00\xEB\xAE\x96\x00\xEB\xAE\x97\x00\xEB\xAE\x98\x00\xEB\xAE"
-     "\x99\x00"
-     "\xEB\xAE\x9A\x00\xEB\xAE\x9B\x00\xEB\xAE\x9C\x00\xEB\xAE\x9D\x00\xEB\xAE"
-     "\x9E\x00"
-     "\xEB\xAE\x9F\x00\xEB\xAE\xA0\x00\xEB\xAE\xA1\x00\xEB\xAE\xA2\x00\xEB\xAE"
-     "\xA3\x00"
-     "\xEB\xAE\xA4\x00\xEB\xAE\xA5\x00\xEB\xAE\xA6\x00\xEB\xAE\xA7\x00\xEB\xAE"
-     "\xA8\x00"
-     "\xEB\xAE\xA9\x00\xEB\xAE\xAA\x00\xEB\xAE\xAB\x00\xEB\xAE\xAC\x00\xEB\xAE"
-     "\xAD\x00"
-     "\xEB\xAE\xAE\x00\xEB\xAE\xAF\x00\xEB\xAE\xB0\x00\xEB\xAE\xB1\x00\xEB\xAE"
-     "\xB2\x00"
-     "\xEB\xAE\xB3\x00\xEB\xAE\xB4\x00\xEB\xAE\xB5\x00\xEB\xAE\xB6\x00\xEB\xAE"
-     "\xB7\x00"
-     "\xEB\xAE\xB8\x00\xEB\xAE\xB9\x00\xEB\xAE\xBA\x00\xEB\xAE\xBB\x00\xEB\xAE"
-     "\xBC\x00"
-     "\xEB\xAE\xBD\x00\xEB\xAE\xBE\x00\xEB\xAE\xBF\x00\xEB\xAF\x80\x00\xEB\xAF"
-     "\x81\x00"
-     "\xEB\xAF\x82\x00\xEB\xAF\x83\x00\xEB\xAF\x84\x00\xEB\xAF\x85\x00\xEB\xAF"
-     "\x86\x00"
-     "\xEB\xAF\x87\x00\xEB\xAF\x88\x00\xEB\xAF\x89\x00\xEB\xAF\x8A\x00\xEB\xAF"
-     "\x8B\x00"
-     "\xEB\xAF\x8C\x00\xEB\xAF\x8D\x00\xEB\xAF\x8E\x00\xEB\xAF\x8F\x00\xEB\xAF"
-     "\x90\x00"
-     "\xEB\xAF\x91\x00\xEB\xAF\x92\x00\xEB\xAF\x93\x00\xEB\xAF\x94\x00\xEB\xAF"
-     "\x95\x00"
-     "\xEB\xAF\x96\x00\xEB\xAF\x97\x00\xEB\xAF\x98\x00\xEB\xAF\x99\x00\xEB\xAF"
-     "\x9A\x00"
-     "\xEB\xAF\x9B\x00\xEB\xAF\x9C\x00\xEB\xAF\x9D\x00\xEB\xAF\x9E\x00\xEB\xAF"
-     "\x9F\x00"
-     "\xEB\xAF\xA0\x00\xEB\xAF\xA1\x00\xEB\xAF\xA2\x00\xEB\xAF\xA3\x00\xEB\xAF"
-     "\xA4\x00"
-     "\xEB\xAF\xA5\x00\xEB\xAF\xA6\x00\xEB\xAF\xA7\x00\xEB\xAF\xA8\x00\xEB\xAF"
-     "\xA9\x00"
-     "\xEB\xAF\xAA\x00\xEB\xAF\xAB\x00\xEB\xAF\xAC\x00\xEB\xAF\xAD\x00\xEB\xAF"
-     "\xAE\x00"
-     "\xEB\xAF\xAF\x00\xEB\xAF\xB0\x00\xEB\xAF\xB1\x00\xEB\xAF\xB2\x00\xEB\xAF"
-     "\xB3\x00"
-     "\xEB\xAF\xB4\x00\xEB\xAF\xB5\x00\xEB\xAF\xB6\x00\xEB\xAF\xB7\x00\xEB\xAF"
-     "\xB8\x00"
-     "\xEB\xAF\xB9\x00\xEB\xAF\xBA\x00\xEB\xAF\xBB\x00\xEB\xAF\xBC\x00\xEB\xAF"
-     "\xBD\x00"
-     "\xEB\xAF\xBE\x00\xEB\xAF\xBF\x00\xEB\xB0\x80\x00\xEB\xB0\x81\x00\xEB\xB0"
-     "\x82\x00"
-     "\xEB\xB0\x83\x00\xEB\xB0\x84\x00\xEB\xB0\x85\x00\xEB\xB0\x86\x00\xEB\xB0"
-     "\x87\x00"
-     "\xEB\xB0\x88\x00\xEB\xB0\x89\x00\xEB\xB0\x8A\x00\xEB\xB0\x8B\x00\xEB\xB0"
-     "\x8C\x00"
-     "\xEB\xB0\x8D\x00\xEB\xB0\x8E\x00\xEB\xB0\x8F\x00\xEB\xB0\x90\x00\xEB\xB0"
-     "\x91\x00"
-     "\xEB\xB0\x92\x00\xEB\xB0\x93\x00\xEB\xB0\x94\x00\xEB\xB0\x95\x00\xEB\xB0"
-     "\x96\x00"
-     "\xEB\xB0\x97\x00\xEB\xB0\x98\x00\xEB\xB0\x99\x00\xEB\xB0\x9A\x00\xEB\xB0"
-     "\x9B\x00"
-     "\xEB\xB0\x9C\x00\xEB\xB0\x9D\x00\xEB\xB0\x9E\x00\xEB\xB0\x9F\x00\xEB\xB0"
-     "\xA0\x00"
-     "\xEB\xB0\xA1\x00\xEB\xB0\xA2\x00\xEB\xB0\xA3\x00\xEB\xB0\xA4\x00\xEB\xB0"
-     "\xA5\x00"
-     "\xEB\xB0\xA6\x00\xEB\xB0\xA7\x00\xEB\xB0\xA8\x00\xEB\xB0\xA9\x00\xEB\xB0"
-     "\xAA\x00"
-     "\xEB\xB0\xAB\x00\xEB\xB0\xAC\x00\xEB\xB0\xAD\x00\xEB\xB0\xAE\x00\xEB\xB0"
-     "\xAF\x00"
-     "\xEB\xB0\xB0\x00\xEB\xB0\xB1\x00\xEB\xB0\xB2\x00\xEB\xB0\xB3\x00\xEB\xB0"
-     "\xB4\x00"
-     "\xEB\xB0\xB5\x00\xEB\xB0\xB6\x00\xEB\xB0\xB7\x00\xEB\xB0\xB8\x00\xEB\xB0"
-     "\xB9\x00"
-     "\xEB\xB0\xBA\x00\xEB\xB0\xBB\x00\xEB\xB0\xBC\x00\xEB\xB0\xBD\x00\xEB\xB0"
-     "\xBE\x00"
-     "\xEB\xB0\xBF\x00\xEB\xB1\x80\x00\xEB\xB1\x81\x00\xEB\xB1\x82\x00\xEB\xB1"
-     "\x83\x00"
-     "\xEB\xB1\x84\x00\xEB\xB1\x85\x00\xEB\xB1\x86\x00\xEB\xB1\x87\x00\xEB\xB1"
-     "\x88\x00"
-     "\xEB\xB1\x89\x00\xEB\xB1\x8A\x00\xEB\xB1\x8B\x00\xEB\xB1\x8C\x00\xEB\xB1"
-     "\x8D\x00"
-     "\xEB\xB1\x8E\x00\xEB\xB1\x8F\x00\xEB\xB1\x90\x00\xEB\xB1\x91\x00\xEB\xB1"
-     "\x92\x00"
-     "\xEB\xB1\x93\x00\xEB\xB1\x94\x00\xEB\xB1\x95\x00\xEB\xB1\x96\x00\xEB\xB1"
-     "\x97\x00"
-     "\xEB\xB1\x98\x00\xEB\xB1\x99\x00\xEB\xB1\x9A\x00\xEB\xB1\x9B\x00\xEB\xB1"
-     "\x9C\x00"
-     "\xEB\xB1\x9D\x00\xEB\xB1\x9E\x00\xEB\xB1\x9F\x00\xEB\xB1\xA0\x00\xEB\xB1"
-     "\xA1\x00"
-     "\xEB\xB1\xA2\x00\xEB\xB1\xA3\x00\xEB\xB1\xA4\x00\xEB\xB1\xA5\x00\xEB\xB1"
-     "\xA6\x00"
-     "\xEB\xB1\xA7\x00\xEB\xB1\xA8\x00\xEB\xB1\xA9\x00\xEB\xB1\xAA\x00\xEB\xB1"
-     "\xAB\x00"
-     "\xEB\xB1\xAC\x00\xEB\xB1\xAD\x00\xEB\xB1\xAE\x00\xEB\xB1\xAF\x00\xEB\xB1"
-     "\xB0\x00"
-     "\xEB\xB1\xB1\x00\xEB\xB1\xB2\x00\xEB\xB1\xB3\x00\xEB\xB1\xB4\x00\xEB\xB1"
-     "\xB5\x00"
-     "\xEB\xB1\xB6\x00\xEB\xB1\xB7\x00\xEB\xB1\xB8\x00\xEB\xB1\xB9\x00\xEB\xB1"
-     "\xBA\x00"
-     "\xEB\xB1\xBB\x00\xEB\xB1\xBC\x00\xEB\xB1\xBD\x00\xEB\xB1\xBE\x00\xEB\xB1"
-     "\xBF\x00"
-     "\xEB\xB2\x80\x00\xEB\xB2\x81\x00\xEB\xB2\x82\x00\xEB\xB2\x83\x00\xEB\xB2"
-     "\x84\x00"
-     "\xEB\xB2\x85\x00\xEB\xB2\x86\x00\xEB\xB2\x87\x00\xEB\xB2\x88\x00\xEB\xB2"
-     "\x89\x00"
-     "\xEB\xB2\x8A\x00\xEB\xB2\x8B\x00\xEB\xB2\x8C\x00\xEB\xB2\x8D\x00\xEB\xB2"
-     "\x8E\x00"
-     "\xEB\xB2\x8F\x00\xEB\xB2\x90\x00\xEB\xB2\x91\x00\xEB\xB2\x92\x00\xEB\xB2"
-     "\x93\x00"
-     "\xEB\xB2\x94\x00\xEB\xB2\x95\x00\xEB\xB2\x96\x00\xEB\xB2\x97\x00\xEB\xB2"
-     "\x98\x00"
-     "\xEB\xB2\x99\x00\xEB\xB2\x9A\x00\xEB\xB2\x9B\x00\xEB\xB2\x9C\x00\xEB\xB2"
-     "\x9D\x00"
-     "\xEB\xB2\x9E\x00\xEB\xB2\x9F\x00\xEB\xB2\xA0\x00\xEB\xB2\xA1\x00\xEB\xB2"
-     "\xA2\x00"
-     "\xEB\xB2\xA3\x00\xEB\xB2\xA4\x00\xEB\xB2\xA5\x00\xEB\xB2\xA6\x00\xEB\xB2"
-     "\xA7\x00"
-     "\xEB\xB2\xA8\x00\xEB\xB2\xA9\x00\xEB\xB2\xAA\x00\xEB\xB2\xAB\x00\xEB\xB2"
-     "\xAC\x00"
-     "\xEB\xB2\xAD\x00\xEB\xB2\xAE\x00\xEB\xB2\xAF\x00\xEB\xB2\xB0\x00\xEB\xB2"
-     "\xB1\x00"
-     "\xEB\xB2\xB2\x00\xEB\xB2\xB3\x00\xEB\xB2\xB4\x00\xEB\xB2\xB5\x00\xEB\xB2"
-     "\xB6\x00"
-     "\xEB\xB2\xB7\x00\xEB\xB2\xB8\x00\xEB\xB2\xB9\x00\xEB\xB2\xBA\x00\xEB\xB2"
-     "\xBB\x00"
-     "\xEB\xB2\xBC\x00\xEB\xB2\xBD\x00\xEB\xB2\xBE\x00\xEB\xB2\xBF\x00\xEB\xB3"
-     "\x80\x00"
-     "\xEB\xB3\x81\x00\xEB\xB3\x82\x00\xEB\xB3\x83\x00\xEB\xB3\x84\x00\xEB\xB3"
-     "\x85\x00"
-     "\xEB\xB3\x86\x00\xEB\xB3\x87\x00\xEB\xB3\x88\x00\xEB\xB3\x89\x00\xEB\xB3"
-     "\x8A\x00"
-     "\xEB\xB3\x8B\x00\xEB\xB3\x8C\x00\xEB\xB3\x8D\x00\xEB\xB3\x8E\x00\xEB\xB3"
-     "\x8F\x00"
-     "\xEB\xB3\x90\x00\xEB\xB3\x91\x00\xEB\xB3\x92\x00\xEB\xB3\x93\x00\xEB\xB3"
-     "\x94\x00"
-     "\xEB\xB3\x95\x00\xEB\xB3\x96\x00\xEB\xB3\x97\x00\xEB\xB3\x98\x00\xEB\xB3"
-     "\x99\x00"
-     "\xEB\xB3\x9A\x00\xEB\xB3\x9B\x00\xEB\xB3\x9C\x00\xEB\xB3\x9D\x00\xEB\xB3"
-     "\x9E\x00"
-     "\xEB\xB3\x9F\x00\xEB\xB3\xA0\x00\xEB\xB3\xA1\x00\xEB\xB3\xA2\x00\xEB\xB3"
-     "\xA3\x00"
-     "\xEB\xB3\xA4\x00\xEB\xB3\xA5\x00\xEB\xB3\xA6\x00\xEB\xB3\xA7\x00\xEB\xB3"
-     "\xA8\x00"
-     "\xEB\xB3\xA9\x00\xEB\xB3\xAA\x00\xEB\xB3\xAB\x00\xEB\xB3\xAC\x00\xEB\xB3"
-     "\xAD\x00"
-     "\xEB\xB3\xAE\x00\xEB\xB3\xAF\x00\xEB\xB3\xB0\x00\xEB\xB3\xB1\x00\xEB\xB3"
-     "\xB2\x00"
-     "\xEB\xB3\xB3\x00\xEB\xB3\xB4\x00\xEB\xB3\xB5\x00\xEB\xB3\xB6\x00\xEB\xB3"
-     "\xB7\x00"
-     "\xEB\xB3\xB8\x00\xEB\xB3\xB9\x00\xEB\xB3\xBA\x00\xEB\xB3\xBB\x00\xEB\xB3"
-     "\xBC\x00"
-     "\xEB\xB3\xBD\x00\xEB\xB3\xBE\x00\xEB\xB3\xBF\x00\xEB\xB4\x80\x00\xEB\xB4"
-     "\x81\x00"
-     "\xEB\xB4\x82\x00\xEB\xB4\x83\x00\xEB\xB4\x84\x00\xEB\xB4\x85\x00\xEB\xB4"
-     "\x86\x00"
-     "\xEB\xB4\x87\x00\xEB\xB4\x88\x00\xEB\xB4\x89\x00\xEB\xB4\x8A\x00\xEB\xB4"
-     "\x8B\x00"
-     "\xEB\xB4\x8C\x00\xEB\xB4\x8D\x00\xEB\xB4\x8E\x00\xEB\xB4\x8F\x00\xEB\xB4"
-     "\x90\x00"
-     "\xEB\xB4\x91\x00\xEB\xB4\x92\x00\xEB\xB4\x93\x00\xEB\xB4\x94\x00\xEB\xB4"
-     "\x95\x00"
-     "\xEB\xB4\x96\x00\xEB\xB4\x97\x00\xEB\xB4\x98\x00\xEB\xB4\x99\x00\xEB\xB4"
-     "\x9A\x00"
-     "\xEB\xB4\x9B\x00\xEB\xB4\x9C\x00\xEB\xB4\x9D\x00\xEB\xB4\x9E\x00\xEB\xB4"
-     "\x9F\x00"
-     "\xEB\xB4\xA0\x00\xEB\xB4\xA1\x00\xEB\xB4\xA2\x00\xEB\xB4\xA3\x00\xEB\xB4"
-     "\xA4\x00"
-     "\xEB\xB4\xA5\x00\xEB\xB4\xA6\x00\xEB\xB4\xA7\x00\xEB\xB4\xA8\x00\xEB\xB4"
-     "\xA9\x00"
-     "\xEB\xB4\xAA\x00\xEB\xB4\xAB\x00\xEB\xB4\xAC\x00\xEB\xB4\xAD\x00\xEB\xB4"
-     "\xAE\x00"
-     "\xEB\xB4\xAF\x00\xEB\xB4\xB0\x00\xEB\xB4\xB1\x00\xEB\xB4\xB2\x00\xEB\xB4"
-     "\xB3\x00"
-     "\xEB\xB4\xB4\x00\xEB\xB4\xB5\x00\xEB\xB4\xB6\x00\xEB\xB4\xB7\x00\xEB\xB4"
-     "\xB8\x00"
-     "\xEB\xB4\xB9\x00\xEB\xB4\xBA\x00\xEB\xB4\xBB\x00\xEB\xB4\xBC\x00\xEB\xB4"
-     "\xBD\x00"
-     "\xEB\xB4\xBE\x00\xEB\xB4\xBF\x00\xEB\xB5\x80\x00\xEB\xB5\x81\x00\xEB\xB5"
-     "\x82\x00"
-     "\xEB\xB5\x83\x00\xEB\xB5\x84\x00\xEB\xB5\x85\x00\xEB\xB5\x86\x00\xEB\xB5"
-     "\x87\x00"
-     "\xEB\xB5\x88\x00\xEB\xB5\x89\x00\xEB\xB5\x8A\x00\xEB\xB5\x8B\x00\xEB\xB5"
-     "\x8C\x00"
-     "\xEB\xB5\x8D\x00\xEB\xB5\x8E\x00\xEB\xB5\x8F\x00\xEB\xB5\x90\x00\xEB\xB5"
-     "\x91\x00"
-     "\xEB\xB5\x92\x00\xEB\xB5\x93\x00\xEB\xB5\x94\x00\xEB\xB5\x95\x00\xEB\xB5"
-     "\x96\x00"
-     "\xEB\xB5\x97\x00\xEB\xB5\x98\x00\xEB\xB5\x99\x00\xEB\xB5\x9A\x00\xEB\xB5"
-     "\x9B\x00"
-     "\xEB\xB5\x9C\x00\xEB\xB5\x9D\x00\xEB\xB5\x9E\x00\xEB\xB5\x9F\x00\xEB\xB5"
-     "\xA0\x00"
-     "\xEB\xB5\xA1\x00\xEB\xB5\xA2\x00\xEB\xB5\xA3\x00\xEB\xB5\xA4\x00\xEB\xB5"
-     "\xA5\x00"
-     "\xEB\xB5\xA6\x00\xEB\xB5\xA7\x00\xEB\xB5\xA8\x00\xEB\xB5\xA9\x00\xEB\xB5"
-     "\xAA\x00"
-     "\xEB\xB5\xAB\x00\xEB\xB5\xAC\x00\xEB\xB5\xAD\x00\xEB\xB5\xAE\x00\xEB\xB5"
-     "\xAF\x00"
-     "\xEB\xB5\xB0\x00\xEB\xB5\xB1\x00\xEB\xB5\xB2\x00\xEB\xB5\xB3\x00\xEB\xB5"
-     "\xB4\x00"
-     "\xEB\xB5\xB5\x00\xEB\xB5\xB6\x00\xEB\xB5\xB7\x00\xEB\xB5\xB8\x00\xEB\xB5"
-     "\xB9\x00"
-     "\xEB\xB5\xBA\x00\xEB\xB5\xBB\x00\xEB\xB5\xBC\x00\xEB\xB5\xBD\x00\xEB\xB5"
-     "\xBE\x00"
-     "\xEB\xB5\xBF\x00\xEB\xB6\x80\x00\xEB\xB6\x81\x00\xEB\xB6\x82\x00\xEB\xB6"
-     "\x83\x00"
-     "\xEB\xB6\x84\x00\xEB\xB6\x85\x00\xEB\xB6\x86\x00\xEB\xB6\x87\x00\xEB\xB6"
-     "\x88\x00"
-     "\xEB\xB6\x89\x00\xEB\xB6\x8A\x00\xEB\xB6\x8B\x00\xEB\xB6\x8C\x00\xEB\xB6"
-     "\x8D\x00"
-     "\xEB\xB6\x8E\x00\xEB\xB6\x8F\x00\xEB\xB6\x90\x00\xEB\xB6\x91\x00\xEB\xB6"
-     "\x92\x00"
-     "\xEB\xB6\x93\x00\xEB\xB6\x94\x00\xEB\xB6\x95\x00\xEB\xB6\x96\x00\xEB\xB6"
-     "\x97\x00"
-     "\xEB\xB6\x98\x00\xEB\xB6\x99\x00\xEB\xB6\x9A\x00\xEB\xB6\x9B\x00\xEB\xB6"
-     "\x9C\x00"
-     "\xEB\xB6\x9D\x00\xEB\xB6\x9E\x00\xEB\xB6\x9F\x00\xEB\xB6\xA0\x00\xEB\xB6"
-     "\xA1\x00"
-     "\xEB\xB6\xA2\x00\xEB\xB6\xA3\x00\xEB\xB6\xA4\x00\xEB\xB6\xA5\x00\xEB\xB6"
-     "\xA6\x00"
-     "\xEB\xB6\xA7\x00\xEB\xB6\xA8\x00\xEB\xB6\xA9\x00\xEB\xB6\xAA\x00\xEB\xB6"
-     "\xAB\x00"
-     "\xEB\xB6\xAC\x00\xEB\xB6\xAD\x00\xEB\xB6\xAE\x00\xEB\xB6\xAF\x00\xEB\xB6"
-     "\xB0\x00"
-     "\xEB\xB6\xB1\x00\xEB\xB6\xB2\x00\xEB\xB6\xB3\x00\xEB\xB6\xB4\x00\xEB\xB6"
-     "\xB5\x00"
-     "\xEB\xB6\xB6\x00\xEB\xB6\xB7\x00\xEB\xB6\xB8\x00\xEB\xB6\xB9\x00\xEB\xB6"
-     "\xBA\x00"
-     "\xEB\xB6\xBB\x00\xEB\xB6\xBC\x00\xEB\xB6\xBD\x00\xEB\xB6\xBE\x00\xEB\xB6"
-     "\xBF\x00"
-     "\xEB\xB7\x80\x00\xEB\xB7\x81\x00\xEB\xB7\x82\x00\xEB\xB7\x83\x00\xEB\xB7"
-     "\x84\x00"
-     "\xEB\xB7\x85\x00\xEB\xB7\x86\x00\xEB\xB7\x87\x00\xEB\xB7\x88\x00\xEB\xB7"
-     "\x89\x00"
-     "\xEB\xB7\x8A\x00\xEB\xB7\x8B\x00\xEB\xB7\x8C\x00\xEB\xB7\x8D\x00\xEB\xB7"
-     "\x8E\x00"
-     "\xEB\xB7\x8F\x00\xEB\xB7\x90\x00\xEB\xB7\x91\x00\xEB\xB7\x92\x00\xEB\xB7"
-     "\x93\x00"
-     "\xEB\xB7\x94\x00\xEB\xB7\x95\x00\xEB\xB7\x96\x00\xEB\xB7\x97\x00\xEB\xB7"
-     "\x98\x00"
-     "\xEB\xB7\x99\x00\xEB\xB7\x9A\x00\xEB\xB7\x9B\x00\xEB\xB7\x9C\x00\xEB\xB7"
-     "\x9D\x00"
-     "\xEB\xB7\x9E\x00\xEB\xB7\x9F\x00\xEB\xB7\xA0\x00\xEB\xB7\xA1\x00\xEB\xB7"
-     "\xA2\x00"
-     "\xEB\xB7\xA3\x00\xEB\xB7\xA4\x00\xEB\xB7\xA5\x00\xEB\xB7\xA6\x00\xEB\xB7"
-     "\xA7\x00"
-     "\xEB\xB7\xA8\x00\xEB\xB7\xA9\x00\xEB\xB7\xAA\x00\xEB\xB7\xAB\x00\xEB\xB7"
-     "\xAC\x00"
-     "\xEB\xB7\xAD\x00\xEB\xB7\xAE\x00\xEB\xB7\xAF\x00\xEB\xB7\xB0\x00\xEB\xB7"
-     "\xB1\x00"
-     "\xEB\xB7\xB2\x00\xEB\xB7\xB3\x00\xEB\xB7\xB4\x00\xEB\xB7\xB5\x00\xEB\xB7"
-     "\xB6\x00"
-     "\xEB\xB7\xB7\x00\xEB\xB7\xB8\x00\xEB\xB7\xB9\x00\xEB\xB7\xBA\x00\xEB\xB7"
-     "\xBB\x00"
-     "\xEB\xB7\xBC\x00\xEB\xB7\xBD\x00\xEB\xB7\xBE\x00\xEB\xB7\xBF\x00\xEB\xB8"
-     "\x80\x00"
-     "\xEB\xB8\x81\x00\xEB\xB8\x82\x00\xEB\xB8\x83\x00\xEB\xB8\x84\x00\xEB\xB8"
-     "\x85\x00"
-     "\xEB\xB8\x86\x00\xEB\xB8\x87\x00\xEB\xB8\x88\x00\xEB\xB8\x89\x00\xEB\xB8"
-     "\x8A\x00"
-     "\xEB\xB8\x8B\x00\xEB\xB8\x8C\x00\xEB\xB8\x8D\x00\xEB\xB8\x8E\x00\xEB\xB8"
-     "\x8F\x00"
-     "\xEB\xB8\x90\x00\xEB\xB8\x91\x00\xEB\xB8\x92\x00\xEB\xB8\x93\x00\xEB\xB8"
-     "\x94\x00"
-     "\xEB\xB8\x95\x00\xEB\xB8\x96\x00\xEB\xB8\x97\x00\xEB\xB8\x98\x00\xEB\xB8"
-     "\x99\x00"
-     "\xEB\xB8\x9A\x00\xEB\xB8\x9B\x00\xEB\xB8\x9C\x00\xEB\xB8\x9D\x00\xEB\xB8"
-     "\x9E\x00"
-     "\xEB\xB8\x9F\x00\xEB\xB8\xA0\x00\xEB\xB8\xA1\x00\xEB\xB8\xA2\x00\xEB\xB8"
-     "\xA3\x00"
-     "\xEB\xB8\xA4\x00\xEB\xB8\xA5\x00\xEB\xB8\xA6\x00\xEB\xB8\xA7\x00\xEB\xB8"
-     "\xA8\x00"
-     "\xEB\xB8\xA9\x00\xEB\xB8\xAA\x00\xEB\xB8\xAB\x00\xEB\xB8\xAC\x00\xEB\xB8"
-     "\xAD\x00"
-     "\xEB\xB8\xAE\x00\xEB\xB8\xAF\x00\xEB\xB8\xB0\x00\xEB\xB8\xB1\x00\xEB\xB8"
-     "\xB2\x00"
-     "\xEB\xB8\xB3\x00\xEB\xB8\xB4\x00\xEB\xB8\xB5\x00\xEB\xB8\xB6\x00\xEB\xB8"
-     "\xB7\x00"
-     "\xEB\xB8\xB8\x00\xEB\xB8\xB9\x00\xEB\xB8\xBA\x00\xEB\xB8\xBB\x00\xEB\xB8"
-     "\xBC\x00"
-     "\xEB\xB8\xBD\x00\xEB\xB8\xBE\x00\xEB\xB8\xBF\x00\xEB\xB9\x80\x00\xEB\xB9"
-     "\x81\x00"
-     "\xEB\xB9\x82\x00\xEB\xB9\x83\x00\xEB\xB9\x84\x00\xEB\xB9\x85\x00\xEB\xB9"
-     "\x86\x00"
-     "\xEB\xB9\x87\x00\xEB\xB9\x88\x00\xEB\xB9\x89\x00\xEB\xB9\x8A\x00\xEB\xB9"
-     "\x8B\x00"
-     "\xEB\xB9\x8C\x00\xEB\xB9\x8D\x00\xEB\xB9\x8E\x00\xEB\xB9\x8F\x00\xEB\xB9"
-     "\x90\x00"
-     "\xEB\xB9\x91\x00\xEB\xB9\x92\x00\xEB\xB9\x93\x00\xEB\xB9\x94\x00\xEB\xB9"
-     "\x95\x00"
-     "\xEB\xB9\x96\x00\xEB\xB9\x97\x00\xEB\xB9\x98\x00\xEB\xB9\x99\x00\xEB\xB9"
-     "\x9A\x00"
-     "\xEB\xB9\x9B\x00\xEB\xB9\x9C\x00\xEB\xB9\x9D\x00\xEB\xB9\x9E\x00\xEB\xB9"
-     "\x9F\x00"
-     "\xEB\xB9\xA0\x00\xEB\xB9\xA1\x00\xEB\xB9\xA2\x00\xEB\xB9\xA3\x00\xEB\xB9"
-     "\xA4\x00"
-     "\xEB\xB9\xA5\x00\xEB\xB9\xA6\x00\xEB\xB9\xA7\x00\xEB\xB9\xA8\x00\xEB\xB9"
-     "\xA9\x00"
-     "\xEB\xB9\xAA\x00\xEB\xB9\xAB\x00\xEB\xB9\xAC\x00\xEB\xB9\xAD\x00\xEB\xB9"
-     "\xAE\x00"
-     "\xEB\xB9\xAF\x00\xEB\xB9\xB0\x00\xEB\xB9\xB1\x00\xEB\xB9\xB2\x00\xEB\xB9"
-     "\xB3\x00"
-     "\xEB\xB9\xB4\x00\xEB\xB9\xB5\x00\xEB\xB9\xB6\x00\xEB\xB9\xB7\x00\xEB\xB9"
-     "\xB8\x00"
-     "\xEB\xB9\xB9\x00\xEB\xB9\xBA\x00\xEB\xB9\xBB\x00\xEB\xB9\xBC\x00\xEB\xB9"
-     "\xBD\x00"
-     "\xEB\xB9\xBE\x00\xEB\xB9\xBF\x00\xEB\xBA\x80\x00\xEB\xBA\x81\x00\xEB\xBA"
-     "\x82\x00"
-     "\xEB\xBA\x83\x00\xEB\xBA\x84\x00\xEB\xBA\x85\x00\xEB\xBA\x86\x00\xEB\xBA"
-     "\x87\x00"
-     "\xEB\xBA\x88\x00\xEB\xBA\x89\x00\xEB\xBA\x8A\x00\xEB\xBA\x8B\x00\xEB\xBA"
-     "\x8C\x00"
-     "\xEB\xBA\x8D\x00\xEB\xBA\x8E\x00\xEB\xBA\x8F\x00\xEB\xBA\x90\x00\xEB\xBA"
-     "\x91\x00"
-     "\xEB\xBA\x92\x00\xEB\xBA\x93\x00\xEB\xBA\x94\x00\xEB\xBA\x95\x00\xEB\xBA"
-     "\x96\x00"
-     "\xEB\xBA\x97\x00\xEB\xBA\x98\x00\xEB\xBA\x99\x00\xEB\xBA\x9A\x00\xEB\xBA"
-     "\x9B\x00"
-     "\xEB\xBA\x9C\x00\xEB\xBA\x9D\x00\xEB\xBA\x9E\x00\xEB\xBA\x9F\x00\xEB\xBA"
-     "\xA0\x00"
-     "\xEB\xBA\xA1\x00\xEB\xBA\xA2\x00\xEB\xBA\xA3\x00\xEB\xBA\xA4\x00\xEB\xBA"
-     "\xA5\x00"
-     "\xEB\xBA\xA6\x00\xEB\xBA\xA7\x00\xEB\xBA\xA8\x00\xEB\xBA\xA9\x00\xEB\xBA"
-     "\xAA\x00"
-     "\xEB\xBA\xAB\x00\xEB\xBA\xAC\x00\xEB\xBA\xAD\x00\xEB\xBA\xAE\x00\xEB\xBA"
-     "\xAF\x00"
-     "\xEB\xBA\xB0\x00\xEB\xBA\xB1\x00\xEB\xBA\xB2\x00\xEB\xBA\xB3\x00\xEB\xBA"
-     "\xB4\x00"
-     "\xEB\xBA\xB5\x00\xEB\xBA\xB6\x00\xEB\xBA\xB7\x00\xEB\xBA\xB8\x00\xEB\xBA"
-     "\xB9\x00"
-     "\xEB\xBA\xBA\x00\xEB\xBA\xBB\x00\xEB\xBA\xBC\x00\xEB\xBA\xBD\x00\xEB\xBA"
-     "\xBE\x00"
-     "\xEB\xBA\xBF\x00\xEB\xBB\x80\x00\xEB\xBB\x81\x00\xEB\xBB\x82\x00\xEB\xBB"
-     "\x83\x00"
-     "\xEB\xBB\x84\x00\xEB\xBB\x85\x00\xEB\xBB\x86\x00\xEB\xBB\x87\x00\xEB\xBB"
-     "\x88\x00"
-     "\xEB\xBB\x89\x00\xEB\xBB\x8A\x00\xEB\xBB\x8B\x00\xEB\xBB\x8C\x00\xEB\xBB"
-     "\x8D\x00"
-     "\xEB\xBB\x8E\x00\xEB\xBB\x8F\x00\xEB\xBB\x90\x00\xEB\xBB\x91\x00\xEB\xBB"
-     "\x92\x00"
-     "\xEB\xBB\x93\x00\xEB\xBB\x94\x00\xEB\xBB\x95\x00\xEB\xBB\x96\x00\xEB\xBB"
-     "\x97\x00"
-     "\xEB\xBB\x98\x00\xEB\xBB\x99\x00\xEB\xBB\x9A\x00\xEB\xBB\x9B\x00\xEB\xBB"
-     "\x9C\x00"
-     "\xEB\xBB\x9D\x00\xEB\xBB\x9E\x00\xEB\xBB\x9F\x00\xEB\xBB\xA0\x00\xEB\xBB"
-     "\xA1\x00"
-     "\xEB\xBB\xA2\x00\xEB\xBB\xA3\x00\xEB\xBB\xA4\x00\xEB\xBB\xA5\x00\xEB\xBB"
-     "\xA6\x00"
-     "\xEB\xBB\xA7\x00\xEB\xBB\xA8\x00\xEB\xBB\xA9\x00\xEB\xBB\xAA\x00\xEB\xBB"
-     "\xAB\x00"
-     "\xEB\xBB\xAC\x00\xEB\xBB\xAD\x00\xEB\xBB\xAE\x00\xEB\xBB\xAF\x00\xEB\xBB"
-     "\xB0\x00"
-     "\xEB\xBB\xB1\x00\xEB\xBB\xB2\x00\xEB\xBB\xB3\x00\xEB\xBB\xB4\x00\xEB\xBB"
-     "\xB5\x00"
-     "\xEB\xBB\xB6\x00\xEB\xBB\xB7\x00\xEB\xBB\xB8\x00\xEB\xBB\xB9\x00\xEB\xBB"
-     "\xBA\x00"
-     "\xEB\xBB\xBB\x00\xEB\xBB\xBC\x00\xEB\xBB\xBD\x00\xEB\xBB\xBE\x00\xEB\xBB"
-     "\xBF\x00"
-     "\xEB\xBC\x80\x00\xEB\xBC\x81\x00\xEB\xBC\x82\x00\xEB\xBC\x83\x00\xEB\xBC"
-     "\x84\x00"
-     "\xEB\xBC\x85\x00\xEB\xBC\x86\x00\xEB\xBC\x87\x00\xEB\xBC\x88\x00\xEB\xBC"
-     "\x89\x00"
-     "\xEB\xBC\x8A\x00\xEB\xBC\x8B\x00\xEB\xBC\x8C\x00\xEB\xBC\x8D\x00\xEB\xBC"
-     "\x8E\x00"
-     "\xEB\xBC\x8F\x00\xEB\xBC\x90\x00\xEB\xBC\x91\x00\xEB\xBC\x92\x00\xEB\xBC"
-     "\x93\x00"
-     "\xEB\xBC\x94\x00\xEB\xBC\x95\x00\xEB\xBC\x96\x00\xEB\xBC\x97\x00\xEB\xBC"
-     "\x98\x00"
-     "\xEB\xBC\x99\x00\xEB\xBC\x9A\x00\xEB\xBC\x9B\x00\xEB\xBC\x9C\x00\xEB\xBC"
-     "\x9D\x00"
-     "\xEB\xBC\x9E\x00\xEB\xBC\x9F\x00\xEB\xBC\xA0\x00\xEB\xBC\xA1\x00\xEB\xBC"
-     "\xA2\x00"
-     "\xEB\xBC\xA3\x00\xEB\xBC\xA4\x00\xEB\xBC\xA5\x00\xEB\xBC\xA6\x00\xEB\xBC"
-     "\xA7\x00"
-     "\xEB\xBC\xA8\x00\xEB\xBC\xA9\x00\xEB\xBC\xAA\x00\xEB\xBC\xAB\x00\xEB\xBC"
-     "\xAC\x00"
-     "\xEB\xBC\xAD\x00\xEB\xBC\xAE\x00\xEB\xBC\xAF\x00\xEB\xBC\xB0\x00\xEB\xBC"
-     "\xB1\x00"
-     "\xEB\xBC\xB2\x00\xEB\xBC\xB3\x00\xEB\xBC\xB4\x00\xEB\xBC\xB5\x00\xEB\xBC"
-     "\xB6\x00"
-     "\xEB\xBC\xB7\x00\xEB\xBC\xB8\x00\xEB\xBC\xB9\x00\xEB\xBC\xBA\x00\xEB\xBC"
-     "\xBB\x00"
-     "\xEB\xBC\xBC\x00\xEB\xBC\xBD\x00\xEB\xBC\xBE\x00\xEB\xBC\xBF\x00\xEB\xBD"
-     "\x80\x00"
-     "\xEB\xBD\x81\x00\xEB\xBD\x82\x00\xEB\xBD\x83\x00\xEB\xBD\x84\x00\xEB\xBD"
-     "\x85\x00"
-     "\xEB\xBD\x86\x00\xEB\xBD\x87\x00\xEB\xBD\x88\x00\xEB\xBD\x89\x00\xEB\xBD"
-     "\x8A\x00"
-     "\xEB\xBD\x8B\x00\xEB\xBD\x8C\x00\xEB\xBD\x8D\x00\xEB\xBD\x8E\x00\xEB\xBD"
-     "\x8F\x00"
-     "\xEB\xBD\x90\x00\xEB\xBD\x91\x00\xEB\xBD\x92\x00\xEB\xBD\x93\x00\xEB\xBD"
-     "\x94\x00"
-     "\xEB\xBD\x95\x00\xEB\xBD\x96\x00\xEB\xBD\x97\x00\xEB\xBD\x98\x00\xEB\xBD"
-     "\x99\x00"
-     "\xEB\xBD\x9A\x00\xEB\xBD\x9B\x00\xEB\xBD\x9C\x00\xEB\xBD\x9D\x00\xEB\xBD"
-     "\x9E\x00"
-     "\xEB\xBD\x9F\x00\xEB\xBD\xA0\x00\xEB\xBD\xA1\x00\xEB\xBD\xA2\x00\xEB\xBD"
-     "\xA3\x00"
-     "\xEB\xBD\xA4\x00\xEB\xBD\xA5\x00\xEB\xBD\xA6\x00\xEB\xBD\xA7\x00\xEB\xBD"
-     "\xA8\x00"
-     "\xEB\xBD\xA9\x00\xEB\xBD\xAA\x00\xEB\xBD\xAB\x00\xEB\xBD\xAC\x00\xEB\xBD"
-     "\xAD\x00"
-     "\xEB\xBD\xAE\x00\xEB\xBD\xAF\x00\xEB\xBD\xB0\x00\xEB\xBD\xB1\x00\xEB\xBD"
-     "\xB2\x00"
-     "\xEB\xBD\xB3\x00\xEB\xBD\xB4\x00\xEB\xBD\xB5\x00\xEB\xBD\xB6\x00\xEB\xBD"
-     "\xB7\x00"
-     "\xEB\xBD\xB8\x00\xEB\xBD\xB9\x00\xEB\xBD\xBA\x00\xEB\xBD\xBB\x00\xEB\xBD"
-     "\xBC\x00"
-     "\xEB\xBD\xBD\x00\xEB\xBD\xBE\x00\xEB\xBD\xBF\x00\xEB\xBE\x80\x00\xEB\xBE"
-     "\x81\x00"
-     "\xEB\xBE\x82\x00\xEB\xBE\x83\x00\xEB\xBE\x84\x00\xEB\xBE\x85\x00\xEB\xBE"
-     "\x86\x00"
-     "\xEB\xBE\x87\x00\xEB\xBE\x88\x00\xEB\xBE\x89\x00\xEB\xBE\x8A\x00\xEB\xBE"
-     "\x8B\x00"
-     "\xEB\xBE\x8C\x00\xEB\xBE\x8D\x00\xEB\xBE\x8E\x00\xEB\xBE\x8F\x00\xEB\xBE"
-     "\x90\x00"
-     "\xEB\xBE\x91\x00\xEB\xBE\x92\x00\xEB\xBE\x93\x00\xEB\xBE\x94\x00\xEB\xBE"
-     "\x95\x00"
-     "\xEB\xBE\x96\x00\xEB\xBE\x97\x00\xEB\xBE\x98\x00\xEB\xBE\x99\x00\xEB\xBE"
-     "\x9A\x00"
-     "\xEB\xBE\x9B\x00\xEB\xBE\x9C\x00\xEB\xBE\x9D\x00\xEB\xBE\x9E\x00\xEB\xBE"
-     "\x9F\x00"
-     "\xEB\xBE\xA0\x00\xEB\xBE\xA1\x00\xEB\xBE\xA2\x00\xEB\xBE\xA3\x00\xEB\xBE"
-     "\xA4\x00"
-     "\xEB\xBE\xA5\x00\xEB\xBE\xA6\x00\xEB\xBE\xA7\x00\xEB\xBE\xA8\x00\xEB\xBE"
-     "\xA9\x00"
-     "\xEB\xBE\xAA\x00\xEB\xBE\xAB\x00\xEB\xBE\xAC\x00\xEB\xBE\xAD\x00\xEB\xBE"
-     "\xAE\x00"
-     "\xEB\xBE\xAF\x00\xEB\xBE\xB0\x00\xEB\xBE\xB1\x00\xEB\xBE\xB2\x00\xEB\xBE"
-     "\xB3\x00"
-     "\xEB\xBE\xB4\x00\xEB\xBE\xB5\x00\xEB\xBE\xB6\x00\xEB\xBE\xB7\x00\xEB\xBE"
-     "\xB8\x00"
-     "\xEB\xBE\xB9\x00\xEB\xBE\xBA\x00\xEB\xBE\xBB\x00\xEB\xBE\xBC\x00\xEB\xBE"
-     "\xBD\x00"
-     "\xEB\xBE\xBE\x00\xEB\xBE\xBF\x00\xEB\xBF\x80\x00\xEB\xBF\x81\x00\xEB\xBF"
-     "\x82\x00"
-     "\xEB\xBF\x83\x00\xEB\xBF\x84\x00\xEB\xBF\x85\x00\xEB\xBF\x86\x00\xEB\xBF"
-     "\x87\x00"
-     "\xEB\xBF\x88\x00\xEB\xBF\x89\x00\xEB\xBF\x8A\x00\xEB\xBF\x8B\x00\xEB\xBF"
-     "\x8C\x00"
-     "\xEB\xBF\x8D\x00\xEB\xBF\x8E\x00\xEB\xBF\x8F\x00\xEB\xBF\x90\x00\xEB\xBF"
-     "\x91\x00"
-     "\xEB\xBF\x92\x00\xEB\xBF\x93\x00\xEB\xBF\x94\x00\xEB\xBF\x95\x00\xEB\xBF"
-     "\x96\x00"
-     "\xEB\xBF\x97\x00\xEB\xBF\x98\x00\xEB\xBF\x99\x00\xEB\xBF\x9A\x00\xEB\xBF"
-     "\x9B\x00"
-     "\xEB\xBF\x9C\x00\xEB\xBF\x9D\x00\xEB\xBF\x9E\x00\xEB\xBF\x9F\x00\xEB\xBF"
-     "\xA0\x00"
-     "\xEB\xBF\xA1\x00\xEB\xBF\xA2\x00\xEB\xBF\xA3\x00\xEB\xBF\xA4\x00\xEB\xBF"
-     "\xA5\x00"
-     "\xEB\xBF\xA6\x00\xEB\xBF\xA7\x00\xEB\xBF\xA8\x00\xEB\xBF\xA9\x00\xEB\xBF"
-     "\xAA\x00"
-     "\xEB\xBF\xAB\x00\xEB\xBF\xAC\x00\xEB\xBF\xAD\x00\xEB\xBF\xAE\x00\xEB\xBF"
-     "\xAF\x00"
-     "\xEB\xBF\xB0\x00\xEB\xBF\xB1\x00\xEB\xBF\xB2\x00\xEB\xBF\xB3\x00\xEB\xBF"
-     "\xB4\x00"
-     "\xEB\xBF\xB5\x00\xEB\xBF\xB6\x00\xEB\xBF\xB7\x00\xEB\xBF\xB8\x00\xEB\xBF"
-     "\xB9\x00"
-     "\xEB\xBF\xBA\x00\xEB\xBF\xBB\x00\xEB\xBF\xBC\x00\xEB\xBF\xBD\x00\xEB\xBF"
-     "\xBE\x00"
-     "\xEB\xBF\xBF\x00\xEC\x80\x80\x00\xEC\x80\x81\x00\xEC\x80\x82\x00\xEC\x80"
-     "\x83\x00"
-     "\xEC\x80\x84\x00\xEC\x80\x85\x00\xEC\x80\x86\x00\xEC\x80\x87\x00\xEC\x80"
-     "\x88\x00"
-     "\xEC\x80\x89\x00\xEC\x80\x8A\x00\xEC\x80\x8B\x00\xEC\x80\x8C\x00\xEC\x80"
-     "\x8D\x00"
-     "\xEC\x80\x8E\x00\xEC\x80\x8F\x00\xEC\x80\x90\x00\xEC\x80\x91\x00\xEC\x80"
-     "\x92\x00"
-     "\xEC\x80\x93\x00\xEC\x80\x94\x00\xEC\x80\x95\x00\xEC\x80\x96\x00\xEC\x80"
-     "\x97\x00"
-     "\xEC\x80\x98\x00\xEC\x80\x99\x00\xEC\x80\x9A\x00\xEC\x80\x9B\x00\xEC\x80"
-     "\x9C\x00"
-     "\xEC\x80\x9D\x00\xEC\x80\x9E\x00\xEC\x80\x9F\x00\xEC\x80\xA0\x00\xEC\x80"
-     "\xA1\x00"
-     "\xEC\x80\xA2\x00\xEC\x80\xA3\x00\xEC\x80\xA4\x00\xEC\x80\xA5\x00\xEC\x80"
-     "\xA6\x00"
-     "\xEC\x80\xA7\x00\xEC\x80\xA8\x00\xEC\x80\xA9\x00\xEC\x80\xAA\x00\xEC\x80"
-     "\xAB\x00"
-     "\xEC\x80\xAC\x00\xEC\x80\xAD\x00\xEC\x80\xAE\x00\xEC\x80\xAF\x00\xEC\x80"
-     "\xB0\x00"
-     "\xEC\x80\xB1\x00\xEC\x80\xB2\x00\xEC\x80\xB3\x00\xEC\x80\xB4\x00\xEC\x80"
-     "\xB5\x00"
-     "\xEC\x80\xB6\x00\xEC\x80\xB7\x00\xEC\x80\xB8\x00\xEC\x80\xB9\x00\xEC\x80"
-     "\xBA\x00"
-     "\xEC\x80\xBB\x00\xEC\x80\xBC\x00\xEC\x80\xBD\x00\xEC\x80\xBE\x00\xEC\x80"
-     "\xBF\x00"
-     "\xEC\x81\x80\x00\xEC\x81\x81\x00\xEC\x81\x82\x00\xEC\x81\x83\x00\xEC\x81"
-     "\x84\x00"
-     "\xEC\x81\x85\x00\xEC\x81\x86\x00\xEC\x81\x87\x00\xEC\x81\x88\x00\xEC\x81"
-     "\x89\x00"
-     "\xEC\x81\x8A\x00\xEC\x81\x8B\x00\xEC\x81\x8C\x00\xEC\x81\x8D\x00\xEC\x81"
-     "\x8E\x00"
-     "\xEC\x81\x8F\x00\xEC\x81\x90\x00\xEC\x81\x91\x00\xEC\x81\x92\x00\xEC\x81"
-     "\x93\x00"
-     "\xEC\x81\x94\x00\xEC\x81\x95\x00\xEC\x81\x96\x00\xEC\x81\x97\x00\xEC\x81"
-     "\x98\x00"
-     "\xEC\x81\x99\x00\xEC\x81\x9A\x00\xEC\x81\x9B\x00\xEC\x81\x9C\x00\xEC\x81"
-     "\x9D\x00"
-     "\xEC\x81\x9E\x00\xEC\x81\x9F\x00\xEC\x81\xA0\x00\xEC\x81\xA1\x00\xEC\x81"
-     "\xA2\x00"
-     "\xEC\x81\xA3\x00\xEC\x81\xA4\x00\xEC\x81\xA5\x00\xEC\x81\xA6\x00\xEC\x81"
-     "\xA7\x00"
-     "\xEC\x81\xA8\x00\xEC\x81\xA9\x00\xEC\x81\xAA\x00\xEC\x81\xAB\x00\xEC\x81"
-     "\xAC\x00"
-     "\xEC\x81\xAD\x00\xEC\x81\xAE\x00\xEC\x81\xAF\x00\xEC\x81\xB0\x00\xEC\x81"
-     "\xB1\x00"
-     "\xEC\x81\xB2\x00\xEC\x81\xB3\x00\xEC\x81\xB4\x00\xEC\x81\xB5\x00\xEC\x81"
-     "\xB6\x00"
-     "\xEC\x81\xB7\x00\xEC\x81\xB8\x00\xEC\x81\xB9\x00\xEC\x81\xBA\x00\xEC\x81"
-     "\xBB\x00"
-     "\xEC\x81\xBC\x00\xEC\x81\xBD\x00\xEC\x81\xBE\x00\xEC\x81\xBF\x00\xEC\x82"
-     "\x80\x00"
-     "\xEC\x82\x81\x00\xEC\x82\x82\x00\xEC\x82\x83\x00\xEC\x82\x84\x00\xEC\x82"
-     "\x85\x00"
-     "\xEC\x82\x86\x00\xEC\x82\x87\x00\xEC\x82\x88\x00\xEC\x82\x89\x00\xEC\x82"
-     "\x8A\x00"
-     "\xEC\x82\x8B\x00\xEC\x82\x8C\x00\xEC\x82\x8D\x00\xEC\x82\x8E\x00\xEC\x82"
-     "\x8F\x00"
-     "\xEC\x82\x90\x00\xEC\x82\x91\x00\xEC\x82\x92\x00\xEC\x82\x93\x00\xEC\x82"
-     "\x94\x00"
-     "\xEC\x82\x95\x00\xEC\x82\x96\x00\xEC\x82\x97\x00\xEC\x82\x98\x00\xEC\x82"
-     "\x99\x00"
-     "\xEC\x82\x9A\x00\xEC\x82\x9B\x00\xEC\x82\x9C\x00\xEC\x82\x9D\x00\xEC\x82"
-     "\x9E\x00"
-     "\xEC\x82\x9F\x00\xEC\x82\xA0\x00\xEC\x82\xA1\x00\xEC\x82\xA2\x00\xEC\x82"
-     "\xA3\x00"
-     "\xEC\x82\xA4\x00\xEC\x82\xA5\x00\xEC\x82\xA6\x00\xEC\x82\xA7\x00\xEC\x82"
-     "\xA8\x00"
-     "\xEC\x82\xA9\x00\xEC\x82\xAA\x00\xEC\x82\xAB\x00\xEC\x82\xAC\x00\xEC\x82"
-     "\xAD\x00"
-     "\xEC\x82\xAE\x00\xEC\x82\xAF\x00\xEC\x82\xB0\x00\xEC\x82\xB1\x00\xEC\x82"
-     "\xB2\x00"
-     "\xEC\x82\xB3\x00\xEC\x82\xB4\x00\xEC\x82\xB5\x00\xEC\x82\xB6\x00\xEC\x82"
-     "\xB7\x00"
-     "\xEC\x82\xB8\x00\xEC\x82\xB9\x00\xEC\x82\xBA\x00\xEC\x82\xBB\x00\xEC\x82"
-     "\xBC\x00"
-     "\xEC\x82\xBD\x00\xEC\x82\xBE\x00\xEC\x82\xBF\x00\xEC\x83\x80\x00\xEC\x83"
-     "\x81\x00"
-     "\xEC\x83\x82\x00\xEC\x83\x83\x00\xEC\x83\x84\x00\xEC\x83\x85\x00\xEC\x83"
-     "\x86\x00"
-     "\xEC\x83\x87\x00\xEC\x83\x88\x00\xEC\x83\x89\x00\xEC\x83\x8A\x00\xEC\x83"
-     "\x8B\x00"
-     "\xEC\x83\x8C\x00\xEC\x83\x8D\x00\xEC\x83\x8E\x00\xEC\x83\x8F\x00\xEC\x83"
-     "\x90\x00"
-     "\xEC\x83\x91\x00\xEC\x83\x92\x00\xEC\x83\x93\x00\xEC\x83\x94\x00\xEC\x83"
-     "\x95\x00"
-     "\xEC\x83\x96\x00\xEC\x83\x97\x00\xEC\x83\x98\x00\xEC\x83\x99\x00\xEC\x83"
-     "\x9A\x00"
-     "\xEC\x83\x9B\x00\xEC\x83\x9C\x00\xEC\x83\x9D\x00\xEC\x83\x9E\x00\xEC\x83"
-     "\x9F\x00"
-     "\xEC\x83\xA0\x00\xEC\x83\xA1\x00\xEC\x83\xA2\x00\xEC\x83\xA3\x00\xEC\x83"
-     "\xA4\x00"
-     "\xEC\x83\xA5\x00\xEC\x83\xA6\x00\xEC\x83\xA7\x00\xEC\x83\xA8\x00\xEC\x83"
-     "\xA9\x00"
-     "\xEC\x83\xAA\x00\xEC\x83\xAB\x00\xEC\x83\xAC\x00\xEC\x83\xAD\x00\xEC\x83"
-     "\xAE\x00"
-     "\xEC\x83\xAF\x00\xEC\x83\xB0\x00\xEC\x83\xB1\x00\xEC\x83\xB2\x00\xEC\x83"
-     "\xB3\x00"
-     "\xEC\x83\xB4\x00\xEC\x83\xB5\x00\xEC\x83\xB6\x00\xEC\x83\xB7\x00\xEC\x83"
-     "\xB8\x00"
-     "\xEC\x83\xB9\x00\xEC\x83\xBA\x00\xEC\x83\xBB\x00\xEC\x83\xBC\x00\xEC\x83"
-     "\xBD\x00"
-     "\xEC\x83\xBE\x00\xEC\x83\xBF\x00\xEC\x84\x80\x00\xEC\x84\x81\x00\xEC\x84"
-     "\x82\x00"
-     "\xEC\x84\x83\x00\xEC\x84\x84\x00\xEC\x84\x85\x00\xEC\x84\x86\x00\xEC\x84"
-     "\x87\x00"
-     "\xEC\x84\x88\x00\xEC\x84\x89\x00\xEC\x84\x8A\x00\xEC\x84\x8B\x00\xEC\x84"
-     "\x8C\x00"
-     "\xEC\x84\x8D\x00\xEC\x84\x8E\x00\xEC\x84\x8F\x00\xEC\x84\x90\x00\xEC\x84"
-     "\x91\x00"
-     "\xEC\x84\x92\x00\xEC\x84\x93\x00\xEC\x84\x94\x00\xEC\x84\x95\x00\xEC\x84"
-     "\x96\x00"
-     "\xEC\x84\x97\x00\xEC\x84\x98\x00\xEC\x84\x99\x00\xEC\x84\x9A\x00\xEC\x84"
-     "\x9B\x00"
-     "\xEC\x84\x9C\x00\xEC\x84\x9D\x00\xEC\x84\x9E\x00\xEC\x84\x9F\x00\xEC\x84"
-     "\xA0\x00"
-     "\xEC\x84\xA1\x00\xEC\x84\xA2\x00\xEC\x84\xA3\x00\xEC\x84\xA4\x00\xEC\x84"
-     "\xA5\x00"
-     "\xEC\x84\xA6\x00\xEC\x84\xA7\x00\xEC\x84\xA8\x00\xEC\x84\xA9\x00\xEC\x84"
-     "\xAA\x00"
-     "\xEC\x84\xAB\x00\xEC\x84\xAC\x00\xEC\x84\xAD\x00\xEC\x84\xAE\x00\xEC\x84"
-     "\xAF\x00"
-     "\xEC\x84\xB0\x00\xEC\x84\xB1\x00\xEC\x84\xB2\x00\xEC\x84\xB3\x00\xEC\x84"
-     "\xB4\x00"
-     "\xEC\x84\xB5\x00\xEC\x84\xB6\x00\xEC\x84\xB7\x00\xEC\x84\xB8\x00\xEC\x84"
-     "\xB9\x00"
-     "\xEC\x84\xBA\x00\xEC\x84\xBB\x00\xEC\x84\xBC\x00\xEC\x84\xBD\x00\xEC\x84"
-     "\xBE\x00"
-     "\xEC\x84\xBF\x00\xEC\x85\x80\x00\xEC\x85\x81\x00\xEC\x85\x82\x00\xEC\x85"
-     "\x83\x00"
-     "\xEC\x85\x84\x00\xEC\x85\x85\x00\xEC\x85\x86\x00\xEC\x85\x87\x00\xEC\x85"
-     "\x88\x00"
-     "\xEC\x85\x89\x00\xEC\x85\x8A\x00\xEC\x85\x8B\x00\xEC\x85\x8C\x00\xEC\x85"
-     "\x8D\x00"
-     "\xEC\x85\x8E\x00\xEC\x85\x8F\x00\xEC\x85\x90\x00\xEC\x85\x91\x00\xEC\x85"
-     "\x92\x00"
-     "\xEC\x85\x93\x00\xEC\x85\x94\x00\xEC\x85\x95\x00\xEC\x85\x96\x00\xEC\x85"
-     "\x97\x00"
-     "\xEC\x85\x98\x00\xEC\x85\x99\x00\xEC\x85\x9A\x00\xEC\x85\x9B\x00\xEC\x85"
-     "\x9C\x00"
-     "\xEC\x85\x9D\x00\xEC\x85\x9E\x00\xEC\x85\x9F\x00\xEC\x85\xA0\x00\xEC\x85"
-     "\xA1\x00"
-     "\xEC\x85\xA2\x00\xEC\x85\xA3\x00\xEC\x85\xA4\x00\xEC\x85\xA5\x00\xEC\x85"
-     "\xA6\x00"
-     "\xEC\x85\xA7\x00\xEC\x85\xA8\x00\xEC\x85\xA9\x00\xEC\x85\xAA\x00\xEC\x85"
-     "\xAB\x00"
-     "\xEC\x85\xAC\x00\xEC\x85\xAD\x00\xEC\x85\xAE\x00\xEC\x85\xAF\x00\xEC\x85"
-     "\xB0\x00"
-     "\xEC\x85\xB1\x00\xEC\x85\xB2\x00\xEC\x85\xB3\x00\xEC\x85\xB4\x00\xEC\x85"
-     "\xB5\x00"
-     "\xEC\x85\xB6\x00\xEC\x85\xB7\x00\xEC\x85\xB8\x00\xEC\x85\xB9\x00\xEC\x85"
-     "\xBA\x00"
-     "\xEC\x85\xBB\x00\xEC\x85\xBC\x00\xEC\x85\xBD\x00\xEC\x85\xBE\x00\xEC\x85"
-     "\xBF\x00"
-     "\xEC\x86\x80\x00\xEC\x86\x81\x00\xEC\x86\x82\x00\xEC\x86\x83\x00\xEC\x86"
-     "\x84\x00"
-     "\xEC\x86\x85\x00\xEC\x86\x86\x00\xEC\x86\x87\x00\xEC\x86\x88\x00\xEC\x86"
-     "\x89\x00"
-     "\xEC\x86\x8A\x00\xEC\x86\x8B\x00\xEC\x86\x8C\x00\xEC\x86\x8D\x00\xEC\x86"
-     "\x8E\x00"
-     "\xEC\x86\x8F\x00\xEC\x86\x90\x00\xEC\x86\x91\x00\xEC\x86\x92\x00\xEC\x86"
-     "\x93\x00"
-     "\xEC\x86\x94\x00\xEC\x86\x95\x00\xEC\x86\x96\x00\xEC\x86\x97\x00\xEC\x86"
-     "\x98\x00"
-     "\xEC\x86\x99\x00\xEC\x86\x9A\x00\xEC\x86\x9B\x00\xEC\x86\x9C\x00\xEC\x86"
-     "\x9D\x00"
-     "\xEC\x86\x9E\x00\xEC\x86\x9F\x00\xEC\x86\xA0\x00\xEC\x86\xA1\x00\xEC\x86"
-     "\xA2\x00"
-     "\xEC\x86\xA3\x00\xEC\x86\xA4\x00\xEC\x86\xA5\x00\xEC\x86\xA6\x00\xEC\x86"
-     "\xA7\x00"
-     "\xEC\x86\xA8\x00\xEC\x86\xA9\x00\xEC\x86\xAA\x00\xEC\x86\xAB\x00\xEC\x86"
-     "\xAC\x00"
-     "\xEC\x86\xAD\x00\xEC\x86\xAE\x00\xEC\x86\xAF\x00\xEC\x86\xB0\x00\xEC\x86"
-     "\xB1\x00"
-     "\xEC\x86\xB2\x00\xEC\x86\xB3\x00\xEC\x86\xB4\x00\xEC\x86\xB5\x00\xEC\x86"
-     "\xB6\x00"
-     "\xEC\x86\xB7\x00\xEC\x86\xB8\x00\xEC\x86\xB9\x00\xEC\x86\xBA\x00\xEC\x86"
-     "\xBB\x00"
-     "\xEC\x86\xBC\x00\xEC\x86\xBD\x00\xEC\x86\xBE\x00\xEC\x86\xBF\x00\xEC\x87"
-     "\x80\x00"
-     "\xEC\x87\x81\x00\xEC\x87\x82\x00\xEC\x87\x83\x00\xEC\x87\x84\x00\xEC\x87"
-     "\x85\x00"
-     "\xEC\x87\x86\x00\xEC\x87\x87\x00\xEC\x87\x88\x00\xEC\x87\x89\x00\xEC\x87"
-     "\x8A\x00"
-     "\xEC\x87\x8B\x00\xEC\x87\x8C\x00\xEC\x87\x8D\x00\xEC\x87\x8E\x00\xEC\x87"
-     "\x8F\x00"
-     "\xEC\x87\x90\x00\xEC\x87\x91\x00\xEC\x87\x92\x00\xEC\x87\x93\x00\xEC\x87"
-     "\x94\x00"
-     "\xEC\x87\x95\x00\xEC\x87\x96\x00\xEC\x87\x97\x00\xEC\x87\x98\x00\xEC\x87"
-     "\x99\x00"
-     "\xEC\x87\x9A\x00\xEC\x87\x9B\x00\xEC\x87\x9C\x00\xEC\x87\x9D\x00\xEC\x87"
-     "\x9E\x00"
-     "\xEC\x87\x9F\x00\xEC\x87\xA0\x00\xEC\x87\xA1\x00\xEC\x87\xA2\x00\xEC\x87"
-     "\xA3\x00"
-     "\xEC\x87\xA4\x00\xEC\x87\xA5\x00\xEC\x87\xA6\x00\xEC\x87\xA7\x00\xEC\x87"
-     "\xA8\x00"
-     "\xEC\x87\xA9\x00\xEC\x87\xAA\x00\xEC\x87\xAB\x00\xEC\x87\xAC\x00\xEC\x87"
-     "\xAD\x00"
-     "\xEC\x87\xAE\x00\xEC\x87\xAF\x00\xEC\x87\xB0\x00\xEC\x87\xB1\x00\xEC\x87"
-     "\xB2\x00"
-     "\xEC\x87\xB3\x00\xEC\x87\xB4\x00\xEC\x87\xB5\x00\xEC\x87\xB6\x00\xEC\x87"
-     "\xB7\x00"
-     "\xEC\x87\xB8\x00\xEC\x87\xB9\x00\xEC\x87\xBA\x00\xEC\x87\xBB\x00\xEC\x87"
-     "\xBC\x00"
-     "\xEC\x87\xBD\x00\xEC\x87\xBE\x00\xEC\x87\xBF\x00\xEC\x88\x80\x00\xEC\x88"
-     "\x81\x00"
-     "\xEC\x88\x82\x00\xEC\x88\x83\x00\xEC\x88\x84\x00\xEC\x88\x85\x00\xEC\x88"
-     "\x86\x00"
-     "\xEC\x88\x87\x00\xEC\x88\x88\x00\xEC\x88\x89\x00\xEC\x88\x8A\x00\xEC\x88"
-     "\x8B\x00"
-     "\xEC\x88\x8C\x00\xEC\x88\x8D\x00\xEC\x88\x8E\x00\xEC\x88\x8F\x00\xEC\x88"
-     "\x90\x00"
-     "\xEC\x88\x91\x00\xEC\x88\x92\x00\xEC\x88\x93\x00\xEC\x88\x94\x00\xEC\x88"
-     "\x95\x00"
-     "\xEC\x88\x96\x00\xEC\x88\x97\x00\xEC\x88\x98\x00\xEC\x88\x99\x00\xEC\x88"
-     "\x9A\x00"
-     "\xEC\x88\x9B\x00\xEC\x88\x9C\x00\xEC\x88\x9D\x00\xEC\x88\x9E\x00\xEC\x88"
-     "\x9F\x00"
-     "\xEC\x88\xA0\x00\xEC\x88\xA1\x00\xEC\x88\xA2\x00\xEC\x88\xA3\x00\xEC\x88"
-     "\xA4\x00"
-     "\xEC\x88\xA5\x00\xEC\x88\xA6\x00\xEC\x88\xA7\x00\xEC\x88\xA8\x00\xEC\x88"
-     "\xA9\x00"
-     "\xEC\x88\xAA\x00\xEC\x88\xAB\x00\xEC\x88\xAC\x00\xEC\x88\xAD\x00\xEC\x88"
-     "\xAE\x00"
-     "\xEC\x88\xAF\x00\xEC\x88\xB0\x00\xEC\x88\xB1\x00\xEC\x88\xB2\x00\xEC\x88"
-     "\xB3\x00"
-     "\xEC\x88\xB4\x00\xEC\x88\xB5\x00\xEC\x88\xB6\x00\xEC\x88\xB7\x00\xEC\x88"
-     "\xB8\x00"
-     "\xEC\x88\xB9\x00\xEC\x88\xBA\x00\xEC\x88\xBB\x00\xEC\x88\xBC\x00\xEC\x88"
-     "\xBD\x00"
-     "\xEC\x88\xBE\x00\xEC\x88\xBF\x00\xEC\x89\x80\x00\xEC\x89\x81\x00\xEC\x89"
-     "\x82\x00"
-     "\xEC\x89\x83\x00\xEC\x89\x84\x00\xEC\x89\x85\x00\xEC\x89\x86\x00\xEC\x89"
-     "\x87\x00"
-     "\xEC\x89\x88\x00\xEC\x89\x89\x00\xEC\x89\x8A\x00\xEC\x89\x8B\x00\xEC\x89"
-     "\x8C\x00"
-     "\xEC\x89\x8D\x00\xEC\x89\x8E\x00\xEC\x89\x8F\x00\xEC\x89\x90\x00\xEC\x89"
-     "\x91\x00"
-     "\xEC\x89\x92\x00\xEC\x89\x93\x00\xEC\x89\x94\x00\xEC\x89\x95\x00\xEC\x89"
-     "\x96\x00"
-     "\xEC\x89\x97\x00\xEC\x89\x98\x00\xEC\x89\x99\x00\xEC\x89\x9A\x00\xEC\x89"
-     "\x9B\x00"
-     "\xEC\x89\x9C\x00\xEC\x89\x9D\x00\xEC\x89\x9E\x00\xEC\x89\x9F\x00\xEC\x89"
-     "\xA0\x00"
-     "\xEC\x89\xA1\x00\xEC\x89\xA2\x00\xEC\x89\xA3\x00\xEC\x89\xA4\x00\xEC\x89"
-     "\xA5\x00"
-     "\xEC\x89\xA6\x00\xEC\x89\xA7\x00\xEC\x89\xA8\x00\xEC\x89\xA9\x00\xEC\x89"
-     "\xAA\x00"
-     "\xEC\x89\xAB\x00\xEC\x89\xAC\x00\xEC\x89\xAD\x00\xEC\x89\xAE\x00\xEC\x89"
-     "\xAF\x00"
-     "\xEC\x89\xB0\x00\xEC\x89\xB1\x00\xEC\x89\xB2\x00\xEC\x89\xB3\x00\xEC\x89"
-     "\xB4\x00"
-     "\xEC\x89\xB5\x00\xEC\x89\xB6\x00\xEC\x89\xB7\x00\xEC\x89\xB8\x00\xEC\x89"
-     "\xB9\x00"
-     "\xEC\x89\xBA\x00\xEC\x89\xBB\x00\xEC\x89\xBC\x00\xEC\x89\xBD\x00\xEC\x89"
-     "\xBE\x00"
-     "\xEC\x89\xBF\x00\xEC\x8A\x80\x00\xEC\x8A\x81\x00\xEC\x8A\x82\x00\xEC\x8A"
-     "\x83\x00"
-     "\xEC\x8A\x84\x00\xEC\x8A\x85\x00\xEC\x8A\x86\x00\xEC\x8A\x87\x00\xEC\x8A"
-     "\x88\x00"
-     "\xEC\x8A\x89\x00\xEC\x8A\x8A\x00\xEC\x8A\x8B\x00\xEC\x8A\x8C\x00\xEC\x8A"
-     "\x8D\x00"
-     "\xEC\x8A\x8E\x00\xEC\x8A\x8F\x00\xEC\x8A\x90\x00\xEC\x8A\x91\x00\xEC\x8A"
-     "\x92\x00"
-     "\xEC\x8A\x93\x00\xEC\x8A\x94\x00\xEC\x8A\x95\x00\xEC\x8A\x96\x00\xEC\x8A"
-     "\x97\x00"
-     "\xEC\x8A\x98\x00\xEC\x8A\x99\x00\xEC\x8A\x9A\x00\xEC\x8A\x9B\x00\xEC\x8A"
-     "\x9C\x00"
-     "\xEC\x8A\x9D\x00\xEC\x8A\x9E\x00\xEC\x8A\x9F\x00\xEC\x8A\xA0\x00\xEC\x8A"
-     "\xA1\x00"
-     "\xEC\x8A\xA2\x00\xEC\x8A\xA3\x00\xEC\x8A\xA4\x00\xEC\x8A\xA5\x00\xEC\x8A"
-     "\xA6\x00"
-     "\xEC\x8A\xA7\x00\xEC\x8A\xA8\x00\xEC\x8A\xA9\x00\xEC\x8A\xAA\x00\xEC\x8A"
-     "\xAB\x00"
-     "\xEC\x8A\xAC\x00\xEC\x8A\xAD\x00\xEC\x8A\xAE\x00\xEC\x8A\xAF\x00\xEC\x8A"
-     "\xB0\x00"
-     "\xEC\x8A\xB1\x00\xEC\x8A\xB2\x00\xEC\x8A\xB3\x00\xEC\x8A\xB4\x00\xEC\x8A"
-     "\xB5\x00"
-     "\xEC\x8A\xB6\x00\xEC\x8A\xB7\x00\xEC\x8A\xB8\x00\xEC\x8A\xB9\x00\xEC\x8A"
-     "\xBA\x00"
-     "\xEC\x8A\xBB\x00\xEC\x8A\xBC\x00\xEC\x8A\xBD\x00\xEC\x8A\xBE\x00\xEC\x8A"
-     "\xBF\x00"
-     "\xEC\x8B\x80\x00\xEC\x8B\x81\x00\xEC\x8B\x82\x00\xEC\x8B\x83\x00\xEC\x8B"
-     "\x84\x00"
-     "\xEC\x8B\x85\x00\xEC\x8B\x86\x00\xEC\x8B\x87\x00\xEC\x8B\x88\x00\xEC\x8B"
-     "\x89\x00"
-     "\xEC\x8B\x8A\x00\xEC\x8B\x8B\x00\xEC\x8B\x8C\x00\xEC\x8B\x8D\x00\xEC\x8B"
-     "\x8E\x00"
-     "\xEC\x8B\x8F\x00\xEC\x8B\x90\x00\xEC\x8B\x91\x00\xEC\x8B\x92\x00\xEC\x8B"
-     "\x93\x00"
-     "\xEC\x8B\x94\x00\xEC\x8B\x95\x00\xEC\x8B\x96\x00\xEC\x8B\x97\x00\xEC\x8B"
-     "\x98\x00"
-     "\xEC\x8B\x99\x00\xEC\x8B\x9A\x00\xEC\x8B\x9B\x00\xEC\x8B\x9C\x00\xEC\x8B"
-     "\x9D\x00"
-     "\xEC\x8B\x9E\x00\xEC\x8B\x9F\x00\xEC\x8B\xA0\x00\xEC\x8B\xA1\x00\xEC\x8B"
-     "\xA2\x00"
-     "\xEC\x8B\xA3\x00\xEC\x8B\xA4\x00\xEC\x8B\xA5\x00\xEC\x8B\xA6\x00\xEC\x8B"
-     "\xA7\x00"
-     "\xEC\x8B\xA8\x00\xEC\x8B\xA9\x00\xEC\x8B\xAA\x00\xEC\x8B\xAB\x00\xEC\x8B"
-     "\xAC\x00"
-     "\xEC\x8B\xAD\x00\xEC\x8B\xAE\x00\xEC\x8B\xAF\x00\xEC\x8B\xB0\x00\xEC\x8B"
-     "\xB1\x00"
-     "\xEC\x8B\xB2\x00\xEC\x8B\xB3\x00\xEC\x8B\xB4\x00\xEC\x8B\xB5\x00\xEC\x8B"
-     "\xB6\x00"
-     "\xEC\x8B\xB7\x00\xEC\x8B\xB8\x00\xEC\x8B\xB9\x00\xEC\x8B\xBA\x00\xEC\x8B"
-     "\xBB\x00"
-     "\xEC\x8B\xBC\x00\xEC\x8B\xBD\x00\xEC\x8B\xBE\x00\xEC\x8B\xBF\x00\xEC\x8C"
-     "\x80\x00"
-     "\xEC\x8C\x81\x00\xEC\x8C\x82\x00\xEC\x8C\x83\x00\xEC\x8C\x84\x00\xEC\x8C"
-     "\x85\x00"
-     "\xEC\x8C\x86\x00\xEC\x8C\x87\x00\xEC\x8C\x88\x00\xEC\x8C\x89\x00\xEC\x8C"
-     "\x8A\x00"
-     "\xEC\x8C\x8B\x00\xEC\x8C\x8C\x00\xEC\x8C\x8D\x00\xEC\x8C\x8E\x00\xEC\x8C"
-     "\x8F\x00"
-     "\xEC\x8C\x90\x00\xEC\x8C\x91\x00\xEC\x8C\x92\x00\xEC\x8C\x93\x00\xEC\x8C"
-     "\x94\x00"
-     "\xEC\x8C\x95\x00\xEC\x8C\x96\x00\xEC\x8C\x97\x00\xEC\x8C\x98\x00\xEC\x8C"
-     "\x99\x00"
-     "\xEC\x8C\x9A\x00\xEC\x8C\x9B\x00\xEC\x8C\x9C\x00\xEC\x8C\x9D\x00\xEC\x8C"
-     "\x9E\x00"
-     "\xEC\x8C\x9F\x00\xEC\x8C\xA0\x00\xEC\x8C\xA1\x00\xEC\x8C\xA2\x00\xEC\x8C"
-     "\xA3\x00"
-     "\xEC\x8C\xA4\x00\xEC\x8C\xA5\x00\xEC\x8C\xA6\x00\xEC\x8C\xA7\x00\xEC\x8C"
-     "\xA8\x00"
-     "\xEC\x8C\xA9\x00\xEC\x8C\xAA\x00\xEC\x8C\xAB\x00\xEC\x8C\xAC\x00\xEC\x8C"
-     "\xAD\x00"
-     "\xEC\x8C\xAE\x00\xEC\x8C\xAF\x00\xEC\x8C\xB0\x00\xEC\x8C\xB1\x00\xEC\x8C"
-     "\xB2\x00"
-     "\xEC\x8C\xB3\x00\xEC\x8C\xB4\x00\xEC\x8C\xB5\x00\xEC\x8C\xB6\x00\xEC\x8C"
-     "\xB7\x00"
-     "\xEC\x8C\xB8\x00\xEC\x8C\xB9\x00\xEC\x8C\xBA\x00\xEC\x8C\xBB\x00\xEC\x8C"
-     "\xBC\x00"
-     "\xEC\x8C\xBD\x00\xEC\x8C\xBE\x00\xEC\x8C\xBF\x00\xEC\x8D\x80\x00\xEC\x8D"
-     "\x81\x00"
-     "\xEC\x8D\x82\x00\xEC\x8D\x83\x00\xEC\x8D\x84\x00\xEC\x8D\x85\x00\xEC\x8D"
-     "\x86\x00"
-     "\xEC\x8D\x87\x00\xEC\x8D\x88\x00\xEC\x8D\x89\x00\xEC\x8D\x8A\x00\xEC\x8D"
-     "\x8B\x00"
-     "\xEC\x8D\x8C\x00\xEC\x8D\x8D\x00\xEC\x8D\x8E\x00\xEC\x8D\x8F\x00\xEC\x8D"
-     "\x90\x00"
-     "\xEC\x8D\x91\x00\xEC\x8D\x92\x00\xEC\x8D\x93\x00\xEC\x8D\x94\x00\xEC\x8D"
-     "\x95\x00"
-     "\xEC\x8D\x96\x00\xEC\x8D\x97\x00\xEC\x8D\x98\x00\xEC\x8D\x99\x00\xEC\x8D"
-     "\x9A\x00"
-     "\xEC\x8D\x9B\x00\xEC\x8D\x9C\x00\xEC\x8D\x9D\x00\xEC\x8D\x9E\x00\xEC\x8D"
-     "\x9F\x00"
-     "\xEC\x8D\xA0\x00\xEC\x8D\xA1\x00\xEC\x8D\xA2\x00\xEC\x8D\xA3\x00\xEC\x8D"
-     "\xA4\x00"
-     "\xEC\x8D\xA5\x00\xEC\x8D\xA6\x00\xEC\x8D\xA7\x00\xEC\x8D\xA8\x00\xEC\x8D"
-     "\xA9\x00"
-     "\xEC\x8D\xAA\x00\xEC\x8D\xAB\x00\xEC\x8D\xAC\x00\xEC\x8D\xAD\x00\xEC\x8D"
-     "\xAE\x00"
-     "\xEC\x8D\xAF\x00\xEC\x8D\xB0\x00\xEC\x8D\xB1\x00\xEC\x8D\xB2\x00\xEC\x8D"
-     "\xB3\x00"
-     "\xEC\x8D\xB4\x00\xEC\x8D\xB5\x00\xEC\x8D\xB6\x00\xEC\x8D\xB7\x00\xEC\x8D"
-     "\xB8\x00"
-     "\xEC\x8D\xB9\x00\xEC\x8D\xBA\x00\xEC\x8D\xBB\x00\xEC\x8D\xBC\x00\xEC\x8D"
-     "\xBD\x00"
-     "\xEC\x8D\xBE\x00\xEC\x8D\xBF\x00\xEC\x8E\x80\x00\xEC\x8E\x81\x00\xEC\x8E"
-     "\x82\x00"
-     "\xEC\x8E\x83\x00\xEC\x8E\x84\x00\xEC\x8E\x85\x00\xEC\x8E\x86\x00\xEC\x8E"
-     "\x87\x00"
-     "\xEC\x8E\x88\x00\xEC\x8E\x89\x00\xEC\x8E\x8A\x00\xEC\x8E\x8B\x00\xEC\x8E"
-     "\x8C\x00"
-     "\xEC\x8E\x8D\x00\xEC\x8E\x8E\x00\xEC\x8E\x8F\x00\xEC\x8E\x90\x00\xEC\x8E"
-     "\x91\x00"
-     "\xEC\x8E\x92\x00\xEC\x8E\x93\x00\xEC\x8E\x94\x00\xEC\x8E\x95\x00\xEC\x8E"
-     "\x96\x00"
-     "\xEC\x8E\x97\x00\xEC\x8E\x98\x00\xEC\x8E\x99\x00\xEC\x8E\x9A\x00\xEC\x8E"
-     "\x9B\x00"
-     "\xEC\x8E\x9C\x00\xEC\x8E\x9D\x00\xEC\x8E\x9E\x00\xEC\x8E\x9F\x00\xEC\x8E"
-     "\xA0\x00"
-     "\xEC\x8E\xA1\x00\xEC\x8E\xA2\x00\xEC\x8E\xA3\x00\xEC\x8E\xA4\x00\xEC\x8E"
-     "\xA5\x00"
-     "\xEC\x8E\xA6\x00\xEC\x8E\xA7\x00\xEC\x8E\xA8\x00\xEC\x8E\xA9\x00\xEC\x8E"
-     "\xAA\x00"
-     "\xEC\x8E\xAB\x00\xEC\x8E\xAC\x00\xEC\x8E\xAD\x00\xEC\x8E\xAE\x00\xEC\x8E"
-     "\xAF\x00"
-     "\xEC\x8E\xB0\x00\xEC\x8E\xB1\x00\xEC\x8E\xB2\x00\xEC\x8E\xB3\x00\xEC\x8E"
-     "\xB4\x00"
-     "\xEC\x8E\xB5\x00\xEC\x8E\xB6\x00\xEC\x8E\xB7\x00\xEC\x8E\xB8\x00\xEC\x8E"
-     "\xB9\x00"
-     "\xEC\x8E\xBA\x00\xEC\x8E\xBB\x00\xEC\x8E\xBC\x00\xEC\x8E\xBD\x00\xEC\x8E"
-     "\xBE\x00"
-     "\xEC\x8E\xBF\x00\xEC\x8F\x80\x00\xEC\x8F\x81\x00\xEC\x8F\x82\x00\xEC\x8F"
-     "\x83\x00"
-     "\xEC\x8F\x84\x00\xEC\x8F\x85\x00\xEC\x8F\x86\x00\xEC\x8F\x87\x00\xEC\x8F"
-     "\x88\x00"
-     "\xEC\x8F\x89\x00\xEC\x8F\x8A\x00\xEC\x8F\x8B\x00\xEC\x8F\x8C\x00\xEC\x8F"
-     "\x8D\x00"
-     "\xEC\x8F\x8E\x00\xEC\x8F\x8F\x00\xEC\x8F\x90\x00\xEC\x8F\x91\x00\xEC\x8F"
-     "\x92\x00"
-     "\xEC\x8F\x93\x00\xEC\x8F\x94\x00\xEC\x8F\x95\x00\xEC\x8F\x96\x00\xEC\x8F"
-     "\x97\x00"
-     "\xEC\x8F\x98\x00\xEC\x8F\x99\x00\xEC\x8F\x9A\x00\xEC\x8F\x9B\x00\xEC\x8F"
-     "\x9C\x00"
-     "\xEC\x8F\x9D\x00\xEC\x8F\x9E\x00\xEC\x8F\x9F\x00\xEC\x8F\xA0\x00\xEC\x8F"
-     "\xA1\x00"
-     "\xEC\x8F\xA2\x00\xEC\x8F\xA3\x00\xEC\x8F\xA4\x00\xEC\x8F\xA5\x00\xEC\x8F"
-     "\xA6\x00"
-     "\xEC\x8F\xA7\x00\xEC\x8F\xA8\x00\xEC\x8F\xA9\x00\xEC\x8F\xAA\x00\xEC\x8F"
-     "\xAB\x00"
-     "\xEC\x8F\xAC\x00\xEC\x8F\xAD\x00\xEC\x8F\xAE\x00\xEC\x8F\xAF\x00\xEC\x8F"
-     "\xB0\x00"
-     "\xEC\x8F\xB1\x00\xEC\x8F\xB2\x00\xEC\x8F\xB3\x00\xEC\x8F\xB4\x00\xEC\x8F"
-     "\xB5\x00"
-     "\xEC\x8F\xB6\x00\xEC\x8F\xB7\x00\xEC\x8F\xB8\x00\xEC\x8F\xB9\x00\xEC\x8F"
-     "\xBA\x00"
-     "\xEC\x8F\xBB\x00\xEC\x8F\xBC\x00\xEC\x8F\xBD\x00\xEC\x8F\xBE\x00\xEC\x8F"
-     "\xBF\x00"
-     "\xEC\x90\x80\x00\xEC\x90\x81\x00\xEC\x90\x82\x00\xEC\x90\x83\x00\xEC\x90"
-     "\x84\x00"
-     "\xEC\x90\x85\x00\xEC\x90\x86\x00\xEC\x90\x87\x00\xEC\x90\x88\x00\xEC\x90"
-     "\x89\x00"
-     "\xEC\x90\x8A\x00\xEC\x90\x8B\x00\xEC\x90\x8C\x00\xEC\x90\x8D\x00\xEC\x90"
-     "\x8E\x00"
-     "\xEC\x90\x8F\x00\xEC\x90\x90\x00\xEC\x90\x91\x00\xEC\x90\x92\x00\xEC\x90"
-     "\x93\x00"
-     "\xEC\x90\x94\x00\xEC\x90\x95\x00\xEC\x90\x96\x00\xEC\x90\x97\x00\xEC\x90"
-     "\x98\x00"
-     "\xEC\x90\x99\x00\xEC\x90\x9A\x00\xEC\x90\x9B\x00\xEC\x90\x9C\x00\xEC\x90"
-     "\x9D\x00"
-     "\xEC\x90\x9E\x00\xEC\x90\x9F\x00\xEC\x90\xA0\x00\xEC\x90\xA1\x00\xEC\x90"
-     "\xA2\x00"
-     "\xEC\x90\xA3\x00\xEC\x90\xA4\x00\xEC\x90\xA5\x00\xEC\x90\xA6\x00\xEC\x90"
-     "\xA7\x00"
-     "\xEC\x90\xA8\x00\xEC\x90\xA9\x00\xEC\x90\xAA\x00\xEC\x90\xAB\x00\xEC\x90"
-     "\xAC\x00"
-     "\xEC\x90\xAD\x00\xEC\x90\xAE\x00\xEC\x90\xAF\x00\xEC\x90\xB0\x00\xEC\x90"
-     "\xB1\x00"
-     "\xEC\x90\xB2\x00\xEC\x90\xB3\x00\xEC\x90\xB4\x00\xEC\x90\xB5\x00\xEC\x90"
-     "\xB6\x00"
-     "\xEC\x90\xB7\x00\xEC\x90\xB8\x00\xEC\x90\xB9\x00\xEC\x90\xBA\x00\xEC\x90"
-     "\xBB\x00"
-     "\xEC\x90\xBC\x00\xEC\x90\xBD\x00\xEC\x90\xBE\x00\xEC\x90\xBF\x00\xEC\x91"
-     "\x80\x00"
-     "\xEC\x91\x81\x00\xEC\x91\x82\x00\xEC\x91\x83\x00\xEC\x91\x84\x00\xEC\x91"
-     "\x85\x00"
-     "\xEC\x91\x86\x00\xEC\x91\x87\x00\xEC\x91\x88\x00\xEC\x91\x89\x00\xEC\x91"
-     "\x8A\x00"
-     "\xEC\x91\x8B\x00\xEC\x91\x8C\x00\xEC\x91\x8D\x00\xEC\x91\x8E\x00\xEC\x91"
-     "\x8F\x00"
-     "\xEC\x91\x90\x00\xEC\x91\x91\x00\xEC\x91\x92\x00\xEC\x91\x93\x00\xEC\x91"
-     "\x94\x00"
-     "\xEC\x91\x95\x00\xEC\x91\x96\x00\xEC\x91\x97\x00\xEC\x91\x98\x00\xEC\x91"
-     "\x99\x00"
-     "\xEC\x91\x9A\x00\xEC\x91\x9B\x00\xEC\x91\x9C\x00\xEC\x91\x9D\x00\xEC\x91"
-     "\x9E\x00"
-     "\xEC\x91\x9F\x00\xEC\x91\xA0\x00\xEC\x91\xA1\x00\xEC\x91\xA2\x00\xEC\x91"
-     "\xA3\x00"
-     "\xEC\x91\xA4\x00\xEC\x91\xA5\x00\xEC\x91\xA6\x00\xEC\x91\xA7\x00\xEC\x91"
-     "\xA8\x00"
-     "\xEC\x91\xA9\x00\xEC\x91\xAA\x00\xEC\x91\xAB\x00\xEC\x91\xAC\x00\xEC\x91"
-     "\xAD\x00"
-     "\xEC\x91\xAE\x00\xEC\x91\xAF\x00\xEC\x91\xB0\x00\xEC\x91\xB1\x00\xEC\x91"
-     "\xB2\x00"
-     "\xEC\x91\xB3\x00\xEC\x91\xB4\x00\xEC\x91\xB5\x00\xEC\x91\xB6\x00\xEC\x91"
-     "\xB7\x00"
-     "\xEC\x91\xB8\x00\xEC\x91\xB9\x00\xEC\x91\xBA\x00\xEC\x91\xBB\x00\xEC\x91"
-     "\xBC\x00"
-     "\xEC\x91\xBD\x00\xEC\x91\xBE\x00\xEC\x91\xBF\x00\xEC\x92\x80\x00\xEC\x92"
-     "\x81\x00"
-     "\xEC\x92\x82\x00\xEC\x92\x83\x00\xEC\x92\x84\x00\xEC\x92\x85\x00\xEC\x92"
-     "\x86\x00"
-     "\xEC\x92\x87\x00\xEC\x92\x88\x00\xEC\x92\x89\x00\xEC\x92\x8A\x00\xEC\x92"
-     "\x8B\x00"
-     "\xEC\x92\x8C\x00\xEC\x92\x8D\x00\xEC\x92\x8E\x00\xEC\x92\x8F\x00\xEC\x92"
-     "\x90\x00"
-     "\xEC\x92\x91\x00\xEC\x92\x92\x00\xEC\x92\x93\x00\xEC\x92\x94\x00\xEC\x92"
-     "\x95\x00"
-     "\xEC\x92\x96\x00\xEC\x92\x97\x00\xEC\x92\x98\x00\xEC\x92\x99\x00\xEC\x92"
-     "\x9A\x00"
-     "\xEC\x92\x9B\x00\xEC\x92\x9C\x00\xEC\x92\x9D\x00\xEC\x92\x9E\x00\xEC\x92"
-     "\x9F\x00"
-     "\xEC\x92\xA0\x00\xEC\x92\xA1\x00\xEC\x92\xA2\x00\xEC\x92\xA3\x00\xEC\x92"
-     "\xA4\x00"
-     "\xEC\x92\xA5\x00\xEC\x92\xA6\x00\xEC\x92\xA7\x00\xEC\x92\xA8\x00\xEC\x92"
-     "\xA9\x00"
-     "\xEC\x92\xAA\x00\xEC\x92\xAB\x00\xEC\x92\xAC\x00\xEC\x92\xAD\x00\xEC\x92"
-     "\xAE\x00"
-     "\xEC\x92\xAF\x00\xEC\x92\xB0\x00\xEC\x92\xB1\x00\xEC\x92\xB2\x00\xEC\x92"
-     "\xB3\x00"
-     "\xEC\x92\xB4\x00\xEC\x92\xB5\x00\xEC\x92\xB6\x00\xEC\x92\xB7\x00\xEC\x92"
-     "\xB8\x00"
-     "\xEC\x92\xB9\x00\xEC\x92\xBA\x00\xEC\x92\xBB\x00\xEC\x92\xBC\x00\xEC\x92"
-     "\xBD\x00"
-     "\xEC\x92\xBE\x00\xEC\x92\xBF\x00\xEC\x93\x80\x00\xEC\x93\x81\x00\xEC\x93"
-     "\x82\x00"
-     "\xEC\x93\x83\x00\xEC\x93\x84\x00\xEC\x93\x85\x00\xEC\x93\x86\x00\xEC\x93"
-     "\x87\x00"
-     "\xEC\x93\x88\x00\xEC\x93\x89\x00\xEC\x93\x8A\x00\xEC\x93\x8B\x00\xEC\x93"
-     "\x8C\x00"
-     "\xEC\x93\x8D\x00\xEC\x93\x8E\x00\xEC\x93\x8F\x00\xEC\x93\x90\x00\xEC\x93"
-     "\x91\x00"
-     "\xEC\x93\x92\x00\xEC\x93\x93\x00\xEC\x93\x94\x00\xEC\x93\x95\x00\xEC\x93"
-     "\x96\x00"
-     "\xEC\x93\x97\x00\xEC\x93\x98\x00\xEC\x93\x99\x00\xEC\x93\x9A\x00\xEC\x93"
-     "\x9B\x00"
-     "\xEC\x93\x9C\x00\xEC\x93\x9D\x00\xEC\x93\x9E\x00\xEC\x93\x9F\x00\xEC\x93"
-     "\xA0\x00"
-     "\xEC\x93\xA1\x00\xEC\x93\xA2\x00\xEC\x93\xA3\x00\xEC\x93\xA4\x00\xEC\x93"
-     "\xA5\x00"
-     "\xEC\x93\xA6\x00\xEC\x93\xA7\x00\xEC\x93\xA8\x00\xEC\x93\xA9\x00\xEC\x93"
-     "\xAA\x00"
-     "\xEC\x93\xAB\x00\xEC\x93\xAC\x00\xEC\x93\xAD\x00\xEC\x93\xAE\x00\xEC\x93"
-     "\xAF\x00"
-     "\xEC\x93\xB0\x00\xEC\x93\xB1\x00\xEC\x93\xB2\x00\xEC\x93\xB3\x00\xEC\x93"
-     "\xB4\x00"
-     "\xEC\x93\xB5\x00\xEC\x93\xB6\x00\xEC\x93\xB7\x00\xEC\x93\xB8\x00\xEC\x93"
-     "\xB9\x00"
-     "\xEC\x93\xBA\x00\xEC\x93\xBB\x00\xEC\x93\xBC\x00\xEC\x93\xBD\x00\xEC\x93"
-     "\xBE\x00"
-     "\xEC\x93\xBF\x00\xEC\x94\x80\x00\xEC\x94\x81\x00\xEC\x94\x82\x00\xEC\x94"
-     "\x83\x00"
-     "\xEC\x94\x84\x00\xEC\x94\x85\x00\xEC\x94\x86\x00\xEC\x94\x87\x00\xEC\x94"
-     "\x88\x00"
-     "\xEC\x94\x89\x00\xEC\x94\x8A\x00\xEC\x94\x8B\x00\xEC\x94\x8C\x00\xEC\x94"
-     "\x8D\x00"
-     "\xEC\x94\x8E\x00\xEC\x94\x8F\x00\xEC\x94\x90\x00\xEC\x94\x91\x00\xEC\x94"
-     "\x92\x00"
-     "\xEC\x94\x93\x00\xEC\x94\x94\x00\xEC\x94\x95\x00\xEC\x94\x96\x00\xEC\x94"
-     "\x97\x00"
-     "\xEC\x94\x98\x00\xEC\x94\x99\x00\xEC\x94\x9A\x00\xEC\x94\x9B\x00\xEC\x94"
-     "\x9C\x00"
-     "\xEC\x94\x9D\x00\xEC\x94\x9E\x00\xEC\x94\x9F\x00\xEC\x94\xA0\x00\xEC\x94"
-     "\xA1\x00"
-     "\xEC\x94\xA2\x00\xEC\x94\xA3\x00\xEC\x94\xA4\x00\xEC\x94\xA5\x00\xEC\x94"
-     "\xA6\x00"
-     "\xEC\x94\xA7\x00\xEC\x94\xA8\x00\xEC\x94\xA9\x00\xEC\x94\xAA\x00\xEC\x94"
-     "\xAB\x00"
-     "\xEC\x94\xAC\x00\xEC\x94\xAD\x00\xEC\x94\xAE\x00\xEC\x94\xAF\x00\xEC\x94"
-     "\xB0\x00"
-     "\xEC\x94\xB1\x00\xEC\x94\xB2\x00\xEC\x94\xB3\x00\xEC\x94\xB4\x00\xEC\x94"
-     "\xB5\x00"
-     "\xEC\x94\xB6\x00\xEC\x94\xB7\x00\xEC\x94\xB8\x00\xEC\x94\xB9\x00\xEC\x94"
-     "\xBA\x00"
-     "\xEC\x94\xBB\x00\xEC\x94\xBC\x00\xEC\x94\xBD\x00\xEC\x94\xBE\x00\xEC\x94"
-     "\xBF\x00"
-     "\xEC\x95\x80\x00\xEC\x95\x81\x00\xEC\x95\x82\x00\xEC\x95\x83\x00\xEC\x95"
-     "\x84\x00"
-     "\xEC\x95\x85\x00\xEC\x95\x86\x00\xEC\x95\x87\x00\xEC\x95\x88\x00\xEC\x95"
-     "\x89\x00"
-     "\xEC\x95\x8A\x00\xEC\x95\x8B\x00\xEC\x95\x8C\x00\xEC\x95\x8D\x00\xEC\x95"
-     "\x8E\x00"
-     "\xEC\x95\x8F\x00\xEC\x95\x90\x00\xEC\x95\x91\x00\xEC\x95\x92\x00\xEC\x95"
-     "\x93\x00"
-     "\xEC\x95\x94\x00\xEC\x95\x95\x00\xEC\x95\x96\x00\xEC\x95\x97\x00\xEC\x95"
-     "\x98\x00"
-     "\xEC\x95\x99\x00\xEC\x95\x9A\x00\xEC\x95\x9B\x00\xEC\x95\x9C\x00\xEC\x95"
-     "\x9D\x00"
-     "\xEC\x95\x9E\x00\xEC\x95\x9F\x00\xEC\x95\xA0\x00\xEC\x95\xA1\x00\xEC\x95"
-     "\xA2\x00"
-     "\xEC\x95\xA3\x00\xEC\x95\xA4\x00\xEC\x95\xA5\x00\xEC\x95\xA6\x00\xEC\x95"
-     "\xA7\x00"
-     "\xEC\x95\xA8\x00\xEC\x95\xA9\x00\xEC\x95\xAA\x00\xEC\x95\xAB\x00\xEC\x95"
-     "\xAC\x00"
-     "\xEC\x95\xAD\x00\xEC\x95\xAE\x00\xEC\x95\xAF\x00\xEC\x95\xB0\x00\xEC\x95"
-     "\xB1\x00"
-     "\xEC\x95\xB2\x00\xEC\x95\xB3\x00\xEC\x95\xB4\x00\xEC\x95\xB5\x00\xEC\x95"
-     "\xB6\x00"
-     "\xEC\x95\xB7\x00\xEC\x95\xB8\x00\xEC\x95\xB9\x00\xEC\x95\xBA\x00\xEC\x95"
-     "\xBB\x00"
-     "\xEC\x95\xBC\x00\xEC\x95\xBD\x00\xEC\x95\xBE\x00\xEC\x95\xBF\x00\xEC\x96"
-     "\x80\x00"
-     "\xEC\x96\x81\x00\xEC\x96\x82\x00\xEC\x96\x83\x00\xEC\x96\x84\x00\xEC\x96"
-     "\x85\x00"
-     "\xEC\x96\x86\x00\xEC\x96\x87\x00\xEC\x96\x88\x00\xEC\x96\x89\x00\xEC\x96"
-     "\x8A\x00"
-     "\xEC\x96\x8B\x00\xEC\x96\x8C\x00\xEC\x96\x8D\x00\xEC\x96\x8E\x00\xEC\x96"
-     "\x8F\x00"
-     "\xEC\x96\x90\x00\xEC\x96\x91\x00\xEC\x96\x92\x00\xEC\x96\x93\x00\xEC\x96"
-     "\x94\x00"
-     "\xEC\x96\x95\x00\xEC\x96\x96\x00\xEC\x96\x97\x00\xEC\x96\x98\x00\xEC\x96"
-     "\x99\x00"
-     "\xEC\x96\x9A\x00\xEC\x96\x9B\x00\xEC\x96\x9C\x00\xEC\x96\x9D\x00\xEC\x96"
-     "\x9E\x00"
-     "\xEC\x96\x9F\x00\xEC\x96\xA0\x00\xEC\x96\xA1\x00\xEC\x96\xA2\x00\xEC\x96"
-     "\xA3\x00"
-     "\xEC\x96\xA4\x00\xEC\x96\xA5\x00\xEC\x96\xA6\x00\xEC\x96\xA7\x00\xEC\x96"
-     "\xA8\x00"
-     "\xEC\x96\xA9\x00\xEC\x96\xAA\x00\xEC\x96\xAB\x00\xEC\x96\xAC\x00\xEC\x96"
-     "\xAD\x00"
-     "\xEC\x96\xAE\x00\xEC\x96\xAF\x00\xEC\x96\xB0\x00\xEC\x96\xB1\x00\xEC\x96"
-     "\xB2\x00"
-     "\xEC\x96\xB3\x00\xEC\x96\xB4\x00\xEC\x96\xB5\x00\xEC\x96\xB6\x00\xEC\x96"
-     "\xB7\x00"
-     "\xEC\x96\xB8\x00\xEC\x96\xB9\x00\xEC\x96\xBA\x00\xEC\x96\xBB\x00\xEC\x96"
-     "\xBC\x00"
-     "\xEC\x96\xBD\x00\xEC\x96\xBE\x00\xEC\x96\xBF\x00\xEC\x97\x80\x00\xEC\x97"
-     "\x81\x00"
-     "\xEC\x97\x82\x00\xEC\x97\x83\x00\xEC\x97\x84\x00\xEC\x97\x85\x00\xEC\x97"
-     "\x86\x00"
-     "\xEC\x97\x87\x00\xEC\x97\x88\x00\xEC\x97\x89\x00\xEC\x97\x8A\x00\xEC\x97"
-     "\x8B\x00"
-     "\xEC\x97\x8C\x00\xEC\x97\x8D\x00\xEC\x97\x8E\x00\xEC\x97\x8F\x00\xEC\x97"
-     "\x90\x00"
-     "\xEC\x97\x91\x00\xEC\x97\x92\x00\xEC\x97\x93\x00\xEC\x97\x94\x00\xEC\x97"
-     "\x95\x00"
-     "\xEC\x97\x96\x00\xEC\x97\x97\x00\xEC\x97\x98\x00\xEC\x97\x99\x00\xEC\x97"
-     "\x9A\x00"
-     "\xEC\x97\x9B\x00\xEC\x97\x9C\x00\xEC\x97\x9D\x00\xEC\x97\x9E\x00\xEC\x97"
-     "\x9F\x00"
-     "\xEC\x97\xA0\x00\xEC\x97\xA1\x00\xEC\x97\xA2\x00\xEC\x97\xA3\x00\xEC\x97"
-     "\xA4\x00"
-     "\xEC\x97\xA5\x00\xEC\x97\xA6\x00\xEC\x97\xA7\x00\xEC\x97\xA8\x00\xEC\x97"
-     "\xA9\x00"
-     "\xEC\x97\xAA\x00\xEC\x97\xAB\x00\xEC\x97\xAC\x00\xEC\x97\xAD\x00\xEC\x97"
-     "\xAE\x00"
-     "\xEC\x97\xAF\x00\xEC\x97\xB0\x00\xEC\x97\xB1\x00\xEC\x97\xB2\x00\xEC\x97"
-     "\xB3\x00"
-     "\xEC\x97\xB4\x00\xEC\x97\xB5\x00\xEC\x97\xB6\x00\xEC\x97\xB7\x00\xEC\x97"
-     "\xB8\x00"
-     "\xEC\x97\xB9\x00\xEC\x97\xBA\x00\xEC\x97\xBB\x00\xEC\x97\xBC\x00\xEC\x97"
-     "\xBD\x00"
-     "\xEC\x97\xBE\x00\xEC\x97\xBF\x00\xEC\x98\x80\x00\xEC\x98\x81\x00\xEC\x98"
-     "\x82\x00"
-     "\xEC\x98\x83\x00\xEC\x98\x84\x00\xEC\x98\x85\x00\xEC\x98\x86\x00\xEC\x98"
-     "\x87\x00"
-     "\xEC\x98\x88\x00\xEC\x98\x89\x00\xEC\x98\x8A\x00\xEC\x98\x8B\x00\xEC\x98"
-     "\x8C\x00"
-     "\xEC\x98\x8D\x00\xEC\x98\x8E\x00\xEC\x98\x8F\x00\xEC\x98\x90\x00\xEC\x98"
-     "\x91\x00"
-     "\xEC\x98\x92\x00\xEC\x98\x93\x00\xEC\x98\x94\x00\xEC\x98\x95\x00\xEC\x98"
-     "\x96\x00"
-     "\xEC\x98\x97\x00\xEC\x98\x98\x00\xEC\x98\x99\x00\xEC\x98\x9A\x00\xEC\x98"
-     "\x9B\x00"
-     "\xEC\x98\x9C\x00\xEC\x98\x9D\x00\xEC\x98\x9E\x00\xEC\x98\x9F\x00\xEC\x98"
-     "\xA0\x00"
-     "\xEC\x98\xA1\x00\xEC\x98\xA2\x00\xEC\x98\xA3\x00\xEC\x98\xA4\x00\xEC\x98"
-     "\xA5\x00"
-     "\xEC\x98\xA6\x00\xEC\x98\xA7\x00\xEC\x98\xA8\x00\xEC\x98\xA9\x00\xEC\x98"
-     "\xAA\x00"
-     "\xEC\x98\xAB\x00\xEC\x98\xAC\x00\xEC\x98\xAD\x00\xEC\x98\xAE\x00\xEC\x98"
-     "\xAF\x00"
-     "\xEC\x98\xB0\x00\xEC\x98\xB1\x00\xEC\x98\xB2\x00\xEC\x98\xB3\x00\xEC\x98"
-     "\xB4\x00"
-     "\xEC\x98\xB5\x00\xEC\x98\xB6\x00\xEC\x98\xB7\x00\xEC\x98\xB8\x00\xEC\x98"
-     "\xB9\x00"
-     "\xEC\x98\xBA\x00\xEC\x98\xBB\x00\xEC\x98\xBC\x00\xEC\x98\xBD\x00\xEC\x98"
-     "\xBE\x00"
-     "\xEC\x98\xBF\x00\xEC\x99\x80\x00\xEC\x99\x81\x00\xEC\x99\x82\x00\xEC\x99"
-     "\x83\x00"
-     "\xEC\x99\x84\x00\xEC\x99\x85\x00\xEC\x99\x86\x00\xEC\x99\x87\x00\xEC\x99"
-     "\x88\x00"
-     "\xEC\x99\x89\x00\xEC\x99\x8A\x00\xEC\x99\x8B\x00\xEC\x99\x8C\x00\xEC\x99"
-     "\x8D\x00"
-     "\xEC\x99\x8E\x00\xEC\x99\x8F\x00\xEC\x99\x90\x00\xEC\x99\x91\x00\xEC\x99"
-     "\x92\x00"
-     "\xEC\x99\x93\x00\xEC\x99\x94\x00\xEC\x99\x95\x00\xEC\x99\x96\x00\xEC\x99"
-     "\x97\x00"
-     "\xEC\x99\x98\x00\xEC\x99\x99\x00\xEC\x99\x9A\x00\xEC\x99\x9B\x00\xEC\x99"
-     "\x9C\x00"
-     "\xEC\x99\x9D\x00\xEC\x99\x9E\x00\xEC\x99\x9F\x00\xEC\x99\xA0\x00\xEC\x99"
-     "\xA1\x00"
-     "\xEC\x99\xA2\x00\xEC\x99\xA3\x00\xEC\x99\xA4\x00\xEC\x99\xA5\x00\xEC\x99"
-     "\xA6\x00"
-     "\xEC\x99\xA7\x00\xEC\x99\xA8\x00\xEC\x99\xA9\x00\xEC\x99\xAA\x00\xEC\x99"
-     "\xAB\x00"
-     "\xEC\x99\xAC\x00\xEC\x99\xAD\x00\xEC\x99\xAE\x00\xEC\x99\xAF\x00\xEC\x99"
-     "\xB0\x00"
-     "\xEC\x99\xB1\x00\xEC\x99\xB2\x00\xEC\x99\xB3\x00\xEC\x99\xB4\x00\xEC\x99"
-     "\xB5\x00"
-     "\xEC\x99\xB6\x00\xEC\x99\xB7\x00\xEC\x99\xB8\x00\xEC\x99\xB9\x00\xEC\x99"
-     "\xBA\x00"
-     "\xEC\x99\xBB\x00\xEC\x99\xBC\x00\xEC\x99\xBD\x00\xEC\x99\xBE\x00\xEC\x99"
-     "\xBF\x00"
-     "\xEC\x9A\x80\x00\xEC\x9A\x81\x00\xEC\x9A\x82\x00\xEC\x9A\x83\x00\xEC\x9A"
-     "\x84\x00"
-     "\xEC\x9A\x85\x00\xEC\x9A\x86\x00\xEC\x9A\x87\x00\xEC\x9A\x88\x00\xEC\x9A"
-     "\x89\x00"
-     "\xEC\x9A\x8A\x00\xEC\x9A\x8B\x00\xEC\x9A\x8C\x00\xEC\x9A\x8D\x00\xEC\x9A"
-     "\x8E\x00"
-     "\xEC\x9A\x8F\x00\xEC\x9A\x90\x00\xEC\x9A\x91\x00\xEC\x9A\x92\x00\xEC\x9A"
-     "\x93\x00"
-     "\xEC\x9A\x94\x00\xEC\x9A\x95\x00\xEC\x9A\x96\x00\xEC\x9A\x97\x00\xEC\x9A"
-     "\x98\x00"
-     "\xEC\x9A\x99\x00\xEC\x9A\x9A\x00\xEC\x9A\x9B\x00\xEC\x9A\x9C\x00\xEC\x9A"
-     "\x9D\x00"
-     "\xEC\x9A\x9E\x00\xEC\x9A\x9F\x00\xEC\x9A\xA0\x00\xEC\x9A\xA1\x00\xEC\x9A"
-     "\xA2\x00"
-     "\xEC\x9A\xA3\x00\xEC\x9A\xA4\x00\xEC\x9A\xA5\x00\xEC\x9A\xA6\x00\xEC\x9A"
-     "\xA7\x00"
-     "\xEC\x9A\xA8\x00\xEC\x9A\xA9\x00\xEC\x9A\xAA\x00\xEC\x9A\xAB\x00\xEC\x9A"
-     "\xAC\x00"
-     "\xEC\x9A\xAD\x00\xEC\x9A\xAE\x00\xEC\x9A\xAF\x00\xEC\x9A\xB0\x00\xEC\x9A"
-     "\xB1\x00"
-     "\xEC\x9A\xB2\x00\xEC\x9A\xB3\x00\xEC\x9A\xB4\x00\xEC\x9A\xB5\x00\xEC\x9A"
-     "\xB6\x00"
-     "\xEC\x9A\xB7\x00\xEC\x9A\xB8\x00\xEC\x9A\xB9\x00\xEC\x9A\xBA\x00\xEC\x9A"
-     "\xBB\x00"
-     "\xEC\x9A\xBC\x00\xEC\x9A\xBD\x00\xEC\x9A\xBE\x00\xEC\x9A\xBF\x00\xEC\x9B"
-     "\x80\x00"
-     "\xEC\x9B\x81\x00\xEC\x9B\x82\x00\xEC\x9B\x83\x00\xEC\x9B\x84\x00\xEC\x9B"
-     "\x85\x00"
-     "\xEC\x9B\x86\x00\xEC\x9B\x87\x00\xEC\x9B\x88\x00\xEC\x9B\x89\x00\xEC\x9B"
-     "\x8A\x00"
-     "\xEC\x9B\x8B\x00\xEC\x9B\x8C\x00\xEC\x9B\x8D\x00\xEC\x9B\x8E\x00\xEC\x9B"
-     "\x8F\x00"
-     "\xEC\x9B\x90\x00\xEC\x9B\x91\x00\xEC\x9B\x92\x00\xEC\x9B\x93\x00\xEC\x9B"
-     "\x94\x00"
-     "\xEC\x9B\x95\x00\xEC\x9B\x96\x00\xEC\x9B\x97\x00\xEC\x9B\x98\x00\xEC\x9B"
-     "\x99\x00"
-     "\xEC\x9B\x9A\x00\xEC\x9B\x9B\x00\xEC\x9B\x9C\x00\xEC\x9B\x9D\x00\xEC\x9B"
-     "\x9E\x00"
-     "\xEC\x9B\x9F\x00\xEC\x9B\xA0\x00\xEC\x9B\xA1\x00\xEC\x9B\xA2\x00\xEC\x9B"
-     "\xA3\x00"
-     "\xEC\x9B\xA4\x00\xEC\x9B\xA5\x00\xEC\x9B\xA6\x00\xEC\x9B\xA7\x00\xEC\x9B"
-     "\xA8\x00"
-     "\xEC\x9B\xA9\x00\xEC\x9B\xAA\x00\xEC\x9B\xAB\x00\xEC\x9B\xAC\x00\xEC\x9B"
-     "\xAD\x00"
-     "\xEC\x9B\xAE\x00\xEC\x9B\xAF\x00\xEC\x9B\xB0\x00\xEC\x9B\xB1\x00\xEC\x9B"
-     "\xB2\x00"
-     "\xEC\x9B\xB3\x00\xEC\x9B\xB4\x00\xEC\x9B\xB5\x00\xEC\x9B\xB6\x00\xEC\x9B"
-     "\xB7\x00"
-     "\xEC\x9B\xB8\x00\xEC\x9B\xB9\x00\xEC\x9B\xBA\x00\xEC\x9B\xBB\x00\xEC\x9B"
-     "\xBC\x00"
-     "\xEC\x9B\xBD\x00\xEC\x9B\xBE\x00\xEC\x9B\xBF\x00\xEC\x9C\x80\x00\xEC\x9C"
-     "\x81\x00"
-     "\xEC\x9C\x82\x00\xEC\x9C\x83\x00\xEC\x9C\x84\x00\xEC\x9C\x85\x00\xEC\x9C"
-     "\x86\x00"
-     "\xEC\x9C\x87\x00\xEC\x9C\x88\x00\xEC\x9C\x89\x00\xEC\x9C\x8A\x00\xEC\x9C"
-     "\x8B\x00"
-     "\xEC\x9C\x8C\x00\xEC\x9C\x8D\x00\xEC\x9C\x8E\x00\xEC\x9C\x8F\x00\xEC\x9C"
-     "\x90\x00"
-     "\xEC\x9C\x91\x00\xEC\x9C\x92\x00\xEC\x9C\x93\x00\xEC\x9C\x94\x00\xEC\x9C"
-     "\x95\x00"
-     "\xEC\x9C\x96\x00\xEC\x9C\x97\x00\xEC\x9C\x98\x00\xEC\x9C\x99\x00\xEC\x9C"
-     "\x9A\x00"
-     "\xEC\x9C\x9B\x00\xEC\x9C\x9C\x00\xEC\x9C\x9D\x00\xEC\x9C\x9E\x00\xEC\x9C"
-     "\x9F\x00"
-     "\xEC\x9C\xA0\x00\xEC\x9C\xA1\x00\xEC\x9C\xA2\x00\xEC\x9C\xA3\x00\xEC\x9C"
-     "\xA4\x00"
-     "\xEC\x9C\xA5\x00\xEC\x9C\xA6\x00\xEC\x9C\xA7\x00\xEC\x9C\xA8\x00\xEC\x9C"
-     "\xA9\x00"
-     "\xEC\x9C\xAA\x00\xEC\x9C\xAB\x00\xEC\x9C\xAC\x00\xEC\x9C\xAD\x00\xEC\x9C"
-     "\xAE\x00"
-     "\xEC\x9C\xAF\x00\xEC\x9C\xB0\x00\xEC\x9C\xB1\x00\xEC\x9C\xB2\x00\xEC\x9C"
-     "\xB3\x00"
-     "\xEC\x9C\xB4\x00\xEC\x9C\xB5\x00\xEC\x9C\xB6\x00\xEC\x9C\xB7\x00\xEC\x9C"
-     "\xB8\x00"
-     "\xEC\x9C\xB9\x00\xEC\x9C\xBA\x00\xEC\x9C\xBB\x00\xEC\x9C\xBC\x00\xEC\x9C"
-     "\xBD\x00"
-     "\xEC\x9C\xBE\x00\xEC\x9C\xBF\x00\xEC\x9D\x80\x00\xEC\x9D\x81\x00\xEC\x9D"
-     "\x82\x00"
-     "\xEC\x9D\x83\x00\xEC\x9D\x84\x00\xEC\x9D\x85\x00\xEC\x9D\x86\x00\xEC\x9D"
-     "\x87\x00"
-     "\xEC\x9D\x88\x00\xEC\x9D\x89\x00\xEC\x9D\x8A\x00\xEC\x9D\x8B\x00\xEC\x9D"
-     "\x8C\x00"
-     "\xEC\x9D\x8D\x00\xEC\x9D\x8E\x00\xEC\x9D\x8F\x00\xEC\x9D\x90\x00\xEC\x9D"
-     "\x91\x00"
-     "\xEC\x9D\x92\x00\xEC\x9D\x93\x00\xEC\x9D\x94\x00\xEC\x9D\x95\x00\xEC\x9D"
-     "\x96\x00"
-     "\xEC\x9D\x97\x00\xEC\x9D\x98\x00\xEC\x9D\x99\x00\xEC\x9D\x9A\x00\xEC\x9D"
-     "\x9B\x00"
-     "\xEC\x9D\x9C\x00\xEC\x9D\x9D\x00\xEC\x9D\x9E\x00\xEC\x9D\x9F\x00\xEC\x9D"
-     "\xA0\x00"
-     "\xEC\x9D\xA1\x00\xEC\x9D\xA2\x00\xEC\x9D\xA3\x00\xEC\x9D\xA4\x00\xEC\x9D"
-     "\xA5\x00"
-     "\xEC\x9D\xA6\x00\xEC\x9D\xA7\x00\xEC\x9D\xA8\x00\xEC\x9D\xA9\x00\xEC\x9D"
-     "\xAA\x00"
-     "\xEC\x9D\xAB\x00\xEC\x9D\xAC\x00\xEC\x9D\xAD\x00\xEC\x9D\xAE\x00\xEC\x9D"
-     "\xAF\x00"
-     "\xEC\x9D\xB0\x00\xEC\x9D\xB1\x00\xEC\x9D\xB2\x00\xEC\x9D\xB3\x00\xEC\x9D"
-     "\xB4\x00"
-     "\xEC\x9D\xB5\x00\xEC\x9D\xB6\x00\xEC\x9D\xB7\x00\xEC\x9D\xB8\x00\xEC\x9D"
-     "\xB9\x00"
-     "\xEC\x9D\xBA\x00\xEC\x9D\xBB\x00\xEC\x9D\xBC\x00\xEC\x9D\xBD\x00\xEC\x9D"
-     "\xBE\x00"
-     "\xEC\x9D\xBF\x00\xEC\x9E\x80\x00\xEC\x9E\x81\x00\xEC\x9E\x82\x00\xEC\x9E"
-     "\x83\x00"
-     "\xEC\x9E\x84\x00\xEC\x9E\x85\x00\xEC\x9E\x86\x00\xEC\x9E\x87\x00\xEC\x9E"
-     "\x88\x00"
-     "\xEC\x9E\x89\x00\xEC\x9E\x8A\x00\xEC\x9E\x8B\x00\xEC\x9E\x8C\x00\xEC\x9E"
-     "\x8D\x00"
-     "\xEC\x9E\x8E\x00\xEC\x9E\x8F\x00\xEC\x9E\x90\x00\xEC\x9E\x91\x00\xEC\x9E"
-     "\x92\x00"
-     "\xEC\x9E\x93\x00\xEC\x9E\x94\x00\xEC\x9E\x95\x00\xEC\x9E\x96\x00\xEC\x9E"
-     "\x97\x00"
-     "\xEC\x9E\x98\x00\xEC\x9E\x99\x00\xEC\x9E\x9A\x00\xEC\x9E\x9B\x00\xEC\x9E"
-     "\x9C\x00"
-     "\xEC\x9E\x9D\x00\xEC\x9E\x9E\x00\xEC\x9E\x9F\x00\xEC\x9E\xA0\x00\xEC\x9E"
-     "\xA1\x00"
-     "\xEC\x9E\xA2\x00\xEC\x9E\xA3\x00\xEC\x9E\xA4\x00\xEC\x9E\xA5\x00\xEC\x9E"
-     "\xA6\x00"
-     "\xEC\x9E\xA7\x00\xEC\x9E\xA8\x00\xEC\x9E\xA9\x00\xEC\x9E\xAA\x00\xEC\x9E"
-     "\xAB\x00"
-     "\xEC\x9E\xAC\x00\xEC\x9E\xAD\x00\xEC\x9E\xAE\x00\xEC\x9E\xAF\x00\xEC\x9E"
-     "\xB0\x00"
-     "\xEC\x9E\xB1\x00\xEC\x9E\xB2\x00\xEC\x9E\xB3\x00\xEC\x9E\xB4\x00\xEC\x9E"
-     "\xB5\x00"
-     "\xEC\x9E\xB6\x00\xEC\x9E\xB7\x00\xEC\x9E\xB8\x00\xEC\x9E\xB9\x00\xEC\x9E"
-     "\xBA\x00"
-     "\xEC\x9E\xBB\x00\xEC\x9E\xBC\x00\xEC\x9E\xBD\x00\xEC\x9E\xBE\x00\xEC\x9E"
-     "\xBF\x00"
-     "\xEC\x9F\x80\x00\xEC\x9F\x81\x00\xEC\x9F\x82\x00\xEC\x9F\x83\x00\xEC\x9F"
-     "\x84\x00"
-     "\xEC\x9F\x85\x00\xEC\x9F\x86\x00\xEC\x9F\x87\x00\xEC\x9F\x88\x00\xEC\x9F"
-     "\x89\x00"
-     "\xEC\x9F\x8A\x00\xEC\x9F\x8B\x00\xEC\x9F\x8C\x00\xEC\x9F\x8D\x00\xEC\x9F"
-     "\x8E\x00"
-     "\xEC\x9F\x8F\x00\xEC\x9F\x90\x00\xEC\x9F\x91\x00\xEC\x9F\x92\x00\xEC\x9F"
-     "\x93\x00"
-     "\xEC\x9F\x94\x00\xEC\x9F\x95\x00\xEC\x9F\x96\x00\xEC\x9F\x97\x00\xEC\x9F"
-     "\x98\x00"
-     "\xEC\x9F\x99\x00\xEC\x9F\x9A\x00\xEC\x9F\x9B\x00\xEC\x9F\x9C\x00\xEC\x9F"
-     "\x9D\x00"
-     "\xEC\x9F\x9E\x00\xEC\x9F\x9F\x00\xEC\x9F\xA0\x00\xEC\x9F\xA1\x00\xEC\x9F"
-     "\xA2\x00"
-     "\xEC\x9F\xA3\x00\xEC\x9F\xA4\x00\xEC\x9F\xA5\x00\xEC\x9F\xA6\x00\xEC\x9F"
-     "\xA7\x00"
-     "\xEC\x9F\xA8\x00\xEC\x9F\xA9\x00\xEC\x9F\xAA\x00\xEC\x9F\xAB\x00\xEC\x9F"
-     "\xAC\x00"
-     "\xEC\x9F\xAD\x00\xEC\x9F\xAE\x00\xEC\x9F\xAF\x00\xEC\x9F\xB0\x00\xEC\x9F"
-     "\xB1\x00"
-     "\xEC\x9F\xB2\x00\xEC\x9F\xB3\x00\xEC\x9F\xB4\x00\xEC\x9F\xB5\x00\xEC\x9F"
-     "\xB6\x00"
-     "\xEC\x9F\xB7\x00\xEC\x9F\xB8\x00\xEC\x9F\xB9\x00\xEC\x9F\xBA\x00\xEC\x9F"
-     "\xBB\x00"
-     "\xEC\x9F\xBC\x00\xEC\x9F\xBD\x00\xEC\x9F\xBE\x00\xEC\x9F\xBF\x00\xEC\xA0"
-     "\x80\x00"
-     "\xEC\xA0\x81\x00\xEC\xA0\x82\x00\xEC\xA0\x83\x00\xEC\xA0\x84\x00\xEC\xA0"
-     "\x85\x00"
-     "\xEC\xA0\x86\x00\xEC\xA0\x87\x00\xEC\xA0\x88\x00\xEC\xA0\x89\x00\xEC\xA0"
-     "\x8A\x00"
-     "\xEC\xA0\x8B\x00\xEC\xA0\x8C\x00\xEC\xA0\x8D\x00\xEC\xA0\x8E\x00\xEC\xA0"
-     "\x8F\x00"
-     "\xEC\xA0\x90\x00\xEC\xA0\x91\x00\xEC\xA0\x92\x00\xEC\xA0\x93\x00\xEC\xA0"
-     "\x94\x00"
-     "\xEC\xA0\x95\x00\xEC\xA0\x96\x00\xEC\xA0\x97\x00\xEC\xA0\x98\x00\xEC\xA0"
-     "\x99\x00"
-     "\xEC\xA0\x9A\x00\xEC\xA0\x9B\x00\xEC\xA0\x9C\x00\xEC\xA0\x9D\x00\xEC\xA0"
-     "\x9E\x00"
-     "\xEC\xA0\x9F\x00\xEC\xA0\xA0\x00\xEC\xA0\xA1\x00\xEC\xA0\xA2\x00\xEC\xA0"
-     "\xA3\x00"
-     "\xEC\xA0\xA4\x00\xEC\xA0\xA5\x00\xEC\xA0\xA6\x00\xEC\xA0\xA7\x00\xEC\xA0"
-     "\xA8\x00"
-     "\xEC\xA0\xA9\x00\xEC\xA0\xAA\x00\xEC\xA0\xAB\x00\xEC\xA0\xAC\x00\xEC\xA0"
-     "\xAD\x00"
-     "\xEC\xA0\xAE\x00\xEC\xA0\xAF\x00\xEC\xA0\xB0\x00\xEC\xA0\xB1\x00\xEC\xA0"
-     "\xB2\x00"
-     "\xEC\xA0\xB3\x00\xEC\xA0\xB4\x00\xEC\xA0\xB5\x00\xEC\xA0\xB6\x00\xEC\xA0"
-     "\xB7\x00"
-     "\xEC\xA0\xB8\x00\xEC\xA0\xB9\x00\xEC\xA0\xBA\x00\xEC\xA0\xBB\x00\xEC\xA0"
-     "\xBC\x00"
-     "\xEC\xA0\xBD\x00\xEC\xA0\xBE\x00\xEC\xA0\xBF\x00\xEC\xA1\x80\x00\xEC\xA1"
-     "\x81\x00"
-     "\xEC\xA1\x82\x00\xEC\xA1\x83\x00\xEC\xA1\x84\x00\xEC\xA1\x85\x00\xEC\xA1"
-     "\x86\x00"
-     "\xEC\xA1\x87\x00\xEC\xA1\x88\x00\xEC\xA1\x89\x00\xEC\xA1\x8A\x00\xEC\xA1"
-     "\x8B\x00"
-     "\xEC\xA1\x8C\x00\xEC\xA1\x8D\x00\xEC\xA1\x8E\x00\xEC\xA1\x8F\x00\xEC\xA1"
-     "\x90\x00"
-     "\xEC\xA1\x91\x00\xEC\xA1\x92\x00\xEC\xA1\x93\x00\xEC\xA1\x94\x00\xEC\xA1"
-     "\x95\x00"
-     "\xEC\xA1\x96\x00\xEC\xA1\x97\x00\xEC\xA1\x98\x00\xEC\xA1\x99\x00\xEC\xA1"
-     "\x9A\x00"
-     "\xEC\xA1\x9B\x00\xEC\xA1\x9C\x00\xEC\xA1\x9D\x00\xEC\xA1\x9E\x00\xEC\xA1"
-     "\x9F\x00"
-     "\xEC\xA1\xA0\x00\xEC\xA1\xA1\x00\xEC\xA1\xA2\x00\xEC\xA1\xA3\x00\xEC\xA1"
-     "\xA4\x00"
-     "\xEC\xA1\xA5\x00\xEC\xA1\xA6\x00\xEC\xA1\xA7\x00\xEC\xA1\xA8\x00\xEC\xA1"
-     "\xA9\x00"
-     "\xEC\xA1\xAA\x00\xEC\xA1\xAB\x00\xEC\xA1\xAC\x00\xEC\xA1\xAD\x00\xEC\xA1"
-     "\xAE\x00"
-     "\xEC\xA1\xAF\x00\xEC\xA1\xB0\x00\xEC\xA1\xB1\x00\xEC\xA1\xB2\x00\xEC\xA1"
-     "\xB3\x00"
-     "\xEC\xA1\xB4\x00\xEC\xA1\xB5\x00\xEC\xA1\xB6\x00\xEC\xA1\xB7\x00\xEC\xA1"
-     "\xB8\x00"
-     "\xEC\xA1\xB9\x00\xEC\xA1\xBA\x00\xEC\xA1\xBB\x00\xEC\xA1\xBC\x00\xEC\xA1"
-     "\xBD\x00"
-     "\xEC\xA1\xBE\x00\xEC\xA1\xBF\x00\xEC\xA2\x80\x00\xEC\xA2\x81\x00\xEC\xA2"
-     "\x82\x00"
-     "\xEC\xA2\x83\x00\xEC\xA2\x84\x00\xEC\xA2\x85\x00\xEC\xA2\x86\x00\xEC\xA2"
-     "\x87\x00"
-     "\xEC\xA2\x88\x00\xEC\xA2\x89\x00\xEC\xA2\x8A\x00\xEC\xA2\x8B\x00\xEC\xA2"
-     "\x8C\x00"
-     "\xEC\xA2\x8D\x00\xEC\xA2\x8E\x00\xEC\xA2\x8F\x00\xEC\xA2\x90\x00\xEC\xA2"
-     "\x91\x00"
-     "\xEC\xA2\x92\x00\xEC\xA2\x93\x00\xEC\xA2\x94\x00\xEC\xA2\x95\x00\xEC\xA2"
-     "\x96\x00"
-     "\xEC\xA2\x97\x00\xEC\xA2\x98\x00\xEC\xA2\x99\x00\xEC\xA2\x9A\x00\xEC\xA2"
-     "\x9B\x00"
-     "\xEC\xA2\x9C\x00\xEC\xA2\x9D\x00\xEC\xA2\x9E\x00\xEC\xA2\x9F\x00\xEC\xA2"
-     "\xA0\x00"
-     "\xEC\xA2\xA1\x00\xEC\xA2\xA2\x00\xEC\xA2\xA3\x00\xEC\xA2\xA4\x00\xEC\xA2"
-     "\xA5\x00"
-     "\xEC\xA2\xA6\x00\xEC\xA2\xA7\x00\xEC\xA2\xA8\x00\xEC\xA2\xA9\x00\xEC\xA2"
-     "\xAA\x00"
-     "\xEC\xA2\xAB\x00\xEC\xA2\xAC\x00\xEC\xA2\xAD\x00\xEC\xA2\xAE\x00\xEC\xA2"
-     "\xAF\x00"
-     "\xEC\xA2\xB0\x00\xEC\xA2\xB1\x00\xEC\xA2\xB2\x00\xEC\xA2\xB3\x00\xEC\xA2"
-     "\xB4\x00"
-     "\xEC\xA2\xB5\x00\xEC\xA2\xB6\x00\xEC\xA2\xB7\x00\xEC\xA2\xB8\x00\xEC\xA2"
-     "\xB9\x00"
-     "\xEC\xA2\xBA\x00\xEC\xA2\xBB\x00\xEC\xA2\xBC\x00\xEC\xA2\xBD\x00\xEC\xA2"
-     "\xBE\x00"
-     "\xEC\xA2\xBF\x00\xEC\xA3\x80\x00\xEC\xA3\x81\x00\xEC\xA3\x82\x00\xEC\xA3"
-     "\x83\x00"
-     "\xEC\xA3\x84\x00\xEC\xA3\x85\x00\xEC\xA3\x86\x00\xEC\xA3\x87\x00\xEC\xA3"
-     "\x88\x00"
-     "\xEC\xA3\x89\x00\xEC\xA3\x8A\x00\xEC\xA3\x8B\x00\xEC\xA3\x8C\x00\xEC\xA3"
-     "\x8D\x00"
-     "\xEC\xA3\x8E\x00\xEC\xA3\x8F\x00\xEC\xA3\x90\x00\xEC\xA3\x91\x00\xEC\xA3"
-     "\x92\x00"
-     "\xEC\xA3\x93\x00\xEC\xA3\x94\x00\xEC\xA3\x95\x00\xEC\xA3\x96\x00\xEC\xA3"
-     "\x97\x00"
-     "\xEC\xA3\x98\x00\xEC\xA3\x99\x00\xEC\xA3\x9A\x00\xEC\xA3\x9B\x00\xEC\xA3"
-     "\x9C\x00"
-     "\xEC\xA3\x9D\x00\xEC\xA3\x9E\x00\xEC\xA3\x9F\x00\xEC\xA3\xA0\x00\xEC\xA3"
-     "\xA1\x00"
-     "\xEC\xA3\xA2\x00\xEC\xA3\xA3\x00\xEC\xA3\xA4\x00\xEC\xA3\xA5\x00\xEC\xA3"
-     "\xA6\x00"
-     "\xEC\xA3\xA7\x00\xEC\xA3\xA8\x00\xEC\xA3\xA9\x00\xEC\xA3\xAA\x00\xEC\xA3"
-     "\xAB\x00"
-     "\xEC\xA3\xAC\x00\xEC\xA3\xAD\x00\xEC\xA3\xAE\x00\xEC\xA3\xAF\x00\xEC\xA3"
-     "\xB0\x00"
-     "\xEC\xA3\xB1\x00\xEC\xA3\xB2\x00\xEC\xA3\xB3\x00\xEC\xA3\xB4\x00\xEC\xA3"
-     "\xB5\x00"
-     "\xEC\xA3\xB6\x00\xEC\xA3\xB7\x00\xEC\xA3\xB8\x00\xEC\xA3\xB9\x00\xEC\xA3"
-     "\xBA\x00"
-     "\xEC\xA3\xBB\x00\xEC\xA3\xBC\x00\xEC\xA3\xBC\xEC\x9D\x98\x00\xEC\xA3\xBD"
-     "\x00\xEC"
-     "\xA3\xBE\x00\xEC\xA3\xBF\x00\xEC\xA4\x80\x00\xEC\xA4\x81\x00\xEC\xA4\x82"
-     "\x00\xEC"
-     "\xA4\x83\x00\xEC\xA4\x84\x00\xEC\xA4\x85\x00\xEC\xA4\x86\x00\xEC\xA4\x87"
-     "\x00\xEC"
-     "\xA4\x88\x00\xEC\xA4\x89\x00\xEC\xA4\x8A\x00\xEC\xA4\x8B\x00\xEC\xA4\x8C"
-     "\x00\xEC"
-     "\xA4\x8D\x00\xEC\xA4\x8E\x00\xEC\xA4\x8F\x00\xEC\xA4\x90\x00\xEC\xA4\x91"
-     "\x00\xEC"
-     "\xA4\x92\x00\xEC\xA4\x93\x00\xEC\xA4\x94\x00\xEC\xA4\x95\x00\xEC\xA4\x96"
-     "\x00\xEC"
-     "\xA4\x97\x00\xEC\xA4\x98\x00\xEC\xA4\x99\x00\xEC\xA4\x9A\x00\xEC\xA4\x9B"
-     "\x00\xEC"
-     "\xA4\x9C\x00\xEC\xA4\x9D\x00\xEC\xA4\x9E\x00\xEC\xA4\x9F\x00\xEC\xA4\xA0"
-     "\x00\xEC"
-     "\xA4\xA1\x00\xEC\xA4\xA2\x00\xEC\xA4\xA3\x00\xEC\xA4\xA4\x00\xEC\xA4\xA5"
-     "\x00\xEC"
-     "\xA4\xA6\x00\xEC\xA4\xA7\x00\xEC\xA4\xA8\x00\xEC\xA4\xA9\x00\xEC\xA4\xAA"
-     "\x00\xEC"
-     "\xA4\xAB\x00\xEC\xA4\xAC\x00\xEC\xA4\xAD\x00\xEC\xA4\xAE\x00\xEC\xA4\xAF"
-     "\x00\xEC"
-     "\xA4\xB0\x00\xEC\xA4\xB1\x00\xEC\xA4\xB2\x00\xEC\xA4\xB3\x00\xEC\xA4\xB4"
-     "\x00\xEC"
-     "\xA4\xB5\x00\xEC\xA4\xB6\x00\xEC\xA4\xB7\x00\xEC\xA4\xB8\x00\xEC\xA4\xB9"
-     "\x00\xEC"
-     "\xA4\xBA\x00\xEC\xA4\xBB\x00\xEC\xA4\xBC\x00\xEC\xA4\xBD\x00\xEC\xA4\xBE"
-     "\x00\xEC"
-     "\xA4\xBF\x00\xEC\xA5\x80\x00\xEC\xA5\x81\x00\xEC\xA5\x82\x00\xEC\xA5\x83"
-     "\x00\xEC"
-     "\xA5\x84\x00\xEC\xA5\x85\x00\xEC\xA5\x86\x00\xEC\xA5\x87\x00\xEC\xA5\x88"
-     "\x00\xEC"
-     "\xA5\x89\x00\xEC\xA5\x8A\x00\xEC\xA5\x8B\x00\xEC\xA5\x8C\x00\xEC\xA5\x8D"
-     "\x00\xEC"
-     "\xA5\x8E\x00\xEC\xA5\x8F\x00\xEC\xA5\x90\x00\xEC\xA5\x91\x00\xEC\xA5\x92"
-     "\x00\xEC"
-     "\xA5\x93\x00\xEC\xA5\x94\x00\xEC\xA5\x95\x00\xEC\xA5\x96\x00\xEC\xA5\x97"
-     "\x00\xEC"
-     "\xA5\x98\x00\xEC\xA5\x99\x00\xEC\xA5\x9A\x00\xEC\xA5\x9B\x00\xEC\xA5\x9C"
-     "\x00\xEC"
-     "\xA5\x9D\x00\xEC\xA5\x9E\x00\xEC\xA5\x9F\x00\xEC\xA5\xA0\x00\xEC\xA5\xA1"
-     "\x00\xEC"
-     "\xA5\xA2\x00\xEC\xA5\xA3\x00\xEC\xA5\xA4\x00\xEC\xA5\xA5\x00\xEC\xA5\xA6"
-     "\x00\xEC"
-     "\xA5\xA7\x00\xEC\xA5\xA8\x00\xEC\xA5\xA9\x00\xEC\xA5\xAA\x00\xEC\xA5\xAB"
-     "\x00\xEC"
-     "\xA5\xAC\x00\xEC\xA5\xAD\x00\xEC\xA5\xAE\x00\xEC\xA5\xAF\x00\xEC\xA5\xB0"
-     "\x00\xEC"
-     "\xA5\xB1\x00\xEC\xA5\xB2\x00\xEC\xA5\xB3\x00\xEC\xA5\xB4\x00\xEC\xA5\xB5"
-     "\x00\xEC"
-     "\xA5\xB6\x00\xEC\xA5\xB7\x00\xEC\xA5\xB8\x00\xEC\xA5\xB9\x00\xEC\xA5\xBA"
-     "\x00\xEC"
-     "\xA5\xBB\x00\xEC\xA5\xBC\x00\xEC\xA5\xBD\x00\xEC\xA5\xBE\x00\xEC\xA5\xBF"
-     "\x00\xEC"
-     "\xA6\x80\x00\xEC\xA6\x81\x00\xEC\xA6\x82\x00\xEC\xA6\x83\x00\xEC\xA6\x84"
-     "\x00\xEC"
-     "\xA6\x85\x00\xEC\xA6\x86\x00\xEC\xA6\x87\x00\xEC\xA6\x88\x00\xEC\xA6\x89"
-     "\x00\xEC"
-     "\xA6\x8A\x00\xEC\xA6\x8B\x00\xEC\xA6\x8C\x00\xEC\xA6\x8D\x00\xEC\xA6\x8E"
-     "\x00\xEC"
-     "\xA6\x8F\x00\xEC\xA6\x90\x00\xEC\xA6\x91\x00\xEC\xA6\x92\x00\xEC\xA6\x93"
-     "\x00\xEC"
-     "\xA6\x94\x00\xEC\xA6\x95\x00\xEC\xA6\x96\x00\xEC\xA6\x97\x00\xEC\xA6\x98"
-     "\x00\xEC"
-     "\xA6\x99\x00\xEC\xA6\x9A\x00\xEC\xA6\x9B\x00\xEC\xA6\x9C\x00\xEC\xA6\x9D"
-     "\x00\xEC"
-     "\xA6\x9E\x00\xEC\xA6\x9F\x00\xEC\xA6\xA0\x00\xEC\xA6\xA1\x00\xEC\xA6\xA2"
-     "\x00\xEC"
-     "\xA6\xA3\x00\xEC\xA6\xA4\x00\xEC\xA6\xA5\x00\xEC\xA6\xA6\x00\xEC\xA6\xA7"
-     "\x00\xEC"
-     "\xA6\xA8\x00\xEC\xA6\xA9\x00\xEC\xA6\xAA\x00\xEC\xA6\xAB\x00\xEC\xA6\xAC"
-     "\x00\xEC"
-     "\xA6\xAD\x00\xEC\xA6\xAE\x00\xEC\xA6\xAF\x00\xEC\xA6\xB0\x00\xEC\xA6\xB1"
-     "\x00\xEC"
-     "\xA6\xB2\x00\xEC\xA6\xB3\x00\xEC\xA6\xB4\x00\xEC\xA6\xB5\x00\xEC\xA6\xB6"
-     "\x00\xEC"
-     "\xA6\xB7\x00\xEC\xA6\xB8\x00\xEC\xA6\xB9\x00\xEC\xA6\xBA\x00\xEC\xA6\xBB"
-     "\x00\xEC"
-     "\xA6\xBC\x00\xEC\xA6\xBD\x00\xEC\xA6\xBE\x00\xEC\xA6\xBF\x00\xEC\xA7\x80"
-     "\x00\xEC"
-     "\xA7\x81\x00\xEC\xA7\x82\x00\xEC\xA7\x83\x00\xEC\xA7\x84\x00\xEC\xA7\x85"
-     "\x00\xEC"
-     "\xA7\x86\x00\xEC\xA7\x87\x00\xEC\xA7\x88\x00\xEC\xA7\x89\x00\xEC\xA7\x8A"
-     "\x00\xEC"
-     "\xA7\x8B\x00\xEC\xA7\x8C\x00\xEC\xA7\x8D\x00\xEC\xA7\x8E\x00\xEC\xA7\x8F"
-     "\x00\xEC"
-     "\xA7\x90\x00\xEC\xA7\x91\x00\xEC\xA7\x92\x00\xEC\xA7\x93\x00\xEC\xA7\x94"
-     "\x00\xEC"
-     "\xA7\x95\x00\xEC\xA7\x96\x00\xEC\xA7\x97\x00\xEC\xA7\x98\x00\xEC\xA7\x99"
-     "\x00\xEC"
-     "\xA7\x9A\x00\xEC\xA7\x9B\x00\xEC\xA7\x9C\x00\xEC\xA7\x9D\x00\xEC\xA7\x9E"
-     "\x00\xEC"
-     "\xA7\x9F\x00\xEC\xA7\xA0\x00\xEC\xA7\xA1\x00\xEC\xA7\xA2\x00\xEC\xA7\xA3"
-     "\x00\xEC"
-     "\xA7\xA4\x00\xEC\xA7\xA5\x00\xEC\xA7\xA6\x00\xEC\xA7\xA7\x00\xEC\xA7\xA8"
-     "\x00\xEC"
-     "\xA7\xA9\x00\xEC\xA7\xAA\x00\xEC\xA7\xAB\x00\xEC\xA7\xAC\x00\xEC\xA7\xAD"
-     "\x00\xEC"
-     "\xA7\xAE\x00\xEC\xA7\xAF\x00\xEC\xA7\xB0\x00\xEC\xA7\xB1\x00\xEC\xA7\xB2"
-     "\x00\xEC"
-     "\xA7\xB3\x00\xEC\xA7\xB4\x00\xEC\xA7\xB5\x00\xEC\xA7\xB6\x00\xEC\xA7\xB7"
-     "\x00\xEC"
-     "\xA7\xB8\x00\xEC\xA7\xB9\x00\xEC\xA7\xBA\x00\xEC\xA7\xBB\x00\xEC\xA7\xBC"
-     "\x00\xEC"
-     "\xA7\xBD\x00\xEC\xA7\xBE\x00\xEC\xA7\xBF\x00\xEC\xA8\x80\x00\xEC\xA8\x81"
-     "\x00\xEC"
-     "\xA8\x82\x00\xEC\xA8\x83\x00\xEC\xA8\x84\x00\xEC\xA8\x85\x00\xEC\xA8\x86"
-     "\x00\xEC"
-     "\xA8\x87\x00\xEC\xA8\x88\x00\xEC\xA8\x89\x00\xEC\xA8\x8A\x00\xEC\xA8\x8B"
-     "\x00\xEC"
-     "\xA8\x8C\x00\xEC\xA8\x8D\x00\xEC\xA8\x8E\x00\xEC\xA8\x8F\x00\xEC\xA8\x90"
-     "\x00\xEC"
-     "\xA8\x91\x00\xEC\xA8\x92\x00\xEC\xA8\x93\x00\xEC\xA8\x94\x00\xEC\xA8\x95"
-     "\x00\xEC"
-     "\xA8\x96\x00\xEC\xA8\x97\x00\xEC\xA8\x98\x00\xEC\xA8\x99\x00\xEC\xA8\x9A"
-     "\x00\xEC"
-     "\xA8\x9B\x00\xEC\xA8\x9C\x00\xEC\xA8\x9D\x00\xEC\xA8\x9E\x00\xEC\xA8\x9F"
-     "\x00\xEC"
-     "\xA8\xA0\x00\xEC\xA8\xA1\x00\xEC\xA8\xA2\x00\xEC\xA8\xA3\x00\xEC\xA8\xA4"
-     "\x00\xEC"
-     "\xA8\xA5\x00\xEC\xA8\xA6\x00\xEC\xA8\xA7\x00\xEC\xA8\xA8\x00\xEC\xA8\xA9"
-     "\x00\xEC"
-     "\xA8\xAA\x00\xEC\xA8\xAB\x00\xEC\xA8\xAC\x00\xEC\xA8\xAD\x00\xEC\xA8\xAE"
-     "\x00\xEC"
-     "\xA8\xAF\x00\xEC\xA8\xB0\x00\xEC\xA8\xB1\x00\xEC\xA8\xB2\x00\xEC\xA8\xB3"
-     "\x00\xEC"
-     "\xA8\xB4\x00\xEC\xA8\xB5\x00\xEC\xA8\xB6\x00\xEC\xA8\xB7\x00\xEC\xA8\xB8"
-     "\x00\xEC"
-     "\xA8\xB9\x00\xEC\xA8\xBA\x00\xEC\xA8\xBB\x00\xEC\xA8\xBC\x00\xEC\xA8\xBD"
-     "\x00\xEC"
-     "\xA8\xBE\x00\xEC\xA8\xBF\x00\xEC\xA9\x80\x00\xEC\xA9\x81\x00\xEC\xA9\x82"
-     "\x00\xEC"
-     "\xA9\x83\x00\xEC\xA9\x84\x00\xEC\xA9\x85\x00\xEC\xA9\x86\x00\xEC\xA9\x87"
-     "\x00\xEC"
-     "\xA9\x88\x00\xEC\xA9\x89\x00\xEC\xA9\x8A\x00\xEC\xA9\x8B\x00\xEC\xA9\x8C"
-     "\x00\xEC"
-     "\xA9\x8D\x00\xEC\xA9\x8E\x00\xEC\xA9\x8F\x00\xEC\xA9\x90\x00\xEC\xA9\x91"
-     "\x00\xEC"
-     "\xA9\x92\x00\xEC\xA9\x93\x00\xEC\xA9\x94\x00\xEC\xA9\x95\x00\xEC\xA9\x96"
-     "\x00\xEC"
-     "\xA9\x97\x00\xEC\xA9\x98\x00\xEC\xA9\x99\x00\xEC\xA9\x9A\x00\xEC\xA9\x9B"
-     "\x00\xEC"
-     "\xA9\x9C\x00\xEC\xA9\x9D\x00\xEC\xA9\x9E\x00\xEC\xA9\x9F\x00\xEC\xA9\xA0"
-     "\x00\xEC"
-     "\xA9\xA1\x00\xEC\xA9\xA2\x00\xEC\xA9\xA3\x00\xEC\xA9\xA4\x00\xEC\xA9\xA5"
-     "\x00\xEC"
-     "\xA9\xA6\x00\xEC\xA9\xA7\x00\xEC\xA9\xA8\x00\xEC\xA9\xA9\x00\xEC\xA9\xAA"
-     "\x00\xEC"
-     "\xA9\xAB\x00\xEC\xA9\xAC\x00\xEC\xA9\xAD\x00\xEC\xA9\xAE\x00\xEC\xA9\xAF"
-     "\x00\xEC"
-     "\xA9\xB0\x00\xEC\xA9\xB1\x00\xEC\xA9\xB2\x00\xEC\xA9\xB3\x00\xEC\xA9\xB4"
-     "\x00\xEC"
-     "\xA9\xB5\x00\xEC\xA9\xB6\x00\xEC\xA9\xB7\x00\xEC\xA9\xB8\x00\xEC\xA9\xB9"
-     "\x00\xEC"
-     "\xA9\xBA\x00\xEC\xA9\xBB\x00\xEC\xA9\xBC\x00\xEC\xA9\xBD\x00\xEC\xA9\xBE"
-     "\x00\xEC"
-     "\xA9\xBF\x00\xEC\xAA\x80\x00\xEC\xAA\x81\x00\xEC\xAA\x82\x00\xEC\xAA\x83"
-     "\x00\xEC"
-     "\xAA\x84\x00\xEC\xAA\x85\x00\xEC\xAA\x86\x00\xEC\xAA\x87\x00\xEC\xAA\x88"
-     "\x00\xEC"
-     "\xAA\x89\x00\xEC\xAA\x8A\x00\xEC\xAA\x8B\x00\xEC\xAA\x8C\x00\xEC\xAA\x8D"
-     "\x00\xEC"
-     "\xAA\x8E\x00\xEC\xAA\x8F\x00\xEC\xAA\x90\x00\xEC\xAA\x91\x00\xEC\xAA\x92"
-     "\x00\xEC"
-     "\xAA\x93\x00\xEC\xAA\x94\x00\xEC\xAA\x95\x00\xEC\xAA\x96\x00\xEC\xAA\x97"
-     "\x00\xEC"
-     "\xAA\x98\x00\xEC\xAA\x99\x00\xEC\xAA\x9A\x00\xEC\xAA\x9B\x00\xEC\xAA\x9C"
-     "\x00\xEC"
-     "\xAA\x9D\x00\xEC\xAA\x9E\x00\xEC\xAA\x9F\x00\xEC\xAA\xA0\x00\xEC\xAA\xA1"
-     "\x00\xEC"
-     "\xAA\xA2\x00\xEC\xAA\xA3\x00\xEC\xAA\xA4\x00\xEC\xAA\xA5\x00\xEC\xAA\xA6"
-     "\x00\xEC"
-     "\xAA\xA7\x00\xEC\xAA\xA8\x00\xEC\xAA\xA9\x00\xEC\xAA\xAA\x00\xEC\xAA\xAB"
-     "\x00\xEC"
-     "\xAA\xAC\x00\xEC\xAA\xAD\x00\xEC\xAA\xAE\x00\xEC\xAA\xAF\x00\xEC\xAA\xB0"
-     "\x00\xEC"
-     "\xAA\xB1\x00\xEC\xAA\xB2\x00\xEC\xAA\xB3\x00\xEC\xAA\xB4\x00\xEC\xAA\xB5"
-     "\x00\xEC"
-     "\xAA\xB6\x00\xEC\xAA\xB7\x00\xEC\xAA\xB8\x00\xEC\xAA\xB9\x00\xEC\xAA\xBA"
-     "\x00\xEC"
-     "\xAA\xBB\x00\xEC\xAA\xBC\x00\xEC\xAA\xBD\x00\xEC\xAA\xBE\x00\xEC\xAA\xBF"
-     "\x00\xEC"
-     "\xAB\x80\x00\xEC\xAB\x81\x00\xEC\xAB\x82\x00\xEC\xAB\x83\x00\xEC\xAB\x84"
-     "\x00\xEC"
-     "\xAB\x85\x00\xEC\xAB\x86\x00\xEC\xAB\x87\x00\xEC\xAB\x88\x00\xEC\xAB\x89"
-     "\x00\xEC"
-     "\xAB\x8A\x00\xEC\xAB\x8B\x00\xEC\xAB\x8C\x00\xEC\xAB\x8D\x00\xEC\xAB\x8E"
-     "\x00\xEC"
-     "\xAB\x8F\x00\xEC\xAB\x90\x00\xEC\xAB\x91\x00\xEC\xAB\x92\x00\xEC\xAB\x93"
-     "\x00\xEC"
-     "\xAB\x94\x00\xEC\xAB\x95\x00\xEC\xAB\x96\x00\xEC\xAB\x97\x00\xEC\xAB\x98"
-     "\x00\xEC"
-     "\xAB\x99\x00\xEC\xAB\x9A\x00\xEC\xAB\x9B\x00\xEC\xAB\x9C\x00\xEC\xAB\x9D"
-     "\x00\xEC"
-     "\xAB\x9E\x00\xEC\xAB\x9F\x00\xEC\xAB\xA0\x00\xEC\xAB\xA1\x00\xEC\xAB\xA2"
-     "\x00\xEC"
-     "\xAB\xA3\x00\xEC\xAB\xA4\x00\xEC\xAB\xA5\x00\xEC\xAB\xA6\x00\xEC\xAB\xA7"
-     "\x00\xEC"
-     "\xAB\xA8\x00\xEC\xAB\xA9\x00\xEC\xAB\xAA\x00\xEC\xAB\xAB\x00\xEC\xAB\xAC"
-     "\x00\xEC"
-     "\xAB\xAD\x00\xEC\xAB\xAE\x00\xEC\xAB\xAF\x00\xEC\xAB\xB0\x00\xEC\xAB\xB1"
-     "\x00\xEC"
-     "\xAB\xB2\x00\xEC\xAB\xB3\x00\xEC\xAB\xB4\x00\xEC\xAB\xB5\x00\xEC\xAB\xB6"
-     "\x00\xEC"
-     "\xAB\xB7\x00\xEC\xAB\xB8\x00\xEC\xAB\xB9\x00\xEC\xAB\xBA\x00\xEC\xAB\xBB"
-     "\x00\xEC"
-     "\xAB\xBC\x00\xEC\xAB\xBD\x00\xEC\xAB\xBE\x00\xEC\xAB\xBF\x00\xEC\xAC\x80"
-     "\x00\xEC"
-     "\xAC\x81\x00\xEC\xAC\x82\x00\xEC\xAC\x83\x00\xEC\xAC\x84\x00\xEC\xAC\x85"
-     "\x00\xEC"
-     "\xAC\x86\x00\xEC\xAC\x87\x00\xEC\xAC\x88\x00\xEC\xAC\x89\x00\xEC\xAC\x8A"
-     "\x00\xEC"
-     "\xAC\x8B\x00\xEC\xAC\x8C\x00\xEC\xAC\x8D\x00\xEC\xAC\x8E\x00\xEC\xAC\x8F"
-     "\x00\xEC"
-     "\xAC\x90\x00\xEC\xAC\x91\x00\xEC\xAC\x92\x00\xEC\xAC\x93\x00\xEC\xAC\x94"
-     "\x00\xEC"
-     "\xAC\x95\x00\xEC\xAC\x96\x00\xEC\xAC\x97\x00\xEC\xAC\x98\x00\xEC\xAC\x99"
-     "\x00\xEC"
-     "\xAC\x9A\x00\xEC\xAC\x9B\x00\xEC\xAC\x9C\x00\xEC\xAC\x9D\x00\xEC\xAC\x9E"
-     "\x00\xEC"
-     "\xAC\x9F\x00\xEC\xAC\xA0\x00\xEC\xAC\xA1\x00\xEC\xAC\xA2\x00\xEC\xAC\xA3"
-     "\x00\xEC"
-     "\xAC\xA4\x00\xEC\xAC\xA5\x00\xEC\xAC\xA6\x00\xEC\xAC\xA7\x00\xEC\xAC\xA8"
-     "\x00\xEC"
-     "\xAC\xA9\x00\xEC\xAC\xAA\x00\xEC\xAC\xAB\x00\xEC\xAC\xAC\x00\xEC\xAC\xAD"
-     "\x00\xEC"
-     "\xAC\xAE\x00\xEC\xAC\xAF\x00\xEC\xAC\xB0\x00\xEC\xAC\xB1\x00\xEC\xAC\xB2"
-     "\x00\xEC"
-     "\xAC\xB3\x00\xEC\xAC\xB4\x00\xEC\xAC\xB5\x00\xEC\xAC\xB6\x00\xEC\xAC\xB7"
-     "\x00\xEC"
-     "\xAC\xB8\x00\xEC\xAC\xB9\x00\xEC\xAC\xBA\x00\xEC\xAC\xBB\x00\xEC\xAC\xBC"
-     "\x00\xEC"
-     "\xAC\xBD\x00\xEC\xAC\xBE\x00\xEC\xAC\xBF\x00\xEC\xAD\x80\x00\xEC\xAD\x81"
-     "\x00\xEC"
-     "\xAD\x82\x00\xEC\xAD\x83\x00\xEC\xAD\x84\x00\xEC\xAD\x85\x00\xEC\xAD\x86"
-     "\x00\xEC"
-     "\xAD\x87\x00\xEC\xAD\x88\x00\xEC\xAD\x89\x00\xEC\xAD\x8A\x00\xEC\xAD\x8B"
-     "\x00\xEC"
-     "\xAD\x8C\x00\xEC\xAD\x8D\x00\xEC\xAD\x8E\x00\xEC\xAD\x8F\x00\xEC\xAD\x90"
-     "\x00\xEC"
-     "\xAD\x91\x00\xEC\xAD\x92\x00\xEC\xAD\x93\x00\xEC\xAD\x94\x00\xEC\xAD\x95"
-     "\x00\xEC"
-     "\xAD\x96\x00\xEC\xAD\x97\x00\xEC\xAD\x98\x00\xEC\xAD\x99\x00\xEC\xAD\x9A"
-     "\x00\xEC"
-     "\xAD\x9B\x00\xEC\xAD\x9C\x00\xEC\xAD\x9D\x00\xEC\xAD\x9E\x00\xEC\xAD\x9F"
-     "\x00\xEC"
-     "\xAD\xA0\x00\xEC\xAD\xA1\x00\xEC\xAD\xA2\x00\xEC\xAD\xA3\x00\xEC\xAD\xA4"
-     "\x00\xEC"
-     "\xAD\xA5\x00\xEC\xAD\xA6\x00\xEC\xAD\xA7\x00\xEC\xAD\xA8\x00\xEC\xAD\xA9"
-     "\x00\xEC"
-     "\xAD\xAA\x00\xEC\xAD\xAB\x00\xEC\xAD\xAC\x00\xEC\xAD\xAD\x00\xEC\xAD\xAE"
-     "\x00\xEC"
-     "\xAD\xAF\x00\xEC\xAD\xB0\x00\xEC\xAD\xB1\x00\xEC\xAD\xB2\x00\xEC\xAD\xB3"
-     "\x00\xEC"
-     "\xAD\xB4\x00\xEC\xAD\xB5\x00\xEC\xAD\xB6\x00\xEC\xAD\xB7\x00\xEC\xAD\xB8"
-     "\x00\xEC"
-     "\xAD\xB9\x00\xEC\xAD\xBA\x00\xEC\xAD\xBB\x00\xEC\xAD\xBC\x00\xEC\xAD\xBD"
-     "\x00\xEC"
-     "\xAD\xBE\x00\xEC\xAD\xBF\x00\xEC\xAE\x80\x00\xEC\xAE\x81\x00\xEC\xAE\x82"
-     "\x00\xEC"
-     "\xAE\x83\x00\xEC\xAE\x84\x00\xEC\xAE\x85\x00\xEC\xAE\x86\x00\xEC\xAE\x87"
-     "\x00\xEC"
-     "\xAE\x88\x00\xEC\xAE\x89\x00\xEC\xAE\x8A\x00\xEC\xAE\x8B\x00\xEC\xAE\x8C"
-     "\x00\xEC"
-     "\xAE\x8D\x00\xEC\xAE\x8E\x00\xEC\xAE\x8F\x00\xEC\xAE\x90\x00\xEC\xAE\x91"
-     "\x00\xEC"
-     "\xAE\x92\x00\xEC\xAE\x93\x00\xEC\xAE\x94\x00\xEC\xAE\x95\x00\xEC\xAE\x96"
-     "\x00\xEC"
-     "\xAE\x97\x00\xEC\xAE\x98\x00\xEC\xAE\x99\x00\xEC\xAE\x9A\x00\xEC\xAE\x9B"
-     "\x00\xEC"
-     "\xAE\x9C\x00\xEC\xAE\x9D\x00\xEC\xAE\x9E\x00\xEC\xAE\x9F\x00\xEC\xAE\xA0"
-     "\x00\xEC"
-     "\xAE\xA1\x00\xEC\xAE\xA2\x00\xEC\xAE\xA3\x00\xEC\xAE\xA4\x00\xEC\xAE\xA5"
-     "\x00\xEC"
-     "\xAE\xA6\x00\xEC\xAE\xA7\x00\xEC\xAE\xA8\x00\xEC\xAE\xA9\x00\xEC\xAE\xAA"
-     "\x00\xEC"
-     "\xAE\xAB\x00\xEC\xAE\xAC\x00\xEC\xAE\xAD\x00\xEC\xAE\xAE\x00\xEC\xAE\xAF"
-     "\x00\xEC"
-     "\xAE\xB0\x00\xEC\xAE\xB1\x00\xEC\xAE\xB2\x00\xEC\xAE\xB3\x00\xEC\xAE\xB4"
-     "\x00\xEC"
-     "\xAE\xB5\x00\xEC\xAE\xB6\x00\xEC\xAE\xB7\x00\xEC\xAE\xB8\x00\xEC\xAE\xB9"
-     "\x00\xEC"
-     "\xAE\xBA\x00\xEC\xAE\xBB\x00\xEC\xAE\xBC\x00\xEC\xAE\xBD\x00\xEC\xAE\xBE"
-     "\x00\xEC"
-     "\xAE\xBF\x00\xEC\xAF\x80\x00\xEC\xAF\x81\x00\xEC\xAF\x82\x00\xEC\xAF\x83"
-     "\x00\xEC"
-     "\xAF\x84\x00\xEC\xAF\x85\x00\xEC\xAF\x86\x00\xEC\xAF\x87\x00\xEC\xAF\x88"
-     "\x00\xEC"
-     "\xAF\x89\x00\xEC\xAF\x8A\x00\xEC\xAF\x8B\x00\xEC\xAF\x8C\x00\xEC\xAF\x8D"
-     "\x00\xEC"
-     "\xAF\x8E\x00\xEC\xAF\x8F\x00\xEC\xAF\x90\x00\xEC\xAF\x91\x00\xEC\xAF\x92"
-     "\x00\xEC"
-     "\xAF\x93\x00\xEC\xAF\x94\x00\xEC\xAF\x95\x00\xEC\xAF\x96\x00\xEC\xAF\x97"
-     "\x00\xEC"
-     "\xAF\x98\x00\xEC\xAF\x99\x00\xEC\xAF\x9A\x00\xEC\xAF\x9B\x00\xEC\xAF\x9C"
-     "\x00\xEC"
-     "\xAF\x9D\x00\xEC\xAF\x9E\x00\xEC\xAF\x9F\x00\xEC\xAF\xA0\x00\xEC\xAF\xA1"
-     "\x00\xEC"
-     "\xAF\xA2\x00\xEC\xAF\xA3\x00\xEC\xAF\xA4\x00\xEC\xAF\xA5\x00\xEC\xAF\xA6"
-     "\x00\xEC"
-     "\xAF\xA7\x00\xEC\xAF\xA8\x00\xEC\xAF\xA9\x00\xEC\xAF\xAA\x00\xEC\xAF\xAB"
-     "\x00\xEC"
-     "\xAF\xAC\x00\xEC\xAF\xAD\x00\xEC\xAF\xAE\x00\xEC\xAF\xAF\x00\xEC\xAF\xB0"
-     "\x00\xEC"
-     "\xAF\xB1\x00\xEC\xAF\xB2\x00\xEC\xAF\xB3\x00\xEC\xAF\xB4\x00\xEC\xAF\xB5"
-     "\x00\xEC"
-     "\xAF\xB6\x00\xEC\xAF\xB7\x00\xEC\xAF\xB8\x00\xEC\xAF\xB9\x00\xEC\xAF\xBA"
-     "\x00\xEC"
-     "\xAF\xBB\x00\xEC\xAF\xBC\x00\xEC\xAF\xBD\x00\xEC\xAF\xBE\x00\xEC\xAF\xBF"
-     "\x00\xEC"
-     "\xB0\x80\x00\xEC\xB0\x81\x00\xEC\xB0\x82\x00\xEC\xB0\x83\x00\xEC\xB0\x84"
-     "\x00\xEC"
-     "\xB0\x85\x00\xEC\xB0\x86\x00\xEC\xB0\x87\x00\xEC\xB0\x88\x00\xEC\xB0\x89"
-     "\x00\xEC"
-     "\xB0\x8A\x00\xEC\xB0\x8B\x00\xEC\xB0\x8C\x00\xEC\xB0\x8D\x00\xEC\xB0\x8E"
-     "\x00\xEC"
-     "\xB0\x8F\x00\xEC\xB0\x90\x00\xEC\xB0\x91\x00\xEC\xB0\x92\x00\xEC\xB0\x93"
-     "\x00\xEC"
-     "\xB0\x94\x00\xEC\xB0\x95\x00\xEC\xB0\x96\x00\xEC\xB0\x97\x00\xEC\xB0\x98"
-     "\x00\xEC"
-     "\xB0\x99\x00\xEC\xB0\x9A\x00\xEC\xB0\x9B\x00\xEC\xB0\x9C\x00\xEC\xB0\x9D"
-     "\x00\xEC"
-     "\xB0\x9E\x00\xEC\xB0\x9F\x00\xEC\xB0\xA0\x00\xEC\xB0\xA1\x00\xEC\xB0\xA2"
-     "\x00\xEC"
-     "\xB0\xA3\x00\xEC\xB0\xA4\x00\xEC\xB0\xA5\x00\xEC\xB0\xA6\x00\xEC\xB0\xA7"
-     "\x00\xEC"
-     "\xB0\xA8\x00\xEC\xB0\xA9\x00\xEC\xB0\xAA\x00\xEC\xB0\xAB\x00\xEC\xB0\xAC"
-     "\x00\xEC"
-     "\xB0\xAD\x00\xEC\xB0\xAE\x00\xEC\xB0\xAF\x00\xEC\xB0\xB0\x00\xEC\xB0\xB1"
-     "\x00\xEC"
-     "\xB0\xB2\x00\xEC\xB0\xB3\x00\xEC\xB0\xB4\x00\xEC\xB0\xB5\x00\xEC\xB0\xB6"
-     "\x00\xEC"
-     "\xB0\xB7\x00\xEC\xB0\xB8\x00\xEC\xB0\xB8\xEA\xB3\xA0\x00\xEC\xB0\xB9\x00"
-     "\xEC\xB0"
-     "\xBA\x00\xEC\xB0\xBB\x00\xEC\xB0\xBC\x00\xEC\xB0\xBD\x00\xEC\xB0\xBE\x00"
-     "\xEC\xB0"
-     "\xBF\x00\xEC\xB1\x80\x00\xEC\xB1\x81\x00\xEC\xB1\x82\x00\xEC\xB1\x83\x00"
-     "\xEC\xB1"
-     "\x84\x00\xEC\xB1\x85\x00\xEC\xB1\x86\x00\xEC\xB1\x87\x00\xEC\xB1\x88\x00"
-     "\xEC\xB1"
-     "\x89\x00\xEC\xB1\x8A\x00\xEC\xB1\x8B\x00\xEC\xB1\x8C\x00\xEC\xB1\x8D\x00"
-     "\xEC\xB1"
-     "\x8E\x00\xEC\xB1\x8F\x00\xEC\xB1\x90\x00\xEC\xB1\x91\x00\xEC\xB1\x92\x00"
-     "\xEC\xB1"
-     "\x93\x00\xEC\xB1\x94\x00\xEC\xB1\x95\x00\xEC\xB1\x96\x00\xEC\xB1\x97\x00"
-     "\xEC\xB1"
-     "\x98\x00\xEC\xB1\x99\x00\xEC\xB1\x9A\x00\xEC\xB1\x9B\x00\xEC\xB1\x9C\x00"
-     "\xEC\xB1"
-     "\x9D\x00\xEC\xB1\x9E\x00\xEC\xB1\x9F\x00\xEC\xB1\xA0\x00\xEC\xB1\xA1\x00"
-     "\xEC\xB1"
-     "\xA2\x00\xEC\xB1\xA3\x00\xEC\xB1\xA4\x00\xEC\xB1\xA5\x00\xEC\xB1\xA6\x00"
-     "\xEC\xB1"
-     "\xA7\x00\xEC\xB1\xA8\x00\xEC\xB1\xA9\x00\xEC\xB1\xAA\x00\xEC\xB1\xAB\x00"
-     "\xEC\xB1"
-     "\xAC\x00\xEC\xB1\xAD\x00\xEC\xB1\xAE\x00\xEC\xB1\xAF\x00\xEC\xB1\xB0\x00"
-     "\xEC\xB1"
-     "\xB1\x00\xEC\xB1\xB2\x00\xEC\xB1\xB3\x00\xEC\xB1\xB4\x00\xEC\xB1\xB5\x00"
-     "\xEC\xB1"
-     "\xB6\x00\xEC\xB1\xB7\x00\xEC\xB1\xB8\x00\xEC\xB1\xB9\x00\xEC\xB1\xBA\x00"
-     "\xEC\xB1"
-     "\xBB\x00\xEC\xB1\xBC\x00\xEC\xB1\xBD\x00\xEC\xB1\xBE\x00\xEC\xB1\xBF\x00"
-     "\xEC\xB2"
-     "\x80\x00\xEC\xB2\x81\x00\xEC\xB2\x82\x00\xEC\xB2\x83\x00\xEC\xB2\x84\x00"
-     "\xEC\xB2"
-     "\x85\x00\xEC\xB2\x86\x00\xEC\xB2\x87\x00\xEC\xB2\x88\x00\xEC\xB2\x89\x00"
-     "\xEC\xB2"
-     "\x8A\x00\xEC\xB2\x8B\x00\xEC\xB2\x8C\x00\xEC\xB2\x8D\x00\xEC\xB2\x8E\x00"
-     "\xEC\xB2"
-     "\x8F\x00\xEC\xB2\x90\x00\xEC\xB2\x91\x00\xEC\xB2\x92\x00\xEC\xB2\x93\x00"
-     "\xEC\xB2"
-     "\x94\x00\xEC\xB2\x95\x00\xEC\xB2\x96\x00\xEC\xB2\x97\x00\xEC\xB2\x98\x00"
-     "\xEC\xB2"
-     "\x99\x00\xEC\xB2\x9A\x00\xEC\xB2\x9B\x00\xEC\xB2\x9C\x00\xEC\xB2\x9D\x00"
-     "\xEC\xB2"
-     "\x9E\x00\xEC\xB2\x9F\x00\xEC\xB2\xA0\x00\xEC\xB2\xA1\x00\xEC\xB2\xA2\x00"
-     "\xEC\xB2"
-     "\xA3\x00\xEC\xB2\xA4\x00\xEC\xB2\xA5\x00\xEC\xB2\xA6\x00\xEC\xB2\xA7\x00"
-     "\xEC\xB2"
-     "\xA8\x00\xEC\xB2\xA9\x00\xEC\xB2\xAA\x00\xEC\xB2\xAB\x00\xEC\xB2\xAC\x00"
-     "\xEC\xB2"
-     "\xAD\x00\xEC\xB2\xAE\x00\xEC\xB2\xAF\x00\xEC\xB2\xB0\x00\xEC\xB2\xB1\x00"
-     "\xEC\xB2"
-     "\xB2\x00\xEC\xB2\xB3\x00\xEC\xB2\xB4\x00\xEC\xB2\xB5\x00\xEC\xB2\xB6\x00"
-     "\xEC\xB2"
-     "\xB7\x00\xEC\xB2\xB8\x00\xEC\xB2\xB9\x00\xEC\xB2\xBA\x00\xEC\xB2\xBB\x00"
-     "\xEC\xB2"
-     "\xBC\x00\xEC\xB2\xBD\x00\xEC\xB2\xBE\x00\xEC\xB2\xBF\x00\xEC\xB3\x80\x00"
-     "\xEC\xB3"
-     "\x81\x00\xEC\xB3\x82\x00\xEC\xB3\x83\x00\xEC\xB3\x84\x00\xEC\xB3\x85\x00"
-     "\xEC\xB3"
-     "\x86\x00\xEC\xB3\x87\x00\xEC\xB3\x88\x00\xEC\xB3\x89\x00\xEC\xB3\x8A\x00"
-     "\xEC\xB3"
-     "\x8B\x00\xEC\xB3\x8C\x00\xEC\xB3\x8D\x00\xEC\xB3\x8E\x00\xEC\xB3\x8F\x00"
-     "\xEC\xB3"
-     "\x90\x00\xEC\xB3\x91\x00\xEC\xB3\x92\x00\xEC\xB3\x93\x00\xEC\xB3\x94\x00"
-     "\xEC\xB3"
-     "\x95\x00\xEC\xB3\x96\x00\xEC\xB3\x97\x00\xEC\xB3\x98\x00\xEC\xB3\x99\x00"
-     "\xEC\xB3"
-     "\x9A\x00\xEC\xB3\x9B\x00\xEC\xB3\x9C\x00\xEC\xB3\x9D\x00\xEC\xB3\x9E\x00"
-     "\xEC\xB3"
-     "\x9F\x00\xEC\xB3\xA0\x00\xEC\xB3\xA1\x00\xEC\xB3\xA2\x00\xEC\xB3\xA3\x00"
-     "\xEC\xB3"
-     "\xA4\x00\xEC\xB3\xA5\x00\xEC\xB3\xA6\x00\xEC\xB3\xA7\x00\xEC\xB3\xA8\x00"
-     "\xEC\xB3"
-     "\xA9\x00\xEC\xB3\xAA\x00\xEC\xB3\xAB\x00\xEC\xB3\xAC\x00\xEC\xB3\xAD\x00"
-     "\xEC\xB3"
-     "\xAE\x00\xEC\xB3\xAF\x00\xEC\xB3\xB0\x00\xEC\xB3\xB1\x00\xEC\xB3\xB2\x00"
-     "\xEC\xB3"
-     "\xB3\x00\xEC\xB3\xB4\x00\xEC\xB3\xB5\x00\xEC\xB3\xB6\x00\xEC\xB3\xB7\x00"
-     "\xEC\xB3"
-     "\xB8\x00\xEC\xB3\xB9\x00\xEC\xB3\xBA\x00\xEC\xB3\xBB\x00\xEC\xB3\xBC\x00"
-     "\xEC\xB3"
-     "\xBD\x00\xEC\xB3\xBE\x00\xEC\xB3\xBF\x00\xEC\xB4\x80\x00\xEC\xB4\x81\x00"
-     "\xEC\xB4"
-     "\x82\x00\xEC\xB4\x83\x00\xEC\xB4\x84\x00\xEC\xB4\x85\x00\xEC\xB4\x86\x00"
-     "\xEC\xB4"
-     "\x87\x00\xEC\xB4\x88\x00\xEC\xB4\x89\x00\xEC\xB4\x8A\x00\xEC\xB4\x8B\x00"
-     "\xEC\xB4"
-     "\x8C\x00\xEC\xB4\x8D\x00\xEC\xB4\x8E\x00\xEC\xB4\x8F\x00\xEC\xB4\x90\x00"
-     "\xEC\xB4"
-     "\x91\x00\xEC\xB4\x92\x00\xEC\xB4\x93\x00\xEC\xB4\x94\x00\xEC\xB4\x95\x00"
-     "\xEC\xB4"
-     "\x96\x00\xEC\xB4\x97\x00\xEC\xB4\x98\x00\xEC\xB4\x99\x00\xEC\xB4\x9A\x00"
-     "\xEC\xB4"
-     "\x9B\x00\xEC\xB4\x9C\x00\xEC\xB4\x9D\x00\xEC\xB4\x9E\x00\xEC\xB4\x9F\x00"
-     "\xEC\xB4"
-     "\xA0\x00\xEC\xB4\xA1\x00\xEC\xB4\xA2\x00\xEC\xB4\xA3\x00\xEC\xB4\xA4\x00"
-     "\xEC\xB4"
-     "\xA5\x00\xEC\xB4\xA6\x00\xEC\xB4\xA7\x00\xEC\xB4\xA8\x00\xEC\xB4\xA9\x00"
-     "\xEC\xB4"
-     "\xAA\x00\xEC\xB4\xAB\x00\xEC\xB4\xAC\x00\xEC\xB4\xAD\x00\xEC\xB4\xAE\x00"
-     "\xEC\xB4"
-     "\xAF\x00\xEC\xB4\xB0\x00\xEC\xB4\xB1\x00\xEC\xB4\xB2\x00\xEC\xB4\xB3\x00"
-     "\xEC\xB4"
-     "\xB4\x00\xEC\xB4\xB5\x00\xEC\xB4\xB6\x00\xEC\xB4\xB7\x00\xEC\xB4\xB8\x00"
-     "\xEC\xB4"
-     "\xB9\x00\xEC\xB4\xBA\x00\xEC\xB4\xBB\x00\xEC\xB4\xBC\x00\xEC\xB4\xBD\x00"
-     "\xEC\xB4"
-     "\xBE\x00\xEC\xB4\xBF\x00\xEC\xB5\x80\x00\xEC\xB5\x81\x00\xEC\xB5\x82\x00"
-     "\xEC\xB5"
-     "\x83\x00\xEC\xB5\x84\x00\xEC\xB5\x85\x00\xEC\xB5\x86\x00\xEC\xB5\x87\x00"
-     "\xEC\xB5"
-     "\x88\x00\xEC\xB5\x89\x00\xEC\xB5\x8A\x00\xEC\xB5\x8B\x00\xEC\xB5\x8C\x00"
-     "\xEC\xB5"
-     "\x8D\x00\xEC\xB5\x8E\x00\xEC\xB5\x8F\x00\xEC\xB5\x90\x00\xEC\xB5\x91\x00"
-     "\xEC\xB5"
-     "\x92\x00\xEC\xB5\x93\x00\xEC\xB5\x94\x00\xEC\xB5\x95\x00\xEC\xB5\x96\x00"
-     "\xEC\xB5"
-     "\x97\x00\xEC\xB5\x98\x00\xEC\xB5\x99\x00\xEC\xB5\x9A\x00\xEC\xB5\x9B\x00"
-     "\xEC\xB5"
-     "\x9C\x00\xEC\xB5\x9D\x00\xEC\xB5\x9E\x00\xEC\xB5\x9F\x00\xEC\xB5\xA0\x00"
-     "\xEC\xB5"
-     "\xA1\x00\xEC\xB5\xA2\x00\xEC\xB5\xA3\x00\xEC\xB5\xA4\x00\xEC\xB5\xA5\x00"
-     "\xEC\xB5"
-     "\xA6\x00\xEC\xB5\xA7\x00\xEC\xB5\xA8\x00\xEC\xB5\xA9\x00\xEC\xB5\xAA\x00"
-     "\xEC\xB5"
-     "\xAB\x00\xEC\xB5\xAC\x00\xEC\xB5\xAD\x00\xEC\xB5\xAE\x00\xEC\xB5\xAF\x00"
-     "\xEC\xB5"
-     "\xB0\x00\xEC\xB5\xB1\x00\xEC\xB5\xB2\x00\xEC\xB5\xB3\x00\xEC\xB5\xB4\x00"
-     "\xEC\xB5"
-     "\xB5\x00\xEC\xB5\xB6\x00\xEC\xB5\xB7\x00\xEC\xB5\xB8\x00\xEC\xB5\xB9\x00"
-     "\xEC\xB5"
-     "\xBA\x00\xEC\xB5\xBB\x00\xEC\xB5\xBC\x00\xEC\xB5\xBD\x00\xEC\xB5\xBE\x00"
-     "\xEC\xB5"
-     "\xBF\x00\xEC\xB6\x80\x00\xEC\xB6\x81\x00\xEC\xB6\x82\x00\xEC\xB6\x83\x00"
-     "\xEC\xB6"
-     "\x84\x00\xEC\xB6\x85\x00\xEC\xB6\x86\x00\xEC\xB6\x87\x00\xEC\xB6\x88\x00"
-     "\xEC\xB6"
-     "\x89\x00\xEC\xB6\x8A\x00\xEC\xB6\x8B\x00\xEC\xB6\x8C\x00\xEC\xB6\x8D\x00"
-     "\xEC\xB6"
-     "\x8E\x00\xEC\xB6\x8F\x00\xEC\xB6\x90\x00\xEC\xB6\x91\x00\xEC\xB6\x92\x00"
-     "\xEC\xB6"
-     "\x93\x00\xEC\xB6\x94\x00\xEC\xB6\x95\x00\xEC\xB6\x96\x00\xEC\xB6\x97\x00"
-     "\xEC\xB6"
-     "\x98\x00\xEC\xB6\x99\x00\xEC\xB6\x9A\x00\xEC\xB6\x9B\x00\xEC\xB6\x9C\x00"
-     "\xEC\xB6"
-     "\x9D\x00\xEC\xB6\x9E\x00\xEC\xB6\x9F\x00\xEC\xB6\xA0\x00\xEC\xB6\xA1\x00"
-     "\xEC\xB6"
-     "\xA2\x00\xEC\xB6\xA3\x00\xEC\xB6\xA4\x00\xEC\xB6\xA5\x00\xEC\xB6\xA6\x00"
-     "\xEC\xB6"
-     "\xA7\x00\xEC\xB6\xA8\x00\xEC\xB6\xA9\x00\xEC\xB6\xAA\x00\xEC\xB6\xAB\x00"
-     "\xEC\xB6"
-     "\xAC\x00\xEC\xB6\xAD\x00\xEC\xB6\xAE\x00\xEC\xB6\xAF\x00\xEC\xB6\xB0\x00"
-     "\xEC\xB6"
-     "\xB1\x00\xEC\xB6\xB2\x00\xEC\xB6\xB3\x00\xEC\xB6\xB4\x00\xEC\xB6\xB5\x00"
-     "\xEC\xB6"
-     "\xB6\x00\xEC\xB6\xB7\x00\xEC\xB6\xB8\x00\xEC\xB6\xB9\x00\xEC\xB6\xBA\x00"
-     "\xEC\xB6"
-     "\xBB\x00\xEC\xB6\xBC\x00\xEC\xB6\xBD\x00\xEC\xB6\xBE\x00\xEC\xB6\xBF\x00"
-     "\xEC\xB7"
-     "\x80\x00\xEC\xB7\x81\x00\xEC\xB7\x82\x00\xEC\xB7\x83\x00\xEC\xB7\x84\x00"
-     "\xEC\xB7"
-     "\x85\x00\xEC\xB7\x86\x00\xEC\xB7\x87\x00\xEC\xB7\x88\x00\xEC\xB7\x89\x00"
-     "\xEC\xB7"
-     "\x8A\x00\xEC\xB7\x8B\x00\xEC\xB7\x8C\x00\xEC\xB7\x8D\x00\xEC\xB7\x8E\x00"
-     "\xEC\xB7"
-     "\x8F\x00\xEC\xB7\x90\x00\xEC\xB7\x91\x00\xEC\xB7\x92\x00\xEC\xB7\x93\x00"
-     "\xEC\xB7"
-     "\x94\x00\xEC\xB7\x95\x00\xEC\xB7\x96\x00\xEC\xB7\x97\x00\xEC\xB7\x98\x00"
-     "\xEC\xB7"
-     "\x99\x00\xEC\xB7\x9A\x00\xEC\xB7\x9B\x00\xEC\xB7\x9C\x00\xEC\xB7\x9D\x00"
-     "\xEC\xB7"
-     "\x9E\x00\xEC\xB7\x9F\x00\xEC\xB7\xA0\x00\xEC\xB7\xA1\x00\xEC\xB7\xA2\x00"
-     "\xEC\xB7"
-     "\xA3\x00\xEC\xB7\xA4\x00\xEC\xB7\xA5\x00\xEC\xB7\xA6\x00\xEC\xB7\xA7\x00"
-     "\xEC\xB7"
-     "\xA8\x00\xEC\xB7\xA9\x00\xEC\xB7\xAA\x00\xEC\xB7\xAB\x00\xEC\xB7\xAC\x00"
-     "\xEC\xB7"
-     "\xAD\x00\xEC\xB7\xAE\x00\xEC\xB7\xAF\x00\xEC\xB7\xB0\x00\xEC\xB7\xB1\x00"
-     "\xEC\xB7"
-     "\xB2\x00\xEC\xB7\xB3\x00\xEC\xB7\xB4\x00\xEC\xB7\xB5\x00\xEC\xB7\xB6\x00"
-     "\xEC\xB7"
-     "\xB7\x00\xEC\xB7\xB8\x00\xEC\xB7\xB9\x00\xEC\xB7\xBA\x00\xEC\xB7\xBB\x00"
-     "\xEC\xB7"
-     "\xBC\x00\xEC\xB7\xBD\x00\xEC\xB7\xBE\x00\xEC\xB7\xBF\x00\xEC\xB8\x80\x00"
-     "\xEC\xB8"
-     "\x81\x00\xEC\xB8\x82\x00\xEC\xB8\x83\x00\xEC\xB8\x84\x00\xEC\xB8\x85\x00"
-     "\xEC\xB8"
-     "\x86\x00\xEC\xB8\x87\x00\xEC\xB8\x88\x00\xEC\xB8\x89\x00\xEC\xB8\x8A\x00"
-     "\xEC\xB8"
-     "\x8B\x00\xEC\xB8\x8C\x00\xEC\xB8\x8D\x00\xEC\xB8\x8E\x00\xEC\xB8\x8F\x00"
-     "\xEC\xB8"
-     "\x90\x00\xEC\xB8\x91\x00\xEC\xB8\x92\x00\xEC\xB8\x93\x00\xEC\xB8\x94\x00"
-     "\xEC\xB8"
-     "\x95\x00\xEC\xB8\x96\x00\xEC\xB8\x97\x00\xEC\xB8\x98\x00\xEC\xB8\x99\x00"
-     "\xEC\xB8"
-     "\x9A\x00\xEC\xB8\x9B\x00\xEC\xB8\x9C\x00\xEC\xB8\x9D\x00\xEC\xB8\x9E\x00"
-     "\xEC\xB8"
-     "\x9F\x00\xEC\xB8\xA0\x00\xEC\xB8\xA1\x00\xEC\xB8\xA2\x00\xEC\xB8\xA3\x00"
-     "\xEC\xB8"
-     "\xA4\x00\xEC\xB8\xA5\x00\xEC\xB8\xA6\x00\xEC\xB8\xA7\x00\xEC\xB8\xA8\x00"
-     "\xEC\xB8"
-     "\xA9\x00\xEC\xB8\xAA\x00\xEC\xB8\xAB\x00\xEC\xB8\xAC\x00\xEC\xB8\xAD\x00"
-     "\xEC\xB8"
-     "\xAE\x00\xEC\xB8\xAF\x00\xEC\xB8\xB0\x00\xEC\xB8\xB1\x00\xEC\xB8\xB2\x00"
-     "\xEC\xB8"
-     "\xB3\x00\xEC\xB8\xB4\x00\xEC\xB8\xB5\x00\xEC\xB8\xB6\x00\xEC\xB8\xB7\x00"
-     "\xEC\xB8"
-     "\xB8\x00\xEC\xB8\xB9\x00\xEC\xB8\xBA\x00\xEC\xB8\xBB\x00\xEC\xB8\xBC\x00"
-     "\xEC\xB8"
-     "\xBD\x00\xEC\xB8\xBE\x00\xEC\xB8\xBF\x00\xEC\xB9\x80\x00\xEC\xB9\x81\x00"
-     "\xEC\xB9"
-     "\x82\x00\xEC\xB9\x83\x00\xEC\xB9\x84\x00\xEC\xB9\x85\x00\xEC\xB9\x86\x00"
-     "\xEC\xB9"
-     "\x87\x00\xEC\xB9\x88\x00\xEC\xB9\x89\x00\xEC\xB9\x8A\x00\xEC\xB9\x8B\x00"
-     "\xEC\xB9"
-     "\x8C\x00\xEC\xB9\x8D\x00\xEC\xB9\x8E\x00\xEC\xB9\x8F\x00\xEC\xB9\x90\x00"
-     "\xEC\xB9"
-     "\x91\x00\xEC\xB9\x92\x00\xEC\xB9\x93\x00\xEC\xB9\x94\x00\xEC\xB9\x95\x00"
-     "\xEC\xB9"
-     "\x96\x00\xEC\xB9\x97\x00\xEC\xB9\x98\x00\xEC\xB9\x99\x00\xEC\xB9\x9A\x00"
-     "\xEC\xB9"
-     "\x9B\x00\xEC\xB9\x9C\x00\xEC\xB9\x9D\x00\xEC\xB9\x9E\x00\xEC\xB9\x9F\x00"
-     "\xEC\xB9"
-     "\xA0\x00\xEC\xB9\xA1\x00\xEC\xB9\xA2\x00\xEC\xB9\xA3\x00\xEC\xB9\xA4\x00"
-     "\xEC\xB9"
-     "\xA5\x00\xEC\xB9\xA6\x00\xEC\xB9\xA7\x00\xEC\xB9\xA8\x00\xEC\xB9\xA9\x00"
-     "\xEC\xB9"
-     "\xAA\x00\xEC\xB9\xAB\x00\xEC\xB9\xAC\x00\xEC\xB9\xAD\x00\xEC\xB9\xAE\x00"
-     "\xEC\xB9"
-     "\xAF\x00\xEC\xB9\xB0\x00\xEC\xB9\xB1\x00\xEC\xB9\xB2\x00\xEC\xB9\xB3\x00"
-     "\xEC\xB9"
-     "\xB4\x00\xEC\xB9\xB5\x00\xEC\xB9\xB6\x00\xEC\xB9\xB7\x00\xEC\xB9\xB8\x00"
-     "\xEC\xB9"
-     "\xB9\x00\xEC\xB9\xBA\x00\xEC\xB9\xBB\x00\xEC\xB9\xBC\x00\xEC\xB9\xBD\x00"
-     "\xEC\xB9"
-     "\xBE\x00\xEC\xB9\xBF\x00\xEC\xBA\x80\x00\xEC\xBA\x81\x00\xEC\xBA\x82\x00"
-     "\xEC\xBA"
-     "\x83\x00\xEC\xBA\x84\x00\xEC\xBA\x85\x00\xEC\xBA\x86\x00\xEC\xBA\x87\x00"
-     "\xEC\xBA"
-     "\x88\x00\xEC\xBA\x89\x00\xEC\xBA\x8A\x00\xEC\xBA\x8B\x00\xEC\xBA\x8C\x00"
-     "\xEC\xBA"
-     "\x8D\x00\xEC\xBA\x8E\x00\xEC\xBA\x8F\x00\xEC\xBA\x90\x00\xEC\xBA\x91\x00"
-     "\xEC\xBA"
-     "\x92\x00\xEC\xBA\x93\x00\xEC\xBA\x94\x00\xEC\xBA\x95\x00\xEC\xBA\x96\x00"
-     "\xEC\xBA"
-     "\x97\x00\xEC\xBA\x98\x00\xEC\xBA\x99\x00\xEC\xBA\x9A\x00\xEC\xBA\x9B\x00"
-     "\xEC\xBA"
-     "\x9C\x00\xEC\xBA\x9D\x00\xEC\xBA\x9E\x00\xEC\xBA\x9F\x00\xEC\xBA\xA0\x00"
-     "\xEC\xBA"
-     "\xA1\x00\xEC\xBA\xA2\x00\xEC\xBA\xA3\x00\xEC\xBA\xA4\x00\xEC\xBA\xA5\x00"
-     "\xEC\xBA"
-     "\xA6\x00\xEC\xBA\xA7\x00\xEC\xBA\xA8\x00\xEC\xBA\xA9\x00\xEC\xBA\xAA\x00"
-     "\xEC\xBA"
-     "\xAB\x00\xEC\xBA\xAC\x00\xEC\xBA\xAD\x00\xEC\xBA\xAE\x00\xEC\xBA\xAF\x00"
-     "\xEC\xBA"
-     "\xB0\x00\xEC\xBA\xB1\x00\xEC\xBA\xB2\x00\xEC\xBA\xB3\x00\xEC\xBA\xB4\x00"
-     "\xEC\xBA"
-     "\xB5\x00\xEC\xBA\xB6\x00\xEC\xBA\xB7\x00\xEC\xBA\xB8\x00\xEC\xBA\xB9\x00"
-     "\xEC\xBA"
-     "\xBA\x00\xEC\xBA\xBB\x00\xEC\xBA\xBC\x00\xEC\xBA\xBD\x00\xEC\xBA\xBE\x00"
-     "\xEC\xBA"
-     "\xBF\x00\xEC\xBB\x80\x00\xEC\xBB\x81\x00\xEC\xBB\x82\x00\xEC\xBB\x83\x00"
-     "\xEC\xBB"
-     "\x84\x00\xEC\xBB\x85\x00\xEC\xBB\x86\x00\xEC\xBB\x87\x00\xEC\xBB\x88\x00"
-     "\xEC\xBB"
-     "\x89\x00\xEC\xBB\x8A\x00\xEC\xBB\x8B\x00\xEC\xBB\x8C\x00\xEC\xBB\x8D\x00"
-     "\xEC\xBB"
-     "\x8E\x00\xEC\xBB\x8F\x00\xEC\xBB\x90\x00\xEC\xBB\x91\x00\xEC\xBB\x92\x00"
-     "\xEC\xBB"
-     "\x93\x00\xEC\xBB\x94\x00\xEC\xBB\x95\x00\xEC\xBB\x96\x00\xEC\xBB\x97\x00"
-     "\xEC\xBB"
-     "\x98\x00\xEC\xBB\x99\x00\xEC\xBB\x9A\x00\xEC\xBB\x9B\x00\xEC\xBB\x9C\x00"
-     "\xEC\xBB"
-     "\x9D\x00\xEC\xBB\x9E\x00\xEC\xBB\x9F\x00\xEC\xBB\xA0\x00\xEC\xBB\xA1\x00"
-     "\xEC\xBB"
-     "\xA2\x00\xEC\xBB\xA3\x00\xEC\xBB\xA4\x00\xEC\xBB\xA5\x00\xEC\xBB\xA6\x00"
-     "\xEC\xBB"
-     "\xA7\x00\xEC\xBB\xA8\x00\xEC\xBB\xA9\x00\xEC\xBB\xAA\x00\xEC\xBB\xAB\x00"
-     "\xEC\xBB"
-     "\xAC\x00\xEC\xBB\xAD\x00\xEC\xBB\xAE\x00\xEC\xBB\xAF\x00\xEC\xBB\xB0\x00"
-     "\xEC\xBB"
-     "\xB1\x00\xEC\xBB\xB2\x00\xEC\xBB\xB3\x00\xEC\xBB\xB4\x00\xEC\xBB\xB5\x00"
-     "\xEC\xBB"
-     "\xB6\x00\xEC\xBB\xB7\x00\xEC\xBB\xB8\x00\xEC\xBB\xB9\x00\xEC\xBB\xBA\x00"
-     "\xEC\xBB"
-     "\xBB\x00\xEC\xBB\xBC\x00\xEC\xBB\xBD\x00\xEC\xBB\xBE\x00\xEC\xBB\xBF\x00"
-     "\xEC\xBC"
-     "\x80\x00\xEC\xBC\x81\x00\xEC\xBC\x82\x00\xEC\xBC\x83\x00\xEC\xBC\x84\x00"
-     "\xEC\xBC"
-     "\x85\x00\xEC\xBC\x86\x00\xEC\xBC\x87\x00\xEC\xBC\x88\x00\xEC\xBC\x89\x00"
-     "\xEC\xBC"
-     "\x8A\x00\xEC\xBC\x8B\x00\xEC\xBC\x8C\x00\xEC\xBC\x8D\x00\xEC\xBC\x8E\x00"
-     "\xEC\xBC"
-     "\x8F\x00\xEC\xBC\x90\x00\xEC\xBC\x91\x00\xEC\xBC\x92\x00\xEC\xBC\x93\x00"
-     "\xEC\xBC"
-     "\x94\x00\xEC\xBC\x95\x00\xEC\xBC\x96\x00\xEC\xBC\x97\x00\xEC\xBC\x98\x00"
-     "\xEC\xBC"
-     "\x99\x00\xEC\xBC\x9A\x00\xEC\xBC\x9B\x00\xEC\xBC\x9C\x00\xEC\xBC\x9D\x00"
-     "\xEC\xBC"
-     "\x9E\x00\xEC\xBC\x9F\x00\xEC\xBC\xA0\x00\xEC\xBC\xA1\x00\xEC\xBC\xA2\x00"
-     "\xEC\xBC"
-     "\xA3\x00\xEC\xBC\xA4\x00\xEC\xBC\xA5\x00\xEC\xBC\xA6\x00\xEC\xBC\xA7\x00"
-     "\xEC\xBC"
-     "\xA8\x00\xEC\xBC\xA9\x00\xEC\xBC\xAA\x00\xEC\xBC\xAB\x00\xEC\xBC\xAC\x00"
-     "\xEC\xBC"
-     "\xAD\x00\xEC\xBC\xAE\x00\xEC\xBC\xAF\x00\xEC\xBC\xB0\x00\xEC\xBC\xB1\x00"
-     "\xEC\xBC"
-     "\xB2\x00\xEC\xBC\xB3\x00\xEC\xBC\xB4\x00\xEC\xBC\xB5\x00\xEC\xBC\xB6\x00"
-     "\xEC\xBC"
-     "\xB7\x00\xEC\xBC\xB8\x00\xEC\xBC\xB9\x00\xEC\xBC\xBA\x00\xEC\xBC\xBB\x00"
-     "\xEC\xBC"
-     "\xBC\x00\xEC\xBC\xBD\x00\xEC\xBC\xBE\x00\xEC\xBC\xBF\x00\xEC\xBD\x80\x00"
-     "\xEC\xBD"
-     "\x81\x00\xEC\xBD\x82\x00\xEC\xBD\x83\x00\xEC\xBD\x84\x00\xEC\xBD\x85\x00"
-     "\xEC\xBD"
-     "\x86\x00\xEC\xBD\x87\x00\xEC\xBD\x88\x00\xEC\xBD\x89\x00\xEC\xBD\x8A\x00"
-     "\xEC\xBD"
-     "\x8B\x00\xEC\xBD\x8C\x00\xEC\xBD\x8D\x00\xEC\xBD\x8E\x00\xEC\xBD\x8F\x00"
-     "\xEC\xBD"
-     "\x90\x00\xEC\xBD\x91\x00\xEC\xBD\x92\x00\xEC\xBD\x93\x00\xEC\xBD\x94\x00"
-     "\xEC\xBD"
-     "\x95\x00\xEC\xBD\x96\x00\xEC\xBD\x97\x00\xEC\xBD\x98\x00\xEC\xBD\x99\x00"
-     "\xEC\xBD"
-     "\x9A\x00\xEC\xBD\x9B\x00\xEC\xBD\x9C\x00\xEC\xBD\x9D\x00\xEC\xBD\x9E\x00"
-     "\xEC\xBD"
-     "\x9F\x00\xEC\xBD\xA0\x00\xEC\xBD\xA1\x00\xEC\xBD\xA2\x00\xEC\xBD\xA3\x00"
-     "\xEC\xBD"
-     "\xA4\x00\xEC\xBD\xA5\x00\xEC\xBD\xA6\x00\xEC\xBD\xA7\x00\xEC\xBD\xA8\x00"
-     "\xEC\xBD"
-     "\xA9\x00\xEC\xBD\xAA\x00\xEC\xBD\xAB\x00\xEC\xBD\xAC\x00\xEC\xBD\xAD\x00"
-     "\xEC\xBD"
-     "\xAE\x00\xEC\xBD\xAF\x00\xEC\xBD\xB0\x00\xEC\xBD\xB1\x00\xEC\xBD\xB2\x00"
-     "\xEC\xBD"
-     "\xB3\x00\xEC\xBD\xB4\x00\xEC\xBD\xB5\x00\xEC\xBD\xB6\x00\xEC\xBD\xB7\x00"
-     "\xEC\xBD"
-     "\xB8\x00\xEC\xBD\xB9\x00\xEC\xBD\xBA\x00\xEC\xBD\xBB\x00\xEC\xBD\xBC\x00"
-     "\xEC\xBD"
-     "\xBD\x00\xEC\xBD\xBE\x00\xEC\xBD\xBF\x00\xEC\xBE\x80\x00\xEC\xBE\x81\x00"
-     "\xEC\xBE"
-     "\x82\x00\xEC\xBE\x83\x00\xEC\xBE\x84\x00\xEC\xBE\x85\x00\xEC\xBE\x86\x00"
-     "\xEC\xBE"
-     "\x87\x00\xEC\xBE\x88\x00\xEC\xBE\x89\x00\xEC\xBE\x8A\x00\xEC\xBE\x8B\x00"
-     "\xEC\xBE"
-     "\x8C\x00\xEC\xBE\x8D\x00\xEC\xBE\x8E\x00\xEC\xBE\x8F\x00\xEC\xBE\x90\x00"
-     "\xEC\xBE"
-     "\x91\x00\xEC\xBE\x92\x00\xEC\xBE\x93\x00\xEC\xBE\x94\x00\xEC\xBE\x95\x00"
-     "\xEC\xBE"
-     "\x96\x00\xEC\xBE\x97\x00\xEC\xBE\x98\x00\xEC\xBE\x99\x00\xEC\xBE\x9A\x00"
-     "\xEC\xBE"
-     "\x9B\x00\xEC\xBE\x9C\x00\xEC\xBE\x9D\x00\xEC\xBE\x9E\x00\xEC\xBE\x9F\x00"
-     "\xEC\xBE"
-     "\xA0\x00\xEC\xBE\xA1\x00\xEC\xBE\xA2\x00\xEC\xBE\xA3\x00\xEC\xBE\xA4\x00"
-     "\xEC\xBE"
-     "\xA5\x00\xEC\xBE\xA6\x00\xEC\xBE\xA7\x00\xEC\xBE\xA8\x00\xEC\xBE\xA9\x00"
-     "\xEC\xBE"
-     "\xAA\x00\xEC\xBE\xAB\x00\xEC\xBE\xAC\x00\xEC\xBE\xAD\x00\xEC\xBE\xAE\x00"
-     "\xEC\xBE"
-     "\xAF\x00\xEC\xBE\xB0\x00\xEC\xBE\xB1\x00\xEC\xBE\xB2\x00\xEC\xBE\xB3\x00"
-     "\xEC\xBE"
-     "\xB4\x00\xEC\xBE\xB5\x00\xEC\xBE\xB6\x00\xEC\xBE\xB7\x00\xEC\xBE\xB8\x00"
-     "\xEC\xBE"
-     "\xB9\x00\xEC\xBE\xBA\x00\xEC\xBE\xBB\x00\xEC\xBE\xBC\x00\xEC\xBE\xBD\x00"
-     "\xEC\xBE"
-     "\xBE\x00\xEC\xBE\xBF\x00\xEC\xBF\x80\x00\xEC\xBF\x81\x00\xEC\xBF\x82\x00"
-     "\xEC\xBF"
-     "\x83\x00\xEC\xBF\x84\x00\xEC\xBF\x85\x00\xEC\xBF\x86\x00\xEC\xBF\x87\x00"
-     "\xEC\xBF"
-     "\x88\x00\xEC\xBF\x89\x00\xEC\xBF\x8A\x00\xEC\xBF\x8B\x00\xEC\xBF\x8C\x00"
-     "\xEC\xBF"
-     "\x8D\x00\xEC\xBF\x8E\x00\xEC\xBF\x8F\x00\xEC\xBF\x90\x00\xEC\xBF\x91\x00"
-     "\xEC\xBF"
-     "\x92\x00\xEC\xBF\x93\x00\xEC\xBF\x94\x00\xEC\xBF\x95\x00\xEC\xBF\x96\x00"
-     "\xEC\xBF"
-     "\x97\x00\xEC\xBF\x98\x00\xEC\xBF\x99\x00\xEC\xBF\x9A\x00\xEC\xBF\x9B\x00"
-     "\xEC\xBF"
-     "\x9C\x00\xEC\xBF\x9D\x00\xEC\xBF\x9E\x00\xEC\xBF\x9F\x00\xEC\xBF\xA0\x00"
-     "\xEC\xBF"
-     "\xA1\x00\xEC\xBF\xA2\x00\xEC\xBF\xA3\x00\xEC\xBF\xA4\x00\xEC\xBF\xA5\x00"
-     "\xEC\xBF"
-     "\xA6\x00\xEC\xBF\xA7\x00\xEC\xBF\xA8\x00\xEC\xBF\xA9\x00\xEC\xBF\xAA\x00"
-     "\xEC\xBF"
-     "\xAB\x00\xEC\xBF\xAC\x00\xEC\xBF\xAD\x00\xEC\xBF\xAE\x00\xEC\xBF\xAF\x00"
-     "\xEC\xBF"
-     "\xB0\x00\xEC\xBF\xB1\x00\xEC\xBF\xB2\x00\xEC\xBF\xB3\x00\xEC\xBF\xB4\x00"
-     "\xEC\xBF"
-     "\xB5\x00\xEC\xBF\xB6\x00\xEC\xBF\xB7\x00\xEC\xBF\xB8\x00\xEC\xBF\xB9\x00"
-     "\xEC\xBF"
-     "\xBA\x00\xEC\xBF\xBB\x00\xEC\xBF\xBC\x00\xEC\xBF\xBD\x00\xEC\xBF\xBE\x00"
-     "\xEC\xBF"
-     "\xBF\x00\xED\x80\x80\x00\xED\x80\x81\x00\xED\x80\x82\x00\xED\x80\x83\x00"
-     "\xED\x80"
-     "\x84\x00\xED\x80\x85\x00\xED\x80\x86\x00\xED\x80\x87\x00\xED\x80\x88\x00"
-     "\xED\x80"
-     "\x89\x00\xED\x80\x8A\x00\xED\x80\x8B\x00\xED\x80\x8C\x00\xED\x80\x8D\x00"
-     "\xED\x80"
-     "\x8E\x00\xED\x80\x8F\x00\xED\x80\x90\x00\xED\x80\x91\x00\xED\x80\x92\x00"
-     "\xED\x80"
-     "\x93\x00\xED\x80\x94\x00\xED\x80\x95\x00\xED\x80\x96\x00\xED\x80\x97\x00"
-     "\xED\x80"
-     "\x98\x00\xED\x80\x99\x00\xED\x80\x9A\x00\xED\x80\x9B\x00\xED\x80\x9C\x00"
-     "\xED\x80"
-     "\x9D\x00\xED\x80\x9E\x00\xED\x80\x9F\x00\xED\x80\xA0\x00\xED\x80\xA1\x00"
-     "\xED\x80"
-     "\xA2\x00\xED\x80\xA3\x00\xED\x80\xA4\x00\xED\x80\xA5\x00\xED\x80\xA6\x00"
-     "\xED\x80"
-     "\xA7\x00\xED\x80\xA8\x00\xED\x80\xA9\x00\xED\x80\xAA\x00\xED\x80\xAB\x00"
-     "\xED\x80"
-     "\xAC\x00\xED\x80\xAD\x00\xED\x80\xAE\x00\xED\x80\xAF\x00\xED\x80\xB0\x00"
-     "\xED\x80"
-     "\xB1\x00\xED\x80\xB2\x00\xED\x80\xB3\x00\xED\x80\xB4\x00\xED\x80\xB5\x00"
-     "\xED\x80"
-     "\xB6\x00\xED\x80\xB7\x00\xED\x80\xB8\x00\xED\x80\xB9\x00\xED\x80\xBA\x00"
-     "\xED\x80"
-     "\xBB\x00\xED\x80\xBC\x00\xED\x80\xBD\x00\xED\x80\xBE\x00\xED\x80\xBF\x00"
-     "\xED\x81"
-     "\x80\x00\xED\x81\x81\x00\xED\x81\x82\x00\xED\x81\x83\x00\xED\x81\x84\x00"
-     "\xED\x81"
-     "\x85\x00\xED\x81\x86\x00\xED\x81\x87\x00\xED\x81\x88\x00\xED\x81\x89\x00"
-     "\xED\x81"
-     "\x8A\x00\xED\x81\x8B\x00\xED\x81\x8C\x00\xED\x81\x8D\x00\xED\x81\x8E\x00"
-     "\xED\x81"
-     "\x8F\x00\xED\x81\x90\x00\xED\x81\x91\x00\xED\x81\x92\x00\xED\x81\x93\x00"
-     "\xED\x81"
-     "\x94\x00\xED\x81\x95\x00\xED\x81\x96\x00\xED\x81\x97\x00\xED\x81\x98\x00"
-     "\xED\x81"
-     "\x99\x00\xED\x81\x9A\x00\xED\x81\x9B\x00\xED\x81\x9C\x00\xED\x81\x9D\x00"
-     "\xED\x81"
-     "\x9E\x00\xED\x81\x9F\x00\xED\x81\xA0\x00\xED\x81\xA1\x00\xED\x81\xA2\x00"
-     "\xED\x81"
-     "\xA3\x00\xED\x81\xA4\x00\xED\x81\xA5\x00\xED\x81\xA6\x00\xED\x81\xA7\x00"
-     "\xED\x81"
-     "\xA8\x00\xED\x81\xA9\x00\xED\x81\xAA\x00\xED\x81\xAB\x00\xED\x81\xAC\x00"
-     "\xED\x81"
-     "\xAD\x00\xED\x81\xAE\x00\xED\x81\xAF\x00\xED\x81\xB0\x00\xED\x81\xB1\x00"
-     "\xED\x81"
-     "\xB2\x00\xED\x81\xB3\x00\xED\x81\xB4\x00\xED\x81\xB5\x00\xED\x81\xB6\x00"
-     "\xED\x81"
-     "\xB7\x00\xED\x81\xB8\x00\xED\x81\xB9\x00\xED\x81\xBA\x00\xED\x81\xBB\x00"
-     "\xED\x81"
-     "\xBC\x00\xED\x81\xBD\x00\xED\x81\xBE\x00\xED\x81\xBF\x00\xED\x82\x80\x00"
-     "\xED\x82"
-     "\x81\x00\xED\x82\x82\x00\xED\x82\x83\x00\xED\x82\x84\x00\xED\x82\x85\x00"
-     "\xED\x82"
-     "\x86\x00\xED\x82\x87\x00\xED\x82\x88\x00\xED\x82\x89\x00\xED\x82\x8A\x00"
-     "\xED\x82"
-     "\x8B\x00\xED\x82\x8C\x00\xED\x82\x8D\x00\xED\x82\x8E\x00\xED\x82\x8F\x00"
-     "\xED\x82"
-     "\x90\x00\xED\x82\x91\x00\xED\x82\x92\x00\xED\x82\x93\x00\xED\x82\x94\x00"
-     "\xED\x82"
-     "\x95\x00\xED\x82\x96\x00\xED\x82\x97\x00\xED\x82\x98\x00\xED\x82\x99\x00"
-     "\xED\x82"
-     "\x9A\x00\xED\x82\x9B\x00\xED\x82\x9C\x00\xED\x82\x9D\x00\xED\x82\x9E\x00"
-     "\xED\x82"
-     "\x9F\x00\xED\x82\xA0\x00\xED\x82\xA1\x00\xED\x82\xA2\x00\xED\x82\xA3\x00"
-     "\xED\x82"
-     "\xA4\x00\xED\x82\xA5\x00\xED\x82\xA6\x00\xED\x82\xA7\x00\xED\x82\xA8\x00"
-     "\xED\x82"
-     "\xA9\x00\xED\x82\xAA\x00\xED\x82\xAB\x00\xED\x82\xAC\x00\xED\x82\xAD\x00"
-     "\xED\x82"
-     "\xAE\x00\xED\x82\xAF\x00\xED\x82\xB0\x00\xED\x82\xB1\x00\xED\x82\xB2\x00"
-     "\xED\x82"
-     "\xB3\x00\xED\x82\xB4\x00\xED\x82\xB5\x00\xED\x82\xB6\x00\xED\x82\xB7\x00"
-     "\xED\x82"
-     "\xB8\x00\xED\x82\xB9\x00\xED\x82\xBA\x00\xED\x82\xBB\x00\xED\x82\xBC\x00"
-     "\xED\x82"
-     "\xBD\x00\xED\x82\xBE\x00\xED\x82\xBF\x00\xED\x83\x80\x00\xED\x83\x81\x00"
-     "\xED\x83"
-     "\x82\x00\xED\x83\x83\x00\xED\x83\x84\x00\xED\x83\x85\x00\xED\x83\x86\x00"
-     "\xED\x83"
-     "\x87\x00\xED\x83\x88\x00\xED\x83\x89\x00\xED\x83\x8A\x00\xED\x83\x8B\x00"
-     "\xED\x83"
-     "\x8C\x00\xED\x83\x8D\x00\xED\x83\x8E\x00\xED\x83\x8F\x00\xED\x83\x90\x00"
-     "\xED\x83"
-     "\x91\x00\xED\x83\x92\x00\xED\x83\x93\x00\xED\x83\x94\x00\xED\x83\x95\x00"
-     "\xED\x83"
-     "\x96\x00\xED\x83\x97\x00\xED\x83\x98\x00\xED\x83\x99\x00\xED\x83\x9A\x00"
-     "\xED\x83"
-     "\x9B\x00\xED\x83\x9C\x00\xED\x83\x9D\x00\xED\x83\x9E\x00\xED\x83\x9F\x00"
-     "\xED\x83"
-     "\xA0\x00\xED\x83\xA1\x00\xED\x83\xA2\x00\xED\x83\xA3\x00\xED\x83\xA4\x00"
-     "\xED\x83"
-     "\xA5\x00\xED\x83\xA6\x00\xED\x83\xA7\x00\xED\x83\xA8\x00\xED\x83\xA9\x00"
-     "\xED\x83"
-     "\xAA\x00\xED\x83\xAB\x00\xED\x83\xAC\x00\xED\x83\xAD\x00\xED\x83\xAE\x00"
-     "\xED\x83"
-     "\xAF\x00\xED\x83\xB0\x00\xED\x83\xB1\x00\xED\x83\xB2\x00\xED\x83\xB3\x00"
-     "\xED\x83"
-     "\xB4\x00\xED\x83\xB5\x00\xED\x83\xB6\x00\xED\x83\xB7\x00\xED\x83\xB8\x00"
-     "\xED\x83"
-     "\xB9\x00\xED\x83\xBA\x00\xED\x83\xBB\x00\xED\x83\xBC\x00\xED\x83\xBD\x00"
-     "\xED\x83"
-     "\xBE\x00\xED\x83\xBF\x00\xED\x84\x80\x00\xED\x84\x81\x00\xED\x84\x82\x00"
-     "\xED\x84"
-     "\x83\x00\xED\x84\x84\x00\xED\x84\x85\x00\xED\x84\x86\x00\xED\x84\x87\x00"
-     "\xED\x84"
-     "\x88\x00\xED\x84\x89\x00\xED\x84\x8A\x00\xED\x84\x8B\x00\xED\x84\x8C\x00"
-     "\xED\x84"
-     "\x8D\x00\xED\x84\x8E\x00\xED\x84\x8F\x00\xED\x84\x90\x00\xED\x84\x91\x00"
-     "\xED\x84"
-     "\x92\x00\xED\x84\x93\x00\xED\x84\x94\x00\xED\x84\x95\x00\xED\x84\x96\x00"
-     "\xED\x84"
-     "\x97\x00\xED\x84\x98\x00\xED\x84\x99\x00\xED\x84\x9A\x00\xED\x84\x9B\x00"
-     "\xED\x84"
-     "\x9C\x00\xED\x84\x9D\x00\xED\x84\x9E\x00\xED\x84\x9F\x00\xED\x84\xA0\x00"
-     "\xED\x84"
-     "\xA1\x00\xED\x84\xA2\x00\xED\x84\xA3\x00\xED\x84\xA4\x00\xED\x84\xA5\x00"
-     "\xED\x84"
-     "\xA6\x00\xED\x84\xA7\x00\xED\x84\xA8\x00\xED\x84\xA9\x00\xED\x84\xAA\x00"
-     "\xED\x84"
-     "\xAB\x00\xED\x84\xAC\x00\xED\x84\xAD\x00\xED\x84\xAE\x00\xED\x84\xAF\x00"
-     "\xED\x84"
-     "\xB0\x00\xED\x84\xB1\x00\xED\x84\xB2\x00\xED\x84\xB3\x00\xED\x84\xB4\x00"
-     "\xED\x84"
-     "\xB5\x00\xED\x84\xB6\x00\xED\x84\xB7\x00\xED\x84\xB8\x00\xED\x84\xB9\x00"
-     "\xED\x84"
-     "\xBA\x00\xED\x84\xBB\x00\xED\x84\xBC\x00\xED\x84\xBD\x00\xED\x84\xBE\x00"
-     "\xED\x84"
-     "\xBF\x00\xED\x85\x80\x00\xED\x85\x81\x00\xED\x85\x82\x00\xED\x85\x83\x00"
-     "\xED\x85"
-     "\x84\x00\xED\x85\x85\x00\xED\x85\x86\x00\xED\x85\x87\x00\xED\x85\x88\x00"
-     "\xED\x85"
-     "\x89\x00\xED\x85\x8A\x00\xED\x85\x8B\x00\xED\x85\x8C\x00\xED\x85\x8D\x00"
-     "\xED\x85"
-     "\x8E\x00\xED\x85\x8F\x00\xED\x85\x90\x00\xED\x85\x91\x00\xED\x85\x92\x00"
-     "\xED\x85"
-     "\x93\x00\xED\x85\x94\x00\xED\x85\x95\x00\xED\x85\x96\x00\xED\x85\x97\x00"
-     "\xED\x85"
-     "\x98\x00\xED\x85\x99\x00\xED\x85\x9A\x00\xED\x85\x9B\x00\xED\x85\x9C\x00"
-     "\xED\x85"
-     "\x9D\x00\xED\x85\x9E\x00\xED\x85\x9F\x00\xED\x85\xA0\x00\xED\x85\xA1\x00"
-     "\xED\x85"
-     "\xA2\x00\xED\x85\xA3\x00\xED\x85\xA4\x00\xED\x85\xA5\x00\xED\x85\xA6\x00"
-     "\xED\x85"
-     "\xA7\x00\xED\x85\xA8\x00\xED\x85\xA9\x00\xED\x85\xAA\x00\xED\x85\xAB\x00"
-     "\xED\x85"
-     "\xAC\x00\xED\x85\xAD\x00\xED\x85\xAE\x00\xED\x85\xAF\x00\xED\x85\xB0\x00"
-     "\xED\x85"
-     "\xB1\x00\xED\x85\xB2\x00\xED\x85\xB3\x00\xED\x85\xB4\x00\xED\x85\xB5\x00"
-     "\xED\x85"
-     "\xB6\x00\xED\x85\xB7\x00\xED\x85\xB8\x00\xED\x85\xB9\x00\xED\x85\xBA\x00"
-     "\xED\x85"
-     "\xBB\x00\xED\x85\xBC\x00\xED\x85\xBD\x00\xED\x85\xBE\x00\xED\x85\xBF\x00"
-     "\xED\x86"
-     "\x80\x00\xED\x86\x81\x00\xED\x86\x82\x00\xED\x86\x83\x00\xED\x86\x84\x00"
-     "\xED\x86"
-     "\x85\x00\xED\x86\x86\x00\xED\x86\x87\x00\xED\x86\x88\x00\xED\x86\x89\x00"
-     "\xED\x86"
-     "\x8A\x00\xED\x86\x8B\x00\xED\x86\x8C\x00\xED\x86\x8D\x00\xED\x86\x8E\x00"
-     "\xED\x86"
-     "\x8F\x00\xED\x86\x90\x00\xED\x86\x91\x00\xED\x86\x92\x00\xED\x86\x93\x00"
-     "\xED\x86"
-     "\x94\x00\xED\x86\x95\x00\xED\x86\x96\x00\xED\x86\x97\x00\xED\x86\x98\x00"
-     "\xED\x86"
-     "\x99\x00\xED\x86\x9A\x00\xED\x86\x9B\x00\xED\x86\x9C\x00\xED\x86\x9D\x00"
-     "\xED\x86"
-     "\x9E\x00\xED\x86\x9F\x00\xED\x86\xA0\x00\xED\x86\xA1\x00\xED\x86\xA2\x00"
-     "\xED\x86"
-     "\xA3\x00\xED\x86\xA4\x00\xED\x86\xA5\x00\xED\x86\xA6\x00\xED\x86\xA7\x00"
-     "\xED\x86"
-     "\xA8\x00\xED\x86\xA9\x00\xED\x86\xAA\x00\xED\x86\xAB\x00\xED\x86\xAC\x00"
-     "\xED\x86"
-     "\xAD\x00\xED\x86\xAE\x00\xED\x86\xAF\x00\xED\x86\xB0\x00\xED\x86\xB1\x00"
-     "\xED\x86"
-     "\xB2\x00\xED\x86\xB3\x00\xED\x86\xB4\x00\xED\x86\xB5\x00\xED\x86\xB6\x00"
-     "\xED\x86"
-     "\xB7\x00\xED\x86\xB8\x00\xED\x86\xB9\x00\xED\x86\xBA\x00\xED\x86\xBB\x00"
-     "\xED\x86"
-     "\xBC\x00\xED\x86\xBD\x00\xED\x86\xBE\x00\xED\x86\xBF\x00\xED\x87\x80\x00"
-     "\xED\x87"
-     "\x81\x00\xED\x87\x82\x00\xED\x87\x83\x00\xED\x87\x84\x00\xED\x87\x85\x00"
-     "\xED\x87"
-     "\x86\x00\xED\x87\x87\x00\xED\x87\x88\x00\xED\x87\x89\x00\xED\x87\x8A\x00"
-     "\xED\x87"
-     "\x8B\x00\xED\x87\x8C\x00\xED\x87\x8D\x00\xED\x87\x8E\x00\xED\x87\x8F\x00"
-     "\xED\x87"
-     "\x90\x00\xED\x87\x91\x00\xED\x87\x92\x00\xED\x87\x93\x00\xED\x87\x94\x00"
-     "\xED\x87"
-     "\x95\x00\xED\x87\x96\x00\xED\x87\x97\x00\xED\x87\x98\x00\xED\x87\x99\x00"
-     "\xED\x87"
-     "\x9A\x00\xED\x87\x9B\x00\xED\x87\x9C\x00\xED\x87\x9D\x00\xED\x87\x9E\x00"
-     "\xED\x87"
-     "\x9F\x00\xED\x87\xA0\x00\xED\x87\xA1\x00\xED\x87\xA2\x00\xED\x87\xA3\x00"
-     "\xED\x87"
-     "\xA4\x00\xED\x87\xA5\x00\xED\x87\xA6\x00\xED\x87\xA7\x00\xED\x87\xA8\x00"
-     "\xED\x87"
-     "\xA9\x00\xED\x87\xAA\x00\xED\x87\xAB\x00\xED\x87\xAC\x00\xED\x87\xAD\x00"
-     "\xED\x87"
-     "\xAE\x00\xED\x87\xAF\x00\xED\x87\xB0\x00\xED\x87\xB1\x00\xED\x87\xB2\x00"
-     "\xED\x87"
-     "\xB3\x00\xED\x87\xB4\x00\xED\x87\xB5\x00\xED\x87\xB6\x00\xED\x87\xB7\x00"
-     "\xED\x87"
-     "\xB8\x00\xED\x87\xB9\x00\xED\x87\xBA\x00\xED\x87\xBB\x00\xED\x87\xBC\x00"
-     "\xED\x87"
-     "\xBD\x00\xED\x87\xBE\x00\xED\x87\xBF\x00\xED\x88\x80\x00\xED\x88\x81\x00"
-     "\xED\x88"
-     "\x82\x00\xED\x88\x83\x00\xED\x88\x84\x00\xED\x88\x85\x00\xED\x88\x86\x00"
-     "\xED\x88"
-     "\x87\x00\xED\x88\x88\x00\xED\x88\x89\x00\xED\x88\x8A\x00\xED\x88\x8B\x00"
-     "\xED\x88"
-     "\x8C\x00\xED\x88\x8D\x00\xED\x88\x8E\x00\xED\x88\x8F\x00\xED\x88\x90\x00"
-     "\xED\x88"
-     "\x91\x00\xED\x88\x92\x00\xED\x88\x93\x00\xED\x88\x94\x00\xED\x88\x95\x00"
-     "\xED\x88"
-     "\x96\x00\xED\x88\x97\x00\xED\x88\x98\x00\xED\x88\x99\x00\xED\x88\x9A\x00"
-     "\xED\x88"
-     "\x9B\x00\xED\x88\x9C\x00\xED\x88\x9D\x00\xED\x88\x9E\x00\xED\x88\x9F\x00"
-     "\xED\x88"
-     "\xA0\x00\xED\x88\xA1\x00\xED\x88\xA2\x00\xED\x88\xA3\x00\xED\x88\xA4\x00"
-     "\xED\x88"
-     "\xA5\x00\xED\x88\xA6\x00\xED\x88\xA7\x00\xED\x88\xA8\x00\xED\x88\xA9\x00"
-     "\xED\x88"
-     "\xAA\x00\xED\x88\xAB\x00\xED\x88\xAC\x00\xED\x88\xAD\x00\xED\x88\xAE\x00"
-     "\xED\x88"
-     "\xAF\x00\xED\x88\xB0\x00\xED\x88\xB1\x00\xED\x88\xB2\x00\xED\x88\xB3\x00"
-     "\xED\x88"
-     "\xB4\x00\xED\x88\xB5\x00\xED\x88\xB6\x00\xED\x88\xB7\x00\xED\x88\xB8\x00"
-     "\xED\x88"
-     "\xB9\x00\xED\x88\xBA\x00\xED\x88\xBB\x00\xED\x88\xBC\x00\xED\x88\xBD\x00"
-     "\xED\x88"
-     "\xBE\x00\xED\x88\xBF\x00\xED\x89\x80\x00\xED\x89\x81\x00\xED\x89\x82\x00"
-     "\xED\x89"
-     "\x83\x00\xED\x89\x84\x00\xED\x89\x85\x00\xED\x89\x86\x00\xED\x89\x87\x00"
-     "\xED\x89"
-     "\x88\x00\xED\x89\x89\x00\xED\x89\x8A\x00\xED\x89\x8B\x00\xED\x89\x8C\x00"
-     "\xED\x89"
-     "\x8D\x00\xED\x89\x8E\x00\xED\x89\x8F\x00\xED\x89\x90\x00\xED\x89\x91\x00"
-     "\xED\x89"
-     "\x92\x00\xED\x89\x93\x00\xED\x89\x94\x00\xED\x89\x95\x00\xED\x89\x96\x00"
-     "\xED\x89"
-     "\x97\x00\xED\x89\x98\x00\xED\x89\x99\x00\xED\x89\x9A\x00\xED\x89\x9B\x00"
-     "\xED\x89"
-     "\x9C\x00\xED\x89\x9D\x00\xED\x89\x9E\x00\xED\x89\x9F\x00\xED\x89\xA0\x00"
-     "\xED\x89"
-     "\xA1\x00\xED\x89\xA2\x00\xED\x89\xA3\x00\xED\x89\xA4\x00\xED\x89\xA5\x00"
-     "\xED\x89"
-     "\xA6\x00\xED\x89\xA7\x00\xED\x89\xA8\x00\xED\x89\xA9\x00\xED\x89\xAA\x00"
-     "\xED\x89"
-     "\xAB\x00\xED\x89\xAC\x00\xED\x89\xAD\x00\xED\x89\xAE\x00\xED\x89\xAF\x00"
-     "\xED\x89"
-     "\xB0\x00\xED\x89\xB1\x00\xED\x89\xB2\x00\xED\x89\xB3\x00\xED\x89\xB4\x00"
-     "\xED\x89"
-     "\xB5\x00\xED\x89\xB6\x00\xED\x89\xB7\x00\xED\x89\xB8\x00\xED\x89\xB9\x00"
-     "\xED\x89"
-     "\xBA\x00\xED\x89\xBB\x00\xED\x89\xBC\x00\xED\x89\xBD\x00\xED\x89\xBE\x00"
-     "\xED\x89"
-     "\xBF\x00\xED\x8A\x80\x00\xED\x8A\x81\x00\xED\x8A\x82\x00\xED\x8A\x83\x00"
-     "\xED\x8A"
-     "\x84\x00\xED\x8A\x85\x00\xED\x8A\x86\x00\xED\x8A\x87\x00\xED\x8A\x88\x00"
-     "\xED\x8A"
-     "\x89\x00\xED\x8A\x8A\x00\xED\x8A\x8B\x00\xED\x8A\x8C\x00\xED\x8A\x8D\x00"
-     "\xED\x8A"
-     "\x8E\x00\xED\x8A\x8F\x00\xED\x8A\x90\x00\xED\x8A\x91\x00\xED\x8A\x92\x00"
-     "\xED\x8A"
-     "\x93\x00\xED\x8A\x94\x00\xED\x8A\x95\x00\xED\x8A\x96\x00\xED\x8A\x97\x00"
-     "\xED\x8A"
-     "\x98\x00\xED\x8A\x99\x00\xED\x8A\x9A\x00\xED\x8A\x9B\x00\xED\x8A\x9C\x00"
-     "\xED\x8A"
-     "\x9D\x00\xED\x8A\x9E\x00\xED\x8A\x9F\x00\xED\x8A\xA0\x00\xED\x8A\xA1\x00"
-     "\xED\x8A"
-     "\xA2\x00\xED\x8A\xA3\x00\xED\x8A\xA4\x00\xED\x8A\xA5\x00\xED\x8A\xA6\x00"
-     "\xED\x8A"
-     "\xA7\x00\xED\x8A\xA8\x00\xED\x8A\xA9\x00\xED\x8A\xAA\x00\xED\x8A\xAB\x00"
-     "\xED\x8A"
-     "\xAC\x00\xED\x8A\xAD\x00\xED\x8A\xAE\x00\xED\x8A\xAF\x00\xED\x8A\xB0\x00"
-     "\xED\x8A"
-     "\xB1\x00\xED\x8A\xB2\x00\xED\x8A\xB3\x00\xED\x8A\xB4\x00\xED\x8A\xB5\x00"
-     "\xED\x8A"
-     "\xB6\x00\xED\x8A\xB7\x00\xED\x8A\xB8\x00\xED\x8A\xB9\x00\xED\x8A\xBA\x00"
-     "\xED\x8A"
-     "\xBB\x00\xED\x8A\xBC\x00\xED\x8A\xBD\x00\xED\x8A\xBE\x00\xED\x8A\xBF\x00"
-     "\xED\x8B"
-     "\x80\x00\xED\x8B\x81\x00\xED\x8B\x82\x00\xED\x8B\x83\x00\xED\x8B\x84\x00"
-     "\xED\x8B"
-     "\x85\x00\xED\x8B\x86\x00\xED\x8B\x87\x00\xED\x8B\x88\x00\xED\x8B\x89\x00"
-     "\xED\x8B"
-     "\x8A\x00\xED\x8B\x8B\x00\xED\x8B\x8C\x00\xED\x8B\x8D\x00\xED\x8B\x8E\x00"
-     "\xED\x8B"
-     "\x8F\x00\xED\x8B\x90\x00\xED\x8B\x91\x00\xED\x8B\x92\x00\xED\x8B\x93\x00"
-     "\xED\x8B"
-     "\x94\x00\xED\x8B\x95\x00\xED\x8B\x96\x00\xED\x8B\x97\x00\xED\x8B\x98\x00"
-     "\xED\x8B"
-     "\x99\x00\xED\x8B\x9A\x00\xED\x8B\x9B\x00\xED\x8B\x9C\x00\xED\x8B\x9D\x00"
-     "\xED\x8B"
-     "\x9E\x00\xED\x8B\x9F\x00\xED\x8B\xA0\x00\xED\x8B\xA1\x00\xED\x8B\xA2\x00"
-     "\xED\x8B"
-     "\xA3\x00\xED\x8B\xA4\x00\xED\x8B\xA5\x00\xED\x8B\xA6\x00\xED\x8B\xA7\x00"
-     "\xED\x8B"
-     "\xA8\x00\xED\x8B\xA9\x00\xED\x8B\xAA\x00\xED\x8B\xAB\x00\xED\x8B\xAC\x00"
-     "\xED\x8B"
-     "\xAD\x00\xED\x8B\xAE\x00\xED\x8B\xAF\x00\xED\x8B\xB0\x00\xED\x8B\xB1\x00"
-     "\xED\x8B"
-     "\xB2\x00\xED\x8B\xB3\x00\xED\x8B\xB4\x00\xED\x8B\xB5\x00\xED\x8B\xB6\x00"
-     "\xED\x8B"
-     "\xB7\x00\xED\x8B\xB8\x00\xED\x8B\xB9\x00\xED\x8B\xBA\x00\xED\x8B\xBB\x00"
-     "\xED\x8B"
-     "\xBC\x00\xED\x8B\xBD\x00\xED\x8B\xBE\x00\xED\x8B\xBF\x00\xED\x8C\x80\x00"
-     "\xED\x8C"
-     "\x81\x00\xED\x8C\x82\x00\xED\x8C\x83\x00\xED\x8C\x84\x00\xED\x8C\x85\x00"
-     "\xED\x8C"
-     "\x86\x00\xED\x8C\x87\x00\xED\x8C\x88\x00\xED\x8C\x89\x00\xED\x8C\x8A\x00"
-     "\xED\x8C"
-     "\x8B\x00\xED\x8C\x8C\x00\xED\x8C\x8D\x00\xED\x8C\x8E\x00\xED\x8C\x8F\x00"
-     "\xED\x8C"
-     "\x90\x00\xED\x8C\x91\x00\xED\x8C\x92\x00\xED\x8C\x93\x00\xED\x8C\x94\x00"
-     "\xED\x8C"
-     "\x95\x00\xED\x8C\x96\x00\xED\x8C\x97\x00\xED\x8C\x98\x00\xED\x8C\x99\x00"
-     "\xED\x8C"
-     "\x9A\x00\xED\x8C\x9B\x00\xED\x8C\x9C\x00\xED\x8C\x9D\x00\xED\x8C\x9E\x00"
-     "\xED\x8C"
-     "\x9F\x00\xED\x8C\xA0\x00\xED\x8C\xA1\x00\xED\x8C\xA2\x00\xED\x8C\xA3\x00"
-     "\xED\x8C"
-     "\xA4\x00\xED\x8C\xA5\x00\xED\x8C\xA6\x00\xED\x8C\xA7\x00\xED\x8C\xA8\x00"
-     "\xED\x8C"
-     "\xA9\x00\xED\x8C\xAA\x00\xED\x8C\xAB\x00\xED\x8C\xAC\x00\xED\x8C\xAD\x00"
-     "\xED\x8C"
-     "\xAE\x00\xED\x8C\xAF\x00\xED\x8C\xB0\x00\xED\x8C\xB1\x00\xED\x8C\xB2\x00"
-     "\xED\x8C"
-     "\xB3\x00\xED\x8C\xB4\x00\xED\x8C\xB5\x00\xED\x8C\xB6\x00\xED\x8C\xB7\x00"
-     "\xED\x8C"
-     "\xB8\x00\xED\x8C\xB9\x00\xED\x8C\xBA\x00\xED\x8C\xBB\x00\xED\x8C\xBC\x00"
-     "\xED\x8C"
-     "\xBD\x00\xED\x8C\xBE\x00\xED\x8C\xBF\x00\xED\x8D\x80\x00\xED\x8D\x81\x00"
-     "\xED\x8D"
-     "\x82\x00\xED\x8D\x83\x00\xED\x8D\x84\x00\xED\x8D\x85\x00\xED\x8D\x86\x00"
-     "\xED\x8D"
-     "\x87\x00\xED\x8D\x88\x00\xED\x8D\x89\x00\xED\x8D\x8A\x00\xED\x8D\x8B\x00"
-     "\xED\x8D"
-     "\x8C\x00\xED\x8D\x8D\x00\xED\x8D\x8E\x00\xED\x8D\x8F\x00\xED\x8D\x90\x00"
-     "\xED\x8D"
-     "\x91\x00\xED\x8D\x92\x00\xED\x8D\x93\x00\xED\x8D\x94\x00\xED\x8D\x95\x00"
-     "\xED\x8D"
-     "\x96\x00\xED\x8D\x97\x00\xED\x8D\x98\x00\xED\x8D\x99\x00\xED\x8D\x9A\x00"
-     "\xED\x8D"
-     "\x9B\x00\xED\x8D\x9C\x00\xED\x8D\x9D\x00\xED\x8D\x9E\x00\xED\x8D\x9F\x00"
-     "\xED\x8D"
-     "\xA0\x00\xED\x8D\xA1\x00\xED\x8D\xA2\x00\xED\x8D\xA3\x00\xED\x8D\xA4\x00"
-     "\xED\x8D"
-     "\xA5\x00\xED\x8D\xA6\x00\xED\x8D\xA7\x00\xED\x8D\xA8\x00\xED\x8D\xA9\x00"
-     "\xED\x8D"
-     "\xAA\x00\xED\x8D\xAB\x00\xED\x8D\xAC\x00\xED\x8D\xAD\x00\xED\x8D\xAE\x00"
-     "\xED\x8D"
-     "\xAF\x00\xED\x8D\xB0\x00\xED\x8D\xB1\x00\xED\x8D\xB2\x00\xED\x8D\xB3\x00"
-     "\xED\x8D"
-     "\xB4\x00\xED\x8D\xB5\x00\xED\x8D\xB6\x00\xED\x8D\xB7\x00\xED\x8D\xB8\x00"
-     "\xED\x8D"
-     "\xB9\x00\xED\x8D\xBA\x00\xED\x8D\xBB\x00\xED\x8D\xBC\x00\xED\x8D\xBD\x00"
-     "\xED\x8D"
-     "\xBE\x00\xED\x8D\xBF\x00\xED\x8E\x80\x00\xED\x8E\x81\x00\xED\x8E\x82\x00"
-     "\xED\x8E"
-     "\x83\x00\xED\x8E\x84\x00\xED\x8E\x85\x00\xED\x8E\x86\x00\xED\x8E\x87\x00"
-     "\xED\x8E"
-     "\x88\x00\xED\x8E\x89\x00\xED\x8E\x8A\x00\xED\x8E\x8B\x00\xED\x8E\x8C\x00"
-     "\xED\x8E"
-     "\x8D\x00\xED\x8E\x8E\x00\xED\x8E\x8F\x00\xED\x8E\x90\x00\xED\x8E\x91\x00"
-     "\xED\x8E"
-     "\x92\x00\xED\x8E\x93\x00\xED\x8E\x94\x00\xED\x8E\x95\x00\xED\x8E\x96\x00"
-     "\xED\x8E"
-     "\x97\x00\xED\x8E\x98\x00\xED\x8E\x99\x00\xED\x8E\x9A\x00\xED\x8E\x9B\x00"
-     "\xED\x8E"
-     "\x9C\x00\xED\x8E\x9D\x00\xED\x8E\x9E\x00\xED\x8E\x9F\x00\xED\x8E\xA0\x00"
-     "\xED\x8E"
-     "\xA1\x00\xED\x8E\xA2\x00\xED\x8E\xA3\x00\xED\x8E\xA4\x00\xED\x8E\xA5\x00"
-     "\xED\x8E"
-     "\xA6\x00\xED\x8E\xA7\x00\xED\x8E\xA8\x00\xED\x8E\xA9\x00\xED\x8E\xAA\x00"
-     "\xED\x8E"
-     "\xAB\x00\xED\x8E\xAC\x00\xED\x8E\xAD\x00\xED\x8E\xAE\x00\xED\x8E\xAF\x00"
-     "\xED\x8E"
-     "\xB0\x00\xED\x8E\xB1\x00\xED\x8E\xB2\x00\xED\x8E\xB3\x00\xED\x8E\xB4\x00"
-     "\xED\x8E"
-     "\xB5\x00\xED\x8E\xB6\x00\xED\x8E\xB7\x00\xED\x8E\xB8\x00\xED\x8E\xB9\x00"
-     "\xED\x8E"
-     "\xBA\x00\xED\x8E\xBB\x00\xED\x8E\xBC\x00\xED\x8E\xBD\x00\xED\x8E\xBE\x00"
-     "\xED\x8E"
-     "\xBF\x00\xED\x8F\x80\x00\xED\x8F\x81\x00\xED\x8F\x82\x00\xED\x8F\x83\x00"
-     "\xED\x8F"
-     "\x84\x00\xED\x8F\x85\x00\xED\x8F\x86\x00\xED\x8F\x87\x00\xED\x8F\x88\x00"
-     "\xED\x8F"
-     "\x89\x00\xED\x8F\x8A\x00\xED\x8F\x8B\x00\xED\x8F\x8C\x00\xED\x8F\x8D\x00"
-     "\xED\x8F"
-     "\x8E\x00\xED\x8F\x8F\x00\xED\x8F\x90\x00\xED\x8F\x91\x00\xED\x8F\x92\x00"
-     "\xED\x8F"
-     "\x93\x00\xED\x8F\x94\x00\xED\x8F\x95\x00\xED\x8F\x96\x00\xED\x8F\x97\x00"
-     "\xED\x8F"
-     "\x98\x00\xED\x8F\x99\x00\xED\x8F\x9A\x00\xED\x8F\x9B\x00\xED\x8F\x9C\x00"
-     "\xED\x8F"
-     "\x9D\x00\xED\x8F\x9E\x00\xED\x8F\x9F\x00\xED\x8F\xA0\x00\xED\x8F\xA1\x00"
-     "\xED\x8F"
-     "\xA2\x00\xED\x8F\xA3\x00\xED\x8F\xA4\x00\xED\x8F\xA5\x00\xED\x8F\xA6\x00"
-     "\xED\x8F"
-     "\xA7\x00\xED\x8F\xA8\x00\xED\x8F\xA9\x00\xED\x8F\xAA\x00\xED\x8F\xAB\x00"
-     "\xED\x8F"
-     "\xAC\x00\xED\x8F\xAD\x00\xED\x8F\xAE\x00\xED\x8F\xAF\x00\xED\x8F\xB0\x00"
-     "\xED\x8F"
-     "\xB1\x00\xED\x8F\xB2\x00\xED\x8F\xB3\x00\xED\x8F\xB4\x00\xED\x8F\xB5\x00"
-     "\xED\x8F"
-     "\xB6\x00\xED\x8F\xB7\x00\xED\x8F\xB8\x00\xED\x8F\xB9\x00\xED\x8F\xBA\x00"
-     "\xED\x8F"
-     "\xBB\x00\xED\x8F\xBC\x00\xED\x8F\xBD\x00\xED\x8F\xBE\x00\xED\x8F\xBF\x00"
-     "\xED\x90"
-     "\x80\x00\xED\x90\x81\x00\xED\x90\x82\x00\xED\x90\x83\x00\xED\x90\x84\x00"
-     "\xED\x90"
-     "\x85\x00\xED\x90\x86\x00\xED\x90\x87\x00\xED\x90\x88\x00\xED\x90\x89\x00"
-     "\xED\x90"
-     "\x8A\x00\xED\x90\x8B\x00\xED\x90\x8C\x00\xED\x90\x8D\x00\xED\x90\x8E\x00"
-     "\xED\x90"
-     "\x8F\x00\xED\x90\x90\x00\xED\x90\x91\x00\xED\x90\x92\x00\xED\x90\x93\x00"
-     "\xED\x90"
-     "\x94\x00\xED\x90\x95\x00\xED\x90\x96\x00\xED\x90\x97\x00\xED\x90\x98\x00"
-     "\xED\x90"
-     "\x99\x00\xED\x90\x9A\x00\xED\x90\x9B\x00\xED\x90\x9C\x00\xED\x90\x9D\x00"
-     "\xED\x90"
-     "\x9E\x00\xED\x90\x9F\x00\xED\x90\xA0\x00\xED\x90\xA1\x00\xED\x90\xA2\x00"
-     "\xED\x90"
-     "\xA3\x00\xED\x90\xA4\x00\xED\x90\xA5\x00\xED\x90\xA6\x00\xED\x90\xA7\x00"
-     "\xED\x90"
-     "\xA8\x00\xED\x90\xA9\x00\xED\x90\xAA\x00\xED\x90\xAB\x00\xED\x90\xAC\x00"
-     "\xED\x90"
-     "\xAD\x00\xED\x90\xAE\x00\xED\x90\xAF\x00\xED\x90\xB0\x00\xED\x90\xB1\x00"
-     "\xED\x90"
-     "\xB2\x00\xED\x90\xB3\x00\xED\x90\xB4\x00\xED\x90\xB5\x00\xED\x90\xB6\x00"
-     "\xED\x90"
-     "\xB7\x00\xED\x90\xB8\x00\xED\x90\xB9\x00\xED\x90\xBA\x00\xED\x90\xBB\x00"
-     "\xED\x90"
-     "\xBC\x00\xED\x90\xBD\x00\xED\x90\xBE\x00\xED\x90\xBF\x00\xED\x91\x80\x00"
-     "\xED\x91"
-     "\x81\x00\xED\x91\x82\x00\xED\x91\x83\x00\xED\x91\x84\x00\xED\x91\x85\x00"
-     "\xED\x91"
-     "\x86\x00\xED\x91\x87\x00\xED\x91\x88\x00\xED\x91\x89\x00\xED\x91\x8A\x00"
-     "\xED\x91"
-     "\x8B\x00\xED\x91\x8C\x00\xED\x91\x8D\x00\xED\x91\x8E\x00\xED\x91\x8F\x00"
-     "\xED\x91"
-     "\x90\x00\xED\x91\x91\x00\xED\x91\x92\x00\xED\x91\x93\x00\xED\x91\x94\x00"
-     "\xED\x91"
-     "\x95\x00\xED\x91\x96\x00\xED\x91\x97\x00\xED\x91\x98\x00\xED\x91\x99\x00"
-     "\xED\x91"
-     "\x9A\x00\xED\x91\x9B\x00\xED\x91\x9C\x00\xED\x91\x9D\x00\xED\x91\x9E\x00"
-     "\xED\x91"
-     "\x9F\x00\xED\x91\xA0\x00\xED\x91\xA1\x00\xED\x91\xA2\x00\xED\x91\xA3\x00"
-     "\xED\x91"
-     "\xA4\x00\xED\x91\xA5\x00\xED\x91\xA6\x00\xED\x91\xA7\x00\xED\x91\xA8\x00"
-     "\xED\x91"
-     "\xA9\x00\xED\x91\xAA\x00\xED\x91\xAB\x00\xED\x91\xAC\x00\xED\x91\xAD\x00"
-     "\xED\x91"
-     "\xAE\x00\xED\x91\xAF\x00\xED\x91\xB0\x00\xED\x91\xB1\x00\xED\x91\xB2\x00"
-     "\xED\x91"
-     "\xB3\x00\xED\x91\xB4\x00\xED\x91\xB5\x00\xED\x91\xB6\x00\xED\x91\xB7\x00"
-     "\xED\x91"
-     "\xB8\x00\xED\x91\xB9\x00\xED\x91\xBA\x00\xED\x91\xBB\x00\xED\x91\xBC\x00"
-     "\xED\x91"
-     "\xBD\x00\xED\x91\xBE\x00\xED\x91\xBF\x00\xED\x92\x80\x00\xED\x92\x81\x00"
-     "\xED\x92"
-     "\x82\x00\xED\x92\x83\x00\xED\x92\x84\x00\xED\x92\x85\x00\xED\x92\x86\x00"
-     "\xED\x92"
-     "\x87\x00\xED\x92\x88\x00\xED\x92\x89\x00\xED\x92\x8A\x00\xED\x92\x8B\x00"
-     "\xED\x92"
-     "\x8C\x00\xED\x92\x8D\x00\xED\x92\x8E\x00\xED\x92\x8F\x00\xED\x92\x90\x00"
-     "\xED\x92"
-     "\x91\x00\xED\x92\x92\x00\xED\x92\x93\x00\xED\x92\x94\x00\xED\x92\x95\x00"
-     "\xED\x92"
-     "\x96\x00\xED\x92\x97\x00\xED\x92\x98\x00\xED\x92\x99\x00\xED\x92\x9A\x00"
-     "\xED\x92"
-     "\x9B\x00\xED\x92\x9C\x00\xED\x92\x9D\x00\xED\x92\x9E\x00\xED\x92\x9F\x00"
-     "\xED\x92"
-     "\xA0\x00\xED\x92\xA1\x00\xED\x92\xA2\x00\xED\x92\xA3\x00\xED\x92\xA4\x00"
-     "\xED\x92"
-     "\xA5\x00\xED\x92\xA6\x00\xED\x92\xA7\x00\xED\x92\xA8\x00\xED\x92\xA9\x00"
-     "\xED\x92"
-     "\xAA\x00\xED\x92\xAB\x00\xED\x92\xAC\x00\xED\x92\xAD\x00\xED\x92\xAE\x00"
-     "\xED\x92"
-     "\xAF\x00\xED\x92\xB0\x00\xED\x92\xB1\x00\xED\x92\xB2\x00\xED\x92\xB3\x00"
-     "\xED\x92"
-     "\xB4\x00\xED\x92\xB5\x00\xED\x92\xB6\x00\xED\x92\xB7\x00\xED\x92\xB8\x00"
-     "\xED\x92"
-     "\xB9\x00\xED\x92\xBA\x00\xED\x92\xBB\x00\xED\x92\xBC\x00\xED\x92\xBD\x00"
-     "\xED\x92"
-     "\xBE\x00\xED\x92\xBF\x00\xED\x93\x80\x00\xED\x93\x81\x00\xED\x93\x82\x00"
-     "\xED\x93"
-     "\x83\x00\xED\x93\x84\x00\xED\x93\x85\x00\xED\x93\x86\x00\xED\x93\x87\x00"
-     "\xED\x93"
-     "\x88\x00\xED\x93\x89\x00\xED\x93\x8A\x00\xED\x93\x8B\x00\xED\x93\x8C\x00"
-     "\xED\x93"
-     "\x8D\x00\xED\x93\x8E\x00\xED\x93\x8F\x00\xED\x93\x90\x00\xED\x93\x91\x00"
-     "\xED\x93"
-     "\x92\x00\xED\x93\x93\x00\xED\x93\x94\x00\xED\x93\x95\x00\xED\x93\x96\x00"
-     "\xED\x93"
-     "\x97\x00\xED\x93\x98\x00\xED\x93\x99\x00\xED\x93\x9A\x00\xED\x93\x9B\x00"
-     "\xED\x93"
-     "\x9C\x00\xED\x93\x9D\x00\xED\x93\x9E\x00\xED\x93\x9F\x00\xED\x93\xA0\x00"
-     "\xED\x93"
-     "\xA1\x00\xED\x93\xA2\x00\xED\x93\xA3\x00\xED\x93\xA4\x00\xED\x93\xA5\x00"
-     "\xED\x93"
-     "\xA6\x00\xED\x93\xA7\x00\xED\x93\xA8\x00\xED\x93\xA9\x00\xED\x93\xAA\x00"
-     "\xED\x93"
-     "\xAB\x00\xED\x93\xAC\x00\xED\x93\xAD\x00\xED\x93\xAE\x00\xED\x93\xAF\x00"
-     "\xED\x93"
-     "\xB0\x00\xED\x93\xB1\x00\xED\x93\xB2\x00\xED\x93\xB3\x00\xED\x93\xB4\x00"
-     "\xED\x93"
-     "\xB5\x00\xED\x93\xB6\x00\xED\x93\xB7\x00\xED\x93\xB8\x00\xED\x93\xB9\x00"
-     "\xED\x93"
-     "\xBA\x00\xED\x93\xBB\x00\xED\x93\xBC\x00\xED\x93\xBD\x00\xED\x93\xBE\x00"
-     "\xED\x93"
-     "\xBF\x00\xED\x94\x80\x00\xED\x94\x81\x00\xED\x94\x82\x00\xED\x94\x83\x00"
-     "\xED\x94"
-     "\x84\x00\xED\x94\x85\x00\xED\x94\x86\x00\xED\x94\x87\x00\xED\x94\x88\x00"
-     "\xED\x94"
-     "\x89\x00\xED\x94\x8A\x00\xED\x94\x8B\x00\xED\x94\x8C\x00\xED\x94\x8D\x00"
-     "\xED\x94"
-     "\x8E\x00\xED\x94\x8F\x00\xED\x94\x90\x00\xED\x94\x91\x00\xED\x94\x92\x00"
-     "\xED\x94"
-     "\x93\x00\xED\x94\x94\x00\xED\x94\x95\x00\xED\x94\x96\x00\xED\x94\x97\x00"
-     "\xED\x94"
-     "\x98\x00\xED\x94\x99\x00\xED\x94\x9A\x00\xED\x94\x9B\x00\xED\x94\x9C\x00"
-     "\xED\x94"
-     "\x9D\x00\xED\x94\x9E\x00\xED\x94\x9F\x00\xED\x94\xA0\x00\xED\x94\xA1\x00"
-     "\xED\x94"
-     "\xA2\x00\xED\x94\xA3\x00\xED\x94\xA4\x00\xED\x94\xA5\x00\xED\x94\xA6\x00"
-     "\xED\x94"
-     "\xA7\x00\xED\x94\xA8\x00\xED\x94\xA9\x00\xED\x94\xAA\x00\xED\x94\xAB\x00"
-     "\xED\x94"
-     "\xAC\x00\xED\x94\xAD\x00\xED\x94\xAE\x00\xED\x94\xAF\x00\xED\x94\xB0\x00"
-     "\xED\x94"
-     "\xB1\x00\xED\x94\xB2\x00\xED\x94\xB3\x00\xED\x94\xB4\x00\xED\x94\xB5\x00"
-     "\xED\x94"
-     "\xB6\x00\xED\x94\xB7\x00\xED\x94\xB8\x00\xED\x94\xB9\x00\xED\x94\xBA\x00"
-     "\xED\x94"
-     "\xBB\x00\xED\x94\xBC\x00\xED\x94\xBD\x00\xED\x94\xBE\x00\xED\x94\xBF\x00"
-     "\xED\x95"
-     "\x80\x00\xED\x95\x81\x00\xED\x95\x82\x00\xED\x95\x83\x00\xED\x95\x84\x00"
-     "\xED\x95"
-     "\x85\x00\xED\x95\x86\x00\xED\x95\x87\x00\xED\x95\x88\x00\xED\x95\x89\x00"
-     "\xED\x95"
-     "\x8A\x00\xED\x95\x8B\x00\xED\x95\x8C\x00\xED\x95\x8D\x00\xED\x95\x8E\x00"
-     "\xED\x95"
-     "\x8F\x00\xED\x95\x90\x00\xED\x95\x91\x00\xED\x95\x92\x00\xED\x95\x93\x00"
-     "\xED\x95"
-     "\x94\x00\xED\x95\x95\x00\xED\x95\x96\x00\xED\x95\x97\x00\xED\x95\x98\x00"
-     "\xED\x95"
-     "\x99\x00\xED\x95\x9A\x00\xED\x95\x9B\x00\xED\x95\x9C\x00\xED\x95\x9D\x00"
-     "\xED\x95"
-     "\x9E\x00\xED\x95\x9F\x00\xED\x95\xA0\x00\xED\x95\xA1\x00\xED\x95\xA2\x00"
-     "\xED\x95"
-     "\xA3\x00\xED\x95\xA4\x00\xED\x95\xA5\x00\xED\x95\xA6\x00\xED\x95\xA7\x00"
-     "\xED\x95"
-     "\xA8\x00\xED\x95\xA9\x00\xED\x95\xAA\x00\xED\x95\xAB\x00\xED\x95\xAC\x00"
-     "\xED\x95"
-     "\xAD\x00\xED\x95\xAE\x00\xED\x95\xAF\x00\xED\x95\xB0\x00\xED\x95\xB1\x00"
-     "\xED\x95"
-     "\xB2\x00\xED\x95\xB3\x00\xED\x95\xB4\x00\xED\x95\xB5\x00\xED\x95\xB6\x00"
-     "\xED\x95"
-     "\xB7\x00\xED\x95\xB8\x00\xED\x95\xB9\x00\xED\x95\xBA\x00\xED\x95\xBB\x00"
-     "\xED\x95"
-     "\xBC\x00\xED\x95\xBD\x00\xED\x95\xBE\x00\xED\x95\xBF\x00\xED\x96\x80\x00"
-     "\xED\x96"
-     "\x81\x00\xED\x96\x82\x00\xED\x96\x83\x00\xED\x96\x84\x00\xED\x96\x85\x00"
-     "\xED\x96"
-     "\x86\x00\xED\x96\x87\x00\xED\x96\x88\x00\xED\x96\x89\x00\xED\x96\x8A\x00"
-     "\xED\x96"
-     "\x8B\x00\xED\x96\x8C\x00\xED\x96\x8D\x00\xED\x96\x8E\x00\xED\x96\x8F\x00"
-     "\xED\x96"
-     "\x90\x00\xED\x96\x91\x00\xED\x96\x92\x00\xED\x96\x93\x00\xED\x96\x94\x00"
-     "\xED\x96"
-     "\x95\x00\xED\x96\x96\x00\xED\x96\x97\x00\xED\x96\x98\x00\xED\x96\x99\x00"
-     "\xED\x96"
-     "\x9A\x00\xED\x96\x9B\x00\xED\x96\x9C\x00\xED\x96\x9D\x00\xED\x96\x9E\x00"
-     "\xED\x96"
-     "\x9F\x00\xED\x96\xA0\x00\xED\x96\xA1\x00\xED\x96\xA2\x00\xED\x96\xA3\x00"
-     "\xED\x96"
-     "\xA4\x00\xED\x96\xA5\x00\xED\x96\xA6\x00\xED\x96\xA7\x00\xED\x96\xA8\x00"
-     "\xED\x96"
-     "\xA9\x00\xED\x96\xAA\x00\xED\x96\xAB\x00\xED\x96\xAC\x00\xED\x96\xAD\x00"
-     "\xED\x96"
-     "\xAE\x00\xED\x96\xAF\x00\xED\x96\xB0\x00\xED\x96\xB1\x00\xED\x96\xB2\x00"
-     "\xED\x96"
-     "\xB3\x00\xED\x96\xB4\x00\xED\x96\xB5\x00\xED\x96\xB6\x00\xED\x96\xB7\x00"
-     "\xED\x96"
-     "\xB8\x00\xED\x96\xB9\x00\xED\x96\xBA\x00\xED\x96\xBB\x00\xED\x96\xBC\x00"
-     "\xED\x96"
-     "\xBD\x00\xED\x96\xBE\x00\xED\x96\xBF\x00\xED\x97\x80\x00\xED\x97\x81\x00"
-     "\xED\x97"
-     "\x82\x00\xED\x97\x83\x00\xED\x97\x84\x00\xED\x97\x85\x00\xED\x97\x86\x00"
-     "\xED\x97"
-     "\x87\x00\xED\x97\x88\x00\xED\x97\x89\x00\xED\x97\x8A\x00\xED\x97\x8B\x00"
-     "\xED\x97"
-     "\x8C\x00\xED\x97\x8D\x00\xED\x97\x8E\x00\xED\x97\x8F\x00\xED\x97\x90\x00"
-     "\xED\x97"
-     "\x91\x00\xED\x97\x92\x00\xED\x97\x93\x00\xED\x97\x94\x00\xED\x97\x95\x00"
-     "\xED\x97"
-     "\x96\x00\xED\x97\x97\x00\xED\x97\x98\x00\xED\x97\x99\x00\xED\x97\x9A\x00"
-     "\xED\x97"
-     "\x9B\x00\xED\x97\x9C\x00\xED\x97\x9D\x00\xED\x97\x9E\x00\xED\x97\x9F\x00"
-     "\xED\x97"
-     "\xA0\x00\xED\x97\xA1\x00\xED\x97\xA2\x00\xED\x97\xA3\x00\xED\x97\xA4\x00"
-     "\xED\x97"
-     "\xA5\x00\xED\x97\xA6\x00\xED\x97\xA7\x00\xED\x97\xA8\x00\xED\x97\xA9\x00"
-     "\xED\x97"
-     "\xAA\x00\xED\x97\xAB\x00\xED\x97\xAC\x00\xED\x97\xAD\x00\xED\x97\xAE\x00"
-     "\xED\x97"
-     "\xAF\x00\xED\x97\xB0\x00\xED\x97\xB1\x00\xED\x97\xB2\x00\xED\x97\xB3\x00"
-     "\xED\x97"
-     "\xB4\x00\xED\x97\xB5\x00\xED\x97\xB6\x00\xED\x97\xB7\x00\xED\x97\xB8\x00"
-     "\xED\x97"
-     "\xB9\x00\xED\x97\xBA\x00\xED\x97\xBB\x00\xED\x97\xBC\x00\xED\x97\xBD\x00"
-     "\xED\x97"
-     "\xBE\x00\xED\x97\xBF\x00\xED\x98\x80\x00\xED\x98\x81\x00\xED\x98\x82\x00"
-     "\xED\x98"
-     "\x83\x00\xED\x98\x84\x00\xED\x98\x85\x00\xED\x98\x86\x00\xED\x98\x87\x00"
-     "\xED\x98"
-     "\x88\x00\xED\x98\x89\x00\xED\x98\x8A\x00\xED\x98\x8B\x00\xED\x98\x8C\x00"
-     "\xED\x98"
-     "\x8D\x00\xED\x98\x8E\x00\xED\x98\x8F\x00\xED\x98\x90\x00\xED\x98\x91\x00"
-     "\xED\x98"
-     "\x92\x00\xED\x98\x93\x00\xED\x98\x94\x00\xED\x98\x95\x00\xED\x98\x96\x00"
-     "\xED\x98"
-     "\x97\x00\xED\x98\x98\x00\xED\x98\x99\x00\xED\x98\x9A\x00\xED\x98\x9B\x00"
-     "\xED\x98"
-     "\x9C\x00\xED\x98\x9D\x00\xED\x98\x9E\x00\xED\x98\x9F\x00\xED\x98\xA0\x00"
-     "\xED\x98"
-     "\xA1\x00\xED\x98\xA2\x00\xED\x98\xA3\x00\xED\x98\xA4\x00\xED\x98\xA5\x00"
-     "\xED\x98"
-     "\xA6\x00\xED\x98\xA7\x00\xED\x98\xA8\x00\xED\x98\xA9\x00\xED\x98\xAA\x00"
-     "\xED\x98"
-     "\xAB\x00\xED\x98\xAC\x00\xED\x98\xAD\x00\xED\x98\xAE\x00\xED\x98\xAF\x00"
-     "\xED\x98"
-     "\xB0\x00\xED\x98\xB1\x00\xED\x98\xB2\x00\xED\x98\xB3\x00\xED\x98\xB4\x00"
-     "\xED\x98"
-     "\xB5\x00\xED\x98\xB6\x00\xED\x98\xB7\x00\xED\x98\xB8\x00\xED\x98\xB9\x00"
-     "\xED\x98"
-     "\xBA\x00\xED\x98\xBB\x00\xED\x98\xBC\x00\xED\x98\xBD\x00\xED\x98\xBE\x00"
-     "\xED\x98"
-     "\xBF\x00\xED\x99\x80\x00\xED\x99\x81\x00\xED\x99\x82\x00\xED\x99\x83\x00"
-     "\xED\x99"
-     "\x84\x00\xED\x99\x85\x00\xED\x99\x86\x00\xED\x99\x87\x00\xED\x99\x88\x00"
-     "\xED\x99"
-     "\x89\x00\xED\x99\x8A\x00\xED\x99\x8B\x00\xED\x99\x8C\x00\xED\x99\x8D\x00"
-     "\xED\x99"
-     "\x8E\x00\xED\x99\x8F\x00\xED\x99\x90\x00\xED\x99\x91\x00\xED\x99\x92\x00"
-     "\xED\x99"
-     "\x93\x00\xED\x99\x94\x00\xED\x99\x95\x00\xED\x99\x96\x00\xED\x99\x97\x00"
-     "\xED\x99"
-     "\x98\x00\xED\x99\x99\x00\xED\x99\x9A\x00\xED\x99\x9B\x00\xED\x99\x9C\x00"
-     "\xED\x99"
-     "\x9D\x00\xED\x99\x9E\x00\xED\x99\x9F\x00\xED\x99\xA0\x00\xED\x99\xA1\x00"
-     "\xED\x99"
-     "\xA2\x00\xED\x99\xA3\x00\xED\x99\xA4\x00\xED\x99\xA5\x00\xED\x99\xA6\x00"
-     "\xED\x99"
-     "\xA7\x00\xED\x99\xA8\x00\xED\x99\xA9\x00\xED\x99\xAA\x00\xED\x99\xAB\x00"
-     "\xED\x99"
-     "\xAC\x00\xED\x99\xAD\x00\xED\x99\xAE\x00\xED\x99\xAF\x00\xED\x99\xB0\x00"
-     "\xED\x99"
-     "\xB1\x00\xED\x99\xB2\x00\xED\x99\xB3\x00\xED\x99\xB4\x00\xED\x99\xB5\x00"
-     "\xED\x99"
-     "\xB6\x00\xED\x99\xB7\x00\xED\x99\xB8\x00\xED\x99\xB9\x00\xED\x99\xBA\x00"
-     "\xED\x99"
-     "\xBB\x00\xED\x99\xBC\x00\xED\x99\xBD\x00\xED\x99\xBE\x00\xED\x99\xBF\x00"
-     "\xED\x9A"
-     "\x80\x00\xED\x9A\x81\x00\xED\x9A\x82\x00\xED\x9A\x83\x00\xED\x9A\x84\x00"
-     "\xED\x9A"
-     "\x85\x00\xED\x9A\x86\x00\xED\x9A\x87\x00\xED\x9A\x88\x00\xED\x9A\x89\x00"
-     "\xED\x9A"
-     "\x8A\x00\xED\x9A\x8B\x00\xED\x9A\x8C\x00\xED\x9A\x8D\x00\xED\x9A\x8E\x00"
-     "\xED\x9A"
-     "\x8F\x00\xED\x9A\x90\x00\xED\x9A\x91\x00\xED\x9A\x92\x00\xED\x9A\x93\x00"
-     "\xED\x9A"
-     "\x94\x00\xED\x9A\x95\x00\xED\x9A\x96\x00\xED\x9A\x97\x00\xED\x9A\x98\x00"
-     "\xED\x9A"
-     "\x99\x00\xED\x9A\x9A\x00\xED\x9A\x9B\x00\xED\x9A\x9C\x00\xED\x9A\x9D\x00"
-     "\xED\x9A"
-     "\x9E\x00\xED\x9A\x9F\x00\xED\x9A\xA0\x00\xED\x9A\xA1\x00\xED\x9A\xA2\x00"
-     "\xED\x9A"
-     "\xA3\x00\xED\x9A\xA4\x00\xED\x9A\xA5\x00\xED\x9A\xA6\x00\xED\x9A\xA7\x00"
-     "\xED\x9A"
-     "\xA8\x00\xED\x9A\xA9\x00\xED\x9A\xAA\x00\xED\x9A\xAB\x00\xED\x9A\xAC\x00"
-     "\xED\x9A"
-     "\xAD\x00\xED\x9A\xAE\x00\xED\x9A\xAF\x00\xED\x9A\xB0\x00\xED\x9A\xB1\x00"
-     "\xED\x9A"
-     "\xB2\x00\xED\x9A\xB3\x00\xED\x9A\xB4\x00\xED\x9A\xB5\x00\xED\x9A\xB6\x00"
-     "\xED\x9A"
-     "\xB7\x00\xED\x9A\xB8\x00\xED\x9A\xB9\x00\xED\x9A\xBA\x00\xED\x9A\xBB\x00"
-     "\xED\x9A"
-     "\xBC\x00\xED\x9A\xBD\x00\xED\x9A\xBE\x00\xED\x9A\xBF\x00\xED\x9B\x80\x00"
-     "\xED\x9B"
-     "\x81\x00\xED\x9B\x82\x00\xED\x9B\x83\x00\xED\x9B\x84\x00\xED\x9B\x85\x00"
-     "\xED\x9B"
-     "\x86\x00\xED\x9B\x87\x00\xED\x9B\x88\x00\xED\x9B\x89\x00\xED\x9B\x8A\x00"
-     "\xED\x9B"
-     "\x8B\x00\xED\x9B\x8C\x00\xED\x9B\x8D\x00\xED\x9B\x8E\x00\xED\x9B\x8F\x00"
-     "\xED\x9B"
-     "\x90\x00\xED\x9B\x91\x00\xED\x9B\x92\x00\xED\x9B\x93\x00\xED\x9B\x94\x00"
-     "\xED\x9B"
-     "\x95\x00\xED\x9B\x96\x00\xED\x9B\x97\x00\xED\x9B\x98\x00\xED\x9B\x99\x00"
-     "\xED\x9B"
-     "\x9A\x00\xED\x9B\x9B\x00\xED\x9B\x9C\x00\xED\x9B\x9D\x00\xED\x9B\x9E\x00"
-     "\xED\x9B"
-     "\x9F\x00\xED\x9B\xA0\x00\xED\x9B\xA1\x00\xED\x9B\xA2\x00\xED\x9B\xA3\x00"
-     "\xED\x9B"
-     "\xA4\x00\xED\x9B\xA5\x00\xED\x9B\xA6\x00\xED\x9B\xA7\x00\xED\x9B\xA8\x00"
-     "\xED\x9B"
-     "\xA9\x00\xED\x9B\xAA\x00\xED\x9B\xAB\x00\xED\x9B\xAC\x00\xED\x9B\xAD\x00"
-     "\xED\x9B"
-     "\xAE\x00\xED\x9B\xAF\x00\xED\x9B\xB0\x00\xED\x9B\xB1\x00\xED\x9B\xB2\x00"
-     "\xED\x9B"
-     "\xB3\x00\xED\x9B\xB4\x00\xED\x9B\xB5\x00\xED\x9B\xB6\x00\xED\x9B\xB7\x00"
-     "\xED\x9B"
-     "\xB8\x00\xED\x9B\xB9\x00\xED\x9B\xBA\x00\xED\x9B\xBB\x00\xED\x9B\xBC\x00"
-     "\xED\x9B"
-     "\xBD\x00\xED\x9B\xBE\x00\xED\x9B\xBF\x00\xED\x9C\x80\x00\xED\x9C\x81\x00"
-     "\xED\x9C"
-     "\x82\x00\xED\x9C\x83\x00\xED\x9C\x84\x00\xED\x9C\x85\x00\xED\x9C\x86\x00"
-     "\xED\x9C"
-     "\x87\x00\xED\x9C\x88\x00\xED\x9C\x89\x00\xED\x9C\x8A\x00\xED\x9C\x8B\x00"
-     "\xED\x9C"
-     "\x8C\x00\xED\x9C\x8D\x00\xED\x9C\x8E\x00\xED\x9C\x8F\x00\xED\x9C\x90\x00"
-     "\xED\x9C"
-     "\x91\x00\xED\x9C\x92\x00\xED\x9C\x93\x00\xED\x9C\x94\x00\xED\x9C\x95\x00"
-     "\xED\x9C"
-     "\x96\x00\xED\x9C\x97\x00\xED\x9C\x98\x00\xED\x9C\x99\x00\xED\x9C\x9A\x00"
-     "\xED\x9C"
-     "\x9B\x00\xED\x9C\x9C\x00\xED\x9C\x9D\x00\xED\x9C\x9E\x00\xED\x9C\x9F\x00"
-     "\xED\x9C"
-     "\xA0\x00\xED\x9C\xA1\x00\xED\x9C\xA2\x00\xED\x9C\xA3\x00\xED\x9C\xA4\x00"
-     "\xED\x9C"
-     "\xA5\x00\xED\x9C\xA6\x00\xED\x9C\xA7\x00\xED\x9C\xA8\x00\xED\x9C\xA9\x00"
-     "\xED\x9C"
-     "\xAA\x00\xED\x9C\xAB\x00\xED\x9C\xAC\x00\xED\x9C\xAD\x00\xED\x9C\xAE\x00"
-     "\xED\x9C"
-     "\xAF\x00\xED\x9C\xB0\x00\xED\x9C\xB1\x00\xED\x9C\xB2\x00\xED\x9C\xB3\x00"
-     "\xED\x9C"
-     "\xB4\x00\xED\x9C\xB5\x00\xED\x9C\xB6\x00\xED\x9C\xB7\x00\xED\x9C\xB8\x00"
-     "\xED\x9C"
-     "\xB9\x00\xED\x9C\xBA\x00\xED\x9C\xBB\x00\xED\x9C\xBC\x00\xED\x9C\xBD\x00"
-     "\xED\x9C"
-     "\xBE\x00\xED\x9C\xBF\x00\xED\x9D\x80\x00\xED\x9D\x81\x00\xED\x9D\x82\x00"
-     "\xED\x9D"
-     "\x83\x00\xED\x9D\x84\x00\xED\x9D\x85\x00\xED\x9D\x86\x00\xED\x9D\x87\x00"
-     "\xED\x9D"
-     "\x88\x00\xED\x9D\x89\x00\xED\x9D\x8A\x00\xED\x9D\x8B\x00\xED\x9D\x8C\x00"
-     "\xED\x9D"
-     "\x8D\x00\xED\x9D\x8E\x00\xED\x9D\x8F\x00\xED\x9D\x90\x00\xED\x9D\x91\x00"
-     "\xED\x9D"
-     "\x92\x00\xED\x9D\x93\x00\xED\x9D\x94\x00\xED\x9D\x95\x00\xED\x9D\x96\x00"
-     "\xED\x9D"
-     "\x97\x00\xED\x9D\x98\x00\xED\x9D\x99\x00\xED\x9D\x9A\x00\xED\x9D\x9B\x00"
-     "\xED\x9D"
-     "\x9C\x00\xED\x9D\x9D\x00\xED\x9D\x9E\x00\xED\x9D\x9F\x00\xED\x9D\xA0\x00"
-     "\xED\x9D"
-     "\xA1\x00\xED\x9D\xA2\x00\xED\x9D\xA3\x00\xED\x9D\xA4\x00\xED\x9D\xA5\x00"
-     "\xED\x9D"
-     "\xA6\x00\xED\x9D\xA7\x00\xED\x9D\xA8\x00\xED\x9D\xA9\x00\xED\x9D\xAA\x00"
-     "\xED\x9D"
-     "\xAB\x00\xED\x9D\xAC\x00\xED\x9D\xAD\x00\xED\x9D\xAE\x00\xED\x9D\xAF\x00"
-     "\xED\x9D"
-     "\xB0\x00\xED\x9D\xB1\x00\xED\x9D\xB2\x00\xED\x9D\xB3\x00\xED\x9D\xB4\x00"
-     "\xED\x9D"
-     "\xB5\x00\xED\x9D\xB6\x00\xED\x9D\xB7\x00\xED\x9D\xB8\x00\xED\x9D\xB9\x00"
-     "\xED\x9D"
-     "\xBA\x00\xED\x9D\xBB\x00\xED\x9D\xBC\x00\xED\x9D\xBD\x00\xED\x9D\xBE\x00"
-     "\xED\x9D"
-     "\xBF\x00\xED\x9E\x80\x00\xED\x9E\x81\x00\xED\x9E\x82\x00\xED\x9E\x83\x00"
-     "\xED\x9E"
-     "\x84\x00\xED\x9E\x85\x00\xED\x9E\x86\x00\xED\x9E\x87\x00\xED\x9E\x88\x00"
-     "\xED\x9E"
-     "\x89\x00\xED\x9E\x8A\x00\xED\x9E\x8B\x00\xED\x9E\x8C\x00\xED\x9E\x8D\x00"
-     "\xED\x9E"
-     "\x8E\x00\xED\x9E\x8F\x00\xED\x9E\x90\x00\xED\x9E\x91\x00\xED\x9E\x92\x00"
-     "\xED\x9E"
-     "\x93\x00\xED\x9E\x94\x00\xED\x9E\x95\x00\xED\x9E\x96\x00\xED\x9E\x97\x00"
-     "\xED\x9E"
-     "\x98\x00\xED\x9E\x99\x00\xED\x9E\x9A\x00\xED\x9E\x9B\x00\xED\x9E\x9C\x00"
-     "\xED\x9E"
-     "\x9D\x00\xED\x9E\x9E\x00\xED\x9E\x9F\x00\xED\x9E\xA0\x00\xED\x9E\xA1\x00"
-     "\xED\x9E"
-     "\xA2\x00\xED\x9E\xA3\x00\xF0\x91\x82\x9A\x00\xF0\x91\x82\x9C\x00\xF0\x91"
-     "\x82\xAB"
-     "\x00\xF0\x91\x84\xAE\x00\xF0\x91\x84\xAF\x00\xF0\x91\x8D\x8B\x00\xF0\x91"
-     "\x8D\x8C"
-     "\x00\xF0\x91\x92\xBB\x00\xF0\x91\x92\xBC\x00\xF0\x91\x92\xBE\x00\xF0\x91"
-     "\x96\xBA"
-     "\x00\xF0\x91\x96\xBB\x00\xF0\x91\xA4\xB8\x00\xF0\x9D\x85\x97\xF0\x9D\x85"
-     "\xA5\x00"
-     "\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\x00\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\xF0"
-     "\x9D\x85"
-     "\xAE\x00\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\xF0\x9D\x85\xAF\x00\xF0\x9D\x85"
-     "\x98\xF0"
-     "\x9D\x85\xA5\xF0\x9D\x85\xB0\x00\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\xF0\x9D"
-     "\x85\xB1"
-     "\x00\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\xF0\x9D\x85\xB2\x00\xF0\x9D\x86\xB9"
-     "\xF0\x9D"
-     "\x85\xA5\x00\xF0\x9D\x86\xB9\xF0\x9D\x85\xA5\xF0\x9D\x85\xAE\x00\xF0\x9D"
-     "\x86\xB9"
-     "\xF0\x9D\x85\xA5\xF0\x9D\x85\xAF\x00\xF0\x9D\x86\xBA\xF0\x9D\x85\xA5\x00"
-     "\xF0\x9D"
-     "\x86\xBA\xF0\x9D\x85\xA5\xF0\x9D\x85\xAE\x00\xF0\x9D\x86\xBA\xF0\x9D\x85"
-     "\xA5\xF0"
-     "\x9D\x85\xAF\x00\xF0\x9D\xBC\x84\x00\xF0\x9D\xBC\x85\x00\xF0\x9D\xBC\x86"
-     "\x00\xF0"
-     "\x9D\xBC\x88\x00\xF0\x9D\xBC\x8A\x00\xF0\x9D\xBC\x9E\x00\xF0\xA0\x84\xA2"
-     "\x00\xF0"
-     "\xA0\x94\x9C\x00\xF0\xA0\x94\xA5\x00\xF0\xA0\x95\x8B\x00\xF0\xA0\x98\xBA"
-     "\x00\xF0"
-     "\xA0\xA0\x84\x00\xF0\xA0\xA3\x9E\x00\xF0\xA0\xA8\xAC\x00\xF0\xA0\xAD\xA3"
-     "\x00\xF0"
-     "\xA1\x93\xA4\x00\xF0\xA1\x9A\xA8\x00\xF0\xA1\x9B\xAA\x00\xF0\xA1\xA7\x88"
-     "\x00\xF0"
-     "\xA1\xAC\x98\x00\xF0\xA1\xB4\x8B\x00\xF0\xA1\xB7\xA4\x00\xF0\xA1\xB7\xA6"
-     "\x00\xF0"
-     "\xA2\x86\x83\x00\xF0\xA2\x86\x9F\x00\xF0\xA2\x8C\xB1\x00\xF0\xA2\x9B\x94"
-     "\x00\xF0"
-     "\xA2\xA1\x84\x00\xF0\xA2\xA1\x8A\x00\xF0\xA2\xAC\x8C\x00\xF0\xA2\xAF\xB1"
-     "\x00\xF0"
-     "\xA3\x80\x8A\x00\xF0\xA3\x8A\xB8\x00\xF0\xA3\x8D\x9F\x00\xF0\xA3\x8E\x93"
-     "\x00\xF0"
-     "\xA3\x8E\x9C\x00\xF0\xA3\x8F\x83\x00\xF0\xA3\x8F\x95\x00\xF0\xA3\x91\xAD"
-     "\x00\xF0"
-     "\xA3\x9A\xA3\x00\xF0\xA3\xA2\xA7\x00\xF0\xA3\xAA\x8D\x00\xF0\xA3\xAB\xBA"
-     "\x00\xF0"
-     "\xA3\xB2\xBC\x00\xF0\xA3\xB4\x9E\x00\xF0\xA3\xBB\x91\x00\xF0\xA3\xBD\x9E"
-     "\x00\xF0"
-     "\xA3\xBE\x8E\x00\xF0\xA4\x89\xA3\x00\xF0\xA4\x8B\xAE\x00\xF0\xA4\x8E\xAB"
-     "\x00\xF0"
-     "\xA4\x98\x88\x00\xF0\xA4\x9C\xB5\x00\xF0\xA4\xA0\x94\x00\xF0\xA4\xB0\xB6"
-     "\x00\xF0"
-     "\xA4\xB2\x92\x00\xF0\xA4\xBE\xA1\x00\xF0\xA4\xBE\xB8\x00\xF0\xA5\x81\x84"
-     "\x00\xF0"
-     "\xA5\x83\xB2\x00\xF0\xA5\x83\xB3\x00\xF0\xA5\x84\x99\x00\xF0\xA5\x84\xB3"
-     "\x00\xF0"
-     "\xA5\x89\x89\x00\xF0\xA5\x90\x9D\x00\xF0\xA5\x98\xA6\x00\xF0\xA5\x9A\x9A"
-     "\x00\xF0"
-     "\xA5\x9B\x85\x00\xF0\xA5\xA5\xBC\x00\xF0\xA5\xAA\xA7\x00\xF0\xA5\xAE\xAB"
-     "\x00\xF0"
-     "\xA5\xB2\x80\x00\xF0\xA5\xB3\x90\x00\xF0\xA5\xBE\x86\x00\xF0\xA6\x87\x9A"
-     "\x00\xF0"
-     "\xA6\x88\xA8\x00\xF0\xA6\x89\x87\x00\xF0\xA6\x8B\x99\x00\xF0\xA6\x8C\xBE"
-     "\x00\xF0"
-     "\xA6\x93\x9A\x00\xF0\xA6\x94\xA3\x00\xF0\xA6\x96\xA8\x00\xF0\xA6\x9E\xA7"
-     "\x00\xF0"
-     "\xA6\x9E\xB5\x00\xF0\xA6\xAC\xBC\x00\xF0\xA6\xB0\xB6\x00\xF0\xA6\xB3\x95"
-     "\x00\xF0"
-     "\xA6\xB5\xAB\x00\xF0\xA6\xBC\xAC\x00\xF0\xA6\xBE\xB1\x00\xF0\xA7\x83\x92"
-     "\x00\xF0"
-     "\xA7\x8F\x8A\x00\xF0\xA7\x99\xA7\x00\xF0\xA7\xA2\xAE\x00\xF0\xA7\xA5\xA6"
-     "\x00\xF0"
-     "\xA7\xB2\xA8\x00\xF0\xA7\xBB\x93\x00\xF0\xA7\xBC\xAF\x00\xF0\xA8\x97\x92"
-     "\x00\xF0"
-     "\xA8\x97\xAD\x00\xF0\xA8\x9C\xAE\x00\xF0\xA8\xAF\xBA\x00\xF0\xA8\xB5\xB7"
-     "\x00\xF0"
-     "\xA9\x85\x85\x00\xF0\xA9\x87\x9F\x00\xF0\xA9\x88\x9A\x00\xF0\xA9\x90\x8A"
-     "\x00\xF0"
-     "\xA9\x92\x96\x00\xF0\xA9\x96\xB6\x00\xF0\xA9\xAC\xB0\x00\xF0\xAA\x83\x8E"
-     "\x00\xF0"
-     "\xAA\x84\x85\x00\xF0\xAA\x88\x8E\x00\xF0\xAA\x8A\x91\x00\xF0\xAA\x8E\x92"
-     "\x00\xF0"
-     "\xAA\x98\x80\x00"},
-    {"nmt_nfkc", 238883,
-     "\x00\xB8\x02\x00\x00\x84\x00\x00\x00\x00\x00\x80\x01\x00\x00\x80\xCC\xFC"
-     "\x02\x00"
-     "\xB8\x05\x00\x00\x81\x23\x00\x80\xCC\xE4\x02\x00\xB8\x3D\x00\x00\x75\x23"
-     "\x00\x80"
-     "\xCC\xEC\x02\x00\xB8\x05\x00\x00\x85\x23\x00\x80\xCC\x3C\x00\x00\xCD\xBC"
-     "\x00\x00"
-     "\x9E\x09\x00\x80\xA1\x09\x00\x80\xA4\x09\x00\x80\x83\x1D\x00\x00\x80\x19"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x37\x1E\x00\x80\x2F\x1E\x00\x80\x47\x1E\x00\x80\x3F\x1E"
-     "\x00\x80"
-     "\x80\x31\x00\x00\x81\x31\x00\x00\xA7\x09\x00\x80\x89\x31\x00\x00\x3D\x58"
-     "\x03\x00"
-     "\x3C\x48\x03\x00\x49\x0A\x00\x80\x3E\x68\x03\x00\x01\x85\x00\x00\x84\x01"
-     "\x01\x00"
-     "\x03\x8D\x00\x00\x02\x89\x00\x00\x05\x95\x00\x00\x04\x91\x00\x00\x07\x9D"
-     "\x00\x00"
-     "\x06\x99\x00\x00\x09\xA9\x00\x00\x08\xA1\x00\x00\x0B\xAD\x00\x00\x0A\xA5"
-     "\x00\x00"
-     "\x0D\xB9\x00\x00\x0C\xBD\x00\x00\x0F\xBD\x00\x00\x0E\xB9\x00\x00\x11\xC5"
-     "\x00\x00"
-     "\x10\xC1\x00\x00\x13\xCD\x00\x00\x12\xC9\x00\x00\x15\xD5\x00\x00\x14\xD1"
-     "\x00\x00"
-     "\x17\xDD\x00\x00\x16\xD9\x00\x00\x19\xE5\x00\x00\x18\xE1\x00\x00\x1B\xED"
-     "\x00\x00"
-     "\x1A\xE9\x00\x00\x1D\xF5\x00\x00\x1C\xF1\x00\x00\x1F\xFD\x00\x00\x1E\xF9"
-     "\x00\x00"
-     "\x61\x38\x09\x00\x5F\x1E\x00\x80\x63\x40\x0E\x00\x62\xF0\x08\x00\x65\x28"
-     "\x0F\x00"
-     "\x64\x80\x0E\x00\x67\xD8\x0F\x00\x66\xF0\x0F\x00\x69\x70\x0C\x00\x68\x30"
-     "\x0C\x00"
-     "\x6B\xF0\x0C\x00\x6A\xD8\x0C\x00\x6D\x48\x0D\x00\x6C\x1C\x0D\x00\x6F\x10"
-     "\x12\x00"
-     "\x6E\x6C\x0D\x00\x4F\x0A\x00\x80\x70\x0C\x13\x00\x73\xA8\x13\x00\x72\xEC"
-     "\x13\x00"
-     "\x75\x28\x10\x00\x74\x4C\x10\x00\x77\xD0\x11\x00\x76\x14\x10\x00\x79\xD0"
-     "\x16\x00"
-     "\x78\xF4\x10\x00\x57\x1E\x00\x80\x7A\xF4\x16\x00\x6F\x1E\x00\x80\x67\x1E"
-     "\x00\x80"
-     "\x7F\x7D\x01\x00\x98\x0C\x00\x80\x41\x80\x02\x00\xF0\x0B\x00\x80\x43\x18"
-     "\x00\x00"
-     "\x42\xC0\x00\x00\x45\x48\x00\x00\x44\x60\x00\x00\x47\x90\x06\x00\x46\x84"
-     "\x01\x00"
-     "\x49\x28\x06\x00\x48\x6C\x01\x00\x4B\x38\x07\x00\x4A\xF0\x07\x00\x4D\x70"
-     "\x07\x00"
-     "\x4C\x44\x07\x00\x4F\xF0\x04\x00\x4E\xCC\x04\x00\xAA\x09\x00\x80\x50\x2C"
-     "\x05\x00"
-     "\x53\xA0\x0A\x00\x52\x10\x05\x00\x55\x00\x0A\x00\x54\x50\x0A\x00\x57\xD0"
-     "\x08\x00"
-     "\x56\x10\x0B\x00\x59\x48\x08\x00\x58\xB8\x08\x00\x84\x11\x00\x00\x5A\x3C"
-     "\x08\x00"
-     "\x80\xF5\x00\x00\x81\x9D\x00\x00\xEA\x0B\x00\x80\x27\x1E\x00\x80\x83\x69"
-     "\x02\x00"
-     "\x82\x45\x02\x00\x81\x35\x02\x00\x80\x35\x02\x00\x87\x6D\x03\x00\x86\x55"
-     "\x03\x00"
-     "\x81\x39\x00\x00\x84\x65\x02\x00\x29\x0C\x00\x80\x8A\x01\x04\x00\x89\xD5"
-     "\x03\x00"
-     "\x88\xED\x03\x00\x8F\x09\x00\x00\xBA\x0B\x00\x80\x3B\x0C\x00\x80\x8C\x09"
-     "\x00\x00"
-     "\x41\x0C\x00\x80\x89\x31\x03\x00\x91\x09\x00\x00\xCC\xD8\x00\x00\x4F\x1E"
-     "\x00\x80"
-     "\x77\x1E\x00\x80\xB7\x1B\x00\x80\x55\x0A\x00\x80\x80\x65\x03\x00\x81\x0D"
-     "\x03\x00"
-     "\x86\x3D\x00\x00\x83\x1D\x03\x00\xCC\x10\x02\x00\xCD\x84\x01\x00\x82\x29"
-     "\x00\x00"
-     "\xCC\x74\x03\x00\xA3\x81\x06\x00\xBF\x1B\x00\x80\x80\xB1\x02\x00\x81\xB1"
-     "\x02\x00"
-     "\xC7\x1B\x00\x80\x81\x29\x00\x00\xA5\xC1\x00\x00\xCF\x1B\x00\x80\xCC\xE8"
-     "\x03\x00"
-     "\xCD\x60\x02\x00\x5B\x0A\x00\x80\xA8\xF1\x00\x00\x61\x0A\x00\x80\x67\x0A"
-     "\x00\x80"
-     "\x6D\x0A\x00\x80\xD7\x1B\x00\x80\x81\x69\x00\x00\xCC\xDC\x04\x00\x82\x11"
-     "\x01\x00"
-     "\xDF\x1B\x00\x80\x73\x0A\x00\x80\xE7\x1B\x00\x80\xF7\x1B\x00\x80\xFF\x1B"
-     "\x00\x80"
-     "\xEF\x1B\x00\x80\xB6\x09\x00\x80\xCC\xA0\x04\x00\xCD\xDC\x04\x00\xCC\x08"
-     "\x01\x00"
-     "\xB9\x09\x00\x80\x9F\x1E\x00\x80\x86\x11\x01\x00\xE1\x00\x2B\x00\xE0\x7C"
-     "\x27\x00"
-     "\xE3\x88\x48\x02\xE2\x0C\x38\x02\x97\x1E\x00\x80\x87\x91\x01\x00\xAF\x1E"
-     "\x00\x80"
-     "\xA7\x1E\x00\x80\x80\xD9\x01\x00\x81\x35\x00\x00\xCC\xC4\x02\x00\xEA\x44"
-     "\x64\x02"
-     "\x8F\x1E\x00\x80\x07\x1C\x00\x80\xEF\xF8\x67\x02\x81\x11\x07\x00\x82\xF1"
-     "\x01\x00"
-     "\xF0\x0C\x8D\x02\x89\x55\x00\x00\x81\xE5\x01\x00\x0F\x1C\x00\x80\x87\xE1"
-     "\x01\x00"
-     "\x80\x6D\x00\x00\x81\x0D\x00\x00\x83\x79\x00\x00\x7F\x0A\x00\x80\x81\x79"
-     "\x00\x00"
-     "\x80\x95\x01\x00\xCC\x38\x01\x00\xCD\x14\x01\x00\x8C\xC1\x01\x00\x85\x0A"
-     "\x00\x80"
-     "\xBF\x09\x00\x80\xA3\x15\x01\x00\xC3\x94\x17\x00\xC2\x9C\x14\x00\xC5\x84"
-     "\x17\x00"
-     "\xC4\x50\x17\x00\xC7\xEC\x17\x00\xC6\x80\x12\x00\x87\x1E\x00\x80\x91\x0A"
-     "\x00\x80"
-     "\xCB\xD0\x16\x00\xCA\xE0\x16\x00\xCD\x14\x16\x00\x47\x0C\x00\x80\xCF\xBC"
-     "\x20\x00"
-     "\xCE\x9C\x19\x00\xD1\xCC\x24\x00\xD0\xD8\x25\x00\xD3\xE8\x24\x00\xB1\x51"
-     "\x01\x00"
-     "\x4D\x0C\x00\x80\xA7\xDD\x07\x00\xB7\x1E\x00\x80\xD6\xBC\x24\x00\xD9\xC8"
-     "\x27\x00"
-     "\xD8\xD4\x22\x00\xDB\xF8\x27\x00\x27\x1C\x00\x80\x87\xED\x07\x00\x8B\x0A"
-     "\x00\x80"
-     "\xCC\xF8\x04\x00\x17\x1C\x00\x80\x1F\x1E\x00\x80\x87\xC9\x06\x00\xB3\x09"
-     "\x00\x80"
-     "\x91\xDD\x07\x00\xBC\x09\x00\x80\x1F\x1C\x00\x80\x79\x0A\x00\x80\x2F\x1C"
-     "\x00\x80"
-     "\x7F\x1E\x00\x80\x9E\x0C\x00\x80\x8C\xF9\x06\x00\x1D\x0C\x00\x80\x80\xD5"
-     "\x06\x00"
-     "\x81\xC1\x06\x00\xCC\xC4\x02\x00\xCD\x04\x05\x00\x82\x51\x00\x00\x83\x75"
-     "\x07\x00"
-     "\x80\xAD\x06\x00\x81\xB9\x06\x00\x86\x35\x07\x00\x87\x29\x07\x00\x84\x41"
-     "\x00\x00"
-     "\x97\x0A\x00\x80\xA7\xED\x00\x00\x37\x1C\x00\x80\x88\xE9\x07\x00\x89\xCD"
-     "\x07\x00"
-     "\x9D\x0A\x00\x80\x8F\xC1\x07\x00\x8C\xDD\x07\x00\xA3\x0A\x00\x80\xFC\x0B"
-     "\x00\x80"
-     "\xA7\x5D\x06\x00\xB0\x9D\x00\x00\xA9\x0A\x00\x80\xAF\x0A\x00\x80\xA3\x41"
-     "\x06\x00"
-     "\x3F\x1C\x00\x80\x4F\x1C\x00\x80\x8F\x0C\x00\x80\x47\x1C\x00\x80\x57\x1C"
-     "\x00\x80"
-     "\xAD\x71\x06\x00\x5F\x1C\x00\x80\xC2\x09\x00\x80\xCC\xF8\x03\x00\xCD\x2C"
-     "\x03\x00"
-     "\xC5\x09\x00\x80\xA3\xE5\x00\x00\xC8\x09\x00\x80\x8C\x4D\x00\x00\xBB\x0A"
-     "\x00\x80"
-     "\xA7\xF1\x00\x00\xC1\x0A\x00\x80\xB1\x3D\x06\x00\x87\x9D\x00\x00\x86\x95"
-     "\x00\x00"
-     "\xA8\x1D\x07\x00\x84\x89\x00\x00\xC7\x0A\x00\x80\x82\xA9\x00\x00\x81\xD5"
-     "\x00\x00"
-     "\xAD\x01\x07\x00\xD3\x0A\x00\x80\x91\x3D\x00\x00\x82\x99\x01\x00\xCB\x09"
-     "\x00\x80"
-     "\xCD\x0C\x05\x00\xCC\x08\x05\x00\x81\x3D\x00\x00\x87\x85\x01\x00\x88\xBD"
-     "\x01\x00"
-     "\x6F\x1C\x00\x80\xBF\x1E\x00\x80\xC0\x0B\x00\x80\x8C\x91\x01\x00\x53\x0C"
-     "\x00\x80"
-     "\x59\x0C\x00\x80\xC7\x1E\x00\x80\x80\xF5\x01\x00\x81\x85\x01\x00\x82\x81"
-     "\x01\x00"
-     "\x83\x9D\x01\x00\x84\x89\x01\x00\xCD\x0A\x00\x80\x86\xA9\x01\x00\x87\x5D"
-     "\x00\x00"
-     "\x88\x6D\x00\x00\x89\x4D\x00\x00\x67\x1C\x00\x80\xCC\x10\x02\x00\x8C\x5D"
-     "\x00\x00"
-     "\x82\x0D\x00\x00\xD9\x0A\x00\x80\x8F\x49\x00\x00\xB0\xE9\x00\x00\x77\x1C"
-     "\x00\x80"
-     "\x02\x0C\x00\x80\xA3\x29\x01\x00\x80\x25\x01\x00\x81\x55\x01\x00\x7F\x1C"
-     "\x00\x80"
-     "\xA7\x35\x01\x00\xCC\xA4\x01\x00\xCD\x10\x02\x00\xDF\x0A\x00\x80\x87\x1C"
-     "\x00\x80"
-     "\x81\x35\x00\x00\xE5\x0A\x00\x80\xAE\x09\x01\x00\xF1\x0A\x00\x80\xCC\xE8"
-     "\x01\x00"
-     "\xCD\x28\x02\x00\x8F\x1C\x00\x80\xA3\xF1\x00\x00\x84\x05\x00\x00\x97\x1C"
-     "\x00\x80"
-     "\xEB\x0A\x00\x80\xA7\x1C\x00\x80\xA8\xB5\x00\x00\x9F\x1C\x00\x80\x81\x5D"
-     "\x00\x00"
-     "\xAF\x1C\x00\x80\xCC\xFC\x01\x00\xCD\xC0\x01\x00\xB7\x1C\x00\x80\xBF\x1C"
-     "\x00\x80"
-     "\x81\x85\x03\x00\x23\x0C\x00\x80\x81\xE5\x03\x00\xF7\x0A\x00\x80\x87\xE9"
-     "\x03\x00"
-     "\xCE\x09\x00\x80\x8C\xA5\x03\x00\xC7\x1C\x00\x80\x03\x0B\x00\x80\xAD\x09"
-     "\x00\x80"
-     "\xCF\x1C\x00\x80\x81\x99\x03\x00\x81\xDD\x03\x00\x8C\xBD\x03\x00\xCD\x24"
-     "\x01\x00"
-     "\xCC\x20\x01\x00\xCC\x10\x02\x00\xCD\x30\x02\x00\x81\xF9\x00\x00\x87\x51"
-     "\x00\x00"
-     "\x80\x55\x00\x00\x81\x59\x00\x00\xFD\x0A\x00\x80\x83\x49\x00\x00\x8C\x41"
-     "\x00\x00"
-     "\xDF\x1C\x00\x80\xD7\x1C\x00\x80\xD1\x09\x00\x80\x81\x7D\x00\x00\x80\x71"
-     "\x00\x00"
-     "\xCC\x20\x03\x00\xCD\xB0\x03\x00\xA3\x7D\x03\x00\xD4\x09\x00\x80\xA3\x11"
-     "\x03\x00"
-     "\xE7\x1E\x00\x80\x81\x2D\x01\x00\xB1\xFD\x00\x00\xA7\x31\x03\x00\xAD\x41"
-     "\x03\x00"
-     "\xDF\x1E\x00\x80\xA3\x6D\x03\x00\xF7\x1E\x00\x80\xEF\x1E\x00\x80\xE7\x1C"
-     "\x00\x80"
-     "\xA7\x6D\x03\x00\x80\x35\x00\x00\x81\x1D\x00\x00\xB1\x6D\x03\x00\x88\x2D"
-     "\x00\x00"
-     "\xAA\x0C\x00\x80\xA7\x95\x00\x00\x81\x71\x00\x00\x80\x59\x00\x00\x83\x71"
-     "\x00\x00"
-     "\xA3\xF5\x00\x00\x81\x51\x00\x00\xAD\x81\x00\x00\xEF\x1C\x00\x80\xB1\x09"
-     "\x03\x00"
-     "\x89\x5D\x00\x00\x84\x3D\x00\x00\xCC\x38\x01\x00\x84\x9D\x01\x00\x81\x19"
-     "\x00\x00"
-     "\x80\x09\x00\x00\x84\x65\x00\x00\xF7\x1C\x00\x80\xCC\xD0\x07\x00\xCC\xF0"
-     "\x07\x00"
-     "\xFF\x1C\x00\x80\x91\x89\x00\x00\xCC\x4C\x06\x00\xCD\x04\x06\x00\xCC\x70"
-     "\x06\x00"
-     "\xCD\x74\x06\x00\xCC\x40\x07\x00\x9B\x2D\x0F\x00\xCC\xA8\x07\x00\xCD\xAC"
-     "\x07\x00"
-     "\x86\x0D\x00\x00\x87\x55\x0F\x00\x84\x41\x0F\x00\x12\x0B\x00\x80\x82\x01"
-     "\x0C\x00"
-     "\x83\x55\x0F\x00\x80\x35\x01\x00\x81\xD9\x01\x00\xB6\x0C\x00\x80\x8F\xE9"
-     "\x00\x00"
-     "\x8C\xED\x00\x00\xA4\x0C\x00\x80\xD7\x1E\x00\x80\x8B\xF5\x00\x00\x88\x6D"
-     "\x0F\x00"
-     "\x89\x0D\x00\x00\x18\x0B\x00\x80\xC6\x0B\x00\x80\x82\x25\x00\x00\x5F\x0C"
-     "\x00\x80"
-     "\x81\x41\x00\x00\x65\x0C\x00\x80\x0F\x1F\x00\x80\x07\x1F\x00\x80\x1F\x1F"
-     "\x00\x80"
-     "\x17\x1F\x00\x80\x27\x1F\x00\x80\xFF\x1E\x00\x80\x80\x29\x00\x00\x81\x29"
-     "\x00\x00"
-     "\x0E\x0C\x00\x80\x07\x1D\x00\x80\x84\x79\x00\x00\x0F\x1D\x00\x80\x81\x4D"
-     "\x01\x00"
-     "\x80\xA1\x01\x00\x21\x0B\x00\x80\xA3\xFD\x0F\x00\xCC\x38\x02\x00\xCD\x48"
-     "\x03\x00"
-     "\x17\x1D\x00\x80\x81\x59\x00\x00\xCD\x7C\x02\x00\xCC\xA4\x0D\x00\x2D\x0B"
-     "\x00\x80"
-     "\x6B\x0C\x00\x80\xA8\xC9\x0F\x00\x87\x39\x00\x00\xDA\x09\x00\x80\x89\xA1"
-     "\x0F\x00"
-     "\x0C\x0B\x00\x80\x91\x11\x00\x00\xB0\x0C\x00\x80\xDD\x09\x00\x80\xA2\x0B"
-     "\x00\x80"
-     "\x71\x0C\x00\x80\x80\xB9\x0F\x00\x81\xB9\x0F\x00\xCF\x1E\x00\x80\x83\x8D"
-     "\x0F\x00"
-     "\x08\x0C\x00\x80\x1F\x1D\x00\x80\x84\x05\x00\x00\x27\x1D\x00\x80\x27\x0B"
-     "\x00\x80"
-     "\x2F\x1D\x00\x80\x33\x0B\x00\x80\x81\x9D\x0F\x00\x87\x21\x00\x00\x87\xB5"
-     "\x0F\x00"
-     "\xCC\xA8\x02\x00\xCD\xE8\x02\x00\xCC\xB4\x0C\x00\xCD\xDC\x0C\x00\xA6\xCD"
-     "\x00\x00"
-     "\xA7\xC5\x00\x00\x47\x1D\x00\x80\x8F\x81\x0F\x00\x8C\x89\x0F\x00\xA3\xE5"
-     "\x00\x00"
-     "\x39\x0B\x00\x80\x37\x1D\x00\x80\x45\x0B\x00\x80\xB1\xC9\x00\x00\x87\x05"
-     "\x00\x00"
-     "\x4F\x1D\x00\x80\x3F\x1D\x00\x80\x87\x3D\x00\x00\x57\x1D\x00\x80\x83\x0C"
-     "\x00\x80"
-     "\x3F\x0B\x00\x80\xA3\x05\x0F\x00\x81\xF9\x00\x00\xCC\xA8\x03\x00\x5F\x1D"
-     "\x00\x80"
-     "\x51\x0B\x00\x80\x8C\x49\x00\x00\xA3\xF1\x00\x00\x67\x1D\x00\x80\x89\x0C"
-     "\x00\x80"
-     "\x4B\x0B\x00\x80\xA7\x95\x00\x00\x77\x1D\x00\x80\x6F\x1D\x00\x80\xCC\xAC"
-     "\x03\x00"
-     "\xCD\xC8\x00\x00\xE0\x09\x00\x80\x87\x69\x00\x00\xE3\x09\x00\x80\x81\xBD"
-     "\x00\x00"
-     "\x82\x79\x00\x00\xE6\x09\x00\x80\x87\xB9\x01\x00\x57\x0B\x00\x80\x91\xA5"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x97\x1D\x00\x80\x5D\x0B\x00\x80\x88\x05\x00\x00\x9F\x1D"
-     "\x00\x80"
-     "\x9B\x91\x00\x00\x63\x0B\x00\x80\xE9\x09\x00\x80\x8C\x91\x01\x00\xE4\x0B"
-     "\x00\x80"
-     "\xD8\x0B\x00\x80\xD2\x0B\x00\x80\xDE\x0B\x00\x80\x83\x45\x00\x00\x82\xB9"
-     "\x01\x00"
-     "\x81\xB9\x01\x00\x80\xA5\x01\x00\x37\x1F\x00\x80\x86\x71\x00\x00\x69\x0B"
-     "\x00\x80"
-     "\x84\x49\x00\x00\x8B\x15\x00\x00\x8A\x3D\x00\x00\x89\x39\x00\x00\x88\x45"
-     "\x00\x00"
-     "\x8F\xF9\x00\x00\x6F\x0B\x00\x80\xCC\x0B\x00\x80\x8C\x05\x00\x00\xA7\x51"
-     "\x01\x00"
-     "\xA6\x49\x01\x00\x77\x0C\x00\x80\xB0\x79\x00\x00\xA3\x59\x01\x00\x8C\xA9"
-     "\x00\x00"
-     "\x80\xA9\x00\x00\x81\xA9\x00\x00\x81\x95\x00\x00\x80\x95\x00\x00\xAD\x71"
-     "\x01\x00"
-     "\x7D\x0C\x00\x80\xAB\x0B\x00\x80\x84\x8D\x00\x00\x47\x1F\x00\x80\x3F\x1F"
-     "\x00\x80"
-     "\xA3\x21\x00\x00\x57\x1F\x00\x80\x4F\x1F\x00\x80\x5F\x1F\x00\x80\x81\x6D"
-     "\x00\x00"
-     "\x80\x6D\x00\x00\xB1\x05\x01\x00\xA4\x39\x00\x00\x2F\x1F\x00\x80\x7F\x1D"
-     "\x00\x80"
-     "\x75\x0B\x00\x80\xA8\x05\x00\x00\x8F\x1D\x00\x80\x87\x1D\x00\x80\xAD\x09"
-     "\x00\x00"
-     "\xCC\xB0\x01\x00\x81\xBD\x03\x00\x80\xBD\x03\x00\x83\xCD\x03\x00\xA7\x1D"
-     "\x00\x80"
-     "\xAF\x1D\x00\x80\xB7\x1D\x00\x80\xCC\xBC\x01\x00\xCD\x84\x01\x00\x89\xE9"
-     "\x03\x00"
-     "\xCC\x1C\x01\x00\x81\xD9\x02\x00\x80\xC5\x02\x00\xCD\x38\x01\x00\xCC\x3C"
-     "\x01\x00"
-     "\xCC\x68\x02\x00\xCD\x44\x02\x00\x83\x4D\x00\x00\xBF\x1D\x00\x80\x87\xD9"
-     "\x00\x00"
-     "\x87\x2D\x00\x00\x80\x45\x00\x00\x81\x11\x00\x00\x82\x05\x00\x00\x7B\x0B"
-     "\x00\x80"
-     "\xCF\x1D\x00\x80\xC7\x1D\x00\x80\xD7\x1D\x00\x80\xCC\x38\x02\x00\x88\x15"
-     "\x00\x00"
-     "\x88\xE1\x00\x00\x80\x6D\x00\x00\x81\x39\x00\x00\xCC\x84\x02\x00\xCD\x50"
-     "\x01\x00"
-     "\xA3\x45\x03\x00\x84\x39\x01\x00\xDF\x1D\x00\x80\xE7\x1D\x00\x80\xCC\xDC"
-     "\x03\x00"
-     "\xCD\x48\x02\x00\x67\x1F\x00\x80\xEC\x09\x00\x80\x81\x0B\x00\x80\x7F\x1F"
-     "\x00\x80"
-     "\xBC\x0C\x00\x80\x81\x6D\x00\x00\xEF\x1D\x00\x80\x87\x0B\x00\x80\xA3\xD1"
-     "\x00\x00"
-     "\x77\x1F\x00\x80\x6F\x1F\x00\x80\xCC\x88\x04\x00\x81\x75\x00\x00\x80\x75"
-     "\x00\x00"
-     "\x8A\x0B\x00\x80\xA3\xB5\x00\x00\xCC\x00\x04\x00\xCD\x54\x02\x00\xF7\x1D"
-     "\x00\x80"
-     "\x90\x0B\x00\x80\x84\x4D\x01\x00\x96\x0B\x00\x80\xFF\x1D\x00\x80\x07\x1E"
-     "\x00\x80"
-     "\xCC\xD0\x0E\x00\xCC\x2C\x01\x00\xCC\x00\x05\x00\xCD\x5C\x05\x00\xEF\x09"
-     "\x00\x80"
-     "\xF2\x09\x00\x80\xCC\x98\x0E\x00\x81\xC1\x00\x00\xCC\xCC\x0F\x00\xCC\x3C"
-     "\x0E\x00"
-     "\xCC\x08\x01\x00\xCD\x9C\x0E\x00\xCC\xD4\x0F\x00\xCD\x78\x0F\x00\xCC\x3C"
-     "\x0E\x00"
-     "\xCD\x38\x0E\x00\x81\xE5\x01\x00\x80\xE5\x01\x00\x83\xE5\x01\x00\x82\xE5"
-     "\x01\x00"
-     "\xD7\x09\x00\x80\x84\xE5\x01\x00\x87\xE1\x01\x00\x3B\x1E\x00\x80\x89\xA5"
-     "\x01\x00"
-     "\x88\xD9\x01\x00\x81\xC9\x07\x00\x33\x1E\x00\x80\x4B\x1E\x00\x80\x43\x1E"
-     "\x00\x80"
-     "\xCC\x34\x01\x00\xF8\x09\x00\x80\x80\xDD\x00\x00\x81\xE9\x00\x00\x4C\x0A"
-     "\x00\x80"
-     "\x83\xFD\x00\x00\x80\xCD\x00\x00\x81\xF9\x00\x00\x81\x11\x07\x00\x63\x1E"
-     "\x00\x80"
-     "\x5B\x1E\x00\x80\x89\xD1\x00\x00\xCC\x28\x01\x00\x73\x1E\x00\x80\x6B\x1E"
-     "\x00\x80"
-     "\xF3\x0B\x00\x80\xCC\x34\x01\x00\xED\x0B\x00\x80\x80\x5D\x00\x00\x81\x65"
-     "\x00\x00"
-     "\xA3\x01\x01\x00\x83\x61\x00\x00\x81\x71\x00\x00\x80\x49\x00\x00\x2B\x1E"
-     "\x00\x80"
-     "\x2C\x0C\x00\x80\xBD\x0B\x00\x80\x89\x55\x00\x00\x3E\x0C\x00\x80\x44\x0C"
-     "\x00\x80"
-     "\x53\x1E\x00\x80\x7B\x1E\x00\x80\xBB\x1B\x00\x80\x58\x0A\x00\x80\x82\x1D"
-     "\x00\x00"
-     "\x83\x79\x07\x00\x80\x19\x07\x00\x81\x19\x07\x00\x86\x21\x00\x00\x87\x29"
-     "\x00\x00"
-     "\x84\x91\x07\x00\xF5\x09\x00\x80\x8A\x69\x00\x00\xB1\xD9\x06\x00\x88\x69"
-     "\x00\x00"
-     "\x89\xF5\x07\x00\x52\x0A\x00\x80\x8F\xDD\x07\x00\x8C\xD9\x07\x00\x9B\x0C"
-     "\x00\x80"
-     "\xFB\x09\x00\x80\x23\x1E\x00\x80\xFE\x09\x00\x80\x91\xA1\x07\x00\x80\x41"
-     "\x07\x00"
-     "\x81\x41\x07\x00\x87\x05\x00\x00\xC3\x1B\x00\x80\x82\x91\x07\x00\xCB\x1B"
-     "\x00\x80"
-     "\xD3\x1B\x00\x80\xA3\x95\x06\x00\x86\x85\x07\x00\xA7\xED\x00\x00\xCC\x90"
-     "\x02\x00"
-     "\xCD\xE0\x05\x00\xB1\xE9\x00\x00\xA3\xC1\x00\x00\x5E\x0A\x00\x80\x64\x0A"
-     "\x00\x80"
-     "\x6A\x0A\x00\x80\x70\x0A\x00\x80\x04\x0A\x00\x80\xA5\x65\x07\x00\xDB\x1B"
-     "\x00\x80"
-     "\xCC\xB8\x03\x00\xA8\x55\x07\x00\xE3\x1B\x00\x80\x76\x0A\x00\x80\xEB\x1B"
-     "\x00\x80"
-     "\xFB\x1B\x00\x80\x03\x1C\x00\x80\xF3\x1B\x00\x80\x07\x0A\x00\x80\xA3\xAD"
-     "\x00\x00"
-     "\x0A\x0A\x00\x80\x8C\x25\x06\x00\x0D\x0A\x00\x80\x8C\x4D\x00\x00\xA3\x1E"
-     "\x00\x80"
-     "\x82\x6D\x00\x00\x81\x3D\x06\x00\x82\x01\x06\x00\x81\x65\x00\x00\x9B\x1E"
-     "\x00\x80"
-     "\x87\x65\x00\x00\xB3\x1E\x00\x80\x87\x11\x06\x00\x87\xAD\x01\x00\xAB\x1E"
-     "\x00\x80"
-     "\xCC\x50\x02\x00\xCD\xC4\x02\x00\x81\xE1\x01\x00\x80\xC9\x01\x00\x83\xE1"
-     "\x01\x00"
-     "\x91\x89\x00\x00\x80\xFD\x01\x00\x81\xD5\x01\x00\x93\x1E\x00\x80\x8C\x9D"
-     "\x01\x00"
-     "\x89\x35\x00\x00\x7C\x0A\x00\x80\x80\x75\x00\x00\x81\x5D\x00\x00\x86\x2D"
-     "\x00\x00"
-     "\x87\x35\x00\x00\x84\x7D\x00\x00\x0B\x1C\x00\x80\x82\x85\x01\x00\x83\x7D"
-     "\x00\x00"
-     "\x80\x9D\x01\x00\x81\x91\x01\x00\x13\x1C\x00\x80\x8F\xE9\x00\x00\x8C\xE1"
-     "\x00\x00"
-     "\x82\x0A\x00\x80\x88\x0A\x00\x80\x10\x0A\x00\x80\x88\x0D\x00\x00\x89\xF9"
-     "\x00\x00"
-     "\xA7\x39\x01\x00\x8B\x1E\x00\x80\x94\x0A\x00\x80\x4A\x0C\x00\x80\xA3\x25"
-     "\x01\x00"
-     "\x50\x0C\x00\x80\xB0\x59\x00\x00\x83\x1E\x00\x80\x82\x05\x00\x00\xBB\x1E"
-     "\x00\x80"
-     "\xAD\x15\x01\x00\xA1\x0C\x00\x80\x2B\x1C\x00\x80\x86\x05\x00\x00\x8E\x0A"
-     "\x00\x80"
-     "\x1B\x1C\x00\x80\x23\x1C\x00\x80\xA7\x69\x00\x00\x80\x0D\x01\x00\x81\x01"
-     "\x01\x00"
-     "\x87\x31\x00\x00\xA3\x49\x00\x00\xB1\x19\x01\x00\xCC\x10\x02\x00\x33\x1C"
-     "\x00\x80"
-     "\x20\x0C\x00\x80\x9A\x0A\x00\x80\xAD\x51\x00\x00\xCC\xD4\x01\x00\xCD\xF8"
-     "\x01\x00"
-     "\xA8\x41\x00\x00\x3B\x1C\x00\x80\xCD\x38\x01\x00\xCC\x3C\x01\x00\x81\xED"
-     "\x03\x00"
-     "\xA0\x0A\x00\x80\xA6\x0A\x00\x80\x8C\x0D\x00\x00\xFF\x0B\x00\x80\xAC\x0A"
-     "\x00\x80"
-     "\x81\xC5\x03\x00\xCC\x68\x02\x00\xB2\x0A\x00\x80\x82\xC5\x03\x00\x43\x1C"
-     "\x00\x80"
-     "\x84\xC9\x03\x00\x87\x29\x00\x00\x86\x31\x00\x00\x53\x1C\x00\x80\x82\x6D"
-     "\x00\x00"
-     "\x92\x0C\x00\x80\x4B\x1C\x00\x80\x87\x61\x00\x00\x5B\x1C\x00\x80\x63\x1C"
-     "\x00\x80"
-     "\x0F\x1E\x00\x80\xCC\xA8\x03\x00\xCD\xAC\x02\x00\x81\xF9\x00\x00\x88\x2D"
-     "\x00\x00"
-     "\x13\x0A\x00\x80\x16\x0A\x00\x80\x19\x0A\x00\x80\x8C\x35\x00\x00\xBE\x0A"
-     "\x00\x80"
-     "\xC4\x0A\x00\x80\xB1\xD5\x00\x00\xCA\x0A\x00\x80\x73\x1C\x00\x80\xC3\x1E"
-     "\x00\x80"
-     "\xC3\x0B\x00\x80\xCC\x30\x01\x00\x56\x0C\x00\x80\x5C\x0C\x00\x80\xCB\x1E"
-     "\x00\x80"
-     "\xCC\x04\x01\x00\xD0\x0A\x00\x80\x6B\x1C\x00\x80\xA7\xA5\x00\x00\xDC\x0A"
-     "\x00\x80"
-     "\xA3\x8D\x00\x00\xCC\x14\x02\x00\x80\xB9\x00\x00\x81\xB9\x00\x00\xA7\x85"
-     "\x00\x00"
-     "\x1A\x0C\x00\x80\x82\x65\x00\x00\x7B\x1C\x00\x80\x8C\x35\x00\x00\x05\x0C"
-     "\x00\x80"
-     "\xCC\xEC\x1C\x00\xCD\xFC\x03\x00\x83\x1C\x00\x80\xAE\xAD\x00\x00\xE2\x0A"
-     "\x00\x80"
-     "\x8B\x1C\x00\x80\xCC\xC0\x06\x00\xCD\xD0\x06\x00\xB0\xBD\x01\x00\xCC\x90"
-     "\x07\x00"
-     "\xE8\x0A\x00\x80\x81\xC9\x01\x00\xCC\x18\x1D\x00\xCD\x20\x02\x00\x84\x11"
-     "\x00\x00"
-     "\xF4\x0A\x00\x80\xCD\xB8\x06\x00\xCC\xAC\x06\x00\x9B\x1C\x00\x80\xEE\x0A"
-     "\x00\x80"
-     "\x81\x29\x00\x00\xAB\x1C\x00\x80\xA3\x1C\x00\x80\xA3\xED\x01\x00\xCC\x40"
-     "\x1D\x00"
-     "\xCD\x10\x02\x00\xB3\x1C\x00\x80\xBB\x1C\x00\x80\x81\x09\x00\x00\xC3\x1C"
-     "\x00\x80"
-     "\xCC\x40\x1D\x00\xCD\xD0\x02\x00\xA8\xD9\x01\x00\x26\x0C\x00\x80\xCC\x90"
-     "\x07\x00"
-     "\xCC\x1C\x01\x00\xCC\x60\x06\x00\xCD\x64\x06\x00\xFA\x0A\x00\x80\x22\x0A"
-     "\x00\x80"
-     "\xCB\x1C\x00\x80\x91\x29\x01\x00\x06\x0B\x00\x80\x81\xCD\x1F\x00\xD3\x1C"
-     "\x00\x80"
-     "\x00\x0B\x00\x80\xE3\x1C\x00\x80\xDB\x1C\x00\x80\xCC\xC4\x06\x00\xCD\xC0"
-     "\x06\x00"
-     "\x81\x31\x00\x00\x80\xD9\x00\x00\x25\x0A\x00\x80\x28\x0A\x00\x80\x82\xB9"
-     "\x01\x00"
-     "\x83\x45\x01\x00\x80\xB9\x01\x00\x81\xB9\x01\x00\x86\x5D\x01\x00\xEB\x1E"
-     "\x00\x80"
-     "\x84\x5D\x01\x00\xE3\x1E\x00\x80\xCD\xC0\x00\x00\xCC\xF0\x00\x00\x88\x01"
-     "\x1C\x00"
-     "\x89\x79\x01\x00\xFB\x1E\x00\x80\x8F\x55\x01\x00\x8C\x61\x01\x00\xF3\x1E"
-     "\x00\x80"
-     "\x81\xDD\x1E\x00\x81\x15\x1F\x00\x93\x1C\x00\x80\x81\x5D\x1F\x00\x8C\x81"
-     "\x1F\x00"
-     "\x87\x41\x1F\x00\xCC\x18\x03\x00\xCD\x68\x03\x00\x80\x4D\x1F\x00\x81\xA5"
-     "\x1F\x00"
-     "\x2B\x0A\x00\x80\x83\xA9\x1F\x00\x8C\x15\x1F\x00\x8C\xD1\x1E\x00\x2E\x0A"
-     "\x00\x80"
-     "\x87\xB5\x1F\x00\x80\x95\x1F\x00\x81\x99\x1F\x00\x81\x11\x00\x00\x83\xBD"
-     "\x1F\x00"
-     "\x80\x85\x1F\x00\x81\x89\x1F\x00\xEB\x1C\x00\x80\x84\x3D\x00\x00\xAD\x0C"
-     "\x00\x80"
-     "\x89\x99\x1F\x00\xF3\x1C\x00\x80\x88\x05\x00\x00\x0F\x0B\x00\x80\xFB\x1C"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\xFD\x00\x00\x03\x1D\x00\x80\xA3\xD9\x1F\x00\xA3\x79"
-     "\x1E\x00"
-     "\xA3\x85\x00\x00\x15\x0B\x00\x80\xAD\x35\x1F\x00\xA7\x61\x1E\x00\xA7\xA9"
-     "\x1F\x00"
-     "\xB3\x0C\x00\x80\x84\x0D\x00\x00\xB9\x0C\x00\x80\xA3\x35\x1F\x00\x31\x0A"
-     "\x00\x80"
-     "\xAD\x89\x1F\x00\x84\x71\x00\x00\xA7\x21\x1F\x00\xB1\x3D\x1E\x00\xB1\x85"
-     "\x1F\x00"
-     "\xA7\x0C\x00\x80\xDB\x1E\x00\x80\x1B\x0B\x00\x80\xC9\x0B\x00\x80\xCC\xB4"
-     "\x1C\x00"
-     "\xCD\xB0\x1C\x00\x62\x0C\x00\x80\xB1\x41\x1F\x00\x68\x0C\x00\x80\xA5\x0B"
-     "\x00\x80"
-     "\x13\x1F\x00\x80\x0B\x1F\x00\x80\x23\x1F\x00\x80\x1B\x1F\x00\x80\x80\xB9"
-     "\x1E\x00"
-     "\x81\xB9\x1E\x00\x82\x21\x01\x00\x83\x35\x01\x00\x84\x61\x01\x00\x2B\x1F"
-     "\x00\x80"
-     "\x86\x89\x01\x00\x87\xBD\x01\x00\x88\x91\x01\x00\x89\xE9\x01\x00\xD3\x1E"
-     "\x00\x80"
-     "\x8B\xFD\x01\x00\x8C\xE5\x01\x00\x82\x0D\x00\x00\x03\x1F\x00\x80\x8F\xDD"
-     "\x01\x00"
-     "\x83\xB9\x01\x00\x91\xAD\x01\x00\x81\xBD\x01\x00\x80\xBD\x01\x00\x80\xA1"
-     "\x01\x00"
-     "\x81\xA1\x01\x00\x0B\x0C\x00\x80\x11\x0C\x00\x80\x84\x3D\x00\x00\x0B\x1D"
-     "\x00\x80"
-     "\x89\x95\x01\x00\x9B\x81\x01\x00\x81\xCD\x1E\x00\x80\xCD\x1E\x00\xCC\xFC"
-     "\x02\x00"
-     "\xCD\xF0\x02\x00\x81\xE5\x00\x00\x13\x1D\x00\x80\x81\xED\x00\x00\xA3\xA5"
-     "\x00\x00"
-     "\xCC\x90\x01\x00\xCD\x5C\x02\x00\x87\x1D\x00\x00\x24\x0B\x00\x80\xA8\xF9"
-     "\x00\x00"
-     "\x1B\x1D\x00\x80\x30\x0B\x00\x80\x6E\x0C\x00\x80\x74\x0C\x00\x80\x23\x1D"
-     "\x00\x80"
-     "\x84\x05\x00\x00\x2B\x1D\x00\x80\xA3\xD5\x00\x00\x2A\x0B\x00\x80\x33\x1D"
-     "\x00\x80"
-     "\x81\x51\x00\x00\xCC\xF4\x01\x00\xCD\xD0\x01\x00\x36\x0B\x00\x80\x87\x3D"
-     "\x00\x00"
-     "\x4B\x1D\x00\x80\x3C\x0B\x00\x80\x3B\x1D\x00\x80\x48\x0B\x00\x80\x87\x05"
-     "\x00\x00"
-     "\x53\x1D\x00\x80\x43\x1D\x00\x80\x87\xF1\x03\x00\x81\xD9\x03\x00\x81\x99"
-     "\x03\x00"
-     "\x81\x91\x00\x00\x5B\x1D\x00\x80\x86\x0C\x00\x80\x8C\xF9\x03\x00\xCC\x24"
-     "\x01\x00"
-     "\x87\xB9\x03\x00\x81\xF9\x03\x00\x42\x0B\x00\x80\xCC\x64\x02\x00\x82\xC9"
-     "\x03\x00"
-     "\x8C\x99\x03\x00\x63\x1D\x00\x80\x87\xD1\x03\x00\x8F\x91\x03\x00\x81\xDD"
-     "\x06\x00"
-     "\x91\xF5\x03\x00\xCC\x00\x04\x00\xCD\xEC\x03\x00\x87\x65\x00\x00\x13\x1E"
-     "\x00\x80"
-     "\x54\x0B\x00\x80\x6B\x1D\x00\x80\x8C\x0C\x00\x80\x4E\x0B\x00\x80\xCC\x18"
-     "\x01\x00"
-     "\x88\x39\x00\x00\x7B\x1D\x00\x80\x73\x1D\x00\x80\xCC\x5C\x03\x00\x8C\x25"
-     "\x00\x00"
-     "\x37\x0A\x00\x80\xCC\x2C\x01\x00\xB1\xFD\x00\x00\xA3\x39\x03\x00\x3A\x0A"
-     "\x00\x80"
-     "\x3D\x0A\x00\x80\x9B\x1D\x00\x80\xA7\x59\x03\x00\xCC\x74\x03\x00\x88\x09"
-     "\x00\x00"
-     "\xA3\x51\x03\x00\xA3\x1D\x00\x80\x66\x0B\x00\x80\x83\x6D\x0D\x00\xA7\x9D"
-     "\x00\x00"
-     "\xA6\xAD\x00\x00\xA3\x9D\x00\x00\xB1\x0D\x03\x00\xCC\x28\x01\x00\xE7\x0B"
-     "\x00\x80"
-     "\xA7\xB5\x00\x00\xA6\xB5\x00\x00\xDB\x0B\x00\x80\xCC\x30\x01\x00\x81\xD5"
-     "\x07\x00"
-     "\xD5\x0B\x00\x80\xCC\x28\x01\x00\xE1\x0B\x00\x80\x3B\x1F\x00\x80\x6C\x0B"
-     "\x00\x80"
-     "\xAD\x89\x00\x00\x72\x0B\x00\x80\x80\xCD\x01\x00\x81\xDD\x01\x00\xCC\x44"
-     "\x01\x00"
-     "\xCD\x9C\x1E\x00\x84\xF5\x01\x00\xCF\x0B\x00\x80\xCC\x58\x01\x00\xCD\x4C"
-     "\x01\x00"
-     "\x80\xED\x01\x00\x81\xFD\x01\x00\x83\xB5\x00\x00\x7A\x0C\x00\x80\x8C\xDD"
-     "\x01\x00"
-     "\x80\x0C\x00\x80\xCC\x08\x1E\x00\x8C\xF1\x06\x00\xCC\x38\x01\x00\xCD\x3C"
-     "\x01\x00"
-     "\x4B\x1F\x00\x80\x89\x11\x00\x00\x81\x05\x06\x00\x43\x1F\x00\x80\x5B\x1F"
-     "\x00\x80"
-     "\x53\x1F\x00\x80\x63\x1F\x00\x80\x83\x3D\x00\x00\x80\x21\x00\x00\x81\x39"
-     "\x00\x00"
-     "\x80\x39\x00\x00\x81\x21\x00\x00\x33\x1F\x00\x80\x83\x1D\x00\x80\xCC\x28"
-     "\x01\x00"
-     "\x81\xD9\x06\x00\x78\x0B\x00\x80\x81\xFD\x06\x00\xCC\x24\x01\x00\x93\x1D"
-     "\x00\x80"
-     "\x8B\x1D\x00\x80\xAB\x1D\x00\x80\x80\x21\x01\x00\x81\x35\x01\x00\xA3\x05"
-     "\x00\x00"
-     "\xB3\x1D\x00\x80\xBB\x1D\x00\x80\xC3\x1D\x00\x80\xCC\x8C\x05\x00\xCD\x40"
-     "\x02\x00"
-     "\xB7\x1C\x03\x00\x7E\x0B\x00\x80\x87\xCD\x07\x00\xD3\x1D\x00\x80\xCB\x1D"
-     "\x00\x80"
-     "\x17\x1E\x00\x80\xCD\x88\x00\x00\xCC\x90\x00\x00\x8C\xDD\x05\x00\xA3\x85"
-     "\x00\x00"
-     "\x1C\x0A\x00\x80\xCC\xE0\x02\x00\xDB\x1D\x00\x80\x88\xD5\x07\x00\x81\x4D"
-     "\x00\x00"
-     "\x80\x4D\x00\x00\x5A\x0B\x00\x80\xE3\x1D\x00\x80\x60\x0B\x00\x80\x91\x39"
-     "\x07\x00"
-     "\x40\x0A\x00\x80\x88\xC5\x07\x00\xAE\x0B\x00\x80\x8A\xC9\x07\x00\xEB\x1D"
-     "\x00\x80"
-     "\x9B\x3D\x00\x00\x87\xE5\x07\x00\x6B\x1F\x00\x80\x81\x85\x07\x00\x80\x85"
-     "\x07\x00"
-     "\x43\x0A\x00\x80\x82\xF9\x07\x00\x82\xD5\x06\x00\x83\x45\x00\x00\x80\xC9"
-     "\x06\x00"
-     "\x81\xDD\x06\x00\x86\xE1\x06\x00\x84\x0B\x00\x80\x84\x51\x00\x00\x83\x1F"
-     "\x00\x80"
-     "\x8A\x95\x06\x00\x8B\x99\x06\x00\x88\x79\x00\x00\x89\x9D\x06\x00\xBF\x0C"
-     "\x00\x80"
-     "\x8F\x59\x07\x00\x8C\x6D\x07\x00\xF3\x1D\x00\x80\xCC\x80\x03\x00\xCD\x24"
-     "\x02\x00"
-     "\xB0\x11\x07\x00\x46\x0A\x00\x80\x7B\x1F\x00\x80\x1B\x1E\x00\x80\x73\x1F"
-     "\x00\x80"
-     "\x8D\x0B\x00\x80\x80\x8D\x00\x00\x81\x9D\x00\x00\xCC\xEC\x03\x00\xCD\xE8"
-     "\x04\x00"
-     "\xFB\x1D\x00\x80\x93\x0B\x00\x80\xA3\x49\x07\x00\x99\x0B\x00\x80\x03\x1E"
-     "\x00\x80"
-     "\xA3\xBD\x07\x00\x0B\x1E\x00\x80\x1B\x00\x00\x80\xEA\x07\x00\x80\x0B\x00"
-     "\x00\x80"
-     "\xA4\xA5\x07\x00\xEB\x04\x00\x80\x8A\x05\x00\x80\x03\x00\x00\x80\xA8\x61"
-     "\x07\x00"
-     "\x60\x0E\x00\x80\x65\x00\x00\x80\xC8\x03\x00\x80\x1E\x09\x00\x80\xAD\x69"
-     "\x07\x00"
-     "\x80\x2D\x01\x00\x81\x3D\x01\x00\x82\x5D\x01\x00\x83\x51\x01\x00\x84\x61"
-     "\x01\x00"
-     "\xB8\x04\x00\x80\xAC\x04\x00\x80\x87\x61\x01\x00\x88\xAD\x01\x00\x89\xB5"
-     "\x01\x00"
-     "\x8A\xBD\x01\x00\x8F\x29\x15\x00\xBC\x05\x00\x80\x2F\x0C\x00\x80\xCC\x78"
-     "\x02\x00"
-     "\xCD\xD8\x05\x00\x81\xDD\x01\x00\x81\x71\x00\x00\xF6\x0B\x00\x80\x82\xFD"
-     "\x01\x00"
-     "\x84\x19\x00\x00\x35\x0C\x00\x80\x87\xED\x01\x00\x32\x0C\x00\x80\xCC\x34"
-     "\x04\x00"
-     "\xCD\x30\x04\x00\xF9\x0B\x00\x80\x9F\x69\x15\x00\x38\x0C\x00\x80\x8C\xC9"
-     "\x01\x00"
-     "\xCD\xF8\x04\x00\xCC\xF0\x02\x00\xB1\x49\x01\x00\x21\x07\x00\x80\x81\xD5"
-     "\x00\x00"
-     "\xA3\x15\x01\x00\xA0\x99\x15\x00\x76\x08\x00\x80\x46\x07\x00\x80\x84\xF5"
-     "\x00\x00"
-     "\xCC\x28\x04\x00\xCD\x2C\x04\x00\xC6\x08\x00\x80\xAF\x79\x01\x00\xA8\x7D"
-     "\x01\x00"
-     "\xB2\x0D\x00\x80\xAA\x69\x01\x00\x55\x09\x00\x80\xB4\x25\x01\x00\xB5\x29"
-     "\x01\x00"
-     "\xA3\x09\x01\x00\x14\x0C\x00\x80\xEA\x06\x00\x80\xEE\x06\x00\x80\xB2\x05"
-     "\x01\x00"
-     "\xB3\x3D\x01\x00\xBC\xF5\x00\x00\xBD\xFD\x00\x00\xBE\xD9\x00\x00\x3B\x08"
-     "\x00\x80"
-     "\xB8\x01\x01\x00\xB9\x01\x01\x00\xBA\x01\x01\x00\x3C\x07\x00\x80\x43\x07"
-     "\x00\x80"
-     "\x86\x0C\x00\x00\xB3\x9D\x03\x00\xB2\x89\x03\x00\xB6\x08\x00\x80\x80\xBD"
-     "\x03\x00"
-     "\x6C\x07\x00\x80\x6F\x07\x00\x80\x15\x09\x00\x80\xE4\x06\x00\x80\xE7\x06"
-     "\x00\x80"
-     "\x38\x08\x00\x80\x89\x85\x03\x00\xCC\xE4\x07\x00\xBF\xA1\x03\x00\x17\x0C"
-     "\x00\x80"
-     "\x01\x0D\x00\x80\x8C\x65\x00\x00\xCD\xE4\x0C\x00\xCC\x24\x0C\x00\x89\x41"
-     "\x00\x00"
-     "\x88\x55\x00\x00\x8B\x45\x00\x00\x8A\x45\x00\x00\x85\xB5\x03\x00\x84\xB5"
-     "\x03\x00"
-     "\x87\x95\x03\x00\x86\x81\x03\x00\x3A\x0D\x00\x80\x40\x0D\x00\x80\x4C\x0D"
-     "\x00\x80"
-     "\x98\x2C\x00\x00\x13\x00\x00\x80\xA6\xC8\x00\x00\xCD\x8C\x06\x00\xCC\xA8"
-     "\x06\x00"
-     "\x85\x69\x00\x00\x17\x00\x00\x80\x31\x00\x00\x80\x69\x00\x00\x80\xCC\xF0"
-     "\x03\x00"
-     "\x07\x00\x00\x80\x35\x00\x00\x80\xF8\x0C\x00\x80\xB1\x95\x00\x00\x79\x0D"
-     "\x00\x80"
-     "\xB3\x95\x00\x00\xB2\x95\x00\x00\xBC\x0D\x00\x80\xBF\x0D\x00\x80\xC7\x0D"
-     "\x00\x80"
-     "\xC2\x0D\x00\x80\xAF\x0D\x00\x80\x75\x00\x00\x80\xA6\x06\x00\x80\x25\x00"
-     "\x00\x80"
-     "\x9B\x09\x00\x80\x17\x22\x00\x80\xBF\x55\x03\x00\xCA\x0D\x00\x80\x13\x22"
-     "\x00\x80"
-     "\x0F\x22\x00\x80\x5B\x21\x00\x80\xB8\x6C\x00\x00\x94\x65\x0D\x00\x92\x00"
-     "\x02\x00"
-     "\x9C\xAD\x01\x00\x9D\xA5\x01\x00\x9A\x89\x01\x00\x9B\x89\x01\x00\x98\x99"
-     "\x01\x00"
-     "\x99\x89\x01\x00\xCC\x20\x06\x00\xCD\x04\x06\x00\xCC\x40\x06\x00\xCD\x5C"
-     "\x06\x00"
-     "\xCC\x3C\x07\x00\xCD\x38\x07\x00\xCC\xBC\x07\x00\x85\x75\x00\x00\x80\x01"
-     "\x0F\x00"
-     "\x81\x0D\x0F\x00\x63\x21\x00\x80\xBA\x99\x01\x00\x85\x05\x00\x00\x6B\x21"
-     "\x00\x80"
-     "\x53\x21\x00\x80\xBE\x85\x01\x00\x81\x29\x0F\x00\x80\x25\x0F\x00\x5F\x21"
-     "\x00\x80"
-     "\x82\x21\x0F\x00\x85\x29\x00\x00\xB4\xA5\x01\x00\x85\x11\x00\x00\x67\x21"
-     "\x00\x80"
-     "\xB3\x89\x0F\x00\xB2\x85\x0F\x00\xB1\xC9\x01\x00\xB0\x01\x0C\x00\xB7\x81"
-     "\x0F\x00"
-     "\xB6\xED\x01\x00\xB5\xED\x01\x00\xB4\xED\x01\x00\x81\x65\x01\x00\x80\x65"
-     "\x01\x00"
-     "\x83\x61\x01\x00\xB8\xB5\x0F\x00\xCC\x3C\x0B\x00\x84\x79\x01\x00\x80\xE1"
-     "\x0F\x00"
-     "\x81\xDD\x0F\x00\x6F\x21\x00\x80\x57\x21\x00\x80\xCC\xC8\x04\x00\xCD\xB8"
-     "\x04\x00"
-     "\x85\xAD\x00\x00\x85\x15\x00\x00\x1B\x22\x00\x80\x33\x22\x00\x80\xCC\xE8"
-     "\x19\x00"
-     "\xCD\xB4\x19\x00\xA4\x5D\x01\x00\xCD\x0D\x00\x80\xA2\x4D\x02\x00\xA3\xF1"
-     "\x0F\x00"
-     "\xA0\x55\x01\x00\xA1\xDD\x0F\x00\x82\x08\x00\x80\x71\x09\x00\x80\x3E\x09"
-     "\x00\x80"
-     "\xE7\x1F\x00\x80\x6F\x09\x00\x80\xEF\x1F\x00\x80\x7A\x09\x00\x80\xEB\x1F"
-     "\x00\x80"
-     "\xB4\x08\x00\x80\x93\x0D\x00\x00\xA7\x1F\x00\x80\xF3\x1F\x00\x80\x84\xD5"
-     "\x0C\x00"
-     "\x85\xE9\x0E\x00\x94\x69\x00\x00\x87\xDD\x0E\x00\xAF\x1F\x00\x80\x99\xB4"
-     "\x02\x00"
-     "\xB7\x1F\x00\x80\xBF\x1F\x00\x80\xAB\x1F\x00\x80\x37\x22\x00\x80\xB3\x1F"
-     "\x00\x80"
-     "\x9F\x74\x01\x00\xBB\x1F\x00\x80\x91\x18\x0D\x00\x80\x8D\x0E\x00\x81\x81"
-     "\x0E\x00"
-     "\x86\x85\x0E\x00\x95\x8C\x03\x00\x84\x89\x0E\x00\x97\x44\x02\x00\x82\x11"
-     "\x00\x00"
-     "\xA9\xB8\x00\x00\x80\xD1\x00\x00\x81\xED\x00\x00\xC3\x1F\x00\x80\xD0\x0D"
-     "\x00\x80"
-     "\xDF\x1F\x00\x80\x85\x59\x0F\x00\x83\x89\x00\x00\xA1\x34\x0D\x00\x81\x45"
-     "\x0E\x00"
-     "\x80\x49\x0E\x00\xE3\x1F\x00\x80\xA5\x34\x01\x00\x85\x61\x0F\x00\xCC\xF0"
-     "\x14\x00"
-     "\x17\x20\x00\x80\xB9\xC4\x05\x00\xCC\xC8\x03\x00\xCD\xDC\x03\x00\x80\xDD"
-     "\x00\x00"
-     "\x81\xC1\x00\x00\x1F\x20\x00\x80\xBF\x90\x05\x00\x85\x11\x00\x00\xB1\xEC"
-     "\x07\x00"
-     "\x80\xF5\x00\x00\x81\xC1\x00\x00\x9B\x21\x00\x80\xB5\x8C\x06\x00\x27\x20"
-     "\x00\x80"
-     "\xB7\x40\x06\x00\x80\xDD\x0E\x00\x81\xE9\x0E\x00\xCC\x28\x02\x00\xCD\xB4"
-     "\x02\x00"
-     "\x80\xCD\x0E\x00\x81\xF9\x0E\x00\x85\x29\x00\x00\x83\x85\x01\x00\x80\x75"
-     "\x01\x00"
-     "\x81\xB1\x01\x00\x80\xF1\x01\x00\x81\xD5\x01\x00\xA3\x21\x00\x80\x2F\x20"
-     "\x00\x80"
-     "\x85\x05\x00\x00\xAB\x21\x00\x80\x80\x99\x01\x00\x81\xBD\x01\x00\x82\x7D"
-     "\x00\x00"
-     "\x93\xD5\x01\x00\x94\xE1\x01\x00\x85\x0D\x00\x00\x93\x21\x00\x80\x1B\x20"
-     "\x00\x80"
-     "\x80\x09\x00\x00\x81\x19\x00\x00\x23\x20\x00\x80\x93\xAD\x01\x00\x94\x2D"
-     "\x00\x00"
-     "\x9F\x21\x00\x80\x85\x0D\x00\x00\x2B\x20\x00\x80\x85\x05\x00\x00\xA7\x21"
-     "\x00\x80"
-     "\x33\x20\x00\x80\x85\x29\x00\x00\x82\x19\x00\x00\x85\x35\x00\x00\x80\xF1"
-     "\x00\x00"
-     "\x81\xE1\x00\x00\xAF\x21\x00\x80\x97\x21\x00\x80\x3B\x22\x00\x80\x85\x05"
-     "\x00\x00"
-     "\x5B\x22\x00\x80\x83\x75\x01\x00\x80\xED\x01\x00\x81\x29\x01\x00\xCC\xF0"
-     "\x01\x00"
-     "\xCD\xB0\x01\x00\xD3\x0D\x00\x80\x57\x22\x00\x80\x53\x22\x00\x80\x2A\x0E"
-     "\x00\x80"
-     "\x57\x20\x00\x80\x5F\x20\x00\x80\x80\x3D\x00\x00\x81\x0D\x00\x00\x67\x20"
-     "\x00\x80"
-     "\x6F\x20\x00\x80\x80\x2D\x00\x00\x81\x1D\x00\x00\x82\x15\x00\x00\x5B\x20"
-     "\x00\x80"
-     "\xCD\x2C\x01\x00\x63\x20\x00\x80\x6B\x20\x00\x80\x73\x20\x00\x80\x88\xC5"
-     "\x03\x00"
-     "\x9F\x22\x00\x80\xCC\x90\x02\x00\xCD\xBC\x02\x00\x84\xED\x03\x00\xD6\x0D"
-     "\x00\x80"
-     "\x86\xF9\x03\x00\x97\x20\x00\x80\x80\x81\x03\x00\x81\xFD\x03\x00\x80\x3D"
-     "\x00\x00"
-     "\x81\x35\x00\x00\x81\x49\x00\x00\x80\x41\x00\x00\xCD\xDC\x01\x00\x82\x41"
-     "\x00\x00"
-     "\x9F\x20\x00\x80\x9B\x20\x00\x80\xA3\x20\x00\x80\xCD\x30\x01\x00\x94\x9D"
-     "\x03\x00"
-     "\x87\x22\x00\x80\xCD\xF0\x01\x00\xCC\x0C\x01\x00\x81\xB9\x03\x00\x80\xC5"
-     "\x03\x00"
-     "\x83\xA1\x03\x00\x93\xA5\x03\x00\x80\xAD\x00\x00\x81\xD5\x00\x00\x80\x9D"
-     "\x00\x00"
-     "\x81\xA9\x00\x00\x83\x22\x00\x80\xD9\x0D\x00\x80\x81\xC1\x00\x00\x80\xC9"
-     "\x00\x00"
-     "\x80\xB5\x00\x00\x81\x81\x00\x00\x7F\x22\x00\x80\x83\x69\x04\x00\xCC\x70"
-     "\x03\x00"
-     "\xCD\xB4\x03\x00\x7B\x22\x00\x80\xCD\x3C\x01\x00\x2D\x0E\x00\x80\x93\x01"
-     "\x04\x00"
-     "\xCD\x8C\x02\x00\xCC\xF4\x02\x00\x80\x0D\x00\x00\x81\x35\x00\x00\x94\xD9"
-     "\x06\x00"
-     "\xCB\x20\x00\x80\xCF\x20\x00\x80\xD3\x20\x00\x80\xCC\x08\x01\x00\xCD\x1C"
-     "\x01\x00"
-     "\x81\x11\x00\x00\x80\x29\x00\x00\xA3\x22\x00\x80\x82\x19\x00\x00\x80\x91"
-     "\x01\x00"
-     "\x81\x91\x01\x00\xCD\x68\x05\x00\xCC\x94\x02\x00\xCC\x10\x09\x00\xCD\x28"
-     "\x16\x00"
-     "\xCC\x58\x0E\x00\xCD\x78\x0E\x00\xCC\x14\x0D\x00\xCD\xD8\x0A\x00\xCC\x28"
-     "\x0C\x00"
-     "\xCD\x8C\x0D\x00\xCC\xE0\x17\x00\xCC\xE0\x0A\x00\xCC\x38\x0B\x00\xCD\xF4"
-     "\x08\x00"
-     "\x85\x11\x00\x00\xDC\x0D\x00\x80\x80\x51\x07\x00\x81\x51\x07\x00\xDB\x21"
-     "\x00\x80"
-     "\xCD\x90\x0E\x00\x85\x05\x00\x00\xE3\x21\x00\x80\xCC\xD8\x0E\x00\xCD\xEC"
-     "\x01\x00"
-     "\xEB\x21\x00\x80\xCD\x00\x0E\x00\x85\x19\x00\x00\xCD\xF0\x0F\x00\xCD\x3C"
-     "\x0E\x00"
-     "\xCD\x54\x0E\x00\xCC\x68\x01\x00\xCD\x6C\x01\x00\xD3\x21\x00\x80\x64\x08"
-     "\x00\x80"
-     "\x94\x99\x07\x00\xCC\xC0\x3B\x00\x80\x61\x01\x00\x81\xD9\x00\x00\x85\x29"
-     "\x00\x00"
-     "\xCD\x64\x0E\x00\xCC\x78\x01\x00\xCD\x7C\x01\x00\x81\xAD\x07\x00\x80\xAD"
-     "\x07\x00"
-     "\x85\x65\x00\x00\x82\x9D\x07\x00\x80\x51\x01\x00\x81\x51\x01\x00\x94\xE1"
-     "\x07\x00"
-     "\xCD\xC0\x00\x00\x84\x79\x01\x00\x93\xC5\x07\x00\x86\x61\x01\x00\xDF\x21"
-     "\x00\x80"
-     "\x88\x21\x01\x00\x85\x0D\x00\x00\xE7\x21\x00\x80\xCD\x18\x01\x00\xCC\xD8"
-     "\x00\x00"
-     "\xCD\xB4\x00\x00\x80\xDD\x07\x00\x81\xCD\x07\x00\x93\x20\x00\x80\x85\x09"
-     "\x00\x00"
-     "\xC7\x20\x00\x80\xEF\x21\x00\x80\xF7\x20\x00\x80\xD7\x21\x00\x80\xFF\x20"
-     "\x00\x80"
-     "\x07\x21\x00\x80\x0F\x21\x00\x80\x03\x21\x00\x80\xFB\x20\x00\x80\xA7\x22"
-     "\x00\x80"
-     "\x0B\x21\x00\x80\x13\x21\x00\x80\xCC\xB8\x02\x00\xCD\x1C\x03\x00\x80\x65"
-     "\x00\x00"
-     "\x81\x75\x00\x00\x82\x7D\x00\x00\x17\x21\x00\x80\x85\x09\x00\x00\x85\x41"
-     "\x00\x00"
-     "\xFB\x21\x00\x80\x30\x0E\x00\x80\x80\x99\x06\x00\x81\x21\x07\x00\x85\x19"
-     "\x00\x00"
-     "\x83\x7D\x00\x00\x03\x22\x00\x80\x85\x59\x00\x00\xF7\x21\x00\x80\xF3\x21"
-     "\x00\x80"
-     "\x80\xCD\x00\x00\x81\xD9\x00\x00\x87\x1F\x00\x80\x85\x11\x00\x00\x84\xE9"
-     "\x00\x00"
-     "\x8F\x1F\x00\x80\x86\xE5\x00\x00\x3B\x21\x00\x80\x80\x35\x00\x00\x81\x0D"
-     "\x00\x00"
-     "\x97\x1F\x00\x80\x85\x1D\x00\x00\x43\x21\x00\x80\x9F\x1F\x00\x80\x85\x05"
-     "\x00\x00"
-     "\x4B\x21\x00\x80\x80\x55\x00\x00\x81\x6D\x00\x00\x82\x7D\x00\x00\x93\x45"
-     "\x00\x00"
-     "\x94\x0D\x00\x00\x85\x0D\x00\x00\x33\x21\x00\x80\x8B\x1F\x00\x80\x80\x09"
-     "\x00\x00"
-     "\x81\x11\x00\x00\x93\x1F\x00\x80\x85\x1D\x00\x00\x3F\x21\x00\x80\x9B\x1F"
-     "\x00\x80"
-     "\x85\x05\x00\x00\x47\x21\x00\x80\x80\xE9\x01\x00\x81\xF1\x01\x00\x82\x05"
-     "\x00\x00"
-     "\xA3\x1F\x00\x80\x85\x09\x00\x00\x85\x09\x00\x00\x4F\x21\x00\x80\x37\x21"
-     "\x00\x80"
-     "\x80\x6D\x01\x00\x81\x79\x01\x00\x82\x19\x00\x00\x83\xA5\x01\x00\x07\x22"
-     "\x00\x80"
-     "\x85\x75\x00\x00\x85\x05\x00\x00\x0B\x22\x00\x80\xFF\x21\x00\x80\x1B\x21"
-     "\x00\x80"
-     "\xCC\xC8\x02\x00\xCD\xDC\x02\x00\x33\x0E\x00\x80\xC7\x1F\x00\x80\x80\x39"
-     "\x00\x00"
-     "\x81\x39\x00\x00\xCF\x1F\x00\x80\xD7\x1F\x00\x80\xCB\x1F\x00\x80\xD3\x1F"
-     "\x00\x80"
-     "\x80\x1D\x00\x00\x81\x0D\x00\x00\xDB\x1F\x00\x80\x1F\x21\x00\x80\x80\xC5"
-     "\x00\x00"
-     "\x81\xD5\x00\x00\xCD\xC0\x00\x00\xCC\x24\x02\x00\x80\xD5\x00\x00\x81\xC5"
-     "\x00\x00"
-     "\x85\x39\x00\x00\x83\xC9\x00\x00\x1F\x22\x00\x80\x36\x0E\x00\x80\x80\xD5"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x85\x05\x00\x00\x27\x22\x00\x80\xF7\x1F\x00\x80\x7B\x21"
-     "\x00\x80"
-     "\x80\x09\x00\x00\x81\x11\x00\x00\xFF\x1F\x00\x80\x93\x99\x00\x00\x94\xB9"
-     "\x00\x00"
-     "\x07\x20\x00\x80\x85\x65\x00\x00\x85\x3D\x00\x00\x83\x21\x00\x80\x93\x5D"
-     "\x00\x00"
-     "\x0F\x20\x00\x80\x85\x11\x00\x00\xCD\x70\x05\x00\xCC\x74\x05\x00\x94\x01"
-     "\x3C\x00"
-     "\x8B\x21\x00\x80\x73\x21\x00\x80\xCD\x28\x01\x00\x7F\x21\x00\x80\x87\x21"
-     "\x00\x80"
-     "\x85\x19\x00\x00\x8F\x21\x00\x80\x77\x21\x00\x80\x2F\x22\x00\x80\x23\x22"
-     "\x00\x80"
-     "\x23\x21\x00\x80\x85\x25\x00\x00\x85\x39\x00\x00\xCC\xF8\x02\x00\xCD\xC4"
-     "\x03\x00"
-     "\xCD\x3C\x01\x00\x39\x0E\x00\x80\x81\x95\x03\x00\x80\x8D\x03\x00\xCD\xC4"
-     "\x01\x00"
-     "\x82\xA5\x03\x00\x85\x51\x00\x00\x85\x49\x00\x00\xCC\x28\x01\x00\xCD\x2C"
-     "\x01\x00"
-     "\xCD\x38\x01\x00\xCC\x3C\x01\x00\x80\x69\x3E\x00\x81\x69\x3E\x00\x43\x22"
-     "\x00\x80"
-     "\x3F\x22\x00\x80\xCD\x38\x3C\x00\xCC\x54\x3C\x00\x81\xD1\x3C\x00\x93\x9D"
-     "\x3E\x00"
-     "\xCC\x48\x01\x00\xCD\xC8\x02\x00\xCD\x34\x01\x00\x47\x22\x00\x80\x94\xB9"
-     "\x3E\x00"
-     "\xDF\x0D\x00\x80\x80\xA1\x3E\x00\x81\xA1\x3E\x00\x82\xA1\x3E\x00\x88\x8D"
-     "\x3C\x00"
-     "\x4F\x22\x00\x80\x85\xAD\x00\x00\x80\x2D\x00\x00\x81\x21\x00\x00\x85\xD5"
-     "\x3F\x00"
-     "\x8F\x20\x00\x80\x80\xED\x00\x00\x81\xF1\x00\x00\x86\xA5\x00\x00\x3F\x20"
-     "\x00\x80"
-     "\x84\xA9\x00\x00\xCD\x24\x01\x00\xCD\x28\x01\x00\x47\x20\x00\x80\x88\xF9"
-     "\x3E\x00"
-     "\x85\xF1\x3F\x00\x4F\x20\x00\x80\x43\x20\x00\x80\x85\xC5\x3F\x00\xCD\x30"
-     "\x01\x00"
-     "\xCD\x10\x01\x00\xCD\xF4\x06\x00\x80\xDD\x01\x00\x81\xE9\x01\x00\xCD\xBC"
-     "\x06\x00"
-     "\xCD\x70\x06\x00\xCC\xE0\x06\x00\xCD\x5C\x01\x00\xCC\x68\x06\x00\xCD\x90"
-     "\x06\x00"
-     "\xCD\x64\x06\x00\xCD\x78\x06\x00\xCC\xAC\x07\x00\xCD\xA8\x07\x00\xCC\xE8"
-     "\x07\x00"
-     "\xCD\xC8\x07\x00\x82\x4D\x3F\x00\x83\xFD\x02\x00\x80\x35\x02\x00\x81\xE9"
-     "\x02\x00"
-     "\x4B\x20\x00\x80\x53\x20\x00\x80\x80\x05\x3D\x00\x85\x7D\x01\x00\x4B\x22"
-     "\x00\x80"
-     "\x27\x21\x00\x80\xCD\x14\x01\x00\xB0\x0E\x00\x80\x81\xED\x01\x00\x80\xE1"
-     "\x01\x00"
-     "\xCD\x3C\x01\x00\x82\xCD\x01\x00\xCD\x2C\x01\x00\x82\x75\x01\x00\x81\x6D"
-     "\x01\x00"
-     "\x80\x59\x01\x00\x80\x65\x01\x00\x81\xC5\x00\x00\x7F\x20\x00\x80\xCD\x24"
-     "\x01\x00"
-     "\xCD\x38\x01\x00\x82\xF1\x00\x00\x81\xF9\x00\x00\x80\x59\x01\x00\x80\x29"
-     "\x00\x00"
-     "\x81\x71\x00\x00\xCC\x18\x01\x00\xCD\x7C\x01\x00\xCD\x2C\x01\x00\x87\x20"
-     "\x00\x80"
-     "\x81\x1D\x00\x00\x80\x1D\x00\x00\x83\x20\x00\x80\x8B\x20\x00\x80\x6B\x22"
-     "\x00\x80"
-     "\xCD\x24\x01\x00\xCC\xE4\x3D\x00\xCD\x5C\x0F\x00\xCD\xE8\x00\x00\xCC\x0C"
-     "\x01\x00"
-     "\x80\xD5\x01\x00\x81\xC9\x01\x00\x82\x99\x00\x00\x83\xE5\x3F\x00\x03\x20"
-     "\x00\x80"
-     "\x0B\x20\x00\x80\x13\x20\x00\x80\x2B\x22\x00\x80\xAA\x0E\x00\x80\x6F\x22"
-     "\x00\x80"
-     "\x37\x20\x00\x80\x2B\x21\x00\x80\x3B\x20\x00\x80\xB3\x0E\x00\x80\x80\x4D"
-     "\x3F\x00"
-     "\x81\x41\x3F\x00\x77\x20\x00\x80\x63\x22\x00\x80\x7B\x20\x00\x80\x5F\x22"
-     "\x00\x80"
-     "\x80\x25\x3F\x00\x81\x29\x3F\x00\x93\x91\x3F\x00\x83\x7D\x00\x00\xAD\x0E"
-     "\x00\x80"
-     "\x94\x41\x00\x00\xCC\xD8\x02\x00\xCD\xAC\x02\x00\x67\x22\x00\x80\x93\x55"
-     "\x00\x00"
-     "\x80\x09\x00\x00\x81\x1D\x00\x00\x3C\x0E\x00\x80\x77\x22\x00\x80\x94\x41"
-     "\x00\x00"
-     "\xA7\x20\x00\x80\x80\x9D\x00\x00\x81\xA1\x00\x00\x80\x1D\x00\x00\x81\x11"
-     "\x00\x00"
-     "\x84\xA5\x00\x00\xAF\x20\x00\x80\x86\xA5\x00\x00\xB7\x20\x00\x80\x88\xF1"
-     "\x00\x00"
-     "\x82\xD1\x00\x00\x81\xD9\x00\x00\x80\xCD\x00\x00\x80\x25\x00\x00\x81\x29"
-     "\x00\x00"
-     "\x82\x05\x00\x00\xBF\x20\x00\x80\xAB\x20\x00\x80\xB3\x20\x00\x80\xBB\x20"
-     "\x00\x80"
-     "\x93\xB1\x00\x00\x94\x11\x00\x00\xC3\x20\x00\x80\x80\x1D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x80\x25\x00\x00\x81\x2D\x00\x00\x82\x3D\x00\x00\x73\x22\x00\x80\x80\xED"
-     "\x00\x00"
-     "\x81\xD1\x00\x00\x82\x15\x00\x00\x83\x81\x00\x00\x81\xD0\x3D\x00\x2F\x21"
-     "\x00\x80"
-     "\xCC\x20\x02\x00\xCD\xDC\x01\x00\x85\x78\x02\x00\x8B\x22\x00\x80\xB6\x0E"
-     "\x00\x80"
-     "\x93\x22\x00\x80\x89\x18\x03\x00\xD7\x20\x00\x80\x80\x2D\x00\x00\x81\x35"
-     "\x00\x00"
-     "\x80\x09\x00\x00\x81\x6D\x00\x00\xDF\x20\x00\x80\xBB\x21\x00\x80\x91\xB1"
-     "\x00\x00"
-     "\x90\xA9\x00\x00\x93\xDD\x3B\x00\x92\x01\x04\x00\x95\xA5\x00\x00\x94\x95"
-     "\x3B\x00"
-     "\xE7\x20\x00\x80\x96\xA1\x00\x00\x85\x09\x00\x00\x93\x41\x00\x00\xC3\x21"
-     "\x00\x80"
-     "\xEF\x20\x00\x80\x85\x05\x00\x00\xCB\x21\x00\x80\x94\xF5\x00\x00\xB3\x21"
-     "\x00\x80"
-     "\x80\xB9\x00\x00\x81\xDD\x00\x00\x82\xE5\x00\x00\xDB\x20\x00\x80\xE3\x20"
-     "\x00\x80"
-     "\x85\xE9\x00\x00\x80\x09\x00\x00\x81\x35\x00\x00\x85\x05\x00\x00\xBF\x21"
-     "\x00\x80"
-     "\xEB\x20\x00\x80\x85\x1D\x00\x00\xC7\x21\x00\x80\xF3\x20\x00\x80\x85\x05"
-     "\x00\x00"
-     "\xCF\x21\x00\x80\xB1\xC1\x05\x00\xB0\xC5\x03\x00\xB7\x21\x00\x80\xB2\xC5"
-     "\x03\x00"
-     "\xB5\xD9\x05\x00\xB4\xCD\x03\x00\x97\x22\x00\x80\x85\x39\x00\x00\xB9\xFD"
-     "\x03\x00"
-     "\x9B\x22\x00\x80\x8F\x22\x00\x80\x42\x0E\x00\x80\x54\x0E\x00\x80\x9E\x0E"
-     "\x00\x80"
-     "\xFB\x1F\x00\x80\x8C\x0E\x00\x80\x5A\x0E\x00\x80\xCC\x88\x02\x00\x92\x0E"
-     "\x00\x80"
-     "\xCD\xE0\x3E\x00\xCD\x90\x01\x00\xCC\x24\x01\x00\xF7\x0D\x00\x80\x15\x0E"
-     "\x00\x80"
-     "\xC8\x0E\x00\x80\x6D\x0F\x00\x80\x80\xB1\x00\x00\xCD\xD4\x3E\x00\xCD\xE4"
-     "\x3E\x00"
-     "\x73\x0F\x00\x80\xCC\xBC\x3E\x00\xCD\xB8\x3E\x00\x80\xD1\x03\x00\x81\xED"
-     "\x03\x00"
-     "\x82\xFD\x03\x00\x86\x69\x00\x00\xC5\x0E\x00\x80\x85\x9D\x03\x00\xCD\x3C"
-     "\x01\x00"
-     "\xBF\x0E\x00\x80\xCC\xE8\x02\x00\xCD\x3C\x3F\x00\x88\xE5\x00\x00\xCD\x18"
-     "\x01\x00"
-     "\x79\x0F\x00\x80\x88\x41\x00\x00\xC2\x0E\x00\x80\x64\x0F\x00\x80\xCD\x2C"
-     "\x01\x00"
-     "\x88\x0F\x00\x80\x80\xD5\x00\x00\x8E\x0F\x00\x80\x86\xE1\x00\x00\x86\xE9"
-     "\x00\x00"
-     "\xCE\x0E\x00\x80\xCD\x24\x01\x00\x94\x0F\x00\x80\xCD\x10\x01\x00\x88\xD1"
-     "\x00\x00"
-     "\x88\x29\x00\x00\xCC\xF8\x02\x00\xD4\x0E\x00\x80\xCD\xF8\x02\x00\xCC\x24"
-     "\x01\x00"
-     "\x9A\x0F\x00\x80\x85\x2D\x03\x00\xCC\xA0\x3E\x00\xCD\x6C\x3E\x00\x80\xD5"
-     "\x03\x00"
-     "\x81\xCD\x03\x00\x82\x01\x03\x00\x83\xF9\x03\x00\xCC\x64\x03\x00\xCD\xCC"
-     "\x02\x00"
-     "\xCB\x0E\x00\x80\xCD\x24\x01\x00\xCC\x0C\x02\x00\xCD\x08\x02\x00\x81\x11"
-     "\x00\x00"
-     "\xCC\x9C\x03\x00\xCC\xB0\x3E\x00\xCD\xB4\x3E\x00\xCC\xC4\x3E\x00\xCD\xC0"
-     "\x3E\x00"
-     "\xCC\x80\x3E\x00\xCD\xB8\x3E\x00\xA0\x0F\x00\x80\xCC\x84\x02\x00\xCC\x98"
-     "\x3F\x00"
-     "\xCD\x50\x3E\x00\xCC\x20\x3E\x00\xCD\xA0\x3E\x00\xCD\x0C\x3F\x00\xCD\x30"
-     "\x3F\x00"
-     "\xCD\x78\x3F\x00\xCD\x04\x3F\x00\x88\x59\x00\x00\xB5\x0F\x00\x80\xCD\xF8"
-     "\x01\x00"
-     "\xCC\xC4\x01\x00\xD1\x0E\x00\x80\xBB\x0F\x00\x80\xC1\x0F\x00\x80\xCC\x14"
-     "\x02\x00"
-     "\xCC\x08\x01\x00\xCD\xC8\x01\x00\x88\x05\x00\x00\xC7\x0F\x00\x80\xCD\x0F"
-     "\x00\x80"
-     "\xCC\x28\x02\x00\xAF\x0F\x00\x80\x88\x0D\x00\x00\x86\xD1\x00\x00\x80\x1D"
-     "\x01\x00"
-     "\x84\xCD\x00\x00\x88\xF5\x00\x00\xCC\x3C\x02\x00\x84\x35\x01\x00\xCC\x44"
-     "\x02\x00"
-     "\x86\x29\x01\x00\x70\x0F\x00\x80\x88\x65\x01\x00\x76\x0F\x00\x80\xA7\x44"
-     "\x05\x00"
-     "\x37\x0F\x00\x80\x8B\xED\x00\x00\x88\xED\x00\x00\x81\x0D\x00\x00\x88\x25"
-     "\x00\x00"
-     "\x86\x65\x00\x00\xCC\x70\x02\x00\xCD\x74\x02\x00\xCC\x30\x02\x00\xCD\xD8"
-     "\x05\x00"
-     "\x31\x0F\x00\x80\x7C\x0F\x00\x80\x80\x39\x00\x00\x34\x0F\x00\x80\xCC\xE0"
-     "\x05\x00"
-     "\x67\x0F\x00\x80\xCC\x28\x01\x00\xCD\x14\x01\x00\x86\x25\x00\x00\x88\x55"
-     "\x00\x00"
-     "\x8F\x0E\x00\x80\x86\x84\x30\x00\x4B\x0E\x00\x80\x80\xD5\x07\x00\x86\xFD"
-     "\x07\x00"
-     "\x8B\x0F\x00\x80\xCC\x24\x02\x00\x88\x3D\x00\x00\x91\x0F\x00\x80\x49\x0F"
-     "\x00\x80"
-     "\x88\x3D\x00\x00\x97\x0F\x00\x80\xCC\x48\x02\x00\xCD\x78\x02\x00\xF2\x0E"
-     "\x00\x80"
-     "\x9D\x0F\x00\x80\x97\xC0\x05\x00\x96\x70\x05\x00\x95\x18\x05\x00\x80\x69"
-     "\x00\x00"
-     "\x93\x58\x05\x00\x81\x39\x00\x00\x88\x65\x00\x00\x90\xF8\x3C\x00\x86\x59"
-     "\x00\x00"
-     "\x9E\xA8\x05\x00\x84\x45\x00\x00\x46\x0F\x00\x80\xCC\xD4\x02\x00\x9A\xB4"
-     "\x05\x00"
-     "\x80\x5D\x00\x00\x98\xAC\x05\x00\xA7\xEC\x04\x00\x88\x11\x00\x00\xCC\xD8"
-     "\x02\x00"
-     "\xCD\xDC\x02\x00\xA3\xBC\x04\x00\xA3\x0F\x00\x80\xCC\x60\x02\x00\xB8\x0F"
-     "\x00\x80"
-     "\x4C\x0F\x00\x80\xBE\x0F\x00\x80\xAD\x08\x04\x00\xC4\x0F\x00\x80\xAB\xF4"
-     "\x04\x00"
-     "\xCC\x2C\x02\x00\x88\x05\x00\x00\xCA\x0F\x00\x80\xB7\xE8\x03\x00\xB6\x1C"
-     "\x04\x00"
-     "\xB5\x28\x04\x00\xCC\x00\x02\x00\xB3\x28\x04\x00\x8B\x79\x00\x00\x88\x7D"
-     "\x00\x00"
-     "\xB0\x74\x04\x00\x86\x41\x00\x00\xBE\xA4\x03\x00\x84\x75\x00\x00\x88\x1D"
-     "\x00\x00"
-     "\xD0\x0F\x00\x80\xBA\x4C\x03\x00\xCC\xDC\x03\x00\xB8\xFC\x03\x00\x83\xA8"
-     "\x02\x00"
-     "\x88\x0D\x00\x00\xB2\x0F\x00\x80\x88\x15\x00\x00\x87\x94\x02\x00\xCC\x38"
-     "\x02\x00"
-     "\x40\x0F\x00\x80\xCC\x04\x02\x00\x8B\xDC\x02\x00\x8F\x0D\x00\x00\x52\x0F"
-     "\x00\x80"
-     "\x8F\x19\x00\x00\xCC\x20\x02\x00\x55\x0F\x00\x80\x8D\xF0\x02\x00\x88\x75"
-     "\x00\x00"
-     "\x98\x20\x03\x00\x99\x2C\x03\x00\x82\x0F\x00\x80\x1B\x0E\x00\x80\xCC\x4C"
-     "\x02\x00"
-     "\x96\x70\x03\x00\xCC\x24\x02\x00\x88\x3D\x00\x00\x85\x0F\x00\x80\xCC\x2C"
-     "\x02\x00"
-     "\x88\x05\x00\x00\xA9\x0F\x00\x80\xCC\x24\x02\x00\x88\x0D\x00\x00\xAC\x0F"
-     "\x00\x80"
-     "\x87\xF5\x00\x00\xA8\xD4\x03\x00\xA9\xC4\x03\x00\xD3\x0F\x00\x80\xD9\x60"
-     "\x02\x00"
-     "\xC8\x10\x00\x80\xCB\x10\x00\x80\xD1\x10\x00\x80\x94\x35\x00\x00\x93\x31"
-     "\x00\x00"
-     "\xD9\x68\x02\x00\xCE\x10\x00\x80\xD9\x4C\x02\x00\x94\x05\x00\x00\xD4\x10"
-     "\x00\x80"
-     "\x95\x21\x00\x00\x94\x29\x00\x00\x46\x11\x00\x80\x6A\x17\x00\x80\x39\x18"
-     "\x00\x80"
-     "\xC8\x17\x00\x80\xD9\x60\x02\x00\x2D\x18\x00\x80\xB5\xD8\x03\x00\xB4\xF0"
-     "\x03\x00"
-     "\x94\x35\x00\x00\xD9\x58\x02\x00\x50\x18\x00\x80\x94\x05\x00\x00\xD9\x54"
-     "\x02\x00"
-     "\x94\x0D\x00\x00\x27\x18\x00\x80\xE0\x74\x01\x00\x8A\xC8\x00\x00\xBC\x15"
-     "\x00\x00"
-     "\x88\xC8\x00\x00\xE0\x80\x02\x00\x7D\x18\x00\x80\x81\xA0\x00\x00\xA4\xEC"
-     "\x02\x00"
-     "\xA4\xC8\x02\x00\xA8\x5C\x00\x00\xBC\x0D\x00\x00\x8F\x18\x00\x80\xE0\x84"
-     "\x02\x00"
-     "\xBC\x05\x00\x00\x93\x18\x00\x80\xA4\xF8\x02\x00\xE0\xF4\x02\x00\xB0\xCC"
-     "\x03\x00"
-     "\x95\xD0\x00\x00\x53\x18\x00\x80\xB3\xE0\x03\x00\xA6\xC8\x02\x00\xA7\x60"
-     "\x02\x00"
-     "\x92\xD8\x00\x00\x5A\x18\x00\x80\xBE\xC1\x00\x00\x61\x18\x00\x80\x97\xC1"
-     "\x00\x00"
-     "\x68\x18\x00\x80\x6F\x18\x00\x80\x76\x18\x00\x80\xCD\x78\x3F\x00\xCC\xBC"
-     "\x3F\x00"
-     "\xBE\x80\x0D\x00\x81\x18\x00\x80\xBC\x78\x0C\x00\xBD\x80\x0D\x00\xBA\xF4"
-     "\x0C\x00"
-     "\xBB\x54\x0C\x00\xB8\xF4\x0C\x00\x88\x18\x00\x80\xAC\x18\x00\x80\xB7\xB8"
-     "\x0C\x00"
-     "\xB0\x18\x00\x80\xB5\x8C\x0C\x00\xB2\xA0\x03\x00\xB3\xA0\x0C\x00\x97\x18"
-     "\x00\x80"
-     "\xB1\x40\x03\x00\xAE\x70\x02\x00\xAF\x64\x03\x00\xB8\x05\x03\x00\xAD\x48"
-     "\x03\x00"
-     "\x9E\x18\x00\x80\xA5\x18\x00\x80\xA8\x44\x03\x00\xA9\xD8\x03\x00\xD0\x18"
-     "\x00\x80"
-     "\xA7\xD8\x03\x00\xA4\x68\x02\x00\xA5\x88\x03\x00\xB6\x35\x03\x00\xB7\x3D"
-     "\x03\x00"
-     "\x92\xC8\x02\x00\xB5\x3D\x03\x00\x99\x5D\x01\x00\x98\x4D\x01\x00\x9B\x65"
-     "\x01\x00"
-     "\x9A\x69\x01\x00\x9D\x65\x01\x00\x9C\x65\x01\x00\x9F\x85\x01\x00\x9E\x99"
-     "\x01\x00"
-     "\x87\x9C\x02\x00\xBE\xAD\x00\x00\x96\xA5\x00\x00\x97\xBD\x00\x00\xCC\x34"
-     "\x05\x00"
-     "\xCD\x8C\x37\x00\xCC\xB8\x38\x00\xCD\xAC\x38\x00\x9F\x95\x01\x00\xB6\x1D"
-     "\x00\x00"
-     "\x9D\x99\x01\x00\x9C\xF5\x01\x00\xB3\xB1\x01\x00\xAE\x78\x02\x00\xD7\x18"
-     "\x00\x80"
-     "\xB4\x18\x00\x80\x99\x3D\x00\x00\xBB\x18\x00\x80\x9B\x19\x00\x00\x9A\x09"
-     "\x00\x00"
-     "\xC2\x18\x00\x80\xC9\x18\x00\x80\xE0\x48\x02\x00\x9E\x09\x00\x00\xAC\x5C"
-     "\x02\x00"
-     "\xAD\xF4\x02\x00\xF0\x18\x00\x80\xEC\x18\x00\x80\xF4\x18\x00\x80\xDE\x18"
-     "\x00\x80"
-     "\x87\x60\x03\x00\xE5\x18\x00\x80\xAF\x54\x02\x00\xBE\x11\x00\x00\x97\x05"
-     "\x00\x00"
-     "\xF8\x18\x00\x80\xE0\xAC\x02\x00\xFC\x18\x00\x80\x86\xF8\x03\x00\x87\xEC"
-     "\x03\x00"
-     "\xE0\xB4\x02\x00\x04\x19\x00\x80\xAF\x48\x02\x00\xAE\x90\x02\x00\xE0\x3C"
-     "\x02\x00"
-     "\xBE\x0D\x00\x00\x00\x19\x00\x80\x97\x19\x00\x00\xE0\xD8\x02\x00\x86\x84"
-     "\x03\x00"
-     "\x96\x11\x00\x00\xBF\x00\x0C\x00\x9D\x6D\x00\x00\x9C\x61\x00\x00\x08\x19"
-     "\x00\x80"
-     "\xB1\x4C\x02\x00\xB3\x50\x02\x00\x95\x0D\x00\x00\x0C\x19\x00\x80\x86\x9C"
-     "\x03\x00"
-     "\xE0\xC8\x02\x00\xB3\x04\x02\x00\x82\x05\x00\x00\x18\x19\x00\x80\xB3\x50"
-     "\x02\x00"
-     "\x95\x0D\x00\x00\x1C\x19\x00\x80\x10\x19\x00\x80\x14\x19\x00\x80\xE0\xB4"
-     "\x02\x00"
-     "\x86\x8C\x03\x00\x87\xDC\x03\x00\xBE\x0D\x00\x00\x95\x69\x00\x00\x96\x79"
-     "\x00\x00"
-     "\x20\x19\x00\x80\xB4\xE8\x02\x00\xB5\x50\x02\x00\x97\x05\x00\x00\x28\x19"
-     "\x00\x80"
-     "\xE0\xD4\x02\x00\xB4\xF4\x02\x00\xBE\x19\x00\x00\xE0\xA0\x02\x00\x24\x19"
-     "\x00\x80"
-     "\xE0\xD4\x02\x00\x99\x8C\x03\x00\xB7\xD4\x02\x00\x8A\x05\x00\x00\x2C\x19"
-     "\x00\x80"
-     "\x30\x19\x00\x80\x8A\x15\x00\x00\xB7\x34\x02\x00\x8F\x1D\x00\x00\x34\x19"
-     "\x00\x80"
-     "\x38\x19\x00\x80\xB3\x05\x00\x00\x3C\x19\x00\x80\xB3\x05\x00\x00\x51\x19"
-     "\x00\x80"
-     "\x9C\x09\x00\x00\x9D\x09\x00\x00\x43\x19\x00\x80\x4A\x19\x00\x80\x8C\x05"
-     "\x00\x00"
-     "\x58\x19\x00\x80\x63\x19\x00\x80\x6A\x19\x00\x80\x71\x19\x00\x80\x9F\x49"
-     "\x00\x00"
-     "\x78\x19\x00\x80\x7F\x19\x00\x80\x5C\x19\x00\x80\x86\x19\x00\x80\x8D\x19"
-     "\x00\x80"
-     "\xCF\x19\x00\x80\xC5\x19\x00\x80\xE0\x19\x00\x80\xD6\x19\x00\x80\x94\x19"
-     "\x00\x80"
-     "\x83\xC9\x01\x00\x81\xF9\x01\x00\xA2\x19\x00\x80\xA9\x19\x00\x80\xB0\x19"
-     "\x00\x80"
-     "\xB7\x19\x00\x80\xBE\x19\x00\x80\x9B\x19\x00\x80\x80\xB4\x02\x00\xA5\x88"
-     "\x03\x00"
-     "\xE1\x08\x02\x00\xAE\x1D\x00\x00\xE7\x19\x00\x80\xBC\x09\x00\x00\x8D\xF5"
-     "\x01\x00"
-     "\xEB\x19\x00\x80\xE1\x00\x02\x00\x92\x95\x01\x00\xE3\x94\x10\x00\x93\x45"
-     "\x00\x00"
-     "\x97\x89\x01\x00\x85\x14\x00\x00\x87\x78\x01\x00\x86\x00\x04\x00\x63\x3B"
-     "\x00\x80"
-     "\x67\x3B\x00\x80\x6B\x3B\x00\x80\x6F\x3B\x00\x80\x73\x3B\x00\x80\x9D\x79"
-     "\x00\x00"
-     "\xEF\x8C\x68\x00\x9C\xA1\x01\x00\x77\x3B\x00\x80\x7B\x3B\x00\x80\xA2\x99"
-     "\x00\x00"
-     "\x7F\x3B\x00\x80\x83\x3B\x00\x80\x87\x3B\x00\x80\x8B\x3B\x00\x80\xA7\x89"
-     "\x00\x00"
-     "\x8F\x3B\x00\x80\x93\x3B\x00\x80\xA9\x49\x01\x00\x97\x3B\x00\x80\xAC\xA9"
-     "\x00\x00"
-     "\x9B\x3B\x00\x80\x9F\x3B\x00\x80\xA3\x3B\x00\x80\xB3\x25\x01\x00\xA7\x3B"
-     "\x00\x80"
-     "\xAB\x3B\x00\x80\xAF\x3B\x00\x80\xB7\x21\x01\x00\xB6\x39\x01\x00\xB5\x31"
-     "\x01\x00"
-     "\xB3\x3B\x00\x80\xB7\x3B\x00\x80\xB9\xF9\x00\x00\xB9\x11\x01\x00\xB8\x19"
-     "\x01\x00"
-     "\xBB\x3B\x00\x80\xBF\x3B\x00\x80\xC3\x3B\x00\x80\xC7\x3B\x00\x80\x80\xB0"
-     "\x01\x00"
-     "\x84\x88\x02\x00\xCB\x3B\x00\x80\x83\xC8\x01\x00\x84\x54\x03\x00\x84\x5C"
-     "\x04\x00"
-     "\xCF\x3B\x00\x80\x84\x5C\x05\x00\x80\xDD\x03\x00\x81\x2D\x00\x00\x82\x31"
-     "\x00\x00"
-     "\xBE\x3C\x02\x00\xD7\x3B\x00\x80\xDB\x3B\x00\x80\x87\x88\x03\x00\x86\x9C"
-     "\x04\x00"
-     "\xB3\x2D\x03\x00\xDF\x3B\x00\x80\xE3\x3B\x00\x80\xBE\x00\x04\x00\xBE\x1C"
-     "\x05\x00"
-     "\xB6\xD1\x03\x00\xB5\xD9\x03\x00\xE7\x3B\x00\x80\xBB\xF9\x03\x00\xBA\xF1"
-     "\x03\x00"
-     "\x9A\x58\xD3\x01\x84\xE0\x07\x00\xBF\xC5\x03\x00\xBE\xD9\x03\x00\xBD\xDD"
-     "\x03\x00"
-     "\xBC\xDD\x03\x00\xBE\x00\x18\x00\xA5\x05\x03\x00\xA6\x0D\x03\x00\xEB\x3B"
-     "\x00\x80"
-     "\x84\x1C\x18\x00\xEF\x3B\x00\x80\xF3\x3B\x00\x80\xA3\xF1\x03\x00\xAC\x01"
-     "\x03\x00"
-     "\xAD\x01\x03\x00\xAE\x05\x03\x00\xAF\x19\x03\x00\xAC\xA4\x1B\x02\xAD\xDC"
-     "\x1A\x02"
-     "\xAA\x2D\x03\x00\xAB\x25\x03\x00\xBE\x4C\x19\x00\xBE\x48\x1A\x00\xF7\x3B"
-     "\x00\x80"
-     "\xBE\x80\x1B\x00\xB4\xE0\x1A\x02\xB5\xD4\x1D\x02\xB6\x30\x1E\x02\xEF\x08"
-     "\x02\x00"
-     "\xFB\x3B\x00\x80\xE1\xA0\x01\x00\xBA\x38\x1A\x02\xE3\xF8\x02\x00\xBA\x00"
-     "\x00\x00"
-     "\xBD\x64\x1C\x02\xBE\xF4\x1C\x02\xBF\x00\x10\x02\x91\x04\xD3\x01\x90\xE4"
-     "\xF6\x01"
-     "\xE0\x44\x01\x00\x92\x08\x3E\x02\xFF\x3B\x00\x80\x03\x3C\x00\x80\x07\x3C"
-     "\x00\x80"
-     "\x0B\x3C\x00\x80\xBE\xAC\x1C\x00\x0F\x3C\x00\x80\x13\x3C\x00\x80\x17\x3C"
-     "\x00\x80"
-     "\x1B\x3C\x00\x80\x1F\x3C\x00\x80\x23\x3C\x00\x80\x27\x3C\x00\x80\x81\xB0"
-     "\x6D\x00"
-     "\x80\x80\x01\x00\x83\x1C\x52\x00\x82\xD8\x77\x00\x85\x20\x9A\x00\x84\x90"
-     "\xBE\x00"
-     "\x87\x08\xCF\x00\x86\x8C\xE6\x00\x89\x6C\x37\x01\x88\xE0\x2C\x01\x8B\x18"
-     "\x7E\x01"
-     "\x8A\xD8\x13\x01\x8D\xE0\xA5\x01\x8C\xF0\x5A\x01\x8F\xFC\xEB\x01\x8E\x58"
-     "\x8F\x01"
-     "\xB0\xD5\x17\x00\xB1\x01\x68\x00\xB2\x8D\x6B\x00\xB3\x9D\x6B\x00\xB4\x49"
-     "\x6B\x00"
-     "\xB5\x95\x6F\x00\x2B\x3C\x00\x80\xE0\x70\x01\x00\x2F\x3C\x00\x80\x33\x3C"
-     "\x00\x80"
-     "\x37\x3C\x00\x80\x3B\x3C\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x3F\x3C\x00\x80\x47\x3C\x00\x80\xA1\xA5\x02\x00\xA2\x49\x07\x00\xA3\x41"
-     "\x07\x00"
-     "\xA4\x41\x06\x00\xA5\xD5\x1B\x00\xA6\xDD\x1B\x00\xA7\xC1\x1A\x00\xA8\x01"
-     "\x1C\x00"
-     "\xA9\xE1\x1F\x00\xAA\x49\x1F\x00\xAB\x01\x10\x00\xAC\xF5\x13\x00\xAD\xAD"
-     "\x13\x00"
-     "\xAE\x01\x14\x00\xAF\xF9\x17\x00\xA8\x31\x06\x00\xA9\x31\x06\x00\xAA\x4D"
-     "\x06\x00"
-     "\xAB\x45\x06\x00\xAC\x4D\x06\x00\xAD\x99\x06\x00\xAE\x8D\x06\x00\xAF\x85"
-     "\x06\x00"
-     "\x86\x80\x03\x00\x87\x18\x03\x00\x4B\x3C\x00\x80\x4F\x3C\x00\x80\x53\x3C"
-     "\x00\x80"
-     "\x57\x3C\x00\x80\x5B\x3C\x00\x80\x5F\x3C\x00\x80\xB8\x6D\x07\x00\xB9\x75"
-     "\x07\x00"
-     "\xBA\x7D\x07\x00\xBB\x75\x07\x00\xBC\x6D\x07\x00\xBD\xCD\x07\x00\xBE\xF9"
-     "\x07\x00"
-     "\xBF\xF9\x07\x00\xB0\xA9\x06\x00\xB1\x85\x06\x00\xB2\x79\x07\x00\xB3\x79"
-     "\x07\x00"
-     "\xB4\x69\x07\x00\xB5\x69\x07\x00\xB6\x5D\x07\x00\xB7\x55\x07\x00\xD3\x3B"
-     "\x00\x80"
-     "\xB3\xC1\x06\x00\x63\x3C\x00\x80\x43\x3C\x00\x80\xB6\x1D\x06\x00\x67\x3C"
-     "\x00\x80"
-     "\x6B\x3C\x00\x80\xB5\xC1\x06\x00\xBA\x69\x06\x00\xBB\x45\x06\x00\x6F\x3C"
-     "\x00\x80"
-     "\x73\x3C\x00\x80\xBE\xA9\x07\x00\xBF\xA9\x07\x00\xBC\xA9\x07\x00\xBD\xA9"
-     "\x07\x00"
-     "\xA3\x85\x06\x00\x77\x3C\x00\x80\x7B\x3C\x00\x80\x7F\x3C\x00\x80\x83\x3C"
-     "\x00\x80"
-     "\xA6\x59\x06\x00\xA5\x85\x06\x00\x87\x3C\x00\x80\xAB\x01\x06\x00\xAA\x2D"
-     "\x06\x00"
-     "\x8B\x3C\x00\x80\x8F\x3C\x00\x80\xAF\xED\x07\x00\xAE\xED\x07\x00\xAD\xED"
-     "\x07\x00"
-     "\xAC\xED\x07\x00\xA8\xC1\x06\x00\xA9\x2D\x01\x00\xAA\x25\x01\x00\xAB\x3D"
-     "\x01\x00"
-     "\xAC\x25\x01\x00\xAD\x2D\x01\x00\xAE\x25\x01\x00\xAF\x95\x01\x00\x93\x3C"
-     "\x00\x80"
-     "\x97\x3C\x00\x80\x9B\x3C\x00\x80\x9F\x3C\x00\x80\xA3\x3C\x00\x80\x82\xBD"
-     "\x00\x00"
-     "\x81\xBD\x00\x00\x80\xBD\x00\x00\xB8\x9D\x01\x00\xB9\xAD\x01\x00\xBA\xA5"
-     "\x01\x00"
-     "\xBB\x6D\x00\x00\xBC\x75\x00\x00\xBD\x7D\x00\x00\xBE\x75\x00\x00\xBF\x6D"
-     "\x00\x00"
-     "\xB0\xF5\x01\x00\xB1\xFD\x01\x00\xB2\xC1\x01\x00\xB3\xC1\x01\x00\xB4\xB5"
-     "\x01\x00"
-     "\xB5\xBD\x01\x00\xB6\xB5\x01\x00\xB7\xAD\x01\x00\xA7\x3C\x00\x80\xAB\x3C"
-     "\x00\x80"
-     "\xAF\x3C\x00\x80\xB3\xA1\x01\x00\xB3\x3C\x00\x80\xB5\xA1\x01\x00\xB6\xA1"
-     "\x01\x00"
-     "\xB7\x3C\x00\x80\x86\x80\x01\x00\x87\xC4\x01\x00\xBA\x3D\x01\x00\xBB\x35"
-     "\x01\x00"
-     "\xBC\x19\x01\x00\xBD\x19\x01\x00\xBE\x7D\x01\x00\xBF\x75\x01\x00\xA3\xED"
-     "\x01\x00"
-     "\xBB\x3C\x00\x80\xBF\x3C\x00\x80\xC3\x3C\x00\x80\xC7\x3C\x00\x80\xA6\xED"
-     "\x01\x00"
-     "\xA5\xED\x01\x00\xCB\x3C\x00\x80\xAB\x79\x01\x00\xAA\x71\x01\x00\xCF\x3C"
-     "\x00\x80"
-     "\xD3\x3C\x00\x80\xAF\x39\x01\x00\xAE\x31\x01\x00\xAD\x55\x01\x00\xAC\x55"
-     "\x01\x00"
-     "\xD7\x3C\x00\x80\xDB\x3C\x00\x80\xDF\x3C\x00\x80\xE3\x3C\x00\x80\xE7\x3C"
-     "\x00\x80"
-     "\xE1\xAC\x01\x00\xEB\x3C\x00\x80\xE3\x60\x06\x00\xEF\x3C\x00\x80\xF3\x3C"
-     "\x00\x80"
-     "\xF7\x3C\x00\x80\xEF\x54\x06\x00\xFB\x3C\x00\x80\xFF\x3C\x00\x80\xBE\xB4"
-     "\x1A\x00"
-     "\x03\x3D\x00\x80\x07\x3D\x00\x80\x0B\x3D\x00\x80\x86\x68\x1C\x00\x87\x8C"
-     "\x03\x00"
-     "\x0F\x3D\x00\x80\x13\x3D\x00\x80\x17\x3D\x00\x80\x1B\x3D\x00\x80\x80\x39"
-     "\x00\x00"
-     "\x81\x39\x00\x00\x82\x05\x00\x00\x1F\x3D\x00\x80\x27\x3D\x00\x80\x2B\x3D"
-     "\x00\x80"
-     "\x2F\x3D\x00\x80\x33\x3D\x00\x80\xA8\x1D\x03\x00\xA9\x41\x03\x00\xAA\x41"
-     "\x03\x00"
-     "\xAB\x41\x03\x00\xAC\x41\x03\x00\xAD\x49\x03\x00\xAE\x71\x03\x00\xAF\x71"
-     "\x03\x00"
-     "\x84\x20\x1D\x00\x37\x3D\x00\x80\x3B\x3D\x00\x80\x3F\x3D\x00\x80\x43\x3D"
-     "\x00\x80"
-     "\x47\x3D\x00\x80\x4B\x3D\x00\x80\x4F\x3D\x00\x80\xB8\xE9\x00\x00\xB9\xF5"
-     "\x00\x00"
-     "\xBA\xFD\x00\x00\xBB\xF1\x00\x00\xBC\x91\x00\x00\xBD\x91\x00\x00\xBE\x89"
-     "\x00\x00"
-     "\xBF\x89\x00\x00\xB0\xE1\x00\x00\xB1\xE1\x00\x00\xB2\xE1\x00\x00\xB3\xE1"
-     "\x00\x00"
-     "\xB4\xE1\x00\x00\xB5\xED\x00\x00\xB6\xD9\x00\x00\xB7\xD9\x00\x00\xE3\x0C"
-     "\x07\x00"
-     "\xE1\x20\x07\x00\xE1\x30\x01\x00\xE3\x08\x07\x00\x53\x3D\x00\x80\x57\x3D"
-     "\x00\x80"
-     "\x5B\x3D\x00\x80\x5F\x3D\x00\x80\x63\x3D\x00\x80\x67\x3D\x00\x80\x6B\x3D"
-     "\x00\x80"
-     "\x6F\x3D\x00\x80\xEF\x98\x07\x00\x73\x3D\x00\x80\x77\x3D\x00\x80\xEF\x88"
-     "\x07\x00"
-     "\xB3\x89\x02\x00\x7B\x3D\x00\x80\x7F\x3D\x00\x80\xBE\x80\x1A\x00\x83\x3D"
-     "\x00\x80"
-     "\xB6\x89\x02\x00\xB5\x89\x02\x00\x87\x3D\x00\x80\xBB\x65\x01\x00\xBA\x65"
-     "\x01\x00"
-     "\x8B\x3D\x00\x80\x8F\x3D\x00\x80\xBF\x69\x01\x00\xBE\x65\x01\x00\xBD\x75"
-     "\x01\x00"
-     "\xBC\x75\x01\x00\xB7\x3D\x06\x00\xB6\x3D\x06\x00\xB5\x3D\x06\x00\xB4\x21"
-     "\x06\x00"
-     "\xB3\x35\x06\x00\xB2\x35\x06\x00\xB1\x01\x06\x00\xB0\x09\x06\x00\xBF\x59"
-     "\x06\x00"
-     "\xBE\x51\x06\x00\xBD\x59\x06\x00\xBC\x4D\x06\x00\xBB\x6D\x06\x00\xBA\x79"
-     "\x06\x00"
-     "\xB9\x71\x06\x00\xB8\x79\x06\x00\x80\x9D\x00\x00\x81\xAD\x00\x00\x82\xA5"
-     "\x00\x00"
-     "\x97\x3D\x00\x80\x9B\x3D\x00\x80\x9F\x3D\x00\x80\xA3\x3D\x00\x80\xA7\x3D"
-     "\x00\x80"
-     "\xAF\x71\x06\x00\xAE\x69\x06\x00\xAD\x6D\x06\x00\xAC\x6D\x06\x00\xAB\x81"
-     "\x06\x00"
-     "\xAA\x99\x06\x00\xA9\x91\x06\x00\xA8\x99\x06\x00\x23\x3D\x00\x80\x93\x3D"
-     "\x00\x80"
-     "\xAB\x3D\x00\x80\xA3\xC5\x1D\x00\xAF\x3D\x00\x80\xA5\xC5\x1D\x00\xA6\xC5"
-     "\x1D\x00"
-     "\xB3\x3D\x00\x80\x86\x00\x03\x00\x87\x64\x03\x00\xAA\x29\x1E\x00\xAB\x29"
-     "\x1E\x00"
-     "\xAC\x39\x1E\x00\xAD\x39\x1E\x00\xAE\x29\x1E\x00\xAF\x25\x1E\x00\xB3\x39"
-     "\x1E\x00"
-     "\xB7\x3D\x00\x80\xBB\x3D\x00\x80\xBF\x3D\x00\x80\xC3\x3D\x00\x80\xB6\xFD"
-     "\x1E\x00"
-     "\xB5\xFD\x1E\x00\xC7\x3D\x00\x80\xBB\xD9\x1E\x00\xBA\xD1\x1E\x00\xCB\x3D"
-     "\x00\x80"
-     "\xCF\x3D\x00\x80\xBF\x69\x1F\x00\xBE\x61\x1F\x00\xBD\x69\x1F\x00\xBC\xC1"
-     "\x1E\x00"
-     "\xA8\xF1\x1E\x00\xA9\xF1\x1E\x00\xAA\xF1\x1E\x00\xAB\xF1\x1E\x00\xAC\x35"
-     "\x1E\x00"
-     "\xAD\x3D\x1E\x00\xAE\x35\x1E\x00\xAF\x2D\x1E\x00\xD3\x3D\x00\x80\xD7\x3D"
-     "\x00\x80"
-     "\xDB\x3D\x00\x80\xDF\x3D\x00\x80\xE3\x3D\x00\x80\xE7\x3D\x00\x80\xEB\x3D"
-     "\x00\x80"
-     "\xEF\x3D\x00\x80\xB8\xE5\x1F\x00\xB9\xED\x1F\x00\xBA\xE5\x1F\x00\xBB\xF9"
-     "\x1F\x00"
-     "\xBC\xE9\x1F\x00\xBD\x91\x1F\x00\xBE\x91\x1F\x00\xBF\x8D\x1F\x00\xB0\x55"
-     "\x1E\x00"
-     "\xB1\x5D\x1E\x00\xB2\x55\x1E\x00\xB3\xFD\x1F\x00\xB4\xE5\x1F\x00\xB5\xED"
-     "\x1F\x00"
-     "\xB6\xE5\x1F\x00\xB7\xDD\x1F\x00\xA3\x79\x1F\x00\xF3\x3D\x00\x80\xF7\x3D"
-     "\x00\x80"
-     "\xFB\x3D\x00\x80\xFF\x3D\x00\x80\xA6\xBD\x1F\x00\xA5\xBD\x1F\x00\x03\x3E"
-     "\x00\x80"
-     "\xAB\x99\x1F\x00\xAA\x91\x1F\x00\x86\x88\x00\x00\x87\x4C\x01\x00\xAF\x29"
-     "\x1E\x00"
-     "\xAE\x21\x1E\x00\xAD\x29\x1E\x00\xAC\x81\x1F\x00\x80\x49\x00\x00\x81\x49"
-     "\x00\x00"
-     "\x82\x59\x00\x00\xB3\x99\x1E\x00\x07\x3E\x00\x80\xB5\x89\x1E\x00\xB6\x51"
-     "\x01\x00"
-     "\x0B\x3E\x00\x80\x0F\x3E\x00\x80\x13\x3E\x00\x80\xBA\x2D\x01\x00\xBB\x25"
-     "\x01\x00"
-     "\xBC\x3D\x01\x00\xBD\x25\x01\x00\xBE\x25\x01\x00\xBF\x15\x01\x00\xA8\x4D"
-     "\x1E\x00"
-     "\xA9\x55\x1E\x00\xAA\x5D\x1E\x00\xAB\x55\x1E\x00\xAC\x4D\x1E\x00\xAD\x9D"
-     "\x01\x00"
-     "\xAE\x89\x01\x00\xAF\x81\x01\x00\x84\xAC\x01\x00\x17\x3E\x00\x80\x1B\x3E"
-     "\x00\x80"
-     "\x1F\x3E\x00\x80\x23\x3E\x00\x80\x27\x3E\x00\x80\x2B\x3E\x00\x80\x2F\x3E"
-     "\x00\x80"
-     "\xB8\x65\x01\x00\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x7D\x01\x00\xBC\x65"
-     "\x01\x00"
-     "\xBD\x6D\x01\x00\xBE\x65\x01\x00\xBF\xD9\x00\x00\xB0\xA5\x01\x00\xB1\xAD"
-     "\x01\x00"
-     "\xB2\xA5\x01\x00\xB3\xBD\x01\x00\xB4\xAD\x01\x00\xB5\x9D\x01\x00\xB6\x95"
-     "\x01\x00"
-     "\xB7\x5D\x01\x00\xA3\xD5\x1D\x00\x33\x3E\x00\x80\x37\x3E\x00\x80\x3B\x3E"
-     "\x00\x80"
-     "\x3F\x3E\x00\x80\xA6\x1D\x02\x00\xA5\xC5\x1D\x00\x43\x3E\x00\x80\xAB\x69"
-     "\x02\x00"
-     "\xAA\x61\x02\x00\x47\x3E\x00\x80\x4B\x3E\x00\x80\xAF\x59\x02\x00\xAE\x69"
-     "\x02\x00"
-     "\xAD\x69\x02\x00\xAC\x71\x02\x00\x4F\x3E\x00\x80\x53\x3E\x00\x80\x57\x3E"
-     "\x00\x80"
-     "\x5B\x3E\x00\x80\x5F\x3E\x00\x80\x63\x3E\x00\x80\x67\x3E\x00\x80\x6B\x3E"
-     "\x00\x80"
-     "\x80\x39\x00\x00\x81\x39\x00\x00\x82\x05\x00\x00\x6F\x3E\x00\x80\x77\x3E"
-     "\x00\x80"
-     "\x7B\x3E\x00\x80\x87\x40\x03\x00\x86\x5C\x04\x00\x84\x4C\x04\x00\x7F\x3E"
-     "\x00\x80"
-     "\x83\x3E\x00\x80\x84\x04\x05\x00\xE3\x20\x01\x00\x87\x3E\x00\x80\xE1\xA8"
-     "\x01\x00"
-     "\x8B\x3E\x00\x80\xEF\x94\x1A\x00\x8F\x3E\x00\x80\x93\x3E\x00\x80\x97\x3E"
-     "\x00\x80"
-     "\x9B\x3E\x00\x80\x9F\x3E\x00\x80\xA3\x3E\x00\x80\xA7\x3E\x00\x80\xB3\xA1"
-     "\x03\x00"
-     "\xAB\x3E\x00\x80\xAF\x3E\x00\x80\xB3\x3E\x00\x80\xB7\x3E\x00\x80\xB6\x7D"
-     "\x03\x00"
-     "\xB5\x7D\x03\x00\xBB\x3E\x00\x80\xBB\x59\x03\x00\xBA\x51\x03\x00\xBF\x3E"
-     "\x00\x80"
-     "\xC3\x3E\x00\x80\xBF\xFD\x00\x00\xBE\xFD\x00\x00\xBD\xFD\x00\x00\xBC\x41"
-     "\x03\x00"
-     "\xA8\x51\x02\x00\xA9\x59\x02\x00\xAA\x69\x02\x00\xAB\x69\x02\x00\xAC\xB5"
-     "\x02\x00"
-     "\xAD\xBD\x02\x00\xAE\xB5\x02\x00\xAF\xAD\x02\x00\x84\xA8\x07\x00\xC7\x3E"
-     "\x00\x80"
-     "\xCB\x3E\x00\x80\xCF\x3E\x00\x80\x82\xA9\x00\x00\xD3\x3E\x00\x80\x80\xA9"
-     "\x00\x00"
-     "\x81\xA9\x00\x00\xB8\x69\x01\x00\xB9\x69\x01\x00\xBA\x09\x01\x00\xBB\x09"
-     "\x01\x00"
-     "\xBC\x19\x01\x00\xBD\x19\x01\x00\xBE\x09\x01\x00\xBF\x09\x01\x00\xB0\xD5"
-     "\x02\x00"
-     "\xB1\xDD\x02\x00\xB2\xD5\x02\x00\xB3\x69\x01\x00\xB4\x79\x01\x00\xB5\x79"
-     "\x01\x00"
-     "\xB6\x69\x01\x00\xB7\x61\x01\x00\xE1\xB8\x01\x00\xE1\xD4\x1F\x00\xE3\x38"
-     "\x1F\x00"
-     "\xE3\x0C\x1B\x00\xD7\x3E\x00\x80\xDB\x3E\x00\x80\xDF\x3E\x00\x80\xE7\x3E"
-     "\x00\x80"
-     "\xEB\x3E\x00\x80\xEF\x3E\x00\x80\xF3\x3E\x00\x80\xF7\x3E\x00\x80\xBE\x3C"
-     "\x09\x00"
-     "\xFB\x3E\x00\x80\xEF\x84\x1B\x00\xEF\x84\x1B\x00\xA3\xA1\x02\x00\xFF\x3E"
-     "\x00\x80"
-     "\x86\xE8\x04\x00\x87\xBC\x05\x00\x03\x3F\x00\x80\xA6\x7D\x02\x00\xA5\x7D"
-     "\x02\x00"
-     "\x07\x3F\x00\x80\xAB\x59\x02\x00\xAA\x51\x02\x00\x0B\x3F\x00\x80\x0F\x3F"
-     "\x00\x80"
-     "\xAF\xFD\x01\x00\xAE\xFD\x01\x00\xAD\xFD\x01\x00\xAC\x41\x02\x00\xB3\x85"
-     "\x06\x00"
-     "\xE3\x3E\x00\x80\x13\x3F\x00\x80\x17\x3F\x00\x80\x1B\x3F\x00\x80\xB6\x89"
-     "\x06\x00"
-     "\xB5\x8D\x06\x00\x1F\x3F\x00\x80\xBB\x91\x06\x00\xBA\x89\x06\x00\x23\x3F"
-     "\x00\x80"
-     "\x27\x3F\x00\x80\xBF\xF5\x06\x00\xBE\x89\x06\x00\xBD\x81\x06\x00\xBC\x89"
-     "\x06\x00"
-     "\x2B\x3F\x00\x80\x2F\x3F\x00\x80\x33\x3F\x00\x80\x37\x3F\x00\x80\x3B\x3F"
-     "\x00\x80"
-     "\x3F\x3F\x00\x80\x43\x3F\x00\x80\xEF\x84\x1D\x00\x47\x3F\x00\x80\xE1\x00"
-     "\x04\x00"
-     "\x4B\x3F\x00\x80\xE3\xFC\x04\x00\x80\x11\x00\x00\x81\x1D\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x4F\x3F\x00\x80\xA8\xF1\x06\x00\xA9\xF1\x06\x00\xAA\x0D\x06\x00\xAB\x05"
-     "\x06\x00"
-     "\xAC\x05\x06\x00\xAD\x09\x06\x00\xAE\x3D\x06\x00\xAF\x35\x06\x00\x53\x3F"
-     "\x00\x80"
-     "\x57\x3F\x00\x80\x86\x88\x00\x00\x87\x10\x03\x00\x5B\x3F\x00\x80\x5F\x3F"
-     "\x00\x80"
-     "\x63\x3F\x00\x80\x67\x3F\x00\x80\xB8\x11\x06\x00\xB9\x19\x06\x00\xBA\x21"
-     "\x06\x00"
-     "\xBB\x21\x06\x00\xBC\xF5\x07\x00\xBD\xFD\x07\x00\xBE\xF5\x07\x00\xBF\xE9"
-     "\x07\x00"
-     "\xB0\x4D\x06\x00\xB1\x55\x06\x00\xB2\x5D\x06\x00\xB3\x55\x06\x00\xB4\x4D"
-     "\x06\x00"
-     "\xB5\x31\x06\x00\xB6\x31\x06\x00\xB7\x31\x06\x00\xA3\x85\x07\x00\x6B\x3F"
-     "\x00\x80"
-     "\x6F\x3F\x00\x80\x73\x3F\x00\x80\x77\x3F\x00\x80\xA6\x89\x07\x00\xA5\x8D"
-     "\x07\x00"
-     "\x7B\x3F\x00\x80\xAB\x91\x07\x00\xAA\x89\x07\x00\x7F\x3F\x00\x80\x83\x3F"
-     "\x00\x80"
-     "\xAF\xF5\x07\x00\xAE\x89\x07\x00\xAD\x81\x07\x00\xAC\x89\x07\x00\x87\x3F"
-     "\x00\x80"
-     "\xB3\xE1\x06\x00\x8B\x3F\x00\x80\x8F\x3F\x00\x80\xB6\xE5\x06\x00\x93\x3F"
-     "\x00\x80"
-     "\x97\x3F\x00\x80\xB5\xF1\x06\x00\xBA\xBD\x06\x00\xBB\x8D\x06\x00\x9B\x3F"
-     "\x00\x80"
-     "\x9F\x3F\x00\x80\xBE\x7D\x01\x00\xBF\x65\x01\x00\xBC\x95\x06\x00\xBD\x75"
-     "\x01\x00"
-     "\xA8\x1D\x06\x00\xA9\x25\x06\x00\xAA\x2D\x06\x00\xAB\x25\x06\x00\xAC\x3D"
-     "\x06\x00"
-     "\xAD\x21\x06\x00\xAE\x5D\x06\x00\xAF\x4D\x06\x00\xA3\x3F\x00\x80\xA7\x3F"
-     "\x00\x80"
-     "\xAB\x3F\x00\x80\xAF\x3F\x00\x80\xB3\x3F\x00\x80\x82\xB9\x03\x00\x81\xB1"
-     "\x03\x00"
-     "\x80\xB9\x03\x00\xB8\xED\x01\x00\xB9\x85\x01\x00\xBA\x8D\x01\x00\xBB\x85"
-     "\x01\x00"
-     "\xBC\x9D\x01\x00\xBD\x85\x01\x00\xBE\x8D\x01\x00\xBF\x85\x01\x00\xB0\x3D"
-     "\x06\x00"
-     "\xB1\x0D\x06\x00\xB2\x05\x06\x00\xB3\xE5\x01\x00\xB4\xFD\x01\x00\xB5\xE5"
-     "\x01\x00"
-     "\xB6\xE5\x01\x00\xB7\xD5\x01\x00\xA3\xA5\x05\x00\xB7\x3F\x00\x80\xBB\x3F"
-     "\x00\x80"
-     "\xBF\x3F\x00\x80\xC7\x3F\x00\x80\xA6\xA1\x05\x00\xA5\xB5\x05\x00\xCB\x3F"
-     "\x00\x80"
-     "\xAB\xC9\x05\x00\xAA\xF9\x05\x00\x86\x08\x0C\x00\x87\x1C\x03\x00\xAF\x21"
-     "\x02\x00"
-     "\xAE\x39\x02\x00\xAD\x31\x02\x00\xAC\xD1\x05\x00\xCF\x3F\x00\x80\xB3\xF1"
-     "\x02\x00"
-     "\xD3\x3F\x00\x80\xD7\x3F\x00\x80\xB6\x55\x03\x00\xDB\x3F\x00\x80\xDF\x3F"
-     "\x00\x80"
-     "\xB5\xE1\x02\x00\xBA\x71\x03\x00\xBB\x79\x03\x00\xE3\x3F\x00\x80\xE7\x3F"
-     "\x00\x80"
-     "\xBE\x31\x03\x00\xBF\x3D\x03\x00\xBC\x51\x03\x00\xBD\x51\x03\x00\xA8\x25"
-     "\x02\x00"
-     "\xA9\x35\x02\x00\xAA\x3D\x02\x00\xAB\x35\x02\x00\xAC\x2D\x02\x00\xAD\x91"
-     "\x03\x00"
-     "\xAE\x91\x03\x00\xAF\x91\x03\x00\xEB\x3F\x00\x80\xEF\x3F\x00\x80\xF3\x3F"
-     "\x00\x80"
-     "\xF7\x3F\x00\x80\xAC\x00\x00\x00\xFB\x3F\x00\x80\xFF\x3F\x00\x80\x03\x40"
-     "\x00\x80"
-     "\xB8\x99\x03\x00\xB9\xAD\x03\x00\xBA\xA5\x03\x00\xBB\x6D\x03\x00\xBC\x75"
-     "\x03\x00"
-     "\xBD\x7D\x03\x00\xBE\x75\x03\x00\xBF\x6D\x03\x00\xB0\xF1\x03\x00\xB1\xF9"
-     "\x03\x00"
-     "\xB2\xC1\x03\x00\xB3\xC1\x03\x00\xB4\xB1\x03\x00\xB5\xBD\x03\x00\xB6\xB5"
-     "\x03\x00"
-     "\xB7\xA9\x03\x00\x07\x40\x00\x80\x0B\x40\x00\x80\x0F\x40\x00\x80\x13\x40"
-     "\x00\x80"
-     "\x17\x40\x00\x80\x1B\x40\x00\x80\x1F\x40\x00\x80\xEF\xA8\x1A\x00\xBE\x68"
-     "\x0C\x00"
-     "\xE1\x94\x01\x00\x23\x40\x00\x80\xE3\x1C\x06\x00\x80\x0D\x00\x00\x81\x71"
-     "\x00\x00"
-     "\x82\x71\x00\x00\x27\x40\x00\x80\xA3\xF5\x03\x00\x2B\x40\x00\x80\x2F\x40"
-     "\x00\x80"
-     "\x84\x4C\x02\x00\x37\x40\x00\x80\xA6\x51\x02\x00\xA5\xE5\x03\x00\x3B\x40"
-     "\x00\x80"
-     "\xAB\x7D\x02\x00\xAA\x75\x02\x00\x86\xC8\x0C\x00\x87\x2C\x0D\x00\xAF\x39"
-     "\x02\x00"
-     "\xAE\x35\x02\x00\xAD\x55\x02\x00\xAC\x55\x02\x00\xE1\x50\x06\x00\x3F\x40"
-     "\x00\x80"
-     "\xE3\x14\x07\x00\x84\xC0\x0C\x00\x43\x40\x00\x80\x47\x40\x00\x80\x4B\x40"
-     "\x00\x80"
-     "\x4F\x40\x00\x80\x53\x40\x00\x80\x57\x40\x00\x80\x5B\x40\x00\x80\x5F\x40"
-     "\x00\x80"
-     "\x63\x40\x00\x80\x67\x40\x00\x80\xEF\x78\x1B\x00\xBE\xF8\x0F\x00\x6B\x40"
-     "\x00\x80"
-     "\x6F\x40\x00\x80\x73\x40\x00\x80\xB3\x8D\x01\x00\x77\x40\x00\x80\xB5\x99"
-     "\x01\x00"
-     "\xB6\x8D\x01\x00\x7B\x40\x00\x80\x73\x3E\x00\x80\x7F\x40\x00\x80\xBA\x85"
-     "\x01\x00"
-     "\xBB\x4D\x01\x00\xBC\x55\x01\x00\xBD\x5D\x01\x00\xBE\x55\x01\x00\xBF\x49"
-     "\x01\x00"
-     "\xA3\x41\x0E\x00\x33\x40\x00\x80\x83\x40\x00\x80\x87\x40\x00\x80\x8B\x40"
-     "\x00\x80"
-     "\xA6\x41\x0E\x00\xA5\x55\x0E\x00\x8F\x40\x00\x80\xAB\x81\x0E\x00\xAA\x49"
-     "\x0E\x00"
-     "\x93\x40\x00\x80\x97\x40\x00\x80\xAF\x85\x0E\x00\xAE\x99\x0E\x00\xAD\x91"
-     "\x0E\x00"
-     "\xAC\x99\x0E\x00\x80\x6D\x00\x00\x81\x09\x00\x00\x82\x1D\x00\x00\x9B\x40"
-     "\x00\x80"
-     "\xEF\x18\x09\x00\x9F\x40\x00\x80\xA3\x40\x00\x80\xA7\x40\x00\x80\xE3\x3C"
-     "\x0D\x00"
-     "\xAB\x40\x00\x80\xE1\x58\x0C\x00\xAF\x40\x00\x80\x86\xD0\x00\x00\x87\xBC"
-     "\x03\x00"
-     "\xB3\x40\x00\x80\xB7\x40\x00\x80\xA8\x91\x0E\x00\xA9\x99\x0E\x00\xAA\xC9"
-     "\x0E\x00"
-     "\xAB\xC5\x0E\x00\xAC\xDD\x0E\x00\xAD\xC1\x0E\x00\xAE\xC1\x0E\x00\xAF\xF5"
-     "\x0E\x00"
-     "\x84\xE8\x00\x00\xBB\x40\x00\x80\xBF\x40\x00\x80\xC3\x40\x00\x80\xC7\x40"
-     "\x00\x80"
-     "\xCB\x40\x00\x80\xCF\x40\x00\x80\xD3\x40\x00\x80\xB8\xC1\x0F\x00\xB9\xC1"
-     "\x0F\x00"
-     "\xBA\xC1\x0F\x00\xBB\xC1\x0F\x00\xBC\xC1\x0F\x00\xBD\xC1\x0F\x00\xBE\xC1"
-     "\x0F\x00"
-     "\xBF\xF5\x0F\x00\xB0\x8D\x0E\x00\xB1\x45\x0E\x00\xB2\x4D\x0E\x00\xB3\x45"
-     "\x0E\x00"
-     "\xB4\x5D\x0E\x00\xB5\x41\x0E\x00\xB6\x41\x0E\x00\xB7\x41\x0E\x00\xA8\x51"
-     "\x0E\x00"
-     "\xA9\x59\x0E\x00\xAA\x8D\x0E\x00\xAB\x9D\x0E\x00\xAC\x85\x0E\x00\xAD\x8D"
-     "\x0E\x00"
-     "\xAE\x85\x0E\x00\xAF\xBD\x0E\x00\xD7\x40\x00\x80\xDB\x40\x00\x80\xDF\x40"
-     "\x00\x80"
-     "\xE3\x40\x00\x80\xE7\x40\x00\x80\xEB\x40\x00\x80\xEF\x40\x00\x80\xF3\x40"
-     "\x00\x80"
-     "\xB8\x91\x0E\x00\xB9\x99\x0E\x00\xBA\xAD\x0E\x00\xBB\x45\x01\x00\xBC\x5D"
-     "\x01\x00"
-     "\xBD\x45\x01\x00\xBE\x45\x01\x00\xBF\x75\x01\x00\xB0\xC5\x0E\x00\xB1\xCD"
-     "\x0E\x00"
-     "\xB2\xC5\x0E\x00\xB3\xDD\x0E\x00\xB4\xC5\x0E\x00\xB5\xB5\x0E\x00\xB6\xBD"
-     "\x0E\x00"
-     "\xB7\xB5\x0E\x00\xB3\x05\x0E\x00\xF7\x40\x00\x80\xFB\x40\x00\x80\xFF\x40"
-     "\x00\x80"
-     "\x03\x41\x00\x80\xB6\x0D\x0E\x00\xB5\x0D\x0E\x00\x07\x41\x00\x80\xBB\x09"
-     "\x0E\x00"
-     "\xBA\x01\x0E\x00\x0B\x41\x00\x80\x0F\x41\x00\x80\xBF\x71\x0E\x00\xBE\x01"
-     "\x0E\x00"
-     "\xBD\x09\x0E\x00\xBC\x11\x0E\x00\x82\x6D\x00\x00\xA3\x41\x0E\x00\x80\x55"
-     "\x00\x00"
-     "\x81\x65\x00\x00\xA6\x49\x0E\x00\x17\x41\x00\x80\x1B\x41\x00\x80\xA5\x49"
-     "\x0E\x00"
-     "\xAA\x45\x0E\x00\xAB\x4D\x0E\x00\x86\x48\x00\x00\x87\x78\x00\x00\xAE\x45"
-     "\x0E\x00"
-     "\xAF\x35\x0E\x00\xAC\x55\x0E\x00\xAD\x4D\x0E\x00\xA8\x5D\x02\x00\xA9\x61"
-     "\x02\x00"
-     "\xAA\x65\x02\x00\xAB\x75\x02\x00\xAC\x6D\x02\x00\xAD\xB1\x02\x00\xAE\xB1"
-     "\x02\x00"
-     "\xAF\xB1\x02\x00\x84\xEC\x04\x00\x1F\x41\x00\x80\x23\x41\x00\x80\x27\x41"
-     "\x00\x80"
-     "\x2B\x41\x00\x80\x2F\x41\x00\x80\x33\x41\x00\x80\x37\x41\x00\x80\xB8\x71"
-     "\x03\x00"
-     "\xB9\x71\x03\x00\xBA\x71\x03\x00\xBB\x71\x03\x00\xBC\xD5\x03\x00\xBD\xDD"
-     "\x03\x00"
-     "\xBE\xD5\x03\x00\xBF\xCD\x03\x00\xB0\xD1\x02\x00\xB1\xD1\x02\x00\xB2\xD1"
-     "\x02\x00"
-     "\xB3\xD1\x02\x00\xB4\x51\x03\x00\xB5\x51\x03\x00\xB6\x51\x03\x00\xB7\x51"
-     "\x03\x00"
-     "\x3B\x41\x00\x80\xB3\xE9\x02\x00\x3F\x41\x00\x80\xBE\x80\x04\x00\xB6\x35"
-     "\x02\x00"
-     "\x43\x41\x00\x80\x47\x41\x00\x80\xB5\xE1\x02\x00\xBA\x11\x02\x00\xBB\x11"
-     "\x02\x00"
-     "\x4B\x41\x00\x80\x4F\x41\x00\x80\xBE\x91\x03\x00\xBF\x91\x03\x00\xBC\x01"
-     "\x02\x00"
-     "\xBD\x01\x02\x00\x53\x41\x00\x80\x57\x41\x00\x80\xA3\xA5\x02\x00\x5B\x41"
-     "\x00\x80"
-     "\xA5\xAD\x02\x00\x5F\x41\x00\x80\x63\x41\x00\x80\xA6\x79\x02\x00\x67\x41"
-     "\x00\x80"
-     "\x6B\x41\x00\x80\xAB\x5D\x02\x00\xAA\x5D\x02\x00\xAD\x4D\x02\x00\xAC\x4D"
-     "\x02\x00"
-     "\xAF\xDD\x03\x00\xAE\xDD\x03\x00\xA8\xD5\x02\x00\xA9\xDD\x02\x00\xAA\x2D"
-     "\x01\x00"
-     "\xAB\x25\x01\x00\xAC\x3D\x01\x00\xAD\x25\x01\x00\xAE\x2D\x01\x00\xAF\x25"
-     "\x01\x00"
-     "\x6F\x41\x00\x80\x73\x41\x00\x80\x77\x41\x00\x80\x7B\x41\x00\x80\x7F\x41"
-     "\x00\x80"
-     "\x87\x41\x00\x80\x8B\x41\x00\x80\x8F\x41\x00\x80\xB8\x85\x01\x00\xB9\x89"
-     "\x01\x00"
-     "\xBA\x9D\x01\x00\xBB\x95\x01\x00\xBC\xB1\x01\x00\xBD\xB1\x01\x00\xBE\x79"
-     "\x00\x00"
-     "\xBF\x79\x00\x00\xB0\x5D\x01\x00\xB1\xE5\x01\x00\xB2\xE1\x01\x00\xB3\xF9"
-     "\x01\x00"
-     "\xB4\xE9\x01\x00\xB5\xDD\x01\x00\xB6\xD5\x01\x00\xB7\xBD\x01\x00\xE1\xF0"
-     "\x0E\x00"
-     "\x93\x41\x00\x80\xE3\x14\x0E\x00\x97\x41\x00\x80\x81\xBD\x00\x00\x80\xBD"
-     "\x00\x00"
-     "\x9B\x41\x00\x80\x82\xAD\x00\x00\x86\x00\x04\x00\x87\xEC\x05\x00\x9F\x41"
-     "\x00\x80"
-     "\xA3\x41\x00\x80\xA7\x41\x00\x80\xAB\x41\x00\x80\xEF\x60\x0E\x00\xAF\x41"
-     "\x00\x80"
-     "\xB3\x41\x00\x80\xB7\x41\x00\x80\x85\x5C\x7D\x00\xBB\x41\x00\x80\xBF\x41"
-     "\x00\x80"
-     "\xE3\x64\x01\x00\xC3\x41\x00\x80\xE1\xB4\x01\x00\xC7\x41\x00\x80\xEF\xA0"
-     "\x0E\x00"
-     "\xCB\x41\x00\x80\xC3\x3F\x00\x80\x84\xF8\x05\x00\xCF\x41\x00\x80\xD3\x41"
-     "\x00\x80"
-     "\xD7\x41\x00\x80\xB3\x25\x06\x00\x83\x41\x00\x80\xDB\x41\x00\x80\xDF\x41"
-     "\x00\x80"
-     "\xE3\x41\x00\x80\xB6\x25\x06\x00\xB5\x35\x06\x00\xE7\x41\x00\x80\xBB\xA1"
-     "\x06\x00"
-     "\xBA\x19\x06\x00\xEB\x41\x00\x80\xEF\x41\x00\x80\xBF\x99\x06\x00\xBE\xAD"
-     "\x06\x00"
-     "\xBD\xAD\x06\x00\xBC\xB5\x06\x00\x82\x6D\x00\x00\xEF\x30\x04\x00\x80\x55"
-     "\x00\x00"
-     "\x81\x65\x00\x00\xBE\x5C\x03\x00\xF3\x41\x00\x80\x86\xF8\x00\x00\x87\x6C"
-     "\x03\x00"
-     "\xF7\x41\x00\x80\xFB\x41\x00\x80\xFF\x41\x00\x80\x03\x42\x00\x80\x07\x42"
-     "\x00\x80"
-     "\xE3\x44\x04\x00\x0B\x42\x00\x80\xE1\x8C\x07\x00\xA3\xA5\x06\x00\x0F\x42"
-     "\x00\x80"
-     "\x13\x42\x00\x80\x17\x42\x00\x80\x1B\x42\x00\x80\xA6\xA5\x06\x00\xA5\xB5"
-     "\x06\x00"
-     "\x1F\x42\x00\x80\xAB\x21\x06\x00\xAA\x99\x06\x00\x23\x42\x00\x80\x27\x42"
-     "\x00\x80"
-     "\xAF\x19\x06\x00\xAE\x2D\x06\x00\xAD\x2D\x06\x00\xAC\x35\x06\x00\x2B\x42"
-     "\x00\x80"
-     "\xB3\xF9\x07\x00\x2F\x42\x00\x80\x33\x42\x00\x80\xB6\x49\x07\x00\x37\x42"
-     "\x00\x80"
-     "\x3B\x42\x00\x80\xB5\x51\x07\x00\xBA\x51\x07\x00\xBB\x51\x07\x00\x3F\x42"
-     "\x00\x80"
-     "\x43\x42\x00\x80\xBE\x35\x07\x00\xBF\x39\x07\x00\xBC\x45\x07\x00\xBD\x3D"
-     "\x07\x00"
-     "\xA8\x35\x06\x00\xA9\x3D\x06\x00\xAA\x35\x06\x00\xAB\x89\x06\x00\xAC\x9D"
-     "\x06\x00"
-     "\xAD\x8D\x06\x00\xAE\xB1\x06\x00\xAF\xB1\x06\x00\x47\x42\x00\x80\x4B\x42"
-     "\x00\x80"
-     "\x4F\x42\x00\x80\x53\x42\x00\x80\x80\x0D\x00\x00\x81\xB1\x00\x00\x82\xB1"
-     "\x00\x00"
-     "\x57\x42\x00\x80\xB8\xA1\x06\x00\xB9\xAD\x06\x00\xBA\xBD\x06\x00\xBB\xB5"
-     "\x06\x00"
-     "\xBC\xAD\x06\x00\xBD\x5D\x01\x00\xBE\x55\x01\x00\xBF\x4D\x01\x00\xB0\xD1"
-     "\x06\x00"
-     "\xB1\xD1\x06\x00\xB2\xD5\x06\x00\xB3\xAD\x06\x00\xB4\xB5\x06\x00\xB5\xB9"
-     "\x06\x00"
-     "\xB6\xA9\x06\x00\xB7\xA5\x06\x00\xA3\xBD\x06\x00\x5B\x42\x00\x80\x5F\x42"
-     "\x00\x80"
-     "\x84\x84\x02\x00\xBE\x90\x01\x00\xA6\x0D\x06\x00\xA5\x15\x06\x00\x67\x42"
-     "\x00\x80"
-     "\xAB\x15\x06\x00\xAA\x15\x06\x00\x86\x08\x00\x00\x87\x7C\x01\x00\xAF\x7D"
-     "\x06\x00"
-     "\xAE\x71\x06\x00\xAD\x79\x06\x00\xAC\x01\x06\x00\x6B\x42\x00\x80\xB3\xAD"
-     "\x01\x00"
-     "\x6F\x42\x00\x80\x73\x42\x00\x80\xB6\xA9\x01\x00\x77\x42\x00\x80\x7B\x42"
-     "\x00\x80"
-     "\xB5\xAD\x01\x00\xBA\x6D\x01\x00\xBB\x75\x01\x00\x7F\x42\x00\x80\x83\x42"
-     "\x00\x80"
-     "\xBE\x5D\x01\x00\xBF\x35\x01\x00\xBC\x65\x01\x00\xBD\x55\x01\x00\xA8\x61"
-     "\x02\x00"
-     "\xA9\x61\x02\x00\xAA\x61\x02\x00\xAB\x61\x02\x00\xAC\x61\x02\x00\xAD\x6D"
-     "\x02\x00"
-     "\xAE\x9D\x02\x00\xAF\x95\x02\x00\x87\x42\x00\x80\x8B\x42\x00\x80\x8F\x42"
-     "\x00\x80"
-     "\x93\x42\x00\x80\x97\x42\x00\x80\x9B\x42\x00\x80\x9F\x42\x00\x80\xA3\x42"
-     "\x00\x80"
-     "\xB8\x95\x02\x00\xB9\x9D\x02\x00\xBA\xA1\x02\x00\xBB\xA1\x02\x00\xBC\x71"
-     "\x03\x00"
-     "\xBD\x71\x03\x00\xBE\x71\x03\x00\xBF\x71\x03\x00\xB0\xED\x02\x00\xB1\xF5"
-     "\x02\x00"
-     "\xB2\xF5\x02\x00\xB3\xC5\x02\x00\xB4\xDD\x02\x00\xB5\xB5\x02\x00\xB6\xB1"
-     "\x02\x00"
-     "\xB7\xB1\x02\x00\xA7\x42\x00\x80\xAB\x42\x00\x80\xAF\x42\x00\x80\xA3\xE5"
-     "\x02\x00"
-     "\xB3\x42\x00\x80\xA5\xE5\x02\x00\xA6\xE1\x02\x00\xB7\x42\x00\x80\xBB\x42"
-     "\x00\x80"
-     "\xBF\x42\x00\x80\xAA\x25\x02\x00\xAB\x3D\x02\x00\xAC\x2D\x02\x00\xAD\x1D"
-     "\x02\x00"
-     "\xAE\x15\x02\x00\xAF\x7D\x02\x00\xC3\x42\x00\x80\xC7\x42\x00\x80\xCB\x42"
-     "\x00\x80"
-     "\x84\x40\x7C\x00\x80\x15\x00\x00\x81\x1D\x00\x00\x82\x05\x00\x00\xCF\x42"
-     "\x00\x80"
-     "\xBE\xEC\x7C\x00\xD7\x42\x00\x80\x86\x48\x7D\x00\x87\x08\x03\x00\xDB\x42"
-     "\x00\x80"
-     "\xDF\x42\x00\x80\xE3\x42\x00\x80\xE7\x42\x00\x80\xA8\x9D\x02\x00\xA9\xC5"
-     "\x02\x00"
-     "\xAA\xC1\x02\x00\xAB\xC1\x02\x00\xAC\xC5\x02\x00\xAD\xCD\x02\x00\xAE\xF1"
-     "\x02\x00"
-     "\xAF\xF1\x02\x00\xEB\x42\x00\x80\xEF\x42\x00\x80\xF3\x42\x00\x80\xF7\x42"
-     "\x00\x80"
-     "\xC9\x00\x00\x00\xFB\x42\x00\x80\xFF\x42\x00\x80\x03\x43\x00\x80\xB8\xC1"
-     "\x01\x00"
-     "\xB9\xC1\x01\x00\xBA\xC1\x01\x00\xBB\xDD\x01\x00\xBC\xCD\x01\x00\xBD\xF5"
-     "\x01\x00"
-     "\xBE\xFD\x01\x00\xBF\x9D\x01\x00\xB0\x41\x01\x00\xB1\x41\x01\x00\xB2\x41"
-     "\x01\x00"
-     "\xB3\x41\x01\x00\xB4\x41\x01\x00\xB5\x41\x01\x00\xB6\x41\x01\x00\xB7\x41"
-     "\x01\x00"
-     "\xE1\x38\x06\x00\x07\x43\x00\x80\xE3\x68\x06\x00\x0B\x43\x00\x80\x0F\x43"
-     "\x00\x80"
-     "\x13\x43\x00\x80\x17\x43\x00\x80\x84\x94\x7D\x00\xBE\xAC\x7C\x00\x1B\x43"
-     "\x00\x80"
-     "\x1F\x43\x00\x80\x23\x43\x00\x80\xBE\xB8\x7F\x00\x27\x43\x00\x80\xEF\x10"
-     "\x01\x00"
-     "\x2B\x43\x00\x80\x2F\x43\x00\x80\x33\x43\x00\x80\x37\x43\x00\x80\x3B\x43"
-     "\x00\x80"
-     "\xE1\x90\x01\x00\x3F\x43\x00\x80\xE3\x44\x00\x00\x47\x43\x00\x80\x81\x2D"
-     "\x00\x00"
-     "\x80\x2D\x00\x00\xEF\x80\x00\x00\x82\x39\x00\x00\x4B\x43\x00\x80\x4F\x43"
-     "\x00\x80"
-     "\x13\x41\x00\x80\x53\x43\x00\x80\xE1\xB0\x7F\x00\xD3\x42\x00\x80\xE3\xD4"
-     "\x7C\x00"
-     "\x57\x43\x00\x80\x43\x43\x00\x80\x5B\x43\x00\x80\x86\xB8\x00\x00\x87\xD4"
-     "\x02\x00"
-     "\x5F\x43\x00\x80\x63\x43\x00\x80\x67\x43\x00\x80\x6B\x43\x00\x80\x6F\x43"
-     "\x00\x80"
-     "\x73\x43\x00\x80\xEF\x20\x7C\x00\x77\x43\x00\x80\xB3\x89\x7D\x00\x7B\x43"
-     "\x00\x80"
-     "\x7F\x43\x00\x80\x83\x43\x00\x80\x87\x43\x00\x80\xB6\x8D\x7D\x00\xB5\x8D"
-     "\x7D\x00"
-     "\x8B\x43\x00\x80\xBB\x45\x7E\x00\xBA\x45\x7E\x00\x8F\x43\x00\x80\x93\x43"
-     "\x00\x80"
-     "\xBF\x45\x7E\x00\xBE\x45\x7E\x00\xBD\x55\x7E\x00\xBC\x55\x7E\x00\xA3\x49"
-     "\x7D\x00"
-     "\x97\x43\x00\x80\x9B\x43\x00\x80\x9F\x43\x00\x80\xA3\x43\x00\x80\xA6\x4D"
-     "\x7D\x00"
-     "\xA5\x4D\x7D\x00\xA7\x43\x00\x80\xAB\x85\x7E\x00\xAA\x85\x7E\x00\xAB\x43"
-     "\x00\x80"
-     "\xAF\x43\x00\x80\xAF\x85\x7E\x00\xAE\x85\x7E\x00\xAD\x95\x7E\x00\xAC\x95"
-     "\x7E\x00"
-     "\x82\x6D\x00\x00\xB3\x31\x7E\x00\x80\x55\x00\x00\x81\x65\x00\x00\xB6\xF1"
-     "\x7F\x00"
-     "\x84\xDC\x03\x00\xB3\x43\x00\x80\xB5\x21\x7E\x00\xBA\xCD\x7F\x00\xBB\xD1"
-     "\x7F\x00"
-     "\x86\x00\x04\x00\x87\xD4\x00\x00\xBE\x75\x7F\x00\xBF\x79\x7F\x00\xBC\xC1"
-     "\x7F\x00"
-     "\xBD\xC1\x7F\x00\xA8\xE5\x7F\x00\xA9\xF5\x7F\x00\xAA\xFD\x7F\x00\xAB\xF5"
-     "\x7F\x00"
-     "\xAC\xED\x7F\x00\xAD\x35\x7E\x00\xAE\x3D\x7E\x00\xAF\x35\x7E\x00\xB7\x43"
-     "\x00\x80"
-     "\xBB\x43\x00\x80\xBF\x43\x00\x80\xC3\x43\x00\x80\xC7\x43\x00\x80\xCB\x43"
-     "\x00\x80"
-     "\xCF\x43\x00\x80\xD3\x43\x00\x80\xB8\xD9\x7E\x00\xB9\xE1\x7E\x00\xBA\xE1"
-     "\x7E\x00"
-     "\xBB\xE1\x7E\x00\xBC\xE5\x7E\x00\xBD\xE9\x7E\x00\xBE\x99\x7E\x00\xBF\x99"
-     "\x7E\x00"
-     "\xB0\x4D\x7E\x00\xB1\x51\x7E\x00\xB2\x51\x7E\x00\xB3\x51\x7E\x00\xB4\xF5"
-     "\x7E\x00"
-     "\xB5\xF9\x7E\x00\xB6\xE9\x7E\x00\xB7\xE9\x7E\x00\xA3\x75\x7F\x00\xD7\x43"
-     "\x00\x80"
-     "\xDB\x43\x00\x80\xDF\x43\x00\x80\xE3\x43\x00\x80\xA6\xB5\x7E\x00\xA5\x65"
-     "\x7F\x00"
-     "\xE7\x43\x00\x80\xAB\x95\x7E\x00\xAA\x89\x7E\x00\xEB\x43\x00\x80\xEF\x43"
-     "\x00\x80"
-     "\xAF\x3D\x7E\x00\xAE\x31\x7E\x00\xAD\x85\x7E\x00\xAC\x85\x7E\x00\xF3\x43"
-     "\x00\x80"
-     "\xB3\x31\x7E\x00\xF7\x43\x00\x80\xFB\x43\x00\x80\xB6\xC5\x01\x00\xFF\x43"
-     "\x00\x80"
-     "\x03\x44\x00\x80\xB5\xD1\x01\x00\xBA\xC9\x01\x00\xBB\xC9\x01\x00\x07\x44"
-     "\x00\x80"
-     "\x0B\x44\x00\x80\xBE\xCD\x01\x00\xBF\xB1\x01\x00\xBC\xC9\x01\x00\xBD\xC9"
-     "\x01\x00"
-     "\xA8\xDD\x7D\x00\xA9\xF5\x7D\x00\xAA\xFD\x7D\x00\xAB\xF1\x7D\x00\xAC\x1D"
-     "\x02\x00"
-     "\xAD\x01\x02\x00\xAE\x39\x02\x00\xAF\x39\x02\x00\x0F\x44\x00\x80\x13\x44"
-     "\x00\x80"
-     "\x17\x44\x00\x80\x1B\x44\x00\x80\x82\x05\x00\x00\x1F\x44\x00\x80\x80\x11"
-     "\x00\x00"
-     "\x81\x11\x00\x00\xB8\x11\x02\x00\xB9\x19\x02\x00\xBA\x21\x02\x00\xBB\x21"
-     "\x02\x00"
-     "\xBC\xD5\x02\x00\xBD\xDD\x02\x00\xBE\xD5\x02\x00\xBF\xCD\x02\x00\xB0\x49"
-     "\x02\x00"
-     "\xB1\x49\x02\x00\xB2\x59\x02\x00\xB3\x59\x02\x00\xB4\x4D\x02\x00\xB5\x31"
-     "\x02\x00"
-     "\xB6\x31\x02\x00\xB7\x31\x02\x00\xBE\x00\x03\x00\xA3\x71\x7D\x00\x84\x88"
-     "\x02\x00"
-     "\xBE\x80\x04\x00\xA6\x85\x02\x00\x27\x44\x00\x80\x2B\x44\x00\x80\xA5\x91"
-     "\x02\x00"
-     "\xAA\x89\x02\x00\xAB\x89\x02\x00\x86\x28\x04\x00\x87\x0C\x03\x00\xAE\x8D"
-     "\x02\x00"
-     "\xAF\xF1\x02\x00\xAC\x89\x02\x00\xAD\x89\x02\x00\x2F\x44\x00\x80\x84\xC8"
-     "\x03\x00"
-     "\x85\xCC\x05\x00\xB3\xE5\x03\x00\x33\x44\x00\x80\xB5\xE5\x03\x00\xB6\xED"
-     "\x03\x00"
-     "\x37\x44\x00\x80\x3B\x44\x00\x80\x3F\x44\x00\x80\xBA\x6D\x03\x00\xBB\x65"
-     "\x03\x00"
-     "\xBC\x7D\x03\x00\xBD\x65\x03\x00\xBE\x65\x03\x00\xBF\x55\x03\x00\x43\x44"
-     "\x00\x80"
-     "\x47\x44\x00\x80\xBF\x00\x04\x00\xA3\x25\x02\x00\x4B\x44\x00\x80\xA5\x25"
-     "\x02\x00"
-     "\xA6\x2D\x02\x00\x4F\x44\x00\x80\x53\x44\x00\x80\x57\x44\x00\x80\xAA\xAD"
-     "\x02\x00"
-     "\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xA5\x02\x00\xAE\xA5\x02\x00\xAF\x95"
-     "\x02\x00"
-     "\x5B\x44\x00\x80\x5F\x44\x00\x80\x63\x44\x00\x80\x67\x44\x00\x80\x6B\x44"
-     "\x00\x80"
-     "\xE3\xCC\x03\x00\x6F\x44\x00\x80\xE1\xAC\x01\x00\x73\x44\x00\x80\xEF\x1C"
-     "\x03\x00"
-     "\x77\x44\x00\x80\x7B\x44\x00\x80\x7F\x44\x00\x80\x83\x44\x00\x80\x87\x44"
-     "\x00\x80"
-     "\x8B\x44\x00\x80\xE1\x70\x7F\x00\x63\x42\x00\x80\xE3\x04\x7E\x00\x8F\x44"
-     "\x00\x80"
-     "\x97\x44\x00\x80\xE1\x94\x01\x00\x9B\x44\x00\x80\xE3\x58\x01\x00\x80\xD9"
-     "\x00\x00"
-     "\x81\xD9\x00\x00\x82\x25\x00\x00\xEF\xE4\x7E\x00\x9F\x44\x00\x80\xA3\x44"
-     "\x00\x80"
-     "\xEF\xE0\x7E\x00\xA7\x44\x00\x80\xB3\x01\x01\x00\xAB\x44\x00\x80\x86\xE8"
-     "\x07\x00"
-     "\x87\x2C\x04\x00\xAF\x44\x00\x80\xB6\x35\x01\x00\xB5\x05\x01\x00\xB3\x44"
-     "\x00\x80"
-     "\xBB\xF1\x00\x00\xBA\xE1\x00\x00\xB7\x44\x00\x80\xBB\x44\x00\x80\xBF\xB1"
-     "\x00\x00"
-     "\xBE\xD1\x00\x00\xBD\xD9\x00\x00\xBC\xE1\x00\x00\x23\x44\x00\x80\x93\x44"
-     "\x00\x80"
-     "\xBF\x44\x00\x80\xC3\x44\x00\x80\xA1\x01\x04\x00\xA0\x11\x04\x00\xA3\x19"
-     "\x00\x00"
-     "\xA2\xC5\x04\x00\xA8\xB5\x06\x00\xA9\xBD\x06\x00\xAA\xE9\x06\x00\xAB\xFD"
-     "\x06\x00"
-     "\xAC\xED\x06\x00\xAD\xD5\x06\x00\xAE\xDD\x06\x00\xAF\x3D\x07\x00\xB0\x45"
-     "\x07\x00"
-     "\xB1\x55\x07\x00\xB2\x55\x07\x00\xB3\x6D\x07\x00\xB4\x75\x07\x00\xB5\x15"
-     "\x07\x00"
-     "\xB6\x1D\x07\x00\xB7\x15\x07\x00\xB8\x2D\x07\x00\xB9\x35\x07\x00\xBA\x31"
-     "\x07\x00"
-     "\xBB\x0D\x07\x00\xBC\x19\x07\x00\xBD\x19\x07\x00\xBE\x09\x07\x00\xBF\x09"
-     "\x07\x00"
-     "\xA3\x41\x06\x00\xC7\x44\x00\x80\xCB\x44\x00\x80\xCF\x44\x00\x80\xD3\x44"
-     "\x00\x80"
-     "\xA6\x75\x06\x00\xA5\x45\x06\x00\xD7\x44\x00\x80\xAB\xB1\x07\x00\xAA\xA1"
-     "\x07\x00"
-     "\x8F\xC9\x6D\x00\xDB\x44\x00\x80\xAF\xF1\x07\x00\xAE\x91\x07\x00\xAD\x99"
-     "\x07\x00"
-     "\xAC\xA1\x07\x00\x95\xDD\x75\x00\x94\xC1\x74\x00\x97\xCD\x70\x00\x96\x2D"
-     "\x73\x00"
-     "\x91\x5D\x68\x00\x90\x55\x68\x00\x93\xD9\x74\x00\x92\x4D\x69\x00\x9D\xE5"
-     "\x78\x00"
-     "\x9C\x1D\x7B\x00\x9F\x6D\x07\x00\x9E\xB9\x78\x00\x99\x1D\x7F\x00\x98\x55"
-     "\x70\x00"
-     "\x9B\xA1\x7C\x00\x9A\xF9\x7C\x00\x82\x61\x6C\x00\x83\x85\x69\x00\xDF\x44"
-     "\x00\x80"
-     "\xE3\x44\x00\x80\x86\x11\x75\x00\x87\x11\x75\x00\x84\x95\x69\x00\x85\x8D"
-     "\x68\x00"
-     "\x8A\x3D\x74\x00\x8B\xC5\x72\x00\xE7\x44\x00\x80\xEB\x44\x00\x80\x8E\xDD"
-     "\x7E\x00"
-     "\x8F\x31\x7D\x00\x8C\x3D\x71\x00\x8D\x9D\x71\x00\x92\x19\x7D\x00\x93\xBD"
-     "\x7A\x00"
-     "\xEF\x44\x00\x80\xEF\x90\x09\x00\x96\xD5\x06\x00\x97\x51\x05\x00\x94\x5D"
-     "\x79\x00"
-     "\x95\x09\x79\x00\x9A\x71\x05\x00\x9B\xBD\x05\x00\xF3\x44\x00\x80\xF7\x44"
-     "\x00\x80"
-     "\xFB\x44\x00\x80\xE1\xA8\x05\x00\x9C\x79\x01\x00\xE3\xB8\x08\x00\xA1\x85"
-     "\x01\x00"
-     "\xFF\x44\x00\x80\xA3\xA9\x0D\x00\xA2\x01\x0C\x00\xA5\x01\x08\x00\xA4\x39"
-     "\x0D\x00"
-     "\xA7\xA9\x09\x00\xA6\xBD\x09\x00\xA9\xA5\x15\x00\xA8\x01\x14\x00\xAB\x01"
-     "\x14\x00"
-     "\xAA\xFD\x15\x00\xAD\xB9\x11\x00\xAC\xB1\x11\x00\xAF\x01\x1C\x00\xAE\xA1"
-     "\x10\x00"
-     "\xB1\xFD\x1C\x00\xB0\xE5\x1D\x00\xB3\xE9\x19\x00\xB2\x01\x18\x00\xB5\x01"
-     "\x24\x00"
-     "\xB4\x7D\x19\x00\x84\x14\x00\x00\xBE\x14\x00\x00\x80\x8D\x00\x00\x81\x95"
-     "\x00\x00"
-     "\x82\x6D\x00\x00\x07\x45\x00\x80\x86\x50\x0F\x00\x87\x64\x00\x00\x0B\x45"
-     "\x00\x80"
-     "\x0F\x45\x00\x80\xBA\xD5\x07\x00\xBB\xDD\x07\x00\xB8\xC1\x07\x00\xB9\xC1"
-     "\x07\x00"
-     "\xBE\x31\x04\x00\xBF\x31\x04\x00\xBC\xF1\x07\x00\xBD\xF1\x07\x00\xB2\xAD"
-     "\x07\x00"
-     "\xB3\xB5\x07\x00\xB0\xAD\x07\x00\xB1\xA5\x07\x00\xB6\x9D\x07\x00\xB7\xF5"
-     "\x07\x00"
-     "\xB4\xA5\x07\x00\xB5\x95\x07\x00\xAA\x69\x07\x00\xAB\x69\x07\x00\xA8\x69"
-     "\x07\x00"
-     "\xA9\x69\x07\x00\xAE\x69\x07\x00\xAF\x69\x07\x00\xAC\x69\x07\x00\xAD\x69"
-     "\x07\x00"
-     "\x13\x45\x00\x80\x17\x45\x00\x80\x1B\x45\x00\x80\x1F\x45\x00\x80\x23\x45"
-     "\x00\x80"
-     "\x27\x45\x00\x80\x2B\x45\x00\x80\x2F\x45\x00\x80\xA8\x11\x05\x00\xA9\x1D"
-     "\x05\x00"
-     "\xAA\x39\x05\x00\xAB\x39\x05\x00\xAC\x2D\x05\x00\xAD\x51\x05\x00\xAE\x49"
-     "\x05\x00"
-     "\xAF\x41\x05\x00\x33\x45\x00\x80\x37\x45\x00\x80\x3B\x45\x00\x80\x3F\x45"
-     "\x00\x80"
-     "\x43\x45\x00\x80\x47\x45\x00\x80\x4B\x45\x00\x80\x4F\x45\x00\x80\xB8\x5D"
-     "\x02\x00"
-     "\xB9\x69\x02\x00\xBA\xC1\x03\x00\xBB\xC1\x03\x00\xBC\xF9\x03\x00\xBD\xF9"
-     "\x03\x00"
-     "\xBE\x91\x03\x00\xBF\xB5\x03\x00\xB0\x09\x05\x00\xB1\x09\x05\x00\xB2\xE1"
-     "\x02\x00"
-     "\xB3\xE1\x02\x00\xB4\x75\x02\x00\xB5\x7D\x02\x00\xB6\x75\x02\x00\xB7\x6D"
-     "\x02\x00"
-     "\xB3\xB1\x04\x00\x84\x00\x02\x00\xBE\x04\x0D\x00\x53\x45\x00\x80\x57\x45"
-     "\x00\x80"
-     "\xB6\xD1\x04\x00\xB5\xA5\x04\x00\x5B\x45\x00\x80\xBB\xCD\x04\x00\xBA\xCD"
-     "\x04\x00"
-     "\x5F\x45\x00\x80\x63\x45\x00\x80\xBF\xB9\x03\x00\xBE\xB1\x03\x00\xBD\x35"
-     "\x03\x00"
-     "\xBC\x35\x03\x00\x67\x45\x00\x80\xA3\xF5\x04\x00\x6B\x45\x00\x80\x6F\x45"
-     "\x00\x80"
-     "\xA6\x95\x04\x00\x77\x45\x00\x80\x7B\x45\x00\x80\xA5\xE1\x04\x00\xAA\x89"
-     "\x04\x00"
-     "\xAB\x89\x04\x00\x87\xA8\x0D\x00\x86\xCC\x0C\x00\xAE\xF5\x03\x00\xAF\xFD"
-     "\x03\x00"
-     "\xAC\x71\x03\x00\xAD\x71\x03\x00\xE1\x50\x06\x00\xE1\x34\x07\x00\xE3\x40"
-     "\x00\x00"
-     "\xE3\x58\x07\x00\x80\xD1\x00\x00\x81\xDD\x00\x00\x82\xD5\x00\x00\x7F\x45"
-     "\x00\x80"
-     "\x83\x45\x00\x80\x87\x45\x00\x80\x8B\x45\x00\x80\x8F\x45\x00\x80\x93\x45"
-     "\x00\x80"
-     "\x97\x45\x00\x80\xEF\x9C\x00\x00\xEF\xC8\x07\x00\x9B\x45\x00\x80\x9F\x45"
-     "\x00\x80"
-     "\xB3\x35\x02\x00\xA3\x45\x00\x80\xB5\xB5\x01\x00\xA7\x45\x00\x80\xAB\x45"
-     "\x00\x80"
-     "\xB6\xB5\x01\x00\xBE\xEC\x0C\x00\xAF\x45\x00\x80\xBB\x91\x01\x00\xBA\x99"
-     "\x01\x00"
-     "\xBD\x51\x01\x00\xBC\x89\x01\x00\xBF\x51\x01\x00\xBE\x59\x01\x00\xA8\xED"
-     "\x0D\x00"
-     "\xA9\xFD\x0D\x00\xAA\xF5\x0D\x00\xAB\x6D\x0E\x00\xAC\x75\x0E\x00\xAD\x7D"
-     "\x0E\x00"
-     "\xAE\x75\x0E\x00\xAF\x6D\x0E\x00\x73\x45\x00\x80\xB3\x45\x00\x80\xB7\x45"
-     "\x00\x80"
-     "\xBB\x45\x00\x80\xBF\x45\x00\x80\xC3\x45\x00\x80\xC7\x45\x00\x80\xCB\x45"
-     "\x00\x80"
-     "\xB8\xF5\x0E\x00\xB9\xFD\x0E\x00\xBA\xF5\x0E\x00\xBB\x41\x0F\x00\xBC\x41"
-     "\x0F\x00"
-     "\xBD\x49\x0F\x00\xBE\x71\x0F\x00\xBF\x71\x0F\x00\xB0\x15\x0E\x00\xB1\x1D"
-     "\x0E\x00"
-     "\xB2\x15\x0E\x00\xB3\xCD\x0E\x00\xB4\xD5\x0E\x00\xB5\xDD\x0E\x00\xB6\xD5"
-     "\x0E\x00"
-     "\xB7\xCD\x0E\x00\xA3\x7D\x0D\x00\xCF\x45\x00\x80\xD3\x45\x00\x80\xD7\x45"
-     "\x00\x80"
-     "\xDB\x45\x00\x80\xA6\xFD\x0E\x00\xA5\xFD\x0E\x00\xDF\x45\x00\x80\xAB\xD9"
-     "\x0E\x00"
-     "\xAA\xD1\x0E\x00\x84\xA8\x02\x00\xE3\x45\x00\x80\xAF\x19\x0E\x00\xAE\x11"
-     "\x0E\x00"
-     "\xAD\x19\x0E\x00\xAC\xC1\x0E\x00\x80\x4D\x00\x00\x81\x55\x00\x00\x82\x55"
-     "\x00\x00"
-     "\xB3\x51\x0F\x00\xE7\x45\x00\x80\xB5\x71\x0F\x00\xB6\x71\x0F\x00\xEB\x45"
-     "\x00\x80"
-     "\x86\xE0\x00\x00\x87\x04\x03\x00\xBA\x5D\x0F\x00\xBB\x2D\x0F\x00\xBC\x35"
-     "\x0F\x00"
-     "\xBD\x39\x0F\x00\xBE\x2D\x0F\x00\xBF\x25\x0F\x00\xA8\x55\x0E\x00\xA9\x5D"
-     "\x0E\x00"
-     "\xAA\x95\x0E\x00\xAB\xAD\x0E\x00\xAC\xB5\x0E\x00\xAD\xBD\x0E\x00\xAE\xB5"
-     "\x0E\x00"
-     "\xAF\xAD\x0E\x00\xEF\x45\x00\x80\xF3\x45\x00\x80\xF7\x45\x00\x80\xFB\x45"
-     "\x00\x80"
-     "\xFF\x45\x00\x80\x03\x46\x00\x80\x07\x46\x00\x80\x0B\x46\x00\x80\xB8\x69"
-     "\x01\x00"
-     "\xB9\x69\x01\x00\xBA\x79\x01\x00\xBB\x79\x01\x00\xBC\x69\x01\x00\xBD\x69"
-     "\x01\x00"
-     "\xBE\xDD\x01\x00\xBF\xD5\x01\x00\xB0\xD5\x0E\x00\xB1\xA5\x0E\x00\xB2\xAD"
-     "\x0E\x00"
-     "\xB3\xA1\x0E\x00\xB4\xA5\x0E\x00\xB5\xAD\x0E\x00\xB6\x9D\x0E\x00\xB7\x59"
-     "\x01\x00"
-     "\xA3\x1D\x0E\x00\x0F\x46\x00\x80\x13\x46\x00\x80\x03\x45\x00\x80\x17\x46"
-     "\x00\x80"
-     "\xA6\x3D\x0E\x00\xA5\x3D\x0E\x00\x1B\x46\x00\x80\xAB\x61\x0E\x00\xAA\x11"
-     "\x0E\x00"
-     "\x1F\x46\x00\x80\x23\x46\x00\x80\xAF\x69\x0E\x00\xAE\x61\x0E\x00\xAD\x75"
-     "\x0E\x00"
-     "\xAC\x79\x0E\x00\x27\x46\x00\x80\x2B\x46\x00\x80\x2F\x46\x00\x80\x33\x46"
-     "\x00\x80"
-     "\x37\x46\x00\x80\x3B\x46\x00\x80\x3F\x46\x00\x80\x43\x46\x00\x80\x80\x0D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x1D\x00\x00\x47\x46\x00\x80\x4B\x46\x00\x80\x4F\x46"
-     "\x00\x80"
-     "\x84\x78\x01\x00\xBE\x14\x00\x00\xE3\x14\x0F\x00\x57\x46\x00\x80\xE1\xE0"
-     "\x0D\x00"
-     "\x84\x00\x03\x00\x86\xB0\x04\x00\x87\x14\x03\x00\x5B\x46\x00\x80\x5F\x46"
-     "\x00\x80"
-     "\x63\x46\x00\x80\x67\x46\x00\x80\x6B\x46\x00\x80\x6F\x46\x00\x80\xEF\xC0"
-     "\x0F\x00"
-     "\x73\x46\x00\x80\x77\x46\x00\x80\x7B\x46\x00\x80\x7F\x46\x00\x80\x83\x46"
-     "\x00\x80"
-     "\x87\x46\x00\x80\xB3\x6D\x03\x00\x8B\x46\x00\x80\xB5\x7D\x03\x00\xB6\x75"
-     "\x03\x00"
-     "\x8F\x46\x00\x80\x93\x46\x00\x80\x97\x46\x00\x80\xBA\x51\x03\x00\xBB\x51"
-     "\x03\x00"
-     "\xBC\xF5\x03\x00\xBD\xFD\x03\x00\xBE\xE9\x03\x00\xBF\xE1\x03\x00\x9B\x46"
-     "\x00\x80"
-     "\x9F\x46\x00\x80\xA3\x46\x00\x80\xA7\x46\x00\x80\xAB\x46\x00\x80\xAF\x46"
-     "\x00\x80"
-     "\xB3\x46\x00\x80\xB7\x46\x00\x80\xA8\x71\x02\x00\xA9\x79\x02\x00\xAA\x89"
-     "\x03\x00"
-     "\xAB\x89\x03\x00\xAC\x99\x03\x00\xAD\x99\x03\x00\xAE\x89\x03\x00\xAF\x89"
-     "\x03\x00"
-     "\xB0\xF9\x03\x00\xB1\xF9\x03\x00\xB2\x4D\x03\x00\xB3\x45\x03\x00\xB4\x41"
-     "\x03\x00"
-     "\xB5\x49\x03\x00\xB6\x71\x03\x00\xB7\x71\x03\x00\xB8\x21\x03\x00\xB9\x21"
-     "\x03\x00"
-     "\xBA\x21\x03\x00\xBB\x21\x03\x00\xBC\x21\x03\x00\xBD\x21\x03\x00\xBE\x21"
-     "\x03\x00"
-     "\xBF\x21\x03\x00\x80\x9D\x01\x00\x81\x11\x00\x00\x82\x11\x00\x00\x84\x04"
-     "\x05\x00"
-     "\xEF\x14\x00\x00\xBB\x46\x00\x80\xBF\x46\x00\x80\xBE\x10\x05\x00\xE3\xC8"
-     "\x00\x00"
-     "\xC7\x46\x00\x80\xE1\xD0\x01\x00\xCB\x46\x00\x80\xCF\x46\x00\x80\xD3\x46"
-     "\x00\x80"
-     "\xD7\x46\x00\x80\xDB\x46\x00\x80\xAA\x79\x02\x00\xAB\x79\x02\x00\x86\xE8"
-     "\x04\x00"
-     "\x87\x60\x05\x00\xAE\xC1\x02\x00\xAF\xC9\x02\x00\xAC\xDD\x02\x00\xAD\xD5"
-     "\x02\x00"
-     "\xDF\x46\x00\x80\xA3\x45\x02\x00\xE3\x46\x00\x80\xE7\x46\x00\x80\xA6\x5D"
-     "\x02\x00"
-     "\xEB\x46\x00\x80\xEF\x46\x00\x80\xA5\x55\x02\x00\xF3\x46\x00\x80\xF7\x46"
-     "\x00\x80"
-     "\xFB\x46\x00\x80\xFF\x46\x00\x80\x03\x47\x00\x80\x07\x47\x00\x80\x0B\x47"
-     "\x00\x80"
-     "\xEF\x84\x0E\x00\xBE\xAC\x04\x00\xE1\xD0\x0E\x00\x0F\x47\x00\x80\xE3\x14"
-     "\x01\x00"
-     "\x13\x47\x00\x80\x17\x47\x00\x80\x1B\x47\x00\x80\x1F\x47\x00\x80\xB3\xDD"
-     "\x01\x00"
-     "\x23\x47\x00\x80\x27\x47\x00\x80\x2B\x47\x00\x80\x2F\x47\x00\x80\xB6\xFD"
-     "\x01\x00"
-     "\xB5\xFD\x01\x00\x37\x47\x00\x80\xBB\xDD\x01\x00\xBA\xDD\x01\x00\x84\xE0"
-     "\x04\x00"
-     "\x3B\x47\x00\x80\xBF\xA1\x01\x00\xBE\xBD\x01\x00\xBD\xBD\x01\x00\xBC\xBD"
-     "\x01\x00"
-     "\xA8\x05\x06\x00\xA9\x1D\x06\x00\xAA\x15\x06\x00\xAB\x2D\x06\x00\xAC\x35"
-     "\x06\x00"
-     "\xAD\x3D\x06\x00\xAE\x35\x06\x00\xAF\x29\x06\x00\xC3\x46\x00\x80\x82\xF5"
-     "\x07\x00"
-     "\x81\xE5\x07\x00\x80\xE5\x07\x00\x33\x47\x00\x80\x3F\x47\x00\x80\x86\x1C"
-     "\x00\x00"
-     "\x87\xB0\x03\x00\xB8\x25\x06\x00\xB9\xC5\x06\x00\xBA\xCD\x06\x00\xBB\xC5"
-     "\x06\x00"
-     "\xBC\xDD\x06\x00\xBD\xC5\x06\x00\xBE\xCD\x06\x00\xBF\xC5\x06\x00\xB0\x59"
-     "\x06\x00"
-     "\xB1\x59\x06\x00\xB2\x29\x06\x00\xB3\x29\x06\x00\xB4\x39\x06\x00\xB5\x25"
-     "\x06\x00"
-     "\xB6\x25\x06\x00\xB7\x1D\x06\x00\xA3\x9D\x06\x00\x43\x47\x00\x80\x47\x47"
-     "\x00\x80"
-     "\x4B\x47\x00\x80\x4F\x47\x00\x80\xA6\xBD\x06\x00\xA5\xBD\x06\x00\x53\x47"
-     "\x00\x80"
-     "\xAB\x9D\x06\x00\xAA\x9D\x06\x00\x57\x47\x00\x80\x5B\x47\x00\x80\xAF\xE1"
-     "\x06\x00"
-     "\xAE\xFD\x06\x00\xAD\xFD\x06\x00\xAC\xFD\x06\x00\x5F\x47\x00\x80\xB3\xF5"
-     "\x07\x00"
-     "\x63\x47\x00\x80\x67\x47\x00\x80\xB6\xED\x07\x00\x6B\x47\x00\x80\x6F\x47"
-     "\x00\x80"
-     "\xB5\xE5\x07\x00\xBA\x8D\x07\x00\xBB\x91\x07\x00\x73\x47\x00\x80\x77\x47"
-     "\x00\x80"
-     "\xBE\x75\x07\x00\xBF\x7D\x07\x00\xBC\x81\x07\x00\xBD\x7D\x07\x00\xA8\x25"
-     "\x06\x00"
-     "\xA9\x29\x06\x00\xAA\x39\x06\x00\xAB\x39\x06\x00\xAC\x29\x06\x00\xAD\x29"
-     "\x06\x00"
-     "\xAE\x79\x06\x00\xAF\x75\x06\x00\x7B\x47\x00\x80\x7F\x47\x00\x80\x83\x47"
-     "\x00\x80"
-     "\x87\x47\x00\x80\x8B\x47\x00\x80\x8F\x47\x00\x80\x93\x47\x00\x80\x97\x47"
-     "\x00\x80"
-     "\xB8\xD5\x06\x00\xB9\xDD\x06\x00\xBA\xE1\x06\x00\xBB\xFD\x06\x00\xBC\xE5"
-     "\x06\x00"
-     "\xBD\xED\x06\x00\xBE\xE5\x06\x00\xBF\x99\x06\x00\xB0\x0D\x06\x00\xB1\x11"
-     "\x06\x00"
-     "\xB2\x11\x06\x00\xB3\xED\x06\x00\xB4\xF5\x06\x00\xB5\xFD\x06\x00\xB6\xF5"
-     "\x06\x00"
-     "\xB7\xED\x06\x00\xA3\xB1\x06\x00\x82\x2D\x00\x00\x81\x15\x00\x00\x80\xB1"
-     "\x00\x00"
-     "\x53\x46\x00\x80\xA6\xA9\x06\x00\xA5\xA1\x06\x00\x9B\x47\x00\x80\xAB\xD5"
-     "\x06\x00"
-     "\xAA\xC9\x06\x00\x9F\x47\x00\x80\xBE\x68\x01\x00\xAF\x39\x06\x00\xAE\x31"
-     "\x06\x00"
-     "\xAD\x39\x06\x00\xAC\xC5\x06\x00\xA7\x47\x00\x80\xB3\xF1\x01\x00\x86\x68"
-     "\x00\x00"
-     "\x87\x7C\x01\x00\xB6\x5D\x01\x00\xAB\x47\x00\x80\xAF\x47\x00\x80\xB5\x55"
-     "\x01\x00"
-     "\xBA\x49\x01\x00\xBB\x49\x01\x00\xB3\x47\x00\x80\xB7\x47\x00\x80\xBE\x3D"
-     "\x01\x00"
-     "\xBF\x21\x01\x00\xBC\x39\x01\x00\xBD\x35\x01\x00\xBB\x47\x00\x80\xBF\x47"
-     "\x00\x80"
-     "\x84\x04\x03\x00\xBE\x80\x1C\x00\xC3\x47\x00\x80\xE1\x1C\x06\x00\xC7\x47"
-     "\x00\x80"
-     "\xE3\x00\x06\x00\xBF\x08\x2E\x00\xCB\x47\x00\x80\xCF\x47\x00\x80\xEF\xC8"
-     "\x07\x00"
-     "\xD3\x47\x00\x80\xD7\x47\x00\x80\xDB\x47\x00\x80\xDF\x47\x00\x80\xE3\x47"
-     "\x00\x80"
-     "\xE7\x47\x00\x80\xA3\x7D\x02\x00\xEB\x47\x00\x80\xA5\xD9\x02\x00\xEF\x47"
-     "\x00\x80"
-     "\xF3\x47\x00\x80\xA6\xD1\x02\x00\xF7\x47\x00\x80\xFB\x47\x00\x80\xAB\xC5"
-     "\x02\x00"
-     "\xAA\xC5\x02\x00\xAD\xB9\x02\x00\xAC\xB5\x02\x00\xAF\xAD\x02\x00\xAE\xB1"
-     "\x02\x00"
-     "\xA9\x6D\x05\x00\xA8\x59\x05\x00\xAB\x0D\x02\x00\xAA\xB9\x02\x00\xAD\x1D"
-     "\x02\x00"
-     "\xAC\x1D\x02\x00\xAF\x05\x02\x00\xAE\x0D\x02\x00\xBE\x68\x1D\x00\xFF\x47"
-     "\x00\x80"
-     "\x03\x48\x00\x80\x07\x48\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x99"
-     "\x01\x00"
-     "\x0B\x48\x00\x80\xB9\xE1\x03\x00\xB8\x29\x02\x00\xBB\xE1\x03\x00\xBA\xE9"
-     "\x03\x00"
-     "\xBD\xE1\x03\x00\xBC\xF9\x03\x00\xBF\xE1\x03\x00\xBE\xE9\x03\x00\xB1\x4D"
-     "\x02\x00"
-     "\xB0\x4D\x02\x00\xB3\x21\x02\x00\xB2\x2D\x02\x00\xB5\x25\x02\x00\xB4\x39"
-     "\x02\x00"
-     "\xB7\x11\x02\x00\xB6\x25\x02\x00\xA8\xC1\x02\x00\xA9\xD1\x02\x00\xAA\xD1"
-     "\x02\x00"
-     "\xAB\xE5\x02\x00\xAC\xFD\x02\x00\xAD\x15\x01\x00\xAE\x1D\x01\x00\xAF\x0D"
-     "\x01\x00"
-     "\x0F\x48\x00\x80\x17\x48\x00\x80\x1B\x48\x00\x80\x1F\x48\x00\x80\x23\x48"
-     "\x00\x80"
-     "\x27\x48\x00\x80\x2B\x48\x00\x80\x2F\x48\x00\x80\xB8\x05\x01\x00\xB9\x09"
-     "\x01\x00"
-     "\xBA\x1D\x01\x00\xBB\x15\x01\x00\xBC\x31\x01\x00\xBD\x31\x01\x00\xBE\xFD"
-     "\x01\x00"
-     "\xBF\xF5\x01\x00\xB0\x79\x01\x00\xB1\x41\x01\x00\xB2\x41\x01\x00\xB3\x5D"
-     "\x01\x00"
-     "\xB4\x45\x01\x00\xB5\x4D\x01\x00\xB6\x45\x01\x00\xB7\x3D\x01\x00\x86\xA0"
-     "\x1D\x00"
-     "\x87\xC4\x1D\x00\x33\x48\x00\x80\xEF\xD8\x00\x00\x37\x48\x00\x80\x3B\x48"
-     "\x00\x80"
-     "\x3F\x48\x00\x80\xEF\xC4\x06\x00\x84\x6C\x1C\x00\xE1\xF4\x06\x00\x43\x48"
-     "\x00\x80"
-     "\xE3\xB0\x06\x00\x47\x48\x00\x80\xE1\x94\x01\x00\x4B\x48\x00\x80\xE3\x44"
-     "\x06\x00"
-     "\xB3\x19\x02\x00\x4F\x48\x00\x80\x53\x48\x00\x80\x57\x48\x00\x80\x85\xEC"
-     "\x2C\x00"
-     "\xB6\xD5\x01\x00\xB5\x35\x02\x00\x5B\x48\x00\x80\xBB\xC5\x01\x00\xBA\xFD"
-     "\x01\x00"
-     "\x5F\x48\x00\x80\x63\x48\x00\x80\xBF\xC9\x01\x00\xBE\xC1\x01\x00\xBD\xC9"
-     "\x01\x00"
-     "\xBC\xD5\x01\x00\xA3\xD9\x1D\x00\x13\x48\x00\x80\x67\x48\x00\x80\x6B\x48"
-     "\x00\x80"
-     "\x6F\x48\x00\x80\xA6\x15\x1E\x00\xA5\xF5\x1D\x00\x73\x48\x00\x80\xAB\x05"
-     "\x1E\x00"
-     "\xAA\x3D\x1E\x00\x77\x48\x00\x80\x7B\x48\x00\x80\xAF\x09\x1E\x00\xAE\x01"
-     "\x1E\x00"
-     "\xAD\x09\x1E\x00\xAC\x15\x1E\x00\x80\x69\x00\x00\x81\x69\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x7F\x48\x00\x80\x83\x48\x00\x80\x87\x48\x00\x80\x87\x10\x03\x00\x86\x7C"
-     "\x03\x00"
-     "\x8B\x48\x00\x80\x8F\x48\x00\x80\x93\x48\x00\x80\x97\x48\x00\x80\x9B\x48"
-     "\x00\x80"
-     "\x9F\x48\x00\x80\xA3\x48\x00\x80\xA7\x48\x00\x80\xA8\xA5\x1F\x00\xA9\xAD"
-     "\x1F\x00"
-     "\xAA\xA5\x1F\x00\xAB\xBD\x1F\x00\xAC\xA5\x1F\x00\xAD\xAD\x1F\x00\xAE\xA5"
-     "\x1F\x00"
-     "\xAF\x15\x1F\x00\xAB\x48\x00\x80\xAF\x48\x00\x80\xB3\x48\x00\x80\xB7\x48"
-     "\x00\x80"
-     "\xBB\x48\x00\x80\xBF\x48\x00\x80\xC3\x48\x00\x80\xC7\x48\x00\x80\xB8\x0D"
-     "\x1F\x00"
-     "\xB9\x19\x1F\x00\xBA\x21\x1F\x00\xBB\x21\x1F\x00\xBC\xD9\x00\x00\xBD\xD9"
-     "\x00\x00"
-     "\xBE\xC9\x00\x00\xBF\xC1\x00\x00\xB0\x71\x1F\x00\xB1\x71\x1F\x00\xB2\x71"
-     "\x1F\x00"
-     "\xB3\x45\x1F\x00\xB4\x41\x1F\x00\xB5\x4D\x1F\x00\xB6\x3D\x1F\x00\xB7\x35"
-     "\x1F\x00"
-     "\xB3\x2D\x1E\x00\xCB\x48\x00\x80\xCF\x48\x00\x80\xD3\x48\x00\x80\xD7\x48"
-     "\x00\x80"
-     "\xB6\x2D\x1E\x00\xB5\x2D\x1E\x00\xDB\x48\x00\x80\xBB\xB5\x1E\x00\xBA\xB5"
-     "\x1E\x00"
-     "\xDF\x48\x00\x80\xE3\x48\x00\x80\xBF\x89\x1E\x00\xBE\x85\x1E\x00\xBD\x91"
-     "\x1E\x00"
-     "\xBC\xA5\x1E\x00\x82\x29\x00\x00\xA3\x69\x1E\x00\x80\x1D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\xA6\x69\x1E\x00\xE7\x48\x00\x80\xEB\x48\x00\x80\xA5\x69\x1E\x00\xAA\xF1"
-     "\x1E\x00"
-     "\xAB\xF1\x1E\x00\xEF\x48\x00\x80\x84\xE0\x01\x00\xAE\xC1\x1E\x00\xAF\xCD"
-     "\x1E\x00"
-     "\xAC\xE1\x1E\x00\xAD\xD5\x1E\x00\xA8\xD5\x01\x00\xA9\xE5\x01\x00\xAA\xED"
-     "\x01\x00"
-     "\xAB\xE5\x01\x00\xAC\xFD\x01\x00\xAD\xE5\x01\x00\xAE\xED\x01\x00\xAF\xE5"
-     "\x01\x00"
-     "\xBE\xA0\x01\x00\xA3\x47\x00\x80\xF3\x48\x00\x80\xF7\x48\x00\x80\x86\x10"
-     "\x00\x00"
-     "\x87\x74\x01\x00\xFB\x48\x00\x80\xFF\x48\x00\x80\xB8\x7D\x01\x00\xB9\xC1"
-     "\x00\x00"
-     "\xBA\xC1\x00\x00\xBB\xC1\x00\x00\xBC\xC1\x00\x00\xBD\xC9\x00\x00\xBE\xF1"
-     "\x00\x00"
-     "\xBF\xF1\x00\x00\xB0\x9D\x01\x00\xB1\x45\x01\x00\xB2\x4D\x01\x00\xB3\x45"
-     "\x01\x00"
-     "\xB4\x5D\x01\x00\xB5\x45\x01\x00\xB6\x4D\x01\x00\xB7\x45\x01\x00\x03\x49"
-     "\x00\x80"
-     "\x07\x49\x00\x80\x0B\x49\x00\x80\x0F\x49\x00\x80\x13\x49\x00\x80\xEF\x34"
-     "\x02\x00"
-     "\xEF\xEC\x1E\x00\x17\x49\x00\x80\xE1\xF0\x1D\x00\xE3\xE0\x02\x00\xE3\x30"
-     "\x1E\x00"
-     "\xE1\x84\x01\x00\x1B\x49\x00\x80\x1F\x49\x00\x80\x23\x49\x00\x80\x27\x49"
-     "\x00\x80"
-     "\xB3\x25\x02\x00\x94\x00\x00\x00\x2B\x49\x00\x80\x2F\x49\x00\x80\x33\x49"
-     "\x00\x80"
-     "\xB6\x25\x02\x00\xB5\x35\x02\x00\x37\x49\x00\x80\xBB\xC1\x02\x00\xBA\x19"
-     "\x02\x00"
-     "\x3B\x49\x00\x80\x3F\x49\x00\x80\xBF\xC1\x02\x00\xBE\xD9\x02\x00\xBD\xD1"
-     "\x02\x00"
-     "\xBC\xD9\x02\x00\x43\x49\x00\x80\x47\x49\x00\x80\x4B\x49\x00\x80\xA3\xE9"
-     "\x02\x00"
-     "\x4F\x49\x00\x80\xA5\xF9\x02\x00\xA6\xE9\x02\x00\x53\x49\x00\x80\x57\x49"
-     "\x00\x80"
-     "\x5B\x49\x00\x80\xAA\xD5\x02\x00\xAB\x0D\x02\x00\xAC\x15\x02\x00\xAD\x1D"
-     "\x02\x00"
-     "\xAE\x15\x02\x00\xAF\x0D\x02\x00\x80\x61\x00\x00\x81\x61\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x5F\x49\x00\x80\x67\x49\x00\x80\x84\x00\x04\x00\xBE\x14\x04\x00\x6B\x49"
-     "\x00\x80"
-     "\x86\xC0\x04\x00\x87\x50\x03\x00\x6F\x49\x00\x80\x73\x49\x00\x80\x77\x49"
-     "\x00\x80"
-     "\x7B\x49\x00\x80\x7F\x49\x00\x80\x83\x49\x00\x80\xA8\xAD\x02\x00\xA9\xBD"
-     "\x02\x00"
-     "\xAA\xB5\x02\x00\xAB\x0D\x01\x00\xAC\x15\x01\x00\xAD\x1D\x01\x00\xAE\x15"
-     "\x01\x00"
-     "\xAF\x0D\x01\x00\x84\xEC\x04\x00\x87\x49\x00\x80\x8B\x49\x00\x80\x8F\x49"
-     "\x00\x80"
-     "\x93\x49\x00\x80\x97\x49\x00\x80\x9B\x49\x00\x80\x9F\x49\x00\x80\xB8\x1D"
-     "\x01\x00"
-     "\xB9\x2D\x01\x00\xBA\x25\x01\x00\xBB\xCD\x01\x00\xBC\xD5\x01\x00\xBD\xDD"
-     "\x01\x00"
-     "\xBE\xC9\x01\x00\xBF\xC1\x01\x00\xB0\x7D\x01\x00\xB1\x55\x01\x00\xB2\x5D"
-     "\x01\x00"
-     "\xB3\x55\x01\x00\xB4\x4D\x01\x00\xB5\x3D\x01\x00\xB6\x35\x01\x00\xB7\x2D"
-     "\x01\x00"
-     "\xE1\x18\x1E\x00\xA3\x49\x00\x80\xE3\x38\x1E\x00\xA7\x49\x00\x80\xAB\x49"
-     "\x00\x80"
-     "\xAF\x49\x00\x80\xB3\x49\x00\x80\xB7\x49\x00\x80\xBB\x49\x00\x80\xBF\x49"
-     "\x00\x80"
-     "\xBE\x60\x04\x00\xC3\x49\x00\x80\x81\x75\x00\x00\x80\x75\x00\x00\xEF\xE0"
-     "\x1F\x00"
-     "\x82\x6D\x00\x00\xC7\x49\x00\x80\xCB\x49\x00\x80\x86\xE8\x04\x00\x87\x7C"
-     "\x05\x00"
-     "\xCF\x49\x00\x80\xE1\x90\x01\x00\xD7\x49\x00\x80\xE3\xA0\x00\x00\xDB\x49"
-     "\x00\x80"
-     "\xDF\x49\x00\x80\xE3\x49\x00\x80\xEF\xB4\x00\x00\xE7\x49\x00\x80\xEB\x49"
-     "\x00\x80"
-     "\xEF\x49\x00\x80\xF3\x49\x00\x80\xB5\x05\x06\x00\x63\x49\x00\x80\xD3\x49"
-     "\x00\x80"
-     "\xB6\x05\x06\x00\xF7\x49\x00\x80\xFB\x49\x00\x80\xB3\xA5\x05\x00\xFF\x49"
-     "\x00\x80"
-     "\xBD\x19\x06\x00\xBC\x11\x06\x00\xBF\x61\x06\x00\xBE\x11\x06\x00\x03\x4A"
-     "\x00\x80"
-     "\x07\x4A\x00\x80\xBB\x09\x06\x00\xBA\x21\x06\x00\xA3\xFD\x05\x00\x0B\x4A"
-     "\x00\x80"
-     "\x0F\x4A\x00\x80\x13\x4A\x00\x80\x17\x4A\x00\x80\xA6\x5D\x06\x00\xA5\x5D"
-     "\x06\x00"
-     "\x1B\x4A\x00\x80\xAB\x51\x06\x00\xAA\x79\x06\x00\x1F\x4A\x00\x80\x23\x4A"
-     "\x00\x80"
-     "\xAF\x39\x06\x00\xAE\x49\x06\x00\xAD\x41\x06\x00\xAC\x49\x06\x00\xA8\x51"
-     "\x06\x00"
-     "\xA9\x59\x06\x00\xAA\x61\x06\x00\xAB\x61\x06\x00\xAC\x61\x06\x00\xAD\x61"
-     "\x06\x00"
-     "\xAE\x61\x06\x00\xAF\x61\x06\x00\x27\x4A\x00\x80\x2B\x4A\x00\x80\x2F\x4A"
-     "\x00\x80"
-     "\x33\x4A\x00\x80\x80\x0D\x00\x00\x81\xB1\x01\x00\x82\xB1\x01\x00\x37\x4A"
-     "\x00\x80"
-     "\xB8\x4D\x07\x00\xB9\x55\x07\x00\xBA\x5D\x07\x00\xBB\x55\x07\x00\xBC\x4D"
-     "\x07\x00"
-     "\xBD\x75\x07\x00\xBE\x7D\x07\x00\xBF\x71\x07\x00\xB0\xC5\x07\x00\xB1\xCD"
-     "\x07\x00"
-     "\xB2\xC5\x07\x00\xB3\xDD\x07\x00\xB4\xC5\x07\x00\xB5\xCD\x07\x00\xB6\xC5"
-     "\x07\x00"
-     "\xB7\x79\x07\x00\xB3\xE9\x07\x00\x3B\x4A\x00\x80\x3F\x4A\x00\x80\x84\xC0"
-     "\x01\x00"
-     "\xBE\xD8\x01\x00\xB6\xE1\x07\x00\xB5\xE9\x07\x00\x43\x4A\x00\x80\xBB\x09"
-     "\x06\x00"
-     "\xBA\x01\x06\x00\x86\x88\x00\x00\x87\xAC\x01\x00\xBF\x09\x06\x00\xBE\x01"
-     "\x06\x00"
-     "\xBD\x09\x06\x00\xBC\x11\x06\x00\x47\x4A\x00\x80\xA3\xAD\x07\x00\x4B\x4A"
-     "\x00\x80"
-     "\x4F\x4A\x00\x80\xA6\xA5\x07\x00\x53\x4A\x00\x80\x57\x4A\x00\x80\xA5\xAD"
-     "\x07\x00"
-     "\xAA\x45\x06\x00\xAB\x4D\x06\x00\x5B\x4A\x00\x80\x5F\x4A\x00\x80\xAE\x45"
-     "\x06\x00"
-     "\xAF\x4D\x06\x00\xAC\x55\x06\x00\xAD\x4D\x06\x00\xA8\x59\x06\x00\xA9\x65"
-     "\x06\x00"
-     "\xAA\x6D\x06\x00\xAB\x65\x06\x00\xAC\x61\x06\x00\xAD\x61\x06\x00\xAE\x61"
-     "\x06\x00"
-     "\xAF\x61\x06\x00\x84\xAC\x01\x00\x63\x4A\x00\x80\x67\x4A\x00\x80\x6B\x4A"
-     "\x00\x80"
-     "\x6F\x4A\x00\x80\x73\x4A\x00\x80\x77\x4A\x00\x80\x7B\x4A\x00\x80\xB8\x91"
-     "\x01\x00"
-     "\xB9\x99\x01\x00\xBA\xA1\x01\x00\xBB\xA1\x01\x00\xBC\x71\x01\x00\xBD\x71"
-     "\x01\x00"
-     "\xBE\x71\x01\x00\xBF\x71\x01\x00\xB0\xF1\x01\x00\xB1\xF1\x01\x00\xB2\xF5"
-     "\x01\x00"
-     "\xB3\xDD\x01\x00\xB4\xC5\x01\x00\xB5\xB1\x01\x00\xB6\xB1\x01\x00\xB7\xB1"
-     "\x01\x00"
-     "\xB3\xE5\x05\x00\x7F\x4A\x00\x80\x83\x4A\x00\x80\x87\x4A\x00\x80\x8B\x4A"
-     "\x00\x80"
-     "\xB6\xE1\x05\x00\xB5\xE9\x05\x00\x8F\x4A\x00\x80\xBB\x35\x02\x00\xBA\x35"
-     "\x02\x00"
-     "\x93\x4A\x00\x80\x97\x4A\x00\x80\xBF\x75\x02\x00\xBE\x01\x02\x00\xBD\x09"
-     "\x02\x00"
-     "\xBC\x21\x02\x00\x9B\x4A\x00\x80\xA3\xA1\x05\x00\x9F\x4A\x00\x80\xA3\x4A"
-     "\x00\x80"
-     "\xA6\xA5\x05\x00\xA7\x4A\x00\x80\xAB\x4A\x00\x80\xA5\xAD\x05\x00\xAA\x71"
-     "\x02\x00"
-     "\xAB\x71\x02\x00\xAF\x4A\x00\x80\xBE\x28\x03\x00\xAE\x45\x02\x00\xAF\x31"
-     "\x02\x00"
-     "\xAC\x65\x02\x00\xAD\x4D\x02\x00\x80\xD5\x00\x00\x81\xDD\x00\x00\x82\xE1"
-     "\x00\x00"
-     "\xB7\x4A\x00\x80\xE3\x20\x01\x00\xBB\x4A\x00\x80\xE1\xA8\x01\x00\xBF\x4A"
-     "\x00\x80"
-     "\xEF\x34\x02\x00\xC3\x4A\x00\x80\x86\x08\x0C\x00\x87\x68\x03\x00\xAC\x00"
-     "\x00\x00"
-     "\xC7\x4A\x00\x80\xCB\x4A\x00\x80\xCF\x4A\x00\x80\xB3\x8D\x03\x00\xD3\x4A"
-     "\x00\x80"
-     "\xD7\x4A\x00\x80\x84\x80\x0C\x00\xDB\x4A\x00\x80\xB6\xBD\x03\x00\xB5\x81"
-     "\x03\x00"
-     "\xDF\x4A\x00\x80\xBB\x4D\x03\x00\xBA\x4D\x03\x00\xE3\x4A\x00\x80\xE7\x4A"
-     "\x00\x80"
-     "\xBF\x4D\x03\x00\xBE\x4D\x03\x00\xBD\x4D\x03\x00\xBC\x4D\x03\x00\xA8\x41"
-     "\x02\x00"
-     "\xA9\x4D\x02\x00\xAA\x45\x02\x00\xAB\x59\x02\x00\xAC\x49\x02\x00\xAD\x7D"
-     "\x02\x00"
-     "\xAE\x75\x02\x00\xAF\xB9\x02\x00\xBE\x68\x0D\x00\xEB\x4A\x00\x80\xEF\x4A"
-     "\x00\x80"
-     "\xF3\x4A\x00\x80\x84\x6C\x0C\x00\xF7\x4A\x00\x80\xFB\x4A\x00\x80\xFF\x4A"
-     "\x00\x80"
-     "\xB8\x4D\x01\x00\xB9\x55\x01\x00\xBA\x55\x01\x00\xBB\x65\x01\x00\xBC\x7D"
-     "\x01\x00"
-     "\xBD\x15\x01\x00\xBE\x11\x01\x00\xBF\x11\x01\x00\xB0\xC9\x02\x00\xB1\xC9"
-     "\x02\x00"
-     "\xB2\xD9\x02\x00\xB3\xD9\x02\x00\xB4\xC9\x02\x00\xB5\xC9\x02\x00\xB6\x7D"
-     "\x01\x00"
-     "\xB7\x75\x01\x00\xE1\x78\x07\x00\xE3\x98\x00\x00\xE3\xB8\x06\x00\xE1\x5C"
-     "\x06\x00"
-     "\x03\x4B\x00\x80\x07\x4B\x00\x80\x0B\x4B\x00\x80\x0F\x4B\x00\x80\x13\x4B"
-     "\x00\x80"
-     "\x17\x4B\x00\x80\x1B\x4B\x00\x80\x1F\x4B\x00\x80\xEC\x00\x00\x00\xEF\x5C"
-     "\x00\x00"
-     "\xEF\xE8\x06\x00\x27\x4B\x00\x80\x81\x69\x00\x00\x80\x61\x00\x00\xA3\x85"
-     "\x02\x00"
-     "\x82\x61\x00\x00\xA5\x89\x02\x00\x2B\x4B\x00\x80\x2F\x4B\x00\x80\xA6\xB5"
-     "\x02\x00"
-     "\x86\x40\x0C\x00\x87\xC4\x0C\x00\xAB\x45\x02\x00\xAA\x45\x02\x00\xAD\x45"
-     "\x02\x00"
-     "\xAC\x45\x02\x00\xAF\x45\x02\x00\xAE\x45\x02\x00\xA8\x8D\x0E\x00\xA9\x91"
-     "\x0E\x00"
-     "\xAA\x95\x0E\x00\xAB\xA9\x0E\x00\xAC\xA5\x0E\x00\xAD\xAD\x0E\x00\xAE\xA5"
-     "\x0E\x00"
-     "\xAF\xD9\x0E\x00\x23\x4B\x00\x80\x33\x4B\x00\x80\x37\x4B\x00\x80\x3B\x4B"
-     "\x00\x80"
-     "\x3F\x4B\x00\x80\x43\x4B\x00\x80\x47\x4B\x00\x80\x4B\x4B\x00\x80\xB8\x75"
-     "\x0F\x00"
-     "\xB9\x7D\x0F\x00\xBA\x75\x0F\x00\xBB\xDD\x0F\x00\xBC\xC5\x0F\x00\xBD\xCD"
-     "\x0F\x00"
-     "\xBE\xC5\x0F\x00\xBF\xFD\x0F\x00\xB0\xA9\x0E\x00\xB1\xB5\x0E\x00\xB2\xB5"
-     "\x0E\x00"
-     "\xB3\x85\x0E\x00\xB4\x9D\x0E\x00\xB5\x51\x0F\x00\xB6\x51\x0F\x00\xB7\x51"
-     "\x0F\x00"
-     "\xB3\x1D\x0E\x00\x4F\x4B\x00\x80\x53\x4B\x00\x80\x57\x4B\x00\x80\x5B\x4B"
-     "\x00\x80"
-     "\xB6\x2D\x0E\x00\xB5\x2D\x0E\x00\x5F\x4B\x00\x80\xBB\x71\x0E\x00\xBA\x6D"
-     "\x0E\x00"
-     "\x63\x4B\x00\x80\x67\x4B\x00\x80\xBF\x95\x0F\x00\xBE\x59\x0E\x00\xBD\x51"
-     "\x0E\x00"
-     "\xBC\x61\x0E\x00\x6B\x4B\x00\x80\xA3\x59\x0E\x00\x6F\x4B\x00\x80\x73\x4B"
-     "\x00\x80"
-     "\xA6\x69\x0E\x00\x77\x4B\x00\x80\x7B\x4B\x00\x80\xA5\x69\x0E\x00\xAA\x29"
-     "\x0E\x00"
-     "\xAB\x35\x0E\x00\x7F\x4B\x00\x80\x83\x4B\x00\x80\xAE\x1D\x0E\x00\xAF\xD1"
-     "\x0F\x00"
-     "\xAC\x25\x0E\x00\xAD\x15\x0E\x00\xA8\xBD\x0E\x00\xA9\xD1\x0E\x00\xAA\xD1"
-     "\x0E\x00"
-     "\xAB\x29\x01\x00\xAC\x39\x01\x00\xAD\x39\x01\x00\xAE\x29\x01\x00\xAF\x29"
-     "\x01\x00"
-     "\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\x87\x4B\x00\x80\x8B\x4B"
-     "\x00\x80"
-     "\x8F\x4B\x00\x80\xBE\x74\x02\x00\x93\x4B\x00\x80\xB8\xED\x01\x00\xB9\x85"
-     "\x01\x00"
-     "\xBA\x81\x01\x00\xBB\x81\x01\x00\xBC\x85\x01\x00\xBD\x8D\x01\x00\xBE\xB1"
-     "\x01\x00"
-     "\xBF\xB1\x01\x00\xB0\x59\x01\x00\xB1\x59\x01\x00\xB2\xED\x01\x00\xB3\xE5"
-     "\x01\x00"
-     "\xB4\xFD\x01\x00\xB5\xE5\x01\x00\xB6\xE5\x01\x00\xB7\xD5\x01\x00\x97\x4B"
-     "\x00\x80"
-     "\xB6\xA9\x01\x00\xB5\xA1\x01\x00\x9B\x4B\x00\x80\xB3\x49\x0E\x00\x9F\x4B"
-     "\x00\x80"
-     "\x86\x38\x00\x00\x87\xDC\x01\x00\xBF\x31\x01\x00\xBE\x29\x01\x00\xBD\x21"
-     "\x01\x00"
-     "\xBC\x29\x01\x00\xBB\x8D\x01\x00\xBA\x8D\x01\x00\xB3\x4A\x00\x80\xA3\x4B"
-     "\x00\x80"
-     "\xA3\x19\x0E\x00\xA7\x4B\x00\x80\xAB\x4B\x00\x80\xAF\x4B\x00\x80\xB3\x4B"
-     "\x00\x80"
-     "\xA6\xF9\x01\x00\xA5\xF1\x01\x00\xB7\x4B\x00\x80\xAB\xDD\x01\x00\xAA\xDD"
-     "\x01\x00"
-     "\xBB\x4B\x00\x80\xBF\x4B\x00\x80\xAF\x61\x01\x00\xAE\x79\x01\x00\xAD\x71"
-     "\x01\x00"
-     "\xAC\x79\x01\x00\xC3\x4B\x00\x80\xEF\xDC\x0F\x00\xC7\x4B\x00\x80\xCB\x4B"
-     "\x00\x80"
-     "\xCF\x4B\x00\x80\xD3\x4B\x00\x80\xD7\x4B\x00\x80\xDB\x4B\x00\x80\xDF\x4B"
-     "\x00\x80"
-     "\xE3\x4B\x00\x80\xE7\x4B\x00\x80\xEB\x4B\x00\x80\xEF\x4B\x00\x80\xE3\xE8"
-     "\x0E\x00"
-     "\xF3\x4B\x00\x80\xE1\x8C\x0E\x00\x80\x11\x00\x00\x81\x11\x00\x00\x82\x11"
-     "\x00\x00"
-     "\x84\x40\x02\x00\xF7\x4B\x00\x80\xFB\x4B\x00\x80\xFF\x4B\x00\x80\xBE\x10"
-     "\x03\x00"
-     "\x86\xC0\x04\x00\x87\x44\x03\x00\x07\x4C\x00\x80\x0B\x4C\x00\x80\x0F\x4C"
-     "\x00\x80"
-     "\x13\x4C\x00\x80\x17\x4C\x00\x80\x1B\x4C\x00\x80\xEF\x24\x02\x00\x1F\x4C"
-     "\x00\x80"
-     "\x23\x4C\x00\x80\x27\x4C\x00\x80\x2B\x4C\x00\x80\x2F\x4C\x00\x80\x33\x4C"
-     "\x00\x80"
-     "\x37\x4C\x00\x80\x84\xEC\x04\x00\x3B\x4C\x00\x80\x3F\x4C\x00\x80\x43\x4C"
-     "\x00\x80"
-     "\xE3\xEC\x02\x00\x47\x4C\x00\x80\xE1\x38\x01\x00\x4B\x4C\x00\x80\xB3\x55"
-     "\x03\x00"
-     "\x4F\x4C\x00\x80\x53\x4C\x00\x80\x57\x4C\x00\x80\x5B\x4C\x00\x80\xB6\x1D"
-     "\x03\x00"
-     "\xB5\x1D\x03\x00\x5F\x4C\x00\x80\xBB\x09\x03\x00\xBA\x39\x03\x00\x63\x4C"
-     "\x00\x80"
-     "\x67\x4C\x00\x80\xBF\xFD\x00\x00\xBE\xFD\x00\x00\xBD\xF9\x00\x00\xBC\x11"
-     "\x03\x00"
-     "\xA8\x81\x02\x00\xA9\x89\x02\x00\xAA\x9D\x02\x00\xAB\xB1\x02\x00\xAC\xD5"
-     "\x02\x00"
-     "\xAD\xDD\x02\x00\xAE\xD5\x02\x00\xAF\xCD\x02\x00\x80\xCD\x01\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x6B\x4C\x00\x80\x6F\x4C\x00\x80\x77\x4C\x00\x80\xBE\x70"
-     "\x05\x00"
-     "\x7B\x4C\x00\x80\xB8\x59\x01\x00\xB9\x59\x01\x00\xBA\x69\x01\x00\xBB\x69"
-     "\x01\x00"
-     "\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x69\x01\x00\xBF\x65\x01\x00\xB0\xBD"
-     "\x02\x00"
-     "\xB1\x8D\x02\x00\xB2\x85\x02\x00\xB3\x6D\x01\x00\xB4\x79\x01\x00\xB5\x79"
-     "\x01\x00"
-     "\xB6\x69\x01\x00\xB7\x69\x01\x00\x86\x20\x04\x00\x87\x08\x05\x00\x7F\x4C"
-     "\x00\x80"
-     "\x83\x4C\x00\x80\x87\x4C\x00\x80\x8B\x4C\x00\x80\x8F\x4C\x00\x80\xEF\x5C"
-     "\x00\x00"
-     "\x84\xEC\x04\x00\xE1\x5C\x0E\x00\x93\x4C\x00\x80\xE3\x8C\x0E\x00\x97\x4C"
-     "\x00\x80"
-     "\x9B\x4C\x00\x80\x9F\x4C\x00\x80\xA3\x4C\x00\x80\xA3\x55\x02\x00\xA7\x4C"
-     "\x00\x80"
-     "\xAB\x4C\x00\x80\xAF\x4C\x00\x80\xB3\x4C\x00\x80\xA6\x1D\x02\x00\xA5\x1D"
-     "\x02\x00"
-     "\xB7\x4C\x00\x80\xAB\x09\x02\x00\xAA\x39\x02\x00\xBB\x4C\x00\x80\xBF\x4C"
-     "\x00\x80"
-     "\xAF\xFD\x01\x00\xAE\xFD\x01\x00\xAD\xF9\x01\x00\xAC\x11\x02\x00\xA8\x69"
-     "\x06\x00"
-     "\xA9\x69\x06\x00\xAA\x79\x06\x00\xAB\x79\x06\x00\xAC\x69\x06\x00\xAD\x69"
-     "\x06\x00"
-     "\xAE\x9D\x06\x00\xAF\x95\x06\x00\x73\x4C\x00\x80\xC3\x4C\x00\x80\xC7\x4C"
-     "\x00\x80"
-     "\xCB\x4C\x00\x80\xCF\x4C\x00\x80\xD3\x4C\x00\x80\xD7\x4C\x00\x80\xDB\x4C"
-     "\x00\x80"
-     "\xB8\xF5\x06\x00\xB9\xF9\x06\x00\xBA\x8D\x06\x00\xBB\x85\x06\x00\xBC\x9D"
-     "\x06\x00"
-     "\xBD\x85\x06\x00\xBE\x85\x06\x00\xBF\xB5\x06\x00\xB0\xED\x06\x00\xB1\xF5"
-     "\x06\x00"
-     "\xB2\xFD\x06\x00\xB3\xF5\x06\x00\xB4\xED\x06\x00\xB5\xD1\x06\x00\xB6\xD1"
-     "\x06\x00"
-     "\xB7\xD1\x06\x00\xB3\xF1\x06\x00\x82\x15\x00\x00\x81\xB5\x00\x00\x80\xB5"
-     "\x00\x00"
-     "\xDF\x4C\x00\x80\xB6\xE9\x06\x00\xB5\xE1\x06\x00\xBE\xD4\x03\x00\xBB\x31"
-     "\x06\x00"
-     "\xBA\x29\x06\x00\xE3\x4C\x00\x80\xE7\x4C\x00\x80\xBF\x15\x06\x00\xBE\x29"
-     "\x06\x00"
-     "\xBD\x21\x06\x00\xBC\x29\x06\x00\xEB\x4C\x00\x80\xA3\xB5\x06\x00\x86\xC8"
-     "\x00\x00"
-     "\x87\xC8\x00\x00\xA6\xAD\x06\x00\xEF\x4C\x00\x80\xF3\x4C\x00\x80\xA5\xA5"
-     "\x06\x00"
-     "\xAA\x6D\x06\x00\xAB\x75\x06\x00\xF7\x4C\x00\x80\xFB\x4C\x00\x80\xAE\x6D"
-     "\x06\x00"
-     "\xAF\x51\x06\x00\xAC\x6D\x06\x00\xAD\x65\x06\x00\xA8\x35\x06\x00\xA9\x39"
-     "\x06\x00"
-     "\xAA\x81\x06\x00\xAB\x81\x06\x00\xAC\x81\x06\x00\xAD\x81\x06\x00\xAE\x81"
-     "\x06\x00"
-     "\xAF\xB5\x06\x00\xFF\x4C\x00\x80\x03\x4D\x00\x80\x07\x4D\x00\x80\x0B\x4D"
-     "\x00\x80"
-     "\x0F\x4D\x00\x80\x13\x4D\x00\x80\x17\x4D\x00\x80\x1B\x4D\x00\x80\xB8\x9D"
-     "\x06\x00"
-     "\xB9\xAD\x06\x00\xBA\xA5\x06\x00\xBB\x69\x01\x00\xBC\x79\x01\x00\xBD\x79"
-     "\x01\x00"
-     "\xBE\x69\x01\x00\xBF\x69\x01\x00\xB0\xD1\x06\x00\xB1\xD1\x06\x00\xB2\xD1"
-     "\x06\x00"
-     "\xB3\xD1\x06\x00\xB4\xB5\x06\x00\xB5\xBD\x06\x00\xB6\xB5\x06\x00\xB7\xAD"
-     "\x06\x00"
-     "\xB3\x09\x06\x00\x1F\x4D\x00\x80\x23\x4D\x00\x80\x27\x4D\x00\x80\x2B\x4D"
-     "\x00\x80"
-     "\xB6\x01\x06\x00\xB5\x09\x06\x00\x2F\x4D\x00\x80\xBB\x15\x06\x00\xBA\x15"
-     "\x06\x00"
-     "\x33\x4D\x00\x80\x37\x4D\x00\x80\xBF\x79\x06\x00\xBE\x71\x06\x00\xBD\x05"
-     "\x06\x00"
-     "\xBC\x05\x06\x00\x3B\x4D\x00\x80\xA3\x4D\x06\x00\x3F\x4D\x00\x80\x03\x4C"
-     "\x00\x80"
-     "\xA6\x45\x06\x00\x43\x4D\x00\x80\x47\x4D\x00\x80\xA5\x4D\x06\x00\xAA\x51"
-     "\x06\x00"
-     "\xAB\x51\x06\x00\x4B\x4D\x00\x80\x4F\x4D\x00\x80\xAE\x35\x06\x00\xAF\x3D"
-     "\x06\x00"
-     "\xAC\x41\x06\x00\xAD\x41\x06\x00\x81\xE9\x03\x00\x80\xDD\x03\x00\x84\x88"
-     "\x03\x00"
-     "\x82\xE1\x03\x00\x86\xB0\x3C\x00\x87\x88\x02\x00\xBE\x54\x03\x00\x57\x4D"
-     "\x00\x80"
-     "\x5B\x4D\x00\x80\x5F\x4D\x00\x80\x63\x4D\x00\x80\x67\x4D\x00\x80\x6B\x4D"
-     "\x00\x80"
-     "\x6F\x4D\x00\x80\x73\x4D\x00\x80\x77\x4D\x00\x80\xE3\xF0\x06\x00\x7B\x4D"
-     "\x00\x80"
-     "\xE1\x30\x06\x00\x84\x00\x3C\x00\x7F\x4D\x00\x80\x83\x4D\x00\x80\x87\x4D"
-     "\x00\x80"
-     "\x8B\x4D\x00\x80\x8F\x4D\x00\x80\x85\x34\x3D\x00\x93\x4D\x00\x80\x97\x4D"
-     "\x00\x80"
-     "\xEF\xB0\x07\x00\x9B\x4D\x00\x80\x9F\x4D\x00\x80\xA3\x4D\x00\x80\xA7\x4D"
-     "\x00\x80"
-     "\xAB\x4D\x00\x80\xAF\x4D\x00\x80\xBE\xC4\x3C\x00\xB3\x4D\x00\x80\x82\x9D"
-     "\x01\x00"
-     "\x81\x9D\x01\x00\x80\x9D\x01\x00\xA8\x0D\x02\x00\xA9\x65\x02\x00\xAA\x7D"
-     "\x02\x00"
-     "\xAB\x75\x02\x00\xAC\x59\x02\x00\xAD\x59\x02\x00\xAE\x99\x03\x00\xAF\x99"
-     "\x03\x00"
-     "\xB0\xE9\x03\x00\xB1\xE9\x03\x00\xB2\xF9\x03\x00\xB3\xF9\x03\x00\xB4\xE9"
-     "\x03\x00"
-     "\xB5\xE9\x03\x00\xB6\x5D\x03\x00\xB7\x55\x03\x00\xB8\x6D\x03\x00\xB9\x75"
-     "\x03\x00"
-     "\xBA\x75\x03\x00\xBB\x45\x03\x00\xBC\x5D\x03\x00\xBD\x35\x03\x00\xBE\x31"
-     "\x03\x00"
-     "\xBF\x29\x03\x00\xB7\x4D\x00\x80\xBB\x4D\x00\x80\xBF\x4D\x00\x80\xC7\x4D"
-     "\x00\x80"
-     "\xE1\x60\x03\x00\xEF\xF4\x03\x00\xE3\x44\x02\x00\xCB\x4D\x00\x80\xCF\x4D"
-     "\x00\x80"
-     "\xE3\x3C\x03\x00\xEF\xF4\x37\x00\xE1\xFC\x01\x00\xD3\x4D\x00\x80\xD7\x4D"
-     "\x00\x80"
-     "\xDB\x4D\x00\x80\xDF\x4D\x00\x80\x86\x64\x3F\x00\x87\x68\x3D\x00\x85\x34"
-     "\x21\x00"
-     "\xB3\x99\x03\x00\xE3\x4D\x00\x80\xB5\xBD\x03\x00\xB6\xB5\x03\x00\xE7\x4D"
-     "\x00\x80"
-     "\xEB\x4D\x00\x80\xEF\x4D\x00\x80\xBA\x41\x02\x00\xBB\x41\x02\x00\xBC\x41"
-     "\x02\x00"
-     "\xBD\x41\x02\x00\xBE\x41\x02\x00\xBF\x41\x02\x00\xF3\x4D\x00\x80\xF7\x4D"
-     "\x00\x80"
-     "\xFB\x4D\x00\x80\xFF\x4D\x00\x80\x03\x4E\x00\x80\x07\x4E\x00\x80\x0B\x4E"
-     "\x00\x80"
-     "\xEF\xF8\x01\x00\x84\x68\x3C\x00\xE1\x3C\x06\x00\x0F\x4E\x00\x80\xE3\x1C"
-     "\x06\x00"
-     "\x13\x4E\x00\x80\x17\x4E\x00\x80\x1B\x4E\x00\x80\x1F\x4E\x00\x80\xA3\x15"
-     "\x03\x00"
-     "\x23\x4E\x00\x80\x27\x4E\x00\x80\x2B\x4E\x00\x80\x2F\x4E\x00\x80\xA6\x39"
-     "\x03\x00"
-     "\xA5\x31\x03\x00\x37\x4E\x00\x80\xAB\xCD\x02\x00\xAA\xCD\x02\x00\xBE\x64"
-     "\x3E\x00"
-     "\x3B\x4E\x00\x80\xAF\xCD\x02\x00\xAE\xCD\x02\x00\xAD\xCD\x02\x00\xAC\xCD"
-     "\x02\x00"
-     "\xA8\x1D\x3E\x00\xA9\x25\x3E\x00\xAA\x2D\x3E\x00\xAB\x25\x3E\x00\xAC\x3D"
-     "\x3E\x00"
-     "\xAD\x25\x3E\x00\xAE\x2D\x3E\x00\xAF\x25\x3E\x00\xC3\x4D\x00\x80\x82\xF5"
-     "\x3F\x00"
-     "\x81\xE5\x3F\x00\x80\xE5\x3F\x00\x33\x4E\x00\x80\x3F\x4E\x00\x80\x86\x00"
-     "\x04\x00"
-     "\x87\x9C\x03\x00\xB8\x2D\x3E\x00\xB9\x31\x3E\x00\xBA\x31\x3E\x00\xBB\x31"
-     "\x3E\x00"
-     "\xBC\xD1\x3E\x00\xBD\xD1\x3E\x00\xBE\xD1\x3E\x00\xBF\xCD\x3E\x00\xB0\x5D"
-     "\x3E\x00"
-     "\xB1\x21\x3E\x00\xB2\x39\x3E\x00\xB3\x39\x3E\x00\xB4\x29\x3E\x00\xB5\x29"
-     "\x3E\x00"
-     "\xB6\x19\x3E\x00\xB7\x15\x3E\x00\xB3\xA5\x3E\x00\x43\x4E\x00\x80\x47\x4E"
-     "\x00\x80"
-     "\x4B\x4E\x00\x80\x4F\x4E\x00\x80\xB6\xA5\x3E\x00\xB5\xB5\x3E\x00\x53\x4E"
-     "\x00\x80"
-     "\xBB\xE5\x3E\x00\xBA\x99\x3E\x00\x57\x4E\x00\x80\x5B\x4E\x00\x80\xBF\xED"
-     "\x3E\x00"
-     "\xBE\xED\x3E\x00\xBD\xED\x3E\x00\xBC\xED\x3E\x00\x5F\x4E\x00\x80\xA3\xE1"
-     "\x3E\x00"
-     "\x63\x4E\x00\x80\x67\x4E\x00\x80\xA6\xE1\x3E\x00\x6B\x4E\x00\x80\x6F\x4E"
-     "\x00\x80"
-     "\xA5\xF1\x3E\x00\xAA\xDD\x3E\x00\xAB\xA1\x3E\x00\x73\x4E\x00\x80\x77\x4E"
-     "\x00\x80"
-     "\xAE\xA9\x3E\x00\xAF\xA9\x3E\x00\xAC\xA9\x3E\x00\xAD\xA9\x3E\x00\x8F\x05"
-     "\x25\x00"
-     "\xB3\x25\x3E\x00\x7B\x4E\x00\x80\x7F\x4E\x00\x80\xB6\x29\x3E\x00\x83\x4E"
-     "\x00\x80"
-     "\x87\x4E\x00\x80\xB5\x29\x3E\x00\xBA\x7D\x3E\x00\xBB\x45\x3E\x00\x8B\x4E"
-     "\x00\x80"
-     "\x8F\x4E\x00\x80\xBE\xB5\x3E\x00\xBF\xBD\x3E\x00\xBC\x5D\x3E\x00\xBD\xBD"
-     "\x3E\x00"
-     "\x9F\x7D\x38\x00\x9E\x65\x39\x00\x9D\xF1\x38\x00\x9C\x51\x34\x00\x9B\x59"
-     "\x35\x00"
-     "\x9A\x51\x35\x00\x99\xF1\x30\x00\x98\x0D\x31\x00\x97\x65\x31\x00\x96\xC1"
-     "\x30\x00"
-     "\x95\x59\x2D\x00\x94\x4D\x2D\x00\x93\xE1\x2C\x00\x92\xD9\x29\x00\x91\x59"
-     "\x29\x00"
-     "\x90\xF1\x28\x00\xB4\xA5\x19\x00\xB5\xDD\x18\x00\x93\x4E\x00\x80\x84\x08"
-     "\x00\x00"
-     "\xB0\x91\x15\x00\xB1\x01\x15\x00\xB2\x01\x18\x00\xB3\xBD\x19\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\xAD\x03\x00\x82\xA5\x03\x00\x97\x4E\x00\x80\xA3\x61\x00\x00\xA2\x1D"
-     "\x3D\x00"
-     "\xA1\x99\x3D\x00\xA0\x71\x3C\x00\xA4\xC5\x05\x00\xA5\x41\x04\x00\xA6\x01"
-     "\x08\x00"
-     "\xA7\xE1\x09\x00\x53\x4D\x00\x80\xA1\xF5\x01\x00\xA2\xE9\x01\x00\xA3\xDD"
-     "\x05\x00"
-     "\xAC\x01\x10\x00\xAD\xC5\x11\x00\xAE\xD9\x11\x00\xAF\x39\x10\x00\xA8\x65"
-     "\x08\x00"
-     "\xA9\x01\x0C\x00\xAA\xD9\x0D\x00\xAB\x09\x0D\x00\x8A\x31\x2E\x00\x8B\xA1"
-     "\x33\x00"
-     "\x9B\x4E\x00\x80\x9F\x4E\x00\x80\x8E\xB9\x33\x00\x8F\x11\x36\x00\x8C\x1D"
-     "\x32\x00"
-     "\x8D\x4D\x32\x00\x82\x25\x26\x00\x83\xA9\x2B\x00\xBE\x64\x03\x00\x84\x60"
-     "\x04\x00"
-     "\x86\xA1\x2F\x00\x87\x15\x2E\x00\x84\x19\x2A\x00\x85\x91\x2A\x00\x9A\x61"
-     "\x3E\x00"
-     "\x9B\xED\x3E\x00\x86\xC8\x04\x00\x87\xDC\x03\x00\xA7\x4E\x00\x80\xE1\x5C"
-     "\x3E\x00"
-     "\x9C\x89\x03\x00\xE3\x00\x3E\x00\x92\x69\x36\x00\x93\x79\x37\x00\xAB\x4E"
-     "\x00\x80"
-     "\xEF\x18\x3E\x00\x96\x4D\x3B\x00\x97\xB9\x3F\x00\x94\xC1\x3B\x00\x95\x5D"
-     "\x3A\x00"
-     "\xA9\x9D\x3D\x00\xA8\x89\x3D\x00\xAB\xB9\x3D\x00\xAA\xAD\x3D\x00\xAD\xA9"
-     "\x3D\x00"
-     "\xAC\xA1\x3D\x00\xAF\xC9\x3D\x00\xAE\xA1\x3D\x00\xBE\xE8\x04\x00\xAF\x4E"
-     "\x00\x80"
-     "\xB3\x4E\x00\x80\xB7\x4E\x00\x80\xBB\x4E\x00\x80\xBF\x4E\x00\x80\xC3\x4E"
-     "\x00\x80"
-     "\xC7\x4E\x00\x80\xB9\x59\x3D\x00\xB8\x51\x3D\x00\xBB\x79\x3D\x00\xBA\x65"
-     "\x3D\x00"
-     "\xBD\x69\x3D\x00\xBC\x61\x3D\x00\xBF\x1D\x3D\x00\xBE\x61\x3D\x00\xB1\x81"
-     "\x3D\x00"
-     "\xB0\xB9\x3D\x00\xB3\x69\x3D\x00\xB2\x89\x3D\x00\xB5\x79\x3D\x00\xB4\x71"
-     "\x3D\x00"
-     "\xB7\x69\x3D\x00\xB6\x71\x3D\x00\xA3\x21\x3C\x00\xCB\x4E\x00\x80\xCF\x4E"
-     "\x00\x80"
-     "\xD3\x4E\x00\x80\xD7\x4E\x00\x80\xA6\x2D\x3C\x00\xA5\x2D\x3C\x00\xDB\x4E"
-     "\x00\x80"
-     "\xAB\x41\x3C\x00\xAA\x79\x3C\x00\xDF\x4E\x00\x80\xE3\x4E\x00\x80\xAF\xB9"
-     "\x3C\x00"
-     "\xAE\xB1\x3C\x00\xAD\xB9\x3C\x00\xAC\x59\x3C\x00\xE7\x4E\x00\x80\xEB\x4E"
-     "\x00\x80"
-     "\xB3\x7D\x03\x00\xEF\x4E\x00\x80\xB5\xD9\x03\x00\xF3\x4E\x00\x80\xF7\x4E"
-     "\x00\x80"
-     "\xB6\xD1\x03\x00\xFB\x4E\x00\x80\xFF\x4E\x00\x80\xBB\xC5\x03\x00\xBA\xC5"
-     "\x03\x00"
-     "\xBD\xB9\x03\x00\xBC\xB5\x03\x00\xBF\xAD\x03\x00\xBE\xB1\x03\x00\x03\x4F"
-     "\x00\x80"
-     "\x07\x4F\x00\x80\x0B\x4F\x00\x80\xEF\x5C\x03\x00\x80\x15\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x82\x31\x00\x00\xEF\x8C\x3E\x00\x84\xEC\x04\x00\xE1\xFC\x3E\x00\x0F\x4F"
-     "\x00\x80"
-     "\xE3\x1C\x3E\x00\x17\x4F\x00\x80\xE1\x94\x01\x00\x1B\x4F\x00\x80\xE3\x20"
-     "\x00\x00"
-     "\xA3\xF5\x03\x00\x1F\x4F\x00\x80\x87\xE8\x04\x00\x86\x6C\x04\x00\x23\x4F"
-     "\x00\x80"
-     "\xA6\x59\x03\x00\xA5\x51\x03\x00\x27\x4F\x00\x80\xAB\x4D\x03\x00\xAA\x4D"
-     "\x03\x00"
-     "\x2B\x4F\x00\x80\x2F\x4F\x00\x80\xAF\x25\x03\x00\xAE\x39\x03\x00\xAD\x31"
-     "\x03\x00"
-     "\xAC\x3D\x03\x00\xA3\x4E\x00\x80\x13\x4F\x00\x80\x33\x4F\x00\x80\x37\x4F"
-     "\x00\x80"
-     "\x3B\x4F\x00\x80\x3F\x4F\x00\x80\x43\x4F\x00\x80\x47\x4F\x00\x80\xA8\x71"
-     "\x06\x00"
-     "\xA9\x4D\x06\x00\xAA\x8D\x06\x00\xAB\x85\x06\x00\xAC\x9D\x06\x00\xAD\x85"
-     "\x06\x00"
-     "\xAE\x8D\x06\x00\xAF\x85\x06\x00\xB0\xFD\x06\x00\xB1\x41\x07\x00\xB2\x41"
-     "\x07\x00"
-     "\xB3\x41\x07\x00\xB4\x41\x07\x00\xB5\x49\x07\x00\xB6\x71\x07\x00\xB7\x71"
-     "\x07\x00"
-     "\xB8\x21\x07\x00\xB9\x21\x07\x00\xBA\x25\x07\x00\xBB\x39\x07\x00\xBC\x29"
-     "\x07\x00"
-     "\xBD\x15\x07\x00\xBE\x1D\x07\x00\xBF\xFD\x07\x00\xB3\x25\x06\x00\x4B\x4F"
-     "\x00\x80"
-     "\x4F\x4F\x00\x80\x53\x4F\x00\x80\x57\x4F\x00\x80\xB6\x25\x06\x00\xB5\x35"
-     "\x06\x00"
-     "\x5B\x4F\x00\x80\xBB\xA5\x07\x00\xBA\x19\x06\x00\x5F\x4F\x00\x80\x63\x4F"
-     "\x00\x80"
-     "\xBF\xAD\x07\x00\xBE\xA5\x07\x00\xBD\xB5\x07\x00\xBC\xB5\x07\x00\x67\x4F"
-     "\x00\x80"
-     "\xA3\x61\x06\x00\x6B\x4F\x00\x80\x6F\x4F\x00\x80\xA6\x61\x06\x00\x73\x4F"
-     "\x00\x80"
-     "\x77\x4F\x00\x80\xA5\x71\x06\x00\xAA\x5D\x06\x00\xAB\xE1\x07\x00\x7B\x4F"
-     "\x00\x80"
-     "\x7F\x4F\x00\x80\xAE\xE1\x07\x00\xAF\xE9\x07\x00\xAC\xF1\x07\x00\xAD\xF1"
-     "\x07\x00"
-     "\xA8\xB1\x06\x00\xA9\xBD\x06\x00\xAA\xCD\x06\x00\xAB\xDD\x06\x00\xAC\xCD"
-     "\x06\x00"
-     "\xAD\xFD\x06\x00\xAE\xF5\x06\x00\xAF\x15\x01\x00\x80\xF9\x01\x00\x81\xCD"
-     "\x01\x00"
-     "\x82\xC5\x01\x00\xBE\x64\x02\x00\x86\x90\x00\x00\x87\x00\x01\x00\x87\x4F"
-     "\x00\x80"
-     "\x8B\x4F\x00\x80\xB8\xD1\x01\x00\xB9\xD9\x01\x00\xBA\xE1\x01\x00\xBB\xE1"
-     "\x01\x00"
-     "\xBC\x91\x01\x00\xBD\x9D\x01\x00\xBE\x95\x01\x00\xBF\x89\x01\x00\xB0\x6D"
-     "\x01\x00"
-     "\xB1\x75\x01\x00\xB2\x7D\x01\x00\xB3\x75\x01\x00\xB4\x6D\x01\x00\xB5\xF1"
-     "\x01\x00"
-     "\xB6\xF1\x01\x00\xB7\xF1\x01\x00\xB3\x45\x06\x00\x83\x4F\x00\x80\x8F\x4F"
-     "\x00\x80"
-     "\x93\x4F\x00\x80\x97\x4F\x00\x80\xB6\x7D\x06\x00\xB5\x45\x06\x00\x9B\x4F"
-     "\x00\x80"
-     "\xBB\xB1\x01\x00\xBA\xA9\x01\x00\x9F\x4F\x00\x80\xA3\x4F\x00\x80\xBF\x35"
-     "\x01\x00"
-     "\xBE\xA9\x01\x00\xBD\xA1\x01\x00\xBC\xA9\x01\x00\xA7\x4F\x00\x80\xA3\x01"
-     "\x06\x00"
-     "\xAB\x4F\x00\x80\xAF\x4F\x00\x80\xA6\x39\x06\x00\xB3\x4F\x00\x80\xB7\x4F"
-     "\x00\x80"
-     "\xA5\x01\x06\x00\xAA\xED\x01\x00\xAB\xF5\x01\x00\xBB\x4F\x00\x80\xBF\x4F"
-     "\x00\x80"
-     "\xAE\xED\x01\x00\xAF\x71\x01\x00\xAC\xED\x01\x00\xAD\xE5\x01\x00\xE1\x28"
-     "\x01\x00"
-     "\xC3\x4F\x00\x80\xE3\x50\x02\x00\xC7\x4F\x00\x80\xCB\x4F\x00\x80\xCF\x4F"
-     "\x00\x80"
-     "\xD3\x4F\x00\x80\xD7\x4F\x00\x80\xDB\x4F\x00\x80\xDF\x4F\x00\x80\xE3\x4F"
-     "\x00\x80"
-     "\xE7\x4F\x00\x80\x81\x71\x00\x00\x80\x19\x00\x00\xEF\x9C\x02\x00\x82\x79"
-     "\x00\x00"
-     "\xEB\x4F\x00\x80\xEF\x4F\x00\x80\x84\xC8\x02\x00\xB3\xC5\x03\x00\xF7\x4F"
-     "\x00\x80"
-     "\xB5\xC5\x03\x00\xB6\xC5\x03\x00\xBE\x10\x03\x00\x86\xC0\x0C\x00\x87\x44"
-     "\x0C\x00"
-     "\xBA\xA9\x03\x00\xBB\xA5\x03\x00\xBC\xBD\x03\x00\xBD\xA1\x03\x00\xBE\xA1"
-     "\x03\x00"
-     "\xBF\x95\x03\x00\xAE\x11\x06\x00\xAF\x19\x06\x00\xAC\x01\x06\x00\xAD\x01"
-     "\x06\x00"
-     "\xAA\xA5\x06\x00\xAB\x11\x06\x00\xA8\x45\x39\x00\xA9\x71\x39\x00\xFB\x4F"
-     "\x00\x80"
-     "\xFF\x4F\x00\x80\x03\x50\x00\x80\x07\x50\x00\x80\x0B\x50\x00\x80\x0F\x50"
-     "\x00\x80"
-     "\x13\x50\x00\x80\x17\x50\x00\x80\xBE\xED\x07\x00\xBF\x4D\x07\x00\xBC\xD1"
-     "\x07\x00"
-     "\xBD\xE5\x07\x00\xBA\xDD\x07\x00\xBB\xC1\x07\x00\xB8\x35\x06\x00\xB9\xD5"
-     "\x07\x00"
-     "\xB6\x39\x06\x00\xB7\x0D\x06\x00\xB4\x25\x06\x00\xB5\x39\x06\x00\xB2\x31"
-     "\x06\x00"
-     "\xB3\x3D\x06\x00\xB0\x51\x06\x00\xB1\x51\x06\x00\xA8\x39\x02\x00\xA9\x39"
-     "\x02\x00"
-     "\xAA\x81\x02\x00\xAB\x81\x02\x00\xAC\x81\x02\x00\xAD\x89\x02\x00\xAE\xB1"
-     "\x02\x00"
-     "\xAF\xB1\x02\x00\x84\x6C\x0D\x00\x1B\x50\x00\x80\xBE\x60\x0D\x00\x1F\x50"
-     "\x00\x80"
-     "\x23\x50\x00\x80\x27\x50\x00\x80\x2B\x50\x00\x80\x2F\x50\x00\x80\xB8\x4D"
-     "\x01\x00"
-     "\xB9\x55\x01\x00\xBA\x5D\x01\x00\xBB\x55\x01\x00\xBC\x4D\x01\x00\xBD\x75"
-     "\x01\x00"
-     "\xBE\x7D\x01\x00\xBF\x75\x01\x00\xB0\xA1\x02\x00\xB1\xAD\x02\x00\xB2\xA5"
-     "\x02\x00"
-     "\xB3\xB9\x02\x00\xB4\xA9\x02\x00\xB5\x9D\x02\x00\xB6\x95\x02\x00\xB7\x79"
-     "\x01\x00"
-     "\xE1\x54\x06\x00\xE1\xD4\x07\x00\xE3\x38\x06\x00\xE3\xB0\x07\x00\x33\x50"
-     "\x00\x80"
-     "\x37\x50\x00\x80\x3B\x50\x00\x80\x3F\x50\x00\x80\x84\xE4\x0C\x00\x43\x50"
-     "\x00\x80"
-     "\x47\x50\x00\x80\x4B\x50\x00\x80\x4F\x50\x00\x80\x53\x50\x00\x80\xEF\x6C"
-     "\x00\x00"
-     "\xEF\xE4\x07\x00\xA3\x49\x02\x00\x57\x50\x00\x80\x5B\x50\x00\x80\x5F\x50"
-     "\x00\x80"
-     "\x67\x50\x00\x80\xA6\x49\x02\x00\xA5\x49\x02\x00\x6B\x50\x00\x80\xAB\x29"
-     "\x02\x00"
-     "\xAA\x25\x02\x00\x86\x48\x0C\x00\x87\xDC\x0C\x00\xAF\x19\x02\x00\xAE\x2D"
-     "\x02\x00"
-     "\xAD\x2D\x02\x00\xAC\x31\x02\x00\xA8\x51\x0E\x00\xA9\xA5\x0E\x00\xAA\xAD"
-     "\x0E\x00"
-     "\xAB\xA5\x0E\x00\xAC\xBD\x0E\x00\xAD\xA5\x0E\x00\xAE\xAD\x0E\x00\xAF\xA5"
-     "\x0E\x00"
-     "\x80\xE5\x0F\x00\x81\xED\x0F\x00\x82\xE5\x0F\x00\x63\x50\x00\x80\x6F\x50"
-     "\x00\x80"
-     "\x73\x50\x00\x80\x77\x50\x00\x80\x7B\x50\x00\x80\xB8\xD5\x0F\x00\xB9\xDD"
-     "\x0F\x00"
-     "\xBA\xD5\x0F\x00\xBB\xE9\x0F\x00\xBC\xF9\x0F\x00\xBD\xF9\x0F\x00\xBE\xE9"
-     "\x0F\x00"
-     "\xBF\xE9\x0F\x00\xB0\xDD\x0E\x00\xB1\x41\x0F\x00\xB2\x45\x0F\x00\xB3\x5D"
-     "\x0F\x00"
-     "\xB4\x45\x0F\x00\xB5\x4D\x0F\x00\xB6\x45\x0F\x00\xB7\xED\x0F\x00\xB3\x25"
-     "\x0E\x00"
-     "\x7F\x50\x00\x80\x83\x50\x00\x80\x87\x50\x00\x80\x8B\x50\x00\x80\xB6\x25"
-     "\x0E\x00"
-     "\xB5\x35\x0E\x00\x8F\x50\x00\x80\xBB\x85\x0F\x00\xBA\x19\x0E\x00\x93\x50"
-     "\x00\x80"
-     "\x97\x50\x00\x80\xBF\x89\x0F\x00\xBE\x81\x0F\x00\xBD\x89\x0F\x00\xBC\x91"
-     "\x0F\x00"
-     "\x9B\x50\x00\x80\xA3\x61\x0E\x00\x9F\x50\x00\x80\xA3\x50\x00\x80\xA6\x61"
-     "\x0E\x00"
-     "\xA7\x50\x00\x80\xAB\x50\x00\x80\xA5\x71\x0E\x00\xAA\x5D\x0E\x00\xAB\xC1"
-     "\x0F\x00"
-     "\xAF\x50\x00\x80\xB3\x50\x00\x80\xAE\xC5\x0F\x00\xAF\xCD\x0F\x00\xAC\xD5"
-     "\x0F\x00"
-     "\xAD\xCD\x0F\x00\xA8\xD1\x0E\x00\xA9\xD9\x0E\x00\xAA\x39\x01\x00\xAB\x39"
-     "\x01\x00"
-     "\xAC\x29\x01\x00\xAD\x29\x01\x00\xAE\x9D\x01\x00\xAF\x95\x01\x00\xB7\x50"
-     "\x00\x80"
-     "\xBB\x50\x00\x80\xBF\x50\x00\x80\xC3\x50\x00\x80\x80\x0D\x00\x00\x81\xB5"
-     "\x00\x00"
-     "\x82\xBD\x00\x00\xC7\x50\x00\x80\xB8\x95\x01\x00\xB9\x9D\x01\x00\xBA\xA1"
-     "\x01\x00"
-     "\xBB\xA1\x01\x00\xBC\x71\x00\x00\xBD\x71\x00\x00\xBE\x71\x00\x00\xBF\x71"
-     "\x00\x00"
-     "\xB0\xED\x01\x00\xB1\xF5\x01\x00\xB2\xF5\x01\x00\xB3\xC5\x01\x00\xB4\xDD"
-     "\x01\x00"
-     "\xB5\xB5\x01\x00\xB6\xB1\x01\x00\xB7\xB1\x01\x00\xCB\x50\x00\x80\xCF\x50"
-     "\x00\x80"
-     "\xB3\xB9\x01\x00\xBE\xC0\x02\x00\xB5\xA9\x01\x00\xD3\x50\x00\x80\xD7\x50"
-     "\x00\x80"
-     "\xB6\xA1\x01\x00\x86\x80\x01\x00\x87\xC4\x01\x00\xBB\x39\x01\x00\xBA\x21"
-     "\x01\x00"
-     "\xBD\x19\x01\x00\xBC\x29\x01\x00\xBF\x79\x01\x00\xBE\x11\x01\x00\xA3\xF1"
-     "\x01\x00"
-     "\xDB\x50\x00\x80\xF3\x4F\x00\x80\xDF\x50\x00\x80\xE3\x50\x00\x80\xA6\xE9"
-     "\x01\x00"
-     "\xA5\xE1\x01\x00\xE7\x50\x00\x80\xAB\x71\x01\x00\xAA\x69\x01\x00\xEB\x50"
-     "\x00\x80"
-     "\xEF\x50\x00\x80\xAF\x31\x01\x00\xAE\x59\x01\x00\xAD\x51\x01\x00\xAC\x61"
-     "\x01\x00"
-     "\xF3\x50\x00\x80\xF7\x50\x00\x80\xFB\x50\x00\x80\xFF\x50\x00\x80\xE1\xA8"
-     "\x01\x00"
-     "\x03\x51\x00\x80\xE3\x40\x02\x00\x07\x51\x00\x80\xBF\x28\x15\x00\x0B\x51"
-     "\x00\x80"
-     "\xEF\x74\x02\x00\x0F\x51\x00\x80\x13\x51\x00\x80\x17\x51\x00\x80\x1B\x51"
-     "\x00\x80"
-     "\x1F\x51\x00\x80\x23\x51\x00\x80\xE3\x74\x0F\x00\x84\x88\x03\x00\xE1\x34"
-     "\x0E\x00"
-     "\x27\x51\x00\x80\x2B\x51\x00\x80\x2F\x51\x00\x80\x33\x51\x00\x80\x80\x0D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x11\x00\x00\x37\x51\x00\x80\x3B\x51\x00\x80\xEF\xA4"
-     "\x0F\x00"
-     "\x3F\x51\x00\x80\x47\x51\x00\x80\xA8\x19\x03\x00\xA9\x41\x03\x00\xAA\x45"
-     "\x03\x00"
-     "\xAB\x5D\x03\x00\xAC\x4D\x03\x00\xAD\x7D\x03\x00\xAE\x75\x03\x00\xAF\x9D"
-     "\x00\x00"
-     "\x85\xA4\x15\x00\xBE\x7C\x03\x00\x86\x08\x04\x00\x87\x1C\x03\x00\x4B\x51"
-     "\x00\x80"
-     "\x4F\x51\x00\x80\x53\x51\x00\x80\x57\x51\x00\x80\xB8\xF5\x00\x00\xB9\xFD"
-     "\x00\x00"
-     "\xBA\xF5\x00\x00\xBB\x8D\x00\x00\xBC\x81\x00\x00\xBD\x81\x00\x00\xBE\x81"
-     "\x00\x00"
-     "\xBF\x81\x00\x00\xB0\xE5\x00\x00\xB1\xED\x00\x00\xB2\xE5\x00\x00\xB3\xF9"
-     "\x00\x00"
-     "\xB4\xED\x00\x00\xB5\xD1\x00\x00\xB6\xD5\x00\x00\xB7\xCD\x00\x00\x5B\x51"
-     "\x00\x80"
-     "\x5F\x51\x00\x80\x63\x51\x00\x80\xB3\xC1\x02\x00\x67\x51\x00\x80\xB5\xC9"
-     "\x02\x00"
-     "\xB6\xF1\x02\x00\x6B\x51\x00\x80\x6F\x51\x00\x80\x73\x51\x00\x80\xBA\x2D"
-     "\x01\x00"
-     "\xBB\x25\x01\x00\xBC\x3D\x01\x00\xBD\x21\x01\x00\xBE\x25\x01\x00\xBF\x19"
-     "\x01\x00"
-     "\xA6\xA9\x02\x00\x84\x48\x02\x00\x77\x51\x00\x80\xA5\x91\x02\x00\x7B\x51"
-     "\x00\x80"
-     "\xA3\x99\x02\x00\x7F\x51\x00\x80\x83\x51\x00\x80\xAE\x7D\x01\x00\xAF\x41"
-     "\x01\x00"
-     "\xAC\x65\x01\x00\xAD\x79\x01\x00\xAA\x75\x01\x00\xAB\x7D\x01\x00\x87\x51"
-     "\x00\x80"
-     "\x8B\x51\x00\x80\x8F\x51\x00\x80\x93\x51\x00\x80\x97\x51\x00\x80\x9B\x51"
-     "\x00\x80"
-     "\xEF\xE4\x00\x00\x9F\x51\x00\x80\xA3\x51\x00\x80\xA7\x51\x00\x80\xE3\x10"
-     "\x0E\x00"
-     "\xAB\x51\x00\x80\xE1\x58\x0E\x00\xAF\x51\x00\x80\x80\x2D\x00\x00\x81\x11"
-     "\x00\x00"
-     "\x82\x15\x00\x00\xBE\xB0\x05\x00\xB3\x75\x01\x00\xB7\x51\x00\x80\x87\x14"
-     "\x05\x00"
-     "\x86\x6C\x04\x00\xBB\x51\x00\x80\xB6\xD5\x00\x00\xB5\x65\x01\x00\xBF\x51"
-     "\x00\x80"
-     "\xBB\xFD\x00\x00\xBA\xF5\x00\x00\xC3\x51\x00\x80\xC7\x51\x00\x80\xBF\xA1"
-     "\x00\x00"
-     "\xBE\xBD\x00\x00\xBD\xDD\x00\x00\xBC\xDD\x00\x00\xA8\x7D\x06\x00\xA9\xB5"
-     "\x06\x00"
-     "\xAA\xBD\x06\x00\xAB\xB5\x06\x00\xAC\xAD\x06\x00\xAD\x15\x07\x00\xAE\x1D"
-     "\x07\x00"
-     "\xAF\x15\x07\x00\xB3\x51\x00\x80\xCB\x51\x00\x80\xCF\x51\x00\x80\xD3\x51"
-     "\x00\x80"
-     "\xD7\x51\x00\x80\xDB\x51\x00\x80\xDF\x51\x00\x80\xE3\x51\x00\x80\xB8\x39"
-     "\x07\x00"
-     "\xB9\x39\x07\x00\xBA\xC9\x07\x00\xBB\xC9\x07\x00\xBC\xD9\x07\x00\xBD\xD9"
-     "\x07\x00"
-     "\xBE\xCD\x07\x00\xBF\xC5\x07\x00\xB0\x71\x07\x00\xB1\x79\x07\x00\xB2\x49"
-     "\x07\x00"
-     "\xB3\x49\x07\x00\xB4\x39\x07\x00\xB5\x25\x07\x00\xB6\x21\x07\x00\xB7\x21"
-     "\x07\x00"
-     "\xA3\x35\x06\x00\xE7\x51\x00\x80\xEB\x51\x00\x80\xEF\x51\x00\x80\xF3\x51"
-     "\x00\x80"
-     "\xA6\x95\x07\x00\xA5\x25\x06\x00\xF7\x51\x00\x80\xAB\xBD\x07\x00\xAA\xB5"
-     "\x07\x00"
-     "\xFB\x51\x00\x80\xFF\x51\x00\x80\xAF\xE1\x07\x00\xAE\xFD\x07\x00\xAD\x9D"
-     "\x07\x00"
-     "\xAC\x9D\x07\x00\x03\x52\x00\x80\x07\x52\x00\x80\x0B\x52\x00\x80\x0F\x52"
-     "\x00\x80"
-     "\x13\x52\x00\x80\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D\x00\x00\x17\x52"
-     "\x00\x80"
-     "\x1B\x52\x00\x80\x1F\x52\x00\x80\x84\xA0\x03\x00\xBE\xA4\x03\x00\x23\x52"
-     "\x00\x80"
-     "\x86\xF8\x00\x00\x87\xE0\x00\x00\xA8\xC5\x06\x00\xA9\xD5\x06\x00\xAA\xD5"
-     "\x06\x00"
-     "\xAB\xE5\x06\x00\xAC\xFD\x06\x00\xAD\x31\x01\x00\xAE\x31\x01\x00\xAF\x31"
-     "\x01\x00"
-     "\x27\x52\x00\x80\x2B\x52\x00\x80\x2F\x52\x00\x80\x33\x52\x00\x80\x37\x52"
-     "\x00\x80"
-     "\x3B\x52\x00\x80\x3F\x52\x00\x80\x43\x52\x00\x80\xB8\xDD\x01\x00\xB9\xED"
-     "\x01\x00"
-     "\xBA\xE5\x01\x00\xBB\x8D\x01\x00\xBC\x95\x01\x00\xBD\x9D\x01\x00\xBE\x95"
-     "\x01\x00"
-     "\xBF\x8D\x01\x00\xB0\x51\x01\x00\xB1\x51\x01\x00\xB2\x51\x01\x00\xB3\x51"
-     "\x01\x00"
-     "\xB4\xF5\x01\x00\xB5\xFD\x01\x00\xB6\xF5\x01\x00\xB7\xED\x01\x00\xB3\x5D"
-     "\x06\x00"
-     "\x47\x52\x00\x80\x4B\x52\x00\x80\x4F\x52\x00\x80\x53\x52\x00\x80\xB6\xB1"
-     "\x01\x00"
-     "\xB5\x75\x06\x00\x57\x52\x00\x80\xBB\x95\x01\x00\xBA\x95\x01\x00\x5B\x52"
-     "\x00\x80"
-     "\x5F\x52\x00\x80\xBF\x39\x01\x00\xBE\x31\x01\x00\xBD\x85\x01\x00\xBC\x85"
-     "\x01\x00"
-     "\xA5\x2D\x06\x00\x63\x52\x00\x80\x67\x52\x00\x80\xA6\xE9\x01\x00\x6B\x52"
-     "\x00\x80"
-     "\x6F\x52\x00\x80\xA3\x05\x06\x00\x73\x52\x00\x80\xAD\xDD\x01\x00\xAC\xDD"
-     "\x01\x00"
-     "\xAF\x61\x01\x00\xAE\x69\x01\x00\x77\x52\x00\x80\x43\x51\x00\x80\xAB\xCD"
-     "\x01\x00"
-     "\xAA\xCD\x01\x00\x7B\x52\x00\x80\x7F\x52\x00\x80\x84\xC4\x03\x00\xBF\x00"
-     "\xF4\x00"
-     "\x83\x52\x00\x80\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D\x00\x00\x87\x52"
-     "\x00\x80"
-     "\x8B\x52\x00\x80\x8F\x52\x00\x80\xBE\x60\x03\x00\x97\x52\x00\x80\x9B\x52"
-     "\x00\x80"
-     "\x9F\x52\x00\x80\xA3\x52\x00\x80\x86\xE0\x1C\x00\x87\x00\x03\x00\xEF\x0C"
-     "\x07\x00"
-     "\xA7\x52\x00\x80\xAB\x52\x00\x80\xAF\x52\x00\x80\xB3\x52\x00\x80\xB7\x52"
-     "\x00\x80"
-     "\xBB\x52\x00\x80\xBF\x52\x00\x80\xC3\x52\x00\x80\xC7\x52\x00\x80\xE1\xC0"
-     "\x06\x00"
-     "\xCB\x52\x00\x80\xE3\x2C\x07\x00\xCF\x52\x00\x80\xD3\x52\x00\x80\xD7\x52"
-     "\x00\x80"
-     "\xDB\x52\x00\x80\xDF\x52\x00\x80\xE3\x52\x00\x80\xE7\x52\x00\x80\xEB\x52"
-     "\x00\x80"
-     "\xEF\x52\x00\x80\xA8\x81\x03\x00\xA9\x81\x03\x00\xAA\x81\x03\x00\xAB\x81"
-     "\x03\x00"
-     "\xAC\x81\x03\x00\xAD\x81\x03\x00\xAE\x81\x03\x00\xAF\x81\x03\x00\xB0\x45"
-     "\x03\x00"
-     "\xB1\x4D\x03\x00\xB2\x45\x03\x00\xB3\x5D\x03\x00\xB4\x4D\x03\x00\xB5\x7D"
-     "\x03\x00"
-     "\xB6\x75\x03\x00\xB7\x19\x03\x00\xB8\x29\x03\x00\xB9\x35\x03\x00\xBA\x3D"
-     "\x03\x00"
-     "\xBB\x31\x03\x00\xBC\x01\x03\x00\xBD\xF5\x00\x00\xBE\xFD\x00\x00\xBF\xED"
-     "\x00\x00"
-     "\xB3\x29\x02\x00\xF3\x52\x00\x80\xF7\x52\x00\x80\xFB\x52\x00\x80\xFF\x52"
-     "\x00\x80"
-     "\xB6\x21\x02\x00\xB5\x29\x02\x00\x84\x50\x1D\x00\xBB\xA9\x02\x00\xBA\xA1"
-     "\x02\x00"
-     "\x07\x53\x00\x80\x0B\x53\x00\x80\xBF\x99\x02\x00\xBE\xA9\x02\x00\xBD\xA9"
-     "\x02\x00"
-     "\xBC\xB1\x02\x00\x81\x4D\x00\x00\x80\x4D\x00\x00\xEF\x9C\x03\x00\x82\x5D"
-     "\x00\x00"
-     "\x86\xF0\x1C\x00\x87\x78\x1D\x00\xBE\x10\x1D\x00\x0F\x53\x00\x80\x13\x53"
-     "\x00\x80"
-     "\x17\x53\x00\x80\x1B\x53\x00\x80\x1F\x53\x00\x80\xE1\x90\x01\x00\x23\x53"
-     "\x00\x80"
-     "\xE3\x60\x03\x00\x27\x53\x00\x80\x2B\x53\x00\x80\x2F\x53\x00\x80\x33\x53"
-     "\x00\x80"
-     "\x37\x53\x00\x80\x3B\x53\x00\x80\x3F\x53\x00\x80\x43\x53\x00\x80\xEF\x94"
-     "\x01\x00"
-     "\x84\xEC\x1C\x00\xE1\x70\x06\x00\x47\x53\x00\x80\xE3\x50\x01\x00\x4B\x53"
-     "\x00\x80"
-     "\x4F\x53\x00\x80\x53\x53\x00\x80\x57\x53\x00\x80\xA3\xE9\x02\x00\x5B\x53"
-     "\x00\x80"
-     "\x5F\x53\x00\x80\x63\x53\x00\x80\x67\x53\x00\x80\xA6\xE1\x02\x00\xA5\xE9"
-     "\x02\x00"
-     "\x6B\x53\x00\x80\xAB\x69\x02\x00\xAA\x61\x02\x00\x6F\x53\x00\x80\xBE\xA8"
-     "\x1C\x00"
-     "\xAF\x59\x02\x00\xAE\x69\x02\x00\xAD\x69\x02\x00\xAC\x71\x02\x00\xA8\x31"
-     "\x1E\x00"
-     "\xA9\x31\x1E\x00\xAA\x31\x1E\x00\xAB\x31\x1E\x00\xAC\x5D\x1E\x00\xAD\x45"
-     "\x1E\x00"
-     "\xAE\x4D\x1E\x00\xAF\x45\x1E\x00\x03\x53\x00\x80\x82\xCD\x1F\x00\x81\xF5"
-     "\x1F\x00"
-     "\x80\xFD\x1F\x00\x73\x53\x00\x80\x77\x53\x00\x80\x86\x1C\x00\x00\x87\xF8"
-     "\x03\x00"
-     "\xB8\xC5\x1E\x00\xB9\xCD\x1E\x00\xBA\xC5\x1E\x00\xBB\xDD\x1E\x00\xBC\xC5"
-     "\x1E\x00"
-     "\xBD\xCD\x1E\x00\xBE\xC5\x1E\x00\xBF\x59\x1F\x00\xB0\x3D\x1E\x00\xB1\x05"
-     "\x1E\x00"
-     "\xB2\x0D\x1E\x00\xB3\x05\x1E\x00\xB4\x1D\x1E\x00\xB5\x01\x1E\x00\xB6\x05"
-     "\x1E\x00"
-     "\xB7\xFD\x1E\x00\xB3\xB9\x1E\x00\x7B\x53\x00\x80\x7F\x53\x00\x80\x83\x53"
-     "\x00\x80"
-     "\x87\x53\x00\x80\xB6\xC5\x1E\x00\xB5\xD5\x1E\x00\x8B\x53\x00\x80\xBB\xC1"
-     "\x1E\x00"
-     "\xBA\xF9\x1E\x00\x8F\x53\x00\x80\x93\x53\x00\x80\xBF\xC5\x1E\x00\xBE\xD9"
-     "\x1E\x00"
-     "\xBD\xD1\x1E\x00\xBC\xD9\x1E\x00\x97\x53\x00\x80\xA3\xFD\x1E\x00\x9B\x53"
-     "\x00\x80"
-     "\x9F\x53\x00\x80\xA6\x81\x1E\x00\xA3\x53\x00\x80\xA7\x53\x00\x80\xA5\x91"
-     "\x1E\x00"
-     "\xAA\xBD\x1E\x00\xAB\x85\x1E\x00\xAB\x53\x00\x80\xAF\x53\x00\x80\xAE\x9D"
-     "\x1E\x00"
-     "\xAF\x81\x1E\x00\xAC\x9D\x1E\x00\xAD\x95\x1E\x00\xA8\x29\x1E\x00\xA9\x29"
-     "\x1E\x00"
-     "\xAA\x55\x1E\x00\xAB\x6D\x1E\x00\xAC\x75\x1E\x00\xAD\x7D\x1E\x00\xAE\x75"
-     "\x1E\x00"
-     "\xAF\x69\x1E\x00\xB3\x53\x00\x80\xB7\x53\x00\x80\xBB\x53\x00\x80\xBF\x53"
-     "\x00\x80"
-     "\xC3\x53\x00\x80\xC7\x53\x00\x80\xCB\x53\x00\x80\xCF\x53\x00\x80\xB8\xE9"
-     "\x1E\x00"
-     "\xB9\xF5\x1E\x00\xBA\xFD\x1E\x00\xBB\xF5\x1E\x00\xBC\xED\x1E\x00\xBD\x91"
-     "\x1E\x00"
-     "\xBE\x91\x1E\x00\xBF\x91\x1E\x00\xB0\x1D\x1E\x00\xB1\xE5\x1E\x00\xB2\xED"
-     "\x1E\x00"
-     "\xB3\xE5\x1E\x00\xB4\xFD\x1E\x00\xB5\xE5\x1E\x00\xB6\xED\x1E\x00\xB7\xE5"
-     "\x1E\x00"
-     "\xB3\xDD\x1E\x00\xD3\x53\x00\x80\xD7\x53\x00\x80\xDB\x53\x00\x80\xDF\x53"
-     "\x00\x80"
-     "\xB6\xFD\x1E\x00\xB5\xFD\x1E\x00\x84\x58\x01\x00\xBB\x21\x1E\x00\xBA\xD9"
-     "\x1E\x00"
-     "\xBE\x28\x00\x00\xE7\x53\x00\x80\xBF\x21\x1E\x00\xBE\x39\x1E\x00\xBD\x31"
-     "\x1E\x00"
-     "\xBC\x39\x1E\x00\x81\x4D\x00\x00\x80\x4D\x00\x00\xA3\x95\x1E\x00\x82\x5D"
-     "\x00\x00"
-     "\xA5\xB5\x1E\x00\xE3\x53\x00\x80\xEB\x53\x00\x80\xA6\xB5\x1E\x00\x93\x52"
-     "\x00\x80"
-     "\xEF\x53\x00\x80\xAB\x69\x1E\x00\xAA\x91\x1E\x00\xAD\x79\x1E\x00\xAC\x71"
-     "\x1E\x00"
-     "\xAF\x69\x1E\x00\xAE\x71\x1E\x00\x86\x00\x04\x00\x87\x44\x03\x00\xB3\x81"
-     "\x02\x00"
-     "\xF3\x53\x00\x80\xB5\x81\x02\x00\xF7\x53\x00\x80\xFB\x53\x00\x80\xB6\x81"
-     "\x02\x00"
-     "\x88\x00\x00\x00\xFF\x53\x00\x80\xBB\xE1\x02\x00\xBA\xED\x02\x00\xBD\xE5"
-     "\x02\x00"
-     "\xBC\xF9\x02\x00\xBF\xD1\x02\x00\xBE\xE5\x02\x00\x03\x54\x00\x80\x07\x54"
-     "\x00\x80"
-     "\x84\x44\x03\x00\xBE\x8C\x03\x00\xE1\x48\x02\x00\x0B\x54\x00\x80\xE3\x00"
-     "\x02\x00"
-     "\xEF\xFC\x1F\x00\x0F\x54\x00\x80\xE1\x3C\x1E\x00\xEF\xDC\x02\x00\xE3\x60"
-     "\x1F\x00"
-     "\x13\x54\x00\x80\x17\x54\x00\x80\x1B\x54\x00\x80\x1F\x54\x00\x80\xA9\x05"
-     "\x02\x00"
-     "\xA8\xB1\x02\x00\xAB\x05\x02\x00\xAA\x05\x02\x00\xAD\x0D\x02\x00\xAC\x05"
-     "\x02\x00"
-     "\xAF\x35\x02\x00\xAE\x35\x02\x00\x84\x6C\x05\x00\x23\x54\x00\x80\x27\x54"
-     "\x00\x80"
-     "\x2B\x54\x00\x80\x2F\x54\x00\x80\x33\x54\x00\x80\x37\x54\x00\x80\x3B\x54"
-     "\x00\x80"
-     "\xB9\xE9\x03\x00\xB8\xE1\x03\x00\xBB\xF9\x03\x00\xBA\xE1\x03\x00\xBD\xE9"
-     "\x03\x00"
-     "\xBC\xE1\x03\x00\xBF\x5D\x03\x00\xBE\xE1\x03\x00\xB1\x29\x02\x00\xB0\x25"
-     "\x02\x00"
-     "\xB3\x3D\x02\x00\xB2\x21\x02\x00\xB5\x19\x02\x00\xB4\x2D\x02\x00\xB7\xD9"
-     "\x03\x00"
-     "\xB6\x11\x02\x00\x3F\x54\x00\x80\x43\x54\x00\x80\x47\x54\x00\x80\xA3\x85"
-     "\x03\x00"
-     "\x4B\x54\x00\x80\xA5\x85\x03\x00\xA6\x85\x03\x00\x4F\x54\x00\x80\x57\x54"
-     "\x00\x80"
-     "\x5B\x54\x00\x80\xAA\xE9\x03\x00\xAB\xE5\x03\x00\xAC\xFD\x03\x00\xAD\xE1"
-     "\x03\x00"
-     "\xAE\xE1\x03\x00\xAF\xD5\x03\x00\x80\x49\x00\x00\x81\x55\x00\x00\x82\x55"
-     "\x00\x00"
-     "\xA3\xA9\x02\x00\xBE\x98\x04\x00\xA5\x41\x01\x00\xA6\x41\x01\x00\x5F\x54"
-     "\x00\x80"
-     "\x86\xE0\x05\x00\x87\xE0\x05\x00\xAA\x2D\x01\x00\xAB\x39\x01\x00\xAC\x11"
-     "\x01\x00"
-     "\xAD\x05\x01\x00\xAE\x0D\x01\x00\xAF\x05\x01\x00\x63\x54\x00\x80\x67\x54"
-     "\x00\x80"
-     "\x6B\x54\x00\x80\xEF\xDC\x00\x00\x6F\x54\x00\x80\x73\x54\x00\x80\x77\x54"
-     "\x00\x80"
-     "\xEF\x88\x1E\x00\x84\x2C\x07\x00\xE1\xEC\x1E\x00\x7B\x54\x00\x80\xE3\x1C"
-     "\x1E\x00"
-     "\x7F\x54\x00\x80\xE1\x94\x01\x00\x83\x54\x00\x80\xE3\x30\x00\x00\xB3\x25"
-     "\x02\x00"
-     "\x85\x60\xE6\x00\x87\x54\x00\x80\x8B\x54\x00\x80\x8F\x54\x00\x80\xB6\xCD"
-     "\x01\x00"
-     "\xB5\xCD\x01\x00\x93\x54\x00\x80\xBB\xB5\x01\x00\xBA\xA1\x01\x00\x97\x54"
-     "\x00\x80"
-     "\x9B\x54\x00\x80\xBF\x89\x01\x00\xBE\x81\x01\x00\xBD\x89\x01\x00\xBC\x9D"
-     "\x01\x00"
-     "\x53\x54\x00\x80\x9F\x54\x00\x80\xA3\x54\x00\x80\xA7\x54\x00\x80\xAB\x54"
-     "\x00\x80"
-     "\xAF\x54\x00\x80\xB3\x54\x00\x80\xB7\x54\x00\x80\xA8\x01\x07\x00\xA9\x01"
-     "\x07\x00"
-     "\xAA\x75\x07\x00\xAB\xAD\x07\x00\xAC\xB5\x07\x00\xAD\xBD\x07\x00\xAE\xA9"
-     "\x07\x00"
-     "\xAF\xA9\x07\x00\xB0\xD9\x07\x00\xB1\xED\x07\x00\xB2\xF9\x07\x00\xB3\xF5"
-     "\x07\x00"
-     "\xB4\x99\x07\x00\xB5\x99\x07\x00\xB6\x89\x07\x00\xB7\x81\x07\x00\xB8\x89"
-     "\x07\x00"
-     "\xB9\x89\x07\x00\xBA\x6D\x00\x00\xBB\x65\x00\x00\xBC\x7D\x00\x00\xBD\x65"
-     "\x00\x00"
-     "\xBE\x6D\x00\x00\xBF\x65\x00\x00\x81\x09\x00\x00\x80\x99\x00\x00\xBB\x54"
-     "\x00\x80"
-     "\x82\x1D\x00\x00\xBF\x54\x00\x80\xC3\x54\x00\x80\xC7\x54\x00\x80\xCB\x54"
-     "\x00\x80"
-     "\xA8\x0D\x05\x00\xA9\x7D\x05\x00\xAA\x4D\x05\x00\xAB\xA1\x06\x00\xAC\xA5"
-     "\x06\x00"
-     "\xAD\xAD\x06\x00\xAE\x9D\x06\x00\xAF\xFD\x06\x00\xB0\x85\x06\x00\xB1\x91"
-     "\x06\x00"
-     "\xB2\xA9\x06\x00\xB3\xBD\x06\x00\xB4\xA5\x06\x00\xB5\xAD\x06\x00\xB6\xA5"
-     "\x06\x00"
-     "\xB7\x79\x06\x00\xB8\x49\x06\x00\xB9\x49\x06\x00\xBA\x59\x06\x00\xBB\x59"
-     "\x06\x00"
-     "\xBC\x49\x06\x00\xBD\x49\x06\x00\xBE\xF9\x07\x00\xBF\xF9\x07\x00\xB3\x5D"
-     "\x06\x00"
-     "\xCF\x54\x00\x80\x86\x28\x02\x00\x87\x2C\x01\x00\xD3\x54\x00\x80\xB6\x9D"
-     "\x06\x00"
-     "\xB5\x9D\x06\x00\xD7\x54\x00\x80\xBB\x89\x06\x00\xBA\xBD\x06\x00\xDB\x54"
-     "\x00\x80"
-     "\xDF\x54\x00\x80\xBF\xFD\x06\x00\xBE\xFD\x06\x00\xBD\x81\x06\x00\xBC\x8D"
-     "\x06\x00"
-     "\xE3\x54\x00\x80\xA3\x19\x06\x00\xE7\x54\x00\x80\xEB\x54\x00\x80\xA6\xD9"
-     "\x06\x00"
-     "\xEF\x54\x00\x80\xF3\x54\x00\x80\xA5\xD9\x06\x00\xAA\xF9\x06\x00\xAB\xCD"
-     "\x06\x00"
-     "\xF7\x54\x00\x80\xFB\x54\x00\x80\xAE\xB9\x06\x00\xAF\xB9\x06\x00\xAC\xC9"
-     "\x06\x00"
-     "\xAD\xC5\x06\x00\xA8\x19\x01\x00\xA9\x19\x01\x00\xAA\x8D\x00\x00\xAB\x9D"
-     "\x00\x00"
-     "\xAC\x8D\x00\x00\xAD\xBD\x00\x00\xAE\xB5\x00\x00\xAF\xDD\x00\x00\xFF\x54"
-     "\x00\x80"
-     "\x03\x55\x00\x80\x07\x55\x00\x80\x0B\x55\x00\x80\x0F\x55\x00\x80\x13\x55"
-     "\x00\x80"
-     "\x17\x55\x00\x80\x1B\x55\x00\x80\xB8\x69\x00\x00\xB9\x69\x00\x00\xBA\x79"
-     "\x00\x00"
-     "\xBB\x79\x00\x00\xBC\x69\x00\x00\xBD\x69\x00\x00\xBE\xDD\x03\x00\xBF\xD5"
-     "\x03\x00"
-     "\xB0\xA9\x00\x00\xB1\xA9\x00\x00\xB2\xBD\x00\x00\xB3\xB5\x00\x00\xB4\x99"
-     "\x00\x00"
-     "\xB5\x99\x00\x00\xB6\x59\x00\x00\xB7\x59\x00\x00\xBE\x2C\x02\x00\x1F\x55"
-     "\x00\x80"
-     "\x23\x55\x00\x80\x27\x55\x00\x80\x2B\x55\x00\x80\x2F\x55\x00\x80\x37\x55"
-     "\x00\x80"
-     "\x3B\x55\x00\x80\x80\x2D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\x3F\x55"
-     "\x00\x80"
-     "\x86\x90\x0C\x00\x87\xE8\x02\x00\x43\x55\x00\x80\x47\x55\x00\x80\xB3\x45"
-     "\x03\x00"
-     "\x4B\x55\x00\x80\x4F\x55\x00\x80\x53\x55\x00\x80\x57\x55\x00\x80\xB6\x7D"
-     "\x03\x00"
-     "\xB5\x45\x03\x00\x5B\x55\x00\x80\xBB\x2D\x03\x00\xBA\x2D\x03\x00\x5F\x55"
-     "\x00\x80"
-     "\x63\x55\x00\x80\xBF\x1D\x03\x00\xBE\x1D\x03\x00\xBD\x21\x03\x00\xBC\x29"
-     "\x03\x00"
-     "\xAB\xCD\x03\x00\xAA\xCD\x03\x00\x67\x55\x00\x80\x6B\x55\x00\x80\xAF\xFD"
-     "\x03\x00"
-     "\xAE\xFD\x03\x00\xAD\xC1\x03\x00\xAC\xC9\x03\x00\xA3\xA5\x03\x00\x6F\x55"
-     "\x00\x80"
-     "\x73\x55\x00\x80\x77\x55\x00\x80\x7B\x55\x00\x80\xA6\x9D\x03\x00\xA5\xA5"
-     "\x03\x00"
-     "\x7F\x55\x00\x80\x83\x55\x00\x80\x87\x55\x00\x80\x8B\x55\x00\x80\x8F\x55"
-     "\x00\x80"
-     "\x93\x55\x00\x80\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D\x00\x00\x97\x55"
-     "\x00\x80"
-     "\x9B\x55\x00\x80\x9F\x55\x00\x80\x84\x60\x03\x00\x86\xD0\x0C\x00\x87\x30"
-     "\x03\x00"
-     "\xA7\x55\x00\x80\xAB\x55\x00\x80\xBE\x44\x02\x00\xAF\x55\x00\x80\xB3\x55"
-     "\x00\x80"
-     "\xB7\x55\x00\x80\xE1\x00\x00\x00\xBB\x55\x00\x80\xE3\xA8\x06\x00\xBF\x55"
-     "\x00\x80"
-     "\x84\xEC\x0C\x00\xC3\x55\x00\x80\xEF\xD0\x03\x00\xC7\x55\x00\x80\xCB\x55"
-     "\x00\x80"
-     "\xCF\x55\x00\x80\xD3\x55\x00\x80\xD7\x55\x00\x80\xB3\x6D\x01\x00\xDB\x55"
-     "\x00\x80"
-     "\xDF\x55\x00\x80\xE3\x55\x00\x80\xE7\x55\x00\x80\xB6\x11\x01\x00\xB5\x65"
-     "\x01\x00"
-     "\xEB\x55\x00\x80\xBB\x3D\x01\x00\xBA\x35\x01\x00\xEF\x55\x00\x80\xF3\x55"
-     "\x00\x80"
-     "\xBF\xFD\x01\x00\xBE\xFD\x01\x00\xBD\x15\x01\x00\xBC\x15\x01\x00\xF7\x55"
-     "\x00\x80"
-     "\xE1\xFC\x06\x00\xFB\x55\x00\x80\xE3\x3C\x07\x00\xFF\x55\x00\x80\x03\x56"
-     "\x00\x80"
-     "\x07\x56\x00\x80\x0B\x56\x00\x80\x0F\x56\x00\x80\xBE\x6C\x0C\x00\x17\x56"
-     "\x00\x80"
-     "\x1B\x56\x00\x80\x1F\x56\x00\x80\x23\x56\x00\x80\x27\x56\x00\x80\xEF\x14"
-     "\x06\x00"
-     "\x81\x5D\x00\x00\x80\x5D\x00\x00\xA3\xE5\x01\x00\x82\x6D\x00\x00\xA5\xED"
-     "\x01\x00"
-     "\x2B\x56\x00\x80\x2F\x56\x00\x80\xA6\x99\x01\x00\x87\xA8\x0C\x00\x86\xE4"
-     "\x0C\x00"
-     "\xAB\xB5\x01\x00\xAA\xBD\x01\x00\xAD\x9D\x01\x00\xAC\x9D\x01\x00\xAF\x75"
-     "\x01\x00"
-     "\xAE\x75\x01\x00\xA8\x19\x0E\x00\xA9\x19\x0E\x00\xAA\x25\x0E\x00\xAB\x35"
-     "\x0E\x00"
-     "\xAC\x2D\x0E\x00\xAD\x51\x0E\x00\xAE\x51\x0E\x00\xAF\x51\x0E\x00\xA3\x55"
-     "\x00\x80"
-     "\x13\x56\x00\x80\x33\x56\x00\x80\x37\x56\x00\x80\x3B\x56\x00\x80\x3F\x56"
-     "\x00\x80"
-     "\x43\x56\x00\x80\x47\x56\x00\x80\xB8\xED\x0E\x00\xB9\xF5\x0E\x00\xBA\xF5"
-     "\x0E\x00"
-     "\xBB\x8D\x0E\x00\xBC\x95\x0E\x00\xBD\x9D\x0E\x00\xBE\x95\x0E\x00\xBF\x8D"
-     "\x0E\x00"
-     "\xB0\x31\x0E\x00\xB1\x39\x0E\x00\xB2\x01\x0E\x00\xB3\x01\x0E\x00\xB4\xF9"
-     "\x0E\x00"
-     "\xB5\xF9\x0E\x00\xB6\xDD\x0E\x00\xB7\xD5\x0E\x00\xA8\x79\x0E\x00\xA9\x79"
-     "\x0E\x00"
-     "\xAA\x8D\x0F\x00\xAB\x85\x0F\x00\xAC\x9D\x0F\x00\xAD\x81\x0F\x00\xAE\x85"
-     "\x0F\x00"
-     "\xAF\xB9\x0F\x00\x4B\x56\x00\x80\x4F\x56\x00\x80\x53\x56\x00\x80\x57\x56"
-     "\x00\x80"
-     "\x5B\x56\x00\x80\x5F\x56\x00\x80\x63\x56\x00\x80\x67\x56\x00\x80\xB8\x91"
-     "\x0F\x00"
-     "\xB9\x99\x0F\x00\xBA\xA1\x0F\x00\xBB\xA1\x0F\x00\xBC\x51\x0F\x00\xBD\x5D"
-     "\x0F\x00"
-     "\xBE\x49\x0F\x00\xBF\x49\x0F\x00\xB0\xCD\x0F\x00\xB1\xD5\x0F\x00\xB2\xDD"
-     "\x0F\x00"
-     "\xB3\xD5\x0F\x00\xB4\xCD\x0F\x00\xB5\xB1\x0F\x00\xB6\xB1\x0F\x00\xB7\xB1"
-     "\x0F\x00"
-     "\xB3\x05\x0E\x00\x6B\x56\x00\x80\x6F\x56\x00\x80\x73\x56\x00\x80\x77\x56"
-     "\x00\x80"
-     "\xB6\x01\x0E\x00\xB5\x15\x0E\x00\x7B\x56\x00\x80\xBB\x11\x0E\x00\xBA\x09"
-     "\x0E\x00"
-     "\x7F\x56\x00\x80\x84\xA0\x01\x00\xBF\x75\x0E\x00\xBE\x09\x0E\x00\xBD\x01"
-     "\x0E\x00"
-     "\xBC\x09\x0E\x00\x82\x69\x00\x00\xA3\x41\x0E\x00\x80\x59\x00\x00\x81\x51"
-     "\x00\x00"
-     "\xA6\x45\x0E\x00\xBE\x58\x01\x00\x83\x56\x00\x80\xA5\x51\x0E\x00\xAA\x4D"
-     "\x0E\x00"
-     "\xAB\x55\x0E\x00\x86\xC8\x00\x00\x87\xAC\x01\x00\xAE\x4D\x0E\x00\xAF\x31"
-     "\x0E\x00"
-     "\xAC\x4D\x0E\x00\xAD\x45\x0E\x00\x87\x56\x00\x80\x8B\x56\x00\x80\x8F\x56"
-     "\x00\x80"
-     "\x93\x56\x00\x80\x97\x56\x00\x80\x9B\x56\x00\x80\x33\x55\x00\x80\x9F\x56"
-     "\x00\x80"
-     "\xA8\x09\x0E\x00\xA9\x09\x0E\x00\xAA\x19\x0E\x00\xAB\x19\x0E\x00\xAC\x09"
-     "\x0E\x00"
-     "\xAD\x61\x0E\x00\xAE\x61\x0E\x00\xAF\x95\x01\x00\xB0\xED\x01\x00\xB1\xF5"
-     "\x01\x00"
-     "\xB2\xFD\x01\x00\xB3\xF5\x01\x00\xB4\xED\x01\x00\xB5\x75\x01\x00\xB6\x7D"
-     "\x01\x00"
-     "\xB7\x75\x01\x00\xB8\x4D\x01\x00\xB9\x55\x01\x00\xBA\x5D\x01\x00\xBB\x55"
-     "\x01\x00"
-     "\xBC\x4D\x01\x00\xBD\xF1\x00\x00\xBE\xF1\x00\x00\xBF\xF1\x00\x00\xA3\x56"
-     "\x00\x80"
-     "\xA7\x56\x00\x80\xAB\x56\x00\x80\xAF\x56\x00\x80\xB3\x56\x00\x80\xE3\xE8"
-     "\x0E\x00"
-     "\xB7\x56\x00\x80\xE1\x34\x0E\x00\xBE\x00\x04\x00\xEF\xDC\x0F\x00\xBB\x56"
-     "\x00\x80"
-     "\xBF\x56\x00\x80\xC3\x56\x00\x80\xC7\x56\x00\x80\xCB\x56\x00\x80\xCF\x56"
-     "\x00\x80"
-     "\xB3\xF1\x0D\x00\xD3\x56\x00\x80\xD7\x56\x00\x80\xDB\x56\x00\x80\xDF\x56"
-     "\x00\x80"
-     "\xB6\x81\x0D\x00\xB5\xE1\x0D\x00\xE3\x56\x00\x80\xBB\x51\x02\x00\xBA\x49"
-     "\x02\x00"
-     "\xE7\x56\x00\x80\xEB\x56\x00\x80\xBF\xD1\x02\x00\xBE\x49\x02\x00\xBD\x41"
-     "\x02\x00"
-     "\xBC\x49\x02\x00\xA3\x31\x0D\x00\xEF\x56\x00\x80\x84\x88\x03\x00\xF7\x56"
-     "\x00\x80"
-     "\xFB\x56\x00\x80\xA6\x41\x0D\x00\xA5\x21\x0D\x00\xFF\x56\x00\x80\xAB\x91"
-     "\x02\x00"
-     "\xAA\x89\x02\x00\x03\x57\x00\x80\x07\x57\x00\x80\xAF\x11\x02\x00\xAE\x89"
-     "\x02\x00"
-     "\xAD\x81\x02\x00\xAC\x89\x02\x00\x80\xA9\x00\x00\x81\xA9\x00\x00\x82\x4D"
-     "\x00\x00"
-     "\x0B\x57\x00\x80\xE1\x64\x12\x00\xE3\x4C\x02\x00\xE3\x08\x0B\x00\xE1\xAC"
-     "\x01\x00"
-     "\x0F\x57\x00\x80\xEF\x3C\x02\x00\xEF\x18\x16\x00\x13\x57\x00\x80\x86\x50"
-     "\x06\x00"
-     "\x87\x48\x03\x00\x17\x57\x00\x80\x1B\x57\x00\x80\xA8\x81\x02\x00\xA9\x81"
-     "\x02\x00"
-     "\xAA\x81\x02\x00\xAB\x81\x02\x00\xAC\x81\x02\x00\xAD\x81\x02\x00\xAE\x85"
-     "\x02\x00"
-     "\xAF\x1D\x01\x00\x1F\x57\x00\x80\x23\x57\x00\x80\x27\x57\x00\x80\x2B\x57"
-     "\x00\x80"
-     "\x2F\x57\x00\x80\x33\x57\x00\x80\x37\x57\x00\x80\x84\xB8\x05\x00\xB8\x75"
-     "\x01\x00"
-     "\xB9\x7D\x01\x00\xBA\x75\x01\x00\xBB\x09\x01\x00\xBC\x19\x01\x00\xBD\x19"
-     "\x01\x00"
-     "\xBE\x09\x01\x00\xBF\x01\x01\x00\xB0\x65\x01\x00\xB1\x6D\x01\x00\xB2\x65"
-     "\x01\x00"
-     "\xB3\x7D\x01\x00\xB4\x69\x01\x00\xB5\x5D\x01\x00\xB6\x55\x01\x00\xB7\x4D"
-     "\x01\x00"
-     "\x3B\x57\x00\x80\x3F\x57\x00\x80\x43\x57\x00\x80\x47\x57\x00\x80\x4B\x57"
-     "\x00\x80"
-     "\x4F\x57\x00\x80\xEF\x34\x00\x00\xEF\xC0\x0E\x00\xE1\x5C\x0E\x00\xE1\x4C"
-     "\x0F\x00"
-     "\xE3\x94\x00\x00\xE3\x9C\x0E\x00\x53\x57\x00\x80\x82\x65\x00\x00\x81\x7D"
-     "\x00\x00"
-     "\x80\x7D\x00\x00\x57\x57\x00\x80\x5B\x57\x00\x80\xBE\xC4\x07\x00\xB3\x45"
-     "\x02\x00"
-     "\x5F\x57\x00\x80\xB5\x45\x02\x00\xB6\x4D\x02\x00\x67\x57\x00\x80\x86\x40"
-     "\x06\x00"
-     "\x87\x90\x04\x00\xBA\xED\x01\x00\xBB\xE5\x01\x00\xBC\xFD\x01\x00\xBD\xE5"
-     "\x01\x00"
-     "\xBE\xE1\x01\x00\xBF\xD5\x01\x00\x9F\x95\x08\x00\x9E\x05\x08\x00\x9D\xDD"
-     "\x0D\x00"
-     "\x9C\x3D\x0C\x00\x9B\x31\x0C\x00\x9A\xF5\x0D\x00\x99\xED\x10\x00\x98\x3D"
-     "\x10\x00"
-     "\x97\xD5\x11\x00\x96\xB1\x15\x00\x95\x05\x14\x00\x94\xE5\x15\x00\x93\xB5"
-     "\x19\x00"
-     "\x92\x31\x18\x00\x91\x39\x18\x00\x90\x0D\x1C\x00\x8F\x61\x1C\x00\xF3\x56"
-     "\x00\x80"
-     "\xB3\xD5\x06\x00\x6B\x57\x00\x80\xB5\xFD\x06\x00\x63\x57\x00\x80\x6F\x57"
-     "\x00\x80"
-     "\xB6\x91\x06\x00\x73\x57\x00\x80\x77\x57\x00\x80\xBB\x95\x06\x00\xBA\x95"
-     "\x06\x00"
-     "\xBD\x55\x07\x00\xBC\x55\x07\x00\xBF\x55\x07\x00\xBE\x55\x07\x00\x7B\x57"
-     "\x00\x80"
-     "\x7F\x57\x00\x80\xAA\x8D\x06\x00\xAB\x85\x06\x00\xAC\x9D\x06\x00\xAD\x85"
-     "\x06\x00"
-     "\xAE\x81\x06\x00\xAF\xB5\x06\x00\x84\xA8\x00\x00\x83\x57\x00\x80\x87\x57"
-     "\x00\x80"
-     "\xA3\x25\x05\x00\x8B\x57\x00\x80\xA5\x25\x05\x00\xA6\x2D\x05\x00\x8F\x57"
-     "\x00\x80"
-     "\x93\x57\x00\x80\x97\x57\x00\x80\x9B\x57\x00\x80\x9F\x57\x00\x80\xA3\x57"
-     "\x00\x80"
-     "\xA7\x57\x00\x80\xAB\x57\x00\x80\xAF\x57\x00\x80\xB3\x57\x00\x80\xB7\x57"
-     "\x00\x80"
-     "\xBB\x57\x00\x80\xBF\x57\x00\x80\xA3\xA9\x05\x00\xA2\xD1\x04\x00\xA1\xD9"
-     "\x04\x00"
-     "\xA0\x65\x05\x00\x82\x21\x1D\x00\x83\x35\x1D\x00\xC3\x57\x00\x80\xC7\x57"
-     "\x00\x80"
-     "\x86\x95\x18\x00\x87\xDD\x14\x00\x84\x19\x19\x00\x85\x19\x19\x00\x8A\x0D"
-     "\x15\x00"
-     "\x8B\xB1\x14\x00\xCB\x57\x00\x80\xCF\x57\x00\x80\x8E\xC5\x11\x00\x8F\xD5"
-     "\x0C\x00"
-     "\x8C\xCD\x10\x00\x8D\x1D\x11\x00\x92\x61\x0D\x00\x93\x75\x0D\x00\xBE\x4C"
-     "\x00\x00"
-     "\xD7\x57\x00\x80\x96\xC5\x09\x00\x97\xCD\x04\x00\x94\x8D\x0C\x00\x95\x5D"
-     "\x09\x00"
-     "\x9A\x41\x05\x00\x9B\x41\x05\x00\x86\xC8\xFF\x00\x87\x4C\x00\x00\x81\x59"
-     "\x00\x00"
-     "\x80\x79\x00\x00\x9C\x21\x04\x00\x82\x51\x00\x00\xA1\xC5\x01\x00\xDB\x57"
-     "\x00\x80"
-     "\xA3\x01\xFC\x00\xA2\xD9\x01\x00\xA5\x15\xFD\x00\xA4\xB5\xFD\x00\xA7\xB9"
-     "\xF9\x00"
-     "\xA6\x01\xF8\x00\xA9\x09\xF8\x00\xA8\x95\xF9\x00\xAB\x05\xF5\x00\xAA\xAD"
-     "\xF5\x00"
-     "\xAD\xB1\xF1\x00\xAC\x01\xF0\x00\xAF\x1D\xF0\x00\xAE\xAD\xF1\x00\xB1\xE1"
-     "\xED\x00"
-     "\xB0\x01\xEC\x00\xB3\x01\xE8\x00\xB2\xFD\xEC\x00\xB5\x5D\xE9\x00\xB4\xF5"
-     "\xE9\x00"
-     "\xDF\x57\x00\x80\xE3\x57\x00\x80\xE7\x57\x00\x80\xEB\x57\x00\x80\xEF\x57"
-     "\x00\x80"
-     "\xF3\x57\x00\x80\xF7\x57\x00\x80\xFB\x57\x00\x80\xFF\x57\x00\x80\x03\x58"
-     "\x00\x80"
-     "\xA8\x8D\x04\x00\xA9\x95\x04\x00\xAA\x95\x04\x00\xAB\xA5\x04\x00\xAC\xBD"
-     "\x04\x00"
-     "\xAD\xD9\x04\x00\xAE\xF9\x04\x00\xAF\xF1\x04\x00\x84\x6C\xFC\x00\x07\x58"
-     "\x00\x80"
-     "\x0B\x58\x00\x80\x0F\x58\x00\x80\x13\x58\x00\x80\x17\x58\x00\x80\x1B\x58"
-     "\x00\x80"
-     "\x1F\x58\x00\x80\xB8\x79\x05\x00\xB9\xC5\x05\x00\xBA\xCD\x05\x00\xBB\xC5"
-     "\x05\x00"
-     "\xBC\xDD\x05\x00\xBD\xC5\x05\x00\xBE\xCD\x05\x00\xBF\xED\x05\x00\xB0\x99"
-     "\x04\x00"
-     "\xB1\x99\x04\x00\xB2\x49\x05\x00\xB3\x49\x05\x00\xB4\x59\x05\x00\xB5\x59"
-     "\x05\x00"
-     "\xB6\x49\x05\x00\xB7\x49\x05\x00\xA3\x1D\x04\x00\xBE\xCC\xFC\x00\x23\x58"
-     "\x00\x80"
-     "\x27\x58\x00\x80\x2B\x58\x00\x80\xA6\x59\x04\x00\xA5\x35\x04\x00\x2F\x58"
-     "\x00\x80"
-     "\xAB\x5D\x04\x00\xAA\x5D\x04\x00\x33\x58\x00\x80\x37\x58\x00\x80\xAF\x9D"
-     "\x05\x00"
-     "\xAE\x9D\x05\x00\xAD\x9D\x05\x00\xAC\x9D\x05\x00\x3B\x58\x00\x80\xB3\x9D"
-     "\x02\x00"
-     "\x3F\x58\x00\x80\x47\x58\x00\x80\xB6\x51\x02\x00\x4B\x58\x00\x80\x4F\x58"
-     "\x00\x80"
-     "\xB5\xB9\x02\x00\xBA\x49\x02\x00\xBB\x55\x02\x00\x86\x48\xFD\x00\x87\xC0"
-     "\xFC\x00"
-     "\xBE\x35\x02\x00\xBF\x3D\x02\x00\xBC\x45\x02\x00\xBD\x3D\x02\x00\xA8\xDD"
-     "\x04\x00"
-     "\xA9\x49\x03\x00\xAA\x51\x03\x00\xAB\x6D\x03\x00\xAC\x75\x03\x00\xAD\x95"
-     "\x03\x00"
-     "\xAE\x9D\x03\x00\xAF\xB9\x03\x00\x80\x8D\x01\x00\x81\xE5\x01\x00\x82\xE1"
-     "\x01\x00"
-     "\x53\x58\x00\x80\x57\x58\x00\x80\x5B\x58\x00\x80\x5F\x58\x00\x80\x63\x58"
-     "\x00\x80"
-     "\xB8\x95\x03\x00\xB9\x9D\x03\x00\xBA\x95\x03\x00\xBB\xAD\x03\x00\xBC\xB5"
-     "\x03\x00"
-     "\xBD\xBD\x03\x00\xBE\xB5\x03\x00\xBF\x55\x02\x00\xB0\xC9\x03\x00\xB1\xD5"
-     "\x03\x00"
-     "\xB2\xD5\x03\x00\xB3\xAD\x03\x00\xB4\xB5\x03\x00\xB5\xBD\x03\x00\xB6\xB5"
-     "\x03\x00"
-     "\xB7\xAD\x03\x00\x67\x58\x00\x80\x6B\x58\x00\x80\xA3\xD1\x03\x00\x6F\x58"
-     "\x00\x80"
-     "\xA5\xF5\x03\x00\x73\x58\x00\x80\x77\x58\x00\x80\xA6\x1D\x03\x00\x7B\x58"
-     "\x00\x80"
-     "\x7F\x58\x00\x80\xAB\x19\x03\x00\xAA\x05\x03\x00\xAD\x71\x03\x00\xAC\x09"
-     "\x03\x00"
-     "\xAF\x71\x03\x00\xAE\x79\x03\x00\xE1\x28\x07\x00\x83\x58\x00\x80\xE3\xE4"
-     "\x06\x00"
-     "\x87\x58\x00\x80\xE1\x28\x06\x00\x8B\x58\x00\x80\xE3\x68\x01\x00\x8F\x58"
-     "\x00\x80"
-     "\x93\x58\x00\x80\x97\x58\x00\x80\xEF\x58\x00\x00\x9B\x58\x00\x80\x9F\x58"
-     "\x00\x80"
-     "\xA3\x58\x00\x80\xEF\xC8\x06\x00\xA7\x58\x00\x80\xA8\x4D\xFD\x00\xA9\x81"
-     "\xFD\x00"
-     "\xAA\xD1\xFD\x00\xAB\xD1\xFD\x00\xAC\xF1\xFD\x00\xAD\xF1\xFD\x00\xAE\xF1"
-     "\xFD\x00"
-     "\xAF\xF1\xFD\x00\x43\x58\x00\x80\x82\x11\x00\x00\x81\x19\x00\x00\x80\xD1"
-     "\xFF\x00"
-     "\xAB\x58\x00\x80\xAF\x58\x00\x80\x84\x74\x03\x00\xBE\x74\x03\x00\xB8\x75"
-     "\xFE\x00"
-     "\xB9\x7D\xFE\x00\xBA\x75\xFE\x00\xBB\xC5\xFE\x00\xBC\xDD\xFE\x00\xBD\xC5"
-     "\xFE\x00"
-     "\xBE\xC5\xFE\x00\xBF\xF5\xFE\x00\xB0\x91\xFD\x00\xB1\x91\xFD\x00\xB2\x91"
-     "\xFD\x00"
-     "\xB3\x91\xFD\x00\xB4\x55\xFE\x00\xB5\x5D\xFE\x00\xB6\x55\xFE\x00\xB7\x4D"
-     "\xFE\x00"
-     "\xB3\x59\xFD\x00\xB3\x58\x00\x80\x86\xAC\x00\x00\x87\x44\x03\x00\xB7\x58"
-     "\x00\x80"
-     "\xB6\x71\xFD\x00\xB5\x79\xFD\x00\xBB\x58\x00\x80\xBB\x55\xFD\x00\xBA\x55"
-     "\xFD\x00"
-     "\xBF\x58\x00\x80\xC3\x58\x00\x80\xBF\x99\xFE\x00\xBE\x91\xFE\x00\xBD\x45"
-     "\xFD\x00"
-     "\xBC\x45\xFD\x00\xC7\x58\x00\x80\xA3\x1D\xFD\x00\xCB\x58\x00\x80\xCF\x58"
-     "\x00\x80"
-     "\xA6\x35\xFD\x00\xD3\x58\x00\x80\xD7\x58\x00\x80\xA5\x3D\xFD\x00\xAA\x11"
-     "\xFD\x00"
-     "\xAB\x11\xFD\x00\xDB\x58\x00\x80\xDF\x58\x00\x80\xAE\xD5\xFE\x00\xAF\xDD"
-     "\xFE\x00"
-     "\xAC\x01\xFD\x00\xAD\x01\xFD\x00\xA8\xCD\xFF\x00\xA9\xD1\xFF\x00\xAA\xD1"
-     "\xFF\x00"
-     "\xAB\x21\xFE\x00\xAC\x21\xFE\x00\xAD\x21\xFE\x00\xAE\x21\xFE\x00\xAF\x21"
-     "\xFE\x00"
-     "\xE3\x58\x00\x80\xE7\x58\x00\x80\xEB\x58\x00\x80\xEF\x58\x00\x80\xF3\x58"
-     "\x00\x80"
-     "\xF7\x58\x00\x80\xFB\x58\x00\x80\xFF\x58\x00\x80\xB8\x8D\xFE\x00\xB9\x91"
-     "\xFE\x00"
-     "\xBA\x95\xFE\x00\xBB\xAD\xFE\x00\xBC\xB5\xFE\x00\xBD\xB9\xFE\x00\xBE\xA9"
-     "\xFE\x00"
-     "\xBF\xA9\xFE\x00\xB0\xE1\xFE\x00\xB1\xE1\xFE\x00\xB2\xE5\xFE\x00\xB3\xF9"
-     "\xFE\x00"
-     "\xB4\xE9\xFE\x00\xB5\xD5\xFE\x00\xB6\xDD\xFE\x00\xB7\xB9\xFE\x00\xB3\x59"
-     "\xFF\x00"
-     "\x03\x59\x00\x80\xD3\x57\x00\x80\x07\x59\x00\x80\x0B\x59\x00\x80\xB6\xA1"
-     "\xFE\x00"
-     "\xB5\xA9\xFE\x00\x0F\x59\x00\x80\xBB\x25\xFE\x00\xBA\x25\xFE\x00\x13\x59"
-     "\x00\x80"
-     "\x17\x59\x00\x80\xBF\x11\xFE\x00\xBE\x2D\xFE\x00\xBD\x2D\xFE\x00\xBC\x31"
-     "\xFE\x00"
-     "\x82\x19\x00\x00\xA3\x1D\xFF\x00\x80\x65\x00\x00\x81\x19\x00\x00\xA6\xE5"
-     "\xFE\x00"
-     "\x1B\x59\x00\x80\x1F\x59\x00\x80\xA5\xED\xFE\x00\xAA\x61\xFE\x00\xAB\x61"
-     "\xFE\x00"
-     "\x84\x64\x01\x00\xBE\x20\x00\x00\xAE\x69\xFE\x00\xAF\x55\xFE\x00\xAC\x75"
-     "\xFE\x00"
-     "\xAD\x69\xFE\x00\x27\x59\x00\x80\xE3\x34\xFE\x00\x2B\x59\x00\x80\xE1\x7C"
-     "\xFD\x00"
-     "\x86\xB0\x04\x00\x87\x08\x03\x00\x2F\x59\x00\x80\x33\x59\x00\x80\x37\x59"
-     "\x00\x80"
-     "\x3B\x59\x00\x80\x84\x24\x03\x00\x84\x24\x04\x00\x3F\x59\x00\x80\xEF\x48"
-     "\xFE\x00"
-     "\x43\x59\x00\x80\x47\x59\x00\x80\xB3\xE9\x02\x00\x4B\x59\x00\x80\xBE\x44"
-     "\x04\x00"
-     "\xBE\x40\x05\x00\x4F\x59\x00\x80\xB6\x9D\x02\x00\xB5\x99\x02\x00\x53\x59"
-     "\x00\x80"
-     "\xBB\x89\x02\x00\xBA\xBD\x02\x00\x57\x59\x00\x80\x5B\x59\x00\x80\xBF\x59"
-     "\x03\x00"
-     "\xBE\x51\x03\x00\xBD\x59\x03\x00\xBC\x91\x02\x00\xA9\x1D\x02\x00\xA8\x15"
-     "\x02\x00"
-     "\xAB\x25\x02\x00\xAA\x25\x02\x00\xAD\x59\x02\x00\xAC\x55\x02\x00\xAF\x4D"
-     "\x02\x00"
-     "\xAE\x51\x02\x00\xBE\x64\x06\x00\x5F\x59\x00\x80\x63\x59\x00\x80\x67\x59"
-     "\x00\x80"
-     "\x6B\x59\x00\x80\x6F\x59\x00\x80\x73\x59\x00\x80\x77\x59\x00\x80\xB9\xF9"
-     "\x03\x00"
-     "\xB8\xF1\x03\x00\xBB\x4D\x03\x00\xBA\xF1\x03\x00\xBD\x41\x03\x00\xBC\x59"
-     "\x03\x00"
-     "\xBF\x71\x03\x00\xBE\x41\x03\x00\xB1\x09\x02\x00\xB0\x3D\x02\x00\xB3\xC9"
-     "\x03\x00"
-     "\xB2\x01\x02\x00\xB5\xD9\x03\x00\xB4\xD1\x03\x00\xB7\xC9\x03\x00\xB6\xD1"
-     "\x03\x00"
-     "\xE1\x90\x01\x00\x7B\x59\x00\x80\xE3\xF0\x00\x00\x7F\x59\x00\x80\x83\x59"
-     "\x00\x80"
-     "\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D\x00\x00\x87\x59\x00\x80\x8B\x59"
-     "\x00\x80"
-     "\x8F\x59\x00\x80\x97\x59\x00\x80\x9B\x59\x00\x80\x9F\x59\x00\x80\xEF\x2C"
-     "\x00\x00"
-     "\xA3\x59\x00\x80\xA3\xE9\x03\x00\xA7\x59\x00\x80\x86\xE8\x04\x00\x87\x60"
-     "\x05\x00"
-     "\xAB\x59\x00\x80\xA6\x9D\x03\x00\xA5\x99\x03\x00\xAF\x59\x00\x80\xAB\x89"
-     "\x03\x00"
-     "\xAA\xBD\x03\x00\xB3\x59\x00\x80\xB7\x59\x00\x80\xAF\x59\x02\x00\xAE\x51"
-     "\x02\x00"
-     "\xAD\x59\x02\x00\xAC\x91\x03\x00\xBB\x59\x00\x80\xBF\x59\x00\x80\xC3\x59"
-     "\x00\x80"
-     "\xC7\x59\x00\x80\xCB\x59\x00\x80\xCF\x59\x00\x80\xD3\x59\x00\x80\xEF\x58"
-     "\x01\x00"
-     "\x84\xA0\x04\x00\xE1\x54\xFF\x00\xD7\x59\x00\x80\xE3\x84\x01\x00\xDB\x59"
-     "\x00\x80"
-     "\xDF\x59\x00\x80\xE3\x59\x00\x80\xE7\x59\x00\x80\xB3\xD9\x01\x00\xEB\x59"
-     "\x00\x80"
-     "\x85\xCC\x19\x00\xEF\x59\x00\x80\xF3\x59\x00\x80\xB6\xF1\x01\x00\xB5\xF9"
-     "\x01\x00"
-     "\xF7\x59\x00\x80\xBB\xA5\x01\x00\xBA\xD9\x01\x00\xFB\x59\x00\x80\xFF\x59"
-     "\x00\x80"
-     "\xBF\x9D\x01\x00\xBE\x9D\x01\x00\xBD\xA5\x01\x00\xBC\xAD\x01\x00\xA8\x01"
-     "\x06\x00"
-     "\xA9\x0D\x06\x00\xAA\x11\x06\x00\xAB\x11\x06\x00\xAC\x31\x06\x00\xAD\x31"
-     "\x06\x00"
-     "\xAE\x29\x06\x00\xAF\x25\x06\x00\x93\x59\x00\x80\x82\xC9\x07\x00\x81\xC1"
-     "\x07\x00"
-     "\x80\xF1\x07\x00\x03\x5A\x00\x80\x07\x5A\x00\x80\x86\x1C\x00\x00\x87\xFC"
-     "\x03\x00"
-     "\xB8\xED\x06\x00\xB9\xF5\x06\x00\xBA\xFD\x06\x00\xBB\xF5\x06\x00\xBC\xED"
-     "\x06\x00"
-     "\xBD\x51\x07\x00\xBE\x55\x07\x00\xBF\x4D\x07\x00\xB0\x5D\x06\x00\xB1\x21"
-     "\x06\x00"
-     "\xB2\x39\x06\x00\xB3\x31\x06\x00\xB4\x19\x06\x00\xB5\x19\x06\x00\xB6\xDD"
-     "\x06\x00"
-     "\xB7\xD5\x06\x00\xA3\x99\x06\x00\x0B\x5A\x00\x80\x0F\x5A\x00\x80\x13\x5A"
-     "\x00\x80"
-     "\x17\x5A\x00\x80\xA6\xB1\x06\x00\xA5\xB9\x06\x00\x1B\x5A\x00\x80\xAB\xE5"
-     "\x06\x00"
-     "\xAA\x99\x06\x00\x1F\x5A\x00\x80\x23\x5A\x00\x80\xAF\xDD\x06\x00\xAE\xDD"
-     "\x06\x00"
-     "\xAD\xE5\x06\x00\xAC\xED\x06\x00\x27\x5A\x00\x80\xB3\xF1\x07\x00\x2B\x5A"
-     "\x00\x80"
-     "\x2F\x5A\x00\x80\xB6\x81\x07\x00\x33\x5A\x00\x80\x37\x5A\x00\x80\xB5\x99"
-     "\x07\x00"
-     "\xBA\x8D\x07\x00\xBB\x65\x07\x00\x3B\x5A\x00\x80\x3F\x5A\x00\x80\xBE\x7D"
-     "\x07\x00"
-     "\xBF\x65\x07\x00\xBC\x7D\x07\x00\xBD\x75\x07\x00\xA8\x2D\x06\x00\xA9\x35"
-     "\x06\x00"
-     "\xAA\x3D\x06\x00\xAB\x31\x06\x00\xAC\x55\x06\x00\xAD\x45\x06\x00\xAE\x45"
-     "\x06\x00"
-     "\xAF\x75\x06\x00\x43\x5A\x00\x80\x47\x5A\x00\x80\x4B\x5A\x00\x80\x4F\x5A"
-     "\x00\x80"
-     "\x53\x5A\x00\x80\x57\x5A\x00\x80\x5B\x5A\x00\x80\x5F\x5A\x00\x80\xB8\xE9"
-     "\x06\x00"
-     "\xB9\xF5\x06\x00\xBA\xFD\x06\x00\xBB\xF5\x06\x00\xBC\xED\x06\x00\xBD\x91"
-     "\x06\x00"
-     "\xBE\x95\x06\x00\xBF\x8D\x06\x00\xB0\x0D\x06\x00\xB1\xE5\x06\x00\xB2\xED"
-     "\x06\x00"
-     "\xB3\xE5\x06\x00\xB4\xFD\x06\x00\xB5\xE5\x06\x00\xB6\xED\x06\x00\xB7\xE5"
-     "\x06\x00"
-     "\xA3\xB5\x06\x00\x63\x5A\x00\x80\x67\x5A\x00\x80\x6B\x5A\x00\x80\x6F\x5A"
-     "\x00\x80"
-     "\xA6\xC5\x06\x00\xA5\xDD\x06\x00\x23\x59\x00\x80\xAB\x21\x06\x00\xAA\xC9"
-     "\x06\x00"
-     "\x73\x5A\x00\x80\x77\x5A\x00\x80\xAF\x21\x06\x00\xAE\x39\x06\x00\xAD\x31"
-     "\x06\x00"
-     "\xAC\x39\x06\x00\x80\x49\x00\x00\x81\x49\x00\x00\x82\x59\x00\x00\xB3\x45"
-     "\x01\x00"
-     "\x7B\x5A\x00\x80\xB5\x45\x01\x00\xB6\x45\x01\x00\x7F\x5A\x00\x80\x86\x40"
-     "\x00\x00"
-     "\x87\x64\x00\x00\xBA\x29\x01\x00\xBB\x25\x01\x00\xBC\x3D\x01\x00\xBD\x21"
-     "\x01\x00"
-     "\xBE\x21\x01\x00\xBF\x15\x01\x00\x83\x5A\x00\x80\x87\x5A\x00\x80\x84\x04"
-     "\x03\x00"
-     "\xBE\x00\x0C\x00\xE3\x28\x06\x00\xEF\xE0\x02\x00\xE1\x10\x06\x00\x8B\x5A"
-     "\x00\x80"
-     "\xEF\x90\x06\x00\xE3\x3C\x02\x00\x8F\x5A\x00\x80\xE1\xD4\x01\x00\x93\x5A"
-     "\x00\x80"
-     "\x97\x5A\x00\x80\x9B\x5A\x00\x80\x9F\x5A\x00\x80\xA3\x5A\x00\x80\xA7\x5A"
-     "\x00\x80"
-     "\xA3\xC1\x02\x00\xAB\x5A\x00\x80\xA5\xC1\x02\x00\xAF\x5A\x00\x80\xB3\x5A"
-     "\x00\x80"
-     "\xA6\xC1\x02\x00\xB7\x5A\x00\x80\xBB\x5A\x00\x80\xAB\xA1\x02\x00\xAA\xAD"
-     "\x02\x00"
-     "\xAD\xA5\x02\x00\xAC\xB9\x02\x00\xAF\x91\x02\x00\xAE\xA5\x02\x00\xA9\x05"
-     "\x02\x00"
-     "\xA8\xB1\x02\x00\xAB\x05\x02\x00\xAA\x05\x02\x00\xAD\x0D\x02\x00\xAC\x05"
-     "\x02\x00"
-     "\xAF\x35\x02\x00\xAE\x35\x02\x00\x84\xA8\x0C\x00\xBF\x5A\x00\x80\xC3\x5A"
-     "\x00\x80"
-     "\xC7\x5A\x00\x80\xCB\x5A\x00\x80\xCF\x5A\x00\x80\xD3\x5A\x00\x80\xD7\x5A"
-     "\x00\x80"
-     "\xB9\xE9\x03\x00\xB8\xE1\x03\x00\xBB\xF9\x03\x00\xBA\xE1\x03\x00\xBD\xE9"
-     "\x03\x00"
-     "\xBC\xE1\x03\x00\xBF\x5D\x03\x00\xBE\xE1\x03\x00\xB1\x29\x02\x00\xB0\x25"
-     "\x02\x00"
-     "\xB3\x3D\x02\x00\xB2\x21\x02\x00\xB5\x19\x02\x00\xB4\x2D\x02\x00\xB7\xD9"
-     "\x03\x00"
-     "\xB6\x11\x02\x00\xA8\xAD\x02\x00\xA9\xD5\x02\x00\xAA\xD5\x02\x00\xAB\x0D"
-     "\x01\x00"
-     "\xAC\x15\x01\x00\xAD\x09\x01\x00\xAE\x31\x01\x00\xAF\x2D\x01\x00\xDB\x5A"
-     "\x00\x80"
-     "\xDF\x5A\x00\x80\xE7\x5A\x00\x80\xEB\x5A\x00\x80\xEF\x5A\x00\x80\xF3\x5A"
-     "\x00\x80"
-     "\xF7\x5A\x00\x80\xFB\x5A\x00\x80\xB8\x21\x01\x00\xB9\x21\x01\x00\xBA\xED"
-     "\x01\x00"
-     "\xBB\xE5\x01\x00\xBC\xFD\x01\x00\xBD\xE5\x01\x00\xBE\xED\x01\x00\xBF\xE5"
-     "\x01\x00"
-     "\xB0\x55\x01\x00\xB1\x5D\x01\x00\xB2\x55\x01\x00\xB3\x2D\x01\x00\xB4\x35"
-     "\x01\x00"
-     "\xB5\x39\x01\x00\xB6\x2D\x01\x00\xB7\x25\x01\x00\x80\x3D\x01\x00\x81\xA5"
-     "\x00\x00"
-     "\x82\xAD\x00\x00\xEF\xD4\x07\x00\xFF\x5A\x00\x80\x03\x5B\x00\x80\x07\x5B"
-     "\x00\x80"
-     "\xEF\x28\x07\x00\xBE\x2C\x0C\x00\xE1\xF4\x06\x00\x0B\x5B\x00\x80\xE3\x90"
-     "\x07\x00"
-     "\x0F\x5B\x00\x80\xE1\x94\x01\x00\x13\x5B\x00\x80\xE3\x0C\x06\x00\xB3\x1D"
-     "\x02\x00"
-     "\x17\x5B\x00\x80\x87\x44\x0D\x00\x86\x4C\x0D\x00\x1B\x5B\x00\x80\xB6\xC9"
-     "\x01\x00"
-     "\xB5\xDD\x01\x00\x1F\x5B\x00\x80\xBB\xD9\x01\x00\xBA\xD1\x01\x00\x23\x5B"
-     "\x00\x80"
-     "\x27\x5B\x00\x80\xBF\xBD\x01\x00\xBE\xB1\x01\x00\xBD\xB9\x01\x00\xBC\xC1"
-     "\x01\x00"
-     "\xE3\x5A\x00\x80\x2B\x5B\x00\x80\x2F\x5B\x00\x80\x33\x5B\x00\x80\x37\x5B"
-     "\x00\x80"
-     "\x3B\x5B\x00\x80\x3F\x5B\x00\x80\x43\x5B\x00\x80\xA8\x09\x0F\x00\xA9\x09"
-     "\x0F\x00"
-     "\xAA\x19\x0F\x00\xAB\x19\x0F\x00\xAC\x09\x0F\x00\xAD\x09\x0F\x00\xAE\xA9"
-     "\x0F\x00"
-     "\xAF\xA9\x0F\x00\xB0\xD9\x0F\x00\xB1\xED\x0F\x00\xB2\xF9\x0F\x00\xB3\xF5"
-     "\x0F\x00"
-     "\xB4\x95\x0F\x00\xB5\x85\x0F\x00\xB6\x85\x0F\x00\xB7\xB5\x0F\x00\xB8\x8D"
-     "\x0F\x00"
-     "\xB9\x61\x00\x00\xBA\x61\x00\x00\xBB\x61\x00\x00\xBC\x61\x00\x00\xBD\x61"
-     "\x00\x00"
-     "\xBE\x61\x00\x00\xBF\x61\x00\x00\xA3\x5D\x0D\x00\x82\x2D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x47\x5B\x00\x80\xA6\x89\x0E\x00\xA5\x9D\x0E\x00\x4B\x5B"
-     "\x00\x80"
-     "\xAB\x99\x0E\x00\xAA\x91\x0E\x00\x4F\x5B\x00\x80\x53\x5B\x00\x80\xAF\xFD"
-     "\x0E\x00"
-     "\xAE\xF1\x0E\x00\xAD\xF9\x0E\x00\xAC\x81\x0E\x00\x57\x5B\x00\x80\xB3\xF5"
-     "\x0F\x00"
-     "\x86\xE8\x03\x00\x87\xBC\x03\x00\xB6\xED\x0F\x00\x5B\x5B\x00\x80\x5F\x5B"
-     "\x00\x80"
-     "\xB5\xE5\x0F\x00\xBA\x79\x0F\x00\xBB\x4D\x0F\x00\x63\x5B\x00\x80\x67\x5B"
-     "\x00\x80"
-     "\xBE\x35\x0F\x00\xBF\x25\x0F\x00\xBC\x49\x0F\x00\xBD\x45\x0F\x00\xA3\x31"
-     "\x0E\x00"
-     "\x6B\x5B\x00\x80\x6F\x5B\x00\x80\x73\x5B\x00\x80\x77\x5B\x00\x80\xA6\x29"
-     "\x0E\x00"
-     "\xA5\x21\x0E\x00\x7B\x5B\x00\x80\xAB\x89\x0E\x00\xAA\xBD\x0E\x00\x7F\x5B"
-     "\x00\x80"
-     "\x83\x5B\x00\x80\xAF\xE1\x0E\x00\xAE\xF1\x0E\x00\xAD\x81\x0E\x00\xAC\x8D"
-     "\x0E\x00"
-     "\x87\x5B\x00\x80\x8B\x5B\x00\x80\x8F\x5B\x00\x80\x93\x5B\x00\x80\x97\x5B"
-     "\x00\x80"
-     "\x9B\x5B\x00\x80\x9F\x5B\x00\x80\xA3\x5B\x00\x80\xA7\x5B\x00\x80\xAB\x5B"
-     "\x00\x80"
-     "\xAF\x5B\x00\x80\xB3\x5B\x00\x80\x80\x0D\x00\x00\x81\xD5\x00\x00\x82\xDD"
-     "\x00\x00"
-     "\xB7\x5B\x00\x80\xA8\x41\x01\x00\xA9\x51\x01\x00\xAA\x51\x01\x00\xAB\x65"
-     "\x01\x00"
-     "\xAC\x7D\x01\x00\xAD\x91\x00\x00\xAE\x91\x00\x00\xAF\x91\x00\x00\xBB\x5B"
-     "\x00\x80"
-     "\xBF\x5B\x00\x80\x84\x64\x01\x00\xBE\x64\x01\x00\x86\x90\x01\x00\x87\x84"
-     "\x00\x00"
-     "\xC7\x5B\x00\x80\xCB\x5B\x00\x80\xB8\x91\x00\x00\xB9\x91\x00\x00\xBA\x91"
-     "\x00\x00"
-     "\xBB\x91\x00\x00\xBC\xB1\x00\x00\xBD\xB1\x00\x00\xBE\xB1\x00\x00\xBF\xB1"
-     "\x00\x00"
-     "\xB0\xF1\x00\x00\xB1\xF9\x00\x00\xB2\xC1\x00\x00\xB3\xC1\x00\x00\xB4\xB1"
-     "\x00\x00"
-     "\xB5\xB1\x00\x00\xB6\xB1\x00\x00\xB7\xB1\x00\x00\xB3\xD9\x02\x00\xCF\x5B"
-     "\x00\x80"
-     "\xBE\x70\x03\x00\xBE\x44\x04\x00\xD3\x5B\x00\x80\xB6\x11\x03\x00\xB5\xF5"
-     "\x02\x00"
-     "\xD7\x5B\x00\x80\xBB\x3D\x03\x00\xBA\x35\x03\x00\xDB\x5B\x00\x80\xDF\x5B"
-     "\x00\x80"
-     "\xBF\x75\x03\x00\xBE\x75\x03\x00\xBD\x15\x03\x00\xBC\x15\x03\x00\xE3\x5B"
-     "\x00\x80"
-     "\xA3\x9D\x02\x00\xE7\x5B\x00\x80\xEB\x5B\x00\x80\xA6\x55\x03\x00\xEF\x5B"
-     "\x00\x80"
-     "\xF3\x5B\x00\x80\xA5\xB1\x02\x00\xAA\x71\x03\x00\xAB\x79\x03\x00\xF7\x5B"
-     "\x00\x80"
-     "\xFB\x5B\x00\x80\xAE\x31\x03\x00\xAF\x31\x03\x00\xAC\x51\x03\x00\xAD\x51"
-     "\x03\x00"
-     "\xA8\x39\x03\x00\xA9\x39\x03\x00\xAA\x8D\x00\x00\xAB\x9D\x00\x00\xAC\x8D"
-     "\x00\x00"
-     "\xAD\xBD\x00\x00\xAE\xB5\x00\x00\xAF\xDD\x00\x00\xFF\x5B\x00\x80\x03\x5C"
-     "\x00\x80"
-     "\x07\x5C\x00\x80\x0B\x5C\x00\x80\x0F\x5C\x00\x80\x13\x5C\x00\x80\x17\x5C"
-     "\x00\x80"
-     "\x1B\x5C\x00\x80\xB8\x69\x00\x00\xB9\x69\x00\x00\xBA\x79\x00\x00\xBB\x79"
-     "\x00\x00"
-     "\xBC\x69\x00\x00\xBD\x69\x00\x00\xBE\xD9\x01\x00\xBF\xD9\x01\x00\xB0\xA9"
-     "\x00\x00"
-     "\xB1\xA9\x00\x00\xB2\xBD\x00\x00\xB3\xB5\x00\x00\xB4\x99\x00\x00\xB5\x99"
-     "\x00\x00"
-     "\xB6\x59\x00\x00\xB7\x59\x00\x00\x1F\x5C\x00\x80\x23\x5C\x00\x80\x27\x5C"
-     "\x00\x80"
-     "\x2B\x5C\x00\x80\xEF\x44\x00\x00\x2F\x5C\x00\x80\x86\x98\x05\x00\x87\xE4"
-     "\x02\x00"
-     "\xE3\x98\x00\x00\x84\xA8\x02\x00\xE1\xF8\x01\x00\x37\x5C\x00\x80\x80\x39"
-     "\x00\x00"
-     "\x81\x39\x00\x00\x82\x2D\x00\x00\x3B\x5C\x00\x80\xB3\x45\x01\x00\x3F\x5C"
-     "\x00\x80"
-     "\x43\x5C\x00\x80\x47\x5C\x00\x80\x4B\x5C\x00\x80\xB6\x7D\x01\x00\xB5\x45"
-     "\x01\x00"
-     "\x4F\x5C\x00\x80\xBB\x2D\x01\x00\xBA\x2D\x01\x00\x53\x5C\x00\x80\x57\x5C"
-     "\x00\x80"
-     "\xBF\x1D\x01\x00\xBE\x1D\x01\x00\xBD\x21\x01\x00\xBC\x29\x01\x00\x5B\x5C"
-     "\x00\x80"
-     "\xE1\x50\x0E\x00\x5F\x5C\x00\x80\xE3\x3C\x0F\x00\x63\x5C\x00\x80\x67\x5C"
-     "\x00\x80"
-     "\x6B\x5C\x00\x80\x6F\x5C\x00\x80\x73\x5C\x00\x80\x77\x5C\x00\x80\xE3\x00"
-     "\x00\x00"
-     "\x7B\x5C\x00\x80\x7F\x5C\x00\x80\x83\x5C\x00\x80\x84\xF4\x05\x00\xEF\xE4"
-     "\x0E\x00"
-     "\xAE\xA9\x01\x00\xAF\xA9\x01\x00\xAC\x9D\x01\x00\xAD\x95\x01\x00\xAA\x99"
-     "\x01\x00"
-     "\xAB\x99\x01\x00\x87\x5C\x00\x80\x8B\x5C\x00\x80\xA6\xC9\x01\x00\x8F\x5C"
-     "\x00\x80"
-     "\x93\x5C\x00\x80\xA5\xF1\x01\x00\x82\xFD\x07\x00\xA3\xF1\x01\x00\x80\xFD"
-     "\x07\x00"
-     "\x81\xF5\x07\x00\x33\x5C\x00\x80\x97\x5C\x00\x80\x9B\x5C\x00\x80\x9F\x5C"
-     "\x00\x80"
-     "\xA3\x5C\x00\x80\xA7\x5C\x00\x80\x86\xB8\x03\x00\x87\x90\x03\x00\xA8\x0D"
-     "\x07\x00"
-     "\xA9\x19\x07\x00\xAA\x6D\x07\x00\xAB\x65\x07\x00\xAC\x7D\x07\x00\xAD\x65"
-     "\x07\x00"
-     "\xAE\x65\x07\x00\xAF\x55\x07\x00\xB0\x2D\x07\x00\xB1\xC5\x07\x00\xB2\xC1"
-     "\x07\x00"
-     "\xB3\xDD\x07\x00\xB4\xC5\x07\x00\xB5\xCD\x07\x00\xB6\xC5\x07\x00\xB7\xFD"
-     "\x07\x00"
-     "\xB8\xC5\x07\x00\xB9\xC9\x07\x00\xBA\xD9\x07\x00\xBB\xD9\x07\x00\xBC\xA9"
-     "\x07\x00"
-     "\xBD\xA9\x07\x00\xBE\x9D\x07\x00\xBF\x95\x07\x00\xB3\xC5\x07\x00\xAB\x5C"
-     "\x00\x80"
-     "\xAF\x5C\x00\x80\xB3\x5C\x00\x80\xB7\x5C\x00\x80\xB6\xC5\x07\x00\xB5\xD5"
-     "\x07\x00"
-     "\xBB\x5C\x00\x80\xBB\x21\x07\x00\xBA\xC9\x07\x00\xBF\x5C\x00\x80\xC3\x5C"
-     "\x00\x80"
-     "\xBF\x29\x07\x00\xBE\x21\x07\x00\xBD\x29\x07\x00\xBC\x35\x07\x00\xC7\x5C"
-     "\x00\x80"
-     "\xA3\x81\x07\x00\xCB\x5C\x00\x80\xCF\x5C\x00\x80\xA6\x81\x07\x00\xD3\x5C"
-     "\x00\x80"
-     "\xD7\x5C\x00\x80\xA5\x91\x07\x00\xAA\x8D\x07\x00\xAB\x65\x07\x00\xDB\x5C"
-     "\x00\x80"
-     "\xDF\x5C\x00\x80\xAE\x65\x07\x00\xAF\x6D\x07\x00\xAC\x71\x07\x00\xAD\x6D"
-     "\x07\x00"
-     "\xA8\x15\x01\x00\xA9\x81\x01\x00\xAA\x81\x01\x00\xAB\x81\x01\x00\xAC\x81"
-     "\x01\x00"
-     "\xAD\x89\x01\x00\xAE\xB1\x01\x00\xAF\xB1\x01\x00\xE3\x5C\x00\x80\xE7\x5C"
-     "\x00\x80"
-     "\xEB\x5C\x00\x80\xEF\x5C\x00\x80\xF3\x5C\x00\x80\xF7\x5C\x00\x80\xFB\x5C"
-     "\x00\x80"
-     "\xFF\x5C\x00\x80\xB8\x65\x00\x00\xB9\x6D\x00\x00\xBA\x65\x00\x00\xBB\x7D"
-     "\x00\x00"
-     "\xBC\x65\x00\x00\xBD\x6D\x00\x00\xBE\x65\x00\x00\xBF\xDD\x00\x00\xB0\xA1"
-     "\x01\x00"
-     "\xB1\xAD\x01\x00\xB2\xA5\x01\x00\xB3\xB9\x01\x00\xB4\xA9\x01\x00\xB5\x9D"
-     "\x01\x00"
-     "\xB6\x95\x01\x00\xB7\x5D\x00\x00\x03\x5D\x00\x80\x82\x1D\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x07\x5D\x00\x80\x0B\x5D\x00\x80\x0F\x5D\x00\x80\xBE\x58"
-     "\x01\x00"
-     "\x84\xAC\x02\x00\x13\x5D\x00\x80\x87\x08\x01\x00\x86\x8C\x01\x00\x17\x5D"
-     "\x00\x80"
-     "\xC3\x5B\x00\x80\x1B\x5D\x00\x80\x1F\x5D\x00\x80\xB3\x49\x01\x00\x23\x5D"
-     "\x00\x80"
-     "\x27\x5D\x00\x80\x2B\x5D\x00\x80\x2F\x5D\x00\x80\xB6\x49\x01\x00\xB5\x49"
-     "\x01\x00"
-     "\x33\x5D\x00\x80\xBB\x29\x01\x00\xBA\x25\x01\x00\x37\x5D\x00\x80\x3B\x5D"
-     "\x00\x80"
-     "\xBF\x19\x01\x00\xBE\x2D\x01\x00\xBD\x2D\x01\x00\xBC\x31\x01\x00\xBE\xD8"
-     "\x03\x00"
-     "\x3F\x5D\x00\x80\xEF\xF8\x06\x00\x43\x5D\x00\x80\x47\x5D\x00\x80\x4B\x5D"
-     "\x00\x80"
-     "\xEF\xE0\x02\x00\x4F\x5D\x00\x80\xE1\x94\x01\x00\x53\x5D\x00\x80\xE3\x74"
-     "\x02\x00"
-     "\x57\x5D\x00\x80\xE1\x98\x05\x00\x5B\x5D\x00\x80\xE3\x0C\x07\x00\x5F\x5D"
-     "\x00\x80"
-     "\x63\x5D\x00\x80\x67\x5D\x00\x80\xA3\xC1\x02\x00\x84\x8C\x03\x00\xA5\xC1"
-     "\x02\x00"
-     "\x6B\x5D\x00\x80\x6F\x5D\x00\x80\xA6\xC1\x02\x00\x73\x5D\x00\x80\x77\x5D"
-     "\x00\x80"
-     "\xAB\xA1\x02\x00\xAA\xAD\x02\x00\xAD\xA5\x02\x00\xAC\xB9\x02\x00\xAF\x91"
-     "\x02\x00"
-     "\xAE\xA5\x02\x00\xA8\x31\x03\x00\xA9\x3D\x03\x00\xAA\x35\x03\x00\xAB\x49"
-     "\x03\x00"
-     "\xAC\x59\x03\x00\xAD\x59\x03\x00\xAE\x49\x03\x00\xAF\x41\x03\x00\x80\xC5"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x19\x00\x00\x7B\x5D\x00\x80\x7F\x5D\x00\x80\x87\x5D"
-     "\x00\x80"
-     "\x87\x6C\x03\x00\x86\x1C\x1C\x00\xB8\xED\x00\x00\xB9\xF1\x00\x00\xBA\xF5"
-     "\x00\x00"
-     "\xBB\x8D\x00\x00\xBC\x95\x00\x00\xBD\x81\x00\x00\xBE\x81\x00\x00\xBF\xBD"
-     "\x00\x00"
-     "\xB0\x09\x03\x00\xB1\x09\x03\x00\xB2\xED\x00\x00\xB3\xE1\x00\x00\xB4\xE1"
-     "\x00\x00"
-     "\xB5\xE1\x00\x00\xB6\xE5\x00\x00\xB7\xD9\x00\x00\x8B\x5D\x00\x80\x8F\x5D"
-     "\x00\x80"
-     "\x93\x5D\x00\x80\xB3\xB1\x02\x00\x97\x5D\x00\x80\xB5\xDD\x02\x00\xB6\xD5"
-     "\x02\x00"
-     "\x9B\x5D\x00\x80\x9F\x5D\x00\x80\xA3\x5D\x00\x80\xBA\xC1\x02\x00\xBB\xC1"
-     "\x02\x00"
-     "\xBC\x35\x01\x00\xBD\x39\x01\x00\xBE\x29\x01\x00\xBF\x29\x01\x00\xA6\x8D"
-     "\x02\x00"
-     "\xA7\x5D\x00\x80\xAB\x5D\x00\x80\xA5\x85\x02\x00\xAF\x5D\x00\x80\xA3\xE9"
-     "\x02\x00"
-     "\xB3\x5D\x00\x80\xB7\x5D\x00\x80\xAE\x71\x01\x00\xAF\x71\x01\x00\xAC\x6D"
-     "\x01\x00"
-     "\xAD\x61\x01\x00\xAA\x99\x02\x00\xAB\x99\x02\x00\xBB\x5D\x00\x80\xBF\x5D"
-     "\x00\x80"
-     "\xC3\x5D\x00\x80\xE1\x84\x06\x00\xC7\x5D\x00\x80\xE3\x24\x06\x00\xCB\x5D"
-     "\x00\x80"
-     "\xE1\x94\x01\x00\xCF\x5D\x00\x80\xE3\x2C\x00\x00\xBE\xE8\x1D\x00\xD3\x5D"
-     "\x00\x80"
-     "\xD7\x5D\x00\x80\xEF\xC8\x00\x00\x84\xFC\x1D\x00\xBE\xF0\x1C\x00\xDB\x5D"
-     "\x00\x80"
-     "\xEF\x48\x07\x00\xDF\x5D\x00\x80\xE3\x5D\x00\x80\xE7\x5D\x00\x80\xEB\x5D"
-     "\x00\x80"
-     "\x81\x1D\x00\x00\x80\x1D\x00\x00\xEF\x5D\x00\x80\x82\x05\x00\x00\x86\x40"
-     "\x1C\x00"
-     "\x87\xC4\x1C\x00\xF7\x5D\x00\x80\xFB\x5D\x00\x80\xFF\x5D\x00\x80\x03\x5E"
-     "\x00\x80"
-     "\x07\x5E\x00\x80\x0B\x5E\x00\x80\xA8\xB5\x1E\x00\xA9\x05\x1F\x00\xAA\x0D"
-     "\x1F\x00"
-     "\xAB\x05\x1F\x00\xAC\x01\x1F\x00\xAD\x09\x1F\x00\xAE\x39\x1F\x00\xAF\x39"
-     "\x1F\x00"
-     "\xF3\x5D\x00\x80\x0F\x5E\x00\x80\x13\x5E\x00\x80\x17\x5E\x00\x80\x1B\x5E"
-     "\x00\x80"
-     "\x1F\x5E\x00\x80\x23\x5E\x00\x80\x27\x5E\x00\x80\xB8\xC9\x1F\x00\xB9\xD5"
-     "\x1F\x00"
-     "\xBA\xD1\x1F\x00\xBB\xE9\x1F\x00\xBC\xF9\x1F\x00\xBD\xED\x1F\x00\xBE\x99"
-     "\x1F\x00"
-     "\xBF\x99\x1F\x00\xB0\x25\x1F\x00\xB1\x2D\x1F\x00\xB2\x39\x1F\x00\xB3\x35"
-     "\x1F\x00"
-     "\xB4\x2D\x1F\x00\xB5\x0D\x1F\x00\xB6\x05\x1F\x00\xB7\xFD\x1F\x00\xB3\x85"
-     "\x1F\x00"
-     "\x2B\x5E\x00\x80\x2F\x5E\x00\x80\x33\x5E\x00\x80\x37\x5E\x00\x80\xB6\x89"
-     "\x1F\x00"
-     "\xB5\x89\x1F\x00\x3B\x5E\x00\x80\xBB\xE9\x1F\x00\xBA\xE1\x1F\x00\x3F\x5E"
-     "\x00\x80"
-     "\x43\x5E\x00\x80\xBF\xC9\x1F\x00\xBE\xE9\x1F\x00\xBD\xE1\x1F\x00\xBC\xED"
-     "\x1F\x00"
-     "\x47\x5E\x00\x80\xA3\xC1\x1F\x00\x4B\x5E\x00\x80\x4F\x5E\x00\x80\xA6\xCD"
-     "\x1F\x00"
-     "\x53\x5E\x00\x80\x57\x5E\x00\x80\xA5\xCD\x1F\x00\xAA\xA5\x1F\x00\xAB\xAD"
-     "\x1F\x00"
-     "\x5B\x5E\x00\x80\x5F\x5E\x00\x80\xAE\xAD\x1F\x00\xAF\x8D\x1F\x00\xAC\xA9"
-     "\x1F\x00"
-     "\xAD\xA5\x1F\x00\xA8\xE9\x1E\x00\xA9\xE9\x1E\x00\xAA\xF9\x1E\x00\xAB\xF9"
-     "\x1E\x00"
-     "\xAC\xE9\x1E\x00\xAD\xE9\x1E\x00\xAE\x3D\x01\x00\xAF\x35\x01\x00\x80\xF9"
-     "\x01\x00"
-     "\x81\xCD\x01\x00\x82\xC5\x01\x00\x84\x60\x02\x00\x63\x5E\x00\x80\x67\x5E"
-     "\x00\x80"
-     "\x87\x68\x01\x00\x86\x9C\x00\x00\xB8\xD1\x01\x00\xB9\xD9\x01\x00\xBA\xE1"
-     "\x01\x00"
-     "\xBB\xE1\x01\x00\xBC\x91\x01\x00\xBD\x9D\x01\x00\xBE\x95\x01\x00\xBF\x89"
-     "\x01\x00"
-     "\xB0\x4D\x01\x00\xB1\x55\x01\x00\xB2\x5D\x01\x00\xB3\x55\x01\x00\xB4\x4D"
-     "\x01\x00"
-     "\xB5\xF1\x01\x00\xB6\xF1\x01\x00\xB7\xF1\x01\x00\xB3\x71\x1E\x00\x6B\x5E"
-     "\x00\x80"
-     "\x6F\x5E\x00\x80\x73\x5E\x00\x80\x77\x5E\x00\x80\xB6\x69\x1E\x00\xB5\x61"
-     "\x1E\x00"
-     "\x7B\x5E\x00\x80\xBB\x91\x01\x00\xBA\x89\x01\x00\x7F\x5E\x00\x80\x83\x5E"
-     "\x00\x80"
-     "\xBF\x35\x01\x00\xBE\x89\x01\x00\xBD\x81\x01\x00\xBC\x89\x01\x00\x87\x5E"
-     "\x00\x80"
-     "\x83\x5D\x00\x80\xA3\x39\x1E\x00\x8B\x5E\x00\x80\xA5\x29\x1E\x00\x8F\x5E"
-     "\x00\x80"
-     "\x93\x5E\x00\x80\xA6\x21\x1E\x00\x97\x5E\x00\x80\x9B\x5E\x00\x80\xAB\xD9"
-     "\x01\x00"
-     "\xAA\xC1\x01\x00\xAD\xC9\x01\x00\xAC\xC1\x01\x00\xAF\x7D\x01\x00\xAE\xC1"
-     "\x01\x00"
-     "\x9F\x5E\x00\x80\xA3\x5E\x00\x80\xA7\x5E\x00\x80\xAB\x5E\x00\x80\xAF\x5E"
-     "\x00\x80"
-     "\xB3\x5E\x00\x80\xB7\x5E\x00\x80\xBB\x5E\x00\x80\xBF\x5E\x00\x80\xC3\x5E"
-     "\x00\x80"
-     "\xC7\x5E\x00\x80\xCB\x5E\x00\x80\xCF\x5E\x00\x80\xD7\x5E\x00\x80\xDB\x5E"
-     "\x00\x80"
-     "\xBE\x70\x03\x00\xE1\xE4\x1E\x00\x84\x48\x02\x00\xE3\xE8\x1F\x00\x84\x00"
-     "\x04\x00"
-     "\x80\x79\x00\x00\x81\x79\x00\x00\x82\x69\x00\x00\xDF\x5E\x00\x80\x86\xC0"
-     "\x04\x00"
-     "\x87\x44\x03\x00\xE3\x5E\x00\x80\xE7\x5E\x00\x80\xEB\x5E\x00\x80\xEF\x5E"
-     "\x00\x80"
-     "\xEF\x20\x1F\x00\xF3\x5E\x00\x80\xF7\x5E\x00\x80\xFB\x5E\x00\x80\xFF\x5E"
-     "\x00\x80"
-     "\xEF\x48\x02\x00\x03\x5F\x00\x80\x07\x5F\x00\x80\x0B\x5F\x00\x80\x0F\x5F"
-     "\x00\x80"
-     "\xBE\xE8\x04\x00\x13\x5F\x00\x80\x17\x5F\x00\x80\x1B\x5F\x00\x80\x1F\x5F"
-     "\x00\x80"
-     "\xE1\x90\x01\x00\x23\x5F\x00\x80\xE3\xE8\x02\x00\xB3\x49\x03\x00\x27\x5F"
-     "\x00\x80"
-     "\x2B\x5F\x00\x80\x2F\x5F\x00\x80\x33\x5F\x00\x80\xB6\x49\x03\x00\xB5\x49"
-     "\x03\x00"
-     "\x37\x5F\x00\x80\xBB\x2D\x03\x00\xBA\x25\x03\x00\x3B\x5F\x00\x80\x3F\x5F"
-     "\x00\x80"
-     "\xBF\x15\x03\x00\xBE\x15\x03\x00\xBD\x21\x03\x00\xBC\x29\x03\x00\xA8\x35"
-     "\x02\x00"
-     "\xA9\x81\x02\x00\xAA\x81\x02\x00\xAB\x81\x02\x00\xAC\x81\x02\x00\xAD\x89"
-     "\x02\x00"
-     "\xAE\xB1\x02\x00\xAF\xB1\x02\x00\x80\xFD\x01\x00\x81\xCD\x01\x00\x82\xC5"
-     "\x01\x00"
-     "\x47\x5F\x00\x80\x86\x90\x04\x00\x87\x04\x05\x00\x4B\x5F\x00\x80\x84\x70"
-     "\x04\x00"
-     "\xB8\x49\x01\x00\xB9\x49\x01\x00\xBA\x59\x01\x00\xBB\x59\x01\x00\xBC\x49"
-     "\x01\x00"
-     "\xBD\x49\x01\x00\xBE\x79\x01\x00\xBF\x79\x01\x00\xB0\xA1\x02\x00\xB1\xA9"
-     "\x02\x00"
-     "\xB2\xBD\x02\x00\xB3\xB5\x02\x00\xB4\x91\x02\x00\xB5\x91\x02\x00\xB6\x79"
-     "\x01\x00"
-     "\xB7\x79\x01\x00\x4F\x5F\x00\x80\x53\x5F\x00\x80\x57\x5F\x00\x80\x5B\x5F"
-     "\x00\x80"
-     "\x5F\x5F\x00\x80\x63\x5F\x00\x80\x67\x5F\x00\x80\xEF\xD0\x1E\x00\xBE\xE8"
-     "\x04\x00"
-     "\xE1\x5C\x1E\x00\x6B\x5F\x00\x80\xE3\x90\x00\x00\x6F\x5F\x00\x80\x73\x5F"
-     "\x00\x80"
-     "\x77\x5F\x00\x80\x7B\x5F\x00\x80\xA3\x49\x02\x00\x7F\x5F\x00\x80\x83\x5F"
-     "\x00\x80"
-     "\x87\x5F\x00\x80\x8B\x5F\x00\x80\xA6\x49\x02\x00\xA5\x49\x02\x00\x8F\x5F"
-     "\x00\x80"
-     "\xAB\x2D\x02\x00\xAA\x25\x02\x00\x93\x5F\x00\x80\x97\x5F\x00\x80\xAF\x15"
-     "\x02\x00"
-     "\xAE\x15\x02\x00\xAD\x21\x02\x00\xAC\x29\x02\x00\xA8\x35\x06\x00\xA9\x3D"
-     "\x06\x00"
-     "\xAA\x55\x06\x00\xAB\x65\x06\x00\xAC\x7D\x06\x00\xAD\x65\x06\x00\xAE\x6D"
-     "\x06\x00"
-     "\xAF\x61\x06\x00\x43\x5F\x00\x80\x9B\x5F\x00\x80\x9F\x5F\x00\x80\xA3\x5F"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\xB1\x00\x00\x82\xB1\x00\x00\xA7\x5F\x00\x80\xB8\xE9"
-     "\x06\x00"
-     "\xB9\xE9\x06\x00\xBA\xF9\x06\x00\xBB\xF5\x06\x00\xBC\x95\x06\x00\xBD\x9D"
-     "\x06\x00"
-     "\xBE\x95\x06\x00\xBF\x8D\x06\x00\xB0\xE1\x06\x00\xB1\xE1\x06\x00\xB2\xE1"
-     "\x06\x00"
-     "\xB3\xFD\x06\x00\xB4\xE5\x06\x00\xB5\xED\x06\x00\xB6\xE5\x06\x00\xB7\xD9"
-     "\x06\x00"
-     "\xB3\xDD\x06\x00\xAB\x5F\x00\x80\xAF\x5F\x00\x80\xB3\x5F\x00\x80\xB7\x5F"
-     "\x00\x80"
-     "\xB6\xE5\x06\x00\xB5\xF5\x06\x00\xBB\x5F\x00\x80\xBB\x25\x06\x00\xBA\x25"
-     "\x06\x00"
-     "\x86\x98\x00\x00\x87\xAC\x00\x00\xBF\x29\x06\x00\xBE\x21\x06\x00\xBD\x29"
-     "\x06\x00"
-     "\xBC\x35\x06\x00\xBF\x5F\x00\x80\xA3\x99\x06\x00\xC3\x5F\x00\x80\xC7\x5F"
-     "\x00\x80"
-     "\xA6\xA1\x06\x00\xCB\x5F\x00\x80\xCF\x5F\x00\x80\xA5\xB1\x06\x00\xAA\x61"
-     "\x06\x00"
-     "\xAB\x61\x06\x00\xD3\x5F\x00\x80\xD7\x5F\x00\x80\xAE\x65\x06\x00\xAF\x6D"
-     "\x06\x00"
-     "\xAC\x71\x06\x00\xAD\x6D\x06\x00\xA8\x2D\x06\x00\xA9\x3D\x06\x00\xAA\x89"
-     "\x06\x00"
-     "\xAB\x89\x06\x00\xAC\x99\x06\x00\xAD\x99\x06\x00\xAE\x89\x06\x00\xAF\x89"
-     "\x06\x00"
-     "\xDB\x5F\x00\x80\xDF\x5F\x00\x80\xE3\x5F\x00\x80\xE7\x5F\x00\x80\xEB\x5F"
-     "\x00\x80"
-     "\xEF\x5F\x00\x80\xF3\x5F\x00\x80\xF7\x5F\x00\x80\xB8\x8D\x06\x00\xB9\x95"
-     "\x06\x00"
-     "\xBA\x95\x06\x00\xBB\xA5\x06\x00\xBC\xBD\x06\x00\xBD\x71\x01\x00\xBE\x71"
-     "\x01\x00"
-     "\xBF\x71\x01\x00\xB0\xF9\x06\x00\xB1\xCD\x06\x00\xB2\xD9\x06\x00\xB3\xD9"
-     "\x06\x00"
-     "\xB4\xC9\x06\x00\xB5\xC9\x06\x00\xB6\xBD\x06\x00\xB7\xB5\x06\x00\xB3\x01"
-     "\x06\x00"
-     "\xFB\x5F\x00\x80\xFF\x5F\x00\x80\x03\x60\x00\x80\x07\x60\x00\x80\xB6\x19"
-     "\x06\x00"
-     "\xB5\x11\x06\x00\x0B\x60\x00\x80\xBB\x09\x06\x00\xBA\x3D\x06\x00\x0F\x60"
-     "\x00\x80"
-     "\x13\x60\x00\x80\xBF\x0D\x06\x00\xBE\x0D\x06\x00\xBD\x0D\x06\x00\xBC\x0D"
-     "\x06\x00"
-     "\x17\x60\x00\x80\xA3\x45\x06\x00\xD3\x5E\x00\x80\x1B\x60\x00\x80\xA6\x5D"
-     "\x06\x00"
-     "\x1F\x60\x00\x80\x84\x58\x00\x00\xA5\x55\x06\x00\xAA\x79\x06\x00\xAB\x4D"
-     "\x06\x00"
-     "\xBE\x68\x01\x00\x23\x60\x00\x80\xAE\x49\x06\x00\xAF\x49\x06\x00\xAC\x49"
-     "\x06\x00"
-     "\xAD\x49\x06\x00\x80\xC1\x03\x00\x81\xC9\x03\x00\x82\xDD\x03\x00\xA3\xCD"
-     "\x02\x00"
-     "\x27\x60\x00\x80\xA5\xD9\x02\x00\xA6\xCD\x02\x00\x2B\x60\x00\x80\x86\x80"
-     "\x0D\x00"
-     "\x87\x94\x03\x00\xAA\xC5\x02\x00\xAB\x0D\x03\x00\xAC\x15\x03\x00\xAD\x1D"
-     "\x03\x00"
-     "\xAE\x15\x03\x00\xAF\x0D\x03\x00\xE1\x9C\x17\x00\xE3\x18\x06\x00\xE3\x14"
-     "\x03\x00"
-     "\xE1\x34\x06\x00\xEF\x18\x02\x00\x2F\x60\x00\x80\x33\x60\x00\x80\x37\x60"
-     "\x00\x80"
-     "\xE3\xD0\x02\x00\x3B\x60\x00\x80\xE1\x50\x02\x00\x3F\x60\x00\x80\x43\x60"
-     "\x00\x80"
-     "\xEF\x2C\x06\x00\xEF\xE4\x25\x00\x47\x60\x00\x80\xAC\x4D\x02\x00\xAD\x51"
-     "\x02\x00"
-     "\xAE\x51\x02\x00\xAF\x65\x02\x00\xA8\x01\x02\x00\xA9\x09\x02\x00\xAA\x59"
-     "\x02\x00"
-     "\xAB\x55\x02\x00\x84\xEC\x0D\x00\x4B\x60\x00\x80\x4F\x60\x00\x80\x53\x60"
-     "\x00\x80"
-     "\xBE\xF8\x0D\x00\x57\x60\x00\x80\x5B\x60\x00\x80\x5F\x60\x00\x80\xBC\x51"
-     "\x03\x00"
-     "\xBD\x59\x03\x00\xBE\x61\x03\x00\xBF\x61\x03\x00\xB8\xED\x03\x00\xB9\x51"
-     "\x03\x00"
-     "\xBA\x51\x03\x00\xBB\x51\x03\x00\xB4\xCD\x03\x00\xB5\xD5\x03\x00\xB6\xDD"
-     "\x03\x00"
-     "\xB7\xD5\x03\x00\xB0\x1D\x02\x00\xB1\xD5\x03\x00\xB2\xDD\x03\x00\xB3\xD5"
-     "\x03\x00"
-     "\xE3\xC8\x00\x00\x63\x60\x00\x80\xE1\xB8\x01\x00\x67\x60\x00\x80\x84\x54"
-     "\x0F\x00"
-     "\x6B\x60\x00\x80\x6F\x60\x00\x80\x73\x60\x00\x80\xA1\xE9\x02\x00\xA0\x15"
-     "\x06\x00"
-     "\xA3\xA5\x03\x00\xA2\x0D\x03\x00\xEF\x20\x00\x00\x77\x60\x00\x80\x7B\x60"
-     "\x00\x80"
-     "\x7F\x60\x00\x80\x83\x60\x00\x80\x87\x60\x00\x80\x85\x34\x26\x00\xB3\x8D"
-     "\x03\x00"
-     "\x8B\x60\x00\x80\xB5\x99\x03\x00\xB6\x8D\x03\x00\x8F\x60\x00\x80\x86\xC0"
-     "\x0F\x00"
-     "\x87\x94\x0D\x00\xBA\x85\x03\x00\xBB\x4D\x02\x00\xBC\x55\x02\x00\xBD\x5D"
-     "\x02\x00"
-     "\xBE\x55\x02\x00\xBF\x4D\x02\x00\x97\x60\x00\x80\x9B\x60\x00\x80\x9F\x60"
-     "\x00\x80"
-     "\xA3\x60\x00\x80\xA7\x60\x00\x80\xAB\x60\x00\x80\x8F\xDD\xEB\x00\xEF\xC4"
-     "\x06\x00"
-     "\xBE\xE0\x0F\x00\xE1\x8C\x06\x00\xAF\x60\x00\x80\xE3\x80\x06\x00\x80\xF5"
-     "\x00\x00"
-     "\x81\xE5\x00\x00\x82\xE5\x00\x00\xB3\x60\x00\x80\x99\x6D\x1F\x00\x98\xC5"
-     "\x1F\x00"
-     "\x9B\xC9\x1B\x00\x9A\x79\x1A\x00\x9D\x75\x1A\x00\x9C\xC5\x1B\x00\x9F\xF9"
-     "\x07\x00"
-     "\x9E\x19\x06\x00\x91\x69\x16\x00\x90\xB1\xEB\x00\x93\x6D\x17\x00\x92\xCD"
-     "\x17\x00"
-     "\x95\xD1\x13\x00\x94\x69\x12\x00\x97\x51\x12\x00\x96\xCD\x13\x00\x83\x55"
-     "\xE4\x00"
-     "\x82\x41\xE4\x00\x93\x60\x00\x80\xB7\x60\x00\x80\x87\x8D\x1D\x00\x86\x91"
-     "\x18\x00"
-     "\x85\x39\x18\x00\x84\x95\x19\x00\x8B\x11\x1C\x00\x8A\x05\x1C\x00\xBB\x60"
-     "\x00\x80"
-     "\xBF\x60\x00\x80\x8F\x85\x15\x00\x8E\x99\x10\x00\x8D\x39\x10\x00\x8C\x95"
-     "\x1D\x00"
-     "\x93\x51\x14\x00\x92\x45\x14\x00\xC3\x60\x00\x80\xC7\x60\x00\x80\x97\x61"
-     "\x09\x00"
-     "\x96\x75\x08\x00\x95\x9D\x09\x00\x94\xF9\x15\x00\x9B\x41\x0C\x00\x9A\xAD"
-     "\x0D\x00"
-     "\xCB\x60\x00\x80\xCF\x60\x00\x80\xD3\x60\x00\x80\xD7\x60\x00\x80\xDB\x60"
-     "\x00\x80"
-     "\x9C\xF1\x0C\x00\xA1\x6D\x0D\x00\xDF\x60\x00\x80\xA3\x01\x04\x00\xA2\x85"
-     "\x00\x00"
-     "\xA5\x99\x04\x00\xA4\x91\x04\x00\xA7\x19\x38\x00\xA6\xC5\x05\x00\xA9\x09"
-     "\x38\x00"
-     "\xA8\x29\x38\x00\xAB\x89\x3C\x00\xAA\x01\x3C\x00\xAD\x01\x30\x00\xAC\x1D"
-     "\x3C\x00"
-     "\xAF\x29\x30\x00\xAE\x9D\x30\x00\xB1\xE1\x34\x00\xB0\x01\x34\x00\xB3\x01"
-     "\x28\x00"
-     "\xB2\xFD\x34\x00\xB5\xD9\x28\x00\xB4\xD1\x28\x00\xE3\x60\x00\x80\xE7\x60"
-     "\x00\x80"
-     "\xEB\x60\x00\x80\xEF\x60\x00\x80\xF3\x60\x00\x80\xF7\x60\x00\x80\x80\x1D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\xD9\x01\x00\xFB\x60\x00\x80\xA8\x11\x0F\x00\xA9\x19"
-     "\x0F\x00"
-     "\xAA\x5D\x0F\x00\xAB\x55\x0F\x00\xAC\x4D\x0F\x00\xAD\x71\x0F\x00\xAE\x75"
-     "\x0F\x00"
-     "\xAF\x6D\x0F\x00\xFF\x60\x00\x80\x07\x61\x00\x80\x86\x88\x00\x00\x87\x10"
-     "\x01\x00"
-     "\x0B\x61\x00\x80\x0F\x61\x00\x80\x13\x61\x00\x80\x17\x61\x00\x80\xB8\x4D"
-     "\x0E\x00"
-     "\xB9\x51\x0E\x00\xBA\x51\x0E\x00\xBB\x51\x0E\x00\xBC\x65\x0E\x00\xBD\x6D"
-     "\x0E\x00"
-     "\xBE\x65\x0E\x00\xBF\x1D\x0E\x00\xB0\x1D\x0F\x00\xB1\xC1\x0F\x00\xB2\xC1"
-     "\x0F\x00"
-     "\xB3\xC1\x0F\x00\xB4\xC5\x0F\x00\xB5\xCD\x0F\x00\xB6\xC5\x0F\x00\xB7\x79"
-     "\x0E\x00"
-     "\xB3\xD5\x0F\x00\x1B\x61\x00\x80\x1F\x61\x00\x80\x23\x61\x00\x80\x27\x61"
-     "\x00\x80"
-     "\xB6\xF1\x0F\x00\xB5\xC5\x0F\x00\x2B\x61\x00\x80\xBB\x05\x0F\x00\xBA\xD9"
-     "\x0F\x00"
-     "\x2F\x61\x00\x80\x33\x61\x00\x80\xBF\x09\x0F\x00\xBE\x01\x0F\x00\xBD\x15"
-     "\x0F\x00"
-     "\xBC\x15\x0F\x00\x37\x61\x00\x80\xA3\x91\x0F\x00\x3B\x61\x00\x80\x3F\x61"
-     "\x00\x80"
-     "\xA6\xB5\x0F\x00\x43\x61\x00\x80\x47\x61\x00\x80\xA5\x81\x0F\x00\xAA\x9D"
-     "\x0F\x00"
-     "\xAB\x41\x0F\x00\x4B\x61\x00\x80\x4F\x61\x00\x80\xAE\x45\x0F\x00\xAF\x4D"
-     "\x0F\x00"
-     "\xAC\x51\x0F\x00\xAD\x51\x0F\x00\xA8\x81\x0D\x00\xA9\x81\x0D\x00\xAA\x81"
-     "\x0D\x00"
-     "\xAB\x81\x0D\x00\xAC\x81\x0D\x00\xAD\x81\x0D\x00\xAE\xB1\x0D\x00\xAF\xA1"
-     "\x0D\x00"
-     "\x53\x61\x00\x80\x57\x61\x00\x80\x5B\x61\x00\x80\x5F\x61\x00\x80\x63\x61"
-     "\x00\x80"
-     "\x82\xB9\x00\x00\x81\xBD\x00\x00\x80\xBD\x00\x00\xB8\x35\x02\x00\xB9\x3D"
-     "\x02\x00"
-     "\xBA\xCD\x02\x00\xBB\x95\x02\x00\xBC\x8D\x02\x00\xBD\xB5\x02\x00\xBE\xBD"
-     "\x02\x00"
-     "\xBF\xB5\x02\x00\xB0\x6D\x02\x00\xB1\x4D\x02\x00\xB2\x45\x02\x00\xB3\x25"
-     "\x02\x00"
-     "\xB4\x3D\x02\x00\xB5\x1D\x02\x00\xB6\x15\x02\x00\xB7\x0D\x02\x00\x67\x61"
-     "\x00\x80"
-     "\x6B\x61\x00\x80\xB3\x01\x0D\x00\x6F\x61\x00\x80\xB5\x01\x0D\x00\x77\x61"
-     "\x00\x80"
-     "\x84\x94\x03\x00\xB6\x09\x0D\x00\xBE\x2C\x04\x00\x7B\x61\x00\x80\xBB\x81"
-     "\x02\x00"
-     "\xBA\xA1\x02\x00\xBD\xF9\x02\x00\xBC\x99\x02\x00\xBF\xD1\x02\x00\xBE\xF1"
-     "\x02\x00"
-     "\x7F\x61\x00\x80\x83\x61\x00\x80\x87\x61\x00\x80\xA3\x45\x0D\x00\x8B\x61"
-     "\x00\x80"
-     "\xA5\x45\x0D\x00\xA6\x4D\x0D\x00\x8F\x61\x00\x80\x86\xE0\x04\x00\x87\xA4"
-     "\x04\x00"
-     "\xAA\xE5\x02\x00\xAB\xC5\x02\x00\xAC\xDD\x02\x00\xAD\xBD\x02\x00\xAE\xB5"
-     "\x02\x00"
-     "\xAF\x95\x02\x00\xA8\x25\x02\x00\xA9\x35\x02\x00\xAA\x3D\x02\x00\xAB\x35"
-     "\x02\x00"
-     "\xAC\x2D\x02\x00\xAD\x91\x02\x00\xAE\x91\x02\x00\xAF\x91\x02\x00\x93\x61"
-     "\x00\x80"
-     "\x97\x61\x00\x80\x9B\x61\x00\x80\x9F\x61\x00\x80\xCC\x00\x00\x00\xA3\x61"
-     "\x00\x80"
-     "\xA7\x61\x00\x80\xAB\x61\x00\x80\xB8\x99\x02\x00\xB9\xAD\x02\x00\xBA\xA5"
-     "\x02\x00"
-     "\xBB\x6D\x01\x00\xBC\x75\x01\x00\xBD\x7D\x01\x00\xBE\x75\x01\x00\xBF\x6D"
-     "\x01\x00"
-     "\xB0\xF1\x02\x00\xB1\xF9\x02\x00\xB2\xC1\x02\x00\xB3\xC1\x02\x00\xB4\xB1"
-     "\x02\x00"
-     "\xB5\xBD\x02\x00\xB6\xB5\x02\x00\xB7\xA9\x02\x00\xAF\x61\x00\x80\xE3\x84"
-     "\x0E\x00"
-     "\xB3\x61\x00\x80\xE1\xF4\x0E\x00\xB7\x61\x00\x80\xBB\x61\x00\x80\xBF\x61"
-     "\x00\x80"
-     "\xC3\x61\x00\x80\x84\x20\x05\x00\xC7\x61\x00\x80\xCB\x61\x00\x80\xCF\x61"
-     "\x00\x80"
-     "\xD3\x61\x00\x80\xEF\xEC\x0E\x00\xD7\x61\x00\x80\xDB\x61\x00\x80\xB3\xF5"
-     "\x02\x00"
-     "\xDF\x61\x00\x80\x86\xE8\x04\x00\x87\x8C\x04\x00\xBE\x5C\x04\x00\xB6\x51"
-     "\x02\x00"
-     "\xB5\xE5\x02\x00\xE7\x61\x00\x80\xBB\x7D\x02\x00\xBA\x75\x02\x00\xEB\x61"
-     "\x00\x80"
-     "\xEF\x61\x00\x80\xBF\x39\x02\x00\xBE\x35\x02\x00\xBD\x55\x02\x00\xBC\x55"
-     "\x02\x00"
-     "\xA3\x35\x05\x00\x73\x61\x00\x80\xE3\x61\x00\x80\xF3\x61\x00\x80\xF7\x61"
-     "\x00\x80"
-     "\xA6\x91\x05\x00\xA5\x25\x05\x00\xFB\x61\x00\x80\xAB\xBD\x05\x00\xAA\xB5"
-     "\x05\x00"
-     "\xFF\x61\x00\x80\x03\x62\x00\x80\xAF\xF9\x05\x00\xAE\xF5\x05\x00\xAD\x95"
-     "\x05\x00"
-     "\xAC\x95\x05\x00\x80\xF9\x07\x00\x81\xF9\x07\x00\x82\x8D\x07\x00\xB3\x8D"
-     "\x06\x00"
-     "\x07\x62\x00\x80\xB5\x9D\x06\x00\xB6\x89\x06\x00\x0B\x62\x00\x80\x0F\x62"
-     "\x00\x80"
-     "\x13\x62\x00\x80\xBA\x4D\x07\x00\xBB\x45\x07\x00\xBC\x5D\x07\x00\xBD\x41"
-     "\x07\x00"
-     "\xBE\x41\x07\x00\xBF\x41\x07\x00\x17\x62\x00\x80\x1B\x62\x00\x80\x03\x61"
-     "\x00\x80"
-     "\x1F\x62\x00\x80\x23\x62\x00\x80\x27\x62\x00\x80\x2B\x62\x00\x80\x2F\x62"
-     "\x00\x80"
-     "\xA8\x35\x06\x00\xA9\x01\x06\x00\xAA\x69\x06\x00\xAB\x69\x06\x00\xAC\x79"
-     "\x06\x00"
-     "\xAD\x65\x06\x00\xAE\x65\x06\x00\xAF\x9D\x07\x00\xB0\xE5\x07\x00\xB1\xED"
-     "\x07\x00"
-     "\xB2\xE5\x07\x00\xB3\xF9\x07\x00\xB4\xE9\x07\x00\xB5\xE9\x07\x00\xB6\x59"
-     "\x07\x00"
-     "\xB7\x55\x07\x00\xB8\x71\x07\x00\xB9\x71\x07\x00\xBA\x71\x07\x00\xBB\x71"
-     "\x07\x00"
-     "\xBC\x55\x07\x00\xBD\x5D\x07\x00\xBE\x55\x07\x00\xBF\x4D\x07\x00\xA3\xC1"
-     "\x07\x00"
-     "\x33\x62\x00\x80\x37\x62\x00\x80\x3B\x62\x00\x80\x3F\x62\x00\x80\xA6\xC5"
-     "\x07\x00"
-     "\xA5\xD1\x07\x00\x43\x62\x00\x80\xAB\x09\x06\x00\xAA\x01\x06\x00\x47\x62"
-     "\x00\x80"
-     "\x4B\x62\x00\x80\xAF\x0D\x06\x00\xAE\x0D\x06\x00\xAD\x0D\x06\x00\xAC\x11"
-     "\x06\x00"
-     "\x80\x69\x00\x00\x81\x69\x00\x00\x82\x05\x00\x00\x4F\x62\x00\x80\xBE\x98"
-     "\x01\x00"
-     "\x84\x98\x01\x00\x53\x62\x00\x80\x57\x62\x00\x80\x86\x00\x3C\x00\x87\xC4"
-     "\x01\x00"
-     "\x5B\x62\x00\x80\x5F\x62\x00\x80\x63\x62\x00\x80\x67\x62\x00\x80\x6B\x62"
-     "\x00\x80"
-     "\x6F\x62\x00\x80\xA8\x5D\x06\x00\xA9\x6D\x06\x00\xAA\x65\x06\x00\xAB\x81"
-     "\x01\x00"
-     "\xAC\x81\x01\x00\xAD\x89\x01\x00\xAE\xB1\x01\x00\xAF\xB1\x01\x00\x73\x62"
-     "\x00\x80"
-     "\x77\x62\x00\x80\x7B\x62\x00\x80\x7F\x62\x00\x80\x83\x62\x00\x80\x87\x62"
-     "\x00\x80"
-     "\x8B\x62\x00\x80\x8F\x62\x00\x80\xB8\x55\x01\x00\xB9\x5D\x01\x00\xBA\x55"
-     "\x01\x00"
-     "\xBB\xC9\x00\x00\xBC\xD9\x00\x00\xBD\xD9\x00\x00\xBE\xC9\x00\x00\xBF\xC1"
-     "\x00\x00"
-     "\xB0\xB1\x01\x00\xB1\xB9\x01\x00\xB2\x89\x01\x00\xB3\x89\x01\x00\xB4\x71"
-     "\x01\x00"
-     "\xB5\x71\x01\x00\xB6\x75\x01\x00\xB7\x6D\x01\x00\xB3\xED\x05\x00\x93\x62"
-     "\x00\x80"
-     "\x97\x62\x00\x80\x9B\x62\x00\x80\x9F\x62\x00\x80\xB6\x09\x02\x00\xB5\x09"
-     "\x02\x00"
-     "\xA3\x62\x00\x80\xBB\x7D\x02\x00\xBA\x75\x02\x00\xA7\x62\x00\x80\xAB\x62"
-     "\x00\x80"
-     "\xBF\xB5\x02\x00\xBE\xB5\x02\x00\xBD\x5D\x02\x00\xBC\x5D\x02\x00\xBE\x60"
-     "\x02\x00"
-     "\xA3\xA9\x05\x00\xAF\x62\x00\x80\xB3\x62\x00\x80\xA6\x4D\x02\x00\xB7\x62"
-     "\x00\x80"
-     "\xBB\x62\x00\x80\xA5\x4D\x02\x00\xAA\x31\x02\x00\xAB\x39\x02\x00\xBF\x62"
-     "\x00\x80"
-     "\x84\xE0\x03\x00\xAE\xF1\x02\x00\xAF\xF1\x02\x00\xAC\x19\x02\x00\xAD\x19"
-     "\x02\x00"
-     "\xBE\x88\x3C\x00\xC7\x62\x00\x80\xAA\x2D\x03\x00\xAB\x25\x03\x00\xAC\x3D"
-     "\x03\x00"
-     "\xAD\x25\x03\x00\xAE\x2D\x03\x00\xAF\x25\x03\x00\x80\xF5\x00\x00\x81\xFD"
-     "\x00\x00"
-     "\x82\xC1\x00\x00\xA3\xC1\x03\x00\xCB\x62\x00\x80\xA5\xC1\x03\x00\xA6\xC1"
-     "\x03\x00"
-     "\xCF\x62\x00\x80\x86\x60\x3C\x00\x87\x54\x03\x00\xD3\x62\x00\x80\xD7\x62"
-     "\x00\x80"
-     "\xDB\x62\x00\x80\xE3\xA8\x02\x00\xDF\x62\x00\x80\xE1\xA4\x01\x00\xE3\x62"
-     "\x00\x80"
-     "\xEF\x5C\x02\x00\xE7\x62\x00\x80\xEB\x62\x00\x80\xEF\x62\x00\x80\xF3\x62"
-     "\x00\x80"
-     "\xF7\x62\x00\x80\xFB\x62\x00\x80\xFF\x62\x00\x80\xE3\x8C\x07\x00\x03\x63"
-     "\x00\x80"
-     "\xE1\x3C\x04\x00\x07\x63\x00\x80\x0B\x63\x00\x80\x0F\x63\x00\x80\x13\x63"
-     "\x00\x80"
-     "\x84\x20\x02\x00\x17\x63\x00\x80\x1B\x63\x00\x80\x1F\x63\x00\x80\x23\x63"
-     "\x00\x80"
-     "\xEF\x6C\x07\x00\x27\x63\x00\x80\x2B\x63\x00\x80\xB3\x2D\x02\x00\x84\x44"
-     "\x3D\x00"
-     "\x2F\x63\x00\x80\x37\x63\x00\x80\x3B\x63\x00\x80\xB6\x2D\x02\x00\xB5\x2D"
-     "\x02\x00"
-     "\x3F\x63\x00\x80\xBB\xC9\x02\x00\xBA\xC1\x02\x00\x43\x63\x00\x80\x47\x63"
-     "\x00\x80"
-     "\xBF\xC9\x02\x00\xBE\xC1\x02\x00\xBD\xC9\x02\x00\xBC\xD1\x02\x00\xE1\x78"
-     "\x07\x00"
-     "\xE3\xC0\x00\x00\xE3\x38\x06\x00\xE1\x5C\x06\x00\x80\xA9\x00\x00\x81\xA9"
-     "\x00\x00"
-     "\x82\xD1\x00\x00\x4B\x63\x00\x80\x4F\x63\x00\x80\x53\x63\x00\x80\xBE\xA4"
-     "\x3C\x00"
-     "\x57\x63\x00\x80\x5B\x63\x00\x80\xEF\x1C\x00\x00\xEF\x90\x06\x00\x5F\x63"
-     "\x00\x80"
-     "\x86\x60\x3C\x00\x87\x04\x3D\x00\x63\x63\x00\x80\xB3\x71\x01\x00\x67\x63"
-     "\x00\x80"
-     "\xB5\x19\x01\x00\xB6\x09\x01\x00\x6B\x63\x00\x80\x6F\x63\x00\x80\x73\x63"
-     "\x00\x80"
-     "\xBA\x01\x01\x00\xBB\x01\x01\x00\xBC\x01\x01\x00\xBD\x01\x01\x00\xBE\x01"
-     "\x01\x00"
-     "\xBF\x01\x01\x00\xA8\x85\x3E\x00\xA9\xB5\x3E\x00\xAA\xB5\x3E\x00\xAB\xC5"
-     "\x3E\x00"
-     "\xAC\xDD\x3E\x00\xAD\xC5\x3E\x00\xAE\xC1\x3E\x00\xAF\xFD\x3E\x00\x77\x63"
-     "\x00\x80"
-     "\x7B\x63\x00\x80\x7F\x63\x00\x80\x83\x63\x00\x80\x87\x63\x00\x80\x8B\x63"
-     "\x00\x80"
-     "\x8F\x63\x00\x80\x93\x63\x00\x80\xB8\x51\x3F\x00\xB9\x51\x3F\x00\xBA\x51"
-     "\x3F\x00"
-     "\xBB\x51\x3F\x00\xBC\x75\x3F\x00\xBD\x7D\x3F\x00\xBE\x75\x3F\x00\xBF\x6D"
-     "\x3F\x00"
-     "\xB0\x89\x3E\x00\xB1\x89\x3E\x00\xB2\x99\x3E\x00\xB3\x99\x3E\x00\xB4\x89"
-     "\x3E\x00"
-     "\xB5\x89\x3E\x00\xB6\x79\x3F\x00\xB7\x75\x3F\x00\xC3\x62\x00\x80\xA3\x39"
-     "\x3E\x00"
-     "\x97\x63\x00\x80\x33\x63\x00\x80\xA6\x41\x3E\x00\x9B\x63\x00\x80\x9F\x63"
-     "\x00\x80"
-     "\xA5\x51\x3E\x00\xAA\x49\x3E\x00\xAB\x49\x3E\x00\xA3\x63\x00\x80\xA7\x63"
-     "\x00\x80"
-     "\xAE\x49\x3E\x00\xAF\x49\x3E\x00\xAC\x49\x3E\x00\xAD\x49\x3E\x00\x80\x49"
-     "\x00\x00"
-     "\x81\x51\x00\x00\x82\x51\x00\x00\xB3\x91\x3F\x00\xAB\x63\x00\x80\xB5\xB9"
-     "\x3F\x00"
-     "\xB6\x45\x3F\x00\xAF\x63\x00\x80\x86\x40\x00\x00\x87\x04\x03\x00\xBA\x45"
-     "\x3F\x00"
-     "\xBB\x5D\x3F\x00\xBC\x4D\x3F\x00\xBD\x3D\x3F\x00\xBE\x29\x3F\x00\xBF\x21"
-     "\x3F\x00"
-     "\xA8\x4D\x3E\x00\xA9\x55\x3E\x00\xAA\x55\x3E\x00\xAB\x65\x3E\x00\xAC\x7D"
-     "\x3E\x00"
-     "\xAD\x89\x3E\x00\xAE\xB9\x3E\x00\xAF\xB9\x3E\x00\xB3\x63\x00\x80\xB7\x63"
-     "\x00\x80"
-     "\xBB\x63\x00\x80\xBF\x63\x00\x80\xC3\x63\x00\x80\xC7\x63\x00\x80\xCB\x63"
-     "\x00\x80"
-     "\xCF\x63\x00\x80\xB8\x61\x01\x00\xB9\x61\x01\x00\xBA\x61\x01\x00\xBB\x61"
-     "\x01\x00"
-     "\xBC\x61\x01\x00\xBD\x61\x01\x00\xBE\x61\x01\x00\xBF\x61\x01\x00\xB0\xCD"
-     "\x3E\x00"
-     "\xB1\xD5\x3E\x00\xB2\xD5\x3E\x00\xB3\xA5\x3E\x00\xB4\xA1\x3E\x00\xB5\xA9"
-     "\x3E\x00"
-     "\xB6\x99\x3E\x00\xB7\x99\x3E\x00\xA3\xDD\x3E\x00\xD3\x63\x00\x80\xD7\x63"
-     "\x00\x80"
-     "\xDB\x63\x00\x80\xDF\x63\x00\x80\xA6\x09\x3E\x00\xA5\xF5\x3E\x00\xE3\x63"
-     "\x00\x80"
-     "\xAB\x11\x3E\x00\xAA\x09\x3E\x00\xE7\x63\x00\x80\xEB\x63\x00\x80\xAF\x6D"
-     "\x3E\x00"
-     "\xAE\x65\x3E\x00\xAD\x71\x3E\x00\xAC\x01\x3E\x00\xEF\x63\x00\x80\xF3\x63"
-     "\x00\x80"
-     "\xF7\x63\x00\x80\xFB\x63\x00\x80\xFF\x63\x00\x80\x03\x64\x00\x80\x07\x64"
-     "\x00\x80"
-     "\x0B\x64\x00\x80\x80\x39\x00\x00\x81\x39\x00\x00\x82\x05\x00\x00\x0F\x64"
-     "\x00\x80"
-     "\xBE\xB8\x01\x00\x84\xB8\x01\x00\x17\x64\x00\x80\x1B\x64\x00\x80\xA8\xAD"
-     "\x02\x00"
-     "\xA9\xD5\x02\x00\xAA\xD5\x02\x00\xAB\x2D\x03\x00\xAC\x35\x03\x00\xAD\x3D"
-     "\x03\x00"
-     "\xAE\x35\x03\x00\xAF\x2D\x03\x00\x1F\x64\x00\x80\x23\x64\x00\x80\x27\x64"
-     "\x00\x80"
-     "\x2B\x64\x00\x80\x2F\x64\x00\x80\x33\x64\x00\x80\x37\x64\x00\x80\x3B\x64"
-     "\x00\x80"
-     "\xB8\xE9\x03\x00\xB9\xE9\x03\x00\xBA\x89\x03\x00\xBB\x89\x03\x00\xBC\x99"
-     "\x03\x00"
-     "\xBD\x99\x03\x00\xBE\x89\x03\x00\xBF\x89\x03\x00\xB0\x55\x03\x00\xB1\x5D"
-     "\x03\x00"
-     "\xB2\x55\x03\x00\xB3\xE9\x03\x00\xB4\xF9\x03\x00\xB5\xF9\x03\x00\xB6\xE9"
-     "\x03\x00"
-     "\xB7\xE1\x03\x00\xB3\x5D\x02\x00\x3F\x64\x00\x80\x86\x28\x04\x00\x87\xCC"
-     "\x03\x00"
-     "\x43\x64\x00\x80\xB6\xBD\x03\x00\xB5\xBD\x03\x00\x47\x64\x00\x80\xBB\x99"
-     "\x03\x00"
-     "\xBA\x91\x03\x00\x4B\x64\x00\x80\x4F\x64\x00\x80\xBF\x3D\x03\x00\xBE\x3D"
-     "\x03\x00"
-     "\xBD\x3D\x03\x00\xBC\x81\x03\x00\x85\x00\x14\x00\xA3\x19\x02\x00\x53\x64"
-     "\x00\x80"
-     "\x57\x64\x00\x80\xA6\xF9\x03\x00\x5B\x64\x00\x80\x5F\x64\x00\x80\xA5\xF9"
-     "\x03\x00"
-     "\xAA\xD5\x03\x00\xAB\xDD\x03\x00\x63\x64\x00\x80\x67\x64\x00\x80\xAE\x79"
-     "\x03\x00"
-     "\xAF\x79\x03\x00\xAC\xC5\x03\x00\xAD\x79\x03\x00\xE3\x54\x3E\x00\xE1\xDC"
-     "\x3F\x00"
-     "\xE1\xD0\x3E\x00\xE3\x3C\x3E\x00\x6B\x64\x00\x80\xEF\x1C\x00\x00\x6F\x64"
-     "\x00\x80"
-     "\x73\x64\x00\x80\x77\x64\x00\x80\xE3\xC0\x00\x00\x7B\x64\x00\x80\xE1\xD4"
-     "\x01\x00"
-     "\xEF\x60\x3E\x00\x7F\x64\x00\x80\x87\x64\x00\x80\xEF\x44\x3F\x00\x80\x61"
-     "\x00\x00"
-     "\x81\x6D\x00\x00\x82\x7D\x00\x00\x84\x00\x05\x00\x86\xF0\x04\x00\x87\x9C"
-     "\x05\x00"
-     "\xBE\x10\x05\x00\x8B\x64\x00\x80\x8F\x64\x00\x80\x93\x64\x00\x80\x97\x64"
-     "\x00\x80"
-     "\x9B\x64\x00\x80\x9F\x64\x00\x80\xA3\x64\x00\x80\xA7\x64\x00\x80\xAB\x64"
-     "\x00\x80"
-     "\xB8\x89\x3D\x00\xB9\x89\x3D\x00\xBA\x99\x3D\x00\xBB\x91\x3D\x00\xBC\xB9"
-     "\x3D\x00"
-     "\xBD\xB9\x3D\x00\xBE\xD1\x3D\x00\xBF\xD1\x3D\x00\xB0\x05\x3E\x00\xB1\x0D"
-     "\x3E\x00"
-     "\xB2\x05\x3E\x00\xB3\x1D\x3E\x00\xB4\x05\x3E\x00\xB5\x0D\x3E\x00\xB6\x05"
-     "\x3E\x00"
-     "\xB7\xB9\x3D\x00\xA8\x9D\x3E\x00\xA9\xAD\x3E\x00\xAA\xA5\x3E\x00\xAB\xBD"
-     "\x3E\x00"
-     "\xAC\xA5\x3E\x00\xAD\xAD\x3E\x00\xAE\xA5\x3E\x00\xAF\x7D\x3E\x00\x84\xAC"
-     "\x04\x00"
-     "\xBE\xAC\x04\x00\xAF\x64\x00\x80\xB3\x64\x00\x80\xB7\x64\x00\x80\xBB\x64"
-     "\x00\x80"
-     "\xBF\x64\x00\x80\xC3\x64\x00\x80\xA8\xF9\x05\x00\xA9\xF9\x05\x00\xAA\x29"
-     "\x06\x00"
-     "\xAB\x29\x06\x00\xAC\x39\x06\x00\xAD\x39\x06\x00\xAE\x29\x06\x00\xAF\x29"
-     "\x06\x00"
-     "\x83\x64\x00\x80\xC7\x64\x00\x80\xCB\x64\x00\x80\xCF\x64\x00\x80\xD3\x64"
-     "\x00\x80"
-     "\xD7\x64\x00\x80\xDB\x64\x00\x80\xDF\x64\x00\x80\xB8\x8D\x06\x00\xB9\x91"
-     "\x06\x00"
-     "\xBA\x91\x06\x00\xBB\xA5\x06\x00\xBC\xBD\x06\x00\xBD\x45\x07\x00\xBE\x41"
-     "\x07\x00"
-     "\xBF\x41\x07\x00\xB0\x59\x06\x00\xB1\x59\x06\x00\xB2\xED\x06\x00\xB3\xFD"
-     "\x06\x00"
-     "\xB4\xED\x06\x00\xB5\xDD\x06\x00\xB6\xD5\x06\x00\xB7\xB5\x06\x00\xB3\xA1"
-     "\x06\x00"
-     "\xE3\x64\x00\x80\xE7\x64\x00\x80\xEB\x64\x00\x80\xEF\x64\x00\x80\xB6\xB9"
-     "\x06\x00"
-     "\xB5\xB1\x06\x00\xF7\x64\x00\x80\xBB\x9D\x06\x00\xBA\x9D\x06\x00\xF3\x64"
-     "\x00\x80"
-     "\x13\x64\x00\x80\xBF\x19\x06\x00\xBE\x29\x06\x00\xBD\x29\x06\x00\xBC\x39"
-     "\x06\x00"
-     "\x82\x51\x00\x00\xA3\xE5\x06\x00\x80\x41\x00\x00\x81\x41\x00\x00\xA6\xFD"
-     "\x06\x00"
-     "\xFB\x64\x00\x80\xFF\x64\x00\x80\xA5\xF5\x06\x00\xAA\xD9\x06\x00\xAB\xD9"
-     "\x06\x00"
-     "\x86\x48\x00\x00\x87\x6C\x00\x00\xAE\x6D\x06\x00\xAF\x5D\x06\x00\xAC\x7D"
-     "\x06\x00"
-     "\xAD\x6D\x06\x00\xA8\x39\x06\x00\xA9\x59\x06\x00\xAA\x69\x06\x00\xAB\x69"
-     "\x06\x00"
-     "\xAC\x79\x06\x00\xAD\x79\x06\x00\xAE\x69\x06\x00\xAF\x69\x06\x00\x03\x65"
-     "\x00\x80"
-     "\x07\x65\x00\x80\x0B\x65\x00\x80\x0F\x65\x00\x80\x13\x65\x00\x80\x17\x65"
-     "\x00\x80"
-     "\x1B\x65\x00\x80\x1F\x65\x00\x80\xB8\x65\x01\x00\xB9\x6D\x01\x00\xBA\x65"
-     "\x01\x00"
-     "\xBB\x7D\x01\x00\xBC\x65\x01\x00\xBD\x6D\x01\x00\xBE\x65\x01\x00\xBF\xD9"
-     "\x01\x00"
-     "\xB0\x19\x06\x00\xB1\x19\x06\x00\xB2\x81\x06\x00\xB3\x81\x06\x00\xB4\x81"
-     "\x06\x00"
-     "\xB5\x81\x06\x00\xB6\x81\x06\x00\xB7\x81\x06\x00\xB3\xE1\x06\x00\x23\x65"
-     "\x00\x80"
-     "\x27\x65\x00\x80\x2B\x65\x00\x80\x2F\x65\x00\x80\xB6\xF9\x06\x00\xB5\xF1"
-     "\x06\x00"
-     "\x33\x65\x00\x80\xBB\xDD\x06\x00\xBA\xDD\x06\x00\x37\x65\x00\x80\x3B\x65"
-     "\x00\x80"
-     "\xBF\x45\x06\x00\xBE\x45\x06\x00\xBD\x55\x06\x00\xBC\x55\x06\x00\x3F\x65"
-     "\x00\x80"
-     "\xA3\xA5\x06\x00\x43\x65\x00\x80\x47\x65\x00\x80\xA6\xBD\x06\x00\x4B\x65"
-     "\x00\x80"
-     "\x4F\x65\x00\x80\xA5\xB5\x06\x00\xAA\x99\x06\x00\xAB\x99\x06\x00\x53\x65"
-     "\x00\x80"
-     "\x57\x65\x00\x80\xAE\x01\x06\x00\xAF\x01\x06\x00\xAC\x11\x06\x00\xAD\x11"
-     "\x06\x00"
-     "\xA8\x9D\x02\x00\xA9\xD1\x02\x00\xAA\xD1\x02\x00\xAB\x2D\x03\x00\xAC\x35"
-     "\x03\x00"
-     "\xAD\x3D\x03\x00\xAE\x35\x03\x00\xAF\x2D\x03\x00\x5B\x65\x00\x80\x5F\x65"
-     "\x00\x80"
-     "\xBE\x64\x02\x00\x67\x65\x00\x80\x6B\x65\x00\x80\x6F\x65\x00\x80\x73\x65"
-     "\x00\x80"
-     "\x77\x65\x00\x80\xB8\xE9\x03\x00\xB9\xE9\x03\x00\xBA\x89\x03\x00\xBB\x85"
-     "\x03\x00"
-     "\xBC\x9D\x03\x00\xBD\x81\x03\x00\xBE\x81\x03\x00\xBF\xB5\x03\x00\xB0\x55"
-     "\x03\x00"
-     "\xB1\x5D\x03\x00\xB2\x55\x03\x00\xB3\xE9\x03\x00\xB4\xF9\x03\x00\xB5\xF9"
-     "\x03\x00"
-     "\xB6\xE9\x03\x00\xB7\xE1\x03\x00\x80\x6D\x03\x00\x81\xA5\x00\x00\x82\xAD"
-     "\x00\x00"
-     "\xB3\x55\x02\x00\x7B\x65\x00\x80\xB5\xB1\x03\x00\xB6\xB1\x03\x00\x7F\x65"
-     "\x00\x80"
-     "\x84\xE0\x02\x00\x83\x65\x00\x80\xBA\x9D\x03\x00\xBB\x95\x03\x00\xBC\x8D"
-     "\x03\x00"
-     "\xBD\x31\x03\x00\xBE\x31\x03\x00\xBF\x31\x03\x00\xA3\x19\x02\x00\x87\x65"
-     "\x00\x80"
-     "\x85\x70\x69\x00\x8B\x65\x00\x80\x8F\x65\x00\x80\xA6\xFD\x03\x00\xA5\xFD"
-     "\x03\x00"
-     "\x93\x65\x00\x80\xAB\xD9\x03\x00\xAA\xD1\x03\x00\x86\x48\x0C\x00\x87\xBC"
-     "\x03\x00"
-     "\xAF\x7D\x03\x00\xAE\x7D\x03\x00\xAD\x7D\x03\x00\xAC\xC1\x03\x00\x97\x65"
-     "\x00\x80"
-     "\x9B\x65\x00\x80\x9F\x65\x00\x80\xA3\x65\x00\x80\xEF\xB0\x06\x00\xEF\xC4"
-     "\x03\x00"
-     "\xA7\x65\x00\x80\xAB\x65\x00\x80\xE3\x7C\x06\x00\xE3\x94\x03\x00\xE1\xB8"
-     "\x07\x00"
-     "\xE1\xDC\x01\x00\xAF\x65\x00\x80\xB3\x65\x00\x80\xB7\x65\x00\x80\xBB\x65"
-     "\x00\x80"
-     "\xBF\x65\x00\x80\xC3\x65\x00\x80\x84\x44\x02\x00\xBE\x58\x0D\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x35\x00\x00\x82\x3D\x00\x00\xC7\x65\x00\x80\xCB\x65\x00\x80\xCF\x65"
-     "\x00\x80"
-     "\x86\xC8\x0C\x00\x87\x5C\x0D\x00\xD7\x65\x00\x80\xDB\x65\x00\x80\xDF\x65"
-     "\x00\x80"
-     "\xE3\x65\x00\x80\xE7\x65\x00\x80\xEB\x65\x00\x80\xEF\x65\x00\x80\xF3\x65"
-     "\x00\x80"
-     "\xF7\x65\x00\x80\xFB\x65\x00\x80\xFF\x65\x00\x80\xEF\x80\x06\x00\x84\xAC"
-     "\x0D\x00"
-     "\xE1\xEC\x06\x00\x03\x66\x00\x80\xE3\x5C\x06\x00\x07\x66\x00\x80\x0B\x66"
-     "\x00\x80"
-     "\x0F\x66\x00\x80\x13\x66\x00\x80\xB3\xF5\x01\x00\x17\x66\x00\x80\x1B\x66"
-     "\x00\x80"
-     "\x1F\x66\x00\x80\x23\x66\x00\x80\xB6\x45\x01\x00\xB5\xE5\x01\x00\x27\x66"
-     "\x00\x80"
-     "\xBB\x2D\x01\x00\xBA\x21\x01\x00\x2B\x66\x00\x80\x2F\x66\x00\x80\xBF\xF5"
-     "\x00\x00"
-     "\xBE\xF5\x00\x00\xBD\x25\x01\x00\xBC\x2D\x01\x00\xA8\x2D\x0E\x00\xA9\x35"
-     "\x0E\x00"
-     "\xAA\x3D\x0E\x00\xAB\x35\x0E\x00\xAC\x2D\x0E\x00\xAD\x85\x0E\x00\xAE\x85"
-     "\x0E\x00"
-     "\xAF\xB9\x0E\x00\xD3\x65\x00\x80\x33\x66\x00\x80\x37\x66\x00\x80\x3B\x66"
-     "\x00\x80"
-     "\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x3F\x66\x00\x80\xB8\x59"
-     "\x0F\x00"
-     "\xB9\x59\x0F\x00\xBA\x79\x0F\x00\xBB\x79\x0F\x00\xBC\x69\x0F\x00\xBD\x69"
-     "\x0F\x00"
-     "\xBE\x19\x0F\x00\xBF\x19\x0F\x00\xB0\xA5\x0E\x00\xB1\xA9\x0E\x00\xB2\xB9"
-     "\x0E\x00"
-     "\xB3\xB1\x0E\x00\xB4\x71\x0F\x00\xB5\x71\x0F\x00\xB6\x71\x0F\x00\xB7\x71"
-     "\x0F\x00"
-     "\xA5\xBD\x0E\x00\xBE\x88\x03\x00\x47\x66\x00\x80\xA6\x1D\x0E\x00\x43\x66"
-     "\x00\x80"
-     "\x4B\x66\x00\x80\xA3\xAD\x0E\x00\x4F\x66\x00\x80\xAD\x7D\x0E\x00\xAC\x75"
-     "\x0E\x00"
-     "\xAF\xAD\x0F\x00\xAE\xAD\x0F\x00\x63\x65\x00\x80\x53\x66\x00\x80\xAB\x75"
-     "\x0E\x00"
-     "\xAA\x79\x0E\x00\xB3\xB9\x0F\x00\x57\x66\x00\x80\x86\x68\x00\x00\x87\x0C"
-     "\x03\x00"
-     "\x5B\x66\x00\x80\xB6\x51\x0F\x00\xB5\x59\x0F\x00\x5F\x66\x00\x80\xBB\x75"
-     "\x0F\x00"
-     "\xBA\x75\x0F\x00\x63\x66\x00\x80\x67\x66\x00\x80\xBF\x45\x0F\x00\xBE\x45"
-     "\x0F\x00"
-     "\xBD\x51\x0F\x00\xBC\x65\x0F\x00\xA8\x71\x0E\x00\xA9\x71\x0E\x00\xAA\x71"
-     "\x0E\x00"
-     "\xAB\x71\x0E\x00\xAC\x91\x0E\x00\xAD\x91\x0E\x00\xAE\x91\x0E\x00\xAF\x91"
-     "\x0E\x00"
-     "\x6B\x66\x00\x80\x6F\x66\x00\x80\x73\x66\x00\x80\x77\x66\x00\x80\x7B\x66"
-     "\x00\x80"
-     "\x7F\x66\x00\x80\x83\x66\x00\x80\x87\x66\x00\x80\xB8\x85\x0E\x00\xB9\x8D"
-     "\x0E\x00"
-     "\xBA\x85\x0E\x00\xBB\x9D\x0E\x00\xBC\x8D\x0E\x00\xBD\xBD\x0E\x00\xBE\xB5"
-     "\x0E\x00"
-     "\xBF\x79\x01\x00\xB0\xF1\x0E\x00\xB1\xF1\x0E\x00\xB2\xF1\x0E\x00\xB3\xC5"
-     "\x0E\x00"
-     "\xB4\xC1\x0E\x00\xB5\xC1\x0E\x00\xB6\xC1\x0E\x00\xB7\xC1\x0E\x00\xA3\xF9"
-     "\x0E\x00"
-     "\x8B\x66\x00\x80\x8F\x66\x00\x80\x93\x66\x00\x80\x97\x66\x00\x80\xA6\x11"
-     "\x0E\x00"
-     "\xA5\x19\x0E\x00\x9B\x66\x00\x80\xAB\x35\x0E\x00\xAA\x35\x0E\x00\x9F\x66"
-     "\x00\x80"
-     "\xA3\x66\x00\x80\xAF\x05\x0E\x00\xAE\x05\x0E\x00\xAD\x11\x0E\x00\xAC\x25"
-     "\x0E\x00"
-     "\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\xA7\x66\x00\x80\xAB\x66"
-     "\x00\x80"
-     "\xAF\x66\x00\x80\x84\x94\x01\x00\xBE\x94\x01\x00\x86\x40\x07\x00\x87\xE4"
-     "\x00\x00"
-     "\xB7\x66\x00\x80\xBB\x66\x00\x80\xBF\x66\x00\x80\xC3\x66\x00\x80\xC7\x66"
-     "\x00\x80"
-     "\xCB\x66\x00\x80\xA8\x89\x02\x00\xA9\x91\x02\x00\xAA\x95\x02\x00\xAB\xB9"
-     "\x02\x00"
-     "\xAC\xD5\x02\x00\xAD\xC5\x02\x00\xAE\xC5\x02\x00\xAF\xF5\x02\x00\xCF\x66"
-     "\x00\x80"
-     "\xD3\x66\x00\x80\xD7\x66\x00\x80\xDB\x66\x00\x80\xBE\x7C\x03\x00\xDF\x66"
-     "\x00\x80"
-     "\xE3\x66\x00\x80\xE7\x66\x00\x80\xB8\x7D\x03\x00\xB9\xC1\x03\x00\xBA\xC1"
-     "\x03\x00"
-     "\xBB\xC1\x03\x00\xBC\xC1\x03\x00\xBD\xC9\x03\x00\xBE\xF1\x03\x00\xBF\xF1"
-     "\x03\x00"
-     "\xB0\x8D\x02\x00\xB1\x45\x03\x00\xB2\x4D\x03\x00\xB3\x45\x03\x00\xB4\x5D"
-     "\x03\x00"
-     "\xB5\x45\x03\x00\xB6\x4D\x03\x00\xB7\x45\x03\x00\xB3\x1D\x02\x00\xEB\x66"
-     "\x00\x80"
-     "\xEF\x66\x00\x80\xF3\x66\x00\x80\xF7\x66\x00\x80\xB6\x45\x02\x00\xB5\x5D"
-     "\x02\x00"
-     "\xFB\x66\x00\x80\xBB\x81\x03\x00\xBA\x49\x02\x00\xFF\x66\x00\x80\x03\x67"
-     "\x00\x80"
-     "\xBF\x81\x03\x00\xBE\x99\x03\x00\xBD\x91\x03\x00\xBC\x99\x03\x00\x07\x67"
-     "\x00\x80"
-     "\xA3\x59\x02\x00\x0B\x67\x00\x80\x0F\x67\x00\x80\xA6\x01\x02\x00\x13\x67"
-     "\x00\x80"
-     "\x17\x67\x00\x80\xA5\x19\x02\x00\xAA\x0D\x02\x00\xAB\xC5\x03\x00\x1B\x67"
-     "\x00\x80"
-     "\x1F\x67\x00\x80\xAE\xDD\x03\x00\xAF\xC5\x03\x00\xAC\xDD\x03\x00\xAD\xD5"
-     "\x03\x00"
-     "\x80\xD9\x01\x00\x81\xED\x01\x00\x82\xE5\x01\x00\xEF\xB8\x0E\x00\x27\x67"
-     "\x00\x80"
-     "\xE1\xC4\x01\x00\x84\x98\x02\x00\xE3\xD4\x00\x00\x2B\x67\x00\x80\xBE\xEC"
-     "\x04\x00"
-     "\x2F\x67\x00\x80\xEF\x08\x00\x00\x33\x67\x00\x80\xE1\xC4\x0F\x00\x37\x67"
-     "\x00\x80"
-     "\xE3\x64\x0E\x00\x86\x00\x05\x00\x87\x68\x05\x00\x3B\x67\x00\x80\xB3\xBD"
-     "\x02\x00"
-     "\x3F\x67\x00\x80\xB5\xAD\x02\x00\xB6\xA5\x02\x00\x43\x67\x00\x80\x47\x67"
-     "\x00\x80"
-     "\x4B\x67\x00\x80\xBA\x41\x01\x00\xBB\x41\x01\x00\xBC\x45\x01\x00\xBD\x4D"
-     "\x01\x00"
-     "\xBE\x45\x01\x00\xBF\xF9\x01\x00\x4F\x67\x00\x80\x53\x67\x00\x80\x57\x67"
-     "\x00\x80"
-     "\x5B\x67\x00\x80\x5F\x67\x00\x80\x63\x67\x00\x80\x67\x67\x00\x80\xEF\xE0"
-     "\x01\x00"
-     "\x84\x6C\x04\x00\xE1\xD4\x0E\x00\x6B\x67\x00\x80\xE3\x1C\x0E\x00\x6F\x67"
-     "\x00\x80"
-     "\x73\x67\x00\x80\x77\x67\x00\x80\x7B\x67\x00\x80\xA3\x31\x02\x00\x7F\x67"
-     "\x00\x80"
-     "\x84\x24\x07\x00\x83\x67\x00\x80\x87\x67\x00\x80\xA6\x29\x02\x00\xA5\x21"
-     "\x02\x00"
-     "\x8B\x67\x00\x80\xAB\xCD\x01\x00\xAA\xCD\x01\x00\x8F\x67\x00\x80\x97\x67"
-     "\x00\x80"
-     "\xAF\x75\x01\x00\xAE\xC9\x01\x00\xAD\xC1\x01\x00\xAC\xC9\x01\x00\xA8\xE9"
-     "\x05\x00"
-     "\xA9\xE9\x05\x00\xAA\xF9\x05\x00\xAB\xF9\x05\x00\xAC\xE9\x05\x00\xAD\xE9"
-     "\x05\x00"
-     "\xAE\x39\x06\x00\xAF\x39\x06\x00\x23\x67\x00\x80\x82\xCD\x07\x00\x81\xF5"
-     "\x07\x00"
-     "\x80\xFD\x07\x00\x93\x67\x00\x80\x9B\x67\x00\x80\x86\x18\x03\x00\x87\x90"
-     "\x03\x00"
-     "\xB8\xD1\x06\x00\xB9\xD9\x06\x00\xBA\xE1\x06\x00\xBB\xE1\x06\x00\xBC\x91"
-     "\x06\x00"
-     "\xBD\x9D\x06\x00\xBE\x95\x06\x00\xBF\x89\x06\x00\xB0\x49\x06\x00\xB1\x49"
-     "\x06\x00"
-     "\xB2\x5D\x06\x00\xB3\x55\x06\x00\xB4\x4D\x06\x00\xB5\xF1\x06\x00\xB6\xF1"
-     "\x06\x00"
-     "\xB7\xF1\x06\x00\xB0\xE1\x07\x00\xB1\xE1\x07\x00\xB2\x09\x07\x00\xB3\x09"
-     "\x07\x00"
-     "\xB4\x19\x07\x00\xB5\x19\x07\x00\xB6\x09\x07\x00\xB7\x09\x07\x00\xB8\x39"
-     "\x07\x00"
-     "\xB9\x0D\x07\x00\xBA\x19\x07\x00\xBB\x19\x07\x00\xBC\x09\x07\x00\xBD\x09"
-     "\x07\x00"
-     "\xBE\x7D\x07\x00\xBF\x71\x07\x00\x9F\x67\x00\x80\xB3\x66\x00\x80\xA3\x67"
-     "\x00\x80"
-     "\xA7\x67\x00\x80\xAB\x67\x00\x80\xAF\x67\x00\x80\xB3\x67\x00\x80\xB7\x67"
-     "\x00\x80"
-     "\xA8\xF1\x07\x00\xA9\xC5\x07\x00\xAA\xC1\x07\x00\xAB\xDD\x07\x00\xAC\xC9"
-     "\x07\x00"
-     "\xAD\xBD\x07\x00\xAE\xA9\x07\x00\xAF\xA1\x07\x00\xB3\x29\x06\x00\xBB\x67"
-     "\x00\x80"
-     "\xBF\x67\x00\x80\xC3\x67\x00\x80\xC7\x67\x00\x80\xB6\x5D\x06\x00\xB5\x21"
-     "\x06\x00"
-     "\xCB\x67\x00\x80\xBB\x45\x06\x00\xBA\x45\x06\x00\xCF\x67\x00\x80\xD3\x67"
-     "\x00\x80"
-     "\xBF\xBD\x06\x00\xBE\xBD\x06\x00\xBD\xBD\x06\x00\xBC\xBD\x06\x00\xD7\x67"
-     "\x00\x80"
-     "\xA3\x6D\x06\x00\xDB\x67\x00\x80\xDF\x67\x00\x80\xA6\x19\x06\x00\xE3\x67"
-     "\x00\x80"
-     "\xE7\x67\x00\x80\xA5\x65\x06\x00\xAA\x01\x06\x00\xAB\x01\x06\x00\xEB\x67"
-     "\x00\x80"
-     "\xEF\x67\x00\x80\xAE\xF9\x06\x00\xAF\xF9\x06\x00\xAC\xF9\x06\x00\xAD\xF9"
-     "\x06\x00"
-     "\xA8\x6D\x06\x00\xA9\xB1\x01\x00\xAA\x49\x01\x00\xAB\x45\x01\x00\xAC\x5D"
-     "\x01\x00"
-     "\xAD\x45\x01\x00\xAE\x4D\x01\x00\xAF\x45\x01\x00\xF3\x67\x00\x80\x82\x1D"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x80\x1D\x00\x00\xF7\x67\x00\x80\xFB\x67\x00\x80\xFF\x67"
-     "\x00\x80"
-     "\xBE\x54\x01\x00\xB8\x81\x00\x00\xB9\x8D\x00\x00\xBA\x85\x00\x00\xBB\x99"
-     "\x00\x00"
-     "\xBC\x89\x00\x00\xBD\xBD\x00\x00\xBE\xB5\x00\x00\xBF\x7D\x00\x00\xB0\x3D"
-     "\x01\x00"
-     "\xB1\xE1\x00\x00\xB2\xE1\x00\x00\xB3\xE1\x00\x00\xB4\xE1\x00\x00\xB5\xE9"
-     "\x00\x00"
-     "\xB6\xD1\x00\x00\xB7\xD1\x00\x00\xBB\x05\x03\x00\xBA\xD9\x02\x00\x86\x2C"
-     "\x02\x00"
-     "\x87\x2C\x03\x00\xBF\x0D\x03\x00\xBE\x05\x03\x00\xBD\x15\x03\x00\xBC\x15"
-     "\x03\x00"
-     "\xB3\xE1\x02\x00\x07\x68\x00\x80\x0B\x68\x00\x80\x84\x2C\x03\x00\x0F\x68"
-     "\x00\x80"
-     "\xB6\xE5\x02\x00\xB5\xF5\x02\x00\x13\x68\x00\x80\xAA\x9D\x02\x00\xAB\x41"
-     "\x03\x00"
-     "\x17\x68\x00\x80\x1B\x68\x00\x80\xAE\x41\x03\x00\xAF\x49\x03\x00\xAC\x51"
-     "\x03\x00"
-     "\xAD\x51\x03\x00\x1F\x68\x00\x80\xA3\xA5\x02\x00\x23\x68\x00\x80\x27\x68"
-     "\x00\x80"
-     "\xA6\xA1\x02\x00\x2B\x68\x00\x80\x2F\x68\x00\x80\xA5\xB1\x02\x00\xA9\xA9"
-     "\x00\x00"
-     "\xA8\xA1\x00\x00\xAB\xB5\x00\x00\xAA\xBD\x00\x00\xAD\xDD\x00\x00\xAC\xDD"
-     "\x00\x00"
-     "\xAF\xF1\x00\x00\xAE\xFD\x00\x00\xBE\x2C\x1C\x00\x33\x68\x00\x80\x37\x68"
-     "\x00\x80"
-     "\x3B\x68\x00\x80\x3F\x68\x00\x80\x43\x68\x00\x80\x47\x68\x00\x80\x4B\x68"
-     "\x00\x80"
-     "\xB9\x7D\x00\x00\xB8\x7D\x00\x00\xBB\xCD\x01\x00\xBA\xCD\x01\x00\xBD\xDD"
-     "\x01\x00"
-     "\xBC\xDD\x01\x00\xBF\xCD\x01\x00\xBE\xCD\x01\x00\xB1\x95\x00\x00\xB0\x89"
-     "\x00\x00"
-     "\xB3\x4D\x00\x00\xB2\x95\x00\x00\xB5\x5D\x00\x00\xB4\x5D\x00\x00\xB7\x4D"
-     "\x00\x00"
-     "\xB6\x4D\x00\x00\x4F\x68\x00\x80\x53\x68\x00\x80\x57\x68\x00\x80\x5B\x68"
-     "\x00\x80"
-     "\x5F\x68\x00\x80\x63\x68\x00\x80\x67\x68\x00\x80\x6B\x68\x00\x80\x80\x39"
-     "\x00\x00"
-     "\x81\x39\x00\x00\x82\x05\x00\x00\x6F\x68\x00\x80\x77\x68\x00\x80\x7B\x68"
-     "\x00\x80"
-     "\x87\xF8\x02\x00\x86\x7C\x1D\x00\xE1\xB8\x04\x00\xBE\xC8\x1C\x00\xE3\x40"
-     "\x06\x00"
-     "\x7F\x68\x00\x80\x83\x68\x00\x80\x87\x68\x00\x80\x8B\x68\x00\x80\x8F\x68"
-     "\x00\x80"
-     "\x93\x68\x00\x80\x97\x68\x00\x80\x9B\x68\x00\x80\x9F\x68\x00\x80\xA3\x68"
-     "\x00\x80"
-     "\xA7\x68\x00\x80\xEF\xB0\x07\x00\xAB\x68\x00\x80\xAF\x68\x00\x80\xB3\x68"
-     "\x00\x80"
-     "\xB7\x68\x00\x80\xEF\xC8\x00\x00\xBB\x68\x00\x80\xBF\x68\x00\x80\xC3\x68"
-     "\x00\x80"
-     "\xEF\x40\x06\x00\xC7\x68\x00\x80\xE1\xFC\x06\x00\xCB\x68\x00\x80\xE3\x1C"
-     "\x06\x00"
-     "\xCF\x68\x00\x80\xE1\x94\x01\x00\xD3\x68\x00\x80\xE3\x64\x06\x00\x80\x11"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x82\x29\x00\x00\xB3\xFD\x01\x00\xD7\x68\x00\x80\xB5\x9D"
-     "\x01\x00"
-     "\xB6\x95\x01\x00\xDB\x68\x00\x80\xDF\x68\x00\x80\x84\x6C\x1D\x00\xBA\x85"
-     "\x01\x00"
-     "\xBB\x99\x01\x00\xBC\x89\x01\x00\xBD\x51\x01\x00\xBE\x51\x01\x00\xBF\x51"
-     "\x01\x00"
-     "\xA3\x31\x1E\x00\x73\x68\x00\x80\xE3\x68\x00\x80\xE7\x68\x00\x80\xEB\x68"
-     "\x00\x80"
-     "\xA6\x59\x1E\x00\xA5\x51\x1E\x00\xEF\x68\x00\x80\xAB\x55\x1E\x00\xAA\x49"
-     "\x1E\x00"
-     "\x86\x08\x03\x00\x87\x6C\x03\x00\xAF\x9D\x1E\x00\xAE\x9D\x1E\x00\xAD\x9D"
-     "\x1E\x00"
-     "\xAC\x45\x1E\x00\xF3\x68\x00\x80\xB3\x09\x1F\x00\xF7\x68\x00\x80\xFB\x68"
-     "\x00\x80"
-     "\xB6\x09\x1F\x00\xFF\x68\x00\x80\x03\x69\x00\x80\xB5\x09\x1F\x00\xBA\x05"
-     "\x1F\x00"
-     "\xBB\x0D\x1F\x00\x07\x69\x00\x80\x0B\x69\x00\x80\xBE\x05\x1F\x00\xBF\x09"
-     "\x1F\x00"
-     "\xBC\x15\x1F\x00\xBD\x0D\x1F\x00\xB0\xE5\x1F\x00\xB1\xED\x1F\x00\xB2\xE5"
-     "\x1F\x00"
-     "\xB3\xFD\x1F\x00\xB4\xE5\x1F\x00\xB5\xE9\x1F\x00\xB6\x19\x1F\x00\xB7\x19"
-     "\x1F\x00"
-     "\xB8\x29\x1F\x00\xB9\x35\x1F\x00\xBA\x3D\x1F\x00\xBB\x35\x1F\x00\xBC\x11"
-     "\x1F\x00"
-     "\xBD\x1D\x1F\x00\xBE\x09\x1F\x00\xBF\x05\x1F\x00\x0F\x69\x00\x80\x13\x69"
-     "\x00\x80"
-     "\x03\x68\x00\x80\x17\x69\x00\x80\x1B\x69\x00\x80\x1F\x69\x00\x80\x23\x69"
-     "\x00\x80"
-     "\x27\x69\x00\x80\xA8\xD1\x1F\x00\xA9\xD1\x1F\x00\xAA\xA5\x1F\x00\xAB\xBD"
-     "\x1F\x00"
-     "\xAC\xA5\x1F\x00\xAD\xAD\x1F\x00\xAE\xA5\x1F\x00\xAF\x9D\x1F\x00\xA3\x4D"
-     "\x1E\x00"
-     "\x2B\x69\x00\x80\x2F\x69\x00\x80\x33\x69\x00\x80\x37\x69\x00\x80\xA6\x4D"
-     "\x1E\x00"
-     "\xA5\x4D\x1E\x00\x3B\x69\x00\x80\xAB\x49\x1E\x00\xAA\x41\x1E\x00\x3F\x69"
-     "\x00\x80"
-     "\x43\x69\x00\x80\xAF\x4D\x1E\x00\xAE\x41\x1E\x00\xAD\x49\x1E\x00\xAC\x51"
-     "\x1E\x00"
-     "\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\x47\x69\x00\x80\x4B\x69"
-     "\x00\x80"
-     "\x4F\x69\x00\x80\x84\xB4\x01\x00\xBE\xB4\x01\x00\xBF\xE8\x01\x00\x57\x69"
-     "\x00\x80"
-     "\x86\x48\x07\x00\x87\x34\x00\x00\x84\xBC\x06\x00\x5B\x69\x00\x80\x5F\x69"
-     "\x00\x80"
-     "\xBE\xB4\x06\x00\xA8\x8D\x01\x00\xA9\x95\x01\x00\xAA\x95\x01\x00\xAB\xCD"
-     "\x01\x00"
-     "\xAC\xD9\x01\x00\xAD\xD9\x01\x00\xAE\xCD\x01\x00\xAF\xC5\x01\x00\x63\x69"
-     "\x00\x80"
-     "\x67\x69\x00\x80\x6B\x69\x00\x80\x6F\x69\x00\x80\x73\x69\x00\x80\x77\x69"
-     "\x00\x80"
-     "\x7B\x69\x00\x80\x7F\x69\x00\x80\xB8\x1D\x01\x00\xB9\xC1\x00\x00\xBA\xC1"
-     "\x00\x00"
-     "\xBB\xC1\x00\x00\xBC\xC1\x00\x00\xBD\xC9\x00\x00\xBE\xF1\x00\x00\xBF\xF1"
-     "\x00\x00"
-     "\xB0\x89\x01\x00\xB1\x89\x01\x00\xB2\x29\x01\x00\xB3\x29\x01\x00\xB4\x3D"
-     "\x01\x00"
-     "\xB5\x25\x01\x00\xB6\x2D\x01\x00\xB7\x25\x01\x00\xBB\x6D\x02\x00\xBA\x6D"
-     "\x02\x00"
-     "\x83\x69\x00\x80\x87\x69\x00\x80\xBF\xC1\x02\x00\xBE\xD9\x02\x00\xBD\xDD"
-     "\x02\x00"
-     "\xBC\xDD\x02\x00\xB3\x3D\x02\x00\x8B\x69\x00\x80\x8F\x69\x00\x80\x93\x69"
-     "\x00\x80"
-     "\x84\xFC\x06\x00\xB6\x79\x02\x00\xB5\x71\x02\x00\x97\x69\x00\x80\xAA\x29"
-     "\x02\x00"
-     "\xAB\x29\x02\x00\x9B\x69\x00\x80\x9F\x69\x00\x80\xAE\x9D\x02\x00\xAF\x85"
-     "\x02\x00"
-     "\xAC\x99\x02\x00\xAD\x99\x02\x00\xA3\x69\x00\x80\xA3\x79\x02\x00\xA7\x69"
-     "\x00\x80"
-     "\xAB\x69\x00\x80\xA6\x3D\x02\x00\xAF\x69\x00\x80\xB3\x69\x00\x80\xA5\x35"
-     "\x02\x00"
-     "\x82\x6D\x27\x00\x83\x8D\x2A\x00\x86\xA8\x05\x00\x87\x6C\x03\x00\x86\x99"
-     "\x2E\x00"
-     "\x87\xCD\x2F\x00\x84\x11\x2E\x00\x85\x99\x2E\x00\x8A\x21\x12\x00\x8B\x29"
-     "\x12\x00"
-     "\xB7\x69\x00\x80\xBB\x69\x00\x80\x8E\x91\x16\x00\x8F\x1D\x16\x00\x8C\x11"
-     "\x12\x00"
-     "\x8D\x11\x16\x00\x92\x71\x1A\x00\x93\xE5\x1A\x00\xBF\x69\x00\x80\xEF\x94"
-     "\x76\x00"
-     "\x96\xF1\x1E\x00\x97\xE5\x1E\x00\x94\x49\x1A\x00\x95\x19\x1E\x00\x9A\x29"
-     "\x02\x00"
-     "\x9B\xE1\x02\x00\xC7\x69\x00\x80\xCB\x69\x00\x80\xCF\x69\x00\x80\xE1\x20"
-     "\x12\x00"
-     "\x9C\xF1\x02\x00\xE3\x20\x16\x00\x9F\x21\x1F\x00\x9E\xC1\x1F\x00\x9D\x99"
-     "\x1B\x00"
-     "\x9C\x2D\x1B\x00\x9B\xA1\x1B\x00\x9A\xBD\x17\x00\x99\x39\x17\x00\x98\xB1"
-     "\x17\x00"
-     "\x97\x89\x13\x00\x96\xA9\x13\x00\x95\xA9\x13\x00\x94\x75\x2E\x00\x93\x39"
-     "\x2F\x00"
-     "\x92\x31\x2F\x00\x91\xB1\x2F\x00\x90\x35\x2B\x00\x8F\xAD\x26\x00\xE3\x78"
-     "\x1F\x00"
-     "\xD2\x00\x00\x00\xE1\x5C\x1F\x00\x82\x99\x01\x00\xD3\x69\x00\x80\x80\xF1"
-     "\x01\x00"
-     "\x81\xF1\x01\x00\xBE\xA8\x07\x00\xD7\x69\x00\x80\xDB\x69\x00\x80\xDF\x69"
-     "\x00\x80"
-     "\x84\xBC\x06\x00\xEF\x2C\x1F\x00\xE3\x69\x00\x80\xE7\x69\x00\x80\xE1\xA4"
-     "\x1E\x00"
-     "\xE3\xCC\x00\x00\xE3\x7C\x1E\x00\xE1\xBC\x01\x00\xEB\x69\x00\x80\xEF\x69"
-     "\x00\x80"
-     "\xF3\x69\x00\x80\x84\x9C\x06\x00\xF7\x69\x00\x80\xBE\x6C\x06\x00\xFB\x69"
-     "\x00\x80"
-     "\xFF\x69\x00\x80\x03\x6A\x00\x80\xEF\x10\x00\x00\xEF\x04\x1E\x00\x07\x6A"
-     "\x00\x80"
-     "\x0B\x6A\x00\x80\x0F\x6A\x00\x80\x13\x6A\x00\x80\x17\x6A\x00\x80\x1B\x6A"
-     "\x00\x80"
-     "\x1F\x6A\x00\x80\x23\x6A\x00\x80\x27\x6A\x00\x80\x80\x3D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\xC9\x07\x00\x2B\x6A\x00\x80\xA3\xF9\x03\x00\xA2\xF1\x03\x00\xA1\x31"
-     "\x03\x00"
-     "\xA0\xCD\x1F\x00\xB0\x49\x71\x00\xB1\x01\x7C\x00\xB2\x09\x7C\x00\xB3\x21"
-     "\x7D\x00"
-     "\xB4\x01\x78\x00\xB5\x15\x78\x00\x53\x69\x00\x80\xC3\x69\x00\x80\x2F\x6A"
-     "\x00\x80"
-     "\xBE\x28\x0E\x00\x86\x0C\x00\x00\x87\x8C\x03\x00\x33\x6A\x00\x80\x37\x6A"
-     "\x00\x80"
-     "\x3B\x6A\x00\x80\x3F\x6A\x00\x80\x43\x6A\x00\x80\xA1\x5D\x00\x00\xA2\x55"
-     "\x00\x00"
-     "\xA3\x7D\x00\x00\xA4\x01\x0C\x00\xA5\x15\x0C\x00\xA6\xF5\x0C\x00\xA7\x01"
-     "\x08\x00"
-     "\xA8\x21\x08\x00\xA9\xC5\x08\x00\xAA\x01\x74\x00\xAB\x09\x74\x00\xAC\x01"
-     "\x74\x00"
-     "\xAD\x1D\x75\x00\xAE\x79\x70\x00\xAF\x79\x70\x00\xA8\xE5\x05\x00\xA9\xF1"
-     "\x05\x00"
-     "\xAA\xF1\x05\x00\xAB\x2D\x05\x00\xAC\x35\x05\x00\xAD\x3D\x05\x00\xAE\x35"
-     "\x05\x00"
-     "\xAF\x2D\x05\x00\x47\x6A\x00\x80\x4B\x6A\x00\x80\x4F\x6A\x00\x80\x53\x6A"
-     "\x00\x80"
-     "\x57\x6A\x00\x80\x5B\x6A\x00\x80\x5F\x6A\x00\x80\x63\x6A\x00\x80\xB8\xFD"
-     "\x06\x00"
-     "\xB9\x8D\x06\x00\xBA\x85\x06\x00\xBB\xAD\x06\x00\xBC\xB9\x06\x00\xBD\xB9"
-     "\x06\x00"
-     "\xBE\xAD\x06\x00\xBF\xA5\x06\x00\xB0\x55\x05\x00\xB1\x5D\x05\x00\xB2\x55"
-     "\x05\x00"
-     "\xB3\xE5\x06\x00\xB4\xFD\x06\x00\xB5\xD1\x06\x00\xB6\xD1\x06\x00\xB7\xD1"
-     "\x06\x00"
-     "\xB3\x79\x04\x00\x67\x6A\x00\x80\x6B\x6A\x00\x80\x6F\x6A\x00\x80\x73\x6A"
-     "\x00\x80"
-     "\xB6\xBD\x04\x00\xB5\xBD\x04\x00\x77\x6A\x00\x80\xBB\x99\x04\x00\xBA\x91"
-     "\x04\x00"
-     "\x7B\x6A\x00\x80\x7F\x6A\x00\x80\xBF\x15\x07\x00\xBE\x39\x07\x00\xBD\x31"
-     "\x07\x00"
-     "\xBC\x81\x04\x00\x83\x6A\x00\x80\xA3\x3D\x04\x00\x87\x6A\x00\x80\x8B\x6A"
-     "\x00\x80"
-     "\xA6\xF9\x04\x00\x8F\x6A\x00\x80\x93\x6A\x00\x80\xA5\xF9\x04\x00\xAA\xD5"
-     "\x04\x00"
-     "\xAB\xDD\x04\x00\x97\x6A\x00\x80\x9B\x6A\x00\x80\xAE\x7D\x07\x00\xAF\x51"
-     "\x07\x00"
-     "\xAC\xC5\x04\x00\xAD\x75\x07\x00\xA8\x69\x07\x00\xA9\x69\x07\x00\xAA\x79"
-     "\x07\x00"
-     "\xAB\xD9\x06\x00\xAC\xF5\x06\x00\xAD\xFD\x06\x00\xAE\xF5\x06\x00\xAF\xE5"
-     "\x06\x00"
-     "\x80\xC9\x00\x00\x81\xC9\x00\x00\x82\x05\x00\x00\x9F\x6A\x00\x80\x86\x70"
-     "\x0F\x00"
-     "\x87\x34\x00\x00\xA7\x6A\x00\x80\xAB\x6A\x00\x80\xB8\x7D\x06\x00\xB9\x05"
-     "\x06\x00"
-     "\xBA\x0D\x06\x00\xBB\x05\x06\x00\xBC\x1D\x06\x00\xBD\x05\x06\x00\xBE\x0D"
-     "\x06\x00"
-     "\xBF\x05\x06\x00\xB0\x9D\x06\x00\xB1\x75\x06\x00\xB2\x7D\x06\x00\xB3\x75"
-     "\x06\x00"
-     "\xB4\x51\x06\x00\xB5\x5D\x06\x00\xB6\x55\x06\x00\xB7\x4D\x06\x00\xB3\xF1"
-     "\x04\x00"
-     "\xAF\x6A\x00\x80\xB3\x6A\x00\x80\xB7\x6A\x00\x80\xBB\x6A\x00\x80\xB6\x7D"
-     "\x05\x00"
-     "\xB5\x7D\x05\x00\xBF\x6A\x00\x80\xBB\xB1\x05\x00\xBA\x59\x05\x00\xC3\x6A"
-     "\x00\x80"
-     "\xC7\x6A\x00\x80\xBF\x99\x05\x00\xBE\x95\x05\x00\xBD\xA1\x05\x00\xBC\xA9"
-     "\x05\x00"
-     "\xCB\x6A\x00\x80\xA3\xB5\x04\x00\xCF\x6A\x00\x80\xD3\x6A\x00\x80\xA6\x39"
-     "\x05\x00"
-     "\xD7\x6A\x00\x80\xDB\x6A\x00\x80\xA5\x39\x05\x00\xAA\x1D\x05\x00\xAB\xF5"
-     "\x05\x00"
-     "\xDF\x6A\x00\x80\xE3\x6A\x00\x80\xAE\xD1\x05\x00\xAF\xDD\x05\x00\xAC\xED"
-     "\x05\x00"
-     "\xAD\xE5\x05\x00\xA9\xB9\x02\x00\xA8\xB1\x02\x00\xAB\xC9\x02\x00\xAA\xB1"
-     "\x02\x00"
-     "\xAD\x35\x02\x00\xAC\x35\x02\x00\xAF\x35\x02\x00\xAE\x35\x02\x00\xE7\x6A"
-     "\x00\x80"
-     "\xEB\x6A\x00\x80\xEF\x6A\x00\x80\xF3\x6A\x00\x80\xF7\x6A\x00\x80\xFB\x6A"
-     "\x00\x80"
-     "\xFF\x6A\x00\x80\x03\x6B\x00\x80\xB9\xE9\x03\x00\xB8\xD9\x03\x00\xBB\x89"
-     "\x03\x00"
-     "\xBA\xE1\x03\x00\xBD\x9D\x03\x00\xBC\x9D\x03\x00\xBF\x81\x03\x00\xBE\x89"
-     "\x03\x00"
-     "\xB1\x55\x02\x00\xB0\x55\x02\x00\xB3\x55\x02\x00\xB2\x55\x02\x00\xB5\xF9"
-     "\x03\x00"
-     "\xB4\xF1\x03\x00\xB7\xE9\x03\x00\xB6\xF1\x03\x00\xB3\x3D\x03\x00\x07\x6B"
-     "\x00\x80"
-     "\x0B\x6B\x00\x80\x0F\x6B\x00\x80\x17\x6B\x00\x80\xB6\xB1\x03\x00\xB5\xB9"
-     "\x03\x00"
-     "\x1B\x6B\x00\x80\xBB\x95\x03\x00\xBA\x95\x03\x00\x86\x88\x0C\x00\x87\xA0"
-     "\x0D\x00"
-     "\xBF\x39\x02\x00\xBE\x31\x02\x00\xBD\x85\x03\x00\xBC\x85\x03\x00\x1F\x6B"
-     "\x00\x80"
-     "\xA3\x79\x03\x00\x23\x6B\x00\x80\x27\x6B\x00\x80\xA6\xF5\x03\x00\x2B\x6B"
-     "\x00\x80"
-     "\x2F\x6B\x00\x80\xA5\xFD\x03\x00\xAA\xD1\x03\x00\xAB\xD1\x03\x00\x33\x6B"
-     "\x00\x80"
-     "\x37\x6B\x00\x80\xAE\x75\x02\x00\xAF\x7D\x02\x00\xAC\xC1\x03\x00\xAD\xC1"
-     "\x03\x00"
-     "\x80\x85\x00\x00\x81\x8D\x00\x00\x82\x85\x00\x00\xEF\xD0\x06\x00\xE3\xB0"
-     "\x07\x00"
-     "\xE3\xF4\x04\x00\xE1\x08\x07\x00\xE1\xEC\x04\x00\xEF\x38\x06\x00\xEF\x20"
-     "\x04\x00"
-     "\xBE\xA4\x0C\x00\x3B\x6B\x00\x80\x3F\x6B\x00\x80\xE1\x84\x01\x00\x43\x6B"
-     "\x00\x80"
-     "\xE3\xDC\x06\x00\x47\x6B\x00\x80\x4B\x6B\x00\x80\x84\xC0\x0D\x00\xB3\xC9"
-     "\x01\x00"
-     "\x4F\x6B\x00\x80\xB5\xD9\x01\x00\xB6\xC9\x01\x00\x53\x6B\x00\x80\x57\x6B"
-     "\x00\x80"
-     "\x5B\x6B\x00\x80\xBA\xC5\x01\x00\xBB\xAD\x01\x00\xBC\xB9\x01\x00\xBD\xB9"
-     "\x01\x00"
-     "\xBE\xAD\x01\x00\xBF\xA5\x01\x00\xB0\x2D\x0E\x00\xB1\x45\x0E\x00\xB2\x41"
-     "\x0E\x00"
-     "\xB3\x41\x0E\x00\xB4\x45\x0E\x00\xB5\x4D\x0E\x00\xB6\x71\x0E\x00\xB7\x71"
-     "\x0E\x00"
-     "\xB8\x81\x0E\x00\xB9\x81\x0E\x00\xBA\x81\x0E\x00\xBB\x81\x0E\x00\xBC\x81"
-     "\x0E\x00"
-     "\xBD\x81\x0E\x00\xBE\x81\x0E\x00\xBF\x81\x0E\x00\x13\x6B\x00\x80\x5F\x6B"
-     "\x00\x80"
-     "\x63\x6B\x00\x80\x67\x6B\x00\x80\xA3\x6A\x00\x80\x6B\x6B\x00\x80\x6F\x6B"
-     "\x00\x80"
-     "\x73\x6B\x00\x80\xA8\xD9\x0D\x00\xA9\xD9\x0D\x00\xAA\x6D\x0E\x00\xAB\x65"
-     "\x0E\x00"
-     "\xAC\x7D\x0E\x00\xAD\x65\x0E\x00\xAE\x65\x0E\x00\xAF\x55\x0E\x00\xA3\x85"
-     "\x0E\x00"
-     "\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D\x00\x00\x77\x6B\x00\x80\xA6\x85"
-     "\x0E\x00"
-     "\xA5\x95\x0E\x00\x7B\x6B\x00\x80\xAB\xE1\x0E\x00\xAA\x89\x0E\x00\x7F\x6B"
-     "\x00\x80"
-     "\x83\x6B\x00\x80\xAF\xE9\x0E\x00\xAE\xE1\x0E\x00\xAD\xF5\x0E\x00\xAC\xF5"
-     "\x0E\x00"
-     "\x87\x6B\x00\x80\xB3\x85\x0F\x00\x86\x68\x00\x00\x87\x1C\x03\x00\xB6\x85"
-     "\x0F\x00"
-     "\x8B\x6B\x00\x80\x8F\x6B\x00\x80\xB5\x91\x0F\x00\xBA\x8D\x0F\x00\xBB\x49"
-     "\x0F\x00"
-     "\x93\x6B\x00\x80\x97\x6B\x00\x80\xBE\x31\x0F\x00\xBF\x31\x0F\x00\xBC\x49"
-     "\x0F\x00"
-     "\xBD\x45\x0F\x00\xA8\x11\x0E\x00\xA9\x19\x0E\x00\xAA\x49\x0E\x00\xAB\x45"
-     "\x0E\x00"
-     "\xAC\x5D\x0E\x00\xAD\x41\x0E\x00\xAE\x41\x0E\x00\xAF\x75\x0E\x00\x9B\x6B"
-     "\x00\x80"
-     "\x9F\x6B\x00\x80\xA3\x6B\x00\x80\xA7\x6B\x00\x80\xAB\x6B\x00\x80\xAF\x6B"
-     "\x00\x80"
-     "\xB3\x6B\x00\x80\xB7\x6B\x00\x80\xB8\xA1\x0E\x00\xB9\xA1\x0E\x00\xBA\x0D"
-     "\x01\x00"
-     "\xBB\x05\x01\x00\xBC\x1D\x01\x00\xBD\x01\x01\x00\xBE\x01\x01\x00\xBF\x01"
-     "\x01\x00"
-     "\xB0\x0D\x0E\x00\xB1\xC9\x0E\x00\xB2\xD9\x0E\x00\xB3\xD5\x0E\x00\xB4\xB1"
-     "\x0E\x00"
-     "\xB5\xB1\x0E\x00\xB6\xA9\x0E\x00\xB7\xA1\x0E\x00\xA3\xC1\x0E\x00\xBB\x6B"
-     "\x00\x80"
-     "\xBF\x6B\x00\x80\xC3\x6B\x00\x80\xC7\x6B\x00\x80\xA6\xC1\x0E\x00\xA5\xD5"
-     "\x0E\x00"
-     "\xCB\x6B\x00\x80\xAB\x0D\x0E\x00\xAA\xC9\x0E\x00\xCF\x6B\x00\x80\xD3\x6B"
-     "\x00\x80"
-     "\xAF\x75\x0E\x00\xAE\x75\x0E\x00\xAD\x01\x0E\x00\xAC\x0D\x0E\x00\xD7\x6B"
-     "\x00\x80"
-     "\xDB\x6B\x00\x80\xDF\x6B\x00\x80\xE3\x6B\x00\x80\x80\x0D\x00\x00\x81\x35"
-     "\x00\x00"
-     "\x82\x3D\x00\x00\xE7\x6B\x00\x80\xEB\x6B\x00\x80\xEF\x6B\x00\x80\x84\x84"
-     "\x01\x00"
-     "\xBE\x84\x01\x00\x86\x30\x07\x00\x87\xF8\x00\x00\xF7\x6B\x00\x80\xFB\x6B"
-     "\x00\x80"
-     "\xA8\xC1\x02\x00\xA9\xD1\x02\x00\xAA\xD1\x02\x00\xAB\xE5\x02\x00\xAC\xFD"
-     "\x02\x00"
-     "\xAD\x35\x03\x00\xAE\x3D\x03\x00\xAF\x35\x03\x00\xFF\x6B\x00\x80\x03\x6C"
-     "\x00\x80"
-     "\x07\x6C\x00\x80\x0B\x6C\x00\x80\x0F\x6C\x00\x80\x13\x6C\x00\x80\x17\x6C"
-     "\x00\x80"
-     "\x1B\x6C\x00\x80\xB8\xD1\x03\x00\xB9\xD9\x03\x00\xBA\xE1\x03\x00\xBB\xE1"
-     "\x03\x00"
-     "\xBC\x91\x03\x00\xBD\x91\x03\x00\xBE\x91\x03\x00\xBF\x91\x03\x00\xB0\x4D"
-     "\x03\x00"
-     "\xB1\x55\x03\x00\xB2\x5D\x03\x00\xB3\x55\x03\x00\xB4\x4D\x03\x00\xB5\xF1"
-     "\x03\x00"
-     "\xB6\xF1\x03\x00\xB7\xF1\x03\x00\xBB\xB1\x03\x00\xBA\xA9\x03\x00\x1F\x6C"
-     "\x00\x80"
-     "\xBE\x84\x03\x00\xBF\x15\x03\x00\xBE\xA9\x03\x00\xBD\xA1\x03\x00\xBC\xA9"
-     "\x03\x00"
-     "\xB3\x79\x02\x00\x23\x6C\x00\x80\x27\x6C\x00\x80\x2B\x6C\x00\x80\x2F\x6C"
-     "\x00\x80"
-     "\xB6\x95\x03\x00\xB5\x55\x02\x00\x33\x6C\x00\x80\xAA\xED\x03\x00\xAB\xF5"
-     "\x03\x00"
-     "\x37\x6C\x00\x80\x3B\x6C\x00\x80\xAE\xED\x03\x00\xAF\x51\x03\x00\xAC\xED"
-     "\x03\x00"
-     "\xAD\xE5\x03\x00\x3F\x6C\x00\x80\xA3\x3D\x02\x00\x43\x6C\x00\x80\x47\x6C"
-     "\x00\x80"
-     "\xA6\xD1\x03\x00\x4B\x6C\x00\x80\x4F\x6C\x00\x80\xA5\x11\x02\x00\x53\x6C"
-     "\x00\x80"
-     "\x82\x21\x00\x00\x81\x15\x00\x00\x80\x15\x00\x00\xEF\x04\x00\x00\x84\x94"
-     "\x02\x00"
-     "\x57\x6C\x00\x80\x5B\x6C\x00\x80\xE3\xD8\x00\x00\x5F\x6C\x00\x80\xE1\xF8"
-     "\x01\x00"
-     "\x67\x6C\x00\x80\x6B\x6C\x00\x80\x6F\x6C\x00\x80\x73\x6C\x00\x80\x77\x6C"
-     "\x00\x80"
-     "\x86\x60\x05\x00\x87\x08\x05\x00\x7B\x6C\x00\x80\xB3\x6D\x01\x00\x7F\x6C"
-     "\x00\x80"
-     "\xB5\x7D\x01\x00\xB6\x75\x01\x00\x83\x6C\x00\x80\x87\x6C\x00\x80\x8B\x6C"
-     "\x00\x80"
-     "\xBA\x51\x01\x00\xBB\x51\x01\x00\xBC\xF9\x01\x00\xBD\xD1\x01\x00\xBE\xD1"
-     "\x01\x00"
-     "\xBF\xD1\x01\x00\x8F\x6C\x00\x80\xA3\xA5\x01\x00\x93\x6C\x00\x80\x97\x6C"
-     "\x00\x80"
-     "\xA6\xBD\x01\x00\x9B\x6C\x00\x80\x9F\x6C\x00\x80\xA5\xB5\x01\x00\xAA\x99"
-     "\x01\x00"
-     "\xAB\x99\x01\x00\xA3\x6C\x00\x80\xA7\x6C\x00\x80\xAE\x19\x01\x00\xAF\x19"
-     "\x01\x00"
-     "\xAC\x31\x01\x00\xAD\x19\x01\x00\xAB\x6C\x00\x80\xE1\xF4\x0E\x00\xAF\x6C"
-     "\x00\x80"
-     "\xE3\x14\x0E\x00\xF4\x00\x00\x00\xE1\x7C\x0C\x00\xB3\x6C\x00\x80\xE3\x50"
-     "\x0A\x00"
-     "\xB7\x6C\x00\x80\xBB\x6C\x00\x80\xBE\x20\x04\x00\xEF\x30\x0D\x00\xBF\x6C"
-     "\x00\x80"
-     "\xC3\x6C\x00\x80\x84\x34\x04\x00\xEF\xAC\x0E\x00\xB0\x39\x06\x00\xB1\x39"
-     "\x06\x00"
-     "\xB2\x81\x06\x00\xB3\xA9\x06\x00\xB4\xB9\x06\x00\xB5\xB9\x06\x00\xB6\xA9"
-     "\x06\x00"
-     "\xB7\xA1\x06\x00\xB8\xE9\x06\x00\xB9\xE9\x06\x00\xBA\xC9\x06\x00\xBB\xC5"
-     "\x06\x00"
-     "\xBC\xDD\x06\x00\xBD\xC1\x06\x00\xBE\xC1\x06\x00\xBF\x3D\x07\x00\x63\x6C"
-     "\x00\x80"
-     "\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\xC7\x6C\x00\x80\xCB\x6C"
-     "\x00\x80"
-     "\xCF\x6C\x00\x80\xF3\x6B\x00\x80\xA8\x99\x05\x00\xA9\x99\x05\x00\xAA\x49"
-     "\x06\x00"
-     "\xAB\x49\x06\x00\xAC\x59\x06\x00\xAD\x59\x06\x00\xAE\x49\x06\x00\xAF\x49"
-     "\x06\x00"
-     "\xA8\xAD\x07\x00\xA9\xB5\x07\x00\xAA\xBD\x07\x00\xAB\xB5\x07\x00\xAC\xAD"
-     "\x07\x00"
-     "\xAD\xDD\x07\x00\xAE\xC9\x07\x00\xAF\xC1\x07\x00\xD3\x6C\x00\x80\xD7\x6C"
-     "\x00\x80"
-     "\x86\x88\x03\x00\x87\x10\x03\x00\xDB\x6C\x00\x80\xDF\x6C\x00\x80\xE3\x6C"
-     "\x00\x80"
-     "\xE7\x6C\x00\x80\xB8\x6D\x07\x00\xB9\x05\x07\x00\xBA\x01\x07\x00\xBB\x15"
-     "\x07\x00"
-     "\xBC\x31\x07\x00\xBD\x31\x07\x00\xBE\x29\x07\x00\xBF\x29\x07\x00\xB0\x81"
-     "\x07\x00"
-     "\xB1\x81\x07\x00\xB2\x69\x07\x00\xB3\x65\x07\x00\xB4\x7D\x07\x00\xB5\x61"
-     "\x07\x00"
-     "\xB6\x61\x07\x00\xB7\x55\x07\x00\xB3\x35\x06\x00\xEB\x6C\x00\x80\xEF\x6C"
-     "\x00\x80"
-     "\xF3\x6C\x00\x80\xF7\x6C\x00\x80\xB6\x5D\x06\x00\xB5\x25\x06\x00\xFB\x6C"
-     "\x00\x80"
-     "\xBB\x45\x06\x00\xBA\x45\x06\x00\xFF\x6C\x00\x80\x03\x6D\x00\x80\xBF\xA5"
-     "\x06\x00"
-     "\xBE\xB9\x06\x00\xBD\xB1\x06\x00\xBC\xBD\x06\x00\x07\x6D\x00\x80\xA3\x71"
-     "\x06\x00"
-     "\x0B\x6D\x00\x80\x0F\x6D\x00\x80\xA6\x19\x06\x00\x13\x6D\x00\x80\x17\x6D"
-     "\x00\x80"
-     "\xA5\x61\x06\x00\xAA\x01\x06\x00\xAB\x01\x06\x00\x1B\x6D\x00\x80\x1F\x6D"
-     "\x00\x80"
-     "\xAE\xFD\x06\x00\xAF\xE1\x06\x00\xAC\xF9\x06\x00\xAD\xF5\x06\x00\xA8\x25"
-     "\x01\x00"
-     "\xA9\x35\x01\x00\xAA\x3D\x01\x00\xAB\x35\x01\x00\xAC\x2D\x01\x00\xAD\x91"
-     "\x00\x00"
-     "\xAE\x91\x00\x00\xAF\x91\x00\x00\x23\x6D\x00\x80\x27\x6D\x00\x80\x2B\x6D"
-     "\x00\x80"
-     "\x2F\x6D\x00\x80\x33\x6D\x00\x80\x82\xBD\x03\x00\x81\xBD\x03\x00\x80\xBD"
-     "\x03\x00"
-     "\xB8\x99\x00\x00\xB9\xAD\x00\x00\xBA\xA5\x00\x00\xBB\x6D\x00\x00\xBC\x75"
-     "\x00\x00"
-     "\xBD\x7D\x00\x00\xBE\x75\x00\x00\xBF\x6D\x00\x00\xB0\xF1\x00\x00\xB1\xF9"
-     "\x00\x00"
-     "\xB2\xC1\x00\x00\xB3\xC1\x00\x00\xB4\xB1\x00\x00\xB5\xBD\x00\x00\xB6\xB5"
-     "\x00\x00"
-     "\xB7\xA9\x00\x00\x37\x6D\x00\x80\x3B\x6D\x00\x80\x3F\x6D\x00\x80\x84\x80"
-     "\x02\x00"
-     "\xBE\x1C\x02\x00\x47\x6D\x00\x80\x86\xF8\x7C\x00\x87\xCC\x02\x00\x84\xAC"
-     "\x03\x00"
-     "\x4B\x6D\x00\x80\x4F\x6D\x00\x80\x53\x6D\x00\x80\x57\x6D\x00\x80\x5B\x6D"
-     "\x00\x80"
-     "\x5F\x6D\x00\x80\x63\x6D\x00\x80\xB3\xF5\x02\x00\x67\x6D\x00\x80\x6B\x6D"
-     "\x00\x80"
-     "\x92\x00\x00\x00\x6F\x6D\x00\x80\xB6\x51\x03\x00\xB5\xE5\x02\x00\x73\x6D"
-     "\x00\x80"
-     "\xBB\x7D\x03\x00\xBA\x75\x03\x00\x77\x6D\x00\x80\x7B\x6D\x00\x80\xBF\x39"
-     "\x03\x00"
-     "\xBE\x35\x03\x00\xBD\x55\x03\x00\xBC\x55\x03\x00\xA3\x35\x02\x00\x7F\x6D"
-     "\x00\x80"
-     "\x83\x6D\x00\x80\x87\x6D\x00\x80\x8B\x6D\x00\x80\xA6\x91\x03\x00\xA5\x25"
-     "\x02\x00"
-     "\x8F\x6D\x00\x80\xAB\xBD\x03\x00\xAA\xB5\x03\x00\x93\x6D\x00\x80\x97\x6D"
-     "\x00\x80"
-     "\xAF\xF9\x03\x00\xAE\xF5\x03\x00\xAD\x95\x03\x00\xAC\x95\x03\x00\xBE\xC0"
-     "\x03\x00"
-     "\x9B\x6D\x00\x80\x9F\x6D\x00\x80\xA3\x6D\x00\x80\x80\x0D\x00\x00\x81\x35"
-     "\x00\x00"
-     "\x82\x3D\x00\x00\xA7\x6D\x00\x80\xAB\x6D\x00\x80\xAF\x6D\x00\x80\x86\xC8"
-     "\x7C\x00"
-     "\x87\x00\x03\x00\xB7\x6D\x00\x80\xBB\x6D\x00\x80\xBF\x6D\x00\x80\xC3\x6D"
-     "\x00\x80"
-     "\xC7\x6D\x00\x80\xCB\x6D\x00\x80\xCF\x6D\x00\x80\xD3\x6D\x00\x80\xD7\x6D"
-     "\x00\x80"
-     "\xDB\x6D\x00\x80\xDF\x6D\x00\x80\xEF\xF4\x03\x00\x84\xEC\x7C\x00\xE1\x94"
-     "\x01\x00"
-     "\xE3\x6D\x00\x80\xE3\x30\x03\x00\xE7\x6D\x00\x80\xEB\x6D\x00\x80\xEF\x6D"
-     "\x00\x80"
-     "\xF3\x6D\x00\x80\xB3\x69\x01\x00\xF7\x6D\x00\x80\xFB\x6D\x00\x80\xFF\x6D"
-     "\x00\x80"
-     "\x03\x6E\x00\x80\xB6\x61\x01\x00\xB5\x69\x01\x00\x07\x6E\x00\x80\xBB\x29"
-     "\x01\x00"
-     "\xBA\x21\x01\x00\x0B\x6E\x00\x80\x0F\x6E\x00\x80\xBF\x1D\x01\x00\xBE\x1D"
-     "\x01\x00"
-     "\xBD\x25\x01\x00\xBC\x2D\x01\x00\x13\x6E\x00\x80\x17\x6E\x00\x80\x1B\x6E"
-     "\x00\x80"
-     "\xA3\xA5\x01\x00\x1F\x6E\x00\x80\xA5\xA5\x01\x00\xA6\xAD\x01\x00\xBE\x54"
-     "\x7C\x00"
-     "\x86\x80\x7C\x00\x87\xEC\x7C\x00\xAA\xED\x01\x00\xAB\xE5\x01\x00\xAC\xE1"
-     "\x01\x00"
-     "\xAD\xE9\x01\x00\xAE\xD1\x01\x00\xAF\xD1\x01\x00\x27\x6E\x00\x80\xE1\x9C"
-     "\x06\x00"
-     "\x84\x04\x7F\x00\xE3\x24\x06\x00\xE3\xD4\x06\x00\x2B\x6E\x00\x80\xE1\x30"
-     "\x04\x00"
-     "\x2F\x6E\x00\x80\xEF\x94\x07\x00\x82\x75\x00\x00\x81\x61\x00\x00\x80\x69"
-     "\x00\x00"
-     "\x33\x6E\x00\x80\x37\x6E\x00\x80\x3B\x6E\x00\x80\xEF\xEC\x06\x00\xB8\x8D"
-     "\x7E\x00"
-     "\xB9\x95\x7E\x00\xBA\x95\x7E\x00\xBB\xA5\x7E\x00\xBC\xBD\x7E\x00\xBD\xD1"
-     "\x7E\x00"
-     "\xBE\xD1\x7E\x00\xBF\xD1\x7E\x00\xB0\x65\x7E\x00\xB1\x6D\x7E\x00\xB2\x79"
-     "\x7E\x00"
-     "\xB3\x71\x7E\x00\xB4\x59\x7E\x00\xB5\x59\x7E\x00\xB6\xBD\x7E\x00\xB7\xB5"
-     "\x7E\x00"
-     "\xA8\x55\x7E\x00\xA9\x61\x7E\x00\xAA\x61\x7E\x00\xAB\x61\x7E\x00\xAC\x61"
-     "\x7E\x00"
-     "\xAD\x61\x7E\x00\xAE\x61\x7E\x00\xAF\x61\x7E\x00\x3F\x6E\x00\x80\xB3\x6D"
-     "\x00\x80"
-     "\x43\x6D\x00\x80\x43\x6E\x00\x80\x23\x6E\x00\x80\x47\x6E\x00\x80\x4B\x6E"
-     "\x00\x80"
-     "\x4F\x6E\x00\x80\xA8\x71\x7E\x00\xA9\x71\x7E\x00\xAA\x71\x7E\x00\xAB\x71"
-     "\x7E\x00"
-     "\xAC\x91\x7F\x00\xAD\x91\x7F\x00\xAE\x91\x7F\x00\xAF\x91\x7F\x00\x53\x6E"
-     "\x00\x80"
-     "\x57\x6E\x00\x80\x5B\x6E\x00\x80\x5F\x6E\x00\x80\x63\x6E\x00\x80\x67\x6E"
-     "\x00\x80"
-     "\x6B\x6E\x00\x80\x6F\x6E\x00\x80\xB8\x85\x7F\x00\xB9\x8D\x7F\x00\xBA\x85"
-     "\x7F\x00"
-     "\xBB\x9D\x7F\x00\xBC\x8D\x7F\x00\xBD\xBD\x7F\x00\xBE\xB5\x7F\x00\xBF\x5D"
-     "\x7F\x00"
-     "\xB0\xF1\x7F\x00\xB1\xF1\x7F\x00\xB2\xF1\x7F\x00\xB3\xC5\x7F\x00\xB4\xC1"
-     "\x7F\x00"
-     "\xB5\xC1\x7F\x00\xB6\xC1\x7F\x00\xB7\xC1\x7F\x00\xB3\xF9\x7F\x00\x73\x6E"
-     "\x00\x80"
-     "\x77\x6E\x00\x80\x7B\x6E\x00\x80\x7F\x6E\x00\x80\xB6\x11\x7E\x00\xB5\x19"
-     "\x7E\x00"
-     "\x83\x6E\x00\x80\xBB\x35\x7E\x00\xBA\x35\x7E\x00\x87\x6E\x00\x80\x8B\x6E"
-     "\x00\x80"
-     "\xBF\x05\x7E\x00\xBE\x05\x7E\x00\xBD\x11\x7E\x00\xBC\x25\x7E\x00\x82\x15"
-     "\x00\x00"
-     "\xA3\xBD\x7F\x00\x80\x61\x00\x00\x81\x61\x00\x00\xA6\x55\x7E\x00\x8F\x6E"
-     "\x00\x80"
-     "\xBE\x90\x01\x00\xA5\x5D\x7E\x00\xAA\x71\x7E\x00\xAB\x71\x7E\x00\x93\x6E"
-     "\x00\x80"
-     "\x97\x6E\x00\x80\xAE\x41\x7E\x00\xAF\x41\x7E\x00\xAC\x61\x7E\x00\xAD\x55"
-     "\x7E\x00"
-     "\xA8\x41\x7E\x00\xA9\x51\x7E\x00\xAA\x55\x7E\x00\xAB\x7D\x7E\x00\xAC\x65"
-     "\x7E\x00"
-     "\xAD\x6D\x7E\x00\xAE\xF9\x01\x00\xAF\xF1\x01\x00\x86\x00\x00\x00\x87\x34"
-     "\x01\x00"
-     "\x9B\x6E\x00\x80\x9F\x6E\x00\x80\xA3\x6E\x00\x80\xA7\x6E\x00\x80\xAB\x6E"
-     "\x00\x80"
-     "\xAF\x6E\x00\x80\xB8\x75\x01\x00\xB9\x7D\x01\x00\xBA\x75\x01\x00\xBB\xC9"
-     "\x00\x00"
-     "\xBC\xD9\x00\x00\xBD\xD9\x00\x00\xBE\xC9\x00\x00\xBF\xC1\x00\x00\xB0\x95"
-     "\x01\x00"
-     "\xB1\x9D\x01\x00\xB2\x95\x01\x00\xB3\x4D\x01\x00\xB4\x55\x01\x00\xB5\x5D"
-     "\x01\x00"
-     "\xB6\x55\x01\x00\xB7\x4D\x01\x00\xB3\xDD\x7D\x00\xB3\x6E\x00\x80\xB7\x6E"
-     "\x00\x80"
-     "\xBB\x6E\x00\x80\xBF\x6E\x00\x80\xB6\xED\x7D\x00\xB5\xED\x7D\x00\xC3\x6E"
-     "\x00\x80"
-     "\xBB\x59\x02\x00\xBA\x51\x02\x00\xC7\x6E\x00\x80\xCB\x6E\x00\x80\xBF\x99"
-     "\x02\x00"
-     "\xBE\x91\x02\x00\xBD\x99\x02\x00\xBC\x41\x02\x00\xCF\x6E\x00\x80\xA3\x99"
-     "\x7D\x00"
-     "\xD3\x6E\x00\x80\xD7\x6E\x00\x80\xA6\xA9\x7D\x00\xDB\x6E\x00\x80\xDF\x6E"
-     "\x00\x80"
-     "\xA5\xA9\x7D\x00\xAA\x15\x02\x00\xAB\x1D\x02\x00\xE3\x6E\x00\x80\xE7\x6E"
-     "\x00\x80"
-     "\xAE\xD5\x02\x00\xAF\xDD\x02\x00\xAC\x05\x02\x00\xAD\xDD\x02\x00\xEB\x6E"
-     "\x00\x80"
-     "\xEF\x6E\x00\x80\xF3\x6E\x00\x80\xF7\x6E\x00\x80\x80\x1D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x39\x00\x00\xFB\x6E\x00\x80\xFF\x6E\x00\x80\xBE\x00\x04\x00\x07\x6F"
-     "\x00\x80"
-     "\x0B\x6F\x00\x80\x0F\x6F\x00\x80\x13\x6F\x00\x80\x17\x6F\x00\x80\x1B\x6F"
-     "\x00\x80"
-     "\x84\x8C\x03\x00\x1F\x6F\x00\x80\x87\x08\x03\x00\x86\xEC\x04\x00\x23\x6F"
-     "\x00\x80"
-     "\xEF\x88\x02\x00\x27\x6F\x00\x80\x2B\x6F\x00\x80\x84\x6C\x04\x00\xE3\x34"
-     "\x02\x00"
-     "\x2F\x6F\x00\x80\xE1\x54\x01\x00\x33\x6F\x00\x80\x37\x6F\x00\x80\x3B\x6F"
-     "\x00\x80"
-     "\x3F\x6F\x00\x80\xA8\x6D\x02\x00\xA9\xAD\x02\x00\xAA\xA5\x02\x00\xAB\xBD"
-     "\x02\x00"
-     "\xAC\xA5\x02\x00\xAD\xAD\x02\x00\xAE\xA5\x02\x00\xAF\x19\x01\x00\xBE\xAC"
-     "\x04\x00"
-     "\x43\x6F\x00\x80\x47\x6F\x00\x80\x4B\x6F\x00\x80\x4F\x6F\x00\x80\x53\x6F"
-     "\x00\x80"
-     "\x57\x6F\x00\x80\x5B\x6F\x00\x80\xB8\x0D\x01\x00\xB9\x11\x01\x00\xBA\x11"
-     "\x01\x00"
-     "\xBB\x25\x01\x00\xBC\x3D\x01\x00\xBD\xD5\x01\x00\xBE\xDD\x01\x00\xBF\xD5"
-     "\x01\x00"
-     "\xB0\x69\x01\x00\xB1\x69\x01\x00\xB2\x79\x01\x00\xB3\x71\x01\x00\xB4\x59"
-     "\x01\x00"
-     "\xB5\x59\x01\x00\xB6\x39\x01\x00\xB7\x35\x01\x00\xB3\x2D\x02\x00\x5F\x6F"
-     "\x00\x80"
-     "\x63\x6F\x00\x80\x67\x6F\x00\x80\x6B\x6F\x00\x80\xB6\x2D\x02\x00\xB5\x2D"
-     "\x02\x00"
-     "\x6F\x6F\x00\x80\xBB\xAD\x01\x00\xBA\xAD\x01\x00\x77\x6F\x00\x80\x7B\x6F"
-     "\x00\x80"
-     "\xBF\x9D\x01\x00\xBE\x9D\x01\x00\xBD\xA5\x01\x00\xBC\xAD\x01\x00\x80\x4D"
-     "\x00\x00"
-     "\x81\x55\x00\x00\x82\x55\x00\x00\xEF\x6C\x00\x00\x7F\x6F\x00\x80\xEF\xEC"
-     "\x7F\x00"
-     "\xEF\xBC\x7E\x00\x83\x6F\x00\x80\xE1\x10\x7F\x00\xE3\xD4\x7F\x00\xE3\xD8"
-     "\x7E\x00"
-     "\xE1\xEC\x7F\x00\x87\x6F\x00\x80\xE1\x4C\x7E\x00\x8B\x6F\x00\x80\xE3\x24"
-     "\x7E\x00"
-     "\x03\x6F\x00\x80\x73\x6F\x00\x80\xAB\x05\x06\x00\xAA\x05\x06\x00\xAD\x0D"
-     "\x06\x00"
-     "\xAC\x05\x06\x00\xAF\x35\x06\x00\xAE\x35\x06\x00\x86\x00\x03\x00\x87\x28"
-     "\x03\x00"
-     "\xA3\x85\x05\x00\x8F\x6F\x00\x80\xA5\x85\x05\x00\x93\x6F\x00\x80\x97\x6F"
-     "\x00\x80"
-     "\xA6\x85\x05\x00\xB3\xF1\x06\x00\x9B\x6F\x00\x80\x9F\x6F\x00\x80\xA3\x6F"
-     "\x00\x80"
-     "\xA7\x6F\x00\x80\xB6\xE9\x06\x00\xB5\xE1\x06\x00\xAB\x6F\x00\x80\xBB\xBD"
-     "\x06\x00"
-     "\xBA\xBD\x06\x00\xAF\x6F\x00\x80\xB3\x6F\x00\x80\xBF\x89\x06\x00\xBE\x81"
-     "\x06\x00"
-     "\xBD\x89\x06\x00\xBC\x95\x06\x00\xA8\x29\x06\x00\xA9\x29\x06\x00\xAA\x39"
-     "\x06\x00"
-     "\xAB\x39\x06\x00\xAC\x29\x06\x00\xAD\x29\x06\x00\xAE\x5D\x06\x00\xAF\x4D"
-     "\x06\x00"
-     "\xB7\x6F\x00\x80\xBB\x6F\x00\x80\xBF\x6F\x00\x80\xC3\x6F\x00\x80\xC7\x6F"
-     "\x00\x80"
-     "\xCB\x6F\x00\x80\xCF\x6F\x00\x80\xD3\x6F\x00\x80\xB8\xE9\x07\x00\xB9\xE9"
-     "\x07\x00"
-     "\xBA\xF9\x07\x00\xBB\xF9\x07\x00\xBC\xE9\x07\x00\xBD\xE9\x07\x00\xBE\x5D"
-     "\x07\x00"
-     "\xBF\x55\x07\x00\xB0\x39\x06\x00\xB1\x39\x06\x00\xB2\x01\x06\x00\xB3\x1D"
-     "\x06\x00"
-     "\xB4\x05\x06\x00\xB5\x0D\x06\x00\xB6\x05\x06\x00\xB7\xD9\x07\x00\xA3\xB1"
-     "\x07\x00"
-     "\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D\x00\x00\xD7\x6F\x00\x80\xA6\xA9"
-     "\x07\x00"
-     "\xA5\xA1\x07\x00\xDB\x6F\x00\x80\xAB\xFD\x07\x00\xAA\xFD\x07\x00\xDF\x6F"
-     "\x00\x80"
-     "\x84\xA4\x02\x00\xAF\xC9\x07\x00\xAE\xC1\x07\x00\xAD\xC9\x07\x00\xAC\xD5"
-     "\x07\x00"
-     "\xBE\xCC\x01\x00\xB3\x95\x06\x00\xE3\x6F\x00\x80\xE7\x6F\x00\x80\xB6\xA9"
-     "\x06\x00"
-     "\xEB\x6F\x00\x80\xEF\x6F\x00\x80\xB5\xAD\x06\x00\xBA\x59\x01\x00\xBB\x21"
-     "\x01\x00"
-     "\x86\xC8\x00\x00\x87\x0C\x01\x00\xBE\x21\x01\x00\xBF\x29\x01\x00\xBC\x31"
-     "\x01\x00"
-     "\xBD\x31\x01\x00\xA8\x29\x06\x00\xA9\x29\x06\x00\xAA\x59\x06\x00\xAB\x51"
-     "\x06\x00"
-     "\xAC\x61\x06\x00\xAD\x6D\x06\x00\xAE\xB5\x01\x00\xAF\xA9\x01\x00\x84\xE0"
-     "\x01\x00"
-     "\xF3\x6F\x00\x80\xF7\x6F\x00\x80\xFB\x6F\x00\x80\xFF\x6F\x00\x80\x03\x70"
-     "\x00\x80"
-     "\x07\x70\x00\x80\x0B\x70\x00\x80\xB8\x61\x01\x00\xB9\x61\x01\x00\xBA\x61"
-     "\x01\x00"
-     "\xBB\x61\x01\x00\xBC\x61\x01\x00\xBD\x61\x01\x00\xBE\x61\x01\x00\xBF\x61"
-     "\x01\x00"
-     "\xB0\xD9\x01\x00\xB1\xA1\x01\x00\xB2\xA1\x01\x00\xB3\xA1\x01\x00\xB4\xA1"
-     "\x01\x00"
-     "\xB5\xA9\x01\x00\xB6\x91\x01\x00\xB7\x91\x01\x00\xA3\xD1\x05\x00\x0F\x70"
-     "\x00\x80"
-     "\x13\x70\x00\x80\x17\x70\x00\x80\x1B\x70\x00\x80\xA6\xED\x05\x00\xA5\xE9"
-     "\x05\x00"
-     "\x1F\x70\x00\x80\xAB\x65\x02\x00\xAA\x1D\x02\x00\x23\x70\x00\x80\x27\x70"
-     "\x00\x80"
-     "\xAF\x6D\x02\x00\xAE\x65\x02\x00\xAD\x75\x02\x00\xAC\x75\x02\x00\x2B\x70"
-     "\x00\x80"
-     "\x2F\x70\x00\x80\x33\x70\x00\x80\x37\x70\x00\x80\x3B\x70\x00\x80\x3F\x70"
-     "\x00\x80"
-     "\x43\x70\x00\x80\x47\x70\x00\x80\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19"
-     "\x00\x00"
-     "\x4B\x70\x00\x80\x4F\x70\x00\x80\x57\x70\x00\x80\xBE\x3C\x03\x00\x5B\x70"
-     "\x00\x80"
-     "\x86\x00\x0C\x00\x87\x14\x03\x00\x5F\x70\x00\x80\xB3\xD5\x03\x00\x63\x70"
-     "\x00\x80"
-     "\xB5\x3D\x03\x00\xB6\x35\x03\x00\x67\x70\x00\x80\x6B\x70\x00\x80\xBF\x8C"
-     "\x0A\x00"
-     "\xBA\x11\x03\x00\xBB\x11\x03\x00\xBC\xB5\x00\x00\xBD\xBD\x00\x00\xBE\xB5"
-     "\x00\x00"
-     "\xBF\xAD\x00\x00\x6F\x70\x00\x80\xE3\x74\x01\x00\x73\x70\x00\x80\xE1\xBC"
-     "\x01\x00"
-     "\x77\x70\x00\x80\x7B\x70\x00\x80\x7F\x70\x00\x80\x83\x70\x00\x80\x87\x70"
-     "\x00\x80"
-     "\x8B\x70\x00\x80\x8F\x70\x00\x80\x93\x70\x00\x80\x97\x70\x00\x80\xEF\x74"
-     "\x02\x00"
-     "\x9B\x70\x00\x80\x9F\x70\x00\x80\xA8\x4D\x02\x00\xA9\x51\x02\x00\xAA\x51"
-     "\x02\x00"
-     "\xAB\xA9\x02\x00\xAC\xB9\x02\x00\xAD\xB9\x02\x00\xAE\xA9\x02\x00\xAF\xA9"
-     "\x02\x00"
-     "\x84\x6C\x0D\x00\xA3\x70\x00\x80\xA7\x70\x00\x80\xAB\x70\x00\x80\xAF\x70"
-     "\x00\x80"
-     "\xB3\x70\x00\x80\xB7\x70\x00\x80\xBE\x74\x0D\x00\xB8\x6D\x01\x00\xB9\x05"
-     "\x01\x00"
-     "\xBA\x0D\x01\x00\xBB\x05\x01\x00\xBC\x1D\x01\x00\xBD\x05\x01\x00\xBE\x0D"
-     "\x01\x00"
-     "\xBF\x05\x01\x00\xB0\xD9\x02\x00\xB1\xD9\x02\x00\xB2\x6D\x01\x00\xB3\x65"
-     "\x01\x00"
-     "\xB4\x7D\x01\x00\xB5\x65\x01\x00\xB6\x65\x01\x00\xB7\x55\x01\x00\xE1\xB8"
-     "\x01\x00"
-     "\xE1\x50\x07\x00\xE3\xB4\x00\x00\xE3\x7C\x07\x00\x80\xA9\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x39\x00\x00\xBB\x70\x00\x80\xBF\x70\x00\x80\xC7\x70\x00\x80\xCB\x70"
-     "\x00\x80"
-     "\xCF\x70\x00\x80\xEE\x00\x00\x00\xD3\x70\x00\x80\xEF\x00\x00\x00\xEF\x64"
-     "\x06\x00"
-     "\x86\x60\x0C\x00\x87\xE4\x0C\x00\xA3\x91\x02\x00\xD7\x70\x00\x80\xA5\x79"
-     "\x02\x00"
-     "\xDB\x70\x00\x80\xDF\x70\x00\x80\xA6\x71\x02\x00\xE3\x70\x00\x80\xE7\x70"
-     "\x00\x80"
-     "\xAB\x55\x02\x00\xAA\x55\x02\x00\xAD\xF9\x01\x00\xAC\xF1\x01\x00\xAF\xE9"
-     "\x01\x00"
-     "\xAE\xF1\x01\x00\xC3\x70\x00\x80\x53\x70\x00\x80\xEB\x70\x00\x80\xEF\x70"
-     "\x00\x80"
-     "\xF3\x70\x00\x80\xF7\x70\x00\x80\xFB\x70\x00\x80\xFF\x70\x00\x80\xA8\x55"
-     "\x0E\x00"
-     "\xA9\x59\x0E\x00\xAA\xA1\x0E\x00\xAB\xBD\x0E\x00\xAC\xAD\x0E\x00\xAD\x95"
-     "\x0E\x00"
-     "\xAE\xF9\x0E\x00\xAF\xF5\x0E\x00\xB0\x91\x0E\x00\xB1\x91\x0E\x00\xB2\x91"
-     "\x0E\x00"
-     "\xB3\x91\x0E\x00\xB4\xB1\x0E\x00\xB5\xB1\x0E\x00\xB6\xB1\x0E\x00\xB7\xB1"
-     "\x0E\x00"
-     "\xB8\x91\x0E\x00\xB9\x9D\x0E\x00\xBA\x95\x0E\x00\xBB\x49\x0F\x00\xBC\x59"
-     "\x0F\x00"
-     "\xBD\x59\x0F\x00\xBE\x49\x0F\x00\xBF\x49\x0F\x00\xB3\x09\x0E\x00\x03\x71"
-     "\x00\x80"
-     "\x07\x71\x00\x80\x0B\x71\x00\x80\x0F\x71\x00\x80\xB6\x35\x0E\x00\xB5\x05"
-     "\x0E\x00"
-     "\x13\x71\x00\x80\xBB\x75\x0E\x00\xBA\x75\x0E\x00\x17\x71\x00\x80\x1B\x71"
-     "\x00\x80"
-     "\xBF\x55\x0E\x00\xBE\x55\x0E\x00\xBD\x65\x0E\x00\xBC\x65\x0E\x00\x1F\x71"
-     "\x00\x80"
-     "\xA3\x4D\x0E\x00\x23\x71\x00\x80\x27\x71\x00\x80\xA6\x71\x0E\x00\x2B\x71"
-     "\x00\x80"
-     "\x2F\x71\x00\x80\xA5\x41\x0E\x00\xAA\x31\x0E\x00\xAB\x31\x0E\x00\x84\xA4"
-     "\x03\x00"
-     "\xBE\xA4\x03\x00\xAE\x11\x0E\x00\xAF\x11\x0E\x00\xAC\x21\x0E\x00\xAD\x21"
-     "\x0E\x00"
-     "\xA8\xA5\x0E\x00\xA9\xAD\x0E\x00\xAA\xA5\x0E\x00\xAB\xB9\x0E\x00\xAC\xDD"
-     "\x0E\x00"
-     "\xAD\xC1\x0E\x00\xAE\xC1\x0E\x00\xAF\xFD\x0E\x00\x80\xED\x01\x00\x81\xF1"
-     "\x01\x00"
-     "\x82\xF1\x01\x00\x33\x71\x00\x80\x86\x90\x01\x00\x87\xB4\x01\x00\x37\x71"
-     "\x00\x80"
-     "\x3B\x71\x00\x80\xB8\xC9\x01\x00\xB9\xC9\x01\x00\xBA\xD9\x01\x00\xBB\xD1"
-     "\x01\x00"
-     "\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\x99\x01\x00\xBF\x95\x01\x00\xB0\x85"
-     "\x0E\x00"
-     "\xB1\x6D\x01\x00\xB2\x65\x01\x00\xB3\x7D\x01\x00\xB4\x65\x01\x00\xB5\x6D"
-     "\x01\x00"
-     "\xB6\x65\x01\x00\xB7\xF9\x01\x00\xB3\x2D\x0E\x00\x3F\x71\x00\x80\x43\x71"
-     "\x00\x80"
-     "\x47\x71\x00\x80\x4B\x71\x00\x80\xB6\x41\x0E\x00\xB5\x55\x0E\x00\x4F\x71"
-     "\x00\x80"
-     "\xBB\xA9\x01\x00\xBA\x41\x0E\x00\x53\x71\x00\x80\x57\x71\x00\x80\xBF\xA9"
-     "\x01\x00"
-     "\xBE\xA1\x01\x00\xBD\xA9\x01\x00\xBC\xB1\x01\x00\x5B\x71\x00\x80\xA3\x69"
-     "\x0E\x00"
-     "\x5F\x71\x00\x80\x63\x71\x00\x80\xA6\x05\x0E\x00\x67\x71\x00\x80\x6B\x71"
-     "\x00\x80"
-     "\xA5\x11\x0E\x00\xAA\x05\x0E\x00\xAB\xED\x01\x00\x6F\x71\x00\x80\x73\x71"
-     "\x00\x80"
-     "\xAE\xE5\x01\x00\xAF\xED\x01\x00\xAC\xF5\x01\x00\xAD\xED\x01\x00\xA8\x39"
-     "\x03\x00"
-     "\xA9\x39\x03\x00\xAA\x8D\x03\x00\xAB\x85\x03\x00\xAC\x9D\x03\x00\xAD\x85"
-     "\x03\x00"
-     "\xAE\x85\x03\x00\xAF\xB5\x03\x00\x77\x71\x00\x80\x7B\x71\x00\x80\x7F\x71"
-     "\x00\x80"
-     "\x83\x71\x00\x80\x87\x71\x00\x80\x8B\x71\x00\x80\x8F\x71\x00\x80\x93\x71"
-     "\x00\x80"
-     "\xB8\x61\x00\x00\xB9\x61\x00\x00\xBA\x61\x00\x00\xBB\x61\x00\x00\xBC\x61"
-     "\x00\x00"
-     "\xBD\x61\x00\x00\xBE\x61\x00\x00\xBF\x61\x00\x00\xB0\xCD\x03\x00\xB1\xA5"
-     "\x03\x00"
-     "\xB2\xA1\x03\x00\xB3\xA1\x03\x00\xB4\xA5\x03\x00\xB5\xAD\x03\x00\xB6\x91"
-     "\x03\x00"
-     "\xB7\x91\x03\x00\x80\x0D\x00\x00\x81\x11\x00\x00\x82\x11\x00\x00\x97\x71"
-     "\x00\x80"
-     "\xEF\xF4\x02\x00\x9B\x71\x00\x80\x9F\x71\x00\x80\xBE\x1C\x03\x00\xE3\x14"
-     "\x02\x00"
-     "\x84\x88\x02\x00\xE1\x80\x01\x00\xA7\x71\x00\x80\xAB\x71\x00\x80\xAF\x71"
-     "\x00\x80"
-     "\x87\xC8\x03\x00\x86\x3C\x04\x00\xBB\x01\x03\x00\xBA\x69\x03\x00\xB3\x71"
-     "\x00\x80"
-     "\xB7\x71\x00\x80\xBF\x09\x03\x00\xBE\x01\x03\x00\xBD\x15\x03\x00\xBC\x15"
-     "\x03\x00"
-     "\xB3\x65\x03\x00\xBB\x71\x00\x80\xBF\x71\x00\x80\xC3\x71\x00\x80\xC7\x71"
-     "\x00\x80"
-     "\xB6\x65\x03\x00\xB5\x75\x03\x00\xCB\x71\x00\x80\xCF\x71\x00\x80\xD3\x71"
-     "\x00\x80"
-     "\xD7\x71\x00\x80\xA3\x89\x02\x00\xDB\x71\x00\x80\xA5\x99\x02\x00\xA6\x89"
-     "\x02\x00"
-     "\xDF\x71\x00\x80\x84\x2C\x02\x00\xE3\x71\x00\x80\xAA\x85\x02\x00\xAB\xED"
-     "\x02\x00"
-     "\xAC\xF9\x02\x00\xAD\xF9\x02\x00\xAE\xED\x02\x00\xAF\xE5\x02\x00\xE7\x71"
-     "\x00\x80"
-     "\xEB\x71\x00\x80\xBE\x44\x05\x00\x84\x4C\x05\x00\xEF\x71\x00\x80\xF3\x71"
-     "\x00\x80"
-     "\xF7\x71\x00\x80\xFB\x71\x00\x80\xFF\x71\x00\x80\x03\x72\x00\x80\x07\x72"
-     "\x00\x80"
-     "\x0B\x72\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x0F\x72"
-     "\x00\x80"
-     "\xE1\x18\x0F\x00\xE1\x5C\x0E\x00\xE3\xB8\x0E\x00\xE3\x74\x01\x00\x17\x72"
-     "\x00\x80"
-     "\x1B\x72\x00\x80\x1F\x72\x00\x80\x23\x72\x00\x80\x86\x00\x04\x00\x87\x64"
-     "\x05\x00"
-     "\x27\x72\x00\x80\x2B\x72\x00\x80\x2F\x72\x00\x80\x33\x72\x00\x80\xEF\x7C"
-     "\x0E\x00"
-     "\xEF\xA8\x01\x00\xB3\x75\x01\x00\x37\x72\x00\x80\x3B\x72\x00\x80\x3F\x72"
-     "\x00\x80"
-     "\x43\x72\x00\x80\xB6\x31\x01\x00\xB5\x15\x01\x00\x47\x72\x00\x80\xBB\x1D"
-     "\x01\x00"
-     "\xBA\x15\x01\x00\x4B\x72\x00\x80\x4F\x72\x00\x80\xBF\xE1\x00\x00\xBE\xFD"
-     "\x00\x00"
-     "\xBD\xFD\x00\x00\xBC\xFD\x00\x00\x13\x72\x00\x80\x53\x72\x00\x80\x57\x72"
-     "\x00\x80"
-     "\x5B\x72\x00\x80\xA3\x71\x00\x80\x5F\x72\x00\x80\x63\x72\x00\x80\x67\x72"
-     "\x00\x80"
-     "\xA8\x8D\x06\x00\xA9\x95\x06\x00\xAA\x9D\x06\x00\xAB\xE5\x06\x00\xAC\xFD"
-     "\x06\x00"
-     "\xAD\xD1\x06\x00\xAE\xD1\x06\x00\xAF\xD1\x06\x00\xB0\xB1\x06\x00\xB1\xB9"
-     "\x06\x00"
-     "\xB2\x49\x07\x00\xB3\x49\x07\x00\xB4\x59\x07\x00\xB5\x45\x07\x00\xB6\x45"
-     "\x07\x00"
-     "\xB7\x79\x07\x00\xB8\x21\x07\x00\xB9\x21\x07\x00\xBA\x39\x07\x00\xBB\x39"
-     "\x07\x00"
-     "\xBC\x29\x07\x00\xBD\x29\x07\x00\xBE\x19\x07\x00\xBF\x19\x07\x00\xA3\x35"
-     "\x06\x00"
-     "\x6B\x72\x00\x80\x6F\x72\x00\x80\x73\x72\x00\x80\x77\x72\x00\x80\xA6\x71"
-     "\x06\x00"
-     "\xA5\x55\x06\x00\x7B\x72\x00\x80\xAB\x5D\x06\x00\xAA\x55\x06\x00\x7F\x72"
-     "\x00\x80"
-     "\xBE\xA0\x03\x00\xAF\xA1\x07\x00\xAE\xBD\x07\x00\xAD\xBD\x07\x00\xAC\xBD"
-     "\x07\x00"
-     "\x80\x51\x00\x00\x81\x59\x00\x00\x82\x61\x00\x00\xB3\x55\x07\x00\x85\xF4"
-     "\x00\x00"
-     "\xB5\x7D\x07\x00\xB6\x75\x07\x00\x83\x72\x00\x80\x86\x00\x1C\x00\x87\xE4"
-     "\x01\x00"
-     "\xBA\x2D\x07\x00\xBB\x25\x07\x00\xBC\x3D\x07\x00\xBD\x25\x07\x00\xBE\x25"
-     "\x07\x00"
-     "\xBF\x15\x07\x00\xA8\x91\x06\x00\xA9\x91\x06\x00\xAA\x91\x06\x00\xAB\x91"
-     "\x06\x00"
-     "\xAC\xB9\x06\x00\xAD\xB9\x06\x00\xAE\xA9\x06\x00\xAF\xA9\x06\x00\x87\x72"
-     "\x00\x80"
-     "\x8B\x72\x00\x80\x8F\x72\x00\x80\x93\x72\x00\x80\xA2\x21\x01\x00\xA3\x35"
-     "\x01\x00"
-     "\xA0\x39\x05\x00\xA1\x11\x04\x00\xB8\x49\x01\x00\xB9\x49\x01\x00\xBA\x5D"
-     "\x01\x00"
-     "\xBB\x55\x01\x00\xBC\x4D\x01\x00\xBD\xD1\x01\x00\xBE\xD1\x01\x00\xBF\xD1"
-     "\x01\x00"
-     "\xB0\xA5\x06\x00\xB1\xAD\x06\x00\xB2\xA5\x06\x00\xB3\xBD\x06\x00\xB4\xAD"
-     "\x06\x00"
-     "\xB5\x9D\x06\x00\xB6\x95\x06\x00\xB7\x79\x01\x00\xA3\x19\x06\x00\x8F\x9D"
-     "\x79\x00"
-     "\x97\x72\x00\x80\x9B\x72\x00\x80\x9F\x72\x00\x80\xA6\x39\x06\x00\xA5\x31"
-     "\x06\x00"
-     "\xA3\x72\x00\x80\xAB\x69\x06\x00\xAA\x61\x06\x00\xA7\x72\x00\x80\xAB\x72"
-     "\x00\x80"
-     "\xAF\x59\x06\x00\xAE\x69\x06\x00\xAD\x69\x06\x00\xAC\x71\x06\x00\x9E\x89"
-     "\x08\x00"
-     "\x9F\xC1\x05\x00\x9C\xC9\x09\x00\x9D\xC9\x09\x00\x9A\xA1\x0D\x00\x9B\xBD"
-     "\x0C\x00"
-     "\x98\xB1\x0D\x00\x99\xB9\x0D\x00\x96\xA1\x71\x00\x97\x45\x71\x00\x94\x45"
-     "\x75\x00"
-     "\x95\xB1\x71\x00\x92\xA1\x75\x00\x93\xB5\x75\x00\x90\xE5\x79\x00\x91\xCD"
-     "\x79\x00"
-     "\x8A\x5D\x72\x00\x8B\x45\x72\x00\xAF\x72\x00\x80\xBE\x80\x1C\x00\x8E\x75"
-     "\x0E\x00"
-     "\x8F\x65\x0E\x00\x8C\xB9\x72\x00\x8D\x75\x0E\x00\x82\x39\x7A\x00\x83\x39"
-     "\x7A\x00"
-     "\xB3\x72\x00\x80\xB7\x72\x00\x80\x86\x71\x76\x00\x87\x99\x76\x00\x84\x09"
-     "\x7A\x00"
-     "\x85\x6D\x76\x00\x9A\x6D\x07\x00\x9B\x55\x02\x00\xBB\x72\x00\x80\xBF\x72"
-     "\x00\x80"
-     "\xC3\x72\x00\x80\xE1\x90\x00\x00\x9C\x59\x02\x00\xE3\x08\x1A\x00\x92\x09"
-     "\x0F\x00"
-     "\x93\x65\x0A\x00\xC7\x72\x00\x80\xEF\x38\x16\x00\x96\x75\x06\x00\x97\x75"
-     "\x06\x00"
-     "\x94\x7D\x0A\x00\x95\x35\x0B\x00\xA9\x8D\x16\x00\xA8\x85\x16\x00\xAB\x01"
-     "\x10\x00"
-     "\xAA\x31\x16\x00\xAD\x71\x12\x00\xAC\xB5\x12\x00\xAF\xB9\x2E\x00\xAE\x01"
-     "\x2C\x00"
-     "\xA1\x7D\x02\x00\xCB\x72\x00\x80\xA3\xA1\x1E\x00\xA2\xA9\x1E\x00\xA5\xB1"
-     "\x1A\x00"
-     "\xA4\xF5\x1F\x00\xA7\xE5\x1B\x00\xA6\xB1\x1A\x00\x84\xCC\x03\x00\x84\x4C"
-     "\x1C\x00"
-     "\xCF\x72\x00\x80\xD3\x72\x00\x80\xD7\x72\x00\x80\xDB\x72\x00\x80\xDF\x72"
-     "\x00\x80"
-     "\xE3\x72\x00\x80\xB1\x01\x28\x00\xB0\xD9\x2E\x00\xB3\x8D\x2A\x00\xB2\xE9"
-     "\x2A\x00"
-     "\xB5\xF5\x26\x00\xB4\x01\x24\x00\x84\x70\x1D\x00\xE7\x72\x00\x80\x80\xFD"
-     "\x01\x00"
-     "\x81\x15\x00\x00\x82\x1D\x00\x00\xBE\x80\x1C\x00\xEB\x72\x00\x80\xEF\x72"
-     "\x00\x80"
-     "\x87\xB8\x02\x00\x86\x3C\x1D\x00\xF7\x72\x00\x80\xFB\x72\x00\x80\xFF\x72"
-     "\x00\x80"
-     "\x03\x73\x00\x80\x07\x73\x00\x80\x0B\x73\x00\x80\x0F\x73\x00\x80\x13\x73"
-     "\x00\x80"
-     "\x17\x73\x00\x80\x1B\x73\x00\x80\x1F\x73\x00\x80\x23\x73\x00\x80\xE3\x80"
-     "\x03\x00"
-     "\x27\x73\x00\x80\xE1\xA0\x01\x00\x2B\x73\x00\x80\xEF\x94\x03\x00\x2F\x73"
-     "\x00\x80"
-     "\x33\x73\x00\x80\x37\x73\x00\x80\x3B\x73\x00\x80\x3F\x73\x00\x80\x43\x73"
-     "\x00\x80"
-     "\x47\x73\x00\x80\x4B\x73\x00\x80\xE1\x3C\x06\x00\x4F\x73\x00\x80\xE3\xD0"
-     "\x06\x00"
-     "\x53\x73\x00\x80\xE1\x30\x07\x00\x57\x73\x00\x80\xE3\xAC\x06\x00\x80\x39"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x1D\x00\x00\xEF\x1C\x06\x00\x5B\x73\x00\x80\x5F\x73"
-     "\x00\x80"
-     "\xBE\xB8\x1F\x00\xEF\xE8\x01\x00\xB3\xE9\x02\x00\x67\x73\x00\x80\x87\xC4"
-     "\x1C\x00"
-     "\x86\xEC\x1C\x00\x6B\x73\x00\x80\xB6\x59\x02\x00\xB5\x51\x02\x00\x6F\x73"
-     "\x00\x80"
-     "\xBB\x4D\x02\x00\xBA\x4D\x02\x00\x73\x73\x00\x80\x77\x73\x00\x80\xBF\xB9"
-     "\x01\x00"
-     "\xBE\xD9\x01\x00\xBD\xD1\x01\x00\xBC\xF5\x01\x00\xA3\x29\x1D\x00\xF3\x72"
-     "\x00\x80"
-     "\x63\x73\x00\x80\x7B\x73\x00\x80\x7F\x73\x00\x80\xA6\x99\x1D\x00\xA5\x91"
-     "\x1D\x00"
-     "\x83\x73\x00\x80\xAB\x8D\x1D\x00\xAA\x8D\x1D\x00\x87\x73\x00\x80\x8B\x73"
-     "\x00\x80"
-     "\xAF\x79\x1E\x00\xAE\x19\x1E\x00\xAD\x11\x1E\x00\xAC\x35\x1E\x00\x8F\x73"
-     "\x00\x80"
-     "\xB3\x6D\x1F\x00\x93\x73\x00\x80\x97\x73\x00\x80\xB6\x65\x1F\x00\x9B\x73"
-     "\x00\x80"
-     "\x9F\x73\x00\x80\xB5\x6D\x1F\x00\xBA\x21\x1F\x00\xBB\x21\x1F\x00\xA3\x73"
-     "\x00\x80"
-     "\xA7\x73\x00\x80\xBE\x25\x1F\x00\xBF\x29\x1F\x00\xBC\x31\x1F\x00\xBD\x31"
-     "\x1F\x00"
-     "\xA8\xA1\x1F\x00\xA9\xA1\x1F\x00\xAA\xA1\x1F\x00\xAB\xA1\x1F\x00\xAC\xA1"
-     "\x1F\x00"
-     "\xAD\xA1\x1F\x00\xAE\xA1\x1F\x00\xAF\xA1\x1F\x00\xAB\x73\x00\x80\xAF\x73"
-     "\x00\x80"
-     "\xB3\x73\x00\x80\xB7\x73\x00\x80\xBB\x73\x00\x80\xBF\x73\x00\x80\xC3\x73"
-     "\x00\x80"
-     "\xC7\x73\x00\x80\xB8\xAD\x1F\x00\xB9\xB5\x1F\x00\xBA\xBD\x1F\x00\xBB\xB5"
-     "\x1F\x00"
-     "\xBC\xAD\x1F\x00\xBD\x55\x1F\x00\xBE\x51\x1F\x00\xBF\x4D\x1F\x00\xB0\xA1"
-     "\x1F\x00"
-     "\xB1\xA1\x1F\x00\xB2\xA1\x1F\x00\xB3\xA1\x1F\x00\xB4\xA5\x1F\x00\xB5\xA9"
-     "\x1F\x00"
-     "\xB6\x9D\x1F\x00\xB7\x95\x1F\x00\xA3\x29\x1E\x00\x82\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x80\xB1\x01\x00\xCB\x73\x00\x80\xA6\x21\x1E\x00\xA5\x29\x1E\x00\xCF\x73"
-     "\x00\x80"
-     "\xAB\x65\x1E\x00\xAA\x65\x1E\x00\x86\x88\x00\x00\x87\xFC\x01\x00\xAF\x6D"
-     "\x1E\x00"
-     "\xAE\x61\x1E\x00\xAD\x75\x1E\x00\xAC\x75\x1E\x00\xD3\x73\x00\x80\xB3\x99"
-     "\x1E\x00"
-     "\xD7\x73\x00\x80\xDB\x73\x00\x80\xB6\x5D\x01\x00\xDF\x73\x00\x80\xE3\x73"
-     "\x00\x80"
-     "\xB5\xB1\x1E\x00\xBA\x49\x01\x00\xBB\x49\x01\x00\xE7\x73\x00\x80\xEB\x73"
-     "\x00\x80"
-     "\xBE\x3D\x01\x00\xBF\x21\x01\x00\xBC\x39\x01\x00\xBD\x35\x01\x00\xA8\x45"
-     "\x1E\x00"
-     "\xA9\x55\x1E\x00\xAA\x55\x1E\x00\xAB\x65\x1E\x00\xAC\x7D\x1E\x00\xAD\x99"
-     "\x01\x00"
-     "\xAE\x89\x01\x00\xAF\x81\x01\x00\x84\xAC\x00\x00\xEF\x73\x00\x80\xF3\x73"
-     "\x00\x80"
-     "\xF7\x73\x00\x80\xFB\x73\x00\x80\xFF\x73\x00\x80\x03\x74\x00\x80\x07\x74"
-     "\x00\x80"
-     "\xB8\xAD\x01\x00\xB9\x65\x01\x00\xBA\x6D\x01\x00\xBB\x65\x01\x00\xBC\x7D"
-     "\x01\x00"
-     "\xBD\x65\x01\x00\xBE\x6D\x01\x00\xBF\x65\x01\x00\xB0\xC9\x01\x00\xB1\xC9"
-     "\x01\x00"
-     "\xB2\xA9\x01\x00\xB3\xA5\x01\x00\xB4\xBD\x01\x00\xB5\xA1\x01\x00\xB6\xA1"
-     "\x01\x00"
-     "\xB7\x95\x01\x00\xB8\x69\x1C\x00\xB9\xA1\x1C\x00\xBA\xC1\x1C\x00\xBB\xC1"
-     "\x1C\x00"
-     "\xBC\xC1\x1C\x00\xBD\xC1\x1C\x00\xBE\xC1\x1C\x00\xBF\xC1\x1C\x00\xB0\x89"
-     "\x1F\x00"
-     "\xB1\x89\x1F\x00\xB2\x21\x1C\x00\xB3\x05\x1C\x00\xB4\x1D\x1C\x00\xB5\x7D"
-     "\x1C\x00"
-     "\xB6\x75\x1C\x00\xB7\x6D\x1C\x00\xA8\x61\x1F\x00\xA9\x61\x1F\x00\xAA\x61"
-     "\x1F\x00"
-     "\xAB\x61\x1F\x00\xAC\xD9\x1F\x00\xAD\xD9\x1F\x00\xAE\xC9\x1F\x00\xAF\xC1"
-     "\x1F\x00"
-     "\x0B\x74\x00\x80\x0F\x74\x00\x80\x13\x74\x00\x80\x17\x74\x00\x80\x1B\x74"
-     "\x00\x80"
-     "\x1F\x74\x00\x80\x23\x74\x00\x80\x27\x74\x00\x80\x2B\x74\x00\x80\x2F\x74"
-     "\x00\x80"
-     "\xBE\x00\x04\x00\xA3\x51\x1D\x00\x33\x74\x00\x80\xA5\x79\x1D\x00\xA6\x95"
-     "\x02\x00"
-     "\x37\x74\x00\x80\x3B\x74\x00\x80\x3F\x74\x00\x80\xAA\x81\x02\x00\xAB\x81"
-     "\x02\x00"
-     "\xAC\xF1\x02\x00\xAD\xFD\x02\x00\xAE\xF5\x02\x00\xAF\xE9\x02\x00\x47\x74"
-     "\x00\x80"
-     "\xE1\xF4\x01\x00\x4B\x74\x00\x80\xE3\x7C\x01\x00\x80\x4D\x00\x00\x81\x75"
-     "\x00\x00"
-     "\x82\x7D\x00\x00\x4F\x74\x00\x80\x86\xC0\x04\x00\x87\xA4\x04\x00\x53\x74"
-     "\x00\x80"
-     "\x57\x74\x00\x80\x5B\x74\x00\x80\x5F\x74\x00\x80\x63\x74\x00\x80\xEF\x8C"
-     "\x02\x00"
-     "\xA8\x49\x02\x00\xA9\x49\x02\x00\xAA\x5D\x02\x00\xAB\x55\x02\x00\xAC\x79"
-     "\x02\x00"
-     "\xAD\x79\x02\x00\xAE\xBD\x02\x00\xAF\xB5\x02\x00\x84\xA0\x05\x00\x67\x74"
-     "\x00\x80"
-     "\x6B\x74\x00\x80\x6F\x74\x00\x80\xBE\xBC\x04\x00\x73\x74\x00\x80\x77\x74"
-     "\x00\x80"
-     "\x7B\x74\x00\x80\xB8\x2D\x01\x00\xB9\x35\x01\x00\xBA\x3D\x01\x00\xBB\x35"
-     "\x01\x00"
-     "\xBC\x2D\x01\x00\xBD\xD5\x01\x00\xBE\xDD\x01\x00\xBF\xCD\x01\x00\xB0\xCD"
-     "\x02\x00"
-     "\xB1\xD5\x02\x00\xB2\xDD\x02\x00\xB3\xD5\x02\x00\xB4\xCD\x02\x00\xB5\x15"
-     "\x01\x00"
-     "\xB6\x1D\x01\x00\xB7\x15\x01\x00\xE1\x84\x1E\x00\xE3\x6C\x1F\x00\xE3\x5C"
-     "\x1F\x00"
-     "\xE1\x58\x1E\x00\x7F\x74\x00\x80\x83\x74\x00\x80\x87\x74\x00\x80\x8B\x74"
-     "\x00\x80"
-     "\x8F\x74\x00\x80\x93\x74\x00\x80\x97\x74\x00\x80\x9B\x74\x00\x80\xE9\x00"
-     "\x00\x00"
-     "\xEF\xF4\x1E\x00\xEF\x4C\x1E\x00\x9F\x74\x00\x80\xB3\x95\x02\x00\xA3\x74"
-     "\x00\x80"
-     "\xA7\x74\x00\x80\xAB\x74\x00\x80\xAF\x74\x00\x80\xB6\xB9\x02\x00\xB5\xB1"
-     "\x02\x00"
-     "\xB7\x74\x00\x80\xBB\x51\x02\x00\xBA\x49\x02\x00\x86\xC8\x04\x00\x87\xAC"
-     "\x04\x00"
-     "\xBF\x91\x01\x00\xBE\x49\x02\x00\xBD\x41\x02\x00\xBC\x49\x02\x00\x43\x74"
-     "\x00\x80"
-     "\xA3\x51\x05\x00\xBB\x74\x00\x80\xB3\x74\x00\x80\xA6\x7D\x05\x00\xBF\x74"
-     "\x00\x80"
-     "\xC3\x74\x00\x80\xA5\x75\x05\x00\xAA\x8D\x05\x00\xAB\x95\x05\x00\xC7\x74"
-     "\x00\x80"
-     "\xCB\x74\x00\x80\xAE\x8D\x05\x00\xAF\x55\x06\x00\xAC\x8D\x05\x00\xAD\x85"
-     "\x05\x00"
-     "\x80\x89\x07\x00\x81\x89\x07\x00\x82\x99\x07\x00\xB3\x91\x06\x00\xCF\x74"
-     "\x00\x80"
-     "\xB5\xB9\x06\x00\xB6\xA9\x06\x00\xD3\x74\x00\x80\xD7\x74\x00\x80\xDB\x74"
-     "\x00\x80"
-     "\xBA\x4D\x07\x00\xBB\x45\x07\x00\xBC\x5D\x07\x00\xBD\x41\x07\x00\xBE\x41"
-     "\x07\x00"
-     "\xBF\x41\x07\x00\xA8\x41\x06\x00\xA9\x4D\x06\x00\xAA\x55\x06\x00\xAB\x65"
-     "\x06\x00"
-     "\xAC\x7D\x06\x00\xAD\x65\x06\x00\xAE\x6D\x06\x00\xAF\x65\x06\x00\xDF\x74"
-     "\x00\x80"
-     "\xE3\x74\x00\x80\xE7\x74\x00\x80\xEB\x74\x00\x80\xEF\x74\x00\x80\xF3\x74"
-     "\x00\x80"
-     "\xF7\x74\x00\x80\xFB\x74\x00\x80\xB8\x59\x07\x00\xB9\x59\x07\x00\xBA\x69"
-     "\x07\x00"
-     "\xBB\x69\x07\x00\xBC\x79\x07\x00\xBD\x79\x07\x00\xBE\x65\x07\x00\xBF\x19"
-     "\x07\x00"
-     "\xB0\xC5\x07\x00\xB1\xCD\x07\x00\xB2\xC5\x07\x00\xB3\xD9\x07\x00\xB4\xC9"
-     "\x07\x00"
-     "\xB5\xC9\x07\x00\xB6\x69\x07\x00\xB7\x69\x07\x00\xA3\xDD\x07\x00\xFF\x74"
-     "\x00\x80"
-     "\x03\x75\x00\x80\x07\x75\x00\x80\x0B\x75\x00\x80\xA6\xE5\x07\x00\xA5\xF5"
-     "\x07\x00"
-     "\x0F\x75\x00\x80\xAB\x09\x06\x00\xAA\x01\x06\x00\x13\x75\x00\x80\x17\x75"
-     "\x00\x80"
-     "\xAF\x0D\x06\x00\xAE\x0D\x06\x00\xAD\x0D\x06\x00\xAC\x11\x06\x00\x80\x6D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x19\x00\x00\x1B\x75\x00\x80\x1F\x75\x00\x80\x84\x98"
-     "\x01\x00"
-     "\xBE\x90\x01\x00\x23\x75\x00\x80\x86\xC0\x00\x00\x87\xE4\x01\x00\x27\x75"
-     "\x00\x80"
-     "\x2B\x75\x00\x80\x2F\x75\x00\x80\x33\x75\x00\x80\x37\x75\x00\x80\x3B\x75"
-     "\x00\x80"
-     "\xA8\x5D\x06\x00\xA9\x8D\x01\x00\xAA\x9D\x01\x00\xAB\x95\x01\x00\xAC\xB9"
-     "\x01\x00"
-     "\xAD\xB9\x01\x00\xAE\xC9\x01\x00\xAF\xC1\x01\x00\x84\xA0\x00\x00\x3F\x75"
-     "\x00\x80"
-     "\x43\x75\x00\x80\x47\x75\x00\x80\x4B\x75\x00\x80\x4F\x75\x00\x80\x53\x75"
-     "\x00\x80"
-     "\x57\x75\x00\x80\xB8\x79\x01\x00\xB9\x79\x01\x00\xBA\xCD\x00\x00\xBB\xC5"
-     "\x00\x00"
-     "\xBC\xDD\x00\x00\xBD\xC5\x00\x00\xBE\xC5\x00\x00\xBF\xF5\x00\x00\xB0\x81"
-     "\x01\x00"
-     "\xB1\x81\x01\x00\xB2\x49\x01\x00\xB3\x49\x01\x00\xB4\x59\x01\x00\xB5\x59"
-     "\x01\x00"
-     "\xB6\x49\x01\x00\xB7\x49\x01\x00\xB3\x15\x02\x00\x5B\x75\x00\x80\x5F\x75"
-     "\x00\x80"
-     "\x63\x75\x00\x80\x67\x75\x00\x80\xB6\x39\x02\x00\xB5\x31\x02\x00\x6B\x75"
-     "\x00\x80"
-     "\xBB\x45\x02\x00\xBA\x45\x02\x00\x6F\x75\x00\x80\x73\x75\x00\x80\xBF\x9D"
-     "\x02\x00"
-     "\xBE\x9D\x02\x00\xBD\x9D\x02\x00\xBC\x9D\x02\x00\x85\x7C\x3E\x00\xA3\x51"
-     "\x02\x00"
-     "\x77\x75\x00\x80\x7B\x75\x00\x80\xA6\x7D\x02\x00\x7F\x75\x00\x80\x83\x75"
-     "\x00\x80"
-     "\xA5\x75\x02\x00\xAA\x01\x02\x00\xAB\x01\x02\x00\x87\x75\x00\x80\x8B\x75"
-     "\x00\x80"
-     "\xAE\xD9\x02\x00\xAF\xD9\x02\x00\xAC\xD9\x02\x00\xAD\xD9\x02\x00\x80\xE9"
-     "\x00\x00"
-     "\x81\xE9\x00\x00\x82\x05\x00\x00\x8F\x75\x00\x80\xBE\x00\x0C\x00\x97\x75"
-     "\x00\x80"
-     "\x87\xA8\x03\x00\x86\xBC\x0C\x00\x9B\x75\x00\x80\x9F\x75\x00\x80\xA3\x75"
-     "\x00\x80"
-     "\xA7\x75\x00\x80\xAB\x75\x00\x80\xAF\x75\x00\x80\xB3\x75\x00\x80\xB7\x75"
-     "\x00\x80"
-     "\xBB\x75\x00\x80\xBF\x75\x00\x80\xC3\x75\x00\x80\xC7\x75\x00\x80\xE3\x60"
-     "\x01\x00"
-     "\xCB\x75\x00\x80\xE1\xA0\x01\x00\xCF\x75\x00\x80\xEF\x88\x02\x00\xD3\x75"
-     "\x00\x80"
-     "\xD7\x75\x00\x80\xDB\x75\x00\x80\xDF\x75\x00\x80\xE3\x75\x00\x80\xE7\x75"
-     "\x00\x80"
-     "\xEB\x75\x00\x80\xA8\x69\x02\x00\xA9\x69\x02\x00\xAA\x79\x02\x00\xAB\x79"
-     "\x02\x00"
-     "\xAC\x69\x02\x00\xAD\x69\x02\x00\xAE\xBD\x02\x00\xAF\xB5\x02\x00\xBE\xAC"
-     "\x0C\x00"
-     "\xEF\x75\x00\x80\xF3\x75\x00\x80\xF7\x75\x00\x80\x80\x1D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\xA9\x00\x00\xFB\x75\x00\x80\xB8\x51\x01\x00\xB9\x59\x01\x00\xBA\x61"
-     "\x01\x00"
-     "\xBB\x61\x01\x00\xBC\x19\x01\x00\xBD\x19\x01\x00\xBE\x0D\x01\x00\xBF\x05"
-     "\x01\x00"
-     "\xB0\xCD\x02\x00\xB1\xD5\x02\x00\xB2\xDD\x02\x00\xB3\xD5\x02\x00\xB4\xCD"
-     "\x02\x00"
-     "\xB5\x71\x01\x00\xB6\x71\x01\x00\xB7\x71\x01\x00\xE3\xC4\x00\x00\xE1\x7C"
-     "\x07\x00"
-     "\xE1\x78\x06\x00\xE3\xBC\x06\x00\xFF\x75\x00\x80\x84\x18\x0D\x00\x86\xB8"
-     "\x0C\x00"
-     "\x87\x3C\x0D\x00\xBE\x2C\x0F\x00\x07\x76\x00\x80\x0B\x76\x00\x80\x0F\x76"
-     "\x00\x80"
-     "\xEF\x10\x00\x00\x13\x76\x00\x80\x17\x76\x00\x80\xEF\x74\x06\x00\x1B\x76"
-     "\x00\x80"
-     "\x1F\x76\x00\x80\x23\x76\x00\x80\xB3\xBD\x02\x00\x27\x76\x00\x80\xB5\xAD"
-     "\x02\x00"
-     "\xB6\xA5\x02\x00\x2B\x76\x00\x80\x2F\x76\x00\x80\x33\x76\x00\x80\xBA\x45"
-     "\x02\x00"
-     "\xBB\x5D\x02\x00\xBC\x45\x02\x00\xBD\x4D\x02\x00\xBE\x45\x02\x00\xBF\xF9"
-     "\x01\x00"
-     "\x93\x75\x00\x80\xA5\x7D\x0D\x00\xA6\x75\x0D\x00\x03\x76\x00\x80\x37\x76"
-     "\x00\x80"
-     "\x3B\x76\x00\x80\x3F\x76\x00\x80\xA3\x6D\x0D\x00\xAC\x95\x0D\x00\xAD\x9D"
-     "\x0D\x00"
-     "\xAE\x95\x0D\x00\xAF\x29\x0E\x00\x43\x76\x00\x80\x47\x76\x00\x80\xAA\x95"
-     "\x0D\x00"
-     "\xAB\x8D\x0D\x00\xB3\xE5\x0E\x00\x4B\x76\x00\x80\x4F\x76\x00\x80\x53\x76"
-     "\x00\x80"
-     "\x57\x76\x00\x80\xB6\xE5\x0E\x00\xB5\xF5\x0E\x00\x5B\x76\x00\x80\xBB\xA1"
-     "\x0E\x00"
-     "\xBA\xD9\x0E\x00\x5F\x76\x00\x80\x63\x76\x00\x80\xBF\xA5\x0E\x00\xBE\xB9"
-     "\x0E\x00"
-     "\xBD\xB1\x0E\x00\xBC\xB9\x0E\x00\xA8\x15\x0E\x00\xA9\x25\x0E\x00\xAA\x2D"
-     "\x0E\x00"
-     "\xAB\x25\x0E\x00\xAC\x3D\x0E\x00\xAD\x25\x0E\x00\xAE\x2D\x0E\x00\xAF\x25"
-     "\x0E\x00"
-     "\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\x67\x76\x00\x80\x6B\x76"
-     "\x00\x80"
-     "\x6F\x76\x00\x80\x84\x30\x03\x00\x73\x76\x00\x80\xB8\x29\x0E\x00\xB9\x29"
-     "\x0E\x00"
-     "\xBA\x39\x0E\x00\xBB\x39\x0E\x00\xBC\x29\x0E\x00\xBD\x29\x0E\x00\xBE\xFD"
-     "\x0F\x00"
-     "\xBF\xF5\x0F\x00\xB0\x5D\x0E\x00\xB1\x25\x0E\x00\xB2\x2D\x0E\x00\xB3\x25"
-     "\x0E\x00"
-     "\xB4\x3D\x0E\x00\xB5\x21\x0E\x00\xB6\x25\x0E\x00\xB7\x19\x0E\x00\xA3\xA5"
-     "\x0F\x00"
-     "\x77\x76\x00\x80\x86\x28\x01\x00\x87\x4C\x01\x00\x7B\x76\x00\x80\xA6\xA5"
-     "\x0F\x00"
-     "\xA5\xB5\x0F\x00\x7F\x76\x00\x80\xAB\xE1\x0F\x00\xAA\x99\x0F\x00\x83\x76"
-     "\x00\x80"
-     "\x87\x76\x00\x80\xAF\xE5\x0F\x00\xAE\xF9\x0F\x00\xAD\xF1\x0F\x00\xAC\xF9"
-     "\x0F\x00"
-     "\x8B\x76\x00\x80\xB3\xE9\x0E\x00\x8F\x76\x00\x80\x93\x76\x00\x80\xB6\x91"
-     "\x0E\x00"
-     "\x97\x76\x00\x80\x9B\x76\x00\x80\xB5\xE5\x0E\x00\xBA\xB1\x0E\x00\xBB\xB9"
-     "\x0E\x00"
-     "\x9F\x76\x00\x80\xA3\x76\x00\x80\xBE\x61\x01\x00\xBF\x61\x01\x00\xBC\x99"
-     "\x0E\x00"
-     "\xBD\x99\x0E\x00\xA8\x25\x0E\x00\xA9\x2D\x0E\x00\xAA\x25\x0E\x00\xAB\x39"
-     "\x0E\x00"
-     "\xAC\x29\x0E\x00\xAD\x55\x0E\x00\xAE\x5D\x0E\x00\xAF\x55\x0E\x00\xA7\x76"
-     "\x00\x80"
-     "\xAB\x76\x00\x80\xAF\x76\x00\x80\xB3\x76\x00\x80\xB7\x76\x00\x80\xBB\x76"
-     "\x00\x80"
-     "\xBF\x76\x00\x80\xC3\x76\x00\x80\xB8\xF5\x01\x00\xB9\x81\x01\x00\xBA\x81"
-     "\x01\x00"
-     "\xBB\x81\x01\x00\xBC\x81\x01\x00\xBD\x89\x01\x00\xBE\xB1\x01\x00\xBF\xB1"
-     "\x01\x00"
-     "\xB0\x31\x0E\x00\xB1\x39\x0E\x00\xB2\x09\x0E\x00\xB3\x09\x0E\x00\xB4\xE1"
-     "\x01\x00"
-     "\xB5\xE1\x01\x00\xB6\xE1\x01\x00\xB7\xCD\x01\x00\xA3\xAD\x0D\x00\xC7\x76"
-     "\x00\x80"
-     "\xCB\x76\x00\x80\xCF\x76\x00\x80\xD3\x76\x00\x80\xA6\xD5\x0D\x00\xA5\xA1"
-     "\x0D\x00"
-     "\xD7\x76\x00\x80\xAB\xFD\x0D\x00\xAA\xF5\x0D\x00\xDB\x76\x00\x80\xDF\x76"
-     "\x00\x80"
-     "\xAF\x25\x02\x00\xAE\x25\x02\x00\xAD\xDD\x0D\x00\xAC\xDD\x0D\x00\x80\x5D"
-     "\x00\x00"
-     "\x81\x6D\x00\x00\x82\x65\x00\x00\xB3\x51\x03\x00\xBE\x9C\x03\x00\xB5\x79"
-     "\x03\x00"
-     "\xB6\x19\x03\x00\xE7\x76\x00\x80\x84\xE0\x02\x00\xEB\x76\x00\x80\xBA\x3D"
-     "\x03\x00"
-     "\xBB\x35\x03\x00\xBC\x19\x03\x00\xBD\x19\x03\x00\xBE\xD9\x03\x00\xBF\xD9"
-     "\x03\x00"
-     "\xA8\x85\x03\x00\xA9\x95\x03\x00\xAA\x95\x03\x00\xAB\xA5\x03\x00\xAC\xBD"
-     "\x03\x00"
-     "\xAD\xD5\x03\x00\xAE\xD1\x03\x00\xAF\xD1\x03\x00\x86\x00\x04\x00\x87\x34"
-     "\x03\x00"
-     "\xBF\xA0\x33\x00\xEF\x76\x00\x80\xF3\x76\x00\x80\xF7\x76\x00\x80\xFB\x76"
-     "\x00\x80"
-     "\xFF\x76\x00\x80\xB8\x71\x03\x00\xB9\x71\x03\x00\xBA\x71\x03\x00\xBB\x71"
-     "\x03\x00"
-     "\xBC\xD5\x00\x00\xBD\xDD\x00\x00\xBE\xD5\x00\x00\xBF\xCD\x00\x00\xB0\xB5"
-     "\x03\x00"
-     "\xB1\xBD\x03\x00\xB2\x81\x03\x00\xB3\x81\x03\x00\xB4\x51\x03\x00\xB5\x51"
-     "\x03\x00"
-     "\xB6\x51\x03\x00\xB7\x51\x03\x00\xEF\xA8\x03\x00\x03\x77\x00\x80\x07\x77"
-     "\x00\x80"
-     "\x0B\x77\x00\x80\x84\x1C\x02\x00\x0F\x77\x00\x80\x13\x77\x00\x80\x17\x77"
-     "\x00\x80"
-     "\xBE\x2C\x05\x00\x1B\x77\x00\x80\x1F\x77\x00\x80\x23\x77\x00\x80\xE3\x40"
-     "\x03\x00"
-     "\x27\x77\x00\x80\xE1\x28\x00\x00\x2B\x77\x00\x80\xA3\x5D\x02\x00\x2F\x77"
-     "\x00\x80"
-     "\x33\x77\x00\x80\x37\x77\x00\x80\x3B\x77\x00\x80\xA6\x15\x02\x00\xA5\x75"
-     "\x02\x00"
-     "\x3F\x77\x00\x80\xAB\x39\x02\x00\xAA\x31\x02\x00\x43\x77\x00\x80\x47\x77"
-     "\x00\x80"
-     "\xAF\xD5\x02\x00\xAE\xD5\x02\x00\xAD\x15\x02\x00\xAC\x15\x02\x00\xE3\x28"
-     "\x01\x00"
-     "\xE1\x00\x0F\x00\xE1\x08\x0E\x00\xE3\x08\x0E\x00\x80\xFD\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x39\x00\x00\x4B\x77\x00\x80\x4F\x77\x00\x80\x57\x77\x00\x80\x5B\x77"
-     "\x00\x80"
-     "\x5F\x77\x00\x80\xEF\xE8\x0E\x00\x63\x77\x00\x80\x67\x77\x00\x80\xEF\x64"
-     "\x0E\x00"
-     "\xB3\x6D\x01\x00\x6B\x77\x00\x80\x86\xE8\x04\x00\x87\x0C\x05\x00\x6F\x77"
-     "\x00\x80"
-     "\xB6\x6D\x01\x00\xB5\x6D\x01\x00\x73\x77\x00\x80\xBB\xED\x00\x00\xBA\xED"
-     "\x00\x00"
-     "\x77\x77\x00\x80\x7B\x77\x00\x80\xBF\xD5\x00\x00\xBE\xE9\x00\x00\xBD\xE1"
-     "\x00\x00"
-     "\xBC\xE9\x00\x00\xA8\x5D\x06\x00\xA9\x61\x06\x00\xAA\xA5\x06\x00\xAB\xBD"
-     "\x06\x00"
-     "\xAC\xA5\x06\x00\xAD\xAD\x06\x00\xAE\xA5\x06\x00\xAF\x19\x07\x00\x53\x77"
-     "\x00\x80"
-     "\x7F\x77\x00\x80\x83\x77\x00\x80\x87\x77\x00\x80\x8B\x77\x00\x80\x8F\x77"
-     "\x00\x80"
-     "\x93\x77\x00\x80\x97\x77\x00\x80\xB8\x75\x07\x00\xB9\x79\x07\x00\xBA\x0D"
-     "\x07\x00"
-     "\xBB\x05\x07\x00\xBC\x1D\x07\x00\xBD\x05\x07\x00\xBE\x05\x07\x00\xBF\x35"
-     "\x07\x00"
-     "\xB0\x69\x07\x00\xB1\x69\x07\x00\xB2\x7D\x07\x00\xB3\x75\x07\x00\xB4\x6D"
-     "\x07\x00"
-     "\xB5\x51\x07\x00\xB6\x51\x07\x00\xB7\x51\x07\x00\xA3\x2D\x06\x00\x9B\x77"
-     "\x00\x80"
-     "\x9F\x77\x00\x80\xA3\x77\x00\x80\xA7\x77\x00\x80\xA6\x2D\x06\x00\xA5\x2D"
-     "\x06\x00"
-     "\xAB\x77\x00\x80\xAB\xAD\x07\x00\xAA\xAD\x07\x00\xAF\x77\x00\x80\xB3\x77"
-     "\x00\x80"
-     "\xAF\x95\x07\x00\xAE\xA9\x07\x00\xAD\xA1\x07\x00\xAC\xA9\x07\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x1D\x00\x00\xB7\x77\x00\x80\xBB\x77\x00\x80\xBF\x77"
-     "\x00\x80"
-     "\x84\x54\x03\x00\xBE\x5C\x00\x00\xC3\x77\x00\x80\xC7\x77\x00\x80\x86\xE8"
-     "\x00\x00"
-     "\x87\x4C\x03\x00\xCB\x77\x00\x80\xCF\x77\x00\x80\xD3\x77\x00\x80\xD7\x77"
-     "\x00\x80"
-     "\xDB\x77\x00\x80\xE3\x04\x04\x00\xDF\x77\x00\x80\xE1\xB4\x05\x00\xE3\x77"
-     "\x00\x80"
-     "\xE7\x77\x00\x80\xEB\x77\x00\x80\xEF\x77\x00\x80\xF3\x77\x00\x80\xF7\x77"
-     "\x00\x80"
-     "\xFB\x77\x00\x80\xFF\x77\x00\x80\x03\x78\x00\x80\xEF\xEC\x04\x00\x07\x78"
-     "\x00\x80"
-     "\x0B\x78\x00\x80\xB3\xED\x06\x00\x0F\x78\x00\x80\x13\x78\x00\x80\x17\x78"
-     "\x00\x80"
-     "\x1B\x78\x00\x80\xB6\x91\x06\x00\xB5\xE1\x06\x00\x1F\x78\x00\x80\xBB\x8D"
-     "\x06\x00"
-     "\xBA\x8D\x06\x00\x23\x78\x00\x80\x27\x78\x00\x80\xBF\x41\x01\x00\xBE\x59"
-     "\x01\x00"
-     "\xBD\x51\x01\x00\xBC\x59\x01\x00\xA8\x25\x06\x00\xA9\x2D\x06\x00\xAA\x25"
-     "\x06\x00"
-     "\xAB\x39\x06\x00\xAC\x29\x06\x00\xAD\x51\x06\x00\xAE\x49\x06\x00\xAF\x41"
-     "\x06\x00"
-     "\x80\xCD\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x2B\x78\x00\x80\x2F\x78"
-     "\x00\x80"
-     "\x84\x2C\x01\x00\xBE\x34\x00\x00\x37\x78\x00\x80\xB8\xFD\x01\x00\xB9\x41"
-     "\x01\x00"
-     "\xBA\x41\x01\x00\xBB\x41\x01\x00\xBC\x41\x01\x00\xBD\x49\x01\x00\xBE\x71"
-     "\x01\x00"
-     "\xBF\x71\x01\x00\xB0\x09\x06\x00\xB1\x09\x06\x00\xB2\xCD\x01\x00\xB3\xC5"
-     "\x01\x00"
-     "\xB4\xDD\x01\x00\xB5\xC5\x01\x00\xB6\xCD\x01\x00\xB7\xC5\x01\x00\x86\xA0"
-     "\x3C\x00"
-     "\x87\x44\x03\x00\x3B\x78\x00\x80\xA3\xA1\x05\x00\x3F\x78\x00\x80\xA5\xAD"
-     "\x05\x00"
-     "\xA6\xDD\x05\x00\x43\x78\x00\x80\x47\x78\x00\x80\xBE\x28\x3C\x00\xAA\xC1"
-     "\x05\x00"
-     "\xAB\xC1\x05\x00\xAC\x15\x02\x00\xAD\x1D\x02\x00\xAE\x15\x02\x00\xAF\x0D"
-     "\x02\x00"
-     "\xB6\x41\x03\x00\x4B\x78\x00\x80\x4F\x78\x00\x80\xB5\xB1\x02\x00\x53\x78"
-     "\x00\x80"
-     "\xB3\xA1\x02\x00\x57\x78\x00\x80\x5B\x78\x00\x80\xBE\x45\x03\x00\xBF\x4D"
-     "\x03\x00"
-     "\xBC\x75\x03\x00\xBD\x4D\x03\x00\xBA\x65\x03\x00\xBB\x6D\x03\x00\x5F\x78"
-     "\x00\x80"
-     "\x63\x78\x00\x80\x67\x78\x00\x80\x6B\x78\x00\x80\xE3\x76\x00\x80\x6F\x78"
-     "\x00\x80"
-     "\x73\x78\x00\x80\x77\x78\x00\x80\x7B\x78\x00\x80\x7F\x78\x00\x80\xA8\x45"
-     "\x02\x00"
-     "\xA9\x55\x02\x00\xAA\x5D\x02\x00\xAB\x55\x02\x00\xAC\x4D\x02\x00\xAD\xB5"
-     "\x03\x00"
-     "\xAE\xB1\x03\x00\xAF\xAD\x03\x00\xB0\xD5\x03\x00\xB1\xDD\x03\x00\xB2\xD5"
-     "\x03\x00"
-     "\xB3\xED\x03\x00\xB4\xF5\x03\x00\xB5\xFD\x03\x00\xB6\xF5\x03\x00\xB7\xED"
-     "\x03\x00"
-     "\xB8\xD9\x03\x00\xB9\xD9\x03\x00\xBA\xAD\x03\x00\xBB\xA5\x03\x00\xBC\xBD"
-     "\x03\x00"
-     "\xBD\xA5\x03\x00\xBE\xA5\x03\x00\xBF\x95\x03\x00\xA3\xF5\x03\x00\x83\x78"
-     "\x00\x80"
-     "\x87\x78\x00\x80\x8B\x78\x00\x80\x8F\x78\x00\x80\xA6\x15\x02\x00\xA5\xE5"
-     "\x03\x00"
-     "\x93\x78\x00\x80\xAB\x39\x02\x00\xAA\x31\x02\x00\x97\x78\x00\x80\x9B\x78"
-     "\x00\x80"
-     "\xAF\x19\x02\x00\xAE\x11\x02\x00\xAD\x19\x02\x00\xAC\x21\x02\x00\x80\x69"
-     "\x00\x00"
-     "\x81\x69\x00\x00\x82\x05\x00\x00\x9F\x78\x00\x80\xA7\x78\x00\x80\xAB\x78"
-     "\x00\x80"
-     "\xAF\x78\x00\x80\xEF\x1C\x00\x00\x84\x6C\x02\x00\xE1\x94\x01\x00\xB3\x78"
-     "\x00\x80"
-     "\xE3\xC8\x00\x00\xB7\x78\x00\x80\xBB\x78\x00\x80\x86\x58\x3C\x00\x87\x50"
-     "\x3D\x00"
-     "\xBF\x78\x00\x80\xC3\x78\x00\x80\xC7\x78\x00\x80\x84\x84\x3D\x00\xCB\x78"
-     "\x00\x80"
-     "\xCF\x78\x00\x80\xD3\x78\x00\x80\xEF\xB8\x01\x00\xBE\x6C\x3C\x00\xE1\x74"
-     "\x06\x00"
-     "\xD7\x78\x00\x80\xE3\x64\x01\x00\xDB\x78\x00\x80\xDF\x78\x00\x80\xE3\x78"
-     "\x00\x80"
-     "\xE7\x78\x00\x80\xB3\xD1\x01\x00\xEB\x78\x00\x80\xEF\x78\x00\x80\xF3\x78"
-     "\x00\x80"
-     "\xF7\x78\x00\x80\xB6\x91\x01\x00\xB5\xF9\x01\x00\xFB\x78\x00\x80\xBB\xBD"
-     "\x01\x00"
-     "\xBA\xBD\x01\x00\xFF\x78\x00\x80\x03\x79\x00\x80\xBF\x75\x01\x00\xBE\x75"
-     "\x01\x00"
-     "\xBD\x85\x01\x00\xBC\x85\x01\x00\xA8\xBD\x3D\x00\xA9\x0D\x3E\x00\xAA\x19"
-     "\x3E\x00"
-     "\xAB\x11\x3E\x00\xAC\x31\x3E\x00\xAD\x51\x3E\x00\xAE\x51\x3E\x00\xAF\x4D"
-     "\x3E\x00"
-     "\xA3\x78\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\x07\x79"
-     "\x00\x80"
-     "\x0B\x79\x00\x80\x0F\x79\x00\x80\x13\x79\x00\x80\xB8\xD5\x3E\x00\xB9\xDD"
-     "\x3E\x00"
-     "\xBA\xD5\x3E\x00\xBB\x49\x3F\x00\xBC\x59\x3F\x00\xBD\x59\x3F\x00\xBE\x49"
-     "\x3F\x00"
-     "\xBF\x41\x3F\x00\xB0\x39\x3E\x00\xB1\x39\x3E\x00\xB2\x11\x3E\x00\xB3\x11"
-     "\x3E\x00"
-     "\xB4\xF1\x3E\x00\xB5\xF1\x3E\x00\xB6\xF5\x3E\x00\xB7\xED\x3E\x00\xA3\x91"
-     "\x3E\x00"
-     "\x17\x79\x00\x80\x86\x28\x00\x00\x87\xC0\x03\x00\x1B\x79\x00\x80\xA6\xD1"
-     "\x3E\x00"
-     "\xA5\xB9\x3E\x00\x1F\x79\x00\x80\xAB\xFD\x3E\x00\xAA\xFD\x3E\x00\x23\x79"
-     "\x00\x80"
-     "\x27\x79\x00\x80\xAF\x35\x3E\x00\xAE\x35\x3E\x00\xAD\xC5\x3E\x00\xAC\xC5"
-     "\x3E\x00"
-     "\x2B\x79\x00\x80\xB3\x9D\x3F\x00\x2F\x79\x00\x80\x33\x79\x00\x80\xB6\xA5"
-     "\x3F\x00"
-     "\x37\x79\x00\x80\x3B\x79\x00\x80\xB5\xAD\x3F\x00\xBA\x69\x3F\x00\xBB\x75"
-     "\x3F\x00"
-     "\x3F\x79\x00\x80\x43\x79\x00\x80\xBE\x59\x3F\x00\xBF\x45\x3F\x00\xBC\x6D"
-     "\x3F\x00"
-     "\xBD\x65\x3F\x00\x47\x79\x00\x80\x4B\x79\x00\x80\x4F\x79\x00\x80\x53\x79"
-     "\x00\x80"
-     "\xE3\x60\x3C\x00\x57\x79\x00\x80\xE1\x00\x3D\x00\x5B\x79\x00\x80\xEF\xFC"
-     "\x3D\x00"
-     "\x5F\x79\x00\x80\x63\x79\x00\x80\x67\x79\x00\x80\x6B\x79\x00\x80\x6F\x79"
-     "\x00\x80"
-     "\x73\x79\x00\x80\x77\x79\x00\x80\xA3\x19\x3E\x00\x82\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x80\x71\x00\x00\x7B\x79\x00\x80\xA6\x21\x3E\x00\xA5\x29\x3E\x00\x7F\x79"
-     "\x00\x80"
-     "\xAB\xF1\x3E\x00\xAA\xED\x3E\x00\x84\x24\x01\x00\xBE\x24\x01\x00\xAF\xC1"
-     "\x3E\x00"
-     "\xAE\xDD\x3E\x00\xAD\xE1\x3E\x00\xAC\xE9\x3E\x00\xA8\xD1\x3E\x00\xA9\xD1"
-     "\x3E\x00"
-     "\xAA\xD1\x3E\x00\xAB\xE5\x3E\x00\xAC\xE1\x3E\x00\xAD\xE1\x3E\x00\xAE\x19"
-     "\x3E\x00"
-     "\xAF\x19\x3E\x00\x86\x00\x00\x00\x87\x84\x00\x00\x87\x79\x00\x80\x8B\x79"
-     "\x00\x80"
-     "\x8F\x79\x00\x80\x93\x79\x00\x80\x97\x79\x00\x80\x9B\x79\x00\x80\xB8\x7D"
-     "\x3E\x00"
-     "\xB9\x01\x3E\x00\xBA\x01\x3E\x00\xBB\x01\x3E\x00\xBC\x01\x3E\x00\xBD\x09"
-     "\x3E\x00"
-     "\xBE\x31\x3E\x00\xBF\x31\x3E\x00\xB0\x69\x3E\x00\xB1\x75\x3E\x00\xB2\x7D"
-     "\x3E\x00"
-     "\xB3\x75\x3E\x00\xB4\x59\x3E\x00\xB5\x45\x3E\x00\xB6\x4D\x3E\x00\xB7\x45"
-     "\x3E\x00"
-     "\xA8\x85\x02\x00\xA9\x95\x02\x00\xAA\x95\x02\x00\xAB\xA5\x02\x00\xAC\xBD"
-     "\x02\x00"
-     "\xAD\xD5\x02\x00\xAE\xD1\x02\x00\xAF\xD1\x02\x00\x9F\x79\x00\x80\xA3\x79"
-     "\x00\x80"
-     "\xA7\x79\x00\x80\xBF\x24\xE6\x01\xAB\x79\x00\x80\xAF\x79\x00\x80\xB3\x79"
-     "\x00\x80"
-     "\xB7\x79\x00\x80\xB8\x55\x03\x00\xB9\x59\x03\x00\xBA\x6D\x03\x00\xBB\x65"
-     "\x03\x00"
-     "\xBC\x7D\x03\x00\xBD\x65\x03\x00\xBE\x6D\x03\x00\xBF\x65\x03\x00\xB0\xB5"
-     "\x02\x00"
-     "\xB1\xBD\x02\x00\xB2\x81\x02\x00\xB3\x81\x02\x00\xB4\x71\x03\x00\xB5\x71"
-     "\x03\x00"
-     "\xB6\x71\x03\x00\xB7\x71\x03\x00\xB3\x1D\x02\x00\xBB\x79\x00\x80\xBF\x79"
-     "\x00\x80"
-     "\xC3\x79\x00\x80\x84\x88\x03\x00\xB6\x55\x02\x00\xB5\x35\x02\x00\x33\x78"
-     "\x00\x80"
-     "\xBB\x79\x02\x00\xBA\x71\x02\x00\xC7\x79\x00\x80\xCB\x79\x00\x80\xBF\xB5"
-     "\x03\x00"
-     "\xBE\xB5\x03\x00\xBD\x55\x02\x00\xBC\x55\x02\x00\xCF\x79\x00\x80\xA3\x59"
-     "\x02\x00"
-     "\xD3\x79\x00\x80\xD7\x79\x00\x80\xA6\x11\x02\x00\xDB\x79\x00\x80\xDF\x79"
-     "\x00\x80"
-     "\xA5\x71\x02\x00\xAA\x35\x02\x00\xAB\x3D\x02\x00\xE3\x79\x00\x80\xE7\x79"
-     "\x00\x80"
-     "\xAE\xF1\x03\x00\xAF\xF1\x03\x00\xAC\x11\x02\x00\xAD\x11\x02\x00\xA8\xA9"
-     "\x02\x00"
-     "\xA9\xA9\x02\x00\xAA\xB9\x02\x00\xAB\xB9\x02\x00\xAC\xA9\x02\x00\xAD\xA9"
-     "\x02\x00"
-     "\xAE\x39\x01\x00\xAF\x39\x01\x00\x80\xCD\x01\x00\x81\x09\x00\x00\x82\x19"
-     "\x00\x00"
-     "\xEB\x79\x00\x80\xEF\x79\x00\x80\xBE\xB8\x05\x00\xF7\x79\x00\x80\xFB\x79"
-     "\x00\x80"
-     "\xB8\xE9\x01\x00\xB9\xE9\x01\x00\xBA\x89\x01\x00\xBB\x85\x01\x00\xBC\x9D"
-     "\x01\x00"
-     "\xBD\x81\x01\x00\xBE\x81\x01\x00\xBF\xB5\x01\x00\xB0\x49\x01\x00\xB1\x55"
-     "\x01\x00"
-     "\xB2\x5D\x01\x00\xB3\x55\x01\x00\xB4\x4D\x01\x00\xB5\xF1\x01\x00\xB6\xF1"
-     "\x01\x00"
-     "\xB7\xF1\x01\x00\xEF\x14\x00\x00\xFF\x79\x00\x80\x86\xA8\x05\x00\x87\xDC"
-     "\x05\x00"
-     "\x03\x7A\x00\x80\x84\x58\x04\x00\x07\x7A\x00\x80\xEF\xC4\x3E\x00\x0B\x7A"
-     "\x00\x80"
-     "\xE1\xC4\x3E\x00\x0F\x7A\x00\x80\xE3\x30\x3E\x00\xE3\xC8\x00\x00\x13\x7A"
-     "\x00\x80"
-     "\xE1\x28\x01\x00\x17\x7A\x00\x80\xB6\x7D\x02\x00\x1B\x7A\x00\x80\x1F\x7A"
-     "\x00\x80"
-     "\xB5\x75\x02\x00\x23\x7A\x00\x80\xB3\x65\x02\x00\x27\x7A\x00\x80\x2B\x7A"
-     "\x00\x80"
-     "\xBE\xDD\x01\x00\xBF\x61\x01\x00\xBC\xDD\x01\x00\xBD\xD5\x01\x00\xBA\xD9"
-     "\x01\x00"
-     "\xBB\xC5\x01\x00\x2F\x7A\x00\x80\x33\x7A\x00\x80\xA3\xB1\x05\x00\xF3\x79"
-     "\x00\x80"
-     "\x37\x7A\x00\x80\x3B\x7A\x00\x80\x3F\x7A\x00\x80\xA6\xA9\x05\x00\xA5\xA1"
-     "\x05\x00"
-     "\x43\x7A\x00\x80\xAB\x11\x06\x00\xAA\x0D\x06\x00\x47\x7A\x00\x80\x4B\x7A"
-     "\x00\x80"
-     "\xAF\xB5\x06\x00\xAE\x09\x06\x00\xAD\x01\x06\x00\xAC\x09\x06\x00\x4F\x7A"
-     "\x00\x80"
-     "\x53\x7A\x00\x80\x57\x7A\x00\x80\x5B\x7A\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x5F\x7A\x00\x80\xBE\x6C\x03\x00\x63\x7A\x00\x80\x86\xC8"
-     "\x00\x00"
-     "\x87\x1C\x03\x00\x67\x7A\x00\x80\x6B\x7A\x00\x80\x6F\x7A\x00\x80\x73\x7A"
-     "\x00\x80"
-     "\xA8\xB9\x07\x00\xA9\xB9\x07\x00\xAA\x0D\x07\x00\xAB\x1D\x07\x00\xAC\x09"
-     "\x07\x00"
-     "\xAD\x35\x07\x00\xAE\x31\x07\x00\xAF\x29\x07\x00\x84\xA8\x03\x00\x77\x7A"
-     "\x00\x80"
-     "\x7B\x7A\x00\x80\x7F\x7A\x00\x80\x83\x7A\x00\x80\x87\x7A\x00\x80\x8B\x7A"
-     "\x00\x80"
-     "\x8F\x7A\x00\x80\xB8\xC9\x00\x00\xB9\xC9\x00\x00\xBA\xD9\x00\x00\xBB\xD1"
-     "\x00\x00"
-     "\xBC\xF9\x00\x00\xBD\xF9\x00\x00\xBE\x99\x00\x00\xBF\x99\x00\x00\xB0\x5D"
-     "\x07\x00"
-     "\xB1\x21\x07\x00\xB2\x21\x07\x00\xB3\x3D\x07\x00\xB4\x29\x07\x00\xB5\x29"
-     "\x07\x00"
-     "\xB6\x01\x07\x00\xB7\x01\x07\x00\xB3\x85\x06\x00\x93\x7A\x00\x80\x97\x7A"
-     "\x00\x80"
-     "\x9B\x7A\x00\x80\x9F\x7A\x00\x80\xB6\xB5\x06\x00\xB5\x81\x06\x00\xA3\x7A"
-     "\x00\x80"
-     "\xBB\xE5\x06\x00\xBA\x99\x06\x00\xA7\x7A\x00\x80\xAB\x7A\x00\x80\xBF\xED"
-     "\x06\x00"
-     "\xBE\xED\x06\x00\xBD\xE9\x06\x00\xBC\xF5\x06\x00\xAF\x7A\x00\x80\xB3\x7A"
-     "\x00\x80"
-     "\xB7\x7A\x00\x80\xBB\x7A\x00\x80\xBF\x7A\x00\x80\xC3\x7A\x00\x80\xC7\x7A"
-     "\x00\x80"
-     "\xEF\x90\x04\x00\xCB\x7A\x00\x80\xE1\xDC\x06\x00\xCF\x7A\x00\x80\xE3\xEC"
-     "\x05\x00"
-     "\x80\x29\x00\x00\x81\x15\x00\x00\x82\x11\x00\x00\xBE\x7C\x01\x00\xA3\x05"
-     "\x06\x00"
-     "\xD7\x7A\x00\x80\x86\x28\x00\x00\x87\x4C\x01\x00\xDB\x7A\x00\x80\xA6\x35"
-     "\x06\x00"
-     "\xA5\x01\x06\x00\xDF\x7A\x00\x80\xAB\x65\x06\x00\xAA\x19\x06\x00\xE3\x7A"
-     "\x00\x80"
-     "\xE7\x7A\x00\x80\xAF\x6D\x06\x00\xAE\x6D\x06\x00\xAD\x69\x06\x00\xAC\x75"
-     "\x06\x00"
-     "\xEB\x7A\x00\x80\xB3\xBD\x01\x00\xEF\x7A\x00\x80\xF3\x7A\x00\x80\xB6\x79"
-     "\x01\x00"
-     "\xF7\x7A\x00\x80\xFB\x7A\x00\x80\xB5\x79\x01\x00\xBA\x55\x01\x00\xBB\x5D"
-     "\x01\x00"
-     "\xFF\x7A\x00\x80\x03\x7B\x00\x80\xBE\xF9\x00\x00\xBF\xF9\x00\x00\xBC\x45"
-     "\x01\x00"
-     "\xBD\xF9\x00\x00\xA8\x71\x02\x00\xA9\x71\x02\x00\xAA\x71\x02\x00\xAB\x71"
-     "\x02\x00"
-     "\xAC\xB5\x02\x00\xAD\xBD\x02\x00\xAE\xB5\x02\x00\xAF\xAD\x02\x00\x84\xEC"
-     "\x0C\x00"
-     "\x07\x7B\x00\x80\x0B\x7B\x00\x80\x0F\x7B\x00\x80\x13\x7B\x00\x80\x17\x7B"
-     "\x00\x80"
-     "\x1B\x7B\x00\x80\x1F\x7B\x00\x80\xB8\x69\x03\x00\xB9\x69\x03\x00\xBA\x09"
-     "\x03\x00"
-     "\xBB\x09\x03\x00\xBC\x19\x03\x00\xBD\x19\x03\x00\xBE\x09\x03\x00\xBF\x09"
-     "\x03\x00"
-     "\xB0\xD5\x02\x00\xB1\xDD\x02\x00\xB2\xD5\x02\x00\xB3\x69\x03\x00\xB4\x79"
-     "\x03\x00"
-     "\xB5\x79\x03\x00\xB6\x69\x03\x00\xB7\x61\x03\x00\x23\x7B\x00\x80\x27\x7B"
-     "\x00\x80"
-     "\x2B\x7B\x00\x80\xA3\xF5\x02\x00\x2F\x7B\x00\x80\xA5\x31\x02\x00\xA6\x31"
-     "\x02\x00"
-     "\x33\x7B\x00\x80\x37\x7B\x00\x80\x3B\x7B\x00\x80\xAA\x1D\x02\x00\xAB\x15"
-     "\x02\x00"
-     "\xAC\x0D\x02\x00\xAD\xB1\x03\x00\xAE\xB1\x03\x00\xAF\xB1\x03\x00\x80\x61"
-     "\x00\x00"
-     "\x81\x61\x00\x00\x82\x05\x00\x00\x3F\x7B\x00\x80\x86\xF0\x0C\x00\x87\x60"
-     "\x03\x00"
-     "\xBE\x10\x0C\x00\x47\x7B\x00\x80\x83\x79\x00\x80\x4B\x7B\x00\x80\x4F\x7B"
-     "\x00\x80"
-     "\x53\x7B\x00\x80\x57\x7B\x00\x80\x5B\x7B\x00\x80\x5F\x7B\x00\x80\x63\x7B"
-     "\x00\x80"
-     "\xA8\x85\x02\x00\xA9\x95\x02\x00\xAA\x95\x02\x00\xAB\xA5\x02\x00\xAC\xBD"
-     "\x02\x00"
-     "\xAD\xD5\x02\x00\xAE\xD1\x02\x00\xAF\xD1\x02\x00\x67\x7B\x00\x80\x6B\x7B"
-     "\x00\x80"
-     "\x6F\x7B\x00\x80\x73\x7B\x00\x80\x77\x7B\x00\x80\x7B\x7B\x00\x80\x7F\x7B"
-     "\x00\x80"
-     "\x83\x7B\x00\x80\xB8\x75\x01\x00\xB9\x7D\x01\x00\xBA\x75\x01\x00\xBB\xCD"
-     "\x01\x00"
-     "\xBC\xD5\x01\x00\xBD\xDD\x01\x00\xBE\xC9\x01\x00\xBF\xC1\x01\x00\xB0\xB5"
-     "\x02\x00"
-     "\xB1\xBD\x02\x00\xB2\x81\x02\x00\xB3\x81\x02\x00\xB4\x55\x01\x00\xB5\x5D"
-     "\x01\x00"
-     "\xB6\x55\x01\x00\xB7\x4D\x01\x00\xE1\x10\x06\x00\x84\x48\x0C\x00\xE3\x0C"
-     "\x06\x00"
-     "\x87\x7B\x00\x80\x84\x98\x0C\x00\x8B\x7B\x00\x80\x8F\x7B\x00\x80\x93\x7B"
-     "\x00\x80"
-     "\x97\x7B\x00\x80\x9B\x7B\x00\x80\x9F\x7B\x00\x80\xA3\x7B\x00\x80\x81\x75"
-     "\x00\x00"
-     "\x80\x75\x00\x00\xEF\x20\x01\x00\x82\x75\x00\x00\xA7\x7B\x00\x80\xAB\x7B"
-     "\x00\x80"
-     "\xAF\x7B\x00\x80\xBE\xC0\x0C\x00\x85\xB4\x0E\x00\xE1\x10\x02\x00\xEF\x5C"
-     "\x00\x00"
-     "\xE3\x00\x16\x00\xE1\x90\x01\x00\xB7\x7B\x00\x80\xE3\x58\x01\x00\xEF\x3C"
-     "\x07\x00"
-     "\xBB\x7B\x00\x80\xBF\x7B\x00\x80\x86\x00\x08\x00\x87\xB8\x0C\x00\xB3\x9D"
-     "\x0D\x00"
-     "\x43\x7B\x00\x80\xC3\x7B\x00\x80\xC7\x7B\x00\x80\xCB\x7B\x00\x80\xB6\xD5"
-     "\x0D\x00"
-     "\xB5\xB5\x0D\x00\xCF\x7B\x00\x80\xBB\xF9\x0D\x00\xBA\xF1\x0D\x00\xD3\x7B"
-     "\x00\x80"
-     "\xD7\x7B\x00\x80\xBF\x19\x0E\x00\xBE\x11\x0E\x00\xBD\xD5\x0D\x00\xBC\xD5"
-     "\x0D\x00"
-     "\xDB\x7B\x00\x80\xA3\xD9\x0D\x00\xDF\x7B\x00\x80\xE3\x7B\x00\x80\xA6\x91"
-     "\x0D\x00"
-     "\xE7\x7B\x00\x80\xEB\x7B\x00\x80\xA5\xF1\x0D\x00\xAA\xB5\x0D\x00\xAB\xBD"
-     "\x0D\x00"
-     "\xEF\x7B\x00\x80\xF3\x7B\x00\x80\xAE\x55\x0E\x00\xAF\x5D\x0E\x00\xAC\x91"
-     "\x0D\x00"
-     "\xAD\x91\x0D\x00\xA8\x5D\x0E\x00\xA9\x61\x0E\x00\xAA\x61\x0E\x00\xAB\x61"
-     "\x0E\x00"
-     "\xAC\x61\x0E\x00\xAD\x61\x0E\x00\xAE\x61\x0E\x00\xAF\x61\x0E\x00\xF7\x7B"
-     "\x00\x80"
-     "\xFB\x7B\x00\x80\xFF\x7B\x00\x80\x03\x7C\x00\x80\x07\x7C\x00\x80\x0B\x7C"
-     "\x00\x80"
-     "\x0F\x7C\x00\x80\x13\x7C\x00\x80\xB8\x4D\x0F\x00\xB9\x51\x0F\x00\xBA\x51"
-     "\x0F\x00"
-     "\xBB\x51\x0F\x00\xBC\x71\x0F\x00\xBD\x71\x0F\x00\xBE\x71\x0F\x00\xBF\x71"
-     "\x0F\x00"
-     "\xB0\xC1\x0F\x00\xB1\xC1\x0F\x00\xB2\xC1\x0F\x00\xB3\xC1\x0F\x00\xB4\xC1"
-     "\x0F\x00"
-     "\xB5\xC1\x0F\x00\xB6\xC1\x0F\x00\xB7\xC1\x0F\x00\xB3\xE9\x0F\x00\x17\x7C"
-     "\x00\x80"
-     "\xBE\x80\x01\x00\x1B\x7C\x00\x80\xB3\x7B\x00\x80\xB6\xE1\x0F\x00\xB5\xE9"
-     "\x0F\x00"
-     "\x1F\x7C\x00\x80\xBB\x05\x0E\x00\xBA\x05\x0E\x00\x27\x7C\x00\x80\x23\x7C"
-     "\x00\x80"
-     "\xBF\x05\x0E\x00\xBE\x05\x0E\x00\xBD\x15\x0E\x00\xBC\x15\x0E\x00\x81\x4D"
-     "\x00\x00"
-     "\x80\x41\x00\x00\xEF\x68\x0D\x00\x82\x51\x00\x00\x86\xF0\x07\x00\x87\xD4"
-     "\x01\x00"
-     "\x2B\x7C\x00\x80\x2F\x7C\x00\x80\x33\x7C\x00\x80\x84\x70\x01\x00\x37\x7C"
-     "\x00\x80"
-     "\x3B\x7C\x00\x80\xE1\xE0\x0E\x00\x3F\x7C\x00\x80\xE3\x48\x0D\x00\x43\x7C"
-     "\x00\x80"
-     "\xA3\x69\x0F\x00\x47\x7C\x00\x80\x4B\x7C\x00\x80\x4F\x7C\x00\x80\x53\x7C"
-     "\x00\x80"
-     "\xA6\x61\x0F\x00\xA5\x69\x0F\x00\x57\x7C\x00\x80\xAB\x85\x0E\x00\xAA\x85"
-     "\x0E\x00"
-     "\x5B\x7C\x00\x80\x5F\x7C\x00\x80\xAF\x85\x0E\x00\xAE\x85\x0E\x00\xAD\x95"
-     "\x0E\x00"
-     "\xAC\x95\x0E\x00\x63\x7C\x00\x80\xB3\x31\x0E\x00\x67\x7C\x00\x80\x6B\x7C"
-     "\x00\x80"
-     "\xB6\xC1\x01\x00\x6F\x7C\x00\x80\x73\x7C\x00\x80\xB5\xD1\x01\x00\xBA\xCD"
-     "\x01\x00"
-     "\xBB\xA5\x01\x00\x77\x7C\x00\x80\x7B\x7C\x00\x80\xBE\xA5\x01\x00\xBF\xAD"
-     "\x01\x00"
-     "\xBC\xB1\x01\x00\xBD\xB1\x01\x00\x8F\xDD\x26\x00\xA3\xF1\x0D\x00\x7F\x7C"
-     "\x00\x80"
-     "\x83\x7C\x00\x80\xA6\x01\x02\x00\x87\x7C\x00\x80\x8B\x7C\x00\x80\xA5\x11"
-     "\x02\x00"
-     "\xAA\x0D\x02\x00\xAB\x65\x02\x00\x8F\x7C\x00\x80\xBE\x20\x04\x00\xAE\x65"
-     "\x02\x00"
-     "\xAF\x6D\x02\x00\xAC\x71\x02\x00\xAD\x71\x02\x00\x9F\xA1\x0C\x00\x9E\x79"
-     "\x0A\x00"
-     "\x9D\x69\x0A\x00\x9C\xD1\x08\x00\x9B\xB1\x36\x00\x9A\x75\x36\x00\x99\xD1"
-     "\x34\x00"
-     "\x98\xE1\x32\x00\x97\x6D\x32\x00\x96\x65\x32\x00\x95\x35\x3F\x00\x94\x61"
-     "\x3E\x00"
-     "\x93\x71\x3E\x00\x92\x35\x3B\x00\x91\x71\x3A\x00\x90\x79\x3A\x00\x80\x95"
-     "\x00\x00"
-     "\x81\x9D\x00\x00\x82\xA1\x00\x00\x97\x7C\x00\x80\xEF\x44\x02\x00\xE1\x74"
-     "\x0F\x00"
-     "\x9B\x7C\x00\x80\xE3\x1C\x0F\x00\xE3\xD4\x01\x00\x9F\x7C\x00\x80\xE1\xE0"
-     "\x01\x00"
-     "\xEF\x5C\x01\x00\xA3\xB5\x02\x00\xA2\x41\x00\x00\xA1\xDD\x0E\x00\xA0\xB9"
-     "\x0E\x00"
-     "\xB5\xA9\x03\x00\xA3\x7C\x00\x80\x84\xC0\x04\x00\xB6\xA1\x03\x00\x86\xF0"
-     "\x05\x00"
-     "\x87\xE4\x04\x00\xB3\x85\x03\x00\xA7\x7C\x00\x80\xBD\x71\x03\x00\xBC\x69"
-     "\x03\x00"
-     "\xBF\x41\x03\x00\xBE\x71\x03\x00\xAB\x7C\x00\x80\xD3\x7A\x00\x80\xBB\x71"
-     "\x03\x00"
-     "\xBA\x79\x03\x00\x82\xDD\x27\x00\x83\x01\x3B\x00\xBE\x84\x07\x00\xBE\xC0"
-     "\x06\x00"
-     "\x86\x11\x3F\x00\x87\x19\x3F\x00\x84\x11\x3B\x00\x85\x5D\x3A\x00\x8A\x7D"
-     "\x3E\x00"
-     "\x8B\x25\x33\x00\xAF\x7C\x00\x80\xB3\x7C\x00\x80\x8E\xB9\x35\x00\x8F\x15"
-     "\x37\x00"
-     "\x8C\x35\x33\x00\x8D\x81\x33\x00\x92\xA1\x37\x00\x93\xD9\x09\x00\xBE\xC4"
-     "\x19\x00"
-     "\xB7\x7C\x00\x80\x96\xB1\x0D\x00\x97\x51\x0F\x00\x94\x79\x0B\x00\x95\x61"
-     "\x0B\x00"
-     "\x9A\x05\x0F\x00\x9B\x91\x01\x00\xBB\x7C\x00\x80\xBF\x7C\x00\x80\xC3\x7C"
-     "\x00\x80"
-     "\xDD\x00\x00\x00\x9C\x7D\x03\x00\xC7\x7C\x00\x80\xE1\x48\x0F\x00\xCB\x7C"
-     "\x00\x80"
-     "\xE3\x1C\x0E\x00\xCF\x7C\x00\x80\xD3\x7C\x00\x80\xD7\x7C\x00\x80\xDB\x7C"
-     "\x00\x80"
-     "\xDF\x7C\x00\x80\xB1\x41\x17\x00\xB0\xA1\x17\x00\xB3\xA9\xEB\x01\xB2\x01"
-     "\xE8\x01"
-     "\xB5\x01\xEC\x01\xB4\x11\xEB\x01\xEF\x8C\x0E\x00\xE3\x7C\x00\x80\xA9\xC5"
-     "\x1F\x00"
-     "\xA8\x01\x1C\x00\xAB\x01\x10\x00\xAA\x91\x1F\x00\xAD\xD9\x13\x00\xAC\xD1"
-     "\x13\x00"
-     "\xAF\xD9\x17\x00\xAE\x05\x13\x00\xA1\xF1\x02\x00\xE7\x7C\x00\x80\xA3\xC9"
-     "\x07\x00"
-     "\xA2\xC1\x02\x00\xA5\x01\x18\x00\xA4\x65\x07\x00\xA7\xA1\x1B\x00\xA6\xF9"
-     "\x1B\x00"
-     "\xA8\x29\x05\x00\xA9\x5D\x05\x00\xAA\x55\x05\x00\xAB\x6D\x05\x00\xAC\x79"
-     "\x05\x00"
-     "\xAD\x79\x05\x00\xAE\x6D\x05\x00\xAF\x61\x05\x00\x93\x7C\x00\x80\xEB\x7C"
-     "\x00\x80"
-     "\xEF\x7C\x00\x80\xF3\x7C\x00\x80\x80\x0D\x00\x00\x81\xB1\x00\x00\x82\xB1"
-     "\x00\x00"
-     "\xF7\x7C\x00\x80\xB8\x89\x05\x00\xB9\x89\x05\x00\xBA\x9D\x05\x00\xBB\x95"
-     "\x05\x00"
-     "\xBC\xB9\x05\x00\xBD\xB9\x05\x00\xBE\x51\x06\x00\xBF\x51\x06\x00\xB0\xE5"
-     "\x05\x00"
-     "\xB1\xED\x05\x00\xB2\xE5\x05\x00\xB3\xFD\x05\x00\xB4\xED\x05\x00\xB5\xDD"
-     "\x05\x00"
-     "\xB6\xD5\x05\x00\xB7\xBD\x05\x00\xA3\xDD\x05\x00\xFB\x7C\x00\x80\xFF\x7C"
-     "\x00\x80"
-     "\x84\x0C\x00\x00\x03\x7D\x00\x80\xA6\xF9\x05\x00\xA5\xF1\x05\x00\x07\x7D"
-     "\x00\x80"
-     "\xAB\x29\x05\x00\xAA\x21\x05\x00\x86\x98\x00\x00\x87\xA0\x00\x00\xAF\x19"
-     "\x05\x00"
-     "\xAE\x29\x05\x00\xAD\x29\x05\x00\xAC\x31\x05\x00\x0B\x7D\x00\x80\xB3\x61"
-     "\x06\x00"
-     "\x0F\x7D\x00\x80\x13\x7D\x00\x80\xB6\x21\x06\x00\x17\x7D\x00\x80\x1B\x7D"
-     "\x00\x80"
-     "\xB5\x01\x06\x00\xBA\xAD\x07\x00\xBB\x8D\x07\x00\x1F\x7D\x00\x80\x23\x7D"
-     "\x00\x80"
-     "\xBE\x8D\x07\x00\xBF\x71\x07\x00\xBC\x95\x07\x00\xBD\x8D\x07\x00\xBE\xB9"
-     "\x05\x00"
-     "\xBF\xB9\x05\x00\xBC\xB9\x05\x00\xBD\xB9\x05\x00\xBA\xB9\x05\x00\xBB\xB9"
-     "\x05\x00"
-     "\xB8\xB9\x05\x00\xB9\xB9\x05\x00\xB6\x49\x05\x00\xB7\x49\x05\x00\xB4\x7D"
-     "\x05\x00"
-     "\xB5\x75\x05\x00\xB2\x79\x05\x00\xB3\x79\x05\x00\xB0\x15\x05\x00\xB1\x7D"
-     "\x05\x00"
-     "\xAE\x5D\x05\x00\xAF\x6D\x05\x00\xAC\x45\x05\x00\xAD\x5D\x05\x00\xAA\xA5"
-     "\x0A\x00"
-     "\xAB\x5D\x05\x00\xA8\xBD\x0A\x00\xA9\xAD\x0A\x00\x27\x7D\x00\x80\x2B\x7D"
-     "\x00\x80"
-     "\x2F\x7D\x00\x80\x33\x7D\x00\x80\x37\x7D\x00\x80\x3B\x7D\x00\x80\x3F\x7D"
-     "\x00\x80"
-     "\x43\x7D\x00\x80\xA8\x0D\x07\x00\xA9\x1D\x07\x00\xAA\x2D\x07\x00\xAB\x49"
-     "\x07\x00"
-     "\xAC\x4D\x07\x00\xAD\x65\x07\x00\xAE\xB1\x06\x00\xAF\xB1\x06\x00\x47\x7D"
-     "\x00\x80"
-     "\x4B\x7D\x00\x80\x4F\x7D\x00\x80\x53\x7D\x00\x80\x57\x7D\x00\x80\x5B\x7D"
-     "\x00\x80"
-     "\x5F\x7D\x00\x80\x63\x7D\x00\x80\xB8\x55\x06\x00\xB9\x5D\x06\x00\xBA\x55"
-     "\x06\x00"
-     "\xBB\x71\x06\x00\xBC\x35\x06\x00\xBD\xF1\x01\x00\xBE\xF1\x01\x00\xBF\xF1"
-     "\x01\x00"
-     "\xB0\xAD\x06\x00\xB1\x8D\x06\x00\xB2\x85\x06\x00\xB3\x9D\x06\x00\xB4\x8D"
-     "\x06\x00"
-     "\xB5\x71\x06\x00\xB6\x75\x06\x00\xB7\x6D\x06\x00\xA3\xA5\x04\x00\x82\x2D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x80\x1D\x00\x00\x67\x7D\x00\x80\xA6\xE5\x04\x00\xA5\xC5"
-     "\x04\x00"
-     "\x6B\x7D\x00\x80\xAB\x49\x05\x00\xAA\x69\x05\x00\x6F\x7D\x00\x80\x77\x7D"
-     "\x00\x80"
-     "\xAF\xB5\x05\x00\xAE\x49\x05\x00\xAD\x49\x05\x00\xAC\x51\x05\x00\x86\x60"
-     "\x1C\x00"
-     "\x87\x08\x03\x00\x7B\x7D\x00\x80\xB3\x85\x02\x00\x7F\x7D\x00\x80\xB5\x81"
-     "\x02\x00"
-     "\xB6\x81\x02\x00\x83\x7D\x00\x80\x87\x7D\x00\x80\x8B\x7D\x00\x80\xBA\x09"
-     "\x03\x00"
-     "\xBB\x09\x03\x00\xBC\x19\x03\x00\xBD\x19\x03\x00\xBE\x09\x03\x00\xBF\x09"
-     "\x03\x00"
-     "\xAC\x55\x02\x00\xAD\x5D\x02\x00\xAE\x61\x02\x00\xAF\x61\x02\x00\xA8\x0D"
-     "\x02\x00"
-     "\xA9\x55\x02\x00\xAA\x51\x02\x00\xAB\x51\x02\x00\x84\xAC\x03\x00\x8F\x7D"
-     "\x00\x80"
-     "\x93\x7D\x00\x80\x97\x7D\x00\x80\x84\xFC\x1D\x00\x9B\x7D\x00\x80\x9F\x7D"
-     "\x00\x80"
-     "\xA3\x7D\x00\x80\xBC\x71\x03\x00\xBD\x71\x03\x00\xBE\x71\x03\x00\xBF\x71"
-     "\x03\x00"
-     "\xB8\x71\x03\x00\xB9\x71\x03\x00\xBA\x71\x03\x00\xBB\x71\x03\x00\xB4\x91"
-     "\x03\x00"
-     "\xB5\x91\x03\x00\xB6\x91\x03\x00\xB7\x91\x03\x00\xB0\x91\x03\x00\xB1\x91"
-     "\x03\x00"
-     "\xB2\x91\x03\x00\xB3\x91\x03\x00\xA7\x7D\x00\x80\xAB\x7D\x00\x80\xAF\x7D"
-     "\x00\x80"
-     "\xB3\x7D\x00\x80\xB7\x7D\x00\x80\xE1\xA4\x01\x00\xBB\x7D\x00\x80\xE3\x80"
-     "\x01\x00"
-     "\xBE\x68\x1C\x00\xBF\x7D\x00\x80\xC3\x7D\x00\x80\xEF\xD8\x06\x00\xC7\x7D"
-     "\x00\x80"
-     "\xCB\x7D\x00\x80\xCF\x7D\x00\x80\xD3\x7D\x00\x80\xA3\x89\x03\x00\x82\x2D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x80\x1D\x00\x00\xD7\x7D\x00\x80\xA6\x8D\x03\x00\xA5\x8D"
-     "\x03\x00"
-     "\xDB\x7D\x00\x80\xAB\x05\x02\x00\xAA\x05\x02\x00\xDF\x7D\x00\x80\xE7\x7D"
-     "\x00\x80"
-     "\xAF\x05\x02\x00\xAE\x05\x02\x00\xAD\x15\x02\x00\xAC\x15\x02\x00\x86\x20"
-     "\x1C\x00"
-     "\x87\xC4\x1D\x00\xEB\x7D\x00\x80\xEF\x7D\x00\x80\xF3\x7D\x00\x80\xF7\x7D"
-     "\x00\x80"
-     "\xFB\x7D\x00\x80\xEF\x6C\x06\x00\xFF\x7D\x00\x80\xE1\x6C\x07\x00\x03\x7E"
-     "\x00\x80"
-     "\xE3\x74\x07\x00\x07\x7E\x00\x80\x0B\x7E\x00\x80\x0F\x7E\x00\x80\x13\x7E"
-     "\x00\x80"
-     "\xB3\x91\x01\x00\x17\x7E\x00\x80\x1B\x7E\x00\x80\x1F\x7E\x00\x80\x23\x7E"
-     "\x00\x80"
-     "\xB6\xB1\x01\x00\xB5\xB9\x01\x00\x27\x7E\x00\x80\xBB\x55\x01\x00\xBA\x49"
-     "\x01\x00"
-     "\x2B\x7E\x00\x80\x2F\x7E\x00\x80\xBF\xF5\x00\x00\xBE\xF5\x00\x00\xBD\x45"
-     "\x01\x00"
-     "\xBC\x45\x01\x00\xA3\x51\x1E\x00\xE3\x7D\x00\x80\x33\x7E\x00\x80\x37\x7E"
-     "\x00\x80"
-     "\x3B\x7E\x00\x80\xA6\x71\x1E\x00\xA5\x79\x1E\x00\x3F\x7E\x00\x80\xAB\x95"
-     "\x1E\x00"
-     "\xAA\x89\x1E\x00\x43\x7E\x00\x80\x47\x7E\x00\x80\xAF\x35\x1F\x00\xAE\x35"
-     "\x1F\x00"
-     "\xAD\x85\x1E\x00\xAC\x85\x1E\x00\x80\x6D\x00\x00\x81\x15\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\xEF\xFC\x19\x00\x4B\x7E\x00\x80\x4F\x7E\x00\x80\x53\x7E\x00\x80\x57\x7E"
-     "\x00\x80"
-     "\x86\xC0\x00\x00\x87\xAC\x03\x00\x5B\x7E\x00\x80\x5F\x7E\x00\x80\x63\x7E"
-     "\x00\x80"
-     "\xE1\x2C\x1C\x00\x67\x7E\x00\x80\xE3\xCC\x1C\x00\xA8\xAD\x1E\x00\xA9\xCD"
-     "\x1E\x00"
-     "\xAA\xD9\x1E\x00\xAB\xD1\x1E\x00\xAC\xF1\x1E\x00\xAD\xF1\x1E\x00\xAE\x3D"
-     "\x1E\x00"
-     "\xAF\x35\x1E\x00\x84\xEC\x00\x00\x6B\x7E\x00\x80\x6F\x7E\x00\x80\x73\x7E"
-     "\x00\x80"
-     "\x77\x7E\x00\x80\x7B\x7E\x00\x80\x7F\x7E\x00\x80\x83\x7E\x00\x80\xB8\xD1"
-     "\x1F\x00"
-     "\xB9\xDD\x1F\x00\xBA\xED\x1F\x00\xBB\xE5\x1F\x00\xBC\xE1\x1F\x00\xBD\xE1"
-     "\x1F\x00"
-     "\xBE\xE1\x1F\x00\xBF\xE1\x1F\x00\xB0\x4D\x1E\x00\xB1\x51\x1E\x00\xB2\x51"
-     "\x1E\x00"
-     "\xB3\x51\x1E\x00\xB4\xF1\x1F\x00\xB5\xF1\x1F\x00\xB6\xF1\x1F\x00\xB7\xF1"
-     "\x1F\x00"
-     "\xA8\x6D\x1E\x00\xA9\x8D\x1E\x00\xAA\x85\x1E\x00\xAB\x9D\x1E\x00\xAC\x85"
-     "\x1E\x00"
-     "\xAD\x8D\x1E\x00\xAE\xB9\x1E\x00\xAF\xB5\x1E\x00\x87\x7E\x00\x80\x8B\x7E"
-     "\x00\x80"
-     "\x8F\x7E\x00\x80\x93\x7E\x00\x80\x97\x7E\x00\x80\x9B\x7E\x00\x80\x9F\x7E"
-     "\x00\x80"
-     "\xA3\x7E\x00\x80\xB8\x9D\x1E\x00\xB9\xAD\x1E\x00\xBA\xA5\x1E\x00\xBB\x45"
-     "\x01\x00"
-     "\xBC\x5D\x01\x00\xBD\x45\x01\x00\xBE\x45\x01\x00\xBF\x75\x01\x00\xB0\xD1"
-     "\x1E\x00"
-     "\xB1\xD1\x1E\x00\xB2\xD1\x1E\x00\xB3\xD1\x1E\x00\xB4\xB5\x1E\x00\xB5\xBD"
-     "\x1E\x00"
-     "\xB6\xB5\x1E\x00\xB7\xAD\x1E\x00\xB3\x0D\x1E\x00\xA7\x7E\x00\x80\xAB\x7E"
-     "\x00\x80"
-     "\xAF\x7E\x00\x80\xB3\x7E\x00\x80\xB6\x0D\x1E\x00\xB5\x0D\x1E\x00\xB7\x7E"
-     "\x00\x80"
-     "\xBB\x15\x1E\x00\xBA\x15\x1E\x00\xBB\x7E\x00\x80\xBF\x7E\x00\x80\xBF\x79"
-     "\x1E\x00"
-     "\xBE\x71\x1E\x00\xBD\x05\x1E\x00\xBC\x05\x1E\x00\x82\x6D\x00\x00\xA3\x49"
-     "\x1E\x00"
-     "\x80\x55\x00\x00\x81\x65\x00\x00\xA6\x49\x1E\x00\xBE\x9C\x01\x00\xC7\x7E"
-     "\x00\x80"
-     "\xA5\x49\x1E\x00\xAA\x51\x1E\x00\xAB\x51\x1E\x00\x87\x7C\x00\x00\x86\x4C"
-     "\x00\x00"
-     "\xAE\x35\x1E\x00\xAF\x3D\x1E\x00\xAC\x41\x1E\x00\xAD\x41\x1E\x00\xA8\x5D"
-     "\x02\x00"
-     "\xA9\x6D\x02\x00\xAA\x65\x02\x00\xAB\x7D\x02\x00\xAC\x69\x02\x00\xAD\xB1"
-     "\x02\x00"
-     "\xAE\xB1\x02\x00\xAF\xB1\x02\x00\x84\xEC\x04\x00\xCB\x7E\x00\x80\xCF\x7E"
-     "\x00\x80"
-     "\xD3\x7E\x00\x80\xD7\x7E\x00\x80\xDB\x7E\x00\x80\xDF\x7E\x00\x80\xE3\x7E"
-     "\x00\x80"
-     "\xB8\x71\x03\x00\xB9\x71\x03\x00\xBA\x71\x03\x00\xBB\x71\x03\x00\xBC\xD5"
-     "\x03\x00"
-     "\xBD\xDD\x03\x00\xBE\xD5\x03\x00\xBF\xCD\x03\x00\xB0\xD1\x02\x00\xB1\xD1"
-     "\x02\x00"
-     "\xB2\xD1\x02\x00\xB3\xD1\x02\x00\xB4\x51\x03\x00\xB5\x51\x03\x00\xB6\x51"
-     "\x03\x00"
-     "\xB7\x51\x03\x00\xB3\xED\x02\x00\xE7\x7E\x00\x80\xEB\x7E\x00\x80\xBE\x80"
-     "\x04\x00"
-     "\xEF\x7E\x00\x80\xB6\x31\x02\x00\xB5\xE1\x02\x00\xF3\x7E\x00\x80\xBB\x15"
-     "\x02\x00"
-     "\xBA\x15\x02\x00\xF7\x7E\x00\x80\xFB\x7E\x00\x80\xBF\x95\x03\x00\xBE\x95"
-     "\x03\x00"
-     "\xBD\x05\x02\x00\xBC\x05\x02\x00\xFF\x7E\x00\x80\xA3\xA9\x02\x00\x03\x7F"
-     "\x00\x80"
-     "\x07\x7F\x00\x80\xA6\x75\x02\x00\x0B\x7F\x00\x80\x0F\x7F\x00\x80\xA5\xA5"
-     "\x02\x00"
-     "\xAA\x51\x02\x00\xAB\x51\x02\x00\x13\x7F\x00\x80\x17\x7F\x00\x80\xAE\xD1"
-     "\x03\x00"
-     "\xAF\xD1\x03\x00\xAC\x41\x02\x00\xAD\x41\x02\x00\xA8\xD9\x02\x00\xA9\x21"
-     "\x01\x00"
-     "\xAA\x21\x01\x00\xAB\x21\x01\x00\xAC\x21\x01\x00\xAD\x21\x01\x00\xAE\x21"
-     "\x01\x00"
-     "\xAF\x21\x01\x00\x1B\x7F\x00\x80\x1F\x7F\x00\x80\x23\x7F\x00\x80\xBE\x20"
-     "\x04\x00"
-     "\x27\x7F\x00\x80\x2B\x7F\x00\x80\x2F\x7F\x00\x80\x37\x7F\x00\x80\xB8\x8D"
-     "\x01\x00"
-     "\xB9\x91\x01\x00\xBA\x91\x01\x00\xBB\xA5\x01\x00\xBC\xBD\x01\x00\xBD\x75"
-     "\x00\x00"
-     "\xBE\x7D\x00\x00\xBF\x75\x00\x00\xB0\xE5\x01\x00\xB1\xED\x01\x00\xB2\xF9"
-     "\x01\x00"
-     "\xB3\xF1\x01\x00\xB4\xD9\x01\x00\xB5\xD9\x01\x00\xB6\xB9\x01\x00\xB7\xB5"
-     "\x01\x00"
-     "\xE1\x18\x1E\x00\x3B\x7F\x00\x80\xE3\x28\x1F\x00\x3F\x7F\x00\x80\x81\xA5"
-     "\x00\x00"
-     "\x80\xA5\x00\x00\x43\x7F\x00\x80\x82\xA5\x00\x00\x86\x00\x04\x00\x87\xF4"
-     "\x05\x00"
-     "\x47\x7F\x00\x80\x4B\x7F\x00\x80\x4F\x7F\x00\x80\x53\x7F\x00\x80\xEF\x60"
-     "\x1E\x00"
-     "\x57\x7F\x00\x80\x5B\x7F\x00\x80\x5F\x7F\x00\x80\x85\xF0\xF4\x01\x63\x7F"
-     "\x00\x80"
-     "\x67\x7F\x00\x80\xE3\x64\x01\x00\x6B\x7F\x00\x80\xE1\xA4\x01\x00\x6F\x7F"
-     "\x00\x80"
-     "\xEF\xC8\x00\x00\x73\x7F\x00\x80\x77\x7F\x00\x80\x73\x7D\x00\x80\x84\xFC"
-     "\x05\x00"
-     "\x7B\x7F\x00\x80\x7F\x7F\x00\x80\xB3\x29\x06\x00\x33\x7F\x00\x80\x83\x7F"
-     "\x00\x80"
-     "\x87\x7F\x00\x80\x8B\x7F\x00\x80\xB6\x21\x06\x00\xB5\x29\x06\x00\x8F\x7F"
-     "\x00\x80"
-     "\xBB\xA9\x06\x00\xBA\xA1\x06\x00\x93\x7F\x00\x80\x97\x7F\x00\x80\xBF\x9D"
-     "\x06\x00"
-     "\xBE\x9D\x06\x00\xBD\xA5\x06\x00\xBC\xAD\x06\x00\xE1\xB4\x07\x00\x9B\x7F"
-     "\x00\x80"
-     "\xE3\x78\x04\x00\x9F\x7F\x00\x80\x80\x7D\x00\x00\x81\x11\x00\x00\x82\x15"
-     "\x00\x00"
-     "\xA3\x7F\x00\x80\x86\xC0\x00\x00\x87\x58\x03\x00\xA7\x7F\x00\x80\xAB\x7F"
-     "\x00\x80"
-     "\xAF\x7F\x00\x80\xB3\x7F\x00\x80\xEF\x0C\x04\x00\xB7\x7F\x00\x80\xA3\xA9"
-     "\x06\x00"
-     "\xBB\x7F\x00\x80\xBF\x7F\x00\x80\xC3\x7F\x00\x80\xC7\x7F\x00\x80\xA6\xA1"
-     "\x06\x00"
-     "\xA5\xA9\x06\x00\xCB\x7F\x00\x80\xAB\x29\x06\x00\xAA\x21\x06\x00\xCF\x7F"
-     "\x00\x80"
-     "\xD3\x7F\x00\x80\xAF\x1D\x06\x00\xAE\x1D\x06\x00\xAD\x25\x06\x00\xAC\x2D"
-     "\x06\x00"
-     "\xD7\x7F\x00\x80\xB3\x49\x07\x00\xDB\x7F\x00\x80\xDF\x7F\x00\x80\xB6\x7D"
-     "\x07\x00"
-     "\xE3\x7F\x00\x80\xE7\x7F\x00\x80\xB5\x75\x07\x00\xBA\x5D\x07\x00\xBB\x25"
-     "\x07\x00"
-     "\xEB\x7F\x00\x80\xEF\x7F\x00\x80\xBE\x21\x07\x00\xBF\x2D\x07\x00\xBC\x3D"
-     "\x07\x00"
-     "\xBD\x31\x07\x00\xA8\x3D\x06\x00\xA9\x81\x06\x00\xAA\x85\x06\x00\xAB\x95"
-     "\x06\x00"
-     "\xAC\xB9\x06\x00\xAD\xB9\x06\x00\xAE\xA9\x06\x00\xAF\xA9\x06\x00\xF3\x7F"
-     "\x00\x80"
-     "\xF7\x7F\x00\x80\xFB\x7F\x00\x80\xFF\x7F\x00\x80\x03\x80\x00\x80\x82\xB9"
-     "\x00\x00"
-     "\x81\xB1\x00\x00\x80\xB9\x00\x00\xB8\xAD\x06\x00\xB9\xBD\x06\x00\xBA\xB5"
-     "\x06\x00"
-     "\xBB\x45\x01\x00\xBC\x5D\x01\x00\xBD\x45\x01\x00\xBE\x45\x01\x00\xBF\x75"
-     "\x01\x00"
-     "\xB0\xDD\x06\x00\xB1\xA5\x06\x00\xB2\xAD\x06\x00\xB3\xA1\x06\x00\xB4\xA1"
-     "\x06\x00"
-     "\xB5\xAD\x06\x00\xB6\x99\x06\x00\xB7\x95\x06\x00\xA3\x0D\x06\x00\x07\x80"
-     "\x00\x80"
-     "\x0B\x80\x00\x80\x0F\x80\x00\x80\x84\x98\x02\x00\xA6\x39\x06\x00\xA5\x31"
-     "\x06\x00"
-     "\xBE\x9C\x01\x00\xAB\x61\x06\x00\xAA\x19\x06\x00\x86\x08\x00\x00\x87\x7C"
-     "\x01\x00"
-     "\xAF\x69\x06\x00\xAE\x65\x06\x00\xAD\x75\x06\x00\xAC\x79\x06\x00\x17\x80"
-     "\x00\x80"
-     "\xB3\xB5\x01\x00\x1B\x80\x00\x80\x1F\x80\x00\x80\xB6\x55\x01\x00\x23\x80"
-     "\x00\x80"
-     "\x27\x80\x00\x80\xB5\xA1\x01\x00\xBA\x71\x01\x00\xBB\x79\x01\x00\x2B\x80"
-     "\x00\x80"
-     "\x2F\x80\x00\x80\xBE\x31\x01\x00\xBF\x3D\x01\x00\xBC\x51\x01\x00\xBD\x51"
-     "\x01\x00"
-     "\xA8\x69\x02\x00\xA9\x69\x02\x00\xAA\x79\x02\x00\xAB\x79\x02\x00\xAC\x6D"
-     "\x02\x00"
-     "\xAD\x91\x02\x00\xAE\x91\x02\x00\xAF\x91\x02\x00\x33\x80\x00\x80\x37\x80"
-     "\x00\x80"
-     "\x3B\x80\x00\x80\x3F\x80\x00\x80\x43\x80\x00\x80\x47\x80\x00\x80\x4B\x80"
-     "\x00\x80"
-     "\x4F\x80\x00\x80\xB8\x99\x02\x00\xB9\xAD\x02\x00\xBA\xA5\x02\x00\xBB\x6D"
-     "\x03\x00"
-     "\xBC\x75\x03\x00\xBD\x7D\x03\x00\xBE\x75\x03\x00\xBF\x6D\x03\x00\xB0\xF1"
-     "\x02\x00"
-     "\xB1\xF9\x02\x00\xB2\xC1\x02\x00\xB3\xC1\x02\x00\xB4\xB1\x02\x00\xB5\xBD"
-     "\x02\x00"
-     "\xB6\xB5\x02\x00\xB7\xA9\x02\x00\x53\x80\x00\x80\x57\x80\x00\x80\x5B\x80"
-     "\x00\x80"
-     "\xA3\xFD\x02\x00\x5F\x80\x00\x80\xA5\xE9\x02\x00\xA6\x1D\x02\x00\x63\x80"
-     "\x00\x80"
-     "\x67\x80\x00\x80\x6B\x80\x00\x80\xAA\x39\x02\x00\xAB\x31\x02\x00\xAC\x19"
-     "\x02\x00"
-     "\xAD\x19\x02\x00\xAE\x79\x02\x00\xAF\x75\x02\x00\x6F\x80\x00\x80\x73\x80"
-     "\x00\x80"
-     "\x77\x80\x00\x80\x84\x00\x0C\x00\x80\x19\x00\x00\x81\x09\x00\x00\x82\x39"
-     "\x00\x00"
-     "\x7B\x80\x00\x80\x7F\x80\x00\x80\x87\x80\x00\x80\x8B\x80\x00\x80\xBE\xE0"
-     "\x0C\x00"
-     "\x8F\x80\x00\x80\x93\x80\x00\x80\x86\x58\x0D\x00\x87\x0C\x03\x00\xA8\xC1"
-     "\x02\x00"
-     "\xA9\xCD\x02\x00\xAA\xC5\x02\x00\xAB\xD9\x02\x00\xAC\xC9\x02\x00\xAD\xFD"
-     "\x02\x00"
-     "\xAE\xF5\x02\x00\xAF\x3D\x01\x00\x97\x80\x00\x80\x9B\x80\x00\x80\x9F\x80"
-     "\x00\x80"
-     "\xA3\x80\x00\x80\xA7\x80\x00\x80\xAB\x80\x00\x80\xAF\x80\x00\x80\xBE\x30"
-     "\x0C\x00"
-     "\xB8\xC5\x01\x00\xB9\xCD\x01\x00\xBA\xD9\x01\x00\xBB\xD1\x01\x00\xBC\xF1"
-     "\x01\x00"
-     "\xBD\xF1\x01\x00\xBE\x99\x01\x00\xBF\x99\x01\x00\xB0\x45\x01\x00\xB1\x4D"
-     "\x01\x00"
-     "\xB2\x45\x01\x00\xB3\x5D\x01\x00\xB4\x45\x01\x00\xB5\x4D\x01\x00\xB6\x45"
-     "\x01\x00"
-     "\xB7\xFD\x01\x00\xE1\x38\x06\x00\xB3\x80\x00\x80\xE3\x6C\x06\x00\xB7\x80"
-     "\x00\x80"
-     "\xBB\x80\x00\x80\xBF\x80\x00\x80\xC3\x80\x00\x80\xC7\x80\x00\x80\x84\xA8"
-     "\x0D\x00"
-     "\xCB\x80\x00\x80\xCF\x80\x00\x80\xD3\x80\x00\x80\xBE\xB0\x0F\x00\xD7\x80"
-     "\x00\x80"
-     "\xEF\x6C\x06\x00\xDB\x80\x00\x80\xDF\x80\x00\x80\xC3\x7E\x00\x80\xE3\x80"
-     "\x00\x80"
-     "\xE7\x80\x00\x80\xE3\x50\x00\x00\xEB\x80\x00\x80\xE1\xA0\x01\x00\xEF\x80"
-     "\x00\x80"
-     "\xEF\x84\x00\x00\xF7\x80\x00\x80\x87\x20\x0D\x00\x86\x4C\x0F\x00\x80\x3D"
-     "\x00\x00"
-     "\x81\x21\x00\x00\x82\x25\x00\x00\xFB\x80\x00\x80\xB3\xCD\x0D\x00\x83\x80"
-     "\x00\x80"
-     "\xF3\x80\x00\x80\xFF\x80\x00\x80\x03\x81\x00\x80\xB6\xFD\x0D\x00\xB5\xC1"
-     "\x0D\x00"
-     "\x07\x81\x00\x80\xBB\x09\x0E\x00\xBA\x01\x0E\x00\x0B\x81\x00\x80\x0F\x81"
-     "\x00\x80"
-     "\xBF\x09\x0E\x00\xBE\x01\x0E\x00\xBD\x09\x0E\x00\xBC\x11\x0E\x00\x13\x81"
-     "\x00\x80"
-     "\xE3\x98\x0C\x00\x17\x81\x00\x80\xE1\xFC\x0F\x00\x1B\x81\x00\x80\x1F\x81"
-     "\x00\x80"
-     "\x23\x81\x00\x80\x27\x81\x00\x80\x2B\x81\x00\x80\x2F\x81\x00\x80\x33\x81"
-     "\x00\x80"
-     "\x37\x81\x00\x80\x3B\x81\x00\x80\xEF\x60\x0C\x00\x3F\x81\x00\x80\x43\x81"
-     "\x00\x80"
-     "\xA3\x4D\x0D\x00\x47\x81\x00\x80\x4B\x81\x00\x80\x4F\x81\x00\x80\x53\x81"
-     "\x00\x80"
-     "\xA6\x7D\x0D\x00\xA5\x41\x0D\x00\x57\x81\x00\x80\xAB\x89\x0E\x00\xAA\x81"
-     "\x0E\x00"
-     "\x5B\x81\x00\x80\x5F\x81\x00\x80\xAF\x89\x0E\x00\xAE\x81\x0E\x00\xAD\x89"
-     "\x0E\x00"
-     "\xAC\x91\x0E\x00\x82\x6D\x00\x00\xB3\x35\x0E\x00\x80\x55\x00\x00\x81\x65"
-     "\x00\x00"
-     "\xB6\xF5\x0F\x00\x84\xDC\x03\x00\x63\x81\x00\x80\xB5\xFD\x0F\x00\xBA\xD1"
-     "\x0F\x00"
-     "\xBB\xD1\x0F\x00\x86\x00\x04\x00\x87\xDC\x00\x00\xBE\x7D\x0F\x00\xBF\x65"
-     "\x0F\x00"
-     "\xBC\xC1\x0F\x00\xBD\x79\x0F\x00\xA8\xE5\x0F\x00\xA9\xED\x0F\x00\xAA\xF9"
-     "\x0F\x00"
-     "\xAB\xF9\x0F\x00\xAC\x31\x0E\x00\xAD\x31\x0E\x00\xAE\x31\x0E\x00\xAF\x31"
-     "\x0E\x00"
-     "\x67\x81\x00\x80\x6B\x81\x00\x80\x6F\x81\x00\x80\x73\x81\x00\x80\x77\x81"
-     "\x00\x80"
-     "\x7B\x81\x00\x80\x7F\x81\x00\x80\x83\x81\x00\x80\xB8\xDD\x0E\x00\xB9\xE1"
-     "\x0E\x00"
-     "\xBA\xE1\x0E\x00\xBB\xE1\x0E\x00\xBC\xE5\x0E\x00\xBD\xE9\x0E\x00\xBE\x99"
-     "\x0E\x00"
-     "\xBF\x95\x0E\x00\xB0\x51\x0E\x00\xB1\x51\x0E\x00\xB2\x51\x0E\x00\xB3\xE9"
-     "\x0E\x00"
-     "\xB4\xFD\x0E\x00\xB5\xE5\x0E\x00\xB6\xED\x0E\x00\xB7\xE5\x0E\x00\xA3\x71"
-     "\x0F\x00"
-     "\x87\x81\x00\x80\x8B\x81\x00\x80\x8F\x81\x00\x80\x93\x81\x00\x80\xA6\xB1"
-     "\x0E\x00"
-     "\xA5\xB9\x0E\x00\x97\x81\x00\x80\xAB\x95\x0E\x00\xAA\x95\x0E\x00\x9B\x81"
-     "\x00\x80"
-     "\x9F\x81\x00\x80\xAF\x21\x0E\x00\xAE\x39\x0E\x00\xAD\x3D\x0E\x00\xAC\x85"
-     "\x0E\x00"
-     "\xA3\x81\x00\x80\xB3\xC9\x01\x00\xA7\x81\x00\x80\xAB\x81\x00\x80\xB6\xF9"
-     "\x01\x00"
-     "\xAF\x81\x00\x80\xB3\x81\x00\x80\xB5\xC1\x01\x00\xBA\xA9\x01\x00\xBB\xB5"
-     "\x01\x00"
-     "\xB7\x81\x00\x80\xBB\x81\x00\x80\xBE\xAD\x01\x00\xBF\x95\x01\x00\xBC\xAD"
-     "\x01\x00"
-     "\xBD\xA5\x01\x00\xA8\xE5\x0D\x00\xA9\xF9\x0D\x00\xAA\x05\x02\x00\xAB\x1D"
-     "\x02\x00"
-     "\xAC\x0D\x02\x00\xAD\x3D\x02\x00\xAE\x35\x02\x00\xAF\x5D\x02\x00\xBF\x81"
-     "\x00\x80"
-     "\xC3\x81\x00\x80\xC7\x81\x00\x80\xCB\x81\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x05\x00\x00\xCF\x81\x00\x80\xB8\x2D\x02\x00\xB9\x35\x02\x00\xBA\x31"
-     "\x02\x00"
-     "\xBB\x31\x02\x00\xBC\xD5\x02\x00\xBD\xDD\x02\x00\xBE\xD5\x02\x00\xBF\xCD"
-     "\x02\x00"
-     "\xB0\x29\x02\x00\xB1\x35\x02\x00\xB2\x3D\x02\x00\xB3\x35\x02\x00\xB4\x2D"
-     "\x02\x00"
-     "\xB5\x15\x02\x00\xB6\x1D\x02\x00\xB7\x15\x02\x00\xD7\x81\x00\x80\x84\x9C"
-     "\x02\x00"
-     "\xDB\x81\x00\x80\xA3\x81\x02\x00\xDF\x81\x00\x80\xA5\x89\x02\x00\xA6\xB1"
-     "\x02\x00"
-     "\xE3\x81\x00\x80\x86\x20\x04\x00\x87\xD4\x03\x00\xAA\xE1\x02\x00\xAB\xFD"
-     "\x02\x00"
-     "\xAC\xE5\x02\x00\xAD\xED\x02\x00\xAE\xE5\x02\x00\xAF\xDD\x02\x00\xB6\xF5"
-     "\x03\x00"
-     "\xBE\x44\x03\x00\x85\x8C\xFD\x01\xB5\xFD\x03\x00\xE7\x81\x00\x80\xB3\xFD"
-     "\x03\x00"
-     "\xEB\x81\x00\x80\xEF\x81\x00\x80\xBE\x7D\x03\x00\xBF\x4D\x03\x00\xBC\x65"
-     "\x03\x00"
-     "\xBD\x7D\x03\x00\xBA\x75\x03\x00\xBB\x7D\x03\x00\xF3\x81\x00\x80\xF7\x81"
-     "\x00\x80"
-     "\xFB\x81\x00\x80\xFF\x81\x00\x80\x84\x04\x02\x00\xA3\x25\x02\x00\x03\x82"
-     "\x00\x80"
-     "\xA5\x25\x02\x00\xA6\x2D\x02\x00\x07\x82\x00\x80\x0B\x82\x00\x80\x0F\x82"
-     "\x00\x80"
-     "\xAA\xAD\x02\x00\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xA5\x02\x00\xAE\xA5"
-     "\x02\x00"
-     "\xAF\x95\x02\x00\x13\x82\x00\x80\x17\x82\x00\x80\x1B\x82\x00\x80\x1F\x82"
-     "\x00\x80"
-     "\x23\x82\x00\x80\xE3\xC0\x03\x00\x27\x82\x00\x80\xE1\xAC\x01\x00\x2B\x82"
-     "\x00\x80"
-     "\xEF\x58\x03\x00\x2F\x82\x00\x80\x33\x82\x00\x80\x80\x0D\x00\x00\x81\xE5"
-     "\x00\x00"
-     "\x82\xED\x00\x00\x37\x82\x00\x80\xE1\x60\x0F\x00\xE3\x40\x01\x00\xE3\x38"
-     "\x0E\x00"
-     "\xE1\x78\x0E\x00\x3B\x82\x00\x80\x3F\x82\x00\x80\xBE\x94\x05\x00\x47\x82"
-     "\x00\x80"
-     "\x86\x00\x04\x00\x87\x64\x05\x00\x4B\x82\x00\x80\x4F\x82\x00\x80\x53\x82"
-     "\x00\x80"
-     "\xEF\xFC\x0E\x00\xEF\x7C\x0E\x00\x57\x82\x00\x80\xB3\x51\x01\x00\x5B\x82"
-     "\x00\x80"
-     "\x13\x80\x00\x80\x5F\x82\x00\x80\x63\x82\x00\x80\xB6\x0D\x01\x00\xB5\x09"
-     "\x01\x00"
-     "\x67\x82\x00\x80\xBB\xE1\x00\x00\xBA\x19\x01\x00\x6B\x82\x00\x80\x6F\x82"
-     "\x00\x80"
-     "\xBF\xD1\x00\x00\xBE\xE9\x00\x00\xBD\xE9\x00\x00\xBC\xF9\x00\x00\xD3\x81"
-     "\x00\x80"
-     "\x43\x82\x00\x80\x73\x82\x00\x80\x77\x82\x00\x80\x7B\x82\x00\x80\x7F\x82"
-     "\x00\x80"
-     "\x83\x82\x00\x80\x87\x82\x00\x80\xA8\xA1\x06\x00\xA9\xAD\x06\x00\xAA\xB9"
-     "\x06\x00"
-     "\xAB\xB1\x06\x00\xAC\xE1\x06\x00\xAD\xED\x06\x00\xAE\xE5\x06\x00\xAF\xC5"
-     "\x06\x00"
-     "\xB0\xBD\x06\x00\xB1\x45\x07\x00\xB2\x4D\x07\x00\xB3\x5D\x07\x00\xB4\x4D"
-     "\x07\x00"
-     "\xB5\x75\x07\x00\xB6\x7D\x07\x00\xB7\x1D\x07\x00\xB8\x25\x07\x00\xB9\x2D"
-     "\x07\x00"
-     "\xBA\x25\x07\x00\xBB\x3D\x07\x00\xBC\x29\x07\x00\xBD\x15\x07\x00\xBE\x11"
-     "\x07\x00"
-     "\xBF\x11\x07\x00\xA3\x11\x06\x00\x8B\x82\x00\x80\x8F\x82\x00\x80\x93\x82"
-     "\x00\x80"
-     "\x97\x82\x00\x80\xA6\x4D\x06\x00\xA5\x49\x06\x00\x9B\x82\x00\x80\xAB\xA1"
-     "\x07\x00"
-     "\xAA\x59\x06\x00\x9F\x82\x00\x80\xA3\x82\x00\x80\xAF\x91\x07\x00\xAE\xA9"
-     "\x07\x00"
-     "\xAD\xA9\x07\x00\xAC\xB9\x07\x00\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\xA7\x82\x00\x80\xAB\x82\x00\x80\xAF\x82\x00\x80\x84\x94\x03\x00\xBE\x94"
-     "\x03\x00"
-     "\xB3\x82\x00\x80\xB7\x82\x00\x80\x86\xC8\x00\x00\x87\x8C\x00\x00\xBB\x82"
-     "\x00\x80"
-     "\xBF\x82\x00\x80\xC3\x82\x00\x80\xC7\x82\x00\x80\xA8\x9D\x06\x00\xA9\xAD"
-     "\x06\x00"
-     "\xAA\xA5\x06\x00\xAB\xBD\x06\x00\xAC\xAD\x06\x00\xAD\xD1\x06\x00\xAE\xD5"
-     "\x06\x00"
-     "\xAF\xCD\x06\x00\xCB\x82\x00\x80\xCF\x82\x00\x80\xD3\x82\x00\x80\xD7\x82"
-     "\x00\x80"
-     "\xDB\x82\x00\x80\xDF\x82\x00\x80\xE3\x82\x00\x80\xE7\x82\x00\x80\xB8\x5D"
-     "\x01\x00"
-     "\xB9\xC1\x01\x00\xBA\xC1\x01\x00\xBB\xC1\x01\x00\xBC\xC1\x01\x00\xBD\xC9"
-     "\x01\x00"
-     "\xBE\xF1\x01\x00\xBF\xF1\x01\x00\xB0\xBD\x06\x00\xB1\x8D\x06\x00\xB2\x85"
-     "\x06\x00"
-     "\xB3\x65\x01\x00\xB4\x7D\x01\x00\xB5\x65\x01\x00\xB6\x6D\x01\x00\xB7\x65"
-     "\x01\x00"
-     "\xB3\x2D\x06\x00\xEB\x82\x00\x80\xEF\x82\x00\x80\xF3\x82\x00\x80\xF7\x82"
-     "\x00\x80"
-     "\xB6\x51\x06\x00\xB5\x25\x06\x00\xFB\x82\x00\x80\xBB\x49\x06\x00\xBA\x79"
-     "\x06\x00"
-     "\xFF\x82\x00\x80\x03\x83\x00\x80\xBF\xA1\x01\x00\xBE\xB9\x01\x00\xBD\xB1"
-     "\x01\x00"
-     "\xBC\x51\x06\x00\x07\x83\x00\x80\xA3\x69\x06\x00\x0B\x83\x00\x80\x0F\x83"
-     "\x00\x80"
-     "\xA6\x15\x06\x00\x13\x83\x00\x80\x17\x83\x00\x80\xA5\x61\x06\x00\xAA\x3D"
-     "\x06\x00"
-     "\xAB\x0D\x06\x00\x1B\x83\x00\x80\x1F\x83\x00\x80\xAE\xFD\x01\x00\xAF\xE5"
-     "\x01\x00"
-     "\xAC\x15\x06\x00\xAD\xF5\x01\x00\xBA\xD5\x07\x00\xBB\xDD\x07\x00\xB8\xC1"
-     "\x07\x00"
-     "\xB9\xC1\x07\x00\xBE\x31\x04\x00\xBF\x31\x04\x00\xBC\xF1\x07\x00\xBD\xF1"
-     "\x07\x00"
-     "\xB2\xAD\x07\x00\xB3\xB5\x07\x00\xB0\xAD\x07\x00\xB1\xA5\x07\x00\xB6\x9D"
-     "\x07\x00"
-     "\xB7\xF5\x07\x00\xB4\xA5\x07\x00\xB5\x95\x07\x00\xAA\x69\x07\x00\xAB\x69"
-     "\x07\x00"
-     "\xA8\x69\x07\x00\xA9\x69\x07\x00\xAE\x69\x07\x00\xAF\x69\x07\x00\xAC\x69"
-     "\x07\x00"
-     "\xAD\x69\x07\x00\x80\xB9\x03\x00\x81\x8D\x03\x00\x82\x85\x03\x00\x84\xA8"
-     "\x03\x00"
-     "\x86\x50\xFC\x01\x87\x08\x03\x00\xBE\x34\x03\x00\x27\x83\x00\x80\xA8\x65"
-     "\x02\x00"
-     "\xA9\x75\x02\x00\xAA\x7D\x02\x00\xAB\x75\x02\x00\xAC\x6D\x02\x00\xAD\xB5"
-     "\x03\x00"
-     "\xAE\xBD\x03\x00\xAF\xB5\x03\x00\x2B\x83\x00\x80\x2F\x83\x00\x80\x33\x83"
-     "\x00\x80"
-     "\x37\x83\x00\x80\x3B\x83\x00\x80\x3F\x83\x00\x80\x43\x83\x00\x80\x47\x83"
-     "\x00\x80"
-     "\xB8\x51\x03\x00\xB9\x59\x03\x00\xBA\x61\x03\x00\xBB\x61\x03\x00\xBC\x11"
-     "\x03\x00"
-     "\xBD\x1D\x03\x00\xBE\x15\x03\x00\xBF\x09\x03\x00\xB0\xCD\x03\x00\xB1\xD5"
-     "\x03\x00"
-     "\xB2\xDD\x03\x00\xB3\xD5\x03\x00\xB4\xCD\x03\x00\xB5\x71\x03\x00\xB6\x71"
-     "\x03\x00"
-     "\xB7\x71\x03\x00\x4B\x83\x00\x80\x4F\x83\x00\x80\xB3\xFD\x03\x00\x53\x83"
-     "\x00\x80"
-     "\xB5\xED\x03\x00\x57\x83\x00\x80\x5B\x83\x00\x80\xB6\x3D\x02\x00\x5F\x83"
-     "\x00\x80"
-     "\x63\x83\x00\x80\xBB\x19\x02\x00\xBA\x01\x02\x00\xBD\x09\x02\x00\xBC\x01"
-     "\x02\x00"
-     "\xBF\xBD\x02\x00\xBE\x01\x02\x00\x67\x83\x00\x80\x6B\x83\x00\x80\x84\xC4"
-     "\xFD\x01"
-     "\xBE\xC0\xFC\x01\x6F\x83\x00\x80\x73\x83\x00\x80\x77\x83\x00\x80\xEF\xDC"
-     "\x03\x00"
-     "\x7B\x83\x00\x80\xE1\x94\x01\x00\x7F\x83\x00\x80\xE3\x10\x03\x00\x83\x83"
-     "\x00\x80"
-     "\x82\xED\x00\x00\x81\xED\x00\x00\x80\xED\x00\x00\xE1\x38\x06\x00\xE1\x3C"
-     "\x07\x00"
-     "\xE3\x40\x01\x00\xE3\x94\x06\x00\x87\x83\x00\x80\x8B\x83\x00\x80\x8F\x83"
-     "\x00\x80"
-     "\x97\x83\x00\x80\x86\x80\xFC\x01\x87\xE8\xFD\x01\x9B\x83\x00\x80\x9F\x83"
-     "\x00\x80"
-     "\xA3\x83\x00\x80\xA7\x83\x00\x80\xEF\x9C\x01\x00\xEF\xDC\x06\x00\xA3\x35"
-     "\x03\x00"
-     "\xAB\x83\x00\x80\xAF\x83\x00\x80\xB3\x83\x00\x80\xB7\x83\x00\x80\xA6\xF5"
-     "\x02\x00"
-     "\xA5\x25\x03\x00\xBB\x83\x00\x80\xAB\xD1\x02\x00\xAA\xC9\x02\x00\xBF\x83"
-     "\x00\x80"
-     "\xC3\x83\x00\x80\xAF\x75\x02\x00\xAE\xC9\x02\x00\xAD\xC1\x02\x00\xAC\xC9"
-     "\x02\x00"
-     "\x93\x83\x00\x80\xC7\x83\x00\x80\xCB\x83\x00\x80\xCF\x83\x00\x80\xEF\xA4"
-     "\xFD\x01"
-     "\xD3\x83\x00\x80\xD7\x83\x00\x80\xDB\x83\x00\x80\xE3\x78\xFD\x01\xDF\x83"
-     "\x00\x80"
-     "\xE1\x40\xFC\x01\xE3\x83\x00\x80\xE7\x83\x00\x80\xEB\x83\x00\x80\xEF\x83"
-     "\x00\x80"
-     "\xF3\x83\x00\x80\xB3\xE5\xFE\x01\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xF7\x83\x00\x80\xB6\xE5\xFE\x01\xB5\xF5\xFE\x01\xFB\x83\x00\x80\xBB\x61"
-     "\xFF\x01"
-     "\xBA\xD9\xFE\x01\xFF\x83\x00\x80\x84\xE4\x03\x00\xBF\x69\xFF\x01\xBE\x61"
-     "\xFF\x01"
-     "\xBD\x75\xFF\x01\xBC\x79\xFF\x01\xA8\xFD\xFE\x01\xA9\x25\xFE\x01\xAA\x2D"
-     "\xFE\x01"
-     "\xAB\x25\xFE\x01\xAC\x3D\xFE\x01\xAD\x25\xFE\x01\xAE\x2D\xFE\x01\xAF\x25"
-     "\xFE\x01"
-     "\xBE\x2C\x00\x00\x03\x84\x00\x80\x86\x88\x00\x00\x87\xEC\x00\x00\x07\x84"
-     "\x00\x80"
-     "\x0B\x84\x00\x80\x0F\x84\x00\x80\x13\x84\x00\x80\xB8\x81\xFF\x01\xB9\x81"
-     "\xFF\x01"
-     "\xBA\x99\xFF\x01\xBB\x99\xFF\x01\xBC\x89\xFF\x01\xBD\xB5\xFF\x01\xBE\xB1"
-     "\xFF\x01"
-     "\xBF\xAD\xFF\x01\xB0\x5D\xFE\x01\xB1\xE5\xFF\x01\xB2\xED\xFF\x01\xB3\xE1"
-     "\xFF\x01"
-     "\xB4\xE5\xFF\x01\xB5\xED\xFF\x01\xB6\xD9\xFF\x01\xB7\xD9\xFF\x01\xA3\xA5"
-     "\xFF\x01"
-     "\x17\x84\x00\x80\x1B\x84\x00\x80\x1F\x84\x00\x80\x23\x84\x00\x80\xA6\xA5"
-     "\xFF\x01"
-     "\xA5\xB5\xFF\x01\x27\x84\x00\x80\xAB\x21\xFE\x01\xAA\x99\xFF\x01\x2B\x84"
-     "\x00\x80"
-     "\x2F\x84\x00\x80\xAF\x29\xFE\x01\xAE\x21\xFE\x01\xAD\x35\xFE\x01\xAC\x39"
-     "\xFE\x01"
-     "\x33\x84\x00\x80\xB3\xE9\xFE\x01\x37\x84\x00\x80\x3B\x84\x00\x80\xB6\x95"
-     "\xFE\x01"
-     "\x3F\x84\x00\x80\x43\x84\x00\x80\xB5\xE9\xFE\x01\xBA\xB1\xFE\x01\xBB\xB9"
-     "\xFE\x01"
-     "\x47\x84\x00\x80\x4B\x84\x00\x80\xBE\x75\x01\x00\xBF\x7D\x01\x00\xBC\x91"
-     "\xFE\x01"
-     "\xBD\x91\xFE\x01\xA8\x1D\xFE\x01\xA9\x2D\xFE\x01\xAA\x25\xFE\x01\xAB\x3D"
-     "\xFE\x01"
-     "\xAC\x25\xFE\x01\xAD\x51\xFE\x01\xAE\x51\xFE\x01\xAF\x51\xFE\x01\x4F\x84"
-     "\x00\x80"
-     "\x53\x84\x00\x80\x57\x84\x00\x80\x5B\x84\x00\x80\x5F\x84\x00\x80\x63\x84"
-     "\x00\x80"
-     "\x67\x84\x00\x80\x6B\x84\x00\x80\xB8\xD9\x01\x00\xB9\xD9\x01\x00\xBA\xED"
-     "\x01\x00"
-     "\xBB\xE1\x01\x00\xBC\xE1\x01\x00\xBD\xE1\x01\x00\xBE\xE1\x01\x00\xBF\xE1"
-     "\x01\x00"
-     "\xB0\x31\xFE\x01\xB1\x39\xFE\x01\xB2\x01\xFE\x01\xB3\x01\xFE\x01\xB4\xF5"
-     "\x01\x00"
-     "\xB5\xFD\x01\x00\xB6\xF5\x01\x00\xB7\xE9\x01\x00\xA3\xAD\xFD\x01\x6F\x84"
-     "\x00\x80"
-     "\xBE\x4C\x03\x00\x77\x84\x00\x80\x7B\x84\x00\x80\xA6\xD1\xFD\x01\xA5\xAD"
-     "\xFD\x01"
-     "\x7F\x84\x00\x80\xAB\xFD\xFD\x01\xAA\xF5\xFD\x01\x83\x84\x00\x80\x87\x84"
-     "\x00\x80"
-     "\xAF\x39\x02\x00\xAE\x31\x02\x00\xAD\xD5\xFD\x01\xAC\xD5\xFD\x01\x80\xF9"
-     "\x03\x00"
-     "\x81\xF9\x03\x00\x82\x4D\x00\x00\x85\x74\x20\x00\x8B\x84\x00\x80\x84\xD8"
-     "\x03\x00"
-     "\x84\xD4\x04\x00\x8F\x84\x00\x80\x86\x40\x04\x00\x87\x54\x03\x00\x93\x84"
-     "\x00\x80"
-     "\x97\x84\x00\x80\x9B\x84\x00\x80\x9F\x84\x00\x80\xA3\x84\x00\x80\xBE\xF0"
-     "\x05\x00"
-     "\xA8\x31\x02\x00\xA9\x31\x02\x00\xAA\x31\x02\x00\xAB\x31\x02\x00\xAC\x95"
-     "\x03\x00"
-     "\xAD\x9D\x03\x00\xAE\x95\x03\x00\xAF\x8D\x03\x00\xA7\x84\x00\x80\xAB\x84"
-     "\x00\x80"
-     "\xAF\x84\x00\x80\xB3\x84\x00\x80\x84\x7C\x07\x00\xB7\x84\x00\x80\xBB\x84"
-     "\x00\x80"
-     "\xBF\x84\x00\x80\xB8\xA9\x03\x00\xB9\xA9\x03\x00\xBA\x69\x03\x00\xBB\x69"
-     "\x03\x00"
-     "\xBC\x79\x03\x00\xBD\x79\x03\x00\xBE\x69\x03\x00\xBF\x69\x03\x00\xB0\xFD"
-     "\x03\x00"
-     "\xB1\xCD\x03\x00\xB2\xC5\x03\x00\xB3\xAD\x03\x00\xB4\xB9\x03\x00\xB5\xB9"
-     "\x03\x00"
-     "\xB6\xAD\x03\x00\xB7\xA5\x03\x00\xC3\x84\x00\x80\xC7\x84\x00\x80\xCB\x84"
-     "\x00\x80"
-     "\xCF\x84\x00\x80\xD3\x84\x00\x80\xD7\x84\x00\x80\xEF\xE8\x03\x00\xDB\x84"
-     "\x00\x80"
-     "\xE1\x90\x01\x00\xDF\x84\x00\x80\xE3\x6C\x03\x00\xE7\x84\x00\x80\x80\x25"
-     "\x00\x00"
-     "\x81\x29\x00\x00\x82\x1D\x00\x00\xEB\x84\x00\x80\xB3\xF9\x03\x00\xEF\x84"
-     "\x00\x80"
-     "\x86\x68\x07\x00\x87\x5C\x05\x00\xF3\x84\x00\x80\xB6\x5D\x02\x00\xB5\x5D"
-     "\x02\x00"
-     "\xF7\x84\x00\x80\xBB\x49\x02\x00\xBA\x79\x02\x00\xFB\x84\x00\x80\xFF\x84"
-     "\x00\x80"
-     "\xBF\x3D\x02\x00\xBE\x3D\x02\x00\xBD\x39\x02\x00\xBC\x51\x02\x00\x03\x85"
-     "\x00\x80"
-     "\xE1\x3C\xFE\x01\xBE\x40\x06\x00\xE3\xF0\x01\x00\x07\x85\x00\x80\x0B\x85"
-     "\x00\x80"
-     "\x0F\x85\x00\x80\x13\x85\x00\x80\x17\x85\x00\x80\x1B\x85\x00\x80\x1F\x85"
-     "\x00\x80"
-     "\x23\x83\x00\x80\x23\x85\x00\x80\x27\x85\x00\x80\x2B\x85\x00\x80\xEF\xE4"
-     "\x01\x00"
-     "\x2F\x85\x00\x80\x33\x85\x00\x80\xA3\x71\x03\x00\x37\x85\x00\x80\xA5\xD5"
-     "\x02\x00"
-     "\x3B\x85\x00\x80\x3F\x85\x00\x80\xA6\xD5\x02\x00\x43\x85\x00\x80\x47\x85"
-     "\x00\x80"
-     "\xAB\xC1\x02\x00\xAA\xF1\x02\x00\xAD\xB1\x02\x00\xAC\xD9\x02\x00\xAF\xB5"
-     "\x02\x00"
-     "\xAE\xB5\x02\x00\xE1\xDC\xFC\x01\xE3\x84\x00\x80\xE3\x50\x04\x00\xEF\x88"
-     "\x04\x00"
-     "\x80\xF5\x07\x00\x81\x09\x00\x00\x82\x3D\x00\x00\x4B\x85\x00\x80\x84\x24"
-     "\x01\x00"
-     "\x4F\x85\x00\x80\x53\x85\x00\x80\x57\x85\x00\x80\x5B\x85\x00\x80\xE1\x4C"
-     "\x04\x00"
-     "\xEF\xE4\x1C\x00\xE3\x74\x04\x00\xB3\x5D\x06\x00\x5F\x85\x00\x80\x86\x00"
-     "\x0C\x00"
-     "\x87\xE0\x03\x00\x63\x85\x00\x80\xB6\x05\x06\x00\xB5\x75\x06\x00\x67\x85"
-     "\x00\x80"
-     "\xBB\x11\x06\x00\xBA\x09\x06\x00\x6B\x85\x00\x80\x6F\x85\x00\x80\xBF\xD5"
-     "\x06\x00"
-     "\xBE\xD5\x06\x00\xBD\x01\x06\x00\xBC\x09\x06\x00\xA8\x8D\x06\x00\xA9\x95"
-     "\x06\x00"
-     "\xAA\x95\x06\x00\xAB\xA5\x06\x00\xAC\xBD\x06\x00\xAD\xC5\x06\x00\xAE\xC5"
-     "\x06\x00"
-     "\xAF\xF5\x06\x00\x73\x85\x00\x80\x77\x85\x00\x80\x7B\x85\x00\x80\x7F\x85"
-     "\x00\x80"
-     "\x83\x85\x00\x80\x87\x85\x00\x80\x8B\x85\x00\x80\x8F\x85\x00\x80\xB8\x75"
-     "\x06\x00"
-     "\xB9\x7D\x06\x00\xBA\x75\x06\x00\xBB\xCD\x07\x00\xBC\xD5\x07\x00\xBD\xDD"
-     "\x07\x00"
-     "\xBE\xD5\x07\x00\xBF\xCD\x07\x00\xB0\x8D\x06\x00\xB1\x95\x06\x00\xB2\x9D"
-     "\x06\x00"
-     "\xB3\x95\x06\x00\xB4\x51\x06\x00\xB5\x51\x06\x00\xB6\x51\x06\x00\xB7\x51"
-     "\x06\x00"
-     "\xA3\x1D\x07\x00\x8F\x15\xEC\x01\x93\x85\x00\x80\x97\x85\x00\x80\x9B\x85"
-     "\x00\x80"
-     "\xA6\x45\x07\x00\xA5\x35\x07\x00\x9F\x85\x00\x80\xAB\x51\x07\x00\xAA\x49"
-     "\x07\x00"
-     "\xA3\x85\x00\x80\xA7\x85\x00\x80\xAF\x95\x07\x00\xAE\x95\x07\x00\xAD\x41"
-     "\x07\x00"
-     "\xAC\x49\x07\x00\x9E\x45\xF9\x01\x9F\xA5\xF9\x01\x9C\x91\xFD\x01\x9D\x4D"
-     "\xF9\x01"
-     "\x9A\x55\xFD\x01\x9B\x5D\xFD\x01\x98\x05\xF1\x01\x99\x95\xFE\x01\x96\xA5"
-     "\xF2\x01"
-     "\x97\x61\xF1\x01\x94\x6D\xF5\x01\x95\x35\xF5\x01\x92\xE1\xE9\x01\x93\x85"
-     "\xF6\x01"
-     "\x90\x55\xEC\x01\x91\x5D\xE9\x01\xB1\xB1\x1D\x00\xB0\xA5\x1D\x00\xB3\x89"
-     "\x19\x00"
-     "\xB2\x41\x1C\x00\xB5\x01\x24\x00\xB4\xF5\x19\x00\xAB\x85\x00\x80\xAF\x85"
-     "\x00\x80"
-     "\xB3\x85\x00\x80\x82\xA9\x03\x00\x81\xA1\x03\x00\x80\x69\x00\x00\xA2\x15"
-     "\x05\x00"
-     "\xA3\x05\x02\x00\xA0\x15\x06\x00\xA1\xBD\x05\x00\xA1\xC5\x01\x00\xB7\x85"
-     "\x00\x80"
-     "\xA3\xCD\x0D\x00\xA2\xE5\x01\x00\xA5\x01\x08\x00\xA4\xDD\x0D\x00\xA7\xD1"
-     "\x09\x00"
-     "\xA6\xD9\x09\x00\xA9\x01\x14\x00\xA8\xA5\x08\x00\xAB\xC5\x14\x00\xAA\xCD"
-     "\x15\x00"
-     "\xAD\xD1\x11\x00\xAC\x01\x10\x00\xAF\x01\x1C\x00\xAE\x75\x11\x00\x82\x11"
-     "\xEF\x01"
-     "\x83\x29\xEF\x01\xBB\x85\x00\x80\xBF\x85\x00\x80\x86\xE1\xF5\x01\x87\x11"
-     "\xF6\x01"
-     "\x84\x39\xEA\x01\x85\x8D\xEA\x01\x8A\x79\xF6\x01\x8B\xE1\xF1\x01\xBE\xA4"
-     "\x0C\x00"
-     "\xC7\x85\x00\x80\x8E\xF9\xFD\x01\x8F\x31\xFE\x01\x8C\x3D\xF2\x01\x8D\x61"
-     "\xF2\x01"
-     "\x92\x49\xFE\x01\x93\x9D\xFE\x01\x87\x08\x0C\x00\x86\x6C\x0C\x00\x96\xB1"
-     "\xFA\x01"
-     "\x97\xF9\x05\x00\x94\x59\xFA\x01\x95\x59\xFA\x01\x9A\x61\x06\x00\x9B\xC1"
-     "\x06\x00"
-     "\xCB\x85\x00\x80\xCF\x85\x00\x80\xD3\x85\x00\x80\xD7\x85\x00\x80\x9C\xC9"
-     "\x01\x00"
-     "\xDB\x85\x00\x80\xA8\xAD\x05\x00\xA9\xB9\x05\x00\xAA\xCD\x05\x00\xAB\xDD"
-     "\x05\x00"
-     "\xAC\xCD\x05\x00\xAD\xFD\x05\x00\xAE\xF5\x05\x00\xAF\x1D\x05\x00\xDF\x85"
-     "\x00\x80"
-     "\xE3\x85\x00\x80\xE7\x85\x00\x80\xEB\x85\x00\x80\xEF\x85\x00\x80\xF3\x85"
-     "\x00\x80"
-     "\xF7\x85\x00\x80\xFB\x85\x00\x80\xB8\x75\x05\x00\xB9\x7D\x05\x00\xBA\x09"
-     "\x05\x00"
-     "\xBB\x09\x05\x00\xBC\x1D\x05\x00\xBD\x01\x05\x00\xBE\x01\x05\x00\xBF\x3D"
-     "\x05\x00"
-     "\xB0\x71\x05\x00\xB1\x71\x05\x00\xB2\x71\x05\x00\xB3\x71\x05\x00\xB4\x51"
-     "\x05\x00"
-     "\xB5\x51\x05\x00\xB6\x51\x05\x00\xB7\x4D\x05\x00\xB3\x45\x04\x00\xFF\x85"
-     "\x00\x80"
-     "\x03\x86\x00\x80\x07\x86\x00\x80\x0B\x86\x00\x80\xB6\x7D\x04\x00\xB5\x45"
-     "\x04\x00"
-     "\x0F\x86\x00\x80\xBB\xB5\x04\x00\xBA\xB5\x04\x00\x13\x86\x00\x80\x17\x86"
-     "\x00\x80"
-     "\xBF\x95\x04\x00\xBE\x95\x04\x00\xBD\xA5\x04\x00\xBC\xA5\x04\x00\x1B\x86"
-     "\x00\x80"
-     "\xA3\x01\x04\x00\x1F\x86\x00\x80\x23\x86\x00\x80\xA6\x39\x04\x00\x27\x86"
-     "\x00\x80"
-     "\x2B\x86\x00\x80\xA5\x01\x04\x00\xAA\xF1\x04\x00\xAB\xF1\x04\x00\x2F\x86"
-     "\x00\x80"
-     "\x84\xEC\x0D\x00\xAE\xD1\x04\x00\xAF\xD1\x04\x00\xAC\xE1\x04\x00\xAD\xE1"
-     "\x04\x00"
-     "\xE1\xD0\x06\x00\x84\x0C\x0C\x00\xE3\x28\x07\x00\xBE\x00\x0C\x00\x37\x86"
-     "\x00\x80"
-     "\xEF\x44\x03\x00\x86\xB8\x0C\x00\x87\x2C\x0D\x00\x3B\x86\x00\x80\xE3\x94"
-     "\x01\x00"
-     "\x3F\x86\x00\x80\xE1\xFC\x01\x00\x73\x84\x00\x80\x43\x86\x00\x80\xEF\xC8"
-     "\x06\x00"
-     "\x47\x86\x00\x80\x4B\x86\x00\x80\x4F\x86\x00\x80\xB3\x8D\x03\x00\x53\x86"
-     "\x00\x80"
-     "\xB5\x8D\x03\x00\x57\x86\x00\x80\x5B\x86\x00\x80\xB6\xB5\x03\x00\x5F\x86"
-     "\x00\x80"
-     "\x63\x86\x00\x80\xBB\x41\x03\x00\xBA\x49\x03\x00\xBD\x41\x03\x00\xBC\x59"
-     "\x03\x00"
-     "\xBF\x41\x03\x00\xBE\x49\x03\x00\xA3\x45\x0C\x00\xC3\x85\x00\x80\x33\x86"
-     "\x00\x80"
-     "\x67\x86\x00\x80\x6B\x86\x00\x80\xA6\x7D\x0C\x00\xA5\x45\x0C\x00\x6F\x86"
-     "\x00\x80"
-     "\xAB\x89\x0C\x00\xAA\x81\x0C\x00\x73\x86\x00\x80\x77\x86\x00\x80\xAF\x89"
-     "\x0C\x00"
-     "\xAE\x81\x0C\x00\xAD\x89\x0C\x00\xAC\x91\x0C\x00\x80\x15\x0F\x00\x81\x1D"
-     "\x0F\x00"
-     "\x82\x21\x0F\x00\xB3\x21\x0E\x00\x7B\x86\x00\x80\xB5\x21\x0E\x00\xB6\x25"
-     "\x0E\x00"
-     "\x7F\x86\x00\x80\x83\x86\x00\x80\x87\x86\x00\x80\xBA\xC1\x0E\x00\xBB\xC1"
-     "\x0E\x00"
-     "\xBC\xC1\x0E\x00\xBD\xC1\x0E\x00\xBE\xC1\x0E\x00\xBF\xC1\x0E\x00\xA8\xAD"
-     "\x0E\x00"
-     "\xA9\xED\x0E\x00\xAA\xE5\x0E\x00\xAB\xFD\x0E\x00\xAC\xE5\x0E\x00\xAD\xE9"
-     "\x0E\x00"
-     "\xAE\x39\x0E\x00\xAF\x39\x0E\x00\x8B\x86\x00\x80\x8F\x86\x00\x80\x93\x86"
-     "\x00\x80"
-     "\x97\x86\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xBD\x01\x00\x9B\x86"
-     "\x00\x80"
-     "\xB8\xCD\x0F\x00\xB9\xD5\x0F\x00\xBA\xD5\x0F\x00\xBB\xE5\x0F\x00\xBC\xFD"
-     "\x0F\x00"
-     "\xBD\x95\x0F\x00\xBE\x91\x0F\x00\xBF\x91\x0F\x00\xB0\x49\x0E\x00\xB1\x49"
-     "\x0E\x00"
-     "\xB2\x59\x0E\x00\xB3\x59\x0E\x00\xB4\x49\x0E\x00\xB5\x49\x0E\x00\xB6\xFD"
-     "\x0F\x00"
-     "\xB7\xF5\x0F\x00\xA3\x6D\x0F\x00\x9F\x86\x00\x80\xBE\x84\x01\x00\xA7\x86"
-     "\x00\x80"
-     "\xAB\x86\x00\x80\xA6\x69\x0F\x00\xA5\x6D\x0F\x00\xAF\x86\x00\x80\xAB\x8D"
-     "\x0F\x00"
-     "\xAA\x8D\x0F\x00\x86\x88\x00\x00\x87\x6C\x01\x00\xAF\x8D\x0F\x00\xAE\x8D"
-     "\x0F\x00"
-     "\xAD\x8D\x0F\x00\xAC\x8D\x0F\x00\xB3\x86\x00\x80\xB3\xED\x0E\x00\xB7\x86"
-     "\x00\x80"
-     "\xBB\x86\x00\x80\xB6\x91\x0E\x00\xBF\x86\x00\x80\xC3\x86\x00\x80\xB5\xE1"
-     "\x0E\x00"
-     "\xBA\xB5\x0E\x00\xBB\xBD\x0E\x00\xC7\x86\x00\x80\xCB\x86\x00\x80\xBE\x7D"
-     "\x01\x00"
-     "\xBF\x65\x01\x00\xBC\x99\x0E\x00\xBD\x99\x0E\x00\xA8\x11\x0E\x00\xA9\x25"
-     "\x0E\x00"
-     "\xAA\x21\x0E\x00\xAB\x39\x0E\x00\xAC\x2D\x0E\x00\xAD\x55\x0E\x00\xAE\x5D"
-     "\x0E\x00"
-     "\xAF\x51\x0E\x00\x84\xA8\x00\x00\xCF\x86\x00\x80\xD3\x86\x00\x80\xD7\x86"
-     "\x00\x80"
-     "\xDB\x86\x00\x80\xDF\x86\x00\x80\xE3\x86\x00\x80\xE7\x86\x00\x80\xB8\xED"
-     "\x01\x00"
-     "\xB9\x95\x01\x00\xBA\x95\x01\x00\xBB\xAD\x01\x00\xBC\xB5\x01\x00\xBD\x75"
-     "\x01\x00"
-     "\xBE\x7D\x01\x00\xBF\x75\x01\x00\xB0\x35\x0E\x00\xB1\x3D\x0E\x00\xB2\x09"
-     "\x0E\x00"
-     "\xB3\x09\x0E\x00\xB4\xFD\x01\x00\xB5\xE5\x01\x00\xB6\xE5\x01\x00\xB7\xD5"
-     "\x01\x00"
-     "\xA3\xA9\x0D\x00\xEB\x86\x00\x80\xEF\x86\x00\x80\xF3\x86\x00\x80\xF7\x86"
-     "\x00\x80"
-     "\xA6\xD5\x0D\x00\xA5\xA5\x0D\x00\xFB\x86\x00\x80\xAB\xF9\x0D\x00\xAA\xF1"
-     "\x0D\x00"
-     "\xFF\x86\x00\x80\x03\x87\x00\x80\xAF\x21\x02\x00\xAE\x39\x02\x00\xAD\xDD"
-     "\x0D\x00"
-     "\xAC\xDD\x0D\x00\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\x07\x87"
-     "\x00\x80"
-     "\x0B\x87\x00\x80\x0F\x87\x00\x80\x87\x90\x03\x00\x86\x7C\x04\x00\xBE\xEC"
-     "\x04\x00"
-     "\x17\x87\x00\x80\x1B\x87\x00\x80\x1F\x87\x00\x80\x23\x87\x00\x80\x27\x87"
-     "\x00\x80"
-     "\x2B\x87\x00\x80\x2F\x87\x00\x80\xB2\x2D\x0E\x00\xB3\x35\x0E\x00\xB0\x2D"
-     "\x0E\x00"
-     "\xB1\x25\x0E\x00\xB6\x2D\x0E\x00\xB7\x9D\x0F\x00\xB4\x2D\x0E\x00\xB5\x25"
-     "\x0E\x00"
-     "\xBA\xBD\x0F\x00\xBB\x8D\x0F\x00\xB8\xA5\x0F\x00\xB9\xBD\x0F\x00\xBE\x2D"
-     "\x0F\x00"
-     "\xBF\x15\x0F\x00\xBC\x95\x0F\x00\xBD\x25\x0F\x00\x33\x87\x00\x80\x37\x87"
-     "\x00\x80"
-     "\x3B\x87\x00\x80\x3F\x87\x00\x80\x43\x87\x00\x80\x47\x87\x00\x80\x4B\x87"
-     "\x00\x80"
-     "\x4F\x87\x00\x80\xAA\xA5\x0E\x00\xAB\xB5\x0E\x00\xA8\xC5\x0E\x00\xA9\xDD"
-     "\x0E\x00"
-     "\xAE\x9D\x0E\x00\xAF\x55\x0E\x00\xAC\xA5\x0E\x00\xAD\x95\x0E\x00\xA8\x0D"
-     "\x02\x00"
-     "\xA9\x15\x02\x00\xAA\x15\x02\x00\xAB\x4D\x02\x00\xAC\x59\x02\x00\xAD\x59"
-     "\x02\x00"
-     "\xAE\x4D\x02\x00\xAF\x45\x02\x00\x84\xA8\x05\x00\x53\x87\x00\x80\x57\x87"
-     "\x00\x80"
-     "\x5B\x87\x00\x80\x84\xB8\x04\x00\x5F\x87\x00\x80\x63\x87\x00\x80\x67\x87"
-     "\x00\x80"
-     "\xB8\xFD\x02\x00\xB9\x41\x01\x00\xBA\x41\x01\x00\xBB\x41\x01\x00\xBC\x41"
-     "\x01\x00"
-     "\xBD\x49\x01\x00\xBE\x71\x01\x00\xBF\x71\x01\x00\xB0\x09\x02\x00\xB1\x09"
-     "\x02\x00"
-     "\xB2\xCD\x02\x00\xB3\xC5\x02\x00\xB4\xDD\x02\x00\xB5\xC5\x02\x00\xB6\xCD"
-     "\x02\x00"
-     "\xB7\xC5\x02\x00\xE1\xD4\x0F\x00\xE3\x10\x0E\x00\xE3\xF4\x0E\x00\xE1\x0C"
-     "\x0E\x00"
-     "\x6B\x87\x00\x80\x6F\x87\x00\x80\x73\x87\x00\x80\x77\x87\x00\x80\x7B\x87"
-     "\x00\x80"
-     "\x7F\x87\x00\x80\xBE\x24\x04\x00\x87\x87\x00\x80\xEC\x00\x00\x00\xEF\x44"
-     "\x00\x00"
-     "\xEF\xCC\x0E\x00\x8B\x87\x00\x80\x82\x65\x00\x00\xB3\xD9\x02\x00\x80\x55"
-     "\x00\x00"
-     "\x81\x6D\x00\x00\xB6\x9D\x02\x00\x8F\x87\x00\x80\x93\x87\x00\x80\xB5\x95"
-     "\x02\x00"
-     "\xBA\x89\x02\x00\xBB\x89\x02\x00\x86\xA8\x04\x00\x87\xE0\x04\x00\xBE\x5D"
-     "\x02\x00"
-     "\xBF\x45\x02\x00\xBC\x5D\x02\x00\xBD\x55\x02\x00\xA3\x1D\x05\x00\x13\x87"
-     "\x00\x80"
-     "\x83\x87\x00\x80\x97\x87\x00\x80\x9B\x87\x00\x80\xA6\x59\x05\x00\xA5\x51"
-     "\x05\x00"
-     "\x9F\x87\x00\x80\xAB\x4D\x05\x00\xAA\x4D\x05\x00\xA3\x87\x00\x80\xA7\x87"
-     "\x00\x80"
-     "\xAF\x81\x05\x00\xAE\x99\x05\x00\xAD\x91\x05\x00\xAC\x99\x05\x00\xAB\x87"
-     "\x00\x80"
-     "\xB3\x29\x06\x00\xAF\x87\x00\x80\xB3\x87\x00\x80\xB6\x29\x06\x00\xB7\x87"
-     "\x00\x80"
-     "\xBB\x87\x00\x80\xB5\x29\x06\x00\xBA\xA5\x06\x00\xBB\xAD\x06\x00\xBF\x87"
-     "\x00\x80"
-     "\xC3\x87\x00\x80\xBE\xA5\x06\x00\xBF\xAD\x06\x00\xBC\xB5\x06\x00\xBD\xAD"
-     "\x06\x00"
-     "\xA8\xE5\x06\x00\xA9\xED\x06\x00\xAA\xE5\x06\x00\xAB\xFD\x06\x00\xAC\xE5"
-     "\x06\x00"
-     "\xAD\xED\x06\x00\xAE\xE5\x06\x00\xAF\x5D\x06\x00\xC7\x87\x00\x80\xCB\x87"
-     "\x00\x80"
-     "\xCF\x87\x00\x80\xD3\x87\x00\x80\xD7\x87\x00\x80\xDB\x87\x00\x80\xDF\x87"
-     "\x00\x80"
-     "\xE3\x87\x00\x80\xB8\xE9\x07\x00\xB9\xE9\x07\x00\xBA\xFD\x07\x00\xBB\xF5"
-     "\x07\x00"
-     "\xBC\xED\x07\x00\xBD\x45\x07\x00\xBE\x4D\x07\x00\xBF\x45\x07\x00\xB0\x25"
-     "\x06\x00"
-     "\xB1\x2D\x06\x00\xB2\x25\x06\x00\xB3\x3D\x06\x00\xB4\x25\x06\x00\xB5\x2D"
-     "\x06\x00"
-     "\xB6\x25\x06\x00\xB7\xD9\x07\x00\xA3\x6D\x07\x00\x82\x2D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\xE7\x87\x00\x80\xA6\x6D\x07\x00\xA5\x6D\x07\x00\xEB\x87"
-     "\x00\x80"
-     "\xAB\xE9\x07\x00\xAA\xE1\x07\x00\xEF\x87\x00\x80\xBE\xA0\x01\x00\xAF\xE9"
-     "\x07\x00"
-     "\xAE\xE1\x07\x00\xAD\xE9\x07\x00\xAC\xF1\x07\x00\xF3\x87\x00\x80\xB3\x91"
-     "\x06\x00"
-     "\x86\xE8\x00\x00\x87\x2C\x01\x00\xB6\x41\x01\x00\xF7\x87\x00\x80\xFB\x87"
-     "\x00\x80"
-     "\xB5\x51\x01\x00\xBA\x4D\x01\x00\xBB\x25\x01\x00\xFF\x87\x00\x80\x03\x88"
-     "\x00\x80"
-     "\xBE\x25\x01\x00\xBF\x2D\x01\x00\xBC\x31\x01\x00\xBD\x31\x01\x00\xB0\xAD"
-     "\x01\x00"
-     "\xB1\xC5\x01\x00\xB2\xC1\x01\x00\xB3\xC1\x01\x00\xB4\xC5\x01\x00\xB5\xCD"
-     "\x01\x00"
-     "\xB6\xF1\x01\x00\xB7\xF1\x01\x00\xB8\x01\x01\x00\xB9\x01\x01\x00\xBA\x01"
-     "\x01\x00"
-     "\xBB\x01\x01\x00\xBC\x01\x01\x00\xBD\x01\x01\x00\xBE\x01\x01\x00\xBF\x01"
-     "\x01\x00"
-     "\x07\x88\x00\x80\x0B\x88\x00\x80\x0F\x88\x00\x80\x13\x88\x00\x80\xA3\x86"
-     "\x00\x80"
-     "\x17\x88\x00\x80\x1B\x88\x00\x80\x1F\x88\x00\x80\xA8\x4D\x06\x00\xA9\x59"
-     "\x06\x00"
-     "\xAA\x3D\x06\x00\xAB\x35\x06\x00\xAC\xFD\x01\x00\xAD\xE5\x01\x00\xAE\xE5"
-     "\x01\x00"
-     "\xAF\xD5\x01\x00\xA3\xD5\x05\x00\x23\x88\x00\x80\x27\x88\x00\x80\x2B\x88"
-     "\x00\x80"
-     "\x2F\x88\x00\x80\xA6\x05\x02\x00\xA5\x15\x02\x00\x33\x88\x00\x80\xAB\x61"
-     "\x02\x00"
-     "\xAA\x09\x02\x00\x37\x88\x00\x80\x3B\x88\x00\x80\xAF\x69\x02\x00\xAE\x61"
-     "\x02\x00"
-     "\xAD\x75\x02\x00\xAC\x75\x02\x00\x3F\x88\x00\x80\x43\x88\x00\x80\x47\x88"
-     "\x00\x80"
-     "\x4B\x88\x00\x80\x4F\x88\x00\x80\xE1\x64\x05\x00\x53\x88\x00\x80\xE3\xEC"
-     "\x05\x00"
-     "\x80\x11\x00\x00\x81\x11\x00\x00\x82\x11\x00\x00\xEF\xF4\x06\x00\x57\x88"
-     "\x00\x80"
-     "\x5B\x88\x00\x80\x5F\x88\x00\x80\xBE\x30\x03\x00\x84\xC4\x02\x00\x67\x88"
-     "\x00\x80"
-     "\xB3\xE1\x03\x00\x84\xC0\x1C\x00\xB5\x51\x03\x00\x6B\x88\x00\x80\x6F\x88"
-     "\x00\x80"
-     "\xB6\x59\x03\x00\x73\x88\x00\x80\x77\x88\x00\x80\xBB\x71\x03\x00\xBA\x79"
-     "\x03\x00"
-     "\xBD\xB5\x00\x00\xBC\x69\x03\x00\xBF\xB5\x00\x00\xBE\xB5\x00\x00\x7B\x88"
-     "\x00\x80"
-     "\xE1\x94\x01\x00\x7F\x88\x00\x80\xE3\x5C\x02\x00\x86\x70\x1C\x00\x87\x44"
-     "\x03\x00"
-     "\x83\x88\x00\x80\x87\x88\x00\x80\x8B\x88\x00\x80\x8F\x88\x00\x80\x93\x88"
-     "\x00\x80"
-     "\x97\x88\x00\x80\x9B\x88\x00\x80\x9F\x88\x00\x80\xA3\x88\x00\x80\xEF\x78"
-     "\x02\x00"
-     "\xA8\x55\x02\x00\xA9\x5D\x02\x00\xAA\x61\x02\x00\xAB\x61\x02\x00\xAC\xD1"
-     "\x02\x00"
-     "\xAD\xD1\x02\x00\xAE\xD1\x02\x00\xAF\xD1\x02\x00\xA7\x88\x00\x80\xAB\x88"
-     "\x00\x80"
-     "\xAF\x88\x00\x80\xB3\x88\x00\x80\xB7\x88\x00\x80\xBB\x88\x00\x80\xBF\x88"
-     "\x00\x80"
-     "\xC3\x88\x00\x80\xB8\x69\x01\x00\xB9\x69\x01\x00\xBA\x09\x01\x00\xBB\x09"
-     "\x01\x00"
-     "\xBC\x19\x01\x00\xBD\x19\x01\x00\xBE\x09\x01\x00\xBF\x05\x01\x00\xB0\xB5"
-     "\x02\x00"
-     "\xB1\xBD\x02\x00\xB2\xB5\x02\x00\xB3\x69\x01\x00\xB4\x79\x01\x00\xB5\x79"
-     "\x01\x00"
-     "\xB6\x69\x01\x00\xB7\x61\x01\x00\xE1\xC4\x07\x00\xE3\xA4\x06\x00\xE3\xB8"
-     "\x06\x00"
-     "\xE1\x7C\x06\x00\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\xC7\x88"
-     "\x00\x80"
-     "\xCB\x88\x00\x80\xCF\x88\x00\x80\xBE\xE0\x1D\x00\xD7\x88\x00\x80\xDB\x88"
-     "\x00\x80"
-     "\xEF\x60\x00\x00\xEF\xE8\x06\x00\xDF\x88\x00\x80\xA3\xA9\x02\x00\xE3\x88"
-     "\x00\x80"
-     "\xE7\x88\x00\x80\xEB\x88\x00\x80\xEF\x88\x00\x80\xA6\x11\x02\x00\xA5\x19"
-     "\x02\x00"
-     "\xF3\x88\x00\x80\xAB\x39\x02\x00\xAA\x31\x02\x00\x86\x48\x1C\x00\x87\xCC"
-     "\x1C\x00"
-     "\xAF\xFD\x01\x00\xAE\xFD\x01\x00\xAD\xFD\x01\x00\xAC\x21\x02\x00\xA8\x85"
-     "\x1E\x00"
-     "\xA9\x91\x1E\x00\xAA\x91\x1E\x00\xAB\xAD\x1E\x00\xAC\xB5\x1E\x00\xAD\xD5"
-     "\x1E\x00"
-     "\xAE\xD1\x1E\x00\xAF\xC5\x1E\x00\xD3\x88\x00\x80\xF7\x88\x00\x80\xFB\x88"
-     "\x00\x80"
-     "\xFF\x88\x00\x80\x03\x89\x00\x80\x07\x89\x00\x80\x0B\x89\x00\x80\x0F\x89"
-     "\x00\x80"
-     "\xB8\x61\x1F\x00\xB9\x61\x1F\x00\xBA\x61\x1F\x00\xBB\x61\x1F\x00\xBC\x61"
-     "\x1F\x00"
-     "\xBD\x61\x1F\x00\xBE\x61\x1F\x00\xBF\x61\x1F\x00\xB0\xBD\x1E\x00\xB1\x85"
-     "\x1E\x00"
-     "\xB2\x8D\x1E\x00\xB3\x85\x1E\x00\xB4\x9D\x1E\x00\xB5\x85\x1E\x00\xB6\x8D"
-     "\x1E\x00"
-     "\xB7\x85\x1E\x00\xB3\x19\x1E\x00\x13\x89\x00\x80\x17\x89\x00\x80\x1B\x89"
-     "\x00\x80"
-     "\x1F\x89\x00\x80\xB6\x55\x1E\x00\xB5\x3D\x1E\x00\x23\x89\x00\x80\xBB\x41"
-     "\x1E\x00"
-     "\xBA\x79\x1E\x00\x27\x89\x00\x80\x2B\x89\x00\x80\xBF\x41\x1E\x00\xBE\x59"
-     "\x1E\x00"
-     "\xBD\x51\x1E\x00\xBC\x59\x1E\x00\x2F\x89\x00\x80\xA3\x5D\x1E\x00\x33\x89"
-     "\x00\x80"
-     "\x37\x89\x00\x80\xA6\x11\x1E\x00\x3B\x89\x00\x80\x3F\x89\x00\x80\xA5\x79"
-     "\x1E\x00"
-     "\xAA\x3D\x1E\x00\xAB\x05\x1E\x00\x84\xA4\x03\x00\xBE\xA8\x03\x00\xAE\x1D"
-     "\x1E\x00"
-     "\xAF\x05\x1E\x00\xAC\x1D\x1E\x00\xAD\x15\x1E\x00\xA8\xAD\x1E\x00\xA9\xB5"
-     "\x1E\x00"
-     "\xAA\xB5\x1E\x00\xAB\xC9\x1E\x00\xAC\xD9\x1E\x00\xAD\xD9\x1E\x00\xAE\xC9"
-     "\x1E\x00"
-     "\xAF\xC1\x1E\x00\x80\xED\x01\x00\x81\xF1\x01\x00\x82\xF1\x01\x00\x43\x89"
-     "\x00\x80"
-     "\x86\x90\x00\x00\x87\x74\x01\x00\x47\x89\x00\x80\x4B\x89\x00\x80\xB8\xC9"
-     "\x01\x00"
-     "\xB9\xC9\x01\x00\xBA\xD9\x01\x00\xBB\xD1\x01\x00\xBC\xF9\x01\x00\xBD\xF9"
-     "\x01\x00"
-     "\xBE\x99\x01\x00\xBF\x95\x01\x00\xB0\x45\x01\x00\xB1\x4D\x01\x00\xB2\x45"
-     "\x01\x00"
-     "\xB3\x5D\x01\x00\xB4\x45\x01\x00\xB5\x4D\x01\x00\xB6\x45\x01\x00\xB7\xF9"
-     "\x01\x00"
-     "\xB3\x3D\x1E\x00\x4F\x89\x00\x80\x53\x89\x00\x80\x57\x89\x00\x80\x5B\x89"
-     "\x00\x80"
-     "\xB6\x59\x1E\x00\xB5\x51\x1E\x00\x5F\x89\x00\x80\xBB\x89\x01\x00\xBA\x81"
-     "\x01\x00"
-     "\x63\x89\x00\x80\x67\x89\x00\x80\xBF\x89\x01\x00\xBE\x81\x01\x00\xBD\x89"
-     "\x01\x00"
-     "\xBC\x91\x01\x00\x6B\x89\x00\x80\x6F\x89\x00\x80\xA3\x75\x1E\x00\x73\x89"
-     "\x00\x80"
-     "\xA5\x19\x1E\x00\x77\x89\x00\x80\x7B\x89\x00\x80\xA6\x11\x1E\x00\x63\x88"
-     "\x00\x80"
-     "\x7F\x89\x00\x80\xAB\xC1\x01\x00\xAA\xC9\x01\x00\xAD\xC1\x01\x00\xAC\xD9"
-     "\x01\x00"
-     "\xAF\xC1\x01\x00\xAE\xC9\x01\x00\x83\x89\x00\x80\x87\x89\x00\x80\x8B\x89"
-     "\x00\x80"
-     "\x8F\x89\x00\x80\x93\x89\x00\x80\x84\x18\x02\x00\x97\x89\x00\x80\x9B\x89"
-     "\x00\x80"
-     "\x9F\x89\x00\x80\xA3\x89\x00\x80\xA7\x89\x00\x80\xAB\x89\x00\x80\xAF\x89"
-     "\x00\x80"
-     "\xB7\x89\x00\x80\xBB\x89\x00\x80\xBE\x70\x03\x00\x80\x69\x00\x00\x81\x69"
-     "\x00\x00"
-     "\x82\x79\x00\x00\x84\x00\x04\x00\x86\xF0\x04\x00\x87\x74\x03\x00\xBF\x89"
-     "\x00\x80"
-     "\xEF\x0C\x1F\x00\xC3\x89\x00\x80\xE1\xAC\x1E\x00\xC7\x89\x00\x80\xE3\xF0"
-     "\x1E\x00"
-     "\xCB\x89\x00\x80\xCF\x89\x00\x80\xD3\x89\x00\x80\xD7\x89\x00\x80\xDB\x89"
-     "\x00\x80"
-     "\xDF\x89\x00\x80\xE3\x89\x00\x80\xE7\x89\x00\x80\xEF\x54\x02\x00\xEB\x89"
-     "\x00\x80"
-     "\xEF\x89\x00\x80\xF3\x89\x00\x80\xE3\xA4\x02\x00\xF7\x89\x00\x80\xE1\x80"
-     "\x01\x00"
-     "\xFB\x89\x00\x80\xFF\x89\x00\x80\x03\x8A\x00\x80\x07\x8A\x00\x80\x0B\x8A"
-     "\x00\x80"
-     "\xB3\x45\x03\x00\x0F\x8A\x00\x80\x13\x8A\x00\x80\x17\x8A\x00\x80\x1B\x8A"
-     "\x00\x80"
-     "\xB6\x45\x03\x00\xB5\x55\x03\x00\x1F\x8A\x00\x80\xBB\x21\x03\x00\xBA\x49"
-     "\x03\x00"
-     "\xBE\xA0\x04\x00\x27\x8A\x00\x80\xBF\x29\x03\x00\xBE\x21\x03\x00\xBD\x35"
-     "\x03\x00"
-     "\xBC\x39\x03\x00\xA8\x39\x02\x00\xA9\x39\x02\x00\xAA\x8D\x02\x00\xAB\x85"
-     "\x02\x00"
-     "\xAC\x9D\x02\x00\xAD\x85\x02\x00\xAE\x85\x02\x00\xAF\xB5\x02\x00\x80\xED"
-     "\x01\x00"
-     "\x81\xF5\x01\x00\x82\xF5\x01\x00\x2B\x8A\x00\x80\x86\x90\x04\x00\x87\x04"
-     "\x05\x00"
-     "\x2F\x8A\x00\x80\x33\x8A\x00\x80\xB8\x45\x01\x00\xB9\x4D\x01\x00\xBA\x45"
-     "\x01\x00"
-     "\xBB\x5D\x01\x00\xBC\x49\x01\x00\xBD\x49\x01\x00\xBE\x79\x01\x00\xBF\x79"
-     "\x01\x00"
-     "\xB0\xCD\x02\x00\xB1\xA5\x02\x00\xB2\xAD\x02\x00\xB3\xA1\x02\x00\xB4\xA5"
-     "\x02\x00"
-     "\xB5\xAD\x02\x00\xB6\x9D\x02\x00\xB7\x7D\x01\x00\x37\x8A\x00\x80\x3B\x8A"
-     "\x00\x80"
-     "\x3F\x8A\x00\x80\x43\x8A\x00\x80\x47\x8A\x00\x80\x4B\x8A\x00\x80\x4F\x8A"
-     "\x00\x80"
-     "\xEF\x88\x01\x00\x84\xEC\x04\x00\xE1\x54\x1E\x00\x53\x8A\x00\x80\xE3\x54"
-     "\x01\x00"
-     "\x57\x8A\x00\x80\x5B\x8A\x00\x80\x5F\x8A\x00\x80\x63\x8A\x00\x80\xA3\x45"
-     "\x02\x00"
-     "\x67\x8A\x00\x80\x6B\x8A\x00\x80\x6F\x8A\x00\x80\x73\x8A\x00\x80\xA6\x45"
-     "\x02\x00"
-     "\xA5\x55\x02\x00\x77\x8A\x00\x80\xAB\x21\x02\x00\xAA\x49\x02\x00\x7B\x8A"
-     "\x00\x80"
-     "\x7F\x8A\x00\x80\xAF\x29\x02\x00\xAE\x21\x02\x00\xAD\x35\x02\x00\xAC\x39"
-     "\x02\x00"
-     "\xA8\x35\x06\x00\xA9\x3D\x06\x00\xAA\x51\x06\x00\xAB\x6D\x06\x00\xAC\x79"
-     "\x06\x00"
-     "\xAD\x65\x06\x00\xAE\x6D\x06\x00\xAF\x65\x06\x00\x23\x8A\x00\x80\x83\x8A"
-     "\x00\x80"
-     "\x87\x8A\x00\x80\x8B\x8A\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x8F\x8A\x00\x80\xB8\xE5\x06\x00\xB9\xE9\x06\x00\xBA\xF9\x06\x00\xBB\xF9"
-     "\x06\x00"
-     "\xBC\xE9\x06\x00\xBD\xE9\x06\x00\xBE\x9D\x06\x00\xBF\x95\x06\x00\xB0\x1D"
-     "\x06\x00"
-     "\xB1\xE5\x06\x00\xB2\xED\x06\x00\xB3\xE5\x06\x00\xB4\xFD\x06\x00\xB5\xE1"
-     "\x06\x00"
-     "\xB6\xE1\x06\x00\xB7\xE1\x06\x00\xB3\xD9\x06\x00\xBE\xD0\x03\x00\x93\x8A"
-     "\x00\x80"
-     "\x97\x8A\x00\x80\x9B\x8A\x00\x80\xB6\xE5\x06\x00\xB5\xF1\x06\x00\x9F\x8A"
-     "\x00\x80"
-     "\xBB\x21\x06\x00\xBA\xD9\x06\x00\x86\x98\x00\x00\x87\x78\x03\x00\xBF\x25"
-     "\x06\x00"
-     "\xBE\x39\x06\x00\xBD\x31\x06\x00\xBC\x39\x06\x00\xA3\x8A\x00\x80\xA3\x9D"
-     "\x06\x00"
-     "\xA7\x8A\x00\x80\xAB\x8A\x00\x80\xA6\xA1\x06\x00\xAF\x8A\x00\x80\xB3\x8A"
-     "\x00\x80"
-     "\xA5\xB5\x06\x00\xAA\x9D\x06\x00\xAB\x65\x06\x00\xB7\x8A\x00\x80\xBB\x8A"
-     "\x00\x80"
-     "\xAE\x7D\x06\x00\xAF\x61\x06\x00\xAC\x7D\x06\x00\xAD\x75\x06\x00\xA8\xED"
-     "\x07\x00"
-     "\xA9\x29\x06\x00\xAA\x31\x06\x00\xAB\x31\x06\x00\xAC\x91\x06\x00\xAD\x91"
-     "\x06\x00"
-     "\xAE\x91\x06\x00\xAF\x91\x06\x00\xBF\x8A\x00\x80\xC3\x8A\x00\x80\xC7\x8A"
-     "\x00\x80"
-     "\xCB\x8A\x00\x80\xCF\x8A\x00\x80\xD3\x8A\x00\x80\xD7\x8A\x00\x80\xDB\x8A"
-     "\x00\x80"
-     "\xB8\x85\x06\x00\xB9\x8D\x06\x00\xBA\x85\x06\x00\xBB\x9D\x06\x00\xBC\x8D"
-     "\x06\x00"
-     "\xBD\xBD\x06\x00\xBE\xB5\x06\x00\xBF\x79\x01\x00\xB0\xF1\x06\x00\xB1\xF1"
-     "\x06\x00"
-     "\xB2\xF1\x06\x00\xB3\xC5\x06\x00\xB4\xC1\x06\x00\xB5\xC1\x06\x00\xB6\xC1"
-     "\x06\x00"
-     "\xB7\xC1\x06\x00\xB3\xB9\x06\x00\xDF\x8A\x00\x80\xE3\x8A\x00\x80\xE7\x8A"
-     "\x00\x80"
-     "\xEB\x8A\x00\x80\xB6\x11\x06\x00\xB5\x19\x06\x00\xEF\x8A\x00\x80\xBB\x35"
-     "\x06\x00"
-     "\xBA\x35\x06\x00\xF3\x8A\x00\x80\xF7\x8A\x00\x80\xBF\x05\x06\x00\xBE\x05"
-     "\x06\x00"
-     "\xBD\x11\x06\x00\xBC\x25\x06\x00\xA5\x41\x06\x00\xFB\x8A\x00\x80\xFF\x8A"
-     "\x00\x80"
-     "\xA6\x49\x06\x00\x81\x15\x00\x00\x80\x79\x00\x00\xA3\xE1\x06\x00\x82\x15"
-     "\x00\x00"
-     "\xAD\x49\x06\x00\xAC\x7D\x06\x00\xAF\x5D\x06\x00\xAE\x5D\x06\x00\x84\x34"
-     "\x01\x00"
-     "\xB3\x89\x00\x80\xAB\x6D\x06\x00\xAA\x6D\x06\x00\xBE\xCC\x03\x00\x07\x8B"
-     "\x00\x80"
-     "\xB3\x95\x02\x00\x0B\x8B\x00\x80\xB5\xD9\x02\x00\x0F\x8B\x00\x80\x13\x8B"
-     "\x00\x80"
-     "\xB6\xD1\x02\x00\x86\x80\x0C\x00\x87\x38\x03\x00\xBB\xC5\x02\x00\xBA\xC5"
-     "\x02\x00"
-     "\xBD\x15\x03\x00\xBC\x15\x03\x00\xBF\x15\x03\x00\xBE\x15\x03\x00\x17\x8B"
-     "\x00\x80"
-     "\x1B\x8B\x00\x80\xEF\x58\x06\x00\x84\x40\x03\x00\x1F\x8B\x00\x80\x23\x8B"
-     "\x00\x80"
-     "\x27\x8B\x00\x80\x2B\x8B\x00\x80\x2F\x8B\x00\x80\x33\x8B\x00\x80\x37\x8B"
-     "\x00\x80"
-     "\x3B\x8B\x00\x80\xE1\x38\x06\x00\x3F\x8B\x00\x80\xE3\x24\x06\x00\xBE\x70"
-     "\x0C\x00"
-     "\xAC\x49\x02\x00\xAD\x49\x02\x00\xAE\x5D\x02\x00\xAF\x55\x02\x00\xA8\x1D"
-     "\x02\x00"
-     "\xA9\x05\x02\x00\xAA\x05\x02\x00\xAB\x5D\x02\x00\x84\xA8\x0C\x00\x43\x8B"
-     "\x00\x80"
-     "\x47\x8B\x00\x80\x4B\x8B\x00\x80\xBE\xBC\x0D\x00\x4F\x8B\x00\x80\x53\x8B"
-     "\x00\x80"
-     "\x57\x8B\x00\x80\xBC\x4D\x03\x00\xBD\x55\x03\x00\xBE\x55\x03\x00\xBF\x65"
-     "\x03\x00"
-     "\xB8\xE9\x03\x00\xB9\xE9\x03\x00\xBA\x5D\x03\x00\xBB\x55\x03\x00\xB4\xC9"
-     "\x03\x00"
-     "\xB5\xC9\x03\x00\xB6\xD9\x03\x00\xB7\xD9\x03\x00\xB0\x19\x02\x00\xB1\x19"
-     "\x02\x00"
-     "\xB2\xD9\x03\x00\xB3\xD9\x03\x00\x5B\x8B\x00\x80\xE3\xE4\x00\x00\x5F\x8B"
-     "\x00\x80"
-     "\xE1\xBC\x01\x00\x63\x8B\x00\x80\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D"
-     "\x00\x00"
-     "\x67\x8B\x00\x80\x6B\x8B\x00\x80\x6F\x8B\x00\x80\x77\x8B\x00\x80\x7B\x8B"
-     "\x00\x80"
-     "\xEF\xCC\x03\x00\x7F\x8B\x00\x80\x83\x8B\x00\x80\xA3\xDD\x03\x00\x87\x8B"
-     "\x00\x80"
-     "\x86\xE8\x0C\x00\x87\x60\x0D\x00\x8B\x8B\x00\x80\xA6\x99\x03\x00\xA5\x91"
-     "\x03\x00"
-     "\x8F\x8B\x00\x80\xAB\x8D\x03\x00\xAA\x8D\x03\x00\x93\x8B\x00\x80\x97\x8B"
-     "\x00\x80"
-     "\xAF\x5D\x02\x00\xAE\x5D\x02\x00\xAD\x5D\x02\x00\xAC\x5D\x02\x00\x9B\x8B"
-     "\x00\x80"
-     "\x9F\x8B\x00\x80\xA3\x8B\x00\x80\xA7\x8B\x00\x80\xAB\x8B\x00\x80\xAF\x8B"
-     "\x00\x80"
-     "\xB3\x8B\x00\x80\xEF\xE0\x01\x00\x84\xBC\x0C\x00\xE1\x8C\x06\x00\xB7\x8B"
-     "\x00\x80"
-     "\xE3\x1C\x06\x00\xBB\x8B\x00\x80\xBF\x8B\x00\x80\xC3\x8B\x00\x80\xC7\x8B"
-     "\x00\x80"
-     "\xB3\xD5\x01\x00\xCB\x8B\x00\x80\xCF\x8B\x00\x80\xD3\x8B\x00\x80\xD7\x8B"
-     "\x00\x80"
-     "\xB6\x91\x01\x00\xB5\x99\x01\x00\xDB\x8B\x00\x80\xBB\xBD\x01\x00\xBA\xBD"
-     "\x01\x00"
-     "\xDF\x8B\x00\x80\xE7\x8B\x00\x80\xBF\x9D\x01\x00\xBE\x9D\x01\x00\xBD\x9D"
-     "\x01\x00"
-     "\xBC\x9D\x01\x00\xA8\x05\x0E\x00\xA9\x09\x0E\x00\xAA\x1D\x0E\x00\xAB\x15"
-     "\x0E\x00"
-     "\xAC\x51\x0E\x00\xAD\x51\x0E\x00\xAE\x49\x0E\x00\xAF\x49\x0E\x00\x73\x8B"
-     "\x00\x80"
-     "\x82\xCD\x0F\x00\x81\xF5\x0F\x00\x80\xFD\x0F\x00\xE3\x8B\x00\x80\xEB\x8B"
-     "\x00\x80"
-     "\x86\x1C\x00\x00\x87\xB0\x03\x00\xB8\xE9\x0E\x00\xB9\xE9\x0E\x00\xBA\xFD"
-     "\x0E\x00"
-     "\xBB\xF5\x0E\x00\xBC\xED\x0E\x00\xBD\x55\x0F\x00\xBE\x51\x0F\x00\xBF\x4D"
-     "\x0F\x00"
-     "\xB0\x39\x0E\x00\xB1\x39\x0E\x00\xB2\x09\x0E\x00\xB3\x09\x0E\x00\xB4\x19"
-     "\x0E\x00"
-     "\xB5\x19\x0E\x00\xB6\x0D\x0E\x00\xB7\xD9\x0E\x00\xA3\x95\x0E\x00\xEF\x8B"
-     "\x00\x80"
-     "\xF3\x8B\x00\x80\xF7\x8B\x00\x80\xFB\x8B\x00\x80\xA6\xD1\x0E\x00\xA5\xD9"
-     "\x0E\x00"
-     "\xFF\x8B\x00\x80\xAB\xFD\x0E\x00\xAA\xFD\x0E\x00\x03\x8C\x00\x80\x07\x8C"
-     "\x00\x80"
-     "\xAF\xDD\x0E\x00\xAE\xDD\x0E\x00\xAD\xDD\x0E\x00\xAC\xDD\x0E\x00\x0B\x8C"
-     "\x00\x80"
-     "\xB3\xFD\x0F\x00\x0F\x8C\x00\x80\x13\x8C\x00\x80\xB6\x81\x0F\x00\x17\x8C"
-     "\x00\x80"
-     "\x1B\x8C\x00\x80\xB5\x99\x0F\x00\xBA\x8D\x0F\x00\xBB\x65\x0F\x00\x1F\x8C"
-     "\x00\x80"
-     "\x23\x8C\x00\x80\xBE\x7D\x0F\x00\xBF\x65\x0F\x00\xBC\x7D\x0F\x00\xBD\x75"
-     "\x0F\x00"
-     "\xA8\x2D\x0E\x00\xA9\x35\x0E\x00\xAA\x31\x0E\x00\xAB\x31\x0E\x00\xAC\x55"
-     "\x0E\x00"
-     "\xAD\x45\x0E\x00\xAE\x45\x0E\x00\xAF\x75\x0E\x00\x27\x8C\x00\x80\x2B\x8C"
-     "\x00\x80"
-     "\x2F\x8C\x00\x80\x33\x8C\x00\x80\x37\x8C\x00\x80\x3B\x8C\x00\x80\x3F\x8C"
-     "\x00\x80"
-     "\x43\x8C\x00\x80\xB8\xE9\x0E\x00\xB9\xF5\x0E\x00\xBA\xFD\x0E\x00\xBB\xF5"
-     "\x0E\x00"
-     "\xBC\xED\x0E\x00\xBD\x91\x0E\x00\xBE\x91\x0E\x00\xBF\x91\x0E\x00\xB0\x0D"
-     "\x0E\x00"
-     "\xB1\xE5\x0E\x00\xB2\xED\x0E\x00\xB3\xE5\x0E\x00\xB4\xFD\x0E\x00\xB5\xE5"
-     "\x0E\x00"
-     "\xB6\xED\x0E\x00\xB7\xE5\x0E\x00\xA3\xB9\x0E\x00\x82\x2D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x47\x8C\x00\x80\xA6\xC5\x0E\x00\xA5\xDD\x0E\x00\x4B\x8C"
-     "\x00\x80"
-     "\xAB\x21\x0E\x00\xAA\xC9\x0E\x00\x4F\x8C\x00\x80\xBE\x2C\x01\x00\xAF\x21"
-     "\x0E\x00"
-     "\xAE\x39\x0E\x00\xAD\x31\x0E\x00\xAC\x39\x0E\x00\x57\x8C\x00\x80\xB6\x55"
-     "\x01\x00"
-     "\xB5\x45\x01\x00\x53\x8C\x00\x80\xB3\x55\x01\x00\x5B\x8C\x00\x80\x86\x78"
-     "\x00\x00"
-     "\x87\x5C\x00\x00\xBF\x39\x01\x00\xBE\x31\x01\x00\xBD\x25\x01\x00\xBC\x25"
-     "\x01\x00"
-     "\xBB\x31\x01\x00\xBA\x59\x01\x00\x03\x8B\x00\x80\x5F\x8C\x00\x80\x63\x8C"
-     "\x00\x80"
-     "\x67\x8C\x00\x80\x84\x04\x03\x00\xA3\x89\x02\x00\x6B\x8C\x00\x80\xA5\x99"
-     "\x02\x00"
-     "\xA6\x89\x02\x00\x6F\x8C\x00\x80\xBF\x28\x39\x00\x73\x8C\x00\x80\xAA\x85"
-     "\x02\x00"
-     "\xAB\xED\x02\x00\xAC\xF9\x02\x00\xAD\xF9\x02\x00\xAE\xED\x02\x00\xAF\xE5"
-     "\x02\x00"
-     "\xE3\x58\x02\x00\xEF\xC0\x0E\x00\xE1\x88\x01\x00\x77\x8C\x00\x80\x7B\x8C"
-     "\x00\x80"
-     "\x7F\x8C\x00\x80\x83\x8C\x00\x80\x87\x8C\x00\x80\x8B\x8C\x00\x80\x8F\x8C"
-     "\x00\x80"
-     "\x93\x8C\x00\x80\x97\x8C\x00\x80\xEF\x28\x02\x00\xE3\x28\x0E\x00\x9B\x8C"
-     "\x00\x80"
-     "\xE1\x44\x0E\x00\xA9\xB5\x02\x00\xA8\x69\x0D\x00\xAB\x01\x02\x00\xAA\x09"
-     "\x02\x00"
-     "\xAD\x01\x02\x00\xAC\x19\x02\x00\xAF\x31\x02\x00\xAE\x01\x02\x00\xBE\x00"
-     "\x04\x00"
-     "\x9F\x8C\x00\x80\xA3\x8C\x00\x80\xA7\x8C\x00\x80\xAB\x8C\x00\x80\xAF\x8C"
-     "\x00\x80"
-     "\xB3\x8C\x00\x80\xB7\x8C\x00\x80\xB9\xE5\x03\x00\xB8\xE5\x03\x00\xBB\xE5"
-     "\x03\x00"
-     "\xBA\xE5\x03\x00\xBD\xE5\x03\x00\xBC\xE5\x03\x00\xBF\xE5\x03\x00\xBE\xE5"
-     "\x03\x00"
-     "\xB1\x21\x02\x00\xB0\x49\x02\x00\xB3\x25\x02\x00\xB2\x25\x02\x00\xB5\x29"
-     "\x02\x00"
-     "\xB4\x21\x02\x00\xB7\x15\x02\x00\xB6\x15\x02\x00\xA8\xC1\x02\x00\xA9\xD1"
-     "\x02\x00"
-     "\xAA\xF5\x02\x00\xAB\x0D\x01\x00\xAC\x15\x01\x00\xAD\x05\x01\x00\xAE\x05"
-     "\x01\x00"
-     "\xAF\x39\x01\x00\xBB\x8C\x00\x80\xBF\x8C\x00\x80\xC7\x8C\x00\x80\xCB\x8C"
-     "\x00\x80"
-     "\xCF\x8C\x00\x80\xD3\x8C\x00\x80\xD7\x8C\x00\x80\xDB\x8C\x00\x80\xB8\x2D"
-     "\x01\x00"
-     "\xB9\x3D\x01\x00\xBA\xED\x01\x00\xBB\xE5\x01\x00\xBC\xFD\x01\x00\xBD\xE5"
-     "\x01\x00"
-     "\xBE\xED\x01\x00\xBF\xE5\x01\x00\xB0\x2D\x01\x00\xB1\x35\x01\x00\xB2\x3D"
-     "\x01\x00"
-     "\xB3\x35\x01\x00\xB4\x2D\x01\x00\xB5\x15\x01\x00\xB6\x1D\x01\x00\xB7\x15"
-     "\x01\x00"
-     "\x80\x3D\x01\x00\x81\xA5\x00\x00\x82\xAD\x00\x00\xEF\xD8\x00\x00\x86\xB0"
-     "\x05\x00"
-     "\x87\xD8\x05\x00\xDF\x8C\x00\x80\xEF\xD4\x0F\x00\x84\x6C\x04\x00\xE1\xF4"
-     "\x0E\x00"
-     "\xE3\x8C\x00\x80\xE3\x1C\x0F\x00\xE7\x8C\x00\x80\xE1\x94\x01\x00\xEB\x8C"
-     "\x00\x80"
-     "\xE3\x0C\x0E\x00\xB3\x3D\x02\x00\xEF\x8C\x00\x80\xF3\x8C\x00\x80\xF7\x8C"
-     "\x00\x80"
-     "\xFB\x8C\x00\x80\xB6\xC5\x01\x00\xB5\xDD\x01\x00\xFF\x8C\x00\x80\xBB\xB1"
-     "\x01\x00"
-     "\xBA\xA9\x01\x00\x03\x8D\x00\x80\x07\x8D\x00\x80\xBF\x91\x01\x00\xBE\xA9"
-     "\x01\x00"
-     "\xBD\xA1\x01\x00\xBC\xA9\x01\x00\xC3\x8C\x00\x80\x0B\x8D\x00\x80\xAA\x45"
-     "\x06\x00"
-     "\xAB\x5D\x06\x00\xAC\x45\x06\x00\xAD\x4D\x06\x00\xAE\x45\x06\x00\xAF\x7D"
-     "\x06\x00"
-     "\x0F\x8D\x00\x80\x13\x8D\x00\x80\x17\x8D\x00\x80\xA3\xD1\x05\x00\x1B\x8D"
-     "\x00\x80"
-     "\xA5\x31\x06\x00\xA6\x29\x06\x00\x1F\x8D\x00\x80\x23\x8D\x00\x80\x82\x1D"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x80\x1D\x00\x00\x27\x8D\x00\x80\x2B\x8D\x00\x80\x2F\x8D"
-     "\x00\x80"
-     "\xBE\x94\x03\x00\x33\x8D\x00\x80\x37\x8D\x00\x80\x86\x48\x03\x00\x87\xCC"
-     "\x03\x00"
-     "\x3B\x8D\x00\x80\x3F\x8D\x00\x80\x43\x8D\x00\x80\x47\x8D\x00\x80\xA8\xA9"
-     "\x07\x00"
-     "\xA9\xA9\x07\x00\xAA\xB9\x07\x00\xAB\xB9\x07\x00\xAC\xA9\x07\x00\xAD\xA9"
-     "\x07\x00"
-     "\xAE\x01\x07\x00\xAF\x35\x07\x00\x4B\x8D\x00\x80\x4F\x8D\x00\x80\x53\x8D"
-     "\x00\x80"
-     "\x57\x8D\x00\x80\x5B\x8D\x00\x80\x5F\x8D\x00\x80\x63\x8D\x00\x80\x67\x8D"
-     "\x00\x80"
-     "\xB8\x2D\x07\x00\xB9\xC1\x00\x00\xBA\xE9\x00\x00\xBB\xE9\x00\x00\xBC\xF9"
-     "\x00\x00"
-     "\xBD\xE5\x00\x00\xBE\xE5\x00\x00\xBF\x9D\x00\x00\xB0\x51\x07\x00\xB1\x5D"
-     "\x07\x00"
-     "\xB2\x2D\x07\x00\xB3\x25\x07\x00\xB4\x3D\x07\x00\xB5\x25\x07\x00\xB6\x25"
-     "\x07\x00"
-     "\xB7\x15\x07\x00\xB3\x31\x06\x00\x6B\x8D\x00\x80\x6F\x8D\x00\x80\x73\x8D"
-     "\x00\x80"
-     "\x77\x8D\x00\x80\xB6\x29\x06\x00\xB5\x21\x06\x00\x7B\x8D\x00\x80\xBB\x99"
-     "\x06\x00"
-     "\xBA\x95\x06\x00\x7F\x8D\x00\x80\x83\x8D\x00\x80\xBF\xE1\x06\x00\xBE\xF9"
-     "\x06\x00"
-     "\xBD\xF1\x06\x00\xBC\xF9\x06\x00\x87\x8D\x00\x80\xA3\x75\x06\x00\x8B\x8D"
-     "\x00\x80"
-     "\x8F\x8D\x00\x80\xA6\x6D\x06\x00\x93\x8D\x00\x80\x97\x8D\x00\x80\xA5\x65"
-     "\x06\x00"
-     "\xAA\xD1\x06\x00\xAB\xDD\x06\x00\x9B\x8D\x00\x80\x9F\x8D\x00\x80\xAE\xBD"
-     "\x06\x00"
-     "\xAF\xA5\x06\x00\xAC\xBD\x06\x00\xAD\xB5\x06\x00\xA8\xE5\x01\x00\xA9\xF5"
-     "\x01\x00"
-     "\xAA\xFD\x01\x00\xAB\xF5\x01\x00\xAC\xED\x01\x00\xAD\x35\x01\x00\xAE\x3D"
-     "\x01\x00"
-     "\xAF\x35\x01\x00\x80\xF9\x00\x00\x81\xCD\x00\x00\x82\xC5\x00\x00\x84\x60"
-     "\x01\x00"
-     "\xBE\x78\x01\x00\xA7\x8D\x00\x80\x87\xAC\x00\x00\x86\x90\x01\x00\xB8\xD1"
-     "\x00\x00"
-     "\xB9\xD9\x00\x00\xBA\xE1\x00\x00\xBB\xE1\x00\x00\xBC\x91\x00\x00\xBD\x9D"
-     "\x00\x00"
-     "\xBE\x95\x00\x00\xBF\x89\x00\x00\xB0\x4D\x01\x00\xB1\x55\x01\x00\xB2\x5D"
-     "\x01\x00"
-     "\xB3\x55\x01\x00\xB4\x4D\x01\x00\xB5\xF1\x00\x00\xB6\xF1\x00\x00\xB7\xF1"
-     "\x00\x00"
-     "\xB3\x75\x02\x00\xAB\x8D\x00\x80\xAF\x8D\x00\x80\xB3\x8D\x00\x80\xB7\x8D"
-     "\x00\x80"
-     "\xB6\xB5\x02\x00\xB5\x65\x02\x00\xBB\x8D\x00\x80\xBB\x91\x02\x00\xBA\x89"
-     "\x02\x00"
-     "\xBF\x8D\x00\x80\xC3\x8D\x00\x80\xBF\x35\x03\x00\xBE\x89\x02\x00\xBD\x81"
-     "\x02\x00"
-     "\xBC\x89\x02\x00\xC7\x8D\x00\x80\xA3\x31\x02\x00\xCB\x8D\x00\x80\x84\xC0"
-     "\x03\x00"
-     "\xA6\xF1\x02\x00\xCF\x8D\x00\x80\xD3\x8D\x00\x80\xA5\x21\x02\x00\xAA\xCD"
-     "\x02\x00"
-     "\xAB\xD5\x02\x00\xD7\x8D\x00\x80\xDB\x8D\x00\x80\xAE\xCD\x02\x00\xAF\x71"
-     "\x03\x00"
-     "\xAC\xCD\x02\x00\xAD\xC5\x02\x00\xAB\x8D\x00\x00\xAA\x8D\x00\x00\xA9\x8D"
-     "\x00\x00"
-     "\xA8\x39\x03\x00\xAF\xBD\x00\x00\xAE\xBD\x00\x00\xAD\x85\x00\x00\xAC\x8D"
-     "\x00\x00"
-     "\xAA\x00\x00\x00\xAB\x00\x00\x00\xDF\x8D\x00\x80\xE3\x8D\x00\x80\xE7\x8D"
-     "\x00\x80"
-     "\xEB\x8D\x00\x80\xEF\x8D\x00\x80\xF3\x8D\x00\x80\xBB\x7D\x00\x00\xBA\x7D"
-     "\x00\x00"
-     "\xB9\x7D\x00\x00\xB8\x7D\x00\x00\xBF\xDD\x01\x00\xBE\xDD\x01\x00\xBD\xDD"
-     "\x01\x00"
-     "\xBC\xDD\x01\x00\xB3\xB9\x00\x00\xB2\xB1\x00\x00\xB1\xA1\x00\x00\xB0\xAD"
-     "\x00\x00"
-     "\xB7\x5D\x00\x00\xB6\x5D\x00\x00\xB5\x95\x00\x00\xB4\x95\x00\x00\xF7\x8D"
-     "\x00\x80"
-     "\xFB\x8D\x00\x80\xFF\x8D\x00\x80\x03\x8E\x00\x80\x81\x35\x00\x00\x80\x0D"
-     "\x00\x00"
-     "\x07\x8E\x00\x80\x82\x35\x00\x00\xBE\xAC\x3D\x00\x0B\x8E\x00\x80\x0F\x8E"
-     "\x00\x80"
-     "\x85\x68\x3D\x00\x17\x8E\x00\x80\x1B\x8E\x00\x80\x86\x38\x3C\x00\x87\xC0"
-     "\x02\x00"
-     "\xB3\x49\x01\x00\x1F\x8E\x00\x80\xD0\x00\x00\x00\x23\x8E\x00\x80\x27\x8E"
-     "\x00\x80"
-     "\xB6\x49\x01\x00\xB5\x49\x01\x00\x2B\x8E\x00\x80\xBB\x29\x01\x00\xBA\x25"
-     "\x01\x00"
-     "\x2F\x8E\x00\x80\x33\x8E\x00\x80\xBF\x1D\x01\x00\xBE\x1D\x01\x00\xBD\x21"
-     "\x01\x00"
-     "\xBC\x29\x01\x00\xE3\x34\x36\x00\xE1\x0C\x06\x00\xE1\xB0\x02\x00\xE3\x3C"
-     "\x06\x00"
-     "\x37\x8E\x00\x80\x3B\x8E\x00\x80\x3F\x8E\x00\x80\x43\x8E\x00\x80\x84\x2C"
-     "\x3F\x00"
-     "\xBE\xA0\x3F\x00\x47\x8E\x00\x80\x4B\x8E\x00\x80\xEF\x7C\x37\x00\x4F\x8E"
-     "\x00\x80"
-     "\x53\x8E\x00\x80\xEF\x18\x01\x00\x57\x8E\x00\x80\x5B\x8E\x00\x80\x86\x68"
-     "\x3E\x00"
-     "\x87\xCC\x3F\x00\x5F\x8E\x00\x80\x63\x8E\x00\x80\x67\x8E\x00\x80\xEF\x94"
-     "\x00\x00"
-     "\x6B\x8E\x00\x80\xE1\x94\x01\x00\x6F\x8E\x00\x80\xE3\x50\x00\x00\x73\x8E"
-     "\x00\x80"
-     "\x82\xE9\x3F\x00\x81\xE9\x3F\x00\x80\xF1\x3F\x00\xA3\x09\x3E\x00\x8F\x01"
-     "\x24\x00"
-     "\x13\x8E\x00\x80\x77\x8E\x00\x80\x7B\x8E\x00\x80\xA6\x09\x3E\x00\xA5\x09"
-     "\x3E\x00"
-     "\x7F\x8E\x00\x80\xAB\x69\x3E\x00\xAA\x65\x3E\x00\x83\x8E\x00\x80\x87\x8E"
-     "\x00\x80"
-     "\xAF\x5D\x3E\x00\xAE\x5D\x3E\x00\xAD\x61\x3E\x00\xAC\x69\x3E\x00\x9E\x61"
-     "\x38\x00"
-     "\x9F\x75\x38\x00\x9C\xC1\x34\x00\x9D\xB5\x39\x00\x9A\xA5\x35\x00\x9B\x75"
-     "\x34\x00"
-     "\x98\x79\x30\x00\x99\x71\x31\x00\x96\x21\x2D\x00\x97\x85\x31\x00\x94\x6D"
-     "\x2C\x00"
-     "\x95\x65\x2C\x00\x92\x79\x28\x00\x93\xA5\x2D\x00\x90\x51\x24\x00\x91\x79"
-     "\x28\x00"
-     "\xB1\x0D\x14\x00\xB0\x05\x14\x00\xB3\x01\x18\x00\xB2\x55\x14\x00\xB5\x79"
-     "\x18\x00"
-     "\xB4\xB5\x18\x00\x8B\x8E\x00\x80\x8F\x8E\x00\x80\x93\x8E\x00\x80\x97\x8E"
-     "\x00\x80"
-     "\x9B\x8E\x00\x80\x9F\x8E\x00\x80\xA2\xD1\x3C\x00\xA3\x25\x01\x00\xA0\x75"
-     "\x39\x00"
-     "\xA1\xBD\x3C\x00\xA1\xC9\x00\x00\xA3\x8E\x00\x80\xA3\x01\x04\x00\xA2\xE5"
-     "\x00\x00"
-     "\xA5\x1D\x04\x00\xA4\xF5\x04\x00\xA7\xF9\x08\x00\xA6\x01\x08\x00\xA9\x01"
-     "\x0C\x00"
-     "\xA8\x6D\x08\x00\xAB\xCD\x0C\x00\xAA\xCD\x0C\x00\xAD\xD1\x10\x00\xAC\x01"
-     "\x10\x00"
-     "\xAF\xD5\x11\x00\xAE\xD9\x10\x00\x82\x05\x25\x00\x83\x2D\x25\x00\xA7\x8E"
-     "\x00\x80"
-     "\xAB\x8E\x00\x80\x86\xC1\x2C\x00\x87\x11\x2D\x00\x84\x1D\x29\x00\x85\x15"
-     "\x29\x00"
-     "\x8A\x29\x2D\x00\x8B\x65\x2C\x00\xAF\x8E\x00\x80\xB3\x8E\x00\x80\x8E\x1D"
-     "\x30\x00"
-     "\x8F\xC1\x34\x00\x8C\xD9\x30\x00\x8D\x1D\x31\x00\x92\x61\x35\x00\x93\xCD"
-     "\x35\x00"
-     "\xB7\x8E\x00\x80\xBB\x8E\x00\x80\x96\x61\x39\x00\x97\x99\x38\x00\x94\xA1"
-     "\x38\x00"
-     "\x95\x7D\x39\x00\x9A\x61\x3D\x00\x9B\x05\x3D\x00\xBF\x8E\x00\x80\xC3\x8E"
-     "\x00\x80"
-     "\xC7\x8E\x00\x80\xCB\x8E\x00\x80\x9C\xE9\x00\x00\xCF\x8E\x00\x80\xD3\x8E"
-     "\x00\x80"
-     "\xD7\x8E\x00\x80\xDB\x8E\x00\x80\xDF\x8E\x00\x80\xA3\x8D\x00\x80\xE3\x8E"
-     "\x00\x80"
-     "\xE7\x8E\x00\x80\xEB\x8E\x00\x80\xA8\x91\x3E\x00\xA9\x91\x3E\x00\xAA\xED"
-     "\x3E\x00"
-     "\xAB\xE1\x3E\x00\xAC\xE1\x3E\x00\xAD\xE9\x3E\x00\xAE\xD1\x3E\x00\xAF\xD1"
-     "\x3E\x00"
-     "\xB0\x51\x3E\x00\xB1\x51\x3E\x00\xB2\x51\x3E\x00\xB3\x51\x3E\x00\xB4\x79"
-     "\x3E\x00"
-     "\xB5\x79\x3E\x00\xB6\x6D\x3E\x00\xB7\x65\x3E\x00\xB8\x21\x3E\x00\xB9\x21"
-     "\x3E\x00"
-     "\xBA\x39\x3E\x00\xBB\x39\x3E\x00\xBC\x29\x3E\x00\xBD\x15\x3E\x00\xBE\x11"
-     "\x3E\x00"
-     "\xBF\x0D\x3E\x00\x80\x99\x03\x00\x81\x99\x03\x00\x82\x05\x00\x00\xEF\x8E"
-     "\x00\x80"
-     "\xBE\x54\x03\x00\xE1\xB0\x3D\x00\xF7\x8E\x00\x80\xE3\x40\x3E\x00\x84\x38"
-     "\x02\x00"
-     "\xFB\x8E\x00\x80\xFF\x8E\x00\x80\xEF\xF4\x3F\x00\x03\x8F\x00\x80\x07\x8F"
-     "\x00\x80"
-     "\x86\x98\x04\x00\x87\x1C\x03\x00\xB3\x05\x3D\x00\x84\x08\x04\x00\x0B\x8F"
-     "\x00\x80"
-     "\x0F\x8F\x00\x80\x13\x8F\x00\x80\xB6\x09\x3D\x00\xB5\x09\x3D\x00\x17\x8F"
-     "\x00\x80"
-     "\xBB\xF5\x3D\x00\xBA\xF5\x3D\x00\x1B\x8F\x00\x80\x1F\x8F\x00\x80\xBF\xDD"
-     "\x3D\x00"
-     "\xBE\xDD\x3D\x00\xBD\xE5\x3D\x00\xBC\xE5\x3D\x00\x23\x8F\x00\x80\x27\x8F"
-     "\x00\x80"
-     "\xA3\xCD\x3D\x00\xBE\xC4\x04\x00\xA5\xC1\x3D\x00\x2B\x8F\x00\x80\x2F\x8F"
-     "\x00\x80"
-     "\xA6\xC1\x3D\x00\x33\x8F\x00\x80\x37\x8F\x00\x80\xAB\x3D\x3D\x00\xAA\x3D"
-     "\x3D\x00"
-     "\xAD\x2D\x3D\x00\xAC\x2D\x3D\x00\xAF\x15\x3D\x00\xAE\x15\x3D\x00\xB6\x69"
-     "\x02\x00"
-     "\x3B\x8F\x00\x80\x3F\x8F\x00\x80\xB5\x69\x02\x00\x43\x8F\x00\x80\xB3\x49"
-     "\x02\x00"
-     "\x47\x8F\x00\x80\x4B\x8F\x00\x80\xBE\xA9\x03\x00\xBF\xA9\x03\x00\xBC\xC1"
-     "\x03\x00"
-     "\xBD\xC1\x03\x00\xBA\xF9\x03\x00\xBB\xF9\x03\x00\x4F\x8F\x00\x80\x53\x8F"
-     "\x00\x80"
-     "\xA8\x2D\x03\x00\xA9\x9D\x03\x00\xAA\x95\x03\x00\xAB\xAD\x03\x00\xAC\xB5"
-     "\x03\x00"
-     "\xAD\xBD\x03\x00\xAE\xB5\x03\x00\xAF\xD9\x03\x00\x80\x0D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x1D\x00\x00\x57\x8F\x00\x80\x5B\x8F\x00\x80\x5F\x8F\x00\x80\x87\xB4"
-     "\x05\x00"
-     "\x86\x9C\x04\x00\xB8\x31\x02\x00\xB9\x31\x02\x00\xBA\x35\x02\x00\xBB\xCD"
-     "\x02\x00"
-     "\xBC\xD5\x02\x00\xBD\xDD\x02\x00\xBE\xD5\x02\x00\xBF\xC9\x02\x00\xB0\x69"
-     "\x02\x00"
-     "\xB1\x69\x02\x00\xB2\x41\x02\x00\xB3\x41\x02\x00\xB4\x39\x02\x00\xB5\x39"
-     "\x02\x00"
-     "\xB6\x11\x02\x00\xB7\x11\x02\x00\x67\x8F\x00\x80\xE3\x34\x3E\x00\x6B\x8F"
-     "\x00\x80"
-     "\xE1\xAC\x3E\x00\x6F\x8F\x00\x80\xEF\x7C\x03\x00\x73\x8F\x00\x80\x77\x8F"
-     "\x00\x80"
-     "\x7B\x8F\x00\x80\xE3\x94\x03\x00\x7F\x8F\x00\x80\xE1\x7C\x3E\x00\x83\x8F"
-     "\x00\x80"
-     "\xEF\xE8\x3E\x00\x87\x8F\x00\x80\x8B\x8F\x00\x80\x8F\x8F\x00\x80\x93\x8F"
-     "\x00\x80"
-     "\xA3\x55\x03\x00\x97\x8F\x00\x80\xA5\x75\x03\x00\x9B\x8F\x00\x80\x9F\x8F"
-     "\x00\x80"
-     "\xA6\x75\x03\x00\xA3\x8F\x00\x80\xA7\x8F\x00\x80\xAB\xE5\x02\x00\xAA\xE5"
-     "\x02\x00"
-     "\xAD\xDD\x02\x00\xAC\xDD\x02\x00\xAF\xB5\x02\x00\xAE\xB5\x02\x00\xA8\x19"
-     "\x06\x00"
-     "\xA9\x21\x06\x00\xAA\x21\x06\x00\xAB\x3D\x06\x00\xAC\x25\x06\x00\xAD\x5D"
-     "\x06\x00"
-     "\xAE\x55\x06\x00\xAF\x4D\x06\x00\x63\x8F\x00\x80\xAB\x8F\x00\x80\xAF\x8F"
-     "\x00\x80"
-     "\xB3\x8F\x00\x80\xB7\x8F\x00\x80\xBB\x8F\x00\x80\xBF\x8F\x00\x80\xC3\x8F"
-     "\x00\x80"
-     "\xB8\xE5\x06\x00\xB9\x81\x06\x00\xBA\x81\x06\x00\xBB\x9D\x06\x00\xBC\x89"
-     "\x06\x00"
-     "\xBD\x89\x06\x00\xBE\xA1\x06\x00\xBF\xA1\x06\x00\xB0\x3D\x06\x00\xB1\x0D"
-     "\x06\x00"
-     "\xB2\x05\x06\x00\xB3\xED\x06\x00\xB4\xF5\x06\x00\xB5\xE1\x06\x00\xB6\xE1"
-     "\x06\x00"
-     "\xB7\xDD\x06\x00\xB3\xA9\x06\x00\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xC7\x8F\x00\x80\xB6\xDD\x06\x00\xB5\xAD\x06\x00\xCB\x8F\x00\x80\xBB\xC9"
-     "\x06\x00"
-     "\xBA\xF9\x06\x00\xCF\x8F\x00\x80\x84\xE0\x03\x00\xBF\x25\x06\x00\xBE\x31"
-     "\x06\x00"
-     "\xBD\x39\x06\x00\xBC\xD1\x06\x00\xBE\x88\x03\x00\xA3\xED\x06\x00\xF3\x8E"
-     "\x00\x80"
-     "\xD3\x8F\x00\x80\xA6\x99\x06\x00\xD7\x8F\x00\x80\xDB\x8F\x00\x80\xA5\xE9"
-     "\x06\x00"
-     "\xAA\xBD\x06\x00\xAB\x8D\x06\x00\x86\x48\x00\x00\x87\x6C\x00\x00\xAE\x75"
-     "\x06\x00"
-     "\xAF\x61\x06\x00\xAC\x95\x06\x00\xAD\x7D\x06\x00\xA8\x81\x06\x00\xA9\x8D"
-     "\x06\x00"
-     "\xAA\x99\x06\x00\xAB\x95\x06\x00\xAC\x8D\x06\x00\xAD\xB5\x06\x00\xAE\xB1"
-     "\x06\x00"
-     "\xAF\xAD\x06\x00\xDF\x8F\x00\x80\xE3\x8F\x00\x80\xE7\x8F\x00\x80\xEB\x8F"
-     "\x00\x80"
-     "\xEF\x8F\x00\x80\xF3\x8F\x00\x80\xF7\x8F\x00\x80\xFB\x8F\x00\x80\xB8\xA5"
-     "\x06\x00"
-     "\xB9\x61\x01\x00\xBA\x61\x01\x00\xBB\x61\x01\x00\xBC\x61\x01\x00\xBD\x61"
-     "\x01\x00"
-     "\xBE\x61\x01\x00\xBF\x61\x01\x00\xB0\xD9\x06\x00\xB1\xD9\x06\x00\xB2\xA9"
-     "\x06\x00"
-     "\xB3\xA9\x06\x00\xB4\xBD\x06\x00\xB5\xA1\x06\x00\xB6\xA1\x06\x00\xB7\x9D"
-     "\x06\x00"
-     "\xB3\x11\x06\x00\xFF\x8F\x00\x80\x03\x90\x00\x80\x07\x90\x00\x80\x0B\x90"
-     "\x00\x80"
-     "\xB6\x35\x06\x00\xB5\x05\x06\x00\x0F\x90\x00\x80\xBB\x1D\x06\x00\xBA\x1D"
-     "\x06\x00"
-     "\x13\x90\x00\x80\x17\x90\x00\x80\xBF\x65\x06\x00\xBE\x79\x06\x00\xBD\x7D"
-     "\x06\x00"
-     "\xBC\x7D\x06\x00\x1B\x90\x00\x80\xA3\x55\x06\x00\x1F\x90\x00\x80\x23\x90"
-     "\x00\x80"
-     "\xA6\x71\x06\x00\x27\x90\x00\x80\x2B\x90\x00\x80\xA5\x41\x06\x00\xAA\x59"
-     "\x06\x00"
-     "\xAB\x59\x06\x00\x2F\x90\x00\x80\x33\x90\x00\x80\xAE\x3D\x06\x00\xAF\x21"
-     "\x06\x00"
-     "\xAC\x39\x06\x00\xAD\x39\x06\x00\xA8\xD5\x02\x00\xA9\xDD\x02\x00\xAA\x29"
-     "\x03\x00"
-     "\xAB\x29\x03\x00\xAC\x39\x03\x00\xAD\x39\x03\x00\xAE\x29\x03\x00\xAF\x29"
-     "\x03\x00"
-     "\x37\x90\x00\x80\x3B\x90\x00\x80\x3F\x90\x00\x80\x47\x90\x00\x80\x4B\x90"
-     "\x00\x80"
-     "\x4F\x90\x00\x80\xBE\xB8\x03\x00\x53\x90\x00\x80\xB8\xED\x03\x00\xB9\x85"
-     "\x03\x00"
-     "\xBA\x81\x03\x00\xBB\x81\x03\x00\xBC\x85\x03\x00\xBD\x8D\x03\x00\xBE\xB1"
-     "\x03\x00"
-     "\xBF\xB1\x03\x00\xB0\x59\x03\x00\xB1\x59\x03\x00\xB2\xED\x03\x00\xB3\xE5"
-     "\x03\x00"
-     "\xB4\xFD\x03\x00\xB5\xE5\x03\x00\xB6\xE5\x03\x00\xB7\xD5\x03\x00\x80\xA1"
-     "\x00\x00"
-     "\x81\xA1\x00\x00\x82\xA1\x00\x00\xBE\x80\x0C\x00\x57\x90\x00\x80\x84\x98"
-     "\x02\x00"
-     "\x5B\x90\x00\x80\x5F\x90\x00\x80\x86\x00\x0C\x00\x87\xF4\x03\x00\x63\x90"
-     "\x00\x80"
-     "\x67\x90\x00\x80\x6B\x90\x00\x80\x6F\x90\x00\x80\x73\x90\x00\x80\x84\xB0"
-     "\x03\x00"
-     "\xB3\xE1\x03\x00\x77\x90\x00\x80\x7B\x90\x00\x80\x7F\x90\x00\x80\x83\x90"
-     "\x00\x80"
-     "\xB6\xF9\x03\x00\xB5\xF1\x03\x00\x87\x90\x00\x80\xBB\xDD\x03\x00\xBA\xDD"
-     "\x03\x00"
-     "\x8B\x90\x00\x80\x8F\x90\x00\x80\xBF\x61\x03\x00\xBE\x79\x03\x00\xBD\x71"
-     "\x03\x00"
-     "\xBC\x79\x03\x00\x93\x90\x00\x80\x97\x90\x00\x80\x9B\x90\x00\x80\xA3\x2D"
-     "\x02\x00"
-     "\x9F\x90\x00\x80\xA5\x3D\x02\x00\xA6\x35\x02\x00\xA3\x90\x00\x80\xA7\x90"
-     "\x00\x80"
-     "\xAB\x90\x00\x80\xAA\x11\x02\x00\xAB\x11\x02\x00\xAC\xB5\x02\x00\xAD\xBD"
-     "\x02\x00"
-     "\xAE\xB5\x02\x00\xAF\xAD\x02\x00\xE3\xC4\x03\x00\xE3\x10\x07\x00\xE1\xB8"
-     "\x01\x00"
-     "\xE1\x6C\x07\x00\x80\x71\x00\x00\x81\x71\x00\x00\x82\x05\x00\x00\xAF\x90"
-     "\x00\x80"
-     "\x86\xC0\x0C\x00\x87\x54\x0D\x00\xB7\x90\x00\x80\xBB\x90\x00\x80\xEF\xB0"
-     "\x03\x00"
-     "\xEF\x00\x07\x00\xBF\x90\x00\x80\xC3\x90\x00\x80\xC7\x90\x00\x80\xCB\x90"
-     "\x00\x80"
-     "\xCF\x90\x00\x80\xD3\x90\x00\x80\xD7\x90\x00\x80\xDB\x90\x00\x80\xDF\x90"
-     "\x00\x80"
-     "\xEF\xA4\x01\x00\x84\xA0\x0D\x00\xE1\x80\x06\x00\xE3\x90\x00\x80\xE3\x10"
-     "\x01\x00"
-     "\xE7\x90\x00\x80\xEB\x90\x00\x80\xEF\x90\x00\x80\xF3\x90\x00\x80\xB3\xF5"
-     "\x01\x00"
-     "\xF7\x90\x00\x80\xFB\x90\x00\x80\xFF\x90\x00\x80\x03\x91\x00\x80\xB6\x4D"
-     "\x01\x00"
-     "\xB5\x49\x01\x00\x07\x91\x00\x80\xBB\x51\x01\x00\xBA\x49\x01\x00\x0B\x91"
-     "\x00\x80"
-     "\x0F\x91\x00\x80\xBF\x39\x01\x00\xBE\x31\x01\x00\xBD\x41\x01\x00\xBC\x49"
-     "\x01\x00"
-     "\xA8\x2D\x0E\x00\xA9\x35\x0E\x00\xAA\x3D\x0E\x00\xAB\x31\x0E\x00\xAC\x81"
-     "\x0E\x00"
-     "\xAD\x8D\x0E\x00\xAE\x85\x0E\x00\xAF\xB5\x0E\x00\xB3\x90\x00\x80\x13\x91"
-     "\x00\x80"
-     "\x17\x91\x00\x80\x1B\x91\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x1F\x91\x00\x80\xB8\x9D\x0E\x00\xB9\xAD\x0E\x00\xBA\xA5\x0E\x00\xBB\x4D"
-     "\x0F\x00"
-     "\xBC\x55\x0F\x00\xBD\x5D\x0F\x00\xBE\x49\x0F\x00\xBF\x41\x0F\x00\xB0\xCD"
-     "\x0E\x00"
-     "\xB1\xD5\x0E\x00\xB2\xDD\x0E\x00\xB3\xD5\x0E\x00\xB4\xB5\x0E\x00\xB5\xBD"
-     "\x0E\x00"
-     "\xB6\xB5\x0E\x00\xB7\xAD\x0E\x00\xA3\xB5\x0E\x00\xBE\x88\x03\x00\x23\x91"
-     "\x00\x80"
-     "\x27\x91\x00\x80\x2B\x91\x00\x80\xA6\x0D\x0E\x00\xA5\x09\x0E\x00\x2F\x91"
-     "\x00\x80"
-     "\xAB\x11\x0E\x00\xAA\x09\x0E\x00\x86\x08\x00\x00\x87\x6C\x03\x00\xAF\x79"
-     "\x0E\x00"
-     "\xAE\x71\x0E\x00\xAD\x01\x0E\x00\xAC\x09\x0E\x00\x33\x91\x00\x80\x37\x91"
-     "\x00\x80"
-     "\x3B\x91\x00\x80\xB3\xB5\x0F\x00\x3F\x91\x00\x80\xB5\x55\x0F\x00\xB6\x5D"
-     "\x0F\x00"
-     "\x43\x90\x00\x80\x43\x91\x00\x80\x47\x91\x00\x80\xBA\x79\x0F\x00\xBB\x79"
-     "\x0F\x00"
-     "\xBC\x69\x0F\x00\xBD\x5D\x0F\x00\xBE\x49\x0F\x00\xBF\x49\x0F\x00\xA8\x69"
-     "\x0E\x00"
-     "\xA9\x69\x0E\x00\xAA\x71\x0E\x00\xAB\x71\x0E\x00\xAC\x91\x0E\x00\xAD\x91"
-     "\x0E\x00"
-     "\xAE\x91\x0E\x00\xAF\x91\x0E\x00\x4B\x91\x00\x80\x4F\x91\x00\x80\x53\x91"
-     "\x00\x80"
-     "\x57\x91\x00\x80\x5B\x91\x00\x80\x5F\x91\x00\x80\x63\x91\x00\x80\x67\x91"
-     "\x00\x80"
-     "\xB8\x85\x0E\x00\xB9\x8D\x0E\x00\xBA\x85\x0E\x00\xBB\x9D\x0E\x00\xBC\x8D"
-     "\x0E\x00"
-     "\xBD\xBD\x0E\x00\xBE\xB5\x0E\x00\xBF\x79\x01\x00\xB0\xF1\x0E\x00\xB1\xF1"
-     "\x0E\x00"
-     "\xB2\xF1\x0E\x00\xB3\xC5\x0E\x00\xB4\xC1\x0E\x00\xB5\xC1\x0E\x00\xB6\xC1"
-     "\x0E\x00"
-     "\xB7\xC1\x0E\x00\xA3\xF9\x0E\x00\x6B\x91\x00\x80\x6F\x91\x00\x80\x73\x91"
-     "\x00\x80"
-     "\x77\x91\x00\x80\xA6\x11\x0E\x00\xA5\x19\x0E\x00\x7B\x91\x00\x80\xAB\x35"
-     "\x0E\x00"
-     "\xAA\x35\x0E\x00\x7F\x91\x00\x80\x83\x91\x00\x80\xAF\x05\x0E\x00\xAE\x05"
-     "\x0E\x00"
-     "\xAD\x11\x0E\x00\xAC\x25\x0E\x00\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\x87\x91\x00\x80\x8B\x91\x00\x80\x8F\x91\x00\x80\x84\x94\x01\x00\xBE\x94"
-     "\x01\x00"
-     "\x86\x40\x07\x00\x87\xF4\x00\x00\x97\x91\x00\x80\x9B\x91\x00\x80\x9F\x91"
-     "\x00\x80"
-     "\xA3\x91\x00\x80\xA7\x91\x00\x80\xAB\x91\x00\x80\xA8\x8D\x02\x00\xA9\x95"
-     "\x02\x00"
-     "\xAA\x95\x02\x00\xAB\xCD\x02\x00\xAC\xD5\x02\x00\xAD\xDD\x02\x00\xAE\xC9"
-     "\x02\x00"
-     "\xAF\xFD\x02\x00\xAF\x91\x00\x80\xB3\x91\x00\x80\xB7\x91\x00\x80\xBB\x91"
-     "\x00\x80"
-     "\xBF\x00\x14\x00\xBF\x91\x00\x80\xC3\x91\x00\x80\xC7\x91\x00\x80\xB8\x7D"
-     "\x03\x00"
-     "\xB9\xC1\x03\x00\xBA\xC1\x03\x00\xBB\xC1\x03\x00\xBC\xC1\x03\x00\xBD\xC9"
-     "\x03\x00"
-     "\xBE\xF1\x03\x00\xBF\xF1\x03\x00\xB0\x85\x02\x00\xB1\x45\x03\x00\xB2\x4D"
-     "\x03\x00"
-     "\xB3\x45\x03\x00\xB4\x5D\x03\x00\xB5\x45\x03\x00\xB6\x4D\x03\x00\xB7\x45"
-     "\x03\x00"
-     "\xB3\x1D\x02\x00\xCB\x91\x00\x80\xCF\x91\x00\x80\xD3\x91\x00\x80\xD7\x91"
-     "\x00\x80"
-     "\xB6\x5D\x02\x00\xB5\x5D\x02\x00\xDB\x91\x00\x80\xBB\x81\x03\x00\xBA\x41"
-     "\x02\x00"
-     "\xDF\x91\x00\x80\xE3\x91\x00\x80\xBF\x81\x03\x00\xBE\x99\x03\x00\xBD\x91"
-     "\x03\x00"
-     "\xBC\x99\x03\x00\xE7\x91\x00\x80\xA3\x59\x02\x00\xEB\x91\x00\x80\xEF\x91"
-     "\x00\x80"
-     "\xA6\x19\x02\x00\xF3\x91\x00\x80\xF7\x91\x00\x80\xA5\x19\x02\x00\xAA\x05"
-     "\x02\x00"
-     "\xAB\xC5\x03\x00\xFB\x91\x00\x80\xFF\x91\x00\x80\xAE\xDD\x03\x00\xAF\xC5"
-     "\x03\x00"
-     "\xAC\xDD\x03\x00\xAD\xD5\x03\x00\x07\x92\x00\x80\xE3\xCC\x00\x00\x84\x04"
-     "\x02\x00"
-     "\xE1\xBC\x01\x00\x80\xC9\x01\x00\x81\xFD\x01\x00\x82\xF5\x01\x00\xBE\x10"
-     "\x05\x00"
-     "\x0B\x92\x00\x80\xBE\x28\x04\x00\x0F\x92\x00\x80\x13\x92\x00\x80\x17\x92"
-     "\x00\x80"
-     "\xEF\x10\x00\x00\x1B\x92\x00\x80\x1F\x92\x00\x80\x86\xE0\x04\x00\x87\xF4"
-     "\x02\x00"
-     "\x23\x92\x00\x80\x27\x92\x00\x80\xE3\xFC\x0F\x00\x2B\x92\x00\x80\xE1\xE0"
-     "\x0F\x00"
-     "\x2F\x92\x00\x80\xEF\x14\x0F\x00\x33\x92\x00\x80\x37\x92\x00\x80\x3B\x92"
-     "\x00\x80"
-     "\x3F\x92\x00\x80\x43\x92\x00\x80\x47\x92\x00\x80\x4B\x92\x00\x80\x4F\x92"
-     "\x00\x80"
-     "\x53\x92\x00\x80\x57\x92\x00\x80\x5B\x92\x00\x80\x5F\x92\x00\x80\x63\x92"
-     "\x00\x80"
-     "\x67\x92\x00\x80\xEF\xE0\x01\x00\x85\x04\x12\x00\xE1\xDC\x0E\x00\x6B\x92"
-     "\x00\x80"
-     "\xE3\x1C\x0E\x00\x80\x29\x00\x00\x81\x1D\x00\x00\x82\x05\x00\x00\x6F\x92"
-     "\x00\x80"
-     "\xB3\x31\x02\x00\x77\x92\x00\x80\x84\xCC\x05\x00\x7B\x92\x00\x80\x7F\x92"
-     "\x00\x80"
-     "\xB6\x29\x02\x00\xB5\x21\x02\x00\x83\x92\x00\x80\xBB\xCD\x01\x00\xBA\xCD"
-     "\x01\x00"
-     "\x87\x92\x00\x80\x8B\x92\x00\x80\xBF\x75\x01\x00\xBE\xC9\x01\x00\xBD\xC1"
-     "\x01\x00"
-     "\xBC\xC9\x01\x00\xA8\xE9\x05\x00\xA9\xE9\x05\x00\xAA\xF9\x05\x00\xAB\xF9"
-     "\x05\x00"
-     "\xAC\xE9\x05\x00\xAD\xE9\x05\x00\xAE\x39\x06\x00\xAF\x39\x06\x00\x03\x92"
-     "\x00\x80"
-     "\x73\x92\x00\x80\x86\x88\x00\x00\x87\x00\x03\x00\x8F\x92\x00\x80\x93\x92"
-     "\x00\x80"
-     "\x97\x92\x00\x80\x9B\x92\x00\x80\xB8\xD1\x06\x00\xB9\xD9\x06\x00\xBA\xE1"
-     "\x06\x00"
-     "\xBB\xE1\x06\x00\xBC\x91\x06\x00\xBD\x9D\x06\x00\xBE\x95\x06\x00\xBF\x89"
-     "\x06\x00"
-     "\xB0\x49\x06\x00\xB1\x49\x06\x00\xB2\x5D\x06\x00\xB3\x55\x06\x00\xB4\x4D"
-     "\x06\x00"
-     "\xB5\xF1\x06\x00\xB6\xF1\x06\x00\xB7\xF1\x06\x00\xA3\x71\x05\x00\x9F\x92"
-     "\x00\x80"
-     "\xA3\x92\x00\x80\xA7\x92\x00\x80\xAB\x92\x00\x80\xA6\x69\x05\x00\xA5\x61"
-     "\x05\x00"
-     "\xAF\x92\x00\x80\xAB\x8D\x06\x00\xAA\x8D\x06\x00\xB3\x92\x00\x80\xB7\x92"
-     "\x00\x80"
-     "\xAF\x35\x06\x00\xAE\x89\x06\x00\xAD\x81\x06\x00\xAC\x89\x06\x00\xBB\x92"
-     "\x00\x80"
-     "\xBF\x92\x00\x80\xB3\xE1\x07\x00\xC3\x92\x00\x80\xB5\xE1\x07\x00\xC7\x92"
-     "\x00\x80"
-     "\xCB\x92\x00\x80\xB6\xE5\x07\x00\x93\x91\x00\x80\xCF\x92\x00\x80\xBB\xBD"
-     "\x07\x00"
-     "\xBA\xA1\x07\x00\xBD\x95\x07\x00\xBC\xA9\x07\x00\xBF\x95\x07\x00\xBE\x95"
-     "\x07\x00"
-     "\xA8\x01\x06\x00\xA9\x25\x06\x00\xAA\x21\x06\x00\xAB\x21\x06\x00\xAC\x21"
-     "\x06\x00"
-     "\xAD\x2D\x06\x00\xAE\x25\x06\x00\xAF\x55\x06\x00\xD3\x92\x00\x80\x82\x1D"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x80\x1D\x00\x00\xD7\x92\x00\x80\xDB\x92\x00\x80\xDF\x92"
-     "\x00\x80"
-     "\xBE\x30\x01\x00\xB8\x39\x06\x00\xB9\x39\x06\x00\xBA\xC9\x06\x00\xBB\xC9"
-     "\x06\x00"
-     "\xBC\xD9\x06\x00\xBD\xD9\x06\x00\xBE\xC9\x06\x00\xBF\xC9\x06\x00\xB0\x2D"
-     "\x06\x00"
-     "\xB1\x31\x06\x00\xB2\x35\x06\x00\xB3\x09\x06\x00\xB4\x19\x06\x00\xB5\x19"
-     "\x06\x00"
-     "\xB6\x09\x06\x00\xB7\x09\x06\x00\xA3\xA9\x06\x00\x84\x8C\x02\x00\x86\x28"
-     "\x1F\x00"
-     "\x87\x44\x01\x00\xE7\x92\x00\x80\xA6\xAD\x06\x00\xA5\xA9\x06\x00\xEB\x92"
-     "\x00\x80"
-     "\xAB\xF5\x06\x00\xAA\xE9\x06\x00\xEF\x92\x00\x80\xF3\x92\x00\x80\xAF\xDD"
-     "\x06\x00"
-     "\xAE\xDD\x06\x00\xAD\xDD\x06\x00\xAC\xE1\x06\x00\xF7\x92\x00\x80\xB3\x15"
-     "\x06\x00"
-     "\xFB\x92\x00\x80\xFF\x92\x00\x80\xB6\x3D\x06\x00\x03\x93\x00\x80\x07\x93"
-     "\x00\x80"
-     "\xB5\x35\x06\x00\xBA\xD9\x01\x00\xBB\xD9\x01\x00\x0B\x93\x00\x80\x0F\x93"
-     "\x00\x80"
-     "\xBE\x7D\x01\x00\xBF\x65\x01\x00\xBC\x7D\x01\x00\xBD\x75\x01\x00\xA8\xC5"
-     "\x05\x00"
-     "\xA9\xC9\x05\x00\xAA\xD9\x05\x00\xAB\xD1\x05\x00\xAC\xF9\x05\x00\xAD\xF9"
-     "\x05\x00"
-     "\xAE\x29\x02\x00\xAF\x29\x02\x00\x13\x93\x00\x80\x17\x93\x00\x80\x1B\x93"
-     "\x00\x80"
-     "\x1F\x93\x00\x80\x8C\x00\x00\x00\x23\x93\x00\x80\x27\x93\x00\x80\x2B\x93"
-     "\x00\x80"
-     "\xB8\xED\x02\x00\xB9\x85\x02\x00\xBA\x8D\x02\x00\xBB\x81\x02\x00\xBC\x85"
-     "\x02\x00"
-     "\xBD\x8D\x02\x00\xBE\xBD\x02\x00\xBF\x7D\x03\x00\xB0\x59\x02\x00\xB1\x59"
-     "\x02\x00"
-     "\xB2\xED\x02\x00\xB3\xE5\x02\x00\xB4\xFD\x02\x00\xB5\xE5\x02\x00\xB6\xE5"
-     "\x02\x00"
-     "\xB7\xD5\x02\x00\xA3\x51\x05\x00\x2F\x93\x00\x80\x33\x93\x00\x80\x37\x93"
-     "\x00\x80"
-     "\x3B\x93\x00\x80\xA6\x79\x05\x00\xA5\x71\x05\x00\x3F\x93\x00\x80\xAB\x9D"
-     "\x02\x00"
-     "\xAA\x9D\x02\x00\x43\x93\x00\x80\x47\x93\x00\x80\xAF\x21\x02\x00\xAE\x39"
-     "\x02\x00"
-     "\xAD\x31\x02\x00\xAC\x39\x02\x00\x82\x11\x00\x00\x4B\x93\x00\x80\x80\x65"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x4F\x93\x00\x80\xBE\x98\x03\x00\x57\x93\x00\x80\x5B\x93"
-     "\x00\x80"
-     "\x84\x24\x03\x00\x5F\x93\x00\x80\x87\x68\x03\x00\x86\x8C\x1C\x00\x63\x93"
-     "\x00\x80"
-     "\x67\x93\x00\x80\x6B\x93\x00\x80\x6F\x93\x00\x80\x73\x93\x00\x80\x77\x93"
-     "\x00\x80"
-     "\xB3\xA1\x02\x00\x84\xC0\x1C\x00\xB5\xD1\x02\x00\x7B\x93\x00\x80\x7F\x93"
-     "\x00\x80"
-     "\xB6\xD5\x02\x00\x83\x93\x00\x80\x87\x93\x00\x80\xBB\xC1\x02\x00\xBA\xF5"
-     "\x02\x00"
-     "\xBD\x11\x01\x00\xBC\xD9\x02\x00\xBF\x11\x01\x00\xBE\x19\x01\x00\x8B\x93"
-     "\x00\x80"
-     "\x8F\x93\x00\x80\x93\x93\x00\x80\x97\x93\x00\x80\x9B\x93\x00\x80\x9F\x93"
-     "\x00\x80"
-     "\xA3\x93\x00\x80\xEF\xB8\x06\x00\xA7\x93\x00\x80\xE1\x9C\x04\x00\xAB\x93"
-     "\x00\x80"
-     "\xE3\x74\x06\x00\xAF\x93\x00\x80\xB3\x93\x00\x80\xB7\x93\x00\x80\xBB\x93"
-     "\x00\x80"
-     "\x80\xF9\x00\x00\x81\xF9\x00\x00\x82\x05\x00\x00\xBF\x93\x00\x80\xBE\x58"
-     "\x1C\x00"
-     "\x84\x58\x1F\x00\xEF\x5C\x00\x00\xEF\x40\x06\x00\xE1\x90\x01\x00\xE1\xFC"
-     "\x06\x00"
-     "\xE3\x3C\x00\x00\xE3\x74\x06\x00\xC7\x93\x00\x80\xCB\x93\x00\x80\x86\x98"
-     "\x1C\x00"
-     "\x87\xF4\x1C\x00\xA3\x69\x02\x00\xBE\x0C\x1F\x00\xCF\x93\x00\x80\xD3\x93"
-     "\x00\x80"
-     "\xD7\x93\x00\x80\xA6\x1D\x02\x00\xA5\x19\x02\x00\xDB\x93\x00\x80\xAB\x09"
-     "\x02\x00"
-     "\xAA\x3D\x02\x00\xDF\x93\x00\x80\xE3\x93\x00\x80\xAF\xD9\x01\x00\xAE\xD1"
-     "\x01\x00"
-     "\xAD\xD9\x01\x00\xAC\x11\x02\x00\xA8\x91\x1D\x00\xA9\x99\x1D\x00\xAA\xA1"
-     "\x1D\x00"
-     "\xAB\xA1\x1D\x00\xAC\xD1\x1D\x00\xAD\xDD\x1D\x00\xAE\xD5\x1D\x00\xAF\xC9"
-     "\x1D\x00"
-     "\x53\x93\x00\x80\xC3\x93\x00\x80\xE7\x93\x00\x80\xEB\x93\x00\x80\xEF\x93"
-     "\x00\x80"
-     "\xF3\x93\x00\x80\xF7\x93\x00\x80\xFB\x93\x00\x80\xB8\x79\x1E\x00\xB9\x79"
-     "\x1E\x00"
-     "\xBA\xCD\x1E\x00\xBB\xC5\x1E\x00\xBC\xDD\x1E\x00\xBD\xC5\x1E\x00\xBE\xC5"
-     "\x1E\x00"
-     "\xBF\xF5\x1E\x00\xB0\xB9\x1D\x00\xB1\x8D\x1D\x00\xB2\x85\x1D\x00\xB3\x4D"
-     "\x1E\x00"
-     "\xB4\x55\x1E\x00\xB5\x5D\x1E\x00\xB6\x55\x1E\x00\xB7\x49\x1E\x00\xB8\xCD"
-     "\x1F\x00"
-     "\xB9\xD5\x1F\x00\xBA\xDD\x1F\x00\xBB\xD5\x1F\x00\xBC\xF1\x1F\x00\xBD\xFD"
-     "\x1F\x00"
-     "\xBE\xE9\x1F\x00\xBF\xE9\x1F\x00\xB0\xA5\x1F\x00\xB1\xB1\x1F\x00\xB2\xB1"
-     "\x1F\x00"
-     "\xB3\x8D\x1F\x00\xB4\x95\x1F\x00\xB5\xF5\x1F\x00\xB6\xFD\x1F\x00\xB7\xF5"
-     "\x1F\x00"
-     "\xA8\x19\x1E\x00\xA9\x19\x1E\x00\xAA\x2D\x1E\x00\xAB\x3D\x1E\x00\xAC\x25"
-     "\x1E\x00"
-     "\xAD\x2D\x1E\x00\xAE\x25\x1E\x00\xAF\xDD\x1F\x00\xFF\x93\x00\x80\x03\x94"
-     "\x00\x80"
-     "\x07\x94\x00\x80\x0B\x94\x00\x80\x0F\x94\x00\x80\xE3\x92\x00\x80\x13\x94"
-     "\x00\x80"
-     "\x17\x94\x00\x80\xB3\xE5\x1F\x00\x1B\x94\x00\x80\x1F\x94\x00\x80\x23\x94"
-     "\x00\x80"
-     "\x27\x94\x00\x80\xB6\xED\x1F\x00\xB5\xED\x1F\x00\x2B\x94\x00\x80\xBB\x35"
-     "\x1E\x00"
-     "\xBA\x21\x1E\x00\x2F\x94\x00\x80\x33\x94\x00\x80\xBF\x71\x1E\x00\xBE\x11"
-     "\x1E\x00"
-     "\xBD\x19\x1E\x00\xBC\x25\x1E\x00\x82\x69\x00\x00\xA3\xA1\x1F\x00\x80\x59"
-     "\x00\x00"
-     "\x81\x51\x00\x00\xA6\xA9\x1F\x00\x37\x94\x00\x80\x3B\x94\x00\x80\xA5\xA9"
-     "\x1F\x00"
-     "\xAA\x65\x1E\x00\xAB\x71\x1E\x00\x86\x00\x04\x00\x87\xEC\x01\x00\xAE\x55"
-     "\x1E\x00"
-     "\xAF\x35\x1E\x00\xAC\x61\x1E\x00\xAD\x5D\x1E\x00\xA8\x31\x1E\x00\xA9\x31"
-     "\x1E\x00"
-     "\xAA\x41\x1E\x00\xAB\x41\x1E\x00\xAC\x41\x1E\x00\xAD\x49\x1E\x00\xAE\x71"
-     "\x1E\x00"
-     "\xAF\x71\x1E\x00\x3F\x94\x00\x80\x43\x94\x00\x80\x47\x94\x00\x80\x4B\x94"
-     "\x00\x80"
-     "\x4F\x94\x00\x80\x53\x94\x00\x80\x57\x94\x00\x80\x5B\x94\x00\x80\xB8\x29"
-     "\x01\x00"
-     "\xB9\x29\x01\x00\xBA\x39\x01\x00\xBB\x35\x01\x00\xBC\x2D\x01\x00\xBD\xD1"
-     "\x00\x00"
-     "\xBE\xD1\x00\x00\xBF\xD1\x00\x00\xB0\xC9\x01\x00\xB1\xC9\x01\x00\xB2\xD9"
-     "\x01\x00"
-     "\xB3\xD9\x01\x00\xB4\xC9\x01\x00\xB5\xC9\x01\x00\xB6\x19\x01\x00\xB7\x19"
-     "\x01\x00"
-     "\xB3\xC9\x1D\x00\x5F\x94\x00\x80\x63\x94\x00\x80\x67\x94\x00\x80\x6B\x94"
-     "\x00\x80"
-     "\xB6\xC9\x1D\x00\xB5\xC9\x1D\x00\x6F\x94\x00\x80\xBB\x0D\x02\x00\xBA\x0D"
-     "\x02\x00"
-     "\x73\x94\x00\x80\x77\x94\x00\x80\xBF\x0D\x02\x00\xBE\x0D\x02\x00\xBD\x0D"
-     "\x02\x00"
-     "\xBC\x0D\x02\x00\x7B\x94\x00\x80\xA3\x8D\x1D\x00\x7F\x94\x00\x80\x83\x94"
-     "\x00\x80"
-     "\xA6\x8D\x1D\x00\x87\x94\x00\x80\x8B\x94\x00\x80\xA5\x8D\x1D\x00\xAA\x49"
-     "\x02\x00"
-     "\xAB\x49\x02\x00\x8F\x94\x00\x80\x93\x94\x00\x80\xAE\x49\x02\x00\xAF\x49"
-     "\x02\x00"
-     "\xAC\x49\x02\x00\xAD\x49\x02\x00\x80\x0D\x00\x00\x81\x11\x00\x00\x82\x11"
-     "\x00\x00"
-     "\x97\x94\x00\x80\xEF\xCC\x02\x00\x9B\x94\x00\x80\x9F\x94\x00\x80\x84\x90"
-     "\x02\x00"
-     "\xE3\x2C\x02\x00\xBE\x28\x03\x00\xE1\xD8\x01\x00\xA7\x94\x00\x80\x86\x10"
-     "\x04\x00"
-     "\x87\xD4\x03\x00\xAB\x94\x00\x80\xAF\x94\x00\x80\xB3\x61\x03\x00\xB3\x94"
-     "\x00\x80"
-     "\xB7\x94\x00\x80\xBB\x94\x00\x80\xBF\x94\x00\x80\xB6\x79\x03\x00\xB5\x71"
-     "\x03\x00"
-     "\xC3\x94\x00\x80\xBB\x5D\x03\x00\xBA\x5D\x03\x00\xC7\x94\x00\x80\xCB\x94"
-     "\x00\x80"
-     "\xBF\xE1\x00\x00\xBE\xF9\x00\x00\xBD\xF1\x00\x00\xBC\xF9\x00\x00\xA3\xA1"
-     "\x02\x00"
-     "\xCF\x94\x00\x80\xD3\x94\x00\x80\xD7\x94\x00\x80\xDB\x94\x00\x80\xA6\xB9"
-     "\x02\x00"
-     "\xA5\xB1\x02\x00\xDF\x94\x00\x80\xAB\x9D\x02\x00\xAA\x9D\x02\x00\xE3\x94"
-     "\x00\x80"
-     "\xE7\x94\x00\x80\xAF\x21\x01\x00\xAE\x39\x01\x00\xAD\x31\x01\x00\xAC\x39"
-     "\x01\x00"
-     "\xEB\x94\x00\x80\xEF\x94\x00\x80\xEF\x64\x1F\x00\xF3\x94\x00\x80\xF7\x94"
-     "\x00\x80"
-     "\xFB\x94\x00\x80\xFF\x94\x00\x80\x03\x95\x00\x80\x80\x0D\x00\x00\x81\x11"
-     "\x00\x00"
-     "\x82\x15\x00\x00\x07\x95\x00\x80\xE1\xE0\x1C\x00\x0B\x95\x00\x80\xE3\x88"
-     "\x1F\x00"
-     "\x0F\x95\x00\x80\x84\x80\x02\x00\xBE\x8C\x05\x00\x87\x48\x05\x00\x86\x2C"
-     "\x04\x00"
-     "\x17\x95\x00\x80\x1B\x95\x00\x80\xEF\xA4\x1E\x00\xEF\xF4\x1E\x00\xE1\x00"
-     "\x1E\x00"
-     "\xE1\x50\x1F\x00\xE3\x2C\x1E\x00\xE3\xB0\x1E\x00\x1F\x95\x00\x80\x23\x95"
-     "\x00\x80"
-     "\x27\x95\x00\x80\x2B\x95\x00\x80\x2F\x95\x00\x80\x33\x95\x00\x80\x84\x84"
-     "\x04\x00"
-     "\xB3\x71\x01\x00\x37\x95\x00\x80\xB5\x1D\x01\x00\xB6\x15\x01\x00\x3B\x95"
-     "\x00\x80"
-     "\x3F\x95\x00\x80\x43\x95\x00\x80\xBA\x01\x01\x00\xBB\x01\x01\x00\xBC\xF5"
-     "\x00\x00"
-     "\xBD\xFD\x00\x00\xBE\xF5\x00\x00\xBF\xED\x00\x00\xA8\xAD\x06\x00\xA9\xBD"
-     "\x06\x00"
-     "\xAA\xB5\x06\x00\xAB\xC9\x06\x00\xAC\xD9\x06\x00\xAD\xD9\x06\x00\xAE\xC9"
-     "\x06\x00"
-     "\xAF\xC1\x06\x00\x47\x95\x00\x80\x4B\x95\x00\x80\x4F\x95\x00\x80\x53\x95"
-     "\x00\x80"
-     "\x57\x95\x00\x80\x5B\x95\x00\x80\x5F\x95\x00\x80\x63\x95\x00\x80\xB8\x6D"
-     "\x07\x00"
-     "\xB9\x05\x07\x00\xBA\x0D\x07\x00\xBB\x01\x07\x00\xBC\x01\x07\x00\xBD\x01"
-     "\x07\x00"
-     "\xBE\x01\x07\x00\xBF\x01\x07\x00\xB0\x89\x06\x00\xB1\x89\x06\x00\xB2\x6D"
-     "\x07\x00"
-     "\xB3\x65\x07\x00\xB4\x7D\x07\x00\xB5\x65\x07\x00\xB6\x65\x07\x00\xB7\x55"
-     "\x07\x00"
-     "\xA3\x94\x00\x80\xA3\x39\x06\x00\x67\x95\x00\x80\x13\x95\x00\x80\xA6\x5D"
-     "\x06\x00"
-     "\x6B\x95\x00\x80\x6F\x95\x00\x80\xA5\x55\x06\x00\xAA\x49\x06\x00\xAB\x49"
-     "\x06\x00"
-     "\x73\x95\x00\x80\x77\x95\x00\x80\xAE\xBD\x07\x00\xAF\xA5\x07\x00\xAC\xBD"
-     "\x07\x00"
-     "\xAD\xB5\x07\x00\x80\x6D\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x7B\x95"
-     "\x00\x80"
-     "\x7F\x95\x00\x80\xBE\x9C\x03\x00\x83\x95\x00\x80\x87\x95\x00\x80\x86\x40"
-     "\x00\x00"
-     "\x87\x60\x00\x00\x8B\x95\x00\x80\x8F\x95\x00\x80\x93\x95\x00\x80\x97\x95"
-     "\x00\x80"
-     "\x9B\x95\x00\x80\x9F\x95\x00\x80\xA8\x91\x06\x00\xA9\x91\x06\x00\xAA\xB9"
-     "\x06\x00"
-     "\xAB\xB9\x06\x00\xAC\xA9\x06\x00\xAD\xA9\x06\x00\xAE\xD9\x06\x00\xAF\xD9"
-     "\x06\x00"
-     "\xA3\x95\x00\x80\xA7\x95\x00\x80\xAB\x95\x00\x80\xAF\x95\x00\x80\xB3\x95"
-     "\x00\x80"
-     "\xB7\x95\x00\x80\xBB\x95\x00\x80\xBF\x95\x00\x80\xB8\x71\x01\x00\xB9\x71"
-     "\x01\x00"
-     "\xBA\x71\x01\x00\xBB\x71\x01\x00\xBC\xD9\x01\x00\xBD\xC1\x01\x00\xBE\xC1"
-     "\x01\x00"
-     "\xBF\xF5\x01\x00\xB0\xB1\x06\x00\xB1\xB9\x06\x00\xB2\x89\x06\x00\xB3\x89"
-     "\x06\x00"
-     "\xB4\x51\x01\x00\xB5\x51\x01\x00\xB6\x51\x01\x00\xB7\x51\x01\x00\xB3\x31"
-     "\x06\x00"
-     "\xC3\x95\x00\x80\xC7\x95\x00\x80\xCB\x95\x00\x80\xCF\x95\x00\x80\xB6\x29"
-     "\x06\x00"
-     "\xB5\x21\x06\x00\xD3\x95\x00\x80\xBB\x7D\x06\x00\xBA\x75\x06\x00\xD7\x95"
-     "\x00\x80"
-     "\xDB\x95\x00\x80\xBF\x95\x01\x00\xBE\x95\x01\x00\xBD\x5D\x06\x00\xBC\x5D"
-     "\x06\x00"
-     "\xDF\x95\x00\x80\xA3\x75\x06\x00\xE3\x95\x00\x80\xE7\x95\x00\x80\xA6\x6D"
-     "\x06\x00"
-     "\xEB\x95\x00\x80\xEF\x95\x00\x80\xA5\x65\x06\x00\xAA\x31\x06\x00\xAB\x39"
-     "\x06\x00"
-     "\x84\xAC\x01\x00\xBE\xA0\x01\x00\xAE\xD1\x01\x00\xAF\xD1\x01\x00\xAC\x19"
-     "\x06\x00"
-     "\xAD\x19\x06\x00\xA8\xDD\x02\x00\xA9\xED\x02\x00\xAA\xE5\x02\x00\xAB\xFD"
-     "\x02\x00"
-     "\xAC\xE5\x02\x00\xAD\xED\x02\x00\xAE\xE5\x02\x00\xAF\x3D\x03\x00\xF7\x95"
-     "\x00\x80"
-     "\xFB\x95\x00\x80\xFF\x95\x00\x80\xBE\x64\x0C\x00\x03\x96\x00\x80\x07\x96"
-     "\x00\x80"
-     "\x0B\x96\x00\x80\x0F\x96\x00\x80\xB8\xC9\x03\x00\xB9\xC9\x03\x00\xBA\xD9"
-     "\x03\x00"
-     "\xBB\xD1\x03\x00\xBC\xF9\x03\x00\xBD\xF9\x03\x00\xBE\x99\x03\x00\xBF\x95"
-     "\x03\x00"
-     "\xB0\x45\x03\x00\xB1\x4D\x03\x00\xB2\x45\x03\x00\xB3\x5D\x03\x00\xB4\x45"
-     "\x03\x00"
-     "\xB5\x4D\x03\x00\xB6\x45\x03\x00\xB7\xF9\x03\x00\x81\x55\x03\x00\x80\x49"
-     "\x03\x00"
-     "\xB3\x65\x02\x00\x82\x55\x03\x00\xB5\x65\x02\x00\x13\x96\x00\x80\x17\x96"
-     "\x00\x80"
-     "\xB6\x65\x02\x00\x86\x00\x0C\x00\x87\xE4\x03\x00\xBB\x81\x03\x00\xBA\x89"
-     "\x03\x00"
-     "\xBD\x81\x03\x00\xBC\x99\x03\x00\xBF\x81\x03\x00\xBE\x89\x03\x00\xA3\x2D"
-     "\x02\x00"
-     "\x1B\x96\x00\x80\x1F\x96\x00\x80\x23\x96\x00\x80\x27\x96\x00\x80\xA6\x2D"
-     "\x02\x00"
-     "\xA5\x2D\x02\x00\x2B\x96\x00\x80\xAB\xC9\x03\x00\xAA\xC1\x03\x00\x2F\x96"
-     "\x00\x80"
-     "\x33\x96\x00\x80\xAF\xC9\x03\x00\xAE\xC1\x03\x00\xAD\xC9\x03\x00\xAC\xD1"
-     "\x03\x00"
-     "\xE3\xD8\x06\x00\xE1\xAC\x07\x00\xE1\x9C\x06\x00\xE3\x9C\x06\x00\x37\x96"
-     "\x00\x80"
-     "\x84\x58\x0D\x00\x3B\x96\x00\x80\x3F\x96\x00\x80\x43\x96\x00\x80\x47\x96"
-     "\x00\x80"
-     "\x4B\x96\x00\x80\x4F\x96\x00\x80\xEF\x1C\x01\x00\x53\x96\x00\x80\x57\x96"
-     "\x00\x80"
-     "\xEF\x40\x06\x00\x80\x79\x00\x00\x81\x15\x00\x00\x82\x11\x00\x00\x84\x00"
-     "\x0C\x00"
-     "\x5B\x96\x00\x80\xE3\xAC\x00\x00\x5F\x96\x00\x80\xE1\xA4\x01\x00\x67\x96"
-     "\x00\x80"
-     "\xEF\x70\x00\x00\x86\xC8\x0C\x00\x87\xA4\x0D\x00\x6B\x96\x00\x80\x6F\x96"
-     "\x00\x80"
-     "\x73\x96\x00\x80\x77\x96\x00\x80\xBA\xC9\x05\x00\xBB\xC9\x05\x00\xB8\xA5"
-     "\x05\x00"
-     "\xB9\xCD\x05\x00\xBE\xF9\x05\x00\xBF\xF9\x05\x00\xBC\xCD\x05\x00\xBD\xC5"
-     "\x05\x00"
-     "\xB2\xA5\x05\x00\xB3\xAD\x05\x00\xB0\x11\x06\x00\xB1\x11\x06\x00\xB6\xAD"
-     "\x05\x00"
-     "\xB7\x9D\x05\x00\xB4\xB5\x05\x00\xB5\xAD\x05\x00\xAA\x61\x06\x00\xAB\x61"
-     "\x06\x00"
-     "\xA8\x9D\x06\x00\xA9\x95\x06\x00\xAE\x61\x06\x00\xAF\x61\x06\x00\xAC\x71"
-     "\x06\x00"
-     "\xAD\x71\x06\x00\x7B\x96\x00\x80\x7F\x96\x00\x80\x83\x96\x00\x80\x87\x96"
-     "\x00\x80"
-     "\x8B\x96\x00\x80\x8F\x96\x00\x80\xBE\xB0\x0C\x00\x93\x96\x00\x80\xA8\x21"
-     "\x0E\x00"
-     "\xA9\x21\x0E\x00\xAA\x21\x0E\x00\xAB\x3D\x0E\x00\xAC\x25\x0E\x00\xAD\x2D"
-     "\x0E\x00"
-     "\xAE\x25\x0E\x00\xAF\x89\x0E\x00\x63\x96\x00\x80\x97\x96\x00\x80\x9B\x96"
-     "\x00\x80"
-     "\x9F\x96\x00\x80\xA3\x96\x00\x80\xA7\x96\x00\x80\xAB\x96\x00\x80\xAF\x96"
-     "\x00\x80"
-     "\xB8\x51\x0F\x00\xB9\x5D\x0F\x00\xBA\x55\x0F\x00\xBB\x6D\x0F\x00\xBC\x75"
-     "\x0F\x00"
-     "\xBD\x7D\x0F\x00\xBE\x75\x0F\x00\xBF\x69\x0F\x00\xB0\xF9\x0E\x00\xB1\xA1"
-     "\x0E\x00"
-     "\xB2\xA1\x0E\x00\xB3\xA1\x0E\x00\xB4\xA1\x0E\x00\xB5\xA9\x0E\x00\xB6\x91"
-     "\x0E\x00"
-     "\xB7\x91\x0E\x00\xB3\xA9\x0E\x00\xB3\x96\x00\x80\xF3\x95\x00\x80\xB7\x96"
-     "\x00\x80"
-     "\xBB\x96\x00\x80\xB6\xAD\x0E\x00\xB5\xAD\x0E\x00\xBF\x96\x00\x80\xBB\x65"
-     "\x0E\x00"
-     "\xBA\x3D\x0E\x00\xC3\x96\x00\x80\xC7\x96\x00\x80\xBF\x6D\x0E\x00\xBE\x65"
-     "\x0E\x00"
-     "\xBD\x75\x0E\x00\xBC\x75\x0E\x00\x82\x19\x00\x00\xA3\xED\x0E\x00\x80\x65"
-     "\x00\x00"
-     "\x81\x19\x00\x00\xA6\xE9\x0E\x00\xCB\x96\x00\x80\xCF\x96\x00\x80\xA5\xE9"
-     "\x0E\x00"
-     "\xAA\x79\x0E\x00\xAB\x21\x0E\x00\xD3\x96\x00\x80\xD7\x96\x00\x80\xAE\x21"
-     "\x0E\x00"
-     "\xAF\x29\x0E\x00\xAC\x31\x0E\x00\xAD\x31\x0E\x00\xA8\x61\x0E\x00\xA9\x75"
-     "\x0E\x00"
-     "\xAA\x7D\x0E\x00\xAB\x75\x0E\x00\xAC\x6D\x0E\x00\xAD\xF5\x0E\x00\xAE\xFD"
-     "\x0E\x00"
-     "\xAF\xF5\x0E\x00\x86\x80\x01\x00\x87\xA4\x01\x00\xDB\x96\x00\x80\xDF\x96"
-     "\x00\x80"
-     "\xE3\x96\x00\x80\xE7\x96\x00\x80\xEB\x96\x00\x80\xEF\x96\x00\x80\xB8\x75"
-     "\x01\x00"
-     "\xB9\x7D\x01\x00\xBA\x75\x01\x00\xBB\xC9\x01\x00\xBC\xDD\x01\x00\xBD\xC5"
-     "\x01\x00"
-     "\xBE\xC5\x01\x00\xBF\xF5\x01\x00\xB0\x8D\x0E\x00\xB1\x95\x0E\x00\xB2\x9D"
-     "\x0E\x00"
-     "\xB3\x91\x0E\x00\xB4\x55\x01\x00\xB5\x5D\x01\x00\xB6\x55\x01\x00\xB7\x4D"
-     "\x01\x00"
-     "\xB3\xF5\x0E\x00\xF3\x96\x00\x80\xF7\x96\x00\x80\xFB\x96\x00\x80\xFF\x96"
-     "\x00\x80"
-     "\xB6\x75\x0E\x00\xB5\xE5\x0E\x00\x03\x97\x00\x80\xBB\x51\x0E\x00\xBA\x49"
-     "\x0E\x00"
-     "\x07\x97\x00\x80\x0B\x97\x00\x80\xBF\x99\x01\x00\xBE\x91\x01\x00\xBD\x45"
-     "\x0E\x00"
-     "\xBC\x49\x0E\x00\x0F\x97\x00\x80\xA3\xB1\x0E\x00\x13\x97\x00\x80\x17\x97"
-     "\x00\x80"
-     "\xA6\x31\x0E\x00\x1B\x97\x00\x80\x1F\x97\x00\x80\xA5\xA1\x0E\x00\xAA\x0D"
-     "\x0E\x00"
-     "\xAB\x15\x0E\x00\x23\x97\x00\x80\x27\x97\x00\x80\xAE\xD5\x01\x00\xAF\xDD"
-     "\x01\x00"
-     "\xAC\x0D\x0E\x00\xAD\x01\x0E\x00\xA8\xED\x02\x00\xA9\x2D\x03\x00\xAA\x25"
-     "\x03\x00"
-     "\xAB\x3D\x03\x00\xAC\x25\x03\x00\xAD\x2D\x03\x00\xAE\x25\x03\x00\xAF\x99"
-     "\x03\x00"
-     "\x2B\x97\x00\x80\x2F\x97\x00\x80\x33\x97\x00\x80\x37\x97\x00\x80\x3B\x97"
-     "\x00\x80"
-     "\x3F\x97\x00\x80\xBE\x74\x02\x00\x47\x97\x00\x80\xB8\x8D\x03\x00\xB9\x91"
-     "\x03\x00"
-     "\xBA\x91\x03\x00\xBB\xA5\x03\x00\xBC\xBD\x03\x00\xBD\x75\x00\x00\xBE\x7D"
-     "\x00\x00"
-     "\xBF\x75\x00\x00\xB0\xE9\x03\x00\xB1\xE9\x03\x00\xB2\xF9\x03\x00\xB3\xF1"
-     "\x03\x00"
-     "\xB4\xD9\x03\x00\xB5\xD9\x03\x00\xB6\xB9\x03\x00\xB7\xB5\x03\x00\x80\xAD"
-     "\x00\x00"
-     "\x81\xB5\x00\x00\x82\xBD\x00\x00\xB3\xA1\x03\x00\x4B\x97\x00\x80\xB5\xA1"
-     "\x03\x00"
-     "\xB6\xA1\x03\x00\x4F\x97\x00\x80\x84\xE0\x02\x00\x53\x97\x00\x80\xBA\x21"
-     "\x03\x00"
-     "\xBB\x21\x03\x00\xBC\x21\x03\x00\xBD\x29\x03\x00\xBE\x11\x03\x00\xBF\x11"
-     "\x03\x00"
-     "\xA3\xED\x03\x00\x85\xC0\x04\x00\x85\xB4\x6F\x00\x57\x97\x00\x80\x5B\x97"
-     "\x00\x80"
-     "\xA6\xED\x03\x00\xA5\xED\x03\x00\x5F\x97\x00\x80\xAB\x6D\x03\x00\xAA\x6D"
-     "\x03\x00"
-     "\x86\x48\x05\x00\x87\x6C\x03\x00\xAF\x5D\x03\x00\xAE\x5D\x03\x00\xAD\x65"
-     "\x03\x00"
-     "\xAC\x6D\x03\x00\x63\x97\x00\x80\xE3\x00\x0E\x00\xEF\x58\x6C\x00\xE1\xB4"
-     "\x0F\x00"
-     "\x67\x97\x00\x80\x6B\x97\x00\x80\x6F\x97\x00\x80\x73\x97\x00\x80\xA1\xA9"
-     "\x03\x00"
-     "\xA0\xFD\x0F\x00\xA3\xC1\x03\x00\xA2\x0D\x03\x00\xE1\xE0\x03\x00\xEF\xE0"
-     "\x0F\x00"
-     "\xE3\xE4\x03\x00\x77\x97\x00\x80\x7B\x97\x00\x80\x7F\x97\x00\x80\x84\x04"
-     "\x04\x00"
-     "\xBE\x04\x04\x00\x83\x97\x00\x80\xEF\x94\x03\x00\x87\x97\x00\x80\x8B\x97"
-     "\x00\x80"
-     "\x8F\x97\x00\x80\xE3\xD4\x03\x00\x93\x97\x00\x80\xE1\x54\x00\x00\x97\x97"
-     "\x00\x80"
-     "\x9B\x97\x00\x80\x9F\x97\x00\x80\xA3\x97\x00\x80\x80\x0D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x1D\x00\x00\xA7\x97\x00\x80\xAB\x97\x00\x80\xAF\x97\x00\x80\x8F\x91"
-     "\x1B\x00"
-     "\xEF\x9C\x0E\x00\x84\xE0\x07\x00\xE1\xD4\x0E\x00\xB7\x97\x00\x80\xE3\xF0"
-     "\x0E\x00"
-     "\xBB\x97\x00\x80\xBF\x97\x00\x80\x86\x18\x07\x00\x87\x90\x04\x00\x99\xE5"
-     "\x17\x00"
-     "\x98\xE5\x17\x00\x9B\xE9\x0B\x00\x9A\x39\x0B\x00\x9D\xFD\x0A\x00\x9C\xF1"
-     "\x0B\x00"
-     "\x9F\x55\x0F\x00\x9E\x5D\x0F\x00\x91\x29\x1F\x00\x90\xCD\x1B\x00\x93\x25"
-     "\x1F\x00"
-     "\x92\xCD\x1F\x00\x95\xD1\x13\x00\x94\x29\x13\x00\x97\x19\x17\x00\x96\x75"
-     "\x13\x00"
-     "\x8C\xE1\x10\x00\x8D\x25\x10\x00\x8E\x2D\x10\x00\x8F\xF9\x0C\x00\x43\x97"
-     "\x00\x80"
-     "\xB3\x97\x00\x80\x8A\x39\x14\x00\x8B\x95\x14\x00\x84\xE9\x18\x00\x85\x05"
-     "\x18\x00"
-     "\x86\xE5\x18\x00\x87\xF1\x14\x00\xC3\x97\x00\x80\xC7\x97\x00\x80\x82\x31"
-     "\x1C\x00"
-     "\x83\x15\x1C\x00\x9C\xA9\x04\x00\xCB\x97\x00\x80\xCF\x97\x00\x80\xD3\x97"
-     "\x00\x80"
-     "\xD7\x97\x00\x80\xDB\x97\x00\x80\x9A\xD1\x04\x00\x9B\x7D\x04\x00\x94\x4D"
-     "\x0D\x00"
-     "\x95\xE5\x08\x00\x96\xE5\x08\x00\x97\xB5\x08\x00\xDF\x97\x00\x80\xE3\x97"
-     "\x00\x80"
-     "\x92\x59\x0C\x00\x93\x59\x0C\x00\xA1\x91\x00\x00\xE7\x97\x00\x80\xA3\x01"
-     "\x7C\x00"
-     "\xA2\x99\x00\x00\xA5\xB9\x7C\x00\xA4\x91\x7C\x00\xA7\x99\x78\x00\xA6\xE1"
-     "\x7D\x00"
-     "\xA9\x81\x79\x00\xA8\xA1\x78\x00\xAB\x89\x74\x00\xAA\x01\x74\x00\xAD\x01"
-     "\x70\x00"
-     "\xAC\x59\x74\x00\xAF\x85\x70\x00\xAE\x9D\x70\x00\xB1\xE1\x6C\x00\xB0\x01"
-     "\x6C\x00"
-     "\xB3\x01\x68\x00\xB2\x1D\x6C\x00\xB5\xF5\x68\x00\xB4\xF5\x68\x00\xEB\x97"
-     "\x00\x80"
-     "\xEF\x97\x00\x80\xF3\x97\x00\x80\xF7\x97\x00\x80\xFB\x97\x00\x80\xFF\x97"
-     "\x00\x80"
-     "\x03\x98\x00\x80\x07\x98\x00\x80\x0B\x98\x00\x80\x0F\x98\x00\x80\xA8\x3D"
-     "\x07\x00"
-     "\xA9\x95\x07\x00\xAA\x95\x07\x00\xAB\xA9\x07\x00\xAC\xDD\x07\x00\xAD\xC5"
-     "\x07\x00"
-     "\xAE\xC5\x07\x00\xAF\x1D\x06\x00\x13\x98\x00\x80\x82\x1D\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x17\x98\x00\x80\x1B\x98\x00\x80\x1F\x98\x00\x80\xBE\x60"
-     "\x01\x00"
-     "\xB8\x19\x06\x00\xB9\x19\x06\x00\xBA\x29\x06\x00\xBB\x25\x06\x00\xBC\x21"
-     "\x06\x00"
-     "\xBD\x21\x06\x00\xBE\x21\x06\x00\xBF\x21\x06\x00\xB0\x71\x06\x00\xB1\x71"
-     "\x06\x00"
-     "\xB2\x71\x06\x00\xB3\x71\x06\x00\xB4\x4D\x06\x00\xB5\x35\x06\x00\xB6\x3D"
-     "\x06\x00"
-     "\xB7\x2D\x06\x00\xB3\x1D\x07\x00\x27\x98\x00\x80\x86\x28\x00\x00\x87\xA8"
-     "\x00\x00"
-     "\x2B\x98\x00\x80\xB6\x45\x07\x00\xB5\x55\x07\x00\x2F\x98\x00\x80\xBB\xB5"
-     "\x06\x00"
-     "\xBA\xB5\x06\x00\x33\x98\x00\x80\x37\x98\x00\x80\xBF\xF1\x06\x00\xBE\x89"
-     "\x06\x00"
-     "\xBD\xA5\x06\x00\xBC\xA5\x06\x00\x3B\x98\x00\x80\xA3\x59\x07\x00\x3F\x98"
-     "\x00\x80"
-     "\x43\x98\x00\x80\xA6\x01\x07\x00\x47\x98\x00\x80\x4B\x98\x00\x80\xA5\x11"
-     "\x07\x00"
-     "\xAA\xF1\x06\x00\xAB\xF1\x06\x00\x4F\x98\x00\x80\x53\x98\x00\x80\xAE\xCD"
-     "\x06\x00"
-     "\xAF\xB5\x06\x00\xAC\xE1\x06\x00\xAD\xE1\x06\x00\xA8\xA9\x05\x00\xA9\xB5"
-     "\x05\x00"
-     "\xAA\xBD\x05\x00\xAB\x3D\x02\x00\xAC\x25\x02\x00\xAD\x51\x02\x00\xAE\x51"
-     "\x02\x00"
-     "\xAF\x51\x02\x00\x57\x98\x00\x80\x5B\x98\x00\x80\x5F\x98\x00\x80\x63\x98"
-     "\x00\x80"
-     "\x84\x3C\x03\x00\x67\x98\x00\x80\x6B\x98\x00\x80\x6F\x98\x00\x80\xB8\xA5"
-     "\x02\x00"
-     "\xB9\xAD\x02\x00\xBA\xA5\x02\x00\xBB\xBD\x02\x00\xBC\xA5\x02\x00\xBD\xAD"
-     "\x02\x00"
-     "\xBE\xA5\x02\x00\xBF\x7D\x03\x00\xB0\x31\x02\x00\xB1\x31\x02\x00\xB2\x19"
-     "\x02\x00"
-     "\xB3\x19\x02\x00\xB4\xF5\x02\x00\xB5\xAD\x02\x00\xB6\xA5\x02\x00\xB7\x9D"
-     "\x02\x00"
-     "\x73\x98\x00\x80\x77\x98\x00\x80\x7B\x98\x00\x80\xB3\x39\x05\x00\x7F\x98"
-     "\x00\x80"
-     "\xB5\xA1\x02\x00\xB6\xDD\x02\x00\x83\x98\x00\x80\x87\x98\x00\x80\x8B\x98"
-     "\x00\x80"
-     "\xBA\xF9\x02\x00\xBB\xF9\x02\x00\xBC\xC1\x02\x00\xBD\xC1\x02\x00\xBE\x3D"
-     "\x02\x00"
-     "\xBF\x65\x02\x00\x8F\x98\x00\x80\xA6\x81\x02\x00\xA5\xFD\x02\x00\x97\x98"
-     "\x00\x80"
-     "\xA3\x65\x05\x00\xBE\x58\x7C\x00\x86\xD8\x7C\x00\x87\x9C\x7C\x00\xAF\x39"
-     "\x02\x00"
-     "\xAE\x61\x02\x00\xAD\x9D\x02\x00\xAC\x9D\x02\x00\xAB\xA5\x02\x00\xAA\xA5"
-     "\x02\x00"
-     "\x9B\x98\x00\x80\x9F\x98\x00\x80\xA8\x85\x02\x00\xA9\x95\x02\x00\xAA\x95"
-     "\x02\x00"
-     "\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xD5\x02\x00\xAE\xD1\x02\x00\xAF\xD1"
-     "\x02\x00"
-     "\x81\x85\x01\x00\x80\x85\x01\x00\xA3\x98\x00\x80\x82\xED\x01\x00\xA7\x98"
-     "\x00\x80"
-     "\xAB\x98\x00\x80\xAF\x98\x00\x80\xB3\x98\x00\x80\xB8\x75\x01\x00\xB9\x7D"
-     "\x01\x00"
-     "\xBA\x75\x01\x00\xBB\xCD\x01\x00\xBC\xD5\x01\x00\xBD\xDD\x01\x00\xBE\xC9"
-     "\x01\x00"
-     "\xBF\xC1\x01\x00\xB0\xB5\x02\x00\xB1\xBD\x02\x00\xB2\x81\x02\x00\xB3\x81"
-     "\x02\x00"
-     "\xB4\x51\x01\x00\xB5\x51\x01\x00\xB6\x51\x01\x00\xB7\x51\x01\x00\xB7\x98"
-     "\x00\x80"
-     "\xBB\x98\x00\x80\xBF\x98\x00\x80\xC3\x98\x00\x80\xE1\x30\x06\x00\xE1\x64"
-     "\x07\x00"
-     "\xE3\x28\x06\x00\xE3\xC4\x06\x00\x84\x20\x7D\x00\xC7\x98\x00\x80\xEF\x6C"
-     "\x00\x00"
-     "\xEF\x18\x06\x00\xCB\x98\x00\x80\xCF\x98\x00\x80\xD3\x98\x00\x80\xD7\x98"
-     "\x00\x80"
-     "\xB3\x5D\x02\x00\xBE\x48\x7C\x00\xDB\x98\x00\x80\xDF\x98\x00\x80\xE3\x98"
-     "\x00\x80"
-     "\xB6\x15\x02\x00\xB5\x75\x02\x00\xE7\x98\x00\x80\xBB\x39\x02\x00\xBA\x31"
-     "\x02\x00"
-     "\xEB\x98\x00\x80\xEF\x98\x00\x80\xBF\xD5\x01\x00\xBE\xD5\x01\x00\xBD\x15"
-     "\x02\x00"
-     "\xBC\x15\x02\x00\xA3\x9D\x7D\x00\x93\x98\x00\x80\xF3\x98\x00\x80\xF7\x98"
-     "\x00\x80"
-     "\xFB\x98\x00\x80\xA6\xD5\x7D\x00\xA5\xB5\x7D\x00\xFF\x98\x00\x80\xAB\xF9"
-     "\x7D\x00"
-     "\xAA\xF1\x7D\x00\x03\x99\x00\x80\x07\x99\x00\x80\xAF\x15\x7E\x00\xAE\x15"
-     "\x7E\x00"
-     "\xAD\xD5\x7D\x00\xAC\xD5\x7D\x00\x80\x4D\x00\x00\x81\x55\x00\x00\x82\x55"
-     "\x00\x00"
-     "\xB3\xB1\x7E\x00\x0B\x99\x00\x80\xB5\x65\x7F\x00\xB6\x6D\x7F\x00\x0F\x99"
-     "\x00\x80"
-     "\x86\x40\x03\x00\x87\x04\x03\x00\xBA\xE9\x7F\x00\xBB\xE9\x7F\x00\xBC\xF9"
-     "\x7F\x00"
-     "\xBD\xF9\x7F\x00\xBE\xDD\x7F\x00\xBF\xCD\x7F\x00\x13\x99\x00\x80\x17\x99"
-     "\x00\x80"
-     "\x23\x98\x00\x80\x1B\x99\x00\x80\x1F\x99\x00\x80\x23\x99\x00\x80\x27\x99"
-     "\x00\x80"
-     "\x2B\x99\x00\x80\xA8\x6D\x7E\x00\xA9\x5D\x7E\x00\xAA\x55\x7E\x00\xAB\x85"
-     "\x7F\x00"
-     "\xAC\x81\x7F\x00\xAD\x81\x7F\x00\xAE\x81\x7F\x00\xAF\x81\x7F\x00\xB0\x41"
-     "\x7F\x00"
-     "\xB1\x41\x7F\x00\xB2\x41\x7F\x00\xB3\x41\x7F\x00\xB4\x75\x7F\x00\xB5\x65"
-     "\x7F\x00"
-     "\xB6\x6D\x7F\x00\xB7\x65\x7F\x00\xB8\x5D\x7F\x00\xB9\x2D\x7F\x00\xBA\x25"
-     "\x7F\x00"
-     "\xBB\x3D\x7F\x00\xBC\x2D\x7F\x00\xBD\x1D\x7F\x00\xBE\x15\x7F\x00\xBF\xF5"
-     "\x00\x00"
-     "\xA3\xFD\x7F\x00\x2F\x99\x00\x80\x33\x99\x00\x80\x37\x99\x00\x80\x3B\x99"
-     "\x00\x80"
-     "\xA6\x21\x7E\x00\xA5\x29\x7E\x00\x3F\x99\x00\x80\xAB\xA5\x7E\x00\xAA\xA5"
-     "\x7E\x00"
-     "\x43\x99\x00\x80\x47\x99\x00\x80\xAF\x81\x7E\x00\xAE\x91\x7E\x00\xAD\xB5"
-     "\x7E\x00"
-     "\xAC\xB5\x7E\x00\x4B\x99\x00\x80\x4F\x99\x00\x80\x53\x99\x00\x80\x57\x99"
-     "\x00\x80"
-     "\x5B\x99\x00\x80\x5F\x99\x00\x80\x63\x99\x00\x80\x67\x99\x00\x80\x80\x3D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x19\x00\x00\x6B\x99\x00\x80\x6F\x99\x00\x80\x84\xB8"
-     "\x01\x00"
-     "\xBE\xB0\x01\x00\x73\x99\x00\x80\xA8\xAD\x01\x00\xA9\xD5\x01\x00\xAA\xD5"
-     "\x01\x00"
-     "\xAB\x0D\x01\x00\xAC\x15\x01\x00\xAD\x19\x01\x00\xAE\x09\x01\x00\xAF\x09"
-     "\x01\x00"
-     "\x86\x00\x04\x00\x87\x04\x01\x00\x77\x99\x00\x80\x7B\x99\x00\x80\x7F\x99"
-     "\x00\x80"
-     "\x83\x99\x00\x80\x87\x99\x00\x80\x8B\x99\x00\x80\xB8\xED\x00\x00\xB9\x85"
-     "\x00\x00"
-     "\xBA\x8D\x00\x00\xBB\x85\x00\x00\xBC\x9D\x00\x00\xBD\x85\x00\x00\xBE\x8D"
-     "\x00\x00"
-     "\xBF\x85\x00\x00\xB0\x79\x01\x00\xB1\x79\x01\x00\xB2\xED\x00\x00\xB3\xE5"
-     "\x00\x00"
-     "\xB4\xFD\x00\x00\xB5\xE5\x00\x00\xB6\xE5\x00\x00\xB7\xD5\x00\x00\xB3\x5D"
-     "\x02\x00"
-     "\x8F\x99\x00\x80\x93\x99\x00\x80\x97\x99\x00\x80\x9B\x99\x00\x80\xB6\x99"
-     "\x02\x00"
-     "\xB5\x9D\x02\x00\x9F\x99\x00\x80\xBB\xBD\x02\x00\xBA\xBD\x02\x00\xA3\x99"
-     "\x00\x80"
-     "\xA7\x99\x00\x80\xBF\x21\x03\x00\xBE\x39\x03\x00\xBD\x31\x03\x00\xBC\x39"
-     "\x03\x00"
-     "\xBE\x28\x03\x00\xA3\x19\x02\x00\xAB\x99\x00\x80\xAF\x99\x00\x80\xA6\xDD"
-     "\x02\x00"
-     "\xB3\x99\x00\x80\xB7\x99\x00\x80\xA5\xD9\x02\x00\xAA\xF9\x02\x00\xAB\xF9"
-     "\x02\x00"
-     "\xBB\x99\x00\x80\xBF\x99\x00\x80\xAE\x7D\x03\x00\xAF\x65\x03\x00\xAC\x7D"
-     "\x03\x00"
-     "\xAD\x75\x03\x00\xBE\xC8\x04\x00\xC3\x99\x00\x80\xC7\x99\x00\x80\xBE\xC4"
-     "\x05\x00"
-     "\xCB\x99\x00\x80\xCF\x99\x00\x80\xD3\x99\x00\x80\xD7\x99\x00\x80\x80\x3D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x19\x00\x00\xDB\x99\x00\x80\xDF\x99\x00\x80\x84\x38"
-     "\x03\x00"
-     "\xE7\x99\x00\x80\xEB\x99\x00\x80\xEF\x78\x02\x00\xEF\x99\x00\x80\x86\x48"
-     "\x04\x00"
-     "\x87\x54\x03\x00\xF3\x99\x00\x80\xF7\x99\x00\x80\xFB\x99\x00\x80\xFF\x99"
-     "\x00\x80"
-     "\x03\x9A\x00\x80\x07\x9A\x00\x80\x0B\x9A\x00\x80\x0F\x9A\x00\x80\xE3\x54"
-     "\x02\x00"
-     "\x13\x9A\x00\x80\xE1\x40\x01\x00\x17\x9A\x00\x80\x1B\x9A\x00\x80\xE3\x24"
-     "\x7F\x00"
-     "\x1F\x9A\x00\x80\xE1\x9C\x7C\x00\x23\x9A\x00\x80\x27\x9A\x00\x80\x2B\x9A"
-     "\x00\x80"
-     "\x2F\x9A\x00\x80\x84\x6C\x05\x00\x33\x9A\x00\x80\x37\x9A\x00\x80\x3B\x9A"
-     "\x00\x80"
-     "\x3F\x9A\x00\x80\xEF\x18\x7F\x00\x43\x9A\x00\x80\x47\x9A\x00\x80\xB3\xF1"
-     "\x02\x00"
-     "\x4B\x9A\x00\x80\x4F\x9A\x00\x80\x57\x9A\x00\x80\x5B\x9A\x00\x80\xB6\xE9"
-     "\x02\x00"
-     "\xB5\xE1\x02\x00\x5F\x9A\x00\x80\xBB\x71\x01\x00\xBA\x69\x01\x00\x87\xA0"
-     "\x05\x00"
-     "\x86\xCC\x04\x00\xBF\x39\x01\x00\xBE\x59\x01\x00\xBD\x51\x01\x00\xBC\x61"
-     "\x01\x00"
-     "\xE1\x40\x7F\x00\x63\x9A\x00\x80\xE3\x38\x7E\x00\x84\xC0\x04\x00\x82\xD9"
-     "\x00\x00"
-     "\xEF\x14\x00\x00\x80\xA5\x00\x00\x81\xD9\x00\x00\x67\x9A\x00\x80\xE3\xC0"
-     "\x00\x00"
-     "\x6B\x9A\x00\x80\xE1\xD4\x01\x00\x6F\x9A\x00\x80\x73\x9A\x00\x80\xEF\x84"
-     "\x7E\x00"
-     "\x77\x9A\x00\x80\xAA\xCD\x01\x00\xAB\xD5\x01\x00\x7B\x9A\x00\x80\x7F\x9A"
-     "\x00\x80"
-     "\xAE\xFD\x01\x00\xAF\x9D\x01\x00\xAC\xC5\x01\x00\xAD\xF5\x01\x00\x83\x9A"
-     "\x00\x80"
-     "\xA3\x55\x02\x00\x87\x9A\x00\x80\x8B\x9A\x00\x80\xA6\x4D\x02\x00\x8F\x9A"
-     "\x00\x80"
-     "\x93\x9A\x00\x80\xA5\x45\x02\x00\xE3\x99\x00\x80\x53\x9A\x00\x80\x97\x9A"
-     "\x00\x80"
-     "\x9B\x9A\x00\x80\x9F\x9A\x00\x80\xA3\x9A\x00\x80\xA7\x9A\x00\x80\xAB\x9A"
-     "\x00\x80"
-     "\xA8\x99\x06\x00\xA9\x99\x06\x00\xAA\xED\x06\x00\xAB\xFD\x06\x00\xAC\xE5"
-     "\x06\x00"
-     "\xAD\xED\x06\x00\xAE\xE5\x06\x00\xAF\xDD\x06\x00\xB0\xA5\x06\x00\xB1\xAD"
-     "\x06\x00"
-     "\xB2\xA5\x06\x00\xB3\xB9\x06\x00\xB4\xAD\x06\x00\xB5\x55\x07\x00\xB6\x51"
-     "\x07\x00"
-     "\xB7\x4D\x07\x00\xB8\x75\x07\x00\xB9\x7D\x07\x00\xBA\x75\x07\x00\xBB\x49"
-     "\x07\x00"
-     "\xBC\x59\x07\x00\xBD\x59\x07\x00\xBE\x49\x07\x00\xBF\x45\x07\x00\xB3\x45"
-     "\x06\x00"
-     "\xAF\x9A\x00\x80\xB3\x9A\x00\x80\xB7\x9A\x00\x80\xBB\x9A\x00\x80\xB6\x4D"
-     "\x06\x00"
-     "\xB5\x4D\x06\x00\xBF\x9A\x00\x80\xBB\x49\x06\x00\xBA\x41\x06\x00\x86\x08"
-     "\x00\x00"
-     "\x87\x8C\x00\x00\xBF\xB1\x07\x00\xBE\x49\x06\x00\xBD\x4D\x06\x00\xBC\x51"
-     "\x06\x00"
-     "\x82\x5D\x00\x00\xA3\x01\x06\x00\x80\x45\x00\x00\x81\x5D\x00\x00\xA6\x09"
-     "\x06\x00"
-     "\xC7\x9A\x00\x80\xCB\x9A\x00\x80\xA5\x09\x06\x00\xAA\x05\x06\x00\xAB\x0D"
-     "\x06\x00"
-     "\xCF\x9A\x00\x80\xD3\x9A\x00\x80\xAE\x0D\x06\x00\xAF\xF5\x07\x00\xAC\x15"
-     "\x06\x00"
-     "\xAD\x09\x06\x00\xA8\x4D\x06\x00\xA9\x55\x06\x00\xAA\x55\x06\x00\xAB\x89"
-     "\x06\x00"
-     "\xAC\xB1\x06\x00\xAD\xBD\x06\x00\xAE\xA9\x06\x00\xAF\xA9\x06\x00\xC3\x9A"
-     "\x00\x80"
-     "\xD7\x9A\x00\x80\xDB\x9A\x00\x80\xDF\x9A\x00\x80\xE3\x9A\x00\x80\xE7\x9A"
-     "\x00\x80"
-     "\xEB\x9A\x00\x80\xEF\x9A\x00\x80\xB8\x49\x01\x00\xB9\x49\x01\x00\xBA\x59"
-     "\x01\x00"
-     "\xBB\x59\x01\x00\xBC\x49\x01\x00\xBD\x49\x01\x00\xBE\xDD\x01\x00\xBF\xD5"
-     "\x01\x00"
-     "\xB0\xDD\x06\x00\xB1\xAD\x06\x00\xB2\xA5\x06\x00\xB3\x8D\x06\x00\xB4\x99"
-     "\x06\x00"
-     "\xB5\x99\x06\x00\xB6\x8D\x06\x00\xB7\x85\x06\x00\xB3\xDD\x06\x00\xF3\x9A"
-     "\x00\x80"
-     "\xF7\x9A\x00\x80\xFB\x9A\x00\x80\xFF\x9A\x00\x80\xB6\x3D\x06\x00\xB5\x39"
-     "\x06\x00"
-     "\x03\x9B\x00\x80\xBB\x69\x06\x00\xBA\x19\x06\x00\x07\x9B\x00\x80\x0B\x9B"
-     "\x00\x80"
-     "\xBF\x5D\x06\x00\xBE\x5D\x06\x00\xBD\x59\x06\x00\xBC\x71\x06\x00\x0F\x9B"
-     "\x00\x80"
-     "\xA3\x99\x06\x00\x13\x9B\x00\x80\x17\x9B\x00\x80\xA6\x79\x06\x00\x1B\x9B"
-     "\x00\x80"
-     "\x1F\x9B\x00\x80\xA5\x7D\x06\x00\xAA\x5D\x06\x00\xAB\x2D\x06\x00\x23\x9B"
-     "\x00\x80"
-     "\x27\x9B\x00\x80\xAE\x19\x06\x00\xAF\x19\x06\x00\xAC\x35\x06\x00\xAD\x1D"
-     "\x06\x00"
-     "\xA8\xD5\x02\x00\xA9\xDD\x02\x00\xAA\xE1\x02\x00\xAB\xE1\x02\x00\xAC\x35"
-     "\x03\x00"
-     "\xAD\x3D\x03\x00\xAE\x35\x03\x00\xAF\x2D\x03\x00\x80\xCD\x03\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x2B\x9B\x00\x80\x2F\x9B\x00\x80\x84\x18\x02\x00\xBE\x74"
-     "\x03\x00"
-     "\x37\x9B\x00\x80\xB8\xE9\x03\x00\xB9\xE9\x03\x00\xBA\x89\x03\x00\xBB\x85"
-     "\x03\x00"
-     "\xBC\x9D\x03\x00\xBD\x81\x03\x00\xBE\x81\x03\x00\xBF\xB5\x03\x00\xB0\x55"
-     "\x03\x00"
-     "\xB1\x5D\x03\x00\xB2\x55\x03\x00\xB3\xE9\x03\x00\xB4\xF9\x03\x00\xB5\xF9"
-     "\x03\x00"
-     "\xB6\xE9\x03\x00\xB7\xE1\x03\x00\x86\x20\x0C\x00\x87\x10\x03\x00\x3B\x9B"
-     "\x00\x80"
-     "\x3F\x9B\x00\x80\x43\x9B\x00\x80\x47\x9B\x00\x80\x4B\x9B\x00\x80\xEF\x5C"
-     "\x02\x00"
-     "\x4F\x9B\x00\x80\xE1\x14\x00\x00\x53\x9B\x00\x80\xE3\x88\x02\x00\xBE\xF8"
-     "\x0C\x00"
-     "\x57\x9B\x00\x80\x5B\x9B\x00\x80\x5F\x9B\x00\x80\xBB\xF9\x03\x00\xBA\xF1"
-     "\x03\x00"
-     "\xBE\x80\x0D\x00\x63\x9B\x00\x80\xBF\x5D\x03\x00\xBE\x5D\x03\x00\xBD\x5D"
-     "\x03\x00"
-     "\xBC\xE1\x03\x00\xB3\x09\x02\x00\x67\x9B\x00\x80\x6B\x9B\x00\x80\x6F\x9B"
-     "\x00\x80"
-     "\x73\x9B\x00\x80\xB6\xDD\x03\x00\xB5\xDD\x03\x00\x77\x9B\x00\x80\xA8\xA9"
-     "\x06\x00"
-     "\xA9\xA9\x06\x00\xAA\xB9\x06\x00\xAB\xB9\x06\x00\xAC\xA9\x06\x00\xAD\xA9"
-     "\x06\x00"
-     "\xAE\x1D\x05\x00\xAF\x15\x05\x00\x7B\x9B\x00\x80\x7F\x9B\x00\x80\x83\x9B"
-     "\x00\x80"
-     "\x87\x9B\x00\x80\x8B\x9B\x00\x80\x8F\x9B\x00\x80\x93\x9B\x00\x80\x97\x9B"
-     "\x00\x80"
-     "\xB8\x19\x05\x00\xB9\x2D\x05\x00\xBA\x25\x05\x00\xBB\xC9\x05\x00\xBC\xD9"
-     "\x05\x00"
-     "\xBD\xC5\x05\x00\xBE\xCD\x05\x00\xBF\xC5\x05\x00\xB0\x6D\x05\x00\xB1\x75"
-     "\x05\x00"
-     "\xB2\x75\x05\x00\xB3\x45\x05\x00\xB4\x5D\x05\x00\xB5\x3D\x05\x00\xB6\x35"
-     "\x05\x00"
-     "\xB7\x29\x05\x00\xE1\xF4\x06\x00\xE1\x54\x07\x00\xE3\x14\x06\x00\xE3\xBC"
-     "\x06\x00"
-     "\x81\x09\x00\x00\x80\xA9\x00\x00\x9B\x9B\x00\x80\x82\x39\x00\x00\x84\xEC"
-     "\x0D\x00"
-     "\x9F\x9B\x00\x80\x87\x88\x0C\x00\x86\x0C\x0C\x00\xA7\x9B\x00\x80\xAB\x9B"
-     "\x00\x80"
-     "\xEF\xCC\x07\x00\xEF\xC4\x07\x00\xA3\x29\x03\x00\xAF\x9B\x00\x80\xB3\x9B"
-     "\x00\x80"
-     "\xB7\x9B\x00\x80\xBB\x9B\x00\x80\xA6\xFD\x02\x00\xA5\xFD\x02\x00\xBF\x9B"
-     "\x00\x80"
-     "\xAB\xD9\x02\x00\xAA\xD1\x02\x00\xC3\x9B\x00\x80\xC7\x9B\x00\x80\xAF\x7D"
-     "\x02\x00"
-     "\xAE\x7D\x02\x00\xAD\x7D\x02\x00\xAC\xC1\x02\x00\xA8\x3D\x0E\x00\xA9\x8D"
-     "\x0E\x00"
-     "\xAA\x85\x0E\x00\xAB\x9D\x0E\x00\xAC\x85\x0E\x00\xAD\x8D\x0E\x00\xAE\xB9"
-     "\x0E\x00"
-     "\xAF\xB5\x0E\x00\xA3\x9B\x00\x80\xCB\x9B\x00\x80\xCF\x9B\x00\x80\xD3\x9B"
-     "\x00\x80"
-     "\xD7\x9B\x00\x80\xDB\x9B\x00\x80\xDF\x9B\x00\x80\xE3\x9B\x00\x80\xB8\xBD"
-     "\x0E\x00"
-     "\xB9\x65\x0F\x00\xBA\x6D\x0F\x00\xBB\x65\x0F\x00\xBC\x7D\x0F\x00\xBD\x65"
-     "\x0F\x00"
-     "\xBE\x6D\x0F\x00\xBF\x65\x0F\x00\xB0\xD5\x0E\x00\xB1\xDD\x0E\x00\xB2\xD5"
-     "\x0E\x00"
-     "\xB3\xA1\x0E\x00\xB4\x95\x0E\x00\xB5\x9D\x0E\x00\xB6\x95\x0E\x00\xB7\x8D"
-     "\x0E\x00"
-     "\xB3\x0D\x0E\x00\xE7\x9B\x00\x80\xEB\x9B\x00\x80\xEF\x9B\x00\x80\xF3\x9B"
-     "\x00\x80"
-     "\xB6\x0D\x0E\x00\xB5\x0D\x0E\x00\xF7\x9B\x00\x80\xBB\x19\x0E\x00\xBA\x11"
-     "\x0E\x00"
-     "\xFB\x9B\x00\x80\x33\x9B\x00\x80\xBF\x59\x0E\x00\xBE\x51\x0E\x00\xBD\x75"
-     "\x0E\x00"
-     "\xBC\x01\x0E\x00\xFF\x9B\x00\x80\xA3\x49\x0E\x00\x03\x9C\x00\x80\x07\x9C"
-     "\x00\x80"
-     "\xA6\x49\x0E\x00\x0B\x9C\x00\x80\x0F\x9C\x00\x80\xA5\x49\x0E\x00\xAA\x55"
-     "\x0E\x00"
-     "\xAB\x5D\x0E\x00\x84\xA4\x03\x00\x13\x9C\x00\x80\xAE\x15\x0E\x00\xAF\x1D"
-     "\x0E\x00"
-     "\xAC\x45\x0E\x00\xAD\x31\x0E\x00\xA8\xB1\x0E\x00\xA9\xB1\x0E\x00\xAA\xCD"
-     "\x0E\x00"
-     "\xAB\xC5\x0E\x00\xAC\xDD\x0E\x00\xAD\xC5\x0E\x00\xAE\xC5\x0E\x00\xAF\xF5"
-     "\x0E\x00"
-     "\x80\xED\x01\x00\x81\xF1\x01\x00\x82\xF1\x01\x00\x17\x9C\x00\x80\x86\x90"
-     "\x01\x00"
-     "\x87\xB4\x01\x00\x1B\x9C\x00\x80\x1F\x9C\x00\x80\xB8\xC5\x01\x00\xB9\xCD"
-     "\x01\x00"
-     "\xBA\xC5\x01\x00\xBB\xDD\x01\x00\xBC\xCD\x01\x00\xBD\xFD\x01\x00\xBE\x99"
-     "\x01\x00"
-     "\xBF\x95\x01\x00\xB0\x8D\x0E\x00\xB1\x41\x01\x00\xB2\x41\x01\x00\xB3\x41"
-     "\x01\x00"
-     "\xB4\x41\x01\x00\xB5\x41\x01\x00\xB6\x41\x01\x00\xB7\x41\x01\x00\xB3\x45"
-     "\x0E\x00"
-     "\x23\x9C\x00\x80\x27\x9C\x00\x80\x2B\x9C\x00\x80\x2F\x9C\x00\x80\xB6\x45"
-     "\x0E\x00"
-     "\xB5\x55\x0E\x00\x33\x9C\x00\x80\xBB\x85\x01\x00\xBA\x49\x0E\x00\x37\x9C"
-     "\x00\x80"
-     "\x3B\x9C\x00\x80\xBF\x85\x01\x00\xBE\x85\x01\x00\xBD\x95\x01\x00\xBC\x95"
-     "\x01\x00"
-     "\x3F\x9C\x00\x80\xA3\x01\x0E\x00\x43\x9C\x00\x80\x47\x9C\x00\x80\xA6\x01"
-     "\x0E\x00"
-     "\x4B\x9C\x00\x80\x4F\x9C\x00\x80\xA5\x11\x0E\x00\xAA\x0D\x0E\x00\xAB\xC1"
-     "\x01\x00"
-     "\x53\x9C\x00\x80\x57\x9C\x00\x80\xAE\xC1\x01\x00\xAF\xC1\x01\x00\xAC\xD1"
-     "\x01\x00"
-     "\xAD\xD1\x01\x00\xA8\x2D\x03\x00\xA9\x3D\x03\x00\xAA\x35\x03\x00\xAB\x89"
-     "\x03\x00"
-     "\xAC\x99\x03\x00\xAD\x99\x03\x00\xAE\x89\x03\x00\xAF\x81\x03\x00\x5B\x9C"
-     "\x00\x80"
-     "\x5F\x9C\x00\x80\x63\x9C\x00\x80\x67\x9C\x00\x80\x6B\x9C\x00\x80\x6F\x9C"
-     "\x00\x80"
-     "\x73\x9C\x00\x80\x77\x9C\x00\x80\xB8\xAD\x03\x00\xB9\x65\x00\x00\xBA\x6D"
-     "\x00\x00"
-     "\xBB\x65\x00\x00\xBC\x7D\x00\x00\xBD\x65\x00\x00\xBE\x6D\x00\x00\xBF\x65"
-     "\x00\x00"
-     "\xB0\xC9\x03\x00\xB1\xC9\x03\x00\xB2\xA9\x03\x00\xB3\xA5\x03\x00\xB4\xBD"
-     "\x03\x00"
-     "\xB5\xA1\x03\x00\xB6\xA1\x03\x00\xB7\x95\x03\x00\x80\xBD\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x7B\x9C\x00\x80\x7F\x9C\x00\x80\xBE\xD8\x03\x00\x87\x9C"
-     "\x00\x80"
-     "\x8B\x9C\x00\x80\x84\xAC\x02\x00\x8F\x9C\x00\x80\x87\xE8\x03\x00\x86\x0C"
-     "\x04\x00"
-     "\x93\x9C\x00\x80\x97\x9C\x00\x80\x9B\x9C\x00\x80\x9F\x9C\x00\x80\xB3\xFD"
-     "\x03\x00"
-     "\xA3\x9C\x00\x80\xA7\x9C\x00\x80\xAB\x9C\x00\x80\xAF\x9C\x00\x80\xB6\x59"
-     "\x03\x00"
-     "\xB5\x51\x03\x00\xB3\x9C\x00\x80\xBB\x4D\x03\x00\xBA\x4D\x03\x00\xB7\x9C"
-     "\x00\x80"
-     "\xBB\x9C\x00\x80\xBF\x25\x03\x00\xBE\x39\x03\x00\xBD\x31\x03\x00\xBC\x3D"
-     "\x03\x00"
-     "\xBF\x9C\x00\x80\xC3\x9C\x00\x80\xC7\x9C\x00\x80\xCB\x9C\x00\x80\xEF\x58"
-     "\x0F\x00"
-     "\xCF\x9C\x00\x80\xD3\x9C\x00\x80\xD7\x9C\x00\x80\xE3\x90\x0E\x00\xDB\x9C"
-     "\x00\x80"
-     "\xE1\xB0\x0F\x00\xDF\x9C\x00\x80\xE3\x9C\x00\x80\xE7\x9C\x00\x80\xEB\x9C"
-     "\x00\x80"
-     "\xEF\x9C\x00\x80\x80\x75\x00\x00\x81\x7D\x00\x00\x82\x75\x00\x00\x84\x18"
-     "\x05\x00"
-     "\xEF\x3C\x03\x00\xF7\x9C\x00\x80\xBE\x14\x05\x00\xFB\x9C\x00\x80\xE3\xD0"
-     "\x03\x00"
-     "\xFF\x9C\x00\x80\xE1\x40\x00\x00\x03\x9D\x00\x80\x86\xD0\x04\x00\x87\x58"
-     "\x05\x00"
-     "\x07\x9D\x00\x80\x0B\x9D\x00\x80\x0F\x9D\x00\x80\x13\x9D\x00\x80\x17\x9D"
-     "\x00\x80"
-     "\x1B\x9D\x00\x80\x1F\x9D\x00\x80\x23\x9D\x00\x80\x27\x9D\x00\x80\xEF\xAC"
-     "\x0F\x00"
-     "\x84\xEC\x04\x00\xE1\x10\x0E\x00\x2B\x9D\x00\x80\xE3\x54\x01\x00\x2F\x9D"
-     "\x00\x80"
-     "\x33\x9D\x00\x80\x37\x9D\x00\x80\x3B\x9D\x00\x80\xA3\xFD\x02\x00\x3F\x9D"
-     "\x00\x80"
-     "\x43\x9D\x00\x80\x47\x9D\x00\x80\x4B\x9D\x00\x80\xA6\x59\x02\x00\xA5\x51"
-     "\x02\x00"
-     "\x4F\x9D\x00\x80\xAB\x4D\x02\x00\xAA\x4D\x02\x00\x53\x9D\x00\x80\x57\x9D"
-     "\x00\x80"
-     "\xAF\x25\x02\x00\xAE\x39\x02\x00\xAD\x31\x02\x00\xAC\x3D\x02\x00\xA8\x99"
-     "\x06\x00"
-     "\xA9\x99\x06\x00\xAA\xAD\x06\x00\xAB\xBD\x06\x00\xAC\xA5\x06\x00\xAD\xAD"
-     "\x06\x00"
-     "\xAE\xA5\x06\x00\xAF\xD9\x06\x00\xF3\x9C\x00\x80\x82\x11\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x80\xC1\x07\x00\x5B\x9D\x00\x80\x5F\x9D\x00\x80\xBE\x70\x03\x00\x63\x9D"
-     "\x00\x80"
-     "\xB8\x49\x07\x00\xB9\x49\x07\x00\xBA\x5D\x07\x00\xBB\x55\x07\x00\xBC\x4D"
-     "\x07\x00"
-     "\xBD\x71\x07\x00\xBE\x75\x07\x00\xBF\x6D\x07\x00\xB0\xA9\x06\x00\xB1\xA9"
-     "\x06\x00"
-     "\xB2\xB9\x06\x00\xB3\xB1\x06\x00\xB4\x99\x06\x00\xB5\x99\x06\x00\xB6\x79"
-     "\x07\x00"
-     "\xB7\x79\x07\x00\xB5\x35\x06\x00\x67\x9D\x00\x80\x6B\x9D\x00\x80\xB6\x35"
-     "\x06\x00"
-     "\x86\x30\x00\x00\x87\x5C\x03\x00\xB3\x3D\x06\x00\x6F\x9D\x00\x80\xBD\x7D"
-     "\x06\x00"
-     "\xBC\x75\x06\x00\xBF\x45\x06\x00\xBE\x45\x06\x00\x83\x9C\x00\x80\x73\x9D"
-     "\x00\x80"
-     "\xBB\x75\x06\x00\xBA\x75\x06\x00\xA3\x65\x06\x00\x77\x9D\x00\x80\x7B\x9D"
-     "\x00\x80"
-     "\x7F\x9D\x00\x80\x83\x9D\x00\x80\xA6\x6D\x06\x00\xA5\x6D\x06\x00\x87\x9D"
-     "\x00\x80"
-     "\xAB\x2D\x06\x00\xAA\x2D\x06\x00\x8B\x9D\x00\x80\x8F\x9D\x00\x80\xAF\x1D"
-     "\x06\x00"
-     "\xAE\x1D\x06\x00\xAD\x25\x06\x00\xAC\x2D\x06\x00\xA8\x55\x06\x00\xA9\x59"
-     "\x06\x00"
-     "\xAA\x6D\x06\x00\xAB\x61\x06\x00\xAC\x69\x06\x00\xAD\x69\x06\x00\xAE\x99"
-     "\x06\x00"
-     "\xAF\x99\x06\x00\x93\x9D\x00\x80\x97\x9D\x00\x80\x9B\x9D\x00\x80\x9F\x9D"
-     "\x00\x80"
-     "\xA3\x9D\x00\x80\xA7\x9D\x00\x80\xAB\x9D\x00\x80\xAF\x9D\x00\x80\xB8\xF9"
-     "\x06\x00"
-     "\xB9\xF9\x06\x00\xBA\x8D\x06\x00\xBB\x85\x06\x00\xBC\x9D\x06\x00\xBD\x85"
-     "\x06\x00"
-     "\xBE\x85\x06\x00\xBF\xB5\x06\x00\xB0\xE9\x06\x00\xB1\xE9\x06\x00\xB2\xF9"
-     "\x06\x00"
-     "\xB3\xF9\x06\x00\xB4\xE9\x06\x00\xB5\xDD\x06\x00\xB6\xC9\x06\x00\xB7\xC9"
-     "\x06\x00"
-     "\xB3\xE5\x06\x00\xB3\x9D\x00\x80\xB7\x9D\x00\x80\xBB\x9D\x00\x80\xBF\x9D"
-     "\x00\x80"
-     "\xB6\xE9\x06\x00\xB5\xE1\x06\x00\xC3\x9D\x00\x80\xBB\x2D\x06\x00\xBA\x2D"
-     "\x06\x00"
-     "\xC7\x9D\x00\x80\xCB\x9D\x00\x80\xBF\x19\x06\x00\xBE\x2D\x06\x00\xBD\x2D"
-     "\x06\x00"
-     "\xBC\x2D\x06\x00\x82\x15\x00\x00\xA3\xA1\x06\x00\x80\x61\x00\x00\x81\x61"
-     "\x00\x00"
-     "\xA6\xAD\x06\x00\xCF\x9D\x00\x80\xBE\x90\x01\x00\xA5\xA5\x06\x00\xAA\x69"
-     "\x06\x00"
-     "\xAB\x69\x06\x00\x84\xA4\x01\x00\xD7\x9D\x00\x80\xAE\x69\x06\x00\xAF\x5D"
-     "\x06\x00"
-     "\xAC\x69\x06\x00\xAD\x69\x06\x00\xA8\x85\x02\x00\xA9\x8D\x02\x00\xAA\x95"
-     "\x02\x00"
-     "\xAB\xB9\x02\x00\xAC\xD5\x02\x00\xAD\xDD\x02\x00\xAE\xD5\x02\x00\xAF\xCD"
-     "\x02\x00"
-     "\x86\x80\x1C\x00\x87\x64\x03\x00\xDB\x9D\x00\x80\xBE\x60\x03\x00\xDF\x9D"
-     "\x00\x80"
-     "\xE3\x9D\x00\x80\xE7\x9D\x00\x80\xEB\x9D\x00\x80\xB8\x75\x03\x00\xB9\x7D"
-     "\x03\x00"
-     "\xBA\x75\x03\x00\xBB\xC9\x03\x00\xBC\xD9\x03\x00\xBD\xD9\x03\x00\xBE\xC9"
-     "\x03\x00"
-     "\xBF\xC1\x03\x00\xB0\xBD\x02\x00\xB1\x8D\x02\x00\xB2\x85\x02\x00\xB3\x4D"
-     "\x03\x00"
-     "\xB4\x55\x03\x00\xB5\x5D\x03\x00\xB6\x55\x03\x00\xB7\x4D\x03\x00\xB3\x1D"
-     "\x02\x00"
-     "\xEF\x9D\x00\x80\x84\x08\x03\x00\xF3\x9D\x00\x80\xF7\x9D\x00\x80\xB6\x5D"
-     "\x02\x00"
-     "\xB5\x5D\x02\x00\xFB\x9D\x00\x80\xBB\x49\x02\x00\xBA\x79\x02\x00\xFF\x9D"
-     "\x00\x80"
-     "\x03\x9E\x00\x80\xBF\x99\x03\x00\xBE\x91\x03\x00\xBD\x99\x03\x00\xBC\x51"
-     "\x02\x00"
-     "\xB0\x00\x00\x00\xA3\x59\x02\x00\x07\x9E\x00\x80\x0B\x9E\x00\x80\xA6\x19"
-     "\x02\x00"
-     "\x0F\x9E\x00\x80\x13\x9E\x00\x80\xA5\x19\x02\x00\xAA\x3D\x02\x00\xAB\x0D"
-     "\x02\x00"
-     "\x17\x9E\x00\x80\x1B\x9E\x00\x80\xAE\xD5\x03\x00\xAF\xDD\x03\x00\xAC\x15"
-     "\x02\x00"
-     "\xAD\xDD\x03\x00\x1F\x9E\x00\x80\x23\x9E\x00\x80\x27\x9E\x00\x80\xEF\xAC"
-     "\x06\x00"
-     "\x2B\x9E\x00\x80\x2F\x9E\x00\x80\x33\x9E\x00\x80\x37\x9E\x00\x80\xBE\xE8"
-     "\x1C\x00"
-     "\x3B\x9E\x00\x80\x3F\x9E\x00\x80\x47\x9E\x00\x80\x4B\x9E\x00\x80\xE1\x80"
-     "\x07\x00"
-     "\x4F\x9E\x00\x80\xE3\x60\x06\x00\x80\x5D\x00\x00\x81\x61\x00\x00\x82\x61"
-     "\x00\x00"
-     "\xB3\x7D\x01\x00\x53\x9E\x00\x80\xB5\x6D\x01\x00\xB6\x65\x01\x00\x57\x9E"
-     "\x00\x80"
-     "\x86\x20\x1D\x00\x87\x58\x1D\x00\xBA\xF9\x01\x00\xBB\xF1\x01\x00\xBC\xD9"
-     "\x01\x00"
-     "\xBD\xD9\x01\x00\xBE\xB1\x01\x00\xBF\xB1\x01\x00\xEF\xA0\x00\x00\x5B\x9E"
-     "\x00\x80"
-     "\x5F\x9E\x00\x80\x63\x9E\x00\x80\x67\x9E\x00\x80\x6B\x9E\x00\x80\x6F\x9E"
-     "\x00\x80"
-     "\xEF\x5C\x01\x00\x84\x6C\x1C\x00\xE1\xCC\x06\x00\x73\x9E\x00\x80\xE3\x1C"
-     "\x06\x00"
-     "\xE3\x48\x00\x00\x77\x9E\x00\x80\xE1\x30\x01\x00\x7B\x9E\x00\x80\xA3\xF1"
-     "\x01\x00"
-     "\x7F\x9E\x00\x80\x85\x00\x14\x00\x83\x9E\x00\x80\x87\x9E\x00\x80\xA6\xE9"
-     "\x01\x00"
-     "\xA5\xE1\x01\x00\x8B\x9E\x00\x80\xAB\x7D\x01\x00\xAA\x75\x01\x00\x8F\x9E"
-     "\x00\x80"
-     "\x93\x9E\x00\x80\xAF\x3D\x01\x00\xAE\x3D\x01\x00\xAD\x55\x01\x00\xAC\x55"
-     "\x01\x00"
-     "\xA8\xED\x1D\x00\xA9\x2D\x1E\x00\xAA\x39\x1E\x00\xAB\x39\x1E\x00\xAC\x29"
-     "\x1E\x00"
-     "\xAD\x29\x1E\x00\xAE\x9D\x1E\x00\xAF\x91\x1E\x00\x43\x9E\x00\x80\x97\x9E"
-     "\x00\x80"
-     "\x9B\x9E\x00\x80\x9F\x9E\x00\x80\xA3\x9E\x00\x80\x82\xF9\x00\x00\x81\xF1"
-     "\x00\x00"
-     "\x80\xFD\x00\x00\xB8\xA9\x1E\x00\xB9\xA9\x1E\x00\xBA\x49\x1F\x00\xBB\x49"
-     "\x1F\x00"
-     "\xBC\x59\x1F\x00\xBD\x45\x1F\x00\xBE\x4D\x1F\x00\xBF\x45\x1F\x00\xB0\xF1"
-     "\x1E\x00"
-     "\xB1\xF9\x1E\x00\xB2\xC1\x1E\x00\xB3\xC1\x1E\x00\xB4\xB9\x1E\x00\xB5\xB9"
-     "\x1E\x00"
-     "\xB6\xAD\x1E\x00\xB7\xA5\x1E\x00\xB0\x11\x1F\x00\xB1\x11\x1F\x00\xB2\x11"
-     "\x1F\x00"
-     "\xB3\x25\x1F\x00\xB4\x25\x1F\x00\xB5\x29\x1F\x00\xB6\x5D\x1F\x00\xB7\x51"
-     "\x1F\x00"
-     "\xB8\x71\x1F\x00\xB9\x79\x1F\x00\xBA\x41\x1F\x00\xBB\x41\x1F\x00\xBC\x95"
-     "\x00\x00"
-     "\xBD\x9D\x00\x00\xBE\x95\x00\x00\xBF\x8D\x00\x00\xA7\x9E\x00\x80\xD3\x9D"
-     "\x00\x80"
-     "\xAB\x9E\x00\x80\xAF\x9E\x00\x80\xB3\x9E\x00\x80\xB7\x9E\x00\x80\x86\xF8"
-     "\x03\x00"
-     "\x87\xD0\x00\x00\xA8\xCD\x1F\x00\xA9\xD5\x1F\x00\xAA\xD1\x1F\x00\xAB\xBD"
-     "\x1F\x00"
-     "\xAC\xAD\x1F\x00\xAD\x71\x1F\x00\xAE\x71\x1F\x00\xAF\x71\x1F\x00\xB3\x39"
-     "\x1E\x00"
-     "\xBB\x9E\x00\x80\xBF\x9E\x00\x80\xC3\x9E\x00\x80\xC7\x9E\x00\x80\xB6\x91"
-     "\x1E\x00"
-     "\xB5\x45\x1E\x00\xCB\x9E\x00\x80\xBB\xB5\x1E\x00\xBA\xB5\x1E\x00\xCF\x9E"
-     "\x00\x80"
-     "\xD3\x9E\x00\x80\xBF\x8D\x1E\x00\xBE\x81\x1E\x00\xBD\x91\x1E\x00\xBC\xA5"
-     "\x1E\x00"
-     "\xD7\x9E\x00\x80\xA3\x7D\x1E\x00\xDB\x9E\x00\x80\xDF\x9E\x00\x80\xA6\xD5"
-     "\x1E\x00"
-     "\xE3\x9E\x00\x80\xE7\x9E\x00\x80\xA5\x01\x1E\x00\xAA\xF1\x1E\x00\xAB\xF1"
-     "\x1E\x00"
-     "\xEB\x9E\x00\x80\xEF\x9E\x00\x80\xAE\xC5\x1E\x00\xAF\xC9\x1E\x00\xAC\xE1"
-     "\x1E\x00"
-     "\xAD\xD5\x1E\x00\xA8\x55\x01\x00\xA9\x81\x00\x00\xAA\x81\x00\x00\xAB\x81"
-     "\x00\x00"
-     "\xAC\x81\x00\x00\xAD\x89\x00\x00\xAE\xB1\x00\x00\xAF\xB1\x00\x00\xF3\x9E"
-     "\x00\x80"
-     "\xF7\x9E\x00\x80\xFB\x9E\x00\x80\xFF\x9E\x00\x80\x03\x9F\x00\x80\x07\x9F"
-     "\x00\x80"
-     "\x0B\x9F\x00\x80\x0F\x9F\x00\x80\xB8\x65\x00\x00\xB9\x6D\x00\x00\xBA\x65"
-     "\x00\x00"
-     "\xBB\x7D\x00\x00\xBC\x65\x00\x00\xBD\x6D\x00\x00\xBE\x65\x00\x00\xBF\xDD"
-     "\x03\x00"
-     "\xB0\xA1\x00\x00\xB1\xAD\x00\x00\xB2\xA5\x00\x00\xB3\xB9\x00\x00\xB4\xA9"
-     "\x00\x00"
-     "\xB5\x9D\x00\x00\xB6\x95\x00\x00\xB7\x5D\x00\x00\x13\x9F\x00\x80\x82\x1D"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x80\x1D\x00\x00\x17\x9F\x00\x80\x1B\x9F\x00\x80\x1F\x9F"
-     "\x00\x80"
-     "\xBE\x14\x02\x00\x27\x9F\x00\x80\x84\xA8\x02\x00\x2B\x9F\x00\x80\x2F\x9F"
-     "\x00\x80"
-     "\x33\x9F\x00\x80\x37\x9F\x00\x80\x3B\x9F\x00\x80\x8F\x00\x00\x00\xB3\x49"
-     "\x03\x00"
-     "\x3F\x9F\x00\x80\x86\xE8\x04\x00\x87\xAC\x02\x00\x43\x9F\x00\x80\xB6\x49"
-     "\x03\x00"
-     "\xB5\x49\x03\x00\x47\x9F\x00\x80\xBB\x29\x03\x00\xBA\x25\x03\x00\x4B\x9F"
-     "\x00\x80"
-     "\x4F\x9F\x00\x80\xBF\x19\x03\x00\xBE\x2D\x03\x00\xBD\x2D\x03\x00\xBC\x31"
-     "\x03\x00"
-     "\x53\x9F\x00\x80\xA3\x8D\x03\x00\x57\x9F\x00\x80\x5B\x9F\x00\x80\xA6\x8D"
-     "\x03\x00"
-     "\x5F\x9F\x00\x80\x63\x9F\x00\x80\xA5\x8D\x03\x00\xAA\xE1\x03\x00\xAB\xED"
-     "\x03\x00"
-     "\x67\x9F\x00\x80\x6B\x9F\x00\x80\xAE\xE9\x03\x00\xAF\xDD\x03\x00\xAC\xF5"
-     "\x03\x00"
-     "\xAD\xE9\x03\x00\xBE\x84\x03\x00\x6F\x9F\x00\x80\x73\x9F\x00\x80\x77\x9F"
-     "\x00\x80"
-     "\x7B\x9F\x00\x80\x7F\x9F\x00\x80\x83\x9F\x00\x80\x87\x9F\x00\x80\x80\x3D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x19\x00\x00\x8B\x9F\x00\x80\x8F\x9F\x00\x80\x97\x9F"
-     "\x00\x80"
-     "\x84\x34\x03\x00\x9B\x9F\x00\x80\xB3\x2D\x01\x00\x9F\x9F\x00\x80\x87\xCC"
-     "\x02\x00"
-     "\x86\x4C\x05\x00\xA3\x9F\x00\x80\xB6\x2D\x01\x00\xB5\x2D\x01\x00\xA7\x9F"
-     "\x00\x80"
-     "\xBB\x49\x01\x00\xBA\x79\x01\x00\xAB\x9F\x00\x80\xAF\x9F\x00\x80\xBF\xBD"
-     "\x01\x00"
-     "\xBE\xBD\x01\x00\xBD\xB9\x01\x00\xBC\x51\x01\x00\xE1\x78\x1F\x00\xB3\x9F"
-     "\x00\x80"
-     "\xE3\xD0\x1F\x00\xB7\x9F\x00\x80\xBB\x9F\x00\x80\xE1\x94\x01\x00\xBF\x9F"
-     "\x00\x80"
-     "\xE3\x68\x03\x00\xC3\x9F\x00\x80\xC7\x9F\x00\x80\xCB\x9F\x00\x80\xEF\x88"
-     "\x03\x00"
-     "\xCF\x9F\x00\x80\xD3\x9F\x00\x80\xEF\xAC\x1F\x00\xD7\x9F\x00\x80\xDB\x9F"
-     "\x00\x80"
-     "\xDF\x9F\x00\x80\xE3\x9F\x00\x80\xE7\x9F\x00\x80\xEB\x9F\x00\x80\xEF\x9F"
-     "\x00\x80"
-     "\xF3\x9F\x00\x80\xEF\x44\x1E\x00\xF7\x9F\x00\x80\xE1\xDC\x1E\x00\xFB\x9F"
-     "\x00\x80"
-     "\xE3\x1C\x1E\x00\xFF\x9F\x00\x80\x07\xA0\x00\x80\x0B\xA0\x00\x80\x0F\xA0"
-     "\x00\x80"
-     "\x81\x69\x00\x00\x80\x65\x00\x00\xA3\xE5\x01\x00\x82\x7D\x00\x00\xA5\xE5"
-     "\x01\x00"
-     "\x13\xA0\x00\x80\x84\x14\x04\x00\xA6\xE5\x01\x00\xBE\x28\x04\x00\x17\xA0"
-     "\x00\x80"
-     "\xAB\x81\x01\x00\xAA\xB1\x01\x00\xAD\x71\x01\x00\xAC\x99\x01\x00\xAF\x75"
-     "\x01\x00"
-     "\xAE\x75\x01\x00\xA8\x21\x06\x00\xA9\x2D\x06\x00\xAA\x25\x06\x00\xAB\x3D"
-     "\x06\x00"
-     "\xAC\x25\x06\x00\xAD\x2D\x06\x00\xAE\x5D\x06\x00\xAF\x4D\x06\x00\x93\x9F"
-     "\x00\x80"
-     "\x03\xA0\x00\x80\x86\x08\x03\x00\x87\x8C\x03\x00\x1B\xA0\x00\x80\x1F\xA0"
-     "\x00\x80"
-     "\x23\xA0\x00\x80\x27\xA0\x00\x80\xB8\xE9\x06\x00\xB9\xE9\x06\x00\xBA\x8D"
-     "\x06\x00"
-     "\xBB\x85\x06\x00\xBC\x9D\x06\x00\xBD\x85\x06\x00\xBE\x8D\x06\x00\xBF\x85"
-     "\x06\x00"
-     "\xB0\x3D\x06\x00\xB1\x0D\x06\x00\xB2\x05\x06\x00\xB3\xED\x06\x00\xB4\xF9"
-     "\x06\x00"
-     "\xB5\xF9\x06\x00\xB6\xED\x06\x00\xB7\xE5\x06\x00\xB0\xCD\x07\x00\xB1\xD5"
-     "\x07\x00"
-     "\xB2\xD1\x07\x00\xB3\xED\x07\x00\xB4\xF5\x07\x00\xB5\xFD\x07\x00\xB6\xE9"
-     "\x07\x00"
-     "\xB7\xE9\x07\x00\xB8\xDD\x07\x00\xB9\x25\x07\x00\xBA\x2D\x07\x00\xBB\x25"
-     "\x07\x00"
-     "\xBC\x3D\x07\x00\xBD\x25\x07\x00\xBE\x2D\x07\x00\xBF\x25\x07\x00\x2B\xA0"
-     "\x00\x80"
-     "\x2F\xA0\x00\x80\x23\x9F\x00\x80\x33\xA0\x00\x80\x37\xA0\x00\x80\x3B\xA0"
-     "\x00\x80"
-     "\x3F\xA0\x00\x80\x43\xA0\x00\x80\xA8\x15\x06\x00\xA9\x19\x06\x00\xAA\xED"
-     "\x07\x00"
-     "\xAB\xFD\x07\x00\xAC\xED\x07\x00\xAD\xDD\x07\x00\xAE\xD5\x07\x00\xAF\xB9"
-     "\x07\x00"
-     "\xB3\x05\x06\x00\x47\xA0\x00\x80\x4B\xA0\x00\x80\x4F\xA0\x00\x80\x53\xA0"
-     "\x00\x80"
-     "\xB6\x3D\x06\x00\xB5\x05\x06\x00\x57\xA0\x00\x80\xBB\x71\x06\x00\xBA\x69"
-     "\x06\x00"
-     "\x5B\xA0\x00\x80\x5F\xA0\x00\x80\xBF\x59\x06\x00\xBE\x51\x06\x00\xBD\x59"
-     "\x06\x00"
-     "\xBC\x65\x06\x00\x63\xA0\x00\x80\xA3\x41\x06\x00\x67\xA0\x00\x80\x6B\xA0"
-     "\x00\x80"
-     "\xA6\x79\x06\x00\x6F\xA0\x00\x80\x84\xB4\x01\x00\xA5\x41\x06\x00\xAA\x2D"
-     "\x06\x00"
-     "\xAB\x35\x06\x00\xBE\x80\x01\x00\x77\xA0\x00\x80\xAE\x15\x06\x00\xAF\x1D"
-     "\x06\x00"
-     "\xAC\x21\x06\x00\xAD\x1D\x06\x00\xA8\x35\x06\x00\xA9\x3D\x06\x00\xAA\x35"
-     "\x06\x00"
-     "\xAB\x59\x06\x00\xAC\x75\x06\x00\xAD\xA5\x01\x00\xAE\xAD\x01\x00\xAF\xA5"
-     "\x01\x00"
-     "\x80\xE9\x00\x00\x81\xE9\x00\x00\x82\xFD\x00\x00\xBF\x24\x01\x00\x86\x30"
-     "\x0F\x00"
-     "\x87\xE4\x00\x00\x7B\xA0\x00\x80\x7F\xA0\x00\x80\xB8\xC5\x00\x00\xB9\xCD"
-     "\x00\x00"
-     "\xBA\xC5\x00\x00\xBB\xDD\x00\x00\xBC\xCD\x00\x00\xBD\xFD\x00\x00\xBE\xF5"
-     "\x00\x00"
-     "\xBF\x9D\x00\x00\xB0\xDD\x01\x00\xB1\x25\x01\x00\xB2\x2D\x01\x00\xB3\x21"
-     "\x01\x00"
-     "\xB4\x21\x01\x00\xB5\x21\x01\x00\xB6\x21\x01\x00\xB7\x21\x01\x00\xBB\xC1"
-     "\x02\x00"
-     "\xBA\x39\x02\x00\x83\xA0\x00\x80\x87\xA0\x00\x80\xBF\xC5\x02\x00\xBE\xC5"
-     "\x02\x00"
-     "\xBD\xD5\x02\x00\xBC\xD9\x02\x00\xB3\x9D\x05\x00\x8B\xA0\x00\x80\x8F\xA0"
-     "\x00\x80"
-     "\x93\xA0\x00\x80\x8C\x00\x00\x00\xB6\x05\x02\x00\xB5\xDD\x05\x00\x97\xA0"
-     "\x00\x80"
-     "\xAA\x7D\x02\x00\xAB\x85\x02\x00\x9B\xA0\x00\x80\x9F\xA0\x00\x80\xAE\x81"
-     "\x02\x00"
-     "\xAF\x81\x02\x00\xAC\x9D\x02\x00\xAD\x91\x02\x00\xA3\xA0\x00\x80\xA3\xD9"
-     "\x05\x00"
-     "\xA7\xA0\x00\x80\xAB\xA0\x00\x80\xA6\x41\x02\x00\xAF\xA0\x00\x80\xB3\xA0"
-     "\x00\x80"
-     "\xA5\x99\x05\x00\x82\x91\x6A\x00\x83\x91\x6A\x00\xB7\xA0\x00\x80\xBB\xA0"
-     "\x00\x80"
-     "\x86\xB9\x16\x00\x87\xE9\x17\x00\x84\x11\x16\x00\x85\x99\x16\x00\x8A\xA1"
-     "\x12\x00"
-     "\x8B\xA1\x12\x00\xBF\xA0\x00\x80\xC3\xA0\x00\x80\x8E\x91\x1E\x00\x8F\x59"
-     "\x1E\x00"
-     "\x8C\x99\x13\x00\x8D\x11\x1E\x00\x92\x71\x1A\x00\x93\xE5\x1A\x00\xC7\xA0"
-     "\x00\x80"
-     "\xEF\xE8\x24\x00\x96\xF1\x06\x00\x97\x05\x06\x00\x94\xE5\x1A\x00\x95\x19"
-     "\x06\x00"
-     "\x9A\x29\x02\x00\x9B\xC5\x02\x00\xCB\xA0\x00\x80\xCF\xA0\x00\x80\xD3\xA0"
-     "\x00\x80"
-     "\xE1\x28\x1B\x00\x9C\xDD\x02\x00\xE3\x20\x0F\x00\x9F\x21\x07\x00\x9E\xC1"
-     "\x07\x00"
-     "\x9D\x35\x1B\x00\x9C\x2D\x1B\x00\x9B\xA1\x1B\x00\x9A\xF9\x1F\x00\x99\x39"
-     "\x1F\x00"
-     "\x98\xB1\x1F\x00\x97\x01\x12\x00\x96\x21\x13\x00\x95\x29\x13\x00\x94\x51"
-     "\x16\x00"
-     "\x93\x19\x17\x00\x92\x31\x17\x00\x91\xB1\x17\x00\x90\x29\x6B\x00\x8F\x51"
-     "\x6B\x00"
-     "\xE3\xAC\x07\x00\x84\x04\x0D\x00\xE1\x1C\x07\x00\x80\x0D\x00\x00\x81\x35"
-     "\x00\x00"
-     "\x82\x3D\x00\x00\xD7\xA0\x00\x80\xDB\xA0\x00\x80\xDF\xA0\x00\x80\xBE\x20"
-     "\x0D\x00"
-     "\xE7\xA0\x00\x80\xEB\xA0\x00\x80\xEF\x4C\x07\x00\x86\x58\x0C\x00\x87\x60"
-     "\x0D\x00"
-     "\xEF\xA0\x00\x80\xF3\xA0\x00\x80\xF7\xA0\x00\x80\xFB\xA0\x00\x80\x84\x5C"
-     "\x0F\x00"
-     "\xFF\xA0\x00\x80\xEF\x08\x00\x00\xEF\x84\x06\x00\xE1\x90\x01\x00\xE1\xB0"
-     "\x06\x00"
-     "\xE3\xE0\x00\x00\xE3\x64\x06\x00\x03\xA1\x00\x80\x07\xA1\x00\x80\x0B\xA1"
-     "\x00\x80"
-     "\x0F\xA1\x00\x80\x13\xA1\x00\x80\x17\xA1\x00\x80\xBE\x00\x0F\x00\x84\x40"
-     "\x0E\x00"
-     "\x1B\xA1\x00\x80\x1F\xA1\x00\x80\x23\xA1\x00\x80\x27\xA1\x00\x80\x2B\xA1"
-     "\x00\x80"
-     "\x2F\xA1\x00\x80\x33\xA1\x00\x80\x37\xA1\x00\x80\xA3\xD5\x03\x00\xA2\xD5"
-     "\x03\x00"
-     "\xA1\x31\x03\x00\xA0\x2D\x07\x00\x73\xA0\x00\x80\xE3\xA0\x00\x80\x3B\xA1"
-     "\x00\x80"
-     "\x3F\xA1\x00\x80\x43\xA1\x00\x80\x82\x99\x00\x00\x81\x91\x00\x00\x80\x99"
-     "\x00\x00"
-     "\xA8\x4D\x0D\x00\xA9\x9D\x0D\x00\xAA\x95\x0D\x00\xAB\x25\x0E\x00\xAC\x3D"
-     "\x0E\x00"
-     "\xAD\x11\x0E\x00\xAE\x11\x0E\x00\xAF\x11\x0E\x00\xB0\x7D\x0E\x00\xB1\x0D"
-     "\x0E\x00"
-     "\xB2\x05\x0E\x00\xB3\x2D\x0E\x00\xB4\x39\x0E\x00\xB5\x39\x0E\x00\xB6\x2D"
-     "\x0E\x00"
-     "\xB7\x25\x0E\x00\xB8\xE9\x0E\x00\xB9\xE9\x0E\x00\xBA\xC1\x0E\x00\xBB\xC1"
-     "\x0E\x00"
-     "\xBC\xB9\x0E\x00\xBD\x9D\x0E\x00\xBE\x95\x0E\x00\xBF\x8D\x0E\x00\xB3\x3D"
-     "\x0D\x00"
-     "\x47\xA1\x00\x80\x4B\xA1\x00\x80\x4F\xA1\x00\x80\x53\xA1\x00\x80\xB6\xB1"
-     "\x0E\x00"
-     "\xB5\x95\x0E\x00\x57\xA1\x00\x80\xBB\xE9\x0E\x00\xBA\x99\x0E\x00\x86\x88"
-     "\x00\x00"
-     "\x87\xE4\x00\x00\xBF\xDD\x0E\x00\xBE\xDD\x0E\x00\xBD\xD9\x0E\x00\xBC\xF1"
-     "\x0E\x00"
-     "\x5B\xA1\x00\x80\xA3\x79\x0D\x00\xBE\x84\x01\x00\x84\x80\x06\x00\xA6\xF5"
-     "\x0E\x00"
-     "\x5F\xA1\x00\x80\x63\xA1\x00\x80\xA5\xD1\x0E\x00\xAA\xDD\x0E\x00\xAB\xAD"
-     "\x0E\x00"
-     "\x67\xA1\x00\x80\x6B\xA1\x00\x80\xAE\x99\x0E\x00\xAF\x99\x0E\x00\xAC\xB5"
-     "\x0E\x00"
-     "\xAD\x9D\x0E\x00\xB2\x05\x35\x00\xB3\x19\x34\x00\xB0\x6D\x30\x00\xB1\x0D"
-     "\x35\x00"
-     "\x6F\xA1\x00\x80\x73\xA1\x00\x80\xB4\x01\x28\x00\xB5\x3D\x29\x00\x77\xA1"
-     "\x00\x80"
-     "\x7B\xA1\x00\x80\x7F\xA1\x00\x80\x83\xA1\x00\x80\x87\xA1\x00\x80\x8B\xA1"
-     "\x00\x80"
-     "\x8F\xA1\x00\x80\x93\xA1\x00\x80\xA2\x45\x01\x00\xA3\xD5\x01\x00\x97\xA1"
-     "\x00\x80"
-     "\xA1\x4D\x01\x00\xA6\xCD\x05\x00\xA7\x01\x38\x00\xA4\x01\x04\x00\xA5\x7D"
-     "\x05\x00"
-     "\xAA\x01\x3C\x00\xAB\x45\x3D\x00\xA8\x49\x39\x00\xA9\xE5\x39\x00\xAE\x75"
-     "\x31\x00"
-     "\xAF\x7D\x31\x00\xAC\x5D\x3D\x00\xAD\x01\x30\x00\xA8\xED\x0E\x00\xA9\xF5"
-     "\x0E\x00"
-     "\xAA\x09\x0E\x00\xAB\x09\x0E\x00\xAC\x19\x0E\x00\xAD\x19\x0E\x00\xAE\x0D"
-     "\x0E\x00"
-     "\xAF\x2D\x0E\x00\x9B\xA1\x00\x80\x9F\xA1\x00\x80\xA3\xA1\x00\x80\xA7\xA1"
-     "\x00\x80"
-     "\xAB\xA1\x00\x80\xAF\xA1\x00\x80\xB3\xA1\x00\x80\xB7\xA1\x00\x80\xB8\x1D"
-     "\x0E\x00"
-     "\xB9\x25\x0E\x00\xBA\x2D\x0E\x00\xBB\x25\x0E\x00\xBC\x3D\x0E\x00\xBD\xDD"
-     "\x01\x00"
-     "\xBE\xD5\x01\x00\xBF\xCD\x01\x00\xB0\x55\x0E\x00\xB1\x5D\x0E\x00\xB2\x55"
-     "\x0E\x00"
-     "\xB3\x2D\x0E\x00\xB4\x35\x0E\x00\xB5\x25\x0E\x00\xB6\x2D\x0E\x00\xB7\x25"
-     "\x0E\x00"
-     "\xB3\x81\x0D\x00\xBB\xA1\x00\x80\xBF\xA1\x00\x80\xC7\xA1\x00\x80\xCB\xA1"
-     "\x00\x80"
-     "\xB6\x99\x0D\x00\xB5\x91\x0D\x00\xBE\x54\x04\x00\xBB\x99\x0D\x00\xBA\x91"
-     "\x0D\x00"
-     "\x86\x88\x04\x00\x87\xBC\x03\x00\xBF\xE1\x0D\x00\xBE\xF1\x0D\x00\xBD\xF9"
-     "\x0D\x00"
-     "\xBC\x81\x0D\x00\x82\x49\x00\x00\xA3\xC5\x0D\x00\x80\xF5\x03\x00\x81\x49"
-     "\x00\x00"
-     "\xA6\xDD\x0D\x00\xCF\xA1\x00\x80\xD3\xA1\x00\x80\xA5\xD5\x0D\x00\xAA\xD5"
-     "\x0D\x00"
-     "\xAB\xDD\x0D\x00\xD7\xA1\x00\x80\xDB\xA1\x00\x80\xAE\xB5\x0D\x00\xAF\xA5"
-     "\x0D\x00"
-     "\xAC\xC5\x0D\x00\xAD\xBD\x0D\x00\xA8\x1D\x02\x00\xA9\x45\x02\x00\xAA\x5D"
-     "\x02\x00"
-     "\xAB\x55\x02\x00\xAC\x79\x02\x00\xAD\x79\x02\x00\xAE\x89\x03\x00\xAF\x89"
-     "\x03\x00"
-     "\xDF\xA1\x00\x80\xE3\xA1\x00\x80\xE7\xA1\x00\x80\xEB\xA1\x00\x80\x84\xFC"
-     "\x05\x00"
-     "\xEF\xA1\x00\x80\xF3\xA1\x00\x80\xF7\xA1\x00\x80\xB8\x89\x03\x00\xB9\x65"
-     "\x03\x00"
-     "\xBA\x6D\x03\x00\xBB\x65\x03\x00\xBC\x7D\x03\x00\xBD\x65\x03\x00\xBE\x6D"
-     "\x03\x00"
-     "\xBF\x65\x03\x00\xB0\xC1\x03\x00\xB1\xC1\x03\x00\xB2\xC1\x03\x00\xB3\xC1"
-     "\x03\x00"
-     "\xB4\xC1\x03\x00\xB5\xC1\x03\x00\xB6\xC1\x03\x00\xB7\xC1\x03\x00\xFB\xA1"
-     "\x00\x80"
-     "\xFF\xA1\x00\x80\x03\xA2\x00\x80\x07\xA2\x00\x80\x0B\xA2\x00\x80\xE1\xA4"
-     "\x01\x00"
-     "\x0F\xA2\x00\x80\xE3\xC0\x0E\x00\xBE\x68\x04\x00\x13\xA2\x00\x80\x17\xA2"
-     "\x00\x80"
-     "\xEF\x1C\x01\x00\x1B\xA2\x00\x80\x1F\xA2\x00\x80\x23\xA2\x00\x80\x27\xA2"
-     "\x00\x80"
-     "\xB3\x95\x03\x00\x2B\xA2\x00\x80\x2F\xA2\x00\x80\x37\xA2\x00\x80\x3B\xA2"
-     "\x00\x80"
-     "\xB6\xB9\x03\x00\xB5\xB1\x03\x00\x3F\xA2\x00\x80\xBB\x45\x02\x00\xBA\x45"
-     "\x02\x00"
-     "\x86\xA8\x04\x00\x87\xA4\x05\x00\xBF\x45\x02\x00\xBE\x45\x02\x00\xBD\x55"
-     "\x02\x00"
-     "\xBC\x55\x02\x00\xE1\xE0\x0E\x00\xE1\x2C\x0C\x00\xE3\x08\x0E\x00\xE3\xD4"
-     "\x0E\x00"
-     "\x80\xAD\x00\x00\x81\xD1\x00\x00\x82\xD1\x00\x00\x43\xA2\x00\x80\x47\xA2"
-     "\x00\x80"
-     "\x4B\xA2\x00\x80\x4F\xA2\x00\x80\x53\xA2\x00\x80\x57\xA2\x00\x80\x5B\xA2"
-     "\x00\x80"
-     "\xEF\x88\x0E\x00\xEF\x2C\x0E\x00\xA3\x15\x03\x00\x5F\xA2\x00\x80\x85\xC4"
-     "\x2B\x00"
-     "\x63\xA2\x00\x80\x67\xA2\x00\x80\xA6\x39\x03\x00\xA5\x31\x03\x00\x6B\xA2"
-     "\x00\x80"
-     "\xAB\xC5\x02\x00\xAA\xC5\x02\x00\x6F\xA2\x00\x80\x73\xA2\x00\x80\xAF\xC5"
-     "\x02\x00"
-     "\xAE\xC5\x02\x00\xAD\xD5\x02\x00\xAC\xD5\x02\x00\xA8\x0D\x06\x00\xA9\x15"
-     "\x06\x00"
-     "\xAA\x5D\x06\x00\xAB\x55\x06\x00\xAC\x79\x06\x00\xAD\x79\x06\x00\xAE\xB9"
-     "\x06\x00"
-     "\xAF\xB9\x06\x00\x33\xA2\x00\x80\x77\xA2\x00\x80\x7B\xA2\x00\x80\x7F\xA2"
-     "\x00\x80"
-     "\x83\xA2\x00\x80\x87\xA2\x00\x80\x8B\xA2\x00\x80\x8F\xA2\x00\x80\xB8\x4D"
-     "\x07\x00"
-     "\xB9\x55\x07\x00\xBA\x51\x07\x00\xBB\x69\x07\x00\xBC\x79\x07\x00\xBD\x65"
-     "\x07\x00"
-     "\xBE\x6D\x07\x00\xBF\x65\x07\x00\xB0\xC9\x06\x00\xB1\xC9\x06\x00\xB2\xDD"
-     "\x06\x00"
-     "\xB3\xD5\x06\x00\xB4\xCD\x06\x00\xB5\x75\x07\x00\xB6\x7D\x07\x00\xB7\x75"
-     "\x07\x00"
-     "\xB3\xD5\x06\x00\x93\xA2\x00\x80\x97\xA2\x00\x80\x9B\xA2\x00\x80\x9F\xA2"
-     "\x00\x80"
-     "\xB6\xF9\x06\x00\xB5\xF1\x06\x00\xA3\xA2\x00\x80\xBB\x0D\x06\x00\xBA\x0D"
-     "\x06\x00"
-     "\x86\x08\x00\x00\x87\x2C\x00\x00\xBF\xB1\x07\x00\xBE\x09\x06\x00\xBD\x01"
-     "\x06\x00"
-     "\xBC\x09\x06\x00\x82\x51\x00\x00\xA3\x91\x06\x00\x80\x41\x00\x00\x81\x41"
-     "\x00\x00"
-     "\xA6\xBD\x06\x00\xA7\xA2\x00\x80\xAB\xA2\x00\x80\xA5\xB5\x06\x00\xAA\x49"
-     "\x06\x00"
-     "\xAB\x49\x06\x00\xAF\xA2\x00\x80\xB3\xA2\x00\x80\xAE\x4D\x06\x00\xAF\xF5"
-     "\x07\x00"
-     "\xAC\x4D\x06\x00\xAD\x45\x06\x00\xB0\xB1\x06\x00\xB1\xB1\x06\x00\xB2\xCD"
-     "\x06\x00"
-     "\xB3\xC1\x06\x00\xB4\xC1\x06\x00\xB5\xC9\x06\x00\xB6\xF1\x06\x00\xB7\xF1"
-     "\x06\x00"
-     "\xB8\x05\x01\x00\xB9\x0D\x01\x00\xBA\x05\x01\x00\xBB\x1D\x01\x00\xBC\x05"
-     "\x01\x00"
-     "\xBD\x0D\x01\x00\xBE\x05\x01\x00\xBF\xB9\x01\x00\xB7\xA2\x00\x80\xBB\xA2"
-     "\x00\x80"
-     "\xBF\xA2\x00\x80\xC3\xA2\x00\x80\xC7\xA2\x00\x80\xCB\xA2\x00\x80\xC3\xA1"
-     "\x00\x80"
-     "\xCF\xA2\x00\x80\xA8\x2D\x06\x00\xA9\x35\x06\x00\xAA\x35\x06\x00\xAB\xF1"
-     "\x06\x00"
-     "\xAC\xD1\x06\x00\xAD\xD1\x06\x00\xAE\xD1\x06\x00\xAF\xD1\x06\x00\xB3\xDD"
-     "\x06\x00"
-     "\xD3\xA2\x00\x80\xD7\xA2\x00\x80\xDB\xA2\x00\x80\xDF\xA2\x00\x80\xB6\x31"
-     "\x06\x00"
-     "\xB5\x39\x06\x00\xE3\xA2\x00\x80\xBB\x15\x06\x00\xBA\x15\x06\x00\xE7\xA2"
-     "\x00\x80"
-     "\xEB\xA2\x00\x80\xBF\x6D\x06\x00\xBE\x65\x06\x00\xBD\x75\x06\x00\xBC\x79"
-     "\x06\x00"
-     "\xEF\xA2\x00\x80\xA3\x99\x06\x00\xF3\xA2\x00\x80\xF7\xA2\x00\x80\xA6\x75"
-     "\x06\x00"
-     "\xFB\xA2\x00\x80\xFF\xA2\x00\x80\xA5\x7D\x06\x00\xAA\x51\x06\x00\xAB\x51"
-     "\x06\x00"
-     "\x03\xA3\x00\x80\x07\xA3\x00\x80\xAE\x21\x06\x00\xAF\x29\x06\x00\xAC\x3D"
-     "\x06\x00"
-     "\xAD\x31\x06\x00\xA8\xD5\x02\x00\xA9\xDD\x02\x00\xAA\xE1\x02\x00\xAB\xE1"
-     "\x02\x00"
-     "\xAC\x51\x03\x00\xAD\x51\x03\x00\xAE\x51\x03\x00\xAF\x51\x03\x00\x0B\xA3"
-     "\x00\x80"
-     "\x0F\xA3\x00\x80\xBE\xEC\x03\x00\x17\xA3\x00\x80\x1B\xA3\x00\x80\x1F\xA3"
-     "\x00\x80"
-     "\x23\xA3\x00\x80\x27\xA3\x00\x80\xB8\xE9\x03\x00\xB9\xE9\x03\x00\xBA\x89"
-     "\x03\x00"
-     "\xBB\x85\x03\x00\xBC\x9D\x03\x00\xBD\x81\x03\x00\xBE\x81\x03\x00\xBF\xB5"
-     "\x03\x00"
-     "\xB0\x31\x03\x00\xB1\x31\x03\x00\xB2\x35\x03\x00\xB3\xE9\x03\x00\xB4\xF9"
-     "\x03\x00"
-     "\xB5\xF9\x03\x00\xB6\xE9\x03\x00\xB7\xE1\x03\x00\x80\x6D\x03\x00\x81\xA5"
-     "\x00\x00"
-     "\x82\xAD\x00\x00\xB3\x65\x02\x00\x2B\xA3\x00\x80\xB5\xD5\x03\x00\xB6\xDD"
-     "\x03\x00"
-     "\x2F\xA3\x00\x80\x84\xE0\x02\x00\x33\xA3\x00\x80\xBA\xF9\x03\x00\xBB\xF9"
-     "\x03\x00"
-     "\xBC\xED\x03\x00\xBD\x31\x03\x00\xBE\x31\x03\x00\xBF\x31\x03\x00\x87\xEC"
-     "\x03\x00"
-     "\x86\x64\x3C\x00\xB2\x00\x00\x00\x37\xA3\x00\x80\x3B\xA3\x00\x80\xE3\x08"
-     "\x04\x00"
-     "\x3F\xA3\x00\x80\xE1\xEC\x06\x00\x43\xA3\x00\x80\xEF\x00\x06\x00\x47\xA3"
-     "\x00\x80"
-     "\x4B\xA3\x00\x80\x4F\xA3\x00\x80\x53\xA3\x00\x80\x57\xA3\x00\x80\x5B\xA3"
-     "\x00\x80"
-     "\x5F\xA3\x00\x80\x63\xA3\x00\x80\x67\xA3\x00\x80\x6B\xA3\x00\x80\xE3\xA0"
-     "\x03\x00"
-     "\x6F\xA3\x00\x80\xE1\xA8\x01\x00\x73\xA3\x00\x80\xEF\xF0\x03\x00\x82\x1D"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x80\x1D\x00\x00\x77\xA3\x00\x80\x7B\xA3\x00\x80\x7F\xA3"
-     "\x00\x80"
-     "\x87\xA3\x00\x80\xBE\x4C\x3D\x00\x8B\xA3\x00\x80\xA3\xA1\x03\x00\xBE\x40"
-     "\x3C\x00"
-     "\xA5\x11\x02\x00\x8F\xA3\x00\x80\x93\xA3\x00\x80\xA6\x19\x02\x00\x84\x6C"
-     "\x02\x00"
-     "\x97\xA3\x00\x80\xAB\x3D\x02\x00\xAA\x3D\x02\x00\xAD\xF5\x02\x00\xAC\x29"
-     "\x02\x00"
-     "\xAF\xF5\x02\x00\xAE\xF5\x02\x00\x86\x40\x3C\x00\x87\xB4\x3D\x00\x9B\xA3"
-     "\x00\x80"
-     "\x9F\xA3\x00\x80\xA3\xA3\x00\x80\xA7\xA3\x00\x80\xAB\xA3\x00\x80\xEF\x44"
-     "\x06\x00"
-     "\xAF\xA3\x00\x80\xE1\xD4\x06\x00\xB3\xA3\x00\x80\xE3\x0C\x07\x00\xB7\xA3"
-     "\x00\x80"
-     "\xBB\xA3\x00\x80\xBF\xA3\x00\x80\xC3\xA3\x00\x80\xB3\xF5\x01\x00\xC7\xA3"
-     "\x00\x80"
-     "\xCB\xA3\x00\x80\xCF\xA3\x00\x80\xD3\xA3\x00\x80\xB6\x45\x01\x00\xB5\xE5"
-     "\x01\x00"
-     "\xD7\xA3\x00\x80\xBB\x31\x01\x00\xBA\x29\x01\x00\xDB\xA3\x00\x80\xDF\xA3"
-     "\x00\x80"
-     "\xBF\x1D\x01\x00\xBE\x1D\x01\x00\xBD\x19\x01\x00\xBC\x25\x01\x00\xA8\x2D"
-     "\x3E\x00"
-     "\xA9\x35\x3E\x00\xAA\x3D\x3E\x00\xAB\x35\x3E\x00\xAC\x2D\x3E\x00\xAD\x85"
-     "\x3E\x00"
-     "\xAE\x85\x3E\x00\xAF\xB9\x3E\x00\x83\xA3\x00\x80\xE3\xA3\x00\x80\xE7\xA3"
-     "\x00\x80"
-     "\xEB\xA3\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\xEF\xA3"
-     "\x00\x80"
-     "\xB8\xB9\x3E\x00\xB9\xB9\x3E\x00\xBA\x49\x3F\x00\xBB\x49\x3F\x00\xBC\x59"
-     "\x3F\x00"
-     "\xBD\x59\x3F\x00\xBE\x4D\x3F\x00\xBF\x41\x3F\x00\xB0\xAD\x3E\x00\xB1\xB5"
-     "\x3E\x00"
-     "\xB2\xB1\x3E\x00\xB3\x8D\x3E\x00\xB4\x99\x3E\x00\xB5\x99\x3E\x00\xB6\x89"
-     "\x3E\x00"
-     "\xB7\x89\x3E\x00\xA3\xB5\x3E\x00\x84\x8C\x02\x00\xF3\xA3\x00\x80\xF7\xA3"
-     "\x00\x80"
-     "\xFB\xA3\x00\x80\xA6\x05\x3E\x00\xA5\xA5\x3E\x00\xFF\xA3\x00\x80\xAB\x71"
-     "\x3E\x00"
-     "\xAA\x69\x3E\x00\x86\x08\x00\x00\x87\x68\x03\x00\xAF\x5D\x3E\x00\xAE\x5D"
-     "\x3E\x00"
-     "\xAD\x59\x3E\x00\xAC\x65\x3E\x00\x03\xA4\x00\x80\xB3\x91\x3F\x00\x07\xA4"
-     "\x00\x80"
-     "\x0B\xA4\x00\x80\xB6\x59\x3F\x00\x0F\xA4\x00\x80\x13\xA4\x00\x80\xB5\xB9"
-     "\x3F\x00"
-     "\xBA\x75\x3F\x00\xBB\x7D\x3F\x00\x17\xA4\x00\x80\x1B\xA4\x00\x80\xBE\x41"
-     "\x3F\x00"
-     "\xBF\x41\x3F\x00\xBC\x59\x3F\x00\xBD\x55\x3F\x00\xB0\x95\x3E\x00\xB1\x9D"
-     "\x3E\x00"
-     "\xB2\xA9\x3E\x00\xB3\xA5\x3E\x00\xB4\xA1\x3E\x00\xB5\xA1\x3E\x00\xB6\xA1"
-     "\x3E\x00"
-     "\xB7\xA1\x3E\x00\xB8\xE5\x3E\x00\xB9\xED\x3E\x00\xBA\xE5\x3E\x00\xBB\xFD"
-     "\x3E\x00"
-     "\xBC\xED\x3E\x00\xBD\xDD\x3E\x00\xBE\xD5\x3E\x00\xBF\x19\x01\x00\x1F\xA4"
-     "\x00\x80"
-     "\x23\xA4\x00\x80\x27\xA4\x00\x80\x2B\xA4\x00\x80\x2F\xA4\x00\x80\x13\xA3"
-     "\x00\x80"
-     "\x33\xA4\x00\x80\x37\xA4\x00\x80\xA8\x5D\x3E\x00\xA9\x21\x3E\x00\xAA\x3D"
-     "\x3E\x00"
-     "\xAB\x35\x3E\x00\xAC\x15\x3E\x00\xAD\xFD\x3E\x00\xAE\xF5\x3E\x00\xAF\xED"
-     "\x3E\x00"
-     "\xA3\xD5\x3E\x00\x3B\xA4\x00\x80\x3F\xA4\x00\x80\x43\xA4\x00\x80\x47\xA4"
-     "\x00\x80"
-     "\xA6\x1D\x3E\x00\xA5\xFD\x3E\x00\x4B\xA4\x00\x80\xAB\x39\x3E\x00\xAA\x31"
-     "\x3E\x00"
-     "\x4F\xA4\x00\x80\x53\xA4\x00\x80\xAF\x05\x3E\x00\xAE\x05\x3E\x00\xAD\x11"
-     "\x3E\x00"
-     "\xAC\x1D\x3E\x00\x81\x11\x00\x00\x80\x0D\x00\x00\x57\xA4\x00\x80\x82\x19"
-     "\x00\x00"
-     "\x5B\xA4\x00\x80\x5F\xA4\x00\x80\x84\x94\x01\x00\xBE\x10\x00\x00\x86\x40"
-     "\x07\x00"
-     "\x87\x00\x01\x00\x67\xA4\x00\x80\x6B\xA4\x00\x80\x6F\xA4\x00\x80\x73\xA4"
-     "\x00\x80"
-     "\x77\xA4\x00\x80\x7B\xA4\x00\x80\xA8\x8D\x02\x00\xA9\x95\x02\x00\xAA\x95"
-     "\x02\x00"
-     "\xAB\xCD\x02\x00\xAC\xD9\x02\x00\xAD\xD9\x02\x00\xAE\xCD\x02\x00\xAF\xC5"
-     "\x02\x00"
-     "\x7F\xA4\x00\x80\x83\xA4\x00\x80\x87\xA4\x00\x80\x8B\xA4\x00\x80\x8C\x00"
-     "\x00\x00"
-     "\x8F\xA4\x00\x80\x93\xA4\x00\x80\x97\xA4\x00\x80\xB8\x1D\x03\x00\xB9\xC1"
-     "\x03\x00"
-     "\xBA\xC1\x03\x00\xBB\xC1\x03\x00\xBC\xC1\x03\x00\xBD\xC9\x03\x00\xBE\xF1"
-     "\x03\x00"
-     "\xBF\xF1\x03\x00\xB0\x89\x02\x00\xB1\x89\x02\x00\xB2\x29\x03\x00\xB3\x29"
-     "\x03\x00"
-     "\xB4\x39\x03\x00\xB5\x39\x03\x00\xB6\x29\x03\x00\xB7\x25\x03\x00\xB3\x1D"
-     "\x02\x00"
-     "\x9B\xA4\x00\x80\x9F\xA4\x00\x80\xA3\xA4\x00\x80\xA7\xA4\x00\x80\xB6\x59"
-     "\x02\x00"
-     "\xB5\x51\x02\x00\xAB\xA4\x00\x80\xBB\x4D\x02\x00\xBA\x4D\x02\x00\xAF\xA4"
-     "\x00\x80"
-     "\xB3\xA4\x00\x80\xBF\xFD\x03\x00\xBE\xFD\x03\x00\xBD\xFD\x03\x00\xBC\xFD"
-     "\x03\x00"
-     "\xB7\xA4\x00\x80\xBB\xA4\x00\x80\xBF\xA4\x00\x80\xC3\xA4\x00\x80\xE1\x0C"
-     "\x3E\x00"
-     "\xC7\xA4\x00\x80\xE3\xA8\x3F\x00\xCB\xA4\x00\x80\x81\x3D\x00\x00\x80\x31"
-     "\x00\x00"
-     "\xEF\x50\x3F\x00\x82\x1D\x00\x00\xCF\xA4\x00\x80\xBE\xF8\x04\x00\x86\x18"
-     "\x05\x00"
-     "\x87\x4C\x03\x00\x84\x0C\x02\x00\xE3\xCC\x00\x00\xD7\xA4\x00\x80\xE1\xBC"
-     "\x01\x00"
-     "\xDB\xA4\x00\x80\xDF\xA4\x00\x80\xE3\xA4\x00\x80\xE7\xA4\x00\x80\xEB\xA4"
-     "\x00\x80"
-     "\x84\x2C\x05\x00\xEF\xA4\x00\x80\xF3\xA4\x00\x80\xF7\xA4\x00\x80\xEF\x10"
-     "\x00\x00"
-     "\xFB\xA4\x00\x80\xFF\xA4\x00\x80\xA3\xDD\x03\x00\x03\xA5\x00\x80\x07\xA5"
-     "\x00\x80"
-     "\x0B\xA5\x00\x80\x0F\xA5\x00\x80\xA6\x99\x03\x00\xA5\x91\x03\x00\x13\xA5"
-     "\x00\x80"
-     "\xAB\x8D\x03\x00\xAA\x8D\x03\x00\x17\xA5\x00\x80\x1B\xA5\x00\x80\xAF\x3D"
-     "\x02\x00"
-     "\xAE\x3D\x02\x00\xAD\x3D\x02\x00\xAC\x3D\x02\x00\x1F\xA5\x00\x80\x23\xA5"
-     "\x00\x80"
-     "\x27\xA5\x00\x80\x2B\xA5\x00\x80\x2F\xA5\x00\x80\x33\xA5\x00\x80\x37\xA5"
-     "\x00\x80"
-     "\xEF\x28\x3E\x00\x3B\xA5\x00\x80\xE1\x3C\x3E\x00\x3F\xA5\x00\x80\xE3\x38"
-     "\x01\x00"
-     "\x80\x29\x00\x00\x81\x15\x00\x00\x82\x11\x00\x00\x47\xA5\x00\x80\xB3\x31"
-     "\x02\x00"
-     "\xBE\xC8\x04\x00\x84\xC0\x04\x00\x4B\xA5\x00\x80\x4F\xA5\x00\x80\xB6\x29"
-     "\x02\x00"
-     "\xB5\x21\x02\x00\x53\xA5\x00\x80\xBB\xCD\x01\x00\xBA\xCD\x01\x00\x57\xA5"
-     "\x00\x80"
-     "\x5B\xA5\x00\x80\xBF\x75\x01\x00\xBE\xC9\x01\x00\xBD\xC1\x01\x00\xBC\xC9"
-     "\x01\x00"
-     "\xA8\xE9\x05\x00\xA9\xE9\x05\x00\xAA\xF9\x05\x00\xAB\xF9\x05\x00\xAC\xE9"
-     "\x05\x00"
-     "\xAD\xE9\x05\x00\xAE\x39\x06\x00\xAF\x39\x06\x00\xD3\xA4\x00\x80\x43\xA5"
-     "\x00\x80"
-     "\x86\x88\x00\x00\x87\x40\x03\x00\x5F\xA5\x00\x80\x63\xA5\x00\x80\x67\xA5"
-     "\x00\x80"
-     "\x6B\xA5\x00\x80\xB8\xD1\x06\x00\xB9\xD9\x06\x00\xBA\xE1\x06\x00\xBB\xE1"
-     "\x06\x00"
-     "\xBC\x91\x06\x00\xBD\x91\x06\x00\xBE\x91\x06\x00\xBF\x91\x06\x00\xB0\x49"
-     "\x06\x00"
-     "\xB1\x49\x06\x00\xB2\x5D\x06\x00\xB3\x55\x06\x00\xB4\x4D\x06\x00\xB5\xF1"
-     "\x06\x00"
-     "\xB6\xF1\x06\x00\xB7\xF1\x06\x00\xA3\x71\x05\x00\x6F\xA5\x00\x80\x73\xA5"
-     "\x00\x80"
-     "\x77\xA5\x00\x80\x7B\xA5\x00\x80\xA6\x69\x05\x00\xA5\x61\x05\x00\x7F\xA5"
-     "\x00\x80"
-     "\xAB\x8D\x06\x00\xAA\x8D\x06\x00\x83\xA5\x00\x80\x87\xA5\x00\x80\xAF\x35"
-     "\x06\x00"
-     "\xAE\x89\x06\x00\xAD\x81\x06\x00\xAC\x89\x06\x00\x8B\xA5\x00\x80\xB3\xD1"
-     "\x07\x00"
-     "\x8F\xA5\x00\x80\x93\xA5\x00\x80\xB6\xF1\x07\x00\x97\xA5\x00\x80\x9B\xA5"
-     "\x00\x80"
-     "\xB5\xC1\x07\x00\xBA\xD1\x07\x00\xBB\xDD\x07\x00\x9F\xA5\x00\x80\xA3\xA5"
-     "\x00\x80"
-     "\xBE\xB9\x07\x00\xBF\xB9\x07\x00\xBC\xC5\x07\x00\xBD\xB9\x07\x00\xB8\x69"
-     "\x06\x00"
-     "\xB9\x69\x06\x00\xBA\x89\x06\x00\xBB\x89\x06\x00\xBC\x99\x06\x00\xBD\x99"
-     "\x06\x00"
-     "\xBE\x89\x06\x00\xBF\x89\x06\x00\xB0\x11\x06\x00\xB1\x1D\x06\x00\xB2\x15"
-     "\x06\x00"
-     "\xB3\x69\x06\x00\xB4\x79\x06\x00\xB5\x79\x06\x00\xB6\x69\x06\x00\xB7\x61"
-     "\x06\x00"
-     "\xA8\x49\x06\x00\xA9\x55\x06\x00\xAA\x5D\x06\x00\xAB\x55\x06\x00\xAC\x4D"
-     "\x06\x00"
-     "\xAD\x75\x06\x00\xAE\x71\x06\x00\xAF\x71\x06\x00\x63\xA4\x00\x80\x82\x1D"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x80\x1D\x00\x00\xA7\xA5\x00\x80\xAB\xA5\x00\x80\xAF\xA5"
-     "\x00\x80"
-     "\xBE\x70\x01\x00\xA3\x95\x06\x00\xB7\xA5\x00\x80\x86\x28\x00\x00\x87\x48"
-     "\x01\x00"
-     "\xBB\xA5\x00\x80\xA6\xB5\x06\x00\xA5\x85\x06\x00\xBF\xA5\x00\x80\xAB\x99"
-     "\x06\x00"
-     "\xAA\x95\x06\x00\xC3\xA5\x00\x80\xC7\xA5\x00\x80\xAF\xFD\x06\x00\xAE\xFD"
-     "\x06\x00"
-     "\xAD\xFD\x06\x00\xAC\x81\x06\x00\xCB\xA5\x00\x80\xB3\x15\x06\x00\xCF\xA5"
-     "\x00\x80"
-     "\xD3\xA5\x00\x80\xB6\x3D\x06\x00\xD7\xA5\x00\x80\xDB\xA5\x00\x80\xB5\x35"
-     "\x06\x00"
-     "\xBA\xD9\x01\x00\xBB\xD9\x01\x00\xDF\xA5\x00\x80\xE3\xA5\x00\x80\xBE\x7D"
-     "\x01\x00"
-     "\xBF\x65\x01\x00\xBC\x7D\x01\x00\xBD\x75\x01\x00\xA8\xBD\x05\x00\xA9\xC9"
-     "\x05\x00"
-     "\xAA\xD9\x05\x00\xAB\xD1\x05\x00\xAC\xF9\x05\x00\xAD\xF9\x05\x00\xAE\x29"
-     "\x02\x00"
-     "\xAF\x29\x02\x00\xE7\xA5\x00\x80\xEB\xA5\x00\x80\xEF\xA5\x00\x80\xF3\xA5"
-     "\x00\x80"
-     "\x8C\x00\x00\x00\xF7\xA5\x00\x80\xFB\xA5\x00\x80\xFF\xA5\x00\x80\xB8\xED"
-     "\x02\x00"
-     "\xB9\x85\x02\x00\xBA\x81\x02\x00\xBB\x81\x02\x00\xBC\x85\x02\x00\xBD\x8D"
-     "\x02\x00"
-     "\xBE\xB1\x02\x00\xBF\xB1\x02\x00\xB0\x59\x02\x00\xB1\x59\x02\x00\xB2\xED"
-     "\x02\x00"
-     "\xB3\xE5\x02\x00\xB4\xFD\x02\x00\xB5\xE5\x02\x00\xB6\xE5\x02\x00\xB7\xD5"
-     "\x02\x00"
-     "\xA3\x51\x05\x00\x03\xA6\x00\x80\x07\xA6\x00\x80\x0B\xA6\x00\x80\x0F\xA6"
-     "\x00\x80"
-     "\xA6\x79\x05\x00\xA5\x71\x05\x00\x13\xA6\x00\x80\xAB\x9D\x02\x00\xAA\x9D"
-     "\x02\x00"
-     "\x17\xA6\x00\x80\x1B\xA6\x00\x80\xAF\x21\x02\x00\xAE\x39\x02\x00\xAD\x31"
-     "\x02\x00"
-     "\xAC\x39\x02\x00\x81\x6D\x00\x00\x80\x6D\x00\x00\x1F\xA6\x00\x80\x82\x05"
-     "\x00\x00"
-     "\xBE\x5C\x0C\x00\x27\xA6\x00\x80\x2B\xA6\x00\x80\xEF\xD0\x06\x00\x84\xEC"
-     "\x03\x00"
-     "\xE1\x1C\x05\x00\x2F\xA6\x00\x80\xE3\xFC\x07\x00\x33\xA6\x00\x80\x37\xA6"
-     "\x00\x80"
-     "\x86\xD8\x0C\x00\x87\xBC\x0C\x00\xA8\x85\x02\x00\xA9\x95\x02\x00\xAA\x95"
-     "\x02\x00"
-     "\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xD5\x02\x00\xAE\xD1\x02\x00\xAF\xD1"
-     "\x02\x00"
-     "\x3B\xA6\x00\x80\x3F\xA6\x00\x80\x43\xA6\x00\x80\x47\xA6\x00\x80\x4B\xA6"
-     "\x00\x80"
-     "\x4F\xA6\x00\x80\x53\xA6\x00\x80\x57\xA6\x00\x80\xB8\x75\x01\x00\xB9\x7D"
-     "\x01\x00"
-     "\xBA\x75\x01\x00\xBB\xC9\x01\x00\xBC\xD9\x01\x00\xBD\xD9\x01\x00\xBE\xC9"
-     "\x01\x00"
-     "\xBF\xC1\x01\x00\xB0\xB5\x02\x00\xB1\xBD\x02\x00\xB2\x81\x02\x00\xB3\x81"
-     "\x02\x00"
-     "\xB4\x51\x01\x00\xB5\x51\x01\x00\xB6\x51\x01\x00\xB7\x51\x01\x00\x5B\xA6"
-     "\x00\x80"
-     "\x84\x04\x0D\x00\x5F\xA6\x00\x80\x63\xA6\x00\x80\xBE\x1C\x0C\x00\x67\xA6"
-     "\x00\x80"
-     "\xEF\x1C\x00\x00\xEF\xA0\x06\x00\xE1\x90\x01\x00\xE1\x44\x07\x00\xE3\x70"
-     "\x06\x00"
-     "\xE3\x98\x06\x00\x6B\xA6\x00\x80\x6F\xA6\x00\x80\x73\xA6\x00\x80\x77\xA6"
-     "\x00\x80"
-     "\xB3\x5D\x02\x00\x7B\xA6\x00\x80\x7F\xA6\x00\x80\x83\xA6\x00\x80\x87\xA6"
-     "\x00\x80"
-     "\xB6\x15\x02\x00\xB5\x75\x02\x00\x8B\xA6\x00\x80\xBB\x39\x02\x00\xBA\x31"
-     "\x02\x00"
-     "\x8F\xA6\x00\x80\x97\xA6\x00\x80\xBF\xD5\x01\x00\xBE\xD5\x01\x00\xBD\x15"
-     "\x02\x00"
-     "\xBC\x15\x02\x00\xA3\x9D\x0D\x00\x23\xA6\x00\x80\x93\xA6\x00\x80\x9B\xA6"
-     "\x00\x80"
-     "\x9F\xA6\x00\x80\xA6\xD5\x0D\x00\xA5\xB5\x0D\x00\xA3\xA6\x00\x80\xAB\xF9"
-     "\x0D\x00"
-     "\xAA\xF1\x0D\x00\x86\x08\x03\x00\x87\x60\x03\x00\xAF\x15\x0E\x00\xAE\x15"
-     "\x0E\x00"
-     "\xAD\xD5\x0D\x00\xAC\xD5\x0D\x00\x80\x91\x0F\x00\x81\x99\x0F\x00\x82\xA1"
-     "\x0F\x00"
-     "\xB3\xA5\x0E\x00\xA7\xA6\x00\x80\xB5\xA1\x0E\x00\xB6\x79\x0F\x00\xAB\xA6"
-     "\x00\x80"
-     "\xAF\xA6\x00\x80\xB3\xA6\x00\x80\xBA\x45\x0F\x00\xBB\x5D\x0F\x00\xBC\x45"
-     "\x0F\x00"
-     "\xBD\x4D\x0F\x00\xBE\x45\x0F\x00\xBF\xFD\x0F\x00\xA8\x55\x0E\x00\xA9\x5D"
-     "\x0E\x00"
-     "\xAA\x61\x0E\x00\xAB\x7D\x0E\x00\xAC\x65\x0E\x00\xAD\xB5\x0F\x00\xAE\xBD"
-     "\x0F\x00"
-     "\xAF\xB5\x0F\x00\xB7\xA6\x00\x80\xBB\xA6\x00\x80\xBF\xA6\x00\x80\xC3\xA6"
-     "\x00\x80"
-     "\xC7\xA6\x00\x80\xCB\xA6\x00\x80\xCF\xA6\x00\x80\xD3\xA6\x00\x80\xB8\x55"
-     "\x0F\x00"
-     "\xB9\x75\x0F\x00\xBA\x7D\x0F\x00\xBB\x75\x0F\x00\xBC\x6D\x0F\x00\xBD\x11"
-     "\x0F\x00"
-     "\xBE\x11\x0F\x00\xBF\x11\x0F\x00\xB0\xCD\x0F\x00\xB1\xD5\x0F\x00\xB2\xDD"
-     "\x0F\x00"
-     "\xB3\xD5\x0F\x00\xB4\xCD\x0F\x00\xB5\x75\x0F\x00\xB6\x71\x0F\x00\xB7\x71"
-     "\x0F\x00"
-     "\xA3\xE9\x0F\x00\xD7\xA6\x00\x80\xDB\xA6\x00\x80\xDF\xA6\x00\x80\xE3\xA6"
-     "\x00\x80"
-     "\xA6\x35\x0E\x00\xA5\xED\x0F\x00\xE7\xA6\x00\x80\xAB\x11\x0E\x00\xAA\x09"
-     "\x0E\x00"
-     "\xEB\xA6\x00\x80\xEF\xA6\x00\x80\xAF\xB1\x0E\x00\xAE\x09\x0E\x00\xAD\x01"
-     "\x0E\x00"
-     "\xAC\x09\x0E\x00\xF3\xA6\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xF7\xA6\x00\x80\xFB\xA6\x00\x80\xFF\xA6\x00\x80\xBE\x94\x01\x00\x84\xAC"
-     "\x01\x00"
-     "\x03\xA7\x00\x80\x87\xE0\x01\x00\x86\xCC\x00\x00\x07\xA7\x00\x80\x0B\xA7"
-     "\x00\x80"
-     "\x0F\xA7\x00\x80\xB3\xA5\x00\x80\xA8\x6D\x0E\x00\xA9\x89\x01\x00\xAA\x99"
-     "\x01\x00"
-     "\xAB\x91\x01\x00\xAC\xC1\x01\x00\xAD\xC9\x01\x00\xAE\xF9\x01\x00\xAF\xF9"
-     "\x01\x00"
-     "\x84\xA0\x00\x00\x13\xA7\x00\x80\x17\xA7\x00\x80\x1B\xA7\x00\x80\x1F\xA7"
-     "\x00\x80"
-     "\x23\xA7\x00\x80\x27\xA7\x00\x80\x2B\xA7\x00\x80\xB8\xC5\x00\x00\xB9\xCD"
-     "\x00\x00"
-     "\xBA\xC5\x00\x00\xBB\xDD\x00\x00\xBC\xCD\x00\x00\xBD\xFD\x00\x00\xBE\xF5"
-     "\x00\x00"
-     "\xBF\x9D\x00\x00\xB0\x41\x01\x00\xB1\x41\x01\x00\xB2\x41\x01\x00\xB3\x41"
-     "\x01\x00"
-     "\xB4\x41\x01\x00\xB5\x41\x01\x00\xB6\x41\x01\x00\xB7\x41\x01\x00\xB3\x11"
-     "\x02\x00"
-     "\x2F\xA7\x00\x80\x33\xA7\x00\x80\x37\xA7\x00\x80\x3B\xA7\x00\x80\xB6\x49"
-     "\x02\x00"
-     "\xB5\x49\x02\x00\x3F\xA7\x00\x80\xBB\x85\x02\x00\xBA\x85\x02\x00\x43\xA7"
-     "\x00\x80"
-     "\x47\xA7\x00\x80\xBF\x85\x02\x00\xBE\x85\x02\x00\xBD\x95\x02\x00\xBC\x95"
-     "\x02\x00"
-     "\x85\x3C\x1A\x00\xA3\x55\x02\x00\x4B\xA7\x00\x80\x4F\xA7\x00\x80\xA6\x0D"
-     "\x02\x00"
-     "\x53\xA7\x00\x80\x57\xA7\x00\x80\xA5\x0D\x02\x00\xAA\xC1\x02\x00\xAB\xC1"
-     "\x02\x00"
-     "\x5B\xA7\x00\x80\x5F\xA7\x00\x80\xAE\xC1\x02\x00\xAF\xC1\x02\x00\xAC\xD1"
-     "\x02\x00"
-     "\xAD\xD1\x02\x00\x82\x19\x00\x00\x63\xA7\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x67\xA7\x00\x80\x6B\xA7\x00\x80\x6F\xA7\x00\x80\x77\xA7\x00\x80\xBE\x00"
-     "\x04\x00"
-     "\x7B\xA7\x00\x80\x7F\xA7\x00\x80\x83\xA7\x00\x80\x87\xA7\x00\x80\x8B\xA7"
-     "\x00\x80"
-     "\x8F\xA7\x00\x80\x93\xA7\x00\x80\xEF\xE8\x0E\x00\x97\xA7\x00\x80\x86\xE8"
-     "\x04\x00"
-     "\x87\x50\x03\x00\x9B\xA7\x00\x80\x9F\xA7\x00\x80\xEF\x80\x02\x00\xA3\xA7"
-     "\x00\x80"
-     "\xE1\x94\x01\x00\xA7\xA7\x00\x80\xE3\x58\x01\x00\xAB\xA7\x00\x80\xE3\x00"
-     "\x0E\x00"
-     "\xAF\xA7\x00\x80\xE1\x68\x0D\x00\xB3\xA7\x00\x80\xA8\x71\x02\x00\xA9\x71"
-     "\x02\x00"
-     "\xAA\x71\x02\x00\xAB\xA9\x02\x00\xAC\xB9\x02\x00\xAD\xB9\x02\x00\xAE\xA9"
-     "\x02\x00"
-     "\xAF\xA9\x02\x00\x84\xAC\x05\x00\xB7\xA7\x00\x80\xBB\xA7\x00\x80\xBF\xA7"
-     "\x00\x80"
-     "\xC3\xA7\x00\x80\xC7\xA7\x00\x80\xCB\xA7\x00\x80\xCF\xA7\x00\x80\xB8\x6D"
-     "\x01\x00"
-     "\xB9\x0D\x01\x00\xBA\x05\x01\x00\xBB\x19\x01\x00\xBC\x09\x01\x00\xBD\x3D"
-     "\x01\x00"
-     "\xBE\x35\x01\x00\xBF\xD9\x01\x00\xB0\xD9\x02\x00\xB1\xD9\x02\x00\xB2\x6D"
-     "\x01\x00"
-     "\xB3\x65\x01\x00\xB4\x7D\x01\x00\xB5\x65\x01\x00\xB6\x65\x01\x00\xB7\x55"
-     "\x01\x00"
-     "\xE1\x60\x0F\x00\xE3\xF4\x00\x00\xE3\x1C\x0E\x00\xE1\xBC\x01\x00\xD3\xA7"
-     "\x00\x80"
-     "\x82\x39\x00\x00\x81\x31\x00\x00\x80\x3D\x00\x00\xD7\xA7\x00\x80\xBE\x28"
-     "\x04\x00"
-     "\xDB\xA7\x00\x80\xDF\xA7\x00\x80\xBE\x3C\x07\x00\xEF\x10\x00\x00\xEF\xD0"
-     "\x0E\x00"
-     "\xE7\xA7\x00\x80\x86\xE0\x04\x00\x87\xC8\x04\x00\xEB\xA7\x00\x80\xB3\xB5"
-     "\x02\x00"
-     "\xEF\xA7\x00\x80\xB5\x7D\x02\x00\xB6\x75\x02\x00\xF3\xA7\x00\x80\xF7\xA7"
-     "\x00\x80"
-     "\xFB\xA7\x00\x80\xBA\x51\x02\x00\xBB\x51\x02\x00\xBC\xF5\x01\x00\xBD\xFD"
-     "\x01\x00"
-     "\xBE\xF5\x01\x00\xBF\xED\x01\x00\x73\xA7\x00\x80\xE3\xA7\x00\x80\xAA\xB1"
-     "\x05\x00"
-     "\xAB\xB1\x05\x00\xAC\x15\x06\x00\xAD\x1D\x06\x00\xAE\x15\x06\x00\xAF\x0D"
-     "\x06\x00"
-     "\xFF\xA7\x00\x80\x03\xA8\x00\x80\x07\xA8\x00\x80\xA3\x55\x05\x00\x0B\xA8"
-     "\x00\x80"
-     "\xA5\x9D\x05\x00\xA6\x95\x05\x00\x0F\xA8\x00\x80\xB3\xE9\x06\x00\x13\xA8"
-     "\x00\x80"
-     "\x17\xA8\x00\x80\x1B\xA8\x00\x80\x1F\xA8\x00\x80\xB6\xE1\x06\x00\xB5\xE9"
-     "\x06\x00"
-     "\x23\xA8\x00\x80\xBB\xB1\x06\x00\xBA\xA1\x06\x00\x27\xA8\x00\x80\x2B\xA8"
-     "\x00\x80"
-     "\xBF\x9D\x06\x00\xBE\x91\x06\x00\xBD\xA5\x06\x00\xBC\xA9\x06\x00\xA8\x1D"
-     "\x06\x00"
-     "\xA9\x21\x06\x00\xAA\x21\x06\x00\xAB\x21\x06\x00\xAC\x21\x06\x00\xAD\x21"
-     "\x06\x00"
-     "\xAE\x21\x06\x00\xAF\x21\x06\x00\x2F\xA8\x00\x80\x33\xA8\x00\x80\x37\xA8"
-     "\x00\x80"
-     "\x3B\xA8\x00\x80\x3F\xA8\x00\x80\x43\xA8\x00\x80\x47\xA8\x00\x80\x4B\xA8"
-     "\x00\x80"
-     "\xB8\xE5\x07\x00\xB9\xED\x07\x00\xBA\xE5\x07\x00\xBB\xFD\x07\x00\xBC\xE5"
-     "\x07\x00"
-     "\xBD\xED\x07\x00\xBE\xE5\x07\x00\xBF\x4D\x07\x00\xB0\x25\x06\x00\xB1\x35"
-     "\x06\x00"
-     "\xB2\x3D\x06\x00\xB3\x31\x06\x00\xB4\x15\x06\x00\xB5\x19\x06\x00\xB6\x0D"
-     "\x06\x00"
-     "\xB7\x01\x06\x00\xA3\xA9\x07\x00\x82\x15\x00\x00\x81\xB5\x01\x00\x80\xB5"
-     "\x01\x00"
-     "\x4F\xA8\x00\x80\xA6\xA1\x07\x00\xA5\xA9\x07\x00\x53\xA8\x00\x80\xAB\xF1"
-     "\x07\x00"
-     "\xAA\xE1\x07\x00\x84\xA0\x02\x00\x57\xA8\x00\x80\xAF\xDD\x07\x00\xAE\xD1"
-     "\x07\x00"
-     "\xAD\xE5\x07\x00\xAC\xE9\x07\x00\x5B\xA8\x00\x80\xB3\x95\x06\x00\x86\xE8"
-     "\x00\x00"
-     "\x87\x18\x01\x00\xB6\xB5\x06\x00\x5F\xA8\x00\x80\x63\xA8\x00\x80\xB5\xBD"
-     "\x06\x00"
-     "\xBA\x49\x01\x00\xBB\x55\x01\x00\x67\xA8\x00\x80\x6B\xA8\x00\x80\xBE\x39"
-     "\x01\x00"
-     "\xBF\x39\x01\x00\xBC\x45\x01\x00\xBD\x39\x01\x00\xA8\x3D\x06\x00\xA9\x4D"
-     "\x06\x00"
-     "\xAA\x59\x06\x00\xAB\x51\x06\x00\xAC\x71\x06\x00\xAD\x71\x06\x00\xAE\xB9"
-     "\x01\x00"
-     "\xAF\xB9\x01\x00\x84\xAC\x01\x00\x6F\xA8\x00\x80\x73\xA8\x00\x80\x77\xA8"
-     "\x00\x80"
-     "\x7B\xA8\x00\x80\x7F\xA8\x00\x80\x83\xA8\x00\x80\x87\xA8\x00\x80\xB8\xA9"
-     "\x01\x00"
-     "\xB9\xA9\x01\x00\xBA\x69\x01\x00\xBB\x69\x01\x00\xBC\x79\x01\x00\xBD\x79"
-     "\x01\x00"
-     "\xBE\x69\x01\x00\xBF\x69\x01\x00\xB0\xC9\x01\x00\xB1\xD5\x01\x00\xB2\xD5"
-     "\x01\x00"
-     "\xB3\xA9\x01\x00\xB4\xB9\x01\x00\xB5\xB9\x01\x00\xB6\xA9\x01\x00\xB7\xA1"
-     "\x01\x00"
-     "\xA3\xD1\x05\x00\x8B\xA8\x00\x80\x8F\xA8\x00\x80\x93\xA8\x00\x80\x97\xA8"
-     "\x00\x80"
-     "\xA6\xF1\x05\x00\xA5\xF9\x05\x00\x9B\xA8\x00\x80\xAB\x11\x02\x00\xAA\x0D"
-     "\x02\x00"
-     "\x9F\xA8\x00\x80\xA3\xA8\x00\x80\xAF\x7D\x02\x00\xAE\x7D\x02\x00\xAD\x7D"
-     "\x02\x00"
-     "\xAC\x01\x02\x00\xA7\xA8\x00\x80\xAB\xA8\x00\x80\xAF\xA8\x00\x80\xB3\xA8"
-     "\x00\x80"
-     "\x81\x31\x00\x00\x80\x0D\x00\x00\xB7\xA8\x00\x80\x82\x39\x00\x00\xBB\xA8"
-     "\x00\x80"
-     "\xBF\xA8\x00\x80\xBE\x24\x03\x00\xC7\xA8\x00\x80\xCB\xA8\x00\x80\xCF\xA8"
-     "\x00\x80"
-     "\x86\xD8\x1C\x00\x87\x4C\x03\x00\xD3\xA8\x00\x80\xD7\xA8\x00\x80\xDB\xA8"
-     "\x00\x80"
-     "\x84\xC0\x1C\x00\xE3\x20\x01\x00\xDF\xA8\x00\x80\xE1\xC8\x01\x00\xE3\xA8"
-     "\x00\x80"
-     "\xEF\x30\x02\x00\xE7\xA8\x00\x80\xEB\xA8\x00\x80\xEF\xA8\x00\x80\xF3\xA8"
-     "\x00\x80"
-     "\xF7\xA8\x00\x80\xFB\xA8\x00\x80\xFF\xA8\x00\x80\xB3\x95\x03\x00\x03\xA9"
-     "\x00\x80"
-     "\x07\xA9\x00\x80\x0B\xA9\x00\x80\x0F\xA9\x00\x80\xB6\xB9\x03\x00\xB5\xB1"
-     "\x03\x00"
-     "\x13\xA9\x00\x80\xBB\x51\x03\x00\xBA\x49\x03\x00\x17\xA9\x00\x80\x1B\xA9"
-     "\x00\x80"
-     "\xBF\xF5\x00\x00\xBE\x49\x03\x00\xBD\x41\x03\x00\xBC\x49\x03\x00\xA8\x2D"
-     "\x02\x00"
-     "\xA9\x45\x02\x00\xAA\x5D\x02\x00\xAB\x55\x02\x00\xAC\x79\x02\x00\xAD\x79"
-     "\x02\x00"
-     "\xAE\xBD\x02\x00\xAF\xB5\x02\x00\xBE\x68\x1D\x00\x1F\xA9\x00\x80\x23\xA9"
-     "\x00\x80"
-     "\x27\xA9\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xA9\x00\x00\x2B\xA9"
-     "\x00\x80"
-     "\xB8\x51\x01\x00\xB9\x59\x01\x00\xBA\x61\x01\x00\xBB\x61\x01\x00\xBC\x11"
-     "\x01\x00"
-     "\xBD\x11\x01\x00\xBE\x11\x01\x00\xBF\x11\x01\x00\xB0\xCD\x02\x00\xB1\xD5"
-     "\x02\x00"
-     "\xB2\xDD\x02\x00\xB3\xD5\x02\x00\xB4\xCD\x02\x00\xB5\x71\x01\x00\xB6\x71"
-     "\x01\x00"
-     "\xB7\x71\x01\x00\xE1\x58\x06\x00\xE1\x54\x07\x00\xE3\xB0\x00\x00\xE3\xBC"
-     "\x06\x00"
-     "\x2F\xA9\x00\x80\x37\xA9\x00\x80\x86\x18\x1C\x00\x87\x54\x1D\x00\x3B\xA9"
-     "\x00\x80"
-     "\x3F\xA9\x00\x80\x43\xA9\x00\x80\x47\xA9\x00\x80\xBE\xF8\x1C\x00\x4B\xA9"
-     "\x00\x80"
-     "\xEF\xF0\x06\x00\xEF\xE0\x06\x00\xA3\x95\x02\x00\x4F\xA9\x00\x80\x53\xA9"
-     "\x00\x80"
-     "\x57\xA9\x00\x80\x5B\xA9\x00\x80\xA6\xB9\x02\x00\xA5\xB1\x02\x00\x5F\xA9"
-     "\x00\x80"
-     "\xAB\x51\x02\x00\xAA\x49\x02\x00\x63\xA9\x00\x80\x67\xA9\x00\x80\xAF\xF5"
-     "\x01\x00"
-     "\xAE\x49\x02\x00\xAD\x41\x02\x00\xAC\x49\x02\x00\xA8\x6D\x1E\x00\xA9\x75"
-     "\x1E\x00"
-     "\xAA\x7D\x1E\x00\xAB\x8D\x1E\x00\xAC\x95\x1E\x00\xAD\x9D\x1E\x00\xAE\x8D"
-     "\x1E\x00"
-     "\xAF\x81\x1E\x00\x33\xA9\x00\x80\x6B\xA9\x00\x80\x6F\xA9\x00\x80\x73\xA9"
-     "\x00\x80"
-     "\x77\xA9\x00\x80\x7B\xA9\x00\x80\x7F\xA9\x00\x80\x83\xA9\x00\x80\xB8\x89"
-     "\x1E\x00"
-     "\xB9\x89\x1E\x00\xBA\x99\x1E\x00\xBB\x91\x1E\x00\xBC\xB9\x1E\x00\xBD\xB9"
-     "\x1E\x00"
-     "\xBE\x7D\x1F\x00\xBF\x75\x1F\x00\xB0\xC5\x1E\x00\xB1\xCD\x1E\x00\xB2\xC5"
-     "\x1E\x00"
-     "\xB3\xDD\x1E\x00\xB4\xC5\x1E\x00\xB5\xCD\x1E\x00\xB6\xC5\x1E\x00\xB7\xB9"
-     "\x1E\x00"
-     "\xB3\xF5\x1E\x00\x87\xA9\x00\x80\x8B\xA9\x00\x80\x8F\xA9\x00\x80\x93\xA9"
-     "\x00\x80"
-     "\xB6\x1D\x1E\x00\xB5\x1D\x1E\x00\x97\xA9\x00\x80\xBB\x09\x1E\x00\xBA\x01"
-     "\x1E\x00"
-     "\x9B\xA9\x00\x80\x9F\xA9\x00\x80\xBF\x09\x1E\x00\xBE\x01\x1E\x00\xBD\x09"
-     "\x1E\x00"
-     "\xBC\x11\x1E\x00\x82\x6D\x00\x00\xA3\xB1\x1E\x00\x80\x55\x00\x00\x81\x65"
-     "\x00\x00"
-     "\xA6\x59\x1E\x00\x84\x98\x03\x00\xBF\xD0\x01\x00\xA5\x59\x1E\x00\xAA\x45"
-     "\x1E\x00"
-     "\xAB\x4D\x1E\x00\x86\x00\x04\x00\x87\x98\x01\x00\xAE\x45\x1E\x00\xAF\x4D"
-     "\x1E\x00"
-     "\xAC\x55\x1E\x00\xAD\x4D\x1E\x00\xA7\xA9\x00\x80\xAB\xA9\x00\x80\x84\x24"
-     "\x00\x00"
-     "\xAF\xA9\x00\x80\xB3\xA9\x00\x80\xB7\xA9\x00\x80\xC3\xA8\x00\x80\xA3\xA9"
-     "\x00\x80"
-     "\xA8\xB5\x1E\x00\xA9\x85\x1E\x00\xAA\x8D\x1E\x00\xAB\x85\x1E\x00\xAC\x9D"
-     "\x1E\x00"
-     "\xAD\x81\x1E\x00\xAE\xFD\x1E\x00\xAF\xF5\x1E\x00\xB0\x8D\x1E\x00\xB1\x95"
-     "\x1E\x00"
-     "\xB2\x95\x1E\x00\xB3\xA5\x1E\x00\xB4\xBD\x1E\x00\xB5\x71\x01\x00\xB6\x71"
-     "\x01\x00"
-     "\xB7\x71\x01\x00\xB8\x51\x01\x00\xB9\x51\x01\x00\xBA\x51\x01\x00\xBB\x51"
-     "\x01\x00"
-     "\xBC\xF5\x01\x00\xBD\xFD\x01\x00\xBE\xF5\x01\x00\xBF\xED\x01\x00\xB3\x25"
-     "\x1E\x00"
-     "\xBE\x08\x07\x00\xBB\xA9\x00\x80\xBF\xA9\x00\x80\xC3\xA9\x00\x80\xB6\x21"
-     "\x1E\x00"
-     "\xB5\x35\x1E\x00\xC7\xA9\x00\x80\xBB\x71\x1E\x00\xBA\x69\x1E\x00\xCB\xA9"
-     "\x00\x80"
-     "\xCF\xA9\x00\x80\xBF\x95\x01\x00\xBE\x59\x1E\x00\xBD\x51\x1E\x00\xBC\x61"
-     "\x1E\x00"
-     "\xD3\xA9\x00\x80\xA3\x61\x1E\x00\xD7\xA9\x00\x80\xDB\xA9\x00\x80\xA6\x65"
-     "\x1E\x00"
-     "\xDF\xA9\x00\x80\xE3\xA9\x00\x80\xA5\x71\x1E\x00\xAA\x2D\x1E\x00\xAB\x35"
-     "\x1E\x00"
-     "\xE7\xA9\x00\x80\xEB\xA9\x00\x80\xAE\x1D\x1E\x00\xAF\xD1\x01\x00\xAC\x25"
-     "\x1E\x00"
-     "\xAD\x15\x1E\x00\xE1\x54\x1A\x00\xEF\xA9\x00\x80\xE3\x5C\x0A\x00\xF3\xA9"
-     "\x00\x80"
-     "\xF7\xA9\x00\x80\xFB\xA9\x00\x80\xFF\xA9\x00\x80\x03\xAA\x00\x80\x07\xAA"
-     "\x00\x80"
-     "\xBE\xA8\x05\x00\x0B\xAA\x00\x80\x0F\xAA\x00\x80\x8F\x31\x2A\x00\x17\xAA"
-     "\x00\x80"
-     "\xEF\xC4\xFB\x00\x1B\xAA\x00\x80\x93\x61\x2E\x00\x92\x1D\x2F\x00\x91\xD9"
-     "\x2A\x00"
-     "\x90\x49\x2A\x00\x97\xD9\x12\x00\x96\x75\x12\x00\x95\x0D\x13\x00\x94\xC1"
-     "\x2E\x00"
-     "\x9B\x1D\x1B\x00\x9A\x41\x16\x00\x99\x49\x16\x00\x98\x0D\x17\x00\x9F\x71"
-     "\x1E\x00"
-     "\x9E\x11\x1B\x00\x9D\x71\x1A\x00\x9C\x79\x1A\x00\xA3\xA1\x02\x00\xA2\x9D"
-     "\x03\x00"
-     "\xA1\x95\x1F\x00\xA0\x89\x1E\x00\xE3\x88\x01\x00\xE3\x08\x1E\x00\xE1\x68"
-     "\x00\x00"
-     "\xE1\xFC\x1E\x00\xEF\xDC\x01\x00\xEF\x7C\x1F\x00\xB5\x89\xFE\x00\xB4\x01"
-     "\xFC\x00"
-     "\xB3\x01\xFA\x00\xB2\x95\xFA\x00\xB1\x01\xF8\x00\xB0\x11\xF6\x00\xAF\xE1"
-     "\xF6\x00"
-     "\xAE\x01\xF4\x00\xAD\x25\xF2\x00\xAC\xED\xF2\x00\xAB\x01\xF0\x00\xAA\xE9"
-     "\x0F\x00"
-     "\xA9\xD5\x0E\x00\xA8\xDD\x0E\x00\xA7\x01\x0C\x00\xA6\xC9\x0A\x00\xA5\xED"
-     "\x0A\x00"
-     "\xA4\x01\x08\x00\xA3\xE1\x06\x00\xA2\xF1\x06\x00\xA1\xE5\x03\x00\x1F\xAA"
-     "\x00\x80"
-     "\x82\x01\x2B\x00\x83\x01\x2B\x00\x23\xAA\x00\x80\x27\xAA\x00\x80\x86\x31"
-     "\x2F\x00"
-     "\x87\x89\x2F\x00\x84\x89\x2B\x00\x85\x45\x2E\x00\x8A\x75\x12\x00\x8B\x05"
-     "\x13\x00"
-     "\x86\x08\x05\x00\x87\x6C\x05\x00\x8E\x11\x17\x00\x8F\x11\x17\x00\x8C\xB1"
-     "\x13\x00"
-     "\x8D\x5D\x16\x00\x92\x3D\x1A\x00\x93\x41\x1B\x00\x84\xC8\x05\x00\x84\x00"
-     "\x07\x00"
-     "\x96\x51\x1F\x00\x97\x51\x1F\x00\x94\x51\x1B\x00\x95\x39\x1A\x00\x9A\x7D"
-     "\x1E\x00"
-     "\x9B\x7D\x02\x00\x2B\xAA\x00\x80\x2F\xAA\x00\x80\x81\x59\x01\x00\x80\x55"
-     "\x01\x00"
-     "\x9C\x59\x03\x00\x82\x51\x01\x00\xBE\xC8\x07\x00\x33\xAA\x00\x80\x37\xAA"
-     "\x00\x80"
-     "\x3B\xAA\x00\x80\x3F\xAA\x00\x80\x43\xAA\x00\x80\x47\xAA\x00\x80\xEF\xD4"
-     "\x1E\x00"
-     "\x4B\xAA\x00\x80\xE1\x24\x1E\x00\x4F\xAA\x00\x80\xE3\x68\x01\x00\x53\xAA"
-     "\x00\x80"
-     "\x57\xAA\x00\x80\x5B\xAA\x00\x80\x5F\xAA\x00\x80\xBB\x69\x02\x00\xBA\x59"
-     "\x02\x00"
-     "\x63\xAA\x00\x80\x67\xAA\x00\x80\xBF\x1D\x02\x00\xBE\x1D\x02\x00\xBD\x19"
-     "\x02\x00"
-     "\xBC\x71\x02\x00\xB3\xED\x02\x00\x6B\xAA\x00\x80\x6F\xAA\x00\x80\x73\xAA"
-     "\x00\x80"
-     "\x77\xAA\x00\x80\xB6\x7D\x02\x00\xB5\xED\x02\x00\x7B\xAA\x00\x80\xA3\x0D"
-     "\x05\x00"
-     "\x13\xAA\x00\x80\x7F\xAA\x00\x80\x87\xAA\x00\x80\x83\xAA\x00\x80\xA6\x9D"
-     "\x05\x00"
-     "\xA5\x0D\x05\x00\x8B\xAA\x00\x80\xAB\x89\x05\x00\xAA\xB9\x05\x00\x86\x08"
-     "\x03\x00"
-     "\x87\x7C\x03\x00\xAF\xFD\x05\x00\xAE\xFD\x05\x00\xAD\xF9\x05\x00\xAC\x91"
-     "\x05\x00"
-     "\x80\xB1\x07\x00\x81\xB9\x07\x00\x82\x41\x00\x00\xB3\xB1\x06\x00\x8F\xAA"
-     "\x00\x80"
-     "\xB5\x59\x07\x00\xB6\x31\x07\x00\x93\xAA\x00\x80\x97\xAA\x00\x80\x9B\xAA"
-     "\x00\x80"
-     "\xBA\xE1\x07\x00\xBB\xE1\x07\x00\xBC\xE1\x07\x00\xBD\xE1\x07\x00\xBE\xE1"
-     "\x07\x00"
-     "\xBF\xDD\x07\x00\xA8\xB5\x06\x00\xA9\xB9\x06\x00\xAA\x75\x06\x00\xAB\x85"
-     "\x07\x00"
-     "\xAC\x9D\x07\x00\xAD\xFD\x07\x00\xAE\xF5\x07\x00\xAF\x19\x07\x00\x9F\xAA"
-     "\x00\x80"
-     "\xA3\xAA\x00\x80\xA7\xAA\x00\x80\xAB\xAA\x00\x80\xAF\xAA\x00\x80\xB3\xAA"
-     "\x00\x80"
-     "\xB7\xAA\x00\x80\xBB\xAA\x00\x80\xB8\x75\x07\x00\xB9\x7D\x07\x00\xBA\x75"
-     "\x07\x00"
-     "\xBB\x05\x07\x00\xBC\x1D\x07\x00\xBD\x31\x07\x00\xBE\x31\x07\x00\xBF\x31"
-     "\x07\x00"
-     "\xB0\x69\x07\x00\xB1\x69\x07\x00\xB2\x79\x07\x00\xB3\x79\x07\x00\xB4\x69"
-     "\x07\x00"
-     "\xB5\x55\x07\x00\xB6\x51\x07\x00\xB7\x4D\x07\x00\xA3\xFD\x07\x00\xBF\xAA"
-     "\x00\x80"
-     "\xC3\xAA\x00\x80\xC7\xAA\x00\x80\xCB\xAA\x00\x80\xA6\x7D\x06\x00\xA5\x15"
-     "\x06\x00"
-     "\xCF\xAA\x00\x80\xAB\xAD\x06\x00\xAA\xAD\x06\x00\xD3\xAA\x00\x80\xD7\xAA"
-     "\x00\x80"
-     "\xAF\x91\x06\x00\xAE\xAD\x06\x00\xAD\xAD\x06\x00\xAC\xAD\x06\x00\xDB\xAA"
-     "\x00\x80"
-     "\xDF\xAA\x00\x80\xE3\xAA\x00\x80\xE7\xAA\x00\x80\x80\x1D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x39\x00\x00\xEB\xAA\x00\x80\xEF\xAA\x00\x80\xF7\xAA\x00\x80\x86\xC8"
-     "\x00\x00"
-     "\x87\xA4\x01\x00\xFB\xAA\x00\x80\xFF\xAA\x00\x80\x03\xAB\x00\x80\x07\xAB"
-     "\x00\x80"
-     "\xA8\x8D\x01\x00\xA9\x99\x01\x00\xAA\xD9\x01\x00\xAB\xD1\x01\x00\xAC\xF1"
-     "\x01\x00"
-     "\xAD\xF1\x01\x00\xAE\x39\x01\x00\xAF\x39\x01\x00\x84\xA0\x00\x00\x0B\xAB"
-     "\x00\x80"
-     "\x0F\xAB\x00\x80\x13\xAB\x00\x80\x17\xAB\x00\x80\x1B\xAB\x00\x80\x1F\xAB"
-     "\x00\x80"
-     "\x23\xAB\x00\x80\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5\x00\x00\xBB\xE5"
-     "\x00\x00"
-     "\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x9D\x00\x00\xBF\x95\x00\x00\xB0\x49"
-     "\x01\x00"
-     "\xB1\x49\x01\x00\xB2\x59\x01\x00\xB3\x59\x01\x00\xB4\x49\x01\x00\xB5\x49"
-     "\x01\x00"
-     "\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xBA\x05\x04\x00\xBB\x09\x04\x00\xB8\xE1"
-     "\x07\x00"
-     "\xB9\xE1\x07\x00\xBE\x09\x04\x00\xBF\x09\x04\x00\xBC\x09\x04\x00\xBD\x09"
-     "\x04\x00"
-     "\xB2\x8D\x07\x00\xB3\xE5\x07\x00\xB0\xB5\x07\x00\xB1\x85\x07\x00\xB6\xE5"
-     "\x07\x00"
-     "\xB7\xED\x07\x00\xB4\xF1\x07\x00\xB5\xF1\x07\x00\xAA\x4D\x07\x00\xAB\x55"
-     "\x07\x00"
-     "\xA8\x49\x07\x00\xA9\x49\x07\x00\xAE\xDD\x07\x00\xAF\xC5\x07\x00\xAC\x4D"
-     "\x07\x00"
-     "\xAD\xD5\x07\x00\x27\xAB\x00\x80\x2B\xAB\x00\x80\x2F\xAB\x00\x80\x33\xAB"
-     "\x00\x80"
-     "\x37\xAB\x00\x80\x3B\xAB\x00\x80\x3F\xAB\x00\x80\x43\xAB\x00\x80\xB3\xD1"
-     "\x02\x00"
-     "\x47\xAB\x00\x80\x4B\xAB\x00\x80\xBE\x00\x0C\x00\x4F\xAB\x00\x80\xB6\xF1"
-     "\x02\x00"
-     "\xB5\xF9\x02\x00\x53\xAB\x00\x80\xBB\x0D\x02\x00\xBA\x0D\x02\x00\x57\xAB"
-     "\x00\x80"
-     "\x5B\xAB\x00\x80\xBF\x0D\x02\x00\xBE\x0D\x02\x00\xBD\x0D\x02\x00\xBC\x0D"
-     "\x02\x00"
-     "\x82\x15\x00\x00\xA3\x95\x02\x00\x80\x61\x00\x00\x81\x61\x00\x00\xA6\xB5"
-     "\x02\x00"
-     "\x5F\xAB\x00\x80\x67\xAB\x00\x80\xA5\xBD\x02\x00\xAA\x49\x02\x00\xAB\x49"
-     "\x02\x00"
-     "\x86\xC8\x0C\x00\x87\xAC\x0C\x00\xAE\x49\x02\x00\xAF\x49\x02\x00\xAC\x49"
-     "\x02\x00"
-     "\xAD\x49\x02\x00\xA8\x65\x02\x00\xA9\x75\x02\x00\xAA\x7D\x02\x00\xAB\x75"
-     "\x02\x00"
-     "\xAC\x6D\x02\x00\xAD\xB1\x02\x00\xAE\xB1\x02\x00\xAF\xB1\x02\x00\x84\xA0"
-     "\x0D\x00"
-     "\x6B\xAB\x00\x80\x6F\xAB\x00\x80\x73\xAB\x00\x80\x77\xAB\x00\x80\x7B\xAB"
-     "\x00\x80"
-     "\x7F\xAB\x00\x80\x83\xAB\x00\x80\xB8\x31\x01\x00\xB9\x31\x01\x00\xBA\x31"
-     "\x01\x00"
-     "\xBB\x31\x01\x00\xBC\xD5\x01\x00\xBD\xDD\x01\x00\xBE\xC9\x01\x00\xBF\xC1"
-     "\x01\x00"
-     "\xB0\xD1\x02\x00\xB1\xD1\x02\x00\xB2\xD1\x02\x00\xB3\xD1\x02\x00\xB4\x11"
-     "\x01\x00"
-     "\xB5\x11\x01\x00\xB6\x11\x01\x00\xB7\x11\x01\x00\xE1\xB0\x06\x00\x87\xAB"
-     "\x00\x80"
-     "\xE3\xD0\x06\x00\x84\x40\x0F\x00\x8B\xAB\x00\x80\xE1\xA4\x01\x00\x8F\xAB"
-     "\x00\x80"
-     "\xE3\xC0\x06\x00\x93\xAB\x00\x80\x97\xAB\x00\x80\x9B\xAB\x00\x80\xEF\xD4"
-     "\x06\x00"
-     "\xEC\x00\x00\x00\x9F\xAB\x00\x80\xEF\x64\x07\x00\xA3\xAB\x00\x80\xA7\xAB"
-     "\x00\x80"
-     "\xAB\xAB\x00\x80\xAF\xAB\x00\x80\xB3\xB9\x02\x00\xB3\xAB\x00\x80\xB5\xA9"
-     "\x02\x00"
-     "\xB6\x7D\x02\x00\xB7\xAB\x00\x80\xBB\xAB\x00\x80\xBF\xAB\x00\x80\xBA\x59"
-     "\x02\x00"
-     "\xBB\x59\x02\x00\xBC\x49\x02\x00\xBD\x49\x02\x00\xBE\x99\x01\x00\xBF\x99"
-     "\x01\x00"
-     "\xA3\x75\x0D\x00\x63\xAB\x00\x80\xC3\xAB\x00\x80\xC7\xAB\x00\x80\xCB\xAB"
-     "\x00\x80"
-     "\xA6\xB1\x0D\x00\xA5\x65\x0D\x00\xCF\xAB\x00\x80\xAB\x95\x0D\x00\xAA\x95"
-     "\x0D\x00"
-     "\xBE\xA4\x03\x00\xD3\xAB\x00\x80\xAF\x55\x0E\x00\xAE\x55\x0E\x00\xAD\x85"
-     "\x0D\x00"
-     "\xAC\x85\x0D\x00\x80\x4D\x00\x00\x81\x55\x00\x00\x82\x55\x00\x00\xB3\x65"
-     "\x0F\x00"
-     "\xD7\xAB\x00\x80\xB5\x65\x0F\x00\xB6\x6D\x0F\x00\xDB\xAB\x00\x80\x86\x40"
-     "\x03\x00"
-     "\x87\x14\x03\x00\xBA\xED\x0F\x00\xBB\xFD\x0F\x00\xBC\xE9\x0F\x00\xBD\xD5"
-     "\x0F\x00"
-     "\xBE\xDD\x0F\x00\xBF\xD5\x0F\x00\xA8\x59\x0E\x00\xA9\xA1\x0F\x00\xAA\xA1"
-     "\x0F\x00"
-     "\xAB\xA1\x0F\x00\xAC\xA1\x0F\x00\xAD\xA1\x0F\x00\xAE\xA1\x0F\x00\xAF\xA1"
-     "\x0F\x00"
-     "\xDF\xAB\x00\x80\xE3\xAB\x00\x80\xE7\xAB\x00\x80\xEB\xAB\x00\x80\xEF\xAB"
-     "\x00\x80"
-     "\xF3\xAB\x00\x80\xF7\xAB\x00\x80\xFB\xAB\x00\x80\xB8\x01\x0F\x00\xB9\x01"
-     "\x0F\x00"
-     "\xBA\x01\x0F\x00\xBB\x1D\x0F\x00\xBC\x0D\x0F\x00\xBD\x35\x0F\x00\xBE\x3D"
-     "\x0F\x00"
-     "\xBF\xD5\x00\x00\xB0\x65\x0F\x00\xB1\x75\x0F\x00\xB2\x71\x0F\x00\xB3\x4D"
-     "\x0F\x00"
-     "\xB4\x55\x0F\x00\xB5\x5D\x0F\x00\xB6\x4D\x0F\x00\xB7\x41\x0F\x00\xA3\x29"
-     "\x0E\x00"
-     "\xFF\xAB\x00\x80\x03\xAC\x00\x80\x07\xAC\x00\x80\x0B\xAC\x00\x80\xA6\x21"
-     "\x0E\x00"
-     "\xA5\x29\x0E\x00\x0F\xAC\x00\x80\xAB\xB1\x0E\x00\xAA\xA1\x0E\x00\x13\xAC"
-     "\x00\x80"
-     "\x17\xAC\x00\x80\xAF\x99\x0E\x00\xAE\x91\x0E\x00\xAD\x99\x0E\x00\xAC\xA5"
-     "\x0E\x00"
-     "\x1B\xAC\x00\x80\x1F\xAC\x00\x80\x23\xAC\x00\x80\x27\xAC\x00\x80\xEF\x24"
-     "\x0D\x00"
-     "\x2B\xAC\x00\x80\x2F\xAC\x00\x80\x33\xAC\x00\x80\xE3\xD0\x0E\x00\x37\xAC"
-     "\x00\x80"
-     "\xE1\x18\x0E\x00\x3B\xAC\x00\x80\x80\x15\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x3F\xAC\x00\x80\xA8\xD1\x01\x00\xA9\xD9\x01\x00\xAA\x29\x01\x00\xAB\x29"
-     "\x01\x00"
-     "\xAC\x39\x01\x00\xAD\x39\x01\x00\xAE\x29\x01\x00\xAF\x29\x01\x00\xBE\x68"
-     "\x01\x00"
-     "\x47\xAC\x00\x80\x86\xC8\x01\x00\x87\xAC\x00\x00\x4B\xAC\x00\x80\x4F\xAC"
-     "\x00\x80"
-     "\x53\xAC\x00\x80\x57\xAC\x00\x80\xB8\xED\x00\x00\xB9\x85\x00\x00\xBA\x8D"
-     "\x00\x00"
-     "\xBB\x85\x00\x00\xBC\x9D\x00\x00\xBD\x81\x00\x00\xBE\x81\x00\x00\xBF\x81"
-     "\x00\x00"
-     "\xB0\x59\x01\x00\xB1\x59\x01\x00\xB2\xED\x00\x00\xB3\xE5\x00\x00\xB4\xFD"
-     "\x00\x00"
-     "\xB5\xE5\x00\x00\xB6\xE5\x00\x00\xB7\xD5\x00\x00\xB3\xA1\x02\x00\x5B\xAC"
-     "\x00\x80"
-     "\x5F\xAC\x00\x80\x63\xAC\x00\x80\x67\xAC\x00\x80\xB6\xB9\x02\x00\xB5\xB1"
-     "\x02\x00"
-     "\x6B\xAC\x00\x80\xBB\x9D\x02\x00\xBA\x9D\x02\x00\x6F\xAC\x00\x80\x73\xAC"
-     "\x00\x80"
-     "\xBF\x21\x03\x00\xBE\x39\x03\x00\xBD\x31\x03\x00\xBC\x39\x03\x00\x85\xF8"
-     "\xF5\x00"
-     "\xA3\xE5\x02\x00\x77\xAC\x00\x80\x7B\xAC\x00\x80\xA6\xFD\x02\x00\x7F\xAC"
-     "\x00\x80"
-     "\x83\xAC\x00\x80\xA5\xF5\x02\x00\xAA\xD9\x02\x00\xAB\xD9\x02\x00\x87\xAC"
-     "\x00\x80"
-     "\x8B\xAC\x00\x80\xAE\x7D\x03\x00\xAF\x65\x03\x00\xAC\x7D\x03\x00\xAD\x75"
-     "\x03\x00"
-     "\xB8\xE9\x00\x00\xB9\xE9\x00\x00\xBA\x69\x00\x00\xBB\x69\x00\x00\xBC\x79"
-     "\x00\x00"
-     "\xBD\x65\x00\x00\xBE\x6D\x00\x00\xBF\x65\x00\x00\xB0\xB1\x00\x00\xB1\xB9"
-     "\x00\x00"
-     "\xB2\x81\x00\x00\xB3\x81\x00\x00\xB4\xF9\x00\x00\xB5\xF9\x00\x00\xB6\xED"
-     "\x00\x00"
-     "\xB7\xE5\x00\x00\xA8\x65\x03\x00\xA9\x75\x03\x00\xAA\x7D\x03\x00\xAB\x75"
-     "\x03\x00"
-     "\xAC\x6D\x03\x00\xAD\xD1\x00\x00\xAE\xD1\x00\x00\xAF\xD1\x00\x00\x8F\xAC"
-     "\x00\x80"
-     "\x93\xAC\x00\x80\x97\xAC\x00\x80\xF3\xAA\x00\x80\x9B\xAC\x00\x80\x9F\xAC"
-     "\x00\x80"
-     "\xA3\xAC\x00\x80\xA7\xAC\x00\x80\x80\xFD\x01\x00\x81\x09\x00\x00\x82\x19"
-     "\x00\x00"
-     "\xAB\xAC\x00\x80\xAF\xAC\x00\x80\xBE\x44\x02\x00\xB7\xAC\x00\x80\xBB\xAC"
-     "\x00\x80"
-     "\x84\xAC\x02\x00\xBF\xAC\x00\x80\x87\xF8\x02\x00\x86\xAC\x05\x00\xC3\xAC"
-     "\x00\x80"
-     "\xC7\xAC\x00\x80\xCB\xAC\x00\x80\xCF\xAC\x00\x80\xB3\xF5\x02\x00\xD3\xAC"
-     "\x00\x80"
-     "\xD7\xAC\x00\x80\xDB\xAC\x00\x80\xDF\xAC\x00\x80\xB6\x51\x01\x00\xB5\xE5"
-     "\x02\x00"
-     "\xE3\xAC\x00\x80\xBB\x7D\x01\x00\xBA\x75\x01\x00\xE7\xAC\x00\x80\xEB\xAC"
-     "\x00\x80"
-     "\xBF\x3D\x01\x00\xBE\x3D\x01\x00\xBD\x55\x01\x00\xBC\x55\x01\x00\xE1\x70"
-     "\x0F\x00"
-     "\xEF\xAC\x00\x80\xE3\xB8\x0E\x00\x84\xC0\x05\x00\xEF\xC8\x00\x00\xF3\xAC"
-     "\x00\x80"
-     "\xF7\xAC\x00\x80\xFB\xAC\x00\x80\xE3\x3C\x0E\x00\xFF\xAC\x00\x80\xE1\xD0"
-     "\x01\x00"
-     "\x03\xAD\x00\x80\x84\x74\x07\x00\x07\xAD\x00\x80\xEF\x68\x01\x00\x0B\xAD"
-     "\x00\x80"
-     "\x0F\xAD\x00\x80\xA5\x79\x02\x00\xA6\xCD\x01\x00\x13\xAD\x00\x80\x80\x21"
-     "\x00\x00"
-     "\x81\x21\x00\x00\x82\xDD\x07\x00\xA3\x69\x02\x00\xAC\xC9\x01\x00\xAD\xC9"
-     "\x01\x00"
-     "\xAE\xA1\x01\x00\xAF\xA1\x01\x00\x17\xAD\x00\x80\x1B\xAD\x00\x80\xAA\xE9"
-     "\x01\x00"
-     "\xAB\xE1\x01\x00\xB3\xAC\x00\x80\x1F\xAD\x00\x80\xBE\x40\x02\x00\x23\xAD"
-     "\x00\x80"
-     "\x86\x30\x03\x00\x87\x30\x03\x00\x27\xAD\x00\x80\x2B\xAD\x00\x80\xA8\x39"
-     "\x07\x00"
-     "\xA9\x39\x07\x00\xAA\x0D\x07\x00\xAB\x1D\x07\x00\xAC\x05\x07\x00\xAD\x0D"
-     "\x07\x00"
-     "\xAE\x05\x07\x00\xAF\x79\x07\x00\xB0\x09\x07\x00\xB1\x09\x07\x00\xB2\x19"
-     "\x07\x00"
-     "\xB3\x11\x07\x00\xB4\x39\x07\x00\xB5\x39\x07\x00\xB6\xDD\x07\x00\xB7\xC9"
-     "\x07\x00"
-     "\xB8\xF9\x07\x00\xB9\xF9\x07\x00\xBA\xCD\x07\x00\xBB\xC1\x07\x00\xBC\xC5"
-     "\x07\x00"
-     "\xBD\xC9\x07\x00\xBE\xB9\x07\x00\xBF\xB1\x07\x00\xB3\xA5\x07\x00\x2F\xAD"
-     "\x00\x80"
-     "\x33\xAD\x00\x80\x37\xAD\x00\x80\x3B\xAD\x00\x80\xB6\xAD\x07\x00\xB5\xAD"
-     "\x07\x00"
-     "\x3F\xAD\x00\x80\xBB\xED\x07\x00\xBA\xED\x07\x00\x43\xAD\x00\x80\x47\xAD"
-     "\x00\x80"
-     "\xBF\xDD\x07\x00\xBE\xDD\x07\x00\xBD\xE5\x07\x00\xBC\xED\x07\x00\x4B\xAD"
-     "\x00\x80"
-     "\xA3\xE1\x07\x00\x4F\xAD\x00\x80\x53\xAD\x00\x80\xA6\xE9\x07\x00\x57\xAD"
-     "\x00\x80"
-     "\x5B\xAD\x00\x80\xA5\xE9\x07\x00\xAA\xA9\x07\x00\xAB\xA9\x07\x00\x5F\xAD"
-     "\x00\x80"
-     "\x63\xAD\x00\x80\xAE\x99\x07\x00\xAF\x99\x07\x00\xAC\xA9\x07\x00\xAD\xA1"
-     "\x07\x00"
-     "\x67\xAD\x00\x80\x6B\xAD\x00\x80\x6F\xAD\x00\x80\x73\xAD\x00\x80\x77\xAD"
-     "\x00\x80"
-     "\x7B\xAD\x00\x80\x7F\xAD\x00\x80\x83\xAD\x00\x80\x81\x11\x00\x00\x80\x0D"
-     "\x00\x00"
-     "\x87\xAD\x00\x80\x82\x19\x00\x00\x8B\xAD\x00\x80\x8F\xAD\x00\x80\xBE\xE4"
-     "\x01\x00"
-     "\x93\xAD\x00\x80\x86\xE0\x00\x00\x87\x18\x01\x00\x97\xAD\x00\x80\x9B\xAD"
-     "\x00\x80"
-     "\x9F\xAD\x00\x80\xA3\xAD\x00\x80\xA7\xAD\x00\x80\xEF\xB0\x04\x00\xAB\xAD"
-     "\x00\x80"
-     "\xE1\xD4\x06\x00\xAF\xAD\x00\x80\xE3\x5C\x04\x00\xB3\xAD\x00\x80\xB7\xAD"
-     "\x00\x80"
-     "\xBB\xAD\x00\x80\xBF\xAD\x00\x80\xA8\x99\x01\x00\xA9\x99\x01\x00\xAA\x0D"
-     "\x01\x00"
-     "\xAB\x05\x01\x00\xAC\x1D\x01\x00\xAD\x05\x01\x00\xAE\x05\x01\x00\xAF\x35"
-     "\x01\x00"
-     "\x84\x88\x01\x00\xC3\xAD\x00\x80\xC7\xAD\x00\x80\xCB\xAD\x00\x80\xCF\xAD"
-     "\x00\x80"
-     "\xD3\xAD\x00\x80\xD7\xAD\x00\x80\xDB\xAD\x00\x80\xB8\xC1\x00\x00\xB9\xC1"
-     "\x00\x00"
-     "\xBA\xC1\x00\x00\xBB\xC1\x00\x00\xBC\xC1\x00\x00\xBD\xC1\x00\x00\xBE\xC1"
-     "\x00\x00"
-     "\xBF\xC1\x00\x00\xB0\x4D\x01\x00\xB1\x25\x01\x00\xB2\x21\x01\x00\xB3\x21"
-     "\x01\x00"
-     "\xB4\x25\x01\x00\xB5\x2D\x01\x00\xB6\x11\x01\x00\xB7\x11\x01\x00\xDF\xAD"
-     "\x00\x80"
-     "\xE3\xAD\x00\x80\xB3\x8D\x02\x00\xE7\xAD\x00\x80\xB5\x9D\x02\x00\xEB\xAD"
-     "\x00\x80"
-     "\xEF\xAD\x00\x80\xB6\x8D\x02\x00\xF3\xAD\x00\x80\xF7\xAD\x00\x80\xBB\xE9"
-     "\x02\x00"
-     "\xBA\x81\x02\x00\xBD\xFD\x02\x00\xBC\xFD\x02\x00\xBF\xE1\x02\x00\xBE\xE9"
-     "\x02\x00"
-     "\xFB\xAD\x00\x80\xA6\xD5\x02\x00\xA5\xC5\x02\x00\xBE\x08\x03\x00\xA3\xD5"
-     "\x02\x00"
-     "\x82\x2D\x00\x00\x81\x19\x00\x00\x80\x79\x00\x00\xAF\xB9\x02\x00\xAE\xB1"
-     "\x02\x00"
-     "\xAD\xA5\x02\x00\xAC\xA5\x02\x00\xAB\xB1\x02\x00\xAA\xD9\x02\x00\xFF\xAD"
-     "\x00\x80"
-     "\x07\xAE\x00\x80\xEF\x34\x02\x00\x0B\xAE\x00\x80\x87\x18\x03\x00\x86\x2C"
-     "\xFC\x00"
-     "\x0F\xAE\x00\x80\x13\xAE\x00\x80\x17\xAE\x00\x80\x1B\xAE\x00\x80\x1F\xAE"
-     "\x00\x80"
-     "\x23\xAE\x00\x80\x27\xAE\x00\x80\x2B\xAE\x00\x80\xE3\x00\x01\x00\x2F\xAE"
-     "\x00\x80"
-     "\xE1\xE0\x01\x00\x33\xAE\x00\x80\xBA\xB5\x03\x00\xBB\xBD\x03\x00\x37\xAE"
-     "\x00\x80"
-     "\x3B\xAE\x00\x80\xBE\x79\x03\x00\xBF\x79\x03\x00\xBC\xA5\x03\x00\xBD\x79"
-     "\x03\x00"
-     "\x43\xAC\x00\x80\xB3\xB5\x03\x00\x3F\xAE\x00\x80\x43\xAE\x00\x80\xB6\x91"
-     "\x03\x00"
-     "\x47\xAE\x00\x80\x4B\xAE\x00\x80\xB5\xA5\x03\x00\xA8\x49\x02\x00\xA9\x49"
-     "\x02\x00"
-     "\xAA\x59\x02\x00\xAB\x59\x02\x00\xAC\x49\x02\x00\xAD\x75\x02\x00\xAE\x71"
-     "\x02\x00"
-     "\xAF\x6D\x02\x00\xBE\x68\xFD\x00\xBE\xA4\xFF\x00\x4F\xAE\x00\x80\x53\xAE"
-     "\x00\x80"
-     "\x57\xAE\x00\x80\x5B\xAE\x00\x80\x5F\xAE\x00\x80\x63\xAE\x00\x80\xB8\xF9"
-     "\x02\x00"
-     "\xB9\xF9\x02\x00\xBA\x49\x01\x00\xBB\x49\x01\x00\xBC\x5D\x01\x00\xBD\x41"
-     "\x01\x00"
-     "\xBE\x41\x01\x00\xBF\x7D\x01\x00\xB0\x15\x02\x00\xB1\x1D\x02\x00\xB2\x15"
-     "\x02\x00"
-     "\xB3\xC9\x02\x00\xB4\xD9\x02\x00\xB5\xD9\x02\x00\xB6\xC9\x02\x00\xB7\xC9"
-     "\x02\x00"
-     "\xE3\x20\x06\x00\xE1\xB0\x06\x00\xE1\x80\x01\x00\xE3\x10\x06\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x35\x00\x00\x82\x3D\x00\x00\x67\xAE\x00\x80\x6B\xAE\x00\x80\x6F\xAE"
-     "\x00\x80"
-     "\x77\xAE\x00\x80\x7B\xAE\x00\x80\xEF\x70\x00\x00\x7F\xAE\x00\x80\x83\xAE"
-     "\x00\x80"
-     "\xEF\x4C\x01\x00\x84\x8C\xFD\x00\x87\xAE\x00\x80\xA3\x99\x02\x00\x8B\xAE"
-     "\x00\x80"
-     "\xA5\x89\x02\x00\x8F\xAE\x00\x80\x93\xAE\x00\x80\xA6\xBD\x02\x00\x86\xC0"
-     "\xFC\x00"
-     "\x87\xE4\xFC\x00\xAB\x91\x02\x00\xAA\x99\x02\x00\xAD\x55\x02\x00\xAC\x89"
-     "\x02\x00"
-     "\xAF\x55\x02\x00\xAE\x55\x02\x00\xA8\x7D\xFE\x00\xA9\x81\xFE\x00\xAA\x99"
-     "\xFE\x00"
-     "\xAB\x99\xFE\x00\xAC\x89\xFE\x00\xAD\x89\xFE\x00\xAE\xB9\xFE\x00\xAF\xB9"
-     "\xFE\x00"
-     "\x73\xAE\x00\x80\x97\xAE\x00\x80\x9B\xAE\x00\x80\x9F\xAE\x00\x80\xA3\xAE"
-     "\x00\x80"
-     "\xA7\xAE\x00\x80\xAB\xAE\x00\x80\xAF\xAE\x00\x80\xB8\xB5\xFE\x00\xB9\xBD"
-     "\xFE\x00"
-     "\xBA\x61\xFF\x00\xBB\x61\xFF\x00\xBC\x61\xFF\x00\xBD\x61\xFF\x00\xBE\x61"
-     "\xFF\x00"
-     "\xBF\x61\xFF\x00\xB0\xC9\xFE\x00\xB1\xC9\xFE\x00\xB2\xDD\xFE\x00\xB3\xD1"
-     "\xFE\x00"
-     "\xB4\xB9\xFE\x00\xB5\xB9\xFE\x00\xB6\x91\xFE\x00\xB7\x91\xFE\x00\xB3\x11"
-     "\xFE\x00"
-     "\xB3\xAE\x00\x80\xB7\xAE\x00\x80\xBB\xAE\x00\x80\xBF\xAE\x00\x80\xB6\x09"
-     "\xFE\x00"
-     "\xB5\x01\xFE\x00\xC3\xAE\x00\x80\xBB\x0D\xFE\x00\xBA\x0D\xFE\x00\xC7\xAE"
-     "\x00\x80"
-     "\xCB\xAE\x00\x80\xBF\x7D\xFE\x00\xBE\x7D\xFE\x00\xBD\x05\xFE\x00\xBC\x09"
-     "\xFE\x00"
-     "\xCF\xAE\x00\x80\xA3\x55\xFE\x00\xD3\xAE\x00\x80\xD7\xAE\x00\x80\xA6\x4D"
-     "\xFE\x00"
-     "\xDB\xAE\x00\x80\xDF\xAE\x00\x80\xA5\x45\xFE\x00\xAA\x49\xFE\x00\xAB\x49"
-     "\xFE\x00"
-     "\x84\x28\x03\x00\xE3\xAE\x00\x80\xAE\x39\xFE\x00\xAF\x39\xFE\x00\xAC\x4D"
-     "\xFE\x00"
-     "\xAD\x41\xFE\x00\x80\xCD\x01\x00\x81\xD1\x01\x00\x82\xD1\x01\x00\xB3\xB9"
-     "\xFE\x00"
-     "\xE7\xAE\x00\x80\xB5\xD1\xFE\x00\xB6\xD5\xFE\x00\xEB\xAE\x00\x80\x86\x60"
-     "\x01\x00"
-     "\x87\x60\x01\x00\xBA\x0D\x01\x00\xBB\x05\x01\x00\xBC\x1D\x01\x00\xBD\x05"
-     "\x01\x00"
-     "\xBE\x0D\x01\x00\xBF\x05\x01\x00\xEF\xAE\x00\x80\xF3\xAE\x00\x80\xF7\xAE"
-     "\x00\x80"
-     "\xFB\xAE\x00\x80\xFF\xAE\x00\x80\xE1\xC0\xFD\x00\x03\xAF\x00\x80\xE3\x8C"
-     "\xFC\x00"
-     "\x07\xAF\x00\x80\x0B\xAF\x00\x80\x0F\xAF\x00\x80\xEF\xB4\xFC\x00\x13\xAF"
-     "\x00\x80"
-     "\x17\xAF\x00\x80\x1B\xAF\x00\x80\x1F\xAF\x00\x80\xA8\x29\xFE\x00\xA9\x29"
-     "\xFE\x00"
-     "\xAA\x3D\xFE\x00\xAB\x35\xFE\x00\xAC\x55\xFE\x00\xAD\x59\xFE\x00\xAE\x4D"
-     "\xFE\x00"
-     "\xAF\x45\xFE\x00\x23\xAF\x00\x80\x27\xAF\x00\x80\x2B\xAF\x00\x80\x2F\xAF"
-     "\x00\x80"
-     "\x33\xAF\x00\x80\x37\xAF\x00\x80\x3B\xAF\x00\x80\x3F\xAF\x00\x80\xB8\x49"
-     "\x01\x00"
-     "\xB9\x49\x01\x00\xBA\x59\x01\x00\xBB\x51\x01\x00\xBC\x79\x01\x00\xBD\x79"
-     "\x01\x00"
-     "\xBE\x19\x01\x00\xBF\x15\x01\x00\xB0\xC5\x01\x00\xB1\xCD\x01\x00\xB2\xC5"
-     "\x01\x00"
-     "\xB3\xDD\x01\x00\xB4\xC5\x01\x00\xB5\xCD\x01\x00\xB6\xC5\x01\x00\xB7\x79"
-     "\x01\x00"
-     "\x43\xAF\x00\x80\x47\xAF\x00\x80\x4B\xAF\x00\x80\xA3\xB9\xFD\x00\x4F\xAF"
-     "\x00\x80"
-     "\xA5\xD1\xFD\x00\xA6\xD5\xFD\x00\x84\xD0\x03\x00\x5E\xAF\x00\x80\xBE\xE0"
-     "\x02\x00"
-     "\xAA\x0D\x02\x00\xAB\x05\x02\x00\xAC\x1D\x02\x00\xAD\x05\x02\x00\xAE\x0D"
-     "\x02\x00"
-     "\xAF\x05\x02\x00\x81\x49\x00\x00\x80\x41\x00\x00\xA3\x09\x03\x00\x82\x5D"
-     "\x00\x00"
-     "\xA5\x19\x03\x00\x62\xAF\x00\x80\x66\xAF\x00\x80\xA6\x11\x03\x00\x86\xC0"
-     "\x04\x00"
-     "\x87\xE4\x03\x00\xAB\x0D\x03\x00\xAA\x0D\x03\x00\xAD\x01\x03\x00\xAC\x1D"
-     "\x03\x00"
-     "\xAF\x01\x03\x00\xAE\x09\x03\x00\xB0\xE1\x03\x00\xB1\xE1\x03\x00\xB2\xE1"
-     "\x03\x00"
-     "\xB3\xFD\x03\x00\xB4\xE5\x03\x00\xB5\xED\x03\x00\xB6\xE5\x03\x00\xB7\x3D"
-     "\x03\x00"
-     "\xB8\x05\x03\x00\xB9\x0D\x03\x00\xBA\x05\x03\x00\xBB\x1D\x03\x00\xBC\x05"
-     "\x03\x00"
-     "\xBD\x0D\x03\x00\xBE\x05\x03\x00\xBF\xBD\x00\x00\x6A\xAF\x00\x80\x6E\xAF"
-     "\x00\x80"
-     "\x72\xAF\x00\x80\x76\xAF\x00\x80\x03\xAE\x00\x80\x7A\xAF\x00\x80\x7E\xAF"
-     "\x00\x80"
-     "\x82\xAF\x00\x80\xA8\xF1\x03\x00\xA9\xF9\x03\x00\xAA\xA9\x03\x00\xAB\xA9"
-     "\x03\x00"
-     "\xAC\xB9\x03\x00\xAD\xB9\x03\x00\xAE\xA9\x03\x00\xAF\xA5\x03\x00\xB3\x41"
-     "\x02\x00"
-     "\x86\xAF\x00\x80\x8A\xAF\x00\x80\x8E\xAF\x00\x80\x92\xAF\x00\x80\xB6\x59"
-     "\x02\x00"
-     "\xB5\x51\x02\x00\x96\xAF\x00\x80\xBB\x45\x02\x00\xBA\x45\x02\x00\x9A\xAF"
-     "\x00\x80"
-     "\x9E\xAF\x00\x80\xBF\x49\x02\x00\xBE\x41\x02\x00\xBD\x49\x02\x00\xBC\x55"
-     "\x02\x00"
-     "\xA2\xAF\x00\x80\xA6\xAF\x00\x80\xAA\xAF\x00\x80\xAE\xAF\x00\x80\xEF\x8C"
-     "\x03\x00"
-     "\xB2\xAF\x00\x80\xB6\xAF\x00\x80\xBA\xAF\x00\x80\xE3\x6C\x03\x00\xBE\xAF"
-     "\x00\x80"
-     "\xE1\x50\x00\x00\xC2\xAF\x00\x80\xC6\xAF\x00\x80\xBE\x78\x05\x00\xCE\xAF"
-     "\x00\x80"
-     "\x84\x70\x02\x00\x80\xE5\x00\x00\x81\xE9\x00\x00\x82\xF9\x00\x00\xD2\xAF"
-     "\x00\x80"
-     "\x86\xB0\x04\x00\x87\x54\x05\x00\xD6\xAF\x00\x80\xEF\x40\xFE\x00\xDA\xAF"
-     "\x00\x80"
-     "\xE1\x5C\xFE\x00\xDE\xAF\x00\x80\xE3\x54\x01\x00\xE2\xAF\x00\x80\xE6\xAF"
-     "\x00\x80"
-     "\xEA\xAF\x00\x80\xEE\xAF\x00\x80\xB3\x99\x01\x00\xF2\xAF\x00\x80\xF6\xAF"
-     "\x00\x80"
-     "\xFA\xAF\x00\x80\xFE\xAF\x00\x80\xB6\x1D\x01\x00\xB5\x1D\x01\x00\x02\xB0"
-     "\x00\x80"
-     "\xBB\x3D\x01\x00\xBA\x3D\x01\x00\x06\xB0\x00\x80\x0A\xB0\x00\x80\xBF\xE1"
-     "\x00\x00"
-     "\xBE\xF9\x00\x00\xBD\xF1\x00\x00\xBC\xF9\x00\x00\xA8\x21\x06\x00\xA9\x51"
-     "\x06\x00"
-     "\xAA\x51\x06\x00\xAB\xCD\x06\x00\xAC\xD5\x06\x00\xAD\xDD\x06\x00\xAE\xD5"
-     "\x06\x00"
-     "\xAF\xC9\x06\x00\xCA\xAF\x00\x80\x0E\xB0\x00\x80\x12\xB0\x00\x80\x16\xB0"
-     "\x00\x80"
-     "\x1A\xB0\x00\x80\x1E\xB0\x00\x80\x22\xB0\x00\x80\x26\xB0\x00\x80\xB8\x6D"
-     "\x07\x00"
-     "\xB9\x05\x07\x00\xBA\x0D\x07\x00\xBB\x05\x07\x00\xBC\x1D\x07\x00\xBD\x01"
-     "\x07\x00"
-     "\xBE\x01\x07\x00\xBF\x01\x07\x00\xB0\xB9\x06\x00\xB1\xB9\x06\x00\xB2\x6D"
-     "\x07\x00"
-     "\xB3\x65\x07\x00\xB4\x7D\x07\x00\xB5\x65\x07\x00\xB6\x65\x07\x00\xB7\x55"
-     "\x07\x00"
-     "\xA3\xD9\x06\x00\x2A\xB0\x00\x80\x2E\xB0\x00\x80\x32\xB0\x00\x80\x36\xB0"
-     "\x00\x80"
-     "\xA6\x5D\x06\x00\xA5\x5D\x06\x00\x84\x9C\x02\x00\xAB\x7D\x06\x00\xAA\x7D"
-     "\x06\x00"
-     "\xBE\x24\x03\x00\x3A\xB0\x00\x80\xAF\xA1\x07\x00\xAE\xB9\x07\x00\xAD\xB1"
-     "\x07\x00"
-     "\xAC\xB9\x07\x00\x80\x49\x00\x00\x81\x49\x00\x00\x82\x59\x00\x00\xB3\x55"
-     "\x07\x00"
-     "\x3E\xB0\x00\x80\xB5\x7D\x07\x00\xB6\x69\x07\x00\x42\xB0\x00\x80\x86\x40"
-     "\x00\x00"
-     "\x87\x54\x03\x00\xBA\x55\x07\x00\xBB\x29\x07\x00\xBC\x39\x07\x00\xBD\x39"
-     "\x07\x00"
-     "\xBE\x29\x07\x00\xBF\x21\x07\x00\xA3\x99\x06\x00\x46\xB0\x00\x80\x4A\xB0"
-     "\x00\x80"
-     "\x4E\xB0\x00\x80\x52\xB0\x00\x80\xA6\xA5\x06\x00\xA5\xB1\x06\x00\x56\xB0"
-     "\x00\x80"
-     "\xAB\xE5\x06\x00\xAA\x99\x06\x00\x5A\xB0\x00\x80\x5E\xB0\x00\x80\xAF\xED"
-     "\x06\x00"
-     "\xAE\xE5\x06\x00\xAD\xF5\x06\x00\xAC\xF5\x06\x00\xE1\x38\x05\x00\x62\xB0"
-     "\x00\x80"
-     "\xE3\x24\x04\x00\x66\xB0\x00\x80\x6A\xB0\x00\x80\x6E\xB0\x00\x80\x72\xB0"
-     "\x00\x80"
-     "\x76\xB0\x00\x80\x7A\xB0\x00\x80\x7E\xB0\x00\x80\x82\xB0\x00\x80\x86\xB0"
-     "\x00\x80"
-     "\x8A\xB0\x00\x80\x8E\xB0\x00\x80\xEF\xF4\x04\x00\x92\xB0\x00\x80\xA8\xF9"
-     "\x06\x00"
-     "\xA9\x09\x06\x00\xAA\x11\x06\x00\xAB\x2D\x06\x00\xAC\x39\x06\x00\xAD\x25"
-     "\x06\x00"
-     "\xAE\x2D\x06\x00\xAF\x25\x06\x00\x96\xB0\x00\x80\x9A\xB0\x00\x80\x9E\xB0"
-     "\x00\x80"
-     "\xA2\xB0\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\xA6\xB0"
-     "\x00\x80"
-     "\xB8\xE5\x01\x00\xB9\xED\x01\x00\xBA\xE5\x01\x00\xBB\xFD\x01\x00\xBC\xE5"
-     "\x01\x00"
-     "\xBD\xED\x01\x00\xBE\xE5\x01\x00\xBF\x59\x01\x00\xB0\x5D\x06\x00\xB1\x21"
-     "\x06\x00"
-     "\xB2\x21\x06\x00\xB3\x21\x06\x00\xB4\x21\x06\x00\xB5\x29\x06\x00\xB6\x11"
-     "\x06\x00"
-     "\xB7\x11\x06\x00\xA8\xD1\x02\x00\xA9\xD9\x02\x00\xAA\x0D\x03\x00\xAB\x05"
-     "\x03\x00"
-     "\xAC\x1D\x03\x00\xAD\x05\x03\x00\xAE\x05\x03\x00\xAF\x35\x03\x00\xBE\x64"
-     "\x02\x00"
-     "\xAE\xB0\x00\x80\xB2\xB0\x00\x80\xB6\xB0\x00\x80\xBA\xB0\x00\x80\xBE\xB0"
-     "\x00\x80"
-     "\xC2\xB0\x00\x80\xC6\xB0\x00\x80\xB8\x25\x03\x00\xB9\x2D\x03\x00\xBA\x25"
-     "\x03\x00"
-     "\xBB\x3D\x03\x00\xBC\x25\x03\x00\xBD\x29\x03\x00\xBE\xF9\x03\x00\xBF\xF9"
-     "\x03\x00"
-     "\xB0\x4D\x03\x00\xB1\x21\x03\x00\xB2\x25\x03\x00\xB3\x3D\x03\x00\xB4\x25"
-     "\x03\x00"
-     "\xB5\x2D\x03\x00\xB6\x25\x03\x00\xB7\x1D\x03\x00\xB3\x85\x03\x00\x84\xC8"
-     "\x02\x00"
-     "\xCA\xB0\x00\x80\x84\x08\x03\x00\xCE\xB0\x00\x80\xB6\x85\x03\x00\xB5\x95"
-     "\x03\x00"
-     "\xD2\xB0\x00\x80\xBB\xE5\x03\x00\xBA\x89\x03\x00\x86\x08\x0C\x00\x87\x9C"
-     "\x03\x00"
-     "\xBF\xE9\x03\x00\xBE\xE1\x03\x00\xBD\xE9\x03\x00\xBC\xF1\x03\x00\x85\xEC"
-     "\x0A\x00"
-     "\x53\xAF\x00\x80\xA3\xCD\x03\x00\xD6\xB0\x00\x80\xA5\xDD\x03\x00\xDA\xB0"
-     "\x00\x80"
-     "\xDE\xB0\x00\x80\xA6\xCD\x03\x00\xE2\xB0\x00\x80\xE6\xB0\x00\x80\xAB\xAD"
-     "\x03\x00"
-     "\xAA\xC1\x03\x00\xAD\xA1\x03\x00\xAC\xB9\x03\x00\xAF\xA1\x03\x00\xAE\xA9"
-     "\x03\x00"
-     "\xEA\xB0\x00\x80\xEE\xB0\x00\x80\xF2\xB0\x00\x80\xF6\xB0\x00\x80\xEF\xC8"
-     "\x03\x00"
-     "\xFA\xB0\x00\x80\xFE\xB0\x00\x80\x02\xB1\x00\x80\xE3\xB4\x03\x00\x06\xB1"
-     "\x00\x80"
-     "\xE1\xD0\x01\x00\x0A\xB1\x00\x80\x80\x0D\x00\x00\x81\x75\x00\x00\x82\x7D"
-     "\x00\x00"
-     "\x0E\xB1\x00\x80\x12\xB1\x00\x80\x16\xB1\x00\x80\xB3\x65\x01\x00\xBE\x04"
-     "\x02\x00"
-     "\xB5\x65\x01\x00\x1E\xB1\x00\x80\x22\xB1\x00\x80\xB6\x65\x01\x00\x86\x40"
-     "\x0D\x00"
-     "\x87\x58\x0D\x00\xBB\xF5\x01\x00\xBA\xFD\x01\x00\xBD\xA5\x01\x00\xBC\xB9"
-     "\x01\x00"
-     "\xBF\x99\x01\x00\xBE\xA5\x01\x00\x26\xB1\x00\x80\x2A\xB1\x00\x80\x2E\xB1"
-     "\x00\x80"
-     "\x84\x00\x0C\x00\x32\xB1\x00\x80\x36\xB1\x00\x80\x3A\xB1\x00\x80\xEF\xCC"
-     "\x01\x00"
-     "\x3E\xB1\x00\x80\xE1\x2C\x06\x00\x42\xB1\x00\x80\xE3\x20\x01\x00\xEC\x00"
-     "\x00\x00"
-     "\x46\xB1\x00\x80\x4A\xB1\x00\x80\x4E\xB1\x00\x80\x52\xB1\x00\x80\xA3\xE9"
-     "\x01\x00"
-     "\x56\xB1\x00\x80\x5A\xB1\x00\x80\xA6\xE9\x01\x00\x5E\xB1\x00\x80\x62\xB1"
-     "\x00\x80"
-     "\xA5\xE9\x01\x00\xAA\x71\x01\x00\xAB\x79\x01\x00\x66\xB1\x00\x80\x6A\xB1"
-     "\x00\x80"
-     "\xAE\x29\x01\x00\xAF\x15\x01\x00\xAC\x35\x01\x00\xAD\x29\x01\x00\xA8\x25"
-     "\x0E\x00"
-     "\xA9\x2D\x0E\x00\xAA\x25\x0E\x00\xAB\x3D\x0E\x00\xAC\x25\x0E\x00\xAD\x2D"
-     "\x0E\x00"
-     "\xAE\x25\x0E\x00\xAF\x95\x0E\x00\x1A\xB1\x00\x80\x6E\xB1\x00\x80\x72\xB1"
-     "\x00\x80"
-     "\x76\xB1\x00\x80\x7A\xB1\x00\x80\x82\x9D\x00\x00\x81\x9D\x00\x00\x80\x9D"
-     "\x00\x00"
-     "\xB8\x45\x0F\x00\xB9\x4D\x0F\x00\xBA\x45\x0F\x00\xBB\x59\x0F\x00\xBC\x49"
-     "\x0F\x00"
-     "\xBD\x49\x0F\x00\xBE\x7D\x0F\x00\xBF\x71\x0F\x00\xB0\xF1\x0E\x00\xB1\xF9"
-     "\x0E\x00"
-     "\xB2\xA5\x0E\x00\xB3\xB9\x0E\x00\xB4\xA9\x0E\x00\xB5\x95\x0E\x00\xB6\x9D"
-     "\x0E\x00"
-     "\xB7\x7D\x0F\x00\xA8\xD5\x0F\x00\xA9\xDD\x0F\x00\xAA\x09\x0F\x00\xAB\x09"
-     "\x0F\x00"
-     "\xAC\x19\x0F\x00\xAD\x05\x0F\x00\xAE\x0D\x0F\x00\xAF\x05\x0F\x00\x7E\xB1"
-     "\x00\x80"
-     "\x82\xB1\x00\x80\x86\xB1\x00\x80\xBE\xA0\x03\x00\x8A\xB1\x00\x80\x8E\xB1"
-     "\x00\x80"
-     "\x87\x78\x03\x00\x86\x10\x00\x00\xB8\x15\x0F\x00\xB9\x1D\x0F\x00\xBA\x21"
-     "\x0F\x00"
-     "\xBB\x21\x0F\x00\xBC\xF5\x00\x00\xBD\xFD\x00\x00\xBE\xF5\x00\x00\xBF\xED"
-     "\x00\x00"
-     "\xB0\x41\x0F\x00\xB1\x4D\x0F\x00\xB2\x5D\x0F\x00\xB3\x55\x0F\x00\xB4\x4D"
-     "\x0F\x00"
-     "\xB5\x35\x0F\x00\xB6\x31\x0F\x00\xB7\x31\x0F\x00\x92\xB1\x00\x80\xEF\xB0"
-     "\x0C\x00"
-     "\x96\xB1\x00\x80\x9A\xB1\x00\x80\x9E\xB1\x00\x80\xA2\xB1\x00\x80\xA6\xB1"
-     "\x00\x80"
-     "\xAA\xB1\x00\x80\xAE\xB1\x00\x80\xB2\xB1\x00\x80\xB6\xB1\x00\x80\xBA\xB1"
-     "\x00\x80"
-     "\xBE\xB1\x00\x80\xE3\xA8\x0D\x00\xC2\xB1\x00\x80\xE1\x8C\x0D\x00\xB3\xC1"
-     "\x0E\x00"
-     "\xC6\xB1\x00\x80\xCA\xB1\x00\x80\xCE\xB1\x00\x80\xD2\xB1\x00\x80\xB6\xC5"
-     "\x0E\x00"
-     "\xB5\xD1\x0E\x00\xD6\xB1\x00\x80\xBB\xC9\x0E\x00\xBA\xC5\x0E\x00\xDA\xB1"
-     "\x00\x80"
-     "\xDE\xB1\x00\x80\xBF\xB1\x0E\x00\xBE\xC9\x0E\x00\xBD\xC1\x0E\x00\xBC\xC9"
-     "\x0E\x00"
-     "\xA3\x01\x0E\x00\xE2\xB1\x00\x80\xE6\xB1\x00\x80\xEA\xB1\x00\x80\xEE\xB1"
-     "\x00\x80"
-     "\xA6\x05\x0E\x00\xA5\x11\x0E\x00\xF2\xB1\x00\x80\xAB\x09\x0E\x00\xAA\x05"
-     "\x0E\x00"
-     "\xF6\xB1\x00\x80\x84\xAC\x02\x00\xAF\x71\x0E\x00\xAE\x09\x0E\x00\xAD\x01"
-     "\x0E\x00"
-     "\xAC\x09\x0E\x00\x80\x51\x00\x00\x81\x59\x00\x00\x82\x61\x00\x00\xB3\xC5"
-     "\x00\x00"
-     "\xBE\xCC\x01\x00\xB5\xC5\x00\x00\xB6\xCD\x00\x00\xFE\xB1\x00\x80\x86\x40"
-     "\x07\x00"
-     "\x87\x14\x01\x00\xBA\xC9\x00\x00\xBB\xC9\x00\x00\xBC\xD9\x00\x00\xBD\xD9"
-     "\x00\x00"
-     "\xBE\xC9\x00\x00\xBF\xC5\x00\x00\xAB\x0D\x03\x00\xAA\x0D\x03\x00\xA9\x09"
-     "\x03\x00"
-     "\xA8\xB9\x02\x00\xAF\x0D\x03\x00\xAE\x0D\x03\x00\xAD\x0D\x03\x00\xAC\x0D"
-     "\x03\x00"
-     "\xBE\x60\x03\x00\x02\xB2\x00\x80\x06\xB2\x00\x80\x0A\xB2\x00\x80\x0E\xB2"
-     "\x00\x80"
-     "\x12\xB2\x00\x80\x16\xB2\x00\x80\xBE\x30\x05\x00\xBB\x29\x03\x00\xBA\x19"
-     "\x03\x00"
-     "\xB9\x19\x03\x00\xB8\x01\x03\x00\xBF\xDD\x03\x00\xBE\xDD\x03\x00\xBD\xDD"
-     "\x03\x00"
-     "\xBC\x31\x03\x00\xB3\x4D\x03\x00\xB2\x4D\x03\x00\xB1\x4D\x03\x00\xB0\x4D"
-     "\x03\x00"
-     "\xB7\x39\x03\x00\xB6\x31\x03\x00\xB5\x41\x03\x00\xB4\x4D\x03\x00\x1A\xB2"
-     "\x00\x80"
-     "\xA6\x91\x03\x00\xA5\x99\x03\x00\x1E\xB2\x00\x80\xA3\x99\x03\x00\x22\xB2"
-     "\x00\x80"
-     "\x26\xB2\x00\x80\x2A\xB2\x00\x80\xAF\x99\x03\x00\xAE\x95\x03\x00\xAD\x85"
-     "\x03\x00"
-     "\xAC\x85\x03\x00\xAB\x95\x03\x00\xAA\x95\x03\x00\xAA\xB0\x00\x80\x2E\xB2"
-     "\x00\x80"
-     "\x32\xB2\x00\x80\x36\xB2\x00\x80\x3A\xB2\x00\x80\x3E\xB2\x00\x80\x42\xB2"
-     "\x00\x80"
-     "\x46\xB2\x00\x80\x4A\xB2\x00\x80\x4E\xB2\x00\x80\x52\xB2\x00\x80\x56\xB2"
-     "\x00\x80"
-     "\x5A\xB2\x00\x80\x5E\xB2\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xFD"
-     "\x01\x00"
-     "\x62\xB2\x00\x80\xBF\x08\x07\x00\x66\xB2\x00\x80\x6E\xB2\x00\x80\xEF\x24"
-     "\x00\x00"
-     "\x72\xB2\x00\x80\x84\x94\x02\x00\x76\xB2\x00\x80\x7A\xB2\x00\x80\x87\xE0"
-     "\x02\x00"
-     "\x86\x04\x05\x00\xBE\x00\x18\x00\x7E\xB2\x00\x80\x82\xB2\x00\x80\xE1\x90"
-     "\x01\x00"
-     "\x86\xB2\x00\x80\xE3\x80\x00\x00\x8A\xB2\x00\x80\x8E\xB2\x00\x80\x92\xB2"
-     "\x00\x80"
-     "\xB3\x65\x01\x00\x96\xB2\x00\x80\xB5\x65\x01\x00\xB6\x6D\x01\x00\x9A\xB2"
-     "\x00\x80"
-     "\x9E\xB2\x00\x80\xA2\xB2\x00\x80\xBA\xE5\x01\x00\xBB\xF9\x01\x00\xBC\xE9"
-     "\x01\x00"
-     "\xBD\xE9\x01\x00\xBE\xC5\x01\x00\xBF\xBD\x01\x00\xA6\xB2\x00\x80\xAA\xB2"
-     "\x00\x80"
-     "\xAE\xB2\x00\x80\xBE\xC4\x19\x00\xB2\xB2\x00\x80\xB6\xB2\x00\x80\xBA\xB2"
-     "\x00\x80"
-     "\xEF\xC8\x01\x00\xBE\xB2\x00\x80\xE1\xDC\x0E\x00\xC2\xB2\x00\x80\xE3\x30"
-     "\x0E\x00"
-     "\xC6\xB2\x00\x80\xCA\xB2\x00\x80\xCE\xB2\x00\x80\x84\x30\x04\x00\x80\x79"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x15\x00\x00\xA3\xE5\x01\x00\xD2\xB2\x00\x80\xA5\xE5"
-     "\x01\x00"
-     "\xA6\xED\x01\x00\xD6\xB2\x00\x80\x86\x40\x06\x00\x87\x90\x07\x00\xAA\x65"
-     "\x01\x00"
-     "\xAB\x79\x01\x00\xAC\x69\x01\x00\xAD\x69\x01\x00\xAE\x45\x01\x00\xAF\x3D"
-     "\x01\x00"
-     "\xA8\xDD\x05\x00\xA9\x21\x06\x00\xAA\x21\x06\x00\xAB\x21\x06\x00\xAC\x21"
-     "\x06\x00"
-     "\xAD\x21\x06\x00\xAE\x21\x06\x00\xAF\x9D\x06\x00\x6A\xB2\x00\x80\xDA\xB2"
-     "\x00\x80"
-     "\xDE\xB2\x00\x80\x84\x30\x01\x00\xE2\xB2\x00\x80\xE6\xB2\x00\x80\xEA\xB2"
-     "\x00\x80"
-     "\xEE\xB2\x00\x80\xB8\x8D\x06\x00\xB9\x95\x06\x00\xBA\x9D\x06\x00\xBB\x95"
-     "\x06\x00"
-     "\xBC\x8D\x06\x00\xBD\xB5\x06\x00\xBE\xBD\x06\x00\xBF\xB5\x06\x00\xB0\xED"
-     "\x06\x00"
-     "\xB1\xF1\x06\x00\xB2\xF1\x06\x00\xB3\xF1\x06\x00\xB4\xCD\x06\x00\xB5\xB5"
-     "\x06\x00"
-     "\xB6\xBD\x06\x00\xB7\xB5\x06\x00\xA8\x89\x07\x00\xA9\x95\x07\x00\xAA\x91"
-     "\x07\x00"
-     "\xAB\x91\x07\x00\xAC\xBD\x07\x00\xAD\xA5\x07\x00\xAE\xA1\x07\x00\xAF\xDD"
-     "\x07\x00"
-     "\xF2\xB2\x00\x80\xF6\xB2\x00\x80\xFA\xB2\x00\x80\xFE\xB2\x00\x80\x02\xB3"
-     "\x00\x80"
-     "\x06\xB3\x00\x80\x0A\xB3\x00\x80\x0E\xB3\x00\x80\xB8\x49\x07\x00\xB9\x55"
-     "\x07\x00"
-     "\xBA\x5D\x07\x00\xBB\x55\x07\x00\xBC\x71\x07\x00\xBD\x7D\x07\x00\xBE\x69"
-     "\x07\x00"
-     "\xBF\x69\x07\x00\xB0\xA5\x07\x00\xB1\xAD\x07\x00\xB2\xB9\x07\x00\xB3\xB1"
-     "\x07\x00"
-     "\xB4\x91\x07\x00\xB5\x91\x07\x00\xB6\x79\x07\x00\xB7\x79\x07\x00\x12\xB3"
-     "\x00\x80"
-     "\x16\xB3\x00\x80\x1A\xB3\x00\x80\x1E\xB3\x00\x80\xEF\xC8\x05\x00\xE1\xC0"
-     "\x09\x00"
-     "\x22\xB3\x00\x80\xE3\xC0\x19\x00\xE3\x24\x04\x00\x26\xB3\x00\x80\xE1\xD0"
-     "\x06\x00"
-     "\xEF\xDC\x28\x00\xA2\x9D\x03\x00\xA3\x15\x01\x00\xA0\x19\x05\x00\xA1\x8D"
-     "\x05\x00"
-     "\xB3\x59\x06\x00\x2A\xB3\x00\x80\x2E\xB3\x00\x80\x32\xB3\x00\x80\x36\xB3"
-     "\x00\x80"
-     "\xB6\x65\x06\x00\xB5\x75\x06\x00\x3A\xB3\x00\x80\xBB\x29\x06\x00\xBA\x21"
-     "\x06\x00"
-     "\x3E\xB3\x00\x80\x42\xB3\x00\x80\xBF\x15\x06\x00\xBE\x15\x06\x00\xBD\x25"
-     "\x06\x00"
-     "\xBC\x2D\x06\x00\xA3\x99\x06\x00\x8F\x99\xFC\x00\x46\xB3\x00\x80\x4E\xB3"
-     "\x00\x80"
-     "\x52\xB3\x00\x80\xA6\xA5\x06\x00\xA5\xB5\x06\x00\x56\xB3\x00\x80\xAB\xE9"
-     "\x06\x00"
-     "\xAA\xE1\x06\x00\x86\x28\x1F\x00\x87\x9C\x00\x00\xAF\xD5\x06\x00\xAE\xD5"
-     "\x06\x00"
-     "\xAD\xE5\x06\x00\xAC\xED\x06\x00\x9E\x6D\x09\x00\x9F\x7D\x07\x00\x9C\x0D"
-     "\x0B\x00"
-     "\x9D\xED\x09\x00\x9A\xF1\x0D\x00\x9B\x39\x0D\x00\x98\xE5\xF0\x00\x99\x0D"
-     "\x0F\x00"
-     "\x96\xE1\xF1\x00\x97\xE9\xF1\x00\x94\xC5\xF5\x00\x95\x0D\xF3\x00\x92\x1D"
-     "\xF7\x00"
-     "\x93\xF1\xF5\x00\x90\xFD\xF9\x00\x91\xED\xF9\x00\x82\x1D\xFF\x00\x83\x01"
-     "\xFA\x00"
-     "\x5A\xB3\x00\x80\x5E\xB3\x00\x80\x86\x15\xF6\x00\x87\x39\xF6\x00\x84\x09"
-     "\xFA\x00"
-     "\x85\xF1\xF4\x00\x8A\xC1\xF0\x00\x8B\x25\xF2\x00\x62\xB3\x00\x80\x66\xB3"
-     "\x00\x80"
-     "\x8E\xE1\x0C\x00\x8F\x15\x0E\x00\x8C\x35\xF2\x00\x8D\x01\xF3\x00\x92\xAD"
-     "\x0E\x00"
-     "\x93\x81\x08\x00\x6A\xB3\x00\x80\x6E\xB3\x00\x80\x96\xE9\x04\x00\x97\x75"
-     "\x06\x00"
-     "\x94\x79\x0A\x00\x95\xF1\x0A\x00\x9A\xD1\x06\x00\x9B\xC9\x00\x00\x72\xB3"
-     "\x00\x80"
-     "\x76\xB3\x00\x80\x81\x1D\x03\x00\x80\x1D\x03\x00\x9C\x59\x02\x00\x82\xF5"
-     "\x03\x00"
-     "\xAB\x01\x10\x00\xAA\x95\x16\x00\xA9\x8D\x16\x00\xA8\x8D\x16\x00\xAF\x95"
-     "\x2E\x00"
-     "\xAE\x01\x2C\x00\xAD\xFD\x12\x00\xAC\x99\x12\x00\xA3\xA5\x1E\x00\xA2\xA5"
-     "\x1E\x00"
-     "\xA1\x8D\x02\x00\xFA\xB1\x00\x80\xA7\x19\x1A\x00\xA6\x95\x1A\x00\xA5\x01"
-     "\x18\x00"
-     "\xA4\x9D\x1F\x00\x7A\xB3\x00\x80\x7E\xB3\x00\x80\x82\xB3\x00\x80\x86\xB3"
-     "\x00\x80"
-     "\x8A\xB3\x00\x80\x8E\xB3\x00\x80\x92\xB3\x00\x80\x96\xB3\x00\x80\xB3\xE5"
-     "\x2A\x00"
-     "\xB2\xE5\x2A\x00\xB1\xAD\x2F\x00\xB0\xE5\x2E\x00\x9A\xB3\x00\x80\x9E\xB3"
-     "\x00\x80"
-     "\xB5\x01\x24\x00\xB4\x11\x2A\x00\xA8\x29\x03\x00\xA9\x35\x03\x00\xAA\x3D"
-     "\x03\x00"
-     "\xAB\x35\x03\x00\xAC\x2D\x03\x00\xAD\xB5\x03\x00\xAE\xBD\x03\x00\xAF\xB5"
-     "\x03\x00"
-     "\xA2\xB3\x00\x80\xA6\xB3\x00\x80\xAA\xB3\x00\x80\xAE\xB3\x00\x80\x80\x1D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\xB9\x00\x00\xB2\xB3\x00\x80\xB8\x4D\x02\x00\xB9\x5D"
-     "\x02\x00"
-     "\xBA\x6D\x02\x00\xBB\x09\x02\x00\xBC\x19\x02\x00\xBD\x19\x02\x00\xBE\x09"
-     "\x02\x00"
-     "\xBF\x01\x02\x00\xB0\xCD\x03\x00\xB1\xD5\x03\x00\xB2\xDD\x03\x00\xB3\xD5"
-     "\x03\x00"
-     "\xB4\xCD\x03\x00\xB5\x75\x02\x00\xB6\x7D\x02\x00\xB7\x75\x02\x00\xB6\xB3"
-     "\x00\x80"
-     "\x84\xC8\x1D\x00\xBE\xB3\x00\x80\xBE\x0C\x1F\x00\xC2\xB3\x00\x80\xC6\xB3"
-     "\x00\x80"
-     "\xEF\x48\x06\x00\xEF\x58\x07\x00\xE1\x58\x06\x00\xE1\x98\x06\x00\xE3\x94"
-     "\x01\x00"
-     "\xE3\x00\x06\x00\x86\x10\x1C\x00\x87\x7C\x1D\x00\xBE\xF4\x1E\x00\xCA\xB3"
-     "\x00\x80"
-     "\xCE\xB3\x00\x80\xB6\x65\x03\x00\xB5\xF5\x03\x00\xD2\xB3\x00\x80\xB3\xE5"
-     "\x03\x00"
-     "\xD6\xB3\x00\x80\xDA\xB3\x00\x80\xDE\xB3\x00\x80\xBF\xE1\x02\x00\xBE\x59"
-     "\x03\x00"
-     "\xBD\x51\x03\x00\xBC\x59\x03\x00\xBB\x41\x03\x00\xBA\x59\x03\x00\xE2\xB3"
-     "\x00\x80"
-     "\xE6\xB3\x00\x80\x4A\xB3\x00\x80\xBA\xB3\x00\x80\xEA\xB3\x00\x80\xEE\xB3"
-     "\x00\x80"
-     "\xF2\xB3\x00\x80\xF6\xB3\x00\x80\xFA\xB3\x00\x80\xFE\xB3\x00\x80\xA8\xAD"
-     "\x1D\x00"
-     "\xA9\xB5\x1D\x00\xAA\xB5\x1D\x00\xAB\x25\x1E\x00\xAC\x3D\x1E\x00\xAD\x1D"
-     "\x1E\x00"
-     "\xAE\x15\x1E\x00\xAF\x75\x1E\x00\xB0\x0D\x1E\x00\xB1\x2D\x1E\x00\xB2\x25"
-     "\x1E\x00"
-     "\xB3\x8D\x1E\x00\xB4\x95\x1E\x00\xB5\x9D\x1E\x00\xB6\x95\x1E\x00\xB7\x8D"
-     "\x1E\x00"
-     "\xB8\xB5\x1E\x00\xB9\xBD\x1E\x00\xBA\xB5\x1E\x00\xBB\x9D\x1E\x00\xBC\x85"
-     "\x1E\x00"
-     "\xBD\x55\x1F\x00\xBE\x5D\x1F\x00\xBF\x55\x1F\x00\xB3\x1D\x1D\x00\x02\xB4"
-     "\x00\x80"
-     "\x06\xB4\x00\x80\x0A\xB4\x00\x80\x0E\xB4\x00\x80\xB6\xBD\x1E\x00\xB5\x95"
-     "\x1E\x00"
-     "\x12\xB4\x00\x80\xBB\xC9\x1E\x00\xBA\xE9\x1E\x00\x16\xB4\x00\x80\x1A\xB4"
-     "\x00\x80"
-     "\xBF\x79\x1E\x00\xBE\x71\x1E\x00\xBD\x79\x1E\x00\xBC\xD1\x1E\x00\x82\x29"
-     "\x00\x00"
-     "\xA3\x59\x1D\x00\x80\x1D\x00\x00\x81\x15\x00\x00\xA6\xF9\x1E\x00\x1E\xB4"
-     "\x00\x80"
-     "\x22\xB4\x00\x80\xA5\xD1\x1E\x00\xAA\xAD\x1E\x00\xAB\x8D\x1E\x00\x26\xB4"
-     "\x00\x80"
-     "\x84\xE0\x03\x00\xAE\x35\x1E\x00\xAF\x3D\x1E\x00\xAC\x95\x1E\x00\xAD\x3D"
-     "\x1E\x00"
-     "\xA8\x89\x1E\x00\xA9\x95\x1E\x00\xAA\x9D\x1E\x00\xAB\xB1\x1E\x00\xAC\xD1"
-     "\x1E\x00"
-     "\xAD\xD9\x1E\x00\xAE\xCD\x1E\x00\xAF\xC5\x1E\x00\x2A\xB4\x00\x80\x2E\xB4"
-     "\x00\x80"
-     "\x86\x88\x00\x00\x87\x6C\x01\x00\x32\xB4\x00\x80\x36\xB4\x00\x80\x3A\xB4"
-     "\x00\x80"
-     "\x3E\xB4\x00\x80\xB8\x5D\x01\x00\xB9\xC1\x01\x00\xBA\xC1\x01\x00\xBB\xC1"
-     "\x01\x00"
-     "\xBC\xC1\x01\x00\xBD\xC9\x01\x00\xBE\xF1\x01\x00\xBF\xF1\x01\x00\xB0\xBD"
-     "\x1E\x00"
-     "\xB1\x9D\x1E\x00\xB2\x95\x1E\x00\xB3\x65\x01\x00\xB4\x7D\x01\x00\xB5\x65"
-     "\x01\x00"
-     "\xB6\x6D\x01\x00\xB7\x65\x01\x00\xAA\x2D\x1D\x00\xAB\x35\x1D\x00\x42\xB4"
-     "\x00\x80"
-     "\x46\xB4\x00\x80\xAE\x2D\x1D\x00\xAF\x95\x1C\x00\xAC\x2D\x1D\x00\xAD\x25"
-     "\x1D\x00"
-     "\x84\x8C\x01\x00\xA3\x91\x1D\x00\x4A\xB4\x00\x80\x4E\xB4\x00\x80\xA6\x11"
-     "\x1D\x00"
-     "\x52\xB4\x00\x80\x56\xB4\x00\x80\xA5\x81\x1D\x00\xB3\x55\x1E\x00\x5A\xB4"
-     "\x00\x80"
-     "\x5E\xB4\x00\x80\x62\xB4\x00\x80\x66\xB4\x00\x80\xB6\x19\x1E\x00\xB5\x19"
-     "\x1E\x00"
-     "\x6A\xB4\x00\x80\xBB\x19\x1E\x00\xBA\x39\x1E\x00\x6E\xB4\x00\x80\x72\xB4"
-     "\x00\x80"
-     "\xBF\xE1\x01\x00\xBE\xF9\x01\x00\xBD\xF1\x01\x00\xBC\x01\x1E\x00\x76\xB4"
-     "\x00\x80"
-     "\x7A\xB4\x00\x80\x7E\xB4\x00\x80\xA3\x99\x1D\x00\x82\xB4\x00\x80\xA5\xD5"
-     "\x1D\x00"
-     "\xA6\xD5\x1D\x00\x86\xB4\x00\x80\x8A\xB4\x00\x80\x8E\xB4\x00\x80\xAA\xF5"
-     "\x1D\x00"
-     "\xAB\xD5\x1D\x00\xAC\xCD\x1D\x00\xAD\x3D\x02\x00\xAE\x35\x02\x00\xAF\x2D"
-     "\x02\x00"
-     "\x80\x65\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\x84\x00\x04\x00\x92\xB4"
-     "\x00\x80"
-     "\x96\xB4\x00\x80\x87\x70\x03\x00\x86\xFC\x04\x00\x9E\xB4\x00\x80\xA2\xB4"
-     "\x00\x80"
-     "\xA6\xB4\x00\x80\xAA\xB4\x00\x80\xAE\xB4\x00\x80\xB2\xB4\x00\x80\xB6\xB4"
-     "\x00\x80"
-     "\xBA\xB4\x00\x80\xBE\xC8\x04\x00\xBE\xB4\x00\x80\xC2\xB4\x00\x80\xC6\xB4"
-     "\x00\x80"
-     "\xCA\xB4\x00\x80\xCE\xB4\x00\x80\xD2\xB4\x00\x80\xEF\xDC\x1F\x00\xD6\xB4"
-     "\x00\x80"
-     "\xE1\x94\x01\x00\xDA\xB4\x00\x80\xE3\x1C\x01\x00\xDE\xB4\x00\x80\xE2\xB4"
-     "\x00\x80"
-     "\xE6\xB4\x00\x80\xEA\xB4\x00\x80\xBB\x75\x03\x00\xBA\x69\x03\x00\xBE\x48"
-     "\x06\x00"
-     "\xEE\xB4\x00\x80\xBF\x1D\x03\x00\xBE\x1D\x03\x00\xBD\x1D\x03\x00\xBC\x65"
-     "\x03\x00"
-     "\xB3\xD5\x03\x00\xF2\xB4\x00\x80\xF6\xB4\x00\x80\xFA\xB4\x00\x80\xFE\xB4"
-     "\x00\x80"
-     "\xB6\x7D\x03\x00\xB5\xC5\x03\x00\x84\x70\x05\x00\xA8\x25\x02\x00\xA9\x35"
-     "\x02\x00"
-     "\xAA\x3D\x02\x00\xAB\x35\x02\x00\xAC\x2D\x02\x00\xAD\x9D\x02\x00\xAE\x95"
-     "\x02\x00"
-     "\xAF\xB5\x02\x00\x82\x15\x00\x00\x02\xB5\x00\x80\x80\xD9\x01\x00\x81\x09"
-     "\x00\x00"
-     "\xC4\x00\x00\x00\x06\xB5\x00\x80\x0E\xB5\x00\x80\x12\xB5\x00\x80\xB8\xA9"
-     "\x02\x00"
-     "\xB9\xA9\x02\x00\xBA\x49\x01\x00\xBB\x49\x01\x00\xBC\x59\x01\x00\xBD\x45"
-     "\x01\x00"
-     "\xBE\x45\x01\x00\xBF\x7D\x01\x00\xB0\xCD\x02\x00\xB1\xD1\x02\x00\xB2\xD1"
-     "\x02\x00"
-     "\xB3\xA9\x02\x00\xB4\xB9\x02\x00\xB5\xB9\x02\x00\xB6\xA9\x02\x00\xB7\xA1"
-     "\x02\x00"
-     "\xE1\x28\x1E\x00\xE1\x34\x1C\x00\xE3\x74\x01\x00\xE3\x18\x1E\x00\x16\xB5"
-     "\x00\x80"
-     "\x1A\xB5\x00\x80\x86\xB8\x04\x00\x87\x54\x05\x00\x84\x38\x07\x00\x1E\xB5"
-     "\x00\x80"
-     "\x22\xB5\x00\x80\x26\xB5\x00\x80\xBE\xAC\x07\x00\x2A\xB5\x00\x80\xEF\xCC"
-     "\x1E\x00"
-     "\xEF\xC8\x1A\x00\xA3\xF5\x02\x00\x2E\xB5\x00\x80\x32\xB5\x00\x80\x36\xB5"
-     "\x00\x80"
-     "\x3A\xB5\x00\x80\xA6\x5D\x02\x00\xA5\xE5\x02\x00\x3E\xB5\x00\x80\xAB\x55"
-     "\x02\x00"
-     "\xAA\x49\x02\x00\x42\xB5\x00\x80\x46\xB5\x00\x80\xAF\x3D\x02\x00\xAE\x3D"
-     "\x02\x00"
-     "\xAD\x3D\x02\x00\xAC\x45\x02\x00\xA8\x61\x06\x00\xA9\x61\x06\x00\xAA\x61"
-     "\x06\x00"
-     "\xAB\x61\x06\x00\xAC\x61\x06\x00\xAD\x61\x06\x00\xAE\x61\x06\x00\xAF\x61"
-     "\x06\x00"
-     "\x0A\xB5\x00\x80\x4A\xB5\x00\x80\x4E\xB5\x00\x80\x52\xB5\x00\x80\x56\xB5"
-     "\x00\x80"
-     "\x5A\xB5\x00\x80\x5E\xB5\x00\x80\x62\xB5\x00\x80\xB8\xF1\x06\x00\xB9\xF1"
-     "\x06\x00"
-     "\xBA\xF1\x06\x00\xBB\xF1\x06\x00\xBC\x9D\x06\x00\xBD\xB1\x06\x00\xBE\xB1"
-     "\x06\x00"
-     "\xBF\xB1\x06\x00\xB0\xE5\x06\x00\xB1\xED\x06\x00\xB2\xE5\x06\x00\xB3\xFD"
-     "\x06\x00"
-     "\xB4\xE5\x06\x00\xB5\xED\x06\x00\xB6\xD9\x06\x00\xB7\xD5\x06\x00\xB3\xE9"
-     "\x06\x00"
-     "\x66\xB5\x00\x80\x6A\xB5\x00\x80\x6E\xB5\x00\x80\x72\xB5\x00\x80\xB6\xE1"
-     "\x06\x00"
-     "\xB5\xE9\x06\x00\x76\xB5\x00\x80\xBB\x29\x06\x00\xBA\x21\x06\x00\x7A\xB5"
-     "\x00\x80"
-     "\x7E\xB5\x00\x80\xBF\x29\x06\x00\xBE\x21\x06\x00\xBD\x29\x06\x00\xBC\x31"
-     "\x06\x00"
-     "\x82\x5D\x00\x00\xA3\xAD\x06\x00\x80\x45\x00\x00\x81\x5D\x00\x00\xA6\xA5"
-     "\x06\x00"
-     "\x82\xB5\x00\x80\x86\xB5\x00\x80\xA5\xAD\x06\x00\xAA\x65\x06\x00\xAB\x6D"
-     "\x06\x00"
-     "\x86\x00\x0C\x00\x87\x40\x03\x00\xAE\x65\x06\x00\xAF\x6D\x06\x00\xAC\x75"
-     "\x06\x00"
-     "\xAD\x6D\x06\x00\x8A\xB5\x00\x80\xEF\x7C\x05\x00\x8E\xB5\x00\x80\x92\xB5"
-     "\x00\x80"
-     "\x96\xB5\x00\x80\x9A\xB5\x00\x80\x9E\xB5\x00\x80\xA2\xB5\x00\x80\xA6\xB5"
-     "\x00\x80"
-     "\xAA\xB5\x00\x80\xAE\xB5\x00\x80\xB2\xB5\x00\x80\xB6\xB5\x00\x80\xE3\x68"
-     "\x05\x00"
-     "\xBA\xB5\x00\x80\xE1\x78\x05\x00\xB3\xD1\x06\x00\xBE\xB5\x00\x80\xC2\xB5"
-     "\x00\x80"
-     "\xC6\xB5\x00\x80\xCA\xB5\x00\x80\xB6\xFD\x06\x00\xB5\xFD\x06\x00\xCE\xB5"
-     "\x00\x80"
-     "\xBB\xA9\x06\x00\xBA\xA1\x06\x00\xD2\xB5\x00\x80\xD6\xB5\x00\x80\xBF\x99"
-     "\x06\x00"
-     "\xBE\xA9\x06\x00\xBD\xA9\x06\x00\xBC\xB1\x06\x00\xA8\xB9\x06\x00\xA9\xB9"
-     "\x06\x00"
-     "\xAA\x19\x06\x00\xAB\x19\x06\x00\xAC\x35\x06\x00\xAD\x3D\x06\x00\xAE\x35"
-     "\x06\x00"
-     "\xAF\x29\x06\x00\xDA\xB5\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xDE\xB5\x00\x80\xE2\xB5\x00\x80\xE6\xB5\x00\x80\xEE\xB5\x00\x80\xB8\xE9"
-     "\x01\x00"
-     "\xB9\xE9\x01\x00\xBA\xF9\x01\x00\xBB\xF9\x01\x00\xBC\xE9\x01\x00\xBD\xE9"
-     "\x01\x00"
-     "\xBE\x5D\x01\x00\xBF\x55\x01\x00\xB0\x25\x06\x00\xB1\x2D\x06\x00\xB2\x25"
-     "\x06\x00"
-     "\xB3\x3D\x06\x00\xB4\x2D\x06\x00\xB5\x1D\x06\x00\xB6\x15\x06\x00\xB7\xD9"
-     "\x01\x00"
-     "\x86\x80\x0C\x00\x87\xE4\x02\x00\xF2\xB5\x00\x80\xA3\x9D\x05\x00\xF6\xB5"
-     "\x00\x80"
-     "\xA5\xB1\x05\x00\xA6\xB1\x05\x00\xFA\xB5\x00\x80\xFE\xB5\x00\x80\x02\xB6"
-     "\x00\x80"
-     "\xAA\xED\x05\x00\xAB\xE5\x05\x00\xAC\xFD\x05\x00\xAD\xE5\x05\x00\xAE\xE5"
-     "\x05\x00"
-     "\xAF\xD5\x05\x00\xB6\x4D\x03\x00\x06\xB6\x00\x80\x84\xC4\x03\x00\xB5\x45"
-     "\x03\x00"
-     "\x0A\xB6\x00\x80\xB3\x8D\x02\x00\x0E\xB6\x00\x80\x12\xB6\x00\x80\xBE\x49"
-     "\x03\x00"
-     "\xBF\x49\x03\x00\xBC\x49\x03\x00\xBD\x49\x03\x00\xBA\x69\x03\x00\xBB\x69"
-     "\x03\x00"
-     "\x16\xB6\x00\x80\x1A\xB6\x00\x80\x1E\xB6\x00\x80\xA6\x89\x03\x00\xA5\x81"
-     "\x03\x00"
-     "\x22\xB6\x00\x80\xA3\x49\x02\x00\x26\xB6\x00\x80\x2A\xB6\x00\x80\x2E\xB6"
-     "\x00\x80"
-     "\xAF\x8D\x03\x00\xAE\x8D\x03\x00\xAD\x8D\x03\x00\xAC\x8D\x03\x00\xAB\xAD"
-     "\x03\x00"
-     "\xAA\xAD\x03\x00\x9A\xB4\x00\x80\x32\xB6\x00\x80\x36\xB6\x00\x80\x3A\xB6"
-     "\x00\x80"
-     "\x85\xB4\x3D\x00\x3E\xB6\x00\x80\x42\xB6\x00\x80\x46\xB6\x00\x80\x4A\xB6"
-     "\x00\x80"
-     "\x4E\xB6\x00\x80\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x1D\x00\x00\x52\xB6"
-     "\x00\x80"
-     "\xBE\xB0\x03\x00\x56\xB6\x00\x80\x87\x38\x03\x00\x86\xDC\x0C\x00\x5E\xB6"
-     "\x00\x80"
-     "\x62\xB6\x00\x80\x66\xB6\x00\x80\x6A\xB6\x00\x80\x6E\xB6\x00\x80\xEF\x5C"
-     "\x06\x00"
-     "\x72\xB6\x00\x80\x76\xB6\x00\x80\xBE\xE8\x0C\x00\xE3\x94\x06\x00\x7A\xB6"
-     "\x00\x80"
-     "\xE1\xDC\x01\x00\x7E\xB6\x00\x80\x82\xB6\x00\x80\x86\xB6\x00\x80\x8A\xB6"
-     "\x00\x80"
-     "\xB3\x51\x01\x00\x8E\xB6\x00\x80\x92\xB6\x00\x80\x96\xB6\x00\x80\x9A\xB6"
-     "\x00\x80"
-     "\xB6\x71\x01\x00\xB5\x79\x01\x00\x9E\xB6\x00\x80\xBB\x3D\x01\x00\xBA\x3D"
-     "\x01\x00"
-     "\xA2\xB6\x00\x80\xA6\xB6\x00\x80\xBF\xFD\x01\x00\xBE\xF5\x01\x00\xBD\x05"
-     "\x01\x00"
-     "\xBC\x05\x01\x00\xAA\xB6\x00\x80\xAE\xB6\x00\x80\xB2\xB6\x00\x80\x84\x40"
-     "\x0C\x00"
-     "\xB6\xB6\x00\x80\xBA\xB6\x00\x80\xBE\xB6\x00\x80\xEF\xAC\x07\x00\xC2\xB6"
-     "\x00\x80"
-     "\xE1\x24\x06\x00\xC6\xB6\x00\x80\xE3\x40\x07\x00\x86\x90\x0C\x00\x87\xFC"
-     "\x0C\x00"
-     "\xCE\xB6\x00\x80\xD2\xB6\x00\x80\x80\x59\x00\x00\x81\x65\x00\x00\x82\x61"
-     "\x00\x00"
-     "\xA3\xDD\x01\x00\xD6\xB6\x00\x80\xA5\xF5\x01\x00\xA6\xFD\x01\x00\xDA\xB6"
-     "\x00\x80"
-     "\xDE\xB6\x00\x80\xE2\xB6\x00\x80\xAA\xB1\x01\x00\xAB\xB1\x01\x00\xAC\x89"
-     "\x01\x00"
-     "\xAD\x89\x01\x00\xAE\x79\x01\x00\xAF\x71\x01\x00\xEA\xB5\x00\x80\x5A\xB6"
-     "\x00\x80"
-     "\xE6\xB6\x00\x80\xEA\xB6\x00\x80\xCA\xB6\x00\x80\xEE\xB6\x00\x80\xF2\xB6"
-     "\x00\x80"
-     "\xF6\xB6\x00\x80\xA8\x9D\x0D\x00\xA9\x2D\x0E\x00\xAA\x39\x0E\x00\xAB\x31"
-     "\x0E\x00"
-     "\xAC\x11\x0E\x00\xAD\x11\x0E\x00\xAE\x7D\x0E\x00\xAF\x6D\x0E\x00\xB0\x19"
-     "\x0E\x00"
-     "\xB1\x19\x0E\x00\xB2\x31\x0E\x00\xB3\x31\x0E\x00\xB4\xD1\x0E\x00\xB5\xD9"
-     "\x0E\x00"
-     "\xB6\xCD\x0E\x00\xB7\xC5\x0E\x00\xB8\xFD\x0E\x00\xB9\xD9\x0E\x00\xBA\xA9"
-     "\x0E\x00"
-     "\xBB\xA9\x0E\x00\xBC\xBD\x0E\x00\xBD\xA5\x0E\x00\xBE\xAD\x0E\x00\xBF\xA5"
-     "\x0E\x00"
-     "\xA8\x81\x0F\x00\xA9\x81\x0F\x00\xAA\x81\x0F\x00\xAB\x81\x0F\x00\xAC\x81"
-     "\x0F\x00"
-     "\xAD\x8D\x0F\x00\xAE\x85\x0F\x00\xAF\xB5\x0F\x00\xFA\xB6\x00\x80\xFE\xB6"
-     "\x00\x80"
-     "\x02\xB7\x00\x80\x06\xB7\x00\x80\x0A\xB7\x00\x80\x0E\xB7\x00\x80\x12\xB7"
-     "\x00\x80"
-     "\x16\xB7\x00\x80\xB8\x9D\x0F\x00\xB9\xAD\x0F\x00\xBA\xA5\x0F\x00\xBB\x4D"
-     "\x0F\x00"
-     "\xBC\x55\x0F\x00\xBD\x5D\x0F\x00\xBE\x49\x0F\x00\xBF\x49\x0F\x00\xB0\xD1"
-     "\x0F\x00"
-     "\xB1\xD1\x0F\x00\xB2\xD1\x0F\x00\xB3\xD1\x0F\x00\xB4\xB5\x0F\x00\xB5\xBD"
-     "\x0F\x00"
-     "\xB6\xB5\x0F\x00\xB7\xAD\x0F\x00\xB3\x09\x0E\x00\x1A\xB7\x00\x80\x1E\xB7"
-     "\x00\x80"
-     "\x22\xB7\x00\x80\x26\xB7\x00\x80\xB6\x0D\x0E\x00\xB5\x09\x0E\x00\x2A\xB7"
-     "\x00\x80"
-     "\xBB\x15\x0E\x00\xBA\x15\x0E\x00\x2E\xB7\x00\x80\x32\xB7\x00\x80\xBF\x79"
-     "\x0E\x00"
-     "\xBE\x71\x0E\x00\xBD\x05\x0E\x00\xBC\x05\x0E\x00\x82\x15\x00\x00\xA3\x4D"
-     "\x0E\x00"
-     "\x80\x61\x00\x00\x81\x61\x00\x00\xA6\x49\x0E\x00\x36\xB7\x00\x80\xBE\x10"
-     "\x01\x00"
-     "\xA5\x4D\x0E\x00\xAA\x51\x0E\x00\xAB\x51\x0E\x00\x84\x24\x01\x00\x3E\xB7"
-     "\x00\x80"
-     "\xAE\x35\x0E\x00\xAF\x3D\x0E\x00\xAC\x41\x0E\x00\xAD\x41\x0E\x00\xA8\x39"
-     "\x0E\x00"
-     "\xA9\x39\x0E\x00\xAA\x59\x0E\x00\xAB\x51\x0E\x00\xAC\x71\x0E\x00\xAD\x71"
-     "\x0E\x00"
-     "\xAE\x91\x01\x00\xAF\x91\x01\x00\x86\x00\x00\x00\x87\x84\x00\x00\x42\xB7"
-     "\x00\x80"
-     "\x46\xB7\x00\x80\x4A\xB7\x00\x80\x4E\xB7\x00\x80\x52\xB7\x00\x80\x56\xB7"
-     "\x00\x80"
-     "\xB8\x75\x01\x00\xB9\x7D\x01\x00\xBA\x75\x01\x00\xBB\xC9\x00\x00\xBC\xD9"
-     "\x00\x00"
-     "\xBD\xD9\x00\x00\xBE\xC9\x00\x00\xBF\xC1\x00\x00\xB0\xF5\x01\x00\xB1\xFD"
-     "\x01\x00"
-     "\xB2\xF5\x01\x00\xB3\x4D\x01\x00\xB4\x55\x01\x00\xB5\x5D\x01\x00\xB6\x55"
-     "\x01\x00"
-     "\xB7\x4D\x01\x00\xBA\x4D\x0F\x00\xBB\x55\x0F\x00\xB8\x4D\x0F\x00\xB9\x45"
-     "\x0F\x00"
-     "\xBE\x7D\x0F\x00\xBF\xB5\x0F\x00\xBC\x45\x0F\x00\xBD\x75\x0F\x00\xB2\x01"
-     "\x0F\x00"
-     "\xB3\x01\x0F\x00\xB0\x31\x0F\x00\xB1\x31\x0F\x00\xB6\x01\x0F\x00\xB7\x0D"
-     "\x0F\x00"
-     "\xB4\x11\x0F\x00\xB5\x11\x0F\x00\xAA\x99\x0E\x00\xAB\x45\x0F\x00\xA8\xE5"
-     "\x0E\x00"
-     "\xA9\x99\x0E\x00\xAE\x41\x0F\x00\xAF\x41\x0F\x00\xAC\x51\x0F\x00\xAD\x51"
-     "\x0F\x00"
-     "\x5A\xB7\x00\x80\x5E\xB7\x00\x80\x62\xB7\x00\x80\x66\xB7\x00\x80\x6A\xB7"
-     "\x00\x80"
-     "\x6E\xB7\x00\x80\x72\xB7\x00\x80\x76\xB7\x00\x80\xB3\x51\x0D\x00\x7A\xB7"
-     "\x00\x80"
-     "\x7E\xB7\x00\x80\x82\xB7\x00\x80\x86\xB7\x00\x80\xB6\x71\x0D\x00\xB5\x79"
-     "\x0D\x00"
-     "\x8A\xB7\x00\x80\xBB\xB9\x02\x00\xBA\xB1\x02\x00\x8E\xB7\x00\x80\x92\xB7"
-     "\x00\x80"
-     "\xBF\x19\x02\x00\xBE\x11\x02\x00\xBD\x19\x02\x00\xBC\xA1\x02\x00\x96\xB7"
-     "\x00\x80"
-     "\xA3\x15\x0D\x00\x9A\xB7\x00\x80\x9E\xB7\x00\x80\xA6\x35\x0D\x00\xA2\xB7"
-     "\x00\x80"
-     "\xA6\xB7\x00\x80\xA5\x3D\x0D\x00\xAA\xF5\x02\x00\xAB\xFD\x02\x00\x84\xE8"
-     "\x03\x00"
-     "\xAE\xB7\x00\x80\xAE\x55\x02\x00\xAF\x5D\x02\x00\xAC\xE5\x02\x00\xAD\x5D"
-     "\x02\x00"
-     "\xA8\x6D\x02\x00\xA9\xAD\x02\x00\xAA\xA5\x02\x00\xAB\xBD\x02\x00\xAC\xA5"
-     "\x02\x00"
-     "\xAD\xAD\x02\x00\xAE\xA5\x02\x00\xAF\x7D\x01\x00\x80\xED\x01\x00\x81\xF1"
-     "\x01\x00"
-     "\x82\xF1\x01\x00\xBE\xA0\x05\x00\xB2\xB7\x00\x80\xB6\xB7\x00\x80\x87\x68"
-     "\x05\x00"
-     "\x86\x1C\x05\x00\xB8\xC9\x01\x00\xB9\xC9\x01\x00\xBA\xD9\x01\x00\xBB\xD1"
-     "\x01\x00"
-     "\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\x99\x01\x00\xBF\x95\x01\x00\xB0\x05"
-     "\x01\x00"
-     "\xB1\x0D\x01\x00\xB2\x05\x01\x00\xB3\x1D\x01\x00\xB4\x05\x01\x00\xB5\x0D"
-     "\x01\x00"
-     "\xB6\x05\x01\x00\xB7\xF9\x01\x00\xE1\x64\x0F\x00\xE1\x9C\x0F\x00\xE3\x14"
-     "\x0E\x00"
-     "\xE3\xD4\x0F\x00\xBA\xB7\x00\x80\xE1\x3C\x0E\x00\xBE\xB7\x00\x80\xE3\xE4"
-     "\x00\x00"
-     "\xBE\xAC\x04\x00\xC2\xB7\x00\x80\xC6\xB7\x00\x80\xEF\x0C\x00\x00\xCA\xB7"
-     "\x00\x80"
-     "\xCE\xB7\x00\x80\xEF\x60\x0E\x00\xEF\xB4\x0F\x00\xD2\xB7\x00\x80\xD6\xB7"
-     "\x00\x80"
-     "\x84\x44\x04\x00\xB3\x61\x02\x00\xDA\xB7\x00\x80\xB5\x61\x02\x00\xB6\x61"
-     "\x02\x00"
-     "\xDE\xB7\x00\x80\xE2\xB7\x00\x80\xE6\xB7\x00\x80\xBA\x8D\x01\x00\xBB\x85"
-     "\x01\x00"
-     "\xBC\x9D\x01\x00\xBD\x85\x01\x00\xBE\x8D\x01\x00\xBF\x85\x01\x00\xA3\xAD"
-     "\x05\x00"
-     "\xAA\xB7\x00\x80\xEA\xB7\x00\x80\xEE\xB7\x00\x80\xF2\xB7\x00\x80\xA6\xAD"
-     "\x05\x00"
-     "\xA5\xAD\x05\x00\xF6\xB7\x00\x80\xAB\x49\x06\x00\xAA\x41\x06\x00\xFA\xB7"
-     "\x00\x80"
-     "\xFE\xB7\x00\x80\xAF\x49\x06\x00\xAE\x41\x06\x00\xAD\x49\x06\x00\xAC\x51"
-     "\x06\x00"
-     "\x02\xB8\x00\x80\x06\xB8\x00\x80\x0A\xB8\x00\x80\x0E\xB8\x00\x80\x80\x1D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x39\x00\x00\x12\xB8\x00\x80\x16\xB8\x00\x80\x1A\xB8"
-     "\x00\x80"
-     "\x86\xC8\x00\x00\x87\x20\x03\x00\x1E\xB8\x00\x80\x22\xB8\x00\x80\x26\xB8"
-     "\x00\x80"
-     "\x2A\xB8\x00\x80\xA8\x6D\x06\x00\xA9\xB5\x07\x00\xAA\xBD\x07\x00\xAB\x1D"
-     "\x07\x00"
-     "\xAC\x09\x07\x00\xAD\x31\x07\x00\xAE\x31\x07\x00\xAF\x2D\x07\x00\x84\xA8"
-     "\x03\x00"
-     "\x2E\xB8\x00\x80\x32\xB8\x00\x80\x36\xB8\x00\x80\x3A\xB8\x00\x80\x3E\xB8"
-     "\x00\x80"
-     "\x42\xB8\x00\x80\x46\xB8\x00\x80\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5"
-     "\x00\x00"
-     "\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x9D\x00\x00\xBF\x95"
-     "\x00\x00"
-     "\xB0\x55\x07\x00\xB1\x25\x07\x00\xB2\x2D\x07\x00\xB3\x3D\x07\x00\xB4\x2D"
-     "\x07\x00"
-     "\xB5\x15\x07\x00\xB6\x1D\x07\x00\xB7\xF5\x00\x00\x4A\xB8\x00\x80\xE1\xBC"
-     "\x06\x00"
-     "\x4E\xB8\x00\x80\xE3\xF4\x05\x00\x52\xB8\x00\x80\x56\xB8\x00\x80\x5A\xB8"
-     "\x00\x80"
-     "\x5E\xB8\x00\x80\x62\xB8\x00\x80\x66\xB8\x00\x80\x6A\xB8\x00\x80\x6E\xB8"
-     "\x00\x80"
-     "\x72\xB8\x00\x80\x76\xB8\x00\x80\x7A\xB8\x00\x80\xEF\xE8\x04\x00\xB3\x75"
-     "\x06\x00"
-     "\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D\x00\x00\x7E\xB8\x00\x80\xB6\xF1"
-     "\x06\x00"
-     "\xB5\xC1\x06\x00\x82\xB8\x00\x80\xBB\xA1\x06\x00\xBA\xD1\x06\x00\x86\xB8"
-     "\x00\x80"
-     "\xBE\x6C\x01\x00\xBF\x91\x06\x00\xBE\xA9\x06\x00\xBD\xA9\x06\x00\xBC\xB9"
-     "\x06\x00"
-     "\xA3\xB5\x06\x00\x8E\xB8\x00\x80\x86\x28\x00\x00\x87\x4C\x01\x00\x92\xB8"
-     "\x00\x80"
-     "\xA6\x31\x06\x00\xA5\x01\x06\x00\x96\xB8\x00\x80\xAB\x61\x06\x00\xAA\x11"
-     "\x06\x00"
-     "\x9A\xB8\x00\x80\x9E\xB8\x00\x80\xAF\x51\x06\x00\xAE\x69\x06\x00\xAD\x69"
-     "\x06\x00"
-     "\xAC\x79\x06\x00\xA2\xB8\x00\x80\xB3\xBD\x01\x00\xA6\xB8\x00\x80\xAA\xB8"
-     "\x00\x80"
-     "\xB6\x79\x01\x00\xAE\xB8\x00\x80\xB2\xB8\x00\x80\xB5\x79\x01\x00\xBA\x55"
-     "\x01\x00"
-     "\xBB\x5D\x01\x00\xB6\xB8\x00\x80\xBA\xB8\x00\x80\xBE\xF9\x00\x00\xBF\xE5"
-     "\x00\x00"
-     "\xBC\x45\x01\x00\xBD\xFD\x00\x00\xA8\x71\x02\x00\xA9\x71\x02\x00\xAA\x71"
-     "\x02\x00"
-     "\xAB\x71\x02\x00\xAC\xB5\x02\x00\xAD\xBD\x02\x00\xAE\xB5\x02\x00\xAF\xAD"
-     "\x02\x00"
-     "\x84\xEC\x3C\x00\xBE\xB8\x00\x80\xC2\xB8\x00\x80\xC6\xB8\x00\x80\xCA\xB8"
-     "\x00\x80"
-     "\xCE\xB8\x00\x80\xD2\xB8\x00\x80\xD6\xB8\x00\x80\xB8\x5D\x03\x00\xB9\x65"
-     "\x03\x00"
-     "\xBA\x6D\x03\x00\xBB\x65\x03\x00\xBC\x7D\x03\x00\xBD\x65\x03\x00\xBE\x6D"
-     "\x03\x00"
-     "\xBF\x65\x03\x00\xB0\xD5\x02\x00\xB1\xDD\x02\x00\xB2\xD5\x02\x00\xB3\x6D"
-     "\x03\x00"
-     "\xB4\x79\x03\x00\xB5\x65\x03\x00\xB6\x6D\x03\x00\xB7\x65\x03\x00\x3A\xB7"
-     "\x00\x80"
-     "\xDA\xB8\x00\x80\xDE\xB8\x00\x80\xA3\xF5\x02\x00\xE2\xB8\x00\x80\xA5\x31"
-     "\x02\x00"
-     "\xA6\x31\x02\x00\xE6\xB8\x00\x80\xEA\xB8\x00\x80\xEE\xB8\x00\x80\xAA\x1D"
-     "\x02\x00"
-     "\xAB\x15\x02\x00\xAC\x0D\x02\x00\xAD\xB5\x03\x00\xAE\xB1\x03\x00\xAF\xAD"
-     "\x03\x00"
-     "\x80\x65\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\xF2\xB8\x00\x80\xF6\xB8"
-     "\x00\x80"
-     "\xFE\xB8\x00\x80\xBE\x10\x3C\x00\x02\xB9\x00\x80\x86\xC0\x3C\x00\x87\x20"
-     "\x03\x00"
-     "\x06\xB9\x00\x80\x0A\xB9\x00\x80\x0E\xB9\x00\x80\x12\xB9\x00\x80\x16\xB9"
-     "\x00\x80"
-     "\x1A\xB9\x00\x80\xA8\x85\x02\x00\xA9\x95\x02\x00\xAA\x95\x02\x00\xAB\xA5"
-     "\x02\x00"
-     "\xAC\xBD\x02\x00\xAD\xD5\x02\x00\xAE\xD1\x02\x00\xAF\xD1\x02\x00\x1E\xB9"
-     "\x00\x80"
-     "\x22\xB9\x00\x80\x26\xB9\x00\x80\x2A\xB9\x00\x80\x2E\xB9\x00\x80\x32\xB9"
-     "\x00\x80"
-     "\x36\xB9\x00\x80\x3A\xB9\x00\x80\xB8\x75\x01\x00\xB9\x7D\x01\x00\xBA\x75"
-     "\x01\x00"
-     "\xBB\xC9\x01\x00\xBC\xD9\x01\x00\xBD\xC5\x01\x00\xBE\xC5\x01\x00\xBF\xFD"
-     "\x01\x00"
-     "\xB0\xB5\x02\x00\xB1\xBD\x02\x00\xB2\x81\x02\x00\xB3\x81\x02\x00\xB4\x55"
-     "\x01\x00"
-     "\xB5\x5D\x01\x00\xB6\x55\x01\x00\xB7\x4D\x01\x00\xE1\xA4\x06\x00\x3E\xB9"
-     "\x00\x80"
-     "\xE3\x50\x06\x00\xBE\x80\x3C\x00\x84\x1C\x3C\x00\xBE\x80\x3F\x00\x42\xB9"
-     "\x00\x80"
-     "\x46\xB9\x00\x80\x4A\xB9\x00\x80\x4E\xB9\x00\x80\x52\xB9\x00\x80\x56\xB9"
-     "\x00\x80"
-     "\x5A\xB9\x00\x80\x5E\xB9\x00\x80\xEF\xE0\x06\x00\x62\xB9\x00\x80\x81\x7D"
-     "\x00\x00"
-     "\x80\x71\x00\x00\x66\xB9\x00\x80\x82\x05\x00\x00\x6E\xB9\x00\x80\x72\xB9"
-     "\x00\x80"
-     "\xEF\x4C\x00\x00\x76\xB9\x00\x80\xE1\x90\x01\x00\x7A\xB9\x00\x80\xE3\x58"
-     "\x01\x00"
-     "\x7E\xB9\x00\x80\x82\xB9\x00\x80\x86\xB9\x00\x80\x86\x58\x3F\x00\x87\xFC"
-     "\x3C\x00"
-     "\xB3\x9D\x3D\x00\xFA\xB8\x00\x80\x6A\xB9\x00\x80\x8A\xB9\x00\x80\x8E\xB9"
-     "\x00\x80"
-     "\xB6\xD5\x3D\x00\xB5\xB5\x3D\x00\x92\xB9\x00\x80\xBB\xF9\x3D\x00\xBA\xF1"
-     "\x3D\x00"
-     "\x96\xB9\x00\x80\x9A\xB9\x00\x80\xBF\x19\x3E\x00\xBE\x11\x3E\x00\xBD\xD5"
-     "\x3D\x00"
-     "\xBC\xD5\x3D\x00\x9E\xB9\x00\x80\xA3\xD9\x3D\x00\xA2\xB9\x00\x80\xA6\xB9"
-     "\x00\x80"
-     "\xA6\x91\x3D\x00\xAA\xB9\x00\x80\xAE\xB9\x00\x80\xA5\xF1\x3D\x00\xAA\xB5"
-     "\x3D\x00"
-     "\xAB\xBD\x3D\x00\xB2\xB9\x00\x80\xB6\xB9\x00\x80\xAE\x55\x3E\x00\xAF\x5D"
-     "\x3E\x00"
-     "\xAC\x91\x3D\x00\xAD\x91\x3D\x00\xA8\x55\x3E\x00\xA9\x59\x3E\x00\xAA\x61"
-     "\x3E\x00"
-     "\xAB\x61\x3E\x00\xAC\x61\x3E\x00\xAD\x61\x3E\x00\xAE\x61\x3E\x00\xAF\x61"
-     "\x3E\x00"
-     "\x84\xA8\x03\x00\xBA\xB9\x00\x80\xBE\xB9\x00\x80\xC2\xB9\x00\x80\xC6\xB9"
-     "\x00\x80"
-     "\xCA\xB9\x00\x80\xCE\xB9\x00\x80\xD2\xB9\x00\x80\xB8\x45\x3F\x00\xB9\x5D"
-     "\x3F\x00"
-     "\xBA\x55\x3F\x00\xBB\x6D\x3F\x00\xBC\x75\x3F\x00\xBD\x7D\x3F\x00\xBE\x75"
-     "\x3F\x00"
-     "\xBF\x6D\x3F\x00\xB0\xC1\x3F\x00\xB1\xC1\x3F\x00\xB2\xC1\x3F\x00\xB3\xC1"
-     "\x3F\x00"
-     "\xB4\xC1\x3F\x00\xB5\xC1\x3F\x00\xB6\xC1\x3F\x00\xB7\xC1\x3F\x00\x80\xB9"
-     "\x01\x00"
-     "\x81\xB9\x01\x00\x82\x05\x00\x00\xD6\xB9\x00\x80\xE1\x80\x3E\x00\xDE\xB9"
-     "\x00\x80"
-     "\xE3\x28\x3D\x00\xE2\xB9\x00\x80\x86\x80\x00\x00\x87\x04\x01\x00\xEF\x08"
-     "\x3D\x00"
-     "\xE6\xB9\x00\x80\xEA\xB9\x00\x80\xEE\xB9\x00\x80\xF2\xB9\x00\x80\xF6\xB9"
-     "\x00\x80"
-     "\xB3\xA9\x3F\x00\xDA\xB9\x00\x80\xFA\xB9\x00\x80\xFE\xB9\x00\x80\x02\xBA"
-     "\x00\x80"
-     "\xB6\xA1\x3F\x00\xB5\xA9\x3F\x00\x06\xBA\x00\x80\xBB\x45\x3E\x00\xBA\x45"
-     "\x3E\x00"
-     "\x0A\xBA\x00\x80\x0E\xBA\x00\x80\xBF\x45\x3E\x00\xBE\x45\x3E\x00\xBD\x55"
-     "\x3E\x00"
-     "\xBC\x55\x3E\x00\xA3\x69\x3F\x00\x12\xBA\x00\x80\x16\xBA\x00\x80\x1A\xBA"
-     "\x00\x80"
-     "\x1E\xBA\x00\x80\xA6\x61\x3F\x00\xA5\x69\x3F\x00\x22\xBA\x00\x80\xAB\x85"
-     "\x3E\x00"
-     "\xAA\x85\x3E\x00\x26\xBA\x00\x80\x2A\xBA\x00\x80\xAF\x85\x3E\x00\xAE\x85"
-     "\x3E\x00"
-     "\xAD\x95\x3E\x00\xAC\x95\x3E\x00\x2E\xBA\x00\x80\xB3\x19\x3E\x00\x32\xBA"
-     "\x00\x80"
-     "\x36\xBA\x00\x80\xB6\x21\x3E\x00\x3A\xBA\x00\x80\x3E\xBA\x00\x80\xB5\x31"
-     "\x3E\x00"
-     "\xBA\xF1\x01\x00\xBB\xF9\x01\x00\x42\xBA\x00\x80\x46\xBA\x00\x80\xBE\xB1"
-     "\x01\x00"
-     "\xBF\xBD\x01\x00\xBC\xD1\x01\x00\xBD\xD1\x01\x00\xA8\xD1\x3D\x00\xA9\xD1"
-     "\x3D\x00"
-     "\xAA\xD5\x3D\x00\xAB\xE9\x3D\x00\xAC\xFD\x3D\x00\xAD\xE5\x3D\x00\xAE\xED"
-     "\x3D\x00"
-     "\xAF\x11\x02\x00\x80\xF9\x03\x00\x81\xCD\x03\x00\x82\xC5\x03\x00\x84\x24"
-     "\x03\x00"
-     "\xBE\x00\x04\x00\x4E\xBA\x00\x80\x87\xAC\x03\x00\x86\xBC\x04\x00\xB8\x19"
-     "\x02\x00"
-     "\xB9\x2D\x02\x00\xBA\x25\x02\x00\xBB\xE9\x02\x00\xBC\xF9\x02\x00\xBD\xF9"
-     "\x02\x00"
-     "\xBE\xE9\x02\x00\xBF\xE9\x02\x00\xB0\x71\x02\x00\xB1\x79\x02\x00\xB2\x41"
-     "\x02\x00"
-     "\xB3\x41\x02\x00\xB4\x31\x02\x00\xB5\x3D\x02\x00\xB6\x35\x02\x00\xB7\x29"
-     "\x02\x00"
-     "\xA5\x6D\x3D\x00\x52\xBA\x00\x80\x56\xBA\x00\x80\xA6\x7D\x3D\x00\x5A\xBA"
-     "\x00\x80"
-     "\x8A\xB8\x00\x80\xA3\x45\x3D\x00\x5E\xBA\x00\x80\xAD\x8D\x02\x00\xAC\x8D"
-     "\x02\x00"
-     "\xAF\xE1\x02\x00\xAE\xED\x02\x00\xAC\x00\x00\x00\x62\xBA\x00\x80\xAB\xA5"
-     "\x02\x00"
-     "\xAA\xAD\x02\x00\xE1\xF8\x01\x00\x66\xBA\x00\x80\xE3\xF4\x02\x00\x84\xC0"
-     "\x04\x00"
-     "\x6A\xBA\x00\x80\x6E\xBA\x00\x80\x72\xBA\x00\x80\x76\xBA\x00\x80\x7A\xBA"
-     "\x00\x80"
-     "\x7E\xBA\x00\x80\x82\xBA\x00\x80\x86\xBA\x00\x80\x8A\xBA\x00\x80\x8E\xBA"
-     "\x00\x80"
-     "\xEF\x30\x02\x00\x92\xBA\x00\x80\xA8\x15\x02\x00\xA9\x19\x02\x00\xAA\x25"
-     "\x02\x00"
-     "\xAB\x3D\x02\x00\xAC\x25\x02\x00\xAD\x2D\x02\x00\xAE\x25\x02\x00\xAF\x55"
-     "\x02\x00"
-     "\x96\xBA\x00\x80\x9A\xBA\x00\x80\x9E\xBA\x00\x80\xA2\xBA\x00\x80\xA6\xBA"
-     "\x00\x80"
-     "\xAA\xBA\x00\x80\x84\xB0\x04\x00\xAE\xBA\x00\x80\xB8\xD1\x02\x00\xB9\xD9"
-     "\x02\x00"
-     "\xBA\xE1\x02\x00\xBB\xE1\x02\x00\xBC\x91\x02\x00\xBD\x9D\x02\x00\xBE\x95"
-     "\x02\x00"
-     "\xBF\x89\x02\x00\xB0\x2D\x02\x00\xB1\x35\x02\x00\xB2\x35\x02\x00\xB3\x05"
-     "\x02\x00"
-     "\xB4\x1D\x02\x00\xB5\xF1\x02\x00\xB6\xF1\x02\x00\xB7\xF1\x02\x00\xE1\x78"
-     "\x3F\x00"
-     "\xE3\x34\x01\x00\xE3\x08\x3E\x00\xE1\x6C\x3E\x00\x81\x09\x00\x00\x80\xA9"
-     "\x00\x00"
-     "\xB2\xBA\x00\x80\x82\x3D\x00\x00\xB6\xBA\x00\x80\xBE\xBA\x00\x80\xC2\xBA"
-     "\x00\x80"
-     "\xBE\x20\x04\x00\xC6\xBA\x00\x80\xEF\xD8\x3E\x00\xEF\xCC\x3E\x00\xCA\xBA"
-     "\x00\x80"
-     "\xCE\xBA\x00\x80\xB3\xE9\x02\x00\x86\xE8\x04\x00\x87\xC0\x04\x00\xB6\xE9"
-     "\x02\x00"
-     "\xD2\xBA\x00\x80\xD6\xBA\x00\x80\xB5\xE9\x02\x00\xBA\xAD\x02\x00\xBB\xB5"
-     "\x02\x00"
-     "\xDA\xBA\x00\x80\xDE\xBA\x00\x80\xBE\x9D\x02\x00\xBF\x71\x02\x00\xBC\xA5"
-     "\x02\x00"
-     "\xBD\x95\x02\x00\x4A\xBA\x00\x80\xBA\xBA\x00\x80\xE2\xBA\x00\x80\xE6\xBA"
-     "\x00\x80"
-     "\xEA\xBA\x00\x80\xEE\xBA\x00\x80\xF2\xBA\x00\x80\xF6\xBA\x00\x80\xA8\x15"
-     "\x06\x00"
-     "\xA9\xA1\x06\x00\xAA\xA1\x06\x00\xAB\xBD\x06\x00\xAC\xAD\x06\x00\xAD\x81"
-     "\x06\x00"
-     "\xAE\xFD\x06\x00\xAF\xED\x06\x00\xB0\x95\x06\x00\xB1\x9D\x06\x00\xB2\x95"
-     "\x06\x00"
-     "\xB3\xAD\x06\x00\xB4\xB5\x06\x00\xB5\xBD\x06\x00\xB6\xB5\x06\x00\xB7\xAD"
-     "\x06\x00"
-     "\xB8\x95\x06\x00\xB9\x99\x06\x00\xBA\x49\x07\x00\xBB\x49\x07\x00\xBC\x59"
-     "\x07\x00"
-     "\xBD\x59\x07\x00\xBE\x49\x07\x00\xBF\x49\x07\x00\xAC\xDD\x05\x00\xAD\xED"
-     "\x05\x00"
-     "\xAE\xE5\x05\x00\xAF\x09\x05\x00\xFA\xBA\x00\x80\xFE\xBA\x00\x80\xAA\xD5"
-     "\x05\x00"
-     "\xAB\xCD\x05\x00\x02\xBB\x00\x80\xA5\x91\x05\x00\xA6\x91\x05\x00\x06\xBB"
-     "\x00\x80"
-     "\x0A\xBB\x00\x80\x0E\xBB\x00\x80\x12\xBB\x00\x80\xA3\x91\x05\x00\xB3\x49"
-     "\x06\x00"
-     "\x16\xBB\x00\x80\x1A\xBB\x00\x80\x1E\xBB\x00\x80\x22\xBB\x00\x80\xB6\x61"
-     "\x06\x00"
-     "\xB5\x45\x06\x00\x26\xBB\x00\x80\xBB\x39\x06\x00\xBA\x31\x06\x00\xBE\x64"
-     "\x00\x00"
-     "\x2A\xBB\x00\x80\xBF\x19\x06\x00\xBE\x11\x06\x00\xBD\x19\x06\x00\xBC\x21"
-     "\x06\x00"
-     "\xA3\x89\x07\x00\x82\xD9\x01\x00\x81\xD9\x01\x00\x80\xC1\x01\x00\x2E\xBB"
-     "\x00\x80"
-     "\xA6\xA1\x07\x00\xA5\x85\x07\x00\x32\xBB\x00\x80\xAB\xF9\x07\x00\xAA\xF1"
-     "\x07\x00"
-     "\x86\x08\x01\x00\x87\x7C\x01\x00\xAF\xD9\x07\x00\xAE\xD1\x07\x00\xAD\xD9"
-     "\x07\x00"
-     "\xAC\xE1\x07\x00\x36\xBB\x00\x80\xB3\xF5\x06\x00\x3A\xBB\x00\x80\x3E\xBB"
-     "\x00\x80"
-     "\xB6\x85\x06\x00\x42\xBB\x00\x80\x46\xBB\x00\x80\xB5\x9D\x06\x00\xBA\x8D"
-     "\x06\x00"
-     "\xBB\x6D\x01\x00\x4A\xBB\x00\x80\x4E\xBB\x00\x80\xBE\x65\x01\x00\xBF\x6D"
-     "\x01\x00"
-     "\xBC\x75\x01\x00\xBD\x6D\x01\x00\xA8\x25\x06\x00\xA9\x2D\x06\x00\xAA\x39"
-     "\x06\x00"
-     "\xAB\x31\x06\x00\xAC\x51\x06\x00\xAD\x41\x06\x00\xAE\x41\x06\x00\xAF\x75"
-     "\x06\x00"
-     "\x52\xBB\x00\x80\x56\xBB\x00\x80\x5A\xBB\x00\x80\x5E\xBB\x00\x80\x62\xBB"
-     "\x00\x80"
-     "\x66\xBB\x00\x80\x6A\xBB\x00\x80\x6E\xBB\x00\x80\xB8\x55\x01\x00\xB9\x65"
-     "\x01\x00"
-     "\xBA\x65\x01\x00\xBB\x7D\x01\x00\xBC\x65\x01\x00\xBD\x6D\x01\x00\xBE\x1D"
-     "\x01\x00"
-     "\xBF\x15\x01\x00\xB0\x0D\x06\x00\xB1\xED\x01\x00\xB2\xE5\x01\x00\xB3\xFD"
-     "\x01\x00"
-     "\xB4\xE5\x01\x00\xB5\xED\x01\x00\xB6\xE5\x01\x00\xB7\x6D\x01\x00\xA3\xB1"
-     "\x05\x00"
-     "\x72\xBB\x00\x80\x76\xBB\x00\x80\xBE\x48\x03\x00\xBE\x58\x0C\x00\xA6\xC1"
-     "\x05\x00"
-     "\xA5\xD9\x05\x00\x7A\xBB\x00\x80\xAB\x29\x02\x00\xAA\xC9\x05\x00\x7E\xBB"
-     "\x00\x80"
-     "\x82\xBB\x00\x80\xAF\x29\x02\x00\xAE\x21\x02\x00\xAD\x29\x02\x00\xAC\x31"
-     "\x02\x00"
-     "\x86\xBB\x00\x80\x8A\xBB\x00\x80\x8E\xBB\x00\x80\x92\xBB\x00\x80\x80\x19"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x82\x05\x00\x00\x96\xBB\x00\x80\x84\xAC\x03\x00\x9E\xBB"
-     "\x00\x80"
-     "\x87\x18\x03\x00\x86\xCC\x0C\x00\xA2\xBB\x00\x80\xA6\xBB\x00\x80\xAA\xBB"
-     "\x00\x80"
-     "\xAE\xBB\x00\x80\xA8\x91\x03\x00\xA9\x99\x03\x00\xAA\xC9\x03\x00\xAB\xC5"
-     "\x03\x00"
-     "\xAC\xDD\x03\x00\xAD\xC1\x03\x00\xAE\xC1\x03\x00\xAF\xF5\x03\x00\xB2\xBB"
-     "\x00\x80"
-     "\xB6\xBB\x00\x80\xBA\xBB\x00\x80\xBE\xBB\x00\x80\xC2\xBB\x00\x80\xC6\xBB"
-     "\x00\x80"
-     "\xCA\xBB\x00\x80\xCE\xBB\x00\x80\xB8\x7D\x03\x00\xB9\xC1\x00\x00\xBA\xC1"
-     "\x00\x00"
-     "\xBB\xD1\x00\x00\xBC\xF9\x00\x00\xBD\xF9\x00\x00\xBE\x99\x00\x00\xBF\x99"
-     "\x00\x00"
-     "\xB0\x8D\x03\x00\xB1\x45\x03\x00\xB2\x4D\x03\x00\xB3\x45\x03\x00\xB4\x5D"
-     "\x03\x00"
-     "\xB5\x45\x03\x00\xB6\x4D\x03\x00\xB7\x45\x03\x00\xB3\x41\x02\x00\xD2\xBB"
-     "\x00\x80"
-     "\xD6\xBB\x00\x80\xBF\x04\x0F\x00\xDA\xBB\x00\x80\xB6\x41\x02\x00\xB5\x55"
-     "\x02\x00"
-     "\xDE\xBB\x00\x80\xBB\x81\x02\x00\xBA\x49\x02\x00\xE2\xBB\x00\x80\xE6\xBB"
-     "\x00\x80"
-     "\xBF\x81\x02\x00\xBE\x99\x02\x00\xBD\x91\x02\x00\xBC\x99\x02\x00\xEA\xBB"
-     "\x00\x80"
-     "\xEE\xBB\x00\x80\xF2\xBB\x00\x80\xF6\xBB\x00\x80\xEF\xA4\x03\x00\xFA\xBB"
-     "\x00\x80"
-     "\xFE\xBB\x00\x80\x02\xBC\x00\x80\xE3\x10\x03\x00\x06\xBC\x00\x80\xE1\x58"
-     "\x00\x00"
-     "\x84\x20\x0D\x00\x80\x29\x00\x00\x81\x29\x00\x00\x82\x1D\x00\x00\x0E\xBC"
-     "\x00\x80"
-     "\xE1\x50\x06\x00\xE1\xA0\x07\x00\xE3\xA0\x06\x00\xE3\x50\x07\x00\x85\x94"
-     "\x0C\x00"
-     "\x12\xBC\x00\x80\xEF\x48\x1B\x00\x16\xBC\x00\x80\xE1\x24\x02\x00\x1A\xBC"
-     "\x00\x80"
-     "\xE3\x70\x1A\x00\x1E\xBC\x00\x80\x22\xBC\x00\x80\x26\xBC\x00\x80\xEF\xA8"
-     "\x01\x00"
-     "\xEF\xE8\x06\x00\x86\xA0\x0F\x00\x87\x0C\x0D\x00\xA3\x89\x02\x00\x2A\xBC"
-     "\x00\x80"
-     "\xA5\x9D\x02\x00\x2E\xBC\x00\x80\x32\xBC\x00\x80\xA6\x89\x02\x00\x36\xBC"
-     "\x00\x80"
-     "\x3A\xBC\x00\x80\xAB\x49\x02\x00\xAA\x81\x02\x00\xAD\x59\x02\x00\xAC\x51"
-     "\x02\x00"
-     "\xAF\x49\x02\x00\xAE\x51\x02\x00\xA8\x65\x0E\x00\xA9\x75\x0E\x00\xAA\x7D"
-     "\x0E\x00"
-     "\xAB\x75\x0E\x00\xAC\x6D\x0E\x00\xAD\xB5\x0E\x00\xAE\xBD\x0E\x00\xAF\xB5"
-     "\x0E\x00"
-     "\x0A\xBC\x00\x80\x3E\xBC\x00\x80\x42\xBC\x00\x80\x46\xBC\x00\x80\x4A\xBC"
-     "\x00\x80"
-     "\x55\xBC\x00\x80\x59\xBC\x00\x80\x5D\xBC\x00\x80\xB8\x5D\x0F\x00\xB9\x6D"
-     "\x0F\x00"
-     "\xBA\x65\x0F\x00\xBB\x0D\x0F\x00\xBC\x15\x0F\x00\xBD\x1D\x0F\x00\xBE\x15"
-     "\x0F\x00"
-     "\xBF\x09\x0F\x00\xB0\xCD\x0E\x00\xB1\xD5\x0E\x00\xB2\xDD\x0E\x00\xB3\xD5"
-     "\x0E\x00"
-     "\xB4\xCD\x0E\x00\xB5\x71\x0F\x00\xB6\x71\x0F\x00\xB7\x6D\x0F\x00\xB3\xF5"
-     "\x0E\x00"
-     "\x61\xBC\x00\x80\x65\xBC\x00\x80\x69\xBC\x00\x80\x6D\xBC\x00\x80\xB6\x35"
-     "\x0E\x00"
-     "\xB5\xE5\x0E\x00\x71\xBC\x00\x80\xBB\x11\x0E\x00\xBA\x09\x0E\x00\x75\xBC"
-     "\x00\x80"
-     "\x79\xBC\x00\x80\xBF\xB5\x0F\x00\xBE\x09\x0E\x00\xBD\x01\x0E\x00\xBC\x09"
-     "\x0E\x00"
-     "\x82\x15\x00\x00\xA3\xB1\x0E\x00\x80\x61\x00\x00\x81\x61\x00\x00\xA6\x71"
-     "\x0E\x00"
-     "\x7D\xBC\x00\x80\xBE\x10\x01\x00\xA5\xA1\x0E\x00\xAA\x4D\x0E\x00\xAB\x55"
-     "\x0E\x00"
-     "\x85\xBC\x00\x80\x84\x20\x01\x00\xAE\x4D\x0E\x00\xAF\xF1\x0F\x00\xAC\x4D"
-     "\x0E\x00"
-     "\xAD\x45\x0E\x00\x9F\x45\x08\x00\x9E\x0D\x09\x00\x9D\x0D\x09\x00\x9C\x99"
-     "\x0B\x00"
-     "\x9B\x75\x35\x00\x9A\x11\x35\x00\x99\x99\x37\x00\x98\x0D\x31\x00\x97\x25"
-     "\x31\x00"
-     "\x96\x25\x31\x00\x95\x81\x3D\x00\x94\x0D\x3D\x00\x93\x89\x3F\x00\x92\x15"
-     "\x38\x00"
-     "\x91\x3D\x39\x00\x90\x3D\x39\x00\x8F\x65\x25\x00\xEF\xAC\x0D\x00\x86\x00"
-     "\x04\x00"
-     "\x87\xA0\x01\x00\x89\xBC\x00\x80\x8D\xBC\x00\x80\x91\xBC\x00\x80\xEF\xE8"
-     "\x01\x00"
-     "\x95\xBC\x00\x80\xE1\x34\x02\x00\x99\xBC\x00\x80\xE3\x34\x01\x00\x9D\xBC"
-     "\x00\x80"
-     "\xE3\x08\x0C\x00\xA1\xBC\x00\x80\xE1\x08\x0D\x00\xA1\xA1\x01\x00\xA5\xBC"
-     "\x00\x80"
-     "\xA3\x09\x05\x00\xA2\x6D\x03\x00\xA5\xCD\x04\x00\xA4\x11\x05\x00\xA7\x1D"
-     "\x19\x00"
-     "\xA6\x1D\x19\x00\xA9\xA1\x1D\x00\xA8\x39\x19\x00\xAB\xE9\x1C\x00\xAA\xA9"
-     "\x1D\x00"
-     "\xAD\x91\x11\x00\xAC\x01\x10\x00\xAF\x01\x14\x00\xAE\x51\x11\x00\xB1\xF9"
-     "\x15\x00"
-     "\xB0\xE5\x15\x00\xB3\xE9\x69\x00\xB2\x01\x68\x00\xB5\x01\x6C\x00\xB4\x79"
-     "\x69\x00"
-     "\xA9\xBC\x00\x80\xAD\xBC\x00\x80\xB1\xBC\x00\x80\xB5\xBC\x00\x80\xB9\xBC"
-     "\x00\x80"
-     "\xBD\xBC\x00\x80\xA3\x09\x03\x00\xA2\x19\x0D\x00\xA1\xFD\x0D\x00\xA0\xFD"
-     "\x0D\x00"
-     "\x82\x25\x26\x00\x83\x05\x3A\x00\xC1\xBC\x00\x80\xC5\xBC\x00\x80\x86\xA9"
-     "\x3C\x00"
-     "\x87\x35\x3E\x00\x84\x1D\x3A\x00\x85\x3D\x3B\x00\x8A\x89\x3E\x00\x8B\x25"
-     "\x32\x00"
-     "\xC9\xBC\x00\x80\xCD\xBC\x00\x80\x8E\xB1\x34\x00\x8F\x31\x36\x00\x8C\x3D"
-     "\x32\x00"
-     "\x8D\x2D\x32\x00\x92\x25\x36\x00\x93\xD1\x08\x00\x84\x44\x03\x00\xBE\xC0"
-     "\x04\x00"
-     "\x96\x15\x0B\x00\x97\x55\x0E\x00\x94\x5D\x0A\x00\x95\x55\x0A\x00\x9A\x65"
-     "\x0E\x00"
-     "\x9B\x89\x0E\x00\xD1\xBC\x00\x80\xD5\xBC\x00\x80\xD9\xBC\x00\x80\xDD\xBC"
-     "\x00\x80"
-     "\x9C\x81\x00\x00\xE1\xBC\x00\x80\xB8\xB5\x02\x00\xB9\xBD\x02\x00\xBA\xB5"
-     "\x02\x00"
-     "\xBB\x09\x02\x00\xBC\x19\x02\x00\xBD\x19\x02\x00\xBE\x09\x02\x00\xBF\x01"
-     "\x02\x00"
-     "\xB0\x75\x0D\x00\xB1\x7D\x0D\x00\xB2\x49\x0D\x00\xB3\x49\x0D\x00\xB4\x95"
-     "\x02\x00"
-     "\xB5\x9D\x02\x00\xB6\x95\x02\x00\xB7\x8D\x02\x00\xA8\xBD\x0D\x00\xA9\x51"
-     "\x0D\x00"
-     "\xAA\x55\x0D\x00\xAB\x69\x0D\x00\xAC\x7D\x0D\x00\xAD\x65\x0D\x00\xAE\x6D"
-     "\x0D\x00"
-     "\xAF\x11\x0D\x00\x81\xBC\x00\x80\x82\xED\x01\x00\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xE5\xBC\x00\x80\xE9\xBC\x00\x80\x9A\xBB\x00\x80\xBE\x70\x05\x00\xB3\x9D"
-     "\x0C\x00"
-     "\x84\x8C\x05\x00\xED\xBC\x00\x80\xF5\xBC\x00\x80\xF9\xBC\x00\x80\xB6\xA5"
-     "\x0C\x00"
-     "\xB5\xB5\x0C\x00\xFD\xBC\x00\x80\xBB\xF9\x0C\x00\xBA\xF1\x0C\x00\x86\x28"
-     "\x05\x00"
-     "\x87\x20\x05\x00\xBF\x19\x03\x00\xBE\x11\x03\x00\xBD\xDD\x0C\x00\xBC\xDD"
-     "\x0C\x00"
-     "\x01\xBD\x00\x80\xA3\xD9\x0C\x00\x05\xBD\x00\x80\x09\xBD\x00\x80\xA6\xE1"
-     "\x0C\x00"
-     "\x0D\xBD\x00\x80\x11\xBD\x00\x80\xA5\xF1\x0C\x00\xAA\xB5\x0C\x00\xAB\xBD"
-     "\x0C\x00"
-     "\x15\xBD\x00\x80\x19\xBD\x00\x80\xAE\x55\x03\x00\xAF\x5D\x03\x00\xAC\x99"
-     "\x0C\x00"
-     "\xAD\x99\x0C\x00\x1D\xBD\x00\x80\x21\xBD\x00\x80\x25\xBD\x00\x80\x29\xBD"
-     "\x00\x80"
-     "\x2D\xBD\x00\x80\x31\xBD\x00\x80\x35\xBD\x00\x80\xEF\xBC\x01\x00\x39\xBD"
-     "\x00\x80"
-     "\xE1\x7C\x0E\x00\x3D\xBD\x00\x80\xE3\x50\x01\x00\x41\xBD\x00\x80\x45\xBD"
-     "\x00\x80"
-     "\x49\xBD\x00\x80\x4D\xBD\x00\x80\xB3\x95\x02\x00\x51\xBD\x00\x80\x55\xBD"
-     "\x00\x80"
-     "\x59\xBD\x00\x80\x5D\xBD\x00\x80\xB6\xBD\x02\x00\xB5\xB9\x02\x00\x65\xBD"
-     "\x00\x80"
-     "\xBB\x39\x02\x00\xBA\x61\x02\x00\x86\xC8\x04\x00\x87\xAC\x04\x00\xBF\x19"
-     "\x02\x00"
-     "\xBE\x11\x02\x00\xBD\x19\x02\x00\xBC\x21\x02\x00\xA3\x55\x05\x00\x82\xD5"
-     "\x07\x00"
-     "\x81\xC5\x07\x00\x80\xC5\x07\x00\x69\xBD\x00\x80\xA6\x7D\x05\x00\xA5\x79"
-     "\x05\x00"
-     "\x6D\xBD\x00\x80\xAB\xF9\x05\x00\xAA\xA1\x05\x00\x71\xBD\x00\x80\x75\xBD"
-     "\x00\x80"
-     "\xAF\xD9\x05\x00\xAE\xD1\x05\x00\xAD\xD9\x05\x00\xAC\xE1\x05\x00\x79\xBD"
-     "\x00\x80"
-     "\xB3\x59\x07\x00\x7D\xBD\x00\x80\x81\xBD\x00\x80\xB6\x1D\x07\x00\x85\xBD"
-     "\x00\x80"
-     "\x89\xBD\x00\x80\xB5\x15\x07\x00\xBA\x09\x07\x00\xBB\x09\x07\x00\x8D\xBD"
-     "\x00\x80"
-     "\x91\xBD\x00\x80\xBE\xF9\x07\x00\xBF\xF9\x07\x00\xBC\xF9\x07\x00\xBD\xF9"
-     "\x07\x00"
-     "\xF1\xBC\x00\x80\x61\xBD\x00\x80\x95\xBD\x00\x80\x99\xBD\x00\x80\x9D\xBD"
-     "\x00\x80"
-     "\xA1\xBD\x00\x80\xA5\xBD\x00\x80\xA9\xBD\x00\x80\xA8\xAD\x07\x00\xA9\xB5"
-     "\x07\x00"
-     "\xAA\xB5\x07\x00\xAB\xED\x07\x00\xAC\xF9\x07\x00\xAD\xF9\x07\x00\xAE\xED"
-     "\x07\x00"
-     "\xAF\xE5\x07\x00\xB0\xA9\x07\x00\xB1\xA9\x07\x00\xB2\x49\x07\x00\xB3\x49"
-     "\x07\x00"
-     "\xB4\x59\x07\x00\xB5\x59\x07\x00\xB6\x49\x07\x00\xB7\x49\x07\x00\xB8\x79"
-     "\x07\x00"
-     "\xB9\x45\x07\x00\xBA\x41\x07\x00\xBB\x5D\x07\x00\xBC\x45\x07\x00\xBD\x4D"
-     "\x07\x00"
-     "\xBE\x45\x07\x00\xBF\x39\x07\x00\xA3\x1D\x06\x00\xAD\xBD\x00\x80\xB1\xBD"
-     "\x00\x80"
-     "\xB5\xBD\x00\x80\xB9\xBD\x00\x80\xA6\x59\x06\x00\xA5\x51\x06\x00\xBD\xBD"
-     "\x00\x80"
-     "\xAB\x4D\x06\x00\xAA\x4D\x06\x00\xC1\xBD\x00\x80\xC5\xBD\x00\x80\xAF\xBD"
-     "\x06\x00"
-     "\xAE\xBD\x06\x00\xAD\xBD\x06\x00\xAC\xBD\x06\x00\x80\x6D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\xC9\xBD\x00\x80\xCD\xBD\x00\x80\x84\x98\x01\x00\xBE\x90"
-     "\x01\x00"
-     "\xD1\xBD\x00\x80\x86\x00\x1C\x00\x87\xC4\x01\x00\xD5\xBD\x00\x80\xD9\xBD"
-     "\x00\x80"
-     "\xDD\xBD\x00\x80\xE1\xBD\x00\x80\xE5\xBD\x00\x80\xE9\xBD\x00\x80\xA8\x5D"
-     "\x06\x00"
-     "\xA9\x95\x01\x00\xAA\x95\x01\x00\xAB\xA5\x01\x00\xAC\xBD\x01\x00\xAD\xD5"
-     "\x01\x00"
-     "\xAE\xD1\x01\x00\xAF\xD1\x01\x00\xED\xBD\x00\x80\xF1\xBD\x00\x80\xF5\xBD"
-     "\x00\x80"
-     "\xF9\xBD\x00\x80\xFD\xBD\x00\x80\x01\xBE\x00\x80\x05\xBE\x00\x80\x09\xBE"
-     "\x00\x80"
-     "\xB8\x59\x01\x00\xB9\x59\x01\x00\xBA\xCD\x00\x00\xBB\xC5\x00\x00\xBC\xDD"
-     "\x00\x00"
-     "\xBD\xC5\x00\x00\xBE\xC5\x00\x00\xBF\xF5\x00\x00\xB0\xB5\x01\x00\xB1\xBD"
-     "\x01\x00"
-     "\xB2\x81\x01\x00\xB3\x81\x01\x00\xB4\x79\x01\x00\xB5\x79\x01\x00\xB6\x69"
-     "\x01\x00"
-     "\xB7\x69\x01\x00\xB3\x1D\x02\x00\x0D\xBE\x00\x80\x11\xBE\x00\x80\xBE\x80"
-     "\x1C\x00"
-     "\x15\xBE\x00\x80\xB6\x55\x02\x00\xB5\x35\x02\x00\x19\xBE\x00\x80\xBB\x79"
-     "\x02\x00"
-     "\xBA\x71\x02\x00\x1D\xBE\x00\x80\x21\xBE\x00\x80\xBF\xBD\x02\x00\xBE\xBD"
-     "\x02\x00"
-     "\xBD\x55\x02\x00\xBC\x55\x02\x00\x25\xBE\x00\x80\xA3\x59\x02\x00\x29\xBE"
-     "\x00\x80"
-     "\x2D\xBE\x00\x80\xA6\x11\x02\x00\x31\xBE\x00\x80\x35\xBE\x00\x80\xA5\x71"
-     "\x02\x00"
-     "\xAA\x35\x02\x00\xAB\x3D\x02\x00\x39\xBE\x00\x80\x3D\xBE\x00\x80\xAE\xF9"
-     "\x02\x00"
-     "\xAF\xF9\x02\x00\xAC\x11\x02\x00\xAD\x11\x02\x00\x45\xBE\x00\x80\x49\xBE"
-     "\x00\x80"
-     "\xBE\x04\x1D\x00\xBE\x00\x1E\x00\x4D\xBE\x00\x80\x51\xBE\x00\x80\x55\xBE"
-     "\x00\x80"
-     "\x59\xBE\x00\x80\x80\xF9\x00\x00\x81\xCD\x00\x00\x82\xC5\x00\x00\x84\x20"
-     "\x03\x00"
-     "\x86\xB0\x1C\x00\x87\x94\x03\x00\x5D\xBE\x00\x80\x61\xBE\x00\x80\x65\xBE"
-     "\x00\x80"
-     "\x69\xBE\x00\x80\x6D\xBE\x00\x80\x71\xBE\x00\x80\xE3\x6C\x02\x00\x75\xBE"
-     "\x00\x80"
-     "\xE1\xA0\x01\x00\x79\xBE\x00\x80\xEF\x14\x02\x00\x7D\xBE\x00\x80\x81\xBE"
-     "\x00\x80"
-     "\x85\xBE\x00\x80\x89\xBE\x00\x80\x8D\xBE\x00\x80\x91\xBE\x00\x80\x95\xBE"
-     "\x00\x80"
-     "\xE1\xF0\x06\x00\xE1\x34\x06\x00\xE3\x4C\x00\x00\xE3\x18\x06\x00\x99\xBE"
-     "\x00\x80"
-     "\x9D\xBE\x00\x80\xA1\xBE\x00\x80\xA5\xBE\x00\x80\x80\x3D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\xA9\xBE\x00\x80\xAD\xBE\x00\x80\x84\xBC\x1D\x00\xEF\x98"
-     "\x00\x00"
-     "\xEF\x38\x07\x00\xB3\x31\x02\x00\xD1\x00\x00\x00\x87\xD8\x1D\x00\x86\x6C"
-     "\x1C\x00"
-     "\xB5\xBE\x00\x80\xB6\x29\x02\x00\xB5\x21\x02\x00\xB9\xBE\x00\x80\xBB\xCD"
-     "\x02\x00"
-     "\xBA\xCD\x02\x00\xBD\xBE\x00\x80\xC1\xBE\x00\x80\xBF\xCD\x02\x00\xBE\xCD"
-     "\x02\x00"
-     "\xBD\xCD\x02\x00\xBC\xCD\x02\x00\xB2\x5D\x06\x00\xB3\x65\x06\x00\xB0\x0D"
-     "\x06\x00"
-     "\xB1\x55\x06\x00\xB6\x7D\x06\x00\xB7\x9D\x05\x00\xB4\x7D\x06\x00\xB5\x75"
-     "\x06\x00"
-     "\xBA\x8D\x05\x00\xBB\xCD\x05\x00\xB8\xA5\x05\x00\xB9\x85\x05\x00\xBE\xC5"
-     "\x05\x00"
-     "\xBF\xC9\x05\x00\xBC\xD5\x05\x00\xBD\xCD\x05\x00\xC5\xBE\x00\x80\xC9\xBE"
-     "\x00\x80"
-     "\xCD\xBE\x00\x80\xD1\xBE\x00\x80\xD5\xBE\x00\x80\xD9\xBE\x00\x80\xDD\xBE"
-     "\x00\x80"
-     "\xE1\xBE\x00\x80\xAA\xB5\x06\x00\xAB\xBD\x06\x00\xA8\x01\x07\x00\xA9\xBD"
-     "\x06\x00"
-     "\xAE\x81\x06\x00\xAF\x8D\x06\x00\xAC\x99\x06\x00\xAD\x95\x06\x00\xA3\x71"
-     "\x1D\x00"
-     "\xE5\xBE\x00\x80\xE9\xBE\x00\x80\xED\xBE\x00\x80\xF1\xBE\x00\x80\xA6\x69"
-     "\x1D\x00"
-     "\xA5\x61\x1D\x00\xF5\xBE\x00\x80\xAB\x8D\x1D\x00\xAA\x8D\x1D\x00\xF9\xBE"
-     "\x00\x80"
-     "\xFD\xBE\x00\x80\xAF\x8D\x1D\x00\xAE\x8D\x1D\x00\xAD\x8D\x1D\x00\xAC\x8D"
-     "\x1D\x00"
-     "\x01\xBF\x00\x80\xB3\xD5\x1E\x00\x05\xBF\x00\x80\x09\xBF\x00\x80\xB6\xCD"
-     "\x1E\x00"
-     "\x0D\xBF\x00\x80\x11\xBF\x00\x80\xB5\xC5\x1E\x00\xBA\xA1\x1E\x00\xBB\xA1"
-     "\x1E\x00"
-     "\x15\xBF\x00\x80\x19\xBF\x00\x80\xBE\xA5\x1E\x00\xBF\xA9\x1E\x00\xBC\xB1"
-     "\x1E\x00"
-     "\xBD\xB1\x1E\x00\x41\xBE\x00\x80\xB1\xBE\x00\x80\x1D\xBF\x00\x80\x84\x04"
-     "\x03\x00"
-     "\x80\xF9\x00\x00\x81\xF9\x00\x00\x82\x11\x00\x00\x21\xBF\x00\x80\xA8\x21"
-     "\x1E\x00"
-     "\xA9\x21\x1E\x00\xAA\x39\x1E\x00\xAB\x39\x1E\x00\xAC\x29\x1E\x00\xAD\x29"
-     "\x1E\x00"
-     "\xAE\x01\x1E\x00\xAF\x01\x1E\x00\xB0\x01\x1E\x00\xB1\x01\x1E\x00\xB2\x01"
-     "\x1E\x00"
-     "\xB3\x01\x1E\x00\xB4\x05\x1E\x00\xB5\x09\x1E\x00\xB6\x3D\x1E\x00\xB7\x35"
-     "\x1E\x00"
-     "\xB8\x0D\x1E\x00\xB9\x15\x1E\x00\xBA\x1D\x1E\x00\xBB\x15\x1E\x00\xBC\x0D"
-     "\x1E\x00"
-     "\xBD\xE5\x1F\x00\xBE\xED\x1F\x00\xBF\xE5\x1F\x00\xA3\x91\x1F\x00\x25\xBF"
-     "\x00\x80"
-     "\x86\x28\x01\x00\x87\x48\x01\x00\x29\xBF\x00\x80\xA6\x89\x1F\x00\xA5\x81"
-     "\x1F\x00"
-     "\x2D\xBF\x00\x80\xAB\xE5\x1F\x00\xAA\xE5\x1F\x00\x31\xBF\x00\x80\x35\xBF"
-     "\x00\x80"
-     "\xAF\xED\x1F\x00\xAE\xE1\x1F\x00\xAD\xF5\x1F\x00\xAC\xF5\x1F\x00\x39\xBF"
-     "\x00\x80"
-     "\xB3\x2D\x1E\x00\x3D\xBF\x00\x80\x41\xBF\x00\x80\xB6\x95\x1E\x00\x45\xBF"
-     "\x00\x80"
-     "\x49\xBF\x00\x80\xB5\x9D\x1E\x00\xBA\xB1\x1E\x00\xBB\xB1\x1E\x00\x4D\xBF"
-     "\x00\x80"
-     "\x51\xBF\x00\x80\xBE\x75\x01\x00\xBF\x7D\x01\x00\xBC\xA1\x1E\x00\xBD\xA1"
-     "\x1E\x00"
-     "\xA8\xD1\x1E\x00\xA9\xD9\x1E\x00\xAA\xE1\x1E\x00\xAB\xE1\x1E\x00\xAC\x51"
-     "\x1E\x00"
-     "\xAD\x51\x1E\x00\xAE\x51\x1E\x00\xAF\x51\x1E\x00\x55\xBF\x00\x80\x59\xBF"
-     "\x00\x80"
-     "\x5D\xBF\x00\x80\x61\xBF\x00\x80\x65\xBF\x00\x80\x69\xBF\x00\x80\x6D\xBF"
-     "\x00\x80"
-     "\x71\xBF\x00\x80\xB8\xDD\x01\x00\xB9\xED\x01\x00\xBA\xE5\x01\x00\xBB\x8D"
-     "\x01\x00"
-     "\xBC\x99\x01\x00\xBD\x99\x01\x00\xBE\x8D\x01\x00\xBF\x85\x01\x00\xB0\x31"
-     "\x1E\x00"
-     "\xB1\x31\x1E\x00\xB2\x31\x1E\x00\xB3\x31\x1E\x00\xB4\xF5\x01\x00\xB5\xFD"
-     "\x01\x00"
-     "\xB6\xF5\x01\x00\xB7\xED\x01\x00\xA3\x69\x1D\x00\x75\xBF\x00\x80\x79\xBF"
-     "\x00\x80"
-     "\x7D\xBF\x00\x80\x81\xBF\x00\x80\xA6\xD1\x1D\x00\xA5\xD9\x1D\x00\x85\xBF"
-     "\x00\x80"
-     "\xAB\xF5\x1D\x00\xAA\xF5\x1D\x00\x89\xBF\x00\x80\x8D\xBF\x00\x80\xAF\x39"
-     "\x02\x00"
-     "\xAE\x31\x02\x00\xAD\xE5\x1D\x00\xAC\xE5\x1D\x00\x81\x69\x00\x00\x80\x59"
-     "\x00\x00"
-     "\xBE\x00\x04\x00\x82\x61\x00\x00\x95\xBF\x00\x80\x99\xBF\x00\x80\x9D\xBF"
-     "\x00\x80"
-     "\xA1\xBF\x00\x80\x84\xEC\x03\x00\xA5\xBF\x00\x80\x87\x88\x03\x00\x86\xEC"
-     "\x04\x00"
-     "\xA9\xBF\x00\x80\xAD\xBF\x00\x80\xB1\xBF\x00\x80\xB5\xBF\x00\x80\xA8\x85"
-     "\x03\x00"
-     "\xA9\x95\x03\x00\xAA\x95\x03\x00\xAB\xA5\x03\x00\xAC\xBD\x03\x00\xAD\xD5"
-     "\x03\x00"
-     "\xAE\xD1\x03\x00\xAF\xD1\x03\x00\xB9\xBF\x00\x80\xBD\xBF\x00\x80\xC1\xBF"
-     "\x00\x80"
-     "\xC5\xBF\x00\x80\xC9\xBF\x00\x80\xCD\xBF\x00\x80\xD1\xBF\x00\x80\xD5\xBF"
-     "\x00\x80"
-     "\xB8\x71\x03\x00\xB9\x71\x03\x00\xBA\x71\x03\x00\xBB\x71\x03\x00\xBC\xD5"
-     "\x00\x00"
-     "\xBD\xDD\x00\x00\xBE\xD5\x00\x00\xBF\xCD\x00\x00\xB0\xB5\x03\x00\xB1\xBD"
-     "\x03\x00"
-     "\xB2\x81\x03\x00\xB3\x81\x03\x00\xB4\x51\x03\x00\xB5\x51\x03\x00\xB6\x51"
-     "\x03\x00"
-     "\xB7\x51\x03\x00\xE1\x54\x1E\x00\xE1\xAC\x1F\x00\xE3\x94\x01\x00\xE3\x28"
-     "\x1E\x00"
-     "\xE3\x60\x03\x00\xD9\xBF\x00\x80\xE1\x08\x00\x00\xDD\xBF\x00\x80\xEF\x90"
-     "\x03\x00"
-     "\xE1\xBF\x00\x80\xE5\xBF\x00\x80\xE9\xBF\x00\x80\xED\xBF\x00\x80\xF1\xBF"
-     "\x00\x80"
-     "\xEF\x4C\x1F\x00\xEF\x4C\x1F\x00\xB3\x5D\x02\x00\xF5\xBF\x00\x80\xF9\xBF"
-     "\x00\x80"
-     "\xFD\xBF\x00\x80\x05\xC0\x00\x80\xB6\x15\x02\x00\xB5\x75\x02\x00\x09\xC0"
-     "\x00\x80"
-     "\xBB\x39\x02\x00\xBA\x31\x02\x00\x84\x24\x05\x00\xBE\xE0\x04\x00\xBF\xD5"
-     "\x02\x00"
-     "\xBE\xD5\x02\x00\xBD\x15\x02\x00\xBC\x15\x02\x00\xB8\x91\x1D\x00\xB9\x99"
-     "\x1D\x00"
-     "\xBA\xA1\x1D\x00\xBB\xA1\x1D\x00\xBC\xD1\x1D\x00\xBD\xDD\x1D\x00\xBE\xD5"
-     "\x1D\x00"
-     "\xBF\xC9\x1D\x00\xB0\x09\x1E\x00\xB1\x09\x1E\x00\xB2\x19\x1E\x00\xB3\x19"
-     "\x1E\x00"
-     "\xB4\x09\x1E\x00\xB5\x09\x1E\x00\xB6\xBD\x1D\x00\xB7\xB5\x1D\x00\xA8\xA9"
-     "\x1E\x00"
-     "\xA9\xA9\x1E\x00\xAA\xB9\x1E\x00\xAB\xB9\x1E\x00\xAC\xA9\x1E\x00\xAD\xA9"
-     "\x1E\x00"
-     "\xAE\x79\x1E\x00\xAF\x79\x1E\x00\x80\xA5\x00\x00\x81\xAD\x00\x00\x82\xA5"
-     "\x00\x00"
-     "\x0D\xC0\x00\x80\x86\xD0\x04\x00\x87\xF8\x04\x00\x11\xC0\x00\x80\x15\xC0"
-     "\x00\x80"
-     "\x91\xBF\x00\x80\x01\xC0\x00\x80\x19\xC0\x00\x80\x1D\xC0\x00\x80\x21\xC0"
-     "\x00\x80"
-     "\x25\xC0\x00\x80\x29\xC0\x00\x80\x2D\xC0\x00\x80\xA8\x71\x06\x00\xA9\x71"
-     "\x06\x00"
-     "\xAA\x71\x06\x00\xAB\x71\x06\x00\xAC\x55\x06\x00\xAD\x45\x06\x00\xAE\x4D"
-     "\x06\x00"
-     "\xAF\x45\x06\x00\xB0\x3D\x06\x00\xB1\xE5\x06\x00\xB2\xED\x06\x00\xB3\xE5"
-     "\x06\x00"
-     "\xB4\xFD\x06\x00\xB5\xE5\x06\x00\xB6\xED\x06\x00\xB7\xE5\x06\x00\xB8\xDD"
-     "\x06\x00"
-     "\xB9\x71\x07\x00\xBA\x75\x07\x00\xBB\x49\x07\x00\xBC\x59\x07\x00\xBD\x59"
-     "\x07\x00"
-     "\xBE\x49\x07\x00\xBF\x49\x07\x00\xB3\xD9\x06\x00\x31\xC0\x00\x80\x35\xC0"
-     "\x00\x80"
-     "\x39\xC0\x00\x80\x3D\xC0\x00\x80\xB6\xE5\x06\x00\xB5\xFD\x06\x00\x41\xC0"
-     "\x00\x80"
-     "\xBB\x01\x06\x00\xBA\xD9\x06\x00\x45\xC0\x00\x80\x49\xC0\x00\x80\xBF\x01"
-     "\x06\x00"
-     "\xBE\x19\x06\x00\xBD\x11\x06\x00\xBC\x19\x06\x00\x4D\xC0\x00\x80\xA3\xD9"
-     "\x05\x00"
-     "\x51\xC0\x00\x80\x55\xC0\x00\x80\xA6\x91\x05\x00\x59\xC0\x00\x80\x5D\xC0"
-     "\x00\x80"
-     "\xA5\xF1\x05\x00\xAA\xB5\x05\x00\xAB\xBD\x05\x00\x61\xC0\x00\x80\x65\xC0"
-     "\x00\x80"
-     "\xAE\x51\x05\x00\xAF\x51\x05\x00\xAC\x91\x05\x00\xAD\x91\x05\x00\xA3\x59"
-     "\x07\x00"
-     "\x82\x19\x00\x00\x81\x19\x00\x00\x80\xE1\x01\x00\x69\xC0\x00\x80\xA6\x65"
-     "\x07\x00"
-     "\xA5\x7D\x07\x00\x6D\xC0\x00\x80\xAB\x81\x07\x00\xAA\x59\x07\x00\x84\xA0"
-     "\x02\x00"
-     "\xBE\xAC\x01\x00\xAF\x81\x07\x00\xAE\x99\x07\x00\xAD\x91\x07\x00\xAC\x99"
-     "\x07\x00"
-     "\x71\xC0\x00\x80\xB3\xA9\x06\x00\x86\xE8\x00\x00\x87\x2C\x01\x00\xB6\x59"
-     "\x01\x00"
-     "\x75\xC0\x00\x80\x79\xC0\x00\x80\xB5\xA1\x06\x00\xBA\x75\x01\x00\xBB\x7D"
-     "\x01\x00"
-     "\x7D\xC0\x00\x80\x81\xC0\x00\x80\xBE\xF9\x01\x00\xBF\xF9\x01\x00\xBC\x65"
-     "\x01\x00"
-     "\xBD\xF9\x01\x00\xA8\xD1\x06\x00\xA9\xD9\x06\x00\xAA\x65\x06\x00\xAB\x75"
-     "\x06\x00"
-     "\xAC\x6D\x06\x00\xAD\x9D\x01\x00\xAE\x95\x01\x00\xAF\x8D\x01\x00\x84\xEC"
-     "\x01\x00"
-     "\x85\xC0\x00\x80\x89\xC0\x00\x80\x8D\xC0\x00\x80\x91\xC0\x00\x80\x95\xC0"
-     "\x00\x80"
-     "\x99\xC0\x00\x80\x9D\xC0\x00\x80\xB8\x69\x01\x00\xB9\x69\x01\x00\xBA\x09"
-     "\x01\x00"
-     "\xBB\x05\x01\x00\xBC\x1D\x01\x00\xBD\x01\x01\x00\xBE\x01\x01\x00\xBF\x35"
-     "\x01\x00"
-     "\xB0\xF5\x01\x00\xB1\xFD\x01\x00\xB2\xF5\x01\x00\xB3\x69\x01\x00\xB4\x79"
-     "\x01\x00"
-     "\xB5\x79\x01\x00\xB6\x69\x01\x00\xB7\x61\x01\x00\xA1\xC0\x00\x80\xA5\xC0"
-     "\x00\x80"
-     "\xA9\xC0\x00\x80\xA3\xE1\x05\x00\xAD\xC0\x00\x80\xA5\xE9\x05\x00\xA6\x11"
-     "\x02\x00"
-     "\xB1\xC0\x00\x80\xB5\xC0\x00\x80\xB9\xC0\x00\x80\xAA\x3D\x02\x00\xAB\x35"
-     "\x02\x00"
-     "\xAC\x2D\x02\x00\xAD\xB1\x02\x00\xAE\xB1\x02\x00\xAF\xB1\x02\x00\xBD\xC0"
-     "\x00\x80"
-     "\xC1\xC0\x00\x80\xBE\x04\x03\x00\x84\x00\x0C\x00\xC5\xC0\x00\x80\xC9\xC0"
-     "\x00\x80"
-     "\xCD\xC0\x00\x80\xD1\xC0\x00\x80\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\xD5\xC0\x00\x80\xD9\xC0\x00\x80\xDD\xC0\x00\x80\x87\x44\x03\x00\x86\xDC"
-     "\x0C\x00"
-     "\xB3\xE9\x03\x00\xE5\xC0\x00\x80\xE9\xC0\x00\x80\xED\xC0\x00\x80\xF1\xC0"
-     "\x00\x80"
-     "\xB6\x3D\x03\x00\xB5\x3D\x03\x00\xF5\xC0\x00\x80\xBB\x19\x03\x00\xBA\x11"
-     "\x03\x00"
-     "\xF9\xC0\x00\x80\xFD\xC0\x00\x80\xBF\xB9\x00\x00\xBE\xB1\x00\x00\xBD\xB9"
-     "\x00\x00"
-     "\xBC\x01\x03\x00\x01\xC1\x00\x80\xE1\x94\x01\x00\x05\xC1\x00\x80\xE3\x7C"
-     "\x01\x00"
-     "\x09\xC1\x00\x80\x0D\xC1\x00\x80\x11\xC1\x00\x80\x15\xC1\x00\x80\x19\xC1"
-     "\x00\x80"
-     "\x1D\xC1\x00\x80\x21\xC1\x00\x80\x25\xC1\x00\x80\x29\xC1\x00\x80\x2D\xC1"
-     "\x00\x80"
-     "\x31\xC1\x00\x80\xEF\x4C\x02\x00\xA8\x55\x02\x00\xA9\x5D\x02\x00\xAA\x61"
-     "\x02\x00"
-     "\xAB\x61\x02\x00\xAC\xB5\x02\x00\xAD\xBD\x02\x00\xAE\xB5\x02\x00\xAF\xAD"
-     "\x02\x00"
-     "\xBE\x68\x0D\x00\x35\xC1\x00\x80\x39\xC1\x00\x80\x3D\xC1\x00\x80\x41\xC1"
-     "\x00\x80"
-     "\x82\xAD\x00\x00\x81\xAD\x00\x00\x80\xAD\x00\x00\xB8\x61\x01\x00\xB9\x61"
-     "\x01\x00"
-     "\xBA\x09\x01\x00\xBB\x09\x01\x00\xBC\x01\x01\x00\xBD\x01\x01\x00\xBE\x01"
-     "\x01\x00"
-     "\xBF\x01\x01\x00\xB0\xD5\x02\x00\xB1\xDD\x02\x00\xB2\xD5\x02\x00\xB3\x6D"
-     "\x01\x00"
-     "\xB4\x75\x01\x00\xB5\x7D\x01\x00\xB6\x69\x01\x00\xB7\x61\x01\x00\xE1\x68"
-     "\x06\x00"
-     "\xE1\xF0\x07\x00\xE3\xB0\x00\x00\xE3\xF4\x06\x00\x45\xC1\x00\x80\x49\xC1"
-     "\x00\x80"
-     "\x4D\xC1\x00\x80\x55\xC1\x00\x80\x59\xC1\x00\x80\x5D\xC1\x00\x80\x61\xC1"
-     "\x00\x80"
-     "\x65\xC1\x00\x80\xBE\xFC\x0C\x00\x69\xC1\x00\x80\xEF\x6C\x00\x00\xEF\x28"
-     "\x06\x00"
-     "\xA3\xA9\x02\x00\x6D\xC1\x00\x80\x86\x68\x0C\x00\x87\x04\x0D\x00\x71\xC1"
-     "\x00\x80"
-     "\xA6\x7D\x02\x00\xA5\x7D\x02\x00\x75\xC1\x00\x80\xAB\x59\x02\x00\xAA\x51"
-     "\x02\x00"
-     "\x79\xC1\x00\x80\x7D\xC1\x00\x80\xAF\xF9\x01\x00\xAE\xF1\x01\x00\xAD\xF9"
-     "\x01\x00"
-     "\xAC\x41\x02\x00\xA8\x85\x0E\x00\xA9\x8D\x0E\x00\xAA\x85\x0E\x00\xAB\x9D"
-     "\x0E\x00"
-     "\xAC\x8D\x0E\x00\xAD\xBD\x0E\x00\xAE\xB5\x0E\x00\xAF\xDD\x0E\x00\x51\xC1"
-     "\x00\x80"
-     "\x81\xC1\x00\x80\x85\xC1\x00\x80\x89\xC1\x00\x80\x8D\xC1\x00\x80\x91\xC1"
-     "\x00\x80"
-     "\x95\xC1\x00\x80\x99\xC1\x00\x80\xB8\xAD\x0E\x00\xB9\xB5\x0E\x00\xBA\xBD"
-     "\x0E\x00"
-     "\xBB\xB5\x0E\x00\xBC\x75\x0F\x00\xBD\x7D\x0F\x00\xBE\x75\x0F\x00\xBF\x6D"
-     "\x0F\x00"
-     "\xB0\xA9\x0E\x00\xB1\xB5\x0E\x00\xB2\xBD\x0E\x00\xB3\xB5\x0E\x00\xB4\xAD"
-     "\x0E\x00"
-     "\xB5\x95\x0E\x00\xB6\x9D\x0E\x00\xB7\x95\x0E\x00\xB3\x0D\x0E\x00\x9D\xC1"
-     "\x00\x80"
-     "\xA1\xC1\x00\x80\xA5\xC1\x00\x80\xA9\xC1\x00\x80\xB6\x3D\x0E\x00\xB5\x05"
-     "\x0E\x00"
-     "\xAD\xC1\x00\x80\xBB\x71\x0E\x00\xBA\x6D\x0E\x00\xB1\xC1\x00\x80\xB5\xC1"
-     "\x00\x80"
-     "\xBF\x51\x0E\x00\xBE\x69\x0E\x00\xBD\x61\x0E\x00\xBC\x69\x0E\x00\x82\x19"
-     "\x00\x00"
-     "\xA3\x49\x0E\x00\x80\x65\x00\x00\x81\x19\x00\x00\xA6\x79\x0E\x00\xB9\xC1"
-     "\x00\x80"
-     "\xBD\xC1\x00\x80\xA5\x41\x0E\x00\xAA\x29\x0E\x00\xAB\x35\x0E\x00\x84\xBC"
-     "\x03\x00"
-     "\xC1\xC1\x00\x80\xAE\x2D\x0E\x00\xAF\x15\x0E\x00\xAC\x2D\x0E\x00\xAD\x25"
-     "\x0E\x00"
-     "\xA8\x9D\x0E\x00\xA9\xA5\x0E\x00\xAA\xAD\x0E\x00\xAB\xA5\x0E\x00\xAC\xBD"
-     "\x0E\x00"
-     "\xAD\xA1\x0E\x00\xAE\xDD\x0E\x00\xAF\xCD\x0E\x00\x86\x20\x01\x00\x87\x64"
-     "\x01\x00"
-     "\xC5\xC1\x00\x80\xC9\xC1\x00\x80\xCD\xC1\x00\x80\xD1\xC1\x00\x80\xD5\xC1"
-     "\x00\x80"
-     "\xD9\xC1\x00\x80\xB8\x79\x01\x00\xB9\x79\x01\x00\xBA\xCD\x01\x00\xBB\xC5"
-     "\x01\x00"
-     "\xBC\xDD\x01\x00\xBD\xC5\x01\x00\xBE\xC5\x01\x00\xBF\xF5\x01\x00\xB0\xBD"
-     "\x0E\x00"
-     "\xB1\x8D\x0E\x00\xB2\x85\x0E\x00\xB3\x49\x01\x00\xB4\x59\x01\x00\xB5\x59"
-     "\x01\x00"
-     "\xB6\x49\x01\x00\xB7\x49\x01\x00\xB5\x2D\x0E\x00\xDD\xC1\x00\x80\xE1\xC1"
-     "\x00\x80"
-     "\xB6\x39\x0E\x00\xE5\xC1\x00\x80\xE9\xC1\x00\x80\xB3\x3D\x0E\x00\xED\xC1"
-     "\x00\x80"
-     "\xBD\x85\x01\x00\xBC\x49\x0E\x00\xBF\x85\x01\x00\xBE\x85\x01\x00\xF1\xC1"
-     "\x00\x80"
-     "\xE1\xC0\x00\x80\xBB\x51\x0E\x00\xBA\x61\x0E\x00\xA3\x65\x0E\x00\xF5\xC1"
-     "\x00\x80"
-     "\xF9\xC1\x00\x80\xFD\xC1\x00\x80\x01\xC2\x00\x80\xA6\x61\x0E\x00\xA5\x75"
-     "\x0E\x00"
-     "\x05\xC2\x00\x80\xAB\x09\x0E\x00\xAA\x39\x0E\x00\x09\xC2\x00\x80\x0D\xC2"
-     "\x00\x80"
-     "\xAF\xDD\x01\x00\xAE\xDD\x01\x00\xAD\xDD\x01\x00\xAC\x11\x0E\x00\x11\xC2"
-     "\x00\x80"
-     "\x15\xC2\x00\x80\xEF\xD0\x0F\x00\x19\xC2\x00\x80\x1D\xC2\x00\x80\x21\xC2"
-     "\x00\x80"
-     "\x25\xC2\x00\x80\x29\xC2\x00\x80\x2D\xC2\x00\x80\xBE\x68\x03\x00\x35\xC2"
-     "\x00\x80"
-     "\x39\xC2\x00\x80\xE1\x54\x0E\x00\x3D\xC2\x00\x80\xE3\x64\x0E\x00\x41\xC2"
-     "\x00\x80"
-     "\x80\x59\x00\x00\x81\x59\x00\x00\x82\x69\x00\x00\x84\x8C\x03\x00\x86\xF0"
-     "\x04\x00"
-     "\x87\x14\x03\x00\x45\xC2\x00\x80\x49\xC2\x00\x80\x4D\xC2\x00\x80\x51\xC2"
-     "\x00\x80"
-     "\x55\xC2\x00\x80\x59\xC2\x00\x80\x5D\xC2\x00\x80\x61\xC2\x00\x80\x65\xC2"
-     "\x00\x80"
-     "\x69\xC2\x00\x80\x6D\xC2\x00\x80\x71\xC2\x00\x80\x75\xC2\x00\x80\x79\xC2"
-     "\x00\x80"
-     "\x7D\xC2\x00\x80\x81\xC2\x00\x80\x85\xC2\x00\x80\x89\xC2\x00\x80\xA8\x89"
-     "\x03\x00"
-     "\xA9\x89\x03\x00\xAA\x99\x03\x00\xAB\x99\x03\x00\xAC\x89\x03\x00\xAD\x89"
-     "\x03\x00"
-     "\xAE\x3D\x03\x00\xAF\x35\x03\x00\xB0\x51\x03\x00\xB1\x51\x03\x00\xB2\x55"
-     "\x03\x00"
-     "\xB3\x7D\x03\x00\xB4\x15\x03\x00\xB5\x1D\x03\x00\xB6\x15\x03\x00\xB7\x0D"
-     "\x03\x00"
-     "\xB8\x3D\x03\x00\xB9\x0D\x03\x00\xBA\x05\x03\x00\xBB\xED\x00\x00\xBC\xF5"
-     "\x00\x00"
-     "\xBD\xF9\x00\x00\xBE\xE9\x00\x00\xBF\xE9\x00\x00\x8D\xC2\x00\x80\x91\xC2"
-     "\x00\x80"
-     "\x95\xC2\x00\x80\xB3\x3D\x02\x00\x99\xC2\x00\x80\xB5\x2D\x02\x00\xB6\x25"
-     "\x02\x00"
-     "\x9D\xC2\x00\x80\xBE\x68\x05\x00\xA5\xC2\x00\x80\xBA\xB9\x02\x00\xBB\xB9"
-     "\x02\x00"
-     "\xBC\xAD\x02\x00\xBD\x85\x02\x00\xBE\xFD\x02\x00\xBF\xF5\x02\x00\x80\x49"
-     "\x00\x00"
-     "\x81\x55\x00\x00\x82\x55\x00\x00\x84\x00\x05\x00\xEF\x8C\x03\x00\xBE\x18"
-     "\x04\x00"
-     "\x87\x74\x05\x00\x86\xFC\x04\x00\xE3\x3C\x03\x00\xA9\xC2\x00\x80\xE1\x50"
-     "\x00\x00"
-     "\xAD\xC2\x00\x80\xB1\xC2\x00\x80\xB5\xC2\x00\x80\xB9\xC2\x00\x80\xBD\xC2"
-     "\x00\x80"
-     "\xC1\xC2\x00\x80\xC5\xC2\x00\x80\xC9\xC2\x00\x80\xCD\xC2\x00\x80\xD1\xC2"
-     "\x00\x80"
-     "\xD5\xC2\x00\x80\xD9\xC2\x00\x80\xEF\xD4\x0E\x00\xBE\x28\x06\x00\xE1\x74"
-     "\x0E\x00"
-     "\xDD\xC2\x00\x80\xE3\x54\x01\x00\xE1\xC2\x00\x80\xE5\xC2\x00\x80\xE9\xC2"
-     "\x00\x80"
-     "\xED\xC2\x00\x80\xA3\xF1\x02\x00\xF1\xC2\x00\x80\xF5\xC2\x00\x80\xF9\xC2"
-     "\x00\x80"
-     "\xFD\xC2\x00\x80\xA6\xE9\x02\x00\xA5\xE1\x02\x00\x01\xC3\x00\x80\xAB\x75"
-     "\x02\x00"
-     "\xAA\x75\x02\x00\x05\xC3\x00\x80\x09\xC3\x00\x80\xAF\x39\x02\x00\xAE\x31"
-     "\x02\x00"
-     "\xAD\x49\x02\x00\xAC\x61\x02\x00\xA8\x29\x06\x00\xA9\x29\x06\x00\xAA\x3D"
-     "\x06\x00"
-     "\xAB\x31\x06\x00\xAC\x49\x06\x00\xAD\x49\x06\x00\xAE\x79\x06\x00\xAF\x79"
-     "\x06\x00"
-     "\xA1\xC2\x00\x80\x82\x15\x00\x00\x81\xC5\x07\x00\x80\xC5\x07\x00\x0D\xC3"
-     "\x00\x80"
-     "\x84\x68\x03\x00\x11\xC3\x00\x80\x15\xC3\x00\x80\xB8\xC9\x06\x00\xB9\xC9"
-     "\x06\x00"
-     "\xBA\xD9\x06\x00\xBB\xD9\x06\x00\xBC\xC9\x06\x00\xBD\xC9\x06\x00\xBE\x59"
-     "\x07\x00"
-     "\xBF\x59\x07\x00\xB0\x09\x06\x00\xB1\x09\x06\x00\xB2\x19\x06\x00\xB3\x19"
-     "\x06\x00"
-     "\xB4\x09\x06\x00\xB5\x09\x06\x00\xB6\xF9\x06\x00\xB7\xF9\x06\x00\xB3\xB5"
-     "\x06\x00"
-     "\x19\xC3\x00\x80\x86\xAC\x00\x00\x87\x40\x03\x00\x1D\xC3\x00\x80\xB6\xC9"
-     "\x06\x00"
-     "\xB5\xC1\x06\x00\x21\xC3\x00\x80\xBB\xCD\x06\x00\xBA\xCD\x06\x00\x25\xC3"
-     "\x00\x80"
-     "\x29\xC3\x00\x80\xBF\xCD\x06\x00\xBE\xCD\x06\x00\xBD\xCD\x06\x00\xBC\xCD"
-     "\x06\x00"
-     "\x2D\xC3\x00\x80\xA3\xF1\x06\x00\x31\xC3\x00\x80\x35\xC3\x00\x80\xA6\x8D"
-     "\x06\x00"
-     "\x39\xC3\x00\x80\x3D\xC3\x00\x80\xA5\x85\x06\x00\xAA\x89\x06\x00\xAB\x89"
-     "\x06\x00"
-     "\x41\xC3\x00\x80\x45\xC3\x00\x80\xAE\x89\x06\x00\xAF\x89\x06\x00\xAC\x89"
-     "\x06\x00"
-     "\xAD\x89\x06\x00\xA8\x25\x06\x00\xA9\x61\x06\x00\xAA\x65\x06\x00\xAB\x7D"
-     "\x06\x00"
-     "\xAC\x65\x06\x00\xAD\x6D\x06\x00\xAE\x65\x06\x00\xAF\x9D\x06\x00\x49\xC3"
-     "\x00\x80"
-     "\x4D\xC3\x00\x80\x51\xC3\x00\x80\x55\xC3\x00\x80\x59\xC3\x00\x80\x5D\xC3"
-     "\x00\x80"
-     "\x61\xC3\x00\x80\x65\xC3\x00\x80\xB8\xF5\x06\x00\xB9\xFD\x06\x00\xBA\xF5"
-     "\x06\x00"
-     "\xBB\x89\x06\x00\xBC\x99\x06\x00\xBD\x99\x06\x00\xBE\x89\x06\x00\xBF\x81"
-     "\x06\x00"
-     "\xB0\xE5\x06\x00\xB1\xED\x06\x00\xB2\xE5\x06\x00\xB3\xFD\x06\x00\xB4\xE5"
-     "\x06\x00"
-     "\xB5\xED\x06\x00\xB6\xD1\x06\x00\xB7\xCD\x06\x00\x69\xC3\x00\x80\xB6\xFD"
-     "\x06\x00"
-     "\xB5\xFD\x06\x00\x6D\xC3\x00\x80\xB3\xFD\x06\x00\x71\xC3\x00\x80\x75\xC3"
-     "\x00\x80"
-     "\x79\xC3\x00\x80\xBF\x39\x06\x00\xBE\x31\x06\x00\xBD\x39\x06\x00\xBC\x21"
-     "\x06\x00"
-     "\xBB\x39\x06\x00\xBA\x31\x06\x00\x31\xC2\x00\x80\x7D\xC3\x00\x80\xA3\xAD"
-     "\x06\x00"
-     "\x82\x79\x00\x00\x81\x55\x00\x00\x80\x55\x00\x00\x84\x5C\x01\x00\xA6\xAD"
-     "\x06\x00"
-     "\xA5\xAD\x06\x00\x85\xC3\x00\x80\xAB\x69\x06\x00\xAA\x61\x06\x00\x86\x48"
-     "\x7F\x00"
-     "\x87\xE4\x00\x00\xAF\x69\x06\x00\xAE\x61\x06\x00\xAD\x69\x06\x00\xAC\x71"
-     "\x06\x00"
-     "\x89\xC3\x00\x80\xEF\xDC\x07\x00\x8D\xC3\x00\x80\x91\xC3\x00\x80\x95\xC3"
-     "\x00\x80"
-     "\x99\xC3\x00\x80\x9D\xC3\x00\x80\xA1\xC3\x00\x80\xA5\xC3\x00\x80\x84\xA0"
-     "\x03\x00"
-     "\xA9\xC3\x00\x80\xBF\x24\x79\x00\xAD\xC3\x00\x80\xE3\x68\x07\x00\xB1\xC3"
-     "\x00\x80"
-     "\xE1\x74\x06\x00\xB3\xD1\x02\x00\xB5\xC3\x00\x80\xBE\x04\x03\x00\x84\x80"
-     "\x7D\x00"
-     "\xB9\xC3\x00\x80\xB6\xF9\x02\x00\xB5\xF1\x02\x00\xBD\xC3\x00\x80\xBB\xB5"
-     "\x02\x00"
-     "\xBA\xA9\x02\x00\xC1\xC3\x00\x80\xC5\xC3\x00\x80\xBF\x51\x03\x00\xBE\x99"
-     "\x02\x00"
-     "\xBD\x91\x02\x00\xBC\xA5\x02\x00\xA9\x05\x02\x00\xA8\xB9\x02\x00\xAB\x15"
-     "\x02\x00"
-     "\xAA\x1D\x02\x00\xAD\x3D\x02\x00\xAC\x3D\x02\x00\xAF\x51\x02\x00\xAE\x5D"
-     "\x02\x00"
-     "\xBE\x68\x7D\x00\xC9\xC3\x00\x80\xCD\xC3\x00\x80\xD1\xC3\x00\x80\xD5\xC3"
-     "\x00\x80"
-     "\xD9\xC3\x00\x80\xDD\xC3\x00\x80\xE1\xC3\x00\x80\xB9\xF1\x03\x00\xB8\xE9"
-     "\x03\x00"
-     "\xBB\xF1\x03\x00\xBA\xF9\x03\x00\xBD\x51\x03\x00\xBC\xE9\x03\x00\xBF\x4D"
-     "\x03\x00"
-     "\xBE\x51\x03\x00\xB1\x35\x02\x00\xB0\x29\x02\x00\xB3\x01\x02\x00\xB2\x35"
-     "\x02\x00"
-     "\xB5\xD1\x03\x00\xB4\x19\x02\x00\xB7\xD1\x03\x00\xB6\xD9\x03\x00\x82\x29"
-     "\x00\x00"
-     "\xA3\x95\x03\x00\x80\x1D\x00\x00\x81\x15\x00\x00\xA6\xBD\x03\x00\xE5\xC3"
-     "\x00\x80"
-     "\xE9\xC3\x00\x80\xA5\xB5\x03\x00\xAA\xED\x03\x00\xAB\xF1\x03\x00\xED\xC3"
-     "\x00\x80"
-     "\xF5\xC3\x00\x80\xAE\xDD\x03\x00\xAF\x15\x02\x00\xAC\xE1\x03\x00\xAD\xD5"
-     "\x03\x00"
-     "\x86\x60\x7D\x00\x87\x78\x7D\x00\xB3\x41\x01\x00\x84\x00\x7F\x00\xB5\x41"
-     "\x01\x00"
-     "\xF9\xC3\x00\x80\xFD\xC3\x00\x80\xB6\x49\x01\x00\x01\xC4\x00\x80\x05\xC4"
-     "\x00\x80"
-     "\xBB\x41\x01\x00\xBA\x4D\x01\x00\xBD\x49\x01\x00\xBC\x45\x01\x00\xBF\x29"
-     "\x01\x00"
-     "\xBE\x39\x01\x00\x09\xC4\x00\x80\xEF\xDC\x06\x00\x0D\xC4\x00\x80\x11\xC4"
-     "\x00\x80"
-     "\x15\xC4\x00\x80\x19\xC4\x00\x80\x1D\xC4\x00\x80\xEF\x30\x06\x00\x84\x2C"
-     "\x7E\x00"
-     "\xE1\xE0\x06\x00\x21\xC4\x00\x80\xE3\x88\x01\x00\x25\xC4\x00\x80\xE3\x74"
-     "\x00\x00"
-     "\x29\xC4\x00\x80\xE1\x2C\x01\x00\xA3\xC9\x01\x00\x2D\xC4\x00\x80\x31\xC4"
-     "\x00\x80"
-     "\x85\x70\x79\x00\x35\xC4\x00\x80\xA6\xC1\x01\x00\xA5\xC9\x01\x00\x39\xC4"
-     "\x00\x80"
-     "\xAB\xC9\x01\x00\xAA\xC5\x01\x00\x3D\xC4\x00\x80\x41\xC4\x00\x80\xAF\xA1"
-     "\x01\x00"
-     "\xAE\xB1\x01\x00\xAD\xC1\x01\x00\xAC\xCD\x01\x00\xA8\xDD\x7D\x00\xA9\x05"
-     "\x7E\x00"
-     "\xAA\x01\x7E\x00\xAB\x01\x7E\x00\xAC\x01\x7E\x00\xAD\x01\x7E\x00\xAE\x01"
-     "\x7E\x00"
-     "\xAF\x01\x7E\x00\xF1\xC3\x00\x80\x45\xC4\x00\x80\x49\xC4\x00\x80\x4D\xC4"
-     "\x00\x80"
-     "\x51\xC4\x00\x80\x82\x9D\x00\x00\x81\x9D\x00\x00\x80\x9D\x00\x00\xB8\x2D"
-     "\x7E\x00"
-     "\xB9\xE1\x7E\x00\xBA\xE1\x7E\x00\xBB\xE1\x7E\x00\xBC\xE1\x7E\x00\xBD\xE1"
-     "\x7E\x00"
-     "\xBE\xE1\x7E\x00\xBF\xE1\x7E\x00\xB0\x41\x7E\x00\xB1\x4D\x7E\x00\xB2\x59"
-     "\x7E\x00"
-     "\xB3\x55\x7E\x00\xB4\x35\x7E\x00\xB5\x25\x7E\x00\xB6\x25\x7E\x00\xB7\x15"
-     "\x7E\x00"
-     "\xA8\xAD\x7F\x00\xA9\xD1\x7F\x00\xAA\xD1\x7F\x00\xAB\xED\x7F\x00\xAC\xF5"
-     "\x7F\x00"
-     "\xAD\x15\x7F\x00\xAE\x11\x7F\x00\xAF\x11\x7F\x00\x55\xC4\x00\x80\x59\xC4"
-     "\x00\x80"
-     "\x5D\xC4\x00\x80\x61\xC4\x00\x80\x86\xF0\x03\x00\x87\xB8\x00\x00\x65\xC4"
-     "\x00\x80"
-     "\x69\xC4\x00\x80\xB8\x11\x7F\x00\xB9\x19\x7F\x00\xBA\x21\x7F\x00\xBB\x21"
-     "\x7F\x00"
-     "\xBC\xF5\x00\x00\xBD\xFD\x00\x00\xBE\xF5\x00\x00\xBF\xED\x00\x00\xB0\x71"
-     "\x7F\x00"
-     "\xB1\x71\x7F\x00\xB2\x71\x7F\x00\xB3\x45\x7F\x00\xB4\x41\x7F\x00\xB5\x4D"
-     "\x7F\x00"
-     "\xB6\x3D\x7F\x00\xB7\x35\x7F\x00\xB3\x59\x7E\x00\x6D\xC4\x00\x80\x71\xC4"
-     "\x00\x80"
-     "\x75\xC4\x00\x80\x79\xC4\x00\x80\xB6\x95\x7E\x00\xB5\x7D\x7E\x00\x7D\xC4"
-     "\x00\x80"
-     "\xBB\xB5\x7E\x00\xBA\xB5\x7E\x00\x81\xC4\x00\x80\x85\xC4\x00\x80\xBF\x89"
-     "\x7E\x00"
-     "\xBE\x85\x7E\x00\xBD\x91\x7E\x00\xBC\xA5\x7E\x00\x89\xC4\x00\x80\xA3\x1D"
-     "\x7E\x00"
-     "\x8D\xC4\x00\x80\x91\xC4\x00\x80\xA6\xD1\x7E\x00\x95\xC4\x00\x80\x99\xC4"
-     "\x00\x80"
-     "\xA5\x39\x7E\x00\xAA\xF1\x7E\x00\xAB\xF1\x7E\x00\x9D\xC4\x00\x80\xA1\xC4"
-     "\x00\x80"
-     "\xAE\xC1\x7E\x00\xAF\xCD\x7E\x00\xAC\xE1\x7E\x00\xAD\xD5\x7E\x00\xB0\xAD"
-     "\x00\x00"
-     "\xB1\xC5\x00\x00\xB2\xC1\x00\x00\xB3\xC1\x00\x00\xB4\xC5\x00\x00\xB5\xCD"
-     "\x00\x00"
-     "\xB6\xF1\x00\x00\xB7\xF1\x00\x00\xB8\x61\x00\x00\xB9\x61\x00\x00\xBA\x61"
-     "\x00\x00"
-     "\xBB\x7D\x00\x00\xBC\x65\x00\x00\xBD\x6D\x00\x00\xBE\x65\x00\x00\xBF\xBD"
-     "\x03\x00"
-     "\xA5\xC4\x00\x80\xA9\xC4\x00\x80\xAD\xC4\x00\x80\x81\xC3\x00\x80\xB1\xC4"
-     "\x00\x80"
-     "\xB5\xC4\x00\x80\xB9\xC4\x00\x80\xBD\xC4\x00\x80\xA8\x59\x01\x00\xA9\x59"
-     "\x01\x00"
-     "\xAA\xED\x00\x00\xAB\xE5\x00\x00\xAC\xFD\x00\x00\xAD\xE5\x00\x00\xAE\xE5"
-     "\x00\x00"
-     "\xAF\xD5\x00\x00\xC1\xC4\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xC5\xC4\x00\x80\xC9\xC4\x00\x80\xCD\xC4\x00\x80\xBE\x54\x02\x00\x86\x80"
-     "\x04\x00"
-     "\x87\xEC\x02\x00\xD5\xC4\x00\x80\xD9\xC4\x00\x80\xDD\xC4\x00\x80\xE1\xC4"
-     "\x00\x80"
-     "\xE5\xC4\x00\x80\xBE\x78\x03\x00\xE3\x74\x7E\x00\xE9\xC4\x00\x80\xE1\xB8"
-     "\x7D\x00"
-     "\xED\xC4\x00\x80\xF1\xC4\x00\x80\xF5\xC4\x00\x80\xF9\xC4\x00\x80\xFD\xC4"
-     "\x00\x80"
-     "\x01\xC5\x00\x80\x05\xC5\x00\x80\x09\xC5\x00\x80\x0D\xC5\x00\x80\xEF\xC0"
-     "\x7E\x00"
-     "\x11\xC5\x00\x80\x15\xC5\x00\x80\x19\xC5\x00\x80\xB3\x85\x03\x00\x1D\xC5"
-     "\x00\x80"
-     "\x21\xC5\x00\x80\x25\xC5\x00\x80\x29\xC5\x00\x80\xB6\x85\x03\x00\xB5\x95"
-     "\x03\x00"
-     "\x2D\xC5\x00\x80\xBB\xE1\x03\x00\xBA\x89\x03\x00\xBE\x24\x06\x00\x31\xC5"
-     "\x00\x80"
-     "\xBF\xE9\x03\x00\xBE\xE1\x03\x00\xBD\xF5\x03\x00\xBC\xF5\x03\x00\x82\x29"
-     "\x00\x00"
-     "\xA3\xC1\x03\x00\x80\x1D\x00\x00\x81\x15\x00\x00\xA6\xC1\x03\x00\x35\xC5"
-     "\x00\x80"
-     "\x39\xC5\x00\x80\xA5\xD1\x03\x00\xAA\xCD\x03\x00\xAB\xA5\x03\x00\x3D\xC5"
-     "\x00\x80"
-     "\x85\xE0\x05\x00\xAE\xA5\x03\x00\xAF\xAD\x03\x00\xAC\xB1\x03\x00\xAD\xB1"
-     "\x03\x00"
-     "\xE1\xF8\x03\x00\x45\xC5\x00\x80\xE3\x5C\x1F\x00\x49\xC5\x00\x80\xEF\xF4"
-     "\x03\x00"
-     "\x4D\xC5\x00\x80\x86\x3C\x07\x00\x87\xAC\x02\x00\xE3\x7C\x7E\x00\x51\xC5"
-     "\x00\x80"
-     "\xE1\x80\x01\x00\x55\xC5\x00\x80\x59\xC5\x00\x80\x5D\xC5\x00\x80\xEF\xE4"
-     "\x13\x00"
-     "\x61\xC5\x00\x80\xB3\x71\x01\x00\x65\xC5\x00\x80\x69\xC5\x00\x80\x6D\xC5"
-     "\x00\x80"
-     "\x71\xC5\x00\x80\xB6\x11\x01\x00\xB5\x61\x01\x00\x75\xC5\x00\x80\xBB\x39"
-     "\x01\x00"
-     "\xBA\x31\x01\x00\x79\xC5\x00\x80\x7D\xC5\x00\x80\xBF\x19\x01\x00\xBE\x11"
-     "\x01\x00"
-     "\xBD\x19\x01\x00\xBC\x21\x01\x00\x81\xC5\x00\x80\x85\xC5\x00\x80\x89\xC5"
-     "\x00\x80"
-     "\x8D\xC5\x00\x80\x91\xC5\x00\x80\x95\xC5\x00\x80\x99\xC5\x00\x80\xEF\xC4"
-     "\x7F\x00"
-     "\x9D\xC5\x00\x80\xE1\xFC\x7E\x00\xA1\xC5\x00\x80\xE3\xF0\x7F\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x75\x00\x00\x82\x7D\x00\x00\xA5\xC5\x00\x80\xA9\xC5\x00\x80\xAD\xC5"
-     "\x00\x80"
-     "\xA3\xF9\x01\x00\xBE\x00\x08\x00\xA5\xE9\x01\x00\xB5\xC5\x00\x80\xB9\xC5"
-     "\x00\x80"
-     "\xA6\x99\x01\x00\x84\xA8\x05\x00\xBD\xC5\x00\x80\xAB\xB1\x01\x00\xAA\xB9"
-     "\x01\x00"
-     "\xAD\x91\x01\x00\xAC\xA9\x01\x00\xAF\x91\x01\x00\xAE\x99\x01\x00\xA8\x29"
-     "\x06\x00"
-     "\xA9\x29\x06\x00\xAA\x39\x06\x00\xAB\x39\x06\x00\xAC\x29\x06\x00\xAD\x51"
-     "\x06\x00"
-     "\xAE\x55\x06\x00\xAF\x4D\x06\x00\x41\xC5\x00\x80\x84\x20\x01\x00\xC1\xC5"
-     "\x00\x80"
-     "\xB1\xC5\x00\x80\xA3\xE1\x01\x00\xA2\x99\x04\x00\xA1\x19\x04\x00\xA0\xF1"
-     "\x05\x00"
-     "\xB8\x39\x06\x00\xB9\x39\x06\x00\xBA\xCD\x06\x00\xBB\xC5\x06\x00\xBC\xDD"
-     "\x06\x00"
-     "\xBD\xC5\x06\x00\xBE\xC5\x06\x00\xBF\xF1\x06\x00\xB0\x35\x06\x00\xB1\x3D"
-     "\x06\x00"
-     "\xB2\x35\x06\x00\xB3\x0D\x06\x00\xB4\x15\x06\x00\xB5\x1D\x06\x00\xB6\x15"
-     "\x06\x00"
-     "\xB7\x09\x06\x00\x8F\xA1\x6C\x00\xB3\x91\x07\x00\x86\x28\x01\x00\x87\x7C"
-     "\x03\x00"
-     "\xB6\xA1\x07\x00\xC5\xC5\x00\x80\xC9\xC5\x00\x80\xB5\xB1\x07\x00\xBA\xE5"
-     "\x07\x00"
-     "\xBB\xED\x07\x00\xCD\xC5\x00\x80\xD1\xC5\x00\x80\xBE\xED\x07\x00\xBF\xDD"
-     "\x07\x00"
-     "\xBC\xF5\x07\x00\xBD\xED\x07\x00\x9F\xF9\x78\x00\x9E\xE5\x78\x00\x9D\x71"
-     "\x79\x00"
-     "\x9C\x21\x7C\x00\x9B\xF1\x7C\x00\x9A\x61\x7D\x00\x99\x99\x71\x00\x98\xD9"
-     "\x70\x00"
-     "\x97\xE1\x70\x00\x96\x79\x74\x00\x95\x6D\x74\x00\x94\x6D\x74\x00\x93\xAD"
-     "\x69\x00"
-     "\x92\x71\x68\x00\x91\x79\x68\x00\x90\x1D\x6E\x00\x82\x21\x6D\x00\x83\xE5"
-     "\x6F\x00"
-     "\xD5\xC5\x00\x80\xD9\xC5\x00\x80\x86\x4D\x68\x00\x87\x95\x75\x00\x84\x99"
-     "\x69\x00"
-     "\x85\x99\x69\x00\x8A\xA5\x75\x00\x8B\xB9\x75\x00\xDD\xC5\x00\x80\xE1\xC5"
-     "\x00\x80"
-     "\x8E\x71\x70\x00\x8F\x81\x7C\x00\x8C\x39\x71\x00\x8D\x39\x71\x00\x92\x61"
-     "\x7D\x00"
-     "\x93\xA9\x7D\x00\xE5\xC5\x00\x80\xE9\xC5\x00\x80\x96\x69\x79\x00\x97\x99"
-     "\x04\x00"
-     "\x94\xE1\x78\x00\x95\x7D\x79\x00\x9A\x41\x05\x00\x9B\xC9\x05\x00\xED\xC5"
-     "\x00\x80"
-     "\xF1\xC5\x00\x80\xF5\xC5\x00\x80\xF9\xC5\x00\x80\x9C\xA9\x00\x00\xFD\xC5"
-     "\x00\x80"
-     "\xA3\x81\x0D\x00\xA2\xA9\x01\x00\xA1\xA9\x01\x00\x01\xC6\x00\x80\xA7\xB1"
-     "\x09\x00"
-     "\xA6\x01\x08\x00\xA5\x4D\x0D\x00\xA4\x99\x0D\x00\xAB\x91\x15\x00\xAA\x85"
-     "\x15\x00"
-     "\xA9\x01\x14\x00\xA8\x71\x09\x00\xAF\x1D\x10\x00\xAE\xE9\x11\x00\xAD\xBD"
-     "\x11\x00"
-     "\xAC\x01\x10\x00\xB3\x01\x18\x00\xB2\xF1\x1C\x00\xB1\xC1\x1D\x00\xB0\xC9"
-     "\x1D\x00"
-     "\xD1\xC4\x00\x80\x05\xC6\x00\x80\xB5\xE1\x18\x00\xB4\xFD\x19\x00\x09\xC6"
-     "\x00\x80"
-     "\x0D\xC6\x00\x80\x11\xC6\x00\x80\x15\xC6\x00\x80\x80\x1D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\xFD\x03\x00\x19\xC6\x00\x80\xA3\x15\x05\x00\x1D\xC6\x00\x80\x86\x88"
-     "\x0C\x00"
-     "\x87\x3C\x03\x00\x25\xC6\x00\x80\xA6\x25\x05\x00\xA5\x35\x05\x00\x29\xC6"
-     "\x00\x80"
-     "\xAB\x69\x05\x00\xAA\x61\x05\x00\x2D\xC6\x00\x80\x31\xC6\x00\x80\xAF\x59"
-     "\x05\x00"
-     "\xAE\x69\x05\x00\xAD\x69\x05\x00\xAC\x71\x05\x00\x35\xC6\x00\x80\x39\xC6"
-     "\x00\x80"
-     "\x84\x04\x0C\x00\x3D\xC6\x00\x80\x41\xC6\x00\x80\xE1\x6C\x06\x00\x45\xC6"
-     "\x00\x80"
-     "\xE3\xEC\x7B\x00\x49\xC6\x00\x80\x4D\xC6\x00\x80\x51\xC6\x00\x80\xEF\xA8"
-     "\x06\x00"
-     "\x55\xC6\x00\x80\x59\xC6\x00\x80\x5D\xC6\x00\x80\x61\xC6\x00\x80\xA9\x8D"
-     "\x05\x00"
-     "\xA8\x81\x05\x00\xAB\xAD\x05\x00\xAA\x99\x05\x00\xAD\xA1\x05\x00\xAC\xB9"
-     "\x05\x00"
-     "\xAF\xA5\x05\x00\xAE\xA9\x05\x00\x84\x68\x0D\x00\x65\xC6\x00\x80\x69\xC6"
-     "\x00\x80"
-     "\x6D\xC6\x00\x80\x71\xC6\x00\x80\x75\xC6\x00\x80\x79\xC6\x00\x80\xBE\xF4"
-     "\x0C\x00"
-     "\xB9\x49\x05\x00\xB8\x41\x05\x00\xBB\x59\x05\x00\xBA\x41\x05\x00\xBD\x49"
-     "\x05\x00"
-     "\xBC\x41\x05\x00\xBF\x71\x05\x00\xBE\x7D\x05\x00\xB1\xA9\x05\x00\xB0\xA1"
-     "\x05\x00"
-     "\xB3\xB9\x05\x00\xB2\xA1\x05\x00\xB5\x99\x05\x00\xB4\xA9\x05\x00\xB7\x79"
-     "\x05\x00"
-     "\xB6\x91\x05\x00\xA8\xD5\x04\x00\xA9\xDD\x04\x00\xAA\xED\x04\x00\xAB\x25"
-     "\x03\x00"
-     "\xAC\x85\x03\x00\xAD\x8D\x03\x00\xAE\xB1\x03\x00\xAF\xB1\x03\x00\x7D\xC6"
-     "\x00\x80"
-     "\x81\xC6\x00\x80\x85\xC6\x00\x80\x89\xC6\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x8D\xC6\x00\x80\xB8\x31\x02\x00\xB9\x31\x02\x00\xBA\x35"
-     "\x02\x00"
-     "\xBB\xC1\x02\x00\xBC\x85\x02\x00\xBD\xB5\x02\x00\xBE\xBD\x02\x00\xBF\xB5"
-     "\x02\x00"
-     "\xB0\x69\x02\x00\xB1\x69\x02\x00\xB2\x41\x02\x00\xB3\x41\x02\x00\xB4\x39"
-     "\x02\x00"
-     "\xB5\x39\x02\x00\xB6\x11\x02\x00\xB7\x11\x02\x00\x86\xA0\x0C\x00\x87\x4C"
-     "\x0D\x00"
-     "\x95\xC6\x00\x80\x99\xC6\x00\x80\xEF\xA4\x06\x00\x9D\xC6\x00\x80\xA1\xC6"
-     "\x00\x80"
-     "\xEF\xCC\x07\x00\xE3\x94\x01\x00\xE1\xA4\x06\x00\xE1\x38\x01\x00\xE3\x5C"
-     "\x06\x00"
-     "\xA5\xC6\x00\x80\xA9\xC6\x00\x80\xAD\xC6\x00\x80\xB1\xC6\x00\x80\xB5\xC6"
-     "\x00\x80"
-     "\xB9\xC6\x00\x80\xB3\x2D\x04\x00\xBD\xC6\x00\x80\xB5\x45\x03\x00\xC1\xC6"
-     "\x00\x80"
-     "\xC5\xC6\x00\x80\xB6\x45\x03\x00\xC9\xC6\x00\x80\xCD\xC6\x00\x80\xBB\xE5"
-     "\x02\x00"
-     "\xBA\xE5\x02\x00\xBD\xDD\x02\x00\xBC\xDD\x02\x00\xBF\xB5\x02\x00\xBE\xB5"
-     "\x02\x00"
-     "\x21\xC6\x00\x80\x91\xC6\x00\x80\xD1\xC6\x00\x80\xD5\xC6\x00\x80\xD9\xC6"
-     "\x00\x80"
-     "\xDD\xC6\x00\x80\xE1\xC6\x00\x80\xE5\xC6\x00\x80\xA8\x31\x0E\x00\xA9\x39"
-     "\x0E\x00"
-     "\xAA\x01\x0E\x00\xAB\x01\x0E\x00\xAC\x71\x0E\x00\xAD\x71\x0E\x00\xAE\x75"
-     "\x0E\x00"
-     "\xAF\x6D\x0E\x00\xB0\x19\x0E\x00\xB1\x25\x0E\x00\xB2\x2D\x0E\x00\xB3\x25"
-     "\x0E\x00"
-     "\xB4\x21\x0E\x00\xB5\x21\x0E\x00\xB6\x21\x0E\x00\xB7\x21\x0E\x00\xB8\xC5"
-     "\x0E\x00"
-     "\xB9\xCD\x0E\x00\xBA\xC5\x0E\x00\xBB\xDD\x0E\x00\xBC\xC5\x0E\x00\xBD\xCD"
-     "\x0E\x00"
-     "\xBE\x59\x0F\x00\xBF\x59\x0F\x00\xB3\xA9\x0E\x00\xE9\xC6\x00\x80\xED\xC6"
-     "\x00\x80"
-     "\xF1\xC6\x00\x80\xF5\xC6\x00\x80\xB6\xD1\x0E\x00\xB5\xD9\x0E\x00\xF9\xC6"
-     "\x00\x80"
-     "\xBB\xC1\x0E\x00\xBA\xFD\x0E\x00\xFD\xC6\x00\x80\xBE\x2C\x00\x00\xBF\xC5"
-     "\x0E\x00"
-     "\xBE\xC5\x0E\x00\xBD\xD1\x0E\x00\xBC\xD9\x0E\x00\x82\x69\x00\x00\xA3\xED"
-     "\x0E\x00"
-     "\x80\x59\x00\x00\x81\x51\x00\x00\xA6\x95\x0E\x00\x01\xC7\x00\x80\x05\xC7"
-     "\x00\x80"
-     "\xA5\x9D\x0E\x00\xAA\xB9\x0E\x00\xAB\x85\x0E\x00\x86\xC8\x00\x00\x87\xAC"
-     "\x00\x00"
-     "\xAE\x81\x0E\x00\xAF\x81\x0E\x00\xAC\x9D\x0E\x00\xAD\x95\x0E\x00\x09\xC7"
-     "\x00\x80"
-     "\xB3\x91\x0E\x00\x0D\xC7\x00\x80\x11\xC7\x00\x80\xB6\xA5\x0E\x00\x15\xC7"
-     "\x00\x80"
-     "\x19\xC7\x00\x80\xB5\xAD\x0E\x00\xBA\xE1\x0E\x00\xBB\xE1\x0E\x00\x1D\xC7"
-     "\x00\x80"
-     "\x21\xC7\x00\x80\xBE\xE9\x0E\x00\xBF\xD5\x0E\x00\xBC\xF5\x0E\x00\xBD\xE9"
-     "\x0E\x00"
-     "\xA3\xA5\x0A\x00\x25\xC7\x00\x80\x29\xC7\x00\x80\x2D\xC7\x00\x80\x31\xC7"
-     "\x00\x80"
-     "\xA6\xCD\x0D\x00\xA5\xCD\x0D\x00\x35\xC7\x00\x80\xAB\x6D\x0C\x00\xAA\x6D"
-     "\x0C\x00"
-     "\x39\xC7\x00\x80\x3D\xC7\x00\x80\xAF\x3D\x0C\x00\xAE\x3D\x0C\x00\xAD\x55"
-     "\x0C\x00"
-     "\xAC\x55\x0C\x00\xA8\x09\x0E\x00\xA9\x09\x0E\x00\xAA\x1D\x0E\x00\xAB\x15"
-     "\x0E\x00"
-     "\xAC\x21\x0E\x00\xAD\x21\x0E\x00\xAE\x21\x0E\x00\xAF\x21\x0E\x00\x41\xC7"
-     "\x00\x80"
-     "\x45\xC7\x00\x80\x49\xC7\x00\x80\x4D\xC7\x00\x80\x51\xC7\x00\x80\x55\xC7"
-     "\x00\x80"
-     "\x59\xC7\x00\x80\x5D\xC7\x00\x80\xB8\xCD\x01\x00\xB9\xD5\x01\x00\xBA\xDD"
-     "\x01\x00"
-     "\xBB\xD5\x01\x00\xBC\xCD\x01\x00\xBD\x51\x01\x00\xBE\x51\x01\x00\xBF\x51"
-     "\x01\x00"
-     "\xB0\x21\x0E\x00\xB1\x21\x0E\x00\xB2\x25\x0E\x00\xB3\x39\x0E\x00\xB4\x29"
-     "\x0E\x00"
-     "\xB5\x15\x0E\x00\xB6\x1D\x0E\x00\xB7\xF5\x01\x00\x61\xC7\x00\x80\x65\xC7"
-     "\x00\x80"
-     "\x69\xC7\x00\x80\xA3\x99\x0D\x00\x6D\xC7\x00\x80\xA5\xA5\x0D\x00\xA6\xAD"
-     "\x0D\x00"
-     "\xBE\xDC\x02\x00\x84\xEC\x03\x00\x75\xC7\x00\x80\xAA\xE9\x0D\x00\xAB\xE9"
-     "\x0D\x00"
-     "\xAC\xFD\x0D\x00\xAD\xE1\x0D\x00\xAE\xE1\x0D\x00\xAF\xDD\x0D\x00\x80\x45"
-     "\x00\x00"
-     "\x81\x4D\x00\x00\x82\x59\x00\x00\xA3\x45\x03\x00\x79\xC7\x00\x80\xA5\x41"
-     "\x03\x00"
-     "\xA6\x41\x03\x00\x7D\xC7\x00\x80\x86\xC0\x04\x00\x87\x00\x03\x00\xAA\x2D"
-     "\x03\x00"
-     "\xAB\x25\x03\x00\xAC\x3D\x03\x00\xAD\x25\x03\x00\xAE\x25\x03\x00\xAF\x15"
-     "\x03\x00"
-     "\xA8\x59\x02\x00\xA9\x85\x03\x00\xAA\x81\x03\x00\xAB\x81\x03\x00\xAC\x85"
-     "\x03\x00"
-     "\xAD\x8D\x03\x00\xAE\xB1\x03\x00\xAF\xB1\x03\x00\x81\xC7\x00\x80\x85\xC7"
-     "\x00\x80"
-     "\x89\xC7\x00\x80\x8D\xC7\x00\x80\x91\xC7\x00\x80\x95\xC7\x00\x80\x99\xC7"
-     "\x00\x80"
-     "\x9D\xC7\x00\x80\xB8\x65\x03\x00\xB9\x6D\x03\x00\xBA\x65\x03\x00\xBB\x7D"
-     "\x03\x00"
-     "\xBC\x65\x03\x00\xBD\x6D\x03\x00\xBE\x65\x03\x00\xBF\xDD\x00\x00\xB0\xA5"
-     "\x03\x00"
-     "\xB1\xAD\x03\x00\xB2\xA5\x03\x00\xB3\xBD\x03\x00\xB4\xAD\x03\x00\xB5\x9D"
-     "\x03\x00"
-     "\xB6\x95\x03\x00\xB7\x5D\x03\x00\xB3\x09\x02\x00\xA1\xC7\x00\x80\xA5\xC7"
-     "\x00\x80"
-     "\xA9\xC7\x00\x80\xAD\xC7\x00\x80\xB6\x0D\x02\x00\xB5\x0D\x02\x00\xB1\xC7"
-     "\x00\x80"
-     "\xBB\x69\x02\x00\xBA\x61\x02\x00\xB5\xC7\x00\x80\xB9\xC7\x00\x80\xBF\x59"
-     "\x02\x00"
-     "\xBE\x69\x02\x00\xBD\x69\x02\x00\xBC\x71\x02\x00\xBD\xC7\x00\x80\xC1\xC7"
-     "\x00\x80"
-     "\xC5\xC7\x00\x80\xC9\xC7\x00\x80\xE1\xA0\x01\x00\xCD\xC7\x00\x80\xE3\x68"
-     "\x03\x00"
-     "\xD1\xC7\x00\x80\x81\x15\x00\x00\x80\x15\x00\x00\xEF\x8C\x03\x00\x82\x15"
-     "\x00\x00"
-     "\xD5\xC7\x00\x80\xD9\xC7\x00\x80\xDD\xC7\x00\x80\xBE\x70\x05\x00\xE1\x18"
-     "\x0E\x00"
-     "\xE1\x94\x0F\x00\xE3\x38\x0F\x00\xE3\xD4\x0F\x00\x84\x94\x02\x00\xE5\xC7"
-     "\x00\x80"
-     "\xE9\xC7\x00\x80\xED\xC7\x00\x80\xF1\xC7\x00\x80\xF5\xC7\x00\x80\xF9\xC7"
-     "\x00\x80"
-     "\xFD\xC7\x00\x80\x01\xC8\x00\x80\x05\xC8\x00\x80\xEF\xEC\x01\x00\xEF\xF8"
-     "\x0F\x00"
-     "\x86\x60\x04\x00\x87\x04\x05\x00\xB3\x95\x01\x00\x84\xCC\x05\x00\xB5\x75"
-     "\x01\x00"
-     "\x09\xC8\x00\x80\x0D\xC8\x00\x80\xB6\x75\x01\x00\x11\xC8\x00\x80\x15\xC8"
-     "\x00\x80"
-     "\xBB\x51\x01\x00\xBA\x59\x01\x00\xBD\xF5\x00\x00\xBC\x49\x01\x00\xBF\xF5"
-     "\x00\x00"
-     "\xBE\xF5\x00\x00\xA8\x25\x06\x00\xA9\x55\x06\x00\xAA\x55\x06\x00\xAB\xAD"
-     "\x06\x00"
-     "\xAC\xB5\x06\x00\xAD\xBD\x06\x00\xAE\xB5\x06\x00\xAF\xAD\x06\x00\xE1\xC7"
-     "\x00\x80"
-     "\x19\xC8\x00\x80\x1D\xC8\x00\x80\x21\xC8\x00\x80\x25\xC8\x00\x80\x29\xC8"
-     "\x00\x80"
-     "\x2D\xC8\x00\x80\x31\xC8\x00\x80\xB8\x69\x07\x00\xB9\x69\x07\x00\xBA\x09"
-     "\x07\x00"
-     "\xBB\x09\x07\x00\xBC\x19\x07\x00\xBD\x19\x07\x00\xBE\x0D\x07\x00\xBF\x01"
-     "\x07\x00"
-     "\xB0\xD5\x06\x00\xB1\xDD\x06\x00\xB2\xD5\x06\x00\xB3\x69\x07\x00\xB4\x79"
-     "\x07\x00"
-     "\xB5\x79\x07\x00\xB6\x69\x07\x00\xB7\x61\x07\x00\xA3\xDD\x06\x00\x35\xC8"
-     "\x00\x80"
-     "\x39\xC8\x00\x80\x3D\xC8\x00\x80\x41\xC8\x00\x80\xA6\x3D\x06\x00\xA5\x3D"
-     "\x06\x00"
-     "\x45\xC8\x00\x80\xAB\x19\x06\x00\xAA\x11\x06\x00\x49\xC8\x00\x80\x4D\xC8"
-     "\x00\x80"
-     "\xAF\xBD\x07\x00\xAE\xBD\x07\x00\xAD\xBD\x07\x00\xAC\x01\x06\x00\x80\x5D"
-     "\x00\x00"
-     "\x81\x6D\x00\x00\x82\x65\x00\x00\xB3\x51\x07\x00\xBE\xD8\x03\x00\xB5\x71"
-     "\x07\x00"
-     "\xB6\x71\x07\x00\x51\xC8\x00\x80\x86\xE0\x00\x00\x87\x14\x03\x00\xBA\x5D"
-     "\x07\x00"
-     "\xBB\x39\x07\x00\xBC\x29\x07\x00\xBD\x15\x07\x00\xBE\x1D\x07\x00\xBF\xD9"
-     "\x00\x00"
-     "\xA8\x95\x06\x00\xA9\x9D\x06\x00\xAA\x95\x06\x00\xAB\xAD\x06\x00\xAC\xB5"
-     "\x06\x00"
-     "\xAD\xBD\x06\x00\xAE\xB5\x06\x00\xAF\xAD\x06\x00\x55\xC8\x00\x80\x59\xC8"
-     "\x00\x80"
-     "\x5D\xC8\x00\x80\x61\xC8\x00\x80\x65\xC8\x00\x80\x69\xC8\x00\x80\x6D\xC8"
-     "\x00\x80"
-     "\x71\xC8\x00\x80\xB8\x71\x01\x00\xB9\x71\x01\x00\xBA\x71\x01\x00\xBB\x71"
-     "\x01\x00"
-     "\xBC\xD5\x01\x00\xBD\xDD\x01\x00\xBE\xD5\x01\x00\xBF\xCD\x01\x00\xB0\xD5"
-     "\x06\x00"
-     "\xB1\xB1\x06\x00\xB2\xB1\x06\x00\xB3\x8D\x06\x00\xB4\x95\x06\x00\xB5\x51"
-     "\x01\x00"
-     "\xB6\x51\x01\x00\xB7\x51\x01\x00\x75\xC8\x00\x80\xA3\x19\x06\x00\x79\xC8"
-     "\x00\x80"
-     "\x7D\xC8\x00\x80\xA6\x39\x06\x00\x71\xC7\x00\x80\x81\xC8\x00\x80\xA5\x39"
-     "\x06\x00"
-     "\xAA\x15\x06\x00\xAB\x71\x06\x00\x85\xC8\x00\x80\x89\xC8\x00\x80\xAE\x55"
-     "\x06\x00"
-     "\xAF\x91\x01\x00\xAC\x61\x06\x00\xAD\x5D\x06\x00\x8D\xC8\x00\x80\x91\xC8"
-     "\x00\x80"
-     "\x95\xC8\x00\x80\x99\xC8\x00\x80\x9D\xC8\x00\x80\xA1\xC8\x00\x80\xA5\xC8"
-     "\x00\x80"
-     "\xA9\xC8\x00\x80\xAD\xC8\x00\x80\xB1\xC8\x00\x80\xB5\xC8\x00\x80\xB9\xC8"
-     "\x00\x80"
-     "\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\xBD\xC8\x00\x80\x84\x80"
-     "\x02\x00"
-     "\xBE\x80\x03\x00\x87\x0C\x03\x00\x86\xAC\x1C\x00\xE1\x68\x06\x00\xC5\xC8"
-     "\x00\x80"
-     "\xE3\x98\x07\x00\xC9\xC8\x00\x80\xCD\xC8\x00\x80\xD1\xC8\x00\x80\xEF\xAC"
-     "\x07\x00"
-     "\xD5\xC8\x00\x80\xD9\xC8\x00\x80\xDD\xC8\x00\x80\xE1\xC8\x00\x80\xE5\xC8"
-     "\x00\x80"
-     "\xE9\xC8\x00\x80\xED\xC8\x00\x80\xF1\xC8\x00\x80\xB3\x65\x03\x00\xF5\xC8"
-     "\x00\x80"
-     "\xB5\x65\x03\x00\xB6\x6D\x03\x00\xF9\xC8\x00\x80\xFD\xC8\x00\x80\x01\xC9"
-     "\x00\x80"
-     "\xBA\xE9\x03\x00\xBB\xE5\x03\x00\xBC\xFD\x03\x00\xBD\xED\x03\x00\xBE\xD1"
-     "\x03\x00"
-     "\xBF\xD1\x03\x00\x05\xC9\x00\x80\x09\xC9\x00\x80\x0D\xC9\x00\x80\x11\xC9"
-     "\x00\x80"
-     "\x15\xC9\x00\x80\x19\xC9\x00\x80\x1D\xC9\x00\x80\x21\xC9\x00\x80\xA8\x81"
-     "\x03\x00"
-     "\xA9\x81\x03\x00\xAA\x81\x03\x00\xAB\x81\x03\x00\xAC\x81\x03\x00\xAD\x81"
-     "\x03\x00"
-     "\xAE\x81\x03\x00\xAF\x81\x03\x00\xB0\x41\x03\x00\xB1\x4D\x03\x00\xB2\x45"
-     "\x03\x00"
-     "\xB3\x55\x03\x00\xB4\x79\x03\x00\xB5\x79\x03\x00\xB6\x19\x03\x00\xB7\x19"
-     "\x03\x00"
-     "\xB8\x29\x03\x00\xB9\x29\x03\x00\xBA\x39\x03\x00\xBB\x39\x03\x00\xBC\x29"
-     "\x03\x00"
-     "\xBD\x29\x03\x00\xBE\x19\x03\x00\xBF\x19\x03\x00\x81\x19\x00\x00\x80\x11"
-     "\x00\x00"
-     "\xA3\x21\x02\x00\x82\x2D\x00\x00\xA5\x21\x02\x00\x25\xC9\x00\x80\x29\xC9"
-     "\x00\x80"
-     "\xA6\x29\x02\x00\x2D\xC9\x00\x80\x35\xC9\x00\x80\xAB\xA1\x02\x00\xAA\xAD"
-     "\x02\x00"
-     "\xAD\xA9\x02\x00\xAC\xB9\x02\x00\xAF\x95\x02\x00\xAE\x95\x02\x00\x84\x4C"
-     "\x02\x00"
-     "\xBE\x48\x1D\x00\x87\x64\x1D\x00\x86\xEC\x1C\x00\xE3\x40\x03\x00\x39\xC9"
-     "\x00\x80"
-     "\xE1\xA0\x01\x00\x3D\xC9\x00\x80\xEF\x9C\x03\x00\x41\xC9\x00\x80\x45\xC9"
-     "\x00\x80"
-     "\x49\xC9\x00\x80\x4D\xC9\x00\x80\x51\xC9\x00\x80\x55\xC9\x00\x80\x59\xC9"
-     "\x00\x80"
-     "\x5D\xC9\x00\x80\x61\xC9\x00\x80\x65\xC9\x00\x80\x69\xC9\x00\x80\x6D\xC9"
-     "\x00\x80"
-     "\x71\xC9\x00\x80\x75\xC9\x00\x80\xEF\xB4\x01\x00\x84\xA8\x1D\x00\xE1\x7C"
-     "\x06\x00"
-     "\x79\xC9\x00\x80\xE3\x70\x06\x00\x7D\xC9\x00\x80\x81\xC9\x00\x80\x85\xC9"
-     "\x00\x80"
-     "\x89\xC9\x00\x80\xB3\xE1\x01\x00\x8D\xC9\x00\x80\x91\xC9\x00\x80\x95\xC9"
-     "\x00\x80"
-     "\x99\xC9\x00\x80\xB6\xE5\x01\x00\xB5\xF5\x01\x00\x9D\xC9\x00\x80\xBB\xA1"
-     "\x01\x00"
-     "\xBA\xD9\x01\x00\xBE\xE4\x1C\x00\xA5\xC9\x00\x80\xBF\xAD\x01\x00\xBE\xA5"
-     "\x01\x00"
-     "\xBD\xB1\x01\x00\xBC\xB9\x01\x00\xA8\x15\x1E\x00\xA9\x19\x1E\x00\xAA\x29"
-     "\x1E\x00"
-     "\xAB\x29\x1E\x00\xAC\x3D\x1E\x00\xAD\x25\x1E\x00\xAE\x2D\x1E\x00\xAF\x25"
-     "\x1E\x00"
-     "\x31\xC9\x00\x80\x82\xF9\x1F\x00\x81\xF9\x1F\x00\x80\xE1\x1F\x00\xA1\xC9"
-     "\x00\x80"
-     "\xA9\xC9\x00\x80\x86\x1C\x00\x00\x87\xB0\x03\x00\xB8\xC1\x1E\x00\xB9\xC1"
-     "\x1E\x00"
-     "\xBA\xC1\x1E\x00\xBB\xC1\x1E\x00\xBC\xC1\x1E\x00\xBD\xC1\x1E\x00\xBE\xC1"
-     "\x1E\x00"
-     "\xBF\xC1\x1E\x00\xB0\x5D\x1E\x00\xB1\x25\x1E\x00\xB2\x2D\x1E\x00\xB3\x25"
-     "\x1E\x00"
-     "\xB4\x21\x1E\x00\xB5\x29\x1E\x00\xB6\x19\x1E\x00\xB7\x19\x1E\x00\xA3\xA1"
-     "\x1E\x00"
-     "\xAD\xC9\x00\x80\xB1\xC9\x00\x80\xB5\xC9\x00\x80\xB9\xC9\x00\x80\xA6\xA5"
-     "\x1E\x00"
-     "\xA5\xB5\x1E\x00\xBD\xC9\x00\x80\xAB\xE1\x1E\x00\xAA\x99\x1E\x00\xC1\xC9"
-     "\x00\x80"
-     "\xC5\xC9\x00\x80\xAF\xED\x1E\x00\xAE\xE5\x1E\x00\xAD\xF1\x1E\x00\xAC\xF9"
-     "\x1E\x00"
-     "\xC9\xC9\x00\x80\xB3\x99\x1F\x00\xCD\xC9\x00\x80\xD1\xC9\x00\x80\xB6\xBD"
-     "\x1F\x00"
-     "\xD5\xC9\x00\x80\xD9\xC9\x00\x80\xB5\xB5\x1F\x00\xBA\x99\x1F\x00\xBB\x99"
-     "\x1F\x00"
-     "\xDD\xC9\x00\x80\xE1\xC9\x00\x80\xBE\x79\x1F\x00\xBF\x79\x1F\x00\xBC\x79"
-     "\x1F\x00"
-     "\xBD\x79\x1F\x00\xA8\x25\x1E\x00\xA9\x51\x1E\x00\xAA\x55\x1E\x00\xAB\x69"
-     "\x1E\x00"
-     "\xAC\x79\x1E\x00\xAD\x79\x1E\x00\xAE\x69\x1E\x00\xAF\x69\x1E\x00\xE5\xC9"
-     "\x00\x80"
-     "\xE9\xC9\x00\x80\xED\xC9\x00\x80\xF1\xC9\x00\x80\xF5\xC9\x00\x80\xF9\xC9"
-     "\x00\x80"
-     "\xFD\xC9\x00\x80\x01\xCA\x00\x80\xB8\xD9\x1E\x00\xB9\xED\x1E\x00\xBA\xF9"
-     "\x1E\x00"
-     "\xBB\xF9\x1E\x00\xBC\xE9\x1E\x00\xBD\xE9\x1E\x00\xBE\x9D\x1E\x00\xBF\x95"
-     "\x1E\x00"
-     "\xB0\x19\x1E\x00\xB1\x19\x1E\x00\xB2\xE9\x1E\x00\xB3\xE9\x1E\x00\xB4\xF9"
-     "\x1E\x00"
-     "\xB5\xF9\x1E\x00\xB6\xE9\x1E\x00\xB7\xE9\x1E\x00\xA3\xDD\x1E\x00\x82\x29"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x80\x1D\x00\x00\x05\xCA\x00\x80\xA6\xF9\x1E\x00\xA5\xF1"
-     "\x1E\x00"
-     "\x09\xCA\x00\x80\xAB\xDD\x1E\x00\xAA\xDD\x1E\x00\xC1\xC8\x00\x80\x0D\xCA"
-     "\x00\x80"
-     "\xAF\x3D\x1E\x00\xAE\x3D\x1E\x00\xAD\x3D\x1E\x00\xAC\x3D\x1E\x00\x84\xC8"
-     "\x02\x00"
-     "\xB3\x41\x01\x00\xBE\x0C\x01\x00\x15\xCA\x00\x80\xB6\x41\x01\x00\x19\xCA"
-     "\x00\x80"
-     "\x1D\xCA\x00\x80\xB5\x51\x01\x00\xBA\x4D\x01\x00\xBB\x25\x01\x00\x86\x48"
-     "\x00\x00"
-     "\x87\x50\x01\x00\xBE\x25\x01\x00\xBF\x2D\x01\x00\xBC\x31\x01\x00\xBD\x31"
-     "\x01\x00"
-     "\x21\xCA\x00\x80\x25\xCA\x00\x80\x84\x04\x03\x00\xBE\x80\x04\x00\x29\xCA"
-     "\x00\x80"
-     "\xEF\xA8\x1F\x00\x2D\xCA\x00\x80\x31\xCA\x00\x80\xBF\x28\x31\x00\xE3\x74"
-     "\x1F\x00"
-     "\x35\xCA\x00\x80\xE1\x38\x1E\x00\x39\xCA\x00\x80\x3D\xCA\x00\x80\x41\xCA"
-     "\x00\x80"
-     "\x45\xCA\x00\x80\x49\xCA\x00\x80\x4D\xCA\x00\x80\xA3\xCD\x02\x00\x51\xCA"
-     "\x00\x80"
-     "\xA5\xDD\x02\x00\x55\xCA\x00\x80\x59\xCA\x00\x80\xA6\xCD\x02\x00\x5D\xCA"
-     "\x00\x80"
-     "\x61\xCA\x00\x80\xAB\xA9\x02\x00\xAA\xC1\x02\x00\xAD\xBD\x02\x00\xAC\xBD"
-     "\x02\x00"
-     "\xAF\xA1\x02\x00\xAE\xA9\x02\x00\xA9\xB5\x02\x00\xA8\x69\x1D\x00\xAB\x01"
-     "\x02\x00"
-     "\xAA\x09\x02\x00\xAD\x01\x02\x00\xAC\x19\x02\x00\xAF\x31\x02\x00\xAE\x01"
-     "\x02\x00"
-     "\x84\x6C\x05\x00\x65\xCA\x00\x80\x69\xCA\x00\x80\x6D\xCA\x00\x80\x71\xCA"
-     "\x00\x80"
-     "\x82\x9D\x01\x00\x81\x9D\x01\x00\x80\x9D\x01\x00\xB9\xE5\x03\x00\xB8\xE5"
-     "\x03\x00"
-     "\xBB\xE5\x03\x00\xBA\xE5\x03\x00\xBD\xE5\x03\x00\xBC\xE5\x03\x00\xBF\xE5"
-     "\x03\x00"
-     "\xBE\xE5\x03\x00\xB1\x21\x02\x00\xB0\x49\x02\x00\xB3\x25\x02\x00\xB2\x25"
-     "\x02\x00"
-     "\xB5\x29\x02\x00\xB4\x21\x02\x00\xB7\x15\x02\x00\xB6\x15\x02\x00\xA8\xCD"
-     "\x02\x00"
-     "\xA9\xD1\x02\x00\xAA\xD1\x02\x00\xAB\x0D\x01\x00\xAC\x15\x01\x00\xAD\x05"
-     "\x01\x00"
-     "\xAE\x01\x01\x00\xAF\x01\x01\x00\x75\xCA\x00\x80\x79\xCA\x00\x80\x7D\xCA"
-     "\x00\x80"
-     "\x85\xCA\x00\x80\xBE\xF8\x04\x00\x89\xCA\x00\x80\x8D\xCA\x00\x80\x91\xCA"
-     "\x00\x80"
-     "\xB8\x15\x01\x00\xB9\x1D\x01\x00\xBA\x29\x01\x00\xBB\x29\x01\x00\xBC\xF5"
-     "\x01\x00"
-     "\xBD\xFD\x01\x00\xBE\xF5\x01\x00\xBF\xED\x01\x00\xB0\x49\x01\x00\xB1\x55"
-     "\x01\x00"
-     "\xB2\x5D\x01\x00\xB3\x55\x01\x00\xB4\x4D\x01\x00\xB5\x35\x01\x00\xB6\x3D"
-     "\x01\x00"
-     "\xB7\x31\x01\x00\x86\xA0\x05\x00\x87\xC8\x05\x00\x95\xCA\x00\x80\xEF\xBC"
-     "\x00\x00"
-     "\x99\xCA\x00\x80\x9D\xCA\x00\x80\xA1\xCA\x00\x80\xEF\x8C\x1E\x00\x84\x2C"
-     "\x07\x00"
-     "\xE1\xF0\x1E\x00\xA5\xCA\x00\x80\xE3\x1C\x1E\x00\xA9\xCA\x00\x80\xE1\x94"
-     "\x01\x00"
-     "\xAD\xCA\x00\x80\xE3\x6C\x00\x00\xB3\x19\x02\x00\xB1\xCA\x00\x80\xB5\xCA"
-     "\x00\x80"
-     "\xB9\xCA\x00\x80\x84\x00\x08\x00\xB6\xC5\x01\x00\xB5\xDD\x01\x00\xBD\xCA"
-     "\x00\x80"
-     "\xBB\xD1\x01\x00\xBA\xCD\x01\x00\xC1\xCA\x00\x80\xC5\xCA\x00\x80\xBF\xB1"
-     "\x01\x00"
-     "\xBE\xC9\x01\x00\xBD\xC1\x01\x00\xBC\xC9\x01\x00\xA3\xD9\x05\x00\x81\xCA"
-     "\x00\x80"
-     "\xC9\xCA\x00\x80\xCD\xCA\x00\x80\xD1\xCA\x00\x80\xA6\x05\x06\x00\xA5\x1D"
-     "\x06\x00"
-     "\xD5\xCA\x00\x80\xAB\x11\x06\x00\xAA\x0D\x06\x00\xD9\xCA\x00\x80\xDD\xCA"
-     "\x00\x80"
-     "\xAF\x71\x06\x00\xAE\x09\x06\x00\xAD\x01\x06\x00\xAC\x09\x06\x00\xE1\xCA"
-     "\x00\x80"
-     "\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\xE5\xCA\x00\x80\xE9\xCA"
-     "\x00\x80"
-     "\xED\xCA\x00\x80\xF1\xCA\x00\x80\x86\x40\x03\x00\x87\xC4\x03\x00\xF5\xCA"
-     "\x00\x80"
-     "\xF9\xCA\x00\x80\xFD\xCA\x00\x80\x01\xCB\x00\x80\x05\xCB\x00\x80\x09\xCB"
-     "\x00\x80"
-     "\xA8\xAD\x07\x00\xA9\xB1\x07\x00\xAA\xB1\x07\x00\xAB\xB1\x07\x00\xAC\x19"
-     "\x07\x00"
-     "\xAD\x05\x07\x00\xAE\x0D\x07\x00\xAF\x05\x07\x00\x0D\xCB\x00\x80\x11\xCB"
-     "\x00\x80"
-     "\x15\xCB\x00\x80\x19\xCB\x00\x80\x1D\xCB\x00\x80\x21\xCB\x00\x80\x25\xCB"
-     "\x00\x80"
-     "\x29\xCB\x00\x80\xB8\x2D\x07\x00\xB9\xCD\x00\x00\xBA\xC5\x00\x00\xBB\xDD"
-     "\x00\x00"
-     "\xBC\xCD\x00\x00\xBD\xFD\x00\x00\xBE\xF5\x00\x00\xBF\x9D\x00\x00\xB0\x49"
-     "\x07\x00"
-     "\xB1\x55\x07\x00\xB2\x51\x07\x00\xB3\x29\x07\x00\xB4\x39\x07\x00\xB5\x39"
-     "\x07\x00"
-     "\xB6\x25\x07\x00\xB7\x15\x07\x00\xB3\x39\x06\x00\x2D\xCB\x00\x80\x31\xCB"
-     "\x00\x80"
-     "\x35\xCB\x00\x80\x39\xCB\x00\x80\xB6\x85\x06\x00\xB5\x91\x06\x00\x3D\xCB"
-     "\x00\x80"
-     "\xBB\x91\x06\x00\xBA\x8D\x06\x00\x41\xCB\x00\x80\x45\xCB\x00\x80\xBF\xFD"
-     "\x06\x00"
-     "\xBE\xFD\x06\x00\xBD\xFD\x06\x00\xBC\x85\x06\x00\x49\xCB\x00\x80\xA3\x7D"
-     "\x06\x00"
-     "\x4D\xCB\x00\x80\x51\xCB\x00\x80\xA6\xC1\x06\x00\x55\xCB\x00\x80\x59\xCB"
-     "\x00\x80"
-     "\xA5\xD5\x06\x00\xAA\xC9\x06\x00\xAB\xD5\x06\x00\x5D\xCB\x00\x80\xBE\x6C"
-     "\x01\x00"
-     "\xAE\xB9\x06\x00\xAF\xB9\x06\x00\xAC\xC1\x06\x00\xAD\xB9\x06\x00\xA8\xE9"
-     "\x01\x00"
-     "\xA9\xE9\x01\x00\xAA\xF9\x01\x00\xAB\xF9\x01\x00\xAC\xE9\x01\x00\xAD\xE9"
-     "\x01\x00"
-     "\xAE\x39\x01\x00\xAF\x39\x01\x00\x80\xF5\x00\x00\x81\xFD\x00\x00\x82\xC1"
-     "\x00\x00"
-     "\x61\xCB\x00\x80\x86\x10\x00\x00\x87\x74\x01\x00\x65\xCB\x00\x80\x11\xCA"
-     "\x00\x80"
-     "\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5\x00\x00\xBB\xE5\x00\x00\xBC\xFD"
-     "\x00\x00"
-     "\xBD\x95\x00\x00\xBE\x91\x00\x00\xBF\x91\x00\x00\xB0\x49\x01\x00\xB1\x49"
-     "\x01\x00"
-     "\xB2\x59\x01\x00\xB3\x59\x01\x00\xB4\x49\x01\x00\xB5\x49\x01\x00\xB6\xFD"
-     "\x00\x00"
-     "\xB7\xF5\x00\x00\xEF\xF4\x06\x00\x69\xCB\x00\x80\x6D\xCB\x00\x80\x71\xCB"
-     "\x00\x80"
-     "\xEF\x30\x02\x00\x75\xCB\x00\x80\x79\xCB\x00\x80\x7D\xCB\x00\x80\xE3\xE0"
-     "\x07\x00"
-     "\x81\xCB\x00\x80\xE1\x80\x01\x00\x85\xCB\x00\x80\xE3\x28\x06\x00\x89\xCB"
-     "\x00\x80"
-     "\xE1\xC8\x05\x00\x8D\xCB\x00\x80\xB3\x31\x02\x00\x91\xCB\x00\x80\x95\xCB"
-     "\x00\x80"
-     "\x96\x00\x00\x00\x99\xCB\x00\x80\xB6\x29\x02\x00\xB5\x21\x02\x00\x9D\xCB"
-     "\x00\x80"
-     "\xBB\xCD\x02\x00\xBA\xCD\x02\x00\xA1\xCB\x00\x80\xA5\xCB\x00\x80\xBF\xCD"
-     "\x02\x00"
-     "\xBE\xCD\x02\x00\xBD\xCD\x02\x00\xBC\xCD\x02\x00\xA9\xCB\x00\x80\xAD\xCB"
-     "\x00\x80"
-     "\xB1\xCB\x00\x80\xA3\xFD\x02\x00\xB5\xCB\x00\x80\xA5\xED\x02\x00\xA6\xE5"
-     "\x02\x00"
-     "\xB9\xCB\x00\x80\xBD\xCB\x00\x80\xC1\xCB\x00\x80\xAA\x01\x02\x00\xAB\x01"
-     "\x02\x00"
-     "\xAC\x01\x02\x00\xAD\x01\x02\x00\xAE\x01\x02\x00\xAF\x01\x02\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x1D\x00\x00\xC5\xCB\x00\x80\xC9\xCB\x00\x80\xCD\xCB"
-     "\x00\x80"
-     "\xBE\x54\x0C\x00\xD5\xCB\x00\x80\x86\xC0\x0C\x00\x87\x24\x03\x00\xD9\xCB"
-     "\x00\x80"
-     "\xDD\xCB\x00\x80\xE1\xCB\x00\x80\xE5\xCB\x00\x80\xE9\xCB\x00\x80\xED\xCB"
-     "\x00\x80"
-     "\xA8\xB9\x02\x00\xA9\x01\x01\x00\xAA\x01\x01\x00\xAB\x01\x01\x00\xAC\x05"
-     "\x01\x00"
-     "\xAD\x0D\x01\x00\xAE\x05\x01\x00\xAF\x39\x01\x00\x84\xA8\x0D\x00\xF1\xCB"
-     "\x00\x80"
-     "\xF5\xCB\x00\x80\xF9\xCB\x00\x80\xFD\xCB\x00\x80\x01\xCC\x00\x80\x05\xCC"
-     "\x00\x80"
-     "\x09\xCC\x00\x80\xB8\x2D\x01\x00\xB9\xC5\x01\x00\xBA\xCD\x01\x00\xBB\xC5"
-     "\x01\x00"
-     "\xBC\xC1\x01\x00\xBD\xC9\x01\x00\xBE\xF9\x01\x00\xBF\xF9\x01\x00\xB0\x4D"
-     "\x01\x00"
-     "\xB1\x51\x01\x00\xB2\x55\x01\x00\xB3\x29\x01\x00\xB4\x39\x01\x00\xB5\x25"
-     "\x01\x00"
-     "\xB6\x25\x01\x00\xB7\x15\x01\x00\xE1\x18\x06\x00\x0D\xCC\x00\x80\xE3\x38"
-     "\x07\x00"
-     "\x11\xCC\x00\x80\x15\xCC\x00\x80\xBE\x58\x0C\x00\x19\xCC\x00\x80\x1D\xCC"
-     "\x00\x80"
-     "\x84\x6C\x0F\x00\x21\xCC\x00\x80\xBE\x60\x0F\x00\x25\xCC\x00\x80\x29\xCC"
-     "\x00\x80"
-     "\x2D\xCC\x00\x80\xEF\x70\x06\x00\x31\xCC\x00\x80\x80\x15\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x2D\x00\x00\x84\xCC\x0F\x00\xE3\x60\x06\x00\x35\xCC\x00\x80\xE1\xA0"
-     "\x01\x00"
-     "\x39\xCC\x00\x80\xEF\x74\x00\x00\x3D\xCC\x00\x80\x86\xC8\x0C\x00\x87\xFC"
-     "\x0C\x00"
-     "\x45\xCC\x00\x80\x49\xCC\x00\x80\x4D\xCC\x00\x80\x51\xCC\x00\x80\xA3\x09"
-     "\x0E\x00"
-     "\xD1\xCB\x00\x80\x41\xCC\x00\x80\x55\xCC\x00\x80\x59\xCC\x00\x80\xA6\x0D"
-     "\x0E\x00"
-     "\xA5\x0D\x0E\x00\x5D\xCC\x00\x80\xAB\x15\x0E\x00\xAA\x09\x0E\x00\x61\xCC"
-     "\x00\x80"
-     "\x65\xCC\x00\x80\xAF\x61\x0E\x00\xAE\x7D\x0E\x00\xAD\x7D\x0E\x00\xAC\x01"
-     "\x0E\x00"
-     "\x69\xCC\x00\x80\xB3\xA9\x0E\x00\x6D\xCC\x00\x80\x71\xCC\x00\x80\xB6\xA9"
-     "\x0E\x00"
-     "\x75\xCC\x00\x80\x79\xCC\x00\x80\xB5\xA9\x0E\x00\xBA\x49\x0F\x00\xBB\x49"
-     "\x0F\x00"
-     "\x7D\xCC\x00\x80\x81\xCC\x00\x80\xBE\x49\x0F\x00\xBF\x49\x0F\x00\xBC\x49"
-     "\x0F\x00"
-     "\xBD\x49\x0F\x00\xA8\x5D\x0E\x00\xA9\x6D\x0E\x00\xAA\x65\x0E\x00\xAB\x7D"
-     "\x0E\x00"
-     "\xAC\x65\x0E\x00\xAD\x6D\x0E\x00\xAE\x65\x0E\x00\xAF\xB9\x0F\x00\x85\xCC"
-     "\x00\x80"
-     "\x89\xCC\x00\x80\x8D\xCC\x00\x80\x91\xCC\x00\x80\x95\xCC\x00\x80\x99\xCC"
-     "\x00\x80"
-     "\x9D\xCC\x00\x80\xA1\xCC\x00\x80\xB8\x51\x0F\x00\xB9\x5D\x0F\x00\xBA\x55"
-     "\x0F\x00"
-     "\xBB\x69\x0F\x00\xBC\x7D\x0F\x00\xBD\x65\x0F\x00\xBE\x6D\x0F\x00\xBF\x61"
-     "\x0F\x00"
-     "\xB0\xC9\x0F\x00\xB1\xC9\x0F\x00\xB2\xD9\x0F\x00\xB3\xD9\x0F\x00\xB4\xC9"
-     "\x0F\x00"
-     "\xB5\xC9\x0F\x00\xB6\x7D\x0F\x00\xB7\x71\x0F\x00\xA5\xCC\x00\x80\xB5\x11"
-     "\x0F\x00"
-     "\xB6\x11\x0F\x00\xA9\xCC\x00\x80\x80\x11\x00\x00\x81\x19\x00\x00\x82\x29"
-     "\x00\x00"
-     "\xB3\x15\x0F\x00\xBC\x1D\x0F\x00\xBD\x61\x0F\x00\xBE\x61\x0F\x00\xBF\x7D"
-     "\x0F\x00"
-     "\xAD\xCC\x00\x80\xB1\xCC\x00\x80\xBA\x15\x0F\x00\xBB\x09\x0F\x00\xA3\xAD"
-     "\x0F\x00"
-     "\xB5\xCC\x00\x80\x86\xE8\x00\x00\x87\xC8\x01\x00\xB9\xCC\x00\x80\xA6\xAD"
-     "\x0F\x00"
-     "\xA5\xAD\x0F\x00\xBD\xCC\x00\x80\xAB\x4D\x0E\x00\xAA\x4D\x0E\x00\xC1\xCC"
-     "\x00\x80"
-     "\xC5\xCC\x00\x80\xAF\x4D\x0E\x00\xAE\x4D\x0E\x00\xAD\x4D\x0E\x00\xAC\x4D"
-     "\x0E\x00"
-     "\xA8\x71\x0E\x00\xA9\x71\x0E\x00\xAA\x71\x0E\x00\xAB\x71\x0E\x00\xAC\x9D"
-     "\x01\x00"
-     "\xAD\x85\x01\x00\xAE\x85\x01\x00\xAF\xB5\x01\x00\xBE\xEC\x00\x00\xC9\xCC"
-     "\x00\x80"
-     "\xCD\xCC\x00\x80\xD1\xCC\x00\x80\xD5\xCC\x00\x80\xD9\xCC\x00\x80\xDD\xCC"
-     "\x00\x80"
-     "\xE1\xCC\x00\x80\xB8\x61\x01\x00\xB9\x61\x01\x00\xBA\x61\x01\x00\xBB\x61"
-     "\x01\x00"
-     "\xBC\x61\x01\x00\xBD\x61\x01\x00\xBE\x61\x01\x00\xBF\x61\x01\x00\xB0\xCD"
-     "\x01\x00"
-     "\xB1\xA5\x01\x00\xB2\xA1\x01\x00\xB3\xA1\x01\x00\xB4\xA5\x01\x00\xB5\xAD"
-     "\x01\x00"
-     "\xB6\x91\x01\x00\xB7\x91\x01\x00\xB3\xF9\x0D\x00\xE5\xCC\x00\x80\xE9\xCC"
-     "\x00\x80"
-     "\xED\xCC\x00\x80\xF1\xCC\x00\x80\xB6\x05\x02\x00\xB5\x15\x02\x00\xF5\xCC"
-     "\x00\x80"
-     "\xBB\x61\x02\x00\xBA\x09\x02\x00\xF9\xCC\x00\x80\xFD\xCC\x00\x80\xBF\x69"
-     "\x02\x00"
-     "\xBE\x61\x02\x00\xBD\x75\x02\x00\xBC\x75\x02\x00\x01\xCD\x00\x80\xA3\xBD"
-     "\x0D\x00"
-     "\x05\xCD\x00\x80\x09\xCD\x00\x80\xA6\x41\x02\x00\x0D\xCD\x00\x80\x11\xCD"
-     "\x00\x80"
-     "\xA5\x51\x02\x00\xAA\x4D\x02\x00\xAB\x25\x02\x00\x15\xCD\x00\x80\x19\xCD"
-     "\x00\x80"
-     "\xAE\x25\x02\x00\xAF\x2D\x02\x00\xAC\x31\x02\x00\xAD\x31\x02\x00\x81\xED"
-     "\x00\x00"
-     "\x80\xED\x00\x00\xEF\xD0\x01\x00\x82\x1D\x00\x00\x1D\xCD\x00\x80\x25\xCD"
-     "\x00\x80"
-     "\x86\x38\x04\x00\x87\x50\x03\x00\x29\xCD\x00\x80\x2D\xCD\x00\x80\x31\xCD"
-     "\x00\x80"
-     "\x35\xCD\x00\x80\xE1\xE0\x01\x00\x39\xCD\x00\x80\xE3\x64\x0F\x00\x3D\xCD"
-     "\x00\x80"
-     "\x41\xCD\x00\x80\x45\xCD\x00\x80\x49\xCD\x00\x80\xB3\x91\x03\x00\x4D\xCD"
-     "\x00\x80"
-     "\xB5\xB9\x03\x00\xB6\x7D\x03\x00\x51\xCD\x00\x80\x55\xCD\x00\x80\x59\xCD"
-     "\x00\x80"
-     "\xBA\x59\x03\x00\xBB\x59\x03\x00\xBC\x49\x03\x00\xBD\x49\x03\x00\xBE\xFD"
-     "\x00\x00"
-     "\xBF\xF5\x00\x00\xA8\x45\x02\x00\xA9\x55\x02\x00\xAA\x55\x02\x00\xAB\x65"
-     "\x02\x00"
-     "\xAC\x7D\x02\x00\xAD\xB1\x02\x00\xAE\xB1\x02\x00\xAF\xB1\x02\x00\xBE\x68"
-     "\x05\x00"
-     "\x5D\xCD\x00\x80\x61\xCD\x00\x80\x65\xCD\x00\x80\x69\xCD\x00\x80\x6D\xCD"
-     "\x00\x80"
-     "\x71\xCD\x00\x80\x75\xCD\x00\x80\xB8\x5D\x01\x00\xB9\x6D\x01\x00\xBA\x65"
-     "\x01\x00"
-     "\xBB\x0D\x01\x00\xBC\x19\x01\x00\xBD\x19\x01\x00\xBE\x0D\x01\x00\xBF\x05"
-     "\x01\x00"
-     "\xB0\xD1\x02\x00\xB1\xD1\x02\x00\xB2\xD1\x02\x00\xB3\xD1\x02\x00\xB4\x75"
-     "\x01\x00"
-     "\xB5\x7D\x01\x00\xB6\x75\x01\x00\xB7\x6D\x01\x00\xE1\x78\x0F\x00\xE3\x34"
-     "\x0E\x00"
-     "\xE3\xB8\x0E\x00\xE1\x7C\x0E\x00\x79\xCD\x00\x80\x7D\xCD\x00\x80\x81\xCD"
-     "\x00\x80"
-     "\x85\xCD\x00\x80\x89\xCD\x00\x80\x8D\xCD\x00\x80\x95\xCD\x00\x80\x99\xCD"
-     "\x00\x80"
-     "\x9D\xCD\x00\x80\xEF\xE4\x0E\x00\xEF\xD4\x0E\x00\xA1\xCD\x00\x80\xA3\x9D"
-     "\x02\x00"
-     "\x82\x61\x00\x00\x81\x69\x00\x00\x80\x51\x00\x00\x84\x9C\x05\x00\xA6\x71"
-     "\x02\x00"
-     "\xA5\xB5\x02\x00\xA5\xCD\x00\x80\xAB\x55\x02\x00\xAA\x55\x02\x00\x86\x48"
-     "\x04\x00"
-     "\x87\xCC\x04\x00\xAF\xF9\x01\x00\xAE\xF1\x01\x00\xAD\x45\x02\x00\xAC\x45"
-     "\x02\x00"
-     "\xA8\x95\x06\x00\xA9\xA5\x06\x00\xAA\xAD\x06\x00\xAB\xA5\x06\x00\xAC\xBD"
-     "\x06\x00"
-     "\xAD\xA1\x06\x00\xAE\xA5\x06\x00\xAF\xDD\x06\x00\x91\xCD\x00\x80\xA9\xCD"
-     "\x00\x80"
-     "\xAD\xCD\x00\x80\xB1\xCD\x00\x80\xB5\xCD\x00\x80\xB9\xCD\x00\x80\xBD\xCD"
-     "\x00\x80"
-     "\xC1\xCD\x00\x80\xB8\x6D\x07\x00\xB9\x75\x07\x00\xBA\x7D\x07\x00\xBB\x75"
-     "\x07\x00"
-     "\xBC\x6D\x07\x00\xBD\xC5\x07\x00\xBE\xCD\x07\x00\xBF\xC5\x07\x00\xB0\xA5"
-     "\x06\x00"
-     "\xB1\xAD\x06\x00\xB2\xB9\x06\x00\xB3\xB1\x06\x00\xB4\x91\x06\x00\xB5\x91"
-     "\x06\x00"
-     "\xB6\x5D\x07\x00\xB7\x55\x07\x00\xB3\x25\x06\x00\xC5\xCD\x00\x80\xC9\xCD"
-     "\x00\x80"
-     "\xCD\xCD\x00\x80\xD1\xCD\x00\x80\xB6\x21\x06\x00\xB5\x35\x06\x00\xD5\xCD"
-     "\x00\x80"
-     "\xBB\x69\x06\x00\xBA\x61\x06\x00\xD9\xCD\x00\x80\xDD\xCD\x00\x80\xBF\x55"
-     "\x06\x00"
-     "\xBE\x55\x06\x00\xBD\x65\x06\x00\xBC\x6D\x06\x00\xE1\xCD\x00\x80\xA3\x61"
-     "\x06\x00"
-     "\xE5\xCD\x00\x80\xE9\xCD\x00\x80\xA6\x65\x06\x00\xED\xCD\x00\x80\xF1\xCD"
-     "\x00\x80"
-     "\xA5\x71\x06\x00\xAA\x25\x06\x00\xAB\x2D\x06\x00\xF5\xCD\x00\x80\xF9\xCD"
-     "\x00\x80"
-     "\xAE\x11\x06\x00\xAF\x11\x06\x00\xAC\x29\x06\x00\xAD\x21\x06\x00\xA8\xA9"
-     "\x06\x00"
-     "\xA9\xA9\x06\x00\xAA\xB9\x06\x00\xAB\xB1\x06\x00\xAC\xCD\x06\x00\xAD\x31"
-     "\x01\x00"
-     "\xAE\x31\x01\x00\xAF\x31\x01\x00\x80\xC9\x01\x00\x81\xC9\x01\x00\x82\x05"
-     "\x00\x00"
-     "\xFD\xCD\x00\x80\xBE\x78\x02\x00\x84\x78\x02\x00\x01\xCE\x00\x80\x05\xCE"
-     "\x00\x80"
-     "\xB8\xDD\x01\x00\xB9\xED\x01\x00\xBA\xE5\x01\x00\xBB\x8D\x01\x00\xBC\x99"
-     "\x01\x00"
-     "\xBD\x99\x01\x00\xBE\x8D\x01\x00\xBF\x85\x01\x00\xB0\x51\x01\x00\xB1\x51"
-     "\x01\x00"
-     "\xB2\x51\x01\x00\xB3\x51\x01\x00\xB4\xF5\x01\x00\xB5\xFD\x01\x00\xB6\xF5"
-     "\x01\x00"
-     "\xB7\xED\x01\x00\xB3\x31\x06\x00\x09\xCE\x00\x80\x86\x28\x00\x00\x87\xDC"
-     "\x01\x00"
-     "\x0D\xCE\x00\x80\xB6\xB1\x01\x00\xB5\x45\x06\x00\x11\xCE\x00\x80\xBB\x95"
-     "\x01\x00"
-     "\xBA\x95\x01\x00\x15\xCE\x00\x80\x19\xCE\x00\x80\xBF\x39\x01\x00\xBE\x31"
-     "\x01\x00"
-     "\xBD\x85\x01\x00\xBC\x85\x01\x00\x21\xCD\x00\x80\xA3\x75\x06\x00\x1D\xCE"
-     "\x00\x80"
-     "\x21\xCE\x00\x80\xA6\xF5\x01\x00\x25\xCE\x00\x80\x29\xCE\x00\x80\xA5\x01"
-     "\x06\x00"
-     "\xAA\xD1\x01\x00\xAB\xD1\x01\x00\x2D\xCE\x00\x80\x31\xCE\x00\x80\xAE\x75"
-     "\x01\x00"
-     "\xAF\x7D\x01\x00\xAC\xC1\x01\x00\xAD\xC1\x01\x00\x35\xCE\x00\x80\x39\xCE"
-     "\x00\x80"
-     "\x3D\xCE\x00\x80\x41\xCE\x00\x80\x45\xCE\x00\x80\x49\xCE\x00\x80\x4D\xCE"
-     "\x00\x80"
-     "\x51\xCE\x00\x80\x55\xCE\x00\x80\x59\xCE\x00\x80\x5D\xCE\x00\x80\x61\xCE"
-     "\x00\x80"
-     "\x65\xCE\x00\x80\x69\xCE\x00\x80\x6D\xCE\x00\x80\xBE\x70\x03\x00\x85\x00"
-     "\x3C\x00"
-     "\xE1\xC4\x06\x00\x84\x44\x02\x00\xE3\x8C\x07\x00\x80\x61\x00\x00\x81\x61"
-     "\x00\x00"
-     "\x82\x61\x00\x00\xEF\x68\x03\x00\x85\x44\x3C\x00\xE1\x10\x02\x00\x75\xCE"
-     "\x00\x80"
-     "\xE3\xD8\x2B\x00\x86\x50\x3D\x00\x87\xF4\x03\x00\x79\xCE\x00\x80\xEF\xA4"
-     "\x07\x00"
-     "\x7D\xCE\x00\x80\xEF\x40\x02\x00\x81\xCE\x00\x80\x85\xCE\x00\x80\x89\xCE"
-     "\x00\x80"
-     "\x8D\xCE\x00\x80\x91\xCE\x00\x80\x95\xCE\x00\x80\x84\x3C\x3C\x00\x99\xCE"
-     "\x00\x80"
-     "\x9D\xCE\x00\x80\xA1\xCE\x00\x80\xA5\xCE\x00\x80\xE3\xEC\x02\x00\xA9\xCE"
-     "\x00\x80"
-     "\xE1\x2C\x01\x00\xB3\x51\x03\x00\xAD\xCE\x00\x80\xB1\xCE\x00\x80\xB5\xCE"
-     "\x00\x80"
-     "\xB9\xCE\x00\x80\xB6\x79\x03\x00\xB5\x71\x03\x00\xBD\xCE\x00\x80\xBB\x39"
-     "\x03\x00"
-     "\xBA\x31\x03\x00\xC1\xCE\x00\x80\xC5\xCE\x00\x80\xBF\xF5\x00\x00\xBE\xF5"
-     "\x00\x00"
-     "\xBD\x15\x03\x00\xBC\x15\x03\x00\xA8\x3D\x02\x00\xA9\x81\x02\x00\xAA\x99"
-     "\x02\x00"
-     "\xAB\x91\x02\x00\xAC\xB9\x02\x00\xAD\xB9\x02\x00\xAE\xD1\x02\x00\xAF\xD1"
-     "\x02\x00"
-     "\x84\xA8\x3F\x00\xBE\xA8\x3F\x00\xC9\xCE\x00\x80\xCD\xCE\x00\x80\xD1\xCE"
-     "\x00\x80"
-     "\xD5\xCE\x00\x80\xD9\xCE\x00\x80\xDD\xCE\x00\x80\xB8\x51\x01\x00\xB9\x51"
-     "\x01\x00"
-     "\xBA\x51\x01\x00\xBB\x51\x01\x00\xBC\x71\x01\x00\xBD\x71\x01\x00\xBE\x71"
-     "\x01\x00"
-     "\xBF\x71\x01\x00\xB0\xB5\x02\x00\xB1\xBD\x02\x00\xB2\x81\x02\x00\xB3\x81"
-     "\x02\x00"
-     "\xB4\x71\x01\x00\xB5\x71\x01\x00\xB6\x71\x01\x00\xB7\x71\x01\x00\x80\xB5"
-     "\x00\x00"
-     "\x81\xBD\x00\x00\x82\xB5\x00\x00\xE5\xCE\x00\x80\x86\xB0\x3F\x00\x87\xE0"
-     "\x3C\x00"
-     "\xE9\xCE\x00\x80\xEF\x54\x00\x00\xBE\x2C\x3E\x00\xE1\x54\x06\x00\xED\xCE"
-     "\x00\x80"
-     "\xE3\x88\x00\x00\xF1\xCE\x00\x80\xF5\xCE\x00\x80\xF9\xCE\x00\x80\xFD\xCE"
-     "\x00\x80"
-     "\xA3\x51\x02\x00\x01\xCF\x00\x80\xBF\xD8\x26\x00\x05\xCF\x00\x80\x09\xCF"
-     "\x00\x80"
-     "\xA6\x79\x02\x00\xA5\x71\x02\x00\x0D\xCF\x00\x80\xAB\x39\x02\x00\xAA\x31"
-     "\x02\x00"
-     "\x11\xCF\x00\x80\x15\xCF\x00\x80\xAF\xF5\x01\x00\xAE\xF5\x01\x00\xAD\x15"
-     "\x02\x00"
-     "\xAC\x15\x02\x00\x90\x2D\x24\x00\x91\x05\x28\x00\x92\x0D\x28\x00\x93\xD9"
-     "\x28\x00"
-     "\x94\x85\x2D\x00\x95\x35\x2C\x00\x96\xC5\x2C\x00\x97\xB5\x31\x00\x98\x01"
-     "\x30\x00"
-     "\x99\x15\x30\x00\x9A\x95\x35\x00\x9B\x29\x34\x00\x9C\x6D\x34\x00\x9D\x99"
-     "\x35\x00"
-     "\x9E\x3D\x38\x00\x9F\x35\x38\x00\x71\xCE\x00\x80\xB6\xD5\x3E\x00\xB5\xC5"
-     "\x3E\x00"
-     "\xE1\xCE\x00\x80\xB3\xD1\x3E\x00\x19\xCF\x00\x80\x1D\xCF\x00\x80\x21\xCF"
-     "\x00\x80"
-     "\xBF\xD9\x3E\x00\xBE\xD5\x3E\x00\xBD\xC5\x3E\x00\xBC\xC5\x3E\x00\xBB\xD5"
-     "\x3E\x00"
-     "\xBA\xE9\x3E\x00\x25\xCF\x00\x80\x8F\x5D\x24\x00\xA9\xE5\x09\x00\xA8\x15"
-     "\x08\x00"
-     "\xAB\x05\x0C\x00\xAA\x0D\x0C\x00\xAD\x01\x10\x00\xAC\x01\x0C\x00\xAF\x41"
-     "\x10\x00"
-     "\xAE\xBD\x10\x00\xA1\xE1\x00\x00\x29\xCF\x00\x80\xA3\x01\x04\x00\xA2\xE1"
-     "\x00\x00"
-     "\xA5\x9D\x04\x00\xA4\x95\x04\x00\xA7\xB9\x08\x00\xA6\x01\x08\x00\xA0\xF5"
-     "\x39\x00"
-     "\xA1\x05\x3D\x00\xA2\xE9\x3C\x00\xA3\xF5\x3D\x00\x2D\xCF\x00\x80\x31\xCF"
-     "\x00\x80"
-     "\x35\xCF\x00\x80\x39\xCF\x00\x80\xB1\xC1\x14\x00\xB0\x01\x14\x00\xB3\x01"
-     "\x18\x00"
-     "\xB2\x7D\x14\x00\xB5\xD5\x18\x00\xB4\xD5\x18\x00\x3D\xCF\x00\x80\x41\xCF"
-     "\x00\x80"
-     "\x82\x21\x25\x00\x83\x29\x25\x00\x45\xCF\x00\x80\x49\xCF\x00\x80\x86\xC5"
-     "\x29\x00"
-     "\x87\x81\x2C\x00\x84\x19\x29\x00\x85\x19\x29\x00\x8A\x01\x2D\x00\x8B\xF9"
-     "\x2D\x00"
-     "\x4D\xCF\x00\x80\x55\xCF\x00\x80\x8E\x01\x31\x00\x8F\x89\x34\x00\x8C\x91"
-     "\x30\x00"
-     "\x8D\x1D\x31\x00\x92\x45\x35\x00\x93\x19\x35\x00\x86\xE8\x07\x00\x87\xEC"
-     "\x01\x00"
-     "\x96\x59\x39\x00\x97\x61\x38\x00\x94\xF5\x34\x00\x95\x59\x39\x00\x9A\xA1"
-     "\x3C\x00"
-     "\x9B\x45\x3D\x00\x59\xCF\x00\x80\x5D\xCF\x00\x80\x81\x7D\x00\x00\x80\x7D"
-     "\x00\x00"
-     "\x9C\x41\x3C\x00\x82\x55\x00\x00\xA8\xE9\x3F\x00\xA9\xFD\x3F\x00\xAA\x09"
-     "\x3F\x00"
-     "\xAB\x05\x3F\x00\xAC\x1D\x3F\x00\xAD\x05\x3F\x00\xAE\x0D\x3F\x00\xAF\x05"
-     "\x3F\x00"
-     "\x61\xCF\x00\x80\x65\xCF\x00\x80\x69\xCF\x00\x80\x6D\xCF\x00\x80\x71\xCF"
-     "\x00\x80"
-     "\x75\xCF\x00\x80\x79\xCF\x00\x80\x7D\xCF\x00\x80\xB8\x0D\x3F\x00\xB9\x15"
-     "\x3F\x00"
-     "\xBA\x15\x3F\x00\xBB\x25\x3F\x00\xBC\x3D\x3F\x00\xBD\xFD\x3E\x00\xBE\xF5"
-     "\x3E\x00"
-     "\xBF\xED\x3E\x00\xB0\x7D\x3F\x00\xB1\x41\x3F\x00\xB2\x41\x3F\x00\xB3\x41"
-     "\x3F\x00"
-     "\xB4\x41\x3F\x00\xB5\x4D\x3F\x00\xB6\x3D\x3F\x00\xB7\x35\x3F\x00\xA3\x81"
-     "\x3C\x00"
-     "\x81\xCF\x00\x80\x85\xCF\x00\x80\x89\xCF\x00\x80\x8D\xCF\x00\x80\xA6\x85"
-     "\x3C\x00"
-     "\xA5\x95\x3C\x00\x91\xCF\x00\x80\xAB\x85\x3C\x00\xAA\xB9\x3C\x00\x95\xCF"
-     "\x00\x80"
-     "\x99\xCF\x00\x80\xAF\x89\x3C\x00\xAE\x85\x3C\x00\xAD\x95\x3C\x00\xAC\x95"
-     "\x3C\x00"
-     "\x84\xC8\x03\x00\xB3\xED\x3D\x00\x9D\xCF\x00\x80\xA1\xCF\x00\x80\xB6\xE9"
-     "\x3D\x00"
-     "\xA5\xCF\x00\x80\xA9\xCF\x00\x80\xB5\xE9\x3D\x00\xBA\xAD\x3D\x00\xBB\xB5"
-     "\x3D\x00"
-     "\xAD\xCF\x00\x80\xB1\xCF\x00\x80\xBE\x9D\x3D\x00\xBF\x21\x02\x00\xBC\xA5"
-     "\x3D\x00"
-     "\xBD\x95\x3D\x00\xA8\x0D\x3D\x00\xA9\x1D\x3D\x00\xAA\x21\x3D\x00\xAB\x3D"
-     "\x3D\x00"
-     "\xAC\x25\x3D\x00\xAD\x2D\x3D\x00\xAE\x25\x3D\x00\xAF\x59\x3D\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x1D\x00\x00\xB5\xCF\x00\x80\xB9\xCF\x00\x80\xBD\xCF"
-     "\x00\x80"
-     "\xC5\xCF\x00\x80\xBE\xB8\x03\x00\xB8\xB9\x02\x00\xB9\x61\x02\x00\xBA\x19"
-     "\x02\x00"
-     "\xBB\x19\x02\x00\xBC\x09\x02\x00\xBD\x09\x02\x00\xBE\x21\x02\x00\xBF\x21"
-     "\x02\x00"
-     "\xB0\x2D\x3D\x00\xB1\x35\x3D\x00\xB2\x35\x3D\x00\xB3\x05\x3D\x00\xB4\x1D"
-     "\x3D\x00"
-     "\xB5\xA1\x02\x00\xB6\xA1\x02\x00\xB7\xA1\x02\x00\xA3\xA9\x3C\x00\xC9\xCF"
-     "\x00\x80"
-     "\x86\x28\x05\x00\x87\xEC\x02\x00\xCD\xCF\x00\x80\xA6\xAD\x3C\x00\xA5\xAD"
-     "\x3C\x00"
-     "\xD1\xCF\x00\x80\xAB\xF1\x3C\x00\xAA\xE9\x3C\x00\xD5\xCF\x00\x80\xD9\xCF"
-     "\x00\x80"
-     "\xAF\x65\x03\x00\xAE\xD9\x3C\x00\xAD\xD1\x3C\x00\xAC\xE1\x3C\x00\xDD\xCF"
-     "\x00\x80"
-     "\xB3\x29\x02\x00\xE1\xCF\x00\x80\xE5\xCF\x00\x80\xB6\xF9\x02\x00\xE9\xCF"
-     "\x00\x80"
-     "\xED\xCF\x00\x80\xB5\xF9\x02\x00\xBA\xD5\x02\x00\xBB\xDD\x02\x00\xF1\xCF"
-     "\x00\x80"
-     "\xF5\xCF\x00\x80\xBE\x79\x01\x00\xBF\x79\x01\x00\xBC\xC5\x02\x00\xBD\x79"
-     "\x01\x00"
-     "\xF9\xCF\x00\x80\xFD\xCF\x00\x80\xA3\xE5\x02\x00\x01\xD0\x00\x80\xA5\x35"
-     "\x02\x00"
-     "\x05\xD0\x00\x80\x09\xD0\x00\x80\xA6\x35\x02\x00\x0D\xD0\x00\x80\x11\xD0"
-     "\x00\x80"
-     "\xAB\x11\x02\x00\xAA\x19\x02\x00\xAD\xB5\x01\x00\xAC\x09\x02\x00\xAF\xB5"
-     "\x01\x00"
-     "\xAE\xB5\x01\x00\xE3\xF0\x3E\x00\xE1\xAC\x3F\x00\xE1\x40\x3E\x00\xE3\x7C"
-     "\x3F\x00"
-     "\x15\xD0\x00\x80\x19\xD0\x00\x80\x1D\xD0\x00\x80\x21\xD0\x00\x80\x80\x0D"
-     "\x00\x00"
-     "\x81\x11\x00\x00\x82\x11\x00\x00\x25\xD0\x00\x80\xEF\xA8\x3E\x00\x29\xD0"
-     "\x00\x80"
-     "\x2D\xD0\x00\x80\xEF\x20\x3E\x00\xA8\x2D\x05\x00\xA9\x6D\x05\x00\xAA\x65"
-     "\x05\x00"
-     "\xAB\xAD\x05\x00\xAC\xB5\x05\x00\xAD\xBD\x05\x00\xAE\xB5\x05\x00\xAF\xAD"
-     "\x05\x00"
-     "\xC1\xCF\x00\x80\x84\xE8\x03\x00\xBE\xE0\x03\x00\x31\xD0\x00\x80\x86\x10"
-     "\x03\x00"
-     "\x87\x98\x03\x00\x35\xD0\x00\x80\x39\xD0\x00\x80\xB8\x69\x06\x00\xB9\x69"
-     "\x06\x00"
-     "\xBA\x01\x06\x00\xBB\x01\x06\x00\xBC\x05\x06\x00\xBD\x0D\x06\x00\xBE\x31"
-     "\x06\x00"
-     "\xBF\x31\x06\x00\xB0\xD5\x05\x00\xB1\xDD\x05\x00\xB2\xD5\x05\x00\xB3\x69"
-     "\x06\x00"
-     "\xB4\x79\x06\x00\xB5\x79\x06\x00\xB6\x69\x06\x00\xB7\x61\x06\x00\xA8\x39"
-     "\x06\x00"
-     "\xA9\x81\x07\x00\xAA\x99\x07\x00\xAB\x91\x07\x00\xAC\xB9\x07\x00\xAD\xB9"
-     "\x07\x00"
-     "\xAE\xD9\x07\x00\xAF\xD5\x07\x00\x3D\xD0\x00\x80\x41\xD0\x00\x80\x51\xCF"
-     "\x00\x80"
-     "\x45\xD0\x00\x80\x49\xD0\x00\x80\x4D\xD0\x00\x80\x51\xD0\x00\x80\x55\xD0"
-     "\x00\x80"
-     "\xB8\x55\x07\x00\xB9\x5D\x07\x00\xBA\x69\x07\x00\xBB\x69\x07\x00\xBC\x01"
-     "\x07\x00"
-     "\xBD\x01\x07\x00\xBE\x01\x07\x00\xBF\x01\x07\x00\xB0\xAD\x07\x00\xB1\xB1"
-     "\x07\x00"
-     "\xB2\xB1\x07\x00\xB3\x85\x07\x00\xB4\x9D\x07\x00\xB5\x75\x07\x00\xB6\x7D"
-     "\x07\x00"
-     "\xB7\x71\x07\x00\xB3\x11\x06\x00\x59\xD0\x00\x80\x5D\xD0\x00\x80\x61\xD0"
-     "\x00\x80"
-     "\x65\xD0\x00\x80\xB6\x39\x06\x00\xB5\x31\x06\x00\x69\xD0\x00\x80\xBB\x75"
-     "\x06\x00"
-     "\xBA\x69\x06\x00\x6D\xD0\x00\x80\x71\xD0\x00\x80\xBF\xB1\x06\x00\xBE\x59"
-     "\x06\x00"
-     "\xBD\x51\x06\x00\xBC\x65\x06\x00\x75\xD0\x00\x80\xA3\x55\x06\x00\x79\xD0"
-     "\x00\x80"
-     "\x7D\xD0\x00\x80\xA6\x7D\x06\x00\x81\xD0\x00\x80\x85\xD0\x00\x80\xA5\x75"
-     "\x06\x00"
-     "\xAA\x2D\x06\x00\xAB\x31\x06\x00\x89\xD0\x00\x80\x8D\xD0\x00\x80\xAE\x1D"
-     "\x06\x00"
-     "\xAF\xF5\x06\x00\xAC\x21\x06\x00\xAD\x15\x06\x00\xA8\xB9\x01\x00\xA9\xB9"
-     "\x01\x00"
-     "\xAA\x29\x01\x00\xAB\x29\x01\x00\xAC\x3D\x01\x00\xAD\x25\x01\x00\xAE\x2D"
-     "\x01\x00"
-     "\xAF\x25\x01\x00\x91\xD0\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\x95\xD0\x00\x80\x99\xD0\x00\x80\x9D\xD0\x00\x80\xBE\x70\x01\x00\xB8\x81"
-     "\x00\x00"
-     "\xB9\x8D\x00\x00\xBA\x85\x00\x00\xBB\x99\x00\x00\xBC\x89\x00\x00\xBD\xBD"
-     "\x00\x00"
-     "\xBE\xB5\x00\x00\xBF\x7D\x00\x00\xB0\x5D\x01\x00\xB1\xE1\x00\x00\xB2\xE1"
-     "\x00\x00"
-     "\xB3\xE1\x00\x00\xB4\xE1\x00\x00\xB5\xE9\x00\x00\xB6\xD1\x00\x00\xB7\xD1"
-     "\x00\x00"
-     "\x84\xC8\x02\x00\xB3\xA5\x02\x00\x87\x38\x03\x00\x86\x28\x02\x00\xB6\xA1"
-     "\x02\x00"
-     "\xA5\xD0\x00\x80\xA9\xD0\x00\x80\xB5\xB1\x02\x00\xBA\x9D\x02\x00\xBB\x21"
-     "\x03\x00"
-     "\xBE\x6C\x03\x00\xAD\xD0\x00\x80\xBE\x21\x03\x00\xBF\x29\x03\x00\xBC\x31"
-     "\x03\x00"
-     "\xBD\x31\x03\x00\xA3\xE1\x02\x00\xB1\xD0\x00\x80\xB5\xD0\x00\x80\xB9\xD0"
-     "\x00\x80"
-     "\xBD\xD0\x00\x80\xA6\xE5\x02\x00\xA5\xF5\x02\x00\xC1\xD0\x00\x80\xAB\x65"
-     "\x03\x00"
-     "\xAA\xD9\x02\x00\xC5\xD0\x00\x80\xC9\xD0\x00\x80\xAF\x6D\x03\x00\xAE\x65"
-     "\x03\x00"
-     "\xAD\x75\x03\x00\xAC\x75\x03\x00\xA9\x99\x00\x00\xA8\x91\x00\x00\xAB\xCD"
-     "\x00\x00"
-     "\xAA\xA1\x00\x00\xAD\xDD\x00\x00\xAC\xDD\x00\x00\xAF\xC5\x00\x00\xAE\xCD"
-     "\x00\x00"
-     "\xBE\x2C\x0D\x00\xCD\xD0\x00\x80\xD1\xD0\x00\x80\xD5\xD0\x00\x80\xD9\xD0"
-     "\x00\x80"
-     "\xDD\xD0\x00\x80\xE1\xD0\x00\x80\xE5\xD0\x00\x80\xB9\xC1\x01\x00\xB8\x79"
-     "\x00\x00"
-     "\xBB\xC1\x01\x00\xBA\xC9\x01\x00\xBD\xC1\x01\x00\xBC\xD9\x01\x00\xBF\xC5"
-     "\x01\x00"
-     "\xBE\xC5\x01\x00\xB1\x8D\x00\x00\xB0\x8D\x00\x00\xB3\x41\x00\x00\xB2\x49"
-     "\x00\x00"
-     "\xB5\x41\x00\x00\xB4\x59\x00\x00\xB7\x41\x00\x00\xB6\x49\x00\x00\xE9\xD0"
-     "\x00\x80"
-     "\xED\xD0\x00\x80\xF1\xD0\x00\x80\xF5\xD0\x00\x80\xF9\xD0\x00\x80\xEF\x50"
-     "\x07\x00"
-     "\xFD\xD0\x00\x80\x01\xD1\x00\x80\xBE\xF8\x0F\x00\xE3\x74\x07\x00\x05\xD1"
-     "\x00\x80"
-     "\xE1\x7C\x04\x00\x80\x19\x00\x00\x81\x09\x00\x00\x82\x79\x00\x00\x09\xD1"
-     "\x00\x80"
-     "\x0D\xD1\x00\x80\xB3\x69\x01\x00\x15\xD1\x00\x80\x84\xC4\x02\x00\xB6\x1D"
-     "\x01\x00"
-     "\x19\xD1\x00\x80\x1D\xD1\x00\x80\xB5\x15\x01\x00\xBA\x09\x01\x00\xBB\x09"
-     "\x01\x00"
-     "\x86\xE8\x0D\x00\x87\xE8\x0D\x00\xBE\xDD\x01\x00\xBF\xC5\x01\x00\xBC\xDD"
-     "\x01\x00"
-     "\xBD\xD5\x01\x00\x21\xD1\x00\x80\x25\xD1\x00\x80\x29\xD1\x00\x80\x2D\xD1"
-     "\x00\x80"
-     "\xEF\xD4\x00\x00\x31\xD1\x00\x80\x35\xD1\x00\x80\xEF\x4C\x01\x00\xE3\xB0"
-     "\x03\x00"
-     "\xE1\xB4\x06\x00\xE1\x80\x01\x00\xE3\x98\x01\x00\x39\xD1\x00\x80\x3D\xD1"
-     "\x00\x80"
-     "\x41\xD1\x00\x80\x45\xD1\x00\x80\x49\xD1\x00\x80\x4D\xD1\x00\x80\xA3\xE5"
-     "\x01\x00"
-     "\x84\xC0\x0D\x00\xA5\x99\x01\x00\x51\xD1\x00\x80\x55\xD1\x00\x80\xA6\x91"
-     "\x01\x00"
-     "\x59\xD1\x00\x80\x5D\xD1\x00\x80\xAB\x85\x01\x00\xAA\x85\x01\x00\xAD\x59"
-     "\x01\x00"
-     "\xAC\x51\x01\x00\xAF\x49\x01\x00\xAE\x51\x01\x00\x11\xD1\x00\x80\x61\xD1"
-     "\x00\x80"
-     "\x65\xD1\x00\x80\x69\xD1\x00\x80\x6D\xD1\x00\x80\x71\xD1\x00\x80\x75\xD1"
-     "\x00\x80"
-     "\x79\xD1\x00\x80\xA8\x69\x0F\x00\xA9\x71\x0F\x00\xAA\x71\x0F\x00\xAB\xAD"
-     "\x0F\x00"
-     "\xAC\xB5\x0F\x00\xAD\xBD\x0F\x00\xAE\xB5\x0F\x00\xAF\xA9\x0F\x00\xB0\xD9"
-     "\x0F\x00"
-     "\xB1\xF5\x0F\x00\xB2\xFD\x0F\x00\xB3\xF5\x0F\x00\xB4\xED\x0F\x00\xB5\x95"
-     "\x0F\x00"
-     "\xB6\x9D\x0F\x00\xB7\x89\x0F\x00\xB8\xB9\x0F\x00\xB9\x85\x0F\x00\xBA\x8D"
-     "\x0F\x00"
-     "\xBB\x69\x00\x00\xBC\x79\x00\x00\xBD\x79\x00\x00\xBE\x69\x00\x00\xBF\x69"
-     "\x00\x00"
-     "\x81\x9D\x00\x00\x80\x9D\x00\x00\x7D\xD1\x00\x80\x82\x05\x00\x00\x81\xD1"
-     "\x00\x80"
-     "\x85\xD1\x00\x80\x89\xD1\x00\x80\x8D\xD1\x00\x80\x86\x80\x03\x00\x87\xF4"
-     "\x03\x00"
-     "\x91\xD1\x00\x80\x95\xD1\x00\x80\x99\xD1\x00\x80\x9D\xD1\x00\x80\xA1\xD1"
-     "\x00\x80"
-     "\xA1\xD0\x00\x80\xB3\x99\x0F\x00\xA5\xD1\x00\x80\xA9\xD1\x00\x80\xAD\xD1"
-     "\x00\x80"
-     "\xB1\xD1\x00\x80\xB6\x5D\x0F\x00\xB5\x5D\x0F\x00\xB5\xD1\x00\x80\xBB\x51"
-     "\x0F\x00"
-     "\xBA\x7D\x0F\x00\xB9\xD1\x00\x80\xBD\xD1\x00\x80\xBF\x31\x0F\x00\xBE\x49"
-     "\x0F\x00"
-     "\xBD\x41\x0F\x00\xBC\x49\x0F\x00\xA3\x59\x0E\x00\xC1\xD1\x00\x80\xC5\xD1"
-     "\x00\x80"
-     "\xC9\xD1\x00\x80\xCD\xD1\x00\x80\xA6\x9D\x0E\x00\xA5\x9D\x0E\x00\xD1\xD1"
-     "\x00\x80"
-     "\xAB\x91\x0E\x00\xAA\xBD\x0E\x00\xD5\xD1\x00\x80\xD9\xD1\x00\x80\xAF\xF1"
-     "\x0E\x00"
-     "\xAE\x89\x0E\x00\xAD\x81\x0E\x00\xAC\x89\x0E\x00\xDD\xD1\x00\x80\xE1\xD1"
-     "\x00\x80"
-     "\xE5\xD1\x00\x80\xE9\xD1\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xED\xD1\x00\x80\x84\xA0\x01\x00\xF1\xD1\x00\x80\x87\xE8\x01\x00\x86\x00"
-     "\x04\x00"
-     "\xF5\xD1\x00\x80\xF9\xD1\x00\x80\xFD\xD1\x00\x80\x01\xD2\x00\x80\xA8\x15"
-     "\x01\x00"
-     "\xA9\x1D\x01\x00\xAA\x15\x01\x00\xAB\x25\x01\x00\xAC\x3D\x01\x00\xAD\x25"
-     "\x01\x00"
-     "\xAE\x2D\x01\x00\xAF\x25\x01\x00\x05\xD2\x00\x80\x09\xD2\x00\x80\x0D\xD2"
-     "\x00\x80"
-     "\x11\xD2\x00\x80\x15\xD2\x00\x80\x19\xD2\x00\x80\x1D\xD2\x00\x80\x21\xD2"
-     "\x00\x80"
-     "\xB8\xC9\x00\x00\xB9\xC9\x00\x00\xBA\xD9\x00\x00\xBB\xD1\x00\x00\xBC\xF9"
-     "\x00\x00"
-     "\xBD\xF9\x00\x00\xBE\x99\x00\x00\xBF\x99\x00\x00\xB0\x25\x01\x00\xB1\x2D"
-     "\x01\x00"
-     "\xB2\x25\x01\x00\xB3\x3D\x01\x00\xB4\x2D\x01\x00\xB5\x1D\x01\x00\xB6\x15"
-     "\x01\x00"
-     "\xB7\xF9\x00\x00\x25\xD2\x00\x80\x29\xD2\x00\x80\x2D\xD2\x00\x80\xB3\x91"
-     "\x02\x00"
-     "\x31\xD2\x00\x80\xB5\xB9\x02\x00\xB6\xA9\x02\x00\x35\xD2\x00\x80\x39\xD2"
-     "\x00\x80"
-     "\x3D\xD2\x00\x80\xBA\xED\x02\x00\xBB\xE5\x02\x00\xBC\xFD\x02\x00\xBD\xE5"
-     "\x02\x00"
-     "\xBE\xE5\x02\x00\xBF\xD5\x02\x00\xA6\xF1\x02\x00\x41\xD2\x00\x80\x45\xD2"
-     "\x00\x80"
-     "\xA5\xE1\x02\x00\x49\xD2\x00\x80\xA3\xC9\x02\x00\x4D\xD2\x00\x80\x51\xD2"
-     "\x00\x80"
-     "\xAE\xBD\x02\x00\xAF\x8D\x02\x00\xAC\xA5\x02\x00\xAD\xBD\x02\x00\xAA\xB5"
-     "\x02\x00"
-     "\xAB\xBD\x02\x00\x55\xD2\x00\x80\x59\xD2\x00\x80\x80\xF9\x00\x00\x81\xF9"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x5D\xD2\x00\x80\xBE\xC8\x03\x00\x84\x18\x03\x00\x65\xD2"
-     "\x00\x80"
-     "\x69\xD2\x00\x80\x6D\xD2\x00\x80\x71\xD2\x00\x80\x75\xD2\x00\x80\x79\xD2"
-     "\x00\x80"
-     "\x7D\xD2\x00\x80\x81\xD2\x00\x80\x86\x18\x04\x00\x87\x9C\x03\x00\x85\xD2"
-     "\x00\x80"
-     "\x89\xD2\x00\x80\x8D\xD2\x00\x80\x91\xD2\x00\x80\x95\xD2\x00\x80\x99\xD2"
-     "\x00\x80"
-     "\xEF\xB0\x02\x00\x9D\xD2\x00\x80\xE1\x94\x01\x00\xA1\xD2\x00\x80\xE3\x6C"
-     "\x02\x00"
-     "\xA5\xD2\x00\x80\xA9\xD2\x00\x80\xAD\xD2\x00\x80\xB1\xD2\x00\x80\xB5\xD2"
-     "\x00\x80"
-     "\xEF\xD0\x0F\x00\xB9\xD2\x00\x80\xBD\xD2\x00\x80\xC1\xD2\x00\x80\xC5\xD2"
-     "\x00\x80"
-     "\xE1\xAC\x01\x00\xC9\xD2\x00\x80\xE3\x6C\x00\x00\x80\x31\x00\x00\x81\x3D"
-     "\x00\x00"
-     "\x82\x1D\x00\x00\xEF\xF4\x0E\x00\xE3\x6C\x0E\x00\xCD\xD2\x00\x80\xE1\x2C"
-     "\x0F\x00"
-     "\xBE\x70\x05\x00\xB3\x39\x02\x00\x84\x0C\x05\x00\x86\xE8\x04\x00\x87\x60"
-     "\x05\x00"
-     "\xDC\x00\x00\x00\xB6\xF1\x02\x00\xB5\xF9\x02\x00\xD5\xD2\x00\x80\xBB\xD5"
-     "\x02\x00"
-     "\xBA\xD5\x02\x00\xD9\xD2\x00\x80\xDD\xD2\x00\x80\xBF\x75\x01\x00\xBE\x75"
-     "\x01\x00"
-     "\xBD\xC5\x02\x00\xBC\xC5\x02\x00\xE1\xD2\x00\x80\xE1\xF4\x0E\x00\xE5\xD2"
-     "\x00\x80"
-     "\xE3\x50\x0E\x00\xE9\xD2\x00\x80\xED\xD2\x00\x80\xF1\xD2\x00\x80\xF5\xD2"
-     "\x00\x80"
-     "\xF9\xD2\x00\x80\xFD\xD2\x00\x80\x01\xD3\x00\x80\x05\xD3\x00\x80\x09\xD3"
-     "\x00\x80"
-     "\x0D\xD3\x00\x80\x11\xD3\x00\x80\xEF\xE4\x0F\x00\xA6\x65\x02\x00\x15\xD3"
-     "\x00\x80"
-     "\x19\xD3\x00\x80\xA5\x6D\x02\x00\x1D\xD3\x00\x80\xA3\xAD\x02\x00\x21\xD3"
-     "\x00\x80"
-     "\x25\xD3\x00\x80\xAE\xE1\x01\x00\xAF\xE1\x01\x00\xAC\x51\x02\x00\xAD\x51"
-     "\x02\x00"
-     "\xAA\x41\x02\x00\xAB\x41\x02\x00\x29\xD3\x00\x80\x2D\xD3\x00\x80\xA8\x99"
-     "\x06\x00"
-     "\xA9\x99\x06\x00\xAA\xA9\x06\x00\xAB\xA9\x06\x00\xAC\xB9\x06\x00\xAD\xB9"
-     "\x06\x00"
-     "\xAE\xA9\x06\x00\xAF\xA9\x06\x00\x31\xD3\x00\x80\x82\x1D\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x35\xD3\x00\x80\x39\xD3\x00\x80\x3D\xD3\x00\x80\xBE\x74"
-     "\x03\x00"
-     "\xB8\xAD\x06\x00\xB9\xB5\x06\x00\xBA\xBD\x06\x00\xBB\xB5\x06\x00\xBC\xAD"
-     "\x06\x00"
-     "\xBD\x51\x07\x00\xBE\x51\x07\x00\xBF\x51\x07\x00\xB0\xA1\x06\x00\xB1\xA1"
-     "\x06\x00"
-     "\xB2\xA1\x06\x00\xB3\xA1\x06\x00\xB4\xA1\x06\x00\xB5\xA1\x06\x00\xB6\xA5"
-     "\x06\x00"
-     "\xB7\x99\x06\x00\x61\xD2\x00\x80\xB3\x25\x06\x00\x84\xC4\x03\x00\xD1\xD2"
-     "\x00\x80"
-     "\xB6\x3D\x06\x00\x41\xD3\x00\x80\x45\xD3\x00\x80\xB5\x35\x06\x00\xBA\x61"
-     "\x06\x00"
-     "\xBB\x61\x06\x00\x86\x08\x00\x00\x87\x88\x00\x00\xBE\x61\x06\x00\xBF\x61"
-     "\x06\x00"
-     "\xBC\x71\x06\x00\xBD\x71\x06\x00\xA3\x61\x06\x00\x49\xD3\x00\x80\x4D\xD3"
-     "\x00\x80"
-     "\x51\xD3\x00\x80\x55\xD3\x00\x80\xA6\x79\x06\x00\xA5\x71\x06\x00\x59\xD3"
-     "\x00\x80"
-     "\xAB\x25\x06\x00\xAA\x25\x06\x00\x5D\xD3\x00\x80\x61\xD3\x00\x80\xAF\x25"
-     "\x06\x00"
-     "\xAE\x25\x06\x00\xAD\x35\x06\x00\xAC\x35\x06\x00\xA8\x5D\x06\x00\xA9\x6D"
-     "\x06\x00"
-     "\xAA\x65\x06\x00\xAB\x8D\x06\x00\xAC\x99\x06\x00\xAD\x85\x06\x00\xAE\x8D"
-     "\x06\x00"
-     "\xAF\x85\x06\x00\x65\xD3\x00\x80\x69\xD3\x00\x80\x6D\xD3\x00\x80\x71\xD3"
-     "\x00\x80"
-     "\x75\xD3\x00\x80\x79\xD3\x00\x80\x7D\xD3\x00\x80\x81\xD3\x00\x80\xB8\x85"
-     "\x06\x00"
-     "\xB9\x8D\x06\x00\xBA\x99\x06\x00\xBB\x95\x06\x00\xBC\x8D\x06\x00\xBD\xAD"
-     "\x06\x00"
-     "\xBE\xA5\x06\x00\xBF\x7D\x01\x00\xB0\xFD\x06\x00\xB1\xC5\x06\x00\xB2\xCD"
-     "\x06\x00"
-     "\xB3\xC5\x06\x00\xB4\xDD\x06\x00\xB5\xC5\x06\x00\xB6\xCD\x06\x00\xB7\xC5"
-     "\x06\x00"
-     "\xB3\xED\x06\x00\x85\xD3\x00\x80\x89\xD3\x00\x80\x8D\xD3\x00\x80\x91\xD3"
-     "\x00\x80"
-     "\xB6\x05\x06\x00\xB5\x11\x06\x00\x95\xD3\x00\x80\xBB\x01\x06\x00\xBA\x39"
-     "\x06\x00"
-     "\x99\xD3\x00\x80\x9D\xD3\x00\x80\xBF\x01\x06\x00\xBE\x19\x06\x00\xBD\x11"
-     "\x06\x00"
-     "\xBC\x19\x06\x00\xA1\xD3\x00\x80\xA3\xA9\x06\x00\xA5\xD3\x00\x80\xA9\xD3"
-     "\x00\x80"
-     "\xA6\x41\x06\x00\xAD\xD3\x00\x80\x84\x94\x01\x00\xA5\x55\x06\x00\xAA\x7D"
-     "\x06\x00"
-     "\xAB\x45\x06\x00\xBE\xA0\x01\x00\xB5\xD3\x00\x80\xAE\x5D\x06\x00\xAF\x45"
-     "\x06\x00"
-     "\xAC\x5D\x06\x00\xAD\x55\x06\x00\xA8\x99\x02\x00\xA9\xC1\x02\x00\xAA\xC1"
-     "\x02\x00"
-     "\xAB\xC1\x02\x00\xAC\xC1\x02\x00\xAD\xC9\x02\x00\xAE\xF1\x02\x00\xAF\xF1"
-     "\x02\x00"
-     "\x81\xED\x03\x00\x80\xED\x03\x00\xB9\xD3\x00\x80\x82\xF9\x03\x00\x86\x90"
-     "\x1C\x00"
-     "\x87\x74\x03\x00\xBD\xD3\x00\x80\xC1\xD3\x00\x80\xB8\xC5\x03\x00\xB9\xCD"
-     "\x03\x00"
-     "\xBA\xC5\x03\x00\xBB\xDD\x03\x00\xBC\xCD\x03\x00\xBD\xFD\x03\x00\xBE\xF5"
-     "\x03\x00"
-     "\xBF\x9D\x03\x00\xB0\x41\x03\x00\xB1\x41\x03\x00\xB2\x41\x03\x00\xB3\x41"
-     "\x03\x00"
-     "\xB4\x41\x03\x00\xB5\x41\x03\x00\xB6\x41\x03\x00\xB7\x41\x03\x00\xB3\x49"
-     "\x02\x00"
-     "\xC5\xD3\x00\x80\xC9\xD3\x00\x80\xCD\xD3\x00\x80\xD1\xD3\x00\x80\xB6\x49"
-     "\x02\x00"
-     "\xB5\x49\x02\x00\xD5\xD3\x00\x80\xBB\x85\x03\x00\xBA\x85\x03\x00\xD9\xD3"
-     "\x00\x80"
-     "\xDD\xD3\x00\x80\xBF\x85\x03\x00\xBE\x85\x03\x00\xBD\x95\x03\x00\xBC\x95"
-     "\x03\x00"
-     "\xE1\xD3\x00\x80\xA3\x0D\x02\x00\xE5\xD3\x00\x80\xE9\xD3\x00\x80\xA6\x0D"
-     "\x02\x00"
-     "\xED\xD3\x00\x80\xF1\xD3\x00\x80\xA5\x0D\x02\x00\xAA\xC1\x03\x00\xAB\xC1"
-     "\x03\x00"
-     "\xF5\xD3\x00\x80\xF9\xD3\x00\x80\xAE\xC1\x03\x00\xAF\xC1\x03\x00\xAC\xD1"
-     "\x03\x00"
-     "\xAD\xD1\x03\x00\xE3\x98\x01\x00\xE1\xA4\x07\x00\xE1\x58\x06\x00\xE3\x58"
-     "\x06\x00"
-     "\xE1\xA0\x01\x00\xFD\xD3\x00\x80\xE3\xD0\x00\x00\x01\xD4\x00\x80\x05\xD4"
-     "\x00\x80"
-     "\x09\xD4\x00\x80\xEF\x38\x00\x00\x0D\xD4\x00\x80\xEF\xF4\x01\x00\x11\xD4"
-     "\x00\x80"
-     "\x15\xD4\x00\x80\xEF\xF8\x06\x00\x80\x79\x00\x00\x81\x15\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\x84\x00\x1D\x00\x19\xD4\x00\x80\x1D\xD4\x00\x80\xBE\x10\x1D\x00\x25\xD4"
-     "\x00\x80"
-     "\x86\xC0\x1C\x00\x87\xAC\x1D\x00\x29\xD4\x00\x80\x2D\xD4\x00\x80\x31\xD4"
-     "\x00\x80"
-     "\x35\xD4\x00\x80\x39\xD4\x00\x80\x3D\xD4\x00\x80\xBB\xC5\x05\x00\xBA\xA1"
-     "\x05\x00"
-     "\xB9\xA9\x05\x00\xB8\x91\x05\x00\xBF\xCD\x05\x00\xBE\xCD\x05\x00\xBD\xC9"
-     "\x05\x00"
-     "\xBC\xD5\x05\x00\xB3\x1D\x06\x00\xB2\x1D\x06\x00\xB1\x1D\x06\x00\xB0\x1D"
-     "\x06\x00"
-     "\xB7\xA1\x05\x00\xB6\xBD\x05\x00\xB5\xBD\x05\x00\xB4\xBD\x05\x00\xAB\xBD"
-     "\x06\x00"
-     "\xAA\xBD\x06\x00\xA9\xBD\x06\x00\xA8\xBD\x06\x00\xAF\x7D\x06\x00\xAE\x7D"
-     "\x06\x00"
-     "\xAD\x7D\x06\x00\xAC\x7D\x06\x00\x41\xD4\x00\x80\x45\xD4\x00\x80\x49\xD4"
-     "\x00\x80"
-     "\x4D\xD4\x00\x80\x51\xD4\x00\x80\x55\xD4\x00\x80\x59\xD4\x00\x80\x5D\xD4"
-     "\x00\x80"
-     "\xA8\xED\x1D\x00\xA9\x2D\x1E\x00\xAA\x31\x1E\x00\xAB\x31\x1E\x00\xAC\x95"
-     "\x1E\x00"
-     "\xAD\x9D\x1E\x00\xAE\x95\x1E\x00\xAF\x8D\x1E\x00\x21\xD4\x00\x80\x61\xD4"
-     "\x00\x80"
-     "\x65\xD4\x00\x80\x69\xD4\x00\x80\x6D\xD4\x00\x80\x71\xD4\x00\x80\x75\xD4"
-     "\x00\x80"
-     "\x79\xD4\x00\x80\xB8\xA9\x1E\x00\xB9\xA9\x1E\x00\xBA\x5D\x1F\x00\xBB\x51"
-     "\x1F\x00"
-     "\xBC\x71\x1F\x00\xBD\x71\x1F\x00\xBE\x75\x1F\x00\xBF\x69\x1F\x00\xB0\xFD"
-     "\x1E\x00"
-     "\xB1\xCD\x1E\x00\xB2\xC5\x1E\x00\xB3\xAD\x1E\x00\xB4\xB9\x1E\x00\xB5\xB9"
-     "\x1E\x00"
-     "\xB6\xAD\x1E\x00\xB7\xA5\x1E\x00\xB3\xB9\x1E\x00\x7D\xD4\x00\x80\x81\xD4"
-     "\x00\x80"
-     "\x85\xD4\x00\x80\xB1\xD3\x00\x80\xB6\x1D\x1E\x00\xB5\x1D\x1E\x00\x89\xD4"
-     "\x00\x80"
-     "\xBB\x09\x1E\x00\xBA\x39\x1E\x00\x8D\xD4\x00\x80\x84\xE0\x03\x00\xBF\x7D"
-     "\x1E\x00"
-     "\xBE\x7D\x1E\x00\xBD\x79\x1E\x00\xBC\x11\x1E\x00\x82\x69\x00\x00\xA3\xFD"
-     "\x1E\x00"
-     "\x80\x45\x00\x00\x81\x51\x00\x00\xA6\x59\x1E\x00\xBE\x9C\x03\x00\x91\xD4"
-     "\x00\x80"
-     "\xA5\x59\x1E\x00\xAA\x7D\x1E\x00\xAB\x4D\x1E\x00\x86\x48\x00\x00\x87\x6C"
-     "\x00\x00"
-     "\xAE\x39\x1E\x00\xAF\x39\x1E\x00\xAC\x55\x1E\x00\xAD\x3D\x1E\x00\xA8\x5D"
-     "\x1E\x00"
-     "\xA9\x6D\x1E\x00\xAA\x65\x1E\x00\xAB\x7D\x1E\x00\xAC\x65\x1E\x00\xAD\x6D"
-     "\x1E\x00"
-     "\xAE\x65\x1E\x00\xAF\xFD\x1E\x00\x95\xD4\x00\x80\x99\xD4\x00\x80\x9D\xD4"
-     "\x00\x80"
-     "\xA1\xD4\x00\x80\xA5\xD4\x00\x80\xA9\xD4\x00\x80\xAD\xD4\x00\x80\xB1\xD4"
-     "\x00\x80"
-     "\xB8\x69\x01\x00\xB9\x69\x01\x00\xBA\x79\x01\x00\xBB\x79\x01\x00\xBC\x69"
-     "\x01\x00"
-     "\xBD\x69\x01\x00\xBE\xDD\x01\x00\xBF\xD5\x01\x00\xB0\x85\x1E\x00\xB1\x8D"
-     "\x1E\x00"
-     "\xB2\x85\x1E\x00\xB3\x9D\x1E\x00\xB4\x85\x1E\x00\xB5\x8D\x1E\x00\xB6\x85"
-     "\x1E\x00"
-     "\xB7\x59\x01\x00\xB3\xED\x1E\x00\xB5\xD4\x00\x80\xB9\xD4\x00\x80\xBD\xD4"
-     "\x00\x80"
-     "\xC1\xD4\x00\x80\xB6\xED\x1E\x00\xB5\xED\x1E\x00\xC5\xD4\x00\x80\xBB\x49"
-     "\x1E\x00"
-     "\xBA\x41\x1E\x00\xC9\xD4\x00\x80\xCD\xD4\x00\x80\xBF\x49\x1E\x00\xBE\x41"
-     "\x1E\x00"
-     "\xBD\x49\x1E\x00\xBC\x51\x1E\x00\xD1\xD4\x00\x80\xA3\xA9\x1E\x00\xD5\xD4"
-     "\x00\x80"
-     "\xD9\xD4\x00\x80\xA6\xA9\x1E\x00\xDD\xD4\x00\x80\xE1\xD4\x00\x80\xA5\xA9"
-     "\x1E\x00"
-     "\xAA\x05\x1E\x00\xAB\x0D\x1E\x00\xE5\xD4\x00\x80\xE9\xD4\x00\x80\xAE\x05"
-     "\x1E\x00"
-     "\xAF\x0D\x1E\x00\xAC\x15\x1E\x00\xAD\x0D\x1E\x00\xA8\x21\x03\x00\xA9\x21"
-     "\x03\x00"
-     "\xAA\x21\x03\x00\xAB\x21\x03\x00\xAC\x21\x03\x00\xAD\x21\x03\x00\xAE\x21"
-     "\x03\x00"
-     "\xAF\x21\x03\x00\xED\xD4\x00\x80\xF1\xD4\x00\x80\xF5\xD4\x00\x80\xBE\x60"
-     "\x02\x00"
-     "\xF9\xD4\x00\x80\xFD\xD4\x00\x80\x05\xD5\x00\x80\x09\xD5\x00\x80\xB8\x89"
-     "\x03\x00"
-     "\xB9\x89\x03\x00\xBA\x9D\x03\x00\xBB\x95\x03\x00\xBC\xB9\x03\x00\xBD\xB9"
-     "\x03\x00"
-     "\xBE\x79\x00\x00\xBF\x79\x00\x00\xB0\xE5\x03\x00\xB1\xED\x03\x00\xB2\xE5"
-     "\x03\x00"
-     "\xB3\xFD\x03\x00\xB4\xED\x03\x00\xB5\xDD\x03\x00\xB6\xD5\x03\x00\xB7\xBD"
-     "\x03\x00"
-     "\x80\xA9\x00\x00\x81\xB5\x00\x00\x82\xBD\x00\x00\xB3\xA5\x03\x00\x0D\xD5"
-     "\x00\x80"
-     "\xB5\xA5\x03\x00\xB6\xAD\x03\x00\x11\xD5\x00\x80\x84\xE0\x02\x00\x15\xD5"
-     "\x00\x80"
-     "\xBA\x2D\x03\x00\xBB\x25\x03\x00\xBC\x3D\x03\x00\xBD\x25\x03\x00\xBE\x25"
-     "\x03\x00"
-     "\xBF\x15\x03\x00\xA3\xE9\x03\x00\x19\xD5\x00\x80\x86\x68\x04\x00\x87\x80"
-     "\x03\x00"
-     "\x1D\xD5\x00\x80\xA6\xE1\x03\x00\xA5\xE9\x03\x00\x21\xD5\x00\x80\xAB\x69"
-     "\x03\x00"
-     "\xAA\x61\x03\x00\x25\xD5\x00\x80\x29\xD5\x00\x80\xAF\x59\x03\x00\xAE\x69"
-     "\x03\x00"
-     "\xAD\x69\x03\x00\xAC\x71\x03\x00\x2D\xD5\x00\x80\x31\xD5\x00\x80\x35\xD5"
-     "\x00\x80"
-     "\x39\xD5\x00\x80\x3D\xD5\x00\x80\xE1\x3C\x1F\x00\x41\xD5\x00\x80\xE3\x40"
-     "\x1E\x00"
-     "\x45\xD5\x00\x80\x49\xD5\x00\x80\x4D\xD5\x00\x80\xEF\x8C\x1E\x00\x51\xD5"
-     "\x00\x80"
-     "\x55\xD5\x00\x80\x59\xD5\x00\x80\x5D\xD5\x00\x80\x61\xD5\x00\x80\x82\x25"
-     "\x00\x00"
-     "\x81\x11\x00\x00\x80\x1D\x00\x00\x65\xD5\x00\x80\xE3\xE4\x03\x00\x69\xD5"
-     "\x00\x80"
-     "\xE1\xAC\x01\x00\x6D\xD5\x00\x80\xEF\xB0\x03\x00\x84\x64\x02\x00\xBE\x60"
-     "\x05\x00"
-     "\x86\xD0\x04\x00\x87\x44\x05\x00\x75\xD5\x00\x80\x79\xD5\x00\x80\x7D\xD5"
-     "\x00\x80"
-     "\x81\xD5\x00\x80\x85\xD5\x00\x80\x89\xD5\x00\x80\x8D\xD5\x00\x80\x91\xD5"
-     "\x00\x80"
-     "\x95\xD5\x00\x80\xEF\xB0\x01\x00\x84\xA4\x05\x00\xE1\xDC\x1E\x00\x99\xD5"
-     "\x00\x80"
-     "\xE3\x10\x01\x00\x9D\xD5\x00\x80\xA1\xD5\x00\x80\xA5\xD5\x00\x80\xA9\xD5"
-     "\x00\x80"
-     "\xB3\x51\x01\x00\xAD\xD5\x00\x80\xB1\xD5\x00\x80\xB5\xD5\x00\x80\xB9\xD5"
-     "\x00\x80"
-     "\xB6\x11\x01\x00\xB5\x7D\x01\x00\xBD\xD5\x00\x80\xBB\x0D\x01\x00\xBA\x0D"
-     "\x01\x00"
-     "\xC1\xD5\x00\x80\xC5\xD5\x00\x80\xBF\xFD\x00\x00\xBE\xFD\x00\x00\xBD\xFD"
-     "\x00\x00"
-     "\xBC\xFD\x00\x00\xA8\x39\x06\x00\xA9\x39\x06\x00\xAA\x99\x06\x00\xAB\x91"
-     "\x06\x00"
-     "\xAC\xB5\x06\x00\xAD\xD1\x06\x00\xAE\xC9\x06\x00\xAF\xC1\x06\x00\x71\xD5"
-     "\x00\x80"
-     "\xC9\xD5\x00\x80\xCD\xD5\x00\x80\xD1\xD5\x00\x80\x80\x0D\x00\x00\x81\xB1"
-     "\x00\x00"
-     "\x82\xB1\x00\x00\xD5\xD5\x00\x80\xB8\x61\x07\x00\xB9\x61\x07\x00\xBA\x61"
-     "\x07\x00"
-     "\xBB\x7D\x07\x00\xBC\x65\x07\x00\xBD\x6D\x07\x00\xBE\x65\x07\x00\xBF\x1D"
-     "\x07\x00"
-     "\xB0\x89\x06\x00\xB1\x89\x06\x00\xB2\x69\x07\x00\xB3\x69\x07\x00\xB4\x79"
-     "\x07\x00"
-     "\xB5\x79\x07\x00\xB6\x69\x07\x00\xB7\x65\x07\x00\xA3\x11\x06\x00\xD9\xD5"
-     "\x00\x80"
-     "\xDD\xD5\x00\x80\xBE\x80\x03\x00\xE1\xD5\x00\x80\xA6\x51\x06\x00\xA5\x3D"
-     "\x06\x00"
-     "\xE5\xD5\x00\x80\xAB\x4D\x06\x00\xAA\x4D\x06\x00\x86\x08\x00\x00\x87\x7C"
-     "\x03\x00"
-     "\xAF\xBD\x07\x00\xAE\xBD\x07\x00\xAD\xBD\x07\x00\xAC\xBD\x07\x00\xE9\xD5"
-     "\x00\x80"
-     "\xED\xD5\x00\x80\xB3\x49\x07\x00\xF1\xD5\x00\x80\xB5\x59\x07\x00\xF5\xD5"
-     "\x00\x80"
-     "\xF9\xD5\x00\x80\xB6\x51\x07\x00\xFD\xD5\x00\x80\x01\xD5\x00\x80\xBB\x41"
-     "\x07\x00"
-     "\xBA\x75\x07\x00\xBD\x45\x07\x00\xBC\x45\x07\x00\xBF\x45\x07\x00\xBE\x45"
-     "\x07\x00"
-     "\xA8\x79\x06\x00\xA9\x79\x06\x00\xAA\x89\x06\x00\xAB\x89\x06\x00\xAC\x99"
-     "\x06\x00"
-     "\xAD\x99\x06\x00\xAE\x89\x06\x00\xAF\x89\x06\x00\x01\xD6\x00\x80\x05\xD6"
-     "\x00\x80"
-     "\x09\xD6\x00\x80\x0D\xD6\x00\x80\x11\xD6\x00\x80\x15\xD6\x00\x80\x19\xD6"
-     "\x00\x80"
-     "\x1D\xD6\x00\x80\xB8\x8D\x06\x00\xB9\x95\x06\x00\xBA\x95\x06\x00\xBB\xA5"
-     "\x06\x00"
-     "\xBC\xBD\x06\x00\xBD\x71\x01\x00\xBE\x71\x01\x00\xBF\x71\x01\x00\xB0\xF9"
-     "\x06\x00"
-     "\xB1\xCD\x06\x00\xB2\xD9\x06\x00\xB3\xD9\x06\x00\xB4\xC9\x06\x00\xB5\xC9"
-     "\x06\x00"
-     "\xB6\xBD\x06\x00\xB7\xB5\x06\x00\xA3\x01\x06\x00\x21\xD6\x00\x80\x25\xD6"
-     "\x00\x80"
-     "\x29\xD6\x00\x80\x2D\xD6\x00\x80\xA6\x19\x06\x00\xA5\x11\x06\x00\x31\xD6"
-     "\x00\x80"
-     "\xAB\x09\x06\x00\xAA\x3D\x06\x00\x35\xD6\x00\x80\x39\xD6\x00\x80\xAF\x0D"
-     "\x06\x00"
-     "\xAE\x0D\x06\x00\xAD\x0D\x06\x00\xAC\x0D\x06\x00\x3D\xD6\x00\x80\x41\xD6"
-     "\x00\x80"
-     "\x45\xD6\x00\x80\x49\xD6\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x4D\xD6\x00\x80\x84\xAC\x01\x00\xBE\xAC\x01\x00\x87\xE8\x00\x00\x86\x4C"
-     "\x0F\x00"
-     "\x55\xD6\x00\x80\x59\xD6\x00\x80\x5D\xD6\x00\x80\x61\xD6\x00\x80\xA8\x9D"
-     "\x02\x00"
-     "\xA9\xC5\x02\x00\xAA\xCD\x02\x00\xAB\xC1\x02\x00\xAC\xC5\x02\x00\xAD\xCD"
-     "\x02\x00"
-     "\xAE\xF9\x02\x00\xAF\x3D\x03\x00\x65\xD6\x00\x80\x69\xD6\x00\x80\x6D\xD6"
-     "\x00\x80"
-     "\x71\xD6\x00\x80\xBE\x3C\x0C\x00\x75\xD6\x00\x80\x79\xD6\x00\x80\x7D\xD6"
-     "\x00\x80"
-     "\xB8\xC9\x03\x00\xB9\xC9\x03\x00\xBA\xD9\x03\x00\xBB\xD1\x03\x00\xBC\xF9"
-     "\x03\x00"
-     "\xBD\xF9\x03\x00\xBE\x99\x03\x00\xBF\x99\x03\x00\xB0\x45\x03\x00\xB1\x4D"
-     "\x03\x00"
-     "\xB2\x45\x03\x00\xB3\x5D\x03\x00\xB4\x45\x03\x00\xB5\x4D\x03\x00\xB6\x45"
-     "\x03\x00"
-     "\xB7\xF9\x03\x00\xB3\x45\x02\x00\x81\xD6\x00\x80\x85\xD6\x00\x80\x89\xD6"
-     "\x00\x80"
-     "\x8D\xD6\x00\x80\xB6\x4D\x02\x00\xB5\x4D\x02\x00\x91\xD6\x00\x80\xBB\x89"
-     "\x03\x00"
-     "\xBA\x81\x03\x00\x95\xD6\x00\x80\x99\xD6\x00\x80\xBF\x89\x03\x00\xBE\x81"
-     "\x03\x00"
-     "\xBD\x89\x03\x00\xBC\x91\x03\x00\x9D\xD6\x00\x80\xA3\x01\x02\x00\xA1\xD6"
-     "\x00\x80"
-     "\xA5\xD6\x00\x80\xA6\x09\x02\x00\xA9\xD6\x00\x80\xAD\xD6\x00\x80\xA5\x09"
-     "\x02\x00"
-     "\xAA\xC5\x03\x00\xAB\xCD\x03\x00\xB1\xD6\x00\x80\xB5\xD6\x00\x80\xAE\xC5"
-     "\x03\x00"
-     "\xAF\xCD\x03\x00\xAC\xD5\x03\x00\xAD\xCD\x03\x00\x80\xED\x01\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x11\x00\x00\x84\x00\x02\x00\xB9\xD6\x00\x80\xE1\xA4\x01\x00\xBD\xD6"
-     "\x00\x80"
-     "\xE3\xEC\x00\x00\xC5\xD6\x00\x80\xC9\xD6\x00\x80\xCD\xD6\x00\x80\xEF\x30"
-     "\x00\x00"
-     "\xD1\xD6\x00\x80\xD5\xD6\x00\x80\xD9\xD6\x00\x80\xDD\xD6\x00\x80\x86\xE0"
-     "\x0C\x00"
-     "\x87\xF4\x02\x00\xE1\xD6\x00\x80\xE5\xD6\x00\x80\xE9\xD6\x00\x80\xED\xD6"
-     "\x00\x80"
-     "\xEF\xCC\x06\x00\xF1\xD6\x00\x80\xE1\xB0\x07\x00\xF5\xD6\x00\x80\xE3\x10"
-     "\x06\x00"
-     "\xF9\xD6\x00\x80\xFD\xD6\x00\x80\x01\xD7\x00\x80\x05\xD7\x00\x80\x09\xD7"
-     "\x00\x80"
-     "\x0D\xD7\x00\x80\x11\xD7\x00\x80\x15\xD7\x00\x80\x19\xD7\x00\x80\x1D\xD7"
-     "\x00\x80"
-     "\x21\xD7\x00\x80\x25\xD7\x00\x80\xEF\xE8\x01\x00\x85\x18\x0F\x00\xE1\xCC"
-     "\x06\x00"
-     "\x29\xD7\x00\x80\xE3\x1C\x06\x00\x80\x29\x00\x00\x81\x1D\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x2D\xD7\x00\x80\xB3\x39\x02\x00\x84\xCC\x0D\x00\x86\x68\x0F\x00\x87\xFC"
-     "\x0C\x00"
-     "\xE1\xD0\xD2\x03\xB6\xF1\x01\x00\xB5\xF9\x01\x00\x35\xD7\x00\x80\xBB\xD9"
-     "\x01\x00"
-     "\xBA\xD1\x01\x00\xBE\xE4\x0C\x00\x39\xD7\x00\x80\xBF\x7D\x01\x00\xBE\x7D"
-     "\x01\x00"
-     "\xBD\x7D\x01\x00\xBC\xC1\x01\x00\xA8\xF1\x0D\x00\xA9\xF1\x0D\x00\xAA\xF1"
-     "\x0D\x00"
-     "\xAB\xF1\x0D\x00\xAC\x31\x0E\x00\xAD\x31\x0E\x00\xAE\x31\x0E\x00\xAF\x31"
-     "\x0E\x00"
-     "\xC1\xD6\x00\x80\x31\xD7\x00\x80\x3D\xD7\x00\x80\x41\xD7\x00\x80\x45\xD7"
-     "\x00\x80"
-     "\x49\xD7\x00\x80\x4D\xD7\x00\x80\x51\xD7\x00\x80\xB8\xE9\x0E\x00\xB9\xE9"
-     "\x0E\x00"
-     "\xBA\x89\x0E\x00\xBB\x85\x0E\x00\xBC\x9D\x0E\x00\xBD\x81\x0E\x00\xBE\x81"
-     "\x0E\x00"
-     "\xBF\xB5\x0E\x00\xB0\x55\x0E\x00\xB1\x5D\x0E\x00\xB2\x55\x0E\x00\xB3\xE9"
-     "\x0E\x00"
-     "\xB4\xF9\x0E\x00\xB5\xF9\x0E\x00\xB6\xE9\x0E\x00\xB7\xE1\x0E\x00\xA3\x79"
-     "\x0D\x00"
-     "\x55\xD7\x00\x80\x59\xD7\x00\x80\x5D\xD7\x00\x80\x61\xD7\x00\x80\xA6\xB1"
-     "\x0E\x00"
-     "\xA5\xB9\x0E\x00\x65\xD7\x00\x80\xAB\x99\x0E\x00\xAA\x91\x0E\x00\x69\xD7"
-     "\x00\x80"
-     "\x6D\xD7\x00\x80\xAF\x3D\x0E\x00\xAE\x3D\x0E\x00\xAD\x3D\x0E\x00\xAC\x81"
-     "\x0E\x00"
-     "\x71\xD7\x00\x80\xB3\xED\x0F\x00\x75\xD7\x00\x80\x79\xD7\x00\x80\xB6\xE9"
-     "\x0F\x00"
-     "\x7D\xD7\x00\x80\x81\xD7\x00\x80\xB5\xE9\x0F\x00\xBA\xAD\x0F\x00\xBB\xB5"
-     "\x0F\x00"
-     "\x51\xD6\x00\x80\x85\xD7\x00\x80\xBE\x95\x0F\x00\xBF\x99\x0F\x00\xBC\xAD"
-     "\x0F\x00"
-     "\xBD\xA1\x0F\x00\xA8\x21\x0E\x00\xA9\x21\x0E\x00\xAA\x21\x0E\x00\xAB\x3D"
-     "\x0E\x00"
-     "\xAC\x25\x0E\x00\xAD\x2D\x0E\x00\xAE\x25\x0E\x00\xAF\x55\x0E\x00\x89\xD7"
-     "\x00\x80"
-     "\x8D\xD7\x00\x80\x91\xD7\x00\x80\x95\xD7\x00\x80\x80\x1D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\xBD\x00\x00\x99\xD7\x00\x80\xB8\x39\x0E\x00\xB9\x39\x0E\x00\xBA\xC9"
-     "\x0E\x00"
-     "\xBB\xC9\x0E\x00\xBC\xD9\x0E\x00\xBD\xD9\x0E\x00\xBE\xC9\x0E\x00\xBF\xC9"
-     "\x0E\x00"
-     "\xB0\x2D\x0E\x00\xB1\x35\x0E\x00\xB2\x3D\x0E\x00\xB3\x31\x0E\x00\xB4\x15"
-     "\x0E\x00"
-     "\xB5\x19\x0E\x00\xB6\x09\x0E\x00\xB7\x09\x0E\x00\xA3\xA9\x0E\x00\x9D\xD7"
-     "\x00\x80"
-     "\x84\x80\x02\x00\xBE\x80\x01\x00\x85\x00\x04\x00\xA6\xAD\x0E\x00\xA5\xAD"
-     "\x0E\x00"
-     "\xA5\xD7\x00\x80\xAB\xF1\x0E\x00\xAA\xE9\x0E\x00\x86\x28\x07\x00\x87\x18"
-     "\x00\x00"
-     "\xAF\xDD\x0E\x00\xAE\xD1\x0E\x00\xAD\xE5\x0E\x00\xAC\xE9\x0E\x00\xA9\xD7"
-     "\x00\x80"
-     "\xB3\xED\x01\x00\xAD\xD7\x00\x80\xB1\xD7\x00\x80\xB6\xE5\x01\x00\xB5\xD7"
-     "\x00\x80"
-     "\xB9\xD7\x00\x80\xB5\xED\x01\x00\xBA\x65\x01\x00\xBB\x6D\x01\x00\xBD\xD7"
-     "\x00\x80"
-     "\xC1\xD7\x00\x80\xBE\x6D\x01\x00\xBF\x5D\x01\x00\xBC\x75\x01\x00\xBD\x6D"
-     "\x01\x00"
-     "\xA8\xDD\x0D\x00\xA9\xE9\x0D\x00\xAA\x21\x02\x00\xAB\x21\x02\x00\xAC\x21"
-     "\x02\x00"
-     "\xAD\x21\x02\x00\xAE\x21\x02\x00\xAF\x21\x02\x00\xC5\xD7\x00\x80\xC9\xD7"
-     "\x00\x80"
-     "\xCD\xD7\x00\x80\xD1\xD7\x00\x80\xA2\x11\x02\x00\xA3\x11\x02\x00\xA0\xA9"
-     "\x0E\x00"
-     "\xA1\xD5\x02\x00\xB8\x89\x02\x00\xB9\x89\x02\x00\xBA\x9D\x02\x00\xBB\x95"
-     "\x02\x00"
-     "\xBC\xBD\x02\x00\xBD\x75\x03\x00\xBE\x7D\x03\x00\xBF\x75\x03\x00\xB0\xE5"
-     "\x02\x00"
-     "\xB1\xED\x02\x00\xB2\xE5\x02\x00\xB3\xFD\x02\x00\xB4\xED\x02\x00\xB5\xDD"
-     "\x02\x00"
-     "\xB6\xD5\x02\x00\xB7\xBD\x02\x00\xA3\xA9\x02\x00\x8F\xC5\x1A\x00\xD5\xD7"
-     "\x00\x80"
-     "\xD9\xD7\x00\x80\xDD\xD7\x00\x80\xA6\xA1\x02\x00\xA5\xA9\x02\x00\xE1\xD7"
-     "\x00\x80"
-     "\xAB\x29\x02\x00\xAA\x21\x02\x00\xE5\xD7\x00\x80\xE9\xD7\x00\x80\xAF\x19"
-     "\x02\x00"
-     "\xAE\x29\x02\x00\xAD\x29\x02\x00\xAC\x31\x02\x00\x9E\x25\x0E\x00\x9F\xE5"
-     "\x0E\x00"
-     "\x9C\xE9\x0A\x00\x9D\x15\x0A\x00\x9A\x45\x16\x00\x9B\x45\x0A\x00\x98\x59"
-     "\x16\x00"
-     "\x99\x51\x16\x00\x96\x71\x12\x00\x97\x81\x13\x00\x94\x55\x12\x00\x95\xED"
-     "\x12\x00"
-     "\x92\xD1\x1E\x00\x93\xD9\x1E\x00\x90\xB5\x1A\x00\x91\x55\x1E\x00\x84\xA9"
-     "\x1F\x00"
-     "\x85\x25\x1F\x00\x86\x25\x1F\x00\x87\xB1\x13\x00\xED\xD7\x00\x80\xF1\xD7"
-     "\x00\x80"
-     "\x82\x59\x1B\x00\x83\x51\x1B\x00\x8C\x45\x12\x00\x8D\xA5\x17\x00\x8E\xA5"
-     "\x17\x00"
-     "\x8F\xB9\x17\x00\x80\x39\xFB\x01\xF5\xD7\x00\x80\x8A\x39\x13\x00\x8B\xAD"
-     "\x13\x00"
-     "\x94\x99\x0B\x00\x95\xA1\x0F\x00\x96\x69\x0F\x00\x97\xDD\x0F\x00\xF9\xD7"
-     "\x00\x80"
-     "\xEF\x9C\x0F\x00\x92\xC9\x0B\x00\x93\x7D\x0B\x00\x9C\x45\x03\x00\xE3\x78"
-     "\x0E\x00"
-     "\xFD\xD7\x00\x80\xE1\x98\x0C\x00\x01\xD8\x00\x80\x84\x78\x02\x00\x9A\x91"
-     "\x03\x00"
-     "\x9B\x5D\x03\x00\xE1\x04\x00\x00\xBE\x88\x05\x00\xE3\xDC\xEA\x03\x05\xD8"
-     "\x00\x80"
-     "\x09\xD8\x00\x80\x0D\xD8\x00\x80\xEF\xEC\x00\x00\xEF\x8C\x0E\x00\xE1\x70"
-     "\x0E\x00"
-     "\xE1\xFC\x0E\x00\xE3\x30\x00\x00\xE3\x78\x0E\x00\x81\x21\x00\x00\x80\x39"
-     "\x00\x00"
-     "\xEF\xB4\xED\x03\x82\x29\x00\x00\xB3\x09\x02\x00\x15\xD8\x00\x80\x86\x68"
-     "\x04\x00"
-     "\x87\x2C\x05\x00\x19\xD8\x00\x80\xB6\x0D\x02\x00\xB5\x0D\x02\x00\x1D\xD8"
-     "\x00\x80"
-     "\xBB\xC5\x01\x00\xBA\xC5\x01\x00\x21\xD8\x00\x80\x25\xD8\x00\x80\xBF\x7D"
-     "\x01\x00"
-     "\xBE\x7D\x01\x00\xBD\xD5\x01\x00\xBC\xD5\x01\x00\xA1\xD7\x00\x80\x11\xD8"
-     "\x00\x80"
-     "\x29\xD8\x00\x80\x2D\xD8\x00\x80\x31\xD8\x00\x80\x35\xD8\x00\x80\x39\xD8"
-     "\x00\x80"
-     "\x3D\xD8\x00\x80\xA8\xBD\x05\x00\xA9\xE5\x05\x00\xAA\xE1\x05\x00\xAB\xE1"
-     "\x05\x00"
-     "\xAC\xE5\x05\x00\xAD\xED\x05\x00\xAE\xD1\x05\x00\xAF\xD1\x05\x00\xB0\x61"
-     "\x06\x00"
-     "\xB1\x61\x06\x00\xB2\x61\x06\x00\xB3\x61\x06\x00\xB4\xD9\x06\x00\xB5\xD9"
-     "\x06\x00"
-     "\xB6\xC9\x06\x00\xB7\xC1\x06\x00\xB8\xC9\x06\x00\xB9\xC9\x06\x00\xBA\x79"
-     "\x07\x00"
-     "\xBB\x79\x07\x00\xBC\x45\x07\x00\xBD\x25\x07\x00\xBE\x11\x07\x00\xBF\x0D"
-     "\x07\x00"
-     "\xA3\x49\x05\x00\x41\xD8\x00\x80\x45\xD8\x00\x80\x49\xD8\x00\x80\x4D\xD8"
-     "\x00\x80"
-     "\xA6\x4D\x05\x00\xA5\x4D\x05\x00\x51\xD8\x00\x80\xAB\x85\x06\x00\xAA\x85"
-     "\x06\x00"
-     "\x55\xD8\x00\x80\x59\xD8\x00\x80\xAF\x3D\x06\x00\xAE\x3D\x06\x00\xAD\x95"
-     "\x06\x00"
-     "\xAC\x95\x06\x00\x5D\xD8\x00\x80\x61\xD8\x00\x80\x65\xD8\x00\x80\x69\xD8"
-     "\x00\x80"
-     "\x6D\xD8\x00\x80\x71\xD8\x00\x80\x75\xD8\x00\x80\x79\xD8\x00\x80\x80\x39"
-     "\x00\x00"
-     "\x81\x39\x00\x00\x82\x05\x00\x00\x7D\xD8\x00\x80\xBE\xB8\x03\x00\x84\xB8"
-     "\x03\x00"
-     "\x85\xD8\x00\x80\x89\xD8\x00\x80\xA8\xC5\x06\x00\xA9\xD5\x06\x00\xAA\xD5"
-     "\x06\x00"
-     "\xAB\xE5\x06\x00\xAC\xFD\x06\x00\xAD\x1D\x01\x00\xAE\x15\x01\x00\xAF\x0D"
-     "\x01\x00"
-     "\x81\xD8\x00\x80\x8D\xD8\x00\x80\x86\x88\x01\x00\x87\x1C\x01\x00\x91\xD8"
-     "\x00\x80"
-     "\x95\xD8\x00\x80\x99\xD8\x00\x80\x9D\xD8\x00\x80\xB8\xE9\x01\x00\xB9\xE9"
-     "\x01\x00"
-     "\xBA\x89\x01\x00\xBB\x89\x01\x00\xBC\x99\x01\x00\xBD\x99\x01\x00\xBE\x89"
-     "\x01\x00"
-     "\xBF\x89\x01\x00\xB0\x75\x01\x00\xB1\x7D\x01\x00\xB2\x75\x01\x00\xB3\xE9"
-     "\x01\x00"
-     "\xB4\xF9\x01\x00\xB5\xF9\x01\x00\xB6\xE9\x01\x00\xB7\xE1\x01\x00\xB3\x5D"
-     "\x06\x00"
-     "\xA1\xD8\x00\x80\xA5\xD8\x00\x80\xA9\xD8\x00\x80\x84\xBC\x01\x00\xB6\x9D"
-     "\x01\x00"
-     "\xB5\x75\x06\x00\xAD\xD8\x00\x80\xBB\xB9\x01\x00\xBA\xB1\x01\x00\xB1\xD8"
-     "\x00\x80"
-     "\xB5\xD8\x00\x80\xBF\x3D\x01\x00\xBE\x3D\x01\x00\xBD\x3D\x01\x00\xBC\xA1"
-     "\x01\x00"
-     "\xB9\xD8\x00\x80\xA3\x19\x06\x00\xBD\xD8\x00\x80\xC1\xD8\x00\x80\xA6\xD9"
-     "\x01\x00"
-     "\xC5\xD8\x00\x80\xC9\xD8\x00\x80\xA5\x31\x06\x00\xAA\xF5\x01\x00\xAB\xFD"
-     "\x01\x00"
-     "\xCD\xD8\x00\x80\xD1\xD8\x00\x80\xAE\x79\x01\x00\xAF\x79\x01\x00\xAC\xE5"
-     "\x01\x00"
-     "\xAD\x79\x01\x00\xA8\xF9\x02\x00\xA9\xF9\x02\x00\xAA\x2D\x03\x00\xAB\x3D"
-     "\x03\x00"
-     "\xAC\x25\x03\x00\xAD\x2D\x03\x00\xAE\x25\x03\x00\xAF\x99\x03\x00\xD5\xD8"
-     "\x00\x80"
-     "\xD9\xD8\x00\x80\xDD\xD8\x00\x80\xE1\xD8\x00\x80\x80\x0D\x00\x00\x81\xB1"
-     "\x00\x00"
-     "\x82\xB1\x00\x00\xE5\xD8\x00\x80\xB8\x95\x03\x00\xB9\x9D\x03\x00\xBA\xA1"
-     "\x03\x00"
-     "\xBB\xA1\x03\x00\xBC\x71\x00\x00\xBD\x71\x00\x00\xBE\x71\x00\x00\xBF\x71"
-     "\x00\x00"
-     "\xB0\xE9\x03\x00\xB1\xE9\x03\x00\xB2\xF5\x03\x00\xB3\xC5\x03\x00\xB4\xDD"
-     "\x03\x00"
-     "\xB5\xB5\x03\x00\xB6\xB1\x03\x00\xB7\xB1\x03\x00\xBE\xCC\x03\x00\xE9\xD8"
-     "\x00\x80"
-     "\xED\xD8\x00\x80\xF5\xD8\x00\x80\xF9\xD8\x00\x80\xFD\xD8\x00\x80\x01\xD9"
-     "\x00\x80"
-     "\xEF\xE4\x02\x00\x05\xD9\x00\x80\xE1\x94\x01\x00\x09\xD9\x00\x80\xE3\x2C"
-     "\x01\x00"
-     "\x0D\xD9\x00\x80\x11\xD9\x00\x80\x87\x18\x03\x00\x86\x1C\xFC\x03\xBB\x4D"
-     "\x03\x00"
-     "\xBA\x4D\x03\x00\x15\xD9\x00\x80\x19\xD9\x00\x80\xBF\x11\x03\x00\xBE\x79"
-     "\x03\x00"
-     "\xBD\x71\x03\x00\xBC\x41\x03\x00\xB3\xC5\x03\x00\x84\xE8\xFC\x03\x1D\xD9"
-     "\x00\x80"
-     "\x21\xD9\x00\x80\x25\xD9\x00\x80\xB6\xCD\x03\x00\xB5\xCD\x03\x00\x29\xD9"
-     "\x00\x80"
-     "\xA4\x01\xFC\x03\xA5\x25\xFF\x03\xA6\xD9\xFF\x03\xA7\x01\xF8\x03\x2D\xD9"
-     "\x00\x80"
-     "\xA1\x15\x03\x00\xA2\x1D\x03\x00\xA3\x3D\x02\x00\xAC\x11\xF7\x03\xAD\x01"
-     "\xF0\x03"
-     "\xAE\x2D\xF3\x03\xAF\x01\xF3\x03\xA8\x11\xFB\x03\xA9\x99\xFB\x03\xAA\x01"
-     "\xF4\x03"
-     "\xAB\x1D\xF7\x03\xB4\x01\xE8\x03\xB5\x65\xEB\x03\xBE\xC4\xFC\x03\x84\xC4"
-     "\xFC\x03"
-     "\xB0\x01\xEC\x03\xB1\x55\xEF\x03\xB2\x4D\xEF\x03\xB3\x71\xEE\x03\x31\xD9"
-     "\x00\x80"
-     "\x35\xD9\x00\x80\x39\xD9\x00\x80\x3D\xD9\x00\x80\x41\xD9\x00\x80\x45\xD9"
-     "\x00\x80"
-     "\x49\xD9\x00\x80\x4D\xD9\x00\x80\xE1\x50\x06\x00\xE1\x34\x04\x00\xE3\x6C"
-     "\x01\x00"
-     "\xE3\xE8\x06\x00\x51\xD9\x00\x80\x55\xD9\x00\x80\x59\xD9\x00\x80\x5D\xD9"
-     "\x00\x80"
-     "\x80\x35\x00\x00\x81\x3D\x00\x00\x82\x35\x00\x00\x65\xD9\x00\x80\x69\xD9"
-     "\x00\x80"
-     "\x6D\xD9\x00\x80\xEF\xB0\x01\x00\xEF\xC0\x06\x00\xA3\xE5\x02\x00\x71\xD9"
-     "\x00\x80"
-     "\x86\xE8\xFC\x03\x87\x7C\xFD\x03\x75\xD9\x00\x80\xA6\xED\x02\x00\xA5\xED"
-     "\x02\x00"
-     "\x79\xD9\x00\x80\xAB\x6D\x02\x00\xAA\x6D\x02\x00\x7D\xD9\x00\x80\x81\xD9"
-     "\x00\x80"
-     "\xAF\x31\x02\x00\xAE\x59\x02\x00\xAD\x51\x02\x00\xAC\x61\x02\x00\xA8\x8D"
-     "\xFE\x03"
-     "\xA9\x95\xFE\x03\xAA\x9D\xFE\x03\xAB\x95\xFE\x03\xAC\xB1\xFE\x03\xAD\xBD"
-     "\xFE\x03"
-     "\xAE\xA9\xFE\x03\xAF\xA9\xFE\x03\x61\xD9\x00\x80\x85\xD9\x00\x80\x89\xD9"
-     "\x00\x80"
-     "\x8D\xD9\x00\x80\x91\xD9\x00\x80\x95\xD9\x00\x80\x99\xD9\x00\x80\x9D\xD9"
-     "\x00\x80"
-     "\xB8\x25\xFF\x03\xB9\x2D\xFF\x03\xBA\x25\xFF\x03\xBB\x3D\xFF\x03\xBC\x25"
-     "\xFF\x03"
-     "\xBD\x2D\xFF\x03\xBE\x25\xFF\x03\xBF\xCD\xFF\x03\xB0\xA9\xFE\x03\xB1\xA9"
-     "\xFE\x03"
-     "\xB2\x81\xFE\x03\xB3\x81\xFE\x03\xB4\x81\xFE\x03\xB5\x89\xFE\x03\xB6\x69"
-     "\xFF\x03"
-     "\xB7\x1D\xFF\x03\xA1\xD9\x00\x80\xE1\x20\xFC\x03\xA5\xD9\x00\x80\xE3\x78"
-     "\xFC\x03"
-     "\xA9\xD9\x00\x80\xAD\xD9\x00\x80\xB1\xD9\x00\x80\xB5\xD9\x00\x80\xB9\xD9"
-     "\x00\x80"
-     "\xBD\xD9\x00\x80\xC1\xD9\x00\x80\xC5\xD9\x00\x80\x80\x1D\x00\x00\x81\x71"
-     "\x00\x00"
-     "\x82\x71\x00\x00\xEF\x0C\xFD\x03\xB3\x55\xFE\x03\xC9\xD9\x00\x80\xCD\xD9"
-     "\x00\x80"
-     "\xBE\x40\x00\x00\xD1\xD9\x00\x80\xB6\x7D\xFE\x03\xB5\x79\xFE\x03\xD5\xD9"
-     "\x00\x80"
-     "\xBB\x2D\xFE\x03\xBA\x2D\xFE\x03\x86\xE8\x00\x00\x87\xAC\x00\x00\xBF\x0D"
-     "\xFE\x03"
-     "\xBE\x05\xFE\x03\xBD\x15\xFE\x03\xBC\x15\xFE\x03\xA3\x95\xFF\x03\xD9\xD9"
-     "\x00\x80"
-     "\xDD\xD9\x00\x80\xE1\xD9\x00\x80\xE5\xD9\x00\x80\xA6\xBD\xFF\x03\xA5\xB9"
-     "\xFF\x03"
-     "\xE9\xD9\x00\x80\xAB\xED\xFF\x03\xAA\xED\xFF\x03\xED\xD9\x00\x80\xF1\xD9"
-     "\x00\x80"
-     "\xAF\xCD\xFF\x03\xAE\xC5\xFF\x03\xAD\xD5\xFF\x03\xAC\xD5\xFF\x03\xF5\xD9"
-     "\x00\x80"
-     "\xB3\xF1\xFE\x03\xF9\xD9\x00\x80\xFD\xD9\x00\x80\xB6\x8D\xFE\x03\x01\xDA"
-     "\x00\x80"
-     "\x05\xDA\x00\x80\xB5\x8D\xFE\x03\xBA\x45\x01\x00\xBB\x4D\x01\x00\x09\xDA"
-     "\x00\x80"
-     "\x0D\xDA\x00\x80\xBE\x45\x01\x00\xBF\x4D\x01\x00\xBC\x55\x01\x00\xBD\x4D"
-     "\x01\x00"
-     "\xA8\x2D\xFE\x03\xA9\x35\xFE\x03\xAA\x3D\xFE\x03\xAB\x49\xFE\x03\xAC\x41"
-     "\xFE\x03"
-     "\xAD\x49\xFE\x03\xAE\x71\xFE\x03\xAF\x71\xFE\x03\x11\xDA\x00\x80\x15\xDA"
-     "\x00\x80"
-     "\x19\xDA\x00\x80\x1D\xDA\x00\x80\x21\xDA\x00\x80\x25\xDA\x00\x80\x29\xDA"
-     "\x00\x80"
-     "\x2D\xDA\x00\x80\xB8\x49\x01\x00\xB9\x55\x01\x00\xBA\x5D\x01\x00\xBB\x55"
-     "\x01\x00"
-     "\xBC\x4D\x01\x00\xBD\x75\x01\x00\xBE\x7D\x01\x00\xBF\x75\x01\x00\xB0\xC5"
-     "\x01\x00"
-     "\xB1\xCD\x01\x00\xB2\xC5\x01\x00\xB3\xDD\x01\x00\xB4\xC5\x01\x00\xB5\xCD"
-     "\x01\x00"
-     "\xB6\xC5\x01\x00\xB7\x7D\x01\x00\xA3\xB5\xFD\x03\x31\xDA\x00\x80\x35\xDA"
-     "\x00\x80"
-     "\x84\xC4\x03\x00\x39\xDA\x00\x80\xA6\xC9\xFD\x03\xA5\xC9\xFD\x03\x3D\xDA"
-     "\x00\x80"
-     "\xAB\x09\x02\x00\xAA\x01\x02\x00\x45\xDA\x00\x80\xBE\xEC\x02\x00\xAF\x09"
-     "\x02\x00"
-     "\xAE\x01\x02\x00\xAD\x09\x02\x00\xAC\x11\x02\x00\x80\x49\x00\x00\x81\x55"
-     "\x00\x00"
-     "\x82\x55\x00\x00\xA3\x45\x03\x00\x49\xDA\x00\x80\xA5\x45\x03\x00\xA6\x45"
-     "\x03\x00"
-     "\x4D\xDA\x00\x80\x86\xC0\x04\x00\x87\x14\x03\x00\xAA\x29\x03\x00\xAB\x25"
-     "\x03\x00"
-     "\xAC\x3D\x03\x00\xAD\x21\x03\x00\xAE\x21\x03\x00\xAF\x15\x03\x00\x51\xDA"
-     "\x00\x80"
-     "\x55\xDA\x00\x80\x59\xDA\x00\x80\x5D\xDA\x00\x80\x61\xDA\x00\x80\x65\xDA"
-     "\x00\x80"
-     "\x69\xDA\x00\x80\x6D\xDA\x00\x80\xA8\x7D\x02\x00\xA9\xA1\x03\x00\xAA\xA1"
-     "\x03\x00"
-     "\xAB\xA1\x03\x00\xAC\xA1\x03\x00\xAD\xA9\x03\x00\xAE\x91\x03\x00\xAF\x91"
-     "\x03\x00"
-     "\xB0\x81\x03\x00\xB1\x8D\x03\x00\xB2\x85\x03\x00\xB3\x99\x03\x00\xB4\x89"
-     "\x03\x00"
-     "\xB5\xBD\x03\x00\xB6\xB5\x03\x00\xB7\x7D\x03\x00\xB8\x45\x03\x00\xB9\x4D"
-     "\x03\x00"
-     "\xBA\x45\x03\x00\xBB\x5D\x03\x00\xBC\x45\x03\x00\xBD\x4D\x03\x00\xBE\x45"
-     "\x03\x00"
-     "\xBF\xF9\x00\x00\xF1\xD8\x00\x80\xB3\x0D\x02\x00\x71\xDA\x00\x80\x75\xDA"
-     "\x00\x80"
-     "\xB6\x0D\x02\x00\x79\xDA\x00\x80\x7D\xDA\x00\x80\xB5\x0D\x02\x00\xBA\x61"
-     "\x02\x00"
-     "\xBB\x6D\x02\x00\x81\xDA\x00\x80\x85\xDA\x00\x80\xBE\x69\x02\x00\xBF\x5D"
-     "\x02\x00"
-     "\xBC\x75\x02\x00\xBD\x69\x02\x00\x89\xDA\x00\x80\x8D\xDA\x00\x80\x91\xDA"
-     "\x00\x80"
-     "\x95\xDA\x00\x80\x99\xDA\x00\x80\xE1\xA4\x01\x00\x9D\xDA\x00\x80\xE3\x40"
-     "\x03\x00"
-     "\xA1\xDA\x00\x80\xA5\xDA\x00\x80\xA9\xDA\x00\x80\xEF\xB8\x03\x00\x80\x15"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x82\x05\x00\x00\xAD\xDA\x00\x80\x84\x80\x02\x00\xBE\xC8"
-     "\x05\x00"
-     "\x87\x18\x05\x00\x86\x2C\x04\x00\xB5\xDA\x00\x80\xB9\xDA\x00\x80\xBD\xDA"
-     "\x00\x80"
-     "\xEF\xA8\x01\x00\xC1\xDA\x00\x80\xE1\x74\xFE\x03\xC5\xDA\x00\x80\xE3\xF0"
-     "\xFE\x03"
-     "\xC9\xDA\x00\x80\xCD\xDA\x00\x80\xD1\xDA\x00\x80\xD5\xDA\x00\x80\xD9\xDA"
-     "\x00\x80"
-     "\xDD\xDA\x00\x80\xE1\xDA\x00\x80\xB3\x91\x01\x00\xE5\xDA\x00\x80\xB5\x51"
-     "\x01\x00"
-     "\xB6\x51\x01\x00\xE9\xDA\x00\x80\xED\xDA\x00\x80\xF1\xDA\x00\x80\xBA\x7D"
-     "\x01\x00"
-     "\xBB\x75\x01\x00\xBC\x6D\x01\x00\xBD\xFD\x00\x00\xBE\xF5\x00\x00\xBF\xE9"
-     "\x00\x00"
-     "\xA8\x29\x06\x00\xA9\x55\x06\x00\xAA\x55\x06\x00\xAB\x8D\x06\x00\xAC\x95"
-     "\x06\x00"
-     "\xAD\x9D\x06\x00\xAE\x95\x06\x00\xAF\x8D\x06\x00\xB1\xDA\x00\x80\xF5\xDA"
-     "\x00\x80"
-     "\xF9\xDA\x00\x80\xFD\xDA\x00\x80\x01\xDB\x00\x80\x05\xDB\x00\x80\x09\xDB"
-     "\x00\x80"
-     "\x0D\xDB\x00\x80\xB8\x6D\x07\x00\xB9\x05\x07\x00\xBA\x0D\x07\x00\xBB\x05"
-     "\x07\x00"
-     "\xBC\x1D\x07\x00\xBD\x05\x07\x00\xBE\x01\x07\x00\xBF\x3D\x07\x00\xB0\xF5"
-     "\x06\x00"
-     "\xB1\xFD\x06\x00\xB2\xF5\x06\x00\xB3\x65\x07\x00\xB4\x7D\x07\x00\xB5\x61"
-     "\x07\x00"
-     "\xB6\x61\x07\x00\xB7\x55\x07\x00\xE3\x10\x05\x00\x11\xDB\x00\x80\xE1\xF0"
-     "\x04\x00"
-     "\x15\xDB\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39\x00\x00\x19\xDB"
-     "\x00\x80"
-     "\x1D\xDB\x00\x80\x84\xE8\x03\x00\xBE\xE0\x03\x00\x21\xDB\x00\x80\xEF\xCC"
-     "\x05\x00"
-     "\x25\xDB\x00\x80\x87\x38\x00\x00\x86\x18\x00\x00\xA3\x9D\x06\x00\x29\xDB"
-     "\x00\x80"
-     "\x2D\xDB\x00\x80\x31\xDB\x00\x80\x35\xDB\x00\x80\xA6\x5D\x06\x00\xA5\x5D"
-     "\x06\x00"
-     "\x39\xDB\x00\x80\xAB\x79\x06\x00\xAA\x71\x06\x00\x3D\xDB\x00\x80\x41\xDB"
-     "\x00\x80"
-     "\xAF\xE5\x07\x00\xAE\xF9\x07\x00\xAD\xF1\x07\x00\xAC\x61\x06\x00\xA8\x91"
-     "\x06\x00"
-     "\xA9\x91\x06\x00\xAA\x91\x06\x00\xAB\xAD\x06\x00\xAC\xB9\x06\x00\xAD\xA5"
-     "\x06\x00"
-     "\xAE\xAD\x06\x00\xAF\xA5\x06\x00\x45\xDB\x00\x80\x49\xDB\x00\x80\x4D\xDB"
-     "\x00\x80"
-     "\x51\xDB\x00\x80\x55\xDB\x00\x80\x59\xDB\x00\x80\x5D\xDB\x00\x80\x61\xDB"
-     "\x00\x80"
-     "\xB8\x65\x01\x00\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x7D\x01\x00\xBC\x65"
-     "\x01\x00"
-     "\xBD\x6D\x01\x00\xBE\x65\x01\x00\xBF\xD9\x01\x00\xB0\xDD\x06\x00\xB1\xA5"
-     "\x06\x00"
-     "\xB2\xAD\x06\x00\xB3\xA5\x06\x00\xB4\xA1\x06\x00\xB5\xA9\x06\x00\xB6\x99"
-     "\x06\x00"
-     "\xB7\x99\x06\x00\xB3\x19\x06\x00\x65\xDB\x00\x80\x69\xDB\x00\x80\x6D\xDB"
-     "\x00\x80"
-     "\x71\xDB\x00\x80\xB6\x25\x06\x00\xB5\x31\x06\x00\x75\xDB\x00\x80\xBB\x61"
-     "\x06\x00"
-     "\xBA\x19\x06\x00\x79\xDB\x00\x80\x7D\xDB\x00\x80\xBF\x6D\x06\x00\xBE\x65"
-     "\x06\x00"
-     "\xBD\x71\x06\x00\xBC\x79\x06\x00\x81\xDB\x00\x80\xA3\x5D\x06\x00\x85\xDB"
-     "\x00\x80"
-     "\x89\xDB\x00\x80\xA6\x61\x06\x00\x8D\xDB\x00\x80\x84\x98\x01\x00\xA5\x75"
-     "\x06\x00"
-     "\xAA\x5D\x06\x00\xAB\x25\x06\x00\xBE\xA4\x01\x00\x95\xDB\x00\x80\xAE\x21"
-     "\x06\x00"
-     "\xAF\x29\x06\x00\xAC\x3D\x06\x00\xAD\x35\x06\x00\xA9\xC5\x02\x00\xA8\xB1"
-     "\x02\x00"
-     "\xAB\xC5\x02\x00\xAA\xC5\x02\x00\xAD\xCD\x02\x00\xAC\xC5\x02\x00\xAF\xF5"
-     "\x02\x00"
-     "\xAE\xF5\x02\x00\x99\xDB\x00\x80\x9D\xDB\x00\x80\xA1\xDB\x00\x80\xA5\xDB"
-     "\x00\x80"
-     "\xA9\xDB\x00\x80\xAD\xDB\x00\x80\xB1\xDB\x00\x80\xB5\xDB\x00\x80\xB9\xC9"
-     "\x03\x00"
-     "\xB8\xC1\x03\x00\xBB\xD9\x03\x00\xBA\xC1\x03\x00\xBD\xF9\x03\x00\xBC\xC9"
-     "\x03\x00"
-     "\xBF\x99\x03\x00\xBE\xF1\x03\x00\xB1\x45\x03\x00\xB0\x45\x03\x00\xB3\x45"
-     "\x03\x00"
-     "\xB2\x45\x03\x00\xB5\x45\x03\x00\xB4\x45\x03\x00\xB7\x45\x03\x00\xB6\x45"
-     "\x03\x00"
-     "\x80\x49\x03\x00\x81\x49\x03\x00\x82\x5D\x03\x00\xB3\x45\x02\x00\xBE\xDC"
-     "\x0C\x00"
-     "\xB5\x45\x02\x00\xB6\x45\x02\x00\xB9\xDB\x00\x80\x86\x00\x0C\x00\x87\xE4"
-     "\x03\x00"
-     "\xBA\x89\x03\x00\xBB\x89\x03\x00\xBC\x99\x03\x00\xBD\x99\x03\x00\xBE\x89"
-     "\x03\x00"
-     "\xBF\x89\x03\x00\xA3\x09\x02\x00\xBD\xDB\x00\x80\xC1\xDB\x00\x80\xC5\xDB"
-     "\x00\x80"
-     "\xC9\xDB\x00\x80\xA6\x09\x02\x00\xA5\x09\x02\x00\xCD\xDB\x00\x80\xAB\xC5"
-     "\x03\x00"
-     "\xAA\xC5\x03\x00\xD1\xDB\x00\x80\xD5\xDB\x00\x80\xAF\xC5\x03\x00\xAE\xC5"
-     "\x03\x00"
-     "\xAD\xD5\x03\x00\xAC\xD5\x03\x00\xD9\xDB\x00\x80\xDD\xDB\x00\x80\xE1\xDB"
-     "\x00\x80"
-     "\x41\xDA\x00\x80\xEF\x00\x00\x00\xE5\xDB\x00\x80\xE9\xDB\x00\x80\xED\xDB"
-     "\x00\x80"
-     "\xE3\xE8\x00\x00\xF1\xDB\x00\x80\xE1\x8C\x01\x00\xF5\xDB\x00\x80\xF9\xDB"
-     "\x00\x80"
-     "\xFD\xDB\x00\x80\x05\xDC\x00\x80\x09\xDC\x00\x80\x80\x6D\x00\x00\x81\x75"
-     "\x00\x00"
-     "\x82\x7D\x00\x00\x84\x40\x02\x00\x86\xF0\x0C\x00\x87\x78\x0D\x00\x0D\xDC"
-     "\x00\x80"
-     "\x11\xDC\x00\x80\x15\xDC\x00\x80\x19\xDC\x00\x80\x1D\xDC\x00\x80\x21\xDC"
-     "\x00\x80"
-     "\x25\xDC\x00\x80\x29\xDC\x00\x80\x2D\xDC\x00\x80\x31\xDC\x00\x80\x35\xDC"
-     "\x00\x80"
-     "\x39\xDC\x00\x80\x3D\xDC\x00\x80\x41\xDC\x00\x80\x45\xDC\x00\x80\x49\xDC"
-     "\x00\x80"
-     "\x4D\xDC\x00\x80\xEF\xCC\x01\x00\x84\xEC\x0C\x00\xE1\x30\x06\x00\x51\xDC"
-     "\x00\x80"
-     "\xE3\x18\x01\x00\x55\xDC\x00\x80\x59\xDC\x00\x80\x5D\xDC\x00\x80\x61\xDC"
-     "\x00\x80"
-     "\xB3\xE5\x01\x00\x65\xDC\x00\x80\x84\x84\x0F\x00\x69\xDC\x00\x80\x6D\xDC"
-     "\x00\x80"
-     "\xB6\xE5\x01\x00\xB5\xF5\x01\x00\x75\xDC\x00\x80\xBB\x7D\x01\x00\xBA\xD9"
-     "\x01\x00"
-     "\xBE\xA0\x0C\x00\x79\xDC\x00\x80\xBF\x21\x01\x00\xBE\x39\x01\x00\xBD\x31"
-     "\x01\x00"
-     "\xBC\x39\x01\x00\xA8\xED\x0D\x00\xA9\x25\x0E\x00\xAA\x2D\x0E\x00\xAB\x25"
-     "\x0E\x00"
-     "\xAC\x3D\x0E\x00\xAD\x25\x0E\x00\xAE\x2D\x0E\x00\xAF\x25\x0E\x00\x01\xDC"
-     "\x00\x80"
-     "\x82\xF5\x0F\x00\x81\xE5\x0F\x00\x80\xE9\x0F\x00\x71\xDC\x00\x80\x7D\xDC"
-     "\x00\x80"
-     "\x86\x98\x00\x00\x87\x0C\x03\x00\xB8\xAD\x0E\x00\xB9\x45\x0F\x00\xBA\x4D"
-     "\x0F\x00"
-     "\xBB\x45\x0F\x00\xBC\x45\x0F\x00\xBD\x4D\x0F\x00\xBE\x45\x0F\x00\xBF\x79"
-     "\x0F\x00"
-     "\xB0\x5D\x0E\x00\xB1\xF9\x0E\x00\xB2\xAD\x0E\x00\xB3\xA5\x0E\x00\xB4\xBD"
-     "\x0E\x00"
-     "\xB5\xA5\x0E\x00\xB6\xA5\x0E\x00\xB7\x95\x0E\x00\x81\xDC\x00\x80\xEF\xEC"
-     "\x0C\x00"
-     "\x85\xDC\x00\x80\x89\xDC\x00\x80\x8D\xDC\x00\x80\x91\xDC\x00\x80\x95\xDC"
-     "\x00\x80"
-     "\x99\xDC\x00\x80\xBE\xE8\x00\x00\x9D\xDC\x00\x80\xA1\xDC\x00\x80\xA5\xDC"
-     "\x00\x80"
-     "\xA9\xDC\x00\x80\xE3\xE8\x0D\x00\xAD\xDC\x00\x80\xE1\x04\x0C\x00\xA3\xE5"
-     "\x0E\x00"
-     "\xB1\xDC\x00\x80\xB5\xDC\x00\x80\xB9\xDC\x00\x80\xBD\xDC\x00\x80\xA6\xE5"
-     "\x0E\x00"
-     "\xA5\xF5\x0E\x00\xC1\xDC\x00\x80\xAB\x7D\x0E\x00\xAA\xD9\x0E\x00\xC5\xDC"
-     "\x00\x80"
-     "\xC9\xDC\x00\x80\xAF\x21\x0E\x00\xAE\x39\x0E\x00\xAD\x31\x0E\x00\xAC\x39"
-     "\x0E\x00"
-     "\xA8\x39\x0E\x00\xA9\x39\x0E\x00\xAA\x51\x0E\x00\xAB\x51\x0E\x00\xAC\x71"
-     "\x0E\x00"
-     "\xAD\x71\x0E\x00\xAE\x71\x0E\x00\xAF\x71\x0E\x00\xCD\xDC\x00\x80\xD1\xDC"
-     "\x00\x80"
-     "\xD5\xDC\x00\x80\xD9\xDC\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xDD\xDC\x00\x80\xB8\xD1\x0E\x00\xB9\xD1\x0E\x00\xBA\xD1\x0E\x00\xBB\xE5"
-     "\x0E\x00"
-     "\xBC\xE1\x0E\x00\xBD\xE1\x0E\x00\xBE\xE1\x0E\x00\xBF\xE1\x0E\x00\xB0\x11"
-     "\x0E\x00"
-     "\xB1\x11\x0E\x00\xB2\x11\x0E\x00\xB3\x11\x0E\x00\xB4\xF1\x0E\x00\xB5\xF1"
-     "\x0E\x00"
-     "\xB6\xF1\x0E\x00\xB7\xF1\x0E\x00\xB3\xD9\x0E\x00\xE5\xDC\x00\x80\x86\x28"
-     "\x00\x00"
-     "\x87\xB8\x00\x00\xE9\xDC\x00\x80\xB6\xF1\x0E\x00\xB5\xF9\x0E\x00\xED\xDC"
-     "\x00\x80"
-     "\xBB\xD5\x0E\x00\xBA\xD5\x0E\x00\xF1\xDC\x00\x80\xF5\xDC\x00\x80\xBF\x35"
-     "\x0E\x00"
-     "\xBE\x35\x0E\x00\xBD\xC5\x0E\x00\xBC\xC5\x0E\x00\xF9\xDC\x00\x80\xA3\x9D"
-     "\x0E\x00"
-     "\xFD\xDC\x00\x80\x01\xDD\x00\x80\xA6\xB5\x0E\x00\x05\xDD\x00\x80\x09\xDD"
-     "\x00\x80"
-     "\xA5\xBD\x0E\x00\xAA\x91\x0E\x00\xAB\x91\x0E\x00\x0D\xDD\x00\x80\x11\xDD"
-     "\x00\x80"
-     "\xAE\x71\x0E\x00\xAF\x71\x0E\x00\xAC\x81\x0E\x00\xAD\x81\x0E\x00\xA8\xDD"
-     "\x0D\x00"
-     "\xA9\xE9\x0D\x00\xAA\x3D\x02\x00\xAB\x8D\x02\x00\xAC\x99\x02\x00\xAD\x99"
-     "\x02\x00"
-     "\xAE\x89\x02\x00\xAF\x89\x02\x00\xBE\xAC\x04\x00\x15\xDD\x00\x80\x19\xDD"
-     "\x00\x80"
-     "\x84\x20\x03\x00\x1D\xDD\x00\x80\x21\xDD\x00\x80\x25\xDD\x00\x80\x29\xDD"
-     "\x00\x80"
-     "\xB8\x89\x02\x00\xB9\x89\x02\x00\xBA\x99\x02\x00\xBB\x91\x02\x00\xBC\xB9"
-     "\x02\x00"
-     "\xBD\xB9\x02\x00\xBE\x79\x03\x00\xBF\x79\x03\x00\xB0\xF9\x02\x00\xB1\xF9"
-     "\x02\x00"
-     "\xB2\xCD\x02\x00\xB3\xC5\x02\x00\xB4\xDD\x02\x00\xB5\xC5\x02\x00\xB6\xC1"
-     "\x02\x00"
-     "\xB7\xB9\x02\x00\xB3\xB5\x02\x00\x2D\xDD\x00\x80\x31\xDD\x00\x80\x35\xDD"
-     "\x00\x80"
-     "\x39\xDD\x00\x80\xB6\x19\x02\x00\xB5\x11\x02\x00\x3D\xDD\x00\x80\xBB\x3D"
-     "\x02\x00"
-     "\xBA\x3D\x02\x00\x41\xDD\x00\x80\x45\xDD\x00\x80\xBF\x01\x02\x00\xBE\x19"
-     "\x02\x00"
-     "\xBD\x11\x02\x00\xBC\x19\x02\x00\x49\xDD\x00\x80\xA3\xF1\x02\x00\x4D\xDD"
-     "\x00\x80"
-     "\x55\xDD\x00\x80\xA6\x5D\x02\x00\x59\xDD\x00\x80\x5D\xDD\x00\x80\xA5\x55"
-     "\x02\x00"
-     "\xAA\x79\x02\x00\xAB\x79\x02\x00\x86\x48\x05\x00\x87\xAC\x04\x00\xAE\x5D"
-     "\x02\x00"
-     "\xAF\x45\x02\x00\xAC\x5D\x02\x00\xAD\x55\x02\x00\xA8\x85\x02\x00\xA9\x95"
-     "\x02\x00"
-     "\xAA\x95\x02\x00\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xD5\x02\x00\xAE\xD1"
-     "\x02\x00"
-     "\xAF\xD1\x02\x00\x61\xDD\x00\x80\x65\xDD\x00\x80\x69\xDD\x00\x80\x6D\xDD"
-     "\x00\x80"
-     "\x81\xF1\x01\x00\x80\x99\x01\x00\x91\xDB\x00\x80\x82\xF5\x01\x00\xB8\x79"
-     "\x01\x00"
-     "\xB9\x79\x01\x00\xBA\xCD\x01\x00\xBB\xC5\x01\x00\xBC\xDD\x01\x00\xBD\xC5"
-     "\x01\x00"
-     "\xBE\xC5\x01\x00\xBF\xF5\x01\x00\xB0\xB5\x02\x00\xB1\xBD\x02\x00\xB2\x81"
-     "\x02\x00"
-     "\xB3\x81\x02\x00\xB4\x55\x01\x00\xB5\x5D\x01\x00\xB6\x49\x01\x00\xB7\x49"
-     "\x01\x00"
-     "\x71\xDD\x00\x80\x75\xDD\x00\x80\x79\xDD\x00\x80\xEF\xD4\x01\x00\x84\x10"
-     "\x05\x00"
-     "\x7D\xDD\x00\x80\x81\xDD\x00\x80\xEF\x8C\x0E\x00\xBE\xEC\x05\x00\xE1\xEC"
-     "\x0E\x00"
-     "\x85\xDD\x00\x80\xE3\x1C\x0E\x00\x89\xDD\x00\x80\xE1\x94\x01\x00\x8D\xDD"
-     "\x00\x80"
-     "\xE3\x64\x0E\x00\xB3\x5D\x02\x00\x91\xDD\x00\x80\x95\xDD\x00\x80\x99\xDD"
-     "\x00\x80"
-     "\x9D\xDD\x00\x80\xB6\x15\x02\x00\xB5\x75\x02\x00\xA1\xDD\x00\x80\xBB\x39"
-     "\x02\x00"
-     "\xBA\x31\x02\x00\xA5\xDD\x00\x80\xA9\xDD\x00\x80\xBF\xD9\x01\x00\xBE\xD1"
-     "\x01\x00"
-     "\xBD\x15\x02\x00\xBC\x15\x02\x00\xA3\x9D\x05\x00\x51\xDD\x00\x80\xAD\xDD"
-     "\x00\x80"
-     "\xB1\xDD\x00\x80\xB5\xDD\x00\x80\xA6\xD5\x05\x00\xA5\xB5\x05\x00\xB9\xDD"
-     "\x00\x80"
-     "\xAB\xF9\x05\x00\xAA\xF1\x05\x00\xBD\xDD\x00\x80\xC1\xDD\x00\x80\xAF\x19"
-     "\x06\x00"
-     "\xAE\x11\x06\x00\xAD\xD5\x05\x00\xAC\xD5\x05\x00\x80\x51\x00\x00\x81\x59"
-     "\x00\x00"
-     "\x82\x61\x00\x00\xB3\x95\x06\x00\xC5\xDD\x00\x80\xB5\x71\x07\x00\xB6\x71"
-     "\x07\x00"
-     "\xC9\xDD\x00\x80\x86\x40\x03\x00\x87\x54\x03\x00\xBA\xED\x07\x00\xBB\xE5"
-     "\x07\x00"
-     "\xBC\xE5\x07\x00\xBD\xED\x07\x00\xBE\xD1\x07\x00\xBF\xCD\x07\x00\xCD\xDD"
-     "\x00\x80"
-     "\xD1\xDD\x00\x80\xD5\xDD\x00\x80\xD9\xDD\x00\x80\xDD\xDD\x00\x80\xE1\xDD"
-     "\x00\x80"
-     "\xEF\x40\x04\x00\xE5\xDD\x00\x80\xE1\x30\x07\x00\xE9\xDD\x00\x80\xE3\x94"
-     "\x04\x00"
-     "\xED\xDD\x00\x80\xF1\xDD\x00\x80\xF5\xDD\x00\x80\xF9\xDD\x00\x80\xFD\xDD"
-     "\x00\x80"
-     "\xA3\x19\x06\x00\x01\xDE\x00\x80\x05\xDE\x00\x80\x09\xDE\x00\x80\x0D\xDE"
-     "\x00\x80"
-     "\xA6\xFD\x07\x00\xA5\xFD\x07\x00\x11\xDE\x00\x80\xAB\x69\x07\x00\xAA\x61"
-     "\x07\x00"
-     "\x15\xDE\x00\x80\x19\xDE\x00\x80\xAF\x41\x07\x00\xAE\x5D\x07\x00\xAD\x61"
-     "\x07\x00"
-     "\xAC\x69\x07\x00\xA8\xCD\x07\x00\xA9\xD1\x07\x00\xAA\xD1\x07\x00\xAB\x2D"
-     "\x06\x00"
-     "\xAC\x35\x06\x00\xAD\x3D\x06\x00\xAE\x35\x06\x00\xAF\x9D\x06\x00\x1D\xDE"
-     "\x00\x80"
-     "\x21\xDE\x00\x80\x25\xDE\x00\x80\x29\xDE\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x2D\xDE\x00\x80\xB8\x89\x06\x00\xB9\x89\x06\x00\xBA\x99"
-     "\x06\x00"
-     "\xBB\x91\x06\x00\xBC\xB9\x06\x00\xBD\xB9\x06\x00\xBE\x51\x01\x00\xBF\x51"
-     "\x01\x00"
-     "\xB0\xE5\x06\x00\xB1\xED\x06\x00\xB2\xFD\x06\x00\xB3\xF5\x06\x00\xB4\xD9"
-     "\x06\x00"
-     "\xB5\xC5\x06\x00\xB6\xC1\x06\x00\xB7\xB9\x06\x00\xA8\xD1\x01\x00\xA9\xD9"
-     "\x01\x00"
-     "\xAA\x09\x01\x00\xAB\x09\x01\x00\xAC\x19\x01\x00\xAD\x19\x01\x00\xAE\x09"
-     "\x01\x00"
-     "\xAF\x09\x01\x00\x84\x60\x01\x00\xBE\x7C\x01\x00\x87\xA8\x00\x00\x86\x8C"
-     "\x01\x00"
-     "\x35\xDE\x00\x80\x39\xDE\x00\x80\x3D\xDE\x00\x80\x41\xDE\x00\x80\xB8\x09"
-     "\x01\x00"
-     "\xB9\x09\x01\x00\xBA\x19\x01\x00\xBB\x11\x01\x00\xBC\x39\x01\x00\xBD\x39"
-     "\x01\x00"
-     "\xBE\xF9\x00\x00\xBF\xF9\x00\x00\xB0\x7D\x01\x00\xB1\x41\x01\x00\xB2\x45"
-     "\x01\x00"
-     "\xB3\x5D\x01\x00\xB4\x45\x01\x00\xB5\x4D\x01\x00\xB6\x45\x01\x00\xB7\x39"
-     "\x01\x00"
-     "\x45\xDE\x00\x80\x49\xDE\x00\x80\x4D\xDE\x00\x80\xB3\x8D\x02\x00\x51\xDE"
-     "\x00\x80"
-     "\xB5\x9D\x02\x00\xB6\x95\x02\x00\x55\xDE\x00\x80\x59\xDE\x00\x80\x5D\xDE"
-     "\x00\x80"
-     "\xBA\xB5\x02\x00\xBB\x89\x02\x00\xBC\x9D\x02\x00\xBD\x85\x02\x00\xBE\x8D"
-     "\x02\x00"
-     "\xBF\x85\x02\x00\xA6\xCD\x02\x00\x61\xDE\x00\x80\x65\xDE\x00\x80\xA5\xC5"
-     "\x02\x00"
-     "\x69\xDE\x00\x80\xA3\xD5\x02\x00\x6D\xDE\x00\x80\x71\xDE\x00\x80\xAE\xD5"
-     "\x02\x00"
-     "\xAF\xDD\x02\x00\xAC\xC5\x02\x00\xAD\xDD\x02\x00\xAA\xED\x02\x00\xAB\xD1"
-     "\x02\x00"
-     "\x84\xF4\x03\x00\x75\xDE\x00\x80\xA8\x31\x03\x00\xA9\x31\x03\x00\xAA\x31"
-     "\x03\x00"
-     "\xAB\x31\x03\x00\xAC\x91\x00\x00\xAD\x91\x00\x00\xAE\x91\x00\x00\xAF\x8D"
-     "\x00\x00"
-     "\x79\xDE\x00\x80\x7D\xDE\x00\x80\x81\xDE\x00\x80\x85\xDE\x00\x80\x89\xDE"
-     "\x00\x80"
-     "\x8D\xDE\x00\x80\x91\xDE\x00\x80\x95\xDE\x00\x80\xB8\xBD\x00\x00\xB9\x65"
-     "\x00\x00"
-     "\xBA\x6D\x00\x00\xBB\x65\x00\x00\xBC\x7D\x00\x00\xBD\x65\x00\x00\xBE\x6D"
-     "\x00\x00"
-     "\xBF\x65\x00\x00\xB0\xFD\x00\x00\xB1\xC5\x00\x00\xB2\xCD\x00\x00\xB3\xA9"
-     "\x00\x00"
-     "\xB4\xB9\x00\x00\xB5\xA5\x00\x00\xB6\xA1\x00\x00\xB7\xA1\x00\x00\x80\xBD"
-     "\x01\x00"
-     "\x81\x09\x00\x00\x82\x19\x00\x00\x99\xDE\x00\x80\x9D\xDE\x00\x80\xBE\x58"
-     "\x02\x00"
-     "\x87\x14\x1D\x00\x86\x9C\x1D\x00\x84\x6C\x1D\x00\xE1\xDC\x00\x80\xA5\xDE"
-     "\x00\x80"
-     "\xA9\xDE\x00\x80\xBE\xBC\x1C\x00\xAD\xDE\x00\x80\xB1\xDE\x00\x80\xB5\xDE"
-     "\x00\x80"
-     "\xB3\xF9\x02\x00\xB9\xDE\x00\x80\xBD\xDE\x00\x80\xC1\xDE\x00\x80\xC5\xDE"
-     "\x00\x80"
-     "\xB6\x51\x01\x00\xB5\x59\x01\x00\xBE\xDC\x1F\x00\xBB\x41\x01\x00\xBA\x79"
-     "\x01\x00"
-     "\xC9\xDE\x00\x80\xCD\xDE\x00\x80\xBF\x21\x01\x00\xBE\x3D\x01\x00\xBD\x3D"
-     "\x01\x00"
-     "\xBC\x59\x01\x00\xE1\x70\x07\x00\xD1\xDE\x00\x80\xE3\x08\x06\x00\xD5\xDE"
-     "\x00\x80"
-     "\xEF\xCC\x00\x00\xD9\xDE\x00\x80\xDD\xDE\x00\x80\xE1\xDE\x00\x80\xE3\x10"
-     "\x00\x00"
-     "\xE5\xDE\x00\x80\xE1\xD0\x01\x00\xE9\xDE\x00\x80\x86\x90\x1C\x00\x87\xF4"
-     "\x1C\x00"
-     "\xEF\xE0\x06\x00\xED\xDE\x00\x80\xA3\x79\x02\x00\xF1\xDE\x00\x80\xF5\xDE"
-     "\x00\x80"
-     "\xF9\xDE\x00\x80\xFD\xDE\x00\x80\xA6\xD1\x01\x00\xA5\xD9\x01\x00\x01\xDF"
-     "\x00\x80"
-     "\xAB\xC1\x01\x00\xAA\xF9\x01\x00\x05\xDF\x00\x80\x09\xDF\x00\x80\xAF\xA1"
-     "\x01\x00"
-     "\xAE\xBD\x01\x00\xAD\xBD\x01\x00\xAC\xD9\x01\x00\xA1\xDE\x00\x80\x82\x15"
-     "\x00\x00"
-     "\x81\xE5\x1F\x00\x80\xE5\x1F\x00\x0D\xDF\x00\x80\x11\xDF\x00\x80\x15\xDF"
-     "\x00\x80"
-     "\x19\xDF\x00\x80\xA8\x09\x1F\x00\xA9\x09\x1F\x00\xAA\x1D\x1F\x00\xAB\x15"
-     "\x1F\x00"
-     "\xAC\x0D\x1F\x00\xAD\x71\x1F\x00\xAE\x71\x1F\x00\xAF\x71\x1F\x00\xB0\x11"
-     "\x1F\x00"
-     "\xB1\x2D\x1F\x00\xB2\x25\x1F\x00\xB3\xC9\x1F\x00\xB4\xDD\x1F\x00\xB5\xC1"
-     "\x1F\x00"
-     "\xB6\xC1\x1F\x00\xB7\xC1\x1F\x00\xB8\xC5\x1F\x00\xB9\xC9\x1F\x00\xBA\xD5"
-     "\x1F\x00"
-     "\xBB\xA9\x1F\x00\xBC\xB9\x1F\x00\xBD\xB9\x1F\x00\xBE\xA9\x1F\x00\xBF\xA1"
-     "\x1F\x00"
-     "\xB3\xB5\x1F\x00\x1D\xDF\x00\x80\x21\xDF\x00\x80\x25\xDF\x00\x80\x29\xDF"
-     "\x00\x80"
-     "\xB6\xD1\x1F\x00\xB5\xA5\x1F\x00\x2D\xDF\x00\x80\xBB\xC9\x1F\x00\xBA\xF5"
-     "\x1F\x00"
-     "\x31\xDF\x00\x80\x35\xDF\x00\x80\xBF\x25\x1F\x00\xBE\x39\x1F\x00\xBD\x3D"
-     "\x1F\x00"
-     "\xBC\xD1\x1F\x00\x39\xDF\x00\x80\x3D\xDF\x00\x80\x41\xDF\x00\x80\x45\xDF"
-     "\x00\x80"
-     "\x49\xDF\x00\x80\xE1\x60\x1F\x00\x4D\xDF\x00\x80\xE3\xB4\x1C\x00\x51\xDF"
-     "\x00\x80"
-     "\x55\xDF\x00\x80\x59\xDF\x00\x80\xEF\x00\x1D\x00\x5D\xDF\x00\x80\x61\xDF"
-     "\x00\x80"
-     "\x65\xDF\x00\x80\x69\xDF\x00\x80\xA3\x35\x1E\x00\x6D\xDF\x00\x80\x71\xDF"
-     "\x00\x80"
-     "\x75\xDF\x00\x80\x79\xDF\x00\x80\xA6\x51\x1E\x00\xA5\x25\x1E\x00\x7D\xDF"
-     "\x00\x80"
-     "\xAB\x49\x1E\x00\xAA\x75\x1E\x00\x84\xA8\x02\x00\x81\xDF\x00\x80\xAF\xA5"
-     "\x1E\x00"
-     "\xAE\xB9\x1E\x00\xAD\xBD\x1E\x00\xAC\x51\x1E\x00\x80\x4D\x00\x00\x81\x55"
-     "\x00\x00"
-     "\x82\x55\x00\x00\xB3\xC9\x01\x00\x85\xDF\x00\x80\xB5\xD9\x01\x00\xB6\xC9"
-     "\x01\x00"
-     "\x89\xDF\x00\x80\x86\xA0\x00\x00\x87\x04\x01\x00\xBA\xC5\x01\x00\xBB\xAD"
-     "\x01\x00"
-     "\xBC\xB5\x01\x00\xBD\xBD\x01\x00\xBE\xB5\x01\x00\xBF\xAD\x01\x00\xA8\x99"
-     "\x01\x00"
-     "\xA9\x99\x01\x00\xAA\x0D\x01\x00\xAB\x05\x01\x00\xAC\x1D\x01\x00\xAD\x05"
-     "\x01\x00"
-     "\xAE\x05\x01\x00\xAF\x35\x01\x00\x8D\xDF\x00\x80\x91\xDF\x00\x80\x95\xDF"
-     "\x00\x80"
-     "\x99\xDF\x00\x80\x9D\xDF\x00\x80\xA1\xDF\x00\x80\xA5\xDF\x00\x80\xA9\xDF"
-     "\x00\x80"
-     "\xB8\x25\x01\x00\xB9\x2D\x01\x00\xBA\x39\x01\x00\xBB\x39\x01\x00\xBC\x29"
-     "\x01\x00"
-     "\xBD\x29\x01\x00\xBE\xDD\x00\x00\xBF\xD5\x00\x00\xB0\x4D\x01\x00\xB1\x25"
-     "\x01\x00"
-     "\xB2\x2D\x01\x00\xB3\x25\x01\x00\xB4\x3D\x01\x00\xB5\x25\x01\x00\xB6\x21"
-     "\x01\x00"
-     "\xB7\x1D\x01\x00\xAD\xDF\x00\x80\xB1\xDF\x00\x80\xB5\xDF\x00\x80\xA3\x89"
-     "\x02\x00"
-     "\xB9\xDF\x00\x80\xA5\x99\x02\x00\xA6\x89\x02\x00\xBD\xDF\x00\x80\xC1\xDF"
-     "\x00\x80"
-     "\xC5\xDF\x00\x80\xAA\x85\x02\x00\xAB\xED\x02\x00\xAC\xF5\x02\x00\xAD\xFD"
-     "\x02\x00"
-     "\xAE\xF5\x02\x00\xAF\xED\x02\x00\xC9\xDF\x00\x80\xCD\xDF\x00\x80\xD1\xDF"
-     "\x00\x80"
-     "\x84\x40\x02\x00\xD5\xDF\x00\x80\xD9\xDF\x00\x80\xDD\xDF\x00\x80\xE1\xDF"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\xE5\xDF\x00\x80\xE9\xDF"
-     "\x00\x80"
-     "\xED\xDF\x00\x80\x87\xB4\x03\x00\x86\xDC\x04\x00\xBE\xCC\x03\x00\xF5\xDF"
-     "\x00\x80"
-     "\xF9\xDF\x00\x80\xFD\xDF\x00\x80\xEF\xE8\x02\x00\x01\xE0\x00\x80\x05\xE0"
-     "\x00\x80"
-     "\x09\xE0\x00\x80\xE3\xFC\x02\x00\x0D\xE0\x00\x80\xE1\xD0\x01\x00\x11\xE0"
-     "\x00\x80"
-     "\x15\xE0\x00\x80\x19\xE0\x00\x80\x1D\xE0\x00\x80\x21\xE0\x00\x80\xB3\x61"
-     "\x03\x00"
-     "\x25\xE0\x00\x80\x29\xE0\x00\x80\x2D\xE0\x00\x80\x31\xE0\x00\x80\xB6\x79"
-     "\x03\x00"
-     "\xB5\x71\x03\x00\x35\xE0\x00\x80\xBB\x5D\x03\x00\xBA\x5D\x03\x00\x39\xE0"
-     "\x00\x80"
-     "\x3D\xE0\x00\x80\xBF\xE1\x00\x00\xBE\xFD\x00\x00\xBD\xFD\x00\x00\xBC\xFD"
-     "\x00\x00"
-     "\xB0\xB9\x02\x00\xB1\xB9\x02\x00\xB2\x09\x01\x00\xB3\x09\x01\x00\xB4\x19"
-     "\x01\x00"
-     "\xB5\x05\x01\x00\xB6\x05\x01\x00\xB7\x3D\x01\x00\xB8\x05\x01\x00\xB9\x65"
-     "\x01\x00"
-     "\xBA\x6D\x01\x00\xBB\x65\x01\x00\xBC\x61\x01\x00\xBD\x61\x01\x00\xBE\x61"
-     "\x01\x00"
-     "\xBF\x61\x01\x00\x85\x5C\x07\x00\x41\xE0\x00\x80\x45\xE0\x00\x80\x49\xE0"
-     "\x00\x80"
-     "\x31\xDE\x00\x80\x4D\xE0\x00\x80\x51\xE0\x00\x80\x55\xE0\x00\x80\xA8\x31"
-     "\x02\x00"
-     "\xA9\x39\x02\x00\xAA\xC9\x02\x00\xAB\xC9\x02\x00\xAC\xD9\x02\x00\xAD\xD9"
-     "\x02\x00"
-     "\xAE\xC9\x02\x00\xAF\xC9\x02\x00\x84\xCC\x05\x00\xE1\x80\x1E\x00\x59\xE0"
-     "\x00\x80"
-     "\xE3\xBC\x1E\x00\xE1\x38\x1E\x00\x5D\xE0\x00\x80\xE3\xA0\x00\x00\xBE\x10"
-     "\x04\x00"
-     "\x65\xE0\x00\x80\x69\xE0\x00\x80\xEF\x0C\x1E\x00\x6D\xE0\x00\x80\x71\xE0"
-     "\x00\x80"
-     "\x75\xE0\x00\x80\x79\xE0\x00\x80\xEF\x74\x1E\x00\xA3\x61\x02\x00\x82\x51"
-     "\x00\x00"
-     "\x81\x41\x00\x00\x80\x91\x00\x00\x7D\xE0\x00\x80\xA6\x79\x02\x00\xA5\x71"
-     "\x02\x00"
-     "\x81\xE0\x00\x80\xAB\x5D\x02\x00\xAA\x5D\x02\x00\x86\xC8\x04\x00\x87\x3C"
-     "\x05\x00"
-     "\xAF\xE1\x01\x00\xAE\xFD\x01\x00\xAD\xFD\x01\x00\xAC\xFD\x01\x00\xA8\x85"
-     "\x06\x00"
-     "\xA9\x8D\x06\x00\xAA\x85\x06\x00\xAB\x99\x06\x00\xAC\x89\x06\x00\xAD\x89"
-     "\x06\x00"
-     "\xAE\xBD\x06\x00\xAF\xB1\x06\x00\x61\xE0\x00\x80\x85\xE0\x00\x80\x89\xE0"
-     "\x00\x80"
-     "\x8D\xE0\x00\x80\x91\xE0\x00\x80\x95\xE0\x00\x80\x99\xE0\x00\x80\x9D\xE0"
-     "\x00\x80"
-     "\xB8\x9D\x06\x00\xB9\xAD\x06\x00\xBA\xA5\x06\x00\xBB\x09\x07\x00\xBC\x19"
-     "\x07\x00"
-     "\xBD\x19\x07\x00\xBE\x0D\x07\x00\xBF\x05\x07\x00\xB0\xD1\x06\x00\xB1\xD1"
-     "\x06\x00"
-     "\xB2\xD1\x06\x00\xB3\xD1\x06\x00\xB4\xB5\x06\x00\xB5\xBD\x06\x00\xB6\xB5"
-     "\x06\x00"
-     "\xB7\xAD\x06\x00\xB3\x0D\x06\x00\xA1\xE0\x00\x80\xA5\xE0\x00\x80\xA9\xE0"
-     "\x00\x80"
-     "\xAD\xE0\x00\x80\xB6\x09\x06\x00\xB5\x01\x06\x00\xB1\xE0\x00\x80\xBB\x15"
-     "\x06\x00"
-     "\xBA\x15\x06\x00\xB5\xE0\x00\x80\xB9\xE0\x00\x80\xBF\x79\x06\x00\xBE\x71"
-     "\x06\x00"
-     "\xBD\x05\x06\x00\xBC\x05\x06\x00\xBD\xE0\x00\x80\xE1\xA0\x04\x00\xC1\xE0"
-     "\x00\x80"
-     "\xE3\x5C\x05\x00\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\xC5\xE0"
-     "\x00\x80"
-     "\xC9\xE0\x00\x80\xCD\xE0\x00\x80\x84\x60\x03\x00\xBE\x6C\x00\x00\xBF\xF0"
-     "\x01\x00"
-     "\x85\x90\x00\x00\xD1\xE0\x00\x80\xEF\x98\x05\x00\xA3\x8D\x07\x00\x84\x08"
-     "\x00\x00"
-     "\x86\x00\x0C\x00\x87\x8C\x00\x00\xD5\xE0\x00\x80\xA6\x89\x07\x00\xA5\x81"
-     "\x07\x00"
-     "\xD9\xE0\x00\x80\xAB\x95\x07\x00\xAA\x95\x07\x00\xDD\xE0\x00\x80\xE1\xE0"
-     "\x00\x80"
-     "\xAF\xF9\x07\x00\xAE\xF1\x07\x00\xAD\x85\x07\x00\xAC\x85\x07\x00\xE5\xE0"
-     "\x00\x80"
-     "\xB3\xE9\x06\x00\xE9\xE0\x00\x80\xED\xE0\x00\x80\xB6\xE9\x06\x00\xF1\xE0"
-     "\x00\x80"
-     "\xF5\xE0\x00\x80\xB5\xE9\x06\x00\xBA\x45\x01\x00\xBB\x4D\x01\x00\xF9\xE0"
-     "\x00\x80"
-     "\xFD\xE0\x00\x80\xBE\x45\x01\x00\xBF\x4D\x01\x00\xBC\x55\x01\x00\xBD\x4D"
-     "\x01\x00"
-     "\xA8\x21\x06\x00\xA9\x21\x06\x00\xAA\x25\x06\x00\xAB\x3D\x06\x00\xAC\x25"
-     "\x06\x00"
-     "\xAD\x2D\x06\x00\xAE\x49\x06\x00\xAF\x41\x06\x00\x01\xE1\x00\x80\x05\xE1"
-     "\x00\x80"
-     "\x09\xE1\x00\x80\x0D\xE1\x00\x80\x11\xE1\x00\x80\x15\xE1\x00\x80\x19\xE1"
-     "\x00\x80"
-     "\x1D\xE1\x00\x80\xB8\x49\x01\x00\xB9\x49\x01\x00\xBA\x59\x01\x00\xBB\x51"
-     "\x01\x00"
-     "\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x19\x01\x00\xBF\x15\x01\x00\xB0\xC5"
-     "\x01\x00"
-     "\xB1\xCD\x01\x00\xB2\xC5\x01\x00\xB3\xDD\x01\x00\xB4\xC5\x01\x00\xB5\xCD"
-     "\x01\x00"
-     "\xB6\xC5\x01\x00\xB7\x79\x01\x00\x21\xE1\x00\x80\x25\xE1\x00\x80\x29\xE1"
-     "\x00\x80"
-     "\xA3\xA1\x05\x00\x2D\xE1\x00\x80\xA5\xA1\x05\x00\xA6\xA1\x05\x00\x31\xE1"
-     "\x00\x80"
-     "\x8F\x21\xEA\x03\x35\xE1\x00\x80\xAA\x0D\x02\x00\xAB\x05\x02\x00\xAC\x1D"
-     "\x02\x00"
-     "\xAD\x05\x02\x00\xAE\x0D\x02\x00\xAF\x05\x02\x00\x97\xED\x12\x00\x96\x65"
-     "\x12\x00"
-     "\x95\x45\x11\x00\x94\x9D\x16\x00\x93\x71\x16\x00\x92\x55\x15\x00\x91\x79"
-     "\xEB\x03"
-     "\x90\x59\xEA\x03\x9F\x61\x06\x00\x9E\x35\x05\x00\x9D\x45\x1A\x00\x9C\x69"
-     "\x1A\x00"
-     "\x9B\x55\x19\x00\x9A\x45\x1E\x00\x99\x59\x1E\x00\x98\x45\x1D\x00\xE1\x60"
-     "\x00\x00"
-     "\x39\xE1\x00\x80\xE3\x4C\x3E\x00\x3D\xE1\x00\x80\xA3\xB1\x02\x00\xA2\xD5"
-     "\x01\x00"
-     "\xA1\xB5\x07\x00\xA0\x89\x06\x00\xB1\x01\x38\x00\xB0\x09\x3E\x00\xB3\x95"
-     "\x3A\x00"
-     "\xB2\x8D\x3A\x00\xB5\xB5\x26\x00\xB4\x01\x24\x00\xEF\x68\x3A\x00\xBE\x30"
-     "\x0C\x00"
-     "\xA9\xC9\x36\x00\xA8\xC1\x36\x00\xAB\x01\x30\x00\xAA\xE1\x37\x00\xAD\xCD"
-     "\x33\x00"
-     "\xAC\xF5\x32\x00\xAF\xF9\x3E\x00\xAE\x01\x3C\x00\xA1\x19\x02\x00\x45\xE1"
-     "\x00\x80"
-     "\xA3\x6D\x0E\x00\xA2\x6D\x0E\x00\xA5\xF5\x0A\x00\xA4\x01\x08\x00\xA7\x85"
-     "\x0B\x00"
-     "\xA6\x99\x0A\x00\x86\x00\x0D\x00\x87\x44\x0D\x00\x82\x09\xEB\x03\x83\x09"
-     "\xEB\x03"
-     "\x84\x31\xEA\x03\x85\x15\x14\x00\x86\x39\x17\x00\x87\xCD\x17\x00\x84\xA0"
-     "\x0D\x00"
-     "\x49\xE1\x00\x80\x8A\x25\x10\x00\x8B\x0D\x13\x00\x8C\x9D\x13\x00\x8D\x0D"
-     "\x1C\x00"
-     "\x8E\x19\x1F\x00\x8F\x0D\x1F\x00\xF1\xDF\x00\x80\xEF\x00\x03\x00\x92\x6D"
-     "\x18\x00"
-     "\x93\x49\x1B\x00\x94\x7D\x1B\x00\x95\x05\x04\x00\x96\x59\x07\x00\x97\x49"
-     "\x07\x00"
-     "\x4D\xE1\x00\x80\x51\xE1\x00\x80\x9A\x45\x06\x00\x9B\x2D\x00\x00\x9C\x51"
-     "\x03\x00"
-     "\xE3\x60\x00\x00\x55\xE1\x00\x80\xE1\x6C\x00\x00\x80\xA5\x01\x00\x81\x01"
-     "\x01\x00"
-     "\x82\x05\x01\x00\xBE\x00\x0C\x00\x59\xE1\x00\x80\x5D\xE1\x00\x80\x61\xE1"
-     "\x00\x80"
-     "\xEF\x88\x01\x00\x65\xE1\x00\x80\xE1\x54\x06\x00\x69\xE1\x00\x80\xE3\x54"
-     "\x01\x00"
-     "\x6D\xE1\x00\x80\x71\xE1\x00\x80\x75\xE1\x00\x80\x79\xE1\x00\x80\xB3\xE9"
-     "\x02\x00"
-     "\x7D\xE1\x00\x80\x81\xE1\x00\x80\x85\xE1\x00\x80\x89\xE1\x00\x80\xB6\x9D"
-     "\x02\x00"
-     "\xB5\x99\x02\x00\x8D\xE1\x00\x80\xBB\x89\x02\x00\xBA\xBD\x02\x00\x91\xE1"
-     "\x00\x80"
-     "\x95\xE1\x00\x80\xBF\x59\x02\x00\xBE\x51\x02\x00\xBD\x59\x02\x00\xBC\x91"
-     "\x02\x00"
-     "\xA3\x29\x0D\x00\x99\xE1\x00\x80\x9D\xE1\x00\x80\xA1\xE1\x00\x80\xA5\xE1"
-     "\x00\x80"
-     "\xA6\x5D\x0D\x00\xA5\x59\x0D\x00\xA9\xE1\x00\x80\xAB\x49\x0D\x00\xAA\x7D"
-     "\x0D\x00"
-     "\xAD\xE1\x00\x80\xB5\xE1\x00\x80\xAF\x99\x0D\x00\xAE\x91\x0D\x00\xAD\x99"
-     "\x0D\x00"
-     "\xAC\x51\x0D\x00\x80\x51\x00\x00\x81\x59\x00\x00\x82\x61\x00\x00\xB3\x2D"
-     "\x0F\x00"
-     "\xB9\xE1\x00\x80\xB5\x2D\x0F\x00\xB6\xC9\x0F\x00\xBD\xE1\x00\x80\x86\x40"
-     "\x03\x00"
-     "\x87\x08\x03\x00\xBA\xC9\x0F\x00\xBB\xC5\x0F\x00\xBC\xC1\x0F\x00\xBD\xC1"
-     "\x0F\x00"
-     "\xBE\xC1\x0F\x00\xBF\xC1\x0F\x00\x41\xE1\x00\x80\xB1\xE1\x00\x80\xC1\xE1"
-     "\x00\x80"
-     "\xC5\xE1\x00\x80\xC9\xE1\x00\x80\xCD\xE1\x00\x80\xD1\xE1\x00\x80\xD5\xE1"
-     "\x00\x80"
-     "\xA8\x45\x0E\x00\xA9\x81\x0F\x00\xAA\xC9\x0F\x00\xAB\xC9\x0F\x00\xAC\xC9"
-     "\x0F\x00"
-     "\xAD\x25\x0F\x00\xAE\x2D\x0F\x00\xAF\x25\x0F\x00\xB0\x61\x0F\x00\xB1\x6D"
-     "\x0F\x00"
-     "\xB2\x79\x0F\x00\xB3\x79\x0F\x00\xB4\x69\x0F\x00\xB5\x69\x0F\x00\xB6\x1D"
-     "\x0F\x00"
-     "\xB7\x15\x0F\x00\xB8\x2D\x0F\x00\xB9\x35\x0F\x00\xBA\x35\x0F\x00\xBB\x05"
-     "\x0F\x00"
-     "\xBC\x1D\x0F\x00\xBD\xF1\x00\x00\xBE\xF1\x00\x00\xBF\xF1\x00\x00\xA3\x61"
-     "\x0E\x00"
-     "\xD9\xE1\x00\x80\x84\xC4\x01\x00\xDD\xE1\x00\x80\xE1\xE1\x00\x80\xA6\x85"
-     "\x0E\x00"
-     "\xA5\x61\x0E\x00\xE5\xE1\x00\x80\xAB\x89\x0E\x00\xAA\x85\x0E\x00\xE9\xE1"
-     "\x00\x80"
-     "\xED\xE1\x00\x80\xAF\x8D\x0E\x00\xAE\x8D\x0E\x00\xAD\x8D\x0E\x00\xAC\x8D"
-     "\x0E\x00"
-     "\xF1\xE1\x00\x80\xF5\xE1\x00\x80\xF9\xE1\x00\x80\xFD\xE1\x00\x80\x01\xE2"
-     "\x00\x80"
-     "\x05\xE2\x00\x80\x09\xE2\x00\x80\x0D\xE2\x00\x80\x11\xE2\x00\x80\x82\x1D"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x80\x1D\x00\x00\x15\xE2\x00\x80\x19\xE2\x00\x80\x1D\xE2"
-     "\x00\x80"
-     "\xBE\xB4\x01\x00\xA8\xAD\x01\x00\xA9\xD5\x01\x00\xAA\xD5\x01\x00\xAB\x05"
-     "\x01\x00"
-     "\xAC\x1D\x01\x00\xAD\x05\x01\x00\xAE\x0D\x01\x00\xAF\x05\x01\x00\x86\x80"
-     "\x01\x00"
-     "\x87\x18\x01\x00\x25\xE2\x00\x80\x29\xE2\x00\x80\x2D\xE2\x00\x80\x31\xE2"
-     "\x00\x80"
-     "\x35\xE2\x00\x80\x39\xE2\x00\x80\xB8\x85\x00\x00\xB9\x8D\x00\x00\xBA\x85"
-     "\x00\x00"
-     "\xBB\x9D\x00\x00\xBC\x85\x00\x00\xBD\x8D\x00\x00\xBE\x85\x00\x00\xBF\xBD"
-     "\x00\x00"
-     "\xB0\x7D\x01\x00\xB1\xE1\x00\x00\xB2\xE5\x00\x00\xB3\xFD\x00\x00\xB4\xED"
-     "\x00\x00"
-     "\xB5\xDD\x00\x00\xB6\xD5\x00\x00\xB7\xBD\x00\x00\xB3\x5D\x02\x00\x3D\xE2"
-     "\x00\x80"
-     "\x41\xE2\x00\x80\x45\xE2\x00\x80\x49\xE2\x00\x80\xB6\x85\x02\x00\xB5\x95"
-     "\x02\x00"
-     "\x4D\xE2\x00\x80\xBB\x25\x03\x00\xBA\xB9\x02\x00\x51\xE2\x00\x80\x55\xE2"
-     "\x00\x80"
-     "\xBF\x19\x03\x00\xBE\x29\x03\x00\xBD\x29\x03\x00\xBC\x31\x03\x00\xBE\xCC"
-     "\x04\x00"
-     "\xA3\x19\x02\x00\x59\xE2\x00\x80\x5D\xE2\x00\x80\xA6\xC1\x02\x00\x61\xE2"
-     "\x00\x80"
-     "\x65\xE2\x00\x80\xA5\xD1\x02\x00\xAA\xFD\x02\x00\xAB\x61\x03\x00\x69\xE2"
-     "\x00\x80"
-     "\x6D\xE2\x00\x80\xAE\x6D\x03\x00\xAF\x5D\x03\x00\xAC\x75\x03\x00\xAD\x6D"
-     "\x03\x00"
-     "\xA8\x29\x03\x00\xA9\x29\x03\x00\xAA\x39\x03\x00\xAB\x39\x03\x00\xAC\x29"
-     "\x03\x00"
-     "\xAD\x29\x03\x00\xAE\x9D\x00\x00\xAF\x95\x00\x00\x71\xE2\x00\x80\x75\xE2"
-     "\x00\x80"
-     "\x79\xE2\x00\x80\x7D\xE2\x00\x80\x81\xE2\x00\x80\x82\xA9\x01\x00\x81\xAD"
-     "\x01\x00"
-     "\x80\xAD\x01\x00\xB8\x99\x00\x00\xB9\xAD\x00\x00\xBA\xA5\x00\x00\xBB\x6D"
-     "\x00\x00"
-     "\xBC\x75\x00\x00\xBD\x7D\x00\x00\xBE\x75\x00\x00\xBF\x6D\x00\x00\xB0\xED"
-     "\x00\x00"
-     "\xB1\xF5\x00\x00\xB2\xF5\x00\x00\xB3\xC5\x00\x00\xB4\xDD\x00\x00\xB5\xBD"
-     "\x00\x00"
-     "\xB6\xB5\x00\x00\xB7\xA9\x00\x00\xE1\x78\x01\x00\xE1\x1C\x0E\x00\xE3\x10"
-     "\x00\x00"
-     "\xE3\x3C\x0E\x00\x85\xE2\x00\x80\x89\xE2\x00\x80\xBE\x14\x04\x00\x8D\xE2"
-     "\x00\x80"
-     "\x84\xAC\x02\x00\x95\xE2\x00\x80\x87\x78\x05\x00\x86\x0C\x05\x00\x99\xE2"
-     "\x00\x80"
-     "\x9D\xE2\x00\x80\xEF\xBC\x00\x00\xEF\x48\x0E\x00\xB3\xF1\x02\x00\xA1\xE2"
-     "\x00\x80"
-     "\xA5\xE2\x00\x80\xA9\xE2\x00\x80\xAD\xE2\x00\x80\xB6\xE9\x02\x00\xB5\xE1"
-     "\x02\x00"
-     "\xB1\xE2\x00\x80\xBB\x71\x01\x00\xBA\x69\x01\x00\xB5\xE2\x00\x80\x84\xA0"
-     "\x04\x00"
-     "\xBF\x39\x01\x00\xBE\x59\x01\x00\xBD\x51\x01\x00\xBC\x61\x01\x00\xB9\xE2"
-     "\x00\x80"
-     "\x84\x8C\x04\x00\xBD\xE2\x00\x80\x84\x00\x38\x00\xC1\xE2\x00\x80\xC5\xE2"
-     "\x00\x80"
-     "\xC9\xE2\x00\x80\xCD\xE2\x00\x80\xAA\x89\x0E\x00\xAB\x89\x0E\x00\xA8\xB9"
-     "\x0E\x00"
-     "\xA9\xB1\x0E\x00\xAE\xFD\x0E\x00\xAF\xE1\x0E\x00\xAC\xF9\x0E\x00\xAD\xF5"
-     "\x0E\x00"
-     "\xB2\xAD\x0E\x00\xB3\x65\x0E\x00\xB0\x91\x0E\x00\xB1\xA5\x0E\x00\xB6\x7D"
-     "\x0E\x00"
-     "\xB7\x65\x0E\x00\xB4\x7D\x0E\x00\xB5\x75\x0E\x00\xBA\x5D\x0E\x00\xBB\xE5"
-     "\x0D\x00"
-     "\xB8\x5D\x0E\x00\xB9\x55\x0E\x00\xBE\xE1\x0D\x00\xBF\xE9\x0D\x00\xBC\xFD"
-     "\x0D\x00"
-     "\xBD\xF5\x0D\x00\xA3\xB1\x05\x00\x91\xE2\x00\x80\xD1\xE2\x00\x80\xD5\xE2"
-     "\x00\x80"
-     "\xD9\xE2\x00\x80\xA6\xA9\x05\x00\xA5\xA1\x05\x00\xDD\xE2\x00\x80\xAB\x31"
-     "\x06\x00"
-     "\xAA\x29\x06\x00\xE1\xE2\x00\x80\xE5\xE2\x00\x80\xAF\x79\x06\x00\xAE\x19"
-     "\x06\x00"
-     "\xAD\x11\x06\x00\xAC\x21\x06\x00\xE9\xE2\x00\x80\xED\xE2\x00\x80\xF1\xE2"
-     "\x00\x80"
-     "\xF5\xE2\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39\x00\x00\xF9\xE2"
-     "\x00\x80"
-     "\xFD\xE2\x00\x80\x01\xE3\x00\x80\x86\xC8\x00\x00\x87\x0C\x03\x00\x05\xE3"
-     "\x00\x80"
-     "\x09\xE3\x00\x80\x0D\xE3\x00\x80\x11\xE3\x00\x80\xA8\xA5\x07\x00\xA9\xB5"
-     "\x07\x00"
-     "\xAA\xBD\x07\x00\xAB\xC9\x07\x00\xAC\xD9\x07\x00\xAD\xD9\x07\x00\xAE\xC9"
-     "\x07\x00"
-     "\xAF\xC1\x07\x00\xBE\xA0\x00\x00\x15\xE3\x00\x80\x19\xE3\x00\x80\x1D\xE3"
-     "\x00\x80"
-     "\x21\xE3\x00\x80\x25\xE3\x00\x80\x29\xE3\x00\x80\x2D\xE3\x00\x80\xB8\xCD"
-     "\x00\x00"
-     "\xB9\xD5\x00\x00\xBA\xD5\x00\x00\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\x95"
-     "\x00\x00"
-     "\xBE\x9D\x00\x00\xBF\x95\x00\x00\xB0\x89\x07\x00\xB1\x65\x07\x00\xB2\x61"
-     "\x07\x00"
-     "\xB3\x7D\x07\x00\xB4\x65\x07\x00\xB5\x6D\x07\x00\xB6\x65\x07\x00\xB7\xF5"
-     "\x00\x00"
-     "\xB3\x35\x06\x00\x31\xE3\x00\x80\x35\xE3\x00\x80\x39\xE3\x00\x80\x3D\xE3"
-     "\x00\x80"
-     "\xB6\x59\x06\x00\xB5\x51\x06\x00\x41\xE3\x00\x80\xBB\xA1\x06\x00\xBA\x4D"
-     "\x06\x00"
-     "\x45\xE3\x00\x80\x49\xE3\x00\x80\xBF\xA9\x06\x00\xBE\xA1\x06\x00\xBD\xA9"
-     "\x06\x00"
-     "\xBC\xB5\x06\x00\x4D\xE3\x00\x80\x51\xE3\x00\x80\xEF\xF0\x05\x00\x55\xE3"
-     "\x00\x80"
-     "\x59\xE3\x00\x80\x5D\xE3\x00\x80\x61\xE3\x00\x80\x65\xE3\x00\x80\x80\x3D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x1D\x00\x00\x69\xE3\x00\x80\xE1\xC8\x06\x00\x6D\xE3"
-     "\x00\x80"
-     "\xE3\x48\x04\x00\x71\xE3\x00\x80\xA3\xB5\x06\x00\x75\xE3\x00\x80\x86\x28"
-     "\x00\x00"
-     "\x87\x40\x01\x00\x79\xE3\x00\x80\xA6\xD9\x06\x00\xA5\xD1\x06\x00\x7D\xE3"
-     "\x00\x80"
-     "\xAB\x21\x06\x00\xAA\xCD\x06\x00\x81\xE3\x00\x80\x85\xE3\x00\x80\xAF\x29"
-     "\x06\x00"
-     "\xAE\x21\x06\x00\xAD\x29\x06\x00\xAC\x35\x06\x00\x89\xE3\x00\x80\xB3\xBD"
-     "\x01\x00"
-     "\x8D\xE3\x00\x80\x91\xE3\x00\x80\xB6\x79\x01\x00\x95\xE3\x00\x80\x99\xE3"
-     "\x00\x80"
-     "\xB5\x79\x01\x00\xBA\x55\x01\x00\xBB\x5D\x01\x00\x9D\xE3\x00\x80\xA1\xE3"
-     "\x00\x80"
-     "\xBE\xF9\x00\x00\xBF\xF9\x00\x00\xBC\x45\x01\x00\xBD\xF9\x00\x00\xA8\x71"
-     "\x02\x00"
-     "\xA9\x71\x02\x00\xAA\x71\x02\x00\xAB\x71\x02\x00\xAC\xB5\x02\x00\xAD\xBD"
-     "\x02\x00"
-     "\xAE\xB5\x02\x00\xAF\xAD\x02\x00\xBE\xAC\x3C\x00\xA5\xE3\x00\x80\xA9\xE3"
-     "\x00\x80"
-     "\xAD\xE3\x00\x80\xB1\xE3\x00\x80\xB5\xE3\x00\x80\xB9\xE3\x00\x80\xBD\xE3"
-     "\x00\x80"
-     "\xB8\x69\x03\x00\xB9\x69\x03\x00\xBA\x09\x03\x00\xBB\x09\x03\x00\xBC\x1D"
-     "\x03\x00"
-     "\xBD\x05\x03\x00\xBE\x0D\x03\x00\xBF\x05\x03\x00\xB0\xD5\x02\x00\xB1\xDD"
-     "\x02\x00"
-     "\xB2\xD5\x02\x00\xB3\x69\x03\x00\xB4\x79\x03\x00\xB5\x79\x03\x00\xB6\x69"
-     "\x03\x00"
-     "\xB7\x61\x03\x00\xC1\xE3\x00\x80\xC5\xE3\x00\x80\xC9\xE3\x00\x80\xA3\xF5"
-     "\x02\x00"
-     "\xCD\xE3\x00\x80\xA5\x31\x02\x00\xA6\x31\x02\x00\xD1\xE3\x00\x80\xD5\xE3"
-     "\x00\x80"
-     "\xD9\xE3\x00\x80\xAA\x1D\x02\x00\xAB\x15\x02\x00\xAC\x0D\x02\x00\xAD\xB1"
-     "\x03\x00"
-     "\xAE\xB1\x03\x00\xAF\xB1\x03\x00\xEF\x18\x02\x00\x82\x15\x00\x00\x81\x6D"
-     "\x00\x00"
-     "\x80\x6D\x00\x00\xDD\xE3\x00\x80\xE5\xE3\x00\x80\x86\xF8\x3C\x00\x87\x18"
-     "\x03\x00"
-     "\xE9\xE3\x00\x80\xED\xE3\x00\x80\xF1\xE3\x00\x80\xF5\xE3\x00\x80\xE3\x6C"
-     "\x07\x00"
-     "\x21\xE2\x00\x80\xE1\x68\x01\x00\xF9\xE3\x00\x80\xA8\x85\x02\x00\xA9\x95"
-     "\x02\x00"
-     "\xAA\x95\x02\x00\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xD5\x02\x00\xAE\xD1"
-     "\x02\x00"
-     "\xAF\xD1\x02\x00\xFD\xE3\x00\x80\x01\xE4\x00\x80\x05\xE4\x00\x80\x09\xE4"
-     "\x00\x80"
-     "\x0D\xE4\x00\x80\x11\xE4\x00\x80\x15\xE4\x00\x80\x19\xE4\x00\x80\xB8\x75"
-     "\x01\x00"
-     "\xB9\x7D\x01\x00\xBA\x75\x01\x00\xBB\xCD\x01\x00\xBC\xD5\x01\x00\xBD\xDD"
-     "\x01\x00"
-     "\xBE\xC9\x01\x00\xBF\xC1\x01\x00\xB0\xB5\x02\x00\xB1\xBD\x02\x00\xB2\x81"
-     "\x02\x00"
-     "\xB3\x81\x02\x00\xB4\x55\x01\x00\xB5\x5D\x01\x00\xB6\x55\x01\x00\xB7\x4D"
-     "\x01\x00"
-     "\xE1\xB4\x06\x00\x1D\xE4\x00\x80\xE3\xF4\x06\x00\x21\xE4\x00\x80\x84\x18"
-     "\x3D\x00"
-     "\x25\xE4\x00\x80\x29\xE4\x00\x80\x2D\xE4\x00\x80\x31\xE4\x00\x80\x35\xE4"
-     "\x00\x80"
-     "\x39\xE4\x00\x80\x3D\xE4\x00\x80\x41\xE4\x00\x80\x45\xE4\x00\x80\xEF\x58"
-     "\x06\x00"
-     "\x49\xE4\x00\x80\x81\x7D\x00\x00\x80\x71\x00\x00\x4D\xE4\x00\x80\x82\x05"
-     "\x00\x00"
-     "\x55\xE4\x00\x80\x59\xE4\x00\x80\xEF\x80\x01\x00\xBE\x54\x3C\x00\xE1\x90"
-     "\x01\x00"
-     "\x5D\xE4\x00\x80\xE3\x7C\x06\x00\x61\xE4\x00\x80\x65\xE4\x00\x80\x69\xE4"
-     "\x00\x80"
-     "\x86\xD8\x3C\x00\x87\xF4\x3C\x00\xA3\x9D\x3D\x00\xE1\xE3\x00\x80\x51\xE4"
-     "\x00\x80"
-     "\x6D\xE4\x00\x80\x71\xE4\x00\x80\xA6\xD5\x3D\x00\xA5\xB5\x3D\x00\x75\xE4"
-     "\x00\x80"
-     "\xAB\xF9\x3D\x00\xAA\xF1\x3D\x00\x79\xE4\x00\x80\x7D\xE4\x00\x80\xAF\x19"
-     "\x3E\x00"
-     "\xAE\x11\x3E\x00\xAD\xD5\x3D\x00\xAC\xD5\x3D\x00\x81\xE4\x00\x80\xB3\xA1"
-     "\x3E\x00"
-     "\x85\xE4\x00\x80\x89\xE4\x00\x80\xB6\xAD\x3E\x00\x8D\xE4\x00\x80\x91\xE4"
-     "\x00\x80"
-     "\xB5\xB1\x3E\x00\xBA\x49\x3F\x00\xBB\x49\x3F\x00\x95\xE4\x00\x80\x99\xE4"
-     "\x00\x80"
-     "\xBE\x49\x3F\x00\xBF\x49\x3F\x00\xBC\x49\x3F\x00\xBD\x49\x3F\x00\xA8\x55"
-     "\x3E\x00"
-     "\xA9\x65\x3E\x00\xAA\x6D\x3E\x00\xAB\x65\x3E\x00\xAC\x7D\x3E\x00\xAD\x69"
-     "\x3E\x00"
-     "\xAE\xB9\x3F\x00\xAF\xB9\x3F\x00\x9D\xE4\x00\x80\xA1\xE4\x00\x80\xA5\xE4"
-     "\x00\x80"
-     "\xA9\xE4\x00\x80\xAD\xE4\x00\x80\xB1\xE4\x00\x80\xB5\xE4\x00\x80\xB9\xE4"
-     "\x00\x80"
-     "\xB8\x55\x3F\x00\xB9\x5D\x3F\x00\xBA\x55\x3F\x00\xBB\x6D\x3F\x00\xBC\x75"
-     "\x3F\x00"
-     "\xBD\x7D\x3F\x00\xBE\x75\x3F\x00\xBF\x6D\x3F\x00\xB0\xC9\x3F\x00\xB1\xC9"
-     "\x3F\x00"
-     "\xB2\xD9\x3F\x00\xB3\xD9\x3F\x00\xB4\xC9\x3F\x00\xB5\xC9\x3F\x00\xB6\x7D"
-     "\x3F\x00"
-     "\xB7\x71\x3F\x00\x82\x15\x00\x00\xA3\xE1\x3F\x00\x80\xB1\x01\x00\x81\xB1"
-     "\x01\x00"
-     "\xA6\xED\x3F\x00\xBD\xE4\x00\x80\xBE\xD0\x01\x00\xA5\xF1\x3F\x00\xAA\x09"
-     "\x3E\x00"
-     "\xAB\x09\x3E\x00\x84\xE4\x01\x00\xC1\xE4\x00\x80\xAE\x09\x3E\x00\xAF\x09"
-     "\x3E\x00"
-     "\xAC\x09\x3E\x00\xAD\x09\x3E\x00\xB3\xDD\x3C\x00\xC5\xE4\x00\x80\x86\xE8"
-     "\x00\x00"
-     "\x87\xCC\x01\x00\xC9\xE4\x00\x80\xB6\x95\x3C\x00\xB5\xF5\x3C\x00\xCD\xE4"
-     "\x00\x80"
-     "\xBB\xB9\x3C\x00\xBA\xB1\x3C\x00\xD1\xE4\x00\x80\xD5\xE4\x00\x80\xBF\x59"
-     "\x3F\x00"
-     "\xBE\x51\x3F\x00\xBD\x95\x3C\x00\xBC\x95\x3C\x00\xA8\x51\x3E\x00\xA9\x51"
-     "\x3E\x00"
-     "\xAA\x6D\x3E\x00\xAB\x61\x3E\x00\xAC\x61\x3E\x00\xAD\x61\x3E\x00\xAE\x95"
-     "\x01\x00"
-     "\xAF\x8D\x01\x00\x84\xA0\x01\x00\xD9\xE4\x00\x80\xDD\xE4\x00\x80\xE1\xE4"
-     "\x00\x80"
-     "\xE5\xE4\x00\x80\xE9\xE4\x00\x80\xED\xE4\x00\x80\xF1\xE4\x00\x80\xB8\xA9"
-     "\x01\x00"
-     "\xB9\xA9\x01\x00\xBA\x69\x01\x00\xBB\x69\x01\x00\xBC\x79\x01\x00\xBD\x79"
-     "\x01\x00"
-     "\xBE\x69\x01\x00\xBF\x69\x01\x00\xB0\xFD\x01\x00\xB1\xCD\x01\x00\xB2\xC5"
-     "\x01\x00"
-     "\xB3\xAD\x01\x00\xB4\xB9\x01\x00\xB5\xB9\x01\x00\xB6\xAD\x01\x00\xB7\xA5"
-     "\x01\x00"
-     "\xB3\xE5\x3D\x00\xF5\xE4\x00\x80\xF9\xE4\x00\x80\xFD\xE4\x00\x80\x01\xE5"
-     "\x00\x80"
-     "\xB6\xE1\x3D\x00\xB5\xE9\x3D\x00\x05\xE5\x00\x80\xBB\x09\x02\x00\xBA\x39"
-     "\x02\x00"
-     "\x09\xE5\x00\x80\x0D\xE5\x00\x80\xBF\x7D\x02\x00\xBE\x7D\x02\x00\xBD\x79"
-     "\x02\x00"
-     "\xBC\x11\x02\x00\x11\xE5\x00\x80\xA3\xA1\x3D\x00\x15\xE5\x00\x80\x19\xE5"
-     "\x00\x80"
-     "\xA6\xA5\x3D\x00\x1D\xE5\x00\x80\x21\xE5\x00\x80\xA5\xAD\x3D\x00\xAA\x7D"
-     "\x02\x00"
-     "\xAB\x4D\x02\x00\x25\xE5\x00\x80\x29\xE5\x00\x80\xAE\x39\x02\x00\xAF\x39"
-     "\x02\x00"
-     "\xAC\x55\x02\x00\xAD\x3D\x02\x00\x80\xE9\x00\x00\x81\xE9\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\xBE\xC0\x03\x00\xEF\xE4\x02\x00\x2D\xE5\x00\x80\x87\x54\x03\x00\x86\x3C"
-     "\x04\x00"
-     "\xE3\x10\x01\x00\x35\xE5\x00\x80\xE1\xF8\x01\x00\x39\xE5\x00\x80\x3D\xE5"
-     "\x00\x80"
-     "\x41\xE5\x00\x80\x45\xE5\x00\x80\x49\xE5\x00\x80\x4D\xE5\x00\x80\x51\xE5"
-     "\x00\x80"
-     "\x55\xE5\x00\x80\xB3\x91\x03\x00\x59\xE5\x00\x80\xB5\xB9\x03\x00\xB6\x7D"
-     "\x03\x00"
-     "\x5D\xE5\x00\x80\x61\xE5\x00\x80\x65\xE5\x00\x80\xBA\x59\x03\x00\xBB\x59"
-     "\x03\x00"
-     "\xBC\x49\x03\x00\xBD\x49\x03\x00\xBE\xFD\x00\x00\xBF\xF5\x00\x00\xA8\x45"
-     "\x02\x00"
-     "\xA9\x55\x02\x00\xAA\x55\x02\x00\xAB\x65\x02\x00\xAC\x7D\x02\x00\xAD\xB1"
-     "\x02\x00"
-     "\xAE\xB1\x02\x00\xAF\xB1\x02\x00\x84\x6C\x05\x00\x69\xE5\x00\x80\x6D\xE5"
-     "\x00\x80"
-     "\x71\xE5\x00\x80\x75\xE5\x00\x80\x79\xE5\x00\x80\xBE\x70\x05\x00\x7D\xE5"
-     "\x00\x80"
-     "\xB8\x5D\x01\x00\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x0D\x01\x00\xBC\x19"
-     "\x01\x00"
-     "\xBD\x19\x01\x00\xBE\x0D\x01\x00\xBF\x05\x01\x00\xB0\xD1\x02\x00\xB1\xD1"
-     "\x02\x00"
-     "\xB2\xD1\x02\x00\xB3\xD1\x02\x00\xB4\x75\x01\x00\xB5\x7D\x01\x00\xB6\x75"
-     "\x01\x00"
-     "\xB7\x6D\x01\x00\xE1\x40\x3F\x00\xE3\xBC\x00\x00\xE3\x08\x3E\x00\xE1\x6C"
-     "\x3E\x00"
-     "\x81\xE5\x00\x80\x85\xE5\x00\x80\x89\xE5\x00\x80\x8D\xE5\x00\x80\x91\xE5"
-     "\x00\x80"
-     "\x95\xE5\x00\x80\x99\xE5\x00\x80\x9D\xE5\x00\x80\xBE\x6C\x07\x00\xEF\x54"
-     "\x00\x00"
-     "\xEF\x9C\x3E\x00\xA5\xE5\x00\x80\xA3\x9D\x02\x00\x82\x69\x00\x00\x81\x61"
-     "\x00\x00"
-     "\x80\x69\x00\x00\xA9\xE5\x00\x80\xA6\x71\x02\x00\xA5\xB5\x02\x00\xAD\xE5"
-     "\x00\x80"
-     "\xAB\x55\x02\x00\xAA\x55\x02\x00\x86\xC8\x04\x00\x87\xEC\x04\x00\xAF\xF9"
-     "\x01\x00"
-     "\xAE\xF1\x01\x00\xAD\x45\x02\x00\xAC\x45\x02\x00\xA8\xA5\x06\x00\xA9\xA9"
-     "\x06\x00"
-     "\xAA\xB9\x06\x00\xAB\xB9\x06\x00\xAC\xA9\x06\x00\xAD\xA9\x06\x00\xAE\xD9"
-     "\x06\x00"
-     "\xAF\xD9\x06\x00\xA1\xE5\x00\x80\xB1\xE5\x00\x80\xB5\xE5\x00\x80\xB9\xE5"
-     "\x00\x80"
-     "\xBD\xE5\x00\x80\xC1\xE5\x00\x80\xC5\xE5\x00\x80\xC9\xE5\x00\x80\xB8\x71"
-     "\x07\x00"
-     "\xB9\x71\x07\x00\xBA\x75\x07\x00\xBB\xDD\x07\x00\xBC\xC5\x07\x00\xBD\xCD"
-     "\x07\x00"
-     "\xBE\xC5\x07\x00\xBF\xFD\x07\x00\xB0\xA9\x06\x00\xB1\xB5\x06\x00\xB2\xB5"
-     "\x06\x00"
-     "\xB3\x8D\x06\x00\xB4\x95\x06\x00\xB5\x51\x07\x00\xB6\x51\x07\x00\xB7\x51"
-     "\x07\x00"
-     "\xB3\x31\x06\x00\xCD\xE5\x00\x80\xD1\xE5\x00\x80\xD5\xE5\x00\x80\xD9\xE5"
-     "\x00\x80"
-     "\xB6\x29\x06\x00\xB5\x21\x06\x00\xDD\xE5\x00\x80\xBB\x71\x06\x00\xBA\x6D"
-     "\x06\x00"
-     "\xE1\xE5\x00\x80\xE5\xE5\x00\x80\xBF\x95\x07\x00\xBE\x51\x06\x00\xBD\x59"
-     "\x06\x00"
-     "\xBC\x61\x06\x00\xE9\xE5\x00\x80\xA3\x75\x06\x00\xED\xE5\x00\x80\xF1\xE5"
-     "\x00\x80"
-     "\xA6\x6D\x06\x00\xF5\xE5\x00\x80\xF9\xE5\x00\x80\xA5\x65\x06\x00\xAA\x29"
-     "\x06\x00"
-     "\xAB\x35\x06\x00\xFD\xE5\x00\x80\x01\xE6\x00\x80\xAE\x15\x06\x00\xAF\xD1"
-     "\x07\x00"
-     "\xAC\x25\x06\x00\xAD\x1D\x06\x00\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\x05\xE6\x00\x80\x09\xE6\x00\x80\x0D\xE6\x00\x80\x84\xDC\x01\x00\x11\xE6"
-     "\x00\x80"
-     "\x86\x80\x00\x00\x87\x20\x01\x00\x15\xE6\x00\x80\x19\xE6\x00\x80\x1D\xE6"
-     "\x00\x80"
-     "\x21\xE6\x00\x80\x25\xE6\x00\x80\x29\xE6\x00\x80\x2D\xE6\x00\x80\xE3\x74"
-     "\x04\x00"
-     "\x31\xE6\x00\x80\xE1\xC8\x05\x00\x35\xE6\x00\x80\x39\xE6\x00\x80\x3D\xE6"
-     "\x00\x80"
-     "\x41\xE6\x00\x80\x45\xE6\x00\x80\x49\xE6\x00\x80\x4D\xE6\x00\x80\x51\xE6"
-     "\x00\x80"
-     "\x55\xE6\x00\x80\xEF\xB4\x04\x00\x59\xE6\x00\x80\x5D\xE6\x00\x80\xA8\x3D"
-     "\x06\x00"
-     "\xA9\x55\x06\x00\xAA\x55\x06\x00\xAB\xA9\x01\x00\xAC\xB9\x01\x00\xAD\xB9"
-     "\x01\x00"
-     "\xAE\xA9\x01\x00\xAF\xA9\x01\x00\x84\xAC\x01\x00\x61\xE6\x00\x80\x65\xE6"
-     "\x00\x80"
-     "\x69\xE6\x00\x80\x6D\xE6\x00\x80\x71\xE6\x00\x80\x75\xE6\x00\x80\x79\xE6"
-     "\x00\x80"
-     "\xB8\x6D\x01\x00\xB9\x05\x01\x00\xBA\x01\x01\x00\xBB\x01\x01\x00\xBC\x05"
-     "\x01\x00"
-     "\xBD\x0D\x01\x00\xBE\x31\x01\x00\xBF\x31\x01\x00\xB0\xD9\x01\x00\xB1\xD9"
-     "\x01\x00"
-     "\xB2\x6D\x01\x00\xB3\x65\x01\x00\xB4\x7D\x01\x00\xB5\x65\x01\x00\xB6\x65"
-     "\x01\x00"
-     "\xB7\x55\x01\x00\x81\xBD\x03\x00\x80\xBD\x03\x00\xB3\xD5\x05\x00\x82\x19"
-     "\x00\x00"
-     "\xB5\x39\x02\x00\x7D\xE6\x00\x80\xBE\x54\x03\x00\xB6\x31\x02\x00\x85\xE6"
-     "\x00\x80"
-     "\x89\xE6\x00\x80\xBB\x15\x02\x00\xBA\x15\x02\x00\xBD\xB9\x02\x00\xBC\xB1"
-     "\x02\x00"
-     "\xBF\xA9\x02\x00\xBE\xB1\x02\x00\x8D\xE6\x00\x80\xA6\x69\x02\x00\xA5\x61"
-     "\x02\x00"
-     "\x84\x00\x0C\x00\xA3\x8D\x05\x00\x91\xE6\x00\x80\x86\xF8\x0C\x00\x87\x7C"
-     "\x03\x00"
-     "\xAF\xF1\x02\x00\xAE\xE9\x02\x00\xAD\xE1\x02\x00\xAC\xE9\x02\x00\xAB\x4D"
-     "\x02\x00"
-     "\xAA\x4D\x02\x00\x95\xE6\x00\x80\x99\xE6\x00\x80\x9D\xE6\x00\x80\xA1\xE6"
-     "\x00\x80"
-     "\xA5\xE6\x00\x80\xA9\xE6\x00\x80\xE3\x20\x01\x00\xAD\xE6\x00\x80\xE1\xA0"
-     "\x01\x00"
-     "\xB1\xE6\x00\x80\xEF\x40\x02\x00\xB5\xE6\x00\x80\xB9\xE6\x00\x80\xBD\xE6"
-     "\x00\x80"
-     "\xC1\xE6\x00\x80\xC5\xE6\x00\x80\xC9\xE6\x00\x80\xCD\xE6\x00\x80\xB3\xF1"
-     "\x03\x00"
-     "\xD1\xE6\x00\x80\x31\xE5\x00\x80\xD5\xE6\x00\x80\xD9\xE6\x00\x80\xB6\xE9"
-     "\x03\x00"
-     "\xB5\xE1\x03\x00\xDD\xE6\x00\x80\xBB\xB5\x03\x00\xBA\xB5\x03\x00\xE1\xE6"
-     "\x00\x80"
-     "\xE5\xE6\x00\x80\xBF\x95\x03\x00\xBE\x95\x03\x00\xBD\xA5\x03\x00\xBC\xA5"
-     "\x03\x00"
-     "\xA8\x29\x02\x00\xA9\x29\x02\x00\xAA\x39\x02\x00\xAB\x39\x02\x00\xAC\x29"
-     "\x02\x00"
-     "\xAD\x29\x02\x00\xAE\x59\x02\x00\xAF\x55\x02\x00\x80\xCD\x01\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\xE9\xE6\x00\x80\xED\xE6\x00\x80\xBE\x7C\x0D\x00\x87\xB4"
-     "\x0D\x00"
-     "\x86\x1C\x0C\x00\xB8\x31\x02\x00\xB9\x3D\x02\x00\xBA\x35\x02\x00\xBB\xE9"
-     "\x02\x00"
-     "\xBC\xF9\x02\x00\xBD\xF9\x02\x00\xBE\xE9\x02\x00\xBF\xE9\x02\x00\xB0\x31"
-     "\x02\x00"
-     "\xB1\x31\x02\x00\xB2\x31\x02\x00\xB3\x31\x02\x00\xB4\x11\x02\x00\xB5\x11"
-     "\x02\x00"
-     "\xB6\x11\x02\x00\xB7\x11\x02\x00\xF5\xE6\x00\x80\xF9\xE6\x00\x80\xFD\xE6"
-     "\x00\x80"
-     "\x01\xE7\x00\x80\x05\xE7\x00\x80\x09\xE7\x00\x80\x0D\xE7\x00\x80\xEF\xD4"
-     "\x06\x00"
-     "\x11\xE7\x00\x80\xE1\x54\x06\x00\x15\xE7\x00\x80\xE3\xA4\x00\x00\xAC\x0C"
-     "\x15\x00"
-     "\x19\xE7\x00\x80\x1D\xE7\x00\x80\x21\xE7\x00\x80\xA3\xF1\x02\x00\x25\xE7"
-     "\x00\x80"
-     "\x29\xE7\x00\x80\x2D\xE7\x00\x80\x31\xE7\x00\x80\xA6\xE9\x02\x00\xA5\xE1"
-     "\x02\x00"
-     "\x35\xE7\x00\x80\xAB\xB5\x02\x00\xAA\xB5\x02\x00\x39\xE7\x00\x80\x3D\xE7"
-     "\x00\x80"
-     "\xAF\x95\x02\x00\xAE\x95\x02\x00\xAD\xA5\x02\x00\xAC\xA5\x02\x00\xA8\x21"
-     "\x0E\x00"
-     "\xA9\x21\x0E\x00\xAA\x49\x0E\x00\xAB\x59\x0E\x00\xAC\x69\x0E\x00\xAD\x69"
-     "\x0E\x00"
-     "\xAE\x99\x0E\x00\xAF\x99\x0E\x00\xF1\xE6\x00\x80\x41\xE7\x00\x80\x45\xE7"
-     "\x00\x80"
-     "\x49\xE7\x00\x80\x4D\xE7\x00\x80\x51\xE7\x00\x80\x55\xE7\x00\x80\x59\xE7"
-     "\x00\x80"
-     "\xB8\xF5\x0E\x00\xB9\xFD\x0E\x00\xBA\xF5\x0E\x00\xBB\x89\x0E\x00\xBC\x9D"
-     "\x0E\x00"
-     "\xBD\x85\x0E\x00\xBE\x85\x0E\x00\xBF\xB5\x0E\x00\xB0\xE9\x0E\x00\xB1\xE9"
-     "\x0E\x00"
-     "\xB2\xFD\x0E\x00\xB3\xF1\x0E\x00\xB4\xD5\x0E\x00\xB5\xDD\x0E\x00\xB6\xD5"
-     "\x0E\x00"
-     "\xB7\xCD\x0E\x00\xB3\xC1\x0E\x00\x82\x15\x00\x00\x81\xB5\x00\x00\x80\xB5"
-     "\x00\x00"
-     "\x5D\xE7\x00\x80\xB6\xE9\x0E\x00\xB5\xE1\x0E\x00\xBE\x10\x00\x00\xBB\x2D"
-     "\x0E\x00"
-     "\xBA\x2D\x0E\x00\x84\x64\x03\x00\x61\xE7\x00\x80\xBF\x19\x0E\x00\xBE\x11"
-     "\x0E\x00"
-     "\xBD\x25\x0E\x00\xBC\x29\x0E\x00\x65\xE7\x00\x80\xA3\x85\x0E\x00\x86\x88"
-     "\x00\x00"
-     "\x87\x6C\x03\x00\xA6\xAD\x0E\x00\x69\xE7\x00\x80\x6D\xE7\x00\x80\xA5\xA5"
-     "\x0E\x00"
-     "\xAA\x69\x0E\x00\xAB\x69\x0E\x00\x71\xE7\x00\x80\x75\xE7\x00\x80\xAE\x55"
-     "\x0E\x00"
-     "\xAF\x5D\x0E\x00\xAC\x6D\x0E\x00\xAD\x61\x0E\x00\xB3\x89\x0E\x00\x79\xE7"
-     "\x00\x80"
-     "\x7D\xE7\x00\x80\x81\xE7\x00\x80\x85\xE7\x00\x80\xB6\x81\x0E\x00\xB5\x89"
-     "\x0E\x00"
-     "\x89\xE7\x00\x80\xBB\x95\x0E\x00\xBA\x8D\x0E\x00\x8D\xE7\x00\x80\x91\xE7"
-     "\x00\x80"
-     "\xBF\xF9\x0E\x00\xBE\xF1\x0E\x00\xBD\x85\x0E\x00\xBC\x85\x0E\x00\x95\xE7"
-     "\x00\x80"
-     "\x99\xE7\x00\x80\x9D\xE7\x00\x80\xA1\xE7\x00\x80\xE3\x0C\x0D\x00\xA5\xE7"
-     "\x00\x80"
-     "\xE1\x18\x0D\x00\xA9\xE7\x00\x80\xEF\xAC\x0C\x00\xAD\xE7\x00\x80\xB1\xE7"
-     "\x00\x80"
-     "\xB5\xE7\x00\x80\xB9\xE7\x00\x80\xBD\xE7\x00\x80\xC1\xE7\x00\x80\xC5\xE7"
-     "\x00\x80"
-     "\xA8\x01\x0E\x00\xA9\x01\x0E\x00\xAA\x01\x0E\x00\xAB\x01\x0E\x00\xAC\x01"
-     "\x0E\x00"
-     "\xAD\x01\x0E\x00\xAE\x01\x0E\x00\xAF\x3D\x0E\x00\x80\xDD\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\xC9\xE7\x00\x80\xCD\xE7\x00\x80\x84\x3C\x01\x00\xBE\x74"
-     "\x00\x00"
-     "\xD5\xE7\x00\x80\xB8\x1D\x0E\x00\xB9\x2D\x0E\x00\xBA\x25\x0E\x00\xBB\xE9"
-     "\x01\x00"
-     "\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\xE9\x01\x00\xBF\xE9\x01\x00\xB0\x49"
-     "\x0E\x00"
-     "\xB1\x51\x0E\x00\xB2\x51\x0E\x00\xB3\x51\x0E\x00\xB4\x35\x0E\x00\xB5\x3D"
-     "\x0E\x00"
-     "\xB6\x35\x0E\x00\xB7\x2D\x0E\x00\xA3\x89\x0D\x00\xD9\xE7\x00\x80\x86\xAC"
-     "\x04\x00"
-     "\x87\x3C\x03\x00\xDD\xE7\x00\x80\xA6\x81\x0D\x00\xA5\x89\x0D\x00\xE1\xE7"
-     "\x00\x80"
-     "\xAB\x95\x0D\x00\xAA\x8D\x0D\x00\xE5\xE7\x00\x80\xE9\xE7\x00\x80\xAF\xF9"
-     "\x0D\x00"
-     "\xAE\xF1\x0D\x00\xAD\x85\x0D\x00\xAC\x85\x0D\x00\xED\xE7\x00\x80\xB3\x9D"
-     "\x02\x00"
-     "\x84\x48\x03\x00\xBE\x40\x04\x00\xB6\x55\x03\x00\xF1\xE7\x00\x80\xF5\xE7"
-     "\x00\x80"
-     "\xB5\xB1\x02\x00\xBA\x71\x03\x00\xBB\x79\x03\x00\xF9\xE7\x00\x80\xFD\xE7"
-     "\x00\x80"
-     "\xBE\x31\x03\x00\xBF\x31\x03\x00\xBC\x51\x03\x00\xBD\x51\x03\x00\xB0\x91"
-     "\x03\x00"
-     "\xB1\x99\x03\x00\xB2\xA1\x03\x00\xB3\xA1\x03\x00\xB4\xD1\x03\x00\xB5\xD1"
-     "\x03\x00"
-     "\xB6\xD1\x03\x00\xB7\xD1\x03\x00\xB8\xF5\x03\x00\xB9\xF9\x03\x00\xBA\xCD"
-     "\x03\x00"
-     "\xBB\xC5\x03\x00\xBC\xDD\x03\x00\xBD\xC5\x03\x00\xBE\xCD\x03\x00\xBF\xC5"
-     "\x03\x00"
-     "\x01\xE8\x00\x80\x05\xE8\x00\x80\x09\xE8\x00\x80\x0D\xE8\x00\x80\x85\x7C"
-     "\x19\x00"
-     "\x11\xE8\x00\x80\x15\xE8\x00\x80\x81\xE6\x00\x80\xA8\x21\x02\x00\xA9\x31"
-     "\x02\x00"
-     "\xAA\x31\x02\x00\xAB\x05\x02\x00\xAC\x1D\x02\x00\xAD\xF1\x03\x00\xAE\xF1"
-     "\x03\x00"
-     "\xAF\xF1\x03\x00\x19\xE8\x00\x80\x1D\xE8\x00\x80\x21\xE8\x00\x80\x25\xE8"
-     "\x00\x80"
-     "\xEF\x50\x00\x00\x29\xE8\x00\x80\x2D\xE8\x00\x80\x31\xE8\x00\x80\xE3\x84"
-     "\x00\x00"
-     "\x35\xE8\x00\x80\xE1\xF8\x01\x00\x39\xE8\x00\x80\x80\x15\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x3D\xE8\x00\x80\xA3\x99\x03\x00\x45\xE8\x00\x80\x86\x68"
-     "\x04\x00"
-     "\x87\x60\x05\x00\x49\xE8\x00\x80\xA6\x51\x02\x00\xA5\xB5\x03\x00\x4D\xE8"
-     "\x00\x80"
-     "\xAB\x7D\x02\x00\xAA\x75\x02\x00\x51\xE8\x00\x80\x55\xE8\x00\x80\xAF\x35"
-     "\x02\x00"
-     "\xAE\x35\x02\x00\xAD\x55\x02\x00\xAC\x55\x02\x00\x59\xE8\x00\x80\x5D\xE8"
-     "\x00\x80"
-     "\x61\xE8\x00\x80\x65\xE8\x00\x80\x69\xE8\x00\x80\x6D\xE8\x00\x80\x71\xE8"
-     "\x00\x80"
-     "\xEF\xF8\x01\x00\xBE\x6C\x04\x00\xE1\x80\x0E\x00\x75\xE8\x00\x80\xE3\x14"
-     "\x01\x00"
-     "\x79\xE8\x00\x80\x7D\xE8\x00\x80\x81\xE8\x00\x80\x85\xE8\x00\x80\x89\xE8"
-     "\x00\x80"
-     "\x8D\xE8\x00\x80\x91\xE8\x00\x80\xB3\xDD\x01\x00\x95\xE8\x00\x80\xB5\xFD"
-     "\x01\x00"
-     "\xB6\xF5\x01\x00\x99\xE8\x00\x80\x9D\xE8\x00\x80\xA1\xE8\x00\x80\xBA\xB1"
-     "\x01\x00"
-     "\xBB\x85\x01\x00\xBC\x9D\x01\x00\xBD\x35\x01\x00\xBE\x3D\x01\x00\xBF\x35"
-     "\x01\x00"
-     "\xA9\x05\x06\x00\xA8\xB9\x05\x00\xAB\x15\x06\x00\xAA\x1D\x06\x00\xAD\x3D"
-     "\x06\x00"
-     "\xAC\x3D\x06\x00\xAF\x4D\x06\x00\xAE\x5D\x06\x00\x41\xE8\x00\x80\x82\x1D"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x80\x1D\x00\x00\xA5\xE8\x00\x80\xA9\xE8\x00\x80\xAD\xE8"
-     "\x00\x80"
-     "\xB1\xE8\x00\x80\xB9\x41\x07\x00\xB8\x9D\x06\x00\xBB\x41\x07\x00\xBA\x49"
-     "\x07\x00"
-     "\xBD\x45\x07\x00\xBC\x59\x07\x00\xBF\x45\x07\x00\xBE\x45\x07\x00\xB1\x09"
-     "\x06\x00"
-     "\xB0\x3D\x06\x00\xB3\xA9\x06\x00\xB2\x01\x06\x00\xB5\xB9\x06\x00\xB4\xB1"
-     "\x06\x00"
-     "\xB7\xAD\x06\x00\xB6\xB1\x06\x00\xA3\x91\x06\x00\x84\x8C\x02\x00\x86\x28"
-     "\x00\x00"
-     "\x87\xC0\x03\x00\xB5\xE8\x00\x80\xA6\xB9\x06\x00\xA5\xB1\x06\x00\xB9\xE8"
-     "\x00\x80"
-     "\xAB\xC9\x06\x00\xAA\xFD\x06\x00\xBD\xE8\x00\x80\xC1\xE8\x00\x80\xAF\x79"
-     "\x06\x00"
-     "\xAE\x71\x06\x00\xAD\x79\x06\x00\xAC\xD1\x06\x00\xC5\xE8\x00\x80\xB3\x99"
-     "\x07\x00"
-     "\xC9\xE8\x00\x80\xCD\xE8\x00\x80\xB6\x51\x07\x00\xD1\xE8\x00\x80\xD5\xE8"
-     "\x00\x80"
-     "\xB5\xB1\x07\x00\xBA\x6D\x07\x00\xBB\x75\x07\x00\xD9\xE8\x00\x80\xDD\xE8"
-     "\x00\x80"
-     "\xBE\x59\x07\x00\xBF\x45\x07\x00\xBC\x6D\x07\x00\xBD\x65\x07\x00\xE1\xE8"
-     "\x00\x80"
-     "\xE5\xE8\x00\x80\xE9\xE8\x00\x80\xED\xE8\x00\x80\xF1\xE8\x00\x80\xF5\xE8"
-     "\x00\x80"
-     "\xF9\xE8\x00\x80\xEF\xA8\x05\x00\xFD\xE8\x00\x80\xE1\x34\x05\x00\x01\xE9"
-     "\x00\x80"
-     "\xE3\x74\x05\x00\x05\xE9\x00\x80\x09\xE9\x00\x80\x0D\xE9\x00\x80\x11\xE9"
-     "\x00\x80"
-     "\xA3\x1D\x06\x00\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D\x00\x00\x15\xE9"
-     "\x00\x80"
-     "\xA6\xD5\x06\x00\xA5\x35\x06\x00\x19\xE9\x00\x80\xAB\xF1\x06\x00\xAA\xE9"
-     "\x06\x00"
-     "\x1D\xE9\x00\x80\x84\x28\x01\x00\xAF\xC1\x06\x00\xAE\xDD\x06\x00\xAD\xE1"
-     "\x06\x00"
-     "\xAC\xE9\x06\x00\xA8\xC5\x06\x00\xA9\xD5\x06\x00\xAA\xD5\x06\x00\xAB\xE5"
-     "\x06\x00"
-     "\xAC\xFD\x06\x00\xAD\x15\x06\x00\xAE\x1D\x06\x00\xAF\x15\x06\x00\xBE\xEC"
-     "\x01\x00"
-     "\x25\xE9\x00\x80\x86\x08\x00\x00\x87\x20\x00\x00\x29\xE9\x00\x80\x2D\xE9"
-     "\x00\x80"
-     "\x31\xE9\x00\x80\x35\xE9\x00\x80\xB8\x7D\x06\x00\xB9\x05\x06\x00\xBA\x0D"
-     "\x06\x00"
-     "\xBB\x05\x06\x00\xBC\x01\x06\x00\xBD\x09\x06\x00\xBE\x39\x06\x00\xBF\x39"
-     "\x06\x00"
-     "\xB0\x6D\x06\x00\xB1\x75\x06\x00\xB2\x7D\x06\x00\xB3\x75\x06\x00\xB4\x59"
-     "\x06\x00"
-     "\xB5\x45\x06\x00\xB6\x4D\x06\x00\xB7\x45\x06\x00\xA8\x91\x02\x00\xA9\x99"
-     "\x02\x00"
-     "\xAA\xA1\x02\x00\xAB\xA1\x02\x00\xAC\xD1\x02\x00\xAD\xDD\x02\x00\xAE\xD5"
-     "\x02\x00"
-     "\xAF\xC9\x02\x00\x39\xE9\x00\x80\x3D\xE9\x00\x80\x41\xE9\x00\x80\xBF\x2C"
-     "\x1E\x00"
-     "\x45\xE9\x00\x80\x49\xE9\x00\x80\x4D\xE9\x00\x80\x51\xE9\x00\x80\xB8\x55"
-     "\x03\x00"
-     "\xB9\x5D\x03\x00\xBA\x69\x03\x00\xBB\x65\x03\x00\xBC\x61\x03\x00\xBD\x61"
-     "\x03\x00"
-     "\xBE\x61\x03\x00\xBF\x61\x03\x00\xB0\xB9\x02\x00\xB1\x8D\x02\x00\xB2\x85"
-     "\x02\x00"
-     "\xB3\x6D\x03\x00\xB4\x75\x03\x00\xB5\x7D\x03\x00\xB6\x75\x03\x00\xB7\x6D"
-     "\x03\x00"
-     "\x55\xE9\x00\x80\x59\xE9\x00\x80\xB3\x21\x02\x00\x5D\xE9\x00\x80\xB5\x51"
-     "\x02\x00"
-     "\x84\x88\x03\x00\x61\xE9\x00\x80\xB6\x55\x02\x00\xD1\xE7\x00\x80\xBE\x28"
-     "\x1C\x00"
-     "\xBB\x41\x02\x00\xBA\x75\x02\x00\xBD\xB1\x03\x00\xBC\x59\x02\x00\xBF\xB1"
-     "\x03\x00"
-     "\xBE\xB9\x03\x00\xA3\x69\x02\x00\x65\xE9\x00\x80\x69\xE9\x00\x80\x6D\xE9"
-     "\x00\x80"
-     "\x71\xE9\x00\x80\xA6\x1D\x02\x00\xA5\x19\x02\x00\x75\xE9\x00\x80\xAB\x09"
-     "\x02\x00"
-     "\xAA\x3D\x02\x00\x79\xE9\x00\x80\x7D\xE9\x00\x80\xAF\xF9\x03\x00\xAE\xF1"
-     "\x03\x00"
-     "\xAD\xF9\x03\x00\xAC\x11\x02\x00\xA8\xA5\x02\x00\xA9\xB5\x02\x00\xAA\xBD"
-     "\x02\x00"
-     "\xAB\xB5\x02\x00\xAC\xAD\x02\x00\xAD\x35\x01\x00\xAE\x3D\x01\x00\xAF\x35"
-     "\x01\x00"
-     "\xBE\x2C\x1C\x00\x81\xE9\x00\x80\x85\xE9\x00\x80\x89\xE9\x00\x80\x8D\xE9"
-     "\x00\x80"
-     "\x95\xE9\x00\x80\x87\x68\x1D\x00\x86\x1C\x1D\x00\xB8\x85\x01\x00\xB9\x8D"
-     "\x01\x00"
-     "\xBA\x85\x01\x00\xBB\x9D\x01\x00\xBC\x8D\x01\x00\xBD\xBD\x01\x00\xBE\xB5"
-     "\x01\x00"
-     "\xBF\x79\x00\x00\xB0\x51\x01\x00\xB1\x51\x01\x00\xB2\x51\x01\x00\xB3\x51"
-     "\x01\x00"
-     "\xB4\xF1\x01\x00\xB5\xF1\x01\x00\xB6\xF5\x01\x00\xB7\xE5\x01\x00\xEF\xD8"
-     "\x00\x00"
-     "\x82\xB5\x00\x00\x81\xA5\x00\x00\x80\xA5\x00\x00\x99\xE9\x00\x80\x9D\xE9"
-     "\x00\x80"
-     "\xA1\xE9\x00\x80\xEF\xC4\x06\x00\xA5\xE9\x00\x80\xE1\xF4\x06\x00\xA9\xE9"
-     "\x00\x80"
-     "\xE3\x38\x01\x00\xE3\xCC\x00\x00\xAD\xE9\x00\x80\xE1\x28\x01\x00\xB1\xE9"
-     "\x00\x80"
-     "\xB5\xE9\x00\x80\xB6\xE5\x01\x00\xB5\x75\x02\x00\x84\x40\x1C\x00\xB3\x65"
-     "\x02\x00"
-     "\xB9\xE9\x00\x80\xBD\xE9\x00\x80\xC1\xE9\x00\x80\xBF\x65\x01\x00\xBE\x65"
-     "\x01\x00"
-     "\xBD\xD5\x01\x00\xBC\xD5\x01\x00\xBB\xC5\x01\x00\xBA\xC5\x01\x00\xC5\xE9"
-     "\x00\x80"
-     "\xC9\xE9\x00\x80\xA3\xB5\x1D\x00\x91\xE9\x00\x80\xCD\xE9\x00\x80\xD1\xE9"
-     "\x00\x80"
-     "\xD5\xE9\x00\x80\xA6\x35\x1E\x00\xA5\xA5\x1D\x00\xD9\xE9\x00\x80\xAB\x15"
-     "\x1E\x00"
-     "\xAA\x15\x1E\x00\xDD\xE9\x00\x80\xE1\xE9\x00\x80\xAF\xB5\x1E\x00\xAE\xB5"
-     "\x1E\x00"
-     "\xAD\x05\x1E\x00\xAC\x05\x1E\x00\xE5\xE9\x00\x80\xE9\xE9\x00\x80\xED\xE9"
-     "\x00\x80"
-     "\xF1\xE9\x00\x80\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\xF5\xE9"
-     "\x00\x80"
-     "\xF9\xE9\x00\x80\xFD\xE9\x00\x80\xB5\x05\x00\x00\x67\x1B\x00\x80\xE1\xB4"
-     "\x02\x00"
-     "\xAC\xD8\x02\x00\xB5\x05\x00\x00\x6B\x1B\x00\x80\xA8\xB5\x1F\x00\xA9\x15"
-     "\x1F\x00"
-     "\xAA\x1D\x1F\x00\xAB\x15\x1F\x00\xAC\x31\x1F\x00\xAD\x3D\x1F\x00\xAE\x2D"
-     "\x1F\x00"
-     "\xAF\x21\x1F\x00\xE1\xB4\x02\x00\xAC\xD8\x02\x00\xB5\x05\x00\x00\x6F\x1B"
-     "\x00\x80"
-     "\xE1\xB4\x02\x00\xAC\xD8\x02\x00\xB5\x05\x00\x00\x73\x1B\x00\x80\xB8\xD1"
-     "\x00\x00"
-     "\xB9\xD9\x00\x00\xBA\xE1\x00\x00\xBB\xE1\x00\x00\xBC\x91\x00\x00\xBD\x91"
-     "\x00\x00"
-     "\xBE\x91\x00\x00\xBF\x91\x00\x00\xB0\x21\x1F\x00\xB1\x31\x1F\x00\xB2\x31"
-     "\x1F\x00"
-     "\xB3\x31\x1F\x00\xB4\x09\x1F\x00\xB5\x09\x1F\x00\xB6\xF1\x00\x00\xB7\xF1"
-     "\x00\x00"
-     "\xE1\xB4\x02\x00\xAC\xDC\x02\x00\xEF\x54\x1D\x00\xB5\x1D\x00\x00\x77\x1B"
-     "\x00\x80"
-     "\xE1\xBC\x02\x00\xAC\xD0\x02\x00\xB5\x29\x00\x00\xA3\x25\x01\x00\xA2\x91"
-     "\x03\x00"
-     "\xA1\x15\x1D\x00\xA0\x0D\x1D\x00\xE1\x80\x1E\x00\x7B\x1B\x00\x80\xE3\xBC"
-     "\x1D\x00"
-     "\xE1\xC4\x02\x00\xB3\xD5\x1E\x00\xB5\x09\x00\x00\xAC\xD8\x02\x00\x7F\x1B"
-     "\x00\x80"
-     "\xE1\xBC\x02\x00\xB6\xFD\x1E\x00\xB5\xF9\x1E\x00\xAC\xE0\x02\x00\xBB\xB5"
-     "\x1E\x00"
-     "\xBA\xA5\x1E\x00\xB5\x05\x00\x00\x83\x1B\x00\x80\xBF\x8D\x1E\x00\xBE\x8D"
-     "\x1E\x00"
-     "\xBD\xA5\x1E\x00\xBC\xA5\x1E\x00\xA3\x15\x1E\x00\xE1\xBC\x02\x00\xAC\xD0"
-     "\x02\x00"
-     "\xB5\x11\x00\x00\x8F\x69\x25\x00\xA6\x3D\x1E\x00\xA5\x39\x1E\x00\x87\x1B"
-     "\x00\x80"
-     "\xAB\x75\x1E\x00\xAA\x65\x1E\x00\xE1\xB4\x02\x00\xAC\x78\x01\x00\xAF\x4D"
-     "\x1E\x00"
-     "\xAE\x4D\x1E\x00\xAD\x65\x1E\x00\xAC\x65\x1E\x00\x9B\xDD\x14\x00\x9A\xE5"
-     "\x15\x00"
-     "\x99\x01\x17\x00\x98\xE1\x10\x00\x9F\x71\x1F\x00\x9E\x79\x19\x00\x9D\xF9"
-     "\x19\x00"
-     "\x9C\x01\x1B\x00\x93\xE5\x2D\x00\x92\x11\x2F\x00\x91\x6D\x29\x00\x90\x6D"
-     "\x29\x00"
-     "\x97\xF9\x11\x00\x96\xF1\x11\x00\x95\x85\x2C\x00\x94\x99\x2D\x00\xB5\x25"
-     "\x00\x00"
-     "\xE1\x94\x02\x00\x82\xF1\x26\x00\x83\x8D\x2A\x00\x84\x95\x2A\x00\x85\xE1"
-     "\x2C\x00"
-     "\x86\x1D\x2E\x00\x87\x79\x2E\x00\xAC\xB4\x02\x00\x8B\x1B\x00\x80\x8A\x55"
-     "\x2F\x00"
-     "\x8B\x29\x12\x00\x8C\x39\x12\x00\x8D\x19\x13\x00\x8E\xF1\x14\x00\x8F\x1D"
-     "\x16\x00"
-     "\xB5\x05\x00\x00\x8F\x1B\x00\x80\x92\x55\x17\x00\x93\x91\x18\x00\x94\x71"
-     "\x1A\x00"
-     "\x95\xF9\x1A\x00\x96\xF9\x1C\x00\x97\x7D\x1E\x00\x82\xE0\x03\x00\x9C\x0B"
-     "\x00\x80"
-     "\x9A\x55\x1E\x00\x9B\xD9\x00\x00\x9C\x75\x02\x00\x95\x0C\x00\x80\xC5\x0C"
-     "\x00\x80"
-     "\xB8\x89\x0A\x00\xAC\x01\x04\x00\xAD\x85\x06\x00\xAE\x81\x06\x00\xCC\x10"
-     "\x02\x00"
-     "\xCC\x7C\x03\x00\xC8\x0C\x00\x80\x93\x1B\x00\x80\xE0\x0C\x00\x80\xE3\x0C"
-     "\x00\x80"
-     "\xE6\x0C\x00\x80\x09\x0B\x00\x80\x81\xA5\x32\x02\xD1\x0C\x00\x80\x21\xE9"
-     "\x00\x80"
-     "\x9A\x95\x06\x00\x9B\x55\x23\x02\xBC\x91\x06\x00\xBD\xB1\x00\x00\xBE\x91"
-     "\x06\x00"
-     "\xBF\xAD\x06\x00\xB8\xE9\x06\x00\xB9\x95\x06\x00\xBA\x91\x06\x00\x97\x1B"
-     "\x00\x80"
-     "\xB4\xC1\x06\x00\xB5\xCD\x06\x00\xB6\xCD\x06\x00\xB7\xDD\x06\x00\xB0\xFD"
-     "\x06\x00"
-     "\xB1\xE5\x06\x00\xB2\x9D\x00\x00\xB3\xE5\x06\x00\x85\x54\xC7\x03\x9B\x1B"
-     "\x00\x80"
-     "\x87\xFC\x00\x00\xB8\x01\x0A\x00\xA3\x1B\x00\x80\x1C\x0D\x00\x80\xA7\x1B"
-     "\x00\x80"
-     "\x8C\xE5\x0C\x00\x8D\xA4\x01\x00\xCC\xF0\x02\x00\x45\x0E\x00\x80\x48\x0E"
-     "\x00\x80"
-     "\x89\x14\x00\x00\xB8\x19\x0A\x00\x8B\x0C\x00\x00\xA1\x0E\x00\x80\xFE\x0E"
-     "\x00\x80"
-     "\xCB\x0C\x00\x80\xD4\x0C\x00\x80\x1F\x0A\x00\x80\x91\xC0\x01\x00\xEC\x0C"
-     "\x00\x80"
-     "\xB8\x6D\x0A\x00\xF2\x0C\x00\x80\xFE\x0C\x00\x80\x07\x0D\x00\x80\x0A\x0D"
-     "\x00\x80"
-     "\x0D\x0D\x00\x80\xAB\x1B\x00\x80\x88\x0D\x00\x80\x16\x0D\x00\x80\xAF\x1B"
-     "\x00\x80"
-     "\xE1\xA4\x1E\x00\x8B\x0D\x00\x80\xE3\x54\x1E\x00\xAF\x25\x73\x00\xCC\x28"
-     "\x02\x00"
-     "\x22\x0D\x00\x80\x1F\x0D\x00\x80\x25\x0D\x00\x80\x28\x0D\x00\x80\x2B\x0D"
-     "\x00\x80"
-     "\xCC\x80\x02\x00\x94\xB8\x03\x00\x2E\x0D\x00\x80\x92\x14\x02\x00\xEF\x60"
-     "\x1E\x00"
-     "\x90\x00\x02\x00\x37\x0D\x00\x80\x4F\x0D\x00\x80\xB8\xF1\x0A\x00\x52\x0D"
-     "\x00\x80"
-     "\xA8\x0B\x00\x80\x5B\x0D\x00\x80\x89\x29\x0B\x00\x5E\x0D\x00\x80\x9F\x1B"
-     "\x00\x80"
-     "\xBC\x30\x01\x00\xBF\xC4\x01\x00\xBE\xEC\x01\x00\x61\x0D\x00\x80\xCC\xEC"
-     "\x02\x00"
-     "\xB8\xC5\x0A\x00\xBA\x44\x01\x00\xB0\x09\x00\x80\x64\x0D\x00\x80\xCA\x06"
-     "\x00\x80"
-     "\xDF\x06\x00\x80\xF2\x06\x00\x80\x70\x0D\x00\x80\xFA\x06\x00\x80\x73\x0D"
-     "\x00\x80"
-     "\x0A\x07\x00\x80\x2D\x07\x00\x80\x18\x07\x00\x80\xF9\x07\x00\x80\x2F\x07"
-     "\x00\x80"
-     "\x30\x0E\x00\x80\x3A\x07\x00\x80\x36\x0E\x00\x80\x4A\x07\x00\x80\xB5\x79"
-     "\x00\x00"
-     "\x6A\x07\x00\x80\xB7\x71\x2A\x02\x75\x07\x00\x80\xB1\x61\x00\x00\x77\x07"
-     "\x00\x80"
-     "\xB3\x0D\x29\x02\x90\x07\x00\x80\xBD\xE9\x00\x00\xA3\x07\x00\x80\xFD\x07"
-     "\x00\x80"
-     "\xB0\x07\x00\x80\xB9\x19\x2B\x02\xC6\x07\x00\x80\xBB\x59\x14\x02\x22\x08"
-     "\x00\x80"
-     "\x5D\x09\x00\x80\x3F\x08\x00\x80\xB6\x0E\x00\x80\x5E\x08\x00\x80\x39\x00"
-     "\x00\x80"
-     "\x84\x08\x00\x80\x71\x00\x00\x80\xCA\x08\x00\x80\x2B\x00\x00\x80\x23\x09"
-     "\x00\x80"
-     "\x3D\x00\x00\x80\x5F\x09\x00\x80\x43\x00\x00\x80\x61\x09\x00\x80\x48\x08"
-     "\x00\x80"
-     "\x6D\x08\x00\x80\x49\x00\x00\x80\x03\x08\x00\x80\x53\x00\x00\x80\x7C\x09"
-     "\x00\x80"
-     "\x59\x00\x00\x80\x76\x0D\x00\x80\x5F\x00\x00\x80\xBB\x0D\x22\x02\x57\x0E"
-     "\x00\x80"
-     "\xCC\x14\x36\x02\x1F\x00\x00\x80\xBF\x65\x00\x00\xBE\x11\x00\x00\xBD\x6D"
-     "\x00\x00"
-     "\xE8\x07\x00\x80\x80\x69\x01\x00\x81\x75\x01\x00\x82\x71\x01\x00\x83\xDD"
-     "\x21\x02"
-     "\x84\x69\x07\x00\x85\x81\x07\x00\x86\x81\x07\x00\x87\x71\x01\x00\x88\xA1"
-     "\x01\x00"
-     "\x89\xAD\x01\x00\x8A\xB5\x07\x00\x8B\x8D\x07\x00\x8C\x95\x07\x00\x8D\xA5"
-     "\x01\x00"
-     "\x4F\x00\x00\x80\x8F\xA5\x01\x00\x90\xE1\x01\x00\x91\xED\x07\x00\x92\xB1"
-     "\x21\x02"
-     "\x93\xFD\x07\x00\x94\x8D\x07\x00\x95\x51\x06\x00\x96\xF1\x01\x00\x97\xD9"
-     "\x01\x00"
-     "\x98\xD1\x01\x00\x99\x75\x06\x00\x9A\x7D\x06\x00\x9B\xD5\x01\x00\x9C\x69"
-     "\x06\x00"
-     "\x9D\x99\x14\x02\x9E\x51\x06\x00\x9F\x51\x06\x00\xA0\x75\x14\x02\xA1\xB9"
-     "\x06\x00"
-     "\xA2\x49\x01\x00\xA3\x85\x2D\x02\xA4\x21\x01\x00\xA5\x2D\x01\x00\xA6\x75"
-     "\x14\x02"
-     "\xA7\xB5\x06\x00\xA8\xA1\x11\x02\xA9\x51\x14\x02\xAA\x95\x06\x00\xAB\x22"
-     "\x00\x80"
-     "\xCC\xBC\x35\x02\xCD\x3C\x35\x02\x6D\x00\x00\x80\xAA\x03\x00\x80\xAC\x03"
-     "\x00\x80"
-     "\xAF\x03\x00\x80\xB7\x22\x00\x80\xBE\x22\x00\x80\xD5\x22\x00\x80\xDC\x22"
-     "\x00\x80"
-     "\xC9\x00\x00\x80\x0F\x00\x00\x80\xB8\xA1\x06\x00\xBA\x06\x00\x80\xB7\x06"
-     "\x00\x80"
-     "\xCC\x00\x00\x80\xC8\x22\x00\x80\xB5\x03\x00\x80\xDD\x05\x00\x80\x18\x06"
-     "\x00\x80"
-     "\xBA\x05\x02\x00\xBB\xD5\x02\x00\xB8\xE9\x05\x00\xB9\xFD\x05\x00\xBE\xC9"
-     "\x02\x00"
-     "\xBF\xE5\x17\x02\xBC\x0D\x02\x00\xBD\x01\x02\x00\xB2\xE1\x05\x00\xB3\xE1"
-     "\x05\x00"
-     "\xB0\x8D\x05\x00\xB1\x9D\x05\x00\xB6\xE5\x05\x00\xB7\xE9\x05\x00\xB4\xF5"
-     "\x05\x00"
-     "\xB5\xED\x05\x00\xAA\x3D\x05\x00\xAB\xC1\x05\x00\xA8\x3D\x05\x00\xA9\x35"
-     "\x05\x00"
-     "\xAE\xCD\x05\x00\xAF\xF5\x05\x00\xAC\xCD\x05\x00\xAD\xC5\x05\x00\xA2\x3D"
-     "\x05\x00"
-     "\xA3\x05\x05\x00\xA0\xD5\x02\x00\xA1\x39\x05\x00\xA6\x1D\x05\x00\xA7\x05"
-     "\x05\x00"
-     "\xA4\x1D\x05\x00\xA5\x15\x05\x00\xBF\x06\x00\x80\x9B\xC1\x05\x00\x3E\x06"
-     "\x00\x80"
-     "\x55\x06\x00\x80\x9E\xDD\x05\x00\x9F\x31\x04\x00\x9C\x51\x02\x00\x9D\xD5"
-     "\x05\x00"
-     "\x72\x06\x00\x80\x89\x06\x00\x80\xA4\x03\x00\x80\x2A\x23\x00\x80\xDB\x00"
-     "\x00\x80"
-     "\xA0\x03\x00\x80\x92\x07\x00\x80\xF1\x07\x00\x80\xF5\x07\x00\x80\x93\x09"
-     "\x00\x80"
-     "\x05\x08\x00\x80\x09\x08\x00\x80\x9F\x0B\x00\x80\x97\x09\x00\x80\xB8\x0A"
-     "\x00\x80"
-     "\x72\x07\x00\x80\x8E\x07\x00\x80\x9A\x07\x00\x80\xA5\x07\x00\x80\xAD\x07"
-     "\x00\x80"
-     "\xAD\x09\x00\x80\xF1\x0F\x00\x80\x08\x10\x00\x80\x15\x10\x00\x80\xCC\x04"
-     "\x33\x02"
-     "\xCD\xB0\x30\x02\xCC\x20\x33\x02\xCD\xE0\x30\x02\xCC\x10\x30\x02\xCC\x68"
-     "\x30\x02"
-     "\xCC\x58\x30\x02\xCD\x8C\x30\x02\xCC\x68\x31\x02\xCD\x14\x31\x02\xCC\xD4"
-     "\x31\x02"
-     "\xCD\x14\x36\x02\xCC\x70\x20\x02\xCD\xD0\x27\x02\xCC\x30\x36\x02\xCC\x24"
-     "\x31\x02"
-     "\xCC\x0C\x3C\x02\xCD\x68\x3F\x02\xCC\x18\x3F\x02\xCD\x34\x3F\x02\xCC\x18"
-     "\x3D\x02"
-     "\xCD\xC0\x32\x02\xCC\x44\x3C\x02\xCC\x18\x39\x02\xCC\x38\x32\x02\xCD\xA8"
-     "\x32\x02"
-     "\xCC\x88\x32\x02\xCD\xF8\x33\x02\xCC\x7C\x33\x02\xCD\x40\x33\x02\xCC\x28"
-     "\x33\x02"
-     "\xCD\x08\x33\x02\xCC\xC8\x21\x02\xCD\x24\x26\x02\xCC\xAC\x26\x02\xCC\x40"
-     "\x38\x02"
-     "\xCC\x98\x25\x02\xCD\xC8\x3A\x02\xCC\x1C\x24\x02\xCD\x10\x24\x02\xCC\x84"
-     "\x3B\x02"
-     "\xCD\xA8\x3B\x02\xCC\xAC\x25\x02\xCD\xBC\x3A\x02\xCC\xAC\x38\x02\xCC\xF8"
-     "\x27\x02"
-     "\xCC\xE0\x38\x02\xCD\x74\x38\x02\x1D\x10\x00\x80\xF6\x06\x00\x80\xE8\x0D"
-     "\x00\x80"
-     "\x0F\x0E\x00\x80\xCD\x20\x2A\x02\xCC\x1C\x2B\x02\xAA\x06\x00\x80\x26\x23"
-     "\x00\x80"
-     "\xCC\xA4\x20\x02\xCD\xA0\x27\x02\xCC\x38\x26\x02\xCA\x04\x00\x80\xCC\x38"
-     "\x3A\x02"
-     "\xCD\x3C\x3B\x02\xCC\x10\x39\x02\xCD\x60\x3E\x02\xCC\xA0\x03\x00\xBE\x3D"
-     "\x0D\x00"
-     "\xBD\xED\x2C\x02\xD6\x04\x00\x80\xBB\x55\x23\x02\x07\x09\x00\x80\xB9\x3D"
-     "\x22\x02"
-     "\xCF\x06\x00\x80\xDC\x07\x00\x80\xA5\x04\x00\x80\x27\x0E\x00\x80\xB2\x04"
-     "\x00\x80"
-     "\x6F\x05\x00\x80\xF4\x06\x00\x80\xBE\x04\x00\x80\x75\x05\x00\x80\xAF\xBD"
-     "\x0C\x00"
-     "\xAE\x99\x2E\x02\xAD\xA5\x0C\x00\xC2\x05\x00\x80\xAB\xC5\x22\x02\x03\x06"
-     "\x00\x80"
-     "\xC4\x04\x00\x80\x23\x06\x00\x80\xD0\x04\x00\x80\xC8\x05\x00\x80\x29\x06"
-     "\x00\x80"
-     "\x5D\x06\x00\x80\xA3\x01\x18\x02\xA0\x04\x00\x80\x1A\x07\x00\x80\x1D\x07"
-     "\x00\x80"
-     "\x9F\x5D\x0C\x00\x9E\x51\x0C\x00\x9D\x45\x0C\x00\x27\x07\x00\x80\x9B\x59"
-     "\x21\x02"
-     "\xB2\x07\x00\x80\xB4\x07\x00\x80\xB7\x07\x00\x80\xBB\x07\x00\x80\x2A\x07"
-     "\x00\x80"
-     "\xD1\x07\x00\x80\xD3\x07\x00\x80\x93\x2D\x26\x02\xD6\x07\x00\x80\x6F\x08"
-     "\x00\x80"
-     "\x72\x08\x00\x80\x8F\x05\x0C\x00\x8E\x71\x0C\x00\x8D\x65\x0C\x00\x7C\x08"
-     "\x00\x80"
-     "\x8B\x45\x20\x02\x63\x09\x00\x80\x89\x35\x2F\x02\x66\x09\x00\x80\x6A\x09"
-     "\x00\x80"
-     "\x7F\x08\x00\x80\x73\x09\x00\x80\x76\x09\x00\x80\xBD\x03\x00\x80\xFA\x22"
-     "\x00\x80"
-     "\x81\x5D\x0C\x00\x80\x61\x0C\x00\x80\x00\x01\x00\x81\x18\x00\x00\x82\x00"
-     "\x04\x00"
-     "\xFE\x22\x00\x80\x84\x10\x07\x00\x85\x14\x06\x00\x86\xE4\x08\x00\x87\x3C"
-     "\x02\x00"
-     "\x88\x2C\x05\x00\x89\xA4\x05\x00\x8A\x00\x78\x00\x02\x23\x00\x80\x8C\x90"
-     "\x06\x00"
-     "\x06\x23\x00\x80\x0E\x23\x00\x80\x0A\x23\x00\x80\xB8\x11\x00\x00\x91\xC4"
-     "\x7B\x00"
-     "\x92\x48\x7A\x00\x93\x4C\x79\x00\x16\x23\x00\x80\xCC\xE8\x02\x00\x96\xF0"
-     "\x09\x00"
-     "\xB8\x39\x00\x00\x90\xC0\x09\x00\x1E\x23\x00\x80\x92\xF0\x09\x00\xCC\xF8"
-     "\x02\x00"
-     "\x94\xB4\x09\x00\xB8\x0D\x00\x00\x22\x23\x00\x80\xCC\x1C\x02\x00\xB8\x05"
-     "\x00\x00"
-     "\x2E\x23\x00\x80\xCC\xE4\x02\x00\xB8\x1D\x00\x00\x32\x23\x00\x80\x36\x23"
-     "\x00\x80"
-     "\x3D\x23\x00\x80\x54\x23\x00\x80\xA8\x8C\x08\x00\xA9\xE4\x7B\x00\x5B\x23"
-     "\x00\x80"
-     "\xAB\x7C\x04\x00\xCC\xE4\x02\x00\xB8\x0D\x00\x00\x65\x23\x00\x80\xCC\x94"
-     "\x02\x00"
-     "\x69\x23\x00\x80\xB1\x94\x7B\x00\xB8\x09\x00\x00\xB8\x15\x00\x00\xCC\xFC"
-     "\x02\x00"
-     "\xB5\xE4\x08\x00\x6D\x23\x00\x80\xCC\xD8\x02\x00\x71\x23\x00\x80\xB8\x05"
-     "\x00\x00"
-     "\xBA\xE8\x05\x00\xBB\x84\x04\x00\xBC\x00\x7C\x00\xBD\xF4\x7F\x00\xBE\xF8"
-     "\x7D\x00"
-     "\xBF\xFC\x72\x00\x80\x09\x3A\x02\x81\x0D\x3A\x02\x82\x01\x3A\x02\x83\x05"
-     "\x3A\x02"
-     "\x84\x19\x3A\x02\x85\x1D\x3A\x02\x86\x11\x3A\x02\x87\x15\x3A\x02\x88\x29"
-     "\x3A\x02"
-     "\x89\x2D\x3A\x02\x8A\x21\x3A\x02\x8B\x25\x3A\x02\x8C\x39\x3A\x02\xCC\x84"
-     "\x02\x00"
-     "\x8E\x31\x3A\x02\x8F\x35\x3A\x02\x7D\x23\x00\x80\x91\xE9\x0F\x00\x79\x23"
-     "\x00\x80"
-     "\xCC\xCC\x02\x00\xB8\x11\x00\x00\x89\x23\x00\x80\xCC\xDC\x02\x00\x97\xE5"
-     "\x0F\x00"
-     "\xB8\x29\x00\x00\xB8\x31\x00\x00\xCC\xF8\x02\x00\x8D\x23\x00\x80\xB8\x1D"
-     "\x00\x00"
-     "\xCC\x3C\x02\x00\x91\x23\x00\x80\xCC\x2C\x02\x00\xCC\x84\x02\x00\x95\x23"
-     "\x00\x80"
-     "\x99\x23\x00\x80\xCC\x84\x02\x00\xA4\x2D\x0F\x00\xA5\x55\x0F\x00\xA6\x5D"
-     "\x0F\x00"
-     "\x9D\x23\x00\x80\xA8\xA9\x3A\x02\xA9\xAD\x3A\x02\xB8\xD1\x00\x00\xA1\x23"
-     "\x00\x80"
-     "\xB8\x35\x00\x00\xA5\x23\x00\x80\xCC\x50\x03\x00\xAF\xB5\x3A\x02\xCC\x2C"
-     "\x03\x00"
-     "\xCC\x14\x03\x00\xC9\x23\x00\x80\xB3\x05\x0F\x00\xB4\x1D\x0F\x00\xCC\x00"
-     "\x03\x00"
-     "\xB6\x09\x0F\x00\xB7\x09\x0F\x00\xBE\x68\x7D\x00\xB8\x6D\x00\x00\xB8\x75"
-     "\x00\x00"
-     "\xCC\x48\x03\x00\xBC\x29\x0F\x00\xCD\x23\x00\x80\xBE\x25\x0F\x00\xCC\x50"
-     "\x03\x00"
-     "\x87\xE5\x0E\x00\xCC\xE8\x3A\x02\xB8\xA5\x01\x00\xB8\xB9\x01\x00\xCC\xF0"
-     "\x35\x02"
-     "\xCD\xA4\x33\x02\xCC\x80\x22\x02\xCD\x7C\x25\x02\xCD\x94\x36\x02\xCC\x04"
-     "\x29\x02"
-     "\xCD\xEC\x3B\x02\xCC\x64\x3A\x02\xB8\xF9\x01\x00\xB8\xC1\x01\x00\x89\xD5"
-     "\x0E\x00"
-     "\x88\xD5\x0E\x00\x97\xB1\x0E\x00\xB8\x0D\x00\x00\xA9\x23\x00\x80\xB8\x1D"
-     "\x00\x00"
-     "\xAD\x23\x00\x80\xB8\x0D\x00\x00\xB1\x23\x00\x80\xB5\x23\x00\x80\x9F\x69"
-     "\x3B\x02"
-     "\xD1\x23\x00\x80\xD5\x23\x00\x80\xB8\x31\x00\x00\xCC\xC4\x02\x00\xCC\xF8"
-     "\x02\x00"
-     "\xB9\x23\x00\x80\xBD\x23\x00\x80\xB8\xE5\x00\x00\xC1\x23\x00\x80\xC5\x23"
-     "\x00\x80"
-     "\xB8\xE5\x00\x00\xB8\xE9\x00\x00\xB8\xE9\x00\x00\xCC\xCC\x33\x02\xD9\x23"
-     "\x00\x80"
-     "\xB8\x25\x00\x00\xCC\xE8\x33\x02\xB8\x21\x00\x00\xDD\x23\x00\x80\xE1\x23"
-     "\x00\x80"
-     "\xB8\x29\x00\x00\xCC\xD8\x02\x00\xB8\x11\x00\x00\xB7\x4D\x0D\x00\xB6\x4D"
-     "\x0D\x00"
-     "\xB5\x35\x0E\x00\xB4\x35\x0E\x00\xB8\x61\x00\x00\xE5\x23\x00\x80\xB1\x19"
-     "\x0F\x00"
-     "\xB0\x29\x0E\x00\xBF\xF5\x03\x00\xBE\x51\x0D\x00\xBD\x59\x0D\x00\xBC\x35"
-     "\x0C\x00"
-     "\xBB\x5D\x0D\x00\xBA\x5D\x0D\x00\xB9\x5D\x0D\x00\xB8\x5D\x0D\x00\x80\xBD"
-     "\x0A\x00"
-     "\x81\xC5\x0A\x00\x82\x15\x04\x00\x83\xC9\x0A\x00\xCC\x7C\x05\x00\x85\xDD"
-     "\x0A\x00"
-     "\x86\xD5\x0A\x00\x87\xCD\x0A\x00\xCC\x50\x05\x00\x89\xF1\x0A\x00\x8A\xB9"
-     "\x08\x00"
-     "\x8B\x0D\x08\x00\x8C\x11\x08\x00\x8D\x15\x08\x00\x8E\xB5\x27\x02\x8F\xE5"
-     "\x0A\x00"
-     "\x90\x69\x08\x00\x91\x6D\x08\x00\x92\x71\x08\x00\x93\x6D\x24\x02\xCC\x24"
-     "\x05\x00"
-     "\x95\x1D\x08\x00\x96\x85\x0A\x00\xCC\x1C\x05\x00\xCC\x28\x05\x00\x99\x3D"
-     "\x08\x00"
-     "\x9A\x89\x0A\x00\x9B\x0D\x08\x00\x9C\x11\x08\x00\x9D\x15\x08\x00\xCC\x2C"
-     "\x05\x00"
-     "\xCC\x20\x05\x00\xA0\x65\x0A\x00\xA1\x6D\x0A\x00\xA2\x65\x0A\x00\xCC\xBC"
-     "\x05\x00"
-     "\xA4\xB1\x04\x00\xCC\xE8\x02\x00\xA6\xB1\x04\x00\xB8\x49\x07\x00\xA8\x81"
-     "\x04\x00"
-     "\x13\x24\x00\x80\xAA\x99\x08\x00\xAB\x5D\x0A\x00\xAC\xB9\x08\x00\xAD\xA9"
-     "\x04\x00"
-     "\x17\x24\x00\x80\xAF\x35\x08\x00\xB0\xD1\x08\x00\xB1\xF1\x04\x00\xD0\x03"
-     "\x00\x80"
-     "\xB3\x8D\x08\x00\xB4\x29\x28\x02\xB5\x21\x0A\x00\xB6\x21\x0A\x00\xB7\x21"
-     "\x0A\x00"
-     "\xB8\x21\x0B\x00\xB9\x25\x08\x00\xE9\x03\x00\x80\xBB\x29\x0B\x00\xBC\x0D"
-     "\x1D\x02"
-     "\xBD\xDD\x0F\x00\xBE\x31\x0B\x00\xBF\x31\x0B\x00\x80\xDD\x0A\x00\xA9\xA1"
-     "\x01\x00"
-     "\xAA\xB1\x01\x00\x02\x04\x00\x80\x1B\x04\x00\x80\x85\x19\x09\x00\x86\x99"
-     "\x09\x00"
-     "\x87\x95\x09\x00\x88\xE1\x09\x00\x89\x25\x25\x02\x2E\x04\x00\x80\x41\x04"
-     "\x00\x80"
-     "\x54\x04\x00\x80\x67\x04\x00\x80\x7A\x04\x00\x80\x8D\x04\x00\x80\x90\xAD"
-     "\x0A\x00"
-     "\x91\x49\x05\x00\x92\x6D\x05\x00\x93\x61\x05\x00\x94\x61\x05\x00\x95\x6D"
-     "\x05\x00"
-     "\x96\x65\x05\x00\x97\x11\x05\x00\x98\x35\x05\x00\x99\x3D\x05\x00\x9A\x35"
-     "\x05\x00"
-     "\x9B\x0D\x05\x00\x9C\x15\x05\x00\x9D\x1D\x05\x00\x9E\x15\x05\x00\x9F\x09"
-     "\x05\x00"
-     "\xA0\xA9\x09\x00\xA1\xFD\x05\x00\xA2\xF5\x05\x00\xA3\x01\x05\x00\xA4\x05"
-     "\x05\x00"
-     "\xA5\x0D\x05\x00\xA6\x05\x05\x00\xA7\x3D\x05\x00\xA8\x05\x05\x00\xA9\x0D"
-     "\x05\x00"
-     "\xAA\x05\x05\x00\xAB\x19\x05\x00\xAC\x89\x09\x00\xAD\xA9\x05\x00\xAE\xB5"
-     "\x09\x00"
-     "\xAF\xFD\x09\x00\xB0\x01\x09\x00\xB1\x7D\x05\x00\xB2\x75\x05\x00\xB3\x01"
-     "\x05\x00"
-     "\xB4\x69\x09\x00\xB5\x01\x05\x00\xB6\x05\x05\x00\xB7\x3D\x05\x00\xB8\x05"
-     "\x05\x00"
-     "\xB9\xE1\x25\x02\xBA\x01\x05\x00\xBB\x01\x05\x00\xBC\xD1\x25\x02\xBD\x3D"
-     "\x09\x00"
-     "\xBE\x79\x09\x00\xBF\x5D\x09\x00\x83\x30\x05\x00\xA1\x78\x07\x00\xF3\x04"
-     "\x00\x80"
-     "\xDD\x00\x00\x80\xA4\x78\x07\x00\xA5\x4C\x07\x00\x13\x01\x00\x80\x1C\x01"
-     "\x00\x80"
-     "\x8B\x7C\x04\x00\x20\x01\x00\x80\x24\x01\x00\x80\x88\x48\x04\x00\x28\x01"
-     "\x00\x80"
-     "\x2C\x01\x00\x80\x30\x01\x00\x80\x34\x01\x00\x80\xE1\x00\x00\x80\xE6\x00"
-     "\x00\x80"
-     "\xB2\x0C\x07\x00\xB3\x00\x07\x00\xEB\x00\x00\x80\xF0\x00\x00\x80\xB6\xE8"
-     "\x07\x00"
-     "\xB7\xE0\x07\x00\xF5\x00\x00\x80\xFA\x00\x00\x80\xBA\xE0\x07\x00\xBB\x9C"
-     "\x07\x00"
-     "\xBC\x84\x07\x00\xBD\x90\x07\x00\xFF\x00\x00\x80\x9F\xD1\x7E\x00\xA3\xCC"
-     "\x04\x00"
-     "\x04\x01\x00\x80\x09\x01\x00\x80\x83\x00\x04\x00\x0E\x01\x00\x80\x85\x74"
-     "\x04\x00"
-     "\xA5\x20\x04\x00\x17\x01\x00\x80\x88\x4C\x04\x00\xCD\x03\x00\x80\xF0\x04"
-     "\x00\x80"
-     "\x8F\x05\x00\x80\xAF\x2D\x07\x00\x8D\xB0\x07\x00\xAD\x21\x07\x00\xAC\x29"
-     "\x07\x00"
-     "\xE2\x05\x00\x80\x1D\x06\x00\x80\x43\x06\x00\x80\xB0\x65\x07\x00\x5A\x06"
-     "\x00\x80"
-     "\x77\x06\x00\x80\x8E\x06\x00\x80\xD3\x03\x00\x80\xEC\x03\x00\x80\x05\x04"
-     "\x00\x80"
-     "\x1E\x04\x00\x80\x31\x04\x00\x80\x44\x04\x00\x80\x57\x04\x00\x80\x6A\x04"
-     "\x00\x80"
-     "\xBC\x7C\x04\x00\x82\xDD\x2B\x02\x83\xE5\x2B\x02\x80\xFD\x0A\x00\x81\xF9"
-     "\x2B\x02"
-     "\x86\x99\x09\x00\x87\x99\x09\x00\x84\xE1\x0A\x00\x85\xE1\x0A\x00\x8A\x89"
-     "\x09\x00"
-     "\x8B\x89\x09\x00\x88\x89\x09\x00\x89\x89\x09\x00\x8E\x85\x09\x00\x7D\x04"
-     "\x00\x80"
-     "\x8C\xE1\x08\x00\x8D\x8D\x09\x00\x92\xB9\x2B\x02\x93\x41\x27\x02\x90\x99"
-     "\x2B\x02"
-     "\x91\xC5\x0B\x00\x96\xC9\x0B\x00\x97\x75\x27\x02\x94\xC5\x0D\x00\x95\xD1"
-     "\x24\x02"
-     "\x9A\xC9\x0B\x00\x9B\xD9\x2A\x02\x98\xC9\x0B\x00\x99\x79\x07\x00\x90\x04"
-     "\x00\x80"
-     "\xF6\x04\x00\x80\x9C\xFD\x0B\x00\xA0\x01\x00\x80\xA4\x01\x00\x80\xA8\x01"
-     "\x00\x80"
-     "\xAC\x01\x00\x80\xB0\x01\x00\x80\xB4\x01\x00\x80\xB8\x01\x00\x80\xBC\x01"
-     "\x00\x80"
-     "\xC0\x01\x00\x80\x9C\x3D\x7F\x00\xB3\x15\x7F\x00\xA8\x11\x09\x00\xC4\x01"
-     "\x00\x80"
-     "\xC8\x01\x00\x80\xCC\x01\x00\x80\xD0\x01\x00\x80\x82\xD8\x7E\x00\xD4\x01"
-     "\x00\x80"
-     "\xD8\x01\x00\x80\xDC\x01\x00\x80\xE0\x01\x00\x80\xE4\x01\x00\x80\xE8\x01"
-     "\x00\x80"
-     "\xEC\x01\x00\x80\xF0\x01\x00\x80\xF4\x01\x00\x80\xF8\x01\x00\x80\xFC\x01"
-     "\x00\x80"
-     "\x00\x02\x00\x80\x04\x02\x00\x80\x3C\x09\x00\x80\x47\x23\x00\x80\xA2\x06"
-     "\x00\x80"
-     "\xA0\xF5\x54\x02\xA1\xD9\x54\x02\xA2\xE9\x54\x02\xA3\xF5\x75\x00\xA4\xED"
-     "\x75\x00"
-     "\xA5\x9D\x76\x00\xA6\x95\x76\x00\xA7\xBD\x76\x00\xA8\x85\x76\x00\xA9\xED"
-     "\x7C\x00"
-     "\xAA\xC1\x7C\x00\xAB\x21\x7D\x00\xAC\x21\x7D\x00\xAD\x1D\x7D\x00\xAE\x15"
-     "\x7D\x00"
-     "\xAF\x0D\x7D\x00\xB0\x75\x7D\x00\xB1\x7D\x7D\x00\xB2\x75\x7D\x00\xB3\x45"
-     "\x7E\x00"
-     "\xB4\x5D\x7E\x00\xB5\x4D\x7E\x00\xB6\x45\x7E\x00\xB7\x79\x7E\x00\xB8\x49"
-     "\x7E\x00"
-     "\xB9\x55\x7E\x00\xBA\x5D\x7E\x00\xBB\x55\x7E\x00\xBC\x4D\x7E\x00\xBD\x35"
-     "\x7E\x00"
-     "\xBE\x3D\x7E\x00\xBF\x2D\x7E\x00\x95\x07\x00\x80\xAC\x06\x00\x80\xDA\x07"
-     "\x00\x80"
-     "\xAF\x06\x00\x80\xB4\x6D\x7F\x00\xB5\x11\x00\x00\xB6\x15\x00\x00\xFD\x23"
-     "\x00\x80"
-     "\xBC\xA1\x58\x02\xBD\x31\x78\x00\x07\x24\x00\x80\xCB\x32\x00\x80\x37\x3B"
-     "\x00\x80"
-     "\x97\x2B\x00\x80\xBB\x2B\x00\x80\xC7\x2B\x00\x80\xD7\x2B\x00\x80\xDF\x2B"
-     "\x00\x80"
-     "\xEB\x2B\x00\x80\xF3\x2B\x00\x80\xFB\x2B\x00\x80\x07\x2C\x00\x80\x62\x2C"
-     "\x00\x80"
-     "\x7A\x2C\x00\x80\x8A\x2C\x00\x80\x9A\x2C\x00\x80\xAA\x2C\x00\x80\xBA\x2C"
-     "\x00\x80"
-     "\xDA\x2C\x00\x80\xDE\x2C\x00\x80\xE2\x2C\x00\x80\x16\x2D\x00\x80\x80\x55"
-     "\x7F\x00"
-     "\x81\x65\x7F\x00\x82\x6D\x7F\x00\x83\x75\x7F\x00\x84\x95\x7F\x00\x85\x9D"
-     "\x7F\x00"
-     "\x86\x89\x7F\x00\x87\x81\x7F\x00\x88\x85\x7F\x00\x89\x8D\x7F\x00\x8A\x85"
-     "\x7F\x00"
-     "\x8B\xED\x7F\x00\x8C\xF5\x7F\x00\x8D\xFD\x7F\x00\x8E\xE9\x7F\x00\x8F\xE9"
-     "\x7F\x00"
-     "\x90\x99\x7F\x00\x91\x99\x7F\x00\x92\xA9\x7F\x00\x93\xA9\x7F\x00\x94\xB9"
-     "\x7F\x00"
-     "\x95\xB9\x7F\x00\x96\xA9\x7F\x00\x97\xA9\x7F\x00\x98\x99\x7F\x00\x99\x51"
-     "\x7E\x00"
-     "\x9A\x19\x7E\x00\x9B\x19\x7E\x00\x9C\x0D\x7E\x00\x9D\xF5\x7E\x00\x9E\xFD"
-     "\x7E\x00"
-     "\x9F\xF5\x7E\x00\xA0\x0D\x7E\x00\xA1\x15\x7E\x00\xA2\x1D\x7E\x00\xA3\x15"
-     "\x7E\x00"
-     "\xA4\x0D\x7E\x00\xA5\x39\x7E\x00\xA6\x29\x7E\x00\xA7\x29\x7E\x00\xA8\x19"
-     "\x7E\x00"
-     "\xA9\x65\x7E\x00\xAA\x6D\x7E\x00\xAB\x65\x7E\x00\xAC\x7D\x7E\x00\xAD\x65"
-     "\x7E\x00"
-     "\xAE\x6D\x7E\x00\xAF\x65\x7E\x00\xB0\x1D\x7E\x00\xB1\x25\x7E\x00\xB2\x2D"
-     "\x7E\x00"
-     "\xB3\x25\x7E\x00\xB4\x3D\x7E\x00\xB5\x25\x7E\x00\xB6\x59\x77\x00\xB7\xD5"
-     "\x75\x00"
-     "\xB8\xFD\x79\x00\xB9\xE9\x75\x00\xBA\xF9\x75\x00\xBB\xD9\x79\x00\xBC\xE9"
-     "\x75\x00"
-     "\xBD\xD5\x75\x00\xBE\xD1\x75\x00\xBF\xD9\x75\x00\x80\x31\x76\x00\x81\x3D"
-     "\x76\x00"
-     "\x82\x49\x76\x00\x83\x45\x76\x00\x84\x41\x76\x00\x85\x4D\x76\x00\x86\xF9"
-     "\x74\x00"
-     "\x87\x7D\x76\x00\x88\xA1\x02\x00\x89\x4D\x76\x00\x8A\x59\x76\x00\x8B\xB9"
-     "\x7A\x00"
-     "\x8C\x49\x76\x00\x8D\x9D\x7A\x00\x8E\xB1\x02\x00\x8F\x1D\x7A\x00\x90\x91"
-     "\x56\x02"
-     "\x91\x29\x76\x00\x92\x81\x76\x00\x93\xCD\x76\x00\x94\xD9\x76\x00\x95\xE9"
-     "\x76\x00"
-     "\x96\xC9\x76\x00\x97\xD1\x59\x02\x98\xA1\x76\x00\x99\x65\x5A\x02\x9A\xF1"
-     "\x76\x00"
-     "\x9B\x4D\x5A\x02\x9C\xD1\x76\x00\x9D\x61\x7A\x00\x9E\x81\x56\x02\x9F\xD5"
-     "\x76\x00"
-     "\xA0\x75\x7D\x00\xA1\x8D\x5A\x02\xA2\x35\x57\x02\xA3\x09\x76\x00\xA4\x21"
-     "\x76\x00"
-     "\xA5\x2D\x76\x00\xA6\x89\x5A\x02\xA7\x99\x5A\x02\xA8\xB9\x5A\x02\xA9\x75"
-     "\x76\x00"
-     "\xAA\x5D\x77\x00\x3E\x2D\x00\x80\x4E\x2D\x00\x80\x56\x2D\x00\x80\x5A\x2D"
-     "\x00\x80"
-     "\x66\x2D\x00\x80\x82\x2D\x00\x80\x86\x2D\x00\x80\x9E\x2D\x00\x80\xA2\x2D"
-     "\x00\x80"
-     "\xAA\x2D\x00\x80\xBA\x2D\x00\x80\x56\x2E\x00\x80\x6A\x2E\x00\x80\xAA\x2E"
-     "\x00\x80"
-     "\xBE\x2E\x00\x80\xC6\x2E\x00\x80\xCA\x2E\x00\x80\xDA\x2E\x00\x80\xFD\x2E"
-     "\x00\x80"
-     "\x29\x2F\x00\x80\x35\x2F\x00\x80\x99\x69\x0A\x00\x55\x2F\x00\x80\x61\x2F"
-     "\x00\x80"
-     "\x65\x2F\x00\x80\x69\x2F\x00\x80\x81\x2F\x00\x80\x85\x2F\x00\x80\xB1\x2F"
-     "\x00\x80"
-     "\x82\xFC\x7A\x00\x83\xE8\x7A\x00\xBD\x2F\x00\x80\xC5\x2F\x00\x80\x86\x70"
-     "\x7B\x00"
-     "\x87\x70\x7B\x00\xCD\x2F\x00\x80\xDD\x2F\x00\x80\xEC\x2F\x00\x80\xF4\x2F"
-     "\x00\x80"
-     "\x20\x30\x00\x80\x24\x30\x00\x80\x2C\x30\x00\x80\x30\x30\x00\x80\x38\x30"
-     "\x00\x80"
-     "\x40\x30\x00\x80\x92\x40\x7C\x00\x50\x30\x00\x80\x6C\x30\x00\x80\x91\xE4"
-     "\x7B\x00"
-     "\xE4\x30\x00\x80\xF8\x30\x00\x80\xFC\x30\x00\x80\x7C\x31\x00\x80\x80\x31"
-     "\x00\x80"
-     "\xAB\xD4\x7C\x00\xA8\xE8\x7C\x00\xA9\xE0\x7C\x00\x94\x31\x00\x80\x98\x31"
-     "\x00\x80"
-     "\xA0\x31\x00\x80\xA8\x31\x00\x80\xA2\xB0\x7C\x00\xB0\x31\x00\x80\xBC\x31"
-     "\x00\x80"
-     "\xF2\x31\x00\x80\xCC\x40\x49\x02\xCD\x00\x49\x02\xCC\xFC\x4A\x02\xCD\x68"
-     "\x4B\x02"
-     "\x02\x32\x00\x80\x16\x32\x00\x80\x93\x32\x00\x80\x9F\x32\x00\x80\xAF\x32"
-     "\x00\x80"
-     "\xBB\x32\x00\x80\xC7\x32\x00\x80\xCF\x32\x00\x80\xB2\xA8\x7C\x00\xB3\xA8"
-     "\x7C\x00"
-     "\xD3\x32\x00\x80\xD7\x32\x00\x80\xDB\x32\x00\x80\xDF\x32\x00\x80\xB4\x44"
-     "\x7C\x00"
-     "\xB5\x58\x7C\x00\x80\x7D\x07\x00\x81\x85\x07\x00\x82\x99\x07\x00\x83\x99"
-     "\x07\x00"
-     "\x84\x89\x07\x00\x85\x89\x07\x00\x86\xBD\x07\x00\x87\xB5\x07\x00\x88\x8D"
-     "\x07\x00"
-     "\x89\xE5\x07\x00\x8A\xED\x07\x00\x8B\xE5\x07\x00\x8C\xFD\x07\x00\x8D\xE5"
-     "\x07\x00"
-     "\x8E\xED\x07\x00\x8F\xE5\x07\x00\x90\x9D\x07\x00\x91\xAD\x07\x00\x92\xA5"
-     "\x07\x00"
-     "\x93\xBD\x07\x00\x94\xA5\x07\x00\x95\x55\x01\x00\x96\x5D\x01\x00\x97\x55"
-     "\x01\x00"
-     "\x98\x6D\x01\x00\x99\x75\x01\x00\x9A\x7D\x01\x00\x9B\x75\x01\x00\x9C\x6D"
-     "\x01\x00"
-     "\x9D\x55\x01\x00\x9E\x5D\x01\x00\x9F\x55\x01\x00\xA0\xAD\x01\x00\xA1\xB5"
-     "\x01\x00"
-     "\xA2\xBD\x01\x00\xA3\xB5\x01\x00\xA4\xAD\x01\x00\xA5\xD5\x01\x00\xA6\xDD"
-     "\x01\x00"
-     "\xA7\xD5\x01\x00\xA8\xED\x01\x00\xA9\xF5\x01\x00\xAA\xFD\x01\x00\xAB\xF5"
-     "\x01\x00"
-     "\xAC\xED\x01\x00\xAD\xD5\x01\x00\xAE\xDD\x01\x00\xAF\xD5\x01\x00\xB0\xAD"
-     "\x01\x00"
-     "\xB1\xB5\x01\x00\xB2\xBD\x01\x00\xB3\xB5\x01\x00\xB4\xAD\x01\x00\xB5\x55"
-     "\x01\x00"
-     "\xB6\x5D\x01\x00\xB7\x55\x01\x00\xB8\x6D\x01\x00\xB9\x75\x01\x00\xBA\x7D"
-     "\x01\x00"
-     "\xBB\x75\x01\x00\xBC\x6D\x01\x00\xBD\x55\x01\x00\xBE\x5D\x01\x00\xBF\x55"
-     "\x01\x00"
-     "\x97\x33\x00\x80\xDF\x33\x00\x80\xEB\x33\x00\x80\xEF\x33\x00\x80\xF3\x33"
-     "\x00\x80"
-     "\xF7\x33\x00\x80\xFF\x33\x00\x80\x03\x34\x00\x80\x17\x34\x00\x80\x33\x34"
-     "\x00\x80"
-     "\x3B\x34\x00\x80\x7B\x34\x00\x80\x7F\x34\x00\x80\x87\x34\x00\x80\x8B\x34"
-     "\x00\x80"
-     "\x93\x34\x00\x80\x97\x34\x00\x80\xBB\x34\x00\x80\xBF\x34\x00\x80\xDB\x34"
-     "\x00\x80"
-     "\xDF\x34\x00\x80\xE3\x34\x00\x80\xE7\x34\x00\x80\xFB\x34\x00\x80\x1F\x35"
-     "\x00\x80"
-     "\x23\x35\x00\x80\x27\x35\x00\x80\x4B\x35\x00\x80\x8B\x35\x00\x80\x8F\x35"
-     "\x00\x80"
-     "\xAF\x35\x00\x80\xBF\x35\x00\x80\xC7\x35\x00\x80\xE7\x35\x00\x80\x0B\x36"
-     "\x00\x80"
-     "\x4F\x36\x00\x80\x57\x36\x00\x80\x6B\x36\x00\x80\x77\x36\x00\x80\x7F\x36"
-     "\x00\x80"
-     "\x87\x36\x00\x80\x8B\x36\x00\x80\x8F\x36\x00\x80\xA7\x36\x00\x80\xAB\x36"
-     "\x00\x80"
-     "\xC7\x36\x00\x80\xCF\x36\x00\x80\xD7\x36\x00\x80\xDB\x36\x00\x80\xE7\x36"
-     "\x00\x80"
-     "\xEF\x36\x00\x80\xF3\x36\x00\x80\xF7\x36\x00\x80\xFF\x36\x00\x80\x03\x37"
-     "\x00\x80"
-     "\xA3\x37\x00\x80\xB7\x37\x00\x80\xEB\x37\x00\x80\xEF\x37\x00\x80\xF7\x37"
-     "\x00\x80"
-     "\x9E\x99\x0C\x00\x23\x38\x00\x80\x2B\x38\x00\x80\x33\x38\x00\x80\x80\xAD"
-     "\x03\x00"
-     "\x81\xB5\x03\x00\x82\xBD\x03\x00\x83\xB5\x03\x00\x84\xAD\x03\x00\x85\xD5"
-     "\x03\x00"
-     "\x86\xDD\x03\x00\x87\xD5\x03\x00\x88\xED\x03\x00\x89\xF5\x03\x00\x8A\xFD"
-     "\x03\x00"
-     "\x8B\xF5\x03\x00\x8C\xED\x03\x00\x8D\xD5\x03\x00\x8E\xDD\x03\x00\x8F\xD5"
-     "\x03\x00"
-     "\x90\xAD\x03\x00\x91\xB1\x03\x00\x92\xB1\x03\x00\x93\xB1\x03\x00\x94\x01"
-     "\x0C\x00"
-     "\x95\x55\x0E\x00\x96\x5D\x0E\x00\x97\x55\x0E\x00\x98\x6D\x0E\x00\x99\x75"
-     "\x0E\x00"
-     "\x9A\x7D\x0E\x00\x9B\x75\x0E\x00\x9C\x6D\x0E\x00\x9D\x55\x0E\x00\x9E\x5D"
-     "\x0E\x00"
-     "\x9F\x55\x0E\x00\xA0\xAD\x0E\x00\xA1\xB5\x0E\x00\xA2\xBD\x0E\x00\xA3\xB5"
-     "\x0E\x00"
-     "\xA4\xAD\x0E\x00\xA5\xD5\x0E\x00\xA6\xDD\x0E\x00\xA7\xD5\x0E\x00\xA8\xED"
-     "\x0E\x00"
-     "\xA9\xF5\x0E\x00\xAA\xFD\x0E\x00\xAB\xF5\x0E\x00\xAC\xED\x0E\x00\xAD\xD5"
-     "\x0E\x00"
-     "\xAE\xDD\x0E\x00\xAF\xD5\x0E\x00\xB0\xAD\x0E\x00\xB1\xB5\x0E\x00\xB2\xBD"
-     "\x0E\x00"
-     "\xB3\xB5\x0E\x00\xB4\xAD\x0E\x00\xB5\x55\x0E\x00\xB6\x5D\x0E\x00\xB7\x55"
-     "\x0E\x00"
-     "\xB8\x6D\x0E\x00\xB9\x75\x0E\x00\xBA\x7D\x0E\x00\xBB\x75\x0E\x00\xBC\x6D"
-     "\x0E\x00"
-     "\xBD\x55\x0E\x00\xBE\x5D\x0E\x00\xBF\x55\x0E\x00\xBC\x69\x04\x00\xBD\x75"
-     "\x04\x00"
-     "\xBE\x7D\x04\x00\xBF\x75\x04\x00\xB8\x45\x04\x00\xB9\x4D\x04\x00\xBA\x45"
-     "\x04\x00"
-     "\xBB\x79\x04\x00\xB4\x65\x04\x00\xB5\x6D\x04\x00\xB6\x65\x04\x00\xB7\x7D"
-     "\x04\x00"
-     "\xB0\x1D\x04\x00\xB1\x0D\x04\x00\xB2\x05\x04\x00\xB3\x7D\x04\x00\xAC\x55"
-     "\x04\x00"
-     "\xAD\x5D\x04\x00\xAE\x55\x04\x00\xAF\x65\x04\x00\xA8\x55\x04\x00\xA9\x5D"
-     "\x04\x00"
-     "\xAA\x55\x04\x00\xAB\x4D\x04\x00\xA4\x91\x05\x00\xA5\x91\x05\x00\xA6\x91"
-     "\x05\x00"
-     "\xA7\x91\x05\x00\xA0\x91\x05\x00\xA1\x91\x05\x00\xA2\x91\x05\x00\xA3\x91"
-     "\x05\x00"
-     "\x9C\x51\x05\x00\x9D\x51\x05\x00\x9E\x51\x05\x00\x9F\x51\x05\x00\x98\x51"
-     "\x05\x00"
-     "\x99\x51\x05\x00\x9A\x51\x05\x00\x9B\x51\x05\x00\x94\x11\x05\x00\x95\x11"
-     "\x05\x00"
-     "\x96\x11\x05\x00\x97\x11\x05\x00\x90\x45\x07\x00\x91\x4D\x07\x00\x92\x45"
-     "\x07\x00"
-     "\x93\x11\x05\x00\x8C\x25\x07\x00\x8D\x2D\x07\x00\x8E\x25\x07\x00\x8F\x3D"
-     "\x07\x00"
-     "\x88\x05\x07\x00\x89\x0D\x07\x00\x8A\x05\x07\x00\x8B\x3D\x07\x00\x84\x25"
-     "\x07\x00"
-     "\x85\x2D\x07\x00\x86\x25\x07\x00\x87\x3D\x07\x00\x80\x45\x07\x00\x81\x4D"
-     "\x07\x00"
-     "\x82\x45\x07\x00\x83\x3D\x07\x00\x77\x38\x00\x80\x7B\x38\x00\x80\x83\x38"
-     "\x00\x80"
-     "\x87\x38\x00\x80\x8B\x38\x00\x80\xB7\x38\x00\x80\xBB\x38\x00\x80\xC3\x38"
-     "\x00\x80"
-     "\xD7\x38\x00\x80\xDB\x38\x00\x80\xF7\x38\x00\x80\xFF\x38\x00\x80\x03\x39"
-     "\x00\x80"
-     "\x27\x39\x00\x80\x47\x39\x00\x80\x5B\x39\x00\x80\x5F\x39\x00\x80\x67\x39"
-     "\x00\x80"
-     "\x93\x39\x00\x80\x97\x39\x00\x80\xA7\x39\x00\x80\xCB\x39\x00\x80\xD7\x39"
-     "\x00\x80"
-     "\xE7\x39\x00\x80\xFF\x39\x00\x80\x07\x3A\x00\x80\x0B\x3A\x00\x80\x0F\x3A"
-     "\x00\x80"
-     "\x13\x3A\x00\x80\x1F\x3A\x00\x80\x23\x3A\x00\x80\x2B\x3A\x00\x80\x47\x3A"
-     "\x00\x80"
-     "\x4B\x3A\x00\x80\x4F\x3A\x00\x80\x67\x3A\x00\x80\x6B\x3A\x00\x80\x73\x3A"
-     "\x00\x80"
-     "\x87\x3A\x00\x80\x8B\x3A\x00\x80\x8F\x3A\x00\x80\x97\x3A\x00\x80\x9B\x3A"
-     "\x00\x80"
-     "\x9F\x3A\x00\x80\xA3\x3A\x00\x80\xA7\x3A\x00\x80\xB7\x3A\x00\x80\xCF\x3A"
-     "\x00\x80"
-     "\xD3\x3A\x00\x80\xDF\x3A\x00\x80\xE3\x3A\x00\x80\xE7\x3A\x00\x80\xEB\x3A"
-     "\x00\x80"
-     "\xF3\x3A\x00\x80\xF7\x3A\x00\x80\xFB\x3A\x00\x80\x07\x3B\x00\x80\x0F\x3B"
-     "\x00\x80"
-     "\x17\x3B\x00\x80\x1B\x3B\x00\x80\x23\x3B\x00\x80\x27\x3B\x00\x80\x2B\x3B"
-     "\x00\x80"
-     "\x33\x3B\x00\x80\x80\xAD\x01\x00\x81\xB5\x01\x00\x82\xBD\x01\x00\x83\xB5"
-     "\x01\x00"
-     "\x84\xAD\x01\x00\x85\xD5\x01\x00\x86\xDD\x01\x00\x87\xD5\x01\x00\x88\xED"
-     "\x01\x00"
-     "\x89\xF5\x01\x00\x8A\xFD\x01\x00\x8B\xF5\x01\x00\x8C\xED\x01\x00\x8D\xD5"
-     "\x01\x00"
-     "\x8E\xDD\x01\x00\x8F\xD5\x01\x00\x90\xAD\x01\x00\x91\xB5\x01\x00\x92\xBD"
-     "\x01\x00"
-     "\x93\xB5\x01\x00\x94\xAD\x01\x00\x95\x0D\x00\x00\x3B\x3B\x00\x80\x3B\x24"
-     "\x00\x80"
-     "\x6A\x2D\x00\x80\x6E\x2D\x00\x80\x23\x25\x00\x80\x82\x78\x02\x00\x99\x05"
-     "\x00\x00"
-     "\xB1\x24\x00\x80\x82\xD0\x02\x00\xBE\x98\x02\x00\x80\x80\x00\x00\x81\x98"
-     "\x00\x00"
-     "\x82\xE4\x06\x00\x83\x88\x04\x00\x84\xD4\x1B\x00\x85\x94\x1A\x00\x86\x18"
-     "\x1F\x00"
-     "\x9E\xE9\x00\x00\x88\xC4\x1E\x00\x89\x00\x10\x00\x8A\xA8\x13\x00\x8B\xAC"
-     "\x11\x00"
-     "\x8C\x00\x28\x00\x8D\xB4\x2B\x00\x8E\xB8\x2A\x00\x8F\xBC\x29\x00\xE3\x74"
-     "\x74\x00"
-     "\xE3\xA0\x02\x00\xE3\xB8\x02\x00\x99\x1D\x00\x00\xB5\x24\x00\x80\xBE\x74"
-     "\x02\x00"
-     "\x9E\x09\x00\x00\xEF\xD8\x02\x00\x82\x6C\x02\x00\xB9\x24\x00\x80\x99\x0D"
-     "\x00\x00"
-     "\xE3\x04\x02\x00\xEF\xE8\x71\x00\xEF\x70\x02\x00\xEF\xD4\x02\x00\xBE\x8C"
-     "\x00\x00"
-     "\xE3\x04\x03\x00\xBD\x24\x00\x80\xE3\x1C\x03\x00\xC1\x24\x00\x80\xE3\x20"
-     "\x03\x00"
-     "\xC5\x24\x00\x80\xE3\x4C\x03\x00\xC9\x24\x00\x80\xEF\xE4\x03\x00\xCD\x24"
-     "\x00\x80"
-     "\xEF\x84\x03\x00\xD1\x24\x00\x80\xEF\xD8\x03\x00\xD5\x24\x00\x80\xEF\xEC"
-     "\x03\x00"
-     "\xD9\x24\x00\x80\xE3\xC8\x03\x00\xE3\xDC\x03\x00\xEF\xA0\x04\x00\xE3\xE8"
-     "\x03\x00"
-     "\x3B\x38\x00\x80\xE3\xA8\x03\x00\xDD\x24\x00\x80\xEF\xA4\x04\x00\xD7\x00"
-     "\x00\x00"
-     "\xEF\x4C\x03\x00\xE1\x24\x00\x80\xE3\x00\x04\x00\xEF\x20\x03\x00\xEF\x3C"
-     "\x03\x00"
-     "\xE3\x78\x04\x00\xEF\x44\x03\x00\xE5\x24\x00\x80\xE9\x24\x00\x80\xED\x24"
-     "\x00\x80"
-     "\xF1\x24\x00\x80\xEF\x58\x04\x00\xF5\x24\x00\x80\xEF\x5C\x04\x00\xE3\x80"
-     "\x04\x00"
-     "\xE3\x80\x04\x00\xF9\x24\x00\x80\xE3\x98\x04\x00\xEF\x74\x04\x00\xFD\x24"
-     "\x00\x80"
-     "\x01\x25\x00\x80\x05\x25\x00\x80\x09\x25\x00\x80\x14\x25\x00\x80\xEF\x60"
-     "\x04\x00"
-     "\xE3\x88\x04\x00\x18\x25\x00\x80\xBD\x00\x00\x80\xC2\x00\x00\x80\x2F\x25"
-     "\x00\x80"
-     "\x1C\x25\x00\x80\x6B\x2A\x00\x80\x06\x26\x00\x80\x69\x26\x00\x80\xAC\x26"
-     "\x00\x80"
-     "\xD8\x26\x00\x80\xE3\xA4\x04\x00\xEF\x44\x04\x00\x02\x27\x00\x80\x80\x09"
-     "\x4B\x02"
-     "\x86\x48\x01\x00\x9E\xED\x02\x00\x82\x70\x02\x00\xBE\x7C\x02\x00\x9E\x11"
-     "\x01\x00"
-     "\x99\x1D\x01\x00\x82\x5C\x02\x00\x8F\x40\x01\x00\x99\x29\x01\x00\x8D\x6C"
-     "\x01\x00"
-     "\xBE\x60\x02\x00\x8B\x78\x01\x00\x9E\x39\x01\x00\x82\x18\x02\x00\xBE\x08"
-     "\x02\x00"
-     "\x97\xD4\x02\x00\x99\x51\x01\x00\x95\xD4\x02\x00\x9E\x59\x01\x00\x93\xD4"
-     "\x02\x00"
-     "\xBE\x58\x02\x00\x91\xD4\x02\x00\x82\x54\x02\x00\x9F\xB8\x02\x00\x96\x00"
-     "\x00\x00"
-     "\x9D\xA8\x02\x00\x99\x71\x01\x00\x9B\xB8\x02\x00\x9E\x69\x01\x00\x99\xB4"
-     "\x02\x00"
-     "\x99\x59\x01\x00\x82\x98\x02\x00\xA6\x94\x02\x00\x9E\xAD\x01\x00\xA4\x78"
-     "\x02\x00"
-     "\x82\x68\x02\x00\x9E\xB9\x01\x00\xA1\x78\x02\x00\x99\xB1\x01\x00\xAF\x84"
-     "\x02\x00"
-     "\x27\x27\x00\x80\xBE\x58\x02\x00\x82\x70\x02\x00\xBE\x5C\x02\x00\x3C\x27"
-     "\x00\x80"
-     "\x99\x8D\x01\x00\xA8\x90\x02\x00\xBE\x7C\x02\x00\x9E\xF1\x01\x00\xB5\xB0"
-     "\x02\x00"
-     "\x99\xF1\x01\x00\xBE\x50\x02\x00\xB2\x90\x02\x00\xB6\x85\x02\x00\x99\x85"
-     "\x01\x00"
-     "\xB8\x5D\x09\x00\xB9\x85\x02\x00\xBA\x8D\x02\x00\x44\x27\x00\x80\xBB\xF4"
-     "\x04\x00"
-     "\x82\x5C\x02\x00\x9E\xC5\x01\x00\xB8\xF4\x04\x00\xBE\x7C\x02\x00\x9E\x5D"
-     "\x06\x00"
-     "\x82\x60\x02\x00\x99\x65\x06\x00\xBE\x70\x02\x00\x82\x6C\x02\x00\x9E\x71"
-     "\x06\x00"
-     "\x99\x9D\x06\x00\xBE\x7C\x02\x00\x9E\xA5\x06\x00\x82\x60\x02\x00\x99\xA9"
-     "\x06\x00"
-     "\xBE\x7C\x02\x00\x9E\xB1\x06\x00\x82\x1C\x02\x00\xBE\x14\x02\x00\x99\xC9"
-     "\x06\x00"
-     "\xBE\x4C\x02\x00\x82\x4C\x02\x00\x9A\xC1\x06\x00\x9E\xD9\x06\x00\x9F\xD9"
-     "\x06\x00"
-     "\x82\xBC\x02\x00\xBE\xB0\x02\x00\x99\xF5\x06\x00\x82\xDC\x05\x00\xBE\x88"
-     "\x02\x00"
-     "\x9A\xC9\x06\x00\x9E\xE5\x06\x00\x9F\xD1\x06\x00\x82\xB8\x02\x00\x82\xA8"
-     "\x02\x00"
-     "\x99\x31\x06\x00\xE3\xF4\x02\x00\xE3\xFC\x02\x00\x9A\x39\x06\x00\x9F\x25"
-     "\x06\x00"
-     "\x9E\x25\x06\x00\x9E\x21\x06\x00\x9F\x21\x06\x00\xBE\x18\x02\x00\x99\x1D"
-     "\x06\x00"
-     "\x9A\x15\x06\x00\xBE\x0C\x02\x00\x99\x71\x06\x00\xEF\x24\x02\x00\xEF\xE0"
-     "\x02\x00"
-     "\x9A\x71\x06\x00\x9E\x01\x06\x00\x9F\x09\x06\x00\xE3\x14\x02\x00\x99\x15"
-     "\x06\x00"
-     "\xE3\x2C\x02\x00\x82\x7C\x02\x00\x9E\x01\x06\x00\x60\x27\x00\x80\xEF\xF0"
-     "\x02\x00"
-     "\x99\x05\x06\x00\xBE\x7C\x02\x00\x9E\x0D\x06\x00\x68\x27\x00\x80\x82\x74"
-     "\x02\x00"
-     "\xEF\xAC\x02\x00\x82\x94\x03\x00\x9B\xD1\x07\x00\xE3\xF0\x03\x00\xBE\x50"
-     "\x02\x00"
-     "\x99\xE9\x07\x00\x9F\xCD\x07\x00\xE3\x34\x02\x00\x9D\x5C\x02\x00\x9C\xC9"
-     "\x07\x00"
-     "\x9E\xC5\x07\x00\xEF\xEC\x02\x00\x82\x60\x02\x00\x99\xC5\x07\x00\xBE\x70"
-     "\x02\x00"
-     "\xEF\xC4\x02\x00\x9E\xD1\x07\x00\x82\x10\x02\x00\xE3\x74\x02\x00\xBE\x6C"
-     "\x02\x00"
-     "\x99\xA5\x07\x00\xBE\x54\x02\x00\x9E\xBD\x07\x00\x99\xA5\x07\x00\x82\x68"
-     "\x02\x00"
-     "\xE3\x0C\x02\x00\x9E\x91\x07\x00\x99\x99\x07\x00\x82\x68\x02\x00\xEF\xD8"
-     "\x02\x00"
-     "\xEF\xCC\x02\x00\xBE\x88\x02\x00\x82\x90\x02\x00\xBE\x98\x02\x00\xBB\x04"
-     "\x00\x00"
-     "\x9E\x79\x07\x00\xB9\x00\x00\x00\x99\x29\x06\x00\xBF\x5C\x00\x00\xE3\xF4"
-     "\x02\x00"
-     "\xBD\x04\x00\x00\x9E\x39\x06\x00\xB3\xB8\x03\x00\xEF\x28\x02\x00\xB1\x78"
-     "\x03\x00"
-     "\x82\x7C\x02\x00\xB7\x3C\x00\x00\x99\x01\x06\x00\xB5\xFC\x03\x00\x9E\x09"
-     "\x06\x00"
-     "\xAB\x48\x03\x00\xBE\x00\x02\x00\xE3\xB8\x02\x00\x99\x21\x06\x00\xAF\x1C"
-     "\x03\x00"
-     "\x82\x7C\x02\x00\xAD\x40\x03\x00\xBE\x44\x02\x00\x9E\x0D\x06\x00\x82\x6C"
-     "\x02\x00"
-     "\xBE\x9C\x01\x00\x99\x31\x01\x00\x82\x80\x01\x00\xA6\xA8\x02\x00\xEF\x7C"
-     "\x02\x00"
-     "\x82\xBC\x01\x00\xE3\xA4\x01\x00\xE3\xD8\x01\x00\xE3\xE4\x01\x00\x9E\xD1"
-     "\x02\x00"
-     "\xE3\xE8\x01\x00\x99\xE5\x02\x00\xBE\x7C\x02\x00\x9E\xED\x02\x00\xEF\xC0"
-     "\x02\x00"
-     "\x99\x09\x00\x00\xBE\x6C\x02\x00\x70\x27\x00\x80\xEF\x0C\x01\x00\xEF\x10"
-     "\x01\x00"
-     "\xEF\x28\x01\x00\x9E\x11\x00\x00\xE3\xB4\x01\x00\xE3\x1C\x02\x00\x99\x2D"
-     "\x00\x00"
-     "\xBE\x6C\x02\x00\x82\x7C\x02\x00\x82\x50\x02\x00\x9E\x3D\x00\x00\xE3\x00"
-     "\x02\x00"
-     "\x99\xE5\x00\x00\x82\x27\x00\x80\xBE\x68\x02\x00\xEF\xFC\x02\x00\xEF\xA0"
-     "\x02\x00"
-     "\xEF\xC0\x02\x00\x9E\xFD\x00\x00\x99\xF1\x00\x00\xE3\x00\x02\x00\x8E\x27"
-     "\x00\x80"
-     "\x82\x90\x02\x00\x96\x27\x00\x80\xBE\x7C\x02\x00\x9E\x19\x00\x00\xE3\x0C"
-     "\x02\x00"
-     "\x82\x58\x02\x00\x99\x0D\x00\x00\x82\x04\x02\x00\xEF\xE4\x02\x00\xAC\x27"
-     "\x00\x80"
-     "\xEF\xD8\x02\x00\xBE\x74\x02\x00\x9E\x15\x00\x00\x99\x65\x01\x00\xE3\x6C"
-     "\x02\x00"
-     "\x99\xC9\x01\x00\x9A\x9D\x01\x00\xBE\x64\x02\x00\x9E\xDD\x01\x00\x9F\x81"
-     "\x01\x00"
-     "\xE3\x0C\x02\x00\x82\x58\x02\x00\x99\x95\x01\x00\xBE\xB4\x03\x00\xEF\xF4"
-     "\x02\x00"
-     "\x9A\x95\x01\x00\xEF\xC4\x02\x00\x82\xAC\x03\x00\xB0\x45\x2A\x00\x82\xF4"
-     "\x03\x00"
-     "\x81\xC8\x02\x00\xE3\xF8\x03\x00\xBE\x30\x02\x00\xEA\x27\x00\x80\xE3\xE4"
-     "\x03\x00"
-     "\x84\xC4\x02\x00\xBE\x38\x02\x00\x86\x10\x02\x00\xEF\xC0\x02\x00\x88\x30"
-     "\x02\x00"
-     "\x9E\x5D\x00\x00\x9F\x55\x00\x00\x9A\xB1\x00\x00\xEF\xC4\x02\x00\x8F\x74"
-     "\x02\x00"
-     "\x99\x41\x00\x00\x9E\x3D\x00\x00\x9F\xC5\x00\x00\x82\x8C\x02\x00\x92\x18"
-     "\x02\x00"
-     "\x95\xD8\x03\x00\xFB\x27\x00\x80\x9E\xC9\x00\x00\x9F\xC9\x00\x00\x9A\xF5"
-     "\x00\x00"
-     "\x98\xF8\x03\x00\x9B\x60\x03\x00\x99\xFD\x00\x00\x82\xA8\x03\x00\x1D\x28"
-     "\x00\x80"
-     "\x28\x28\x00\x80\x78\x28\x00\x80\x8C\x28\x00\x80\xE3\xC4\x02\x00\xBE\x20"
-     "\x02\x00"
-     "\xBE\x50\x02\x00\xAE\x28\x00\x80\xBC\x28\x00\x80\xE3\xE0\x02\x00\x04\x29"
-     "\x00\x80"
-     "\x99\x3D\x00\x00\x9E\x7D\x00\x00\xEF\xDC\x02\x00\x9A\x1D\x00\x00\x12\x29"
-     "\x00\x80"
-     "\xEF\xC4\x02\x00\xAF\xAC\x03\x00\x82\x10\x02\x00\xB1\x00\x1C\x00\xB0\x94"
-     "\x03\x00"
-     "\x99\x29\x00\x00\xB2\x4C\x1C\x00\x9E\x4D\x00\x00\x9F\x65\x00\x00\x6F\x2A"
-     "\x00\x80"
-     "\xE3\x68\x02\x00\x73\x2A\x00\x80\xBE\x74\x02\x00\x9E\x09\x00\x00\xEF\xF0"
-     "\x02\x00"
-     "\xBD\x3C\x1C\x00\x82\x6C\x02\x00\xBF\xCD\x1F\x00\x99\xFD\x1F\x00\x77\x2A"
-     "\x00\x80"
-     "\xBE\x74\x02\x00\x9E\x09\x00\x00\xE3\x14\x02\x00\x82\x6C\x02\x00\x7B\x2A"
-     "\x00\x80"
-     "\x99\x0D\x00\x00\xBE\x44\x02\x00\x9E\x35\x00\x00\x82\x68\x02\x00\x99\x05"
-     "\x00\x00"
-     "\x87\x2A\x00\x80\xBE\x70\x02\x00\xE7\x26\x00\x80\x9E\x15\x00\x00\xEF\xD8"
-     "\x02\x00"
-     "\xE3\x6C\x02\x00\xEF\x19\x00\x80\xF3\x19\x00\x80\xE3\xB0\x1D\x00\x17\x1B"
-     "\x00\x80"
-     "\xF7\x19\x00\x80\x1B\x1B\x00\x80\x1F\x1B\x00\x80\xFB\x19\x00\x80\x23\x1B"
-     "\x00\x80"
-     "\x27\x1B\x00\x80\x2B\x1B\x00\x80\xEF\xB8\x02\x00\x2F\x1B\x00\x80\x33\x1B"
-     "\x00\x80"
-     "\xEF\x80\x02\x00\xB1\xC5\x00\x00\xFF\x19\x00\x80\xB3\xD9\x00\x00\xB2\xC5"
-     "\x00\x00"
-     "\xB5\xC9\x00\x00\xB4\xC1\x00\x00\xB7\xF9\x00\x00\xB6\xC1\x00\x00\xB9\x61"
-     "\x00\x00"
-     "\xB8\x21\x00\x00\xBB\xC1\x00\x00\xBA\xC9\x00\x00\xBD\xC5\x00\x00\xBC\xD9"
-     "\x00\x00"
-     "\xBF\x7D\x00\x00\xBE\xC5\x00\x00\x03\x1A\x00\x80\x37\x1B\x00\x80\x43\x1A"
-     "\x00\x80"
-     "\x07\x1A\x00\x80\x0B\x1A\x00\x80\x0F\x1A\x00\x80\x57\x1A\x00\x80\xEF\x10"
-     "\x78\x03"
-     "\xA3\x1A\x00\x80\xE1\xDC\x4D\x03\xA7\x1A\x00\x80\xE3\x58\x79\x03\xAB\x1A"
-     "\x00\x80"
-     "\xAF\x1A\x00\x80\xB3\x1A\x00\x80\xB7\x1A\x00\x80\x80\xC9\x01\x00\x81\xC9"
-     "\x01\x00"
-     "\x82\xD9\x01\x00\x83\xD9\x01\x00\x84\xC9\x01\x00\x85\x75\x02\x00\x86\x01"
-     "\x04\x00"
-     "\x87\x1D\x05\x00\x88\x25\x05\x00\x89\x35\x05\x00\x8A\x3D\x05\x00\x8B\x6D"
-     "\x05\x00"
-     "\x8C\x75\x05\x00\x8D\x65\x05\x00\x8E\x6D\x05\x00\x8F\xDD\x01\x00\x90\xA9"
-     "\x01\x00"
-     "\x91\xB5\x01\x00\x92\xBD\x01\x00\x93\xB5\x01\x00\x94\xAD\x01\x00\x95\x55"
-     "\x03\x00"
-     "\x96\x5D\x03\x00\x97\x55\x03\x00\x98\x6D\x03\x00\x99\x75\x03\x00\x9A\x7D"
-     "\x03\x00"
-     "\x9B\x71\x03\x00\x9C\x51\x03\x00\x9D\x51\x03\x00\x9E\x51\x03\x00\x9F\x4D"
-     "\x03\x00"
-     "\xA0\xB5\x03\x00\xA1\xBD\x03\x00\xA2\xB5\x03\x00\xA3\xC5\x03\x00\xA4\xC1"
-     "\x03\x00"
-     "\xA5\xC1\x03\x00\xA6\xC5\x03\x00\xA7\xFD\x03\x00\xA8\xC5\x03\x00\xA9\xCD"
-     "\x03\x00"
-     "\xAA\xC5\x03\x00\xAB\xD9\x03\x00\xAC\xC9\x03\x00\xAD\x35\x03\x00\xAE\x3D"
-     "\x03\x00"
-     "\xAF\x35\x03\x00\xB0\x4D\x03\x00\xB1\x55\x03\x00\xB2\x5D\x03\x00\xB3\x55"
-     "\x03\x00"
-     "\xB4\x4D\x03\x00\xB5\x75\x03\x00\xB6\x7D\x03\x00\xB7\x75\x03\x00\xB8\x4D"
-     "\x03\x00"
-     "\xB9\x55\x03\x00\xBA\x5D\x03\x00\xBB\x55\x03\x00\xBC\x4D\x03\x00\xBD\x35"
-     "\x03\x00"
-     "\xBE\x3D\x03\x00\xBF\x31\x03\x00\xBB\x1A\x00\x80\xBF\x1A\x00\x80\xC3\x1A"
-     "\x00\x80"
-     "\xC7\x1A\x00\x80\xCB\x1A\x00\x80\xCF\x1A\x00\x80\xF0\x10\x02\x00\xD3\x1A"
-     "\x00\x80"
-     "\xD7\x1A\x00\x80\xDB\x1A\x00\x80\xDF\x1A\x00\x80\x82\xB4\x02\x00\xE3\x1A"
-     "\x00\x80"
-     "\xE7\x1A\x00\x80\xEB\x1A\x00\x80\xEF\x1A\x00\x80\x9C\xF5\x36\x00\x9D\xFD"
-     "\x36\x00"
-     "\xF3\x1A\x00\x80\x91\x68\x02\x00\x9F\x1A\x00\x80\x3B\x1A\x00\x80\x9E\x5C"
-     "\xDD\x00"
-     "\x3F\x1A\x00\x80\x3B\x1B\x00\x80\x3F\x1B\x00\x80\x43\x1B\x00\x80\x47\x1B"
-     "\x00\x80"
-     "\xBA\xED\xDE\x00\x4B\x1B\x00\x80\x4F\x1B\x00\x80\xF0\xF0\x02\x00\x53\x1B"
-     "\x00\x80"
-     "\x47\x1A\x00\x80\x57\x1B\x00\x80\x5B\x1B\x00\x80\x4B\x1A\x00\x80\x4F\x1A"
-     "\x00\x80"
-     "\x53\x1A\x00\x80\x5B\x1A\x00\x80\x5F\x1A\x00\x80\x63\x1A\x00\x80\x67\x1A"
-     "\x00\x80"
-     "\x6B\x1A\x00\x80\x6F\x1A\x00\x80\x73\x1A\x00\x80\x77\x1A\x00\x80\x7B\x1A"
-     "\x00\x80"
-     "\x7F\x1A\x00\x80\x83\x1A\x00\x80\x91\x24\x02\x00\x87\x1A\x00\x80\x8B\x1A"
-     "\x00\x80"
-     "\x8F\x1A\x00\x80\x5F\x1B\x00\x80\x63\x1B\x00\x80\x93\x1A\x00\x80\x82\x8C"
-     "\x03\x00"
-     "\x97\x1A\x00\x80\x9B\x1A\x00\x80\xF7\x1A\x00\x80\xFB\x1A\x00\x80\xFF\x1A"
-     "\x00\x80"
-     "\xF0\xDC\x07\x00\xE1\xDC\xD2\x02\x03\x1B\x00\x80\xE3\x54\xC6\x02\x07\x1B"
-     "\x00\x80"
-     "\x0B\x1B\x00\x80\x0F\x1B\x00\x80\x13\x1B\x00\x80\xA3\x2B\x00\x80\x5A\x2E"
-     "\x00\x80"
-     "\xA7\x2B\x00\x80\xBF\x2B\x00\x80\xAB\x2B\x00\x80\x9B\x34\x00\x80\xB3\x2B"
-     "\x00\x80"
-     "\xEF\xA4\xE0\x02\x9B\x2B\x00\x80\xE3\x14\x87\x02\xE1\x90\xF3\x02\xE1\x8C"
-     "\x9D\x02"
-     "\xE3\x90\xF7\x02\xE1\x8C\xAB\x02\xE1\x90\xB6\x02\xE3\x00\x53\x03\xE3\x80"
-     "\xA2\x02"
-     "\x13\x1A\x00\x80\xE5\x2E\x00\x80\xEF\xB4\x4B\x03\xEF\xBC\xB3\x02\xEF\x24"
-     "\x8B\x02"
-     "\x6E\x2E\x00\x80\x08\x02\x00\x80\xEF\x24\x97\x02\xBA\x21\xDF\x00\x0E\x02"
-     "\x00\x80"
-     "\x14\x02\x00\x80\x1A\x02\x00\x80\x20\x02\x00\x80\x26\x02\x00\x80\x2C\x02"
-     "\x00\x80"
-     "\x32\x02\x00\x80\x38\x02\x00\x80\x3E\x02\x00\x80\x44\x02\x00\x80\x4A\x02"
-     "\x00\x80"
-     "\x50\x02\x00\x80\x56\x02\x00\x80\x34\x03\x00\x80\x3A\x03\x00\x80\xE1\x80"
-     "\x78\x02"
-     "\xE3\x00\x1E\x02\xE3\x14\x6A\x02\xE1\x10\x13\x02\xE1\xA0\x0F\x02\x40\x03"
-     "\x00\x80"
-     "\xE3\x84\x0E\x02\xEF\x2C\x3F\x02\x46\x03\x00\x80\xE1\x50\x3B\x02\xEF\x34"
-     "\x3B\x02"
-     "\xE3\x3C\x22\x02\x4C\x03\x00\x80\xEF\x28\x1F\x02\xEF\x0C\x12\x02\x52\x03"
-     "\x00\x80"
-     "\x1B\x1A\x00\x80\x1F\x1A\x00\x80\x58\x03\x00\x80\x5E\x03\x00\x80\x23\x1A"
-     "\x00\x80"
-     "\x27\x1A\x00\x80\x76\x03\x00\x80\x82\x03\x00\x80\x88\x03\x00\x80\x8E\x03"
-     "\x00\x80"
-     "\x94\x03\x00\x80\x9A\x03\x00\x80\x7C\x03\x00\x80\x64\x03\x00\x80\x6D\x03"
-     "\x00\x80"
-     "\x5C\x02\x00\x80\x2B\x1A\x00\x80\x2F\x1A\x00\x80\x74\x02\x00\x80\x68\x02"
-     "\x00\x80"
-     "\x33\x1A\x00\x80\x37\x1A\x00\x80\xBC\x02\x00\x80\x7A\x02\x00\x80\x98\x02"
-     "\x00\x80"
-     "\x62\x02\x00\x80\x92\x02\x00\x80\x6E\x02\x00\x80\xA4\x02\x00\x80\xD4\x02"
-     "\x00\x80"
-     "\xF2\x02\x00\x80\xEC\x02\x00\x80\x80\x51\x06\x00\x81\x51\x06\x00\x82\x51"
-     "\x06\x00"
-     "\x83\x51\x06\x00\x84\x71\x06\x00\x85\x7D\x06\x00\x86\x75\x06\x00\x87\x6D"
-     "\x06\x00"
-     "\x88\x55\x06\x00\x89\x5D\x06\x00\x8A\xAD\x07\x00\x8B\xA1\x07\x00\x8C\xA1"
-     "\x07\x00"
-     "\x8D\xA1\x07\x00\x8E\xA1\x07\x00\xE0\x02\x00\x80\x2E\x03\x00\x80\xC2\x02"
-     "\x00\x80"
-     "\x92\x7D\x14\x00\x93\x01\x14\x00\x94\xCD\x07\x00\x95\xF5\x07\x00\x96\xFD"
-     "\x07\x00"
-     "\x97\xF5\x07\x00\x98\xCD\x07\x00\x99\xD5\x07\x00\x9A\x2D\x14\x00\x9B\xD9"
-     "\x07\x00"
-     "\x9C\xCD\x07\x00\x9D\x91\x07\x00\x9E\x8D\x07\x00\x9F\x29\x14\x00\x99\x25"
-     "\x01\x00"
-     "\x98\x25\x01\x00\x9B\x25\x01\x00\x9A\x25\x01\x00\x9D\x25\x01\x00\x9C\x25"
-     "\x01\x00"
-     "\x17\x1A\x00\x80\x9E\x25\x01\x00\x91\xC1\x06\x00\x90\xF1\x06\x00\x93\xD5"
-     "\x06\x00"
-     "\x92\xDD\x06\x00\x95\x3D\x01\x00\x94\x3D\x01\x00\x97\x25\x01\x00\x96\x25"
-     "\x01\x00"
-     "\x89\xE5\x06\x00\x88\xE5\x06\x00\x8B\xE5\x06\x00\x8A\xE5\x06\x00\x8D\xE5"
-     "\x06\x00"
-     "\x8C\xE5\x06\x00\x8F\xE5\x06\x00\x8E\xE5\x06\x00\x81\xE5\x06\x00\x80\x1D"
-     "\x06\x00"
-     "\x83\xE5\x06\x00\x82\xE5\x06\x00\x85\xE5\x06\x00\x84\xE5\x06\x00\x87\xE5"
-     "\x06\x00"
-     "\x86\xE5\x06\x00\xB9\xA5\x03\x00\xB8\xA5\x03\x00\xBB\xA5\x03\x00\xBA\xA5"
-     "\x03\x00"
-     "\xBD\xA5\x03\x00\xBC\xA5\x03\x00\xBF\xA5\x03\x00\xBE\xA5\x03\x00\xB1\xA5"
-     "\x03\x00"
-     "\xB0\xA5\x03\x00\xB3\xA5\x03\x00\xB2\xA5\x03\x00\xB5\xA5\x03\x00\xB4\xA5"
-     "\x03\x00"
-     "\xB7\xA5\x03\x00\xB6\xA5\x03\x00\xA9\xB1\x01\x00\xA8\xA9\x01\x00\xAB\xB1"
-     "\x01\x00"
-     "\xAA\xB9\x01\x00\xAD\x6D\x01\x00\xAC\xA9\x01\x00\xAF\x1D\x01\x00\xAE\x1D"
-     "\x01\x00"
-     "\xA1\xA9\x01\x00\xA0\xD9\x01\x00\xA3\xB1\x01\x00\xA2\xA1\x01\x00\xA5\x91"
-     "\x01\x00"
-     "\xA4\xA9\x01\x00\xA7\x91\x01\x00\xA6\x99\x01\x00\xCE\x02\x00\x80\xE6\x02"
-     "\x00\x80"
-     "\xDA\x02\x00\x80\x04\x03\x00\x80\xB0\x02\x00\x80\xF8\x02\x00\x80\x22\x03"
-     "\x00\x80"
-     "\x0A\x03\x00\x80\x9E\x02\x00\x80\x80\x02\x00\x80\xB6\x02\x00\x80\xC8\x02"
-     "\x00\x80"
-     "\xFE\x02\x00\x80\x86\x02\x00\x80\x28\x03\x00\x80\xAA\x02\x00\x80\x10\x03"
-     "\x00\x80"
-     "\x8C\x02\x00\x80\x16\x03\x00\x80\x1C\x03\x00\x80\x0E\x2E\x00\x80\xF0\x2F"
-     "\x00\x80"
-     "\xCF\x35\x00\x80\x87\x07\x00\x80\x06\x05\x00\x80\x15\x05\x00\x80\x24\x05"
-     "\x00\x80"
-     "\x33\x05\x00\x80\x42\x05\x00\x80\x4B\x05\x00\x80\x54\x05\x00\x80\x5D\x05"
-     "\x00\x80"
-     "\x82\xCC\x0D\x00\x66\x05\x00\x80\x92\x05\x00\x80\x9B\x05\x00\x80\xA4\x05"
-     "\x00\x80"
-     "\xE3\x48\x6E\x03\xA7\x05\x00\x80\xE1\x4C\x6E\x03\xAA\x05\x00\x80\xEF\xF4"
-     "\x01\x03"
-     "\xAD\x05\x00\x80\x63\x3B\x00\x80\xC3\x4D\x00\x80\xF3\x56\x00\x80\x53\x69"
-     "\x00\x80"
-     "\x83\x72\x00\x80\xB3\x7B\x00\x80\x13\x8E\x00\x80\x73\xA0\x00\x80\xA3\xA9"
-     "\x00\x80"
-     "\x0A\xBC\x00\x80\x41\xC5\x00\x80\x71\xCE\x00\x80\xA1\xD7\x00\x80\xD1\xE0"
-     "\x00\x80"
-     "\x4E\xBC\x00\x80\x57\xAF\x00\x80\x23\xA6\x00\x80\xF7\x2B\x00\x80\x8E\x2C"
-     "\x00\x80"
-     "\x9F\x2B\x00\x80\xE3\x2B\x00\x80\x46\x32\x00\x80\x06\x32\x00\x80\x53\x35"
-     "\x00\x80"
-     "\xDA\x2D\x00\x80\x0B\x34\x00\x80\x8F\x38\x00\x80\x67\x35\x00\x80\x22\x2D"
-     "\x00\x80"
-     "\x97\x35\x00\x80\xA3\x34\x00\x80\x17\x39\x00\x80\x5E\x2C\x00\x80\x8A\x2D"
-     "\x00\x80"
-     "\xFF\x32\x00\x80\x2F\x3A\x00\x80\x22\x2C\x00\x80\xB2\x2C\x00\x80\xA3\x32"
-     "\x00\x80"
-     "\xC1\x2F\x00\x80\xCE\x2D\x00\x80\x5E\x2D\x00\x80\x3D\x2F\x00\x80\x31\x2F"
-     "\x00\x80"
-     "\xF3\x34\x00\x80\x1E\x2C\x00\x80\x91\xEC\x03\x00\xA3\x38\x00\x80\x7A\x2D"
-     "\x00\x80"
-     "\xD2\x2E\x00\x80\xB0\x05\x00\x80\xB3\x05\x00\x80\xB6\x05\x00\x80\xE1\xD4"
-     "\x3F\x03"
-     "\xE1\x58\x1A\x03\xE3\xDC\x2F\x03\xE3\x50\x0E\x03\xE1\x14\xF2\x03\xE1\x40"
-     "\xD3\x03"
-     "\xE3\x40\xEA\x03\xE3\x40\xC3\x03\xB9\x05\x00\x80\xE5\x05\x00\x80\xEF\x78"
-     "\xEB\x03"
-     "\xEF\x5C\xF2\x03\xE8\x05\x00\x80\xEB\x05\x00\x80\xEF\x44\xDE\x03\xEF\x98"
-     "\x25\x03"
-     "\xE1\xB4\x8B\x03\xE1\x3C\x97\x03\xE3\x7C\xA2\x03\xE3\x98\x8B\x03\xE1\x30"
-     "\x41\x00"
-     "\xE1\x50\xAC\x03\xE3\x1C\x7F\x00\xE3\x88\x46\x00\xEE\x05\x00\x80\xF1\x05"
-     "\x00\x80"
-     "\xEF\x38\x7B\x00\xEF\xE0\x41\x00\xF4\x05\x00\x80\xF7\x05\x00\x80\xEF\xCC"
-     "\x8A\x03"
-     "\xEF\x20\x87\x03\x80\xB5\x18\x00\x81\xC9\x18\x00\x82\x15\x0B\x00\x83\x2D"
-     "\x0B\x00"
-     "\x84\xE1\x0E\x00\x85\xE1\x0E\x00\x86\xE1\x0E\x00\x87\xFD\x18\x00\x88\xDD"
-     "\x0E\x00"
-     "\x89\x99\x1A\x00\x8A\xE9\x1B\x00\x8B\x2D\x1D\x00\x8C\xC5\x1B\x00\x8D\xF1"
-     "\x1B\x00"
-     "\x8E\x7D\x10\x00\x8F\xE1\x1A\x00\x90\xE9\x1B\x00\x91\x45\x0F\x00\x92\x4D"
-     "\x0F\x00"
-     "\x93\x45\x0F\x00\x94\x5D\x0F\x00\x95\x45\x0F\x00\x96\x4D\x0F\x00\x97\x45"
-     "\x0F\x00"
-     "\x98\x7D\x0F\x00\x99\x45\x0F\x00\x9A\x4D\x0F\x00\x9B\x99\x1A\x00\x9C\x59"
-     "\x0F\x00"
-     "\x9D\x59\x0F\x00\x9E\x49\x0F\x00\x9F\x49\x0F\x00\xA0\xB9\x0F\x00\xA1\xB9"
-     "\x0F\x00"
-     "\xA2\xC9\x0F\x00\xA3\xC9\x0F\x00\xA4\xB5\x0B\x00\xA5\xBD\x0B\x00\xA6\xB5"
-     "\x0B\x00"
-     "\xA7\xD5\x0F\x00\xA8\xED\x0F\x00\xA9\xF5\x0F\x00\xAA\xFD\x0F\x00\xAB\xF5"
-     "\x0F\x00"
-     "\xAC\xED\x0F\x00\xAD\xD5\x0F\x00\xAE\xD1\x0F\x00\xAF\xD1\x0F\x00\xB0\xB1"
-     "\x0F\x00"
-     "\xB1\xB1\x0F\x00\xB2\xB1\x0F\x00\xB3\xB1\x0F\x00\xB4\x71\x0F\x00\xB5\x71"
-     "\x0F\x00"
-     "\xB6\x69\x0F\x00\xB7\x69\x0F\x00\xB8\x01\x0F\x00\xB9\x01\x0F\x00\xBA\x19"
-     "\x0F\x00"
-     "\xBB\x19\x0F\x00\xBC\xF1\x01\x00\xBD\xF1\x01\x00\xBE\xF1\x01\x00\xBF\xF1"
-     "\x01\x00"
-     "\xFA\x05\x00\x80\xFD\x05\x00\x80\x00\x06\x00\x80\x20\x06\x00\x80\xE1\x04"
-     "\x00\x80"
-     "\x80\x05\x00\x80\xD3\x05\x00\x80\x0E\x06\x00\x80\x34\x06\x00\x80\x4B\x06"
-     "\x00\x80"
-     "\x68\x06\x00\x80\x7F\x06\x00\x80\x96\x06\x00\x80\xDD\x03\x00\x80\xF6\x03"
-     "\x00\x80"
-     "\x0F\x04\x00\x80\x12\x07\x00\x80\x44\x08\x00\x80\x41\x08\x00\x80\x3F\x07"
-     "\x00\x80"
-     "\x37\x25\x00\x80\x70\x25\x00\x80\xA1\x25\x00\x80\xC6\x25\x00\x80\xB7\x27"
-     "\x00\x80"
-     "\xC2\x27\x00\x80\xC6\x27\x00\x80\xCA\x27\x00\x80\xCE\x27\x00\x80\x2D\x29"
-     "\x00\x80"
-     "\x6C\x29\x00\x80\x94\x29\x00\x80\x98\x29\x00\x80\xBD\x29\x00\x80\xC5\x29"
-     "\x00\x80"
-     "\xE1\x29\x00\x80\xF3\x29\x00\x80\xF7\x29\x00\x80\xFB\x29\x00\x80\x13\x2A"
-     "\x00\x80"
-     "\x2E\x2A\x00\x80\xBA\x01\xD5\x00\x49\x2A\x00\x80\x17\x2C\x00\x80\x3B\x25"
-     "\x00\x80"
-     "\x48\x25\x00\x80\x55\x25\x00\x80\x62\x25\x00\x80\x74\x25\x00\x80\x81\x25"
-     "\x00\x80"
-     "\x93\x25\x00\x80\xA5\x25\x00\x80\xB5\x25\x00\x80\xCA\x25\x00\x80\xD4\x25"
-     "\x00\x80"
-     "\xE2\x25\x00\x80\xEC\x25\x00\x80\xF9\x25\x00\x80\x0A\x26\x00\x80\x14\x26"
-     "\x00\x80"
-     "\x6D\x26\x00\x80\x74\x26\x00\x80\x22\x26\x00\x80\x7E\x26\x00\x80\x80\x11"
-     "\x03\x00"
-     "\x81\x11\x03\x00\x82\x11\x03\x00\x83\x11\x03\x00\x84\x31\x03\x00\x85\x31"
-     "\x03\x00"
-     "\x86\x31\x03\x00\x87\x31\x03\x00\x88\x11\x03\x00\x89\x11\x03\x00\x8A\x11"
-     "\x03\x00"
-     "\x8B\x11\x03\x00\x8C\x71\x03\x00\x8D\x71\x03\x00\x8E\x71\x03\x00\x8F\x71"
-     "\x03\x00"
-     "\x90\x11\x03\x00\x91\x11\x03\x00\x92\x01\x04\x00\x93\x15\x03\x00\x94\x0D"
-     "\x03\x00"
-     "\x95\x55\x06\x00\x96\x5D\x06\x00\x97\x55\x06\x00\x98\x6D\x06\x00\x99\x75"
-     "\x06\x00"
-     "\x9A\x7D\x06\x00\x9B\x75\x06\x00\x9C\x6D\x06\x00\x9D\x35\x06\x00\x9E\x3D"
-     "\x06\x00"
-     "\x9F\x35\x06\x00\xA0\xCD\x06\x00\xA1\xD5\x06\x00\xA2\xDD\x06\x00\xA3\xD5"
-     "\x06\x00"
-     "\xA4\xF1\x03\x00\xA5\xF1\x03\x00\xA6\xF1\x03\x00\xA7\xF1\x03\x00\xA8\xD1"
-     "\x03\x00"
-     "\xA9\xF9\x06\x00\xAA\x29\x06\x00\xAB\x29\x06\x00\xAC\x39\x06\x00\xAD\x39"
-     "\x06\x00"
-     "\xAE\xCD\x03\x00\xAF\xC5\x03\x00\xB0\xBD\x03\x00\xB1\x45\x03\x00\xB2\x4D"
-     "\x03\x00"
-     "\xB3\x45\x03\x00\xB4\x5D\x03\x00\xB5\x45\x03\x00\xB6\x4D\x03\x00\xB7\x45"
-     "\x03\x00"
-     "\xB8\x7D\x03\x00\xB9\x45\x03\x00\xBA\x41\x03\x00\xBB\x7D\x06\x00\xBC\x65"
-     "\x06\x00"
-     "\xBD\x6D\x06\x00\xBE\x65\x06\x00\xBF\x51\x03\x00\xA9\x15\x0F\x00\xA8\x01"
-     "\x0F\x00"
-     "\xAB\x4D\x0F\x00\xAA\x4D\x0F\x00\xAD\x45\x0F\x00\xAC\x45\x0F\x00\xAF\xA9"
-     "\x0D\x00"
-     "\xAE\xB1\x0D\x00\xA1\x75\x0B\x00\xA0\x61\x0B\x00\xA3\xE9\x0B\x00\xA2\x4D"
-     "\x0B\x00"
-     "\xA5\xE1\x0B\x00\xA4\xF9\x0B\x00\xA7\x39\x0F\x00\xA6\x5D\x08\x00\xB9\x99"
-     "\x0D\x00"
-     "\xB8\x91\x0D\x00\xBB\xA9\x0D\x00\xBA\x91\x0D\x00\xBD\xB9\x0D\x00\xBC\xB1"
-     "\x0D\x00"
-     "\x2F\x26\x00\x80\xBE\xB1\x0D\x00\xB1\xD9\x0D\x00\xB0\xD1\x0D\x00\xB3\xA9"
-     "\x0D\x00"
-     "\xB2\xD1\x0D\x00\xB5\xB9\x0D\x00\xB4\xB1\x0D\x00\xB7\xA9\x0D\x00\xB6\xB1"
-     "\x0D\x00"
-     "\x36\x26\x00\x80\x46\x26\x00\x80\x59\x26\x00\x80\xB0\x26\x00\x80\xBA\x26"
-     "\x00\x80"
-     "\x8F\x26\x00\x80\x9F\x26\x00\x80\xCE\x26\x00\x80\x81\xE5\x0D\x00\x80\xE5"
-     "\x0D\x00"
-     "\x83\xE5\x0D\x00\x82\xE5\x0D\x00\x85\xE5\x0D\x00\x84\xE5\x0D\x00\x87\xE1"
-     "\x0D\x00"
-     "\x86\x2D\x18\x00\x99\x45\x0D\x00\x98\xB9\x0D\x00\x9B\x45\x0D\x00\x9A\x45"
-     "\x0D\x00"
-     "\x9D\x49\x0D\x00\x9C\x45\x0D\x00\x9F\x71\x0D\x00\x9E\x7D\x0D\x00\x91\x81"
-     "\x0D\x00"
-     "\x90\xB9\x0D\x00\x93\x81\x0D\x00\x92\x89\x0D\x00\x95\x81\x0D\x00\x94\x99"
-     "\x0D\x00"
-     "\x97\x81\x0D\x00\x96\x89\x0D\x00\xDE\x25\x00\x80\x1E\x26\x00\x80\x8B\x26"
-     "\x00\x80"
-     "\xCA\x26\x00\x80\xDC\x26\x00\x80\x06\x27\x00\x80\x2B\x27\x00\x80\x40\x27"
-     "\x00\x80"
-     "\xEE\x26\x00\x80\xF8\x26\x00\x80\x0A\x27\x00\x80\x17\x27\x00\x80\x2F\x27"
-     "\x00\x80"
-     "\x4C\x27\x00\x80\x56\x27\x00\x80\x74\x27\x00\x80\x48\x27\x00\x80\x64\x27"
-     "\x00\x80"
-     "\x6C\x27\x00\x80\x7E\x27\x00\x80\x8A\x27\x00\x80\x92\x27\x00\x80\xA1\x27"
-     "\x00\x80"
-     "\xDC\x27\x00\x80\x9A\x27\x00\x80\xB0\x27\x00\x80\xA5\x27\x00\x80\xBB\x27"
-     "\x00\x80"
-     "\xD2\x27\x00\x80\xE0\x27\x00\x80\xFF\x27\x00\x80\x0F\x28\x00\x80\x19\x28"
-     "\x00\x80"
-     "\x4D\x28\x00\x80\x90\x28\x00\x80\xE5\x28\x00\x80\x4D\x2A\x00\x80\x5B\x2A"
-     "\x00\x80"
-     "\x5F\x2A\x00\x80\x63\x2A\x00\x80\xEE\x27\x00\x80\x2C\x28\x00\x80\x3C\x28"
-     "\x00\x80"
-     "\x46\x28\x00\x80\x21\x28\x00\x80\x51\x28\x00\x80\x61\x28\x00\x80\x7C\x28"
-     "\x00\x80"
-     "\x6E\x28\x00\x80\x94\x28\x00\x80\xC0\x28\x00\x80\xC7\x28\x00\x80\xA4\x28"
-     "\x00\x80"
-     "\xD1\x28\x00\x80\xDB\x28\x00\x80\xB2\x28\x00\x80\x16\x29\x00\x80\x08\x29"
-     "\x00\x80"
-     "\xE9\x28\x00\x80\x23\x29\x00\x80\xF0\x28\x00\x80\xFA\x28\x00\x80\x31\x29"
-     "\x00\x80"
-     "\x3E\x29\x00\x80\x48\x29\x00\x80\x52\x29\x00\x80\x5C\x29\x00\x80\x70\x29"
-     "\x00\x80"
-     "\x7D\x29\x00\x80\x84\x29\x00\x80\x9C\x29\x00\x80\xA3\x29\x00\x80\xB0\x29"
-     "\x00\x80"
-     "\xC9\x29\x00\x80\xD3\x29\x00\x80\xE5\x29\x00\x80\xFF\x29\x00\x80\x0C\x2A"
-     "\x00\x80"
-     "\x17\x2A\x00\x80\x21\x2A\x00\x80\x32\x2A\x00\x80\x39\x2A\x00\x80\x51\x2A"
-     "\x00\x80"
-     "\xC3\x03\x00\x80\xE6\x04\x00\x80\x85\x05\x00\x80\xD8\x05\x00\x80\x13\x06"
-     "\x00\x80"
-     "\x39\x06\x00\x80\x50\x06\x00\x80\x6D\x06\x00\x80\x84\x06\x00\x80\x9B\x06"
-     "\x00\x80"
-     "\xE3\x03\x00\x80\xFC\x03\x00\x80\x15\x04\x00\x80\x28\x04\x00\x80\x3B\x04"
-     "\x00\x80"
-     "\x4E\x04\x00\x80\x61\x04\x00\x80\x74\x04\x00\x80\x87\x04\x00\x80\x9A\x04"
-     "\x00\x80"
-     "\x00\x05\x00\x80\x0F\x05\x00\x80\x1E\x05\x00\x80\x2D\x05\x00\x80\x3C\x05"
-     "\x00\x80"
-     "\x66\x08\x00\x80\x27\x08\x00\x80\xC1\x06\x00\x80\xFF\x07\x00\x80\x20\x09"
-     "\x00\x80"
-     "\xE3\x38\x13\x00\x36\x09\x00\x80\x2D\x08\x00\x80\x30\x08\x00\x80\x34\x08"
-     "\x00\x80"
-     "\x24\x07\x00\x80\xE1\x2F\x00\x80\xCF\x31\x00\x80\xDE\x2E\x00\x80\xC0\x31"
-     "\x00\x80"
-     "\x4A\x32\x00\x80\x2A\x09\x00\x80\xEF\xE4\x13\x00\x09\x09\x00\x80\x63\x0E"
-     "\x00\x80"
-     "\xD2\x08\x00\x80\x86\x08\x00\x80\x31\x07\x00\x80\x4C\x07\x00\x80\xFC\x06"
-     "\x00\x80"
-     "\x0D\x08\x00\x80\x97\x08\x00\x80\x2D\x09\x00\x80\x0C\x09\x00\x80\x67\x0E"
-     "\x00\x80"
-     "\x73\x0E\x00\x80\xDE\x08\x00\x80\x9C\x08\x00\x80\x15\x07\x00\x80\x89\x08"
-     "\x00\x80"
-     "\x55\x07\x00\x80\xFF\x06\x00\x80\xA9\x07\x00\x80\x8F\x25\x00\x80\x77\x0E"
-     "\x00\x80"
-     "\xE5\x08\x00\x80\x2A\x08\x00\x80\x9F\x08\x00\x80\x5B\x08\x00\x80\x18\x09"
-     "\x00\x80"
-     "\x7B\x0E\x00\x80\xE8\x08\x00\x80\x17\x08\x00\x80\xA2\x08\x00\x80\xEB\x08"
-     "\x00\x80"
-     "\x1A\x08\x00\x80\xCC\x08\x00\x80\xA5\x08\x00\x80\xEF\x08\x00\x80\x1E\x08"
-     "\x00\x80"
-     "\xCF\x08\x00\x80\xA9\x08\x00\x80\xFA\x08\x00\x80\x00\x09\x00\x80\x8B\x07"
-     "\x00\x80"
-     "\x8D\x08\x00\x80\x59\x07\x00\x80\x03\x07\x00\x80\x40\x09\x00\x80\x44\x09"
-     "\x00\x80"
-     "\x4C\x09\x00\x80\x39\x09\x00\x80\x1B\x09\x00\x80\x7F\x0E\x00\x80\xF7\x08"
-     "\x00\x80"
-     "\x30\x09\x00\x80\x0F\x09\x00\x80\x6B\x0E\x00\x80\xD5\x08\x00\x80\x91\x08"
-     "\x00\x80"
-     "\x60\x07\x00\x80\x33\x09\x00\x80\x12\x09\x00\x80\x6F\x0E\x00\x80\xD8\x08"
-     "\x00\x80"
-     "\x94\x08\x00\x80\x63\x07\x00\x80\xB0\x08\x00\x80\x66\x07\x00\x80\xE3\x40"
-     "\x12\x00"
-     "\xE3\xF0\x13\x00\xE3\xF4\x13\x00\xE3\xD8\x13\x00\xE3\x08\x0D\x00\xE3\xE8"
-     "\x12\x00"
-     "\xE3\xAC\x12\x00\xE3\x64\x12\x00\xEF\xC4\x0D\x00\xEF\xD8\x0D\x00\xEF\x34"
-     "\x12\x00"
-     "\xEF\x58\x12\x00\xEF\x40\x12\x00\xEF\xD4\x12\x00\xEF\xC8\x12\x00\xEF\x20"
-     "\x13\x00"
-     "\xEC\x07\x00\x80\xCC\x06\x00\x80\x11\x08\x00\x80\x14\x08\x00\x80\xD8\x06"
-     "\x00\x80"
-     "\xD4\x06\x00\x80\x24\x08\x00\x80\x07\x07\x00\x80\x6A\x08\x00\x80\x0C\x07"
-     "\x00\x80"
-     "\x79\x08\x00\x80\x34\x07\x00\x80\x37\x07\x00\x80\xAD\x08\x00\x80\xB9\x08"
-     "\x00\x80"
-     "\xBC\x08\x00\x80\xE3\xD0\x10\x00\xE3\xB0\x10\x00\xE3\xA8\x10\x00\xE3\x78"
-     "\x11\x00"
-     "\xE3\x4C\x10\x00\xE3\x30\x10\x00\xE3\xE8\x10\x00\xE3\xFC\x10\x00\xEF\x84"
-     "\x10\x00"
-     "\xEF\x98\x10\x00\xEF\x2C\x10\x00\xEF\x20\x10\x00\xEF\x04\x10\x00\xEF\x08"
-     "\x10\x00"
-     "\xEF\x70\x10\x00\xEF\x4C\x10\x00\xE3\x98\x13\x00\xE3\x30\x10\x00\xE3\x30"
-     "\x10\x00"
-     "\xE3\x20\x10\x00\xE3\x7C\x13\x00\xE3\x40\x13\x00\xE3\x58\x13\x00\xE3\x80"
-     "\x13\x00"
-     "\xEF\xD4\x13\x00\xEF\xA8\x13\x00\xEF\x80\x13\x00\xEF\x7C\x13\x00\xEF\x44"
-     "\x13\x00"
-     "\xEF\x24\x13\x00\xEF\x18\x13\x00\xEF\xEC\x10\x00\x80\xED\x3C\x00\x81\xF5"
-     "\x3C\x00"
-     "\x82\xFD\x3C\x00\x83\xF5\x3C\x00\x84\xED\x3C\x00\x85\x15\x3D\x00\x86\x1D"
-     "\x3D\x00"
-     "\x87\x15\x3D\x00\x88\x2D\x3D\x00\x89\x35\x3D\x00\x8A\x3D\x3D\x00\x8B\x35"
-     "\x3D\x00"
-     "\x8C\x2D\x3D\x00\x8D\x15\x3D\x00\x8E\x1D\x3D\x00\x8F\x15\x3D\x00\x90\x6D"
-     "\x3D\x00"
-     "\x91\x75\x3D\x00\x92\x7D\x3D\x00\x93\x75\x3D\x00\x94\x6D\x3D\x00\x95\x15"
-     "\x3D\x00"
-     "\x96\x1D\x3D\x00\x97\x15\x3D\x00\x98\x2D\x3D\x00\x99\x35\x3D\x00\x9A\x3D"
-     "\x3D\x00"
-     "\x9B\x35\x3D\x00\x9C\x2D\x3D\x00\x9D\x15\x3D\x00\x9E\x1D\x3D\x00\x9F\x15"
-     "\x3D\x00"
-     "\xA0\xED\x3D\x00\xA1\xF5\x3D\x00\xA2\xFD\x3D\x00\xA3\xF5\x3D\x00\xA4\xED"
-     "\x3D\x00"
-     "\xA5\x15\x3D\x00\xA6\x1D\x3D\x00\xA7\x15\x3D\x00\xA8\x2D\x3D\x00\xA9\x35"
-     "\x3D\x00"
-     "\xAA\x3D\x3D\x00\xAB\x35\x3D\x00\xAC\x2D\x3D\x00\xAD\x15\x3D\x00\xAE\x1D"
-     "\x3D\x00"
-     "\xAF\x15\x3D\x00\xB0\x6D\x3D\x00\xB1\x75\x3D\x00\xB2\x7D\x3D\x00\xB3\x75"
-     "\x3D\x00"
-     "\xB4\x6D\x3D\x00\xB5\x15\x3D\x00\xB6\x11\x3D\x00\xB7\x11\x3D\x00\xB8\x31"
-     "\x3D\x00"
-     "\xB9\x31\x3D\x00\xBA\x31\x3D\x00\xBB\x31\x3D\x00\xBC\x11\x3D\x00\xBD\x11"
-     "\x3D\x00"
-     "\xBE\x11\x3D\x00\xBF\x11\x3D\x00\x80\xF1\x3C\x00\x81\xFD\x3C\x00\x82\xF5"
-     "\x3C\x00"
-     "\x83\x0D\x3F\x00\x84\x15\x3F\x00\x85\x1D\x3F\x00\x86\x15\x3F\x00\x87\x0D"
-     "\x3F\x00"
-     "\x88\x35\x3F\x00\x89\x3D\x3F\x00\x8A\x35\x3F\x00\x8B\x0D\x3F\x00\x8C\x15"
-     "\x3F\x00"
-     "\x8D\x1D\x3F\x00\x8E\x15\x3F\x00\x8F\x0D\x3F\x00\x90\x75\x3F\x00\x91\x7D"
-     "\x3F\x00"
-     "\x92\x75\x3F\x00\x93\x0D\x3F\x00\x94\x15\x3F\x00\x95\x19\x3F\x00\x96\x09"
-     "\x3F\x00"
-     "\x97\x09\x3F\x00\x98\x39\x3F\x00\x99\x39\x3F\x00\x9A\x09\x3F\x00\x9B\x09"
-     "\x3F\x00"
-     "\x9C\x19\x3F\x00\x9D\x19\x3F\x00\x9E\x09\x3F\x00\x9F\x09\x3F\x00\xA0\xF9"
-     "\x3F\x00"
-     "\xA1\xF9\x3F\x00\xA2\x09\x3F\x00\xA3\x09\x3F\x00\xA4\x19\x3F\x00\xA5\x19"
-     "\x3F\x00"
-     "\xA6\x09\x3F\x00\xA7\x09\x3F\x00\xA8\x39\x3F\x00\xA9\x39\x3F\x00\xAA\x09"
-     "\x3F\x00"
-     "\xAB\x09\x3F\x00\xAC\x19\x3F\x00\xAD\x19\x3F\x00\xAE\x09\x3F\x00\xAF\x09"
-     "\x3F\x00"
-     "\xB0\x79\x3F\x00\xB1\x79\x3F\x00\xB2\x09\x3F\x00\xB3\x09\x3F\x00\xB4\x19"
-     "\x3F\x00"
-     "\xB5\x19\x3F\x00\xB6\x09\x3F\x00\xB7\x09\x3F\x00\xB8\x39\x3F\x00\xB9\x39"
-     "\x3F\x00"
-     "\xBA\x09\x3F\x00\xBB\x09\x3F\x00\xBC\x19\x3F\x00\xBD\x19\x3F\x00\xBE\x09"
-     "\x3F\x00"
-     "\xBF\x09\x3F\x00\x80\xF9\x3C\x00\x81\xF9\x3C\x00\x82\x49\x3D\x00\x83\x49"
-     "\x3D\x00"
-     "\x84\x59\x3D\x00\x85\x59\x3D\x00\x86\x49\x3D\x00\x87\x49\x3D\x00\x88\x79"
-     "\x3D\x00"
-     "\x89\x79\x3D\x00\x8A\x49\x3D\x00\x8B\x49\x3D\x00\x8C\x59\x3D\x00\x8D\x59"
-     "\x3D\x00"
-     "\x8E\x49\x3D\x00\x8F\x49\x3D\x00\x90\x39\x3D\x00\x91\x39\x3D\x00\x92\x01"
-     "\x04\x00"
-     "\x93\x4D\x06\x00\x94\x55\x06\x00\x95\x5D\x06\x00\x96\x55\x06\x00\x97\x4D"
-     "\x06\x00"
-     "\x98\x75\x06\x00\x99\x7D\x06\x00\x9A\x75\x06\x00\x9B\x4D\x06\x00\x9C\x55"
-     "\x06\x00"
-     "\x9D\x5D\x06\x00\x9E\x55\x06\x00\x9F\x4D\x06\x00\xA0\xB5\x06\x00\xA1\xBD"
-     "\x06\x00"
-     "\xA2\xB5\x06\x00\xA3\xCD\x06\x00\xA4\xD5\x06\x00\xA5\xDD\x06\x00\xA6\xD5"
-     "\x06\x00"
-     "\xA7\xCD\x06\x00\xA8\xF5\x06\x00\xA9\xFD\x06\x00\xAA\xF5\x06\x00\xAB\xCD"
-     "\x06\x00"
-     "\xAC\xD5\x06\x00\xAD\xDD\x06\x00\xAE\xD5\x06\x00\xAF\xCD\x06\x00\xB0\xB5"
-     "\x06\x00"
-     "\xB1\xBD\x06\x00\xB2\xB5\x06\x00\xB3\x4D\x06\x00\xB4\x55\x06\x00\xB5\x5D"
-     "\x06\x00"
-     "\xB6\x55\x06\x00\xB7\x4D\x06\x00\xB8\x75\x06\x00\xB9\x7D\x06\x00\xBA\x75"
-     "\x06\x00"
-     "\xBB\x4D\x06\x00\xBC\x55\x06\x00\xBD\x5D\x06\x00\xBE\x55\x06\x00\xBF\x4D"
-     "\x06\x00"
-     "\xAC\x7D\x3F\x00\xAD\xA5\x3F\x00\xAE\xAD\x3F\x00\xAF\xA5\x3F\x00\xA8\x7D"
-     "\x3F\x00"
-     "\xA9\x65\x3F\x00\xAA\x6D\x3F\x00\xAB\x65\x3F\x00\xA4\x1D\x3F\x00\xA5\x45"
-     "\x3F\x00"
-     "\xA6\x4D\x3F\x00\xA7\x45\x3F\x00\xA0\x1D\x3F\x00\xA1\x05\x3F\x00\xA2\x0D"
-     "\x3F\x00"
-     "\xA3\x05\x3F\x00\xBC\x9D\x3F\x00\xBD\xA5\x3F\x00\xBE\xAD\x3F\x00\xBF\xA5"
-     "\x3F\x00"
-     "\xB8\x9D\x3F\x00\xB9\x85\x3F\x00\xBA\x8D\x3F\x00\xBB\x85\x3F\x00\xB4\xDD"
-     "\x3F\x00"
-     "\xB5\xA5\x3F\x00\xB6\xAD\x3F\x00\xB7\xA5\x3F\x00\xB0\xDD\x3F\x00\xB1\xC5"
-     "\x3F\x00"
-     "\xB2\xCD\x3F\x00\xB3\xC5\x3F\x00\x8C\x61\x3A\x00\x8D\x61\x3A\x00\x8E\x61"
-     "\x3A\x00"
-     "\x8F\x61\x3A\x00\x88\x41\x3A\x00\x89\x41\x3A\x00\x8A\x41\x3A\x00\x8B\x41"
-     "\x3A\x00"
-     "\x84\x61\x3A\x00\x85\x61\x3A\x00\x86\x61\x3A\x00\x87\x61\x3A\x00\x80\x01"
-     "\x3A\x00"
-     "\x81\x01\x3A\x00\x82\x01\x3A\x00\x83\x01\x3A\x00\x9C\x5D\x38\x00\x9D\xE5"
-     "\x3F\x00"
-     "\x9E\xED\x3F\x00\x9F\xE5\x3F\x00\x98\x5D\x38\x00\x99\x45\x38\x00\x9A\x4D"
-     "\x38\x00"
-     "\x9B\x45\x38\x00\x94\xB9\x38\x00\x95\x65\x38\x00\x96\x6D\x38\x00\x97\x65"
-     "\x38\x00"
-     "\x90\x01\x3A\x00\x91\x01\x3A\x00\x92\x01\x3A\x00\x93\x01\x3A\x00\xC3\x08"
-     "\x00\x80"
-     "\xDB\x08\x00\x80\xE1\x08\x00\x80\xF3\x08\x00\x80\x79\x07\x00\x80\x25\x09"
-     "\x00\x80"
-     "\x7C\x07\x00\x80\x84\x07\x00\x80\x57\x09\x00\x80\xA0\x07\x00\x80\xCE\x07"
-     "\x00\x80"
-     "\xC0\x07\x00\x80\xC4\x06\x00\x80\xDC\x04\x00\x80\x7B\x05\x00\x80\xCE\x05"
-     "\x00\x80"
-     "\x09\x06\x00\x80\x2F\x06\x00\x80\x46\x06\x00\x80\x63\x06\x00\x80\x7A\x06"
-     "\x00\x80"
-     "\x91\x06\x00\x80\xD7\x03\x00\x80\xF0\x03\x00\x80\x09\x04\x00\x80\x22\x04"
-     "\x00\x80"
-     "\x35\x04\x00\x80\x48\x04\x00\x80\x5B\x04\x00\x80\x6E\x04\x00\x80\x81\x04"
-     "\x00\x80"
-     "\x94\x04\x00\x80\xFA\x04\x00\x80\x09\x05\x00\x80\x18\x05\x00\x80\x27\x05"
-     "\x00\x80"
-     "\x36\x05\x00\x80\x45\x05\x00\x80\x4E\x05\x00\x80\x57\x05\x00\x80\x60\x05"
-     "\x00\x80"
-     "\x69\x05\x00\x80\x95\x05\x00\x80\x9E\x05\x00\x80\x60\x08\x00\x80\x22\x0F"
-     "\x00\x80"
-     "\x28\x0F\x00\x80\x4B\x3B\x00\x80\xB2\xDD\x5C\x00\xB3\xF1\x5C\x00\xB5\x0A"
-     "\x00\x80"
-     "\x1E\x0B\x00\x80\x47\x3B\x00\x80\x53\x3B\x00\x80\xB4\x39\x52\x00\x10\x0D"
-     "\x00\x80"
-     "\x57\x3B\x00\x80\x67\x0D\x00\x80\xB8\x21\x00\x00\xB9\x21\x00\x00\x7F\x38"
-     "\x00\x80"
-     "\xF6\x31\x00\x80\xA7\x38\x00\x80\x5F\x33\x00\x80\x9A\x30\x35\x00\xC3\x2B"
-     "\x00\x80"
-     "\xBE\x2D\x00\x80\xF1\x2E\x00\x80\x22\x2E\x00\x80\x9F\x1C\x03\x00\xED\x2E"
-     "\x00\x80"
-     "\x9D\x24\x29\x00\x9C\xF1\x01\x00\x9D\xF1\x01\x00\x9E\xF5\x01\x00\x9F\xCD"
-     "\x01\x00"
-     "\xDC\x30\x00\x80\xD7\x34\x00\x80\x67\x36\x00\x80\x9F\x37\x00\x80\xDB\x37"
-     "\x00\x80"
-     "\x0F\x38\x00\x80\x23\x39\x00\x80\x7B\x32\x00\x80\x07\x33\x00\x80\xB3\x33"
-     "\x00\x80"
-     "\x4B\x34\x00\x80\x67\x37\x00\x80\x4F\x39\x00\x80\x7B\x3A\x00\x80\xE7\x2B"
-     "\x00\x80"
-     "\x92\x2D\x00\x80\x8F\x32\x00\x80\xD7\x33\x00\x80\x9B\x37\x00\x80\xCB\x3A"
-     "\x00\x80"
-     "\x99\x2F\x00\x80\x7F\x33\x00\x80\x8B\x37\x00\x80\x1F\x38\x00\x80\x10\x31"
-     "\x00\x80"
-     "\xC3\x36\x00\x80\xE7\x37\x00\x80\x0C\x30\x00\x80\x0A\x32\x00\x80\x1F\x33"
-     "\x00\x80"
-     "\x23\x34\x00\x80\x2F\x39\x00\x80\x2E\x2C\x00\x80\xC6\x2C\x00\x80\x32\x2D"
-     "\x00\x80"
-     "\x78\x31\x00\x80\x87\x32\x00\x80\xD3\x33\x00\x80\xF7\x34\x00\x80\x93\x37"
-     "\x00\x80"
-     "\xAB\x37\x00\x80\xCF\x38\x00\x80\xF7\x39\x00\x80\xAD\x8C\x02\x00\xAB\x3A"
-     "\x00\x80"
-     "\xC7\x3A\x00\x80\x33\x35\x00\x80\x73\x35\x00\x80\x2B\x36\x00\x80\xA9\x11"
-     "\x01\x00"
-     "\x4B\x37\x00\x80\x83\x39\x00\x80\xB0\x00\x1C\x00\xB1\x04\x1E\x00\xB2\x00"
-     "\x18\x00"
-     "\xB3\x0C\x19\x00\xB4\x10\x1B\x00\xB5\x00\xE4\x00\xB6\x18\xE5\x00\xB7\x1C"
-     "\xE7\x00"
-     "\xB8\x00\xE0\x00\xB9\x24\xE1\x00\xBA\x20\xE2\x00\xBB\x00\xEC\x00\xBC\x34"
-     "\xED\x00"
-     "\xBD\x34\xEE\x00\xBE\x38\xEF\x00\xBF\x74\xE3\x00\x53\x38\x00\x80\xA6\x2E"
-     "\x00\x80"
-     "\xBF\x08\x00\x80\x28\x30\x00\x80\xA4\x50\x03\x00\xA5\x00\x0C\x00\xA6\x58"
-     "\x0E\x00"
-     "\xA7\x00\x08\x00\xA8\x60\x0A\x00\xA9\x00\x14\x00\xAA\x68\x15\x00\xAB\xEC"
-     "\x16\x00"
-     "\xAC\x50\x17\x00\xAD\x74\x11\x00\xAE\xB0\x12\x00\xAF\xF0\x13\x00\xE3\x35"
-     "\x00\x80"
-     "\xA3\x36\x00\x80\x07\x34\x00\x80\x3B\x35\x00\x80\x9F\x38\x00\x80\xEF\x39"
-     "\x00\x80"
-     "\xA2\x2E\x00\x80\x75\x2F\x00\x80\x7F\x32\x00\x80\x33\x33\x00\x80\x67\x33"
-     "\x00\x80"
-     "\x1B\x36\x00\x80\x43\x36\x00\x80\xAF\x39\x00\x80\x2E\x2D\x00\x80\xAF\x36"
-     "\x00\x80"
-     "\xD2\x2C\x00\x80\xCE\x2C\x00\x80\x9F\x35\x00\x80\x2F\x36\x00\x80\x63\x37"
-     "\x00\x80"
-     "\xB7\x39\x00\x80\x6F\x38\x00\x80\x14\x31\x00\x80\x5F\x38\x00\x80\xCF\x2B"
-     "\x00\x80"
-     "\x49\x2F\x00\x80\x80\x30\x00\x80\x37\x34\x00\x80\xB7\x34\x00\x80\x52\x2D"
-     "\x00\x80"
-     "\x43\x35\x00\x80\x80\xF1\x03\x00\x81\xF1\x03\x00\x82\x2D\x2F\x00\x83\xF5"
-     "\x03\x00"
-     "\x84\xED\x03\x00\x85\x11\x02\x00\x86\x11\x02\x00\x87\xE1\x2F\x00\x88\xDD"
-     "\x2F\x00"
-     "\x89\x29\x02\x00\x8A\x6D\x2F\x00\x8B\x3D\x02\x00\x8C\x29\x02\x00\x8D\x05"
-     "\x02\x00"
-     "\x8E\x0D\x02\x00\x8F\x05\x02\x00\x90\x7D\x02\x00\x91\x05\x02\x00\x92\x0D"
-     "\x02\x00"
-     "\x93\x05\x02\x00\x94\x1D\x02\x00\x95\x05\x02\x00\x96\x0D\x02\x00\x97\x05"
-     "\x02\x00"
-     "\x98\x3D\x02\x00\x99\x05\x02\x00\x9A\x0D\x02\x00\x9B\x05\x02\x00\x9C\x1D"
-     "\x02\x00"
-     "\x9D\x05\x02\x00\x9E\x0D\x02\x00\x9F\x05\x02\x00\xA0\xFD\x02\x00\xA1\x05"
-     "\x02\x00"
-     "\xA2\x0D\x02\x00\xA3\x05\x02\x00\xA4\x1D\x02\x00\xA5\x05\x02\x00\xA6\x0D"
-     "\x02\x00"
-     "\xA7\x05\x02\x00\xA8\x3D\x02\x00\xA9\x05\x02\x00\xAA\x0D\x02\x00\xAB\x05"
-     "\x02\x00"
-     "\xAC\x1D\x02\x00\xAD\x05\x02\x00\xAE\x0D\x02\x00\xAF\x05\x02\x00\xB0\x7D"
-     "\x02\x00"
-     "\xB1\x05\x02\x00\xB2\x0D\x02\x00\xB3\x05\x02\x00\xB4\x4D\x22\x00\xB5\x19"
-     "\x02\x00"
-     "\xB6\x09\x02\x00\xB7\x09\x02\x00\xB8\x39\x02\x00\xB9\x05\x02\x00\xBA\x0D"
-     "\x02\x00"
-     "\xBB\x05\x02\x00\xBC\x1D\x02\x00\xBD\x05\x02\x00\xBE\x01\x02\x00\xBF\x01"
-     "\x02\x00"
-     "\x95\xC5\x0E\x00\x94\xC5\x0E\x00\x97\xC5\x0E\x00\x96\xC5\x0E\x00\x91\xA5"
-     "\x0E\x00"
-     "\x90\xA5\x0E\x00\x93\xA5\x0E\x00\x92\xA5\x0E\x00\x9D\xC1\x0E\x00\x9C\x69"
-     "\x0D\x00"
-     "\x9F\xC1\x0E\x00\x9E\xC9\x0E\x00\x99\xE5\x0E\x00\x98\xE5\x0E\x00\x9B\xE5"
-     "\x0E\x00"
-     "\x9A\xE5\x0E\x00\x85\xC5\x0E\x00\x84\xC5\x0E\x00\x87\xC5\x0E\x00\x86\xC5"
-     "\x0E\x00"
-     "\x81\x11\x0E\x00\x80\x95\x20\x00\x83\x15\x0E\x00\x82\x19\x0E\x00\x8D\xC5"
-     "\x0E\x00"
-     "\x8C\xC5\x0E\x00\x8F\xC5\x0E\x00\x8E\xC5\x0E\x00\x89\xE5\x0E\x00\x88\xE5"
-     "\x0E\x00"
-     "\x8B\xE5\x0E\x00\x8A\xE5\x0E\x00\xB5\xBD\x01\x00\xB4\xBD\x01\x00\xB7\xCD"
-     "\x01\x00"
-     "\xB6\xCD\x01\x00\xB1\xBD\x20\x00\xB0\xB9\x01\x00\xB3\xAD\x01\x00\xB2\xAD"
-     "\x01\x00"
-     "\xBD\xFD\x01\x00\xBC\xFD\x01\x00\xBF\xCD\x01\x00\xBE\xCD\x01\x00\xB9\xFD"
-     "\x01\x00"
-     "\xB8\xFD\x01\x00\xBB\xED\x01\x00\xBA\xED\x01\x00\xA5\xC1\x01\x00\xA4\x39"
-     "\x0E\x00"
-     "\xA7\xC1\x01\x00\xA6\xC9\x01\x00\xA1\x21\x0E\x00\xA0\x39\x0E\x00\xA3\x21"
-     "\x0E\x00"
-     "\xA2\x29\x0E\x00\xAD\xC1\x01\x00\xAC\xF9\x01\x00\xAF\xC1\x01\x00\xAE\xC9"
-     "\x01\x00"
-     "\xA9\xE1\x01\x00\xA8\xF9\x01\x00\xAB\xE1\x01\x00\xAA\xE9\x01\x00\x3E\x2C"
-     "\x00\x80"
-     "\x64\x30\x00\x80\xAF\x2B\x00\x80\xF3\x32\x00\x80\x90\x31\x00\x80\x17\x36"
-     "\x00\x80"
-     "\xA6\x2D\x00\x80\x92\x2E\x00\x80\x03\x35\x00\x80\x6B\x37\x00\x80\x43\x38"
-     "\x00\x80"
-     "\xBF\x32\x00\x80\xEB\x32\x00\x80\x24\x31\x00\x80\x23\x37\x00\x80\x44\x31"
-     "\x00\x80"
-     "\xB3\x34\x00\x80\x03\x2C\x00\x80\x86\x2C\x00\x80\xCA\x2C\x00\x80\x5B\x32"
-     "\x00\x80"
-     "\x03\x36\x00\x80\xFB\x36\x00\x80\x4F\x38\x00\x80\x63\x39\x00\x80\x3A\x2D"
-     "\x00\x80"
-     "\xEE\x2D\x00\x80\x1C\x30\x00\x80\xAC\x31\x00\x80\x83\x33\x00\x80\x47\x35"
-     "\x00\x80"
-     "\x9F\x39\x00\x80\x83\x3A\x00\x80\xD7\x3A\x00\x80\xEF\x3A\x00\x80\xEE\x31"
-     "\x00\x80"
-     "\xAB\x32\x00\x80\xF3\x38\x00\x80\xE8\x2F\x00\x80\xC4\x30\x00\x80\xE0\x30"
-     "\x00\x80"
-     "\x70\x31\x00\x80\x73\x33\x00\x80\xBF\x33\x00\x80\x6F\x34\x00\x80\x93\x35"
-     "\x00\x80"
-     "\x37\x36\x00\x80\x9B\x36\x00\x80\xE3\x38\x00\x80\x73\x37\x00\x80\x0B\x39"
-     "\x00\x80"
-     "\x87\x39\x00\x80\xEE\x2C\x00\x80\x1A\x2D\x00\x80\x02\x2E\x00\x80\xAF\x33"
-     "\x00\x80"
-     "\xFB\x37\x00\x80\x08\x30\x00\x80\x78\x30\x00\x80\x3C\x31\x00\x80\xB7\x32"
-     "\x00\x80"
-     "\xDF\x35\x00\x80\x2B\x34\x00\x80\x13\x2C\x00\x80\x80\x69\x03\x00\x81\xF1"
-     "\x28\x00"
-     "\x82\x65\x03\x00\x83\x7D\x03\x00\x84\x65\x03\x00\x85\x6D\x03\x00\x86\x65"
-     "\x03\x00"
-     "\x87\x9D\x03\x00\x88\xA5\x03\x00\x89\xAD\x03\x00\x8A\xD5\x2A\x00\x8B\xA1"
-     "\x03\x00"
-     "\x8C\xA1\x03\x00\x8D\xA1\x03\x00\x8E\xA1\x03\x00\x8F\xA1\x03\x00\x90\xE1"
-     "\x03\x00"
-     "\x91\xE1\x03\x00\x92\xE1\x03\x00\x93\xE1\x03\x00\x94\xE1\x03\x00\x95\xE1"
-     "\x03\x00"
-     "\x96\xE1\x03\x00\x97\xE1\x03\x00\x98\xA1\x03\x00\x99\xA1\x03\x00\x9A\xA1"
-     "\x03\x00"
-     "\x9B\xA1\x03\x00\x9C\xA1\x03\x00\x9D\xA1\x03\x00\x9E\xA1\x03\x00\x9F\xA1"
-     "\x03\x00"
-     "\xA0\x61\x03\x00\xA1\xAD\x03\x00\xA2\x6D\x03\x00\xA3\x65\x03\x00\xA4\x7D"
-     "\x03\x00"
-     "\xA5\x65\x03\x00\xA6\x6D\x03\x00\xA7\x65\x03\x00\xA8\x5D\x03\x00\xA9\xA5"
-     "\x01\x00"
-     "\xAA\x41\x0C\x00\xAB\xA9\x01\x00\xAC\xB9\x01\x00\xAD\xB9\x01\x00\xAE\xA9"
-     "\x01\x00"
-     "\xAF\xA9\x01\x00\xB0\xD9\x01\x00\xB1\xD9\x01\x00\xB2\xE9\x01\x00\xB3\xE9"
-     "\x01\x00"
-     "\xB4\xF9\x01\x00\xB5\xF9\x01\x00\xB6\xE9\x01\x00\xB7\xE9\x01\x00\xB8\xD9"
-     "\x01\x00"
-     "\xB9\xD9\x01\x00\xBA\xA9\x01\x00\xBB\xA9\x01\x00\xBC\xB9\x01\x00\xBD\xB9"
-     "\x01\x00"
-     "\xBE\xA9\x01\x00\xBF\xB1\x0F\x00\x5E\x2E\x00\x80\xAD\x2F\x00\x80\x84\x30"
-     "\x00\x80"
-     "\x47\x34\x00\x80\x63\x34\x00\x80\x73\x36\x00\x80\x93\x36\x00\x80\x6B\x39"
-     "\x00\x80"
-     "\xEB\x39\x00\x80\xFB\x39\x00\x80\x37\x3A\x00\x80\x32\x2C\x00\x80\x87\x35"
-     "\x00\x80"
-     "\x57\x39\x00\x80\xD3\x39\x00\x80\x9C\x30\x00\x80\xEF\x2B\x00\x80\x56\x2C"
-     "\x00\x80"
-     "\x4D\x2F\x00\x80\x6D\x2F\x00\x80\x9C\x31\x00\x80\xCB\x33\x00\x80\xD3\x34"
-     "\x00\x80"
-     "\x7F\x37\x00\x80\x1F\x39\x00\x80\xDE\x31\x00\x80\xAB\x39\x00\x80\x12\x2D"
-     "\x00\x80"
-     "\x2A\x32\x00\x80\x36\x32\x00\x80\x17\x33\x00\x80\x4F\x33\x00\x80\x53\x34"
-     "\x00\x80"
-     "\xAF\x34\x00\x80\x2F\x35\x00\x80\x13\x36\x00\x80\x43\x3A\x00\x80\xF0\x30"
-     "\x00\x80"
-     "\xBB\x39\x00\x80\x46\x2C\x00\x80\x91\x2F\x00\x80\x37\x33\x00\x80\xE7\x38"
-     "\x00\x80"
-     "\x54\x30\x00\x80\xA4\x30\x00\x80\x3E\x32\x00\x80\xC3\x39\x00\x80\xF6\x2C"
-     "\x00\x80"
-     "\xDE\x2D\x00\x80\x7D\x2F\x00\x80\xC7\x31\x00\x80\x1A\x32\x00\x80\x67\x32"
-     "\x00\x80"
-     "\xFB\x32\x00\x80\x4F\x34\x00\x80\xC3\x34\x00\x80\x5F\x36\x00\x80\xFF\x37"
-     "\x00\x80"
-     "\x0B\x38\x00\x80\xDF\x39\x00\x80\x62\x2D\x00\x80\x53\x33\x00\x80\xE2\x2D"
-     "\x00\x80"
-     "\xCF\x33\x00\x80\x97\xA9\x07\x00\x96\xA1\x07\x00\x95\xA9\x07\x00\x94\x91"
-     "\x07\x00"
-     "\x93\x89\x07\x00\x92\x81\x07\x00\x91\xED\x3A\x00\x90\x8D\x07\x00\x9F\xE5"
-     "\x07\x00"
-     "\x9E\xE5\x07\x00\x9D\x95\x07\x00\x9C\x95\x07\x00\x9B\xC1\x0A\x00\x9A\x81"
-     "\x07\x00"
-     "\x99\x89\x07\x00\x98\x91\x07\x00\x87\x2D\x07\x00\x86\x2D\x07\x00\x85\x2D"
-     "\x07\x00"
-     "\x84\xB5\x24\x00\x83\x31\x07\x00\x82\x09\x07\x00\x81\x01\x07\x00\x80\x09"
-     "\x07\x00"
-     "\x8F\xED\x07\x00\x8E\xED\x07\x00\x8D\xED\x07\x00\x8C\xED\x07\x00\x8B\x0D"
-     "\x07\x00"
-     "\x8A\x0D\x07\x00\x89\x0D\x07\x00\x88\x0D\x07\x00\xB7\xFD\x29\x00\xB6\x99"
-     "\x01\x00"
-     "\xB5\x91\x01\x00\xB4\x99\x01\x00\xB3\x81\x01\x00\xB2\x99\x01\x00\xB1\x91"
-     "\x01\x00"
-     "\xB0\x99\x01\x00\xBF\x9D\x01\x00\xBE\x9D\x01\x00\xBD\x9D\x01\x00\xBC\x9D"
-     "\x01\x00"
-     "\xBB\x9D\x01\x00\xBA\x9D\x01\x00\xB9\x9D\x01\x00\xB8\x9D\x01\x00\xA7\x25"
-     "\x07\x00"
-     "\xA6\x25\x07\x00\xA5\x15\x07\x00\xA4\x15\x07\x00\xA3\x05\x07\x00\xA2\x05"
-     "\x07\x00"
-     "\xA1\x15\x07\x00\xA0\x15\x07\x00\xAF\xE1\x01\x00\xAE\x19\x07\x00\xAD\x11"
-     "\x07\x00"
-     "\xAC\x19\x07\x00\xAB\x01\x07\x00\xAA\x19\x07\x00\xA9\xB9\x24\x00\xA8\x15"
-     "\x07\x00"
-     "\x73\x34\x00\x80\x8F\x37\x00\x80\xC7\x39\x00\x80\xB3\x3A\x00\x80\xDB\x3A"
-     "\x00\x80"
-     "\x32\x32\x00\x80\x2F\x33\x00\x80\xCB\x36\x00\x80\xCB\x35\x00\x80\xEF\x35"
-     "\x00\x80"
-     "\x1F\x34\x00\x80\xA7\x33\x00\x80\x6B\x38\x00\x80\x0B\x2C\x00\x80\x33\x37"
-     "\x00\x80"
-     "\xFA\x2C\x00\x80\xEA\x2C\x00\x80\xF8\x2F\x00\x80\x18\x31\x00\x80\xFB\x35"
-     "\x00\x80"
-     "\x39\x2F\x00\x80\x0B\x33\x00\x80\xEA\x31\x00\x80\xEF\x38\x00\x80\xB3\x39"
-     "\x00\x80"
-     "\x14\x30\x00\x80\x66\x2C\x00\x80\x3A\x2E\x00\x80\x8E\x2E\x00\x80\xDA\x31"
-     "\x00\x80"
-     "\xD6\x2C\x00\x80\x27\x34\x00\x80\xEB\x34\x00\x80\x4F\x35\x00\x80\x6B\x35"
-     "\x00\x80"
-     "\x6F\x35\x00\x80\x83\x35\x00\x80\x03\x3A\x00\x80\xF3\x35\x00\x80\x97\x37"
-     "\x00\x80"
-     "\x5B\x38\x00\x80\x0F\x39\x00\x80\x3B\x39\x00\x80\x57\x3A\x00\x80\x5B\x3A"
-     "\x00\x80"
-     "\x5F\x3A\x00\x80\xC3\x3A\x00\x80\x33\x39\x00\x80\xCF\x39\x00\x80\x36\x2C"
-     "\x00\x80"
-     "\x5A\x2C\x00\x80\x6E\x2C\x00\x80\x2A\x2D\x00\x80\x36\x2D\x00\x80\x76\x2D"
-     "\x00\x80"
-     "\x2A\x2E\x00\x80\x46\x2E\x00\x80\x4E\x2E\x00\x80\x8A\x2E\x00\x80\x96\x2E"
-     "\x00\x80"
-     "\x79\x2F\x00\x80\x90\x30\x00\x80\xB8\x30\x00\x80\xC0\x30\x00\x80\x80\x41"
-     "\x03\x00"
-     "\x81\x41\x03\x00\x82\x41\x03\x00\x83\x41\x03\x00\x84\x41\x03\x00\x85\x41"
-     "\x03\x00"
-     "\x86\x41\x03\x00\x87\x41\x03\x00\x88\xC9\x2B\x00\x89\x45\x03\x00\x8A\xD1"
-     "\x2B\x00"
-     "\x8B\x49\x03\x00\x8C\x59\x03\x00\x8D\x59\x03\x00\xD8\x30\x00\x80\x84\x31"
-     "\x00\x80"
-     "\x90\x31\x03\x00\xB4\x31\x00\x80\x92\x3D\x03\x00\xE2\x31\x00\x80\x5F\x32"
-     "\x00\x80"
-     "\x95\x2D\x03\x00\x96\x25\x03\x00\x97\xDD\x03\x00\x98\xE5\x03\x00\x99\xED"
-     "\x03\x00"
-     "\x9A\xE5\x03\x00\x9B\xFD\x03\x00\x9C\xE5\x03\x00\x9D\xED\x03\x00\x9E\xE1"
-     "\x2B\x00"
-     "\x23\x33\x00\x80\xA0\x1D\x03\x00\x3F\x33\x00\x80\xA2\x29\x03\x00\x6F\x33"
-     "\x00\x80"
-     "\xC3\x33\x00\x80\xA5\x31\x03\x00\xA6\x31\x03\x00\xE3\x33\x00\x80\x57\x34"
-     "\x00\x80"
-     "\x37\x35\x00\x80\xAA\x05\x03\x00\xAB\x1D\x03\x00\xAC\x05\x03\x00\xAD\x0D"
-     "\x03\x00"
-     "\xAE\x05\x03\x00\xAF\x7D\x03\x00\xB0\x05\x03\x00\xB1\x0D\x03\x00\xB2\x05"
-     "\x03\x00"
-     "\xB3\x1D\x03\x00\xB4\x05\x03\x00\xB5\x0D\x03\x00\xB6\x05\x03\x00\xB7\x3D"
-     "\x03\x00"
-     "\xB8\x05\x03\x00\xB9\x0D\x03\x00\xBA\x05\x03\x00\xBB\x1D\x03\x00\xBC\xE5"
-     "\x2D\x00"
-     "\xBD\x01\x03\x00\xBE\x01\x03\x00\xBF\x01\x03\x00\x5B\x35\x00\x80\x57\x35"
-     "\x00\x80"
-     "\x5F\x35\x00\x80\x63\x35\x00\x80\x7B\x35\x00\x80\x7F\x35\x00\x80\xA3\x35"
-     "\x00\x80"
-     "\xB7\x35\x00\x80\xD3\x35\x00\x80\x3F\x36\x00\x80\x47\x36\x00\x80\x5B\x36"
-     "\x00\x80"
-     "\x83\x36\x00\x80\xD3\x36\x00\x80\x07\x37\x00\x80\x6F\x37\x00\x80\x17\x38"
-     "\x00\x80"
-     "\x2F\x38\x00\x80\x63\x38\x00\x80\x67\x38\x00\x80\xAB\x38\x00\x80\xAF\x38"
-     "\x00\x80"
-     "\x07\x39\x00\x80\xE3\x39\x00\x80\x27\x3A\x00\x80\x3F\x3A\x00\x80\x88\x30"
-     "\x00\x80"
-     "\xD0\xEB\x00\x80\xEB\x36\x00\x80\xB7\x2B\x00\x80\xC2\x2C\x00\x80\x82\x2C"
-     "\x00\x80"
-     "\x2A\x2C\x00\x80\x6A\x2C\x00\x80\x96\x2C\x00\x80\x26\x2D\x00\x80\x42\x2D"
-     "\x00\x80"
-     "\x16\x2E\x00\x80\x26\x2E\x00\x80\x42\x2E\x00\x80\x9E\x2E\x00\x80\xE9\x2E"
-     "\x00\x80"
-     "\xF5\x2E\x00\x80\x11\x2F\x00\x80\x21\x2F\x00\x80\x10\x30\x00\x80\x18\x30"
-     "\x00\x80"
-     "\x48\x30\x00\x80\x68\x30\x00\x80\x98\x30\x00\x80\xB0\x30\x00\x80\xA0\x30"
-     "\x00\x80"
-     "\xB4\x30\x00\x80\xF4\x30\x00\x80\x4C\x31\x00\x80\x58\x31\x00\x80\x60\x31"
-     "\x00\x80"
-     "\x88\x31\x00\x80\x0E\x32\x00\x80\x22\x32\x00\x80\x63\x33\x00\x80\x5B\x33"
-     "\x00\x80"
-     "\x8B\x33\x00\x80\x1B\x35\x00\x80\x99\x29\x1F\x00\x98\x21\x1F\x00\x9B\x39"
-     "\x1F\x00"
-     "\x9A\x21\x1F\x00\x9D\x29\x1F\x00\x9C\x21\x1F\x00\x9F\x25\x1F\x00\x9E\x25"
-     "\x1F\x00"
-     "\x91\x39\x27\x00\x90\x29\x1F\x00\x93\x3D\x1F\x00\x92\x3D\x1F\x00\x95\x2D"
-     "\x1F\x00"
-     "\x94\x2D\x1F\x00\x97\x81\x19\x00\x96\x1D\x1F\x00\x89\xDD\x1E\x00\x88\xD5"
-     "\x1E\x00"
-     "\x8B\xED\x1E\x00\x8A\xED\x1E\x00\x8D\x51\x1F\x00\x8C\x5D\x27\x00\x8F\x51"
-     "\x1F\x00"
-     "\x8E\x59\x1F\x00\x81\x5D\x1E\x00\x80\x5D\x1E\x00\x83\x25\x1E\x00\x82\x29"
-     "\x1E\x00"
-     "\x85\x0D\x1E\x00\x84\x25\x1E\x00\x87\xF5\x1E\x00\x86\xF9\x1E\x00\xB9\x15"
-     "\x1F\x00"
-     "\xB8\x59\x1D\x00\xBB\x15\x1F\x00\xBA\x15\x1F\x00\xBD\x31\x1F\x00\xBC\xE5"
-     "\x1D\x00"
-     "\xBF\x31\x1F\x00\xBE\x39\x1F\x00\xB1\xD9\x1F\x00\xB0\xD1\x1F\x00\xB3\x29"
-     "\x1F\x00"
-     "\xB2\xD1\x1F\x00\xB5\x39\x1F\x00\xB4\x31\x1F\x00\xB7\x29\x1F\x00\xB6\x31"
-     "\x1F\x00"
-     "\xA9\xA1\x1F\x00\xA8\x99\x1F\x00\xAB\xA1\x1F\x00\xAA\xA9\x1F\x00\xAD\xA1"
-     "\x1F\x00"
-     "\xAC\xB9\x1F\x00\xE7\x33\x00\x80\x1B\x34\x00\x80\xA1\xA5\x1F\x00\xA0\xA5"
-     "\x1F\x00"
-     "\xA3\xA5\x1F\x00\xA2\xA5\x1F\x00\xA5\xA5\x1F\x00\xA4\xA5\x1F\x00\xA7\x5D"
-     "\x1D\x00"
-     "\xA6\xA5\x1F\x00\x80\xAD\x1E\x00\x81\xB5\x1E\x00\x82\xBD\x1E\x00\x83\xB5"
-     "\x1E\x00"
-     "\x84\xAD\x1E\x00\x85\xD5\x1E\x00\x86\xDD\x1E\x00\x87\xD5\x1E\x00\x88\xED"
-     "\x1E\x00"
-     "\x89\xD9\x1C\x00\x8A\xF9\x1E\x00\x8B\x15\x1F\x00\x8C\xE5\x1E\x00\x8D\xED"
-     "\x1E\x00"
-     "\x8E\xE5\x1E\x00\x8F\xDD\x1E\x00\x90\xA5\x1E\x00\x91\x31\x1C\x00\x92\xE9"
-     "\x14\x00"
-     "\x93\xA5\x1E\x00\x94\xBD\x1E\x00\x95\x85\x32\x00\x96\x75\x18\x00\x97\x25"
-     "\x1A\x00"
-     "\x98\x91\x1E\x00\x99\x91\x1E\x00\x9A\xE5\x1F\x00\x9B\x95\x1E\x00\x9C\xE1"
-     "\x1F\x00"
-     "\x9D\x89\x1E\x00\x9E\xA1\x01\x00\x9F\xA1\x01\x00\xA0\xFD\x1C\x00\xA1\x85"
-     "\x03\x00"
-     "\xA2\x8D\x03\x00\xA3\x85\x03\x00\xA4\x9D\x03\x00\xA5\x85\x03\x00\xA6\x19"
-     "\x1C\x00"
-     "\xA7\x89\x03\x00\xA8\xB9\x03\x00\xA9\xB9\x03\x00\xAA\x89\x03\x00\xAB\x89"
-     "\x03\x00"
-     "\xAC\x99\x03\x00\xAD\x95\x1E\x00\xAE\x95\x03\x00\xAF\x8D\x03\x00\xB0\x79"
-     "\x18\x00"
-     "\xB1\xF1\x03\x00\xB2\xF9\x1E\x00\xB3\xF5\x03\x00\xB4\xED\x03\x00\xB5\x95"
-     "\x03\x00"
-     "\xB6\x9D\x03\x00\xB7\x95\x03\x00\xB8\xA5\x1E\x00\xB9\xA9\x03\x00\xBA\x49"
-     "\x1C\x00"
-     "\xBB\xBD\x03\x00\xBC\xB9\x1E\x00\xBD\x15\x17\x00\xBE\xAD\x03\x00\xBF\xB1"
-     "\x1E\x00"
-     "\x67\x34\x00\x80\x83\x34\x00\x80\xA7\x34\x00\x80\xCB\x34\x00\x80\xCF\x34"
-     "\x00\x80"
-     "\xEF\x34\x00\x80\xFF\x34\x00\x80\x13\x35\x00\x80\x0F\x35\x00\x80\x3F\x35"
-     "\x00\x80"
-     "\xBB\x35\x00\x80\xEB\x35\x00\x80\x23\x36\x00\x80\x53\x36\x00\x80\x37\x37"
-     "\x00\x80"
-     "\x5F\x37\x00\x80\xD3\x37\x00\x80\x1B\x38\x00\x80\x27\x38\x00\x80\x47\x38"
-     "\x00\x80"
-     "\x4B\x38\x00\x80\x57\x38\x00\x80\x73\x38\x00\x80\xEB\x38\x00\x80\x1B\x39"
-     "\x00\x80"
-     "\xF0\x20\x03\x00\x53\x39\x00\x80\x73\x39\x00\x80\xBF\x39\x00\x80\x17\x3A"
-     "\x00\x80"
-     "\x33\x3A\x00\x80\x93\x3A\x00\x80\x80\xD9\x00\x00\x81\xD1\x1D\x00\x82\xD5"
-     "\x00\x00"
-     "\x83\xED\x00\x00\x84\xF9\x00\x00\x85\xF9\x00\x00\x86\xE9\x00\x00\x87\xC1"
-     "\x1D\x00"
-     "\x88\x35\x1F\x00\x89\xD1\x00\x00\x8A\xC9\x1D\x00\x8B\xD5\x00\x00\x8C\xD5"
-     "\x1D\x00"
-     "\x8D\xC9\x00\x00\x8E\x45\x3A\x00\x8F\x55\x00\x00\x90\x2D\x00\x00\x91\x35"
-     "\x00\x00"
-     "\x92\x3D\x00\x00\x93\x35\x00\x00\x94\x2D\x00\x00\x95\xD1\x01\x00\x96\xD1"
-     "\x01\x00"
-     "\x97\xD1\x01\x00\x98\xF1\x01\x00\x99\xF1\x01\x00\x67\xEB\x00\x80\x62\xEB"
-     "\x00\x80"
-     "\x94\xEB\x00\x80\xDF\x2A\x00\x80\x07\x2B\x00\x80\x0B\x2B\x00\x80\x91\x18"
-     "\x0F\x00"
-     "\x16\xEC\x00\x80\x43\xEC\x00\x80\xBB\xEC\x00\x80\x1F\x3B\x00\x80\x2F\x3B"
-     "\x00\x80"
-     "\x4A\x08\x00\x80\x55\x08\x00\x80\x80\x39\x00\x00\x81\x39\x00\x00\x82\x15"
-     "\x00\x00"
-     "\x83\x6D\x00\x00\x84\x75\x00\x00\x85\x7D\x00\x00\x86\x71\x00\x00\x58\x08"
-     "\x00\x80"
-     "\x4D\x08\x00\x80\x51\x08\x00\x80\x5A\x09\x00\x80\xE7\x0F\x00\x80\xD8\x0F"
-     "\x00\x80"
-     "\x9D\x3D\x00\x00\xDD\x0F\x00\x80\x9F\x89\x00\x00\xEC\x0F\x00\x80\xE2\x0F"
-     "\x00\x80"
-     "\x41\x10\x00\x80\x93\x21\x00\x00\x94\x21\x00\x00\x95\x2D\x00\x00\x96\x19"
-     "\x00\x00"
-     "\x97\x19\x00\x00\xA8\x8D\x07\x00\xA9\x2D\x45\x00\xAA\x99\x07\x00\xAB\x99"
-     "\x07\x00"
-     "\xAC\x89\x07\x00\xAD\x89\x07\x00\xAE\xB9\x07\x00\xAF\xB9\x07\x00\xA0\x7D"
-     "\x00\x00"
-     "\xA1\x51\x45\x00\xA2\x69\x44\x00\xA3\x51\x00\x00\xA4\x01\x04\x00\xA5\xB5"
-     "\x07\x00"
-     "\xA6\xBD\x07\x00\xA7\xB5\x07\x00\xB8\x85\x07\x00\xB9\x8D\x07\x00\xBA\x85"
-     "\x07\x00"
-     "\xBB\x9D\x07\x00\xBC\x85\x07\x00\xC0\x10\x00\x80\xBE\x81\x07\x00\x77\x10"
-     "\x00\x80"
-     "\xB0\xC9\x07\x00\xB1\xC9\x07\x00\xB2\xD9\x07\x00\xB3\xD9\x07\x00\xB4\xC9"
-     "\x07\x00"
-     "\xB5\xC9\x07\x00\xB6\xB9\x07\x00\x25\x10\x00\x80\x5D\x10\x00\x80\x65\x10"
-     "\x00\x80"
-     "\x93\x10\x00\x80\xB8\x10\x00\x80\xAE\x10\x00\x80\xB3\x10\x00\x80\xA0\x10"
-     "\x00\x80"
-     "\xA7\x10\x00\x80\xF4\x0F\x00\x80\xF9\x0F\x00\x80\xFE\x0F\x00\x80\x0B\x10"
-     "\x00\x80"
-     "\x18\x10\x00\x80\x20\x10\x00\x80\x28\x10\x00\x80\x32\x10\x00\x80\x37\x10"
-     "\x00\x80"
-     "\x3C\x10\x00\x80\x46\x10\x00\x80\x4B\x10\x00\x80\x53\x10\x00\x80\x60\x10"
-     "\x00\x80"
-     "\x68\x10\x00\x80\x6D\x10\x00\x80\x72\x10\x00\x80\x7A\x10\x00\x80\x7F\x10"
-     "\x00\x80"
-     "\x89\x10\x00\x80\x8E\x10\x00\x80\x96\x10\x00\x80\x9B\x10\x00\x80\xBB\x10"
-     "\x00\x80"
-     "\x2D\x10\x00\x80\x10\x10\x00\x80\x58\x10\x00\x80\x84\x10\x00\x80\x03\x10"
-     "\x00\x80"
-     "\xD3\x17\x00\x80\xDC\x17\x00\x80\xDF\x17\x00\x80\xE5\x17\x00\x80\xD9\x17"
-     "\x00\x80"
-     "\xE2\x17\x00\x80\xD6\x17\x00\x80\x09\x18\x00\x80\x0C\x18\x00\x80\xEB\x17"
-     "\x00\x80"
-     "\xE8\x17\x00\x80\xEE\x17\x00\x80\xF1\x17\x00\x80\xD9\x74\x53\x02\xFA\x17"
-     "\x00\x80"
-     "\xF7\x17\x00\x80\xFD\x17\x00\x80\xF4\x17\x00\x80\x03\x18\x00\x80\x00\x18"
-     "\x00\x80"
-     "\x0F\x18\x00\x80\x15\x18\x00\x80\x1B\x18\x00\x80\x18\x18\x00\x80\x1E\x18"
-     "\x00\x80"
-     "\x21\x18\x00\x80\x2A\x18\x00\x80\x80\x5D\x01\x00\x81\x65\x01\x00\x24\x18"
-     "\x00\x80"
-     "\x83\x69\x01\x00\x84\x79\x01\x00\x12\x18\x00\x80\x86\x75\x01\x00\x87\x6D"
-     "\x01\x00"
-     "\x88\x55\x01\x00\x89\x5D\x01\x00\x8A\x55\x01\x00\x8B\xAD\x01\x00\x8C\xB5"
-     "\x01\x00"
-     "\x8D\xBD\x01\x00\x8E\xB5\x01\x00\x8F\xAD\x01\x00\x90\xD5\x01\x00\x91\xD1"
-     "\x01\x00"
-     "\x92\xD1\x01\x00\x93\xD5\x01\x00\x94\xC9\x01\x00\x95\xCD\x01\x00\x96\xC5"
-     "\x01\x00"
-     "\x97\xC1\x01\x00\x98\xFD\x01\x00\x99\xF9\x01\x00\x9A\xC9\x01\x00\x9B\xCD"
-     "\x01\x00"
-     "\x9C\xD1\x01\x00\x9D\xD5\x01\x00\x9E\xDD\x01\x00\x9F\xD9\x01\x00\xA0\x25"
-     "\x01\x00"
-     "\xA1\x21\x01\x00\xA2\x21\x01\x00\xA3\x25\x01\x00\xA4\x39\x01\x00\xA5\x3D"
-     "\x01\x00"
-     "\xA6\x35\x01\x00\xA7\x31\x01\x00\xA8\x0D\x01\x00\xA9\x09\x01\x00\xAA\x19"
-     "\x01\x00"
-     "\xAB\x1D\x01\x00\xAC\x01\x01\x00\xAD\x05\x01\x00\xAE\x0D\x01\x00\xAF\x09"
-     "\x01\x00"
-     "\xB0\x75\x01\x00\xB1\x71\x01\x00\xB2\x71\x01\x00\xB3\x75\x01\x00\xB4\x69"
-     "\x01\x00"
-     "\xB5\x6D\x01\x00\xB6\x65\x01\x00\xB7\x61\x01\x00\xB8\x5D\x01\x00\xB9\x59"
-     "\x01\x00"
-     "\xBA\x29\x01\x00\xBB\x2D\x01\x00\xBC\x31\x01\x00\xBD\x35\x01\x00\xBE\x3D"
-     "\x01\x00"
-     "\xBF\x39\x01\x00\x36\x18\x00\x80\x33\x18\x00\x80\xB0\xE1\x53\x02\xB1\xE5"
-     "\x53\x02"
-     "\x3E\x18\x00\x80\x44\x18\x00\x80\x30\x18\x00\x80\x41\x18\x00\x80\x4A\x18"
-     "\x00\x80"
-     "\x6F\x17\x00\x80\xD7\x10\x00\x80\x27\x11\x00\x80\x04\x11\x00\x80\x18\x11"
-     "\x00\x80"
-     "\x13\x11\x00\x80\x1D\x11\x00\x80\xA2\xB9\x02\x00\xA3\xBD\x02\x00\xA0\xB1"
-     "\x02\x00"
-     "\xA1\xB5\x02\x00\xA6\xAD\x02\x00\xA7\xA9\x02\x00\xA4\xA1\x53\x02\xA5\xA5"
-     "\x53\x02"
-     "\xAA\x69\x03\x00\xAB\x6D\x03\x00\xA8\x95\x02\x00\xA9\x91\x02\x00\xAE\xDD"
-     "\x01\x00"
-     "\xAF\xD9\x01\x00\xAC\x71\x03\x00\xAD\x75\x03\x00\x92\x69\x02\x00\x93\x6D"
-     "\x02\x00"
-     "\x90\x6D\x02\x00\x91\x69\x02\x00\x96\x7D\x02\x00\x97\x79\x02\x00\x94\x71"
-     "\x02\x00"
-     "\x95\x75\x02\x00\x9A\x51\x02\x00\x9B\x55\x02\x00\x98\x45\x02\x00\x99\x41"
-     "\x02\x00"
-     "\x9E\x45\x02\x00\x9F\x41\x02\x00\x9C\x49\x02\x00\x9D\x4D\x02\x00\x82\x0D"
-     "\x02\x00"
-     "\x83\x09\x02\x00\x80\x0D\x02\x00\x81\x09\x02\x00\x86\x05\x02\x00\x87\x01"
-     "\x02\x00"
-     "\x84\x09\x02\x00\x85\x0D\x02\x00\x8A\x3D\x02\x00\x8B\x39\x02\x00\x88\x31"
-     "\x02\x00"
-     "\x89\x35\x02\x00\x8E\x15\x02\x00\x8F\x11\x02\x00\x8C\x19\x02\x00\x8D\x1D"
-     "\x02\x00"
-     "\xA3\x1D\x01\x00\xA2\x19\x01\x00\xA1\x11\x01\x00\xA0\x15\x01\x00\xA7\x31"
-     "\x01\x00"
-     "\xA6\x35\x01\x00\xA5\x39\x01\x00\xA4\x3D\x01\x00\xAB\x09\x01\x00\xAA\x0D"
-     "\x01\x00"
-     "\xA9\x0D\x01\x00\xA8\x09\x01\x00\xAF\x01\x01\x00\xAE\x05\x01\x00\xAD\x15"
-     "\x01\x00"
-     "\xAC\x11\x01\x00\xB3\x79\x01\x00\xB2\x7D\x01\x00\xB1\x7D\x01\x00\xB0\x79"
-     "\x01\x00"
-     "\xB7\xD1\x00\x00\xB6\xD5\x00\x00\xB5\x65\x01\x00\xB4\x61\x01\x00\xBB\xE5"
-     "\x00\x00"
-     "\xBA\xE1\x00\x00\xB9\xED\x00\x00\xB8\xED\x00\x00\xBF\xF9\x00\x00\xBE\xFD"
-     "\x00\x00"
-     "\xBD\xF1\x00\x00\xBC\xF5\x00\x00\x22\x11\x00\x80\x09\x11\x00\x80\x47\x18"
-     "\x00\x80"
-     "\xDC\x10\x00\x80\xE1\x10\x00\x80\xF5\x10\x00\x80\xD9\x28\x50\x02\x0E\x11"
-     "\x00\x80"
-     "\x4E\x11\x00\x80\x53\x11\x00\x80\x5F\x11\x00\x80\x4D\x18\x00\x80\x50\x10"
-     "\x00\x80"
-     "\x75\x11\x00\x80\x84\x11\x00\x80\x89\x11\x00\x80\x93\xD5\x02\x00\x94\x11"
-     "\x00\x80"
-     "\xAE\x11\x00\x80\xC1\x11\x00\x80\x97\xD1\x01\x00\x96\xC1\x02\x00\x95\xCD"
-     "\x02\x00"
-     "\x94\xC9\x02\x00\x9B\xF1\x01\x00\x9A\xE1\x01\x00\x99\xED\x01\x00\x98\xED"
-     "\x01\x00"
-     "\x9F\xE5\x01\x00\x9E\xE1\x01\x00\x9D\x31\x52\x02\x9C\xED\x01\x00\x80\x8D"
-     "\x0F\x00"
-     "\x81\x95\x0F\x00\x82\x9D\x0F\x00\x83\x89\x0F\x00\x84\x8D\x0F\x00\x85\xB5"
-     "\x0F\x00"
-     "\x86\xBD\x0F\x00\x87\xB5\x0F\x00\x88\x95\x0F\x00\x89\x9D\x0F\x00\x8A\x95"
-     "\x0F\x00"
-     "\x8B\xE9\x0F\x00\x8C\xF9\x0F\x00\x8D\xF9\x0F\x00\x8E\x39\x01\x00\x8F\x39"
-     "\x01\x00"
-     "\x90\x49\x01\x00\x91\x49\x01\x00\x92\x59\x01\x00\x93\x59\x01\x00\x94\x49"
-     "\x01\x00"
-     "\x95\x49\x01\x00\x96\x79\x01\x00\x97\x79\x01\x00\x98\x49\x01\x00\x99\x49"
-     "\x01\x00"
-     "\x9A\x59\x01\x00\x9B\x59\x01\x00\x9C\x49\x01\x00\x9D\x49\x01\x00\x9E\x39"
-     "\x01\x00"
-     "\x9F\x39\x01\x00\xA0\xC9\x01\x00\xA1\xC9\x01\x00\xA2\xD9\x01\x00\xA3\xD9"
-     "\x01\x00"
-     "\xA4\xC9\x01\x00\xA5\xC9\x01\x00\xA6\xF9\x01\x00\xA7\xF9\x01\x00\xA8\xC9"
-     "\x01\x00"
-     "\xA9\xC9\x01\x00\xAA\xD9\x01\x00\xAB\xD9\x01\x00\xAC\xC9\x01\x00\xAD\xC9"
-     "\x01\x00"
-     "\xAE\x39\x01\x00\xAF\x39\x01\x00\xB0\x49\x01\x00\xB1\x49\x01\x00\xB2\x59"
-     "\x01\x00"
-     "\xB3\x59\x01\x00\xB4\x49\x01\x00\xB5\x49\x01\x00\xB6\x79\x01\x00\xB7\x79"
-     "\x01\x00"
-     "\xB8\x49\x01\x00\xB9\x49\x01\x00\xBA\x59\x01\x00\xBB\x59\x01\x00\xBC\x49"
-     "\x01\x00"
-     "\xBD\x49\x01\x00\xBE\x39\x01\x00\xBF\x39\x01\x00\xE5\x11\x00\x80\x17\x12"
-     "\x00\x80"
-     "\x1C\x12\x00\x80\x24\x12\x00\x80\x33\x12\x00\x80\x42\x12\x00\x80\x47\x12"
-     "\x00\x80"
-     "\x4F\x12\x00\x80\x72\x12\x00\x80\x99\x12\x00\x80\xA5\x12\x00\x80\xC5\x12"
-     "\x00\x80"
-     "\xCA\x12\x00\x80\xCF\x12\x00\x80\x09\x13\x00\x80\x1C\x13\x00\x80\x28\x13"
-     "\x00\x80"
-     "\x40\x13\x00\x80\xBA\x13\x00\x80\x10\x14\x00\x80\x29\x14\x00\x80\x2E\x14"
-     "\x00\x80"
-     "\x41\x14\x00\x80\x52\x14\x00\x80\x64\x14\x00\x80\x69\x14\x00\x80\x90\x14"
-     "\x00\x80"
-     "\x98\x14\x00\x80\xAD\x14\x00\x80\xD4\x14\x00\x80\xD9\x14\x00\x80\x33\x15"
-     "\x00\x80"
-     "\x38\x15\x00\x80\x3D\x15\x00\x80\x49\x15\x00\x80\x55\x15\x00\x80\x5A\x15"
-     "\x00\x80"
-     "\x62\x15\x00\x80\x6E\x15\x00\x80\x88\x15\x00\x80\x8D\x15\x00\x80\x95\x15"
-     "\x00\x80"
-     "\x9A\x15\x00\x80\x9F\x15\x00\x80\xA4\x15\x00\x80\xA9\x15\x00\x80\xAE\x15"
-     "\x00\x80"
-     "\xC1\x15\x00\x80\xC6\x15\x00\x80\xE2\x15\x00\x80\xFC\x15\x00\x80\x16\x16"
-     "\x00\x80"
-     "\x22\x16\x00\x80\x3A\x16\x00\x80\x3F\x16\x00\x80\x4C\x16\x00\x80\x6D\x16"
-     "\x00\x80"
-     "\x90\x16\x00\x80\xAA\x16\x00\x80\xB6\x16\x00\x80\xBB\x16\x00\x80\xC3\x16"
-     "\x00\x80"
-     "\xE4\x16\x00\x80\xFE\x16\x00\x80\x80\xC9\x03\x00\x81\xC9\x03\x00\x82\xD9"
-     "\x03\x00"
-     "\x83\xD9\x03\x00\x84\xC9\x03\x00\x85\xC9\x03\x00\x86\xF9\x03\x00\x87\xF9"
-     "\x03\x00"
-     "\x88\xC9\x03\x00\x89\xC9\x03\x00\x8A\xD9\x03\x00\x8B\xD9\x03\x00\x8C\xC9"
-     "\x03\x00"
-     "\x8D\xC9\x03\x00\x8E\x39\x01\x00\x8F\x39\x01\x00\x90\x49\x01\x00\x91\x49"
-     "\x01\x00"
-     "\x92\x59\x01\x00\x93\x59\x01\x00\x94\x49\x01\x00\x95\x49\x01\x00\x96\x79"
-     "\x01\x00"
-     "\x97\x79\x01\x00\x98\x49\x01\x00\x99\x49\x01\x00\x9A\x59\x01\x00\x9B\x59"
-     "\x01\x00"
-     "\x9C\x49\x01\x00\x9D\x49\x01\x00\x9E\x39\x01\x00\x9F\x39\x01\x00\xA0\xC9"
-     "\x01\x00"
-     "\xA1\xC9\x01\x00\xA2\xD9\x01\x00\xA3\xD9\x01\x00\xA4\xC9\x01\x00\xA5\xC9"
-     "\x01\x00"
-     "\xA6\x0D\x0D\x00\xA7\xFD\x01\x00\xA8\x25\x0D\x00\xA9\x39\x0D\x00\xAA\xCD"
-     "\x01\x00"
-     "\xAB\xC5\x01\x00\xAC\x05\x0D\x00\xAD\xD9\x01\x00\xAE\x01\x0D\x00\xAF\x01"
-     "\x0D\x00"
-     "\xB0\xB1\x01\x00\xB1\xB1\x01\x00\xB2\xC9\x03\x00\xB3\xB5\x01\x00\xB4\xD5"
-     "\x03\x00"
-     "\xB5\xDD\x03\x00\xB6\xA5\x01\x00\xB7\x5D\x01\x00\xB8\xF1\x03\x00\xB9\x61"
-     "\x01\x00"
-     "\xBA\xFD\x03\x00\xBB\xF5\x03\x00\xBC\x7D\x03\x00\xBD\x65\x03\x00\xBE\x65"
-     "\x03\x00"
-     "\xBF\x5D\x03\x00\x0D\x17\x00\x80\x2A\x17\x00\x80\x2F\x17\x00\x80\x37\x17"
-     "\x00\x80"
-     "\x3C\x17\x00\x80\x4F\x17\x00\x80\x54\x17\x00\x80\x7A\x17\x00\x80\x86\x17"
-     "\x00\x80"
-     "\x92\x17\x00\x80\xA1\x17\x00\x80\xBE\x17\x00\x80\xC3\x17\x00\x80\xE4\x12"
-     "\x00\x80"
-     "\xF5\x12\x00\x80\x72\x17\x00\x80\x81\x00\x00\x80\x8B\x00\x00\x80\x95\x00"
-     "\x00\x80"
-     "\x9F\x00\x00\x80\xA9\x00\x00\x80\xB3\x00\x00\x80\xEB\x10\x00\x80\xF0\x10"
-     "\x00\x80"
-     "\xFA\x10\x00\x80\x6B\x11\x00\x80\x70\x11\x00\x80\x7A\x11\x00\x80\xDB\x11"
-     "\x00\x80"
-     "\xE0\x11\x00\x80\x0D\x12\x00\x80\x29\x12\x00\x80\x2E\x12\x00\x80\x38\x12"
-     "\x00\x80"
-     "\x47\x16\x00\x80\x03\x17\x00\x80\x08\x17\x00\x80\x20\x17\x00\x80\x97\x17"
-     "\x00\x80"
-     "\x9C\x17\x00\x80\xB4\x17\x00\x80\xE6\x10\x00\x80\xFF\x10\x00\x80\x7F\x11"
-     "\x00\x80"
-     "\x12\x12\x00\x80\xCD\x13\x00\x80\x35\x16\x00\x80\x25\x17\x00\x80\x59\x17"
-     "\x00\x80"
-     "\xB9\x17\x00\x80\x3D\x12\x00\x80\x5A\x13\x00\x80\x95\x13\x00\x80\xA8\x13"
-     "\x00\x80"
-     "\x07\x15\x00\x80\x13\x15\x00\x80\x1F\x15\x00\x80\x83\x14\x00\x80\x88\x14"
-     "\x00\x80"
-     "\xC7\x14\x00\x80\xCC\x14\x00\x80\xF3\x14\x00\x80\xF8\x14\x00\x80\x5F\x13"
-     "\x00\x80"
-     "\x82\xB9\x05\x00\x83\xB9\x05\x00\x80\xAD\x05\x00\x81\xB9\x05\x00\x86\xC5"
-     "\x05\x00"
-     "\x87\xCD\x05\x00\x84\xD9\x05\x00\x85\xCD\x05\x00\x8A\xAD\x07\x00\x8B\xB5"
-     "\x07\x00"
-     "\x88\xA1\x07\x00\x89\xC5\x05\x00\x8E\x35\x07\x00\x8F\x3D\x07\x00\x8C\x39"
-     "\x07\x00"
-     "\x8D\xA9\x07\x00\x92\xDD\x07\x00\x93\x6D\x07\x00\x90\x75\x07\x00\x91\xD5"
-     "\x07\x00"
-     "\x96\x61\x07\x00\x97\xD9\x0B\x00\x94\xF9\x07\x00\x95\x71\x07\x00\x9A\xF5"
-     "\x0B\x00"
-     "\x9B\xCD\x07\x00\x98\xF9\x0B\x00\x99\xC9\x07\x00\x9E\xD9\x0B\x00\x9F\xC1"
-     "\x0B\x00"
-     "\x9C\xD9\x0B\x00\x9D\xD9\x0B\x00\xA2\x49\x0B\x00\xA3\x49\x0B\x00\xA0\x25"
-     "\x07\x00"
-     "\xA1\x49\x0B\x00\xA6\x81\x05\x00\xA7\x89\x05\x00\xA4\x99\x05\x00\xA5\x2D"
-     "\x07\x00"
-     "\xAA\x89\x05\x00\xAB\x89\x05\x00\xA8\x89\x05\x00\xA9\x89\x05\x00\xAE\x8D"
-     "\x05\x00"
-     "\xAF\xF5\x05\x00\xAC\x8D\x05\x00\xAD\x85\x05\x00\xB2\x75\x07\x00\xB3\x89"
-     "\x05\x00"
-     "\xB0\x8D\x05\x00\xB1\x85\x05\x00\xB6\x89\x05\x00\xB7\x89\x05\x00\xB4\x89"
-     "\x05\x00"
-     "\xB5\x89\x05\x00\xBA\x8D\x05\x00\xBB\x95\x05\x00\xB8\x89\x05\x00\xB9\x85"
-     "\x05\x00"
-     "\xBE\x8D\x05\x00\xBF\x75\x05\x00\xBC\x8D\x05\x00\xBD\x85\x05\x00\x80\x85"
-     "\x04\x00"
-     "\x81\x8D\x04\x00\x82\x9D\x04\x00\x83\x95\x04\x00\x84\xB9\x04\x00\x85\xB9"
-     "\x04\x00"
-     "\x86\xA9\x04\x00\x87\xA9\x04\x00\x88\x99\x04\x00\x89\xE1\x04\x00\x8A\xE1"
-     "\x04\x00"
-     "\x8B\xE1\x04\x00\x8C\xE1\x04\x00\x8D\x8D\x06\x00\x8E\xE5\x04\x00\x8F\xDD"
-     "\x04\x00"
-     "\x90\xA5\x04\x00\x91\xAD\x04\x00\x92\xBD\x04\x00\x93\xB5\x04\x00\x94\xAD"
-     "\x04\x00"
-     "\x95\x55\x06\x00\x96\xE5\x06\x00\x97\x51\x06\x00\x98\x71\x06\x00\x99\xA5"
-     "\x06\x00"
-     "\x9A\x75\x06\x00\x9B\x4D\x06\x00\x9C\x55\x06\x00\x9D\x5D\x06\x00\x9E\xBD"
-     "\x06\x00"
-     "\x9F\xE9\x0A\x00\xA0\x29\x06\x00\xA1\x31\x0A\x00\xA2\x25\x06\x00\xA3\x8D"
-     "\x04\x00"
-     "\xA4\x21\x06\x00\xA5\x85\x04\x00\xA6\x51\x06\x00\xA7\xD9\x04\x00\xA8\x6D"
-     "\x06\x00"
-     "\xA9\x75\x06\x00\xAA\x7D\x06\x00\xAB\x19\x04\x00\xAC\x09\x04\x00\xAD\x65"
-     "\x04\x00"
-     "\xAE\xB9\x06\x00\xAF\x01\x06\x00\xB0\xE9\x06\x00\xB1\x01\x06\x00\xB2\x11"
-     "\x06\x00"
-     "\xB3\x11\x06\x00\xB4\x31\x06\x00\xB5\x31\x06\x00\xB6\x31\x06\x00\xB7\x31"
-     "\x06\x00"
-     "\xB8\x11\x06\x00\xB9\x11\x06\x00\xBA\x11\x06\x00\xBB\x11\x06\x00\xBC\xF1"
-     "\x03\x00"
-     "\xBD\xF1\x03\x00\xBE\xF5\x03\x00\xBF\xED\x03\x00\x64\x13\x00\x80\xAD\x13"
-     "\x00\x80"
-     "\x84\x80\xFE\x00\xB2\x13\x00\x80\xB8\x12\x00\x80\xBD\x12\x00\x80\x8C\x12"
-     "\x00\x80"
-     "\x91\x12\x00\x80\xD4\x12\x00\x80\xD9\x12\x00\x80\x1C\x14\x00\x80\x21\x14"
-     "\x00\x80"
-     "\x57\x14\x00\x80\x5C\x14\x00\x80\x6C\x13\x00\x80\x78\x13\x00\x80\x8B\x13"
-     "\x00\x80"
-     "\x90\x13\x00\x80\x3B\x13\x00\x80\xD2\x13\x00\x80\x4D\x14\x00\x80\x41\x11"
-     "\x00\x80"
-     "\x99\x11\x00\x80\xB3\x11\x00\x80\xBA\x11\x00\x80\xC6\x11\x00\x80\xEA\x11"
-     "\x00\x80"
-     "\xF1\x11\x00\x80\xF8\x11\x00\x80\x77\x12\x00\x80\xB1\x12\x00\x80\xAA\x12"
-     "\x00\x80"
-     "\x21\x13\x00\x80\x0E\x13\x00\x80\x15\x13\x00\x80\x4C\x13\x00\x80\x45\x13"
-     "\x00\x80"
-     "\x53\x13\x00\x80\xBF\x13\x00\x80\x15\x14\x00\x80\x7D\x13\x00\x80\x71\x13"
-     "\x00\x80"
-     "\x9A\x13\x00\x80\xA1\x13\x00\x80\x33\x14\x00\x80\x46\x14\x00\x80\x6E\x14"
-     "\x00\x80"
-     "\x75\x14\x00\x80\x7C\x14\x00\x80\x9D\x14\x00\x80\xB2\x14\x00\x80\xB9\x14"
-     "\x00\x80"
-     "\xDE\x14\x00\x80\xEC\x14\x00\x80\xE5\x14\x00\x80\x42\x15\x00\x80\x73\x15"
-     "\x00\x80"
-     "\x7A\x15\x00\x80\x01\x16\x00\x80\x0F\x16\x00\x80\x08\x16\x00\x80\xE7\x15"
-     "\x00\x80"
-     "\x1B\x16\x00\x80\x27\x16\x00\x80\x80\x15\x01\x00\x81\x1D\x01\x00\x82\x15"
-     "\x01\x00"
-     "\x83\x2D\x01\x00\x84\x35\x01\x00\x85\x3D\x01\x00\x86\x35\x01\x00\x87\x2D"
-     "\x01\x00"
-     "\x88\x15\x01\x00\x89\x1D\x01\x00\x8A\x15\x01\x00\x8B\x6D\x01\x00\x8C\xE1"
-     "\x04\x00"
-     "\x8D\x71\x01\x00\x8E\x71\x01\x00\x8F\x71\x01\x00\x90\x11\x01\x00\x91\xA1"
-     "\x04\x00"
-     "\x92\xA1\x04\x00\x93\xA1\x04\x00\x94\xA1\x04\x00\x95\xA1\x04\x00\x96\xA1"
-     "\x04\x00"
-     "\x97\xA1\x04\x00\x98\x61\x01\x00\x99\x61\x01\x00\x9A\x65\x01\x00\x9B\x7D"
-     "\x01\x00"
-     "\x9C\x65\x01\x00\x9D\x6D\x01\x00\x9E\x65\x01\x00\x9F\x5D\x01\x00\xA0\xA5"
-     "\x01\x00"
-     "\xA1\xAD\x01\x00\xA2\xA5\x01\x00\xA3\xBD\x01\x00\xA4\xA5\x01\x00\xA5\xAD"
-     "\x01\x00"
-     "\xA6\xA5\x01\x00\xA7\xDD\x01\x00\xA8\x71\x04\x00\xA9\xE1\x01\x00\xAA\xE1"
-     "\x01\x00"
-     "\xAB\xE1\x01\x00\xAC\xE1\x01\x00\xAD\x8D\x01\x00\xAE\x85\x01\x00\xAF\xFD"
-     "\x01\x00"
-     "\xB0\x11\x04\x00\xB1\x09\x04\x00\xB2\x1D\x04\x00\xB3\xA9\x06\x00\xB4\xE9"
-     "\x06\x00"
-     "\xB5\xE9\x06\x00\xB6\x99\x06\x00\xB7\xE5\x01\x00\xB8\xDD\x01\x00\xB9\xA5"
-     "\x01\x00"
-     "\xBA\x8D\x06\x00\xBB\x85\x06\x00\xBC\xA5\x01\x00\xBD\xA1\x01\x00\x72\x16"
-     "\x00\x80"
-     "\x79\x16\x00\x80\x89\x16\x00\x80\x51\x16\x00\x80\x5F\x16\x00\x80\x95\x16"
-     "\x00\x80"
-     "\x9C\x16\x00\x80\x58\x16\x00\x80\x41\x17\x00\x80\x48\x17\x00\x80\xE9\x16"
-     "\x00\x80"
-     "\xF0\x16\x00\x80\xCF\x16\x00\x80\xD6\x16\x00\x80\x19\x17\x00\x80\x12\x17"
-     "\x00\x80"
-     "\xA6\x17\x00\x80\x64\x11\x00\x80\x95\x29\xFE\x00\x94\x21\xFE\x00\x97\x39"
-     "\xFE\x00"
-     "\x96\x21\xFE\x00\x91\x0D\xFE\x00\x90\x0D\xFE\x00\x93\x39\xFE\x00\x92\x01"
-     "\xFE\x00"
-     "\x9D\xE5\xFE\x00\x9C\xE5\xFE\x00\x9F\xE5\xFE\x00\x9E\xE5\xFE\x00\x99\x05"
-     "\xFE\x00"
-     "\x98\x01\xFE\x00\x9B\x05\xFE\x00\x9A\x05\xFE\x00\xA5\x19\xFE\x00\xA4\x1D"
-     "\xFE\x00"
-     "\xA7\x29\xFE\x00\xA6\x11\xFE\x00\xA1\x01\xFE\x00\xA0\x19\xFE\x00\xA3\x0D"
-     "\xFE\x00"
-     "\xA2\x09\xFE\x00\xAD\x0D\xFE\x00\xAC\x09\xFE\x00\xAF\x1D\xFE\x00\xAE\x1D"
-     "\xFE\x00"
-     "\xA9\x15\xFE\x00\xA8\x15\xFE\x00\xAB\x11\xFE\x00\xAA\x15\xFE\x00\xB5\x05"
-     "\xFE\x00"
-     "\xB4\x05\xFE\x00\xB7\x01\xFE\x00\xB6\x05\xFE\x00\xB1\x69\xFE\x00\xB0\x61"
-     "\xFE\x00"
-     "\xB3\x65\xFE\x00\xB2\x65\xFE\x00\xBD\x3D\xFE\x00\xBC\x39\xFE\x00\xBF\x0D"
-     "\xFE\x00"
-     "\xBE\x0D\xFE\x00\xB9\x21\xFE\x00\xB8\x39\xFE\x00\xBB\x21\xFE\x00\xBA\x29"
-     "\xFE\x00"
-     "\x80\xE9\xFF\x00\x81\xE9\xFF\x00\x82\xF9\xFF\x00\x83\xF9\xFF\x00\x84\xE5"
-     "\xFF\x00"
-     "\x85\xED\xFF\x00\x86\xE1\xFF\x00\x87\xE1\xFF\x00\x88\xDD\xFF\x00\x89\xDD"
-     "\xFE\x00"
-     "\x8A\xD5\xFE\x00\x8B\x2D\x01\x00\x8C\x35\x01\x00\x8D\x3D\x01\x00\x8E\x35"
-     "\x01\x00"
-     "\x8F\x2D\x01\x00\xA7\x11\x00\x80\xA0\x11\x00\x80\x92\x5D\x01\x00\x93\x55"
-     "\x01\x00"
-     "\x94\x4D\x01\x00\x95\x75\x01\x00\x96\x7D\x01\x00\x97\x75\x01\x00\x98\x49"
-     "\x01\x00"
-     "\x99\x49\x01\x00\x9A\x59\x01\x00\x9B\x59\x01\x00\x9C\x49\x01\x00\x9D\x4D"
-     "\x01\x00"
-     "\x9E\x45\x01\x00\x9F\x3D\x00\x00\xA0\xC5\x00\x00\xA1\x85\x03\x00\xA2\x8D"
-     "\x03\x00"
-     "\xA3\x85\x03\x00\xA4\x9D\x03\x00\xA5\x85\x03\x00\xA6\x81\x03\x00\xA7\x81"
-     "\x03\x00"
-     "\xA8\x81\x03\x00\xA9\x81\x03\x00\xAA\x81\x03\x00\xAB\x81\x03\x00\xAC\x81"
-     "\x03\x00"
-     "\xAD\x81\x03\x00\xAE\x81\x03\x00\xAF\x81\x03\x00\xB0\x81\x03\x00\xB1\x81"
-     "\x03\x00"
-     "\xB2\x81\x03\x00\xB3\x81\x03\x00\xB4\x31\xFF\x00\xB5\x3D\xFF\x00\xB6\x89"
-     "\x03\x00"
-     "\xB7\x89\x03\x00\xB8\xB9\x03\x00\xB9\xB9\x03\x00\xBA\x89\x03\x00\xBB\x89"
-     "\x03\x00"
-     "\xBC\x91\x03\x00\xBD\xAD\x03\x00\xBE\x91\x03\x00\xBF\x91\x03\x00\xD4\x11"
-     "\x00\x80"
-     "\xCD\x11\x00\x80\x06\x12\x00\x80\xFF\x11\x00\x80\x85\x12\x00\x80\xA7\xBD"
-     "\x0E\x00"
-     "\x54\x12\x00\x80\x7E\x12\x00\x80\x2D\x13\x00\x80\xC6\x13\x00\x80\x84\x13"
-     "\x00\x80"
-     "\x3A\x14\x00\x80\xF5\x15\x00\x80\x2E\x16\x00\x80\x8B\x17\x00\x80\x7F\x17"
-     "\x00\x80"
-     "\xAD\x17\x00\x80\xAF\x16\x00\x80\x81\x15\x00\x80\xF7\x16\x00\x80\xC0\x14"
-     "\x00\x80"
-     "\xBA\x15\x00\x80\xC8\x16\x00\x80\xA3\x16\x00\x80\xEE\x15\x00\x80\xB3\x15"
-     "\x00\x80"
-     "\x5B\x12\x00\x80\x9E\x12\x00\x80\x66\x16\x00\x80\x4E\x15\x00\x80\x58\x11"
-     "\x00\x80"
-     "\x34\x13\x00\x80\xDD\x16\x00\x80\x09\x14\x00\x80\x67\x15\x00\x80\x38\x11"
-     "\x00\x80"
-     "\x2F\x11\x00\x80\x80\x16\x00\x80\xD7\x13\x00\x80\xEC\x12\x00\x80\xA4\x14"
-     "\x00\x80"
-     "\x61\x17\x00\x80\xE0\x13\x00\x80\xE7\x13\x00\x80\x62\x12\x00\x80\xFA\x12"
-     "\x00\x80"
-     "\xA6\x03\x00\x80\x0B\x24\x00\x80\x0F\x24\x00\x80\xA0\x06\x00\x80\xC7\x00"
-     "\x00\x80"
-     "\xB5\x06\x00\x80\xA9\x24\x00\x80\xAD\x24\x00\x80\xB3\x22\x00\x80\xAF\x22"
-     "\x00\x80"
-     "\xE6\x07\x00\x80\x7E\x09\x00\x80\x82\x09\x00\x80\x3F\x24\x00\x80\xA5\x24"
-     "\x00\x80"
-     "\x33\x24\x00\x80\x37\x24\x00\x80\x1B\x24\x00\x80\x80\x71\x01\x00\x81\x71"
-     "\x01\x00"
-     "\x82\x71\x01\x00\x83\x69\x01\x00\x84\xD5\xFD\x00\x85\x89\xFD\x00\x86\x65"
-     "\x01\x00"
-     "\x87\x9D\x01\x00\x1F\x24\x00\x80\x23\x24\x00\x80\x27\x24\x00\x80\x2B\x24"
-     "\x00\x80"
-     "\x2F\x24\x00\x80\xDE\x07\x00\x80\xE2\x07\x00\x80\xD1\x00\x00\x80\xD7\x00"
-     "\x00\x80"
-     "\xA2\x03\x00\x80\xA8\x03\x00\x80\xE0\x07\x00\x80\xD3\x00\x00\x80\xD5\x00"
-     "\x00\x80"
-     "\xBD\x06\x00\x80\x79\x00\x00\x80\xFD\x14\x00\x80\x7D\x00\x00\x80\x87\x00"
-     "\x00\x80"
-     "\x91\x00\x00\x80\x02\x15\x00\x80\x9B\x00\x00\x80\x0E\x15\x00\x80\xA5\x00"
-     "\x00\x80"
-     "\x1A\x15\x00\x80\xAF\x00\x00\x80\x26\x15\x00\x80\xB9\x00\x00\x80\x2B\x15"
-     "\x00\x80"
-     "\xC5\x10\x00\x80\x4B\x11\x00\x80\x8E\x11\x00\x80\x91\x11\x00\x80\x21\x12"
-     "\x00\x80"
-     "\x4C\x12\x00\x80\x96\x12\x00\x80\xC2\x12\x00\x80\xDE\x12\x00\x80\xE1\x12"
-     "\x00\x80"
-     "\xE9\x12\x00\x80\xB0\x45\x01\x00\xB1\x4D\x01\x00\xB2\x45\x01\x00\xB3\x5D"
-     "\x01\x00"
-     "\xB4\x45\x01\x00\xB5\x4D\x01\x00\xB6\x45\x01\x00\xB7\x7D\x01\x00\xB8\x45"
-     "\x01\x00"
-     "\xB9\x4D\x01\x00\xBA\x45\x01\x00\xBB\x5D\x01\x00\xBC\x45\x01\x00\x03\x13"
-     "\x00\x80"
-     "\x06\x13\x00\x80\x69\x13\x00\x80\xA8\x06\x00\x80\xB3\x06\x00\x80\xF0\xE0"
-     "\x0A\x00"
-     "\x2C\x11\x00\x80\xB7\x13\x00\x80\x26\x14\x00\x80\x61\x14\x00\x80\x8D\x14"
-     "\x00\x80"
-     "\x95\x14\x00\x80\xD1\x14\x00\x80\x30\x15\x00\x80\x5F\x15\x00\x80\x92\x15"
-     "\x00\x80"
-     "\xCB\x15\x00\x80\x44\x16\x00\x80\xC0\x16\x00\x80\x98\x95\x06\x00\x99\x81"
-     "\xB1\x00"
-     "\x34\x17\x00\x80\x5E\x17\x00\x80\xCE\x15\x00\x80\xD3\x15\x00\x80\x77\x17"
-     "\x00\x80"
-     "\xD8\x15\x00\x80\x90\xD9\x06\x00\x91\xD9\x06\x00\x92\xA9\x06\x00\x93\xA9"
-     "\x06\x00"
-     "\x94\xA9\x92\x02\x95\xBD\x06\x00\x96\xB5\x06\x00\x97\xAD\x06\x00\xDD\x15"
-     "\x00\x80"
-     "\xCF\x00\x00\x80\xD9\x00\x00\x80\xB3\x03\x00\x80\xE4\x07\x00\x80\x80\x09"
-     "\x00\x80"
-     "\xF5\x23\x00\x80\xF9\x23\x00\x80\x7F\x2A\x00\x80\x33\x25\x00\x80\x6C\x25"
-     "\x00\x80"
-     "\x8B\x25\x00\x80\x9D\x25\x00\x80\xC2\x25\x00\x80\xC1\x29\x00\x80\xDD\x29"
-     "\x00\x80"
-     "\xB8\x29\x06\x00\xB9\x29\x06\x00\xBA\x39\x06\x00\xBB\x39\x06\x00\xBC\x29"
-     "\x06\x00"
-     "\xBD\x29\x06\x00\xBE\xF9\x07\x00\xBF\x79\xB8\x00\xB0\x21\xB1\x00\xB1\x3D"
-     "\x06\x00"
-     "\xB2\x35\x06\x00\xB3\x0D\x06\x00\xB4\x11\x06\x00\xB5\x45\xB2\x00\xB6\x4D"
-     "\xB2\x00"
-     "\xB7\x19\x06\x00\x80\xB9\xB9\x00\x81\x21\x06\x00\x82\x21\x06\x00\x83\x21"
-     "\x06\x00"
-     "\x84\x21\x06\x00\xEF\x29\x00\x80\x86\x27\x00\x80\x87\x29\x06\x00\x88\x19"
-     "\x06\x00"
-     "\x89\xF9\xB0\x00\x8A\xF5\xB0\x00\x8B\xF1\xB0\x00\x8C\xED\xB0\x00\x8D\xD5"
-     "\x07\x00"
-     "\x8E\xD9\x07\x00\x8F\xDD\x07\x00\x90\xF9\x07\x00\x91\xF9\x07\x00\x92\xF5"
-     "\xB0\x00"
-     "\x83\x2A\x00\x80\x94\x89\x93\x02\x95\x91\x07\x00\x96\x95\x07\x00\x97\x95"
-     "\x07\x00"
-     "\x98\xB9\x07\x00\x99\xD5\xB3\x00\x9A\xDD\xB3\x00\x9B\x89\x07\x00\x9C\x99"
-     "\x07\x00"
-     "\x9D\x99\x07\x00\x9E\x89\x07\x00\x9F\x41\x06\x00\xA0\xC1\x06\x00\xA1\xC1"
-     "\x06\x00"
-     "\xA2\x21\xB3\x00\xA3\xC5\x06\x00\xA4\x91\x01\x00\xA5\x91\x01\x00\xA6\x41"
-     "\xB2\x00"
-     "\x67\x2A\x00\x80\xA8\xED\x06\x00\xA9\xF5\x06\x00\xAA\xFD\x06\x00\xAB\xF5"
-     "\x06\x00"
-     "\x27\x25\x00\x80\x2B\x25\x00\x80\x84\x09\x00\x80\x87\x09\x00\x80\xB0\x9D"
-     "\x06\x00"
-     "\xB1\xA5\x06\x00\xB2\xAD\x06\x00\x90\x09\x00\x80\xB4\xB9\x06\x00\x8D\x09"
-     "\x00\x80"
-     "\xB6\xB5\x06\x00\xB7\xAD\x06\x00\xB8\x95\x06\x00\xB9\x9D\x06\x00\xBA\x95"
-     "\x06\x00"
-     "\xBB\x6D\x06\x00\xBC\x75\x06\x00\xBD\x7D\x06\x00\xBE\x75\x06\x00\xBF\x6D"
-     "\x06\x00"
-     "\x82\x91\xA3\x02\x83\x91\xA3\x02\x80\x9D\x05\x00\x81\x9D\xA3\x02\x86\xB9"
-     "\xA3\x02"
-     "\x87\x8D\xA3\x02\x84\x8D\xA3\x02\x85\xB5\xA3\x02\x8A\x95\xA3\x02\x8B\x91"
-     "\xA3\x02"
-     "\x88\xB1\xA3\x02\x89\x99\xA3\x02\x8E\x3D\x02\x00\x8F\xAD\x05\x00\x8C\x8D"
-     "\xA3\x02"
-     "\x8D\x31\x02\x00\x92\xD9\x05\x00\x93\xD9\x05\x00\x90\xD1\x05\x00\x91\xD5"
-     "\x05\x00"
-     "\x96\xF1\x05\x00\x97\xF5\x05\x00\x94\xC5\x05\x00\x95\xFD\x05\x00\x9A\xC5"
-     "\x05\x00"
-     "\x9B\xC1\x05\x00\x98\xC9\x05\x00\x99\xC9\x05\x00\x9E\xD9\x05\x00\x9F\xDD"
-     "\x05\x00"
-     "\x9C\xDD\x05\x00\x9D\xD5\x05\x00\xA2\x2D\x05\x00\xA3\x29\x05\x00\xA0\x21"
-     "\x05\x00"
-     "\xA1\x21\x05\x00\xA6\x21\x05\x00\xA7\x25\x05\x00\xA4\x35\x05\x00\xA5\x2D"
-     "\x05\x00"
-     "\xAA\x15\x05\x00\xAB\x1D\x05\x00\xA8\x19\x05\x00\xA9\x19\x05\x00\xAE\x0D"
-     "\x05\x00"
-     "\xAF\x41\x05\x00\xAC\x01\x05\x00\xAD\x01\x05\x00\xB2\x39\x05\x00\xB3\x3D"
-     "\x05\x00"
-     "\xB0\x3D\x05\x00\xB1\x35\x05\x00\xB6\x2D\x05\x00\xB7\x29\x05\x00\xB4\x21"
-     "\x05\x00"
-     "\xB5\x21\x05\x00\xBA\xF1\x02\x00\xBB\xF5\x02\x00\xB8\x15\x05\x00\xB9\xFD"
-     "\x02\x00"
-     "\xBE\xE5\x02\x00\xBF\xE1\x02\x00\xBC\xE9\x02\x00\xBD\xE9\x02\x00\x92\x1D"
-     "\xD5\x00"
-     "\x93\x25\xD5\x00\x8A\x09\x00\x80\xE6\x22\x00\x80\x96\x21\xD5\x00\x97\x21"
-     "\xD5\x00"
-     "\x94\x3D\xD5\x00\x95\x35\xD5\x00\x9A\x19\xD5\x00\x9B\x19\xD5\x00\xD9\xE4"
-     "\xA0\x02"
-     "\xE9\x23\x00\x80\xCC\xC0\x85\x02\xCC\xFC\x85\x02\x9C\x71\xD5\x00\xD9\xDC"
-     "\xA0\x02"
-     "\x82\xA1\xD7\x00\x83\xAD\xD7\x00\xEE\x22\x00\x80\xF6\x22\x00\x80\x86\xA9"
-     "\xD7\x00"
-     "\x87\xA9\xD7\x00\x84\xA9\xD7\x00\x85\xA9\xD7\x00\x8A\x61\xD5\x00\x8B\x61"
-     "\xD5\x00"
-     "\xD9\x84\xA0\x02\xED\x23\x00\x80\x8E\x61\xD5\x00\x8F\x61\xD5\x00\x8C\x71"
-     "\xD5\x00"
-     "\x8D\x71\xD5\x00\xF1\x23\x00\x80\xB6\x0D\x00\x80\xB9\x0D\x00\x80\x7F\x0D"
-     "\x00\x80"
-     "\xCE\x0C\x00\x80\x94\x0D\x00\x80\x5B\x3B\x00\x80\x9A\x0D\x00\x80\x97\x0D"
-     "\x00\x80"
-     "\xD7\x0C\x00\x80\xDA\x0C\x00\x80\xB3\x1B\x00\x80\xDD\x0C\x00\x80\xE9\x0C"
-     "\x00\x80"
-     "\xA6\x0D\x00\x80\xBD\xB5\x87\x02\xA2\x0D\x02\x00\xA3\x11\x93\x02\xA0\x71"
-     "\x02\x00"
-     "\xA1\x71\x02\x00\xA6\xDD\x00\x00\xFB\x0C\x00\x80\xA4\x0D\x02\x00\xA5\xD5"
-     "\x00\x00"
-     "\xAA\xA9\x00\x00\xAB\xE5\x0B\x00\xA8\xC5\x00\x00\xA9\x01\x08\x00\xAE\x71"
-     "\x00\x00"
-     "\xF5\x0C\x00\x80\xAC\xB5\x00\x00\xAD\x91\x00\x00\x8B\x21\x0C\x00\x8A\x25"
-     "\x0C\x00"
-     "\x89\x29\x0C\x00\x88\x11\x0C\x00\x8F\x35\x0C\x00\x8E\x31\x0C\x00\x8D\x3D"
-     "\x0C\x00"
-     "\x8C\x3D\x0C\x00\x83\x39\x0C\x00\x82\x3D\x0C\x00\x81\x31\x0C\x00\x80\x39"
-     "\x0C\x00"
-     "\x87\x2D\x0C\x00\x86\x29\x0C\x00\x85\x25\x0C\x00\x84\x25\x0C\x00\x9B\x71"
-     "\x0C\x00"
-     "\x9A\x75\x0C\x00\x99\x79\x0C\x00\x98\x61\x0C\x00\x9F\x65\x0C\x00\x9E\x61"
-     "\x0C\x00"
-     "\x9D\x6D\x0C\x00\x9C\x6D\x0C\x00\x93\x49\x0C\x00\x92\x4D\x0C\x00\x91\x41"
-     "\x0C\x00"
-     "\x90\x49\x0C\x00\x97\x5D\x0C\x00\x96\x59\x0C\x00\x95\x55\x0C\x00\x94\x55"
-     "\x0C\x00"
-     "\xAB\xC9\x0C\x00\xAA\xCD\x0C\x00\xA9\xC1\x0C\x00\xA8\xB1\x0C\x00\xAF\xB5"
-     "\xFE\x00"
-     "\xAE\xD9\x0C\x00\xAD\xD5\x0C\x00\xAC\xD5\x0C\x00\xA3\x99\x0C\x00\xA2\x9D"
-     "\x0C\x00"
-     "\xA1\x91\x0C\x00\xA0\x99\x0C\x00\xA7\x8D\x0C\x00\xA6\x89\x0C\x00\xA5\x85"
-     "\x0C\x00"
-     "\xA4\x85\x0C\x00\xBB\x41\x0C\x00\xBA\x99\x0C\x00\xB9\x95\x0C\x00\xB8\x99"
-     "\x0C\x00"
-     "\xBF\xD9\x88\x02\xEF\x0C\x00\x80\x82\x0D\x00\x80\xBC\x5D\x0C\x00\xB3\xB9"
-     "\x0C\x00"
-     "\xB2\xBD\x0C\x00\xB1\xB1\x0C\x00\xB0\xC9\xFE\x00\xB7\xA5\x0C\x00\xB6\xA5"
-     "\x0C\x00"
-     "\xB5\xA9\x0C\x00\xB4\xA5\x0C\x00\x81\xC9\x0B\x00\x85\x0D\x00\x80\x83\x35"
-     "\x0A\x00"
-     "\x82\x89\x0D\x00\x85\x45\x0A\x00\x84\x45\x0A\x00\x87\x91\x0D\x00\x86\x5D"
-     "\x0A\x00"
-     "\x89\x95\xBF\x00\x88\x95\xBF\x00\x8B\x81\xBF\x00\x8A\x69\x0A\x00\x8D\x79"
-     "\x0A\x00"
-     "\x8C\x8D\x0B\x00\x8F\xB5\x0D\x00\x8E\x81\xBC\x00\x91\xB5\x9D\x02\x90\xB1"
-     "\xBC\x00"
-     "\x93\xA9\x9D\x02\x92\x91\x9D\x02\x95\xAD\xBC\x00\x94\xAD\xBC\x00\x97\xD5"
-     "\xBF\x00"
-     "\x96\xD5\xBF\x00\x99\xC5\xBF\x00\x98\xC5\xBF\x00\x9B\xB1\x9F\x02\x9A\xA9"
-     "\x0B\x00"
-     "\x9D\xA9\xBE\x00\x9C\x75\x0D\x00\x9F\xB5\x0B\x00\x9E\x79\x0D\x00\xA1\x01"
-     "\xBF\x00"
-     "\xA0\xDD\x0A\x00\xA3\x95\xB1\x00\xA2\x85\xBD\x00\xA5\x81\xBD\x00\xA4\x99"
-     "\xBD\x00"
-     "\xA7\x81\xBD\x00\xA6\xA9\xB1\x00\xA9\x89\xBD\x00\xA8\x85\xBD\x00\xAB\x99"
-     "\xBD\x00"
-     "\xAA\x89\xBD\x00\xAD\xF1\xBD\x00\xAC\x8D\xBD\x00\xAF\xE9\xBD\x00\xAE\xF5"
-     "\xBD\x00"
-     "\xB1\x21\xBF\x00\xB0\x9D\xBD\x00\xB3\x49\xC9\x00\xB2\xAD\xBD\x00\xB5\xA1"
-     "\xBD\x00"
-     "\xB4\xBD\xBD\x00\xB7\xA1\xBD\x00\xB6\x49\xB1\x00\xB9\x69\xC9\x00\xB8\x4D"
-     "\xB1\x00"
-     "\xBB\xDD\x0A\x00\xBA\xCD\xB1\x00\xBD\xC9\x0A\x00\xBC\xB9\x0A\x00\xBF\x19"
-     "\x0B\x00"
-     "\xBE\x6D\x0D\x00\x80\xFD\xBE\x00\x81\xD5\x9E\x02\x82\x65\xBE\x00\x83\xC5"
-     "\xBE\x00"
-     "\x84\x91\xBE\x00\x85\x9D\xBE\x00\x86\xA5\xBE\x00\x87\x8D\xBE\x00\x88\xAD"
-     "\x91\x02"
-     "\x89\xE5\xBE\x00\x8A\x29\x92\x02\x8B\xB5\xBE\x00\x8C\x11\x92\x02\x8D\x95"
-     "\xBE\x00"
-     "\x8E\x2D\xB2\x00\x8F\xC5\x9E\x02\x90\xE9\xBE\x00\x91\xB1\xB5\x00\x92\x41"
-     "\x92\x02"
-     "\x93\xF1\x9F\x02\x94\xD5\xBE\x00\x95\xE5\xBE\x00\x96\xE1\xBE\x00\x97\x4D"
-     "\x92\x02"
-     "\x98\x65\x92\x02\x99\x7D\x92\x02\x9A\xB9\xBE\x00\x9B\x89\x08\x00\x9C\xE5"
-     "\x0E\x00"
-     "\x9D\x9D\x08\x00\x04\x0D\x00\x80\x9F\xE5\x0E\x00\xA0\x1D\x0E\x00\xA1\x45"
-     "\x08\x00"
-     "\xA2\xAD\x09\x00\xA3\xA5\x09\x00\xA4\x2D\x08\x00\xA5\x35\x0E\x00\xA6\x05"
-     "\xD6\x00"
-     "\xA7\x39\x0E\x00\xA8\x09\x0E\x00\xA9\x09\x0E\x00\xAA\x19\x0E\x00\xAB\x19"
-     "\x0E\x00"
-     "\xAC\x09\x0E\x00\xAD\x09\x0E\x00\xAE\xAD\x0F\x00\xAF\xA5\x0F\x00\xB0\x8D"
-     "\x0F\x00"
-     "\xB1\x95\xD3\x00\xB2\x9D\xD3\x00\xB3\x85\xD4\x00\xB4\x89\xD3\x00\xB5\x89"
-     "\xD3\x00"
-     "\xB6\xD9\xD6\x00\xB7\xD9\xD6\x00\xB8\xE9\xD6\x00\xB9\xE9\xD6\x00\xBA\xF9"
-     "\xD6\x00"
-     "\xBB\xF9\xD6\x00\xBC\xE9\xD6\x00\xBD\xE9\xD6\x00\xBE\xB9\xD6\x00\xBF\xB9"
-     "\xD6\x00"
-     "\x80\x49\xD7\x00\x81\x49\xD7\x00\x82\x59\xD7\x00\x83\x59\xD7\x00\x84\x49"
-     "\xD7\x00"
-     "\x85\x75\xD2\x00\x86\x7D\xD2\x00\x87\x75\xD2\x00\x88\x4D\xD2\x00\x89\x55"
-     "\xD2\x00"
-     "\x8A\x75\xD7\x00\x8B\xAD\xD7\x00\x8C\xB5\xD7\x00\x8D\xBD\xD7\x00\x8E\xB5"
-     "\xD7\x00"
-     "\x8F\x49\xD2\x00\x90\x39\xD2\x00\x91\x39\xD2\x00\x92\xC9\xD2\x00\x93\xC9"
-     "\xD2\x00"
-     "\x94\xD9\xD2\x00\x95\xD9\xD2\x00\x96\xC9\xD2\x00\x97\xC9\xD2\x00\x98\xF9"
-     "\xD2\x00"
-     "\x99\xF9\xD2\x00\x9A\xC9\xD2\x00\x9B\xC5\xD2\x00\x9C\xE1\xD7\x00\x9D\xE9"
-     "\x0E\x00"
-     "\x9E\xC9\x0E\x00\x9F\xC9\x0E\x00\xA0\xD1\xD9\x00\xA1\xC1\xDB\x00\xA2\xC1"
-     "\xDB\x00"
-     "\xA3\xDD\xDB\x00\xA4\xC5\xDB\x00\xA5\xCD\xDB\x00\xA6\xC5\xDB\x00\xA7\xFD"
-     "\xDB\x00"
-     "\xA8\x25\xDB\x00\xA9\x65\xDB\x00\xAA\xCD\xDB\x00\xAB\xC5\xDB\x00\xAC\xDD"
-     "\xDB\x00"
-     "\xAD\xC1\xDB\x00\xAE\xC1\xDB\x00\xAF\x79\xDB\x00\xB0\x09\xDB\x00\xB1\x09"
-     "\xDB\x00"
-     "\xB2\x19\xDB\x00\xB3\x19\xDB\x00\xB4\x09\xDB\x00\xB5\xB5\xD8\x00\xB6\xC1"
-     "\xDE\x00"
-     "\xB7\xDD\xDF\x00\xB8\xE5\xDF\x00\xB9\xF5\xDF\x00\xBA\xFD\xDF\x00\xBB\xAD"
-     "\xDF\x00"
-     "\xBC\xB5\xDF\x00\xBD\xA5\xDF\x00\xBE\xAD\xDF\x00\x55\x0D\x00\x80\xF2\x22"
-     "\x00\x80"
-     "\xA9\x0D\x00\x80\x90\x28\xDC\x00\x91\x04\x00\x00\x01\x0A\x00\x80\xAC\x0D"
-     "\x00\x80"
-     "\x13\x0D\x00\x80\xDB\xEA\x00\x80\x4F\x3B\x00\x80\x19\x0D\x00\x80\xE0\xEA"
-     "\x00\x80"
-     "\x6A\x0D\x00\x80\x9E\x60\x12\x00\x9F\xEC\x11\x00\xE5\xEA\x00\x80\x9D\x88"
-     "\x00\x00"
-     "\x34\x0A\x00\x80\x31\x0D\x00\x80\x34\x0D\x00\x80\x3D\x0D\x00\x80\xEA\xEA"
-     "\x00\x80"
-     "\x43\x0D\x00\x80\x46\x0D\x00\x80\x49\x0D\x00\x80\xA3\x0D\x00\x80\x9D\x0D"
-     "\x00\x80"
-     "\x5F\x3B\x00\x80\xA0\x0D\x00\x80\x58\x0D\x00\x80\x03\x24\x00\x80\x6D\x0D"
-     "\x00\x80"
-     "\x8E\x0D\x00\x80\x91\x0D\x00\x80\x7C\x0D\x00\x80\xB1\x0B\x00\x80\xB4\x0B"
-     "\x00\x80"
-     "\xB7\x0B\x00\x80\xEF\xEA\x00\x80\xF4\xEA\x00\x80\x01\xEA\x00\x80\x06\xEA"
-     "\x00\x80"
-     "\x0B\xEA\x00\x80\x10\xEA\x00\x80\x91\x08\x02\x00\xB1\x14\xFD\x00\x84\x90"
-     "\x02\x00"
-     "\xA7\x79\x00\x00\xB2\x74\x0B\x00\x15\xEA\x00\x80\x87\xCC\x03\x00\xF0\x40"
-     "\x02\x00"
-     "\x91\x2C\x02\x00\x8C\xF0\x02\x00\x8D\x6C\x02\x00\xBE\x39\x00\x00\xCC\x40"
-     "\xBC\x00"
-     "\x1A\xEA\x00\x80\x97\x0D\x00\x00\x1F\xEA\x00\x80\xB9\xDC\x03\x00\xF0\x4C"
-     "\x02\x00"
-     "\xAF\x9C\x29\x00\x91\x4C\x02\x00\x92\x38\x00\x00\x81\x85\x0A\x00\xA5\x38"
-     "\xDC\x00"
-     "\x83\x19\x01\x00\x82\x85\x0A\x00\x85\x81\x0A\x00\x84\x99\x0A\x00\x87\x8D"
-     "\x0A\x00"
-     "\x29\xEA\x00\x80\x89\xBD\x0A\x00\x88\xBD\x0A\x00\x8B\x8D\x0A\x00\x8A\x8D"
-     "\x0A\x00"
-     "\x8D\x9D\x0A\x00\x8C\x9D\x0A\x00\x8F\x8D\x0A\x00\x8E\x8D\x0A\x00\x91\x95"
-     "\x0A\x00"
-     "\x90\xFD\x0A\x00\x93\x9D\x05\x00\x92\xB9\x0A\x00\x95\x89\xBB\x00\x94\x8D"
-     "\x05\x00"
-     "\x97\x21\x07\x00\x96\x59\x04\x00\x99\x65\x01\x00\x98\x99\x06\x00\x9B\x71"
-     "\x01\x00"
-     "\x9A\x79\x01\x00\x9D\x51\x01\x00\x9C\x69\x01\x00\x9F\x51\x01\x00\x9E\x59"
-     "\x01\x00"
-     "\xA1\xB9\x01\x00\xA0\xA9\x01\x00\xA3\xCD\x01\x00\xA2\xCD\x01\x00\xA5\x65"
-     "\xB2\x00"
-     "\xA4\xDD\x01\x00\xA7\xC9\x01\x00\xA6\xD1\x01\x00\xA9\xF9\x01\x00\xA8\xF1"
-     "\x01\x00"
-     "\xAB\xC9\x01\x00\xAA\xF1\x01\x00\xAD\xD9\x01\x00\xAC\xD1\x01\x00\xAF\xC9"
-     "\x01\x00"
-     "\xAE\xD1\x01\x00\x24\xEA\x00\x80\xB0\xB1\x01\x00\xB3\xB5\x01\x00\xB2\xB5"
-     "\x01\x00"
-     "\xB5\x55\x01\x00\xB4\x55\x01\x00\xB7\x55\x01\x00\xB6\x55\x01\x00\xB9\x75"
-     "\x01\x00"
-     "\xB8\x75\x01\x00\x2E\xEA\x00\x80\xBA\x75\x01\x00\xF0\x4C\x02\x00\x99\x2C"
-     "\xDC\x00"
-     "\x91\xF4\x01\x00\x9B\x40\xDC\x00\x82\x90\x01\x00\xB0\x19\x03\x00\xB9\xD4"
-     "\x03\x00"
-     "\xB8\x3C\x00\x00\xAF\x05\x00\x00\x33\xEA\x00\x80\x84\x6C\x00\x00\xF0\x28"
-     "\x02\x00"
-     "\xB5\xC4\x03\x00\xF0\x14\x02\x00\x38\xEA\x00\x80\xBA\xED\x03\x00\xBD\xD9"
-     "\x03\x00"
-     "\x3D\xEA\x00\x80\x42\xEA\x00\x80\x8D\xE4\x00\x00\x92\x48\x00\x00\xAF\x6D"
-     "\x00\x00"
-     "\xA4\xF0\x02\x00\x4B\xEA\x00\x80\x96\xBC\x02\x00\x96\xF8\x02\x00\x54\xEA"
-     "\x00\x80"
-     "\x61\xEA\x00\x80\x6E\xEA\x00\x80\x96\x9C\x02\x00\x7B\xEA\x00\x80\x88\xEA"
-     "\x00\x80"
-     "\x91\x4C\x02\x00\x91\xAC\x02\x00\xB0\x4D\x00\x00\x95\xEA\x00\x80\xB8\xEA"
-     "\x00\x80"
-     "\x9E\xEA\x00\x80\x90\x79\x10\x00\xC1\xEA\x00\x80\xAB\xEA\x00\x80\xCE\xEA"
-     "\x00\x80"
-     "\xA4\x6C\x02\x00\xD6\x0A\x00\x80\x85\xD8\x03\x00\x86\x00\x00\x00\x87\x18"
-     "\x02\x00"
-     "\x86\x4C\x03\x00\x80\x25\x00\x00\x82\x00\x00\x00\x81\x00\x00\x00\x80\x00"
-     "\x00\x00"
-     "\x9D\xC0\x2D\x00\x9C\x00\x2C\x00\x9F\x68\x11\x00\x9E\xCC\x2E\x00\x99\xD0"
-     "\x15\x00"
-     "\x98\x00\x14\x00\x9B\xD8\x17\x00\x9A\xDC\x16\x00\x95\xE0\x19\x00\x94\x00"
-     "\x18\x00"
-     "\x97\xE8\x1B\x00\x96\xEC\x1A\x00\x91\xF0\x1D\x00\x90\x00\x1C\x00\x93\xF8"
-     "\x1F\x00"
-     "\x92\xFC\x1E\x00\x9E\x39\x01\x00\x9F\x49\x01\x00\xED\x00\x00\x00\xBB\x71"
-     "\x01\x00"
-     "\xBC\x91\x01\x00\xBD\x91\x01\x00\xBE\x95\x01\x00\xBF\x8D\x01\x00\xCC\xAC"
-     "\x98\x02"
-     "\xCD\xA4\x9E\x02\xC2\x0C\x00\x80\xE5\x0D\x00\x80\xCC\xDC\xBE\x00\xEB\x0D"
-     "\x00\x80"
-     "\xF1\x0D\x00\x80\xFD\x0D\x00\x80\x00\x0E\x00\x80\x03\x0E\x00\x80\x06\x0E"
-     "\x00\x80"
-     "\x09\x0E\x00\x80\x18\x0E\x00\x80\x1E\x0E\x00\x80\x21\x0E\x00\x80\x24\x0E"
-     "\x00\x80"
-     "\x1A\x23\x00\x80\x4E\x0E\x00\x80\x5D\x0E\x00\x80\x83\x0E\x00\x80\xCC\xF8"
-     "\x9E\x02"
-     "\xCD\xFC\x9E\x02\x86\x0E\x00\x80\x95\x0E\x00\x80\x98\x0E\x00\x80\xA4\x0E"
-     "\x00\x80"
-     "\x12\x23\x00\x80\xB9\x0E\x00\x80\xCB\x00\x00\x00\xBC\x0E\x00\x80\xDA\x0E"
-     "\x00\x80"
-     "\xDD\x0E\x00\x80\xE3\x0E\x00\x80\xF8\x0E\x00\x80\xFB\x0E\x00\x80\x04\x0F"
-     "\x00\x80"
-     "\x07\x0F\x00\x80\x0A\x0F\x00\x80\x0D\x0F\x00\x80\xC0\x00\x00\x00\x13\x0F"
-     "\x00\x80"
-     "\x16\x0F\x00\x80\x19\x0F\x00\x80\x1F\x0F\x00\x80\x2E\x0F\x00\x80\x43\x3B"
-     "\x00\x80"
-     "\x43\x0F\x00\x80\x5E\x0F\x00\x80\x6A\x0F\x00\x80\x58\x0F\x00\x80\xA4\x85"
-     "\x01\x00"
-     "\xEA\x22\x00\x80\xA2\x91\x01\x00\xA3\x9D\x01\x00\xA0\xA1\x01\x00\xA1\xA1"
-     "\x01\x00"
-     "\xBD\x25\xA4\x00\xBC\x2D\xA4\x00\xBF\x31\xA4\x00\xBE\x3D\xA4\x00\xB9\x25"
-     "\xA8\x00"
-     "\xB8\x05\xA4\x00\xBB\x39\xA4\x00\xBA\x05\xA4\x00\xB5\x29\xA4\x00\xB4\xA5"
-     "\xA6\x00"
-     "\xB7\x25\xA4\x00\xB6\x31\xA8\x00\xB1\x35\x88\x02\xB0\x25\x88\x02\xB3\xE9"
-     "\xA4\x00"
-     "\xB2\x25\x88\x02\xAD\xC5\xA4\x00\xAC\xF9\x85\x02\xAF\xF1\xA4\x00\xAE\xFD"
-     "\xA4\x00"
-     "\xA9\xF9\xA4\x00\xA8\xAD\x84\x02\xAB\x51\x88\x02\xAA\xA9\xAF\x00\xA5\x41"
-     "\x88\x02"
-     "\xA4\xFD\xA4\x00\xA7\x7D\xA8\x00\xA6\xCD\xA4\x00\xA1\xFD\x8B\x02\xA0\xE5"
-     "\xA4\x00"
-     "\xA3\x79\x88\x02\xA2\xBD\xA4\x00\x9D\x01\xA4\x00\x9C\x4D\xA4\x00\x9F\x35"
-     "\xA4\x00"
-     "\x9E\x05\xA4\x00\x99\xB1\xA8\x00\x98\x1D\xD0\x00\x9B\xF5\xA4\x00\x9A\x4D"
-     "\x84\x02"
-     "\x95\x35\xA8\x00\x94\xD5\xA4\x00\x97\x01\xA8\x00\x96\xD5\xA4\x00\x91\xD1"
-     "\xA4\x00"
-     "\x90\x55\xA6\x00\x93\xD1\xA4\x00\x92\x3D\xD0\x00\x8D\x89\xA4\x00\x8C\x85"
-     "\xA4\x00"
-     "\x8F\x91\xA4\x00\x8E\x9D\xA4\x00\x89\xF5\xA4\x00\x88\xFD\xA4\x00\x8B\xE1"
-     "\xA4\x00"
-     "\x8A\xED\xA4\x00\x85\xD5\xA8\x00\x84\xF5\xA4\x00\x87\xC9\xA4\x00\x86\xF5"
-     "\xA4\x00"
-     "\x81\xF9\xA4\x00\x80\x75\xA6\x00\x83\xF5\xA4\x00\x82\xE1\xA8\x00\x80\x41"
-     "\xA5\x00"
-     "\x81\x4D\xA5\x00\x82\x59\xA5\x00\x83\x55\xA5\x00\x84\xF1\xA7\x00\x85\x75"
-     "\xA5\x00"
-     "\x86\x99\xD1\x00\x87\x75\xA5\x00\x88\x51\xA5\x00\x89\xB1\xA9\x00\x8A\x51"
-     "\xA5\x00"
-     "\x8B\x85\xA9\x00\x8C\xB9\xD1\x00\x8D\x15\xA9\x00\x8E\xE9\x85\x02\x8F\x51"
-     "\xA5\x00"
-     "\x90\x89\xA5\x00\x91\xC5\xA5\x00\x92\xC1\xA5\x00\x93\xF1\xA5\x00\x94\xC1"
-     "\xA5\x00"
-     "\x3A\x0F\x00\x80\x96\x99\xA5\x00\x97\x5D\x89\x02\x98\xF9\xA5\x00\x99\x45"
-     "\x89\x02"
-     "\x9A\xC9\xA5\x00\x9B\x79\xA9\x00\x9C\x89\x85\x02\x9D\xDD\xA5\x00\x9E\x8D"
-     "\xAE\x00"
-     "\x9F\x75\x89\x02\xA0\x3D\x84\x02\xA1\x01\xA5\x00\xA2\x39\xA5\x00\xA3\x35"
-     "\xA5\x00"
-     "\xA4\x81\x89\x02\xA5\x91\x89\x02\xA6\x81\x89\x02\xA7\x4D\xA5\x00\xA8\x21"
-     "\xA7\x00"
-     "\xA9\xAD\xA5\x00\xAA\xB5\xA9\x00\xAB\xA1\xA5\x00\xAC\xA1\xA5\x00\xAD\x81"
-     "\xA9\x00"
-     "\xAE\xA1\xA5\x00\xAF\x9D\xA5\x00\xB0\xE9\xA5\x00\xB1\xE1\xA5\x00\xB2\xF9"
-     "\xA5\x00"
-     "\xB3\xF5\xA5\x00\xB4\x91\xA5\x00\xB5\x9D\xA5\x00\xB6\x89\xA5\x00\xB7\x85"
-     "\xA5\x00"
-     "\xB8\x01\xA7\x00\xB9\x85\xA5\x00\xBA\x69\xD1\x00\xBB\x85\xA5\x00\xBC\x81"
-     "\xA5\x00"
-     "\xBD\x61\xA9\x00\xBE\x81\xA5\x00\xBF\x55\xA9\x00\x80\x89\xD2\x00\x81\x25"
-     "\xAA\x00"
-     "\x82\xD9\x86\x02\x83\x61\xA6\x00\x84\xD9\xA6\x00\x85\x95\xA6\x00\x86\x91"
-     "\xA6\x00"
-     "\x87\xA1\xA6\x00\x88\xB1\xA6\x00\x89\xA9\x89\x02\x8A\xE9\xA6\x00\x8B\x2D"
-     "\x8A\x02"
-     "\x8C\xA9\xA6\x00\x8D\x15\x8A\x02\x8E\x99\xA6\x00\x8F\x29\xAA\x00\x90\xB9"
-     "\x86\x02"
-     "\x91\xED\xA6\x00\x92\xBD\xAD\x00\x93\x45\x8A\x02\x94\xED\x87\x02\x95\xD1"
-     "\xA6\x00"
-     "\x96\xE9\xA6\x00\x97\xE5\xA6\x00\x98\x71\x8A\x02\x99\x61\x8A\x02\x9A\x71"
-     "\x8A\x02"
-     "\x9B\xBD\xA6\x00\x9C\xF1\xA4\x00\x4F\x0F\x00\x80\x9E\x65\xAA\x00\x9F\x71"
-     "\xA6\x00"
-     "\x5B\x0F\x00\x80\x3F\x3B\x00\x80\xA2\x91\xA6\x00\x61\x0F\x00\x80\xCD\x17"
-     "\x00\x80"
-     "\xA5\xB1\xA6\x00\xA6\xA9\xA6\x00\x06\x18\x00\x80\xD0\x17\x00\x80\xA9\xED"
-     "\xA6\x00"
-     "\xAA\xF9\xA6\x00\xAB\xF5\xA6\x00\xAC\x51\xA4\x00\xBA\x03\x00\x80\xAE\x39"
-     "\xD2\x00"
-     "\xAF\xD5\xA6\x00\xB0\xB1\xA6\x00\xB1\x51\xAA\x00\xB2\xB1\xA6\x00\xB3\x65"
-     "\xAA\x00"
-     "\xB4\x59\xD2\x00\xB5\xF5\xAA\x00\xB6\x09\x86\x02\xB7\xB1\xA6\x00\xB8\x29"
-     "\xA6\x00"
-     "\xB9\x65\xA6\x00\xB7\x03\x00\x80\xBB\x51\xA6\x00\xCA\x03\x00\x80\xBD\x79"
-     "\x89\x02"
-     "\xBE\x39\xA6\x00\xBF\xFD\x8A\x02\x80\x99\xA7\x00\x81\x25\x8B\x02\x82\xA9"
-     "\xA7\x00"
-     "\x83\x19\xAB\x00\xED\x04\x00\x80\x85\xBD\xA7\x00\x86\xED\xAC\x00\x87\x15"
-     "\x8B\x02"
-     "\x88\x9D\x86\x02\x89\xA1\xA7\x00\x8A\x99\xA7\x00\x8B\x95\xA7\x00\x8C\x21"
-     "\x8B\x02"
-     "\x8D\x31\x8B\x02\x8E\x21\x8B\x02\x8F\xED\xA7\x00\x90\xC1\xA5\x00\x91\x4D"
-     "\xA7\x00"
-     "\x92\x55\xAB\x00\x93\x41\xA7\x00\x94\x41\xA7\x00\x95\x61\xAB\x00\x96\x41"
-     "\xA7\x00"
-     "\x97\x7D\xA7\x00\x98\x49\xA7\x00\x99\x41\xA7\x00\x9A\x59\xA7\x00\x9B\x55"
-     "\xA7\x00"
-     "\x9C\x31\xA7\x00\x9D\x3D\xA7\x00\x9E\x29\xA7\x00\x9F\x25\xA7\x00\xA0\x61"
-     "\xA5\x00"
-     "\xA1\xE5\xA7\x00\xA2\x09\xD3\x00\xA3\xE5\xA7\x00\xA4\xE1\xA7\x00\xA5\x01"
-     "\xAB\x00"
-     "\xA6\xE1\xA7\x00\xA7\x35\xAB\x00\xA8\x29\xD3\x00\xA9\x85\xAB\x00\xAA\x79"
-     "\x87\x02"
-     "\xAB\xC1\xA7\x00\xAC\x79\xA7\x00\xAD\x35\xA7\x00\xAE\x31\xA7\x00\xAF\x01"
-     "\xA7\x00"
-     "\xB0\x51\xA7\x00\xB1\x49\x88\x02\xB2\x09\xA7\x00\xB3\xCD\x8B\x02\xB4\x49"
-     "\xA7\x00"
-     "\xB5\xF5\x8B\x02\xB6\x79\xA7\x00\xB7\xC9\xAB\x00\xB8\x19\x87\x02\xB9\x4D"
-     "\xA7\x00"
-     "\xBA\x1D\xAC\x00\xBB\xE5\x8B\x02\xBC\x4D\x86\x02\xBD\x71\xA7\x00\xBE\x49"
-     "\xA7\x00"
-     "\xBF\x45\xA7\x00\xB9\x09\xA0\x00\xB8\x85\xA2\x00\xBB\x05\xA0\x00\x8C\x05"
-     "\x00\x80"
-     "\xBD\x25\xAC\x00\xBC\x05\xA0\x00\xDF\x05\x00\x80\xBE\x05\xA0\x00\xB1\xA5"
-     "\xA0\x00"
-     "\xB0\x99\x81\x02\xB3\x91\xA0\x00\xB2\x9D\xA0\x00\xB5\x35\x8C\x02\xB4\x25"
-     "\x8C\x02"
-     "\xB7\xE9\xA0\x00\xB6\x25\x8C\x02\xA9\x61\x8C\x02\xA8\xDD\xA0\x00\xAB\x5D"
-     "\xAC\x00"
-     "\xAA\xED\xA0\x00\xAD\xF9\xA0\x00\xAC\xAD\x80\x02\xAF\x51\x8C\x02\xAE\xA9"
-     "\xAB\x00"
-     "\xA1\xE1\xA0\x00\xA0\xAD\xA0\x00\xA3\xD5\xA0\x00\xA2\xE5\xA0\x00\xA5\xFD"
-     "\x8F\x02"
-     "\xA4\xE5\xA0\x00\xA7\x79\x8C\x02\xA6\xBD\xA0\x00\x99\x15\xAC\x00\x98\xF5"
-     "\xA0\x00"
-     "\x9B\x21\xAC\x00\x9A\xF5\xA0\x00\x1A\x06\x00\x80\x9C\x1D\xD4\x00\x9F\xF5"
-     "\xA0\x00"
-     "\x9E\x4D\x80\x02\x91\xE9\xA0\x00\x90\xE5\xA0\x00\x93\xF1\xA0\x00\x92\xFD"
-     "\xA0\x00"
-     "\x40\x06\x00\x80\x94\x55\xA2\x00\x97\xD1\xA0\x00\x96\x3D\xD4\x00\x89\xF5"
-     "\xAC\x00"
-     "\x88\xD5\xA0\x00\x57\x06\x00\x80\x8A\xD5\xA0\x00\x8D\xF5\xA0\x00\x74\x06"
-     "\x00\x80"
-     "\x8F\xE1\xA0\x00\x8E\xED\xA0\x00\x81\xE5\x8C\x02\x80\xF5\x8C\x02\x83\x39"
-     "\xA0\x00"
-     "\x82\xF5\x8C\x02\x85\xF9\xA0\x00\x84\x75\xA2\x00\x87\xF5\xA0\x00\x8B\x06"
-     "\x00\x80"
-     "\x80\x29\xA1\x00\x81\x21\xA1\x00\x82\x39\xA1\x00\x83\x35\xA1\x00\x84\x51"
-     "\xA1\x00"
-     "\x38\x01\x00\x80\x86\x49\xA1\x00\x3C\x01\x00\x80\x40\x01\x00\x80\x44\x01"
-     "\x00\x80"
-     "\x8A\xA9\xD5\x00\x8B\x45\xA1\x00\x8C\x41\xA1\x00\x8D\xA1\xAD\x00\x8E\x41"
-     "\xA1\x00"
-     "\x8F\x95\xAD\x00\x90\xC9\xD5\x00\x48\x01\x00\x80\x92\x99\x81\x02\x93\x21"
-     "\xA1\x00"
-     "\x94\x99\xA1\x00\x95\xD5\xA1\x00\x96\xD1\xA1\x00\x97\xE1\xA1\x00\x98\xF1"
-     "\xA1\x00"
-     "\x99\xE9\x8E\x02\x9A\xA9\xA1\x00\x9B\x6D\x8D\x02\x9C\xE9\xA1\x00\x9D\x55"
-     "\x8D\x02"
-     "\x9E\xD9\xA1\x00\x9F\x69\xAD\x00\xA0\x79\x81\x02\xA1\x2D\xA1\x00\xA2\x7D"
-     "\xAA\x00"
-     "\xA3\x85\x8D\x02\xA4\x2D\x80\x02\xA5\x11\xA1\x00\xA6\x29\xA1\x00\xA7\x25"
-     "\xA1\x00"
-     "\xA8\xB1\x8D\x02\xA9\xA1\x8D\x02\xAA\xB1\x8D\x02\xAB\x7D\xA1\x00\xAC\x31"
-     "\xA3\x00"
-     "\xAD\xBD\xA1\x00\xAE\xA5\xAD\x00\xAF\xB1\xA1\x00\xB0\xD1\xA1\x00\xB1\xF1"
-     "\xAD\x00"
-     "\xB2\xD1\xA1\x00\xB3\xED\xA1\x00\xB4\xF9\xA1\x00\xB5\xF1\xA1\x00\xB6\xE9"
-     "\xA1\x00"
-     "\xB7\xE5\xA1\x00\xB8\xA1\xA1\x00\xB9\xAD\xA1\x00\xBA\xB9\xA1\x00\xBB\xB5"
-     "\xA1\x00"
-     "\xBC\x11\xA3\x00\xBD\x95\xA1\x00\xBE\x79\xD5\x00\xBF\x95\xA1\x00\x80\x71"
-     "\xA2\x00"
-     "\x81\x91\xAE\x00\x82\x71\xA2\x00\x83\xA5\xAE\x00\x84\x99\xD6\x00\x85\x35"
-     "\xAE\x00"
-     "\x86\xC9\x82\x02\x87\x71\xA2\x00\x88\xE9\xA2\x00\x89\xA5\xA2\x00\x8A\xA1"
-     "\xA2\x00"
-     "\x8B\x91\xA2\x00\x8C\xA1\xA2\x00\x8D\xB9\x8D\x02\x8E\xF9\xA2\x00\x8F\x3D"
-     "\x8E\x02"
-     "\x90\xD9\xA2\x00\x91\x65\x8E\x02\x92\xE9\xA2\x00\x93\x59\xAE\x00\x94\xA9"
-     "\x82\x02"
-     "\x95\xFD\xA2\x00\x96\xAD\xA9\x00\x97\x55\x8E\x02\x98\xDD\x83\x02\x99\xE1"
-     "\xA2\x00"
-     "\x9A\xD9\xA2\x00\x9B\xD5\xA2\x00\x9C\x61\x8E\x02\x9D\x71\x8E\x02\x9E\x61"
-     "\x8E\x02"
-     "\x9F\xAD\xA2\x00\xA0\x01\xA0\x00\xA1\x8D\xA2\x00\xA2\x95\xAE\x00\xA3\x81"
-     "\xA2\x00"
-     "\xA4\x81\xA2\x00\xA5\xA1\xAE\x00\xA6\x81\xA2\x00\xA7\xBD\xA2\x00\xA8\x89"
-     "\xA2\x00"
-     "\xA9\x81\xA2\x00\xAA\x99\xA2\x00\xAB\x95\xA2\x00\xAC\xF1\xA2\x00\xAD\xFD"
-     "\xA2\x00"
-     "\xAE\xE9\xA2\x00\xAF\xE5\xA2\x00\xB0\x21\xA0\x00\xB1\xA5\xA2\x00\xB2\x49"
-     "\xD6\x00"
-     "\xB3\xA5\xA2\x00\xB4\xA1\xA2\x00\xB5\x41\xAE\x00\xB6\xA1\xA2\x00\xB7\x75"
-     "\xAE\x00"
-     "\xB8\x69\xD6\x00\xB9\xC5\xAE\x00\xBA\x39\x82\x02\xBB\x81\xA2\x00\xBC\x39"
-     "\xA2\x00"
-     "\xBD\x75\xA2\x00\xBE\x71\xA2\x00\xBF\x41\xA2\x00\x80\x91\xA3\x00\x81\x89"
-     "\x8C\x02"
-     "\x82\xC9\xA3\x00\x83\x0D\x8F\x02\x84\x89\xA3\x00\x85\x35\x8F\x02\x86\xB9"
-     "\xA3\x00"
-     "\x87\x09\xAF\x00\x88\xD9\x83\x02\x89\x8D\xA3\x00\x8A\xDD\xA8\x00\x8B\x25"
-     "\x8F\x02"
-     "\x8C\x8D\x82\x02\x8D\xB1\xA3\x00\x8E\x89\xA3\x00\x8F\x85\xA3\x00\x90\x51"
-     "\x8F\x02"
-     "\x91\x41\x8F\x02\x92\x51\x8F\x02\x93\x9D\xA3\x00\x94\xD1\xA1\x00\x95\x5D"
-     "\xA3\x00"
-     "\x96\x45\xAF\x00\x97\x51\xA3\x00\x98\x71\xA3\x00\x99\x51\xAF\x00\x9A\x71"
-     "\xA3\x00"
-     "\x9B\x4D\xA3\x00\x9C\x59\xA3\x00\x9D\x51\xA3\x00\x9E\x49\xA3\x00\x9F\x45"
-     "\xA3\x00"
-     "\xA0\xC1\xA3\x00\xA1\xCD\xA3\x00\xA2\xD9\xA3\x00\xA3\xD5\xA3\x00\xA4\x71"
-     "\xA1\x00"
-     "\xA5\xF5\xA3\x00\xA6\x19\xD7\x00\xA7\xF5\xA3\x00\xA8\xD1\xA3\x00\xA9\x31"
-     "\xAF\x00"
-     "\xAA\xD1\xA3\x00\xAB\x05\xAF\x00\xAC\x39\xD7\x00\xAD\x95\xAF\x00\xAE\x69"
-     "\x83\x02"
-     "\xAF\xD1\xA3\x00\xB0\x09\xA3\x00\xB1\x45\xA3\x00\xB2\x41\xA3\x00\xB3\x71"
-     "\xA3\x00"
-     "\xB4\x41\xA3\x00\xB5\x59\x8C\x02\xB6\x19\xA3\x00\xB7\xDD\x8F\x02\xB8\x79"
-     "\xA3\x00"
-     "\xB9\xC5\x8F\x02\xBA\x49\xA3\x00\xBB\xF9\xAF\x00\xBC\x09\x83\x02\xBD\x5D"
-     "\xA3\x00"
-     "\xBE\x0D\xA8\x00\xBF\xF5\x8F\x02\xB5\x85\xAC\x00\xB4\xB9\x8D\x02\xB7\xB1"
-     "\xAC\x00"
-     "\xB6\xBD\xAC\x00\xB1\xB9\xAC\x00\xB0\xED\x8C\x02\xB3\x11\x80\x02\xB2\xE9"
-     "\xA7\x00"
-     "\xBD\x29\xAC\x00\xBC\xA5\xAE\x00\xBF\x25\xAC\x00\xBE\x31\xA0\x00\xB9\x35"
-     "\x80\x02"
-     "\xB8\x25\x80\x02\xBB\xE9\xAC\x00\xBA\x25\x80\x02\xA5\xC1\xAC\x00\xA4\x8D"
-     "\xAC\x00"
-     "\xA7\xF5\xAC\x00\xA6\xC5\xAC\x00\xA1\x71\xA0\x00\xA0\xDD\xD8\x00\xA3\x35"
-     "\xAC\x00"
-     "\xA2\x8D\x8C\x02\xAD\x41\x80\x02\xAC\xFD\xAC\x00\xAF\x7D\xA0\x00\xAE\xCD"
-     "\xAC\x00"
-     "\xA9\xFD\x83\x02\xA8\xE5\xAC\x00\xAB\x79\x80\x02\xAA\xBD\xAC\x00\x95\xC9"
-     "\xAC\x00"
-     "\x94\xC5\xAC\x00\x97\xD1\xAC\x00\x96\xDD\xAC\x00\x91\xB5\xAC\x00\x90\xBD"
-     "\xAC\x00"
-     "\x93\xA1\xAC\x00\x92\xAD\xAC\x00\x9D\x35\xA0\x00\x9C\xD5\xAC\x00\x9F\x01"
-     "\xA0\x00"
-     "\x9E\xD5\xAC\x00\x99\xD1\xAC\x00\x98\x55\xAE\x00\x9B\xD1\xAC\x00\x9A\x3D"
-     "\xD8\x00"
-     "\x85\xC5\x80\x02\x84\xD5\x80\x02\x87\x19\xAC\x00\x86\xD5\x80\x02\x81\x55"
-     "\xAC\x00"
-     "\x80\x69\x8D\x02\x83\x61\xAC\x00\x82\x6D\xAC\x00\x8D\xD5\xA0\x00\x8C\xF5"
-     "\xAC\x00"
-     "\x8F\xC9\xAC\x00\x8E\xF5\xAC\x00\x89\xF9\xAC\x00\x88\x75\xAE\x00\x8B\xF5"
-     "\xAC\x00"
-     "\x8A\xE1\xA0\x00\x80\x11\xAD\x00\x81\x31\xA1\x00\x82\x11\xAD\x00\x83\x2D"
-     "\xAD\x00"
-     "\x84\x39\xAD\x00\x85\x31\xAD\x00\x86\x29\xAD\x00\x87\x25\xAD\x00\x88\x61"
-     "\xAD\x00"
-     "\x89\x6D\xAD\x00\x8A\x79\xAD\x00\x8B\x75\xAD\x00\x8C\xD1\xAF\x00\x8D\x55"
-     "\xAD\x00"
-     "\x8E\xB9\xD9\x00\x8F\x55\xAD\x00\x90\x31\xAD\x00\x91\xD1\xA1\x00\x92\x31"
-     "\xAD\x00"
-     "\x93\xE5\xA1\x00\x94\xD9\xD9\x00\x95\x75\xA1\x00\x96\x89\x8D\x02\x97\x31"
-     "\xAD\x00"
-     "\x98\xA9\xAD\x00\x99\xE5\xAD\x00\x9A\xE1\xAD\x00\x9B\xD1\xAD\x00\x9C\xE1"
-     "\xAD\x00"
-     "\x9D\xF9\x82\x02\x9E\xB9\xAD\x00\x9F\x7D\x81\x02\xA0\x19\xAD\x00\xA1\xA5"
-     "\x81\x02"
-     "\xA2\x29\xAD\x00\xA3\x99\xA1\x00\xA4\x69\x8D\x02\xA5\x3D\xAD\x00\xA6\x6D"
-     "\xA6\x00"
-     "\xA7\x95\x81\x02\xA8\x1D\x8C\x02\xA9\x21\xAD\x00\xAA\x19\xAD\x00\xAB\x15"
-     "\xAD\x00"
-     "\xAC\xA1\x81\x02\xAD\xB1\x81\x02\xAE\xA1\x81\x02\xAF\x6D\xAD\x00\xB0\x41"
-     "\xAF\x00"
-     "\xB1\xCD\xAD\x00\xB2\xD5\xA1\x00\xB3\xC1\xAD\x00\xB4\xC1\xAD\x00\xB5\xE1"
-     "\xA1\x00"
-     "\xB6\xC1\xAD\x00\xB7\xFD\xAD\x00\xB8\xC9\xAD\x00\xB9\xC1\xAD\x00\xBA\xD9"
-     "\xAD\x00"
-     "\xBB\xD5\xAD\x00\xBC\xB1\xAD\x00\xBD\xBD\xAD\x00\xBE\xA9\xAD\x00\xBF\xA5"
-     "\xAD\x00"
-     "\x80\xE1\xAC\x00\x81\x65\xAE\x00\x82\x89\xDA\x00\x83\x65\xAE\x00\x84\x61"
-     "\xAE\x00"
-     "\x85\x81\xA2\x00\x86\x61\xAE\x00\x87\xB5\xA2\x00\x88\xA9\xDA\x00\x89\x05"
-     "\xA2\x00"
-     "\x8A\xF9\x8E\x02\x8B\x41\xAE\x00\x8C\xF9\xAE\x00\x8D\xB5\xAE\x00\x8E\xB1"
-     "\xAE\x00"
-     "\x8F\x81\xAE\x00\x90\xD1\xAE\x00\x91\xC9\x81\x02\x92\x89\xAE\x00\x93\x4D"
-     "\x82\x02"
-     "\x94\xC9\xAE\x00\x95\x75\x82\x02\x96\xF9\xAE\x00\x97\x49\xA2\x00\x98\x99"
-     "\x8E\x02"
-     "\x99\xCD\xAE\x00\x9A\x9D\xA5\x00\x9B\x65\x82\x02\x9C\xCD\x8F\x02\x9D\xF1"
-     "\xAE\x00"
-     "\x9E\xC9\xAE\x00\x9F\xC5\xAE\x00\xA0\x91\x82\x02\xA1\x81\x82\x02\xA2\x91"
-     "\x82\x02"
-     "\xA3\x5D\xAE\x00\xA4\x0D\x14\x00\xA5\xBD\x15\x00\x4C\x01\x00\x80\x50\x01"
-     "\x00\x80"
-     "\xA8\x01\x04\x00\xA9\x89\x15\x00\xAA\x61\xAD\x00\xAB\x99\x15\x00\xAC\xB1"
-     "\x06\x00"
-     "\xAD\x8D\x15\x00\xAE\xB1\x06\x00\xAF\xA5\xB8\x02\xB0\xD1\x06\x00\xB1\xF9"
-     "\x15\x00"
-     "\xB2\x89\x15\x00\xB3\x89\x15\x00\xB4\x99\x15\x00\xB5\x99\x15\x00\xB6\xC1"
-     "\x06\x00"
-     "\xB7\x11\xAD\x00\xB8\xC1\x06\x00\xB9\xFD\xB8\x02\xBA\xF5\xB8\x02\xBB\xBD"
-     "\x15\x00"
-     "\xBC\xBD\xBF\x02\xBD\xA1\x15\x00\xBE\xA1\x15\x00\xBF\xA1\x15\x00\x80\x29"
-     "\xA3\x00"
-     "\x81\x65\x14\x00\x82\x21\x07\x00\x83\xE1\xBA\x02\x84\x39\xA8\x00\x85\x39"
-     "\xA8\x00"
-     "\x86\x11\xBF\x02\x87\x79\x14\x00\x88\x11\x07\x00\x89\xD5\xBA\x02\x8A\xCD"
-     "\xAF\x00"
-     "\x8B\x2D\xB9\x02\x8C\x59\x14\x00\x8D\xD9\x8F\x02\x8E\x55\x14\x00\x8F\x45"
-     "\x14\x00"
-     "\x90\x79\x07\x00\x91\xB9\xBA\x02\x92\x49\xBF\x02\x93\x45\xB9\x02\x94\x2D"
-     "\x14\x00"
-     "\x95\xD5\x14\x00\x96\x19\x07\x00\x97\xA9\xBA\x02\x98\x5D\xA8\x00\x99\xE1"
-     "\x14\x00"
-     "\x9A\x21\x07\x00\x9B\xE5\x14\x00\x9C\x79\xBF\x02\x9D\x85\xBA\x02\x9E\x79"
-     "\xB9\x02"
-     "\x9F\x89\xBA\x02\xA0\x81\xBF\x02\xA1\x79\xBA\x02\xA2\x29\x05\x00\xA3\xA1"
-     "\x14\x00"
-     "\xA4\x59\xAC\x00\xA5\xA1\x14\x00\xA6\x99\x07\x00\xA7\xA5\x14\x00\xA8\xA9"
-     "\x07\x00"
-     "\xA9\xBD\xB9\x02\xAA\xB9\x07\x00\xAB\x91\x14\x00\xAC\xF1\x14\x00\xAD\xF1"
-     "\x14\x00"
-     "\xAE\xF1\x14\x00\xAF\xF1\x14\x00\xB0\xD9\x07\x00\xB1\x09\xAC\x00\xB2\xE9"
-     "\x07\x00"
-     "\xB3\xD5\xB9\x02\xB4\xCD\xB9\x02\xB5\x85\x14\x00\xB6\x95\xBE\x02\xB7\x89"
-     "\x14\x00"
-     "\xB8\xB9\x14\x00\xB9\xB9\x14\x00\xBA\xC1\xA3\x00\xBB\x8D\x14\x00\xBC\xD9"
-     "\x07\x00"
-     "\xBD\x19\xBA\x02\xBE\xD1\xA8\x00\xBF\xD1\xA8\x00\xEE\x0D\x00\x80\x54\x01"
-     "\x00\x80"
-     "\xF4\x0D\x00\x80\x58\x01\x00\x80\xFA\x0D\x00\x80\x5C\x01\x00\x80\x0C\x0E"
-     "\x00\x80"
-     "\x13\x00\x00\x00\x12\x0E\x00\x80\xD7\x0E\x00\x80\x3F\x0E\x00\x80\x60\x01"
-     "\x00\x80"
-     "\x51\x0E\x00\x80\x64\x01\x00\x80\x89\x0E\x00\x80\x68\x01\x00\x80\x9B\x0E"
-     "\x00\x80"
-     "\x6C\x01\x00\x80\xA7\x0E\x00\x80\x70\x01\x00\x80\x80\x50\x02\x00\x81\x48"
-     "\x00\x00"
-     "\xE0\x0E\x00\x80\x74\x01\x00\x80\xE6\x0E\x00\x80\x78\x01\x00\x80\xE9\x0E"
-     "\x00\x80"
-     "\xEC\x0E\x00\x80\xEF\x0E\x00\x80\x7C\x01\x00\x80\xF5\x0E\x00\x80\x80\x01"
-     "\x00\x80"
-     "\x01\x0F\x00\x80\x10\x0F\x00\x80\x1C\x0F\x00\x80\x25\x0F\x00\x80\x2B\x0F"
-     "\x00\x80"
-     "\x7F\x0F\x00\x80\x3D\x0F\x00\x80\xA6\x0F\x00\x80\x84\x01\x00\x80\x88\x01"
-     "\x00\x80"
-     "\x8C\x01\x00\x80\x90\x01\x00\x80\xB8\xB0\x3D\x00\xB9\x00\x38\x00\xBA\xB8"
-     "\x3A\x00"
-     "\x94\x01\x00\x80\xB0\xE5\x00\x00\xB1\x5D\x13\x00\xB2\x55\x13\x00\xB3\x95"
-     "\x00\x00"
-     "\xB4\x71\x13\x00\xB5\xB5\x00\x00\xB6\xB1\x00\x00\xB7\xB1\x00\x00\xB8\x91"
-     "\x00\x00"
-     "\xB9\x9D\x00\x00\xBA\x4D\x13\x00\xBB\x45\x13\x00\xBC\x71\x00\x00\xBD\x59"
-     "\x13\x00"
-     "\xBE\x49\x13\x00\xBF\x49\x13\x00\xA8\x89\xAB\x00\xA9\x89\xAB\x00\xAA\xA9"
-     "\xAB\x00"
-     "\xAB\xA9\xAB\x00\xAC\xE5\xA8\x00\xAD\xE1\x89\x02\xAE\xC5\x89\x02\xAF\xFD"
-     "\x89\x02"
-     "\xA0\x81\xAB\x00\xA1\x81\xAB\x00\xA2\xDD\xA8\x00\xA3\xD9\x89\x02\xA4\xED"
-     "\x89\x02"
-     "\xA5\xD5\x89\x02\xA6\xC1\xA8\x00\xA7\xC1\xA8\x00\xB8\x9D\x89\x02\xB9\xA5"
-     "\x89\x02"
-     "\xBA\xB1\xA8\x00\xBB\xB1\xA8\x00\xBC\xD9\xAB\x00\xBD\xD9\xAB\x00\xBE\xF9"
-     "\xAB\x00"
-     "\xBF\xF9\xAB\x00\xB0\x99\xA8\x00\xB1\x99\xA8\x00\xB2\xE1\xAB\x00\xB3\xE1"
-     "\xAB\x00"
-     "\xB4\xD1\xAB\x00\xB5\xD1\xAB\x00\xB6\x8D\xA8\x00\xB7\x89\x89\x02\x88\x81"
-     "\xB9\x02"
-     "\x89\x79\xBC\x02\x8A\x05\x12\x00\x8B\x19\x12\x00\xCC\xF8\x87\x02\xCD\x38"
-     "\x87\x02"
-     "\x8E\x6D\xA8\x00\x8F\x69\x89\x02\x80\x9D\xAE\x00\x81\x21\x12\x00\x82\xE1"
-     "\x01\x00"
-     "\x83\x25\x12\x00\x84\xB9\xB9\x02\x85\x45\xBC\x02\x86\xB9\xBF\x02\x87\x49"
-     "\xBC\x02"
-     "\x98\x35\xA8\x00\x99\x31\x89\x02\x9A\x15\x89\x02\x9B\x2D\x89\x02\x9C\x29"
-     "\xA8\x00"
-     "\x9D\x29\xA8\x00\x9E\x51\xAB\x00\x9F\x51\xAB\x00\x90\x3D\x89\x02\x91\x05"
-     "\x89\x02"
-     "\x92\x11\xA8\x00\x93\x11\xA8\x00\x94\x79\xAB\x00\x95\x79\xAB\x00\x96\x59"
-     "\xAB\x00"
-     "\x97\x59\xAB\x00\x87\x8D\x02\x00\x86\xC9\x11\x00\x85\x81\x02\x00\x84\xC5"
-     "\x11\x00"
-     "\x83\xD5\x11\x00\x82\xD5\x11\x00\x81\x9D\x02\x00\x80\xA5\x11\x00\x8F\xFD"
-     "\x11\x00"
-     "\x8E\xB9\x02\x00\x8D\xF1\x11\x00\x8C\xB5\x02\x00\x8B\xA5\x02\x00\x8A\xE1"
-     "\x11\x00"
-     "\x89\xE9\x11\x00\x88\xAD\x02\x00\x97\x29\x02\x00\x96\x25\x02\x00\x95\xE9"
-     "\x11\x00"
-     "\x94\x15\x02\x00\x93\xCD\x11\x00\x92\xCD\x11\x00\x91\x7D\x02\x00\x90\xBD"
-     "\x11\x00"
-     "\x9F\xE5\x02\x00\x9E\xD1\x11\x00\x9D\xC5\x11\x00\x9C\xED\x02\x00\x9B\xD5"
-     "\x11\x00"
-     "\x9A\x0D\x02\x00\x99\x09\x02\x00\x98\x11\x02\x00\xA7\x65\x11\x00\xA6\x09"
-     "\x02\x00"
-     "\xA5\x3D\xFA\x00\xA4\x41\x11\x00\xA3\x19\x02\x00\xA2\x5D\x11\x00\xA1\x5D"
-     "\x11\x00"
-     "\xA0\x5D\x11\x00\xE2\x0D\x00\x80\x98\x01\x00\x80\xAD\x25\x0C\x00\xAC\x29"
-     "\x0C\x00"
-     "\xAB\x31\x0C\x00\xAA\xC1\x0C\x00\xA9\xED\x0F\x00\xA8\x25\x02\x00\x9C\x01"
-     "\x00\x80"
-     "\x43\x24\x00\x80\xD1\x06\x00\x80\xCB\x07\x00\x80\x0F\x07\x00\x80\x97\x07"
-     "\x00\x80"
-     "\x9D\x07\x00\x80\x80\x07\x00\x80\xC8\x07\x00\x80\x4F\x07\x00\x80\x52\x07"
-     "\x00\x80"
-     "\x5D\x07\x00\x80\xE1\x06\x00\x80\x0D\x25\x00\x80\xE0\x26\x00\x80\x2D\x2F"
-     "\x00\x80"
-     "\xCC\xA8\x80\x02\xCD\x4C\x82\x02\xCC\x20\xBD\x02\xCD\x4C\xBD\x02\xCC\x40"
-     "\x8A\x02"
-     "\xCC\xC8\xBE\x02\xCC\x98\x81\x02\xCD\x3C\x80\x02\xCC\x10\x87\x02\xCD\x50"
-     "\x87\x02"
-     "\xCC\xA0\x87\x02\xCD\x8C\x87\x02\xCC\x30\x84\x02\xCD\x24\x84\x02\xCC\x18"
-     "\x84\x02"
-     "\xCD\x84\x84\x02\xCC\x38\xB9\x02\xCD\x08\xB9\x02\xCC\x18\xB9\x02\xCD\x1C"
-     "\xB9\x02"
-     "\xCC\xB0\xBE\x02\xCD\x58\xB9\x02\xCC\x50\xB9\x02\xCC\x60\xB9\x02\xB2\x2D"
-     "\x00\x80"
-     "\xAF\xE0\x3A\x00\xCC\xD0\xBE\x02\xCD\xD4\xBE\x02\xCC\xF0\xBD\x02\xCD\x84"
-     "\xBD\x02"
-     "\xCC\x50\xBC\x02\xCD\xAC\xBF\x02\xCE\x2E\x00\x80\x37\x38\x00\x80\xFF\x2B"
-     "\x00\x80"
-     "\xCB\x31\x00\x80\xB7\x33\x00\x80\x06\x2D\x00\x80\x58\x30\x00\x80\x9E\x2C"
-     "\x00\x80"
-     "\xCC\x84\xB8\x02\xCC\xD8\xB8\x02\xCC\x84\xB8\x02\xCC\xF8\xB8\x02\xCC\x14"
-     "\xB9\x02"
-     "\xCC\x4C\xB9\x02\xCC\x20\xB9\x02\xCC\x70\xB9\x02\x86\x08\x13\x00\xA4\x31"
-     "\x00\x80"
-     "\x84\xC8\x38\x00\x85\xCC\x3A\x00\xF2\x2C\x00\x80\x1F\x36\x00\x80\x93\x38"
-     "\x00\x80"
-     "\xB2\x2E\x00\x80\xEA\x2D\x00\x80\x6B\x33\x00\x80\x08\x31\x00\x80\x34\x31"
-     "\x00\x80"
-     "\x13\x39\x00\x80\x28\x31\x00\x80\x88\x00\x34\x00\x89\x7C\x3A\x00\xB1\x55"
-     "\x2F\x00"
-     "\xB0\x49\x85\x02\xB3\x55\x2F\x00\xB2\x55\x2F\x00\xB5\x71\x2F\x00\xB4\x3D"
-     "\x98\x00"
-     "\xB7\xF5\x81\x02\xB6\x35\x3C\x00\xB9\x0D\x93\x00\xB8\x0D\x93\x00\xBB\x4D"
-     "\x2F\x00"
-     "\xBA\x25\x84\x02\xBD\xC1\x81\x02\xBC\x05\x3C\x00\xBF\x39\x82\x02\xBE\xD9"
-     "\x94\x00"
-     "\xA1\x79\x2F\x00\xA0\x45\x3C\x00\xA3\x51\x82\x02\xA2\x45\x3C\x00\xA5\x6D"
-     "\x2F\x00"
-     "\xA4\x45\x3C\x00\xA7\x1D\x2F\x00\xA6\x1D\x2F\x00\xA9\x2D\x2F\x00\xA8\x2D"
-     "\x2F\x00"
-     "\xAB\xA5\x97\x00\xAA\x75\x3C\x00\xAD\x69\x82\x02\xAC\x55\x3C\x00\xAF\x29"
-     "\x2F\x00"
-     "\xAE\x61\x82\x02\x91\x75\x81\x02\x90\xC5\x3C\x00\x93\x0D\x2F\x00\x92\xB1"
-     "\x93\x00"
-     "\x95\x19\x2F\x00\x94\xDD\x3C\x00\x97\x69\x81\x02\x96\x95\x84\x02\x99\x55"
-     "\x81\x02"
-     "\x98\xA5\x82\x02\x9B\x55\x81\x02\x9A\xAD\x84\x02\x9D\x9D\x2F\x00\x9C\x15"
-     "\x3E\x00"
-     "\x9F\x8D\x2F\x00\x9E\x75\x97\x00\x81\xA5\x2F\x00\x80\xCD\x84\x02\x83\x39"
-     "\x81\x02"
-     "\x82\xFD\x3C\x00\x85\xD1\x82\x02\x84\x31\x94\x00\x87\x35\xB4\x02\x86\xB5"
-     "\x2F\x00"
-     "\x89\x99\x2F\x00\x88\x89\x2F\x00\x8B\x15\x81\x02\x8A\xD5\x3C\x00\x8D\xF9"
-     "\x82\x02"
-     "\x8C\xF5\x84\x02\x8F\x79\x2F\x00\x8E\x81\x2F\x00\x80\x69\x2E\x00\x81\xE9"
-     "\xB5\x02"
-     "\x82\x65\x2E\x00\x83\x75\x2E\x00\x84\x29\x3D\x00\x85\xE9\x80\x02\x86\x19"
-     "\x85\x02"
-     "\x87\x15\x83\x02\x88\x5D\x2E\x00\x89\xA5\x2E\x00\x8A\x69\x3D\x00\x8B\xD9"
-     "\x80\x02"
-     "\x8C\x0D\x92\x00\x8D\xB1\x2E\x00\x8E\x71\x3D\x00\x8F\xB5\x2E\x00\x90\x49"
-     "\x85\x02"
-     "\x91\xB5\x80\x02\x92\x49\x83\x02\x93\xB9\x80\x02\x94\x51\x85\x02\x95\xA9"
-     "\x80\x02"
-     "\x96\xF9\x3F\x00\x97\x71\x2E\x00\x98\xA9\x96\x00\x99\x51\x2E\x00\x9A\x69"
-     "\x3D\x00"
-     "\x9B\x55\x2E\x00\x9C\x79\x3D\x00\x9D\x6D\x83\x02\x9E\x69\x3D\x00\x9F\x41"
-     "\x2E\x00"
-     "\xA0\xC1\x2E\x00\xA1\xC1\x2E\x00\xA2\xC1\x2E\x00\xA3\xC1\x2E\x00\xA4\x89"
-     "\x3D\x00"
-     "\xA5\x59\x96\x00\xA6\xB9\x3D\x00\xA7\x85\x83\x02\xA8\xBD\x83\x02\xA9\xF5"
-     "\x2E\x00"
-     "\xAA\xE5\x84\x02\xAB\xF9\x2E\x00\xAC\xE9\x2E\x00\xAD\xE9\x2E\x00\xAE\x91"
-     "\x99\x00"
-     "\xAF\xDD\x2E\x00\xB0\xE9\x3D\x00\xB1\x29\x80\x02\xB2\xE1\x92\x00\xB3\xE1"
-     "\x92\x00"
-     "\xB4\xD9\x85\x02\xB5\xB1\x2E\x00\xB6\xE9\x3D\x00\xB7\x2D\x80\x02\xB8\x05"
-     "\x95\x00"
-     "\xB9\xE5\x83\x02\xBA\x81\x2E\x00\xBB\x01\xB5\x02\xBC\x9D\x2E\x00\xBD\x8D"
-     "\x2E\x00"
-     "\xBE\xC1\x3D\x00\xBF\x01\x80\x02\x80\x01\x86\x02\x81\x0D\x80\x02\x82\x75"
-     "\x2D\x00"
-     "\x83\x8D\x2D\x00\x84\x51\x3E\x00\x85\xE1\x83\x02\x86\x25\x91\x00\x87\x99"
-     "\x2D\x00"
-     "\x88\x69\x3E\x00\x89\xAD\x2D\x00\x8A\x21\x86\x02\x8B\xDD\x83\x02\x8C\x31"
-     "\x80\x02"
-     "\x8D\xC1\x83\x02\x8E\x39\x86\x02\x8F\xC1\x83\x02\x90\xE1\x3C\x00\x91\x69"
-     "\x2D\x00"
-     "\x92\x81\x95\x00\x93\x79\x2D\x00\x94\x51\x3E\x00\x95\x6D\x2D\x00\x96\x51"
-     "\x3E\x00"
-     "\x97\x45\x80\x02\x98\x71\x3E\x00\x99\x59\x2D\x00\x9A\x29\x2D\x00\x9B\x29"
-     "\x2D\x00"
-     "\x9C\x39\x2D\x00\x9D\x39\x2D\x00\x9E\x61\x3E\x00\x9F\xB1\x95\x00\xA0\xA1"
-     "\x3E\x00"
-     "\xA1\x9D\x80\x02\xA2\x95\x80\x02\xA3\xDD\x2D\x00\xA4\xDD\x87\x02\xA5\xC1"
-     "\x2D\x00"
-     "\xA6\xC1\x2D\x00\xA7\xC1\x2D\x00\xA8\x89\x9A\x00\xA9\xC5\x2D\x00\xAA\x81"
-     "\x3E\x00"
-     "\xAB\x41\x83\x02\xAC\x99\x91\x00\xAD\x99\x91\x00\xAE\xB1\x86\x02\xAF\xD9"
-     "\x2D\x00"
-     "\xB0\xF1\x3E\x00\xB1\x35\x83\x02\xB2\x2D\x96\x00\xB3\xCD\x80\x02\xB4\xB9"
-     "\x2D\x00"
-     "\xB5\x39\xB6\x02\xB6\xB5\x2D\x00\xB7\xA5\x2D\x00\xB8\xD9\x3E\x00\xB9\x19"
-     "\x83\x02"
-     "\xBA\xE9\x86\x02\xBB\xE5\x80\x02\xBC\x8D\x2D\x00\xBD\x75\x2D\x00\xBE\xB9"
-     "\x3E\x00"
-     "\xBF\x09\x83\x02\x80\x0D\x27\x00\x81\x9D\x20\x00\x82\xA1\x20\x00\x83\xB9"
-     "\x20\x00"
-     "\x00\x31\x00\x80\x85\x59\x27\x00\x86\xED\x20\x00\x87\xB1\x20\x00\x88\x39"
-     "\x27\x00"
-     "\x89\x7D\x27\x00\x8A\x19\x27\x00\x8B\x19\x27\x00\x8C\x09\x27\x00\x8D\x09"
-     "\x27\x00"
-     "\x8E\xC1\x20\x00\x8F\x1D\x27\x00\x90\x69\x27\x00\x91\x55\x27\x00\x92\x65"
-     "\x27\x00"
-     "\x93\xF1\x20\x00\x94\xAD\x20\x00\x95\xF5\x20\x00\x96\xFD\x20\x00\x97\xED"
-     "\x20\x00"
-     "\x98\xD9\x20\x00\x99\x71\x27\x00\x9A\x75\x27\x00\x9B\x49\x27\x00\x9C\xE1"
-     "\x31\x00"
-     "\x9D\x81\xD7\x00\x9E\xE5\x31\x00\x9F\xDD\x31\x00\x77\x35\x00\x80\xA1\x1D"
-     "\x20\x00"
-     "\xA2\x21\x20\x00\xB3\x35\x00\x80\xA4\xD9\x27\x00\xD6\x2D\x00\x80\x4B\x33"
-     "\x00\x80"
-     "\xA7\x31\x20\x00\x9F\x34\x00\x80\xA9\xFD\x27\x00\xAA\x99\x27\x00\xAB\x99"
-     "\x27\x00"
-     "\xAC\x89\x27\x00\xAD\x89\x27\x00\xAE\x41\x20\x00\xAF\x9D\x27\x00\xB0\xE9"
-     "\x27\x00"
-     "\xB1\xD5\x27\x00\xB2\xE5\x27\x00\x0A\x2D\x00\x80\xB4\x2D\x20\x00\xB5\x75"
-     "\x20\x00"
-     "\xB6\x7D\x20\x00\xB7\x6D\x20\x00\x36\x2E\x00\x80\xB9\xF1\x27\x00\x4B\x39"
-     "\x00\x80"
-     "\xBB\xC9\x27\x00\x87\x24\x00\x80\x4B\x24\x00\x80\x55\x24\x00\x80\x69\x24"
-     "\x00\x80"
-     "\x80\x0D\x20\x00\x81\x9D\x27\x00\x82\xA1\x27\x00\x83\xB9\x27\x00\x84\x59"
-     "\x20\x00"
-     "\x85\x59\x20\x00\x86\xED\x27\x00\x87\xB1\x27\x00\x88\x39\x20\x00\x89\x7D"
-     "\x20\x00"
-     "\x91\x24\x00\x80\x8B\x19\x20\x00\x8C\x09\x20\x00\x8D\x09\x20\x00\x8E\xC1"
-     "\x27\x00"
-     "\x8F\x1D\x20\x00\x90\x69\x20\x00\x91\x55\x20\x00\x92\x65\x20\x00\x93\xF1"
-     "\x27\x00"
-     "\x94\xAD\x27\x00\x95\xF5\x27\x00\x96\xFD\x27\x00\x97\xED\x27\x00\x98\xD9"
-     "\x27\x00"
-     "\x99\x71\x20\x00\x9A\x75\x20\x00\x9B\x49\x20\x00\x73\x24\x00\x80\x9B\x24"
-     "\x00\x80"
-     "\x5F\x24\x00\x80\x7D\x24\x00\x80\x5C\x30\x00\x80\xA1\x1D\x27\x00\xA2\x21"
-     "\x27\x00"
-     "\xA3\x39\x27\x00\xC6\x2D\x00\x80\xA5\xD9\x20\x00\xA6\x6D\x27\x00\xA7\x31"
-     "\x27\x00"
-     "\xA8\xB9\x20\x00\xA9\xFD\x20\x00\xD3\x2B\x00\x80\xAB\x99\x20\x00\xAC\x89"
-     "\x20\x00"
-     "\xAD\x89\x20\x00\xAE\x41\x27\x00\xAF\x9D\x20\x00\xB0\xE9\x20\x00\xB1\xD5"
-     "\x20\x00"
-     "\xB2\xE5\x20\x00\xB3\x71\x27\x00\xB4\x2D\x27\x00\xB5\x75\x27\x00\xB6\x7D"
-     "\x27\x00"
-     "\xB7\x6D\x27\x00\xB8\x59\x27\x00\xB9\xF1\x20\x00\xBA\xF5\x20\x00\xBB\xC9"
-     "\x20\x00"
-     "\xCB\x2B\x00\x80\xDB\x2B\x00\x80\xF9\xEA\x00\x80\x26\x2C\x00\x80\x80\xB5"
-     "\x37\x00"
-     "\x81\xED\x37\x00\x82\xF9\x37\x00\x83\x1D\x36\x00\x84\x1D\x31\x00\x85\x0D"
-     "\x31\x00"
-     "\x86\x89\x31\x00\x87\xAD\x31\x00\x88\xE9\x31\x00\x89\xF1\x31\x00\x8A\xD5"
-     "\x31\x00"
-     "\x3A\x2C\x00\x80\x42\x2C\x00\x80\x4A\x2C\x00\x80\x4E\x2C\x00\x80\x52\x2C"
-     "\x00\x80"
-     "\x90\x55\x30\x00\x91\x59\x30\x00\x92\x69\x30\x00\x93\x69\x30\x00\x94\x15"
-     "\x30\x00"
-     "\x95\xCD\x3F\x00\x96\xC5\x3F\x00\x97\x59\x39\x00\x98\x6D\x39\x00\x99\x71"
-     "\x39\x00"
-     "\x9A\x45\x39\x00\x9B\x59\x39\x00\x9C\x4D\x39\x00\x9D\x31\x39\x00\x9E\x35"
-     "\x39\x00"
-     "\x9F\x21\x39\x00\xA0\xE5\x39\x00\xA1\xF1\x39\x00\xA2\xF5\x39\x00\xA3\x2D"
-     "\x39\x00"
-     "\xA4\x35\x39\x00\xA5\x3D\x39\x00\xA6\x35\x39\x00\xA7\x21\x39\x00\xA8\x05"
-     "\x3B\x00"
-     "\xA9\x65\x3B\x00\xAA\x6D\x3B\x00\xAB\xB1\x9D\x00\xAC\xD1\x91\x00\xAD\x7D"
-     "\x3B\x00"
-     "\xAE\x75\x3B\x00\x8B\x2A\x00\x80\xB0\x41\x93\x00\xB1\xCD\x91\x00\xB2\xD5"
-     "\x9D\x00"
-     "\xB3\xC1\x91\x00\xB4\xC1\x91\x00\xB5\xE1\x9D\x00\xB6\xC1\x91\x00\xB7\xFD"
-     "\x91\x00"
-     "\xB8\xC9\x91\x00\xB9\xC1\x91\x00\xBA\xD9\x91\x00\xBB\xD5\x91\x00\xBC\xB1"
-     "\x91\x00"
-     "\xBD\xBD\x91\x00\xBE\xA9\x91\x00\xBF\xA5\x91\x00\xAD\x3D\x22\x00\xAC\x3D"
-     "\x22\x00"
-     "\xAF\x29\x22\x00\xAE\xF5\x25\x00\xA9\x49\x22\x00\xA8\x0D\x22\x00\x0D\xEB"
-     "\x00\x80"
-     "\xAA\x2D\x22\x00\x72\x2C\x00\x80\xA4\x6D\x22\x00\xA7\x85\x25\x00\x76\x2C"
-     "\x00\x80"
-     "\xA1\xA9\x25\x00\x92\x2C\x00\x80\xFE\xEA\x00\x80\xA2\x95\x25\x00\x8F\x2A"
-     "\x00\x80"
-     "\xBC\xD5\x34\x00\x7E\x2C\x00\x80\xBE\xD1\x34\x00\xB9\x45\x22\x00\x08\xEB"
-     "\x00\x80"
-     "\xBB\x7D\x22\x00\xBA\x41\x22\x00\xB5\xC1\x25\x00\xB4\x99\x25\x00\xB7\xD9"
-     "\x25\x00"
-     "\xB6\xC9\x25\x00\xB1\x61\x22\x00\xB0\x5D\x22\x00\xAE\x2C\x00\x80\xB2\x51"
-     "\x22\x00"
-     "\x8D\xBD\x22\x00\xB6\x2C\x00\x80\x8F\xA9\x22\x00\x8E\x75\x25\x00\x89\xC9"
-     "\x22\x00"
-     "\x0F\x2C\x00\x80\x8B\xAD\x22\x00\xBE\x2C\x00\x80\xE3\xEC\x00\x80\xE6\x2C"
-     "\x00\x80"
-     "\x87\x05\x25\x00\x97\x2A\x00\x80\xFE\x2C\x00\x80\x02\x2D\x00\x80\x0E\x2D"
-     "\x00\x80"
-     "\x82\x15\x25\x00\x9D\x35\xD2\x00\x9B\x2A\x00\x80\x9F\x69\x34\x00\x46\x2D"
-     "\x00\x80"
-     "\x99\xC5\x22\x00\x4A\x2D\x00\x80\x9B\xFD\x22\x00\x72\x2D\x00\x80\x7E\x2D"
-     "\x00\x80"
-     "\x94\x19\x25\x00\x97\x59\x25\x00\x8E\x2D\x00\x80\x91\xE1\x22\x00\x96\x2D"
-     "\x00\x80"
-     "\x9A\x2D\x00\x80\x92\xD1\x22\x00\x80\xE1\x91\x00\x81\x65\x93\x00\x82\x89"
-     "\xE7\x00"
-     "\x83\x65\x93\x00\x84\x61\x93\x00\x85\x81\x9F\x00\x86\x61\x93\x00\x87\xB5"
-     "\x9F\x00"
-     "\x88\xA9\xE7\x00\x89\x05\x9F\x00\x8A\xF9\x39\x00\x8B\xF1\xF9\x00\x8C\xE5"
-     "\x39\x00"
-     "\x8D\xED\x39\x00\x8E\xE5\x39\x00\x8F\xDD\x39\x00\xB1\xB5\xB3\x02\xB0\xB1"
-     "\x92\x00"
-     "\xB3\xA9\xB3\x02\xB2\x91\xB3\x02\xB5\xAD\x92\x00\xB4\xAD\x92\x00\xB7\xD5"
-     "\x91\x00"
-     "\xB6\xD5\x91\x00\xB9\xC5\x91\x00\xB8\xC5\x91\x00\x1C\xEB\x00\x80\x9B\x33"
-     "\x00\x80"
-     "\xAE\x2D\x00\x80\xB6\x2D\x00\x80\x21\xEB\x00\x80\xC2\x2D\x00\x80\x80\x71"
-     "\x04\x00"
-     "\x81\x71\x04\x00\x82\x71\x04\x00\x83\x71\x04\x00\x84\xB9\x03\x00\x85\xE5"
-     "\x03\x00"
-     "\x86\xED\x03\x00\x87\xE5\x03\x00\x88\xDD\x03\x00\xCA\x2D\x00\x80\xAA\x4D"
-     "\x39\x00"
-     "\xAB\x45\x39\x00\xAC\x5D\x39\x00\xD2\x2D\x00\x80\xFA\x2D\x00\x80\xE6\x2D"
-     "\x00\x80"
-     "\x90\x41\x03\x00\x91\x55\x03\x00\xA8\xC8\x12\x00\xF2\x2D\x00\x80\xF6\x2D"
-     "\x00\x80"
-     "\xFE\x2D\x00\x80\x06\x2E\x00\x80\x0A\x2E\x00\x80\xA2\x00\x08\x00\xA3\x8C"
-     "\x0B\x00"
-     "\xA0\xE8\x0D\x00\xA1\xEC\x0F\x00\xA6\x00\x10\x00\xA7\xF4\x11\x00\xA4\x00"
-     "\x14\x00"
-     "\xA5\xC8\x17\x00\x80\xA9\x0C\x00\x81\xF5\x0C\x00\x82\xFD\x0C\x00\x83\xF5"
-     "\x0C\x00"
-     "\x84\xED\x0C\x00\x85\xD1\xC5\x00\x86\x35\x0D\x00\x87\x2D\x0D\x00\x88\x15"
-     "\x0D\x00"
-     "\x89\x1D\x0D\x00\x8A\xE1\xC5\x00\x8B\x11\x0D\x00\x8C\x8D\x0D\x00\x8D\x2D"
-     "\x0E\x00"
-     "\x8E\xF5\xC5\x00\x8F\x1D\x0E\x00\x90\x6D\x0E\x00\x91\x71\x0E\x00\x92\x71"
-     "\x0E\x00"
-     "\x93\x0D\x0E\x00\x94\x19\x0E\x00\x95\xC9\x37\x00\x96\x0D\x0E\x00\x97\x25"
-     "\x0E\x00"
-     "\x98\xE5\x0E\x00\x99\xF1\x0E\x00\x9A\xF5\x0E\x00\x9B\x15\xC7\x00\x9C\xD1"
-     "\x0E\x00"
-     "\x9D\x81\xE8\x00\x9E\xDD\x0E\x00\x9F\xD1\x0E\x00\xA0\x31\x0E\x00\xA1\x31"
-     "\x0E\x00"
-     "\xA2\x45\x0B\x00\xA3\x35\x0E\x00\xA4\x55\x0E\x00\xA5\x19\xC7\x00\xA6\x49"
-     "\xC5\x00"
-     "\xA7\x49\xC5\x00\xA8\x31\xC7\x00\xA9\x69\x0E\x00\xAA\x7D\x0E\x00\xAB\x55"
-     "\xCE\x00"
-     "\xAC\x6D\x0E\x00\xAD\x6D\xC5\x00\xAE\x59\x0E\x00\xAF\x51\x0E\x00\xB0\x35"
-     "\x0E\x00"
-     "\xB1\x3D\x0E\x00\xB2\x31\x0E\x00\xB3\x35\x0E\x00\xB4\xB9\x0F\x00\xB5\xB9"
-     "\x0F\x00"
-     "\xB6\xA9\x0F\x00\xB7\xA9\x0F\x00\xB8\x99\x0F\x00\xB9\x99\x0F\x00\xBA\x4D"
-     "\x0F\x00"
-     "\xBB\x59\x0F\x00\xBC\x49\x0F\x00\xBD\x49\x0F\x00\xBE\x35\x0F\x00\xBF\x2D"
-     "\x0F\x00"
-     "\x12\x2E\x00\x80\x1A\x2E\x00\x80\x1E\x2E\x00\x80\x2E\x2E\x00\x80\xBA\xF1"
-     "\x0A\x00"
-     "\xBB\xFD\x0A\x00\xB8\xB9\x0A\x00\xB9\xD1\x0A\x00\xB6\x25\xFE\x00\xB7\xE5"
-     "\xEA\x00"
-     "\xB4\xBD\x0A\x00\xB5\xB5\x0A\x00\xB2\xB1\x0A\x00\xB3\xB9\x0A\x00\xB0\x21"
-     "\xE1\x00"
-     "\xB1\x2D\x0A\x00\xAE\x0D\xFE\x00\xAF\xB1\x36\x00\xAC\x01\xFE\x00\xAD\x65"
-     "\xFA\x00"
-     "\xAA\x4D\xFA\x00\xAB\xA5\x36\x00\xA8\xB5\x36\x00\xA9\xE9\xE9\x00\xA6\x85"
-     "\x36\x00"
-     "\xA7\x8D\x36\x00\xA4\xB5\x36\x00\xA5\x8D\x36\x00\xA2\x49\xEC\x00\xA3\xAD"
-     "\x36\x00"
-     "\xA0\xA9\x36\x00\xA1\xA9\x36\x00\x9E\x1D\x36\x00\x9F\x41\x36\x00\x9C\x1D"
-     "\x36\x00"
-     "\x9D\x15\x36\x00\x9A\x01\x36\x00\x9B\xC5\xC2\x00\x98\x11\x36\x00\x99\x11"
-     "\x36\x00"
-     "\x96\x25\x36\x00\x97\xC1\xFA\x00\x94\x65\xE9\x00\x95\x2D\x36\x00\x92\xD1"
-     "\x36\x00"
-     "\x93\x39\xE7\x00\x90\xC1\xEA\x00\x91\x41\x37\x00\x32\x2E\x00\x80\x3E\x2E"
-     "\x00\x80"
-     "\x62\x2E\x00\x80\x66\x2E\x00\x80\x4A\x2E\x00\x80\x52\x2E\x00\x80\xAE\x2E"
-     "\x00\x80"
-     "\x76\x2E\x00\x80\x7A\x2E\x00\x80\x7E\x2E\x00\x80\x72\x2E\x00\x80\x86\x2E"
-     "\x00\x80"
-     "\x82\xE5\xF5\x00\x82\x2E\x00\x80\x80\x0D\x37\x00\x81\x05\x37\x00\x80\xD5"
-     "\x0D\x00"
-     "\x81\xDD\x0D\x00\x82\xD5\x0D\x00\x83\x0D\x03\x00\x84\x15\x03\x00\x85\x1D"
-     "\x03\x00"
-     "\x86\x11\x03\x00\x87\x85\xC5\x00\x88\x2D\x03\x00\x89\xE5\x03\x00\x8A\xED"
-     "\x03\x00"
-     "\x8B\xE5\x03\x00\x8C\xD1\xC7\x00\x8D\xF9\x03\x00\x8E\xE9\x03\x00\x8F\xE9"
-     "\x03\x00"
-     "\x90\x01\xC7\x00\x91\x9D\x03\x00\x92\x95\x03\x00\x93\xAD\x03\x00\x94\xB5"
-     "\x03\x00"
-     "\x95\xBD\x03\x00\x96\xB5\x03\x00\x97\xA9\x03\x00\x98\x61\x01\x00\x99\x61"
-     "\x01\x00"
-     "\x9A\x61\x01\x00\x9B\x61\x01\x00\x9C\x61\x01\x00\x9D\xF5\x35\x00\x9E\x6D"
-     "\x01\x00"
-     "\x9F\x65\x01\x00\xA0\x9D\x01\x00\xA1\xA5\x01\x00\xA2\xAD\x01\x00\xA3\xA5"
-     "\x01\x00"
-     "\xA4\xBD\x01\x00\xA5\xA5\x01\x00\xA6\xAD\x01\x00\xA7\xA5\x01\x00\xA8\x9D"
-     "\x01\x00"
-     "\xA9\xE5\x01\x00\xAA\xED\x01\x00\xAB\xE9\x01\x00\xAC\xF9\x01\x00\xAD\xF9"
-     "\x01\x00"
-     "\xAE\xE9\x01\x00\xAF\x55\xCC\x00\xB0\x95\x01\x00\xB1\x9D\x01\x00\xB2\x95"
-     "\x01\x00"
-     "\xB3\xAD\x01\x00\xB4\xB5\x01\x00\xB5\xD9\xE9\x00\xB6\xB1\x01\x00\xB7\xB1"
-     "\x01\x00"
-     "\xB8\xF5\xE9\x00\xB9\x95\x01\x00\xBA\x9D\x01\x00\xBB\x95\x01\x00\xBC\x8D"
-     "\x01\x00"
-     "\xBD\x75\x01\x00\xBE\x7D\x01\x00\xBF\x75\x01\x00\x9A\x2E\x00\x80\x26\xEB"
-     "\x00\x80"
-     "\xB6\x2E\x00\x80\xBA\x2E\x00\x80\xC2\x2E\x00\x80\xD6\x2E\x00\x80\xF9\x2E"
-     "\x00\x80"
-     "\x2B\xEB\x00\x80\x30\xEB\x00\x80\x05\x2F\x00\x80\x09\x2F\x00\x80\x0D\x2F"
-     "\x00\x80"
-     "\x01\x2F\x00\x80\x15\x2F\x00\x80\x9F\x2A\x00\x80\xA3\x2A\x00\x80\x1D\x2F"
-     "\x00\x80"
-     "\x25\x2F\x00\x80\x35\xEB\x00\x80\x41\x2F\x00\x80\x45\x2F\x00\x80\x51\x2F"
-     "\x00\x80"
-     "\x3A\xEB\x00\x80\x59\x2F\x00\x80\x5D\x2F\x00\x80\x3C\x30\x00\x80\xA7\x2A"
-     "\x00\x80"
-     "\x71\x2F\x00\x80\x8D\x2F\x00\x80\x89\x2F\x00\x80\x44\xEB\x00\x80\x95\x2F"
-     "\x00\x80"
-     "\x49\xEB\x00\x80\xA1\x2F\x00\x80\x9D\x2F\x00\x80\xA5\x2F\x00\x80\xB5\x2F"
-     "\x00\x80"
-     "\xB9\x2F\x00\x80\xAB\x2A\x00\x80\xC9\x2F\x00\x80\xD1\x2F\x00\x80\xD5\x2F"
-     "\x00\x80"
-     "\xD9\x2F\x00\x80\xAF\x2A\x00\x80\x4E\xEB\x00\x80\xB3\x2A\x00\x80\xFC\x2F"
-     "\x00\x80"
-     "\x00\x30\x00\x80\x04\x30\x00\x80\x15\xED\x00\x80\x58\xEB\x00\x80\xDF\x36"
-     "\x00\x80"
-     "\x34\x30\x00\x80\xB7\x2A\x00\x80\x7B\xEB\x00\x80\x4D\xEC\x00\x80\x44\x30"
-     "\x00\x80"
-     "\x4C\x30\x00\x80\xBB\x2A\x00\x80\x60\x30\x00\x80\x70\x30\x00\x80\x74\x30"
-     "\x00\x80"
-     "\xFF\x00\x00\x00\x7C\x30\x00\x80\x8C\x30\x00\x80\xC3\x2A\x00\x80\xBF\x2A"
-     "\x00\x80"
-     "\x5D\xEB\x00\x80\x94\x30\x00\x80\xA8\x30\x00\x80\xAC\x30\x00\x80\xBC\x30"
-     "\x00\x80"
-     "\xD0\x30\x00\x80\xC8\x30\x00\x80\xCC\x30\x00\x80\xD4\x30\x00\x80\xE8\x30"
-     "\x00\x80"
-     "\xEC\x30\x00\x80\x04\x31\x00\x80\x0C\x31\x00\x80\x1C\x31\x00\x80\x30\x31"
-     "\x00\x80"
-     "\x6C\xEB\x00\x80\x2C\x31\x00\x80\x20\x31\x00\x80\x38\x31\x00\x80\x40\x31"
-     "\x00\x80"
-     "\x54\x31\x00\x80\x71\xEB\x00\x80\x5C\x31\x00\x80\x5A\x00\x00\x00\x50\x31"
-     "\x00\x80"
-     "\x48\x31\x00\x80\xC7\x2A\x00\x80\x64\x31\x00\x80\x6C\x31\x00\x80\x74\x31"
-     "\x00\x80"
-     "\x68\x31\x00\x80\xCB\x2A\x00\x80\x8C\x31\x00\x80\x76\xEB\x00\x80\xB8\x31"
-     "\x00\x80"
-     "\xFA\x31\x00\x80\xD6\x31\x00\x80\xD7\x2A\x00\x80\xD3\x2A\x00\x80\xCF\x2A"
-     "\x00\x80"
-     "\xA2\x2C\x00\x80\xA6\x2C\x00\x80\xFE\x31\x00\x80\xE6\x31\x00\x80\xB3\x36"
-     "\x00\x80"
-     "\x3B\x2B\x00\x80\x12\x32\x00\x80\x26\x32\x00\x80\x1E\x32\x00\x80\x8F\xEB"
-     "\x00\x80"
-     "\xDB\x2A\x00\x80\x3A\x32\x00\x80\x2E\x32\x00\x80\x57\x32\x00\x80\x99\xEB"
-     "\x00\x80"
-     "\x63\x32\x00\x80\x42\x32\x00\x80\x6B\x32\x00\x80\x6F\x32\x00\x80\x73\x32"
-     "\x00\x80"
-     "\x77\x32\x00\x80\xB8\x49\x01\x00\xB9\x49\x01\x00\xBA\x59\x01\x00\xBB\x59"
-     "\x01\x00"
-     "\xBC\x49\x01\x00\xBD\x49\x01\x00\xBE\x79\x01\x00\xBF\x79\x01\x00\xB0\x19"
-     "\xC3\x00"
-     "\xB1\x15\xC8\x00\xB2\x19\x01\x00\xB3\x19\x01\x00\xB4\x09\x01\x00\xB5\x09"
-     "\x01\x00"
-     "\xB6\x79\x01\x00\xB7\x79\x01\x00\xA8\x89\xC2\x00\xA9\x5D\x01\x00\xAA\x55"
-     "\x01\x00"
-     "\xAB\xB1\xC0\x00\xAC\x71\x01\x00\xAD\x71\x01\x00\xAE\x71\x01\x00\xAF\x71"
-     "\x01\x00"
-     "\xA0\x61\x01\x00\xA1\x61\x01\x00\xA2\x61\x01\x00\xA3\x99\xC0\x00\xA4\x7D"
-     "\x01\x00"
-     "\xA5\x65\x01\x00\xA6\x6D\x01\x00\xA7\x65\x01\x00\x98\x5D\x06\x00\x99\x65"
-     "\x06\x00"
-     "\x9A\x6D\x06\x00\x9B\x65\x06\x00\x9C\x7D\x06\x00\x9D\x65\x06\x00\x9E\x6D"
-     "\x06\x00"
-     "\x9F\x61\x06\x00\x90\xED\xEE\x00\x91\x6D\x06\x00\x92\x61\x06\x00\x93\x61"
-     "\x06\x00"
-     "\x94\x61\x06\x00\x95\x65\x06\x00\x96\x6D\x06\x00\x97\x65\x06\x00\x88\x6D"
-     "\x06\x00"
-     "\x89\x75\x06\x00\x8A\x7D\x06\x00\x8B\x75\x06\x00\x8C\x6D\x06\x00\x8D\x15"
-     "\x06\x00"
-     "\x8E\x45\xC8\x00\x8F\x19\x06\x00\x80\x6D\x06\x00\x81\x75\x06\x00\x82\x7D"
-     "\x06\x00"
-     "\x83\x75\x06\x00\x84\x6D\x06\x00\x85\x55\x06\x00\x86\x5D\x06\x00\x87\x55"
-     "\x06\x00"
-     "\x9A\xC5\x02\x00\x9B\xCD\x02\x00\x98\xA9\xCB\x00\x99\xE5\xC1\x00\x9E\xC5"
-     "\x02\x00"
-     "\x9F\xCD\x02\x00\x9C\xD5\x02\x00\x9D\xCD\x02\x00\x92\x8D\x02\x00\x93\x95"
-     "\x02\x00"
-     "\x90\x8D\x02\x00\x91\x85\x02\x00\x96\x8D\x02\x00\x97\xF5\x02\x00\x94\x8D"
-     "\x02\x00"
-     "\x95\x85\x02\x00\x8A\xD1\x02\x00\x8B\xD1\x02\x00\x88\x75\xC0\x00\x89\xC1"
-     "\x02\x00"
-     "\x8E\xF1\x02\x00\x8F\x19\xC0\x00\x8C\xC1\x02\x00\x8D\xC1\x02\x00\x82\x15"
-     "\x02\x00"
-     "\x83\x1D\x02\x00\x80\x05\x02\x00\x81\x1D\x02\x00\x86\xF5\x02\x00\x87\xFD"
-     "\x02\x00"
-     "\x84\x05\x02\x00\x85\xFD\x02\x00\xBA\xB9\x01\x00\xBB\xB9\x01\x00\xB8\xB9"
-     "\x01\x00"
-     "\xB9\xB9\x01\x00\xBE\xD9\x01\x00\xBF\xD9\x01\x00\xBC\xD9\x01\x00\xBD\xD9"
-     "\x01\x00"
-     "\xB2\xBD\x01\x00\xB3\x85\x01\x00\xB0\xBD\x01\x00\xB1\xB5\x01\x00\xB6\x99"
-     "\x01\x00"
-     "\xB7\x99\x01\x00\xB4\x9D\x01\x00\xB5\x91\xC7\x00\xAA\x3D\x02\x00\xAB\xC5"
-     "\x01\x00"
-     "\xA8\x3D\x02\x00\xA9\x35\x02\x00\xAE\xDD\x01\x00\xAF\xC5\x01\x00\xAC\xDD"
-     "\x01\x00"
-     "\xAD\xD5\x01\x00\xA2\xB1\xC0\x00\xA3\x01\x02\x00\xA0\x35\x02\x00\xA1\x0D"
-     "\x02\x00"
-     "\xA6\x01\x02\x00\xA7\x01\xDC\x00\xA4\x11\x02\x00\xA5\x11\x02\x00\x83\x32"
-     "\x00\x80"
-     "\x9E\xEB\x00\x80\x8B\x32\x00\x80\xE3\x2A\x00\x80\x97\x32\x00\x80\xA3\xEB"
-     "\x00\x80"
-     "\x9B\x32\x00\x80\xE7\x2A\x00\x80\xA7\x32\x00\x80\xB3\x32\x00\x80\xC3\x32"
-     "\x00\x80"
-     "\xA8\xEB\x00\x80\x3F\xEB\x00\x80\xAD\xEB\x00\x80\xE3\x32\x00\x80\xB2\xEB"
-     "\x00\x80"
-     "\xEF\x32\x00\x80\xF7\x32\x00\x80\xE7\x32\x00\x80\x03\x33\x00\x80\x13\x33"
-     "\x00\x80"
-     "\x1B\x33\x00\x80\x27\x33\x00\x80\x2B\x33\x00\x80\xB7\xEB\x00\x80\x0F\x33"
-     "\x00\x80"
-     "\x43\x33\x00\x80\x47\x33\x00\x80\xEB\x2A\x00\x80\x57\x33\x00\x80\xBC\xEB"
-     "\x00\x80"
-     "\x3B\x33\x00\x80\x77\x33\x00\x80\xC1\xEB\x00\x80\xC6\xEB\x00\x80\x7B\x33"
-     "\x00\x80"
-     "\x8F\x33\x00\x80\x87\x33\x00\x80\xEF\x2A\x00\x80\x93\x33\x00\x80\xA3\x33"
-     "\x00\x80"
-     "\x9F\x33\x00\x80\xAB\x33\x00\x80\x03\xEB\x00\x80\xBB\x33\x00\x80\xCB\xEB"
-     "\x00\x80"
-     "\xC7\x33\x00\x80\xD5\xEB\x00\x80\xDB\x33\x00\x80\xFB\x33\x00\x80\xDA\xEB"
-     "\x00\x80"
-     "\x0F\x34\x00\x80\x13\x34\x00\x80\xDF\xEB\x00\x80\xE4\xEB\x00\x80\x2F\x34"
-     "\x00\x80"
-     "\xB8\x00\x00\x00\x3F\x34\x00\x80\xF3\x2A\x00\x80\x43\x34\x00\x80\xF7\x2A"
-     "\x00\x80"
-     "\x5B\x34\x00\x80\x5F\x34\x00\x80\x6B\x34\x00\x80\x77\x34\x00\x80\xFB\x2A"
-     "\x00\x80"
-     "\x8F\x34\x00\x80\xE9\xEB\x00\x80\xAB\x34\x00\x80\xEE\xEB\x00\x80\x53\xEB"
-     "\x00\x80"
-     "\xC7\x34\x00\x80\xF3\xEB\x00\x80\xF8\xEB\x00\x80\xFD\xEB\x00\x80\xFF\x2A"
-     "\x00\x80"
-     "\x03\x2B\x00\x80\x4D\x00\x00\x00\x07\xEC\x00\x80\x02\xEC\x00\x80\x0C\xEC"
-     "\x00\x80"
-     "\x11\xEC\x00\x80\x07\x35\x00\x80\x0B\x35\x00\x80\x17\x35\x00\x80\x0F\x2B"
-     "\x00\x80"
-     "\x13\x2B\x00\x80\x1B\xEC\x00\x80\x2B\x35\x00\x80\x17\x2B\x00\x80\x20\xEC"
-     "\x00\x80"
-     "\x25\xEC\x00\x80\x2A\xEC\x00\x80\x9B\x35\x00\x80\x1B\x2B\x00\x80\xA7\x35"
-     "\x00\x80"
-     "\xAB\x35\x00\x80\x2F\xEC\x00\x80\x34\xEC\x00\x80\xC3\x35\x00\x80\x1F\x2B"
-     "\x00\x80"
-     "\x39\xEC\x00\x80\xD7\x35\x00\x80\xDB\x35\x00\x80\x23\x2B\x00\x80\x3E\xEC"
-     "\x00\x80"
-     "\xF7\x35\x00\x80\x27\x2B\x00\x80\x07\x36\x00\x80\xFF\x35\x00\x80\x0F\x36"
-     "\x00\x80"
-     "\x48\xEC\x00\x80\x27\x36\x00\x80\x2B\x2B\x00\x80\x33\x36\x00\x80\x3B\x36"
-     "\x00\x80"
-     "\x4B\x36\x00\x80\x2F\x2B\x00\x80\x52\xEC\x00\x80\x57\xEC\x00\x80\x33\x2B"
-     "\x00\x80"
-     "\x5C\xEC\x00\x80\x63\x36\x00\x80\x61\xEC\x00\x80\x6F\x36\x00\x80\x7B\x36"
-     "\x00\x80"
-     "\x66\xEC\x00\x80\x6B\xEC\x00\x80\xBE\x2D\x01\x00\xBF\x35\x01\x00\xBC\x2D"
-     "\x01\x00"
-     "\xBD\x25\x01\x00\xBA\x0D\x01\x00\xBB\x35\x01\x00\xB8\xED\xD7\x00\xB9\x05"
-     "\x01\x00"
-     "\xB6\x31\x01\x00\xB7\x31\x01\x00\xB4\x21\x01\x00\xB5\x21\x01\x00\xB2\x31"
-     "\x01\x00"
-     "\xB3\x31\x01\x00\xB0\x39\xC3\x00\xB1\xC1\x1E\x00\xAE\xB5\x1E\x00\xAF\xBD"
-     "\x1E\x00"
-     "\xAC\xB9\x1E\x00\xAD\xBD\x1E\x00\xAA\xB9\x1E\x00\xAB\xB9\x1E\x00\xA8\x85"
-     "\x1E\x00"
-     "\xA9\xB9\x1E\x00\xA6\xB5\x1E\x00\xA7\xBD\x1E\x00\xA4\xA5\x1E\x00\xA5\xBD"
-     "\x1E\x00"
-     "\xA2\xB5\x1E\x00\xA3\xBD\x1E\x00\xA0\xB9\x1E\x00\xA1\xA5\xC1\x00\x9E\x39"
-     "\x1E\x00"
-     "\x9F\x39\x1E\x00\x9C\x39\x1E\x00\x9D\x39\x1E\x00\x9A\x39\x1E\x00\x9B\x39"
-     "\x1E\x00"
-     "\x98\x39\x1E\x00\x99\x39\x1E\x00\x96\x39\x1E\x00\x97\x39\x1E\x00\x94\x51"
-     "\xDE\x00"
-     "\x95\x39\x1E\x00\x92\x3D\x1E\x00\x93\x25\x1E\x00\x90\x3D\x1E\x00\x91\x35"
-     "\x1E\x00"
-     "\x8E\xBD\x1E\x00\x8F\x45\x1E\x00\x8C\xBD\x1E\x00\x8D\xB5\x1E\x00\x8A\xA1"
-     "\x1E\x00"
-     "\x8B\x21\xDE\x00\x88\xB1\x1E\x00\x89\xB1\x1E\x00\x86\x81\x1E\x00\x87\x81"
-     "\x1E\x00"
-     "\x84\xB1\x1E\x00\x85\xB1\x1E\x00\x82\x89\xDA\x00\x83\xA1\x1E\x00\x80\xA9"
-     "\x1E\x00"
-     "\x81\x81\xDF\x00\x8D\x5D\x02\x00\x8C\x5D\x02\x00\x8F\xE1\xD7\x00\x8E\x6D"
-     "\x02\x00"
-     "\x89\xED\xC0\x00\x88\x09\xC0\x00\x8B\x4D\x02\x00\x8A\x4D\x02\x00\x85\x69"
-     "\x02\x00"
-     "\x84\x61\x02\x00\x87\x65\x02\x00\x86\x61\x02\x00\x81\x09\x02\x00\x80\x2D"
-     "\xC0\x00"
-     "\x83\x79\x02\x00\x82\x01\x02\x00\x9D\xC9\x02\x00\x9C\xC1\x02\x00\x9F\xC5"
-     "\x02\x00"
-     "\x9E\xC5\x02\x00\x99\x49\xDD\x00\x98\x2D\x02\x00\x9B\xD9\x02\x00\x9A\x21"
-     "\x02\x00"
-     "\x95\x11\x02\x00\x94\x09\x02\x00\x97\x1D\x02\x00\x96\xFD\xDF\x00\x91\x15"
-     "\x02\x00"
-     "\x90\x51\xC0\x00\x93\x75\xDD\x00\x92\x15\x02\x00\xAD\x45\x02\x00\xAC\x45"
-     "\x02\x00"
-     "\xAF\x45\x02\x00\xAE\x45\x02\x00\xA9\x25\x02\x00\xA8\x25\x02\x00\xAB\x25"
-     "\x02\x00"
-     "\xAA\x25\x02\x00\xA5\x05\x02\x00\xA4\x05\x02\x00\xA7\x05\x02\x00\xA6\x05"
-     "\x02\x00"
-     "\xA1\x25\x02\x00\xA0\x25\x02\x00\xA3\x25\x02\x00\xA2\x25\x02\x00\xBD\xC1"
-     "\x01\x00"
-     "\xBC\x39\x02\x00\xBF\xC1\x01\x00\xBE\xC9\x01\x00\xB9\x25\x02\x00\xB8\x25"
-     "\x02\x00"
-     "\xBB\x21\x02\x00\xBA\xED\xDD\x00\xB5\x05\x02\x00\xB4\x05\x02\x00\xB7\x05"
-     "\x02\x00"
-     "\xB6\x05\x02\x00\xB1\x25\x02\x00\xB0\x25\x02\x00\xB3\x25\x02\x00\xB2\x25"
-     "\x02\x00"
-     "\x97\x36\x00\x80\x70\xEC\x00\x80\x9F\x36\x00\x80\x80\xEB\x00\x80\x37\x2B"
-     "\x00\x80"
-     "\xB7\x36\x00\x80\xBB\x36\x00\x80\x3F\x2B\x00\x80\xBF\x36\x00\x80\x19\x2F"
-     "\x00\x80"
-     "\x75\xEC\x00\x80\x7A\xEC\x00\x80\x85\xEB\x00\x80\x8A\xEB\x00\x80\xE3\x36"
-     "\x00\x80"
-     "\xFB\x38\x00\x80\x43\x2B\x00\x80\x0F\x37\x00\x80\x0B\x37\x00\x80\x13\x37"
-     "\x00\x80"
-     "\x1E\x2D\x00\x80\x17\x37\x00\x80\x1B\x37\x00\x80\x1F\x37\x00\x80\x27\x37"
-     "\x00\x80"
-     "\x7F\xEC\x00\x80\x2B\x37\x00\x80\x3F\x37\x00\x80\x43\x37\x00\x80\x2F\x37"
-     "\x00\x80"
-     "\x47\x37\x00\x80\x5B\x37\x00\x80\x3B\x37\x00\x80\x4F\x37\x00\x80\x53\x37"
-     "\x00\x80"
-     "\x57\x37\x00\x80\x84\xEC\x00\x80\x8E\xEC\x00\x80\x89\xEC\x00\x80\x47\x2B"
-     "\x00\x80"
-     "\x77\x37\x00\x80\x7B\x37\x00\x80\x83\x37\x00\x80\xA2\xEC\x00\x80\x87\x37"
-     "\x00\x80"
-     "\x93\xEC\x00\x80\x4B\x2B\x00\x80\x4F\x2B\x00\x80\x98\xEC\x00\x80\x9D\xEC"
-     "\x00\x80"
-     "\x53\x2B\x00\x80\xA7\x37\x00\x80\xAF\x37\x00\x80\xB3\x37\x00\x80\xBF\x37"
-     "\x00\x80"
-     "\xBB\x37\x00\x80\xC7\x37\x00\x80\xC3\x37\x00\x80\xCB\x37\x00\x80\xCF\x37"
-     "\x00\x80"
-     "\xD7\x37\x00\x80\x57\x2B\x00\x80\xBE\x00\x00\x00\xDF\x37\x00\x80\x80\xFD"
-     "\x07\x00"
-     "\x81\x01\x02\x00\x82\x01\x02\x00\x83\x01\x02\x00\x84\xFD\xF1\x00\x85\x05"
-     "\x02\x00"
-     "\x86\x0D\x02\x00\x87\x05\x02\x00\x88\x3D\x02\x00\x89\x05\x02\x00\x8A\x0D"
-     "\x02\x00"
-     "\x8B\x05\x02\x00\x8C\x1D\x02\x00\x8D\x05\x02\x00\x8E\x0D\x02\x00\x8F\x05"
-     "\x02\x00"
-     "\x90\x15\xC6\x00\x91\x1D\xC6\x00\x92\x41\xFC\x00\x93\x71\x02\x00\x94\x11"
-     "\x02\x00"
-     "\x95\x11\x02\x00\x96\x11\x02\x00\x97\x11\x02\x00\x98\x31\x02\x00\x99\x31"
-     "\x02\x00"
-     "\x9A\x31\x02\x00\x9B\x31\x02\x00\x9C\x11\x02\x00\x9D\x11\x02\x00\x9E\x11"
-     "\x02\x00"
-     "\x9F\x21\xC6\x00\xA0\xED\x02\x00\xA1\xF5\x02\x00\xA2\xFD\x02\x00\xA3\xF5"
-     "\x02\x00"
-     "\xA4\xED\x02\x00\xA5\x15\x02\x00\xA6\x1D\x02\x00\xA7\x15\x02\x00\xA8\x2D"
-     "\x02\x00"
-     "\xA9\x35\x02\x00\xAA\x3D\x02\x00\xAB\x35\x02\x00\xAC\x2D\x02\x00\xAD\x15"
-     "\x02\x00"
-     "\xAE\x1D\x02\x00\xAF\x15\x02\x00\xB0\x6D\x02\x00\xB1\x75\x02\x00\xB2\x7D"
-     "\x02\x00"
-     "\xB3\x75\x02\x00\xB4\x6D\x02\x00\xB5\x15\x02\x00\xB6\x1D\x02\x00\xB7\x15"
-     "\x02\x00"
-     "\xB8\x2D\x02\x00\xB9\x35\x02\x00\xBA\x3D\x02\x00\xBB\x35\x02\x00\xBC\x2D"
-     "\x02\x00"
-     "\xBD\x15\x02\x00\xBE\x81\xC6\x00\xBF\x85\xC6\x00\xBC\xE9\x06\x00\xBD\xE9"
-     "\x06\x00"
-     "\xBE\xF9\x06\x00\xBF\xF9\x06\x00\xB8\xCD\x06\x00\xB9\xF5\x06\x00\xBA\xFD"
-     "\x06\x00"
-     "\xBB\x5D\xC7\x00\xB4\xF1\xD0\x00\xB5\xF5\x06\x00\xB6\xFD\x06\x00\xB7\xF5"
-     "\x06\x00"
-     "\xB0\xF1\x06\x00\xB1\xF1\x06\x00\xB2\xF1\x06\x00\xB3\xF1\x06\x00\xAC\xF1"
-     "\x06\x00"
-     "\xAD\xF1\x06\x00\xAE\xF1\x06\x00\xAF\xF1\x06\x00\xA8\xF1\x06\x00\xA9\xF1"
-     "\x06\x00"
-     "\xAA\xF1\x06\x00\xAB\xF1\x06\x00\xA4\xF1\x06\x00\xA5\xF1\x06\x00\xA6\xF1"
-     "\x06\x00"
-     "\xA7\xF1\x06\x00\xA0\xF1\x06\x00\xA1\xF1\x06\x00\xA2\xF1\x06\x00\xA3\xF1"
-     "\x06\x00"
-     "\x9C\xF5\x06\x00\x9D\xFD\x06\x00\x9E\xF5\x06\x00\x9F\xB1\xDA\x00\x98\xA9"
-     "\xDC\x00"
-     "\x99\xFD\x06\x00\x9A\xF5\x06\x00\x9B\xED\x06\x00\x94\xF9\x06\x00\x95\xF9"
-     "\x06\x00"
-     "\x96\xC9\x06\x00\x97\xC9\x06\x00\x90\xF9\x06\x00\x91\xF9\x06\x00\x92\xE9"
-     "\x06\x00"
-     "\x93\xE9\x06\x00\x8C\xF9\x06\x00\x8D\xF9\x06\x00\x8E\x89\x06\x00\x8F\x89"
-     "\x06\x00"
-     "\x88\xFD\x06\x00\x89\xE5\x06\x00\x8A\xED\x06\x00\x8B\x11\x1E\x00\x84\xFD"
-     "\x06\x00"
-     "\x85\xC5\x06\x00\x86\xCD\x06\x00\x87\xC5\x06\x00\x80\xFD\x06\x00\x81\xE5"
-     "\x06\x00"
-     "\x82\xED\x06\x00\x83\xE5\x06\x00\x80\x00\x00\x00\xE3\x37\x00\x80\x5B\x2B"
-     "\x00\x80"
-     "\xF3\x37\x00\x80\xA7\xEC\x00\x80\x03\x38\x00\x80\x07\x38\x00\x80\x5F\x2B"
-     "\x00\x80"
-     "\x13\x38\x00\x80\x93\x2A\x00\x80\xAC\xEC\x00\x80\xB1\xEC\x00\x80\x63\x2B"
-     "\x00\x80"
-     "\x67\x2B\x00\x80\x3F\x38\x00\x80\xB6\xEC\x00\x80\x97\x38\x00\x80\x9B\x38"
-     "\x00\x80"
-     "\xB3\x38\x00\x80\xBF\x38\x00\x80\xC0\xEC\x00\x80\x12\xEB\x00\x80\xCB\x38"
-     "\x00\x80"
-     "\xC7\x38\x00\x80\xD3\x38\x00\x80\x17\xEB\x00\x80\xDF\x38\x00\x80\xC5\xEC"
-     "\x00\x80"
-     "\xCA\xEC\x00\x80\x2B\x39\x00\x80\x37\x39\x00\x80\x3F\x39\x00\x80\xCF\xEC"
-     "\x00\x80"
-     "\x43\x39\x00\x80\x6F\x39\x00\x80\x7B\x39\x00\x80\x7F\x39\x00\x80\x77\x39"
-     "\x00\x80"
-     "\x8B\x39\x00\x80\x8F\x39\x00\x80\xD4\xEC\x00\x80\x9B\x39\x00\x80\x6B\x2B"
-     "\x00\x80"
-     "\xA3\x39\x00\x80\xD9\xEC\x00\x80\x6F\x2B\x00\x80\xDB\x39\x00\x80\xA9\x2F"
-     "\x00\x80"
-     "\xF3\x39\x00\x80\xDE\xEC\x00\x80\xE8\xEC\x00\x80\x73\x2B\x00\x80\x77\x2B"
-     "\x00\x80"
-     "\x1B\x3A\x00\x80\xED\xEC\x00\x80\x7B\x2B\x00\x80\xF2\xEC\x00\x80\x3B\x3A"
-     "\x00\x80"
-     "\xF7\xEC\x00\x80\x53\x3A\x00\x80\x7F\x2B\x00\x80\x63\x3A\x00\x80\x6F\x3A"
-     "\x00\x80"
-     "\x77\x3A\x00\x80\x80\xE5\x01\x00\x81\xED\x01\x00\x82\xE5\x01\x00\x83\xFD"
-     "\x01\x00"
-     "\x84\xE5\x01\x00\x85\xED\x01\x00\x86\xE5\x01\x00\x87\x65\x00\x00\x88\x5D"
-     "\x00\x00"
-     "\x89\xA5\x00\x00\x8A\x55\xC5\x00\x8B\xA9\x00\x00\x8C\xB9\x00\x00\x8D\xB9"
-     "\x00\x00"
-     "\x8E\xA9\x00\x00\x8F\xA9\x00\x00\x90\xD9\x00\x00\x91\xD9\x00\x00\x92\xE9"
-     "\x00\x00"
-     "\x93\xE9\x00\x00\x94\xF9\x00\x00\x95\x9D\xFF\x00\x96\xF5\x00\x00\x97\x85"
-     "\xFF\x00"
-     "\x98\xD1\x00\x00\x99\xD1\x00\x00\x9A\xD1\x00\x00\x9B\xD1\x00\x00\x9C\x9D"
-     "\xFF\x00"
-     "\x9D\xB5\x00\x00\x7F\x3A\x00\x80\x83\x2B\x00\x80\xFC\xEC\x00\x80\xAF\x3A"
-     "\x00\x80"
-     "\xBB\x3A\x00\x80\x87\x2B\x00\x80\x8B\x2B\x00\x80\xBF\x3A\x00\x80\x01\xED"
-     "\x00\x80"
-     "\x8F\x2B\x00\x80\x06\xED\x00\x80\x0B\xED\x00\x80\x10\xED\x00\x80\x93\x2B"
-     "\x00\x80"
-     "\xFF\x3A\x00\x80\x03\x3B\x00\x80\x0B\x3B\x00\x80\x13\x3B\x00\x80\x1A\xED"
-     "\x00\x80"
-     "\xF1\x00\x00\x00\xF2\x00\x00\x00\xF3\x00\x00\x00\xF4\x00\x00\x00\xF5\x00"
-     "\x00\x00"
-     "\xF6\x00\x00\x00\xF7\x00\x00\x00\xF8\x00\x00\x00\xF9\x00\x00\x00\xFA\x00"
-     "\x00\x00"
-     "\xFB\x00\x00\x00\xFC\x00\x00\x00\xFD\x00\x00\x00\xFE\x00\x00\x00\xFF\x00"
-     "\x00\x00"
-     "\x00\x20\x00\x20\xCC\x81\x00\x20\xCC\x83\x00\x20\xCC\x84\x00\x20\xCC\x85"
-     "\x00\x20"
-     "\xCC\x86\x00\x20\xCC\x87\x00\x20\xCC\x88\x00\x20\xCC\x88\xCC\x80\x00\x20"
-     "\xCC\x88"
-     "\xCC\x81\x00\x20\xCC\x88\xCD\x82\x00\x20\xCC\x8A\x00\x20\xCC\x8B\x00\x20"
-     "\xCC\x93"
-     "\x00\x20\xCC\x93\xCC\x80\x00\x20\xCC\x93\xCC\x81\x00\x20\xCC\x93\xCD\x82"
-     "\x00\x20"
-     "\xCC\x94\x00\x20\xCC\x94\xCC\x80\x00\x20\xCC\x94\xCC\x81\x00\x20\xCC\x94"
-     "\xCD\x82"
-     "\x00\x20\xCC\xA7\x00\x20\xCC\xA8\x00\x20\xCC\xB3\x00\x20\xCD\x82\x00\x20"
-     "\xCD\x85"
-     "\x00\x20\xD9\x8B\x00\x20\xD9\x8C\x00\x20\xD9\x8C\xD9\x91\x00\x20\xD9\x8D"
-     "\x00\x20"
-     "\xD9\x8D\xD9\x91\x00\x20\xD9\x8E\x00\x20\xD9\x8E\xD9\x91\x00\x20\xD9\x8F"
-     "\x00\x20"
-     "\xD9\x8F\xD9\x91\x00\x20\xD9\x90\x00\x20\xD9\x90\xD9\x91\x00\x20\xD9\x91"
-     "\x00\x20"
-     "\xD9\x91\xD9\xB0\x00\x20\xD9\x92\x00\x20\xE3\x82\x99\x00\x20\xE3\x82\x9A"
-     "\x00\x21"
-     "\x00\x21\x21\x00\x21\x3F\x00\x22\x00\x23\x00\x24\x00\x25\x00\x26\x00\x27"
-     "\x00\x28"
-     "\x00\x28\x31\x29\x00\x28\x31\x30\x29\x00\x28\x31\x31\x29\x00\x28\x31\x32"
-     "\x29\x00"
-     "\x28\x31\x33\x29\x00\x28\x31\x34\x29\x00\x28\x31\x35\x29\x00\x28\x31\x36"
-     "\x29\x00"
-     "\x28\x31\x37\x29\x00\x28\x31\x38\x29\x00\x28\x31\x39\x29\x00\x28\x32\x29"
-     "\x00\x28"
-     "\x32\x30\x29\x00\x28\x33\x29\x00\x28\x34\x29\x00\x28\x35\x29\x00\x28\x36"
-     "\x29\x00"
-     "\x28\x37\x29\x00\x28\x38\x29\x00\x28\x39\x29\x00\x28\x41\x29\x00\x28\x42"
-     "\x29\x00"
-     "\x28\x43\x29\x00\x28\x44\x29\x00\x28\x45\x29\x00\x28\x46\x29\x00\x28\x47"
-     "\x29\x00"
-     "\x28\x48\x29\x00\x28\x49\x29\x00\x28\x4A\x29\x00\x28\x4B\x29\x00\x28\x4C"
-     "\x29\x00"
-     "\x28\x4D\x29\x00\x28\x4E\x29\x00\x28\x4F\x29\x00\x28\x50\x29\x00\x28\x51"
-     "\x29\x00"
-     "\x28\x52\x29\x00\x28\x53\x29\x00\x28\x54\x29\x00\x28\x55\x29\x00\x28\x56"
-     "\x29\x00"
-     "\x28\x57\x29\x00\x28\x58\x29\x00\x28\x59\x29\x00\x28\x5A\x29\x00\x28\x61"
-     "\x29\x00"
-     "\x28\x62\x29\x00\x28\x63\x29\x00\x28\x64\x29\x00\x28\x65\x29\x00\x28\x66"
-     "\x29\x00"
-     "\x28\x67\x29\x00\x28\x68\x29\x00\x28\x69\x29\x00\x28\x6A\x29\x00\x28\x6B"
-     "\x29\x00"
-     "\x28\x6C\x29\x00\x28\x6D\x29\x00\x28\x6E\x29\x00\x28\x6F\x29\x00\x28\x70"
-     "\x29\x00"
-     "\x28\x71\x29\x00\x28\x72\x29\x00\x28\x73\x29\x00\x28\x74\x29\x00\x28\x75"
-     "\x29\x00"
-     "\x28\x76\x29\x00\x28\x77\x29\x00\x28\x78\x29\x00\x28\x79\x29\x00\x28\x7A"
-     "\x29\x00"
-     "\x28\xE1\x84\x80\x29\x00\x28\xE1\x84\x82\x29\x00\x28\xE1\x84\x83\x29\x00"
-     "\x28\xE1"
-     "\x84\x85\x29\x00\x28\xE1\x84\x86\x29\x00\x28\xE1\x84\x87\x29\x00\x28\xE1"
-     "\x84\x89"
-     "\x29\x00\x28\xE1\x84\x8B\x29\x00\x28\xE1\x84\x8C\x29\x00\x28\xE1\x84\x8E"
-     "\x29\x00"
-     "\x28\xE1\x84\x8F\x29\x00\x28\xE1\x84\x90\x29\x00\x28\xE1\x84\x91\x29\x00"
-     "\x28\xE1"
-     "\x84\x92\x29\x00\x28\xE4\xB8\x80\x29\x00\x28\xE4\xB8\x83\x29\x00\x28\xE4"
-     "\xB8\x89"
-     "\x29\x00\x28\xE4\xB9\x9D\x29\x00\x28\xE4\xBA\x8C\x29\x00\x28\xE4\xBA\x94"
-     "\x29\x00"
-     "\x28\xE4\xBB\xA3\x29\x00\x28\xE4\xBC\x81\x29\x00\x28\xE4\xBC\x91\x29\x00"
-     "\x28\xE5"
-     "\x85\xAB\x29\x00\x28\xE5\x85\xAD\x29\x00\x28\xE5\x8A\xB4\x29\x00\x28\xE5"
-     "\x8D\x81"
-     "\x29\x00\x28\xE5\x8D\x94\x29\x00\x28\xE5\x90\x8D\x29\x00\x28\xE5\x91\xBC"
-     "\x29\x00"
-     "\x28\xE5\x9B\x9B\x29\x00\x28\xE5\x9C\x9F\x29\x00\x28\xE5\xAD\xA6\x29\x00"
-     "\x28\xE6"
-     "\x97\xA5\x29\x00\x28\xE6\x9C\x88\x29\x00\x28\xE6\x9C\x89\x29\x00\x28\xE6"
-     "\x9C\xA8"
-     "\x29\x00\x28\xE6\xA0\xAA\x29\x00\x28\xE6\xB0\xB4\x29\x00\x28\xE7\x81\xAB"
-     "\x29\x00"
-     "\x28\xE7\x89\xB9\x29\x00\x28\xE7\x9B\xA3\x29\x00\x28\xE7\xA4\xBE\x29\x00"
-     "\x28\xE7"
-     "\xA5\x9D\x29\x00\x28\xE7\xA5\xAD\x29\x00\x28\xE8\x87\xAA\x29\x00\x28\xE8"
-     "\x87\xB3"
-     "\x29\x00\x28\xE8\xB2\xA1\x29\x00\x28\xE8\xB3\x87\x29\x00\x28\xE9\x87\x91"
-     "\x29\x00"
-     "\x28\xEA\xB0\x80\x29\x00\x28\xEB\x82\x98\x29\x00\x28\xEB\x8B\xA4\x29\x00"
-     "\x28\xEB"
-     "\x9D\xBC\x29\x00\x28\xEB\xA7\x88\x29\x00\x28\xEB\xB0\x94\x29\x00\x28\xEC"
-     "\x82\xAC"
-     "\x29\x00\x28\xEC\x95\x84\x29\x00\x28\xEC\x98\xA4\xEC\xA0\x84\x29\x00\x28"
-     "\xEC\x98"
-     "\xA4\xED\x9B\x84\x29\x00\x28\xEC\x9E\x90\x29\x00\x28\xEC\xA3\xBC\x29\x00"
-     "\x28\xEC"
-     "\xB0\xA8\x29\x00\x28\xEC\xB9\xB4\x29\x00\x28\xED\x83\x80\x29\x00\x28\xED"
-     "\x8C\x8C"
-     "\x29\x00\x28\xED\x95\x98\x29\x00\x29\x00\x2A\x00\x2B\x00\x2C\x00\x2D\x00"
-     "\x2E\x00"
-     "\x2E\x2E\x00\x2E\x2E\x2E\x00\x2F\x00\x30\x00\x30\x2C\x00\x30\x2E\x00\x30"
-     "\xE2\x81"
-     "\x84\x33\x00\x30\xE7\x82\xB9\x00\x31\x00\x31\x2C\x00\x31\x2E\x00\x31\x30"
-     "\x00\x31"
-     "\x30\x2E\x00\x31\x30\xE6\x97\xA5\x00\x31\x30\xE6\x9C\x88\x00\x31\x30\xE7"
-     "\x82\xB9"
-     "\x00\x31\x31\x00\x31\x31\x2E\x00\x31\x31\xE6\x97\xA5\x00\x31\x31\xE6\x9C"
-     "\x88\x00"
-     "\x31\x31\xE7\x82\xB9\x00\x31\x32\x00\x31\x32\x2E\x00\x31\x32\xE6\x97\xA5"
-     "\x00\x31"
-     "\x32\xE6\x9C\x88\x00\x31\x32\xE7\x82\xB9\x00\x31\x33\x00\x31\x33\x2E\x00"
-     "\x31\x33"
-     "\xE6\x97\xA5\x00\x31\x33\xE7\x82\xB9\x00\x31\x34\x00\x31\x34\x2E\x00\x31"
-     "\x34\xE6"
-     "\x97\xA5\x00\x31\x34\xE7\x82\xB9\x00\x31\x35\x00\x31\x35\x2E\x00\x31\x35"
-     "\xE6\x97"
-     "\xA5\x00\x31\x35\xE7\x82\xB9\x00\x31\x36\x00\x31\x36\x2E\x00\x31\x36\xE6"
-     "\x97\xA5"
-     "\x00\x31\x36\xE7\x82\xB9\x00\x31\x37\x00\x31\x37\x2E\x00\x31\x37\xE6\x97"
-     "\xA5\x00"
-     "\x31\x37\xE7\x82\xB9\x00\x31\x38\x00\x31\x38\x2E\x00\x31\x38\xE6\x97\xA5"
-     "\x00\x31"
-     "\x38\xE7\x82\xB9\x00\x31\x39\x00\x31\x39\x2E\x00\x31\x39\xE6\x97\xA5\x00"
-     "\x31\x39"
-     "\xE7\x82\xB9\x00\x31\xE2\x81\x84\x00\x31\xE2\x81\x84\x31\x30\x00\x31\xE2"
-     "\x81\x84"
-     "\x32\x00\x31\xE2\x81\x84\x33\x00\x31\xE2\x81\x84\x34\x00\x31\xE2\x81\x84"
-     "\x35\x00"
-     "\x31\xE2\x81\x84\x36\x00\x31\xE2\x81\x84\x37\x00\x31\xE2\x81\x84\x38\x00"
-     "\x31\xE2"
-     "\x81\x84\x39\x00\x31\xE6\x97\xA5\x00\x31\xE6\x9C\x88\x00\x31\xE7\x82\xB9"
-     "\x00\x32"
-     "\x00\x32\x2C\x00\x32\x2E\x00\x32\x30\x00\x32\x30\x2E\x00\x32\x30\xE6\x97"
-     "\xA5\x00"
-     "\x32\x30\xE7\x82\xB9\x00\x32\x31\x00\x32\x31\xE6\x97\xA5\x00\x32\x31\xE7"
-     "\x82\xB9"
-     "\x00\x32\x32\x00\x32\x32\xE6\x97\xA5\x00\x32\x32\xE7\x82\xB9\x00\x32\x33"
-     "\x00\x32"
-     "\x33\xE6\x97\xA5\x00\x32\x33\xE7\x82\xB9\x00\x32\x34\x00\x32\x34\xE6\x97"
-     "\xA5\x00"
-     "\x32\x34\xE7\x82\xB9\x00\x32\x35\x00\x32\x35\xE6\x97\xA5\x00\x32\x36\x00"
-     "\x32\x36"
-     "\xE6\x97\xA5\x00\x32\x37\x00\x32\x37\xE6\x97\xA5\x00\x32\x38\x00\x32\x38"
-     "\xE6\x97"
-     "\xA5\x00\x32\x39\x00\x32\x39\xE6\x97\xA5\x00\x32\xE2\x81\x84\x33\x00\x32"
-     "\xE2\x81"
-     "\x84\x35\x00\x32\xE6\x97\xA5\x00\x32\xE6\x9C\x88\x00\x32\xE7\x82\xB9\x00"
-     "\x33\x00"
-     "\x33\x2C\x00\x33\x2E\x00\x33\x30\x00\x33\x30\xE6\x97\xA5\x00\x33\x31\x00"
-     "\x33\x31"
-     "\xE6\x97\xA5\x00\x33\x32\x00\x33\x33\x00\x33\x34\x00\x33\x35\x00\x33\x36"
-     "\x00\x33"
-     "\x37\x00\x33\x38\x00\x33\x39\x00\x33\xE2\x81\x84\x34\x00\x33\xE2\x81\x84"
-     "\x35\x00"
-     "\x33\xE2\x81\x84\x38\x00\x33\xE6\x97\xA5\x00\x33\xE6\x9C\x88\x00\x33\xE7"
-     "\x82\xB9"
-     "\x00\x34\x00\x34\x2C\x00\x34\x2E\x00\x34\x30\x00\x34\x31\x00\x34\x32\x00"
-     "\x34\x33"
-     "\x00\x34\x34\x00\x34\x35\x00\x34\x36\x00\x34\x37\x00\x34\x38\x00\x34\x39"
-     "\x00\x34"
-     "\xE2\x81\x84\x35\x00\x34\xE6\x97\xA5\x00\x34\xE6\x9C\x88\x00\x34\xE7\x82"
-     "\xB9\x00"
-     "\x35\x00\x35\x2C\x00\x35\x2E\x00\x35\x30\x00\x35\xE2\x81\x84\x36\x00\x35"
-     "\xE2\x81"
-     "\x84\x38\x00\x35\xE6\x97\xA5\x00\x35\xE6\x9C\x88\x00\x35\xE7\x82\xB9\x00"
-     "\x36\x00"
-     "\x36\x2C\x00\x36\x2E\x00\x36\xE6\x97\xA5\x00\x36\xE6\x9C\x88\x00\x36\xE7"
-     "\x82\xB9"
-     "\x00\x37\x00\x37\x2C\x00\x37\x2E\x00\x37\xE2\x81\x84\x38\x00\x37\xE6\x97"
-     "\xA5\x00"
-     "\x37\xE6\x9C\x88\x00\x37\xE7\x82\xB9\x00\x38\x00\x38\x2C\x00\x38\x2E\x00"
-     "\x38\xE6"
-     "\x97\xA5\x00\x38\xE6\x9C\x88\x00\x38\xE7\x82\xB9\x00\x39\x00\x39\x2C\x00"
-     "\x39\x2E"
-     "\x00\x39\xE6\x97\xA5\x00\x39\xE6\x9C\x88\x00\x39\xE7\x82\xB9\x00\x3A\x00"
-     "\x3A\x3A"
-     "\x3D\x00\x3B\x00\x3C\x00\x3D\x00\x3D\x3D\x00\x3D\x3D\x3D\x00\x3E\x00\x3F"
-     "\x00\x3F"
-     "\x21\x00\x3F\x3F\x00\x40\x00\x41\x00\x41\x55\x00\x41\xE2\x88\x95\x6D\x00"
-     "\x42\x00"
-     "\x42\x71\x00\x43\x00\x43\x44\x00\x43\x6F\x2E\x00\x43\xE2\x88\x95\x6B\x67"
-     "\x00\x44"
-     "\x00\x44\x4A\x00\x44\x5A\x00\x44\x7A\x00\x44\xC5\xBD\x00\x44\xC5\xBE\x00"
-     "\x45\x00"
-     "\x46\x00\x46\x41\x58\x00\x47\x00\x47\x42\x00\x47\x48\x7A\x00\x47\x50\x61"
-     "\x00\x47"
-     "\x79\x00\x48\x00\x48\x50\x00\x48\x56\x00\x48\x67\x00\x48\x7A\x00\x49\x00"
-     "\x49\x49"
-     "\x00\x49\x49\x49\x00\x49\x4A\x00\x49\x55\x00\x49\x56\x00\x49\x58\x00\x4A"
-     "\x00\x4B"
-     "\x00\x4B\x42\x00\x4B\x4B\x00\x4B\x4D\x00\x4C\x00\x4C\x4A\x00\x4C\x54\x44"
-     "\x00\x4C"
-     "\x6A\x00\x4C\xC2\xB7\x00\x4D\x00\x4D\x42\x00\x4D\x43\x00\x4D\x44\x00\x4D"
-     "\x48\x7A"
-     "\x00\x4D\x50\x61\x00\x4D\x52\x00\x4D\x56\x00\x4D\x57\x00\x4D\xCE\xA9\x00"
-     "\x4E\x00"
-     "\x4E\x4A\x00\x4E\x6A\x00\x4E\x6F\x00\x4F\x00\x50\x00\x50\x48\x00\x50\x50"
-     "\x4D\x00"
-     "\x50\x50\x56\x00\x50\x52\x00\x50\x54\x45\x00\x50\x61\x00\x51\x00\x52\x00"
-     "\x52\x73"
-     "\x00\x53\x00\x53\x44\x00\x53\x4D\x00\x53\x53\x00\x53\x76\x00\x54\x00\x54"
-     "\x45\x4C"
-     "\x00\x54\x48\x7A\x00\x54\x4D\x00\x55\x00\x56\x00\x56\x49\x00\x56\x49\x49"
-     "\x00\x56"
-     "\x49\x49\x49\x00\x56\xE2\x88\x95\x6D\x00\x57\x00\x57\x43\x00\x57\x5A\x00"
-     "\x57\x62"
-     "\x00\x58\x00\x58\x49\x00\x58\x49\x49\x00\x59\x00\x5A\x00\x5B\x00\x5C\x00"
-     "\x5D\x00"
-     "\x5E\x00\x5F\x00\x60\x00\x61\x00\x61\x2E\x6D\x2E\x00\x61\x2F\x63\x00\x61"
-     "\x2F\x73"
-     "\x00\x61\xCA\xBE\x00\x62\x00\x62\x61\x72\x00\x63\x00\x63\x2F\x6F\x00\x63"
-     "\x2F\x75"
-     "\x00\x63\x61\x6C\x00\x63\x63\x00\x63\x64\x00\x63\x6D\x00\x63\x6D\x32\x00"
-     "\x63\x6D"
-     "\x33\x00\x64\x00\x64\x42\x00\x64\x61\x00\x64\x6C\x00\x64\x6D\x00\x64\x6D"
-     "\x32\x00"
-     "\x64\x6D\x33\x00\x64\x7A\x00\x64\xC5\xBE\x00\x65\x00\x65\x56\x00\x65\x72"
-     "\x67\x00"
-     "\x66\x00\x66\x66\x00\x66\x66\x69\x00\x66\x66\x6C\x00\x66\x69\x00\x66\x6C"
-     "\x00\x66"
-     "\x6D\x00\x67\x00\x67\x61\x6C\x00\x68\x00\x68\x50\x61\x00\x68\x61\x00\x69"
-     "\x00\x69"
-     "\x69\x00\x69\x69\x69\x00\x69\x6A\x00\x69\x6E\x00\x69\x76\x00\x69\x78\x00"
-     "\x6A\x00"
-     "\x6B\x00\x6B\x41\x00\x6B\x48\x7A\x00\x6B\x50\x61\x00\x6B\x56\x00\x6B\x57"
-     "\x00\x6B"
-     "\x63\x61\x6C\x00\x6B\x67\x00\x6B\x6C\x00\x6B\x6D\x00\x6B\x6D\x32\x00\x6B"
-     "\x6D\x33"
-     "\x00\x6B\x74\x00\x6B\xCE\xA9\x00\x6C\x00\x6C\x6A\x00\x6C\x6D\x00\x6C\x6E"
-     "\x00\x6C"
-     "\x6F\x67\x00\x6C\x78\x00\x6C\xC2\xB7\x00\x6D\x00\x6D\x32\x00\x6D\x33\x00"
-     "\x6D\x41"
-     "\x00\x6D\x56\x00\x6D\x57\x00\x6D\x62\x00\x6D\x67\x00\x6D\x69\x6C\x00\x6D"
-     "\x6C\x00"
-     "\x6D\x6D\x00\x6D\x6D\x32\x00\x6D\x6D\x33\x00\x6D\x6F\x6C\x00\x6D\x73\x00"
-     "\x6D\xE2"
-     "\x88\x95\x73\x00\x6D\xE2\x88\x95\x73\x32\x00\x6E\x00\x6E\x41\x00\x6E\x46"
-     "\x00\x6E"
-     "\x56\x00\x6E\x57\x00\x6E\x6A\x00\x6E\x6D\x00\x6E\x73\x00\x6F\x00\x6F\x56"
-     "\x00\x70"
-     "\x00\x70\x2E\x6D\x2E\x00\x70\x41\x00\x70\x46\x00\x70\x56\x00\x70\x57\x00"
-     "\x70\x63"
-     "\x00\x70\x73\x00\x71\x00\x72\x00\x72\x61\x64\x00\x72\x61\x64\xE2\x88\x95"
-     "\x73\x00"
-     "\x72\x61\x64\xE2\x88\x95\x73\x32\x00\x73\x00\x73\x72\x00\x73\x74\x00\x74"
-     "\x00\x75"
-     "\x00\x76\x00\x76\x69\x00\x76\x69\x69\x00\x76\x69\x69\x69\x00\x77\x00\x78"
-     "\x00\x78"
-     "\x69\x00\x78\x69\x69\x00\x79\x00\x7A\x00\x7B\x00\x7C\x00\x7D\x00\xC2\xA2"
-     "\x00\xC2"
-     "\xA3\x00\xC2\xA5\x00\xC2\xA6\x00\xC2\xAC\x00\xC2\xB0\x43\x00\xC2\xB0\x46"
-     "\x00\xC2"
-     "\xB7\x00\xC3\x80\x00\xC3\x81\x00\xC3\x82\x00\xC3\x83\x00\xC3\x84\x00\xC3"
-     "\x85\x00"
-     "\xC3\x86\x00\xC3\x87\x00\xC3\x88\x00\xC3\x89\x00\xC3\x8A\x00\xC3\x8B\x00"
-     "\xC3\x8C"
-     "\x00\xC3\x8D\x00\xC3\x8E\x00\xC3\x8F\x00\xC3\x91\x00\xC3\x92\x00\xC3\x93"
-     "\x00\xC3"
-     "\x94\x00\xC3\x95\x00\xC3\x96\x00\xC3\x99\x00\xC3\x9A\x00\xC3\x9B\x00\xC3"
-     "\x9C\x00"
-     "\xC3\x9D\x00\xC3\xA0\x00\xC3\xA1\x00\xC3\xA2\x00\xC3\xA3\x00\xC3\xA4\x00"
-     "\xC3\xA5"
-     "\x00\xC3\xA6\x00\xC3\xA7\x00\xC3\xA8\x00\xC3\xA9\x00\xC3\xAA\x00\xC3\xAB"
-     "\x00\xC3"
-     "\xAC\x00\xC3\xAD\x00\xC3\xAE\x00\xC3\xAF\x00\xC3\xB0\x00\xC3\xB1\x00\xC3"
-     "\xB2\x00"
-     "\xC3\xB3\x00\xC3\xB4\x00\xC3\xB5\x00\xC3\xB6\x00\xC3\xB8\x00\xC3\xB9\x00"
-     "\xC3\xBA"
-     "\x00\xC3\xBB\x00\xC3\xBC\x00\xC3\xBD\x00\xC3\xBF\x00\xC4\x80\x00\xC4\x81"
-     "\x00\xC4"
-     "\x82\x00\xC4\x83\x00\xC4\x84\x00\xC4\x85\x00\xC4\x86\x00\xC4\x87\x00\xC4"
-     "\x88\x00"
-     "\xC4\x89\x00\xC4\x8A\x00\xC4\x8B\x00\xC4\x8C\x00\xC4\x8D\x00\xC4\x8E\x00"
-     "\xC4\x8F"
-     "\x00\xC4\x92\x00\xC4\x93\x00\xC4\x94\x00\xC4\x95\x00\xC4\x96\x00\xC4\x97"
-     "\x00\xC4"
-     "\x98\x00\xC4\x99\x00\xC4\x9A\x00\xC4\x9B\x00\xC4\x9C\x00\xC4\x9D\x00\xC4"
-     "\x9E\x00"
-     "\xC4\x9F\x00\xC4\xA0\x00\xC4\xA1\x00\xC4\xA2\x00\xC4\xA3\x00\xC4\xA4\x00"
-     "\xC4\xA5"
-     "\x00\xC4\xA6\x00\xC4\xA7\x00\xC4\xA8\x00\xC4\xA9\x00\xC4\xAA\x00\xC4\xAB"
-     "\x00\xC4"
-     "\xAC\x00\xC4\xAD\x00\xC4\xAE\x00\xC4\xAF\x00\xC4\xB0\x00\xC4\xB1\x00\xC4"
-     "\xB4\x00"
-     "\xC4\xB5\x00\xC4\xB6\x00\xC4\xB7\x00\xC4\xB9\x00\xC4\xBA\x00\xC4\xBB\x00"
-     "\xC4\xBC"
-     "\x00\xC4\xBD\x00\xC4\xBE\x00\xC5\x83\x00\xC5\x84\x00\xC5\x85\x00\xC5\x86"
-     "\x00\xC5"
-     "\x87\x00\xC5\x88\x00\xC5\x8B\x00\xC5\x8C\x00\xC5\x8D\x00\xC5\x8E\x00\xC5"
-     "\x8F\x00"
-     "\xC5\x90\x00\xC5\x91\x00\xC5\x93\x00\xC5\x94\x00\xC5\x95\x00\xC5\x96\x00"
-     "\xC5\x97"
-     "\x00\xC5\x98\x00\xC5\x99\x00\xC5\x9A\x00\xC5\x9B\x00\xC5\x9C\x00\xC5\x9D"
-     "\x00\xC5"
-     "\x9E\x00\xC5\x9F\x00\xC5\xA0\x00\xC5\xA1\x00\xC5\xA2\x00\xC5\xA3\x00\xC5"
-     "\xA4\x00"
-     "\xC5\xA5\x00\xC5\xA8\x00\xC5\xA9\x00\xC5\xAA\x00\xC5\xAB\x00\xC5\xAC\x00"
-     "\xC5\xAD"
-     "\x00\xC5\xAE\x00\xC5\xAF\x00\xC5\xB0\x00\xC5\xB1\x00\xC5\xB2\x00\xC5\xB3"
-     "\x00\xC5"
-     "\xB4\x00\xC5\xB5\x00\xC5\xB6\x00\xC5\xB7\x00\xC5\xB8\x00\xC5\xB9\x00\xC5"
-     "\xBA\x00"
-     "\xC5\xBB\x00\xC5\xBC\x00\xC5\xBD\x00\xC5\xBE\x00\xC6\x8E\x00\xC6\x90\x00"
-     "\xC6\xA0"
-     "\x00\xC6\xA1\x00\xC6\xAB\x00\xC6\xAF\x00\xC6\xB0\x00\xC7\x80\x00\xC7\x81"
-     "\x00\xC7"
-     "\x82\x00\xC7\x8D\x00\xC7\x8E\x00\xC7\x8F\x00\xC7\x90\x00\xC7\x91\x00\xC7"
-     "\x92\x00"
-     "\xC7\x93\x00\xC7\x94\x00\xC7\x95\x00\xC7\x96\x00\xC7\x97\x00\xC7\x98\x00"
-     "\xC7\x99"
-     "\x00\xC7\x9A\x00\xC7\x9B\x00\xC7\x9C\x00\xC7\x9E\x00\xC7\x9F\x00\xC7\xA0"
-     "\x00\xC7"
-     "\xA1\x00\xC7\xA2\x00\xC7\xA3\x00\xC7\xA6\x00\xC7\xA7\x00\xC7\xA8\x00\xC7"
-     "\xA9\x00"
-     "\xC7\xAA\x00\xC7\xAB\x00\xC7\xAC\x00\xC7\xAD\x00\xC7\xAE\x00\xC7\xAF\x00"
-     "\xC7\xB0"
-     "\x00\xC7\xB4\x00\xC7\xB5\x00\xC7\xB8\x00\xC7\xB9\x00\xC7\xBA\x00\xC7\xBB"
-     "\x00\xC7"
-     "\xBC\x00\xC7\xBD\x00\xC7\xBE\x00\xC7\xBF\x00\xC8\x80\x00\xC8\x81\x00\xC8"
-     "\x82\x00"
-     "\xC8\x83\x00\xC8\x84\x00\xC8\x85\x00\xC8\x86\x00\xC8\x87\x00\xC8\x88\x00"
-     "\xC8\x89"
-     "\x00\xC8\x8A\x00\xC8\x8B\x00\xC8\x8C\x00\xC8\x8D\x00\xC8\x8E\x00\xC8\x8F"
-     "\x00\xC8"
-     "\x90\x00\xC8\x91\x00\xC8\x92\x00\xC8\x93\x00\xC8\x94\x00\xC8\x95\x00\xC8"
-     "\x96\x00"
-     "\xC8\x97\x00\xC8\x98\x00\xC8\x99\x00\xC8\x9A\x00\xC8\x9B\x00\xC8\x9E\x00"
-     "\xC8\x9F"
-     "\x00\xC8\xA2\x00\xC8\xA6\x00\xC8\xA7\x00\xC8\xA8\x00\xC8\xA9\x00\xC8\xAA"
-     "\x00\xC8"
-     "\xAB\x00\xC8\xAC\x00\xC8\xAD\x00\xC8\xAE\x00\xC8\xAF\x00\xC8\xB0\x00\xC8"
-     "\xB1\x00"
-     "\xC8\xB2\x00\xC8\xB3\x00\xC8\xB7\x00\xC9\x90\x00\xC9\x91\x00\xC9\x92\x00"
-     "\xC9\x93"
-     "\x00\xC9\x94\x00\xC9\x95\x00\xC9\x96\x00\xC9\x97\x00\xC9\x98\x00\xC9\x99"
-     "\x00\xC9"
-     "\x9B\x00\xC9\x9C\x00\xC9\x9E\x00\xC9\x9F\x00\xC9\xA0\x00\xC9\xA1\x00\xC9"
-     "\xA2\x00"
-     "\xC9\xA3\x00\xC9\xA4\x00\xC9\xA5\x00\xC9\xA6\x00\xC9\xA7\x00\xC9\xA8\x00"
-     "\xC9\xA9"
-     "\x00\xC9\xAA\x00\xC9\xAB\x00\xC9\xAC\x00\xC9\xAD\x00\xC9\xAE\x00\xC9\xAF"
-     "\x00\xC9"
-     "\xB0\x00\xC9\xB1\x00\xC9\xB2\x00\xC9\xB3\x00\xC9\xB4\x00\xC9\xB5\x00\xC9"
-     "\xB6\x00"
-     "\xC9\xB7\x00\xC9\xB8\x00\xC9\xB9\x00\xC9\xBA\x00\xC9\xBB\x00\xC9\xBD\x00"
-     "\xC9\xBE"
-     "\x00\xCA\x80\x00\xCA\x81\x00\xCA\x82\x00\xCA\x83\x00\xCA\x84\x00\xCA\x88"
-     "\x00\xCA"
-     "\x89\x00\xCA\x8A\x00\xCA\x8B\x00\xCA\x8C\x00\xCA\x8D\x00\xCA\x8E\x00\xCA"
-     "\x8F\x00"
-     "\xCA\x90\x00\xCA\x91\x00\xCA\x92\x00\xCA\x95\x00\xCA\x98\x00\xCA\x99\x00"
-     "\xCA\x9B"
-     "\x00\xCA\x9C\x00\xCA\x9D\x00\xCA\x9F\x00\xCA\xA1\x00\xCA\xA2\x00\xCA\xA3"
-     "\x00\xCA"
-     "\xA4\x00\xCA\xA5\x00\xCA\xA6\x00\xCA\xA7\x00\xCA\xA8\x00\xCA\xA9\x00\xCA"
-     "\xAA\x00"
-     "\xCA\xAB\x00\xCA\xB9\x00\xCA\xBC\x6E\x00\xCB\x90\x00\xCB\x91\x00\xCC\x80"
-     "\x00\xCC"
-     "\x81\x00\xCC\x88\xCC\x81\x00\xCC\x93\x00\xCE\x86\x00\xCE\x88\x00\xCE\x89"
-     "\x00\xCE"
-     "\x8A\x00\xCE\x8C\x00\xCE\x8E\x00\xCE\x8F\x00\xCE\x90\x00\xCE\x91\x00\xCE"
-     "\x92\x00"
-     "\xCE\x93\x00\xCE\x94\x00\xCE\x95\x00\xCE\x96\x00\xCE\x97\x00\xCE\x98\x00"
-     "\xCE\x99"
-     "\x00\xCE\x9A\x00\xCE\x9B\x00\xCE\x9C\x00\xCE\x9D\x00\xCE\x9E\x00\xCE\x9F"
-     "\x00\xCE"
-     "\xA0\x00\xCE\xA1\x00\xCE\xA3\x00\xCE\xA4\x00\xCE\xA5\x00\xCE\xA6\x00\xCE"
-     "\xA7\x00"
-     "\xCE\xA8\x00\xCE\xA9\x00\xCE\xAA\x00\xCE\xAB\x00\xCE\xAC\x00\xCE\xAD\x00"
-     "\xCE\xAE"
-     "\x00\xCE\xAF\x00\xCE\xB0\x00\xCE\xB1\x00\xCE\xB2\x00\xCE\xB3\x00\xCE\xB4"
-     "\x00\xCE"
-     "\xB5\x00\xCE\xB6\x00\xCE\xB7\x00\xCE\xB8\x00\xCE\xB9\x00\xCE\xBA\x00\xCE"
-     "\xBB\x00"
-     "\xCE\xBC\x00\xCE\xBC\x41\x00\xCE\xBC\x46\x00\xCE\xBC\x56\x00\xCE\xBC\x57"
-     "\x00\xCE"
-     "\xBC\x67\x00\xCE\xBC\x6C\x00\xCE\xBC\x6D\x00\xCE\xBC\x73\x00\xCE\xBD\x00"
-     "\xCE\xBE"
-     "\x00\xCE\xBF\x00\xCF\x80\x00\xCF\x81\x00\xCF\x82\x00\xCF\x83\x00\xCF\x84"
-     "\x00\xCF"
-     "\x85\x00\xCF\x86\x00\xCF\x87\x00\xCF\x88\x00\xCF\x89\x00\xCF\x8A\x00\xCF"
-     "\x8B\x00"
-     "\xCF\x8C\x00\xCF\x8D\x00\xCF\x8E\x00\xCF\x9C\x00\xCF\x9D\x00\xD0\x80\x00"
-     "\xD0\x81"
-     "\x00\xD0\x83\x00\xD0\x87\x00\xD0\x8C\x00\xD0\x8D\x00\xD0\x8E\x00\xD0\x99"
-     "\x00\xD0"
-     "\xB0\x00\xD0\xB1\x00\xD0\xB2\x00\xD0\xB3\x00\xD0\xB4\x00\xD0\xB5\x00\xD0"
-     "\xB6\x00"
-     "\xD0\xB7\x00\xD0\xB8\x00\xD0\xB9\x00\xD0\xBA\x00\xD0\xBB\x00\xD0\xBC\x00"
-     "\xD0\xBD"
-     "\x00\xD0\xBE\x00\xD0\xBF\x00\xD1\x80\x00\xD1\x81\x00\xD1\x82\x00\xD1\x83"
-     "\x00\xD1"
-     "\x84\x00\xD1\x85\x00\xD1\x86\x00\xD1\x87\x00\xD1\x88\x00\xD1\x8A\x00\xD1"
-     "\x8B\x00"
-     "\xD1\x8C\x00\xD1\x8D\x00\xD1\x8E\x00\xD1\x90\x00\xD1\x91\x00\xD1\x93\x00"
-     "\xD1\x95"
-     "\x00\xD1\x96\x00\xD1\x97\x00\xD1\x98\x00\xD1\x9C\x00\xD1\x9D\x00\xD1\x9E"
-     "\x00\xD1"
-     "\x9F\x00\xD1\xB6\x00\xD1\xB7\x00\xD2\x91\x00\xD2\xAB\x00\xD2\xAF\x00\xD2"
-     "\xB1\x00"
-     "\xD3\x81\x00\xD3\x82\x00\xD3\x8F\x00\xD3\x90\x00\xD3\x91\x00\xD3\x92\x00"
-     "\xD3\x93"
-     "\x00\xD3\x96\x00\xD3\x97\x00\xD3\x99\x00\xD3\x9A\x00\xD3\x9B\x00\xD3\x9C"
-     "\x00\xD3"
-     "\x9D\x00\xD3\x9E\x00\xD3\x9F\x00\xD3\xA2\x00\xD3\xA3\x00\xD3\xA4\x00\xD3"
-     "\xA5\x00"
-     "\xD3\xA6\x00\xD3\xA7\x00\xD3\xA9\x00\xD3\xAA\x00\xD3\xAB\x00\xD3\xAC\x00"
-     "\xD3\xAD"
-     "\x00\xD3\xAE\x00\xD3\xAF\x00\xD3\xB0\x00\xD3\xB1\x00\xD3\xB2\x00\xD3\xB3"
-     "\x00\xD3"
-     "\xB4\x00\xD3\xB5\x00\xD3\xB8\x00\xD3\xB9\x00\xD5\xA5\xD6\x82\x00\xD5\xB4"
-     "\xD5\xA5"
-     "\x00\xD5\xB4\xD5\xAB\x00\xD5\xB4\xD5\xAD\x00\xD5\xB4\xD5\xB6\x00\xD5\xBE"
-     "\xD5\xB6"
-     "\x00\xD7\x90\x00\xD7\x90\xD6\xB7\x00\xD7\x90\xD6\xB8\x00\xD7\x90\xD6\xBC"
-     "\x00\xD7"
-     "\x90\xD7\x9C\x00\xD7\x91\x00\xD7\x91\xD6\xBC\x00\xD7\x91\xD6\xBF\x00\xD7"
-     "\x92\x00"
-     "\xD7\x92\xD6\xBC\x00\xD7\x93\x00\xD7\x93\xD6\xBC\x00\xD7\x94\x00\xD7\x94"
-     "\xD6\xBC"
-     "\x00\xD7\x95\xD6\xB9\x00\xD7\x95\xD6\xBC\x00\xD7\x96\xD6\xBC\x00\xD7\x98"
-     "\xD6\xBC"
-     "\x00\xD7\x99\xD6\xB4\x00\xD7\x99\xD6\xBC\x00\xD7\x9A\xD6\xBC\x00\xD7\x9B"
-     "\x00\xD7"
-     "\x9B\xD6\xBC\x00\xD7\x9B\xD6\xBF\x00\xD7\x9C\x00\xD7\x9C\xD6\xBC\x00\xD7"
-     "\x9D\x00"
-     "\xD7\x9E\xD6\xBC\x00\xD7\xA0\xD6\xBC\x00\xD7\xA1\xD6\xBC\x00\xD7\xA2\x00"
-     "\xD7\xA3"
-     "\xD6\xBC\x00\xD7\xA4\xD6\xBC\x00\xD7\xA4\xD6\xBF\x00\xD7\xA6\xD6\xBC\x00"
-     "\xD7\xA7"
-     "\xD6\xBC\x00\xD7\xA8\x00\xD7\xA8\xD6\xBC\x00\xD7\xA9\xD6\xBC\x00\xD7\xA9"
-     "\xD6\xBC"
-     "\xD7\x81\x00\xD7\xA9\xD6\xBC\xD7\x82\x00\xD7\xA9\xD7\x81\x00\xD7\xA9\xD7"
-     "\x82\x00"
-     "\xD7\xAA\x00\xD7\xAA\xD6\xBC\x00\xD7\xB2\xD6\xB7\x00\xD8\xA1\x00\xD8\xA2"
-     "\x00\xD8"
-     "\xA3\x00\xD8\xA4\x00\xD8\xA5\x00\xD8\xA6\x00\xD8\xA6\xD8\xA7\x00\xD8\xA6"
-     "\xD8\xAC"
-     "\x00\xD8\xA6\xD8\xAD\x00\xD8\xA6\xD8\xAE\x00\xD8\xA6\xD8\xB1\x00\xD8\xA6"
-     "\xD8\xB2"
-     "\x00\xD8\xA6\xD9\x85\x00\xD8\xA6\xD9\x86\x00\xD8\xA6\xD9\x87\x00\xD8\xA6"
-     "\xD9\x88"
-     "\x00\xD8\xA6\xD9\x89\x00\xD8\xA6\xD9\x8A\x00\xD8\xA6\xDB\x86\x00\xD8\xA6"
-     "\xDB\x87"
-     "\x00\xD8\xA6\xDB\x88\x00\xD8\xA6\xDB\x90\x00\xD8\xA6\xDB\x95\x00\xD8\xA7"
-     "\x00\xD8"
-     "\xA7\xD9\x83\xD8\xA8\xD8\xB1\x00\xD8\xA7\xD9\x84\xD9\x84\xD9\x87\x00\xD8"
-     "\xA7\xD9"
-     "\x8B\x00\xD8\xA7\xD9\xB4\x00\xD8\xA8\x00\xD8\xA8\xD8\xAC\x00\xD8\xA8\xD8"
-     "\xAD\x00"
-     "\xD8\xA8\xD8\xAD\xD9\x8A\x00\xD8\xA8\xD8\xAE\x00\xD8\xA8\xD8\xAE\xD9\x8A"
-     "\x00\xD8"
-     "\xA8\xD8\xB1\x00\xD8\xA8\xD8\xB2\x00\xD8\xA8\xD9\x85\x00\xD8\xA8\xD9\x86"
-     "\x00\xD8"
-     "\xA8\xD9\x87\x00\xD8\xA8\xD9\x89\x00\xD8\xA8\xD9\x8A\x00\xD8\xA9\x00\xD8"
-     "\xAA\x00"
-     "\xD8\xAA\xD8\xAC\x00\xD8\xAA\xD8\xAC\xD9\x85\x00\xD8\xAA\xD8\xAC\xD9\x89"
-     "\x00\xD8"
-     "\xAA\xD8\xAC\xD9\x8A\x00\xD8\xAA\xD8\xAD\x00\xD8\xAA\xD8\xAD\xD8\xAC\x00"
-     "\xD8\xAA"
-     "\xD8\xAD\xD9\x85\x00\xD8\xAA\xD8\xAE\x00\xD8\xAA\xD8\xAE\xD9\x85\x00\xD8"
-     "\xAA\xD8"
-     "\xAE\xD9\x89\x00\xD8\xAA\xD8\xAE\xD9\x8A\x00\xD8\xAA\xD8\xB1\x00\xD8\xAA"
-     "\xD8\xB2"
-     "\x00\xD8\xAA\xD9\x85\x00\xD8\xAA\xD9\x85\xD8\xAC\x00\xD8\xAA\xD9\x85\xD8"
-     "\xAD\x00"
-     "\xD8\xAA\xD9\x85\xD8\xAE\x00\xD8\xAA\xD9\x85\xD9\x89\x00\xD8\xAA\xD9\x85"
-     "\xD9\x8A"
-     "\x00\xD8\xAA\xD9\x86\x00\xD8\xAA\xD9\x87\x00\xD8\xAA\xD9\x89\x00\xD8\xAA"
-     "\xD9\x8A"
-     "\x00\xD8\xAB\x00\xD8\xAB\xD8\xAC\x00\xD8\xAB\xD8\xB1\x00\xD8\xAB\xD8\xB2"
-     "\x00\xD8"
-     "\xAB\xD9\x85\x00\xD8\xAB\xD9\x86\x00\xD8\xAB\xD9\x87\x00\xD8\xAB\xD9\x89"
-     "\x00\xD8"
-     "\xAB\xD9\x8A\x00\xD8\xAC\x00\xD8\xAC\xD8\xAD\x00\xD8\xAC\xD8\xAD\xD9\x89"
-     "\x00\xD8"
-     "\xAC\xD8\xAD\xD9\x8A\x00\xD8\xAC\xD9\x84\x20\xD8\xAC\xD9\x84\xD8\xA7\xD9"
-     "\x84\xD9"
-     "\x87\x00\xD8\xAC\xD9\x85\x00\xD8\xAC\xD9\x85\xD8\xAD\x00\xD8\xAC\xD9\x85"
-     "\xD9\x89"
-     "\x00\xD8\xAC\xD9\x85\xD9\x8A\x00\xD8\xAC\xD9\x89\x00\xD8\xAC\xD9\x8A\x00"
-     "\xD8\xAD"
-     "\x00\xD8\xAD\xD8\xAC\x00\xD8\xAD\xD8\xAC\xD9\x8A\x00\xD8\xAD\xD9\x85\x00"
-     "\xD8\xAD"
-     "\xD9\x85\xD9\x89\x00\xD8\xAD\xD9\x85\xD9\x8A\x00\xD8\xAD\xD9\x89\x00\xD8"
-     "\xAD\xD9"
-     "\x8A\x00\xD8\xAE\x00\xD8\xAE\xD8\xAC\x00\xD8\xAE\xD8\xAD\x00\xD8\xAE\xD9"
-     "\x85\x00"
-     "\xD8\xAE\xD9\x89\x00\xD8\xAE\xD9\x8A\x00\xD8\xAF\x00\xD8\xB0\x00\xD8\xB0"
-     "\xD9\xB0"
-     "\x00\xD8\xB1\x00\xD8\xB1\xD8\xB3\xD9\x88\xD9\x84\x00\xD8\xB1\xD9\xB0\x00"
-     "\xD8\xB1"
-     "\xDB\x8C\xD8\xA7\xD9\x84\x00\xD8\xB2\x00\xD8\xB3\x00\xD8\xB3\xD8\xAC\x00"
-     "\xD8\xB3"
-     "\xD8\xAC\xD8\xAD\x00\xD8\xB3\xD8\xAC\xD9\x89\x00\xD8\xB3\xD8\xAD\x00\xD8"
-     "\xB3\xD8"
-     "\xAD\xD8\xAC\x00\xD8\xB3\xD8\xAE\x00\xD8\xB3\xD8\xAE\xD9\x89\x00\xD8\xB3"
-     "\xD8\xAE"
-     "\xD9\x8A\x00\xD8\xB3\xD8\xB1\x00\xD8\xB3\xD9\x85\x00\xD8\xB3\xD9\x85\xD8"
-     "\xAC\x00"
-     "\xD8\xB3\xD9\x85\xD8\xAD\x00\xD8\xB3\xD9\x85\xD9\x85\x00\xD8\xB3\xD9\x87"
-     "\x00\xD8"
-     "\xB3\xD9\x89\x00\xD8\xB3\xD9\x8A\x00\xD8\xB4\x00\xD8\xB4\xD8\xAC\x00\xD8"
-     "\xB4\xD8"
-     "\xAC\xD9\x8A\x00\xD8\xB4\xD8\xAD\x00\xD8\xB4\xD8\xAD\xD9\x85\x00\xD8\xB4"
-     "\xD8\xAD"
-     "\xD9\x8A\x00\xD8\xB4\xD8\xAE\x00\xD8\xB4\xD8\xB1\x00\xD8\xB4\xD9\x85\x00"
-     "\xD8\xB4"
-     "\xD9\x85\xD8\xAE\x00\xD8\xB4\xD9\x85\xD9\x85\x00\xD8\xB4\xD9\x87\x00\xD8"
-     "\xB4\xD9"
-     "\x89\x00\xD8\xB4\xD9\x8A\x00\xD8\xB5\x00\xD8\xB5\xD8\xAD\x00\xD8\xB5\xD8"
-     "\xAD\xD8"
-     "\xAD\x00\xD8\xB5\xD8\xAD\xD9\x8A\x00\xD8\xB5\xD8\xAE\x00\xD8\xB5\xD8\xB1"
-     "\x00\xD8"
-     "\xB5\xD9\x84\xD8\xB9\xD9\x85\x00\xD8\xB5\xD9\x84\xD9\x89\x00\xD8\xB5\xD9"
-     "\x84\xD9"
-     "\x89\x20\xD8\xA7\xD9\x84\xD9\x84\xD9\x87\x20\xD8\xB9\xD9\x84\xD9\x8A\xD9"
-     "\x87\x20"
-     "\xD9\x88\xD8\xB3\xD9\x84\xD9\x85\x00\xD8\xB5\xD9\x84\xDB\x92\x00\xD8\xB5"
-     "\xD9\x85"
-     "\x00\xD8\xB5\xD9\x85\xD9\x85\x00\xD8\xB5\xD9\x89\x00\xD8\xB5\xD9\x8A\x00"
-     "\xD8\xB6"
-     "\x00\xD8\xB6\xD8\xAC\x00\xD8\xB6\xD8\xAD\x00\xD8\xB6\xD8\xAD\xD9\x89\x00"
-     "\xD8\xB6"
-     "\xD8\xAD\xD9\x8A\x00\xD8\xB6\xD8\xAE\x00\xD8\xB6\xD8\xAE\xD9\x85\x00\xD8"
-     "\xB6\xD8"
-     "\xB1\x00\xD8\xB6\xD9\x85\x00\xD8\xB6\xD9\x89\x00\xD8\xB6\xD9\x8A\x00\xD8"
-     "\xB7\x00"
-     "\xD8\xB7\xD8\xAD\x00\xD8\xB7\xD9\x85\x00\xD8\xB7\xD9\x85\xD8\xAD\x00\xD8"
-     "\xB7\xD9"
-     "\x85\xD9\x85\x00\xD8\xB7\xD9\x85\xD9\x8A\x00\xD8\xB7\xD9\x89\x00\xD8\xB7"
-     "\xD9\x8A"
-     "\x00\xD8\xB8\x00\xD8\xB8\xD9\x85\x00\xD8\xB9\x00\xD8\xB9\xD8\xAC\x00\xD8"
-     "\xB9\xD8"
-     "\xAC\xD9\x85\x00\xD8\xB9\xD9\x84\xD9\x8A\xD9\x87\x00\xD8\xB9\xD9\x85\x00"
-     "\xD8\xB9"
-     "\xD9\x85\xD9\x85\x00\xD8\xB9\xD9\x85\xD9\x89\x00\xD8\xB9\xD9\x85\xD9\x8A"
-     "\x00\xD8"
-     "\xB9\xD9\x89\x00\xD8\xB9\xD9\x8A\x00\xD8\xBA\x00\xD8\xBA\xD8\xAC\x00\xD8"
-     "\xBA\xD9"
-     "\x85\x00\xD8\xBA\xD9\x85\xD9\x85\x00\xD8\xBA\xD9\x85\xD9\x89\x00\xD8\xBA"
-     "\xD9\x85"
-     "\xD9\x8A\x00\xD8\xBA\xD9\x89\x00\xD8\xBA\xD9\x8A\x00\xD9\x80\xD9\x8B\x00"
-     "\xD9\x80"
-     "\xD9\x8E\x00\xD9\x80\xD9\x8E\xD9\x91\x00\xD9\x80\xD9\x8F\x00\xD9\x80\xD9"
-     "\x8F\xD9"
-     "\x91\x00\xD9\x80\xD9\x90\x00\xD9\x80\xD9\x90\xD9\x91\x00\xD9\x80\xD9\x91"
-     "\x00\xD9"
-     "\x80\xD9\x92\x00\xD9\x81\x00\xD9\x81\xD8\xAC\x00\xD9\x81\xD8\xAD\x00\xD9"
-     "\x81\xD8"
-     "\xAE\x00\xD9\x81\xD8\xAE\xD9\x85\x00\xD9\x81\xD9\x85\x00\xD9\x81\xD9\x85"
-     "\xD9\x8A"
-     "\x00\xD9\x81\xD9\x89\x00\xD9\x81\xD9\x8A\x00\xD9\x82\x00\xD9\x82\xD8\xAD"
-     "\x00\xD9"
-     "\x82\xD9\x84\xDB\x92\x00\xD9\x82\xD9\x85\x00\xD9\x82\xD9\x85\xD8\xAD\x00"
-     "\xD9\x82"
-     "\xD9\x85\xD9\x85\x00\xD9\x82\xD9\x85\xD9\x8A\x00\xD9\x82\xD9\x89\x00\xD9"
-     "\x82\xD9"
-     "\x8A\x00\xD9\x83\x00\xD9\x83\xD8\xA7\x00\xD9\x83\xD8\xAC\x00\xD9\x83\xD8"
-     "\xAD\x00"
-     "\xD9\x83\xD8\xAE\x00\xD9\x83\xD9\x84\x00\xD9\x83\xD9\x85\x00\xD9\x83\xD9"
-     "\x85\xD9"
-     "\x85\x00\xD9\x83\xD9\x85\xD9\x8A\x00\xD9\x83\xD9\x89\x00\xD9\x83\xD9\x8A"
-     "\x00\xD9"
-     "\x84\x00\xD9\x84\xD8\xA2\x00\xD9\x84\xD8\xA3\x00\xD9\x84\xD8\xA5\x00\xD9"
-     "\x84\xD8"
-     "\xA7\x00\xD9\x84\xD8\xAC\x00\xD9\x84\xD8\xAC\xD8\xAC\x00\xD9\x84\xD8\xAC"
-     "\xD9\x85"
-     "\x00\xD9\x84\xD8\xAC\xD9\x8A\x00\xD9\x84\xD8\xAD\x00\xD9\x84\xD8\xAD\xD9"
-     "\x85\x00"
-     "\xD9\x84\xD8\xAD\xD9\x89\x00\xD9\x84\xD8\xAD\xD9\x8A\x00\xD9\x84\xD8\xAE"
-     "\x00\xD9"
-     "\x84\xD8\xAE\xD9\x85\x00\xD9\x84\xD9\x85\x00\xD9\x84\xD9\x85\xD8\xAD\x00"
-     "\xD9\x84"
-     "\xD9\x85\xD9\x8A\x00\xD9\x84\xD9\x87\x00\xD9\x84\xD9\x89\x00\xD9\x84\xD9"
-     "\x8A\x00"
-     "\xD9\x85\x00\xD9\x85\xD8\xA7\x00\xD9\x85\xD8\xAC\x00\xD9\x85\xD8\xAC\xD8"
-     "\xAD\x00"
-     "\xD9\x85\xD8\xAC\xD8\xAE\x00\xD9\x85\xD8\xAC\xD9\x85\x00\xD9\x85\xD8\xAC"
-     "\xD9\x8A"
-     "\x00\xD9\x85\xD8\xAD\x00\xD9\x85\xD8\xAD\xD8\xAC\x00\xD9\x85\xD8\xAD\xD9"
-     "\x85\x00"
-     "\xD9\x85\xD8\xAD\xD9\x85\xD8\xAF\x00\xD9\x85\xD8\xAD\xD9\x8A\x00\xD9\x85"
-     "\xD8\xAE"
-     "\x00\xD9\x85\xD8\xAE\xD8\xAC\x00\xD9\x85\xD8\xAE\xD9\x85\x00\xD9\x85\xD8"
-     "\xAE\xD9"
-     "\x8A\x00\xD9\x85\xD9\x85\x00\xD9\x85\xD9\x85\xD9\x8A\x00\xD9\x85\xD9\x89"
-     "\x00\xD9"
-     "\x85\xD9\x8A\x00\xD9\x86\x00\xD9\x86\xD8\xAC\x00\xD9\x86\xD8\xAC\xD8\xAD"
-     "\x00\xD9"
-     "\x86\xD8\xAC\xD9\x85\x00\xD9\x86\xD8\xAC\xD9\x89\x00\xD9\x86\xD8\xAC\xD9"
-     "\x8A\x00"
-     "\xD9\x86\xD8\xAD\x00\xD9\x86\xD8\xAD\xD9\x85\x00\xD9\x86\xD8\xAD\xD9\x89"
-     "\x00\xD9"
-     "\x86\xD8\xAD\xD9\x8A\x00\xD9\x86\xD8\xAE\x00\xD9\x86\xD8\xB1\x00\xD9\x86"
-     "\xD8\xB2"
-     "\x00\xD9\x86\xD9\x85\x00\xD9\x86\xD9\x85\xD9\x89\x00\xD9\x86\xD9\x85\xD9"
-     "\x8A\x00"
-     "\xD9\x86\xD9\x86\x00\xD9\x86\xD9\x87\x00\xD9\x86\xD9\x89\x00\xD9\x86\xD9"
-     "\x8A\x00"
-     "\xD9\x87\x00\xD9\x87\xD8\xAC\x00\xD9\x87\xD9\x85\x00\xD9\x87\xD9\x85\xD8"
-     "\xAC\x00"
-     "\xD9\x87\xD9\x85\xD9\x85\x00\xD9\x87\xD9\x89\x00\xD9\x87\xD9\x8A\x00\xD9"
-     "\x87\xD9"
-     "\xB0\x00\xD9\x88\x00\xD9\x88\xD8\xB3\xD9\x84\xD9\x85\x00\xD9\x88\xD9\xB4"
-     "\x00\xD9"
-     "\x89\x00\xD9\x89\xD9\xB0\x00\xD9\x8A\x00\xD9\x8A\xD8\xAC\x00\xD9\x8A\xD8"
-     "\xAC\xD9"
-     "\x8A\x00\xD9\x8A\xD8\xAD\x00\xD9\x8A\xD8\xAD\xD9\x8A\x00\xD9\x8A\xD8\xAE"
-     "\x00\xD9"
-     "\x8A\xD8\xB1\x00\xD9\x8A\xD8\xB2\x00\xD9\x8A\xD9\x85\x00\xD9\x8A\xD9\x85"
-     "\xD9\x85"
-     "\x00\xD9\x8A\xD9\x85\xD9\x8A\x00\xD9\x8A\xD9\x86\x00\xD9\x8A\xD9\x87\x00"
-     "\xD9\x8A"
-     "\xD9\x89\x00\xD9\x8A\xD9\x8A\x00\xD9\x8A\xD9\xB4\x00\xD9\xAE\x00\xD9\xAF"
-     "\x00\xD9"
-     "\xB1\x00\xD9\xB9\x00\xD9\xBA\x00\xD9\xBB\x00\xD9\xBE\x00\xD9\xBF\x00\xDA"
-     "\x80\x00"
-     "\xDA\x83\x00\xDA\x84\x00\xDA\x86\x00\xDA\x87\x00\xDA\x88\x00\xDA\x8C\x00"
-     "\xDA\x8D"
-     "\x00\xDA\x8E\x00\xDA\x91\x00\xDA\x98\x00\xDA\xA1\x00\xDA\xA4\x00\xDA\xA6"
-     "\x00\xDA"
-     "\xA9\x00\xDA\xAD\x00\xDA\xAF\x00\xDA\xB1\x00\xDA\xB3\x00\xDA\xBA\x00\xDA"
-     "\xBB\x00"
-     "\xDA\xBE\x00\xDB\x80\x00\xDB\x81\x00\xDB\x82\x00\xDB\x85\x00\xDB\x86\x00"
-     "\xDB\x87"
-     "\x00\xDB\x87\xD9\xB4\x00\xDB\x88\x00\xDB\x89\x00\xDB\x8B\x00\xDB\x8C\x00"
-     "\xDB\x90"
-     "\x00\xDB\x92\x00\xDB\x93\x00\xE0\xA4\x95\xE0\xA4\xBC\x00\xE0\xA4\x96\xE0"
-     "\xA4\xBC"
-     "\x00\xE0\xA4\x97\xE0\xA4\xBC\x00\xE0\xA4\x9C\xE0\xA4\xBC\x00\xE0\xA4\xA1"
-     "\xE0\xA4"
-     "\xBC\x00\xE0\xA4\xA2\xE0\xA4\xBC\x00\xE0\xA4\xA9\x00\xE0\xA4\xAB\xE0\xA4"
-     "\xBC\x00"
-     "\xE0\xA4\xAF\xE0\xA4\xBC\x00\xE0\xA4\xB1\x00\xE0\xA4\xB4\x00\xE0\xA6\xA1"
-     "\xE0\xA6"
-     "\xBC\x00\xE0\xA6\xA2\xE0\xA6\xBC\x00\xE0\xA6\xAF\xE0\xA6\xBC\x00\xE0\xA7"
-     "\x8B\x00"
-     "\xE0\xA7\x8C\x00\xE0\xA8\x96\xE0\xA8\xBC\x00\xE0\xA8\x97\xE0\xA8\xBC\x00"
-     "\xE0\xA8"
-     "\x9C\xE0\xA8\xBC\x00\xE0\xA8\xAB\xE0\xA8\xBC\x00\xE0\xA8\xB2\xE0\xA8\xBC"
-     "\x00\xE0"
-     "\xA8\xB8\xE0\xA8\xBC\x00\xE0\xAC\xA1\xE0\xAC\xBC\x00\xE0\xAC\xA2\xE0\xAC"
-     "\xBC\x00"
-     "\xE0\xAD\x88\x00\xE0\xAD\x8B\x00\xE0\xAD\x8C\x00\xE0\xAE\x94\x00\xE0\xAF"
-     "\x8A\x00"
-     "\xE0\xAF\x8B\x00\xE0\xAF\x8C\x00\xE0\xB1\x88\x00\xE0\xB3\x80\x00\xE0\xB3"
-     "\x87\x00"
-     "\xE0\xB3\x88\x00\xE0\xB3\x8A\x00\xE0\xB3\x8B\x00\xE0\xB5\x8A\x00\xE0\xB5"
-     "\x8B\x00"
-     "\xE0\xB5\x8C\x00\xE0\xB7\x9A\x00\xE0\xB7\x9C\x00\xE0\xB7\x9D\x00\xE0\xB7"
-     "\x9E\x00"
-     "\xE0\xB9\x8D\xE0\xB8\xB2\x00\xE0\xBA\xAB\xE0\xBA\x99\x00\xE0\xBA\xAB\xE0"
-     "\xBA\xA1"
-     "\x00\xE0\xBB\x8D\xE0\xBA\xB2\x00\xE0\xBC\x8B\x00\xE0\xBD\x80\xE0\xBE\xB5"
-     "\x00\xE0"
-     "\xBD\x82\xE0\xBE\xB7\x00\xE0\xBD\x8C\xE0\xBE\xB7\x00\xE0\xBD\x91\xE0\xBE"
-     "\xB7\x00"
-     "\xE0\xBD\x96\xE0\xBE\xB7\x00\xE0\xBD\x9B\xE0\xBE\xB7\x00\xE0\xBD\xB1\xE0"
-     "\xBD\xB2"
-     "\x00\xE0\xBD\xB1\xE0\xBD\xB4\x00\xE0\xBD\xB1\xE0\xBE\x80\x00\xE0\xBE\x90"
-     "\xE0\xBE"
-     "\xB5\x00\xE0\xBE\x92\xE0\xBE\xB7\x00\xE0\xBE\x9C\xE0\xBE\xB7\x00\xE0\xBE"
-     "\xA1\xE0"
-     "\xBE\xB7\x00\xE0\xBE\xA6\xE0\xBE\xB7\x00\xE0\xBE\xAB\xE0\xBE\xB7\x00\xE0"
-     "\xBE\xB2"
-     "\xE0\xBD\xB1\xE0\xBE\x80\x00\xE0\xBE\xB2\xE0\xBE\x80\x00\xE0\xBE\xB3\xE0"
-     "\xBD\xB1"
-     "\xE0\xBE\x80\x00\xE0\xBE\xB3\xE0\xBE\x80\x00\xE1\x80\xA6\x00\xE1\x83\x9C"
-     "\x00\xE1"
-     "\x84\x80\x00\xE1\x84\x81\x00\xE1\x84\x82\x00\xE1\x84\x83\x00\xE1\x84\x84"
-     "\x00\xE1"
-     "\x84\x85\x00\xE1\x84\x86\x00\xE1\x84\x87\x00\xE1\x84\x88\x00\xE1\x84\x89"
-     "\x00\xE1"
-     "\x84\x8A\x00\xE1\x84\x8B\x00\xE1\x84\x8C\x00\xE1\x84\x8D\x00\xE1\x84\x8E"
-     "\x00\xE1"
-     "\x84\x8F\x00\xE1\x84\x90\x00\xE1\x84\x91\x00\xE1\x84\x92\x00\xE1\x84\x94"
-     "\x00\xE1"
-     "\x84\x95\x00\xE1\x84\x9A\x00\xE1\x84\x9C\x00\xE1\x84\x9D\x00\xE1\x84\x9E"
-     "\x00\xE1"
-     "\x84\xA0\x00\xE1\x84\xA1\x00\xE1\x84\xA2\x00\xE1\x84\xA3\x00\xE1\x84\xA7"
-     "\x00\xE1"
-     "\x84\xA9\x00\xE1\x84\xAB\x00\xE1\x84\xAC\x00\xE1\x84\xAD\x00\xE1\x84\xAE"
-     "\x00\xE1"
-     "\x84\xAF\x00\xE1\x84\xB2\x00\xE1\x84\xB6\x00\xE1\x85\x80\x00\xE1\x85\x87"
-     "\x00\xE1"
-     "\x85\x8C\x00\xE1\x85\x97\x00\xE1\x85\x98\x00\xE1\x85\x99\x00\xE1\x85\xA0"
-     "\x00\xE1"
-     "\x85\xA1\x00\xE1\x85\xA2\x00\xE1\x85\xA3\x00\xE1\x85\xA4\x00\xE1\x85\xA5"
-     "\x00\xE1"
-     "\x85\xA6\x00\xE1\x85\xA7\x00\xE1\x85\xA8\x00\xE1\x85\xA9\x00\xE1\x85\xAA"
-     "\x00\xE1"
-     "\x85\xAB\x00\xE1\x85\xAC\x00\xE1\x85\xAD\x00\xE1\x85\xAE\x00\xE1\x85\xAF"
-     "\x00\xE1"
-     "\x85\xB0\x00\xE1\x85\xB1\x00\xE1\x85\xB2\x00\xE1\x85\xB3\x00\xE1\x85\xB4"
-     "\x00\xE1"
-     "\x85\xB5\x00\xE1\x86\x84\x00\xE1\x86\x85\x00\xE1\x86\x88\x00\xE1\x86\x91"
-     "\x00\xE1"
-     "\x86\x92\x00\xE1\x86\x94\x00\xE1\x86\x9E\x00\xE1\x86\xA1\x00\xE1\x86\xAA"
-     "\x00\xE1"
-     "\x86\xAC\x00\xE1\x86\xAD\x00\xE1\x86\xB0\x00\xE1\x86\xB1\x00\xE1\x86\xB2"
-     "\x00\xE1"
-     "\x86\xB3\x00\xE1\x86\xB4\x00\xE1\x86\xB5\x00\xE1\x87\x87\x00\xE1\x87\x88"
-     "\x00\xE1"
-     "\x87\x8C\x00\xE1\x87\x8E\x00\xE1\x87\x93\x00\xE1\x87\x97\x00\xE1\x87\x99"
-     "\x00\xE1"
-     "\x87\x9D\x00\xE1\x87\x9F\x00\xE1\x87\xB1\x00\xE1\x87\xB2\x00\xE1\xAC\x86"
-     "\x00\xE1"
-     "\xAC\x88\x00\xE1\xAC\x8A\x00\xE1\xAC\x8C\x00\xE1\xAC\x8E\x00\xE1\xAC\x92"
-     "\x00\xE1"
-     "\xAC\xBB\x00\xE1\xAC\xBD\x00\xE1\xAD\x80\x00\xE1\xAD\x81\x00\xE1\xAD\x83"
-     "\x00\xE1"
-     "\xB4\x82\x00\xE1\xB4\x96\x00\xE1\xB4\x97\x00\xE1\xB4\x9C\x00\xE1\xB4\x9D"
-     "\x00\xE1"
-     "\xB4\xA5\x00\xE1\xB5\xBB\x00\xE1\xB6\x85\x00\xE1\xB6\x91\x00\xE1\xB8\x80"
-     "\x00\xE1"
-     "\xB8\x81\x00\xE1\xB8\x82\x00\xE1\xB8\x83\x00\xE1\xB8\x84\x00\xE1\xB8\x85"
-     "\x00\xE1"
-     "\xB8\x86\x00\xE1\xB8\x87\x00\xE1\xB8\x88\x00\xE1\xB8\x89\x00\xE1\xB8\x8A"
-     "\x00\xE1"
-     "\xB8\x8B\x00\xE1\xB8\x8C\x00\xE1\xB8\x8D\x00\xE1\xB8\x8E\x00\xE1\xB8\x8F"
-     "\x00\xE1"
-     "\xB8\x90\x00\xE1\xB8\x91\x00\xE1\xB8\x92\x00\xE1\xB8\x93\x00\xE1\xB8\x94"
-     "\x00\xE1"
-     "\xB8\x95\x00\xE1\xB8\x96\x00\xE1\xB8\x97\x00\xE1\xB8\x98\x00\xE1\xB8\x99"
-     "\x00\xE1"
-     "\xB8\x9A\x00\xE1\xB8\x9B\x00\xE1\xB8\x9C\x00\xE1\xB8\x9D\x00\xE1\xB8\x9E"
-     "\x00\xE1"
-     "\xB8\x9F\x00\xE1\xB8\xA0\x00\xE1\xB8\xA1\x00\xE1\xB8\xA2\x00\xE1\xB8\xA3"
-     "\x00\xE1"
-     "\xB8\xA4\x00\xE1\xB8\xA5\x00\xE1\xB8\xA6\x00\xE1\xB8\xA7\x00\xE1\xB8\xA8"
-     "\x00\xE1"
-     "\xB8\xA9\x00\xE1\xB8\xAA\x00\xE1\xB8\xAB\x00\xE1\xB8\xAC\x00\xE1\xB8\xAD"
-     "\x00\xE1"
-     "\xB8\xAE\x00\xE1\xB8\xAF\x00\xE1\xB8\xB0\x00\xE1\xB8\xB1\x00\xE1\xB8\xB2"
-     "\x00\xE1"
-     "\xB8\xB3\x00\xE1\xB8\xB4\x00\xE1\xB8\xB5\x00\xE1\xB8\xB6\x00\xE1\xB8\xB7"
-     "\x00\xE1"
-     "\xB8\xB8\x00\xE1\xB8\xB9\x00\xE1\xB8\xBA\x00\xE1\xB8\xBB\x00\xE1\xB8\xBC"
-     "\x00\xE1"
-     "\xB8\xBD\x00\xE1\xB8\xBE\x00\xE1\xB8\xBF\x00\xE1\xB9\x80\x00\xE1\xB9\x81"
-     "\x00\xE1"
-     "\xB9\x82\x00\xE1\xB9\x83\x00\xE1\xB9\x84\x00\xE1\xB9\x85\x00\xE1\xB9\x86"
-     "\x00\xE1"
-     "\xB9\x87\x00\xE1\xB9\x88\x00\xE1\xB9\x89\x00\xE1\xB9\x8A\x00\xE1\xB9\x8B"
-     "\x00\xE1"
-     "\xB9\x8C\x00\xE1\xB9\x8D\x00\xE1\xB9\x8E\x00\xE1\xB9\x8F\x00\xE1\xB9\x90"
-     "\x00\xE1"
-     "\xB9\x91\x00\xE1\xB9\x92\x00\xE1\xB9\x93\x00\xE1\xB9\x94\x00\xE1\xB9\x95"
-     "\x00\xE1"
-     "\xB9\x96\x00\xE1\xB9\x97\x00\xE1\xB9\x98\x00\xE1\xB9\x99\x00\xE1\xB9\x9A"
-     "\x00\xE1"
-     "\xB9\x9B\x00\xE1\xB9\x9C\x00\xE1\xB9\x9D\x00\xE1\xB9\x9E\x00\xE1\xB9\x9F"
-     "\x00\xE1"
-     "\xB9\xA0\x00\xE1\xB9\xA1\x00\xE1\xB9\xA2\x00\xE1\xB9\xA3\x00\xE1\xB9\xA4"
-     "\x00\xE1"
-     "\xB9\xA5\x00\xE1\xB9\xA6\x00\xE1\xB9\xA7\x00\xE1\xB9\xA8\x00\xE1\xB9\xA9"
-     "\x00\xE1"
-     "\xB9\xAA\x00\xE1\xB9\xAB\x00\xE1\xB9\xAC\x00\xE1\xB9\xAD\x00\xE1\xB9\xAE"
-     "\x00\xE1"
-     "\xB9\xAF\x00\xE1\xB9\xB0\x00\xE1\xB9\xB1\x00\xE1\xB9\xB2\x00\xE1\xB9\xB3"
-     "\x00\xE1"
-     "\xB9\xB4\x00\xE1\xB9\xB5\x00\xE1\xB9\xB6\x00\xE1\xB9\xB7\x00\xE1\xB9\xB8"
-     "\x00\xE1"
-     "\xB9\xB9\x00\xE1\xB9\xBA\x00\xE1\xB9\xBB\x00\xE1\xB9\xBC\x00\xE1\xB9\xBD"
-     "\x00\xE1"
-     "\xB9\xBE\x00\xE1\xB9\xBF\x00\xE1\xBA\x80\x00\xE1\xBA\x81\x00\xE1\xBA\x82"
-     "\x00\xE1"
-     "\xBA\x83\x00\xE1\xBA\x84\x00\xE1\xBA\x85\x00\xE1\xBA\x86\x00\xE1\xBA\x87"
-     "\x00\xE1"
-     "\xBA\x88\x00\xE1\xBA\x89\x00\xE1\xBA\x8A\x00\xE1\xBA\x8B\x00\xE1\xBA\x8C"
-     "\x00\xE1"
-     "\xBA\x8D\x00\xE1\xBA\x8E\x00\xE1\xBA\x8F\x00\xE1\xBA\x90\x00\xE1\xBA\x91"
-     "\x00\xE1"
-     "\xBA\x92\x00\xE1\xBA\x93\x00\xE1\xBA\x94\x00\xE1\xBA\x95\x00\xE1\xBA\x96"
-     "\x00\xE1"
-     "\xBA\x97\x00\xE1\xBA\x98\x00\xE1\xBA\x99\x00\xE1\xBA\xA0\x00\xE1\xBA\xA1"
-     "\x00\xE1"
-     "\xBA\xA2\x00\xE1\xBA\xA3\x00\xE1\xBA\xA4\x00\xE1\xBA\xA5\x00\xE1\xBA\xA6"
-     "\x00\xE1"
-     "\xBA\xA7\x00\xE1\xBA\xA8\x00\xE1\xBA\xA9\x00\xE1\xBA\xAA\x00\xE1\xBA\xAB"
-     "\x00\xE1"
-     "\xBA\xAC\x00\xE1\xBA\xAD\x00\xE1\xBA\xAE\x00\xE1\xBA\xAF\x00\xE1\xBA\xB0"
-     "\x00\xE1"
-     "\xBA\xB1\x00\xE1\xBA\xB2\x00\xE1\xBA\xB3\x00\xE1\xBA\xB4\x00\xE1\xBA\xB5"
-     "\x00\xE1"
-     "\xBA\xB6\x00\xE1\xBA\xB7\x00\xE1\xBA\xB8\x00\xE1\xBA\xB9\x00\xE1\xBA\xBA"
-     "\x00\xE1"
-     "\xBA\xBB\x00\xE1\xBA\xBC\x00\xE1\xBA\xBD\x00\xE1\xBA\xBE\x00\xE1\xBA\xBF"
-     "\x00\xE1"
-     "\xBB\x80\x00\xE1\xBB\x81\x00\xE1\xBB\x82\x00\xE1\xBB\x83\x00\xE1\xBB\x84"
-     "\x00\xE1"
-     "\xBB\x85\x00\xE1\xBB\x86\x00\xE1\xBB\x87\x00\xE1\xBB\x88\x00\xE1\xBB\x89"
-     "\x00\xE1"
-     "\xBB\x8A\x00\xE1\xBB\x8B\x00\xE1\xBB\x8C\x00\xE1\xBB\x8D\x00\xE1\xBB\x8E"
-     "\x00\xE1"
-     "\xBB\x8F\x00\xE1\xBB\x90\x00\xE1\xBB\x91\x00\xE1\xBB\x92\x00\xE1\xBB\x93"
-     "\x00\xE1"
-     "\xBB\x94\x00\xE1\xBB\x95\x00\xE1\xBB\x96\x00\xE1\xBB\x97\x00\xE1\xBB\x98"
-     "\x00\xE1"
-     "\xBB\x99\x00\xE1\xBB\x9A\x00\xE1\xBB\x9B\x00\xE1\xBB\x9C\x00\xE1\xBB\x9D"
-     "\x00\xE1"
-     "\xBB\x9E\x00\xE1\xBB\x9F\x00\xE1\xBB\xA0\x00\xE1\xBB\xA1\x00\xE1\xBB\xA2"
-     "\x00\xE1"
-     "\xBB\xA3\x00\xE1\xBB\xA4\x00\xE1\xBB\xA5\x00\xE1\xBB\xA6\x00\xE1\xBB\xA7"
-     "\x00\xE1"
-     "\xBB\xA8\x00\xE1\xBB\xA9\x00\xE1\xBB\xAA\x00\xE1\xBB\xAB\x00\xE1\xBB\xAC"
-     "\x00\xE1"
-     "\xBB\xAD\x00\xE1\xBB\xAE\x00\xE1\xBB\xAF\x00\xE1\xBB\xB0\x00\xE1\xBB\xB1"
-     "\x00\xE1"
-     "\xBB\xB2\x00\xE1\xBB\xB3\x00\xE1\xBB\xB4\x00\xE1\xBB\xB5\x00\xE1\xBB\xB6"
-     "\x00\xE1"
-     "\xBB\xB7\x00\xE1\xBB\xB8\x00\xE1\xBB\xB9\x00\xE1\xBC\x80\x00\xE1\xBC\x81"
-     "\x00\xE1"
-     "\xBC\x82\x00\xE1\xBC\x83\x00\xE1\xBC\x84\x00\xE1\xBC\x85\x00\xE1\xBC\x86"
-     "\x00\xE1"
-     "\xBC\x87\x00\xE1\xBC\x88\x00\xE1\xBC\x89\x00\xE1\xBC\x8A\x00\xE1\xBC\x8B"
-     "\x00\xE1"
-     "\xBC\x8C\x00\xE1\xBC\x8D\x00\xE1\xBC\x8E\x00\xE1\xBC\x8F\x00\xE1\xBC\x90"
-     "\x00\xE1"
-     "\xBC\x91\x00\xE1\xBC\x92\x00\xE1\xBC\x93\x00\xE1\xBC\x94\x00\xE1\xBC\x95"
-     "\x00\xE1"
-     "\xBC\x98\x00\xE1\xBC\x99\x00\xE1\xBC\x9A\x00\xE1\xBC\x9B\x00\xE1\xBC\x9C"
-     "\x00\xE1"
-     "\xBC\x9D\x00\xE1\xBC\xA0\x00\xE1\xBC\xA1\x00\xE1\xBC\xA2\x00\xE1\xBC\xA3"
-     "\x00\xE1"
-     "\xBC\xA4\x00\xE1\xBC\xA5\x00\xE1\xBC\xA6\x00\xE1\xBC\xA7\x00\xE1\xBC\xA8"
-     "\x00\xE1"
-     "\xBC\xA9\x00\xE1\xBC\xAA\x00\xE1\xBC\xAB\x00\xE1\xBC\xAC\x00\xE1\xBC\xAD"
-     "\x00\xE1"
-     "\xBC\xAE\x00\xE1\xBC\xAF\x00\xE1\xBC\xB0\x00\xE1\xBC\xB1\x00\xE1\xBC\xB2"
-     "\x00\xE1"
-     "\xBC\xB3\x00\xE1\xBC\xB4\x00\xE1\xBC\xB5\x00\xE1\xBC\xB6\x00\xE1\xBC\xB7"
-     "\x00\xE1"
-     "\xBC\xB8\x00\xE1\xBC\xB9\x00\xE1\xBC\xBA\x00\xE1\xBC\xBB\x00\xE1\xBC\xBC"
-     "\x00\xE1"
-     "\xBC\xBD\x00\xE1\xBC\xBE\x00\xE1\xBC\xBF\x00\xE1\xBD\x80\x00\xE1\xBD\x81"
-     "\x00\xE1"
-     "\xBD\x82\x00\xE1\xBD\x83\x00\xE1\xBD\x84\x00\xE1\xBD\x85\x00\xE1\xBD\x88"
-     "\x00\xE1"
-     "\xBD\x89\x00\xE1\xBD\x8A\x00\xE1\xBD\x8B\x00\xE1\xBD\x8C\x00\xE1\xBD\x8D"
-     "\x00\xE1"
-     "\xBD\x90\x00\xE1\xBD\x91\x00\xE1\xBD\x92\x00\xE1\xBD\x93\x00\xE1\xBD\x94"
-     "\x00\xE1"
-     "\xBD\x95\x00\xE1\xBD\x96\x00\xE1\xBD\x97\x00\xE1\xBD\x99\x00\xE1\xBD\x9B"
-     "\x00\xE1"
-     "\xBD\x9D\x00\xE1\xBD\x9F\x00\xE1\xBD\xA0\x00\xE1\xBD\xA1\x00\xE1\xBD\xA2"
-     "\x00\xE1"
-     "\xBD\xA3\x00\xE1\xBD\xA4\x00\xE1\xBD\xA5\x00\xE1\xBD\xA6\x00\xE1\xBD\xA7"
-     "\x00\xE1"
-     "\xBD\xA8\x00\xE1\xBD\xA9\x00\xE1\xBD\xAA\x00\xE1\xBD\xAB\x00\xE1\xBD\xAC"
-     "\x00\xE1"
-     "\xBD\xAD\x00\xE1\xBD\xAE\x00\xE1\xBD\xAF\x00\xE1\xBD\xB0\x00\xE1\xBD\xB2"
-     "\x00\xE1"
-     "\xBD\xB4\x00\xE1\xBD\xB6\x00\xE1\xBD\xB8\x00\xE1\xBD\xBA\x00\xE1\xBD\xBC"
-     "\x00\xE1"
-     "\xBE\x80\x00\xE1\xBE\x81\x00\xE1\xBE\x82\x00\xE1\xBE\x83\x00\xE1\xBE\x84"
-     "\x00\xE1"
-     "\xBE\x85\x00\xE1\xBE\x86\x00\xE1\xBE\x87\x00\xE1\xBE\x88\x00\xE1\xBE\x89"
-     "\x00\xE1"
-     "\xBE\x8A\x00\xE1\xBE\x8B\x00\xE1\xBE\x8C\x00\xE1\xBE\x8D\x00\xE1\xBE\x8E"
-     "\x00\xE1"
-     "\xBE\x8F\x00\xE1\xBE\x90\x00\xE1\xBE\x91\x00\xE1\xBE\x92\x00\xE1\xBE\x93"
-     "\x00\xE1"
-     "\xBE\x94\x00\xE1\xBE\x95\x00\xE1\xBE\x96\x00\xE1\xBE\x97\x00\xE1\xBE\x98"
-     "\x00\xE1"
-     "\xBE\x99\x00\xE1\xBE\x9A\x00\xE1\xBE\x9B\x00\xE1\xBE\x9C\x00\xE1\xBE\x9D"
-     "\x00\xE1"
-     "\xBE\x9E\x00\xE1\xBE\x9F\x00\xE1\xBE\xA0\x00\xE1\xBE\xA1\x00\xE1\xBE\xA2"
-     "\x00\xE1"
-     "\xBE\xA3\x00\xE1\xBE\xA4\x00\xE1\xBE\xA5\x00\xE1\xBE\xA6\x00\xE1\xBE\xA7"
-     "\x00\xE1"
-     "\xBE\xA8\x00\xE1\xBE\xA9\x00\xE1\xBE\xAA\x00\xE1\xBE\xAB\x00\xE1\xBE\xAC"
-     "\x00\xE1"
-     "\xBE\xAD\x00\xE1\xBE\xAE\x00\xE1\xBE\xAF\x00\xE1\xBE\xB0\x00\xE1\xBE\xB1"
-     "\x00\xE1"
-     "\xBE\xB2\x00\xE1\xBE\xB3\x00\xE1\xBE\xB4\x00\xE1\xBE\xB6\x00\xE1\xBE\xB7"
-     "\x00\xE1"
-     "\xBE\xB8\x00\xE1\xBE\xB9\x00\xE1\xBE\xBA\x00\xE1\xBE\xBC\x00\xE1\xBF\x82"
-     "\x00\xE1"
-     "\xBF\x83\x00\xE1\xBF\x84\x00\xE1\xBF\x86\x00\xE1\xBF\x87\x00\xE1\xBF\x88"
-     "\x00\xE1"
-     "\xBF\x8A\x00\xE1\xBF\x8C\x00\xE1\xBF\x90\x00\xE1\xBF\x91\x00\xE1\xBF\x92"
-     "\x00\xE1"
-     "\xBF\x96\x00\xE1\xBF\x97\x00\xE1\xBF\x98\x00\xE1\xBF\x99\x00\xE1\xBF\x9A"
-     "\x00\xE1"
-     "\xBF\xA0\x00\xE1\xBF\xA1\x00\xE1\xBF\xA2\x00\xE1\xBF\xA4\x00\xE1\xBF\xA5"
-     "\x00\xE1"
-     "\xBF\xA6\x00\xE1\xBF\xA7\x00\xE1\xBF\xA8\x00\xE1\xBF\xA9\x00\xE1\xBF\xAA"
-     "\x00\xE1"
-     "\xBF\xAC\x00\xE1\xBF\xB2\x00\xE1\xBF\xB3\x00\xE1\xBF\xB4\x00\xE1\xBF\xB6"
-     "\x00\xE1"
-     "\xBF\xB7\x00\xE1\xBF\xB8\x00\xE1\xBF\xBA\x00\xE1\xBF\xBC\x00\xE2\x80\x90"
-     "\x00\xE2"
-     "\x80\x93\x00\xE2\x80\x94\x00\xE2\x80\xB2\xE2\x80\xB2\x00\xE2\x80\xB2\xE2"
-     "\x80\xB2"
-     "\xE2\x80\xB2\x00\xE2\x80\xB2\xE2\x80\xB2\xE2\x80\xB2\xE2\x80\xB2\x00\xE2"
-     "\x80\xB5"
-     "\xE2\x80\xB5\x00\xE2\x80\xB5\xE2\x80\xB5\xE2\x80\xB5\x00\xE2\x82\xA9\x00"
-     "\xE2\x86"
-     "\x90\x00\xE2\x86\x91\x00\xE2\x86\x92\x00\xE2\x86\x93\x00\xE2\x86\x9A\x00"
-     "\xE2\x86"
-     "\x9B\x00\xE2\x86\xAE\x00\xE2\x87\x8D\x00\xE2\x87\x8E\x00\xE2\x87\x8F\x00"
-     "\xE2\x88"
-     "\x82\x00\xE2\x88\x84\x00\xE2\x88\x87\x00\xE2\x88\x89\x00\xE2\x88\x8C\x00"
-     "\xE2\x88"
-     "\x91\x00\xE2\x88\x92\x00\xE2\x88\xA4\x00\xE2\x88\xA6\x00\xE2\x88\xAB\xE2"
-     "\x88\xAB"
-     "\x00\xE2\x88\xAB\xE2\x88\xAB\xE2\x88\xAB\x00\xE2\x88\xAB\xE2\x88\xAB\xE2"
-     "\x88\xAB"
-     "\xE2\x88\xAB\x00\xE2\x88\xAE\xE2\x88\xAE\x00\xE2\x88\xAE\xE2\x88\xAE\xE2"
-     "\x88\xAE"
-     "\x00\xE2\x89\x81\x00\xE2\x89\x84\x00\xE2\x89\x87\x00\xE2\x89\x89\x00\xE2"
-     "\x89\xA0"
-     "\x00\xE2\x89\xA2\x00\xE2\x89\xAD\x00\xE2\x89\xAE\x00\xE2\x89\xAF\x00\xE2"
-     "\x89\xB0"
-     "\x00\xE2\x89\xB1\x00\xE2\x89\xB4\x00\xE2\x89\xB5\x00\xE2\x89\xB8\x00\xE2"
-     "\x89\xB9"
-     "\x00\xE2\x8A\x80\x00\xE2\x8A\x81\x00\xE2\x8A\x84\x00\xE2\x8A\x85\x00\xE2"
-     "\x8A\x88"
-     "\x00\xE2\x8A\x89\x00\xE2\x8A\xAC\x00\xE2\x8A\xAD\x00\xE2\x8A\xAE\x00\xE2"
-     "\x8A\xAF"
-     "\x00\xE2\x8B\xA0\x00\xE2\x8B\xA1\x00\xE2\x8B\xA2\x00\xE2\x8B\xA3\x00\xE2"
-     "\x8B\xAA"
-     "\x00\xE2\x8B\xAB\x00\xE2\x8B\xAC\x00\xE2\x8B\xAD\x00\xE2\x94\x82\x00\xE2"
-     "\x96\xA0"
-     "\x00\xE2\x97\x8B\x00\xE2\xA6\x85\x00\xE2\xA6\x86\x00\xE2\xAB\x9D\xCC\xB8"
-     "\x00\xE2"
-     "\xB1\xB1\x00\xE2\xB5\xA1\x00\xE3\x80\x81\x00\xE3\x80\x82\x00\xE3\x80\x88"
-     "\x00\xE3"
-     "\x80\x89\x00\xE3\x80\x8A\x00\xE3\x80\x8B\x00\xE3\x80\x8C\x00\xE3\x80\x8D"
-     "\x00\xE3"
-     "\x80\x8E\x00\xE3\x80\x8F\x00\xE3\x80\x90\x00\xE3\x80\x91\x00\xE3\x80\x92"
-     "\x00\xE3"
-     "\x80\x94\x00\xE3\x80\x94\x53\xE3\x80\x95\x00\xE3\x80\x94\xE4\xB8\x89\xE3"
-     "\x80\x95"
-     "\x00\xE3\x80\x94\xE4\xBA\x8C\xE3\x80\x95\x00\xE3\x80\x94\xE5\x8B\x9D\xE3"
-     "\x80\x95"
-     "\x00\xE3\x80\x94\xE5\xAE\x89\xE3\x80\x95\x00\xE3\x80\x94\xE6\x89\x93\xE3"
-     "\x80\x95"
-     "\x00\xE3\x80\x94\xE6\x95\x97\xE3\x80\x95\x00\xE3\x80\x94\xE6\x9C\xAC\xE3"
-     "\x80\x95"
-     "\x00\xE3\x80\x94\xE7\x82\xB9\xE3\x80\x95\x00\xE3\x80\x94\xE7\x9B\x97\xE3"
-     "\x80\x95"
-     "\x00\xE3\x80\x95\x00\xE3\x80\x96\x00\xE3\x80\x97\x00\xE3\x81\x8C\x00\xE3"
-     "\x81\x8E"
-     "\x00\xE3\x81\x90\x00\xE3\x81\x92\x00\xE3\x81\x94\x00\xE3\x81\x96\x00\xE3"
-     "\x81\x98"
-     "\x00\xE3\x81\x9A\x00\xE3\x81\x9C\x00\xE3\x81\x9E\x00\xE3\x81\xA0\x00\xE3"
-     "\x81\xA2"
-     "\x00\xE3\x81\xA5\x00\xE3\x81\xA7\x00\xE3\x81\xA9\x00\xE3\x81\xB0\x00\xE3"
-     "\x81\xB1"
-     "\x00\xE3\x81\xB3\x00\xE3\x81\xB4\x00\xE3\x81\xB6\x00\xE3\x81\xB7\x00\xE3"
-     "\x81\xB9"
-     "\x00\xE3\x81\xBA\x00\xE3\x81\xBB\xE3\x81\x8B\x00\xE3\x81\xBC\x00\xE3\x81"
-     "\xBD\x00"
-     "\xE3\x82\x88\xE3\x82\x8A\x00\xE3\x82\x94\x00\xE3\x82\x99\x00\xE3\x82\x9A"
-     "\x00\xE3"
-     "\x82\x9E\x00\xE3\x82\xA1\x00\xE3\x82\xA2\x00\xE3\x82\xA2\xE3\x83\x91\xE3"
-     "\x83\xBC"
-     "\xE3\x83\x88\x00\xE3\x82\xA2\xE3\x83\xAB\xE3\x83\x95\xE3\x82\xA1\x00\xE3"
-     "\x82\xA2"
-     "\xE3\x83\xB3\xE3\x83\x9A\xE3\x82\xA2\x00\xE3\x82\xA2\xE3\x83\xBC\xE3\x83"
-     "\xAB\x00"
-     "\xE3\x82\xA3\x00\xE3\x82\xA4\x00\xE3\x82\xA4\xE3\x83\x8B\xE3\x83\xB3\xE3"
-     "\x82\xB0"
-     "\x00\xE3\x82\xA4\xE3\x83\xB3\xE3\x83\x81\x00\xE3\x82\xA5\x00\xE3\x82\xA6"
-     "\x00\xE3"
-     "\x82\xA6\xE3\x82\xA9\xE3\x83\xB3\x00\xE3\x82\xA7\x00\xE3\x82\xA8\x00\xE3"
-     "\x82\xA8"
-     "\xE3\x82\xB9\xE3\x82\xAF\xE3\x83\xBC\xE3\x83\x89\x00\xE3\x82\xA8\xE3\x83"
-     "\xBC\xE3"
-     "\x82\xAB\xE3\x83\xBC\x00\xE3\x82\xA9\x00\xE3\x82\xAA\x00\xE3\x82\xAA\xE3"
-     "\x83\xB3"
-     "\xE3\x82\xB9\x00\xE3\x82\xAA\xE3\x83\xBC\xE3\x83\xA0\x00\xE3\x82\xAB\x00"
-     "\xE3\x82"
-     "\xAB\xE3\x82\xA4\xE3\x83\xAA\x00\xE3\x82\xAB\xE3\x83\xA9\xE3\x83\x83\xE3"
-     "\x83\x88"
-     "\x00\xE3\x82\xAB\xE3\x83\xAD\xE3\x83\xAA\xE3\x83\xBC\x00\xE3\x82\xAC\x00"
-     "\xE3\x82"
-     "\xAC\xE3\x83\xAD\xE3\x83\xB3\x00\xE3\x82\xAC\xE3\x83\xB3\xE3\x83\x9E\x00"
-     "\xE3\x82"
-     "\xAD\x00\xE3\x82\xAD\xE3\x83\xA5\xE3\x83\xAA\xE3\x83\xBC\x00\xE3\x82\xAD"
-     "\xE3\x83"
-     "\xAD\x00\xE3\x82\xAD\xE3\x83\xAD\xE3\x82\xB0\xE3\x83\xA9\xE3\x83\xA0\x00"
-     "\xE3\x82"
-     "\xAD\xE3\x83\xAD\xE3\x83\xA1\xE3\x83\xBC\xE3\x83\x88\xE3\x83\xAB\x00\xE3"
-     "\x82\xAD"
-     "\xE3\x83\xAD\xE3\x83\xAF\xE3\x83\x83\xE3\x83\x88\x00\xE3\x82\xAE\x00\xE3"
-     "\x82\xAE"
-     "\xE3\x82\xAC\x00\xE3\x82\xAE\xE3\x83\x8B\xE3\x83\xBC\x00\xE3\x82\xAE\xE3"
-     "\x83\xAB"
-     "\xE3\x83\x80\xE3\x83\xBC\x00\xE3\x82\xAF\x00\xE3\x82\xAF\xE3\x83\xAB\xE3"
-     "\x82\xBC"
-     "\xE3\x82\xA4\xE3\x83\xAD\x00\xE3\x82\xAF\xE3\x83\xAD\xE3\x83\xBC\xE3\x83"
-     "\x8D\x00"
-     "\xE3\x82\xB0\x00\xE3\x82\xB0\xE3\x83\xA9\xE3\x83\xA0\x00\xE3\x82\xB0\xE3"
-     "\x83\xA9"
-     "\xE3\x83\xA0\xE3\x83\x88\xE3\x83\xB3\x00\xE3\x82\xB1\x00\xE3\x82\xB1\xE3"
-     "\x83\xBC"
-     "\xE3\x82\xB9\x00\xE3\x82\xB2\x00\xE3\x82\xB3\x00\xE3\x82\xB3\xE3\x82\xB3"
-     "\x00\xE3"
-     "\x82\xB3\xE3\x83\x88\x00\xE3\x82\xB3\xE3\x83\xAB\xE3\x83\x8A\x00\xE3\x82"
-     "\xB3\xE3"
-     "\x83\xBC\xE3\x83\x9D\x00\xE3\x82\xB4\x00\xE3\x82\xB5\x00\xE3\x82\xB5\xE3"
-     "\x82\xA4"
-     "\xE3\x82\xAF\xE3\x83\xAB\x00\xE3\x82\xB5\xE3\x83\xB3\xE3\x83\x81\xE3\x83"
-     "\xBC\xE3"
-     "\x83\xA0\x00\xE3\x82\xB6\x00\xE3\x82\xB7\x00\xE3\x82\xB7\xE3\x83\xAA\xE3"
-     "\x83\xB3"
-     "\xE3\x82\xB0\x00\xE3\x82\xB8\x00\xE3\x82\xB9\x00\xE3\x82\xBA\x00\xE3\x82"
-     "\xBB\x00"
-     "\xE3\x82\xBB\xE3\x83\xB3\xE3\x83\x81\x00\xE3\x82\xBB\xE3\x83\xB3\xE3\x83"
-     "\x88\x00"
-     "\xE3\x82\xBC\x00\xE3\x82\xBD\x00\xE3\x82\xBE\x00\xE3\x82\xBF\x00\xE3\x83"
-     "\x80\x00"
-     "\xE3\x83\x80\xE3\x83\xBC\xE3\x82\xB9\x00\xE3\x83\x81\x00\xE3\x83\x82\x00"
-     "\xE3\x83"
-     "\x83\x00\xE3\x83\x84\x00\xE3\x83\x85\x00\xE3\x83\x86\x00\xE3\x83\x87\x00"
-     "\xE3\x83"
-     "\x87\xE3\x82\xB7\x00\xE3\x83\x88\x00\xE3\x83\x88\xE3\x83\xB3\x00\xE3\x83"
-     "\x89\x00"
-     "\xE3\x83\x89\xE3\x83\xAB\x00\xE3\x83\x8A\x00\xE3\x83\x8A\xE3\x83\x8E\x00"
-     "\xE3\x83"
-     "\x8B\x00\xE3\x83\x8C\x00\xE3\x83\x8D\x00\xE3\x83\x8E\x00\xE3\x83\x8E\xE3"
-     "\x83\x83"
-     "\xE3\x83\x88\x00\xE3\x83\x8F\x00\xE3\x83\x8F\xE3\x82\xA4\xE3\x83\x84\x00"
-     "\xE3\x83"
-     "\x90\x00\xE3\x83\x90\xE3\x83\xBC\xE3\x83\xAC\xE3\x83\xAB\x00\xE3\x83\x91"
-     "\x00\xE3"
-     "\x83\x91\xE3\x83\xBC\xE3\x82\xBB\xE3\x83\xB3\xE3\x83\x88\x00\xE3\x83\x91"
-     "\xE3\x83"
-     "\xBC\xE3\x83\x84\x00\xE3\x83\x92\x00\xE3\x83\x93\x00\xE3\x83\x93\xE3\x83"
-     "\xAB\x00"
-     "\xE3\x83\x94\x00\xE3\x83\x94\xE3\x82\xA2\xE3\x82\xB9\xE3\x83\x88\xE3\x83"
-     "\xAB\x00"
-     "\xE3\x83\x94\xE3\x82\xAF\xE3\x83\xAB\x00\xE3\x83\x94\xE3\x82\xB3\x00\xE3"
-     "\x83\x95"
-     "\x00\xE3\x83\x95\xE3\x82\xA1\xE3\x83\xA9\xE3\x83\x83\xE3\x83\x89\x00\xE3"
-     "\x83\x95"
-     "\xE3\x82\xA3\xE3\x83\xBC\xE3\x83\x88\x00\xE3\x83\x95\xE3\x83\xA9\xE3\x83"
-     "\xB3\x00"
-     "\xE3\x83\x96\x00\xE3\x83\x96\xE3\x83\x83\xE3\x82\xB7\xE3\x82\xA7\xE3\x83"
-     "\xAB\x00"
-     "\xE3\x83\x97\x00\xE3\x83\x98\x00\xE3\x83\x98\xE3\x82\xAF\xE3\x82\xBF\xE3"
-     "\x83\xBC"
-     "\xE3\x83\xAB\x00\xE3\x83\x98\xE3\x83\xAB\xE3\x83\x84\x00\xE3\x83\x99\x00"
-     "\xE3\x83"
-     "\x99\xE3\x83\xBC\xE3\x82\xBF\x00\xE3\x83\x9A\x00\xE3\x83\x9A\xE3\x82\xBD"
-     "\x00\xE3"
-     "\x83\x9A\xE3\x83\x8B\xE3\x83\x92\x00\xE3\x83\x9A\xE3\x83\xB3\xE3\x82\xB9"
-     "\x00\xE3"
-     "\x83\x9A\xE3\x83\xBC\xE3\x82\xB8\x00\xE3\x83\x9B\x00\xE3\x83\x9B\xE3\x83"
-     "\xB3\x00"
-     "\xE3\x83\x9B\xE3\x83\xBC\xE3\x83\xAB\x00\xE3\x83\x9B\xE3\x83\xBC\xE3\x83"
-     "\xB3\x00"
-     "\xE3\x83\x9C\x00\xE3\x83\x9C\xE3\x83\xAB\xE3\x83\x88\x00\xE3\x83\x9D\x00"
-     "\xE3\x83"
-     "\x9D\xE3\x82\xA4\xE3\x83\xB3\xE3\x83\x88\x00\xE3\x83\x9D\xE3\x83\xB3\xE3"
-     "\x83\x89"
-     "\x00\xE3\x83\x9E\x00\xE3\x83\x9E\xE3\x82\xA4\xE3\x82\xAF\xE3\x83\xAD\x00"
-     "\xE3\x83"
-     "\x9E\xE3\x82\xA4\xE3\x83\xAB\x00\xE3\x83\x9E\xE3\x83\x83\xE3\x83\x8F\x00"
-     "\xE3\x83"
-     "\x9E\xE3\x83\xAB\xE3\x82\xAF\x00\xE3\x83\x9E\xE3\x83\xB3\xE3\x82\xB7\xE3"
-     "\x83\xA7"
-     "\xE3\x83\xB3\x00\xE3\x83\x9F\x00\xE3\x83\x9F\xE3\x82\xAF\xE3\x83\xAD\xE3"
-     "\x83\xB3"
-     "\x00\xE3\x83\x9F\xE3\x83\xAA\x00\xE3\x83\x9F\xE3\x83\xAA\xE3\x83\x90\xE3"
-     "\x83\xBC"
-     "\xE3\x83\xAB\x00\xE3\x83\xA0\x00\xE3\x83\xA1\x00\xE3\x83\xA1\xE3\x82\xAC"
-     "\x00\xE3"
-     "\x83\xA1\xE3\x82\xAC\xE3\x83\x88\xE3\x83\xB3\x00\xE3\x83\xA1\xE3\x83\xBC"
-     "\xE3\x83"
-     "\x88\xE3\x83\xAB\x00\xE3\x83\xA2\x00\xE3\x83\xA3\x00\xE3\x83\xA4\x00\xE3"
-     "\x83\xA4"
-     "\xE3\x83\xBC\xE3\x83\x89\x00\xE3\x83\xA4\xE3\x83\xBC\xE3\x83\xAB\x00\xE3"
-     "\x83\xA5"
-     "\x00\xE3\x83\xA6\x00\xE3\x83\xA6\xE3\x82\xA2\xE3\x83\xB3\x00\xE3\x83\xA7"
-     "\x00\xE3"
-     "\x83\xA8\x00\xE3\x83\xA9\x00\xE3\x83\xAA\x00\xE3\x83\xAA\xE3\x83\x83\xE3"
-     "\x83\x88"
-     "\xE3\x83\xAB\x00\xE3\x83\xAA\xE3\x83\xA9\x00\xE3\x83\xAB\x00\xE3\x83\xAB"
-     "\xE3\x83"
-     "\x94\xE3\x83\xBC\x00\xE3\x83\xAB\xE3\x83\xBC\xE3\x83\x96\xE3\x83\xAB\x00"
-     "\xE3\x83"
-     "\xAC\x00\xE3\x83\xAC\xE3\x83\xA0\x00\xE3\x83\xAC\xE3\x83\xB3\xE3\x83\x88"
-     "\xE3\x82"
-     "\xB2\xE3\x83\xB3\x00\xE3\x83\xAD\x00\xE3\x83\xAF\x00\xE3\x83\xAF\xE3\x83"
-     "\x83\xE3"
-     "\x83\x88\x00\xE3\x83\xB0\x00\xE3\x83\xB1\x00\xE3\x83\xB2\x00\xE3\x83\xB3"
-     "\x00\xE3"
-     "\x83\xB4\x00\xE3\x83\xB7\x00\xE3\x83\xB8\x00\xE3\x83\xB9\x00\xE3\x83\xBA"
-     "\x00\xE3"
-     "\x83\xBB\x00\xE3\x83\xBC\x00\xE3\x83\xBE\x00\xE3\x92\x9E\x00\xE3\x92\xB9"
-     "\x00\xE3"
-     "\x92\xBB\x00\xE3\x93\x9F\x00\xE3\x94\x95\x00\xE3\x9B\xAE\x00\xE3\x9B\xBC"
-     "\x00\xE3"
-     "\x9E\x81\x00\xE3\xA0\xAF\x00\xE3\xA1\xA2\x00\xE3\xA1\xBC\x00\xE3\xA3\x87"
-     "\x00\xE3"
-     "\xA3\xA3\x00\xE3\xA4\x9C\x00\xE3\xA4\xBA\x00\xE3\xA8\xAE\x00\xE3\xA9\xAC"
-     "\x00\xE3"
-     "\xAB\xA4\x00\xE3\xAC\x88\x00\xE3\xAC\x99\x00\xE3\xAD\x89\x00\xE3\xAE\x9D"
-     "\x00\xE3"
-     "\xB0\x98\x00\xE3\xB1\x8E\x00\xE3\xB4\xB3\x00\xE3\xB6\x96\x00\xE3\xBA\xAC"
-     "\x00\xE3"
-     "\xBA\xB8\x00\xE3\xBC\x9B\x00\xE3\xBF\xBC\x00\xE4\x80\x88\x00\xE4\x80\x98"
-     "\x00\xE4"
-     "\x80\xB9\x00\xE4\x81\x86\x00\xE4\x82\x96\x00\xE4\x83\xA3\x00\xE4\x84\xAF"
-     "\x00\xE4"
-     "\x88\x82\x00\xE4\x88\xA7\x00\xE4\x8A\xA0\x00\xE4\x8C\x81\x00\xE4\x8C\xB4"
-     "\x00\xE4"
-     "\x8D\x99\x00\xE4\x8F\x95\x00\xE4\x8F\x99\x00\xE4\x90\x8B\x00\xE4\x91\xAB"
-     "\x00\xE4"
-     "\x94\xAB\x00\xE4\x95\x9D\x00\xE4\x95\xA1\x00\xE4\x95\xAB\x00\xE4\x97\x97"
-     "\x00\xE4"
-     "\x97\xB9\x00\xE4\x98\xB5\x00\xE4\x9A\xBE\x00\xE4\x9B\x87\x00\xE4\xA6\x95"
-     "\x00\xE4"
-     "\xA7\xA6\x00\xE4\xA9\xAE\x00\xE4\xA9\xB6\x00\xE4\xAA\xB2\x00\xE4\xAC\xB3"
-     "\x00\xE4"
-     "\xAF\x8E\x00\xE4\xB3\x8E\x00\xE4\xB3\xAD\x00\xE4\xB3\xB8\x00\xE4\xB5\x96"
-     "\x00\xE4"
-     "\xB8\x80\x00\xE4\xB8\x81\x00\xE4\xB8\x83\x00\xE4\xB8\x89\x00\xE4\xB8\x8A"
-     "\x00\xE4"
-     "\xB8\x8B\x00\xE4\xB8\x8D\x00\xE4\xB8\x99\x00\xE4\xB8\xA6\x00\xE4\xB8\xA8"
-     "\x00\xE4"
-     "\xB8\xAD\x00\xE4\xB8\xB2\x00\xE4\xB8\xB6\x00\xE4\xB8\xB8\x00\xE4\xB8\xB9"
-     "\x00\xE4"
-     "\xB8\xBD\x00\xE4\xB8\xBF\x00\xE4\xB9\x81\x00\xE4\xB9\x99\x00\xE4\xB9\x9D"
-     "\x00\xE4"
-     "\xBA\x82\x00\xE4\xBA\x85\x00\xE4\xBA\x86\x00\xE4\xBA\x8C\x00\xE4\xBA\x94"
-     "\x00\xE4"
-     "\xBA\xA0\x00\xE4\xBA\xA4\x00\xE4\xBA\xAE\x00\xE4\xBA\xBA\x00\xE4\xBB\x80"
-     "\x00\xE4"
-     "\xBB\x8C\x00\xE4\xBB\xA4\x00\xE4\xBB\xA4\xE5\x92\x8C\x00\xE4\xBC\x81\x00"
-     "\xE4\xBC"
-     "\x91\x00\xE4\xBD\xA0\x00\xE4\xBE\x80\x00\xE4\xBE\x86\x00\xE4\xBE\x8B\x00"
-     "\xE4\xBE"
-     "\xAE\x00\xE4\xBE\xBB\x00\xE4\xBE\xBF\x00\xE5\x80\x82\x00\xE5\x80\xAB\x00"
-     "\xE5\x81"
-     "\xBA\x00\xE5\x82\x99\x00\xE5\x83\x8F\x00\xE5\x83\x9A\x00\xE5\x83\xA7\x00"
-     "\xE5\x84"
-     "\xAA\x00\xE5\x84\xBF\x00\xE5\x85\x80\x00\xE5\x85\x85\x00\xE5\x85\x8D\x00"
-     "\xE5\x85"
-     "\x94\x00\xE5\x85\xA4\x00\xE5\x85\xA5\x00\xE5\x85\xA7\x00\xE5\x85\xA8\x00"
-     "\xE5\x85"
-     "\xA9\x00\xE5\x85\xAB\x00\xE5\x85\xAD\x00\xE5\x85\xB7\x00\xE5\x86\x80\x00"
-     "\xE5\x86"
-     "\x82\x00\xE5\x86\x8D\x00\xE5\x86\x92\x00\xE5\x86\x95\x00\xE5\x86\x96\x00"
-     "\xE5\x86"
-     "\x97\x00\xE5\x86\x99\x00\xE5\x86\xA4\x00\xE5\x86\xAB\x00\xE5\x86\xAC\x00"
-     "\xE5\x86"
-     "\xB5\x00\xE5\x86\xB7\x00\xE5\x87\x89\x00\xE5\x87\x8C\x00\xE5\x87\x9C\x00"
-     "\xE5\x87"
-     "\x9E\x00\xE5\x87\xA0\x00\xE5\x87\xB5\x00\xE5\x88\x80\x00\xE5\x88\x83\x00"
-     "\xE5\x88"
-     "\x87\x00\xE5\x88\x97\x00\xE5\x88\x9D\x00\xE5\x88\xA9\x00\xE5\x88\xBA\x00"
-     "\xE5\x88"
-     "\xBB\x00\xE5\x89\x86\x00\xE5\x89\x8D\x00\xE5\x89\xB2\x00\xE5\x89\xB7\x00"
-     "\xE5\x8A"
-     "\x89\x00\xE5\x8A\x9B\x00\xE5\x8A\xA3\x00\xE5\x8A\xB3\x00\xE5\x8A\xB4\x00"
-     "\xE5\x8B"
-     "\x87\x00\xE5\x8B\x89\x00\xE5\x8B\x92\x00\xE5\x8B\x9E\x00\xE5\x8B\xA4\x00"
-     "\xE5\x8B"
-     "\xB5\x00\xE5\x8B\xB9\x00\xE5\x8B\xBA\x00\xE5\x8C\x85\x00\xE5\x8C\x86\x00"
-     "\xE5\x8C"
-     "\x95\x00\xE5\x8C\x97\x00\xE5\x8C\x9A\x00\xE5\x8C\xB8\x00\xE5\x8C\xBB\x00"
-     "\xE5\x8C"
-     "\xBF\x00\xE5\x8D\x81\x00\xE5\x8D\x84\x00\xE5\x8D\x85\x00\xE5\x8D\x89\x00"
-     "\xE5\x8D"
-     "\x91\x00\xE5\x8D\x94\x00\xE5\x8D\x9A\x00\xE5\x8D\x9C\x00\xE5\x8D\xA9\x00"
-     "\xE5\x8D"
-     "\xB0\x00\xE5\x8D\xB3\x00\xE5\x8D\xB5\x00\xE5\x8D\xBD\x00\xE5\x8D\xBF\x00"
-     "\xE5\x8E"
-     "\x82\x00\xE5\x8E\xB6\x00\xE5\x8F\x83\x00\xE5\x8F\x88\x00\xE5\x8F\x8A\x00"
-     "\xE5\x8F"
-     "\x8C\x00\xE5\x8F\x9F\x00\xE5\x8F\xA3\x00\xE5\x8F\xA5\x00\xE5\x8F\xAB\x00"
-     "\xE5\x8F"
-     "\xAF\x00\xE5\x8F\xB1\x00\xE5\x8F\xB3\x00\xE5\x90\x86\x00\xE5\x90\x88\x00"
-     "\xE5\x90"
-     "\x8D\x00\xE5\x90\x8F\x00\xE5\x90\x9D\x00\xE5\x90\xB8\x00\xE5\x90\xB9\x00"
-     "\xE5\x91"
-     "\x82\x00\xE5\x91\x88\x00\xE5\x91\xA8\x00\xE5\x92\x9E\x00\xE5\x92\xA2\x00"
-     "\xE5\x92"
-     "\xBD\x00\xE5\x93\xB6\x00\xE5\x94\x90\x00\xE5\x95\x8F\x00\xE5\x95\x93\x00"
-     "\xE5\x95"
-     "\x95\x00\xE5\x95\xA3\x00\xE5\x96\x84\x00\xE5\x96\x87\x00\xE5\x96\x99\x00"
-     "\xE5\x96"
-     "\x9D\x00\xE5\x96\xAB\x00\xE5\x96\xB3\x00\xE5\x96\xB6\x00\xE5\x97\x80\x00"
-     "\xE5\x97"
-     "\x82\x00\xE5\x97\xA2\x00\xE5\x98\x86\x00\xE5\x99\x91\x00\xE5\x99\xA8\x00"
-     "\xE5\x99"
-     "\xB4\x00\xE5\x9B\x97\x00\xE5\x9B\x9B\x00\xE5\x9B\xB9\x00\xE5\x9C\x96\x00"
-     "\xE5\x9C"
-     "\x97\x00\xE5\x9C\x9F\x00\xE5\x9C\xB0\x00\xE5\x9E\x8B\x00\xE5\x9F\x8E\x00"
-     "\xE5\x9F"
-     "\xB4\x00\xE5\xA0\x8D\x00\xE5\xA0\xB1\x00\xE5\xA0\xB2\x00\xE5\xA1\x80\x00"
-     "\xE5\xA1"
-     "\x9A\x00\xE5\xA1\x9E\x00\xE5\xA2\xA8\x00\xE5\xA2\xAC\x00\xE5\xA2\xB3\x00"
-     "\xE5\xA3"
-     "\x98\x00\xE5\xA3\x9F\x00\xE5\xA3\xAB\x00\xE5\xA3\xAE\x00\xE5\xA3\xB0\x00"
-     "\xE5\xA3"
-     "\xB2\x00\xE5\xA3\xB7\x00\xE5\xA4\x82\x00\xE5\xA4\x86\x00\xE5\xA4\x8A\x00"
-     "\xE5\xA4"
-     "\x95\x00\xE5\xA4\x9A\x00\xE5\xA4\x9C\x00\xE5\xA4\xA2\x00\xE5\xA4\xA7\x00"
-     "\xE5\xA4"
-     "\xA7\xE6\xAD\xA3\x00\xE5\xA4\xA9\x00\xE5\xA5\x84\x00\xE5\xA5\x88\x00\xE5"
-     "\xA5\x91"
-     "\x00\xE5\xA5\x94\x00\xE5\xA5\xA2\x00\xE5\xA5\xB3\x00\xE5\xA7\x98\x00\xE5"
-     "\xA7\xAC"
-     "\x00\xE5\xA8\x9B\x00\xE5\xA8\xA7\x00\xE5\xA9\xA2\x00\xE5\xA9\xA6\x00\xE5"
-     "\xAA\xB5"
-     "\x00\xE5\xAC\x88\x00\xE5\xAC\xA8\x00\xE5\xAC\xBE\x00\xE5\xAD\x90\x00\xE5"
-     "\xAD\x97"
-     "\x00\xE5\xAD\xA6\x00\xE5\xAE\x80\x00\xE5\xAE\x85\x00\xE5\xAE\x97\x00\xE5"
-     "\xAF\x83"
-     "\x00\xE5\xAF\x98\x00\xE5\xAF\xA7\x00\xE5\xAF\xAE\x00\xE5\xAF\xB3\x00\xE5"
-     "\xAF\xB8"
-     "\x00\xE5\xAF\xBF\x00\xE5\xB0\x86\x00\xE5\xB0\x8F\x00\xE5\xB0\xA2\x00\xE5"
-     "\xB0\xB8"
-     "\x00\xE5\xB0\xBF\x00\xE5\xB1\xA0\x00\xE5\xB1\xA2\x00\xE5\xB1\xA4\x00\xE5"
-     "\xB1\xA5"
-     "\x00\xE5\xB1\xAE\x00\xE5\xB1\xB1\x00\xE5\xB2\x8D\x00\xE5\xB3\x80\x00\xE5"
-     "\xB4\x99"
-     "\x00\xE5\xB5\x83\x00\xE5\xB5\x90\x00\xE5\xB5\xAB\x00\xE5\xB5\xAE\x00\xE5"
-     "\xB5\xBC"
-     "\x00\xE5\xB6\xB2\x00\xE5\xB6\xBA\x00\xE5\xB7\x9B\x00\xE5\xB7\xA1\x00\xE5"
-     "\xB7\xA2"
-     "\x00\xE5\xB7\xA5\x00\xE5\xB7\xA6\x00\xE5\xB7\xB1\x00\xE5\xB7\xBD\x00\xE5"
-     "\xB7\xBE"
-     "\x00\xE5\xB8\xA8\x00\xE5\xB8\xBD\x00\xE5\xB9\xA9\x00\xE5\xB9\xB2\x00\xE5"
-     "\xB9\xB3"
-     "\xE6\x88\x90\x00\xE5\xB9\xB4\x00\xE5\xB9\xBA\x00\xE5\xB9\xBC\x00\xE5\xB9"
-     "\xBF\x00"
-     "\xE5\xBA\xA6\x00\xE5\xBA\xB0\x00\xE5\xBA\xB3\x00\xE5\xBA\xB6\x00\xE5\xBB"
-     "\x89\x00"
-     "\xE5\xBB\x8A\x00\xE5\xBB\x92\x00\xE5\xBB\x93\x00\xE5\xBB\x99\x00\xE5\xBB"
-     "\xAC\x00"
-     "\xE5\xBB\xB4\x00\xE5\xBB\xBE\x00\xE5\xBC\x84\x00\xE5\xBC\x8B\x00\xE5\xBC"
-     "\x93\x00"
-     "\xE5\xBC\xA2\x00\xE5\xBD\x90\x00\xE5\xBD\x93\x00\xE5\xBD\xA1\x00\xE5\xBD"
-     "\xA2\x00"
-     "\xE5\xBD\xA9\x00\xE5\xBD\xAB\x00\xE5\xBD\xB3\x00\xE5\xBE\x8B\x00\xE5\xBE"
-     "\x8C\x00"
-     "\xE5\xBE\x97\x00\xE5\xBE\x9A\x00\xE5\xBE\xA9\x00\xE5\xBE\xAD\x00\xE5\xBF"
-     "\x83\x00"
-     "\xE5\xBF\x8D\x00\xE5\xBF\x97\x00\xE5\xBF\xB5\x00\xE5\xBF\xB9\x00\xE6\x80"
-     "\x92\x00"
-     "\xE6\x80\x9C\x00\xE6\x81\xB5\x00\xE6\x82\x81\x00\xE6\x82\x94\x00\xE6\x83"
-     "\x87\x00"
-     "\xE6\x83\x98\x00\xE6\x83\xA1\x00\xE6\x84\x88\x00\xE6\x85\x84\x00\xE6\x85"
-     "\x88\x00"
-     "\xE6\x85\x8C\x00\xE6\x85\x8E\x00\xE6\x85\xA0\x00\xE6\x85\xA8\x00\xE6\x85"
-     "\xBA\x00"
-     "\xE6\x86\x8E\x00\xE6\x86\x90\x00\xE6\x86\xA4\x00\xE6\x86\xAF\x00\xE6\x86"
-     "\xB2\x00"
-     "\xE6\x87\x9E\x00\xE6\x87\xB2\x00\xE6\x87\xB6\x00\xE6\x88\x80\x00\xE6\x88"
-     "\x88\x00"
-     "\xE6\x88\x90\x00\xE6\x88\x9B\x00\xE6\x88\xAE\x00\xE6\x88\xB4\x00\xE6\x88"
-     "\xB6\x00"
-     "\xE6\x89\x8B\x00\xE6\x89\x93\x00\xE6\x89\x9D\x00\xE6\x8A\x95\x00\xE6\x8A"
-     "\xB1\x00"
-     "\xE6\x8B\x89\x00\xE6\x8B\x8F\x00\xE6\x8B\x93\x00\xE6\x8B\x94\x00\xE6\x8B"
-     "\xBC\x00"
-     "\xE6\x8B\xBE\x00\xE6\x8C\x87\x00\xE6\x8C\xBD\x00\xE6\x8D\x90\x00\xE6\x8D"
-     "\x95\x00"
-     "\xE6\x8D\xA8\x00\xE6\x8D\xBB\x00\xE6\x8E\x83\x00\xE6\x8E\xA0\x00\xE6\x8E"
-     "\xA9\x00"
-     "\xE6\x8F\x84\x00\xE6\x8F\x85\x00\xE6\x8F\xA4\x00\xE6\x90\x9C\x00\xE6\x90"
-     "\xA2\x00"
-     "\xE6\x91\x92\x00\xE6\x91\xA9\x00\xE6\x91\xB7\x00\xE6\x91\xBE\x00\xE6\x92"
-     "\x9A\x00"
-     "\xE6\x92\x9D\x00\xE6\x93\x84\x00\xE6\x94\xAF\x00\xE6\x94\xB4\x00\xE6\x95"
-     "\x8F\x00"
-     "\xE6\x95\x96\x00\xE6\x95\xAC\x00\xE6\x95\xB8\x00\xE6\x96\x87\x00\xE6\x96"
-     "\x97\x00"
-     "\xE6\x96\x99\x00\xE6\x96\xA4\x00\xE6\x96\xB0\x00\xE6\x96\xB9\x00\xE6\x97"
-     "\x85\x00"
-     "\xE6\x97\xA0\x00\xE6\x97\xA2\x00\xE6\x97\xA3\x00\xE6\x97\xA5\x00\xE6\x98"
-     "\x8E\xE6"
-     "\xB2\xBB\x00\xE6\x98\x93\x00\xE6\x98\xA0\x00\xE6\x98\xAD\xE5\x92\x8C\x00"
-     "\xE6\x99"
-     "\x89\x00\xE6\x99\xB4\x00\xE6\x9A\x88\x00\xE6\x9A\x91\x00\xE6\x9A\x9C\x00"
-     "\xE6\x9A"
-     "\xB4\x00\xE6\x9B\x86\x00\xE6\x9B\xB0\x00\xE6\x9B\xB4\x00\xE6\x9B\xB8\x00"
-     "\xE6\x9C"
-     "\x80\x00\xE6\x9C\x88\x00\xE6\x9C\x89\x00\xE6\x9C\x97\x00\xE6\x9C\x9B\x00"
-     "\xE6\x9C"
-     "\xA1\x00\xE6\x9C\xA8\x00\xE6\x9D\x8E\x00\xE6\x9D\x93\x00\xE6\x9D\x96\x00"
-     "\xE6\x9D"
-     "\x9E\x00\xE6\x9D\xBB\x00\xE6\x9E\x85\x00\xE6\x9E\x97\x00\xE6\x9F\xB3\x00"
-     "\xE6\x9F"
-     "\xBA\x00\xE6\xA0\x97\x00\xE6\xA0\x9F\x00\xE6\xA0\xAA\x00\xE6\xA0\xAA\xE5"
-     "\xBC\x8F"
-     "\xE4\xBC\x9A\xE7\xA4\xBE\x00\xE6\xA1\x92\x00\xE6\xA2\x81\x00\xE6\xA2\x85"
-     "\x00\xE6"
-     "\xA2\x8E\x00\xE6\xA2\xA8\x00\xE6\xA4\x94\x00\xE6\xA5\x82\x00\xE6\xA6\xA3"
-     "\x00\xE6"
-     "\xA7\xAA\x00\xE6\xA8\x82\x00\xE6\xA8\x93\x00\xE6\xAA\xA8\x00\xE6\xAB\x93"
-     "\x00\xE6"
-     "\xAB\x9B\x00\xE6\xAC\x84\x00\xE6\xAC\xA0\x00\xE6\xAC\xA1\x00\xE6\xAD\x94"
-     "\x00\xE6"
-     "\xAD\xA2\x00\xE6\xAD\xA3\x00\xE6\xAD\xB2\x00\xE6\xAD\xB7\x00\xE6\xAD\xB9"
-     "\x00\xE6"
-     "\xAE\x9F\x00\xE6\xAE\xAE\x00\xE6\xAE\xB3\x00\xE6\xAE\xBA\x00\xE6\xAE\xBB"
-     "\x00\xE6"
-     "\xAF\x8B\x00\xE6\xAF\x8D\x00\xE6\xAF\x94\x00\xE6\xAF\x9B\x00\xE6\xB0\x8F"
-     "\x00\xE6"
-     "\xB0\x94\x00\xE6\xB0\xB4\x00\xE6\xB1\x8E\x00\xE6\xB1\xA7\x00\xE6\xB2\x88"
-     "\x00\xE6"
-     "\xB2\xBF\x00\xE6\xB3\x8C\x00\xE6\xB3\x8D\x00\xE6\xB3\xA5\x00\xE6\xB3\xA8"
-     "\x00\xE6"
-     "\xB4\x96\x00\xE6\xB4\x9B\x00\xE6\xB4\x9E\x00\xE6\xB4\xB4\x00\xE6\xB4\xBE"
-     "\x00\xE6"
-     "\xB5\x81\x00\xE6\xB5\xA9\x00\xE6\xB5\xAA\x00\xE6\xB5\xB7\x00\xE6\xB5\xB8"
-     "\x00\xE6"
-     "\xB6\x85\x00\xE6\xB7\x8B\x00\xE6\xB7\x9A\x00\xE6\xB7\xAA\x00\xE6\xB7\xB9"
-     "\x00\xE6"
-     "\xB8\x9A\x00\xE6\xB8\xAF\x00\xE6\xB9\xAE\x00\xE6\xBA\x80\x00\xE6\xBA\x9C"
-     "\x00\xE6"
-     "\xBA\xBA\x00\xE6\xBB\x87\x00\xE6\xBB\x8B\x00\xE6\xBB\x91\x00\xE6\xBB\x9B"
-     "\x00\xE6"
-     "\xBC\x8F\x00\xE6\xBC\x94\x00\xE6\xBC\xA2\x00\xE6\xBC\xA3\x00\xE6\xBD\xAE"
-     "\x00\xE6"
-     "\xBF\x86\x00\xE6\xBF\xAB\x00\xE6\xBF\xBE\x00\xE7\x80\x9B\x00\xE7\x80\x9E"
-     "\x00\xE7"
-     "\x80\xB9\x00\xE7\x81\x8A\x00\xE7\x81\xAB\x00\xE7\x81\xB0\x00\xE7\x81\xB7"
-     "\x00\xE7"
-     "\x81\xBD\x00\xE7\x82\x99\x00\xE7\x82\xAD\x00\xE7\x83\x88\x00\xE7\x83\x99"
-     "\x00\xE7"
-     "\x84\xA1\x00\xE7\x85\x85\x00\xE7\x85\x89\x00\xE7\x85\xAE\x00\xE7\x86\x9C"
-     "\x00\xE7"
-     "\x87\x8E\x00\xE7\x87\x90\x00\xE7\x88\x90\x00\xE7\x88\x9B\x00\xE7\x88\xA8"
-     "\x00\xE7"
-     "\x88\xAA\x00\xE7\x88\xAB\x00\xE7\x88\xB5\x00\xE7\x88\xB6\x00\xE7\x88\xBB"
-     "\x00\xE7"
-     "\x88\xBF\x00\xE7\x89\x87\x00\xE7\x89\x90\x00\xE7\x89\x99\x00\xE7\x89\x9B"
-     "\x00\xE7"
-     "\x89\xA2\x00\xE7\x89\xB9\x00\xE7\x8A\x80\x00\xE7\x8A\x95\x00\xE7\x8A\xAC"
-     "\x00\xE7"
-     "\x8A\xAF\x00\xE7\x8B\x80\x00\xE7\x8B\xBC\x00\xE7\x8C\xAA\x00\xE7\x8D\xB5"
-     "\x00\xE7"
-     "\x8D\xBA\x00\xE7\x8E\x84\x00\xE7\x8E\x87\x00\xE7\x8E\x89\x00\xE7\x8E\x8B"
-     "\x00\xE7"
-     "\x8E\xA5\x00\xE7\x8E\xB2\x00\xE7\x8F\x9E\x00\xE7\x90\x86\x00\xE7\x90\x89"
-     "\x00\xE7"
-     "\x90\xA2\x00\xE7\x91\x87\x00\xE7\x91\x9C\x00\xE7\x91\xA9\x00\xE7\x91\xB1"
-     "\x00\xE7"
-     "\x92\x85\x00\xE7\x92\x89\x00\xE7\x92\x98\x00\xE7\x93\x8A\x00\xE7\x93\x9C"
-     "\x00\xE7"
-     "\x93\xA6\x00\xE7\x94\x86\x00\xE7\x94\x98\x00\xE7\x94\x9F\x00\xE7\x94\xA4"
-     "\x00\xE7"
-     "\x94\xA8\x00\xE7\x94\xB0\x00\xE7\x94\xB2\x00\xE7\x94\xB3\x00\xE7\x94\xB7"
-     "\x00\xE7"
-     "\x94\xBB\x00\xE7\x94\xBE\x00\xE7\x95\x99\x00\xE7\x95\xA5\x00\xE7\x95\xB0"
-     "\x00\xE7"
-     "\x96\x8B\x00\xE7\x96\x92\x00\xE7\x97\xA2\x00\xE7\x98\x90\x00\xE7\x98\x9D"
-     "\x00\xE7"
-     "\x98\x9F\x00\xE7\x99\x82\x00\xE7\x99\xA9\x00\xE7\x99\xB6\x00\xE7\x99\xBD"
-     "\x00\xE7"
-     "\x9A\xAE\x00\xE7\x9A\xBF\x00\xE7\x9B\x8A\x00\xE7\x9B\x9B\x00\xE7\x9B\xA3"
-     "\x00\xE7"
-     "\x9B\xA7\x00\xE7\x9B\xAE\x00\xE7\x9B\xB4\x00\xE7\x9C\x81\x00\xE7\x9C\x9E"
-     "\x00\xE7"
-     "\x9C\x9F\x00\xE7\x9D\x80\x00\xE7\x9D\x8A\x00\xE7\x9E\x8B\x00\xE7\x9E\xA7"
-     "\x00\xE7"
-     "\x9F\x9B\x00\xE7\x9F\xA2\x00\xE7\x9F\xB3\x00\xE7\xA1\x8E\x00\xE7\xA1\xAB"
-     "\x00\xE7"
-     "\xA2\x8C\x00\xE7\xA2\x91\x00\xE7\xA3\x8A\x00\xE7\xA3\x8C\x00\xE7\xA3\xBB"
-     "\x00\xE7"
-     "\xA4\xAA\x00\xE7\xA4\xBA\x00\xE7\xA4\xBC\x00\xE7\xA4\xBE\x00\xE7\xA5\x88"
-     "\x00\xE7"
-     "\xA5\x89\x00\xE7\xA5\x90\x00\xE7\xA5\x96\x00\xE7\xA5\x9D\x00\xE7\xA5\x9E"
-     "\x00\xE7"
-     "\xA5\xA5\x00\xE7\xA5\xBF\x00\xE7\xA6\x81\x00\xE7\xA6\x8D\x00\xE7\xA6\x8E"
-     "\x00\xE7"
-     "\xA6\x8F\x00\xE7\xA6\xAE\x00\xE7\xA6\xB8\x00\xE7\xA6\xBE\x00\xE7\xA7\x8A"
-     "\x00\xE7"
-     "\xA7\x98\x00\xE7\xA7\xAB\x00\xE7\xA8\x9C\x00\xE7\xA9\x80\x00\xE7\xA9\x8A"
-     "\x00\xE7"
-     "\xA9\x8F\x00\xE7\xA9\xB4\x00\xE7\xA9\xBA\x00\xE7\xAA\x81\x00\xE7\xAA\xB1"
-     "\x00\xE7"
-     "\xAB\x8B\x00\xE7\xAB\xAE\x00\xE7\xAB\xB9\x00\xE7\xAC\xA0\x00\xE7\xAE\x8F"
-     "\x00\xE7"
-     "\xAF\x80\x00\xE7\xAF\x86\x00\xE7\xAF\x89\x00\xE7\xB0\xBE\x00\xE7\xB1\xA0"
-     "\x00\xE7"
-     "\xB1\xB3\x00\xE7\xB1\xBB\x00\xE7\xB2\x92\x00\xE7\xB2\xBE\x00\xE7\xB3\x92"
-     "\x00\xE7"
-     "\xB3\x96\x00\xE7\xB3\xA3\x00\xE7\xB3\xA7\x00\xE7\xB3\xA8\x00\xE7\xB3\xB8"
-     "\x00\xE7"
-     "\xB4\x80\x00\xE7\xB4\x90\x00\xE7\xB4\xA2\x00\xE7\xB4\xAF\x00\xE7\xB5\x82"
-     "\x00\xE7"
-     "\xB5\x9B\x00\xE7\xB5\xA3\x00\xE7\xB6\xA0\x00\xE7\xB6\xBE\x00\xE7\xB7\x87"
-     "\x00\xE7"
-     "\xB7\xB4\x00\xE7\xB8\x82\x00\xE7\xB8\x89\x00\xE7\xB8\xB7\x00\xE7\xB9\x81"
-     "\x00\xE7"
-     "\xB9\x85\x00\xE7\xBC\xB6\x00\xE7\xBC\xBE\x00\xE7\xBD\x91\x00\xE7\xBD\xB2"
-     "\x00\xE7"
-     "\xBD\xB9\x00\xE7\xBD\xBA\x00\xE7\xBE\x85\x00\xE7\xBE\x8A\x00\xE7\xBE\x95"
-     "\x00\xE7"
-     "\xBE\x9A\x00\xE7\xBE\xBD\x00\xE7\xBF\xBA\x00\xE8\x80\x81\x00\xE8\x80\x85"
-     "\x00\xE8"
-     "\x80\x8C\x00\xE8\x80\x92\x00\xE8\x80\xB3\x00\xE8\x81\x86\x00\xE8\x81\xA0"
-     "\x00\xE8"
-     "\x81\xAF\x00\xE8\x81\xB0\x00\xE8\x81\xBE\x00\xE8\x81\xBF\x00\xE8\x82\x89"
-     "\x00\xE8"
-     "\x82\x8B\x00\xE8\x82\xAD\x00\xE8\x82\xB2\x00\xE8\x84\x83\x00\xE8\x84\xBE"
-     "\x00\xE8"
-     "\x87\x98\x00\xE8\x87\xA3\x00\xE8\x87\xA8\x00\xE8\x87\xAA\x00\xE8\x87\xAD"
-     "\x00\xE8"
-     "\x87\xB3\x00\xE8\x87\xBC\x00\xE8\x88\x81\x00\xE8\x88\x84\x00\xE8\x88\x8C"
-     "\x00\xE8"
-     "\x88\x98\x00\xE8\x88\x9B\x00\xE8\x88\x9F\x00\xE8\x89\xAE\x00\xE8\x89\xAF"
-     "\x00\xE8"
-     "\x89\xB2\x00\xE8\x89\xB8\x00\xE8\x89\xB9\x00\xE8\x8A\x8B\x00\xE8\x8A\x91"
-     "\x00\xE8"
-     "\x8A\x9D\x00\xE8\x8A\xB1\x00\xE8\x8A\xB3\x00\xE8\x8A\xBD\x00\xE8\x8B\xA5"
-     "\x00\xE8"
-     "\x8B\xA6\x00\xE8\x8C\x9D\x00\xE8\x8C\xA3\x00\xE8\x8C\xB6\x00\xE8\x8D\x92"
-     "\x00\xE8"
-     "\x8D\x93\x00\xE8\x8D\xA3\x00\xE8\x8E\xAD\x00\xE8\x8E\xBD\x00\xE8\x8F\x89"
-     "\x00\xE8"
-     "\x8F\x8A\x00\xE8\x8F\x8C\x00\xE8\x8F\x9C\x00\xE8\x8F\xA7\x00\xE8\x8F\xAF"
-     "\x00\xE8"
-     "\x8F\xB1\x00\xE8\x90\xBD\x00\xE8\x91\x89\x00\xE8\x91\x97\x00\xE8\x93\xAE"
-     "\x00\xE8"
-     "\x93\xB1\x00\xE8\x93\xB3\x00\xE8\x93\xBC\x00\xE8\x94\x96\x00\xE8\x95\xA4"
-     "\x00\xE8"
-     "\x97\x8D\x00\xE8\x97\xBA\x00\xE8\x98\x86\x00\xE8\x98\x92\x00\xE8\x98\xAD"
-     "\x00\xE8"
-     "\x98\xBF\x00\xE8\x99\x8D\x00\xE8\x99\x90\x00\xE8\x99\x9C\x00\xE8\x99\xA7"
-     "\x00\xE8"
-     "\x99\xA9\x00\xE8\x99\xAB\x00\xE8\x9A\x88\x00\xE8\x9A\xA9\x00\xE8\x9B\xA2"
-     "\x00\xE8"
-     "\x9C\x8E\x00\xE8\x9C\xA8\x00\xE8\x9D\xAB\x00\xE8\x9D\xB9\x00\xE8\x9E\x86"
-     "\x00\xE8"
-     "\x9E\xBA\x00\xE8\x9F\xA1\x00\xE8\xA0\x81\x00\xE8\xA0\x9F\x00\xE8\xA1\x80"
-     "\x00\xE8"
-     "\xA1\x8C\x00\xE8\xA1\xA0\x00\xE8\xA1\xA3\x00\xE8\xA3\x82\x00\xE8\xA3\x8F"
-     "\x00\xE8"
-     "\xA3\x97\x00\xE8\xA3\x9E\x00\xE8\xA3\xA1\x00\xE8\xA3\xB8\x00\xE8\xA3\xBA"
-     "\x00\xE8"
-     "\xA4\x90\x00\xE8\xA5\x81\x00\xE8\xA5\xA4\x00\xE8\xA5\xBE\x00\xE8\xA6\x86"
-     "\x00\xE8"
-     "\xA6\x8B\x00\xE8\xA6\x96\x00\xE8\xA7\x92\x00\xE8\xA7\xA3\x00\xE8\xA8\x80"
-     "\x00\xE8"
-     "\xAA\xA0\x00\xE8\xAA\xAA\x00\xE8\xAA\xBF\x00\xE8\xAB\x8B\x00\xE8\xAB\x92"
-     "\x00\xE8"
-     "\xAB\x96\x00\xE8\xAB\xAD\x00\xE8\xAB\xB8\x00\xE8\xAB\xBE\x00\xE8\xAC\x81"
-     "\x00\xE8"
-     "\xAC\xB9\x00\xE8\xAD\x98\x00\xE8\xAE\x80\x00\xE8\xAE\x8A\x00\xE8\xB0\xB7"
-     "\x00\xE8"
-     "\xB1\x86\x00\xE8\xB1\x88\x00\xE8\xB1\x95\x00\xE8\xB1\xB8\x00\xE8\xB2\x9D"
-     "\x00\xE8"
-     "\xB2\xA1\x00\xE8\xB2\xA9\x00\xE8\xB2\xAB\x00\xE8\xB3\x81\x00\xE8\xB3\x82"
-     "\x00\xE8"
-     "\xB3\x87\x00\xE8\xB3\x88\x00\xE8\xB3\x93\x00\xE8\xB4\x88\x00\xE8\xB4\x9B"
-     "\x00\xE8"
-     "\xB5\xA4\x00\xE8\xB5\xB0\x00\xE8\xB5\xB7\x00\xE8\xB6\xB3\x00\xE8\xB6\xBC"
-     "\x00\xE8"
-     "\xB7\x8B\x00\xE8\xB7\xAF\x00\xE8\xB7\xB0\x00\xE8\xBA\xAB\x00\xE8\xBB\x8A"
-     "\x00\xE8"
-     "\xBB\x94\x00\xE8\xBC\xA6\x00\xE8\xBC\xAA\x00\xE8\xBC\xB8\x00\xE8\xBC\xBB"
-     "\x00\xE8"
-     "\xBD\xA2\x00\xE8\xBE\x9B\x00\xE8\xBE\x9E\x00\xE8\xBE\xB0\x00\xE8\xBE\xB5"
-     "\x00\xE8"
-     "\xBE\xB6\x00\xE9\x80\xA3\x00\xE9\x80\xB8\x00\xE9\x81\x8A\x00\xE9\x81\xA9"
-     "\x00\xE9"
-     "\x81\xB2\x00\xE9\x81\xBC\x00\xE9\x82\x8F\x00\xE9\x82\x91\x00\xE9\x82\x94"
-     "\x00\xE9"
-     "\x83\x8E\x00\xE9\x83\x9E\x00\xE9\x83\xB1\x00\xE9\x83\xBD\x00\xE9\x84\x91"
-     "\x00\xE9"
-     "\x84\x9B\x00\xE9\x85\x89\x00\xE9\x85\x8D\x00\xE9\x85\xAA\x00\xE9\x86\x99"
-     "\x00\xE9"
-     "\x86\xB4\x00\xE9\x87\x86\x00\xE9\x87\x8C\x00\xE9\x87\x8F\x00\xE9\x87\x91"
-     "\x00\xE9"
-     "\x88\xB4\x00\xE9\x88\xB8\x00\xE9\x89\xB6\x00\xE9\x89\xBC\x00\xE9\x8B\x97"
-     "\x00\xE9"
-     "\x8B\x98\x00\xE9\x8C\x84\x00\xE9\x8D\x8A\x00\xE9\x8F\xB9\x00\xE9\x90\x95"
-     "\x00\xE9"
-     "\x95\xB7\x00\xE9\x96\x80\x00\xE9\x96\x8B\x00\xE9\x96\xAD\x00\xE9\x96\xB7"
-     "\x00\xE9"
-     "\x98\x9C\x00\xE9\x98\xAE\x00\xE9\x99\x8B\x00\xE9\x99\x8D\x00\xE9\x99\xB5"
-     "\x00\xE9"
-     "\x99\xB8\x00\xE9\x99\xBC\x00\xE9\x9A\x86\x00\xE9\x9A\xA3\x00\xE9\x9A\xB6"
-     "\x00\xE9"
-     "\x9A\xB7\x00\xE9\x9A\xB8\x00\xE9\x9A\xB9\x00\xE9\x9B\x83\x00\xE9\x9B\xA2"
-     "\x00\xE9"
-     "\x9B\xA3\x00\xE9\x9B\xA8\x00\xE9\x9B\xB6\x00\xE9\x9B\xB7\x00\xE9\x9C\xA3"
-     "\x00\xE9"
-     "\x9C\xB2\x00\xE9\x9D\x88\x00\xE9\x9D\x91\x00\xE9\x9D\x96\x00\xE9\x9D\x9E"
-     "\x00\xE9"
-     "\x9D\xA2\x00\xE9\x9D\xA9\x00\xE9\x9F\x8B\x00\xE9\x9F\x9B\x00\xE9\x9F\xA0"
-     "\x00\xE9"
-     "\x9F\xAD\x00\xE9\x9F\xB3\x00\xE9\x9F\xBF\x00\xE9\xA0\x81\x00\xE9\xA0\x85"
-     "\x00\xE9"
-     "\xA0\x8B\x00\xE9\xA0\x98\x00\xE9\xA0\xA9\x00\xE9\xA0\xBB\x00\xE9\xA1\x9E"
-     "\x00\xE9"
-     "\xA2\xA8\x00\xE9\xA3\x9B\x00\xE9\xA3\x9F\x00\xE9\xA3\xA2\x00\xE9\xA3\xAF"
-     "\x00\xE9"
-     "\xA3\xBC\x00\xE9\xA4\xA8\x00\xE9\xA4\xA9\x00\xE9\xA6\x96\x00\xE9\xA6\x99"
-     "\x00\xE9"
-     "\xA6\xA7\x00\xE9\xA6\xAC\x00\xE9\xA7\x82\x00\xE9\xA7\xB1\x00\xE9\xA7\xBE"
-     "\x00\xE9"
-     "\xA9\xAA\x00\xE9\xAA\xA8\x00\xE9\xAB\x98\x00\xE9\xAB\x9F\x00\xE9\xAC\x92"
-     "\x00\xE9"
-     "\xAC\xA5\x00\xE9\xAC\xAF\x00\xE9\xAC\xB2\x00\xE9\xAC\xBC\x00\xE9\xAD\x9A"
-     "\x00\xE9"
-     "\xAD\xAF\x00\xE9\xB1\x80\x00\xE9\xB1\x97\x00\xE9\xB3\xA5\x00\xE9\xB3\xBD"
-     "\x00\xE9"
-     "\xB5\xA7\x00\xE9\xB6\xB4\x00\xE9\xB7\xBA\x00\xE9\xB8\x9E\x00\xE9\xB9\xB5"
-     "\x00\xE9"
-     "\xB9\xBF\x00\xE9\xBA\x97\x00\xE9\xBA\x9F\x00\xE9\xBA\xA5\x00\xE9\xBA\xBB"
-     "\x00\xE9"
-     "\xBB\x83\x00\xE9\xBB\x8D\x00\xE9\xBB\x8E\x00\xE9\xBB\x91\x00\xE9\xBB\xB9"
-     "\x00\xE9"
-     "\xBB\xBD\x00\xE9\xBB\xBE\x00\xE9\xBC\x85\x00\xE9\xBC\x8E\x00\xE9\xBC\x8F"
-     "\x00\xE9"
-     "\xBC\x93\x00\xE9\xBC\x96\x00\xE9\xBC\xA0\x00\xE9\xBC\xBB\x00\xE9\xBD\x83"
-     "\x00\xE9"
-     "\xBD\x8A\x00\xE9\xBD\x92\x00\xE9\xBE\x8D\x00\xE9\xBE\x8E\x00\xE9\xBE\x9C"
-     "\x00\xE9"
-     "\xBE\x9F\x00\xE9\xBE\xA0\x00\xEA\x99\x91\x00\xEA\x9A\x89\x00\xEA\x9C\xA7"
-     "\x00\xEA"
-     "\x9D\xAF\x00\xEA\x9E\x8E\x00\xEA\xAC\xB7\x00\xEA\xAD\x92\x00\xEA\xAD\xA6"
-     "\x00\xEA"
-     "\xAD\xA7\x00\xEA\xB0\x80\x00\xEA\xB0\x81\x00\xEA\xB0\x82\x00\xEA\xB0\x83"
-     "\x00\xEA"
-     "\xB0\x84\x00\xEA\xB0\x85\x00\xEA\xB0\x86\x00\xEA\xB0\x87\x00\xEA\xB0\x88"
-     "\x00\xEA"
-     "\xB0\x89\x00\xEA\xB0\x8A\x00\xEA\xB0\x8B\x00\xEA\xB0\x8C\x00\xEA\xB0\x8D"
-     "\x00\xEA"
-     "\xB0\x8E\x00\xEA\xB0\x8F\x00\xEA\xB0\x90\x00\xEA\xB0\x91\x00\xEA\xB0\x92"
-     "\x00\xEA"
-     "\xB0\x93\x00\xEA\xB0\x94\x00\xEA\xB0\x95\x00\xEA\xB0\x96\x00\xEA\xB0\x97"
-     "\x00\xEA"
-     "\xB0\x98\x00\xEA\xB0\x99\x00\xEA\xB0\x9A\x00\xEA\xB0\x9B\x00\xEA\xB0\x9C"
-     "\x00\xEA"
-     "\xB0\x9D\x00\xEA\xB0\x9E\x00\xEA\xB0\x9F\x00\xEA\xB0\xA0\x00\xEA\xB0\xA1"
-     "\x00\xEA"
-     "\xB0\xA2\x00\xEA\xB0\xA3\x00\xEA\xB0\xA4\x00\xEA\xB0\xA5\x00\xEA\xB0\xA6"
-     "\x00\xEA"
-     "\xB0\xA7\x00\xEA\xB0\xA8\x00\xEA\xB0\xA9\x00\xEA\xB0\xAA\x00\xEA\xB0\xAB"
-     "\x00\xEA"
-     "\xB0\xAC\x00\xEA\xB0\xAD\x00\xEA\xB0\xAE\x00\xEA\xB0\xAF\x00\xEA\xB0\xB0"
-     "\x00\xEA"
-     "\xB0\xB1\x00\xEA\xB0\xB2\x00\xEA\xB0\xB3\x00\xEA\xB0\xB4\x00\xEA\xB0\xB5"
-     "\x00\xEA"
-     "\xB0\xB6\x00\xEA\xB0\xB7\x00\xEA\xB0\xB8\x00\xEA\xB0\xB9\x00\xEA\xB0\xBA"
-     "\x00\xEA"
-     "\xB0\xBB\x00\xEA\xB0\xBC\x00\xEA\xB0\xBD\x00\xEA\xB0\xBE\x00\xEA\xB0\xBF"
-     "\x00\xEA"
-     "\xB1\x80\x00\xEA\xB1\x81\x00\xEA\xB1\x82\x00\xEA\xB1\x83\x00\xEA\xB1\x84"
-     "\x00\xEA"
-     "\xB1\x85\x00\xEA\xB1\x86\x00\xEA\xB1\x87\x00\xEA\xB1\x88\x00\xEA\xB1\x89"
-     "\x00\xEA"
-     "\xB1\x8A\x00\xEA\xB1\x8B\x00\xEA\xB1\x8C\x00\xEA\xB1\x8D\x00\xEA\xB1\x8E"
-     "\x00\xEA"
-     "\xB1\x8F\x00\xEA\xB1\x90\x00\xEA\xB1\x91\x00\xEA\xB1\x92\x00\xEA\xB1\x93"
-     "\x00\xEA"
-     "\xB1\x94\x00\xEA\xB1\x95\x00\xEA\xB1\x96\x00\xEA\xB1\x97\x00\xEA\xB1\x98"
-     "\x00\xEA"
-     "\xB1\x99\x00\xEA\xB1\x9A\x00\xEA\xB1\x9B\x00\xEA\xB1\x9C\x00\xEA\xB1\x9D"
-     "\x00\xEA"
-     "\xB1\x9E\x00\xEA\xB1\x9F\x00\xEA\xB1\xA0\x00\xEA\xB1\xA1\x00\xEA\xB1\xA2"
-     "\x00\xEA"
-     "\xB1\xA3\x00\xEA\xB1\xA4\x00\xEA\xB1\xA5\x00\xEA\xB1\xA6\x00\xEA\xB1\xA7"
-     "\x00\xEA"
-     "\xB1\xA8\x00\xEA\xB1\xA9\x00\xEA\xB1\xAA\x00\xEA\xB1\xAB\x00\xEA\xB1\xAC"
-     "\x00\xEA"
-     "\xB1\xAD\x00\xEA\xB1\xAE\x00\xEA\xB1\xAF\x00\xEA\xB1\xB0\x00\xEA\xB1\xB1"
-     "\x00\xEA"
-     "\xB1\xB2\x00\xEA\xB1\xB3\x00\xEA\xB1\xB4\x00\xEA\xB1\xB5\x00\xEA\xB1\xB6"
-     "\x00\xEA"
-     "\xB1\xB7\x00\xEA\xB1\xB8\x00\xEA\xB1\xB9\x00\xEA\xB1\xBA\x00\xEA\xB1\xBB"
-     "\x00\xEA"
-     "\xB1\xBC\x00\xEA\xB1\xBD\x00\xEA\xB1\xBE\x00\xEA\xB1\xBF\x00\xEA\xB2\x80"
-     "\x00\xEA"
-     "\xB2\x81\x00\xEA\xB2\x82\x00\xEA\xB2\x83\x00\xEA\xB2\x84\x00\xEA\xB2\x85"
-     "\x00\xEA"
-     "\xB2\x86\x00\xEA\xB2\x87\x00\xEA\xB2\x88\x00\xEA\xB2\x89\x00\xEA\xB2\x8A"
-     "\x00\xEA"
-     "\xB2\x8B\x00\xEA\xB2\x8C\x00\xEA\xB2\x8D\x00\xEA\xB2\x8E\x00\xEA\xB2\x8F"
-     "\x00\xEA"
-     "\xB2\x90\x00\xEA\xB2\x91\x00\xEA\xB2\x92\x00\xEA\xB2\x93\x00\xEA\xB2\x94"
-     "\x00\xEA"
-     "\xB2\x95\x00\xEA\xB2\x96\x00\xEA\xB2\x97\x00\xEA\xB2\x98\x00\xEA\xB2\x99"
-     "\x00\xEA"
-     "\xB2\x9A\x00\xEA\xB2\x9B\x00\xEA\xB2\x9C\x00\xEA\xB2\x9D\x00\xEA\xB2\x9E"
-     "\x00\xEA"
-     "\xB2\x9F\x00\xEA\xB2\xA0\x00\xEA\xB2\xA1\x00\xEA\xB2\xA2\x00\xEA\xB2\xA3"
-     "\x00\xEA"
-     "\xB2\xA4\x00\xEA\xB2\xA5\x00\xEA\xB2\xA6\x00\xEA\xB2\xA7\x00\xEA\xB2\xA8"
-     "\x00\xEA"
-     "\xB2\xA9\x00\xEA\xB2\xAA\x00\xEA\xB2\xAB\x00\xEA\xB2\xAC\x00\xEA\xB2\xAD"
-     "\x00\xEA"
-     "\xB2\xAE\x00\xEA\xB2\xAF\x00\xEA\xB2\xB0\x00\xEA\xB2\xB1\x00\xEA\xB2\xB2"
-     "\x00\xEA"
-     "\xB2\xB3\x00\xEA\xB2\xB4\x00\xEA\xB2\xB5\x00\xEA\xB2\xB6\x00\xEA\xB2\xB7"
-     "\x00\xEA"
-     "\xB2\xB8\x00\xEA\xB2\xB9\x00\xEA\xB2\xBA\x00\xEA\xB2\xBB\x00\xEA\xB2\xBC"
-     "\x00\xEA"
-     "\xB2\xBD\x00\xEA\xB2\xBE\x00\xEA\xB2\xBF\x00\xEA\xB3\x80\x00\xEA\xB3\x81"
-     "\x00\xEA"
-     "\xB3\x82\x00\xEA\xB3\x83\x00\xEA\xB3\x84\x00\xEA\xB3\x85\x00\xEA\xB3\x86"
-     "\x00\xEA"
-     "\xB3\x87\x00\xEA\xB3\x88\x00\xEA\xB3\x89\x00\xEA\xB3\x8A\x00\xEA\xB3\x8B"
-     "\x00\xEA"
-     "\xB3\x8C\x00\xEA\xB3\x8D\x00\xEA\xB3\x8E\x00\xEA\xB3\x8F\x00\xEA\xB3\x90"
-     "\x00\xEA"
-     "\xB3\x91\x00\xEA\xB3\x92\x00\xEA\xB3\x93\x00\xEA\xB3\x94\x00\xEA\xB3\x95"
-     "\x00\xEA"
-     "\xB3\x96\x00\xEA\xB3\x97\x00\xEA\xB3\x98\x00\xEA\xB3\x99\x00\xEA\xB3\x9A"
-     "\x00\xEA"
-     "\xB3\x9B\x00\xEA\xB3\x9C\x00\xEA\xB3\x9D\x00\xEA\xB3\x9E\x00\xEA\xB3\x9F"
-     "\x00\xEA"
-     "\xB3\xA0\x00\xEA\xB3\xA1\x00\xEA\xB3\xA2\x00\xEA\xB3\xA3\x00\xEA\xB3\xA4"
-     "\x00\xEA"
-     "\xB3\xA5\x00\xEA\xB3\xA6\x00\xEA\xB3\xA7\x00\xEA\xB3\xA8\x00\xEA\xB3\xA9"
-     "\x00\xEA"
-     "\xB3\xAA\x00\xEA\xB3\xAB\x00\xEA\xB3\xAC\x00\xEA\xB3\xAD\x00\xEA\xB3\xAE"
-     "\x00\xEA"
-     "\xB3\xAF\x00\xEA\xB3\xB0\x00\xEA\xB3\xB1\x00\xEA\xB3\xB2\x00\xEA\xB3\xB3"
-     "\x00\xEA"
-     "\xB3\xB4\x00\xEA\xB3\xB5\x00\xEA\xB3\xB6\x00\xEA\xB3\xB7\x00\xEA\xB3\xB8"
-     "\x00\xEA"
-     "\xB3\xB9\x00\xEA\xB3\xBA\x00\xEA\xB3\xBB\x00\xEA\xB3\xBC\x00\xEA\xB3\xBD"
-     "\x00\xEA"
-     "\xB3\xBE\x00\xEA\xB3\xBF\x00\xEA\xB4\x80\x00\xEA\xB4\x81\x00\xEA\xB4\x82"
-     "\x00\xEA"
-     "\xB4\x83\x00\xEA\xB4\x84\x00\xEA\xB4\x85\x00\xEA\xB4\x86\x00\xEA\xB4\x87"
-     "\x00\xEA"
-     "\xB4\x88\x00\xEA\xB4\x89\x00\xEA\xB4\x8A\x00\xEA\xB4\x8B\x00\xEA\xB4\x8C"
-     "\x00\xEA"
-     "\xB4\x8D\x00\xEA\xB4\x8E\x00\xEA\xB4\x8F\x00\xEA\xB4\x90\x00\xEA\xB4\x91"
-     "\x00\xEA"
-     "\xB4\x92\x00\xEA\xB4\x93\x00\xEA\xB4\x94\x00\xEA\xB4\x95\x00\xEA\xB4\x96"
-     "\x00\xEA"
-     "\xB4\x97\x00\xEA\xB4\x98\x00\xEA\xB4\x99\x00\xEA\xB4\x9A\x00\xEA\xB4\x9B"
-     "\x00\xEA"
-     "\xB4\x9C\x00\xEA\xB4\x9D\x00\xEA\xB4\x9E\x00\xEA\xB4\x9F\x00\xEA\xB4\xA0"
-     "\x00\xEA"
-     "\xB4\xA1\x00\xEA\xB4\xA2\x00\xEA\xB4\xA3\x00\xEA\xB4\xA4\x00\xEA\xB4\xA5"
-     "\x00\xEA"
-     "\xB4\xA6\x00\xEA\xB4\xA7\x00\xEA\xB4\xA8\x00\xEA\xB4\xA9\x00\xEA\xB4\xAA"
-     "\x00\xEA"
-     "\xB4\xAB\x00\xEA\xB4\xAC\x00\xEA\xB4\xAD\x00\xEA\xB4\xAE\x00\xEA\xB4\xAF"
-     "\x00\xEA"
-     "\xB4\xB0\x00\xEA\xB4\xB1\x00\xEA\xB4\xB2\x00\xEA\xB4\xB3\x00\xEA\xB4\xB4"
-     "\x00\xEA"
-     "\xB4\xB5\x00\xEA\xB4\xB6\x00\xEA\xB4\xB7\x00\xEA\xB4\xB8\x00\xEA\xB4\xB9"
-     "\x00\xEA"
-     "\xB4\xBA\x00\xEA\xB4\xBB\x00\xEA\xB4\xBC\x00\xEA\xB4\xBD\x00\xEA\xB4\xBE"
-     "\x00\xEA"
-     "\xB4\xBF\x00\xEA\xB5\x80\x00\xEA\xB5\x81\x00\xEA\xB5\x82\x00\xEA\xB5\x83"
-     "\x00\xEA"
-     "\xB5\x84\x00\xEA\xB5\x85\x00\xEA\xB5\x86\x00\xEA\xB5\x87\x00\xEA\xB5\x88"
-     "\x00\xEA"
-     "\xB5\x89\x00\xEA\xB5\x8A\x00\xEA\xB5\x8B\x00\xEA\xB5\x8C\x00\xEA\xB5\x8D"
-     "\x00\xEA"
-     "\xB5\x8E\x00\xEA\xB5\x8F\x00\xEA\xB5\x90\x00\xEA\xB5\x91\x00\xEA\xB5\x92"
-     "\x00\xEA"
-     "\xB5\x93\x00\xEA\xB5\x94\x00\xEA\xB5\x95\x00\xEA\xB5\x96\x00\xEA\xB5\x97"
-     "\x00\xEA"
-     "\xB5\x98\x00\xEA\xB5\x99\x00\xEA\xB5\x9A\x00\xEA\xB5\x9B\x00\xEA\xB5\x9C"
-     "\x00\xEA"
-     "\xB5\x9D\x00\xEA\xB5\x9E\x00\xEA\xB5\x9F\x00\xEA\xB5\xA0\x00\xEA\xB5\xA1"
-     "\x00\xEA"
-     "\xB5\xA2\x00\xEA\xB5\xA3\x00\xEA\xB5\xA4\x00\xEA\xB5\xA5\x00\xEA\xB5\xA6"
-     "\x00\xEA"
-     "\xB5\xA7\x00\xEA\xB5\xA8\x00\xEA\xB5\xA9\x00\xEA\xB5\xAA\x00\xEA\xB5\xAB"
-     "\x00\xEA"
-     "\xB5\xAC\x00\xEA\xB5\xAD\x00\xEA\xB5\xAE\x00\xEA\xB5\xAF\x00\xEA\xB5\xB0"
-     "\x00\xEA"
-     "\xB5\xB1\x00\xEA\xB5\xB2\x00\xEA\xB5\xB3\x00\xEA\xB5\xB4\x00\xEA\xB5\xB5"
-     "\x00\xEA"
-     "\xB5\xB6\x00\xEA\xB5\xB7\x00\xEA\xB5\xB8\x00\xEA\xB5\xB9\x00\xEA\xB5\xBA"
-     "\x00\xEA"
-     "\xB5\xBB\x00\xEA\xB5\xBC\x00\xEA\xB5\xBD\x00\xEA\xB5\xBE\x00\xEA\xB5\xBF"
-     "\x00\xEA"
-     "\xB6\x80\x00\xEA\xB6\x81\x00\xEA\xB6\x82\x00\xEA\xB6\x83\x00\xEA\xB6\x84"
-     "\x00\xEA"
-     "\xB6\x85\x00\xEA\xB6\x86\x00\xEA\xB6\x87\x00\xEA\xB6\x88\x00\xEA\xB6\x89"
-     "\x00\xEA"
-     "\xB6\x8A\x00\xEA\xB6\x8B\x00\xEA\xB6\x8C\x00\xEA\xB6\x8D\x00\xEA\xB6\x8E"
-     "\x00\xEA"
-     "\xB6\x8F\x00\xEA\xB6\x90\x00\xEA\xB6\x91\x00\xEA\xB6\x92\x00\xEA\xB6\x93"
-     "\x00\xEA"
-     "\xB6\x94\x00\xEA\xB6\x95\x00\xEA\xB6\x96\x00\xEA\xB6\x97\x00\xEA\xB6\x98"
-     "\x00\xEA"
-     "\xB6\x99\x00\xEA\xB6\x9A\x00\xEA\xB6\x9B\x00\xEA\xB6\x9C\x00\xEA\xB6\x9D"
-     "\x00\xEA"
-     "\xB6\x9E\x00\xEA\xB6\x9F\x00\xEA\xB6\xA0\x00\xEA\xB6\xA1\x00\xEA\xB6\xA2"
-     "\x00\xEA"
-     "\xB6\xA3\x00\xEA\xB6\xA4\x00\xEA\xB6\xA5\x00\xEA\xB6\xA6\x00\xEA\xB6\xA7"
-     "\x00\xEA"
-     "\xB6\xA8\x00\xEA\xB6\xA9\x00\xEA\xB6\xAA\x00\xEA\xB6\xAB\x00\xEA\xB6\xAC"
-     "\x00\xEA"
-     "\xB6\xAD\x00\xEA\xB6\xAE\x00\xEA\xB6\xAF\x00\xEA\xB6\xB0\x00\xEA\xB6\xB1"
-     "\x00\xEA"
-     "\xB6\xB2\x00\xEA\xB6\xB3\x00\xEA\xB6\xB4\x00\xEA\xB6\xB5\x00\xEA\xB6\xB6"
-     "\x00\xEA"
-     "\xB6\xB7\x00\xEA\xB6\xB8\x00\xEA\xB6\xB9\x00\xEA\xB6\xBA\x00\xEA\xB6\xBB"
-     "\x00\xEA"
-     "\xB6\xBC\x00\xEA\xB6\xBD\x00\xEA\xB6\xBE\x00\xEA\xB6\xBF\x00\xEA\xB7\x80"
-     "\x00\xEA"
-     "\xB7\x81\x00\xEA\xB7\x82\x00\xEA\xB7\x83\x00\xEA\xB7\x84\x00\xEA\xB7\x85"
-     "\x00\xEA"
-     "\xB7\x86\x00\xEA\xB7\x87\x00\xEA\xB7\x88\x00\xEA\xB7\x89\x00\xEA\xB7\x8A"
-     "\x00\xEA"
-     "\xB7\x8B\x00\xEA\xB7\x8C\x00\xEA\xB7\x8D\x00\xEA\xB7\x8E\x00\xEA\xB7\x8F"
-     "\x00\xEA"
-     "\xB7\x90\x00\xEA\xB7\x91\x00\xEA\xB7\x92\x00\xEA\xB7\x93\x00\xEA\xB7\x94"
-     "\x00\xEA"
-     "\xB7\x95\x00\xEA\xB7\x96\x00\xEA\xB7\x97\x00\xEA\xB7\x98\x00\xEA\xB7\x99"
-     "\x00\xEA"
-     "\xB7\x9A\x00\xEA\xB7\x9B\x00\xEA\xB7\x9C\x00\xEA\xB7\x9D\x00\xEA\xB7\x9E"
-     "\x00\xEA"
-     "\xB7\x9F\x00\xEA\xB7\xA0\x00\xEA\xB7\xA1\x00\xEA\xB7\xA2\x00\xEA\xB7\xA3"
-     "\x00\xEA"
-     "\xB7\xA4\x00\xEA\xB7\xA5\x00\xEA\xB7\xA6\x00\xEA\xB7\xA7\x00\xEA\xB7\xA8"
-     "\x00\xEA"
-     "\xB7\xA9\x00\xEA\xB7\xAA\x00\xEA\xB7\xAB\x00\xEA\xB7\xAC\x00\xEA\xB7\xAD"
-     "\x00\xEA"
-     "\xB7\xAE\x00\xEA\xB7\xAF\x00\xEA\xB7\xB0\x00\xEA\xB7\xB1\x00\xEA\xB7\xB2"
-     "\x00\xEA"
-     "\xB7\xB3\x00\xEA\xB7\xB4\x00\xEA\xB7\xB5\x00\xEA\xB7\xB6\x00\xEA\xB7\xB7"
-     "\x00\xEA"
-     "\xB7\xB8\x00\xEA\xB7\xB9\x00\xEA\xB7\xBA\x00\xEA\xB7\xBB\x00\xEA\xB7\xBC"
-     "\x00\xEA"
-     "\xB7\xBD\x00\xEA\xB7\xBE\x00\xEA\xB7\xBF\x00\xEA\xB8\x80\x00\xEA\xB8\x81"
-     "\x00\xEA"
-     "\xB8\x82\x00\xEA\xB8\x83\x00\xEA\xB8\x84\x00\xEA\xB8\x85\x00\xEA\xB8\x86"
-     "\x00\xEA"
-     "\xB8\x87\x00\xEA\xB8\x88\x00\xEA\xB8\x89\x00\xEA\xB8\x8A\x00\xEA\xB8\x8B"
-     "\x00\xEA"
-     "\xB8\x8C\x00\xEA\xB8\x8D\x00\xEA\xB8\x8E\x00\xEA\xB8\x8F\x00\xEA\xB8\x90"
-     "\x00\xEA"
-     "\xB8\x91\x00\xEA\xB8\x92\x00\xEA\xB8\x93\x00\xEA\xB8\x94\x00\xEA\xB8\x95"
-     "\x00\xEA"
-     "\xB8\x96\x00\xEA\xB8\x97\x00\xEA\xB8\x98\x00\xEA\xB8\x99\x00\xEA\xB8\x9A"
-     "\x00\xEA"
-     "\xB8\x9B\x00\xEA\xB8\x9C\x00\xEA\xB8\x9D\x00\xEA\xB8\x9E\x00\xEA\xB8\x9F"
-     "\x00\xEA"
-     "\xB8\xA0\x00\xEA\xB8\xA1\x00\xEA\xB8\xA2\x00\xEA\xB8\xA3\x00\xEA\xB8\xA4"
-     "\x00\xEA"
-     "\xB8\xA5\x00\xEA\xB8\xA6\x00\xEA\xB8\xA7\x00\xEA\xB8\xA8\x00\xEA\xB8\xA9"
-     "\x00\xEA"
-     "\xB8\xAA\x00\xEA\xB8\xAB\x00\xEA\xB8\xAC\x00\xEA\xB8\xAD\x00\xEA\xB8\xAE"
-     "\x00\xEA"
-     "\xB8\xAF\x00\xEA\xB8\xB0\x00\xEA\xB8\xB1\x00\xEA\xB8\xB2\x00\xEA\xB8\xB3"
-     "\x00\xEA"
-     "\xB8\xB4\x00\xEA\xB8\xB5\x00\xEA\xB8\xB6\x00\xEA\xB8\xB7\x00\xEA\xB8\xB8"
-     "\x00\xEA"
-     "\xB8\xB9\x00\xEA\xB8\xBA\x00\xEA\xB8\xBB\x00\xEA\xB8\xBC\x00\xEA\xB8\xBD"
-     "\x00\xEA"
-     "\xB8\xBE\x00\xEA\xB8\xBF\x00\xEA\xB9\x80\x00\xEA\xB9\x81\x00\xEA\xB9\x82"
-     "\x00\xEA"
-     "\xB9\x83\x00\xEA\xB9\x84\x00\xEA\xB9\x85\x00\xEA\xB9\x86\x00\xEA\xB9\x87"
-     "\x00\xEA"
-     "\xB9\x88\x00\xEA\xB9\x89\x00\xEA\xB9\x8A\x00\xEA\xB9\x8B\x00\xEA\xB9\x8C"
-     "\x00\xEA"
-     "\xB9\x8D\x00\xEA\xB9\x8E\x00\xEA\xB9\x8F\x00\xEA\xB9\x90\x00\xEA\xB9\x91"
-     "\x00\xEA"
-     "\xB9\x92\x00\xEA\xB9\x93\x00\xEA\xB9\x94\x00\xEA\xB9\x95\x00\xEA\xB9\x96"
-     "\x00\xEA"
-     "\xB9\x97\x00\xEA\xB9\x98\x00\xEA\xB9\x99\x00\xEA\xB9\x9A\x00\xEA\xB9\x9B"
-     "\x00\xEA"
-     "\xB9\x9C\x00\xEA\xB9\x9D\x00\xEA\xB9\x9E\x00\xEA\xB9\x9F\x00\xEA\xB9\xA0"
-     "\x00\xEA"
-     "\xB9\xA1\x00\xEA\xB9\xA2\x00\xEA\xB9\xA3\x00\xEA\xB9\xA4\x00\xEA\xB9\xA5"
-     "\x00\xEA"
-     "\xB9\xA6\x00\xEA\xB9\xA7\x00\xEA\xB9\xA8\x00\xEA\xB9\xA9\x00\xEA\xB9\xAA"
-     "\x00\xEA"
-     "\xB9\xAB\x00\xEA\xB9\xAC\x00\xEA\xB9\xAD\x00\xEA\xB9\xAE\x00\xEA\xB9\xAF"
-     "\x00\xEA"
-     "\xB9\xB0\x00\xEA\xB9\xB1\x00\xEA\xB9\xB2\x00\xEA\xB9\xB3\x00\xEA\xB9\xB4"
-     "\x00\xEA"
-     "\xB9\xB5\x00\xEA\xB9\xB6\x00\xEA\xB9\xB7\x00\xEA\xB9\xB8\x00\xEA\xB9\xB9"
-     "\x00\xEA"
-     "\xB9\xBA\x00\xEA\xB9\xBB\x00\xEA\xB9\xBC\x00\xEA\xB9\xBD\x00\xEA\xB9\xBE"
-     "\x00\xEA"
-     "\xB9\xBF\x00\xEA\xBA\x80\x00\xEA\xBA\x81\x00\xEA\xBA\x82\x00\xEA\xBA\x83"
-     "\x00\xEA"
-     "\xBA\x84\x00\xEA\xBA\x85\x00\xEA\xBA\x86\x00\xEA\xBA\x87\x00\xEA\xBA\x88"
-     "\x00\xEA"
-     "\xBA\x89\x00\xEA\xBA\x8A\x00\xEA\xBA\x8B\x00\xEA\xBA\x8C\x00\xEA\xBA\x8D"
-     "\x00\xEA"
-     "\xBA\x8E\x00\xEA\xBA\x8F\x00\xEA\xBA\x90\x00\xEA\xBA\x91\x00\xEA\xBA\x92"
-     "\x00\xEA"
-     "\xBA\x93\x00\xEA\xBA\x94\x00\xEA\xBA\x95\x00\xEA\xBA\x96\x00\xEA\xBA\x97"
-     "\x00\xEA"
-     "\xBA\x98\x00\xEA\xBA\x99\x00\xEA\xBA\x9A\x00\xEA\xBA\x9B\x00\xEA\xBA\x9C"
-     "\x00\xEA"
-     "\xBA\x9D\x00\xEA\xBA\x9E\x00\xEA\xBA\x9F\x00\xEA\xBA\xA0\x00\xEA\xBA\xA1"
-     "\x00\xEA"
-     "\xBA\xA2\x00\xEA\xBA\xA3\x00\xEA\xBA\xA4\x00\xEA\xBA\xA5\x00\xEA\xBA\xA6"
-     "\x00\xEA"
-     "\xBA\xA7\x00\xEA\xBA\xA8\x00\xEA\xBA\xA9\x00\xEA\xBA\xAA\x00\xEA\xBA\xAB"
-     "\x00\xEA"
-     "\xBA\xAC\x00\xEA\xBA\xAD\x00\xEA\xBA\xAE\x00\xEA\xBA\xAF\x00\xEA\xBA\xB0"
-     "\x00\xEA"
-     "\xBA\xB1\x00\xEA\xBA\xB2\x00\xEA\xBA\xB3\x00\xEA\xBA\xB4\x00\xEA\xBA\xB5"
-     "\x00\xEA"
-     "\xBA\xB6\x00\xEA\xBA\xB7\x00\xEA\xBA\xB8\x00\xEA\xBA\xB9\x00\xEA\xBA\xBA"
-     "\x00\xEA"
-     "\xBA\xBB\x00\xEA\xBA\xBC\x00\xEA\xBA\xBD\x00\xEA\xBA\xBE\x00\xEA\xBA\xBF"
-     "\x00\xEA"
-     "\xBB\x80\x00\xEA\xBB\x81\x00\xEA\xBB\x82\x00\xEA\xBB\x83\x00\xEA\xBB\x84"
-     "\x00\xEA"
-     "\xBB\x85\x00\xEA\xBB\x86\x00\xEA\xBB\x87\x00\xEA\xBB\x88\x00\xEA\xBB\x89"
-     "\x00\xEA"
-     "\xBB\x8A\x00\xEA\xBB\x8B\x00\xEA\xBB\x8C\x00\xEA\xBB\x8D\x00\xEA\xBB\x8E"
-     "\x00\xEA"
-     "\xBB\x8F\x00\xEA\xBB\x90\x00\xEA\xBB\x91\x00\xEA\xBB\x92\x00\xEA\xBB\x93"
-     "\x00\xEA"
-     "\xBB\x94\x00\xEA\xBB\x95\x00\xEA\xBB\x96\x00\xEA\xBB\x97\x00\xEA\xBB\x98"
-     "\x00\xEA"
-     "\xBB\x99\x00\xEA\xBB\x9A\x00\xEA\xBB\x9B\x00\xEA\xBB\x9C\x00\xEA\xBB\x9D"
-     "\x00\xEA"
-     "\xBB\x9E\x00\xEA\xBB\x9F\x00\xEA\xBB\xA0\x00\xEA\xBB\xA1\x00\xEA\xBB\xA2"
-     "\x00\xEA"
-     "\xBB\xA3\x00\xEA\xBB\xA4\x00\xEA\xBB\xA5\x00\xEA\xBB\xA6\x00\xEA\xBB\xA7"
-     "\x00\xEA"
-     "\xBB\xA8\x00\xEA\xBB\xA9\x00\xEA\xBB\xAA\x00\xEA\xBB\xAB\x00\xEA\xBB\xAC"
-     "\x00\xEA"
-     "\xBB\xAD\x00\xEA\xBB\xAE\x00\xEA\xBB\xAF\x00\xEA\xBB\xB0\x00\xEA\xBB\xB1"
-     "\x00\xEA"
-     "\xBB\xB2\x00\xEA\xBB\xB3\x00\xEA\xBB\xB4\x00\xEA\xBB\xB5\x00\xEA\xBB\xB6"
-     "\x00\xEA"
-     "\xBB\xB7\x00\xEA\xBB\xB8\x00\xEA\xBB\xB9\x00\xEA\xBB\xBA\x00\xEA\xBB\xBB"
-     "\x00\xEA"
-     "\xBB\xBC\x00\xEA\xBB\xBD\x00\xEA\xBB\xBE\x00\xEA\xBB\xBF\x00\xEA\xBC\x80"
-     "\x00\xEA"
-     "\xBC\x81\x00\xEA\xBC\x82\x00\xEA\xBC\x83\x00\xEA\xBC\x84\x00\xEA\xBC\x85"
-     "\x00\xEA"
-     "\xBC\x86\x00\xEA\xBC\x87\x00\xEA\xBC\x88\x00\xEA\xBC\x89\x00\xEA\xBC\x8A"
-     "\x00\xEA"
-     "\xBC\x8B\x00\xEA\xBC\x8C\x00\xEA\xBC\x8D\x00\xEA\xBC\x8E\x00\xEA\xBC\x8F"
-     "\x00\xEA"
-     "\xBC\x90\x00\xEA\xBC\x91\x00\xEA\xBC\x92\x00\xEA\xBC\x93\x00\xEA\xBC\x94"
-     "\x00\xEA"
-     "\xBC\x95\x00\xEA\xBC\x96\x00\xEA\xBC\x97\x00\xEA\xBC\x98\x00\xEA\xBC\x99"
-     "\x00\xEA"
-     "\xBC\x9A\x00\xEA\xBC\x9B\x00\xEA\xBC\x9C\x00\xEA\xBC\x9D\x00\xEA\xBC\x9E"
-     "\x00\xEA"
-     "\xBC\x9F\x00\xEA\xBC\xA0\x00\xEA\xBC\xA1\x00\xEA\xBC\xA2\x00\xEA\xBC\xA3"
-     "\x00\xEA"
-     "\xBC\xA4\x00\xEA\xBC\xA5\x00\xEA\xBC\xA6\x00\xEA\xBC\xA7\x00\xEA\xBC\xA8"
-     "\x00\xEA"
-     "\xBC\xA9\x00\xEA\xBC\xAA\x00\xEA\xBC\xAB\x00\xEA\xBC\xAC\x00\xEA\xBC\xAD"
-     "\x00\xEA"
-     "\xBC\xAE\x00\xEA\xBC\xAF\x00\xEA\xBC\xB0\x00\xEA\xBC\xB1\x00\xEA\xBC\xB2"
-     "\x00\xEA"
-     "\xBC\xB3\x00\xEA\xBC\xB4\x00\xEA\xBC\xB5\x00\xEA\xBC\xB6\x00\xEA\xBC\xB7"
-     "\x00\xEA"
-     "\xBC\xB8\x00\xEA\xBC\xB9\x00\xEA\xBC\xBA\x00\xEA\xBC\xBB\x00\xEA\xBC\xBC"
-     "\x00\xEA"
-     "\xBC\xBD\x00\xEA\xBC\xBE\x00\xEA\xBC\xBF\x00\xEA\xBD\x80\x00\xEA\xBD\x81"
-     "\x00\xEA"
-     "\xBD\x82\x00\xEA\xBD\x83\x00\xEA\xBD\x84\x00\xEA\xBD\x85\x00\xEA\xBD\x86"
-     "\x00\xEA"
-     "\xBD\x87\x00\xEA\xBD\x88\x00\xEA\xBD\x89\x00\xEA\xBD\x8A\x00\xEA\xBD\x8B"
-     "\x00\xEA"
-     "\xBD\x8C\x00\xEA\xBD\x8D\x00\xEA\xBD\x8E\x00\xEA\xBD\x8F\x00\xEA\xBD\x90"
-     "\x00\xEA"
-     "\xBD\x91\x00\xEA\xBD\x92\x00\xEA\xBD\x93\x00\xEA\xBD\x94\x00\xEA\xBD\x95"
-     "\x00\xEA"
-     "\xBD\x96\x00\xEA\xBD\x97\x00\xEA\xBD\x98\x00\xEA\xBD\x99\x00\xEA\xBD\x9A"
-     "\x00\xEA"
-     "\xBD\x9B\x00\xEA\xBD\x9C\x00\xEA\xBD\x9D\x00\xEA\xBD\x9E\x00\xEA\xBD\x9F"
-     "\x00\xEA"
-     "\xBD\xA0\x00\xEA\xBD\xA1\x00\xEA\xBD\xA2\x00\xEA\xBD\xA3\x00\xEA\xBD\xA4"
-     "\x00\xEA"
-     "\xBD\xA5\x00\xEA\xBD\xA6\x00\xEA\xBD\xA7\x00\xEA\xBD\xA8\x00\xEA\xBD\xA9"
-     "\x00\xEA"
-     "\xBD\xAA\x00\xEA\xBD\xAB\x00\xEA\xBD\xAC\x00\xEA\xBD\xAD\x00\xEA\xBD\xAE"
-     "\x00\xEA"
-     "\xBD\xAF\x00\xEA\xBD\xB0\x00\xEA\xBD\xB1\x00\xEA\xBD\xB2\x00\xEA\xBD\xB3"
-     "\x00\xEA"
-     "\xBD\xB4\x00\xEA\xBD\xB5\x00\xEA\xBD\xB6\x00\xEA\xBD\xB7\x00\xEA\xBD\xB8"
-     "\x00\xEA"
-     "\xBD\xB9\x00\xEA\xBD\xBA\x00\xEA\xBD\xBB\x00\xEA\xBD\xBC\x00\xEA\xBD\xBD"
-     "\x00\xEA"
-     "\xBD\xBE\x00\xEA\xBD\xBF\x00\xEA\xBE\x80\x00\xEA\xBE\x81\x00\xEA\xBE\x82"
-     "\x00\xEA"
-     "\xBE\x83\x00\xEA\xBE\x84\x00\xEA\xBE\x85\x00\xEA\xBE\x86\x00\xEA\xBE\x87"
-     "\x00\xEA"
-     "\xBE\x88\x00\xEA\xBE\x89\x00\xEA\xBE\x8A\x00\xEA\xBE\x8B\x00\xEA\xBE\x8C"
-     "\x00\xEA"
-     "\xBE\x8D\x00\xEA\xBE\x8E\x00\xEA\xBE\x8F\x00\xEA\xBE\x90\x00\xEA\xBE\x91"
-     "\x00\xEA"
-     "\xBE\x92\x00\xEA\xBE\x93\x00\xEA\xBE\x94\x00\xEA\xBE\x95\x00\xEA\xBE\x96"
-     "\x00\xEA"
-     "\xBE\x97\x00\xEA\xBE\x98\x00\xEA\xBE\x99\x00\xEA\xBE\x9A\x00\xEA\xBE\x9B"
-     "\x00\xEA"
-     "\xBE\x9C\x00\xEA\xBE\x9D\x00\xEA\xBE\x9E\x00\xEA\xBE\x9F\x00\xEA\xBE\xA0"
-     "\x00\xEA"
-     "\xBE\xA1\x00\xEA\xBE\xA2\x00\xEA\xBE\xA3\x00\xEA\xBE\xA4\x00\xEA\xBE\xA5"
-     "\x00\xEA"
-     "\xBE\xA6\x00\xEA\xBE\xA7\x00\xEA\xBE\xA8\x00\xEA\xBE\xA9\x00\xEA\xBE\xAA"
-     "\x00\xEA"
-     "\xBE\xAB\x00\xEA\xBE\xAC\x00\xEA\xBE\xAD\x00\xEA\xBE\xAE\x00\xEA\xBE\xAF"
-     "\x00\xEA"
-     "\xBE\xB0\x00\xEA\xBE\xB1\x00\xEA\xBE\xB2\x00\xEA\xBE\xB3\x00\xEA\xBE\xB4"
-     "\x00\xEA"
-     "\xBE\xB5\x00\xEA\xBE\xB6\x00\xEA\xBE\xB7\x00\xEA\xBE\xB8\x00\xEA\xBE\xB9"
-     "\x00\xEA"
-     "\xBE\xBA\x00\xEA\xBE\xBB\x00\xEA\xBE\xBC\x00\xEA\xBE\xBD\x00\xEA\xBE\xBE"
-     "\x00\xEA"
-     "\xBE\xBF\x00\xEA\xBF\x80\x00\xEA\xBF\x81\x00\xEA\xBF\x82\x00\xEA\xBF\x83"
-     "\x00\xEA"
-     "\xBF\x84\x00\xEA\xBF\x85\x00\xEA\xBF\x86\x00\xEA\xBF\x87\x00\xEA\xBF\x88"
-     "\x00\xEA"
-     "\xBF\x89\x00\xEA\xBF\x8A\x00\xEA\xBF\x8B\x00\xEA\xBF\x8C\x00\xEA\xBF\x8D"
-     "\x00\xEA"
-     "\xBF\x8E\x00\xEA\xBF\x8F\x00\xEA\xBF\x90\x00\xEA\xBF\x91\x00\xEA\xBF\x92"
-     "\x00\xEA"
-     "\xBF\x93\x00\xEA\xBF\x94\x00\xEA\xBF\x95\x00\xEA\xBF\x96\x00\xEA\xBF\x97"
-     "\x00\xEA"
-     "\xBF\x98\x00\xEA\xBF\x99\x00\xEA\xBF\x9A\x00\xEA\xBF\x9B\x00\xEA\xBF\x9C"
-     "\x00\xEA"
-     "\xBF\x9D\x00\xEA\xBF\x9E\x00\xEA\xBF\x9F\x00\xEA\xBF\xA0\x00\xEA\xBF\xA1"
-     "\x00\xEA"
-     "\xBF\xA2\x00\xEA\xBF\xA3\x00\xEA\xBF\xA4\x00\xEA\xBF\xA5\x00\xEA\xBF\xA6"
-     "\x00\xEA"
-     "\xBF\xA7\x00\xEA\xBF\xA8\x00\xEA\xBF\xA9\x00\xEA\xBF\xAA\x00\xEA\xBF\xAB"
-     "\x00\xEA"
-     "\xBF\xAC\x00\xEA\xBF\xAD\x00\xEA\xBF\xAE\x00\xEA\xBF\xAF\x00\xEA\xBF\xB0"
-     "\x00\xEA"
-     "\xBF\xB1\x00\xEA\xBF\xB2\x00\xEA\xBF\xB3\x00\xEA\xBF\xB4\x00\xEA\xBF\xB5"
-     "\x00\xEA"
-     "\xBF\xB6\x00\xEA\xBF\xB7\x00\xEA\xBF\xB8\x00\xEA\xBF\xB9\x00\xEA\xBF\xBA"
-     "\x00\xEA"
-     "\xBF\xBB\x00\xEA\xBF\xBC\x00\xEA\xBF\xBD\x00\xEA\xBF\xBE\x00\xEA\xBF\xBF"
-     "\x00\xEB"
-     "\x80\x80\x00\xEB\x80\x81\x00\xEB\x80\x82\x00\xEB\x80\x83\x00\xEB\x80\x84"
-     "\x00\xEB"
-     "\x80\x85\x00\xEB\x80\x86\x00\xEB\x80\x87\x00\xEB\x80\x88\x00\xEB\x80\x89"
-     "\x00\xEB"
-     "\x80\x8A\x00\xEB\x80\x8B\x00\xEB\x80\x8C\x00\xEB\x80\x8D\x00\xEB\x80\x8E"
-     "\x00\xEB"
-     "\x80\x8F\x00\xEB\x80\x90\x00\xEB\x80\x91\x00\xEB\x80\x92\x00\xEB\x80\x93"
-     "\x00\xEB"
-     "\x80\x94\x00\xEB\x80\x95\x00\xEB\x80\x96\x00\xEB\x80\x97\x00\xEB\x80\x98"
-     "\x00\xEB"
-     "\x80\x99\x00\xEB\x80\x9A\x00\xEB\x80\x9B\x00\xEB\x80\x9C\x00\xEB\x80\x9D"
-     "\x00\xEB"
-     "\x80\x9E\x00\xEB\x80\x9F\x00\xEB\x80\xA0\x00\xEB\x80\xA1\x00\xEB\x80\xA2"
-     "\x00\xEB"
-     "\x80\xA3\x00\xEB\x80\xA4\x00\xEB\x80\xA5\x00\xEB\x80\xA6\x00\xEB\x80\xA7"
-     "\x00\xEB"
-     "\x80\xA8\x00\xEB\x80\xA9\x00\xEB\x80\xAA\x00\xEB\x80\xAB\x00\xEB\x80\xAC"
-     "\x00\xEB"
-     "\x80\xAD\x00\xEB\x80\xAE\x00\xEB\x80\xAF\x00\xEB\x80\xB0\x00\xEB\x80\xB1"
-     "\x00\xEB"
-     "\x80\xB2\x00\xEB\x80\xB3\x00\xEB\x80\xB4\x00\xEB\x80\xB5\x00\xEB\x80\xB6"
-     "\x00\xEB"
-     "\x80\xB7\x00\xEB\x80\xB8\x00\xEB\x80\xB9\x00\xEB\x80\xBA\x00\xEB\x80\xBB"
-     "\x00\xEB"
-     "\x80\xBC\x00\xEB\x80\xBD\x00\xEB\x80\xBE\x00\xEB\x80\xBF\x00\xEB\x81\x80"
-     "\x00\xEB"
-     "\x81\x81\x00\xEB\x81\x82\x00\xEB\x81\x83\x00\xEB\x81\x84\x00\xEB\x81\x85"
-     "\x00\xEB"
-     "\x81\x86\x00\xEB\x81\x87\x00\xEB\x81\x88\x00\xEB\x81\x89\x00\xEB\x81\x8A"
-     "\x00\xEB"
-     "\x81\x8B\x00\xEB\x81\x8C\x00\xEB\x81\x8D\x00\xEB\x81\x8E\x00\xEB\x81\x8F"
-     "\x00\xEB"
-     "\x81\x90\x00\xEB\x81\x91\x00\xEB\x81\x92\x00\xEB\x81\x93\x00\xEB\x81\x94"
-     "\x00\xEB"
-     "\x81\x95\x00\xEB\x81\x96\x00\xEB\x81\x97\x00\xEB\x81\x98\x00\xEB\x81\x99"
-     "\x00\xEB"
-     "\x81\x9A\x00\xEB\x81\x9B\x00\xEB\x81\x9C\x00\xEB\x81\x9D\x00\xEB\x81\x9E"
-     "\x00\xEB"
-     "\x81\x9F\x00\xEB\x81\xA0\x00\xEB\x81\xA1\x00\xEB\x81\xA2\x00\xEB\x81\xA3"
-     "\x00\xEB"
-     "\x81\xA4\x00\xEB\x81\xA5\x00\xEB\x81\xA6\x00\xEB\x81\xA7\x00\xEB\x81\xA8"
-     "\x00\xEB"
-     "\x81\xA9\x00\xEB\x81\xAA\x00\xEB\x81\xAB\x00\xEB\x81\xAC\x00\xEB\x81\xAD"
-     "\x00\xEB"
-     "\x81\xAE\x00\xEB\x81\xAF\x00\xEB\x81\xB0\x00\xEB\x81\xB1\x00\xEB\x81\xB2"
-     "\x00\xEB"
-     "\x81\xB3\x00\xEB\x81\xB4\x00\xEB\x81\xB5\x00\xEB\x81\xB6\x00\xEB\x81\xB7"
-     "\x00\xEB"
-     "\x81\xB8\x00\xEB\x81\xB9\x00\xEB\x81\xBA\x00\xEB\x81\xBB\x00\xEB\x81\xBC"
-     "\x00\xEB"
-     "\x81\xBD\x00\xEB\x81\xBE\x00\xEB\x81\xBF\x00\xEB\x82\x80\x00\xEB\x82\x81"
-     "\x00\xEB"
-     "\x82\x82\x00\xEB\x82\x83\x00\xEB\x82\x84\x00\xEB\x82\x85\x00\xEB\x82\x86"
-     "\x00\xEB"
-     "\x82\x87\x00\xEB\x82\x88\x00\xEB\x82\x89\x00\xEB\x82\x8A\x00\xEB\x82\x8B"
-     "\x00\xEB"
-     "\x82\x8C\x00\xEB\x82\x8D\x00\xEB\x82\x8E\x00\xEB\x82\x8F\x00\xEB\x82\x90"
-     "\x00\xEB"
-     "\x82\x91\x00\xEB\x82\x92\x00\xEB\x82\x93\x00\xEB\x82\x94\x00\xEB\x82\x95"
-     "\x00\xEB"
-     "\x82\x96\x00\xEB\x82\x97\x00\xEB\x82\x98\x00\xEB\x82\x99\x00\xEB\x82\x9A"
-     "\x00\xEB"
-     "\x82\x9B\x00\xEB\x82\x9C\x00\xEB\x82\x9D\x00\xEB\x82\x9E\x00\xEB\x82\x9F"
-     "\x00\xEB"
-     "\x82\xA0\x00\xEB\x82\xA1\x00\xEB\x82\xA2\x00\xEB\x82\xA3\x00\xEB\x82\xA4"
-     "\x00\xEB"
-     "\x82\xA5\x00\xEB\x82\xA6\x00\xEB\x82\xA7\x00\xEB\x82\xA8\x00\xEB\x82\xA9"
-     "\x00\xEB"
-     "\x82\xAA\x00\xEB\x82\xAB\x00\xEB\x82\xAC\x00\xEB\x82\xAD\x00\xEB\x82\xAE"
-     "\x00\xEB"
-     "\x82\xAF\x00\xEB\x82\xB0\x00\xEB\x82\xB1\x00\xEB\x82\xB2\x00\xEB\x82\xB3"
-     "\x00\xEB"
-     "\x82\xB4\x00\xEB\x82\xB5\x00\xEB\x82\xB6\x00\xEB\x82\xB7\x00\xEB\x82\xB8"
-     "\x00\xEB"
-     "\x82\xB9\x00\xEB\x82\xBA\x00\xEB\x82\xBB\x00\xEB\x82\xBC\x00\xEB\x82\xBD"
-     "\x00\xEB"
-     "\x82\xBE\x00\xEB\x82\xBF\x00\xEB\x83\x80\x00\xEB\x83\x81\x00\xEB\x83\x82"
-     "\x00\xEB"
-     "\x83\x83\x00\xEB\x83\x84\x00\xEB\x83\x85\x00\xEB\x83\x86\x00\xEB\x83\x87"
-     "\x00\xEB"
-     "\x83\x88\x00\xEB\x83\x89\x00\xEB\x83\x8A\x00\xEB\x83\x8B\x00\xEB\x83\x8C"
-     "\x00\xEB"
-     "\x83\x8D\x00\xEB\x83\x8E\x00\xEB\x83\x8F\x00\xEB\x83\x90\x00\xEB\x83\x91"
-     "\x00\xEB"
-     "\x83\x92\x00\xEB\x83\x93\x00\xEB\x83\x94\x00\xEB\x83\x95\x00\xEB\x83\x96"
-     "\x00\xEB"
-     "\x83\x97\x00\xEB\x83\x98\x00\xEB\x83\x99\x00\xEB\x83\x9A\x00\xEB\x83\x9B"
-     "\x00\xEB"
-     "\x83\x9C\x00\xEB\x83\x9D\x00\xEB\x83\x9E\x00\xEB\x83\x9F\x00\xEB\x83\xA0"
-     "\x00\xEB"
-     "\x83\xA1\x00\xEB\x83\xA2\x00\xEB\x83\xA3\x00\xEB\x83\xA4\x00\xEB\x83\xA5"
-     "\x00\xEB"
-     "\x83\xA6\x00\xEB\x83\xA7\x00\xEB\x83\xA8\x00\xEB\x83\xA9\x00\xEB\x83\xAA"
-     "\x00\xEB"
-     "\x83\xAB\x00\xEB\x83\xAC\x00\xEB\x83\xAD\x00\xEB\x83\xAE\x00\xEB\x83\xAF"
-     "\x00\xEB"
-     "\x83\xB0\x00\xEB\x83\xB1\x00\xEB\x83\xB2\x00\xEB\x83\xB3\x00\xEB\x83\xB4"
-     "\x00\xEB"
-     "\x83\xB5\x00\xEB\x83\xB6\x00\xEB\x83\xB7\x00\xEB\x83\xB8\x00\xEB\x83\xB9"
-     "\x00\xEB"
-     "\x83\xBA\x00\xEB\x83\xBB\x00\xEB\x83\xBC\x00\xEB\x83\xBD\x00\xEB\x83\xBE"
-     "\x00\xEB"
-     "\x83\xBF\x00\xEB\x84\x80\x00\xEB\x84\x81\x00\xEB\x84\x82\x00\xEB\x84\x83"
-     "\x00\xEB"
-     "\x84\x84\x00\xEB\x84\x85\x00\xEB\x84\x86\x00\xEB\x84\x87\x00\xEB\x84\x88"
-     "\x00\xEB"
-     "\x84\x89\x00\xEB\x84\x8A\x00\xEB\x84\x8B\x00\xEB\x84\x8C\x00\xEB\x84\x8D"
-     "\x00\xEB"
-     "\x84\x8E\x00\xEB\x84\x8F\x00\xEB\x84\x90\x00\xEB\x84\x91\x00\xEB\x84\x92"
-     "\x00\xEB"
-     "\x84\x93\x00\xEB\x84\x94\x00\xEB\x84\x95\x00\xEB\x84\x96\x00\xEB\x84\x97"
-     "\x00\xEB"
-     "\x84\x98\x00\xEB\x84\x99\x00\xEB\x84\x9A\x00\xEB\x84\x9B\x00\xEB\x84\x9C"
-     "\x00\xEB"
-     "\x84\x9D\x00\xEB\x84\x9E\x00\xEB\x84\x9F\x00\xEB\x84\xA0\x00\xEB\x84\xA1"
-     "\x00\xEB"
-     "\x84\xA2\x00\xEB\x84\xA3\x00\xEB\x84\xA4\x00\xEB\x84\xA5\x00\xEB\x84\xA6"
-     "\x00\xEB"
-     "\x84\xA7\x00\xEB\x84\xA8\x00\xEB\x84\xA9\x00\xEB\x84\xAA\x00\xEB\x84\xAB"
-     "\x00\xEB"
-     "\x84\xAC\x00\xEB\x84\xAD\x00\xEB\x84\xAE\x00\xEB\x84\xAF\x00\xEB\x84\xB0"
-     "\x00\xEB"
-     "\x84\xB1\x00\xEB\x84\xB2\x00\xEB\x84\xB3\x00\xEB\x84\xB4\x00\xEB\x84\xB5"
-     "\x00\xEB"
-     "\x84\xB6\x00\xEB\x84\xB7\x00\xEB\x84\xB8\x00\xEB\x84\xB9\x00\xEB\x84\xBA"
-     "\x00\xEB"
-     "\x84\xBB\x00\xEB\x84\xBC\x00\xEB\x84\xBD\x00\xEB\x84\xBE\x00\xEB\x84\xBF"
-     "\x00\xEB"
-     "\x85\x80\x00\xEB\x85\x81\x00\xEB\x85\x82\x00\xEB\x85\x83\x00\xEB\x85\x84"
-     "\x00\xEB"
-     "\x85\x85\x00\xEB\x85\x86\x00\xEB\x85\x87\x00\xEB\x85\x88\x00\xEB\x85\x89"
-     "\x00\xEB"
-     "\x85\x8A\x00\xEB\x85\x8B\x00\xEB\x85\x8C\x00\xEB\x85\x8D\x00\xEB\x85\x8E"
-     "\x00\xEB"
-     "\x85\x8F\x00\xEB\x85\x90\x00\xEB\x85\x91\x00\xEB\x85\x92\x00\xEB\x85\x93"
-     "\x00\xEB"
-     "\x85\x94\x00\xEB\x85\x95\x00\xEB\x85\x96\x00\xEB\x85\x97\x00\xEB\x85\x98"
-     "\x00\xEB"
-     "\x85\x99\x00\xEB\x85\x9A\x00\xEB\x85\x9B\x00\xEB\x85\x9C\x00\xEB\x85\x9D"
-     "\x00\xEB"
-     "\x85\x9E\x00\xEB\x85\x9F\x00\xEB\x85\xA0\x00\xEB\x85\xA1\x00\xEB\x85\xA2"
-     "\x00\xEB"
-     "\x85\xA3\x00\xEB\x85\xA4\x00\xEB\x85\xA5\x00\xEB\x85\xA6\x00\xEB\x85\xA7"
-     "\x00\xEB"
-     "\x85\xA8\x00\xEB\x85\xA9\x00\xEB\x85\xAA\x00\xEB\x85\xAB\x00\xEB\x85\xAC"
-     "\x00\xEB"
-     "\x85\xAD\x00\xEB\x85\xAE\x00\xEB\x85\xAF\x00\xEB\x85\xB0\x00\xEB\x85\xB1"
-     "\x00\xEB"
-     "\x85\xB2\x00\xEB\x85\xB3\x00\xEB\x85\xB4\x00\xEB\x85\xB5\x00\xEB\x85\xB6"
-     "\x00\xEB"
-     "\x85\xB7\x00\xEB\x85\xB8\x00\xEB\x85\xB9\x00\xEB\x85\xBA\x00\xEB\x85\xBB"
-     "\x00\xEB"
-     "\x85\xBC\x00\xEB\x85\xBD\x00\xEB\x85\xBE\x00\xEB\x85\xBF\x00\xEB\x86\x80"
-     "\x00\xEB"
-     "\x86\x81\x00\xEB\x86\x82\x00\xEB\x86\x83\x00\xEB\x86\x84\x00\xEB\x86\x85"
-     "\x00\xEB"
-     "\x86\x86\x00\xEB\x86\x87\x00\xEB\x86\x88\x00\xEB\x86\x89\x00\xEB\x86\x8A"
-     "\x00\xEB"
-     "\x86\x8B\x00\xEB\x86\x8C\x00\xEB\x86\x8D\x00\xEB\x86\x8E\x00\xEB\x86\x8F"
-     "\x00\xEB"
-     "\x86\x90\x00\xEB\x86\x91\x00\xEB\x86\x92\x00\xEB\x86\x93\x00\xEB\x86\x94"
-     "\x00\xEB"
-     "\x86\x95\x00\xEB\x86\x96\x00\xEB\x86\x97\x00\xEB\x86\x98\x00\xEB\x86\x99"
-     "\x00\xEB"
-     "\x86\x9A\x00\xEB\x86\x9B\x00\xEB\x86\x9C\x00\xEB\x86\x9D\x00\xEB\x86\x9E"
-     "\x00\xEB"
-     "\x86\x9F\x00\xEB\x86\xA0\x00\xEB\x86\xA1\x00\xEB\x86\xA2\x00\xEB\x86\xA3"
-     "\x00\xEB"
-     "\x86\xA4\x00\xEB\x86\xA5\x00\xEB\x86\xA6\x00\xEB\x86\xA7\x00\xEB\x86\xA8"
-     "\x00\xEB"
-     "\x86\xA9\x00\xEB\x86\xAA\x00\xEB\x86\xAB\x00\xEB\x86\xAC\x00\xEB\x86\xAD"
-     "\x00\xEB"
-     "\x86\xAE\x00\xEB\x86\xAF\x00\xEB\x86\xB0\x00\xEB\x86\xB1\x00\xEB\x86\xB2"
-     "\x00\xEB"
-     "\x86\xB3\x00\xEB\x86\xB4\x00\xEB\x86\xB5\x00\xEB\x86\xB6\x00\xEB\x86\xB7"
-     "\x00\xEB"
-     "\x86\xB8\x00\xEB\x86\xB9\x00\xEB\x86\xBA\x00\xEB\x86\xBB\x00\xEB\x86\xBC"
-     "\x00\xEB"
-     "\x86\xBD\x00\xEB\x86\xBE\x00\xEB\x86\xBF\x00\xEB\x87\x80\x00\xEB\x87\x81"
-     "\x00\xEB"
-     "\x87\x82\x00\xEB\x87\x83\x00\xEB\x87\x84\x00\xEB\x87\x85\x00\xEB\x87\x86"
-     "\x00\xEB"
-     "\x87\x87\x00\xEB\x87\x88\x00\xEB\x87\x89\x00\xEB\x87\x8A\x00\xEB\x87\x8B"
-     "\x00\xEB"
-     "\x87\x8C\x00\xEB\x87\x8D\x00\xEB\x87\x8E\x00\xEB\x87\x8F\x00\xEB\x87\x90"
-     "\x00\xEB"
-     "\x87\x91\x00\xEB\x87\x92\x00\xEB\x87\x93\x00\xEB\x87\x94\x00\xEB\x87\x95"
-     "\x00\xEB"
-     "\x87\x96\x00\xEB\x87\x97\x00\xEB\x87\x98\x00\xEB\x87\x99\x00\xEB\x87\x9A"
-     "\x00\xEB"
-     "\x87\x9B\x00\xEB\x87\x9C\x00\xEB\x87\x9D\x00\xEB\x87\x9E\x00\xEB\x87\x9F"
-     "\x00\xEB"
-     "\x87\xA0\x00\xEB\x87\xA1\x00\xEB\x87\xA2\x00\xEB\x87\xA3\x00\xEB\x87\xA4"
-     "\x00\xEB"
-     "\x87\xA5\x00\xEB\x87\xA6\x00\xEB\x87\xA7\x00\xEB\x87\xA8\x00\xEB\x87\xA9"
-     "\x00\xEB"
-     "\x87\xAA\x00\xEB\x87\xAB\x00\xEB\x87\xAC\x00\xEB\x87\xAD\x00\xEB\x87\xAE"
-     "\x00\xEB"
-     "\x87\xAF\x00\xEB\x87\xB0\x00\xEB\x87\xB1\x00\xEB\x87\xB2\x00\xEB\x87\xB3"
-     "\x00\xEB"
-     "\x87\xB4\x00\xEB\x87\xB5\x00\xEB\x87\xB6\x00\xEB\x87\xB7\x00\xEB\x87\xB8"
-     "\x00\xEB"
-     "\x87\xB9\x00\xEB\x87\xBA\x00\xEB\x87\xBB\x00\xEB\x87\xBC\x00\xEB\x87\xBD"
-     "\x00\xEB"
-     "\x87\xBE\x00\xEB\x87\xBF\x00\xEB\x88\x80\x00\xEB\x88\x81\x00\xEB\x88\x82"
-     "\x00\xEB"
-     "\x88\x83\x00\xEB\x88\x84\x00\xEB\x88\x85\x00\xEB\x88\x86\x00\xEB\x88\x87"
-     "\x00\xEB"
-     "\x88\x88\x00\xEB\x88\x89\x00\xEB\x88\x8A\x00\xEB\x88\x8B\x00\xEB\x88\x8C"
-     "\x00\xEB"
-     "\x88\x8D\x00\xEB\x88\x8E\x00\xEB\x88\x8F\x00\xEB\x88\x90\x00\xEB\x88\x91"
-     "\x00\xEB"
-     "\x88\x92\x00\xEB\x88\x93\x00\xEB\x88\x94\x00\xEB\x88\x95\x00\xEB\x88\x96"
-     "\x00\xEB"
-     "\x88\x97\x00\xEB\x88\x98\x00\xEB\x88\x99\x00\xEB\x88\x9A\x00\xEB\x88\x9B"
-     "\x00\xEB"
-     "\x88\x9C\x00\xEB\x88\x9D\x00\xEB\x88\x9E\x00\xEB\x88\x9F\x00\xEB\x88\xA0"
-     "\x00\xEB"
-     "\x88\xA1\x00\xEB\x88\xA2\x00\xEB\x88\xA3\x00\xEB\x88\xA4\x00\xEB\x88\xA5"
-     "\x00\xEB"
-     "\x88\xA6\x00\xEB\x88\xA7\x00\xEB\x88\xA8\x00\xEB\x88\xA9\x00\xEB\x88\xAA"
-     "\x00\xEB"
-     "\x88\xAB\x00\xEB\x88\xAC\x00\xEB\x88\xAD\x00\xEB\x88\xAE\x00\xEB\x88\xAF"
-     "\x00\xEB"
-     "\x88\xB0\x00\xEB\x88\xB1\x00\xEB\x88\xB2\x00\xEB\x88\xB3\x00\xEB\x88\xB4"
-     "\x00\xEB"
-     "\x88\xB5\x00\xEB\x88\xB6\x00\xEB\x88\xB7\x00\xEB\x88\xB8\x00\xEB\x88\xB9"
-     "\x00\xEB"
-     "\x88\xBA\x00\xEB\x88\xBB\x00\xEB\x88\xBC\x00\xEB\x88\xBD\x00\xEB\x88\xBE"
-     "\x00\xEB"
-     "\x88\xBF\x00\xEB\x89\x80\x00\xEB\x89\x81\x00\xEB\x89\x82\x00\xEB\x89\x83"
-     "\x00\xEB"
-     "\x89\x84\x00\xEB\x89\x85\x00\xEB\x89\x86\x00\xEB\x89\x87\x00\xEB\x89\x88"
-     "\x00\xEB"
-     "\x89\x89\x00\xEB\x89\x8A\x00\xEB\x89\x8B\x00\xEB\x89\x8C\x00\xEB\x89\x8D"
-     "\x00\xEB"
-     "\x89\x8E\x00\xEB\x89\x8F\x00\xEB\x89\x90\x00\xEB\x89\x91\x00\xEB\x89\x92"
-     "\x00\xEB"
-     "\x89\x93\x00\xEB\x89\x94\x00\xEB\x89\x95\x00\xEB\x89\x96\x00\xEB\x89\x97"
-     "\x00\xEB"
-     "\x89\x98\x00\xEB\x89\x99\x00\xEB\x89\x9A\x00\xEB\x89\x9B\x00\xEB\x89\x9C"
-     "\x00\xEB"
-     "\x89\x9D\x00\xEB\x89\x9E\x00\xEB\x89\x9F\x00\xEB\x89\xA0\x00\xEB\x89\xA1"
-     "\x00\xEB"
-     "\x89\xA2\x00\xEB\x89\xA3\x00\xEB\x89\xA4\x00\xEB\x89\xA5\x00\xEB\x89\xA6"
-     "\x00\xEB"
-     "\x89\xA7\x00\xEB\x89\xA8\x00\xEB\x89\xA9\x00\xEB\x89\xAA\x00\xEB\x89\xAB"
-     "\x00\xEB"
-     "\x89\xAC\x00\xEB\x89\xAD\x00\xEB\x89\xAE\x00\xEB\x89\xAF\x00\xEB\x89\xB0"
-     "\x00\xEB"
-     "\x89\xB1\x00\xEB\x89\xB2\x00\xEB\x89\xB3\x00\xEB\x89\xB4\x00\xEB\x89\xB5"
-     "\x00\xEB"
-     "\x89\xB6\x00\xEB\x89\xB7\x00\xEB\x89\xB8\x00\xEB\x89\xB9\x00\xEB\x89\xBA"
-     "\x00\xEB"
-     "\x89\xBB\x00\xEB\x89\xBC\x00\xEB\x89\xBD\x00\xEB\x89\xBE\x00\xEB\x89\xBF"
-     "\x00\xEB"
-     "\x8A\x80\x00\xEB\x8A\x81\x00\xEB\x8A\x82\x00\xEB\x8A\x83\x00\xEB\x8A\x84"
-     "\x00\xEB"
-     "\x8A\x85\x00\xEB\x8A\x86\x00\xEB\x8A\x87\x00\xEB\x8A\x88\x00\xEB\x8A\x89"
-     "\x00\xEB"
-     "\x8A\x8A\x00\xEB\x8A\x8B\x00\xEB\x8A\x8C\x00\xEB\x8A\x8D\x00\xEB\x8A\x8E"
-     "\x00\xEB"
-     "\x8A\x8F\x00\xEB\x8A\x90\x00\xEB\x8A\x91\x00\xEB\x8A\x92\x00\xEB\x8A\x93"
-     "\x00\xEB"
-     "\x8A\x94\x00\xEB\x8A\x95\x00\xEB\x8A\x96\x00\xEB\x8A\x97\x00\xEB\x8A\x98"
-     "\x00\xEB"
-     "\x8A\x99\x00\xEB\x8A\x9A\x00\xEB\x8A\x9B\x00\xEB\x8A\x9C\x00\xEB\x8A\x9D"
-     "\x00\xEB"
-     "\x8A\x9E\x00\xEB\x8A\x9F\x00\xEB\x8A\xA0\x00\xEB\x8A\xA1\x00\xEB\x8A\xA2"
-     "\x00\xEB"
-     "\x8A\xA3\x00\xEB\x8A\xA4\x00\xEB\x8A\xA5\x00\xEB\x8A\xA6\x00\xEB\x8A\xA7"
-     "\x00\xEB"
-     "\x8A\xA8\x00\xEB\x8A\xA9\x00\xEB\x8A\xAA\x00\xEB\x8A\xAB\x00\xEB\x8A\xAC"
-     "\x00\xEB"
-     "\x8A\xAD\x00\xEB\x8A\xAE\x00\xEB\x8A\xAF\x00\xEB\x8A\xB0\x00\xEB\x8A\xB1"
-     "\x00\xEB"
-     "\x8A\xB2\x00\xEB\x8A\xB3\x00\xEB\x8A\xB4\x00\xEB\x8A\xB5\x00\xEB\x8A\xB6"
-     "\x00\xEB"
-     "\x8A\xB7\x00\xEB\x8A\xB8\x00\xEB\x8A\xB9\x00\xEB\x8A\xBA\x00\xEB\x8A\xBB"
-     "\x00\xEB"
-     "\x8A\xBC\x00\xEB\x8A\xBD\x00\xEB\x8A\xBE\x00\xEB\x8A\xBF\x00\xEB\x8B\x80"
-     "\x00\xEB"
-     "\x8B\x81\x00\xEB\x8B\x82\x00\xEB\x8B\x83\x00\xEB\x8B\x84\x00\xEB\x8B\x85"
-     "\x00\xEB"
-     "\x8B\x86\x00\xEB\x8B\x87\x00\xEB\x8B\x88\x00\xEB\x8B\x89\x00\xEB\x8B\x8A"
-     "\x00\xEB"
-     "\x8B\x8B\x00\xEB\x8B\x8C\x00\xEB\x8B\x8D\x00\xEB\x8B\x8E\x00\xEB\x8B\x8F"
-     "\x00\xEB"
-     "\x8B\x90\x00\xEB\x8B\x91\x00\xEB\x8B\x92\x00\xEB\x8B\x93\x00\xEB\x8B\x94"
-     "\x00\xEB"
-     "\x8B\x95\x00\xEB\x8B\x96\x00\xEB\x8B\x97\x00\xEB\x8B\x98\x00\xEB\x8B\x99"
-     "\x00\xEB"
-     "\x8B\x9A\x00\xEB\x8B\x9B\x00\xEB\x8B\x9C\x00\xEB\x8B\x9D\x00\xEB\x8B\x9E"
-     "\x00\xEB"
-     "\x8B\x9F\x00\xEB\x8B\xA0\x00\xEB\x8B\xA1\x00\xEB\x8B\xA2\x00\xEB\x8B\xA3"
-     "\x00\xEB"
-     "\x8B\xA4\x00\xEB\x8B\xA5\x00\xEB\x8B\xA6\x00\xEB\x8B\xA7\x00\xEB\x8B\xA8"
-     "\x00\xEB"
-     "\x8B\xA9\x00\xEB\x8B\xAA\x00\xEB\x8B\xAB\x00\xEB\x8B\xAC\x00\xEB\x8B\xAD"
-     "\x00\xEB"
-     "\x8B\xAE\x00\xEB\x8B\xAF\x00\xEB\x8B\xB0\x00\xEB\x8B\xB1\x00\xEB\x8B\xB2"
-     "\x00\xEB"
-     "\x8B\xB3\x00\xEB\x8B\xB4\x00\xEB\x8B\xB5\x00\xEB\x8B\xB6\x00\xEB\x8B\xB7"
-     "\x00\xEB"
-     "\x8B\xB8\x00\xEB\x8B\xB9\x00\xEB\x8B\xBA\x00\xEB\x8B\xBB\x00\xEB\x8B\xBC"
-     "\x00\xEB"
-     "\x8B\xBD\x00\xEB\x8B\xBE\x00\xEB\x8B\xBF\x00\xEB\x8C\x80\x00\xEB\x8C\x81"
-     "\x00\xEB"
-     "\x8C\x82\x00\xEB\x8C\x83\x00\xEB\x8C\x84\x00\xEB\x8C\x85\x00\xEB\x8C\x86"
-     "\x00\xEB"
-     "\x8C\x87\x00\xEB\x8C\x88\x00\xEB\x8C\x89\x00\xEB\x8C\x8A\x00\xEB\x8C\x8B"
-     "\x00\xEB"
-     "\x8C\x8C\x00\xEB\x8C\x8D\x00\xEB\x8C\x8E\x00\xEB\x8C\x8F\x00\xEB\x8C\x90"
-     "\x00\xEB"
-     "\x8C\x91\x00\xEB\x8C\x92\x00\xEB\x8C\x93\x00\xEB\x8C\x94\x00\xEB\x8C\x95"
-     "\x00\xEB"
-     "\x8C\x96\x00\xEB\x8C\x97\x00\xEB\x8C\x98\x00\xEB\x8C\x99\x00\xEB\x8C\x9A"
-     "\x00\xEB"
-     "\x8C\x9B\x00\xEB\x8C\x9C\x00\xEB\x8C\x9D\x00\xEB\x8C\x9E\x00\xEB\x8C\x9F"
-     "\x00\xEB"
-     "\x8C\xA0\x00\xEB\x8C\xA1\x00\xEB\x8C\xA2\x00\xEB\x8C\xA3\x00\xEB\x8C\xA4"
-     "\x00\xEB"
-     "\x8C\xA5\x00\xEB\x8C\xA6\x00\xEB\x8C\xA7\x00\xEB\x8C\xA8\x00\xEB\x8C\xA9"
-     "\x00\xEB"
-     "\x8C\xAA\x00\xEB\x8C\xAB\x00\xEB\x8C\xAC\x00\xEB\x8C\xAD\x00\xEB\x8C\xAE"
-     "\x00\xEB"
-     "\x8C\xAF\x00\xEB\x8C\xB0\x00\xEB\x8C\xB1\x00\xEB\x8C\xB2\x00\xEB\x8C\xB3"
-     "\x00\xEB"
-     "\x8C\xB4\x00\xEB\x8C\xB5\x00\xEB\x8C\xB6\x00\xEB\x8C\xB7\x00\xEB\x8C\xB8"
-     "\x00\xEB"
-     "\x8C\xB9\x00\xEB\x8C\xBA\x00\xEB\x8C\xBB\x00\xEB\x8C\xBC\x00\xEB\x8C\xBD"
-     "\x00\xEB"
-     "\x8C\xBE\x00\xEB\x8C\xBF\x00\xEB\x8D\x80\x00\xEB\x8D\x81\x00\xEB\x8D\x82"
-     "\x00\xEB"
-     "\x8D\x83\x00\xEB\x8D\x84\x00\xEB\x8D\x85\x00\xEB\x8D\x86\x00\xEB\x8D\x87"
-     "\x00\xEB"
-     "\x8D\x88\x00\xEB\x8D\x89\x00\xEB\x8D\x8A\x00\xEB\x8D\x8B\x00\xEB\x8D\x8C"
-     "\x00\xEB"
-     "\x8D\x8D\x00\xEB\x8D\x8E\x00\xEB\x8D\x8F\x00\xEB\x8D\x90\x00\xEB\x8D\x91"
-     "\x00\xEB"
-     "\x8D\x92\x00\xEB\x8D\x93\x00\xEB\x8D\x94\x00\xEB\x8D\x95\x00\xEB\x8D\x96"
-     "\x00\xEB"
-     "\x8D\x97\x00\xEB\x8D\x98\x00\xEB\x8D\x99\x00\xEB\x8D\x9A\x00\xEB\x8D\x9B"
-     "\x00\xEB"
-     "\x8D\x9C\x00\xEB\x8D\x9D\x00\xEB\x8D\x9E\x00\xEB\x8D\x9F\x00\xEB\x8D\xA0"
-     "\x00\xEB"
-     "\x8D\xA1\x00\xEB\x8D\xA2\x00\xEB\x8D\xA3\x00\xEB\x8D\xA4\x00\xEB\x8D\xA5"
-     "\x00\xEB"
-     "\x8D\xA6\x00\xEB\x8D\xA7\x00\xEB\x8D\xA8\x00\xEB\x8D\xA9\x00\xEB\x8D\xAA"
-     "\x00\xEB"
-     "\x8D\xAB\x00\xEB\x8D\xAC\x00\xEB\x8D\xAD\x00\xEB\x8D\xAE\x00\xEB\x8D\xAF"
-     "\x00\xEB"
-     "\x8D\xB0\x00\xEB\x8D\xB1\x00\xEB\x8D\xB2\x00\xEB\x8D\xB3\x00\xEB\x8D\xB4"
-     "\x00\xEB"
-     "\x8D\xB5\x00\xEB\x8D\xB6\x00\xEB\x8D\xB7\x00\xEB\x8D\xB8\x00\xEB\x8D\xB9"
-     "\x00\xEB"
-     "\x8D\xBA\x00\xEB\x8D\xBB\x00\xEB\x8D\xBC\x00\xEB\x8D\xBD\x00\xEB\x8D\xBE"
-     "\x00\xEB"
-     "\x8D\xBF\x00\xEB\x8E\x80\x00\xEB\x8E\x81\x00\xEB\x8E\x82\x00\xEB\x8E\x83"
-     "\x00\xEB"
-     "\x8E\x84\x00\xEB\x8E\x85\x00\xEB\x8E\x86\x00\xEB\x8E\x87\x00\xEB\x8E\x88"
-     "\x00\xEB"
-     "\x8E\x89\x00\xEB\x8E\x8A\x00\xEB\x8E\x8B\x00\xEB\x8E\x8C\x00\xEB\x8E\x8D"
-     "\x00\xEB"
-     "\x8E\x8E\x00\xEB\x8E\x8F\x00\xEB\x8E\x90\x00\xEB\x8E\x91\x00\xEB\x8E\x92"
-     "\x00\xEB"
-     "\x8E\x93\x00\xEB\x8E\x94\x00\xEB\x8E\x95\x00\xEB\x8E\x96\x00\xEB\x8E\x97"
-     "\x00\xEB"
-     "\x8E\x98\x00\xEB\x8E\x99\x00\xEB\x8E\x9A\x00\xEB\x8E\x9B\x00\xEB\x8E\x9C"
-     "\x00\xEB"
-     "\x8E\x9D\x00\xEB\x8E\x9E\x00\xEB\x8E\x9F\x00\xEB\x8E\xA0\x00\xEB\x8E\xA1"
-     "\x00\xEB"
-     "\x8E\xA2\x00\xEB\x8E\xA3\x00\xEB\x8E\xA4\x00\xEB\x8E\xA5\x00\xEB\x8E\xA6"
-     "\x00\xEB"
-     "\x8E\xA7\x00\xEB\x8E\xA8\x00\xEB\x8E\xA9\x00\xEB\x8E\xAA\x00\xEB\x8E\xAB"
-     "\x00\xEB"
-     "\x8E\xAC\x00\xEB\x8E\xAD\x00\xEB\x8E\xAE\x00\xEB\x8E\xAF\x00\xEB\x8E\xB0"
-     "\x00\xEB"
-     "\x8E\xB1\x00\xEB\x8E\xB2\x00\xEB\x8E\xB3\x00\xEB\x8E\xB4\x00\xEB\x8E\xB5"
-     "\x00\xEB"
-     "\x8E\xB6\x00\xEB\x8E\xB7\x00\xEB\x8E\xB8\x00\xEB\x8E\xB9\x00\xEB\x8E\xBA"
-     "\x00\xEB"
-     "\x8E\xBB\x00\xEB\x8E\xBC\x00\xEB\x8E\xBD\x00\xEB\x8E\xBE\x00\xEB\x8E\xBF"
-     "\x00\xEB"
-     "\x8F\x80\x00\xEB\x8F\x81\x00\xEB\x8F\x82\x00\xEB\x8F\x83\x00\xEB\x8F\x84"
-     "\x00\xEB"
-     "\x8F\x85\x00\xEB\x8F\x86\x00\xEB\x8F\x87\x00\xEB\x8F\x88\x00\xEB\x8F\x89"
-     "\x00\xEB"
-     "\x8F\x8A\x00\xEB\x8F\x8B\x00\xEB\x8F\x8C\x00\xEB\x8F\x8D\x00\xEB\x8F\x8E"
-     "\x00\xEB"
-     "\x8F\x8F\x00\xEB\x8F\x90\x00\xEB\x8F\x91\x00\xEB\x8F\x92\x00\xEB\x8F\x93"
-     "\x00\xEB"
-     "\x8F\x94\x00\xEB\x8F\x95\x00\xEB\x8F\x96\x00\xEB\x8F\x97\x00\xEB\x8F\x98"
-     "\x00\xEB"
-     "\x8F\x99\x00\xEB\x8F\x9A\x00\xEB\x8F\x9B\x00\xEB\x8F\x9C\x00\xEB\x8F\x9D"
-     "\x00\xEB"
-     "\x8F\x9E\x00\xEB\x8F\x9F\x00\xEB\x8F\xA0\x00\xEB\x8F\xA1\x00\xEB\x8F\xA2"
-     "\x00\xEB"
-     "\x8F\xA3\x00\xEB\x8F\xA4\x00\xEB\x8F\xA5\x00\xEB\x8F\xA6\x00\xEB\x8F\xA7"
-     "\x00\xEB"
-     "\x8F\xA8\x00\xEB\x8F\xA9\x00\xEB\x8F\xAA\x00\xEB\x8F\xAB\x00\xEB\x8F\xAC"
-     "\x00\xEB"
-     "\x8F\xAD\x00\xEB\x8F\xAE\x00\xEB\x8F\xAF\x00\xEB\x8F\xB0\x00\xEB\x8F\xB1"
-     "\x00\xEB"
-     "\x8F\xB2\x00\xEB\x8F\xB3\x00\xEB\x8F\xB4\x00\xEB\x8F\xB5\x00\xEB\x8F\xB6"
-     "\x00\xEB"
-     "\x8F\xB7\x00\xEB\x8F\xB8\x00\xEB\x8F\xB9\x00\xEB\x8F\xBA\x00\xEB\x8F\xBB"
-     "\x00\xEB"
-     "\x8F\xBC\x00\xEB\x8F\xBD\x00\xEB\x8F\xBE\x00\xEB\x8F\xBF\x00\xEB\x90\x80"
-     "\x00\xEB"
-     "\x90\x81\x00\xEB\x90\x82\x00\xEB\x90\x83\x00\xEB\x90\x84\x00\xEB\x90\x85"
-     "\x00\xEB"
-     "\x90\x86\x00\xEB\x90\x87\x00\xEB\x90\x88\x00\xEB\x90\x89\x00\xEB\x90\x8A"
-     "\x00\xEB"
-     "\x90\x8B\x00\xEB\x90\x8C\x00\xEB\x90\x8D\x00\xEB\x90\x8E\x00\xEB\x90\x8F"
-     "\x00\xEB"
-     "\x90\x90\x00\xEB\x90\x91\x00\xEB\x90\x92\x00\xEB\x90\x93\x00\xEB\x90\x94"
-     "\x00\xEB"
-     "\x90\x95\x00\xEB\x90\x96\x00\xEB\x90\x97\x00\xEB\x90\x98\x00\xEB\x90\x99"
-     "\x00\xEB"
-     "\x90\x9A\x00\xEB\x90\x9B\x00\xEB\x90\x9C\x00\xEB\x90\x9D\x00\xEB\x90\x9E"
-     "\x00\xEB"
-     "\x90\x9F\x00\xEB\x90\xA0\x00\xEB\x90\xA1\x00\xEB\x90\xA2\x00\xEB\x90\xA3"
-     "\x00\xEB"
-     "\x90\xA4\x00\xEB\x90\xA5\x00\xEB\x90\xA6\x00\xEB\x90\xA7\x00\xEB\x90\xA8"
-     "\x00\xEB"
-     "\x90\xA9\x00\xEB\x90\xAA\x00\xEB\x90\xAB\x00\xEB\x90\xAC\x00\xEB\x90\xAD"
-     "\x00\xEB"
-     "\x90\xAE\x00\xEB\x90\xAF\x00\xEB\x90\xB0\x00\xEB\x90\xB1\x00\xEB\x90\xB2"
-     "\x00\xEB"
-     "\x90\xB3\x00\xEB\x90\xB4\x00\xEB\x90\xB5\x00\xEB\x90\xB6\x00\xEB\x90\xB7"
-     "\x00\xEB"
-     "\x90\xB8\x00\xEB\x90\xB9\x00\xEB\x90\xBA\x00\xEB\x90\xBB\x00\xEB\x90\xBC"
-     "\x00\xEB"
-     "\x90\xBD\x00\xEB\x90\xBE\x00\xEB\x90\xBF\x00\xEB\x91\x80\x00\xEB\x91\x81"
-     "\x00\xEB"
-     "\x91\x82\x00\xEB\x91\x83\x00\xEB\x91\x84\x00\xEB\x91\x85\x00\xEB\x91\x86"
-     "\x00\xEB"
-     "\x91\x87\x00\xEB\x91\x88\x00\xEB\x91\x89\x00\xEB\x91\x8A\x00\xEB\x91\x8B"
-     "\x00\xEB"
-     "\x91\x8C\x00\xEB\x91\x8D\x00\xEB\x91\x8E\x00\xEB\x91\x8F\x00\xEB\x91\x90"
-     "\x00\xEB"
-     "\x91\x91\x00\xEB\x91\x92\x00\xEB\x91\x93\x00\xEB\x91\x94\x00\xEB\x91\x95"
-     "\x00\xEB"
-     "\x91\x96\x00\xEB\x91\x97\x00\xEB\x91\x98\x00\xEB\x91\x99\x00\xEB\x91\x9A"
-     "\x00\xEB"
-     "\x91\x9B\x00\xEB\x91\x9C\x00\xEB\x91\x9D\x00\xEB\x91\x9E\x00\xEB\x91\x9F"
-     "\x00\xEB"
-     "\x91\xA0\x00\xEB\x91\xA1\x00\xEB\x91\xA2\x00\xEB\x91\xA3\x00\xEB\x91\xA4"
-     "\x00\xEB"
-     "\x91\xA5\x00\xEB\x91\xA6\x00\xEB\x91\xA7\x00\xEB\x91\xA8\x00\xEB\x91\xA9"
-     "\x00\xEB"
-     "\x91\xAA\x00\xEB\x91\xAB\x00\xEB\x91\xAC\x00\xEB\x91\xAD\x00\xEB\x91\xAE"
-     "\x00\xEB"
-     "\x91\xAF\x00\xEB\x91\xB0\x00\xEB\x91\xB1\x00\xEB\x91\xB2\x00\xEB\x91\xB3"
-     "\x00\xEB"
-     "\x91\xB4\x00\xEB\x91\xB5\x00\xEB\x91\xB6\x00\xEB\x91\xB7\x00\xEB\x91\xB8"
-     "\x00\xEB"
-     "\x91\xB9\x00\xEB\x91\xBA\x00\xEB\x91\xBB\x00\xEB\x91\xBC\x00\xEB\x91\xBD"
-     "\x00\xEB"
-     "\x91\xBE\x00\xEB\x91\xBF\x00\xEB\x92\x80\x00\xEB\x92\x81\x00\xEB\x92\x82"
-     "\x00\xEB"
-     "\x92\x83\x00\xEB\x92\x84\x00\xEB\x92\x85\x00\xEB\x92\x86\x00\xEB\x92\x87"
-     "\x00\xEB"
-     "\x92\x88\x00\xEB\x92\x89\x00\xEB\x92\x8A\x00\xEB\x92\x8B\x00\xEB\x92\x8C"
-     "\x00\xEB"
-     "\x92\x8D\x00\xEB\x92\x8E\x00\xEB\x92\x8F\x00\xEB\x92\x90\x00\xEB\x92\x91"
-     "\x00\xEB"
-     "\x92\x92\x00\xEB\x92\x93\x00\xEB\x92\x94\x00\xEB\x92\x95\x00\xEB\x92\x96"
-     "\x00\xEB"
-     "\x92\x97\x00\xEB\x92\x98\x00\xEB\x92\x99\x00\xEB\x92\x9A\x00\xEB\x92\x9B"
-     "\x00\xEB"
-     "\x92\x9C\x00\xEB\x92\x9D\x00\xEB\x92\x9E\x00\xEB\x92\x9F\x00\xEB\x92\xA0"
-     "\x00\xEB"
-     "\x92\xA1\x00\xEB\x92\xA2\x00\xEB\x92\xA3\x00\xEB\x92\xA4\x00\xEB\x92\xA5"
-     "\x00\xEB"
-     "\x92\xA6\x00\xEB\x92\xA7\x00\xEB\x92\xA8\x00\xEB\x92\xA9\x00\xEB\x92\xAA"
-     "\x00\xEB"
-     "\x92\xAB\x00\xEB\x92\xAC\x00\xEB\x92\xAD\x00\xEB\x92\xAE\x00\xEB\x92\xAF"
-     "\x00\xEB"
-     "\x92\xB0\x00\xEB\x92\xB1\x00\xEB\x92\xB2\x00\xEB\x92\xB3\x00\xEB\x92\xB4"
-     "\x00\xEB"
-     "\x92\xB5\x00\xEB\x92\xB6\x00\xEB\x92\xB7\x00\xEB\x92\xB8\x00\xEB\x92\xB9"
-     "\x00\xEB"
-     "\x92\xBA\x00\xEB\x92\xBB\x00\xEB\x92\xBC\x00\xEB\x92\xBD\x00\xEB\x92\xBE"
-     "\x00\xEB"
-     "\x92\xBF\x00\xEB\x93\x80\x00\xEB\x93\x81\x00\xEB\x93\x82\x00\xEB\x93\x83"
-     "\x00\xEB"
-     "\x93\x84\x00\xEB\x93\x85\x00\xEB\x93\x86\x00\xEB\x93\x87\x00\xEB\x93\x88"
-     "\x00\xEB"
-     "\x93\x89\x00\xEB\x93\x8A\x00\xEB\x93\x8B\x00\xEB\x93\x8C\x00\xEB\x93\x8D"
-     "\x00\xEB"
-     "\x93\x8E\x00\xEB\x93\x8F\x00\xEB\x93\x90\x00\xEB\x93\x91\x00\xEB\x93\x92"
-     "\x00\xEB"
-     "\x93\x93\x00\xEB\x93\x94\x00\xEB\x93\x95\x00\xEB\x93\x96\x00\xEB\x93\x97"
-     "\x00\xEB"
-     "\x93\x98\x00\xEB\x93\x99\x00\xEB\x93\x9A\x00\xEB\x93\x9B\x00\xEB\x93\x9C"
-     "\x00\xEB"
-     "\x93\x9D\x00\xEB\x93\x9E\x00\xEB\x93\x9F\x00\xEB\x93\xA0\x00\xEB\x93\xA1"
-     "\x00\xEB"
-     "\x93\xA2\x00\xEB\x93\xA3\x00\xEB\x93\xA4\x00\xEB\x93\xA5\x00\xEB\x93\xA6"
-     "\x00\xEB"
-     "\x93\xA7\x00\xEB\x93\xA8\x00\xEB\x93\xA9\x00\xEB\x93\xAA\x00\xEB\x93\xAB"
-     "\x00\xEB"
-     "\x93\xAC\x00\xEB\x93\xAD\x00\xEB\x93\xAE\x00\xEB\x93\xAF\x00\xEB\x93\xB0"
-     "\x00\xEB"
-     "\x93\xB1\x00\xEB\x93\xB2\x00\xEB\x93\xB3\x00\xEB\x93\xB4\x00\xEB\x93\xB5"
-     "\x00\xEB"
-     "\x93\xB6\x00\xEB\x93\xB7\x00\xEB\x93\xB8\x00\xEB\x93\xB9\x00\xEB\x93\xBA"
-     "\x00\xEB"
-     "\x93\xBB\x00\xEB\x93\xBC\x00\xEB\x93\xBD\x00\xEB\x93\xBE\x00\xEB\x93\xBF"
-     "\x00\xEB"
-     "\x94\x80\x00\xEB\x94\x81\x00\xEB\x94\x82\x00\xEB\x94\x83\x00\xEB\x94\x84"
-     "\x00\xEB"
-     "\x94\x85\x00\xEB\x94\x86\x00\xEB\x94\x87\x00\xEB\x94\x88\x00\xEB\x94\x89"
-     "\x00\xEB"
-     "\x94\x8A\x00\xEB\x94\x8B\x00\xEB\x94\x8C\x00\xEB\x94\x8D\x00\xEB\x94\x8E"
-     "\x00\xEB"
-     "\x94\x8F\x00\xEB\x94\x90\x00\xEB\x94\x91\x00\xEB\x94\x92\x00\xEB\x94\x93"
-     "\x00\xEB"
-     "\x94\x94\x00\xEB\x94\x95\x00\xEB\x94\x96\x00\xEB\x94\x97\x00\xEB\x94\x98"
-     "\x00\xEB"
-     "\x94\x99\x00\xEB\x94\x9A\x00\xEB\x94\x9B\x00\xEB\x94\x9C\x00\xEB\x94\x9D"
-     "\x00\xEB"
-     "\x94\x9E\x00\xEB\x94\x9F\x00\xEB\x94\xA0\x00\xEB\x94\xA1\x00\xEB\x94\xA2"
-     "\x00\xEB"
-     "\x94\xA3\x00\xEB\x94\xA4\x00\xEB\x94\xA5\x00\xEB\x94\xA6\x00\xEB\x94\xA7"
-     "\x00\xEB"
-     "\x94\xA8\x00\xEB\x94\xA9\x00\xEB\x94\xAA\x00\xEB\x94\xAB\x00\xEB\x94\xAC"
-     "\x00\xEB"
-     "\x94\xAD\x00\xEB\x94\xAE\x00\xEB\x94\xAF\x00\xEB\x94\xB0\x00\xEB\x94\xB1"
-     "\x00\xEB"
-     "\x94\xB2\x00\xEB\x94\xB3\x00\xEB\x94\xB4\x00\xEB\x94\xB5\x00\xEB\x94\xB6"
-     "\x00\xEB"
-     "\x94\xB7\x00\xEB\x94\xB8\x00\xEB\x94\xB9\x00\xEB\x94\xBA\x00\xEB\x94\xBB"
-     "\x00\xEB"
-     "\x94\xBC\x00\xEB\x94\xBD\x00\xEB\x94\xBE\x00\xEB\x94\xBF\x00\xEB\x95\x80"
-     "\x00\xEB"
-     "\x95\x81\x00\xEB\x95\x82\x00\xEB\x95\x83\x00\xEB\x95\x84\x00\xEB\x95\x85"
-     "\x00\xEB"
-     "\x95\x86\x00\xEB\x95\x87\x00\xEB\x95\x88\x00\xEB\x95\x89\x00\xEB\x95\x8A"
-     "\x00\xEB"
-     "\x95\x8B\x00\xEB\x95\x8C\x00\xEB\x95\x8D\x00\xEB\x95\x8E\x00\xEB\x95\x8F"
-     "\x00\xEB"
-     "\x95\x90\x00\xEB\x95\x91\x00\xEB\x95\x92\x00\xEB\x95\x93\x00\xEB\x95\x94"
-     "\x00\xEB"
-     "\x95\x95\x00\xEB\x95\x96\x00\xEB\x95\x97\x00\xEB\x95\x98\x00\xEB\x95\x99"
-     "\x00\xEB"
-     "\x95\x9A\x00\xEB\x95\x9B\x00\xEB\x95\x9C\x00\xEB\x95\x9D\x00\xEB\x95\x9E"
-     "\x00\xEB"
-     "\x95\x9F\x00\xEB\x95\xA0\x00\xEB\x95\xA1\x00\xEB\x95\xA2\x00\xEB\x95\xA3"
-     "\x00\xEB"
-     "\x95\xA4\x00\xEB\x95\xA5\x00\xEB\x95\xA6\x00\xEB\x95\xA7\x00\xEB\x95\xA8"
-     "\x00\xEB"
-     "\x95\xA9\x00\xEB\x95\xAA\x00\xEB\x95\xAB\x00\xEB\x95\xAC\x00\xEB\x95\xAD"
-     "\x00\xEB"
-     "\x95\xAE\x00\xEB\x95\xAF\x00\xEB\x95\xB0\x00\xEB\x95\xB1\x00\xEB\x95\xB2"
-     "\x00\xEB"
-     "\x95\xB3\x00\xEB\x95\xB4\x00\xEB\x95\xB5\x00\xEB\x95\xB6\x00\xEB\x95\xB7"
-     "\x00\xEB"
-     "\x95\xB8\x00\xEB\x95\xB9\x00\xEB\x95\xBA\x00\xEB\x95\xBB\x00\xEB\x95\xBC"
-     "\x00\xEB"
-     "\x95\xBD\x00\xEB\x95\xBE\x00\xEB\x95\xBF\x00\xEB\x96\x80\x00\xEB\x96\x81"
-     "\x00\xEB"
-     "\x96\x82\x00\xEB\x96\x83\x00\xEB\x96\x84\x00\xEB\x96\x85\x00\xEB\x96\x86"
-     "\x00\xEB"
-     "\x96\x87\x00\xEB\x96\x88\x00\xEB\x96\x89\x00\xEB\x96\x8A\x00\xEB\x96\x8B"
-     "\x00\xEB"
-     "\x96\x8C\x00\xEB\x96\x8D\x00\xEB\x96\x8E\x00\xEB\x96\x8F\x00\xEB\x96\x90"
-     "\x00\xEB"
-     "\x96\x91\x00\xEB\x96\x92\x00\xEB\x96\x93\x00\xEB\x96\x94\x00\xEB\x96\x95"
-     "\x00\xEB"
-     "\x96\x96\x00\xEB\x96\x97\x00\xEB\x96\x98\x00\xEB\x96\x99\x00\xEB\x96\x9A"
-     "\x00\xEB"
-     "\x96\x9B\x00\xEB\x96\x9C\x00\xEB\x96\x9D\x00\xEB\x96\x9E\x00\xEB\x96\x9F"
-     "\x00\xEB"
-     "\x96\xA0\x00\xEB\x96\xA1\x00\xEB\x96\xA2\x00\xEB\x96\xA3\x00\xEB\x96\xA4"
-     "\x00\xEB"
-     "\x96\xA5\x00\xEB\x96\xA6\x00\xEB\x96\xA7\x00\xEB\x96\xA8\x00\xEB\x96\xA9"
-     "\x00\xEB"
-     "\x96\xAA\x00\xEB\x96\xAB\x00\xEB\x96\xAC\x00\xEB\x96\xAD\x00\xEB\x96\xAE"
-     "\x00\xEB"
-     "\x96\xAF\x00\xEB\x96\xB0\x00\xEB\x96\xB1\x00\xEB\x96\xB2\x00\xEB\x96\xB3"
-     "\x00\xEB"
-     "\x96\xB4\x00\xEB\x96\xB5\x00\xEB\x96\xB6\x00\xEB\x96\xB7\x00\xEB\x96\xB8"
-     "\x00\xEB"
-     "\x96\xB9\x00\xEB\x96\xBA\x00\xEB\x96\xBB\x00\xEB\x96\xBC\x00\xEB\x96\xBD"
-     "\x00\xEB"
-     "\x96\xBE\x00\xEB\x96\xBF\x00\xEB\x97\x80\x00\xEB\x97\x81\x00\xEB\x97\x82"
-     "\x00\xEB"
-     "\x97\x83\x00\xEB\x97\x84\x00\xEB\x97\x85\x00\xEB\x97\x86\x00\xEB\x97\x87"
-     "\x00\xEB"
-     "\x97\x88\x00\xEB\x97\x89\x00\xEB\x97\x8A\x00\xEB\x97\x8B\x00\xEB\x97\x8C"
-     "\x00\xEB"
-     "\x97\x8D\x00\xEB\x97\x8E\x00\xEB\x97\x8F\x00\xEB\x97\x90\x00\xEB\x97\x91"
-     "\x00\xEB"
-     "\x97\x92\x00\xEB\x97\x93\x00\xEB\x97\x94\x00\xEB\x97\x95\x00\xEB\x97\x96"
-     "\x00\xEB"
-     "\x97\x97\x00\xEB\x97\x98\x00\xEB\x97\x99\x00\xEB\x97\x9A\x00\xEB\x97\x9B"
-     "\x00\xEB"
-     "\x97\x9C\x00\xEB\x97\x9D\x00\xEB\x97\x9E\x00\xEB\x97\x9F\x00\xEB\x97\xA0"
-     "\x00\xEB"
-     "\x97\xA1\x00\xEB\x97\xA2\x00\xEB\x97\xA3\x00\xEB\x97\xA4\x00\xEB\x97\xA5"
-     "\x00\xEB"
-     "\x97\xA6\x00\xEB\x97\xA7\x00\xEB\x97\xA8\x00\xEB\x97\xA9\x00\xEB\x97\xAA"
-     "\x00\xEB"
-     "\x97\xAB\x00\xEB\x97\xAC\x00\xEB\x97\xAD\x00\xEB\x97\xAE\x00\xEB\x97\xAF"
-     "\x00\xEB"
-     "\x97\xB0\x00\xEB\x97\xB1\x00\xEB\x97\xB2\x00\xEB\x97\xB3\x00\xEB\x97\xB4"
-     "\x00\xEB"
-     "\x97\xB5\x00\xEB\x97\xB6\x00\xEB\x97\xB7\x00\xEB\x97\xB8\x00\xEB\x97\xB9"
-     "\x00\xEB"
-     "\x97\xBA\x00\xEB\x97\xBB\x00\xEB\x97\xBC\x00\xEB\x97\xBD\x00\xEB\x97\xBE"
-     "\x00\xEB"
-     "\x97\xBF\x00\xEB\x98\x80\x00\xEB\x98\x81\x00\xEB\x98\x82\x00\xEB\x98\x83"
-     "\x00\xEB"
-     "\x98\x84\x00\xEB\x98\x85\x00\xEB\x98\x86\x00\xEB\x98\x87\x00\xEB\x98\x88"
-     "\x00\xEB"
-     "\x98\x89\x00\xEB\x98\x8A\x00\xEB\x98\x8B\x00\xEB\x98\x8C\x00\xEB\x98\x8D"
-     "\x00\xEB"
-     "\x98\x8E\x00\xEB\x98\x8F\x00\xEB\x98\x90\x00\xEB\x98\x91\x00\xEB\x98\x92"
-     "\x00\xEB"
-     "\x98\x93\x00\xEB\x98\x94\x00\xEB\x98\x95\x00\xEB\x98\x96\x00\xEB\x98\x97"
-     "\x00\xEB"
-     "\x98\x98\x00\xEB\x98\x99\x00\xEB\x98\x9A\x00\xEB\x98\x9B\x00\xEB\x98\x9C"
-     "\x00\xEB"
-     "\x98\x9D\x00\xEB\x98\x9E\x00\xEB\x98\x9F\x00\xEB\x98\xA0\x00\xEB\x98\xA1"
-     "\x00\xEB"
-     "\x98\xA2\x00\xEB\x98\xA3\x00\xEB\x98\xA4\x00\xEB\x98\xA5\x00\xEB\x98\xA6"
-     "\x00\xEB"
-     "\x98\xA7\x00\xEB\x98\xA8\x00\xEB\x98\xA9\x00\xEB\x98\xAA\x00\xEB\x98\xAB"
-     "\x00\xEB"
-     "\x98\xAC\x00\xEB\x98\xAD\x00\xEB\x98\xAE\x00\xEB\x98\xAF\x00\xEB\x98\xB0"
-     "\x00\xEB"
-     "\x98\xB1\x00\xEB\x98\xB2\x00\xEB\x98\xB3\x00\xEB\x98\xB4\x00\xEB\x98\xB5"
-     "\x00\xEB"
-     "\x98\xB6\x00\xEB\x98\xB7\x00\xEB\x98\xB8\x00\xEB\x98\xB9\x00\xEB\x98\xBA"
-     "\x00\xEB"
-     "\x98\xBB\x00\xEB\x98\xBC\x00\xEB\x98\xBD\x00\xEB\x98\xBE\x00\xEB\x98\xBF"
-     "\x00\xEB"
-     "\x99\x80\x00\xEB\x99\x81\x00\xEB\x99\x82\x00\xEB\x99\x83\x00\xEB\x99\x84"
-     "\x00\xEB"
-     "\x99\x85\x00\xEB\x99\x86\x00\xEB\x99\x87\x00\xEB\x99\x88\x00\xEB\x99\x89"
-     "\x00\xEB"
-     "\x99\x8A\x00\xEB\x99\x8B\x00\xEB\x99\x8C\x00\xEB\x99\x8D\x00\xEB\x99\x8E"
-     "\x00\xEB"
-     "\x99\x8F\x00\xEB\x99\x90\x00\xEB\x99\x91\x00\xEB\x99\x92\x00\xEB\x99\x93"
-     "\x00\xEB"
-     "\x99\x94\x00\xEB\x99\x95\x00\xEB\x99\x96\x00\xEB\x99\x97\x00\xEB\x99\x98"
-     "\x00\xEB"
-     "\x99\x99\x00\xEB\x99\x9A\x00\xEB\x99\x9B\x00\xEB\x99\x9C\x00\xEB\x99\x9D"
-     "\x00\xEB"
-     "\x99\x9E\x00\xEB\x99\x9F\x00\xEB\x99\xA0\x00\xEB\x99\xA1\x00\xEB\x99\xA2"
-     "\x00\xEB"
-     "\x99\xA3\x00\xEB\x99\xA4\x00\xEB\x99\xA5\x00\xEB\x99\xA6\x00\xEB\x99\xA7"
-     "\x00\xEB"
-     "\x99\xA8\x00\xEB\x99\xA9\x00\xEB\x99\xAA\x00\xEB\x99\xAB\x00\xEB\x99\xAC"
-     "\x00\xEB"
-     "\x99\xAD\x00\xEB\x99\xAE\x00\xEB\x99\xAF\x00\xEB\x99\xB0\x00\xEB\x99\xB1"
-     "\x00\xEB"
-     "\x99\xB2\x00\xEB\x99\xB3\x00\xEB\x99\xB4\x00\xEB\x99\xB5\x00\xEB\x99\xB6"
-     "\x00\xEB"
-     "\x99\xB7\x00\xEB\x99\xB8\x00\xEB\x99\xB9\x00\xEB\x99\xBA\x00\xEB\x99\xBB"
-     "\x00\xEB"
-     "\x99\xBC\x00\xEB\x99\xBD\x00\xEB\x99\xBE\x00\xEB\x99\xBF\x00\xEB\x9A\x80"
-     "\x00\xEB"
-     "\x9A\x81\x00\xEB\x9A\x82\x00\xEB\x9A\x83\x00\xEB\x9A\x84\x00\xEB\x9A\x85"
-     "\x00\xEB"
-     "\x9A\x86\x00\xEB\x9A\x87\x00\xEB\x9A\x88\x00\xEB\x9A\x89\x00\xEB\x9A\x8A"
-     "\x00\xEB"
-     "\x9A\x8B\x00\xEB\x9A\x8C\x00\xEB\x9A\x8D\x00\xEB\x9A\x8E\x00\xEB\x9A\x8F"
-     "\x00\xEB"
-     "\x9A\x90\x00\xEB\x9A\x91\x00\xEB\x9A\x92\x00\xEB\x9A\x93\x00\xEB\x9A\x94"
-     "\x00\xEB"
-     "\x9A\x95\x00\xEB\x9A\x96\x00\xEB\x9A\x97\x00\xEB\x9A\x98\x00\xEB\x9A\x99"
-     "\x00\xEB"
-     "\x9A\x9A\x00\xEB\x9A\x9B\x00\xEB\x9A\x9C\x00\xEB\x9A\x9D\x00\xEB\x9A\x9E"
-     "\x00\xEB"
-     "\x9A\x9F\x00\xEB\x9A\xA0\x00\xEB\x9A\xA1\x00\xEB\x9A\xA2\x00\xEB\x9A\xA3"
-     "\x00\xEB"
-     "\x9A\xA4\x00\xEB\x9A\xA5\x00\xEB\x9A\xA6\x00\xEB\x9A\xA7\x00\xEB\x9A\xA8"
-     "\x00\xEB"
-     "\x9A\xA9\x00\xEB\x9A\xAA\x00\xEB\x9A\xAB\x00\xEB\x9A\xAC\x00\xEB\x9A\xAD"
-     "\x00\xEB"
-     "\x9A\xAE\x00\xEB\x9A\xAF\x00\xEB\x9A\xB0\x00\xEB\x9A\xB1\x00\xEB\x9A\xB2"
-     "\x00\xEB"
-     "\x9A\xB3\x00\xEB\x9A\xB4\x00\xEB\x9A\xB5\x00\xEB\x9A\xB6\x00\xEB\x9A\xB7"
-     "\x00\xEB"
-     "\x9A\xB8\x00\xEB\x9A\xB9\x00\xEB\x9A\xBA\x00\xEB\x9A\xBB\x00\xEB\x9A\xBC"
-     "\x00\xEB"
-     "\x9A\xBD\x00\xEB\x9A\xBE\x00\xEB\x9A\xBF\x00\xEB\x9B\x80\x00\xEB\x9B\x81"
-     "\x00\xEB"
-     "\x9B\x82\x00\xEB\x9B\x83\x00\xEB\x9B\x84\x00\xEB\x9B\x85\x00\xEB\x9B\x86"
-     "\x00\xEB"
-     "\x9B\x87\x00\xEB\x9B\x88\x00\xEB\x9B\x89\x00\xEB\x9B\x8A\x00\xEB\x9B\x8B"
-     "\x00\xEB"
-     "\x9B\x8C\x00\xEB\x9B\x8D\x00\xEB\x9B\x8E\x00\xEB\x9B\x8F\x00\xEB\x9B\x90"
-     "\x00\xEB"
-     "\x9B\x91\x00\xEB\x9B\x92\x00\xEB\x9B\x93\x00\xEB\x9B\x94\x00\xEB\x9B\x95"
-     "\x00\xEB"
-     "\x9B\x96\x00\xEB\x9B\x97\x00\xEB\x9B\x98\x00\xEB\x9B\x99\x00\xEB\x9B\x9A"
-     "\x00\xEB"
-     "\x9B\x9B\x00\xEB\x9B\x9C\x00\xEB\x9B\x9D\x00\xEB\x9B\x9E\x00\xEB\x9B\x9F"
-     "\x00\xEB"
-     "\x9B\xA0\x00\xEB\x9B\xA1\x00\xEB\x9B\xA2\x00\xEB\x9B\xA3\x00\xEB\x9B\xA4"
-     "\x00\xEB"
-     "\x9B\xA5\x00\xEB\x9B\xA6\x00\xEB\x9B\xA7\x00\xEB\x9B\xA8\x00\xEB\x9B\xA9"
-     "\x00\xEB"
-     "\x9B\xAA\x00\xEB\x9B\xAB\x00\xEB\x9B\xAC\x00\xEB\x9B\xAD\x00\xEB\x9B\xAE"
-     "\x00\xEB"
-     "\x9B\xAF\x00\xEB\x9B\xB0\x00\xEB\x9B\xB1\x00\xEB\x9B\xB2\x00\xEB\x9B\xB3"
-     "\x00\xEB"
-     "\x9B\xB4\x00\xEB\x9B\xB5\x00\xEB\x9B\xB6\x00\xEB\x9B\xB7\x00\xEB\x9B\xB8"
-     "\x00\xEB"
-     "\x9B\xB9\x00\xEB\x9B\xBA\x00\xEB\x9B\xBB\x00\xEB\x9B\xBC\x00\xEB\x9B\xBD"
-     "\x00\xEB"
-     "\x9B\xBE\x00\xEB\x9B\xBF\x00\xEB\x9C\x80\x00\xEB\x9C\x81\x00\xEB\x9C\x82"
-     "\x00\xEB"
-     "\x9C\x83\x00\xEB\x9C\x84\x00\xEB\x9C\x85\x00\xEB\x9C\x86\x00\xEB\x9C\x87"
-     "\x00\xEB"
-     "\x9C\x88\x00\xEB\x9C\x89\x00\xEB\x9C\x8A\x00\xEB\x9C\x8B\x00\xEB\x9C\x8C"
-     "\x00\xEB"
-     "\x9C\x8D\x00\xEB\x9C\x8E\x00\xEB\x9C\x8F\x00\xEB\x9C\x90\x00\xEB\x9C\x91"
-     "\x00\xEB"
-     "\x9C\x92\x00\xEB\x9C\x93\x00\xEB\x9C\x94\x00\xEB\x9C\x95\x00\xEB\x9C\x96"
-     "\x00\xEB"
-     "\x9C\x97\x00\xEB\x9C\x98\x00\xEB\x9C\x99\x00\xEB\x9C\x9A\x00\xEB\x9C\x9B"
-     "\x00\xEB"
-     "\x9C\x9C\x00\xEB\x9C\x9D\x00\xEB\x9C\x9E\x00\xEB\x9C\x9F\x00\xEB\x9C\xA0"
-     "\x00\xEB"
-     "\x9C\xA1\x00\xEB\x9C\xA2\x00\xEB\x9C\xA3\x00\xEB\x9C\xA4\x00\xEB\x9C\xA5"
-     "\x00\xEB"
-     "\x9C\xA6\x00\xEB\x9C\xA7\x00\xEB\x9C\xA8\x00\xEB\x9C\xA9\x00\xEB\x9C\xAA"
-     "\x00\xEB"
-     "\x9C\xAB\x00\xEB\x9C\xAC\x00\xEB\x9C\xAD\x00\xEB\x9C\xAE\x00\xEB\x9C\xAF"
-     "\x00\xEB"
-     "\x9C\xB0\x00\xEB\x9C\xB1\x00\xEB\x9C\xB2\x00\xEB\x9C\xB3\x00\xEB\x9C\xB4"
-     "\x00\xEB"
-     "\x9C\xB5\x00\xEB\x9C\xB6\x00\xEB\x9C\xB7\x00\xEB\x9C\xB8\x00\xEB\x9C\xB9"
-     "\x00\xEB"
-     "\x9C\xBA\x00\xEB\x9C\xBB\x00\xEB\x9C\xBC\x00\xEB\x9C\xBD\x00\xEB\x9C\xBE"
-     "\x00\xEB"
-     "\x9C\xBF\x00\xEB\x9D\x80\x00\xEB\x9D\x81\x00\xEB\x9D\x82\x00\xEB\x9D\x83"
-     "\x00\xEB"
-     "\x9D\x84\x00\xEB\x9D\x85\x00\xEB\x9D\x86\x00\xEB\x9D\x87\x00\xEB\x9D\x88"
-     "\x00\xEB"
-     "\x9D\x89\x00\xEB\x9D\x8A\x00\xEB\x9D\x8B\x00\xEB\x9D\x8C\x00\xEB\x9D\x8D"
-     "\x00\xEB"
-     "\x9D\x8E\x00\xEB\x9D\x8F\x00\xEB\x9D\x90\x00\xEB\x9D\x91\x00\xEB\x9D\x92"
-     "\x00\xEB"
-     "\x9D\x93\x00\xEB\x9D\x94\x00\xEB\x9D\x95\x00\xEB\x9D\x96\x00\xEB\x9D\x97"
-     "\x00\xEB"
-     "\x9D\x98\x00\xEB\x9D\x99\x00\xEB\x9D\x9A\x00\xEB\x9D\x9B\x00\xEB\x9D\x9C"
-     "\x00\xEB"
-     "\x9D\x9D\x00\xEB\x9D\x9E\x00\xEB\x9D\x9F\x00\xEB\x9D\xA0\x00\xEB\x9D\xA1"
-     "\x00\xEB"
-     "\x9D\xA2\x00\xEB\x9D\xA3\x00\xEB\x9D\xA4\x00\xEB\x9D\xA5\x00\xEB\x9D\xA6"
-     "\x00\xEB"
-     "\x9D\xA7\x00\xEB\x9D\xA8\x00\xEB\x9D\xA9\x00\xEB\x9D\xAA\x00\xEB\x9D\xAB"
-     "\x00\xEB"
-     "\x9D\xAC\x00\xEB\x9D\xAD\x00\xEB\x9D\xAE\x00\xEB\x9D\xAF\x00\xEB\x9D\xB0"
-     "\x00\xEB"
-     "\x9D\xB1\x00\xEB\x9D\xB2\x00\xEB\x9D\xB3\x00\xEB\x9D\xB4\x00\xEB\x9D\xB5"
-     "\x00\xEB"
-     "\x9D\xB6\x00\xEB\x9D\xB7\x00\xEB\x9D\xB8\x00\xEB\x9D\xB9\x00\xEB\x9D\xBA"
-     "\x00\xEB"
-     "\x9D\xBB\x00\xEB\x9D\xBC\x00\xEB\x9D\xBD\x00\xEB\x9D\xBE\x00\xEB\x9D\xBF"
-     "\x00\xEB"
-     "\x9E\x80\x00\xEB\x9E\x81\x00\xEB\x9E\x82\x00\xEB\x9E\x83\x00\xEB\x9E\x84"
-     "\x00\xEB"
-     "\x9E\x85\x00\xEB\x9E\x86\x00\xEB\x9E\x87\x00\xEB\x9E\x88\x00\xEB\x9E\x89"
-     "\x00\xEB"
-     "\x9E\x8A\x00\xEB\x9E\x8B\x00\xEB\x9E\x8C\x00\xEB\x9E\x8D\x00\xEB\x9E\x8E"
-     "\x00\xEB"
-     "\x9E\x8F\x00\xEB\x9E\x90\x00\xEB\x9E\x91\x00\xEB\x9E\x92\x00\xEB\x9E\x93"
-     "\x00\xEB"
-     "\x9E\x94\x00\xEB\x9E\x95\x00\xEB\x9E\x96\x00\xEB\x9E\x97\x00\xEB\x9E\x98"
-     "\x00\xEB"
-     "\x9E\x99\x00\xEB\x9E\x9A\x00\xEB\x9E\x9B\x00\xEB\x9E\x9C\x00\xEB\x9E\x9D"
-     "\x00\xEB"
-     "\x9E\x9E\x00\xEB\x9E\x9F\x00\xEB\x9E\xA0\x00\xEB\x9E\xA1\x00\xEB\x9E\xA2"
-     "\x00\xEB"
-     "\x9E\xA3\x00\xEB\x9E\xA4\x00\xEB\x9E\xA5\x00\xEB\x9E\xA6\x00\xEB\x9E\xA7"
-     "\x00\xEB"
-     "\x9E\xA8\x00\xEB\x9E\xA9\x00\xEB\x9E\xAA\x00\xEB\x9E\xAB\x00\xEB\x9E\xAC"
-     "\x00\xEB"
-     "\x9E\xAD\x00\xEB\x9E\xAE\x00\xEB\x9E\xAF\x00\xEB\x9E\xB0\x00\xEB\x9E\xB1"
-     "\x00\xEB"
-     "\x9E\xB2\x00\xEB\x9E\xB3\x00\xEB\x9E\xB4\x00\xEB\x9E\xB5\x00\xEB\x9E\xB6"
-     "\x00\xEB"
-     "\x9E\xB7\x00\xEB\x9E\xB8\x00\xEB\x9E\xB9\x00\xEB\x9E\xBA\x00\xEB\x9E\xBB"
-     "\x00\xEB"
-     "\x9E\xBC\x00\xEB\x9E\xBD\x00\xEB\x9E\xBE\x00\xEB\x9E\xBF\x00\xEB\x9F\x80"
-     "\x00\xEB"
-     "\x9F\x81\x00\xEB\x9F\x82\x00\xEB\x9F\x83\x00\xEB\x9F\x84\x00\xEB\x9F\x85"
-     "\x00\xEB"
-     "\x9F\x86\x00\xEB\x9F\x87\x00\xEB\x9F\x88\x00\xEB\x9F\x89\x00\xEB\x9F\x8A"
-     "\x00\xEB"
-     "\x9F\x8B\x00\xEB\x9F\x8C\x00\xEB\x9F\x8D\x00\xEB\x9F\x8E\x00\xEB\x9F\x8F"
-     "\x00\xEB"
-     "\x9F\x90\x00\xEB\x9F\x91\x00\xEB\x9F\x92\x00\xEB\x9F\x93\x00\xEB\x9F\x94"
-     "\x00\xEB"
-     "\x9F\x95\x00\xEB\x9F\x96\x00\xEB\x9F\x97\x00\xEB\x9F\x98\x00\xEB\x9F\x99"
-     "\x00\xEB"
-     "\x9F\x9A\x00\xEB\x9F\x9B\x00\xEB\x9F\x9C\x00\xEB\x9F\x9D\x00\xEB\x9F\x9E"
-     "\x00\xEB"
-     "\x9F\x9F\x00\xEB\x9F\xA0\x00\xEB\x9F\xA1\x00\xEB\x9F\xA2\x00\xEB\x9F\xA3"
-     "\x00\xEB"
-     "\x9F\xA4\x00\xEB\x9F\xA5\x00\xEB\x9F\xA6\x00\xEB\x9F\xA7\x00\xEB\x9F\xA8"
-     "\x00\xEB"
-     "\x9F\xA9\x00\xEB\x9F\xAA\x00\xEB\x9F\xAB\x00\xEB\x9F\xAC\x00\xEB\x9F\xAD"
-     "\x00\xEB"
-     "\x9F\xAE\x00\xEB\x9F\xAF\x00\xEB\x9F\xB0\x00\xEB\x9F\xB1\x00\xEB\x9F\xB2"
-     "\x00\xEB"
-     "\x9F\xB3\x00\xEB\x9F\xB4\x00\xEB\x9F\xB5\x00\xEB\x9F\xB6\x00\xEB\x9F\xB7"
-     "\x00\xEB"
-     "\x9F\xB8\x00\xEB\x9F\xB9\x00\xEB\x9F\xBA\x00\xEB\x9F\xBB\x00\xEB\x9F\xBC"
-     "\x00\xEB"
-     "\x9F\xBD\x00\xEB\x9F\xBE\x00\xEB\x9F\xBF\x00\xEB\xA0\x80\x00\xEB\xA0\x81"
-     "\x00\xEB"
-     "\xA0\x82\x00\xEB\xA0\x83\x00\xEB\xA0\x84\x00\xEB\xA0\x85\x00\xEB\xA0\x86"
-     "\x00\xEB"
-     "\xA0\x87\x00\xEB\xA0\x88\x00\xEB\xA0\x89\x00\xEB\xA0\x8A\x00\xEB\xA0\x8B"
-     "\x00\xEB"
-     "\xA0\x8C\x00\xEB\xA0\x8D\x00\xEB\xA0\x8E\x00\xEB\xA0\x8F\x00\xEB\xA0\x90"
-     "\x00\xEB"
-     "\xA0\x91\x00\xEB\xA0\x92\x00\xEB\xA0\x93\x00\xEB\xA0\x94\x00\xEB\xA0\x95"
-     "\x00\xEB"
-     "\xA0\x96\x00\xEB\xA0\x97\x00\xEB\xA0\x98\x00\xEB\xA0\x99\x00\xEB\xA0\x9A"
-     "\x00\xEB"
-     "\xA0\x9B\x00\xEB\xA0\x9C\x00\xEB\xA0\x9D\x00\xEB\xA0\x9E\x00\xEB\xA0\x9F"
-     "\x00\xEB"
-     "\xA0\xA0\x00\xEB\xA0\xA1\x00\xEB\xA0\xA2\x00\xEB\xA0\xA3\x00\xEB\xA0\xA4"
-     "\x00\xEB"
-     "\xA0\xA5\x00\xEB\xA0\xA6\x00\xEB\xA0\xA7\x00\xEB\xA0\xA8\x00\xEB\xA0\xA9"
-     "\x00\xEB"
-     "\xA0\xAA\x00\xEB\xA0\xAB\x00\xEB\xA0\xAC\x00\xEB\xA0\xAD\x00\xEB\xA0\xAE"
-     "\x00\xEB"
-     "\xA0\xAF\x00\xEB\xA0\xB0\x00\xEB\xA0\xB1\x00\xEB\xA0\xB2\x00\xEB\xA0\xB3"
-     "\x00\xEB"
-     "\xA0\xB4\x00\xEB\xA0\xB5\x00\xEB\xA0\xB6\x00\xEB\xA0\xB7\x00\xEB\xA0\xB8"
-     "\x00\xEB"
-     "\xA0\xB9\x00\xEB\xA0\xBA\x00\xEB\xA0\xBB\x00\xEB\xA0\xBC\x00\xEB\xA0\xBD"
-     "\x00\xEB"
-     "\xA0\xBE\x00\xEB\xA0\xBF\x00\xEB\xA1\x80\x00\xEB\xA1\x81\x00\xEB\xA1\x82"
-     "\x00\xEB"
-     "\xA1\x83\x00\xEB\xA1\x84\x00\xEB\xA1\x85\x00\xEB\xA1\x86\x00\xEB\xA1\x87"
-     "\x00\xEB"
-     "\xA1\x88\x00\xEB\xA1\x89\x00\xEB\xA1\x8A\x00\xEB\xA1\x8B\x00\xEB\xA1\x8C"
-     "\x00\xEB"
-     "\xA1\x8D\x00\xEB\xA1\x8E\x00\xEB\xA1\x8F\x00\xEB\xA1\x90\x00\xEB\xA1\x91"
-     "\x00\xEB"
-     "\xA1\x92\x00\xEB\xA1\x93\x00\xEB\xA1\x94\x00\xEB\xA1\x95\x00\xEB\xA1\x96"
-     "\x00\xEB"
-     "\xA1\x97\x00\xEB\xA1\x98\x00\xEB\xA1\x99\x00\xEB\xA1\x9A\x00\xEB\xA1\x9B"
-     "\x00\xEB"
-     "\xA1\x9C\x00\xEB\xA1\x9D\x00\xEB\xA1\x9E\x00\xEB\xA1\x9F\x00\xEB\xA1\xA0"
-     "\x00\xEB"
-     "\xA1\xA1\x00\xEB\xA1\xA2\x00\xEB\xA1\xA3\x00\xEB\xA1\xA4\x00\xEB\xA1\xA5"
-     "\x00\xEB"
-     "\xA1\xA6\x00\xEB\xA1\xA7\x00\xEB\xA1\xA8\x00\xEB\xA1\xA9\x00\xEB\xA1\xAA"
-     "\x00\xEB"
-     "\xA1\xAB\x00\xEB\xA1\xAC\x00\xEB\xA1\xAD\x00\xEB\xA1\xAE\x00\xEB\xA1\xAF"
-     "\x00\xEB"
-     "\xA1\xB0\x00\xEB\xA1\xB1\x00\xEB\xA1\xB2\x00\xEB\xA1\xB3\x00\xEB\xA1\xB4"
-     "\x00\xEB"
-     "\xA1\xB5\x00\xEB\xA1\xB6\x00\xEB\xA1\xB7\x00\xEB\xA1\xB8\x00\xEB\xA1\xB9"
-     "\x00\xEB"
-     "\xA1\xBA\x00\xEB\xA1\xBB\x00\xEB\xA1\xBC\x00\xEB\xA1\xBD\x00\xEB\xA1\xBE"
-     "\x00\xEB"
-     "\xA1\xBF\x00\xEB\xA2\x80\x00\xEB\xA2\x81\x00\xEB\xA2\x82\x00\xEB\xA2\x83"
-     "\x00\xEB"
-     "\xA2\x84\x00\xEB\xA2\x85\x00\xEB\xA2\x86\x00\xEB\xA2\x87\x00\xEB\xA2\x88"
-     "\x00\xEB"
-     "\xA2\x89\x00\xEB\xA2\x8A\x00\xEB\xA2\x8B\x00\xEB\xA2\x8C\x00\xEB\xA2\x8D"
-     "\x00\xEB"
-     "\xA2\x8E\x00\xEB\xA2\x8F\x00\xEB\xA2\x90\x00\xEB\xA2\x91\x00\xEB\xA2\x92"
-     "\x00\xEB"
-     "\xA2\x93\x00\xEB\xA2\x94\x00\xEB\xA2\x95\x00\xEB\xA2\x96\x00\xEB\xA2\x97"
-     "\x00\xEB"
-     "\xA2\x98\x00\xEB\xA2\x99\x00\xEB\xA2\x9A\x00\xEB\xA2\x9B\x00\xEB\xA2\x9C"
-     "\x00\xEB"
-     "\xA2\x9D\x00\xEB\xA2\x9E\x00\xEB\xA2\x9F\x00\xEB\xA2\xA0\x00\xEB\xA2\xA1"
-     "\x00\xEB"
-     "\xA2\xA2\x00\xEB\xA2\xA3\x00\xEB\xA2\xA4\x00\xEB\xA2\xA5\x00\xEB\xA2\xA6"
-     "\x00\xEB"
-     "\xA2\xA7\x00\xEB\xA2\xA8\x00\xEB\xA2\xA9\x00\xEB\xA2\xAA\x00\xEB\xA2\xAB"
-     "\x00\xEB"
-     "\xA2\xAC\x00\xEB\xA2\xAD\x00\xEB\xA2\xAE\x00\xEB\xA2\xAF\x00\xEB\xA2\xB0"
-     "\x00\xEB"
-     "\xA2\xB1\x00\xEB\xA2\xB2\x00\xEB\xA2\xB3\x00\xEB\xA2\xB4\x00\xEB\xA2\xB5"
-     "\x00\xEB"
-     "\xA2\xB6\x00\xEB\xA2\xB7\x00\xEB\xA2\xB8\x00\xEB\xA2\xB9\x00\xEB\xA2\xBA"
-     "\x00\xEB"
-     "\xA2\xBB\x00\xEB\xA2\xBC\x00\xEB\xA2\xBD\x00\xEB\xA2\xBE\x00\xEB\xA2\xBF"
-     "\x00\xEB"
-     "\xA3\x80\x00\xEB\xA3\x81\x00\xEB\xA3\x82\x00\xEB\xA3\x83\x00\xEB\xA3\x84"
-     "\x00\xEB"
-     "\xA3\x85\x00\xEB\xA3\x86\x00\xEB\xA3\x87\x00\xEB\xA3\x88\x00\xEB\xA3\x89"
-     "\x00\xEB"
-     "\xA3\x8A\x00\xEB\xA3\x8B\x00\xEB\xA3\x8C\x00\xEB\xA3\x8D\x00\xEB\xA3\x8E"
-     "\x00\xEB"
-     "\xA3\x8F\x00\xEB\xA3\x90\x00\xEB\xA3\x91\x00\xEB\xA3\x92\x00\xEB\xA3\x93"
-     "\x00\xEB"
-     "\xA3\x94\x00\xEB\xA3\x95\x00\xEB\xA3\x96\x00\xEB\xA3\x97\x00\xEB\xA3\x98"
-     "\x00\xEB"
-     "\xA3\x99\x00\xEB\xA3\x9A\x00\xEB\xA3\x9B\x00\xEB\xA3\x9C\x00\xEB\xA3\x9D"
-     "\x00\xEB"
-     "\xA3\x9E\x00\xEB\xA3\x9F\x00\xEB\xA3\xA0\x00\xEB\xA3\xA1\x00\xEB\xA3\xA2"
-     "\x00\xEB"
-     "\xA3\xA3\x00\xEB\xA3\xA4\x00\xEB\xA3\xA5\x00\xEB\xA3\xA6\x00\xEB\xA3\xA7"
-     "\x00\xEB"
-     "\xA3\xA8\x00\xEB\xA3\xA9\x00\xEB\xA3\xAA\x00\xEB\xA3\xAB\x00\xEB\xA3\xAC"
-     "\x00\xEB"
-     "\xA3\xAD\x00\xEB\xA3\xAE\x00\xEB\xA3\xAF\x00\xEB\xA3\xB0\x00\xEB\xA3\xB1"
-     "\x00\xEB"
-     "\xA3\xB2\x00\xEB\xA3\xB3\x00\xEB\xA3\xB4\x00\xEB\xA3\xB5\x00\xEB\xA3\xB6"
-     "\x00\xEB"
-     "\xA3\xB7\x00\xEB\xA3\xB8\x00\xEB\xA3\xB9\x00\xEB\xA3\xBA\x00\xEB\xA3\xBB"
-     "\x00\xEB"
-     "\xA3\xBC\x00\xEB\xA3\xBD\x00\xEB\xA3\xBE\x00\xEB\xA3\xBF\x00\xEB\xA4\x80"
-     "\x00\xEB"
-     "\xA4\x81\x00\xEB\xA4\x82\x00\xEB\xA4\x83\x00\xEB\xA4\x84\x00\xEB\xA4\x85"
-     "\x00\xEB"
-     "\xA4\x86\x00\xEB\xA4\x87\x00\xEB\xA4\x88\x00\xEB\xA4\x89\x00\xEB\xA4\x8A"
-     "\x00\xEB"
-     "\xA4\x8B\x00\xEB\xA4\x8C\x00\xEB\xA4\x8D\x00\xEB\xA4\x8E\x00\xEB\xA4\x8F"
-     "\x00\xEB"
-     "\xA4\x90\x00\xEB\xA4\x91\x00\xEB\xA4\x92\x00\xEB\xA4\x93\x00\xEB\xA4\x94"
-     "\x00\xEB"
-     "\xA4\x95\x00\xEB\xA4\x96\x00\xEB\xA4\x97\x00\xEB\xA4\x98\x00\xEB\xA4\x99"
-     "\x00\xEB"
-     "\xA4\x9A\x00\xEB\xA4\x9B\x00\xEB\xA4\x9C\x00\xEB\xA4\x9D\x00\xEB\xA4\x9E"
-     "\x00\xEB"
-     "\xA4\x9F\x00\xEB\xA4\xA0\x00\xEB\xA4\xA1\x00\xEB\xA4\xA2\x00\xEB\xA4\xA3"
-     "\x00\xEB"
-     "\xA4\xA4\x00\xEB\xA4\xA5\x00\xEB\xA4\xA6\x00\xEB\xA4\xA7\x00\xEB\xA4\xA8"
-     "\x00\xEB"
-     "\xA4\xA9\x00\xEB\xA4\xAA\x00\xEB\xA4\xAB\x00\xEB\xA4\xAC\x00\xEB\xA4\xAD"
-     "\x00\xEB"
-     "\xA4\xAE\x00\xEB\xA4\xAF\x00\xEB\xA4\xB0\x00\xEB\xA4\xB1\x00\xEB\xA4\xB2"
-     "\x00\xEB"
-     "\xA4\xB3\x00\xEB\xA4\xB4\x00\xEB\xA4\xB5\x00\xEB\xA4\xB6\x00\xEB\xA4\xB7"
-     "\x00\xEB"
-     "\xA4\xB8\x00\xEB\xA4\xB9\x00\xEB\xA4\xBA\x00\xEB\xA4\xBB\x00\xEB\xA4\xBC"
-     "\x00\xEB"
-     "\xA4\xBD\x00\xEB\xA4\xBE\x00\xEB\xA4\xBF\x00\xEB\xA5\x80\x00\xEB\xA5\x81"
-     "\x00\xEB"
-     "\xA5\x82\x00\xEB\xA5\x83\x00\xEB\xA5\x84\x00\xEB\xA5\x85\x00\xEB\xA5\x86"
-     "\x00\xEB"
-     "\xA5\x87\x00\xEB\xA5\x88\x00\xEB\xA5\x89\x00\xEB\xA5\x8A\x00\xEB\xA5\x8B"
-     "\x00\xEB"
-     "\xA5\x8C\x00\xEB\xA5\x8D\x00\xEB\xA5\x8E\x00\xEB\xA5\x8F\x00\xEB\xA5\x90"
-     "\x00\xEB"
-     "\xA5\x91\x00\xEB\xA5\x92\x00\xEB\xA5\x93\x00\xEB\xA5\x94\x00\xEB\xA5\x95"
-     "\x00\xEB"
-     "\xA5\x96\x00\xEB\xA5\x97\x00\xEB\xA5\x98\x00\xEB\xA5\x99\x00\xEB\xA5\x9A"
-     "\x00\xEB"
-     "\xA5\x9B\x00\xEB\xA5\x9C\x00\xEB\xA5\x9D\x00\xEB\xA5\x9E\x00\xEB\xA5\x9F"
-     "\x00\xEB"
-     "\xA5\xA0\x00\xEB\xA5\xA1\x00\xEB\xA5\xA2\x00\xEB\xA5\xA3\x00\xEB\xA5\xA4"
-     "\x00\xEB"
-     "\xA5\xA5\x00\xEB\xA5\xA6\x00\xEB\xA5\xA7\x00\xEB\xA5\xA8\x00\xEB\xA5\xA9"
-     "\x00\xEB"
-     "\xA5\xAA\x00\xEB\xA5\xAB\x00\xEB\xA5\xAC\x00\xEB\xA5\xAD\x00\xEB\xA5\xAE"
-     "\x00\xEB"
-     "\xA5\xAF\x00\xEB\xA5\xB0\x00\xEB\xA5\xB1\x00\xEB\xA5\xB2\x00\xEB\xA5\xB3"
-     "\x00\xEB"
-     "\xA5\xB4\x00\xEB\xA5\xB5\x00\xEB\xA5\xB6\x00\xEB\xA5\xB7\x00\xEB\xA5\xB8"
-     "\x00\xEB"
-     "\xA5\xB9\x00\xEB\xA5\xBA\x00\xEB\xA5\xBB\x00\xEB\xA5\xBC\x00\xEB\xA5\xBD"
-     "\x00\xEB"
-     "\xA5\xBE\x00\xEB\xA5\xBF\x00\xEB\xA6\x80\x00\xEB\xA6\x81\x00\xEB\xA6\x82"
-     "\x00\xEB"
-     "\xA6\x83\x00\xEB\xA6\x84\x00\xEB\xA6\x85\x00\xEB\xA6\x86\x00\xEB\xA6\x87"
-     "\x00\xEB"
-     "\xA6\x88\x00\xEB\xA6\x89\x00\xEB\xA6\x8A\x00\xEB\xA6\x8B\x00\xEB\xA6\x8C"
-     "\x00\xEB"
-     "\xA6\x8D\x00\xEB\xA6\x8E\x00\xEB\xA6\x8F\x00\xEB\xA6\x90\x00\xEB\xA6\x91"
-     "\x00\xEB"
-     "\xA6\x92\x00\xEB\xA6\x93\x00\xEB\xA6\x94\x00\xEB\xA6\x95\x00\xEB\xA6\x96"
-     "\x00\xEB"
-     "\xA6\x97\x00\xEB\xA6\x98\x00\xEB\xA6\x99\x00\xEB\xA6\x9A\x00\xEB\xA6\x9B"
-     "\x00\xEB"
-     "\xA6\x9C\x00\xEB\xA6\x9D\x00\xEB\xA6\x9E\x00\xEB\xA6\x9F\x00\xEB\xA6\xA0"
-     "\x00\xEB"
-     "\xA6\xA1\x00\xEB\xA6\xA2\x00\xEB\xA6\xA3\x00\xEB\xA6\xA4\x00\xEB\xA6\xA5"
-     "\x00\xEB"
-     "\xA6\xA6\x00\xEB\xA6\xA7\x00\xEB\xA6\xA8\x00\xEB\xA6\xA9\x00\xEB\xA6\xAA"
-     "\x00\xEB"
-     "\xA6\xAB\x00\xEB\xA6\xAC\x00\xEB\xA6\xAD\x00\xEB\xA6\xAE\x00\xEB\xA6\xAF"
-     "\x00\xEB"
-     "\xA6\xB0\x00\xEB\xA6\xB1\x00\xEB\xA6\xB2\x00\xEB\xA6\xB3\x00\xEB\xA6\xB4"
-     "\x00\xEB"
-     "\xA6\xB5\x00\xEB\xA6\xB6\x00\xEB\xA6\xB7\x00\xEB\xA6\xB8\x00\xEB\xA6\xB9"
-     "\x00\xEB"
-     "\xA6\xBA\x00\xEB\xA6\xBB\x00\xEB\xA6\xBC\x00\xEB\xA6\xBD\x00\xEB\xA6\xBE"
-     "\x00\xEB"
-     "\xA6\xBF\x00\xEB\xA7\x80\x00\xEB\xA7\x81\x00\xEB\xA7\x82\x00\xEB\xA7\x83"
-     "\x00\xEB"
-     "\xA7\x84\x00\xEB\xA7\x85\x00\xEB\xA7\x86\x00\xEB\xA7\x87\x00\xEB\xA7\x88"
-     "\x00\xEB"
-     "\xA7\x89\x00\xEB\xA7\x8A\x00\xEB\xA7\x8B\x00\xEB\xA7\x8C\x00\xEB\xA7\x8D"
-     "\x00\xEB"
-     "\xA7\x8E\x00\xEB\xA7\x8F\x00\xEB\xA7\x90\x00\xEB\xA7\x91\x00\xEB\xA7\x92"
-     "\x00\xEB"
-     "\xA7\x93\x00\xEB\xA7\x94\x00\xEB\xA7\x95\x00\xEB\xA7\x96\x00\xEB\xA7\x97"
-     "\x00\xEB"
-     "\xA7\x98\x00\xEB\xA7\x99\x00\xEB\xA7\x9A\x00\xEB\xA7\x9B\x00\xEB\xA7\x9C"
-     "\x00\xEB"
-     "\xA7\x9D\x00\xEB\xA7\x9E\x00\xEB\xA7\x9F\x00\xEB\xA7\xA0\x00\xEB\xA7\xA1"
-     "\x00\xEB"
-     "\xA7\xA2\x00\xEB\xA7\xA3\x00\xEB\xA7\xA4\x00\xEB\xA7\xA5\x00\xEB\xA7\xA6"
-     "\x00\xEB"
-     "\xA7\xA7\x00\xEB\xA7\xA8\x00\xEB\xA7\xA9\x00\xEB\xA7\xAA\x00\xEB\xA7\xAB"
-     "\x00\xEB"
-     "\xA7\xAC\x00\xEB\xA7\xAD\x00\xEB\xA7\xAE\x00\xEB\xA7\xAF\x00\xEB\xA7\xB0"
-     "\x00\xEB"
-     "\xA7\xB1\x00\xEB\xA7\xB2\x00\xEB\xA7\xB3\x00\xEB\xA7\xB4\x00\xEB\xA7\xB5"
-     "\x00\xEB"
-     "\xA7\xB6\x00\xEB\xA7\xB7\x00\xEB\xA7\xB8\x00\xEB\xA7\xB9\x00\xEB\xA7\xBA"
-     "\x00\xEB"
-     "\xA7\xBB\x00\xEB\xA7\xBC\x00\xEB\xA7\xBD\x00\xEB\xA7\xBE\x00\xEB\xA7\xBF"
-     "\x00\xEB"
-     "\xA8\x80\x00\xEB\xA8\x81\x00\xEB\xA8\x82\x00\xEB\xA8\x83\x00\xEB\xA8\x84"
-     "\x00\xEB"
-     "\xA8\x85\x00\xEB\xA8\x86\x00\xEB\xA8\x87\x00\xEB\xA8\x88\x00\xEB\xA8\x89"
-     "\x00\xEB"
-     "\xA8\x8A\x00\xEB\xA8\x8B\x00\xEB\xA8\x8C\x00\xEB\xA8\x8D\x00\xEB\xA8\x8E"
-     "\x00\xEB"
-     "\xA8\x8F\x00\xEB\xA8\x90\x00\xEB\xA8\x91\x00\xEB\xA8\x92\x00\xEB\xA8\x93"
-     "\x00\xEB"
-     "\xA8\x94\x00\xEB\xA8\x95\x00\xEB\xA8\x96\x00\xEB\xA8\x97\x00\xEB\xA8\x98"
-     "\x00\xEB"
-     "\xA8\x99\x00\xEB\xA8\x9A\x00\xEB\xA8\x9B\x00\xEB\xA8\x9C\x00\xEB\xA8\x9D"
-     "\x00\xEB"
-     "\xA8\x9E\x00\xEB\xA8\x9F\x00\xEB\xA8\xA0\x00\xEB\xA8\xA1\x00\xEB\xA8\xA2"
-     "\x00\xEB"
-     "\xA8\xA3\x00\xEB\xA8\xA4\x00\xEB\xA8\xA5\x00\xEB\xA8\xA6\x00\xEB\xA8\xA7"
-     "\x00\xEB"
-     "\xA8\xA8\x00\xEB\xA8\xA9\x00\xEB\xA8\xAA\x00\xEB\xA8\xAB\x00\xEB\xA8\xAC"
-     "\x00\xEB"
-     "\xA8\xAD\x00\xEB\xA8\xAE\x00\xEB\xA8\xAF\x00\xEB\xA8\xB0\x00\xEB\xA8\xB1"
-     "\x00\xEB"
-     "\xA8\xB2\x00\xEB\xA8\xB3\x00\xEB\xA8\xB4\x00\xEB\xA8\xB5\x00\xEB\xA8\xB6"
-     "\x00\xEB"
-     "\xA8\xB7\x00\xEB\xA8\xB8\x00\xEB\xA8\xB9\x00\xEB\xA8\xBA\x00\xEB\xA8\xBB"
-     "\x00\xEB"
-     "\xA8\xBC\x00\xEB\xA8\xBD\x00\xEB\xA8\xBE\x00\xEB\xA8\xBF\x00\xEB\xA9\x80"
-     "\x00\xEB"
-     "\xA9\x81\x00\xEB\xA9\x82\x00\xEB\xA9\x83\x00\xEB\xA9\x84\x00\xEB\xA9\x85"
-     "\x00\xEB"
-     "\xA9\x86\x00\xEB\xA9\x87\x00\xEB\xA9\x88\x00\xEB\xA9\x89\x00\xEB\xA9\x8A"
-     "\x00\xEB"
-     "\xA9\x8B\x00\xEB\xA9\x8C\x00\xEB\xA9\x8D\x00\xEB\xA9\x8E\x00\xEB\xA9\x8F"
-     "\x00\xEB"
-     "\xA9\x90\x00\xEB\xA9\x91\x00\xEB\xA9\x92\x00\xEB\xA9\x93\x00\xEB\xA9\x94"
-     "\x00\xEB"
-     "\xA9\x95\x00\xEB\xA9\x96\x00\xEB\xA9\x97\x00\xEB\xA9\x98\x00\xEB\xA9\x99"
-     "\x00\xEB"
-     "\xA9\x9A\x00\xEB\xA9\x9B\x00\xEB\xA9\x9C\x00\xEB\xA9\x9D\x00\xEB\xA9\x9E"
-     "\x00\xEB"
-     "\xA9\x9F\x00\xEB\xA9\xA0\x00\xEB\xA9\xA1\x00\xEB\xA9\xA2\x00\xEB\xA9\xA3"
-     "\x00\xEB"
-     "\xA9\xA4\x00\xEB\xA9\xA5\x00\xEB\xA9\xA6\x00\xEB\xA9\xA7\x00\xEB\xA9\xA8"
-     "\x00\xEB"
-     "\xA9\xA9\x00\xEB\xA9\xAA\x00\xEB\xA9\xAB\x00\xEB\xA9\xAC\x00\xEB\xA9\xAD"
-     "\x00\xEB"
-     "\xA9\xAE\x00\xEB\xA9\xAF\x00\xEB\xA9\xB0\x00\xEB\xA9\xB1\x00\xEB\xA9\xB2"
-     "\x00\xEB"
-     "\xA9\xB3\x00\xEB\xA9\xB4\x00\xEB\xA9\xB5\x00\xEB\xA9\xB6\x00\xEB\xA9\xB7"
-     "\x00\xEB"
-     "\xA9\xB8\x00\xEB\xA9\xB9\x00\xEB\xA9\xBA\x00\xEB\xA9\xBB\x00\xEB\xA9\xBC"
-     "\x00\xEB"
-     "\xA9\xBD\x00\xEB\xA9\xBE\x00\xEB\xA9\xBF\x00\xEB\xAA\x80\x00\xEB\xAA\x81"
-     "\x00\xEB"
-     "\xAA\x82\x00\xEB\xAA\x83\x00\xEB\xAA\x84\x00\xEB\xAA\x85\x00\xEB\xAA\x86"
-     "\x00\xEB"
-     "\xAA\x87\x00\xEB\xAA\x88\x00\xEB\xAA\x89\x00\xEB\xAA\x8A\x00\xEB\xAA\x8B"
-     "\x00\xEB"
-     "\xAA\x8C\x00\xEB\xAA\x8D\x00\xEB\xAA\x8E\x00\xEB\xAA\x8F\x00\xEB\xAA\x90"
-     "\x00\xEB"
-     "\xAA\x91\x00\xEB\xAA\x92\x00\xEB\xAA\x93\x00\xEB\xAA\x94\x00\xEB\xAA\x95"
-     "\x00\xEB"
-     "\xAA\x96\x00\xEB\xAA\x97\x00\xEB\xAA\x98\x00\xEB\xAA\x99\x00\xEB\xAA\x9A"
-     "\x00\xEB"
-     "\xAA\x9B\x00\xEB\xAA\x9C\x00\xEB\xAA\x9D\x00\xEB\xAA\x9E\x00\xEB\xAA\x9F"
-     "\x00\xEB"
-     "\xAA\xA0\x00\xEB\xAA\xA1\x00\xEB\xAA\xA2\x00\xEB\xAA\xA3\x00\xEB\xAA\xA4"
-     "\x00\xEB"
-     "\xAA\xA5\x00\xEB\xAA\xA6\x00\xEB\xAA\xA7\x00\xEB\xAA\xA8\x00\xEB\xAA\xA9"
-     "\x00\xEB"
-     "\xAA\xAA\x00\xEB\xAA\xAB\x00\xEB\xAA\xAC\x00\xEB\xAA\xAD\x00\xEB\xAA\xAE"
-     "\x00\xEB"
-     "\xAA\xAF\x00\xEB\xAA\xB0\x00\xEB\xAA\xB1\x00\xEB\xAA\xB2\x00\xEB\xAA\xB3"
-     "\x00\xEB"
-     "\xAA\xB4\x00\xEB\xAA\xB5\x00\xEB\xAA\xB6\x00\xEB\xAA\xB7\x00\xEB\xAA\xB8"
-     "\x00\xEB"
-     "\xAA\xB9\x00\xEB\xAA\xBA\x00\xEB\xAA\xBB\x00\xEB\xAA\xBC\x00\xEB\xAA\xBD"
-     "\x00\xEB"
-     "\xAA\xBE\x00\xEB\xAA\xBF\x00\xEB\xAB\x80\x00\xEB\xAB\x81\x00\xEB\xAB\x82"
-     "\x00\xEB"
-     "\xAB\x83\x00\xEB\xAB\x84\x00\xEB\xAB\x85\x00\xEB\xAB\x86\x00\xEB\xAB\x87"
-     "\x00\xEB"
-     "\xAB\x88\x00\xEB\xAB\x89\x00\xEB\xAB\x8A\x00\xEB\xAB\x8B\x00\xEB\xAB\x8C"
-     "\x00\xEB"
-     "\xAB\x8D\x00\xEB\xAB\x8E\x00\xEB\xAB\x8F\x00\xEB\xAB\x90\x00\xEB\xAB\x91"
-     "\x00\xEB"
-     "\xAB\x92\x00\xEB\xAB\x93\x00\xEB\xAB\x94\x00\xEB\xAB\x95\x00\xEB\xAB\x96"
-     "\x00\xEB"
-     "\xAB\x97\x00\xEB\xAB\x98\x00\xEB\xAB\x99\x00\xEB\xAB\x9A\x00\xEB\xAB\x9B"
-     "\x00\xEB"
-     "\xAB\x9C\x00\xEB\xAB\x9D\x00\xEB\xAB\x9E\x00\xEB\xAB\x9F\x00\xEB\xAB\xA0"
-     "\x00\xEB"
-     "\xAB\xA1\x00\xEB\xAB\xA2\x00\xEB\xAB\xA3\x00\xEB\xAB\xA4\x00\xEB\xAB\xA5"
-     "\x00\xEB"
-     "\xAB\xA6\x00\xEB\xAB\xA7\x00\xEB\xAB\xA8\x00\xEB\xAB\xA9\x00\xEB\xAB\xAA"
-     "\x00\xEB"
-     "\xAB\xAB\x00\xEB\xAB\xAC\x00\xEB\xAB\xAD\x00\xEB\xAB\xAE\x00\xEB\xAB\xAF"
-     "\x00\xEB"
-     "\xAB\xB0\x00\xEB\xAB\xB1\x00\xEB\xAB\xB2\x00\xEB\xAB\xB3\x00\xEB\xAB\xB4"
-     "\x00\xEB"
-     "\xAB\xB5\x00\xEB\xAB\xB6\x00\xEB\xAB\xB7\x00\xEB\xAB\xB8\x00\xEB\xAB\xB9"
-     "\x00\xEB"
-     "\xAB\xBA\x00\xEB\xAB\xBB\x00\xEB\xAB\xBC\x00\xEB\xAB\xBD\x00\xEB\xAB\xBE"
-     "\x00\xEB"
-     "\xAB\xBF\x00\xEB\xAC\x80\x00\xEB\xAC\x81\x00\xEB\xAC\x82\x00\xEB\xAC\x83"
-     "\x00\xEB"
-     "\xAC\x84\x00\xEB\xAC\x85\x00\xEB\xAC\x86\x00\xEB\xAC\x87\x00\xEB\xAC\x88"
-     "\x00\xEB"
-     "\xAC\x89\x00\xEB\xAC\x8A\x00\xEB\xAC\x8B\x00\xEB\xAC\x8C\x00\xEB\xAC\x8D"
-     "\x00\xEB"
-     "\xAC\x8E\x00\xEB\xAC\x8F\x00\xEB\xAC\x90\x00\xEB\xAC\x91\x00\xEB\xAC\x92"
-     "\x00\xEB"
-     "\xAC\x93\x00\xEB\xAC\x94\x00\xEB\xAC\x95\x00\xEB\xAC\x96\x00\xEB\xAC\x97"
-     "\x00\xEB"
-     "\xAC\x98\x00\xEB\xAC\x99\x00\xEB\xAC\x9A\x00\xEB\xAC\x9B\x00\xEB\xAC\x9C"
-     "\x00\xEB"
-     "\xAC\x9D\x00\xEB\xAC\x9E\x00\xEB\xAC\x9F\x00\xEB\xAC\xA0\x00\xEB\xAC\xA1"
-     "\x00\xEB"
-     "\xAC\xA2\x00\xEB\xAC\xA3\x00\xEB\xAC\xA4\x00\xEB\xAC\xA5\x00\xEB\xAC\xA6"
-     "\x00\xEB"
-     "\xAC\xA7\x00\xEB\xAC\xA8\x00\xEB\xAC\xA9\x00\xEB\xAC\xAA\x00\xEB\xAC\xAB"
-     "\x00\xEB"
-     "\xAC\xAC\x00\xEB\xAC\xAD\x00\xEB\xAC\xAE\x00\xEB\xAC\xAF\x00\xEB\xAC\xB0"
-     "\x00\xEB"
-     "\xAC\xB1\x00\xEB\xAC\xB2\x00\xEB\xAC\xB3\x00\xEB\xAC\xB4\x00\xEB\xAC\xB5"
-     "\x00\xEB"
-     "\xAC\xB6\x00\xEB\xAC\xB7\x00\xEB\xAC\xB8\x00\xEB\xAC\xB9\x00\xEB\xAC\xBA"
-     "\x00\xEB"
-     "\xAC\xBB\x00\xEB\xAC\xBC\x00\xEB\xAC\xBD\x00\xEB\xAC\xBE\x00\xEB\xAC\xBF"
-     "\x00\xEB"
-     "\xAD\x80\x00\xEB\xAD\x81\x00\xEB\xAD\x82\x00\xEB\xAD\x83\x00\xEB\xAD\x84"
-     "\x00\xEB"
-     "\xAD\x85\x00\xEB\xAD\x86\x00\xEB\xAD\x87\x00\xEB\xAD\x88\x00\xEB\xAD\x89"
-     "\x00\xEB"
-     "\xAD\x8A\x00\xEB\xAD\x8B\x00\xEB\xAD\x8C\x00\xEB\xAD\x8D\x00\xEB\xAD\x8E"
-     "\x00\xEB"
-     "\xAD\x8F\x00\xEB\xAD\x90\x00\xEB\xAD\x91\x00\xEB\xAD\x92\x00\xEB\xAD\x93"
-     "\x00\xEB"
-     "\xAD\x94\x00\xEB\xAD\x95\x00\xEB\xAD\x96\x00\xEB\xAD\x97\x00\xEB\xAD\x98"
-     "\x00\xEB"
-     "\xAD\x99\x00\xEB\xAD\x9A\x00\xEB\xAD\x9B\x00\xEB\xAD\x9C\x00\xEB\xAD\x9D"
-     "\x00\xEB"
-     "\xAD\x9E\x00\xEB\xAD\x9F\x00\xEB\xAD\xA0\x00\xEB\xAD\xA1\x00\xEB\xAD\xA2"
-     "\x00\xEB"
-     "\xAD\xA3\x00\xEB\xAD\xA4\x00\xEB\xAD\xA5\x00\xEB\xAD\xA6\x00\xEB\xAD\xA7"
-     "\x00\xEB"
-     "\xAD\xA8\x00\xEB\xAD\xA9\x00\xEB\xAD\xAA\x00\xEB\xAD\xAB\x00\xEB\xAD\xAC"
-     "\x00\xEB"
-     "\xAD\xAD\x00\xEB\xAD\xAE\x00\xEB\xAD\xAF\x00\xEB\xAD\xB0\x00\xEB\xAD\xB1"
-     "\x00\xEB"
-     "\xAD\xB2\x00\xEB\xAD\xB3\x00\xEB\xAD\xB4\x00\xEB\xAD\xB5\x00\xEB\xAD\xB6"
-     "\x00\xEB"
-     "\xAD\xB7\x00\xEB\xAD\xB8\x00\xEB\xAD\xB9\x00\xEB\xAD\xBA\x00\xEB\xAD\xBB"
-     "\x00\xEB"
-     "\xAD\xBC\x00\xEB\xAD\xBD\x00\xEB\xAD\xBE\x00\xEB\xAD\xBF\x00\xEB\xAE\x80"
-     "\x00\xEB"
-     "\xAE\x81\x00\xEB\xAE\x82\x00\xEB\xAE\x83\x00\xEB\xAE\x84\x00\xEB\xAE\x85"
-     "\x00\xEB"
-     "\xAE\x86\x00\xEB\xAE\x87\x00\xEB\xAE\x88\x00\xEB\xAE\x89\x00\xEB\xAE\x8A"
-     "\x00\xEB"
-     "\xAE\x8B\x00\xEB\xAE\x8C\x00\xEB\xAE\x8D\x00\xEB\xAE\x8E\x00\xEB\xAE\x8F"
-     "\x00\xEB"
-     "\xAE\x90\x00\xEB\xAE\x91\x00\xEB\xAE\x92\x00\xEB\xAE\x93\x00\xEB\xAE\x94"
-     "\x00\xEB"
-     "\xAE\x95\x00\xEB\xAE\x96\x00\xEB\xAE\x97\x00\xEB\xAE\x98\x00\xEB\xAE\x99"
-     "\x00\xEB"
-     "\xAE\x9A\x00\xEB\xAE\x9B\x00\xEB\xAE\x9C\x00\xEB\xAE\x9D\x00\xEB\xAE\x9E"
-     "\x00\xEB"
-     "\xAE\x9F\x00\xEB\xAE\xA0\x00\xEB\xAE\xA1\x00\xEB\xAE\xA2\x00\xEB\xAE\xA3"
-     "\x00\xEB"
-     "\xAE\xA4\x00\xEB\xAE\xA5\x00\xEB\xAE\xA6\x00\xEB\xAE\xA7\x00\xEB\xAE\xA8"
-     "\x00\xEB"
-     "\xAE\xA9\x00\xEB\xAE\xAA\x00\xEB\xAE\xAB\x00\xEB\xAE\xAC\x00\xEB\xAE\xAD"
-     "\x00\xEB"
-     "\xAE\xAE\x00\xEB\xAE\xAF\x00\xEB\xAE\xB0\x00\xEB\xAE\xB1\x00\xEB\xAE\xB2"
-     "\x00\xEB"
-     "\xAE\xB3\x00\xEB\xAE\xB4\x00\xEB\xAE\xB5\x00\xEB\xAE\xB6\x00\xEB\xAE\xB7"
-     "\x00\xEB"
-     "\xAE\xB8\x00\xEB\xAE\xB9\x00\xEB\xAE\xBA\x00\xEB\xAE\xBB\x00\xEB\xAE\xBC"
-     "\x00\xEB"
-     "\xAE\xBD\x00\xEB\xAE\xBE\x00\xEB\xAE\xBF\x00\xEB\xAF\x80\x00\xEB\xAF\x81"
-     "\x00\xEB"
-     "\xAF\x82\x00\xEB\xAF\x83\x00\xEB\xAF\x84\x00\xEB\xAF\x85\x00\xEB\xAF\x86"
-     "\x00\xEB"
-     "\xAF\x87\x00\xEB\xAF\x88\x00\xEB\xAF\x89\x00\xEB\xAF\x8A\x00\xEB\xAF\x8B"
-     "\x00\xEB"
-     "\xAF\x8C\x00\xEB\xAF\x8D\x00\xEB\xAF\x8E\x00\xEB\xAF\x8F\x00\xEB\xAF\x90"
-     "\x00\xEB"
-     "\xAF\x91\x00\xEB\xAF\x92\x00\xEB\xAF\x93\x00\xEB\xAF\x94\x00\xEB\xAF\x95"
-     "\x00\xEB"
-     "\xAF\x96\x00\xEB\xAF\x97\x00\xEB\xAF\x98\x00\xEB\xAF\x99\x00\xEB\xAF\x9A"
-     "\x00\xEB"
-     "\xAF\x9B\x00\xEB\xAF\x9C\x00\xEB\xAF\x9D\x00\xEB\xAF\x9E\x00\xEB\xAF\x9F"
-     "\x00\xEB"
-     "\xAF\xA0\x00\xEB\xAF\xA1\x00\xEB\xAF\xA2\x00\xEB\xAF\xA3\x00\xEB\xAF\xA4"
-     "\x00\xEB"
-     "\xAF\xA5\x00\xEB\xAF\xA6\x00\xEB\xAF\xA7\x00\xEB\xAF\xA8\x00\xEB\xAF\xA9"
-     "\x00\xEB"
-     "\xAF\xAA\x00\xEB\xAF\xAB\x00\xEB\xAF\xAC\x00\xEB\xAF\xAD\x00\xEB\xAF\xAE"
-     "\x00\xEB"
-     "\xAF\xAF\x00\xEB\xAF\xB0\x00\xEB\xAF\xB1\x00\xEB\xAF\xB2\x00\xEB\xAF\xB3"
-     "\x00\xEB"
-     "\xAF\xB4\x00\xEB\xAF\xB5\x00\xEB\xAF\xB6\x00\xEB\xAF\xB7\x00\xEB\xAF\xB8"
-     "\x00\xEB"
-     "\xAF\xB9\x00\xEB\xAF\xBA\x00\xEB\xAF\xBB\x00\xEB\xAF\xBC\x00\xEB\xAF\xBD"
-     "\x00\xEB"
-     "\xAF\xBE\x00\xEB\xAF\xBF\x00\xEB\xB0\x80\x00\xEB\xB0\x81\x00\xEB\xB0\x82"
-     "\x00\xEB"
-     "\xB0\x83\x00\xEB\xB0\x84\x00\xEB\xB0\x85\x00\xEB\xB0\x86\x00\xEB\xB0\x87"
-     "\x00\xEB"
-     "\xB0\x88\x00\xEB\xB0\x89\x00\xEB\xB0\x8A\x00\xEB\xB0\x8B\x00\xEB\xB0\x8C"
-     "\x00\xEB"
-     "\xB0\x8D\x00\xEB\xB0\x8E\x00\xEB\xB0\x8F\x00\xEB\xB0\x90\x00\xEB\xB0\x91"
-     "\x00\xEB"
-     "\xB0\x92\x00\xEB\xB0\x93\x00\xEB\xB0\x94\x00\xEB\xB0\x95\x00\xEB\xB0\x96"
-     "\x00\xEB"
-     "\xB0\x97\x00\xEB\xB0\x98\x00\xEB\xB0\x99\x00\xEB\xB0\x9A\x00\xEB\xB0\x9B"
-     "\x00\xEB"
-     "\xB0\x9C\x00\xEB\xB0\x9D\x00\xEB\xB0\x9E\x00\xEB\xB0\x9F\x00\xEB\xB0\xA0"
-     "\x00\xEB"
-     "\xB0\xA1\x00\xEB\xB0\xA2\x00\xEB\xB0\xA3\x00\xEB\xB0\xA4\x00\xEB\xB0\xA5"
-     "\x00\xEB"
-     "\xB0\xA6\x00\xEB\xB0\xA7\x00\xEB\xB0\xA8\x00\xEB\xB0\xA9\x00\xEB\xB0\xAA"
-     "\x00\xEB"
-     "\xB0\xAB\x00\xEB\xB0\xAC\x00\xEB\xB0\xAD\x00\xEB\xB0\xAE\x00\xEB\xB0\xAF"
-     "\x00\xEB"
-     "\xB0\xB0\x00\xEB\xB0\xB1\x00\xEB\xB0\xB2\x00\xEB\xB0\xB3\x00\xEB\xB0\xB4"
-     "\x00\xEB"
-     "\xB0\xB5\x00\xEB\xB0\xB6\x00\xEB\xB0\xB7\x00\xEB\xB0\xB8\x00\xEB\xB0\xB9"
-     "\x00\xEB"
-     "\xB0\xBA\x00\xEB\xB0\xBB\x00\xEB\xB0\xBC\x00\xEB\xB0\xBD\x00\xEB\xB0\xBE"
-     "\x00\xEB"
-     "\xB0\xBF\x00\xEB\xB1\x80\x00\xEB\xB1\x81\x00\xEB\xB1\x82\x00\xEB\xB1\x83"
-     "\x00\xEB"
-     "\xB1\x84\x00\xEB\xB1\x85\x00\xEB\xB1\x86\x00\xEB\xB1\x87\x00\xEB\xB1\x88"
-     "\x00\xEB"
-     "\xB1\x89\x00\xEB\xB1\x8A\x00\xEB\xB1\x8B\x00\xEB\xB1\x8C\x00\xEB\xB1\x8D"
-     "\x00\xEB"
-     "\xB1\x8E\x00\xEB\xB1\x8F\x00\xEB\xB1\x90\x00\xEB\xB1\x91\x00\xEB\xB1\x92"
-     "\x00\xEB"
-     "\xB1\x93\x00\xEB\xB1\x94\x00\xEB\xB1\x95\x00\xEB\xB1\x96\x00\xEB\xB1\x97"
-     "\x00\xEB"
-     "\xB1\x98\x00\xEB\xB1\x99\x00\xEB\xB1\x9A\x00\xEB\xB1\x9B\x00\xEB\xB1\x9C"
-     "\x00\xEB"
-     "\xB1\x9D\x00\xEB\xB1\x9E\x00\xEB\xB1\x9F\x00\xEB\xB1\xA0\x00\xEB\xB1\xA1"
-     "\x00\xEB"
-     "\xB1\xA2\x00\xEB\xB1\xA3\x00\xEB\xB1\xA4\x00\xEB\xB1\xA5\x00\xEB\xB1\xA6"
-     "\x00\xEB"
-     "\xB1\xA7\x00\xEB\xB1\xA8\x00\xEB\xB1\xA9\x00\xEB\xB1\xAA\x00\xEB\xB1\xAB"
-     "\x00\xEB"
-     "\xB1\xAC\x00\xEB\xB1\xAD\x00\xEB\xB1\xAE\x00\xEB\xB1\xAF\x00\xEB\xB1\xB0"
-     "\x00\xEB"
-     "\xB1\xB1\x00\xEB\xB1\xB2\x00\xEB\xB1\xB3\x00\xEB\xB1\xB4\x00\xEB\xB1\xB5"
-     "\x00\xEB"
-     "\xB1\xB6\x00\xEB\xB1\xB7\x00\xEB\xB1\xB8\x00\xEB\xB1\xB9\x00\xEB\xB1\xBA"
-     "\x00\xEB"
-     "\xB1\xBB\x00\xEB\xB1\xBC\x00\xEB\xB1\xBD\x00\xEB\xB1\xBE\x00\xEB\xB1\xBF"
-     "\x00\xEB"
-     "\xB2\x80\x00\xEB\xB2\x81\x00\xEB\xB2\x82\x00\xEB\xB2\x83\x00\xEB\xB2\x84"
-     "\x00\xEB"
-     "\xB2\x85\x00\xEB\xB2\x86\x00\xEB\xB2\x87\x00\xEB\xB2\x88\x00\xEB\xB2\x89"
-     "\x00\xEB"
-     "\xB2\x8A\x00\xEB\xB2\x8B\x00\xEB\xB2\x8C\x00\xEB\xB2\x8D\x00\xEB\xB2\x8E"
-     "\x00\xEB"
-     "\xB2\x8F\x00\xEB\xB2\x90\x00\xEB\xB2\x91\x00\xEB\xB2\x92\x00\xEB\xB2\x93"
-     "\x00\xEB"
-     "\xB2\x94\x00\xEB\xB2\x95\x00\xEB\xB2\x96\x00\xEB\xB2\x97\x00\xEB\xB2\x98"
-     "\x00\xEB"
-     "\xB2\x99\x00\xEB\xB2\x9A\x00\xEB\xB2\x9B\x00\xEB\xB2\x9C\x00\xEB\xB2\x9D"
-     "\x00\xEB"
-     "\xB2\x9E\x00\xEB\xB2\x9F\x00\xEB\xB2\xA0\x00\xEB\xB2\xA1\x00\xEB\xB2\xA2"
-     "\x00\xEB"
-     "\xB2\xA3\x00\xEB\xB2\xA4\x00\xEB\xB2\xA5\x00\xEB\xB2\xA6\x00\xEB\xB2\xA7"
-     "\x00\xEB"
-     "\xB2\xA8\x00\xEB\xB2\xA9\x00\xEB\xB2\xAA\x00\xEB\xB2\xAB\x00\xEB\xB2\xAC"
-     "\x00\xEB"
-     "\xB2\xAD\x00\xEB\xB2\xAE\x00\xEB\xB2\xAF\x00\xEB\xB2\xB0\x00\xEB\xB2\xB1"
-     "\x00\xEB"
-     "\xB2\xB2\x00\xEB\xB2\xB3\x00\xEB\xB2\xB4\x00\xEB\xB2\xB5\x00\xEB\xB2\xB6"
-     "\x00\xEB"
-     "\xB2\xB7\x00\xEB\xB2\xB8\x00\xEB\xB2\xB9\x00\xEB\xB2\xBA\x00\xEB\xB2\xBB"
-     "\x00\xEB"
-     "\xB2\xBC\x00\xEB\xB2\xBD\x00\xEB\xB2\xBE\x00\xEB\xB2\xBF\x00\xEB\xB3\x80"
-     "\x00\xEB"
-     "\xB3\x81\x00\xEB\xB3\x82\x00\xEB\xB3\x83\x00\xEB\xB3\x84\x00\xEB\xB3\x85"
-     "\x00\xEB"
-     "\xB3\x86\x00\xEB\xB3\x87\x00\xEB\xB3\x88\x00\xEB\xB3\x89\x00\xEB\xB3\x8A"
-     "\x00\xEB"
-     "\xB3\x8B\x00\xEB\xB3\x8C\x00\xEB\xB3\x8D\x00\xEB\xB3\x8E\x00\xEB\xB3\x8F"
-     "\x00\xEB"
-     "\xB3\x90\x00\xEB\xB3\x91\x00\xEB\xB3\x92\x00\xEB\xB3\x93\x00\xEB\xB3\x94"
-     "\x00\xEB"
-     "\xB3\x95\x00\xEB\xB3\x96\x00\xEB\xB3\x97\x00\xEB\xB3\x98\x00\xEB\xB3\x99"
-     "\x00\xEB"
-     "\xB3\x9A\x00\xEB\xB3\x9B\x00\xEB\xB3\x9C\x00\xEB\xB3\x9D\x00\xEB\xB3\x9E"
-     "\x00\xEB"
-     "\xB3\x9F\x00\xEB\xB3\xA0\x00\xEB\xB3\xA1\x00\xEB\xB3\xA2\x00\xEB\xB3\xA3"
-     "\x00\xEB"
-     "\xB3\xA4\x00\xEB\xB3\xA5\x00\xEB\xB3\xA6\x00\xEB\xB3\xA7\x00\xEB\xB3\xA8"
-     "\x00\xEB"
-     "\xB3\xA9\x00\xEB\xB3\xAA\x00\xEB\xB3\xAB\x00\xEB\xB3\xAC\x00\xEB\xB3\xAD"
-     "\x00\xEB"
-     "\xB3\xAE\x00\xEB\xB3\xAF\x00\xEB\xB3\xB0\x00\xEB\xB3\xB1\x00\xEB\xB3\xB2"
-     "\x00\xEB"
-     "\xB3\xB3\x00\xEB\xB3\xB4\x00\xEB\xB3\xB5\x00\xEB\xB3\xB6\x00\xEB\xB3\xB7"
-     "\x00\xEB"
-     "\xB3\xB8\x00\xEB\xB3\xB9\x00\xEB\xB3\xBA\x00\xEB\xB3\xBB\x00\xEB\xB3\xBC"
-     "\x00\xEB"
-     "\xB3\xBD\x00\xEB\xB3\xBE\x00\xEB\xB3\xBF\x00\xEB\xB4\x80\x00\xEB\xB4\x81"
-     "\x00\xEB"
-     "\xB4\x82\x00\xEB\xB4\x83\x00\xEB\xB4\x84\x00\xEB\xB4\x85\x00\xEB\xB4\x86"
-     "\x00\xEB"
-     "\xB4\x87\x00\xEB\xB4\x88\x00\xEB\xB4\x89\x00\xEB\xB4\x8A\x00\xEB\xB4\x8B"
-     "\x00\xEB"
-     "\xB4\x8C\x00\xEB\xB4\x8D\x00\xEB\xB4\x8E\x00\xEB\xB4\x8F\x00\xEB\xB4\x90"
-     "\x00\xEB"
-     "\xB4\x91\x00\xEB\xB4\x92\x00\xEB\xB4\x93\x00\xEB\xB4\x94\x00\xEB\xB4\x95"
-     "\x00\xEB"
-     "\xB4\x96\x00\xEB\xB4\x97\x00\xEB\xB4\x98\x00\xEB\xB4\x99\x00\xEB\xB4\x9A"
-     "\x00\xEB"
-     "\xB4\x9B\x00\xEB\xB4\x9C\x00\xEB\xB4\x9D\x00\xEB\xB4\x9E\x00\xEB\xB4\x9F"
-     "\x00\xEB"
-     "\xB4\xA0\x00\xEB\xB4\xA1\x00\xEB\xB4\xA2\x00\xEB\xB4\xA3\x00\xEB\xB4\xA4"
-     "\x00\xEB"
-     "\xB4\xA5\x00\xEB\xB4\xA6\x00\xEB\xB4\xA7\x00\xEB\xB4\xA8\x00\xEB\xB4\xA9"
-     "\x00\xEB"
-     "\xB4\xAA\x00\xEB\xB4\xAB\x00\xEB\xB4\xAC\x00\xEB\xB4\xAD\x00\xEB\xB4\xAE"
-     "\x00\xEB"
-     "\xB4\xAF\x00\xEB\xB4\xB0\x00\xEB\xB4\xB1\x00\xEB\xB4\xB2\x00\xEB\xB4\xB3"
-     "\x00\xEB"
-     "\xB4\xB4\x00\xEB\xB4\xB5\x00\xEB\xB4\xB6\x00\xEB\xB4\xB7\x00\xEB\xB4\xB8"
-     "\x00\xEB"
-     "\xB4\xB9\x00\xEB\xB4\xBA\x00\xEB\xB4\xBB\x00\xEB\xB4\xBC\x00\xEB\xB4\xBD"
-     "\x00\xEB"
-     "\xB4\xBE\x00\xEB\xB4\xBF\x00\xEB\xB5\x80\x00\xEB\xB5\x81\x00\xEB\xB5\x82"
-     "\x00\xEB"
-     "\xB5\x83\x00\xEB\xB5\x84\x00\xEB\xB5\x85\x00\xEB\xB5\x86\x00\xEB\xB5\x87"
-     "\x00\xEB"
-     "\xB5\x88\x00\xEB\xB5\x89\x00\xEB\xB5\x8A\x00\xEB\xB5\x8B\x00\xEB\xB5\x8C"
-     "\x00\xEB"
-     "\xB5\x8D\x00\xEB\xB5\x8E\x00\xEB\xB5\x8F\x00\xEB\xB5\x90\x00\xEB\xB5\x91"
-     "\x00\xEB"
-     "\xB5\x92\x00\xEB\xB5\x93\x00\xEB\xB5\x94\x00\xEB\xB5\x95\x00\xEB\xB5\x96"
-     "\x00\xEB"
-     "\xB5\x97\x00\xEB\xB5\x98\x00\xEB\xB5\x99\x00\xEB\xB5\x9A\x00\xEB\xB5\x9B"
-     "\x00\xEB"
-     "\xB5\x9C\x00\xEB\xB5\x9D\x00\xEB\xB5\x9E\x00\xEB\xB5\x9F\x00\xEB\xB5\xA0"
-     "\x00\xEB"
-     "\xB5\xA1\x00\xEB\xB5\xA2\x00\xEB\xB5\xA3\x00\xEB\xB5\xA4\x00\xEB\xB5\xA5"
-     "\x00\xEB"
-     "\xB5\xA6\x00\xEB\xB5\xA7\x00\xEB\xB5\xA8\x00\xEB\xB5\xA9\x00\xEB\xB5\xAA"
-     "\x00\xEB"
-     "\xB5\xAB\x00\xEB\xB5\xAC\x00\xEB\xB5\xAD\x00\xEB\xB5\xAE\x00\xEB\xB5\xAF"
-     "\x00\xEB"
-     "\xB5\xB0\x00\xEB\xB5\xB1\x00\xEB\xB5\xB2\x00\xEB\xB5\xB3\x00\xEB\xB5\xB4"
-     "\x00\xEB"
-     "\xB5\xB5\x00\xEB\xB5\xB6\x00\xEB\xB5\xB7\x00\xEB\xB5\xB8\x00\xEB\xB5\xB9"
-     "\x00\xEB"
-     "\xB5\xBA\x00\xEB\xB5\xBB\x00\xEB\xB5\xBC\x00\xEB\xB5\xBD\x00\xEB\xB5\xBE"
-     "\x00\xEB"
-     "\xB5\xBF\x00\xEB\xB6\x80\x00\xEB\xB6\x81\x00\xEB\xB6\x82\x00\xEB\xB6\x83"
-     "\x00\xEB"
-     "\xB6\x84\x00\xEB\xB6\x85\x00\xEB\xB6\x86\x00\xEB\xB6\x87\x00\xEB\xB6\x88"
-     "\x00\xEB"
-     "\xB6\x89\x00\xEB\xB6\x8A\x00\xEB\xB6\x8B\x00\xEB\xB6\x8C\x00\xEB\xB6\x8D"
-     "\x00\xEB"
-     "\xB6\x8E\x00\xEB\xB6\x8F\x00\xEB\xB6\x90\x00\xEB\xB6\x91\x00\xEB\xB6\x92"
-     "\x00\xEB"
-     "\xB6\x93\x00\xEB\xB6\x94\x00\xEB\xB6\x95\x00\xEB\xB6\x96\x00\xEB\xB6\x97"
-     "\x00\xEB"
-     "\xB6\x98\x00\xEB\xB6\x99\x00\xEB\xB6\x9A\x00\xEB\xB6\x9B\x00\xEB\xB6\x9C"
-     "\x00\xEB"
-     "\xB6\x9D\x00\xEB\xB6\x9E\x00\xEB\xB6\x9F\x00\xEB\xB6\xA0\x00\xEB\xB6\xA1"
-     "\x00\xEB"
-     "\xB6\xA2\x00\xEB\xB6\xA3\x00\xEB\xB6\xA4\x00\xEB\xB6\xA5\x00\xEB\xB6\xA6"
-     "\x00\xEB"
-     "\xB6\xA7\x00\xEB\xB6\xA8\x00\xEB\xB6\xA9\x00\xEB\xB6\xAA\x00\xEB\xB6\xAB"
-     "\x00\xEB"
-     "\xB6\xAC\x00\xEB\xB6\xAD\x00\xEB\xB6\xAE\x00\xEB\xB6\xAF\x00\xEB\xB6\xB0"
-     "\x00\xEB"
-     "\xB6\xB1\x00\xEB\xB6\xB2\x00\xEB\xB6\xB3\x00\xEB\xB6\xB4\x00\xEB\xB6\xB5"
-     "\x00\xEB"
-     "\xB6\xB6\x00\xEB\xB6\xB7\x00\xEB\xB6\xB8\x00\xEB\xB6\xB9\x00\xEB\xB6\xBA"
-     "\x00\xEB"
-     "\xB6\xBB\x00\xEB\xB6\xBC\x00\xEB\xB6\xBD\x00\xEB\xB6\xBE\x00\xEB\xB6\xBF"
-     "\x00\xEB"
-     "\xB7\x80\x00\xEB\xB7\x81\x00\xEB\xB7\x82\x00\xEB\xB7\x83\x00\xEB\xB7\x84"
-     "\x00\xEB"
-     "\xB7\x85\x00\xEB\xB7\x86\x00\xEB\xB7\x87\x00\xEB\xB7\x88\x00\xEB\xB7\x89"
-     "\x00\xEB"
-     "\xB7\x8A\x00\xEB\xB7\x8B\x00\xEB\xB7\x8C\x00\xEB\xB7\x8D\x00\xEB\xB7\x8E"
-     "\x00\xEB"
-     "\xB7\x8F\x00\xEB\xB7\x90\x00\xEB\xB7\x91\x00\xEB\xB7\x92\x00\xEB\xB7\x93"
-     "\x00\xEB"
-     "\xB7\x94\x00\xEB\xB7\x95\x00\xEB\xB7\x96\x00\xEB\xB7\x97\x00\xEB\xB7\x98"
-     "\x00\xEB"
-     "\xB7\x99\x00\xEB\xB7\x9A\x00\xEB\xB7\x9B\x00\xEB\xB7\x9C\x00\xEB\xB7\x9D"
-     "\x00\xEB"
-     "\xB7\x9E\x00\xEB\xB7\x9F\x00\xEB\xB7\xA0\x00\xEB\xB7\xA1\x00\xEB\xB7\xA2"
-     "\x00\xEB"
-     "\xB7\xA3\x00\xEB\xB7\xA4\x00\xEB\xB7\xA5\x00\xEB\xB7\xA6\x00\xEB\xB7\xA7"
-     "\x00\xEB"
-     "\xB7\xA8\x00\xEB\xB7\xA9\x00\xEB\xB7\xAA\x00\xEB\xB7\xAB\x00\xEB\xB7\xAC"
-     "\x00\xEB"
-     "\xB7\xAD\x00\xEB\xB7\xAE\x00\xEB\xB7\xAF\x00\xEB\xB7\xB0\x00\xEB\xB7\xB1"
-     "\x00\xEB"
-     "\xB7\xB2\x00\xEB\xB7\xB3\x00\xEB\xB7\xB4\x00\xEB\xB7\xB5\x00\xEB\xB7\xB6"
-     "\x00\xEB"
-     "\xB7\xB7\x00\xEB\xB7\xB8\x00\xEB\xB7\xB9\x00\xEB\xB7\xBA\x00\xEB\xB7\xBB"
-     "\x00\xEB"
-     "\xB7\xBC\x00\xEB\xB7\xBD\x00\xEB\xB7\xBE\x00\xEB\xB7\xBF\x00\xEB\xB8\x80"
-     "\x00\xEB"
-     "\xB8\x81\x00\xEB\xB8\x82\x00\xEB\xB8\x83\x00\xEB\xB8\x84\x00\xEB\xB8\x85"
-     "\x00\xEB"
-     "\xB8\x86\x00\xEB\xB8\x87\x00\xEB\xB8\x88\x00\xEB\xB8\x89\x00\xEB\xB8\x8A"
-     "\x00\xEB"
-     "\xB8\x8B\x00\xEB\xB8\x8C\x00\xEB\xB8\x8D\x00\xEB\xB8\x8E\x00\xEB\xB8\x8F"
-     "\x00\xEB"
-     "\xB8\x90\x00\xEB\xB8\x91\x00\xEB\xB8\x92\x00\xEB\xB8\x93\x00\xEB\xB8\x94"
-     "\x00\xEB"
-     "\xB8\x95\x00\xEB\xB8\x96\x00\xEB\xB8\x97\x00\xEB\xB8\x98\x00\xEB\xB8\x99"
-     "\x00\xEB"
-     "\xB8\x9A\x00\xEB\xB8\x9B\x00\xEB\xB8\x9C\x00\xEB\xB8\x9D\x00\xEB\xB8\x9E"
-     "\x00\xEB"
-     "\xB8\x9F\x00\xEB\xB8\xA0\x00\xEB\xB8\xA1\x00\xEB\xB8\xA2\x00\xEB\xB8\xA3"
-     "\x00\xEB"
-     "\xB8\xA4\x00\xEB\xB8\xA5\x00\xEB\xB8\xA6\x00\xEB\xB8\xA7\x00\xEB\xB8\xA8"
-     "\x00\xEB"
-     "\xB8\xA9\x00\xEB\xB8\xAA\x00\xEB\xB8\xAB\x00\xEB\xB8\xAC\x00\xEB\xB8\xAD"
-     "\x00\xEB"
-     "\xB8\xAE\x00\xEB\xB8\xAF\x00\xEB\xB8\xB0\x00\xEB\xB8\xB1\x00\xEB\xB8\xB2"
-     "\x00\xEB"
-     "\xB8\xB3\x00\xEB\xB8\xB4\x00\xEB\xB8\xB5\x00\xEB\xB8\xB6\x00\xEB\xB8\xB7"
-     "\x00\xEB"
-     "\xB8\xB8\x00\xEB\xB8\xB9\x00\xEB\xB8\xBA\x00\xEB\xB8\xBB\x00\xEB\xB8\xBC"
-     "\x00\xEB"
-     "\xB8\xBD\x00\xEB\xB8\xBE\x00\xEB\xB8\xBF\x00\xEB\xB9\x80\x00\xEB\xB9\x81"
-     "\x00\xEB"
-     "\xB9\x82\x00\xEB\xB9\x83\x00\xEB\xB9\x84\x00\xEB\xB9\x85\x00\xEB\xB9\x86"
-     "\x00\xEB"
-     "\xB9\x87\x00\xEB\xB9\x88\x00\xEB\xB9\x89\x00\xEB\xB9\x8A\x00\xEB\xB9\x8B"
-     "\x00\xEB"
-     "\xB9\x8C\x00\xEB\xB9\x8D\x00\xEB\xB9\x8E\x00\xEB\xB9\x8F\x00\xEB\xB9\x90"
-     "\x00\xEB"
-     "\xB9\x91\x00\xEB\xB9\x92\x00\xEB\xB9\x93\x00\xEB\xB9\x94\x00\xEB\xB9\x95"
-     "\x00\xEB"
-     "\xB9\x96\x00\xEB\xB9\x97\x00\xEB\xB9\x98\x00\xEB\xB9\x99\x00\xEB\xB9\x9A"
-     "\x00\xEB"
-     "\xB9\x9B\x00\xEB\xB9\x9C\x00\xEB\xB9\x9D\x00\xEB\xB9\x9E\x00\xEB\xB9\x9F"
-     "\x00\xEB"
-     "\xB9\xA0\x00\xEB\xB9\xA1\x00\xEB\xB9\xA2\x00\xEB\xB9\xA3\x00\xEB\xB9\xA4"
-     "\x00\xEB"
-     "\xB9\xA5\x00\xEB\xB9\xA6\x00\xEB\xB9\xA7\x00\xEB\xB9\xA8\x00\xEB\xB9\xA9"
-     "\x00\xEB"
-     "\xB9\xAA\x00\xEB\xB9\xAB\x00\xEB\xB9\xAC\x00\xEB\xB9\xAD\x00\xEB\xB9\xAE"
-     "\x00\xEB"
-     "\xB9\xAF\x00\xEB\xB9\xB0\x00\xEB\xB9\xB1\x00\xEB\xB9\xB2\x00\xEB\xB9\xB3"
-     "\x00\xEB"
-     "\xB9\xB4\x00\xEB\xB9\xB5\x00\xEB\xB9\xB6\x00\xEB\xB9\xB7\x00\xEB\xB9\xB8"
-     "\x00\xEB"
-     "\xB9\xB9\x00\xEB\xB9\xBA\x00\xEB\xB9\xBB\x00\xEB\xB9\xBC\x00\xEB\xB9\xBD"
-     "\x00\xEB"
-     "\xB9\xBE\x00\xEB\xB9\xBF\x00\xEB\xBA\x80\x00\xEB\xBA\x81\x00\xEB\xBA\x82"
-     "\x00\xEB"
-     "\xBA\x83\x00\xEB\xBA\x84\x00\xEB\xBA\x85\x00\xEB\xBA\x86\x00\xEB\xBA\x87"
-     "\x00\xEB"
-     "\xBA\x88\x00\xEB\xBA\x89\x00\xEB\xBA\x8A\x00\xEB\xBA\x8B\x00\xEB\xBA\x8C"
-     "\x00\xEB"
-     "\xBA\x8D\x00\xEB\xBA\x8E\x00\xEB\xBA\x8F\x00\xEB\xBA\x90\x00\xEB\xBA\x91"
-     "\x00\xEB"
-     "\xBA\x92\x00\xEB\xBA\x93\x00\xEB\xBA\x94\x00\xEB\xBA\x95\x00\xEB\xBA\x96"
-     "\x00\xEB"
-     "\xBA\x97\x00\xEB\xBA\x98\x00\xEB\xBA\x99\x00\xEB\xBA\x9A\x00\xEB\xBA\x9B"
-     "\x00\xEB"
-     "\xBA\x9C\x00\xEB\xBA\x9D\x00\xEB\xBA\x9E\x00\xEB\xBA\x9F\x00\xEB\xBA\xA0"
-     "\x00\xEB"
-     "\xBA\xA1\x00\xEB\xBA\xA2\x00\xEB\xBA\xA3\x00\xEB\xBA\xA4\x00\xEB\xBA\xA5"
-     "\x00\xEB"
-     "\xBA\xA6\x00\xEB\xBA\xA7\x00\xEB\xBA\xA8\x00\xEB\xBA\xA9\x00\xEB\xBA\xAA"
-     "\x00\xEB"
-     "\xBA\xAB\x00\xEB\xBA\xAC\x00\xEB\xBA\xAD\x00\xEB\xBA\xAE\x00\xEB\xBA\xAF"
-     "\x00\xEB"
-     "\xBA\xB0\x00\xEB\xBA\xB1\x00\xEB\xBA\xB2\x00\xEB\xBA\xB3\x00\xEB\xBA\xB4"
-     "\x00\xEB"
-     "\xBA\xB5\x00\xEB\xBA\xB6\x00\xEB\xBA\xB7\x00\xEB\xBA\xB8\x00\xEB\xBA\xB9"
-     "\x00\xEB"
-     "\xBA\xBA\x00\xEB\xBA\xBB\x00\xEB\xBA\xBC\x00\xEB\xBA\xBD\x00\xEB\xBA\xBE"
-     "\x00\xEB"
-     "\xBA\xBF\x00\xEB\xBB\x80\x00\xEB\xBB\x81\x00\xEB\xBB\x82\x00\xEB\xBB\x83"
-     "\x00\xEB"
-     "\xBB\x84\x00\xEB\xBB\x85\x00\xEB\xBB\x86\x00\xEB\xBB\x87\x00\xEB\xBB\x88"
-     "\x00\xEB"
-     "\xBB\x89\x00\xEB\xBB\x8A\x00\xEB\xBB\x8B\x00\xEB\xBB\x8C\x00\xEB\xBB\x8D"
-     "\x00\xEB"
-     "\xBB\x8E\x00\xEB\xBB\x8F\x00\xEB\xBB\x90\x00\xEB\xBB\x91\x00\xEB\xBB\x92"
-     "\x00\xEB"
-     "\xBB\x93\x00\xEB\xBB\x94\x00\xEB\xBB\x95\x00\xEB\xBB\x96\x00\xEB\xBB\x97"
-     "\x00\xEB"
-     "\xBB\x98\x00\xEB\xBB\x99\x00\xEB\xBB\x9A\x00\xEB\xBB\x9B\x00\xEB\xBB\x9C"
-     "\x00\xEB"
-     "\xBB\x9D\x00\xEB\xBB\x9E\x00\xEB\xBB\x9F\x00\xEB\xBB\xA0\x00\xEB\xBB\xA1"
-     "\x00\xEB"
-     "\xBB\xA2\x00\xEB\xBB\xA3\x00\xEB\xBB\xA4\x00\xEB\xBB\xA5\x00\xEB\xBB\xA6"
-     "\x00\xEB"
-     "\xBB\xA7\x00\xEB\xBB\xA8\x00\xEB\xBB\xA9\x00\xEB\xBB\xAA\x00\xEB\xBB\xAB"
-     "\x00\xEB"
-     "\xBB\xAC\x00\xEB\xBB\xAD\x00\xEB\xBB\xAE\x00\xEB\xBB\xAF\x00\xEB\xBB\xB0"
-     "\x00\xEB"
-     "\xBB\xB1\x00\xEB\xBB\xB2\x00\xEB\xBB\xB3\x00\xEB\xBB\xB4\x00\xEB\xBB\xB5"
-     "\x00\xEB"
-     "\xBB\xB6\x00\xEB\xBB\xB7\x00\xEB\xBB\xB8\x00\xEB\xBB\xB9\x00\xEB\xBB\xBA"
-     "\x00\xEB"
-     "\xBB\xBB\x00\xEB\xBB\xBC\x00\xEB\xBB\xBD\x00\xEB\xBB\xBE\x00\xEB\xBB\xBF"
-     "\x00\xEB"
-     "\xBC\x80\x00\xEB\xBC\x81\x00\xEB\xBC\x82\x00\xEB\xBC\x83\x00\xEB\xBC\x84"
-     "\x00\xEB"
-     "\xBC\x85\x00\xEB\xBC\x86\x00\xEB\xBC\x87\x00\xEB\xBC\x88\x00\xEB\xBC\x89"
-     "\x00\xEB"
-     "\xBC\x8A\x00\xEB\xBC\x8B\x00\xEB\xBC\x8C\x00\xEB\xBC\x8D\x00\xEB\xBC\x8E"
-     "\x00\xEB"
-     "\xBC\x8F\x00\xEB\xBC\x90\x00\xEB\xBC\x91\x00\xEB\xBC\x92\x00\xEB\xBC\x93"
-     "\x00\xEB"
-     "\xBC\x94\x00\xEB\xBC\x95\x00\xEB\xBC\x96\x00\xEB\xBC\x97\x00\xEB\xBC\x98"
-     "\x00\xEB"
-     "\xBC\x99\x00\xEB\xBC\x9A\x00\xEB\xBC\x9B\x00\xEB\xBC\x9C\x00\xEB\xBC\x9D"
-     "\x00\xEB"
-     "\xBC\x9E\x00\xEB\xBC\x9F\x00\xEB\xBC\xA0\x00\xEB\xBC\xA1\x00\xEB\xBC\xA2"
-     "\x00\xEB"
-     "\xBC\xA3\x00\xEB\xBC\xA4\x00\xEB\xBC\xA5\x00\xEB\xBC\xA6\x00\xEB\xBC\xA7"
-     "\x00\xEB"
-     "\xBC\xA8\x00\xEB\xBC\xA9\x00\xEB\xBC\xAA\x00\xEB\xBC\xAB\x00\xEB\xBC\xAC"
-     "\x00\xEB"
-     "\xBC\xAD\x00\xEB\xBC\xAE\x00\xEB\xBC\xAF\x00\xEB\xBC\xB0\x00\xEB\xBC\xB1"
-     "\x00\xEB"
-     "\xBC\xB2\x00\xEB\xBC\xB3\x00\xEB\xBC\xB4\x00\xEB\xBC\xB5\x00\xEB\xBC\xB6"
-     "\x00\xEB"
-     "\xBC\xB7\x00\xEB\xBC\xB8\x00\xEB\xBC\xB9\x00\xEB\xBC\xBA\x00\xEB\xBC\xBB"
-     "\x00\xEB"
-     "\xBC\xBC\x00\xEB\xBC\xBD\x00\xEB\xBC\xBE\x00\xEB\xBC\xBF\x00\xEB\xBD\x80"
-     "\x00\xEB"
-     "\xBD\x81\x00\xEB\xBD\x82\x00\xEB\xBD\x83\x00\xEB\xBD\x84\x00\xEB\xBD\x85"
-     "\x00\xEB"
-     "\xBD\x86\x00\xEB\xBD\x87\x00\xEB\xBD\x88\x00\xEB\xBD\x89\x00\xEB\xBD\x8A"
-     "\x00\xEB"
-     "\xBD\x8B\x00\xEB\xBD\x8C\x00\xEB\xBD\x8D\x00\xEB\xBD\x8E\x00\xEB\xBD\x8F"
-     "\x00\xEB"
-     "\xBD\x90\x00\xEB\xBD\x91\x00\xEB\xBD\x92\x00\xEB\xBD\x93\x00\xEB\xBD\x94"
-     "\x00\xEB"
-     "\xBD\x95\x00\xEB\xBD\x96\x00\xEB\xBD\x97\x00\xEB\xBD\x98\x00\xEB\xBD\x99"
-     "\x00\xEB"
-     "\xBD\x9A\x00\xEB\xBD\x9B\x00\xEB\xBD\x9C\x00\xEB\xBD\x9D\x00\xEB\xBD\x9E"
-     "\x00\xEB"
-     "\xBD\x9F\x00\xEB\xBD\xA0\x00\xEB\xBD\xA1\x00\xEB\xBD\xA2\x00\xEB\xBD\xA3"
-     "\x00\xEB"
-     "\xBD\xA4\x00\xEB\xBD\xA5\x00\xEB\xBD\xA6\x00\xEB\xBD\xA7\x00\xEB\xBD\xA8"
-     "\x00\xEB"
-     "\xBD\xA9\x00\xEB\xBD\xAA\x00\xEB\xBD\xAB\x00\xEB\xBD\xAC\x00\xEB\xBD\xAD"
-     "\x00\xEB"
-     "\xBD\xAE\x00\xEB\xBD\xAF\x00\xEB\xBD\xB0\x00\xEB\xBD\xB1\x00\xEB\xBD\xB2"
-     "\x00\xEB"
-     "\xBD\xB3\x00\xEB\xBD\xB4\x00\xEB\xBD\xB5\x00\xEB\xBD\xB6\x00\xEB\xBD\xB7"
-     "\x00\xEB"
-     "\xBD\xB8\x00\xEB\xBD\xB9\x00\xEB\xBD\xBA\x00\xEB\xBD\xBB\x00\xEB\xBD\xBC"
-     "\x00\xEB"
-     "\xBD\xBD\x00\xEB\xBD\xBE\x00\xEB\xBD\xBF\x00\xEB\xBE\x80\x00\xEB\xBE\x81"
-     "\x00\xEB"
-     "\xBE\x82\x00\xEB\xBE\x83\x00\xEB\xBE\x84\x00\xEB\xBE\x85\x00\xEB\xBE\x86"
-     "\x00\xEB"
-     "\xBE\x87\x00\xEB\xBE\x88\x00\xEB\xBE\x89\x00\xEB\xBE\x8A\x00\xEB\xBE\x8B"
-     "\x00\xEB"
-     "\xBE\x8C\x00\xEB\xBE\x8D\x00\xEB\xBE\x8E\x00\xEB\xBE\x8F\x00\xEB\xBE\x90"
-     "\x00\xEB"
-     "\xBE\x91\x00\xEB\xBE\x92\x00\xEB\xBE\x93\x00\xEB\xBE\x94\x00\xEB\xBE\x95"
-     "\x00\xEB"
-     "\xBE\x96\x00\xEB\xBE\x97\x00\xEB\xBE\x98\x00\xEB\xBE\x99\x00\xEB\xBE\x9A"
-     "\x00\xEB"
-     "\xBE\x9B\x00\xEB\xBE\x9C\x00\xEB\xBE\x9D\x00\xEB\xBE\x9E\x00\xEB\xBE\x9F"
-     "\x00\xEB"
-     "\xBE\xA0\x00\xEB\xBE\xA1\x00\xEB\xBE\xA2\x00\xEB\xBE\xA3\x00\xEB\xBE\xA4"
-     "\x00\xEB"
-     "\xBE\xA5\x00\xEB\xBE\xA6\x00\xEB\xBE\xA7\x00\xEB\xBE\xA8\x00\xEB\xBE\xA9"
-     "\x00\xEB"
-     "\xBE\xAA\x00\xEB\xBE\xAB\x00\xEB\xBE\xAC\x00\xEB\xBE\xAD\x00\xEB\xBE\xAE"
-     "\x00\xEB"
-     "\xBE\xAF\x00\xEB\xBE\xB0\x00\xEB\xBE\xB1\x00\xEB\xBE\xB2\x00\xEB\xBE\xB3"
-     "\x00\xEB"
-     "\xBE\xB4\x00\xEB\xBE\xB5\x00\xEB\xBE\xB6\x00\xEB\xBE\xB7\x00\xEB\xBE\xB8"
-     "\x00\xEB"
-     "\xBE\xB9\x00\xEB\xBE\xBA\x00\xEB\xBE\xBB\x00\xEB\xBE\xBC\x00\xEB\xBE\xBD"
-     "\x00\xEB"
-     "\xBE\xBE\x00\xEB\xBE\xBF\x00\xEB\xBF\x80\x00\xEB\xBF\x81\x00\xEB\xBF\x82"
-     "\x00\xEB"
-     "\xBF\x83\x00\xEB\xBF\x84\x00\xEB\xBF\x85\x00\xEB\xBF\x86\x00\xEB\xBF\x87"
-     "\x00\xEB"
-     "\xBF\x88\x00\xEB\xBF\x89\x00\xEB\xBF\x8A\x00\xEB\xBF\x8B\x00\xEB\xBF\x8C"
-     "\x00\xEB"
-     "\xBF\x8D\x00\xEB\xBF\x8E\x00\xEB\xBF\x8F\x00\xEB\xBF\x90\x00\xEB\xBF\x91"
-     "\x00\xEB"
-     "\xBF\x92\x00\xEB\xBF\x93\x00\xEB\xBF\x94\x00\xEB\xBF\x95\x00\xEB\xBF\x96"
-     "\x00\xEB"
-     "\xBF\x97\x00\xEB\xBF\x98\x00\xEB\xBF\x99\x00\xEB\xBF\x9A\x00\xEB\xBF\x9B"
-     "\x00\xEB"
-     "\xBF\x9C\x00\xEB\xBF\x9D\x00\xEB\xBF\x9E\x00\xEB\xBF\x9F\x00\xEB\xBF\xA0"
-     "\x00\xEB"
-     "\xBF\xA1\x00\xEB\xBF\xA2\x00\xEB\xBF\xA3\x00\xEB\xBF\xA4\x00\xEB\xBF\xA5"
-     "\x00\xEB"
-     "\xBF\xA6\x00\xEB\xBF\xA7\x00\xEB\xBF\xA8\x00\xEB\xBF\xA9\x00\xEB\xBF\xAA"
-     "\x00\xEB"
-     "\xBF\xAB\x00\xEB\xBF\xAC\x00\xEB\xBF\xAD\x00\xEB\xBF\xAE\x00\xEB\xBF\xAF"
-     "\x00\xEB"
-     "\xBF\xB0\x00\xEB\xBF\xB1\x00\xEB\xBF\xB2\x00\xEB\xBF\xB3\x00\xEB\xBF\xB4"
-     "\x00\xEB"
-     "\xBF\xB5\x00\xEB\xBF\xB6\x00\xEB\xBF\xB7\x00\xEB\xBF\xB8\x00\xEB\xBF\xB9"
-     "\x00\xEB"
-     "\xBF\xBA\x00\xEB\xBF\xBB\x00\xEB\xBF\xBC\x00\xEB\xBF\xBD\x00\xEB\xBF\xBE"
-     "\x00\xEB"
-     "\xBF\xBF\x00\xEC\x80\x80\x00\xEC\x80\x81\x00\xEC\x80\x82\x00\xEC\x80\x83"
-     "\x00\xEC"
-     "\x80\x84\x00\xEC\x80\x85\x00\xEC\x80\x86\x00\xEC\x80\x87\x00\xEC\x80\x88"
-     "\x00\xEC"
-     "\x80\x89\x00\xEC\x80\x8A\x00\xEC\x80\x8B\x00\xEC\x80\x8C\x00\xEC\x80\x8D"
-     "\x00\xEC"
-     "\x80\x8E\x00\xEC\x80\x8F\x00\xEC\x80\x90\x00\xEC\x80\x91\x00\xEC\x80\x92"
-     "\x00\xEC"
-     "\x80\x93\x00\xEC\x80\x94\x00\xEC\x80\x95\x00\xEC\x80\x96\x00\xEC\x80\x97"
-     "\x00\xEC"
-     "\x80\x98\x00\xEC\x80\x99\x00\xEC\x80\x9A\x00\xEC\x80\x9B\x00\xEC\x80\x9C"
-     "\x00\xEC"
-     "\x80\x9D\x00\xEC\x80\x9E\x00\xEC\x80\x9F\x00\xEC\x80\xA0\x00\xEC\x80\xA1"
-     "\x00\xEC"
-     "\x80\xA2\x00\xEC\x80\xA3\x00\xEC\x80\xA4\x00\xEC\x80\xA5\x00\xEC\x80\xA6"
-     "\x00\xEC"
-     "\x80\xA7\x00\xEC\x80\xA8\x00\xEC\x80\xA9\x00\xEC\x80\xAA\x00\xEC\x80\xAB"
-     "\x00\xEC"
-     "\x80\xAC\x00\xEC\x80\xAD\x00\xEC\x80\xAE\x00\xEC\x80\xAF\x00\xEC\x80\xB0"
-     "\x00\xEC"
-     "\x80\xB1\x00\xEC\x80\xB2\x00\xEC\x80\xB3\x00\xEC\x80\xB4\x00\xEC\x80\xB5"
-     "\x00\xEC"
-     "\x80\xB6\x00\xEC\x80\xB7\x00\xEC\x80\xB8\x00\xEC\x80\xB9\x00\xEC\x80\xBA"
-     "\x00\xEC"
-     "\x80\xBB\x00\xEC\x80\xBC\x00\xEC\x80\xBD\x00\xEC\x80\xBE\x00\xEC\x80\xBF"
-     "\x00\xEC"
-     "\x81\x80\x00\xEC\x81\x81\x00\xEC\x81\x82\x00\xEC\x81\x83\x00\xEC\x81\x84"
-     "\x00\xEC"
-     "\x81\x85\x00\xEC\x81\x86\x00\xEC\x81\x87\x00\xEC\x81\x88\x00\xEC\x81\x89"
-     "\x00\xEC"
-     "\x81\x8A\x00\xEC\x81\x8B\x00\xEC\x81\x8C\x00\xEC\x81\x8D\x00\xEC\x81\x8E"
-     "\x00\xEC"
-     "\x81\x8F\x00\xEC\x81\x90\x00\xEC\x81\x91\x00\xEC\x81\x92\x00\xEC\x81\x93"
-     "\x00\xEC"
-     "\x81\x94\x00\xEC\x81\x95\x00\xEC\x81\x96\x00\xEC\x81\x97\x00\xEC\x81\x98"
-     "\x00\xEC"
-     "\x81\x99\x00\xEC\x81\x9A\x00\xEC\x81\x9B\x00\xEC\x81\x9C\x00\xEC\x81\x9D"
-     "\x00\xEC"
-     "\x81\x9E\x00\xEC\x81\x9F\x00\xEC\x81\xA0\x00\xEC\x81\xA1\x00\xEC\x81\xA2"
-     "\x00\xEC"
-     "\x81\xA3\x00\xEC\x81\xA4\x00\xEC\x81\xA5\x00\xEC\x81\xA6\x00\xEC\x81\xA7"
-     "\x00\xEC"
-     "\x81\xA8\x00\xEC\x81\xA9\x00\xEC\x81\xAA\x00\xEC\x81\xAB\x00\xEC\x81\xAC"
-     "\x00\xEC"
-     "\x81\xAD\x00\xEC\x81\xAE\x00\xEC\x81\xAF\x00\xEC\x81\xB0\x00\xEC\x81\xB1"
-     "\x00\xEC"
-     "\x81\xB2\x00\xEC\x81\xB3\x00\xEC\x81\xB4\x00\xEC\x81\xB5\x00\xEC\x81\xB6"
-     "\x00\xEC"
-     "\x81\xB7\x00\xEC\x81\xB8\x00\xEC\x81\xB9\x00\xEC\x81\xBA\x00\xEC\x81\xBB"
-     "\x00\xEC"
-     "\x81\xBC\x00\xEC\x81\xBD\x00\xEC\x81\xBE\x00\xEC\x81\xBF\x00\xEC\x82\x80"
-     "\x00\xEC"
-     "\x82\x81\x00\xEC\x82\x82\x00\xEC\x82\x83\x00\xEC\x82\x84\x00\xEC\x82\x85"
-     "\x00\xEC"
-     "\x82\x86\x00\xEC\x82\x87\x00\xEC\x82\x88\x00\xEC\x82\x89\x00\xEC\x82\x8A"
-     "\x00\xEC"
-     "\x82\x8B\x00\xEC\x82\x8C\x00\xEC\x82\x8D\x00\xEC\x82\x8E\x00\xEC\x82\x8F"
-     "\x00\xEC"
-     "\x82\x90\x00\xEC\x82\x91\x00\xEC\x82\x92\x00\xEC\x82\x93\x00\xEC\x82\x94"
-     "\x00\xEC"
-     "\x82\x95\x00\xEC\x82\x96\x00\xEC\x82\x97\x00\xEC\x82\x98\x00\xEC\x82\x99"
-     "\x00\xEC"
-     "\x82\x9A\x00\xEC\x82\x9B\x00\xEC\x82\x9C\x00\xEC\x82\x9D\x00\xEC\x82\x9E"
-     "\x00\xEC"
-     "\x82\x9F\x00\xEC\x82\xA0\x00\xEC\x82\xA1\x00\xEC\x82\xA2\x00\xEC\x82\xA3"
-     "\x00\xEC"
-     "\x82\xA4\x00\xEC\x82\xA5\x00\xEC\x82\xA6\x00\xEC\x82\xA7\x00\xEC\x82\xA8"
-     "\x00\xEC"
-     "\x82\xA9\x00\xEC\x82\xAA\x00\xEC\x82\xAB\x00\xEC\x82\xAC\x00\xEC\x82\xAD"
-     "\x00\xEC"
-     "\x82\xAE\x00\xEC\x82\xAF\x00\xEC\x82\xB0\x00\xEC\x82\xB1\x00\xEC\x82\xB2"
-     "\x00\xEC"
-     "\x82\xB3\x00\xEC\x82\xB4\x00\xEC\x82\xB5\x00\xEC\x82\xB6\x00\xEC\x82\xB7"
-     "\x00\xEC"
-     "\x82\xB8\x00\xEC\x82\xB9\x00\xEC\x82\xBA\x00\xEC\x82\xBB\x00\xEC\x82\xBC"
-     "\x00\xEC"
-     "\x82\xBD\x00\xEC\x82\xBE\x00\xEC\x82\xBF\x00\xEC\x83\x80\x00\xEC\x83\x81"
-     "\x00\xEC"
-     "\x83\x82\x00\xEC\x83\x83\x00\xEC\x83\x84\x00\xEC\x83\x85\x00\xEC\x83\x86"
-     "\x00\xEC"
-     "\x83\x87\x00\xEC\x83\x88\x00\xEC\x83\x89\x00\xEC\x83\x8A\x00\xEC\x83\x8B"
-     "\x00\xEC"
-     "\x83\x8C\x00\xEC\x83\x8D\x00\xEC\x83\x8E\x00\xEC\x83\x8F\x00\xEC\x83\x90"
-     "\x00\xEC"
-     "\x83\x91\x00\xEC\x83\x92\x00\xEC\x83\x93\x00\xEC\x83\x94\x00\xEC\x83\x95"
-     "\x00\xEC"
-     "\x83\x96\x00\xEC\x83\x97\x00\xEC\x83\x98\x00\xEC\x83\x99\x00\xEC\x83\x9A"
-     "\x00\xEC"
-     "\x83\x9B\x00\xEC\x83\x9C\x00\xEC\x83\x9D\x00\xEC\x83\x9E\x00\xEC\x83\x9F"
-     "\x00\xEC"
-     "\x83\xA0\x00\xEC\x83\xA1\x00\xEC\x83\xA2\x00\xEC\x83\xA3\x00\xEC\x83\xA4"
-     "\x00\xEC"
-     "\x83\xA5\x00\xEC\x83\xA6\x00\xEC\x83\xA7\x00\xEC\x83\xA8\x00\xEC\x83\xA9"
-     "\x00\xEC"
-     "\x83\xAA\x00\xEC\x83\xAB\x00\xEC\x83\xAC\x00\xEC\x83\xAD\x00\xEC\x83\xAE"
-     "\x00\xEC"
-     "\x83\xAF\x00\xEC\x83\xB0\x00\xEC\x83\xB1\x00\xEC\x83\xB2\x00\xEC\x83\xB3"
-     "\x00\xEC"
-     "\x83\xB4\x00\xEC\x83\xB5\x00\xEC\x83\xB6\x00\xEC\x83\xB7\x00\xEC\x83\xB8"
-     "\x00\xEC"
-     "\x83\xB9\x00\xEC\x83\xBA\x00\xEC\x83\xBB\x00\xEC\x83\xBC\x00\xEC\x83\xBD"
-     "\x00\xEC"
-     "\x83\xBE\x00\xEC\x83\xBF\x00\xEC\x84\x80\x00\xEC\x84\x81\x00\xEC\x84\x82"
-     "\x00\xEC"
-     "\x84\x83\x00\xEC\x84\x84\x00\xEC\x84\x85\x00\xEC\x84\x86\x00\xEC\x84\x87"
-     "\x00\xEC"
-     "\x84\x88\x00\xEC\x84\x89\x00\xEC\x84\x8A\x00\xEC\x84\x8B\x00\xEC\x84\x8C"
-     "\x00\xEC"
-     "\x84\x8D\x00\xEC\x84\x8E\x00\xEC\x84\x8F\x00\xEC\x84\x90\x00\xEC\x84\x91"
-     "\x00\xEC"
-     "\x84\x92\x00\xEC\x84\x93\x00\xEC\x84\x94\x00\xEC\x84\x95\x00\xEC\x84\x96"
-     "\x00\xEC"
-     "\x84\x97\x00\xEC\x84\x98\x00\xEC\x84\x99\x00\xEC\x84\x9A\x00\xEC\x84\x9B"
-     "\x00\xEC"
-     "\x84\x9C\x00\xEC\x84\x9D\x00\xEC\x84\x9E\x00\xEC\x84\x9F\x00\xEC\x84\xA0"
-     "\x00\xEC"
-     "\x84\xA1\x00\xEC\x84\xA2\x00\xEC\x84\xA3\x00\xEC\x84\xA4\x00\xEC\x84\xA5"
-     "\x00\xEC"
-     "\x84\xA6\x00\xEC\x84\xA7\x00\xEC\x84\xA8\x00\xEC\x84\xA9\x00\xEC\x84\xAA"
-     "\x00\xEC"
-     "\x84\xAB\x00\xEC\x84\xAC\x00\xEC\x84\xAD\x00\xEC\x84\xAE\x00\xEC\x84\xAF"
-     "\x00\xEC"
-     "\x84\xB0\x00\xEC\x84\xB1\x00\xEC\x84\xB2\x00\xEC\x84\xB3\x00\xEC\x84\xB4"
-     "\x00\xEC"
-     "\x84\xB5\x00\xEC\x84\xB6\x00\xEC\x84\xB7\x00\xEC\x84\xB8\x00\xEC\x84\xB9"
-     "\x00\xEC"
-     "\x84\xBA\x00\xEC\x84\xBB\x00\xEC\x84\xBC\x00\xEC\x84\xBD\x00\xEC\x84\xBE"
-     "\x00\xEC"
-     "\x84\xBF\x00\xEC\x85\x80\x00\xEC\x85\x81\x00\xEC\x85\x82\x00\xEC\x85\x83"
-     "\x00\xEC"
-     "\x85\x84\x00\xEC\x85\x85\x00\xEC\x85\x86\x00\xEC\x85\x87\x00\xEC\x85\x88"
-     "\x00\xEC"
-     "\x85\x89\x00\xEC\x85\x8A\x00\xEC\x85\x8B\x00\xEC\x85\x8C\x00\xEC\x85\x8D"
-     "\x00\xEC"
-     "\x85\x8E\x00\xEC\x85\x8F\x00\xEC\x85\x90\x00\xEC\x85\x91\x00\xEC\x85\x92"
-     "\x00\xEC"
-     "\x85\x93\x00\xEC\x85\x94\x00\xEC\x85\x95\x00\xEC\x85\x96\x00\xEC\x85\x97"
-     "\x00\xEC"
-     "\x85\x98\x00\xEC\x85\x99\x00\xEC\x85\x9A\x00\xEC\x85\x9B\x00\xEC\x85\x9C"
-     "\x00\xEC"
-     "\x85\x9D\x00\xEC\x85\x9E\x00\xEC\x85\x9F\x00\xEC\x85\xA0\x00\xEC\x85\xA1"
-     "\x00\xEC"
-     "\x85\xA2\x00\xEC\x85\xA3\x00\xEC\x85\xA4\x00\xEC\x85\xA5\x00\xEC\x85\xA6"
-     "\x00\xEC"
-     "\x85\xA7\x00\xEC\x85\xA8\x00\xEC\x85\xA9\x00\xEC\x85\xAA\x00\xEC\x85\xAB"
-     "\x00\xEC"
-     "\x85\xAC\x00\xEC\x85\xAD\x00\xEC\x85\xAE\x00\xEC\x85\xAF\x00\xEC\x85\xB0"
-     "\x00\xEC"
-     "\x85\xB1\x00\xEC\x85\xB2\x00\xEC\x85\xB3\x00\xEC\x85\xB4\x00\xEC\x85\xB5"
-     "\x00\xEC"
-     "\x85\xB6\x00\xEC\x85\xB7\x00\xEC\x85\xB8\x00\xEC\x85\xB9\x00\xEC\x85\xBA"
-     "\x00\xEC"
-     "\x85\xBB\x00\xEC\x85\xBC\x00\xEC\x85\xBD\x00\xEC\x85\xBE\x00\xEC\x85\xBF"
-     "\x00\xEC"
-     "\x86\x80\x00\xEC\x86\x81\x00\xEC\x86\x82\x00\xEC\x86\x83\x00\xEC\x86\x84"
-     "\x00\xEC"
-     "\x86\x85\x00\xEC\x86\x86\x00\xEC\x86\x87\x00\xEC\x86\x88\x00\xEC\x86\x89"
-     "\x00\xEC"
-     "\x86\x8A\x00\xEC\x86\x8B\x00\xEC\x86\x8C\x00\xEC\x86\x8D\x00\xEC\x86\x8E"
-     "\x00\xEC"
-     "\x86\x8F\x00\xEC\x86\x90\x00\xEC\x86\x91\x00\xEC\x86\x92\x00\xEC\x86\x93"
-     "\x00\xEC"
-     "\x86\x94\x00\xEC\x86\x95\x00\xEC\x86\x96\x00\xEC\x86\x97\x00\xEC\x86\x98"
-     "\x00\xEC"
-     "\x86\x99\x00\xEC\x86\x9A\x00\xEC\x86\x9B\x00\xEC\x86\x9C\x00\xEC\x86\x9D"
-     "\x00\xEC"
-     "\x86\x9E\x00\xEC\x86\x9F\x00\xEC\x86\xA0\x00\xEC\x86\xA1\x00\xEC\x86\xA2"
-     "\x00\xEC"
-     "\x86\xA3\x00\xEC\x86\xA4\x00\xEC\x86\xA5\x00\xEC\x86\xA6\x00\xEC\x86\xA7"
-     "\x00\xEC"
-     "\x86\xA8\x00\xEC\x86\xA9\x00\xEC\x86\xAA\x00\xEC\x86\xAB\x00\xEC\x86\xAC"
-     "\x00\xEC"
-     "\x86\xAD\x00\xEC\x86\xAE\x00\xEC\x86\xAF\x00\xEC\x86\xB0\x00\xEC\x86\xB1"
-     "\x00\xEC"
-     "\x86\xB2\x00\xEC\x86\xB3\x00\xEC\x86\xB4\x00\xEC\x86\xB5\x00\xEC\x86\xB6"
-     "\x00\xEC"
-     "\x86\xB7\x00\xEC\x86\xB8\x00\xEC\x86\xB9\x00\xEC\x86\xBA\x00\xEC\x86\xBB"
-     "\x00\xEC"
-     "\x86\xBC\x00\xEC\x86\xBD\x00\xEC\x86\xBE\x00\xEC\x86\xBF\x00\xEC\x87\x80"
-     "\x00\xEC"
-     "\x87\x81\x00\xEC\x87\x82\x00\xEC\x87\x83\x00\xEC\x87\x84\x00\xEC\x87\x85"
-     "\x00\xEC"
-     "\x87\x86\x00\xEC\x87\x87\x00\xEC\x87\x88\x00\xEC\x87\x89\x00\xEC\x87\x8A"
-     "\x00\xEC"
-     "\x87\x8B\x00\xEC\x87\x8C\x00\xEC\x87\x8D\x00\xEC\x87\x8E\x00\xEC\x87\x8F"
-     "\x00\xEC"
-     "\x87\x90\x00\xEC\x87\x91\x00\xEC\x87\x92\x00\xEC\x87\x93\x00\xEC\x87\x94"
-     "\x00\xEC"
-     "\x87\x95\x00\xEC\x87\x96\x00\xEC\x87\x97\x00\xEC\x87\x98\x00\xEC\x87\x99"
-     "\x00\xEC"
-     "\x87\x9A\x00\xEC\x87\x9B\x00\xEC\x87\x9C\x00\xEC\x87\x9D\x00\xEC\x87\x9E"
-     "\x00\xEC"
-     "\x87\x9F\x00\xEC\x87\xA0\x00\xEC\x87\xA1\x00\xEC\x87\xA2\x00\xEC\x87\xA3"
-     "\x00\xEC"
-     "\x87\xA4\x00\xEC\x87\xA5\x00\xEC\x87\xA6\x00\xEC\x87\xA7\x00\xEC\x87\xA8"
-     "\x00\xEC"
-     "\x87\xA9\x00\xEC\x87\xAA\x00\xEC\x87\xAB\x00\xEC\x87\xAC\x00\xEC\x87\xAD"
-     "\x00\xEC"
-     "\x87\xAE\x00\xEC\x87\xAF\x00\xEC\x87\xB0\x00\xEC\x87\xB1\x00\xEC\x87\xB2"
-     "\x00\xEC"
-     "\x87\xB3\x00\xEC\x87\xB4\x00\xEC\x87\xB5\x00\xEC\x87\xB6\x00\xEC\x87\xB7"
-     "\x00\xEC"
-     "\x87\xB8\x00\xEC\x87\xB9\x00\xEC\x87\xBA\x00\xEC\x87\xBB\x00\xEC\x87\xBC"
-     "\x00\xEC"
-     "\x87\xBD\x00\xEC\x87\xBE\x00\xEC\x87\xBF\x00\xEC\x88\x80\x00\xEC\x88\x81"
-     "\x00\xEC"
-     "\x88\x82\x00\xEC\x88\x83\x00\xEC\x88\x84\x00\xEC\x88\x85\x00\xEC\x88\x86"
-     "\x00\xEC"
-     "\x88\x87\x00\xEC\x88\x88\x00\xEC\x88\x89\x00\xEC\x88\x8A\x00\xEC\x88\x8B"
-     "\x00\xEC"
-     "\x88\x8C\x00\xEC\x88\x8D\x00\xEC\x88\x8E\x00\xEC\x88\x8F\x00\xEC\x88\x90"
-     "\x00\xEC"
-     "\x88\x91\x00\xEC\x88\x92\x00\xEC\x88\x93\x00\xEC\x88\x94\x00\xEC\x88\x95"
-     "\x00\xEC"
-     "\x88\x96\x00\xEC\x88\x97\x00\xEC\x88\x98\x00\xEC\x88\x99\x00\xEC\x88\x9A"
-     "\x00\xEC"
-     "\x88\x9B\x00\xEC\x88\x9C\x00\xEC\x88\x9D\x00\xEC\x88\x9E\x00\xEC\x88\x9F"
-     "\x00\xEC"
-     "\x88\xA0\x00\xEC\x88\xA1\x00\xEC\x88\xA2\x00\xEC\x88\xA3\x00\xEC\x88\xA4"
-     "\x00\xEC"
-     "\x88\xA5\x00\xEC\x88\xA6\x00\xEC\x88\xA7\x00\xEC\x88\xA8\x00\xEC\x88\xA9"
-     "\x00\xEC"
-     "\x88\xAA\x00\xEC\x88\xAB\x00\xEC\x88\xAC\x00\xEC\x88\xAD\x00\xEC\x88\xAE"
-     "\x00\xEC"
-     "\x88\xAF\x00\xEC\x88\xB0\x00\xEC\x88\xB1\x00\xEC\x88\xB2\x00\xEC\x88\xB3"
-     "\x00\xEC"
-     "\x88\xB4\x00\xEC\x88\xB5\x00\xEC\x88\xB6\x00\xEC\x88\xB7\x00\xEC\x88\xB8"
-     "\x00\xEC"
-     "\x88\xB9\x00\xEC\x88\xBA\x00\xEC\x88\xBB\x00\xEC\x88\xBC\x00\xEC\x88\xBD"
-     "\x00\xEC"
-     "\x88\xBE\x00\xEC\x88\xBF\x00\xEC\x89\x80\x00\xEC\x89\x81\x00\xEC\x89\x82"
-     "\x00\xEC"
-     "\x89\x83\x00\xEC\x89\x84\x00\xEC\x89\x85\x00\xEC\x89\x86\x00\xEC\x89\x87"
-     "\x00\xEC"
-     "\x89\x88\x00\xEC\x89\x89\x00\xEC\x89\x8A\x00\xEC\x89\x8B\x00\xEC\x89\x8C"
-     "\x00\xEC"
-     "\x89\x8D\x00\xEC\x89\x8E\x00\xEC\x89\x8F\x00\xEC\x89\x90\x00\xEC\x89\x91"
-     "\x00\xEC"
-     "\x89\x92\x00\xEC\x89\x93\x00\xEC\x89\x94\x00\xEC\x89\x95\x00\xEC\x89\x96"
-     "\x00\xEC"
-     "\x89\x97\x00\xEC\x89\x98\x00\xEC\x89\x99\x00\xEC\x89\x9A\x00\xEC\x89\x9B"
-     "\x00\xEC"
-     "\x89\x9C\x00\xEC\x89\x9D\x00\xEC\x89\x9E\x00\xEC\x89\x9F\x00\xEC\x89\xA0"
-     "\x00\xEC"
-     "\x89\xA1\x00\xEC\x89\xA2\x00\xEC\x89\xA3\x00\xEC\x89\xA4\x00\xEC\x89\xA5"
-     "\x00\xEC"
-     "\x89\xA6\x00\xEC\x89\xA7\x00\xEC\x89\xA8\x00\xEC\x89\xA9\x00\xEC\x89\xAA"
-     "\x00\xEC"
-     "\x89\xAB\x00\xEC\x89\xAC\x00\xEC\x89\xAD\x00\xEC\x89\xAE\x00\xEC\x89\xAF"
-     "\x00\xEC"
-     "\x89\xB0\x00\xEC\x89\xB1\x00\xEC\x89\xB2\x00\xEC\x89\xB3\x00\xEC\x89\xB4"
-     "\x00\xEC"
-     "\x89\xB5\x00\xEC\x89\xB6\x00\xEC\x89\xB7\x00\xEC\x89\xB8\x00\xEC\x89\xB9"
-     "\x00\xEC"
-     "\x89\xBA\x00\xEC\x89\xBB\x00\xEC\x89\xBC\x00\xEC\x89\xBD\x00\xEC\x89\xBE"
-     "\x00\xEC"
-     "\x89\xBF\x00\xEC\x8A\x80\x00\xEC\x8A\x81\x00\xEC\x8A\x82\x00\xEC\x8A\x83"
-     "\x00\xEC"
-     "\x8A\x84\x00\xEC\x8A\x85\x00\xEC\x8A\x86\x00\xEC\x8A\x87\x00\xEC\x8A\x88"
-     "\x00\xEC"
-     "\x8A\x89\x00\xEC\x8A\x8A\x00\xEC\x8A\x8B\x00\xEC\x8A\x8C\x00\xEC\x8A\x8D"
-     "\x00\xEC"
-     "\x8A\x8E\x00\xEC\x8A\x8F\x00\xEC\x8A\x90\x00\xEC\x8A\x91\x00\xEC\x8A\x92"
-     "\x00\xEC"
-     "\x8A\x93\x00\xEC\x8A\x94\x00\xEC\x8A\x95\x00\xEC\x8A\x96\x00\xEC\x8A\x97"
-     "\x00\xEC"
-     "\x8A\x98\x00\xEC\x8A\x99\x00\xEC\x8A\x9A\x00\xEC\x8A\x9B\x00\xEC\x8A\x9C"
-     "\x00\xEC"
-     "\x8A\x9D\x00\xEC\x8A\x9E\x00\xEC\x8A\x9F\x00\xEC\x8A\xA0\x00\xEC\x8A\xA1"
-     "\x00\xEC"
-     "\x8A\xA2\x00\xEC\x8A\xA3\x00\xEC\x8A\xA4\x00\xEC\x8A\xA5\x00\xEC\x8A\xA6"
-     "\x00\xEC"
-     "\x8A\xA7\x00\xEC\x8A\xA8\x00\xEC\x8A\xA9\x00\xEC\x8A\xAA\x00\xEC\x8A\xAB"
-     "\x00\xEC"
-     "\x8A\xAC\x00\xEC\x8A\xAD\x00\xEC\x8A\xAE\x00\xEC\x8A\xAF\x00\xEC\x8A\xB0"
-     "\x00\xEC"
-     "\x8A\xB1\x00\xEC\x8A\xB2\x00\xEC\x8A\xB3\x00\xEC\x8A\xB4\x00\xEC\x8A\xB5"
-     "\x00\xEC"
-     "\x8A\xB6\x00\xEC\x8A\xB7\x00\xEC\x8A\xB8\x00\xEC\x8A\xB9\x00\xEC\x8A\xBA"
-     "\x00\xEC"
-     "\x8A\xBB\x00\xEC\x8A\xBC\x00\xEC\x8A\xBD\x00\xEC\x8A\xBE\x00\xEC\x8A\xBF"
-     "\x00\xEC"
-     "\x8B\x80\x00\xEC\x8B\x81\x00\xEC\x8B\x82\x00\xEC\x8B\x83\x00\xEC\x8B\x84"
-     "\x00\xEC"
-     "\x8B\x85\x00\xEC\x8B\x86\x00\xEC\x8B\x87\x00\xEC\x8B\x88\x00\xEC\x8B\x89"
-     "\x00\xEC"
-     "\x8B\x8A\x00\xEC\x8B\x8B\x00\xEC\x8B\x8C\x00\xEC\x8B\x8D\x00\xEC\x8B\x8E"
-     "\x00\xEC"
-     "\x8B\x8F\x00\xEC\x8B\x90\x00\xEC\x8B\x91\x00\xEC\x8B\x92\x00\xEC\x8B\x93"
-     "\x00\xEC"
-     "\x8B\x94\x00\xEC\x8B\x95\x00\xEC\x8B\x96\x00\xEC\x8B\x97\x00\xEC\x8B\x98"
-     "\x00\xEC"
-     "\x8B\x99\x00\xEC\x8B\x9A\x00\xEC\x8B\x9B\x00\xEC\x8B\x9C\x00\xEC\x8B\x9D"
-     "\x00\xEC"
-     "\x8B\x9E\x00\xEC\x8B\x9F\x00\xEC\x8B\xA0\x00\xEC\x8B\xA1\x00\xEC\x8B\xA2"
-     "\x00\xEC"
-     "\x8B\xA3\x00\xEC\x8B\xA4\x00\xEC\x8B\xA5\x00\xEC\x8B\xA6\x00\xEC\x8B\xA7"
-     "\x00\xEC"
-     "\x8B\xA8\x00\xEC\x8B\xA9\x00\xEC\x8B\xAA\x00\xEC\x8B\xAB\x00\xEC\x8B\xAC"
-     "\x00\xEC"
-     "\x8B\xAD\x00\xEC\x8B\xAE\x00\xEC\x8B\xAF\x00\xEC\x8B\xB0\x00\xEC\x8B\xB1"
-     "\x00\xEC"
-     "\x8B\xB2\x00\xEC\x8B\xB3\x00\xEC\x8B\xB4\x00\xEC\x8B\xB5\x00\xEC\x8B\xB6"
-     "\x00\xEC"
-     "\x8B\xB7\x00\xEC\x8B\xB8\x00\xEC\x8B\xB9\x00\xEC\x8B\xBA\x00\xEC\x8B\xBB"
-     "\x00\xEC"
-     "\x8B\xBC\x00\xEC\x8B\xBD\x00\xEC\x8B\xBE\x00\xEC\x8B\xBF\x00\xEC\x8C\x80"
-     "\x00\xEC"
-     "\x8C\x81\x00\xEC\x8C\x82\x00\xEC\x8C\x83\x00\xEC\x8C\x84\x00\xEC\x8C\x85"
-     "\x00\xEC"
-     "\x8C\x86\x00\xEC\x8C\x87\x00\xEC\x8C\x88\x00\xEC\x8C\x89\x00\xEC\x8C\x8A"
-     "\x00\xEC"
-     "\x8C\x8B\x00\xEC\x8C\x8C\x00\xEC\x8C\x8D\x00\xEC\x8C\x8E\x00\xEC\x8C\x8F"
-     "\x00\xEC"
-     "\x8C\x90\x00\xEC\x8C\x91\x00\xEC\x8C\x92\x00\xEC\x8C\x93\x00\xEC\x8C\x94"
-     "\x00\xEC"
-     "\x8C\x95\x00\xEC\x8C\x96\x00\xEC\x8C\x97\x00\xEC\x8C\x98\x00\xEC\x8C\x99"
-     "\x00\xEC"
-     "\x8C\x9A\x00\xEC\x8C\x9B\x00\xEC\x8C\x9C\x00\xEC\x8C\x9D\x00\xEC\x8C\x9E"
-     "\x00\xEC"
-     "\x8C\x9F\x00\xEC\x8C\xA0\x00\xEC\x8C\xA1\x00\xEC\x8C\xA2\x00\xEC\x8C\xA3"
-     "\x00\xEC"
-     "\x8C\xA4\x00\xEC\x8C\xA5\x00\xEC\x8C\xA6\x00\xEC\x8C\xA7\x00\xEC\x8C\xA8"
-     "\x00\xEC"
-     "\x8C\xA9\x00\xEC\x8C\xAA\x00\xEC\x8C\xAB\x00\xEC\x8C\xAC\x00\xEC\x8C\xAD"
-     "\x00\xEC"
-     "\x8C\xAE\x00\xEC\x8C\xAF\x00\xEC\x8C\xB0\x00\xEC\x8C\xB1\x00\xEC\x8C\xB2"
-     "\x00\xEC"
-     "\x8C\xB3\x00\xEC\x8C\xB4\x00\xEC\x8C\xB5\x00\xEC\x8C\xB6\x00\xEC\x8C\xB7"
-     "\x00\xEC"
-     "\x8C\xB8\x00\xEC\x8C\xB9\x00\xEC\x8C\xBA\x00\xEC\x8C\xBB\x00\xEC\x8C\xBC"
-     "\x00\xEC"
-     "\x8C\xBD\x00\xEC\x8C\xBE\x00\xEC\x8C\xBF\x00\xEC\x8D\x80\x00\xEC\x8D\x81"
-     "\x00\xEC"
-     "\x8D\x82\x00\xEC\x8D\x83\x00\xEC\x8D\x84\x00\xEC\x8D\x85\x00\xEC\x8D\x86"
-     "\x00\xEC"
-     "\x8D\x87\x00\xEC\x8D\x88\x00\xEC\x8D\x89\x00\xEC\x8D\x8A\x00\xEC\x8D\x8B"
-     "\x00\xEC"
-     "\x8D\x8C\x00\xEC\x8D\x8D\x00\xEC\x8D\x8E\x00\xEC\x8D\x8F\x00\xEC\x8D\x90"
-     "\x00\xEC"
-     "\x8D\x91\x00\xEC\x8D\x92\x00\xEC\x8D\x93\x00\xEC\x8D\x94\x00\xEC\x8D\x95"
-     "\x00\xEC"
-     "\x8D\x96\x00\xEC\x8D\x97\x00\xEC\x8D\x98\x00\xEC\x8D\x99\x00\xEC\x8D\x9A"
-     "\x00\xEC"
-     "\x8D\x9B\x00\xEC\x8D\x9C\x00\xEC\x8D\x9D\x00\xEC\x8D\x9E\x00\xEC\x8D\x9F"
-     "\x00\xEC"
-     "\x8D\xA0\x00\xEC\x8D\xA1\x00\xEC\x8D\xA2\x00\xEC\x8D\xA3\x00\xEC\x8D\xA4"
-     "\x00\xEC"
-     "\x8D\xA5\x00\xEC\x8D\xA6\x00\xEC\x8D\xA7\x00\xEC\x8D\xA8\x00\xEC\x8D\xA9"
-     "\x00\xEC"
-     "\x8D\xAA\x00\xEC\x8D\xAB\x00\xEC\x8D\xAC\x00\xEC\x8D\xAD\x00\xEC\x8D\xAE"
-     "\x00\xEC"
-     "\x8D\xAF\x00\xEC\x8D\xB0\x00\xEC\x8D\xB1\x00\xEC\x8D\xB2\x00\xEC\x8D\xB3"
-     "\x00\xEC"
-     "\x8D\xB4\x00\xEC\x8D\xB5\x00\xEC\x8D\xB6\x00\xEC\x8D\xB7\x00\xEC\x8D\xB8"
-     "\x00\xEC"
-     "\x8D\xB9\x00\xEC\x8D\xBA\x00\xEC\x8D\xBB\x00\xEC\x8D\xBC\x00\xEC\x8D\xBD"
-     "\x00\xEC"
-     "\x8D\xBE\x00\xEC\x8D\xBF\x00\xEC\x8E\x80\x00\xEC\x8E\x81\x00\xEC\x8E\x82"
-     "\x00\xEC"
-     "\x8E\x83\x00\xEC\x8E\x84\x00\xEC\x8E\x85\x00\xEC\x8E\x86\x00\xEC\x8E\x87"
-     "\x00\xEC"
-     "\x8E\x88\x00\xEC\x8E\x89\x00\xEC\x8E\x8A\x00\xEC\x8E\x8B\x00\xEC\x8E\x8C"
-     "\x00\xEC"
-     "\x8E\x8D\x00\xEC\x8E\x8E\x00\xEC\x8E\x8F\x00\xEC\x8E\x90\x00\xEC\x8E\x91"
-     "\x00\xEC"
-     "\x8E\x92\x00\xEC\x8E\x93\x00\xEC\x8E\x94\x00\xEC\x8E\x95\x00\xEC\x8E\x96"
-     "\x00\xEC"
-     "\x8E\x97\x00\xEC\x8E\x98\x00\xEC\x8E\x99\x00\xEC\x8E\x9A\x00\xEC\x8E\x9B"
-     "\x00\xEC"
-     "\x8E\x9C\x00\xEC\x8E\x9D\x00\xEC\x8E\x9E\x00\xEC\x8E\x9F\x00\xEC\x8E\xA0"
-     "\x00\xEC"
-     "\x8E\xA1\x00\xEC\x8E\xA2\x00\xEC\x8E\xA3\x00\xEC\x8E\xA4\x00\xEC\x8E\xA5"
-     "\x00\xEC"
-     "\x8E\xA6\x00\xEC\x8E\xA7\x00\xEC\x8E\xA8\x00\xEC\x8E\xA9\x00\xEC\x8E\xAA"
-     "\x00\xEC"
-     "\x8E\xAB\x00\xEC\x8E\xAC\x00\xEC\x8E\xAD\x00\xEC\x8E\xAE\x00\xEC\x8E\xAF"
-     "\x00\xEC"
-     "\x8E\xB0\x00\xEC\x8E\xB1\x00\xEC\x8E\xB2\x00\xEC\x8E\xB3\x00\xEC\x8E\xB4"
-     "\x00\xEC"
-     "\x8E\xB5\x00\xEC\x8E\xB6\x00\xEC\x8E\xB7\x00\xEC\x8E\xB8\x00\xEC\x8E\xB9"
-     "\x00\xEC"
-     "\x8E\xBA\x00\xEC\x8E\xBB\x00\xEC\x8E\xBC\x00\xEC\x8E\xBD\x00\xEC\x8E\xBE"
-     "\x00\xEC"
-     "\x8E\xBF\x00\xEC\x8F\x80\x00\xEC\x8F\x81\x00\xEC\x8F\x82\x00\xEC\x8F\x83"
-     "\x00\xEC"
-     "\x8F\x84\x00\xEC\x8F\x85\x00\xEC\x8F\x86\x00\xEC\x8F\x87\x00\xEC\x8F\x88"
-     "\x00\xEC"
-     "\x8F\x89\x00\xEC\x8F\x8A\x00\xEC\x8F\x8B\x00\xEC\x8F\x8C\x00\xEC\x8F\x8D"
-     "\x00\xEC"
-     "\x8F\x8E\x00\xEC\x8F\x8F\x00\xEC\x8F\x90\x00\xEC\x8F\x91\x00\xEC\x8F\x92"
-     "\x00\xEC"
-     "\x8F\x93\x00\xEC\x8F\x94\x00\xEC\x8F\x95\x00\xEC\x8F\x96\x00\xEC\x8F\x97"
-     "\x00\xEC"
-     "\x8F\x98\x00\xEC\x8F\x99\x00\xEC\x8F\x9A\x00\xEC\x8F\x9B\x00\xEC\x8F\x9C"
-     "\x00\xEC"
-     "\x8F\x9D\x00\xEC\x8F\x9E\x00\xEC\x8F\x9F\x00\xEC\x8F\xA0\x00\xEC\x8F\xA1"
-     "\x00\xEC"
-     "\x8F\xA2\x00\xEC\x8F\xA3\x00\xEC\x8F\xA4\x00\xEC\x8F\xA5\x00\xEC\x8F\xA6"
-     "\x00\xEC"
-     "\x8F\xA7\x00\xEC\x8F\xA8\x00\xEC\x8F\xA9\x00\xEC\x8F\xAA\x00\xEC\x8F\xAB"
-     "\x00\xEC"
-     "\x8F\xAC\x00\xEC\x8F\xAD\x00\xEC\x8F\xAE\x00\xEC\x8F\xAF\x00\xEC\x8F\xB0"
-     "\x00\xEC"
-     "\x8F\xB1\x00\xEC\x8F\xB2\x00\xEC\x8F\xB3\x00\xEC\x8F\xB4\x00\xEC\x8F\xB5"
-     "\x00\xEC"
-     "\x8F\xB6\x00\xEC\x8F\xB7\x00\xEC\x8F\xB8\x00\xEC\x8F\xB9\x00\xEC\x8F\xBA"
-     "\x00\xEC"
-     "\x8F\xBB\x00\xEC\x8F\xBC\x00\xEC\x8F\xBD\x00\xEC\x8F\xBE\x00\xEC\x8F\xBF"
-     "\x00\xEC"
-     "\x90\x80\x00\xEC\x90\x81\x00\xEC\x90\x82\x00\xEC\x90\x83\x00\xEC\x90\x84"
-     "\x00\xEC"
-     "\x90\x85\x00\xEC\x90\x86\x00\xEC\x90\x87\x00\xEC\x90\x88\x00\xEC\x90\x89"
-     "\x00\xEC"
-     "\x90\x8A\x00\xEC\x90\x8B\x00\xEC\x90\x8C\x00\xEC\x90\x8D\x00\xEC\x90\x8E"
-     "\x00\xEC"
-     "\x90\x8F\x00\xEC\x90\x90\x00\xEC\x90\x91\x00\xEC\x90\x92\x00\xEC\x90\x93"
-     "\x00\xEC"
-     "\x90\x94\x00\xEC\x90\x95\x00\xEC\x90\x96\x00\xEC\x90\x97\x00\xEC\x90\x98"
-     "\x00\xEC"
-     "\x90\x99\x00\xEC\x90\x9A\x00\xEC\x90\x9B\x00\xEC\x90\x9C\x00\xEC\x90\x9D"
-     "\x00\xEC"
-     "\x90\x9E\x00\xEC\x90\x9F\x00\xEC\x90\xA0\x00\xEC\x90\xA1\x00\xEC\x90\xA2"
-     "\x00\xEC"
-     "\x90\xA3\x00\xEC\x90\xA4\x00\xEC\x90\xA5\x00\xEC\x90\xA6\x00\xEC\x90\xA7"
-     "\x00\xEC"
-     "\x90\xA8\x00\xEC\x90\xA9\x00\xEC\x90\xAA\x00\xEC\x90\xAB\x00\xEC\x90\xAC"
-     "\x00\xEC"
-     "\x90\xAD\x00\xEC\x90\xAE\x00\xEC\x90\xAF\x00\xEC\x90\xB0\x00\xEC\x90\xB1"
-     "\x00\xEC"
-     "\x90\xB2\x00\xEC\x90\xB3\x00\xEC\x90\xB4\x00\xEC\x90\xB5\x00\xEC\x90\xB6"
-     "\x00\xEC"
-     "\x90\xB7\x00\xEC\x90\xB8\x00\xEC\x90\xB9\x00\xEC\x90\xBA\x00\xEC\x90\xBB"
-     "\x00\xEC"
-     "\x90\xBC\x00\xEC\x90\xBD\x00\xEC\x90\xBE\x00\xEC\x90\xBF\x00\xEC\x91\x80"
-     "\x00\xEC"
-     "\x91\x81\x00\xEC\x91\x82\x00\xEC\x91\x83\x00\xEC\x91\x84\x00\xEC\x91\x85"
-     "\x00\xEC"
-     "\x91\x86\x00\xEC\x91\x87\x00\xEC\x91\x88\x00\xEC\x91\x89\x00\xEC\x91\x8A"
-     "\x00\xEC"
-     "\x91\x8B\x00\xEC\x91\x8C\x00\xEC\x91\x8D\x00\xEC\x91\x8E\x00\xEC\x91\x8F"
-     "\x00\xEC"
-     "\x91\x90\x00\xEC\x91\x91\x00\xEC\x91\x92\x00\xEC\x91\x93\x00\xEC\x91\x94"
-     "\x00\xEC"
-     "\x91\x95\x00\xEC\x91\x96\x00\xEC\x91\x97\x00\xEC\x91\x98\x00\xEC\x91\x99"
-     "\x00\xEC"
-     "\x91\x9A\x00\xEC\x91\x9B\x00\xEC\x91\x9C\x00\xEC\x91\x9D\x00\xEC\x91\x9E"
-     "\x00\xEC"
-     "\x91\x9F\x00\xEC\x91\xA0\x00\xEC\x91\xA1\x00\xEC\x91\xA2\x00\xEC\x91\xA3"
-     "\x00\xEC"
-     "\x91\xA4\x00\xEC\x91\xA5\x00\xEC\x91\xA6\x00\xEC\x91\xA7\x00\xEC\x91\xA8"
-     "\x00\xEC"
-     "\x91\xA9\x00\xEC\x91\xAA\x00\xEC\x91\xAB\x00\xEC\x91\xAC\x00\xEC\x91\xAD"
-     "\x00\xEC"
-     "\x91\xAE\x00\xEC\x91\xAF\x00\xEC\x91\xB0\x00\xEC\x91\xB1\x00\xEC\x91\xB2"
-     "\x00\xEC"
-     "\x91\xB3\x00\xEC\x91\xB4\x00\xEC\x91\xB5\x00\xEC\x91\xB6\x00\xEC\x91\xB7"
-     "\x00\xEC"
-     "\x91\xB8\x00\xEC\x91\xB9\x00\xEC\x91\xBA\x00\xEC\x91\xBB\x00\xEC\x91\xBC"
-     "\x00\xEC"
-     "\x91\xBD\x00\xEC\x91\xBE\x00\xEC\x91\xBF\x00\xEC\x92\x80\x00\xEC\x92\x81"
-     "\x00\xEC"
-     "\x92\x82\x00\xEC\x92\x83\x00\xEC\x92\x84\x00\xEC\x92\x85\x00\xEC\x92\x86"
-     "\x00\xEC"
-     "\x92\x87\x00\xEC\x92\x88\x00\xEC\x92\x89\x00\xEC\x92\x8A\x00\xEC\x92\x8B"
-     "\x00\xEC"
-     "\x92\x8C\x00\xEC\x92\x8D\x00\xEC\x92\x8E\x00\xEC\x92\x8F\x00\xEC\x92\x90"
-     "\x00\xEC"
-     "\x92\x91\x00\xEC\x92\x92\x00\xEC\x92\x93\x00\xEC\x92\x94\x00\xEC\x92\x95"
-     "\x00\xEC"
-     "\x92\x96\x00\xEC\x92\x97\x00\xEC\x92\x98\x00\xEC\x92\x99\x00\xEC\x92\x9A"
-     "\x00\xEC"
-     "\x92\x9B\x00\xEC\x92\x9C\x00\xEC\x92\x9D\x00\xEC\x92\x9E\x00\xEC\x92\x9F"
-     "\x00\xEC"
-     "\x92\xA0\x00\xEC\x92\xA1\x00\xEC\x92\xA2\x00\xEC\x92\xA3\x00\xEC\x92\xA4"
-     "\x00\xEC"
-     "\x92\xA5\x00\xEC\x92\xA6\x00\xEC\x92\xA7\x00\xEC\x92\xA8\x00\xEC\x92\xA9"
-     "\x00\xEC"
-     "\x92\xAA\x00\xEC\x92\xAB\x00\xEC\x92\xAC\x00\xEC\x92\xAD\x00\xEC\x92\xAE"
-     "\x00\xEC"
-     "\x92\xAF\x00\xEC\x92\xB0\x00\xEC\x92\xB1\x00\xEC\x92\xB2\x00\xEC\x92\xB3"
-     "\x00\xEC"
-     "\x92\xB4\x00\xEC\x92\xB5\x00\xEC\x92\xB6\x00\xEC\x92\xB7\x00\xEC\x92\xB8"
-     "\x00\xEC"
-     "\x92\xB9\x00\xEC\x92\xBA\x00\xEC\x92\xBB\x00\xEC\x92\xBC\x00\xEC\x92\xBD"
-     "\x00\xEC"
-     "\x92\xBE\x00\xEC\x92\xBF\x00\xEC\x93\x80\x00\xEC\x93\x81\x00\xEC\x93\x82"
-     "\x00\xEC"
-     "\x93\x83\x00\xEC\x93\x84\x00\xEC\x93\x85\x00\xEC\x93\x86\x00\xEC\x93\x87"
-     "\x00\xEC"
-     "\x93\x88\x00\xEC\x93\x89\x00\xEC\x93\x8A\x00\xEC\x93\x8B\x00\xEC\x93\x8C"
-     "\x00\xEC"
-     "\x93\x8D\x00\xEC\x93\x8E\x00\xEC\x93\x8F\x00\xEC\x93\x90\x00\xEC\x93\x91"
-     "\x00\xEC"
-     "\x93\x92\x00\xEC\x93\x93\x00\xEC\x93\x94\x00\xEC\x93\x95\x00\xEC\x93\x96"
-     "\x00\xEC"
-     "\x93\x97\x00\xEC\x93\x98\x00\xEC\x93\x99\x00\xEC\x93\x9A\x00\xEC\x93\x9B"
-     "\x00\xEC"
-     "\x93\x9C\x00\xEC\x93\x9D\x00\xEC\x93\x9E\x00\xEC\x93\x9F\x00\xEC\x93\xA0"
-     "\x00\xEC"
-     "\x93\xA1\x00\xEC\x93\xA2\x00\xEC\x93\xA3\x00\xEC\x93\xA4\x00\xEC\x93\xA5"
-     "\x00\xEC"
-     "\x93\xA6\x00\xEC\x93\xA7\x00\xEC\x93\xA8\x00\xEC\x93\xA9\x00\xEC\x93\xAA"
-     "\x00\xEC"
-     "\x93\xAB\x00\xEC\x93\xAC\x00\xEC\x93\xAD\x00\xEC\x93\xAE\x00\xEC\x93\xAF"
-     "\x00\xEC"
-     "\x93\xB0\x00\xEC\x93\xB1\x00\xEC\x93\xB2\x00\xEC\x93\xB3\x00\xEC\x93\xB4"
-     "\x00\xEC"
-     "\x93\xB5\x00\xEC\x93\xB6\x00\xEC\x93\xB7\x00\xEC\x93\xB8\x00\xEC\x93\xB9"
-     "\x00\xEC"
-     "\x93\xBA\x00\xEC\x93\xBB\x00\xEC\x93\xBC\x00\xEC\x93\xBD\x00\xEC\x93\xBE"
-     "\x00\xEC"
-     "\x93\xBF\x00\xEC\x94\x80\x00\xEC\x94\x81\x00\xEC\x94\x82\x00\xEC\x94\x83"
-     "\x00\xEC"
-     "\x94\x84\x00\xEC\x94\x85\x00\xEC\x94\x86\x00\xEC\x94\x87\x00\xEC\x94\x88"
-     "\x00\xEC"
-     "\x94\x89\x00\xEC\x94\x8A\x00\xEC\x94\x8B\x00\xEC\x94\x8C\x00\xEC\x94\x8D"
-     "\x00\xEC"
-     "\x94\x8E\x00\xEC\x94\x8F\x00\xEC\x94\x90\x00\xEC\x94\x91\x00\xEC\x94\x92"
-     "\x00\xEC"
-     "\x94\x93\x00\xEC\x94\x94\x00\xEC\x94\x95\x00\xEC\x94\x96\x00\xEC\x94\x97"
-     "\x00\xEC"
-     "\x94\x98\x00\xEC\x94\x99\x00\xEC\x94\x9A\x00\xEC\x94\x9B\x00\xEC\x94\x9C"
-     "\x00\xEC"
-     "\x94\x9D\x00\xEC\x94\x9E\x00\xEC\x94\x9F\x00\xEC\x94\xA0\x00\xEC\x94\xA1"
-     "\x00\xEC"
-     "\x94\xA2\x00\xEC\x94\xA3\x00\xEC\x94\xA4\x00\xEC\x94\xA5\x00\xEC\x94\xA6"
-     "\x00\xEC"
-     "\x94\xA7\x00\xEC\x94\xA8\x00\xEC\x94\xA9\x00\xEC\x94\xAA\x00\xEC\x94\xAB"
-     "\x00\xEC"
-     "\x94\xAC\x00\xEC\x94\xAD\x00\xEC\x94\xAE\x00\xEC\x94\xAF\x00\xEC\x94\xB0"
-     "\x00\xEC"
-     "\x94\xB1\x00\xEC\x94\xB2\x00\xEC\x94\xB3\x00\xEC\x94\xB4\x00\xEC\x94\xB5"
-     "\x00\xEC"
-     "\x94\xB6\x00\xEC\x94\xB7\x00\xEC\x94\xB8\x00\xEC\x94\xB9\x00\xEC\x94\xBA"
-     "\x00\xEC"
-     "\x94\xBB\x00\xEC\x94\xBC\x00\xEC\x94\xBD\x00\xEC\x94\xBE\x00\xEC\x94\xBF"
-     "\x00\xEC"
-     "\x95\x80\x00\xEC\x95\x81\x00\xEC\x95\x82\x00\xEC\x95\x83\x00\xEC\x95\x84"
-     "\x00\xEC"
-     "\x95\x85\x00\xEC\x95\x86\x00\xEC\x95\x87\x00\xEC\x95\x88\x00\xEC\x95\x89"
-     "\x00\xEC"
-     "\x95\x8A\x00\xEC\x95\x8B\x00\xEC\x95\x8C\x00\xEC\x95\x8D\x00\xEC\x95\x8E"
-     "\x00\xEC"
-     "\x95\x8F\x00\xEC\x95\x90\x00\xEC\x95\x91\x00\xEC\x95\x92\x00\xEC\x95\x93"
-     "\x00\xEC"
-     "\x95\x94\x00\xEC\x95\x95\x00\xEC\x95\x96\x00\xEC\x95\x97\x00\xEC\x95\x98"
-     "\x00\xEC"
-     "\x95\x99\x00\xEC\x95\x9A\x00\xEC\x95\x9B\x00\xEC\x95\x9C\x00\xEC\x95\x9D"
-     "\x00\xEC"
-     "\x95\x9E\x00\xEC\x95\x9F\x00\xEC\x95\xA0\x00\xEC\x95\xA1\x00\xEC\x95\xA2"
-     "\x00\xEC"
-     "\x95\xA3\x00\xEC\x95\xA4\x00\xEC\x95\xA5\x00\xEC\x95\xA6\x00\xEC\x95\xA7"
-     "\x00\xEC"
-     "\x95\xA8\x00\xEC\x95\xA9\x00\xEC\x95\xAA\x00\xEC\x95\xAB\x00\xEC\x95\xAC"
-     "\x00\xEC"
-     "\x95\xAD\x00\xEC\x95\xAE\x00\xEC\x95\xAF\x00\xEC\x95\xB0\x00\xEC\x95\xB1"
-     "\x00\xEC"
-     "\x95\xB2\x00\xEC\x95\xB3\x00\xEC\x95\xB4\x00\xEC\x95\xB5\x00\xEC\x95\xB6"
-     "\x00\xEC"
-     "\x95\xB7\x00\xEC\x95\xB8\x00\xEC\x95\xB9\x00\xEC\x95\xBA\x00\xEC\x95\xBB"
-     "\x00\xEC"
-     "\x95\xBC\x00\xEC\x95\xBD\x00\xEC\x95\xBE\x00\xEC\x95\xBF\x00\xEC\x96\x80"
-     "\x00\xEC"
-     "\x96\x81\x00\xEC\x96\x82\x00\xEC\x96\x83\x00\xEC\x96\x84\x00\xEC\x96\x85"
-     "\x00\xEC"
-     "\x96\x86\x00\xEC\x96\x87\x00\xEC\x96\x88\x00\xEC\x96\x89\x00\xEC\x96\x8A"
-     "\x00\xEC"
-     "\x96\x8B\x00\xEC\x96\x8C\x00\xEC\x96\x8D\x00\xEC\x96\x8E\x00\xEC\x96\x8F"
-     "\x00\xEC"
-     "\x96\x90\x00\xEC\x96\x91\x00\xEC\x96\x92\x00\xEC\x96\x93\x00\xEC\x96\x94"
-     "\x00\xEC"
-     "\x96\x95\x00\xEC\x96\x96\x00\xEC\x96\x97\x00\xEC\x96\x98\x00\xEC\x96\x99"
-     "\x00\xEC"
-     "\x96\x9A\x00\xEC\x96\x9B\x00\xEC\x96\x9C\x00\xEC\x96\x9D\x00\xEC\x96\x9E"
-     "\x00\xEC"
-     "\x96\x9F\x00\xEC\x96\xA0\x00\xEC\x96\xA1\x00\xEC\x96\xA2\x00\xEC\x96\xA3"
-     "\x00\xEC"
-     "\x96\xA4\x00\xEC\x96\xA5\x00\xEC\x96\xA6\x00\xEC\x96\xA7\x00\xEC\x96\xA8"
-     "\x00\xEC"
-     "\x96\xA9\x00\xEC\x96\xAA\x00\xEC\x96\xAB\x00\xEC\x96\xAC\x00\xEC\x96\xAD"
-     "\x00\xEC"
-     "\x96\xAE\x00\xEC\x96\xAF\x00\xEC\x96\xB0\x00\xEC\x96\xB1\x00\xEC\x96\xB2"
-     "\x00\xEC"
-     "\x96\xB3\x00\xEC\x96\xB4\x00\xEC\x96\xB5\x00\xEC\x96\xB6\x00\xEC\x96\xB7"
-     "\x00\xEC"
-     "\x96\xB8\x00\xEC\x96\xB9\x00\xEC\x96\xBA\x00\xEC\x96\xBB\x00\xEC\x96\xBC"
-     "\x00\xEC"
-     "\x96\xBD\x00\xEC\x96\xBE\x00\xEC\x96\xBF\x00\xEC\x97\x80\x00\xEC\x97\x81"
-     "\x00\xEC"
-     "\x97\x82\x00\xEC\x97\x83\x00\xEC\x97\x84\x00\xEC\x97\x85\x00\xEC\x97\x86"
-     "\x00\xEC"
-     "\x97\x87\x00\xEC\x97\x88\x00\xEC\x97\x89\x00\xEC\x97\x8A\x00\xEC\x97\x8B"
-     "\x00\xEC"
-     "\x97\x8C\x00\xEC\x97\x8D\x00\xEC\x97\x8E\x00\xEC\x97\x8F\x00\xEC\x97\x90"
-     "\x00\xEC"
-     "\x97\x91\x00\xEC\x97\x92\x00\xEC\x97\x93\x00\xEC\x97\x94\x00\xEC\x97\x95"
-     "\x00\xEC"
-     "\x97\x96\x00\xEC\x97\x97\x00\xEC\x97\x98\x00\xEC\x97\x99\x00\xEC\x97\x9A"
-     "\x00\xEC"
-     "\x97\x9B\x00\xEC\x97\x9C\x00\xEC\x97\x9D\x00\xEC\x97\x9E\x00\xEC\x97\x9F"
-     "\x00\xEC"
-     "\x97\xA0\x00\xEC\x97\xA1\x00\xEC\x97\xA2\x00\xEC\x97\xA3\x00\xEC\x97\xA4"
-     "\x00\xEC"
-     "\x97\xA5\x00\xEC\x97\xA6\x00\xEC\x97\xA7\x00\xEC\x97\xA8\x00\xEC\x97\xA9"
-     "\x00\xEC"
-     "\x97\xAA\x00\xEC\x97\xAB\x00\xEC\x97\xAC\x00\xEC\x97\xAD\x00\xEC\x97\xAE"
-     "\x00\xEC"
-     "\x97\xAF\x00\xEC\x97\xB0\x00\xEC\x97\xB1\x00\xEC\x97\xB2\x00\xEC\x97\xB3"
-     "\x00\xEC"
-     "\x97\xB4\x00\xEC\x97\xB5\x00\xEC\x97\xB6\x00\xEC\x97\xB7\x00\xEC\x97\xB8"
-     "\x00\xEC"
-     "\x97\xB9\x00\xEC\x97\xBA\x00\xEC\x97\xBB\x00\xEC\x97\xBC\x00\xEC\x97\xBD"
-     "\x00\xEC"
-     "\x97\xBE\x00\xEC\x97\xBF\x00\xEC\x98\x80\x00\xEC\x98\x81\x00\xEC\x98\x82"
-     "\x00\xEC"
-     "\x98\x83\x00\xEC\x98\x84\x00\xEC\x98\x85\x00\xEC\x98\x86\x00\xEC\x98\x87"
-     "\x00\xEC"
-     "\x98\x88\x00\xEC\x98\x89\x00\xEC\x98\x8A\x00\xEC\x98\x8B\x00\xEC\x98\x8C"
-     "\x00\xEC"
-     "\x98\x8D\x00\xEC\x98\x8E\x00\xEC\x98\x8F\x00\xEC\x98\x90\x00\xEC\x98\x91"
-     "\x00\xEC"
-     "\x98\x92\x00\xEC\x98\x93\x00\xEC\x98\x94\x00\xEC\x98\x95\x00\xEC\x98\x96"
-     "\x00\xEC"
-     "\x98\x97\x00\xEC\x98\x98\x00\xEC\x98\x99\x00\xEC\x98\x9A\x00\xEC\x98\x9B"
-     "\x00\xEC"
-     "\x98\x9C\x00\xEC\x98\x9D\x00\xEC\x98\x9E\x00\xEC\x98\x9F\x00\xEC\x98\xA0"
-     "\x00\xEC"
-     "\x98\xA1\x00\xEC\x98\xA2\x00\xEC\x98\xA3\x00\xEC\x98\xA4\x00\xEC\x98\xA5"
-     "\x00\xEC"
-     "\x98\xA6\x00\xEC\x98\xA7\x00\xEC\x98\xA8\x00\xEC\x98\xA9\x00\xEC\x98\xAA"
-     "\x00\xEC"
-     "\x98\xAB\x00\xEC\x98\xAC\x00\xEC\x98\xAD\x00\xEC\x98\xAE\x00\xEC\x98\xAF"
-     "\x00\xEC"
-     "\x98\xB0\x00\xEC\x98\xB1\x00\xEC\x98\xB2\x00\xEC\x98\xB3\x00\xEC\x98\xB4"
-     "\x00\xEC"
-     "\x98\xB5\x00\xEC\x98\xB6\x00\xEC\x98\xB7\x00\xEC\x98\xB8\x00\xEC\x98\xB9"
-     "\x00\xEC"
-     "\x98\xBA\x00\xEC\x98\xBB\x00\xEC\x98\xBC\x00\xEC\x98\xBD\x00\xEC\x98\xBE"
-     "\x00\xEC"
-     "\x98\xBF\x00\xEC\x99\x80\x00\xEC\x99\x81\x00\xEC\x99\x82\x00\xEC\x99\x83"
-     "\x00\xEC"
-     "\x99\x84\x00\xEC\x99\x85\x00\xEC\x99\x86\x00\xEC\x99\x87\x00\xEC\x99\x88"
-     "\x00\xEC"
-     "\x99\x89\x00\xEC\x99\x8A\x00\xEC\x99\x8B\x00\xEC\x99\x8C\x00\xEC\x99\x8D"
-     "\x00\xEC"
-     "\x99\x8E\x00\xEC\x99\x8F\x00\xEC\x99\x90\x00\xEC\x99\x91\x00\xEC\x99\x92"
-     "\x00\xEC"
-     "\x99\x93\x00\xEC\x99\x94\x00\xEC\x99\x95\x00\xEC\x99\x96\x00\xEC\x99\x97"
-     "\x00\xEC"
-     "\x99\x98\x00\xEC\x99\x99\x00\xEC\x99\x9A\x00\xEC\x99\x9B\x00\xEC\x99\x9C"
-     "\x00\xEC"
-     "\x99\x9D\x00\xEC\x99\x9E\x00\xEC\x99\x9F\x00\xEC\x99\xA0\x00\xEC\x99\xA1"
-     "\x00\xEC"
-     "\x99\xA2\x00\xEC\x99\xA3\x00\xEC\x99\xA4\x00\xEC\x99\xA5\x00\xEC\x99\xA6"
-     "\x00\xEC"
-     "\x99\xA7\x00\xEC\x99\xA8\x00\xEC\x99\xA9\x00\xEC\x99\xAA\x00\xEC\x99\xAB"
-     "\x00\xEC"
-     "\x99\xAC\x00\xEC\x99\xAD\x00\xEC\x99\xAE\x00\xEC\x99\xAF\x00\xEC\x99\xB0"
-     "\x00\xEC"
-     "\x99\xB1\x00\xEC\x99\xB2\x00\xEC\x99\xB3\x00\xEC\x99\xB4\x00\xEC\x99\xB5"
-     "\x00\xEC"
-     "\x99\xB6\x00\xEC\x99\xB7\x00\xEC\x99\xB8\x00\xEC\x99\xB9\x00\xEC\x99\xBA"
-     "\x00\xEC"
-     "\x99\xBB\x00\xEC\x99\xBC\x00\xEC\x99\xBD\x00\xEC\x99\xBE\x00\xEC\x99\xBF"
-     "\x00\xEC"
-     "\x9A\x80\x00\xEC\x9A\x81\x00\xEC\x9A\x82\x00\xEC\x9A\x83\x00\xEC\x9A\x84"
-     "\x00\xEC"
-     "\x9A\x85\x00\xEC\x9A\x86\x00\xEC\x9A\x87\x00\xEC\x9A\x88\x00\xEC\x9A\x89"
-     "\x00\xEC"
-     "\x9A\x8A\x00\xEC\x9A\x8B\x00\xEC\x9A\x8C\x00\xEC\x9A\x8D\x00\xEC\x9A\x8E"
-     "\x00\xEC"
-     "\x9A\x8F\x00\xEC\x9A\x90\x00\xEC\x9A\x91\x00\xEC\x9A\x92\x00\xEC\x9A\x93"
-     "\x00\xEC"
-     "\x9A\x94\x00\xEC\x9A\x95\x00\xEC\x9A\x96\x00\xEC\x9A\x97\x00\xEC\x9A\x98"
-     "\x00\xEC"
-     "\x9A\x99\x00\xEC\x9A\x9A\x00\xEC\x9A\x9B\x00\xEC\x9A\x9C\x00\xEC\x9A\x9D"
-     "\x00\xEC"
-     "\x9A\x9E\x00\xEC\x9A\x9F\x00\xEC\x9A\xA0\x00\xEC\x9A\xA1\x00\xEC\x9A\xA2"
-     "\x00\xEC"
-     "\x9A\xA3\x00\xEC\x9A\xA4\x00\xEC\x9A\xA5\x00\xEC\x9A\xA6\x00\xEC\x9A\xA7"
-     "\x00\xEC"
-     "\x9A\xA8\x00\xEC\x9A\xA9\x00\xEC\x9A\xAA\x00\xEC\x9A\xAB\x00\xEC\x9A\xAC"
-     "\x00\xEC"
-     "\x9A\xAD\x00\xEC\x9A\xAE\x00\xEC\x9A\xAF\x00\xEC\x9A\xB0\x00\xEC\x9A\xB1"
-     "\x00\xEC"
-     "\x9A\xB2\x00\xEC\x9A\xB3\x00\xEC\x9A\xB4\x00\xEC\x9A\xB5\x00\xEC\x9A\xB6"
-     "\x00\xEC"
-     "\x9A\xB7\x00\xEC\x9A\xB8\x00\xEC\x9A\xB9\x00\xEC\x9A\xBA\x00\xEC\x9A\xBB"
-     "\x00\xEC"
-     "\x9A\xBC\x00\xEC\x9A\xBD\x00\xEC\x9A\xBE\x00\xEC\x9A\xBF\x00\xEC\x9B\x80"
-     "\x00\xEC"
-     "\x9B\x81\x00\xEC\x9B\x82\x00\xEC\x9B\x83\x00\xEC\x9B\x84\x00\xEC\x9B\x85"
-     "\x00\xEC"
-     "\x9B\x86\x00\xEC\x9B\x87\x00\xEC\x9B\x88\x00\xEC\x9B\x89\x00\xEC\x9B\x8A"
-     "\x00\xEC"
-     "\x9B\x8B\x00\xEC\x9B\x8C\x00\xEC\x9B\x8D\x00\xEC\x9B\x8E\x00\xEC\x9B\x8F"
-     "\x00\xEC"
-     "\x9B\x90\x00\xEC\x9B\x91\x00\xEC\x9B\x92\x00\xEC\x9B\x93\x00\xEC\x9B\x94"
-     "\x00\xEC"
-     "\x9B\x95\x00\xEC\x9B\x96\x00\xEC\x9B\x97\x00\xEC\x9B\x98\x00\xEC\x9B\x99"
-     "\x00\xEC"
-     "\x9B\x9A\x00\xEC\x9B\x9B\x00\xEC\x9B\x9C\x00\xEC\x9B\x9D\x00\xEC\x9B\x9E"
-     "\x00\xEC"
-     "\x9B\x9F\x00\xEC\x9B\xA0\x00\xEC\x9B\xA1\x00\xEC\x9B\xA2\x00\xEC\x9B\xA3"
-     "\x00\xEC"
-     "\x9B\xA4\x00\xEC\x9B\xA5\x00\xEC\x9B\xA6\x00\xEC\x9B\xA7\x00\xEC\x9B\xA8"
-     "\x00\xEC"
-     "\x9B\xA9\x00\xEC\x9B\xAA\x00\xEC\x9B\xAB\x00\xEC\x9B\xAC\x00\xEC\x9B\xAD"
-     "\x00\xEC"
-     "\x9B\xAE\x00\xEC\x9B\xAF\x00\xEC\x9B\xB0\x00\xEC\x9B\xB1\x00\xEC\x9B\xB2"
-     "\x00\xEC"
-     "\x9B\xB3\x00\xEC\x9B\xB4\x00\xEC\x9B\xB5\x00\xEC\x9B\xB6\x00\xEC\x9B\xB7"
-     "\x00\xEC"
-     "\x9B\xB8\x00\xEC\x9B\xB9\x00\xEC\x9B\xBA\x00\xEC\x9B\xBB\x00\xEC\x9B\xBC"
-     "\x00\xEC"
-     "\x9B\xBD\x00\xEC\x9B\xBE\x00\xEC\x9B\xBF\x00\xEC\x9C\x80\x00\xEC\x9C\x81"
-     "\x00\xEC"
-     "\x9C\x82\x00\xEC\x9C\x83\x00\xEC\x9C\x84\x00\xEC\x9C\x85\x00\xEC\x9C\x86"
-     "\x00\xEC"
-     "\x9C\x87\x00\xEC\x9C\x88\x00\xEC\x9C\x89\x00\xEC\x9C\x8A\x00\xEC\x9C\x8B"
-     "\x00\xEC"
-     "\x9C\x8C\x00\xEC\x9C\x8D\x00\xEC\x9C\x8E\x00\xEC\x9C\x8F\x00\xEC\x9C\x90"
-     "\x00\xEC"
-     "\x9C\x91\x00\xEC\x9C\x92\x00\xEC\x9C\x93\x00\xEC\x9C\x94\x00\xEC\x9C\x95"
-     "\x00\xEC"
-     "\x9C\x96\x00\xEC\x9C\x97\x00\xEC\x9C\x98\x00\xEC\x9C\x99\x00\xEC\x9C\x9A"
-     "\x00\xEC"
-     "\x9C\x9B\x00\xEC\x9C\x9C\x00\xEC\x9C\x9D\x00\xEC\x9C\x9E\x00\xEC\x9C\x9F"
-     "\x00\xEC"
-     "\x9C\xA0\x00\xEC\x9C\xA1\x00\xEC\x9C\xA2\x00\xEC\x9C\xA3\x00\xEC\x9C\xA4"
-     "\x00\xEC"
-     "\x9C\xA5\x00\xEC\x9C\xA6\x00\xEC\x9C\xA7\x00\xEC\x9C\xA8\x00\xEC\x9C\xA9"
-     "\x00\xEC"
-     "\x9C\xAA\x00\xEC\x9C\xAB\x00\xEC\x9C\xAC\x00\xEC\x9C\xAD\x00\xEC\x9C\xAE"
-     "\x00\xEC"
-     "\x9C\xAF\x00\xEC\x9C\xB0\x00\xEC\x9C\xB1\x00\xEC\x9C\xB2\x00\xEC\x9C\xB3"
-     "\x00\xEC"
-     "\x9C\xB4\x00\xEC\x9C\xB5\x00\xEC\x9C\xB6\x00\xEC\x9C\xB7\x00\xEC\x9C\xB8"
-     "\x00\xEC"
-     "\x9C\xB9\x00\xEC\x9C\xBA\x00\xEC\x9C\xBB\x00\xEC\x9C\xBC\x00\xEC\x9C\xBD"
-     "\x00\xEC"
-     "\x9C\xBE\x00\xEC\x9C\xBF\x00\xEC\x9D\x80\x00\xEC\x9D\x81\x00\xEC\x9D\x82"
-     "\x00\xEC"
-     "\x9D\x83\x00\xEC\x9D\x84\x00\xEC\x9D\x85\x00\xEC\x9D\x86\x00\xEC\x9D\x87"
-     "\x00\xEC"
-     "\x9D\x88\x00\xEC\x9D\x89\x00\xEC\x9D\x8A\x00\xEC\x9D\x8B\x00\xEC\x9D\x8C"
-     "\x00\xEC"
-     "\x9D\x8D\x00\xEC\x9D\x8E\x00\xEC\x9D\x8F\x00\xEC\x9D\x90\x00\xEC\x9D\x91"
-     "\x00\xEC"
-     "\x9D\x92\x00\xEC\x9D\x93\x00\xEC\x9D\x94\x00\xEC\x9D\x95\x00\xEC\x9D\x96"
-     "\x00\xEC"
-     "\x9D\x97\x00\xEC\x9D\x98\x00\xEC\x9D\x99\x00\xEC\x9D\x9A\x00\xEC\x9D\x9B"
-     "\x00\xEC"
-     "\x9D\x9C\x00\xEC\x9D\x9D\x00\xEC\x9D\x9E\x00\xEC\x9D\x9F\x00\xEC\x9D\xA0"
-     "\x00\xEC"
-     "\x9D\xA1\x00\xEC\x9D\xA2\x00\xEC\x9D\xA3\x00\xEC\x9D\xA4\x00\xEC\x9D\xA5"
-     "\x00\xEC"
-     "\x9D\xA6\x00\xEC\x9D\xA7\x00\xEC\x9D\xA8\x00\xEC\x9D\xA9\x00\xEC\x9D\xAA"
-     "\x00\xEC"
-     "\x9D\xAB\x00\xEC\x9D\xAC\x00\xEC\x9D\xAD\x00\xEC\x9D\xAE\x00\xEC\x9D\xAF"
-     "\x00\xEC"
-     "\x9D\xB0\x00\xEC\x9D\xB1\x00\xEC\x9D\xB2\x00\xEC\x9D\xB3\x00\xEC\x9D\xB4"
-     "\x00\xEC"
-     "\x9D\xB5\x00\xEC\x9D\xB6\x00\xEC\x9D\xB7\x00\xEC\x9D\xB8\x00\xEC\x9D\xB9"
-     "\x00\xEC"
-     "\x9D\xBA\x00\xEC\x9D\xBB\x00\xEC\x9D\xBC\x00\xEC\x9D\xBD\x00\xEC\x9D\xBE"
-     "\x00\xEC"
-     "\x9D\xBF\x00\xEC\x9E\x80\x00\xEC\x9E\x81\x00\xEC\x9E\x82\x00\xEC\x9E\x83"
-     "\x00\xEC"
-     "\x9E\x84\x00\xEC\x9E\x85\x00\xEC\x9E\x86\x00\xEC\x9E\x87\x00\xEC\x9E\x88"
-     "\x00\xEC"
-     "\x9E\x89\x00\xEC\x9E\x8A\x00\xEC\x9E\x8B\x00\xEC\x9E\x8C\x00\xEC\x9E\x8D"
-     "\x00\xEC"
-     "\x9E\x8E\x00\xEC\x9E\x8F\x00\xEC\x9E\x90\x00\xEC\x9E\x91\x00\xEC\x9E\x92"
-     "\x00\xEC"
-     "\x9E\x93\x00\xEC\x9E\x94\x00\xEC\x9E\x95\x00\xEC\x9E\x96\x00\xEC\x9E\x97"
-     "\x00\xEC"
-     "\x9E\x98\x00\xEC\x9E\x99\x00\xEC\x9E\x9A\x00\xEC\x9E\x9B\x00\xEC\x9E\x9C"
-     "\x00\xEC"
-     "\x9E\x9D\x00\xEC\x9E\x9E\x00\xEC\x9E\x9F\x00\xEC\x9E\xA0\x00\xEC\x9E\xA1"
-     "\x00\xEC"
-     "\x9E\xA2\x00\xEC\x9E\xA3\x00\xEC\x9E\xA4\x00\xEC\x9E\xA5\x00\xEC\x9E\xA6"
-     "\x00\xEC"
-     "\x9E\xA7\x00\xEC\x9E\xA8\x00\xEC\x9E\xA9\x00\xEC\x9E\xAA\x00\xEC\x9E\xAB"
-     "\x00\xEC"
-     "\x9E\xAC\x00\xEC\x9E\xAD\x00\xEC\x9E\xAE\x00\xEC\x9E\xAF\x00\xEC\x9E\xB0"
-     "\x00\xEC"
-     "\x9E\xB1\x00\xEC\x9E\xB2\x00\xEC\x9E\xB3\x00\xEC\x9E\xB4\x00\xEC\x9E\xB5"
-     "\x00\xEC"
-     "\x9E\xB6\x00\xEC\x9E\xB7\x00\xEC\x9E\xB8\x00\xEC\x9E\xB9\x00\xEC\x9E\xBA"
-     "\x00\xEC"
-     "\x9E\xBB\x00\xEC\x9E\xBC\x00\xEC\x9E\xBD\x00\xEC\x9E\xBE\x00\xEC\x9E\xBF"
-     "\x00\xEC"
-     "\x9F\x80\x00\xEC\x9F\x81\x00\xEC\x9F\x82\x00\xEC\x9F\x83\x00\xEC\x9F\x84"
-     "\x00\xEC"
-     "\x9F\x85\x00\xEC\x9F\x86\x00\xEC\x9F\x87\x00\xEC\x9F\x88\x00\xEC\x9F\x89"
-     "\x00\xEC"
-     "\x9F\x8A\x00\xEC\x9F\x8B\x00\xEC\x9F\x8C\x00\xEC\x9F\x8D\x00\xEC\x9F\x8E"
-     "\x00\xEC"
-     "\x9F\x8F\x00\xEC\x9F\x90\x00\xEC\x9F\x91\x00\xEC\x9F\x92\x00\xEC\x9F\x93"
-     "\x00\xEC"
-     "\x9F\x94\x00\xEC\x9F\x95\x00\xEC\x9F\x96\x00\xEC\x9F\x97\x00\xEC\x9F\x98"
-     "\x00\xEC"
-     "\x9F\x99\x00\xEC\x9F\x9A\x00\xEC\x9F\x9B\x00\xEC\x9F\x9C\x00\xEC\x9F\x9D"
-     "\x00\xEC"
-     "\x9F\x9E\x00\xEC\x9F\x9F\x00\xEC\x9F\xA0\x00\xEC\x9F\xA1\x00\xEC\x9F\xA2"
-     "\x00\xEC"
-     "\x9F\xA3\x00\xEC\x9F\xA4\x00\xEC\x9F\xA5\x00\xEC\x9F\xA6\x00\xEC\x9F\xA7"
-     "\x00\xEC"
-     "\x9F\xA8\x00\xEC\x9F\xA9\x00\xEC\x9F\xAA\x00\xEC\x9F\xAB\x00\xEC\x9F\xAC"
-     "\x00\xEC"
-     "\x9F\xAD\x00\xEC\x9F\xAE\x00\xEC\x9F\xAF\x00\xEC\x9F\xB0\x00\xEC\x9F\xB1"
-     "\x00\xEC"
-     "\x9F\xB2\x00\xEC\x9F\xB3\x00\xEC\x9F\xB4\x00\xEC\x9F\xB5\x00\xEC\x9F\xB6"
-     "\x00\xEC"
-     "\x9F\xB7\x00\xEC\x9F\xB8\x00\xEC\x9F\xB9\x00\xEC\x9F\xBA\x00\xEC\x9F\xBB"
-     "\x00\xEC"
-     "\x9F\xBC\x00\xEC\x9F\xBD\x00\xEC\x9F\xBE\x00\xEC\x9F\xBF\x00\xEC\xA0\x80"
-     "\x00\xEC"
-     "\xA0\x81\x00\xEC\xA0\x82\x00\xEC\xA0\x83\x00\xEC\xA0\x84\x00\xEC\xA0\x85"
-     "\x00\xEC"
-     "\xA0\x86\x00\xEC\xA0\x87\x00\xEC\xA0\x88\x00\xEC\xA0\x89\x00\xEC\xA0\x8A"
-     "\x00\xEC"
-     "\xA0\x8B\x00\xEC\xA0\x8C\x00\xEC\xA0\x8D\x00\xEC\xA0\x8E\x00\xEC\xA0\x8F"
-     "\x00\xEC"
-     "\xA0\x90\x00\xEC\xA0\x91\x00\xEC\xA0\x92\x00\xEC\xA0\x93\x00\xEC\xA0\x94"
-     "\x00\xEC"
-     "\xA0\x95\x00\xEC\xA0\x96\x00\xEC\xA0\x97\x00\xEC\xA0\x98\x00\xEC\xA0\x99"
-     "\x00\xEC"
-     "\xA0\x9A\x00\xEC\xA0\x9B\x00\xEC\xA0\x9C\x00\xEC\xA0\x9D\x00\xEC\xA0\x9E"
-     "\x00\xEC"
-     "\xA0\x9F\x00\xEC\xA0\xA0\x00\xEC\xA0\xA1\x00\xEC\xA0\xA2\x00\xEC\xA0\xA3"
-     "\x00\xEC"
-     "\xA0\xA4\x00\xEC\xA0\xA5\x00\xEC\xA0\xA6\x00\xEC\xA0\xA7\x00\xEC\xA0\xA8"
-     "\x00\xEC"
-     "\xA0\xA9\x00\xEC\xA0\xAA\x00\xEC\xA0\xAB\x00\xEC\xA0\xAC\x00\xEC\xA0\xAD"
-     "\x00\xEC"
-     "\xA0\xAE\x00\xEC\xA0\xAF\x00\xEC\xA0\xB0\x00\xEC\xA0\xB1\x00\xEC\xA0\xB2"
-     "\x00\xEC"
-     "\xA0\xB3\x00\xEC\xA0\xB4\x00\xEC\xA0\xB5\x00\xEC\xA0\xB6\x00\xEC\xA0\xB7"
-     "\x00\xEC"
-     "\xA0\xB8\x00\xEC\xA0\xB9\x00\xEC\xA0\xBA\x00\xEC\xA0\xBB\x00\xEC\xA0\xBC"
-     "\x00\xEC"
-     "\xA0\xBD\x00\xEC\xA0\xBE\x00\xEC\xA0\xBF\x00\xEC\xA1\x80\x00\xEC\xA1\x81"
-     "\x00\xEC"
-     "\xA1\x82\x00\xEC\xA1\x83\x00\xEC\xA1\x84\x00\xEC\xA1\x85\x00\xEC\xA1\x86"
-     "\x00\xEC"
-     "\xA1\x87\x00\xEC\xA1\x88\x00\xEC\xA1\x89\x00\xEC\xA1\x8A\x00\xEC\xA1\x8B"
-     "\x00\xEC"
-     "\xA1\x8C\x00\xEC\xA1\x8D\x00\xEC\xA1\x8E\x00\xEC\xA1\x8F\x00\xEC\xA1\x90"
-     "\x00\xEC"
-     "\xA1\x91\x00\xEC\xA1\x92\x00\xEC\xA1\x93\x00\xEC\xA1\x94\x00\xEC\xA1\x95"
-     "\x00\xEC"
-     "\xA1\x96\x00\xEC\xA1\x97\x00\xEC\xA1\x98\x00\xEC\xA1\x99\x00\xEC\xA1\x9A"
-     "\x00\xEC"
-     "\xA1\x9B\x00\xEC\xA1\x9C\x00\xEC\xA1\x9D\x00\xEC\xA1\x9E\x00\xEC\xA1\x9F"
-     "\x00\xEC"
-     "\xA1\xA0\x00\xEC\xA1\xA1\x00\xEC\xA1\xA2\x00\xEC\xA1\xA3\x00\xEC\xA1\xA4"
-     "\x00\xEC"
-     "\xA1\xA5\x00\xEC\xA1\xA6\x00\xEC\xA1\xA7\x00\xEC\xA1\xA8\x00\xEC\xA1\xA9"
-     "\x00\xEC"
-     "\xA1\xAA\x00\xEC\xA1\xAB\x00\xEC\xA1\xAC\x00\xEC\xA1\xAD\x00\xEC\xA1\xAE"
-     "\x00\xEC"
-     "\xA1\xAF\x00\xEC\xA1\xB0\x00\xEC\xA1\xB1\x00\xEC\xA1\xB2\x00\xEC\xA1\xB3"
-     "\x00\xEC"
-     "\xA1\xB4\x00\xEC\xA1\xB5\x00\xEC\xA1\xB6\x00\xEC\xA1\xB7\x00\xEC\xA1\xB8"
-     "\x00\xEC"
-     "\xA1\xB9\x00\xEC\xA1\xBA\x00\xEC\xA1\xBB\x00\xEC\xA1\xBC\x00\xEC\xA1\xBD"
-     "\x00\xEC"
-     "\xA1\xBE\x00\xEC\xA1\xBF\x00\xEC\xA2\x80\x00\xEC\xA2\x81\x00\xEC\xA2\x82"
-     "\x00\xEC"
-     "\xA2\x83\x00\xEC\xA2\x84\x00\xEC\xA2\x85\x00\xEC\xA2\x86\x00\xEC\xA2\x87"
-     "\x00\xEC"
-     "\xA2\x88\x00\xEC\xA2\x89\x00\xEC\xA2\x8A\x00\xEC\xA2\x8B\x00\xEC\xA2\x8C"
-     "\x00\xEC"
-     "\xA2\x8D\x00\xEC\xA2\x8E\x00\xEC\xA2\x8F\x00\xEC\xA2\x90\x00\xEC\xA2\x91"
-     "\x00\xEC"
-     "\xA2\x92\x00\xEC\xA2\x93\x00\xEC\xA2\x94\x00\xEC\xA2\x95\x00\xEC\xA2\x96"
-     "\x00\xEC"
-     "\xA2\x97\x00\xEC\xA2\x98\x00\xEC\xA2\x99\x00\xEC\xA2\x9A\x00\xEC\xA2\x9B"
-     "\x00\xEC"
-     "\xA2\x9C\x00\xEC\xA2\x9D\x00\xEC\xA2\x9E\x00\xEC\xA2\x9F\x00\xEC\xA2\xA0"
-     "\x00\xEC"
-     "\xA2\xA1\x00\xEC\xA2\xA2\x00\xEC\xA2\xA3\x00\xEC\xA2\xA4\x00\xEC\xA2\xA5"
-     "\x00\xEC"
-     "\xA2\xA6\x00\xEC\xA2\xA7\x00\xEC\xA2\xA8\x00\xEC\xA2\xA9\x00\xEC\xA2\xAA"
-     "\x00\xEC"
-     "\xA2\xAB\x00\xEC\xA2\xAC\x00\xEC\xA2\xAD\x00\xEC\xA2\xAE\x00\xEC\xA2\xAF"
-     "\x00\xEC"
-     "\xA2\xB0\x00\xEC\xA2\xB1\x00\xEC\xA2\xB2\x00\xEC\xA2\xB3\x00\xEC\xA2\xB4"
-     "\x00\xEC"
-     "\xA2\xB5\x00\xEC\xA2\xB6\x00\xEC\xA2\xB7\x00\xEC\xA2\xB8\x00\xEC\xA2\xB9"
-     "\x00\xEC"
-     "\xA2\xBA\x00\xEC\xA2\xBB\x00\xEC\xA2\xBC\x00\xEC\xA2\xBD\x00\xEC\xA2\xBE"
-     "\x00\xEC"
-     "\xA2\xBF\x00\xEC\xA3\x80\x00\xEC\xA3\x81\x00\xEC\xA3\x82\x00\xEC\xA3\x83"
-     "\x00\xEC"
-     "\xA3\x84\x00\xEC\xA3\x85\x00\xEC\xA3\x86\x00\xEC\xA3\x87\x00\xEC\xA3\x88"
-     "\x00\xEC"
-     "\xA3\x89\x00\xEC\xA3\x8A\x00\xEC\xA3\x8B\x00\xEC\xA3\x8C\x00\xEC\xA3\x8D"
-     "\x00\xEC"
-     "\xA3\x8E\x00\xEC\xA3\x8F\x00\xEC\xA3\x90\x00\xEC\xA3\x91\x00\xEC\xA3\x92"
-     "\x00\xEC"
-     "\xA3\x93\x00\xEC\xA3\x94\x00\xEC\xA3\x95\x00\xEC\xA3\x96\x00\xEC\xA3\x97"
-     "\x00\xEC"
-     "\xA3\x98\x00\xEC\xA3\x99\x00\xEC\xA3\x9A\x00\xEC\xA3\x9B\x00\xEC\xA3\x9C"
-     "\x00\xEC"
-     "\xA3\x9D\x00\xEC\xA3\x9E\x00\xEC\xA3\x9F\x00\xEC\xA3\xA0\x00\xEC\xA3\xA1"
-     "\x00\xEC"
-     "\xA3\xA2\x00\xEC\xA3\xA3\x00\xEC\xA3\xA4\x00\xEC\xA3\xA5\x00\xEC\xA3\xA6"
-     "\x00\xEC"
-     "\xA3\xA7\x00\xEC\xA3\xA8\x00\xEC\xA3\xA9\x00\xEC\xA3\xAA\x00\xEC\xA3\xAB"
-     "\x00\xEC"
-     "\xA3\xAC\x00\xEC\xA3\xAD\x00\xEC\xA3\xAE\x00\xEC\xA3\xAF\x00\xEC\xA3\xB0"
-     "\x00\xEC"
-     "\xA3\xB1\x00\xEC\xA3\xB2\x00\xEC\xA3\xB3\x00\xEC\xA3\xB4\x00\xEC\xA3\xB5"
-     "\x00\xEC"
-     "\xA3\xB6\x00\xEC\xA3\xB7\x00\xEC\xA3\xB8\x00\xEC\xA3\xB9\x00\xEC\xA3\xBA"
-     "\x00\xEC"
-     "\xA3\xBB\x00\xEC\xA3\xBC\x00\xEC\xA3\xBC\xEC\x9D\x98\x00\xEC\xA3\xBD\x00"
-     "\xEC\xA3"
-     "\xBE\x00\xEC\xA3\xBF\x00\xEC\xA4\x80\x00\xEC\xA4\x81\x00\xEC\xA4\x82\x00"
-     "\xEC\xA4"
-     "\x83\x00\xEC\xA4\x84\x00\xEC\xA4\x85\x00\xEC\xA4\x86\x00\xEC\xA4\x87\x00"
-     "\xEC\xA4"
-     "\x88\x00\xEC\xA4\x89\x00\xEC\xA4\x8A\x00\xEC\xA4\x8B\x00\xEC\xA4\x8C\x00"
-     "\xEC\xA4"
-     "\x8D\x00\xEC\xA4\x8E\x00\xEC\xA4\x8F\x00\xEC\xA4\x90\x00\xEC\xA4\x91\x00"
-     "\xEC\xA4"
-     "\x92\x00\xEC\xA4\x93\x00\xEC\xA4\x94\x00\xEC\xA4\x95\x00\xEC\xA4\x96\x00"
-     "\xEC\xA4"
-     "\x97\x00\xEC\xA4\x98\x00\xEC\xA4\x99\x00\xEC\xA4\x9A\x00\xEC\xA4\x9B\x00"
-     "\xEC\xA4"
-     "\x9C\x00\xEC\xA4\x9D\x00\xEC\xA4\x9E\x00\xEC\xA4\x9F\x00\xEC\xA4\xA0\x00"
-     "\xEC\xA4"
-     "\xA1\x00\xEC\xA4\xA2\x00\xEC\xA4\xA3\x00\xEC\xA4\xA4\x00\xEC\xA4\xA5\x00"
-     "\xEC\xA4"
-     "\xA6\x00\xEC\xA4\xA7\x00\xEC\xA4\xA8\x00\xEC\xA4\xA9\x00\xEC\xA4\xAA\x00"
-     "\xEC\xA4"
-     "\xAB\x00\xEC\xA4\xAC\x00\xEC\xA4\xAD\x00\xEC\xA4\xAE\x00\xEC\xA4\xAF\x00"
-     "\xEC\xA4"
-     "\xB0\x00\xEC\xA4\xB1\x00\xEC\xA4\xB2\x00\xEC\xA4\xB3\x00\xEC\xA4\xB4\x00"
-     "\xEC\xA4"
-     "\xB5\x00\xEC\xA4\xB6\x00\xEC\xA4\xB7\x00\xEC\xA4\xB8\x00\xEC\xA4\xB9\x00"
-     "\xEC\xA4"
-     "\xBA\x00\xEC\xA4\xBB\x00\xEC\xA4\xBC\x00\xEC\xA4\xBD\x00\xEC\xA4\xBE\x00"
-     "\xEC\xA4"
-     "\xBF\x00\xEC\xA5\x80\x00\xEC\xA5\x81\x00\xEC\xA5\x82\x00\xEC\xA5\x83\x00"
-     "\xEC\xA5"
-     "\x84\x00\xEC\xA5\x85\x00\xEC\xA5\x86\x00\xEC\xA5\x87\x00\xEC\xA5\x88\x00"
-     "\xEC\xA5"
-     "\x89\x00\xEC\xA5\x8A\x00\xEC\xA5\x8B\x00\xEC\xA5\x8C\x00\xEC\xA5\x8D\x00"
-     "\xEC\xA5"
-     "\x8E\x00\xEC\xA5\x8F\x00\xEC\xA5\x90\x00\xEC\xA5\x91\x00\xEC\xA5\x92\x00"
-     "\xEC\xA5"
-     "\x93\x00\xEC\xA5\x94\x00\xEC\xA5\x95\x00\xEC\xA5\x96\x00\xEC\xA5\x97\x00"
-     "\xEC\xA5"
-     "\x98\x00\xEC\xA5\x99\x00\xEC\xA5\x9A\x00\xEC\xA5\x9B\x00\xEC\xA5\x9C\x00"
-     "\xEC\xA5"
-     "\x9D\x00\xEC\xA5\x9E\x00\xEC\xA5\x9F\x00\xEC\xA5\xA0\x00\xEC\xA5\xA1\x00"
-     "\xEC\xA5"
-     "\xA2\x00\xEC\xA5\xA3\x00\xEC\xA5\xA4\x00\xEC\xA5\xA5\x00\xEC\xA5\xA6\x00"
-     "\xEC\xA5"
-     "\xA7\x00\xEC\xA5\xA8\x00\xEC\xA5\xA9\x00\xEC\xA5\xAA\x00\xEC\xA5\xAB\x00"
-     "\xEC\xA5"
-     "\xAC\x00\xEC\xA5\xAD\x00\xEC\xA5\xAE\x00\xEC\xA5\xAF\x00\xEC\xA5\xB0\x00"
-     "\xEC\xA5"
-     "\xB1\x00\xEC\xA5\xB2\x00\xEC\xA5\xB3\x00\xEC\xA5\xB4\x00\xEC\xA5\xB5\x00"
-     "\xEC\xA5"
-     "\xB6\x00\xEC\xA5\xB7\x00\xEC\xA5\xB8\x00\xEC\xA5\xB9\x00\xEC\xA5\xBA\x00"
-     "\xEC\xA5"
-     "\xBB\x00\xEC\xA5\xBC\x00\xEC\xA5\xBD\x00\xEC\xA5\xBE\x00\xEC\xA5\xBF\x00"
-     "\xEC\xA6"
-     "\x80\x00\xEC\xA6\x81\x00\xEC\xA6\x82\x00\xEC\xA6\x83\x00\xEC\xA6\x84\x00"
-     "\xEC\xA6"
-     "\x85\x00\xEC\xA6\x86\x00\xEC\xA6\x87\x00\xEC\xA6\x88\x00\xEC\xA6\x89\x00"
-     "\xEC\xA6"
-     "\x8A\x00\xEC\xA6\x8B\x00\xEC\xA6\x8C\x00\xEC\xA6\x8D\x00\xEC\xA6\x8E\x00"
-     "\xEC\xA6"
-     "\x8F\x00\xEC\xA6\x90\x00\xEC\xA6\x91\x00\xEC\xA6\x92\x00\xEC\xA6\x93\x00"
-     "\xEC\xA6"
-     "\x94\x00\xEC\xA6\x95\x00\xEC\xA6\x96\x00\xEC\xA6\x97\x00\xEC\xA6\x98\x00"
-     "\xEC\xA6"
-     "\x99\x00\xEC\xA6\x9A\x00\xEC\xA6\x9B\x00\xEC\xA6\x9C\x00\xEC\xA6\x9D\x00"
-     "\xEC\xA6"
-     "\x9E\x00\xEC\xA6\x9F\x00\xEC\xA6\xA0\x00\xEC\xA6\xA1\x00\xEC\xA6\xA2\x00"
-     "\xEC\xA6"
-     "\xA3\x00\xEC\xA6\xA4\x00\xEC\xA6\xA5\x00\xEC\xA6\xA6\x00\xEC\xA6\xA7\x00"
-     "\xEC\xA6"
-     "\xA8\x00\xEC\xA6\xA9\x00\xEC\xA6\xAA\x00\xEC\xA6\xAB\x00\xEC\xA6\xAC\x00"
-     "\xEC\xA6"
-     "\xAD\x00\xEC\xA6\xAE\x00\xEC\xA6\xAF\x00\xEC\xA6\xB0\x00\xEC\xA6\xB1\x00"
-     "\xEC\xA6"
-     "\xB2\x00\xEC\xA6\xB3\x00\xEC\xA6\xB4\x00\xEC\xA6\xB5\x00\xEC\xA6\xB6\x00"
-     "\xEC\xA6"
-     "\xB7\x00\xEC\xA6\xB8\x00\xEC\xA6\xB9\x00\xEC\xA6\xBA\x00\xEC\xA6\xBB\x00"
-     "\xEC\xA6"
-     "\xBC\x00\xEC\xA6\xBD\x00\xEC\xA6\xBE\x00\xEC\xA6\xBF\x00\xEC\xA7\x80\x00"
-     "\xEC\xA7"
-     "\x81\x00\xEC\xA7\x82\x00\xEC\xA7\x83\x00\xEC\xA7\x84\x00\xEC\xA7\x85\x00"
-     "\xEC\xA7"
-     "\x86\x00\xEC\xA7\x87\x00\xEC\xA7\x88\x00\xEC\xA7\x89\x00\xEC\xA7\x8A\x00"
-     "\xEC\xA7"
-     "\x8B\x00\xEC\xA7\x8C\x00\xEC\xA7\x8D\x00\xEC\xA7\x8E\x00\xEC\xA7\x8F\x00"
-     "\xEC\xA7"
-     "\x90\x00\xEC\xA7\x91\x00\xEC\xA7\x92\x00\xEC\xA7\x93\x00\xEC\xA7\x94\x00"
-     "\xEC\xA7"
-     "\x95\x00\xEC\xA7\x96\x00\xEC\xA7\x97\x00\xEC\xA7\x98\x00\xEC\xA7\x99\x00"
-     "\xEC\xA7"
-     "\x9A\x00\xEC\xA7\x9B\x00\xEC\xA7\x9C\x00\xEC\xA7\x9D\x00\xEC\xA7\x9E\x00"
-     "\xEC\xA7"
-     "\x9F\x00\xEC\xA7\xA0\x00\xEC\xA7\xA1\x00\xEC\xA7\xA2\x00\xEC\xA7\xA3\x00"
-     "\xEC\xA7"
-     "\xA4\x00\xEC\xA7\xA5\x00\xEC\xA7\xA6\x00\xEC\xA7\xA7\x00\xEC\xA7\xA8\x00"
-     "\xEC\xA7"
-     "\xA9\x00\xEC\xA7\xAA\x00\xEC\xA7\xAB\x00\xEC\xA7\xAC\x00\xEC\xA7\xAD\x00"
-     "\xEC\xA7"
-     "\xAE\x00\xEC\xA7\xAF\x00\xEC\xA7\xB0\x00\xEC\xA7\xB1\x00\xEC\xA7\xB2\x00"
-     "\xEC\xA7"
-     "\xB3\x00\xEC\xA7\xB4\x00\xEC\xA7\xB5\x00\xEC\xA7\xB6\x00\xEC\xA7\xB7\x00"
-     "\xEC\xA7"
-     "\xB8\x00\xEC\xA7\xB9\x00\xEC\xA7\xBA\x00\xEC\xA7\xBB\x00\xEC\xA7\xBC\x00"
-     "\xEC\xA7"
-     "\xBD\x00\xEC\xA7\xBE\x00\xEC\xA7\xBF\x00\xEC\xA8\x80\x00\xEC\xA8\x81\x00"
-     "\xEC\xA8"
-     "\x82\x00\xEC\xA8\x83\x00\xEC\xA8\x84\x00\xEC\xA8\x85\x00\xEC\xA8\x86\x00"
-     "\xEC\xA8"
-     "\x87\x00\xEC\xA8\x88\x00\xEC\xA8\x89\x00\xEC\xA8\x8A\x00\xEC\xA8\x8B\x00"
-     "\xEC\xA8"
-     "\x8C\x00\xEC\xA8\x8D\x00\xEC\xA8\x8E\x00\xEC\xA8\x8F\x00\xEC\xA8\x90\x00"
-     "\xEC\xA8"
-     "\x91\x00\xEC\xA8\x92\x00\xEC\xA8\x93\x00\xEC\xA8\x94\x00\xEC\xA8\x95\x00"
-     "\xEC\xA8"
-     "\x96\x00\xEC\xA8\x97\x00\xEC\xA8\x98\x00\xEC\xA8\x99\x00\xEC\xA8\x9A\x00"
-     "\xEC\xA8"
-     "\x9B\x00\xEC\xA8\x9C\x00\xEC\xA8\x9D\x00\xEC\xA8\x9E\x00\xEC\xA8\x9F\x00"
-     "\xEC\xA8"
-     "\xA0\x00\xEC\xA8\xA1\x00\xEC\xA8\xA2\x00\xEC\xA8\xA3\x00\xEC\xA8\xA4\x00"
-     "\xEC\xA8"
-     "\xA5\x00\xEC\xA8\xA6\x00\xEC\xA8\xA7\x00\xEC\xA8\xA8\x00\xEC\xA8\xA9\x00"
-     "\xEC\xA8"
-     "\xAA\x00\xEC\xA8\xAB\x00\xEC\xA8\xAC\x00\xEC\xA8\xAD\x00\xEC\xA8\xAE\x00"
-     "\xEC\xA8"
-     "\xAF\x00\xEC\xA8\xB0\x00\xEC\xA8\xB1\x00\xEC\xA8\xB2\x00\xEC\xA8\xB3\x00"
-     "\xEC\xA8"
-     "\xB4\x00\xEC\xA8\xB5\x00\xEC\xA8\xB6\x00\xEC\xA8\xB7\x00\xEC\xA8\xB8\x00"
-     "\xEC\xA8"
-     "\xB9\x00\xEC\xA8\xBA\x00\xEC\xA8\xBB\x00\xEC\xA8\xBC\x00\xEC\xA8\xBD\x00"
-     "\xEC\xA8"
-     "\xBE\x00\xEC\xA8\xBF\x00\xEC\xA9\x80\x00\xEC\xA9\x81\x00\xEC\xA9\x82\x00"
-     "\xEC\xA9"
-     "\x83\x00\xEC\xA9\x84\x00\xEC\xA9\x85\x00\xEC\xA9\x86\x00\xEC\xA9\x87\x00"
-     "\xEC\xA9"
-     "\x88\x00\xEC\xA9\x89\x00\xEC\xA9\x8A\x00\xEC\xA9\x8B\x00\xEC\xA9\x8C\x00"
-     "\xEC\xA9"
-     "\x8D\x00\xEC\xA9\x8E\x00\xEC\xA9\x8F\x00\xEC\xA9\x90\x00\xEC\xA9\x91\x00"
-     "\xEC\xA9"
-     "\x92\x00\xEC\xA9\x93\x00\xEC\xA9\x94\x00\xEC\xA9\x95\x00\xEC\xA9\x96\x00"
-     "\xEC\xA9"
-     "\x97\x00\xEC\xA9\x98\x00\xEC\xA9\x99\x00\xEC\xA9\x9A\x00\xEC\xA9\x9B\x00"
-     "\xEC\xA9"
-     "\x9C\x00\xEC\xA9\x9D\x00\xEC\xA9\x9E\x00\xEC\xA9\x9F\x00\xEC\xA9\xA0\x00"
-     "\xEC\xA9"
-     "\xA1\x00\xEC\xA9\xA2\x00\xEC\xA9\xA3\x00\xEC\xA9\xA4\x00\xEC\xA9\xA5\x00"
-     "\xEC\xA9"
-     "\xA6\x00\xEC\xA9\xA7\x00\xEC\xA9\xA8\x00\xEC\xA9\xA9\x00\xEC\xA9\xAA\x00"
-     "\xEC\xA9"
-     "\xAB\x00\xEC\xA9\xAC\x00\xEC\xA9\xAD\x00\xEC\xA9\xAE\x00\xEC\xA9\xAF\x00"
-     "\xEC\xA9"
-     "\xB0\x00\xEC\xA9\xB1\x00\xEC\xA9\xB2\x00\xEC\xA9\xB3\x00\xEC\xA9\xB4\x00"
-     "\xEC\xA9"
-     "\xB5\x00\xEC\xA9\xB6\x00\xEC\xA9\xB7\x00\xEC\xA9\xB8\x00\xEC\xA9\xB9\x00"
-     "\xEC\xA9"
-     "\xBA\x00\xEC\xA9\xBB\x00\xEC\xA9\xBC\x00\xEC\xA9\xBD\x00\xEC\xA9\xBE\x00"
-     "\xEC\xA9"
-     "\xBF\x00\xEC\xAA\x80\x00\xEC\xAA\x81\x00\xEC\xAA\x82\x00\xEC\xAA\x83\x00"
-     "\xEC\xAA"
-     "\x84\x00\xEC\xAA\x85\x00\xEC\xAA\x86\x00\xEC\xAA\x87\x00\xEC\xAA\x88\x00"
-     "\xEC\xAA"
-     "\x89\x00\xEC\xAA\x8A\x00\xEC\xAA\x8B\x00\xEC\xAA\x8C\x00\xEC\xAA\x8D\x00"
-     "\xEC\xAA"
-     "\x8E\x00\xEC\xAA\x8F\x00\xEC\xAA\x90\x00\xEC\xAA\x91\x00\xEC\xAA\x92\x00"
-     "\xEC\xAA"
-     "\x93\x00\xEC\xAA\x94\x00\xEC\xAA\x95\x00\xEC\xAA\x96\x00\xEC\xAA\x97\x00"
-     "\xEC\xAA"
-     "\x98\x00\xEC\xAA\x99\x00\xEC\xAA\x9A\x00\xEC\xAA\x9B\x00\xEC\xAA\x9C\x00"
-     "\xEC\xAA"
-     "\x9D\x00\xEC\xAA\x9E\x00\xEC\xAA\x9F\x00\xEC\xAA\xA0\x00\xEC\xAA\xA1\x00"
-     "\xEC\xAA"
-     "\xA2\x00\xEC\xAA\xA3\x00\xEC\xAA\xA4\x00\xEC\xAA\xA5\x00\xEC\xAA\xA6\x00"
-     "\xEC\xAA"
-     "\xA7\x00\xEC\xAA\xA8\x00\xEC\xAA\xA9\x00\xEC\xAA\xAA\x00\xEC\xAA\xAB\x00"
-     "\xEC\xAA"
-     "\xAC\x00\xEC\xAA\xAD\x00\xEC\xAA\xAE\x00\xEC\xAA\xAF\x00\xEC\xAA\xB0\x00"
-     "\xEC\xAA"
-     "\xB1\x00\xEC\xAA\xB2\x00\xEC\xAA\xB3\x00\xEC\xAA\xB4\x00\xEC\xAA\xB5\x00"
-     "\xEC\xAA"
-     "\xB6\x00\xEC\xAA\xB7\x00\xEC\xAA\xB8\x00\xEC\xAA\xB9\x00\xEC\xAA\xBA\x00"
-     "\xEC\xAA"
-     "\xBB\x00\xEC\xAA\xBC\x00\xEC\xAA\xBD\x00\xEC\xAA\xBE\x00\xEC\xAA\xBF\x00"
-     "\xEC\xAB"
-     "\x80\x00\xEC\xAB\x81\x00\xEC\xAB\x82\x00\xEC\xAB\x83\x00\xEC\xAB\x84\x00"
-     "\xEC\xAB"
-     "\x85\x00\xEC\xAB\x86\x00\xEC\xAB\x87\x00\xEC\xAB\x88\x00\xEC\xAB\x89\x00"
-     "\xEC\xAB"
-     "\x8A\x00\xEC\xAB\x8B\x00\xEC\xAB\x8C\x00\xEC\xAB\x8D\x00\xEC\xAB\x8E\x00"
-     "\xEC\xAB"
-     "\x8F\x00\xEC\xAB\x90\x00\xEC\xAB\x91\x00\xEC\xAB\x92\x00\xEC\xAB\x93\x00"
-     "\xEC\xAB"
-     "\x94\x00\xEC\xAB\x95\x00\xEC\xAB\x96\x00\xEC\xAB\x97\x00\xEC\xAB\x98\x00"
-     "\xEC\xAB"
-     "\x99\x00\xEC\xAB\x9A\x00\xEC\xAB\x9B\x00\xEC\xAB\x9C\x00\xEC\xAB\x9D\x00"
-     "\xEC\xAB"
-     "\x9E\x00\xEC\xAB\x9F\x00\xEC\xAB\xA0\x00\xEC\xAB\xA1\x00\xEC\xAB\xA2\x00"
-     "\xEC\xAB"
-     "\xA3\x00\xEC\xAB\xA4\x00\xEC\xAB\xA5\x00\xEC\xAB\xA6\x00\xEC\xAB\xA7\x00"
-     "\xEC\xAB"
-     "\xA8\x00\xEC\xAB\xA9\x00\xEC\xAB\xAA\x00\xEC\xAB\xAB\x00\xEC\xAB\xAC\x00"
-     "\xEC\xAB"
-     "\xAD\x00\xEC\xAB\xAE\x00\xEC\xAB\xAF\x00\xEC\xAB\xB0\x00\xEC\xAB\xB1\x00"
-     "\xEC\xAB"
-     "\xB2\x00\xEC\xAB\xB3\x00\xEC\xAB\xB4\x00\xEC\xAB\xB5\x00\xEC\xAB\xB6\x00"
-     "\xEC\xAB"
-     "\xB7\x00\xEC\xAB\xB8\x00\xEC\xAB\xB9\x00\xEC\xAB\xBA\x00\xEC\xAB\xBB\x00"
-     "\xEC\xAB"
-     "\xBC\x00\xEC\xAB\xBD\x00\xEC\xAB\xBE\x00\xEC\xAB\xBF\x00\xEC\xAC\x80\x00"
-     "\xEC\xAC"
-     "\x81\x00\xEC\xAC\x82\x00\xEC\xAC\x83\x00\xEC\xAC\x84\x00\xEC\xAC\x85\x00"
-     "\xEC\xAC"
-     "\x86\x00\xEC\xAC\x87\x00\xEC\xAC\x88\x00\xEC\xAC\x89\x00\xEC\xAC\x8A\x00"
-     "\xEC\xAC"
-     "\x8B\x00\xEC\xAC\x8C\x00\xEC\xAC\x8D\x00\xEC\xAC\x8E\x00\xEC\xAC\x8F\x00"
-     "\xEC\xAC"
-     "\x90\x00\xEC\xAC\x91\x00\xEC\xAC\x92\x00\xEC\xAC\x93\x00\xEC\xAC\x94\x00"
-     "\xEC\xAC"
-     "\x95\x00\xEC\xAC\x96\x00\xEC\xAC\x97\x00\xEC\xAC\x98\x00\xEC\xAC\x99\x00"
-     "\xEC\xAC"
-     "\x9A\x00\xEC\xAC\x9B\x00\xEC\xAC\x9C\x00\xEC\xAC\x9D\x00\xEC\xAC\x9E\x00"
-     "\xEC\xAC"
-     "\x9F\x00\xEC\xAC\xA0\x00\xEC\xAC\xA1\x00\xEC\xAC\xA2\x00\xEC\xAC\xA3\x00"
-     "\xEC\xAC"
-     "\xA4\x00\xEC\xAC\xA5\x00\xEC\xAC\xA6\x00\xEC\xAC\xA7\x00\xEC\xAC\xA8\x00"
-     "\xEC\xAC"
-     "\xA9\x00\xEC\xAC\xAA\x00\xEC\xAC\xAB\x00\xEC\xAC\xAC\x00\xEC\xAC\xAD\x00"
-     "\xEC\xAC"
-     "\xAE\x00\xEC\xAC\xAF\x00\xEC\xAC\xB0\x00\xEC\xAC\xB1\x00\xEC\xAC\xB2\x00"
-     "\xEC\xAC"
-     "\xB3\x00\xEC\xAC\xB4\x00\xEC\xAC\xB5\x00\xEC\xAC\xB6\x00\xEC\xAC\xB7\x00"
-     "\xEC\xAC"
-     "\xB8\x00\xEC\xAC\xB9\x00\xEC\xAC\xBA\x00\xEC\xAC\xBB\x00\xEC\xAC\xBC\x00"
-     "\xEC\xAC"
-     "\xBD\x00\xEC\xAC\xBE\x00\xEC\xAC\xBF\x00\xEC\xAD\x80\x00\xEC\xAD\x81\x00"
-     "\xEC\xAD"
-     "\x82\x00\xEC\xAD\x83\x00\xEC\xAD\x84\x00\xEC\xAD\x85\x00\xEC\xAD\x86\x00"
-     "\xEC\xAD"
-     "\x87\x00\xEC\xAD\x88\x00\xEC\xAD\x89\x00\xEC\xAD\x8A\x00\xEC\xAD\x8B\x00"
-     "\xEC\xAD"
-     "\x8C\x00\xEC\xAD\x8D\x00\xEC\xAD\x8E\x00\xEC\xAD\x8F\x00\xEC\xAD\x90\x00"
-     "\xEC\xAD"
-     "\x91\x00\xEC\xAD\x92\x00\xEC\xAD\x93\x00\xEC\xAD\x94\x00\xEC\xAD\x95\x00"
-     "\xEC\xAD"
-     "\x96\x00\xEC\xAD\x97\x00\xEC\xAD\x98\x00\xEC\xAD\x99\x00\xEC\xAD\x9A\x00"
-     "\xEC\xAD"
-     "\x9B\x00\xEC\xAD\x9C\x00\xEC\xAD\x9D\x00\xEC\xAD\x9E\x00\xEC\xAD\x9F\x00"
-     "\xEC\xAD"
-     "\xA0\x00\xEC\xAD\xA1\x00\xEC\xAD\xA2\x00\xEC\xAD\xA3\x00\xEC\xAD\xA4\x00"
-     "\xEC\xAD"
-     "\xA5\x00\xEC\xAD\xA6\x00\xEC\xAD\xA7\x00\xEC\xAD\xA8\x00\xEC\xAD\xA9\x00"
-     "\xEC\xAD"
-     "\xAA\x00\xEC\xAD\xAB\x00\xEC\xAD\xAC\x00\xEC\xAD\xAD\x00\xEC\xAD\xAE\x00"
-     "\xEC\xAD"
-     "\xAF\x00\xEC\xAD\xB0\x00\xEC\xAD\xB1\x00\xEC\xAD\xB2\x00\xEC\xAD\xB3\x00"
-     "\xEC\xAD"
-     "\xB4\x00\xEC\xAD\xB5\x00\xEC\xAD\xB6\x00\xEC\xAD\xB7\x00\xEC\xAD\xB8\x00"
-     "\xEC\xAD"
-     "\xB9\x00\xEC\xAD\xBA\x00\xEC\xAD\xBB\x00\xEC\xAD\xBC\x00\xEC\xAD\xBD\x00"
-     "\xEC\xAD"
-     "\xBE\x00\xEC\xAD\xBF\x00\xEC\xAE\x80\x00\xEC\xAE\x81\x00\xEC\xAE\x82\x00"
-     "\xEC\xAE"
-     "\x83\x00\xEC\xAE\x84\x00\xEC\xAE\x85\x00\xEC\xAE\x86\x00\xEC\xAE\x87\x00"
-     "\xEC\xAE"
-     "\x88\x00\xEC\xAE\x89\x00\xEC\xAE\x8A\x00\xEC\xAE\x8B\x00\xEC\xAE\x8C\x00"
-     "\xEC\xAE"
-     "\x8D\x00\xEC\xAE\x8E\x00\xEC\xAE\x8F\x00\xEC\xAE\x90\x00\xEC\xAE\x91\x00"
-     "\xEC\xAE"
-     "\x92\x00\xEC\xAE\x93\x00\xEC\xAE\x94\x00\xEC\xAE\x95\x00\xEC\xAE\x96\x00"
-     "\xEC\xAE"
-     "\x97\x00\xEC\xAE\x98\x00\xEC\xAE\x99\x00\xEC\xAE\x9A\x00\xEC\xAE\x9B\x00"
-     "\xEC\xAE"
-     "\x9C\x00\xEC\xAE\x9D\x00\xEC\xAE\x9E\x00\xEC\xAE\x9F\x00\xEC\xAE\xA0\x00"
-     "\xEC\xAE"
-     "\xA1\x00\xEC\xAE\xA2\x00\xEC\xAE\xA3\x00\xEC\xAE\xA4\x00\xEC\xAE\xA5\x00"
-     "\xEC\xAE"
-     "\xA6\x00\xEC\xAE\xA7\x00\xEC\xAE\xA8\x00\xEC\xAE\xA9\x00\xEC\xAE\xAA\x00"
-     "\xEC\xAE"
-     "\xAB\x00\xEC\xAE\xAC\x00\xEC\xAE\xAD\x00\xEC\xAE\xAE\x00\xEC\xAE\xAF\x00"
-     "\xEC\xAE"
-     "\xB0\x00\xEC\xAE\xB1\x00\xEC\xAE\xB2\x00\xEC\xAE\xB3\x00\xEC\xAE\xB4\x00"
-     "\xEC\xAE"
-     "\xB5\x00\xEC\xAE\xB6\x00\xEC\xAE\xB7\x00\xEC\xAE\xB8\x00\xEC\xAE\xB9\x00"
-     "\xEC\xAE"
-     "\xBA\x00\xEC\xAE\xBB\x00\xEC\xAE\xBC\x00\xEC\xAE\xBD\x00\xEC\xAE\xBE\x00"
-     "\xEC\xAE"
-     "\xBF\x00\xEC\xAF\x80\x00\xEC\xAF\x81\x00\xEC\xAF\x82\x00\xEC\xAF\x83\x00"
-     "\xEC\xAF"
-     "\x84\x00\xEC\xAF\x85\x00\xEC\xAF\x86\x00\xEC\xAF\x87\x00\xEC\xAF\x88\x00"
-     "\xEC\xAF"
-     "\x89\x00\xEC\xAF\x8A\x00\xEC\xAF\x8B\x00\xEC\xAF\x8C\x00\xEC\xAF\x8D\x00"
-     "\xEC\xAF"
-     "\x8E\x00\xEC\xAF\x8F\x00\xEC\xAF\x90\x00\xEC\xAF\x91\x00\xEC\xAF\x92\x00"
-     "\xEC\xAF"
-     "\x93\x00\xEC\xAF\x94\x00\xEC\xAF\x95\x00\xEC\xAF\x96\x00\xEC\xAF\x97\x00"
-     "\xEC\xAF"
-     "\x98\x00\xEC\xAF\x99\x00\xEC\xAF\x9A\x00\xEC\xAF\x9B\x00\xEC\xAF\x9C\x00"
-     "\xEC\xAF"
-     "\x9D\x00\xEC\xAF\x9E\x00\xEC\xAF\x9F\x00\xEC\xAF\xA0\x00\xEC\xAF\xA1\x00"
-     "\xEC\xAF"
-     "\xA2\x00\xEC\xAF\xA3\x00\xEC\xAF\xA4\x00\xEC\xAF\xA5\x00\xEC\xAF\xA6\x00"
-     "\xEC\xAF"
-     "\xA7\x00\xEC\xAF\xA8\x00\xEC\xAF\xA9\x00\xEC\xAF\xAA\x00\xEC\xAF\xAB\x00"
-     "\xEC\xAF"
-     "\xAC\x00\xEC\xAF\xAD\x00\xEC\xAF\xAE\x00\xEC\xAF\xAF\x00\xEC\xAF\xB0\x00"
-     "\xEC\xAF"
-     "\xB1\x00\xEC\xAF\xB2\x00\xEC\xAF\xB3\x00\xEC\xAF\xB4\x00\xEC\xAF\xB5\x00"
-     "\xEC\xAF"
-     "\xB6\x00\xEC\xAF\xB7\x00\xEC\xAF\xB8\x00\xEC\xAF\xB9\x00\xEC\xAF\xBA\x00"
-     "\xEC\xAF"
-     "\xBB\x00\xEC\xAF\xBC\x00\xEC\xAF\xBD\x00\xEC\xAF\xBE\x00\xEC\xAF\xBF\x00"
-     "\xEC\xB0"
-     "\x80\x00\xEC\xB0\x81\x00\xEC\xB0\x82\x00\xEC\xB0\x83\x00\xEC\xB0\x84\x00"
-     "\xEC\xB0"
-     "\x85\x00\xEC\xB0\x86\x00\xEC\xB0\x87\x00\xEC\xB0\x88\x00\xEC\xB0\x89\x00"
-     "\xEC\xB0"
-     "\x8A\x00\xEC\xB0\x8B\x00\xEC\xB0\x8C\x00\xEC\xB0\x8D\x00\xEC\xB0\x8E\x00"
-     "\xEC\xB0"
-     "\x8F\x00\xEC\xB0\x90\x00\xEC\xB0\x91\x00\xEC\xB0\x92\x00\xEC\xB0\x93\x00"
-     "\xEC\xB0"
-     "\x94\x00\xEC\xB0\x95\x00\xEC\xB0\x96\x00\xEC\xB0\x97\x00\xEC\xB0\x98\x00"
-     "\xEC\xB0"
-     "\x99\x00\xEC\xB0\x9A\x00\xEC\xB0\x9B\x00\xEC\xB0\x9C\x00\xEC\xB0\x9D\x00"
-     "\xEC\xB0"
-     "\x9E\x00\xEC\xB0\x9F\x00\xEC\xB0\xA0\x00\xEC\xB0\xA1\x00\xEC\xB0\xA2\x00"
-     "\xEC\xB0"
-     "\xA3\x00\xEC\xB0\xA4\x00\xEC\xB0\xA5\x00\xEC\xB0\xA6\x00\xEC\xB0\xA7\x00"
-     "\xEC\xB0"
-     "\xA8\x00\xEC\xB0\xA9\x00\xEC\xB0\xAA\x00\xEC\xB0\xAB\x00\xEC\xB0\xAC\x00"
-     "\xEC\xB0"
-     "\xAD\x00\xEC\xB0\xAE\x00\xEC\xB0\xAF\x00\xEC\xB0\xB0\x00\xEC\xB0\xB1\x00"
-     "\xEC\xB0"
-     "\xB2\x00\xEC\xB0\xB3\x00\xEC\xB0\xB4\x00\xEC\xB0\xB5\x00\xEC\xB0\xB6\x00"
-     "\xEC\xB0"
-     "\xB7\x00\xEC\xB0\xB8\x00\xEC\xB0\xB8\xEA\xB3\xA0\x00\xEC\xB0\xB9\x00\xEC"
-     "\xB0\xBA"
-     "\x00\xEC\xB0\xBB\x00\xEC\xB0\xBC\x00\xEC\xB0\xBD\x00\xEC\xB0\xBE\x00\xEC"
-     "\xB0\xBF"
-     "\x00\xEC\xB1\x80\x00\xEC\xB1\x81\x00\xEC\xB1\x82\x00\xEC\xB1\x83\x00\xEC"
-     "\xB1\x84"
-     "\x00\xEC\xB1\x85\x00\xEC\xB1\x86\x00\xEC\xB1\x87\x00\xEC\xB1\x88\x00\xEC"
-     "\xB1\x89"
-     "\x00\xEC\xB1\x8A\x00\xEC\xB1\x8B\x00\xEC\xB1\x8C\x00\xEC\xB1\x8D\x00\xEC"
-     "\xB1\x8E"
-     "\x00\xEC\xB1\x8F\x00\xEC\xB1\x90\x00\xEC\xB1\x91\x00\xEC\xB1\x92\x00\xEC"
-     "\xB1\x93"
-     "\x00\xEC\xB1\x94\x00\xEC\xB1\x95\x00\xEC\xB1\x96\x00\xEC\xB1\x97\x00\xEC"
-     "\xB1\x98"
-     "\x00\xEC\xB1\x99\x00\xEC\xB1\x9A\x00\xEC\xB1\x9B\x00\xEC\xB1\x9C\x00\xEC"
-     "\xB1\x9D"
-     "\x00\xEC\xB1\x9E\x00\xEC\xB1\x9F\x00\xEC\xB1\xA0\x00\xEC\xB1\xA1\x00\xEC"
-     "\xB1\xA2"
-     "\x00\xEC\xB1\xA3\x00\xEC\xB1\xA4\x00\xEC\xB1\xA5\x00\xEC\xB1\xA6\x00\xEC"
-     "\xB1\xA7"
-     "\x00\xEC\xB1\xA8\x00\xEC\xB1\xA9\x00\xEC\xB1\xAA\x00\xEC\xB1\xAB\x00\xEC"
-     "\xB1\xAC"
-     "\x00\xEC\xB1\xAD\x00\xEC\xB1\xAE\x00\xEC\xB1\xAF\x00\xEC\xB1\xB0\x00\xEC"
-     "\xB1\xB1"
-     "\x00\xEC\xB1\xB2\x00\xEC\xB1\xB3\x00\xEC\xB1\xB4\x00\xEC\xB1\xB5\x00\xEC"
-     "\xB1\xB6"
-     "\x00\xEC\xB1\xB7\x00\xEC\xB1\xB8\x00\xEC\xB1\xB9\x00\xEC\xB1\xBA\x00\xEC"
-     "\xB1\xBB"
-     "\x00\xEC\xB1\xBC\x00\xEC\xB1\xBD\x00\xEC\xB1\xBE\x00\xEC\xB1\xBF\x00\xEC"
-     "\xB2\x80"
-     "\x00\xEC\xB2\x81\x00\xEC\xB2\x82\x00\xEC\xB2\x83\x00\xEC\xB2\x84\x00\xEC"
-     "\xB2\x85"
-     "\x00\xEC\xB2\x86\x00\xEC\xB2\x87\x00\xEC\xB2\x88\x00\xEC\xB2\x89\x00\xEC"
-     "\xB2\x8A"
-     "\x00\xEC\xB2\x8B\x00\xEC\xB2\x8C\x00\xEC\xB2\x8D\x00\xEC\xB2\x8E\x00\xEC"
-     "\xB2\x8F"
-     "\x00\xEC\xB2\x90\x00\xEC\xB2\x91\x00\xEC\xB2\x92\x00\xEC\xB2\x93\x00\xEC"
-     "\xB2\x94"
-     "\x00\xEC\xB2\x95\x00\xEC\xB2\x96\x00\xEC\xB2\x97\x00\xEC\xB2\x98\x00\xEC"
-     "\xB2\x99"
-     "\x00\xEC\xB2\x9A\x00\xEC\xB2\x9B\x00\xEC\xB2\x9C\x00\xEC\xB2\x9D\x00\xEC"
-     "\xB2\x9E"
-     "\x00\xEC\xB2\x9F\x00\xEC\xB2\xA0\x00\xEC\xB2\xA1\x00\xEC\xB2\xA2\x00\xEC"
-     "\xB2\xA3"
-     "\x00\xEC\xB2\xA4\x00\xEC\xB2\xA5\x00\xEC\xB2\xA6\x00\xEC\xB2\xA7\x00\xEC"
-     "\xB2\xA8"
-     "\x00\xEC\xB2\xA9\x00\xEC\xB2\xAA\x00\xEC\xB2\xAB\x00\xEC\xB2\xAC\x00\xEC"
-     "\xB2\xAD"
-     "\x00\xEC\xB2\xAE\x00\xEC\xB2\xAF\x00\xEC\xB2\xB0\x00\xEC\xB2\xB1\x00\xEC"
-     "\xB2\xB2"
-     "\x00\xEC\xB2\xB3\x00\xEC\xB2\xB4\x00\xEC\xB2\xB5\x00\xEC\xB2\xB6\x00\xEC"
-     "\xB2\xB7"
-     "\x00\xEC\xB2\xB8\x00\xEC\xB2\xB9\x00\xEC\xB2\xBA\x00\xEC\xB2\xBB\x00\xEC"
-     "\xB2\xBC"
-     "\x00\xEC\xB2\xBD\x00\xEC\xB2\xBE\x00\xEC\xB2\xBF\x00\xEC\xB3\x80\x00\xEC"
-     "\xB3\x81"
-     "\x00\xEC\xB3\x82\x00\xEC\xB3\x83\x00\xEC\xB3\x84\x00\xEC\xB3\x85\x00\xEC"
-     "\xB3\x86"
-     "\x00\xEC\xB3\x87\x00\xEC\xB3\x88\x00\xEC\xB3\x89\x00\xEC\xB3\x8A\x00\xEC"
-     "\xB3\x8B"
-     "\x00\xEC\xB3\x8C\x00\xEC\xB3\x8D\x00\xEC\xB3\x8E\x00\xEC\xB3\x8F\x00\xEC"
-     "\xB3\x90"
-     "\x00\xEC\xB3\x91\x00\xEC\xB3\x92\x00\xEC\xB3\x93\x00\xEC\xB3\x94\x00\xEC"
-     "\xB3\x95"
-     "\x00\xEC\xB3\x96\x00\xEC\xB3\x97\x00\xEC\xB3\x98\x00\xEC\xB3\x99\x00\xEC"
-     "\xB3\x9A"
-     "\x00\xEC\xB3\x9B\x00\xEC\xB3\x9C\x00\xEC\xB3\x9D\x00\xEC\xB3\x9E\x00\xEC"
-     "\xB3\x9F"
-     "\x00\xEC\xB3\xA0\x00\xEC\xB3\xA1\x00\xEC\xB3\xA2\x00\xEC\xB3\xA3\x00\xEC"
-     "\xB3\xA4"
-     "\x00\xEC\xB3\xA5\x00\xEC\xB3\xA6\x00\xEC\xB3\xA7\x00\xEC\xB3\xA8\x00\xEC"
-     "\xB3\xA9"
-     "\x00\xEC\xB3\xAA\x00\xEC\xB3\xAB\x00\xEC\xB3\xAC\x00\xEC\xB3\xAD\x00\xEC"
-     "\xB3\xAE"
-     "\x00\xEC\xB3\xAF\x00\xEC\xB3\xB0\x00\xEC\xB3\xB1\x00\xEC\xB3\xB2\x00\xEC"
-     "\xB3\xB3"
-     "\x00\xEC\xB3\xB4\x00\xEC\xB3\xB5\x00\xEC\xB3\xB6\x00\xEC\xB3\xB7\x00\xEC"
-     "\xB3\xB8"
-     "\x00\xEC\xB3\xB9\x00\xEC\xB3\xBA\x00\xEC\xB3\xBB\x00\xEC\xB3\xBC\x00\xEC"
-     "\xB3\xBD"
-     "\x00\xEC\xB3\xBE\x00\xEC\xB3\xBF\x00\xEC\xB4\x80\x00\xEC\xB4\x81\x00\xEC"
-     "\xB4\x82"
-     "\x00\xEC\xB4\x83\x00\xEC\xB4\x84\x00\xEC\xB4\x85\x00\xEC\xB4\x86\x00\xEC"
-     "\xB4\x87"
-     "\x00\xEC\xB4\x88\x00\xEC\xB4\x89\x00\xEC\xB4\x8A\x00\xEC\xB4\x8B\x00\xEC"
-     "\xB4\x8C"
-     "\x00\xEC\xB4\x8D\x00\xEC\xB4\x8E\x00\xEC\xB4\x8F\x00\xEC\xB4\x90\x00\xEC"
-     "\xB4\x91"
-     "\x00\xEC\xB4\x92\x00\xEC\xB4\x93\x00\xEC\xB4\x94\x00\xEC\xB4\x95\x00\xEC"
-     "\xB4\x96"
-     "\x00\xEC\xB4\x97\x00\xEC\xB4\x98\x00\xEC\xB4\x99\x00\xEC\xB4\x9A\x00\xEC"
-     "\xB4\x9B"
-     "\x00\xEC\xB4\x9C\x00\xEC\xB4\x9D\x00\xEC\xB4\x9E\x00\xEC\xB4\x9F\x00\xEC"
-     "\xB4\xA0"
-     "\x00\xEC\xB4\xA1\x00\xEC\xB4\xA2\x00\xEC\xB4\xA3\x00\xEC\xB4\xA4\x00\xEC"
-     "\xB4\xA5"
-     "\x00\xEC\xB4\xA6\x00\xEC\xB4\xA7\x00\xEC\xB4\xA8\x00\xEC\xB4\xA9\x00\xEC"
-     "\xB4\xAA"
-     "\x00\xEC\xB4\xAB\x00\xEC\xB4\xAC\x00\xEC\xB4\xAD\x00\xEC\xB4\xAE\x00\xEC"
-     "\xB4\xAF"
-     "\x00\xEC\xB4\xB0\x00\xEC\xB4\xB1\x00\xEC\xB4\xB2\x00\xEC\xB4\xB3\x00\xEC"
-     "\xB4\xB4"
-     "\x00\xEC\xB4\xB5\x00\xEC\xB4\xB6\x00\xEC\xB4\xB7\x00\xEC\xB4\xB8\x00\xEC"
-     "\xB4\xB9"
-     "\x00\xEC\xB4\xBA\x00\xEC\xB4\xBB\x00\xEC\xB4\xBC\x00\xEC\xB4\xBD\x00\xEC"
-     "\xB4\xBE"
-     "\x00\xEC\xB4\xBF\x00\xEC\xB5\x80\x00\xEC\xB5\x81\x00\xEC\xB5\x82\x00\xEC"
-     "\xB5\x83"
-     "\x00\xEC\xB5\x84\x00\xEC\xB5\x85\x00\xEC\xB5\x86\x00\xEC\xB5\x87\x00\xEC"
-     "\xB5\x88"
-     "\x00\xEC\xB5\x89\x00\xEC\xB5\x8A\x00\xEC\xB5\x8B\x00\xEC\xB5\x8C\x00\xEC"
-     "\xB5\x8D"
-     "\x00\xEC\xB5\x8E\x00\xEC\xB5\x8F\x00\xEC\xB5\x90\x00\xEC\xB5\x91\x00\xEC"
-     "\xB5\x92"
-     "\x00\xEC\xB5\x93\x00\xEC\xB5\x94\x00\xEC\xB5\x95\x00\xEC\xB5\x96\x00\xEC"
-     "\xB5\x97"
-     "\x00\xEC\xB5\x98\x00\xEC\xB5\x99\x00\xEC\xB5\x9A\x00\xEC\xB5\x9B\x00\xEC"
-     "\xB5\x9C"
-     "\x00\xEC\xB5\x9D\x00\xEC\xB5\x9E\x00\xEC\xB5\x9F\x00\xEC\xB5\xA0\x00\xEC"
-     "\xB5\xA1"
-     "\x00\xEC\xB5\xA2\x00\xEC\xB5\xA3\x00\xEC\xB5\xA4\x00\xEC\xB5\xA5\x00\xEC"
-     "\xB5\xA6"
-     "\x00\xEC\xB5\xA7\x00\xEC\xB5\xA8\x00\xEC\xB5\xA9\x00\xEC\xB5\xAA\x00\xEC"
-     "\xB5\xAB"
-     "\x00\xEC\xB5\xAC\x00\xEC\xB5\xAD\x00\xEC\xB5\xAE\x00\xEC\xB5\xAF\x00\xEC"
-     "\xB5\xB0"
-     "\x00\xEC\xB5\xB1\x00\xEC\xB5\xB2\x00\xEC\xB5\xB3\x00\xEC\xB5\xB4\x00\xEC"
-     "\xB5\xB5"
-     "\x00\xEC\xB5\xB6\x00\xEC\xB5\xB7\x00\xEC\xB5\xB8\x00\xEC\xB5\xB9\x00\xEC"
-     "\xB5\xBA"
-     "\x00\xEC\xB5\xBB\x00\xEC\xB5\xBC\x00\xEC\xB5\xBD\x00\xEC\xB5\xBE\x00\xEC"
-     "\xB5\xBF"
-     "\x00\xEC\xB6\x80\x00\xEC\xB6\x81\x00\xEC\xB6\x82\x00\xEC\xB6\x83\x00\xEC"
-     "\xB6\x84"
-     "\x00\xEC\xB6\x85\x00\xEC\xB6\x86\x00\xEC\xB6\x87\x00\xEC\xB6\x88\x00\xEC"
-     "\xB6\x89"
-     "\x00\xEC\xB6\x8A\x00\xEC\xB6\x8B\x00\xEC\xB6\x8C\x00\xEC\xB6\x8D\x00\xEC"
-     "\xB6\x8E"
-     "\x00\xEC\xB6\x8F\x00\xEC\xB6\x90\x00\xEC\xB6\x91\x00\xEC\xB6\x92\x00\xEC"
-     "\xB6\x93"
-     "\x00\xEC\xB6\x94\x00\xEC\xB6\x95\x00\xEC\xB6\x96\x00\xEC\xB6\x97\x00\xEC"
-     "\xB6\x98"
-     "\x00\xEC\xB6\x99\x00\xEC\xB6\x9A\x00\xEC\xB6\x9B\x00\xEC\xB6\x9C\x00\xEC"
-     "\xB6\x9D"
-     "\x00\xEC\xB6\x9E\x00\xEC\xB6\x9F\x00\xEC\xB6\xA0\x00\xEC\xB6\xA1\x00\xEC"
-     "\xB6\xA2"
-     "\x00\xEC\xB6\xA3\x00\xEC\xB6\xA4\x00\xEC\xB6\xA5\x00\xEC\xB6\xA6\x00\xEC"
-     "\xB6\xA7"
-     "\x00\xEC\xB6\xA8\x00\xEC\xB6\xA9\x00\xEC\xB6\xAA\x00\xEC\xB6\xAB\x00\xEC"
-     "\xB6\xAC"
-     "\x00\xEC\xB6\xAD\x00\xEC\xB6\xAE\x00\xEC\xB6\xAF\x00\xEC\xB6\xB0\x00\xEC"
-     "\xB6\xB1"
-     "\x00\xEC\xB6\xB2\x00\xEC\xB6\xB3\x00\xEC\xB6\xB4\x00\xEC\xB6\xB5\x00\xEC"
-     "\xB6\xB6"
-     "\x00\xEC\xB6\xB7\x00\xEC\xB6\xB8\x00\xEC\xB6\xB9\x00\xEC\xB6\xBA\x00\xEC"
-     "\xB6\xBB"
-     "\x00\xEC\xB6\xBC\x00\xEC\xB6\xBD\x00\xEC\xB6\xBE\x00\xEC\xB6\xBF\x00\xEC"
-     "\xB7\x80"
-     "\x00\xEC\xB7\x81\x00\xEC\xB7\x82\x00\xEC\xB7\x83\x00\xEC\xB7\x84\x00\xEC"
-     "\xB7\x85"
-     "\x00\xEC\xB7\x86\x00\xEC\xB7\x87\x00\xEC\xB7\x88\x00\xEC\xB7\x89\x00\xEC"
-     "\xB7\x8A"
-     "\x00\xEC\xB7\x8B\x00\xEC\xB7\x8C\x00\xEC\xB7\x8D\x00\xEC\xB7\x8E\x00\xEC"
-     "\xB7\x8F"
-     "\x00\xEC\xB7\x90\x00\xEC\xB7\x91\x00\xEC\xB7\x92\x00\xEC\xB7\x93\x00\xEC"
-     "\xB7\x94"
-     "\x00\xEC\xB7\x95\x00\xEC\xB7\x96\x00\xEC\xB7\x97\x00\xEC\xB7\x98\x00\xEC"
-     "\xB7\x99"
-     "\x00\xEC\xB7\x9A\x00\xEC\xB7\x9B\x00\xEC\xB7\x9C\x00\xEC\xB7\x9D\x00\xEC"
-     "\xB7\x9E"
-     "\x00\xEC\xB7\x9F\x00\xEC\xB7\xA0\x00\xEC\xB7\xA1\x00\xEC\xB7\xA2\x00\xEC"
-     "\xB7\xA3"
-     "\x00\xEC\xB7\xA4\x00\xEC\xB7\xA5\x00\xEC\xB7\xA6\x00\xEC\xB7\xA7\x00\xEC"
-     "\xB7\xA8"
-     "\x00\xEC\xB7\xA9\x00\xEC\xB7\xAA\x00\xEC\xB7\xAB\x00\xEC\xB7\xAC\x00\xEC"
-     "\xB7\xAD"
-     "\x00\xEC\xB7\xAE\x00\xEC\xB7\xAF\x00\xEC\xB7\xB0\x00\xEC\xB7\xB1\x00\xEC"
-     "\xB7\xB2"
-     "\x00\xEC\xB7\xB3\x00\xEC\xB7\xB4\x00\xEC\xB7\xB5\x00\xEC\xB7\xB6\x00\xEC"
-     "\xB7\xB7"
-     "\x00\xEC\xB7\xB8\x00\xEC\xB7\xB9\x00\xEC\xB7\xBA\x00\xEC\xB7\xBB\x00\xEC"
-     "\xB7\xBC"
-     "\x00\xEC\xB7\xBD\x00\xEC\xB7\xBE\x00\xEC\xB7\xBF\x00\xEC\xB8\x80\x00\xEC"
-     "\xB8\x81"
-     "\x00\xEC\xB8\x82\x00\xEC\xB8\x83\x00\xEC\xB8\x84\x00\xEC\xB8\x85\x00\xEC"
-     "\xB8\x86"
-     "\x00\xEC\xB8\x87\x00\xEC\xB8\x88\x00\xEC\xB8\x89\x00\xEC\xB8\x8A\x00\xEC"
-     "\xB8\x8B"
-     "\x00\xEC\xB8\x8C\x00\xEC\xB8\x8D\x00\xEC\xB8\x8E\x00\xEC\xB8\x8F\x00\xEC"
-     "\xB8\x90"
-     "\x00\xEC\xB8\x91\x00\xEC\xB8\x92\x00\xEC\xB8\x93\x00\xEC\xB8\x94\x00\xEC"
-     "\xB8\x95"
-     "\x00\xEC\xB8\x96\x00\xEC\xB8\x97\x00\xEC\xB8\x98\x00\xEC\xB8\x99\x00\xEC"
-     "\xB8\x9A"
-     "\x00\xEC\xB8\x9B\x00\xEC\xB8\x9C\x00\xEC\xB8\x9D\x00\xEC\xB8\x9E\x00\xEC"
-     "\xB8\x9F"
-     "\x00\xEC\xB8\xA0\x00\xEC\xB8\xA1\x00\xEC\xB8\xA2\x00\xEC\xB8\xA3\x00\xEC"
-     "\xB8\xA4"
-     "\x00\xEC\xB8\xA5\x00\xEC\xB8\xA6\x00\xEC\xB8\xA7\x00\xEC\xB8\xA8\x00\xEC"
-     "\xB8\xA9"
-     "\x00\xEC\xB8\xAA\x00\xEC\xB8\xAB\x00\xEC\xB8\xAC\x00\xEC\xB8\xAD\x00\xEC"
-     "\xB8\xAE"
-     "\x00\xEC\xB8\xAF\x00\xEC\xB8\xB0\x00\xEC\xB8\xB1\x00\xEC\xB8\xB2\x00\xEC"
-     "\xB8\xB3"
-     "\x00\xEC\xB8\xB4\x00\xEC\xB8\xB5\x00\xEC\xB8\xB6\x00\xEC\xB8\xB7\x00\xEC"
-     "\xB8\xB8"
-     "\x00\xEC\xB8\xB9\x00\xEC\xB8\xBA\x00\xEC\xB8\xBB\x00\xEC\xB8\xBC\x00\xEC"
-     "\xB8\xBD"
-     "\x00\xEC\xB8\xBE\x00\xEC\xB8\xBF\x00\xEC\xB9\x80\x00\xEC\xB9\x81\x00\xEC"
-     "\xB9\x82"
-     "\x00\xEC\xB9\x83\x00\xEC\xB9\x84\x00\xEC\xB9\x85\x00\xEC\xB9\x86\x00\xEC"
-     "\xB9\x87"
-     "\x00\xEC\xB9\x88\x00\xEC\xB9\x89\x00\xEC\xB9\x8A\x00\xEC\xB9\x8B\x00\xEC"
-     "\xB9\x8C"
-     "\x00\xEC\xB9\x8D\x00\xEC\xB9\x8E\x00\xEC\xB9\x8F\x00\xEC\xB9\x90\x00\xEC"
-     "\xB9\x91"
-     "\x00\xEC\xB9\x92\x00\xEC\xB9\x93\x00\xEC\xB9\x94\x00\xEC\xB9\x95\x00\xEC"
-     "\xB9\x96"
-     "\x00\xEC\xB9\x97\x00\xEC\xB9\x98\x00\xEC\xB9\x99\x00\xEC\xB9\x9A\x00\xEC"
-     "\xB9\x9B"
-     "\x00\xEC\xB9\x9C\x00\xEC\xB9\x9D\x00\xEC\xB9\x9E\x00\xEC\xB9\x9F\x00\xEC"
-     "\xB9\xA0"
-     "\x00\xEC\xB9\xA1\x00\xEC\xB9\xA2\x00\xEC\xB9\xA3\x00\xEC\xB9\xA4\x00\xEC"
-     "\xB9\xA5"
-     "\x00\xEC\xB9\xA6\x00\xEC\xB9\xA7\x00\xEC\xB9\xA8\x00\xEC\xB9\xA9\x00\xEC"
-     "\xB9\xAA"
-     "\x00\xEC\xB9\xAB\x00\xEC\xB9\xAC\x00\xEC\xB9\xAD\x00\xEC\xB9\xAE\x00\xEC"
-     "\xB9\xAF"
-     "\x00\xEC\xB9\xB0\x00\xEC\xB9\xB1\x00\xEC\xB9\xB2\x00\xEC\xB9\xB3\x00\xEC"
-     "\xB9\xB4"
-     "\x00\xEC\xB9\xB5\x00\xEC\xB9\xB6\x00\xEC\xB9\xB7\x00\xEC\xB9\xB8\x00\xEC"
-     "\xB9\xB9"
-     "\x00\xEC\xB9\xBA\x00\xEC\xB9\xBB\x00\xEC\xB9\xBC\x00\xEC\xB9\xBD\x00\xEC"
-     "\xB9\xBE"
-     "\x00\xEC\xB9\xBF\x00\xEC\xBA\x80\x00\xEC\xBA\x81\x00\xEC\xBA\x82\x00\xEC"
-     "\xBA\x83"
-     "\x00\xEC\xBA\x84\x00\xEC\xBA\x85\x00\xEC\xBA\x86\x00\xEC\xBA\x87\x00\xEC"
-     "\xBA\x88"
-     "\x00\xEC\xBA\x89\x00\xEC\xBA\x8A\x00\xEC\xBA\x8B\x00\xEC\xBA\x8C\x00\xEC"
-     "\xBA\x8D"
-     "\x00\xEC\xBA\x8E\x00\xEC\xBA\x8F\x00\xEC\xBA\x90\x00\xEC\xBA\x91\x00\xEC"
-     "\xBA\x92"
-     "\x00\xEC\xBA\x93\x00\xEC\xBA\x94\x00\xEC\xBA\x95\x00\xEC\xBA\x96\x00\xEC"
-     "\xBA\x97"
-     "\x00\xEC\xBA\x98\x00\xEC\xBA\x99\x00\xEC\xBA\x9A\x00\xEC\xBA\x9B\x00\xEC"
-     "\xBA\x9C"
-     "\x00\xEC\xBA\x9D\x00\xEC\xBA\x9E\x00\xEC\xBA\x9F\x00\xEC\xBA\xA0\x00\xEC"
-     "\xBA\xA1"
-     "\x00\xEC\xBA\xA2\x00\xEC\xBA\xA3\x00\xEC\xBA\xA4\x00\xEC\xBA\xA5\x00\xEC"
-     "\xBA\xA6"
-     "\x00\xEC\xBA\xA7\x00\xEC\xBA\xA8\x00\xEC\xBA\xA9\x00\xEC\xBA\xAA\x00\xEC"
-     "\xBA\xAB"
-     "\x00\xEC\xBA\xAC\x00\xEC\xBA\xAD\x00\xEC\xBA\xAE\x00\xEC\xBA\xAF\x00\xEC"
-     "\xBA\xB0"
-     "\x00\xEC\xBA\xB1\x00\xEC\xBA\xB2\x00\xEC\xBA\xB3\x00\xEC\xBA\xB4\x00\xEC"
-     "\xBA\xB5"
-     "\x00\xEC\xBA\xB6\x00\xEC\xBA\xB7\x00\xEC\xBA\xB8\x00\xEC\xBA\xB9\x00\xEC"
-     "\xBA\xBA"
-     "\x00\xEC\xBA\xBB\x00\xEC\xBA\xBC\x00\xEC\xBA\xBD\x00\xEC\xBA\xBE\x00\xEC"
-     "\xBA\xBF"
-     "\x00\xEC\xBB\x80\x00\xEC\xBB\x81\x00\xEC\xBB\x82\x00\xEC\xBB\x83\x00\xEC"
-     "\xBB\x84"
-     "\x00\xEC\xBB\x85\x00\xEC\xBB\x86\x00\xEC\xBB\x87\x00\xEC\xBB\x88\x00\xEC"
-     "\xBB\x89"
-     "\x00\xEC\xBB\x8A\x00\xEC\xBB\x8B\x00\xEC\xBB\x8C\x00\xEC\xBB\x8D\x00\xEC"
-     "\xBB\x8E"
-     "\x00\xEC\xBB\x8F\x00\xEC\xBB\x90\x00\xEC\xBB\x91\x00\xEC\xBB\x92\x00\xEC"
-     "\xBB\x93"
-     "\x00\xEC\xBB\x94\x00\xEC\xBB\x95\x00\xEC\xBB\x96\x00\xEC\xBB\x97\x00\xEC"
-     "\xBB\x98"
-     "\x00\xEC\xBB\x99\x00\xEC\xBB\x9A\x00\xEC\xBB\x9B\x00\xEC\xBB\x9C\x00\xEC"
-     "\xBB\x9D"
-     "\x00\xEC\xBB\x9E\x00\xEC\xBB\x9F\x00\xEC\xBB\xA0\x00\xEC\xBB\xA1\x00\xEC"
-     "\xBB\xA2"
-     "\x00\xEC\xBB\xA3\x00\xEC\xBB\xA4\x00\xEC\xBB\xA5\x00\xEC\xBB\xA6\x00\xEC"
-     "\xBB\xA7"
-     "\x00\xEC\xBB\xA8\x00\xEC\xBB\xA9\x00\xEC\xBB\xAA\x00\xEC\xBB\xAB\x00\xEC"
-     "\xBB\xAC"
-     "\x00\xEC\xBB\xAD\x00\xEC\xBB\xAE\x00\xEC\xBB\xAF\x00\xEC\xBB\xB0\x00\xEC"
-     "\xBB\xB1"
-     "\x00\xEC\xBB\xB2\x00\xEC\xBB\xB3\x00\xEC\xBB\xB4\x00\xEC\xBB\xB5\x00\xEC"
-     "\xBB\xB6"
-     "\x00\xEC\xBB\xB7\x00\xEC\xBB\xB8\x00\xEC\xBB\xB9\x00\xEC\xBB\xBA\x00\xEC"
-     "\xBB\xBB"
-     "\x00\xEC\xBB\xBC\x00\xEC\xBB\xBD\x00\xEC\xBB\xBE\x00\xEC\xBB\xBF\x00\xEC"
-     "\xBC\x80"
-     "\x00\xEC\xBC\x81\x00\xEC\xBC\x82\x00\xEC\xBC\x83\x00\xEC\xBC\x84\x00\xEC"
-     "\xBC\x85"
-     "\x00\xEC\xBC\x86\x00\xEC\xBC\x87\x00\xEC\xBC\x88\x00\xEC\xBC\x89\x00\xEC"
-     "\xBC\x8A"
-     "\x00\xEC\xBC\x8B\x00\xEC\xBC\x8C\x00\xEC\xBC\x8D\x00\xEC\xBC\x8E\x00\xEC"
-     "\xBC\x8F"
-     "\x00\xEC\xBC\x90\x00\xEC\xBC\x91\x00\xEC\xBC\x92\x00\xEC\xBC\x93\x00\xEC"
-     "\xBC\x94"
-     "\x00\xEC\xBC\x95\x00\xEC\xBC\x96\x00\xEC\xBC\x97\x00\xEC\xBC\x98\x00\xEC"
-     "\xBC\x99"
-     "\x00\xEC\xBC\x9A\x00\xEC\xBC\x9B\x00\xEC\xBC\x9C\x00\xEC\xBC\x9D\x00\xEC"
-     "\xBC\x9E"
-     "\x00\xEC\xBC\x9F\x00\xEC\xBC\xA0\x00\xEC\xBC\xA1\x00\xEC\xBC\xA2\x00\xEC"
-     "\xBC\xA3"
-     "\x00\xEC\xBC\xA4\x00\xEC\xBC\xA5\x00\xEC\xBC\xA6\x00\xEC\xBC\xA7\x00\xEC"
-     "\xBC\xA8"
-     "\x00\xEC\xBC\xA9\x00\xEC\xBC\xAA\x00\xEC\xBC\xAB\x00\xEC\xBC\xAC\x00\xEC"
-     "\xBC\xAD"
-     "\x00\xEC\xBC\xAE\x00\xEC\xBC\xAF\x00\xEC\xBC\xB0\x00\xEC\xBC\xB1\x00\xEC"
-     "\xBC\xB2"
-     "\x00\xEC\xBC\xB3\x00\xEC\xBC\xB4\x00\xEC\xBC\xB5\x00\xEC\xBC\xB6\x00\xEC"
-     "\xBC\xB7"
-     "\x00\xEC\xBC\xB8\x00\xEC\xBC\xB9\x00\xEC\xBC\xBA\x00\xEC\xBC\xBB\x00\xEC"
-     "\xBC\xBC"
-     "\x00\xEC\xBC\xBD\x00\xEC\xBC\xBE\x00\xEC\xBC\xBF\x00\xEC\xBD\x80\x00\xEC"
-     "\xBD\x81"
-     "\x00\xEC\xBD\x82\x00\xEC\xBD\x83\x00\xEC\xBD\x84\x00\xEC\xBD\x85\x00\xEC"
-     "\xBD\x86"
-     "\x00\xEC\xBD\x87\x00\xEC\xBD\x88\x00\xEC\xBD\x89\x00\xEC\xBD\x8A\x00\xEC"
-     "\xBD\x8B"
-     "\x00\xEC\xBD\x8C\x00\xEC\xBD\x8D\x00\xEC\xBD\x8E\x00\xEC\xBD\x8F\x00\xEC"
-     "\xBD\x90"
-     "\x00\xEC\xBD\x91\x00\xEC\xBD\x92\x00\xEC\xBD\x93\x00\xEC\xBD\x94\x00\xEC"
-     "\xBD\x95"
-     "\x00\xEC\xBD\x96\x00\xEC\xBD\x97\x00\xEC\xBD\x98\x00\xEC\xBD\x99\x00\xEC"
-     "\xBD\x9A"
-     "\x00\xEC\xBD\x9B\x00\xEC\xBD\x9C\x00\xEC\xBD\x9D\x00\xEC\xBD\x9E\x00\xEC"
-     "\xBD\x9F"
-     "\x00\xEC\xBD\xA0\x00\xEC\xBD\xA1\x00\xEC\xBD\xA2\x00\xEC\xBD\xA3\x00\xEC"
-     "\xBD\xA4"
-     "\x00\xEC\xBD\xA5\x00\xEC\xBD\xA6\x00\xEC\xBD\xA7\x00\xEC\xBD\xA8\x00\xEC"
-     "\xBD\xA9"
-     "\x00\xEC\xBD\xAA\x00\xEC\xBD\xAB\x00\xEC\xBD\xAC\x00\xEC\xBD\xAD\x00\xEC"
-     "\xBD\xAE"
-     "\x00\xEC\xBD\xAF\x00\xEC\xBD\xB0\x00\xEC\xBD\xB1\x00\xEC\xBD\xB2\x00\xEC"
-     "\xBD\xB3"
-     "\x00\xEC\xBD\xB4\x00\xEC\xBD\xB5\x00\xEC\xBD\xB6\x00\xEC\xBD\xB7\x00\xEC"
-     "\xBD\xB8"
-     "\x00\xEC\xBD\xB9\x00\xEC\xBD\xBA\x00\xEC\xBD\xBB\x00\xEC\xBD\xBC\x00\xEC"
-     "\xBD\xBD"
-     "\x00\xEC\xBD\xBE\x00\xEC\xBD\xBF\x00\xEC\xBE\x80\x00\xEC\xBE\x81\x00\xEC"
-     "\xBE\x82"
-     "\x00\xEC\xBE\x83\x00\xEC\xBE\x84\x00\xEC\xBE\x85\x00\xEC\xBE\x86\x00\xEC"
-     "\xBE\x87"
-     "\x00\xEC\xBE\x88\x00\xEC\xBE\x89\x00\xEC\xBE\x8A\x00\xEC\xBE\x8B\x00\xEC"
-     "\xBE\x8C"
-     "\x00\xEC\xBE\x8D\x00\xEC\xBE\x8E\x00\xEC\xBE\x8F\x00\xEC\xBE\x90\x00\xEC"
-     "\xBE\x91"
-     "\x00\xEC\xBE\x92\x00\xEC\xBE\x93\x00\xEC\xBE\x94\x00\xEC\xBE\x95\x00\xEC"
-     "\xBE\x96"
-     "\x00\xEC\xBE\x97\x00\xEC\xBE\x98\x00\xEC\xBE\x99\x00\xEC\xBE\x9A\x00\xEC"
-     "\xBE\x9B"
-     "\x00\xEC\xBE\x9C\x00\xEC\xBE\x9D\x00\xEC\xBE\x9E\x00\xEC\xBE\x9F\x00\xEC"
-     "\xBE\xA0"
-     "\x00\xEC\xBE\xA1\x00\xEC\xBE\xA2\x00\xEC\xBE\xA3\x00\xEC\xBE\xA4\x00\xEC"
-     "\xBE\xA5"
-     "\x00\xEC\xBE\xA6\x00\xEC\xBE\xA7\x00\xEC\xBE\xA8\x00\xEC\xBE\xA9\x00\xEC"
-     "\xBE\xAA"
-     "\x00\xEC\xBE\xAB\x00\xEC\xBE\xAC\x00\xEC\xBE\xAD\x00\xEC\xBE\xAE\x00\xEC"
-     "\xBE\xAF"
-     "\x00\xEC\xBE\xB0\x00\xEC\xBE\xB1\x00\xEC\xBE\xB2\x00\xEC\xBE\xB3\x00\xEC"
-     "\xBE\xB4"
-     "\x00\xEC\xBE\xB5\x00\xEC\xBE\xB6\x00\xEC\xBE\xB7\x00\xEC\xBE\xB8\x00\xEC"
-     "\xBE\xB9"
-     "\x00\xEC\xBE\xBA\x00\xEC\xBE\xBB\x00\xEC\xBE\xBC\x00\xEC\xBE\xBD\x00\xEC"
-     "\xBE\xBE"
-     "\x00\xEC\xBE\xBF\x00\xEC\xBF\x80\x00\xEC\xBF\x81\x00\xEC\xBF\x82\x00\xEC"
-     "\xBF\x83"
-     "\x00\xEC\xBF\x84\x00\xEC\xBF\x85\x00\xEC\xBF\x86\x00\xEC\xBF\x87\x00\xEC"
-     "\xBF\x88"
-     "\x00\xEC\xBF\x89\x00\xEC\xBF\x8A\x00\xEC\xBF\x8B\x00\xEC\xBF\x8C\x00\xEC"
-     "\xBF\x8D"
-     "\x00\xEC\xBF\x8E\x00\xEC\xBF\x8F\x00\xEC\xBF\x90\x00\xEC\xBF\x91\x00\xEC"
-     "\xBF\x92"
-     "\x00\xEC\xBF\x93\x00\xEC\xBF\x94\x00\xEC\xBF\x95\x00\xEC\xBF\x96\x00\xEC"
-     "\xBF\x97"
-     "\x00\xEC\xBF\x98\x00\xEC\xBF\x99\x00\xEC\xBF\x9A\x00\xEC\xBF\x9B\x00\xEC"
-     "\xBF\x9C"
-     "\x00\xEC\xBF\x9D\x00\xEC\xBF\x9E\x00\xEC\xBF\x9F\x00\xEC\xBF\xA0\x00\xEC"
-     "\xBF\xA1"
-     "\x00\xEC\xBF\xA2\x00\xEC\xBF\xA3\x00\xEC\xBF\xA4\x00\xEC\xBF\xA5\x00\xEC"
-     "\xBF\xA6"
-     "\x00\xEC\xBF\xA7\x00\xEC\xBF\xA8\x00\xEC\xBF\xA9\x00\xEC\xBF\xAA\x00\xEC"
-     "\xBF\xAB"
-     "\x00\xEC\xBF\xAC\x00\xEC\xBF\xAD\x00\xEC\xBF\xAE\x00\xEC\xBF\xAF\x00\xEC"
-     "\xBF\xB0"
-     "\x00\xEC\xBF\xB1\x00\xEC\xBF\xB2\x00\xEC\xBF\xB3\x00\xEC\xBF\xB4\x00\xEC"
-     "\xBF\xB5"
-     "\x00\xEC\xBF\xB6\x00\xEC\xBF\xB7\x00\xEC\xBF\xB8\x00\xEC\xBF\xB9\x00\xEC"
-     "\xBF\xBA"
-     "\x00\xEC\xBF\xBB\x00\xEC\xBF\xBC\x00\xEC\xBF\xBD\x00\xEC\xBF\xBE\x00\xEC"
-     "\xBF\xBF"
-     "\x00\xED\x80\x80\x00\xED\x80\x81\x00\xED\x80\x82\x00\xED\x80\x83\x00\xED"
-     "\x80\x84"
-     "\x00\xED\x80\x85\x00\xED\x80\x86\x00\xED\x80\x87\x00\xED\x80\x88\x00\xED"
-     "\x80\x89"
-     "\x00\xED\x80\x8A\x00\xED\x80\x8B\x00\xED\x80\x8C\x00\xED\x80\x8D\x00\xED"
-     "\x80\x8E"
-     "\x00\xED\x80\x8F\x00\xED\x80\x90\x00\xED\x80\x91\x00\xED\x80\x92\x00\xED"
-     "\x80\x93"
-     "\x00\xED\x80\x94\x00\xED\x80\x95\x00\xED\x80\x96\x00\xED\x80\x97\x00\xED"
-     "\x80\x98"
-     "\x00\xED\x80\x99\x00\xED\x80\x9A\x00\xED\x80\x9B\x00\xED\x80\x9C\x00\xED"
-     "\x80\x9D"
-     "\x00\xED\x80\x9E\x00\xED\x80\x9F\x00\xED\x80\xA0\x00\xED\x80\xA1\x00\xED"
-     "\x80\xA2"
-     "\x00\xED\x80\xA3\x00\xED\x80\xA4\x00\xED\x80\xA5\x00\xED\x80\xA6\x00\xED"
-     "\x80\xA7"
-     "\x00\xED\x80\xA8\x00\xED\x80\xA9\x00\xED\x80\xAA\x00\xED\x80\xAB\x00\xED"
-     "\x80\xAC"
-     "\x00\xED\x80\xAD\x00\xED\x80\xAE\x00\xED\x80\xAF\x00\xED\x80\xB0\x00\xED"
-     "\x80\xB1"
-     "\x00\xED\x80\xB2\x00\xED\x80\xB3\x00\xED\x80\xB4\x00\xED\x80\xB5\x00\xED"
-     "\x80\xB6"
-     "\x00\xED\x80\xB7\x00\xED\x80\xB8\x00\xED\x80\xB9\x00\xED\x80\xBA\x00\xED"
-     "\x80\xBB"
-     "\x00\xED\x80\xBC\x00\xED\x80\xBD\x00\xED\x80\xBE\x00\xED\x80\xBF\x00\xED"
-     "\x81\x80"
-     "\x00\xED\x81\x81\x00\xED\x81\x82\x00\xED\x81\x83\x00\xED\x81\x84\x00\xED"
-     "\x81\x85"
-     "\x00\xED\x81\x86\x00\xED\x81\x87\x00\xED\x81\x88\x00\xED\x81\x89\x00\xED"
-     "\x81\x8A"
-     "\x00\xED\x81\x8B\x00\xED\x81\x8C\x00\xED\x81\x8D\x00\xED\x81\x8E\x00\xED"
-     "\x81\x8F"
-     "\x00\xED\x81\x90\x00\xED\x81\x91\x00\xED\x81\x92\x00\xED\x81\x93\x00\xED"
-     "\x81\x94"
-     "\x00\xED\x81\x95\x00\xED\x81\x96\x00\xED\x81\x97\x00\xED\x81\x98\x00\xED"
-     "\x81\x99"
-     "\x00\xED\x81\x9A\x00\xED\x81\x9B\x00\xED\x81\x9C\x00\xED\x81\x9D\x00\xED"
-     "\x81\x9E"
-     "\x00\xED\x81\x9F\x00\xED\x81\xA0\x00\xED\x81\xA1\x00\xED\x81\xA2\x00\xED"
-     "\x81\xA3"
-     "\x00\xED\x81\xA4\x00\xED\x81\xA5\x00\xED\x81\xA6\x00\xED\x81\xA7\x00\xED"
-     "\x81\xA8"
-     "\x00\xED\x81\xA9\x00\xED\x81\xAA\x00\xED\x81\xAB\x00\xED\x81\xAC\x00\xED"
-     "\x81\xAD"
-     "\x00\xED\x81\xAE\x00\xED\x81\xAF\x00\xED\x81\xB0\x00\xED\x81\xB1\x00\xED"
-     "\x81\xB2"
-     "\x00\xED\x81\xB3\x00\xED\x81\xB4\x00\xED\x81\xB5\x00\xED\x81\xB6\x00\xED"
-     "\x81\xB7"
-     "\x00\xED\x81\xB8\x00\xED\x81\xB9\x00\xED\x81\xBA\x00\xED\x81\xBB\x00\xED"
-     "\x81\xBC"
-     "\x00\xED\x81\xBD\x00\xED\x81\xBE\x00\xED\x81\xBF\x00\xED\x82\x80\x00\xED"
-     "\x82\x81"
-     "\x00\xED\x82\x82\x00\xED\x82\x83\x00\xED\x82\x84\x00\xED\x82\x85\x00\xED"
-     "\x82\x86"
-     "\x00\xED\x82\x87\x00\xED\x82\x88\x00\xED\x82\x89\x00\xED\x82\x8A\x00\xED"
-     "\x82\x8B"
-     "\x00\xED\x82\x8C\x00\xED\x82\x8D\x00\xED\x82\x8E\x00\xED\x82\x8F\x00\xED"
-     "\x82\x90"
-     "\x00\xED\x82\x91\x00\xED\x82\x92\x00\xED\x82\x93\x00\xED\x82\x94\x00\xED"
-     "\x82\x95"
-     "\x00\xED\x82\x96\x00\xED\x82\x97\x00\xED\x82\x98\x00\xED\x82\x99\x00\xED"
-     "\x82\x9A"
-     "\x00\xED\x82\x9B\x00\xED\x82\x9C\x00\xED\x82\x9D\x00\xED\x82\x9E\x00\xED"
-     "\x82\x9F"
-     "\x00\xED\x82\xA0\x00\xED\x82\xA1\x00\xED\x82\xA2\x00\xED\x82\xA3\x00\xED"
-     "\x82\xA4"
-     "\x00\xED\x82\xA5\x00\xED\x82\xA6\x00\xED\x82\xA7\x00\xED\x82\xA8\x00\xED"
-     "\x82\xA9"
-     "\x00\xED\x82\xAA\x00\xED\x82\xAB\x00\xED\x82\xAC\x00\xED\x82\xAD\x00\xED"
-     "\x82\xAE"
-     "\x00\xED\x82\xAF\x00\xED\x82\xB0\x00\xED\x82\xB1\x00\xED\x82\xB2\x00\xED"
-     "\x82\xB3"
-     "\x00\xED\x82\xB4\x00\xED\x82\xB5\x00\xED\x82\xB6\x00\xED\x82\xB7\x00\xED"
-     "\x82\xB8"
-     "\x00\xED\x82\xB9\x00\xED\x82\xBA\x00\xED\x82\xBB\x00\xED\x82\xBC\x00\xED"
-     "\x82\xBD"
-     "\x00\xED\x82\xBE\x00\xED\x82\xBF\x00\xED\x83\x80\x00\xED\x83\x81\x00\xED"
-     "\x83\x82"
-     "\x00\xED\x83\x83\x00\xED\x83\x84\x00\xED\x83\x85\x00\xED\x83\x86\x00\xED"
-     "\x83\x87"
-     "\x00\xED\x83\x88\x00\xED\x83\x89\x00\xED\x83\x8A\x00\xED\x83\x8B\x00\xED"
-     "\x83\x8C"
-     "\x00\xED\x83\x8D\x00\xED\x83\x8E\x00\xED\x83\x8F\x00\xED\x83\x90\x00\xED"
-     "\x83\x91"
-     "\x00\xED\x83\x92\x00\xED\x83\x93\x00\xED\x83\x94\x00\xED\x83\x95\x00\xED"
-     "\x83\x96"
-     "\x00\xED\x83\x97\x00\xED\x83\x98\x00\xED\x83\x99\x00\xED\x83\x9A\x00\xED"
-     "\x83\x9B"
-     "\x00\xED\x83\x9C\x00\xED\x83\x9D\x00\xED\x83\x9E\x00\xED\x83\x9F\x00\xED"
-     "\x83\xA0"
-     "\x00\xED\x83\xA1\x00\xED\x83\xA2\x00\xED\x83\xA3\x00\xED\x83\xA4\x00\xED"
-     "\x83\xA5"
-     "\x00\xED\x83\xA6\x00\xED\x83\xA7\x00\xED\x83\xA8\x00\xED\x83\xA9\x00\xED"
-     "\x83\xAA"
-     "\x00\xED\x83\xAB\x00\xED\x83\xAC\x00\xED\x83\xAD\x00\xED\x83\xAE\x00\xED"
-     "\x83\xAF"
-     "\x00\xED\x83\xB0\x00\xED\x83\xB1\x00\xED\x83\xB2\x00\xED\x83\xB3\x00\xED"
-     "\x83\xB4"
-     "\x00\xED\x83\xB5\x00\xED\x83\xB6\x00\xED\x83\xB7\x00\xED\x83\xB8\x00\xED"
-     "\x83\xB9"
-     "\x00\xED\x83\xBA\x00\xED\x83\xBB\x00\xED\x83\xBC\x00\xED\x83\xBD\x00\xED"
-     "\x83\xBE"
-     "\x00\xED\x83\xBF\x00\xED\x84\x80\x00\xED\x84\x81\x00\xED\x84\x82\x00\xED"
-     "\x84\x83"
-     "\x00\xED\x84\x84\x00\xED\x84\x85\x00\xED\x84\x86\x00\xED\x84\x87\x00\xED"
-     "\x84\x88"
-     "\x00\xED\x84\x89\x00\xED\x84\x8A\x00\xED\x84\x8B\x00\xED\x84\x8C\x00\xED"
-     "\x84\x8D"
-     "\x00\xED\x84\x8E\x00\xED\x84\x8F\x00\xED\x84\x90\x00\xED\x84\x91\x00\xED"
-     "\x84\x92"
-     "\x00\xED\x84\x93\x00\xED\x84\x94\x00\xED\x84\x95\x00\xED\x84\x96\x00\xED"
-     "\x84\x97"
-     "\x00\xED\x84\x98\x00\xED\x84\x99\x00\xED\x84\x9A\x00\xED\x84\x9B\x00\xED"
-     "\x84\x9C"
-     "\x00\xED\x84\x9D\x00\xED\x84\x9E\x00\xED\x84\x9F\x00\xED\x84\xA0\x00\xED"
-     "\x84\xA1"
-     "\x00\xED\x84\xA2\x00\xED\x84\xA3\x00\xED\x84\xA4\x00\xED\x84\xA5\x00\xED"
-     "\x84\xA6"
-     "\x00\xED\x84\xA7\x00\xED\x84\xA8\x00\xED\x84\xA9\x00\xED\x84\xAA\x00\xED"
-     "\x84\xAB"
-     "\x00\xED\x84\xAC\x00\xED\x84\xAD\x00\xED\x84\xAE\x00\xED\x84\xAF\x00\xED"
-     "\x84\xB0"
-     "\x00\xED\x84\xB1\x00\xED\x84\xB2\x00\xED\x84\xB3\x00\xED\x84\xB4\x00\xED"
-     "\x84\xB5"
-     "\x00\xED\x84\xB6\x00\xED\x84\xB7\x00\xED\x84\xB8\x00\xED\x84\xB9\x00\xED"
-     "\x84\xBA"
-     "\x00\xED\x84\xBB\x00\xED\x84\xBC\x00\xED\x84\xBD\x00\xED\x84\xBE\x00\xED"
-     "\x84\xBF"
-     "\x00\xED\x85\x80\x00\xED\x85\x81\x00\xED\x85\x82\x00\xED\x85\x83\x00\xED"
-     "\x85\x84"
-     "\x00\xED\x85\x85\x00\xED\x85\x86\x00\xED\x85\x87\x00\xED\x85\x88\x00\xED"
-     "\x85\x89"
-     "\x00\xED\x85\x8A\x00\xED\x85\x8B\x00\xED\x85\x8C\x00\xED\x85\x8D\x00\xED"
-     "\x85\x8E"
-     "\x00\xED\x85\x8F\x00\xED\x85\x90\x00\xED\x85\x91\x00\xED\x85\x92\x00\xED"
-     "\x85\x93"
-     "\x00\xED\x85\x94\x00\xED\x85\x95\x00\xED\x85\x96\x00\xED\x85\x97\x00\xED"
-     "\x85\x98"
-     "\x00\xED\x85\x99\x00\xED\x85\x9A\x00\xED\x85\x9B\x00\xED\x85\x9C\x00\xED"
-     "\x85\x9D"
-     "\x00\xED\x85\x9E\x00\xED\x85\x9F\x00\xED\x85\xA0\x00\xED\x85\xA1\x00\xED"
-     "\x85\xA2"
-     "\x00\xED\x85\xA3\x00\xED\x85\xA4\x00\xED\x85\xA5\x00\xED\x85\xA6\x00\xED"
-     "\x85\xA7"
-     "\x00\xED\x85\xA8\x00\xED\x85\xA9\x00\xED\x85\xAA\x00\xED\x85\xAB\x00\xED"
-     "\x85\xAC"
-     "\x00\xED\x85\xAD\x00\xED\x85\xAE\x00\xED\x85\xAF\x00\xED\x85\xB0\x00\xED"
-     "\x85\xB1"
-     "\x00\xED\x85\xB2\x00\xED\x85\xB3\x00\xED\x85\xB4\x00\xED\x85\xB5\x00\xED"
-     "\x85\xB6"
-     "\x00\xED\x85\xB7\x00\xED\x85\xB8\x00\xED\x85\xB9\x00\xED\x85\xBA\x00\xED"
-     "\x85\xBB"
-     "\x00\xED\x85\xBC\x00\xED\x85\xBD\x00\xED\x85\xBE\x00\xED\x85\xBF\x00\xED"
-     "\x86\x80"
-     "\x00\xED\x86\x81\x00\xED\x86\x82\x00\xED\x86\x83\x00\xED\x86\x84\x00\xED"
-     "\x86\x85"
-     "\x00\xED\x86\x86\x00\xED\x86\x87\x00\xED\x86\x88\x00\xED\x86\x89\x00\xED"
-     "\x86\x8A"
-     "\x00\xED\x86\x8B\x00\xED\x86\x8C\x00\xED\x86\x8D\x00\xED\x86\x8E\x00\xED"
-     "\x86\x8F"
-     "\x00\xED\x86\x90\x00\xED\x86\x91\x00\xED\x86\x92\x00\xED\x86\x93\x00\xED"
-     "\x86\x94"
-     "\x00\xED\x86\x95\x00\xED\x86\x96\x00\xED\x86\x97\x00\xED\x86\x98\x00\xED"
-     "\x86\x99"
-     "\x00\xED\x86\x9A\x00\xED\x86\x9B\x00\xED\x86\x9C\x00\xED\x86\x9D\x00\xED"
-     "\x86\x9E"
-     "\x00\xED\x86\x9F\x00\xED\x86\xA0\x00\xED\x86\xA1\x00\xED\x86\xA2\x00\xED"
-     "\x86\xA3"
-     "\x00\xED\x86\xA4\x00\xED\x86\xA5\x00\xED\x86\xA6\x00\xED\x86\xA7\x00\xED"
-     "\x86\xA8"
-     "\x00\xED\x86\xA9\x00\xED\x86\xAA\x00\xED\x86\xAB\x00\xED\x86\xAC\x00\xED"
-     "\x86\xAD"
-     "\x00\xED\x86\xAE\x00\xED\x86\xAF\x00\xED\x86\xB0\x00\xED\x86\xB1\x00\xED"
-     "\x86\xB2"
-     "\x00\xED\x86\xB3\x00\xED\x86\xB4\x00\xED\x86\xB5\x00\xED\x86\xB6\x00\xED"
-     "\x86\xB7"
-     "\x00\xED\x86\xB8\x00\xED\x86\xB9\x00\xED\x86\xBA\x00\xED\x86\xBB\x00\xED"
-     "\x86\xBC"
-     "\x00\xED\x86\xBD\x00\xED\x86\xBE\x00\xED\x86\xBF\x00\xED\x87\x80\x00\xED"
-     "\x87\x81"
-     "\x00\xED\x87\x82\x00\xED\x87\x83\x00\xED\x87\x84\x00\xED\x87\x85\x00\xED"
-     "\x87\x86"
-     "\x00\xED\x87\x87\x00\xED\x87\x88\x00\xED\x87\x89\x00\xED\x87\x8A\x00\xED"
-     "\x87\x8B"
-     "\x00\xED\x87\x8C\x00\xED\x87\x8D\x00\xED\x87\x8E\x00\xED\x87\x8F\x00\xED"
-     "\x87\x90"
-     "\x00\xED\x87\x91\x00\xED\x87\x92\x00\xED\x87\x93\x00\xED\x87\x94\x00\xED"
-     "\x87\x95"
-     "\x00\xED\x87\x96\x00\xED\x87\x97\x00\xED\x87\x98\x00\xED\x87\x99\x00\xED"
-     "\x87\x9A"
-     "\x00\xED\x87\x9B\x00\xED\x87\x9C\x00\xED\x87\x9D\x00\xED\x87\x9E\x00\xED"
-     "\x87\x9F"
-     "\x00\xED\x87\xA0\x00\xED\x87\xA1\x00\xED\x87\xA2\x00\xED\x87\xA3\x00\xED"
-     "\x87\xA4"
-     "\x00\xED\x87\xA5\x00\xED\x87\xA6\x00\xED\x87\xA7\x00\xED\x87\xA8\x00\xED"
-     "\x87\xA9"
-     "\x00\xED\x87\xAA\x00\xED\x87\xAB\x00\xED\x87\xAC\x00\xED\x87\xAD\x00\xED"
-     "\x87\xAE"
-     "\x00\xED\x87\xAF\x00\xED\x87\xB0\x00\xED\x87\xB1\x00\xED\x87\xB2\x00\xED"
-     "\x87\xB3"
-     "\x00\xED\x87\xB4\x00\xED\x87\xB5\x00\xED\x87\xB6\x00\xED\x87\xB7\x00\xED"
-     "\x87\xB8"
-     "\x00\xED\x87\xB9\x00\xED\x87\xBA\x00\xED\x87\xBB\x00\xED\x87\xBC\x00\xED"
-     "\x87\xBD"
-     "\x00\xED\x87\xBE\x00\xED\x87\xBF\x00\xED\x88\x80\x00\xED\x88\x81\x00\xED"
-     "\x88\x82"
-     "\x00\xED\x88\x83\x00\xED\x88\x84\x00\xED\x88\x85\x00\xED\x88\x86\x00\xED"
-     "\x88\x87"
-     "\x00\xED\x88\x88\x00\xED\x88\x89\x00\xED\x88\x8A\x00\xED\x88\x8B\x00\xED"
-     "\x88\x8C"
-     "\x00\xED\x88\x8D\x00\xED\x88\x8E\x00\xED\x88\x8F\x00\xED\x88\x90\x00\xED"
-     "\x88\x91"
-     "\x00\xED\x88\x92\x00\xED\x88\x93\x00\xED\x88\x94\x00\xED\x88\x95\x00\xED"
-     "\x88\x96"
-     "\x00\xED\x88\x97\x00\xED\x88\x98\x00\xED\x88\x99\x00\xED\x88\x9A\x00\xED"
-     "\x88\x9B"
-     "\x00\xED\x88\x9C\x00\xED\x88\x9D\x00\xED\x88\x9E\x00\xED\x88\x9F\x00\xED"
-     "\x88\xA0"
-     "\x00\xED\x88\xA1\x00\xED\x88\xA2\x00\xED\x88\xA3\x00\xED\x88\xA4\x00\xED"
-     "\x88\xA5"
-     "\x00\xED\x88\xA6\x00\xED\x88\xA7\x00\xED\x88\xA8\x00\xED\x88\xA9\x00\xED"
-     "\x88\xAA"
-     "\x00\xED\x88\xAB\x00\xED\x88\xAC\x00\xED\x88\xAD\x00\xED\x88\xAE\x00\xED"
-     "\x88\xAF"
-     "\x00\xED\x88\xB0\x00\xED\x88\xB1\x00\xED\x88\xB2\x00\xED\x88\xB3\x00\xED"
-     "\x88\xB4"
-     "\x00\xED\x88\xB5\x00\xED\x88\xB6\x00\xED\x88\xB7\x00\xED\x88\xB8\x00\xED"
-     "\x88\xB9"
-     "\x00\xED\x88\xBA\x00\xED\x88\xBB\x00\xED\x88\xBC\x00\xED\x88\xBD\x00\xED"
-     "\x88\xBE"
-     "\x00\xED\x88\xBF\x00\xED\x89\x80\x00\xED\x89\x81\x00\xED\x89\x82\x00\xED"
-     "\x89\x83"
-     "\x00\xED\x89\x84\x00\xED\x89\x85\x00\xED\x89\x86\x00\xED\x89\x87\x00\xED"
-     "\x89\x88"
-     "\x00\xED\x89\x89\x00\xED\x89\x8A\x00\xED\x89\x8B\x00\xED\x89\x8C\x00\xED"
-     "\x89\x8D"
-     "\x00\xED\x89\x8E\x00\xED\x89\x8F\x00\xED\x89\x90\x00\xED\x89\x91\x00\xED"
-     "\x89\x92"
-     "\x00\xED\x89\x93\x00\xED\x89\x94\x00\xED\x89\x95\x00\xED\x89\x96\x00\xED"
-     "\x89\x97"
-     "\x00\xED\x89\x98\x00\xED\x89\x99\x00\xED\x89\x9A\x00\xED\x89\x9B\x00\xED"
-     "\x89\x9C"
-     "\x00\xED\x89\x9D\x00\xED\x89\x9E\x00\xED\x89\x9F\x00\xED\x89\xA0\x00\xED"
-     "\x89\xA1"
-     "\x00\xED\x89\xA2\x00\xED\x89\xA3\x00\xED\x89\xA4\x00\xED\x89\xA5\x00\xED"
-     "\x89\xA6"
-     "\x00\xED\x89\xA7\x00\xED\x89\xA8\x00\xED\x89\xA9\x00\xED\x89\xAA\x00\xED"
-     "\x89\xAB"
-     "\x00\xED\x89\xAC\x00\xED\x89\xAD\x00\xED\x89\xAE\x00\xED\x89\xAF\x00\xED"
-     "\x89\xB0"
-     "\x00\xED\x89\xB1\x00\xED\x89\xB2\x00\xED\x89\xB3\x00\xED\x89\xB4\x00\xED"
-     "\x89\xB5"
-     "\x00\xED\x89\xB6\x00\xED\x89\xB7\x00\xED\x89\xB8\x00\xED\x89\xB9\x00\xED"
-     "\x89\xBA"
-     "\x00\xED\x89\xBB\x00\xED\x89\xBC\x00\xED\x89\xBD\x00\xED\x89\xBE\x00\xED"
-     "\x89\xBF"
-     "\x00\xED\x8A\x80\x00\xED\x8A\x81\x00\xED\x8A\x82\x00\xED\x8A\x83\x00\xED"
-     "\x8A\x84"
-     "\x00\xED\x8A\x85\x00\xED\x8A\x86\x00\xED\x8A\x87\x00\xED\x8A\x88\x00\xED"
-     "\x8A\x89"
-     "\x00\xED\x8A\x8A\x00\xED\x8A\x8B\x00\xED\x8A\x8C\x00\xED\x8A\x8D\x00\xED"
-     "\x8A\x8E"
-     "\x00\xED\x8A\x8F\x00\xED\x8A\x90\x00\xED\x8A\x91\x00\xED\x8A\x92\x00\xED"
-     "\x8A\x93"
-     "\x00\xED\x8A\x94\x00\xED\x8A\x95\x00\xED\x8A\x96\x00\xED\x8A\x97\x00\xED"
-     "\x8A\x98"
-     "\x00\xED\x8A\x99\x00\xED\x8A\x9A\x00\xED\x8A\x9B\x00\xED\x8A\x9C\x00\xED"
-     "\x8A\x9D"
-     "\x00\xED\x8A\x9E\x00\xED\x8A\x9F\x00\xED\x8A\xA0\x00\xED\x8A\xA1\x00\xED"
-     "\x8A\xA2"
-     "\x00\xED\x8A\xA3\x00\xED\x8A\xA4\x00\xED\x8A\xA5\x00\xED\x8A\xA6\x00\xED"
-     "\x8A\xA7"
-     "\x00\xED\x8A\xA8\x00\xED\x8A\xA9\x00\xED\x8A\xAA\x00\xED\x8A\xAB\x00\xED"
-     "\x8A\xAC"
-     "\x00\xED\x8A\xAD\x00\xED\x8A\xAE\x00\xED\x8A\xAF\x00\xED\x8A\xB0\x00\xED"
-     "\x8A\xB1"
-     "\x00\xED\x8A\xB2\x00\xED\x8A\xB3\x00\xED\x8A\xB4\x00\xED\x8A\xB5\x00\xED"
-     "\x8A\xB6"
-     "\x00\xED\x8A\xB7\x00\xED\x8A\xB8\x00\xED\x8A\xB9\x00\xED\x8A\xBA\x00\xED"
-     "\x8A\xBB"
-     "\x00\xED\x8A\xBC\x00\xED\x8A\xBD\x00\xED\x8A\xBE\x00\xED\x8A\xBF\x00\xED"
-     "\x8B\x80"
-     "\x00\xED\x8B\x81\x00\xED\x8B\x82\x00\xED\x8B\x83\x00\xED\x8B\x84\x00\xED"
-     "\x8B\x85"
-     "\x00\xED\x8B\x86\x00\xED\x8B\x87\x00\xED\x8B\x88\x00\xED\x8B\x89\x00\xED"
-     "\x8B\x8A"
-     "\x00\xED\x8B\x8B\x00\xED\x8B\x8C\x00\xED\x8B\x8D\x00\xED\x8B\x8E\x00\xED"
-     "\x8B\x8F"
-     "\x00\xED\x8B\x90\x00\xED\x8B\x91\x00\xED\x8B\x92\x00\xED\x8B\x93\x00\xED"
-     "\x8B\x94"
-     "\x00\xED\x8B\x95\x00\xED\x8B\x96\x00\xED\x8B\x97\x00\xED\x8B\x98\x00\xED"
-     "\x8B\x99"
-     "\x00\xED\x8B\x9A\x00\xED\x8B\x9B\x00\xED\x8B\x9C\x00\xED\x8B\x9D\x00\xED"
-     "\x8B\x9E"
-     "\x00\xED\x8B\x9F\x00\xED\x8B\xA0\x00\xED\x8B\xA1\x00\xED\x8B\xA2\x00\xED"
-     "\x8B\xA3"
-     "\x00\xED\x8B\xA4\x00\xED\x8B\xA5\x00\xED\x8B\xA6\x00\xED\x8B\xA7\x00\xED"
-     "\x8B\xA8"
-     "\x00\xED\x8B\xA9\x00\xED\x8B\xAA\x00\xED\x8B\xAB\x00\xED\x8B\xAC\x00\xED"
-     "\x8B\xAD"
-     "\x00\xED\x8B\xAE\x00\xED\x8B\xAF\x00\xED\x8B\xB0\x00\xED\x8B\xB1\x00\xED"
-     "\x8B\xB2"
-     "\x00\xED\x8B\xB3\x00\xED\x8B\xB4\x00\xED\x8B\xB5\x00\xED\x8B\xB6\x00\xED"
-     "\x8B\xB7"
-     "\x00\xED\x8B\xB8\x00\xED\x8B\xB9\x00\xED\x8B\xBA\x00\xED\x8B\xBB\x00\xED"
-     "\x8B\xBC"
-     "\x00\xED\x8B\xBD\x00\xED\x8B\xBE\x00\xED\x8B\xBF\x00\xED\x8C\x80\x00\xED"
-     "\x8C\x81"
-     "\x00\xED\x8C\x82\x00\xED\x8C\x83\x00\xED\x8C\x84\x00\xED\x8C\x85\x00\xED"
-     "\x8C\x86"
-     "\x00\xED\x8C\x87\x00\xED\x8C\x88\x00\xED\x8C\x89\x00\xED\x8C\x8A\x00\xED"
-     "\x8C\x8B"
-     "\x00\xED\x8C\x8C\x00\xED\x8C\x8D\x00\xED\x8C\x8E\x00\xED\x8C\x8F\x00\xED"
-     "\x8C\x90"
-     "\x00\xED\x8C\x91\x00\xED\x8C\x92\x00\xED\x8C\x93\x00\xED\x8C\x94\x00\xED"
-     "\x8C\x95"
-     "\x00\xED\x8C\x96\x00\xED\x8C\x97\x00\xED\x8C\x98\x00\xED\x8C\x99\x00\xED"
-     "\x8C\x9A"
-     "\x00\xED\x8C\x9B\x00\xED\x8C\x9C\x00\xED\x8C\x9D\x00\xED\x8C\x9E\x00\xED"
-     "\x8C\x9F"
-     "\x00\xED\x8C\xA0\x00\xED\x8C\xA1\x00\xED\x8C\xA2\x00\xED\x8C\xA3\x00\xED"
-     "\x8C\xA4"
-     "\x00\xED\x8C\xA5\x00\xED\x8C\xA6\x00\xED\x8C\xA7\x00\xED\x8C\xA8\x00\xED"
-     "\x8C\xA9"
-     "\x00\xED\x8C\xAA\x00\xED\x8C\xAB\x00\xED\x8C\xAC\x00\xED\x8C\xAD\x00\xED"
-     "\x8C\xAE"
-     "\x00\xED\x8C\xAF\x00\xED\x8C\xB0\x00\xED\x8C\xB1\x00\xED\x8C\xB2\x00\xED"
-     "\x8C\xB3"
-     "\x00\xED\x8C\xB4\x00\xED\x8C\xB5\x00\xED\x8C\xB6\x00\xED\x8C\xB7\x00\xED"
-     "\x8C\xB8"
-     "\x00\xED\x8C\xB9\x00\xED\x8C\xBA\x00\xED\x8C\xBB\x00\xED\x8C\xBC\x00\xED"
-     "\x8C\xBD"
-     "\x00\xED\x8C\xBE\x00\xED\x8C\xBF\x00\xED\x8D\x80\x00\xED\x8D\x81\x00\xED"
-     "\x8D\x82"
-     "\x00\xED\x8D\x83\x00\xED\x8D\x84\x00\xED\x8D\x85\x00\xED\x8D\x86\x00\xED"
-     "\x8D\x87"
-     "\x00\xED\x8D\x88\x00\xED\x8D\x89\x00\xED\x8D\x8A\x00\xED\x8D\x8B\x00\xED"
-     "\x8D\x8C"
-     "\x00\xED\x8D\x8D\x00\xED\x8D\x8E\x00\xED\x8D\x8F\x00\xED\x8D\x90\x00\xED"
-     "\x8D\x91"
-     "\x00\xED\x8D\x92\x00\xED\x8D\x93\x00\xED\x8D\x94\x00\xED\x8D\x95\x00\xED"
-     "\x8D\x96"
-     "\x00\xED\x8D\x97\x00\xED\x8D\x98\x00\xED\x8D\x99\x00\xED\x8D\x9A\x00\xED"
-     "\x8D\x9B"
-     "\x00\xED\x8D\x9C\x00\xED\x8D\x9D\x00\xED\x8D\x9E\x00\xED\x8D\x9F\x00\xED"
-     "\x8D\xA0"
-     "\x00\xED\x8D\xA1\x00\xED\x8D\xA2\x00\xED\x8D\xA3\x00\xED\x8D\xA4\x00\xED"
-     "\x8D\xA5"
-     "\x00\xED\x8D\xA6\x00\xED\x8D\xA7\x00\xED\x8D\xA8\x00\xED\x8D\xA9\x00\xED"
-     "\x8D\xAA"
-     "\x00\xED\x8D\xAB\x00\xED\x8D\xAC\x00\xED\x8D\xAD\x00\xED\x8D\xAE\x00\xED"
-     "\x8D\xAF"
-     "\x00\xED\x8D\xB0\x00\xED\x8D\xB1\x00\xED\x8D\xB2\x00\xED\x8D\xB3\x00\xED"
-     "\x8D\xB4"
-     "\x00\xED\x8D\xB5\x00\xED\x8D\xB6\x00\xED\x8D\xB7\x00\xED\x8D\xB8\x00\xED"
-     "\x8D\xB9"
-     "\x00\xED\x8D\xBA\x00\xED\x8D\xBB\x00\xED\x8D\xBC\x00\xED\x8D\xBD\x00\xED"
-     "\x8D\xBE"
-     "\x00\xED\x8D\xBF\x00\xED\x8E\x80\x00\xED\x8E\x81\x00\xED\x8E\x82\x00\xED"
-     "\x8E\x83"
-     "\x00\xED\x8E\x84\x00\xED\x8E\x85\x00\xED\x8E\x86\x00\xED\x8E\x87\x00\xED"
-     "\x8E\x88"
-     "\x00\xED\x8E\x89\x00\xED\x8E\x8A\x00\xED\x8E\x8B\x00\xED\x8E\x8C\x00\xED"
-     "\x8E\x8D"
-     "\x00\xED\x8E\x8E\x00\xED\x8E\x8F\x00\xED\x8E\x90\x00\xED\x8E\x91\x00\xED"
-     "\x8E\x92"
-     "\x00\xED\x8E\x93\x00\xED\x8E\x94\x00\xED\x8E\x95\x00\xED\x8E\x96\x00\xED"
-     "\x8E\x97"
-     "\x00\xED\x8E\x98\x00\xED\x8E\x99\x00\xED\x8E\x9A\x00\xED\x8E\x9B\x00\xED"
-     "\x8E\x9C"
-     "\x00\xED\x8E\x9D\x00\xED\x8E\x9E\x00\xED\x8E\x9F\x00\xED\x8E\xA0\x00\xED"
-     "\x8E\xA1"
-     "\x00\xED\x8E\xA2\x00\xED\x8E\xA3\x00\xED\x8E\xA4\x00\xED\x8E\xA5\x00\xED"
-     "\x8E\xA6"
-     "\x00\xED\x8E\xA7\x00\xED\x8E\xA8\x00\xED\x8E\xA9\x00\xED\x8E\xAA\x00\xED"
-     "\x8E\xAB"
-     "\x00\xED\x8E\xAC\x00\xED\x8E\xAD\x00\xED\x8E\xAE\x00\xED\x8E\xAF\x00\xED"
-     "\x8E\xB0"
-     "\x00\xED\x8E\xB1\x00\xED\x8E\xB2\x00\xED\x8E\xB3\x00\xED\x8E\xB4\x00\xED"
-     "\x8E\xB5"
-     "\x00\xED\x8E\xB6\x00\xED\x8E\xB7\x00\xED\x8E\xB8\x00\xED\x8E\xB9\x00\xED"
-     "\x8E\xBA"
-     "\x00\xED\x8E\xBB\x00\xED\x8E\xBC\x00\xED\x8E\xBD\x00\xED\x8E\xBE\x00\xED"
-     "\x8E\xBF"
-     "\x00\xED\x8F\x80\x00\xED\x8F\x81\x00\xED\x8F\x82\x00\xED\x8F\x83\x00\xED"
-     "\x8F\x84"
-     "\x00\xED\x8F\x85\x00\xED\x8F\x86\x00\xED\x8F\x87\x00\xED\x8F\x88\x00\xED"
-     "\x8F\x89"
-     "\x00\xED\x8F\x8A\x00\xED\x8F\x8B\x00\xED\x8F\x8C\x00\xED\x8F\x8D\x00\xED"
-     "\x8F\x8E"
-     "\x00\xED\x8F\x8F\x00\xED\x8F\x90\x00\xED\x8F\x91\x00\xED\x8F\x92\x00\xED"
-     "\x8F\x93"
-     "\x00\xED\x8F\x94\x00\xED\x8F\x95\x00\xED\x8F\x96\x00\xED\x8F\x97\x00\xED"
-     "\x8F\x98"
-     "\x00\xED\x8F\x99\x00\xED\x8F\x9A\x00\xED\x8F\x9B\x00\xED\x8F\x9C\x00\xED"
-     "\x8F\x9D"
-     "\x00\xED\x8F\x9E\x00\xED\x8F\x9F\x00\xED\x8F\xA0\x00\xED\x8F\xA1\x00\xED"
-     "\x8F\xA2"
-     "\x00\xED\x8F\xA3\x00\xED\x8F\xA4\x00\xED\x8F\xA5\x00\xED\x8F\xA6\x00\xED"
-     "\x8F\xA7"
-     "\x00\xED\x8F\xA8\x00\xED\x8F\xA9\x00\xED\x8F\xAA\x00\xED\x8F\xAB\x00\xED"
-     "\x8F\xAC"
-     "\x00\xED\x8F\xAD\x00\xED\x8F\xAE\x00\xED\x8F\xAF\x00\xED\x8F\xB0\x00\xED"
-     "\x8F\xB1"
-     "\x00\xED\x8F\xB2\x00\xED\x8F\xB3\x00\xED\x8F\xB4\x00\xED\x8F\xB5\x00\xED"
-     "\x8F\xB6"
-     "\x00\xED\x8F\xB7\x00\xED\x8F\xB8\x00\xED\x8F\xB9\x00\xED\x8F\xBA\x00\xED"
-     "\x8F\xBB"
-     "\x00\xED\x8F\xBC\x00\xED\x8F\xBD\x00\xED\x8F\xBE\x00\xED\x8F\xBF\x00\xED"
-     "\x90\x80"
-     "\x00\xED\x90\x81\x00\xED\x90\x82\x00\xED\x90\x83\x00\xED\x90\x84\x00\xED"
-     "\x90\x85"
-     "\x00\xED\x90\x86\x00\xED\x90\x87\x00\xED\x90\x88\x00\xED\x90\x89\x00\xED"
-     "\x90\x8A"
-     "\x00\xED\x90\x8B\x00\xED\x90\x8C\x00\xED\x90\x8D\x00\xED\x90\x8E\x00\xED"
-     "\x90\x8F"
-     "\x00\xED\x90\x90\x00\xED\x90\x91\x00\xED\x90\x92\x00\xED\x90\x93\x00\xED"
-     "\x90\x94"
-     "\x00\xED\x90\x95\x00\xED\x90\x96\x00\xED\x90\x97\x00\xED\x90\x98\x00\xED"
-     "\x90\x99"
-     "\x00\xED\x90\x9A\x00\xED\x90\x9B\x00\xED\x90\x9C\x00\xED\x90\x9D\x00\xED"
-     "\x90\x9E"
-     "\x00\xED\x90\x9F\x00\xED\x90\xA0\x00\xED\x90\xA1\x00\xED\x90\xA2\x00\xED"
-     "\x90\xA3"
-     "\x00\xED\x90\xA4\x00\xED\x90\xA5\x00\xED\x90\xA6\x00\xED\x90\xA7\x00\xED"
-     "\x90\xA8"
-     "\x00\xED\x90\xA9\x00\xED\x90\xAA\x00\xED\x90\xAB\x00\xED\x90\xAC\x00\xED"
-     "\x90\xAD"
-     "\x00\xED\x90\xAE\x00\xED\x90\xAF\x00\xED\x90\xB0\x00\xED\x90\xB1\x00\xED"
-     "\x90\xB2"
-     "\x00\xED\x90\xB3\x00\xED\x90\xB4\x00\xED\x90\xB5\x00\xED\x90\xB6\x00\xED"
-     "\x90\xB7"
-     "\x00\xED\x90\xB8\x00\xED\x90\xB9\x00\xED\x90\xBA\x00\xED\x90\xBB\x00\xED"
-     "\x90\xBC"
-     "\x00\xED\x90\xBD\x00\xED\x90\xBE\x00\xED\x90\xBF\x00\xED\x91\x80\x00\xED"
-     "\x91\x81"
-     "\x00\xED\x91\x82\x00\xED\x91\x83\x00\xED\x91\x84\x00\xED\x91\x85\x00\xED"
-     "\x91\x86"
-     "\x00\xED\x91\x87\x00\xED\x91\x88\x00\xED\x91\x89\x00\xED\x91\x8A\x00\xED"
-     "\x91\x8B"
-     "\x00\xED\x91\x8C\x00\xED\x91\x8D\x00\xED\x91\x8E\x00\xED\x91\x8F\x00\xED"
-     "\x91\x90"
-     "\x00\xED\x91\x91\x00\xED\x91\x92\x00\xED\x91\x93\x00\xED\x91\x94\x00\xED"
-     "\x91\x95"
-     "\x00\xED\x91\x96\x00\xED\x91\x97\x00\xED\x91\x98\x00\xED\x91\x99\x00\xED"
-     "\x91\x9A"
-     "\x00\xED\x91\x9B\x00\xED\x91\x9C\x00\xED\x91\x9D\x00\xED\x91\x9E\x00\xED"
-     "\x91\x9F"
-     "\x00\xED\x91\xA0\x00\xED\x91\xA1\x00\xED\x91\xA2\x00\xED\x91\xA3\x00\xED"
-     "\x91\xA4"
-     "\x00\xED\x91\xA5\x00\xED\x91\xA6\x00\xED\x91\xA7\x00\xED\x91\xA8\x00\xED"
-     "\x91\xA9"
-     "\x00\xED\x91\xAA\x00\xED\x91\xAB\x00\xED\x91\xAC\x00\xED\x91\xAD\x00\xED"
-     "\x91\xAE"
-     "\x00\xED\x91\xAF\x00\xED\x91\xB0\x00\xED\x91\xB1\x00\xED\x91\xB2\x00\xED"
-     "\x91\xB3"
-     "\x00\xED\x91\xB4\x00\xED\x91\xB5\x00\xED\x91\xB6\x00\xED\x91\xB7\x00\xED"
-     "\x91\xB8"
-     "\x00\xED\x91\xB9\x00\xED\x91\xBA\x00\xED\x91\xBB\x00\xED\x91\xBC\x00\xED"
-     "\x91\xBD"
-     "\x00\xED\x91\xBE\x00\xED\x91\xBF\x00\xED\x92\x80\x00\xED\x92\x81\x00\xED"
-     "\x92\x82"
-     "\x00\xED\x92\x83\x00\xED\x92\x84\x00\xED\x92\x85\x00\xED\x92\x86\x00\xED"
-     "\x92\x87"
-     "\x00\xED\x92\x88\x00\xED\x92\x89\x00\xED\x92\x8A\x00\xED\x92\x8B\x00\xED"
-     "\x92\x8C"
-     "\x00\xED\x92\x8D\x00\xED\x92\x8E\x00\xED\x92\x8F\x00\xED\x92\x90\x00\xED"
-     "\x92\x91"
-     "\x00\xED\x92\x92\x00\xED\x92\x93\x00\xED\x92\x94\x00\xED\x92\x95\x00\xED"
-     "\x92\x96"
-     "\x00\xED\x92\x97\x00\xED\x92\x98\x00\xED\x92\x99\x00\xED\x92\x9A\x00\xED"
-     "\x92\x9B"
-     "\x00\xED\x92\x9C\x00\xED\x92\x9D\x00\xED\x92\x9E\x00\xED\x92\x9F\x00\xED"
-     "\x92\xA0"
-     "\x00\xED\x92\xA1\x00\xED\x92\xA2\x00\xED\x92\xA3\x00\xED\x92\xA4\x00\xED"
-     "\x92\xA5"
-     "\x00\xED\x92\xA6\x00\xED\x92\xA7\x00\xED\x92\xA8\x00\xED\x92\xA9\x00\xED"
-     "\x92\xAA"
-     "\x00\xED\x92\xAB\x00\xED\x92\xAC\x00\xED\x92\xAD\x00\xED\x92\xAE\x00\xED"
-     "\x92\xAF"
-     "\x00\xED\x92\xB0\x00\xED\x92\xB1\x00\xED\x92\xB2\x00\xED\x92\xB3\x00\xED"
-     "\x92\xB4"
-     "\x00\xED\x92\xB5\x00\xED\x92\xB6\x00\xED\x92\xB7\x00\xED\x92\xB8\x00\xED"
-     "\x92\xB9"
-     "\x00\xED\x92\xBA\x00\xED\x92\xBB\x00\xED\x92\xBC\x00\xED\x92\xBD\x00\xED"
-     "\x92\xBE"
-     "\x00\xED\x92\xBF\x00\xED\x93\x80\x00\xED\x93\x81\x00\xED\x93\x82\x00\xED"
-     "\x93\x83"
-     "\x00\xED\x93\x84\x00\xED\x93\x85\x00\xED\x93\x86\x00\xED\x93\x87\x00\xED"
-     "\x93\x88"
-     "\x00\xED\x93\x89\x00\xED\x93\x8A\x00\xED\x93\x8B\x00\xED\x93\x8C\x00\xED"
-     "\x93\x8D"
-     "\x00\xED\x93\x8E\x00\xED\x93\x8F\x00\xED\x93\x90\x00\xED\x93\x91\x00\xED"
-     "\x93\x92"
-     "\x00\xED\x93\x93\x00\xED\x93\x94\x00\xED\x93\x95\x00\xED\x93\x96\x00\xED"
-     "\x93\x97"
-     "\x00\xED\x93\x98\x00\xED\x93\x99\x00\xED\x93\x9A\x00\xED\x93\x9B\x00\xED"
-     "\x93\x9C"
-     "\x00\xED\x93\x9D\x00\xED\x93\x9E\x00\xED\x93\x9F\x00\xED\x93\xA0\x00\xED"
-     "\x93\xA1"
-     "\x00\xED\x93\xA2\x00\xED\x93\xA3\x00\xED\x93\xA4\x00\xED\x93\xA5\x00\xED"
-     "\x93\xA6"
-     "\x00\xED\x93\xA7\x00\xED\x93\xA8\x00\xED\x93\xA9\x00\xED\x93\xAA\x00\xED"
-     "\x93\xAB"
-     "\x00\xED\x93\xAC\x00\xED\x93\xAD\x00\xED\x93\xAE\x00\xED\x93\xAF\x00\xED"
-     "\x93\xB0"
-     "\x00\xED\x93\xB1\x00\xED\x93\xB2\x00\xED\x93\xB3\x00\xED\x93\xB4\x00\xED"
-     "\x93\xB5"
-     "\x00\xED\x93\xB6\x00\xED\x93\xB7\x00\xED\x93\xB8\x00\xED\x93\xB9\x00\xED"
-     "\x93\xBA"
-     "\x00\xED\x93\xBB\x00\xED\x93\xBC\x00\xED\x93\xBD\x00\xED\x93\xBE\x00\xED"
-     "\x93\xBF"
-     "\x00\xED\x94\x80\x00\xED\x94\x81\x00\xED\x94\x82\x00\xED\x94\x83\x00\xED"
-     "\x94\x84"
-     "\x00\xED\x94\x85\x00\xED\x94\x86\x00\xED\x94\x87\x00\xED\x94\x88\x00\xED"
-     "\x94\x89"
-     "\x00\xED\x94\x8A\x00\xED\x94\x8B\x00\xED\x94\x8C\x00\xED\x94\x8D\x00\xED"
-     "\x94\x8E"
-     "\x00\xED\x94\x8F\x00\xED\x94\x90\x00\xED\x94\x91\x00\xED\x94\x92\x00\xED"
-     "\x94\x93"
-     "\x00\xED\x94\x94\x00\xED\x94\x95\x00\xED\x94\x96\x00\xED\x94\x97\x00\xED"
-     "\x94\x98"
-     "\x00\xED\x94\x99\x00\xED\x94\x9A\x00\xED\x94\x9B\x00\xED\x94\x9C\x00\xED"
-     "\x94\x9D"
-     "\x00\xED\x94\x9E\x00\xED\x94\x9F\x00\xED\x94\xA0\x00\xED\x94\xA1\x00\xED"
-     "\x94\xA2"
-     "\x00\xED\x94\xA3\x00\xED\x94\xA4\x00\xED\x94\xA5\x00\xED\x94\xA6\x00\xED"
-     "\x94\xA7"
-     "\x00\xED\x94\xA8\x00\xED\x94\xA9\x00\xED\x94\xAA\x00\xED\x94\xAB\x00\xED"
-     "\x94\xAC"
-     "\x00\xED\x94\xAD\x00\xED\x94\xAE\x00\xED\x94\xAF\x00\xED\x94\xB0\x00\xED"
-     "\x94\xB1"
-     "\x00\xED\x94\xB2\x00\xED\x94\xB3\x00\xED\x94\xB4\x00\xED\x94\xB5\x00\xED"
-     "\x94\xB6"
-     "\x00\xED\x94\xB7\x00\xED\x94\xB8\x00\xED\x94\xB9\x00\xED\x94\xBA\x00\xED"
-     "\x94\xBB"
-     "\x00\xED\x94\xBC\x00\xED\x94\xBD\x00\xED\x94\xBE\x00\xED\x94\xBF\x00\xED"
-     "\x95\x80"
-     "\x00\xED\x95\x81\x00\xED\x95\x82\x00\xED\x95\x83\x00\xED\x95\x84\x00\xED"
-     "\x95\x85"
-     "\x00\xED\x95\x86\x00\xED\x95\x87\x00\xED\x95\x88\x00\xED\x95\x89\x00\xED"
-     "\x95\x8A"
-     "\x00\xED\x95\x8B\x00\xED\x95\x8C\x00\xED\x95\x8D\x00\xED\x95\x8E\x00\xED"
-     "\x95\x8F"
-     "\x00\xED\x95\x90\x00\xED\x95\x91\x00\xED\x95\x92\x00\xED\x95\x93\x00\xED"
-     "\x95\x94"
-     "\x00\xED\x95\x95\x00\xED\x95\x96\x00\xED\x95\x97\x00\xED\x95\x98\x00\xED"
-     "\x95\x99"
-     "\x00\xED\x95\x9A\x00\xED\x95\x9B\x00\xED\x95\x9C\x00\xED\x95\x9D\x00\xED"
-     "\x95\x9E"
-     "\x00\xED\x95\x9F\x00\xED\x95\xA0\x00\xED\x95\xA1\x00\xED\x95\xA2\x00\xED"
-     "\x95\xA3"
-     "\x00\xED\x95\xA4\x00\xED\x95\xA5\x00\xED\x95\xA6\x00\xED\x95\xA7\x00\xED"
-     "\x95\xA8"
-     "\x00\xED\x95\xA9\x00\xED\x95\xAA\x00\xED\x95\xAB\x00\xED\x95\xAC\x00\xED"
-     "\x95\xAD"
-     "\x00\xED\x95\xAE\x00\xED\x95\xAF\x00\xED\x95\xB0\x00\xED\x95\xB1\x00\xED"
-     "\x95\xB2"
-     "\x00\xED\x95\xB3\x00\xED\x95\xB4\x00\xED\x95\xB5\x00\xED\x95\xB6\x00\xED"
-     "\x95\xB7"
-     "\x00\xED\x95\xB8\x00\xED\x95\xB9\x00\xED\x95\xBA\x00\xED\x95\xBB\x00\xED"
-     "\x95\xBC"
-     "\x00\xED\x95\xBD\x00\xED\x95\xBE\x00\xED\x95\xBF\x00\xED\x96\x80\x00\xED"
-     "\x96\x81"
-     "\x00\xED\x96\x82\x00\xED\x96\x83\x00\xED\x96\x84\x00\xED\x96\x85\x00\xED"
-     "\x96\x86"
-     "\x00\xED\x96\x87\x00\xED\x96\x88\x00\xED\x96\x89\x00\xED\x96\x8A\x00\xED"
-     "\x96\x8B"
-     "\x00\xED\x96\x8C\x00\xED\x96\x8D\x00\xED\x96\x8E\x00\xED\x96\x8F\x00\xED"
-     "\x96\x90"
-     "\x00\xED\x96\x91\x00\xED\x96\x92\x00\xED\x96\x93\x00\xED\x96\x94\x00\xED"
-     "\x96\x95"
-     "\x00\xED\x96\x96\x00\xED\x96\x97\x00\xED\x96\x98\x00\xED\x96\x99\x00\xED"
-     "\x96\x9A"
-     "\x00\xED\x96\x9B\x00\xED\x96\x9C\x00\xED\x96\x9D\x00\xED\x96\x9E\x00\xED"
-     "\x96\x9F"
-     "\x00\xED\x96\xA0\x00\xED\x96\xA1\x00\xED\x96\xA2\x00\xED\x96\xA3\x00\xED"
-     "\x96\xA4"
-     "\x00\xED\x96\xA5\x00\xED\x96\xA6\x00\xED\x96\xA7\x00\xED\x96\xA8\x00\xED"
-     "\x96\xA9"
-     "\x00\xED\x96\xAA\x00\xED\x96\xAB\x00\xED\x96\xAC\x00\xED\x96\xAD\x00\xED"
-     "\x96\xAE"
-     "\x00\xED\x96\xAF\x00\xED\x96\xB0\x00\xED\x96\xB1\x00\xED\x96\xB2\x00\xED"
-     "\x96\xB3"
-     "\x00\xED\x96\xB4\x00\xED\x96\xB5\x00\xED\x96\xB6\x00\xED\x96\xB7\x00\xED"
-     "\x96\xB8"
-     "\x00\xED\x96\xB9\x00\xED\x96\xBA\x00\xED\x96\xBB\x00\xED\x96\xBC\x00\xED"
-     "\x96\xBD"
-     "\x00\xED\x96\xBE\x00\xED\x96\xBF\x00\xED\x97\x80\x00\xED\x97\x81\x00\xED"
-     "\x97\x82"
-     "\x00\xED\x97\x83\x00\xED\x97\x84\x00\xED\x97\x85\x00\xED\x97\x86\x00\xED"
-     "\x97\x87"
-     "\x00\xED\x97\x88\x00\xED\x97\x89\x00\xED\x97\x8A\x00\xED\x97\x8B\x00\xED"
-     "\x97\x8C"
-     "\x00\xED\x97\x8D\x00\xED\x97\x8E\x00\xED\x97\x8F\x00\xED\x97\x90\x00\xED"
-     "\x97\x91"
-     "\x00\xED\x97\x92\x00\xED\x97\x93\x00\xED\x97\x94\x00\xED\x97\x95\x00\xED"
-     "\x97\x96"
-     "\x00\xED\x97\x97\x00\xED\x97\x98\x00\xED\x97\x99\x00\xED\x97\x9A\x00\xED"
-     "\x97\x9B"
-     "\x00\xED\x97\x9C\x00\xED\x97\x9D\x00\xED\x97\x9E\x00\xED\x97\x9F\x00\xED"
-     "\x97\xA0"
-     "\x00\xED\x97\xA1\x00\xED\x97\xA2\x00\xED\x97\xA3\x00\xED\x97\xA4\x00\xED"
-     "\x97\xA5"
-     "\x00\xED\x97\xA6\x00\xED\x97\xA7\x00\xED\x97\xA8\x00\xED\x97\xA9\x00\xED"
-     "\x97\xAA"
-     "\x00\xED\x97\xAB\x00\xED\x97\xAC\x00\xED\x97\xAD\x00\xED\x97\xAE\x00\xED"
-     "\x97\xAF"
-     "\x00\xED\x97\xB0\x00\xED\x97\xB1\x00\xED\x97\xB2\x00\xED\x97\xB3\x00\xED"
-     "\x97\xB4"
-     "\x00\xED\x97\xB5\x00\xED\x97\xB6\x00\xED\x97\xB7\x00\xED\x97\xB8\x00\xED"
-     "\x97\xB9"
-     "\x00\xED\x97\xBA\x00\xED\x97\xBB\x00\xED\x97\xBC\x00\xED\x97\xBD\x00\xED"
-     "\x97\xBE"
-     "\x00\xED\x97\xBF\x00\xED\x98\x80\x00\xED\x98\x81\x00\xED\x98\x82\x00\xED"
-     "\x98\x83"
-     "\x00\xED\x98\x84\x00\xED\x98\x85\x00\xED\x98\x86\x00\xED\x98\x87\x00\xED"
-     "\x98\x88"
-     "\x00\xED\x98\x89\x00\xED\x98\x8A\x00\xED\x98\x8B\x00\xED\x98\x8C\x00\xED"
-     "\x98\x8D"
-     "\x00\xED\x98\x8E\x00\xED\x98\x8F\x00\xED\x98\x90\x00\xED\x98\x91\x00\xED"
-     "\x98\x92"
-     "\x00\xED\x98\x93\x00\xED\x98\x94\x00\xED\x98\x95\x00\xED\x98\x96\x00\xED"
-     "\x98\x97"
-     "\x00\xED\x98\x98\x00\xED\x98\x99\x00\xED\x98\x9A\x00\xED\x98\x9B\x00\xED"
-     "\x98\x9C"
-     "\x00\xED\x98\x9D\x00\xED\x98\x9E\x00\xED\x98\x9F\x00\xED\x98\xA0\x00\xED"
-     "\x98\xA1"
-     "\x00\xED\x98\xA2\x00\xED\x98\xA3\x00\xED\x98\xA4\x00\xED\x98\xA5\x00\xED"
-     "\x98\xA6"
-     "\x00\xED\x98\xA7\x00\xED\x98\xA8\x00\xED\x98\xA9\x00\xED\x98\xAA\x00\xED"
-     "\x98\xAB"
-     "\x00\xED\x98\xAC\x00\xED\x98\xAD\x00\xED\x98\xAE\x00\xED\x98\xAF\x00\xED"
-     "\x98\xB0"
-     "\x00\xED\x98\xB1\x00\xED\x98\xB2\x00\xED\x98\xB3\x00\xED\x98\xB4\x00\xED"
-     "\x98\xB5"
-     "\x00\xED\x98\xB6\x00\xED\x98\xB7\x00\xED\x98\xB8\x00\xED\x98\xB9\x00\xED"
-     "\x98\xBA"
-     "\x00\xED\x98\xBB\x00\xED\x98\xBC\x00\xED\x98\xBD\x00\xED\x98\xBE\x00\xED"
-     "\x98\xBF"
-     "\x00\xED\x99\x80\x00\xED\x99\x81\x00\xED\x99\x82\x00\xED\x99\x83\x00\xED"
-     "\x99\x84"
-     "\x00\xED\x99\x85\x00\xED\x99\x86\x00\xED\x99\x87\x00\xED\x99\x88\x00\xED"
-     "\x99\x89"
-     "\x00\xED\x99\x8A\x00\xED\x99\x8B\x00\xED\x99\x8C\x00\xED\x99\x8D\x00\xED"
-     "\x99\x8E"
-     "\x00\xED\x99\x8F\x00\xED\x99\x90\x00\xED\x99\x91\x00\xED\x99\x92\x00\xED"
-     "\x99\x93"
-     "\x00\xED\x99\x94\x00\xED\x99\x95\x00\xED\x99\x96\x00\xED\x99\x97\x00\xED"
-     "\x99\x98"
-     "\x00\xED\x99\x99\x00\xED\x99\x9A\x00\xED\x99\x9B\x00\xED\x99\x9C\x00\xED"
-     "\x99\x9D"
-     "\x00\xED\x99\x9E\x00\xED\x99\x9F\x00\xED\x99\xA0\x00\xED\x99\xA1\x00\xED"
-     "\x99\xA2"
-     "\x00\xED\x99\xA3\x00\xED\x99\xA4\x00\xED\x99\xA5\x00\xED\x99\xA6\x00\xED"
-     "\x99\xA7"
-     "\x00\xED\x99\xA8\x00\xED\x99\xA9\x00\xED\x99\xAA\x00\xED\x99\xAB\x00\xED"
-     "\x99\xAC"
-     "\x00\xED\x99\xAD\x00\xED\x99\xAE\x00\xED\x99\xAF\x00\xED\x99\xB0\x00\xED"
-     "\x99\xB1"
-     "\x00\xED\x99\xB2\x00\xED\x99\xB3\x00\xED\x99\xB4\x00\xED\x99\xB5\x00\xED"
-     "\x99\xB6"
-     "\x00\xED\x99\xB7\x00\xED\x99\xB8\x00\xED\x99\xB9\x00\xED\x99\xBA\x00\xED"
-     "\x99\xBB"
-     "\x00\xED\x99\xBC\x00\xED\x99\xBD\x00\xED\x99\xBE\x00\xED\x99\xBF\x00\xED"
-     "\x9A\x80"
-     "\x00\xED\x9A\x81\x00\xED\x9A\x82\x00\xED\x9A\x83\x00\xED\x9A\x84\x00\xED"
-     "\x9A\x85"
-     "\x00\xED\x9A\x86\x00\xED\x9A\x87\x00\xED\x9A\x88\x00\xED\x9A\x89\x00\xED"
-     "\x9A\x8A"
-     "\x00\xED\x9A\x8B\x00\xED\x9A\x8C\x00\xED\x9A\x8D\x00\xED\x9A\x8E\x00\xED"
-     "\x9A\x8F"
-     "\x00\xED\x9A\x90\x00\xED\x9A\x91\x00\xED\x9A\x92\x00\xED\x9A\x93\x00\xED"
-     "\x9A\x94"
-     "\x00\xED\x9A\x95\x00\xED\x9A\x96\x00\xED\x9A\x97\x00\xED\x9A\x98\x00\xED"
-     "\x9A\x99"
-     "\x00\xED\x9A\x9A\x00\xED\x9A\x9B\x00\xED\x9A\x9C\x00\xED\x9A\x9D\x00\xED"
-     "\x9A\x9E"
-     "\x00\xED\x9A\x9F\x00\xED\x9A\xA0\x00\xED\x9A\xA1\x00\xED\x9A\xA2\x00\xED"
-     "\x9A\xA3"
-     "\x00\xED\x9A\xA4\x00\xED\x9A\xA5\x00\xED\x9A\xA6\x00\xED\x9A\xA7\x00\xED"
-     "\x9A\xA8"
-     "\x00\xED\x9A\xA9\x00\xED\x9A\xAA\x00\xED\x9A\xAB\x00\xED\x9A\xAC\x00\xED"
-     "\x9A\xAD"
-     "\x00\xED\x9A\xAE\x00\xED\x9A\xAF\x00\xED\x9A\xB0\x00\xED\x9A\xB1\x00\xED"
-     "\x9A\xB2"
-     "\x00\xED\x9A\xB3\x00\xED\x9A\xB4\x00\xED\x9A\xB5\x00\xED\x9A\xB6\x00\xED"
-     "\x9A\xB7"
-     "\x00\xED\x9A\xB8\x00\xED\x9A\xB9\x00\xED\x9A\xBA\x00\xED\x9A\xBB\x00\xED"
-     "\x9A\xBC"
-     "\x00\xED\x9A\xBD\x00\xED\x9A\xBE\x00\xED\x9A\xBF\x00\xED\x9B\x80\x00\xED"
-     "\x9B\x81"
-     "\x00\xED\x9B\x82\x00\xED\x9B\x83\x00\xED\x9B\x84\x00\xED\x9B\x85\x00\xED"
-     "\x9B\x86"
-     "\x00\xED\x9B\x87\x00\xED\x9B\x88\x00\xED\x9B\x89\x00\xED\x9B\x8A\x00\xED"
-     "\x9B\x8B"
-     "\x00\xED\x9B\x8C\x00\xED\x9B\x8D\x00\xED\x9B\x8E\x00\xED\x9B\x8F\x00\xED"
-     "\x9B\x90"
-     "\x00\xED\x9B\x91\x00\xED\x9B\x92\x00\xED\x9B\x93\x00\xED\x9B\x94\x00\xED"
-     "\x9B\x95"
-     "\x00\xED\x9B\x96\x00\xED\x9B\x97\x00\xED\x9B\x98\x00\xED\x9B\x99\x00\xED"
-     "\x9B\x9A"
-     "\x00\xED\x9B\x9B\x00\xED\x9B\x9C\x00\xED\x9B\x9D\x00\xED\x9B\x9E\x00\xED"
-     "\x9B\x9F"
-     "\x00\xED\x9B\xA0\x00\xED\x9B\xA1\x00\xED\x9B\xA2\x00\xED\x9B\xA3\x00\xED"
-     "\x9B\xA4"
-     "\x00\xED\x9B\xA5\x00\xED\x9B\xA6\x00\xED\x9B\xA7\x00\xED\x9B\xA8\x00\xED"
-     "\x9B\xA9"
-     "\x00\xED\x9B\xAA\x00\xED\x9B\xAB\x00\xED\x9B\xAC\x00\xED\x9B\xAD\x00\xED"
-     "\x9B\xAE"
-     "\x00\xED\x9B\xAF\x00\xED\x9B\xB0\x00\xED\x9B\xB1\x00\xED\x9B\xB2\x00\xED"
-     "\x9B\xB3"
-     "\x00\xED\x9B\xB4\x00\xED\x9B\xB5\x00\xED\x9B\xB6\x00\xED\x9B\xB7\x00\xED"
-     "\x9B\xB8"
-     "\x00\xED\x9B\xB9\x00\xED\x9B\xBA\x00\xED\x9B\xBB\x00\xED\x9B\xBC\x00\xED"
-     "\x9B\xBD"
-     "\x00\xED\x9B\xBE\x00\xED\x9B\xBF\x00\xED\x9C\x80\x00\xED\x9C\x81\x00\xED"
-     "\x9C\x82"
-     "\x00\xED\x9C\x83\x00\xED\x9C\x84\x00\xED\x9C\x85\x00\xED\x9C\x86\x00\xED"
-     "\x9C\x87"
-     "\x00\xED\x9C\x88\x00\xED\x9C\x89\x00\xED\x9C\x8A\x00\xED\x9C\x8B\x00\xED"
-     "\x9C\x8C"
-     "\x00\xED\x9C\x8D\x00\xED\x9C\x8E\x00\xED\x9C\x8F\x00\xED\x9C\x90\x00\xED"
-     "\x9C\x91"
-     "\x00\xED\x9C\x92\x00\xED\x9C\x93\x00\xED\x9C\x94\x00\xED\x9C\x95\x00\xED"
-     "\x9C\x96"
-     "\x00\xED\x9C\x97\x00\xED\x9C\x98\x00\xED\x9C\x99\x00\xED\x9C\x9A\x00\xED"
-     "\x9C\x9B"
-     "\x00\xED\x9C\x9C\x00\xED\x9C\x9D\x00\xED\x9C\x9E\x00\xED\x9C\x9F\x00\xED"
-     "\x9C\xA0"
-     "\x00\xED\x9C\xA1\x00\xED\x9C\xA2\x00\xED\x9C\xA3\x00\xED\x9C\xA4\x00\xED"
-     "\x9C\xA5"
-     "\x00\xED\x9C\xA6\x00\xED\x9C\xA7\x00\xED\x9C\xA8\x00\xED\x9C\xA9\x00\xED"
-     "\x9C\xAA"
-     "\x00\xED\x9C\xAB\x00\xED\x9C\xAC\x00\xED\x9C\xAD\x00\xED\x9C\xAE\x00\xED"
-     "\x9C\xAF"
-     "\x00\xED\x9C\xB0\x00\xED\x9C\xB1\x00\xED\x9C\xB2\x00\xED\x9C\xB3\x00\xED"
-     "\x9C\xB4"
-     "\x00\xED\x9C\xB5\x00\xED\x9C\xB6\x00\xED\x9C\xB7\x00\xED\x9C\xB8\x00\xED"
-     "\x9C\xB9"
-     "\x00\xED\x9C\xBA\x00\xED\x9C\xBB\x00\xED\x9C\xBC\x00\xED\x9C\xBD\x00\xED"
-     "\x9C\xBE"
-     "\x00\xED\x9C\xBF\x00\xED\x9D\x80\x00\xED\x9D\x81\x00\xED\x9D\x82\x00\xED"
-     "\x9D\x83"
-     "\x00\xED\x9D\x84\x00\xED\x9D\x85\x00\xED\x9D\x86\x00\xED\x9D\x87\x00\xED"
-     "\x9D\x88"
-     "\x00\xED\x9D\x89\x00\xED\x9D\x8A\x00\xED\x9D\x8B\x00\xED\x9D\x8C\x00\xED"
-     "\x9D\x8D"
-     "\x00\xED\x9D\x8E\x00\xED\x9D\x8F\x00\xED\x9D\x90\x00\xED\x9D\x91\x00\xED"
-     "\x9D\x92"
-     "\x00\xED\x9D\x93\x00\xED\x9D\x94\x00\xED\x9D\x95\x00\xED\x9D\x96\x00\xED"
-     "\x9D\x97"
-     "\x00\xED\x9D\x98\x00\xED\x9D\x99\x00\xED\x9D\x9A\x00\xED\x9D\x9B\x00\xED"
-     "\x9D\x9C"
-     "\x00\xED\x9D\x9D\x00\xED\x9D\x9E\x00\xED\x9D\x9F\x00\xED\x9D\xA0\x00\xED"
-     "\x9D\xA1"
-     "\x00\xED\x9D\xA2\x00\xED\x9D\xA3\x00\xED\x9D\xA4\x00\xED\x9D\xA5\x00\xED"
-     "\x9D\xA6"
-     "\x00\xED\x9D\xA7\x00\xED\x9D\xA8\x00\xED\x9D\xA9\x00\xED\x9D\xAA\x00\xED"
-     "\x9D\xAB"
-     "\x00\xED\x9D\xAC\x00\xED\x9D\xAD\x00\xED\x9D\xAE\x00\xED\x9D\xAF\x00\xED"
-     "\x9D\xB0"
-     "\x00\xED\x9D\xB1\x00\xED\x9D\xB2\x00\xED\x9D\xB3\x00\xED\x9D\xB4\x00\xED"
-     "\x9D\xB5"
-     "\x00\xED\x9D\xB6\x00\xED\x9D\xB7\x00\xED\x9D\xB8\x00\xED\x9D\xB9\x00\xED"
-     "\x9D\xBA"
-     "\x00\xED\x9D\xBB\x00\xED\x9D\xBC\x00\xED\x9D\xBD\x00\xED\x9D\xBE\x00\xED"
-     "\x9D\xBF"
-     "\x00\xED\x9E\x80\x00\xED\x9E\x81\x00\xED\x9E\x82\x00\xED\x9E\x83\x00\xED"
-     "\x9E\x84"
-     "\x00\xED\x9E\x85\x00\xED\x9E\x86\x00\xED\x9E\x87\x00\xED\x9E\x88\x00\xED"
-     "\x9E\x89"
-     "\x00\xED\x9E\x8A\x00\xED\x9E\x8B\x00\xED\x9E\x8C\x00\xED\x9E\x8D\x00\xED"
-     "\x9E\x8E"
-     "\x00\xED\x9E\x8F\x00\xED\x9E\x90\x00\xED\x9E\x91\x00\xED\x9E\x92\x00\xED"
-     "\x9E\x93"
-     "\x00\xED\x9E\x94\x00\xED\x9E\x95\x00\xED\x9E\x96\x00\xED\x9E\x97\x00\xED"
-     "\x9E\x98"
-     "\x00\xED\x9E\x99\x00\xED\x9E\x9A\x00\xED\x9E\x9B\x00\xED\x9E\x9C\x00\xED"
-     "\x9E\x9D"
-     "\x00\xED\x9E\x9E\x00\xED\x9E\x9F\x00\xED\x9E\xA0\x00\xED\x9E\xA1\x00\xED"
-     "\x9E\xA2"
-     "\x00\xED\x9E\xA3\x00\xF0\x91\x82\x9A\x00\xF0\x91\x82\x9C\x00\xF0\x91\x82"
-     "\xAB\x00"
-     "\xF0\x91\x84\xAE\x00\xF0\x91\x84\xAF\x00\xF0\x91\x8D\x8B\x00\xF0\x91\x8D"
-     "\x8C\x00"
-     "\xF0\x91\x92\xBB\x00\xF0\x91\x92\xBC\x00\xF0\x91\x92\xBE\x00\xF0\x91\x96"
-     "\xBA\x00"
-     "\xF0\x91\x96\xBB\x00\xF0\x91\xA4\xB8\x00\xF0\x9D\x85\x97\xF0\x9D\x85\xA5"
-     "\x00\xF0"
-     "\x9D\x85\x98\xF0\x9D\x85\xA5\x00\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\xF0\x9D"
-     "\x85\xAE"
-     "\x00\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\xF0\x9D\x85\xAF\x00\xF0\x9D\x85\x98"
-     "\xF0\x9D"
-     "\x85\xA5\xF0\x9D\x85\xB0\x00\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\xF0\x9D\x85"
-     "\xB1\x00"
-     "\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\xF0\x9D\x85\xB2\x00\xF0\x9D\x86\xB9\xF0"
-     "\x9D\x85"
-     "\xA5\x00\xF0\x9D\x86\xB9\xF0\x9D\x85\xA5\xF0\x9D\x85\xAE\x00\xF0\x9D\x86"
-     "\xB9\xF0"
-     "\x9D\x85\xA5\xF0\x9D\x85\xAF\x00\xF0\x9D\x86\xBA\xF0\x9D\x85\xA5\x00\xF0"
-     "\x9D\x86"
-     "\xBA\xF0\x9D\x85\xA5\xF0\x9D\x85\xAE\x00\xF0\x9D\x86\xBA\xF0\x9D\x85\xA5"
-     "\xF0\x9D"
-     "\x85\xAF\x00\xF0\x9D\xBC\x84\x00\xF0\x9D\xBC\x85\x00\xF0\x9D\xBC\x86\x00"
-     "\xF0\x9D"
-     "\xBC\x88\x00\xF0\x9D\xBC\x8A\x00\xF0\x9D\xBC\x9E\x00\xF0\xA0\x84\xA2\x00"
-     "\xF0\xA0"
-     "\x94\x9C\x00\xF0\xA0\x94\xA5\x00\xF0\xA0\x95\x8B\x00\xF0\xA0\x98\xBA\x00"
-     "\xF0\xA0"
-     "\xA0\x84\x00\xF0\xA0\xA3\x9E\x00\xF0\xA0\xA8\xAC\x00\xF0\xA0\xAD\xA3\x00"
-     "\xF0\xA1"
-     "\x93\xA4\x00\xF0\xA1\x9A\xA8\x00\xF0\xA1\x9B\xAA\x00\xF0\xA1\xA7\x88\x00"
-     "\xF0\xA1"
-     "\xAC\x98\x00\xF0\xA1\xB4\x8B\x00\xF0\xA1\xB7\xA4\x00\xF0\xA1\xB7\xA6\x00"
-     "\xF0\xA2"
-     "\x86\x83\x00\xF0\xA2\x86\x9F\x00\xF0\xA2\x8C\xB1\x00\xF0\xA2\x9B\x94\x00"
-     "\xF0\xA2"
-     "\xA1\x84\x00\xF0\xA2\xA1\x8A\x00\xF0\xA2\xAC\x8C\x00\xF0\xA2\xAF\xB1\x00"
-     "\xF0\xA3"
-     "\x80\x8A\x00\xF0\xA3\x8A\xB8\x00\xF0\xA3\x8D\x9F\x00\xF0\xA3\x8E\x93\x00"
-     "\xF0\xA3"
-     "\x8E\x9C\x00\xF0\xA3\x8F\x83\x00\xF0\xA3\x8F\x95\x00\xF0\xA3\x91\xAD\x00"
-     "\xF0\xA3"
-     "\x9A\xA3\x00\xF0\xA3\xA2\xA7\x00\xF0\xA3\xAA\x8D\x00\xF0\xA3\xAB\xBA\x00"
-     "\xF0\xA3"
-     "\xB2\xBC\x00\xF0\xA3\xB4\x9E\x00\xF0\xA3\xBB\x91\x00\xF0\xA3\xBD\x9E\x00"
-     "\xF0\xA3"
-     "\xBE\x8E\x00\xF0\xA4\x89\xA3\x00\xF0\xA4\x8B\xAE\x00\xF0\xA4\x8E\xAB\x00"
-     "\xF0\xA4"
-     "\x98\x88\x00\xF0\xA4\x9C\xB5\x00\xF0\xA4\xA0\x94\x00\xF0\xA4\xB0\xB6\x00"
-     "\xF0\xA4"
-     "\xB2\x92\x00\xF0\xA4\xBE\xA1\x00\xF0\xA4\xBE\xB8\x00\xF0\xA5\x81\x84\x00"
-     "\xF0\xA5"
-     "\x83\xB2\x00\xF0\xA5\x83\xB3\x00\xF0\xA5\x84\x99\x00\xF0\xA5\x84\xB3\x00"
-     "\xF0\xA5"
-     "\x89\x89\x00\xF0\xA5\x90\x9D\x00\xF0\xA5\x98\xA6\x00\xF0\xA5\x9A\x9A\x00"
-     "\xF0\xA5"
-     "\x9B\x85\x00\xF0\xA5\xA5\xBC\x00\xF0\xA5\xAA\xA7\x00\xF0\xA5\xAE\xAB\x00"
-     "\xF0\xA5"
-     "\xB2\x80\x00\xF0\xA5\xB3\x90\x00\xF0\xA5\xBE\x86\x00\xF0\xA6\x87\x9A\x00"
-     "\xF0\xA6"
-     "\x88\xA8\x00\xF0\xA6\x89\x87\x00\xF0\xA6\x8B\x99\x00\xF0\xA6\x8C\xBE\x00"
-     "\xF0\xA6"
-     "\x93\x9A\x00\xF0\xA6\x94\xA3\x00\xF0\xA6\x96\xA8\x00\xF0\xA6\x9E\xA7\x00"
-     "\xF0\xA6"
-     "\x9E\xB5\x00\xF0\xA6\xAC\xBC\x00\xF0\xA6\xB0\xB6\x00\xF0\xA6\xB3\x95\x00"
-     "\xF0\xA6"
-     "\xB5\xAB\x00\xF0\xA6\xBC\xAC\x00\xF0\xA6\xBE\xB1\x00\xF0\xA7\x83\x92\x00"
-     "\xF0\xA7"
-     "\x8F\x8A\x00\xF0\xA7\x99\xA7\x00\xF0\xA7\xA2\xAE\x00\xF0\xA7\xA5\xA6\x00"
-     "\xF0\xA7"
-     "\xB2\xA8\x00\xF0\xA7\xBB\x93\x00\xF0\xA7\xBC\xAF\x00\xF0\xA8\x97\x92\x00"
-     "\xF0\xA8"
-     "\x97\xAD\x00\xF0\xA8\x9C\xAE\x00\xF0\xA8\xAF\xBA\x00\xF0\xA8\xB5\xB7\x00"
-     "\xF0\xA9"
-     "\x85\x85\x00\xF0\xA9\x87\x9F\x00\xF0\xA9\x88\x9A\x00\xF0\xA9\x90\x8A\x00"
-     "\xF0\xA9"
-     "\x92\x96\x00\xF0\xA9\x96\xB6\x00\xF0\xA9\xAC\xB0\x00\xF0\xAA\x83\x8E\x00"
-     "\xF0\xAA"
-     "\x84\x85\x00\xF0\xAA\x88\x8E\x00\xF0\xAA\x8A\x91\x00\xF0\xAA\x8E\x92\x00"
-     "\xF0\xAA"
-     "\x98\x80\x00"},
-    {"nfkc_cf", 245779,
-     "\x00\xCC\x02\x00\x00\xFC\x00\x00\x3E\x1C\x03\x00\x3D\x24\x03\x00\x3C\x2C"
-     "\x03\x00"
-     "\xCC\xE4\x02\x00\xB8\x0D\x00\x00\x5D\x21\x00\x80\xCC\xDC\x02\x00\xB8\x05"
-     "\x00\x00"
-     "\x51\x21\x00\x80\xCC\xE4\x02\x00\xB8\x1D\x00\x00\x61\x21\x00\x80\x62\x06"
-     "\x00\x80"
-     "\xB3\x08\x00\x80\x91\x05\x01\x00\xB6\x08\x00\x80\x8F\x55\x01\x00\x8C\x45"
-     "\x01\x00"
-     "\x0B\x1E\x00\x80\x8A\x59\x01\x00\xB9\x08\x00\x80\x88\x59\x01\x00\x89\xAD"
-     "\x00\x00"
-     "\x86\xD1\x00\x00\x87\xD9\x00\x00\x84\xC9\x00\x00\x07\x1E\x00\x80\x82\x25"
-     "\x00\x00"
-     "\x83\xE1\x00\x00\x80\x41\x00\x00\x81\x3D\x00\x00\x80\xCD\x00\x00\x81\xE9"
-     "\x00\x00"
-     "\x13\x1E\x00\x80\x83\x05\x00\x00\x0F\x1E\x00\x80\xBC\x08\x00\x80\x80\xD5"
-     "\x00\x00"
-     "\x81\xF1\x00\x00\x10\x09\x00\x80\x89\x35\x00\x00\x1F\x1E\x00\x80\x1B\x1E"
-     "\x00\x80"
-     "\x13\x09\x00\x80\x27\x1E\x00\x80\x23\x1E\x00\x80\x90\x0A\x00\x80\x81\x6D"
-     "\x00\x00"
-     "\x80\x6D\x00\x00\x83\x7D\x00\x00\x84\x11\x00\x00\x80\x79\x00\x00\x81\x79"
-     "\x00\x00"
-     "\xA8\x25\x01\x00\x30\x0A\x00\x80\x89\x59\x00\x00\x2D\x0A\x00\x80\x84\x0D"
-     "\x00\x00"
-     "\xA5\x61\x01\x00\x03\x1E\x00\x80\xA3\x89\x01\x00\x81\xFD\x01\x00\xCC\xE8"
-     "\x00\x00"
-     "\xBF\x08\x00\x80\x51\x0A\x00\x80\xC2\x08\x00\x80\x12\x0A\x00\x80\x5A\x0A"
-     "\x00\x80"
-     "\x7A\x2C\x0F\x00\x79\xF0\x0E\x00\x78\xA8\x09\x00\x77\xA8\x0E\x00\x76\x78"
-     "\x09\x00"
-     "\x75\xB0\x0E\x00\x74\x54\x09\x00\x73\xA0\x0E\x00\x72\x74\x0E\x00\x5D\x0A"
-     "\x00\x80"
-     "\x70\x2C\x0E\x00\x6F\x8C\x09\x00\x6E\x98\x09\x00\x6D\xCC\x09\x00\x6C\xE4"
-     "\x09\x00"
-     "\x6B\x00\x09\x00\x6A\xAC\x08\x00\x69\x34\x09\x00\x68\xD8\x08\x00\x67\x10"
-     "\x09\x00"
-     "\x66\xF4\x08\x00\x65\xB0\x08\x00\x64\xD0\x08\x00\x63\xD8\x08\x00\x62\xC0"
-     "\x08\x00"
-     "\x61\x30\x08\x00\xFF\x1D\x00\x80\x82\x05\x00\x00\x17\x1E\x00\x80\x2B\x1E"
-     "\x00\x80"
-     "\xC3\x1C\x00\x80\x86\x19\x00\x00\x5A\x01\x0C\x00\x59\x55\x08\x00\x58\x2D"
-     "\x08\x00"
-     "\x57\x2D\x08\x00\x56\xC1\x08\x00\x55\x3D\x09\x00\x54\xF5\x09\x00\x53\xF1"
-     "\x09\x00"
-     "\x52\x85\x09\x00\x51\x1D\x06\x00\x50\x95\x09\x00\x4F\x39\x04\x00\x4E\x55"
-     "\x04\x00"
-     "\x4D\xF5\x04\x00\x4C\x21\x05\x00\x4B\x15\x05\x00\x4A\x61\x05\x00\x49\x8D"
-     "\x05\x00"
-     "\x48\xE5\x05\x00\x47\xD1\x05\x00\x46\xC9\x05\x00\x45\xF9\x05\x00\x44\xE5"
-     "\x05\x00"
-     "\x43\x01\x04\x00\x42\x39\x02\x00\x41\xCD\x01\x00\x16\x09\x00\x80\x81\x41"
-     "\x02\x00"
-     "\x80\x3D\x02\x00\x87\x05\x00\x00\xC7\x1C\x00\x80\xCB\x1C\x00\x80\xCF\x1C"
-     "\x00\x80"
-     "\x8C\x59\x00\x00\x19\x09\x00\x80\x82\x05\x00\x00\x1C\x09\x00\x80\x1F\x09"
-     "\x00\x80"
-     "\x81\x31\x00\x00\xCC\xC8\x00\x00\x87\x1D\x00\x00\xCC\xC4\x00\x00\xCD\xC0"
-     "\x00\x00"
-     "\x22\x09\x00\x80\xD3\x1C\x00\x80\x8C\x15\x00\x00\xCC\xC4\x01\x00\xD7\x1C"
-     "\x00\x80"
-     "\x81\x11\x00\x00\x81\x45\x00\x00\x25\x09\x00\x80\xDB\x1C\x00\x80\x87\x35"
-     "\x00\x00"
-     "\xE3\x1C\x00\x80\xE7\x1C\x00\x80\xDF\x1C\x00\x80\xCB\x08\x00\x80\x82\x89"
-     "\x00\x00"
-     "\xCE\x08\x00\x80\x80\x09\x00\x00\x81\x15\x00\x00\x3F\x1E\x00\x80\x83\x19"
-     "\x00\x00"
-     "\x3B\x1E\x00\x80\x47\x1E\x00\x80\xA3\x89\x00\x00\x43\x1E\x00\x80\x37\x1E"
-     "\x00\x80"
-     "\x89\x39\x00\x00\xEB\x1C\x00\x80\xEF\x1C\x00\x80\x2E\x09\x00\x80\xA7\xB5"
-     "\x06\x00"
-     "\x31\x09\x00\x80\xB1\xCD\x00\x00\x80\x49\x00\x00\x81\x55\x00\x00\xD4\x08"
-     "\x00\x80"
-     "\xAD\xA1\x00\x00\xB1\xC5\x00\x00\x33\x1E\x00\x80\x80\x71\x00\x00\x81\x71"
-     "\x00\x00"
-     "\x37\x09\x00\x80\xA7\x8D\x00\x00\x60\x0A\x00\x80\x63\x0A\x00\x80\x4B\x1E"
-     "\x00\x80"
-     "\xA3\x95\x00\x00\x86\xF1\x01\x00\x81\xD5\x01\x00\xCD\x04\x03\x00\xCC\x7C"
-     "\x01\x00"
-     "\xFB\x1C\x00\x80\x34\x09\x00\x80\xF3\x1C\x00\x80\xF7\x1C\x00\x80\x80\x6D"
-     "\x01\x00"
-     "\x81\x61\x01\x00\x87\x05\x00\x00\xFF\x1C\x00\x80\x4B\x0A\x00\x80\x81\x59"
-     "\x07\x00"
-     "\x3A\x09\x00\x80\x03\x1D\x00\x80\x3D\x09\x00\x80\xF0\x5C\x95\x02\xEF\x50"
-     "\x71\x02"
-     "\x40\x09\x00\x80\x39\x0A\x00\x80\x43\x09\x00\x80\x46\x09\x00\x80\xEA\x30"
-     "\x6A\x02"
-     "\x07\x1D\x00\x80\x0F\x1D\x00\x80\xCD\xF8\x01\x00\xCC\xE4\x01\x00\x87\x0A"
-     "\x00\x80"
-     "\x0B\x1D\x00\x80\xE3\xB0\x4F\x02\xE2\x00\x3C\x02\xE1\x78\x29\x00\xE0\xEC"
-     "\x25\x00"
-     "\xCC\x44\x01\x00\xCD\x54\x01\x00\x82\x49\x07\x00\xCC\x50\x01\x00\xDB\xB4"
-     "\x22\x00"
-     "\x13\x1D\x00\x80\xD9\x00\x24\x00\xD8\x58\x23\x00\x17\x1D\x00\x80\xD6\xA8"
-     "\x0B\x00"
-     "\xD5\x28\x22\x00\xD4\xA0\x23\x00\xD3\xB0\x20\x00\xD2\xB8\x21\x00\xD1\xF4"
-     "\x1E\x00"
-     "\xD0\x40\x1D\x00\xCF\x00\x1C\x00\xCE\xD4\x16\x00\xCD\xC8\x17\x00\x7E\x06"
-     "\x00\x80"
-     "\xCB\xE0\x15\x00\xCA\x98\x15\x00\xC9\xDC\x15\x00\xC8\x00\x14\x00\xC7\xE0"
-     "\x12\x00"
-     "\xC6\xE0\x13\x00\xC5\xE8\x11\x00\xC4\x00\x10\x00\xC3\xF0\x0E\x00\xC2\x74"
-     "\x0F\x00"
-     "\xD7\x08\x00\x80\xDA\x08\x00\x80\xC8\x08\x00\x80\x91\xE9\x06\x00\xB1\x06"
-     "\x00\x80"
-     "\x2F\x1E\x00\x80\xD1\x06\x00\x80\xDD\x08\x00\x80\xD1\x08\x00\x80\xA7\x51"
-     "\x07\x00"
-     "\x2B\x09\x00\x80\x4C\x09\x00\x80\x93\x0A\x00\x80\xDA\x06\x00\x80\xF4\x06"
-     "\x00\x80"
-     "\x4F\x09\x00\x80\x08\x07\x00\x80\x52\x09\x00\x80\x80\x35\x06\x00\x81\x39"
-     "\x06\x00"
-     "\x82\x51\x00\x00\x83\xED\x06\x00\x84\x71\x00\x00\x1D\x07\x00\x80\x86\xE9"
-     "\x06\x00"
-     "\x87\xE5\x06\x00\x88\xA9\x06\x00\x89\xB1\x06\x00\x58\x09\x00\x80\xE0\x08"
-     "\x00\x80"
-     "\x8C\x91\x06\x00\x1F\x1D\x00\x80\x4F\x1E\x00\x80\x8F\x95\x06\x00\xB0\x95"
-     "\x07\x00"
-     "\x81\xAD\x07\x00\x82\xB9\x07\x00\x15\x0A\x00\x80\x84\xA5\x07\x00\x66\x0A"
-     "\x00\x80"
-     "\x86\xA1\x07\x00\x87\xD9\x07\x00\x69\x0A\x00\x80\x53\x1E\x00\x80\x55\x09"
-     "\x00\x80"
-     "\x1B\x1D\x00\x80\x8C\xF9\x07\x00\x35\x07\x00\x80\x80\x41\x07\x00\x81\x41"
-     "\x07\x00"
-     "\x5E\x09\x00\x80\x37\x07\x00\x80\x23\x1D\x00\x80\xA3\xC1\x00\x00\x3C\x0A"
-     "\x00\x80"
-     "\x27\x1D\x00\x80\x61\x09\x00\x80\xA7\xF5\x00\x00\xA8\xED\x07\x00\x2B\x1D"
-     "\x00\x80"
-     "\x81\x21\x00\x00\x6D\x07\x00\x80\x64\x09\x00\x80\xAD\xE5\x07\x00\x6A\x09"
-     "\x00\x80"
-     "\x2F\x1D\x00\x80\x82\x51\x06\x00\x91\xA5\x01\x00\x84\x05\x00\x00\x33\x1D"
-     "\x00\x80"
-     "\x81\xD9\x01\x00\x87\x4D\x06\x00\x8C\xE1\x01\x00\x67\x09\x00\x80\x3B\x1D"
-     "\x00\x80"
-     "\x8C\xF5\x01\x00\x88\x75\x06\x00\x81\xDD\x01\x00\x37\x1D\x00\x80\x81\xC5"
-     "\x01\x00"
-     "\x8C\x51\x06\x00\x87\x07\x00\x80\x80\xB9\x06\x00\x81\xB9\x06\x00\x82\x5D"
-     "\x01\x00"
-     "\x83\x69\x01\x00\x84\x65\x01\x00\x3F\x1D\x00\x80\x86\x65\x01\x00\x87\x35"
-     "\x01\x00"
-     "\x88\x0D\x01\x00\x89\x1D\x01\x00\x43\x1D\x00\x80\x81\x39\x00\x00\x8C\xFD"
-     "\x01\x00"
-     "\x87\x1D\x00\x00\x47\x1D\x00\x80\x8F\xE9\x01\x00\xB0\x2D\x01\x00\xA3\xE9"
-     "\x06\x00"
-     "\xA7\x51\x01\x00\x81\xD9\x00\x00\xD9\x07\x00\x80\xA7\x01\x06\x00\xA3\x4D"
-     "\x01\x00"
-     "\xAD\xBD\x00\x00\x4E\x0A\x00\x80\xA3\x59\x01\x00\x70\x09\x00\x80\xE6\x08"
-     "\x00\x80"
-     "\xAE\x11\x06\x00\xA7\xA9\x00\x00\x4B\x1D\x00\x80\x76\x09\x00\x80\x4F\x1D"
-     "\x00\x80"
-     "\x73\x09\x00\x80\x57\x1D\x00\x80\xA3\x69\x01\x00\xB1\x35\x01\x00\x53\x1D"
-     "\x00\x80"
-     "\x80\x79\x00\x00\x81\x75\x00\x00\xA8\x49\x01\x00\xA3\x9D\x00\x00\xE9\x08"
-     "\x00\x80"
-     "\xB1\xDD\x00\x00\x87\x06\x00\x80\xEC\x08\x00\x80\xF3\x07\x00\x80\x63\x1E"
-     "\x00\x80"
-     "\x87\xB9\x03\x00\x5F\x1E\x00\x80\xCD\x80\x01\x00\xCC\x64\x01\x00\x83\xBD"
-     "\x03\x00"
-     "\x91\xE1\x00\x00\x81\xB1\x03\x00\x80\xBD\x03\x00\x6B\x1E\x00\x80\x67\x1E"
-     "\x00\x80"
-     "\x5B\x1D\x00\x80\x8C\x91\x03\x00\x99\x0A\x00\x80\x5F\x1D\x00\x80\x63\x1D"
-     "\x00\x80"
-     "\x9B\x25\x0E\x00\x84\x49\x0E\x00\x81\x6D\x00\x00\x86\x21\x00\x00\x87\x5D"
-     "\x0E\x00"
-     "\x80\xB9\x03\x00\x81\xA1\x03\x00\x82\x01\x0C\x00\x83\x5D\x0E\x00\x8C\x81"
-     "\x00\x00"
-     "\x67\x1D\x00\x80\x7F\x09\x00\x80\x8F\x85\x00\x00\x88\x65\x0E\x00\x89\xA5"
-     "\x00\x00"
-     "\x9F\x0A\x00\x80\x8B\xA9\x00\x00\xA7\x45\x03\x00\x83\xA5\x00\x00\x80\x75"
-     "\x03\x00"
-     "\x81\x89\x00\x00\xA3\x51\x03\x00\x84\xED\x00\x00\x80\x65\x03\x00\x81\x99"
-     "\x00\x00"
-     "\xCC\xBC\x02\x00\xCD\x7C\x01\x00\xAD\x61\x03\x00\x89\x89\x00\x00\x80\x89"
-     "\x00\x00"
-     "\x81\xD1\x00\x00\x81\x21\x01\x00\x96\x0A\x00\x80\xCC\xE8\x06\x00\xCD\x98"
-     "\x06\x00"
-     "\xCC\x98\x02\x00\xCD\x58\x03\x00\x5B\x1E\x00\x80\x82\x09\x00\x80\xB1\x0D"
-     "\x03\x00"
-     "\xA3\xFD\x0E\x00\x18\x0A\x00\x80\x84\x59\x00\x00\x6C\x0A\x00\x80\x81\x45"
-     "\x01\x00"
-     "\xA8\xD5\x0E\x00\x6F\x0A\x00\x80\x77\x1E\x00\x80\x73\x1E\x00\x80\xCC\x80"
-     "\x01\x00"
-     "\xCD\xFC\x07\x00\xCC\x00\x00\x00\x7F\x1E\x00\x80\xCC\xC8\x07\x00\xCD\xCC"
-     "\x07\x00"
-     "\xCC\xF8\x07\x00\xCC\x20\x06\x00\xCC\x60\x01\x00\xCD\x3C\x06\x00\xCC\xA8"
-     "\x07\x00"
-     "\xCD\xEC\x07\x00\xCC\x60\x07\x00\xCD\x64\x07\x00\xCC\x40\x07\x00\xCC\x4C"
-     "\x07\x00"
-     "\x84\x71\x01\x00\x83\x49\x00\x00\x80\xB1\x01\x00\x81\xB1\x01\x00\x7B\x1E"
-     "\x00\x80"
-     "\x81\x7D\x01\x00\xCD\x70\x06\x00\xCC\x74\x06\x00\xCD\xD8\x01\x00\xCC\x24"
-     "\x00\x00"
-     "\x82\xE9\x00\x00\x89\x3D\x00\x00\x88\x45\x01\x00\x83\x1E\x00\x80\x81\x85"
-     "\x01\x00"
-     "\x80\x85\x01\x00\x6F\x1E\x00\x80\xCC\x04\x06\x00\x84\x05\x00\x00\x42\x0A"
-     "\x00\x80"
-     "\x81\x65\x02\x00\x80\x7D\x02\x00\x6B\x1D\x00\x80\x6F\x1D\x00\x80\x81\x39"
-     "\x00\x00"
-     "\x23\x08\x00\x80\x88\x09\x00\x80\x73\x1D\x00\x80\x8E\x09\x00\x80\x72\x0A"
-     "\x00\x80"
-     "\x87\x25\x00\x00\x75\x0A\x00\x80\x84\x09\x00\x00\xA3\xB1\x00\x00\x7B\x1D"
-     "\x00\x80"
-     "\xCC\xD4\x00\x00\x8B\x09\x00\x80\x7F\x1D\x00\x80\xCD\xE8\x00\x00\xCC\xF4"
-     "\x00\x00"
-     "\x8B\x1D\x00\x80\x94\x09\x00\x80\x83\x1D\x00\x80\x8F\x1D\x00\x80\xCD\x1C"
-     "\x01\x00"
-     "\xCC\xE0\x00\x00\x87\x19\x00\x00\x87\x11\x00\x00\xF2\x08\x00\x80\x91\xB9"
-     "\x0F\x00"
-     "\x7C\x09\x00\x80\x87\x11\x00\x00\x9C\x0A\x00\x80\xF5\x08\x00\x80\xE5\x09"
-     "\x00\x80"
-     "\x93\x1D\x00\x80\x57\x1E\x00\x80\x3F\x0A\x00\x80\xF8\x07\x00\x80\x7E\x0A"
-     "\x00\x80"
-     "\x25\x08\x00\x80\x97\x09\x00\x80\x77\x1D\x00\x80\x91\x09\x00\x80\x3F\x08"
-     "\x00\x80"
-     "\x81\xED\x0F\x00\x9A\x09\x00\x80\x87\x1D\x00\x80\x81\x6D\x00\x00\x8C\x1D"
-     "\x00\x00"
-     "\x53\x08\x00\x80\x87\xF1\x0F\x00\x81\x5D\x00\x00\x97\x1D\x00\x80\xA0\x09"
-     "\x00\x80"
-     "\x82\x89\x0F\x00\x8C\xC1\x0F\x00\x9B\x1D\x00\x80\x87\x91\x0F\x00\x8F\xC9"
-     "\x0F\x00"
-     "\x81\x0A\x00\x80\xB1\x51\x0F\x00\x9D\x09\x00\x80\xA3\x1D\x00\x80\x9F\x1D"
-     "\x00\x80"
-     "\x60\x08\x00\x80\xFB\x08\x00\x80\xFE\x08\x00\x80\x01\x09\x00\x80\xA6\x09"
-     "\x00\x80"
-     "\x81\x05\x00\x00\xB3\x1D\x00\x80\xA9\x09\x00\x80\x88\x0D\x00\x00\xB7\x1D"
-     "\x00\x80"
-     "\xAC\x09\x00\x80\x04\x09\x00\x80\x84\x25\x00\x00\x27\x0A\x00\x80\xA3\xF5"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x80\x1D\x00\x00\x21\x0A\x00\x80\xA7\x0D\x0F\x00\x1E\x0A"
-     "\x00\x80"
-     "\x8C\x09\x00\x00\xA3\xA5\x00\x00\x24\x0A\x00\x80\x80\x39\x00\x00\x81\x2D"
-     "\x00\x00"
-     "\xA7\xCD\x00\x00\xA6\xD1\x00\x00\x87\x65\x01\x00\x8B\x1E\x00\x80\xAF\x09"
-     "\x00\x80"
-     "\xB2\x09\x00\x80\x1B\x0A\x00\x80\x78\x0A\x00\x80\x7B\x0A\x00\x80\x89\x2D"
-     "\x00\x00"
-     "\x93\x1E\x00\x80\x8F\x1E\x00\x80\x9B\x1E\x00\x80\x8C\x45\x01\x00\x97\x1E"
-     "\x00\x80"
-     "\x83\x1D\x00\x00\x80\x19\x00\x00\x81\x19\x00\x00\xF7\x09\x00\x80\x9F\x1E"
-     "\x00\x80"
-     "\x80\x69\x00\x00\x81\x69\x00\x00\x87\x1E\x00\x80\xA7\x1D\x00\x80\xB5\x09"
-     "\x00\x80"
-     "\xAF\x1D\x00\x80\xAB\x1D\x00\x80\x62\x08\x00\x80\xCC\xE0\x0D\x00\xCD\xF4"
-     "\x0D\x00"
-     "\x80\xE9\x01\x00\x81\xE9\x01\x00\x83\x05\x00\x00\xBB\x1D\x00\x80\xA7\x09"
-     "\x01\x00"
-     "\xA6\x05\x01\x00\xBF\x1D\x00\x80\x76\x08\x00\x80\xA3\xE5\x01\x00\xC3\x1D"
-     "\x00\x80"
-     "\xC7\x1D\x00\x80\xB8\x09\x00\x80\xCF\x1D\x00\x80\xCB\x1D\x00\x80\xAD\x25"
-     "\x01\x00"
-     "\xD3\x1D\x00\x80\x81\x08\x00\x80\xA3\xF1\x00\x00\x80\x2D\x00\x00\x81\x25"
-     "\x00\x00"
-     "\x87\x05\x00\x00\xD7\x1D\x00\x80\xDB\x1D\x00\x80\x8A\x08\x00\x80\xA3\x1E"
-     "\x00\x80"
-     "\x07\x09\x00\x80\xB1\x49\x01\x00\xBB\x09\x00\x80\xAF\x1E\x00\x80\xA2\x0A"
-     "\x00\x80"
-     "\xDF\x1D\x00\x80\x0D\x09\x00\x80\xAB\x1E\x00\x80\xA7\x1E\x00\x80\xA3\x71"
-     "\x00\x00"
-     "\x88\x35\x00\x00\x81\x9D\x02\x00\x80\x9D\x02\x00\x83\xAD\x02\x00\x82\xAD"
-     "\x02\x00"
-     "\x87\x41\x28\x00\x84\xA5\x02\x00\xBE\x09\x00\x80\x86\xA1\x02\x00\x89\x25"
-     "\x03\x00"
-     "\x88\xE5\x02\x00\x8B\x25\x03\x00\x8A\x25\x03\x00\xE3\x1D\x00\x80\x8C\x25"
-     "\x03\x00"
-     "\x8F\x2D\x03\x00\xC1\x09\x00\x80\x91\x59\x03\x00\xC4\x09\x00\x80\xE7\x1D"
-     "\x00\x80"
-     "\xEB\x1D\x00\x80\x81\x49\x00\x00\xCC\x20\x01\x00\xEF\x08\x00\x80\xA3\xF1"
-     "\x03\x00"
-     "\x80\xB1\x03\x00\x81\xB1\x03\x00\x9B\x29\x03\x00\xEF\x1D\x00\x80\xCC\x48"
-     "\x01\x00"
-     "\xCD\x3C\x01\x00\xCC\x00\x08\x00\xCD\x78\x0A\x00\xCC\x5C\x01\x00\xCC\x90"
-     "\x0A\x00"
-     "\xA3\xD9\x03\x00\xCC\xF4\x0A\x00\x48\x0A\x00\x80\xA4\xC1\x03\x00\xCC\xB0"
-     "\x0A\x00"
-     "\xCD\xC0\x0A\x00\xCC\x9C\x0B\x00\xA8\xFD\x03\x00\xF3\x1D\x00\x80\x94\x6D"
-     "\x19\x00"
-     "\xAD\xED\x03\x00\xCC\xB8\x04\x00\xCD\x00\x01\x00\xCC\x6C\x02\x00\xF7\x1D"
-     "\x00\x80"
-     "\xB0\xA5\x03\x00\xCC\x1C\x02\x00\x87\x6D\x03\x00\x80\x45\x03\x00\x81\x4D"
-     "\x03\x00"
-     "\x82\x45\x03\x00\x82\x1D\x0E\x00\xFB\x1D\x00\x80\x8C\x75\x00\x00\xCC\xA8"
-     "\x0C\x00"
-     "\xCD\xF4\x0B\x00\x88\x55\x03\x00\x00\x00\x00\x80\xCD\x98\x02\x00\xCC\x3C"
-     "\x03\x00"
-     "\xCC\x40\x01\x00\xCD\xEC\x0F\x00\xCC\xD8\x0E\x00\xCC\xE8\x0E\x00\xCC\x70"
-     "\x0E\x00"
-     "\xCC\x8C\x0E\x00\xCC\xBC\x0E\x00\xCD\xA0\x0F\x00\xCC\xA4\x0F\x00\xCC\xD0"
-     "\x0F\x00"
-     "\xCC\x60\x0E\x00\xCD\x50\x0E\x00\xCC\x60\x0E\x00\xCD\x74\x0F\x00\xCC\x88"
-     "\x0F\x00"
-     "\xCD\x8C\x0F\x00\xCC\xC0\x0D\x00\xCD\xBC\x0B\x00\xCC\x54\x04\x00\xA3\xB9"
-     "\x02\x00"
-     "\x1A\x00\x00\x80\x0A\x00\x00\x80\xCD\xE0\x0D\x00\xCC\x04\x05\x00\xCC\x74"
-     "\x0C\x00"
-     "\xCD\x88\x0D\x00\xCC\x74\x00\x00\xCC\x18\x00\x00\xCC\x14\x01\x00\xCD\x20"
-     "\x01\x00"
-     "\xCC\x98\x0E\x00\xCC\x6C\x01\x00\xCC\x14\x01\x00\xCC\x58\x00\x00\x82\x04"
-     "\x00\x80"
-     "\x21\x05\x00\x80\xCD\x3C\x01\x00\xCC\x38\x01\x00\xCC\xB0\x0C\x00\xCD\xAC"
-     "\x0C\x00"
-     "\xCD\xD0\x00\x00\xCC\xA0\x03\x00\x02\x00\x00\x80\x13\x0C\x00\x80\xCD\x38"
-     "\x0C\x00"
-     "\xCC\xDC\x0C\x00\x80\xD1\x10\x00\x64\x00\x00\x80\x80\x11\x02\x00\x81\x11"
-     "\x02\x00"
-     "\x82\x15\x02\x00\x83\x2D\x02\x00\x84\x35\x02\x00\x5F\x03\x00\x80\x4F\x04"
-     "\x00\x80"
-     "\x87\x35\x02\x00\x88\x0D\x02\x00\x89\x15\x02\x00\xCD\x30\x0C\x00\xCC\xF0"
-     "\x0F\x00"
-     "\xCC\x24\x01\x00\xCD\x04\x01\x00\xCD\x6C\x0E\x00\xCC\xAC\x0F\x00\x81\x19"
-     "\x06\x00"
-     "\x82\x55\x0F\x00\x43\x04\x00\x80\x82\x3D\x06\x00\x87\x49\x0F\x00\x53\x05"
-     "\x00\x80"
-     "\x87\x25\x06\x00\x91\xBD\x08\x00\xCC\x84\x09\x00\xCD\x98\x08\x00\xB9\x08"
-     "\x00\x80"
-     "\x88\x71\x0F\x00\xBF\x08\x00\x80\x8C\x71\x06\x00\xC2\x08\x00\x80\x8C\x55"
-     "\x0F\x00"
-     "\x86\x7D\x08\x00\x54\x0A\x00\x80\x84\x7D\x08\x00\x33\x0A\x00\x80\x82\x45"
-     "\x08\x00"
-     "\x83\x71\x08\x00\x80\xA1\x0F\x00\x81\xA1\x0F\x00\xA3\xCD\x05\x00\x8F\xF1"
-     "\x08\x00"
-     "\x8C\xE5\x08\x00\x81\x29\x00\x00\xCC\x30\x06\x00\xCD\x9C\x08\x00\x88\x15"
-     "\x08\x00"
-     "\x89\x05\x08\x00\xA3\xED\x0F\x00\xC5\x08\x00\x80\xA3\xC1\x06\x00\xC8\x08"
-     "\x00\x80"
-     "\xA7\x05\x0F\x00\xD1\x08\x00\x80\xB0\x35\x08\x00\xE0\x08\x00\x80\xCC\xC4"
-     "\x06\x00"
-     "\xCD\xF0\x06\x00\xCC\xAC\x04\x00\xCD\xDC\x07\x00\xE3\x08\x00\x80\xAE\x15"
-     "\x0F\x00"
-     "\xCC\x54\x02\x00\xCD\x04\x07\x00\xB1\x8D\x06\x00\x8A\x05\x06\x00\xB1\xA5"
-     "\x06\x00"
-     "\x88\x69\x05\x00\xEF\x08\x00\x80\xA3\x71\x08\x00\xCC\x9C\x03\x00\xCD\xBC"
-     "\x04\x00"
-     "\x81\xA5\x00\x00\x82\x79\x05\x00\x81\x71\x05\x00\x80\x79\x05\x00\x87\x51"
-     "\x05\x00"
-     "\x84\xB9\x00\x00\xA8\x51\x08\x00\xF2\x08\x00\x80\x87\x65\x05\x00\xF5\x08"
-     "\x00\x80"
-     "\xCC\x0C\x03\x00\xCD\x58\x05\x00\x81\x05\x00\x00\x57\x0A\x00\x80\x04\x09"
-     "\x00\x80"
-     "\x0A\x09\x00\x80\xF8\x08\x00\x80\x13\x09\x00\x80\x28\x09\x00\x80\x8C\x45"
-     "\x05\x00"
-     "\xCC\xD4\x03\x00\xCD\x5C\x03\x00\x2B\x09\x00\x80\x88\x95\x07\x00\x81\x95"
-     "\x04\x00"
-     "\x80\x95\x04\x00\x83\xA9\x04\x00\x82\x91\x04\x00\x49\x09\x00\x80\x84\xB1"
-     "\x04\x00"
-     "\x87\xB1\x04\x00\x26\x07\x00\x80\x89\x91\x04\x00\x88\x89\x04\x00\x83\x07"
-     "\x00\x80"
-     "\x8A\x8D\x07\x00\xCC\x64\x02\x00\xCD\x60\x02\x00\x6D\x09\x00\x80\xB6\x0B"
-     "\x00\x80"
-     "\xA7\x09\x05\x00\xA6\x05\x05\x00\x79\x09\x00\x80\x7C\x09\x00\x80\xA3\xE5"
-     "\x05\x00"
-     "\x8C\x08\x00\x80\x85\x09\x00\x80\x91\x09\x00\x80\x9A\x09\x00\x80\xA3\x09"
-     "\x00\x80"
-     "\xAD\x25\x05\x00\xA6\x09\x00\x80\xA9\x09\x00\x80\xCC\x0A\x00\x80\xCA\x09"
-     "\x00\x80"
-     "\xCD\x09\x00\x80\xCF\x0A\x00\x80\xD0\x09\x00\x80\xA3\x3D\x04\x00\xD5\x0A"
-     "\x00\x80"
-     "\xD8\x0A\x00\x80\xD3\x09\x00\x80\xB1\x49\x05\x00\x2A\x0A\x00\x80\xDE\x0A"
-     "\x00\x80"
-     "\xE1\x0A\x00\x80\xD6\x09\x00\x80\xED\x0A\x00\x80\xF6\x0A\x00\x80\x08\x0B"
-     "\x00\x80"
-     "\x05\x0B\x00\x80\xDC\x09\x00\x80\xA0\xF5\x04\x00\x8C\x91\x00\x00\x1A\x0B"
-     "\x00\x80"
-     "\xCC\x38\x02\x00\xCC\x04\x01\x00\xCD\x00\x01\x00\x23\x0B\x00\x80\x2C\x0B"
-     "\x00\x80"
-     "\xA8\x71\x05\x00\xE5\x09\x00\x80\xAA\x1D\x0E\x00\xE8\x09\x00\x80\xEB\x09"
-     "\x00\x80"
-     "\x47\x0B\x00\x80\x85\x95\x1E\x00\xAF\x69\x05\x00\x85\xF5\x1E\x00\xEE\x09"
-     "\x00\x80"
-     "\xB2\xC1\x05\x00\xB3\xC1\x05\x00\xB4\xF9\x05\x00\xB5\xF9\x05\x00\x88\xD5"
-     "\x15\x00"
-     "\x50\x0B\x00\x80\xB8\xDD\x05\x00\xB9\xB1\x05\x00\xBA\x91\x0B\x00\xF4\x09"
-     "\x00\x80"
-     "\xBC\xA9\x05\x00\xBD\x7D\x02\x00\xBE\x6D\x02\x00\x59\x0B\x00\x80\xF7\x09"
-     "\x00\x80"
-     "\x5F\x0B\x00\x80\x62\x0B\x00\x80\xFA\x09\x00\x80\xFD\x09\x00\x80\x45\x0A"
-     "\x00\x80"
-     "\x00\x0A\x00\x80\x03\x0A\x00\x80\xCD\xF4\x06\x00\xCC\xA0\x00\x00\xCD\x06"
-     "\x00\x80"
-     "\x6F\x07\x00\x80\xE1\x07\x00\x80\x36\x0A\x00\x80\x3F\x0A\x00\x80\xCA\x06"
-     "\x00\x80"
-     "\xD9\x09\x00\x80\x06\x0A\x00\x80\x84\x0A\x00\x80\xE2\x09\x00\x80\x8A\x0A"
-     "\x00\x80"
-     "\x8D\x0A\x00\x80\x90\x0A\x00\x80\x93\x0A\x00\x80\x9C\x0A\x00\x80\x85\x59"
-     "\x1E\x00"
-     "\xA3\x22\x00\x80\x85\x21\x1E\x00\xA8\x0A\x00\x80\xDF\x09\x00\x80\xA7\x22"
-     "\x00\x80"
-     "\xB1\x0A\x00\x80\x80\x39\x0F\x00\x81\x45\x0F\x00\x82\x21\x03\x00\x83\x99"
-     "\x0F\x00"
-     "\x84\x21\x03\x00\x85\x2D\x03\x00\x86\x9D\x03\x00\x87\x91\x03\x00\x88\x55"
-     "\x0D\x00"
-     "\x89\x59\x0D\x00\x8A\xBD\x03\x00\x8B\xE5\x0D\x00\x8C\xC9\x0C\x00\x8D\xC9"
-     "\x0C\x00"
-     "\x8E\x2D\x0B\x00\x8F\xA1\x03\x00\x90\xF1\x03\x00\x91\xE9\x0A\x00\x92\xA1"
-     "\x0A\x00"
-     "\x93\xB9\x0A\x00\x94\x81\x03\x00\x95\xA9\x03\x00\x96\x5D\x02\x00\xC7\x09"
-     "\x00\x80"
-     "\x98\x41\x02\x00\x99\xFD\x07\x00\x9A\xF5\x07\x00\x9B\xCD\x07\x00\x9C\x69"
-     "\x02\x00"
-     "\x9D\xA1\x06\x00\x9E\x65\x02\x00\x5C\x0B\x00\x80\x65\x0B\x00\x80\xB4\x0A"
-     "\x00\x80"
-     "\x88\x79\x14\x00\xB7\x0A\x00\x80\x85\x45\x1F\x00\xBA\x0A\x00\x80\xA6\x98"
-     "\x01\x00"
-     "\xBD\x0A\x00\x80\xC0\x0A\x00\x80\xFF\x0A\x00\x80\x38\x0B\x00\x80\x3E\x0B"
-     "\x00\x80"
-     "\xCD\xD8\x01\x00\xCC\x50\x01\x00\x4A\x0B\x00\x80\x12\x00\x00\x80\x16\x00"
-     "\x00\x80"
-     "\x30\x00\x00\x80\x68\x00\x00\x80\x06\x00\x00\x80\x34\x00\x00\x80\x7A\x0B"
-     "\x00\x80"
-     "\xC0\x0B\x00\x80\xC3\x0B\x00\x80\xB8\xA0\x01\x00\xCB\x0B\x00\x80\x88\x25"
-     "\x12\x00"
-     "\xC6\x0B\x00\x80\x0A\x0C\x00\x80\xCE\x0B\x00\x80\xD1\x0B\x00\x80\xB3\x0B"
-     "\x00\x80"
-     "\x80\x69\x1D\x00\x81\x7D\x1D\x00\x77\x0B\x00\x80\x83\xA5\x15\x00\xD4\x0B"
-     "\x00\x80"
-     "\x85\xD1\x15\x00\x74\x00\x00\x80\x87\xA1\x15\x00\x3D\x06\x00\x80\x89\x59"
-     "\x1D\x00"
-     "\x8A\xA1\x1D\x00\x90\x0C\x00\x80\x8C\xBD\x1D\x00\x08\x07\x00\x80\x8E\x51"
-     "\x16\x00"
-     "\x35\x07\x00\x80\x90\x95\x16\x00\x76\x08\x00\x80\x92\xD1\x1D\x00\x24\x00"
-     "\x00\x80"
-     "\x94\xF9\x17\x00\xE3\x0B\x00\x80\x96\x89\x17\x00\xAD\x08\x00\x80\x98\xD1"
-     "\x17\x00"
-     "\xE6\x0B\x00\x80\x9A\xF5\x1D\x00\xE9\x0B\x00\x80\x9C\x95\x17\x00\xEC\x0B"
-     "\x00\x80"
-     "\x9E\x4D\x18\x00\x60\x0C\x00\x80\xA0\x21\x1D\x00\x63\x0C\x00\x80\xA2\x01"
-     "\x18\x00"
-     "\x66\x0C\x00\x80\xA4\xF9\x18\x00\x8F\x1F\x00\x80\xA6\x3D\x1D\x00\x0F\x20"
-     "\x00\x80"
-     "\xA8\x0D\x1D\x00\x0B\x20\x00\x80\xAA\x19\x1D\x00\xB3\x1F\x00\x80\xAC\x0D"
-     "\x18\x00"
-     "\xBB\x1F\x00\x80\xAE\xB1\x19\x00\xC3\x1F\x00\x80\xB0\xCD\x19\x00\xAB\x1F"
-     "\x00\x80"
-     "\xB2\x91\x19\x00\xB7\x1F\x00\x80\xB4\x4D\x19\x00\xBF\x1F\x00\x80\xB6\x05"
-     "\x19\x00"
-     "\x85\x2D\x00\x00\xB8\x0D\x19\x00\xB9\xFD\x1A\x00\x85\x49\x01\x00\xBB\xD1"
-     "\x1A\x00"
-     "\xC7\x1F\x00\x80\xBD\xB1\x1A\x00\xAF\x1F\x00\x80\xBF\xBD\x18\x00\xBB\x1D"
-     "\x14\x00"
-     "\x8A\x08\x00\x80\xB9\xF9\x15\x00\x17\x20\x00\x80\xBF\x79\x1C\x00\xF5\x0B"
-     "\x00\x80"
-     "\xBD\xE1\x15\x00\xF8\x0B\x00\x80\xB3\x7D\x1C\x00\xB2\x79\x1C\x00\x80\x19"
-     "\x00\x00"
-     "\x81\x65\x00\x00\xC7\x1E\x00\x80\xB6\xED\x15\x00\xFB\x0B\x00\x80\xB4\xFD"
-     "\x15\x00"
-     "\x93\x1F\x00\x80\xAA\x71\x15\x00\xCF\x1E\x00\x80\xA8\x69\x15\x00\xE7\x1E"
-     "\x00\x80"
-     "\xAE\xFD\x15\x00\xF2\x0B\x00\x80\xAC\x61\x15\x00\xEF\x1E\x00\x80\xA2\x29"
-     "\x12\x00"
-     "\xEB\x1E\x00\x80\xA0\x29\x12\x00\xF3\x1E\x00\x80\xA6\x25\x1C\x00\xD7\x1E"
-     "\x00\x80"
-     "\xA4\x2D\x12\x00\x01\x0C\x00\x80\x9A\x65\x13\x00\x97\x1F\x00\x80\x98\x81"
-     "\x12\x00"
-     "\xCB\x1E\x00\x80\x9E\xAD\x12\x00\xBF\x1E\x00\x80\x9C\xAD\x12\x00\xD3\x1F"
-     "\x00\x80"
-     "\x92\x9D\x1C\x00\x85\x19\x00\x00\x90\x85\x1C\x00\xDB\x1F\x00\x80\x96\x0D"
-     "\x13\x00"
-     "\x85\x79\x00\x00\x94\x0D\x13\x00\xE3\x1F\x00\x80\x8A\xED\x13\x00\xCB\x1F"
-     "\x00\x80"
-     "\x88\xE9\x13\x00\x85\x09\x00\x00\x8E\x89\x13\x00\xD7\x1F\x00\x80\x8C\x9D"
-     "\x13\x00"
-     "\xD3\x1E\x00\x80\x82\xC1\x1C\x00\xDB\x1E\x00\x80\x80\x4D\x11\x00\xFF\x1E"
-     "\x00\x80"
-     "\x86\xE9\x13\x00\xDF\x1F\x00\x80\x84\x01\x10\x00\x81\xB5\x1F\x00\xCD\x28"
-     "\x1E\x00"
-     "\x85\xF1\x03\x00\x82\xB5\x1F\x00\xE7\x1F\x00\x80\x84\xA9\x1F\x00\x87\xDD"
-     "\x1F\x00"
-     "\x86\xDD\x1F\x00\x89\xED\x1F\x00\x85\x35\x00\x00\x8B\xFD\x1F\x00\x8A\xFD"
-     "\x1F\x00"
-     "\xCD\x50\x1A\x00\xCC\x70\x1D\x00\x8F\xD9\x1F\x00\x8E\xE1\x1F\x00\x91\xA9"
-     "\x1F\x00"
-     "\x90\xA1\x1F\x00\x93\xA5\x1F\x00\xCD\xBC\x1E\x00\xCF\x1F\x00\x80\x94\xA5"
-     "\x1F\x00"
-     "\x97\xA1\x1F\x00\x96\xA9\x1F\x00\x2B\x20\x00\x80\x98\x99\x1F\x00\xCC\x48"
-     "\x06\x00"
-     "\xCD\xAC\x04\x00\x9D\x69\x1C\x00\x9C\x7D\x1C\x00\x9F\x65\x1C\x00\xCD\x28"
-     "\x1E\x00"
-     "\x07\x0C\x00\x80\xA0\xA1\x1C\x00\x9B\x1F\x00\x80\xA2\xA1\x1C\x00\x3F\x20"
-     "\x00\x80"
-     "\xA4\xA5\x1C\x00\xA7\xDD\x1C\x00\xA6\xA9\x1C\x00\xA9\xFD\x1C\x00\xCD\x5C"
-     "\x1E\x00"
-     "\xCC\x80\x01\x00\xCD\x44\x02\x00\x3B\x20\x00\x80\xAC\xD9\x1C\x00\xAF\x39"
-     "\x1C\x00"
-     "\xAE\xC1\x1C\x00\xB1\x45\x1C\x00\xCD\x38\x00\x00\xB3\x45\x1C\x00\xB2\x45"
-     "\x1C\x00"
-     "\xB5\x41\x1C\x00\xCD\x50\x1F\x00\xB7\xD1\x02\x00\xCD\x14\x00\x00\x82\x99"
-     "\x03\x00"
-     "\xB8\x7D\x1C\x00\x80\xD9\x03\x00\x81\xA5\x03\x00\x5A\x0C\x00\x80\xBC\x69"
-     "\x1C\x00"
-     "\x1F\x1F\x00\x80\x85\x39\x02\x00\x27\x1F\x00\x80\xCC\x08\x1D\x00\x2F\x1F"
-     "\x00\x80"
-     "\xCC\xD8\x1B\x00\x84\xB9\x1D\x00\x85\xBD\x1D\x00\x86\xB1\x1D\x00\x87\xB1"
-     "\x1D\x00"
-     "\x88\x8D\x1D\x00\x89\x89\x1D\x00\x8A\x99\x1D\x00\x8B\x9D\x1D\x00\x8C\x81"
-     "\x1D\x00"
-     "\x8D\x39\x12\x00\xCC\x0C\x1F\x00\x8F\xB1\x17\x00\x23\x1F\x00\x80\x91\xA5"
-     "\x15\x00"
-     "\x2B\x1F\x00\x80\x93\x5D\x1A\x00\x33\x1F\x00\x80\x95\xB5\x1B\x00\x0D\x0C"
-     "\x00\x80"
-     "\x97\x85\x1B\x00\x10\x0C\x00\x80\x99\x89\x1B\x00\x36\x0C\x00\x80\x9B\xA5"
-     "\x1B\x00"
-     "\xCD\xC8\x18\x00\xCC\x98\x1D\x00\x9E\x9D\x13\x00\x39\x0C\x00\x80\xA0\x5D"
-     "\x13\x00"
-     "\x9F\x1F\x00\x80\xA2\xC5\x1F\x00\x3F\x1F\x00\x80\xA4\x25\x1D\x00\x47\x1F"
-     "\x00\x80"
-     "\xA6\xD1\x10\x00\x43\x1F\x00\x80\xA8\x71\x17\x00\x4B\x1F\x00\x80\xAA\x11"
-     "\x1D\x00"
-     "\x85\xF5\x06\x00\xAC\x3D\x14\x00\x3F\x0C\x00\x80\xAE\x2D\x1D\x00\x85\xDD"
-     "\x06\x00"
-     "\x5F\x20\x00\x80\xB1\x79\x1D\x00\xB2\x75\x1D\x00\xB3\x71\x1D\x00\xB4\xF9"
-     "\x10\x00"
-     "\x45\x0C\x00\x80\xB6\x19\x1D\x00\xB7\x19\x1D\x00\xB8\x41\x16\x00\x48\x0C"
-     "\x00\x80"
-     "\xBA\xED\x12\x00\x80\x51\x08\x00\xBC\xB5\x1F\x00\xA3\x1F\x00\x80\xBE\xED"
-     "\x1E\x00"
-     "\x53\x20\x00\x80\x81\xFD\x1A\x00\x92\x38\x19\x00\x83\x55\x1B\x00\x42\x0C"
-     "\x00\x80"
-     "\x85\x91\x1B\x00\x84\x69\x1B\x00\x85\x69\x01\x00\x86\x99\x1B\x00\x85\xB1"
-     "\x01\x00"
-     "\x88\xA9\x1B\x00\xC3\x1E\x00\x80\x8A\xB9\x1B\x00\x4F\x20\x00\x80\x8C\xA9"
-     "\x1B\x00"
-     "\x5D\x0C\x00\x80\x8E\x9D\x1B\x00\x9C\x8D\x1B\x00\x9D\x95\x1B\x00\x80\xD9"
-     "\x05\x00"
-     "\x81\xAD\x05\x00\x98\xED\x1B\x00\x99\x95\x1B\x00\x9A\x9D\x1B\x00\x9B\x95"
-     "\x1B\x00"
-     "\x80\xF1\x05\x00\x81\x85\x05\x00\x82\x81\x05\x00\x5B\x1F\x00\x80\x63\x1F"
-     "\x00\x80"
-     "\x85\x8D\x04\x00\x80\x29\x05\x00\x81\x1D\x05\x00\xB3\xB1\x1C\x00\xB2\xB9"
-     "\x04\x00"
-     "\xB1\x2D\x1B\x00\xB0\xB9\x04\x00\xB7\xD5\x04\x00\xB6\x2D\x1B\x00\xB5\x35"
-     "\x1B\x00"
-     "\xB4\x35\x1B\x00\xA4\x75\x1B\x00\x6B\x1F\x00\x80\x4E\x0C\x00\x80\xB8\xA9"
-     "\x05\x00"
-     "\xA0\x41\x19\x00\xA1\x59\x10\x00\xA2\xF9\x1C\x00\xA3\x0D\x1D\x00\x80\x55"
-     "\x04\x00"
-     "\x81\x91\x04\x00\xFE\x0B\x00\x80\x83\x55\x05\x00\xDF\x1E\x00\x80\x85\xD9"
-     "\x05\x00"
-     "\x80\xB1\x05\x00\x81\xAD\x05\x00\x80\xA1\x05\x00\x81\x81\x04\x00\xE3\x1E"
-     "\x00\x80"
-     "\x83\x3D\x00\x00\x80\x01\x05\x00\x81\x09\x00\x00\x51\x0C\x00\x80\x07\x1F"
-     "\x00\x80"
-     "\x81\xFD\x01\x00\x80\xF5\x04\x00\x0F\x1F\x00\x80\x82\x05\x00\x00\x85\xD9"
-     "\x04\x00"
-     "\xB2\xF1\x1C\x00\x04\x0C\x00\x80\xB0\x65\x13\x00\x54\x0C\x00\x80\xBE\xDD"
-     "\x1B\x00"
-     "\xBD\xDD\x1B\x00\xA7\x1F\x00\x80\xBB\xC1\x1B\x00\xBA\xDD\x1B\x00\x80\x79"
-     "\x00\x00"
-     "\x81\x55\x00\x00\x03\x1F\x00\x80\xA6\x9D\x1B\x00\xF7\x1E\x00\x80\xA4\x99"
-     "\x1B\x00"
-     "\xFB\x1E\x00\x80\xA2\x85\x1B\x00\xF3\x1F\x00\x80\xA0\x91\x1B\x00\xFB\x1F"
-     "\x00\x80"
-     "\xAE\x85\x1B\x00\x0B\x1F\x00\x80\xAC\x5D\x13\x00\x03\x20\x00\x80\xAA\xC9"
-     "\x13\x00"
-     "\xEB\x1F\x00\x80\xA8\xA1\x1B\x00\x13\x1F\x00\x80\x96\x9D\x1C\x00\x0A\x0C"
-     "\x00\x80"
-     "\x94\x99\x1C\x00\xF7\x1F\x00\x80\x92\x29\x12\x00\x17\x1F\x00\x80\x90\x29"
-     "\x12\x00"
-     "\x85\x09\x00\x00\x9E\xCD\x16\x00\xFF\x1F\x00\x80\x9C\x65\x1B\x00\x85\x09"
-     "\x00\x00"
-     "\x9A\x35\x1D\x00\x07\x20\x00\x80\x98\x91\x1D\x00\x77\x1F\x00\x80\x86\x29"
-     "\x17\x00"
-     "\xEF\x1F\x00\x80\x84\x25\x17\x00\x1B\x1F\x00\x80\x82\xED\x14\x00\x6F\x20"
-     "\x00\x80"
-     "\x80\xCD\x14\x00\x7F\x1F\x00\x80\x8E\x11\x13\x00\x87\x1F\x00\x80\x8C\x05"
-     "\x13\x00"
-     "\x7B\x1F\x00\x80\x8A\x55\x11\x00\x83\x1F\x00\x80\x88\x69\x11\x00\x80\xD9"
-     "\x19\x00"
-     "\x81\xD9\x19\x00\x85\x51\x00\x00\x83\xE9\x19\x00\x84\xFD\x19\x00\x85\xE5"
-     "\x19\x00"
-     "\x80\x59\x07\x00\x81\x79\x06\x00\x80\xB5\x06\x00\x81\x71\x06\x00\xCC\x88"
-     "\x02\x00"
-     "\xCD\xAC\x02\x00\x84\xAD\x06\x00\x88\x3D\x08\x00\x86\x9D\x06\x00\xCD\x90"
-     "\x05\x00"
-     "\x81\x29\x03\x00\x80\x05\x03\x00\x94\xA1\x02\x00\x82\xCD\x03\x00\x85\x01"
-     "\x04\x00"
-     "\x93\x85\x02\x00\x13\x20\x00\x80\xCD\x40\x06\x00\xCC\x58\x01\x00\xCD\x20"
-     "\x02\x00"
-     "\x85\xE1\x00\x00\x93\xF5\x07\x00\x94\x59\x02\x00\xCD\xB0\x05\x00\xCC\x88"
-     "\x00\x00"
-     "\xCD\x78\x01\x00\x80\x09\x07\x00\x81\xE9\x06\x00\x1B\x20\x00\x80\x83\xC5"
-     "\x03\x00"
-     "\x85\x21\x00\x00\x85\xF5\x04\x00\x80\x61\x04\x00\x81\x9D\x05\x00\x80\x29"
-     "\x07\x00"
-     "\x81\xC9\x06\x00\xCC\x30\x02\x00\xCD\x0C\x03\x00\x23\x20\x00\x80\xCD\x84"
-     "\x06\x00"
-     "\xCC\x00\x0C\x00\xCD\xAC\x0C\x00\xB0\x35\x19\x00\xCD\xF0\x00\x00\xB2\x31"
-     "\x19\x00"
-     "\xCD\x78\x01\x00\xB4\x2D\x19\x00\xCD\x7C\x07\x00\xB6\xC9\x06\x00\xCD\xD0"
-     "\x00\x00"
-     "\xCC\xF4\x03\x00\xCD\xE4\x0C\x00\xBA\xF1\x06\x00\x93\xA5\x03\x00\x94\xA1"
-     "\x03\x00"
-     "\xCD\x4C\x07\x00\xBE\xD9\x06\x00\xBF\xD1\x06\x00\x84\xA9\x1C\x00\x85\x49"
-     "\x07\x00"
-     "\x86\x5D\x07\x00\x87\x51\x07\x00\x81\x59\x04\x00\x80\x55\x04\x00\xCC\x50"
-     "\x01\x00"
-     "\xCD\x7C\x01\x00\x8C\x5D\x07\x00\x85\x75\x00\x00\x8E\xAD\x07\x00\x8F\xA1"
-     "\x07\x00"
-     "\x88\x55\x07\x00\x89\x59\x07\x00\x8A\x4D\x07\x00\xCC\xC0\x1F\x00\x94\x79"
-     "\x06\x00"
-     "\x95\x8D\x03\x00\x96\xC9\x06\x00\x97\x55\x02\x00\x27\x20\x00\x80\x91\x0D"
-     "\x06\x00"
-     "\x92\x4D\x06\x00\x93\x41\x06\x00\x9C\xCD\x1C\x00\x9D\x0D\x04\x00\x9E\x15"
-     "\x04\x00"
-     "\x9F\x61\x0F\x00\x98\xF1\x05\x00\x99\xFD\x02\x00\x9A\x21\x04\x00\x9B\x1D"
-     "\x04\x00"
-     "\xA4\x65\x04\x00\xA5\xCD\x0F\x00\xA6\x21\x03\x00\xA7\x75\x03\x00\xA0\x25"
-     "\x04\x00"
-     "\xA1\x99\x03\x00\x85\x35\x00\x00\xA3\x49\x04\x00\x80\x59\x05\x00\x81\xA5"
-     "\x04\x00"
-     "\x82\xA1\x04\x00\x2F\x20\x00\x80\xA8\x91\x02\x00\xA9\xC1\x0F\x00\xAA\x49"
-     "\x05\x00"
-     "\xAB\x85\x03\x00\x85\x11\x00\x00\xB5\xAC\x0D\x00\xCD\x74\x04\x00\xB7\x9C"
-     "\x0D\x00"
-     "\x37\x20\x00\x80\xB1\xF8\x0D\x00\xCC\x00\x01\x00\xCD\x04\x01\x00\xCC\x98"
-     "\x01\x00"
-     "\xCD\x9C\x01\x00\x43\x20\x00\x80\xBF\x04\x0D\x00\xE0\x0B\x00\x80\xB9\xBC"
-     "\x0D\x00"
-     "\xCC\xE0\x03\x00\xCD\x1C\x01\x00\x80\x41\x0B\x00\x81\x4D\x0B\x00\x82\x59"
-     "\x0B\x00"
-     "\x3C\x0C\x00\x80\x80\x99\x0A\x00\x81\x61\x08\x00\x37\x1F\x00\x80\x83\x85"
-     "\x0D\x00"
-     "\x80\xA5\x0B\x00\x81\x59\x08\x00\x80\xA5\x0B\x00\x81\xB9\x0B\x00\x3B\x1F"
-     "\x00\x80"
-     "\x4B\x20\x00\x80\x80\xA5\x0B\x00\x81\x99\x0B\x00\x80\xC5\x0B\x00\x81\x39"
-     "\x08\x00"
-     "\x4B\x0C\x00\x80\x83\x55\x00\x00\x53\x1F\x00\x80\x47\x20\x00\x80\x80\x35"
-     "\x0C\x00"
-     "\x81\x2D\x0C\x00\x80\x0D\x0C\x00\x81\x15\x0C\x00\x82\xCD\x0C\x00\x93\x75"
-     "\x00\x00"
-     "\x94\x41\x00\x00\x57\x0C\x00\x80\x80\x8D\x0B\x00\x81\xF9\x08\x00\x81\x05"
-     "\x08\x00"
-     "\x80\x71\x0B\x00\x80\xA5\x0D\x00\x81\x01\x08\x00\x6F\x1F\x00\x80\x84\x69"
-     "\x0B\x00"
-     "\x88\x69\x0A\x00\x86\xAD\x0C\x00\x86\x11\x0A\x00\x88\x9D\x0C\x00\x84\x39"
-     "\x0A\x00"
-     "\x73\x1F\x00\x80\x80\x71\x0D\x00\x81\x55\x0D\x00\x80\x31\x0A\x00\x81\xC9"
-     "\x08\x00"
-     "\x82\x29\x0D\x00\x93\x55\x00\x00\x80\x09\x0D\x00\x81\x2D\x0D\x00\x8B\x1F"
-     "\x00\x80"
-     "\x94\x85\x00\x00\x94\x75\x00\x00\x85\xA5\x0D\x00\x80\x11\x0D\x00\x81\x1D"
-     "\x0D\x00"
-     "\x94\x39\x0D\x00\x85\x25\x0D\x00\x82\x21\x00\x00\x93\x6D\x0D\x00\x80\x09"
-     "\x0D\x00"
-     "\x81\x05\x0D\x00\x80\x2D\x0C\x00\x81\x89\x09\x00\x8F\x1F\x00\x80\x83\x9D"
-     "\x01\x00"
-     "\xE3\x0B\x00\x80\x85\xCD\x0C\x00\xCD\x58\x02\x00\xCC\x3C\x02\x00\x80\x29"
-     "\x00\x00"
-     "\x81\x35\x00\x00\x1F\x20\x00\x80\x97\x1F\x00\x80\x84\xAD\x09\x00\xE9\x0B"
-     "\x00\x80"
-     "\x86\x9D\x09\x00\x33\x20\x00\x80\x80\x49\x00\x00\x81\x55\x00\x00\x82\x61"
-     "\x00\x00"
-     "\x83\xD5\x08\x00\x5A\x0C\x00\x80\x85\x59\x08\x00\xCC\x40\x0D\x00\xCD\xBC"
-     "\x0F\x00"
-     "\xCC\xE0\x02\x00\xCD\xFC\x02\x00\x93\xDD\x00\x00\x93\xF5\x08\x00\x94\x59"
-     "\x0D\x00"
-     "\x94\xB5\x0A\x00\xCC\xD0\x02\x00\xCD\x3C\x01\x00\x80\xAD\x00\x00\x81\xB1"
-     "\x00\x00"
-     "\x88\xD9\x00\x00\x85\xED\x0B\x00\x82\xAD\x00\x00\x83\xDD\x0C\x00\x80\xB5"
-     "\x00\x00"
-     "\x81\xA9\x00\x00\x80\x49\x0E\x00\x81\x55\x0E\x00\x80\x21\x0B\x00\x81\xD9"
-     "\x09\x00"
-     "\x86\x01\x0B\x00\x5B\x20\x00\x80\x84\x29\x0B\x00\x83\x09\x00\x00\x80\x29"
-     "\x0E\x00"
-     "\x81\x35\x0E\x00\x82\xFD\x01\x00\x93\x85\x0C\x00\x94\x81\x0C\x00\x5D\x0C"
-     "\x00\x80"
-     "\xCC\xF8\x03\x00\xCD\x9C\x03\x00\x57\x20\x00\x80\x93\x7D\x0C\x00\xCC\xCC"
-     "\x02\x00"
-     "\xCD\x5C\x02\x00\xCC\x70\x01\x00\xCD\x44\x01\x00\x94\x29\x0C\x00\xEF\x0B"
-     "\x00\x80"
-     "\xCD\x8C\x0F\x00\x67\x20\x00\x80\xCD\xDC\x0F\x00\x57\x1F\x00\x80\x80\xF1"
-     "\x03\x00"
-     "\x81\xE9\x03\x00\xCD\x28\x0F\x00\xCD\x98\x0E\x00\x82\x25\x00\x00\xCD\x58"
-     "\x0E\x00"
-     "\x80\xC9\x03\x00\x81\xD1\x03\x00\x5F\x1F\x00\x80\x67\x1F\x00\x80\x85\x45"
-     "\x00\x00"
-     "\xCD\x20\x0E\x00\x80\xC9\x08\x00\x81\x31\x0A\x00\x82\x1D\x02\x00\x83\xD5"
-     "\x0F\x00"
-     "\x85\xC5\x00\x00\x63\x20\x00\x80\x80\x55\x00\x00\x81\x6D\x00\x00\xCC\xC0"
-     "\x01\x00"
-     "\xCD\xA4\x01\x00\xCC\xE0\x02\x00\xCD\x94\x02\x00\x80\x7D\x00\x00\x81\x45"
-     "\x00\x00"
-     "\x82\x4D\x00\x00\x6B\x20\x00\x80\x80\x75\x09\x00\x81\x01\x0A\x00\x80\x81"
-     "\x06\x00"
-     "\x81\x81\x06\x00\x84\x6D\x09\x00\x73\x20\x00\x80\x86\xA9\x0E\x00\x7B\x20"
-     "\x00\x80"
-     "\x88\x75\x03\x00\x69\x0C\x00\x80\x85\x35\x00\x00\x6C\x0C\x00\x80\x80\x45"
-     "\x09\x00"
-     "\x81\x31\x0A\x00\x82\xED\x00\x00\x83\xBD\x06\x00\x81\xB5\x01\x00\x93\x55"
-     "\x02\x00"
-     "\x81\x71\x01\x00\x93\xCD\x06\x00\x94\x81\x02\x00\x6F\x0C\x00\x80\x94\x75"
-     "\x02\x00"
-     "\x72\x0C\x00\x80\xCD\xC8\x02\x00\xCC\xF4\x02\x00\x88\xCD\x04\x00\x85\x35"
-     "\x0F\x00"
-     "\x82\xE1\x06\x00\x83\x65\x02\x00\x80\xF1\x06\x00\x81\xF1\x06\x00\xCC\x38"
-     "\x01\x00"
-     "\xCD\x3C\x01\x00\x75\x0C\x00\x80\x78\x0C\x00\x80\x7B\x0C\x00\x80\x7E\x0C"
-     "\x00\x80"
-     "\x81\x0C\x00\x80\x84\x0C\x00\x80\x88\x11\x05\x00\xCC\xFC\x02\x00\xCC\x00"
-     "\x01\x00"
-     "\xCD\x04\x01\x00\x87\x0C\x00\x80\x8A\x0C\x00\x80\xCD\x78\x01\x00\xCC\x1C"
-     "\x01\x00"
-     "\xCD\x40\x00\x00\xCC\xCC\x01\x00\x8D\x0C\x00\x80\x93\x0C\x00\x80\x96\x0C"
-     "\x00\x80"
-     "\xD7\x0B\x00\x80\xDA\x0B\x00\x80\xDD\x0B\x00\x80\xCC\x38\x01\x00\xCD\x04"
-     "\x01\x00"
-     "\xF9\x0C\x00\x80\xFC\x0C\x00\x80\xFF\x0C\x00\x80\x02\x0D\x00\x80\xCC\xFC"
-     "\x01\x00"
-     "\xCD\x04\x01\x00\x8F\x41\x05\x00\x05\x0D\x00\x80\x08\x0D\x00\x80\x0E\x0D"
-     "\x00\x80"
-     "\x11\x0D\x00\x80\x14\x0D\x00\x80\xCD\x60\x01\x00\xCC\x7C\x01\x00\xCC\x10"
-     "\x01\x00"
-     "\xCD\x6C\x01\x00\xCC\xFC\x04\x00\x17\x0D\x00\x80\x88\xA9\x3B\x00\x1A\x0D"
-     "\x00\x80"
-     "\x1D\x0D\x00\x80\x20\x0D\x00\x80\x86\x0D\x00\x00\x23\x0D\x00\x80\x26\x0D"
-     "\x00\x80"
-     "\xC5\x0D\x00\x80\xC8\x0D\x00\x80\x9C\x0C\x00\x80\x9F\x0C\x00\x80\xA5\x0C"
-     "\x00\x80"
-     "\xCE\x0D\x00\x80\xCD\x40\x0C\x00\xAD\x0D\x00\x80\xD7\x0D\x00\x80\xDA\x0D"
-     "\x00\x80"
-     "\xE0\x0D\x00\x80\x88\x29\x00\x00\xB4\x0C\x00\x80\x4F\x1F\x00\x80\xBE\x58"
-     "\x01\x00"
-     "\xA7\x1F\x00\x80\x66\x0C\x00\x80\x77\x20\x00\x80\xBA\xAC\x01\x00\x42\x0C"
-     "\x00\x80"
-     "\xB8\xAC\x01\x00\xB7\x78\x01\x00\xB6\xBC\x04\x00\xB5\x40\x01\x00\x0B\x0D"
-     "\x00\x80"
-     "\xB3\x50\x01\x00\x99\x0C\x00\x80\xA2\x0C\x00\x80\xB0\x28\x04\x00\xAF\x05"
-     "\x02\x00"
-     "\xAE\x71\x02\x00\xAD\x4D\x3C\x00\xAC\x61\x02\x00\xAB\x51\x3C\x00\xAA\x75"
-     "\x02\x00"
-     "\xA9\x4D\x02\x00\xA8\x51\x02\x00\xA7\x01\x3C\x00\xA6\x61\x02\x00\xA5\x69"
-     "\x02\x00"
-     "\xA4\x95\x01\x00\xA3\x99\x01\x00\xA2\x95\x01\x00\xA1\xA5\x01\x00\xA0\xA5"
-     "\x01\x00"
-     "\x9F\x51\x01\x00\x9E\x45\x01\x00\x9D\x3D\x01\x00\x9C\xF1\x01\x00\x9B\xFD"
-     "\x01\x00"
-     "\x9A\xC1\x01\x00\x99\x65\x07\x00\x98\xB1\x01\x00\x97\x81\x01\x00\x96\xBD"
-     "\x01\x00"
-     "\x95\xA5\x01\x00\x94\x71\x07\x00\x93\x89\x00\x00\x92\x6D\x07\x00\x91\x6D"
-     "\x07\x00"
-     "\x90\x89\x00\x00\x8F\x09\x07\x00\x8E\x01\x07\x00\x8D\x75\x07\x00\x8C\x75"
-     "\x07\x00"
-     "\x8B\x65\x07\x00\x8A\x79\x07\x00\x89\x4D\x07\x00\x88\x4D\x07\x00\x87\x6D"
-     "\x07\x00"
-     "\x86\xC9\x00\x00\x85\x61\x07\x00\x84\x69\x07\x00\x83\x85\x07\x00\x82\x85"
-     "\x07\x00"
-     "\x81\x95\x07\x00\x80\x95\x07\x00\xB0\x59\x0E\x00\xB1\x1D\x01\x00\xB2\xDD"
-     "\x0E\x00"
-     "\xA8\x0C\x00\x80\xB4\x91\x0F\x00\xB5\xDD\x09\x00\xAB\x0C\x00\x80\xB7\x51"
-     "\x06\x00"
-     "\xAE\x0C\x00\x80\xB9\xF1\x0E\x00\xBA\x79\x06\x00\xB1\x0C\x00\x80\xE3\x0D"
-     "\x00\x80"
-     "\xBD\x61\x06\x00\xBE\x61\x06\x00\xBF\x61\x06\x00\xA0\x4D\x06\x00\x80\x2D"
-     "\x06\x00"
-     "\xA2\x59\x06\x00\x86\xA5\x06\x00\xA4\x45\x06\x00\xB7\x0C\x00\x80\xA6\x41"
-     "\x06\x00"
-     "\x86\xBD\x06\x00\xA8\x7D\x06\x00\x88\x09\x06\x00\xAA\x59\x06\x00\xBA\x0C"
-     "\x00\x80"
-     "\xAC\x45\x06\x00\x88\x99\x06\x00\xAE\x31\x06\x00\xBD\x0C\x00\x80\x90\x95"
-     "\x0C\x00"
-     "\x91\x05\x0F\x00\x92\xC1\x05\x00\x93\x09\x0D\x00\x94\xA9\x09\x00\x95\x3D"
-     "\x09\x00"
-     "\x96\x2D\x0E\x00\x88\xAD\x00\x00\x98\x0D\x07\x00\x86\x59\x06\x00\x9A\x7D"
-     "\x07\x00"
-     "\x84\x59\x06\x00\x9C\x6D\x07\x00\xC0\x0C\x00\x80\x9E\xB1\x06\x00\x80\x01"
-     "\x06\x00"
-     "\xC3\x0C\x00\x80\x81\x1C\x07\x00\x82\x1D\x0E\x00\xE6\x0D\x00\x80\xC6\x0C"
-     "\x00\x80"
-     "\x85\x14\x07\x00\xC9\x0C\x00\x80\xCC\x0C\x00\x80\xCF\x0C\x00\x80\x89\xBC"
-     "\x04\x00"
-     "\xD2\x0C\x00\x80\xF2\x0D\x00\x80\xF5\x0D\x00\x80\xF8\x0D\x00\x80\xD5\x0C"
-     "\x00\x80"
-     "\x8F\x59\x07\x00\xD8\x0C\x00\x80\xDB\x0C\x00\x80\xFB\x0D\x00\x80\xE1\x0C"
-     "\x00\x80"
-     "\xCC\xB8\x13\x00\xE4\x0C\x00\x80\xCD\xF4\x0F\x00\xCC\xA0\x01\x00\xE7\x0C"
-     "\x00\x80"
-     "\x01\x0E\x00\x80\xCC\x7C\x01\x00\xED\x0C\x00\x80\xEF\x0D\x00\x80\xF3\x0C"
-     "\x00\x80"
-     "\xCD\x80\x17\x00\xCC\x78\x01\x00\x84\xAD\x03\x00\xF6\x0C\x00\x80\x86\x8D"
-     "\x05\x00"
-     "\x29\x0D\x00\x80\xCC\x94\x06\x00\xCD\x90\x06\x00\xCC\x1C\x01\x00\xCD\xE4"
-     "\x17\x00"
-     "\xCD\xAC\x06\x00\xCC\xA8\x06\x00\xCC\xC0\x13\x00\xCC\xA4\x13\x00\x88\x81"
-     "\x03\x00"
-     "\xCC\x24\x02\x00\xCC\x6C\x0A\x00\x8B\x89\x03\x00\xCC\x3C\x01\x00\xCD\x68"
-     "\x0F\x00"
-     "\xCC\x44\x06\x00\xCD\x40\x06\x00\xCC\x58\x06\x00\xCC\x94\x07\x00\xCC\xE0"
-     "\x12\x00"
-     "\x2C\x0D\x00\x80\x2F\x0D\x00\x80\x32\x0D\x00\x80\x35\x0D\x00\x80\x38\x0D"
-     "\x00\x80"
-     "\x3B\x0D\x00\x80\x3E\x0D\x00\x80\x41\x0D\x00\x80\x44\x0D\x00\x80\x4A\x0D"
-     "\x00\x80"
-     "\x4D\x0D\x00\x80\x50\x0D\x00\x80\xCC\x70\x07\x00\x53\x0D\x00\x80\x56\x0D"
-     "\x00\x80"
-     "\xCC\x88\x02\x00\x59\x0D\x00\x80\x5C\x0D\x00\x80\x5F\x0D\x00\x80\x62\x0D"
-     "\x00\x80"
-     "\x65\x0D\x00\x80\xCC\x38\x13\x00\x68\x0D\x00\x80\x6B\x0D\x00\x80\x88\xD5"
-     "\x3E\x00"
-     "\x6E\x0D\x00\x80\xCC\xDC\x06\x00\xCC\xCC\x07\x00\xCD\xC8\x07\x00\x71\x0D"
-     "\x00\x80"
-     "\x96\xBC\x02\x00\x74\x0D\x00\x80\xCC\x64\x13\x00\x77\x0D\x00\x80\xCD\x98"
-     "\x07\x00"
-     "\xCC\x0C\x04\x00\xCD\x9C\x04\x00\xCC\x68\x07\x00\xCD\x44\x07\x00\xCD\x7C"
-     "\x07\x00"
-     "\xCD\x10\x07\x00\x83\x04\x02\x00\xCC\xD4\x06\x00\x7A\x0D\x00\x80\x7D\x0D"
-     "\x00\x80"
-     "\x87\x18\x02\x00\xCC\x88\x03\x00\x80\x0D\x00\x80\x83\x0D\x00\x80\x8B\xC0"
-     "\x02\x00"
-     "\x86\x0D\x00\x80\x89\x0D\x00\x80\x8C\x0D\x00\x80\x8F\x0D\x00\x80\x92\x0D"
-     "\x00\x80"
-     "\x8D\xDC\x02\x00\x95\x0D\x00\x80\x98\x0D\x00\x80\xB2\x39\x01\x00\x9B\x0D"
-     "\x00\x80"
-     "\xB0\x35\x01\x00\x9E\x0D\x00\x80\xB6\x55\x01\x00\xB5\x64\x02\x00\xB4\x95"
-     "\x3F\x00"
-     "\xA1\x0D\x00\x80\xBA\x6D\x01\x00\xA4\x0D\x00\x80\xB8\x69\x01\x00\xA7\x0D"
-     "\x00\x80"
-     "\xBE\x61\x01\x00\xAA\x0D\x00\x80\xBC\x6D\x01\x00\xD9\x88\x3E\x00\xA2\x59"
-     "\x01\x00"
-     "\xC2\x0D\x00\x80\xA0\x81\x01\x00\xB0\x0D\x00\x80\xA6\x71\x01\x00\xB3\x0D"
-     "\x00\x80"
-     "\xA4\x7D\x01\x00\xB6\x0D\x00\x80\xAA\x49\x01\x00\xB9\x0D\x00\x80\xA8\x45"
-     "\x01\x00"
-     "\xBC\x0D\x00\x80\xAE\x41\x01\x00\xBF\x0D\x00\x80\xAC\x4D\x01\x00\xEA\x0C"
-     "\x00\x80"
-     "\xF0\x0C\x00\x80\x47\x0D\x00\x80\x80\xD1\x3E\x00\xCB\x0D\x00\x80\xD1\x0D"
-     "\x00\x80"
-     "\xD4\x0D\x00\x80\xDD\x0D\x00\x80\xE9\x0D\x00\x80\x8A\xC5\x3E\x00\xEC\x0D"
-     "\x00\x80"
-     "\xFE\x0D\x00\x80\x04\x0E\x00\x80\x8E\xDD\x3E\x00\x07\x0E\x00\x80\x8C\xD9"
-     "\x3E\x00"
-     "\x0A\x0E\x00\x80\x92\xB1\x3E\x00\x0D\x0E\x00\x80\x90\xA1\x3E\x00\x10\x0E"
-     "\x00\x80"
-     "\x96\x55\x3F\x00\x13\x0E\x00\x80\x94\xA5\x3E\x00\xDE\x0C\x00\x80\x9A\x7D"
-     "\x3F\x00"
-     "\x16\x0E\x00\x80\x98\x7D\x3F\x00\x19\x0E\x00\x80\x9E\x3D\x3F\x00\x1C\x0E"
-     "\x00\x80"
-     "\x9C\x5D\x3F\x00\x1F\x0E\x00\x80\xA2\xDD\x3F\x00\x22\x0E\x00\x80\xA0\xC1"
-     "\x3F\x00"
-     "\x25\x0E\x00\x80\xA6\xF1\x3F\x00\x28\x0E\x00\x80\xA4\xC1\x3F\x00\x2B\x0E"
-     "\x00\x80"
-     "\xAA\xC9\x3F\x00\x2E\x0E\x00\x80\xA8\xC5\x3F\x00\x31\x0E\x00\x80\xAE\xC1"
-     "\x3F\x00"
-     "\x34\x0E\x00\x80\xAC\xCD\x3F\x00\x37\x0E\x00\x80\xB2\x45\x3F\x00\x3A\x0E"
-     "\x00\x80"
-     "\xB0\xB1\x3F\x00\x3D\x0E\x00\x80\xB6\x45\x3F\x00\x40\x0E\x00\x80\xB4\x55"
-     "\x3F\x00"
-     "\x43\x0E\x00\x80\xBA\x4D\x3F\x00\x46\x0E\x00\x80\xB8\x4D\x3F\x00\x49\x0E"
-     "\x00\x80"
-     "\xBE\x4D\x3F\x00\x4C\x0E\x00\x80\xBC\x4D\x3F\x00\x80\xC9\x3E\x00\x81\xED"
-     "\x3A\x00"
-     "\x4F\x0E\x00\x80\x83\xDD\x3E\x00\x52\x0E\x00\x80\x85\xCD\x3E\x00\x55\x0E"
-     "\x00\x80"
-     "\x87\xFD\x3E\x00\x58\x0E\x00\x80\x89\xCD\x3E\x00\x5B\x0E\x00\x80\x8B\xDD"
-     "\x3E\x00"
-     "\x5E\x0E\x00\x80\x8D\xCD\x3E\x00\x61\x0E\x00\x80\x69\x0E\x00\x80\x90\x8D"
-     "\x3A\x00"
-     "\x6C\x0E\x00\x80\x92\x99\x3A\x00\x6F\x0E\x00\x80\x94\x41\x01\x00\x72\x0E"
-     "\x00\x80"
-     "\x96\xB9\x3A\x00\x75\x0E\x00\x80\x98\x75\x01\x00\x99\x10\x3D\x00\x9A\x71"
-     "\x01\x00"
-     "\x78\x0E\x00\x80\x9C\x9D\x3A\x00\x7B\x0E\x00\x80\x9E\x89\x3A\x00\x7E\x0E"
-     "\x00\x80"
-     "\xA0\x9D\x01\x00\x81\x0E\x00\x80\xA2\x7D\x3A\x00\x84\x0E\x00\x80\xA4\xA1"
-     "\x3C\x00"
-     "\x87\x0E\x00\x80\xA6\x55\x3C\x00\x8A\x0E\x00\x80\xA8\x81\x01\x00\xA9\xC0"
-     "\x3D\x00"
-     "\xAA\x81\x01\x00\x8D\x0E\x00\x80\xAC\x81\x3F\x00\x90\x0E\x00\x80\xAE\x55"
-     "\x3C\x00"
-     "\x93\x0E\x00\x80\xB0\x31\x3C\x00\xAA\x0E\x00\x80\xB2\x3D\x3C\x00\xAD\x0E"
-     "\x00\x80"
-     "\xB4\xD9\x3F\x00\xB0\x0E\x00\x80\xB6\xF5\x01\x00\xB3\x0E\x00\x80\xB8\xC5"
-     "\x3F\x00"
-     "\xB6\x0E\x00\x80\xBA\xD9\x01\x00\xB9\x0E\x00\x80\xBC\xC9\x01\x00\xBC\x0E"
-     "\x00\x80"
-     "\xBE\xB9\x01\x00\xBF\x0E\x00\x80\xC2\x0E\x00\x80\x82\x55\x02\x00\xC5\x0E"
-     "\x00\x80"
-     "\x80\x45\x02\x00\xCD\x0E\x00\x80\x86\x7D\x02\x00\xD0\x0E\x00\x80\x84\x45"
-     "\x02\x00"
-     "\xD3\x0E\x00\x80\x8A\x5D\x02\x00\xD6\x0E\x00\x80\x88\x5D\x02\x00\xD9\x0E"
-     "\x00\x80"
-     "\x8E\xBD\x02\x00\xDC\x0E\x00\x80\x8C\xBD\x02\x00\xDF\x0E\x00\x80\x92\xDD"
-     "\x02\x00"
-     "\xE2\x0E\x00\x80\x90\xDD\x02\x00\x64\x0E\x00\x80\x96\xFD\x02\x00\xA7\xFC"
-     "\x3E\x00"
-     "\x94\xFD\x02\x00\xBC\x0F\x00\x80\x9A\xDD\x02\x00\xBF\x0F\x00\x80\x98\xDD"
-     "\x02\x00"
-     "\xC5\x0F\x00\x80\x9E\xBD\x02\x00\xD9\xE4\x02\x00\x9C\xBD\x02\x00\xC2\x0F"
-     "\x00\x80"
-     "\xA2\x5D\x02\x00\xC8\x0F\x00\x80\xA0\x5D\x02\x00\x3A\x10\x00\x80\xA6\x7D"
-     "\x02\x00"
-     "\x5E\x16\x00\x80\xA4\x7D\x02\x00\x2D\x17\x00\x80\xAA\x5D\x02\x00\xBC\x16"
-     "\x00\x80"
-     "\xA8\x5D\x02\x00\x21\x17\x00\x80\xAE\xA5\x03\x00\x44\x17\x00\x80\xAC\xB5"
-     "\x03\x00"
-     "\xB3\xE9\x03\x00\xB2\xE5\x03\x00\xB1\xD1\x03\x00\x94\x4D\x00\x00\xB7\x95"
-     "\x03\x00"
-     "\xB6\xE9\x03\x00\xB5\xE1\x03\x00\xB4\xED\x03\x00\xBB\x8D\x03\x00\xBA\xB9"
-     "\x03\x00"
-     "\xB9\xBD\x03\x00\xB8\xA1\x03\x00\xBF\x7D\x03\x00\xBE\x71\x03\x00\xBD\x85"
-     "\x03\x00"
-     "\xBC\x89\x03\x00\xD9\x58\x02\x00\x93\x65\x01\x00\x94\x81\x01\x00\x1B\x17"
-     "\x00\x80"
-     "\xCC\xCC\x3F\x00\xE0\xCC\x02\x00\x94\x71\x01\x00\x95\x6D\x01\x00\x81\x6C"
-     "\x03\x00"
-     "\xCC\xB4\x3A\x00\xD9\x3C\x02\x00\x71\x17\x00\x80\xCC\xF8\x3B\x00\xCC\x34"
-     "\x3A\x00"
-     "\xCC\x84\x3A\x00\xE0\xC4\x02\x00\xD9\x44\x02\x00\x94\xF5\x01\x00\xA4\xD8"
-     "\x02\x00"
-     "\x83\x17\x00\x80\xCC\x44\x37\x00\x94\xED\x01\x00\xD9\x54\x02\x00\x94\x51"
-     "\x00\x00"
-     "\xBC\x4D\x00\x00\xE0\x14\x02\x00\xA4\xE0\x02\x00\x92\x48\x03\x00\x95\x4C"
-     "\x03\x00"
-     "\x87\x17\x00\x80\xBC\x35\x00\x00\x47\x17\x00\x80\x80\x15\x02\x00\x81\x19"
-     "\x02\x00"
-     "\x82\x25\x02\x00\x83\x39\x02\x00\x84\x2D\x02\x00\x85\x51\x02\x00\x86\x55"
-     "\x02\x00"
-     "\x87\x49\x02\x00\x88\x7D\x02\x00\x89\x41\x02\x00\x8A\x45\x02\x00\x8B\x59"
-     "\x02\x00"
-     "\x8C\x4D\x02\x00\x8D\xB5\x01\x00\x8E\xB1\x01\x00\x8F\xAD\x01\x00\x90\xD9"
-     "\x01\x00"
-     "\x91\xE5\x01\x00\x92\xE1\x01\x00\x93\xFD\x01\x00\x94\xE9\x01\x00\x95\x95"
-     "\x01\x00"
-     "\x96\x91\x01\x00\xBC\xA9\x00\x00\xA4\xCC\x02\x00\x4E\x17\x00\x80\x55\x17"
-     "\x00\x80"
-     "\x5C\x17\x00\x80\x63\x17\x00\x80\x6A\x17\x00\x80\x75\x17\x00\x80\x7C\x17"
-     "\x00\x80"
-     "\xA4\xD4\x02\x00\xA5\xE4\x02\x00\xE0\xA8\x02\x00\xA7\xB0\x02\x00\xBE\x0D"
-     "\x00\x00"
-     "\xB0\xA8\x04\x00\xB3\xFC\x04\x00\xA0\x17\x00\x80\x97\x61\x00\x00\xAD\x38"
-     "\x03\x00"
-     "\xAE\x70\x03\x00\xAF\x1C\x03\x00\xA8\x58\x02\x00\xA9\xB0\x02\x00\xA6\xD0"
-     "\x02\x00"
-     "\xA7\x40\x02\x00\xA4\x17\x00\x80\xB5\x94\x03\x00\x8B\x17\x00\x80\xB7\x2C"
-     "\x00\x00"
-     "\x92\x17\x00\x80\xB1\x2C\x03\x00\xB2\x2C\x03\x00\xB3\xDC\x03\x00\xBC\x4C"
-     "\x00\x00"
-     "\xBD\x64\x01\x00\xBE\x74\x01\x00\x99\x17\x00\x80\xB8\x90\x00\x00\xA8\xE0"
-     "\x04\x00"
-     "\xBA\x88\x00\x00\xBB\x44\x00\x00\x98\xFD\x07\x00\x99\x61\x07\x00\x9A\x61"
-     "\x07\x00"
-     "\x9B\x61\x07\x00\x9C\x61\x07\x00\x9D\x61\x07\x00\x9E\x61\x07\x00\x9F\x61"
-     "\x07\x00"
-     "\x87\x28\x03\x00\xB3\x0D\x00\x00\xC4\x17\x00\x80\xCB\x17\x00\x80\xB6\x1D"
-     "\x00\x00"
-     "\xA8\x17\x00\x80\xAF\x17\x00\x80\xB6\x17\x00\x80\x9F\xAD\x00\x00\xBD\x17"
-     "\x00\x80"
-     "\x9D\x99\x00\x00\x9C\x85\x00\x00\xE0\xBC\x02\x00\xBE\x31\x00\x00\xAD\x70"
-     "\x02\x00"
-     "\xAC\xF0\x02\x00\x99\x55\x00\x00\xE4\x17\x00\x80\x9B\x55\x00\x00\x9A\x55"
-     "\x00\x00"
-     "\x96\x09\x00\x00\x97\xF1\x01\x00\xE0\x17\x00\x80\x9E\x39\x00\x00\x87\x50"
-     "\x02\x00"
-     "\xE8\x17\x00\x80\xD2\x17\x00\x80\xD9\x17\x00\x80\x92\x84\x03\x00\xE0\xB0"
-     "\x02\x00"
-     "\xAF\x58\x02\x00\x97\x3D\x00\x00\xEC\x17\x00\x80\xE0\xA8\x02\x00\x86\x8C"
-     "\x03\x00"
-     "\x87\xF4\x03\x00\xAF\x24\x02\x00\xAE\xF4\x02\x00\xBE\x05\x00\x00\xF0\x17"
-     "\x00\x80"
-     "\xB6\xD9\x07\x00\xB7\xE5\x07\x00\x97\x19\x00\x00\xB5\xDD\x07\x00\xF8\x17"
-     "\x00\x80"
-     "\xAE\xC8\x02\x00\xE0\xB4\x02\x00\xF4\x17\x00\x80\x86\x90\x03\x00\xBE\x39"
-     "\x00\x00"
-     "\x9D\x65\x00\x00\x9C\x65\x00\x00\xE0\xC0\x02\x00\xB1\x50\x02\x00\xB8\xD1"
-     "\x07\x00"
-     "\x96\xFD\x00\x00\xFC\x17\x00\x80\xE0\xD8\x02\x00\xBF\x8C\x03\x00\x95\x19"
-     "\x00\x00"
-     "\xB3\x48\x02\x00\x00\x18\x00\x80\x86\x84\x03\x00\xE0\xF0\x02\x00\xB3\x0C"
-     "\x02\x00"
-     "\x82\x0D\x00\x00\x0C\x18\x00\x80\xB3\x4C\x02\x00\x8A\xD4\x05\x00\x95\x09"
-     "\x00\x00"
-     "\x88\xA4\x04\x00\x10\x18\x00\x80\x04\x18\x00\x80\x08\x18\x00\x80\x86\x98"
-     "\x03\x00"
-     "\x87\xBC\x03\x00\xE0\xD8\x02\x00\xBE\x35\x00\x00\xB4\xF4\x02\x00\xB5\x64"
-     "\x02\x00"
-     "\x14\x18\x00\x80\x97\x0D\x00\x00\x1C\x18\x00\x80\xB4\xE8\x02\x00\xE0\xCC"
-     "\x02\x00"
-     "\x96\x31\x00\x00\x95\x39\x00\x00\xBE\xFD\x03\x00\x18\x18\x00\x80\x99\x8C"
-     "\x03\x00"
-     "\xE0\xD8\x02\x00\xB7\x04\x02\x00\x20\x18\x00\x80\xB7\x24\x02\x00\x8A\x3D"
-     "\x00\x00"
-     "\x8C\x45\x00\x00\x8A\x31\x00\x00\x28\x18\x00\x80\xE0\x48\x01\x00\xB3\x15"
-     "\x00\x00"
-     "\x2C\x18\x00\x80\x8F\x01\x0C\x00\x30\x18\x00\x80\xB3\x7D\x00\x00\x45\x18"
-     "\x00\x80"
-     "\x37\x18\x00\x80\x9C\x0D\x00\x00\x9D\x19\x00\x00\xE1\x20\x02\x00\x3E\x18"
-     "\x00\x80"
-     "\x4C\x18\x00\x80\x57\x18\x00\x80\x5E\x18\x00\x80\x65\x18\x00\x80\xA5\xBC"
-     "\x03\x00"
-     "\xAE\x89\x00\x00\x80\xA4\x02\x00\x9F\x45\x00\x00\x6C\x18\x00\x80\x85\xC8"
-     "\x0E\x00"
-     "\x73\x18\x00\x80\x84\x48\x0F\x00\x50\x18\x00\x80\x7A\x18\x00\x80\xEF\x9C"
-     "\x02\x00"
-     "\x84\xD8\x0F\x00\x81\x18\x00\x80\xC3\x18\x00\x80\xE3\x60\x02\x00\xBE\x6C"
-     "\x0F\x00"
-     "\xE1\xC0\x02\x00\xB9\x18\x00\x80\xD4\x18\x00\x80\xCA\x18\x00\x80\x88\x18"
-     "\x00\x80"
-     "\xBE\xB0\x08\x00\x96\x18\x00\x80\x9D\x18\x00\x80\xA4\x18\x00\x80\xAB\x18"
-     "\x00\x80"
-     "\xB2\x18\x00\x80\x85\xE8\x60\x00\xBE\x94\x09\x00\x8F\x18\x00\x80\xBE\xA8"
-     "\x0A\x00"
-     "\xDB\x18\x00\x80\xBE\xA4\x0B\x00\x97\x23\x00\x80\xBF\x78\x61\x00\x9B\x23"
-     "\x00\x80"
-     "\x9F\x23\x00\x80\xA3\x23\x00\x80\xA7\x23\x00\x80\x84\x0C\x09\x00\x84\x44"
-     "\x0B\x00"
-     "\x83\x51\x01\x00\xAB\x23\x00\x80\xBE\xDC\x14\x00\x81\xA1\x01\x00\xBE\x10"
-     "\x17\x00"
-     "\xAF\x23\x00\x80\x84\xA4\x0A\x00\xB3\x23\x00\x80\xBF\x00\x44\x00\xBE\xBC"
-     "\x11\x00"
-     "\x8D\x55\x01\x00\xB7\x23\x00\x80\xCF\x00\x00\x00\xBE\xCC\x10\x00\xBB\x23"
-     "\x00\x80"
-     "\x92\x2D\x01\x00\xBF\x23\x00\x80\x93\x91\x01\x00\xC3\x23\x00\x80\x84\x94"
-     "\x14\x00"
-     "\x97\x25\x01\x00\x84\x6C\x14\x00\xC7\x23\x00\x80\x9D\xAD\x01\x00\x84\x6C"
-     "\x16\x00"
-     "\x9C\xF1\x01\x00\x84\xFC\x11\x00\x84\xF8\x10\x00\xCB\x23\x00\x80\xA7\x4D"
-     "\x01\x00"
-     "\xCF\x23\x00\x80\x86\xB4\x0F\x00\x87\x2C\x0C\x00\xD3\x23\x00\x80\xA2\x5D"
-     "\x01\x00"
-     "\x87\xD0\x0E\x00\x86\x3C\x09\x00\xD7\x23\x00\x80\xA9\x2D\x01\x00\xDB\x23"
-     "\x00\x80"
-     "\xAC\x7D\x01\x00\x84\xB0\x12\x00\xDF\x23\x00\x80\xBE\xF4\x10\x00\xE3\x23"
-     "\x00\x80"
-     "\xBE\xC0\x13\x00\xE7\x23\x00\x80\xEB\x23\x00\x80\xB3\x41\x01\x00\xBE\x98"
-     "\x1D\x00"
-     "\xB5\x4D\x01\x00\xB6\x45\x01\x00\xB7\x71\x01\x00\xB8\x51\x01\x00\xB9\x51"
-     "\x01\x00"
-     "\xEF\x23\x00\x80\x84\xA0\x1D\x00\xF3\x23\x00\x80\xF7\x23\x00\x80\xB9\x25"
-     "\x01\x00"
-     "\xFB\x23\x00\x80\x80\xFC\x0C\x00\xFF\x23\x00\x80\x82\x88\x03\x00\x83\x44"
-     "\x02\x00"
-     "\x84\xD0\x03\x00\x03\x24\x00\x80\x07\x24\x00\x80\x0B\x24\x00\x80\x0F\x24"
-     "\x00\x80"
-     "\x13\x24\x00\x80\x17\x24\x00\x80\x1B\x24\x00\x80\x1F\x24\x00\x80\x23\x24"
-     "\x00\x80"
-     "\x27\x24\x00\x80\x8F\x5C\x1B\x02\x82\x71\x00\x00\x83\x71\x00\x00\x80\x61"
-     "\x00\x00"
-     "\x81\x61\x00\x00\x2B\x24\x00\x80\x87\x35\x00\x00\x84\x61\x00\x00\x85\x0D"
-     "\x00\x00"
-     "\x2F\x24\x00\x80\x33\x24\x00\x80\x0F\x19\x00\x80\x3F\x3D\x00\x80\x43\x3D"
-     "\x00\x80"
-     "\x47\x3D\x00\x80\x4B\x3D\x00\x80\x8D\x19\x00\x00\x4F\x3D\x00\x80\x53\x3D"
-     "\x00\x80"
-     "\x57\x3D\x00\x80\x5B\x3D\x00\x80\x5F\x3D\x00\x80\x63\x3D\x00\x80\x67\x3D"
-     "\x00\x80"
-     "\x6B\x3D\x00\x80\x6F\x3D\x00\x80\x73\x3D\x00\x80\x77\x3D\x00\x80\x7B\x3D"
-     "\x00\x80"
-     "\xAC\xA4\x18\x02\xAD\x0C\x2D\x02\xBC\xD1\x00\x00\x7F\x3D\x00\x80\x83\x3D"
-     "\x00\x80"
-     "\x87\x3D\x00\x80\xB2\xC8\x19\x02\x8B\x3D\x00\x80\xB4\xA0\x18\x02\xB5\x00"
-     "\x1C\x02"
-     "\xB6\xB0\x1E\x02\x8F\x3D\x00\x80\xB8\xE0\x1F\x02\xB9\xE0\x1F\x02\xBA\xA4"
-     "\x1D\x02"
-     "\xBB\x00\x10\x02\xBC\xD0\x12\x02\xBD\xD4\x13\x02\xBE\xF8\x12\x02\xBF\xD4"
-     "\x11\x02"
-     "\xA0\xED\x0F\x00\xA1\xF1\x0F\x00\xA2\xF1\x0F\x00\xA3\xF1\x0F\x00\xA4\x11"
-     "\x0E\x00"
-     "\xA5\x05\x0E\x00\xA6\x39\x0E\x00\xA7\x35\x0E\x00\xA8\x19\x0E\x00\xA9\x61"
-     "\x0E\x00"
-     "\xAA\x65\x0E\x00\xAB\x79\x0E\x00\xAC\x55\x0E\x00\xAD\x49\x0E\x00\xAE\xBD"
-     "\x0E\x00"
-     "\xAF\xAD\x0E\x00\xB0\xE1\x0E\x00\xB1\xED\x0E\x00\xB2\xFD\x0E\x00\xB3\xF1"
-     "\x0E\x00"
-     "\xB4\x95\x0E\x00\xB5\x9D\x0E\x00\xB6\xB1\x0E\x00\xB7\xAD\x0E\x00\xB8\x95"
-     "\x0E\x00"
-     "\xB9\x99\x0E\x00\xBA\x6D\x01\x00\xBB\x79\x01\x00\xBC\x69\x01\x00\xBD\x69"
-     "\x01\x00"
-     "\xBE\x59\x01\x00\xBF\x59\x01\x00\x90\x94\xF4\x01\x91\x84\xD2\x01\x92\x9C"
-     "\x3F\x02"
-     "\x93\x3D\x00\x80\x97\x3D\x00\x80\x9B\x3D\x00\x80\x9F\x3D\x00\x80\xA3\x3D"
-     "\x00\x80"
-     "\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\xA7\x3D\x00\x80\xAB\x3D"
-     "\x00\x80"
-     "\xE0\x7C\x0D\x00\xB3\x3D\x00\x80\xB7\x3D\x00\x80\x80\xEC\x0C\x00\x81\xC4"
-     "\x69\x00"
-     "\x82\x7C\x75\x00\x83\xBC\x50\x00\x84\x50\xBC\x00\x85\x44\x9B\x00\x86\x0C"
-     "\xE7\x00"
-     "\x87\xDC\xCD\x00\x88\x70\x29\x01\x89\x64\x35\x01\x8A\xE8\x10\x01\x8B\x6C"
-     "\x7F\x01"
-     "\x8C\xE4\x5B\x01\x8D\xD0\xA1\x01\x8E\xF8\x8C\x01\x8F\x98\xEB\x01\xAD\x21"
-     "\x12\x00"
-     "\xAC\x01\x10\x00\xAF\x01\x14\x00\xAE\xE1\x12\x00\xA9\x15\x1E\x00\xA8\x01"
-     "\x1C\x00"
-     "\xAB\x49\x1F\x00\xAA\x15\x1E\x00\xA5\x01\x18\x00\xA4\xB5\x06\x00\xA7\x0D"
-     "\x1A\x00"
-     "\xA6\x0D\x1A\x00\xA1\x5D\x02\x00\x24\x18\x00\x80\xA3\x19\x06\x00\xA2\x01"
-     "\x04\x00"
-     "\xBB\x3D\x00\x80\xBF\x3D\x00\x80\x84\x44\x11\x00\xC3\x3D\x00\x80\xC7\x3D"
-     "\x00\x80"
-     "\xCB\x3D\x00\x80\xCF\x3D\x00\x80\xD3\x3D\x00\x80\xB5\x01\x68\x00\xB4\x59"
-     "\x6A\x00"
-     "\xD7\x3D\x00\x80\xDB\x3D\x00\x80\xB1\x45\x16\x00\xB0\x55\x16\x00\xB3\x59"
-     "\x6A\x00"
-     "\xB2\x01\x68\x00\xAF\x11\x02\x00\xAE\x09\x02\x00\xAD\x01\x02\x00\xAC\x09"
-     "\x02\x00"
-     "\xAB\x11\x02\x00\xAA\xED\x02\x00\xA9\xED\x02\x00\xA8\xED\x02\x00\xDF\x3D"
-     "\x00\x80"
-     "\xE3\x3D\x00\x80\xE7\x3D\x00\x80\xEB\x3D\x00\x80\xEF\x3D\x00\x80\xF3\x3D"
-     "\x00\x80"
-     "\xF7\x3D\x00\x80\xFB\x3D\x00\x80\xBF\x59\x03\x00\xBE\x51\x03\x00\xBD\x59"
-     "\x03\x00"
-     "\xBC\x41\x03\x00\xBB\x0D\x02\x00\xBA\x19\x02\x00\xB9\x1D\x02\x00\xB8\x1D"
-     "\x02\x00"
-     "\xB7\x5D\x02\x00\xB6\x49\x02\x00\xB5\x41\x02\x00\xB4\x49\x02\x00\xB3\x51"
-     "\x02\x00"
-     "\xB2\x69\x02\x00\xB1\x61\x02\x00\xB0\x69\x02\x00\xB3\x79\x03\x00\xFF\x3D"
-     "\x00\x80"
-     "\x03\x3E\x00\x80\x07\x3E\x00\x80\x0B\x3E\x00\x80\xB6\x9D\x03\x00\xB5\x9D"
-     "\x03\x00"
-     "\x0F\x3E\x00\x80\xBB\xB9\x03\x00\xBA\xB1\x03\x00\x13\x3E\x00\x80\x17\x3E"
-     "\x00\x80"
-     "\xBF\x99\x03\x00\xBE\x91\x03\x00\xBD\x99\x03\x00\xBC\xA1\x03\x00\x82\x15"
-     "\x00\x00"
-     "\xA3\x3D\x03\x00\x80\x61\x00\x00\x81\x61\x00\x00\xA6\xD9\x03\x00\x1B\x3E"
-     "\x00\x80"
-     "\x23\x3E\x00\x80\xA5\xD9\x03\x00\xAA\xF5\x03\x00\xAB\xFD\x03\x00\x86\xC8"
-     "\x04\x00"
-     "\x87\xE0\x05\x00\xAE\xD5\x03\x00\xAF\xDD\x03\x00\xAC\xE5\x03\x00\xAD\xDD"
-     "\x03\x00"
-     "\xB3\xC1\x01\x00\x27\x3E\x00\x80\x2B\x3E\x00\x80\x2F\x3E\x00\x80\x33\x3E"
-     "\x00\x80"
-     "\xB6\xC5\x01\x00\xB5\xD5\x01\x00\x37\x3E\x00\x80\xBB\xC9\x01\x00\xBA\xF9"
-     "\x01\x00"
-     "\x3B\x3E\x00\x80\x3F\x3E\x00\x80\xBF\x19\x01\x00\xBE\x11\x01\x00\xBD\x19"
-     "\x01\x00"
-     "\xBC\xD1\x01\x00\xBE\x40\x13\x00\x43\x3E\x00\x80\x47\x3E\x00\x80\x4B\x3E"
-     "\x00\x80"
-     "\xEF\x8C\x0F\x00\x4F\x3E\x00\x80\x53\x3E\x00\x80\x57\x3E\x00\x80\xE3\x74"
-     "\x0F\x00"
-     "\x5B\x3E\x00\x80\xE1\x78\x0E\x00\x5F\x3E\x00\x80\x63\x3E\x00\x80\x67\x3E"
-     "\x00\x80"
-     "\x6B\x3E\x00\x80\x6F\x3E\x00\x80\xA8\xB9\x05\x00\xA9\xB9\x05\x00\xAA\x49"
-     "\x06\x00"
-     "\xAB\x49\x06\x00\xAC\x5D\x06\x00\xAD\x45\x06\x00\xAE\x4D\x06\x00\xAF\x45"
-     "\x06\x00"
-     "\x1F\x3E\x00\x80\x73\x3E\x00\x80\x77\x3E\x00\x80\x7B\x3E\x00\x80\x7F\x3E"
-     "\x00\x80"
-     "\x83\x3E\x00\x80\x87\x3E\x00\x80\x8B\x3E\x00\x80\xB8\xF5\x06\x00\xB9\xFD"
-     "\x06\x00"
-     "\xBA\xF5\x06\x00\xBB\x49\x07\x00\xBC\x59\x07\x00\xBD\x59\x07\x00\xBE\x49"
-     "\x07\x00"
-     "\xBF\x41\x07\x00\xB0\x3D\x06\x00\xB1\x0D\x06\x00\xB2\x05\x06\x00\xB3\xCD"
-     "\x06\x00"
-     "\xB4\xD5\x06\x00\xB5\xDD\x06\x00\xB6\xD5\x06\x00\xB7\xCD\x06\x00\xA3\xC1"
-     "\x06\x00"
-     "\x82\xB9\x00\x00\x81\xB1\x00\x00\x80\xB9\x00\x00\x8F\x3E\x00\x80\xA6\xC5"
-     "\x06\x00"
-     "\xA5\xD5\x06\x00\x93\x3E\x00\x80\xAB\xC9\x06\x00\xAA\xF9\x06\x00\x86\x08"
-     "\x00\x00"
-     "\x87\x00\x03\x00\xAF\x19\x06\x00\xAE\x11\x06\x00\xAD\x19\x06\x00\xAC\xD1"
-     "\x06\x00"
-     "\x97\x3E\x00\x80\xB3\x4D\x07\x00\x9B\x3E\x00\x80\x9F\x3E\x00\x80\xB6\x4D"
-     "\x07\x00"
-     "\xA3\x3E\x00\x80\xA7\x3E\x00\x80\xB5\x4D\x07\x00\xBA\x25\x07\x00\xBB\x2D"
-     "\x07\x00"
-     "\xAB\x3E\x00\x80\xAF\x3E\x00\x80\xBE\x29\x07\x00\xBF\x29\x07\x00\xBC\x35"
-     "\x07\x00"
-     "\xBD\x29\x07\x00\xA8\x59\x06\x00\xA9\x81\x06\x00\xAA\x81\x06\x00\xAB\x81"
-     "\x06\x00"
-     "\xAC\x81\x06\x00\xAD\x89\x06\x00\xAE\xB1\x06\x00\xAF\xB1\x06\x00\xB3\x3E"
-     "\x00\x80"
-     "\xB7\x3E\x00\x80\xBB\x3E\x00\x80\xBF\x3E\x00\x80\xC3\x3E\x00\x80\xC7\x3E"
-     "\x00\x80"
-     "\xCB\x3E\x00\x80\xCF\x3E\x00\x80\xB8\xB1\x06\x00\xB9\xB1\x06\x00\xBA\xB1"
-     "\x06\x00"
-     "\xBB\xB1\x06\x00\xBC\x75\x01\x00\xBD\x7D\x01\x00\xBE\x75\x01\x00\xBF\x6D"
-     "\x01\x00"
-     "\xB0\xA5\x06\x00\xB1\xAD\x06\x00\xB2\xA5\x06\x00\xB3\xB9\x06\x00\xB4\xA9"
-     "\x06\x00"
-     "\xB5\xA9\x06\x00\xB6\x9D\x06\x00\xB7\x91\x06\x00\xEF\xC4\x09\x00\xD3\x3E"
-     "\x00\x80"
-     "\xD7\x3E\x00\x80\xDB\x3E\x00\x80\xDF\x3E\x00\x80\xE3\x3E\x00\x80\xE7\x3E"
-     "\x00\x80"
-     "\xEB\x3E\x00\x80\xEF\x3E\x00\x80\x82\x61\x00\x00\x81\x15\x00\x00\x80\x19"
-     "\x00\x00"
-     "\xE3\x34\x09\x00\xF3\x3E\x00\x80\xE1\x3C\x08\x00\xF7\x3E\x00\x80\xA3\x49"
-     "\x06\x00"
-     "\xFB\x3E\x00\x80\x86\x68\x1F\x00\x87\xCC\x00\x00\x03\x3F\x00\x80\xA6\x49"
-     "\x06\x00"
-     "\xA5\x49\x06\x00\x07\x3F\x00\x80\xAB\x29\x06\x00\xAA\x21\x06\x00\x0B\x3F"
-     "\x00\x80"
-     "\x0F\x3F\x00\x80\xAF\x2D\x06\x00\xAE\x2D\x06\x00\xAD\x2D\x06\x00\xAC\x31"
-     "\x06\x00"
-     "\xA8\x9D\x02\x00\xA9\xC5\x02\x00\xAA\xC1\x02\x00\xAB\xC1\x02\x00\xAC\xC5"
-     "\x02\x00"
-     "\xAD\xCD\x02\x00\xAE\xF1\x02\x00\xAF\xF1\x02\x00\x13\x3F\x00\x80\x17\x3F"
-     "\x00\x80"
-     "\x1B\x3F\x00\x80\x1F\x3F\x00\x80\x23\x3F\x00\x80\x27\x3F\x00\x80\x2B\x3F"
-     "\x00\x80"
-     "\xAF\x3D\x00\x80\xB8\xC5\x03\x00\xB9\xCD\x03\x00\xBA\xC5\x03\x00\xBB\xDD"
-     "\x03\x00"
-     "\xBC\xC5\x03\x00\xBD\xCD\x03\x00\xBE\xC5\x03\x00\xBF\xFD\x03\x00\xB0\x41"
-     "\x03\x00"
-     "\xB1\x41\x03\x00\xB2\x41\x03\x00\xB3\x41\x03\x00\xB4\x41\x03\x00\xB5\x41"
-     "\x03\x00"
-     "\xB6\x41\x03\x00\xB7\x41\x03\x00\xB3\x49\x02\x00\x2F\x3F\x00\x80\x33\x3F"
-     "\x00\x80"
-     "\xBE\xC0\x17\x00\x37\x3F\x00\x80\xB6\x4D\x02\x00\xB5\x4D\x02\x00\x3B\x3F"
-     "\x00\x80"
-     "\xBB\x85\x03\x00\xBA\x85\x03\x00\x3F\x3F\x00\x80\x43\x3F\x00\x80\xBF\x85"
-     "\x03\x00"
-     "\xBE\x85\x03\x00\xBD\x95\x03\x00\xBC\x95\x03\x00\x47\x3F\x00\x80\xA3\x0D"
-     "\x02\x00"
-     "\x4B\x3F\x00\x80\x4F\x3F\x00\x80\xA6\x09\x02\x00\x53\x3F\x00\x80\x57\x3F"
-     "\x00\x80"
-     "\xA5\x09\x02\x00\xAA\xC1\x03\x00\xAB\xC1\x03\x00\x5B\x3F\x00\x80\x5F\x3F"
-     "\x00\x80"
-     "\xAE\xC1\x03\x00\xAF\xC1\x03\x00\xAC\xD1\x03\x00\xAD\xD1\x03\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x11\x00\x00\x63\x3F\x00\x80\x67\x3F\x00\x80\xE1\xA4"
-     "\x01\x00"
-     "\x6B\x3F\x00\x80\xE3\x04\x0A\x00\x73\x3F\x00\x80\x77\x3F\x00\x80\x7B\x3F"
-     "\x00\x80"
-     "\xEF\x34\x0B\x00\x7F\x3F\x00\x80\x83\x3F\x00\x80\x87\x3F\x00\x80\x8B\x3F"
-     "\x00\x80"
-     "\x86\x64\x1D\x00\x87\x88\x02\x00\x8F\x3F\x00\x80\xB3\xA5\x02\x00\x93\x3F"
-     "\x00\x80"
-     "\xB5\x75\x01\x00\xB6\x7D\x01\x00\x97\x3F\x00\x80\x9B\x3F\x00\x80\x9F\x3F"
-     "\x00\x80"
-     "\xBA\x59\x01\x00\xBB\x59\x01\x00\xBC\x49\x01\x00\xBD\xF1\x01\x00\xBE\xF1"
-     "\x01\x00"
-     "\xBF\xED\x01\x00\xA3\x3F\x00\x80\xA7\x3F\x00\x80\xAB\x3F\x00\x80\xAF\x3F"
-     "\x00\x80"
-     "\xB3\x3F\x00\x80\xB7\x3F\x00\x80\xBB\x3F\x00\x80\xEF\xBC\x0B\x00\x84\x60"
-     "\x16\x00"
-     "\xE1\xD4\x06\x00\xBF\x3F\x00\x80\xE3\xB8\x0A\x00\xC3\x3F\x00\x80\xC7\x3F"
-     "\x00\x80"
-     "\xCB\x3F\x00\x80\xCF\x3F\x00\x80\xA3\x29\x02\x00\x82\x29\x00\x00\x81\x21"
-     "\x00\x00"
-     "\x80\x11\x00\x00\x84\x9C\x11\x00\xA6\xF1\x01\x00\xA5\xF9\x01\x00\xD3\x3F"
-     "\x00\x80"
-     "\xAB\xD5\x01\x00\xAA\xD5\x01\x00\xD7\x3F\x00\x80\xDB\x3F\x00\x80\xAF\x61"
-     "\x01\x00"
-     "\xAE\x7D\x01\x00\xAD\x7D\x01\x00\xAC\xC5\x01\x00\xA8\xD1\x1D\x00\xA9\xD9"
-     "\x1D\x00"
-     "\xAA\xE1\x1D\x00\xAB\xE1\x1D\x00\xAC\x31\x1E\x00\xAD\x31\x1E\x00\xAE\x31"
-     "\x1E\x00"
-     "\xAF\x31\x1E\x00\x6F\x3F\x00\x80\xDF\x3F\x00\x80\x86\x88\x00\x00\x87\xAC"
-     "\x03\x00"
-     "\xE3\x3F\x00\x80\xE7\x3F\x00\x80\xEB\x3F\x00\x80\xEF\x3F\x00\x80\xB8\xED"
-     "\x1E\x00"
-     "\xB9\x85\x1E\x00\xBA\x81\x1E\x00\xBB\x81\x1E\x00\xBC\x85\x1E\x00\xBD\x8D"
-     "\x1E\x00"
-     "\xBE\xB1\x1E\x00\xBF\xB1\x1E\x00\xB0\x51\x1E\x00\xB1\x51\x1E\x00\xB2\x51"
-     "\x1E\x00"
-     "\xB3\xE9\x1E\x00\xB4\xF9\x1E\x00\xB5\xE5\x1E\x00\xB6\xE5\x1E\x00\xB7\xD5"
-     "\x1E\x00"
-     "\xA9\xA5\x1F\x00\xA8\xA5\x1F\x00\xAB\xB9\x1F\x00\xAA\xA5\x1F\x00\xAD\xA9"
-     "\x1F\x00"
-     "\xAC\xA1\x1F\x00\xAF\xD1\x1F\x00\xAE\xA1\x1F\x00\xBE\xE8\x0E\x00\xF3\x3F"
-     "\x00\x80"
-     "\xF7\x3F\x00\x80\xFB\x3F\x00\x80\xFF\x3F\x00\x80\x03\x40\x00\x80\x07\x40"
-     "\x00\x80"
-     "\x0B\x40\x00\x80\xB9\x55\x1F\x00\xB8\x55\x1F\x00\xBB\x61\x1F\x00\xBA\x55"
-     "\x1F\x00"
-     "\xBD\x65\x1F\x00\xBC\x65\x1F\x00\xBF\x65\x1F\x00\xBE\x65\x1F\x00\xB1\xBD"
-     "\x1F\x00"
-     "\xB0\xB5\x1F\x00\xB3\x8D\x1F\x00\xB2\x8D\x1F\x00\xB5\x75\x1F\x00\xB4\x75"
-     "\x1F\x00"
-     "\xB7\x75\x1F\x00\xB6\x75\x1F\x00\xB3\x29\x1E\x00\x0F\x40\x00\x80\x13\x40"
-     "\x00\x80"
-     "\x17\x40\x00\x80\x1B\x40\x00\x80\xB6\x2D\x1E\x00\xB5\x2D\x1E\x00\x1F\x40"
-     "\x00\x80"
-     "\xBB\x7D\x1E\x00\xBA\x75\x1E\x00\x23\x40\x00\x80\x27\x40\x00\x80\xBF\xB5"
-     "\x1E\x00"
-     "\xBE\xB5\x1E\x00\xBD\x5D\x1E\x00\xBC\x5D\x1E\x00\x2B\x40\x00\x80\xA3\x6D"
-     "\x1E\x00"
-     "\x2F\x40\x00\x80\x33\x40\x00\x80\xA6\x69\x1E\x00\x37\x40\x00\x80\x3B\x40"
-     "\x00\x80"
-     "\xA5\x69\x1E\x00\xAA\x31\x1E\x00\xAB\x39\x1E\x00\x3F\x40\x00\x80\x43\x40"
-     "\x00\x80"
-     "\xAE\xF1\x1E\x00\xAF\xF1\x1E\x00\xAC\x19\x1E\x00\xAD\x19\x1E\x00\x80\xED"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x19\x00\x00\x47\x40\x00\x80\x4B\x40\x00\x80\x53\x40"
-     "\x00\x80"
-     "\x57\x40\x00\x80\xEF\x8C\x15\x00\xBE\x2C\x0D\x00\xE1\x74\x1E\x00\x5B\x40"
-     "\x00\x80"
-     "\xE3\x98\x14\x00\x86\x10\x00\x00\x87\x74\x01\x00\x5F\x40\x00\x80\x63\x40"
-     "\x00\x80"
-     "\xB3\xF9\x01\x00\x67\x40\x00\x80\x6B\x40\x00\x80\x6F\x40\x00\x80\x73\x40"
-     "\x00\x80"
-     "\xB6\x7D\x01\x00\xB5\x7D\x01\x00\x77\x40\x00\x80\xBB\x59\x01\x00\xBA\x51"
-     "\x01\x00"
-     "\x7B\x40\x00\x80\x7F\x40\x00\x80\xBF\xFD\x00\x00\xBE\xFD\x00\x00\xBD\xFD"
-     "\x00\x00"
-     "\xBC\x41\x01\x00\xA8\x31\x02\x00\xA9\x31\x02\x00\xAA\x31\x02\x00\xAB\x31"
-     "\x02\x00"
-     "\xAC\xB5\x02\x00\xAD\xBD\x02\x00\xAE\xB5\x02\x00\xAF\xAD\x02\x00\xFF\x3E"
-     "\x00\x80"
-     "\xBE\xA8\x0F\x00\x83\x40\x00\x80\x87\x40\x00\x80\xBE\x7C\x0F\x00\x8B\x40"
-     "\x00\x80"
-     "\x8F\x40\x00\x80\x93\x40\x00\x80\xB8\x69\x03\x00\xB9\x69\x03\x00\xBA\x09"
-     "\x03\x00"
-     "\xBB\x05\x03\x00\xBC\x1D\x03\x00\xBD\x01\x03\x00\xBE\x01\x03\x00\xBF\x35"
-     "\x03\x00"
-     "\xB0\xD5\x02\x00\xB1\xDD\x02\x00\xB2\xD5\x02\x00\xB3\x69\x03\x00\xB4\x79"
-     "\x03\x00"
-     "\xB5\x79\x03\x00\xB6\x69\x03\x00\xB7\x61\x03\x00\xA3\xB9\x02\x00\x97\x40"
-     "\x00\x80"
-     "\x9B\x40\x00\x80\x9F\x40\x00\x80\xA3\x40\x00\x80\xA6\x3D\x02\x00\xA5\x3D"
-     "\x02\x00"
-     "\xA7\x40\x00\x80\xAB\x19\x02\x00\xAA\x11\x02\x00\xAB\x40\x00\x80\xAF\x40"
-     "\x00\x80"
-     "\xAF\xBD\x03\x00\xAE\xBD\x03\x00\xAD\xBD\x03\x00\xAC\x01\x02\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x1D\x00\x00\xB3\x40\x00\x80\xB7\x40\x00\x80\xBB\x40"
-     "\x00\x80"
-     "\x87\x18\x03\x00\x86\xFC\x04\x00\xC3\x40\x00\x80\xC7\x40\x00\x80\xCB\x40"
-     "\x00\x80"
-     "\xCF\x40\x00\x80\xD3\x40\x00\x80\xD7\x40\x00\x80\xDB\x40\x00\x80\xDF\x40"
-     "\x00\x80"
-     "\xA8\x95\x02\x00\xA9\x9D\x02\x00\xAA\xA1\x02\x00\xAB\xB5\x02\x00\xAC\xAD"
-     "\x02\x00"
-     "\xAD\xD1\x02\x00\xAE\xD1\x02\x00\xAF\xD1\x02\x00\xE3\x40\x00\x80\xE7\x40"
-     "\x00\x80"
-     "\xEB\x40\x00\x80\xEF\x40\x00\x80\xF3\x40\x00\x80\xF7\x40\x00\x80\xFB\x40"
-     "\x00\x80"
-     "\xFF\x40\x00\x80\xB8\x7D\x01\x00\xB9\xC1\x01\x00\xBA\xC1\x01\x00\xBB\xC1"
-     "\x01\x00"
-     "\xBC\xC1\x01\x00\xBD\xC9\x01\x00\xBE\xF1\x01\x00\xBF\xF1\x01\x00\xB0\xB1"
-     "\x02\x00"
-     "\xB1\xB9\x02\x00\xB2\x81\x02\x00\xB3\x81\x02\x00\xB4\x59\x01\x00\xB5\x59"
-     "\x01\x00"
-     "\xB6\x4D\x01\x00\xB7\x45\x01\x00\xBE\x48\x70\x00\x03\x41\x00\x80\x07\x41"
-     "\x00\x80"
-     "\x0B\x41\x00\x80\x0F\x41\x00\x80\x13\x41\x00\x80\xEF\x7C\x16\x00\xEF\x70"
-     "\x16\x00"
-     "\xE1\x90\x01\x00\xE1\xFC\x1E\x00\xE3\xD8\x16\x00\xE3\xE4\x16\x00\x17\x41"
-     "\x00\x80"
-     "\x1B\x41\x00\x80\x1F\x41\x00\x80\x23\x41\x00\x80\x80\x7D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x15\x00\x00\xB3\x65\x02\x00\x27\x41\x00\x80\xB5\x69\x02\x00\xB6\x1D"
-     "\x02\x00"
-     "\x2B\x41\x00\x80\x86\x60\x07\x00\x87\xA4\x04\x00\xBA\x39\x02\x00\xBB\x31"
-     "\x02\x00"
-     "\xBC\x19\x02\x00\xBD\x19\x02\x00\xBE\xD1\x01\x00\xBF\xD1\x01\x00\xA3\xA9"
-     "\x05\x00"
-     "\xBF\x40\x00\x80\x2F\x41\x00\x80\x33\x41\x00\x80\x37\x41\x00\x80\xA6\xD1"
-     "\x05\x00"
-     "\xA5\xA5\x05\x00\x3B\x41\x00\x80\xAB\xFD\x05\x00\xAA\xF5\x05\x00\x3F\x41"
-     "\x00\x80"
-     "\x43\x41\x00\x80\xAF\x1D\x06\x00\xAE\x1D\x06\x00\xAD\xD5\x05\x00\xAC\xD5"
-     "\x05\x00"
-     "\x47\x41\x00\x80\xB3\xAD\x06\x00\x4B\x41\x00\x80\x4F\x41\x00\x80\xB6\xA5"
-     "\x06\x00"
-     "\x53\x41\x00\x80\x57\x41\x00\x80\xB5\xAD\x06\x00\xBA\x41\x07\x00\xBB\x41"
-     "\x07\x00"
-     "\x5B\x41\x00\x80\x5F\x41\x00\x80\xBE\x41\x07\x00\xBF\x41\x07\x00\xBC\x51"
-     "\x07\x00"
-     "\xBD\x51\x07\x00\xA8\x61\x06\x00\xA9\x61\x06\x00\xAA\x61\x06\x00\xAB\x61"
-     "\x06\x00"
-     "\xAC\x61\x06\x00\xAD\x61\x06\x00\xAE\x61\x06\x00\xAF\x61\x06\x00\x63\x41"
-     "\x00\x80"
-     "\x67\x41\x00\x80\x6B\x41\x00\x80\x6F\x41\x00\x80\x73\x41\x00\x80\x77\x41"
-     "\x00\x80"
-     "\x7B\x41\x00\x80\x7F\x41\x00\x80\xB8\x45\x07\x00\xB9\x4D\x07\x00\xBA\x45"
-     "\x07\x00"
-     "\xBB\x5D\x07\x00\xBC\x45\x07\x00\xBD\x65\x07\x00\xBE\x6D\x07\x00\xBF\x65"
-     "\x07\x00"
-     "\xB0\xC1\x07\x00\xB1\xC1\x07\x00\xB2\xC1\x07\x00\xB3\xC1\x07\x00\xB4\xC1"
-     "\x07\x00"
-     "\xB5\xC1\x07\x00\xB6\xC1\x07\x00\xB7\xC1\x07\x00\xA3\xE9\x07\x00\x83\x41"
-     "\x00\x80"
-     "\x87\x41\x00\x80\x8B\x41\x00\x80\x8F\x41\x00\x80\xA6\xE1\x07\x00\xA5\xE9"
-     "\x07\x00"
-     "\x93\x41\x00\x80\xAB\x05\x06\x00\xAA\x05\x06\x00\x97\x41\x00\x80\x9B\x41"
-     "\x00\x80"
-     "\xAF\x05\x06\x00\xAE\x05\x06\x00\xAD\x15\x06\x00\xAC\x15\x06\x00\x80\x5D"
-     "\x00\x00"
-     "\x81\x6D\x00\x00\x82\x65\x00\x00\xB3\x51\x01\x00\x9F\x41\x00\x80\xB5\x79"
-     "\x01\x00"
-     "\xB6\x19\x01\x00\xA3\x41\x00\x80\x86\xC0\x00\x00\x87\x7C\x01\x00\xBA\x3D"
-     "\x01\x00"
-     "\xBB\x35\x01\x00\xBC\x19\x01\x00\xBD\x19\x01\x00\xBE\xD9\x01\x00\xBF\xD9"
-     "\x01\x00"
-     "\xA7\x41\x00\x80\xAB\x41\x00\x80\xAF\x41\x00\x80\xB3\x41\x00\x80\xE1\x38"
-     "\x06\x00"
-     "\xB7\x41\x00\x80\xE3\xE4\x10\x00\xBB\x41\x00\x80\xBE\xE8\x75\x00\xBF\x41"
-     "\x00\x80"
-     "\xEF\x60\x10\x00\xC3\x41\x00\x80\xC7\x41\x00\x80\xCB\x41\x00\x80\xCF\x41"
-     "\x00\x80"
-     "\xD3\x41\x00\x80\xA3\xDD\x01\x00\xD7\x41\x00\x80\xDB\x41\x00\x80\xDF\x41"
-     "\x00\x80"
-     "\xE3\x41\x00\x80\xA6\x95\x01\x00\xA5\xF5\x01\x00\xE7\x41\x00\x80\xAB\xB9"
-     "\x01\x00"
-     "\xAA\xB1\x01\x00\xEB\x41\x00\x80\xEF\x41\x00\x80\xAF\x55\x01\x00\xAE\x55"
-     "\x01\x00"
-     "\xAD\x95\x01\x00\xAC\x95\x01\x00\xF3\x41\x00\x80\xF7\x41\x00\x80\xFB\x41"
-     "\x00\x80"
-     "\xFF\x41\x00\x80\x82\x81\x00\x00\x4F\x40\x00\x80\x80\x99\x00\x00\x81\x99"
-     "\x00\x00"
-     "\xA8\x2D\x02\x00\xA9\x35\x02\x00\xAA\x35\x02\x00\xAB\x05\x02\x00\xAC\x1D"
-     "\x02\x00"
-     "\xAD\x75\x02\x00\xAE\x71\x02\x00\xAF\x71\x02\x00\xB0\x15\x02\x00\xB1\x1D"
-     "\x02\x00"
-     "\xB2\x21\x02\x00\xB3\x21\x02\x00\xB4\xF1\x02\x00\xB5\xF1\x02\x00\xB6\xF1"
-     "\x02\x00"
-     "\xB7\xF1\x02\x00\xB8\xD1\x02\x00\xB9\xD1\x02\x00\xBA\xD1\x02\x00\xBB\xD1"
-     "\x02\x00"
-     "\xBC\x75\x03\x00\xBD\x7D\x03\x00\xBE\x75\x03\x00\xBF\x6D\x03\x00\x03\x42"
-     "\x00\x80"
-     "\x07\x42\x00\x80\x84\x84\x09\x00\xA3\x09\x03\x00\x0B\x42\x00\x80\xA5\x19"
-     "\x03\x00"
-     "\xA6\x0D\x03\x00\x13\x42\x00\x80\x17\x42\x00\x80\x1B\x42\x00\x80\xAA\x05"
-     "\x03\x00"
-     "\xAB\x1D\x03\x00\xAC\x05\x03\x00\xAD\x0D\x03\x00\xAE\x05\x03\x00\xAF\x79"
-     "\x03\x00"
-     "\x86\x60\x0C\x00\x87\x6C\x03\x00\x1F\x42\x00\x80\x23\x42\x00\x80\x27\x42"
-     "\x00\x80"
-     "\xE1\xAC\x01\x00\x2B\x42\x00\x80\xE3\x7C\x13\x00\x2F\x42\x00\x80\x33\x42"
-     "\x00\x80"
-     "\x37\x42\x00\x80\xEF\xD4\x13\x00\x3B\x42\x00\x80\x3F\x42\x00\x80\x43\x42"
-     "\x00\x80"
-     "\x47\x42\x00\x80\xA8\x5D\x02\x00\xA9\x85\x02\x00\xAA\x8D\x02\x00\xAB\x85"
-     "\x02\x00"
-     "\xAC\x9D\x02\x00\xAD\x81\x02\x00\xAE\x85\x02\x00\xAF\xB9\x02\x00\x4B\x42"
-     "\x00\x80"
-     "\x4F\x42\x00\x80\x53\x42\x00\x80\x57\x42\x00\x80\x5B\x42\x00\x80\x5F\x42"
-     "\x00\x80"
-     "\x63\x42\x00\x80\x67\x42\x00\x80\xB8\x91\x02\x00\xB9\x99\x02\x00\xBA\xA1"
-     "\x02\x00"
-     "\xBB\xA1\x02\x00\xBC\x51\x01\x00\xBD\x51\x01\x00\xBE\x51\x01\x00\xBF\x51"
-     "\x01\x00"
-     "\xB0\xCD\x02\x00\xB1\xD5\x02\x00\xB2\xDD\x02\x00\xB3\xD5\x02\x00\xB4\xCD"
-     "\x02\x00"
-     "\xB5\xB1\x02\x00\xB6\xB1\x02\x00\xB7\xB1\x02\x00\xE1\x44\x07\x00\xE3\xEC"
-     "\x12\x00"
-     "\xE3\xEC\x12\x00\xE1\x74\x06\x00\x6B\x42\x00\x80\x6F\x42\x00\x80\x73\x42"
-     "\x00\x80"
-     "\x77\x42\x00\x80\x82\x05\x00\x00\x7B\x42\x00\x80\x80\x31\x00\x00\x81\x31"
-     "\x00\x00"
-     "\x83\x42\x00\x80\xEF\xF4\x12\x00\xEF\x00\x12\x00\x87\x42\x00\x80\x86\x60"
-     "\x0C\x00"
-     "\x87\xE4\x0C\x00\x8B\x42\x00\x80\xB3\x49\x02\x00\x8F\x42\x00\x80\xB5\x59"
-     "\x02\x00"
-     "\xB6\x4D\x02\x00\x93\x42\x00\x80\x97\x42\x00\x80\x9B\x42\x00\x80\xBA\x45"
-     "\x02\x00"
-     "\xBB\x5D\x02\x00\xBC\x45\x02\x00\xBD\x4D\x02\x00\xBE\x45\x02\x00\xBF\x39"
-     "\x02\x00"
-     "\xA8\x9D\x0E\x00\xA9\xA5\x0E\x00\xAA\xAD\x0E\x00\xAB\xC5\x0E\x00\xAC\xDD"
-     "\x0E\x00"
-     "\xAD\xC5\x0E\x00\xAE\xC1\x0E\x00\xAF\xFD\x0E\x00\x7F\x42\x00\x80\x9F\x42"
-     "\x00\x80"
-     "\xA3\x42\x00\x80\xA7\x42\x00\x80\xAB\x42\x00\x80\xAF\x42\x00\x80\xB3\x42"
-     "\x00\x80"
-     "\xB7\x42\x00\x80\xB8\x69\x0F\x00\xB9\x69\x0F\x00\xBA\x79\x0F\x00\xBB\x79"
-     "\x0F\x00"
-     "\xBC\x69\x0F\x00\xBD\x69\x0F\x00\xBE\xC9\x0F\x00\xBF\xC9\x0F\x00\xB0\x85"
-     "\x0E\x00"
-     "\xB1\x8D\x0E\x00\xB2\x99\x0E\x00\xB3\x99\x0E\x00\xB4\x89\x0E\x00\xB5\x89"
-     "\x0E\x00"
-     "\xB6\x59\x0F\x00\xB7\x59\x0F\x00\xB3\x25\x0E\x00\xBB\x42\x00\x80\xBF\x42"
-     "\x00\x80"
-     "\xC3\x42\x00\x80\xC7\x42\x00\x80\xB6\x45\x0E\x00\xB5\x55\x0E\x00\xCB\x42"
-     "\x00\x80"
-     "\xBB\x49\x0E\x00\xBA\x41\x0E\x00\xCF\x42\x00\x80\xD3\x42\x00\x80\xBF\x4D"
-     "\x0E\x00"
-     "\xBE\x4D\x0E\x00\xBD\x4D\x0E\x00\xBC\x4D\x0E\x00\xD7\x42\x00\x80\xA3\x61"
-     "\x0E\x00"
-     "\xDB\x42\x00\x80\xDF\x42\x00\x80\xA6\x01\x0E\x00\xE3\x42\x00\x80\xE7\x42"
-     "\x00\x80"
-     "\xA5\x11\x0E\x00\xAA\x05\x0E\x00\xAB\x0D\x0E\x00\xEB\x42\x00\x80\xEF\x42"
-     "\x00\x80"
-     "\xAE\x09\x0E\x00\xAF\x09\x0E\x00\xAC\x09\x0E\x00\xAD\x09\x0E\x00\xA8\xB1"
-     "\x0E\x00"
-     "\xA9\xB9\x0E\x00\xAA\xC1\x0E\x00\xAB\xDD\x0E\x00\xAC\xCD\x0E\x00\xAD\xF5"
-     "\x0E\x00"
-     "\xAE\xFD\x0E\x00\xAF\x39\x01\x00\x80\xF5\x01\x00\x81\xFD\x01\x00\x82\xC1"
-     "\x01\x00"
-     "\xF3\x42\x00\x80\x86\x30\x01\x00\x87\x74\x01\x00\xF7\x42\x00\x80\xFB\x42"
-     "\x00\x80"
-     "\xB8\xCD\x01\x00\xB9\xD5\x01\x00\xBA\xD5\x01\x00\xBB\xE5\x01\x00\xBC\xFD"
-     "\x01\x00"
-     "\xBD\x95\x01\x00\xBE\x91\x01\x00\xBF\x91\x01\x00\xB0\x49\x01\x00\xB1\x49"
-     "\x01\x00"
-     "\xB2\x59\x01\x00\xB3\x59\x01\x00\xB4\x49\x01\x00\xB5\x49\x01\x00\xB6\xFD"
-     "\x01\x00"
-     "\xB7\xF5\x01\x00\xE1\x58\x0D\x00\xFF\x42\x00\x80\xE3\x74\x1C\x00\x03\x43"
-     "\x00\x80"
-     "\x07\x43\x00\x80\x0B\x43\x00\x80\x0F\x43\x00\x80\x13\x43\x00\x80\x17\x43"
-     "\x00\x80"
-     "\x1B\x43\x00\x80\x1F\x43\x00\x80\x23\x43\x00\x80\x27\x43\x00\x80\x2B\x43"
-     "\x00\x80"
-     "\xEF\x90\x1C\x00\x2F\x43\x00\x80\xB3\x09\x0E\x00\x33\x43\x00\x80\x37\x43"
-     "\x00\x80"
-     "\x3B\x43\x00\x80\x3F\x43\x00\x80\xB6\x35\x0E\x00\xB5\x05\x0E\x00\x43\x43"
-     "\x00\x80"
-     "\xBB\xCD\x01\x00\xBA\xCD\x01\x00\x47\x43\x00\x80\x4B\x43\x00\x80\xBF\xCD"
-     "\x01\x00"
-     "\xBE\xCD\x01\x00\xBD\xCD\x01\x00\xBC\xCD\x01\x00\x4F\x43\x00\x80\x0F\x42"
-     "\x00\x80"
-     "\x53\x43\x00\x80\xA3\xC5\x0D\x00\x57\x43\x00\x80\xA5\xC9\x0D\x00\xA6\xF9"
-     "\x0D\x00"
-     "\x5B\x43\x00\x80\x63\x43\x00\x80\x67\x43\x00\x80\xAA\x01\x02\x00\xAB\x01"
-     "\x02\x00"
-     "\xAC\x01\x02\x00\xAD\x01\x02\x00\xAE\x01\x02\x00\xAF\x01\x02\x00\x80\x49"
-     "\x00\x00"
-     "\x81\x55\x00\x00\x82\x55\x00\x00\x84\xCC\x06\x00\x86\xF0\x04\x00\x87\x14"
-     "\x03\x00"
-     "\x6B\x43\x00\x80\x6F\x43\x00\x80\x73\x43\x00\x80\x77\x43\x00\x80\x7B\x43"
-     "\x00\x80"
-     "\x7F\x43\x00\x80\x83\x43\x00\x80\x87\x43\x00\x80\x8B\x43\x00\x80\x8F\x43"
-     "\x00\x80"
-     "\x93\x43\x00\x80\x97\x43\x00\x80\x9B\x43\x00\x80\x9F\x43\x00\x80\xA3\x43"
-     "\x00\x80"
-     "\xA7\x43\x00\x80\xAB\x43\x00\x80\xAF\x43\x00\x80\xA8\x8D\x03\x00\xA9\x9D"
-     "\x03\x00"
-     "\xAA\x95\x03\x00\xAB\x29\x03\x00\xAC\x3D\x03\x00\xAD\x25\x03\x00\xAE\x2D"
-     "\x03\x00"
-     "\xAF\x25\x03\x00\xB0\x5D\x03\x00\xB1\x65\x03\x00\xB2\x6D\x03\x00\xB3\x65"
-     "\x03\x00"
-     "\xB4\x7D\x03\x00\xB5\x65\x03\x00\xB6\x6D\x03\x00\xB7\x61\x03\x00\xB8\x25"
-     "\x03\x00"
-     "\xB9\x2D\x03\x00\xBA\x25\x03\x00\xBB\x3D\x03\x00\xBC\x2D\x03\x00\xBD\x1D"
-     "\x03\x00"
-     "\xBE\x15\x03\x00\xBF\xFD\x00\x00\xB3\x43\x00\x80\xB7\x43\x00\x80\xBB\x43"
-     "\x00\x80"
-     "\xB3\x31\x02\x00\xBF\x43\x00\x80\xB5\x99\x02\x00\xB6\x89\x02\x00\xC3\x43"
-     "\x00\x80"
-     "\x84\xA4\x06\x00\xC7\x43\x00\x80\xBA\xB5\x02\x00\xBB\x8D\x02\x00\xBC\x95"
-     "\x02\x00"
-     "\xBD\x9D\x02\x00\xBE\x95\x02\x00\xBF\x8D\x02\x00\x80\x5D\x00\x00\x81\x61"
-     "\x00\x00"
-     "\x82\x05\x00\x00\xCB\x43\x00\x80\xEF\x24\x17\x00\xD3\x43\x00\x80\x87\x74"
-     "\x05\x00"
-     "\x86\xFC\x04\x00\xE3\xD0\x16\x00\xD7\x43\x00\x80\xE1\x40\x00\x00\xDB\x43"
-     "\x00\x80"
-     "\xDF\x43\x00\x80\xE3\x43\x00\x80\xE7\x43\x00\x80\xEB\x43\x00\x80\xEF\x43"
-     "\x00\x80"
-     "\xF3\x43\x00\x80\xF7\x43\x00\x80\xFB\x43\x00\x80\xFF\x43\x00\x80\x03\x44"
-     "\x00\x80"
-     "\x07\x44\x00\x80\xEF\xA4\x12\x00\xBE\x00\x7C\x00\xE1\x7C\x0E\x00\x0B\x44"
-     "\x00\x80"
-     "\xE3\xF4\x13\x00\x0F\x44\x00\x80\x13\x44\x00\x80\x17\x44\x00\x80\x1B\x44"
-     "\x00\x80"
-     "\xA3\xFD\x02\x00\x1F\x44\x00\x80\x23\x44\x00\x80\x27\x44\x00\x80\x2B\x44"
-     "\x00\x80"
-     "\xA6\x45\x02\x00\xA5\x55\x02\x00\x2F\x44\x00\x80\xAB\x41\x02\x00\xAA\x79"
-     "\x02\x00"
-     "\x33\x44\x00\x80\x37\x44\x00\x80\xAF\x41\x02\x00\xAE\x59\x02\x00\xAD\x51"
-     "\x02\x00"
-     "\xAC\x59\x02\x00\xA8\x21\x06\x00\xA9\x21\x06\x00\xAA\x51\x06\x00\xAB\x51"
-     "\x06\x00"
-     "\xAC\x71\x06\x00\xAD\x71\x06\x00\xAE\x71\x06\x00\xAF\x71\x06\x00\xCF\x43"
-     "\x00\x80"
-     "\x82\x15\x00\x00\x81\xC5\x07\x00\x80\xC5\x07\x00\x3B\x44\x00\x80\x3F\x44"
-     "\x00\x80"
-     "\x86\x1C\x00\x00\x87\xA0\x03\x00\xB8\xD1\x06\x00\xB9\xD1\x06\x00\xBA\xD1"
-     "\x06\x00"
-     "\xBB\xD1\x06\x00\xBC\x51\x07\x00\xBD\x51\x07\x00\xBE\x51\x07\x00\xBF\x4D"
-     "\x07\x00"
-     "\xB0\x11\x06\x00\xB1\x11\x06\x00\xB2\x11\x06\x00\xB3\x11\x06\x00\xB4\xF1"
-     "\x06\x00"
-     "\xB5\xF1\x06\x00\xB6\xF1\x06\x00\xB7\xF1\x06\x00\xB3\xD9\x06\x00\x43\x44"
-     "\x00\x80"
-     "\x47\x44\x00\x80\x84\x40\x7F\x00\x4B\x44\x00\x80\xB6\xF1\x06\x00\xB5\xF9"
-     "\x06\x00"
-     "\x4F\x44\x00\x80\xBB\xD5\x06\x00\xBA\xD5\x06\x00\x53\x44\x00\x80\x57\x44"
-     "\x00\x80"
-     "\xBF\x35\x06\x00\xBE\x35\x06\x00\xBD\xC5\x06\x00\xBC\xC5\x06\x00\x5B\x44"
-     "\x00\x80"
-     "\xA3\x9D\x06\x00\x5F\x44\x00\x80\x63\x44\x00\x80\xA6\xB5\x06\x00\x67\x44"
-     "\x00\x80"
-     "\x6B\x44\x00\x80\xA5\xBD\x06\x00\xAA\x91\x06\x00\xAB\x91\x06\x00\x6F\x44"
-     "\x00\x80"
-     "\x73\x44\x00\x80\xAE\x71\x06\x00\xAF\x71\x06\x00\xAC\x81\x06\x00\xAD\x81"
-     "\x06\x00"
-     "\xA8\x55\x06\x00\xA9\x61\x06\x00\xAA\x65\x06\x00\xAB\x7D\x06\x00\xAC\x65"
-     "\x06\x00"
-     "\xAD\x6D\x06\x00\xAE\x65\x06\x00\xAF\x9D\x06\x00\x77\x44\x00\x80\x7B\x44"
-     "\x00\x80"
-     "\x7F\x44\x00\x80\x83\x44\x00\x80\x87\x44\x00\x80\x8B\x44\x00\x80\x8F\x44"
-     "\x00\x80"
-     "\x93\x44\x00\x80\xB8\xF5\x06\x00\xB9\xFD\x06\x00\xBA\xF5\x06\x00\xBB\x89"
-     "\x06\x00"
-     "\xBC\x99\x06\x00\xBD\x99\x06\x00\xBE\x89\x06\x00\xBF\x81\x06\x00\xB0\xE5"
-     "\x06\x00"
-     "\xB1\xED\x06\x00\xB2\xE5\x06\x00\xB3\xFD\x06\x00\xB4\xE5\x06\x00\xB5\xED"
-     "\x06\x00"
-     "\xB6\xD1\x06\x00\xB7\xCD\x06\x00\x97\x44\x00\x80\xE3\x44\x10\x00\x9B\x44"
-     "\x00\x80"
-     "\xE1\xC4\x05\x00\x9F\x44\x00\x80\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D"
-     "\x00\x00"
-     "\xA3\x44\x00\x80\xA7\x44\x00\x80\xAB\x44\x00\x80\xB3\x44\x00\x80\xB7\x44"
-     "\x00\x80"
-     "\xEF\xEC\x0F\x00\x87\x04\x01\x00\x86\x1C\x7F\x00\xB3\xAD\x06\x00\xBB\x44"
-     "\x00\x80"
-     "\xBF\x44\x00\x80\xC3\x44\x00\x80\xC7\x44\x00\x80\xB6\xAD\x06\x00\xB5\xAD"
-     "\x06\x00"
-     "\xCB\x44\x00\x80\xBB\x69\x06\x00\xBA\x61\x06\x00\xCF\x44\x00\x80\xD3\x44"
-     "\x00\x80"
-     "\xBF\x69\x06\x00\xBE\x61\x06\x00\xBD\x69\x06\x00\xBC\x71\x06\x00\xA3\x6D"
-     "\x05\x00"
-     "\xD7\x44\x00\x80\x5F\x43\x00\x80\xDB\x44\x00\x80\xDF\x44\x00\x80\xA6\x6D"
-     "\x05\x00"
-     "\xA5\x6D\x05\x00\xE3\x44\x00\x80\xAB\xA9\x05\x00\xAA\xA1\x05\x00\xE7\x44"
-     "\x00\x80"
-     "\x84\x60\x7E\x00\xAF\xA9\x05\x00\xAE\xA1\x05\x00\xAD\xA9\x05\x00\xAC\xB1"
-     "\x05\x00"
-     "\xEB\x44\x00\x80\xEF\x44\x00\x80\x84\x48\x7F\x00\xB3\xDD\x02\x00\xF3\x44"
-     "\x00\x80"
-     "\xB5\xFD\x02\x00\xB6\xF5\x02\x00\xF7\x44\x00\x80\xFB\x44\x00\x80\xFF\x44"
-     "\x00\x80"
-     "\xBA\xA5\x02\x00\xBB\xB9\x02\x00\xBC\xA9\x02\x00\xBD\x9D\x02\x00\xBE\x95"
-     "\x02\x00"
-     "\xBF\x5D\x03\x00\xA9\x09\x02\x00\xA8\x01\x02\x00\xAB\x15\x02\x00\xAA\x1D"
-     "\x02\x00"
-     "\xAD\x3D\x02\x00\xAC\x3D\x02\x00\xAF\x51\x02\x00\xAE\x5D\x02\x00\x85\xE4"
-     "\x78\x00"
-     "\x03\x45\x00\x80\x07\x45\x00\x80\x0B\x45\x00\x80\x0F\x45\x00\x80\x82\xA5"
-     "\x01\x00"
-     "\x81\xB1\x01\x00\x80\xB9\x01\x00\xB9\xF1\x03\x00\xB8\xE9\x03\x00\xBB\xF1"
-     "\x03\x00"
-     "\xBA\xF9\x03\x00\xBD\x51\x03\x00\xBC\xE9\x03\x00\xBF\x49\x03\x00\xBE\x55"
-     "\x03\x00"
-     "\xB1\x35\x02\x00\xB0\x29\x02\x00\xB3\x01\x02\x00\xB2\x35\x02\x00\xB5\xD1"
-     "\x03\x00"
-     "\xB4\x19\x02\x00\xB7\xD1\x03\x00\xB6\xD9\x03\x00\xA3\x91\x03\x00\x13\x45"
-     "\x00\x80"
-     "\x17\x45\x00\x80\x84\x80\x78\x00\x1B\x45\x00\x80\xA6\xB9\x03\x00\xA5\xB1"
-     "\x03\x00"
-     "\x23\x45\x00\x80\xAB\xF5\x03\x00\xAA\xE9\x03\x00\x27\x45\x00\x80\x2B\x45"
-     "\x00\x80"
-     "\xAF\x11\x02\x00\xAE\xD9\x03\x00\xAD\xD1\x03\x00\xAC\xE5\x03\x00\x86\x60"
-     "\x7D\x00"
-     "\x87\x68\x7D\x00\xB3\x8D\x02\x00\x2F\x45\x00\x80\xB5\x9D\x02\x00\x33\x45"
-     "\x00\x80"
-     "\x37\x45\x00\x80\xB6\x8D\x02\x00\x3B\x45\x00\x80\x3F\x45\x00\x80\xBB\x6D"
-     "\x01\x00"
-     "\xBA\x81\x02\x00\xBD\x61\x01\x00\xBC\x79\x01\x00\xBF\x51\x01\x00\xBE\x61"
-     "\x01\x00"
-     "\x43\x45\x00\x80\x47\x45\x00\x80\xEF\x54\x0F\x00\x4B\x45\x00\x80\x4F\x45"
-     "\x00\x80"
-     "\x53\x45\x00\x80\x57\x45\x00\x80\xEF\xC4\x0C\x00\x85\x04\x62\x00\xE1\xE0"
-     "\x06\x00"
-     "\x5B\x45\x00\x80\xE3\x34\x0B\x00\xE1\x90\x01\x00\x5F\x45\x00\x80\xE3\xE4"
-     "\x07\x00"
-     "\x63\x45\x00\x80\xA3\x05\x02\x00\x67\x45\x00\x80\x6B\x45\x00\x80\x6F\x45"
-     "\x00\x80"
-     "\x73\x45\x00\x80\xA6\x05\x02\x00\xA5\x15\x02\x00\x77\x45\x00\x80\xAB\xE5"
-     "\x01\x00"
-     "\xAA\x09\x02\x00\x7B\x45\x00\x80\x7F\x45\x00\x80\xAF\xD9\x01\x00\xAE\xE9"
-     "\x01\x00"
-     "\xAD\xE9\x01\x00\xAC\xF1\x01\x00\xA8\xAD\x7D\x00\xA9\xB5\x7D\x00\xAA\xCD"
-     "\x7D\x00"
-     "\xAB\xC5\x7D\x00\xAC\xDD\x7D\x00\xAD\xC5\x7D\x00\xAE\xC5\x7D\x00\xAF\xF5"
-     "\x7D\x00"
-     "\x1F\x45\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\x83\x45"
-     "\x00\x80"
-     "\x87\x45\x00\x80\x8B\x45\x00\x80\x8F\x45\x00\x80\xB8\x11\x7E\x00\xB9\x19"
-     "\x7E\x00"
-     "\xBA\x29\x7E\x00\xBB\x29\x7E\x00\xBC\xF5\x7E\x00\xBD\xFD\x7E\x00\xBE\xF5"
-     "\x7E\x00"
-     "\xBF\xED\x7E\x00\xB0\x8D\x7D\x00\xB1\x61\x7E\x00\xB2\x65\x7E\x00\xB3\x7D"
-     "\x7E\x00"
-     "\xB4\x6D\x7E\x00\xB5\x5D\x7E\x00\xB6\x39\x7E\x00\xB7\x39\x7E\x00\xA8\x9D"
-     "\x7F\x00"
-     "\xA9\xAD\x7F\x00\xAA\xA5\x7F\x00\xAB\xC1\x7F\x00\xAC\xC5\x7F\x00\xAD\xCD"
-     "\x7F\x00"
-     "\xAE\xF9\x7F\x00\xAF\xF9\x7F\x00\x86\x40\x00\x00\x87\x34\x03\x00\x93\x45"
-     "\x00\x80"
-     "\x97\x45\x00\x80\x9B\x45\x00\x80\x9F\x45\x00\x80\xA3\x45\x00\x80\xA7\x45"
-     "\x00\x80"
-     "\xB8\x7D\x7F\x00\xB9\x01\x7F\x00\xBA\x01\x7F\x00\xBB\x01\x7F\x00\xBC\x01"
-     "\x7F\x00"
-     "\xBD\x09\x7F\x00\xBE\x31\x7F\x00\xBF\x31\x7F\x00\xB0\x61\x7F\x00\xB1\x61"
-     "\x7F\x00"
-     "\xB2\x65\x7F\x00\xB3\x7D\x7F\x00\xB4\x65\x7F\x00\xB5\x6D\x7F\x00\xB6\x51"
-     "\x7F\x00"
-     "\xB7\x49\x7F\x00\xB3\x2D\x7E\x00\xAB\x45\x00\x80\xAF\x45\x00\x80\xB3\x45"
-     "\x00\x80"
-     "\xB7\x45\x00\x80\xB6\x4D\x7E\x00\xB5\x4D\x7E\x00\xBB\x45\x00\x80\xBB\xA5"
-     "\x7E\x00"
-     "\xBA\xA5\x7E\x00\xBF\x45\x00\x80\xC3\x45\x00\x80\xBF\xA9\x7E\x00\xBE\xA1"
-     "\x7E\x00"
-     "\xBD\xA9\x7E\x00\xBC\xB1\x7E\x00\xC7\x45\x00\x80\xA3\x69\x7E\x00\xCB\x45"
-     "\x00\x80"
-     "\xCF\x45\x00\x80\xA6\x09\x7E\x00\xD3\x45\x00\x80\xD7\x45\x00\x80\xA5\x09"
-     "\x7E\x00"
-     "\xAA\xE1\x7E\x00\xAB\xE1\x7E\x00\xDB\x45\x00\x80\xDF\x45\x00\x80\xAE\xE5"
-     "\x7E\x00"
-     "\xAF\xED\x7E\x00\xAC\xF5\x7E\x00\xAD\xED\x7E\x00\xE3\x45\x00\x80\xE7\x45"
-     "\x00\x80"
-     "\xEB\x45\x00\x80\xEF\x45\x00\x80\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D"
-     "\x00\x00"
-     "\xF3\x45\x00\x80\xF7\x45\x00\x80\xFB\x45\x00\x80\x86\x00\x00\x00\x87\x64"
-     "\x01\x00"
-     "\x03\x46\x00\x80\x07\x46\x00\x80\x0B\x46\x00\x80\x0F\x46\x00\x80\xEF\x08"
-     "\x71\x00"
-     "\x13\x46\x00\x80\x17\x46\x00\x80\x1B\x46\x00\x80\x1F\x46\x00\x80\x23\x46"
-     "\x00\x80"
-     "\x27\x46\x00\x80\xEF\x84\x78\x00\x2B\x46\x00\x80\xE1\x54\x02\x00\x2F\x46"
-     "\x00\x80"
-     "\xE3\xB4\x7D\x00\xE3\xC0\x78\x00\x33\x46\x00\x80\xE1\x5C\x7E\x00\x37\x46"
-     "\x00\x80"
-     "\xB8\x55\x03\x00\xB9\x5D\x03\x00\xBA\x61\x03\x00\xBB\x61\x03\x00\xBC\x91"
-     "\x03\x00"
-     "\xBD\x91\x03\x00\xBE\x91\x03\x00\xBF\x91\x03\x00\xB0\x2D\x03\x00\xB1\x35"
-     "\x03\x00"
-     "\xB2\x35\x03\x00\xB3\x05\x03\x00\xB4\x1D\x03\x00\xB5\x75\x03\x00\xB6\x71"
-     "\x03\x00"
-     "\xB7\x71\x03\x00\xA8\xE9\x02\x00\xA9\xE9\x02\x00\xAA\xF9\x02\x00\xAB\xF9"
-     "\x02\x00"
-     "\xAC\xE9\x02\x00\xAD\xE9\x02\x00\xAE\x5D\x03\x00\xAF\x55\x03\x00\xFF\x45"
-     "\x00\x80"
-     "\x3B\x46\x00\x80\x3F\x46\x00\x80\x43\x46\x00\x80\x47\x46\x00\x80\x4B\x46"
-     "\x00\x80"
-     "\x4F\x46\x00\x80\xAF\x44\x00\x80\xB3\x31\x02\x00\x53\x46\x00\x80\x57\x46"
-     "\x00\x80"
-     "\x5B\x46\x00\x80\x5F\x46\x00\x80\xB6\x29\x02\x00\xB5\x21\x02\x00\x63\x46"
-     "\x00\x80"
-     "\xBB\xB1\x03\x00\xBA\xA9\x03\x00\x67\x46\x00\x80\x6B\x46\x00\x80\xBF\xF1"
-     "\x03\x00"
-     "\xBE\x99\x03\x00\xBD\x91\x03\x00\xBC\xA1\x03\x00\x82\x6D\x00\x00\xA3\x75"
-     "\x02\x00"
-     "\x80\x55\x00\x00\x81\x65\x00\x00\xA6\x6D\x02\x00\x73\x46\x00\x80\x77\x46"
-     "\x00\x80"
-     "\xA5\x65\x02\x00\xAA\xED\x03\x00\xAB\xF5\x03\x00\x86\x48\x06\x00\x87\xFC"
-     "\x04\x00"
-     "\xAE\xDD\x03\x00\xAF\xB5\x03\x00\xAC\xE5\x03\x00\xAD\xD5\x03\x00\xAA\xB9"
-     "\x02\x00"
-     "\xAB\xB9\x02\x00\xA8\xB9\x02\x00\xA9\xB9\x02\x00\xAE\xDD\x02\x00\xAF\xC5"
-     "\x02\x00"
-     "\xAC\xDD\x02\x00\xAD\xD5\x02\x00\x7B\x46\x00\x80\x7F\x46\x00\x80\x83\x46"
-     "\x00\x80"
-     "\x87\x46\x00\x80\x8B\x46\x00\x80\x8F\x46\x00\x80\x93\x46\x00\x80\x97\x46"
-     "\x00\x80"
-     "\xBA\x35\x01\x00\xBB\x3D\x01\x00\xB8\x25\x01\x00\xB9\x3D\x01\x00\xBE\xDD"
-     "\x01\x00"
-     "\xBF\xC5\x01\x00\xBC\xDD\x01\x00\xBD\xD5\x01\x00\xB2\x81\x02\x00\xB3\x8D"
-     "\x02\x00"
-     "\xB0\xBD\x02\x00\xB1\xB5\x02\x00\xB6\x15\x01\x00\xB7\x1D\x01\x00\xB4\x8D"
-     "\x02\x00"
-     "\xB5\x81\x02\x00\xEF\xD8\x74\x00\xE1\x30\x7F\x00\x9B\x46\x00\x80\xE3\x38"
-     "\x7E\x00"
-     "\xE5\x00\x00\x00\x9F\x46\x00\x80\xA3\x46\x00\x80\xEF\xE4\x72\x00\xA7\x46"
-     "\x00\x80"
-     "\xE1\xFC\x7E\x00\xAB\x46\x00\x80\xE3\x30\x7D\x00\xE3\x8C\x7F\x00\xAF\x46"
-     "\x00\x80"
-     "\xE1\x00\x01\x00\xEF\x6C\x66\x00\xB3\x46\x00\x80\xB7\x46\x00\x80\xBE\xC8"
-     "\x07\x00"
-     "\xB3\x75\x02\x00\xBB\x46\x00\x80\xB5\x11\x02\x00\xB6\x11\x02\x00\xBF\x46"
-     "\x00\x80"
-     "\xC3\x46\x00\x80\xC7\x46\x00\x80\xBA\x3D\x02\x00\xBB\x35\x02\x00\xBC\x11"
-     "\x02\x00"
-     "\xBD\x1D\x02\x00\xBE\x0D\x02\x00\xBF\x01\x02\x00\x8F\x6D\x6C\x00\x6F\x46"
-     "\x00\x80"
-     "\xCB\x46\x00\x80\xCF\x46\x00\x80\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D"
-     "\x00\x00"
-     "\xD3\x46\x00\x80\xD7\x46\x00\x80\xDB\x46\x00\x80\xDF\x46\x00\x80\xE3\x46"
-     "\x00\x80"
-     "\x86\x90\x01\x00\x87\xC0\x03\x00\xE7\x46\x00\x80\xEB\x46\x00\x80\x9F\x75"
-     "\x78\x00"
-     "\x9E\x65\x78\x00\x9D\x41\x7A\x00\x9C\xA1\x7C\x00\x9B\x71\x7C\x00\x9A\x51"
-     "\x7E\x00"
-     "\x99\x09\x71\x00\x98\x55\x70\x00\x97\x65\x70\x00\x96\x71\x72\x00\x95\x4D"
-     "\x74\x00"
-     "\x94\x4D\x74\x00\x93\x51\x76\x00\x92\xE5\x68\x00\x91\x59\x68\x00\x90\x41"
-     "\x6A\x00"
-     "\xB1\x85\x1C\x00\xB0\x85\x1C\x00\xB3\x01\x18\x00\xB2\xAD\x1D\x00\xB5\xA5"
-     "\x19\x00"
-     "\xB4\xB5\x18\x00\xEF\x46\x00\x80\xF3\x46\x00\x80\x84\xB8\x00\x00\xF7\x46"
-     "\x00\x80"
-     "\xFB\x46\x00\x80\xFF\x46\x00\x80\xA3\xD1\x06\x00\xA2\x89\x04\x00\xA1\x99"
-     "\x04\x00"
-     "\xA0\xC1\x06\x00\xA1\xC5\x00\x00\x03\x47\x00\x80\xA3\x01\x0C\x00\xA2\xE5"
-     "\x00\x00"
-     "\xA5\x05\x0C\x00\xA4\xF5\x0C\x00\xA7\xF9\x08\x00\xA6\x01\x08\x00\xA9\x01"
-     "\x14\x00"
-     "\xA8\x49\x08\x00\xAB\xD9\x14\x00\xAA\xCD\x14\x00\xAD\xD1\x10\x00\xAC\x01"
-     "\x10\x00"
-     "\xAF\x91\x11\x00\xAE\xCD\x10\x00\x07\x47\x00\x80\x0B\x47\x00\x80\xAA\xDD"
-     "\x04\x00"
-     "\xAB\xD5\x04\x00\xAC\xF1\x04\x00\xAD\xFD\x04\x00\xAE\xED\x04\x00\xAF\xE1"
-     "\x04\x00"
-     "\x0F\x47\x00\x80\x13\x47\x00\x80\x17\x47\x00\x80\xA3\x95\x04\x00\x1B\x47"
-     "\x00\x80"
-     "\xA5\xF1\x04\x00\xA6\xF1\x04\x00\x1F\x47\x00\x80\xB3\x61\x06\x00\x23\x47"
-     "\x00\x80"
-     "\x27\x47\x00\x80\x2B\x47\x00\x80\x2F\x47\x00\x80\xB6\x79\x06\x00\xB5\x71"
-     "\x06\x00"
-     "\x33\x47\x00\x80\xBB\x5D\x06\x00\xBA\x5D\x06\x00\x37\x47\x00\x80\x3B\x47"
-     "\x00\x80"
-     "\xBF\xD1\x06\x00\xBE\xED\x06\x00\xBD\xED\x06\x00\xBC\xF5\x06\x00\x82\xED"
-     "\x6D\x00"
-     "\x83\xB9\x6B\x00\x3F\x47\x00\x80\x43\x47\x00\x80\x86\xB1\x77\x00\x87\xB5"
-     "\x75\x00"
-     "\x84\xB9\x69\x00\x85\x0D\x69\x00\x8A\x85\x75\x00\x8B\xB1\x73\x00\x47\x47"
-     "\x00\x80"
-     "\x4B\x47\x00\x80\x8E\xE1\x70\x00\x8F\xB1\x7F\x00\x8C\xBD\x71\x00\x8D\xB5"
-     "\x71\x00"
-     "\x92\xE1\x7D\x00\x93\x29\x7D\x00\xBE\x88\x00\x00\x53\x47\x00\x80\x96\xE5"
-     "\x79\x00"
-     "\x97\xA9\x07\x00\x94\xD1\x7B\x00\x95\xFD\x79\x00\x9A\xC1\x05\x00\x9B\xD9"
-     "\x05\x00"
-     "\x86\x48\x0F\x00\x87\xD8\x00\x00\x81\x59\x00\x00\x80\x79\x00\x00\x9C\x99"
-     "\x07\x00"
-     "\x82\x51\x00\x00\xA8\xD5\x04\x00\xA9\xDD\x04\x00\xAA\x29\x05\x00\xAB\x29"
-     "\x05\x00"
-     "\xAC\x39\x05\x00\xAD\x39\x05\x00\xAE\x29\x05\x00\xAF\x29\x05\x00\x57\x47"
-     "\x00\x80"
-     "\x5B\x47\x00\x80\xBE\x60\x0C\x00\x5F\x47\x00\x80\x63\x47\x00\x80\x67\x47"
-     "\x00\x80"
-     "\x6B\x47\x00\x80\xBF\xB4\x3E\x00\xB8\x81\x05\x00\xB9\x81\x05\x00\xBA\x99"
-     "\x05\x00"
-     "\xBB\x91\x05\x00\xBC\xB1\x05\x00\xBD\xB1\x05\x00\xBE\x51\x02\x00\xBF\x51"
-     "\x02\x00"
-     "\xB0\x59\x05\x00\xB1\x59\x05\x00\xB2\xE1\x05\x00\xB3\xF9\x05\x00\xB4\xE9"
-     "\x05\x00"
-     "\xB5\xD5\x05\x00\xB6\xD1\x05\x00\xB7\xC9\x05\x00\x6F\x47\x00\x80\x73\x47"
-     "\x00\x80"
-     "\x77\x47\x00\x80\xA3\xAD\x05\x00\x7B\x47\x00\x80\xA5\xBD\x05\x00\xA6\xB5"
-     "\x05\x00"
-     "\x7F\x47\x00\x80\x83\x47\x00\x80\x87\x47\x00\x80\xAA\x91\x05\x00\xAB\x91"
-     "\x05\x00"
-     "\xAC\x39\x05\x00\xAD\x21\x05\x00\xAE\x21\x05\x00\xAF\x1D\x05\x00\xB6\x65"
-     "\x02\x00"
-     "\x8B\x47\x00\x80\x8F\x47\x00\x80\xB5\x51\x02\x00\x93\x47\x00\x80\xB3\xE9"
-     "\x02\x00"
-     "\x97\x47\x00\x80\x9B\x47\x00\x80\xBE\xC5\x03\x00\xBF\xCD\x03\x00\xBC\xED"
-     "\x03\x00"
-     "\xBD\xCD\x03\x00\xBA\x35\x02\x00\xBB\xF5\x03\x00\x9F\x47\x00\x80\xA3\x47"
-     "\x00\x80"
-     "\xA8\x2D\x03\x00\xA9\x3D\x03\x00\xAA\x35\x03\x00\xAB\x95\x03\x00\xAC\x8D"
-     "\x03\x00"
-     "\xAD\xB1\x03\x00\xAE\xB1\x03\x00\xAF\xB1\x03\x00\xA7\x47\x00\x80\xAB\x47"
-     "\x00\x80"
-     "\xAF\x47\x00\x80\xB3\x47\x00\x80\x81\x09\x00\x00\x80\x19\x00\x00\xB7\x47"
-     "\x00\x80"
-     "\x82\xB9\x00\x00\xB8\x05\x02\x00\xB9\x0D\x02\x00\xBA\x05\x02\x00\xBB\x1D"
-     "\x02\x00"
-     "\xBC\x0D\x02\x00\xBD\x3D\x02\x00\xBE\x35\x02\x00\xBF\xD5\x02\x00\xB0\xA5"
-     "\x03\x00"
-     "\xB1\x65\x02\x00\xB2\x6D\x02\x00\xB3\x4D\x02\x00\xB4\x55\x02\x00\xB5\x5D"
-     "\x02\x00"
-     "\xB6\x49\x02\x00\xB7\x41\x02\x00\xE1\x70\x07\x00\xBB\x47\x00\x80\xE3\x94"
-     "\x7A\x00"
-     "\x84\xC0\x0D\x00\xC3\x47\x00\x80\xEF\xA4\x00\x00\x86\x38\x0C\x00\x87\xBC"
-     "\x0C\x00"
-     "\x84\xEC\x0C\x00\xE3\x14\x01\x00\xC7\x47\x00\x80\xE1\x84\x01\x00\xCB\x47"
-     "\x00\x80"
-     "\xCF\x47\x00\x80\xEF\x88\x06\x00\xD3\x47\x00\x80\xD7\x47\x00\x80\xDB\x47"
-     "\x00\x80"
-     "\xDF\x47\x00\x80\xA3\xF1\x03\x00\xE3\x47\x00\x80\xA5\x49\x03\x00\xA6\x7D"
-     "\x03\x00"
-     "\xE7\x47\x00\x80\xEB\x47\x00\x80\xEF\x47\x00\x80\xAA\x2D\x03\x00\xAB\xED"
-     "\x02\x00"
-     "\xAC\xF5\x02\x00\xAD\xD5\x02\x00\xAE\xDD\x02\x00\xAF\xD5\x02\x00\xA8\xCD"
-     "\x0D\x00"
-     "\xA9\x25\x0E\x00\xAA\x2D\x0E\x00\xAB\x21\x0E\x00\xAC\x21\x0E\x00\xAD\x21"
-     "\x0E\x00"
-     "\xAE\x21\x0E\x00\xAF\x9D\x0E\x00\xBF\x47\x00\x80\xF3\x47\x00\x80\xF7\x47"
-     "\x00\x80"
-     "\xFB\x47\x00\x80\xFF\x47\x00\x80\x03\x48\x00\x80\x07\x48\x00\x80\x0B\x48"
-     "\x00\x80"
-     "\xB8\x91\x0E\x00\xB9\x9D\x0E\x00\xBA\x95\x0E\x00\xBB\xB5\x0E\x00\xBC\xAD"
-     "\x0E\x00"
-     "\xBD\x75\x0F\x00\xBE\x7D\x0F\x00\xBF\x75\x0F\x00\xB0\xE5\x0E\x00\xB1\xED"
-     "\x0E\x00"
-     "\xB2\xF9\x0E\x00\xB3\xF1\x0E\x00\xB4\xD1\x0E\x00\xB5\xD1\x0E\x00\xB6\xBD"
-     "\x0E\x00"
-     "\xB7\xB5\x0E\x00\xB3\xA5\x0E\x00\x0F\x48\x00\x80\x13\x48\x00\x80\x17\x48"
-     "\x00\x80"
-     "\x1B\x48\x00\x80\xB6\xA1\x0E\x00\xB5\xA9\x0E\x00\x1F\x48\x00\x80\xBB\x29"
-     "\x0E\x00"
-     "\xBA\x21\x0E\x00\x23\x48\x00\x80\x27\x48\x00\x80\xBF\x15\x0E\x00\xBE\x15"
-     "\x0E\x00"
-     "\xBD\x25\x0E\x00\xBC\x2D\x0E\x00\x2B\x48\x00\x80\xA3\xE1\x0E\x00\x4F\x47"
-     "\x00\x80"
-     "\xBE\x84\x03\x00\xA6\xE5\x0E\x00\x2F\x48\x00\x80\x33\x48\x00\x80\xA5\xED"
-     "\x0E\x00"
-     "\xAA\x65\x0E\x00\xAB\x6D\x0E\x00\x86\x48\x00\x00\x87\x6C\x00\x00\xAE\x51"
-     "\x0E\x00"
-     "\xAF\x51\x0E\x00\xAC\x69\x0E\x00\xAD\x61\x0E\x00\xA8\x91\x0E\x00\xA9\xAD"
-     "\x0E\x00"
-     "\xAA\xB9\x0E\x00\xAB\xB9\x0E\x00\xAC\xAD\x0E\x00\xAD\xD5\x0E\x00\xAE\xDD"
-     "\x0E\x00"
-     "\xAF\xD5\x0E\x00\x80\x89\x01\x00\x81\x89\x01\x00\x82\xE9\x01\x00\x37\x48"
-     "\x00\x80"
-     "\x3B\x48\x00\x80\x3F\x48\x00\x80\x43\x48\x00\x80\x47\x48\x00\x80\xB8\x75"
-     "\x01\x00"
-     "\xB9\x7D\x01\x00\xBA\x75\x01\x00\xBB\xC9\x01\x00\xBC\xD9\x01\x00\xBD\xD9"
-     "\x01\x00"
-     "\xBE\xC9\x01\x00\xBF\xC1\x01\x00\xB0\xB1\x0E\x00\xB1\xB9\x0E\x00\xB2\x89"
-     "\x0E\x00"
-     "\xB3\x89\x0E\x00\xB4\x55\x01\x00\xB5\x5D\x01\x00\xB6\x55\x01\x00\xB7\x4D"
-     "\x01\x00"
-     "\xB3\x31\x0E\x00\x4B\x48\x00\x80\x4F\x48\x00\x80\x53\x48\x00\x80\x57\x48"
-     "\x00\x80"
-     "\xB6\x55\x0E\x00\xB5\x25\x0E\x00\x5B\x48\x00\x80\xBB\x7D\x0E\x00\xBA\x75"
-     "\x0E\x00"
-     "\x5F\x48\x00\x80\x63\x48\x00\x80\xBF\x99\x01\x00\xBE\x91\x01\x00\xBD\x5D"
-     "\x0E\x00"
-     "\xBC\x5D\x0E\x00\x67\x48\x00\x80\xA3\x75\x0E\x00\x6B\x48\x00\x80\x6F\x48"
-     "\x00\x80"
-     "\xA6\x11\x0E\x00\x73\x48\x00\x80\x77\x48\x00\x80\xA5\x61\x0E\x00\xAA\x31"
-     "\x0E\x00"
-     "\xAB\x39\x0E\x00\x7B\x48\x00\x80\x7F\x48\x00\x80\xAE\xD5\x01\x00\xAF\xDD"
-     "\x01\x00"
-     "\xAC\x19\x0E\x00\xAD\x19\x0E\x00\xA8\xD9\x02\x00\xA9\x29\x03\x00\xAA\x39"
-     "\x03\x00"
-     "\xAB\x39\x03\x00\xAC\x29\x03\x00\xAD\x29\x03\x00\xAE\x9D\x03\x00\xAF\x95"
-     "\x03\x00"
-     "\x83\x48\x00\x80\x87\x48\x00\x80\x8B\x48\x00\x80\x8F\x48\x00\x80\x93\x48"
-     "\x00\x80"
-     "\xBE\x78\x02\x00\x97\x48\x00\x80\x9B\x48\x00\x80\xB8\x95\x03\x00\xB9\x9D"
-     "\x03\x00"
-     "\xBA\xA1\x03\x00\xBB\xA1\x03\x00\xBC\x71\x00\x00\xBD\x71\x00\x00\xBE\x71"
-     "\x00\x00"
-     "\xBF\x71\x00\x00\xB0\xED\x03\x00\xB1\xF5\x03\x00\xB2\xF5\x03\x00\xB3\xC5"
-     "\x03\x00"
-     "\xB4\xDD\x03\x00\xB5\xB5\x03\x00\xB6\xB1\x03\x00\xB7\xB1\x03\x00\x80\xB1"
-     "\x00\x00"
-     "\x81\xBD\x00\x00\x82\xB5\x00\x00\xB3\xBD\x03\x00\xA3\x48\x00\x80\xB5\xAD"
-     "\x03\x00"
-     "\xB6\xA5\x03\x00\xA7\x48\x00\x80\x84\xE0\x02\x00\xAB\x48\x00\x80\xBA\x25"
-     "\x03\x00"
-     "\xBB\x3D\x03\x00\xBC\x2D\x03\x00\xBD\x1D\x03\x00\xBE\x15\x03\x00\xBF\x7D"
-     "\x03\x00"
-     "\xA3\xF1\x03\x00\xAF\x48\x00\x80\x86\x68\x04\x00\x87\xA8\x03\x00\xB3\x48"
-     "\x00\x80"
-     "\xA6\xE9\x03\x00\xA5\xE1\x03\x00\xB7\x48\x00\x80\xAB\x71\x03\x00\xAA\x69"
-     "\x03\x00"
-     "\xBB\x48\x00\x80\xBF\x48\x00\x80\xAF\x31\x03\x00\xAE\x59\x03\x00\xAD\x51"
-     "\x03\x00"
-     "\xAC\x61\x03\x00\xC3\x48\x00\x80\xC7\x48\x00\x80\xCB\x48\x00\x80\xCF\x48"
-     "\x00\x80"
-     "\xD3\x48\x00\x80\xD7\x48\x00\x80\xDB\x48\x00\x80\xDF\x48\x00\x80\xC8\x00"
-     "\x00\x00"
-     "\xE3\x48\x00\x80\xE7\x48\x00\x80\xEB\x48\x00\x80\xEF\x48\x00\x80\xF3\x48"
-     "\x00\x80"
-     "\xF7\x48\x00\x80\xFB\x48\x00\x80\xFF\x48\x00\x80\xE3\xD8\x0E\x00\x84\x44"
-     "\x02\x00"
-     "\xE1\xBC\x0C\x00\xE1\xA0\x01\x00\x03\x49\x00\x80\xE3\xE8\x03\x00\x07\x49"
-     "\x00\x80"
-     "\x81\x3D\x00\x00\x80\x31\x00\x00\xEF\xA4\x03\x00\x82\x1D\x00\x00\x0B\x49"
-     "\x00\x80"
-     "\xEF\xA4\x0F\x00\xBE\x74\x05\x00\x13\x49\x00\x80\x86\x20\x04\x00\x87\x08"
-     "\x05\x00"
-     "\x17\x49\x00\x80\x1B\x49\x00\x80\x1F\x49\x00\x80\x23\x49\x00\x80\x27\x49"
-     "\x00\x80"
-     "\xEF\xA8\x01\x00\x84\xA0\x05\x00\xE1\x24\x0E\x00\x2B\x49\x00\x80\xE3\x1C"
-     "\x0E\x00"
-     "\x2F\x49\x00\x80\x33\x49\x00\x80\x37\x49\x00\x80\x3B\x49\x00\x80\xB3\x65"
-     "\x01\x00"
-     "\x3F\x49\x00\x80\x43\x49\x00\x80\x47\x49\x00\x80\x4B\x49\x00\x80\xB6\x09"
-     "\x01\x00"
-     "\xB5\x09\x01\x00\x4F\x49\x00\x80\xBB\xE5\x00\x00\xBA\xE5\x00\x00\x53\x49"
-     "\x00\x80"
-     "\x57\x49\x00\x80\xBF\xE5\x00\x00\xBE\xE5\x00\x00\xBD\xF5\x00\x00\xBC\xF5"
-     "\x00\x00"
-     "\xA8\x91\x06\x00\xA9\x99\x06\x00\xAA\xAD\x06\x00\xAB\xC9\x06\x00\xAC\xC1"
-     "\x06\x00"
-     "\xAD\xC9\x06\x00\xAE\xF1\x06\x00\xAF\xF1\x06\x00\x0F\x49\x00\x80\x5B\x49"
-     "\x00\x80"
-     "\x5F\x49\x00\x80\x63\x49\x00\x80\x67\x49\x00\x80\x6B\x49\x00\x80\x6F\x49"
-     "\x00\x80"
-     "\x73\x49\x00\x80\xB8\x65\x07\x00\xB9\x6D\x07\x00\xBA\x65\x07\x00\xBB\x7D"
-     "\x07\x00"
-     "\xBC\x65\x07\x00\xBD\x6D\x07\x00\xBE\x65\x07\x00\xBF\x1D\x07\x00\xB0\x61"
-     "\x07\x00"
-     "\xB1\x61\x07\x00\xB2\x61\x07\x00\xB3\x61\x07\x00\xB4\x61\x07\x00\xB5\x61"
-     "\x07\x00"
-     "\xB6\x61\x07\x00\xB7\x61\x07\x00\xA3\x25\x06\x00\x82\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x80\xB1\x00\x00\x77\x49\x00\x80\xA6\x49\x06\x00\xA5\x49\x06\x00\x7B\x49"
-     "\x00\x80"
-     "\xAB\xA5\x07\x00\xAA\xA5\x07\x00\xBE\xA4\x03\x00\x7F\x49\x00\x80\xAF\xA5"
-     "\x07\x00"
-     "\xAE\xA5\x07\x00\xAD\xB5\x07\x00\xAC\xB5\x07\x00\x86\x60\x00\x00\x87\x18"
-     "\x03\x00"
-     "\x83\x49\x00\x80\xB3\x41\x07\x00\x87\x49\x00\x80\xB5\x41\x07\x00\xB6\x41"
-     "\x07\x00"
-     "\x9F\x48\x00\x80\x8B\x49\x00\x80\x8F\x49\x00\x80\xBA\x5D\x07\x00\xBB\x55"
-     "\x07\x00"
-     "\xBC\x4D\x07\x00\xBD\x31\x07\x00\xBE\x31\x07\x00\xBF\x31\x07\x00\xA8\x7D"
-     "\x06\x00"
-     "\xA9\x8D\x06\x00\xAA\x85\x06\x00\xAB\x9D\x06\x00\xAC\x85\x06\x00\xAD\x8D"
-     "\x06\x00"
-     "\xAE\xB1\x06\x00\xAF\xAD\x06\x00\x93\x49\x00\x80\x97\x49\x00\x80\x9B\x49"
-     "\x00\x80"
-     "\x9F\x49\x00\x80\xA3\x49\x00\x80\xA7\x49\x00\x80\xAB\x49\x00\x80\xAF\x49"
-     "\x00\x80"
-     "\xB8\xA9\x06\x00\xB9\xA9\x06\x00\xBA\x6D\x01\x00\xBB\x65\x01\x00\xBC\x7D"
-     "\x01\x00"
-     "\xBD\x65\x01\x00\xBE\x6D\x01\x00\xBF\x65\x01\x00\xB0\xD5\x06\x00\xB1\xDD"
-     "\x06\x00"
-     "\xB2\xD5\x06\x00\xB3\xA9\x06\x00\xB4\xB9\x06\x00\xB5\xB9\x06\x00\xB6\xA9"
-     "\x06\x00"
-     "\xB7\xA1\x06\x00\xA3\x0D\x06\x00\xB3\x49\x00\x80\xB7\x49\x00\x80\xBB\x49"
-     "\x00\x80"
-     "\xBF\x49\x00\x80\xA6\x0D\x06\x00\xA5\x0D\x06\x00\xC3\x49\x00\x80\xAB\x19"
-     "\x06\x00"
-     "\xAA\x11\x06\x00\xC7\x49\x00\x80\xCB\x49\x00\x80\xAF\x7D\x06\x00\xAE\x7D"
-     "\x06\x00"
-     "\xAD\x7D\x06\x00\xAC\x01\x06\x00\xCF\x49\x00\x80\xD3\x49\x00\x80\xD7\x49"
-     "\x00\x80"
-     "\xDB\x49\x00\x80\xDF\x49\x00\x80\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D"
-     "\x00\x00"
-     "\xE3\x49\x00\x80\xE7\x49\x00\x80\xEB\x49\x00\x80\x84\xA0\x01\x00\xBE\xB8"
-     "\x01\x00"
-     "\xF3\x49\x00\x80\x86\x58\x1F\x00\x87\xBC\x00\x00\xA8\xC9\x02\x00\xA9\xD5"
-     "\x02\x00"
-     "\xAA\xE9\x02\x00\xAB\xE9\x02\x00\xAC\x31\x03\x00\xAD\x31\x03\x00\xAE\x31"
-     "\x03\x00"
-     "\xAF\x31\x03\x00\xF7\x49\x00\x80\xFB\x49\x00\x80\xFF\x49\x00\x80\x03\x4A"
-     "\x00\x80"
-     "\x07\x4A\x00\x80\x0B\x4A\x00\x80\x0F\x4A\x00\x80\x13\x4A\x00\x80\xB8\xDD"
-     "\x03\x00"
-     "\xB9\xED\x03\x00\xBA\xE5\x03\x00\xBB\x8D\x03\x00\xBC\x95\x03\x00\xBD\x9D"
-     "\x03\x00"
-     "\xBE\x95\x03\x00\xBF\x8D\x03\x00\xB0\x51\x03\x00\xB1\x51\x03\x00\xB2\x51"
-     "\x03\x00"
-     "\xB3\x51\x03\x00\xB4\xF5\x03\x00\xB5\xFD\x03\x00\xB6\xF5\x03\x00\xB7\xED"
-     "\x03\x00"
-     "\xB3\x61\x02\x00\x17\x4A\x00\x80\x1B\x4A\x00\x80\xBE\x80\x1C\x00\x1F\x4A"
-     "\x00\x80"
-     "\xB6\xB1\x03\x00\xB5\xB9\x03\x00\x23\x4A\x00\x80\xBB\x95\x03\x00\xBA\x95"
-     "\x03\x00"
-     "\x27\x4A\x00\x80\x2B\x4A\x00\x80\xBF\x39\x03\x00\xBE\x31\x03\x00\xBD\x85"
-     "\x03\x00"
-     "\xBC\x85\x03\x00\x2F\x4A\x00\x80\xA3\x25\x02\x00\x33\x4A\x00\x80\x37\x4A"
-     "\x00\x80"
-     "\xA6\xF5\x03\x00\x3B\x4A\x00\x80\x85\x5C\x2C\x00\xA5\xFD\x03\x00\xAA\xD1"
-     "\x03\x00"
-     "\xAB\xD1\x03\x00\x3F\x4A\x00\x80\x43\x4A\x00\x80\xAE\x75\x03\x00\xAF\x7D"
-     "\x03\x00"
-     "\xAC\xC1\x03\x00\xAD\xC1\x03\x00\x47\x4A\x00\x80\x4B\x4A\x00\x80\x84\x04"
-     "\x02\x00"
-     "\x84\x08\x1D\x00\xEF\x18\x00\x00\x4F\x4A\x00\x80\x53\x4A\x00\x80\x57\x4A"
-     "\x00\x80"
-     "\xE3\xDC\x00\x00\x5B\x4A\x00\x80\xE1\xF0\x01\x00\x63\x4A\x00\x80\x67\x4A"
-     "\x00\x80"
-     "\x6B\x4A\x00\x80\x6F\x4A\x00\x80\x73\x4A\x00\x80\x80\x59\x00\x00\x81\x59"
-     "\x00\x00"
-     "\x82\x6D\x00\x00\x77\x4A\x00\x80\xE3\x08\x06\x00\x7B\x4A\x00\x80\xE1\x88"
-     "\x07\x00"
-     "\x7F\x4A\x00\x80\xEF\xE0\x06\x00\x83\x4A\x00\x80\x86\xC8\x1C\x00\x87\x40"
-     "\x1D\x00"
-     "\x87\x4A\x00\x80\x8B\x4A\x00\x80\x8F\x4A\x00\x80\x93\x4A\x00\x80\x97\x4A"
-     "\x00\x80"
-     "\x9B\x4A\x00\x80\x9F\x4A\x00\x80\xA3\x4A\x00\x80\xA7\x4A\x00\x80\xAB\x4A"
-     "\x00\x80"
-     "\xAF\x4A\x00\x80\xEF\xCC\x01\x00\xBE\xAC\x1C\x00\xE1\x3C\x06\x00\xB3\x4A"
-     "\x00\x80"
-     "\xE3\x1C\x06\x00\xB7\x4A\x00\x80\xBB\x4A\x00\x80\xBF\x4A\x00\x80\xC3\x4A"
-     "\x00\x80"
-     "\xB3\xE5\x01\x00\xC7\x4A\x00\x80\xCB\x4A\x00\x80\xD3\x4A\x00\x80\xD7\x4A"
-     "\x00\x80"
-     "\xB6\xE5\x01\x00\xB5\xF5\x01\x00\xF7\x00\x00\x00\xBB\x65\x01\x00\xBA\xD9"
-     "\x01\x00"
-     "\x86\x48\x1F\x00\x87\x60\x1C\x00\xBF\x59\x01\x00\xBE\x69\x01\x00\xBD\x69"
-     "\x01\x00"
-     "\xBC\x75\x01\x00\xA8\xF5\x1D\x00\xA9\x25\x1E\x00\xAA\x2D\x1E\x00\xAB\x25"
-     "\x1E\x00"
-     "\xAC\x3D\x1E\x00\xAD\x25\x1E\x00\xAE\x2D\x1E\x00\xAF\x25\x1E\x00\x5F\x4A"
-     "\x00\x80"
-     "\x82\xED\x1F\x00\x81\xED\x1F\x00\x80\x91\x1F\x00\xCF\x4A\x00\x80\xDB\x4A"
-     "\x00\x80"
-     "\xDF\x4A\x00\x80\xE3\x4A\x00\x80\xB8\x89\x1E\x00\xB9\x89\x1E\x00\xBA\x9D"
-     "\x1E\x00"
-     "\xBB\xA5\x1E\x00\xBC\xBD\x1E\x00\xBD\x55\x1F\x00\xBE\x5D\x1F\x00\xBF\x49"
-     "\x1F\x00"
-     "\xB0\x5D\x1E\x00\xB1\xE1\x1E\x00\xB2\xE1\x1E\x00\xB3\xFD\x1E\x00\xB4\xED"
-     "\x1E\x00"
-     "\xB5\xDD\x1E\x00\xB6\xD5\x1E\x00\xB7\xBD\x1E\x00\xA3\xA5\x1E\x00\xE7\x4A"
-     "\x00\x80"
-     "\xEB\x4A\x00\x80\xEF\x4A\x00\x80\xF3\x4A\x00\x80\xA6\xA5\x1E\x00\xA5\xB5"
-     "\x1E\x00"
-     "\xF7\x4A\x00\x80\xAB\x25\x1E\x00\xAA\x99\x1E\x00\xFB\x4A\x00\x80\xFF\x4A"
-     "\x00\x80"
-     "\xAF\x19\x1E\x00\xAE\x29\x1E\x00\xAD\x29\x1E\x00\xAC\x35\x1E\x00\x03\x4B"
-     "\x00\x80"
-     "\x07\x4B\x00\x80\xB3\xFD\x1F\x00\x0B\x4B\x00\x80\xB5\xED\x1F\x00\x0F\x4B"
-     "\x00\x80"
-     "\x13\x4B\x00\x80\xB6\x91\x1F\x00\xEF\x49\x00\x80\x17\x4B\x00\x80\xBB\x99"
-     "\x1F\x00"
-     "\xBA\x89\x1F\x00\xBD\x89\x1F\x00\xBC\x81\x1F\x00\xBF\x79\x1F\x00\xBE\x81"
-     "\x1F\x00"
-     "\xA8\x2D\x1E\x00\xA9\x35\x1E\x00\xAA\x3D\x1E\x00\xAB\x35\x1E\x00\xAC\x2D"
-     "\x1E\x00"
-     "\xAD\x59\x1E\x00\xAE\x4D\x1E\x00\xAF\x41\x1E\x00\x1B\x4B\x00\x80\x1F\x4B"
-     "\x00\x80"
-     "\x23\x4B\x00\x80\x27\x4B\x00\x80\x2B\x4B\x00\x80\x2F\x4B\x00\x80\x33\x4B"
-     "\x00\x80"
-     "\x37\x4B\x00\x80\xB8\xCD\x1E\x00\xB9\xD5\x1E\x00\xBA\xDD\x1E\x00\xBB\xD5"
-     "\x1E\x00"
-     "\xBC\xF9\x1E\x00\xBD\xE5\x1E\x00\xBE\xED\x1E\x00\xBF\xE5\x1E\x00\xB0\x05"
-     "\x1E\x00"
-     "\xB1\x15\x1E\x00\xB2\x1D\x1E\x00\xB3\x15\x1E\x00\xB4\x0D\x1E\x00\xB5\xF5"
-     "\x1E\x00"
-     "\xB6\xFD\x1E\x00\xB7\xF5\x1E\x00\xA3\xB5\x1E\x00\x82\x15\x00\x00\x81\xB5"
-     "\x00\x00"
-     "\x80\xB5\x00\x00\x3B\x4B\x00\x80\xA6\xD9\x1E\x00\xA5\xA5\x1E\x00\xBE\x90"
-     "\x01\x00"
-     "\xAB\xD1\x1E\x00\xAA\xC1\x1E\x00\x84\xA0\x02\x00\x43\x4B\x00\x80\xAF\x31"
-     "\x1E\x00"
-     "\xAE\xC9\x1E\x00\xAD\xC1\x1E\x00\xAC\xC9\x1E\x00\x47\x4B\x00\x80\xB3\xF1"
-     "\x01\x00"
-     "\x86\x68\x07\x00\x87\x1C\x01\x00\xB6\x55\x01\x00\x4B\x4B\x00\x80\x4F\x4B"
-     "\x00\x80"
-     "\xB5\xE1\x01\x00\xBA\x71\x01\x00\xBB\x79\x01\x00\x53\x4B\x00\x80\x57\x4B"
-     "\x00\x80"
-     "\xBE\x31\x01\x00\xBF\x31\x01\x00\xBC\x51\x01\x00\xBD\x51\x01\x00\xA8\xCD"
-     "\x1D\x00"
-     "\xA9\xD5\x1D\x00\xAA\x3D\x02\x00\xAB\x35\x02\x00\xAC\x2D\x02\x00\xAD\x91"
-     "\x02\x00"
-     "\xAE\x91\x02\x00\xAF\x91\x02\x00\x5B\x4B\x00\x80\x5F\x4B\x00\x80\x63\x4B"
-     "\x00\x80"
-     "\x8B\x00\x00\x00\x67\x4B\x00\x80\x6B\x4B\x00\x80\x6F\x4B\x00\x80\x73\x4B"
-     "\x00\x80"
-     "\xB8\x99\x02\x00\xB9\xA1\x02\x00\xBA\x69\x03\x00\xBB\x69\x03\x00\xBC\x79"
-     "\x03\x00"
-     "\xBD\x79\x03\x00\xBE\x69\x03\x00\xBF\x69\x03\x00\xB0\xF1\x02\x00\xB1\xF9"
-     "\x02\x00"
-     "\xB2\xC1\x02\x00\xB3\xC1\x02\x00\xB4\xB1\x02\x00\xB5\xB1\x02\x00\xB6\xB5"
-     "\x02\x00"
-     "\xB7\xA9\x02\x00\xA3\xB5\x02\x00\x77\x4B\x00\x80\x7B\x4B\x00\x80\x7F\x4B"
-     "\x00\x80"
-     "\x83\x4B\x00\x80\xA6\x11\x02\x00\xA5\xA5\x02\x00\x87\x4B\x00\x80\xAB\x3D"
-     "\x02\x00"
-     "\xAA\x35\x02\x00\x8B\x4B\x00\x80\x8F\x4B\x00\x80\xAF\x75\x02\x00\xAE\x75"
-     "\x02\x00"
-     "\xAD\x15\x02\x00\xAC\x15\x02\x00\x93\x4B\x00\x80\x97\x4B\x00\x80\x9B\x4B"
-     "\x00\x80"
-     "\x9F\x4B\x00\x80\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\xA3\x4B"
-     "\x00\x80"
-     "\xA7\x4B\x00\x80\xAB\x4B\x00\x80\x84\x24\x03\x00\xBE\xA0\x03\x00\x86\x90"
-     "\x04\x00"
-     "\x87\xD8\x04\x00\xB3\x4B\x00\x80\xB7\x4B\x00\x80\xB3\xD5\x02\x00\xBB\x4B"
-     "\x00\x80"
-     "\x84\xC4\x04\x00\xBE\xC4\x04\x00\xBF\x4B\x00\x80\xB6\xF9\x02\x00\xB5\xF1"
-     "\x02\x00"
-     "\xC3\x4B\x00\x80\xBB\x11\x01\x00\xBA\x0D\x01\x00\xC7\x4B\x00\x80\xCB\x4B"
-     "\x00\x80"
-     "\xBF\xB5\x01\x00\xBE\x09\x01\x00\xBD\x01\x01\x00\xBC\x09\x01\x00\xCF\x4B"
-     "\x00\x80"
-     "\xD3\x4B\x00\x80\xD7\x4B\x00\x80\xDB\x4B\x00\x80\xDF\x4B\x00\x80\xE3\x4B"
-     "\x00\x80"
-     "\xE7\x4B\x00\x80\xEF\x28\x1F\x00\xEB\x4B\x00\x80\xE1\x84\x1C\x00\xEF\x4B"
-     "\x00\x80"
-     "\xE3\x74\x1E\x00\xF3\x4B\x00\x80\xF7\x4B\x00\x80\xFB\x4B\x00\x80\xFF\x4B"
-     "\x00\x80"
-     "\x03\x4C\x00\x80\x07\x4C\x00\x80\x0B\x4C\x00\x80\xEF\xE8\x00\x00\x0F\x4C"
-     "\x00\x80"
-     "\x13\x4C\x00\x80\x17\x4C\x00\x80\xEF\xC4\x1E\x00\x84\x6C\x07\x00\xE1\xD4"
-     "\x1E\x00"
-     "\x1B\x4C\x00\x80\xE3\xB0\x1E\x00\x23\x4C\x00\x80\xE1\x44\x1E\x00\xBE\x70"
-     "\x07\x00"
-     "\xE3\x04\x00\x00\xA3\x15\x02\x00\x27\x4C\x00\x80\x86\xE8\x04\x00\x87\x6C"
-     "\x04\x00"
-     "\x85\xEC\x2A\x00\xA6\x39\x02\x00\xA5\x31\x02\x00\x2B\x4C\x00\x80\xAB\xD1"
-     "\x01\x00"
-     "\xAA\xCD\x01\x00\x2F\x4C\x00\x80\x33\x4C\x00\x80\xAF\x75\x01\x00\xAE\xC9"
-     "\x01\x00"
-     "\xAD\xC1\x01\x00\xAC\xC9\x01\x00\xA8\xAD\x05\x00\xA9\xC5\x05\x00\xAA\xDD"
-     "\x05\x00"
-     "\xAB\xD5\x05\x00\xAC\xF9\x05\x00\xAD\xF9\x05\x00\xAE\x39\x06\x00\xAF\x39"
-     "\x06\x00"
-     "\xAF\x4B\x00\x80\x82\x8D\x07\x00\x81\xB1\x07\x00\x80\xB9\x07\x00\x1F\x4C"
-     "\x00\x80"
-     "\x37\x4C\x00\x80\x3B\x4C\x00\x80\x3F\x4C\x00\x80\xB8\xD1\x06\x00\xB9\xD9"
-     "\x06\x00"
-     "\xBA\xE1\x06\x00\xBB\xE1\x06\x00\xBC\x91\x06\x00\xBD\x91\x06\x00\xBE\x91"
-     "\x06\x00"
-     "\xBF\x91\x06\x00\xB0\x49\x06\x00\xB1\x55\x06\x00\xB2\x5D\x06\x00\xB3\x55"
-     "\x06\x00"
-     "\xB4\x4D\x06\x00\xB5\xF1\x06\x00\xB6\xF1\x06\x00\xB7\xF1\x06\x00\xB8\xF1"
-     "\x07\x00"
-     "\xB9\xF1\x07\x00\xBA\xF1\x07\x00\xBB\xF1\x07\x00\xBC\x11\x07\x00\xBD\x11"
-     "\x07\x00"
-     "\xBE\x11\x07\x00\xBF\x0D\x07\x00\xB0\xE5\x07\x00\xB1\xED\x07\x00\xB2\xE5"
-     "\x07\x00"
-     "\xB3\xFD\x07\x00\xB4\xE5\x07\x00\xB5\xED\x07\x00\xB6\xD9\x07\x00\xB7\xD5"
-     "\x07\x00"
-     "\xA8\x39\x06\x00\xA9\x39\x06\x00\xAA\xCD\x07\x00\xAB\xDD\x07\x00\xAC\xCD"
-     "\x07\x00"
-     "\xAD\xFD\x07\x00\xAE\xF5\x07\x00\xAF\x9D\x07\x00\x43\x4C\x00\x80\x47\x4C"
-     "\x00\x80"
-     "\x4B\x4C\x00\x80\x4F\x4C\x00\x80\x53\x4C\x00\x80\x3F\x4B\x00\x80\x57\x4C"
-     "\x00\x80"
-     "\x5B\x4C\x00\x80\xB3\x05\x06\x00\x5F\x4C\x00\x80\x63\x4C\x00\x80\x67\x4C"
-     "\x00\x80"
-     "\x6B\x4C\x00\x80\xB6\x3D\x06\x00\xB5\x05\x06\x00\x6F\x4C\x00\x80\xBB\x69"
-     "\x06\x00"
-     "\xBA\x61\x06\x00\x73\x4C\x00\x80\x77\x4C\x00\x80\xBF\x69\x06\x00\xBE\x61"
-     "\x06\x00"
-     "\xBD\x69\x06\x00\xBC\x71\x06\x00\x7B\x4C\x00\x80\xA3\x41\x06\x00\x7F\x4C"
-     "\x00\x80"
-     "\x83\x4C\x00\x80\xA6\x79\x06\x00\x87\x4C\x00\x80\x8B\x4C\x00\x80\xA5\x41"
-     "\x06\x00"
-     "\xAA\x25\x06\x00\xAB\x2D\x06\x00\x8F\x4C\x00\x80\x93\x4C\x00\x80\xAE\x25"
-     "\x06\x00"
-     "\xAF\x2D\x06\x00\xAC\x35\x06\x00\xAD\x2D\x06\x00\xA8\x31\x06\x00\xA9\x41"
-     "\x06\x00"
-     "\xAA\x55\x06\x00\xAB\x65\x06\x00\xAC\x7D\x06\x00\xAD\xA1\x01\x00\xAE\xA1"
-     "\x01\x00"
-     "\xAF\xA1\x01\x00\x80\xED\x00\x00\x81\xF1\x00\x00\x82\xF1\x00\x00\x97\x4C"
-     "\x00\x80"
-     "\x86\x00\x0C\x00\x87\x14\x01\x00\x9B\x4C\x00\x80\x9F\x4C\x00\x80\xB8\xC9"
-     "\x00\x00"
-     "\xB9\xC9\x00\x00\xBA\xD9\x00\x00\xBB\xD1\x00\x00\xBC\xF9\x00\x00\xBD\xF9"
-     "\x00\x00"
-     "\xBE\x99\x00\x00\xBF\x95\x00\x00\xB0\x01\x01\x00\xB1\x01\x01\x00\xB2\x01"
-     "\x01\x00"
-     "\xB3\x01\x01\x00\xB4\x01\x01\x00\xB5\x0D\x01\x00\xB6\x05\x01\x00\xB7\xF9"
-     "\x00\x00"
-     "\xB3\xDD\x05\x00\xA3\x4C\x00\x80\xA7\x4C\x00\x80\xAB\x4C\x00\x80\xAF\x4C"
-     "\x00\x80"
-     "\xB6\x21\x02\x00\xB5\xF5\x05\x00\xB3\x4C\x00\x80\xBB\xC5\x02\x00\xBA\xC5"
-     "\x02\x00"
-     "\xB7\x4C\x00\x80\xBE\x68\x03\x00\xBF\xC9\x02\x00\xBE\xC5\x02\x00\xBD\xD5"
-     "\x02\x00"
-     "\xBC\xD5\x02\x00\xBB\x4C\x00\x80\xA3\x99\x05\x00\xBF\x4C\x00\x80\xC3\x4C"
-     "\x00\x80"
-     "\xA6\x65\x02\x00\xC7\x4C\x00\x80\xCB\x4C\x00\x80\xA5\xB1\x05\x00\xAA\x81"
-     "\x02\x00"
-     "\xAB\x81\x02\x00\xCF\x4C\x00\x80\xD3\x4C\x00\x80\xAE\x81\x02\x00\xAF\x8D"
-     "\x02\x00"
-     "\xAC\x91\x02\x00\xAD\x91\x02\x00\xD7\x4C\x00\x80\xDB\x4C\x00\x80\xDF\x4C"
-     "\x00\x80"
-     "\xE3\x4C\x00\x80\xE7\x4C\x00\x80\xEB\x4C\x00\x80\xEF\x4C\x00\x80\xF3\x4C"
-     "\x00\x80"
-     "\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\xF7\x4C\x00\x80\xFB\x4C"
-     "\x00\x80"
-     "\x84\x78\x03\x00\x03\x4D\x00\x80\x07\x4D\x00\x80\x86\x40\x0C\x00\x87\x98"
-     "\x03\x00"
-     "\x0B\x4D\x00\x80\x0F\x4D\x00\x80\x13\x4D\x00\x80\xEF\x00\x02\x00\x17\x4D"
-     "\x00\x80"
-     "\x1B\x4D\x00\x80\x1F\x4D\x00\x80\xE3\x40\x02\x00\x23\x4D\x00\x80\xE1\x44"
-     "\x01\x00"
-     "\x27\x4D\x00\x80\x2B\x4D\x00\x80\x2F\x4D\x00\x80\x33\x4D\x00\x80\xB3\xB9"
-     "\x02\x00"
-     "\x37\x4D\x00\x80\x3B\x4D\x00\x80\x3F\x4D\x00\x80\x43\x4D\x00\x80\xB6\x1D"
-     "\x01\x00"
-     "\xB5\x1D\x01\x00\x47\x4D\x00\x80\xBB\x09\x01\x00\xBA\x39\x01\x00\x4B\x4D"
-     "\x00\x80"
-     "\x4F\x4D\x00\x80\xBF\x65\x01\x00\xBE\x79\x01\x00\xBD\x71\x01\x00\xBC\x7D"
-     "\x01\x00"
-     "\x53\x4D\x00\x80\x57\x4D\x00\x80\xEF\x5C\x07\x00\x5B\x4D\x00\x80\x5F\x4D"
-     "\x00\x80"
-     "\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D\x00\x00\x63\x4D\x00\x80\x67\x4D"
-     "\x00\x80"
-     "\x6B\x4D\x00\x80\x84\x20\x0D\x00\xE1\xA0\x04\x00\xBE\x3C\x0D\x00\xE3\x80"
-     "\x07\x00"
-     "\x73\x4D\x00\x80\x86\x20\x0C\x00\x87\x4C\x0D\x00\x77\x4D\x00\x80\xEF\xCC"
-     "\x06\x00"
-     "\x7B\x4D\x00\x80\x7F\x4D\x00\x80\x83\x4D\x00\x80\xEF\x68\x06\x00\x84\xEC"
-     "\x0C\x00"
-     "\xE1\x4C\x07\x00\xBE\xE0\x0C\x00\xE3\xB4\x06\x00\x87\x4D\x00\x80\xE1\x64"
-     "\x06\x00"
-     "\x8B\x4D\x00\x80\xE3\x0C\x06\x00\x8F\x4D\x00\x80\x93\x4D\x00\x80\xA3\x71"
-     "\x02\x00"
-     "\x84\xE0\x0F\x00\xA5\xD5\x01\x00\x97\x4D\x00\x80\x9B\x4D\x00\x80\xA6\xD5"
-     "\x01\x00"
-     "\x9F\x4D\x00\x80\xA3\x4D\x00\x80\xAB\xC1\x01\x00\xAA\xF1\x01\x00\xAD\xB9"
-     "\x01\x00"
-     "\xAC\xB5\x01\x00\xAF\xAD\x01\x00\xAE\xB1\x01\x00\xFF\x4C\x00\x80\x6F\x4D"
-     "\x00\x80"
-     "\xA7\x4D\x00\x80\xAB\x4D\x00\x80\xAF\x4D\x00\x80\xB3\x4D\x00\x80\xB7\x4D"
-     "\x00\x80"
-     "\xBB\x4D\x00\x80\xA8\x99\x0E\x00\xA9\x99\x0E\x00\xAA\xE1\x0E\x00\xAB\xE1"
-     "\x0E\x00"
-     "\xAC\xE1\x0E\x00\xAD\xED\x0E\x00\xAE\xE5\x0E\x00\xAF\xDD\x0E\x00\xB0\xA9"
-     "\x0E\x00"
-     "\xB1\xB5\x0E\x00\xB2\xBD\x0E\x00\xB3\xB5\x0E\x00\xB4\xAD\x0E\x00\xB5\x51"
-     "\x0F\x00"
-     "\xB6\x51\x0F\x00\xB7\x51\x0F\x00\xB8\x71\x0F\x00\xB9\x79\x0F\x00\xBA\x41"
-     "\x0F\x00"
-     "\xBB\x41\x0F\x00\xBC\x31\x0F\x00\xBD\x31\x0F\x00\xBE\x35\x0F\x00\xBF\x2D"
-     "\x0F\x00"
-     "\xB3\x49\x0E\x00\xBF\x4D\x00\x80\xC3\x4D\x00\x80\xC7\x4D\x00\x80\xCB\x4D"
-     "\x00\x80"
-     "\xB6\x4D\x0E\x00\xB5\x49\x0E\x00\xCF\x4D\x00\x80\xBB\x51\x0E\x00\xBA\x4D"
-     "\x0E\x00"
-     "\xD3\x4D\x00\x80\xD7\x4D\x00\x80\xBF\xB5\x0F\x00\xBE\x49\x0E\x00\xBD\x41"
-     "\x0E\x00"
-     "\xBC\x49\x0E\x00\x82\x15\x00\x00\xA3\x0D\x0E\x00\x80\x61\x00\x00\x81\x61"
-     "\x00\x00"
-     "\xA6\x09\x0E\x00\xDB\x4D\x00\x80\xE3\x4D\x00\x80\xA5\x0D\x0E\x00\xAA\x09"
-     "\x0E\x00"
-     "\xAB\x15\x0E\x00\x86\x48\x00\x00\x87\xDC\x00\x00\xAE\x0D\x0E\x00\xAF\xF1"
-     "\x0F\x00"
-     "\xAC\x0D\x0E\x00\xAD\x05\x0E\x00\xA8\x7D\x0E\x00\xA9\x8D\x0E\x00\xAA\x99"
-     "\x0E\x00"
-     "\xAB\x99\x0E\x00\xAC\x89\x0E\x00\xAD\xA5\x0E\x00\xAE\xA1\x0E\x00\xAF\xDD"
-     "\x0E\x00"
-     "\xDF\x4D\x00\x80\xE7\x4D\x00\x80\xEB\x4D\x00\x80\xEF\x4D\x00\x80\xF3\x4D"
-     "\x00\x80"
-     "\xF7\x4D\x00\x80\xFB\x4D\x00\x80\xFF\x4D\x00\x80\xB8\x55\x01\x00\xB9\x5D"
-     "\x01\x00"
-     "\xBA\x55\x01\x00\xBB\xC9\x01\x00\xBC\xD9\x01\x00\xBD\xD9\x01\x00\xBE\xC9"
-     "\x01\x00"
-     "\xBF\xC1\x01\x00\xB0\xA5\x0E\x00\xB1\xB1\x0E\x00\xB2\xB1\x0E\x00\xB3\x8D"
-     "\x0E\x00"
-     "\xB4\x95\x0E\x00\xB5\x7D\x01\x00\xB6\x75\x01\x00\xB7\x6D\x01\x00\xB3\x11"
-     "\x0E\x00"
-     "\x03\x4E\x00\x80\x07\x4E\x00\x80\x0B\x4E\x00\x80\x0F\x4E\x00\x80\xB6\x25"
-     "\x0E\x00"
-     "\xB5\x01\x0E\x00\x13\x4E\x00\x80\xBB\x75\x0E\x00\xBA\x61\x0E\x00\x17\x4E"
-     "\x00\x80"
-     "\x1B\x4E\x00\x80\xBF\xB9\x01\x00\xBE\x51\x0E\x00\xBD\x59\x0E\x00\xBC\x65"
-     "\x0E\x00"
-     "\x1F\x4E\x00\x80\xA3\x55\x0E\x00\x23\x4E\x00\x80\x27\x4E\x00\x80\xA6\x61"
-     "\x0E\x00"
-     "\x2B\x4E\x00\x80\x2F\x4E\x00\x80\xA5\x45\x0E\x00\xAA\x25\x0E\x00\xAB\x31"
-     "\x0E\x00"
-     "\x33\x4E\x00\x80\x37\x4E\x00\x80\xAE\x15\x0E\x00\xAF\xFD\x01\x00\xAC\x21"
-     "\x0E\x00"
-     "\xAD\x1D\x0E\x00\xA8\xD9\x02\x00\xA9\x21\x03\x00\xAA\x21\x03\x00\xAB\x21"
-     "\x03\x00"
-     "\xAC\x21\x03\x00\xAD\x21\x03\x00\xAE\x21\x03\x00\xAF\x21\x03\x00\x3B\x4E"
-     "\x00\x80"
-     "\x3F\x4E\x00\x80\x43\x4E\x00\x80\x47\x4E\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x4B\x4E\x00\x80\xB8\x89\x03\x00\xB9\x89\x03\x00\xBA\x9D"
-     "\x03\x00"
-     "\xBB\x95\x03\x00\xBC\xB9\x03\x00\xBD\xB9\x03\x00\xBE\x79\x00\x00\xBF\x79"
-     "\x00\x00"
-     "\xB0\xE5\x03\x00\xB1\xED\x03\x00\xB2\xE5\x03\x00\xB3\xFD\x03\x00\xB4\xED"
-     "\x03\x00"
-     "\xB5\xDD\x03\x00\xB6\xD5\x03\x00\xB7\xBD\x03\x00\xBE\xCC\x03\x00\x53\x4E"
-     "\x00\x80"
-     "\x87\x28\x03\x00\x86\x2C\x04\x00\x57\x4E\x00\x80\x5B\x4E\x00\x80\x5F\x4E"
-     "\x00\x80"
-     "\xEF\xE4\x02\x00\x63\x4E\x00\x80\xE1\x54\x00\x00\x67\x4E\x00\x80\xE3\x70"
-     "\x0F\x00"
-     "\x6B\x4E\x00\x80\x6F\x4E\x00\x80\x73\x4E\x00\x80\x77\x4E\x00\x80\xBB\x49"
-     "\x03\x00"
-     "\xBA\x41\x03\x00\x85\x00\x24\x00\x7B\x4E\x00\x80\xBF\x79\x03\x00\xBE\x49"
-     "\x03\x00"
-     "\xBD\x49\x03\x00\xBC\x51\x03\x00\xB3\xC9\x03\x00\x7F\x4E\x00\x80\x83\x4E"
-     "\x00\x80"
-     "\x87\x4E\x00\x80\x84\xBC\x04\x00\xB6\xC1\x03\x00\xB5\xC9\x03\x00\x8B\x4E"
-     "\x00\x80"
-     "\xAC\x05\x0D\x00\xAD\x0D\x0D\x00\xAE\x05\x0D\x00\xAF\x1D\x0D\x00\xA8\x7D"
-     "\x0E\x00"
-     "\xA9\xB9\x0E\x00\xAA\xA1\x0E\x00\xAB\xA1\x0E\x00\x8F\x4E\x00\x80\x93\x4E"
-     "\x00\x80"
-     "\x97\x4E\x00\x80\x9B\x4E\x00\x80\x9F\x4E\x00\x80\xA3\x4E\x00\x80\xA7\x4E"
-     "\x00\x80"
-     "\xAB\x4E\x00\x80\xBC\xCD\x0D\x00\xBD\xD5\x0D\x00\xBE\xDD\x0D\x00\xBF\xD5"
-     "\x0D\x00"
-     "\xB8\x3D\x0D\x00\xB9\xD5\x0D\x00\xBA\xDD\x0D\x00\xBB\xD5\x0D\x00\xB4\x2D"
-     "\x0D\x00"
-     "\xB5\x31\x0D\x00\xB6\x31\x0D\x00\xB7\x05\x0D\x00\xB0\x6D\x0D\x00\xB1\x5D"
-     "\x0D\x00"
-     "\xB2\x39\x0D\x00\xB3\x35\x0D\x00\xE1\x18\x0E\x00\xE1\x5C\x0F\x00\xE3\x38"
-     "\x0E\x00"
-     "\xE3\xF0\x0E\x00\xAF\x4E\x00\x80\xB3\x4E\x00\x80\xB7\x4E\x00\x80\xBB\x4E"
-     "\x00\x80"
-     "\x80\x35\x00\x00\x81\x3D\x00\x00\x82\x35\x00\x00\xBE\xE4\x05\x00\xC3\x4E"
-     "\x00\x80"
-     "\xC7\x4E\x00\x80\xEF\xE8\x0E\x00\xEF\x30\x0E\x00\xA3\xE9\x02\x00\xCB\x4E"
-     "\x00\x80"
-     "\x86\x68\x04\x00\x87\xEC\x04\x00\xCF\x4E\x00\x80\xA6\xE1\x02\x00\xA5\xE9"
-     "\x02\x00"
-     "\xD3\x4E\x00\x80\xAB\x69\x02\x00\xAA\x61\x02\x00\xD7\x4E\x00\x80\xDB\x4E"
-     "\x00\x80"
-     "\xAF\x59\x02\x00\xAE\x69\x02\x00\xAD\x69\x02\x00\xAC\x71\x02\x00\xA8\x85"
-     "\x06\x00"
-     "\xA9\x95\x06\x00\xAA\x9D\x06\x00\xAB\x95\x06\x00\xAC\xB1\x06\x00\xAD\xBD"
-     "\x06\x00"
-     "\xAE\xA9\x06\x00\xAF\xA9\x06\x00\xBF\x4E\x00\x80\xDF\x4E\x00\x80\xE3\x4E"
-     "\x00\x80"
-     "\xE7\x4E\x00\x80\xEB\x4E\x00\x80\xEF\x4E\x00\x80\xF3\x4E\x00\x80\xF7\x4E"
-     "\x00\x80"
-     "\xB8\x65\x07\x00\xB9\x6D\x07\x00\xBA\x65\x07\x00\xBB\x7D\x07\x00\xBC\x65"
-     "\x07\x00"
-     "\xBD\x6D\x07\x00\xBE\x65\x07\x00\xBF\xCD\x07\x00\xB0\xD9\x06\x00\xB1\xD9"
-     "\x06\x00"
-     "\xB2\xA5\x06\x00\xB3\x89\x06\x00\xB4\x99\x06\x00\xB5\x99\x06\x00\xB6\x85"
-     "\x06\x00"
-     "\xB7\x5D\x07\x00\xB3\x15\x06\x00\xFB\x4E\x00\x80\x4F\x4E\x00\x80\xFF\x4E"
-     "\x00\x80"
-     "\x03\x4F\x00\x80\xB6\x3D\x06\x00\xB5\x39\x06\x00\x07\x4F\x00\x80\xBB\x1D"
-     "\x06\x00"
-     "\xBA\x1D\x06\x00\x0B\x4F\x00\x80\x0F\x4F\x00\x80\xBF\x5D\x06\x00\xBE\x5D"
-     "\x06\x00"
-     "\xBD\x5D\x06\x00\xBC\x71\x06\x00\x13\x4F\x00\x80\xA3\x51\x06\x00\x17\x4F"
-     "\x00\x80"
-     "\x1B\x4F\x00\x80\xA6\x79\x06\x00\x1F\x4F\x00\x80\x23\x4F\x00\x80\xA5\x7D"
-     "\x06\x00"
-     "\xAA\x59\x06\x00\xAB\x59\x06\x00\x27\x4F\x00\x80\x2B\x4F\x00\x80\xAE\x19"
-     "\x06\x00"
-     "\xAF\x19\x06\x00\xAC\x35\x06\x00\xAD\x19\x06\x00\xA8\xB1\x06\x00\xA9\xB1"
-     "\x06\x00"
-     "\xAA\xCD\x06\x00\xAB\xDD\x06\x00\xAC\xCD\x06\x00\xAD\xFD\x06\x00\xAE\xF5"
-     "\x06\x00"
-     "\xAF\x39\x01\x00\x80\xF9\x01\x00\x81\xCD\x01\x00\x82\xC5\x01\x00\x84\x60"
-     "\x02\x00"
-     "\xBE\x78\x02\x00\x2F\x4F\x00\x80\x87\xAC\x01\x00\x86\x9C\x01\x00\xB8\xD1"
-     "\x01\x00"
-     "\xB9\xD9\x01\x00\xBA\xE1\x01\x00\xBB\xE1\x01\x00\xBC\x91\x01\x00\xBD\x9D"
-     "\x01\x00"
-     "\xBE\x95\x01\x00\xBF\x89\x01\x00\xB0\x49\x01\x00\xB1\x49\x01\x00\xB2\x59"
-     "\x01\x00"
-     "\xB3\x59\x01\x00\xB4\x49\x01\x00\xB5\x49\x01\x00\xB6\xFD\x01\x00\xB7\xF1"
-     "\x01\x00"
-     "\xB3\x45\x06\x00\x33\x4F\x00\x80\x37\x4F\x00\x80\x3B\x4F\x00\x80\x3F\x4F"
-     "\x00\x80"
-     "\xB6\x7D\x06\x00\xB5\x45\x06\x00\x43\x4F\x00\x80\xBB\x8D\x01\x00\xBA\x8D"
-     "\x01\x00"
-     "\x47\x4F\x00\x80\x4B\x4F\x00\x80\xBF\x8D\x01\x00\xBE\x8D\x01\x00\xBD\x8D"
-     "\x01\x00"
-     "\xBC\x8D\x01\x00\x4F\x4F\x00\x80\xA3\x01\x06\x00\x53\x4F\x00\x80\x57\x4F"
-     "\x00\x80"
-     "\xA6\x39\x06\x00\x5B\x4F\x00\x80\x5F\x4F\x00\x80\xA5\x01\x06\x00\xAA\xC9"
-     "\x01\x00"
-     "\xAB\xC9\x01\x00\x63\x4F\x00\x80\x67\x4F\x00\x80\xAE\xC9\x01\x00\xAF\xC9"
-     "\x01\x00"
-     "\xAC\xC9\x01\x00\xAD\xC9\x01\x00\xA8\x85\x03\x00\xA9\x8D\x03\x00\xAA\x85"
-     "\x03\x00"
-     "\xAB\x9D\x03\x00\xAC\x8D\x03\x00\xAD\xBD\x03\x00\xAE\xB5\x03\x00\xAF\xDD"
-     "\x03\x00"
-     "\x6B\x4F\x00\x80\x6F\x4F\x00\x80\x73\x4F\x00\x80\x77\x4F\x00\x80\x7B\x4F"
-     "\x00\x80"
-     "\x7F\x4F\x00\x80\x83\x4F\x00\x80\x87\x4F\x00\x80\xB8\x69\x00\x00\xB9\x69"
-     "\x00\x00"
-     "\xBA\x79\x00\x00\xBB\x79\x00\x00\xBC\x69\x00\x00\xBD\x69\x00\x00\xBE\xD9"
-     "\x00\x00"
-     "\xBF\xD9\x00\x00\xB0\xA9\x03\x00\xB1\xA9\x03\x00\xB2\xBD\x03\x00\xB3\xB5"
-     "\x03\x00"
-     "\xB4\x99\x03\x00\xB5\x99\x03\x00\xB6\x59\x00\x00\xB7\x59\x00\x00\x8B\x4F"
-     "\x00\x80"
-     "\x8F\x4F\x00\x80\xEF\xD8\x02\x00\x93\x4F\x00\x80\x81\x09\x00\x00\x80\x19"
-     "\x00\x00"
-     "\x97\x4F\x00\x80\x82\x39\x00\x00\x84\xAC\x02\x00\x9B\x4F\x00\x80\xBE\x6C"
-     "\x03\x00"
-     "\xA3\x4F\x00\x80\xE1\x94\x01\x00\xA7\x4F\x00\x80\xE3\x08\x02\x00\xAB\x4F"
-     "\x00\x80"
-     "\xB3\x4D\x03\x00\xAF\x4F\x00\x80\x86\xE8\x3C\x00\x87\x7C\x03\x00\xB3\x4F"
-     "\x00\x80"
-     "\xB6\x7D\x03\x00\xB5\x45\x03\x00\xB7\x4F\x00\x80\xBB\x2D\x03\x00\xBA\x2D"
-     "\x03\x00"
-     "\xBB\x4F\x00\x80\xBF\x4F\x00\x80\xBF\x1D\x03\x00\xBE\x1D\x03\x00\xBD\x21"
-     "\x03\x00"
-     "\xBC\x29\x03\x00\xC3\x4F\x00\x80\xE3\xE4\x07\x00\xEF\xC8\x1B\x00\xE1\x98"
-     "\x07\x00"
-     "\xC7\x4F\x00\x80\xCB\x4F\x00\x80\xCF\x4F\x00\x80\xD3\x4F\x00\x80\xA1\xA9"
-     "\x03\x00"
-     "\xA0\x05\x06\x00\xA3\x1D\x02\x00\xA2\x0D\x03\x00\xE1\xFC\x13\x00\xEF\x60"
-     "\x06\x00"
-     "\xE3\xFC\x0E\x00\x84\x70\x3C\x00\xD7\x4F\x00\x80\xDB\x4F\x00\x80\xA3\xC5"
-     "\x02\x00"
-     "\xBE\x04\x3C\x00\xA5\xCD\x02\x00\xDF\x4F\x00\x80\xE3\x4F\x00\x80\xA6\xF5"
-     "\x02\x00"
-     "\xE7\x4F\x00\x80\xEB\x4F\x00\x80\xAB\xA5\x02\x00\xAA\xA5\x02\x00\xAD\xA9"
-     "\x02\x00"
-     "\xAC\xA1\x02\x00\xAF\x95\x02\x00\xAE\x95\x02\x00\xEF\x4F\x00\x80\xF3\x4F"
-     "\x00\x80"
-     "\xF7\x4F\x00\x80\xFB\x4F\x00\x80\xFF\x4F\x00\x80\x03\x50\x00\x80\x8F\x9D"
-     "\x1B\x00"
-     "\xEF\x54\x00\x00\xBE\xE8\x3F\x00\xE1\x74\x06\x00\x07\x50\x00\x80\xE3\x90"
-     "\x00\x00"
-     "\x80\x25\x00\x00\x81\x1D\x00\x00\x82\x05\x00\x00\x0B\x50\x00\x80\x99\xF1"
-     "\x0F\x00"
-     "\x98\xE5\x0F\x00\x9B\xE9\x0B\x00\x9A\x39\x0B\x00\x9D\xB9\x0A\x00\x9C\xE5"
-     "\x0B\x00"
-     "\x9F\xDD\x07\x00\x9E\xDD\x07\x00\x91\x29\x17\x00\x90\xCD\x1B\x00\x93\x3D"
-     "\x17\x00"
-     "\x92\xCD\x17\x00\x95\xD1\x13\x00\x94\x29\x13\x00\x97\x19\x0F\x00\x96\x51"
-     "\x13\x00"
-     "\xA1\x05\x00\x00\x9F\x4F\x00\x80\xA3\x01\x04\x00\xA2\x55\x00\x00\xA5\x0D"
-     "\x04\x00"
-     "\xA4\x11\x04\x00\xA7\x19\x08\x00\xA6\x01\x08\x00\xA9\xE5\x09\x00\xA8\xB5"
-     "\x08\x00"
-     "\xAB\x09\x0C\x00\xAA\x01\x0C\x00\xAD\x01\x10\x00\xAC\xA9\x0D\x00\xAF\xD5"
-     "\x10\x00"
-     "\xAE\x1D\x10\x00\xB1\x01\x14\x00\xB0\x01\x14\x00\xB3\x01\x18\x00\xB2\xDD"
-     "\x14\x00"
-     "\xB5\x05\x18\x00\xB4\x11\x18\x00\x0F\x50\x00\x80\x13\x50\x00\x80\x86\xC0"
-     "\x00\x00"
-     "\x87\xE8\x03\x00\x17\x50\x00\x80\x1B\x50\x00\x80\x1F\x50\x00\x80\x23\x50"
-     "\x00\x80"
-     "\x27\x50\x00\x80\x2B\x50\x00\x80\xA8\x2D\x3E\x00\xA9\x31\x3E\x00\xAA\x35"
-     "\x3E\x00"
-     "\xAB\x49\x3E\x00\xAC\x41\x3E\x00\xAD\x49\x3E\x00\xAE\x71\x3E\x00\xAF\x71"
-     "\x3E\x00"
-     "\x84\xEC\x00\x00\x2F\x50\x00\x80\x33\x50\x00\x80\x37\x50\x00\x80\x3B\x50"
-     "\x00\x80"
-     "\x3F\x50\x00\x80\x43\x50\x00\x80\x47\x50\x00\x80\xB8\x99\x3F\x00\xB9\xA5"
-     "\x3F\x00"
-     "\xBA\xAD\x3F\x00\xBB\x4D\x3F\x00\xBC\x55\x3F\x00\xBD\x5D\x3F\x00\xBE\x55"
-     "\x3F\x00"
-     "\xBF\x4D\x3F\x00\xB0\xE1\x3F\x00\xB1\xF5\x3F\x00\xB2\xFD\x3F\x00\xB3\xF5"
-     "\x3F\x00"
-     "\xB4\xED\x3F\x00\xB5\xB5\x3F\x00\xB6\xBD\x3F\x00\xB7\xA9\x3F\x00\x8B\xAD"
-     "\x2D\x00"
-     "\x8A\x25\x2D\x00\x4B\x50\x00\x80\x4F\x50\x00\x80\x8F\xE5\x35\x00\x8E\x25"
-     "\x31\x00"
-     "\x8D\x39\x31\x00\x8C\xFD\x31\x00\x83\x09\x25\x00\x82\x21\x25\x00\x53\x50"
-     "\x00\x80"
-     "\x57\x50\x00\x80\x87\xED\x2D\x00\x86\xE1\x29\x00\x85\x19\x29\x00\x84\xF5"
-     "\x29\x00"
-     "\x9B\x61\x3D\x00\x9A\xCD\x3D\x00\x5B\x50\x00\x80\x5F\x50\x00\x80\x63\x50"
-     "\x00\x80"
-     "\x67\x50\x00\x80\x6B\x50\x00\x80\x9C\x69\x3C\x00\x93\x51\x35\x00\x92\x45"
-     "\x35\x00"
-     "\x6F\x50\x00\x80\x73\x50\x00\x80\x97\x51\x38\x00\x96\x71\x39\x00\x95\x79"
-     "\x39\x00"
-     "\x94\x15\x34\x00\xB3\x3D\x3F\x00\x82\x19\x00\x00\x81\x19\x00\x00\x80\xE1"
-     "\x00\x00"
-     "\x77\x50\x00\x80\xB6\x49\x3F\x00\xB5\x49\x3F\x00\x7B\x50\x00\x80\xBB\xB1"
-     "\x3E\x00"
-     "\xBA\xA5\x3E\x00\x86\x08\x00\x00\x87\x8C\x00\x00\xBF\xF1\x3E\x00\xBE\xA9"
-     "\x3E\x00"
-     "\xBD\xA1\x3E\x00\xBC\xA9\x3E\x00\x83\x50\x00\x80\xA3\x79\x3F\x00\x87\x50"
-     "\x00\x80"
-     "\x8B\x50\x00\x80\xA6\x0D\x3F\x00\x8F\x50\x00\x80\x93\x50\x00\x80\xA5\x0D"
-     "\x3F\x00"
-     "\xAA\xE1\x3E\x00\xAB\xF5\x3E\x00\x97\x50\x00\x80\x9B\x50\x00\x80\xAE\xED"
-     "\x3E\x00"
-     "\xAF\xB5\x3E\x00\xAC\xED\x3E\x00\xAD\xE5\x3E\x00\xA8\x5D\x02\x00\xA9\x6D"
-     "\x02\x00"
-     "\xAA\x65\x02\x00\xAB\x7D\x02\x00\xAC\x65\x02\x00\xAD\x6D\x02\x00\xAE\x65"
-     "\x02\x00"
-     "\xAF\xB9\x02\x00\x84\xEC\x04\x00\x9F\x50\x00\x80\xA3\x50\x00\x80\xA7\x50"
-     "\x00\x80"
-     "\xAB\x50\x00\x80\xAF\x50\x00\x80\xB3\x50\x00\x80\xB7\x50\x00\x80\xB8\x69"
-     "\x03\x00"
-     "\xB9\x69\x03\x00\xBA\x01\x03\x00\xBB\x01\x03\x00\xBC\x01\x03\x00\xBD\x01"
-     "\x03\x00"
-     "\xBE\x01\x03\x00\xBF\x35\x03\x00\xB0\xC9\x02\x00\xB1\xC9\x02\x00\xB2\xD9"
-     "\x02\x00"
-     "\xB3\xD9\x02\x00\xB4\xC9\x02\x00\xB5\xC9\x02\x00\xB6\x71\x03\x00\xB7\x71"
-     "\x03\x00"
-     "\xB3\xED\x02\x00\x84\x48\x05\x00\xBB\x50\x00\x80\xBF\x50\x00\x80\xC3\x50"
-     "\x00\x80"
-     "\xB6\xED\x02\x00\xB5\xED\x02\x00\xC7\x50\x00\x80\xBB\x0D\x02\x00\xBA\x0D"
-     "\x02\x00"
-     "\xCB\x50\x00\x80\xCF\x50\x00\x80\xBF\x0D\x02\x00\xBE\x0D\x02\x00\xBD\x0D"
-     "\x02\x00"
-     "\xBC\x0D\x02\x00\xD3\x50\x00\x80\xA3\xA9\x02\x00\xD7\x50\x00\x80\xDB\x50"
-     "\x00\x80"
-     "\xA6\xA9\x02\x00\xDF\x50\x00\x80\xE3\x50\x00\x80\xA5\xA9\x02\x00\xAA\x49"
-     "\x02\x00"
-     "\xAB\x49\x02\x00\xE7\x50\x00\x80\xEB\x50\x00\x80\xAE\x49\x02\x00\xAF\x49"
-     "\x02\x00"
-     "\xAC\x49\x02\x00\xAD\x49\x02\x00\xA8\xC1\x02\x00\xA9\xCD\x02\x00\xAA\xC5"
-     "\x02\x00"
-     "\xAB\xD9\x02\x00\xAC\xC9\x02\x00\xAD\xFD\x02\x00\xAE\xF5\x02\x00\xAF\x39"
-     "\x01\x00"
-     "\x80\xF9\x01\x00\x81\xCD\x01\x00\x82\xC5\x01\x00\x84\xE0\x07\x00\xF3\x50"
-     "\x00\x80"
-     "\xF7\x50\x00\x80\xFB\x50\x00\x80\xBE\xF4\x05\x00\xB8\xD1\x01\x00\xB9\xD9"
-     "\x01\x00"
-     "\xBA\xE1\x01\x00\xBB\xE1\x01\x00\xBC\x91\x01\x00\xBD\x9D\x01\x00\xBE\x95"
-     "\x01\x00"
-     "\xBF\x89\x01\x00\xB0\x49\x01\x00\xB1\x49\x01\x00\xB2\x59\x01\x00\xB3\x59"
-     "\x01\x00"
-     "\xB4\x49\x01\x00\xB5\x49\x01\x00\xB6\xF1\x01\x00\xB7\xF1\x01\x00\xE1\x50"
-     "\x3F\x00"
-     "\xFF\x50\x00\x80\xE3\xA4\x3E\x00\xBE\x44\x04\x00\x86\xB0\x05\x00\x87\xDC"
-     "\x05\x00"
-     "\x03\x51\x00\x80\x07\x51\x00\x80\x0B\x51\x00\x80\x0F\x51\x00\x80\x13\x51"
-     "\x00\x80"
-     "\x17\x51\x00\x80\x85\x8C\xF7\x00\x1B\x51\x00\x80\xEF\x0C\x3E\x00\x1F\x51"
-     "\x00\x80"
-     "\x23\x51\x00\x80\x27\x51\x00\x80\x2B\x51\x00\x80\x2F\x51\x00\x80\x33\x51"
-     "\x00\x80"
-     "\xE3\x64\x01\x00\x37\x51\x00\x80\xE1\xA4\x01\x00\x3B\x51\x00\x80\xEF\xBC"
-     "\x3E\x00"
-     "\x3F\x51\x00\x80\x43\x51\x00\x80\x47\x51\x00\x80\x4B\x51\x00\x80\x4F\x51"
-     "\x00\x80"
-     "\x53\x51\x00\x80\xB3\xCD\x05\x00\xEF\x50\x00\x80\x7F\x50\x00\x80\x57\x51"
-     "\x00\x80"
-     "\x5B\x51\x00\x80\xB6\xFD\x05\x00\xB5\xC1\x05\x00\x5F\x51\x00\x80\xBB\x0D"
-     "\x06\x00"
-     "\xBA\x0D\x06\x00\x86\x08\x03\x00\x87\x8C\x03\x00\xBF\x0D\x06\x00\xBE\x0D"
-     "\x06\x00"
-     "\xBD\x0D\x06\x00\xBC\x0D\x06\x00\x82\x51\x00\x00\xA3\x89\x05\x00\x80\xB5"
-     "\x07\x00"
-     "\x81\x41\x00\x00\xA6\xB9\x05\x00\x63\x51\x00\x80\x67\x51\x00\x80\xA5\x85"
-     "\x05\x00"
-     "\xAA\x49\x06\x00\xAB\x49\x06\x00\x6B\x51\x00\x80\x6F\x51\x00\x80\xAE\x49"
-     "\x06\x00"
-     "\xAF\x49\x06\x00\xAC\x49\x06\x00\xAD\x49\x06\x00\xA8\xB1\x07\x00\xA9\xB1"
-     "\x07\x00"
-     "\xAA\xB1\x07\x00\xAB\x09\x07\x00\xAC\x09\x07\x00\xAD\x09\x07\x00\xAE\x39"
-     "\x07\x00"
-     "\xAF\x39\x07\x00\x73\x51\x00\x80\x77\x51\x00\x80\x7B\x51\x00\x80\x7F\x51"
-     "\x00\x80"
-     "\x83\x51\x00\x80\x87\x51\x00\x80\x8B\x51\x00\x80\x8F\x51\x00\x80\xB8\x11"
-     "\x07\x00"
-     "\xB9\x1D\x07\x00\xBA\x29\x07\x00\xBB\x25\x07\x00\xBC\x3D\x07\x00\xBD\x25"
-     "\x07\x00"
-     "\xBE\x2D\x07\x00\xBF\x25\x07\x00\xB0\x49\x07\x00\xB1\x49\x07\x00\xB2\x5D"
-     "\x07\x00"
-     "\xB3\x51\x07\x00\xB4\x31\x07\x00\xB5\x31\x07\x00\xB6\x31\x07\x00\xB7\x31"
-     "\x07\x00"
-     "\xB3\x39\x06\x00\x93\x51\x00\x80\x97\x51\x00\x80\x9B\x51\x00\x80\x9F\x51"
-     "\x00\x80"
-     "\xB6\x89\x06\x00\xB5\x81\x06\x00\xA3\x51\x00\x80\xBB\x8D\x06\x00\xBA\x8D"
-     "\x06\x00"
-     "\xA7\x51\x00\x80\xAB\x51\x00\x80\xBF\xF5\x06\x00\xBE\xF5\x06\x00\xBD\x85"
-     "\x06\x00"
-     "\xBC\x89\x06\x00\xAF\x51\x00\x80\xA3\x7D\x06\x00\xB3\x51\x00\x80\xB7\x51"
-     "\x00\x80"
-     "\xA6\xCD\x06\x00\xBB\x51\x00\x80\xBF\x51\x00\x80\xA5\xC5\x06\x00\xAA\xC9"
-     "\x06\x00"
-     "\xAB\xC9\x06\x00\xC3\x51\x00\x80\xC7\x51\x00\x80\xAE\xB1\x06\x00\xAF\xB1"
-     "\x06\x00"
-     "\xAC\xCD\x06\x00\xAD\xC1\x06\x00\xA8\xD5\x01\x00\xA9\xDD\x01\x00\xAA\xE1"
-     "\x01\x00"
-     "\xAB\xE1\x01\x00\xAC\x31\x01\x00\xAD\x31\x01\x00\xAE\x31\x01\x00\xAF\x31"
-     "\x01\x00"
-     "\x80\xC9\x00\x00\x81\xC9\x00\x00\x82\x05\x00\x00\xCB\x51\x00\x80\xBE\x78"
-     "\x01\x00"
-     "\xD3\x51\x00\x80\x87\xAC\x00\x00\x86\x9C\x01\x00\xB8\xDD\x00\x00\xB9\xED"
-     "\x00\x00"
-     "\xBA\xE5\x00\x00\xBB\x8D\x00\x00\xBC\x99\x00\x00\xBD\x99\x00\x00\xBE\x8D"
-     "\x00\x00"
-     "\xBF\x85\x00\x00\xB0\x51\x01\x00\xB1\x51\x01\x00\xB2\x51\x01\x00\xB3\x51"
-     "\x01\x00"
-     "\xB4\xF5\x00\x00\xB5\xFD\x00\x00\xB6\xF5\x00\x00\xB7\xED\x00\x00\xB3\x69"
-     "\x02\x00"
-     "\xD7\x51\x00\x80\xDB\x51\x00\x80\xDF\x51\x00\x80\xE3\x51\x00\x80\xB6\xB1"
-     "\x02\x00"
-     "\xB5\xB9\x02\x00\xE7\x51\x00\x80\xBB\x95\x02\x00\xBA\x95\x02\x00\xEB\x51"
-     "\x00\x80"
-     "\xEF\x51\x00\x80\xBF\x39\x03\x00\xBE\x31\x03\x00\xBD\x85\x02\x00\xBC\x85"
-     "\x02\x00"
-     "\xF3\x51\x00\x80\xA3\x2D\x02\x00\x85\x0C\xF2\x00\xF7\x51\x00\x80\xA6\xF5"
-     "\x02\x00"
-     "\xFB\x51\x00\x80\xFF\x51\x00\x80\xA5\xFD\x02\x00\xAA\xD1\x02\x00\xAB\xD1"
-     "\x02\x00"
-     "\x03\x52\x00\x80\x07\x52\x00\x80\xAE\x75\x03\x00\xAF\x7D\x03\x00\xAC\xC1"
-     "\x02\x00"
-     "\xAD\xC1\x02\x00\xA8\x35\x03\x00\xA9\x81\x00\x00\xAA\x81\x00\x00\xAB\x81"
-     "\x00\x00"
-     "\xAC\x81\x00\x00\xAD\x89\x00\x00\xAE\xB1\x00\x00\xAF\xB1\x00\x00\x0B\x52"
-     "\x00\x80"
-     "\x0F\x52\x00\x80\x13\x52\x00\x80\x17\x52\x00\x80\x1B\x52\x00\x80\x1F\x52"
-     "\x00\x80"
-     "\x23\x52\x00\x80\x27\x52\x00\x80\xB8\x69\x00\x00\xB9\x69\x00\x00\xBA\x79"
-     "\x00\x00"
-     "\xBB\x79\x00\x00\xBC\x69\x00\x00\xBD\x69\x00\x00\xBE\xD9\x01\x00\xBF\xD9"
-     "\x01\x00"
-     "\xB0\xA1\x00\x00\xB1\xA9\x00\x00\xB2\xBD\x00\x00\xB3\xB5\x00\x00\xB4\x99"
-     "\x00\x00"
-     "\xB5\x99\x00\x00\xB6\x59\x00\x00\xB7\x59\x00\x00\x2B\x52\x00\x80\x2F\x52"
-     "\x00\x80"
-     "\x33\x52\x00\x80\x37\x52\x00\x80\x81\x1D\x00\x00\x80\x1D\x00\x00\x3B\x52"
-     "\x00\x80"
-     "\x82\x05\x00\x00\x84\xAC\x02\x00\xBE\x60\x02\x00\x87\x38\x0D\x00\x86\xAC"
-     "\x0D\x00"
-     "\x43\x52\x00\x80\x47\x52\x00\x80\x4B\x52\x00\x80\x4F\x52\x00\x80\xB3\x49"
-     "\x01\x00"
-     "\x53\x52\x00\x80\xBE\x04\x0D\x00\x57\x52\x00\x80\x5B\x52\x00\x80\xB6\x49"
-     "\x01\x00"
-     "\xB5\x49\x01\x00\x5F\x52\x00\x80\xBB\x2D\x01\x00\xBA\x25\x01\x00\x63\x52"
-     "\x00\x80"
-     "\x67\x52\x00\x80\xBF\x1D\x01\x00\xBE\x1D\x01\x00\xBD\x21\x01\x00\xBC\x29"
-     "\x01\x00"
-     "\xE1\x08\x06\x00\xE3\xB4\x07\x00\xE3\x1C\x07\x00\xE1\x3C\x04\x00\xE3\xAC"
-     "\x06\x00"
-     "\x6B\x52\x00\x80\xE1\xA8\x01\x00\x6F\x52\x00\x80\xEF\xE8\x00\x00\xBE\x28"
-     "\x0F\x00"
-     "\x73\x52\x00\x80\x77\x52\x00\x80\x7B\x52\x00\x80\xEF\x70\x06\x00\xEF\xCC"
-     "\x06\x00"
-     "\x7F\x52\x00\x80\x83\x52\x00\x80\x87\x52\x00\x80\xA3\xC1\x01\x00\xF3\x00"
-     "\x00\x00"
-     "\xA5\xC1\x01\x00\x8B\x52\x00\x80\x8F\x52\x00\x80\xA6\xC1\x01\x00\x93\x52"
-     "\x00\x80"
-     "\x97\x52\x00\x80\xAB\xA5\x01\x00\xAA\xAD\x01\x00\xAD\xA9\x01\x00\xAC\xA1"
-     "\x01\x00"
-     "\xAF\x95\x01\x00\xAE\x95\x01\x00\x3F\x52\x00\x80\x9B\x52\x00\x80\x9F\x52"
-     "\x00\x80"
-     "\xA3\x52\x00\x80\x80\x1D\x00\x00\x81\x95\x00\x00\x82\x9D\x00\x00\xCF\x51"
-     "\x00\x80"
-     "\xA8\x11\x0F\x00\xA9\x11\x0F\x00\xAA\x11\x0F\x00\xAB\x11\x0F\x00\xAC\x75"
-     "\x0F\x00"
-     "\xAD\x79\x0F\x00\xAE\x69\x0F\x00\xAF\x61\x0F\x00\xB0\x29\x0F\x00\xB1\x29"
-     "\x0F\x00"
-     "\xB2\xDD\x0F\x00\xB3\xD1\x0F\x00\xB4\xF9\x0F\x00\xB5\xF9\x0F\x00\xB6\xE9"
-     "\x0F\x00"
-     "\xB7\xE1\x0F\x00\xB8\xA1\x0F\x00\xB9\xA1\x0F\x00\xBA\xBD\x0F\x00\xBB\xB5"
-     "\x0F\x00"
-     "\xBC\x91\x0F\x00\xBD\x91\x0F\x00\xBE\x7D\x00\x00\xBF\x75\x00\x00\xA7\x52"
-     "\x00\x80"
-     "\xAB\x52\x00\x80\x84\xC8\x02\x00\xB3\xB5\x0F\x00\xAF\x52\x00\x80\xB5\xD1"
-     "\x0F\x00"
-     "\xB6\xD5\x0F\x00\xB3\x52\x00\x80\x86\x00\x03\x00\x87\x00\x03\x00\xBA\xC1"
-     "\x0F\x00"
-     "\xBB\xC1\x0F\x00\xBC\x25\x0F\x00\xBD\x29\x0F\x00\xBE\x11\x0F\x00\xBF\x11"
-     "\x0F\x00"
-     "\xA3\xF9\x0F\x00\xB7\x52\x00\x80\xBB\x52\x00\x80\xBF\x52\x00\x80\xC3\x52"
-     "\x00\x80"
-     "\xA6\x99\x0F\x00\xA5\x9D\x0F\x00\xC7\x52\x00\x80\xAB\x8D\x0F\x00\xAA\x8D"
-     "\x0F\x00"
-     "\xCB\x52\x00\x80\xCF\x52\x00\x80\xAF\x5D\x0F\x00\xAE\x5D\x0F\x00\xAD\x65"
-     "\x0F\x00"
-     "\xAC\x69\x0F\x00\xA8\x9D\x01\x00\xA9\xC1\x01\x00\xAA\xC1\x01\x00\xAB\xC1"
-     "\x01\x00"
-     "\xAC\xC1\x01\x00\xAD\xC9\x01\x00\xAE\xF1\x01\x00\xAF\xF1\x01\x00\xD3\x52"
-     "\x00\x80"
-     "\xD7\x52\x00\x80\xDB\x52\x00\x80\xDF\x52\x00\x80\xE3\x52\x00\x80\xE7\x52"
-     "\x00\x80"
-     "\xEB\x52\x00\x80\xEF\x52\x00\x80\xB8\xC1\x00\x00\xB9\xC1\x00\x00\xBA\xC1"
-     "\x00\x00"
-     "\xBB\xC1\x00\x00\xBC\xC1\x00\x00\xBD\xC1\x00\x00\xBE\xC1\x00\x00\xBF\xC1"
-     "\x00\x00"
-     "\xB0\x41\x00\x00\xB1\x41\x00\x00\xB2\x41\x00\x00\xB3\x41\x00\x00\xB4\x41"
-     "\x00\x00"
-     "\xB5\x41\x00\x00\xB6\x41\x00\x00\xB7\x41\x00\x00\xF3\x52\x00\x80\xF7\x52"
-     "\x00\x80"
-     "\xFB\x52\x00\x80\xFF\x52\x00\x80\x03\x53\x00\x80\x07\x53\x00\x80\x0B\x53"
-     "\x00\x80"
-     "\x0F\x53\x00\x80\x81\x11\x00\x00\x80\x0D\x00\x00\x13\x53\x00\x80\x82\x19"
-     "\x00\x00"
-     "\x17\x53\x00\x80\x1B\x53\x00\x80\x84\xB4\x02\x00\x23\x53\x00\x80\xB3\x09"
-     "\x01\x00"
-     "\x1F\x53\x00\x80\x86\xE8\x01\x00\x87\x60\x01\x00\x27\x53\x00\x80\xB6\x09"
-     "\x01\x00"
-     "\xB5\x09\x01\x00\x2B\x53\x00\x80\xBB\xC5\x00\x00\xBA\xC5\x00\x00\x2F\x53"
-     "\x00\x80"
-     "\x33\x53\x00\x80\xBF\xC5\x00\x00\xBE\xC5\x00\x00\xBD\xD5\x00\x00\xBC\xD5"
-     "\x00\x00"
-     "\xA3\xC9\x02\x00\x37\x53\x00\x80\x3B\x53\x00\x80\x3F\x53\x00\x80\x80\x00"
-     "\x00\x00"
-     "\xA6\xC9\x02\x00\xA5\xC9\x02\x00\x43\x53\x00\x80\xAB\x05\x03\x00\xAA\x05"
-     "\x03\x00"
-     "\x47\x53\x00\x80\x4B\x53\x00\x80\xAF\x05\x03\x00\xAE\x05\x03\x00\xAD\x15"
-     "\x03\x00"
-     "\xAC\x15\x03\x00\xE1\x90\x01\x00\x4F\x53\x00\x80\xE3\x14\x02\x00\x84\xE0"
-     "\x03\x00"
-     "\x53\x53\x00\x80\x57\x53\x00\x80\x5B\x53\x00\x80\x5F\x53\x00\x80\x63\x53"
-     "\x00\x80"
-     "\x67\x53\x00\x80\x6B\x53\x00\x80\x6F\x53\x00\x80\x73\x53\x00\x80\x77\x53"
-     "\x00\x80"
-     "\xEF\x38\x02\x00\x7B\x53\x00\x80\xA8\x3D\x03\x00\xA9\x55\x03\x00\xAA\x55"
-     "\x03\x00"
-     "\xAB\x65\x03\x00\xAC\x7D\x03\x00\xAD\x91\x00\x00\xAE\x91\x00\x00\xAF\x91"
-     "\x00\x00"
-     "\x7F\x53\x00\x80\x82\x19\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\x83\x53"
-     "\x00\x80"
-     "\x87\x53\x00\x80\xBE\xF4\x03\x00\x8B\x53\x00\x80\xB8\xA1\x00\x00\xB9\xA1"
-     "\x00\x00"
-     "\xBA\xA1\x00\x00\xBB\xA1\x00\x00\xBC\xA1\x00\x00\xBD\xA1\x00\x00\xBE\xA5"
-     "\x00\x00"
-     "\xBF\x7D\x00\x00\xB0\xFD\x00\x00\xB1\xCD\x00\x00\xB2\xC5\x00\x00\xB3\xA9"
-     "\x00\x00"
-     "\xB4\xA1\x00\x00\xB5\xA1\x00\x00\xB6\xA1\x00\x00\xB7\xA1\x00\x00\x86\xA0"
-     "\x05\x00"
-     "\x87\xC8\x02\x00\x93\x53\x00\x80\x84\x40\x05\x00\x97\x53\x00\x80\xE1\xAC"
-     "\x01\x00"
-     "\x9B\x53\x00\x80\xE3\xB4\x00\x00\x9F\x53\x00\x80\xA3\x53\x00\x80\xA7\x53"
-     "\x00\x80"
-     "\xEF\x70\x0E\x00\xAB\x53\x00\x80\xAF\x53\x00\x80\xB3\x53\x00\x80\xB7\x53"
-     "\x00\x80"
-     "\xBB\x69\x01\x00\xBA\x59\x01\x00\xBB\x53\x00\x80\xBF\x53\x00\x80\xBF\x05"
-     "\x01\x00"
-     "\xBE\x05\x01\x00\xBD\x1D\x01\x00\xBC\x71\x01\x00\xB3\xBD\x02\x00\xC3\x53"
-     "\x00\x80"
-     "\xC7\x53\x00\x80\xCB\x53\x00\x80\xCF\x53\x00\x80\xB6\x71\x01\x00\xB5\x95"
-     "\x02\x00"
-     "\xD3\x53\x00\x80\xD7\x53\x00\x80\xE1\xB0\x0E\x00\xDB\x53\x00\x80\xE3\x14"
-     "\x01\x00"
-     "\xE3\xA0\x0E\x00\xDF\x53\x00\x80\xE1\x3C\x0C\x00\xE3\x53\x00\x80\xEF\x7C"
-     "\x0E\x00"
-     "\xBE\xEC\x05\x00\x84\xE8\x04\x00\xE7\x53\x00\x80\xEB\x53\x00\x80\xEF\x53"
-     "\x00\x80"
-     "\xF3\x53\x00\x80\xEF\x0C\x0E\x00\xA3\x1D\x02\x00\x82\x2D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x80\x75\x00\x00\x84\x10\x06\x00\xA6\xD1\x01\x00\xA5\x35\x02\x00\xF7\x53"
-     "\x00\x80"
-     "\xAB\xC9\x01\x00\xAA\xF9\x01\x00\xFB\x53\x00\x80\x03\x54\x00\x80\xAF\xA5"
-     "\x01\x00"
-     "\xAE\xA5\x01\x00\xAD\xBD\x01\x00\xAC\xD1\x01\x00\xA8\x09\x07\x00\xA9\x15"
-     "\x07\x00"
-     "\xAA\x11\x07\x00\xAB\x2D\x07\x00\xAC\x35\x07\x00\xAD\x3D\x07\x00\xAE\x29"
-     "\x07\x00"
-     "\xAF\x29\x07\x00\xFF\x53\x00\x80\x07\x54\x00\x80\x0B\x54\x00\x80\x8F\x53"
-     "\x00\x80"
-     "\x86\x90\x03\x00\x87\x08\x03\x00\x0F\x54\x00\x80\x13\x54\x00\x80\xB8\x3D"
-     "\x07\x00"
-     "\xB9\xC5\x07\x00\xBA\xC1\x07\x00\xBB\xD9\x07\x00\xBC\xCD\x07\x00\xBD\xF9"
-     "\x07\x00"
-     "\xBE\xE9\x07\x00\xBF\xE9\x07\x00\xB0\x59\x07\x00\xB1\x59\x07\x00\xB2\x21"
-     "\x07\x00"
-     "\xB3\x21\x07\x00\xB4\x15\x07\x00\xB5\x1D\x07\x00\xB6\x15\x07\x00\xB7\x0D"
-     "\x07\x00"
-     "\xB3\x99\x07\x00\x17\x54\x00\x80\x1B\x54\x00\x80\x1F\x54\x00\x80\x23\x54"
-     "\x00\x80"
-     "\xB6\xBD\x07\x00\xB5\xBD\x07\x00\x27\x54\x00\x80\xBB\x9D\x07\x00\xBA\x9D"
-     "\x07\x00"
-     "\x2B\x54\x00\x80\x2F\x54\x00\x80\xBF\xD9\x07\x00\xBE\xD1\x07\x00\xBD\xF5"
-     "\x07\x00"
-     "\xBC\xF5\x07\x00\x33\x54\x00\x80\xA3\xDD\x07\x00\x37\x54\x00\x80\x3B\x54"
-     "\x00\x80"
-     "\xA6\xF9\x07\x00\x3F\x54\x00\x80\x43\x54\x00\x80\xA5\xF9\x07\x00\xAA\xD9"
-     "\x07\x00"
-     "\xAB\xD9\x07\x00\x47\x54\x00\x80\x4B\x54\x00\x80\xAE\x95\x07\x00\xAF\x9D"
-     "\x07\x00"
-     "\xAC\xB1\x07\x00\xAD\xB1\x07\x00\xA8\xED\x06\x00\xA9\x21\x01\x00\xAA\x21"
-     "\x01\x00"
-     "\xAB\x35\x01\x00\xAC\x2D\x01\x00\xAD\x91\x01\x00\xAE\x91\x01\x00\xAF\x91"
-     "\x01\x00"
-     "\x4F\x54\x00\x80\x53\x54\x00\x80\x57\x54\x00\x80\x5B\x54\x00\x80\x5F\x54"
-     "\x00\x80"
-     "\x82\xBD\x00\x00\x81\xBD\x00\x00\x80\xBD\x00\x00\xB8\x99\x01\x00\xB9\xAD"
-     "\x01\x00"
-     "\xBA\xA5\x01\x00\xBB\x6D\x00\x00\xBC\x75\x00\x00\xBD\x7D\x00\x00\xBE\x75"
-     "\x00\x00"
-     "\xBF\x6D\x00\x00\xB0\xF1\x01\x00\xB1\xF9\x01\x00\xB2\xC1\x01\x00\xB3\xC1"
-     "\x01\x00"
-     "\xB4\xB1\x01\x00\xB5\xBD\x01\x00\xB6\xB5\x01\x00\xB7\xA9\x01\x00\x63\x54"
-     "\x00\x80"
-     "\x67\x54\x00\x80\x6B\x54\x00\x80\xB3\xA5\x01\x00\xBE\xDC\x02\x00\xB5\xA9"
-     "\x01\x00"
-     "\xB6\x1D\x01\x00\x6F\x54\x00\x80\x86\x90\x01\x00\x87\xC4\x01\x00\xBA\x39"
-     "\x01\x00"
-     "\xBB\x31\x01\x00\xBC\x19\x01\x00\xBD\x19\x01\x00\xBE\x79\x01\x00\xBF\x75"
-     "\x01\x00"
-     "\xA3\xE9\x01\x00\x73\x54\x00\x80\x77\x54\x00\x80\x7B\x54\x00\x80\x7F\x54"
-     "\x00\x80"
-     "\xA6\x51\x01\x00\xA5\xE5\x01\x00\x83\x54\x00\x80\xAB\x7D\x01\x00\xAA\x75"
-     "\x01\x00"
-     "\x87\x54\x00\x80\x8B\x54\x00\x80\xAF\x39\x01\x00\xAE\x35\x01\x00\xAD\x55"
-     "\x01\x00"
-     "\xAC\x55\x01\x00\x8F\x54\x00\x80\x93\x54\x00\x80\x97\x54\x00\x80\x9B\x54"
-     "\x00\x80"
-     "\xE3\xD0\x07\x00\x9F\x54\x00\x80\xE1\xA0\x01\x00\xA3\x54\x00\x80\xEF\xF0"
-     "\x02\x00"
-     "\xA7\x54\x00\x80\xBE\xF0\x03\x00\xAB\x54\x00\x80\xAF\x54\x00\x80\xB3\x54"
-     "\x00\x80"
-     "\xB7\x54\x00\x80\xBB\x54\x00\x80\xBF\x54\x00\x80\xC3\x54\x00\x80\x86\xE8"
-     "\x1C\x00"
-     "\x87\xC0\x03\x00\xC7\x54\x00\x80\xCB\x54\x00\x80\xCF\x54\x00\x80\xD3\x54"
-     "\x00\x80"
-     "\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\xD7\x54\x00\x80\xDB\x54"
-     "\x00\x80"
-     "\xE3\x54\x00\x80\xE7\x54\x00\x80\xEB\x54\x00\x80\xAC\x5D\x03\x00\xAD\x6D"
-     "\x03\x00"
-     "\xAE\x65\x03\x00\xAF\x8D\x00\x00\xA8\x55\x03\x00\xA9\x5D\x03\x00\xAA\x55"
-     "\x03\x00"
-     "\xAB\x4D\x03\x00\x84\xAC\x03\x00\xEF\x54\x00\x80\xF3\x54\x00\x80\xF7\x54"
-     "\x00\x80"
-     "\xFB\x54\x00\x80\xFF\x54\x00\x80\x03\x55\x00\x80\x07\x55\x00\x80\xBC\x85"
-     "\x00\x00"
-     "\xBD\x95\x00\x00\xBE\x9D\x00\x00\xBF\x95\x00\x00\xB8\xE5\x00\x00\xB9\xED"
-     "\x00\x00"
-     "\xBA\xE5\x00\x00\xBB\x9D\x00\x00\xB4\xFD\x00\x00\xB5\xC1\x00\x00\xB6\xC5"
-     "\x00\x00"
-     "\xB7\xDD\x00\x00\xB0\xF5\x00\x00\xB1\xFD\x00\x00\xB2\xF5\x00\x00\xB3\xE9"
-     "\x00\x00"
-     "\x0B\x55\x00\x80\x0F\x55\x00\x80\x13\x55\x00\x80\xB3\xC1\x02\x00\x17\x55"
-     "\x00\x80"
-     "\xB5\xC9\x02\x00\xB6\xF1\x02\x00\x1B\x55\x00\x80\x1F\x55\x00\x80\x23\x55"
-     "\x00\x80"
-     "\xBA\x2D\x01\x00\xBB\x25\x01\x00\xBC\x3D\x01\x00\xBD\x21\x01\x00\xBE\x25"
-     "\x01\x00"
-     "\xBF\x19\x01\x00\xA6\xA9\x02\x00\x27\x55\x00\x80\x2B\x55\x00\x80\xA5\x91"
-     "\x02\x00"
-     "\x2F\x55\x00\x80\xA3\x99\x02\x00\x33\x55\x00\x80\x37\x55\x00\x80\xAE\x7D"
-     "\x01\x00"
-     "\xAF\x41\x01\x00\xAC\x65\x01\x00\xAD\x79\x01\x00\xAA\x75\x01\x00\xAB\x7D"
-     "\x01\x00"
-     "\x3B\x55\x00\x80\x3F\x55\x00\x80\x43\x55\x00\x80\x82\x1D\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x80\x0D\x00\x00\xEF\xCC\x06\x00\x47\x55\x00\x80\x4B\x55\x00\x80\xEF\xF4"
-     "\x00\x00"
-     "\xE3\xBC\x06\x00\xE1\x8C\x01\x00\xE1\xB0\x06\x00\xE3\x14\x06\x00\xBE\xBC"
-     "\x1D\x00"
-     "\x53\x55\x00\x80\x86\x18\x1C\x00\x87\x6C\x1D\x00\xB3\xBD\x01\x00\x57\x55"
-     "\x00\x80"
-     "\x5B\x55\x00\x80\x5F\x55\x00\x80\x63\x55\x00\x80\xB6\x79\x01\x00\xB5\x71"
-     "\x01\x00"
-     "\x67\x55\x00\x80\xBB\x5D\x01\x00\xBA\x5D\x01\x00\x6B\x55\x00\x80\x6F\x55"
-     "\x00\x80"
-     "\xBF\xE1\x00\x00\xBE\xFD\x00\x00\xBD\xFD\x00\x00\xBC\xFD\x00\x00\xA8\x75"
-     "\x1E\x00"
-     "\xA9\x7D\x1E\x00\xAA\x75\x1E\x00\xAB\xA9\x1E\x00\xAC\xB9\x1E\x00\xAD\xB9"
-     "\x1E\x00"
-     "\xAE\xA9\x1E\x00\xAF\xA9\x1E\x00\x4F\x55\x00\x80\x73\x55\x00\x80\x77\x55"
-     "\x00\x80"
-     "\x7B\x55\x00\x80\xDF\x54\x00\x80\x7F\x55\x00\x80\x83\x55\x00\x80\x87\x55"
-     "\x00\x80"
-     "\xB8\x65\x1F\x00\xB9\x01\x1F\x00\xBA\x01\x1F\x00\xBB\x1D\x1F\x00\xBC\x09"
-     "\x1F\x00"
-     "\xBD\x09\x1F\x00\xBE\x21\x1F\x00\xBF\x21\x1F\x00\xB0\xD9\x1E\x00\xB1\xD9"
-     "\x1E\x00"
-     "\xB2\x69\x1F\x00\xB3\x69\x1F\x00\xB4\x79\x1F\x00\xB5\x65\x1F\x00\xB6\x65"
-     "\x1F\x00"
-     "\xB7\x5D\x1F\x00\xA3\xFD\x1E\x00\x8B\x55\x00\x80\x8F\x55\x00\x80\x93\x55"
-     "\x00\x80"
-     "\x97\x55\x00\x80\xA6\x39\x1E\x00\xA5\x31\x1E\x00\x9B\x55\x00\x80\xAB\x1D"
-     "\x1E\x00"
-     "\xAA\x1D\x1E\x00\x9F\x55\x00\x80\xA3\x55\x00\x80\xAF\xA1\x1F\x00\xAE\xBD"
-     "\x1F\x00"
-     "\xAD\xBD\x1F\x00\xAC\xBD\x1F\x00\xA7\x55\x00\x80\xAB\x55\x00\x80\xAF\x55"
-     "\x00\x80"
-     "\xB3\x55\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39\x00\x00\xB7\x55"
-     "\x00\x80"
-     "\xBB\x55\x00\x80\x84\xA8\x03\x00\xBE\xA0\x03\x00\xBF\x55\x00\x80\x86\x50"
-     "\x00\x00"
-     "\x87\x74\x00\x00\xC3\x55\x00\x80\xC7\x55\x00\x80\xA8\xB5\x1E\x00\xA9\xC1"
-     "\x1E\x00"
-     "\xAA\xC1\x1E\x00\xAB\xC1\x1E\x00\xAC\xC1\x1E\x00\xAD\xC9\x1E\x00\xAE\xF1"
-     "\x1E\x00"
-     "\xAF\xF1\x1E\x00\xCB\x55\x00\x80\xCF\x55\x00\x80\xD3\x55\x00\x80\xD7\x55"
-     "\x00\x80"
-     "\xDB\x55\x00\x80\xDF\x55\x00\x80\xE3\x55\x00\x80\xE7\x55\x00\x80\xB8\xCD"
-     "\x01\x00"
-     "\xB9\xD5\x01\x00\xBA\xD5\x01\x00\xBB\xE5\x01\x00\xBC\xFD\x01\x00\xBD\x95"
-     "\x01\x00"
-     "\xBE\x9D\x01\x00\xBF\x95\x01\x00\xB0\x45\x01\x00\xB1\x4D\x01\x00\xB2\x45"
-     "\x01\x00"
-     "\xB3\x59\x01\x00\xB4\x49\x01\x00\xB5\x49\x01\x00\xB6\xFD\x01\x00\xB7\xF5"
-     "\x01\x00"
-     "\xB3\x49\x1E\x00\xEB\x55\x00\x80\xEF\x55\x00\x80\xF3\x55\x00\x80\xF7\x55"
-     "\x00\x80"
-     "\xB6\x49\x1E\x00\xB5\x49\x1E\x00\xFB\x55\x00\x80\xBB\x89\x01\x00\xBA\x81"
-     "\x01\x00"
-     "\xFF\x55\x00\x80\x03\x56\x00\x80\xBF\x8D\x01\x00\xBE\x8D\x01\x00\xBD\x8D"
-     "\x01\x00"
-     "\xBC\x91\x01\x00\x07\x56\x00\x80\xA3\x0D\x1E\x00\x0B\x56\x00\x80\x0F\x56"
-     "\x00\x80"
-     "\xA6\x0D\x1E\x00\x13\x56\x00\x80\x17\x56\x00\x80\xA5\x0D\x1E\x00\xAA\xC5"
-     "\x01\x00"
-     "\xAB\xCD\x01\x00\x1B\x56\x00\x80\x1F\x56\x00\x80\xAE\xC9\x01\x00\xAF\xC9"
-     "\x01\x00"
-     "\xAC\xD5\x01\x00\xAD\xC9\x01\x00\xA8\xF9\x02\x00\xA9\xF9\x02\x00\xAA\x29"
-     "\x03\x00"
-     "\xAB\x29\x03\x00\xAC\x39\x03\x00\xAD\x39\x03\x00\xAE\x29\x03\x00\xAF\x29"
-     "\x03\x00"
-     "\x23\x56\x00\x80\x27\x56\x00\x80\x2B\x56\x00\x80\xBE\x60\x02\x00\x33\x56"
-     "\x00\x80"
-     "\x37\x56\x00\x80\x3B\x56\x00\x80\x3F\x56\x00\x80\xB8\xED\x03\x00\xB9\x85"
-     "\x03\x00"
-     "\xBA\x81\x03\x00\xBB\x81\x03\x00\xBC\x85\x03\x00\xBD\x8D\x03\x00\xBE\xB1"
-     "\x03\x00"
-     "\xBF\xB1\x03\x00\xB0\x59\x03\x00\xB1\x59\x03\x00\xB2\xED\x03\x00\xB3\xE5"
-     "\x03\x00"
-     "\xB4\xFD\x03\x00\xB5\xE5\x03\x00\xB6\xE5\x03\x00\xB7\xD5\x03\x00\x80\xA1"
-     "\x00\x00"
-     "\x81\xA1\x00\x00\x82\xA1\x00\x00\xB3\xAD\x03\x00\x43\x56\x00\x80\xB5\xBD"
-     "\x03\x00"
-     "\xB6\xB5\x03\x00\x47\x56\x00\x80\x84\xE0\x02\x00\x4B\x56\x00\x80\xBA\x91"
-     "\x03\x00"
-     "\xBB\x91\x03\x00\xBC\x35\x03\x00\xBD\x3D\x03\x00\xBE\x35\x03\x00\xBF\x2D"
-     "\x03\x00"
-     "\xA3\xE1\x03\x00\x4F\x56\x00\x80\x86\x08\x04\x00\x87\x8C\x03\x00\xB4\x00"
-     "\x00\x00"
-     "\xA6\xF9\x03\x00\xA5\xF1\x03\x00\x53\x56\x00\x80\xAB\xDD\x03\x00\xAA\xDD"
-     "\x03\x00"
-     "\xBE\xE4\x04\x00\x57\x56\x00\x80\xAF\x61\x03\x00\xAE\x79\x03\x00\xAD\x71"
-     "\x03\x00"
-     "\xAC\x79\x03\x00\x5B\x56\x00\x80\x5F\x56\x00\x80\x63\x56\x00\x80\x67\x56"
-     "\x00\x80"
-     "\x6B\x56\x00\x80\x6F\x56\x00\x80\x73\x56\x00\x80\x77\x56\x00\x80\x7B\x56"
-     "\x00\x80"
-     "\x7F\x56\x00\x80\x83\x56\x00\x80\x87\x56\x00\x80\x8B\x56\x00\x80\x8F\x56"
-     "\x00\x80"
-     "\x93\x56\x00\x80\x97\x56\x00\x80\x80\x79\x00\x00\x81\x79\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\x84\x40\x02\x00\xEF\x84\x03\x00\x9B\x56\x00\x80\x87\x08\x05\x00\x86\xFC"
-     "\x04\x00"
-     "\xE3\xD0\x03\x00\xA3\x56\x00\x80\xE1\x80\x01\x00\xA7\x56\x00\x80\xAB\x56"
-     "\x00\x80"
-     "\xAF\x56\x00\x80\xB3\x56\x00\x80\xB7\x56\x00\x80\xBB\x56\x00\x80\xBF\x56"
-     "\x00\x80"
-     "\xC3\x56\x00\x80\xEF\xC4\x1E\x00\xC7\x56\x00\x80\xCB\x56\x00\x80\xCF\x56"
-     "\x00\x80"
-     "\xEF\x8C\x01\x00\x84\xA8\x05\x00\xE1\x04\x1E\x00\xD3\x56\x00\x80\xE3\x58"
-     "\x01\x00"
-     "\xD7\x56\x00\x80\xE1\xE4\x1F\x00\xDB\x56\x00\x80\xE3\x0C\x1F\x00\xB3\xF9"
-     "\x01\x00"
-     "\xDF\x56\x00\x80\xE3\x56\x00\x80\xE7\x56\x00\x80\xEB\x56\x00\x80\xB6\x49"
-     "\x01\x00"
-     "\xB5\x49\x01\x00\xEF\x56\x00\x80\xBB\x0D\x01\x00\xBA\x39\x01\x00\xF3\x56"
-     "\x00\x80"
-     "\xF7\x56\x00\x80\xBF\xFD\x00\x00\xBE\xFD\x00\x00\xBD\xFD\x00\x00\xBC\xFD"
-     "\x00\x00"
-     "\xB0\xDD\x06\x00\xB1\xE9\x06\x00\xB2\x09\x07\x00\xB3\x09\x07\x00\xB4\x19"
-     "\x07\x00"
-     "\xB5\x19\x07\x00\xB6\x09\x07\x00\xB7\x09\x07\x00\xB8\x39\x07\x00\xB9\x39"
-     "\x07\x00"
-     "\xBA\x09\x07\x00\xBB\x09\x07\x00\xBC\x19\x07\x00\xBD\x19\x07\x00\xBE\x09"
-     "\x07\x00"
-     "\xBF\x09\x07\x00\x9F\x56\x00\x80\x2F\x56\x00\x80\xFB\x56\x00\x80\xFF\x56"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\xD1\x00\x00\x82\xD1\x00\x00\x03\x57\x00\x80\xA8\x51"
-     "\x06\x00"
-     "\xA9\x51\x06\x00\xAA\x51\x06\x00\xAB\x51\x06\x00\xAC\xE1\x06\x00\xAD\xE9"
-     "\x06\x00"
-     "\xAE\xDD\x06\x00\xAF\xB9\x06\x00\xA3\xB9\x06\x00\x07\x57\x00\x80\x0B\x57"
-     "\x00\x80"
-     "\xBE\x80\x03\x00\x0F\x57\x00\x80\xA6\x09\x06\x00\xA5\x09\x06\x00\x13\x57"
-     "\x00\x80"
-     "\xAB\x4D\x06\x00\xAA\x79\x06\x00\x86\x08\x00\x00\x87\x60\x03\x00\xAF\xBD"
-     "\x07\x00"
-     "\xAE\xBD\x07\x00\xAD\xBD\x07\x00\xAC\xBD\x07\x00\x17\x57\x00\x80\xB3\xB5"
-     "\x07\x00"
-     "\x1B\x57\x00\x80\x1F\x57\x00\x80\xB6\xAD\x07\x00\x23\x57\x00\x80\x27\x57"
-     "\x00\x80"
-     "\xB5\xA5\x07\x00\xBA\x69\x07\x00\xBB\x75\x07\x00\x2B\x57\x00\x80\x2F\x57"
-     "\x00\x80"
-     "\xBE\x59\x07\x00\xBF\x45\x07\x00\xBC\x6D\x07\x00\xBD\x65\x07\x00\xA8\x65"
-     "\x06\x00"
-     "\xA9\x69\x06\x00\xAA\x79\x06\x00\xAB\x79\x06\x00\xAC\x69\x06\x00\xAD\x69"
-     "\x06\x00"
-     "\xAE\x99\x06\x00\xAF\x99\x06\x00\x33\x57\x00\x80\x37\x57\x00\x80\x3B\x57"
-     "\x00\x80"
-     "\x3F\x57\x00\x80\x43\x57\x00\x80\x47\x57\x00\x80\x4B\x57\x00\x80\x4F\x57"
-     "\x00\x80"
-     "\xB8\x89\x06\x00\xB9\x89\x06\x00\xBA\x99\x06\x00\xBB\x91\x06\x00\xBC\xB9"
-     "\x06\x00"
-     "\xBD\xB9\x06\x00\xBE\x11\x01\x00\xBF\x11\x01\x00\xB0\xE9\x06\x00\xB1\xF5"
-     "\x06\x00"
-     "\xB2\xFD\x06\x00\xB3\xF5\x06\x00\xB4\xD9\x06\x00\xB5\xC5\x06\x00\xB6\xC1"
-     "\x06\x00"
-     "\xB7\xB9\x06\x00\xA3\xF1\x06\x00\x53\x57\x00\x80\x57\x57\x00\x80\x5B\x57"
-     "\x00\x80"
-     "\x5F\x57\x00\x80\xA6\xE9\x06\x00\xA5\xE1\x06\x00\x63\x57\x00\x80\xAB\x31"
-     "\x06\x00"
-     "\xAA\x2D\x06\x00\x67\x57\x00\x80\x6B\x57\x00\x80\xAF\x01\x06\x00\xAE\x1D"
-     "\x06\x00"
-     "\xAD\x21\x06\x00\xAC\x29\x06\x00\x6F\x57\x00\x80\x82\x1D\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x73\x57\x00\x80\x77\x57\x00\x80\x7B\x57\x00\x80\xBE\x88"
-     "\x01\x00"
-     "\x84\xB8\x01\x00\x83\x57\x00\x80\x87\xE8\x00\x00\x86\x2C\x0F\x00\x87\x57"
-     "\x00\x80"
-     "\x8B\x57\x00\x80\x8F\x57\x00\x80\x93\x57\x00\x80\xA8\x9D\x02\x00\xA9\xC5"
-     "\x02\x00"
-     "\xAA\xC1\x02\x00\xAB\xC1\x02\x00\xAC\xC5\x02\x00\xAD\xCD\x02\x00\xAE\xF1"
-     "\x02\x00"
-     "\xAF\xF1\x02\x00\x97\x57\x00\x80\x9B\x57\x00\x80\x9F\x57\x00\x80\xA3\x57"
-     "\x00\x80"
-     "\xA7\x57\x00\x80\xAB\x57\x00\x80\xAF\x57\x00\x80\xB3\x57\x00\x80\xB8\xC5"
-     "\x03\x00"
-     "\xB9\xCD\x03\x00\xBA\xC5\x03\x00\xBB\xDD\x03\x00\xBC\xCD\x03\x00\xBD\xFD"
-     "\x03\x00"
-     "\xBE\xF5\x03\x00\xBF\x9D\x03\x00\xB0\x41\x03\x00\xB1\x41\x03\x00\xB2\x41"
-     "\x03\x00"
-     "\xB3\x41\x03\x00\xB4\x41\x03\x00\xB5\x41\x03\x00\xB6\x41\x03\x00\xB7\x41"
-     "\x03\x00"
-     "\xB6\x59\x02\x00\xB7\x57\x00\x80\xBB\x57\x00\x80\xB5\x59\x02\x00\xBF\x57"
-     "\x00\x80"
-     "\xB3\x5D\x02\x00\xC3\x57\x00\x80\xC7\x57\x00\x80\xBE\x91\x03\x00\xBF\x91"
-     "\x03\x00"
-     "\xBC\x81\x03\x00\xBD\x81\x03\x00\xBA\x91\x03\x00\xBB\x91\x03\x00\xCB\x57"
-     "\x00\x80"
-     "\xCF\x57\x00\x80\xA5\x15\x02\x00\xD3\x57\x00\x80\xD7\x57\x00\x80\xA6\x15"
-     "\x02\x00"
-     "\xDB\x57\x00\x80\xDF\x57\x00\x80\xA3\x11\x02\x00\xE3\x57\x00\x80\xAD\xCD"
-     "\x03\x00"
-     "\xAC\xCD\x03\x00\xAF\xDD\x03\x00\xAE\xDD\x03\x00\xE7\x57\x00\x80\xEB\x57"
-     "\x00\x80"
-     "\xAB\xDD\x03\x00\xAA\xDD\x03\x00\x80\xDD\x01\x00\x81\xF5\x01\x00\x82\xFD"
-     "\x01\x00"
-     "\x84\x8C\x02\x00\xBE\x1C\x02\x00\xE1\xA4\x01\x00\xF3\x57\x00\x80\xE3\xEC"
-     "\x00\x00"
-     "\xF7\x57\x00\x80\xBE\xEC\x0C\x00\xFB\x57\x00\x80\xEF\x08\x00\x00\xFF\x57"
-     "\x00\x80"
-     "\x03\x58\x00\x80\x07\x58\x00\x80\x0B\x58\x00\x80\x86\x60\x0D\x00\x87\xA8"
-     "\x02\x00"
-     "\x0F\x58\x00\x80\xB3\xB9\x02\x00\x13\x58\x00\x80\xB5\xA9\x02\x00\xB6\x79"
-     "\x01\x00"
-     "\x17\x58\x00\x80\x1B\x58\x00\x80\x1F\x58\x00\x80\xBA\x45\x01\x00\xBB\x5D"
-     "\x01\x00"
-     "\xBC\x45\x01\x00\xBD\x4D\x01\x00\xBE\x45\x01\x00\xBF\xF9\x01\x00\xA3\x75"
-     "\x02\x00"
-     "\x23\x58\x00\x80\x27\x58\x00\x80\x2B\x58\x00\x80\x2F\x58\x00\x80\xA6\xB5"
-     "\x01\x00"
-     "\xA5\x65\x02\x00\x33\x58\x00\x80\xAB\x91\x01\x00\xAA\x89\x01\x00\x37\x58"
-     "\x00\x80"
-     "\x3B\x58\x00\x80\xAF\x35\x01\x00\xAE\x89\x01\x00\xAD\x81\x01\x00\xAC\x89"
-     "\x01\x00"
-     "\xE1\x10\x07\x00\xE1\x80\x06\x00\xE3\x94\x06\x00\xE3\x7C\x06\x00\x3F\x58"
-     "\x00\x80"
-     "\x43\x58\x00\x80\x47\x58\x00\x80\x4B\x58\x00\x80\x4F\x58\x00\x80\x82\xB5"
-     "\x0F\x00"
-     "\x81\xA5\x0F\x00\x80\xAD\x0F\x00\x84\x3C\x0C\x00\xBF\xFC\xE2\x00\xEF\xE8"
-     "\x07\x00"
-     "\xEF\x88\x01\x00\xB0\x2D\x0E\x00\xB1\x35\x0E\x00\xB2\x3D\x0E\x00\xB3\x35"
-     "\x0E\x00"
-     "\xB4\x2D\x0E\x00\xB5\x91\x0E\x00\xB6\x91\x0E\x00\xB7\x8D\x0E\x00\xB8\xB9"
-     "\x0E\x00"
-     "\xB9\xB9\x0E\x00\xBA\x91\x0E\x00\xBB\x91\x0E\x00\xBC\xF5\x0E\x00\xBD\xFD"
-     "\x0E\x00"
-     "\xBE\xE9\x0E\x00\xBF\xE9\x0E\x00\x53\x58\x00\x80\xEF\x57\x00\x80\x7F\x57"
-     "\x00\x80"
-     "\x57\x58\x00\x80\x5B\x58\x00\x80\x5F\x58\x00\x80\x86\x78\x03\x00\x87\xEC"
-     "\x03\x00"
-     "\xA8\x85\x0D\x00\xA9\x95\x0D\x00\xAA\x9D\x0D\x00\xAB\x95\x0D\x00\xAC\x8D"
-     "\x0D\x00"
-     "\xAD\x55\x0E\x00\xAE\x5D\x0E\x00\xAF\x55\x0E\x00\xA8\x99\x0F\x00\xA9\xA5"
-     "\x0F\x00"
-     "\xAA\xA1\x0F\x00\xAB\xBD\x0F\x00\xAC\xA5\x0F\x00\xAD\xAD\x0F\x00\xAE\xA5"
-     "\x0F\x00"
-     "\xAF\xD5\x0F\x00\x63\x58\x00\x80\x67\x58\x00\x80\x6B\x58\x00\x80\x6F\x58"
-     "\x00\x80"
-     "\x73\x58\x00\x80\x77\x58\x00\x80\x7B\x58\x00\x80\x7F\x58\x00\x80\xB8\x5D"
-     "\x0F\x00"
-     "\xB9\x6D\x0F\x00\xBA\x65\x0F\x00\xBB\x1D\x0F\x00\xBC\x05\x0F\x00\xBD\x0D"
-     "\x0F\x00"
-     "\xBE\x05\x0F\x00\xBF\x3D\x0F\x00\xB0\xB1\x0F\x00\xB1\xB9\x0F\x00\xB2\x89"
-     "\x0F\x00"
-     "\xB3\x89\x0F\x00\xB4\x75\x0F\x00\xB5\x7D\x0F\x00\xB6\x75\x0F\x00\xB7\x6D"
-     "\x0F\x00"
-     "\xB3\x29\x0E\x00\x83\x58\x00\x80\x87\x58\x00\x80\x8B\x58\x00\x80\x8F\x58"
-     "\x00\x80"
-     "\xB6\x2D\x0E\x00\xB5\x2D\x0E\x00\x93\x58\x00\x80\xBB\x7D\x0E\x00\xBA\x75"
-     "\x0E\x00"
-     "\x97\x58\x00\x80\x9B\x58\x00\x80\xBF\xB9\x0E\x00\xBE\xB1\x0E\x00\xBD\x5D"
-     "\x0E\x00"
-     "\xBC\x5D\x0E\x00\x9F\x58\x00\x80\xA3\x6D\x0E\x00\xA3\x58\x00\x80\xA7\x58"
-     "\x00\x80"
-     "\xA6\x69\x0E\x00\xAB\x58\x00\x80\xAF\x58\x00\x80\xA5\x69\x0E\x00\xAA\x31"
-     "\x0E\x00"
-     "\xAB\x39\x0E\x00\xB3\x58\x00\x80\xB7\x58\x00\x80\xAE\xF5\x0E\x00\xAF\xFD"
-     "\x0E\x00"
-     "\xAC\x19\x0E\x00\xAD\x19\x0E\x00\xA8\x21\x01\x00\xA9\x21\x01\x00\xAA\x21"
-     "\x01\x00"
-     "\xAB\x21\x01\x00\xAC\x21\x01\x00\xAD\x21\x01\x00\xAE\x21\x01\x00\xAF\x21"
-     "\x01\x00"
-     "\xBB\x58\x00\x80\xBF\x58\x00\x80\xC3\x58\x00\x80\xC7\x58\x00\x80\x80\x19"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x82\x05\x00\x00\xCB\x58\x00\x80\xB8\x89\x00\x00\xB9\x89"
-     "\x00\x00"
-     "\xBA\x9D\x00\x00\xBB\x95\x00\x00\xBC\xB9\x00\x00\xBD\xB9\x00\x00\xBE\x79"
-     "\x00\x00"
-     "\xBF\x79\x00\x00\xB0\xE5\x00\x00\xB1\xED\x00\x00\xB2\xE5\x00\x00\xB3\xFD"
-     "\x00\x00"
-     "\xB4\xED\x00\x00\xB5\xDD\x00\x00\xB6\xD5\x00\x00\xB7\xBD\x00\x00\x84\xC8"
-     "\x02\x00"
-     "\xB3\xAD\x02\x00\xBE\xC4\x02\x00\xD3\x58\x00\x80\xB6\xA5\x02\x00\xD7\x58"
-     "\x00\x80"
-     "\xDB\x58\x00\x80\xB5\xAD\x02\x00\xBA\x25\x03\x00\xBB\x2D\x03\x00\x86\x08"
-     "\x04\x00"
-     "\x87\x0C\x03\x00\xBE\x2D\x03\x00\xBF\x1D\x03\x00\xBC\x35\x03\x00\xBD\x2D"
-     "\x03\x00"
-     "\xA3\xE9\x02\x00\xDF\x58\x00\x80\xE3\x58\x00\x80\xE7\x58\x00\x80\xEB\x58"
-     "\x00\x80"
-     "\xA6\xE1\x02\x00\xA5\xE9\x02\x00\xEF\x58\x00\x80\xAB\x69\x03\x00\xAA\x61"
-     "\x03\x00"
-     "\xF3\x58\x00\x80\xF7\x58\x00\x80\xAF\x59\x03\x00\xAE\x69\x03\x00\xAD\x69"
-     "\x03\x00"
-     "\xAC\x71\x03\x00\x82\x85\x1A\x00\x83\xDD\x1A\x00\xFB\x58\x00\x80\xFF\x58"
-     "\x00\x80"
-     "\x86\x8D\x1E\x00\x87\x89\x12\x00\x84\x91\x1E\x00\x85\x99\x1E\x00\x8A\xA1"
-     "\x12\x00"
-     "\x8B\x35\x12\x00\x03\x59\x00\x80\x07\x59\x00\x80\x8E\x91\x16\x00\x8F\x29"
-     "\x17\x00"
-     "\x8C\x7D\x13\x00\x8D\x91\x16\x00\x92\xF1\x0A\x00\x93\xE5\x0A\x00\x0B\x59"
-     "\x00\x80"
-     "\xEF\xC4\x0F\x00\x96\x91\x0E\x00\x97\x45\x0E\x00\x94\x35\x0A\x00\x95\x99"
-     "\x0E\x00"
-     "\x9A\xA9\x02\x00\x9B\xA1\x02\x00\x0F\x59\x00\x80\x13\x59\x00\x80\x17\x59"
-     "\x00\x80"
-     "\xE1\x28\x22\x00\x9C\xAD\x02\x00\xE3\x58\x1B\x00\x8F\x05\x1B\x00\x1B\x59"
-     "\x00\x80"
-     "\x1F\x59\x00\x80\x23\x59\x00\x80\x27\x59\x00\x80\x2B\x59\x00\x80\x2F\x59"
-     "\x00\x80"
-     "\x33\x59\x00\x80\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x37\x59"
-     "\x00\x80"
-     "\x3B\x59\x00\x80\x84\xF8\x04\x00\xBE\xF0\x04\x00\x43\x59\x00\x80\x9F\x01"
-     "\x0F\x00"
-     "\x9E\x01\x0F\x00\x9D\xA5\x0B\x00\x9C\x6D\x0B\x00\x9B\x61\x0B\x00\x9A\xC9"
-     "\x16\x00"
-     "\x99\x79\x17\x00\x98\x71\x17\x00\x97\xA5\x12\x00\x96\xF5\x13\x00\x95\x69"
-     "\x13\x00"
-     "\x94\x71\x13\x00\x93\x6D\x1F\x00\x92\x71\x1F\x00\x91\x71\x1F\x00\x90\x25"
-     "\x1B\x00"
-     "\x86\xA0\x04\x00\x87\x24\x05\x00\x47\x59\x00\x80\x4B\x59\x00\x80\x4F\x59"
-     "\x00\x80"
-     "\x53\x59\x00\x80\x57\x59\x00\x80\x5B\x59\x00\x80\x5F\x59\x00\x80\x63\x59"
-     "\x00\x80"
-     "\x67\x59\x00\x80\x6B\x59\x00\x80\xA3\xE5\x03\x00\xA2\xF1\x03\x00\xA1\xF1"
-     "\x03\x00"
-     "\xA0\x2D\x0F\x00\xE1\x98\x0E\x00\xE3\xD8\x0E\x00\xE3\x28\x0E\x00\xE1\x38"
-     "\x0C\x00"
-     "\xE1\x90\x01\x00\x6F\x59\x00\x80\xE3\xC8\x03\x00\x73\x59\x00\x80\x77\x59"
-     "\x00\x80"
-     "\x7B\x59\x00\x80\xEF\x18\x03\x00\x7F\x59\x00\x80\x83\x59\x00\x80\xEF\x1C"
-     "\x0E\x00"
-     "\xEF\x24\x0E\x00\x87\x59\x00\x80\xB2\x05\xEC\x00\xB3\x0D\xED\x00\xB0\x61"
-     "\xF1\x00"
-     "\xB1\x0D\xEC\x00\xCF\x58\x00\x80\x3F\x59\x00\x80\xB4\x01\xE8\x00\xB5\x1D"
-     "\xE8\x00"
-     "\x8B\x59\x00\x80\x8F\x59\x00\x80\x93\x59\x00\x80\x97\x59\x00\x80\x9B\x59"
-     "\x00\x80"
-     "\x9F\x59\x00\x80\xA3\x59\x00\x80\xA7\x59\x00\x80\xA2\x55\x00\x00\xA3\xF1"
-     "\x00\x00"
-     "\xAB\x59\x00\x80\xA1\x5D\x00\x00\xA6\xBD\xFC\x00\xA7\x01\xF8\x00\xA4\x01"
-     "\xFC\x00"
-     "\xA5\x1D\xFC\x00\xAA\x01\xF4\x00\xAB\x01\xF4\x00\xA8\x29\xF8\x00\xA9\xAD"
-     "\xF8\x00"
-     "\xAE\x75\xF0\x00\xAF\x69\xF0\x00\xAC\x09\xF4\x00\xAD\x01\xF0\x00\xA8\x8D"
-     "\x06\x00"
-     "\xA9\x91\x06\x00\xAA\x91\x06\x00\xAB\xC9\x06\x00\xAC\xD9\x06\x00\xAD\xD9"
-     "\x06\x00"
-     "\xAE\xC9\x06\x00\xAF\xC1\x06\x00\x80\x99\x00\x00\x81\x99\x00\x00\x82\xE1"
-     "\x00\x00"
-     "\x84\x60\x00\x00\xBE\x78\x00\x00\xAF\x59\x00\x80\x87\x38\x00\x00\x86\xBC"
-     "\x00\x00"
-     "\xB8\x2D\x06\x00\xB9\x35\x06\x00\xBA\xCD\x07\x00\xBB\xC5\x07\x00\xBC\xDD"
-     "\x07\x00"
-     "\xBD\xC5\x07\x00\xBE\xCD\x07\x00\xBF\xC5\x07\x00\xB0\x89\x06\x00\xB1\x89"
-     "\x06\x00"
-     "\xB2\x61\x06\x00\xB3\x61\x06\x00\xB4\x59\x06\x00\xB5\x59\x06\x00\xB6\x31"
-     "\x06\x00"
-     "\xB7\x15\x06\x00\xB3\x19\x07\x00\xB3\x59\x00\x80\xB7\x59\x00\x80\xBB\x59"
-     "\x00\x80"
-     "\xBF\x59\x00\x80\xB6\x59\x07\x00\xB5\x51\x07\x00\xC3\x59\x00\x80\xBB\x4D"
-     "\x07\x00"
-     "\xBA\x4D\x07\x00\xC7\x59\x00\x80\xCB\x59\x00\x80\xBF\x9D\x07\x00\xBE\x9D"
-     "\x07\x00"
-     "\xBD\xB5\x07\x00\xBC\xB5\x07\x00\xCF\x59\x00\x80\xA3\x5D\x07\x00\xD3\x59"
-     "\x00\x80"
-     "\xD7\x59\x00\x80\xA6\x1D\x07\x00\xDB\x59\x00\x80\xDF\x59\x00\x80\xA5\x15"
-     "\x07\x00"
-     "\xAA\x09\x07\x00\xAB\x09\x07\x00\xE3\x59\x00\x80\xE7\x59\x00\x80\xAE\xD9"
-     "\x07\x00"
-     "\xAF\xD9\x07\x00\xAC\xF1\x07\x00\xAD\xF1\x07\x00\xA8\xBD\x06\x00\xA9\x0D"
-     "\x06\x00"
-     "\xAA\x05\x06\x00\xAB\x1D\x06\x00\xAC\x05\x06\x00\xAD\x0D\x06\x00\xAE\x05"
-     "\x06\x00"
-     "\xAF\x3D\x06\x00\xEB\x59\x00\x80\xEF\x59\x00\x80\xF3\x59\x00\x80\xF7\x59"
-     "\x00\x80"
-     "\xFB\x59\x00\x80\xFF\x59\x00\x80\x03\x5A\x00\x80\x07\x5A\x00\x80\xB8\x31"
-     "\x06\x00"
-     "\xB9\x39\x06\x00\xBA\xE9\x01\x00\xBB\xE9\x01\x00\xBC\xF9\x01\x00\xBD\xF9"
-     "\x01\x00"
-     "\xBE\xE9\x01\x00\xBF\xE9\x01\x00\xB0\x45\x06\x00\xB1\x4D\x06\x00\xB2\x45"
-     "\x06\x00"
-     "\xB3\x39\x06\x00\xB4\x2D\x06\x00\xB5\x15\x06\x00\xB6\x1D\x06\x00\xB7\x11"
-     "\x06\x00"
-     "\xB3\x8D\x05\x00\x0B\x5A\x00\x80\x0F\x5A\x00\x80\x13\x5A\x00\x80\x17\x5A"
-     "\x00\x80"
-     "\xB6\x8D\x05\x00\xB5\x8D\x05\x00\x1B\x5A\x00\x80\xBB\x89\x05\x00\xBA\x81"
-     "\x05\x00"
-     "\xBE\x6C\x03\x00\x23\x5A\x00\x80\xBF\xD1\x05\x00\xBE\xE9\x05\x00\xBD\xED"
-     "\x05\x00"
-     "\xBC\x91\x05\x00\x84\xCC\x03\x00\xA3\xC9\x05\x00\x87\xC8\x03\x00\x86\x8C"
-     "\xFC\x00"
-     "\xA6\xC9\x05\x00\x27\x5A\x00\x80\x2B\x5A\x00\x80\xA5\xC9\x05\x00\xAA\xC5"
-     "\x05\x00"
-     "\xAB\xCD\x05\x00\x2F\x5A\x00\x80\x33\x5A\x00\x80\xAE\xAD\x05\x00\xAF\x95"
-     "\x05\x00"
-     "\xAC\xD5\x05\x00\xAD\xA9\x05\x00\x80\x8D\x00\x00\x81\x95\x00\x00\x82\x95"
-     "\x00\x00"
-     "\x37\x5A\x00\x80\x3B\x5A\x00\x80\x3F\x5A\x00\x80\x43\x5A\x00\x80\x47\x5A"
-     "\x00\x80"
-     "\xA8\xA5\x03\x00\xA9\xAD\x03\x00\xAA\xA5\x03\x00\xAB\xBD\x03\x00\xAC\xAD"
-     "\x03\x00"
-     "\xAD\x9D\x03\x00\xAE\x95\x03\x00\xAF\xFD\x03\x00\xB0\x89\x03\x00\xB1\x89"
-     "\x03\x00"
-     "\xB2\x9D\x03\x00\xB3\x95\x03\x00\xB4\xB9\x03\x00\xB5\xB9\x03\x00\xB6\x79"
-     "\x03\x00"
-     "\xB7\x79\x03\x00\xB8\x49\x03\x00\xB9\x49\x03\x00\xBA\x59\x03\x00\xBB\x59"
-     "\x03\x00"
-     "\xBC\x49\x03\x00\xBD\x49\x03\x00\xBE\xFD\x00\x00\xBF\xF5\x00\x00\x4B\x5A"
-     "\x00\x80"
-     "\x4F\x5A\x00\x80\xEF\xD8\x03\x00\x53\x5A\x00\x80\x57\x5A\x00\x80\x5B\x5A"
-     "\x00\x80"
-     "\x5F\x5A\x00\x80\x63\x5A\x00\x80\x84\x6C\xFD\x00\x67\x5A\x00\x80\x6B\x5A"
-     "\x00\x80"
-     "\x6F\x5A\x00\x80\xE1\x94\x01\x00\x73\x5A\x00\x80\xE3\x68\x03\x00\x77\x5A"
-     "\x00\x80"
-     "\xB3\x4D\x02\x00\x7B\x5A\x00\x80\x7F\x5A\x00\x80\x83\x5A\x00\x80\x87\x5A"
-     "\x00\x80"
-     "\xB6\x7D\x02\x00\xB5\x45\x02\x00\x8B\x5A\x00\x80\xBB\x2D\x02\x00\xBA\x2D"
-     "\x02\x00"
-     "\xBE\x20\xFD\x00\x93\x5A\x00\x80\xBF\x1D\x02\x00\xBE\x1D\x02\x00\xBD\x21"
-     "\x02\x00"
-     "\xBC\x29\x02\x00\x80\xCD\x00\x00\x81\xD5\x00\x00\x82\xD5\x00\x00\x97\x5A"
-     "\x00\x80"
-     "\xE1\x08\x06\x00\xE1\xA4\x07\x00\xE3\x24\x01\x00\xE3\xB8\x06\x00\x9B\x5A"
-     "\x00\x80"
-     "\x9F\x5A\x00\x80\xEF\xE0\x01\x00\xEF\x6C\x06\x00\xA3\x5A\x00\x80\xA7\x5A"
-     "\x00\x80"
-     "\xAB\x5A\x00\x80\xAF\x5A\x00\x80\x86\x60\xFC\x00\x87\x0C\xFD\x00\xA3\xC5"
-     "\x02\x00"
-     "\x84\x80\xFC\x00\xA5\xCD\x02\x00\xB3\x5A\x00\x80\xB7\x5A\x00\x80\xA6\xF5"
-     "\x02\x00"
-     "\xBB\x5A\x00\x80\xBF\x5A\x00\x80\xAB\xA5\x02\x00\xAA\xA5\x02\x00\xAD\xA9"
-     "\x02\x00"
-     "\xAC\xA1\x02\x00\xAF\x95\x02\x00\xAE\x95\x02\x00\xB0\x99\xFE\x00\xB1\x99"
-     "\xFE\x00"
-     "\xB2\x09\xFF\x00\xB3\x09\xFF\x00\xB4\x1D\xFF\x00\xB5\x05\xFF\x00\xB6\x0D"
-     "\xFF\x00"
-     "\xB7\x05\xFF\x00\xB8\x3D\xFF\x00\xB9\x01\xFF\x00\xBA\x01\xFF\x00\xBB\x01"
-     "\xFF\x00"
-     "\xBC\x05\xFF\x00\xBD\x09\xFF\x00\xBE\x7D\xFF\x00\xBF\x75\xFF\x00\x8F\x5A"
-     "\x00\x80"
-     "\xC3\x5A\x00\x80\xC7\x5A\x00\x80\xCB\x5A\x00\x80\xCF\x5A\x00\x80\xD3\x5A"
-     "\x00\x80"
-     "\x1F\x5A\x00\x80\xD7\x5A\x00\x80\xA8\x4D\xFE\x00\xA9\x31\xFE\x00\xAA\x31"
-     "\xFE\x00"
-     "\xAB\x05\xFE\x00\xAC\x1D\xFE\x00\xAD\xF9\xFE\x00\xAE\xE9\xFE\x00\xAF\xE9"
-     "\xFE\x00"
-     "\xB3\xD9\xFE\x00\xDB\x5A\x00\x80\xDF\x5A\x00\x80\xE3\x5A\x00\x80\xE7\x5A"
-     "\x00\x80"
-     "\xB6\x19\xFE\x00\xB5\xF5\xFE\x00\xEB\x5A\x00\x80\xBB\x3D\xFE\x00\xBA\x3D"
-     "\xFE\x00"
-     "\xEF\x5A\x00\x80\xF3\x5A\x00\x80\xBF\xA1\xFF\x00\xBE\xB9\xFF\x00\xBD\xBD"
-     "\xFF\x00"
-     "\xBC\xBD\xFF\x00\x82\x19\x00\x00\xA3\x9D\xFE\x00\x80\x65\x00\x00\x81\x19"
-     "\x00\x00"
-     "\xA6\x5D\xFE\x00\xF7\x5A\x00\x80\xFB\x5A\x00\x80\xA5\xB1\xFE\x00\xAA\x79"
-     "\xFE\x00"
-     "\xAB\x79\xFE\x00\xBE\xA0\x03\x00\xFF\x5A\x00\x80\xAE\xFD\xFF\x00\xAF\xE5"
-     "\xFF\x00"
-     "\xAC\xF9\xFF\x00\xAD\xF9\xFF\x00\xA8\x6D\xFE\x00\xA9\x89\xFE\x00\xAA\xA9"
-     "\xFE\x00"
-     "\xAB\xA9\xFE\x00\xAC\xA1\xFE\x00\xAD\xA1\xFE\x00\xAE\xA1\xFE\x00\xAF\xA1"
-     "\xFE\x00"
-     "\x86\x80\x01\x00\x87\x14\x01\x00\x03\x5B\x00\x80\x07\x5B\x00\x80\x0B\x5B"
-     "\x00\x80"
-     "\x0F\x5B\x00\x80\x13\x5B\x00\x80\x17\x5B\x00\x80\xB8\x15\x01\x00\xB9\x19"
-     "\x01\x00"
-     "\xBA\xCD\x01\x00\xBB\xC5\x01\x00\xBC\xDD\x01\x00\xBD\xC5\x01\x00\xBE\xC5"
-     "\x01\x00"
-     "\xBF\xF5\x01\x00\xB0\xB5\xFE\x00\xB1\xBD\xFE\x00\xB2\x89\xFE\x00\xB3\x89"
-     "\xFE\x00"
-     "\xB4\x35\x01\x00\xB5\x3D\x01\x00\xB6\x35\x01\x00\xB7\x2D\x01\x00\xB3\x21"
-     "\xFE\x00"
-     "\x1B\x5B\x00\x80\x1F\x5B\x00\x80\x23\x5B\x00\x80\x27\x5B\x00\x80\xB6\x21"
-     "\xFE\x00"
-     "\xB5\x29\xFE\x00\x2B\x5B\x00\x80\xBB\x79\xFE\x00\xBA\x71\xFE\x00\x2F\x5B"
-     "\x00\x80"
-     "\x33\x5B\x00\x80\xBF\xF9\x01\x00\xBE\xF1\x01\x00\xBD\x5D\xFE\x00\xBC\x5D"
-     "\xFE\x00"
-     "\x37\x5B\x00\x80\xA3\x65\xFE\x00\x3B\x5B\x00\x80\x3F\x5B\x00\x80\xA6\x65"
-     "\xFE\x00"
-     "\x43\x5B\x00\x80\x47\x5B\x00\x80\xA5\x6D\xFE\x00\xAA\x35\xFE\x00\xAB\x3D"
-     "\xFE\x00"
-     "\x4B\x5B\x00\x80\x4F\x5B\x00\x80\xAE\xB5\x01\x00\xAF\xBD\x01\x00\xAC\x19"
-     "\xFE\x00"
-     "\xAD\x19\xFE\x00\xA8\x29\x03\x00\xA9\x29\x03\x00\xAA\x39\x03\x00\xAB\x39"
-     "\x03\x00"
-     "\xAC\x29\x03\x00\xAD\x29\x03\x00\xAE\x9D\x03\x00\xAF\x95\x03\x00\x53\x5B"
-     "\x00\x80"
-     "\x57\x5B\x00\x80\x5B\x5B\x00\x80\x5F\x5B\x00\x80\x80\x0D\x00\x00\x81\xB5"
-     "\x00\x00"
-     "\x82\xBD\x00\x00\x63\x5B\x00\x80\xB8\x95\x03\x00\xB9\x9D\x03\x00\xBA\xA1"
-     "\x03\x00"
-     "\xBB\xA1\x03\x00\xBC\x71\x00\x00\xBD\x71\x00\x00\xBE\x71\x00\x00\xBF\x71"
-     "\x00\x00"
-     "\xB0\xED\x03\x00\xB1\xF5\x03\x00\xB2\xF5\x03\x00\xB3\xC5\x03\x00\xB4\xDD"
-     "\x03\x00"
-     "\xB5\xB5\x03\x00\xB6\xB1\x03\x00\xB7\xB1\x03\x00\x67\x5B\x00\x80\x6B\x5B"
-     "\x00\x80"
-     "\x84\x84\x02\x00\xBE\x0C\x03\x00\x73\x5B\x00\x80\x77\x5B\x00\x80\x86\x58"
-     "\x04\x00"
-     "\x87\xD0\x03\x00\x7B\x5B\x00\x80\x7F\x5B\x00\x80\x83\x5B\x00\x80\x87\x5B"
-     "\x00\x80"
-     "\x8B\x5B\x00\x80\x8F\x5B\x00\x80\x93\x5B\x00\x80\x97\x5B\x00\x80\xB3\xF1"
-     "\x03\x00"
-     "\x9B\x5B\x00\x80\x85\x70\x31\x00\x9F\x5B\x00\x80\xA3\x5B\x00\x80\xB6\xE9"
-     "\x03\x00"
-     "\xB5\xE1\x03\x00\xA7\x5B\x00\x80\xBB\x71\x03\x00\xBA\x69\x03\x00\xAB\x5B"
-     "\x00\x80"
-     "\xAF\x5B\x00\x80\xBF\x31\x03\x00\xBE\x59\x03\x00\xBD\x51\x03\x00\xBC\x61"
-     "\x03\x00"
-     "\xB3\x5B\x00\x80\xB7\x5B\x00\x80\xBB\x5B\x00\x80\xA3\x3D\x02\x00\xBF\x5B"
-     "\x00\x80"
-     "\xA5\x2D\x02\x00\xA6\x25\x02\x00\xC3\x5B\x00\x80\xC7\x5B\x00\x80\xCB\x5B"
-     "\x00\x80"
-     "\xAA\xA5\x02\x00\xAB\xBD\x02\x00\xAC\xAD\x02\x00\xAD\x9D\x02\x00\xAE\x95"
-     "\x02\x00"
-     "\xAF\xFD\x02\x00\xCF\x5B\x00\x80\x82\x19\x00\x00\x81\x11\x00\x00\x80\x19"
-     "\x00\x00"
-     "\xEF\x24\x03\x00\xD3\x5B\x00\x80\xD7\x5B\x00\x80\xDB\x5B\x00\x80\xE3\xF8"
-     "\x03\x00"
-     "\x84\x2C\x02\x00\xE1\x50\x00\x00\xBE\x60\x05\x00\x86\xD0\x04\x00\x87\xF8"
-     "\x05\x00"
-     "\xE3\x5B\x00\x80\xE7\x5B\x00\x80\xEB\x5B\x00\x80\xB3\x25\x01\x00\xEF\x5B"
-     "\x00\x80"
-     "\xF3\x5B\x00\x80\xB6\x3D\x01\x00\xF7\x5B\x00\x80\xFB\x5B\x00\x80\xB5\x35"
-     "\x01\x00"
-     "\xBA\x19\x01\x00\xBB\x19\x01\x00\xFF\x5B\x00\x80\x03\x5C\x00\x80\xBE\x61"
-     "\x01\x00"
-     "\xBF\x69\x01\x00\xBC\x7D\x01\x00\xBD\x71\x01\x00\x07\x5C\x00\x80\xE1\x1C"
-     "\xFE\x00"
-     "\x0B\x5C\x00\x80\xE3\x10\xFE\x00\x0F\x5C\x00\x80\xE3\xE4\xFE\x00\x13\x5C"
-     "\x00\x80"
-     "\xE1\x88\xFF\x00\x17\x5C\x00\x80\xEF\x74\xFE\x00\x1B\x5C\x00\x80\x1F\x5C"
-     "\x00\x80"
-     "\x23\x5C\x00\x80\x27\x5C\x00\x80\x2B\x5C\x00\x80\xEF\xEC\x01\x00\xB0\xB9"
-     "\x06\x00"
-     "\xB1\xB9\x06\x00\xB2\xCD\x06\x00\xB3\xC1\x06\x00\xB4\xC1\x06\x00\xB5\xC9"
-     "\x06\x00"
-     "\xB6\xF1\x06\x00\xB7\xF1\x06\x00\xB8\x21\x07\x00\xB9\x21\x07\x00\xBA\x21"
-     "\x07\x00"
-     "\xBB\x3D\x07\x00\xBC\x2D\x07\x00\xBD\x15\x07\x00\xBE\x1D\x07\x00\xBF\x7D"
-     "\x07\x00"
-     "\xDF\x5B\x00\x80\x2F\x5C\x00\x80\x33\x5C\x00\x80\x37\x5C\x00\x80\x3B\x5C"
-     "\x00\x80"
-     "\x6F\x5B\x00\x80\x3F\x5C\x00\x80\x43\x5C\x00\x80\xA8\xF1\x06\x00\xA9\xF1"
-     "\x06\x00"
-     "\xAA\xC9\x06\x00\xAB\xC9\x06\x00\xAC\xD9\x06\x00\xAD\xD9\x06\x00\xAE\xC9"
-     "\x06\x00"
-     "\xAF\xC9\x06\x00\xA3\x21\x06\x00\x82\x19\x00\x00\x81\x19\x00\x00\x80\xD1"
-     "\x00\x00"
-     "\x47\x5C\x00\x80\xA6\x39\x06\x00\xA5\x31\x06\x00\x4B\x5C\x00\x80\xAB\x1D"
-     "\x06\x00"
-     "\xAA\x1D\x06\x00\x84\xA0\x02\x00\xBE\xA0\x03\x00\xAF\x6D\x06\x00\xAE\x65"
-     "\x06\x00"
-     "\xAD\x75\x06\x00\xAC\x79\x06\x00\x4F\x5C\x00\x80\xB3\x45\x07\x00\x86\x68"
-     "\x00\x00"
-     "\x87\x10\x03\x00\xB6\x45\x07\x00\x53\x5C\x00\x80\x57\x5C\x00\x80\xB5\x51"
-     "\x07\x00"
-     "\xBA\x4D\x07\x00\xBB\x25\x07\x00\x5B\x5C\x00\x80\x5F\x5C\x00\x80\xBE\x29"
-     "\x07\x00"
-     "\xBF\x15\x07\x00\xBC\x31\x07\x00\xBD\x2D\x07\x00\xA8\x79\x06\x00\xA9\x79"
-     "\x06\x00"
-     "\xAA\x89\x06\x00\xAB\x85\x06\x00\xAC\x9D\x06\x00\xAD\x81\x06\x00\xAE\x81"
-     "\x06\x00"
-     "\xAF\xB5\x06\x00\x63\x5C\x00\x80\x67\x5C\x00\x80\x6B\x5C\x00\x80\x6F\x5C"
-     "\x00\x80"
-     "\x73\x5C\x00\x80\x77\x5C\x00\x80\x7B\x5C\x00\x80\x7F\x5C\x00\x80\xB8\xB1"
-     "\x06\x00"
-     "\xB9\xB1\x06\x00\xBA\xB1\x06\x00\xBB\x09\x01\x00\xBC\x19\x01\x00\xBD\x19"
-     "\x01\x00"
-     "\xBE\x09\x01\x00\xBF\x05\x01\x00\xB0\xCD\x06\x00\xB1\xA5\x06\x00\xB2\xA1"
-     "\x06\x00"
-     "\xB3\xBD\x06\x00\xB4\xA9\x06\x00\xB5\x95\x06\x00\xB6\x91\x06\x00\xB7\x91"
-     "\x06\x00"
-     "\xA3\x01\x06\x00\x83\x5C\x00\x80\x87\x5C\x00\x80\x8B\x5C\x00\x80\x8F\x5C"
-     "\x00\x80"
-     "\xA6\x01\x06\x00\xA5\x15\x06\x00\x93\x5C\x00\x80\xAB\x61\x06\x00\xAA\x09"
-     "\x06\x00"
-     "\x97\x5C\x00\x80\x9B\x5C\x00\x80\xAF\x51\x06\x00\xAE\x6D\x06\x00\xAD\x69"
-     "\x06\x00"
-     "\xAC\x75\x06\x00\x9F\x5C\x00\x80\xA3\x5C\x00\x80\xA7\x5C\x00\x80\xAB\x5C"
-     "\x00\x80"
-     "\xAF\x5C\x00\x80\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D\x00\x00\xB3\x5C"
-     "\x00\x80"
-     "\xB7\x5C\x00\x80\xBB\x5C\x00\x80\x84\x80\x01\x00\xBE\x98\x01\x00\xC3\x5C"
-     "\x00\x80"
-     "\x86\x18\x0F\x00\x87\xEC\x00\x00\xA8\xD5\x02\x00\xA9\xDD\x02\x00\xAA\xE1"
-     "\x02\x00"
-     "\xAB\xE1\x02\x00\xAC\x31\x03\x00\xAD\x31\x03\x00\xAE\x31\x03\x00\xAF\x31"
-     "\x03\x00"
-     "\xC7\x5C\x00\x80\xCB\x5C\x00\x80\xCF\x5C\x00\x80\xD3\x5C\x00\x80\xD7\x5C"
-     "\x00\x80"
-     "\xDB\x5C\x00\x80\xDF\x5C\x00\x80\xE3\x5C\x00\x80\xB8\xDD\x03\x00\xB9\xED"
-     "\x03\x00"
-     "\xBA\xE5\x03\x00\xBB\x8D\x03\x00\xBC\x95\x03\x00\xBD\x9D\x03\x00\xBE\x95"
-     "\x03\x00"
-     "\xBF\x8D\x03\x00\xB0\x51\x03\x00\xB1\x51\x03\x00\xB2\x51\x03\x00\xB3\x51"
-     "\x03\x00"
-     "\xB4\xF5\x03\x00\xB5\xFD\x03\x00\xB6\xF5\x03\x00\xB7\xED\x03\x00\xBB\xB5"
-     "\x03\x00"
-     "\xBA\xB5\x03\x00\xE7\x5C\x00\x80\xBE\x80\x0C\x00\xBF\x19\x03\x00\xBE\x11"
-     "\x03\x00"
-     "\xBD\xA5\x03\x00\xBC\xA5\x03\x00\xB3\x49\x02\x00\xEB\x5C\x00\x80\xEF\x5C"
-     "\x00\x80"
-     "\xF3\x5C\x00\x80\xF7\x5C\x00\x80\xB6\x91\x03\x00\xB5\x99\x03\x00\xFB\x5C"
-     "\x00\x80"
-     "\xAA\xF1\x03\x00\xAB\xF1\x03\x00\xFF\x5C\x00\x80\x03\x5D\x00\x80\xAE\x55"
-     "\x03\x00"
-     "\xAF\x5D\x03\x00\xAC\xE1\x03\x00\xAD\xE1\x03\x00\x07\x5D\x00\x80\xA3\x0D"
-     "\x02\x00"
-     "\x0B\x5D\x00\x80\x0F\x5D\x00\x80\xA6\xD5\x03\x00\x13\x5D\x00\x80\x17\x5D"
-     "\x00\x80"
-     "\xA5\xDD\x03\x00\x1B\x5D\x00\x80\x1F\x5D\x00\x80\x84\x44\x02\x00\x84\x04"
-     "\x0D\x00"
-     "\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\x23\x5D\x00\x80\x27\x5D"
-     "\x00\x80"
-     "\x2B\x5D\x00\x80\x33\x5D\x00\x80\x37\x5D\x00\x80\x3B\x5D\x00\x80\x3F\x5D"
-     "\x00\x80"
-     "\x43\x5D\x00\x80\x47\x5D\x00\x80\x4B\x5D\x00\x80\xE1\x94\x01\x00\x4F\x5D"
-     "\x00\x80"
-     "\xE3\xB0\x00\x00\x86\xF0\x0C\x00\x87\xA4\x02\x00\x53\x5D\x00\x80\x57\x5D"
-     "\x00\x80"
-     "\x5B\x5D\x00\x80\x5F\x5D\x00\x80\x63\x5D\x00\x80\x67\x5D\x00\x80\x6B\x5D"
-     "\x00\x80"
-     "\x6F\x5D\x00\x80\x73\x5D\x00\x80\xEF\x74\x00\x00\xB3\xA5\x01\x00\x77\x5D"
-     "\x00\x80"
-     "\x7B\x5D\x00\x80\x7F\x5D\x00\x80\x83\x5D\x00\x80\xB6\xA5\x01\x00\xB5\xB5"
-     "\x01\x00"
-     "\x87\x5D\x00\x80\xBB\x29\x01\x00\xBA\x99\x01\x00\x8B\x5D\x00\x80\x8F\x5D"
-     "\x00\x80"
-     "\xBF\x01\x01\x00\xBE\x19\x01\x00\xBD\x11\x01\x00\xBC\x19\x01\x00\x93\x5D"
-     "\x00\x80"
-     "\xE1\xAC\x06\x00\x97\x5D\x00\x80\xE3\x74\x01\x00\xE1\x50\x07\x00\x9B\x5D"
-     "\x00\x80"
-     "\xE3\xAC\x06\x00\xF7\x00\x00\x00\x86\x40\x0F\x00\x87\xD8\x0C\x00\xEF\x7C"
-     "\x06\x00"
-     "\xBE\xE0\x0C\x00\x80\x31\x00\x00\x81\x3D\x00\x00\x82\x2D\x00\x00\xEF\xA8"
-     "\x01\x00"
-     "\xB8\xCD\x0E\x00\xB9\xE5\x0E\x00\xBA\xE1\x0E\x00\xBB\xE1\x0E\x00\xBC\xE5"
-     "\x0E\x00"
-     "\xBD\xED\x0E\x00\xBE\x19\x0F\x00\xBF\x19\x0F\x00\xB0\x1D\x0E\x00\xB1\xAD"
-     "\x0E\x00"
-     "\xB2\x8D\x0E\x00\xB3\x85\x0E\x00\xB4\x9D\x0E\x00\xB5\x85\x0E\x00\xB6\x85"
-     "\x0E\x00"
-     "\xB7\xF5\x0E\x00\xA8\xB5\x0D\x00\xA9\x65\x0E\x00\xAA\x6D\x0E\x00\xAB\x65"
-     "\x0E\x00"
-     "\xAC\x7D\x0E\x00\xAD\x65\x0E\x00\xAE\x6D\x0E\x00\xAF\x65\x0E\x00\xBF\x5C"
-     "\x00\x80"
-     "\x2F\x5D\x00\x80\xA3\x5D\x00\x80\xA7\x5D\x00\x80\x9F\x5D\x00\x80\xAB\x5D"
-     "\x00\x80"
-     "\xAF\x5D\x00\x80\xB3\x5D\x00\x80\xA3\xA5\x0E\x00\xB7\x5D\x00\x80\xBB\x5D"
-     "\x00\x80"
-     "\xBF\x5D\x00\x80\xC3\x5D\x00\x80\xA6\xA5\x0E\x00\xA5\xB5\x0E\x00\xC7\x5D"
-     "\x00\x80"
-     "\xAB\x29\x0E\x00\xAA\x99\x0E\x00\xCB\x5D\x00\x80\xCF\x5D\x00\x80\xAF\x01"
-     "\x0E\x00"
-     "\xAE\x19\x0E\x00\xAD\x11\x0E\x00\xAC\x19\x0E\x00\xD3\x5D\x00\x80\xB3\xF5"
-     "\x0F\x00"
-     "\xD7\x5D\x00\x80\xDB\x5D\x00\x80\xB6\xED\x0F\x00\xDF\x5D\x00\x80\xE3\x5D"
-     "\x00\x80"
-     "\xB5\xE5\x0F\x00\xBA\xB9\x0F\x00\xBB\xB9\x0F\x00\xE7\x5D\x00\x80\xEB\x5D"
-     "\x00\x80"
-     "\xBE\x99\x0F\x00\xBF\x99\x0F\x00\xBC\x99\x0F\x00\xBD\x99\x0F\x00\xA8\x29"
-     "\x0E\x00"
-     "\xA9\x29\x0E\x00\xAA\x39\x0E\x00\xAB\x39\x0E\x00\xAC\x29\x0E\x00\xAD\x29"
-     "\x0E\x00"
-     "\xAE\x59\x0E\x00\xAF\x55\x0E\x00\xEF\x5D\x00\x80\xF3\x5D\x00\x80\xF7\x5D"
-     "\x00\x80"
-     "\xFB\x5D\x00\x80\xFF\x5D\x00\x80\x82\xB9\x00\x00\x81\xB1\x00\x00\x80\xB9"
-     "\x00\x00"
-     "\xB8\x39\x0E\x00\xB9\xC1\x0E\x00\xBA\xC1\x0E\x00\xBB\xC1\x0E\x00\xBC\xC1"
-     "\x0E\x00"
-     "\xBD\xC9\x0E\x00\xBE\xF1\x0E\x00\xBF\xF1\x0E\x00\xB0\x39\x0E\x00\xB1\x39"
-     "\x0E\x00"
-     "\xB2\x09\x0E\x00\xB3\x09\x0E\x00\xB4\x19\x0E\x00\xB5\x19\x0E\x00\xB6\x09"
-     "\x0E\x00"
-     "\xB7\x09\x0E\x00\xA3\xB1\x0E\x00\x03\x5E\x00\x80\x07\x5E\x00\x80\x0B\x5E"
-     "\x00\x80"
-     "\x84\x98\x02\x00\xA6\xA9\x0E\x00\xA5\xA1\x0E\x00\xBE\x90\x01\x00\xAB\xFD"
-     "\x0E\x00"
-     "\xAA\xFD\x0E\x00\x86\x08\x00\x00\x87\x7C\x01\x00\xAF\xDD\x0E\x00\xAE\xDD"
-     "\x0E\x00"
-     "\xAD\xDD\x0E\x00\xAC\xDD\x0E\x00\x13\x5E\x00\x80\xB3\xB1\x01\x00\x17\x5E"
-     "\x00\x80"
-     "\x1B\x5E\x00\x80\xB6\x55\x01\x00\x1F\x5E\x00\x80\x23\x5E\x00\x80\xB5\xA1"
-     "\x01\x00"
-     "\xBA\x71\x01\x00\xBB\x79\x01\x00\x27\x5E\x00\x80\x2B\x5E\x00\x80\xBE\x31"
-     "\x01\x00"
-     "\xBF\x3D\x01\x00\xBC\x51\x01\x00\xBD\x51\x01\x00\xA8\x69\x02\x00\xA9\x69"
-     "\x02\x00"
-     "\xAA\x7D\x02\x00\xAB\x75\x02\x00\xAC\x6D\x02\x00\xAD\x91\x02\x00\xAE\x91"
-     "\x02\x00"
-     "\xAF\x91\x02\x00\x2F\x5E\x00\x80\x33\x5E\x00\x80\x37\x5E\x00\x80\x3B\x5E"
-     "\x00\x80"
-     "\x3F\x5E\x00\x80\x43\x5E\x00\x80\x47\x5E\x00\x80\x4B\x5E\x00\x80\xB8\x99"
-     "\x02\x00"
-     "\xB9\xAD\x02\x00\xBA\xA5\x02\x00\xBB\x6D\x03\x00\xBC\x75\x03\x00\xBD\x7D"
-     "\x03\x00"
-     "\xBE\x75\x03\x00\xBF\x6D\x03\x00\xB0\xF1\x02\x00\xB1\xF9\x02\x00\xB2\xC1"
-     "\x02\x00"
-     "\xB3\xC1\x02\x00\xB4\xB1\x02\x00\xB5\xBD\x02\x00\xB6\xB5\x02\x00\xB7\xA9"
-     "\x02\x00"
-     "\xA3\xF5\x02\x00\x4F\x5E\x00\x80\x53\x5E\x00\x80\x57\x5E\x00\x80\x5B\x5E"
-     "\x00\x80"
-     "\xA6\x11\x02\x00\xA5\xE5\x02\x00\x5F\x5E\x00\x80\xAB\x3D\x02\x00\xAA\x35"
-     "\x02\x00"
-     "\x63\x5E\x00\x80\x67\x5E\x00\x80\xAF\x79\x02\x00\xAE\x75\x02\x00\xAD\x15"
-     "\x02\x00"
-     "\xAC\x15\x02\x00\x6B\x5E\x00\x80\x6F\x5E\x00\x80\x73\x5E\x00\x80\x77\x5E"
-     "\x00\x80"
-     "\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x7B\x5E\x00\x80\x84\xAC"
-     "\x03\x00"
-     "\xBE\x68\x04\x00\x87\xA8\x04\x00\x86\xCC\x04\x00\x83\x5E\x00\x80\x87\x5E"
-     "\x00\x80"
-     "\x8B\x5E\x00\x80\x8F\x5E\x00\x80\xA8\xC1\x02\x00\xA9\xCD\x02\x00\xAA\xC5"
-     "\x02\x00"
-     "\xAB\xD9\x02\x00\xAC\xC9\x02\x00\xAD\xFD\x02\x00\xAE\xF5\x02\x00\xAF\x3D"
-     "\x01\x00"
-     "\x93\x5E\x00\x80\x97\x5E\x00\x80\x9B\x5E\x00\x80\x9F\x5E\x00\x80\xA3\x5E"
-     "\x00\x80"
-     "\xA7\x5E\x00\x80\xAB\x5E\x00\x80\xBE\xF4\x04\x00\xB8\xC9\x01\x00\xB9\xC9"
-     "\x01\x00"
-     "\xBA\xD9\x01\x00\xBB\xD1\x01\x00\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\x99"
-     "\x01\x00"
-     "\xBF\x99\x01\x00\xB0\x45\x01\x00\xB1\x4D\x01\x00\xB2\x45\x01\x00\xB3\x5D"
-     "\x01\x00"
-     "\xB4\x45\x01\x00\xB5\x4D\x01\x00\xB6\x45\x01\x00\xB7\xF9\x01\x00\xB3\x4D"
-     "\x02\x00"
-     "\xAF\x5E\x00\x80\xB3\x5E\x00\x80\xB7\x5E\x00\x80\xBB\x5E\x00\x80\xB6\x7D"
-     "\x02\x00"
-     "\xB5\x41\x02\x00\xBF\x5E\x00\x80\xBB\x89\x01\x00\xBA\x81\x01\x00\xC3\x5E"
-     "\x00\x80"
-     "\xC7\x5E\x00\x80\xBF\x89\x01\x00\xBE\x81\x01\x00\xBD\x89\x01\x00\xBC\x91"
-     "\x01\x00"
-     "\xCB\x5E\x00\x80\xCF\x5E\x00\x80\xD3\x5E\x00\x80\xEF\xA8\x00\x00\xD7\x5E"
-     "\x00\x80"
-     "\xDB\x5E\x00\x80\xEF\xCC\x0E\x00\xEF\xA8\x0F\x00\xE1\xA8\x0E\x00\xE1\x04"
-     "\x0C\x00"
-     "\xE3\xF4\x0E\x00\xE3\xF4\x0E\x00\xDF\x5E\x00\x80\xE1\x0C\x0E\x00\xE3\x5E"
-     "\x00\x80"
-     "\xE3\x7C\x00\x00\xA3\xCD\x05\x00\x7F\x5E\x00\x80\xE7\x5E\x00\x80\xEB\x5E"
-     "\x00\x80"
-     "\x84\xDC\x03\x00\xA6\xFD\x05\x00\xA5\xC1\x05\x00\xEF\x5E\x00\x80\xAB\x09"
-     "\x06\x00"
-     "\xAA\x01\x06\x00\x86\x08\x03\x00\x87\x64\x03\x00\xAF\x09\x06\x00\xAE\x01"
-     "\x06\x00"
-     "\xAD\x09\x06\x00\xAC\x11\x06\x00\x80\xB9\x07\x00\x81\x4D\x00\x00\x82\x45"
-     "\x00\x00"
-     "\xF3\x5E\x00\x80\xF7\x5E\x00\x80\xFB\x5E\x00\x80\xFF\x5E\x00\x80\x03\x5F"
-     "\x00\x80"
-     "\x07\x5F\x00\x80\x0B\x5F\x00\x80\x0F\x5F\x00\x80\x13\x5F\x00\x80\x17\x5F"
-     "\x00\x80"
-     "\x1B\x5F\x00\x80\x1F\x5F\x00\x80\x0F\x5E\x00\x80\xA8\x71\x06\x00\xA9\x71"
-     "\x06\x00"
-     "\xAA\x71\x06\x00\xAB\x71\x06\x00\xAC\xB1\x07\x00\xAD\xB1\x07\x00\xAE\xB1"
-     "\x07\x00"
-     "\xAF\xB1\x07\x00\x23\x5F\x00\x80\x27\x5F\x00\x80\x2B\x5F\x00\x80\x2F\x5F"
-     "\x00\x80"
-     "\x33\x5F\x00\x80\x37\x5F\x00\x80\x3B\x5F\x00\x80\x3F\x5F\x00\x80\xB8\x69"
-     "\x07\x00"
-     "\xB9\x69\x07\x00\xBA\x09\x07\x00\xBB\x09\x07\x00\xBC\x19\x07\x00\xBD\x05"
-     "\x07\x00"
-     "\xBE\x0D\x07\x00\xBF\x2D\x07\x00\xB0\xD1\x07\x00\xB1\xD1\x07\x00\xB2\xD1"
-     "\x07\x00"
-     "\xB3\x69\x07\x00\xB4\x79\x07\x00\xB5\x79\x07\x00\xB6\x69\x07\x00\xB7\x61"
-     "\x07\x00"
-     "\xB3\xF9\x07\x00\x43\x5F\x00\x80\x47\x5F\x00\x80\x4B\x5F\x00\x80\x4F\x5F"
-     "\x00\x80"
-     "\xB6\x31\x06\x00\xB5\x39\x06\x00\x53\x5F\x00\x80\xBB\x15\x06\x00\xBA\x15"
-     "\x06\x00"
-     "\x57\x5F\x00\x80\x5B\x5F\x00\x80\xBF\xBD\x06\x00\xBE\xBD\x06\x00\xBD\xBD"
-     "\x06\x00"
-     "\xBC\x05\x06\x00\x82\x6D\x00\x00\xA3\xBD\x07\x00\x80\x55\x00\x00\x81\x65"
-     "\x00\x00"
-     "\xA6\x75\x06\x00\x84\x98\x01\x00\xBE\x90\x01\x00\xA5\x7D\x06\x00\xAA\x51"
-     "\x06\x00"
-     "\xAB\x51\x06\x00\x5F\x5F\x00\x80\x63\x5F\x00\x80\xAE\xF9\x06\x00\xAF\xF9"
-     "\x06\x00"
-     "\xAC\x41\x06\x00\xAD\xF9\x06\x00\xA8\xCD\x01\x00\xA9\xD5\x01\x00\xAA\xDD"
-     "\x01\x00"
-     "\xAB\xD5\x01\x00\xAC\xCD\x01\x00\xAD\xF5\x01\x00\xAE\xFD\x01\x00\xAF\xF5"
-     "\x01\x00"
-     "\x86\x00\x00\x00\x87\x38\x01\x00\x67\x5F\x00\x80\x6B\x5F\x00\x80\x6F\x5F"
-     "\x00\x80"
-     "\x73\x5F\x00\x80\x77\x5F\x00\x80\x7B\x5F\x00\x80\xB8\x71\x01\x00\xB9\x71"
-     "\x01\x00"
-     "\xBA\x71\x01\x00\xBB\x71\x01\x00\xBC\xD5\x00\x00\xBD\xDD\x00\x00\xBE\xD5"
-     "\x00\x00"
-     "\xBF\xCD\x00\x00\xB0\x8D\x01\x00\xB1\x95\x01\x00\xB2\x9D\x01\x00\xB3\x95"
-     "\x01\x00"
-     "\xB4\x51\x01\x00\xB5\x51\x01\x00\xB6\x51\x01\x00\xB7\x51\x01\x00\xB3\x55"
-     "\x02\x00"
-     "\x7F\x5F\x00\x80\x83\x5F\x00\x80\x87\x5F\x00\x80\x8B\x5F\x00\x80\xB6\x75"
-     "\x02\x00"
-     "\xB5\x45\x02\x00\x8F\x5F\x00\x80\xBB\x51\x02\x00\xBA\x49\x02\x00\x93\x5F"
-     "\x00\x80"
-     "\x97\x5F\x00\x80\xBF\x95\x02\x00\xBE\x95\x02\x00\xBD\x41\x02\x00\xBC\x49"
-     "\x02\x00"
-     "\x90\x00\x00\x00\xA3\x11\x02\x00\x9B\x5F\x00\x80\x9F\x5F\x00\x80\xA6\x31"
-     "\x02\x00"
-     "\xA3\x5F\x00\x80\xA7\x5F\x00\x80\xA5\x01\x02\x00\xAA\x0D\x02\x00\xAB\x15"
-     "\x02\x00"
-     "\xAB\x5F\x00\x80\xAF\x5F\x00\x80\xAE\xD1\x02\x00\xAF\xD1\x02\x00\xAC\x0D"
-     "\x02\x00"
-     "\xAD\x05\x02\x00\xB3\x5F\x00\x80\xB7\x5F\x00\x80\xBB\x5F\x00\x80\x85\x98"
-     "\x3E\x00"
-     "\xBF\x5F\x00\x80\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D\x00\x00\xC3\x5F"
-     "\x00\x80"
-     "\xC7\x5F\x00\x80\xCB\x5F\x00\x80\xD3\x5F\x00\x80\xBE\x7C\x1C\x00\xD7\x5F"
-     "\x00\x80"
-     "\xDB\x5F\x00\x80\xDF\x5F\x00\x80\x84\x00\x1C\x00\xE3\x5F\x00\x80\x87\x04"
-     "\x03\x00"
-     "\x86\xEC\x1C\x00\xE7\x5F\x00\x80\xEF\x9C\x02\x00\xEB\x5F\x00\x80\xEF\x5F"
-     "\x00\x80"
-     "\xF3\x5F\x00\x80\xE3\x34\x02\x00\xF7\x5F\x00\x80\xE1\x54\x01\x00\xFB\x5F"
-     "\x00\x80"
-     "\xFF\x5F\x00\x80\x03\x60\x00\x80\x07\x60\x00\x80\xA8\x6D\x02\x00\xA9\xAD"
-     "\x02\x00"
-     "\xAA\xA5\x02\x00\xAB\xBD\x02\x00\xAC\xA5\x02\x00\xAD\xAD\x02\x00\xAE\xA5"
-     "\x02\x00"
-     "\xAF\x19\x01\x00\x84\xEC\x1C\x00\x0B\x60\x00\x80\x0F\x60\x00\x80\x13\x60"
-     "\x00\x80"
-     "\x17\x60\x00\x80\x1B\x60\x00\x80\x1F\x60\x00\x80\x23\x60\x00\x80\xB8\x15"
-     "\x01\x00"
-     "\xB9\x1D\x01\x00\xBA\x21\x01\x00\xBB\x21\x01\x00\xBC\xD1\x01\x00\xBD\xD1"
-     "\x01\x00"
-     "\xBE\xD1\x01\x00\xBF\xCD\x01\x00\xB0\x69\x01\x00\xB1\x69\x01\x00\xB2\x79"
-     "\x01\x00"
-     "\xB3\x71\x01\x00\xB4\x59\x01\x00\xB5\x59\x01\x00\xB6\x31\x01\x00\xB7\x31"
-     "\x01\x00"
-     "\xE1\x74\x07\x00\x27\x60\x00\x80\xE3\x88\x01\x00\x2B\x60\x00\x80\xE3\xD4"
-     "\x00\x00"
-     "\x2F\x60\x00\x80\xE1\xA8\x01\x00\x33\x60\x00\x80\xEF\x00\x00\x00\x37\x60"
-     "\x00\x80"
-     "\x3B\x60\x00\x80\xBE\x64\x1C\x00\x43\x60\x00\x80\x47\x60\x00\x80\xEF\x7C"
-     "\x06\x00"
-     "\x4B\x60\x00\x80\xEF\xDC\x07\x00\x82\x6D\x00\x00\x81\x6D\x00\x00\x80\x51"
-     "\x00\x00"
-     "\x86\xF0\x1C\x00\x87\x78\x1D\x00\x4F\x60\x00\x80\x53\x60\x00\x80\x57\x60"
-     "\x00\x80"
-     "\x5B\x60\x00\x80\x5F\x60\x00\x80\x63\x60\x00\x80\xE3\x00\x07\x00\x67\x60"
-     "\x00\x80"
-     "\xE1\x48\x04\x00\x6B\x60\x00\x80\xCF\x5F\x00\x80\x3F\x60\x00\x80\x6F\x60"
-     "\x00\x80"
-     "\xB3\xA1\x1D\x00\x73\x60\x00\x80\xB5\xA1\x1D\x00\xB6\xA1\x1D\x00\x77\x60"
-     "\x00\x80"
-     "\x7B\x60\x00\x80\x7F\x60\x00\x80\xBA\x21\x1E\x00\xBB\x21\x1E\x00\xBC\x21"
-     "\x1E\x00"
-     "\xBD\x29\x1E\x00\xBE\x11\x1E\x00\xBF\x11\x1E\x00\xA3\xED\x1D\x00\x83\x60"
-     "\x00\x80"
-     "\x87\x60\x00\x80\x8B\x60\x00\x80\x8F\x60\x00\x80\xA6\xED\x1D\x00\xA5\xED"
-     "\x1D\x00"
-     "\x93\x60\x00\x80\xAB\x6D\x1E\x00\xAA\x6D\x1E\x00\x97\x60\x00\x80\x9B\x60"
-     "\x00\x80"
-     "\xAF\x5D\x1E\x00\xAE\x5D\x1E\x00\xAD\x65\x1E\x00\xAC\x6D\x1E\x00\xA8\x2D"
-     "\x1E\x00"
-     "\xA9\x31\x1E\x00\xAA\x31\x1E\x00\xAB\x31\x1E\x00\xAC\x55\x1E\x00\xAD\x45"
-     "\x1E\x00"
-     "\xAE\x41\x1E\x00\xAF\x41\x1E\x00\x9F\x60\x00\x80\x82\x1D\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\xA3\x60\x00\x80\xA7\x60\x00\x80\xAB\x60\x00\x80\xBE\x74"
-     "\x00\x00"
-     "\xB8\xED\x1F\x00\xB9\xF5\x1F\x00\xBA\xFD\x1F\x00\xBB\xF5\x1F\x00\xBC\xED"
-     "\x1F\x00"
-     "\xBD\x51\x1F\x00\xBE\x55\x1F\x00\xBF\x49\x1F\x00\xB0\x01\x1E\x00\xB1\x0D"
-     "\x1E\x00"
-     "\xB2\x19\x1E\x00\xB3\x19\x1E\x00\xB4\x09\x1E\x00\xB5\x09\x1E\x00\xB6\xDD"
-     "\x1F\x00"
-     "\xB7\xD5\x1F\x00\xB3\xB9\x1F\x00\x84\xC8\x01\x00\xAF\x60\x00\x80\xB3\x60"
-     "\x00\x80"
-     "\xB7\x60\x00\x80\xB6\xC5\x1F\x00\xB5\xDD\x1F\x00\xBB\x60\x00\x80\xBB\xC9"
-     "\x1F\x00"
-     "\xBA\xC5\x1F\x00\x86\x88\x00\x00\x87\xBC\x01\x00\xBF\xCD\x1F\x00\xBE\xCD"
-     "\x1F\x00"
-     "\xBD\xCD\x1F\x00\xBC\xCD\x1F\x00\xBF\x60\x00\x80\xA3\xFD\x1F\x00\xC3\x60"
-     "\x00\x80"
-     "\xC7\x60\x00\x80\xA6\x81\x1F\x00\xCB\x60\x00\x80\xCF\x60\x00\x80\xA5\x99"
-     "\x1F\x00"
-     "\xAA\x81\x1F\x00\xAB\x8D\x1F\x00\xD3\x60\x00\x80\xD7\x60\x00\x80\xAE\x89"
-     "\x1F\x00"
-     "\xAF\x89\x1F\x00\xAC\x89\x1F\x00\xAD\x89\x1F\x00\xA8\x31\x1E\x00\xA9\x31"
-     "\x1E\x00"
-     "\xAA\x35\x1E\x00\xAB\x55\x1E\x00\xAC\x4D\x1E\x00\xAD\x75\x1E\x00\xAE\x7D"
-     "\x1E\x00"
-     "\xAF\x75\x1E\x00\x84\xAC\x01\x00\xDB\x60\x00\x80\xDF\x60\x00\x80\xE3\x60"
-     "\x00\x80"
-     "\xE7\x60\x00\x80\xEB\x60\x00\x80\xEF\x60\x00\x80\xF3\x60\x00\x80\xB8\x85"
-     "\x01\x00"
-     "\xB9\x8D\x01\x00\xBA\x85\x01\x00\xBB\x9D\x01\x00\xBC\x8D\x01\x00\xBD\xBD"
-     "\x01\x00"
-     "\xBE\xB5\x01\x00\xBF\x7D\x01\x00\xB0\x0D\x1E\x00\xB1\x11\x1E\x00\xB2\x15"
-     "\x1E\x00"
-     "\xB3\xE5\x01\x00\xB4\xE1\x01\x00\xB5\xE9\x01\x00\xB6\xD9\x01\x00\xB7\xD9"
-     "\x01\x00"
-     "\xB3\xBD\x1D\x00\xF7\x60\x00\x80\xFB\x60\x00\x80\xFF\x60\x00\x80\x03\x61"
-     "\x00\x80"
-     "\xB6\xF5\x1D\x00\xB5\xC5\x1D\x00\x07\x61\x00\x80\xBB\xD5\x1D\x00\xBA\xC9"
-     "\x1D\x00"
-     "\x0B\x61\x00\x80\x0F\x61\x00\x80\xBF\x2D\x02\x00\xBE\x25\x02\x00\xBD\x31"
-     "\x02\x00"
-     "\xBC\xC1\x1D\x00\x13\x61\x00\x80\xA3\xF9\x1D\x00\x17\x61\x00\x80\x1B\x61"
-     "\x00\x80"
-     "\xA6\xB1\x1D\x00\xBE\x10\x03\x00\x23\x61\x00\x80\xA5\x81\x1D\x00\xAA\x8D"
-     "\x1D\x00"
-     "\xAB\x91\x1D\x00\x86\xE8\x04\x00\x87\xA8\x03\x00\xAE\x61\x02\x00\xAF\x69"
-     "\x02\x00"
-     "\xAC\x85\x1D\x00\xAD\x75\x02\x00\x27\x61\x00\x80\xE3\x34\x01\x00\x2B\x61"
-     "\x00\x80"
-     "\xE1\xBC\x01\x00\x80\xD1\x00\x00\x81\xDD\x00\x00\x82\xD5\x00\x00\x2F\x61"
-     "\x00\x80"
-     "\x33\x61\x00\x80\x37\x61\x00\x80\x3B\x61\x00\x80\x84\xE0\x04\x00\x3F\x61"
-     "\x00\x80"
-     "\xEF\x18\x02\x00\x43\x61\x00\x80\x47\x61\x00\x80\xB3\xDD\x03\x00\x4B\x61"
-     "\x00\x80"
-     "\x4F\x61\x00\x80\xBE\x84\x04\x00\x53\x61\x00\x80\xB6\x99\x03\x00\xB5\x91"
-     "\x03\x00"
-     "\x57\x61\x00\x80\xBB\x8D\x03\x00\xBA\x8D\x03\x00\x5B\x61\x00\x80\x5F\x61"
-     "\x00\x80"
-     "\xBF\x41\x03\x00\xBE\x59\x03\x00\xBD\x51\x03\x00\xBC\x59\x03\x00\xA8\x0D"
-     "\x02\x00"
-     "\xA9\x15\x02\x00\xAA\x15\x02\x00\xAB\x4D\x02\x00\xAC\x59\x02\x00\xAD\x59"
-     "\x02\x00"
-     "\xAE\x4D\x02\x00\xAF\x45\x02\x00\x84\x6C\x04\x00\x63\x61\x00\x80\x67\x61"
-     "\x00\x80"
-     "\x6B\x61\x00\x80\xBE\x78\x04\x00\x6F\x61\x00\x80\x73\x61\x00\x80\x77\x61"
-     "\x00\x80"
-     "\xB8\xFD\x02\x00\xB9\x41\x01\x00\xBA\x41\x01\x00\xBB\x41\x01\x00\xBC\x41"
-     "\x01\x00"
-     "\xBD\x49\x01\x00\xBE\x71\x01\x00\xBF\x71\x01\x00\xB0\x09\x02\x00\xB1\x09"
-     "\x02\x00"
-     "\xB2\xCD\x02\x00\xB3\xC5\x02\x00\xB4\xDD\x02\x00\xB5\xC5\x02\x00\xB6\xCD"
-     "\x02\x00"
-     "\xB7\xC5\x02\x00\xE1\x48\x1E\x00\xE1\x50\x1F\x00\xE3\x9C\x00\x00\xE3\xBC"
-     "\x1E\x00"
-     "\x7B\x61\x00\x80\x7F\x61\x00\x80\x83\x61\x00\x80\x87\x61\x00\x80\x8B\x61"
-     "\x00\x80"
-     "\x93\x61\x00\x80\x97\x61\x00\x80\x9B\x61\x00\x80\xE9\x00\x00\x00\x9F\x61"
-     "\x00\x80"
-     "\xEF\xFC\x01\x00\xEF\xE8\x1E\x00\x81\x5D\x00\x00\x80\x5D\x00\x00\xA3\xD5"
-     "\x02\x00"
-     "\x82\x6D\x00\x00\xA5\x99\x02\x00\xA3\x61\x00\x80\xA7\x61\x00\x80\xA6\x91"
-     "\x02\x00"
-     "\x86\x40\x04\x00\x87\xC4\x04\x00\xAB\x85\x02\x00\xAA\x85\x02\x00\xAD\x59"
-     "\x02\x00"
-     "\xAC\x51\x02\x00\xAF\x49\x02\x00\xAE\x51\x02\x00\x8F\x61\x00\x80\xAB\x61"
-     "\x00\x80"
-     "\xAF\x61\x00\x80\xB3\x61\x00\x80\xB7\x61\x00\x80\xBB\x61\x00\x80\x1F\x61"
-     "\x00\x80"
-     "\xBF\x61\x00\x80\xA8\x79\x06\x00\xA9\xA5\x06\x00\xAA\xA1\x06\x00\xAB\xB1"
-     "\x06\x00"
-     "\xAC\x91\x06\x00\xAD\x91\x06\x00\xAE\x91\x06\x00\xAF\x8D\x06\x00\xB0\xF9"
-     "\x06\x00"
-     "\xB1\xF9\x06\x00\xB2\x8D\x06\x00\xB3\x85\x06\x00\xB4\x81\x06\x00\xB5\x89"
-     "\x06\x00"
-     "\xB6\xB1\x06\x00\xB7\xB1\x06\x00\xB8\x45\x07\x00\xB9\x4D\x07\x00\xBA\x45"
-     "\x07\x00"
-     "\xBB\x59\x07\x00\xBC\x49\x07\x00\xBD\x49\x07\x00\xBE\xE9\x07\x00\xBF\xE9"
-     "\x07\x00"
-     "\xB3\x09\x06\x00\xC3\x61\x00\x80\xC7\x61\x00\x80\xCB\x61\x00\x80\xCF\x61"
-     "\x00\x80"
-     "\xB6\x31\x06\x00\xB5\x39\x06\x00\xD3\x61\x00\x80\xBB\x1D\x06\x00\xBA\x1D"
-     "\x06\x00"
-     "\xD7\x61\x00\x80\xDB\x61\x00\x80\xBF\x6D\x06\x00\xBE\x65\x06\x00\xBD\x71"
-     "\x06\x00"
-     "\xBC\x79\x06\x00\xDF\x61\x00\x80\xA3\x4D\x06\x00\xE3\x61\x00\x80\xE7\x61"
-     "\x00\x80"
-     "\xA6\x75\x06\x00\xEB\x61\x00\x80\xEF\x61\x00\x80\xA5\x7D\x06\x00\xAA\x59"
-     "\x06\x00"
-     "\xAB\x59\x06\x00\xF3\x61\x00\x80\xF7\x61\x00\x80\xAE\x21\x06\x00\xAF\x29"
-     "\x06\x00"
-     "\xAC\x3D\x06\x00\xAD\x35\x06\x00\xA8\xA5\x06\x00\xA9\xAD\x06\x00\xAA\xA5"
-     "\x06\x00"
-     "\xAB\xB9\x06\x00\xAC\xC5\x06\x00\xAD\xCD\x06\x00\xAE\x1D\x01\x00\xAF\x15"
-     "\x01\x00"
-     "\x80\xC9\x01\x00\x81\xC9\x01\x00\x82\x05\x00\x00\xFB\x61\x00\x80\x86\x00"
-     "\x0C\x00"
-     "\x87\x94\x01\x00\x03\x62\x00\x80\x07\x62\x00\x80\xB8\xDD\x01\x00\xB9\xED"
-     "\x01\x00"
-     "\xBA\xE5\x01\x00\xBB\x8D\x01\x00\xBC\x99\x01\x00\xBD\x99\x01\x00\xBE\x8D"
-     "\x01\x00"
-     "\xBF\x85\x01\x00\xB0\x6D\x01\x00\xB1\x75\x01\x00\xB2\x7D\x01\x00\xB3\x71"
-     "\x01\x00"
-     "\xB4\xF5\x01\x00\xB5\xFD\x01\x00\xB6\xF5\x01\x00\xB7\xED\x01\x00\xB3\x2D"
-     "\x06\x00"
-     "\x0B\x62\x00\x80\x0F\x62\x00\x80\x13\x62\x00\x80\xFF\x61\x00\x80\xB6\x4D"
-     "\x06\x00"
-     "\xB5\x4D\x06\x00\x17\x62\x00\x80\xBB\xB1\x01\x00\xBA\xA9\x01\x00\x1B\x62"
-     "\x00\x80"
-     "\x1F\x62\x00\x80\xBF\x39\x01\x00\xBE\x31\x01\x00\xBD\xA5\x01\x00\xBC\xA9"
-     "\x01\x00"
-     "\x23\x62\x00\x80\xA3\x69\x06\x00\x27\x62\x00\x80\x2B\x62\x00\x80\xA6\x09"
-     "\x06\x00"
-     "\x2F\x62\x00\x80\x33\x62\x00\x80\xA5\x09\x06\x00\xAA\xED\x01\x00\xAB\xF5"
-     "\x01\x00"
-     "\x37\x62\x00\x80\x3B\x62\x00\x80\xAE\x75\x01\x00\xAF\x7D\x01\x00\xAC\xED"
-     "\x01\x00"
-     "\xAD\xE1\x01\x00\x8F\x11\xEA\x00\xE1\x58\xF2\x00\x3F\x62\x00\x80\xE3\xD0"
-     "\x16\x00"
-     "\x43\x62\x00\x80\xE1\x3C\x01\x00\x47\x62\x00\x80\xE3\x04\x03\x00\x4B\x62"
-     "\x00\x80"
-     "\x4F\x62\x00\x80\x53\x62\x00\x80\xEF\xD8\x03\x00\x57\x62\x00\x80\x5B\x62"
-     "\x00\x80"
-     "\x5F\x62\x00\x80\xEF\xF8\x37\x00\x9F\x41\x06\x00\x9E\x2D\x1B\x00\x9D\x55"
-     "\x1A\x00"
-     "\x9C\x49\x1A\x00\x9B\x2D\x1A\x00\x9A\x25\x1E\x00\x99\x3D\x1E\x00\x98\x3D"
-     "\x1E\x00"
-     "\x97\xA1\x12\x00\x96\x25\x12\x00\x95\x3D\x12\x00\x94\x81\x16\x00\x93\x31"
-     "\x16\x00"
-     "\x92\x2D\x16\x00\x91\xCD\xEA\x00\x90\x69\xEA\x00\xB1\x01\x38\x00\xB0\x29"
-     "\x3E\x00"
-     "\xB3\x0D\x3A\x00\xB2\xAD\x3A\x00\xB5\xB1\x26\x00\xB4\x01\x24\x00\x63\x62"
-     "\x00\x80"
-     "\x67\x62\x00\x80\x80\x39\x00\x00\x81\x39\x00\x00\x82\x05\x00\x00\x6B\x62"
-     "\x00\x80"
-     "\xA3\xB1\x02\x00\xA2\xAD\x02\x00\xA1\xB1\x07\x00\xA0\xB9\x06\x00\xA1\x51"
-     "\x03\x00"
-     "\x73\x62\x00\x80\xA3\x4D\x0E\x00\xA2\x4D\x0E\x00\xA5\xD5\x0A\x00\xA4\xA1"
-     "\x0F\x00"
-     "\xA7\xF5\x0B\x00\xA6\xD5\x0A\x00\xA9\x6D\x36\x00\xA8\x65\x36\x00\xAB\x01"
-     "\x30\x00"
-     "\xAA\x01\x36\x00\xAD\xE9\x33\x00\xAC\xD5\x32\x00\xAF\xD9\x3E\x00\xAE\x01"
-     "\x3C\x00"
-     "\x82\x11\xEB\x00\x83\x11\xEB\x00\x84\x04\x02\x00\xBE\x80\x0D\x00\x86\x51"
-     "\x17\x00"
-     "\x87\xF9\x17\x00\x84\xAD\xEB\x00\x85\x45\x17\x00\x8A\x75\x13\x00\x8B\x65"
-     "\x13\x00"
-     "\x86\x28\x0D\x00\x87\x6C\x0D\x00\x8E\x55\x1F\x00\x8F\x45\x1F\x00\x8C\xF9"
-     "\x13\x00"
-     "\x8D\x5D\x1F\x00\x92\x3D\x1B\x00\x93\x41\x1B\x00\x84\xC8\x0D\x00\xBE\xC0"
-     "\x0C\x00"
-     "\x96\x51\x07\x00\x97\x51\x07\x00\x94\x45\x1B\x00\x95\x35\x1A\x00\x9A\x69"
-     "\x06\x00"
-     "\x9B\x7D\x03\x00\x77\x62\x00\x80\x7B\x62\x00\x80\x7F\x62\x00\x80\x83\x62"
-     "\x00\x80"
-     "\x9C\x79\x03\x00\x87\x62\x00\x80\x8B\x62\x00\x80\xE1\x7C\x06\x00\x8F\x62"
-     "\x00\x80"
-     "\xE3\xBC\x06\x00\x93\x62\x00\x80\xE3\x28\x07\x00\x97\x62\x00\x80\xE1\xEC"
-     "\x07\x00"
-     "\x84\x6C\x0F\x00\xEF\x90\x07\x00\x9B\x62\x00\x80\x9F\x62\x00\x80\xA3\x62"
-     "\x00\x80"
-     "\xA7\x62\x00\x80\xAB\x62\x00\x80\xEF\x74\x06\x00\xB3\x45\x02\x00\xAF\x62"
-     "\x00\x80"
-     "\xB3\x62\x00\x80\xB7\x62\x00\x80\xBB\x62\x00\x80\xB6\x49\x02\x00\xB5\x55"
-     "\x02\x00"
-     "\xBF\x62\x00\x80\xBB\x29\x02\x00\xBA\x25\x02\x00\xC3\x62\x00\x80\xC7\x62"
-     "\x00\x80"
-     "\xBF\x19\x02\x00\xBE\x2D\x02\x00\xBD\x2D\x02\x00\xBC\x31\x02\x00\xA3\x85"
-     "\x0D\x00"
-     "\xCB\x62\x00\x80\xCF\x62\x00\x80\xD3\x62\x00\x80\xD7\x62\x00\x80\xA6\x89"
-     "\x0D\x00"
-     "\xA5\x95\x0D\x00\xDB\x62\x00\x80\xAB\xE9\x0D\x00\xAA\xE5\x0D\x00\x86\x08"
-     "\x03\x00"
-     "\x87\x4C\x03\x00\xAF\xD9\x0D\x00\xAE\xED\x0D\x00\xAD\xED\x0D\x00\xAC\xF1"
-     "\x0D\x00"
-     "\x80\x4D\x00\x00\x81\x55\x00\x00\x82\x55\x00\x00\xB3\x21\x0F\x00\xE3\x62"
-     "\x00\x80"
-     "\xB5\x21\x0F\x00\xB6\x25\x0F\x00\xE7\x62\x00\x80\xEB\x62\x00\x80\xEF\x62"
-     "\x00\x80"
-     "\xBA\xF9\x0F\x00\xBB\xF5\x0F\x00\xBC\xD9\x0F\x00\xBD\xD9\x0F\x00\xBE\xC9"
-     "\x0F\x00"
-     "\xBF\xC9\x0F\x00\x6F\x62\x00\x80\xDF\x62\x00\x80\xF3\x62\x00\x80\xF7\x62"
-     "\x00\x80"
-     "\xFB\x62\x00\x80\xFF\x62\x00\x80\x03\x63\x00\x80\x07\x63\x00\x80\xA8\x65"
-     "\x0E\x00"
-     "\xA9\xCD\x0F\x00\xAA\xC5\x0F\x00\xAB\xDD\x0F\x00\xAC\xC5\x0F\x00\xAD\xC9"
-     "\x0F\x00"
-     "\xAE\x39\x0F\x00\xAF\x35\x0F\x00\xB0\x51\x0F\x00\xB1\x5D\x0F\x00\xB2\x61"
-     "\x0F\x00"
-     "\xB3\x61\x0F\x00\xB4\x61\x0F\x00\xB5\x61\x0F\x00\xB6\x61\x0F\x00\xB7\x19"
-     "\x0F\x00"
-     "\xB8\x29\x0F\x00\xB9\x29\x0F\x00\xBA\x39\x0F\x00\xBB\x31\x0F\x00\xBC\x19"
-     "\x0F\x00"
-     "\xBD\x19\x0F\x00\xBE\xFD\x00\x00\xBF\xF5\x00\x00\xA3\x6D\x0E\x00\x0B\x63"
-     "\x00\x80"
-     "\x0F\x63\x00\x80\x13\x63\x00\x80\x17\x63\x00\x80\xA6\x69\x0E\x00\xA5\x6D"
-     "\x0E\x00"
-     "\x1B\x63\x00\x80\xAB\xB9\x0E\x00\xAA\xB5\x0E\x00\x1F\x63\x00\x80\x23\x63"
-     "\x00\x80"
-     "\xAF\x85\x0E\x00\xAE\x85\x0E\x00\xAD\x95\x0E\x00\xAC\x95\x0E\x00\x27\x63"
-     "\x00\x80"
-     "\x2B\x63\x00\x80\x2F\x63\x00\x80\x33\x63\x00\x80\x37\x63\x00\x80\x3B\x63"
-     "\x00\x80"
-     "\x3F\x63\x00\x80\x43\x63\x00\x80\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19"
-     "\x00\x00"
-     "\x47\x63\x00\x80\x4B\x63\x00\x80\x4F\x63\x00\x80\xBE\xB0\x01\x00\x53\x63"
-     "\x00\x80"
-     "\xA8\xD1\x01\x00\xA9\xD9\x01\x00\xAA\xE9\x01\x00\xAB\xE9\x01\x00\xAC\x19"
-     "\x01\x00"
-     "\xAD\x19\x01\x00\xAE\x09\x01\x00\xAF\x09\x01\x00\x86\x00\x04\x00\x87\x04"
-     "\x01\x00"
-     "\x57\x63\x00\x80\x5B\x63\x00\x80\x5F\x63\x00\x80\x63\x63\x00\x80\x67\x63"
-     "\x00\x80"
-     "\x6B\x63\x00\x80\xB8\xED\x00\x00\xB9\x85\x00\x00\xBA\x8D\x00\x00\xBB\x85"
-     "\x00\x00"
-     "\xBC\x9D\x00\x00\xBD\x85\x00\x00\xBE\x8D\x00\x00\xBF\x85\x00\x00\xB0\x79"
-     "\x01\x00"
-     "\xB1\x79\x01\x00\xB2\xED\x00\x00\xB3\xE5\x00\x00\xB4\xFD\x00\x00\xB5\xE5"
-     "\x00\x00"
-     "\xB6\xE5\x00\x00\xB7\xD5\x00\x00\xB3\x61\x02\x00\x6F\x63\x00\x80\x73\x63"
-     "\x00\x80"
-     "\x77\x63\x00\x80\x7B\x63\x00\x80\xB6\x99\x02\x00\xB5\x91\x02\x00\x7F\x63"
-     "\x00\x80"
-     "\xBB\xBD\x02\x00\xBA\xBD\x02\x00\x83\x63\x00\x80\x87\x63\x00\x80\xBF\x21"
-     "\x03\x00"
-     "\xBE\x39\x03\x00\xBD\x31\x03\x00\xBC\x39\x03\x00\xBE\xCC\x04\x00\xA3\x25"
-     "\x02\x00"
-     "\x8B\x63\x00\x80\x8F\x63\x00\x80\xA6\xDD\x02\x00\x93\x63\x00\x80\x97\x63"
-     "\x00\x80"
-     "\xA5\xD5\x02\x00\xAA\xF9\x02\x00\xAB\xF9\x02\x00\x9B\x63\x00\x80\x9F\x63"
-     "\x00\x80"
-     "\xAE\x7D\x03\x00\xAF\x65\x03\x00\xAC\x7D\x03\x00\xAD\x75\x03\x00\xBE\x0C"
-     "\x05\x00"
-     "\xA3\x63\x00\x80\xA7\x63\x00\x80\xBE\x00\x06\x00\xAB\x63\x00\x80\xAF\x63"
-     "\x00\x80"
-     "\xB3\x63\x00\x80\xB7\x63\x00\x80\x80\x39\x00\x00\x81\x39\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xBB\x63\x00\x80\xC3\x63\x00\x80\x84\x38\x03\x00\xC7\x63\x00\x80\xCB\x63"
-     "\x00\x80"
-     "\x86\xE0\x04\x00\x87\x64\x03\x00\xEF\x70\x02\x00\xCF\x63\x00\x80\xD3\x63"
-     "\x00\x80"
-     "\xD7\x63\x00\x80\xDB\x63\x00\x80\xDF\x63\x00\x80\xE3\x63\x00\x80\xE7\x63"
-     "\x00\x80"
-     "\xEB\x63\x00\x80\xEF\x63\x00\x80\xE1\x48\x01\x00\xF3\x63\x00\x80\xE3\x5C"
-     "\x02\x00"
-     "\xF7\x63\x00\x80\xE1\x30\x01\x00\xE1\x34\x0E\x00\xE3\xB8\x0E\x00\xE3\x54"
-     "\x0E\x00"
-     "\xFB\x63\x00\x80\xFF\x63\x00\x80\x03\x64\x00\x80\x07\x64\x00\x80\xBE\x6C"
-     "\x06\x00"
-     "\x0B\x64\x00\x80\x0F\x64\x00\x80\x13\x64\x00\x80\x17\x64\x00\x80\x1B\x64"
-     "\x00\x80"
-     "\xEF\xB8\x00\x00\xEF\x88\x0E\x00\xB3\xED\x02\x00\x1F\x64\x00\x80\x23\x64"
-     "\x00\x80"
-     "\x27\x64\x00\x80\x2B\x64\x00\x80\xB6\xED\x02\x00\xB5\xED\x02\x00\x33\x64"
-     "\x00\x80"
-     "\xBB\x6D\x01\x00\xBA\x6D\x01\x00\x84\xE0\x05\x00\x37\x64\x00\x80\xBF\x5D"
-     "\x01\x00"
-     "\xBE\x5D\x01\x00\xBD\x65\x01\x00\xBC\x6D\x01\x00\xAC\x0D\x0E\x00\xAD\x19"
-     "\x0E\x00"
-     "\xAE\x21\x0E\x00\xAF\x21\x0E\x00\xA8\xA9\x0E\x00\xA9\xA9\x0E\x00\xAA\x1D"
-     "\x0E\x00"
-     "\xAB\x15\x0E\x00\x80\xE9\x00\x00\x81\xE9\x00\x00\x82\xF9\x00\x00\x3B\x64"
-     "\x00\x80"
-     "\x3F\x64\x00\x80\x43\x64\x00\x80\x87\xB4\x04\x00\x86\x9C\x04\x00\xBC\xE9"
-     "\x0D\x00"
-     "\xBD\xF5\x0D\x00\xBE\xFD\x0D\x00\xBF\xF5\x0D\x00\xB8\xC1\x0D\x00\xB9\xCD"
-     "\x0D\x00"
-     "\xBA\xF9\x0D\x00\xBB\xF9\x0D\x00\xB4\x39\x0E\x00\xB5\x0D\x0E\x00\xB6\x05"
-     "\x0E\x00"
-     "\xB7\xE5\x0D\x00\xB0\x31\x0E\x00\xB1\x3D\x0E\x00\xB2\x35\x0E\x00\xB3\x29"
-     "\x0E\x00"
-     "\xA3\xAD\x05\x00\xBF\x63\x00\x80\x2F\x64\x00\x80\x47\x64\x00\x80\x4B\x64"
-     "\x00\x80"
-     "\xA6\xAD\x05\x00\xA5\xAD\x05\x00\x4F\x64\x00\x80\xAB\x2D\x06\x00\xAA\x2D"
-     "\x06\x00"
-     "\x53\x64\x00\x80\x57\x64\x00\x80\xAF\x1D\x06\x00\xAE\x1D\x06\x00\xAD\x25"
-     "\x06\x00"
-     "\xAC\x2D\x06\x00\x5B\x64\x00\x80\xB3\xF9\x06\x00\x5F\x64\x00\x80\x63\x64"
-     "\x00\x80"
-     "\xB6\x85\x06\x00\x67\x64\x00\x80\x6B\x64\x00\x80\xB5\x8D\x06\x00\xBA\x81"
-     "\x06\x00"
-     "\xBB\x81\x06\x00\x6F\x64\x00\x80\x73\x64\x00\x80\xBE\x85\x06\x00\xBF\x89"
-     "\x06\x00"
-     "\xBC\x91\x06\x00\xBD\x91\x06\x00\xA8\x31\x06\x00\xA9\x3D\x06\x00\xAA\x35"
-     "\x06\x00"
-     "\xAB\x41\x06\x00\xAC\x41\x06\x00\xAD\x41\x06\x00\xAE\x41\x06\x00\xAF\x41"
-     "\x06\x00"
-     "\x77\x64\x00\x80\x7B\x64\x00\x80\x7F\x64\x00\x80\x83\x64\x00\x80\x87\x64"
-     "\x00\x80"
-     "\x8B\x64\x00\x80\x8F\x64\x00\x80\x93\x64\x00\x80\xB8\xD9\x07\x00\xB9\xD9"
-     "\x07\x00"
-     "\xBA\xED\x07\x00\xBB\xE5\x07\x00\xBC\xFD\x07\x00\xBD\xE5\x07\x00\xBE\xED"
-     "\x07\x00"
-     "\xBF\x99\x07\x00\xB0\x01\x06\x00\xB1\x01\x06\x00\xB2\x01\x06\x00\xB3\x01"
-     "\x06\x00"
-     "\xB4\x05\x06\x00\xB5\x09\x06\x00\xB6\xE9\x07\x00\xB7\xE9\x07\x00\xA3\xBD"
-     "\x07\x00"
-     "\x82\x19\x00\x00\x81\x19\x00\x00\x80\xB1\x01\x00\x97\x64\x00\x80\xA6\xC1"
-     "\x07\x00"
-     "\xA5\xC9\x07\x00\x9B\x64\x00\x80\xAB\xC5\x07\x00\xAA\xC5\x07\x00\x84\xA0"
-     "\x02\x00"
-     "\x9F\x64\x00\x80\xAF\xCD\x07\x00\xAE\xC1\x07\x00\xAD\xD5\x07\x00\xAC\xD5"
-     "\x07\x00"
-     "\xA3\x64\x00\x80\xB3\xE9\x06\x00\x86\x00\x3C\x00\x87\x10\x01\x00\xB6\x5D"
-     "\x01\x00"
-     "\xA7\x64\x00\x80\xAB\x64\x00\x80\xB5\xE9\x06\x00\xBA\x49\x01\x00\xBB\x49"
-     "\x01\x00"
-     "\xAF\x64\x00\x80\xB3\x64\x00\x80\xBE\x3D\x01\x00\xBF\x21\x01\x00\xBC\x39"
-     "\x01\x00"
-     "\xBD\x35\x01\x00\xA8\xDD\x06\x00\xA9\xE9\x06\x00\xAA\x25\x06\x00\xAB\x3D"
-     "\x06\x00"
-     "\xAC\x25\x06\x00\xAD\x99\x01\x00\xAE\x89\x01\x00\xAF\x81\x01\x00\x84\xA0"
-     "\x01\x00"
-     "\xB7\x64\x00\x80\xBB\x64\x00\x80\xBF\x64\x00\x80\xC3\x64\x00\x80\xC7\x64"
-     "\x00\x80"
-     "\xCB\x64\x00\x80\xCF\x64\x00\x80\xB8\xAD\x01\x00\xB9\x65\x01\x00\xBA\x6D"
-     "\x01\x00"
-     "\xBB\x65\x01\x00\xBC\x7D\x01\x00\xBD\x65\x01\x00\xBE\x6D\x01\x00\xBF\x65"
-     "\x01\x00"
-     "\xB0\xC9\x01\x00\xB1\xC9\x01\x00\xB2\xA9\x01\x00\xB3\xA5\x01\x00\xB4\xBD"
-     "\x01\x00"
-     "\xB5\xA1\x01\x00\xB6\xA1\x01\x00\xB7\x95\x01\x00\xA3\xAD\x05\x00\xD3\x64"
-     "\x00\x80"
-     "\xD7\x64\x00\x80\xDB\x64\x00\x80\xDF\x64\x00\x80\xA6\x19\x02\x00\xA5\xAD"
-     "\x05\x00"
-     "\xE3\x64\x00\x80\xAB\x0D\x02\x00\xAA\x0D\x02\x00\xE7\x64\x00\x80\xEB\x64"
-     "\x00\x80"
-     "\xAF\x65\x02\x00\xAE\x79\x02\x00\xAD\x71\x02\x00\xAC\x7D\x02\x00\xEF\x64"
-     "\x00\x80"
-     "\xF3\x64\x00\x80\xF7\x64\x00\x80\xFB\x64\x00\x80\xFF\x64\x00\x80\x82\x3D"
-     "\x00\x00"
-     "\x81\x3D\x00\x00\x80\x3D\x00\x00\x03\x65\x00\x80\x07\x65\x00\x80\x0B\x65"
-     "\x00\x80"
-     "\x13\x65\x00\x80\x17\x65\x00\x80\x1B\x65\x00\x80\x86\xD8\x3C\x00\x87\x4C"
-     "\x03\x00"
-     "\x1F\x65\x00\x80\x23\x65\x00\x80\x84\xC0\x3C\x00\xBE\xC0\x3C\x00\xEF\x48"
-     "\x07\x00"
-     "\x27\x65\x00\x80\x2B\x65\x00\x80\x2F\x65\x00\x80\xE3\x10\x01\x00\x33\x65"
-     "\x00\x80"
-     "\xE1\xF8\x01\x00\x37\x65\x00\x80\x3B\x65\x00\x80\x3F\x65\x00\x80\x43\x65"
-     "\x00\x80"
-     "\x47\x65\x00\x80\x4B\x65\x00\x80\x4F\x65\x00\x80\x53\x65\x00\x80\xB3\x95"
-     "\x03\x00"
-     "\x57\x65\x00\x80\xB5\xB9\x03\x00\xB6\x7D\x03\x00\x5B\x65\x00\x80\x5F\x65"
-     "\x00\x80"
-     "\x63\x65\x00\x80\xBA\x59\x03\x00\xBB\x59\x03\x00\xBC\x49\x03\x00\xBD\x49"
-     "\x03\x00"
-     "\xBE\xFD\x00\x00\xBF\xF5\x00\x00\xA8\x7D\x06\x00\xA9\xAD\x06\x00\xAA\xA5"
-     "\x06\x00"
-     "\xAB\xBD\x06\x00\xAC\xA5\x06\x00\xAD\xAD\x06\x00\xAE\xA5\x06\x00\xAF\x19"
-     "\x05\x00"
-     "\xBE\x68\x3C\x00\x67\x65\x00\x80\x6B\x65\x00\x80\x6F\x65\x00\x80\x80\x0D"
-     "\x00\x00"
-     "\x81\xA5\x00\x00\x82\xAD\x00\x00\x73\x65\x00\x80\xB8\x15\x05\x00\xB9\x1D"
-     "\x05\x00"
-     "\xBA\x21\x05\x00\xBB\x21\x05\x00\xBC\xD5\x05\x00\xBD\xDD\x05\x00\xBE\xC9"
-     "\x05\x00"
-     "\xBF\xC9\x05\x00\xB0\x69\x05\x00\xB1\x69\x05\x00\xB2\x79\x05\x00\xB3\x71"
-     "\x05\x00"
-     "\xB4\x59\x05\x00\xB5\x59\x05\x00\xB6\x31\x05\x00\xB7\x31\x05\x00\xE1\x58"
-     "\x07\x00"
-     "\xE3\xBC\x00\x00\xE3\xF0\x06\x00\xE1\x28\x06\x00\x77\x65\x00\x80\x7B\x65"
-     "\x00\x80"
-     "\x86\x38\x3C\x00\x87\x50\x3D\x00\x83\x65\x00\x80\x87\x65\x00\x80\x8B\x65"
-     "\x00\x80"
-     "\x8F\x65\x00\x80\xBE\xFC\x3F\x00\xEF\xC0\x07\x00\xEF\xC0\x07\x00\x93\x65"
-     "\x00\x80"
-     "\xA3\x99\x02\x00\x97\x65\x00\x80\x9B\x65\x00\x80\x9F\x65\x00\x80\xA3\x65"
-     "\x00\x80"
-     "\xA6\x71\x02\x00\xA5\xB5\x02\x00\xA7\x65\x00\x80\xAB\x55\x02\x00\xAA\x55"
-     "\x02\x00"
-     "\xAB\x65\x00\x80\xAF\x65\x00\x80\xAF\xF9\x01\x00\xAE\xF1\x01\x00\xAD\x45"
-     "\x02\x00"
-     "\xAC\x45\x02\x00\xA8\x6D\x3E\x00\xA9\x75\x3E\x00\xAA\x7D\x3E\x00\xAB\x8D"
-     "\x3E\x00"
-     "\xAC\x95\x3E\x00\xAD\x81\x3E\x00\xAE\x81\x3E\x00\xAF\x81\x3E\x00\xB3\x65"
-     "\x00\x80"
-     "\xB7\x65\x00\x80\xBB\x65\x00\x80\xBF\x65\x00\x80\xC3\x65\x00\x80\xC7\x65"
-     "\x00\x80"
-     "\xCB\x65\x00\x80\xCF\x65\x00\x80\xB8\x89\x3E\x00\xB9\x95\x3E\x00\xBA\x95"
-     "\x3E\x00"
-     "\xBB\xAD\x3E\x00\xBC\xB5\x3E\x00\xBD\x71\x3F\x00\xBE\x71\x3F\x00\xBF\x71"
-     "\x3F\x00"
-     "\xB0\xC5\x3E\x00\xB1\xC9\x3E\x00\xB2\xD9\x3E\x00\xB3\xD9\x3E\x00\xB4\xC9"
-     "\x3E\x00"
-     "\xB5\xC9\x3E\x00\xB6\xB9\x3E\x00\xB7\xB9\x3E\x00\x0F\x65\x00\x80\xB3\xF1"
-     "\x3E\x00"
-     "\xD3\x65\x00\x80\x7F\x65\x00\x80\xB6\x05\x3E\x00\xD7\x65\x00\x80\xDB\x65"
-     "\x00\x80"
-     "\xB5\x19\x3E\x00\xBA\x05\x3E\x00\xBB\x09\x3E\x00\xDF\x65\x00\x80\xE3\x65"
-     "\x00\x80"
-     "\xBE\x09\x3E\x00\xBF\x09\x3E\x00\xBC\x09\x3E\x00\xBD\x09\x3E\x00\xA3\xB5"
-     "\x3E\x00"
-     "\x82\x19\x00\x00\x81\x19\x00\x00\x80\x61\x00\x00\xE7\x65\x00\x80\xA6\x41"
-     "\x3E\x00"
-     "\xA5\x5D\x3E\x00\xEB\x65\x00\x80\xAB\x4D\x3E\x00\xAA\x41\x3E\x00\x84\xA4"
-     "\x03\x00"
-     "\xEF\x65\x00\x80\xAF\x4D\x3E\x00\xAE\x4D\x3E\x00\xAD\x4D\x3E\x00\xAC\x4D"
-     "\x3E\x00"
-     "\xA8\xA1\x3E\x00\xA9\xA1\x3E\x00\xAA\xA1\x3E\x00\xAB\xA1\x3E\x00\xAC\xAD"
-     "\x3E\x00"
-     "\xAD\xD1\x3E\x00\xAE\xD1\x3E\x00\xAF\xD1\x3E\x00\x86\x80\x01\x00\x87\xA0"
-     "\x01\x00"
-     "\xF3\x65\x00\x80\xF7\x65\x00\x80\xFB\x65\x00\x80\xFF\x65\x00\x80\x03\x66"
-     "\x00\x80"
-     "\x07\x66\x00\x80\xB8\x71\x01\x00\xB9\x71\x01\x00\xBA\x71\x01\x00\xBB\x71"
-     "\x01\x00"
-     "\xBC\xD9\x01\x00\xBD\xC1\x01\x00\xBE\xC1\x01\x00\xBF\xF5\x01\x00\xB0\xB1"
-     "\x3E\x00"
-     "\xB1\xB9\x3E\x00\xB2\x81\x3E\x00\xB3\x81\x3E\x00\xB4\x51\x01\x00\xB5\x51"
-     "\x01\x00"
-     "\xB6\x51\x01\x00\xB7\x51\x01\x00\xB3\x29\x3E\x00\x0B\x66\x00\x80\x0F\x66"
-     "\x00\x80"
-     "\x13\x66\x00\x80\x17\x66\x00\x80\xB6\x51\x3E\x00\xB5\x25\x3E\x00\x1B\x66"
-     "\x00\x80"
-     "\xBB\x7D\x3E\x00\xBA\x75\x3E\x00\x1F\x66\x00\x80\x23\x66\x00\x80\xBF\x95"
-     "\x01\x00"
-     "\xBE\x95\x01\x00\xBD\x55\x3E\x00\xBC\x55\x3E\x00\x27\x66\x00\x80\xA3\x6D"
-     "\x3E\x00"
-     "\x2B\x66\x00\x80\x2F\x66\x00\x80\xA6\x15\x3E\x00\x33\x66\x00\x80\x37\x66"
-     "\x00\x80"
-     "\xA5\x61\x3E\x00\xAA\x31\x3E\x00\xAB\x39\x3E\x00\x3B\x66\x00\x80\x3F\x66"
-     "\x00\x80"
-     "\xAE\xD1\x01\x00\xAF\xD1\x01\x00\xAC\x11\x3E\x00\xAD\x11\x3E\x00\xA8\x29"
-     "\x03\x00"
-     "\xA9\x29\x03\x00\xAA\x39\x03\x00\xAB\x39\x03\x00\xAC\x29\x03\x00\xAD\x29"
-     "\x03\x00"
-     "\xAE\x9D\x03\x00\xAF\x95\x03\x00\x43\x66\x00\x80\x47\x66\x00\x80\x4B\x66"
-     "\x00\x80"
-     "\x4F\x66\x00\x80\x53\x66\x00\x80\x57\x66\x00\x80\x5B\x66\x00\x80\x63\x66"
-     "\x00\x80"
-     "\xB8\x95\x03\x00\xB9\x9D\x03\x00\xBA\xA1\x03\x00\xBB\xA1\x03\x00\xBC\x71"
-     "\x00\x00"
-     "\xBD\x71\x00\x00\xBE\x71\x00\x00\xBF\x71\x00\x00\xB0\xED\x03\x00\xB1\xF5"
-     "\x03\x00"
-     "\xB2\xF5\x03\x00\xB3\xC5\x03\x00\xB4\xDD\x03\x00\xB5\xB5\x03\x00\xB6\xB1"
-     "\x03\x00"
-     "\xB7\xB1\x03\x00\x81\xB5\x00\x00\x80\xB5\x00\x00\xB3\xB9\x03\x00\x82\xB5"
-     "\x00\x00"
-     "\xB5\xA9\x03\x00\x67\x66\x00\x80\x6B\x66\x00\x80\xB6\xA1\x03\x00\x84\xE4"
-     "\x02\x00"
-     "\x6F\x66\x00\x80\xBB\x39\x03\x00\xBA\x21\x03\x00\xBD\x19\x03\x00\xBC\x29"
-     "\x03\x00"
-     "\xBF\x79\x03\x00\xBE\x11\x03\x00\xA3\xF1\x03\x00\x73\x66\x00\x80\x87\xB8"
-     "\x03\x00"
-     "\x86\x6C\x04\x00\x77\x66\x00\x80\xA6\xE9\x03\x00\xA5\xE1\x03\x00\x7B\x66"
-     "\x00\x80"
-     "\xAB\x71\x03\x00\xAA\x69\x03\x00\x7F\x66\x00\x80\x83\x66\x00\x80\xAF\x31"
-     "\x03\x00"
-     "\xAE\x59\x03\x00\xAD\x51\x03\x00\xAC\x61\x03\x00\xE3\xF8\x01\x00\xE1\x64"
-     "\x3F\x00"
-     "\xE1\x88\x3E\x00\xE3\x94\x3F\x00\x87\x66\x00\x80\x8B\x66\x00\x80\x8F\x66"
-     "\x00\x80"
-     "\x93\x66\x00\x80\x97\x66\x00\x80\x9B\x66\x00\x80\x9F\x66\x00\x80\xA3\x66"
-     "\x00\x80"
-     "\xEF\xE8\x3E\x00\xA7\x66\x00\x80\xBF\x30\x13\x00\xEF\x48\x3F\x00\xAB\x66"
-     "\x00\x80"
-     "\xAF\x66\x00\x80\x84\x04\x05\x00\xBE\x04\x05\x00\xEF\x18\x3E\x00\xB3\x66"
-     "\x00\x80"
-     "\xB7\x66\x00\x80\xBB\x66\x00\x80\xE3\xD0\x03\x00\xBF\x66\x00\x80\xE1\x50"
-     "\x00\x00"
-     "\xC3\x66\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xB9\x00\x00\xC7\x66"
-     "\x00\x80"
-     "\xB2\x39\x3E\x00\xB3\x39\x3E\x00\xB0\x39\x3E\x00\xB1\x39\x3E\x00\xB6\x9D"
-     "\x3D\x00"
-     "\xB7\x85\x3D\x00\xB4\x9D\x3D\x00\xB5\x95\x3D\x00\xBA\x8D\x3D\x00\xBB\xE5"
-     "\x3D\x00"
-     "\xB8\xB5\x3D\x00\xB9\x85\x3D\x00\xBE\xFD\x3D\x00\xBF\xE5\x3D\x00\xBC\xFD"
-     "\x3D\x00"
-     "\xBD\xF5\x3D\x00\xCB\x66\x00\x80\xD3\x66\x00\x80\x86\xE8\x04\x00\x87\xCC"
-     "\x04\x00"
-     "\xD7\x66\x00\x80\xDB\x66\x00\x80\xDF\x66\x00\x80\xE3\x66\x00\x80\xAA\xB9"
-     "\x3E\x00"
-     "\xAB\x8D\x3E\x00\xA8\xA9\x3E\x00\xA9\xB9\x3E\x00\xAE\x59\x3E\x00\xAF\x59"
-     "\x3E\x00"
-     "\xAC\x95\x3E\x00\xAD\x59\x3E\x00\xA8\x3D\x06\x00\xA9\x91\x06\x00\xAA\x91"
-     "\x06\x00"
-     "\xAB\xA9\x06\x00\xAC\xD5\x06\x00\xAD\xC5\x06\x00\xAE\xC5\x06\x00\xAF\xF5"
-     "\x06\x00"
-     "\xCF\x66\x00\x80\xE7\x66\x00\x80\xEB\x66\x00\x80\xEF\x66\x00\x80\xF3\x66"
-     "\x00\x80"
-     "\xF7\x66\x00\x80\xFB\x66\x00\x80\xFF\x66\x00\x80\xB8\x6D\x07\x00\xB9\x71"
-     "\x07\x00"
-     "\xBA\x09\x07\x00\xBB\x09\x07\x00\xBC\x05\x07\x00\xBD\x0D\x07\x00\xBE\x05"
-     "\x07\x00"
-     "\xBF\x39\x07\x00\xB0\x8D\x06\x00\xB1\x75\x07\x00\xB2\x7D\x07\x00\xB3\x75"
-     "\x07\x00"
-     "\xB4\x6D\x07\x00\xB5\x55\x07\x00\xB6\x5D\x07\x00\xB7\x55\x07\x00\xB3\x19"
-     "\x06\x00"
-     "\x03\x67\x00\x80\x07\x67\x00\x80\x0B\x67\x00\x80\x0F\x67\x00\x80\xB6\x45"
-     "\x06\x00"
-     "\xB5\x5D\x06\x00\x13\x67\x00\x80\xBB\xB1\x07\x00\xBA\x49\x06\x00\x17\x67"
-     "\x00\x80"
-     "\x1B\x67\x00\x80\xBF\x91\x07\x00\xBE\xA9\x07\x00\xBD\xA1\x07\x00\xBC\xA9"
-     "\x07\x00"
-     "\x1F\x67\x00\x80\xA3\x5D\x06\x00\x23\x67\x00\x80\x27\x67\x00\x80\xA6\x01"
-     "\x06\x00"
-     "\x2B\x67\x00\x80\x2F\x67\x00\x80\xA5\x19\x06\x00\xAA\x0D\x06\x00\xAB\xF5"
-     "\x07\x00"
-     "\x33\x67\x00\x80\x37\x67\x00\x80\xAE\xED\x07\x00\xAF\xD5\x07\x00\xAC\xED"
-     "\x07\x00"
-     "\xAD\xE5\x07\x00\x80\xE9\x01\x00\x81\xE9\x01\x00\x82\x05\x00\x00\x3B\x67"
-     "\x00\x80"
-     "\xBE\xCC\x01\x00\x84\xD0\x01\x00\x3F\x67\x00\x80\x43\x67\x00\x80\x86\x80"
-     "\x00\x00"
-     "\x87\x98\x02\x00\x47\x67\x00\x80\x4B\x67\x00\x80\x4F\x67\x00\x80\x53\x67"
-     "\x00\x80"
-     "\x57\x67\x00\x80\x5B\x67\x00\x80\x5F\x67\x00\x80\x63\x67\x00\x80\x67\x67"
-     "\x00\x80"
-     "\x6B\x67\x00\x80\x6F\x67\x00\x80\xEF\xE0\x04\x00\x73\x67\x00\x80\x77\x67"
-     "\x00\x80"
-     "\x7B\x67\x00\x80\xE3\x3C\x04\x00\x7F\x67\x00\x80\xE1\xB8\x05\x00\x83\x67"
-     "\x00\x80"
-     "\x87\x67\x00\x80\x8B\x67\x00\x80\x8F\x67\x00\x80\xA8\x45\x06\x00\xA9\x55"
-     "\x06\x00"
-     "\xAA\x5D\x06\x00\xAB\x55\x06\x00\xAC\x4D\x06\x00\xAD\xB1\x01\x00\xAE\xB1"
-     "\x01\x00"
-     "\xAF\xB1\x01\x00\x84\xAC\x01\x00\x93\x67\x00\x80\x97\x67\x00\x80\x9B\x67"
-     "\x00\x80"
-     "\x9F\x67\x00\x80\x82\xB9\x03\x00\x81\xBD\x03\x00\x80\xBD\x03\x00\xB8\x5D"
-     "\x01\x00"
-     "\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x0D\x01\x00\xBC\x19\x01\x00\xBD\x19"
-     "\x01\x00"
-     "\xBE\x0D\x01\x00\xBF\x05\x01\x00\xB0\xD1\x01\x00\xB1\xD1\x01\x00\xB2\xD1"
-     "\x01\x00"
-     "\xB3\xD1\x01\x00\xB4\x75\x01\x00\xB5\x7D\x01\x00\xB6\x75\x01\x00\xB7\x6D"
-     "\x01\x00"
-     "\xA3\x67\x00\x80\xA7\x67\x00\x80\xB3\xDD\x05\x00\xAB\x67\x00\x80\xB5\xCD"
-     "\x05\x00"
-     "\xB3\x67\x00\x80\xBE\x4C\x03\x00\xB6\x39\x02\x00\xB7\x67\x00\x80\xBB\x67"
-     "\x00\x80"
-     "\xBB\x1D\x02\x00\xBA\x1D\x02\x00\xBD\x0D\x02\x00\xBC\x0D\x02\x00\xBF\xB1"
-     "\x02\x00"
-     "\xBE\xB9\x02\x00\xBF\x67\x00\x80\xA6\x61\x02\x00\xA5\x95\x05\x00\x84\x00"
-     "\x0C\x00"
-     "\xA3\x85\x05\x00\xC3\x67\x00\x80\x86\xF8\x0C\x00\x87\x7C\x03\x00\xAF\xE9"
-     "\x02\x00"
-     "\xAE\xE1\x02\x00\xAD\x55\x02\x00\xAC\x55\x02\x00\xAB\x45\x02\x00\xAA\x45"
-     "\x02\x00"
-     "\xC7\x67\x00\x80\xCB\x67\x00\x80\xCF\x67\x00\x80\xD3\x67\x00\x80\xD7\x67"
-     "\x00\x80"
-     "\xDB\x67\x00\x80\xDF\x67\x00\x80\xE3\x24\x01\x00\xE3\x67\x00\x80\xE1\xA4"
-     "\x01\x00"
-     "\xBE\xE8\x0C\x00\xEF\x44\x02\x00\xE7\x67\x00\x80\xEB\x67\x00\x80\xEF\x67"
-     "\x00\x80"
-     "\xF3\x67\x00\x80\xF7\x67\x00\x80\xFB\x67\x00\x80\xB3\xED\x03\x00\xFF\x67"
-     "\x00\x80"
-     "\x03\x68\x00\x80\x07\x68\x00\x80\x0B\x68\x00\x80\xB6\xED\x03\x00\xB5\xED"
-     "\x03\x00"
-     "\x0F\x68\x00\x80\xBB\xA9\x03\x00\xBA\xA1\x03\x00\x5F\x66\x00\x80\x13\x68"
-     "\x00\x80"
-     "\xBF\xAD\x03\x00\xBE\xAD\x03\x00\xBD\xA9\x03\x00\xBC\xB1\x03\x00\xA8\x1D"
-     "\x02\x00"
-     "\xA9\x2D\x02\x00\xAA\x25\x02\x00\xAB\x3D\x02\x00\xAC\x25\x02\x00\xAD\x2D"
-     "\x02\x00"
-     "\xAE\x25\x02\x00\xAF\x5D\x02\x00\x80\xCD\x01\x00\x81\x09\x00\x00\x82\x19"
-     "\x00\x00"
-     "\x17\x68\x00\x80\x1B\x68\x00\x80\x23\x68\x00\x80\x87\xB4\x0D\x00\x86\x1C"
-     "\x0C\x00"
-     "\xB8\x31\x02\x00\xB9\x31\x02\x00\xBA\x31\x02\x00\xBB\x31\x02\x00\xBC\xF5"
-     "\x02\x00"
-     "\xBD\xFD\x02\x00\xBE\xF5\x02\x00\xBF\xED\x02\x00\xB0\x25\x02\x00\xB1\x2D"
-     "\x02\x00"
-     "\xB2\x25\x02\x00\xB3\x3D\x02\x00\xB4\x29\x02\x00\xB5\x29\x02\x00\xB6\x19"
-     "\x02\x00"
-     "\xB7\x15\x02\x00\x27\x68\x00\x80\x2B\x68\x00\x80\x2F\x68\x00\x80\x33\x68"
-     "\x00\x80"
-     "\x37\x68\x00\x80\x3B\x68\x00\x80\x3F\x68\x00\x80\xEF\x74\x00\x00\x43\x68"
-     "\x00\x80"
-     "\xE1\x74\x06\x00\x47\x68\x00\x80\xE3\xA8\x00\x00\xEC\x00\x00\x00\x4B\x68"
-     "\x00\x80"
-     "\x4F\x68\x00\x80\xEF\x00\x00\x00\xA5\xF5\x02\x00\x53\x68\x00\x80\x57\x68"
-     "\x00\x80"
-     "\xA6\xF5\x02\x00\x5B\x68\x00\x80\x5F\x68\x00\x80\xA3\xF5\x02\x00\x63\x68"
-     "\x00\x80"
-     "\xAD\xB1\x02\x00\xAC\xA9\x02\x00\xAF\xB5\x02\x00\xAE\xB5\x02\x00\x67\x68"
-     "\x00\x80"
-     "\x6B\x68\x00\x80\xAB\xB1\x02\x00\xAA\xB9\x02\x00\xA8\x15\x0E\x00\xA9\x25"
-     "\x0E\x00"
-     "\xAA\x2D\x0E\x00\xAB\x4D\x0E\x00\xAC\x45\x0E\x00\xAD\x6D\x0E\x00\xAE\x65"
-     "\x0E\x00"
-     "\xAF\x9D\x0E\x00\x1F\x68\x00\x80\x6F\x68\x00\x80\x73\x68\x00\x80\x77\x68"
-     "\x00\x80"
-     "\x7B\x68\x00\x80\x7F\x68\x00\x80\x83\x68\x00\x80\x87\x68\x00\x80\xB8\xF5"
-     "\x0E\x00"
-     "\xB9\xFD\x0E\x00\xBA\xF5\x0E\x00\xBB\x89\x0E\x00\xBC\x99\x0E\x00\xBD\x99"
-     "\x0E\x00"
-     "\xBE\x89\x0E\x00\xBF\x81\x0E\x00\xB0\xE5\x0E\x00\xB1\xED\x0E\x00\xB2\xE5"
-     "\x0E\x00"
-     "\xB3\xFD\x0E\x00\xB4\xE9\x0E\x00\xB5\xDD\x0E\x00\xB6\xD5\x0E\x00\xB7\xCD"
-     "\x0E\x00"
-     "\xB3\xA5\x0E\x00\x82\x15\x00\x00\x81\xB5\x00\x00\x80\xB5\x00\x00\x8B\x68"
-     "\x00\x80"
-     "\xB6\xED\x0E\x00\xB5\xCD\x0E\x00\xBE\x10\x00\x00\xBB\x29\x0E\x00\xBA\x21"
-     "\x0E\x00"
-     "\x84\x64\x03\x00\x8F\x68\x00\x80\xBF\x19\x0E\x00\xBE\x2D\x0E\x00\xBD\x29"
-     "\x0E\x00"
-     "\xBC\x31\x0E\x00\x93\x68\x00\x80\xA3\xE1\x0E\x00\x86\xA8\x00\x00\x87\x6C"
-     "\x03\x00"
-     "\xA6\xA9\x0E\x00\x97\x68\x00\x80\x9B\x68\x00\x80\xA5\x89\x0E\x00\xAA\x65"
-     "\x0E\x00"
-     "\xAB\x6D\x0E\x00\x9F\x68\x00\x80\xA3\x68\x00\x80\xAE\x69\x0E\x00\xAF\x5D"
-     "\x0E\x00"
-     "\xAC\x75\x0E\x00\xAD\x6D\x0E\x00\xB3\x85\x0E\x00\xA7\x68\x00\x80\xAB\x68"
-     "\x00\x80"
-     "\xAF\x68\x00\x80\xB3\x68\x00\x80\xB6\x85\x0E\x00\xB5\x95\x0E\x00\xB7\x68"
-     "\x00\x80"
-     "\xBB\x89\x0E\x00\xBA\x85\x0E\x00\xBB\x68\x00\x80\xBF\x68\x00\x80\xBF\xF9"
-     "\x0E\x00"
-     "\xBE\xF1\x0E\x00\xBD\x81\x0E\x00\xBC\x89\x0E\x00\xEF\x18\x0D\x00\xC3\x68"
-     "\x00\x80"
-     "\xC7\x68\x00\x80\xCB\x68\x00\x80\xCF\x68\x00\x80\xD3\x68\x00\x80\xD7\x68"
-     "\x00\x80"
-     "\xDB\x68\x00\x80\xDF\x68\x00\x80\xE3\x68\x00\x80\xE7\x68\x00\x80\xEB\x68"
-     "\x00\x80"
-     "\xE3\x2C\x0D\x00\xEF\x68\x00\x80\xE1\x38\x0D\x00\xF3\x68\x00\x80\xA8\xB5"
-     "\x0E\x00"
-     "\xA9\x05\x0E\x00\xAA\x0D\x0E\x00\xAB\x05\x0E\x00\xAC\x1D\x0E\x00\xAD\x05"
-     "\x0E\x00"
-     "\xAE\x0D\x0E\x00\xAF\x05\x0E\x00\x80\xDD\x00\x00\x81\x09\x00\x00\x82\x19"
-     "\x00\x00"
-     "\xF7\x68\x00\x80\xFB\x68\x00\x80\x84\x3C\x01\x00\xBE\x74\x00\x00\x03\x69"
-     "\x00\x80"
-     "\xB8\x1D\x0E\x00\xB9\x35\x0E\x00\xBA\x3D\x0E\x00\xBB\xE9\x01\x00\xBC\xF9"
-     "\x01\x00"
-     "\xBD\xF9\x01\x00\xBE\xE9\x01\x00\xBF\xE9\x01\x00\xB0\x41\x0E\x00\xB1\x4D"
-     "\x0E\x00"
-     "\xB2\x5D\x0E\x00\xB3\x55\x0E\x00\xB4\x35\x0E\x00\xB5\x3D\x0E\x00\xB6\x29"
-     "\x0E\x00"
-     "\xB7\x29\x0E\x00\xA3\x85\x0D\x00\x07\x69\x00\x80\x86\xA8\x04\x00\x87\x3C"
-     "\x03\x00"
-     "\x0B\x69\x00\x80\xA6\x85\x0D\x00\xA5\x95\x0D\x00\x0F\x69\x00\x80\xAB\x89"
-     "\x0D\x00"
-     "\xAA\x85\x0D\x00\x13\x69\x00\x80\x17\x69\x00\x80\xAF\xF9\x0D\x00\xAE\xF1"
-     "\x0D\x00"
-     "\xAD\x81\x0D\x00\xAC\x89\x0D\x00\x1B\x69\x00\x80\xB3\x9D\x02\x00\x84\x48"
-     "\x03\x00"
-     "\xBE\x40\x04\x00\xB6\x55\x03\x00\x1F\x69\x00\x80\x23\x69\x00\x80\xB5\xB1"
-     "\x02\x00"
-     "\xBA\x71\x03\x00\xBB\x79\x03\x00\x27\x69\x00\x80\x2B\x69\x00\x80\xBE\x35"
-     "\x03\x00"
-     "\xBF\x3D\x03\x00\xBC\x51\x03\x00\xBD\x51\x03\x00\xB8\xD5\x03\x00\xB9\xDD"
-     "\x03\x00"
-     "\xBA\xD5\x03\x00\xBB\xED\x03\x00\xBC\xF5\x03\x00\xBD\xFD\x03\x00\xBE\xF5"
-     "\x03\x00"
-     "\xBF\xE9\x03\x00\xB0\xB1\x03\x00\xB1\xB9\x03\x00\xB2\x81\x03\x00\xB3\x81"
-     "\x03\x00"
-     "\xB4\xF5\x03\x00\xB5\xFD\x03\x00\xB6\xF5\x03\x00\xB7\xED\x03\x00\xA8\x01"
-     "\x02\x00"
-     "\xA9\x11\x02\x00\xAA\x11\x02\x00\xAB\x25\x02\x00\xAC\x3D\x02\x00\xAD\xD1"
-     "\x03\x00"
-     "\xAE\xD1\x03\x00\xAF\xD1\x03\x00\x2F\x69\x00\x80\x33\x69\x00\x80\x37\x69"
-     "\x00\x80"
-     "\x3B\x69\x00\x80\x3F\x69\x00\x80\x43\x69\x00\x80\x47\x69\x00\x80\xAF\x67"
-     "\x00\x80"
-     "\x4B\x69\x00\x80\x4F\x69\x00\x80\x53\x69\x00\x80\x57\x69\x00\x80\xEF\x50"
-     "\x00\x00"
-     "\x5B\x69\x00\x80\x5F\x69\x00\x80\x63\x69\x00\x80\xE3\x84\x00\x00\x67\x69"
-     "\x00\x80"
-     "\xE1\xF8\x01\x00\x6B\x69\x00\x80\x80\x2D\x00\x00\x81\x11\x00\x00\x82\x15"
-     "\x00\x00"
-     "\x73\x69\x00\x80\xA3\x99\x03\x00\x77\x69\x00\x80\x7B\x69\x00\x80\x7F\x69"
-     "\x00\x80"
-     "\x83\x69\x00\x80\xA6\x51\x02\x00\xA5\xB5\x03\x00\x87\x69\x00\x80\xAB\x7D"
-     "\x02\x00"
-     "\xAA\x75\x02\x00\x86\x48\x04\x00\x87\x40\x05\x00\xAF\x39\x02\x00\xAE\x31"
-     "\x02\x00"
-     "\xAD\x55\x02\x00\xAC\x55\x02\x00\x8B\x69\x00\x80\x8F\x69\x00\x80\x93\x69"
-     "\x00\x80"
-     "\x97\x69\x00\x80\x9B\x69\x00\x80\x9F\x69\x00\x80\xA3\x69\x00\x80\xEF\xF8"
-     "\x01\x00"
-     "\x85\xDC\x64\x00\xE1\x80\x0E\x00\xA7\x69\x00\x80\xE3\x14\x01\x00\xAB\x69"
-     "\x00\x80"
-     "\xAF\x69\x00\x80\xB3\x69\x00\x80\xB7\x69\x00\x80\xBB\x69\x00\x80\xBF\x69"
-     "\x00\x80"
-     "\xC3\x69\x00\x80\xB3\xDD\x01\x00\xC7\x69\x00\x80\xB5\xFD\x01\x00\xB6\xF5"
-     "\x01\x00"
-     "\xCB\x69\x00\x80\xCF\x69\x00\x80\xD3\x69\x00\x80\xBA\xB1\x01\x00\xBB\x85"
-     "\x01\x00"
-     "\xBC\x9D\x01\x00\xBD\x15\x01\x00\xBE\x1D\x01\x00\xBF\x15\x01\x00\xA8\xBD"
-     "\x05\x00"
-     "\xA9\x01\x06\x00\xAA\x19\x06\x00\xAB\x11\x06\x00\xAC\x39\x06\x00\xAD\x39"
-     "\x06\x00"
-     "\xAE\x59\x06\x00\xAF\x49\x06\x00\x6F\x69\x00\x80\x82\x11\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x80\xC9\x07\x00\xD7\x69\x00\x80\xDB\x69\x00\x80\x84\x30\x02\x00\xBE\xF0"
-     "\x03\x00"
-     "\xB8\xB5\x06\x00\xB9\xBD\x06\x00\xBA\x49\x07\x00\xBB\x49\x07\x00\xBC\x59"
-     "\x07\x00"
-     "\xBD\x59\x07\x00\xBE\x4D\x07\x00\xBF\x45\x07\x00\xB0\x39\x06\x00\xB1\x0D"
-     "\x06\x00"
-     "\xB2\x05\x06\x00\xB3\x8D\x06\x00\xB4\x95\x06\x00\xB5\x9D\x06\x00\xB6\x95"
-     "\x06\x00"
-     "\xB7\x8D\x06\x00\xA3\x91\x06\x00\xDF\x69\x00\x80\x86\x28\x00\x00\x87\xC0"
-     "\x03\x00"
-     "\xE3\x69\x00\x80\xA6\xB9\x06\x00\xA5\xB1\x06\x00\xE7\x69\x00\x80\xAB\xC9"
-     "\x06\x00"
-     "\xAA\xFD\x06\x00\xEB\x69\x00\x80\xEF\x69\x00\x80\xAF\x59\x06\x00\xAE\x51"
-     "\x06\x00"
-     "\xAD\x59\x06\x00\xAC\xD1\x06\x00\xF3\x69\x00\x80\xB3\x8D\x07\x00\xF7\x69"
-     "\x00\x80"
-     "\xFB\x69\x00\x80\xB6\x8D\x07\x00\xFF\x69\x00\x80\x03\x6A\x00\x80\xB5\x8D"
-     "\x07\x00"
-     "\xBA\x69\x07\x00\xBB\x69\x07\x00\x07\x6A\x00\x80\x0B\x6A\x00\x80\xBE\x69"
-     "\x07\x00"
-     "\xBF\x5D\x07\x00\xBC\x69\x07\x00\xBD\x69\x07\x00\x0F\x6A\x00\x80\xE1\x14"
-     "\x05\x00"
-     "\x13\x6A\x00\x80\xE3\x54\x05\x00\x17\x6A\x00\x80\x1B\x6A\x00\x80\x1F\x6A"
-     "\x00\x80"
-     "\x23\x6A\x00\x80\x27\x6A\x00\x80\x2B\x6A\x00\x80\x2F\x6A\x00\x80\x33\x6A"
-     "\x00\x80"
-     "\x37\x6A\x00\x80\x3B\x6A\x00\x80\x3F\x6A\x00\x80\xEF\x88\x05\x00\xA3\x09"
-     "\x06\x00"
-     "\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D\x00\x00\x43\x6A\x00\x80\xA6\x09"
-     "\x06\x00"
-     "\xA5\x09\x06\x00\x47\x6A\x00\x80\xAB\xED\x06\x00\xAA\xED\x06\x00\x4B\x6A"
-     "\x00\x80"
-     "\x84\x24\x01\x00\xAF\xD9\x06\x00\xAE\xED\x06\x00\xAD\xED\x06\x00\xAC\xED"
-     "\x06\x00"
-     "\xA8\xBD\x06\x00\xA9\xC1\x06\x00\xAA\xC1\x06\x00\xAB\xC1\x06\x00\xAC\xC1"
-     "\x06\x00"
-     "\xAD\xC9\x06\x00\xAE\x19\x06\x00\xAF\x19\x06\x00\xBE\xEC\x01\x00\x53\x6A"
-     "\x00\x80"
-     "\x86\x08\x1F\x00\x87\xEC\x00\x00\x57\x6A\x00\x80\x5B\x6A\x00\x80\x5F\x6A"
-     "\x00\x80"
-     "\x63\x6A\x00\x80\xB8\x79\x06\x00\xB9\x79\x06\x00\xBA\x09\x06\x00\xBB\x09"
-     "\x06\x00"
-     "\xBC\x19\x06\x00\xBD\x05\x06\x00\xBE\x05\x06\x00\xBF\x3D\x06\x00\xB0\x69"
-     "\x06\x00"
-     "\xB1\x69\x06\x00\xB2\x79\x06\x00\xB3\x79\x06\x00\xB4\x69\x06\x00\xB5\x5D"
-     "\x06\x00"
-     "\xB6\x49\x06\x00\xB7\x49\x06\x00\xA8\x91\x02\x00\xA9\x99\x02\x00\xAA\xA1"
-     "\x02\x00"
-     "\xAB\xA1\x02\x00\xAC\xD1\x02\x00\xAD\xDD\x02\x00\xAE\xD5\x02\x00\xAF\xC9"
-     "\x02\x00"
-     "\x67\x6A\x00\x80\x6B\x6A\x00\x80\x84\xA8\x03\x00\xBF\x24\x62\x00\x6F\x6A"
-     "\x00\x80"
-     "\x73\x6A\x00\x80\x77\x6A\x00\x80\x7B\x6A\x00\x80\xB8\x5D\x03\x00\xB9\x65"
-     "\x03\x00"
-     "\xBA\x6D\x03\x00\xBB\x65\x03\x00\xBC\x7D\x03\x00\xBD\x65\x03\x00\xBE\x6D"
-     "\x03\x00"
-     "\xBF\x65\x03\x00\xB0\xB9\x02\x00\xB1\x8D\x02\x00\xB2\x85\x02\x00\xB3\x6D"
-     "\x03\x00"
-     "\xB4\x79\x03\x00\xB5\x65\x03\x00\xB6\x6D\x03\x00\xB7\x65\x03\x00\xFF\x68"
-     "\x00\x80"
-     "\xB3\x2D\x02\x00\x7F\x6A\x00\x80\x83\x6A\x00\x80\xB6\x59\x02\x00\x87\x6A"
-     "\x00\x80"
-     "\x8B\x6A\x00\x80\xB5\x5D\x02\x00\xBA\x79\x02\x00\xBB\x4D\x02\x00\x8F\x6A"
-     "\x00\x80"
-     "\x93\x6A\x00\x80\xBE\xB9\x03\x00\xBF\xA5\x03\x00\xBC\x55\x02\x00\xBD\xBD"
-     "\x03\x00"
-     "\xA3\x69\x02\x00\xBE\x88\x03\x00\x97\x6A\x00\x80\x9B\x6A\x00\x80\x9F\x6A"
-     "\x00\x80"
-     "\xA6\x1D\x02\x00\xA5\x19\x02\x00\xA3\x6A\x00\x80\xAB\x09\x02\x00\xAA\x3D"
-     "\x02\x00"
-     "\xA7\x6A\x00\x80\xAB\x6A\x00\x80\xAF\xE1\x03\x00\xAE\xFD\x03\x00\xAD\xF9"
-     "\x03\x00"
-     "\xAC\x11\x02\x00\xE1\x90\x01\x00\xAF\x6A\x00\x80\xE3\x74\x00\x00\xB3\x6A"
-     "\x00\x80"
-     "\xB7\x6A\x00\x80\xBB\x6A\x00\x80\x87\x34\x1D\x00\x86\x3C\x1D\x00\xBE\x2C"
-     "\x1C\x00"
-     "\xC3\x6A\x00\x80\xC7\x6A\x00\x80\xCB\x6A\x00\x80\xCF\x6A\x00\x80\xD3\x6A"
-     "\x00\x80"
-     "\xEF\x60\x00\x00\xD7\x6A\x00\x80\x80\x4D\x00\x00\x81\x55\x00\x00\x82\x5D"
-     "\x00\x00"
-     "\xB3\xF9\x02\x00\xDB\x6A\x00\x80\xB5\xE9\x02\x00\xB6\x79\x01\x00\xDF\x6A"
-     "\x00\x80"
-     "\xE3\x6A\x00\x80\xE7\x6A\x00\x80\xBA\x5D\x01\x00\xBB\x55\x01\x00\xBC\x4D"
-     "\x01\x00"
-     "\xBD\xFD\x01\x00\xBE\xF5\x01\x00\xBF\xE9\x01\x00\xEB\x6A\x00\x80\xEF\x6A"
-     "\x00\x80"
-     "\xF3\x6A\x00\x80\xF7\x6A\x00\x80\xFB\x6A\x00\x80\xFF\x6A\x00\x80\x03\x6B"
-     "\x00\x80"
-     "\xEF\x58\x01\x00\x07\x6B\x00\x80\xE1\xD4\x06\x00\x0B\x6B\x00\x80\xE3\x84"
-     "\x01\x00"
-     "\x84\x3C\x1C\x00\x0F\x6B\x00\x80\x13\x6B\x00\x80\x17\x6B\x00\x80\xA3\x75"
-     "\x02\x00"
-     "\x1B\x6B\x00\x80\x1F\x6B\x00\x80\x84\xD0\x1E\x00\x23\x6B\x00\x80\xA6\xF5"
-     "\x01\x00"
-     "\xA5\x65\x02\x00\x27\x6B\x00\x80\xAB\xD9\x01\x00\xAA\xD1\x01\x00\x2B\x6B"
-     "\x00\x80"
-     "\x33\x6B\x00\x80\xAF\x65\x01\x00\xAE\x79\x01\x00\xAD\x71\x01\x00\xAC\xC1"
-     "\x01\x00"
-     "\xA8\xA5\x1D\x00\xA9\xB5\x1D\x00\xAA\xBD\x1D\x00\xAB\xB5\x1D\x00\xAC\xAD"
-     "\x1D\x00"
-     "\xAD\x35\x1E\x00\xAE\x3D\x1E\x00\xAF\x2D\x1E\x00\xBF\x6A\x00\x80\x82\xCD"
-     "\x1F\x00"
-     "\x81\xF5\x1F\x00\x80\xFD\x1F\x00\x2F\x6B\x00\x80\x37\x6B\x00\x80\x87\x4C"
-     "\x00\x00"
-     "\x86\x9C\x01\x00\xB8\xED\x1E\x00\xB9\x8D\x1E\x00\xBA\x85\x1E\x00\xBB\x9D"
-     "\x1E\x00"
-     "\xBC\x8D\x1E\x00\xBD\xBD\x1E\x00\xBE\xB5\x1E\x00\xBF\x59\x1F\x00\xB0\x55"
-     "\x1E\x00"
-     "\xB1\x5D\x1E\x00\xB2\x55\x1E\x00\xB3\xE5\x1E\x00\xB4\xFD\x1E\x00\xB5\xE1"
-     "\x1E\x00"
-     "\xB6\xE1\x1E\x00\xB7\xD5\x1E\x00\xA8\xA5\x1F\x00\xA9\xAD\x1F\x00\xAA\xA5"
-     "\x1F\x00"
-     "\xAB\xBD\x1F\x00\xAC\xA5\x1F\x00\xAD\xA9\x1F\x00\xAE\xC9\x1F\x00\xAF\xC1"
-     "\x1F\x00"
-     "\x3B\x6B\x00\x80\x3F\x6B\x00\x80\x43\x6B\x00\x80\x47\x6B\x00\x80\x4B\x6B"
-     "\x00\x80"
-     "\x4F\x6B\x00\x80\x53\x6B\x00\x80\x57\x6B\x00\x80\xB8\x59\x1F\x00\xB9\x65"
-     "\x1F\x00"
-     "\xBA\x61\x1F\x00\xBB\x79\x1F\x00\xBC\x69\x1F\x00\xBD\x69\x1F\x00\xBE\x1D"
-     "\x1F\x00"
-     "\xBF\x15\x1F\x00\xB0\x81\x1F\x00\xB1\x81\x1F\x00\xB2\x69\x1F\x00\xB3\x69"
-     "\x1F\x00"
-     "\xB4\x79\x1F\x00\xB5\x79\x1F\x00\xB6\x69\x1F\x00\xB7\x69\x1F\x00\x5B\x6B"
-     "\x00\x80"
-     "\x5F\x6B\x00\x80\x63\x6B\x00\x80\x67\x6B\x00\x80\xE1\xA0\x1E\x00\xE1\x8C"
-     "\x3D\x00"
-     "\xE3\x20\x1D\x00\xE3\xA8\x18\x00\x6B\x6B\x00\x80\x6F\x6B\x00\x80\xEF\x14"
-     "\x1D\x00"
-     "\xEF\xFC\x1D\x00\xA2\xA5\x03\x00\xA3\x15\x01\x00\xA0\x19\x1D\x00\xA1\xB9"
-     "\x1D\x00"
-     "\xB3\x6D\x1E\x00\x73\x6B\x00\x80\x77\x6B\x00\x80\x7B\x6B\x00\x80\x7F\x6B"
-     "\x00\x80"
-     "\xB6\x69\x1E\x00\xB5\x6D\x1E\x00\x83\x6B\x00\x80\xBB\x05\x1E\x00\xBA\x05"
-     "\x1E\x00"
-     "\x87\x6B\x00\x80\x8B\x6B\x00\x80\xBF\xFD\x1E\x00\xBE\xFD\x1E\x00\xBD\xFD"
-     "\x1E\x00"
-     "\xBC\xFD\x1E\x00\xA3\xAD\x1E\x00\x8F\x15\x24\x00\x8F\x6B\x00\x80\x93\x6B"
-     "\x00\x80"
-     "\x97\x6B\x00\x80\xA6\xA9\x1E\x00\xA5\xAD\x1E\x00\x9B\x6B\x00\x80\xAB\xC5"
-     "\x1E\x00"
-     "\xAA\xC5\x1E\x00\xBE\x00\x04\x00\xA3\x6B\x00\x80\xAF\x3D\x1E\x00\xAE\x3D"
-     "\x1E\x00"
-     "\xAD\x3D\x1E\x00\xAC\x3D\x1E\x00\x9E\x6D\x19\x00\x9F\x45\x1F\x00\x9C\x35"
-     "\x1B\x00"
-     "\x9D\xED\x19\x00\x9A\xF1\x15\x00\x9B\xCD\x14\x00\x98\x35\x11\x00\x99\x35"
-     "\x17\x00"
-     "\x96\x61\x11\x00\x97\x69\x11\x00\x94\xE1\x2D\x00\x95\xC1\x2C\x00\x92\x85"
-     "\x28\x00"
-     "\x93\xF1\x2D\x00\x90\x79\x29\x00\x91\x79\x29\x00\x82\x91\x27\x00\x83\x85"
-     "\x2A\x00"
-     "\x86\x88\x04\x00\x87\x8C\x03\x00\x86\x15\x2E\x00\x87\x1D\x2E\x00\x84\x9D"
-     "\x2A\x00"
-     "\x85\x69\x2B\x00\x8A\x0D\x2F\x00\x8B\xA5\x12\x00\xA7\x6B\x00\x80\x84\x60"
-     "\x04\x00"
-     "\x8E\xD9\x14\x00\x8F\x15\x16\x00\x8C\xB5\x12\x00\x8D\xD1\x12\x00\x92\x59"
-     "\x17\x00"
-     "\x93\xB9\x18\x00\xAB\x6B\x00\x80\xAF\x6B\x00\x80\x96\xD1\x1C\x00\x97\x75"
-     "\x1E\x00"
-     "\x94\x79\x1A\x00\x95\xF1\x1A\x00\x9A\xE5\x1E\x00\x9B\xF1\x00\x00\xB3\x6B"
-     "\x00\x80"
-     "\xB7\x6B\x00\x80\x81\xE1\x03\x00\x80\xF9\x03\x00\x9C\x59\x02\x00\x82\xE1"
-     "\x03\x00"
-     "\xA5\x01\x08\x00\xA4\xB5\x07\x00\xA7\x7D\x0A\x00\xA6\xAD\x0A\x00\xA1\xA1"
-     "\x02\x00"
-     "\x4F\x6A\x00\x80\xA3\x9D\x06\x00\xA2\x9D\x06\x00\xAD\x9D\x72\x00\xAC\x85"
-     "\x72\x00"
-     "\xAF\xAD\x76\x00\xAE\x01\x74\x00\xA9\x91\x0E\x00\xA8\x01\x0C\x00\xAB\x01"
-     "\x70\x00"
-     "\xAA\x15\x0E\x00\xB5\x01\x7C\x00\xB4\xE9\x7B\x00\xBB\x6B\x00\x80\xBF\x6B"
-     "\x00\x80"
-     "\xB1\xB9\x77\x00\xB0\xF9\x76\x00\xB3\xDD\x7A\x00\xB2\xDD\x7A\x00\xC3\x6B"
-     "\x00\x80"
-     "\xC7\x6B\x00\x80\xCB\x6B\x00\x80\xCF\x6B\x00\x80\xD3\x6B\x00\x80\xD7\x6B"
-     "\x00\x80"
-     "\xDB\x6B\x00\x80\xDF\x6B\x00\x80\xA8\xED\x1C\x00\xA9\x31\x03\x00\xAA\xA9"
-     "\x03\x00"
-     "\xAB\xA9\x03\x00\xAC\xB9\x03\x00\xAD\xB9\x03\x00\xAE\xA9\x03\x00\xAF\xA9"
-     "\x03\x00"
-     "\xE3\x6B\x00\x80\xE7\x6B\x00\x80\xEB\x6B\x00\x80\xEF\x6B\x00\x80\xF3\x6B"
-     "\x00\x80"
-     "\xF7\x6B\x00\x80\xFB\x6B\x00\x80\xFF\x6B\x00\x80\xB8\x05\x02\x00\xB9\x0D"
-     "\x02\x00"
-     "\xBA\x05\x02\x00\xBB\x1D\x02\x00\xBC\x0D\x02\x00\xBD\x3D\x02\x00\xBE\x35"
-     "\x02\x00"
-     "\xBF\xD5\x02\x00\xB0\xD9\x03\x00\xB1\xD9\x03\x00\xB2\x69\x02\x00\xB3\x69"
-     "\x02\x00"
-     "\xB4\x79\x02\x00\xB5\x79\x02\x00\xB6\x59\x02\x00\xB7\x59\x02\x00\x80\x11"
-     "\x00\x00"
-     "\x81\x11\x00\x00\x82\x11\x00\x00\xEF\xDC\x00\x00\x03\x6C\x00\x80\x07\x6C"
-     "\x00\x80"
-     "\x0B\x6C\x00\x80\xEF\x44\x1E\x00\xBE\xE8\x05\x00\xE1\xA0\x1E\x00\x13\x6C"
-     "\x00\x80"
-     "\xE3\x0C\x1E\x00\x17\x6C\x00\x80\xE1\x90\x01\x00\x84\xF4\x04\x00\xE3\x60"
-     "\x1E\x00"
-     "\x1B\x6C\x00\x80\xB6\x69\x03\x00\xB5\x61\x03\x00\x1F\x6C\x00\x80\xB3\x71"
-     "\x03\x00"
-     "\x23\x6C\x00\x80\x86\x78\x04\x00\x87\xFC\x04\x00\xBF\xED\x02\x00\xBE\xED"
-     "\x02\x00"
-     "\xBD\xED\x02\x00\xBC\xED\x02\x00\xBB\x4D\x03\x00\xBA\x4D\x03\x00\x27\x6C"
-     "\x00\x80"
-     "\x2B\x6C\x00\x80\x9F\x6B\x00\x80\x0F\x6C\x00\x80\x2F\x6C\x00\x80\x33\x6C"
-     "\x00\x80"
-     "\x37\x6C\x00\x80\x3B\x6C\x00\x80\x3F\x6C\x00\x80\x43\x6C\x00\x80\xA8\x8D"
-     "\x05\x00"
-     "\xA9\x0D\x06\x00\xAA\x61\x06\x00\xAB\x61\x06\x00\xAC\x59\x06\x00\xAD\x59"
-     "\x06\x00"
-     "\xAE\xF1\x06\x00\xAF\xF1\x06\x00\xB0\xA1\x06\x00\xB1\xA1\x06\x00\xB2\xA1"
-     "\x06\x00"
-     "\xB3\xA1\x06\x00\xB4\xA1\x06\x00\xB5\xA1\x06\x00\xB6\xA1\x06\x00\xB7\xA1"
-     "\x06\x00"
-     "\xB8\x41\x07\x00\xB9\x41\x07\x00\xBA\x41\x07\x00\xBB\x41\x07\x00\xBC\x41"
-     "\x07\x00"
-     "\xBD\x41\x07\x00\xBE\x41\x07\x00\xBF\x41\x07\x00\xB3\xC9\x06\x00\x47\x6C"
-     "\x00\x80"
-     "\x4B\x6C\x00\x80\x4F\x6C\x00\x80\x53\x6C\x00\x80\xB6\xF9\x06\x00\xB5\xF1"
-     "\x06\x00"
-     "\x57\x6C\x00\x80\xBB\x45\x06\x00\xBA\x45\x06\x00\x5B\x6C\x00\x80\x5F\x6C"
-     "\x00\x80"
-     "\xBF\x45\x06\x00\xBE\x45\x06\x00\xBD\x55\x06\x00\xBC\x55\x06\x00\x63\x6C"
-     "\x00\x80"
-     "\xA3\x8D\x06\x00\x67\x6C\x00\x80\x6B\x6C\x00\x80\xA6\xBD\x06\x00\x6F\x6C"
-     "\x00\x80"
-     "\x73\x6C\x00\x80\xA5\xB5\x06\x00\xAA\x01\x06\x00\xAB\x01\x06\x00\x77\x6C"
-     "\x00\x80"
-     "\x7B\x6C\x00\x80\xAE\x01\x06\x00\xAF\x01\x06\x00\xAC\x11\x06\x00\xAD\x11"
-     "\x06\x00"
-     "\xA8\xA9\x06\x00\xA9\xB5\x06\x00\xAA\xC9\x06\x00\xAB\xC1\x06\x00\xAC\xC5"
-     "\x06\x00"
-     "\xAD\xED\x06\x00\xAE\xE5\x06\x00\xAF\x15\x01\x00\x80\xF9\x01\x00\x81\xCD"
-     "\x01\x00"
-     "\x82\xC5\x01\x00\xBE\x64\x02\x00\x7F\x6C\x00\x80\x83\x6C\x00\x80\x87\xAC"
-     "\x01\x00"
-     "\x86\x9C\x01\x00\xB8\xD1\x01\x00\xB9\xD9\x01\x00\xBA\xE1\x01\x00\xBB\xE1"
-     "\x01\x00"
-     "\xBC\x91\x01\x00\xBD\x9D\x01\x00\xBE\x95\x01\x00\xBF\x89\x01\x00\xB0\x6D"
-     "\x01\x00"
-     "\xB1\x75\x01\x00\xB2\x7D\x01\x00\xB3\x75\x01\x00\xB4\x6D\x01\x00\xB5\xF1"
-     "\x01\x00"
-     "\xB6\xF1\x01\x00\xB7\xF1\x01\x00\xB3\x41\x06\x00\x87\x6C\x00\x80\x8B\x6C"
-     "\x00\x80"
-     "\x8F\x6C\x00\x80\x93\x6C\x00\x80\xB6\x6D\x06\x00\xB5\x4D\x06\x00\x97\x6C"
-     "\x00\x80"
-     "\xBB\xB1\x01\x00\xBA\xA9\x01\x00\x84\x84\x01\x00\x9B\x6C\x00\x80\xBF\x35"
-     "\x01\x00"
-     "\xBE\xA9\x01\x00\xBD\xA1\x01\x00\xBC\xA9\x01\x00\x9F\x6C\x00\x80\xA3\x05"
-     "\x06\x00"
-     "\xA3\x6C\x00\x80\xA7\x6C\x00\x80\xA6\x29\x06\x00\xAB\x6C\x00\x80\xAF\x6C"
-     "\x00\x80"
-     "\xA5\x09\x06\x00\xAA\xED\x01\x00\xAB\xF5\x01\x00\xB3\x6C\x00\x80\xB7\x6C"
-     "\x00\x80"
-     "\xAE\xED\x01\x00\xAF\x71\x01\x00\xAC\xED\x01\x00\xAD\xE5\x01\x00\xA8\x35"
-     "\x03\x00"
-     "\xA9\x81\x03\x00\xAA\x81\x03\x00\xAB\x81\x03\x00\xAC\x81\x03\x00\xAD\x89"
-     "\x03\x00"
-     "\xAE\xB5\x03\x00\xAF\xDD\x03\x00\xBB\x6C\x00\x80\xBF\x6C\x00\x80\xC3\x6C"
-     "\x00\x80"
-     "\xC7\x6C\x00\x80\xCB\x6C\x00\x80\xCF\x6C\x00\x80\xD3\x6C\x00\x80\xD7\x6C"
-     "\x00\x80"
-     "\xB8\x69\x00\x00\xB9\x69\x00\x00\xBA\x79\x00\x00\xBB\x79\x00\x00\xBC\x69"
-     "\x00\x00"
-     "\xBD\x69\x00\x00\xBE\xDD\x00\x00\xBF\xD5\x00\x00\xB0\xA9\x03\x00\xB1\xA9"
-     "\x03\x00"
-     "\xB2\xBD\x03\x00\xB3\xB5\x03\x00\xB4\x99\x03\x00\xB5\x99\x03\x00\xB6\x59"
-     "\x00\x00"
-     "\xB7\x59\x00\x00\xBE\x44\x03\x00\xDB\x6C\x00\x80\xDF\x6C\x00\x80\xEF\xF4"
-     "\x02\x00"
-     "\x81\x31\x00\x00\x80\x0D\x00\x00\xE3\x6C\x00\x80\x82\x39\x00\x00\xE7\x6C"
-     "\x00\x80"
-     "\xEB\x6C\x00\x80\x86\x08\x0C\x00\x87\x0C\x03\x00\xF3\x6C\x00\x80\xE1\x90"
-     "\x01\x00"
-     "\xF7\x6C\x00\x80\xE3\x04\x01\x00\xBB\x0D\x03\x00\xBA\x0D\x03\x00\xFB\x6C"
-     "\x00\x80"
-     "\xFF\x6C\x00\x80\xBF\x3D\x03\x00\xBE\x3D\x03\x00\xBD\x01\x03\x00\xBC\x09"
-     "\x03\x00"
-     "\xB3\x69\x03\x00\x03\x6D\x00\x80\x07\x6D\x00\x80\x0B\x6D\x00\x80\x0F\x6D"
-     "\x00\x80"
-     "\xB6\x69\x03\x00\xB5\x69\x03\x00\x13\x6D\x00\x80\xAA\xB9\x07\x00\xAB\xB9"
-     "\x07\x00"
-     "\x17\x6D\x00\x80\x1B\x6D\x00\x80\xAE\x19\x06\x00\xAF\x19\x06\x00\xAC\x19"
-     "\x06\x00"
-     "\xAD\x19\x06\x00\xBF\x28\x5C\x00\xA3\x85\x07\x00\x1F\x6D\x00\x80\x23\x6D"
-     "\x00\x80"
-     "\xA6\x9D\x07\x00\x27\x6D\x00\x80\x2B\x6D\x00\x80\xA5\x95\x07\x00\x2F\x6D"
-     "\x00\x80"
-     "\x33\x6D\x00\x80\xA3\xC1\x02\x00\x84\x40\x02\x00\xA5\xC1\x02\x00\x37\x6D"
-     "\x00\x80"
-     "\x3B\x6D\x00\x80\xA6\xC1\x02\x00\x3F\x6D\x00\x80\x43\x6D\x00\x80\xAB\xA5"
-     "\x02\x00"
-     "\xAA\xA5\x02\x00\xAD\xA9\x02\x00\xAC\xA1\x02\x00\xAF\x95\x02\x00\xAE\x95"
-     "\x02\x00"
-     "\xE1\x40\x06\x00\xE1\x94\x07\x00\xE3\x20\x06\x00\xE3\xB0\x07\x00\x47\x6D"
-     "\x00\x80"
-     "\x4B\x6D\x00\x80\x4F\x6D\x00\x80\x53\x6D\x00\x80\x80\x3D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x57\x6D\x00\x80\x5B\x6D\x00\x80\xBE\xB8\x0D\x00\xEF\x6C"
-     "\x00\x00"
-     "\xEF\xE4\x06\x00\xB3\x09\x01\x00\x63\x6D\x00\x80\x86\xE8\x0C\x00\x87\x10"
-     "\x0D\x00"
-     "\x67\x6D\x00\x80\xB6\x3D\x01\x00\xB5\x3D\x01\x00\x6B\x6D\x00\x80\xBB\xE9"
-     "\x00\x00"
-     "\xBA\x19\x01\x00\x6F\x6D\x00\x80\x73\x6D\x00\x80\xBF\xC5\x00\x00\xBE\xD9"
-     "\x00\x00"
-     "\xBD\xD1\x00\x00\xBC\xF1\x00\x00\xA8\xB1\x0E\x00\xA9\xBD\x0E\x00\xAA\xC1"
-     "\x0E\x00"
-     "\xAB\xC1\x0E\x00\xAC\xF5\x0E\x00\xAD\xFD\x0E\x00\xAE\xF5\x0E\x00\xAF\xED"
-     "\x0E\x00"
-     "\x5F\x6D\x00\x80\x77\x6D\x00\x80\x7B\x6D\x00\x80\x7F\x6D\x00\x80\x83\x6D"
-     "\x00\x80"
-     "\x87\x6D\x00\x80\x8B\x6D\x00\x80\x8F\x6D\x00\x80\xB8\x05\x0F\x00\xB9\x11"
-     "\x0F\x00"
-     "\xBA\x11\x0F\x00\xBB\x2D\x0F\x00\xBC\x35\x0F\x00\xBD\xC5\x0F\x00\xBE\xCD"
-     "\x0F\x00"
-     "\xBF\xC5\x0F\x00\xB0\x9D\x0E\x00\xB1\x6D\x0F\x00\xB2\x65\x0F\x00\xB3\x45"
-     "\x0F\x00"
-     "\xB4\x5D\x0F\x00\xB5\x41\x0F\x00\xB6\x41\x0F\x00\xB7\x3D\x0F\x00\xA3\x49"
-     "\x0E\x00"
-     "\x93\x6D\x00\x80\x97\x6D\x00\x80\x9B\x6D\x00\x80\x9F\x6D\x00\x80\xA6\x7D"
-     "\x0E\x00"
-     "\xA5\x7D\x0E\x00\xA3\x6D\x00\x80\xAB\xA9\x0F\x00\xAA\x59\x0E\x00\xA7\x6D"
-     "\x00\x80"
-     "\xAB\x6D\x00\x80\xAF\x85\x0F\x00\xAE\x99\x0F\x00\xAD\x91\x0F\x00\xAC\xB1"
-     "\x0F\x00"
-     "\xAF\x6D\x00\x80\xB3\x6D\x00\x80\xB7\x6D\x00\x80\xBB\x6D\x00\x80\xBF\x6D"
-     "\x00\x80"
-     "\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D\x00\x00\xC3\x6D\x00\x80\xC7\x6D"
-     "\x00\x80"
-     "\xCB\x6D\x00\x80\x84\x60\x03\x00\xBE\x70\x00\x00\xCF\x6D\x00\x80\x86\xF8"
-     "\x00\x00"
-     "\x87\x4C\x03\x00\xD3\x6D\x00\x80\xE1\xB4\x0D\x00\xD7\x6D\x00\x80\xE3\x50"
-     "\x0C\x00"
-     "\xDB\x6D\x00\x80\xDF\x6D\x00\x80\xE3\x6D\x00\x80\xE7\x6D\x00\x80\xEB\x6D"
-     "\x00\x80"
-     "\xEF\x6D\x00\x80\xF3\x6D\x00\x80\xF7\x6D\x00\x80\xFB\x6D\x00\x80\xFF\x6D"
-     "\x00\x80"
-     "\x03\x6E\x00\x80\xEF\x70\x0C\x00\xB3\x9D\x0E\x00\x07\x6E\x00\x80\x0B\x6E"
-     "\x00\x80"
-     "\x0F\x6E\x00\x80\x13\x6E\x00\x80\xB6\x71\x01\x00\xB5\xB5\x0E\x00\x17\x6E"
-     "\x00\x80"
-     "\xBB\x55\x01\x00\xBA\x55\x01\x00\x1B\x6E\x00\x80\x1F\x6E\x00\x80\xBF\xF9"
-     "\x01\x00"
-     "\xBE\xF1\x01\x00\xBD\x45\x01\x00\xBC\x45\x01\x00\xA8\x45\x0E\x00\xA9\x55"
-     "\x0E\x00"
-     "\xAA\x55\x0E\x00\xAB\x65\x0E\x00\xAC\x7D\x0E\x00\xAD\xB1\x01\x00\xAE\xB1"
-     "\x01\x00"
-     "\xAF\xB1\x01\x00\x23\x6E\x00\x80\x27\x6E\x00\x80\x2B\x6E\x00\x80\x2F\x6E"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\xB5\x03\x00\x82\xBD\x03\x00\x33\x6E\x00\x80\xB8\x5D"
-     "\x01\x00"
-     "\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x0D\x01\x00\xBC\x15\x01\x00\xBD\x1D"
-     "\x01\x00"
-     "\xBE\x15\x01\x00\xBF\x0D\x01\x00\xB0\xD1\x01\x00\xB1\xD1\x01\x00\xB2\xD1"
-     "\x01\x00"
-     "\xB3\xD1\x01\x00\xB4\x75\x01\x00\xB5\x7D\x01\x00\xB6\x75\x01\x00\xB7\x6D"
-     "\x01\x00"
-     "\x37\x6E\x00\x80\x3B\x6E\x00\x80\x84\x90\x02\x00\xA3\xD1\x0D\x00\x43\x6E"
-     "\x00\x80"
-     "\xA5\xF9\x0D\x00\xA6\x3D\x02\x00\x47\x6E\x00\x80\xBE\xAC\x03\x00\x4B\x6E"
-     "\x00\x80"
-     "\xAA\x19\x02\x00\xAB\x19\x02\x00\xAC\x09\x02\x00\xAD\x09\x02\x00\xAE\xBD"
-     "\x02\x00"
-     "\xAF\xB5\x02\x00\xB6\x5D\x03\x00\x4F\x6E\x00\x80\x53\x6E\x00\x80\xB5\x55"
-     "\x03\x00"
-     "\x57\x6E\x00\x80\xB3\x45\x03\x00\x86\xF8\x04\x00\x87\x08\x03\x00\xBE\x59"
-     "\x03\x00"
-     "\xBF\x21\x03\x00\xBC\x59\x03\x00\xBD\x59\x03\x00\xBA\x59\x03\x00\xBB\x59"
-     "\x03\x00"
-     "\x5B\x6E\x00\x80\x5F\x6E\x00\x80\xEF\x6C\x00\x80\x63\x6E\x00\x80\x67\x6E"
-     "\x00\x80"
-     "\x6B\x6E\x00\x80\x6F\x6E\x00\x80\x73\x6E\x00\x80\x77\x6E\x00\x80\x7B\x6E"
-     "\x00\x80"
-     "\xA8\xA1\x03\x00\xA9\xAD\x03\x00\xAA\xB9\x03\x00\xAB\xB9\x03\x00\xAC\xA9"
-     "\x03\x00"
-     "\xAD\xA9\x03\x00\xAE\x99\x03\x00\xAF\x99\x03\x00\xB0\xE9\x03\x00\xB1\xE9"
-     "\x03\x00"
-     "\xB2\xF9\x03\x00\xB3\xF9\x03\x00\xB4\xE9\x03\x00\xB5\x91\x03\x00\xB6\x91"
-     "\x03\x00"
-     "\xB7\x91\x03\x00\xB8\xB1\x03\x00\xB9\xB9\x03\x00\xBA\x81\x03\x00\xBB\x81"
-     "\x03\x00"
-     "\xBC\x51\x03\x00\xBD\x51\x03\x00\xBE\x51\x03\x00\xBF\x51\x03\x00\xA3\x11"
-     "\x02\x00"
-     "\x7F\x6E\x00\x80\x83\x6E\x00\x80\x87\x6E\x00\x80\x8B\x6E\x00\x80\xA6\x09"
-     "\x02\x00"
-     "\xA5\x01\x02\x00\x8F\x6E\x00\x80\xAB\x0D\x02\x00\xAA\x0D\x02\x00\x93\x6E"
-     "\x00\x80"
-     "\x97\x6E\x00\x80\xAF\x75\x02\x00\xAE\x0D\x02\x00\xAD\x0D\x02\x00\xAC\x0D"
-     "\x02\x00"
-     "\x9B\x6E\x00\x80\x9F\x6E\x00\x80\xEF\x90\x03\x00\xA3\x6E\x00\x80\x81\x09"
-     "\x00\x00"
-     "\x80\x19\x00\x00\xA7\x6E\x00\x80\x82\x3D\x00\x00\xAB\x6E\x00\x80\x84\x68"
-     "\x02\x00"
-     "\xBE\x64\x05\x00\xB3\x6E\x00\x80\xE1\x90\x01\x00\xB7\x6E\x00\x80\xE3\x60"
-     "\x03\x00"
-     "\xBB\x6E\x00\x80\x86\xA0\x04\x00\x87\x08\x05\x00\xBF\x6E\x00\x80\x84\x84"
-     "\x05\x00"
-     "\xC3\x6E\x00\x80\xC7\x6E\x00\x80\xCB\x6E\x00\x80\xEF\x94\x01\x00\xCF\x6E"
-     "\x00\x80"
-     "\xE1\x74\x0E\x00\xD3\x6E\x00\x80\xE3\x50\x01\x00\xED\x00\x00\x00\xD7\x6E"
-     "\x00\x80"
-     "\xDB\x6E\x00\x80\xDF\x6E\x00\x80\xB3\xE5\x01\x00\xE3\x6E\x00\x80\xE7\x6E"
-     "\x00\x80"
-     "\xEB\x6E\x00\x80\xEF\x6E\x00\x80\xB6\x85\x01\x00\xB5\x91\x01\x00\xF3\x6E"
-     "\x00\x80"
-     "\xBB\x69\x01\x00\xBA\x61\x01\x00\xF7\x6E\x00\x80\xFB\x6E\x00\x80\xBF\x69"
-     "\x01\x00"
-     "\xBE\x61\x01\x00\xBD\x69\x01\x00\xBC\x71\x01\x00\xA8\x11\x06\x00\xA9\x19"
-     "\x06\x00"
-     "\xAA\x2D\x06\x00\xAB\x49\x06\x00\xAC\x59\x06\x00\xAD\x45\x06\x00\xAE\x61"
-     "\x06\x00"
-     "\xAF\x61\x06\x00\xAF\x6E\x00\x80\xFF\x6E\x00\x80\x03\x6F\x00\x80\x07\x6F"
-     "\x00\x80"
-     "\x0B\x6F\x00\x80\x0F\x6F\x00\x80\x13\x6F\x00\x80\x17\x6F\x00\x80\xB8\x49"
-     "\x07\x00"
-     "\xB9\x49\x07\x00\xBA\x59\x07\x00\xBB\x51\x07\x00\xBC\x79\x07\x00\xBD\x79"
-     "\x07\x00"
-     "\xBE\x19\x07\x00\xBF\x19\x07\x00\xB0\xE5\x06\x00\xB1\xED\x06\x00\xB2\xE5"
-     "\x06\x00"
-     "\xB3\xFD\x06\x00\xB4\xE5\x06\x00\xB5\xED\x06\x00\xB6\xE5\x06\x00\xB7\x79"
-     "\x07\x00"
-     "\xA3\xA5\x06\x00\x82\x15\x00\x00\x81\xB5\x00\x00\x80\xB5\x00\x00\x1B\x6F"
-     "\x00\x80"
-     "\xA6\xC5\x06\x00\xA5\xD1\x06\x00\xBE\x50\x03\x00\xAB\x29\x06\x00\xAA\x21"
-     "\x06\x00"
-     "\x1F\x6F\x00\x80\x23\x6F\x00\x80\xAF\x29\x06\x00\xAE\x21\x06\x00\xAD\x29"
-     "\x06\x00"
-     "\xAC\x31\x06\x00\x27\x6F\x00\x80\xB3\xB9\x07\x00\x86\x68\x00\x00\x87\x6C"
-     "\x03\x00"
-     "\xB6\x45\x07\x00\x2B\x6F\x00\x80\x2F\x6F\x00\x80\xB5\x4D\x07\x00\xBA\x49"
-     "\x07\x00"
-     "\xBB\x51\x07\x00\x33\x6F\x00\x80\x37\x6F\x00\x80\xBE\x35\x07\x00\xBF\x3D"
-     "\x07\x00"
-     "\xBC\x41\x07\x00\xBD\x41\x07\x00\x3B\x6F\x00\x80\x3F\x6F\x00\x80\xEF\x98"
-     "\x04\x00"
-     "\x43\x6F\x00\x80\x47\x6F\x00\x80\x4B\x6F\x00\x80\x4F\x6F\x00\x80\x53\x6F"
-     "\x00\x80"
-     "\x57\x6F\x00\x80\x5B\x6F\x00\x80\x5F\x6F\x00\x80\x63\x6F\x00\x80\xE1\x70"
-     "\x05\x00"
-     "\x67\x6F\x00\x80\xE3\x4C\x04\x00\x6B\x6F\x00\x80\xA3\x3D\x06\x00\x6F\x6F"
-     "\x00\x80"
-     "\x73\x6F\x00\x80\x77\x6F\x00\x80\x7B\x6F\x00\x80\xA6\xC1\x06\x00\xA5\xC9"
-     "\x06\x00"
-     "\x7F\x6F\x00\x80\xAB\xD5\x06\x00\xAA\xCD\x06\x00\x83\x6F\x00\x80\x87\x6F"
-     "\x00\x80"
-     "\xAF\xB9\x06\x00\xAE\xB1\x06\x00\xAD\xC5\x06\x00\xAC\xC5\x06\x00\xA8\xED"
-     "\x06\x00"
-     "\xA9\xF1\x06\x00\xAA\xF5\x06\x00\xAB\x05\x06\x00\xAC\x01\x06\x00\xAD\x01"
-     "\x06\x00"
-     "\xAE\x01\x06\x00\xAF\x3D\x06\x00\x80\xC9\x00\x00\x81\xC9\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x8B\x6F\x00\x80\xBE\xFC\x01\x00\x84\xE0\x01\x00\x93\x6F\x00\x80\x97\x6F"
-     "\x00\x80"
-     "\xB8\x1D\x06\x00\xB9\x2D\x06\x00\xBA\x25\x06\x00\xBB\xE9\x01\x00\xBC\xF9"
-     "\x01\x00"
-     "\xBD\xF9\x01\x00\xBE\xE9\x01\x00\xBF\xE9\x01\x00\xB0\x49\x06\x00\xB1\x51"
-     "\x06\x00"
-     "\xB2\x51\x06\x00\xB3\x51\x06\x00\xB4\x35\x06\x00\xB5\x3D\x06\x00\xB6\x35"
-     "\x06\x00"
-     "\xB7\x2D\x06\x00\xA8\xAD\x02\x00\xA9\xC5\x02\x00\xAA\xDD\x02\x00\xAB\xD5"
-     "\x02\x00"
-     "\xAC\xF9\x02\x00\xAD\xF9\x02\x00\xAE\x19\x03\x00\xAF\x19\x03\x00\x9B\x6F"
-     "\x00\x80"
-     "\x9F\x6F\x00\x80\xA3\x6F\x00\x80\xA7\x6F\x00\x80\xAB\x6F\x00\x80\xAF\x6F"
-     "\x00\x80"
-     "\xB3\x6F\x00\x80\xB7\x6F\x00\x80\xB8\x6D\x03\x00\xB9\x71\x03\x00\xBA\x75"
-     "\x03\x00"
-     "\xBB\x09\x03\x00\xBC\x19\x03\x00\xBD\x19\x03\x00\xBE\x09\x03\x00\xBF\x01"
-     "\x03\x00"
-     "\xB0\x6D\x03\x00\xB1\x75\x03\x00\xB2\x7D\x03\x00\xB3\x75\x03\x00\xB4\x6D"
-     "\x03\x00"
-     "\xB5\x55\x03\x00\xB6\x5D\x03\x00\xB7\x55\x03\x00\xB5\x69\x02\x00\xBB\x6F"
-     "\x00\x80"
-     "\xBF\x6F\x00\x80\xB6\x61\x02\x00\x86\x30\x7C\x00\x87\xA4\x03\x00\xB3\x4D"
-     "\x02\x00"
-     "\xC3\x6F\x00\x80\xBD\xB9\x03\x00\xBC\xB1\x03\x00\xBF\x89\x03\x00\xBE\xB1"
-     "\x03\x00"
-     "\xC7\x6F\x00\x80\x3F\x6E\x00\x80\xBB\xA9\x03\x00\xBA\xB1\x03\x00\xA3\x15"
-     "\x02\x00"
-     "\xCB\x6F\x00\x80\xCF\x6F\x00\x80\xD3\x6F\x00\x80\x85\xD4\x71\x00\xA6\x39"
-     "\x02\x00"
-     "\xA5\x31\x02\x00\xD7\x6F\x00\x80\xAB\xF1\x03\x00\xAA\xE9\x03\x00\xDB\x6F"
-     "\x00\x80"
-     "\xDF\x6F\x00\x80\xAF\xD1\x03\x00\xAE\xE9\x03\x00\xAD\xE1\x03\x00\xAC\xE9"
-     "\x03\x00"
-     "\xE3\x6F\x00\x80\xE7\x6F\x00\x80\xEB\x6F\x00\x80\xEF\x6F\x00\x80\xF3\x6F"
-     "\x00\x80"
-     "\xEF\x5C\x00\x00\xF7\x6F\x00\x80\xFB\x6F\x00\x80\xBE\x60\x02\x00\xE3\x80"
-     "\x00\x00"
-     "\x84\xA4\x02\x00\xE1\xFC\x01\x00\x80\x21\x00\x00\x81\x2D\x00\x00\x82\x25"
-     "\x00\x00"
-     "\x03\x70\x00\x80\xB3\x61\x01\x00\x07\x70\x00\x80\x87\x64\x7D\x00\x86\x6C"
-     "\x7D\x00"
-     "\x0B\x70\x00\x80\xB6\x79\x01\x00\xB5\x71\x01\x00\x0F\x70\x00\x80\xBB\x5D"
-     "\x01\x00"
-     "\xBA\x5D\x01\x00\x13\x70\x00\x80\x17\x70\x00\x80\xBF\xC1\x01\x00\xBE\xE9"
-     "\x01\x00"
-     "\xBD\xF1\x01\x00\xBC\xF9\x01\x00\x1B\x70\x00\x80\x1F\x70\x00\x80\x23\x70"
-     "\x00\x80"
-     "\x27\x70\x00\x80\x2B\x70\x00\x80\x2F\x70\x00\x80\x33\x70\x00\x80\xEF\xF8"
-     "\x01\x00"
-     "\x84\x2C\x7C\x00\xE1\x3C\x06\x00\x37\x70\x00\x80\xE3\x30\x06\x00\x3B\x70"
-     "\x00\x80"
-     "\x3F\x70\x00\x80\x43\x70\x00\x80\x47\x70\x00\x80\x4B\x70\x00\x80\x4F\x70"
-     "\x00\x80"
-     "\x53\x70\x00\x80\xA3\xED\x01\x00\x57\x70\x00\x80\xA5\xFD\x01\x00\xA6\xF5"
-     "\x01\x00"
-     "\x5B\x70\x00\x80\x5F\x70\x00\x80\x63\x70\x00\x80\xAA\xD1\x01\x00\xAB\xD1"
-     "\x01\x00"
-     "\xAC\x75\x01\x00\xAD\x7D\x01\x00\xAE\x65\x01\x00\xAF\x4D\x01\x00\xA8\xB9"
-     "\x7D\x00"
-     "\xA9\xB9\x7D\x00\xAA\x29\x7E\x00\xAB\x29\x7E\x00\xAC\x39\x7E\x00\xAD\x39"
-     "\x7E\x00"
-     "\xAE\x29\x7E\x00\xAF\x29\x7E\x00\xFF\x6F\x00\x80\x82\x11\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x80\xC9\x7F\x00\x67\x70\x00\x80\x6B\x70\x00\x80\xBE\x34\x03\x00\x6F\x70"
-     "\x00\x80"
-     "\xB8\x8D\x7E\x00\xB9\x99\x7E\x00\xBA\xA1\x7E\x00\xBB\xA1\x7E\x00\xBC\x51"
-     "\x7F\x00"
-     "\xBD\x51\x7F\x00\xBE\x51\x7F\x00\xBF\x51\x7F\x00\xB0\x59\x7E\x00\xB1\x59"
-     "\x7E\x00"
-     "\xB2\xED\x7E\x00\xB3\xE5\x7E\x00\xB4\xED\x7E\x00\xB5\xC5\x7E\x00\xB6\xBD"
-     "\x7E\x00"
-     "\xB7\xB5\x7E\x00\xA8\xBD\x7F\x00\xA9\xC1\x7F\x00\xAA\xD9\x7F\x00\xAB\xD1"
-     "\x7F\x00"
-     "\xAC\xF9\x7F\x00\xAD\xF9\x7F\x00\xAE\x19\x7F\x00\xAF\x19\x7F\x00\x87\x60"
-     "\x03\x00"
-     "\x86\x04\x00\x00\x73\x70\x00\x80\x77\x70\x00\x80\x7B\x70\x00\x80\x7F\x70"
-     "\x00\x80"
-     "\x83\x70\x00\x80\x87\x70\x00\x80\xB8\x79\x7F\x00\xB9\x79\x7F\x00\xBA\x09"
-     "\x7F\x00"
-     "\xBB\x09\x7F\x00\xBC\x19\x7F\x00\xBD\x05\x7F\x00\xBE\x05\x7F\x00\xBF\x3D"
-     "\x7F\x00"
-     "\xB0\x69\x7F\x00\xB1\x69\x7F\x00\xB2\x79\x7F\x00\xB3\x79\x7F\x00\xB4\x69"
-     "\x7F\x00"
-     "\xB5\x5D\x7F\x00\xB6\x49\x7F\x00\xB7\x49\x7F\x00\x8B\x70\x00\x80\xEF\xDC"
-     "\x7C\x00"
-     "\x8F\x70\x00\x80\x93\x70\x00\x80\x97\x70\x00\x80\x9B\x70\x00\x80\x9F\x70"
-     "\x00\x80"
-     "\xA3\x70\x00\x80\xA7\x70\x00\x80\xAB\x70\x00\x80\xAF\x70\x00\x80\xB3\x70"
-     "\x00\x80"
-     "\xB7\x70\x00\x80\xE3\x0C\x7C\x00\xBB\x70\x00\x80\xE1\xB4\x7D\x00\xB3\x11"
-     "\x7E\x00"
-     "\xBF\x70\x00\x80\xC3\x70\x00\x80\xC7\x70\x00\x80\xCB\x70\x00\x80\xB6\x39"
-     "\x7E\x00"
-     "\xB5\x31\x7E\x00\xCF\x70\x00\x80\xBB\xED\x7E\x00\xBA\xED\x7E\x00\xD3\x70"
-     "\x00\x80"
-     "\xD7\x70\x00\x80\xBF\xD9\x7E\x00\xBE\xED\x7E\x00\xBD\xED\x7E\x00\xBC\xED"
-     "\x7E\x00"
-     "\xA3\xD1\x7E\x00\xDB\x70\x00\x80\x84\x88\x02\x00\xBE\x90\x01\x00\xE3\x70"
-     "\x00\x80"
-     "\xA6\xF9\x7E\x00\xA5\xF1\x7E\x00\xE7\x70\x00\x80\xAB\x2D\x7E\x00\xAA\x2D"
-     "\x7E\x00"
-     "\x86\x08\x00\x00\x87\x4C\x01\x00\xAF\x19\x7E\x00\xAE\x2D\x7E\x00\xAD\x2D"
-     "\x7E\x00"
-     "\xAC\x2D\x7E\x00\x80\xA9\x00\x00\x81\xA9\x00\x00\x82\x4D\x00\x00\xB3\x7D"
-     "\x01\x00"
-     "\xEB\x70\x00\x80\xB5\x6D\x01\x00\xB6\x65\x01\x00\xEF\x70\x00\x80\xF3\x70"
-     "\x00\x80"
-     "\xF7\x70\x00\x80\xBA\xE5\x00\x00\xBB\xFD\x00\x00\xBC\xE5\x00\x00\xBD\xED"
-     "\x00\x00"
-     "\xBE\xE5\x00\x00\xBF\xDD\x00\x00\xA8\xA9\x02\x00\xA9\xA9\x02\x00\xAA\xB9"
-     "\x02\x00"
-     "\xAB\xB9\x02\x00\xAC\xA9\x02\x00\xAD\xA9\x02\x00\xAE\x19\x03\x00\xAF\x15"
-     "\x03\x00"
-     "\xBE\xAC\x04\x00\xFB\x70\x00\x80\xFF\x70\x00\x80\xBE\xE0\x05\x00\x03\x71"
-     "\x00\x80"
-     "\x07\x71\x00\x80\x0B\x71\x00\x80\x0F\x71\x00\x80\xB8\x6D\x03\x00\xB9\x75"
-     "\x03\x00"
-     "\xBA\x71\x03\x00\xBB\x09\x03\x00\xBC\x19\x03\x00\xBD\x19\x03\x00\xBE\x09"
-     "\x03\x00"
-     "\xBF\x01\x03\x00\xB0\x6D\x03\x00\xB1\x75\x03\x00\xB2\x7D\x03\x00\xB3\x75"
-     "\x03\x00"
-     "\xB4\x6D\x03\x00\xB5\x55\x03\x00\xB6\x5D\x03\x00\xB7\x55\x03\x00\x13\x71"
-     "\x00\x80"
-     "\xA6\x39\x02\x00\xA5\x31\x02\x00\x17\x71\x00\x80\xA3\x21\x02\x00\x1B\x71"
-     "\x00\x80"
-     "\x1F\x71\x00\x80\x23\x71\x00\x80\xAF\x81\x03\x00\xAE\xB9\x03\x00\xAD\xB1"
-     "\x03\x00"
-     "\xAC\xB9\x03\x00\xAB\xA1\x03\x00\xAA\xB9\x03\x00\x8F\x6F\x00\x80\xBE\x74"
-     "\x06\x00"
-     "\x84\xC4\x03\x00\x27\x71\x00\x80\x2B\x71\x00\x80\x2F\x71\x00\x80\x33\x71"
-     "\x00\x80"
-     "\x37\x71\x00\x80\x3B\x71\x00\x80\x3F\x71\x00\x80\x80\x0D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x1D\x00\x00\x43\x71\x00\x80\x47\x71\x00\x80\x4B\x71\x00\x80\x87\xB8"
-     "\x04\x00"
-     "\x86\xDC\x04\x00\xA8\x91\x02\x00\xA9\x99\x02\x00\xAA\xD9\x02\x00\xAB\xD1"
-     "\x02\x00"
-     "\xAC\xF1\x02\x00\xAD\xF1\x02\x00\xAE\x3D\x01\x00\xAF\x35\x01\x00\x53\x71"
-     "\x00\x80"
-     "\x57\x71\x00\x80\x5B\x71\x00\x80\x5F\x71\x00\x80\x63\x71\x00\x80\x67\x71"
-     "\x00\x80"
-     "\x6B\x71\x00\x80\x6F\x71\x00\x80\xB8\xDD\x01\x00\xB9\x81\x01\x00\xBA\x81"
-     "\x01\x00"
-     "\xBB\x81\x01\x00\xBC\x81\x01\x00\xBD\x81\x01\x00\xBE\x81\x01\x00\xBF\x81"
-     "\x01\x00"
-     "\xB0\x51\x01\x00\xB1\x51\x01\x00\xB2\x51\x01\x00\xB3\x51\x01\x00\xB4\xF1"
-     "\x01\x00"
-     "\xB5\xF9\x01\x00\xB6\xED\x01\x00\xB7\xE5\x01\x00\xEF\xF4\x00\x00\x73\x71"
-     "\x00\x80"
-     "\x77\x71\x00\x80\xBE\x00\x08\x00\x7B\x71\x00\x80\x7F\x71\x00\x80\x84\x18"
-     "\x05\x00"
-     "\xEF\x5C\x01\x00\x83\x71\x00\x80\xE1\xF0\x7E\x00\x87\x71\x00\x80\xE3\x1C"
-     "\x7E\x00"
-     "\xE3\x28\x00\x00\x8B\x71\x00\x80\xE1\x04\x01\x00\x8F\x71\x00\x80\x85\xFC"
-     "\x29\x00"
-     "\x93\x71\x00\x80\xB3\x19\x02\x00\x97\x71\x00\x80\xB5\x31\x02\x00\x9B\x71"
-     "\x00\x80"
-     "\x9F\x71\x00\x80\xB6\x39\x02\x00\xA3\x71\x00\x80\xA7\x71\x00\x80\xBB\xDD"
-     "\x01\x00"
-     "\xBA\xDD\x01\x00\xBD\xCD\x01\x00\xBC\xCD\x01\x00\xBF\x75\x01\x00\xBE\x7D"
-     "\x01\x00"
-     "\xA3\xD1\x05\x00\x4F\x71\x00\x80\xAB\x71\x00\x80\xAF\x71\x00\x80\xB3\x71"
-     "\x00\x80"
-     "\xA6\xF1\x05\x00\xA5\xF9\x05\x00\xB7\x71\x00\x80\xAB\x15\x06\x00\xAA\x15"
-     "\x06\x00"
-     "\xBB\x71\x00\x80\xBF\x71\x00\x80\xAF\xBD\x06\x00\xAE\xB5\x06\x00\xAD\x05"
-     "\x06\x00"
-     "\xAC\x05\x06\x00\x80\x51\x00\x00\x81\x59\x00\x00\x82\x61\x00\x00\xB3\x65"
-     "\x07\x00"
-     "\xC3\x71\x00\x80\xB5\x65\x07\x00\xB6\x6D\x07\x00\xC7\x71\x00\x80\x86\x40"
-     "\x03\x00"
-     "\x87\xC4\x03\x00\xBA\xED\x07\x00\xBB\xE5\x07\x00\xBC\xE1\x07\x00\xBD\xE1"
-     "\x07\x00"
-     "\xBE\xD9\x07\x00\xBF\xD5\x07\x00\xA8\xA1\x07\x00\xA9\xA1\x07\x00\xAA\xA1"
-     "\x07\x00"
-     "\xAB\xA1\x07\x00\xAC\xA1\x07\x00\xAD\xA1\x07\x00\xAE\xA1\x07\x00\xAF\xA1"
-     "\x07\x00"
-     "\xCB\x71\x00\x80\xCF\x71\x00\x80\xD3\x71\x00\x80\xD7\x71\x00\x80\xDB\x71"
-     "\x00\x80"
-     "\xDF\x71\x00\x80\xE3\x71\x00\x80\xE7\x71\x00\x80\xB8\x05\x07\x00\xB9\x09"
-     "\x07\x00"
-     "\xBA\x1D\x07\x00\xBB\x15\x07\x00\xBC\x31\x07\x00\xBD\x31\x07\x00\xBE\xD1"
-     "\x00\x00"
-     "\xBF\xD1\x00\x00\xB0\x65\x07\x00\xB1\x6D\x07\x00\xB2\x79\x07\x00\xB3\x79"
-     "\x07\x00"
-     "\xB4\x51\x07\x00\xB5\x5D\x07\x00\xB6\x4D\x07\x00\xB7\x41\x07\x00\xEB\x71"
-     "\x00\x80"
-     "\xEF\x71\x00\x80\xF3\x71\x00\x80\xF7\x71\x00\x80\xEF\xC8\x05\x00\xFB\x71"
-     "\x00\x80"
-     "\xFF\x71\x00\x80\x03\x72\x00\x80\xE3\xF0\x05\x00\x07\x72\x00\x80\xE1\x98"
-     "\x06\x00"
-     "\x0B\x72\x00\x80\x0F\x72\x00\x80\x13\x72\x00\x80\x17\x72\x00\x80\x1B\x72"
-     "\x00\x80"
-     "\xA3\x69\x06\x00\x1F\x72\x00\x80\x23\x72\x00\x80\x27\x72\x00\x80\x2B\x72"
-     "\x00\x80"
-     "\xA6\x61\x06\x00\xA5\x69\x06\x00\x84\x90\x01\x00\xAB\xE9\x06\x00\xAA\xE1"
-     "\x06\x00"
-     "\x2F\x72\x00\x80\x33\x72\x00\x80\xAF\xD9\x06\x00\xAE\xD5\x06\x00\xAD\xED"
-     "\x06\x00"
-     "\xAC\xED\x06\x00\xA8\xD1\x01\x00\xA9\xD9\x01\x00\xAA\x29\x01\x00\xAB\x29"
-     "\x01\x00"
-     "\xAC\x39\x01\x00\xAD\x39\x01\x00\xAE\x29\x01\x00\xAF\x29\x01\x00\x80\xE9"
-     "\x00\x00"
-     "\x81\xE9\x00\x00\x82\xF9\x00\x00\x37\x72\x00\x80\x86\x10\x00\x00\x87\x74"
-     "\x01\x00"
-     "\x3B\x72\x00\x80\x3F\x72\x00\x80\xB8\xDD\x00\x00\xB9\xE1\x00\x00\xBA\xE5"
-     "\x00\x00"
-     "\xBB\xFD\x00\x00\xBC\xE5\x00\x00\xBD\xED\x00\x00\xBE\xE5\x00\x00\xBF\x99"
-     "\x00\x00"
-     "\xB0\x59\x01\x00\xB1\x59\x01\x00\xB2\xE9\x00\x00\xB3\xE9\x00\x00\xB4\xF9"
-     "\x00\x00"
-     "\xB5\xF9\x00\x00\xB6\xED\x00\x00\xB7\xE5\x00\x00\xB3\xA1\x02\x00\x43\x72"
-     "\x00\x80"
-     "\x47\x72\x00\x80\x4B\x72\x00\x80\x4F\x72\x00\x80\xB6\xB9\x02\x00\xB5\xB1"
-     "\x02\x00"
-     "\x53\x72\x00\x80\xBB\x9D\x02\x00\xBA\x9D\x02\x00\x57\x72\x00\x80\x5B\x72"
-     "\x00\x80"
-     "\xBF\x3D\x03\x00\xBE\x3D\x03\x00\xBD\x3D\x03\x00\xBC\x3D\x03\x00\x85\x04"
-     "\x2C\x00"
-     "\xA3\xE5\x02\x00\x5F\x72\x00\x80\x63\x72\x00\x80\xA6\xFD\x02\x00\x67\x72"
-     "\x00\x80"
-     "\x6B\x72\x00\x80\xA5\xF5\x02\x00\xAA\xD9\x02\x00\xAB\xD9\x02\x00\x6F\x72"
-     "\x00\x80"
-     "\x73\x72\x00\x80\xAE\x79\x03\x00\xAF\x79\x03\x00\xAC\x79\x03\x00\xAD\x79"
-     "\x03\x00"
-     "\x84\x90\x03\x00\x77\x72\x00\x80\x7B\x72\x00\x80\x84\xC0\x0C\x00\x7F\x72"
-     "\x00\x80"
-     "\x83\x72\x00\x80\x87\x72\x00\x80\x8B\x72\x00\x80\x8F\x72\x00\x80\x93\x72"
-     "\x00\x80"
-     "\x97\x72\x00\x80\xDF\x70\x00\x80\x9B\x72\x00\x80\xA3\x72\x00\x80\xA7\x72"
-     "\x00\x80"
-     "\xAB\x72\x00\x80\x80\x65\x00\x00\x81\x6D\x00\x00\x82\x79\x00\x00\xAF\x72"
-     "\x00\x80"
-     "\xB3\x72\x00\x80\xEF\xA0\x07\x00\x87\x6C\x03\x00\x86\x5C\x0C\x00\xB7\x72"
-     "\x00\x80"
-     "\xE3\x74\x02\x00\xBB\x72\x00\x80\xE1\x44\x01\x00\xBF\x72\x00\x80\xC3\x72"
-     "\x00\x80"
-     "\xC7\x72\x00\x80\xCB\x72\x00\x80\xA8\x6D\x02\x00\xA9\xA9\x02\x00\xAA\xB1"
-     "\x02\x00"
-     "\xAB\xB1\x02\x00\xAC\x15\x01\x00\xAD\x1D\x01\x00\xAE\x15\x01\x00\xAF\x0D"
-     "\x01\x00"
-     "\xBE\x68\x0D\x00\xCF\x72\x00\x80\xD3\x72\x00\x80\xD7\x72\x00\x80\xDB\x72"
-     "\x00\x80"
-     "\xDF\x72\x00\x80\xE3\x72\x00\x80\xE7\x72\x00\x80\xB8\x2D\x01\x00\xB9\xC1"
-     "\x01\x00"
-     "\xBA\xC1\x01\x00\xBB\xDD\x01\x00\xBC\xC5\x01\x00\xBD\xCD\x01\x00\xBE\xC5"
-     "\x01\x00"
-     "\xBF\xFD\x01\x00\xB0\x7D\x01\x00\xB1\x4D\x01\x00\xB2\x45\x01\x00\xB3\x25"
-     "\x01\x00"
-     "\xB4\x3D\x01\x00\xB5\x21\x01\x00\xB6\x21\x01\x00\xB7\x15\x01\x00\xE1\x64"
-     "\x07\x00"
-     "\xEB\x72\x00\x80\xE3\x00\x06\x00\xEF\x72\x00\x80\xEF\xD4\x07\x00\xF3\x72"
-     "\x00\x80"
-     "\xF7\x72\x00\x80\xFB\x72\x00\x80\xE3\xEC\x00\x00\xFF\x72\x00\x80\xE1\xD0"
-     "\x01\x00"
-     "\x03\x73\x00\x80\x81\x71\x00\x00\x80\x19\x00\x00\xEF\x6C\x07\x00\x82\x6D"
-     "\x00\x00"
-     "\x07\x73\x00\x80\xB5\xC1\x01\x00\xB6\xC1\x01\x00\x84\x4C\x0C\x00\x0B\x73"
-     "\x00\x80"
-     "\x13\x73\x00\x80\x17\x73\x00\x80\xB3\x65\x02\x00\xBC\xCD\x01\x00\xBD\xAD"
-     "\x01\x00"
-     "\xBE\xA5\x01\x00\xBF\xB9\x01\x00\x86\xD0\x0C\x00\x87\xB4\x0C\x00\xBA\xE5"
-     "\x01\x00"
-     "\xBB\xD5\x01\x00\x9F\x72\x00\x80\x0F\x73\x00\x80\x1B\x73\x00\x80\x1F\x73"
-     "\x00\x80"
-     "\x23\x73\x00\x80\x27\x73\x00\x80\x2B\x73\x00\x80\x2F\x73\x00\x80\xA8\x95"
-     "\x0E\x00"
-     "\xA9\x9D\x0E\x00\xAA\x95\x0E\x00\xAB\xE1\x0E\x00\xAC\xE1\x0E\x00\xAD\xD5"
-     "\x0E\x00"
-     "\xAE\xD1\x0E\x00\xAF\xCD\x0E\x00\xB0\xB5\x0E\x00\xB1\xBD\x0E\x00\xB2\xB5"
-     "\x0E\x00"
-     "\xB3\x69\x0F\x00\xB4\x79\x0F\x00\xB5\x79\x0F\x00\xB6\x69\x0F\x00\xB7\x69"
-     "\x0F\x00"
-     "\xB8\x59\x0F\x00\xB9\x59\x0F\x00\xBA\xED\x0F\x00\xBB\xE1\x0F\x00\xBC\xE5"
-     "\x0F\x00"
-     "\xBD\xED\x0F\x00\xBE\xE5\x0F\x00\xBF\xD9\x0F\x00\xB3\x3D\x0E\x00\x33\x73"
-     "\x00\x80"
-     "\x37\x73\x00\x80\x3B\x73\x00\x80\x3F\x73\x00\x80\xB6\x75\x0E\x00\xB5\x49"
-     "\x0E\x00"
-     "\x43\x73\x00\x80\xBB\x59\x0E\x00\xBA\x51\x0E\x00\x47\x73\x00\x80\x4B\x73"
-     "\x00\x80"
-     "\xBF\x9D\x0F\x00\xBE\x9D\x0F\x00\xBD\x9D\x0F\x00\xBC\x41\x0E\x00\x4F\x73"
-     "\x00\x80"
-     "\xA3\x79\x0E\x00\x53\x73\x00\x80\x57\x73\x00\x80\xA6\x31\x0E\x00\x5B\x73"
-     "\x00\x80"
-     "\x5F\x73\x00\x80\xA5\x0D\x0E\x00\xAA\x15\x0E\x00\xAB\x1D\x0E\x00\x63\x73"
-     "\x00\x80"
-     "\x67\x73\x00\x80\xAE\xD9\x0F\x00\xAF\xD9\x0F\x00\xAC\x05\x0E\x00\xAD\xD9"
-     "\x0F\x00"
-     "\xA3\xB9\x0C\x00\x6B\x73\x00\x80\x6F\x73\x00\x80\x73\x73\x00\x80\x77\x73"
-     "\x00\x80"
-     "\xA6\x1D\x0F\x00\xA5\x1D\x0F\x00\x7B\x73\x00\x80\xAB\x09\x0F\x00\xAA\x39"
-     "\x0F\x00"
-     "\xBE\xA8\x01\x00\x7F\x73\x00\x80\xAF\x65\x0F\x00\xAE\x79\x0F\x00\xAD\x71"
-     "\x0F\x00"
-     "\xAC\x11\x0F\x00\x80\x4D\x00\x00\x81\x55\x00\x00\x82\x55\x00\x00\xB3\x45"
-     "\x01\x00"
-     "\x83\x73\x00\x80\xB5\x45\x01\x00\xB6\x4D\x01\x00\x87\x73\x00\x80\x86\xC0"
-     "\x00\x00"
-     "\x87\x14\x01\x00\xBA\xED\x01\x00\xBB\xE5\x01\x00\xBC\xFD\x01\x00\xBD\xE5"
-     "\x01\x00"
-     "\xBE\xE5\x01\x00\xBF\xD5\x01\x00\xA8\x55\x0E\x00\xA9\x5D\x0E\x00\xAA\x81"
-     "\x01\x00"
-     "\xAB\x81\x01\x00\xAC\x81\x01\x00\xAD\x81\x01\x00\xAE\x81\x01\x00\xAF\x81"
-     "\x01\x00"
-     "\x84\x8C\x01\x00\x8B\x73\x00\x80\x8F\x73\x00\x80\x93\x73\x00\x80\x97\x73"
-     "\x00\x80"
-     "\x9B\x73\x00\x80\x9F\x73\x00\x80\xA3\x73\x00\x80\xB8\x09\x01\x00\xB9\x09"
-     "\x01\x00"
-     "\xBA\x1D\x01\x00\xBB\x15\x01\x00\xBC\x39\x01\x00\xBD\x39\x01\x00\xBE\xDD"
-     "\x00\x00"
-     "\xBF\xD5\x00\x00\xB0\x65\x01\x00\xB1\x6D\x01\x00\xB2\x65\x01\x00\xB3\x7D"
-     "\x01\x00"
-     "\xB4\x6D\x01\x00\xB5\x5D\x01\x00\xB6\x55\x01\x00\xB7\x3D\x01\x00\xA7\x73"
-     "\x00\x80"
-     "\xAB\x73\x00\x80\xAF\x73\x00\x80\xA3\x05\x02\x00\xB3\x73\x00\x80\xA5\x05"
-     "\x02\x00"
-     "\xA6\x0D\x02\x00\xB7\x73\x00\x80\xBB\x73\x00\x80\xBF\x73\x00\x80\xAA\xAD"
-     "\x02\x00"
-     "\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xA5\x02\x00\xAE\xA5\x02\x00\xAF\x95"
-     "\x02\x00"
-     "\xC3\x73\x00\x80\xC7\x73\x00\x80\xBE\x04\x03\x00\x84\x00\x04\x00\xCB\x73"
-     "\x00\x80"
-     "\xCF\x73\x00\x80\xD3\x73\x00\x80\xD7\x73\x00\x80\xDB\x73\x00\x80\xDF\x73"
-     "\x00\x80"
-     "\xE3\x73\x00\x80\xE7\x73\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xEB\x73\x00\x80\x86\x20\x04\x00\x87\x14\x03\x00\xF3\x73\x00\x80\xF7\x73"
-     "\x00\x80"
-     "\xEF\x20\x02\x00\xFB\x73\x00\x80\xFF\x73\x00\x80\x03\x74\x00\x80\xE3\x10"
-     "\x01\x00"
-     "\x07\x74\x00\x80\xE1\xD0\x01\x00\x0B\x74\x00\x80\x0F\x74\x00\x80\x13\x74"
-     "\x00\x80"
-     "\x17\x74\x00\x80\x1B\x74\x00\x80\xBB\xC1\x00\x00\xBA\x79\x03\x00\x1F\x74"
-     "\x00\x80"
-     "\x23\x74\x00\x80\xBF\xCD\x00\x00\xBE\xC5\x00\x00\xBD\xD1\x00\x00\xBC\xD9"
-     "\x00\x00"
-     "\xB3\x45\x03\x00\x27\x74\x00\x80\x2B\x74\x00\x80\x2F\x74\x00\x80\x33\x74"
-     "\x00\x80"
-     "\xB6\x45\x03\x00\xB5\x55\x03\x00\x37\x74\x00\x80\xA8\x51\x02\x00\xA9\x59"
-     "\x02\x00"
-     "\xAA\xAD\x02\x00\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xA5\x02\x00\xAE\xAD"
-     "\x02\x00"
-     "\xAF\xA5\x02\x00\xBE\x68\x05\x00\x3B\x74\x00\x80\x3F\x74\x00\x80\x43\x74"
-     "\x00\x80"
-     "\x47\x74\x00\x80\x4B\x74\x00\x80\x4F\x74\x00\x80\x53\x74\x00\x80\xB8\x01"
-     "\x01\x00"
-     "\xB9\x01\x01\x00\xBA\x19\x01\x00\xBB\x19\x01\x00\xBC\x09\x01\x00\xBD\x09"
-     "\x01\x00"
-     "\xBE\x39\x01\x00\xBF\x39\x01\x00\xB0\xDD\x02\x00\xB1\x65\x01\x00\xB2\x6D"
-     "\x01\x00"
-     "\xB3\x65\x01\x00\xB4\x61\x01\x00\xB5\x69\x01\x00\xB6\x59\x01\x00\xB7\x59"
-     "\x01\x00"
-     "\xE1\x68\x0E\x00\xE1\x40\x0F\x00\xE3\x54\x0E\x00\xE3\x10\x0F\x00\x80\xAD"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x39\x00\x00\x57\x74\x00\x80\x5B\x74\x00\x80\x63\x74"
-     "\x00\x80"
-     "\x67\x74\x00\x80\x6B\x74\x00\x80\xBE\x78\x06\x00\x6F\x74\x00\x80\xEF\x68"
-     "\x00\x00"
-     "\xEF\x64\x0E\x00\xA3\x65\x02\x00\x73\x74\x00\x80\x86\x88\x07\x00\x87\xA8"
-     "\x04\x00"
-     "\x84\x1C\x06\x00\xA6\x65\x02\x00\xA5\x75\x02\x00\x77\x74\x00\x80\xAB\xE1"
-     "\x01\x00"
-     "\xAA\x59\x02\x00\x7B\x74\x00\x80\x7F\x74\x00\x80\xAF\xED\x01\x00\xAE\xE5"
-     "\x01\x00"
-     "\xAD\xF1\x01\x00\xAC\xF9\x01\x00\xA8\x89\x06\x00\xA9\x89\x06\x00\xAA\x9D"
-     "\x06\x00"
-     "\xAB\x95\x06\x00\xAC\x8D\x06\x00\xAD\xB1\x06\x00\xAE\xB5\x06\x00\xAF\xAD"
-     "\x06\x00"
-     "\x5F\x74\x00\x80\x83\x74\x00\x80\x87\x74\x00\x80\x8B\x74\x00\x80\x8F\x74"
-     "\x00\x80"
-     "\x93\x74\x00\x80\x97\x74\x00\x80\x9B\x74\x00\x80\xB8\xB1\x06\x00\xB9\xB9"
-     "\x06\x00"
-     "\xBA\x6D\x07\x00\xBB\x65\x07\x00\xBC\x7D\x07\x00\xBD\x65\x07\x00\xBE\x6D"
-     "\x07\x00"
-     "\xBF\x65\x07\x00\xB0\xD5\x06\x00\xB1\xDD\x06\x00\xB2\xD5\x06\x00\xB3\xA5"
-     "\x06\x00"
-     "\xB4\xBD\x06\x00\xB5\x91\x06\x00\xB6\x91\x06\x00\xB7\x91\x06\x00\xB3\x15"
-     "\x06\x00"
-     "\x9F\x74\x00\x80\xA3\x74\x00\x80\xA7\x74\x00\x80\xAB\x74\x00\x80\xB6\x31"
-     "\x06\x00"
-     "\xB5\x05\x06\x00\xAF\x74\x00\x80\xBB\x19\x06\x00\xBA\x11\x06\x00\xB3\x74"
-     "\x00\x80"
-     "\xB7\x74\x00\x80\xBF\x55\x06\x00\xBE\x79\x06\x00\xBD\x71\x06\x00\xBC\x01"
-     "\x06\x00"
-     "\xBB\x74\x00\x80\xA3\x51\x06\x00\xBF\x74\x00\x80\xC3\x74\x00\x80\xA6\x75"
-     "\x06\x00"
-     "\xC7\x74\x00\x80\xCB\x74\x00\x80\xA5\x41\x06\x00\xAA\x55\x06\x00\xAB\x5D"
-     "\x06\x00"
-     "\x86\x48\x1E\x00\x87\x3C\x00\x00\xAE\x3D\x06\x00\xAF\x11\x06\x00\xAC\x45"
-     "\x06\x00"
-     "\xAD\x35\x06\x00\x9F\x69\x08\x00\x9E\x29\x09\x00\x9D\x0D\x09\x00\x9C\xA1"
-     "\x0B\x00"
-     "\x9B\x2D\x0D\x00\x9A\x35\x0D\x00\x99\xA1\x0F\x00\x98\xB5\x71\x00\x97\x01"
-     "\x71\x00"
-     "\x96\x91\x73\x00\x95\xDD\x75\x00\x94\x0D\x75\x00\x93\xB1\x77\x00\x92\x05"
-     "\x78\x00"
-     "\x91\x3D\x79\x00\x90\x3D\x79\x00\x8F\x71\x7D\x00\xE1\xDC\x05\x00\xD3\x74"
-     "\x00\x80"
-     "\xE3\x10\x05\x00\xE1\x94\x71\x00\xD7\x74\x00\x80\xE3\x8C\x0C\x00\xDB\x74"
-     "\x00\x80"
-     "\x81\xB5\x01\x00\x80\xA9\x01\x00\xEF\xB4\x78\x00\x82\xB5\x01\x00\xDF\x74"
-     "\x00\x80"
-     "\xE3\x74\x00\x80\xE7\x74\x00\x80\xEF\xA4\x04\x00\xA1\xA1\x01\x00\xEB\x74"
-     "\x00\x80"
-     "\xA3\xA9\x1D\x00\xA2\x01\x1C\x00\xA5\x39\x1C\x00\xA4\xA5\x1D\x00\xA7\xC1"
-     "\x19\x00"
-     "\xA6\x9D\x19\x00\xA9\xA5\x15\x00\xA8\x01\x14\x00\xAB\x85\x14\x00\xAA\xA5"
-     "\x15\x00"
-     "\xAD\x91\x11\x00\xAC\x01\x10\x00\xAF\x01\x2C\x00\xAE\x89\x10\x00\xB1\x51"
-     "\x2D\x00"
-     "\xB0\xE5\x2D\x00\xB3\xE9\x29\x00\xB2\x01\x28\x00\xB5\x01\x28\x00\xB4\xE9"
-     "\x29\x00"
-     "\xEF\x74\x00\x80\xF3\x74\x00\x80\xF7\x74\x00\x80\xFB\x74\x00\x80\xFF\x74"
-     "\x00\x80"
-     "\x03\x75\x00\x80\xA3\x31\x03\x00\xA2\xD9\x04\x00\xA1\xFD\x05\x00\xA0\xFD"
-     "\x05\x00"
-     "\x82\x31\x7E\x00\x83\x05\x7A\x00\x07\x75\x00\x80\x0B\x75\x00\x80\x86\x91"
-     "\x74\x00"
-     "\x87\x35\x76\x00\x84\x1D\x7A\x00\x85\x2D\x7B\x00\x8A\xD5\x76\x00\x8B\x91"
-     "\x70\x00"
-     "\x0F\x75\x00\x80\x13\x75\x00\x80\x8E\x89\x0C\x00\x8F\x15\x0E\x00\x8C\x19"
-     "\x72\x00"
-     "\x8D\x95\x72\x00\x92\x7D\x0E\x00\x93\xE9\x08\x00\x84\x44\x03\x00\xBE\x40"
-     "\x1C\x00"
-     "\x96\x39\x0B\x00\x97\x55\x06\x00\x94\x5D\x0A\x00\x95\x71\x0A\x00\x9A\x65"
-     "\x06\x00"
-     "\x9B\x49\x07\x00\x17\x75\x00\x80\x1B\x75\x00\x80\x1F\x75\x00\x80\x23\x75"
-     "\x00\x80"
-     "\x9C\xB9\x00\x00\x27\x75\x00\x80\xB0\xA5\x02\x00\xB1\xAD\x02\x00\xB2\xA5"
-     "\x02\x00"
-     "\xB3\xBD\x02\x00\xB4\xA5\x02\x00\xB5\xAD\x02\x00\xB6\xA5\x02\x00\xB7\x19"
-     "\x02\x00"
-     "\xB8\x29\x02\x00\xB9\x29\x02\x00\xBA\x39\x02\x00\xBB\x31\x02\x00\xBC\x19"
-     "\x02\x00"
-     "\xBD\x19\x02\x00\xBE\x79\x02\x00\xBF\x75\x02\x00\xEF\x73\x00\x80\x2B\x75"
-     "\x00\x80"
-     "\x2F\x75\x00\x80\x33\x75\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xD1"
-     "\x01\x00"
-     "\x37\x75\x00\x80\xA8\x45\x05\x00\xA9\x4D\x05\x00\xAA\x45\x05\x00\xAB\x59"
-     "\x05\x00"
-     "\xAC\x35\x05\x00\xAD\x29\x05\x00\xAE\x11\x05\x00\xAF\x11\x05\x00\xB3\xAD"
-     "\x04\x00"
-     "\x84\x0C\x1D\x00\x3B\x75\x00\x80\x43\x75\x00\x80\x47\x75\x00\x80\xB6\xC9"
-     "\x04\x00"
-     "\xB5\xDD\x04\x00\x4B\x75\x00\x80\xBB\x09\x03\x00\xBA\x01\x03\x00\x86\x88"
-     "\x1C\x00"
-     "\x87\x00\x1D\x00\xBF\x09\x03\x00\xBE\x01\x03\x00\xBD\x09\x03\x00\xBC\x11"
-     "\x03\x00"
-     "\x4F\x75\x00\x80\xA3\xE9\x04\x00\x53\x75\x00\x80\x57\x75\x00\x80\xA6\x8D"
-     "\x04\x00"
-     "\x5B\x75\x00\x80\x5F\x75\x00\x80\xA5\x99\x04\x00\xAA\x45\x03\x00\xAB\x4D"
-     "\x03\x00"
-     "\x63\x75\x00\x80\x67\x75\x00\x80\xAE\x45\x03\x00\xAF\x4D\x03\x00\xAC\x55"
-     "\x03\x00"
-     "\xAD\x4D\x03\x00\x6B\x75\x00\x80\x6F\x75\x00\x80\x73\x75\x00\x80\x77\x75"
-     "\x00\x80"
-     "\x7B\x75\x00\x80\x7F\x75\x00\x80\x83\x75\x00\x80\xEF\xD4\x06\x00\x87\x75"
-     "\x00\x80"
-     "\xE1\x74\x06\x00\x8B\x75\x00\x80\xE3\x6C\x06\x00\x8F\x75\x00\x80\x93\x75"
-     "\x00\x80"
-     "\x97\x75\x00\x80\x9B\x75\x00\x80\xB3\x9D\x02\x00\x9F\x75\x00\x80\xA3\x75"
-     "\x00\x80"
-     "\xA7\x75\x00\x80\xAB\x75\x00\x80\xB6\xAD\x02\x00\xB5\xAD\x02\x00\xBE\xD0"
-     "\x1C\x00"
-     "\xBB\x01\x02\x00\xBA\x39\x02\x00\x86\xCC\x1F\x00\x87\x2C\x1C\x00\xBF\xF9"
-     "\x01\x00"
-     "\xBE\x19\x02\x00\xBD\x11\x02\x00\xBC\x19\x02\x00\xA8\x49\x1D\x00\xA9\x51"
-     "\x1D\x00"
-     "\xAA\x55\x1D\x00\xAB\x7D\x1D\x00\xAC\x65\x1D\x00\xAD\x6D\x1D\x00\xAE\x9D"
-     "\x1D\x00"
-     "\xAF\xC5\x1D\x00\x3F\x75\x00\x80\x82\xF5\x1F\x00\x81\xE5\x1F\x00\x80\xE5"
-     "\x1F\x00"
-     "\xAF\x75\x00\x80\xB3\x75\x00\x80\xB7\x75\x00\x80\xBB\x75\x00\x80\xB8\x6D"
-     "\x1E\x00"
-     "\xB9\x2D\x1E\x00\xBA\x25\x1E\x00\xBB\x3D\x1E\x00\xBC\x25\x1E\x00\xBD\x29"
-     "\x1E\x00"
-     "\xBE\x9D\x1E\x00\xBF\x95\x1E\x00\xB0\xBD\x1D\x00\xB1\x85\x1D\x00\xB2\x8D"
-     "\x1D\x00"
-     "\xB3\x85\x1D\x00\xB4\x9D\x1D\x00\xB5\x7D\x1E\x00\xB6\x75\x1E\x00\xB7\x55"
-     "\x1E\x00"
-     "\xA3\xDD\x1D\x00\xBF\x75\x00\x80\xC3\x75\x00\x80\xC7\x75\x00\x80\xCB\x75"
-     "\x00\x80"
-     "\xA6\xED\x1D\x00\xA5\xED\x1D\x00\xCF\x75\x00\x80\xAB\x41\x1D\x00\xAA\x79"
-     "\x1D\x00"
-     "\xD3\x75\x00\x80\xD7\x75\x00\x80\xAF\xB9\x1E\x00\xAE\x59\x1D\x00\xAD\x51"
-     "\x1D\x00"
-     "\xAC\x59\x1D\x00\xDB\x75\x00\x80\xB3\xD9\x1F\x00\xDF\x75\x00\x80\xE3\x75"
-     "\x00\x80"
-     "\xB6\x9D\x1F\x00\xE7\x75\x00\x80\xEB\x75\x00\x80\xB5\x95\x1F\x00\xBA\x89"
-     "\x1F\x00"
-     "\xBB\x89\x1F\x00\xEF\x75\x00\x80\xF3\x75\x00\x80\xBE\x79\x1F\x00\xBF\x79"
-     "\x1F\x00"
-     "\xBC\x79\x1F\x00\xBD\x79\x1F\x00\xA8\x0D\x1E\x00\xA9\x15\x1E\x00\xAA\x15"
-     "\x1E\x00"
-     "\xAB\x4D\x1E\x00\xAC\x59\x1E\x00\xAD\x59\x1E\x00\xAE\x4D\x1E\x00\xAF\x45"
-     "\x1E\x00"
-     "\xF7\x75\x00\x80\xFB\x75\x00\x80\xFF\x75\x00\x80\x03\x76\x00\x80\x07\x76"
-     "\x00\x80"
-     "\x0B\x76\x00\x80\x0F\x76\x00\x80\x13\x76\x00\x80\xB8\xD9\x1E\x00\xB9\xE5"
-     "\x1E\x00"
-     "\xBA\xE1\x1E\x00\xBB\xFD\x1E\x00\xBC\xE5\x1E\x00\xBD\xED\x1E\x00\xBE\xE5"
-     "\x1E\x00"
-     "\xBF\x99\x1E\x00\xB0\x09\x1E\x00\xB1\x09\x1E\x00\xB2\xE9\x1E\x00\xB3\xE9"
-     "\x1E\x00"
-     "\xB4\xF9\x1E\x00\xB5\xF9\x1E\x00\xB6\xE9\x1E\x00\xB7\xE9\x1E\x00\xA3\x9D"
-     "\x1E\x00"
-     "\x17\x76\x00\x80\x1B\x76\x00\x80\xCF\x74\x00\x80\x84\x90\x02\x00\xA6\xD9"
-     "\x1E\x00"
-     "\xA5\xD1\x1E\x00\xBE\x58\x01\x00\xAB\xCD\x1E\x00\xAA\xCD\x1E\x00\x1F\x76"
-     "\x00\x80"
-     "\x23\x76\x00\x80\xAF\x3D\x1E\x00\xAE\x3D\x1E\x00\xAD\x3D\x1E\x00\xAC\x3D"
-     "\x1E\x00"
-     "\x80\xFD\x00\x00\x81\x41\x00\x00\x82\x41\x00\x00\xB3\xF1\x01\x00\x27\x76"
-     "\x00\x80"
-     "\xB5\x5D\x01\x00\xB6\x55\x01\x00\x2B\x76\x00\x80\x86\x40\x00\x00\x87\x64"
-     "\x00\x00"
-     "\xBA\x41\x01\x00\xBB\x41\x01\x00\xBC\x31\x01\x00\xBD\x3D\x01\x00\xBE\x35"
-     "\x01\x00"
-     "\xBF\x29\x01\x00\xE3\x58\x1E\x00\x2F\x76\x00\x80\xE1\xF8\x1C\x00\x84\x00"
-     "\x03\x00"
-     "\x33\x76\x00\x80\xEF\x4C\x02\x00\x37\x76\x00\x80\x3B\x76\x00\x80\x3F\x76"
-     "\x00\x80"
-     "\xE3\x38\x02\x00\x43\x76\x00\x80\xE1\xD4\x01\x00\xEF\x84\x1E\x00\x47\x76"
-     "\x00\x80"
-     "\x4B\x76\x00\x80\x4F\x76\x00\x80\x53\x76\x00\x80\x57\x76\x00\x80\xA3\x75"
-     "\x02\x00"
-     "\x5B\x76\x00\x80\xA5\xD9\x02\x00\x5F\x76\x00\x80\x63\x76\x00\x80\xA6\xD1"
-     "\x02\x00"
-     "\x67\x76\x00\x80\x6B\x76\x00\x80\xAB\xC5\x02\x00\xAA\xC5\x02\x00\xAD\xB9"
-     "\x02\x00"
-     "\xAC\xB5\x02\x00\xAF\xAD\x02\x00\xAE\xB1\x02\x00\xA9\xB9\x02\x00\xA8\xB1"
-     "\x02\x00"
-     "\xAB\x0D\x02\x00\xAA\xB1\x02\x00\xAD\x1D\x02\x00\xAC\x1D\x02\x00\xAF\x05"
-     "\x02\x00"
-     "\xAE\x0D\x02\x00\xBE\x00\x04\x00\x6F\x76\x00\x80\x73\x76\x00\x80\x77\x76"
-     "\x00\x80"
-     "\x7B\x76\x00\x80\x7F\x76\x00\x80\x83\x76\x00\x80\x87\x76\x00\x80\xB9\xE1"
-     "\x03\x00"
-     "\xB8\x29\x02\x00\xBB\xE1\x03\x00\xBA\xE9\x03\x00\xBD\xE1\x03\x00\xBC\xF9"
-     "\x03\x00"
-     "\xBF\xE1\x03\x00\xBE\xE9\x03\x00\xB1\x4D\x02\x00\xB0\x4D\x02\x00\xB3\x21"
-     "\x02\x00"
-     "\xB2\x2D\x02\x00\xB5\x25\x02\x00\xB4\x39\x02\x00\xB7\x11\x02\x00\xB6\x25"
-     "\x02\x00"
-     "\xA8\xAD\x02\x00\xA9\xD5\x02\x00\xAA\xD5\x02\x00\xAB\x09\x01\x00\xAC\x01"
-     "\x01\x00"
-     "\xAD\x0D\x01\x00\xAE\x05\x01\x00\xAF\x3D\x01\x00\x8B\x76\x00\x80\x93\x76"
-     "\x00\x80"
-     "\x97\x76\x00\x80\x9B\x76\x00\x80\x9F\x76\x00\x80\xA3\x76\x00\x80\xA7\x76"
-     "\x00\x80"
-     "\xAB\x76\x00\x80\xB8\x19\x01\x00\xB9\x25\x01\x00\xBA\x2D\x01\x00\xBB\xE9"
-     "\x01\x00"
-     "\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\xE9\x01\x00\xBF\xE9\x01\x00\xB0\x49"
-     "\x01\x00"
-     "\xB1\x51\x01\x00\xB2\x51\x01\x00\xB3\x51\x01\x00\xB4\x31\x01\x00\xB5\x31"
-     "\x01\x00"
-     "\xB6\x35\x01\x00\xB7\x29\x01\x00\x80\x39\x01\x00\x81\x39\x01\x00\x82\xA9"
-     "\x00\x00"
-     "\xEF\xE8\x00\x00\xAF\x76\x00\x80\xB3\x76\x00\x80\xB7\x76\x00\x80\xEF\xB8"
-     "\x1E\x00"
-     "\x84\x6C\x04\x00\xE1\x94\x1E\x00\xBB\x76\x00\x80\xE3\xB8\x1E\x00\xBF\x76"
-     "\x00\x80"
-     "\xE1\x44\x1E\x00\xC3\x76\x00\x80\xE3\x3C\x00\x00\xB3\x1D\x02\x00\xC7\x76"
-     "\x00\x80"
-     "\x87\x88\x05\x00\x86\xEC\x05\x00\xCB\x76\x00\x80\xB6\xCD\x01\x00\xB5\xC9"
-     "\x01\x00"
-     "\xCF\x76\x00\x80\xBB\xD5\x01\x00\xBA\xCD\x01\x00\xD3\x76\x00\x80\xD7\x76"
-     "\x00\x80"
-     "\xBF\xB5\x01\x00\xBE\xB5\x01\x00\xBD\xC5\x01\x00\xBC\xC5\x01\x00\x8F\x76"
-     "\x00\x80"
-     "\xDB\x76\x00\x80\xAA\x21\x06\x00\xAB\x39\x06\x00\xAC\x29\x06\x00\xAD\x29"
-     "\x06\x00"
-     "\xAE\x59\x06\x00\xAF\x59\x06\x00\xDF\x76\x00\x80\xE3\x76\x00\x80\xE7\x76"
-     "\x00\x80"
-     "\xA3\xF1\x05\x00\xEB\x76\x00\x80\xA5\x25\x06\x00\xA6\x21\x06\x00\xEF\x76"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\xF3\x76\x00\x80\xF7\x76"
-     "\x00\x80"
-     "\xFB\x76\x00\x80\xFF\x76\x00\x80\xBE\x94\x03\x00\x86\x40\x03\x00\x87\xC4"
-     "\x03\x00"
-     "\x03\x77\x00\x80\x07\x77\x00\x80\x0B\x77\x00\x80\x0F\x77\x00\x80\x13\x77"
-     "\x00\x80"
-     "\x17\x77\x00\x80\xA8\xA5\x07\x00\xA9\xAD\x07\x00\xAA\xA5\x07\x00\xAB\xBD"
-     "\x07\x00"
-     "\xAC\xA5\x07\x00\xAD\xAD\x07\x00\xAE\xA5\x07\x00\xAF\x0D\x07\x00\x1B\x77"
-     "\x00\x80"
-     "\x1F\x77\x00\x80\x23\x77\x00\x80\x27\x77\x00\x80\x2B\x77\x00\x80\x2F\x77"
-     "\x00\x80"
-     "\x33\x77\x00\x80\x37\x77\x00\x80\xB8\x29\x07\x00\xB9\x29\x07\x00\xBA\xCD"
-     "\x00\x00"
-     "\xBB\xED\x00\x00\xBC\xF5\x00\x00\xBD\xFD\x00\x00\xBE\xE9\x00\x00\xBF\xE1"
-     "\x00\x00"
-     "\xB0\x75\x07\x00\xB1\x7D\x07\x00\xB2\x41\x07\x00\xB3\x5D\x07\x00\xB4\x49"
-     "\x07\x00"
-     "\xB5\x31\x07\x00\xB6\x29\x07\x00\xB7\x21\x07\x00\xB3\x2D\x06\x00\x3B\x77"
-     "\x00\x80"
-     "\x3F\x77\x00\x80\x43\x77\x00\x80\x47\x77\x00\x80\xB6\x2D\x06\x00\xB5\x2D"
-     "\x06\x00"
-     "\x4B\x77\x00\x80\xBB\xB9\x06\x00\xBA\xB1\x06\x00\x4F\x77\x00\x80\x53\x77"
-     "\x00\x80"
-     "\xBF\xF5\x06\x00\xBE\x8D\x06\x00\xBD\x89\x06\x00\xBC\x95\x06\x00\x57\x77"
-     "\x00\x80"
-     "\xA3\x69\x06\x00\x5B\x77\x00\x80\x5F\x77\x00\x80\xA6\x69\x06\x00\x63\x77"
-     "\x00\x80"
-     "\x67\x77\x00\x80\xA5\x69\x06\x00\xAA\xF5\x06\x00\xAB\xFD\x06\x00\x6B\x77"
-     "\x00\x80"
-     "\x84\xA0\x01\x00\xAE\xC9\x06\x00\xAF\xB1\x06\x00\xAC\xD1\x06\x00\xAD\xCD"
-     "\x06\x00"
-     "\xA8\xE9\x01\x00\xA9\xE9\x01\x00\xAA\xF9\x01\x00\xAB\xF9\x01\x00\xAC\xE9"
-     "\x01\x00"
-     "\xAD\xE9\x01\x00\xAE\x39\x01\x00\xAF\x39\x01\x00\x80\xF5\x00\x00\x81\xFD"
-     "\x00\x00"
-     "\x82\xC1\x00\x00\xBE\x64\x01\x00\x86\x9C\x01\x00\x87\xA0\x00\x00\x73\x77"
-     "\x00\x80"
-     "\x77\x77\x00\x80\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5\x00\x00\xBB\xE5"
-     "\x00\x00"
-     "\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x91\x00\x00\xBF\x91\x00\x00\xB0\x49"
-     "\x01\x00"
-     "\xB1\x49\x01\x00\xB2\x59\x01\x00\xB3\x59\x01\x00\xB4\x49\x01\x00\xB5\x49"
-     "\x01\x00"
-     "\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xB3\x71\x02\x00\x7B\x77\x00\x80\x7F\x77"
-     "\x00\x80"
-     "\x83\x77\x00\x80\x87\x77\x00\x80\xB6\x69\x02\x00\xB5\x61\x02\x00\x8B\x77"
-     "\x00\x80"
-     "\xBB\x8D\x02\x00\xBA\x8D\x02\x00\x8F\x77\x00\x80\x93\x77\x00\x80\xBF\x8D"
-     "\x02\x00"
-     "\xBE\x8D\x02\x00\xBD\x8D\x02\x00\xBC\x8D\x02\x00\x92\x00\x00\x00\xA3\x35"
-     "\x02\x00"
-     "\x97\x77\x00\x80\x9B\x77\x00\x80\xA6\x2D\x02\x00\x9F\x77\x00\x80\xA3\x77"
-     "\x00\x80"
-     "\xA5\x25\x02\x00\xAA\xC9\x02\x00\xAB\xC9\x02\x00\xA7\x77\x00\x80\xAB\x77"
-     "\x00\x80"
-     "\xAE\xC9\x02\x00\xAF\xC9\x02\x00\xAC\xC9\x02\x00\xAD\xC9\x02\x00\xAB\x89"
-     "\x00\x00"
-     "\xAA\x81\x00\x00\xA9\x35\x03\x00\xA8\x35\x03\x00\xAF\xB9\x00\x00\xAE\x89"
-     "\x00\x00"
-     "\xAD\x89\x00\x00\xAC\x91\x00\x00\xAA\x00\x00\x00\xBE\x28\x0C\x00\xAF\x77"
-     "\x00\x80"
-     "\xB3\x77\x00\x80\xB7\x77\x00\x80\xBB\x77\x00\x80\xBF\x77\x00\x80\xC3\x77"
-     "\x00\x80"
-     "\xBB\x79\x00\x00\xBA\x71\x00\x00\xB9\x79\x00\x00\xB8\x61\x00\x00\xBF\xD9"
-     "\x01\x00"
-     "\xBE\xD1\x01\x00\xBD\xD9\x01\x00\xBC\x61\x00\x00\xB3\xB5\x00\x00\xB2\xA9"
-     "\x00\x00"
-     "\xB1\xA1\x00\x00\xB0\xC1\x00\x00\xB7\x59\x00\x00\xB6\x99\x00\x00\xB5\x91"
-     "\x00\x00"
-     "\xB4\xA5\x00\x00\xC7\x77\x00\x80\xCB\x77\x00\x80\xCF\x77\x00\x80\xD3\x77"
-     "\x00\x80"
-     "\x81\x09\x00\x00\x80\x19\x00\x00\xD7\x77\x00\x80\x82\x39\x00\x00\x84\xAC"
-     "\x02\x00"
-     "\xDB\x77\x00\x80\x87\xC0\x02\x00\x86\x0C\x0D\x00\xE3\x77\x00\x80\xE7\x77"
-     "\x00\x80"
-     "\xEB\x77\x00\x80\xEF\x77\x00\x80\xB3\x45\x01\x00\xF3\x77\x00\x80\xF7\x77"
-     "\x00\x80"
-     "\xFB\x77\x00\x80\xFF\x77\x00\x80\xB6\x45\x01\x00\xB5\x55\x01\x00\x03\x78"
-     "\x00\x80"
-     "\xBB\x29\x01\x00\xBA\x49\x01\x00\x07\x78\x00\x80\x0B\x78\x00\x80\xBF\x19"
-     "\x01\x00"
-     "\xBE\x2D\x01\x00\xBD\x2D\x01\x00\xBC\x31\x01\x00\x0F\x78\x00\x80\xE1\x50"
-     "\x06\x00"
-     "\x13\x78\x00\x80\xE3\x3C\x06\x00\x17\x78\x00\x80\x1B\x78\x00\x80\x1F\x78"
-     "\x00\x80"
-     "\x23\x78\x00\x80\x84\x28\x0D\x00\x27\x78\x00\x80\xBE\x24\x0C\x00\x2B\x78"
-     "\x00\x80"
-     "\x2F\x78\x00\x80\x33\x78\x00\x80\x37\x78\x00\x80\xEF\x18\x01\x00\x3B\x78"
-     "\x00\x80"
-     "\x3F\x78\x00\x80\x86\xE8\x0C\x00\x87\x8C\x0C\x00\x43\x78\x00\x80\xE3\xA8"
-     "\x00\x00"
-     "\x47\x78\x00\x80\xE1\xA4\x01\x00\x82\x11\x00\x00\xEF\xB0\x00\x00\x80\x39"
-     "\x00\x00"
-     "\x81\x35\x00\x00\x4B\x78\x00\x80\x53\x78\x00\x80\x57\x78\x00\x80\x5B\x78"
-     "\x00\x80"
-     "\xA3\x05\x0E\x00\xDF\x77\x00\x80\x4F\x78\x00\x80\x5F\x78\x00\x80\x63\x78"
-     "\x00\x80"
-     "\xA6\x05\x0E\x00\xA5\x15\x0E\x00\x67\x78\x00\x80\xAB\x69\x0E\x00\xAA\x09"
-     "\x0E\x00"
-     "\x6B\x78\x00\x80\x6F\x78\x00\x80\xAF\x59\x0E\x00\xAE\x6D\x0E\x00\xAD\x6D"
-     "\x0E\x00"
-     "\xAC\x71\x0E\x00\x73\x78\x00\x80\xB3\x7D\x0F\x00\x77\x78\x00\x80\x7B\x78"
-     "\x00\x80"
-     "\xB6\x19\x0F\x00\x7F\x78\x00\x80\x83\x78\x00\x80\xB5\x19\x0F\x00\xBA\x3D"
-     "\x0F\x00"
-     "\xBB\x0D\x0F\x00\x87\x78\x00\x80\x8B\x78\x00\x80\xBE\xF9\x0F\x00\xBF\xE5"
-     "\x0F\x00"
-     "\xBC\x15\x0F\x00\xBD\xFD\x0F\x00\xA8\xB1\x0F\x00\xA9\xB1\x0F\x00\xAA\xB1"
-     "\x0F\x00"
-     "\xAB\xB1\x0F\x00\xAC\xD5\x0F\x00\xAD\xDD\x0F\x00\xAE\xD5\x0F\x00\xAF\xCD"
-     "\x0F\x00"
-     "\x8F\x78\x00\x80\x93\x78\x00\x80\x97\x78\x00\x80\x9B\x78\x00\x80\x9F\x78"
-     "\x00\x80"
-     "\xA3\x78\x00\x80\xA7\x78\x00\x80\xAB\x78\x00\x80\xB8\x5D\x0F\x00\xB9\x61"
-     "\x0F\x00"
-     "\xBA\x65\x0F\x00\xBB\x7D\x0F\x00\xBC\x65\x0F\x00\xBD\x6D\x0F\x00\xBE\x19"
-     "\x0F\x00"
-     "\xBF\x19\x0F\x00\xB0\xBD\x0F\x00\xB1\x8D\x0F\x00\xB2\x85\x0F\x00\xB3\x6D"
-     "\x0F\x00"
-     "\xB4\x79\x0F\x00\xB5\x65\x0F\x00\xB6\x6D\x0F\x00\xB7\x65\x0F\x00\xA3\x39"
-     "\x0E\x00"
-     "\xAF\x78\x00\x80\xB3\x78\x00\x80\xB7\x78\x00\x80\xBB\x78\x00\x80\xA6\x5D"
-     "\x0E\x00"
-     "\xA5\x5D\x0E\x00\x84\x9C\x02\x00\xAB\x49\x0E\x00\xAA\x79\x0E\x00\xBE\xE4"
-     "\x01\x00"
-     "\xBF\x78\x00\x80\xAF\xA1\x0E\x00\xAE\xBD\x0E\x00\xAD\xB9\x0E\x00\xAC\x51"
-     "\x0E\x00"
-     "\x80\x49\x00\x00\x81\x49\x00\x00\x82\x59\x00\x00\xB3\xB1\x0E\x00\xC3\x78"
-     "\x00\x80"
-     "\xB5\x45\x01\x00\xB6\x45\x01\x00\xC7\x78\x00\x80\x86\xC0\x00\x00\x87\xE4"
-     "\x01\x00"
-     "\xBA\x29\x01\x00\xBB\x25\x01\x00\xBC\x3D\x01\x00\xBD\x21\x01\x00\xBE\x21"
-     "\x01\x00"
-     "\xBF\x15\x01\x00\xA8\x75\x0E\x00\xA9\x7D\x0E\x00\xAA\x75\x0E\x00\xAB\x81"
-     "\x01\x00"
-     "\xAC\x81\x01\x00\xAD\x89\x01\x00\xAE\xB1\x01\x00\xAF\xB1\x01\x00\x84\xA0"
-     "\x01\x00"
-     "\x6F\x77\x00\x80\xCB\x78\x00\x80\xCF\x78\x00\x80\xD3\x78\x00\x80\xD7\x78"
-     "\x00\x80"
-     "\xDB\x78\x00\x80\xDF\x78\x00\x80\xB8\x65\x01\x00\xB9\x6D\x01\x00\xBA\x65"
-     "\x01\x00"
-     "\xBB\x7D\x01\x00\xBC\x65\x01\x00\xBD\x6D\x01\x00\xBE\x65\x01\x00\xBF\xD9"
-     "\x00\x00"
-     "\xB0\xA1\x01\x00\xB1\xAD\x01\x00\xB2\xA5\x01\x00\xB3\xB9\x01\x00\xB4\xA9"
-     "\x01\x00"
-     "\xB5\x9D\x01\x00\xB6\x95\x01\x00\xB7\x5D\x01\x00\xE3\x78\x00\x80\xE7\x78"
-     "\x00\x80"
-     "\xEB\x78\x00\x80\xEF\x78\x00\x80\xF3\x78\x00\x80\xF7\x78\x00\x80\xEF\x44"
-     "\x0E\x00"
-     "\xFB\x78\x00\x80\xE1\x88\x0D\x00\xFF\x78\x00\x80\xE3\x18\x0F\x00\x03\x79"
-     "\x00\x80"
-     "\x07\x79\x00\x80\x0B\x79\x00\x80\x0F\x79\x00\x80\x13\x79\x00\x80\xA3\xBD"
-     "\x0D\x00"
-     "\x17\x79\x00\x80\x1B\x79\x00\x80\x1F\x79\x00\x80\x23\x79\x00\x80\xA6\x49"
-     "\x02\x00"
-     "\xA5\x49\x02\x00\x27\x79\x00\x80\xAB\x29\x02\x00\xAA\x25\x02\x00\x2B\x79"
-     "\x00\x80"
-     "\xBE\x20\x03\x00\xAF\x19\x02\x00\xAE\x2D\x02\x00\xAD\x2D\x02\x00\xAC\x31"
-     "\x02\x00"
-     "\x80\xD1\x00\x00\x81\xD9\x00\x00\x82\xE1\x00\x00\x84\xC0\x04\x00\xEF\x04"
-     "\x02\x00"
-     "\x33\x79\x00\x80\x87\x54\x03\x00\x86\x3C\x04\x00\xE3\x10\x01\x00\x37\x79"
-     "\x00\x80"
-     "\xE1\xD0\x01\x00\x3B\x79\x00\x80\x3F\x79\x00\x80\x43\x79\x00\x80\x47\x79"
-     "\x00\x80"
-     "\x4B\x79\x00\x80\x4F\x79\x00\x80\x53\x79\x00\x80\x57\x79\x00\x80\xB3\x65"
-     "\x03\x00"
-     "\x5B\x79\x00\x80\xB5\x65\x03\x00\xB6\x6D\x03\x00\x5F\x79\x00\x80\x63\x79"
-     "\x00\x80"
-     "\x67\x79\x00\x80\xBA\xE9\x00\x00\xBB\xE9\x00\x00\xBC\xF9\x00\x00\xBD\xF9"
-     "\x00\x00"
-     "\xBE\xE9\x00\x00\xBF\xE9\x00\x00\xA8\x51\x02\x00\xA9\x59\x02\x00\xAA\xA1"
-     "\x02\x00"
-     "\xAB\xA1\x02\x00\xAC\xA1\x02\x00\xAD\xA1\x02\x00\xAE\xA1\x02\x00\xAF\xA1"
-     "\x02\x00"
-     "\xBE\x68\x05\x00\x6B\x79\x00\x80\x6F\x79\x00\x80\x73\x79\x00\x80\x77\x79"
-     "\x00\x80"
-     "\x7B\x79\x00\x80\x7F\x79\x00\x80\x83\x79\x00\x80\xB8\x6D\x01\x00\xB9\x75"
-     "\x01\x00"
-     "\xBA\x7D\x01\x00\xBB\x75\x01\x00\xBC\x6D\x01\x00\xBD\x11\x01\x00\xBE\x11"
-     "\x01\x00"
-     "\xBF\x11\x01\x00\xB0\x61\x01\x00\xB1\x61\x01\x00\xB2\x61\x01\x00\xB3\x61"
-     "\x01\x00"
-     "\xB4\x61\x01\x00\xB5\x61\x01\x00\xB6\x65\x01\x00\xB7\x59\x01\x00\xE1\xB4"
-     "\x0E\x00"
-     "\xE3\x88\x00\x00\xE3\x38\x0E\x00\xE1\x5C\x0E\x00\x87\x79\x00\x80\x8B\x79"
-     "\x00\x80"
-     "\x8F\x79\x00\x80\x93\x79\x00\x80\x97\x79\x00\x80\x9B\x79\x00\x80\xA3\x79"
-     "\x00\x80"
-     "\xA7\x79\x00\x80\xBE\xFC\x04\x00\xEF\x64\x00\x00\xEF\xE8\x0E\x00\xAB\x79"
-     "\x00\x80"
-     "\xA3\x69\x02\x00\x82\x61\x00\x00\x81\x69\x00\x00\x80\x51\x00\x00\x84\x84"
-     "\x04\x00"
-     "\xA6\x61\x02\x00\xA5\x69\x02\x00\xAF\x79\x00\x80\xAB\xE5\x01\x00\xAA\xE5"
-     "\x01\x00"
-     "\x86\x48\x04\x00\x87\xCC\x04\x00\xAF\xE5\x01\x00\xAE\xE5\x01\x00\xAD\xF5"
-     "\x01\x00"
-     "\xAC\xF5\x01\x00\xA8\x95\x06\x00\xA9\xA1\x06\x00\xAA\xA5\x06\x00\xAB\xBD"
-     "\x06\x00"
-     "\xAC\xA5\x06\x00\xAD\xAD\x06\x00\xAE\xA5\x06\x00\xAF\xDD\x06\x00\x9F\x79"
-     "\x00\x80"
-     "\xB3\x79\x00\x80\xB7\x79\x00\x80\xBB\x79\x00\x80\xBF\x79\x00\x80\xC3\x79"
-     "\x00\x80"
-     "\xC7\x79\x00\x80\xCB\x79\x00\x80\xB8\x6D\x07\x00\xB9\x75\x07\x00\xBA\x7D"
-     "\x07\x00"
-     "\xBB\x75\x07\x00\xBC\x6D\x07\x00\xBD\xC5\x07\x00\xBE\xCD\x07\x00\xBF\xC5"
-     "\x07\x00"
-     "\xB0\xA5\x06\x00\xB1\xAD\x06\x00\xB2\xB9\x06\x00\xB3\xB1\x06\x00\xB4\x91"
-     "\x06\x00"
-     "\xB5\x91\x06\x00\xB6\x5D\x07\x00\xB7\x55\x07\x00\xB3\x2D\x06\x00\xCF\x79"
-     "\x00\x80"
-     "\xD3\x79\x00\x80\xD7\x79\x00\x80\xDB\x79\x00\x80\xB6\x2D\x06\x00\xB5\x2D"
-     "\x06\x00"
-     "\xDF\x79\x00\x80\xBB\x69\x06\x00\xBA\x61\x06\x00\xE3\x79\x00\x80\xE7\x79"
-     "\x00\x80"
-     "\xBF\x55\x06\x00\xBE\x55\x06\x00\xBD\x65\x06\x00\xBC\x6D\x06\x00\xEB\x79"
-     "\x00\x80"
-     "\xA3\x69\x06\x00\xEF\x79\x00\x80\xF3\x79\x00\x80\xA6\x69\x06\x00\xF7\x79"
-     "\x00\x80"
-     "\xFB\x79\x00\x80\xA5\x69\x06\x00\xAA\x25\x06\x00\xAB\x2D\x06\x00\xFF\x79"
-     "\x00\x80"
-     "\x03\x7A\x00\x80\xAE\x11\x06\x00\xAF\x11\x06\x00\xAC\x29\x06\x00\xAD\x21"
-     "\x06\x00"
-     "\xA8\xA9\x06\x00\xA9\xA9\x06\x00\xAA\xB5\x06\x00\xAB\xD1\x06\x00\xAC\x35"
-     "\x01\x00"
-     "\xAD\x3D\x01\x00\xAE\x35\x01\x00\xAF\x2D\x01\x00\x80\xCD\x01\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x07\x7A\x00\x80\x0B\x7A\x00\x80\x0F\x7A\x00\x80\x87\x78"
-     "\x01\x00"
-     "\x86\x9C\x00\x00\xB8\xE9\x01\x00\xB9\xE9\x01\x00\xBA\x89\x01\x00\xBB\x85"
-     "\x01\x00"
-     "\xBC\x9D\x01\x00\xBD\x81\x01\x00\xBE\x81\x01\x00\xBF\xB5\x01\x00\xB0\x55"
-     "\x01\x00"
-     "\xB1\x5D\x01\x00\xB2\x55\x01\x00\xB3\xE9\x01\x00\xB4\xF9\x01\x00\xB5\xF9"
-     "\x01\x00"
-     "\xB6\xE9\x01\x00\xB7\xE1\x01\x00\xB5\xA5\x01\x00\x13\x7A\x00\x80\x17\x7A"
-     "\x00\x80"
-     "\xB6\xA5\x01\x00\x1B\x7A\x00\x80\x1F\x7A\x00\x80\xB3\x25\x06\x00\x23\x7A"
-     "\x00\x80"
-     "\xBD\x25\x01\x00\xBC\x99\x01\x00\xBF\x25\x01\x00\xBE\x25\x01\x00\x27\x7A"
-     "\x00\x80"
-     "\x2F\x79\x00\x80\xBB\x81\x01\x00\xBA\x89\x01\x00\xA3\x7D\x06\x00\x2B\x7A"
-     "\x00\x80"
-     "\x2F\x7A\x00\x80\x33\x7A\x00\x80\x37\x7A\x00\x80\xA6\xFD\x01\x00\xA5\xFD"
-     "\x01\x00"
-     "\x3B\x7A\x00\x80\xAB\xD9\x01\x00\xAA\xD1\x01\x00\x3F\x7A\x00\x80\x43\x7A"
-     "\x00\x80"
-     "\xAF\x7D\x01\x00\xAE\x7D\x01\x00\xAD\x7D\x01\x00\xAC\xC1\x01\x00\x47\x7A"
-     "\x00\x80"
-     "\x4B\x7A\x00\x80\xEF\xD0\x07\x00\x4F\x7A\x00\x80\x53\x7A\x00\x80\x57\x7A"
-     "\x00\x80"
-     "\x5B\x7A\x00\x80\x5F\x7A\x00\x80\x63\x7A\x00\x80\x67\x7A\x00\x80\x6B\x7A"
-     "\x00\x80"
-     "\x6F\x7A\x00\x80\xE1\xB4\x06\x00\x73\x7A\x00\x80\xE3\xA4\x06\x00\x77\x7A"
-     "\x00\x80"
-     "\x80\x75\x00\x00\x81\x79\x00\x00\x82\x19\x00\x00\x84\x40\x02\x00\x7B\x7A"
-     "\x00\x80"
-     "\x83\x7A\x00\x80\xBE\x14\x03\x00\x87\x7A\x00\x80\x86\xC0\x3C\x00\x87\x44"
-     "\x03\x00"
-     "\x8B\x7A\x00\x80\x8F\x7A\x00\x80\x93\x7A\x00\x80\x97\x7A\x00\x80\x9B\x7A"
-     "\x00\x80"
-     "\x9F\x7A\x00\x80\xA3\x7A\x00\x80\xA7\x7A\x00\x80\xEF\x28\x02\x00\xAB\x7A"
-     "\x00\x80"
-     "\xAF\x7A\x00\x80\xB3\x7A\x00\x80\xB7\x7A\x00\x80\xBB\x7A\x00\x80\x84\xEC"
-     "\x3C\x00"
-     "\xBF\x7A\x00\x80\xC3\x7A\x00\x80\xC7\x7A\x00\x80\xE1\x94\x01\x00\xCB\x7A"
-     "\x00\x80"
-     "\xE3\xE4\x02\x00\xCF\x7A\x00\x80\xB3\x51\x03\x00\xD3\x7A\x00\x80\xD7\x7A"
-     "\x00\x80"
-     "\xDB\x7A\x00\x80\xDF\x7A\x00\x80\xB6\x65\x03\x00\xB5\x45\x03\x00\xE3\x7A"
-     "\x00\x80"
-     "\xBB\x31\x03\x00\xBA\x29\x03\x00\xE7\x7A\x00\x80\xEB\x7A\x00\x80\xBF\xF1"
-     "\x00\x00"
-     "\xBE\x19\x03\x00\xBD\x11\x03\x00\xBC\x21\x03\x00\xA8\x85\x02\x00\xA9\x8D"
-     "\x02\x00"
-     "\xAA\x99\x02\x00\xAB\x99\x02\x00\xAC\x8D\x02\x00\xAD\xA5\x02\x00\xAE\xDD"
-     "\x02\x00"
-     "\xAF\xD5\x02\x00\x80\xFD\x01\x00\x81\xCD\x01\x00\x82\xC5\x01\x00\xBE\x64"
-     "\x3D\x00"
-     "\x86\x90\x3C\x00\x87\x04\x3D\x00\xF3\x7A\x00\x80\xF7\x7A\x00\x80\xB8\x51"
-     "\x01\x00"
-     "\xB9\x51\x01\x00\xBA\x51\x01\x00\xBB\x51\x01\x00\xBC\x71\x01\x00\xBD\x71"
-     "\x01\x00"
-     "\xBE\x71\x01\x00\xBF\x71\x01\x00\xB0\xAD\x02\x00\xB1\xB5\x02\x00\xB2\xB5"
-     "\x02\x00"
-     "\xB3\x85\x02\x00\xB4\x9D\x02\x00\xB5\x75\x01\x00\xB6\x7D\x01\x00\xB7\x71"
-     "\x01\x00"
-     "\xFB\x7A\x00\x80\xFF\x7A\x00\x80\x03\x7B\x00\x80\x07\x7B\x00\x80\x0B\x7B"
-     "\x00\x80"
-     "\x0F\x7B\x00\x80\x13\x7B\x00\x80\xEF\xD4\x06\x00\xBE\xEC\x3C\x00\xE1\x54"
-     "\x06\x00"
-     "\x17\x7B\x00\x80\xE3\x6C\x06\x00\x1B\x7B\x00\x80\x1F\x7B\x00\x80\x23\x7B"
-     "\x00\x80"
-     "\x27\x7B\x00\x80\xA3\x51\x02\x00\x2B\x7B\x00\x80\x2F\x7B\x00\x80\x33\x7B"
-     "\x00\x80"
-     "\x37\x7B\x00\x80\xA6\x65\x02\x00\xA5\x45\x02\x00\x3B\x7B\x00\x80\xAB\x31"
-     "\x02\x00"
-     "\xAA\x29\x02\x00\x3F\x7B\x00\x80\x43\x7B\x00\x80\xAF\xF1\x01\x00\xAE\x19"
-     "\x02\x00"
-     "\xAD\x11\x02\x00\xAC\x21\x02\x00\xA8\x55\x3E\x00\xA9\x59\x3E\x00\xAA\x61"
-     "\x3E\x00"
-     "\xAB\x61\x3E\x00\xAC\x61\x3E\x00\xAD\x61\x3E\x00\xAE\x61\x3E\x00\xAF\x61"
-     "\x3E\x00"
-     "\xEF\x7A\x00\x80\x47\x7B\x00\x80\x4B\x7B\x00\x80\x4F\x7B\x00\x80\x80\x0D"
-     "\x00\x00"
-     "\x81\xB1\x00\x00\x82\xB1\x00\x00\x53\x7B\x00\x80\xB8\xE9\x3E\x00\xB9\xE9"
-     "\x3E\x00"
-     "\xBA\xFD\x3E\x00\xBB\xF1\x3E\x00\xBC\x95\x3E\x00\xBD\x9D\x3E\x00\xBE\x95"
-     "\x3E\x00"
-     "\xBF\x8D\x3E\x00\xB0\xE1\x3E\x00\xB1\xE1\x3E\x00\xB2\xE5\x3E\x00\xB3\xFD"
-     "\x3E\x00"
-     "\xB4\xE5\x3E\x00\xB5\xED\x3E\x00\xB6\xE5\x3E\x00\xB7\xD9\x3E\x00\xB3\xE9"
-     "\x3E\x00"
-     "\x57\x7B\x00\x80\x5B\x7B\x00\x80\x84\xD0\x03\x00\x5F\x7B\x00\x80\xB6\xE1"
-     "\x3E\x00"
-     "\xB5\xE9\x3E\x00\x63\x7B\x00\x80\xBB\x25\x3E\x00\xBA\x25\x3E\x00\x86\x98"
-     "\x00\x00"
-     "\x87\xA4\x00\x00\xBF\x29\x3E\x00\xBE\x21\x3E\x00\xBD\x29\x3E\x00\xBC\x31"
-     "\x3E\x00"
-     "\x67\x7B\x00\x80\xA3\xAD\x3E\x00\x6B\x7B\x00\x80\x6F\x7B\x00\x80\xA6\xA5"
-     "\x3E\x00"
-     "\x73\x7B\x00\x80\x77\x7B\x00\x80\xA5\xAD\x3E\x00\xAA\x61\x3E\x00\xAB\x61"
-     "\x3E\x00"
-     "\x7B\x7B\x00\x80\x7F\x7B\x00\x80\xAE\x65\x3E\x00\xAF\x6D\x3E\x00\xAC\x75"
-     "\x3E\x00"
-     "\xAD\x6D\x3E\x00\xA8\x39\x3E\x00\xA9\x39\x3E\x00\xAA\x89\x3E\x00\xAB\x89"
-     "\x3E\x00"
-     "\xAC\x99\x3E\x00\xAD\x99\x3E\x00\xAE\x89\x3E\x00\xAF\x89\x3E\x00\x83\x7B"
-     "\x00\x80"
-     "\x87\x7B\x00\x80\x8B\x7B\x00\x80\x8F\x7B\x00\x80\x93\x7B\x00\x80\x97\x7B"
-     "\x00\x80"
-     "\x9B\x7B\x00\x80\x9F\x7B\x00\x80\xB8\x8D\x3E\x00\xB9\x95\x3E\x00\xBA\x95"
-     "\x3E\x00"
-     "\xBB\xA5\x3E\x00\xBC\xBD\x3E\x00\xBD\x71\x01\x00\xBE\x71\x01\x00\xBF\x71"
-     "\x01\x00"
-     "\xB0\xF9\x3E\x00\xB1\xCD\x3E\x00\xB2\xD9\x3E\x00\xB3\xD9\x3E\x00\xB4\xC9"
-     "\x3E\x00"
-     "\xB5\xC9\x3E\x00\xB6\xBD\x3E\x00\xB7\xB5\x3E\x00\xA3\x7B\x00\x80\xB6\x09"
-     "\x3E\x00"
-     "\xB5\x01\x3E\x00\xA7\x7B\x00\x80\xB3\x11\x3E\x00\xAB\x7B\x00\x80\xAF\x7B"
-     "\x00\x80"
-     "\xB3\x7B\x00\x80\xBF\x1D\x3E\x00\xBE\x1D\x3E\x00\xBD\x1D\x3E\x00\xBC\x1D"
-     "\x3E\x00"
-     "\xBB\x19\x3E\x00\xBA\x2D\x3E\x00\x7F\x7A\x00\x80\xB7\x7B\x00\x80\xA3\x41"
-     "\x3E\x00"
-     "\xBB\x7B\x00\x80\xBF\x7B\x00\x80\xC3\x7B\x00\x80\xC7\x7B\x00\x80\xA6\x59"
-     "\x3E\x00"
-     "\xA5\x51\x3E\x00\xCB\x7B\x00\x80\xAB\x49\x3E\x00\xAA\x7D\x3E\x00\x84\x64"
-     "\x01\x00"
-     "\xBE\x2C\x00\x00\xAF\x4D\x3E\x00\xAE\x4D\x3E\x00\xAD\x4D\x3E\x00\xAC\x4D"
-     "\x3E\x00"
-     "\x82\xDD\x03\x00\xEF\xD0\x3F\x00\x80\xC5\x03\x00\x81\xDD\x03\x00\x87\x50"
-     "\x03\x00"
-     "\x86\xB4\x04\x00\xD3\x7B\x00\x80\xD7\x7B\x00\x80\xDB\x7B\x00\x80\xDF\x7B"
-     "\x00\x80"
-     "\xE3\x7B\x00\x80\xE7\x7B\x00\x80\x8C\x00\x00\x00\xE3\x84\x3E\x00\xEB\x7B"
-     "\x00\x80"
-     "\xE1\x44\x3E\x00\xB3\x89\x02\x00\xEF\x7B\x00\x80\x84\x44\x03\x00\xBE\x40"
-     "\x04\x00"
-     "\xF3\x7B\x00\x80\xB6\x8D\x02\x00\xB5\x8D\x02\x00\xF7\x7B\x00\x80\xBB\x45"
-     "\x03\x00"
-     "\xBA\x45\x03\x00\xFB\x7B\x00\x80\xFF\x7B\x00\x80\xBF\x45\x03\x00\xBE\x45"
-     "\x03\x00"
-     "\xBD\x55\x03\x00\xBC\x55\x03\x00\xAC\x55\x02\x00\xAD\x5D\x02\x00\xAE\x61"
-     "\x02\x00"
-     "\xAF\x61\x02\x00\xA8\x0D\x02\x00\xA9\x55\x02\x00\xAA\x51\x02\x00\xAB\x51"
-     "\x02\x00"
-     "\x84\xA8\x04\x00\x03\x7C\x00\x80\x07\x7C\x00\x80\x0B\x7C\x00\x80\xBE\xBC"
-     "\x05\x00"
-     "\x0F\x7C\x00\x80\x13\x7C\x00\x80\x17\x7C\x00\x80\xBC\x5D\x03\x00\xBD\x79"
-     "\x03\x00"
-     "\xBE\x0D\x03\x00\xBF\x05\x03\x00\xB8\x51\x03\x00\xB9\x59\x03\x00\xBA\x4D"
-     "\x03\x00"
-     "\xBB\x45\x03\x00\xB4\xD1\x03\x00\xB5\xD1\x03\x00\xB6\xD1\x03\x00\xB7\xD1"
-     "\x03\x00"
-     "\xB0\xD1\x03\x00\xB1\xD1\x03\x00\xB2\xD1\x03\x00\xB3\xD1\x03\x00\x1B\x7C"
-     "\x00\x80"
-     "\x1F\x7C\x00\x80\x23\x7C\x00\x80\x27\x7C\x00\x80\x2B\x7C\x00\x80\xE1\xA4"
-     "\x01\x00"
-     "\x2F\x7C\x00\x80\xE3\xE4\x00\x00\x33\x7C\x00\x80\x37\x7C\x00\x80\x3B\x7C"
-     "\x00\x80"
-     "\xEF\x38\x00\x00\x43\x7C\x00\x80\x47\x7C\x00\x80\x4B\x7C\x00\x80\x4F\x7C"
-     "\x00\x80"
-     "\x80\x61\x00\x00\x81\x61\x00\x00\x82\x61\x00\x00\xA3\x85\x03\x00\x53\x7C"
-     "\x00\x80"
-     "\xA5\x81\x03\x00\xA6\x81\x03\x00\x84\x98\x07\x00\x86\xC0\x04\x00\x87\x48"
-     "\x05\x00"
-     "\xAA\x49\x02\x00\xAB\x49\x02\x00\xAC\x59\x02\x00\xAD\x59\x02\x00\xAE\x49"
-     "\x02\x00"
-     "\xAF\x49\x02\x00\x57\x7C\x00\x80\x5B\x7C\x00\x80\x5F\x7C\x00\x80\x63\x7C"
-     "\x00\x80"
-     "\x67\x7C\x00\x80\x6B\x7C\x00\x80\x6F\x7C\x00\x80\xEF\xC4\x3E\x00\xBE\x20"
-     "\x06\x00"
-     "\xE1\x8C\x3E\x00\x73\x7C\x00\x80\xE3\x30\x01\x00\x77\x7C\x00\x80\x7B\x7C"
-     "\x00\x80"
-     "\x7F\x7C\x00\x80\x83\x7C\x00\x80\xB3\xE1\x01\x00\x87\x7C\x00\x80\x8B\x7C"
-     "\x00\x80"
-     "\x8F\x7C\x00\x80\x93\x7C\x00\x80\xB6\xF9\x01\x00\xB5\xF1\x01\x00\x97\x7C"
-     "\x00\x80"
-     "\xBB\xBD\x01\x00\xBA\xB5\x01\x00\x9B\x7C\x00\x80\x9F\x7C\x00\x80\xBF\x59"
-     "\x01\x00"
-     "\xBE\x51\x01\x00\xBD\x95\x01\x00\xBC\x95\x01\x00\xA8\x19\x06\x00\xA9\x19"
-     "\x06\x00"
-     "\xAA\x29\x06\x00\xAB\x29\x06\x00\xAC\x39\x06\x00\xAD\x39\x06\x00\xAE\x21"
-     "\x06\x00"
-     "\xAF\x59\x06\x00\x3F\x7C\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xA3\x7C\x00\x80\xA7\x7C\x00\x80\xAB\x7C\x00\x80\xAF\x7C\x00\x80\xB8\xF9"
-     "\x06\x00"
-     "\xB9\xF9\x06\x00\xBA\x49\x07\x00\xBB\x49\x07\x00\xBC\x59\x07\x00\xBD\x59"
-     "\x07\x00"
-     "\xBE\x49\x07\x00\xBF\x45\x07\x00\xB0\x31\x06\x00\xB1\x39\x06\x00\xB2\x01"
-     "\x06\x00"
-     "\xB3\x01\x06\x00\xB4\xD5\x06\x00\xB5\xDD\x06\x00\xB6\xD5\x06\x00\xB7\xC9"
-     "\x06\x00"
-     "\xA3\xA1\x06\x00\xB3\x7C\x00\x80\x86\xAC\x00\x00\x87\x4C\x03\x00\xB7\x7C"
-     "\x00\x80"
-     "\xA6\xB9\x06\x00\xA5\xB1\x06\x00\xBB\x7C\x00\x80\xAB\xFD\x06\x00\xAA\xF5"
-     "\x06\x00"
-     "\xBF\x7C\x00\x80\xC3\x7C\x00\x80\xAF\x19\x06\x00\xAE\x11\x06\x00\xAD\xD5"
-     "\x06\x00"
-     "\xAC\xD5\x06\x00\xC7\x7C\x00\x80\xB3\xFD\x07\x00\xCB\x7C\x00\x80\xCF\x7C"
-     "\x00\x80"
-     "\xB6\x51\x07\x00\xD3\x7C\x00\x80\xD7\x7C\x00\x80\xB5\x8D\x07\x00\xBA\x6D"
-     "\x07\x00"
-     "\xBB\x75\x07\x00\xDB\x7C\x00\x80\xDF\x7C\x00\x80\xBE\x59\x07\x00\xBF\x45"
-     "\x07\x00"
-     "\xBC\x6D\x07\x00\xBD\x65\x07\x00\xA8\x31\x06\x00\xA9\x31\x06\x00\xAA\x31"
-     "\x06\x00"
-     "\xAB\x31\x06\x00\xAC\x41\x06\x00\xAD\x95\x06\x00\xAE\x9D\x06\x00\xAF\x95"
-     "\x06\x00"
-     "\xE3\x7C\x00\x80\xE7\x7C\x00\x80\xEB\x7C\x00\x80\xEF\x7C\x00\x80\xF3\x7C"
-     "\x00\x80"
-     "\xF7\x7C\x00\x80\xFB\x7C\x00\x80\xFF\x7C\x00\x80\xB8\xFD\x06\x00\xB9\x81"
-     "\x06\x00"
-     "\xBA\x81\x06\x00\xBB\x81\x06\x00\xBC\x81\x06\x00\xBD\x89\x06\x00\xBE\xB1"
-     "\x06\x00"
-     "\xBF\xB1\x06\x00\xB0\xED\x06\x00\xB1\xF5\x06\x00\xB2\xFD\x06\x00\xB3\xF5"
-     "\x06\x00"
-     "\xB4\xD9\x06\x00\xB5\xC5\x06\x00\xB6\xCD\x06\x00\xB7\xC5\x06\x00\xA3\xB9"
-     "\x06\x00"
-     "\x03\x7D\x00\x80\x07\x7D\x00\x80\x0B\x7D\x00\x80\x0F\x7D\x00\x80\xA6\x15"
-     "\x06\x00"
-     "\xA5\xC9\x06\x00\x13\x7D\x00\x80\xAB\x31\x06\x00\xAA\x29\x06\x00\x17\x7D"
-     "\x00\x80"
-     "\x1B\x7D\x00\x80\xAF\x01\x06\x00\xAE\x1D\x06\x00\xAD\x21\x06\x00\xAC\x29"
-     "\x06\x00"
-     "\x80\x5D\x00\x00\x81\x6D\x00\x00\x82\x65\x00\x00\xCF\x7B\x00\x80\x1F\x7D"
-     "\x00\x80"
-     "\x23\x7D\x00\x80\x87\x78\x01\x00\x86\x00\x0C\x00\x27\x7D\x00\x80\x2B\x7D"
-     "\x00\x80"
-     "\x2F\x7D\x00\x80\x33\x7D\x00\x80\x37\x7D\x00\x80\x3B\x7D\x00\x80\x3F\x7D"
-     "\x00\x80"
-     "\x43\x7D\x00\x80\x47\x7D\x00\x80\x4B\x7D\x00\x80\xB3\x95\x02\x00\x84\x94"
-     "\x03\x00"
-     "\xB5\xBD\x02\x00\x4F\x7D\x00\x80\x53\x7D\x00\x80\xB6\xDD\x02\x00\x57\x7D"
-     "\x00\x80"
-     "\x5B\x7D\x00\x80\xBB\xF1\x02\x00\xBA\xF9\x02\x00\xBD\xDD\x02\x00\xBC\xDD"
-     "\x02\x00"
-     "\xBF\x1D\x03\x00\xBE\x1D\x03\x00\xE3\x18\x06\x00\x5F\x7D\x00\x80\xE1\x44"
-     "\x06\x00"
-     "\x63\x7D\x00\x80\x67\x7D\x00\x80\xE1\x90\x01\x00\x6B\x7D\x00\x80\xE3\x40"
-     "\x02\x00"
-     "\x6F\x7D\x00\x80\x73\x7D\x00\x80\x77\x7D\x00\x80\xEF\x20\x02\x00\xEF\xC4"
-     "\x06\x00"
-     "\x7B\x7D\x00\x80\x7F\x7D\x00\x80\x83\x7D\x00\x80\xA9\x11\x02\x00\xA8\x01"
-     "\x02\x00"
-     "\xAB\x21\x02\x00\xAA\x11\x02\x00\xAD\x51\x02\x00\xAC\x39\x02\x00\xAF\x55"
-     "\x02\x00"
-     "\xAE\x55\x02\x00\x80\xDD\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x87\x7D"
-     "\x00\x80"
-     "\x8B\x7D\x00\x80\xBE\xF8\x03\x00\x87\xA8\x0C\x00\x86\x9C\x0D\x00\xB9\xF5"
-     "\x03\x00"
-     "\xB8\xF5\x03\x00\xBB\xF5\x03\x00\xBA\xF5\x03\x00\xBD\x59\x03\x00\xBC\x51"
-     "\x03\x00"
-     "\xBF\x49\x03\x00\xBE\x51\x03\x00\xB1\x39\x02\x00\xB0\x31\x02\x00\xB3\x05"
-     "\x02\x00"
-     "\xB2\x05\x02\x00\xB5\xD5\x03\x00\xB4\xD5\x03\x00\xB7\xD5\x03\x00\xB6\xD5"
-     "\x03\x00"
-     "\x93\x7D\x00\x80\x97\x7D\x00\x80\x9B\x7D\x00\x80\xA3\x91\x03\x00\x9F\x7D"
-     "\x00\x80"
-     "\xA5\xB9\x03\x00\xA6\xD9\x03\x00\xA3\x7D\x00\x80\xA7\x7D\x00\x80\xAB\x7D"
-     "\x00\x80"
-     "\xAA\xFD\x03\x00\xAB\xF5\x03\x00\xAC\xD9\x03\x00\xAD\xD9\x03\x00\xAE\x19"
-     "\x02\x00"
-     "\xAF\x19\x02\x00\xB6\xB1\x02\x00\xAF\x7D\x00\x80\xB3\x7D\x00\x80\xB5\xA1"
-     "\x02\x00"
-     "\xB7\x7D\x00\x80\xB3\xB1\x02\x00\xBB\x7D\x00\x80\xBF\x7D\x00\x80\xBE\x79"
-     "\x01\x00"
-     "\xBF\x45\x01\x00\xBC\x6D\x01\x00\xBD\x79\x01\x00\xBA\x8D\x02\x00\xBB\x75"
-     "\x01\x00"
-     "\xC3\x7D\x00\x80\xC7\x7D\x00\x80\xEF\xD8\x06\x00\xE1\xD0\x23\x00\xCB\x7D"
-     "\x00\x80"
-     "\xE3\x38\x12\x00\xCF\x7D\x00\x80\xD3\x7D\x00\x80\xD7\x7D\x00\x80\xEF\xD4"
-     "\x07\x00"
-     "\x84\x6C\x0F\x00\xE1\xDC\x06\x00\xDB\x7D\x00\x80\xE3\x1C\x07\x00\xE3\xFC"
-     "\x06\x00"
-     "\xDF\x7D\x00\x80\xE1\x28\x01\x00\xEF\xA8\xE4\x01\xA3\x25\x02\x00\xF0\x00"
-     "\x00\x00"
-     "\xE3\x7D\x00\x80\xE7\x7D\x00\x80\xEB\x7D\x00\x80\xA6\x25\x02\x00\xA5\x35"
-     "\x02\x00"
-     "\xEF\x7D\x00\x80\xAB\xE1\x01\x00\xAA\x19\x02\x00\xF3\x7D\x00\x80\xF7\x7D"
-     "\x00\x80"
-     "\xAF\xD1\x01\x00\xAE\xED\x01\x00\xAD\xED\x01\x00\xAC\xF9\x01\x00\x8F\x21"
-     "\x24\x00"
-     "\x8F\x7D\x00\x80\xFB\x7D\x00\x80\xBE\xC4\x00\x00\x80\xF9\x0F\x00\x81\xF9"
-     "\x0F\x00"
-     "\x82\x11\x00\x00\xFF\x7D\x00\x80\x86\x00\x00\x00\x87\x84\x03\x00\x03\x7E"
-     "\x00\x80"
-     "\x07\x7E\x00\x80\x0B\x7E\x00\x80\x0F\x7E\x00\x80\x13\x7E\x00\x80\x17\x7E"
-     "\x00\x80"
-     "\x9F\xD1\x08\x00\x9E\x65\x08\x00\x9D\xF1\x09\x00\x9C\x79\x35\x00\x9B\x71"
-     "\x34\x00"
-     "\x9A\xE1\x35\x00\x99\x75\x31\x00\x98\x55\x30\x00\x97\x65\x30\x00\x96\xC1"
-     "\x31\x00"
-     "\x95\x31\x3C\x00\x94\x6D\x3C\x00\x93\xD9\x39\x00\x92\x45\x38\x00\x91\x59"
-     "\x38\x00"
-     "\x90\xF1\x39\x00\xB1\xA5\x14\x00\xB0\xA5\x14\x00\xB3\x09\xE8\x01\xB2\xE9"
-     "\x15\x00"
-     "\xB5\x05\xE8\x01\xB4\x19\xE8\x01\x1B\x7E\x00\x80\x1F\x7E\x00\x80\x23\x7E"
-     "\x00\x80"
-     "\x27\x7E\x00\x80\x2B\x7E\x00\x80\x2F\x7E\x00\x80\xA3\x61\x0D\x00\xA2\x89"
-     "\x0C\x00"
-     "\xA1\x99\x0C\x00\xA0\x71\x0D\x00\xA1\xC5\x00\x00\x33\x7E\x00\x80\xA3\xC9"
-     "\x04\x00"
-     "\xA2\xD1\x00\x00\xA5\x39\x04\x00\xA4\xD1\x04\x00\xA7\xF9\x18\x00\xA6\x01"
-     "\x18\x00"
-     "\xA9\x01\x1C\x00\xA8\x79\x18\x00\xAB\x6D\x1C\x00\xAA\xCD\x1C\x00\xAD\xD1"
-     "\x10\x00"
-     "\xAC\x01\x10\x00\xAF\x01\x14\x00\xAE\xD9\x10\x00\x82\x21\x25\x00\x83\x89"
-     "\x38\x00"
-     "\x37\x7E\x00\x80\x3B\x7E\x00\x80\x86\xB9\x38\x00\x87\x15\x3D\x00\x84\x39"
-     "\x39\x00"
-     "\x85\x2D\x39\x00\x8A\x79\x3D\x00\x8B\x81\x30\x00\x3F\x7E\x00\x80\x43\x7E"
-     "\x00\x80"
-     "\x8E\x1D\x30\x00\x8F\x81\x34\x00\x8C\x3D\x31\x00\x8D\x35\x31\x00\x92\x61"
-     "\x35\x00"
-     "\x93\x71\x34\x00\x47\x7E\x00\x80\x4B\x7E\x00\x80\x96\xC1\x09\x00\x97\x99"
-     "\x0C\x00"
-     "\x94\xE1\x08\x00\x95\x7D\x09\x00\x9A\x41\x0D\x00\x9B\x59\x0D\x00\x4F\x7E"
-     "\x00\x80"
-     "\x53\x7E\x00\x80\x57\x7E\x00\x80\x5B\x7E\x00\x80\x9C\xA9\x0C\x00\x5F\x7E"
-     "\x00\x80"
-     "\xA8\x81\x0E\x00\xA9\x81\x0E\x00\xAA\x81\x0E\x00\xAB\x9D\x0E\x00\xAC\x8D"
-     "\x0E\x00"
-     "\xAD\xB5\x0E\x00\xAE\xBD\x0E\x00\xAF\xD9\x0E\x00\x80\x0D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x1D\x00\x00\x63\x7E\x00\x80\x67\x7E\x00\x80\x6B\x7E\x00\x80\x84\xBC"
-     "\x00\x00"
-     "\xBE\x00\x04\x00\xB8\x55\x0E\x00\xB9\x5D\x0E\x00\xBA\x61\x0E\x00\xBB\x79"
-     "\x0E\x00"
-     "\xBC\x19\x0E\x00\xBD\x19\x0E\x00\xBE\x09\x0E\x00\xBF\x01\x0E\x00\xB0\xA9"
-     "\x0E\x00"
-     "\xB1\xB5\x0E\x00\xB2\xB1\x0E\x00\xB3\xB1\x0E\x00\xB4\x89\x0E\x00\xB5\x89"
-     "\x0E\x00"
-     "\xB6\x71\x0E\x00\xB7\x6D\x0E\x00\xB3\x09\x0D\x00\x73\x7E\x00\x80\x86\x88"
-     "\x04\x00"
-     "\x87\xC8\x03\x00\x84\x5C\x04\x00\xB6\x35\x0D\x00\xB5\x05\x0D\x00\x77\x7E"
-     "\x00\x80"
-     "\xBB\x71\x0D\x00\xBA\x6D\x0D\x00\x7B\x7E\x00\x80\x7F\x7E\x00\x80\xBF\x4D"
-     "\x0D\x00"
-     "\xBE\x4D\x0D\x00\xBD\x65\x0D\x00\xBC\x65\x0D\x00\x83\x7E\x00\x80\xA3\x4D"
-     "\x0D\x00"
-     "\x87\x7E\x00\x80\x8B\x7E\x00\x80\xA6\x71\x0D\x00\x8F\x7E\x00\x80\x93\x7E"
-     "\x00\x80"
-     "\xA5\x41\x0D\x00\xAA\x29\x0D\x00\xAB\x35\x0D\x00\x97\x7E\x00\x80\x9B\x7E"
-     "\x00\x80"
-     "\xAE\x09\x0D\x00\xAF\x09\x0D\x00\xAC\x21\x0D\x00\xAD\x21\x0D\x00\xA8\xA9"
-     "\x02\x00"
-     "\xA9\xA9\x02\x00\xAA\xB9\x02\x00\xAB\xB9\x02\x00\xAC\xA9\x02\x00\xAD\xA9"
-     "\x02\x00"
-     "\xAE\x01\x02\x00\xAF\x01\x02\x00\x9F\x7E\x00\x80\xA3\x7E\x00\x80\xA7\x7E"
-     "\x00\x80"
-     "\xAB\x7E\x00\x80\xAF\x7E\x00\x80\xB3\x7E\x00\x80\xB7\x7E\x00\x80\xBB\x7E"
-     "\x00\x80"
-     "\xB8\xED\x03\x00\xB9\x8D\x03\x00\xBA\x85\x03\x00\xBB\xA5\x03\x00\xBC\xBD"
-     "\x03\x00"
-     "\xBD\xA5\x03\x00\xBE\xAD\x03\x00\xBF\xA1\x03\x00\xB0\x41\x02\x00\xB1\x41"
-     "\x02\x00"
-     "\xB2\x41\x02\x00\xB3\x41\x02\x00\xB4\x35\x02\x00\xB5\xFD\x03\x00\xB6\xF5"
-     "\x03\x00"
-     "\xB7\xD5\x03\x00\xB7\x45\x0D\x00\xB6\x21\x0D\x00\xB5\x29\x0D\x00\xB4\x15"
-     "\x0D\x00"
-     "\xB3\x05\x0D\x00\xB2\x11\x0D\x00\xB1\x19\x0D\x00\xB0\xE1\x0E\x00\xBF\x59"
-     "\x0D\x00"
-     "\xBE\x51\x0D\x00\xBD\x41\x0D\x00\xBC\x49\x0D\x00\xBB\x51\x0D\x00\xBA\x6D"
-     "\x0D\x00"
-     "\xB9\x69\x0D\x00\xB8\x71\x0D\x00\xBF\x7E\x00\x80\xBE\x0C\x05\x00\xC3\x7E"
-     "\x00\x80"
-     "\xC7\x7E\x00\x80\xCB\x7E\x00\x80\xCF\x7E\x00\x80\xD3\x7E\x00\x80\xD7\x7E"
-     "\x00\x80"
-     "\xAF\x99\x0E\x00\xAE\x91\x0E\x00\xAD\x99\x0E\x00\xAC\x81\x0E\x00\xAB\x99"
-     "\x0E\x00"
-     "\xAA\x91\x0E\x00\xA9\x99\x0E\x00\xA8\xA9\x0E\x00\xE1\x24\x0F\x00\xDB\x7E"
-     "\x00\x80"
-     "\xE3\x38\x0E\x00\xE3\x7E\x00\x80\xE3\x38\x00\x00\xE7\x7E\x00\x80\xE1\xA0"
-     "\x01\x00"
-     "\xEB\x7E\x00\x80\xEF\xE4\x00\x00\xEF\x7E\x00\x80\x84\xE4\x04\x00\xF3\x7E"
-     "\x00\x80"
-     "\x86\x10\x04\x00\x87\x94\x04\x00\xEF\x4C\x01\x00\xF7\x7E\x00\x80\x81\x9D"
-     "\x00\x00"
-     "\x80\x9D\x00\x00\xB3\xF9\x03\x00\x82\x49\x00\x00\xB5\xE9\x03\x00\xFB\x7E"
-     "\x00\x80"
-     "\xFF\x7E\x00\x80\xB6\xE1\x03\x00\x03\x7F\x00\x80\x07\x7F\x00\x80\xBB\x4D"
-     "\x03\x00"
-     "\xBA\x4D\x03\x00\xBD\x5D\x03\x00\xBC\x5D\x03\x00\xBF\xF1\x02\x00\xBE\x39"
-     "\x03\x00"
-     "\x6F\x7E\x00\x80\xDF\x7E\x00\x80\x0B\x7F\x00\x80\x0F\x7F\x00\x80\x13\x7F"
-     "\x00\x80"
-     "\x17\x7F\x00\x80\x1B\x7F\x00\x80\x1F\x7F\x00\x80\xA8\x25\x06\x00\xA9\x39"
-     "\x06\x00"
-     "\xAA\x01\x06\x00\xAB\x01\x06\x00\xAC\x71\x06\x00\xAD\x7D\x06\x00\xAE\x75"
-     "\x06\x00"
-     "\xAF\x69\x06\x00\xB0\x19\x06\x00\xB1\x2D\x06\x00\xB2\x25\x06\x00\xB3\xED"
-     "\x06\x00"
-     "\xB4\xF5\x06\x00\xB5\xFD\x06\x00\xB6\xF5\x06\x00\xB7\xED\x06\x00\xB8\xD5"
-     "\x06\x00"
-     "\xB9\xDD\x06\x00\xBA\xD5\x06\x00\xBB\x2D\x07\x00\xBC\x39\x07\x00\xBD\x39"
-     "\x07\x00"
-     "\xBE\x29\x07\x00\xBF\x29\x07\x00\xB3\xA9\x06\x00\x23\x7F\x00\x80\x27\x7F"
-     "\x00\x80"
-     "\x2B\x7F\x00\x80\x2F\x7F\x00\x80\xB6\xDD\x06\x00\xB5\xD9\x06\x00\x33\x7F"
-     "\x00\x80"
-     "\xBB\xC9\x06\x00\xBA\xFD\x06\x00\x37\x7F\x00\x80\x3B\x7F\x00\x80\xBF\x19"
-     "\x06\x00"
-     "\xBE\x11\x06\x00\xBD\x19\x06\x00\xBC\xD1\x06\x00\x3F\x7F\x00\x80\xA3\xED"
-     "\x06\x00"
-     "\x43\x7F\x00\x80\x47\x7F\x00\x80\xA6\x99\x06\x00\x4B\x7F\x00\x80\xBE\x90"
-     "\x03\x00"
-     "\xA5\x9D\x06\x00\xAA\xB9\x06\x00\xAB\x8D\x06\x00\x4F\x7F\x00\x80\x53\x7F"
-     "\x00\x80"
-     "\xAE\x55\x06\x00\xAF\x5D\x06\x00\xAC\x95\x06\x00\xAD\x5D\x06\x00\xA8\x8D"
-     "\x06\x00"
-     "\xA9\x91\x06\x00\xAA\x95\x06\x00\xAB\xA9\x06\x00\xAC\xBD\x06\x00\xAD\xA9"
-     "\x06\x00"
-     "\xAE\xCD\x06\x00\xAF\xC5\x06\x00\x80\xE9\x01\x00\x81\xE9\x01\x00\x82\xFD"
-     "\x01\x00"
-     "\x57\x7F\x00\x80\x86\x90\x01\x00\x87\xB4\x01\x00\x5B\x7F\x00\x80\x5F\x7F"
-     "\x00\x80"
-     "\xB8\xC5\x01\x00\xB9\xCD\x01\x00\xBA\xC5\x01\x00\xBB\xDD\x01\x00\xBC\xCD"
-     "\x01\x00"
-     "\xBD\xFD\x01\x00\xBE\xF5\x01\x00\xBF\x9D\x01\x00\xB0\x81\x06\x00\xB1\x81"
-     "\x06\x00"
-     "\xB2\x61\x01\x00\xB3\x61\x01\x00\xB4\x61\x01\x00\xB5\x61\x01\x00\xB6\x61"
-     "\x01\x00"
-     "\xB7\x61\x01\x00\xB3\x1D\x06\x00\x63\x7F\x00\x80\x67\x7F\x00\x80\x6B\x7F"
-     "\x00\x80"
-     "\x6F\x7F\x00\x80\xB6\x29\x06\x00\xB5\x35\x06\x00\x73\x7F\x00\x80\xBB\x45"
-     "\x06\x00"
-     "\xBA\x45\x06\x00\x77\x7F\x00\x80\x7B\x7F\x00\x80\xBF\xA5\x01\x00\xBE\xA5"
-     "\x01\x00"
-     "\xBD\xB5\x01\x00\xBC\xB5\x01\x00\x7F\x7F\x00\x80\xA3\x59\x06\x00\x83\x7F"
-     "\x00\x80"
-     "\x87\x7F\x00\x80\xA6\x6D\x06\x00\x8B\x7F\x00\x80\x8F\x7F\x00\x80\xA5\x71"
-     "\x06\x00"
-     "\xAA\x01\x06\x00\xAB\x01\x06\x00\x93\x7F\x00\x80\x97\x7F\x00\x80\xAE\xE1"
-     "\x01\x00"
-     "\xAF\xE1\x01\x00\xAC\xF1\x01\x00\xAD\xF1\x01\x00\xA8\xED\x02\x00\xA9\x29"
-     "\x03\x00"
-     "\xAA\x31\x03\x00\xAB\x31\x03\x00\xAC\x95\x03\x00\xAD\x9D\x03\x00\xAE\x95"
-     "\x03\x00"
-     "\xAF\x8D\x03\x00\x84\x0C\x02\x00\x9B\x7F\x00\x80\x9F\x7F\x00\x80\xA3\x7F"
-     "\x00\x80"
-     "\xA7\x7F\x00\x80\xAB\x7F\x00\x80\xAF\x7F\x00\x80\xB3\x7F\x00\x80\xB8\xA9"
-     "\x03\x00"
-     "\xB9\xA9\x03\x00\xBA\x6D\x00\x00\xBB\x65\x00\x00\xBC\x7D\x00\x00\xBD\x65"
-     "\x00\x00"
-     "\xBE\x6D\x00\x00\xBF\x65\x00\x00\xB0\xFD\x03\x00\xB1\xCD\x03\x00\xB2\xC5"
-     "\x03\x00"
-     "\xB3\xAD\x03\x00\xB4\xB9\x03\x00\xB5\xB9\x03\x00\xB6\xAD\x03\x00\xB7\xA5"
-     "\x03\x00"
-     "\x80\xBD\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\xB7\x7F\x00\x80\xBB\x7F"
-     "\x00\x80"
-     "\xBE\xD8\x03\x00\xC3\x7F\x00\x80\xC7\x7F\x00\x80\x86\x00\x1C\x00\x87\xD4"
-     "\x03\x00"
-     "\xCB\x7F\x00\x80\xCF\x7F\x00\x80\xD3\x7F\x00\x80\xD7\x7F\x00\x80\xDB\x7F"
-     "\x00\x80"
-     "\xDF\x7F\x00\x80\xBB\x69\x03\x00\xBA\x59\x03\x00\xE3\x7F\x00\x80\xE7\x7F"
-     "\x00\x80"
-     "\xBF\x1D\x03\x00\xBE\x1D\x03\x00\xBD\x19\x03\x00\xBC\x71\x03\x00\xB3\xD9"
-     "\x03\x00"
-     "\xEB\x7F\x00\x80\xEF\x7F\x00\x80\xF3\x7F\x00\x80\xF7\x7F\x00\x80\xB6\x7D"
-     "\x03\x00"
-     "\xB5\x7D\x03\x00\xFB\x7F\x00\x80\xAF\x09\x06\x00\xAE\xC1\x07\x00\xAD\xA5"
-     "\x07\x00"
-     "\xAC\xA5\x07\x00\xAB\xB5\x07\x00\xAA\xB5\x07\x00\xFF\x7F\x00\x80\x03\x80"
-     "\x00\x80"
-     "\x84\x6C\x1D\x00\xA6\x19\x07\x00\xA5\x11\x07\x00\x07\x80\x00\x80\xA3\x01"
-     "\x07\x00"
-     "\x0B\x80\x00\x80\x0F\x80\x00\x80\x13\x80\x00\x80\x17\x80\x00\x80\x1B\x80"
-     "\x00\x80"
-     "\xBE\x08\x1D\x00\x84\x00\x1C\x00\x1F\x80\x00\x80\xEF\x38\x03\x00\x23\x80"
-     "\x00\x80"
-     "\x27\x80\x00\x80\x2B\x80\x00\x80\xE3\x54\x03\x00\x33\x80\x00\x80\xE1\x44"
-     "\x00\x00"
-     "\x80\x29\x00\x00\x81\x29\x00\x00\x82\x19\x00\x00\x37\x80\x00\x80\xE1\x28"
-     "\x06\x00"
-     "\xE1\x94\x07\x00\xE3\x30\x06\x00\xE3\x8C\x07\x00\x3B\x80\x00\x80\x3F\x80"
-     "\x00\x80"
-     "\x86\x38\x1C\x00\x87\x5C\x1D\x00\xBE\xE8\x1C\x00\x43\x80\x00\x80\x47\x80"
-     "\x00\x80"
-     "\x4B\x80\x00\x80\x4F\x80\x00\x80\x53\x80\x00\x80\xEF\x98\x01\x00\xEF\x00"
-     "\x06\x00"
-     "\xA3\xF9\x02\x00\x57\x80\x00\x80\x5B\x80\x00\x80\x5F\x80\x00\x80\x63\x80"
-     "\x00\x80"
-     "\xA6\x5D\x02\x00\xA5\x5D\x02\x00\x67\x80\x00\x80\xAB\x49\x02\x00\xAA\x79"
-     "\x02\x00"
-     "\x6B\x80\x00\x80\x6F\x80\x00\x80\xAF\x3D\x02\x00\xAE\x3D\x02\x00\xAD\x39"
-     "\x02\x00"
-     "\xAC\x51\x02\x00\xA8\x99\x1E\x00\xA9\x99\x1E\x00\xAA\xA1\x1E\x00\xAB\xA1"
-     "\x1E\x00"
-     "\xAC\xA1\x1E\x00\xAD\xA1\x1E\x00\xAE\xA1\x1E\x00\xAF\xA1\x1E\x00\x2F\x80"
-     "\x00\x80"
-     "\x73\x80\x00\x80\x77\x80\x00\x80\x7B\x80\x00\x80\x7F\x80\x00\x80\x83\x80"
-     "\x00\x80"
-     "\x87\x80\x00\x80\x8B\x80\x00\x80\xB8\x5D\x1F\x00\xB9\x69\x1F\x00\xBA\x7D"
-     "\x1F\x00"
-     "\xBB\x75\x1F\x00\xBC\x6D\x1F\x00\xBD\x15\x1F\x00\xBE\x1D\x1F\x00\xBF\x0D"
-     "\x1F\x00"
-     "\xB0\xA9\x1E\x00\xB1\xA9\x1E\x00\xB2\x8D\x1E\x00\xB3\x85\x1E\x00\xB4\x9D"
-     "\x1E\x00"
-     "\xB5\x85\x1E\x00\xB6\x85\x1E\x00\xB7\x65\x1F\x00\xB3\x29\x1E\x00\x8F\x80"
-     "\x00\x80"
-     "\x93\x80\x00\x80\x97\x80\x00\x80\x9B\x80\x00\x80\xB6\x21\x1E\x00\xB5\x29"
-     "\x1E\x00"
-     "\x9F\x80\x00\x80\xBB\x6D\x1E\x00\xBA\x6D\x1E\x00\x86\x08\x00\x00\x87\x2C"
-     "\x00\x00"
-     "\xBF\x41\x1E\x00\xBE\x59\x1E\x00\xBD\x51\x1E\x00\xBC\x59\x1E\x00\x80\x49"
-     "\x00\x00"
-     "\x81\x49\x00\x00\x82\x59\x00\x00\xA3\x65\x1E\x00\xA3\x80\x00\x80\xA5\x65"
-     "\x1E\x00"
-     "\xA6\x6D\x1E\x00\xBF\x7F\x00\x80\xA7\x80\x00\x80\xAB\x80\x00\x80\xAA\x21"
-     "\x1E\x00"
-     "\xAB\x21\x1E\x00\xAC\x15\x1E\x00\xAD\x1D\x1E\x00\xAE\x15\x1E\x00\xAF\x0D"
-     "\x1E\x00"
-     "\xA8\x69\x1E\x00\xA9\x79\x1E\x00\xAA\x99\x1E\x00\xAB\x99\x1E\x00\xAC\xB5"
-     "\x1E\x00"
-     "\xAD\xBD\x1E\x00\xAE\xB5\x1E\x00\xAF\xAD\x1E\x00\xAF\x80\x00\x80\xB3\x80"
-     "\x00\x80"
-     "\xB7\x80\x00\x80\xBB\x80\x00\x80\xBF\x80\x00\x80\xC3\x80\x00\x80\xC7\x80"
-     "\x00\x80"
-     "\xCB\x80\x00\x80\xB8\x65\x01\x00\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x7D"
-     "\x01\x00"
-     "\xBC\x65\x01\x00\xBD\x6D\x01\x00\xBE\x65\x01\x00\xBF\xD9\x01\x00\xB0\xD5"
-     "\x1E\x00"
-     "\xB1\xA1\x1E\x00\xB2\xA1\x1E\x00\xB3\xA1\x1E\x00\xB4\xA1\x1E\x00\xB5\xA9"
-     "\x1E\x00"
-     "\xB6\x91\x1E\x00\xB7\x91\x1E\x00\xB3\x11\x1E\x00\xCF\x80\x00\x80\xD3\x80"
-     "\x00\x80"
-     "\xD7\x80\x00\x80\xDB\x80\x00\x80\xB6\x3D\x1E\x00\xB5\x3D\x1E\x00\xDF\x80"
-     "\x00\x80"
-     "\xBB\x65\x1E\x00\xBA\x11\x1E\x00\xE3\x80\x00\x80\xE7\x80\x00\x80\xBF\x6D"
-     "\x1E\x00"
-     "\xBE\x65\x1E\x00\xBD\x75\x1E\x00\xBC\x75\x1E\x00\xEB\x80\x00\x80\xA3\x55"
-     "\x1E\x00"
-     "\xEF\x80\x00\x80\xF3\x80\x00\x80\xA6\x79\x1E\x00\xF7\x80\x00\x80\xFB\x80"
-     "\x00\x80"
-     "\xA5\x79\x1E\x00\xAA\x55\x1E\x00\xAB\x21\x1E\x00\xFF\x80\x00\x80\x03\x81"
-     "\x00\x80"
-     "\xAE\x21\x1E\x00\xAF\x29\x1E\x00\xAC\x31\x1E\x00\xAD\x31\x1E\x00\xA8\xD1"
-     "\x02\x00"
-     "\xA9\xE5\x02\x00\xAA\xED\x02\x00\xAB\x2D\x03\x00\xAC\x35\x03\x00\xAD\x3D"
-     "\x03\x00"
-     "\xAE\x35\x03\x00\xAF\x2D\x03\x00\x07\x81\x00\x80\x0B\x81\x00\x80\x13\x81"
-     "\x00\x80"
-     "\xBE\x60\x02\x00\x17\x81\x00\x80\x1B\x81\x00\x80\x1F\x81\x00\x80\x23\x81"
-     "\x00\x80"
-     "\xB8\xE9\x03\x00\xB9\xE9\x03\x00\xBA\x89\x03\x00\xBB\x85\x03\x00\xBC\x9D"
-     "\x03\x00"
-     "\xBD\x81\x03\x00\xBE\x81\x03\x00\xBF\xB5\x03\x00\xB0\x55\x03\x00\xB1\x5D"
-     "\x03\x00"
-     "\xB2\x55\x03\x00\xB3\xE9\x03\x00\xB4\xF9\x03\x00\xB5\xF9\x03\x00\xB6\xE9"
-     "\x03\x00"
-     "\xB7\xE1\x03\x00\x80\x6D\x03\x00\x81\xA5\x00\x00\x82\xAD\x00\x00\xB3\x69"
-     "\x02\x00"
-     "\x27\x81\x00\x80\xB5\xB1\x03\x00\xB6\xB1\x03\x00\x2B\x81\x00\x80\x84\xE0"
-     "\x02\x00"
-     "\x2F\x81\x00\x80\xBA\x9D\x03\x00\xBB\x95\x03\x00\xBC\x8D\x03\x00\xBD\x31"
-     "\x03\x00"
-     "\xBE\x31\x03\x00\xBF\x31\x03\x00\xA3\x25\x02\x00\x33\x81\x00\x80\x37\x81"
-     "\x00\x80"
-     "\x3B\x81\x00\x80\x3F\x81\x00\x80\xA6\xFD\x03\x00\xA5\xFD\x03\x00\x43\x81"
-     "\x00\x80"
-     "\xAB\xD9\x03\x00\xAA\xD1\x03\x00\x86\x48\x04\x00\x87\xAC\x03\x00\xAF\x7D"
-     "\x03\x00"
-     "\xAE\x7D\x03\x00\xAD\x7D\x03\x00\xAC\xC1\x03\x00\x47\x81\x00\x80\x4B\x81"
-     "\x00\x80"
-     "\x4F\x81\x00\x80\x53\x81\x00\x80\xE3\xA0\x03\x00\x57\x81\x00\x80\xE1\xA8"
-     "\x01\x00"
-     "\x5B\x81\x00\x80\xEF\xF4\x03\x00\x5F\x81\x00\x80\x63\x81\x00\x80\x67\x81"
-     "\x00\x80"
-     "\x6B\x81\x00\x80\x6F\x81\x00\x80\x73\x81\x00\x80\x77\x81\x00\x80\x80\x79"
-     "\x00\x00"
-     "\x81\x79\x00\x00\x82\x1D\x00\x00\x84\x40\x02\x00\xE3\x60\x1E\x00\x7B\x81"
-     "\x00\x80"
-     "\xE1\xE0\x1F\x00\xBE\x50\x05\x00\xEF\xD0\x1E\x00\x83\x81\x00\x80\x86\xC8"
-     "\x04\x00"
-     "\x87\x40\x05\x00\x87\x81\x00\x80\x8B\x81\x00\x80\x8F\x81\x00\x80\x93\x81"
-     "\x00\x80"
-     "\x97\x81\x00\x80\x9B\x81\x00\x80\x9F\x81\x00\x80\xA3\x81\x00\x80\xA7\x81"
-     "\x00\x80"
-     "\xAB\x81\x00\x80\xAF\x81\x00\x80\xEF\xC4\x1E\x00\x84\xA4\x05\x00\xE1\x24"
-     "\x1E\x00"
-     "\xB3\x81\x00\x80\xE3\xF0\x1E\x00\xB7\x81\x00\x80\xBB\x81\x00\x80\xBF\x81"
-     "\x00\x80"
-     "\xC3\x81\x00\x80\xB3\xF5\x01\x00\xC7\x81\x00\x80\xCB\x81\x00\x80\xCF\x81"
-     "\x00\x80"
-     "\xD3\x81\x00\x80\xB6\x45\x01\x00\xB5\xE5\x01\x00\xD7\x81\x00\x80\xBB\x29"
-     "\x01\x00"
-     "\xBA\x21\x01\x00\xDB\x81\x00\x80\xDF\x81\x00\x80\xBF\xF1\x00\x00\xBE\x29"
-     "\x01\x00"
-     "\xBD\x29\x01\x00\xBC\x31\x01\x00\xA8\x29\x06\x00\xA9\x35\x06\x00\xAA\x3D"
-     "\x06\x00"
-     "\xAB\x35\x06\x00\xAC\x2D\x06\x00\xAD\x85\x06\x00\xAE\x85\x06\x00\xAF\xB9"
-     "\x06\x00"
-     "\x7F\x81\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\xE3\x81"
-     "\x00\x80"
-     "\xE7\x81\x00\x80\xEB\x81\x00\x80\xBE\x30\x03\x00\xB8\x4D\x07\x00\xB9\x51"
-     "\x07\x00"
-     "\xBA\x55\x07\x00\xBB\x69\x07\x00\xBC\x79\x07\x00\xBD\x79\x07\x00\xBE\x69"
-     "\x07\x00"
-     "\xBF\x69\x07\x00\xB0\xA5\x06\x00\xB1\xAD\x06\x00\xB2\xA5\x06\x00\xB3\xBD"
-     "\x06\x00"
-     "\xB4\xAD\x06\x00\xB5\x75\x07\x00\xB6\x7D\x07\x00\xB7\x75\x07\x00\xA3\xB5"
-     "\x06\x00"
-     "\xEF\x81\x00\x80\x86\x28\x00\x00\x87\x4C\x03\x00\xF3\x81\x00\x80\xA6\x05"
-     "\x06\x00"
-     "\xA5\xA5\x06\x00\xF7\x81\x00\x80\xAB\x69\x06\x00\xAA\x61\x06\x00\xFB\x81"
-     "\x00\x80"
-     "\xFF\x81\x00\x80\xAF\xB1\x07\x00\xAE\x69\x06\x00\xAD\x69\x06\x00\xAC\x71"
-     "\x06\x00"
-     "\x03\x82\x00\x80\x07\x82\x00\x80\xB3\xBD\x07\x00\x0B\x82\x00\x80\xB5\xAD"
-     "\x07\x00"
-     "\x0F\x82\x00\x80\x13\x82\x00\x80\xB6\x5D\x07\x00\x0F\x81\x00\x80\x17\x82"
-     "\x00\x80"
-     "\xBB\x79\x07\x00\xBA\x61\x07\x00\xBD\x69\x07\x00\xBC\x61\x07\x00\xBF\x49"
-     "\x07\x00"
-     "\xBE\x55\x07\x00\xA8\x65\x06\x00\xA9\x75\x06\x00\xAA\x7D\x06\x00\xAB\x75"
-     "\x06\x00"
-     "\xAC\x6D\x06\x00\xAD\x95\x06\x00\xAE\x9D\x06\x00\xAF\x95\x06\x00\x1B\x82"
-     "\x00\x80"
-     "\x1F\x82\x00\x80\x23\x82\x00\x80\x27\x82\x00\x80\x2B\x82\x00\x80\x2F\x82"
-     "\x00\x80"
-     "\x33\x82\x00\x80\x37\x82\x00\x80\xB8\xFD\x06\x00\xB9\x81\x06\x00\xBA\x81"
-     "\x06\x00"
-     "\xBB\x81\x06\x00\xBC\x81\x06\x00\xBD\x89\x06\x00\xBE\xB1\x06\x00\xBF\xB1"
-     "\x06\x00"
-     "\xB0\xED\x06\x00\xB1\xF5\x06\x00\xB2\xFD\x06\x00\xB3\xF5\x06\x00\xB4\xD9"
-     "\x06\x00"
-     "\xB5\xC5\x06\x00\xB6\xCD\x06\x00\xB7\xC5\x06\x00\xA3\xF5\x06\x00\x3B\x82"
-     "\x00\x80"
-     "\x3F\x82\x00\x80\x43\x82\x00\x80\x47\x82\x00\x80\xA6\x15\x06\x00\xA5\xE5"
-     "\x06\x00"
-     "\x4B\x82\x00\x80\xAB\x31\x06\x00\xAA\x29\x06\x00\x4F\x82\x00\x80\x53\x82"
-     "\x00\x80"
-     "\xAF\x01\x06\x00\xAE\x1D\x06\x00\xAD\x21\x06\x00\xAC\x29\x06\x00\x80\x6D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x19\x00\x00\x57\x82\x00\x80\x5B\x82\x00\x80\x84\x98"
-     "\x01\x00"
-     "\xBE\x90\x01\x00\x63\x82\x00\x80\x86\x40\x0F\x00\x87\xF4\x00\x00\x67\x82"
-     "\x00\x80"
-     "\x6B\x82\x00\x80\x6F\x82\x00\x80\x73\x82\x00\x80\x77\x82\x00\x80\x7B\x82"
-     "\x00\x80"
-     "\xA8\x91\x02\x00\xA9\x99\x02\x00\xAA\xA1\x02\x00\xAB\xA1\x02\x00\xAC\xD1"
-     "\x02\x00"
-     "\xAD\xD1\x02\x00\xAE\xD5\x02\x00\xAF\xC9\x02\x00\x7F\x82\x00\x80\x83\x82"
-     "\x00\x80"
-     "\x87\x82\x00\x80\x84\x20\x03\x00\x8B\x82\x00\x80\x8F\x82\x00\x80\x93\x82"
-     "\x00\x80"
-     "\x97\x82\x00\x80\xB8\x79\x03\x00\xB9\x79\x03\x00\xBA\xCD\x03\x00\xBB\xC5"
-     "\x03\x00"
-     "\xBC\xDD\x03\x00\xBD\xC5\x03\x00\xBE\xC5\x03\x00\xBF\xF5\x03\x00\xB0\xB9"
-     "\x02\x00"
-     "\xB1\x81\x02\x00\xB2\x49\x03\x00\xB3\x49\x03\x00\xB4\x59\x03\x00\xB5\x59"
-     "\x03\x00"
-     "\xB6\x49\x03\x00\xB7\x49\x03\x00\xB3\x29\x02\x00\x9B\x82\x00\x80\x9F\x82"
-     "\x00\x80"
-     "\xA3\x82\x00\x80\xBE\x9C\x0C\x00\xB6\x51\x02\x00\xB5\x59\x02\x00\xA7\x82"
-     "\x00\x80"
-     "\xBB\x45\x02\x00\xBA\x7D\x02\x00\xAB\x82\x00\x80\xAF\x82\x00\x80\xBF\x9D"
-     "\x03\x00"
-     "\xBE\x9D\x03\x00\xBD\x9D\x03\x00\xBC\x9D\x03\x00\xB3\x82\x00\x80\xA3\x6D"
-     "\x02\x00"
-     "\xB7\x82\x00\x80\xBB\x82\x00\x80\xA6\x15\x02\x00\xBF\x82\x00\x80\xC3\x82"
-     "\x00\x80"
-     "\xA5\x1D\x02\x00\xAA\x39\x02\x00\xAB\x01\x02\x00\xC7\x82\x00\x80\xCB\x82"
-     "\x00\x80"
-     "\xAE\xD9\x03\x00\xAF\xD9\x03\x00\xAC\xD9\x03\x00\xAD\xD9\x03\x00\xE1\x90"
-     "\x01\x00"
-     "\xD3\x82\x00\x80\xE3\xCC\x00\x00\x84\x18\x0D\x00\x80\xC9\x01\x00\x81\xFD"
-     "\x01\x00"
-     "\x82\xF5\x01\x00\xD7\x82\x00\x80\xBE\x2C\x0C\x00\xDB\x82\x00\x80\xDF\x82"
-     "\x00\x80"
-     "\xE3\x82\x00\x80\xE7\x82\x00\x80\xEB\x82\x00\x80\xEF\x18\x00\x00\xEF\x82"
-     "\x00\x80"
-     "\x86\xE0\x0C\x00\x87\xF4\x02\x00\xF3\x82\x00\x80\xF7\x82\x00\x80\xFB\x82"
-     "\x00\x80"
-     "\xFF\x82\x00\x80\x03\x83\x00\x80\xEF\xF8\x06\x00\x07\x83\x00\x80\xE1\x94"
-     "\x07\x00"
-     "\x0B\x83\x00\x80\xE3\x30\x06\x00\x0F\x83\x00\x80\x13\x83\x00\x80\x17\x83"
-     "\x00\x80"
-     "\x1B\x83\x00\x80\x1F\x83\x00\x80\x23\x83\x00\x80\x27\x83\x00\x80\x2B\x83"
-     "\x00\x80"
-     "\x2F\x83\x00\x80\x33\x83\x00\x80\x37\x83\x00\x80\xEF\x58\x01\x00\x85\x00"
-     "\xF0\x01"
-     "\xE1\xD4\x06\x00\x3B\x83\x00\x80\xE3\x1C\x06\x00\x80\x25\x00\x00\x81\x2D"
-     "\x00\x00"
-     "\x82\x11\x00\x00\x43\x83\x00\x80\xB3\x2D\x02\x00\xF1\x00\x00\x00\x87\xE8"
-     "\x0C\x00"
-     "\x86\x6C\x0F\x00\x47\x83\x00\x80\xB6\x2D\x02\x00\xB5\x2D\x02\x00\x4B\x83"
-     "\x00\x80"
-     "\xBB\xC9\x01\x00\xBA\xC1\x01\x00\x4F\x83\x00\x80\x53\x83\x00\x80\xBF\xCD"
-     "\x01\x00"
-     "\xBE\xCD\x01\x00\xBD\xCD\x01\x00\xBC\xCD\x01\x00\xA8\xDD\x0D\x00\xA9\xED"
-     "\x0D\x00"
-     "\xAA\xE5\x0D\x00\xAB\xFD\x0D\x00\xAC\xE5\x0D\x00\xAD\xED\x0D\x00\xAE\xE5"
-     "\x0D\x00"
-     "\xAF\x3D\x0E\x00\xCF\x82\x00\x80\x3F\x83\x00\x80\x57\x83\x00\x80\x5B\x83"
-     "\x00\x80"
-     "\x5F\x83\x00\x80\x63\x83\x00\x80\x67\x83\x00\x80\x6B\x83\x00\x80\xB8\xCD"
-     "\x0E\x00"
-     "\xB9\xD9\x0E\x00\xBA\xE1\x0E\x00\xBB\xE1\x0E\x00\xBC\x91\x0E\x00\xBD\x9D"
-     "\x0E\x00"
-     "\xBE\x95\x0E\x00\xBF\x89\x0E\x00\xB0\x45\x0E\x00\xB1\x4D\x0E\x00\xB2\x59"
-     "\x0E\x00"
-     "\xB3\x59\x0E\x00\xB4\x49\x0E\x00\xB5\x49\x0E\x00\xB6\xFD\x0E\x00\xB7\xF5"
-     "\x0E\x00"
-     "\xA3\x6D\x0D\x00\x6F\x83\x00\x80\x73\x83\x00\x80\x77\x83\x00\x80\x7B\x83"
-     "\x00\x80"
-     "\xA6\x6D\x0D\x00\xA5\x6D\x0D\x00\x7F\x83\x00\x80\xAB\x89\x0E\x00\xAA\x81"
-     "\x0E\x00"
-     "\x83\x83\x00\x80\x87\x83\x00\x80\xAF\x8D\x0E\x00\xAE\x8D\x0E\x00\xAD\x8D"
-     "\x0E\x00"
-     "\xAC\x8D\x0E\x00\xB5\xED\x0F\x00\x8B\x83\x00\x80\x8F\x83\x00\x80\xB6\xFD"
-     "\x0F\x00"
-     "\x93\x83\x00\x80\x97\x83\x00\x80\xB3\xFD\x0F\x00\x9B\x83\x00\x80\xBD\xB5"
-     "\x0F\x00"
-     "\xBC\xB5\x0F\x00\xBF\x89\x0F\x00\xBE\xB5\x0F\x00\x9F\x83\x00\x80\x5F\x82"
-     "\x00\x80"
-     "\xBB\xB5\x0F\x00\xBA\xB5\x0F\x00\xA8\x05\x0E\x00\xA9\x19\x0E\x00\xAA\x2D"
-     "\x0E\x00"
-     "\xAB\x25\x0E\x00\xAC\x3D\x0E\x00\xAD\x25\x0E\x00\xAE\x2D\x0E\x00\xAF\x21"
-     "\x0E\x00"
-     "\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\xA3\x83\x00\x80\xA7\x83"
-     "\x00\x80"
-     "\xAB\x83\x00\x80\x84\x30\x02\x00\xBE\x30\x01\x00\xB8\x35\x0E\x00\xB9\x3D"
-     "\x0E\x00"
-     "\xBA\x35\x0E\x00\xBB\xCD\x0E\x00\xBC\xD5\x0E\x00\xBD\xDD\x0E\x00\xBE\xD5"
-     "\x0E\x00"
-     "\xBF\xCD\x0E\x00\xB0\x29\x0E\x00\xB1\x29\x0E\x00\xB2\x39\x0E\x00\xB3\x39"
-     "\x0E\x00"
-     "\xB4\x29\x0E\x00\xB5\x15\x0E\x00\xB6\x11\x0E\x00\xB7\x0D\x0E\x00\xA3\xA5"
-     "\x0E\x00"
-     "\xB3\x83\x00\x80\x86\x28\x07\x00\x87\x4C\x01\x00\xB7\x83\x00\x80\xA6\xA5"
-     "\x0E\x00"
-     "\xA5\xB5\x0E\x00\xBB\x83\x00\x80\xAB\xED\x0E\x00\xAA\xED\x0E\x00\xBF\x83"
-     "\x00\x80"
-     "\xC3\x83\x00\x80\xAF\xD1\x0E\x00\xAE\xED\x0E\x00\xAD\xED\x0E\x00\xAC\xED"
-     "\x0E\x00"
-     "\xC7\x83\x00\x80\xB3\x19\x0E\x00\xCB\x83\x00\x80\xCF\x83\x00\x80\xB6\xF9"
-     "\x01\x00"
-     "\xD3\x83\x00\x80\xD7\x83\x00\x80\xB5\x31\x0E\x00\xBA\xD5\x01\x00\xBB\xDD"
-     "\x01\x00"
-     "\xDB\x83\x00\x80\xDF\x83\x00\x80\xBE\x79\x01\x00\xBF\x79\x01\x00\xBC\xC5"
-     "\x01\x00"
-     "\xBD\x79\x01\x00\xA8\xBD\x0D\x00\xA9\xD5\x0D\x00\xAA\xD5\x0D\x00\xAB\xE5"
-     "\x0D\x00"
-     "\xAC\xFD\x0D\x00\xAD\x3D\x02\x00\xAE\x35\x02\x00\xAF\x2D\x02\x00\xE3\x83"
-     "\x00\x80"
-     "\xE7\x83\x00\x80\xEB\x83\x00\x80\xEF\x83\x00\x80\xF3\x83\x00\x80\xF7\x83"
-     "\x00\x80"
-     "\xFB\x83\x00\x80\xFF\x83\x00\x80\xB8\xE9\x02\x00\xB9\xE9\x02\x00\xBA\x8D"
-     "\x02\x00"
-     "\xBB\x85\x02\x00\xBC\x81\x02\x00\xBD\x81\x02\x00\xBE\x85\x02\x00\xBF\xAD"
-     "\x02\x00"
-     "\xB0\x55\x02\x00\xB1\x5D\x02\x00\xB2\x55\x02\x00\xB3\xE9\x02\x00\xB4\xF9"
-     "\x02\x00"
-     "\xB5\xF9\x02\x00\xB6\xE9\x02\x00\xB7\xE1\x02\x00\xA3\x5D\x0D\x00\x03\x84"
-     "\x00\x80"
-     "\x07\x84\x00\x80\x0B\x84\x00\x80\x0F\x84\x00\x80\xA6\xBD\x02\x00\xA5\x75"
-     "\x0D\x00"
-     "\x13\x84\x00\x80\xAB\x99\x02\x00\xAA\x91\x02\x00\x17\x84\x00\x80\x1B\x84"
-     "\x00\x80"
-     "\xAF\x3D\x02\x00\xAE\x3D\x02\x00\xAD\x3D\x02\x00\xAC\x81\x02\x00\x80\x5D"
-     "\x00\x00"
-     "\x81\x6D\x00\x00\x82\x65\x00\x00\xB3\x51\x03\x00\x23\x84\x00\x80\xB5\x79"
-     "\x03\x00"
-     "\xB6\x19\x03\x00\x84\xDC\x02\x00\xBE\x6C\x03\x00\x27\x84\x00\x80\xBA\x3D"
-     "\x03\x00"
-     "\xBB\x35\x03\x00\xBC\x19\x03\x00\xBD\x19\x03\x00\xBE\xD9\x00\x00\xBF\xD9"
-     "\x00\x00"
-     "\xA3\x9D\x02\x00\x2B\x84\x00\x80\x86\x68\x05\x00\x87\xBC\x02\x00\x2F\x84"
-     "\x00\x80"
-     "\xA6\xD5\x02\x00\xA5\xB5\x02\x00\x33\x84\x00\x80\xAB\xF9\x02\x00\xAA\xF1"
-     "\x02\x00"
-     "\x84\x20\x02\x00\x37\x84\x00\x80\xAF\x15\x01\x00\xAE\x15\x01\x00\xAD\xD5"
-     "\x02\x00"
-     "\xAC\xD5\x02\x00\x3B\x84\x00\x80\x3F\x84\x00\x80\xEF\x20\x0F\x00\x43\x84"
-     "\x00\x80"
-     "\x47\x84\x00\x80\x4B\x84\x00\x80\x4F\x84\x00\x80\x53\x84\x00\x80\x57\x84"
-     "\x00\x80"
-     "\x5B\x84\x00\x80\x5F\x84\x00\x80\x63\x84\x00\x80\xE1\x88\x0C\x00\x67\x84"
-     "\x00\x80"
-     "\xE3\x44\x0F\x00\x6B\x84\x00\x80\x6F\x84\x00\x80\x73\x84\x00\x80\x77\x84"
-     "\x00\x80"
-     "\x7B\x84\x00\x80\xBE\x1C\x04\x00\x7F\x84\x00\x80\xEF\x5C\x00\x00\xEF\x44"
-     "\x0E\x00"
-     "\xE1\x74\x0E\x00\xE1\xFC\x0E\x00\xE3\x80\x00\x00\xE3\xB0\x0E\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x75\x00\x00\x82\x71\x00\x00\x83\x84\x00\x80\x87\x84\x00\x80\xB3\xC1"
-     "\x00\x00"
-     "\x8B\x84\x00\x80\x93\x84\x00\x80\xB6\xFD\x00\x00\x97\x84\x00\x80\x9B\x84"
-     "\x00\x80"
-     "\xB5\xC1\x00\x00\xBA\xAD\x00\x00\xBB\xB5\x00\x00\x86\xC8\x04\x00\x87\xA0"
-     "\x04\x00"
-     "\xBE\xAD\x00\x00\xBF\x95\x00\x00\xBC\xAD\x00\x00\xBD\xA5\x00\x00\xA8\x85"
-     "\x05\x00"
-     "\xA9\x95\x05\x00\xAA\x95\x05\x00\xAB\xA5\x05\x00\xAC\xBD\x05\x00\xAD\xD5"
-     "\x05\x00"
-     "\xAE\xD1\x05\x00\xAF\xD1\x05\x00\x1F\x84\x00\x80\x8F\x84\x00\x80\x9F\x84"
-     "\x00\x80"
-     "\xA3\x84\x00\x80\xA7\x84\x00\x80\xAB\x84\x00\x80\xAF\x84\x00\x80\xB3\x84"
-     "\x00\x80"
-     "\xB8\x71\x06\x00\xB9\x71\x06\x00\xBA\x71\x06\x00\xBB\x71\x06\x00\xBC\xD5"
-     "\x06\x00"
-     "\xBD\xDD\x06\x00\xBE\xD5\x06\x00\xBF\xCD\x06\x00\xB0\xB5\x05\x00\xB1\xBD"
-     "\x05\x00"
-     "\xB2\x81\x05\x00\xB3\x81\x05\x00\xB4\x51\x06\x00\xB5\x51\x06\x00\xB6\x51"
-     "\x06\x00"
-     "\xB7\x51\x06\x00\xB8\xED\x07\x00\xB9\xF1\x07\x00\xBA\xF5\x07\x00\xBB\x0D"
-     "\x07\x00"
-     "\xBC\x15\x07\x00\xBD\x1D\x07\x00\xBE\x15\x07\x00\xBF\x09\x07\x00\xB0\xED"
-     "\x07\x00"
-     "\xB1\xF5\x07\x00\xB2\xFD\x07\x00\xB3\xF5\x07\x00\xB4\xED\x07\x00\xB5\xD5"
-     "\x07\x00"
-     "\xB6\xDD\x07\x00\xB7\xD5\x07\x00\xA8\x11\x06\x00\xA9\x21\x06\x00\xAA\xCD"
-     "\x07\x00"
-     "\xAB\xDD\x07\x00\xAC\xCD\x07\x00\xAD\xF9\x07\x00\xAE\x99\x07\x00\xAF\x99"
-     "\x07\x00"
-     "\xB7\x84\x00\x80\xBB\x84\x00\x80\xBF\x84\x00\x80\xC3\x84\x00\x80\xC7\x84"
-     "\x00\x80"
-     "\xCB\x84\x00\x80\xAF\x83\x00\x80\xCF\x84\x00\x80\xA3\x05\x06\x00\xD3\x84"
-     "\x00\x80"
-     "\xD7\x84\x00\x80\xDB\x84\x00\x80\xDF\x84\x00\x80\xA6\x39\x06\x00\xA5\x05"
-     "\x06\x00"
-     "\xE3\x84\x00\x80\xAB\x71\x06\x00\xAA\x69\x06\x00\xE7\x84\x00\x80\xEB\x84"
-     "\x00\x80"
-     "\xAF\x51\x06\x00\xAE\x69\x06\x00\xAD\x61\x06\x00\xAC\x69\x06\x00\xEF\x84"
-     "\x00\x80"
-     "\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\xF3\x84\x00\x80\xF7\x84"
-     "\x00\x80"
-     "\xFB\x84\x00\x80\xBE\x94\x01\x00\x84\xAC\x01\x00\xFF\x84\x00\x80\x87\xE8"
-     "\x01\x00"
-     "\x86\x00\xFC\x01\x03\x85\x00\x80\x07\x85\x00\x80\x0B\x85\x00\x80\x0F\x85"
-     "\x00\x80"
-     "\xA8\x4D\x06\x00\xA9\x51\x06\x00\xAA\x51\x06\x00\xAB\xA5\x01\x00\xAC\xBD"
-     "\x01\x00"
-     "\xAD\xA5\x01\x00\xAE\xAD\x01\x00\xAF\xA5\x01\x00\x13\x85\x00\x80\x17\x85"
-     "\x00\x80"
-     "\x1B\x85\x00\x80\x1F\x85\x00\x80\x23\x85\x00\x80\x27\x85\x00\x80\x2B\x85"
-     "\x00\x80"
-     "\x2F\x85\x00\x80\xB8\xC5\x00\x00\xB9\xCD\x00\x00\xBA\xC5\x00\x00\xBB\xDD"
-     "\x00\x00"
-     "\xBC\xCD\x00\x00\xBD\xFD\x00\x00\xBE\xF5\x00\x00\xBF\x9D\x00\x00\xB0\xDD"
-     "\x01\x00"
-     "\xB1\x05\x01\x00\xB2\x0D\x01\x00\xB3\x05\x01\x00\xB4\x1D\x01\x00\xB5\x05"
-     "\x01\x00"
-     "\xB6\x0D\x01\x00\xB7\x01\x01\x00\xB3\xD9\x05\x00\x33\x85\x00\x80\x37\x85"
-     "\x00\x80"
-     "\x3B\x85\x00\x80\x3F\x85\x00\x80\xB6\x25\x02\x00\xB5\x35\x02\x00\x43\x85"
-     "\x00\x80"
-     "\xBB\xC1\x02\x00\xBA\x19\x02\x00\x47\x85\x00\x80\x4B\x85\x00\x80\xBF\xC1"
-     "\x02\x00"
-     "\xBE\xD9\x02\x00\xBD\xD1\x02\x00\xBC\xD9\x02\x00\x4F\x85\x00\x80\xA3\x9D"
-     "\x05\x00"
-     "\x53\x85\x00\x80\x57\x85\x00\x80\xA6\x61\x02\x00\x5B\x85\x00\x80\x5F\x85"
-     "\x00\x80"
-     "\xA5\x71\x02\x00\xAA\x5D\x02\x00\xAB\x85\x02\x00\x63\x85\x00\x80\x67\x85"
-     "\x00\x80"
-     "\xAE\x9D\x02\x00\xAF\x85\x02\x00\xAC\x9D\x02\x00\xAD\x95\x02\x00\x80\xE9"
-     "\x00\x00"
-     "\x81\xE9\x00\x00\x82\x05\x00\x00\x6B\x85\x00\x80\xBE\xDC\x03\x00\x84\x58"
-     "\x03\x00"
-     "\x73\x85\x00\x80\x77\x85\x00\x80\x86\x00\xFC\x01\x87\xEC\x03\x00\x7B\x85"
-     "\x00\x80"
-     "\x7F\x85\x00\x80\x83\x85\x00\x80\x87\x85\x00\x80\x8B\x85\x00\x80\x8F\x85"
-     "\x00\x80"
-     "\x93\x85\x00\x80\x97\x85\x00\x80\x9B\x85\x00\x80\x9F\x85\x00\x80\xA3\x85"
-     "\x00\x80"
-     "\xA7\x85\x00\x80\xEF\xB0\x02\x00\xAB\x85\x00\x80\xE1\x94\x01\x00\xAF\x85"
-     "\x00\x80"
-     "\xE3\x6C\x02\x00\xB3\x85\x00\x80\xB7\x85\x00\x80\xBB\x85\x00\x80\xBF\x85"
-     "\x00\x80"
-     "\xC3\x85\x00\x80\xB3\x79\x02\x00\xC7\x85\x00\x80\xCB\x85\x00\x80\xCF\x85"
-     "\x00\x80"
-     "\xD3\x85\x00\x80\xB6\xB1\x02\x00\xB5\xB9\x02\x00\xD7\x85\x00\x80\xBB\x95"
-     "\x02\x00"
-     "\xBA\x95\x02\x00\xDB\x85\x00\x80\x84\x64\xFD\x01\xBF\x39\x01\x00\xBE\x31"
-     "\x01\x00"
-     "\xBD\x85\x02\x00\xBC\x85\x02\x00\x82\x69\x00\x00\xEF\x5C\x07\x00\x80\x59"
-     "\x00\x00"
-     "\x81\x51\x00\x00\xE3\x85\x00\x80\xE7\x85\x00\x80\x87\x68\xFD\x01\x86\xFC"
-     "\xFC\x01"
-     "\xBE\x2C\xFD\x01\xEB\x85\x00\x80\xEF\x85\x00\x80\xF3\x85\x00\x80\x84\x3C"
-     "\xFC\x01"
-     "\xE3\xB8\x07\x00\xF7\x85\x00\x80\xE1\xEC\x04\x00\xFB\x85\x00\x80\xFF\x85"
-     "\x00\x80"
-     "\x03\x86\x00\x80\x84\xC0\xFF\x01\x07\x86\x00\x80\x0B\x86\x00\x80\xEF\x3C"
-     "\x00\x00"
-     "\xEF\xF4\x06\x00\xE1\x50\x06\x00\xE1\xB4\x06\x00\xE3\xD8\x00\x00\xE3\x94"
-     "\x06\x00"
-     "\x0F\x86\x00\x80\x13\x86\x00\x80\x17\x86\x00\x80\x1B\x86\x00\x80\xA3\xB9"
-     "\x02\x00"
-     "\x1F\x86\x00\x80\x23\x86\x00\x80\x27\x86\x00\x80\x2B\x86\x00\x80\xA6\x71"
-     "\x02\x00"
-     "\xA5\x79\x02\x00\x2F\x86\x00\x80\xAB\x55\x02\x00\xAA\x55\x02\x00\x33\x86"
-     "\x00\x80"
-     "\x37\x86\x00\x80\xAF\xF9\x01\x00\xAE\xF1\x01\x00\xAD\x45\x02\x00\xAC\x45"
-     "\x02\x00"
-     "\xA8\x99\xFE\x01\xA9\x99\xFE\x01\xAA\xA1\xFE\x01\xAB\xA1\xFE\x01\xAC\xA1"
-     "\xFE\x01"
-     "\xAD\xA1\xFE\x01\xAE\xA1\xFE\x01\xAF\xA1\xFE\x01\x3B\x86\x00\x80\x3F\x86"
-     "\x00\x80"
-     "\x43\x86\x00\x80\x47\x86\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x4B\x86\x00\x80\xB8\xAD\xFE\x01\xB9\xB5\xFE\x01\xBA\xBD\xFE\x01\xBB\xB5"
-     "\xFE\x01"
-     "\xBC\xAD\xFE\x01\xBD\x51\xFF\x01\xBE\x51\xFF\x01\xBF\x51\xFF\x01\xB0\xA1"
-     "\xFE\x01"
-     "\xB1\xA1\xFE\x01\xB2\xA1\xFE\x01\xB3\xA1\xFE\x01\xB4\xA1\xFE\x01\xB5\xA1"
-     "\xFE\x01"
-     "\xB6\xA5\xFE\x01\xB7\x99\xFE\x01\x6F\x85\x00\x80\xB3\x2D\xFE\x01\xBE\xC0"
-     "\x03\x00"
-     "\xDF\x85\x00\x80\xB6\x25\xFE\x01\x4F\x86\x00\x80\x53\x86\x00\x80\xB5\x2D"
-     "\xFE\x01"
-     "\xBA\x61\xFE\x01\xBB\x61\xFE\x01\x86\x08\x00\x00\x87\x9C\x00\x00\xBE\x61"
-     "\xFE\x01"
-     "\xBF\x61\xFE\x01\xBC\x71\xFE\x01\xBD\x71\xFE\x01\xA3\x69\xFE\x01\x57\x86"
-     "\x00\x80"
-     "\x5B\x86\x00\x80\x5F\x86\x00\x80\x63\x86\x00\x80\xA6\x61\xFE\x01\xA5\x69"
-     "\xFE\x01"
-     "\x67\x86\x00\x80\xAB\x25\xFE\x01\xAA\x25\xFE\x01\x6B\x86\x00\x80\x6F\x86"
-     "\x00\x80"
-     "\xAF\x25\xFE\x01\xAE\x25\xFE\x01\xAD\x35\xFE\x01\xAC\x35\xFE\x01\xA8\x51"
-     "\xFE\x01"
-     "\xA9\x51\xFE\x01\xAA\x6D\xFE\x01\xAB\x65\xFE\x01\xAC\x7D\xFE\x01\xAD\x6D"
-     "\xFE\x01"
-     "\xAE\x99\xFE\x01\xAF\x99\xFE\x01\x73\x86\x00\x80\x77\x86\x00\x80\x7B\x86"
-     "\x00\x80"
-     "\x7F\x86\x00\x80\x83\x86\x00\x80\x87\x86\x00\x80\x8B\x86\x00\x80\x8F\x86"
-     "\x00\x80"
-     "\xB8\x89\xFE\x01\xB9\x89\xFE\x01\xBA\x99\xFE\x01\xBB\x91\xFE\x01\xBC\xB9"
-     "\xFE\x01"
-     "\xBD\xB9\xFE\x01\xBE\x79\x01\x00\xBF\x79\x01\x00\xB0\xE9\xFE\x01\xB1\xF5"
-     "\xFE\x01"
-     "\xB2\xFD\xFE\x01\xB3\xDD\xFE\x01\xB4\xC5\xFE\x01\xB5\xCD\xFE\x01\xB6\xC5"
-     "\xFE\x01"
-     "\xB7\xB9\xFE\x01\xB3\xE5\xFE\x01\x93\x86\x00\x80\x97\x86\x00\x80\x9B\x86"
-     "\x00\x80"
-     "\x9F\x86\x00\x80\xB6\xED\xFE\x01\xB5\xF5\xFE\x01\xA3\x86\x00\x80\xBB\x31"
-     "\xFE\x01"
-     "\xBA\x2D\xFE\x01\xA7\x86\x00\x80\xAB\x86\x00\x80\xBF\x09\xFE\x01\xBE\x01"
-     "\xFE\x01"
-     "\xBD\x09\xFE\x01\xBC\x29\xFE\x01\xAF\x86\x00\x80\xA3\xA1\xFE\x01\xB3\x86"
-     "\x00\x80"
-     "\xB7\x86\x00\x80\xA6\xA9\xFE\x01\xBB\x86\x00\x80\xBE\x90\x01\x00\xA5\xB1"
-     "\xFE\x01"
-     "\xAA\x69\xFE\x01\xAB\x75\xFE\x01\xBF\xE0\x07\x00\x85\xE8\x07\x00\xAE\x45"
-     "\xFE\x01"
-     "\xAF\x4D\xFE\x01\xAC\x6D\xFE\x01\xAD\x4D\xFE\x01\xA8\x99\x02\x00\xA9\xC1"
-     "\x02\x00"
-     "\xAA\xC1\x02\x00\xAB\xC1\x02\x00\xAC\xC1\x02\x00\xAD\xC9\x02\x00\xAE\xF1"
-     "\x02\x00"
-     "\xAF\xF1\x02\x00\x86\x84\x05\x00\x87\x44\x03\x00\xC3\x86\x00\x80\xC7\x86"
-     "\x00\x80"
-     "\x80\xF9\x03\x00\x81\xF9\x03\x00\x82\xED\x03\x00\xCB\x86\x00\x80\xB8\xC5"
-     "\x03\x00"
-     "\xB9\xCD\x03\x00\xBA\xC5\x03\x00\xBB\xDD\x03\x00\xBC\xCD\x03\x00\xBD\xFD"
-     "\x03\x00"
-     "\xBE\xF5\x03\x00\xBF\x9D\x03\x00\xB0\x41\x03\x00\xB1\x41\x03\x00\xB2\x41"
-     "\x03\x00"
-     "\xB3\x41\x03\x00\xB4\x41\x03\x00\xB5\x41\x03\x00\xB6\x41\x03\x00\xB7\x41"
-     "\x03\x00"
-     "\xB3\x49\x02\x00\xCF\x86\x00\x80\xD3\x86\x00\x80\xD7\x86\x00\x80\xDB\x86"
-     "\x00\x80"
-     "\xB6\x49\x02\x00\xB5\x49\x02\x00\xDF\x86\x00\x80\xBB\x85\x03\x00\xBA\x85"
-     "\x03\x00"
-     "\xE3\x86\x00\x80\xE7\x86\x00\x80\xBF\x85\x03\x00\xBE\x85\x03\x00\xBD\x95"
-     "\x03\x00"
-     "\xBC\x95\x03\x00\xEB\x86\x00\x80\xA3\x0D\x02\x00\xEF\x86\x00\x80\xF3\x86"
-     "\x00\x80"
-     "\xA6\x0D\x02\x00\xF7\x86\x00\x80\xFB\x86\x00\x80\xA5\x0D\x02\x00\xAA\xC1"
-     "\x03\x00"
-     "\xAB\xC1\x03\x00\xFF\x86\x00\x80\x03\x87\x00\x80\xAE\xC1\x03\x00\xAF\xC1"
-     "\x03\x00"
-     "\xAC\xD1\x03\x00\xAD\xD1\x03\x00\xE3\x60\xFE\x01\xE1\xBC\xFF\x01\xE1\x4C"
-     "\xFE\x01"
-     "\xE3\x30\xFE\x01\xE1\xA0\x01\x00\xE1\xEC\x00\x00\xE3\x84\xFE\x01\xE3\xC8"
-     "\x1C\x00"
-     "\xBE\xAC\x07\x00\x07\x87\x00\x80\xEF\x08\x00\x00\xEF\xFC\x1B\x00\xEF\x40"
-     "\x01\x00"
-     "\x0B\x87\x00\x80\x0F\x87\x00\x80\xEF\xA4\xFE\x01\x13\x87\x00\x80\x17\x87"
-     "\x00\x80"
-     "\x84\x04\x04\x00\x84\xC4\x07\x00\x1B\x87\x00\x80\x1F\x87\x00\x80\x23\x87"
-     "\x00\x80"
-     "\x27\x87\x00\x80\x80\x39\x00\x00\x81\x39\x00\x00\x82\x05\x00\x00\x2B\x87"
-     "\x00\x80"
-     "\x86\x50\x07\x00\x87\x84\x05\x00\x33\x87\x00\x80\x37\x87\x00\x80\xBB\xC1"
-     "\xFD\x01"
-     "\xBA\xA9\xFD\x01\xB9\xA1\xFD\x01\xB8\x91\xFD\x01\xBF\xC9\xFD\x01\xBE\xC1"
-     "\xFD\x01"
-     "\xBD\xD5\xFD\x01\xBC\xD5\xFD\x01\xB3\x1D\xFE\x01\xB2\x1D\xFE\x01\xB1\x1D"
-     "\xFE\x01"
-     "\xB0\x1D\xFE\x01\xB7\xA1\xFD\x01\xB6\xB9\xFD\x01\xB5\xB1\xFD\x01\xB4\xB9"
-     "\xFD\x01"
-     "\xAB\xBD\xFE\x01\xAA\xBD\xFE\x01\xA9\xBD\xFE\x01\xA8\xBD\xFE\x01\xAF\x7D"
-     "\xFE\x01"
-     "\xAE\x7D\xFE\x01\xAD\x7D\xFE\x01\xAC\x7D\xFE\x01\x3B\x87\x00\x80\x3F\x87"
-     "\x00\x80"
-     "\x43\x87\x00\x80\x47\x87\x00\x80\x4B\x87\x00\x80\x4F\x87\x00\x80\xBE\x74"
-     "\x06\x00"
-     "\x53\x87\x00\x80\x82\x05\xEC\x01\x83\x19\xEC\x01\xBF\x86\x00\x80\x2F\x87"
-     "\x00\x80"
-     "\x86\xF9\xE8\x01\x87\xC9\xF4\x01\x84\x19\xE8\x01\x85\x19\xE8\x01\x8A\x01"
-     "\xF4\x01"
-     "\x8B\xB9\xF4\x01\x57\x87\x00\x80\x5B\x87\x00\x80\x8E\xB5\xF0\x01\x8F\xC1"
-     "\xFC\x01"
-     "\x8C\xD9\xF0\x01\x8D\x1D\xF0\x01\x92\x61\xFC\x01\x93\x61\xFC\x01\x5F\x87"
-     "\x00\x80"
-     "\x63\x87\x00\x80\x96\x75\xF8\x01\x97\x31\xF9\x01\x94\xA1\xF8\x01\x95\x7D"
-     "\xF8\x01"
-     "\x9A\xE1\x04\x00\x9B\xE9\x04\x00\x67\x87\x00\x80\x6B\x87\x00\x80\x6F\x87"
-     "\x00\x80"
-     "\x73\x87\x00\x80\x9C\xED\x04\x00\x77\x87\x00\x80\xA1\x89\x00\x00\x7B\x87"
-     "\x00\x80"
-     "\xA3\x01\x0C\x00\xA2\x81\x00\x00\xA5\x15\x0C\x00\xA4\xB5\x0C\x00\xA7\xB9"
-     "\x08\x00"
-     "\xA6\x01\x08\x00\xA9\xC9\x09\x00\xA8\x81\x08\x00\xAB\xAD\x14\x00\xAA\xAD"
-     "\x14\x00"
-     "\xAD\x01\x10\x00\xAC\x95\x15\x00\xAF\x55\x10\x00\xAE\x9D\x10\x00\xB1\xE1"
-     "\x1C\x00"
-     "\xB0\x01\x1C\x00\xB3\x01\x18\x00\xB2\xF5\x1C\x00\xB5\xF5\x18\x00\xB4\xF5"
-     "\x18\x00"
-     "\x7F\x87\x00\x80\x83\x87\x00\x80\x87\x87\x00\x80\x8B\x87\x00\x80\x8F\x87"
-     "\x00\x80"
-     "\x93\x87\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xC9\x01\x00\x97\x87"
-     "\x00\x80"
-     "\xA9\x2D\x07\x00\xA8\x2D\x07\x00\xAB\x8D\x07\x00\xAA\x31\x07\x00\xAD\x9D"
-     "\x07\x00"
-     "\xAC\x9D\x07\x00\xAF\x85\x07\x00\xAE\x8D\x07\x00\xA0\x45\xF8\x01\xA1\xA5"
-     "\x05\x00"
-     "\xA2\xA5\x05\x00\xA3\xB9\x05\x00\x9B\x87\x00\x80\xA3\x87\x00\x80\x86\x98"
-     "\x0F\x00"
-     "\x87\x30\x00\x00\xB9\x65\x06\x00\xB8\xA9\x07\x00\xBB\x71\x06\x00\xBA\x79"
-     "\x06\x00"
-     "\xBD\x15\x06\x00\xBC\x15\x06\x00\xBF\x0D\x06\x00\xBE\x0D\x06\x00\xB1\xCD"
-     "\x07\x00"
-     "\xB0\xCD\x07\x00\xB3\xA1\x07\x00\xB2\xAD\x07\x00\xB5\xA5\x07\x00\xB4\xB9"
-     "\x07\x00"
-     "\xB7\x91\x07\x00\xB6\xA5\x07\x00\xB3\xBD\x07\x00\xA7\x87\x00\x80\xAB\x87"
-     "\x00\x80"
-     "\x84\x0C\x00\x00\xAF\x87\x00\x80\xB6\x19\x07\x00\xB5\x11\x07\x00\x8F\x99"
-     "\xED\x01"
-     "\xBB\x0D\x07\x00\xBA\x0D\x07\x00\xB3\x87\x00\x80\xB7\x87\x00\x80\xBF\x65"
-     "\x07\x00"
-     "\xBE\x79\x07\x00\xBD\x71\x07\x00\xBC\x7D\x07\x00\x98\xE1\xF1\x01\x99\x41"
-     "\xF1\x01"
-     "\x9A\x3D\xFD\x01\x9B\xED\xFD\x01\x9C\xF5\xFD\x01\x9D\x2D\xF9\x01\x9E\xF9"
-     "\xF9\x01"
-     "\x9F\xF9\xF9\x01\x90\xFD\xED\x01\x91\xE5\xE9\x01\x92\xED\xE9\x01\x93\x05"
-     "\xE9\x01"
-     "\x94\x2D\xF5\x01\x95\xD5\xF5\x01\x96\x65\xF5\x01\x97\x1D\xF1\x01\xA3\x3D"
-     "\x04\x00"
-     "\x84\xC8\x0D\x00\xBE\x44\x03\x00\xBB\x87\x00\x80\xBF\x87\x00\x80\xA6\x99"
-     "\x04\x00"
-     "\xA5\x91\x04\x00\xC3\x87\x00\x80\xAB\x8D\x04\x00\xAA\x8D\x04\x00\xC7\x87"
-     "\x00\x80"
-     "\xCB\x87\x00\x80\xAF\xE5\x04\x00\xAE\xF9\x04\x00\xAD\xF1\x04\x00\xAC\xFD"
-     "\x04\x00"
-     "\xCF\x87\x00\x80\xB3\xA5\x05\x00\xD3\x87\x00\x80\xD7\x87\x00\x80\xB6\xBD"
-     "\x05\x00"
-     "\xDB\x87\x00\x80\xDF\x87\x00\x80\xB5\xB5\x05\x00\xBA\x9D\x05\x00\xBB\x4D"
-     "\x02\x00"
-     "\xE3\x87\x00\x80\xE7\x87\x00\x80\xBE\x3D\x02\x00\xBF\x1D\x02\x00\xBC\x55"
-     "\x02\x00"
-     "\xBD\x35\x02\x00\xA8\xF9\x05\x00\xA9\xF9\x05\x00\xAA\x69\x05\x00\xAB\x69"
-     "\x05\x00"
-     "\xAC\x79\x05\x00\xAD\x79\x05\x00\xAE\x69\x05\x00\xAF\x69\x05\x00\xEB\x87"
-     "\x00\x80"
-     "\xEF\x87\x00\x80\xF3\x87\x00\x80\xF7\x87\x00\x80\xFB\x87\x00\x80\xFF\x87"
-     "\x00\x80"
-     "\x03\x88\x00\x80\x07\x88\x00\x80\xB8\xB5\x02\x00\xB9\xB9\x02\x00\xBA\x4D"
-     "\x02\x00"
-     "\xBB\x15\x02\x00\xBC\x0D\x02\x00\xBD\x35\x02\x00\xBE\x3D\x02\x00\xBF\x35"
-     "\x02\x00"
-     "\xB0\x1D\x05\x00\xB1\xCD\x02\x00\xB2\xC5\x02\x00\xB3\xA5\x02\x00\xB4\xBD"
-     "\x02\x00"
-     "\xB5\x9D\x02\x00\xB6\x95\x02\x00\xB7\x8D\x02\x00\xA3\xE1\x04\x00\x82\x15"
-     "\x00\x00"
-     "\x81\xB5\x01\x00\x80\xB5\x01\x00\x0B\x88\x00\x80\xA6\xF9\x04\x00\xA5\xF1"
-     "\x04\x00"
-     "\x13\x88\x00\x80\xAB\x09\x03\x00\xAA\xD9\x04\x00\x17\x88\x00\x80\x1B\x88"
-     "\x00\x80"
-     "\xAF\x59\x03\x00\xAE\x79\x03\x00\xAD\x71\x03\x00\xAC\x11\x03\x00\x86\x40"
-     "\x0D\x00"
-     "\x87\x74\x0D\x00\x1F\x88\x00\x80\xB3\xC5\x02\x00\x23\x88\x00\x80\xB5\xFD"
-     "\x02\x00"
-     "\xB6\xF5\x02\x00\x27\x88\x00\x80\x2B\x88\x00\x80\x2F\x88\x00\x80\xBA\x75"
-     "\x01\x00"
-     "\xBB\x45\x01\x00\xBC\x5D\x01\x00\xBD\x35\x01\x00\xBE\x31\x01\x00\xBF\x31"
-     "\x01\x00"
-     "\xE3\xDC\x06\x00\xE3\xDC\x06\x00\xE1\xE0\x06\x00\xE1\xB4\x07\x00\x33\x88"
-     "\x00\x80"
-     "\x37\x88\x00\x80\x3B\x88\x00\x80\xBE\x54\x0C\x00\x86\x80\x0C\x00\x87\x18"
-     "\x0C\x00"
-     "\x3F\x88\x00\x80\x43\x88\x00\x80\xEF\xB0\x06\x00\xEF\x6C\x06\x00\x47\x88"
-     "\x00\x80"
-     "\x4B\x88\x00\x80\xA3\x49\x02\x00\x4F\x88\x00\x80\x53\x88\x00\x80\x57\x88"
-     "\x00\x80"
-     "\x5B\x88\x00\x80\xA6\x79\x02\x00\xA5\x71\x02\x00\x5F\x88\x00\x80\xAB\xC9"
-     "\x01\x00"
-     "\xAA\xF9\x01\x00\x63\x88\x00\x80\x67\x88\x00\x80\xAF\xBD\x01\x00\xAE\xBD"
-     "\x01\x00"
-     "\xAD\xB9\x01\x00\xAC\xD1\x01\x00\xA8\x35\x0D\x00\xA9\x81\x0D\x00\xAA\x81"
-     "\x0D\x00"
-     "\xAB\x9D\x0D\x00\xAC\xB9\x0D\x00\xAD\xB9\x0D\x00\xAE\x15\x0E\x00\xAF\x0D"
-     "\x0E\x00"
-     "\x0F\x88\x00\x80\x6B\x88\x00\x80\x6F\x88\x00\x80\x73\x88\x00\x80\x80\x1D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\xB9\x00\x00\x77\x88\x00\x80\xB8\x29\x0E\x00\xB9\x29"
-     "\x0E\x00"
-     "\xBA\xE9\x0E\x00\xBB\xE9\x0E\x00\xBC\xF9\x0E\x00\xBD\xF9\x0E\x00\xBE\xE9"
-     "\x0E\x00"
-     "\xBF\xE9\x0E\x00\xB0\x7D\x0E\x00\xB1\x4D\x0E\x00\xB2\x45\x0E\x00\xB3\x2D"
-     "\x0E\x00"
-     "\xB4\x39\x0E\x00\xB5\x39\x0E\x00\xB6\x2D\x0E\x00\xB7\x25\x0E\x00\x7B\x88"
-     "\x00\x80"
-     "\x7F\x88\x00\x80\x9F\x87\x00\x80\x83\x88\x00\x80\x87\x88\x00\x80\x8B\x88"
-     "\x00\x80"
-     "\x8F\x88\x00\x80\x93\x88\x00\x80\xA8\xBD\x0F\x00\xA9\x8D\x0F\x00\xAA\x85"
-     "\x0F\x00"
-     "\xAB\xE5\x0F\x00\xAC\xE1\x0F\x00\xAD\xE9\x0F\x00\xAE\xD9\x0F\x00\xAF\xD9"
-     "\x0F\x00"
-     "\xB0\x51\x0F\x00\xB1\x51\x0F\x00\xB2\x51\x0F\x00\xB3\x65\x0F\x00\xB4\x7D"
-     "\x0F\x00"
-     "\xB5\x6D\x0F\x00\xB6\x65\x0F\x00\xB7\x19\x0F\x00\xB8\x29\x0F\x00\xB9\x29"
-     "\x0F\x00"
-     "\xBA\x39\x0F\x00\xBB\x31\x0F\x00\xBC\x19\x0F\x00\xBD\x19\x0F\x00\xBE\xDD"
-     "\x00\x00"
-     "\xBF\xD5\x00\x00\xB3\x2D\x0E\x00\x97\x88\x00\x80\x9B\x88\x00\x80\x9F\x88"
-     "\x00\x80"
-     "\xA3\x88\x00\x80\xB6\x49\x0E\x00\xB5\x49\x0E\x00\xA7\x88\x00\x80\xBB\xB5"
-     "\x0E\x00"
-     "\xBA\xB5\x0E\x00\xAB\x88\x00\x80\xAF\x88\x00\x80\xBF\x89\x0E\x00\xBE\x99"
-     "\x0E\x00"
-     "\xBD\x91\x0E\x00\xBC\xA5\x0E\x00\xB3\x88\x00\x80\xA3\x69\x0E\x00\xB7\x88"
-     "\x00\x80"
-     "\xBB\x88\x00\x80\xA6\x0D\x0E\x00\xBF\x88\x00\x80\xC3\x88\x00\x80\xA5\x0D"
-     "\x0E\x00"
-     "\xAA\xF1\x0E\x00\xAB\xF1\x0E\x00\xC7\x88\x00\x80\xCB\x88\x00\x80\xAE\xDD"
-     "\x0E\x00"
-     "\xAF\xCD\x0E\x00\xAC\xE1\x0E\x00\xAD\xD5\x0E\x00\xA8\x0D\x01\x00\xA9\x15"
-     "\x01\x00"
-     "\xAA\x1D\x01\x00\xAB\x15\x01\x00\xAC\x0D\x01\x00\xAD\x91\x00\x00\xAE\x91"
-     "\x00\x00"
-     "\xAF\x91\x00\x00\xCF\x88\x00\x80\xD3\x88\x00\x80\xD7\x88\x00\x80\xDB\x88"
-     "\x00\x80"
-     "\xDF\x88\x00\x80\x82\xB9\x03\x00\x81\xB1\x03\x00\x80\xB9\x03\x00\xB8\x99"
-     "\x00\x00"
-     "\xB9\xAD\x00\x00\xBA\xA5\x00\x00\xBB\x6D\x00\x00\xBC\x75\x00\x00\xBD\x7D"
-     "\x00\x00"
-     "\xBE\x75\x00\x00\xBF\x6D\x00\x00\xB0\xF1\x00\x00\xB1\xF9\x00\x00\xB2\xC1"
-     "\x00\x00"
-     "\xB3\xC1\x00\x00\xB4\xB1\x00\x00\xB5\xBD\x00\x00\xB6\xB5\x00\x00\xB7\xA9"
-     "\x00\x00"
-     "\x84\xAC\x02\x00\xE3\x88\x00\x80\xE7\x88\x00\x80\xEB\x88\x00\x80\xBE\x1C"
-     "\x02\x00"
-     "\xF3\x88\x00\x80\x86\xB8\x04\x00\x87\xCC\x02\x00\xF7\x88\x00\x80\xFB\x88"
-     "\x00\x80"
-     "\xFF\x88\x00\x80\x03\x89\x00\x80\xBE\x74\x03\x00\x07\x89\x00\x80\x0B\x89"
-     "\x00\x80"
-     "\x0F\x89\x00\x80\xBB\x5D\x03\x00\xBA\x55\x03\x00\x13\x89\x00\x80\x17\x89"
-     "\x00\x80"
-     "\xBF\x19\x03\x00\xBE\x15\x03\x00\xBD\x75\x03\x00\xBC\x75\x03\x00\xB3\xF5"
-     "\x02\x00"
-     "\x1B\x89\x00\x80\x1F\x89\x00\x80\x23\x89\x00\x80\x27\x89\x00\x80\xB6\x71"
-     "\x03\x00"
-     "\xB5\xE5\x02\x00\x2B\x89\x00\x80\xA3\x15\x02\x00\x2F\x89\x00\x80\x33\x89"
-     "\x00\x80"
-     "\x37\x89\x00\x80\x3B\x89\x00\x80\xA6\x91\x03\x00\xA5\x05\x02\x00\x3F\x89"
-     "\x00\x80"
-     "\xAB\xBD\x03\x00\xAA\xB5\x03\x00\x43\x89\x00\x80\x47\x89\x00\x80\xAF\xF9"
-     "\x03\x00"
-     "\xAE\xF5\x03\x00\xAD\x95\x03\x00\xAC\x95\x03\x00\x4B\x89\x00\x80\x4F\x89"
-     "\x00\x80"
-     "\x53\x89\x00\x80\x57\x89\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x5B\x89\x00\x80\x84\x6C\x03\x00\x63\x89\x00\x80\x87\x78\x03\x00\x86\xEC"
-     "\x04\x00"
-     "\x67\x89\x00\x80\x6B\x89\x00\x80\x6F\x89\x00\x80\x73\x89\x00\x80\xE1\xB8"
-     "\x0E\x00"
-     "\xE3\x98\x0C\x00\xE3\x18\x03\x00\xE1\x2C\x0C\x00\x77\x89\x00\x80\x7B\x89"
-     "\x00\x80"
-     "\x7F\x89\x00\x80\x83\x89\x00\x80\x84\xEC\x04\x00\x87\x89\x00\x80\x8B\x89"
-     "\x00\x80"
-     "\x8F\x89\x00\x80\x93\x89\x00\x80\xEF\xC4\x0F\x00\xEF\xD0\x03\x00\x97\x89"
-     "\x00\x80"
-     "\xB3\x65\x01\x00\x9B\x89\x00\x80\x9F\x89\x00\x80\xA3\x89\x00\x80\xA7\x89"
-     "\x00\x80"
-     "\xB6\x61\x01\x00\xB5\x75\x01\x00\xAB\x89\x00\x80\xBB\x2D\x01\x00\xBA\x2D"
-     "\x01\x00"
-     "\xAF\x89\x00\x80\xB3\x89\x00\x80\xBF\x01\x01\x00\xBE\x19\x01\x00\xBD\x11"
-     "\x01\x00"
-     "\xBC\x19\x01\x00\xB7\x89\x00\x80\xBB\x89\x00\x80\xBF\x89\x00\x80\xC3\x89"
-     "\x00\x80"
-     "\x80\x1D\x00\x00\x81\x35\x00\x00\x82\x31\x00\x00\xEF\xF8\x0E\x00\xC7\x89"
-     "\x00\x80"
-     "\xE1\xFC\x0E\x00\xCB\x89\x00\x80\xE3\xB8\x0F\x00\xD3\x89\x00\x80\xD7\x89"
-     "\x00\x80"
-     "\xDB\x89\x00\x80\xEE\x00\x00\x00\xBE\x08\x04\x00\xDF\x89\x00\x80\xA3\xED"
-     "\x01\x00"
-     "\x84\x00\x07\x00\xA5\xFD\x01\x00\xE3\x89\x00\x80\xE7\x89\x00\x80\xA6\xE9"
-     "\x01\x00"
-     "\x86\xC0\x04\x00\x87\xAC\x04\x00\xAB\xA5\x01\x00\xAA\xA5\x01\x00\xAD\x99"
-     "\x01\x00"
-     "\xAC\x91\x01\x00\xAF\x89\x01\x00\xAE\x91\x01\x00\xA8\x15\x06\x00\xA9\x25"
-     "\x06\x00"
-     "\xAA\x2D\x06\x00\xAB\x25\x06\x00\xAC\x3D\x06\x00\xAD\x21\x06\x00\xAE\x21"
-     "\x06\x00"
-     "\xAF\x21\x06\x00\x5F\x89\x00\x80\xCF\x89\x00\x80\xEB\x89\x00\x80\xEF\x89"
-     "\x00\x80"
-     "\xF3\x89\x00\x80\xF7\x89\x00\x80\xFB\x89\x00\x80\xFF\x89\x00\x80\xB8\xC9"
-     "\x06\x00"
-     "\xB9\xC9\x06\x00\xBA\xD9\x06\x00\xBB\xD1\x06\x00\xBC\xF9\x06\x00\xBD\xF9"
-     "\x06\x00"
-     "\xBE\x99\x06\x00\xBF\x99\x06\x00\xB0\x29\x06\x00\xB1\x29\x06\x00\xB2\x0D"
-     "\x06\x00"
-     "\xB3\x05\x06\x00\xB4\x1D\x06\x00\xB5\x05\x06\x00\xB6\x05\x06\x00\xB7\xF9"
-     "\x06\x00"
-     "\x03\x8A\x00\x80\x07\x8A\x00\x80\x0B\x8A\x00\x80\x0F\x8A\x00\x80\x13\x8A"
-     "\x00\x80"
-     "\x17\x8A\x00\x80\x1B\x8A\x00\x80\xEF\x88\x00\x80\xA8\x45\x06\x00\xA9\x55"
-     "\x06\x00"
-     "\xAA\x5D\x06\x00\xAB\x55\x06\x00\xAC\x4D\x06\x00\xAD\xA5\x07\x00\xAE\xAD"
-     "\x07\x00"
-     "\xAF\xA5\x07\x00\xB0\xDD\x07\x00\xB1\xE1\x07\x00\xB2\xE1\x07\x00\xB3\xE1"
-     "\x07\x00"
-     "\xB4\xE1\x07\x00\xB5\xE9\x07\x00\xB6\x9D\x07\x00\xB7\x95\x07\x00\xB8\xAD"
-     "\x07\x00"
-     "\xB9\xB5\x07\x00\xBA\xB5\x07\x00\xBB\x85\x07\x00\xBC\x9D\x07\x00\xBD\x75"
-     "\x07\x00"
-     "\xBE\x7D\x07\x00\xBF\x75\x07\x00\xB3\xF5\x07\x00\x1F\x8A\x00\x80\x23\x8A"
-     "\x00\x80"
-     "\x27\x8A\x00\x80\x2B\x8A\x00\x80\xB6\x05\x06\x00\xB5\xE5\x07\x00\x2F\x8A"
-     "\x00\x80"
-     "\xBB\x05\x06\x00\xBA\x39\x06\x00\x33\x8A\x00\x80\x37\x8A\x00\x80\xBF\x0D"
-     "\x06\x00"
-     "\xBE\x05\x06\x00\xBD\x15\x06\x00\xBC\x15\x06\x00\x82\x15\x00\x00\xA3\xB1"
-     "\x07\x00"
-     "\x80\x61\x00\x00\x81\x61\x00\x00\xA6\x41\x06\x00\x3B\x8A\x00\x80\xBE\x90"
-     "\x01\x00"
-     "\xA5\xA1\x07\x00\xAA\x7D\x06\x00\xAB\x41\x06\x00\x3F\x8A\x00\x80\x43\x8A"
-     "\x00\x80"
-     "\xAE\x41\x06\x00\xAF\x49\x06\x00\xAC\x51\x06\x00\xAD\x51\x06\x00\xA8\x31"
-     "\x06\x00"
-     "\xA9\x31\x06\x00\xAA\x31\x06\x00\xAB\x49\x06\x00\xAC\x45\x06\x00\xAD\x4D"
-     "\x06\x00"
-     "\xAE\xB1\x01\x00\xAF\xB1\x01\x00\x86\x00\x1C\x00\x87\x04\x01\x00\x47\x8A"
-     "\x00\x80"
-     "\x4B\x8A\x00\x80\x4F\x8A\x00\x80\x53\x8A\x00\x80\x57\x8A\x00\x80\x5B\x8A"
-     "\x00\x80"
-     "\xB8\x51\x01\x00\xB9\x51\x01\x00\xBA\x51\x01\x00\xBB\xC9\x00\x00\xBC\xD9"
-     "\x00\x00"
-     "\xBD\xD9\x00\x00\xBE\xC9\x00\x00\xBF\xC1\x00\x00\xB0\xD1\x01\x00\xB1\xD1"
-     "\x01\x00"
-     "\xB2\xD1\x01\x00\xB3\xD1\x01\x00\xB4\x71\x01\x00\xB5\x71\x01\x00\xB6\x71"
-     "\x01\x00"
-     "\xB7\x71\x01\x00\xB3\xB9\x05\x00\x5F\x8A\x00\x80\x63\x8A\x00\x80\x67\x8A"
-     "\x00\x80"
-     "\x6B\x8A\x00\x80\xB6\xCD\x05\x00\xB5\xCD\x05\x00\x6F\x8A\x00\x80\xBB\x15"
-     "\x02\x00"
-     "\xBA\x15\x02\x00\x73\x8A\x00\x80\x77\x8A\x00\x80\xBF\xB5\x02\x00\xBE\xB5"
-     "\x02\x00"
-     "\xBD\x05\x02\x00\xBC\x05\x02\x00\x7B\x8A\x00\x80\xA3\xFD\x05\x00\x7F\x8A"
-     "\x00\x80"
-     "\x83\x8A\x00\x80\xA6\x89\x05\x00\x87\x8A\x00\x80\x8B\x8A\x00\x80\xA5\x89"
-     "\x05\x00"
-     "\xAA\x51\x02\x00\xAB\x51\x02\x00\x8F\x8A\x00\x80\x93\x8A\x00\x80\xAE\xF1"
-     "\x02\x00"
-     "\xAF\xF1\x02\x00\xAC\x41\x02\x00\xAD\x41\x02\x00\x97\x8A\x00\x80\x9B\x8A"
-     "\x00\x80"
-     "\x9F\x8A\x00\x80\xA3\x8A\x00\x80\xA7\x8A\x00\x80\xAB\x8A\x00\x80\xBE\xD4"
-     "\x03\x00"
-     "\xB3\x8A\x00\x80\x80\x2D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\x84\x20"
-     "\x03\x00"
-     "\x86\x30\x1C\x00\x87\x84\x03\x00\xB7\x8A\x00\x80\xBB\x8A\x00\x80\xBF\x8A"
-     "\x00\x80"
-     "\xC3\x8A\x00\x80\xC7\x8A\x00\x80\xCB\x8A\x00\x80\xCF\x8A\x00\x80\xEF\xB4"
-     "\x02\x00"
-     "\xD3\x8A\x00\x80\xD7\x8A\x00\x80\xDB\x8A\x00\x80\xE3\x58\x02\x00\xDF\x8A"
-     "\x00\x80"
-     "\xE1\x54\x01\x00\xE3\x8A\x00\x80\xE7\x8A\x00\x80\xEB\x8A\x00\x80\xEF\x8A"
-     "\x00\x80"
-     "\xF3\x8A\x00\x80\xF7\x8A\x00\x80\xFB\x8A\x00\x80\xFF\x8A\x00\x80\xE3\xE0"
-     "\x07\x00"
-     "\x03\x8B\x00\x80\xE1\x18\x07\x00\x07\x8B\x00\x80\xEF\xE8\x07\x00\x0B\x8B"
-     "\x00\x80"
-     "\x0F\x8B\x00\x80\x13\x8B\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x7D"
-     "\x00\x00"
-     "\x17\x8B\x00\x80\xB3\xED\x02\x00\x1B\x8B\x00\x80\x84\x04\x1D\x00\x23\x8B"
-     "\x00\x80"
-     "\x27\x8B\x00\x80\xB6\xED\x02\x00\xB5\xED\x02\x00\xBE\x14\x1D\x00\xBB\x6D"
-     "\x01\x00"
-     "\xBA\x6D\x01\x00\x86\xC8\x1C\x00\x87\xA4\x1D\x00\xBF\x5D\x01\x00\xBE\x5D"
-     "\x01\x00"
-     "\xBD\x65\x01\x00\xBC\x6D\x01\x00\x2B\x8B\x00\x80\x2F\x8B\x00\x80\x33\x8B"
-     "\x00\x80"
-     "\x37\x8B\x00\x80\x3B\x8B\x00\x80\x3F\x8B\x00\x80\xEF\x38\x00\x00\xEF\xA4"
-     "\x06\x00"
-     "\xE1\x58\x06\x00\xE1\x50\x07\x00\xE3\x74\x06\x00\xE3\x9C\x06\x00\x43\x8B"
-     "\x00\x80"
-     "\x47\x8B\x00\x80\xBE\xF0\x1C\x00\x4B\x8B\x00\x80\xA3\x6D\x02\x00\x4F\x8B"
-     "\x00\x80"
-     "\x84\x84\x1C\x00\x84\xE0\x1F\x00\x53\x8B\x00\x80\xA6\x6D\x02\x00\xA5\x6D"
-     "\x02\x00"
-     "\x57\x8B\x00\x80\xAB\xED\x01\x00\xAA\xED\x01\x00\x5B\x8B\x00\x80\x5F\x8B"
-     "\x00\x80"
-     "\xAF\xDD\x01\x00\xAE\xDD\x01\x00\xAD\xE5\x01\x00\xAC\xED\x01\x00\xAF\x8A"
-     "\x00\x80"
-     "\x1F\x8B\x00\x80\x63\x8B\x00\x80\x67\x8B\x00\x80\x6B\x8B\x00\x80\x6F\x8B"
-     "\x00\x80"
-     "\x73\x8B\x00\x80\x77\x8B\x00\x80\xA8\xBD\x1E\x00\xA9\x9D\x1E\x00\xAA\x95"
-     "\x1E\x00"
-     "\xAB\xED\x1E\x00\xAC\xF5\x1E\x00\xAD\xFD\x1E\x00\xAE\xF5\x1E\x00\xAF\xED"
-     "\x1E\x00"
-     "\xB0\x99\x1E\x00\xB1\x99\x1E\x00\xB2\xA9\x1E\x00\xB3\xA5\x1E\x00\xB4\xA1"
-     "\x1E\x00"
-     "\xB5\xA1\x1E\x00\xB6\xA1\x1E\x00\xB7\xA1\x1E\x00\xB8\x61\x1F\x00\xB9\x61"
-     "\x1F\x00"
-     "\xBA\x61\x1F\x00\xBB\x61\x1F\x00\xBC\x65\x1F\x00\xBD\x69\x1F\x00\xBE\x5D"
-     "\x1F\x00"
-     "\xBF\x55\x1F\x00\xB3\x3D\x1E\x00\x7B\x8B\x00\x80\x7F\x8B\x00\x80\x83\x8B"
-     "\x00\x80"
-     "\x87\x8B\x00\x80\xB6\x5D\x1E\x00\xB5\x5D\x1E\x00\x8B\x8B\x00\x80\xBB\x7D"
-     "\x1E\x00"
-     "\xBA\x7D\x1E\x00\x86\x08\x00\x00\x87\x8C\x00\x00\xBF\x45\x1E\x00\xBE\x45"
-     "\x1E\x00"
-     "\xBD\x51\x1E\x00\xBC\x5D\x1E\x00\x82\x5D\x00\x00\xA3\x79\x1E\x00\x80\x45"
-     "\x00\x00"
-     "\x81\x5D\x00\x00\xA6\x19\x1E\x00\x93\x8B\x00\x80\x97\x8B\x00\x80\xA5\x19"
-     "\x1E\x00"
-     "\xAA\x39\x1E\x00\xAB\x39\x1E\x00\x9B\x8B\x00\x80\x9F\x8B\x00\x80\xAE\x01"
-     "\x1E\x00"
-     "\xAF\x01\x1E\x00\xAC\x19\x1E\x00\xAD\x15\x1E\x00\xA8\x4D\x1E\x00\xA9\x55"
-     "\x1E\x00"
-     "\xAA\x89\x1E\x00\xAB\x89\x1E\x00\xAC\x99\x1E\x00\xAD\x99\x1E\x00\xAE\x89"
-     "\x1E\x00"
-     "\xAF\x89\x1E\x00\x8F\x8B\x00\x80\xA3\x8B\x00\x80\xA7\x8B\x00\x80\xAB\x8B"
-     "\x00\x80"
-     "\xAF\x8B\x00\x80\xB3\x8B\x00\x80\xB7\x8B\x00\x80\xBB\x8B\x00\x80\xB8\x49"
-     "\x01\x00"
-     "\xB9\x49\x01\x00\xBA\x59\x01\x00\xBB\x59\x01\x00\xBC\x49\x01\x00\xBD\x49"
-     "\x01\x00"
-     "\xBE\xDD\x01\x00\xBF\xD5\x01\x00\xB0\xD1\x1E\x00\xB1\xD1\x1E\x00\xB2\xD5"
-     "\x1E\x00"
-     "\xB3\xA9\x1E\x00\xB4\xA1\x1E\x00\xB5\xA9\x1E\x00\xB6\x91\x1E\x00\xB7\x91"
-     "\x1E\x00"
-     "\xB3\x01\x1E\x00\xBF\x8B\x00\x80\xC3\x8B\x00\x80\xC7\x8B\x00\x80\xCB\x8B"
-     "\x00\x80"
-     "\xB6\x19\x1E\x00\xB5\x11\x1E\x00\xCF\x8B\x00\x80\xBB\x15\x1E\x00\xBA\x15"
-     "\x1E\x00"
-     "\xD3\x8B\x00\x80\xD7\x8B\x00\x80\xBF\x6D\x1E\x00\xBE\x65\x1E\x00\xBD\x7D"
-     "\x1E\x00"
-     "\xBC\x01\x1E\x00\xDB\x8B\x00\x80\xA3\x45\x1E\x00\xDF\x8B\x00\x80\xE3\x8B"
-     "\x00\x80"
-     "\xA6\x5D\x1E\x00\xE7\x8B\x00\x80\xEB\x8B\x00\x80\xA5\x55\x1E\x00\xAA\x51"
-     "\x1E\x00"
-     "\xAB\x51\x1E\x00\xEF\x8B\x00\x80\xF3\x8B\x00\x80\xAE\x21\x1E\x00\xAF\x29"
-     "\x1E\x00"
-     "\xAC\x45\x1E\x00\xAD\x39\x1E\x00\xA8\xD5\x02\x00\xA9\xDD\x02\x00\xAA\xE1"
-     "\x02\x00"
-     "\xAB\xE1\x02\x00\xAC\x35\x03\x00\xAD\x3D\x03\x00\xAE\x35\x03\x00\xAF\x2D"
-     "\x03\x00"
-     "\x80\xCD\x03\x00\x81\x09\x00\x00\x82\x19\x00\x00\xF7\x8B\x00\x80\xFB\x8B"
-     "\x00\x80"
-     "\xBE\x78\x03\x00\x87\xAC\x03\x00\x86\x9C\x04\x00\xB8\xE9\x03\x00\xB9\xE9"
-     "\x03\x00"
-     "\xBA\x89\x03\x00\xBB\x85\x03\x00\xBC\x9D\x03\x00\xBD\x81\x03\x00\xBE\x81"
-     "\x03\x00"
-     "\xBF\xB5\x03\x00\xB0\x55\x03\x00\xB1\x5D\x03\x00\xB2\x55\x03\x00\xB3\xE9"
-     "\x03\x00"
-     "\xB4\xF9\x03\x00\xB5\xF9\x03\x00\xB6\xE9\x03\x00\xB7\xE1\x03\x00\xBE\xCC"
-     "\x04\x00"
-     "\x03\x8C\x00\x80\x07\x8C\x00\x80\x0B\x8C\x00\x80\x0F\x8C\x00\x80\x13\x8C"
-     "\x00\x80"
-     "\x17\x8C\x00\x80\xEF\x50\x02\x00\x1B\x8C\x00\x80\xE1\x14\x00\x00\x1F\x8C"
-     "\x00\x80"
-     "\xE3\x70\x1F\x00\xBE\xF8\x05\x00\x23\x8C\x00\x80\x27\x8C\x00\x80\x2B\x8C"
-     "\x00\x80"
-     "\xBB\xF9\x03\x00\xBA\xF1\x03\x00\xBE\x84\x07\x00\x2F\x8C\x00\x80\xBF\x5D"
-     "\x03\x00"
-     "\xBE\x5D\x03\x00\xBD\x5D\x03\x00\xBC\xE1\x03\x00\xB3\x09\x02\x00\x33\x8C"
-     "\x00\x80"
-     "\x37\x8C\x00\x80\x3B\x8C\x00\x80\x3F\x8C\x00\x80\xB6\xDD\x03\x00\xB5\xDD"
-     "\x03\x00"
-     "\x43\x8C\x00\x80\xA8\x6D\x1E\x00\xA9\xAD\x1E\x00\xAA\xA5\x1E\x00\xAB\xBD"
-     "\x1E\x00"
-     "\xAC\xA5\x1E\x00\xAD\xAD\x1E\x00\xAE\xA5\x1E\x00\xAF\x19\x1D\x00\x47\x8C"
-     "\x00\x80"
-     "\x4B\x8C\x00\x80\x4F\x8C\x00\x80\x53\x8C\x00\x80\x57\x8C\x00\x80\x5B\x8C"
-     "\x00\x80"
-     "\x5F\x8C\x00\x80\x63\x8C\x00\x80\xB8\x0D\x1D\x00\xB9\x11\x1D\x00\xBA\x11"
-     "\x1D\x00"
-     "\xBB\x25\x1D\x00\xBC\x3D\x1D\x00\xBD\xD5\x1D\x00\xBE\xDD\x1D\x00\xBF\xD5"
-     "\x1D\x00"
-     "\xB0\x69\x1D\x00\xB1\x69\x1D\x00\xB2\x79\x1D\x00\xB3\x71\x1D\x00\xB4\x59"
-     "\x1D\x00"
-     "\xB5\x59\x1D\x00\xB6\x39\x1D\x00\xB7\x35\x1D\x00\xE1\xF0\x1E\x00\xE1\x2C"
-     "\x1F\x00"
-     "\xE3\x50\x1E\x00\xE3\xD4\x1E\x00\x81\x09\x00\x00\x80\xA9\x00\x00\x67\x8C"
-     "\x00\x80"
-     "\x82\x39\x00\x00\x84\xF4\x05\x00\x6B\x8C\x00\x80\x87\x88\x04\x00\x86\x0C"
-     "\x04\x00"
-     "\x73\x8C\x00\x80\x77\x8C\x00\x80\xEF\xF8\x1E\x00\xEF\x18\x1E\x00\xA3\x29"
-     "\x03\x00"
-     "\x7B\x8C\x00\x80\x7F\x8C\x00\x80\x83\x8C\x00\x80\x87\x8C\x00\x80\xA6\xFD"
-     "\x02\x00"
-     "\xA5\xFD\x02\x00\x8B\x8C\x00\x80\xAB\xD9\x02\x00\xAA\xD1\x02\x00\x8F\x8C"
-     "\x00\x80"
-     "\x93\x8C\x00\x80\xAF\x7D\x02\x00\xAE\x7D\x02\x00\xAD\x7D\x02\x00\xAC\xC1"
-     "\x02\x00"
-     "\xA8\x85\x06\x00\xA9\x8D\x06\x00\xAA\x85\x06\x00\xAB\x9D\x06\x00\xAC\x85"
-     "\x06\x00"
-     "\xAD\x8D\x06\x00\xAE\xB9\x06\x00\xAF\xB5\x06\x00\x6F\x8C\x00\x80\x97\x8C"
-     "\x00\x80"
-     "\x9B\x8C\x00\x80\x9F\x8C\x00\x80\xA3\x8C\x00\x80\xA7\x8C\x00\x80\xAB\x8C"
-     "\x00\x80"
-     "\xAF\x8C\x00\x80\xB8\xBD\x06\x00\xB9\x65\x07\x00\xBA\x6D\x07\x00\xBB\x65"
-     "\x07\x00"
-     "\xBC\x7D\x07\x00\xBD\x65\x07\x00\xBE\x6D\x07\x00\xBF\x65\x07\x00\xB0\xD5"
-     "\x06\x00"
-     "\xB1\xDD\x06\x00\xB2\xD5\x06\x00\xB3\xA1\x06\x00\xB4\x95\x06\x00\xB5\x9D"
-     "\x06\x00"
-     "\xB6\x95\x06\x00\xB7\x8D\x06\x00\xB5\x15\x06\x00\xB3\x8C\x00\x80\xB7\x8C"
-     "\x00\x80"
-     "\xB6\x15\x06\x00\xBB\x8C\x00\x80\xBF\x8C\x00\x80\xB3\x15\x06\x00\xC3\x8C"
-     "\x00\x80"
-     "\xBD\x6D\x06\x00\xBC\x19\x06\x00\xBF\x41\x06\x00\xBE\x49\x06\x00\xFF\x8B"
-     "\x00\x80"
-     "\xC7\x8C\x00\x80\xBB\x01\x06\x00\xBA\x09\x06\x00\xA3\x4D\x06\x00\xCB\x8C"
-     "\x00\x80"
-     "\xCF\x8C\x00\x80\xD3\x8C\x00\x80\xD7\x8C\x00\x80\xA6\x4D\x06\x00\xA5\x4D"
-     "\x06\x00"
-     "\xDB\x8C\x00\x80\xAB\x59\x06\x00\xAA\x51\x06\x00\x84\xA4\x03\x00\xDF\x8C"
-     "\x00\x80"
-     "\xAF\x19\x06\x00\xAE\x11\x06\x00\xAD\x35\x06\x00\xAC\x41\x06\x00\xA8\xB1"
-     "\x06\x00"
-     "\xA9\xB1\x06\x00\xAA\xCD\x06\x00\xAB\xC5\x06\x00\xAC\xDD\x06\x00\xAD\xC5"
-     "\x06\x00"
-     "\xAE\xC5\x06\x00\xAF\xF5\x06\x00\x80\xED\x01\x00\x81\xF5\x01\x00\x82\xF5"
-     "\x01\x00"
-     "\xE3\x8C\x00\x80\x86\x90\x01\x00\x87\xB4\x01\x00\xE7\x8C\x00\x80\xEB\x8C"
-     "\x00\x80"
-     "\xB8\xC5\x01\x00\xB9\xCD\x01\x00\xBA\xD9\x01\x00\xBB\xD9\x01\x00\xBC\xCD"
-     "\x01\x00"
-     "\xBD\xE1\x01\x00\xBE\x9D\x01\x00\xBF\x95\x01\x00\xB0\x8D\x06\x00\xB1\x41"
-     "\x01\x00"
-     "\xB2\x41\x01\x00\xB3\x41\x01\x00\xB4\x41\x01\x00\xB5\x41\x01\x00\xB6\x41"
-     "\x01\x00"
-     "\xB7\x41\x01\x00\xB3\x45\x06\x00\xEF\x8C\x00\x80\xF3\x8C\x00\x80\xF7\x8C"
-     "\x00\x80"
-     "\xFB\x8C\x00\x80\xB6\x45\x06\x00\xB5\x55\x06\x00\xFF\x8C\x00\x80\xBB\x85"
-     "\x01\x00"
-     "\xBA\x49\x06\x00\x03\x8D\x00\x80\x07\x8D\x00\x80\xBF\x85\x01\x00\xBE\x85"
-     "\x01\x00"
-     "\xBD\x95\x01\x00\xBC\x95\x01\x00\x0B\x8D\x00\x80\xA3\x01\x06\x00\x0F\x8D"
-     "\x00\x80"
-     "\x13\x8D\x00\x80\xA6\x01\x06\x00\x17\x8D\x00\x80\x1B\x8D\x00\x80\xA5\x11"
-     "\x06\x00"
-     "\xAA\x0D\x06\x00\xAB\xC1\x01\x00\x1F\x8D\x00\x80\x23\x8D\x00\x80\xAE\xC1"
-     "\x01\x00"
-     "\xAF\xC1\x01\x00\xAC\xD1\x01\x00\xAD\xD1\x01\x00\xA8\x2D\x03\x00\xA9\x3D"
-     "\x03\x00"
-     "\xAA\x35\x03\x00\xAB\x89\x03\x00\xAC\x99\x03\x00\xAD\x99\x03\x00\xAE\x89"
-     "\x03\x00"
-     "\xAF\x81\x03\x00\x27\x8D\x00\x80\x2B\x8D\x00\x80\x2F\x8D\x00\x80\x33\x8D"
-     "\x00\x80"
-     "\x37\x8D\x00\x80\x3B\x8D\x00\x80\x3F\x8D\x00\x80\x43\x8D\x00\x80\xB8\xAD"
-     "\x03\x00"
-     "\xB9\x65\x00\x00\xBA\x6D\x00\x00\xBB\x65\x00\x00\xBC\x7D\x00\x00\xBD\x65"
-     "\x00\x00"
-     "\xBE\x6D\x00\x00\xBF\x65\x00\x00\xB0\xC9\x03\x00\xB1\xC9\x03\x00\xB2\xA9"
-     "\x03\x00"
-     "\xB3\xA5\x03\x00\xB4\xBD\x03\x00\xB5\xA1\x03\x00\xB6\xA1\x03\x00\xB7\x95"
-     "\x03\x00"
-     "\x80\xBD\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x47\x8D\x00\x80\x4B\x8D"
-     "\x00\x80"
-     "\x53\x8D\x00\x80\x57\x8D\x00\x80\x84\x90\x02\x00\x86\x00\x0C\x00\x87\xE4"
-     "\x03\x00"
-     "\x5B\x8D\x00\x80\x5F\x8D\x00\x80\x63\x8D\x00\x80\x67\x8D\x00\x80\x6B\x8D"
-     "\x00\x80"
-     "\x6F\x8D\x00\x80\xB3\xFD\x03\x00\x73\x8D\x00\x80\x77\x8D\x00\x80\x7B\x8D"
-     "\x00\x80"
-     "\x7F\x8D\x00\x80\xB6\x59\x03\x00\xB5\x51\x03\x00\x83\x8D\x00\x80\xBB\x4D"
-     "\x03\x00"
-     "\xBA\x4D\x03\x00\x87\x8D\x00\x80\x8B\x8D\x00\x80\xBF\x25\x03\x00\xBE\x39"
-     "\x03\x00"
-     "\xBD\x31\x03\x00\xBC\x3D\x03\x00\x8F\x8D\x00\x80\x93\x8D\x00\x80\x97\x8D"
-     "\x00\x80"
-     "\xEF\x74\x07\x00\x9B\x8D\x00\x80\x9F\x8D\x00\x80\xA3\x8D\x00\x80\xA7\x8D"
-     "\x00\x80"
-     "\x84\x6C\x0D\x00\xAB\x8D\x00\x80\xBE\x6C\x0D\x00\xAF\x8D\x00\x80\xB3\x8D"
-     "\x00\x80"
-     "\xE1\x10\x07\x00\xB7\x8D\x00\x80\xE3\x8C\x06\x00\x80\x71\x00\x00\x81\x7D"
-     "\x00\x00"
-     "\x82\x1D\x00\x00\x84\x00\x0C\x00\xEF\x60\x07\x00\xBB\x8D\x00\x80\x87\x74"
-     "\x0D\x00"
-     "\x86\xFC\x0C\x00\xE3\xEC\x03\x00\xC3\x8D\x00\x80\xE1\x40\x00\x00\xC7\x8D"
-     "\x00\x80"
-     "\xCB\x8D\x00\x80\xCF\x8D\x00\x80\xD3\x8D\x00\x80\xD7\x8D\x00\x80\xDB\x8D"
-     "\x00\x80"
-     "\xDF\x8D\x00\x80\xE3\x8D\x00\x80\xE7\x8D\x00\x80\xEB\x8D\x00\x80\xEF\x8D"
-     "\x00\x80"
-     "\xF3\x8D\x00\x80\xEF\xD4\x07\x00\xBE\xE8\x0C\x00\xE1\x10\x06\x00\xF7\x8D"
-     "\x00\x80"
-     "\xE3\x54\x01\x00\xFB\x8D\x00\x80\xFF\x8D\x00\x80\x03\x8E\x00\x80\x07\x8E"
-     "\x00\x80"
-     "\xA3\xFD\x02\x00\x0B\x8E\x00\x80\x0F\x8E\x00\x80\x13\x8E\x00\x80\x17\x8E"
-     "\x00\x80"
-     "\xA6\x59\x02\x00\xA5\x51\x02\x00\x1B\x8E\x00\x80\xAB\x4D\x02\x00\xAA\x4D"
-     "\x02\x00"
-     "\x1F\x8E\x00\x80\x23\x8E\x00\x80\xAF\x25\x02\x00\xAE\x39\x02\x00\xAD\x31"
-     "\x02\x00"
-     "\xAC\x3D\x02\x00\xA8\x95\x0E\x00\xA9\x9D\x0E\x00\xAA\xA1\x0E\x00\xAB\xA1"
-     "\x0E\x00"
-     "\xAC\xA1\x0E\x00\xAD\xA1\x0E\x00\xAE\xA1\x0E\x00\xAF\xA1\x0E\x00\xBF\x8D"
-     "\x00\x80"
-     "\x82\x11\x00\x00\x81\x19\x00\x00\x80\xC1\x0F\x00\x27\x8E\x00\x80\x2B\x8E"
-     "\x00\x80"
-     "\x2F\x8E\x00\x80\x33\x8E\x00\x80\xB8\x45\x0F\x00\xB9\x4D\x0F\x00\xBA\x59"
-     "\x0F\x00"
-     "\xBB\x59\x0F\x00\xBC\x49\x0F\x00\xBD\x49\x0F\x00\xBE\x7D\x0F\x00\xBF\x71"
-     "\x0F\x00"
-     "\xB0\xA5\x0E\x00\xB1\xAD\x0E\x00\xB2\xA5\x0E\x00\xB3\xBD\x0E\x00\xB4\xAD"
-     "\x0E\x00"
-     "\xB5\x9D\x0E\x00\xB6\x95\x0E\x00\xB7\x7D\x0F\x00\xB3\x29\x0E\x00\x37\x8E"
-     "\x00\x80"
-     "\x86\x28\x00\x00\x87\x44\x03\x00\x3B\x8E\x00\x80\xB6\x21\x0E\x00\xB5\x29"
-     "\x0E\x00"
-     "\x3F\x8E\x00\x80\xBB\x69\x0E\x00\xBA\x61\x0E\x00\x43\x8E\x00\x80\x47\x8E"
-     "\x00\x80"
-     "\xBF\x59\x0E\x00\xBE\x69\x0E\x00\xBD\x69\x0E\x00\xBC\x71\x0E\x00\x4B\x8E"
-     "\x00\x80"
-     "\x4F\x8D\x00\x80\xA3\x61\x0E\x00\x4F\x8E\x00\x80\xA5\x61\x0E\x00\x53\x8E"
-     "\x00\x80"
-     "\x57\x8E\x00\x80\xA6\x69\x0E\x00\x5B\x8E\x00\x80\x5F\x8E\x00\x80\xAB\x21"
-     "\x0E\x00"
-     "\xAA\x29\x0E\x00\xAD\x21\x0E\x00\xAC\x39\x0E\x00\xAF\x11\x0E\x00\xAE\x21"
-     "\x0E\x00"
-     "\xA8\x65\x0E\x00\xA9\x69\x0E\x00\xAA\x79\x0E\x00\xAB\x79\x0E\x00\xAC\x69"
-     "\x0E\x00"
-     "\xAD\x69\x0E\x00\xAE\x99\x0E\x00\xAF\x99\x0E\x00\x63\x8E\x00\x80\x67\x8E"
-     "\x00\x80"
-     "\x6B\x8E\x00\x80\x6F\x8E\x00\x80\x73\x8E\x00\x80\x77\x8E\x00\x80\x7B\x8E"
-     "\x00\x80"
-     "\x7F\x8E\x00\x80\xB8\xF9\x0E\x00\xB9\xF9\x0E\x00\xBA\x8D\x0E\x00\xBB\x85"
-     "\x0E\x00"
-     "\xBC\x9D\x0E\x00\xBD\x85\x0E\x00\xBE\x85\x0E\x00\xBF\xB5\x0E\x00\xB0\xE9"
-     "\x0E\x00"
-     "\xB1\xE9\x0E\x00\xB2\xF9\x0E\x00\xB3\xF9\x0E\x00\xB4\xE9\x0E\x00\xB5\xDD"
-     "\x0E\x00"
-     "\xB6\xC9\x0E\x00\xB7\xC9\x0E\x00\xB3\xF1\x0E\x00\x83\x8E\x00\x80\x87\x8E"
-     "\x00\x80"
-     "\x8B\x8E\x00\x80\x8F\x8E\x00\x80\xB6\xE9\x0E\x00\xB5\xE1\x0E\x00\x93\x8E"
-     "\x00\x80"
-     "\xBB\x2D\x0E\x00\xBA\x2D\x0E\x00\x97\x8E\x00\x80\x9B\x8E\x00\x80\xBF\x19"
-     "\x0E\x00"
-     "\xBE\x2D\x0E\x00\xBD\x2D\x0E\x00\xBC\x2D\x0E\x00\x82\x6D\x00\x00\xA3\xB5"
-     "\x0E\x00"
-     "\x80\x55\x00\x00\x81\x65\x00\x00\xA6\xAD\x0E\x00\x84\x98\x01\x00\xBE\x90"
-     "\x01\x00"
-     "\xA5\xA5\x0E\x00\xAA\x69\x0E\x00\xAB\x69\x0E\x00\x86\x48\x07\x00\x87\xFC"
-     "\x00\x00"
-     "\xAE\x69\x0E\x00\xAF\x5D\x0E\x00\xAC\x69\x0E\x00\xAD\x69\x0E\x00\xA8\x3D"
-     "\x02\x00"
-     "\xA9\x81\x02\x00\xAA\x99\x02\x00\xAB\x91\x02\x00\xAC\xB9\x02\x00\xAD\xB9"
-     "\x02\x00"
-     "\xAE\xD9\x02\x00\xAF\xD1\x02\x00\xA3\x8E\x00\x80\xA7\x8E\x00\x80\xAB\x8E"
-     "\x00\x80"
-     "\xBF\x00\x38\x00\xAF\x8E\x00\x80\xB3\x8E\x00\x80\xB7\x8E\x00\x80\xBB\x8E"
-     "\x00\x80"
-     "\xB8\x71\x03\x00\xB9\x71\x03\x00\xBA\x71\x03\x00\xBB\x71\x03\x00\xBC\xD5"
-     "\x03\x00"
-     "\xBD\xDD\x03\x00\xBE\xD5\x03\x00\xBF\xCD\x03\x00\xB0\xB5\x02\x00\xB1\xBD"
-     "\x02\x00"
-     "\xB2\x89\x02\x00\xB3\x89\x02\x00\xB4\x51\x03\x00\xB5\x51\x03\x00\xB6\x51"
-     "\x03\x00"
-     "\xB7\x51\x03\x00\xB3\x11\x02\x00\xBF\x8E\x00\x80\xC3\x8E\x00\x80\xC7\x8E"
-     "\x00\x80"
-     "\x84\x9C\x03\x00\xB6\x39\x02\x00\xB5\x31\x02\x00\xCB\x8E\x00\x80\xBB\x79"
-     "\x02\x00"
-     "\xBA\x71\x02\x00\xCF\x8E\x00\x80\xD3\x8E\x00\x80\xBF\x95\x03\x00\xBE\x95"
-     "\x03\x00"
-     "\xBD\x5D\x02\x00\xBC\x5D\x02\x00\xD7\x8E\x00\x80\xA3\x55\x02\x00\xDB\x8E"
-     "\x00\x80"
-     "\xDF\x8E\x00\x80\xA6\x7D\x02\x00\xE3\x8E\x00\x80\xE7\x8E\x00\x80\xA5\x75"
-     "\x02\x00"
-     "\xAA\x35\x02\x00\xAB\x3D\x02\x00\xEB\x8E\x00\x80\xEF\x8E\x00\x80\xAE\xD1"
-     "\x03\x00"
-     "\xAF\xD1\x03\x00\xAC\x19\x02\x00\xAD\x19\x02\x00\xF3\x8E\x00\x80\xF7\x8E"
-     "\x00\x80"
-     "\xFB\x8E\x00\x80\xFF\x8E\x00\x80\x80\x0D\x00\x00\x81\x31\x00\x00\x82\x31"
-     "\x00\x00"
-     "\x03\x8F\x00\x80\xBE\x2C\x05\x00\x07\x8F\x00\x80\x0B\x8F\x00\x80\x13\x8F"
-     "\x00\x80"
-     "\x17\x8F\x00\x80\x1B\x8F\x00\x80\x1F\x8F\x00\x80\x23\x8F\x00\x80\xE1\xF8"
-     "\x0F\x00"
-     "\x27\x8F\x00\x80\xE3\x78\x0E\x00\x84\x44\x05\x00\x86\xF0\x04\x00\x87\x98"
-     "\x02\x00"
-     "\x2B\x8F\x00\x80\x2F\x8F\x00\x80\x33\x8F\x00\x80\x37\x8F\x00\x80\x3B\x8F"
-     "\x00\x80"
-     "\x3F\x8F\x00\x80\x43\x8F\x00\x80\x47\x8F\x00\x80\xEF\xA8\x0E\x00\x4B\x8F"
-     "\x00\x80"
-     "\x4F\x8F\x00\x80\xEF\xA4\x00\x00\x53\x8F\x00\x80\x57\x8F\x00\x80\x5B\x8F"
-     "\x00\x80"
-     "\x5F\x8F\x00\x80\x63\x8F\x00\x80\xEF\xC4\x0E\x00\xBE\xAC\x04\x00\xE1\xE0"
-     "\x0E\x00"
-     "\x67\x8F\x00\x80\xE3\xF0\x0E\x00\x6B\x8F\x00\x80\xE3\x70\x00\x00\x6F\x8F"
-     "\x00\x80"
-     "\xE1\x2C\x01\x00\xB3\xE9\x01\x00\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\x73\x8F\x00\x80\xB6\xED\x01\x00\xB5\xED\x01\x00\x77\x8F\x00\x80\xBB\x6D"
-     "\x01\x00"
-     "\xBA\x6D\x01\x00\x7B\x8F\x00\x80\x83\x8F\x00\x80\xBF\x31\x01\x00\xBE\x59"
-     "\x01\x00"
-     "\xBD\x51\x01\x00\xBC\x61\x01\x00\xA8\x21\x06\x00\xA9\x21\x06\x00\xAA\x21"
-     "\x06\x00"
-     "\xAB\x3D\x06\x00\xAC\x25\x06\x00\xAD\x2D\x06\x00\xAE\x25\x06\x00\xAF\x99"
-     "\x06\x00"
-     "\x0F\x8F\x00\x80\x7F\x8F\x00\x80\x86\x88\x00\x00\x87\xA0\x03\x00\x87\x8F"
-     "\x00\x80"
-     "\x8B\x8F\x00\x80\x8F\x8F\x00\x80\x93\x8F\x00\x80\xB8\x95\x06\x00\xB9\x9D"
-     "\x06\x00"
-     "\xBA\xA1\x06\x00\xBB\xA1\x06\x00\xBC\x51\x07\x00\xBD\x51\x07\x00\xBE\x51"
-     "\x07\x00"
-     "\xBF\x51\x07\x00\xB0\xE9\x06\x00\xB1\xE9\x06\x00\xB2\xF5\x06\x00\xB3\xC5"
-     "\x06\x00"
-     "\xB4\xDD\x06\x00\xB5\xB5\x06\x00\xB6\xB1\x06\x00\xB7\xB1\x06\x00\xA3\xA9"
-     "\x06\x00"
-     "\x97\x8F\x00\x80\x9B\x8F\x00\x80\x9F\x8F\x00\x80\xA3\x8F\x00\x80\xA6\xAD"
-     "\x06\x00"
-     "\xA5\xAD\x06\x00\xA7\x8F\x00\x80\xAB\x2D\x06\x00\xAA\x2D\x06\x00\xAB\x8F"
-     "\x00\x80"
-     "\xAF\x8F\x00\x80\xAF\x71\x06\x00\xAE\x19\x06\x00\xAD\x11\x06\x00\xAC\x21"
-     "\x06\x00"
-     "\xB3\x8F\x00\x80\xB3\xF1\x07\x00\x9F\x8E\x00\x80\xB7\x8F\x00\x80\xB6\x95"
-     "\x07\x00"
-     "\xBB\x8F\x00\x80\xBF\x8F\x00\x80\xB5\xE1\x07\x00\xBA\x81\x07\x00\xBB\x81"
-     "\x07\x00"
-     "\xC3\x8F\x00\x80\xC7\x8F\x00\x80\xBE\x81\x07\x00\xBF\x81\x07\x00\xBC\x91"
-     "\x07\x00"
-     "\xBD\x91\x07\x00\xA8\x2D\x06\x00\xA9\x35\x06\x00\xAA\x3D\x06\x00\xAB\x35"
-     "\x06\x00"
-     "\xAC\x2D\x06\x00\xAD\x51\x06\x00\xAE\x41\x06\x00\xAF\x41\x06\x00\xCB\x8F"
-     "\x00\x80"
-     "\xCF\x8F\x00\x80\xD3\x8F\x00\x80\xD7\x8F\x00\x80\xDB\x8F\x00\x80\xDF\x8F"
-     "\x00\x80"
-     "\xE3\x8F\x00\x80\xE7\x8F\x00\x80\xB8\xC9\x06\x00\xB9\xC9\x06\x00\xBA\xD9"
-     "\x06\x00"
-     "\xBB\xD9\x06\x00\xBC\xC9\x06\x00\xBD\xFD\x06\x00\xBE\xE9\x06\x00\xBF\xE5"
-     "\x06\x00"
-     "\xB0\x01\x06\x00\xB1\x01\x06\x00\xB2\x01\x06\x00\xB3\x01\x06\x00\xB4\x01"
-     "\x06\x00"
-     "\xB5\x01\x06\x00\xB6\x05\x06\x00\xB7\xFD\x06\x00\xA3\xB5\x06\x00\x82\x15"
-     "\x00\x00"
-     "\x81\xB5\x00\x00\x80\xB5\x00\x00\xEB\x8F\x00\x80\xA6\xD1\x06\x00\xA5\xA5"
-     "\x06\x00"
-     "\xBE\x90\x01\x00\xAB\xC5\x06\x00\xAA\xC5\x06\x00\x84\xA0\x02\x00\xF3\x8F"
-     "\x00\x80"
-     "\xAF\xC5\x06\x00\xAE\xC5\x06\x00\xAD\xD5\x06\x00\xAC\xD5\x06\x00\x85\x78"
-     "\x01\x00"
-     "\xB3\xF1\x01\x00\x85\x8C\x00\x00\xF7\x8F\x00\x80\xB6\x55\x01\x00\xFB\x8F"
-     "\x00\x80"
-     "\xFF\x8F\x00\x80\xB5\xE1\x01\x00\xBA\x71\x01\x00\xBB\x79\x01\x00\x86\xC8"
-     "\x3F\x00"
-     "\x87\xBC\x00\x00\xBE\x31\x01\x00\xBF\x39\x01\x00\xBC\x51\x01\x00\xBD\x51"
-     "\x01\x00"
-     "\xA8\xED\x05\x00\xA9\x35\x02\x00\xAA\x3D\x02\x00\xAB\x35\x02\x00\xAC\x2D"
-     "\x02\x00"
-     "\xAD\x91\x02\x00\xAE\x91\x02\x00\xAF\x91\x02\x00\x03\x90\x00\x80\x07\x90"
-     "\x00\x80"
-     "\x0B\x90\x00\x80\x0F\x90\x00\x80\xA2\x11\x02\x00\xA3\x11\x02\x00\xA0\x01"
-     "\x06\x00"
-     "\xA1\xF5\x02\x00\xB8\xA9\x02\x00\xB9\xA9\x02\x00\xBA\x69\x03\x00\xBB\x69"
-     "\x03\x00"
-     "\xBC\x79\x03\x00\xBD\x79\x03\x00\xBE\x69\x03\x00\xBF\x69\x03\x00\xB0\xF1"
-     "\x02\x00"
-     "\xB1\xF9\x02\x00\xB2\xC1\x02\x00\xB3\xC1\x02\x00\xB4\xB1\x02\x00\xB5\xB9"
-     "\x02\x00"
-     "\xB6\xAD\x02\x00\xB7\xA5\x02\x00\xA3\xB5\x02\x00\x8F\x8D\x1A\x00\x13\x90"
-     "\x00\x80"
-     "\x17\x90\x00\x80\x1B\x90\x00\x80\xA6\x11\x02\x00\xA5\xA5\x02\x00\x1F\x90"
-     "\x00\x80"
-     "\xAB\x3D\x02\x00\xAA\x35\x02\x00\x23\x90\x00\x80\x27\x90\x00\x80\xAF\x7D"
-     "\x02\x00"
-     "\xAE\x75\x02\x00\xAD\x15\x02\x00\xAC\x15\x02\x00\x9E\x05\x06\x00\x9F\xE5"
-     "\x06\x00"
-     "\x9C\xE9\x0A\x00\x9D\x21\x0A\x00\x9A\xF1\x0F\x00\x9B\x65\x0A\x00\x98\xD9"
-     "\x0E\x00"
-     "\x99\xD9\x0E\x00\x96\xC5\x12\x00\x97\x8D\x13\x00\x94\x75\x12\x00\x95\xCD"
-     "\x12\x00"
-     "\x92\xD1\x16\x00\x93\x71\x16\x00\x90\xF5\x1A\x00\x91\x75\x16\x00\x2B\x90"
-     "\x00\x80"
-     "\xE1\xD8\x00\x00\x84\xC4\x3C\x00\xE3\xC8\x32\x00\x2F\x90\x00\x80\x33\x90"
-     "\x00\x80"
-     "\x37\x90\x00\x80\x3B\x90\x00\x80\x3F\x90\x00\x80\x43\x90\x00\x80\x47\x90"
-     "\x00\x80"
-     "\x4B\x90\x00\x80\x4F\x90\x00\x80\x53\x90\x00\x80\x57\x90\x00\x80\xEF\x68"
-     "\x37\x00"
-     "\x5B\x90\x00\x80\xE3\xC4\x07\x00\xBE\x80\x3C\x00\xE1\xBC\x04\x00\x80\x65"
-     "\x00\x00"
-     "\x81\x6D\x00\x00\x82\x19\x00\x00\x63\x90\x00\x80\x86\x40\x3C\x00\x87\xD8"
-     "\x3D\x00"
-     "\x67\x90\x00\x80\x6B\x90\x00\x80\x6F\x90\x00\x80\xEF\xAC\x06\x00\x73\x90"
-     "\x00\x80"
-     "\x77\x90\x00\x80\xB1\x85\x2B\x00\xB0\x01\x28\x00\xB3\xA1\x2A\x00\xB2\x9D"
-     "\x2B\x00"
-     "\xB5\x91\x27\x00\xB4\x01\x24\x00\xEF\xFC\x00\x00\xEF\xC4\x06\x00\xE1\x90"
-     "\x01\x00"
-     "\xE1\xEC\x06\x00\xE3\x10\x00\x00\xE3\xF0\x06\x00\x7B\x90\x00\x80\x7F\x90"
-     "\x00\x80"
-     "\x84\x74\x3F\x00\x83\x90\x00\x80\xA1\xC9\x3F\x00\x87\x90\x00\x80\xA3\xD5"
-     "\x3E\x00"
-     "\xA2\xC1\x3F\x00\xA5\xD5\x3B\x00\xA4\x01\x38\x00\xA7\x01\x34\x00\xA6\x61"
-     "\x3B\x00"
-     "\xA9\xB9\x37\x00\xA8\xE1\x37\x00\xAB\xED\x33\x00\xAA\x01\x30\x00\xAD\x01"
-     "\x2C\x00"
-     "\xAC\xDD\x33\x00\xAF\x91\x2F\x00\xAE\xF9\x2F\x00\x8B\x25\x2C\x00\x8A\x25"
-     "\x2C\x00"
-     "\xEF\x8F\x00\x80\x5F\x90\x00\x80\x8F\x31\x31\x00\x8E\x11\x30\x00\x8D\x19"
-     "\x30\x00"
-     "\x8C\x75\x2D\x00\x83\x2D\x24\x00\x82\x2D\x24\x00\x8B\x90\x00\x80\x8F\x90"
-     "\x00\x80"
-     "\x87\xAD\x2C\x00\x86\xB1\x28\x00\x85\x19\x28\x00\x84\xB5\x28\x00\x9B\x61"
-     "\x3C\x00"
-     "\x9A\x8D\x3C\x00\x93\x90\x00\x80\x97\x90\x00\x80\x9B\x90\x00\x80\x9F\x90"
-     "\x00\x80"
-     "\xA3\x90\x00\x80\x9C\x79\x3C\x00\x93\x41\x34\x00\x92\xD5\x34\x00\xA7\x90"
-     "\x00\x80"
-     "\xAB\x90\x00\x80\x97\x49\x38\x00\x96\x55\x38\x00\x95\xBD\x38\x00\x94\x91"
-     "\x34\x00"
-     "\xA8\x2D\x3D\x00\xA9\x49\x3D\x00\xAA\x5D\x3D\x00\xAB\x55\x3D\x00\xAC\xB1"
-     "\x3E\x00"
-     "\xAD\xB1\x3E\x00\xAE\xB1\x3E\x00\xAF\xB1\x3E\x00\xAF\x90\x00\x80\xB3\x90"
-     "\x00\x80"
-     "\xB7\x90\x00\x80\xBB\x90\x00\x80\xBF\x90\x00\x80\x82\xB9\x01\x00\x81\xB1"
-     "\x01\x00"
-     "\x80\xB9\x01\x00\xB8\xD5\x3F\x00\xB9\xDD\x3F\x00\xBA\xD5\x3F\x00\xBB\xED"
-     "\x3F\x00"
-     "\xBC\xF5\x3F\x00\xBD\xFD\x3F\x00\xBE\xF5\x3F\x00\xBF\xED\x3F\x00\xB0\x69"
-     "\x3E\x00"
-     "\xB1\x69\x3E\x00\xB2\x79\x3E\x00\xB3\x71\x3E\x00\xB4\x59\x3E\x00\xB5\x59"
-     "\x3E\x00"
-     "\xB6\xF9\x3F\x00\xB7\xED\x3F\x00\xB3\xD5\x3C\x00\xC3\x90\x00\x80\xC7\x90"
-     "\x00\x80"
-     "\xCB\x90\x00\x80\xBE\xD8\x01\x00\xB6\x31\x3F\x00\xB5\x39\x3F\x00\xCF\x90"
-     "\x00\x80"
-     "\xBB\xAD\x3F\x00\xBA\xAD\x3F\x00\x86\x88\x00\x00\x87\xAC\x01\x00\xBF\x9D"
-     "\x3F\x00"
-     "\xBE\x9D\x3F\x00\xBD\xA5\x3F\x00\xBC\xAD\x3F\x00\xD3\x90\x00\x80\xA3\x91"
-     "\x3C\x00"
-     "\xD7\x90\x00\x80\xDB\x90\x00\x80\xA6\x75\x3F\x00\xDF\x90\x00\x80\xE3\x90"
-     "\x00\x80"
-     "\xA5\x7D\x3F\x00\xAA\xE9\x3F\x00\xAB\xE9\x3F\x00\xE7\x90\x00\x80\xEB\x90"
-     "\x00\x80"
-     "\xAE\xD9\x3F\x00\xAF\xD9\x3F\x00\xAC\xE9\x3F\x00\xAD\xE1\x3F\x00\xA8\xDD"
-     "\x3E\x00"
-     "\xA9\xED\x3E\x00\xAA\xE5\x3E\x00\xAB\xFD\x3E\x00\xAC\xE9\x3E\x00\xAD\xE9"
-     "\x3E\x00"
-     "\xAE\x29\x3E\x00\xAF\x29\x3E\x00\x84\xA4\x01\x00\xEF\x90\x00\x80\xF3\x90"
-     "\x00\x80"
-     "\xF7\x90\x00\x80\xFB\x90\x00\x80\xFF\x90\x00\x80\x03\x91\x00\x80\x07\x91"
-     "\x00\x80"
-     "\xB8\x99\x01\x00\xB9\xAD\x01\x00\xBA\xA5\x01\x00\xBB\x6D\x01\x00\xBC\x75"
-     "\x01\x00"
-     "\xBD\x7D\x01\x00\xBE\x75\x01\x00\xBF\x6D\x01\x00\xB0\x5D\x3E\x00\xB1\x21"
-     "\x3E\x00"
-     "\xB2\xC1\x01\x00\xB3\xC1\x01\x00\xB4\xB9\x01\x00\xB5\xB9\x01\x00\xB6\xA9"
-     "\x01\x00"
-     "\xB7\xA9\x01\x00\xB3\x6D\x3D\x00\x0B\x91\x00\x80\x0F\x91\x00\x80\x13\x91"
-     "\x00\x80"
-     "\x17\x91\x00\x80\xB6\x69\x3D\x00\xB5\x61\x3D\x00\x1B\x91\x00\x80\xBB\xE5"
-     "\x3D\x00"
-     "\xBA\x99\x3D\x00\x1F\x91\x00\x80\xBE\x68\x03\x00\xBF\x7D\x02\x00\xBE\x7D"
-     "\x02\x00"
-     "\xBD\x15\x02\x00\xBC\x15\x02\x00\x23\x91\x00\x80\xA3\x29\x3D\x00\x27\x91"
-     "\x00\x80"
-     "\x2B\x91\x00\x80\xA6\x2D\x3D\x00\x2F\x91\x00\x80\x33\x91\x00\x80\xA5\x25"
-     "\x3D\x00"
-     "\xAA\xDD\x3D\x00\xAB\xA1\x3D\x00\x37\x91\x00\x80\x3B\x91\x00\x80\xAE\x39"
-     "\x02\x00"
-     "\xAF\x39\x02\x00\xAC\x51\x02\x00\xAD\x51\x02\x00\x80\xD9\x00\x00\x81\xED"
-     "\x00\x00"
-     "\x82\xE5\x00\x00\x43\x91\x00\x80\xE3\x20\x01\x00\x47\x91\x00\x80\xE1\xA8"
-     "\x01\x00"
-     "\x4B\x91\x00\x80\xEF\x74\x02\x00\x4F\x91\x00\x80\x87\x6C\x03\x00\x86\x8C"
-     "\x04\x00"
-     "\x53\x91\x00\x80\x57\x91\x00\x80\x5B\x91\x00\x80\xBE\x30\x05\x00\x5F\x91"
-     "\x00\x80"
-     "\x63\x91\x00\x80\xB3\x85\x03\x00\x67\x91\x00\x80\xB5\x89\x03\x00\x6B\x91"
-     "\x00\x80"
-     "\x6F\x91\x00\x80\xB6\xB5\x03\x00\x73\x91\x00\x80\x77\x91\x00\x80\xBB\x45"
-     "\x03\x00"
-     "\xBA\x45\x03\x00\xBD\x45\x03\x00\xBC\x45\x03\x00\xBF\x45\x03\x00\xBE\x45"
-     "\x03\x00"
-     "\xA8\x41\x02\x00\xA9\x4D\x02\x00\xAA\x45\x02\x00\xAB\x59\x02\x00\xAC\x49"
-     "\x02\x00"
-     "\xAD\x7D\x02\x00\xAE\x75\x02\x00\xAF\xB9\x02\x00\x84\xEC\x05\x00\x7B\x91"
-     "\x00\x80"
-     "\x7F\x91\x00\x80\x83\x91\x00\x80\x84\xB8\x04\x00\x87\x91\x00\x80\x8B\x91"
-     "\x00\x80"
-     "\x8F\x91\x00\x80\xB8\x4D\x01\x00\xB9\x55\x01\x00\xBA\x55\x01\x00\xBB\x65"
-     "\x01\x00"
-     "\xBC\x11\x01\x00\xBD\x1D\x01\x00\xBE\x15\x01\x00\xBF\x09\x01\x00\xB0\xC9"
-     "\x02\x00"
-     "\xB1\xC9\x02\x00\xB2\xD9\x02\x00\xB3\xD9\x02\x00\xB4\xC9\x02\x00\xB5\xC9"
-     "\x02\x00"
-     "\xB6\x7D\x01\x00\xB7\x75\x01\x00\xE3\x28\x3E\x00\xE3\x9C\x3E\x00\xE1\xF0"
-     "\x3E\x00"
-     "\xE1\x7C\x3C\x00\x93\x91\x00\x80\x97\x91\x00\x80\x9B\x91\x00\x80\x9F\x91"
-     "\x00\x80"
-     "\xA3\x91\x00\x80\xA7\x91\x00\x80\xAB\x91\x00\x80\xB3\x91\x00\x80\xEF\x58"
-     "\x00\x00"
-     "\xEF\x04\x3E\x00\xB7\x91\x00\x80\xBB\x91\x00\x80\x81\x65\x00\x00\x80\x65"
-     "\x00\x00"
-     "\xA3\x85\x02\x00\x82\x65\x00\x00\xA5\x89\x02\x00\xBF\x91\x00\x80\xC3\x91"
-     "\x00\x80"
-     "\xA6\xB5\x02\x00\x86\xC0\x04\x00\x87\x44\x05\x00\xAB\x45\x02\x00\xAA\x45"
-     "\x02\x00"
-     "\xAD\x45\x02\x00\xAC\x45\x02\x00\xAF\x45\x02\x00\xAE\x45\x02\x00\xB3\x9D"
-     "\x06\x00"
-     "\xAF\x91\x00\x80\xC7\x91\x00\x80\xCB\x91\x00\x80\xCF\x91\x00\x80\xB6\xA1"
-     "\x06\x00"
-     "\xB5\xA9\x06\x00\xD3\x91\x00\x80\xBB\xED\x06\x00\xBA\xE1\x06\x00\xD7\x91"
-     "\x00\x80"
-     "\xDB\x91\x00\x80\xBF\xD5\x06\x00\xBE\xD5\x06\x00\xBD\xE1\x06\x00\xBC\xE9"
-     "\x06\x00"
-     "\xDF\x91\x00\x80\xE3\x91\x00\x80\xE7\x91\x00\x80\xEB\x91\x00\x80\xE1\xE4"
-     "\x04\x00"
-     "\xEF\x91\x00\x80\xE3\x68\x05\x00\xF3\x91\x00\x80\xF7\x91\x00\x80\xFB\x91"
-     "\x00\x80"
-     "\xEF\x2C\x04\x00\xFF\x91\x00\x80\x03\x92\x00\x80\x07\x92\x00\x80\x0B\x92"
-     "\x00\x80"
-     "\x0F\x92\x00\x80\xA8\x0D\x06\x00\xA9\x11\x06\x00\xAA\x15\x06\x00\xAB\x29"
-     "\x06\x00"
-     "\xAC\x21\x06\x00\xAD\x21\x06\x00\xAE\x21\x06\x00\xAF\x5D\x06\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x1D\x00\x00\x13\x92\x00\x80\x17\x92\x00\x80\x1B\x92"
-     "\x00\x80"
-     "\xBE\x34\x00\x00\x1F\x92\x00\x80\xB8\xED\x07\x00\xB9\xF1\x07\x00\xBA\xF1"
-     "\x07\x00"
-     "\xBB\xF1\x07\x00\xBC\x41\x07\x00\xBD\x41\x07\x00\xBE\x41\x07\x00\xBF\x41"
-     "\x07\x00"
-     "\xB0\x25\x06\x00\xB1\x29\x06\x00\xB2\x3D\x06\x00\xB3\x35\x06\x00\xB4\x11"
-     "\x06\x00"
-     "\xB5\x11\x06\x00\xB6\xDD\x07\x00\xB7\xD5\x07\x00\xA3\x9D\x07\x00\x23\x92"
-     "\x00\x80"
-     "\x86\x28\x01\x00\x87\x5C\x01\x00\x27\x92\x00\x80\xA6\xA1\x07\x00\xA5\xA9"
-     "\x07\x00"
-     "\x2B\x92\x00\x80\xAB\xED\x07\x00\xAA\xE1\x07\x00\x2F\x92\x00\x80\x33\x92"
-     "\x00\x80"
-     "\xAF\xD5\x07\x00\xAE\xD5\x07\x00\xAD\xE1\x07\x00\xAC\xE9\x07\x00\x37\x92"
-     "\x00\x80"
-     "\xB3\xF9\x06\x00\x3B\x92\x00\x80\x3F\x92\x00\x80\xB6\x59\x01\x00\x43\x92"
-     "\x00\x80"
-     "\x47\x92\x00\x80\xB5\x85\x06\x00\xBA\x7D\x01\x00\xBB\x4D\x01\x00\x4B\x92"
-     "\x00\x80"
-     "\x4F\x92\x00\x80\xBE\x35\x01\x00\xBF\x3D\x01\x00\xBC\x55\x01\x00\xBD\x3D"
-     "\x01\x00"
-     "\xA8\x2D\x06\x00\xA9\x3D\x06\x00\xAA\x35\x06\x00\xAB\x59\x06\x00\xAC\x49"
-     "\x06\x00"
-     "\xAD\x9D\x01\x00\xAE\x95\x01\x00\xAF\x8D\x01\x00\x53\x92\x00\x80\x57\x92"
-     "\x00\x80"
-     "\x5B\x92\x00\x80\x5F\x92\x00\x80\x63\x92\x00\x80\x67\x92\x00\x80\x6B\x92"
-     "\x00\x80"
-     "\x6F\x92\x00\x80\xB8\x9D\x01\x00\xB9\xA5\x01\x00\xBA\xAD\x01\x00\xBB\xA1"
-     "\x01\x00"
-     "\xBC\xA1\x01\x00\xBD\xA1\x01\x00\xBE\xA1\x01\x00\xBF\xA1\x01\x00\xB0\xFD"
-     "\x01\x00"
-     "\xB1\xCD\x01\x00\xB2\xC5\x01\x00\xB3\xAD\x01\x00\xB4\xB5\x01\x00\xB5\xBD"
-     "\x01\x00"
-     "\xB6\xB5\x01\x00\xB7\xA9\x01\x00\xA3\xBD\x05\x00\x73\x92\x00\x80\x77\x92"
-     "\x00\x80"
-     "\x7B\x92\x00\x80\x7F\x92\x00\x80\xA6\x1D\x02\x00\xA5\xC1\x05\x00\x83\x92"
-     "\x00\x80"
-     "\xAB\x09\x02\x00\xAA\x39\x02\x00\x87\x92\x00\x80\x8B\x92\x00\x80\xAF\x79"
-     "\x02\x00"
-     "\xAE\x71\x02\x00\xAD\x79\x02\x00\xAC\x11\x02\x00\xBE\x04\x03\x00\x93\x92"
-     "\x00\x80"
-     "\x84\xC4\x03\x00\x97\x92\x00\x80\x80\x4D\x00\x00\x81\x7D\x00\x00\x82\x75"
-     "\x00\x00"
-     "\x9B\x92\x00\x80\x9F\x92\x00\x80\xA3\x92\x00\x80\xA7\x92\x00\x80\xAB\x92"
-     "\x00\x80"
-     "\x86\x50\x0C\x00\x87\xE4\x03\x00\xAF\x92\x00\x80\xB3\x92\x00\x80\xA8\x85"
-     "\x03\x00"
-     "\xA9\x95\x03\x00\xAA\x95\x03\x00\xAB\xA5\x03\x00\xAC\xBD\x03\x00\xAD\xD5"
-     "\x03\x00"
-     "\xAE\xD1\x03\x00\xAF\xD1\x03\x00\xA8\x00\x00\x00\xB7\x92\x00\x80\xBB\x92"
-     "\x00\x80"
-     "\xBF\x92\x00\x80\xC3\x92\x00\x80\xC7\x92\x00\x80\xCB\x92\x00\x80\xCF\x92"
-     "\x00\x80"
-     "\xB8\x71\x03\x00\xB9\x71\x03\x00\xBA\x71\x03\x00\xBB\x71\x03\x00\xBC\xD5"
-     "\x00\x00"
-     "\xBD\xDD\x00\x00\xBE\xD5\x00\x00\xBF\xCD\x00\x00\xB0\xB5\x03\x00\xB1\xBD"
-     "\x03\x00"
-     "\xB2\x81\x03\x00\xB3\x81\x03\x00\xB4\x51\x03\x00\xB5\x51\x03\x00\xB6\x51"
-     "\x03\x00"
-     "\xB7\x51\x03\x00\xEF\xB8\x03\x00\xD3\x92\x00\x80\xD7\x92\x00\x80\xDB\x92"
-     "\x00\x80"
-     "\x84\xDC\x0D\x00\xDF\x92\x00\x80\xE3\x92\x00\x80\xE7\x92\x00\x80\xBE\xAC"
-     "\x0C\x00"
-     "\xEB\x92\x00\x80\xEF\x92\x00\x80\xF3\x92\x00\x80\xE3\x68\x03\x00\xF7\x92"
-     "\x00\x80"
-     "\xE1\x28\x00\x00\xFB\x92\x00\x80\xB3\x5D\x02\x00\x03\x93\x00\x80\x86\x68"
-     "\x0C\x00"
-     "\x87\x0C\x0D\x00\x07\x93\x00\x80\xB6\x15\x02\x00\xB5\x75\x02\x00\x0B\x93"
-     "\x00\x80"
-     "\xBB\x39\x02\x00\xBA\x31\x02\x00\x0F\x93\x00\x80\x13\x93\x00\x80\xBF\xD5"
-     "\x02\x00"
-     "\xBE\xD5\x02\x00\xBD\x15\x02\x00\xBC\x15\x02\x00\xE1\x78\x06\x00\x17\x93"
-     "\x00\x80"
-     "\xE3\xA8\x06\x00\x84\x04\x0C\x00\x80\xBD\x00\x00\x81\xA1\x00\x00\x82\xA5"
-     "\x00\x00"
-     "\x1B\x93\x00\x80\x1F\x93\x00\x80\x23\x93\x00\x80\x27\x93\x00\x80\x2B\x93"
-     "\x00\x80"
-     "\x2F\x93\x00\x80\x33\x93\x00\x80\xEF\xF8\x01\x00\x37\x93\x00\x80\x3B\x93"
-     "\x00\x80"
-     "\x3F\x93\x00\x80\xA3\xD5\x02\x00\x43\x93\x00\x80\xA5\xFD\x02\x00\x3F\x91"
-     "\x00\x80"
-     "\x47\x93\x00\x80\xA6\x9D\x02\x00\x4B\x93\x00\x80\x4F\x93\x00\x80\xAB\xB1"
-     "\x02\x00"
-     "\xAA\xB9\x02\x00\xAD\x9D\x02\x00\xAC\x9D\x02\x00\xAF\x5D\x02\x00\xAE\x5D"
-     "\x02\x00"
-     "\xB3\x5D\x0E\x00\xFF\x92\x00\x80\x53\x93\x00\x80\x57\x93\x00\x80\x5B\x93"
-     "\x00\x80"
-     "\xB6\x7D\x0E\x00\xB5\x7D\x0E\x00\x5F\x93\x00\x80\xBB\x85\x0E\x00\xBA\x59"
-     "\x0E\x00"
-     "\x63\x93\x00\x80\x67\x93\x00\x80\xBF\x85\x0E\x00\xBE\x85\x0E\x00\xBD\x95"
-     "\x0E\x00"
-     "\xBC\x95\x0E\x00\x80\x69\x00\x00\x81\x69\x00\x00\x82\x05\x00\x00\x6B\x93"
-     "\x00\x80"
-     "\xBE\xDC\x00\x00\xE3\x54\x0D\x00\x6F\x93\x00\x80\xE1\x34\x0C\x00\x73\x93"
-     "\x00\x80"
-     "\xEF\xBC\x0D\x00\x86\x48\x00\x00\x87\xCC\x03\x00\x77\x93\x00\x80\x7B\x93"
-     "\x00\x80"
-     "\x7F\x93\x00\x80\x83\x93\x00\x80\xA8\xC5\x0E\x00\xA9\xC9\x0E\x00\xAA\xD5"
-     "\x0E\x00"
-     "\xAB\xED\x0E\x00\xAC\xF5\x0E\x00\xAD\xFD\x0E\x00\xAE\xF5\x0E\x00\xAF\xE5"
-     "\x0E\x00"
-     "\x87\x93\x00\x80\x8B\x93\x00\x80\x8F\x93\x00\x80\x93\x93\x00\x80\x97\x93"
-     "\x00\x80"
-     "\x9B\x93\x00\x80\x9F\x93\x00\x80\xA3\x93\x00\x80\xB8\xC5\x0F\x00\xB9\xCD"
-     "\x0F\x00"
-     "\xBA\xC5\x0F\x00\xBB\xD9\x0F\x00\xBC\xC9\x0F\x00\xBD\xC9\x0F\x00\xBE\xFD"
-     "\x0F\x00"
-     "\xBF\xF5\x0F\x00\xB0\x9D\x0E\x00\xB1\x41\x0E\x00\xB2\x41\x0E\x00\xB3\x41"
-     "\x0E\x00"
-     "\xB4\x41\x0E\x00\xB5\x41\x0E\x00\xB6\x41\x0E\x00\xB7\x41\x0E\x00\xA3\x5D"
-     "\x0F\x00"
-     "\xA7\x93\x00\x80\xAB\x93\x00\x80\xAF\x93\x00\x80\xB3\x93\x00\x80\xA6\x7D"
-     "\x0F\x00"
-     "\xA5\x7D\x0F\x00\xB7\x93\x00\x80\xAB\x85\x0F\x00\xAA\x59\x0F\x00\xBB\x93"
-     "\x00\x80"
-     "\xBF\x93\x00\x80\xAF\x85\x0F\x00\xAE\x85\x0F\x00\xAD\x95\x0F\x00\xAC\x95"
-     "\x0F\x00"
-     "\xC3\x93\x00\x80\xB3\x19\x0E\x00\xC7\x93\x00\x80\xCB\x93\x00\x80\xB6\xD5"
-     "\x0E\x00"
-     "\xCF\x93\x00\x80\xD3\x93\x00\x80\xB5\x31\x0E\x00\xBA\xC5\x0E\x00\xBB\xCD"
-     "\x0E\x00"
-     "\xD7\x93\x00\x80\xDB\x93\x00\x80\xBE\xC5\x0E\x00\xBF\xC9\x0E\x00\xBC\xD5"
-     "\x0E\x00"
-     "\xBD\xCD\x0E\x00\xA8\xC5\x0E\x00\xA9\xD5\x0E\x00\xAA\xD5\x0E\x00\xAB\xE5"
-     "\x0E\x00"
-     "\xAC\xFD\x0E\x00\xAD\x11\x0E\x00\xAE\x05\x0E\x00\xAF\x3D\x0E\x00\x80\xF9"
-     "\x00\x00"
-     "\x81\xCD\x00\x00\x82\xC5\x00\x00\x84\x24\x01\x00\xBE\x3C\x00\x00\xE3\x93"
-     "\x00\x80"
-     "\x87\xFC\x00\x00\x86\x9C\x07\x00\xB8\x0D\x0E\x00\xB9\x1D\x0E\x00\xBA\x29"
-     "\x0E\x00"
-     "\xBB\x29\x0E\x00\xBC\xD9\x01\x00\xBD\xD9\x01\x00\xBE\xC9\x01\x00\xBF\xC1"
-     "\x01\x00"
-     "\xB0\x45\x0E\x00\xB1\x4D\x0E\x00\xB2\x45\x0E\x00\xB3\x5D\x0E\x00\xB4\x45"
-     "\x0E\x00"
-     "\xB5\x49\x0E\x00\xB6\x39\x0E\x00\xB7\x39\x0E\x00\xA3\x5D\x0D\x00\xE7\x93"
-     "\x00\x80"
-     "\xEB\x93\x00\x80\xEF\x93\x00\x80\xF3\x93\x00\x80\xA6\x91\x0D\x00\xA5\x75"
-     "\x0D\x00"
-     "\xF7\x93\x00\x80\xAB\x89\x0D\x00\xAA\x81\x0D\x00\xFB\x93\x00\x80\xFF\x93"
-     "\x00\x80"
-     "\xAF\x8D\x0D\x00\xAE\x81\x0D\x00\xAD\x89\x0D\x00\xAC\x91\x0D\x00\x84\xC4"
-     "\x03\x00"
-     "\xB3\x31\x02\x00\x03\x94\x00\x80\x07\x94\x00\x80\xB6\xF1\x02\x00\x0B\x94"
-     "\x00\x80"
-     "\x0F\x94\x00\x80\xB5\x21\x02\x00\xBA\xCD\x02\x00\xBB\xD5\x02\x00\x13\x94"
-     "\x00\x80"
-     "\x17\x94\x00\x80\xBE\xCD\x02\x00\xBF\x71\x03\x00\xBC\xCD\x02\x00\xBD\xC5"
-     "\x02\x00"
-     "\xA8\xE1\x02\x00\xA9\xF5\x02\x00\xAA\xFD\x02\x00\xAB\xF5\x02\x00\xAC\xED"
-     "\x02\x00"
-     "\xAD\x35\x02\x00\xAE\x3D\x02\x00\xAF\x35\x02\x00\x1B\x94\x00\x80\x1F\x94"
-     "\x00\x80"
-     "\xBF\xA4\x13\x00\x23\x94\x00\x80\x27\x94\x00\x80\x2B\x94\x00\x80\x2F\x94"
-     "\x00\x80"
-     "\x33\x94\x00\x80\xB8\xD1\x03\x00\xB9\xD9\x03\x00\xBA\xE1\x03\x00\xBB\xE1"
-     "\x03\x00"
-     "\xBC\x91\x03\x00\xBD\x9D\x03\x00\xBE\x95\x03\x00\xBF\x89\x03\x00\xB0\x4D"
-     "\x02\x00"
-     "\xB1\x55\x02\x00\xB2\x5D\x02\x00\xB3\x55\x02\x00\xB4\x4D\x02\x00\xB5\xF1"
-     "\x03\x00"
-     "\xB6\xF1\x03\x00\xB7\xF1\x03\x00\xA3\x75\x03\x00\x37\x94\x00\x80\x3B\x94"
-     "\x00\x80"
-     "\x3F\x94\x00\x80\x43\x94\x00\x80\xA6\xB5\x03\x00\xA5\x65\x03\x00\x47\x94"
-     "\x00\x80"
-     "\xAB\x91\x03\x00\xAA\x89\x03\x00\x4B\x94\x00\x80\xBE\xA0\x05\x00\xAF\x35"
-     "\x02\x00"
-     "\xAE\x89\x03\x00\xAD\x81\x03\x00\xAC\x89\x03\x00\x81\x5D\x00\x00\x80\x55"
-     "\x00\x00"
-     "\xB3\x41\x01\x00\x82\x65\x00\x00\xB5\x41\x01\x00\x53\x94\x00\x80\x57\x94"
-     "\x00\x80"
-     "\xB6\x41\x01\x00\x86\xC0\x04\x00\x87\x24\x05\x00\xBB\x25\x01\x00\xBA\x25"
-     "\x01\x00"
-     "\xBD\x29\x01\x00\xBC\x21\x01\x00\xBF\x15\x01\x00\xBE\x15\x01\x00\xE1\xD0"
-     "\x0E\x00"
-     "\x5B\x94\x00\x80\xE3\x14\x0E\x00\x84\xC0\x04\x00\x5F\x94\x00\x80\x63\x94"
-     "\x00\x80"
-     "\x67\x94\x00\x80\x6B\x94\x00\x80\x84\xE8\x07\x00\xBE\xEC\x04\x00\x6F\x94"
-     "\x00\x80"
-     "\x73\x94\x00\x80\x77\x94\x00\x80\x7B\x94\x00\x80\xEF\x10\x01\x00\x84\xB0"
-     "\x06\x00"
-     "\x7F\x94\x00\x80\x83\x94\x00\x80\x87\x94\x00\x80\x8B\x94\x00\x80\x8F\x94"
-     "\x00\x80"
-     "\xE3\x68\x00\x00\x8F\x92\x00\x80\xE1\xAC\x01\x00\x93\x94\x00\x80\xEF\xAC"
-     "\x00\x00"
-     "\x97\x94\x00\x80\x9B\x94\x00\x80\x9F\x94\x00\x80\xA3\x94\x00\x80\xA7\x94"
-     "\x00\x80"
-     "\xAB\x94\x00\x80\xA3\x09\x06\x00\x4F\x94\x00\x80\xAF\x94\x00\x80\xB3\x94"
-     "\x00\x80"
-     "\xB7\x94\x00\x80\xA6\x09\x06\x00\xA5\x09\x06\x00\xBB\x94\x00\x80\xAB\x6D"
-     "\x06\x00"
-     "\xAA\x6D\x06\x00\xBF\x94\x00\x80\xC3\x94\x00\x80\xAF\x5D\x06\x00\xAE\x5D"
-     "\x06\x00"
-     "\xAD\x61\x06\x00\xAC\x69\x06\x00\x80\x4D\x00\x00\x81\x55\x00\x00\x82\x55"
-     "\x00\x00"
-     "\xC7\x94\x00\x80\xEF\x84\x05\x00\xCB\x94\x00\x80\x87\xD4\x03\x00\x86\xFC"
-     "\x00\x00"
-     "\xE3\x30\x04\x00\xCF\x94\x00\x80\xE1\xCC\x07\x00\xD3\x94\x00\x80\xD7\x94"
-     "\x00\x80"
-     "\xDB\x94\x00\x80\xDF\x94\x00\x80\xE3\x94\x00\x80\xA8\x35\x05\x00\xA9\x81"
-     "\x06\x00"
-     "\xAA\x81\x06\x00\xAB\x81\x06\x00\xAC\x81\x06\x00\xAD\x89\x06\x00\xAE\xB1"
-     "\x06\x00"
-     "\xAF\xB1\x06\x00\xBE\xE4\x00\x00\xE7\x94\x00\x80\xEB\x94\x00\x80\xEF\x94"
-     "\x00\x80"
-     "\xF3\x94\x00\x80\xF7\x94\x00\x80\xFB\x94\x00\x80\xFF\x94\x00\x80\xB8\x6D"
-     "\x06\x00"
-     "\xB9\x75\x06\x00\xBA\x7D\x06\x00\xBB\x75\x06\x00\xBC\x6D\x06\x00\xBD\xD1"
-     "\x07\x00"
-     "\xBE\xD1\x07\x00\xBF\xD1\x07\x00\xB0\xA9\x06\x00\xB1\xA9\x06\x00\xB2\xBD"
-     "\x06\x00"
-     "\xB3\xB5\x06\x00\xB4\x99\x06\x00\xB5\x99\x06\x00\xB6\x59\x06\x00\xB7\x59"
-     "\x06\x00"
-     "\xA8\x55\x06\x00\xA9\x5D\x06\x00\xAA\x8D\x06\x00\xAB\x9D\x06\x00\xAC\x85"
-     "\x06\x00"
-     "\xAD\x8D\x06\x00\xAE\x85\x06\x00\xAF\xB5\x06\x00\x03\x95\x00\x80\x07\x95"
-     "\x00\x80"
-     "\x0B\x95\x00\x80\x0F\x95\x00\x80\x13\x95\x00\x80\x17\x95\x00\x80\x1B\x95"
-     "\x00\x80"
-     "\x1F\x95\x00\x80\xB8\xA1\x06\x00\xB9\xAD\x06\x00\xBA\xA5\x06\x00\xBB\xBD"
-     "\x06\x00"
-     "\xBC\xA5\x06\x00\xBD\xAD\x06\x00\xBE\xA5\x06\x00\xBF\x55\x01\x00\xB0\xCD"
-     "\x06\x00"
-     "\xB1\xD5\x06\x00\xB2\xDD\x06\x00\xB3\xAD\x06\x00\xB4\xB5\x06\x00\xB5\xBD"
-     "\x06\x00"
-     "\xB6\xB5\x06\x00\xB7\xAD\x06\x00\xB3\x05\x06\x00\x82\x2D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x23\x95\x00\x80\xB6\x0D\x06\x00\xB5\x0D\x06\x00\x27\x95"
-     "\x00\x80"
-     "\xBB\x11\x06\x00\xBA\x09\x06\x00\x2B\x95\x00\x80\x33\x95\x00\x80\xBF\x79"
-     "\x06\x00"
-     "\xBE\x71\x06\x00\xBD\x79\x06\x00\xBC\x09\x06\x00\x37\x95\x00\x80\xA3\x41"
-     "\x06\x00"
-     "\x86\x68\x00\x00\x87\x54\x00\x00\xA6\x49\x06\x00\x3B\x95\x00\x80\x3F\x95"
-     "\x00\x80"
-     "\xA5\x49\x06\x00\xAA\x4D\x06\x00\xAB\x55\x06\x00\x43\x95\x00\x80\x47\x95"
-     "\x00\x80"
-     "\xAE\x35\x06\x00\xAF\x3D\x06\x00\xAC\x4D\x06\x00\xAD\x3D\x06\x00\xA8\x2D"
-     "\x02\x00"
-     "\xA9\x49\x02\x00\xAA\x5D\x02\x00\xAB\x69\x02\x00\xAC\x7D\x02\x00\xAD\x65"
-     "\x02\x00"
-     "\xAE\x6D\x02\x00\xAF\x65\x02\x00\xBE\xAC\x1C\x00\x4B\x95\x00\x80\x4F\x95"
-     "\x00\x80"
-     "\x53\x95\x00\x80\x57\x95\x00\x80\x5B\x95\x00\x80\x5F\x95\x00\x80\x63\x95"
-     "\x00\x80"
-     "\xB8\x61\x03\x00\xB9\x61\x03\x00\xBA\x61\x03\x00\xBB\x61\x03\x00\xBC\x61"
-     "\x03\x00"
-     "\xBD\x61\x03\x00\xBE\x61\x03\x00\xBF\x61\x03\x00\xB0\x1D\x02\x00\xB1\xC1"
-     "\x02\x00"
-     "\xB2\xC1\x02\x00\xB3\xC1\x02\x00\xB4\xC1\x02\x00\xB5\xC1\x02\x00\xB6\xC1"
-     "\x02\x00"
-     "\xB7\xC1\x02\x00\x67\x95\x00\x80\xB3\xD1\x02\x00\x6B\x95\x00\x80\x84\x40"
-     "\x1D\x00"
-     "\xB6\xE1\x02\x00\x6F\x95\x00\x80\x73\x95\x00\x80\xB5\xF1\x02\x00\xBA\xDD"
-     "\x02\x00"
-     "\xBB\x01\x02\x00\x77\x95\x00\x80\x7B\x95\x00\x80\xBE\x01\x02\x00\xBF\x01"
-     "\x02\x00"
-     "\xBC\x11\x02\x00\xBD\x11\x02\x00\x7F\x95\x00\x80\x83\x95\x00\x80\xA3\x9D"
-     "\x02\x00"
-     "\x87\x95\x00\x80\xA5\xBD\x02\x00\x8B\x95\x00\x80\x8F\x95\x00\x80\xA6\xAD"
-     "\x02\x00"
-     "\x93\x95\x00\x80\x97\x95\x00\x80\xAB\x4D\x02\x00\xAA\x91\x02\x00\xAD\x5D"
-     "\x02\x00"
-     "\xAC\x5D\x02\x00\xAF\x4D\x02\x00\xAE\x4D\x02\x00\xA8\xAD\x02\x00\xA9\xC5"
-     "\x02\x00"
-     "\xAA\xDD\x02\x00\xAB\xD5\x02\x00\xAC\xF9\x02\x00\xAD\xF9\x02\x00\xAE\x3D"
-     "\x01\x00"
-     "\xAF\x35\x01\x00\x9B\x95\x00\x80\xA3\x95\x00\x80\xA7\x95\x00\x80\xAB\x95"
-     "\x00\x80"
-     "\xAF\x95\x00\x80\xB3\x95\x00\x80\xB7\x95\x00\x80\xBB\x95\x00\x80\xB8\xD9"
-     "\x01\x00"
-     "\xB9\xED\x01\x00\xBA\xE5\x01\x00\xBB\x8D\x01\x00\xBC\x95\x01\x00\xBD\x99"
-     "\x01\x00"
-     "\xBE\x8D\x01\x00\xBF\x85\x01\x00\xB0\x4D\x01\x00\xB1\x55\x01\x00\xB2\x5D"
-     "\x01\x00"
-     "\xB3\x55\x01\x00\xB4\x4D\x01\x00\xB5\xF5\x01\x00\xB6\xF1\x01\x00\xB7\xE9"
-     "\x01\x00"
-     "\xE1\xF8\x06\x00\xBF\x95\x00\x80\xE3\x14\x06\x00\xBE\x44\x1C\x00\x82\xB1"
-     "\x00\x00"
-     "\xEF\xCC\x01\x00\x80\x79\x01\x00\x81\x79\x01\x00\x84\x6C\x1F\x00\xE3\x14"
-     "\x06\x00"
-     "\xC3\x95\x00\x80\xE1\xE8\x01\x00\xC7\x95\x00\x80\xCB\x95\x00\x80\xEF\xE4"
-     "\x06\x00"
-     "\xCF\x95\x00\x80\xB3\x15\x02\x00\xD3\x95\x00\x80\x86\xE8\x1D\x00\x87\x9C"
-     "\x1D\x00"
-     "\x85\xC0\x17\x00\xB6\x39\x02\x00\xB5\x31\x02\x00\xD7\x95\x00\x80\xBB\xD1"
-     "\x01\x00"
-     "\xBA\xC9\x01\x00\xDB\x95\x00\x80\xDF\x93\x00\x80\xBF\x71\x01\x00\xBE\xC9"
-     "\x01\x00"
-     "\xBD\xC1\x01\x00\xBC\xC9\x01\x00\xA3\xD5\x1D\x00\x9F\x95\x00\x80\xDF\x95"
-     "\x00\x80"
-     "\xE3\x95\x00\x80\xE7\x95\x00\x80\xA6\xF9\x1D\x00\xA5\xF1\x1D\x00\xEB\x95"
-     "\x00\x80"
-     "\xAB\x11\x1E\x00\xAA\x09\x1E\x00\xEF\x95\x00\x80\xF3\x95\x00\x80\xAF\xB1"
-     "\x1E\x00"
-     "\xAE\x09\x1E\x00\xAD\x01\x1E\x00\xAC\x09\x1E\x00\xF7\x95\x00\x80\xFB\x95"
-     "\x00\x80"
-     "\xFF\x95\x00\x80\xEF\x94\x1C\x00\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D"
-     "\x00\x00"
-     "\x03\x96\x00\x80\x07\x96\x00\x80\x0B\x96\x00\x80\x0F\x96\x00\x80\x13\x96"
-     "\x00\x80"
-     "\x17\x96\x00\x80\xE1\xA4\x1F\x00\x1B\x96\x00\x80\xE3\xD0\x19\x00\xA8\xA5"
-     "\x1F\x00"
-     "\xA9\xAD\x1F\x00\xAA\xA5\x1F\x00\xAB\xBD\x1F\x00\xAC\xA5\x1F\x00\xAD\xAD"
-     "\x1F\x00"
-     "\xAE\xA5\x1F\x00\xAF\x19\x1F\x00\x86\x80\x03\x00\x87\x08\x03\x00\x1F\x96"
-     "\x00\x80"
-     "\x23\x96\x00\x80\x27\x96\x00\x80\x2B\x96\x00\x80\x2F\x96\x00\x80\x33\x96"
-     "\x00\x80"
-     "\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5\x00\x00\xBB\xE5\x00\x00\xBC\xFD"
-     "\x00\x00"
-     "\xBD\x95\x00\x00\xBE\x9D\x00\x00\xBF\x95\x00\x00\xB0\x49\x1F\x00\xB1\x55"
-     "\x1F\x00"
-     "\xB2\x5D\x1F\x00\xB3\x51\x1F\x00\xB4\x35\x1F\x00\xB5\x29\x1F\x00\xB6\x11"
-     "\x1F\x00"
-     "\xB7\xF5\x00\x00\xB3\x2D\x1E\x00\x37\x96\x00\x80\x3B\x96\x00\x80\x3F\x96"
-     "\x00\x80"
-     "\x43\x96\x00\x80\xB6\x2D\x1E\x00\xB5\x2D\x1E\x00\x47\x96\x00\x80\xBB\x91"
-     "\x1E\x00"
-     "\xBA\x8D\x1E\x00\x4B\x96\x00\x80\x4F\x96\x00\x80\xBF\xED\x1E\x00\xBE\xF1"
-     "\x1E\x00"
-     "\xBD\x85\x1E\x00\xBC\x89\x1E\x00\x53\x96\x00\x80\xA3\x69\x1E\x00\x57\x96"
-     "\x00\x80"
-     "\x5B\x96\x00\x80\xA6\x69\x1E\x00\x5F\x96\x00\x80\x63\x96\x00\x80\xA5\x69"
-     "\x1E\x00"
-     "\xAA\xC9\x1E\x00\xAB\xD5\x1E\x00\x67\x96\x00\x80\x6B\x96\x00\x80\xAE\xB5"
-     "\x1E\x00"
-     "\xAF\xA9\x1E\x00\xAC\xCD\x1E\x00\xAD\xC1\x1E\x00\xA8\xED\x01\x00\xA9\xF5"
-     "\x01\x00"
-     "\xAA\xF1\x01\x00\xAB\x25\x01\x00\xAC\x3D\x01\x00\xAD\x25\x01\x00\xAE\x2D"
-     "\x01\x00"
-     "\xAF\x25\x01\x00\x6F\x96\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\x73\x96\x00\x80\x77\x96\x00\x80\x7B\x96\x00\x80\xBE\x74\x01\x00\xB8\x81"
-     "\x00\x00"
-     "\xB9\x8D\x00\x00\xBA\x85\x00\x00\xBB\x99\x00\x00\xBC\x89\x00\x00\xBD\xBD"
-     "\x00\x00"
-     "\xBE\xB5\x00\x00\xBF\x7D\x00\x00\xB0\x5D\x01\x00\xB1\xE1\x00\x00\xB2\xE1"
-     "\x00\x00"
-     "\xB3\xE1\x00\x00\xB4\xE1\x00\x00\xB5\xE9\x00\x00\xB6\xD1\x00\x00\xB7\xD1"
-     "\x00\x00"
-     "\xB3\x79\x02\x00\x83\x96\x00\x80\x87\x28\x03\x00\x86\x2C\x02\x00\x87\x96"
-     "\x00\x80"
-     "\xB6\xA5\x02\x00\xB5\xB5\x02\x00\x8B\x96\x00\x80\xBB\x25\x03\x00\xBA\x99"
-     "\x02\x00"
-     "\x8F\x96\x00\x80\x93\x96\x00\x80\xBF\x2D\x03\x00\xBE\x25\x03\x00\xBD\x35"
-     "\x03\x00"
-     "\xBC\x35\x03\x00\x97\x96\x00\x80\xA3\x3D\x02\x00\x9B\x96\x00\x80\x9F\x96"
-     "\x00\x80"
-     "\xA6\xE1\x02\x00\xA3\x96\x00\x80\xA7\x96\x00\x80\xA5\xF1\x02\x00\xAA\xDD"
-     "\x02\x00"
-     "\xAB\x61\x03\x00\xAB\x96\x00\x80\xAF\x96\x00\x80\xAE\x61\x03\x00\xAF\x69"
-     "\x03\x00"
-     "\xAC\x71\x03\x00\xAD\x71\x03\x00\xA8\x85\x00\x00\xA9\x95\x00\x00\xAA\x95"
-     "\x00\x00"
-     "\xAB\xA5\x00\x00\xAC\xBD\x00\x00\xAD\xD5\x00\x00\xAE\xD1\x00\x00\xAF\xD1"
-     "\x00\x00"
-     "\xB3\x96\x00\x80\xB7\x96\x00\x80\xBB\x96\x00\x80\xBF\x96\x00\x80\xC3\x96"
-     "\x00\x80"
-     "\xC7\x96\x00\x80\xAE\x00\x00\x00\xCB\x96\x00\x80\xB8\x71\x00\x00\xB9\x71"
-     "\x00\x00"
-     "\xBA\x75\x00\x00\xBB\xCD\x01\x00\xBC\xD5\x01\x00\xBD\xDD\x01\x00\xBE\xD5"
-     "\x01\x00"
-     "\xBF\xCD\x01\x00\xB0\xB5\x00\x00\xB1\xBD\x00\x00\xB2\x81\x00\x00\xB3\x81"
-     "\x00\x00"
-     "\xB4\x51\x00\x00\xB5\x51\x00\x00\xB6\x51\x00\x00\xB7\x51\x00\x00\xCF\x96"
-     "\x00\x80"
-     "\xD3\x96\x00\x80\xD7\x96\x00\x80\xDB\x96\x00\x80\xDF\x96\x00\x80\xE3\x96"
-     "\x00\x80"
-     "\xE7\x96\x00\x80\xEB\x96\x00\x80\x80\x35\x00\x00\x81\x3D\x00\x00\x82\x35"
-     "\x00\x00"
-     "\x84\xA0\x02\x00\xBE\xFC\x05\x00\xF3\x96\x00\x80\x87\xC4\x02\x00\x86\x1C"
-     "\x05\x00"
-     "\xB3\x5D\x01\x00\x85\x20\x0B\x00\xF7\x96\x00\x80\xFB\x96\x00\x80\xFF\x96"
-     "\x00\x80"
-     "\xB6\x15\x01\x00\xB5\x75\x01\x00\x03\x97\x00\x80\xBB\x39\x01\x00\xBA\x31"
-     "\x01\x00"
-     "\x07\x97\x00\x80\x0B\x97\x00\x80\xBF\xD5\x01\x00\xBE\xD5\x01\x00\xBD\x15"
-     "\x01\x00"
-     "\xBC\x15\x01\x00\xE1\x50\x1E\x00\x0F\x97\x00\x80\xE3\x14\x1F\x00\x13\x97"
-     "\x00\x80"
-     "\x17\x97\x00\x80\x1B\x97\x00\x80\x1F\x97\x00\x80\x23\x97\x00\x80\x27\x97"
-     "\x00\x80"
-     "\x2B\x97\x00\x80\x2F\x97\x00\x80\x84\x60\x04\x00\x33\x97\x00\x80\x37\x97"
-     "\x00\x80"
-     "\xEF\xCC\x1E\x00\x3B\x97\x00\x80\x3F\x97\x00\x80\x43\x97\x00\x80\xBE\x00"
-     "\x04\x00"
-     "\x84\x00\x07\x00\x47\x97\x00\x80\xEF\x90\x00\x00\x4B\x97\x00\x80\x4F\x97"
-     "\x00\x80"
-     "\x53\x97\x00\x80\xE3\x54\x1E\x00\x57\x97\x00\x80\xE1\x54\x01\x00\x2F\x95"
-     "\x00\x80"
-     "\x82\xF9\x07\x00\x81\x11\x00\x00\x80\x1D\x00\x00\xE1\x14\x04\x00\xEF\x96"
-     "\x00\x80"
-     "\xE3\x34\x05\x00\x5B\x97\x00\x80\x5F\x97\x00\x80\x63\x97\x00\x80\x86\x38"
-     "\x03\x00"
-     "\x87\x54\x03\x00\x67\x97\x00\x80\x6B\x97\x00\x80\x6F\x97\x00\x80\x73\x97"
-     "\x00\x80"
-     "\x77\x97\x00\x80\x7B\x97\x00\x80\xEF\xF0\x05\x00\x7F\x97\x00\x80\xA3\x5D"
-     "\x06\x00"
-     "\x83\x97\x00\x80\x87\x97\x00\x80\x8B\x97\x00\x80\x8F\x97\x00\x80\xA6\x15"
-     "\x06\x00"
-     "\xA5\x75\x06\x00\x93\x97\x00\x80\xAB\x39\x06\x00\xAA\x31\x06\x00\x97\x97"
-     "\x00\x80"
-     "\x9B\x97\x00\x80\xAF\xD5\x06\x00\xAE\xD5\x06\x00\xAD\x15\x06\x00\xAC\x15"
-     "\x06\x00"
-     "\xA8\xB5\x07\x00\xA9\xCD\x07\x00\xAA\xC5\x07\x00\xAB\xDD\x07\x00\xAC\xCD"
-     "\x07\x00"
-     "\xAD\xFD\x07\x00\xAE\xF5\x07\x00\xAF\x19\x07\x00\x9F\x97\x00\x80\xA3\x97"
-     "\x00\x80"
-     "\xA7\x97\x00\x80\xAB\x97\x00\x80\xAF\x97\x00\x80\xB3\x97\x00\x80\xB7\x97"
-     "\x00\x80"
-     "\xBB\x97\x00\x80\xB8\x71\x07\x00\xB9\x71\x07\x00\xBA\x75\x07\x00\xBB\x0D"
-     "\x07\x00"
-     "\xBC\x15\x07\x00\xBD\x05\x07\x00\xBE\x01\x07\x00\xBF\x01\x07\x00\xB0\x6D"
-     "\x07\x00"
-     "\xB1\x75\x07\x00\xB2\x7D\x07\x00\xB3\x75\x07\x00\xB4\x6D\x07\x00\xB5\x55"
-     "\x07\x00"
-     "\xB6\x5D\x07\x00\xB7\x55\x07\x00\xB3\x4D\x06\x00\xBF\x97\x00\x80\xC3\x97"
-     "\x00\x80"
-     "\xC7\x97\x00\x80\xCB\x97\x00\x80\xB6\x7D\x06\x00\xB5\x45\x06\x00\xBE\xD4"
-     "\x01\x00"
-     "\xBB\xB1\x06\x00\xBA\xA9\x06\x00\xCF\x97\x00\x80\xD3\x97\x00\x80\xBF\x91"
-     "\x06\x00"
-     "\xBE\xA9\x06\x00\xBD\xA1\x06\x00\xBC\xA9\x06\x00\x82\x51\x00\x00\xA3\x09"
-     "\x06\x00"
-     "\x80\x41\x00\x00\x81\x41\x00\x00\xA6\x39\x06\x00\xD7\x97\x00\x80\xDB\x97"
-     "\x00\x80"
-     "\xA5\x01\x06\x00\xAA\xED\x06\x00\xAB\xF5\x06\x00\x86\xC8\x00\x00\x87\xEC"
-     "\x01\x00"
-     "\xAE\xED\x06\x00\xAF\xD5\x06\x00\xAC\xED\x06\x00\xAD\xE5\x06\x00\xA8\x95"
-     "\x01\x00"
-     "\xA9\xA5\x01\x00\xAA\xAD\x01\x00\xAB\xA5\x01\x00\xAC\xBD\x01\x00\xAD\xA5"
-     "\x01\x00"
-     "\xAE\xAD\x01\x00\xAF\xA1\x01\x00\x84\xA8\x01\x00\xDF\x97\x00\x80\xE3\x97"
-     "\x00\x80"
-     "\xE7\x97\x00\x80\xEB\x97\x00\x80\xEF\x97\x00\x80\xF3\x97\x00\x80\xF7\x97"
-     "\x00\x80"
-     "\xB8\x65\x01\x00\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x7D\x01\x00\xBC\x65"
-     "\x01\x00"
-     "\xBD\x6D\x01\x00\xBE\x65\x01\x00\xBF\xD9\x00\x00\xB0\xA5\x01\x00\xB1\xAD"
-     "\x01\x00"
-     "\xB2\xA5\x01\x00\xB3\xBD\x01\x00\xB4\xAD\x01\x00\xB5\x9D\x01\x00\xB6\x95"
-     "\x01\x00"
-     "\xB7\x5D\x01\x00\xB3\x25\x02\x00\xFB\x97\x00\x80\xFF\x97\x00\x80\x03\x98"
-     "\x00\x80"
-     "\x07\x98\x00\x80\xB6\x25\x02\x00\xB5\x35\x02\x00\x0B\x98\x00\x80\xBB\x69"
-     "\x02\x00"
-     "\xBA\x61\x02\x00\x0F\x98\x00\x80\x13\x98\x00\x80\xBF\x59\x02\x00\xBE\x69"
-     "\x02\x00"
-     "\xBD\x69\x02\x00\xBC\x71\x02\x00\x17\x98\x00\x80\xA3\x61\x02\x00\x1B\x98"
-     "\x00\x80"
-     "\x1F\x98\x00\x80\xA6\x61\x02\x00\x23\x98\x00\x80\x27\x98\x00\x80\xA5\x71"
-     "\x02\x00"
-     "\xAA\x25\x02\x00\xAB\x2D\x02\x00\x2B\x98\x00\x80\x2F\x98\x00\x80\xAE\x2D"
-     "\x02\x00"
-     "\xAF\x1D\x02\x00\xAC\x35\x02\x00\xAD\x2D\x02\x00\x80\x0D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x11\x00\x00\x33\x98\x00\x80\x37\x98\x00\x80\xEF\xF0\x02\x00\x3B\x98"
-     "\x00\x80"
-     "\xBE\xD0\x03\x00\x43\x98\x00\x80\xE3\x14\x01\x00\x47\x98\x00\x80\xE1\xDC"
-     "\x01\x00"
-     "\x4B\x98\x00\x80\x4F\x98\x00\x80\x53\x98\x00\x80\x57\x98\x00\x80\x5B\x98"
-     "\x00\x80"
-     "\xB3\x9D\x03\x00\x86\xE8\x0C\x00\x87\x00\x03\x00\xB6\x7D\x03\x00\x5F\x98"
-     "\x00\x80"
-     "\x63\x98\x00\x80\xB5\x75\x03\x00\xBA\x59\x03\x00\xBB\x59\x03\x00\x67\x98"
-     "\x00\x80"
-     "\x6B\x98\x00\x80\xBE\xFD\x00\x00\xBF\xE5\x00\x00\xBC\xFD\x00\x00\xBD\xF5"
-     "\x00\x00"
-     "\xA8\x2D\x02\x00\xA9\x51\x02\x00\xAA\x51\x02\x00\xAB\xA9\x02\x00\xAC\xB9"
-     "\x02\x00"
-     "\xAD\xB9\x02\x00\xAE\xA9\x02\x00\xAF\xA9\x02\x00\x84\xAC\x0D\x00\x6F\x98"
-     "\x00\x80"
-     "\x73\x98\x00\x80\x77\x98\x00\x80\x7B\x98\x00\x80\x7F\x98\x00\x80\x83\x98"
-     "\x00\x80"
-     "\x87\x98\x00\x80\xB8\x6D\x01\x00\xB9\x05\x01\x00\xBA\x01\x01\x00\xBB\x01"
-     "\x01\x00"
-     "\xBC\x05\x01\x00\xBD\x0D\x01\x00\xBE\x31\x01\x00\xBF\x31\x01\x00\xB0\xD9"
-     "\x02\x00"
-     "\xB1\xD9\x02\x00\xB2\x6D\x01\x00\xB3\x65\x01\x00\xB4\x7D\x01\x00\xB5\x65"
-     "\x01\x00"
-     "\xB6\x65\x01\x00\xB7\x55\x01\x00\xE1\xA4\x06\x00\x7F\x96\x00\x80\xE3\x34"
-     "\x06\x00"
-     "\x8B\x98\x00\x80\x8F\x98\x00\x80\xE1\x94\x01\x00\x93\x98\x00\x80\xE3\x40"
-     "\x06\x00"
-     "\x97\x98\x00\x80\x9B\x98\x00\x80\x9F\x98\x00\x80\xEF\x48\x00\x00\xBE\x3C"
-     "\x0C\x00"
-     "\xA3\x98\x00\x80\xEF\x88\x06\x00\xA7\x98\x00\x80\xA3\x99\x02\x00\x82\x15"
-     "\x00\x00"
-     "\x81\x75\x00\x00\x80\x79\x00\x00\xAB\x98\x00\x80\xA6\x79\x02\x00\xA5\x71"
-     "\x02\x00"
-     "\xB3\x98\x00\x80\xAB\x5D\x02\x00\xAA\x5D\x02\x00\x86\xC8\x0C\x00\x87\x5C"
-     "\x0D\x00"
-     "\xAF\xE1\x01\x00\xAE\xF9\x01\x00\xAD\xF1\x01\x00\xAC\xF9\x01\x00\xAF\x98"
-     "\x00\x80"
-     "\xB7\x98\x00\x80\xBB\x98\x00\x80\xBF\x98\x00\x80\xE1\xC0\x0C\x00\xC3\x98"
-     "\x00\x80"
-     "\xE3\x28\x0D\x00\xC7\x98\x00\x80\xCB\x98\x00\x80\xCF\x98\x00\x80\xEF\x60"
-     "\x0D\x00"
-     "\xD3\x98\x00\x80\xD7\x98\x00\x80\xDB\x98\x00\x80\xDF\x98\x00\x80\xE3\x98"
-     "\x00\x80"
-     "\xB3\xF1\x0E\x00\xE7\x98\x00\x80\xEB\x98\x00\x80\xEF\x98\x00\x80\xF3\x98"
-     "\x00\x80"
-     "\xB6\xE9\x0E\x00\xB5\xE1\x0E\x00\xF7\x98\x00\x80\xBB\x89\x0E\x00\xBA\xBD"
-     "\x0E\x00"
-     "\xFB\x98\x00\x80\xFF\x98\x00\x80\xBF\x5D\x0F\x00\xBE\x5D\x0F\x00\xBD\x5D"
-     "\x0F\x00"
-     "\xBC\x91\x0E\x00\xA8\x21\x0E\x00\xA9\x2D\x0E\x00\xAA\x39\x0E\x00\xAB\x39"
-     "\x0E\x00"
-     "\xAC\x29\x0E\x00\xAD\x29\x0E\x00\xAE\x59\x0E\x00\xAF\x49\x0E\x00\x03\x99"
-     "\x00\x80"
-     "\x07\x99\x00\x80\x0B\x99\x00\x80\x0F\x99\x00\x80\x80\x0D\x00\x00\x81\xB1"
-     "\x01\x00"
-     "\x82\xB1\x01\x00\x13\x99\x00\x80\xB8\xF9\x0F\x00\xB9\xF9\x0F\x00\xBA\x45"
-     "\x0F\x00"
-     "\xBB\x5D\x0F\x00\xBC\x49\x0F\x00\xBD\x75\x0F\x00\xBE\x7D\x0F\x00\xBF\x75"
-     "\x0F\x00"
-     "\xB0\x39\x0E\x00\xB1\x0D\x0E\x00\xB2\x05\x0E\x00\xB3\xC9\x0F\x00\xB4\xD9"
-     "\x0F\x00"
-     "\xB5\xD9\x0F\x00\xB6\xC9\x0F\x00\xB7\xC9\x0F\x00\xA3\xB1\x0F\x00\x17\x99"
-     "\x00\x80"
-     "\x1B\x99\x00\x80\x84\x84\x02\x00\xBE\x9C\x01\x00\xA6\xA9\x0F\x00\xA5\xA1"
-     "\x0F\x00"
-     "\x23\x99\x00\x80\xAB\xC9\x0F\x00\xAA\xFD\x0F\x00\x86\x88\x07\x00\x87\x5C"
-     "\x00\x00"
-     "\xAF\x1D\x0E\x00\xAE\x1D\x0E\x00\xAD\x1D\x0E\x00\xAC\xD1\x0F\x00\x27\x99"
-     "\x00\x80"
-     "\xB3\x4D\x01\x00\x2B\x99\x00\x80\x2F\x99\x00\x80\xB6\x4D\x01\x00\x33\x99"
-     "\x00\x80"
-     "\x37\x99\x00\x80\xB5\x49\x01\x00\xBA\x4D\x01\x00\xBB\x55\x01\x00\x3B\x99"
-     "\x00\x80"
-     "\x3F\x99\x00\x80\xBE\x4D\x01\x00\xBF\x31\x01\x00\xBC\x4D\x01\x00\xBD\x45"
-     "\x01\x00"
-     "\xA8\x5D\x0E\x00\xA9\x81\x01\x00\xAA\x81\x01\x00\xAB\x81\x01\x00\xAC\x85"
-     "\x01\x00"
-     "\xAD\x89\x01\x00\xAE\xB9\x01\x00\xAF\xB9\x01\x00\x43\x99\x00\x80\x47\x99"
-     "\x00\x80"
-     "\x4B\x99\x00\x80\x4F\x99\x00\x80\x53\x99\x00\x80\x57\x99\x00\x80\x5B\x99"
-     "\x00\x80"
-     "\x5F\x99\x00\x80\xB8\x91\x01\x00\xB9\x99\x01\x00\xBA\xA1\x01\x00\xBB\xA1"
-     "\x01\x00"
-     "\xBC\x71\x01\x00\xBD\x71\x01\x00\xBE\x71\x01\x00\xBF\x71\x01\x00\xB0\xCD"
-     "\x01\x00"
-     "\xB1\xD5\x01\x00\xB2\xDD\x01\x00\xB3\xD5\x01\x00\xB4\xCD\x01\x00\xB5\xB1"
-     "\x01\x00"
-     "\xB6\xB1\x01\x00\xB7\xB1\x01\x00\xA3\x09\x02\x00\x63\x99\x00\x80\x67\x99"
-     "\x00\x80"
-     "\x6B\x99\x00\x80\x6F\x99\x00\x80\xA6\x09\x02\x00\xA5\x0D\x02\x00\x73\x99"
-     "\x00\x80"
-     "\xAB\x11\x02\x00\xAA\x09\x02\x00\x77\x99\x00\x80\x8F\x25\x1A\x00\xAF\x75"
-     "\x02\x00"
-     "\xAE\x09\x02\x00\xAD\x01\x02\x00\xAC\x09\x02\x00\x94\xFD\x1E\x00\x95\x9D"
-     "\x11\x00"
-     "\x96\x4D\x12\x00\x97\x15\x12\x00\x90\x5D\x1A\x00\x91\x49\x1B\x00\x92\x8D"
-     "\x1D\x00"
-     "\x93\x59\x1E\x00\x9C\x65\x0A\x00\x9D\x0D\x0A\x00\x9E\xED\x0D\x00\x9F\x69"
-     "\x0E\x00"
-     "\x98\x9D\x15\x00\x99\x71\x16\x00\x9A\x51\x16\x00\x9B\x8D\x09\x00\x7B\x99"
-     "\x00\x80"
-     "\xE1\xD4\x0B\x00\x7F\x99\x00\x80\xE3\xD8\x7E\x00\xA0\x81\x0E\x00\xA1\xBD"
-     "\x0F\x00"
-     "\xA2\x0D\x01\x00\xA3\x9D\x02\x00\xB1\x01\x68\x00\xB0\x49\x76\x00\xB3\x69"
-     "\x6A\x00"
-     "\xB2\x8D\x6A\x00\xB5\x91\x6E\x00\xB4\x01\x6C\x00\x83\x99\x00\x80\xEF\x64"
-     "\x6C\x00"
-     "\xA9\xD1\x7E\x00\xA8\xC1\x7E\x00\xAB\xCD\x72\x00\xAA\xBD\x7F\x00\xAD\x91"
-     "\x73\x00"
-     "\xAC\xDD\x72\x00\xAF\xF9\x76\x00\xAE\x01\x74\x00\xA1\x01\x04\x00\x87\x99"
-     "\x00\x80"
-     "\xA3\x79\x06\x00\xA2\xCD\x06\x00\xA5\xD1\x7A\x00\xA4\x01\x78\x00\xA7\x01"
-     "\x7C\x00"
-     "\xA6\xC1\x7A\x00\x83\x0D\x1B\x00\x82\x0D\x1B\x00\x8B\x99\x00\x80\xBE\x00"
-     "\x05\x00"
-     "\x87\xAD\x1F\x00\x86\x11\x1F\x00\x85\xCD\x1C\x00\x84\x01\x1A\x00\x8B\x25"
-     "\x13\x00"
-     "\x8A\xFD\x10\x00\x84\x30\x02\x00\x93\x99\x00\x80\x8F\x19\x17\x00\x8E\x31"
-     "\x17\x00"
-     "\x8D\xD5\x14\x00\x8C\x65\x13\x00\x93\x45\x0B\x00\x92\xB5\x08\x00\x86\x68"
-     "\x04\x00"
-     "\x87\x7C\x05\x00\x97\x41\x0F\x00\x96\x51\x0F\x00\x95\xDD\x0C\x00\x94\x35"
-     "\x0B\x00"
-     "\x9B\xF5\x00\x00\x9A\x4D\x0E\x00\x97\x99\x00\x80\x9B\x99\x00\x80\x80\xC9"
-     "\x01\x00"
-     "\x81\x91\x01\x00\x82\x71\x00\x00\x9C\x7D\x03\x00\x9F\x99\x00\x80\xE1\x74"
-     "\x0E\x00"
-     "\x84\x80\x05\x00\xE3\x90\x00\x00\xA3\x99\x00\x80\x3F\x98\x00\x80\xA7\x99"
-     "\x00\x80"
-     "\xAB\x99\x00\x80\xAF\x99\x00\x80\xB3\x99\x00\x80\xB7\x99\x00\x80\xBB\x99"
-     "\x00\x80"
-     "\xBF\x99\x00\x80\xC3\x99\x00\x80\xC7\x99\x00\x80\xEF\x54\x00\x00\xB6\x99"
-     "\x02\x00"
-     "\xCB\x99\x00\x80\xCF\x99\x00\x80\xB5\x99\x02\x00\xD3\x99\x00\x80\xB3\x9D"
-     "\x02\x00"
-     "\xD7\x99\x00\x80\xDB\x99\x00\x80\xBE\x51\x02\x00\xBF\x51\x02\x00\xBC\x41"
-     "\x02\x00"
-     "\xBD\x41\x02\x00\xBA\x51\x02\x00\xBB\x51\x02\x00\xDF\x99\x00\x80\xE3\x99"
-     "\x00\x80"
-     "\xA8\x1D\x05\x00\xA9\x45\x05\x00\xAA\x41\x05\x00\xAB\x41\x05\x00\xAC\x45"
-     "\x05\x00"
-     "\xAD\x4D\x05\x00\xAE\x71\x05\x00\xAF\x71\x05\x00\x8F\x99\x00\x80\xE7\x99"
-     "\x00\x80"
-     "\xEB\x99\x00\x80\xEF\x99\x00\x80\x80\x0D\x00\x00\x81\xB1\x00\x00\x82\xB1"
-     "\x00\x00"
-     "\xF3\x99\x00\x80\xB8\x45\x06\x00\xB9\x4D\x06\x00\xBA\x45\x06\x00\xBB\x5D"
-     "\x06\x00"
-     "\xBC\x4D\x06\x00\xBD\x7D\x06\x00\xBE\xD9\x06\x00\xBF\xD5\x06\x00\xB0\xC1"
-     "\x05\x00"
-     "\xB1\xC1\x05\x00\xB2\xC1\x05\x00\xB3\xC1\x05\x00\xB4\xC1\x05\x00\xB5\xC1"
-     "\x05\x00"
-     "\xB6\xC1\x05\x00\xB7\xC1\x05\x00\xA3\xC9\x05\x00\xF7\x99\x00\x80\xFB\x99"
-     "\x00\x80"
-     "\xBE\x00\x03\x00\xFF\x99\x00\x80\xA6\xCD\x05\x00\xA5\xCD\x05\x00\x03\x9A"
-     "\x00\x80"
-     "\xAB\x05\x05\x00\xAA\x05\x05\x00\x86\x08\x00\x00\x87\xDC\x03\x00\xAF\x05"
-     "\x05\x00"
-     "\xAE\x05\x05\x00\xAD\x15\x05\x00\xAC\x15\x05\x00\x07\x9A\x00\x80\xB3\xA9"
-     "\x07\x00"
-     "\x0B\x9A\x00\x80\x0F\x9A\x00\x80\xB6\x49\x07\x00\x13\x9A\x00\x80\x17\x9A"
-     "\x00\x80"
-     "\xB5\x4D\x07\x00\xBA\x45\x07\x00\xBB\x4D\x07\x00\x1B\x9A\x00\x80\x1F\x9A"
-     "\x00\x80"
-     "\xBE\x45\x07\x00\xBF\x4D\x07\x00\xBC\x55\x07\x00\xBD\x4D\x07\x00\xA3\x6D"
-     "\x06\x00"
-     "\x23\x9A\x00\x80\x27\x9A\x00\x80\x2B\x9A\x00\x80\x2F\x9A\x00\x80\xA6\x8D"
-     "\x06\x00"
-     "\xA5\x89\x06\x00\x33\x9A\x00\x80\xAB\x89\x06\x00\xAA\x81\x06\x00\x37\x9A"
-     "\x00\x80"
-     "\x3B\x9A\x00\x80\xAF\x89\x06\x00\xAE\x81\x06\x00\xAD\x89\x06\x00\xAC\x91"
-     "\x06\x00"
-     "\x3F\x9A\x00\x80\x43\x9A\x00\x80\x47\x9A\x00\x80\x4B\x9A\x00\x80\x4F\x9A"
-     "\x00\x80"
-     "\x53\x9A\x00\x80\x57\x9A\x00\x80\xEF\xA8\x04\x00\x5B\x9A\x00\x80\xE1\x54"
-     "\x06\x00"
-     "\x5F\x9A\x00\x80\xE3\x54\x04\x00\x80\x0D\x00\x00\x81\xD5\x00\x00\x82\xDD"
-     "\x00\x00"
-     "\x63\x9A\x00\x80\xA8\xAD\x06\x00\xA9\xED\x06\x00\xAA\xE5\x06\x00\xAB\x0D"
-     "\x06\x00"
-     "\xAC\x01\x06\x00\xAD\x0D\x06\x00\xAE\x05\x06\x00\xAF\x3D\x06\x00\x67\x9A"
-     "\x00\x80"
-     "\x6B\x9A\x00\x80\x84\xE4\x01\x00\xBE\xE8\x01\x00\x86\x10\x7F\x00\x87\x94"
-     "\x00\x00"
-     "\x73\x9A\x00\x80\x77\x9A\x00\x80\xB8\x11\x06\x00\xB9\x1D\x06\x00\xBA\x15"
-     "\x06\x00"
-     "\xBB\x35\x06\x00\xBC\x2D\x06\x00\xBD\xF1\x01\x00\xBE\xF1\x01\x00\xBF\xF1"
-     "\x01\x00"
-     "\xB0\x45\x06\x00\xB1\x4D\x06\x00\xB2\x45\x06\x00\xB3\x5D\x06\x00\xB4\x45"
-     "\x06\x00"
-     "\xB5\x4D\x06\x00\xB6\x3D\x06\x00\xB7\x35\x06\x00\xA8\x9D\x02\x00\xA9\xC5"
-     "\x02\x00"
-     "\xAA\xC1\x02\x00\xAB\xC1\x02\x00\xAC\xC5\x02\x00\xAD\xCD\x02\x00\xAE\xF1"
-     "\x02\x00"
-     "\xAF\xF1\x02\x00\x7B\x9A\x00\x80\x7F\x9A\x00\x80\x83\x9A\x00\x80\x87\x9A"
-     "\x00\x80"
-     "\x8C\x00\x00\x00\x8B\x9A\x00\x80\x8F\x9A\x00\x80\x93\x9A\x00\x80\xB8\x05"
-     "\x03\x00"
-     "\xB9\x0D\x03\x00\xBA\x05\x03\x00\xBB\x1D\x03\x00\xBC\x05\x03\x00\xBD\x0D"
-     "\x03\x00"
-     "\xBE\x39\x03\x00\xBF\x35\x03\x00\xB0\x65\x03\x00\xB1\x6D\x03\x00\xB2\x65"
-     "\x03\x00"
-     "\xB3\x7D\x03\x00\xB4\x6D\x03\x00\xB5\x5D\x03\x00\xB6\x55\x03\x00\xB7\x3D"
-     "\x03\x00"
-     "\xB3\x49\x02\x00\x97\x9A\x00\x80\x9B\x9A\x00\x80\x9F\x9A\x00\x80\xA3\x9A"
-     "\x00\x80"
-     "\xB6\x4D\x02\x00\xB5\x4D\x02\x00\xA7\x9A\x00\x80\xBB\xA9\x03\x00\xBA\xA1"
-     "\x03\x00"
-     "\xAB\x9A\x00\x80\xAF\x9A\x00\x80\xBF\x99\x03\x00\xBE\xA9\x03\x00\xBD\xA9"
-     "\x03\x00"
-     "\xBC\xB1\x03\x00\xB3\x9A\x00\x80\xB7\x9A\x00\x80\xA3\x01\x02\x00\xBB\x9A"
-     "\x00\x80"
-     "\xA5\x05\x02\x00\xBF\x9A\x00\x80\xC3\x9A\x00\x80\xA6\x05\x02\x00\x1F\x99"
-     "\x00\x80"
-     "\xC7\x9A\x00\x80\xAB\xE1\x03\x00\xAA\xE9\x03\x00\xAD\xE1\x03\x00\xAC\xF9"
-     "\x03\x00"
-     "\xAF\xD1\x03\x00\xAE\xE1\x03\x00\xCB\x9A\x00\x80\xCF\x9A\x00\x80\xD3\x9A"
-     "\x00\x80"
-     "\xD7\x9A\x00\x80\xDB\x9A\x00\x80\xE1\xA4\x01\x00\xBE\x44\x02\x00\xE3\xA4"
-     "\x00\x00"
-     "\x80\x29\x00\x00\x81\x29\x00\x00\x82\x39\x00\x00\xEF\x78\x00\x00\xE3\x9A"
-     "\x00\x80"
-     "\x84\xB8\x02\x00\xE7\x9A\x00\x80\xEB\x9A\x00\x80\xB3\x61\x01\x00\xEF\x9A"
-     "\x00\x80"
-     "\xF3\x9A\x00\x80\xF7\x9A\x00\x80\xFB\x9A\x00\x80\xB6\x79\x01\x00\xB5\x71"
-     "\x01\x00"
-     "\xFF\x9A\x00\x80\xBB\xE5\x01\x00\xBA\x59\x01\x00\x86\x48\x7D\x00\x87\x40"
-     "\x7D\x00"
-     "\xBF\xED\x01\x00\xBE\xE5\x01\x00\xBD\xF5\x01\x00\xBC\xF5\x01\x00\x03\x9B"
-     "\x00\x80"
-     "\x07\x9B\x00\x80\x0B\x9B\x00\x80\x84\x40\x7C\x00\x0F\x9B\x00\x80\x13\x9B"
-     "\x00\x80"
-     "\x17\x9B\x00\x80\xEF\xE8\x01\x00\x1B\x9B\x00\x80\xE1\xF4\x06\x00\xEA\x00"
-     "\x00\x00"
-     "\xE3\x34\x06\x00\x1F\x9B\x00\x80\x23\x9B\x00\x80\x84\xB0\x7C\x00\x27\x9B"
-     "\x00\x80"
-     "\x2B\x9B\x00\x80\xA6\xE9\x01\x00\xA5\xE1\x01\x00\x2F\x9B\x00\x80\xA3\xF1"
-     "\x01\x00"
-     "\x33\x9B\x00\x80\x37\x9B\x00\x80\x3B\x9B\x00\x80\xAF\x7D\x01\x00\xAE\x75"
-     "\x01\x00"
-     "\xAD\x65\x01\x00\xAC\x65\x01\x00\xAB\x75\x01\x00\xAA\xC9\x01\x00\x3F\x9B"
-     "\x00\x80"
-     "\x43\x9B\x00\x80\xA8\xB9\x7D\x00\xA9\xB9\x7D\x00\xAA\x29\x7E\x00\xAB\x29"
-     "\x7E\x00"
-     "\xAC\x39\x7E\x00\xAD\x39\x7E\x00\xAE\x2D\x7E\x00\xAF\x25\x7E\x00\xDF\x9A"
-     "\x00\x80"
-     "\x82\x11\x00\x00\x81\x19\x00\x00\x80\xD1\x7F\x00\x47\x9B\x00\x80\x4B\x9B"
-     "\x00\x80"
-     "\xBE\x34\x03\x00\x4F\x9B\x00\x80\xB8\x81\x7E\x00\xB9\x81\x7E\x00\xBA\x85"
-     "\x7E\x00"
-     "\xBB\x99\x7E\x00\xBC\x89\x7E\x00\xBD\xB1\x7E\x00\xBE\x59\x7F\x00\xBF\x59"
-     "\x7F\x00"
-     "\xB0\x5D\x7E\x00\xB1\xE1\x7E\x00\xB2\xE1\x7E\x00\xB3\xE1\x7E\x00\xB4\xE1"
-     "\x7E\x00"
-     "\xB5\xE9\x7E\x00\xB6\xD1\x7E\x00\xB7\xD1\x7E\x00\xA8\xB1\x7F\x00\xA9\xBD"
-     "\x7F\x00"
-     "\xAA\xB5\x7F\x00\xAB\xC9\x7F\x00\xAC\xD9\x7F\x00\xAD\xD9\x7F\x00\xAE\xC9"
-     "\x7F\x00"
-     "\xAF\xC1\x7F\x00\x86\x00\x00\x00\x87\x28\x03\x00\x53\x9B\x00\x80\x57\x9B"
-     "\x00\x80"
-     "\x5B\x9B\x00\x80\x5F\x9B\x00\x80\x63\x9B\x00\x80\x67\x9B\x00\x80\xB8\x75"
-     "\x7F\x00"
-     "\xB9\x79\x7F\x00\xBA\x09\x7F\x00\xBB\x01\x7F\x00\xBC\x05\x7F\x00\xBD\x0D"
-     "\x7F\x00"
-     "\xBE\x05\x7F\x00\xBF\x25\x7F\x00\xB0\x61\x7F\x00\xB1\x61\x7F\x00\xB2\x61"
-     "\x7F\x00"
-     "\xB3\x7D\x7F\x00\xB4\x65\x7F\x00\xB5\x6D\x7F\x00\xB6\x59\x7F\x00\xB7\x4D"
-     "\x7F\x00"
-     "\xB3\x3D\x7E\x00\x6B\x9B\x00\x80\x6F\x9B\x00\x80\x73\x9B\x00\x80\x77\x9B"
-     "\x00\x80"
-     "\xB6\x59\x7E\x00\xB5\x51\x7E\x00\x7B\x9B\x00\x80\xBB\xA5\x7E\x00\xBA\xA5"
-     "\x7E\x00"
-     "\x7F\x9B\x00\x80\x83\x9B\x00\x80\xBF\xA9\x7E\x00\xBE\xA1\x7E\x00\xBD\xA9"
-     "\x7E\x00"
-     "\xBC\xB5\x7E\x00\x87\x9B\x00\x80\x8B\x9B\x00\x80\xEF\xA8\x7C\x00\x8F\x9B"
-     "\x00\x80"
-     "\x93\x9B\x00\x80\x97\x9B\x00\x80\x9B\x9B\x00\x80\x9F\x9B\x00\x80\xA3\x9B"
-     "\x00\x80"
-     "\xA7\x9B\x00\x80\xAB\x9B\x00\x80\xAF\x9B\x00\x80\xE1\xD8\x7E\x00\xB3\x9B"
-     "\x00\x80"
-     "\xE3\x5C\x7C\x00\xB7\x9B\x00\x80\xA3\xBD\x7E\x00\x82\x15\x00\x00\x81\xF5"
-     "\x00\x00"
-     "\x80\xF9\x00\x00\xBB\x9B\x00\x80\xA6\xD9\x7E\x00\xA5\xD1\x7E\x00\xBE\x90"
-     "\x01\x00"
-     "\xAB\x25\x7E\x00\xAA\x25\x7E\x00\x86\x08\x00\x00\x87\x70\x01\x00\xAF\x29"
-     "\x7E\x00"
-     "\xAE\x21\x7E\x00\xAD\x29\x7E\x00\xAC\x35\x7E\x00\xC3\x9B\x00\x80\xB3\x75"
-     "\x01\x00"
-     "\xC7\x9B\x00\x80\xCB\x9B\x00\x80\xB6\x6D\x01\x00\xCF\x9B\x00\x80\xD3\x9B"
-     "\x00\x80"
-     "\xB5\x65\x01\x00\xBA\xED\x00\x00\xBB\xF5\x00\x00\xD7\x9B\x00\x80\xDB\x9B"
-     "\x00\x80"
-     "\xBE\xDD\x00\x00\xBF\xB5\x00\x00\xBC\xE5\x00\x00\xBD\xD5\x00\x00\xA8\xA9"
-     "\x02\x00"
-     "\xA9\xA9\x02\x00\xAA\xB9\x02\x00\xAB\xB9\x02\x00\xAC\xA9\x02\x00\xAD\xA9"
-     "\x02\x00"
-     "\xAE\x1D\x03\x00\xAF\x15\x03\x00\xBE\xAC\x04\x00\xDF\x9B\x00\x80\xE3\x9B"
-     "\x00\x80"
-     "\xBE\xA0\x05\x00\xE7\x9B\x00\x80\xEB\x9B\x00\x80\xEF\x9B\x00\x80\xF3\x9B"
-     "\x00\x80"
-     "\xB8\x11\x03\x00\xB9\x11\x03\x00\xBA\x11\x03\x00\xBB\x11\x03\x00\xBC\x31"
-     "\x03\x00"
-     "\xBD\x31\x03\x00\xBE\x31\x03\x00\xBF\x31\x03\x00\xB0\x6D\x03\x00\xB1\x75"
-     "\x03\x00"
-     "\xB2\x75\x03\x00\xB3\x45\x03\x00\xB4\x5D\x03\x00\xB5\x35\x03\x00\xB6\x3D"
-     "\x03\x00"
-     "\xB7\x31\x03\x00\xA3\x31\x02\x00\xF7\x9B\x00\x80\xFB\x9B\x00\x80\xFF\x9B"
-     "\x00\x80"
-     "\x03\x9C\x00\x80\xA6\x29\x02\x00\xA5\x21\x02\x00\x07\x9C\x00\x80\xAB\xB1"
-     "\x03\x00"
-     "\xAA\xA9\x03\x00\x0B\x9C\x00\x80\x0F\x9C\x00\x80\xAF\xF1\x03\x00\xAE\x99"
-     "\x03\x00"
-     "\xAD\x91\x03\x00\xAC\xA1\x03\x00\x84\xC0\x03\x00\x13\x9C\x00\x80\x17\x9C"
-     "\x00\x80"
-     "\x1B\x9C\x00\x80\x1F\x9C\x00\x80\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D"
-     "\x00\x00"
-     "\x23\x9C\x00\x80\x27\x9C\x00\x80\x2B\x9C\x00\x80\x33\x9C\x00\x80\x37\x9C"
-     "\x00\x80"
-     "\x6F\x9A\x00\x80\x86\xD8\x04\x00\x87\xBC\x04\x00\xA8\xC1\x02\x00\xA9\xCD"
-     "\x02\x00"
-     "\xAA\xC5\x02\x00\xAB\xD9\x02\x00\xAC\xC9\x02\x00\xAD\xFD\x02\x00\xAE\xF5"
-     "\x02\x00"
-     "\xAF\x39\x01\x00\x3B\x9C\x00\x80\x3F\x9C\x00\x80\x43\x9C\x00\x80\x47\x9C"
-     "\x00\x80"
-     "\x4B\x9C\x00\x80\x4F\x9C\x00\x80\x53\x9C\x00\x80\x57\x9C\x00\x80\xB8\xD1"
-     "\x01\x00"
-     "\xB9\xD9\x01\x00\xBA\xE1\x01\x00\xBB\xE1\x01\x00\xBC\x95\x01\x00\xBD\x9D"
-     "\x01\x00"
-     "\xBE\x95\x01\x00\xBF\x8D\x01\x00\xB0\x49\x01\x00\xB1\x55\x01\x00\xB2\x5D"
-     "\x01\x00"
-     "\xB3\x55\x01\x00\xB4\x4D\x01\x00\xB5\xF1\x01\x00\xB6\xF1\x01\x00\xB7\xF1"
-     "\x01\x00"
-     "\xEF\xE4\x00\x00\x5B\x9C\x00\x80\x5F\x9C\x00\x80\x84\x0C\x05\x00\x63\x9C"
-     "\x00\x80"
-     "\x67\x9C\x00\x80\x6B\x9C\x00\x80\xEF\x5C\x01\x00\x6F\x9C\x00\x80\xE1\xD4"
-     "\x7E\x00"
-     "\x73\x9C\x00\x80\xE3\x30\x7E\x00\xE3\x18\x7E\x00\x77\x9C\x00\x80\xE1\x00"
-     "\x01\x00"
-     "\x7B\x9C\x00\x80\x7F\x9C\x00\x80\x83\x9C\x00\x80\x87\x9C\x00\x80\xB3\x01"
-     "\x02\x00"
-     "\x8B\x9C\x00\x80\xB5\x0D\x02\x00\xB6\x31\x02\x00\x8F\x9C\x00\x80\x93\x9C"
-     "\x00\x80"
-     "\x97\x9C\x00\x80\xBA\xC1\x01\x00\xBB\xDD\x01\x00\xBC\xC5\x01\x00\xBD\xCD"
-     "\x01\x00"
-     "\xBE\xC5\x01\x00\xBF\x79\x01\x00\xA3\xCD\x05\x00\x2F\x9C\x00\x80\x9B\x9C"
-     "\x00\x80"
-     "\xA3\x9C\x00\x80\x9F\x9C\x00\x80\xA6\xFD\x05\x00\xA5\xC1\x05\x00\xA7\x9C"
-     "\x00\x80"
-     "\xAB\x11\x06\x00\xAA\x0D\x06\x00\x86\x08\x03\x00\x87\x7C\x03\x00\xAF\xB5"
-     "\x06\x00"
-     "\xAE\x09\x06\x00\xAD\x01\x06\x00\xAC\x09\x06\x00\x80\xA1\x07\x00\x81\xA1"
-     "\x07\x00"
-     "\x82\x41\x00\x00\xB3\x6D\x07\x00\xAB\x9C\x00\x80\xB5\x7D\x07\x00\xB6\x75"
-     "\x07\x00"
-     "\xAF\x9C\x00\x80\xB3\x9C\x00\x80\xB7\x9C\x00\x80\xBA\x51\x07\x00\xBB\x51"
-     "\x07\x00"
-     "\xBC\xF5\x07\x00\xBD\xFD\x07\x00\xBE\xE9\x07\x00\xBF\xE9\x07\x00\xA8\x6D"
-     "\x06\x00"
-     "\xA9\x75\x06\x00\xAA\xA9\x07\x00\xAB\xA9\x07\x00\xAC\xB9\x07\x00\xAD\xB9"
-     "\x07\x00"
-     "\xAE\xA9\x07\x00\xAF\xA9\x07\x00\xBB\x9C\x00\x80\xBF\x9C\x00\x80\xC3\x9C"
-     "\x00\x80"
-     "\xC7\x9C\x00\x80\xCB\x9C\x00\x80\xCF\x9C\x00\x80\xD3\x9C\x00\x80\xD7\x9C"
-     "\x00\x80"
-     "\xB8\x69\x07\x00\xB9\x71\x07\x00\xBA\x75\x07\x00\xBB\x0D\x07\x00\xBC\x15"
-     "\x07\x00"
-     "\xBD\x1D\x07\x00\xBE\x09\x07\x00\xBF\x01\x07\x00\xB0\xD9\x07\x00\xB1\xD9"
-     "\x07\x00"
-     "\xB2\x6D\x07\x00\xB3\x65\x07\x00\xB4\x61\x07\x00\xB5\x61\x07\x00\xB6\x61"
-     "\x07\x00"
-     "\xB7\x5D\x07\x00\xA3\x21\x06\x00\xDB\x9C\x00\x80\xDF\x9C\x00\x80\xE3\x9C"
-     "\x00\x80"
-     "\xE7\x9C\x00\x80\xA6\x39\x06\x00\xA5\x31\x06\x00\xEB\x9C\x00\x80\xAB\x1D"
-     "\x06\x00"
-     "\xAA\x1D\x06\x00\xEF\x9C\x00\x80\xF3\x9C\x00\x80\xAF\xA5\x06\x00\xAE\xA5"
-     "\x06\x00"
-     "\xAD\xB1\x06\x00\xAC\xB9\x06\x00\xF7\x9C\x00\x80\xEF\x9C\x05\x00\xFB\x9C"
-     "\x00\x80"
-     "\xFF\x9C\x00\x80\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\x03\x9D"
-     "\x00\x80"
-     "\x07\x9D\x00\x80\x0B\x9D\x00\x80\x84\xA4\x01\x00\xBE\xA4\x01\x00\x0F\x9D"
-     "\x00\x80"
-     "\xE3\xA4\x05\x00\x13\x9D\x00\x80\xE1\xCC\x06\x00\xA8\x8D\x01\x00\xA9\x99"
-     "\x01\x00"
-     "\xAA\xD9\x01\x00\xAB\xD1\x01\x00\xAC\xF1\x01\x00\xAD\xF1\x01\x00\xAE\x39"
-     "\x01\x00"
-     "\xAF\x39\x01\x00\x86\x00\x00\x00\x87\x64\x01\x00\x17\x9D\x00\x80\x1B\x9D"
-     "\x00\x80"
-     "\x1F\x9D\x00\x80\x23\x9D\x00\x80\x27\x9D\x00\x80\x2B\x9D\x00\x80\xB8\xCD"
-     "\x00\x00"
-     "\xB9\xD5\x00\x00\xBA\xD5\x00\x00\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\x95"
-     "\x00\x00"
-     "\xBE\x91\x00\x00\xBF\x91\x00\x00\xB0\x49\x01\x00\xB1\x49\x01\x00\xB2\x59"
-     "\x01\x00"
-     "\xB3\x59\x01\x00\xB4\x49\x01\x00\xB5\x49\x01\x00\xB6\xFD\x00\x00\xB7\xF5"
-     "\x00\x00"
-     "\xB3\x51\x02\x00\x2F\x9D\x00\x80\x33\x9D\x00\x80\x37\x9D\x00\x80\x3B\x9D"
-     "\x00\x80"
-     "\xB6\x71\x02\x00\xB5\x79\x02\x00\x3F\x9D\x00\x80\xBB\x8D\x02\x00\xBA\x8D"
-     "\x02\x00"
-     "\x43\x9D\x00\x80\x47\x9D\x00\x80\xBF\x8D\x02\x00\xBE\x8D\x02\x00\xBD\x8D"
-     "\x02\x00"
-     "\xBC\x8D\x02\x00\x85\x6C\x12\x00\xA3\x15\x02\x00\x4B\x9D\x00\x80\x4F\x9D"
-     "\x00\x80"
-     "\xA6\x35\x02\x00\x53\x9D\x00\x80\x57\x9D\x00\x80\xA5\x3D\x02\x00\xAA\xC9"
-     "\x02\x00"
-     "\xAB\xC9\x02\x00\x5B\x9D\x00\x80\x5F\x9D\x00\x80\xAE\xC9\x02\x00\xAF\xC9"
-     "\x02\x00"
-     "\xAC\xC9\x02\x00\xAD\xC9\x02\x00\x84\x80\x03\x00\x63\x9D\x00\x80\x67\x9D"
-     "\x00\x80"
-     "\xBE\x00\x0C\x00\x6B\x9D\x00\x80\x6F\x9D\x00\x80\x73\x9D\x00\x80\x77\x9D"
-     "\x00\x80"
-     "\x7B\x9D\x00\x80\x83\x9D\x00\x80\x87\x9D\x00\x80\x8B\x9D\x00\x80\x8F\x9D"
-     "\x00\x80"
-     "\x93\x9D\x00\x80\x97\x9D\x00\x80\x9B\x9D\x00\x80\x80\x59\x00\x00\x81\x59"
-     "\x00\x00"
-     "\x82\x69\x00\x00\xEF\xB8\x02\x00\x9F\x9D\x00\x80\xA3\x9D\x00\x80\x87\x18"
-     "\x03\x00"
-     "\x86\xFC\x0C\x00\xA7\x9D\x00\x80\xAB\x9D\x00\x80\xAF\x9D\x00\x80\xB3\x9D"
-     "\x00\x80"
-     "\xB7\x9D\x00\x80\xE1\x4C\x01\x00\xBF\x9B\x00\x80\xE3\x6C\x02\x00\xA8\x79"
-     "\x02\x00"
-     "\xA9\xAD\x02\x00\xAA\xA5\x02\x00\xAB\xBD\x02\x00\xAC\xA5\x02\x00\xAD\xAD"
-     "\x02\x00"
-     "\xAE\xA5\x02\x00\xAF\x19\x01\x00\x84\xA8\x0D\x00\xBB\x9D\x00\x80\xBF\x9D"
-     "\x00\x80"
-     "\xC3\x9D\x00\x80\xC7\x9D\x00\x80\xCB\x9D\x00\x80\xCF\x9D\x00\x80\xD3\x9D"
-     "\x00\x80"
-     "\xB8\x15\x01\x00\xB9\x1D\x01\x00\xBA\x21\x01\x00\xBB\x21\x01\x00\xBC\xD5"
-     "\x01\x00"
-     "\xBD\xDD\x01\x00\xBE\xC9\x01\x00\xBF\xC9\x01\x00\xB0\x69\x01\x00\xB1\x69"
-     "\x01\x00"
-     "\xB2\x79\x01\x00\xB3\x71\x01\x00\xB4\x59\x01\x00\xB5\x59\x01\x00\xB6\x31"
-     "\x01\x00"
-     "\xB7\x31\x01\x00\xE1\x60\x07\x00\xD7\x9D\x00\x80\xE3\x14\x06\x00\xBE\x00"
-     "\x0C\x00"
-     "\xDB\x9D\x00\x80\xE1\xA4\x01\x00\xDF\x9D\x00\x80\xE3\xE4\x00\x00\xE3\x9D"
-     "\x00\x80"
-     "\xE7\x9D\x00\x80\xEB\x9D\x00\x80\xEF\x38\x00\x00\xF3\x9D\x00\x80\xF7\x9D"
-     "\x00\x80"
-     "\xEF\x6C\x07\x00\xFB\x9D\x00\x80\x80\x61\x00\x00\x81\x61\x00\x00\x82\x61"
-     "\x00\x00"
-     "\xB3\x61\x02\x00\xFF\x9D\x00\x80\xB5\x61\x02\x00\xB6\x61\x02\x00\x03\x9E"
-     "\x00\x80"
-     "\x86\xC0\x0C\x00\x87\xE8\x0C\x00\xBA\xE1\x01\x00\xBB\xE1\x01\x00\xBC\xE1"
-     "\x01\x00"
-     "\xBD\xE9\x01\x00\xBE\xD1\x01\x00\xBF\xD1\x01\x00\xA3\xAD\x0D\x00\x7F\x9D"
-     "\x00\x80"
-     "\xEF\x9D\x00\x80\x07\x9E\x00\x80\x0B\x9E\x00\x80\xA6\xAD\x0D\x00\xA5\xAD"
-     "\x0D\x00"
-     "\x0F\x9E\x00\x80\xAB\x2D\x0E\x00\xAA\x2D\x0E\x00\x13\x9E\x00\x80\x17\x9E"
-     "\x00\x80"
-     "\xAF\x1D\x0E\x00\xAE\x1D\x0E\x00\xAD\x25\x0E\x00\xAC\x2D\x0E\x00\x1B\x9E"
-     "\x00\x80"
-     "\xB3\xE9\x0E\x00\x1F\x9E\x00\x80\x23\x9E\x00\x80\xB6\xE9\x0E\x00\x27\x9E"
-     "\x00\x80"
-     "\x2B\x9E\x00\x80\xB5\xE9\x0E\x00\xBA\xB5\x0E\x00\xBB\x85\x0E\x00\x2F\x9E"
-     "\x00\x80"
-     "\x33\x9E\x00\x80\xBE\x9D\x0E\x00\xBF\x85\x0E\x00\xBC\x9D\x0E\x00\xBD\x95"
-     "\x0E\x00"
-     "\xA8\x25\x0E\x00\xA9\x2D\x0E\x00\xAA\x25\x0E\x00\xAB\x3D\x0E\x00\xAC\x25"
-     "\x0E\x00"
-     "\xAD\x2D\x0E\x00\xAE\x25\x0E\x00\xAF\x4D\x0E\x00\x37\x9E\x00\x80\x3B\x9E"
-     "\x00\x80"
-     "\x3F\x9E\x00\x80\x43\x9E\x00\x80\x47\x9E\x00\x80\x4B\x9E\x00\x80\x4F\x9E"
-     "\x00\x80"
-     "\x53\x9E\x00\x80\xB8\xED\x0F\x00\xB9\xF5\x0F\x00\xBA\xFD\x0F\x00\xBB\xF5"
-     "\x0F\x00"
-     "\xBC\xED\x0F\x00\xBD\x51\x0F\x00\xBE\x49\x0F\x00\xBF\x45\x0F\x00\xB0\x35"
-     "\x0E\x00"
-     "\xB1\x05\x0E\x00\xB2\x0D\x0E\x00\xB3\x05\x0E\x00\xB4\x1D\x0E\x00\xB5\x05"
-     "\x0E\x00"
-     "\xB6\xDD\x0F\x00\xB7\xD5\x0F\x00\xA3\xAD\x0F\x00\x82\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x80\xB1\x01\x00\x57\x9E\x00\x80\xA6\xAD\x0F\x00\xA5\xAD\x0F\x00\x5B\x9E"
-     "\x00\x80"
-     "\xAB\xC1\x0F\x00\xAA\xF1\x0F\x00\x84\xA0\x02\x00\xBE\xA0\x01\x00\xAF\xC1"
-     "\x0F\x00"
-     "\xAE\xD9\x0F\x00\xAD\xD1\x0F\x00\xAC\xD9\x0F\x00\x5F\x9E\x00\x80\xB3\xF9"
-     "\x0E\x00"
-     "\x86\xE8\x00\x00\x87\x2C\x01\x00\xB6\x5D\x01\x00\x63\x9E\x00\x80\x67\x9E"
-     "\x00\x80"
-     "\xB5\x81\x0E\x00\xBA\x49\x01\x00\xBB\x49\x01\x00\x6B\x9E\x00\x80\x6F\x9E"
-     "\x00\x80"
-     "\xBE\x3D\x01\x00\xBF\x21\x01\x00\xBC\x39\x01\x00\xBD\x35\x01\x00\xA8\x31"
-     "\x0E\x00"
-     "\xA9\x31\x0E\x00\xAA\x35\x0E\x00\xAB\x5D\x0E\x00\xAC\x4D\x0E\x00\xAD\x99"
-     "\x01\x00"
-     "\xAE\x89\x01\x00\xAF\x81\x01\x00\x84\xE4\x01\x00\x73\x9E\x00\x80\x77\x9E"
-     "\x00\x80"
-     "\x7B\x9E\x00\x80\x7F\x9E\x00\x80\x83\x9E\x00\x80\x87\x9E\x00\x80\x8B\x9E"
-     "\x00\x80"
-     "\xB8\xAD\x01\x00\xB9\x65\x01\x00\xBA\x6D\x01\x00\xBB\x65\x01\x00\xBC\x7D"
-     "\x01\x00"
-     "\xBD\x65\x01\x00\xBE\x6D\x01\x00\xBF\x65\x01\x00\xB0\xC9\x01\x00\xB1\xC9"
-     "\x01\x00"
-     "\xB2\xA9\x01\x00\xB3\xA5\x01\x00\xB4\xBD\x01\x00\xB5\xA1\x01\x00\xB6\xA1"
-     "\x01\x00"
-     "\xB7\x95\x01\x00\xA3\xBD\x0D\x00\x8F\x9E\x00\x80\x93\x9E\x00\x80\x97\x9E"
-     "\x00\x80"
-     "\x9B\x9E\x00\x80\xA6\x19\x02\x00\xA5\xC5\x0D\x00\x9F\x9E\x00\x80\xAB\x0D"
-     "\x02\x00"
-     "\xAA\x0D\x02\x00\xA3\x9E\x00\x80\xA7\x9E\x00\x80\xAF\x65\x02\x00\xAE\x79"
-     "\x02\x00"
-     "\xAD\x71\x02\x00\xAC\x7D\x02\x00\xAB\x9E\x00\x80\xAF\x9E\x00\x80\xB3\x9E"
-     "\x00\x80"
-     "\xB7\x9E\x00\x80\xBB\x9E\x00\x80\xBF\x9E\x00\x80\xC3\x9E\x00\x80\xC7\x9E"
-     "\x00\x80"
-     "\x80\x39\x00\x00\x81\x39\x00\x00\x82\x05\x00\x00\xCB\x9E\x00\x80\xBE\x3C"
-     "\x03\x00"
-     "\xD3\x9E\x00\x80\x87\x44\x03\x00\x86\xDC\x04\x00\xD7\x9E\x00\x80\xDB\x9E"
-     "\x00\x80"
-     "\xB3\xE1\x03\x00\x84\xC0\x04\x00\xB5\x35\x03\x00\xDF\x9E\x00\x80\xE3\x9E"
-     "\x00\x80"
-     "\xB6\x35\x03\x00\xE7\x9E\x00\x80\xEB\x9E\x00\x80\xBB\x11\x03\x00\xBA\x19"
-     "\x03\x00"
-     "\xBD\xB5\x00\x00\xBC\x09\x03\x00\xBF\xB5\x00\x00\xBE\xB5\x00\x00\xEF\x9E"
-     "\x00\x80"
-     "\xE1\x94\x01\x00\xF3\x9E\x00\x80\xE3\x7C\x01\x00\xF7\x9E\x00\x80\xFB\x9E"
-     "\x00\x80"
-     "\xFF\x9E\x00\x80\x03\x9F\x00\x80\x07\x9F\x00\x80\x0B\x9F\x00\x80\x0F\x9F"
-     "\x00\x80"
-     "\x13\x9F\x00\x80\x17\x9F\x00\x80\x1B\x9F\x00\x80\x1F\x9F\x00\x80\xEF\x74"
-     "\x02\x00"
-     "\xA8\x55\x02\x00\xA9\x5D\x02\x00\xAA\x61\x02\x00\xAB\x61\x02\x00\xAC\xB5"
-     "\x02\x00"
-     "\xAD\xBD\x02\x00\xAE\xB5\x02\x00\xAF\xAD\x02\x00\xBE\x68\x05\x00\x23\x9F"
-     "\x00\x80"
-     "\x27\x9F\x00\x80\x2B\x9F\x00\x80\x2F\x9F\x00\x80\x82\xAD\x00\x00\x81\xAD"
-     "\x00\x00"
-     "\x80\xAD\x00\x00\xB8\x69\x01\x00\xB9\x69\x01\x00\xBA\x09\x01\x00\xBB\x09"
-     "\x01\x00"
-     "\xBC\x19\x01\x00\xBD\x19\x01\x00\xBE\x09\x01\x00\xBF\x09\x01\x00\xB0\xD5"
-     "\x02\x00"
-     "\xB1\xDD\x02\x00\xB2\xD5\x02\x00\xB3\x69\x01\x00\xB4\x79\x01\x00\xB5\x79"
-     "\x01\x00"
-     "\xB6\x69\x01\x00\xB7\x61\x01\x00\xE1\x44\x0F\x00\xE3\xBC\x00\x00\xE3\xB8"
-     "\x0E\x00"
-     "\xE1\x74\x0E\x00\x33\x9F\x00\x80\x37\x9F\x00\x80\x3B\x9F\x00\x80\x43\x9F"
-     "\x00\x80"
-     "\x47\x9F\x00\x80\x4B\x9F\x00\x80\x4F\x9F\x00\x80\x53\x9F\x00\x80\xBE\x30"
-     "\x07\x00"
-     "\xEF\x60\x00\x00\xEF\xCC\x0E\x00\x57\x9F\x00\x80\xA3\xA9\x02\x00\x5B\x9F"
-     "\x00\x80"
-     "\x86\xE8\x04\x00\x87\x04\x05\x00\x5F\x9F\x00\x80\xA6\x7D\x02\x00\xA5\x7D"
-     "\x02\x00"
-     "\x63\x9F\x00\x80\xAB\x59\x02\x00\xAA\x51\x02\x00\x67\x9F\x00\x80\x6B\x9F"
-     "\x00\x80"
-     "\xAF\xFD\x01\x00\xAE\xFD\x01\x00\xAD\xFD\x01\x00\xAC\x41\x02\x00\xA8\x75"
-     "\x06\x00"
-     "\xA9\x85\x06\x00\xAA\x8D\x06\x00\xAB\x99\x06\x00\xAC\x89\x06\x00\xAD\xB5"
-     "\x06\x00"
-     "\xAE\xBD\x06\x00\xAF\xDD\x06\x00\x3F\x9F\x00\x80\x6F\x9F\x00\x80\x73\x9F"
-     "\x00\x80"
-     "\x77\x9F\x00\x80\x7B\x9F\x00\x80\x7F\x9F\x00\x80\x83\x9F\x00\x80\x87\x9F"
-     "\x00\x80"
-     "\xB8\xAD\x06\x00\xB9\xB5\x06\x00\xBA\xBD\x06\x00\xBB\xB5\x06\x00\xBC\x75"
-     "\x07\x00"
-     "\xBD\x7D\x07\x00\xBE\x75\x07\x00\xBF\x6D\x07\x00\xB0\xA9\x06\x00\xB1\xB5"
-     "\x06\x00"
-     "\xB2\xBD\x06\x00\xB3\xB5\x06\x00\xB4\xAD\x06\x00\xB5\x95\x06\x00\xB6\x9D"
-     "\x06\x00"
-     "\xB7\x95\x06\x00\xB3\x05\x06\x00\x8B\x9F\x00\x80\x8F\x9F\x00\x80\x93\x9F"
-     "\x00\x80"
-     "\x97\x9F\x00\x80\xB6\x35\x06\x00\xB5\x01\x06\x00\x9B\x9F\x00\x80\xBB\x71"
-     "\x06\x00"
-     "\xBA\x6D\x06\x00\x9F\x9F\x00\x80\xA3\x9F\x00\x80\xBF\x51\x06\x00\xBE\x69"
-     "\x06\x00"
-     "\xBD\x61\x06\x00\xBC\x69\x06\x00\x82\x19\x00\x00\xA3\x41\x06\x00\x80\x65"
-     "\x00\x00"
-     "\x81\x19\x00\x00\xA6\x71\x06\x00\xA7\x9F\x00\x80\xAB\x9F\x00\x80\xA5\x45"
-     "\x06\x00"
-     "\xAA\x29\x06\x00\xAB\x35\x06\x00\x84\x24\x03\x00\xAF\x9F\x00\x80\xAE\x2D"
-     "\x06\x00"
-     "\xAF\x15\x06\x00\xAC\x2D\x06\x00\xAD\x25\x06\x00\xB3\xA5\x06\x00\xB3\x9F"
-     "\x00\x80"
-     "\x86\x48\x01\x00\x87\x0C\x01\x00\xB7\x9F\x00\x80\xB6\xA9\x06\x00\xB5\xB5"
-     "\x06\x00"
-     "\xBB\x9F\x00\x80\xBB\xC9\x06\x00\xBA\xF9\x06\x00\xBF\x9F\x00\x80\xC3\x9F"
-     "\x00\x80"
-     "\xBF\x1D\x01\x00\xBE\x1D\x01\x00\xBD\x1D\x01\x00\xBC\xD1\x06\x00\xC7\x9F"
-     "\x00\x80"
-     "\xCB\x9F\x00\x80\xCF\x9F\x00\x80\xD3\x9F\x00\x80\xE3\xCC\x05\x00\xD7\x9F"
-     "\x00\x80"
-     "\xE1\x3C\x05\x00\xDB\x9F\x00\x80\xEF\xB8\x04\x00\xDF\x9F\x00\x80\xE3\x9F"
-     "\x00\x80"
-     "\xE7\x9F\x00\x80\xEB\x9F\x00\x80\xEF\x9F\x00\x80\xF3\x9F\x00\x80\xF7\x9F"
-     "\x00\x80"
-     "\xA8\x1D\x06\x00\xA9\x25\x06\x00\xAA\x2D\x06\x00\xAB\x25\x06\x00\xAC\x3D"
-     "\x06\x00"
-     "\xAD\x29\x06\x00\xAE\x5D\x06\x00\xAF\x4D\x06\x00\xFB\x9F\x00\x80\xFF\x9F"
-     "\x00\x80"
-     "\x03\xA0\x00\x80\x07\xA0\x00\x80\x0B\xA0\x00\x80\x0F\xA0\x00\x80\x13\xA0"
-     "\x00\x80"
-     "\x17\xA0\x00\x80\xB8\xF9\x01\x00\xB9\xF9\x01\x00\xBA\x4D\x01\x00\xBB\x45"
-     "\x01\x00"
-     "\xBC\x5D\x01\x00\xBD\x45\x01\x00\xBE\x45\x01\x00\xBF\x75\x01\x00\xB0\x3D"
-     "\x06\x00"
-     "\xB1\x0D\x06\x00\xB2\x05\x06\x00\xB3\xC9\x01\x00\xB4\xD9\x01\x00\xB5\xD9"
-     "\x01\x00"
-     "\xB6\xC9\x01\x00\xB7\xC9\x01\x00\x80\xB9\x03\x00\x81\xB9\x03\x00\x82\x19"
-     "\x00\x00"
-     "\xA3\xA9\x05\x00\x1B\xA0\x00\x80\xA5\xB9\x05\x00\xA6\xA5\x05\x00\x84\x84"
-     "\x02\x00"
-     "\x23\xA0\x00\x80\xBE\xA8\x03\x00\xAA\xF5\x05\x00\xAB\xC5\x05\x00\xAC\xDD"
-     "\x05\x00"
-     "\xAD\x11\x02\x00\xAE\x11\x02\x00\xAF\x11\x02\x00\xB6\x01\x03\x00\x27\xA0"
-     "\x00\x80"
-     "\x84\x44\x03\x00\xB5\x61\x03\x00\x2B\xA0\x00\x80\xB3\x49\x03\x00\x87\xD8"
-     "\x03\x00"
-     "\x86\xF8\x1C\x00\xBE\x09\x03\x00\xBF\x09\x03\x00\xBC\x35\x03\x00\xBD\x09"
-     "\x03\x00"
-     "\xBA\x3D\x03\x00\xBB\x21\x03\x00\x2F\xA0\x00\x80\x33\xA0\x00\x80\xB0\xCD"
-     "\x03\x00"
-     "\xB1\xD1\x03\x00\xB2\xD5\x03\x00\xB3\xE9\x03\x00\xB4\xF9\x03\x00\xB5\xF9"
-     "\x03\x00"
-     "\xB6\xE9\x03\x00\xB7\xE9\x03\x00\xB8\xD9\x03\x00\xB9\xD9\x03\x00\xBA\x09"
-     "\x03\x00"
-     "\xBB\x09\x03\x00\xBC\x19\x03\x00\xBD\x19\x03\x00\xBE\x09\x03\x00\xBF\x09"
-     "\x03\x00"
-     "\x37\xA0\x00\x80\x3B\xA0\x00\x80\xB3\x00\x00\x00\x3F\xA0\x00\x80\x43\xA0"
-     "\x00\x80"
-     "\x47\xA0\x00\x80\x4B\xA0\x00\x80\xCF\x9E\x00\x80\xA8\xE5\x03\x00\xA9\xF5"
-     "\x03\x00"
-     "\xAA\xF5\x03\x00\xAB\xC5\x03\x00\xAC\xDD\x03\x00\xAD\xB5\x03\x00\xAE\xBD"
-     "\x03\x00"
-     "\xAF\xB5\x03\x00\xA3\x1D\x02\x00\x4F\xA0\x00\x80\x53\xA0\x00\x80\x57\xA0"
-     "\x00\x80"
-     "\xBE\x1C\x1D\x00\xA6\x55\x02\x00\xA5\x35\x02\x00\x5B\xA0\x00\x80\xAB\x75"
-     "\x02\x00"
-     "\xAA\x69\x02\x00\x5F\xA0\x00\x80\x63\xA0\x00\x80\xAF\x5D\x02\x00\xAE\x5D"
-     "\x02\x00"
-     "\xAD\x5D\x02\x00\xAC\x61\x02\x00\x67\xA0\x00\x80\x6B\xA0\x00\x80\x6F\xA0"
-     "\x00\x80"
-     "\x73\xA0\x00\x80\xEF\x8C\x03\x00\x77\xA0\x00\x80\x7B\xA0\x00\x80\x7F\xA0"
-     "\x00\x80"
-     "\xE3\x6C\x03\x00\x83\xA0\x00\x80\xE1\x78\x00\x00\x87\xA0\x00\x80\x8B\xA0"
-     "\x00\x80"
-     "\x93\xA0\x00\x80\x86\x58\x1C\x00\x87\xF0\x1D\x00\x80\xE5\x00\x00\x81\xE9"
-     "\x00\x00"
-     "\x82\xF9\x00\x00\x97\xA0\x00\x80\x9B\xA0\x00\x80\x9F\xA0\x00\x80\xA3\xA0"
-     "\x00\x80"
-     "\xEF\xD4\x06\x00\xA7\xA0\x00\x80\xE1\x74\x06\x00\xAB\xA0\x00\x80\xE3\x54"
-     "\x01\x00"
-     "\xAF\xA0\x00\x80\xB3\xA0\x00\x80\xB7\xA0\x00\x80\xBB\xA0\x00\x80\xB3\x91"
-     "\x01\x00"
-     "\xBF\xA0\x00\x80\xC3\xA0\x00\x80\x84\xC4\x1D\x00\xC7\xA0\x00\x80\xB6\x15"
-     "\x01\x00"
-     "\xB5\xA5\x01\x00\xCB\xA0\x00\x80\xBB\x35\x01\x00\xBA\x35\x01\x00\xCF\xA0"
-     "\x00\x80"
-     "\xD3\xA0\x00\x80\xBF\xFD\x00\x00\xBE\xFD\x00\x00\xBD\xFD\x00\x00\xBC\x25"
-     "\x01\x00"
-     "\xA9\x45\x1E\x00\xA8\x25\x1E\x00\xAB\x71\x1E\x00\xAA\x5D\x1E\x00\xAD\xD1"
-     "\x1E\x00"
-     "\xAC\x69\x1E\x00\xAF\xD5\x1E\x00\xAE\xD9\x1E\x00\x8F\xA0\x00\x80\xD7\xA0"
-     "\x00\x80"
-     "\xDB\xA0\x00\x80\xDF\xA0\x00\x80\xE3\xA0\x00\x80\xE7\xA0\x00\x80\xEB\xA0"
-     "\x00\x80"
-     "\xEF\xA0\x00\x80\xB9\x01\x1F\x00\xB8\x69\x1F\x00\xBB\x01\x1F\x00\xBA\x09"
-     "\x1F\x00"
-     "\xBD\x05\x1F\x00\xBC\x19\x1F\x00\xBF\x05\x1F\x00\xBE\x05\x1F\x00\xB1\xB5"
-     "\x1E\x00"
-     "\xB0\xB5\x1E\x00\xB3\x6D\x1F\x00\xB2\xB5\x1E\x00\xB5\x7D\x1F\x00\xB4\x7D"
-     "\x1F\x00"
-     "\xB7\x51\x1F\x00\xB6\x61\x1F\x00\xA3\xD1\x1E\x00\x82\x2D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\xF3\xA0\x00\x80\xA6\x55\x1E\x00\xA5\xE5\x1E\x00\xF7\xA0"
-     "\x00\x80"
-     "\xAB\x75\x1E\x00\xAA\x75\x1E\x00\xFB\xA0\x00\x80\xBE\x60\x03\x00\xAF\xBD"
-     "\x1F\x00"
-     "\xAE\xBD\x1F\x00\xAD\xBD\x1F\x00\xAC\x65\x1E\x00\xFF\xA0\x00\x80\xB3\x55"
-     "\x1F\x00"
-     "\x86\x68\x00\x00\x87\x6C\x03\x00\xB6\x71\x1F\x00\x03\xA1\x00\x80\x07\xA1"
-     "\x00\x80"
-     "\xB5\x45\x1F\x00\xBA\x55\x1F\x00\xBB\x5D\x1F\x00\x0B\xA1\x00\x80\x0F\xA1"
-     "\x00\x80"
-     "\xBE\x39\x1F\x00\xBF\x25\x1F\x00\xBC\x45\x1F\x00\xBD\x39\x1F\x00\x13\xA1"
-     "\x00\x80"
-     "\x17\xA1\x00\x80\x1B\xA1\x00\x80\x1F\xA1\x00\x80\x23\xA1\x00\x80\x27\xA1"
-     "\x00\x80"
-     "\x2B\xA1\x00\x80\xEF\xF4\x1C\x00\x2F\xA1\x00\x80\xE1\x60\x1D\x00\x33\xA1"
-     "\x00\x80"
-     "\xE3\x74\x1D\x00\x37\xA1\x00\x80\x3B\xA1\x00\x80\x3F\xA1\x00\x80\x43\xA1"
-     "\x00\x80"
-     "\xA3\xD1\x1E\x00\x47\xA1\x00\x80\x4B\xA1\x00\x80\x4F\xA1\x00\x80\x53\xA1"
-     "\x00\x80"
-     "\xA6\xF5\x1E\x00\xA5\xC1\x1E\x00\x57\xA1\x00\x80\xAB\xD9\x1E\x00\xAA\xD1"
-     "\x1E\x00"
-     "\x5B\xA1\x00\x80\x5F\xA1\x00\x80\xAF\xA1\x1E\x00\xAE\xBD\x1E\x00\xAD\xBD"
-     "\x1E\x00"
-     "\xAC\xC1\x1E\x00\xA8\xF5\x1E\x00\xA9\x09\x1E\x00\xAA\x19\x1E\x00\xAB\x19"
-     "\x1E\x00"
-     "\xAC\x09\x1E\x00\xAD\x35\x1E\x00\xAE\x31\x1E\x00\xAF\x2D\x1E\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x1D\x00\x00\x63\xA1\x00\x80\x67\xA1\x00\x80\x6B\xA1"
-     "\x00\x80"
-     "\x84\xFC\x01\x00\xBE\xF0\x01\x00\xB8\x2D\x1E\x00\xB9\xE1\x01\x00\xBA\xE1"
-     "\x01\x00"
-     "\xBB\xE1\x01\x00\xBC\xE1\x01\x00\xBD\xE1\x01\x00\xBE\xE1\x01\x00\xBF\xE1"
-     "\x01\x00"
-     "\xB0\x55\x1E\x00\xB1\x5D\x1E\x00\xB2\x55\x1E\x00\xB3\x29\x1E\x00\xB4\x39"
-     "\x1E\x00"
-     "\xB5\x25\x1E\x00\xB6\x25\x1E\x00\xB7\x15\x1E\x00\xA8\xD5\x02\x00\xA9\xDD"
-     "\x02\x00"
-     "\xAA\xE1\x02\x00\xAB\xE1\x02\x00\xAC\x11\x03\x00\xAD\x11\x03\x00\xAE\x11"
-     "\x03\x00"
-     "\xAF\x11\x03\x00\x73\xA1\x00\x80\x77\xA1\x00\x80\x7B\xA1\x00\x80\x7F\xA1"
-     "\x00\x80"
-     "\x83\xA1\x00\x80\x87\xA1\x00\x80\x8B\xA1\x00\x80\x8F\xA1\x00\x80\xB8\x7D"
-     "\x03\x00"
-     "\xB9\x01\x03\x00\xBA\x01\x03\x00\xBB\x01\x03\x00\xBC\x01\x03\x00\xBD\x09"
-     "\x03\x00"
-     "\xBE\x31\x03\x00\xBF\x31\x03\x00\xB0\x71\x03\x00\xB1\x71\x03\x00\xB2\x71"
-     "\x03\x00"
-     "\xB3\x4D\x03\x00\xB4\x59\x03\x00\xB5\x45\x03\x00\xB6\x4D\x03\x00\xB7\x45"
-     "\x03\x00"
-     "\x86\x20\x04\x00\x87\xA0\x03\x00\xB3\x61\x02\x00\xBE\x80\x04\x00\xB5\x91"
-     "\x03\x00"
-     "\x93\xA1\x00\x80\x97\xA1\x00\x80\xB6\x99\x03\x00\x1F\xA0\x00\x80\x84\xB8"
-     "\x03\x00"
-     "\xBB\xBD\x03\x00\xBA\xBD\x03\x00\xBD\x91\x03\x00\xBC\xAD\x03\x00\xBF\x89"
-     "\x03\x00"
-     "\xBE\x95\x03\x00\xA3\x29\x02\x00\x9B\xA1\x00\x80\x9F\xA1\x00\x80\xA3\xA1"
-     "\x00\x80"
-     "\x85\x94\x06\x00\xA6\xD1\x03\x00\xA5\xD9\x03\x00\xA7\xA1\x00\x80\xAB\xF5"
-     "\x03\x00"
-     "\xAA\xF5\x03\x00\xAB\xA1\x00\x80\xAF\xA1\x00\x80\xAF\xC1\x03\x00\xAE\xDD"
-     "\x03\x00"
-     "\xAD\xD9\x03\x00\xAC\xE5\x03\x00\xB3\xA1\x00\x80\xB7\xA1\x00\x80\xBB\xA1"
-     "\x00\x80"
-     "\xBF\xA1\x00\x80\xC3\xA1\x00\x80\xC7\xA1\x00\x80\xCB\xA1\x00\x80\xCF\xA1"
-     "\x00\x80"
-     "\x81\x11\x00\x00\x80\x0D\x00\x00\xD3\xA1\x00\x80\x82\x19\x00\x00\xD7\xA1"
-     "\x00\x80"
-     "\xDB\xA1\x00\x80\xE3\xA1\x00\x80\xE7\xA1\x00\x80\xEB\xA1\x00\x80\xE1\x94"
-     "\x01\x00"
-     "\xEF\xA1\x00\x80\xE3\xE4\x00\x00\xF3\xA1\x00\x80\xF7\xA1\x00\x80\x86\x78"
-     "\x04\x00"
-     "\x87\x80\x02\x00\xFB\xA1\x00\x80\xFF\xA1\x00\x80\x03\xA2\x00\x80\x07\xA2"
-     "\x00\x80"
-     "\x0B\xA2\x00\x80\x0F\xA2\x00\x80\x13\xA2\x00\x80\xEF\x38\x00\x00\x17\xA2"
-     "\x00\x80"
-     "\x1B\xA2\x00\x80\x1F\xA2\x00\x80\x23\xA2\x00\x80\x27\xA2\x00\x80\x2B\xA2"
-     "\x00\x80"
-     "\x2F\xA2\x00\x80\xEF\xC8\x01\x00\xBE\x6C\x04\x00\xE1\xC4\x1E\x00\x84\xA0"
-     "\x05\x00"
-     "\xE3\x70\x1E\x00\x33\xA2\x00\x80\x37\xA2\x00\x80\x3B\xA2\x00\x80\x3F\xA2"
-     "\x00\x80"
-     "\x43\xA2\x00\x80\xB6\xF1\x01\x00\xB5\xF9\x01\x00\x85\xF4\x20\x00\xB3\xF9"
-     "\x01\x00"
-     "\x82\x29\x00\x00\x81\x21\x00\x00\x80\x1D\x00\x00\xBF\x3D\x01\x00\xBE\x29"
-     "\x01\x00"
-     "\xBD\x21\x01\x00\xBC\x29\x01\x00\xBB\x69\x01\x00\xBA\x61\x01\x00\x47\xA2"
-     "\x00\x80"
-     "\x4B\xA2\x00\x80\xA8\x21\x06\x00\xA9\x21\x06\x00\xAA\x21\x06\x00\xAB\x21"
-     "\x06\x00"
-     "\xAC\x21\x06\x00\xAD\x21\x06\x00\xAE\x21\x06\x00\xAF\x21\x06\x00\xDF\xA1"
-     "\x00\x80"
-     "\xBF\x2C\x01\x00\x86\x00\x0C\x00\x87\x60\x03\x00\x4F\xA2\x00\x80\x53\xA2"
-     "\x00\x80"
-     "\x57\xA2\x00\x80\x5B\xA2\x00\x80\xB8\x41\x07\x00\xB9\x41\x07\x00\xBA\x41"
-     "\x07\x00"
-     "\xBB\x41\x07\x00\xBC\x41\x07\x00\xBD\x41\x07\x00\xBE\x41\x07\x00\xBF\x41"
-     "\x07\x00"
-     "\xB0\xF5\x06\x00\xB1\xFD\x06\x00\xB2\xAD\x06\x00\xB3\xA5\x06\x00\xB4\xBD"
-     "\x06\x00"
-     "\xB5\xA9\x06\x00\xB6\x91\x06\x00\xB7\x91\x06\x00\xA3\xA9\x06\x00\x5F\xA2"
-     "\x00\x80"
-     "\x63\xA2\x00\x80\x67\xA2\x00\x80\x6B\xA2\x00\x80\xA6\xA1\x06\x00\xA5\xA9"
-     "\x06\x00"
-     "\x6F\xA2\x00\x80\xAB\x39\x06\x00\xAA\x31\x06\x00\x73\xA2\x00\x80\x77\xA2"
-     "\x00\x80"
-     "\xAF\x6D\x06\x00\xAE\x79\x06\x00\xAD\x71\x06\x00\xAC\x79\x06\x00\x7B\xA2"
-     "\x00\x80"
-     "\xB3\x85\x07\x00\x7F\xA2\x00\x80\x83\xA2\x00\x80\xB6\x85\x07\x00\x87\xA2"
-     "\x00\x80"
-     "\x8B\xA2\x00\x80\xB5\x8D\x07\x00\xBA\x89\x07\x00\xBB\x89\x07\x00\x8F\xA2"
-     "\x00\x80"
-     "\x93\xA2\x00\x80\xBE\x89\x07\x00\xBF\x75\x07\x00\xBC\x89\x07\x00\xBD\x89"
-     "\x07\x00"
-     "\x97\xA2\x00\x80\xE1\xC0\x05\x00\xEF\x7C\x05\x00\xE3\xD4\x04\x00\x9B\xA2"
-     "\x00\x80"
-     "\x9F\xA2\x00\x80\xA3\xA2\x00\x80\xA7\xA2\x00\x80\xAB\xA2\x00\x80\xAF\xA2"
-     "\x00\x80"
-     "\xB3\xA2\x00\x80\xB7\xA2\x00\x80\xE1\xA8\x05\x00\xBB\xA2\x00\x80\xE3\x8C"
-     "\x05\x00"
-     "\xEF\xF8\x0C\x00\xA3\x01\x06\x00\x82\x71\x00\x00\x81\x65\x00\x00\x80\x65"
-     "\x00\x00"
-     "\x84\x80\x01\x00\xA6\x01\x06\x00\xA5\x09\x06\x00\xBE\x10\x00\x00\xAB\x0D"
-     "\x06\x00"
-     "\xAA\x0D\x06\x00\x86\xE8\x0F\x00\x87\x18\x00\x00\xAF\xF1\x06\x00\xAE\x0D"
-     "\x06\x00"
-     "\xAD\x0D\x06\x00\xAC\x0D\x06\x00\xA8\xB9\x06\x00\xA9\xB9\x06\x00\xAA\xC9"
-     "\x06\x00"
-     "\xAB\xC5\x06\x00\xAC\xC1\x06\x00\xAD\xC1\x06\x00\xAE\xF9\x06\x00\xAF\xF9"
-     "\x06\x00"
-     "\xC3\xA2\x00\x80\xC7\xA2\x00\x80\xCB\xA2\x00\x80\xCF\xA2\x00\x80\xD3\xA2"
-     "\x00\x80"
-     "\xD7\xA2\x00\x80\xDB\xA2\x00\x80\xDF\xA2\x00\x80\xB8\x4D\x06\x00\xB9\x55"
-     "\x06\x00"
-     "\xBA\x5D\x06\x00\xBB\x55\x06\x00\xBC\x61\x06\x00\xBD\x61\x06\x00\xBE\x61"
-     "\x06\x00"
-     "\xBF\x61\x06\x00\xB0\x89\x06\x00\xB1\x89\x06\x00\xB2\x99\x06\x00\xB3\x99"
-     "\x06\x00"
-     "\xB4\x89\x06\x00\xB5\x75\x06\x00\xB6\x7D\x06\x00\xB7\x75\x06\x00\xE3\xA2"
-     "\x00\x80"
-     "\xB5\x8D\x02\x00\xB6\x85\x02\x00\x84\x40\x0C\x00\x8F\xF1\x68\x00\xE7\xA2"
-     "\x00\x80"
-     "\xEB\xA2\x00\x80\xB3\x21\x02\x00\xBC\xE1\x02\x00\xBD\xED\x02\x00\xBE\xE5"
-     "\x02\x00"
-     "\xBF\xF9\x02\x00\x85\x34\x26\x00\xEF\xA2\x00\x80\xBA\x91\x02\x00\xBB\x91"
-     "\x02\x00"
-     "\x9B\x01\x1F\x00\x9A\x4D\x1E\x00\x99\x4D\x1E\x00\x98\x3D\x13\x00\x9F\xF1"
-     "\x04\x00"
-     "\x9E\xB9\x1A\x00\x9D\x79\x1A\x00\x9C\x81\x18\x00\x93\x41\x16\x00\x92\x91"
-     "\x14\x00"
-     "\x91\xF9\x6A\x00\x90\x4D\x6A\x00\x97\x61\x12\x00\x96\x71\x12\x00\x95\x81"
-     "\x10\x00"
-     "\x94\x41\x16\x00\xB1\x01\x38\x00\xB0\xE1\x3E\x00\xB3\x81\x3A\x00\xB2\x89"
-     "\x3A\x00"
-     "\xB5\x91\x26\x00\xB4\xD1\x3B\x00\xF3\xA2\x00\x80\xF7\xA2\x00\x80\xA3\x81"
-     "\x02\x00"
-     "\xA2\x11\x00\x00\xA1\x69\x06\x00\xA0\x8D\x06\x00\xFB\xA2\x00\x80\x6F\xA1"
-     "\x00\x80"
-     "\xFF\xA2\x00\x80\x03\xA3\x00\x80\xA1\xF1\x02\x00\x07\xA3\x00\x80\xA3\xCD"
-     "\x0E\x00"
-     "\xA2\xCD\x0E\x00\xA5\x01\x08\x00\xA4\x51\x0F\x00\xA7\xC5\x0B\x00\xA6\xFD"
-     "\x0A\x00"
-     "\xA9\xC1\x36\x00\xA8\x01\x34\x00\xAB\x01\x30\x00\xAA\x41\x36\x00\xAD\xCD"
-     "\x32\x00"
-     "\xAC\xD5\x32\x00\xAF\xFD\x3E\x00\xAE\x01\x3C\x00\x82\xE1\x69\x00\x83\x25"
-     "\x6B\x00"
-     "\x0B\xA3\x00\x80\x0F\xA3\x00\x80\x86\x31\x17\x00\x87\x29\x17\x00\x84\x89"
-     "\x6B\x00"
-     "\x85\xE9\x15\x00\x8A\xD9\x11\x00\x8B\x21\x13\x00\x13\xA3\x00\x80\x17\xA3"
-     "\x00\x80"
-     "\x8E\x35\x1F\x00\x8F\x3D\x1F\x00\x8C\x29\x13\x00\x8D\x4D\x12\x00\x92\x01"
-     "\x1E\x00"
-     "\x93\x99\x19\x00\x1B\xA3\x00\x80\x84\x00\x0C\x00\x96\xF1\x05\x00\x97\x75"
-     "\x07\x00"
-     "\x94\x79\x1B\x00\x95\xB1\x1B\x00\x9A\xA1\x07\x00\x9B\xD1\x01\x00\x1F\xA3"
-     "\x00\x80"
-     "\x23\xA3\x00\x80\x81\xF1\x00\x00\x80\x19\x00\x00\x9C\x59\x03\x00\x82\xF9"
-     "\x00\x00"
-     "\x27\xA3\x00\x80\x2B\xA3\x00\x80\xBE\xC0\x0D\x00\xEF\x28\x00\x00\x86\x30"
-     "\x0C\x00"
-     "\x87\x5C\x0D\x00\x33\xA3\x00\x80\xEF\xC4\x06\x00\x37\xA3\x00\x80\xE1\xE4"
-     "\x06\x00"
-     "\x3B\xA3\x00\x80\xE3\x14\x06\x00\x3F\xA3\x00\x80\xE1\x84\x01\x00\x43\xA3"
-     "\x00\x80"
-     "\xE3\xFC\x00\x00\xA3\xFD\x03\x00\x47\xA3\x00\x80\x4B\xA3\x00\x80\x84\x4C"
-     "\x0F\x00"
-     "\x4F\xA3\x00\x80\xA6\x59\x03\x00\xA5\x51\x03\x00\x53\xA3\x00\x80\xAB\x4D"
-     "\x03\x00"
-     "\xAA\x4D\x03\x00\x57\xA3\x00\x80\x5B\xA3\x00\x80\xAF\x25\x03\x00\xAE\x39"
-     "\x03\x00"
-     "\xAD\x31\x03\x00\xAC\x3D\x03\x00\xBF\xA2\x00\x80\x2F\xA3\x00\x80\x5F\xA3"
-     "\x00\x80"
-     "\x63\xA3\x00\x80\x67\xA3\x00\x80\x6B\xA3\x00\x80\x6F\xA3\x00\x80\x73\xA3"
-     "\x00\x80"
-     "\xA8\x81\x0D\x00\xA9\x81\x0D\x00\xAA\x81\x0D\x00\xAB\x81\x0D\x00\xAC\x81"
-     "\x0D\x00"
-     "\xAD\x81\x0D\x00\xAE\x81\x0D\x00\xAF\x81\x0D\x00\xB0\x41\x0E\x00\xB1\x51"
-     "\x0E\x00"
-     "\xB2\x51\x0E\x00\xB3\x79\x0E\x00\xB4\xC9\x0E\x00\xB5\xC9\x0E\x00\xB6\xE9"
-     "\x0E\x00"
-     "\xB7\xE5\x0E\x00\xB8\xDD\x0E\x00\xB9\xA1\x0E\x00\xBA\x61\x0F\x00\xBB\x61"
-     "\x0F\x00"
-     "\xBC\x59\x0F\x00\xBD\x59\x0F\x00\xBE\x31\x0F\x00\xBF\x15\x0F\x00\xB3\x29"
-     "\x0D\x00"
-     "\x77\xA3\x00\x80\x7B\xA3\x00\x80\x7F\xA3\x00\x80\x83\xA3\x00\x80\xB6\x21"
-     "\x0D\x00"
-     "\xB5\x29\x0D\x00\x87\xA3\x00\x80\xBB\xB5\x0E\x00\xBA\xA5\x0E\x00\x8B\xA3"
-     "\x00\x80"
-     "\x8F\xA3\x00\x80\xBF\x2D\x0E\x00\xBE\x2D\x0E\x00\xBD\x8D\x0E\x00\xBC\xA5"
-     "\x0E\x00"
-     "\x82\x29\x00\x00\xA3\x6D\x0D\x00\x80\x1D\x00\x00\x81\x15\x00\x00\xA6\x65"
-     "\x0D\x00"
-     "\x93\xA3\x00\x80\x97\xA3\x00\x80\xA5\x6D\x0D\x00\xAA\xE1\x0E\x00\xAB\xF1"
-     "\x0E\x00"
-     "\x9B\xA3\x00\x80\xBE\xA4\x03\x00\xAE\x69\x0E\x00\xAF\x69\x0E\x00\xAC\xE1"
-     "\x0E\x00"
-     "\xAD\xC9\x0E\x00\xA8\x99\x0E\x00\xA9\xA5\x0E\x00\xAA\xA1\x0E\x00\xAB\xBD"
-     "\x0E\x00"
-     "\xAC\xA9\x0E\x00\xAD\xD1\x0E\x00\xAE\xD1\x0E\x00\xAF\xCD\x0E\x00\x86\x80"
-     "\x00\x00"
-     "\x87\x04\x01\x00\xA3\xA3\x00\x80\xA7\xA3\x00\x80\xAB\xA3\x00\x80\xAF\xA3"
-     "\x00\x80"
-     "\xB3\xA3\x00\x80\xB7\xA3\x00\x80\xB8\x5D\x01\x00\xB9\xC1\x01\x00\xBA\xC1"
-     "\x01\x00"
-     "\xBB\xC1\x01\x00\xBC\xC1\x01\x00\xBD\xC9\x01\x00\xBE\xF1\x01\x00\xBF\xF1"
-     "\x01\x00"
-     "\xB0\xBD\x0E\x00\xB1\x8D\x0E\x00\xB2\x85\x0E\x00\xB3\x65\x01\x00\xB4\x7D"
-     "\x01\x00"
-     "\xB5\x65\x01\x00\xB6\x6D\x01\x00\xB7\x65\x01\x00\xB3\x29\x0E\x00\x9F\xA3"
-     "\x00\x80"
-     "\xBB\xA3\x00\x80\xBF\xA3\x00\x80\xC3\xA3\x00\x80\xB6\x51\x0E\x00\xB5\x21"
-     "\x0E\x00"
-     "\xC7\xA3\x00\x80\xBB\x49\x0E\x00\xBA\x79\x0E\x00\xCB\xA3\x00\x80\xCF\xA3"
-     "\x00\x80"
-     "\xBF\xA1\x01\x00\xBE\xB9\x01\x00\xBD\xB1\x01\x00\xBC\x51\x0E\x00\xD3\xA3"
-     "\x00\x80"
-     "\xA3\x6D\x0E\x00\xD7\xA3\x00\x80\xDB\xA3\x00\x80\xA6\x15\x0E\x00\xDF\xA3"
-     "\x00\x80"
-     "\xE3\xA3\x00\x80\xA5\x65\x0E\x00\xAA\x3D\x0E\x00\xAB\x0D\x0E\x00\xE7\xA3"
-     "\x00\x80"
-     "\xEB\xA3\x00\x80\xAE\xFD\x01\x00\xAF\xE5\x01\x00\xAC\x15\x0E\x00\xAD\xF5"
-     "\x01\x00"
-     "\xE1\x38\x01\x00\xEF\xA3\x00\x80\xE3\xD4\x0F\x00\xF3\xA3\x00\x80\xF7\xA3"
-     "\x00\x80"
-     "\xBE\x58\x03\x00\xFB\xA3\x00\x80\xFF\xA3\x00\x80\x80\x0D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x1D\x00\x00\x03\xA4\x00\x80\x07\xA4\x00\x80\x0B\xA4\x00\x80\xEF\xD4"
-     "\x02\x00"
-     "\x13\xA4\x00\x80\x86\xE0\x04\x00\x87\xE0\x03\x00\xBE\x00\x04\x00\xB3\x7D"
-     "\x03\x00"
-     "\x17\xA4\x00\x80\xB5\x6D\x03\x00\xB6\x65\x03\x00\x1B\xA4\x00\x80\x1F\xA4"
-     "\x00\x80"
-     "\x23\xA4\x00\x80\xBA\xF9\x03\x00\xBB\xF1\x03\x00\xBC\xD9\x03\x00\xBD\xD9"
-     "\x03\x00"
-     "\xBE\xB9\x03\x00\xBF\xB5\x03\x00\xAA\xBD\x0E\x00\xAB\x01\x0E\x00\xA8\xBD"
-     "\x0E\x00"
-     "\xA9\xB5\x0E\x00\xAE\x01\x0E\x00\xAF\x09\x0E\x00\xAC\x11\x0E\x00\xAD\x11"
-     "\x0E\x00"
-     "\x27\xA4\x00\x80\x2B\xA4\x00\x80\x2F\xA4\x00\x80\x33\xA4\x00\x80\x37\xA4"
-     "\x00\x80"
-     "\x3B\xA4\x00\x80\x3F\xA4\x00\x80\x43\xA4\x00\x80\xBA\xD1\x0F\x00\xBB\xD1"
-     "\x0F\x00"
-     "\xB8\x25\x0E\x00\xB9\xCD\x0F\x00\xBE\xF1\x0F\x00\xBF\xF1\x0F\x00\xBC\xC1"
-     "\x0F\x00"
-     "\xBD\xC1\x0F\x00\xB2\x21\x0E\x00\xB3\x2D\x0E\x00\xB0\x41\x0E\x00\xB1\x41"
-     "\x0E\x00"
-     "\xB6\x29\x0E\x00\xB7\x1D\x0E\x00\xB4\x35\x0E\x00\xB5\x29\x0E\x00\xA8\x29"
-     "\x02\x00"
-     "\xA9\x29\x02\x00\xAA\x39\x02\x00\xAB\x39\x02\x00\xAC\x29\x02\x00\xAD\x29"
-     "\x02\x00"
-     "\xAE\x91\x02\x00\xAF\x91\x02\x00\x47\xA4\x00\x80\x4B\xA4\x00\x80\x4F\xA4"
-     "\x00\x80"
-     "\x53\xA4\x00\x80\x57\xA4\x00\x80\x5B\xA4\x00\x80\xBF\x30\x2B\x00\x5F\xA4"
-     "\x00\x80"
-     "\xB8\x99\x02\x00\xB9\xAD\x02\x00\xBA\xA5\x02\x00\xBB\x49\x01\x00\xBC\x5D"
-     "\x01\x00"
-     "\xBD\x45\x01\x00\xBE\x41\x01\x00\xBF\x41\x01\x00\xB0\xF1\x02\x00\xB1\xF9"
-     "\x02\x00"
-     "\xB2\xC1\x02\x00\xB3\xC1\x02\x00\xB4\xB1\x02\x00\xB5\xBD\x02\x00\xB6\xB5"
-     "\x02\x00"
-     "\xB7\xA9\x02\x00\xE1\x08\x0E\x00\xE1\xF4\x0C\x00\xE3\xD4\x0E\x00\xE3\x90"
-     "\x0F\x00"
-     "\x63\xA4\x00\x80\x67\xA4\x00\x80\x6B\xA4\x00\x80\x6F\xA4\x00\x80\x81\x11"
-     "\x00\x00"
-     "\x80\x0D\x00\x00\x73\xA4\x00\x80\x82\x19\x00\x00\x77\xA4\x00\x80\x7B\xA4"
-     "\x00\x80"
-     "\xEF\xAC\x01\x00\xEF\xCC\x0E\x00\xA3\xF1\x02\x00\x83\xA4\x00\x80\x87\xA4"
-     "\x00\x80"
-     "\x84\x80\x05\x00\x8B\xA4\x00\x80\xA6\xE9\x02\x00\xA5\xE1\x02\x00\x8F\xA4"
-     "\x00\x80"
-     "\xAB\x7D\x02\x00\xAA\x75\x02\x00\x86\xC8\x04\x00\x87\xEC\x04\x00\xAF\x39"
-     "\x02\x00"
-     "\xAE\x35\x02\x00\xAD\x55\x02\x00\xAC\x55\x02\x00\xA8\x3D\x06\x00\xA9\x55"
-     "\x06\x00"
-     "\xAA\x55\x06\x00\xAB\x6D\x06\x00\xAC\x75\x06\x00\xAD\xB5\x06\x00\xAE\xBD"
-     "\x06\x00"
-     "\xAF\xB5\x06\x00\x7F\xA4\x00\x80\x93\xA4\x00\x80\x97\xA4\x00\x80\x9B\xA4"
-     "\x00\x80"
-     "\x9F\xA4\x00\x80\xA3\xA4\x00\x80\xA7\xA4\x00\x80\xAB\xA4\x00\x80\xB8\x6D"
-     "\x07\x00"
-     "\xB9\x75\x07\x00\xBA\x7D\x07\x00\xBB\x71\x07\x00\xBC\xC1\x07\x00\xBD\xC1"
-     "\x07\x00"
-     "\xBE\xC1\x07\x00\xBF\xC1\x07\x00\xB0\xCD\x06\x00\xB1\xD5\x06\x00\xB2\xDD"
-     "\x06\x00"
-     "\xB3\xD5\x06\x00\xB4\xCD\x06\x00\xB5\x55\x07\x00\xB6\x5D\x07\x00\xB7\x55"
-     "\x07\x00"
-     "\xB3\xDD\x06\x00\xAF\xA4\x00\x80\xB3\xA4\x00\x80\xB7\xA4\x00\x80\xBB\xA4"
-     "\x00\x80"
-     "\xB6\x35\x06\x00\xB5\xFD\x06\x00\xBF\xA4\x00\x80\xBB\x11\x06\x00\xBA\x09"
-     "\x06\x00"
-     "\xC3\xA4\x00\x80\xC7\xA4\x00\x80\xBF\x91\x07\x00\xBE\x09\x06\x00\xBD\x01"
-     "\x06\x00"
-     "\xBC\x09\x06\x00\xCB\xA4\x00\x80\xA3\x99\x06\x00\xCF\xA4\x00\x80\xD3\xA4"
-     "\x00\x80"
-     "\xA6\x71\x06\x00\xD7\xA4\x00\x80\xDB\xA4\x00\x80\xA5\xB9\x06\x00\xAA\x4D"
-     "\x06\x00"
-     "\xAB\x55\x06\x00\xDF\xA4\x00\x80\xE3\xA4\x00\x80\xAE\x4D\x06\x00\xAF\xD5"
-     "\x07\x00"
-     "\xAC\x4D\x06\x00\xAD\x45\x06\x00\x80\xED\x01\x00\x81\x09\x00\x00\x82\x19"
-     "\x00\x00"
-     "\xE7\xA4\x00\x80\xEB\xA4\x00\x80\x84\xD0\x01\x00\xBE\xC4\x01\x00\xEF\xA4"
-     "\x00\x80"
-     "\x86\x80\x00\x00\x87\x20\x01\x00\xF3\xA4\x00\x80\xF7\xA4\x00\x80\xFB\xA4"
-     "\x00\x80"
-     "\xFF\xA4\x00\x80\x03\xA5\x00\x80\x07\xA5\x00\x80\x0B\xA5\x00\x80\x0F\xA5"
-     "\x00\x80"
-     "\x13\xA5\x00\x80\x17\xA5\x00\x80\x1B\xA5\x00\x80\xE1\x0C\x05\x00\x1F\xA5"
-     "\x00\x80"
-     "\xE3\x80\x04\x00\x23\xA5\x00\x80\x27\xA5\x00\x80\x2B\xA5\x00\x80\xEF\xDC"
-     "\x05\x00"
-     "\x2F\xA5\x00\x80\x33\xA5\x00\x80\x37\xA5\x00\x80\x3B\xA5\x00\x80\xA8\x45"
-     "\x06\x00"
-     "\xA9\x4D\x06\x00\xAA\x59\x06\x00\xAB\x51\x06\x00\xAC\xB5\x01\x00\xAD\xBD"
-     "\x01\x00"
-     "\xAE\xB5\x01\x00\xAF\xAD\x01\x00\x84\xAC\x01\x00\x3F\xA5\x00\x80\x43\xA5"
-     "\x00\x80"
-     "\x47\xA5\x00\x80\x4B\xA5\x00\x80\x4F\xA5\x00\x80\x53\xA5\x00\x80\x57\xA5"
-     "\x00\x80"
-     "\xB8\x69\x01\x00\xB9\x69\x01\x00\xBA\x09\x01\x00\xBB\x05\x01\x00\xBC\x1D"
-     "\x01\x00"
-     "\xBD\x01\x01\x00\xBE\x01\x01\x00\xBF\x35\x01\x00\xB0\xD5\x01\x00\xB1\xDD"
-     "\x01\x00"
-     "\xB2\xD5\x01\x00\xB3\x69\x01\x00\xB4\x79\x01\x00\xB5\x79\x01\x00\xB6\x69"
-     "\x01\x00"
-     "\xB7\x61\x01\x00\x81\xBD\x03\x00\x80\xBD\x03\x00\xB3\xD9\x05\x00\x82\x19"
-     "\x00\x00"
-     "\xB5\x35\x02\x00\x5B\xA5\x00\x80\xBE\x44\x03\x00\xB6\x35\x02\x00\x63\xA5"
-     "\x00\x80"
-     "\x67\xA5\x00\x80\xBB\x11\x02\x00\xBA\x19\x02\x00\xBD\xB5\x02\x00\xBC\x09"
-     "\x02\x00"
-     "\xBF\xB5\x02\x00\xBE\xB5\x02\x00\x6B\xA5\x00\x80\xA6\x6D\x02\x00\xA5\x6D"
-     "\x02\x00"
-     "\x84\x00\x3C\x00\xA3\x81\x05\x00\x6F\xA5\x00\x80\x86\xF8\x3C\x00\x87\x7C"
-     "\x03\x00"
-     "\xAF\xED\x02\x00\xAE\xED\x02\x00\xAD\xED\x02\x00\xAC\x51\x02\x00\xAB\x49"
-     "\x02\x00"
-     "\xAA\x41\x02\x00\x73\xA5\x00\x80\x77\xA5\x00\x80\x7B\xA5\x00\x80\x7F\xA5"
-     "\x00\x80"
-     "\x83\xA5\x00\x80\x87\xA5\x00\x80\xE3\x20\x01\x00\x8B\xA5\x00\x80\xE1\xA0"
-     "\x01\x00"
-     "\x8F\xA5\x00\x80\xEF\x40\x02\x00\x93\xA5\x00\x80\x97\xA5\x00\x80\x9B\xA5"
-     "\x00\x80"
-     "\x9F\xA5\x00\x80\xA3\xA5\x00\x80\xA7\xA5\x00\x80\xAB\xA5\x00\x80\xB3\xF1"
-     "\x03\x00"
-     "\xAF\xA5\x00\x80\xB3\xA5\x00\x80\xB7\xA5\x00\x80\x0F\xA4\x00\x80\xB6\xE9"
-     "\x03\x00"
-     "\xB5\xE1\x03\x00\xBB\xA5\x00\x80\xBB\xAD\x03\x00\xBA\xAD\x03\x00\xBF\xA5"
-     "\x00\x80"
-     "\xC3\xA5\x00\x80\xBF\x95\x03\x00\xBE\x95\x03\x00\xBD\xA1\x03\x00\xBC\xAD"
-     "\x03\x00"
-     "\xA8\x29\x02\x00\xA9\x29\x02\x00\xAA\x39\x02\x00\xAB\x39\x02\x00\xAC\x29"
-     "\x02\x00"
-     "\xAD\x29\x02\x00\xAE\x59\x02\x00\xAF\x59\x02\x00\x80\xCD\x01\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\xC7\xA5\x00\x80\xCB\xA5\x00\x80\xBE\x7C\x3D\x00\x87\xB4"
-     "\x3D\x00"
-     "\x86\x1C\x3C\x00\xB8\x35\x02\x00\xB9\x3D\x02\x00\xBA\x35\x02\x00\xBB\xE9"
-     "\x02\x00"
-     "\xBC\xF9\x02\x00\xBD\xF9\x02\x00\xBE\xE9\x02\x00\xBF\xE9\x02\x00\xB0\x29"
-     "\x02\x00"
-     "\xB1\x29\x02\x00\xB2\x39\x02\x00\xB3\x35\x02\x00\xB4\x11\x02\x00\xB5\x11"
-     "\x02\x00"
-     "\xB6\x11\x02\x00\xB7\x0D\x02\x00\xD3\xA5\x00\x80\xD7\xA5\x00\x80\xDB\xA5"
-     "\x00\x80"
-     "\xDF\xA5\x00\x80\xE3\xA5\x00\x80\xE7\xA5\x00\x80\xEB\xA5\x00\x80\xEF\x7C"
-     "\x00\x00"
-     "\xEF\xA5\x00\x80\xE1\x54\x06\x00\xF3\xA5\x00\x80\xE3\xA8\x00\x00\xEC\x00"
-     "\x00\x00"
-     "\xF7\xA5\x00\x80\xFB\xA5\x00\x80\xFF\xA5\x00\x80\x03\xA6\x00\x80\xA6\xF9"
-     "\x02\x00"
-     "\xA5\xF1\x02\x00\x07\xA6\x00\x80\xA3\xE1\x02\x00\x0B\xA6\x00\x80\x0F\xA6"
-     "\x00\x80"
-     "\x13\xA6\x00\x80\xAF\x85\x02\x00\xAE\x85\x02\x00\xAD\xB1\x02\x00\xAC\xBD"
-     "\x02\x00"
-     "\xAB\xBD\x02\x00\xAA\xBD\x02\x00\x17\xA6\x00\x80\x1B\xA6\x00\x80\xA8\x21"
-     "\x3E\x00"
-     "\xA9\x21\x3E\x00\xAA\x49\x3E\x00\xAB\x59\x3E\x00\xAC\x69\x3E\x00\xAD\x69"
-     "\x3E\x00"
-     "\xAE\x99\x3E\x00\xAF\x99\x3E\x00\xCF\xA5\x00\x80\x1F\xA6\x00\x80\x23\xA6"
-     "\x00\x80"
-     "\x27\xA6\x00\x80\x2B\xA6\x00\x80\x2F\xA6\x00\x80\x33\xA6\x00\x80\x37\xA6"
-     "\x00\x80"
-     "\xB8\xF5\x3E\x00\xB9\xFD\x3E\x00\xBA\xF5\x3E\x00\xBB\x89\x3E\x00\xBC\x99"
-     "\x3E\x00"
-     "\xBD\x99\x3E\x00\xBE\x85\x3E\x00\xBF\xB5\x3E\x00\xB0\xE9\x3E\x00\xB1\xE9"
-     "\x3E\x00"
-     "\xB2\xFD\x3E\x00\xB3\xF1\x3E\x00\xB4\xD5\x3E\x00\xB5\xDD\x3E\x00\xB6\xD5"
-     "\x3E\x00"
-     "\xB7\xCD\x3E\x00\xB3\xC1\x3E\x00\x82\x15\x00\x00\x81\xB5\x00\x00\x80\xB5"
-     "\x00\x00"
-     "\x3B\xA6\x00\x80\xB6\xE9\x3E\x00\xB5\xE1\x3E\x00\xBE\x10\x00\x00\xBB\x2D"
-     "\x3E\x00"
-     "\xBA\x2D\x3E\x00\x84\x64\x03\x00\x3F\xA6\x00\x80\xBF\x19\x3E\x00\xBE\x11"
-     "\x3E\x00"
-     "\xBD\x25\x3E\x00\xBC\x29\x3E\x00\x43\xA6\x00\x80\xA3\x85\x3E\x00\x86\xA8"
-     "\x00\x00"
-     "\x87\x6C\x03\x00\xA6\xAD\x3E\x00\x47\xA6\x00\x80\x4B\xA6\x00\x80\xA5\xA5"
-     "\x3E\x00"
-     "\xAA\x69\x3E\x00\xAB\x69\x3E\x00\x4F\xA6\x00\x80\x53\xA6\x00\x80\xAE\x55"
-     "\x3E\x00"
-     "\xAF\x5D\x3E\x00\xAC\x6D\x3E\x00\xAD\x61\x3E\x00\xB3\x89\x3E\x00\x57\xA6"
-     "\x00\x80"
-     "\x5B\xA6\x00\x80\x5F\xA6\x00\x80\x63\xA6\x00\x80\xB6\x81\x3E\x00\xB5\x89"
-     "\x3E\x00"
-     "\x67\xA6\x00\x80\xBB\x95\x3E\x00\xBA\x8D\x3E\x00\x6B\xA6\x00\x80\x6F\xA6"
-     "\x00\x80"
-     "\xBF\xFD\x3E\x00\xBE\xF5\x3E\x00\xBD\x85\x3E\x00\xBC\x85\x3E\x00\x73\xA6"
-     "\x00\x80"
-     "\x77\xA6\x00\x80\x7B\xA6\x00\x80\x7F\xA6\x00\x80\x83\xA6\x00\x80\x87\xA6"
-     "\x00\x80"
-     "\xEF\x94\x3C\x00\x8B\xA6\x00\x80\xE1\x20\x3D\x00\x8F\xA6\x00\x80\xE3\x34"
-     "\x3D\x00"
-     "\x93\xA6\x00\x80\x97\xA6\x00\x80\x9B\xA6\x00\x80\x9F\xA6\x00\x80\xA3\xA6"
-     "\x00\x80"
-     "\xA8\x01\x3E\x00\xA9\x01\x3E\x00\xAA\x01\x3E\x00\xAB\x01\x3E\x00\xAC\x01"
-     "\x3E\x00"
-     "\xAD\x01\x3E\x00\xAE\x01\x3E\x00\xAF\x3D\x3E\x00\x80\xDD\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\xA7\xA6\x00\x80\xAB\xA6\x00\x80\x84\x3C\x01\x00\xBE\x74"
-     "\x00\x00"
-     "\xB3\xA6\x00\x80\xB8\x1D\x3E\x00\xB9\x3D\x3E\x00\xBA\x35\x3E\x00\xBB\xE9"
-     "\x01\x00"
-     "\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\xE9\x01\x00\xBF\xE9\x01\x00\xB0\x49"
-     "\x3E\x00"
-     "\xB1\x51\x3E\x00\xB2\x51\x3E\x00\xB3\x51\x3E\x00\xB4\x31\x3E\x00\xB5\x39"
-     "\x3E\x00"
-     "\xB6\x2D\x3E\x00\xB7\x25\x3E\x00\xA3\x89\x3D\x00\xB7\xA6\x00\x80\x86\xA8"
-     "\x04\x00"
-     "\x87\x3C\x03\x00\xBB\xA6\x00\x80\xA6\x81\x3D\x00\xA5\x89\x3D\x00\xBF\xA6"
-     "\x00\x80"
-     "\xAB\x95\x3D\x00\xAA\x8D\x3D\x00\xC3\xA6\x00\x80\xC7\xA6\x00\x80\xAF\xFD"
-     "\x3D\x00"
-     "\xAE\xF5\x3D\x00\xAD\x85\x3D\x00\xAC\x85\x3D\x00\xCB\xA6\x00\x80\xB3\x9D"
-     "\x02\x00"
-     "\x84\x48\x03\x00\xBE\x40\x04\x00\xB6\x55\x03\x00\xCF\xA6\x00\x80\xD3\xA6"
-     "\x00\x80"
-     "\xB5\xB1\x02\x00\xBA\x71\x03\x00\xBB\x79\x03\x00\xD7\xA6\x00\x80\xDB\xA6"
-     "\x00\x80"
-     "\xBE\x31\x03\x00\xBF\x31\x03\x00\xBC\x51\x03\x00\xBD\x51\x03\x00\xB8\xD1"
-     "\x03\x00"
-     "\xB9\xD1\x03\x00\xBA\xD5\x03\x00\xBB\xE9\x03\x00\xBC\xFD\x03\x00\xBD\xE5"
-     "\x03\x00"
-     "\xBE\xED\x03\x00\xBF\xE5\x03\x00\xB0\xB1\x03\x00\xB1\xB9\x03\x00\xB2\x81"
-     "\x03\x00"
-     "\xB3\x81\x03\x00\xB4\xF1\x03\x00\xB5\xF1\x03\x00\xB6\xF1\x03\x00\xB7\xF1"
-     "\x03\x00"
-     "\xA8\x01\x02\x00\xA9\x11\x02\x00\xAA\x11\x02\x00\xAB\x25\x02\x00\xAC\x3D"
-     "\x02\x00"
-     "\xAD\xD1\x03\x00\xAE\xD1\x03\x00\xAF\xD1\x03\x00\xDF\xA6\x00\x80\x5F\xA5"
-     "\x00\x80"
-     "\xE3\xA6\x00\x80\xE7\xA6\x00\x80\xEB\xA6\x00\x80\xEF\xA6\x00\x80\xF3\xA6"
-     "\x00\x80"
-     "\xF7\xA6\x00\x80\xFB\xA6\x00\x80\xFF\xA6\x00\x80\x03\xA7\x00\x80\x07\xA7"
-     "\x00\x80"
-     "\xEF\x50\x00\x00\x0B\xA7\x00\x80\x0F\xA7\x00\x80\x13\xA7\x00\x80\xE3\x84"
-     "\x00\x00"
-     "\x17\xA7\x00\x80\xE1\xF8\x01\x00\x1B\xA7\x00\x80\x80\x2D\x00\x00\x81\x11"
-     "\x00\x00"
-     "\x82\x15\x00\x00\x23\xA7\x00\x80\xA3\x99\x03\x00\x27\xA7\x00\x80\x2B\xA7"
-     "\x00\x80"
-     "\x2F\xA7\x00\x80\x33\xA7\x00\x80\xA6\x51\x02\x00\xA5\xB5\x03\x00\x37\xA7"
-     "\x00\x80"
-     "\xAB\x7D\x02\x00\xAA\x75\x02\x00\x3B\xA7\x00\x80\x3F\xA7\x00\x80\xAF\x35"
-     "\x02\x00"
-     "\xAE\x35\x02\x00\xAD\x55\x02\x00\xAC\x55\x02\x00\x86\xA0\x04\x00\x87\xA8"
-     "\x05\x00"
-     "\x43\xA7\x00\x80\x47\xA7\x00\x80\x4B\xA7\x00\x80\x4F\xA7\x00\x80\x53\xA7"
-     "\x00\x80"
-     "\xEF\xF8\x01\x00\xBE\x2C\x04\x00\xE1\x80\x3E\x00\x57\xA7\x00\x80\xE3\x14"
-     "\x01\x00"
-     "\x5B\xA7\x00\x80\x5F\xA7\x00\x80\x63\xA7\x00\x80\x67\xA7\x00\x80\x6B\xA7"
-     "\x00\x80"
-     "\x6F\xA7\x00\x80\x73\xA7\x00\x80\xB3\xDD\x01\x00\x77\xA7\x00\x80\xB5\xFD"
-     "\x01\x00"
-     "\xB6\xF5\x01\x00\x7B\xA7\x00\x80\x7F\xA7\x00\x80\x83\xA7\x00\x80\xBA\xB1"
-     "\x01\x00"
-     "\xBB\x85\x01\x00\xBC\x9D\x01\x00\xBD\x15\x01\x00\xBE\x11\x01\x00\xBF\x11"
-     "\x01\x00"
-     "\xA8\xBD\x05\x00\xA9\x01\x06\x00\xAA\x19\x06\x00\xAB\x11\x06\x00\xAC\x39"
-     "\x06\x00"
-     "\xAD\x39\x06\x00\xAE\x59\x06\x00\xAF\x49\x06\x00\x1F\xA7\x00\x80\x82\x11"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x80\xC9\x07\x00\x87\xA7\x00\x80\x8B\xA7\x00\x80\x84\x30"
-     "\x02\x00"
-     "\x8F\xA7\x00\x80\xB8\xB9\x06\x00\xB9\xB9\x06\x00\xBA\x49\x07\x00\xBB\x49"
-     "\x07\x00"
-     "\xBC\x59\x07\x00\xBD\x59\x07\x00\xBE\x4D\x07\x00\xBF\x45\x07\x00\xB0\x39"
-     "\x06\x00"
-     "\xB1\x0D\x06\x00\xB2\x05\x06\x00\xB3\x8D\x06\x00\xB4\x99\x06\x00\xB5\x99"
-     "\x06\x00"
-     "\xB6\x89\x06\x00\xB7\x89\x06\x00\xA3\x91\x06\x00\x93\xA7\x00\x80\x86\x28"
-     "\x00\x00"
-     "\x87\xFC\x03\x00\x97\xA7\x00\x80\xA6\xB9\x06\x00\xA5\xB1\x06\x00\x9B\xA7"
-     "\x00\x80"
-     "\xAB\xC9\x06\x00\xAA\xFD\x06\x00\x9F\xA7\x00\x80\xA3\xA7\x00\x80\xAF\x5D"
-     "\x06\x00"
-     "\xAE\x5D\x06\x00\xAD\x59\x06\x00\xAC\xD1\x06\x00\xA7\xA7\x00\x80\xB3\x8D"
-     "\x07\x00"
-     "\xAB\xA7\x00\x80\xAF\xA7\x00\x80\xB6\x8D\x07\x00\xB3\xA7\x00\x80\xB7\xA7"
-     "\x00\x80"
-     "\xB5\x8D\x07\x00\xBA\x69\x07\x00\xBB\x69\x07\x00\xBB\xA7\x00\x80\xBF\xA7"
-     "\x00\x80"
-     "\xBE\x69\x07\x00\xBF\x5D\x07\x00\xBC\x69\x07\x00\xBD\x69\x07\x00\xA3\x49"
-     "\x06\x00"
-     "\xC3\xA7\x00\x80\xC7\xA7\x00\x80\xCB\xA7\x00\x80\xCF\xA7\x00\x80\xA6\x49"
-     "\x06\x00"
-     "\xA5\x49\x06\x00\xD3\xA7\x00\x80\xAB\xAD\x06\x00\xAA\xAD\x06\x00\xD7\xA7"
-     "\x00\x80"
-     "\xDB\xA7\x00\x80\xAF\x99\x06\x00\xAE\xAD\x06\x00\xAD\xAD\x06\x00\xAC\xAD"
-     "\x06\x00"
-     "\xDF\xA7\x00\x80\xE3\xA7\x00\x80\xE7\xA7\x00\x80\xEB\xA7\x00\x80\xEF\xA7"
-     "\x00\x80"
-     "\xEF\xE0\x05\x00\xF3\xA7\x00\x80\xF7\xA7\x00\x80\xFB\xA7\x00\x80\xE3\x3C"
-     "\x05\x00"
-     "\x84\x24\x01\x00\xE1\x7C\x05\x00\x80\x29\x00\x00\x81\x29\x00\x00\x82\x19"
-     "\x00\x00"
-     "\xBE\x34\x01\x00\xA8\xBD\x06\x00\xA9\xC1\x06\x00\xAA\xC1\x06\x00\xAB\xC1"
-     "\x06\x00"
-     "\xAC\xC1\x06\x00\xAD\xC9\x06\x00\xAE\xF1\x06\x00\xAF\xF1\x06\x00\x86\x00"
-     "\x00\x00"
-     "\x87\x84\x00\x00\x03\xA8\x00\x80\x07\xA8\x00\x80\x0B\xA8\x00\x80\x0F\xA8"
-     "\x00\x80"
-     "\x13\xA8\x00\x80\x17\xA8\x00\x80\xB8\x79\x06\x00\xB9\x79\x06\x00\xBA\x09"
-     "\x06\x00"
-     "\xBB\x09\x06\x00\xBC\x19\x06\x00\xBD\x05\x06\x00\xBE\x05\x06\x00\xBF\x3D"
-     "\x06\x00"
-     "\xB0\x69\x06\x00\xB1\x69\x06\x00\xB2\x79\x06\x00\xB3\x79\x06\x00\xB4\x69"
-     "\x06\x00"
-     "\xB5\x5D\x06\x00\xB6\x49\x06\x00\xB7\x49\x06\x00\xA8\x85\x02\x00\xA9\x95"
-     "\x02\x00"
-     "\xAA\x95\x02\x00\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xD5\x02\x00\xAE\xD1"
-     "\x02\x00"
-     "\xAF\xD1\x02\x00\x1B\xA8\x00\x80\x1F\xA8\x00\x80\x84\x28\x03\x00\xBF\x04"
-     "\xE9\x00"
-     "\x23\xA8\x00\x80\x27\xA8\x00\x80\x2B\xA8\x00\x80\x2F\xA8\x00\x80\xB8\x5D"
-     "\x03\x00"
-     "\xB9\x6D\x03\x00\xBA\x65\x03\x00\xBB\x0D\x03\x00\xBC\x15\x03\x00\xBD\x1D"
-     "\x03\x00"
-     "\xBE\x15\x03\x00\xBF\x0D\x03\x00\xB0\xB5\x02\x00\xB1\xBD\x02\x00\xB2\x81"
-     "\x02\x00"
-     "\xB3\x81\x02\x00\xB4\x75\x03\x00\xB5\x7D\x03\x00\xB6\x75\x03\x00\xB7\x6D"
-     "\x03\x00"
-     "\xB3\x1D\x02\x00\x33\xA8\x00\x80\x37\xA8\x00\x80\x3B\xA8\x00\x80\xBE\x1C"
-     "\x0C\x00"
-     "\xB6\x55\x02\x00\xB5\x35\x02\x00\x3F\xA8\x00\x80\xBB\x79\x02\x00\xBA\x71"
-     "\x02\x00"
-     "\x43\xA8\x00\x80\x47\xA8\x00\x80\xBF\xB9\x03\x00\xBE\xB1\x03\x00\xBD\x55"
-     "\x02\x00"
-     "\xBC\x55\x02\x00\x4B\xA8\x00\x80\xA3\x59\x02\x00\x4F\xA8\x00\x80\x53\xA8"
-     "\x00\x80"
-     "\xA6\x11\x02\x00\x57\xA8\x00\x80\x5B\xA8\x00\x80\xA5\x71\x02\x00\xAA\x35"
-     "\x02\x00"
-     "\xAB\x3D\x02\x00\xAF\xA6\x00\x80\x5F\xA8\x00\x80\xAE\xF5\x03\x00\xAF\xFD"
-     "\x03\x00"
-     "\xAC\x11\x02\x00\xAD\x11\x02\x00\xA8\xA9\x02\x00\xA9\xA9\x02\x00\xAA\xB9"
-     "\x02\x00"
-     "\xAB\xB9\x02\x00\xAC\xA9\x02\x00\xAD\xA9\x02\x00\xAE\x39\x01\x00\xAF\x39"
-     "\x01\x00"
-     "\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\x63\xA8\x00\x80\x67\xA8"
-     "\x00\x80"
-     "\x6B\xA8\x00\x80\x87\x68\x0D\x00\x86\x1C\x0D\x00\xB8\xE9\x01\x00\xB9\xE9"
-     "\x01\x00"
-     "\xBA\x89\x01\x00\xBB\x85\x01\x00\xBC\x9D\x01\x00\xBD\x81\x01\x00\xBE\x81"
-     "\x01\x00"
-     "\xBF\xB1\x01\x00\xB0\x51\x01\x00\xB1\x51\x01\x00\xB2\x51\x01\x00\xB3\x51"
-     "\x01\x00"
-     "\xB4\xF5\x01\x00\xB5\xFD\x01\x00\xB6\xF5\x01\x00\xB7\xE1\x01\x00\xEF\x04"
-     "\x00\x00"
-     "\x84\x48\x0C\x00\x73\xA8\x00\x80\xBE\x80\x0C\x00\x77\xA8\x00\x80\x7B\xA8"
-     "\x00\x80"
-     "\x7F\xA8\x00\x80\xEF\xC4\x06\x00\x83\xA8\x00\x80\xE1\x3C\x06\x00\x87\xA8"
-     "\x00\x80"
-     "\xE3\x84\x01\x00\xE3\xC8\x00\x00\x8B\xA8\x00\x80\xE1\x28\x01\x00\x8F\xA8"
-     "\x00\x80"
-     "\x93\xA8\x00\x80\x97\xA8\x00\x80\x9B\xA8\x00\x80\xB3\x7D\x02\x00\x9F\xA8"
-     "\x00\x80"
-     "\xB5\x6D\x02\x00\xB6\x65\x02\x00\xA3\xA8\x00\x80\xA7\xA8\x00\x80\xAB\xA8"
-     "\x00\x80"
-     "\xBA\xD9\x01\x00\xBB\xD9\x01\x00\xBC\xC9\x01\x00\xBD\xC9\x01\x00\xBE\x7D"
-     "\x01\x00"
-     "\xBF\x75\x01\x00\xA3\xB1\x0D\x00\x6F\xA8\x00\x80\xAF\xA8\x00\x80\xB3\xA8"
-     "\x00\x80"
-     "\xB7\xA8\x00\x80\xA6\xA9\x0D\x00\xA5\xA1\x0D\x00\xBB\xA8\x00\x80\xAB\x15"
-     "\x0E\x00"
-     "\xAA\x15\x0E\x00\xBF\xA8\x00\x80\xC3\xA8\x00\x80\xAF\xB9\x0E\x00\xAE\xB1"
-     "\x0E\x00"
-     "\xAD\x05\x0E\x00\xAC\x05\x0E\x00\xC7\xA8\x00\x80\xCB\xA8\x00\x80\xCF\xA8"
-     "\x00\x80"
-     "\xD3\xA8\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39\x00\x00\xD7\xA8"
-     "\x00\x80"
-     "\xDB\xA8\x00\x80\xDF\xA8\x00\x80\x86\xC8\x00\x00\x87\x0C\x03\x00\xE3\xA8"
-     "\x00\x80"
-     "\xE7\xA8\x00\x80\xEB\xA8\x00\x80\xEF\xA8\x00\x80\xA8\x09\x0F\x00\xA9\x15"
-     "\x0F\x00"
-     "\xAA\x1D\x0F\x00\xAB\x15\x0F\x00\xAC\x31\x0F\x00\xAD\x3D\x0F\x00\xAE\x2D"
-     "\x0F\x00"
-     "\xAF\x21\x0F\x00\x84\xA8\x03\x00\xF3\xA8\x00\x80\xF7\xA8\x00\x80\xFB\xA8"
-     "\x00\x80"
-     "\xFF\xA8\x00\x80\x03\xA9\x00\x80\x07\xA9\x00\x80\x0B\xA9\x00\x80\xB8\xCD"
-     "\x00\x00"
-     "\xB9\xD5\x00\x00\xBA\xD5\x00\x00\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\x95"
-     "\x00\x00"
-     "\xBE\x9D\x00\x00\xBF\x95\x00\x00\xB0\x21\x0F\x00\xB1\x21\x0F\x00\xB2\x21"
-     "\x0F\x00"
-     "\xB3\x3D\x0F\x00\xB4\x2D\x0F\x00\xB5\x15\x0F\x00\xB6\x1D\x0F\x00\xB7\xF5"
-     "\x00\x00"
-     "\xB3\x95\x0E\x00\x0F\xA9\x00\x80\x13\xA9\x00\x80\x17\xA9\x00\x80\x1B\xA9"
-     "\x00\x80"
-     "\xB6\xBD\x0E\x00\xB5\xB9\x0E\x00\x1F\xA9\x00\x80\xBB\xE5\x0E\x00\xBA\xE5"
-     "\x0E\x00"
-     "\x23\xA9\x00\x80\x27\xA9\x00\x80\xBF\xD1\x0E\x00\xBE\xE9\x0E\x00\xBD\xE9"
-     "\x0E\x00"
-     "\xBC\xF5\x0E\x00\xEF\x28\x0D\x00\x2B\xA9\x00\x80\x2F\xA9\x00\x80\x33\xA9"
-     "\x00\x80"
-     "\x37\xA9\x00\x80\x3B\xA9\x00\x80\x3F\xA9\x00\x80\x43\xA9\x00\x80\x80\x3D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x1D\x00\x00\x47\xA9\x00\x80\xE3\x48\x0D\x00\x4B\xA9"
-     "\x00\x80"
-     "\xE1\xC0\x0E\x00\xBE\x7C\x01\x00\xA3\x15\x0E\x00\x53\xA9\x00\x80\x86\x28"
-     "\x00\x00"
-     "\x87\x4C\x01\x00\x57\xA9\x00\x80\xA6\x3D\x0E\x00\xA5\x39\x0E\x00\x5B\xA9"
-     "\x00\x80"
-     "\xAB\x65\x0E\x00\xAA\x65\x0E\x00\x5F\xA9\x00\x80\x63\xA9\x00\x80\xAF\x51"
-     "\x0E\x00"
-     "\xAE\x69\x0E\x00\xAD\x69\x0E\x00\xAC\x75\x0E\x00\x67\xA9\x00\x80\xB3\xBD"
-     "\x01\x00"
-     "\x6B\xA9\x00\x80\x6F\xA9\x00\x80\xB6\x79\x01\x00\x73\xA9\x00\x80\x77\xA9"
-     "\x00\x80"
-     "\xB5\x79\x01\x00\xBA\x55\x01\x00\xBB\x5D\x01\x00\x7B\xA9\x00\x80\x7F\xA9"
-     "\x00\x80"
-     "\xBE\xF9\x00\x00\xBF\xF9\x00\x00\xBC\x45\x01\x00\xBD\xF9\x00\x00\xA8\x71"
-     "\x02\x00"
-     "\xA9\x71\x02\x00\xAA\x71\x02\x00\xAB\x71\x02\x00\xAC\xB5\x02\x00\xAD\xBD"
-     "\x02\x00"
-     "\xAE\xB5\x02\x00\xAF\xAD\x02\x00\x84\xEC\x04\x00\x83\xA9\x00\x80\x87\xA9"
-     "\x00\x80"
-     "\x8B\xA9\x00\x80\x8F\xA9\x00\x80\x93\xA9\x00\x80\x97\xA9\x00\x80\x9B\xA9"
-     "\x00\x80"
-     "\xB8\x69\x03\x00\xB9\x69\x03\x00\xBA\x0D\x03\x00\xBB\x05\x03\x00\xBC\x1D"
-     "\x03\x00"
-     "\xBD\x05\x03\x00\xBE\x0D\x03\x00\xBF\x05\x03\x00\xB0\xD5\x02\x00\xB1\xDD"
-     "\x02\x00"
-     "\xB2\xD5\x02\x00\xB3\x69\x03\x00\xB4\x79\x03\x00\xB5\x79\x03\x00\xB6\x69"
-     "\x03\x00"
-     "\xB7\x61\x03\x00\x9F\xA9\x00\x80\xA3\xA9\x00\x80\xA7\xA9\x00\x80\xA3\xF5"
-     "\x02\x00"
-     "\xAB\xA9\x00\x80\xA5\x31\x02\x00\xA6\x31\x02\x00\xAF\xA9\x00\x80\xB3\xA9"
-     "\x00\x80"
-     "\xB7\xA9\x00\x80\xAA\x1D\x02\x00\xAB\x15\x02\x00\xAC\x0D\x02\x00\xAD\xB1"
-     "\x03\x00"
-     "\xAE\xB1\x03\x00\xAF\xB1\x03\x00\x80\x61\x00\x00\x81\x61\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xBB\xA9\x00\x80\x86\xF0\x04\x00\x87\x60\x03\x00\xBE\x10\x04\x00\xC3\xA9"
-     "\x00\x80"
-     "\xC7\xA9\x00\x80\xCB\xA9\x00\x80\xCF\xA9\x00\x80\xD3\xA9\x00\x80\xD7\xA9"
-     "\x00\x80"
-     "\xDB\xA9\x00\x80\xDF\xA9\x00\x80\xFF\xA7\x00\x80\xA8\x85\x02\x00\xA9\x95"
-     "\x02\x00"
-     "\xAA\x95\x02\x00\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xD5\x02\x00\xAE\xD1"
-     "\x02\x00"
-     "\xAF\xD1\x02\x00\xE3\xA9\x00\x80\xE7\xA9\x00\x80\xEB\xA9\x00\x80\xEF\xA9"
-     "\x00\x80"
-     "\xF3\xA9\x00\x80\xF7\xA9\x00\x80\xFB\xA9\x00\x80\xFF\xA9\x00\x80\xB8\x75"
-     "\x01\x00"
-     "\xB9\x7D\x01\x00\xBA\x75\x01\x00\xBB\xCD\x01\x00\xBC\xD5\x01\x00\xBD\xDD"
-     "\x01\x00"
-     "\xBE\xC9\x01\x00\xBF\xC1\x01\x00\xB0\xB5\x02\x00\xB1\xBD\x02\x00\xB2\x81"
-     "\x02\x00"
-     "\xB3\x81\x02\x00\xB4\x55\x01\x00\xB5\x5D\x01\x00\xB6\x55\x01\x00\xB7\x4D"
-     "\x01\x00"
-     "\xE1\xA4\x0E\x00\x84\x08\x04\x00\xE3\x1C\x0E\x00\x03\xAA\x00\x80\xBE\x9C"
-     "\x04\x00"
-     "\x07\xAA\x00\x80\x0B\xAA\x00\x80\x0F\xAA\x00\x80\x13\xAA\x00\x80\x17\xAA"
-     "\x00\x80"
-     "\x1B\xAA\x00\x80\x1F\xAA\x00\x80\x81\x75\x00\x00\x80\x75\x00\x00\xEF\xCC"
-     "\x0E\x00"
-     "\x82\x75\x00\x00\x23\xAA\x00\x80\x27\xAA\x00\x80\x2B\xAA\x00\x80\x84\xC4"
-     "\x04\x00"
-     "\x33\xAA\x00\x80\x37\xAA\x00\x80\xEF\x5C\x00\x00\x3B\xAA\x00\x80\xE1\x90"
-     "\x01\x00"
-     "\x3F\xAA\x00\x80\xE3\x58\x01\x00\x43\xAA\x00\x80\x47\xAA\x00\x80\x4B\xAA"
-     "\x00\x80"
-     "\x86\x58\x07\x00\x87\xEC\x04\x00\xB3\x9D\x05\x00\xBF\xA9\x00\x80\x2F\xAA"
-     "\x00\x80"
-     "\x4F\xAA\x00\x80\x53\xAA\x00\x80\xB6\xD5\x05\x00\xB5\xB5\x05\x00\x57\xAA"
-     "\x00\x80"
-     "\xBB\xF9\x05\x00\xBA\xF1\x05\x00\x5B\xAA\x00\x80\x5F\xAA\x00\x80\xBF\x19"
-     "\x06\x00"
-     "\xBE\x11\x06\x00\xBD\xD5\x05\x00\xBC\xD5\x05\x00\x63\xAA\x00\x80\xA3\xD9"
-     "\x05\x00"
-     "\x67\xAA\x00\x80\x6B\xAA\x00\x80\xA6\x91\x05\x00\x6F\xAA\x00\x80\x73\xAA"
-     "\x00\x80"
-     "\xA5\xF1\x05\x00\xAA\xB5\x05\x00\xAB\xBD\x05\x00\x77\xAA\x00\x80\x7B\xAA"
-     "\x00\x80"
-     "\xAE\x55\x06\x00\xAF\x5D\x06\x00\xAC\x91\x05\x00\xAD\x91\x05\x00\xA8\x5D"
-     "\x06\x00"
-     "\xA9\x65\x06\x00\xAA\x6D\x06\x00\xAB\x65\x06\x00\xAC\x7D\x06\x00\xAD\x65"
-     "\x06\x00"
-     "\xAE\x6D\x06\x00\xAF\x65\x06\x00\xBE\xA0\x00\x00\x7F\xAA\x00\x80\x83\xAA"
-     "\x00\x80"
-     "\x87\xAA\x00\x80\x8B\xAA\x00\x80\x8F\xAA\x00\x80\x93\xAA\x00\x80\x97\xAA"
-     "\x00\x80"
-     "\xB8\x4D\x07\x00\xB9\x51\x07\x00\xBA\x51\x07\x00\xBB\x51\x07\x00\xBC\x71"
-     "\x07\x00"
-     "\xBD\x71\x07\x00\xBE\x71\x07\x00\xBF\x71\x07\x00\xB0\xC1\x07\x00\xB1\xC1"
-     "\x07\x00"
-     "\xB2\xC1\x07\x00\xB3\xC1\x07\x00\xB4\xC1\x07\x00\xB5\xC1\x07\x00\xB6\xC1"
-     "\x07\x00"
-     "\xB7\xC1\x07\x00\x9B\xAA\x00\x80\xE1\x94\x06\x00\xA3\xAA\x00\x80\xE3\x3C"
-     "\x05\x00"
-     "\x80\xA9\x01\x00\x81\xA9\x01\x00\x82\x19\x00\x00\xA7\xAA\x00\x80\x86\x80"
-     "\x00\x00"
-     "\x87\x04\x01\x00\xAB\xAA\x00\x80\xAF\xAA\x00\x80\xB3\xAA\x00\x80\xB7\xAA"
-     "\x00\x80"
-     "\xBB\xAA\x00\x80\xEF\x1C\x05\x00\xB3\xA5\x07\x00\x9F\xAA\x00\x80\xBF\xAA"
-     "\x00\x80"
-     "\xC3\xAA\x00\x80\xC7\xAA\x00\x80\xB6\xA5\x07\x00\xB5\xB5\x07\x00\xCB\xAA"
-     "\x00\x80"
-     "\xBB\x45\x06\x00\xBA\x45\x06\x00\xCF\xAA\x00\x80\xD3\xAA\x00\x80\xBF\x45"
-     "\x06\x00"
-     "\xBE\x45\x06\x00\xBD\x55\x06\x00\xBC\x55\x06\x00\xA3\x65\x07\x00\xD7\xAA"
-     "\x00\x80"
-     "\xDB\xAA\x00\x80\xDF\xAA\x00\x80\xE3\xAA\x00\x80\xA6\x65\x07\x00\xA5\x75"
-     "\x07\x00"
-     "\xE7\xAA\x00\x80\xAB\x85\x06\x00\xAA\x85\x06\x00\xEB\xAA\x00\x80\xEF\xAA"
-     "\x00\x80"
-     "\xAF\x85\x06\x00\xAE\x85\x06\x00\xAD\x95\x06\x00\xAC\x95\x06\x00\xF3\xAA"
-     "\x00\x80"
-     "\xB3\x19\x06\x00\xF7\xAA\x00\x80\xFB\xAA\x00\x80\xB6\x21\x06\x00\xFF\xAA"
-     "\x00\x80"
-     "\x03\xAB\x00\x80\xB5\x31\x06\x00\xBA\xF1\x01\x00\xBB\xF9\x01\x00\x07\xAB"
-     "\x00\x80"
-     "\x0B\xAB\x00\x80\xBE\xB1\x01\x00\xBF\xBD\x01\x00\xBC\xD1\x01\x00\xBD\xD1"
-     "\x01\x00"
-     "\xA8\xD1\x05\x00\xA9\xD1\x05\x00\xAA\xD5\x05\x00\xAB\xE9\x05\x00\xAC\xFD"
-     "\x05\x00"
-     "\xAD\xE5\x05\x00\xAE\xED\x05\x00\xAF\x11\x02\x00\x80\xF9\x03\x00\x81\xCD"
-     "\x03\x00"
-     "\x82\xC5\x03\x00\x84\x20\x03\x00\xBE\x00\x1C\x00\x13\xAB\x00\x80\x87\x04"
-     "\x03\x00"
-     "\x86\xBC\x1C\x00\xB8\x19\x02\x00\xB9\x2D\x02\x00\xBA\x25\x02\x00\xBB\xE9"
-     "\x02\x00"
-     "\xBC\xF9\x02\x00\xBD\xF9\x02\x00\xBE\xE9\x02\x00\xBF\xE9\x02\x00\xB0\x71"
-     "\x02\x00"
-     "\xB1\x79\x02\x00\xB2\x41\x02\x00\xB3\x41\x02\x00\xB4\x31\x02\x00\xB5\x3D"
-     "\x02\x00"
-     "\xB6\x35\x02\x00\xB7\x29\x02\x00\x4F\xA9\x00\x80\x17\xAB\x00\x80\x1B\xAB"
-     "\x00\x80"
-     "\xA3\x51\x05\x00\x1F\xAB\x00\x80\xA5\x79\x05\x00\xA6\x69\x05\x00\x23\xAB"
-     "\x00\x80"
-     "\x27\xAB\x00\x80\x2B\xAB\x00\x80\xAA\xB9\x02\x00\xAB\xB1\x02\x00\xAC\x99"
-     "\x02\x00"
-     "\xAD\x99\x02\x00\xAE\xF9\x02\x00\xAF\xF5\x02\x00\xE1\xF8\x01\x00\x2F\xAB"
-     "\x00\x80"
-     "\xE3\xF4\x02\x00\x84\xC0\x1C\x00\x33\xAB\x00\x80\x37\xAB\x00\x80\x3B\xAB"
-     "\x00\x80"
-     "\x3F\xAB\x00\x80\x43\xAB\x00\x80\x47\xAB\x00\x80\x4B\xAB\x00\x80\x4F\xAB"
-     "\x00\x80"
-     "\x53\xAB\x00\x80\x57\xAB\x00\x80\xEF\x30\x02\x00\x5B\xAB\x00\x80\xA8\x09"
-     "\x02\x00"
-     "\xA9\x19\x02\x00\xAA\x21\x02\x00\xAB\x21\x02\x00\xAC\x21\x02\x00\xAD\x21"
-     "\x02\x00"
-     "\xAE\x21\x02\x00\xAF\x55\x02\x00\x5F\xAB\x00\x80\x63\xAB\x00\x80\x67\xAB"
-     "\x00\x80"
-     "\x6B\xAB\x00\x80\x6F\xAB\x00\x80\x82\xAD\x00\x00\x81\xAD\x00\x00\x80\xAD"
-     "\x00\x00"
-     "\xB8\xD1\x02\x00\xB9\xD9\x02\x00\xBA\xE1\x02\x00\xBB\xE1\x02\x00\xBC\x91"
-     "\x02\x00"
-     "\xBD\x9D\x02\x00\xBE\x95\x02\x00\xBF\x89\x02\x00\xB0\x2D\x02\x00\xB1\x35"
-     "\x02\x00"
-     "\xB2\x35\x02\x00\xB3\x05\x02\x00\xB4\x1D\x02\x00\xB5\xF1\x02\x00\xB6\xF1"
-     "\x02\x00"
-     "\xB7\xF1\x02\x00\xE1\x48\x07\x00\xE3\x58\x06\x00\xE3\x1C\x06\x00\xE1\x6C"
-     "\x06\x00"
-     "\x73\xAB\x00\x80\x77\xAB\x00\x80\x7B\xAB\x00\x80\xBE\x10\x1C\x00\x86\x00"
-     "\x1E\x00"
-     "\x87\x24\x1C\x00\x83\xAB\x00\x80\x87\xAB\x00\x80\x8B\xAB\x00\x80\xEF\xEC"
-     "\x01\x00"
-     "\xEF\xD0\x06\x00\x8F\xAB\x00\x80\xB3\xE9\x02\x00\x93\xAB\x00\x80\x97\xAB"
-     "\x00\x80"
-     "\x9B\xAB\x00\x80\x9F\xAB\x00\x80\xB6\xE1\x02\x00\xB5\xE9\x02\x00\xA3\xAB"
-     "\x00\x80"
-     "\xBB\xB1\x02\x00\xBA\xA9\x02\x00\xA7\xAB\x00\x80\xAB\xAB\x00\x80\xBF\x75"
-     "\x02\x00"
-     "\xBE\x99\x02\x00\xBD\x91\x02\x00\xBC\xA1\x02\x00\x0F\xAB\x00\x80\xBE\x48"
-     "\x01\x00"
-     "\x84\x04\x03\x00\x7F\xAB\x00\x80\xAF\xAB\x00\x80\xB3\xAB\x00\x80\xB7\xAB"
-     "\x00\x80"
-     "\xBB\xAB\x00\x80\xA8\x15\x1E\x00\xA9\xA1\x1E\x00\xAA\xA1\x1E\x00\xAB\xBD"
-     "\x1E\x00"
-     "\xAC\xAD\x1E\x00\xAD\x95\x1E\x00\xAE\x9D\x1E\x00\xAF\xF9\x1E\x00\xB0\x91"
-     "\x1E\x00"
-     "\xB1\x91\x1E\x00\xB2\x91\x1E\x00\xB3\x91\x1E\x00\xB4\xB1\x1E\x00\xB5\xB1"
-     "\x1E\x00"
-     "\xB6\xB1\x1E\x00\xB7\xB1\x1E\x00\xB8\x91\x1E\x00\xB9\x91\x1E\x00\xBA\x95"
-     "\x1E\x00"
-     "\xBB\x4D\x1F\x00\xBC\x55\x1F\x00\xBD\x5D\x1F\x00\xBE\x55\x1F\x00\xBF\x4D"
-     "\x1F\x00"
-     "\xAC\xDD\x1D\x00\xAD\xED\x1D\x00\xAE\xE5\x1D\x00\xAF\x09\x1D\x00\xBF\xAB"
-     "\x00\x80"
-     "\xC3\xAB\x00\x80\xAA\xD5\x1D\x00\xAB\xCD\x1D\x00\xC7\xAB\x00\x80\xA5\x95"
-     "\x1D\x00"
-     "\xA6\x9D\x1D\x00\xCB\xAB\x00\x80\xCF\xAB\x00\x80\xD3\xAB\x00\x80\xD7\xAB"
-     "\x00\x80"
-     "\xA3\x95\x1D\x00\xB3\x49\x1E\x00\xDB\xAB\x00\x80\xDF\xAB\x00\x80\xE3\xAB"
-     "\x00\x80"
-     "\xE7\xAB\x00\x80\xB6\x75\x1E\x00\xB5\x45\x1E\x00\xEB\xAB\x00\x80\xBB\x35"
-     "\x1E\x00"
-     "\xBA\x35\x1E\x00\x86\x00\x04\x00\x87\x0C\x00\x00\xBF\x15\x1E\x00\xBE\x15"
-     "\x1E\x00"
-     "\xBD\x25\x1E\x00\xBC\x25\x1E\x00\xA3\x89\x1F\x00\x8F\x35\x25\x00\xF3\xAB"
-     "\x00\x80"
-     "\xF7\xAB\x00\x80\xFB\xAB\x00\x80\xA6\xB5\x1F\x00\xA5\x85\x1F\x00\xFF\xAB"
-     "\x00\x80"
-     "\xAB\xF5\x1F\x00\xAA\xF5\x1F\x00\x03\xAC\x00\x80\x07\xAC\x00\x80\xAF\xD5"
-     "\x1F\x00"
-     "\xAE\xD5\x1F\x00\xAD\xE5\x1F\x00\xAC\xE5\x1F\x00\x9E\x25\x19\x00\x9F\xC5"
-     "\x1F\x00"
-     "\x9C\xB5\x1B\x00\x9D\x49\x19\x00\x9A\x71\x15\x00\x9B\x69\x15\x00\x98\xF5"
-     "\x11\x00"
-     "\x99\xB5\x17\x00\x96\x85\x13\x00\x97\x45\x11\x00\x94\x49\x2D\x00\x95\x71"
-     "\x2C\x00"
-     "\x92\xF5\x28\x00\x93\xA5\x2F\x00\x90\x79\x29\x00\x91\x79\x29\x00\x80\x4D"
-     "\x01\x00"
-     "\x81\x55\x01\x00\x82\x55\x01\x00\x0B\xAC\x00\x80\xE1\xB4\x25\x00\xEF\x60"
-     "\x01\x00"
-     "\xE3\x44\x04\x00\x0F\xAC\x00\x80\x84\x6C\x07\x00\xE3\x14\x02\x00\xEF\x78"
-     "\x15\x00"
-     "\xE1\x5C\x03\x00\xA2\xF5\x1C\x00\xA3\xB5\x01\x00\xA0\xBD\x1D\x00\xA1\xB5"
-     "\x1D\x00"
-     "\xB3\x71\x1E\x00\x13\xAC\x00\x80\x17\xAC\x00\x80\x1B\xAC\x00\x80\x1F\xAC"
-     "\x00\x80"
-     "\xB6\x1D\x1E\x00\xB5\x1D\x1E\x00\x23\xAC\x00\x80\xBB\x09\x1E\x00\xBA\x39"
-     "\x1E\x00"
-     "\x27\xAC\x00\x80\x2B\xAC\x00\x80\xBF\xE9\x01\x00\xBE\xE1\x01\x00\xBD\xE9"
-     "\x01\x00"
-     "\xBC\x11\x1E\x00\x82\x31\x26\x00\x83\x05\x2A\x00\x2F\xAC\x00\x80\xBE\xC4"
-     "\x04\x00"
-     "\x86\xC1\x2C\x00\x87\x35\x2E\x00\x84\x1D\x2A\x00\x85\x99\x2B\x00\x8A\x3D"
-     "\x2F\x00"
-     "\x8B\xC1\x10\x00\xEF\xAB\x00\x80\x33\xAC\x00\x80\x8E\xD9\x14\x00\x8F\x15"
-     "\x16\x00"
-     "\x8C\x19\x12\x00\x8D\x91\x12\x00\x92\x7D\x16\x00\x93\xB9\x18\x00\x37\xAC"
-     "\x00\x80"
-     "\x3B\xAC\x00\x80\x96\xD1\x1C\x00\x97\x51\x1E\x00\x94\x5D\x1A\x00\x95\x39"
-     "\x1A\x00"
-     "\x9A\x69\x1E\x00\x9B\x21\x1F\x00\x3F\xAC\x00\x80\x43\xAC\x00\x80\x47\xAC"
-     "\x00\x80"
-     "\x4B\xAC\x00\x80\x9C\x79\x02\x00\x4F\xAC\x00\x80\xA1\x8D\x02\x00\x53\xAC"
-     "\x00\x80"
-     "\xA3\x8D\x06\x00\xA2\xA5\x00\x00\xA5\x01\x08\x00\xA4\x85\x06\x00\xA7\xED"
-     "\x0A\x00"
-     "\xA6\x99\x0A\x00\xA9\xA5\x0E\x00\xA8\x01\x0C\x00\xAB\xED\x0F\x00\xAA\xA5"
-     "\x0E\x00"
-     "\xAD\xB5\xF2\x00\xAC\xB5\xF2\x00\xAF\x01\xF4\x00\xAE\xA1\xF3\x00\xB1\x19"
-     "\xF6\x00"
-     "\xB0\xE5\xF6\x00\xB3\xCD\xFA\x00\xB2\xCD\xFA\x00\xB5\x01\xFC\x00\xB4\xC5"
-     "\xFA\x00"
-     "\x57\xAC\x00\x80\x5B\xAC\x00\x80\x80\x65\x00\x00\x81\x3D\x00\x00\x82\x35"
-     "\x00\x00"
-     "\x63\xAC\x00\x80\x86\xD0\x04\x00\x87\xBC\x04\x00\x67\xAC\x00\x80\x84\x30"
-     "\x05\x00"
-     "\xA8\xA9\x1C\x00\xA9\xA9\x1C\x00\xAA\xB9\x1C\x00\xAB\xB1\x1C\x00\xAC\xD5"
-     "\x1C\x00"
-     "\xAD\xDD\x1C\x00\xAE\xD5\x1C\x00\xAF\xCD\x1C\x00\x6B\xAC\x00\x80\x6F\xAC"
-     "\x00\x80"
-     "\x73\xAC\x00\x80\x77\xAC\x00\x80\x7B\xAC\x00\x80\x7F\xAC\x00\x80\xBE\xF0"
-     "\x04\x00"
-     "\x83\xAC\x00\x80\xB8\xD5\x03\x00\xB9\xE5\x03\x00\xBA\xE5\x03\x00\xBB\xFD"
-     "\x03\x00"
-     "\xBC\xE5\x03\x00\xBD\xED\x03\x00\xBE\x9D\x03\x00\xBF\x95\x03\x00\xB0\xBD"
-     "\x1C\x00"
-     "\xB1\x8D\x1C\x00\xB2\x85\x1C\x00\xB3\x7D\x03\x00\xB4\x65\x03\x00\xB5\x6D"
-     "\x03\x00"
-     "\xB6\x65\x03\x00\xB7\xED\x03\x00\xE3\x98\x01\x00\xE1\xA0\x1F\x00\xE1\x30"
-     "\x1E\x00"
-     "\xE3\x50\x1F\x00\x87\xAC\x00\x80\x8B\xAC\x00\x80\x8F\xAC\x00\x80\x93\xAC"
-     "\x00\x80"
-     "\x97\xAC\x00\x80\x9B\xAC\x00\x80\x9F\xAC\x00\x80\xA3\xAC\x00\x80\xEF\x4C"
-     "\x01\x00"
-     "\xA7\xAC\x00\x80\xAB\xAC\x00\x80\xEF\xD8\x1E\x00\xA3\x71\x1C\x00\xAF\xAC"
-     "\x00\x80"
-     "\xB3\xAC\x00\x80\xB7\xAC\x00\x80\xBB\xAC\x00\x80\xA6\x1D\x1C\x00\xA5\x1D"
-     "\x1C\x00"
-     "\xBF\xAC\x00\x80\xAB\x09\x1C\x00\xAA\x39\x1C\x00\xC3\xAC\x00\x80\xC7\xAC"
-     "\x00\x80"
-     "\xAF\xE9\x03\x00\xAE\xE1\x03\x00\xAD\xE9\x03\x00\xAC\x11\x1C\x00\xB3\x11"
-     "\x05\x00"
-     "\x5F\xAC\x00\x80\xCB\xAC\x00\x80\xD3\xAC\x00\x80\xCF\xAC\x00\x80\xB6\x39"
-     "\x05\x00"
-     "\xB5\x31\x05\x00\xD7\xAC\x00\x80\xBB\xE9\x05\x00\xBA\xE1\x05\x00\x86\x08"
-     "\x03\x00"
-     "\x87\x8C\x03\x00\xBF\xD9\x05\x00\xBE\xE9\x05\x00\xBD\xE9\x05\x00\xBC\xF1"
-     "\x05\x00"
-     "\x82\x49\x00\x00\xA3\x55\x05\x00\x80\xA1\x07\x00\x81\xA1\x07\x00\xA6\x7D"
-     "\x05\x00"
-     "\xDB\xAC\x00\x80\xDF\xAC\x00\x80\xA5\x75\x05\x00\xAA\xA5\x05\x00\xAB\xAD"
-     "\x05\x00"
-     "\xE3\xAC\x00\x80\xE7\xAC\x00\x80\xAE\xAD\x05\x00\xAF\x9D\x05\x00\xAC\xB5"
-     "\x05\x00"
-     "\xAD\xAD\x05\x00\xA8\x6D\x06\x00\xA9\x7D\x06\x00\xAA\xA9\x07\x00\xAB\xA9"
-     "\x07\x00"
-     "\xAC\xB9\x07\x00\xAD\xB9\x07\x00\xAE\xA9\x07\x00\xAF\xA9\x07\x00\xEB\xAC"
-     "\x00\x80"
-     "\xEF\xAC\x00\x80\xF3\xAC\x00\x80\xF7\xAC\x00\x80\xFB\xAC\x00\x80\xFF\xAC"
-     "\x00\x80"
-     "\x03\xAD\x00\x80\x07\xAD\x00\x80\xB8\x6D\x07\x00\xB9\x71\x07\x00\xBA\x71"
-     "\x07\x00"
-     "\xBB\x09\x07\x00\xBC\x19\x07\x00\xBD\x19\x07\x00\xBE\x09\x07\x00\xBF\x01"
-     "\x07\x00"
-     "\xB0\xDD\x07\x00\xB1\x75\x07\x00\xB2\x7D\x07\x00\xB3\x75\x07\x00\xB4\x6D"
-     "\x07\x00"
-     "\xB5\x55\x07\x00\xB6\x5D\x07\x00\xB7\x55\x07\x00\xB3\x21\x06\x00\x0B\xAD"
-     "\x00\x80"
-     "\x0F\xAD\x00\x80\x13\xAD\x00\x80\x17\xAD\x00\x80\xB6\x39\x06\x00\xB5\x31"
-     "\x06\x00"
-     "\x1B\xAD\x00\x80\xBB\xB1\x06\x00\xBA\x19\x06\x00\x1F\xAD\x00\x80\x23\xAD"
-     "\x00\x80"
-     "\xBF\x91\x06\x00\xBE\xA9\x06\x00\xBD\xA1\x06\x00\xBC\xA9\x06\x00\x27\xAD"
-     "\x00\x80"
-     "\xA3\x65\x06\x00\x2B\xAD\x00\x80\x2F\xAD\x00\x80\xA6\x7D\x06\x00\x33\xAD"
-     "\x00\x80"
-     "\x37\xAD\x00\x80\xA5\x75\x06\x00\xAA\x5D\x06\x00\xAB\xF5\x06\x00\x3B\xAD"
-     "\x00\x80"
-     "\x84\x20\x01\x00\xAE\xED\x06\x00\xAF\xD5\x06\x00\xAC\xED\x06\x00\xAD\xE5"
-     "\x06\x00"
-     "\xA8\x8D\x01\x00\xA9\x99\x01\x00\xAA\xDD\x01\x00\xAB\xD5\x01\x00\xAC\xF9"
-     "\x01\x00"
-     "\xAD\xF9\x01\x00\xAE\x39\x01\x00\xAF\x39\x01\x00\x80\xF5\x00\x00\x81\xFD"
-     "\x00\x00"
-     "\x82\xC1\x00\x00\xBE\xE4\x01\x00\x86\x10\x00\x00\x87\x84\x00\x00\x43\xAD"
-     "\x00\x80"
-     "\x47\xAD\x00\x80\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5\x00\x00\xBB\xE5"
-     "\x00\x00"
-     "\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x91\x00\x00\xBF\x91\x00\x00\xB0\x49"
-     "\x01\x00"
-     "\xB1\x49\x01\x00\xB2\x59\x01\x00\xB3\x59\x01\x00\xB4\x49\x01\x00\xB5\x49"
-     "\x01\x00"
-     "\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xBA\x4D\x04\x00\xBB\x6D\x04\x00\xB8\x25"
-     "\x04\x00"
-     "\xB9\x45\x04\x00\xBE\x65\x04\x00\xBF\x69\x04\x00\xBC\x75\x04\x00\xBD\x6D"
-     "\x04\x00"
-     "\xB2\xED\x07\x00\xB3\xF5\x07\x00\xB0\xED\x07\x00\xB1\xE5\x07\x00\xB6\xDD"
-     "\x07\x00"
-     "\xB7\x3D\x04\x00\xB4\xE5\x07\x00\xB5\xD5\x07\x00\xAA\x51\x07\x00\xAB\x59"
-     "\x07\x00"
-     "\xA8\x41\x07\x00\xA9\x41\x07\x00\xAE\xDD\x07\x00\xAF\xC9\x07\x00\xAC\x71"
-     "\x07\x00"
-     "\xAD\x71\x07\x00\x4B\xAD\x00\x80\x4F\xAD\x00\x80\x53\xAD\x00\x80\x57\xAD"
-     "\x00\x80"
-     "\x5B\xAD\x00\x80\x5F\xAD\x00\x80\x63\xAD\x00\x80\x67\xAD\x00\x80\xB3\xD5"
-     "\x02\x00"
-     "\x6B\xAD\x00\x80\x6F\xAD\x00\x80\x73\xAD\x00\x80\x77\xAD\x00\x80\xB6\xF9"
-     "\x02\x00"
-     "\xB5\xF1\x02\x00\x7B\xAD\x00\x80\xBB\x0D\x02\x00\xBA\x0D\x02\x00\x7F\xAD"
-     "\x00\x80"
-     "\x83\xAD\x00\x80\xBF\x0D\x02\x00\xBE\x0D\x02\x00\xBD\x0D\x02\x00\xBC\x0D"
-     "\x02\x00"
-     "\x87\xAD\x00\x80\xA3\x91\x02\x00\x84\xC8\x03\x00\xBE\x40\x0C\x00\xA6\xBD"
-     "\x02\x00"
-     "\x8B\xAD\x00\x80\x8F\xAD\x00\x80\xA5\xB5\x02\x00\xAA\x49\x02\x00\xAB\x49"
-     "\x02\x00"
-     "\x93\xAD\x00\x80\x97\xAD\x00\x80\xAE\x49\x02\x00\xAF\x49\x02\x00\xAC\x49"
-     "\x02\x00"
-     "\xAD\x49\x02\x00\xA8\xB9\x02\x00\xA9\xB9\x02\x00\xAA\x69\x01\x00\xAB\x69"
-     "\x01\x00"
-     "\xAC\x79\x01\x00\xAD\x79\x01\x00\xAE\x69\x01\x00\xAF\x69\x01\x00\xBE\x2C"
-     "\x0C\x00"
-     "\x9B\xAD\x00\x80\x9F\xAD\x00\x80\xA3\xAD\x00\x80\x80\x1D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\xBD\x00\x00\xA7\xAD\x00\x80\xB8\xED\x01\x00\xB9\x85\x01\x00\xBA\x9D"
-     "\x01\x00"
-     "\xBB\x95\x01\x00\xBC\xB9\x01\x00\xBD\xB9\x01\x00\xBE\x7D\x00\x00\xBF\x75"
-     "\x00\x00"
-     "\xB0\x19\x01\x00\xB1\x19\x01\x00\xB2\xED\x01\x00\xB3\xE5\x01\x00\xB4\xFD"
-     "\x01\x00"
-     "\xB5\xE5\x01\x00\xB6\xE5\x01\x00\xB7\xD5\x01\x00\xE1\x68\x07\x00\xAB\xAD"
-     "\x00\x80"
-     "\xE3\x64\x07\x00\xB3\xAD\x00\x80\xB7\xAD\x00\x80\xE1\xA4\x01\x00\xBB\xAD"
-     "\x00\x80"
-     "\xE3\xE0\x00\x00\x84\x2C\x0C\x00\xBF\xAD\x00\x80\xBE\xA8\x0F\x00\xEF\xF8"
-     "\x00\x00"
-     "\x86\x90\x0D\x00\x87\xD4\x0D\x00\xEF\x78\x06\x00\xC3\xAD\x00\x80\xC7\xAD"
-     "\x00\x80"
-     "\xCB\xAD\x00\x80\x85\xF0\xF9\x00\xB3\xAD\x01\x00\xCF\xAD\x00\x80\xB5\xBD"
-     "\x01\x00"
-     "\xB6\xB5\x01\x00\xD3\xAD\x00\x80\xD7\xAD\x00\x80\xDB\xAD\x00\x80\xBA\x91"
-     "\x01\x00"
-     "\xBB\x91\x01\x00\xBC\x75\x01\x00\xBD\x7D\x01\x00\xBE\x75\x01\x00\xBF\x6D"
-     "\x01\x00"
-     "\xA3\x61\x0E\x00\xAF\xAD\x00\x80\xDF\xAD\x00\x80\xE3\xAD\x00\x80\xE7\xAD"
-     "\x00\x80"
-     "\xA6\x79\x0E\x00\xA5\x71\x0E\x00\xEB\xAD\x00\x80\xAB\x5D\x0E\x00\xAA\x5D"
-     "\x0E\x00"
-     "\xEF\xAD\x00\x80\xF3\xAD\x00\x80\xAF\xA1\x0E\x00\xAE\xB9\x0E\x00\xAD\xB1"
-     "\x0E\x00"
-     "\xAC\xB9\x0E\x00\xF7\xAD\x00\x80\xFB\xAD\x00\x80\xFF\xAD\x00\x80\x03\xAE"
-     "\x00\x80"
-     "\x07\xAE\x00\x80\x0B\xAE\x00\x80\x0F\xAE\x00\x80\x13\xAE\x00\x80\x80\x3D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x19\x00\x00\x17\xAE\x00\x80\x1B\xAE\x00\x80\x1F\xAE"
-     "\x00\x80"
-     "\x87\xB8\x03\x00\x86\xDC\x00\x00\xA8\x0D\x0F\x00\xA9\x15\x0F\x00\xAA\x11"
-     "\x0F\x00"
-     "\xAB\x29\x0F\x00\xAC\x2D\x0F\x00\xAD\x55\x0F\x00\xAE\x5D\x0F\x00\xAF\x4D"
-     "\x0F\x00"
-     "\x84\xA0\x03\x00\x23\xAE\x00\x80\x27\xAE\x00\x80\x2B\xAE\x00\x80\x2F\xAE"
-     "\x00\x80"
-     "\x33\xAE\x00\x80\x37\xAE\x00\x80\x3B\xAE\x00\x80\xB8\xED\x00\x00\xB9\x85"
-     "\x00\x00"
-     "\xBA\x8D\x00\x00\xBB\x85\x00\x00\xBC\x9D\x00\x00\xBD\x85\x00\x00\xBE\x8D"
-     "\x00\x00"
-     "\xBF\x85\x00\x00\xB0\x3D\x0F\x00\xB1\x05\x0F\x00\xB2\x0D\x0F\x00\xB3\xE5"
-     "\x00\x00"
-     "\xB4\xFD\x00\x00\xB5\xE5\x00\x00\xB6\xE5\x00\x00\xB7\xD5\x00\x00\x3F\xAE"
-     "\x00\x80"
-     "\x43\xAE\x00\x80\x47\xAE\x00\x80\x4B\xAE\x00\x80\x4F\xAE\x00\x80\xE1\xBC"
-     "\x0E\x00"
-     "\x53\xAE\x00\x80\xE3\xC4\x0D\x00\x57\xAE\x00\x80\x5B\xAE\x00\x80\x5F\xAE"
-     "\x00\x80"
-     "\xEF\x18\x0D\x00\x63\xAE\x00\x80\x67\xAE\x00\x80\x6B\xAE\x00\x80\x6F\xAE"
-     "\x00\x80"
-     "\xB3\xD9\x0E\x00\x73\xAE\x00\x80\x77\xAE\x00\x80\x7B\xAE\x00\x80\x7F\xAE"
-     "\x00\x80"
-     "\xB6\x95\x0E\x00\xB5\xE5\x0E\x00\x83\xAE\x00\x80\xBB\x81\x0E\x00\xBA\xB9"
-     "\x0E\x00"
-     "\x87\xAE\x00\x80\x8B\xAE\x00\x80\xBF\x61\x01\x00\xBE\x79\x01\x00\xBD\x71"
-     "\x01\x00"
-     "\xBC\x99\x0E\x00\xA3\x19\x0E\x00\x93\xAE\x00\x80\x86\x28\x00\x00\x87\x5C"
-     "\x01\x00"
-     "\x97\xAE\x00\x80\xA6\x55\x0E\x00\xA5\x25\x0E\x00\x9B\xAE\x00\x80\xAB\x41"
-     "\x0E\x00"
-     "\xAA\x79\x0E\x00\x9F\xAE\x00\x80\xA3\xAE\x00\x80\xAF\xA1\x01\x00\xAE\xB9"
-     "\x01\x00"
-     "\xAD\xB1\x01\x00\xAC\x59\x0E\x00\x80\x9D\x00\x00\x81\xAD\x00\x00\x82\xA5"
-     "\x00\x00"
-     "\xB3\x61\x01\x00\xA7\xAE\x00\x80\xB5\x61\x01\x00\xB6\x61\x01\x00\xAB\xAE"
-     "\x00\x80"
-     "\xAF\xAE\x00\x80\xB3\xAE\x00\x80\xBA\xED\x00\x00\xBB\xE5\x00\x00\xBC\xFD"
-     "\x00\x00"
-     "\xBD\xE5\x00\x00\xBE\xE1\x00\x00\xBF\xDD\x00\x00\xA8\xA5\x02\x00\xA9\xAD"
-     "\x02\x00"
-     "\xAA\xA5\x02\x00\xAB\xBD\x02\x00\xAC\xA5\x02\x00\xAD\xAD\x02\x00\xAE\xA5"
-     "\x02\x00"
-     "\xAF\x1D\x03\x00\x84\xEC\x04\x00\xB7\xAE\x00\x80\xBB\xAE\x00\x80\xBF\xAE"
-     "\x00\x80"
-     "\xC3\xAE\x00\x80\xC7\xAE\x00\x80\xCB\xAE\x00\x80\xCF\xAE\x00\x80\xB8\x71"
-     "\x03\x00"
-     "\xB9\x71\x03\x00\xBA\x71\x03\x00\xBB\x71\x03\x00\xBC\x15\x03\x00\xBD\x1D"
-     "\x03\x00"
-     "\xBE\x15\x03\x00\xBF\x0D\x03\x00\xB0\x65\x03\x00\xB1\x6D\x03\x00\xB2\x65"
-     "\x03\x00"
-     "\xB3\x7D\x03\x00\xB4\x69\x03\x00\xB5\x55\x03\x00\xB6\x5D\x03\x00\xB7\x55"
-     "\x03\x00"
-     "\xD3\xAE\x00\x80\xD7\xAE\x00\x80\xDB\xAE\x00\x80\xA3\x21\x02\x00\x3F\xAD"
-     "\x00\x80"
-     "\xA5\x21\x02\x00\xA6\x21\x02\x00\xDF\xAE\x00\x80\xE3\xAE\x00\x80\xE7\xAE"
-     "\x00\x80"
-     "\xAA\xAD\x03\x00\xAB\xA5\x03\x00\xAC\xBD\x03\x00\xAD\xA5\x03\x00\xAE\xA1"
-     "\x03\x00"
-     "\xAF\x9D\x03\x00\xEB\xAE\x00\x80\xEF\xAE\x00\x80\xF3\xAE\x00\x80\xF7\xAE"
-     "\x00\x80"
-     "\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\xFB\xAE\x00\x80\xBE\x28"
-     "\x04\x00"
-     "\x03\xAF\x00\x80\x86\xC8\x04\x00\x87\x28\x03\x00\x07\xAF\x00\x80\x0B\xAF"
-     "\x00\x80"
-     "\x0F\xAF\x00\x80\x13\xAF\x00\x80\xA8\x85\x02\x00\xA9\x95\x02\x00\xAA\x95"
-     "\x02\x00"
-     "\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xC5\x02\x00\xAE\xC5\x02\x00\xAF\xFD"
-     "\x02\x00"
-     "\x17\xAF\x00\x80\x1B\xAF\x00\x80\x1F\xAF\x00\x80\x23\xAF\x00\x80\x27\xAF"
-     "\x00\x80"
-     "\x2B\xAF\x00\x80\x2F\xAF\x00\x80\x33\xAF\x00\x80\xB8\xC1\x01\x00\xB9\xC1"
-     "\x01\x00"
-     "\xBA\xC1\x01\x00\xBB\xD1\x01\x00\xBC\xF1\x01\x00\xBD\xF1\x01\x00\xBE\x99"
-     "\x01\x00"
-     "\xBF\x99\x01\x00\xB0\x85\x02\x00\xB1\x41\x01\x00\xB2\x41\x01\x00\xB3\x41"
-     "\x01\x00"
-     "\xB4\x41\x01\x00\xB5\x41\x01\x00\xB6\x41\x01\x00\xB7\x41\x01\x00\xE1\x28"
-     "\x0E\x00"
-     "\x37\xAF\x00\x80\xE3\x38\x0E\x00\xBE\x80\x04\x00\x84\x1C\x04\x00\x3B\xAF"
-     "\x00\x80"
-     "\x3F\xAF\x00\x80\x43\xAF\x00\x80\x47\xAF\x00\x80\x4B\xAF\x00\x80\x4F\xAF"
-     "\x00\x80"
-     "\x53\xAF\x00\x80\x57\xAF\x00\x80\x5B\xAF\x00\x80\xEF\x60\x0E\x00\x5F\xAF"
-     "\x00\x80"
-     "\x81\x15\x00\x00\x80\x15\x00\x00\xEF\x98\x0E\x00\x82\x15\x00\x00\x63\xAF"
-     "\x00\x80"
-     "\x67\xAF\x00\x80\x6B\xAF\x00\x80\x73\xAF\x00\x80\x86\x40\x07\x00\x87\xE4"
-     "\x04\x00"
-     "\x77\xAF\x00\x80\x7B\xAF\x00\x80\xE1\x94\x01\x00\x7F\xAF\x00\x80\xE3\x48"
-     "\x01\x00"
-     "\x83\xAF\x00\x80\xB3\x9D\x05\x00\xFF\xAE\x00\x80\x6F\xAF\x00\x80\x87\xAF"
-     "\x00\x80"
-     "\x8B\xAF\x00\x80\xB6\xC5\x05\x00\xB5\xB5\x05\x00\x8F\xAF\x00\x80\xBB\x05"
-     "\x06\x00"
-     "\xBA\xC1\x05\x00\x93\xAF\x00\x80\x97\xAF\x00\x80\xBF\x05\x06\x00\xBE\x05"
-     "\x06\x00"
-     "\xBD\x15\x06\x00\xBC\x15\x06\x00\x9B\xAF\x00\x80\xA3\xD9\x05\x00\x9F\xAF"
-     "\x00\x80"
-     "\xA3\xAF\x00\x80\xA6\x81\x05\x00\xA7\xAF\x00\x80\xAB\xAF\x00\x80\xA5\xF1"
-     "\x05\x00"
-     "\xAA\x85\x05\x00\xAB\x41\x06\x00\xAF\xAF\x00\x80\xB3\xAF\x00\x80\xAE\x41"
-     "\x06\x00"
-     "\xAF\x41\x06\x00\xAC\x51\x06\x00\xAD\x51\x06\x00\xA8\x65\x06\x00\xA9\x75"
-     "\x06\x00"
-     "\xAA\x7D\x06\x00\xAB\x75\x06\x00\xAC\x6D\x06\x00\xAD\xB5\x07\x00\xAE\xBD"
-     "\x07\x00"
-     "\xAF\xB5\x07\x00\x84\xA8\x03\x00\xB7\xAF\x00\x80\xBB\xAF\x00\x80\xBF\xAF"
-     "\x00\x80"
-     "\xC3\xAF\x00\x80\xC7\xAF\x00\x80\xCB\xAF\x00\x80\xCF\xAF\x00\x80\xB8\x5D"
-     "\x07\x00"
-     "\xB9\x65\x07\x00\xBA\x6D\x07\x00\xBB\x65\x07\x00\xBC\x7D\x07\x00\xBD\x65"
-     "\x07\x00"
-     "\xBE\x6D\x07\x00\xBF\x61\x07\x00\xB0\xCD\x07\x00\xB1\xD5\x07\x00\xB2\xDD"
-     "\x07\x00"
-     "\xB3\xD5\x07\x00\xB4\xCD\x07\x00\xB5\x71\x07\x00\xB6\x6D\x07\x00\xB7\x65"
-     "\x07\x00"
-     "\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\xD3\xAF\x00\x80\xD7\xAF"
-     "\x00\x80"
-     "\xDB\xAF\x00\x80\xEF\x00\x05\x00\xBE\x1C\x01\x00\xE1\xB0\x06\x00\xE3\xAF"
-     "\x00\x80"
-     "\xE3\x18\x05\x00\xE7\xAF\x00\x80\x86\x90\x00\x00\x87\x14\x01\x00\xEB\xAF"
-     "\x00\x80"
-     "\xEF\xAF\x00\x80\xB3\xB5\x07\x00\xDF\xAF\x00\x80\xF3\xAF\x00\x80\xF7\xAF"
-     "\x00\x80"
-     "\xFB\xAF\x00\x80\xB6\x75\x06\x00\xB5\xA5\x07\x00\xFF\xAF\x00\x80\xBB\x51"
-     "\x06\x00"
-     "\xBA\x49\x06\x00\x03\xB0\x00\x80\x07\xB0\x00\x80\xBF\xF5\x06\x00\xBE\x49"
-     "\x06\x00"
-     "\xBD\x41\x06\x00\xBC\x49\x06\x00\xA3\x75\x07\x00\x0B\xB0\x00\x80\x0F\xB0"
-     "\x00\x80"
-     "\x13\xB0\x00\x80\x17\xB0\x00\x80\xA6\xB5\x06\x00\xA5\x65\x07\x00\x1B\xB0"
-     "\x00\x80"
-     "\xAB\x91\x06\x00\xAA\x89\x06\x00\x1F\xB0\x00\x80\x23\xB0\x00\x80\xAF\x35"
-     "\x06\x00"
-     "\xAE\x89\x06\x00\xAD\x81\x06\x00\xAC\x89\x06\x00\x27\xB0\x00\x80\xB3\x29"
-     "\x06\x00"
-     "\x2B\xB0\x00\x80\x2F\xB0\x00\x80\xB6\xDD\x01\x00\x33\xB0\x00\x80\x37\xB0"
-     "\x00\x80"
-     "\xB5\x21\x06\x00\xBA\xC9\x01\x00\xBB\xC9\x01\x00\x3B\xB0\x00\x80\x3F\xB0"
-     "\x00\x80"
-     "\xBE\xBD\x01\x00\xBF\xA1\x01\x00\xBC\xB9\x01\x00\xBD\xB5\x01\x00\xA8\xDD"
-     "\x05\x00"
-     "\xA9\xED\x05\x00\xAA\xE5\x05\x00\xAB\xF9\x05\x00\xAC\xED\x05\x00\xAD\x19"
-     "\x02\x00"
-     "\xAE\x09\x02\x00\xAF\x01\x02\x00\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\x43\xB0\x00\x80\x47\xB0\x00\x80\x4B\xB0\x00\x80\xBE\x00\xFC\x00\x53\xB0"
-     "\x00\x80"
-     "\xB8\x2D\x02\x00\xB9\xE1\x02\x00\xBA\xE1\x02\x00\xBB\xE1\x02\x00\xBC\xE1"
-     "\x02\x00"
-     "\xBD\xE1\x02\x00\xBE\xE1\x02\x00\xBF\xE1\x02\x00\xB0\x49\x02\x00\xB1\x49"
-     "\x02\x00"
-     "\xB2\x29\x02\x00\xB3\x25\x02\x00\xB4\x3D\x02\x00\xB5\x21\x02\x00\xB6\x21"
-     "\x02\x00"
-     "\xB7\x15\x02\x00\x84\x88\x03\x00\x57\xB0\x00\x80\x5B\xB0\x00\x80\xA3\x61"
-     "\x05\x00"
-     "\x5F\xB0\x00\x80\xA5\x69\x05\x00\xA6\x95\x02\x00\x63\xB0\x00\x80\x67\xB0"
-     "\x00\x80"
-     "\x8F\xAE\x00\x80\xAA\x81\x02\x00\xAB\x81\x02\x00\xAC\xF1\x02\x00\xAD\xFD"
-     "\x02\x00"
-     "\xAE\xF5\x02\x00\xAF\xE9\x02\x00\xE1\xE8\x01\x00\x6B\xB0\x00\x80\xE3\x58"
-     "\x02\x00"
-     "\x84\xC0\xFC\x00\x86\x50\xFC\x00\x87\x70\x03\x00\x6F\xB0\x00\x80\x73\xB0"
-     "\x00\x80"
-     "\x77\xB0\x00\x80\x7B\xB0\x00\x80\x7F\xB0\x00\x80\x83\xB0\x00\x80\x87\xB0"
-     "\x00\x80"
-     "\x8B\xB0\x00\x80\xEF\x38\x02\x00\x8F\xB0\x00\x80\xA8\x25\x02\x00\xA9\x29"
-     "\x02\x00"
-     "\xAA\x31\x02\x00\xAB\x31\x02\x00\xAC\x59\x02\x00\xAD\x59\x02\x00\xAE\x49"
-     "\x02\x00"
-     "\xAF\x41\x02\x00\x93\xB0\x00\x80\x97\xB0\x00\x80\xBE\xA4\xFC\x00\x9B\xB0"
-     "\x00\x80"
-     "\x9F\xB0\x00\x80\xA3\xB0\x00\x80\xA7\xB0\x00\x80\xAB\xB0\x00\x80\xB8\xED"
-     "\x02\x00"
-     "\xB9\x85\x02\x00\xBA\x81\x02\x00\xBB\x81\x02\x00\xBC\x85\x02\x00\xBD\x8D"
-     "\x02\x00"
-     "\xBE\xB1\x02\x00\xBF\xB1\x02\x00\xB0\x09\x02\x00\xB1\x09\x02\x00\xB2\xED"
-     "\x02\x00"
-     "\xB3\xE5\x02\x00\xB4\xFD\x02\x00\xB5\xE5\x02\x00\xB6\xE5\x02\x00\xB7\xD5"
-     "\x02\x00"
-     "\xE1\x70\x07\x00\xE3\x34\x01\x00\xE3\x08\x06\x00\xE1\x74\x06\x00\xAF\xB0"
-     "\x00\x80"
-     "\xB3\xB0\x00\x80\xB7\xB0\x00\x80\xBB\xB0\x00\x80\x81\x39\x00\x00\x80\x31"
-     "\x00\x00"
-     "\xC3\xB0\x00\x80\x82\x31\x00\x00\xC7\xB0\x00\x80\xEF\x50\x00\x00\xEF\xCC"
-     "\x06\x00"
-     "\xCB\xB0\x00\x80\xCF\xB0\x00\x80\xB3\xFD\x02\x00\xD3\xB0\x00\x80\xD7\xB0"
-     "\x00\x80"
-     "\xB6\x9D\x02\x00\xDB\xB0\x00\x80\xDF\xB0\x00\x80\xB5\x95\x02\x00\xBA\x89"
-     "\x02\x00"
-     "\xBB\x89\x02\x00\x86\xC8\xFC\x00\x87\xE0\xFC\x00\xBE\x7D\x02\x00\xBF\x65"
-     "\x02\x00"
-     "\xBC\x7D\x02\x00\xBD\x75\x02\x00\x4F\xB0\x00\x80\xBF\xB0\x00\x80\xE3\xB0"
-     "\x00\x80"
-     "\xE7\xB0\x00\x80\xEB\xB0\x00\x80\xEF\xB0\x00\x80\xF3\xB0\x00\x80\xF7\xB0"
-     "\x00\x80"
-     "\xA8\x1D\xFE\x00\xA9\xA1\xFE\x00\xAA\xA1\xFE\x00\xAB\xBD\xFE\x00\xAC\xAD"
-     "\xFE\x00"
-     "\xAD\x95\xFE\x00\xAE\xFD\xFE\x00\xAF\xE5\xFE\x00\xB0\x9D\xFE\x00\xB1\xA5"
-     "\xFE\x00"
-     "\xB2\xAD\xFE\x00\xB3\xA5\xFE\x00\xB4\xBD\xFE\x00\xB5\xA5\xFE\x00\xB6\xAD"
-     "\xFE\x00"
-     "\xB7\xA1\xFE\x00\xB8\x41\xFF\x00\xB9\x41\xFF\x00\xBA\x45\xFF\x00\xBB\x5D"
-     "\xFF\x00"
-     "\xBC\x45\xFF\x00\xBD\x4D\xFF\x00\xBE\x45\xFF\x00\xBF\xED\xFF\x00\xAE\x0D"
-     "\xFD\x00"
-     "\xAF\x15\xFD\x00\xAC\x0D\xFD\x00\xAD\x05\xFD\x00\xAA\xF9\xFD\x00\xAB\xF9"
-     "\xFD\x00"
-     "\xFB\xB0\x00\x80\xFF\xB0\x00\x80\xA6\xED\xFD\x00\x03\xB1\x00\x80\x07\xB1"
-     "\x00\x80"
-     "\xA5\xE5\xFD\x00\x0B\xB1\x00\x80\xA3\x8D\xFD\x00\x0F\xB1\x00\x80\x13\xB1"
-     "\x00\x80"
-     "\xB3\x49\xFE\x00\x17\xB1\x00\x80\x1B\xB1\x00\x80\x1F\xB1\x00\x80\x23\xB1"
-     "\x00\x80"
-     "\xB6\x75\xFE\x00\xB5\x45\xFE\x00\x27\xB1\x00\x80\xBB\x01\xFE\x00\xBA\x39"
-     "\xFE\x00"
-     "\x2B\xB1\x00\x80\x84\x64\x03\x00\xBF\x01\xFE\x00\xBE\x19\xFE\x00\xBD\x11"
-     "\xFE\x00"
-     "\xBC\x19\xFE\x00\xA3\x89\xFF\x00\x82\xED\x01\x00\x81\xD5\x01\x00\x80\xDD"
-     "\x01\x00"
-     "\xBE\x9C\x01\x00\xA6\xB5\xFF\x00\xA5\x85\xFF\x00\x2F\xB1\x00\x80\xAB\xC1"
-     "\xFF\x00"
-     "\xAA\xF9\xFF\x00\x86\x88\x00\x00\x87\x7C\x01\x00\xAF\xC1\xFF\x00\xAE\xD9"
-     "\xFF\x00"
-     "\xAD\xD1\xFF\x00\xAC\xD9\xFF\x00\x3A\xB1\x00\x80\xB3\xF9\xFE\x00\x3E\xB1"
-     "\x00\x80"
-     "\x42\xB1\x00\x80\xB6\x9D\xFE\x00\x46\xB1\x00\x80\x4A\xB1\x00\x80\xB5\x95"
-     "\xFE\x00"
-     "\xBA\x69\x01\x00\xBB\x69\x01\x00\x4E\xB1\x00\x80\x52\xB1\x00\x80\xBE\x69"
-     "\x01\x00"
-     "\xBF\x69\x01\x00\xBC\x69\x01\x00\xBD\x69\x01\x00\xA8\x29\xFE\x00\xA9\x35"
-     "\xFE\x00"
-     "\xAA\x35\xFE\x00\xAB\x4D\xFE\x00\xAC\x59\xFE\x00\xAD\x59\xFE\x00\xAE\x4D"
-     "\xFE\x00"
-     "\xAF\x45\xFE\x00\x84\x8C\x01\x00\x56\xB1\x00\x80\x5A\xB1\x00\x80\x5E\xB1"
-     "\x00\x80"
-     "\x62\xB1\x00\x80\x66\xB1\x00\x80\x6A\xB1\x00\x80\x6E\xB1\x00\x80\xB8\x65"
-     "\x01\x00"
-     "\xB9\x6D\x01\x00\xBA\x79\x01\x00\xBB\x75\x01\x00\xBC\x6D\x01\x00\xBD\x11"
-     "\x01\x00"
-     "\xBE\x11\x01\x00\xBF\x11\x01\x00\xB0\xE9\x01\x00\xB1\xE9\x01\x00\xB2\xF9"
-     "\x01\x00"
-     "\xB3\xF9\x01\x00\xB4\xE9\x01\x00\xB5\xE9\x01\x00\xB6\x61\x01\x00\xB7\x61"
-     "\x01\x00"
-     "\xA3\xBD\xFD\x00\x72\xB1\x00\x80\x76\xB1\x00\x80\xBE\x48\x03\x00\x7A\xB1"
-     "\x00\x80"
-     "\xA6\xD9\xFD\x00\xA5\xD1\xFD\x00\x7E\xB1\x00\x80\xAB\x2D\x02\x00\xAA\x2D"
-     "\x02\x00"
-     "\x82\xB1\x00\x80\x86\xB1\x00\x80\xAF\x2D\x02\x00\xAE\x2D\x02\x00\xAD\x2D"
-     "\x02\x00"
-     "\xAC\x2D\x02\x00\x8A\xB1\x00\x80\x8E\xB1\x00\x80\x92\xB1\x00\x80\x96\xB1"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\x9A\xB1\x00\x80\x9E\xB1"
-     "\x00\x80"
-     "\xA2\xB1\x00\x80\x86\xC8\x04\x00\x87\x5C\x03\x00\xAA\xB1\x00\x80\xAE\xB1"
-     "\x00\x80"
-     "\xB2\xB1\x00\x80\xB6\xB1\x00\x80\x84\xCC\x04\x00\xBA\xB1\x00\x80\xBE\xB1"
-     "\x00\x80"
-     "\xC2\xB1\x00\x80\xE3\x20\x01\x00\xC6\xB1\x00\x80\xE1\xA8\x01\x00\xCA\xB1"
-     "\x00\x80"
-     "\xEF\x54\x02\x00\xCE\xB1\x00\x80\xD2\xB1\x00\x80\xD6\xB1\x00\x80\xDA\xB1"
-     "\x00\x80"
-     "\xDE\xB1\x00\x80\xE2\xB1\x00\x80\xE6\xB1\x00\x80\xBB\x6D\x03\x00\xBA\x6D"
-     "\x03\x00"
-     "\xEA\xB1\x00\x80\xEE\xB1\x00\x80\xBF\x6D\x03\x00\xBE\x6D\x03\x00\xBD\x6D"
-     "\x03\x00"
-     "\xBC\x6D\x03\x00\xB3\xAD\x03\x00\xF2\xB1\x00\x80\xF6\xB1\x00\x80\xFA\xB1"
-     "\x00\x80"
-     "\xFE\xB1\x00\x80\xB6\x9D\x03\x00\xB5\xA1\x03\x00\x84\x70\x05\x00\xA8\x41"
-     "\x02\x00"
-     "\xA9\x4D\x02\x00\xAA\x45\x02\x00\xAB\x59\x02\x00\xAC\x49\x02\x00\xAD\x7D"
-     "\x02\x00"
-     "\xAE\x75\x02\x00\xAF\xB9\x02\x00\xBE\x68\x05\x00\x02\xB2\x00\x80\x06\xB2"
-     "\x00\x80"
-     "\x0A\xB2\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xA9\x00\x00\x0E\xB2"
-     "\x00\x80"
-     "\xB8\x51\x01\x00\xB9\x59\x01\x00\xBA\x61\x01\x00\xBB\x61\x01\x00\xBC\x11"
-     "\x01\x00"
-     "\xBD\x11\x01\x00\xBE\x11\x01\x00\xBF\x11\x01\x00\xB0\xC9\x02\x00\xB1\xC9"
-     "\x02\x00"
-     "\xB2\xD9\x02\x00\xB3\xD9\x02\x00\xB4\xC9\x02\x00\xB5\xC9\x02\x00\xB6\x7D"
-     "\x01\x00"
-     "\xB7\x75\x01\x00\xE1\x70\xFE\x00\xE1\x4C\xFF\x00\xE3\x74\x01\x00\xE3\x50"
-     "\xFF\x00"
-     "\x12\xB2\x00\x80\x1A\xB2\x00\x80\x86\xB8\x04\x00\x87\x54\x05\x00\x1E\xB2"
-     "\x00\x80"
-     "\x22\xB2\x00\x80\x26\xB2\x00\x80\x2A\xB2\x00\x80\xE4\x00\x00\x00\x2E\xB2"
-     "\x00\x80"
-     "\xEF\x50\x00\x00\xEF\x6C\xFE\x00\x32\xB2\x00\x80\x36\xB2\x00\x80\xA3\x85"
-     "\x02\x00"
-     "\x3A\xB2\x00\x80\xA5\x89\x02\x00\x3E\xB2\x00\x80\x42\xB2\x00\x80\xA6\xB5"
-     "\x02\x00"
-     "\x46\xB2\x00\x80\x4A\xB2\x00\x80\xAB\x45\x02\x00\xAA\x45\x02\x00\xAD\x45"
-     "\x02\x00"
-     "\xAC\x45\x02\x00\xAF\x45\x02\x00\xAE\x45\x02\x00\xA8\x71\x06\x00\xA9\x71"
-     "\x06\x00"
-     "\xAA\x71\x06\x00\xAB\x71\x06\x00\xAC\x95\x06\x00\xAD\x9D\x06\x00\xAE\x95"
-     "\x06\x00"
-     "\xAF\x89\x06\x00\x16\xB2\x00\x80\x4E\xB2\x00\x80\x52\xB2\x00\x80\x56\xB2"
-     "\x00\x80"
-     "\x5A\xB2\x00\x80\x5E\xB2\x00\x80\x62\xB2\x00\x80\x66\xB2\x00\x80\xB8\x89"
-     "\x06\x00"
-     "\xB9\x89\x06\x00\xBA\xAD\x06\x00\xBB\xA5\x06\x00\xBC\xBD\x06\x00\xBD\xA5"
-     "\x06\x00"
-     "\xBE\x7D\x07\x00\xBF\x75\x07\x00\xB0\xFD\x06\x00\xB1\xC1\x06\x00\xB2\xC1"
-     "\x06\x00"
-     "\xB3\xC1\x06\x00\xB4\xC1\x06\x00\xB5\xC1\x06\x00\xB6\xC1\x06\x00\xB7\xC1"
-     "\x06\x00"
-     "\xB3\xF9\x06\x00\x6A\xB2\x00\x80\x6E\xB2\x00\x80\x72\xB2\x00\x80\x76\xB2"
-     "\x00\x80"
-     "\xB6\x1D\x06\x00\xB5\x1D\x06\x00\x7A\xB2\x00\x80\xBB\x05\x06\x00\xBA\x39"
-     "\x06\x00"
-     "\x7E\xB2\x00\x80\x82\xB2\x00\x80\xBF\x05\x06\x00\xBE\x05\x06\x00\xBD\x15"
-     "\x06\x00"
-     "\xBC\x15\x06\x00\x82\x6D\x00\x00\xA3\xBD\x06\x00\x80\x55\x00\x00\x81\x65"
-     "\x00\x00"
-     "\xA6\x59\x06\x00\xBE\x54\x00\x00\x86\xB2\x00\x80\xA5\x59\x06\x00\xAA\x7D"
-     "\x06\x00"
-     "\xAB\x41\x06\x00\x86\x00\x0C\x00\x87\x5C\x03\x00\xAE\x41\x06\x00\xAF\x41"
-     "\x06\x00"
-     "\xAC\x51\x06\x00\xAD\x51\x06\x00\x8A\xB2\x00\x80\xEF\xDC\x04\x00\x8E\xB2"
-     "\x00\x80"
-     "\x92\xB2\x00\x80\x96\xB2\x00\x80\x9A\xB2\x00\x80\x9E\xB2\x00\x80\xA2\xB2"
-     "\x00\x80"
-     "\xA6\xB2\x00\x80\xAA\xB2\x00\x80\xAE\xB2\x00\x80\xB2\xB2\x00\x80\xB6\xB2"
-     "\x00\x80"
-     "\xE3\xA4\x05\x00\xBA\xB2\x00\x80\xE1\x4C\x05\x00\xB3\xE5\x06\x00\xBE\xB2"
-     "\x00\x80"
-     "\xC2\xB2\x00\x80\xC6\xB2\x00\x80\xCA\xB2\x00\x80\xB6\xE1\x06\x00\xB5\xF5"
-     "\x06\x00"
-     "\xCE\xB2\x00\x80\xBB\xB5\x06\x00\xBA\xA9\x06\x00\xD2\xB2\x00\x80\xD6\xB2"
-     "\x00\x80"
-     "\xBF\x55\x01\x00\xBE\x91\x06\x00\xBD\x99\x06\x00\xBC\xA5\x06\x00\xA8\x15"
-     "\x06\x00"
-     "\xA9\x25\x06\x00\xAA\x2D\x06\x00\xAB\x25\x06\x00\xAC\x3D\x06\x00\xAD\x21"
-     "\x06\x00"
-     "\xAE\x59\x06\x00\xAF\x59\x06\x00\xDA\xB2\x00\x80\xDE\xB2\x00\x80\xE2\xB2"
-     "\x00\x80"
-     "\xE6\xB2\x00\x80\x80\x0D\x00\x00\x81\xB5\x03\x00\x82\xBD\x03\x00\xEA\xB2"
-     "\x00\x80"
-     "\xB8\xF1\x01\x00\xB9\xF1\x01\x00\xBA\xF1\x01\x00\xBB\xF1\x01\x00\xBC\x55"
-     "\x01\x00"
-     "\xBD\x5D\x01\x00\xBE\x55\x01\x00\xBF\x4D\x01\x00\xB0\x2D\x06\x00\xB1\x31"
-     "\x06\x00"
-     "\xB2\x31\x06\x00\xB3\x0D\x06\x00\xB4\x15\x06\x00\xB5\xD1\x01\x00\xB6\xD1"
-     "\x01\x00"
-     "\xB7\xD1\x01\x00\xEE\xB2\x00\x80\xF2\xB2\x00\x80\x84\x90\x02\x00\xA3\xA9"
-     "\x05\x00"
-     "\xBE\x0C\x02\x00\xA5\xB9\x05\x00\xA6\xAD\x05\x00\xFA\xB2\x00\x80\xFE\xB2"
-     "\x00\x80"
-     "\x02\xB3\x00\x80\xAA\xE5\x05\x00\xAB\xF9\x05\x00\xAC\xE9\x05\x00\xAD\xD5"
-     "\x05\x00"
-     "\xAE\xDD\x05\x00\xAF\x19\x02\x00\xB6\x59\x03\x00\x06\xB3\x00\x80\x84\x8C"
-     "\x03\x00"
-     "\xB5\x59\x03\x00\x0A\xB3\x00\x80\xB3\x59\x03\x00\x86\xD8\x0C\x00\x87\xAC"
-     "\x02\x00"
-     "\xBE\x59\x03\x00\xBF\x21\x03\x00\xBC\x45\x03\x00\xBD\x5D\x03\x00\xBA\x55"
-     "\x03\x00"
-     "\xBB\x5D\x03\x00\x0E\xB3\x00\x80\x12\xB3\x00\x80\x16\xB3\x00\x80\xA6\x9D"
-     "\x03\x00"
-     "\xA5\x9D\x03\x00\x1A\xB3\x00\x80\xA3\x9D\x03\x00\x1E\xB3\x00\x80\x22\xB3"
-     "\x00\x80"
-     "\x26\xB3\x00\x80\xAF\xE5\x03\x00\xAE\x9D\x03\x00\xAD\x99\x03\x00\xAC\x81"
-     "\x03\x00"
-     "\xAB\x99\x03\x00\xAA\x91\x03\x00\xA6\xB1\x00\x80\x2A\xB3\x00\x80\x2E\xB3"
-     "\x00\x80"
-     "\x32\xB3\x00\x80\x85\x8C\x08\x00\x36\xB3\x00\x80\x3A\xB3\x00\x80\x3E\xB3"
-     "\x00\x80"
-     "\x42\xB3\x00\x80\x46\xB3\x00\x80\x4A\xB3\x00\x80\x4E\xB3\x00\x80\x52\xB3"
-     "\x00\x80"
-     "\x56\xB3\x00\x80\x80\x0D\x00\x00\x81\xF5\x01\x00\x82\xFD\x01\x00\x5A\xB3"
-     "\x00\x80"
-     "\x5E\xB3\x00\x80\x62\xB3\x00\x80\xB3\x2D\x01\x00\xBE\x4C\x02\x00\xB5\x41"
-     "\x01\x00"
-     "\x6A\xB3\x00\x80\x6E\xB3\x00\x80\xB6\x65\x01\x00\x87\x98\x02\x00\x86\x04"
-     "\x0D\x00"
-     "\xBB\xBD\x01\x00\xBA\xA1\x01\x00\xBD\xAD\x01\x00\xBC\xAD\x01\x00\xBF\xF9"
-     "\x01\x00"
-     "\xBE\x81\x01\x00\x72\xB3\x00\x80\x76\xB3\x00\x80\x7A\xB3\x00\x80\x7E\xB3"
-     "\x00\x80"
-     "\x82\xB3\x00\x80\xE3\x4C\x03\x00\x86\xB3\x00\x80\xE1\xAC\x01\x00\x8A\xB3"
-     "\x00\x80"
-     "\xEF\x8C\x03\x00\x8E\xB3\x00\x80\x92\xB3\x00\x80\x96\xB3\x00\x80\x9A\xB3"
-     "\x00\x80"
-     "\x9E\xB3\x00\x80\xA2\xB3\x00\x80\xA6\xB3\x00\x80\xAA\xB3\x00\x80\xAE\xB3"
-     "\x00\x80"
-     "\x84\x40\x0C\x00\xB2\xB3\x00\x80\xB6\xB3\x00\x80\xBA\xB3\x00\x80\xEF\xB0"
-     "\x01\x00"
-     "\xBE\xB3\x00\x80\xE1\x24\x06\x00\xC2\xB3\x00\x80\xE3\x40\x07\x00\x82\x79"
-     "\x00\x00"
-     "\xC6\xB3\x00\x80\x80\x79\x00\x00\x81\x79\x00\x00\xCA\xB3\x00\x80\xCE\xB3"
-     "\x00\x80"
-     "\xD2\xB3\x00\x80\xA3\xE9\x01\x00\xDA\xB3\x00\x80\xA5\x85\x01\x00\xA6\xA1"
-     "\x01\x00"
-     "\xDE\xB3\x00\x80\x86\xC0\x0C\x00\x87\x44\x0C\x00\xAA\x65\x01\x00\xAB\x79"
-     "\x01\x00"
-     "\xAC\x69\x01\x00\xAD\x69\x01\x00\xAE\x45\x01\x00\xAF\x3D\x01\x00\xF6\xB2"
-     "\x00\x80"
-     "\x66\xB3\x00\x80\xBE\xC0\x03\x00\xE2\xB3\x00\x80\xD6\xB3\x00\x80\xE6\xB3"
-     "\x00\x80"
-     "\xEA\xB3\x00\x80\xEE\xB3\x00\x80\xA8\x3D\x0E\x00\xA9\x05\x0E\x00\xAA\x0D"
-     "\x0E\x00"
-     "\xAB\x69\x0E\x00\xAC\x61\x0E\x00\xAD\x4D\x0E\x00\xAE\x45\x0E\x00\xAF\xBD"
-     "\x0E\x00"
-     "\xB0\xCD\x0E\x00\xB1\xD1\x0E\x00\xB2\xD1\x0E\x00\xB3\xD1\x0E\x00\xB4\xED"
-     "\x0E\x00"
-     "\xB5\x95\x0E\x00\xB6\x9D\x0E\x00\xB7\x91\x0E\x00\xB8\xB1\x0E\x00\xB9\xB1"
-     "\x0E\x00"
-     "\xBA\xB1\x0E\x00\xBB\xB1\x0E\x00\xBC\x91\x0E\x00\xBD\x91\x0E\x00\xBE\x91"
-     "\x0E\x00"
-     "\xBF\x91\x0E\x00\xA8\x95\x0F\x00\xA9\x9D\x0F\x00\xAA\xC9\x0F\x00\xAB\xC9"
-     "\x0F\x00"
-     "\xAC\xD9\x0F\x00\xAD\xD9\x0F\x00\xAE\xC9\x0F\x00\xAF\xC5\x0F\x00\xF2\xB3"
-     "\x00\x80"
-     "\xF6\xB3\x00\x80\xFA\xB3\x00\x80\xFE\xB3\x00\x80\x02\xB4\x00\x80\x06\xB4"
-     "\x00\x80"
-     "\x0A\xB4\x00\x80\x0E\xB4\x00\x80\xB8\x45\x0F\x00\xB9\x4D\x0F\x00\xBA\x59"
-     "\x0F\x00"
-     "\xBB\x59\x0F\x00\xBC\x49\x0F\x00\xBD\x75\x0F\x00\xBE\x71\x0F\x00\xBF\x69"
-     "\x0F\x00"
-     "\xB0\x81\x0F\x00\xB1\x8D\x0F\x00\xB2\x85\x0F\x00\xB3\x9D\x0F\x00\xB4\x85"
-     "\x0F\x00"
-     "\xB5\x8D\x0F\x00\xB6\x85\x0F\x00\xB7\x7D\x0F\x00\xB3\x41\x0E\x00\x12\xB4"
-     "\x00\x80"
-     "\x16\xB4\x00\x80\x1A\xB4\x00\x80\x1E\xB4\x00\x80\xB6\x59\x0E\x00\xB5\x51"
-     "\x0E\x00"
-     "\x22\xB4\x00\x80\xBB\x49\x0E\x00\xBA\x45\x0E\x00\x26\xB4\x00\x80\x2A\xB4"
-     "\x00\x80"
-     "\xBF\x49\x0E\x00\xBE\x41\x0E\x00\xBD\x49\x0E\x00\xBC\x51\x0E\x00\x2E\xB4"
-     "\x00\x80"
-     "\xA3\x05\x0E\x00\x32\xB4\x00\x80\x36\xB4\x00\x80\xA6\x1D\x0E\x00\x3A\xB4"
-     "\x00\x80"
-     "\x3E\xB4\x00\x80\xA5\x15\x0E\x00\xAA\x01\x0E\x00\xAB\x0D\x0E\x00\x42\xB4"
-     "\x00\x80"
-     "\x84\x20\x01\x00\xAE\x05\x0E\x00\xAF\x0D\x0E\x00\xAC\x15\x0E\x00\xAD\x0D"
-     "\x0E\x00"
-     "\xA8\x51\x0E\x00\xA9\x59\x0E\x00\xAA\x85\x01\x00\xAB\x99\x01\x00\xAC\x89"
-     "\x01\x00"
-     "\xAD\x89\x01\x00\xAE\x39\x01\x00\xAF\x39\x01\x00\x80\xF5\x00\x00\x81\xFD"
-     "\x00\x00"
-     "\x82\xC1\x00\x00\xBE\xE4\x01\x00\x86\x10\x00\x00\x87\x94\x00\x00\x4A\xB4"
-     "\x00\x80"
-     "\x4E\xB4\x00\x80\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5\x00\x00\xBB\xE5"
-     "\x00\x00"
-     "\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x91\x00\x00\xBF\x91\x00\x00\xB0\x49"
-     "\x01\x00"
-     "\xB1\x49\x01\x00\xB2\x59\x01\x00\xB3\x59\x01\x00\xB4\x49\x01\x00\xB5\x49"
-     "\x01\x00"
-     "\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xBA\x69\x0F\x00\xBB\x69\x0F\x00\xB8\x59"
-     "\x0F\x00"
-     "\xB9\x51\x0F\x00\xBE\xB9\x0F\x00\xBF\xB9\x0F\x00\xBC\xB9\x0F\x00\xBD\xB9"
-     "\x0F\x00"
-     "\xB2\x0D\x0F\x00\xB3\x15\x0F\x00\xB0\x0D\x0F\x00\xB1\x05\x0F\x00\xB6\x79"
-     "\x0F\x00"
-     "\xB7\x79\x0F\x00\xB4\x01\x0F\x00\xB5\x79\x0F\x00\xAA\x4D\x0F\x00\xAB\x55"
-     "\x0F\x00"
-     "\xA8\x49\x0F\x00\xA9\x45\x0F\x00\xAE\x4D\x0F\x00\xAF\x75\x0F\x00\xAC\x4D"
-     "\x0F\x00"
-     "\xAD\x45\x0F\x00\x52\xB4\x00\x80\x56\xB4\x00\x80\x5A\xB4\x00\x80\x5E\xB4"
-     "\x00\x80"
-     "\x62\xB4\x00\x80\x66\xB4\x00\x80\x6A\xB4\x00\x80\x6E\xB4\x00\x80\xB3\x8D"
-     "\x02\x00"
-     "\x72\xB4\x00\x80\x76\xB4\x00\x80\x7A\xB4\x00\x80\x7E\xB4\x00\x80\xB6\x89"
-     "\x02\x00"
-     "\xB5\x81\x02\x00\x82\xB4\x00\x80\xBB\x0D\x02\x00\xBA\x0D\x02\x00\x86\xB4"
-     "\x00\x80"
-     "\x8A\xB4\x00\x80\xBF\x0D\x02\x00\xBE\x0D\x02\x00\xBD\x0D\x02\x00\xBC\x0D"
-     "\x02\x00"
-     "\x8E\xB4\x00\x80\xA3\xC9\x02\x00\x92\xB4\x00\x80\x96\xB4\x00\x80\xA6\xCD"
-     "\x02\x00"
-     "\x9A\xB4\x00\x80\x9E\xB4\x00\x80\xA5\xC5\x02\x00\xAA\x49\x02\x00\xAB\x49"
-     "\x02\x00"
-     "\xA2\xB4\x00\x80\xA6\xB4\x00\x80\xAE\x49\x02\x00\xAF\x49\x02\x00\xAC\x49"
-     "\x02\x00"
-     "\xAD\x49\x02\x00\xA8\xB9\x02\x00\xA9\xB9\x02\x00\xAA\x69\x01\x00\xAB\x69"
-     "\x01\x00"
-     "\xAC\x79\x01\x00\xAD\x79\x01\x00\xAE\x69\x01\x00\xAF\x69\x01\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x1D\x00\x00\xAA\xB4\x00\x80\xAE\xB4\x00\x80\xB2\xB4"
-     "\x00\x80"
-     "\xBE\xB4\x05\x00\xBA\xB4\x00\x80\xB8\xED\x01\x00\xB9\x85\x01\x00\xBA\x81"
-     "\x01\x00"
-     "\xBB\x81\x01\x00\xBC\x85\x01\x00\xBD\x8D\x01\x00\xBE\xB1\x01\x00\xBF\xB1"
-     "\x01\x00"
-     "\xB0\x19\x01\x00\xB1\x19\x01\x00\xB2\xED\x01\x00\xB3\xE5\x01\x00\xB4\xFD"
-     "\x01\x00"
-     "\xB5\xE5\x01\x00\xB6\xE5\x01\x00\xB7\xD5\x01\x00\xE1\xBC\x0E\x00\xE1\xC4"
-     "\x0F\x00"
-     "\xE3\x14\x0E\x00\xE3\xD4\x0F\x00\x84\x90\x02\x00\xE1\x6C\x0E\x00\xEF\xCC"
-     "\x06\x00"
-     "\xE3\xE0\x00\x00\xE1\x7C\x0F\x00\xBE\xB4\x00\x80\xE3\x7C\x13\x00\xEF\xF8"
-     "\x00\x00"
-     "\x86\x90\x04\x00\x87\x14\x04\x00\xEF\x48\x01\x00\xEF\x48\x0E\x00\xC2\xB4"
-     "\x00\x80"
-     "\xC6\xB4\x00\x80\x84\x04\x07\x00\xB3\xAD\x01\x00\xCA\xB4\x00\x80\xB5\xBD"
-     "\x01\x00"
-     "\xB6\xB5\x01\x00\xCE\xB4\x00\x80\xD2\xB4\x00\x80\xD6\xB4\x00\x80\xBA\x91"
-     "\x01\x00"
-     "\xBB\x91\x01\x00\xBC\x75\x01\x00\xBD\x7D\x01\x00\xBE\x75\x01\x00\xBF\x6D"
-     "\x01\x00"
-     "\xA3\x61\x06\x00\x8F\x09\xFC\x00\xB6\xB4\x00\x80\xDA\xB4\x00\x80\xDE\xB4"
-     "\x00\x80"
-     "\xA6\x79\x06\x00\xA5\x71\x06\x00\xE2\xB4\x00\x80\xAB\x5D\x06\x00\xAA\x5D"
-     "\x06\x00"
-     "\xE6\xB4\x00\x80\xEA\xB4\x00\x80\xAF\xA1\x06\x00\xAE\xB9\x06\x00\xAD\xB1"
-     "\x06\x00"
-     "\xAC\xB9\x06\x00\x9E\x61\x08\x00\x9F\x9D\x08\x00\x9C\x55\x0D\x00\x9D\xF5"
-     "\x0A\x00"
-     "\x9A\x51\x0C\x00\x9B\x51\x0C\x00\x98\x51\xF0\x00\x99\x19\xF1\x00\x96\xC5"
-     "\xF2\x00"
-     "\x97\x61\xF0\x00\x94\x6D\xF4\x00\x95\x19\xF4\x00\x92\x61\xF8\x00\x93\xE5"
-     "\xF6\x00"
-     "\x90\x51\xFE\x00\x91\x79\xF8\x00\xB1\xB1\x2C\x00\xB0\xA5\x2C\x00\xB3\x89"
-     "\x28\x00"
-     "\xB2\xE1\x2D\x00\xB5\xB1\x29\x00\xB4\x91\x28\x00\xEE\xB4\x00\x80\xF2\xB4"
-     "\x00\x80"
-     "\xF6\xB4\x00\x80\xFA\xB4\x00\x80\xFE\xB4\x00\x80\x02\xB5\x00\x80\xA2\xA1"
-     "\x04\x00"
-     "\xA3\x65\x02\x00\xA0\xB9\x04\x00\xA1\xB9\x04\x00\xA1\xC9\x00\x00\x06\xB5"
-     "\x00\x80"
-     "\xA3\xC9\x1C\x00\xA2\x01\x1C\x00\xA5\x15\x1D\x00\xA4\xC5\x1C\x00\xA7\xF9"
-     "\x18\x00"
-     "\xA6\x01\x18\x00\xA9\x01\x14\x00\xA8\xC9\x19\x00\xAB\x41\x14\x00\xAA\xCD"
-     "\x14\x00"
-     "\xAD\xD1\x10\x00\xAC\x01\x10\x00\xAF\x01\x2C\x00\xAE\xCD\x10\x00\x8A\x75"
-     "\xF5\x00"
-     "\x8B\xA1\xF3\x00\x0A\xB5\x00\x80\x0E\xB5\x00\x80\x8E\x55\xF0\x00\x8F\x15"
-     "\x0D\x00"
-     "\x8C\x1D\xF1\x00\x8D\x15\xF1\x00\x82\x2D\xFD\x00\x83\x0D\xFF\x00\x12\xB5"
-     "\x00\x80"
-     "\x16\xB5\x00\x80\x86\xA1\xF7\x00\x87\x31\xF5\x00\x84\x3D\xF9\x00\x85\x2D"
-     "\xF9\x00"
-     "\x9A\x45\x05\x00\x9B\x55\x05\x00\x1A\xB5\x00\x80\x1E\xB5\x00\x80\x22\xB5"
-     "\x00\x80"
-     "\xE1\xFC\x06\x00\x9C\x89\x03\x00\xE3\xDC\x04\x00\x92\x65\x0D\x00\x93\x79"
-     "\x0C\x00"
-     "\xBE\xA0\x01\x00\xEF\xAC\x04\x00\x96\xA9\x09\x00\x97\x75\x05\x00\x94\xC1"
-     "\x0B\x00"
-     "\x95\x5D\x09\x00\xA8\xBD\x06\x00\xA9\x21\x06\x00\xAA\x49\x06\x00\xAB\x49"
-     "\x06\x00"
-     "\xAC\x41\x06\x00\xAD\x49\x06\x00\xAE\x79\x06\x00\xAF\x79\x06\x00\x80\xE9"
-     "\x00\x00"
-     "\x81\xE9\x00\x00\x82\xF9\x00\x00\x2A\xB5\x00\x80\x86\xF0\x01\x00\x87\xB4"
-     "\x00\x00"
-     "\x2E\xB5\x00\x80\x32\xB5\x00\x80\xB8\x65\x01\x00\xB9\x6D\x01\x00\xBA\x65"
-     "\x01\x00"
-     "\xBB\x09\x01\x00\xBC\xD9\x00\x00\xBD\xD9\x00\x00\xBE\xF1\x00\x00\xBF\xF1"
-     "\x00\x00"
-     "\xB0\xED\x01\x00\xB1\xF5\x01\x00\xB2\xF5\x01\x00\xB3\xC5\x01\x00\xB4\xDD"
-     "\x01\x00"
-     "\xB5\x6D\x01\x00\xB6\x65\x01\x00\xB7\x5D\x01\x00\xB3\xC1\x05\x00\x36\xB5"
-     "\x00\x80"
-     "\x3A\xB5\x00\x80\x3E\xB5\x00\x80\x42\xB5\x00\x80\xB6\xC9\x05\x00\xB5\xC9"
-     "\x05\x00"
-     "\x46\xB5\x00\x80\xBB\x31\x02\x00\xBA\x29\x02\x00\x4A\xB5\x00\x80\x4E\xB5"
-     "\x00\x80"
-     "\xBF\xA9\x02\x00\xBE\x19\x02\x00\xBD\x11\x02\x00\xBC\x21\x02\x00\x52\xB5"
-     "\x00\x80"
-     "\xA3\x85\x05\x00\x56\xB5\x00\x80\x5A\xB5\x00\x80\xA6\x8D\x05\x00\x5E\xB5"
-     "\x00\x80"
-     "\x62\xB5\x00\x80\xA5\x8D\x05\x00\xAA\x6D\x02\x00\xAB\x75\x02\x00\x66\xB5"
-     "\x00\x80"
-     "\x6A\xB5\x00\x80\xAE\x5D\x02\x00\xAF\xED\x02\x00\xAC\x65\x02\x00\xAD\x55"
-     "\x02\x00"
-     "\xB0\xA9\x00\x00\xB1\xA9\x00\x00\xB2\xC9\x00\x00\xB3\xC5\x00\x00\xB4\xDD"
-     "\x00\x00"
-     "\xB5\xC1\x00\x00\xB6\xC1\x00\x00\xB7\xF5\x00\x00\xB8\xCD\x00\x00\xB9\x65"
-     "\x00\x00"
-     "\xBA\x6D\x00\x00\xBB\x65\x00\x00\xBC\x7D\x00\x00\xBD\x61\x00\x00\xBE\x61"
-     "\x00\x00"
-     "\xBF\x61\x00\x00\x6E\xB5\x00\x80\x72\xB5\x00\x80\x76\xB5\x00\x80\x7A\xB5"
-     "\x00\x80"
-     "\x46\xB4\x00\x80\x7E\xB5\x00\x80\x82\xB5\x00\x80\x86\xB5\x00\x80\xA8\x4D"
-     "\x03\x00"
-     "\xA9\x5D\x03\x00\xAA\x55\x03\x00\xAB\xE9\x00\x00\xAC\xF9\x00\x00\xAD\xF9"
-     "\x00\x00"
-     "\xAE\xE9\x00\x00\xAF\xE1\x00\x00\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\x8A\xB5\x00\x80\x8E\xB5\x00\x80\x92\xB5\x00\x80\xBE\x5C\x02\x00\x9A\xB5"
-     "\x00\x80"
-     "\x84\xAC\x02\x00\x9E\xB5\x00\x80\x87\x24\x1D\x00\x86\xAC\x1D\x00\xA2\xB5"
-     "\x00\x80"
-     "\xA6\xB5\x00\x80\xAA\xB5\x00\x80\xAE\xB5\x00\x80\xB3\xFD\x02\x00\xB2\xB5"
-     "\x00\x80"
-     "\xB6\xB5\x00\x80\xBA\xB5\x00\x80\xBE\xB5\x00\x80\xB6\x59\x01\x00\xB5\x51"
-     "\x01\x00"
-     "\xC2\xB5\x00\x80\xBB\x4D\x01\x00\xBA\x4D\x01\x00\xC6\xB5\x00\x80\xCA\xB5"
-     "\x00\x80"
-     "\xBF\x25\x01\x00\xBE\x39\x01\x00\xBD\x31\x01\x00\xBC\x3D\x01\x00\xCE\xB5"
-     "\x00\x80"
-     "\xE3\x5C\x06\x00\xD2\xB5\x00\x80\xE1\x24\x06\x00\x84\xDC\x1D\x00\xD6\xB5"
-     "\x00\x80"
-     "\xDA\xB5\x00\x80\xDE\xB5\x00\x80\xBE\xE8\x1D\x00\xE2\xB5\x00\x80\xBF\xA0"
-     "\x24\x00"
-     "\xE6\xB5\x00\x80\xEA\xB5\x00\x80\xEF\x24\x00\x00\xEE\xB5\x00\x80\xF2\xB5"
-     "\x00\x80"
-     "\xF6\xB5\x00\x80\xFA\xB5\x00\x80\xA3\x75\x02\x00\xFE\xB5\x00\x80\xA5\xD9"
-     "\x01\x00"
-     "\x02\xB6\x00\x80\x0A\xB6\x00\x80\xA6\xD1\x01\x00\x86\x40\x1C\x00\x87\x64"
-     "\x1C\x00"
-     "\xAB\xC5\x01\x00\xAA\xC5\x01\x00\xAD\xB9\x01\x00\xAC\xB5\x01\x00\xAF\xAD"
-     "\x01\x00"
-     "\xAE\xB1\x01\x00\x96\xB5\x00\x80\x82\xD1\x1F\x00\x81\xC5\x1F\x00\x80\xC9"
-     "\x1F\x00"
-     "\x06\xB6\x00\x80\x0E\xB6\x00\x80\x12\xB6\x00\x80\x16\xB6\x00\x80\xA8\x3D"
-     "\x1F\x00"
-     "\xA9\x01\x1F\x00\xAA\x19\x1F\x00\xAB\x19\x1F\x00\xAC\x09\x1F\x00\xAD\x09"
-     "\x1F\x00"
-     "\xAE\x7D\x1F\x00\xAF\x75\x1F\x00\xB0\x0D\x1F\x00\xB1\x15\x1F\x00\xB2\x15"
-     "\x1F\x00"
-     "\xB3\x25\x1F\x00\xB4\x3D\x1F\x00\xB5\xD5\x1F\x00\xB6\xD1\x1F\x00\xB7\xC9"
-     "\x1F\x00"
-     "\xB8\xF9\x1F\x00\xB9\xF9\x1F\x00\xBA\xCD\x1F\x00\xBB\xC1\x1F\x00\xBC\xC1"
-     "\x1F\x00"
-     "\xBD\xCD\x1F\x00\xBE\xC5\x1F\x00\xBF\xBD\x1F\x00\xB3\xB1\x1F\x00\x1A\xB6"
-     "\x00\x80"
-     "\x1E\xB6\x00\x80\x22\xB6\x00\x80\x26\xB6\x00\x80\xB6\xA9\x1F\x00\xB5\xA1"
-     "\x1F\x00"
-     "\x2A\xB6\x00\x80\xBB\xF1\x1F\x00\xBA\xE9\x1F\x00\x2E\xB6\x00\x80\x32\xB6"
-     "\x00\x80"
-     "\xBF\x31\x1F\x00\xBE\xD9\x1F\x00\xBD\xD1\x1F\x00\xBC\xE1\x1F\x00\x36\xB6"
-     "\x00\x80"
-     "\xA3\xF5\x1F\x00\x3A\xB6\x00\x80\x26\xB5\x00\x80\xA6\xED\x1F\x00\x3E\xB6"
-     "\x00\x80"
-     "\x42\xB6\x00\x80\xA5\xE5\x1F\x00\xAA\xAD\x1F\x00\xAB\xB5\x1F\x00\x46\xB6"
-     "\x00\x80"
-     "\x4A\xB6\x00\x80\xAE\x9D\x1F\x00\xAF\x75\x1F\x00\xAC\xA5\x1F\x00\xAD\x95"
-     "\x1F\x00"
-     "\xA8\xD9\x1E\x00\xA9\x11\x01\x00\xAA\x11\x01\x00\xAB\x11\x01\x00\xAC\x95"
-     "\x01\x00"
-     "\xAD\x9D\x01\x00\xAE\x95\x01\x00\xAF\x8D\x01\x00\x4E\xB6\x00\x80\x52\xB6"
-     "\x00\x80"
-     "\x56\xB6\x00\x80\x5A\xB6\x00\x80\x5E\xB6\x00\x80\x62\xB6\x00\x80\x66\xB6"
-     "\x00\x80"
-     "\x6A\xB6\x00\x80\xB8\xAD\x01\x00\xB9\x65\x00\x00\xBA\x6D\x00\x00\xBB\x65"
-     "\x00\x00"
-     "\xBC\x7D\x00\x00\xBD\x65\x00\x00\xBE\x6D\x00\x00\xBF\x65\x00\x00\xB0\xFD"
-     "\x01\x00"
-     "\xB1\xCD\x01\x00\xB2\xC5\x01\x00\xB3\xAD\x01\x00\xB4\xB9\x01\x00\xB5\xA1"
-     "\x01\x00"
-     "\xB6\xA1\x01\x00\xB7\x95\x01\x00\x80\xBD\x00\x00\x81\x09\x00\x00\x82\x19"
-     "\x00\x00"
-     "\x6E\xB6\x00\x80\x72\xB6\x00\x80\x84\x98\x02\x00\xBE\x18\x01\x00\x7A\xB6"
-     "\x00\x80"
-     "\x86\x80\x01\x00\x87\xC4\x01\x00\x7E\xB6\x00\x80\x82\xB6\x00\x80\x86\xB6"
-     "\x00\x80"
-     "\x8A\xB6\x00\x80\x8E\xB6\x00\x80\x92\xB6\x00\x80\xB3\xD9\x01\x00\x76\xB6"
-     "\x00\x80"
-     "\x96\xB6\x00\x80\x9A\xB6\x00\x80\x9E\xB6\x00\x80\xB6\x5D\x01\x00\xB5\x5D"
-     "\x01\x00"
-     "\xA2\xB6\x00\x80\xBB\x49\x01\x00\xBA\x79\x01\x00\xA6\xB6\x00\x80\xAA\xB6"
-     "\x00\x80"
-     "\xBF\x25\x01\x00\xBE\x3D\x01\x00\xBD\x39\x01\x00\xBC\x51\x01\x00\xAE\xB6"
-     "\x00\x80"
-     "\xB2\xB6\x00\x80\xB6\xB6\x00\x80\xA3\x15\x02\x00\xBA\xB6\x00\x80\xA5\x91"
-     "\x02\x00"
-     "\xA6\x91\x02\x00\xBE\xB6\x00\x80\xC2\xB6\x00\x80\xC6\xB6\x00\x80\xAA\xB5"
-     "\x02\x00"
-     "\xAB\x85\x02\x00\xAC\x9D\x02\x00\xAD\xF5\x02\x00\xAE\xF1\x02\x00\xAF\xE9"
-     "\x02\x00"
-     "\xE1\xF8\x01\x00\xCA\xB6\x00\x80\xE3\x14\x02\x00\x84\x8C\x03\x00\xCE\xB6"
-     "\x00\x80"
-     "\xD2\xB6\x00\x80\xD6\xB6\x00\x80\xDA\xB6\x00\x80\x80\x3D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\xDE\xB6\x00\x80\xE2\xB6\x00\x80\xEA\xB6\x00\x80\xEF\x20"
-     "\x02\x00"
-     "\xEE\xB6\x00\x80\xA8\x1D\x03\x00\xA9\x2D\x03\x00\xAA\x25\x03\x00\xAB\x3D"
-     "\x03\x00"
-     "\xAC\x25\x03\x00\xAD\x2D\x03\x00\xAE\x25\x03\x00\xAF\x55\x03\x00\xBE\xEC"
-     "\x03\x00"
-     "\xF2\xB6\x00\x80\x86\x08\x04\x00\x87\x0C\x03\x00\xF6\xB6\x00\x80\xFA\xB6"
-     "\x00\x80"
-     "\xFE\xB6\x00\x80\x02\xB7\x00\x80\xB8\xD1\x00\x00\xB9\xD9\x00\x00\xBA\xE9"
-     "\x00\x00"
-     "\xBB\xE9\x00\x00\xBC\x95\x00\x00\xBD\x85\x00\x00\xBE\x8D\x00\x00\xBF\x85"
-     "\x00\x00"
-     "\xB0\x2D\x03\x00\xB1\x35\x03\x00\xB2\x35\x03\x00\xB3\x05\x03\x00\xB4\x1D"
-     "\x03\x00"
-     "\xB5\xF5\x00\x00\xB6\xFD\x00\x00\xB7\xF5\x00\x00\x06\xB7\x00\x80\x0A\xB7"
-     "\x00\x80"
-     "\x0E\xB7\x00\x80\xB3\xA1\x02\x00\x12\xB7\x00\x80\xB5\xA1\x02\x00\xB6\xA1"
-     "\x02\x00"
-     "\x16\xB7\x00\x80\x1A\xB7\x00\x80\x1E\xB7\x00\x80\xBA\xE5\x02\x00\xBB\xFD"
-     "\x02\x00"
-     "\xBC\xED\x02\x00\xBD\xDD\x02\x00\xBE\xD5\x02\x00\xBF\x3D\x01\x00\xE3\x70"
-     "\x1E\x00"
-     "\x22\xB7\x00\x80\xE1\xE0\x1C\x00\x26\xB7\x00\x80\x2A\xB7\x00\x80\x2E\xB7"
-     "\x00\x80"
-     "\x32\xB7\x00\x80\x36\xB7\x00\x80\x84\x2C\x05\x00\x3A\xB7\x00\x80\x3E\xB7"
-     "\x00\x80"
-     "\x42\xB7\x00\x80\xEF\x90\x1E\x00\x46\xB7\x00\x80\x4A\xB7\x00\x80\x4E\xB7"
-     "\x00\x80"
-     "\xBE\xC0\x05\x00\xE1\xB0\x1E\x00\x52\xB7\x00\x80\xE3\x0C\x1E\x00\x5A\xB7"
-     "\x00\x80"
-     "\xE1\x6C\x1E\x00\x5E\xB7\x00\x80\xE3\x18\x00\x00\x80\xD9\x00\x00\x81\xD9"
-     "\x00\x00"
-     "\x82\x21\x00\x00\xEF\xCC\x00\x00\x62\xB7\x00\x80\x66\xB7\x00\x80\xBE\xF0"
-     "\x04\x00"
-     "\xEF\xE4\x1E\x00\xA3\x6D\x02\x00\x6A\xB7\x00\x80\x87\xE8\x04\x00\x86\x6C"
-     "\x04\x00"
-     "\x6E\xB7\x00\x80\xA6\x6D\x02\x00\xA5\x6D\x02\x00\x72\xB7\x00\x80\xAB\x31"
-     "\x02\x00"
-     "\xAA\x29\x02\x00\x76\xB7\x00\x80\x7A\xB7\x00\x80\xAF\xF1\x01\x00\xAE\x19"
-     "\x02\x00"
-     "\xAD\x11\x02\x00\xAC\x21\x02\x00\xA8\x75\x06\x00\xA9\x79\x06\x00\xAA\xAD"
-     "\x06\x00"
-     "\xAB\xBD\x06\x00\xAC\xA5\x06\x00\xAD\xAD\x06\x00\xAE\xA5\x06\x00\xAF\x1D"
-     "\x07\x00"
-     "\x56\xB7\x00\x80\x7E\xB7\x00\x80\x82\xB7\x00\x80\x86\xB7\x00\x80\x8A\xB7"
-     "\x00\x80"
-     "\x8E\xB7\x00\x80\x92\xB7\x00\x80\x96\xB7\x00\x80\xB8\x11\x07\x00\xB9\x11"
-     "\x07\x00"
-     "\xBA\x11\x07\x00\xBB\x11\x07\x00\xBC\x35\x07\x00\xBD\x3D\x07\x00\xBE\x35"
-     "\x07\x00"
-     "\xBF\x29\x07\x00\xB0\x65\x07\x00\xB1\x6D\x07\x00\xB2\x79\x07\x00\xB3\x71"
-     "\x07\x00"
-     "\xB4\x51\x07\x00\xB5\x51\x07\x00\xB6\x3D\x07\x00\xB7\x31\x07\x00\xB3\x25"
-     "\x06\x00"
-     "\x9A\xB7\x00\x80\x9E\xB7\x00\x80\xA2\xB7\x00\x80\xA6\xB7\x00\x80\xB6\x2D"
-     "\x06\x00"
-     "\xB5\x2D\x06\x00\xAA\xB7\x00\x80\xBB\xA9\x07\x00\xBA\xA1\x07\x00\xAE\xB7"
-     "\x00\x80"
-     "\xB2\xB7\x00\x80\xBF\x95\x07\x00\xBE\x95\x07\x00\xBD\xA5\x07\x00\xBC\xAD"
-     "\x07\x00"
-     "\xB6\xB7\x00\x80\xBA\xB7\x00\x80\xBE\xB7\x00\x80\xA3\x69\x06\x00\xC2\xB7"
-     "\x00\x80"
-     "\xA5\x61\x06\x00\xA6\x61\x06\x00\xE6\xB6\x00\x80\x84\xAC\x03\x00\xCA\xB7"
-     "\x00\x80"
-     "\xAA\xED\x07\x00\xAB\xE5\x07\x00\xAC\xE1\x07\x00\xAD\xE9\x07\x00\xAE\xD9"
-     "\x07\x00"
-     "\xAF\xD9\x07\x00\xA8\x91\x06\x00\xA9\x9D\x06\x00\xAA\xB9\x06\x00\xAB\xB9"
-     "\x06\x00"
-     "\xAC\xD5\x06\x00\xAD\xC5\x06\x00\xAE\xC5\x06\x00\xAF\xF5\x06\x00\x80\xE5"
-     "\x01\x00"
-     "\x81\xED\x01\x00\x82\xF9\x01\x00\xC6\xB7\x00\x80\x86\x90\x01\x00\x87\xB4"
-     "\x01\x00"
-     "\xCE\xB7\x00\x80\xD2\xB7\x00\x80\xB8\xC5\x01\x00\xB9\xCD\x01\x00\xBA\xC5"
-     "\x01\x00"
-     "\xBB\xDD\x01\x00\xBC\xCD\x01\x00\xBD\xFD\x01\x00\xBE\xF5\x01\x00\xBF\x9D"
-     "\x01\x00"
-     "\xB0\x8D\x06\x00\xB1\x41\x01\x00\xB2\x41\x01\x00\xB3\x41\x01\x00\xB4\x41"
-     "\x01\x00"
-     "\xB5\x41\x01\x00\xB6\x41\x01\x00\xB7\x41\x01\x00\xB3\x31\x06\x00\xD6\xB7"
-     "\x00\x80"
-     "\xDA\xB7\x00\x80\xDE\xB7\x00\x80\xE2\xB7\x00\x80\xB6\x45\x06\x00\xB5\x5D"
-     "\x06\x00"
-     "\xE6\xB7\x00\x80\xBB\x85\x01\x00\xBA\x49\x06\x00\xEA\xB7\x00\x80\xEE\xB7"
-     "\x00\x80"
-     "\xBF\x85\x01\x00\xBE\x85\x01\x00\xBD\x95\x01\x00\xBC\x95\x01\x00\xF2\xB7"
-     "\x00\x80"
-     "\xA3\x75\x06\x00\xF6\xB7\x00\x80\xFA\xB7\x00\x80\xA6\x01\x06\x00\xFE\xB7"
-     "\x00\x80"
-     "\x02\xB8\x00\x80\xA5\x19\x06\x00\xAA\x0D\x06\x00\xAB\xC1\x01\x00\x06\xB8"
-     "\x00\x80"
-     "\x0A\xB8\x00\x80\xAE\xC1\x01\x00\xAF\xC1\x01\x00\xAC\xD1\x01\x00\xAD\xD1"
-     "\x01\x00"
-     "\xA8\xE5\x02\x00\xA9\x3D\x03\x00\xAA\x35\x03\x00\xAB\x89\x03\x00\xAC\x99"
-     "\x03\x00"
-     "\xAD\x99\x03\x00\xAE\x89\x03\x00\xAF\x81\x03\x00\x0E\xB8\x00\x80\x12\xB8"
-     "\x00\x80"
-     "\x16\xB8\x00\x80\x1A\xB8\x00\x80\x1E\xB8\x00\x80\x22\xB8\x00\x80\x26\xB8"
-     "\x00\x80"
-     "\x2A\xB8\x00\x80\xB8\xAD\x03\x00\xB9\x65\x00\x00\xBA\x6D\x00\x00\xBB\x65"
-     "\x00\x00"
-     "\xBC\x7D\x00\x00\xBD\x65\x00\x00\xBE\x6D\x00\x00\xBF\x65\x00\x00\xB0\xC9"
-     "\x03\x00"
-     "\xB1\xC9\x03\x00\xB2\xA9\x03\x00\xB3\xA5\x03\x00\xB4\xBD\x03\x00\xB5\xA1"
-     "\x03\x00"
-     "\xB6\xA1\x03\x00\xB7\x95\x03\x00\x80\xBD\x00\x00\x81\x09\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\x2E\xB8\x00\x80\xEF\xF0\x02\x00\x32\xB8\x00\x80\xBE\x18\x03\x00\x3A\xB8"
-     "\x00\x80"
-     "\xE3\x14\x02\x00\x84\xA8\x02\x00\xE1\x80\x01\x00\x3E\xB8\x00\x80\x42\xB8"
-     "\x00\x80"
-     "\x46\xB8\x00\x80\x87\xC8\x03\x00\x86\x1C\x0C\x00\xB3\xFD\x03\x00\x4A\xB8"
-     "\x00\x80"
-     "\x4E\xB8\x00\x80\x52\xB8\x00\x80\x56\xB8\x00\x80\xB6\x59\x03\x00\xB5\x51"
-     "\x03\x00"
-     "\x5A\xB8\x00\x80\xBB\x4D\x03\x00\xBA\x4D\x03\x00\x5E\xB8\x00\x80\x62\xB8"
-     "\x00\x80"
-     "\xBF\x25\x03\x00\xBE\x39\x03\x00\xBD\x31\x03\x00\xBC\x3D\x03\x00\x66\xB8"
-     "\x00\x80"
-     "\x6A\xB8\x00\x80\x6E\xB8\x00\x80\xA3\x31\x02\x00\x72\xB8\x00\x80\xA5\x9D"
-     "\x02\x00"
-     "\xA6\x95\x02\x00\x76\xB8\x00\x80\x84\x2C\x02\x00\x7A\xB8\x00\x80\xAA\x81"
-     "\x02\x00"
-     "\xAB\x81\x02\x00\xAC\xF1\x02\x00\xAD\xFD\x02\x00\xAE\xF5\x02\x00\xAF\xE9"
-     "\x02\x00"
-     "\x7E\xB8\x00\x80\x82\xB8\x00\x80\xBE\x44\x0D\x00\x84\x4C\x0D\x00\xE3\x20"
-     "\x06\x00"
-     "\x86\xB8\x00\x80\xE1\xE8\x07\x00\x8A\xB8\x00\x80\xEF\x18\x06\x00\x8E\xB8"
-     "\x00\x80"
-     "\x92\xB8\x00\x80\x96\xB8\x00\x80\x80\x0D\x00\x00\x81\xF5\x00\x00\x82\xFD"
-     "\x00\x00"
-     "\x9A\xB8\x00\x80\x9E\xB8\x00\x80\xA2\xB8\x00\x80\x87\x48\x0D\x00\x86\x2C"
-     "\x0C\x00"
-     "\xAA\xB8\x00\x80\xAE\xB8\x00\x80\xB2\xB8\x00\x80\xEF\xDC\x06\x00\xBE\xAC"
-     "\x0C\x00"
-     "\xE1\x70\x06\x00\xB6\xB8\x00\x80\xE3\x5C\x01\x00\xBA\xB8\x00\x80\xBE\xB8"
-     "\x00\x80"
-     "\xC2\xB8\x00\x80\xC6\xB8\x00\x80\xB3\x71\x01\x00\xCA\xB8\x00\x80\xCE\xB8"
-     "\x00\x80"
-     "\xD2\xB8\x00\x80\xD6\xB8\x00\x80\xB6\x1D\x01\x00\xB5\x1D\x01\x00\xDA\xB8"
-     "\x00\x80"
-     "\xBB\x09\x01\x00\xBA\x39\x01\x00\xDE\xB8\x00\x80\xE2\xB8\x00\x80\xBF\xF9"
-     "\x00\x00"
-     "\xBE\xF1\x00\x00\xBD\xF9\x00\x00\xBC\x11\x01\x00\xA8\x9D\x0E\x00\xA9\xA9"
-     "\x0E\x00"
-     "\xAA\xB9\x0E\x00\xAB\xB9\x0E\x00\xAC\xD5\x0E\x00\xAD\xDD\x0E\x00\xAE\xD5"
-     "\x0E\x00"
-     "\xAF\xCD\x0E\x00\xA6\xB8\x00\x80\xE6\xB8\x00\x80\xEA\xB8\x00\x80\xEE\xB8"
-     "\x00\x80"
-     "\xF2\xB8\x00\x80\xF6\xB8\x00\x80\xFA\xB8\x00\x80\xFE\xB8\x00\x80\xB8\x65"
-     "\x0F\x00"
-     "\xB9\x05\x0F\x00\xBA\x0D\x0F\x00\xBB\x19\x0F\x00\xBC\x0D\x0F\x00\xBD\x31"
-     "\x0F\x00"
-     "\xBE\x31\x0F\x00\xBF\x31\x0F\x00\xB0\xBD\x0E\x00\xB1\x8D\x0E\x00\xB2\x85"
-     "\x0E\x00"
-     "\xB3\x6D\x0F\x00\xB4\x75\x0F\x00\xB5\x7D\x0F\x00\xB6\x69\x0F\x00\xB7\x61"
-     "\x0F\x00"
-     "\xA3\x31\x0E\x00\x02\xB9\x00\x80\x06\xB9\x00\x80\x0A\xB9\x00\x80\x0E\xB9"
-     "\x00\x80"
-     "\xA6\x5D\x0E\x00\xA5\x5D\x0E\x00\x12\xB9\x00\x80\xAB\x49\x0E\x00\xAA\x79"
-     "\x0E\x00"
-     "\x16\xB9\x00\x80\x1A\xB9\x00\x80\xAF\xB9\x0F\x00\xAE\xB1\x0F\x00\xAD\xB9"
-     "\x0F\x00"
-     "\xAC\x51\x0E\x00\x80\x4D\x00\x00\x81\x55\x00\x00\x82\x55\x00\x00\xB3\x55"
-     "\x0F\x00"
-     "\x1E\xB9\x00\x80\xB5\x65\x0F\x00\xB6\x6D\x0F\x00\x22\xB9\x00\x80\x36\xB8"
-     "\x00\x80"
-     "\x26\xB9\x00\x80\xBA\x2D\x0F\x00\xBB\x25\x0F\x00\xBC\x3D\x0F\x00\xBD\x25"
-     "\x0F\x00"
-     "\xBE\x25\x0F\x00\xBF\x15\x0F\x00\xA8\x91\x0E\x00\xA9\x91\x0E\x00\xAA\x91"
-     "\x0E\x00"
-     "\xAB\xA5\x0E\x00\xAC\xA1\x0E\x00\xAD\xA1\x0E\x00\xAE\xA1\x0E\x00\xAF\xA1"
-     "\x0E\x00"
-     "\x86\x80\x01\x00\x87\xC4\x02\x00\x2A\xB9\x00\x80\x2E\xB9\x00\x80\x32\xB9"
-     "\x00\x80"
-     "\x36\xB9\x00\x80\x3A\xB9\x00\x80\x3E\xB9\x00\x80\xB8\x69\x01\x00\xB9\x69"
-     "\x01\x00"
-     "\xBA\x79\x01\x00\xBB\x79\x01\x00\xBC\x69\x01\x00\xBD\x69\x01\x00\xBE\xDD"
-     "\x01\x00"
-     "\xBF\xD5\x01\x00\xB0\xA5\x0E\x00\xB1\xAD\x0E\x00\xB2\xA5\x0E\x00\xB3\xBD"
-     "\x0E\x00"
-     "\xB4\xAD\x0E\x00\xB5\x9D\x0E\x00\xB6\x95\x0E\x00\xB7\x59\x01\x00\xA3\x19"
-     "\x0E\x00"
-     "\x42\xB9\x00\x80\x46\xB9\x00\x80\x4A\xB9\x00\x80\x4E\xB9\x00\x80\xA6\x21"
-     "\x0E\x00"
-     "\xA5\x29\x0E\x00\x52\xB9\x00\x80\xAB\x69\x0E\x00\xAA\x61\x0E\x00\x56\xB9"
-     "\x00\x80"
-     "\x5A\xB9\x00\x80\xAF\x59\x0E\x00\xAE\x69\x0E\x00\xAD\x69\x0E\x00\xAC\x71"
-     "\x0E\x00"
-     "\x5E\xB9\x00\x80\x62\xB9\x00\x80\x66\xB9\x00\x80\x6A\xB9\x00\x80\x6E\xB9"
-     "\x00\x80"
-     "\x72\xB9\x00\x80\x76\xB9\x00\x80\x7A\xB9\x00\x80\x80\x3D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x7E\xB9\x00\x80\x82\xB9\x00\x80\x84\xB8\x01\x00\xBE\xB0"
-     "\x01\x00"
-     "\x8A\xB9\x00\x80\xA8\xAD\x02\x00\xA9\xD5\x02\x00\xAA\xD5\x02\x00\xAB\xE9"
-     "\x02\x00"
-     "\xAC\x39\x03\x00\xAD\x39\x03\x00\xAE\x29\x03\x00\xAF\x29\x03\x00\x8E\xB9"
-     "\x00\x80"
-     "\x92\xB9\x00\x80\x96\xB9\x00\x80\x9A\xB9\x00\x80\x9E\xB9\x00\x80\xA2\xB9"
-     "\x00\x80"
-     "\xA6\xB9\x00\x80\xAA\xB9\x00\x80\xB8\xED\x03\x00\xB9\x85\x03\x00\xBA\x8D"
-     "\x03\x00"
-     "\xBB\x85\x03\x00\xBC\x9D\x03\x00\xBD\x85\x03\x00\xBE\x8D\x03\x00\xBF\x85"
-     "\x03\x00"
-     "\xB0\x59\x03\x00\xB1\x59\x03\x00\xB2\xED\x03\x00\xB3\xE5\x03\x00\xB4\xFD"
-     "\x03\x00"
-     "\xB5\xE5\x03\x00\xB6\xE5\x03\x00\xB7\xD5\x03\x00\xB3\x5D\x02\x00\xAE\xB9"
-     "\x00\x80"
-     "\x87\xC8\x03\x00\x86\x2C\x04\x00\xB2\xB9\x00\x80\xB6\xB9\x03\x00\xB5\xB1"
-     "\x03\x00"
-     "\xB6\xB9\x00\x80\xBB\x9D\x03\x00\xBA\x9D\x03\x00\xBA\xB9\x00\x80\xBE\xB9"
-     "\x00\x80"
-     "\xBF\x21\x03\x00\xBE\x39\x03\x00\xBD\x31\x03\x00\xBC\x39\x03\x00\x85\x00"
-     "\x30\x00"
-     "\xA3\x19\x02\x00\xC2\xB9\x00\x80\xC6\xB9\x00\x80\xA6\xFD\x03\x00\xCA\xB9"
-     "\x00\x80"
-     "\xCE\xB9\x00\x80\xA5\xF5\x03\x00\xAA\xD9\x03\x00\xAB\xD9\x03\x00\xD2\xB9"
-     "\x00\x80"
-     "\xD6\xB9\x00\x80\xAE\x7D\x03\x00\xAF\x65\x03\x00\xAC\x7D\x03\x00\xAD\x75"
-     "\x03\x00"
-     "\x84\x0C\x02\x00\xDA\xB9\x00\x80\xDE\xB9\x00\x80\xBE\x00\x05\x00\xE2\xB9"
-     "\x00\x80"
-     "\xEF\x14\x00\x00\xE6\xB9\x00\x80\xEA\xB9\x00\x80\xEE\xB9\x00\x80\xE3\xC0"
-     "\x00\x00"
-     "\xF2\xB9\x00\x80\xE1\x14\x00\x00\xFA\xB9\x00\x80\xFE\xB9\x00\x80\x86\x98"
-     "\x04\x00"
-     "\x87\x10\x05\x00\x80\x5D\x00\x00\x81\x65\x00\x00\x82\x61\x00\x00\x02\xBA"
-     "\x00\x80"
-     "\x06\xBA\x00\x80\x0A\xBA\x00\x80\x0E\xBA\x00\x80\xEF\xEC\x0E\x00\x12\xBA"
-     "\x00\x80"
-     "\xE1\xC0\x0F\x00\x16\xBA\x00\x80\xE3\x5C\x0E\x00\x1A\xBA\x00\x80\x1E\xBA"
-     "\x00\x80"
-     "\x22\xBA\x00\x80\x26\xBA\x00\x80\x2A\xBA\x00\x80\x2E\xBA\x00\x80\x32\xBA"
-     "\x00\x80"
-     "\x36\xBA\x00\x80\x3A\xBA\x00\x80\x3E\xBA\x00\x80\x42\xBA\x00\x80\xEF\xC4"
-     "\x0E\x00"
-     "\x84\xB4\x05\x00\xE1\x30\x0E\x00\x46\xBA\x00\x80\xE3\xF0\x0E\x00\x4A\xBA"
-     "\x00\x80"
-     "\x4E\xBA\x00\x80\x52\xBA\x00\x80\x56\xBA\x00\x80\xB3\xE5\x01\x00\x5A\xBA"
-     "\x00\x80"
-     "\x5E\xBA\x00\x80\x62\xBA\x00\x80\x6A\xBA\x00\x80\xB6\xE5\x01\x00\xB5\xF5"
-     "\x01\x00"
-     "\xBE\xD0\x04\x00\xBB\x65\x01\x00\xBA\xD9\x01\x00\x6E\xBA\x00\x80\x72\xBA"
-     "\x00\x80"
-     "\xBF\x5D\x01\x00\xBE\x5D\x01\x00\xBD\x65\x01\x00\xBC\x6D\x01\x00\xA8\xFD"
-     "\x05\x00"
-     "\xA9\x25\x06\x00\xAA\x2D\x06\x00\xAB\x25\x06\x00\xAC\x3D\x06\x00\xAD\x25"
-     "\x06\x00"
-     "\xAE\x2D\x06\x00\xAF\x25\x06\x00\xF6\xB9\x00\x80\x82\xE9\x07\x00\x81\xE1"
-     "\x07\x00"
-     "\x80\xE9\x07\x00\x66\xBA\x00\x80\x76\xBA\x00\x80\x86\x98\x00\x00\x87\xB0"
-     "\x03\x00"
-     "\xB8\x95\x06\x00\xB9\x9D\x06\x00\xBA\xA1\x06\x00\xBB\xA1\x06\x00\xBC\x55"
-     "\x07\x00"
-     "\xBD\x5D\x07\x00\xBE\x49\x07\x00\xBF\x45\x07\x00\xB0\x5D\x06\x00\xB1\xE1"
-     "\x06\x00"
-     "\xB2\xF9\x06\x00\xB3\xF1\x06\x00\xB4\xD9\x06\x00\xB5\xD9\x06\x00\xB6\xB1"
-     "\x06\x00"
-     "\xB7\xB1\x06\x00\xA3\xA5\x06\x00\x7A\xBA\x00\x80\x7E\xBA\x00\x80\x82\xBA"
-     "\x00\x80"
-     "\x86\xBA\x00\x80\xA6\xA5\x06\x00\xA5\xB5\x06\x00\x8A\xBA\x00\x80\xAB\x25"
-     "\x06\x00"
-     "\xAA\x99\x06\x00\x8E\xBA\x00\x80\x92\xBA\x00\x80\xAF\x1D\x06\x00\xAE\x1D"
-     "\x06\x00"
-     "\xAD\x25\x06\x00\xAC\x2D\x06\x00\xB5\x95\x07\x00\x96\xBA\x00\x80\x9A\xBA"
-     "\x00\x80"
-     "\xB6\x95\x07\x00\x9E\xBA\x00\x80\xA2\xBA\x00\x80\xB3\x95\x07\x00\xA6\xBA"
-     "\x00\x80"
-     "\xBD\x9D\x07\x00\xBC\x9D\x07\x00\xBF\x6D\x07\x00\xBE\x6D\x07\x00\xAA\xBA"
-     "\x00\x80"
-     "\x86\xB9\x00\x80\xBB\x8D\x07\x00\xBA\x8D\x07\x00\xA8\x31\x06\x00\xA9\x31"
-     "\x06\x00"
-     "\xAA\x45\x06\x00\xAB\x5D\x06\x00\xAC\x45\x06\x00\xAD\x4D\x06\x00\xAE\x79"
-     "\x06\x00"
-     "\xAF\x75\x06\x00\xAE\xBA\x00\x80\xB2\xBA\x00\x80\xB6\xBA\x00\x80\xBA\xBA"
-     "\x00\x80"
-     "\xBE\xBA\x00\x80\xC2\xBA\x00\x80\xC6\xBA\x00\x80\xCA\xBA\x00\x80\xB8\xD1"
-     "\x06\x00"
-     "\xB9\xD1\x06\x00\xBA\xD1\x06\x00\xBB\xE5\x06\x00\xBC\xE1\x06\x00\xBD\xE1"
-     "\x06\x00"
-     "\xBE\xE1\x06\x00\xBF\xE1\x06\x00\xB0\x11\x06\x00\xB1\x11\x06\x00\xB2\x11"
-     "\x06\x00"
-     "\xB3\x11\x06\x00\xB4\xF1\x06\x00\xB5\xF1\x06\x00\xB6\xF1\x06\x00\xB7\xF1"
-     "\x06\x00"
-     "\xA3\xCD\x06\x00\x82\x19\x00\x00\x81\x19\x00\x00\x80\xB1\x00\x00\xCE\xBA"
-     "\x00\x80"
-     "\xA6\xCD\x06\x00\xA5\xCD\x06\x00\xD2\xBA\x00\x80\xAB\xD5\x06\x00\xAA\xD5"
-     "\x06\x00"
-     "\x84\xA0\x02\x00\xBE\xA0\x01\x00\xAF\x35\x06\x00\xAE\x35\x06\x00\xAD\xC5"
-     "\x06\x00"
-     "\xAC\xC5\x06\x00\xDA\xBA\x00\x80\xB3\x29\x06\x00\x86\x68\x3F\x00\x87\x2C"
-     "\x01\x00"
-     "\xB6\x59\x01\x00\xDE\xBA\x00\x80\xE2\xBA\x00\x80\xB5\x59\x01\x00\xBA\x7D"
-     "\x01\x00"
-     "\xBB\x4D\x01\x00\xE6\xBA\x00\x80\xEA\xBA\x00\x80\xBE\x39\x01\x00\xBF\x25"
-     "\x01\x00"
-     "\xBC\x55\x01\x00\xBD\x39\x01\x00\xA8\xD1\x05\x00\xA9\xED\x05\x00\xAA\xE5"
-     "\x05\x00"
-     "\xAB\x39\x02\x00\xAC\x95\x02\x00\xAD\x9D\x02\x00\xAE\x95\x02\x00\xAF\x8D"
-     "\x02\x00"
-     "\xBE\x6C\x03\x00\xEE\xBA\x00\x80\xF2\xBA\x00\x80\x84\xE0\x03\x00\xF6\xBA"
-     "\x00\x80"
-     "\xFA\xBA\x00\x80\xFE\xBA\x00\x80\x02\xBB\x00\x80\xB8\xB1\x02\x00\xB9\x65"
-     "\x03\x00"
-     "\xBA\x6D\x03\x00\xBB\x65\x03\x00\xBC\x7D\x03\x00\xBD\x65\x03\x00\xBE\x6D"
-     "\x03\x00"
-     "\xBF\x65\x03\x00\xB0\xFD\x02\x00\xB1\xCD\x02\x00\xB2\xC5\x02\x00\xB3\xA9"
-     "\x02\x00"
-     "\xB4\xB9\x02\x00\xB5\xA5\x02\x00\xB6\xAD\x02\x00\xB7\xA1\x02\x00\xA3\x6D"
-     "\x05\x00"
-     "\x06\xBB\x00\x80\x0A\xBB\x00\x80\x0E\xBB\x00\x80\x12\xBB\x00\x80\xA6\x1D"
-     "\x02\x00"
-     "\xA5\x1D\x02\x00\x16\xBB\x00\x80\xAB\x09\x02\x00\xAA\x39\x02\x00\x1A\xBB"
-     "\x00\x80"
-     "\x1E\xBB\x00\x80\xAF\x61\x02\x00\xAE\x7D\x02\x00\xAD\x7D\x02\x00\xAC\x11"
-     "\x02\x00"
-     "\x22\xBB\x00\x80\x26\xBB\x00\x80\x2A\xBB\x00\x80\x2E\xBB\x00\x80\x32\xBB"
-     "\x00\x80"
-     "\x36\xBB\x00\x80\x3A\xBB\x00\x80\x3E\xBB\x00\x80\x80\x39\x00\x00\x81\x39"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x42\xBB\x00\x80\x86\x90\x3C\x00\x87\xD8\x3C\x00\x4A\xBB"
-     "\x00\x80"
-     "\x4E\xBB\x00\x80\xB3\xD9\x02\x00\x52\xBB\x00\x80\x84\xC4\x3C\x00\xBE\xC4"
-     "\x3C\x00"
-     "\x56\xBB\x00\x80\xB6\x31\x01\x00\xB5\xF5\x02\x00\x5A\xBB\x00\x80\xBB\x19"
-     "\x01\x00"
-     "\xBA\x11\x01\x00\x5E\xBB\x00\x80\x62\xBB\x00\x80\xBF\xBD\x01\x00\xBE\xBD"
-     "\x01\x00"
-     "\xBD\xBD\x01\x00\xBC\x01\x01\x00\x66\xBB\x00\x80\x6A\xBB\x00\x80\xEF\xBC"
-     "\x06\x00"
-     "\x6E\xBB\x00\x80\x72\xBB\x00\x80\x84\x98\x3F\x00\x76\xBB\x00\x80\x7A\xBB"
-     "\x00\x80"
-     "\x7E\xBB\x00\x80\x82\xBB\x00\x80\x86\xBB\x00\x80\x8A\xBB\x00\x80\xE1\xE4"
-     "\x07\x00"
-     "\x8E\xBB\x00\x80\xE3\x68\x06\x00\x92\xBB\x00\x80\x96\xBB\x00\x80\x9A\xBB"
-     "\x00\x80"
-     "\x9E\xBB\x00\x80\xEF\xD8\x00\x00\xA2\xBB\x00\x80\xA6\xBB\x00\x80\xAA\xBB"
-     "\x00\x80"
-     "\xEF\xB8\x06\x00\xBE\x68\x3F\x00\xE1\x84\x06\x00\xAE\xBB\x00\x80\xE3\x1C"
-     "\x06\x00"
-     "\xB2\xBB\x00\x80\xE1\x94\x01\x00\x84\x74\x3E\x00\xE3\x04\x00\x00\xA3\x19"
-     "\x02\x00"
-     "\xBA\xBB\x00\x80\x86\xE8\x3C\x00\x87\x6C\x3C\x00\x85\x74\x35\x00\xA6\xF1"
-     "\x01\x00"
-     "\xA5\x35\x02\x00\xBE\xBB\x00\x80\xAB\xD9\x01\x00\xAA\xD1\x01\x00\xC2\xBB"
-     "\x00\x80"
-     "\xC6\xBB\x00\x80\xAF\x7D\x01\x00\xAE\x7D\x01\x00\xAD\x7D\x01\x00\xAC\xC1"
-     "\x01\x00"
-     "\xA8\xC1\x3D\x00\xA9\xD1\x3D\x00\xAA\xD1\x3D\x00\xAB\xE5\x3D\x00\xAC\xFD"
-     "\x3D\x00"
-     "\xAD\x31\x3E\x00\xAE\x31\x3E\x00\xAF\x2D\x3E\x00\x46\xBB\x00\x80\x82\x95"
-     "\x3F\x00"
-     "\x81\x81\x3F\x00\x80\xB5\x3F\x00\xB6\xBB\x00\x80\xCA\xBB\x00\x80\xCE\xBB"
-     "\x00\x80"
-     "\xD2\xBB\x00\x80\xB8\xE9\x3E\x00\xB9\xE9\x3E\x00\xBA\x89\x3E\x00\xBB\x89"
-     "\x3E\x00"
-     "\xBC\x99\x3E\x00\xBD\x99\x3E\x00\xBE\x89\x3E\x00\xBF\x89\x3E\x00\xB0\x55"
-     "\x3E\x00"
-     "\xB1\x5D\x3E\x00\xB2\x55\x3E\x00\xB3\xE9\x3E\x00\xB4\xF9\x3E\x00\xB5\xF9"
-     "\x3E\x00"
-     "\xB6\xE9\x3E\x00\xB7\xE1\x3E\x00\xB8\xED\x3F\x00\xB9\xF1\x3F\x00\xBA\xF5"
-     "\x3F\x00"
-     "\xBB\x0D\x3F\x00\xBC\x15\x3F\x00\xBD\x1D\x3F\x00\xBE\x09\x3F\x00\xBF\x09"
-     "\x3F\x00"
-     "\xB0\xE5\x3F\x00\xB1\xE9\x3F\x00\xB2\xF9\x3F\x00\xB3\xF5\x3F\x00\xB4\xED"
-     "\x3F\x00"
-     "\xB5\xD5\x3F\x00\xB6\xDD\x3F\x00\xB7\xD5\x3F\x00\xA8\x39\x3E\x00\xA9\x39"
-     "\x3E\x00"
-     "\xAA\xCD\x3F\x00\xAB\xDD\x3F\x00\xAC\xCD\x3F\x00\xAD\xFD\x3F\x00\xAE\xF5"
-     "\x3F\x00"
-     "\xAF\x9D\x3F\x00\xD6\xBB\x00\x80\xDA\xBB\x00\x80\xDE\xBB\x00\x80\xE2\xBB"
-     "\x00\x80"
-     "\xE6\xBB\x00\x80\xEA\xBB\x00\x80\xD6\xBA\x00\x80\xEE\xBB\x00\x80\xB3\x05"
-     "\x3E\x00"
-     "\xF2\xBB\x00\x80\xF6\xBB\x00\x80\xFA\xBB\x00\x80\xFE\xBB\x00\x80\xB6\x3D"
-     "\x3E\x00"
-     "\xB5\x05\x3E\x00\x02\xBC\x00\x80\xBB\x6D\x3E\x00\xBA\x61\x3E\x00\x06\xBC"
-     "\x00\x80"
-     "\x0A\xBC\x00\x80\xBF\x51\x3E\x00\xBE\x69\x3E\x00\xBD\x61\x3E\x00\xBC\x6D"
-     "\x3E\x00"
-     "\x0E\xBC\x00\x80\xA3\x41\x3E\x00\x12\xBC\x00\x80\x16\xBC\x00\x80\xA6\x79"
-     "\x3E\x00"
-     "\x1A\xBC\x00\x80\x1E\xBC\x00\x80\xA5\x41\x3E\x00\xAA\x25\x3E\x00\xAB\x29"
-     "\x3E\x00"
-     "\x22\xBC\x00\x80\x26\xBC\x00\x80\xAE\x2D\x3E\x00\xAF\x15\x3E\x00\xAC\x29"
-     "\x3E\x00"
-     "\xAD\x25\x3E\x00\xA8\x45\x3E\x00\xA9\x59\x3E\x00\xAA\x61\x3E\x00\xAB\x61"
-     "\x3E\x00"
-     "\xAC\xA5\x01\x00\xAD\xAD\x01\x00\xAE\xA5\x01\x00\xAF\x7D\x01\x00\x80\xF5"
-     "\x00\x00"
-     "\x81\xFD\x00\x00\x82\xC1\x00\x00\x2A\xBC\x00\x80\x86\x00\x04\x00\x87\x14"
-     "\x01\x00"
-     "\x2E\xBC\x00\x80\x32\xBC\x00\x80\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5"
-     "\x00\x00"
-     "\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x91\x00\x00\xBF\x91"
-     "\x00\x00"
-     "\xB0\x05\x01\x00\xB1\x0D\x01\x00\xB2\x05\x01\x00\xB3\x1D\x01\x00\xB4\x05"
-     "\x01\x00"
-     "\xB5\x09\x01\x00\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xB3\xE9\x3D\x00\x36\xBC"
-     "\x00\x80"
-     "\x3A\xBC\x00\x80\x3E\xBC\x00\x80\x42\xBC\x00\x80\xB6\x2D\x02\x00\xB5\x2D"
-     "\x02\x00"
-     "\x46\xBC\x00\x80\xBB\xC9\x02\x00\xBA\xC1\x02\x00\x4A\xBC\x00\x80\xBE\x68"
-     "\x03\x00"
-     "\xBF\xCD\x02\x00\xBE\xC1\x02\x00\xBD\xC9\x02\x00\xBC\xD1\x02\x00\x4E\xBC"
-     "\x00\x80"
-     "\xA3\xAD\x3D\x00\x52\xBC\x00\x80\x56\xBC\x00\x80\xA6\x69\x02\x00\x5A\xBC"
-     "\x00\x80"
-     "\x5E\xBC\x00\x80\xA5\x69\x02\x00\xAA\x85\x02\x00\xAB\x8D\x02\x00\x62\xBC"
-     "\x00\x80"
-     "\x66\xBC\x00\x80\xAE\x85\x02\x00\xAF\x89\x02\x00\xAC\x95\x02\x00\xAD\x8D"
-     "\x02\x00"
-     "\x6A\xBC\x00\x80\x6E\xBC\x00\x80\x72\xBC\x00\x80\x76\xBC\x00\x80\x7A\xBC"
-     "\x00\x80"
-     "\x7E\xBC\x00\x80\x82\xBC\x00\x80\x86\xBC\x00\x80\x80\x0D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x1D\x00\x00\x8A\xBC\x00\x80\x8E\xBC\x00\x80\x92\xBC\x00\x80\x87\xE4"
-     "\x03\x00"
-     "\x86\x3C\x04\x00\x9A\xBC\x00\x80\x9E\xBC\x00\x80\xA2\xBC\x00\x80\xA6\xBC"
-     "\x00\x80"
-     "\xEF\x04\x02\x00\xAA\xBC\x00\x80\xAE\xBC\x00\x80\xB2\xBC\x00\x80\xE3\x48"
-     "\x3F\x00"
-     "\xB6\xBC\x00\x80\xE1\x40\x01\x00\xBA\xBC\x00\x80\xBE\xBC\x00\x80\xC2\xBC"
-     "\x00\x80"
-     "\xC6\xBC\x00\x80\xCA\xBC\x00\x80\xB3\xBD\x02\x00\xCE\xBC\x00\x80\xD2\xBC"
-     "\x00\x80"
-     "\xD6\xBC\x00\x80\xDA\xBC\x00\x80\xB6\x19\x01\x00\xB5\x11\x01\x00\xDE\xBC"
-     "\x00\x80"
-     "\xBB\x61\x01\x00\xBA\x0D\x01\x00\xE2\xBC\x00\x80\xE6\xBC\x00\x80\xBF\x69"
-     "\x01\x00"
-     "\xBE\x61\x01\x00\xBD\x75\x01\x00\xBC\x75\x01\x00\xEA\xBC\x00\x80\xEF\x9C"
-     "\x3F\x00"
-     "\xEE\xBC\x00\x80\xF2\xBC\x00\x80\xF6\xBC\x00\x80\x82\x3D\x00\x00\x81\x3D"
-     "\x00\x00"
-     "\x80\x3D\x00\x00\xFA\xBC\x00\x80\xFE\xBC\x00\x80\x02\xBD\x00\x80\x84\x20"
-     "\x05\x00"
-     "\xBE\x38\x05\x00\xE3\x48\x3F\x00\x0A\xBD\x00\x80\xE1\xAC\x3C\x00\x86\x20"
-     "\x04\x00"
-     "\x87\x4C\x05\x00\x0E\xBD\x00\x80\xEF\xD4\x3E\x00\x12\xBD\x00\x80\x16\xBD"
-     "\x00\x80"
-     "\x1A\xBD\x00\x80\xEF\x68\x3E\x00\x84\xEC\x04\x00\xE1\x4C\x3F\x00\xBE\xE0"
-     "\x04\x00"
-     "\xE3\xB4\x3E\x00\x1E\xBD\x00\x80\xE1\x64\x3E\x00\x22\xBD\x00\x80\xE3\xF8"
-     "\x3E\x00"
-     "\x26\xBD\x00\x80\x2A\xBD\x00\x80\xA3\x75\x02\x00\x84\xE0\x07\x00\xA5\xD9"
-     "\x01\x00"
-     "\x2E\xBD\x00\x80\x32\xBD\x00\x80\xA6\xD1\x01\x00\x36\xBD\x00\x80\x3A\xBD"
-     "\x00\x80"
-     "\xAB\xA9\x01\x00\xAA\xC5\x01\x00\xAD\xBD\x01\x00\xAC\xBD\x01\x00\xAF\xA1"
-     "\x01\x00"
-     "\xAE\xA9\x01\x00\x96\xBC\x00\x80\x06\xBD\x00\x80\x3E\xBD\x00\x80\x42\xBD"
-     "\x00\x80"
-     "\x46\xBD\x00\x80\x4A\xBD\x00\x80\x4E\xBD\x00\x80\x52\xBD\x00\x80\xA8\xE1"
-     "\x06\x00"
-     "\xA9\xE1\x06\x00\xAA\xE1\x06\x00\xAB\xE1\x06\x00\xAC\xE5\x06\x00\xAD\xED"
-     "\x06\x00"
-     "\xAE\xE5\x06\x00\xAF\xD9\x06\x00\xB0\xAD\x06\x00\xB1\xB5\x06\x00\xB2\xBD"
-     "\x06\x00"
-     "\xB3\xB5\x06\x00\xB4\xAD\x06\x00\xB5\x51\x07\x00\xB6\x51\x07\x00\xB7\x51"
-     "\x07\x00"
-     "\xB8\x71\x07\x00\xB9\x79\x07\x00\xBA\x41\x07\x00\xBB\x41\x07\x00\xBC\x31"
-     "\x07\x00"
-     "\xBD\x3D\x07\x00\xBE\x35\x07\x00\xBF\x2D\x07\x00\xB3\x49\x06\x00\x56\xBD"
-     "\x00\x80"
-     "\x5A\xBD\x00\x80\x5E\xBD\x00\x80\x62\xBD\x00\x80\xB6\x4D\x06\x00\xB5\x4D"
-     "\x06\x00"
-     "\x66\xBD\x00\x80\xBB\x51\x06\x00\xBA\x49\x06\x00\x6A\xBD\x00\x80\x6E\xBD"
-     "\x00\x80"
-     "\xBF\xB5\x07\x00\xBE\x49\x06\x00\xBD\x41\x06\x00\xBC\x49\x06\x00\x82\x15"
-     "\x00\x00"
-     "\xA3\x0D\x06\x00\x80\x61\x00\x00\x81\x61\x00\x00\xA6\x09\x06\x00\x72\xBD"
-     "\x00\x80"
-     "\x7A\xBD\x00\x80\xA5\x09\x06\x00\xAA\x0D\x06\x00\xAB\x15\x06\x00\x86\x48"
-     "\x00\x00"
-     "\x87\xDC\x00\x00\xAE\x0D\x06\x00\xAF\xF1\x07\x00\xAC\x0D\x06\x00\xAD\x05"
-     "\x06\x00"
-     "\xA8\x79\x06\x00\xA9\x8D\x06\x00\xAA\x99\x06\x00\xAB\x99\x06\x00\xAC\x89"
-     "\x06\x00"
-     "\xAD\xA5\x06\x00\xAE\xA1\x06\x00\xAF\xDD\x06\x00\x76\xBD\x00\x80\x7E\xBD"
-     "\x00\x80"
-     "\x82\xBD\x00\x80\x86\xBD\x00\x80\x8A\xBD\x00\x80\x8E\xBD\x00\x80\x92\xBD"
-     "\x00\x80"
-     "\x96\xBD\x00\x80\xB8\x55\x01\x00\xB9\x5D\x01\x00\xBA\x55\x01\x00\xBB\xC9"
-     "\x01\x00"
-     "\xBC\xD9\x01\x00\xBD\xD9\x01\x00\xBE\xC9\x01\x00\xBF\xC1\x01\x00\xB0\xA5"
-     "\x06\x00"
-     "\xB1\xB1\x06\x00\xB2\xB1\x06\x00\xB3\x8D\x06\x00\xB4\x95\x06\x00\xB5\x7D"
-     "\x01\x00"
-     "\xB6\x75\x01\x00\xB7\x6D\x01\x00\xB3\x11\x06\x00\x9A\xBD\x00\x80\x9E\xBD"
-     "\x00\x80"
-     "\xA2\xBD\x00\x80\xA6\xBD\x00\x80\xB6\x25\x06\x00\xB5\x01\x06\x00\xAA\xBD"
-     "\x00\x80"
-     "\xBB\x75\x06\x00\xBA\x61\x06\x00\xAE\xBD\x00\x80\xB2\xBD\x00\x80\xBF\xB9"
-     "\x01\x00"
-     "\xBE\x51\x06\x00\xBD\x59\x06\x00\xBC\x65\x06\x00\xB6\xBD\x00\x80\xA3\x55"
-     "\x06\x00"
-     "\xBA\xBD\x00\x80\xBE\xBD\x00\x80\xA6\x61\x06\x00\xC2\xBD\x00\x80\xC6\xBD"
-     "\x00\x80"
-     "\xA5\x45\x06\x00\xAA\x25\x06\x00\xAB\x31\x06\x00\xCA\xBD\x00\x80\xCE\xBD"
-     "\x00\x80"
-     "\xAE\x15\x06\x00\xAF\xFD\x01\x00\xAC\x21\x06\x00\xAD\x1D\x06\x00\xA8\xD9"
-     "\x02\x00"
-     "\xA9\x21\x03\x00\xAA\x21\x03\x00\xAB\x21\x03\x00\xAC\x21\x03\x00\xAD\x21"
-     "\x03\x00"
-     "\xAE\x21\x03\x00\xAF\x21\x03\x00\xD2\xBD\x00\x80\xD6\xBD\x00\x80\xDA\xBD"
-     "\x00\x80"
-     "\xDE\xBD\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\xE2\xBD"
-     "\x00\x80"
-     "\xB8\x89\x03\x00\xB9\x89\x03\x00\xBA\x9D\x03\x00\xBB\x95\x03\x00\xBC\xB9"
-     "\x03\x00"
-     "\xBD\xB9\x03\x00\xBE\x79\x00\x00\xBF\x79\x00\x00\xB0\xE5\x03\x00\xB1\xED"
-     "\x03\x00"
-     "\xB2\xE5\x03\x00\xB3\xFD\x03\x00\xB4\xED\x03\x00\xB5\xDD\x03\x00\xB6\xD5"
-     "\x03\x00"
-     "\xB7\xBD\x03\x00\xBE\xCC\x03\x00\xE1\x40\x3B\x00\xEA\xBD\x00\x80\xE3\xD8"
-     "\x16\x00"
-     "\xEE\xBD\x00\x80\xF2\xBD\x00\x80\xF6\xBD\x00\x80\xEF\xE4\x02\x00\xFA\xBD"
-     "\x00\x80"
-     "\xE1\x54\x00\x00\xFE\xBD\x00\x80\xE3\x70\x07\x00\x02\xBE\x00\x80\x06\xBE"
-     "\x00\x80"
-     "\x0A\xBE\x00\x80\xEF\x60\x0A\x00\xBB\x49\x03\x00\xBA\x41\x03\x00\x86\xE8"
-     "\x0D\x00"
-     "\x87\x6C\x03\x00\xBF\x79\x03\x00\xBE\x49\x03\x00\xBD\x49\x03\x00\xBC\x51"
-     "\x03\x00"
-     "\xB3\xC9\x03\x00\x0E\xBE\x00\x80\x12\xBE\x00\x80\x16\xBE\x00\x80\x84\xEC"
-     "\x0C\x00"
-     "\xB6\xC1\x03\x00\xB5\xC9\x03\x00\x1A\xBE\x00\x80\xAC\x09\x05\x00\xAD\x09"
-     "\x05\x00"
-     "\xAE\x19\x05\x00\xAF\x11\x05\x00\xA8\xBD\x06\x00\xA9\xAD\x06\x00\xAA\xA5"
-     "\x06\x00"
-     "\xAB\x19\x05\x00\x1E\xBE\x00\x80\x22\xBE\x00\x80\x26\xBE\x00\x80\x31\xBE"
-     "\x00\x80"
-     "\x35\xBE\x00\x80\x39\xBE\x00\x80\x3D\xBE\x00\x80\x41\xBE\x00\x80\xBC\xD1"
-     "\x05\x00"
-     "\xBD\xD1\x05\x00\xBE\xD1\x05\x00\xBF\xD1\x05\x00\xB8\xD1\x05\x00\xB9\xD1"
-     "\x05\x00"
-     "\xBA\xD1\x05\x00\xBB\xD1\x05\x00\xB4\x35\x05\x00\xB5\x3D\x05\x00\xB6\x01"
-     "\x05\x00"
-     "\xB7\x01\x05\x00\xB0\x59\x05\x00\xB1\x35\x05\x00\xB2\x31\x05\x00\xB3\x31"
-     "\x05\x00"
-     "\xE1\x18\x06\x00\xE1\x20\x07\x00\xE3\x38\x06\x00\xE3\xF0\x06\x00\x45\xBE"
-     "\x00\x80"
-     "\x49\xBE\x00\x80\x4D\xBE\x00\x80\x51\xBE\x00\x80\x80\x3D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x55\xBE\x00\x80\x59\xBE\x00\x80\x61\xBE\x00\x80\xEF\xE8"
-     "\x06\x00"
-     "\xEF\x34\x06\x00\xA3\xE9\x02\x00\x65\xBE\x00\x80\x86\xE8\x0F\x00\x87\xB8"
-     "\x0C\x00"
-     "\x69\xBE\x00\x80\xA6\xE1\x02\x00\xA5\xE9\x02\x00\x6D\xBE\x00\x80\xAB\x69"
-     "\x02\x00"
-     "\xAA\x61\x02\x00\xBE\x64\x0F\x00\x71\xBE\x00\x80\xAF\x59\x02\x00\xAE\x69"
-     "\x02\x00"
-     "\xAD\x69\x02\x00\xAC\x71\x02\x00\xA9\x01\x1C\x00\xA8\x59\x18\x00\xAB\x1D"
-     "\x1C\x00"
-     "\xAA\x01\x1C\x00\xAD\x01\x10\x00\xAC\x41\x1D\x00\xAF\x9D\x11\x00\xAE\x1D"
-     "\x10\x00"
-     "\xA1\x05\x00\x00\xE6\xBD\x00\x80\xA3\x01\x04\x00\xA2\x7D\x00\x00\xA5\x05"
-     "\x04\x00"
-     "\xA4\x11\x04\x00\xA7\x39\x18\x00\xA6\x01\x18\x00\xA0\x2D\x09\x00\xA1\x45"
-     "\x0C\x00"
-     "\xA2\x4D\x0C\x00\xA3\x6D\x0D\x00\x5D\xBE\x00\x80\x75\xBE\x00\x80\x79\xBE"
-     "\x00\x80"
-     "\x7D\xBE\x00\x80\xB1\x01\x14\x00\xB0\x01\x14\x00\xB3\x01\x68\x00\xB2\x89"
-     "\x14\x00"
-     "\xB5\xF5\x68\x00\xB4\x11\x68\x00\x81\xBE\x00\x80\x85\xBE\x00\x80\x90\x4D"
-     "\x38\x00"
-     "\x91\x85\x38\x00\x92\x99\x38\x00\x93\x49\x39\x00\x94\x4D\x3C\x00\x95\xB5"
-     "\x3C\x00"
-     "\x96\xB5\x3D\x00\x97\x7D\x30\x00\x98\x81\x30\x00\x99\x0D\x30\x00\x9A\x5D"
-     "\x34\x00"
-     "\x9B\x8D\x34\x00\x9C\x81\x34\x00\x9D\x4D\x08\x00\x9E\x99\x08\x00\x9F\x8D"
-     "\x08\x00"
-     "\x89\xBE\x00\x80\xB6\x6D\x0E\x00\xB5\x69\x0E\x00\x8D\xBE\x00\x80\xB3\x45"
-     "\x0E\x00"
-     "\x91\xBE\x00\x80\x95\xBE\x00\x80\x99\xBE\x00\x80\xBF\xC1\x0F\x00\xBE\x19"
-     "\x0E\x00"
-     "\xBD\x05\x0E\x00\xBC\x05\x0E\x00\xBB\x15\x0E\x00\xBA\x4D\x0E\x00\x9D\xBE"
-     "\x00\x80"
-     "\x8F\xF5\x24\x00\x82\x09\x25\x00\x83\xC9\x39\x00\xA1\xBE\x00\x80\xA5\xBE"
-     "\x00\x80"
-     "\x86\xD5\x38\x00\x87\xC9\x3D\x00\x84\x19\x39\x00\x85\x0D\x39\x00\x8A\x01"
-     "\x3D\x00"
-     "\x8B\x09\x3C\x00\xA9\xBE\x00\x80\xAD\xBE\x00\x80\x8E\x99\x31\x00\x8F\xC1"
-     "\x35\x00"
-     "\x8C\xD9\x31\x00\x8D\x1D\x31\x00\x92\x61\x35\x00\x93\x75\x35\x00\xB1\xBE"
-     "\x00\x80"
-     "\xB5\xBE\x00\x80\x96\x61\x09\x00\x97\x39\x08\x00\x94\xA1\x09\x00\x95\x7D"
-     "\x09\x00"
-     "\x9A\x61\x0D\x00\x9B\x61\x0D\x00\xB9\xBE\x00\x80\xBD\xBE\x00\x80\x81\xD5"
-     "\x00\x00"
-     "\x80\xD5\x00\x00\x9C\x59\x0C\x00\x82\xD5\x00\x00\xA8\x09\x0F\x00\xA9\x15"
-     "\x0F\x00"
-     "\xAA\x1D\x0F\x00\xAB\x15\x0F\x00\xAC\x31\x0F\x00\xAD\x3D\x0F\x00\xAE\x29"
-     "\x0F\x00"
-     "\xAF\x29\x0F\x00\xC1\xBE\x00\x80\xC5\xBE\x00\x80\xC9\xBE\x00\x80\x84\x24"
-     "\x01\x00"
-     "\xBE\x3C\x00\x00\xD1\xBE\x00\x80\x86\x18\x07\x00\x87\xFC\x00\x00\xB8\xED"
-     "\x0E\x00"
-     "\xB9\xF5\x0E\x00\xBA\xFD\x0E\x00\xBB\xF5\x0E\x00\xBC\xED\x0E\x00\xBD\x45"
-     "\x0E\x00"
-     "\xBE\x4D\x0E\x00\xBF\x45\x0E\x00\xB0\x59\x0F\x00\xB1\x01\x0F\x00\xB2\x01"
-     "\x0F\x00"
-     "\xB3\x01\x0F\x00\xB4\x0D\x0F\x00\xB5\xD5\x0E\x00\xB6\xDD\x0E\x00\xB7\xD5"
-     "\x0E\x00"
-     "\xA3\x95\x0C\x00\xD5\xBE\x00\x80\xD9\xBE\x00\x80\xDD\xBE\x00\x80\xE1\xBE"
-     "\x00\x80"
-     "\xA6\xBD\x0C\x00\xA5\xB9\x0C\x00\x84\x9C\x03\x00\xAB\xC5\x0C\x00\xAA\x9D"
-     "\x0C\x00"
-     "\xE5\xBE\x00\x80\xE9\xBE\x00\x80\xAF\x11\x0D\x00\xAE\xC9\x0C\x00\xAD\xD5"
-     "\x0C\x00"
-     "\xAC\xD5\x0C\x00\xED\xBE\x00\x80\xB3\x95\x0D\x00\xF1\xBE\x00\x80\xF5\xBE"
-     "\x00\x80"
-     "\xB6\x05\x02\x00\xF9\xBE\x00\x80\xFD\xBE\x00\x80\xB5\xB1\x0D\x00\xBA\x19"
-     "\x02\x00"
-     "\xBB\x19\x02\x00\x01\xBF\x00\x80\x05\xBF\x00\x80\xBE\xA5\x02\x00\xBF\xAD"
-     "\x02\x00"
-     "\xBC\xB9\x02\x00\xBD\xB1\x02\x00\xA8\x35\x0D\x00\xA9\x41\x0D\x00\xAA\x59"
-     "\x0D\x00"
-     "\xAB\x51\x0D\x00\xAC\x7D\x0D\x00\xAD\xC1\x02\x00\xAE\xC1\x02\x00\xAF\xC1"
-     "\x02\x00"
-     "\x09\xBF\x00\x80\x84\xAC\x04\x00\x0D\xBF\x00\x80\x11\xBF\x00\x80\x15\xBF"
-     "\x00\x80"
-     "\x19\xBF\x00\x80\x1D\xBF\x00\x80\x21\xBF\x00\x80\xB8\xC9\x03\x00\xB9\xC9"
-     "\x03\x00"
-     "\xBA\xE1\x03\x00\xBB\xE1\x03\x00\xBC\x99\x03\x00\xBD\x99\x03\x00\xBE\xB1"
-     "\x03\x00"
-     "\xBF\xB1\x03\x00\xB0\x99\x02\x00\xB1\x99\x02\x00\xB2\x29\x02\x00\xB3\x21"
-     "\x02\x00"
-     "\xB4\x25\x02\x00\xB5\x2D\x02\x00\xB6\x25\x02\x00\xB7\x05\x02\x00\xA3\xD1"
-     "\x0C\x00"
-     "\x25\xBF\x00\x80\xBE\x04\x05\x00\x29\xBF\x00\x80\x2D\xBF\x00\x80\xA6\x41"
-     "\x03\x00"
-     "\xA5\xF5\x0C\x00\x31\xBF\x00\x80\xAB\x5D\x03\x00\xAA\x5D\x03\x00\x35\xBF"
-     "\x00\x80"
-     "\x39\xBF\x00\x80\xAF\xE9\x03\x00\xAE\xE1\x03\x00\xAD\xF5\x03\x00\xAC\xFD"
-     "\x03\x00"
-     "\x80\x5D\x00\x00\x81\x6D\x00\x00\x82\x65\x00\x00\xB3\x45\x01\x00\x41\xBF"
-     "\x00\x80"
-     "\xB5\x45\x01\x00\xB6\x71\x01\x00\x45\xBF\x00\x80\x86\xC0\x04\x00\x87\x98"
-     "\x02\x00"
-     "\xBA\x21\x01\x00\xBB\x21\x01\x00\xBC\x25\x01\x00\xBD\x2D\x01\x00\xBE\x11"
-     "\x01\x00"
-     "\xBF\x11\x01\x00\x49\xBF\x00\x80\x4D\xBF\x00\x80\x51\xBF\x00\x80\x55\xBF"
-     "\x00\x80"
-     "\x59\xBF\x00\x80\x5D\xBF\x00\x80\xEF\x88\x0E\x00\x61\xBF\x00\x80\xE1\x70"
-     "\x0F\x00"
-     "\x65\xBF\x00\x80\xE3\x48\x0F\x00\x69\xBF\x00\x80\x6D\xBF\x00\x80\x71\xBF"
-     "\x00\x80"
-     "\x75\xBF\x00\x80\xBE\xB0\x04\x00\x79\xBF\x00\x80\x7D\xBF\x00\x80\x81\xBF"
-     "\x00\x80"
-     "\x85\xBF\x00\x80\x89\xBF\x00\x80\xE1\x04\x0E\x00\x8D\xBF\x00\x80\xE3\x60"
-     "\x00\x00"
-     "\x91\xBF\x00\x80\x95\xBF\x00\x80\x99\xBF\x00\x80\xEF\xA4\x00\x00\x9D\xBF"
-     "\x00\x80"
-     "\x82\xC5\x07\x00\x81\xD5\x07\x00\x80\xDD\x07\x00\xA8\xB5\x05\x00\xA9\x01"
-     "\x06\x00"
-     "\xAA\x01\x06\x00\xAB\x01\x06\x00\xAC\x01\x06\x00\xAD\x3D\x06\x00\xAE\x35"
-     "\x06\x00"
-     "\xAF\x5D\x06\x00\xA1\xBF\x00\x80\x3D\xBF\x00\x80\xCD\xBE\x00\x80\xA5\xBF"
-     "\x00\x80"
-     "\xA9\xBF\x00\x80\xAD\xBF\x00\x80\x86\x98\x00\x00\x87\xFC\x03\x00\xB8\xED"
-     "\x06\x00"
-     "\xB9\xF5\x06\x00\xBA\xFD\x06\x00\xBB\xF5\x06\x00\xBC\xED\x06\x00\xBD\x51"
-     "\x07\x00"
-     "\xBE\x55\x07\x00\xBF\x4D\x07\x00\xB0\x29\x06\x00\xB1\x29\x06\x00\xB2\x3D"
-     "\x06\x00"
-     "\xB3\x35\x06\x00\xB4\x19\x06\x00\xB5\x19\x06\x00\xB6\xD9\x06\x00\xB7\xD5"
-     "\x06\x00"
-     "\xA3\x89\x06\x00\xB1\xBF\x00\x80\xB5\xBF\x00\x80\xB9\xBF\x00\x80\xBD\xBF"
-     "\x00\x80"
-     "\xA6\xBD\x06\x00\xA5\x89\x06\x00\xC1\xBF\x00\x80\xAB\xED\x06\x00\xAA\xED"
-     "\x06\x00"
-     "\xC5\xBF\x00\x80\xC9\xBF\x00\x80\xAF\xDD\x06\x00\xAE\xDD\x06\x00\xAD\xE1"
-     "\x06\x00"
-     "\xAC\xE9\x06\x00\xCD\xBF\x00\x80\xB3\x45\x07\x00\xD1\xBF\x00\x80\xD5\xBF"
-     "\x00\x80"
-     "\xB6\x5D\x07\x00\xD9\xBF\x00\x80\xDD\xBF\x00\x80\xB5\x55\x07\x00\xBA\x7D"
-     "\x07\x00"
-     "\xBB\x41\x07\x00\xE1\xBF\x00\x80\xE5\xBF\x00\x80\xBE\x45\x07\x00\xBF\x4D"
-     "\x07\x00"
-     "\xBC\x55\x07\x00\xBD\x4D\x07\x00\xA8\x51\x06\x00\xA9\x59\x06\x00\xAA\x89"
-     "\x06\x00"
-     "\xAB\x89\x06\x00\xAC\x99\x06\x00\xAD\x99\x06\x00\xAE\x89\x06\x00\xAF\x89"
-     "\x06\x00"
-     "\xE9\xBF\x00\x80\xED\xBF\x00\x80\xF1\xBF\x00\x80\xF5\xBF\x00\x80\xF9\xBF"
-     "\x00\x80"
-     "\xFD\xBF\x00\x80\x01\xC0\x00\x80\x05\xC0\x00\x80\xB8\x89\x06\x00\xB9\x89"
-     "\x06\x00"
-     "\xBA\x99\x06\x00\xBB\x95\x06\x00\xBC\xB1\x06\x00\xBD\xB1\x06\x00\xBE\xB1"
-     "\x06\x00"
-     "\xBF\xAD\x06\x00\xB0\xFD\x06\x00\xB1\xC1\x06\x00\xB2\xC5\x06\x00\xB3\xDD"
-     "\x06\x00"
-     "\xB4\xC5\x06\x00\xB5\xCD\x06\x00\xB6\xB9\x06\x00\xB7\xB9\x06\x00\xA3\x01"
-     "\x06\x00"
-     "\x09\xC0\x00\x80\x0D\xC0\x00\x80\x11\xC0\x00\x80\x15\xC0\x00\x80\xA6\x19"
-     "\x06\x00"
-     "\xA5\x11\x06\x00\x19\xC0\x00\x80\xAB\x05\x06\x00\xAA\x39\x06\x00\x84\xA8"
-     "\x02\x00"
-     "\xBE\xA0\x01\x00\xAF\x09\x06\x00\xAE\x01\x06\x00\xAD\x09\x06\x00\xAC\x11"
-     "\x06\x00"
-     "\x80\x4D\x00\x00\x81\x55\x00\x00\x82\x55\x00\x00\xB3\x41\x01\x00\x21\xC0"
-     "\x00\x80"
-     "\xB5\x49\x01\x00\xB6\x71\x01\x00\x25\xC0\x00\x80\x86\x40\x00\x00\x87\x04"
-     "\x01\x00"
-     "\xBA\x21\x01\x00\xBB\x21\x01\x00\xBC\x25\x01\x00\xBD\x2D\x01\x00\xBE\x11"
-     "\x01\x00"
-     "\xBF\x11\x01\x00\xA8\x85\x02\x00\xA9\x8D\x02\x00\xAA\x85\x02\x00\xAB\x9D"
-     "\x02\x00"
-     "\xAC\x8D\x02\x00\xAD\xBD\x02\x00\xAE\xB5\x02\x00\xAF\xDD\x02\x00\xBF\x00"
-     "\x68\x00"
-     "\x29\xC0\x00\x80\x2D\xC0\x00\x80\x31\xC0\x00\x80\x35\xC0\x00\x80\x39\xC0"
-     "\x00\x80"
-     "\x3D\xC0\x00\x80\x41\xC0\x00\x80\xB8\x69\x03\x00\xB9\x69\x03\x00\xBA\x79"
-     "\x03\x00"
-     "\xBB\x79\x03\x00\xBC\x69\x03\x00\xBD\x69\x03\x00\xBE\xDD\x03\x00\xBF\xD5"
-     "\x03\x00"
-     "\xB0\xA9\x02\x00\xB1\xA9\x02\x00\xB2\xBD\x02\x00\xB3\xB5\x02\x00\xB4\x99"
-     "\x02\x00"
-     "\xB5\x99\x02\x00\xB6\x59\x03\x00\xB7\x59\x03\x00\xA3\x0D\x02\x00\x45\xC0"
-     "\x00\x80"
-     "\x49\xC0\x00\x80\x4D\xC0\x00\x80\x51\xC0\x00\x80\xA6\x3D\x02\x00\xA5\x05"
-     "\x02\x00"
-     "\x55\xC0\x00\x80\xAB\x6D\x02\x00\xAA\x6D\x02\x00\x59\xC0\x00\x80\x5D\xC0"
-     "\x00\x80"
-     "\xAF\x5D\x02\x00\xAE\x5D\x02\x00\xAD\x61\x02\x00\xAC\x69\x02\x00\x61\xC0"
-     "\x00\x80"
-     "\x65\xC0\x00\x80\x69\xC0\x00\x80\x6D\xC0\x00\x80\x71\xC0\x00\x80\x75\xC0"
-     "\x00\x80"
-     "\x79\xC0\x00\x80\x7D\xC0\x00\x80\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\x81\xC0\x00\x80\x85\xC0\x00\x80\x89\xC0\x00\x80\xBE\x64\x1C\x00\x91\xC0"
-     "\x00\x80"
-     "\xA8\xD1\x02\x00\xA9\xD9\x02\x00\xAA\x2D\x01\x00\xAB\x25\x01\x00\xAC\x3D"
-     "\x01\x00"
-     "\xAD\x25\x01\x00\xAE\x2D\x01\x00\xAF\x25\x01\x00\x95\xC0\x00\x80\x99\xC0"
-     "\x00\x80"
-     "\x9D\xC0\x00\x80\xA1\xC0\x00\x80\xA5\xC0\x00\x80\xA9\xC0\x00\x80\xAD\xC0"
-     "\x00\x80"
-     "\xB1\xC0\x00\x80\xB8\x81\x01\x00\xB9\x81\x01\x00\xBA\x81\x01\x00\xBB\x81"
-     "\x01\x00"
-     "\xBC\x81\x01\x00\xBD\x81\x01\x00\xBE\x81\x01\x00\xBF\x81\x01\x00\xB0\x5D"
-     "\x01\x00"
-     "\xB1\xE1\x01\x00\xB2\xE1\x01\x00\xB3\xE1\x01\x00\xB4\xE1\x01\x00\xB5\xE9"
-     "\x01\x00"
-     "\xB6\xD1\x01\x00\xB7\xD1\x01\x00\xB5\xC0\x00\x80\x84\x88\x02\x00\xEF\xDC"
-     "\x00\x00"
-     "\xB9\xC0\x00\x80\xBD\xC0\x00\x80\xEF\x64\x06\x00\xEF\xDC\x06\x00\xC1\xC0"
-     "\x00\x80"
-     "\xE1\xD8\x06\x00\xE3\xA8\x07\x00\xE3\x1C\x07\x00\xE1\x0C\x04\x00\xE1\x90"
-     "\x01\x00"
-     "\xC5\xC0\x00\x80\xE3\x00\x00\x00\xC9\xC0\x00\x80\xB3\xE5\x01\x00\xCD\xC0"
-     "\x00\x80"
-     "\x87\x88\x1D\x00\x86\xEC\x1D\x00\xD1\xC0\x00\x80\xB6\xE5\x01\x00\xB5\xF5"
-     "\x01\x00"
-     "\xD5\xC0\x00\x80\xBB\x65\x01\x00\xBA\xD9\x01\x00\xD9\xC0\x00\x80\xDD\xC0"
-     "\x00\x80"
-     "\xBF\x6D\x01\x00\xBE\x65\x01\x00\xBD\x75\x01\x00\xBC\x75\x01\x00\x8D\xC0"
-     "\x00\x80"
-     "\xA6\x35\x1E\x00\xA5\x25\x1E\x00\xE1\xC0\x00\x80\xA3\x35\x1E\x00\xE5\xC0"
-     "\x00\x80"
-     "\xE9\xC0\x00\x80\xED\xC0\x00\x80\xAF\xBD\x1E\x00\xAE\xB5\x1E\x00\xAD\xA5"
-     "\x1E\x00"
-     "\xAC\xA5\x1E\x00\xAB\xB5\x1E\x00\xAA\x09\x1E\x00\xF1\xC0\x00\x80\x1D\xC0"
-     "\x00\x80"
-     "\x80\x79\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\xF5\xC0\x00\x80\xF9\xC0"
-     "\x00\x80"
-     "\x84\x98\x03\x00\xBE\x90\x03\x00\xFD\xC0\x00\x80\x86\x40\x03\x00\x87\xC4"
-     "\x03\x00"
-     "\x01\xC1\x00\x80\x05\xC1\x00\x80\x09\xC1\x00\x80\x0D\xC1\x00\x80\x11\xC1"
-     "\x00\x80"
-     "\x15\xC1\x00\x80\xA8\x7D\x1E\x00\xA9\xA5\x1F\x00\xAA\xAD\x1F\x00\xAB\xA5"
-     "\x1F\x00"
-     "\xAC\xBD\x1F\x00\xAD\xA5\x1F\x00\xAE\xAD\x1F\x00\xAF\xA5\x1F\x00\x19\xC1"
-     "\x00\x80"
-     "\x1D\xC1\x00\x80\x21\xC1\x00\x80\x25\xC1\x00\x80\x29\xC1\x00\x80\x2D\xC1"
-     "\x00\x80"
-     "\x31\xC1\x00\x80\x35\xC1\x00\x80\xB8\x11\x1F\x00\xB9\x19\x1F\x00\xBA\x21"
-     "\x1F\x00"
-     "\xBB\x21\x1F\x00\xBC\xDD\x00\x00\xBD\xC1\x00\x00\xBE\xC1\x00\x00\xBF\xC1"
-     "\x00\x00"
-     "\xB0\xDD\x1F\x00\xB1\x65\x1F\x00\xB2\x65\x1F\x00\xB3\x7D\x1F\x00\xB4\x6D"
-     "\x1F\x00"
-     "\xB5\x41\x1F\x00\xB6\x45\x1F\x00\xB7\x39\x1F\x00\xB3\x25\x1E\x00\x39\xC1"
-     "\x00\x80"
-     "\x3D\xC1\x00\x80\x41\xC1\x00\x80\x45\xC1\x00\x80\xB6\x25\x1E\x00\xB5\x35"
-     "\x1E\x00"
-     "\x49\xC1\x00\x80\xBB\xA1\x1E\x00\xBA\x19\x1E\x00\x4D\xC1\x00\x80\x51\xC1"
-     "\x00\x80"
-     "\xBF\x85\x1E\x00\xBE\xA9\x1E\x00\xBD\xA9\x1E\x00\xBC\xB1\x1E\x00\x55\xC1"
-     "\x00\x80"
-     "\xA3\x61\x1E\x00\x59\xC1\x00\x80\x5D\xC1\x00\x80\xA6\x61\x1E\x00\x61\xC1"
-     "\x00\x80"
-     "\x65\xC1\x00\x80\xA5\x71\x1E\x00\xAA\x5D\x1E\x00\xAB\xE5\x1E\x00\x69\xC1"
-     "\x00\x80"
-     "\x84\xA0\x01\x00\xAE\xED\x1E\x00\xAF\xC1\x1E\x00\xAC\xF5\x1E\x00\xAD\xED"
-     "\x1E\x00"
-     "\xA8\xE9\x01\x00\xA9\xE9\x01\x00\xAA\xF9\x01\x00\xAB\xF9\x01\x00\xAC\xE9"
-     "\x01\x00"
-     "\xAD\xE9\x01\x00\xAE\x39\x01\x00\xAF\x39\x01\x00\x80\xF5\x00\x00\x81\xFD"
-     "\x00\x00"
-     "\x82\xC1\x00\x00\xBE\x64\x01\x00\x86\x90\x01\x00\x87\xA0\x00\x00\x71\xC1"
-     "\x00\x80"
-     "\x75\xC1\x00\x80\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5\x00\x00\xBB\xE5"
-     "\x00\x00"
-     "\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x91\x00\x00\xBF\x91\x00\x00\xB0\x49"
-     "\x01\x00"
-     "\xB1\x49\x01\x00\xB2\x59\x01\x00\xB3\x59\x01\x00\xB4\x49\x01\x00\xB5\x49"
-     "\x01\x00"
-     "\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xB3\x71\x02\x00\x79\xC1\x00\x80\x7D\xC1"
-     "\x00\x80"
-     "\x81\xC1\x00\x80\x85\xC1\x00\x80\xB6\x69\x02\x00\xB5\x61\x02\x00\x89\xC1"
-     "\x00\x80"
-     "\xBB\x8D\x02\x00\xBA\x8D\x02\x00\x8D\xC1\x00\x80\x91\xC1\x00\x80\xBF\x8D"
-     "\x02\x00"
-     "\xBE\x8D\x02\x00\xBD\x8D\x02\x00\xBC\x8D\x02\x00\x85\x5C\x73\x00\xA3\x35"
-     "\x02\x00"
-     "\x95\xC1\x00\x80\x99\xC1\x00\x80\xA6\x2D\x02\x00\x9D\xC1\x00\x80\xA1\xC1"
-     "\x00\x80"
-     "\xA5\x25\x02\x00\xAA\xC9\x02\x00\xAB\xC9\x02\x00\xA5\xC1\x00\x80\xA9\xC1"
-     "\x00\x80"
-     "\xAE\xC9\x02\x00\xAF\xC9\x02\x00\xAC\xC9\x02\x00\xAD\xC9\x02\x00\xA8\x39"
-     "\x03\x00"
-     "\xA9\x39\x03\x00\xAA\x8D\x00\x00\xAB\x85\x00\x00\xAC\x9D\x00\x00\xAD\x85"
-     "\x00\x00"
-     "\xAE\x85\x00\x00\xAF\xB5\x00\x00\xBE\x28\x04\x00\xAD\xC1\x00\x80\xB1\xC1"
-     "\x00\x80"
-     "\xB5\xC1\x00\x80\xB9\xC1\x00\x80\xBD\xC1\x00\x80\xC1\xC1\x00\x80\xC5\xC1"
-     "\x00\x80"
-     "\xB8\x69\x00\x00\xB9\x69\x00\x00\xBA\x79\x00\x00\xBB\x79\x00\x00\xBC\x69"
-     "\x00\x00"
-     "\xBD\x69\x00\x00\xBE\xD9\x01\x00\xBF\xD9\x01\x00\xB0\xCD\x00\x00\xB1\xAD"
-     "\x00\x00"
-     "\xB2\xA5\x00\x00\xB3\xB9\x00\x00\xB4\xA9\x00\x00\xB5\x9D\x00\x00\xB6\x95"
-     "\x00\x00"
-     "\xB7\x59\x00\x00\xC9\xC1\x00\x80\xCD\xC1\x00\x80\xD1\xC1\x00\x80\xD5\xC1"
-     "\x00\x80"
-     "\x81\x1D\x00\x00\x80\x1D\x00\x00\xD9\xC1\x00\x80\x82\x05\x00\x00\x84\xAC"
-     "\x02\x00"
-     "\xE1\xC1\x00\x80\x87\xC4\x02\x00\x86\x0C\x05\x00\xE5\xC1\x00\x80\xE9\xC1"
-     "\x00\x80"
-     "\xED\xC1\x00\x80\xF1\xC1\x00\x80\xB3\x45\x01\x00\xF5\xC1\x00\x80\xF9\xC1"
-     "\x00\x80"
-     "\xFD\xC1\x00\x80\x01\xC2\x00\x80\xB6\x45\x01\x00\xB5\x55\x01\x00\x05\xC2"
-     "\x00\x80"
-     "\xBB\x29\x01\x00\xBA\x49\x01\x00\x09\xC2\x00\x80\x0D\xC2\x00\x80\xBF\x19"
-     "\x01\x00"
-     "\xBE\x2D\x01\x00\xBD\x2D\x01\x00\xBC\x31\x01\x00\xEF\x80\x1F\x00\xE1\x50"
-     "\x1E\x00"
-     "\x11\xC2\x00\x80\xE3\xC8\x01\x00\x15\xC2\x00\x80\x19\xC2\x00\x80\x1D\xC2"
-     "\x00\x80"
-     "\x21\xC2\x00\x80\x84\x2C\x05\x00\x25\xC2\x00\x80\xBE\x24\x04\x00\x29\xC2"
-     "\x00\x80"
-     "\xE3\x24\x1E\x00\x2D\xC2\x00\x80\xE1\x40\x1C\x00\xEF\x1C\x01\x00\x31\xC2"
-     "\x00\x80"
-     "\x35\xC2\x00\x80\x86\xE8\x04\x00\x87\x8C\x04\x00\xE3\xAC\x00\x00\x39\xC2"
-     "\x00\x80"
-     "\xE1\xA0\x01\x00\x3D\xC2\x00\x80\xEF\xB4\x00\x00\x82\x1D\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x41\xC2\x00\x80\x45\xC2\x00\x80\x49\xC2\x00\x80\x51\xC2"
-     "\x00\x80"
-     "\x6D\xC1\x00\x80\xA3\x01\x06\x00\x55\xC2\x00\x80\xDD\xC1\x00\x80\xA6\x01"
-     "\x06\x00"
-     "\x4D\xC2\x00\x80\x59\xC2\x00\x80\xA5\x11\x06\x00\xAA\x0D\x06\x00\xAB\x6D"
-     "\x06\x00"
-     "\x5D\xC2\x00\x80\x61\xC2\x00\x80\xAE\x69\x06\x00\xAF\x5D\x06\x00\xAC\x75"
-     "\x06\x00"
-     "\xAD\x69\x06\x00\xB3\x75\x07\x00\x65\xC2\x00\x80\x69\xC2\x00\x80\x6D\xC2"
-     "\x00\x80"
-     "\x71\xC2\x00\x80\xB6\x11\x07\x00\xB5\x65\x07\x00\x75\xC2\x00\x80\xBB\x3D"
-     "\x07\x00"
-     "\xBA\x35\x07\x00\x79\xC2\x00\x80\x7D\xC2\x00\x80\xBF\xE1\x07\x00\xBE\xFD"
-     "\x07\x00"
-     "\xBD\x15\x07\x00\xBC\x15\x07\x00\xA8\xA5\x07\x00\xA9\xB5\x07\x00\xAA\xBD"
-     "\x07\x00"
-     "\xAB\xB5\x07\x00\xAC\xAD\x07\x00\xAD\xD1\x07\x00\xAE\xD1\x07\x00\xAF\xD1"
-     "\x07\x00"
-     "\x81\xC2\x00\x80\x85\xC2\x00\x80\x89\xC2\x00\x80\x8D\xC2\x00\x80\x91\xC2"
-     "\x00\x80"
-     "\x95\xC2\x00\x80\x99\xC2\x00\x80\x9D\xC2\x00\x80\xB8\x5D\x07\x00\xB9\x61"
-     "\x07\x00"
-     "\xBA\x65\x07\x00\xBB\x79\x07\x00\xBC\x11\x07\x00\xBD\x11\x07\x00\xBE\x0D"
-     "\x07\x00"
-     "\xBF\x01\x07\x00\xB0\xB1\x07\x00\xB1\xB9\x07\x00\xB2\x81\x07\x00\xB3\x81"
-     "\x07\x00"
-     "\xB4\x79\x07\x00\xB5\x65\x07\x00\xB6\x6D\x07\x00\xB7\x65\x07\x00\xA3\x35"
-     "\x06\x00"
-     "\xA1\xC2\x00\x80\xA5\xC2\x00\x80\xA9\xC2\x00\x80\xAD\xC2\x00\x80\xA6\x51"
-     "\x06\x00"
-     "\xA5\x25\x06\x00\xB1\xC2\x00\x80\xAB\x7D\x06\x00\xAA\x75\x06\x00\xB5\xC2"
-     "\x00\x80"
-     "\xB9\xC2\x00\x80\xAF\xA1\x06\x00\xAE\xBD\x06\x00\xAD\x55\x06\x00\xAC\x55"
-     "\x06\x00"
-     "\x80\x5D\x00\x00\x81\x6D\x00\x00\x82\x65\x00\x00\x84\x80\x01\x00\xBE\x98"
-     "\x01\x00"
-     "\xBD\xC2\x00\x80\x87\xD8\x01\x00\x86\xFC\x00\x00\xC1\xC2\x00\x80\xC5\xC2"
-     "\x00\x80"
-     "\xC9\xC2\x00\x80\xCD\xC2\x00\x80\xD1\xC2\x00\x80\xD5\xC2\x00\x80\xD9\xC2"
-     "\x00\x80"
-     "\xDD\xC2\x00\x80\xA8\x7D\x06\x00\xA9\x8D\x01\x00\xAA\x91\x01\x00\xAB\xA5"
-     "\x01\x00"
-     "\xAC\xBD\x01\x00\xAD\xD1\x01\x00\xAE\xD1\x01\x00\xAF\xD1\x01\x00\x84\xE0"
-     "\x01\x00"
-     "\xE1\xC2\x00\x80\xE5\xC2\x00\x80\xE9\xC2\x00\x80\xED\xC2\x00\x80\xF1\xC2"
-     "\x00\x80"
-     "\xF5\xC2\x00\x80\xF9\xC2\x00\x80\xB8\x71\x01\x00\xB9\x71\x01\x00\xBA\x71"
-     "\x01\x00"
-     "\xBB\x71\x01\x00\xBC\xD5\x00\x00\xBD\xDD\x00\x00\xBE\xD5\x00\x00\xBF\xCD"
-     "\x00\x00"
-     "\xB0\xB1\x01\x00\xB1\xB9\x01\x00\xB2\x81\x01\x00\xB3\x81\x01\x00\xB4\x51"
-     "\x01\x00"
-     "\xB5\x51\x01\x00\xB6\x51\x01\x00\xB7\x51\x01\x00\xB3\x19\x02\x00\xFD\xC2"
-     "\x00\x80"
-     "\x01\xC3\x00\x80\x05\xC3\x00\x80\x09\xC3\x00\x80\xB6\x51\x02\x00\xB5\x35"
-     "\x02\x00"
-     "\x0D\xC3\x00\x80\xBB\x7D\x02\x00\xBA\x75\x02\x00\x11\xC3\x00\x80\x15\xC3"
-     "\x00\x80"
-     "\xBF\x95\x02\x00\xBE\x95\x02\x00\xBD\x55\x02\x00\xBC\x55\x02\x00\x19\xC3"
-     "\x00\x80"
-     "\xA3\x5D\x02\x00\x1D\xC3\x00\x80\x21\xC3\x00\x80\xA6\x15\x02\x00\x25\xC3"
-     "\x00\x80"
-     "\x29\xC3\x00\x80\xA5\x71\x02\x00\xAA\x31\x02\x00\xAB\x39\x02\x00\xBE\x24"
-     "\x03\x00"
-     "\x31\xC3\x00\x80\xAE\xD1\x02\x00\xAF\xD1\x02\x00\xAC\x11\x02\x00\xAD\x11"
-     "\x02\x00"
-     "\x80\xCD\x00\x00\x81\xD1\x00\x00\x82\xD1\x00\x00\xB3\xE5\x03\x00\x35\xC3"
-     "\x00\x80"
-     "\xB5\xE5\x03\x00\xB6\xED\x03\x00\x39\xC3\x00\x80\x3D\xC3\x00\x80\x41\xC3"
-     "\x00\x80"
-     "\xBA\x0D\x03\x00\xBB\x05\x03\x00\xBC\x1D\x03\x00\xBD\x05\x03\x00\xBE\x0D"
-     "\x03\x00"
-     "\xBF\x05\x03\x00\x86\x40\x0C\x00\x87\x08\x03\x00\x45\xC3\x00\x80\x49\xC3"
-     "\x00\x80"
-     "\xE1\x94\x01\x00\x4D\xC3\x00\x80\xE3\x68\x01\x00\x51\xC3\x00\x80\x55\xC3"
-     "\x00\x80"
-     "\x59\xC3\x00\x80\xEF\x78\x02\x00\x5D\xC3\x00\x80\x61\xC3\x00\x80\x65\xC3"
-     "\x00\x80"
-     "\x69\xC3\x00\x80\x6D\xC3\x00\x80\xA8\x61\x02\x00\xA9\x61\x02\x00\xAA\x61"
-     "\x02\x00"
-     "\xAB\x61\x02\x00\xAC\x61\x02\x00\xAD\x61\x02\x00\xAE\x61\x02\x00\xAF\x61"
-     "\x02\x00"
-     "\x84\x6C\x0D\x00\x71\xC3\x00\x80\x75\xC3\x00\x80\x79\xC3\x00\x80\x7D\xC3"
-     "\x00\x80"
-     "\x81\xC3\x00\x80\x85\xC3\x00\x80\x89\xC3\x00\x80\xB8\x49\x01\x00\xB9\x49"
-     "\x01\x00"
-     "\xBA\x59\x01\x00\xBB\x51\x01\x00\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x19"
-     "\x01\x00"
-     "\xBF\x15\x01\x00\xB0\xC5\x02\x00\xB1\xCD\x02\x00\xB2\xC5\x02\x00\xB3\xDD"
-     "\x02\x00"
-     "\xB4\xC5\x02\x00\xB5\xCD\x02\x00\xB6\xC5\x02\x00\xB7\x79\x01\x00\xE3\x30"
-     "\x06\x00"
-     "\xE1\x2C\x04\x00\xE1\xBC\x06\x00\xE3\xBC\x06\x00\xE1\x40\x06\x00\x8D\xC3"
-     "\x00\x80"
-     "\xE3\xC4\x00\x00\x91\xC3\x00\x80\x95\xC3\x00\x80\x99\xC3\x00\x80\xEF\x00"
-     "\x00\x00"
-     "\xA1\xC3\x00\x80\xEF\xE0\x06\x00\xA5\xC3\x00\x80\xBE\xF0\x0D\x00\xEF\xE4"
-     "\x07\x00"
-     "\xA3\xA9\x02\x00\x82\x61\x00\x00\x81\x69\x00\x00\x80\x51\x00\x00\xA9\xC3"
-     "\x00\x80"
-     "\xA6\xA1\x02\x00\xA5\xA9\x02\x00\xAD\xC3\x00\x80\xAB\x49\x02\x00\xAA\x41"
-     "\x02\x00"
-     "\x86\x48\x0C\x00\x87\xCC\x0C\x00\xAF\x49\x02\x00\xAE\x41\x02\x00\xAD\x49"
-     "\x02\x00"
-     "\xAC\x51\x02\x00\x9D\xC3\x00\x80\xB1\xC3\x00\x80\xB5\xC3\x00\x80\xB9\xC3"
-     "\x00\x80"
-     "\x2D\xC3\x00\x80\xBD\xC3\x00\x80\xC1\xC3\x00\x80\xC5\xC3\x00\x80\xA8\x4D"
-     "\x0E\x00"
-     "\xA9\xB5\x0E\x00\xAA\xB5\x0E\x00\xAB\x8D\x0E\x00\xAC\x95\x0E\x00\xAD\xFD"
-     "\x0E\x00"
-     "\xAE\xF5\x0E\x00\xAF\xE9\x0E\x00\xB0\x9D\x0E\x00\xB1\xA5\x0E\x00\xB2\xAD"
-     "\x0E\x00"
-     "\xB3\xA1\x0E\x00\xB4\xA1\x0E\x00\xB5\xA1\x0E\x00\xB6\xA1\x0E\x00\xB7\xA1"
-     "\x0E\x00"
-     "\xB8\x45\x0F\x00\xB9\x4D\x0F\x00\xBA\x45\x0F\x00\xBB\x5D\x0F\x00\xBC\x45"
-     "\x0F\x00"
-     "\xBD\x4D\x0F\x00\xBE\x45\x0F\x00\xBF\xE9\x0F\x00\xB3\x1D\x0E\x00\xC9\xC3"
-     "\x00\x80"
-     "\xCD\xC3\x00\x80\xD1\xC3\x00\x80\xD5\xC3\x00\x80\xB6\x5D\x0E\x00\xB5\x3D"
-     "\x0E\x00"
-     "\xD9\xC3\x00\x80\xBB\x41\x0E\x00\xBA\x79\x0E\x00\xDD\xC3\x00\x80\xE1\xC3"
-     "\x00\x80"
-     "\xBF\x45\x0E\x00\xBE\x45\x0E\x00\xBD\x55\x0E\x00\xBC\x59\x0E\x00\xE5\xC3"
-     "\x00\x80"
-     "\xA3\x59\x0E\x00\xE9\xC3\x00\x80\xED\xC3\x00\x80\xA6\x19\x0E\x00\xF1\xC3"
-     "\x00\x80"
-     "\xF5\xC3\x00\x80\xA5\x79\x0E\x00\xAA\x3D\x0E\x00\xAB\x05\x0E\x00\xF9\xC3"
-     "\x00\x80"
-     "\xFD\xC3\x00\x80\xAE\x01\x0E\x00\xAF\x01\x0E\x00\xAC\x1D\x0E\x00\xAD\x11"
-     "\x0E\x00"
-     "\xA8\xAD\x0E\x00\xA9\xB1\x0E\x00\xAA\xD9\x0E\x00\xAB\xD1\x0E\x00\xAC\x15"
-     "\x01\x00"
-     "\xAD\x1D\x01\x00\xAE\x15\x01\x00\xAF\x09\x01\x00\x80\x0D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x1D\x00\x00\x01\xC4\x00\x80\x05\xC4\x00\x80\x09\xC4\x00\x80\x84\x74"
-     "\x02\x00"
-     "\xBE\x74\x02\x00\xB8\xED\x01\x00\xB9\x85\x01\x00\xBA\x81\x01\x00\xBB\x81"
-     "\x01\x00"
-     "\xBC\x85\x01\x00\xBD\x8D\x01\x00\xBE\xB1\x01\x00\xBF\xB1\x01\x00\xB0\x79"
-     "\x01\x00"
-     "\xB1\x79\x01\x00\xB2\xED\x01\x00\xB3\xE5\x01\x00\xB4\xFD\x01\x00\xB5\xE5"
-     "\x01\x00"
-     "\xB6\xE5\x01\x00\xB7\xD5\x01\x00\xB3\x51\x0E\x00\x0D\xC4\x00\x80\x86\xA8"
-     "\x01\x00"
-     "\x87\xCC\x01\x00\x11\xC4\x00\x80\xB6\x9D\x01\x00\xB5\x9D\x01\x00\x15\xC4"
-     "\x00\x80"
-     "\xBB\xBD\x01\x00\xBA\xBD\x01\x00\x19\xC4\x00\x80\x1D\xC4\x00\x80\xBF\x21"
-     "\x01\x00"
-     "\xBE\x39\x01\x00\xBD\x31\x01\x00\xBC\x39\x01\x00\x21\xC4\x00\x80\xA3\x15"
-     "\x0E\x00"
-     "\x25\xC4\x00\x80\x29\xC4\x00\x80\xA6\xD9\x01\x00\x2D\xC4\x00\x80\x31\xC4"
-     "\x00\x80"
-     "\xA5\xD9\x01\x00\xAA\xF9\x01\x00\xAB\xF9\x01\x00\x35\xC4\x00\x80\x39\xC4"
-     "\x00\x80"
-     "\xAE\x7D\x01\x00\xAF\x65\x01\x00\xAC\x7D\x01\x00\xAD\x75\x01\x00\xE1\x90"
-     "\x01\x00"
-     "\x3D\xC4\x00\x80\xE3\x54\x02\x00\x41\xC4\x00\x80\x45\xC4\x00\x80\x49\xC4"
-     "\x00\x80"
-     "\x4D\xC4\x00\x80\x51\xC4\x00\x80\x55\xC4\x00\x80\x59\xC4\x00\x80\x5D\xC4"
-     "\x00\x80"
-     "\x61\xC4\x00\x80\x65\xC4\x00\x80\x69\xC4\x00\x80\xEF\x9C\x02\x00\x6D\xC4"
-     "\x00\x80"
-     "\x80\x11\x00\x00\x81\x11\x00\x00\x82\x11\x00\x00\x84\x88\x03\x00\x71\xC4"
-     "\x00\x80"
-     "\x75\xC4\x00\x80\x79\xC4\x00\x80\xBE\x80\x03\x00\x86\x80\x04\x00\x87\xC4"
-     "\x03\x00"
-     "\x81\xC4\x00\x80\x85\xC4\x00\x80\x89\xC4\x00\x80\x8D\xC4\x00\x80\x91\xC4"
-     "\x00\x80"
-     "\x95\xC4\x00\x80\xA8\x11\x03\x00\xA9\x19\x03\x00\xAA\x21\x03\x00\xAB\x21"
-     "\x03\x00"
-     "\xAC\x51\x03\x00\xAD\x5D\x03\x00\xAE\x55\x03\x00\xAF\x49\x03\x00\x99\xC4"
-     "\x00\x80"
-     "\x9D\xC4\x00\x80\xA1\xC4\x00\x80\xA5\xC4\x00\x80\xA9\xC4\x00\x80\xAD\xC4"
-     "\x00\x80"
-     "\xB1\xC4\x00\x80\xB5\xC4\x00\x80\xB8\xDD\x00\x00\xB9\xE5\x00\x00\xBA\xED"
-     "\x00\x00"
-     "\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\xE5\x00\x00\xBE\xED\x00\x00\xBF\xE1"
-     "\x00\x00"
-     "\xB0\x39\x03\x00\xB1\x0D\x03\x00\xB2\x05\x03\x00\xB3\xE9\x00\x00\xB4\xFD"
-     "\x00\x00"
-     "\xB5\xE5\x00\x00\xB6\xED\x00\x00\xB7\xE5\x00\x00\xB9\xC4\x00\x80\xBD\xC4"
-     "\x00\x80"
-     "\xB3\xA1\x02\x00\xC1\xC4\x00\x80\xB5\xD1\x02\x00\xC5\xC4\x00\x80\xC9\xC4"
-     "\x00\x80"
-     "\xB6\xD5\x02\x00\xC1\x00\x00\x00\xCD\xC4\x00\x80\xBB\xC1\x02\x00\xBA\xF5"
-     "\x02\x00"
-     "\xBD\x35\x01\x00\xBC\xD9\x02\x00\xBF\x29\x01\x00\xBE\x31\x01\x00\xD1\xC4"
-     "\x00\x80"
-     "\xA6\x8D\x02\x00\xA5\x89\x02\x00\x84\x40\x02\x00\xA3\xF9\x02\x00\xD5\xC4"
-     "\x00\x80"
-     "\xD9\xC4\x00\x80\xDD\xC4\x00\x80\xAF\x71\x01\x00\xAE\x69\x01\x00\xAD\x6D"
-     "\x01\x00"
-     "\xAC\x81\x02\x00\xAB\x99\x02\x00\xAA\xAD\x02\x00\xE1\xC4\x00\x80\xE5\xC4"
-     "\x00\x80"
-     "\xE9\xC4\x00\x80\xE1\x74\x0E\x00\xBE\x84\x05\x00\xE3\xA4\x00\x00\x84\xDC"
-     "\x04\x00"
-     "\xE1\x4C\x0F\x00\xF1\xC4\x00\x80\xE3\xAC\x0E\x00\x80\xD9\x00\x00\x81\xD9"
-     "\x00\x00"
-     "\x82\x29\x00\x00\xEF\xEC\x0E\x00\xF5\xC4\x00\x80\xF9\xC4\x00\x80\xFD\xC4"
-     "\x00\x80"
-     "\xEF\x78\x00\x00\xB3\xB1\x01\x00\xBE\x8C\x04\x00\x86\x68\x04\x00\x87\x1C"
-     "\x05\x00"
-     "\x01\xC5\x00\x80\xB6\xA9\x01\x00\xB5\xA1\x01\x00\x05\xC5\x00\x80\xBB\x4D"
-     "\x01\x00"
-     "\xBA\x4D\x01\x00\x09\xC5\x00\x80\x0D\xC5\x00\x80\xBF\x4D\x01\x00\xBE\x4D"
-     "\x01\x00"
-     "\xBD\x4D\x01\x00\xBC\x4D\x01\x00\xB8\x0D\x07\x00\xB9\x11\x07\x00\xBA\x11"
-     "\x07\x00"
-     "\xBB\x29\x07\x00\xBC\x51\x07\x00\xBD\x41\x07\x00\xBE\x41\x07\x00\xBF\x41"
-     "\x07\x00"
-     "\xB0\x89\x06\x00\xB1\x89\x06\x00\xB2\x99\x06\x00\xB3\x99\x06\x00\xB4\x89"
-     "\x06\x00"
-     "\xB5\x89\x06\x00\xB6\x39\x07\x00\xB7\x39\x07\x00\xA8\x29\x06\x00\xA9\x29"
-     "\x06\x00"
-     "\xAA\x39\x06\x00\xAB\x39\x06\x00\xAC\x29\x06\x00\xAD\x29\x06\x00\xAE\xF9"
-     "\x06\x00"
-     "\xAF\xF9\x06\x00\xED\xC4\x00\x80\x7D\xC4\x00\x80\x11\xC5\x00\x80\x15\xC5"
-     "\x00\x80"
-     "\x19\xC5\x00\x80\x1D\xC5\x00\x80\x21\xC5\x00\x80\x25\xC5\x00\x80\xA3\xF1"
-     "\x06\x00"
-     "\x29\xC5\x00\x80\x2D\xC5\x00\x80\x31\xC5\x00\x80\x35\xC5\x00\x80\xA6\xE9"
-     "\x06\x00"
-     "\xA5\xE1\x06\x00\x39\xC5\x00\x80\xAB\x0D\x06\x00\xAA\x0D\x06\x00\x3D\xC5"
-     "\x00\x80"
-     "\x41\xC5\x00\x80\xAF\x0D\x06\x00\xAE\x0D\x06\x00\xAD\x0D\x06\x00\xAC\x0D"
-     "\x06\x00"
-     "\x45\xC5\x00\x80\x49\xC5\x00\x80\x4D\xC5\x00\x80\x51\xC5\x00\x80\x80\x1D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x39\x00\x00\x55\xC5\x00\x80\x59\xC5\x00\x80\x5D\xC5"
-     "\x00\x80"
-     "\x86\x48\x00\x00\x87\xCC\x00\x00\x61\xC5\x00\x80\x65\xC5\x00\x80\x69\xC5"
-     "\x00\x80"
-     "\x6D\xC5\x00\x80\xA8\x99\x06\x00\xA9\xB5\x06\x00\xAA\xBD\x06\x00\xAB\xB5"
-     "\x06\x00"
-     "\xAC\xC1\x06\x00\xAD\xC9\x06\x00\xAE\xF1\x06\x00\xAF\xF1\x06\x00\x71\xC5"
-     "\x00\x80"
-     "\x75\xC5\x00\x80\x79\xC5\x00\x80\x7D\xC5\x00\x80\x81\xC5\x00\x80\x85\xC5"
-     "\x00\x80"
-     "\x89\xC5\x00\x80\x8D\xC5\x00\x80\xB8\xCD\x01\x00\xB9\xD5\x01\x00\xBA\xD5"
-     "\x01\x00"
-     "\xBB\xE5\x01\x00\xBC\xFD\x01\x00\xBD\x95\x01\x00\xBE\x9D\x01\x00\xBF\x95"
-     "\x01\x00"
-     "\xB0\x09\x01\x00\xB1\x09\x01\x00\xB2\x19\x01\x00\xB3\x19\x01\x00\xB4\x09"
-     "\x01\x00"
-     "\xB5\x09\x01\x00\xB6\xFD\x01\x00\xB7\xF5\x01\x00\xB3\x35\x06\x00\x91\xC5"
-     "\x00\x80"
-     "\x95\xC5\x00\x80\x99\xC5\x00\x80\x9D\xC5\x00\x80\xB6\x49\x06\x00\xB5\x49"
-     "\x06\x00"
-     "\xA1\xC5\x00\x80\xBB\xCD\x01\x00\xBA\xCD\x01\x00\xA5\xC5\x00\x80\xA9\xC5"
-     "\x00\x80"
-     "\xBF\xCD\x01\x00\xBE\xCD\x01\x00\xBD\xCD\x01\x00\xBC\xCD\x01\x00\xAD\xC5"
-     "\x00\x80"
-     "\xA3\x71\x06\x00\xB1\xC5\x00\x80\xB5\xC5\x00\x80\xA6\x0D\x06\x00\xB9\xC5"
-     "\x00\x80"
-     "\xBD\xC5\x00\x80\xA5\x0D\x06\x00\xAA\x89\x01\x00\xAB\x89\x01\x00\xC1\xC5"
-     "\x00\x80"
-     "\xC5\xC5\x00\x80\xAE\x89\x01\x00\xAF\x89\x01\x00\xAC\x89\x01\x00\xAD\x89"
-     "\x01\x00"
-     "\xA8\xF1\x02\x00\xA9\xF1\x02\x00\xAA\xF1\x02\x00\xAB\xF1\x02\x00\xAC\x31"
-     "\x03\x00"
-     "\xAD\x31\x03\x00\xAE\x31\x03\x00\xAF\x31\x03\x00\x80\xC9\x03\x00\x81\xC9"
-     "\x03\x00"
-     "\x82\x05\x00\x00\xC9\xC5\x00\x80\xD1\xC5\x00\x80\x84\x78\x02\x00\xD5\xC5"
-     "\x00\x80"
-     "\xD9\xC5\x00\x80\xB8\xDD\x03\x00\xB9\xED\x03\x00\xBA\xE5\x03\x00\xBB\x8D"
-     "\x03\x00"
-     "\xBC\x99\x03\x00\xBD\x99\x03\x00\xBE\x8D\x03\x00\xBF\x85\x03\x00\xB0\x51"
-     "\x03\x00"
-     "\xB1\x51\x03\x00\xB2\x51\x03\x00\xB3\x51\x03\x00\xB4\xF5\x03\x00\xB5\xFD"
-     "\x03\x00"
-     "\xB6\xF5\x03\x00\xB7\xED\x03\x00\xB3\x79\x02\x00\xDD\xC5\x00\x80\x87\xC8"
-     "\x03\x00"
-     "\x86\x6C\x7C\x00\xE1\xC5\x00\x80\xB6\xB1\x03\x00\xB5\xB9\x03\x00\xE5\xC5"
-     "\x00\x80"
-     "\xBB\x95\x03\x00\xBA\x95\x03\x00\xBE\xE0\x02\x00\xE9\xC5\x00\x80\xBF\x39"
-     "\x03\x00"
-     "\xBE\x31\x03\x00\xBD\x85\x03\x00\xBC\x85\x03\x00\xED\xC5\x00\x80\xA3\x3D"
-     "\x02\x00"
-     "\xF1\xC5\x00\x80\xF5\xC5\x00\x80\xA6\xF5\x03\x00\xF9\xC5\x00\x80\xFD\xC5"
-     "\x00\x80"
-     "\xA5\xFD\x03\x00\xAA\xD1\x03\x00\xAB\xD1\x03\x00\x01\xC6\x00\x80\x05\xC6"
-     "\x00\x80"
-     "\xAE\x75\x03\x00\xAF\x7D\x03\x00\xAC\xC1\x03\x00\xAD\xC1\x03\x00\x09\xC6"
-     "\x00\x80"
-     "\x0D\xC6\x00\x80\x11\xC6\x00\x80\x15\xC6\x00\x80\x19\xC6\x00\x80\x1D\xC6"
-     "\x00\x80"
-     "\x21\xC6\x00\x80\x25\xC6\x00\x80\x29\xC6\x00\x80\x2D\xC6\x00\x80\x31\xC6"
-     "\x00\x80"
-     "\x35\xC6\x00\x80\x39\xC6\x00\x80\x41\xC6\x00\x80\x45\xC6\x00\x80\xBE\x30"
-     "\x7D\x00"
-     "\x80\x69\x00\x00\x81\x69\x00\x00\x82\x79\x00\x00\x84\x00\x02\x00\x86\xF0"
-     "\x7C\x00"
-     "\x87\x24\x7D\x00\xEF\x08\x03\x00\x49\xC6\x00\x80\xE1\x90\x01\x00\x4D\xC6"
-     "\x00\x80"
-     "\xE3\x4C\x03\x00\x51\xC6\x00\x80\x55\xC6\x00\x80\x59\xC6\x00\x80\x5D\xC6"
-     "\x00\x80"
-     "\x61\xC6\x00\x80\xB3\xBD\x01\x00\x65\xC6\x00\x80\x69\xC6\x00\x80\x6D\xC6"
-     "\x00\x80"
-     "\x71\xC6\x00\x80\xB6\x09\x01\x00\xB5\x09\x01\x00\x75\xC6\x00\x80\xBB\x6D"
-     "\x01\x00"
-     "\xBA\x6D\x01\x00\x79\xC6\x00\x80\x7D\xC6\x00\x80\xBF\x5D\x01\x00\xBE\x5D"
-     "\x01\x00"
-     "\xBD\x61\x01\x00\xBC\x69\x01\x00\x81\xC6\x00\x80\x85\xC6\x00\x80\x89\xC6"
-     "\x00\x80"
-     "\x8D\xC6\x00\x80\x91\xC6\x00\x80\x95\xC6\x00\x80\xEF\xA8\x06\x00\xEF\xE0"
-     "\x06\x00"
-     "\xE1\x98\x07\x00\xE1\x34\x06\x00\xE3\xB4\x06\x00\xE3\x6C\x06\x00\x99\xC6"
-     "\x00\x80"
-     "\x9D\xC6\x00\x80\xA1\xC6\x00\x80\xA5\xC6\x00\x80\xB8\x09\x7F\x00\xB9\x09"
-     "\x7F\x00"
-     "\xBA\x19\x7F\x00\xBB\x19\x7F\x00\xBC\x09\x7F\x00\xBD\x09\x7F\x00\xBE\x39"
-     "\x7F\x00"
-     "\xBF\x39\x7F\x00\xB0\xE9\x7E\x00\xB1\xE9\x7E\x00\xB2\xFD\x7E\x00\xB3\xF5"
-     "\x7E\x00"
-     "\xB4\xD9\x7E\x00\xB5\xD9\x7E\x00\xB6\x39\x7F\x00\xB7\x39\x7F\x00\xA8\x71"
-     "\x7E\x00"
-     "\xA9\x7D\x7E\x00\xAA\x75\x7E\x00\xAB\xC9\x7E\x00\xAC\xC1\x7E\x00\xAD\xC9"
-     "\x7E\x00"
-     "\xAE\xF5\x7E\x00\xAF\x9D\x7E\x00\xA9\xC6\x00\x80\x3D\xC6\x00\x80\xCD\xC5"
-     "\x00\x80"
-     "\x84\x64\x02\x00\x80\x89\x7F\x00\x81\x89\x7F\x00\x82\x19\x00\x00\xBE\x70"
-     "\x03\x00"
-     "\xA3\xBD\x7E\x00\xAD\xC6\x00\x80\x86\x28\x00\x00\x87\x4C\x03\x00\xB1\xC6"
-     "\x00\x80"
-     "\xA6\x09\x7E\x00\xA5\x09\x7E\x00\xB5\xC6\x00\x80\xAB\x6D\x7E\x00\xAA\x6D"
-     "\x7E\x00"
-     "\xB9\xC6\x00\x80\xBD\xC6\x00\x80\xAF\x5D\x7E\x00\xAE\x5D\x7E\x00\xAD\x61"
-     "\x7E\x00"
-     "\xAC\x69\x7E\x00\xC1\xC6\x00\x80\xB3\x99\x7F\x00\xC5\xC6\x00\x80\xC9\xC6"
-     "\x00\x80"
-     "\xB6\xA1\x7F\x00\xCD\xC6\x00\x80\xD1\xC6\x00\x80\xB5\xB1\x7F\x00\xBA\x9D"
-     "\x7F\x00"
-     "\xBB\x61\x7F\x00\xD5\xC6\x00\x80\xD9\xC6\x00\x80\xBE\x61\x7F\x00\xBF\x69"
-     "\x7F\x00"
-     "\xBC\x71\x7F\x00\xBD\x71\x7F\x00\xA8\x35\x7E\x00\xA9\x3D\x7E\x00\xAA\x55"
-     "\x7E\x00"
-     "\xAB\x69\x7E\x00\xAC\x7D\x7E\x00\xAD\x65\x7E\x00\xAE\x6D\x7E\x00\xAF\x65"
-     "\x7E\x00"
-     "\xDD\xC6\x00\x80\xE1\xC6\x00\x80\xE5\xC6\x00\x80\xE9\xC6\x00\x80\xED\xC6"
-     "\x00\x80"
-     "\xF1\xC6\x00\x80\xF5\xC6\x00\x80\xF9\xC6\x00\x80\xB8\x81\x7E\x00\xB9\x81"
-     "\x7E\x00"
-     "\xBA\x81\x7E\x00\xBB\x81\x7E\x00\xBC\x81\x7E\x00\xBD\x81\x7E\x00\xBE\xA9"
-     "\x7E\x00"
-     "\xBF\xA9\x7E\x00\xB0\x1D\x7E\x00\xB1\xE1\x7E\x00\xB2\xE1\x7E\x00\xB3\xE1"
-     "\x7E\x00"
-     "\xB4\xE1\x7E\x00\xB5\xE9\x7E\x00\xB6\xD1\x7E\x00\xB7\xD1\x7E\x00\xA3\xDD"
-     "\x7E\x00"
-     "\xFD\xC6\x00\x80\x01\xC7\x00\x80\x05\xC7\x00\x80\x09\xC7\x00\x80\xA6\xE5"
-     "\x7E\x00"
-     "\xA5\xF5\x7E\x00\x0D\xC7\x00\x80\xAB\x25\x7E\x00\xAA\xD9\x7E\x00\x11\xC7"
-     "\x00\x80"
-     "\x15\xC7\x00\x80\xAF\x2D\x7E\x00\xAE\x25\x7E\x00\xAD\x35\x7E\x00\xAC\x35"
-     "\x7E\x00"
-     "\x80\x69\x00\x00\x81\x69\x00\x00\x82\x05\x00\x00\x19\xC7\x00\x80\xBE\x98"
-     "\x01\x00"
-     "\x84\x98\x01\x00\xEF\xC0\x18\x00\x85\xD8\x07\x00\xE1\x28\x02\x00\x21\xC7"
-     "\x00\x80"
-     "\xE3\x34\x04\x00\x25\xC7\x00\x80\x86\x50\x06\x00\x87\xE4\x00\x00\x29\xC7"
-     "\x00\x80"
-     "\x2D\xC7\x00\x80\xA8\x91\x02\x00\xA9\x99\x02\x00\xAA\xA1\x02\x00\xAB\xA1"
-     "\x02\x00"
-     "\xAC\xD1\x02\x00\xAD\xDD\x02\x00\xAE\xD5\x02\x00\xAF\xC9\x02\x00\x31\xC7"
-     "\x00\x80"
-     "\x35\xC7\x00\x80\x39\xC7\x00\x80\x3D\xC7\x00\x80\x41\xC7\x00\x80\x45\xC7"
-     "\x00\x80"
-     "\x49\xC7\x00\x80\x4D\xC7\x00\x80\xB8\x75\x03\x00\xB9\x7D\x03\x00\xBA\x75"
-     "\x03\x00"
-     "\xBB\xC9\x03\x00\xBC\xD9\x03\x00\xBD\xD9\x03\x00\xBE\xC9\x03\x00\xBF\xC1"
-     "\x03\x00"
-     "\xB0\xB9\x02\x00\xB1\x8D\x02\x00\xB2\x85\x02\x00\xB3\x4D\x03\x00\xB4\x55"
-     "\x03\x00"
-     "\xB5\x5D\x03\x00\xB6\x55\x03\x00\xB7\x4D\x03\x00\xB3\x29\x02\x00\x84\xC8"
-     "\x05\x00"
-     "\x51\xC7\x00\x80\x55\xC7\x00\x80\x84\xDC\x06\x00\xB6\x5D\x02\x00\xB5\x59"
-     "\x02\x00"
-     "\x59\xC7\x00\x80\xBB\x49\x02\x00\xBA\x7D\x02\x00\x5D\xC7\x00\x80\x61\xC7"
-     "\x00\x80"
-     "\xBF\x99\x03\x00\xBE\x91\x03\x00\xBD\x99\x03\x00\xBC\x51\x02\x00\x65\xC7"
-     "\x00\x80"
-     "\xA3\x6D\x02\x00\x69\xC7\x00\x80\x6D\xC7\x00\x80\xA6\x19\x02\x00\x71\xC7"
-     "\x00\x80"
-     "\x75\xC7\x00\x80\xA5\x1D\x02\x00\xAA\x39\x02\x00\xAB\x0D\x02\x00\x79\xC7"
-     "\x00\x80"
-     "\x7D\xC7\x00\x80\xAE\xD5\x03\x00\xAF\xDD\x03\x00\xAC\x15\x02\x00\xAD\xDD"
-     "\x03\x00"
-     "\xE1\x90\x01\x00\x81\xC7\x00\x80\xE3\xCC\x00\x00\xBE\x44\x04\x00\x80\x15"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x3D\x00\x00\x85\xC7\x00\x80\x8F\x05\x6B\x00\x89\xC7"
-     "\x00\x80"
-     "\xBE\x60\x07\x00\x91\xC7\x00\x80\x95\xC7\x00\x80\x99\xC7\x00\x80\xEF\x10"
-     "\x00\x00"
-     "\x9D\xC7\x00\x80\x97\x41\x77\x00\x96\x19\x73\x00\x95\x49\x73\x00\x94\x71"
-     "\x73\x00"
-     "\x93\x65\x6F\x00\x92\x71\x6F\x00\x91\x71\x6F\x00\x90\x0D\x6B\x00\x9F\x01"
-     "\x7F\x00"
-     "\x9E\x01\x7F\x00\x9D\xED\x7A\x00\x9C\x6D\x7B\x00\x9B\x61\x7B\x00\x9A\x21"
-     "\x76\x00"
-     "\x99\x6D\x77\x00\x98\x71\x77\x00\x86\x20\x07\x00\x87\x44\x02\x00\xA1\xC7"
-     "\x00\x80"
-     "\xA5\xC7\x00\x80\xA3\x15\x03\x00\xA2\xF1\x03\x00\xA1\xF1\x03\x00\xA0\x79"
-     "\x7F\x00"
-     "\x84\x00\x08\x00\xA9\xC7\x00\x80\xAD\xC7\x00\x80\xB1\xC7\x00\x80\xB5\xC7"
-     "\x00\x80"
-     "\xB9\xC7\x00\x80\xBD\xC7\x00\x80\xC1\xC7\x00\x80\x80\x95\x07\x00\x81\x9D"
-     "\x07\x00"
-     "\x82\x95\x07\x00\xC5\xC7\x00\x80\xC9\xC7\x00\x80\xEF\x10\x7E\x00\xEF\x5C"
-     "\x01\x00"
-     "\xCD\xC7\x00\x80\xE1\xA8\x7E\x00\xE3\xB8\x7E\x00\xE3\x94\x01\x00\xE1\x78"
-     "\x7F\x00"
-     "\xD1\xC7\x00\x80\xD5\xC7\x00\x80\xD9\xC7\x00\x80\xDD\xC7\x00\x80\x1D\xC7"
-     "\x00\x80"
-     "\x8D\xC7\x00\x80\x8A\x09\x74\x00\x8B\x51\x74\x00\x8C\x11\x70\x00\x8D\x19"
-     "\x70\x00"
-     "\x8E\x0D\x70\x00\x8F\x49\x71\x00\xE1\xC7\x00\x80\xE5\xC7\x00\x80\x82\x0D"
-     "\x6C\x00"
-     "\x83\x7D\x6C\x00\x84\x19\x68\x00\x85\x11\x68\x00\x86\x0D\x68\x00\x87\x01"
-     "\x74\x00"
-     "\xE9\xC7\x00\x80\xED\xC7\x00\x80\x9A\x21\x04\x00\x9B\x29\x04\x00\x9C\xD5"
-     "\x04\x00"
-     "\xF1\xC7\x00\x80\xF5\xC7\x00\x80\xF9\xC7\x00\x80\xFD\xC7\x00\x80\x01\xC8"
-     "\x00\x80"
-     "\x92\x79\x7C\x00\x93\x6D\x7C\x00\x94\xF5\x7D\x00\x95\x11\x78\x00\x96\x19"
-     "\x78\x00"
-     "\x97\x99\x78\x00\xA1\x81\x00\x00\x05\xC8\x00\x80\xA3\xE9\x00\x00\xA2\x89"
-     "\x00\x00"
-     "\xA5\x99\x0C\x00\xA4\x91\x0C\x00\xA7\x01\x08\x00\xA6\x31\x0C\x00\xA9\x7D"
-     "\x08\x00"
-     "\xA8\x85\x08\x00\xAB\x89\x14\x00\xAA\x01\x14\x00\xAD\x01\x10\x00\xAC\x25"
-     "\x14\x00"
-     "\xAF\x8D\x10\x00\xAE\x9D\x10\x00\xB1\xC1\x1C\x00\xB0\x01\x1C\x00\xB3\xA5"
-     "\x1D\x00"
-     "\xB2\x59\x1C\x00\xB5\x55\x18\x00\xB4\x55\x18\x00\x86\x78\x01\x00\x87\x14"
-     "\x03\x00"
-     "\x09\xC8\x00\x80\x0D\xC8\x00\x80\x11\xC8\x00\x80\x15\xC8\x00\x80\x19\xC8"
-     "\x00\x80"
-     "\x1D\xC8\x00\x80\x21\xC8\x00\x80\x25\xC8\x00\x80\xA8\xE5\x04\x00\xA9\xE9"
-     "\x04\x00"
-     "\xAA\xF1\x04\x00\xAB\xF1\x04\x00\xAC\x31\x07\x00\xAD\x31\x07\x00\xAE\x31"
-     "\x07\x00"
-     "\xAF\x31\x07\x00\x29\xC8\x00\x80\x2D\xC8\x00\x80\x31\xC8\x00\x80\x35\xC8"
-     "\x00\x80"
-     "\x39\xC8\x00\x80\x3D\xC8\x00\x80\x41\xC8\x00\x80\x45\xC8\x00\x80\xB8\xED"
-     "\x07\x00"
-     "\xB9\x85\x07\x00\xBA\x81\x07\x00\xBB\x81\x07\x00\xBC\x85\x07\x00\xBD\x8D"
-     "\x07\x00"
-     "\xBE\xB1\x07\x00\xBF\xB1\x07\x00\xB0\x51\x07\x00\xB1\x51\x07\x00\xB2\x51"
-     "\x07\x00"
-     "\xB3\x51\x07\x00\xB4\xF9\x07\x00\xB5\xF9\x07\x00\xB6\xE5\x07\x00\xB7\xD5"
-     "\x07\x00"
-     "\xB3\x79\x04\x00\x49\xC8\x00\x80\x4D\xC8\x00\x80\x51\xC8\x00\x80\x55\xC8"
-     "\x00\x80"
-     "\xB6\xB1\x07\x00\xB5\xB9\x07\x00\x59\xC8\x00\x80\xBB\x95\x07\x00\xBA\x95"
-     "\x07\x00"
-     "\x5D\xC8\x00\x80\x61\xC8\x00\x80\xBF\x3D\x07\x00\xBE\x3D\x07\x00\xBD\x85"
-     "\x07\x00"
-     "\xBC\x85\x07\x00\x82\x19\x00\x00\xA3\x3D\x04\x00\x80\x65\x00\x00\x81\x19"
-     "\x00\x00"
-     "\xA6\xF5\x07\x00\x65\xC8\x00\x80\x69\xC8\x00\x80\xA5\xFD\x07\x00\xAA\xD1"
-     "\x07\x00"
-     "\xAB\xD1\x07\x00\xBE\x64\x0F\x00\x71\xC8\x00\x80\xAE\x79\x07\x00\xAF\x79"
-     "\x07\x00"
-     "\xAC\xC1\x07\x00\xAD\xC1\x07\x00\xA8\x2D\x05\x00\xA9\x35\x05\x00\xAA\x3D"
-     "\x05\x00"
-     "\xAB\x31\x05\x00\xAC\x99\x05\x00\xAD\x99\x05\x00\xAE\xBD\x05\x00\xAF\xB5"
-     "\x05\x00"
-     "\x86\x00\x0C\x00\x87\xB0\x03\x00\x75\xC8\x00\x80\x79\xC8\x00\x80\x7D\xC8"
-     "\x00\x80"
-     "\x81\xC8\x00\x80\x85\xC8\x00\x80\x89\xC8\x00\x80\xB8\xB1\x05\x00\xB9\xB1"
-     "\x05\x00"
-     "\xBA\xB1\x05\x00\xBB\xB1\x05\x00\xBC\x71\x02\x00\xBD\x71\x02\x00\xBE\x19"
-     "\x02\x00"
-     "\xBF\x19\x02\x00\xB0\xCD\x05\x00\xB1\xD5\x05\x00\xB2\xDD\x05\x00\xB3\xD5"
-     "\x05\x00"
-     "\xB4\xCD\x05\x00\xB5\xB9\x05\x00\xB6\xA9\x05\x00\xB7\xA1\x05\x00\xB3\xB5"
-     "\x05\x00"
-     "\x8D\xC8\x00\x80\x91\xC8\x00\x80\x95\xC8\x00\x80\x99\xC8\x00\x80\xB6\x19"
-     "\x05\x00"
-     "\xB5\x11\x05\x00\x9D\xC8\x00\x80\xBB\x11\x05\x00\xBA\x09\x05\x00\xA1\xC8"
-     "\x00\x80"
-     "\xA5\xC8\x00\x80\xBF\x7D\x05\x00\xBE\x09\x05\x00\xBD\x01\x05\x00\xBC\x09"
-     "\x05\x00"
-     "\x84\x48\x0C\x00\xA3\xF1\x05\x00\xA9\xC8\x00\x80\xAD\xC8\x00\x80\xA6\x5D"
-     "\x05\x00"
-     "\xB1\xC8\x00\x80\xB5\xC8\x00\x80\xA5\x55\x05\x00\xAA\x4D\x05\x00\xAB\x55"
-     "\x05\x00"
-     "\xB9\xC8\x00\x80\xBD\xC8\x00\x80\xAE\x4D\x05\x00\xAF\x39\x05\x00\xAC\x4D"
-     "\x05\x00"
-     "\xAD\x45\x05\x00\xA8\x65\x03\x00\xA9\x81\x03\x00\xAA\xE9\x03\x00\xAB\xE9"
-     "\x03\x00"
-     "\xAC\xF9\x03\x00\xAD\xF9\x03\x00\xAE\xE9\x03\x00\xAF\xE9\x03\x00\xC1\xC8"
-     "\x00\x80"
-     "\xC5\xC8\x00\x80\xC9\xC8\x00\x80\xCD\xC8\x00\x80\xD1\xC8\x00\x80\xD5\xC8"
-     "\x00\x80"
-     "\xD9\xC8\x00\x80\xBE\xB0\x0C\x00\xB8\x79\x02\x00\xB9\x79\x02\x00\xBA\xCD"
-     "\x02\x00"
-     "\xBB\xC5\x02\x00\xBC\xDD\x02\x00\xBD\xC5\x02\x00\xBE\xC5\x02\x00\xBF\xF5"
-     "\x02\x00"
-     "\xB0\x99\x03\x00\xB1\x99\x03\x00\xB2\x49\x02\x00\xB3\x49\x02\x00\xB4\x59"
-     "\x02\x00"
-     "\xB5\x59\x02\x00\xB6\x49\x02\x00\xB7\x49\x02\x00\xE1\x18\x06\x00\xE1\x6C"
-     "\x04\x00"
-     "\xE3\x08\x07\x00\xE3\x18\x07\x00\x81\xA5\x00\x00\x80\xA5\x00\x00\xE1\xC8"
-     "\x00\x80"
-     "\x82\xA5\x00\x00\x86\x80\x0C\x00\x87\x04\x0D\x00\xE5\xC8\x00\x80\xE9\xC8"
-     "\x00\x80"
-     "\xED\xC8\x00\x80\xF1\xC8\x00\x80\xEF\x68\x06\x00\xEF\x48\x06\x00\xF5\xC8"
-     "\x00\x80"
-     "\xF9\xC8\x00\x80\x85\xF0\x1B\x00\xFD\xC8\x00\x80\x01\xC9\x00\x80\x05\xC9"
-     "\x00\x80"
-     "\x09\xC9\x00\x80\xEF\xCC\x00\x00\x0D\xC9\x00\x80\xE1\x24\x01\x00\x11\xC9"
-     "\x00\x80"
-     "\xE3\x5C\x06\x00\x15\xC9\x00\x80\x19\xC9\x00\x80\x1D\xC9\x00\x80\x21\xC9"
-     "\x00\x80"
-     "\xA3\xE1\x0C\x00\x6D\xC8\x00\x80\xDD\xC8\x00\x80\x25\xC9\x00\x80\x29\xC9"
-     "\x00\x80"
-     "\xA6\xF9\x0C\x00\xA5\xF1\x0C\x00\x2D\xC9\x00\x80\xAB\xDD\x0C\x00\xAA\xDD"
-     "\x0C\x00"
-     "\x31\xC9\x00\x80\x35\xC9\x00\x80\xAF\x1D\x0D\x00\xAE\x1D\x0D\x00\xAD\x1D"
-     "\x0D\x00"
-     "\xAC\x1D\x0D\x00\x39\xC9\x00\x80\xB3\x69\x0E\x00\x3D\xC9\x00\x80\x41\xC9"
-     "\x00\x80"
-     "\xB6\x69\x0E\x00\x45\xC9\x00\x80\x49\xC9\x00\x80\xB5\x69\x0E\x00\xBA\xE9"
-     "\x0E\x00"
-     "\xBB\xE9\x0E\x00\x84\xA4\x00\x00\xBE\xA4\x00\x00\xBE\xD9\x0E\x00\xBF\xD9"
-     "\x0E\x00"
-     "\xBC\xE9\x0E\x00\xBD\xE1\x0E\x00\xA8\x6D\x0D\x00\xA9\xAD\x0E\x00\xAA\xA5"
-     "\x0E\x00"
-     "\xAB\xBD\x0E\x00\xAC\xA5\x0E\x00\xAD\xAD\x0E\x00\xAE\xA5\x0E\x00\xAF\x19"
-     "\x0E\x00"
-     "\x80\xED\x00\x00\x81\xF1\x00\x00\x82\xF1\x00\x00\x4D\xC9\x00\x80\x86\x90"
-     "\x00\x00"
-     "\x87\x14\x00\x00\x51\xC9\x00\x80\x55\xC9\x00\x80\xB8\x15\x0E\x00\xB9\x1D"
-     "\x0E\x00"
-     "\xBA\x21\x0E\x00\xBB\x21\x0E\x00\xBC\xF1\x0F\x00\xBD\xF1\x0F\x00\xBE\xF1"
-     "\x0F\x00"
-     "\xBF\xF1\x0F\x00\xB0\x69\x0E\x00\xB1\x69\x0E\x00\xB2\x79\x0E\x00\xB3\x71"
-     "\x0E\x00"
-     "\xB4\x59\x0E\x00\xB5\x59\x0E\x00\xB6\x31\x0E\x00\xB7\x31\x0E\x00\xA3\x2D"
-     "\x0F\x00"
-     "\x59\xC9\x00\x80\x5D\xC9\x00\x80\x61\xC9\x00\x80\x65\xC9\x00\x80\xA6\x2D"
-     "\x0F\x00"
-     "\xA5\x2D\x0F\x00\x69\xC9\x00\x80\xAB\xAD\x0F\x00\xAA\xAD\x0F\x00\x6D\xC9"
-     "\x00\x80"
-     "\x71\xC9\x00\x80\xAF\x9D\x0F\x00\xAE\x9D\x0F\x00\xAD\xA5\x0F\x00\xAC\xAD"
-     "\x0F\x00"
-     "\x75\xC9\x00\x80\x79\xC9\x00\x80\xBA\xB1\x0D\x00\xBB\xB1\x0D\x00\xBC\x71"
-     "\x0C\x00"
-     "\xBD\x71\x0C\x00\xBE\x71\x0C\x00\xBF\x71\x0C\x00\x7D\xC9\x00\x80\x81\xC9"
-     "\x00\x80"
-     "\x85\xC9\x00\x80\xB3\x8D\x0D\x00\x89\xC9\x00\x80\xB5\x9D\x0D\x00\xB6\x95"
-     "\x0D\x00"
-     "\x8D\xC9\x00\x80\xA8\x19\x0E\x00\xA9\x2D\x0E\x00\xAA\x25\x0E\x00\xAB\x3D"
-     "\x0E\x00"
-     "\xAC\x25\x0E\x00\xAD\x55\x0E\x00\xAE\x5D\x0E\x00\xAF\x51\x0E\x00\x91\xC9"
-     "\x00\x80"
-     "\x95\xC9\x00\x80\x99\xC9\x00\x80\x9D\xC9\x00\x80\xA1\xC9\x00\x80\xA5\xC9"
-     "\x00\x80"
-     "\xA9\xC9\x00\x80\xAD\xC9\x00\x80\xB8\x3D\x0E\x00\xB9\xC9\x01\x00\xBA\xD9"
-     "\x01\x00"
-     "\xBB\xD1\x01\x00\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\x99\x01\x00\xBF\x95"
-     "\x01\x00"
-     "\xB0\x31\x0E\x00\xB1\x31\x0E\x00\xB2\x31\x0E\x00\xB3\x0D\x0E\x00\xB4\x19"
-     "\x0E\x00"
-     "\xB5\x05\x0E\x00\xB6\x0D\x0E\x00\xB7\x05\x0E\x00\xB3\xAD\x0D\x00\xB1\xC9"
-     "\x00\x80"
-     "\xB5\xC9\x00\x80\xB9\xC9\x00\x80\x84\x9C\x03\x00\xB6\xD5\x0D\x00\xB5\xAD"
-     "\x0D\x00"
-     "\xBE\x50\x03\x00\xBB\xF5\x0D\x00\xBA\xF5\x0D\x00\x86\x88\x04\x00\x87\xBC"
-     "\x03\x00"
-     "\xBF\xC1\x0D\x00\xBE\xDD\x0D\x00\xBD\xD9\x0D\x00\xBC\xE5\x0D\x00\xC1\xC9"
-     "\x00\x80"
-     "\xA3\xE9\x0D\x00\xC5\xC9\x00\x80\xC9\xC9\x00\x80\xA6\x91\x0D\x00\xCD\xC9"
-     "\x00\x80"
-     "\xD1\xC9\x00\x80\xA5\xE9\x0D\x00\xAA\xB1\x0D\x00\xAB\xB1\x0D\x00\xD5\xC9"
-     "\x00\x80"
-     "\xD9\xC9\x00\x80\xAE\x99\x0D\x00\xAF\x85\x0D\x00\xAC\xA1\x0D\x00\xAD\x9D"
-     "\x0D\x00"
-     "\xA8\x39\x02\x00\xA9\x39\x02\x00\xAA\x8D\x03\x00\xAB\x85\x03\x00\xAC\x9D"
-     "\x03\x00"
-     "\xAD\x85\x03\x00\xAE\x85\x03\x00\xAF\xB5\x03\x00\x80\xA5\x00\x00\x81\xAD"
-     "\x00\x00"
-     "\x82\xA5\x00\x00\xDD\xC9\x00\x80\xE1\xC9\x00\x80\xE5\xC9\x00\x80\xE9\xC9"
-     "\x00\x80"
-     "\xED\xC9\x00\x80\xB8\x61\x03\x00\xB9\x61\x03\x00\xBA\x61\x03\x00\xBB\x61"
-     "\x03\x00"
-     "\xBC\x61\x03\x00\xBD\x61\x03\x00\xBE\x61\x03\x00\xBF\x61\x03\x00\xB0\xCD"
-     "\x03\x00"
-     "\xB1\xA5\x03\x00\xB2\xAD\x03\x00\xB3\x8D\x03\x00\xB4\x95\x03\x00\xB5\x9D"
-     "\x03\x00"
-     "\xB6\x89\x03\x00\xB7\x81\x03\x00\xF1\xC9\x00\x80\xF5\xC9\x00\x80\xF9\xC9"
-     "\x00\x80"
-     "\xFD\xC9\x00\x80\xEF\xF4\x03\x00\x01\xCA\x00\x80\x05\xCA\x00\x80\x09\xCA"
-     "\x00\x80"
-     "\xE3\x20\x03\x00\x0D\xCA\x00\x80\xE1\x80\x01\x00\x11\xCA\x00\x80\x15\xCA"
-     "\x00\x80"
-     "\x19\xCA\x00\x80\x1D\xCA\x00\x80\x21\xCA\x00\x80\xB3\x45\x02\x00\x25\xCA"
-     "\x00\x80"
-     "\x29\xCA\x00\x80\x84\x04\x05\x00\x31\xCA\x00\x80\xB6\x45\x02\x00\xB5\x55"
-     "\x02\x00"
-     "\x35\xCA\x00\x80\xBB\x21\x02\x00\xBA\x49\x02\x00\x39\xCA\x00\x80\xBE\x20"
-     "\x05\x00"
-     "\xBF\x19\x02\x00\xBE\x11\x02\x00\xBD\x19\x02\x00\xBC\x39\x02\x00\xE1\xB8"
-     "\x0F\x00"
-     "\xE3\xFC\x01\x00\xE3\x58\x0F\x00\xE1\x34\x0E\x00\x80\xAD\x00\x00\x81\xD1"
-     "\x00\x00"
-     "\x82\xD1\x00\x00\x3D\xCA\x00\x80\xBE\xE8\x04\x00\x41\xCA\x00\x80\x45\xCA"
-     "\x00\x80"
-     "\x49\xCA\x00\x80\x4D\xCA\x00\x80\xEF\x20\x01\x00\xEF\x7C\x0E\x00\x51\xCA"
-     "\x00\x80"
-     "\xA3\xC5\x02\x00\x55\xCA\x00\x80\x86\x68\x04\x00\x87\x00\x05\x00\x59\xCA"
-     "\x00\x80"
-     "\xA6\xC5\x02\x00\xA5\xD5\x02\x00\x5D\xCA\x00\x80\xAB\xA1\x02\x00\xAA\xC9"
-     "\x02\x00"
-     "\x61\xCA\x00\x80\x65\xCA\x00\x80\xAF\x99\x02\x00\xAE\x91\x02\x00\xAD\x99"
-     "\x02\x00"
-     "\xAC\xB9\x02\x00\xA8\x41\x06\x00\xA9\x4D\x06\x00\xAA\x45\x06\x00\xAB\x59"
-     "\x06\x00"
-     "\xAC\x49\x06\x00\xAD\x7D\x06\x00\xAE\x75\x06\x00\xAF\xB1\x06\x00\x2D\xCA"
-     "\x00\x80"
-     "\x69\xCA\x00\x80\x6D\xCA\x00\x80\x71\xCA\x00\x80\x75\xCA\x00\x80\x79\xCA"
-     "\x00\x80"
-     "\x7D\xCA\x00\x80\x81\xCA\x00\x80\xB8\x55\x07\x00\xB9\x5D\x07\x00\xBA\x55"
-     "\x07\x00"
-     "\xBB\x69\x07\x00\xBC\x7D\x07\x00\xBD\x65\x07\x00\xBE\x6D\x07\x00\xBF\x65"
-     "\x07\x00"
-     "\xB0\xD1\x06\x00\xB1\xD1\x06\x00\xB2\xD1\x06\x00\xB3\xD1\x06\x00\xB4\x71"
-     "\x07\x00"
-     "\xB5\x71\x07\x00\xB6\x71\x07\x00\xB7\x71\x07\x00\x85\xCA\x00\x80\xB3\xC9"
-     "\x06\x00"
-     "\x89\xCA\x00\x80\x8D\xCA\x00\x80\xB6\xF9\x06\x00\x91\xCA\x00\x80\xBD\xC9"
-     "\x00\x80"
-     "\xB5\xC5\x06\x00\xBA\x11\x06\x00\xBB\x11\x06\x00\x95\xCA\x00\x80\x99\xCA"
-     "\x00\x80"
-     "\xBE\xB1\x07\x00\xBF\xB1\x07\x00\xBC\x01\x06\x00\xBD\x01\x06\x00\xA3\x8D"
-     "\x06\x00"
-     "\x82\x69\x00\x00\x81\x61\x00\x00\x80\x59\x00\x00\x84\x9C\x03\x00\xA6\xBD"
-     "\x06\x00"
-     "\xA5\x81\x06\x00\x9D\xCA\x00\x80\xAB\x55\x06\x00\xAA\x55\x06\x00\x86\x48"
-     "\x00\x00"
-     "\x87\x68\x00\x00\xAF\xF5\x07\x00\xAE\xF5\x07\x00\xAD\x45\x06\x00\xAC\x45"
-     "\x06\x00"
-     "\xA8\x51\x06\x00\xA9\x59\x06\x00\xAA\x61\x06\x00\xAB\x91\x06\x00\xAC\xB5"
-     "\x06\x00"
-     "\xAD\xBD\x06\x00\xAE\xB5\x06\x00\xAF\xAD\x06\x00\xA1\xCA\x00\x80\xA5\xCA"
-     "\x00\x80"
-     "\xA9\xCA\x00\x80\xAD\xCA\x00\x80\xB1\xCA\x00\x80\xB5\xCA\x00\x80\xB9\xCA"
-     "\x00\x80"
-     "\xBD\xCA\x00\x80\xB8\x6D\x01\x00\xB9\x75\x01\x00\xBA\x7D\x01\x00\xBB\x75"
-     "\x01\x00"
-     "\xBC\x6D\x01\x00\xBD\xD5\x01\x00\xBE\xD1\x01\x00\xBF\xD1\x01\x00\xB0\xDD"
-     "\x06\x00"
-     "\xB1\xA1\x06\x00\xB2\xB9\x06\x00\xB3\xB1\x06\x00\xB4\x99\x06\x00\xB5\x99"
-     "\x06\x00"
-     "\xB6\x5D\x01\x00\xB7\x55\x01\x00\xB3\xE9\x06\x00\xC1\xCA\x00\x80\xC5\xCA"
-     "\x00\x80"
-     "\xC9\xCA\x00\x80\xCD\xCA\x00\x80\xB6\x3D\x06\x00\xB5\x3D\x06\x00\xD1\xCA"
-     "\x00\x80"
-     "\xBB\x65\x06\x00\xBA\x19\x06\x00\xD5\xCA\x00\x80\xD9\xCA\x00\x80\xBF\x5D"
-     "\x06\x00"
-     "\xBE\x5D\x06\x00\xBD\x65\x06\x00\xBC\x6D\x06\x00\xDD\xCA\x00\x80\xA3\xAD"
-     "\x06\x00"
-     "\xE1\xCA\x00\x80\xE5\xCA\x00\x80\xA6\x79\x06\x00\xE9\xCA\x00\x80\xED\xCA"
-     "\x00\x80"
-     "\xA5\x79\x06\x00\xAA\x5D\x06\x00\xAB\x21\x06\x00\xF1\xCA\x00\x80\xF5\xCA"
-     "\x00\x80"
-     "\xAE\x19\x06\x00\xAF\x19\x06\x00\xAC\x29\x06\x00\xAD\x21\x06\x00\xA8\x21"
-     "\x03\x00"
-     "\xA9\x21\x03\x00\xAA\x21\x03\x00\xAB\x21\x03\x00\xAC\x21\x03\x00\xAD\x21"
-     "\x03\x00"
-     "\xAE\x21\x03\x00\xAF\x21\x03\x00\xF9\xCA\x00\x80\xFD\xCA\x00\x80\x01\xCB"
-     "\x00\x80"
-     "\x05\xCB\x00\x80\x09\xCB\x00\x80\xBE\x78\x02\x00\x11\xCB\x00\x80\x15\xCB"
-     "\x00\x80"
-     "\xB8\x89\x03\x00\xB9\x89\x03\x00\xBA\x9D\x03\x00\xBB\x95\x03\x00\xBC\xB9"
-     "\x03\x00"
-     "\xBD\xB9\x03\x00\xBE\x79\x00\x00\xBF\x79\x00\x00\xB0\xE5\x03\x00\xB1\xED"
-     "\x03\x00"
-     "\xB2\xE5\x03\x00\xB3\xFD\x03\x00\xB4\xED\x03\x00\xB5\xDD\x03\x00\xB6\xD5"
-     "\x03\x00"
-     "\xB7\xBD\x03\x00\x81\xAD\x00\x00\x80\xAD\x00\x00\xB3\xA1\x03\x00\x82\xBD"
-     "\x00\x00"
-     "\xB5\xA1\x03\x00\x19\xCB\x00\x80\x1D\xCB\x00\x80\xB6\xA9\x03\x00\x84\xE4"
-     "\x02\x00"
-     "\x21\xCB\x00\x80\xBB\x21\x03\x00\xBA\x29\x03\x00\xBD\x21\x03\x00\xBC\x39"
-     "\x03\x00"
-     "\xBF\x11\x03\x00\xBE\x21\x03\x00\xA3\xE9\x03\x00\x25\xCB\x00\x80\x86\x68"
-     "\x1C\x00"
-     "\x87\x80\x03\x00\x29\xCB\x00\x80\xA6\xE1\x03\x00\xA5\xE9\x03\x00\x2D\xCB"
-     "\x00\x80"
-     "\xAB\x69\x03\x00\xAA\x61\x03\x00\x31\xCB\x00\x80\x35\xCB\x00\x80\xAF\x59"
-     "\x03\x00"
-     "\xAE\x69\x03\x00\xAD\x69\x03\x00\xAC\x71\x03\x00\x39\xCB\x00\x80\x3D\xCB"
-     "\x00\x80"
-     "\x41\xCB\x00\x80\x45\xCB\x00\x80\x49\xCB\x00\x80\xE3\x48\x06\x00\x4D\xCB"
-     "\x00\x80"
-     "\xE1\x2C\x07\x00\x51\xCB\x00\x80\xEF\xB0\x06\x00\x55\xCB\x00\x80\x59\xCB"
-     "\x00\x80"
-     "\x5D\xCB\x00\x80\x61\xCB\x00\x80\x65\xCB\x00\x80\x69\xCB\x00\x80\x6D\xCB"
-     "\x00\x80"
-     "\x82\x21\x00\x00\x81\x15\x00\x00\x80\x19\x00\x00\xE3\xE0\x03\x00\x71\xCB"
-     "\x00\x80"
-     "\xE1\xA8\x01\x00\x75\xCB\x00\x80\xEF\xAC\x03\x00\x79\xCB\x00\x80\x84\x64"
-     "\x02\x00"
-     "\xBE\x60\x1D\x00\x86\xD0\x1C\x00\x87\x44\x1D\x00\x81\xCB\x00\x80\x85\xCB"
-     "\x00\x80"
-     "\x89\xCB\x00\x80\x8D\xCB\x00\x80\x91\xCB\x00\x80\x95\xCB\x00\x80\x99\xCB"
-     "\x00\x80"
-     "\x9D\xCB\x00\x80\xA1\xCB\x00\x80\xEF\xC4\x06\x00\x84\xA8\x1D\x00\xE1\x74"
-     "\x06\x00"
-     "\xA5\xCB\x00\x80\xE3\x6C\x01\x00\xA9\xCB\x00\x80\xAD\xCB\x00\x80\xB1\xCB"
-     "\x00\x80"
-     "\xB5\xCB\x00\x80\xB3\x55\x01\x00\xB9\xCB\x00\x80\xBD\xCB\x00\x80\xC1\xCB"
-     "\x00\x80"
-     "\xC5\xCB\x00\x80\xB6\x11\x01\x00\xB5\x7D\x01\x00\xC9\xCB\x00\x80\xBB\x0D"
-     "\x01\x00"
-     "\xBA\x0D\x01\x00\xCD\xCB\x00\x80\xD1\xCB\x00\x80\xBF\xFD\x00\x00\xBE\xFD"
-     "\x00\x00"
-     "\xBD\xFD\x00\x00\xBC\xFD\x00\x00\xA8\x39\x1E\x00\xA9\x39\x1E\x00\xAA\x9D"
-     "\x1E\x00"
-     "\xAB\x95\x1E\x00\xAC\xB5\x1E\x00\xAD\xD1\x1E\x00\xAE\xC9\x1E\x00\xAF\xC1"
-     "\x1E\x00"
-     "\x7D\xCB\x00\x80\xD5\xCB\x00\x80\xD9\xCB\x00\x80\xDD\xCB\x00\x80\x80\x0D"
-     "\x00\x00"
-     "\x81\xB1\x00\x00\x82\xB1\x00\x00\xE1\xCB\x00\x80\xB8\x61\x1F\x00\xB9\x6D"
-     "\x1F\x00"
-     "\xBA\x65\x1F\x00\xBB\x7D\x1F\x00\xBC\x65\x1F\x00\xBD\x6D\x1F\x00\xBE\x65"
-     "\x1F\x00"
-     "\xBF\x1D\x1F\x00\xB0\x89\x1E\x00\xB1\x89\x1E\x00\xB2\x69\x1F\x00\xB3\x69"
-     "\x1F\x00"
-     "\xB4\x79\x1F\x00\xB5\x79\x1F\x00\xB6\x69\x1F\x00\xB7\x65\x1F\x00\xA3\x15"
-     "\x1E\x00"
-     "\xE5\xCB\x00\x80\xE9\xCB\x00\x80\xBE\x80\x03\x00\xED\xCB\x00\x80\xA6\x51"
-     "\x1E\x00"
-     "\xA5\x3D\x1E\x00\xF1\xCB\x00\x80\xAB\x4D\x1E\x00\xAA\x4D\x1E\x00\xF5\xCB"
-     "\x00\x80"
-     "\x0D\xCB\x00\x80\xAF\xBD\x1F\x00\xAE\xBD\x1F\x00\xAD\xBD\x1F\x00\xAC\xBD"
-     "\x1F\x00"
-     "\xF9\xCB\x00\x80\xB3\x45\x1F\x00\x86\x68\x00\x00\x87\x1C\x03\x00\xB6\x5D"
-     "\x1F\x00"
-     "\xFD\xCB\x00\x80\x01\xCC\x00\x80\xB5\x55\x1F\x00\xBA\x79\x1F\x00\xBB\x4D"
-     "\x1F\x00"
-     "\x05\xCC\x00\x80\x09\xCC\x00\x80\xBE\x49\x1F\x00\xBF\x49\x1F\x00\xBC\x49"
-     "\x1F\x00"
-     "\xBD\x49\x1F\x00\xA8\x79\x1E\x00\xA9\x79\x1E\x00\xAA\x89\x1E\x00\xAB\x89"
-     "\x1E\x00"
-     "\xAC\x99\x1E\x00\xAD\x99\x1E\x00\xAE\x89\x1E\x00\xAF\x89\x1E\x00\x0D\xCC"
-     "\x00\x80"
-     "\x11\xCC\x00\x80\x15\xCC\x00\x80\x19\xCC\x00\x80\x1D\xCC\x00\x80\x21\xCC"
-     "\x00\x80"
-     "\x25\xCC\x00\x80\x29\xCC\x00\x80\xB8\x8D\x1E\x00\xB9\x95\x1E\x00\xBA\x95"
-     "\x1E\x00"
-     "\xBB\xA5\x1E\x00\xBC\xBD\x1E\x00\xBD\x71\x01\x00\xBE\x71\x01\x00\xBF\x71"
-     "\x01\x00"
-     "\xB0\xF9\x1E\x00\xB1\xCD\x1E\x00\xB2\xD9\x1E\x00\xB3\xD9\x1E\x00\xB4\xC9"
-     "\x1E\x00"
-     "\xB5\xC9\x1E\x00\xB6\xBD\x1E\x00\xB7\xB5\x1E\x00\xA3\x01\x1E\x00\x2D\xCC"
-     "\x00\x80"
-     "\x31\xCC\x00\x80\x35\xCC\x00\x80\x39\xCC\x00\x80\xA6\x19\x1E\x00\xA5\x11"
-     "\x1E\x00"
-     "\x3D\xCC\x00\x80\xAB\x09\x1E\x00\xAA\x3D\x1E\x00\x41\xCC\x00\x80\x45\xCC"
-     "\x00\x80"
-     "\xAF\x0D\x1E\x00\xAE\x0D\x1E\x00\xAD\x0D\x1E\x00\xAC\x0D\x1E\x00\x49\xCC"
-     "\x00\x80"
-     "\x4D\xCC\x00\x80\x51\xCC\x00\x80\x55\xCC\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x59\xCC\x00\x80\x84\xAC\x01\x00\xBE\xAC\x01\x00\x87\xE8"
-     "\x00\x00"
-     "\x86\x4C\x07\x00\x61\xCC\x00\x80\x65\xCC\x00\x80\x69\xCC\x00\x80\x6D\xCC"
-     "\x00\x80"
-     "\xA8\x9D\x02\x00\xA9\xAD\x02\x00\xAA\xC9\x02\x00\xAB\xD9\x02\x00\xAC\xC9"
-     "\x02\x00"
-     "\xAD\xFD\x02\x00\xAE\xF5\x02\x00\xAF\x3D\x03\x00\x71\xCC\x00\x80\x75\xCC"
-     "\x00\x80"
-     "\x79\xCC\x00\x80\x7D\xCC\x00\x80\xBE\x7C\x03\x00\x81\xCC\x00\x80\x85\xCC"
-     "\x00\x80"
-     "\x89\xCC\x00\x80\xB8\xC9\x03\x00\xB9\xC9\x03\x00\xBA\xD9\x03\x00\xBB\xD1"
-     "\x03\x00"
-     "\xBC\xF9\x03\x00\xBD\xF9\x03\x00\xBE\x99\x03\x00\xBF\x99\x03\x00\xB0\x45"
-     "\x03\x00"
-     "\xB1\x4D\x03\x00\xB2\x45\x03\x00\xB3\x5D\x03\x00\xB4\x45\x03\x00\xB5\x4D"
-     "\x03\x00"
-     "\xB6\x45\x03\x00\xB7\xF9\x03\x00\xB3\x41\x02\x00\x8D\xCC\x00\x80\x91\xCC"
-     "\x00\x80"
-     "\x95\xCC\x00\x80\x99\xCC\x00\x80\xB6\x7D\x02\x00\xB5\x41\x02\x00\x9D\xCC"
-     "\x00\x80"
-     "\xBB\x89\x03\x00\xBA\x81\x03\x00\xA1\xCC\x00\x80\xA5\xCC\x00\x80\xBF\x89"
-     "\x03\x00"
-     "\xBE\x81\x03\x00\xBD\x89\x03\x00\xBC\x91\x03\x00\xA9\xCC\x00\x80\xA3\x05"
-     "\x02\x00"
-     "\xAD\xCC\x00\x80\xB1\xCC\x00\x80\xA6\x39\x02\x00\xB5\xCC\x00\x80\xB9\xCC"
-     "\x00\x80"
-     "\xA5\x05\x02\x00\xAA\xC5\x03\x00\xAB\xCD\x03\x00\xBD\xCC\x00\x80\xC1\xCC"
-     "\x00\x80"
-     "\xAE\xC5\x03\x00\xAF\xCD\x03\x00\xAC\xD5\x03\x00\xAD\xCD\x03\x00\x80\xED"
-     "\x01\x00"
-     "\x81\x1D\x00\x00\x82\x2D\x00\x00\x84\x8C\x02\x00\xEF\x84\x1E\x00\xE1\xC4"
-     "\x01\x00"
-     "\xC5\xCC\x00\x80\xE3\xEC\x00\x00\xE3\x58\x1E\x00\xC9\xCC\x00\x80\xE1\xF8"
-     "\x1F\x00"
-     "\xEF\x30\x00\x00\xD1\xCC\x00\x80\xD5\xCC\x00\x80\xD9\xCC\x00\x80\xDD\xCC"
-     "\x00\x80"
-     "\x86\x60\x05\x00\x87\x68\x05\x00\xE1\xCC\x00\x80\xB3\xB5\x02\x00\xE5\xCC"
-     "\x00\x80"
-     "\xB5\x75\x01\x00\xB6\x7D\x01\x00\xE9\xCC\x00\x80\xED\xCC\x00\x80\xF1\xCC"
-     "\x00\x80"
-     "\xBA\x5D\x01\x00\xBB\x55\x01\x00\xBC\x4D\x01\x00\xBD\xF1\x01\x00\xBE\xF1"
-     "\x01\x00"
-     "\xBF\xF1\x01\x00\xF5\xCC\x00\x80\xF9\xCC\x00\x80\xFD\xCC\x00\x80\x01\xCD"
-     "\x00\x80"
-     "\x05\xCD\x00\x80\x09\xCD\x00\x80\x0D\xCD\x00\x80\xEF\xE8\x01\x00\x84\x6C"
-     "\x04\x00"
-     "\xE1\xCC\x1E\x00\x11\xCD\x00\x80\xE3\x1C\x1E\x00\x15\xCD\x00\x80\x19\xCD"
-     "\x00\x80"
-     "\x1D\xCD\x00\x80\x21\xCD\x00\x80\xA3\x39\x02\x00\x25\xCD\x00\x80\x29\xCD"
-     "\x00\x80"
-     "\x2D\xCD\x00\x80\xF4\x00\x00\x00\xA6\xF1\x01\x00\xA5\xF9\x01\x00\x31\xCD"
-     "\x00\x80"
-     "\xAB\xD9\x01\x00\xAA\xD1\x01\x00\x35\xCD\x00\x80\x39\xCD\x00\x80\xAF\x7D"
-     "\x01\x00"
-     "\xAE\x7D\x01\x00\xAD\x7D\x01\x00\xAC\xC1\x01\x00\xA8\xF1\x05\x00\xA9\xF1"
-     "\x05\x00"
-     "\xAA\xF1\x05\x00\xAB\xF1\x05\x00\xAC\x31\x06\x00\xAD\x31\x06\x00\xAE\x31"
-     "\x06\x00"
-     "\xAF\x31\x06\x00\xCD\xCC\x00\x80\x82\xC9\x07\x00\x81\xC1\x07\x00\x80\xF1"
-     "\x07\x00"
-     "\xBE\xF8\x03\x00\x3D\xCD\x00\x80\x87\x88\x03\x00\x86\x1C\x03\x00\xB8\xE9"
-     "\x06\x00"
-     "\xB9\xE9\x06\x00\xBA\x89\x06\x00\xBB\x85\x06\x00\xBC\x9D\x06\x00\xBD\x81"
-     "\x06\x00"
-     "\xBE\x81\x06\x00\xBF\xB5\x06\x00\xB0\x55\x06\x00\xB1\x5D\x06\x00\xB2\x55"
-     "\x06\x00"
-     "\xB3\xE9\x06\x00\xB4\xF9\x06\x00\xB5\xF9\x06\x00\xB6\xE9\x06\x00\xB7\xE1"
-     "\x06\x00"
-     "\xA8\xB1\x07\x00\xA9\xB1\x07\x00\xAA\xB1\x07\x00\xAB\xB1\x07\x00\xAC\xD9"
-     "\x07\x00"
-     "\xAD\xC5\x07\x00\xAE\xC5\x07\x00\xAF\xFD\x07\x00\x41\xCD\x00\x80\x45\xCD"
-     "\x00\x80"
-     "\x49\xCD\x00\x80\x4D\xCD\x00\x80\x5D\xCC\x00\x80\x51\xCD\x00\x80\x55\xCD"
-     "\x00\x80"
-     "\x59\xCD\x00\x80\xB8\x69\x07\x00\xB9\x75\x07\x00\xBA\x7D\x07\x00\xBB\x75"
-     "\x07\x00"
-     "\xBC\x6D\x07\x00\xBD\x11\x07\x00\xBE\x11\x07\x00\xBF\x11\x07\x00\xB0\x85"
-     "\x07\x00"
-     "\xB1\x65\x07\x00\xB2\x6D\x07\x00\xB3\x65\x07\x00\xB4\x7D\x07\x00\xB5\x65"
-     "\x07\x00"
-     "\xB6\x6D\x07\x00\xB7\x65\x07\x00\xB3\x39\x06\x00\x5D\xCD\x00\x80\x61\xCD"
-     "\x00\x80"
-     "\x65\xCD\x00\x80\x69\xCD\x00\x80\xB6\x45\x06\x00\xB5\x51\x06\x00\x6D\xCD"
-     "\x00\x80"
-     "\xBB\xA1\x06\x00\xBA\x41\x06\x00\x71\xCD\x00\x80\x75\xCD\x00\x80\xBF\xA1"
-     "\x06\x00"
-     "\xBE\xB9\x06\x00\xBD\xB1\x06\x00\xBC\xB9\x06\x00\x79\xCD\x00\x80\xA3\x7D"
-     "\x06\x00"
-     "\x7D\xCD\x00\x80\x81\xCD\x00\x80\xA6\x01\x06\x00\x85\xCD\x00\x80\x89\xCD"
-     "\x00\x80"
-     "\xA5\x15\x06\x00\xAA\x05\x06\x00\xAB\xE5\x06\x00\x8D\xCD\x00\x80\x91\xCD"
-     "\x00\x80"
-     "\xAE\xFD\x06\x00\xAF\xE5\x06\x00\xAC\xFD\x06\x00\xAD\xF5\x06\x00\xA8\x21"
-     "\x01\x00"
-     "\xA9\x21\x01\x00\xAA\x21\x01\x00\xAB\x21\x01\x00\xAC\x25\x01\x00\xAD\x2D"
-     "\x01\x00"
-     "\xAE\x25\x01\x00\xAF\x99\x00\x00\x95\xCD\x00\x80\x99\xCD\x00\x80\x9D\xCD"
-     "\x00\x80"
-     "\xA1\xCD\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xB9\x03\x00\xA5\xCD"
-     "\x00\x80"
-     "\xB8\x8D\x00\x00\xB9\x91\x00\x00\xBA\x91\x00\x00\xBB\xA5\x00\x00\xBC\xBD"
-     "\x00\x00"
-     "\xBD\x75\x00\x00\xBE\x7D\x00\x00\xBF\x75\x00\x00\xB0\xE9\x00\x00\xB1\xE9"
-     "\x00\x00"
-     "\xB2\xF9\x00\x00\xB3\xF1\x00\x00\xB4\xD9\x00\x00\xB5\xD9\x00\x00\xB6\xB9"
-     "\x00\x00"
-     "\xB7\xB5\x00\x00\xA9\xCD\x00\x80\x84\xC0\x02\x00\xB3\xA1\x02\x00\xBE\xC0"
-     "\x02\x00"
-     "\xB5\xA5\x02\x00\xB1\xCD\x00\x80\xB5\xCD\x00\x80\xB6\xA5\x02\x00\x86\x04"
-     "\x02\x00"
-     "\x87\x60\x03\x00\xBB\x25\x03\x00\xBA\x25\x03\x00\xBD\x2D\x03\x00\xBC\x25"
-     "\x03\x00"
-     "\xBF\x15\x03\x00\xBE\x15\x03\x00\xA3\xE9\x02\x00\xB9\xCD\x00\x80\xBD\xCD"
-     "\x00\x80"
-     "\xC1\xCD\x00\x80\xC5\xCD\x00\x80\xA6\xED\x02\x00\xA5\xED\x02\x00\xC9\xCD"
-     "\x00\x80"
-     "\xAB\x6D\x03\x00\xAA\x6D\x03\x00\xCD\xCD\x00\x80\xD1\xCD\x00\x80\xAF\x5D"
-     "\x03\x00"
-     "\xAE\x5D\x03\x00\xAD\x65\x03\x00\xAC\x6D\x03\x00\xA9\x99\x00\x00\xA8\x91"
-     "\x00\x00"
-     "\xAB\xC1\x00\x00\xAA\xC9\x00\x00\xAD\xC1\x00\x00\xAC\xD9\x00\x00\xAF\xF1"
-     "\x00\x00"
-     "\xAE\xC1\x00\x00\xBE\xFC\x03\x00\xD5\xCD\x00\x80\xD9\xCD\x00\x80\xDD\xCD"
-     "\x00\x80"
-     "\xE1\xCD\x00\x80\xE5\xCD\x00\x80\xE9\xCD\x00\x80\xED\xCD\x00\x80\xB9\xC5"
-     "\x01\x00"
-     "\xB8\xC5\x01\x00\xBB\xD9\x01\x00\xBA\xC5\x01\x00\xBD\xC9\x01\x00\xBC\xC1"
-     "\x01\x00"
-     "\xBF\xFD\x01\x00\xBE\xC1\x01\x00\xB1\x45\x00\x00\xB0\x89\x00\x00\xB3\x45"
-     "\x00\x00"
-     "\xB2\x45\x00\x00\xB5\x45\x00\x00\xB4\x45\x00\x00\xB7\x45\x00\x00\xB6\x45"
-     "\x00\x00"
-     "\xF1\xCD\x00\x80\xF5\xCD\x00\x80\xF9\xCD\x00\x80\xEF\xC4\x07\x00\xFD\xCD"
-     "\x00\x80"
-     "\x01\xCE\x00\x80\x05\xCE\x00\x80\x09\xCE\x00\x80\x84\xA4\x02\x00\x0D\xCE"
-     "\x00\x80"
-     "\x11\xCE\x00\x80\x15\xCE\x00\x80\x19\xCE\x00\x80\xE1\xC0\x04\x00\x21\xCE"
-     "\x00\x80"
-     "\xE3\x20\x06\x00\x81\x6D\x00\x00\x80\x6D\x00\x00\xB3\x0D\x01\x00\x82\x7D"
-     "\x00\x00"
-     "\xB5\x1D\x01\x00\x25\xCE\x00\x80\x29\xCE\x00\x80\xB6\x0D\x01\x00\x86\xE0"
-     "\x0D\x00"
-     "\x87\xE0\x0D\x00\xBB\xCD\x01\x00\xBA\x01\x01\x00\xBD\xDD\x01\x00\xBC\xDD"
-     "\x01\x00"
-     "\xBF\xCD\x01\x00\xBE\xCD\x01\x00\x2D\xCE\x00\x80\x31\xCE\x00\x80\xEF\x7C"
-     "\x03\x00"
-     "\x35\xCE\x00\x80\x39\xCE\x00\x80\x3D\xCE\x00\x80\x41\xCE\x00\x80\xEF\x54"
-     "\x01\x00"
-     "\x45\xCE\x00\x80\xE1\xFC\x06\x00\x49\xCE\x00\x80\xE3\x1C\x06\x00\xE1\x94"
-     "\x01\x00"
-     "\x4D\xCE\x00\x80\xE3\xAC\x03\x00\x51\xCE\x00\x80\x55\xCE\x00\x80\xA6\x95"
-     "\x01\x00"
-     "\xA5\x85\x01\x00\x84\xC0\x0D\x00\xA3\x95\x01\x00\x59\xCE\x00\x80\x5D\xCE"
-     "\x00\x80"
-     "\x61\xCE\x00\x80\xAF\x55\x01\x00\xAE\x55\x01\x00\xAD\x45\x01\x00\xAC\x45"
-     "\x01\x00"
-     "\xAB\x55\x01\x00\xAA\x99\x01\x00\x65\xCE\x00\x80\x69\xCE\x00\x80\x1D\xCE"
-     "\x00\x80"
-     "\x6D\xCE\x00\x80\x71\xCE\x00\x80\x75\xCE\x00\x80\x79\xCE\x00\x80\x7D\xCE"
-     "\x00\x80"
-     "\x81\xCE\x00\x80\x85\xCE\x00\x80\xA8\x19\x0F\x00\xA9\x71\x0F\x00\xAA\x71"
-     "\x0F\x00"
-     "\xAB\xAD\x0F\x00\xAC\xB5\x0F\x00\xAD\xB9\x0F\x00\xAE\xA9\x0F\x00\xAF\xA9"
-     "\x0F\x00"
-     "\xB0\xD9\x0F\x00\xB1\xE5\x0F\x00\xB2\xE1\x0F\x00\xB3\xF9\x0F\x00\xB4\xED"
-     "\x0F\x00"
-     "\xB5\x95\x0F\x00\xB6\x8D\x0F\x00\xB7\x85\x0F\x00\xB8\xBD\x0F\x00\xB9\x9D"
-     "\x0F\x00"
-     "\xBA\x95\x0F\x00\xBB\x69\x00\x00\xBC\x79\x00\x00\xBD\x79\x00\x00\xBE\x69"
-     "\x00\x00"
-     "\xBF\x69\x00\x00\x81\x9D\x00\x00\x80\x9D\x00\x00\x89\xCE\x00\x80\x82\x05"
-     "\x00\x00"
-     "\xBE\x1C\x03\x00\x8D\xCE\x00\x80\x91\xCE\x00\x80\x95\xCE\x00\x80\x86\x80"
-     "\x03\x00"
-     "\x87\xF4\x03\x00\x99\xCE\x00\x80\x9D\xCE\x00\x80\xA1\xCE\x00\x80\xA5\xCE"
-     "\x00\x80"
-     "\xA9\xCE\x00\x80\xAD\xCE\x00\x80\xB3\x99\x0F\x00\xAD\xCD\x00\x80\xB1\xCE"
-     "\x00\x80"
-     "\xB5\xCE\x00\x80\xB9\xCE\x00\x80\xB6\x59\x0F\x00\xB5\x5D\x0F\x00\xBD\xCE"
-     "\x00\x80"
-     "\xBB\x41\x0F\x00\xBA\x7D\x0F\x00\xC1\xCE\x00\x80\xC5\xCE\x00\x80\xBF\x31"
-     "\x0F\x00"
-     "\xBE\x49\x0F\x00\xBD\x4D\x0F\x00\xBC\x55\x0F\x00\xA3\x59\x0E\x00\xC9\xCE"
-     "\x00\x80"
-     "\xCD\xCE\x00\x80\xD1\xCE\x00\x80\xD5\xCE\x00\x80\xA6\x99\x0E\x00\xA5\x9D"
-     "\x0E\x00"
-     "\xD9\xCE\x00\x80\xAB\x81\x0E\x00\xAA\xBD\x0E\x00\xDD\xCE\x00\x80\xE1\xCE"
-     "\x00\x80"
-     "\xAF\xF1\x0E\x00\xAE\x89\x0E\x00\xAD\x8D\x0E\x00\xAC\x95\x0E\x00\xE5\xCE"
-     "\x00\x80"
-     "\xE9\xCE\x00\x80\xED\xCE\x00\x80\xF1\xCE\x00\x80\x80\x1D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x39\x00\x00\xF5\xCE\x00\x80\xF9\xCE\x00\x80\x84\xA4\x01\x00\xBE\xB0"
-     "\x01\x00"
-     "\xFD\xCE\x00\x80\x86\x00\x04\x00\x87\xF4\x01\x00\x01\xCF\x00\x80\x05\xCF"
-     "\x00\x80"
-     "\xA8\x19\x01\x00\xA9\x19\x01\x00\xAA\x21\x01\x00\xAB\x21\x01\x00\xAC\x21"
-     "\x01\x00"
-     "\xAD\x21\x01\x00\xAE\x21\x01\x00\xAF\x21\x01\x00\x09\xCF\x00\x80\x0D\xCF"
-     "\x00\x80"
-     "\x11\xCF\x00\x80\x15\xCF\x00\x80\x19\xCF\x00\x80\x1D\xCF\x00\x80\x21\xCF"
-     "\x00\x80"
-     "\x25\xCF\x00\x80\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5\x00\x00\xBB\xE5"
-     "\x00\x00"
-     "\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x9D\x00\x00\xBF\x95\x00\x00\xB0\x29"
-     "\x01\x00"
-     "\xB1\x29\x01\x00\xB2\x39\x01\x00\xB3\x31\x01\x00\xB4\x19\x01\x00\xB5\x19"
-     "\x01\x00"
-     "\xB6\xFD\x00\x00\xB7\xF5\x00\x00\x29\xCF\x00\x80\x2D\xCF\x00\x80\x31\xCF"
-     "\x00\x80"
-     "\xB3\xA5\x02\x00\x35\xCF\x00\x80\xB5\xA5\x02\x00\xB6\xAD\x02\x00\x39\xCF"
-     "\x00\x80"
-     "\x3D\xCF\x00\x80\x41\xCF\x00\x80\xBA\xE1\x02\x00\xBB\xE1\x02\x00\xBC\xE1"
-     "\x02\x00"
-     "\xBD\xE9\x02\x00\xBE\xD1\x02\x00\xBF\xD1\x02\x00\xA6\xF5\x02\x00\x45\xCF"
-     "\x00\x80"
-     "\x49\xCF\x00\x80\xA5\xFD\x02\x00\x4D\xCF\x00\x80\xA3\xFD\x02\x00\x51\xCF"
-     "\x00\x80"
-     "\x55\xCF\x00\x80\xAE\x89\x02\x00\xAF\x89\x02\x00\xAC\xB9\x02\x00\xAD\xB1"
-     "\x02\x00"
-     "\xAA\xB9\x02\x00\xAB\xB9\x02\x00\x59\xCF\x00\x80\x5D\xCF\x00\x80\x80\x0D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x1D\x00\x00\x61\xCF\x00\x80\x65\xCF\x00\x80\x69\xCF"
-     "\x00\x80"
-     "\xBE\x14\x04\x00\x84\x9C\x03\x00\x71\xCF\x00\x80\x75\xCF\x00\x80\x79\xCF"
-     "\x00\x80"
-     "\x7D\xCF\x00\x80\x81\xCF\x00\x80\x85\xCF\x00\x80\x86\xB8\x04\x00\x87\x98"
-     "\x03\x00"
-     "\x89\xCF\x00\x80\x8D\xCF\x00\x80\x91\xCF\x00\x80\x95\xCF\x00\x80\xEF\xB0"
-     "\x02\x00"
-     "\x99\xCF\x00\x80\x9D\xCF\x00\x80\xA1\xCF\x00\x80\xE3\x6C\x02\x00\xA5\xCF"
-     "\x00\x80"
-     "\xE1\x80\x01\x00\xA9\xCF\x00\x80\xAD\xCF\x00\x80\xB1\xCF\x00\x80\xB5\xCF"
-     "\x00\x80"
-     "\xB9\xCF\x00\x80\xBD\xCF\x00\x80\xC1\xCF\x00\x80\xC5\xCF\x00\x80\xC9\xCF"
-     "\x00\x80"
-     "\xE1\xA8\x01\x00\xCD\xCF\x00\x80\xE3\x6C\x00\x00\xD1\xCF\x00\x80\x81\x11"
-     "\x00\x00"
-     "\x80\x39\x00\x00\xEF\xB8\x00\x00\x82\x19\x00\x00\xD5\xCF\x00\x80\xD9\xCF"
-     "\x00\x80"
-     "\xBE\xB4\x04\x00\xE1\xCF\x00\x80\xE1\x78\x0F\x00\x84\xCC\x05\x00\xE3\x78"
-     "\x0E\x00"
-     "\xE5\xCF\x00\x80\xDC\x00\x00\x00\xE9\xCF\x00\x80\xED\xCF\x00\x80\xF1\xCF"
-     "\x00\x80"
-     "\xF5\xCF\x00\x80\xF9\xCF\x00\x80\x86\x48\x04\x00\x87\x60\x05\x00\xFD\xCF"
-     "\x00\x80"
-     "\x84\xFC\x04\x00\xEF\x20\x0F\x00\x01\xD0\x00\x80\x05\xD0\x00\x80\xE1\xFC"
-     "\x0E\x00"
-     "\x09\xD0\x00\x80\xE3\x50\x0E\x00\xBE\x1C\x07\x00\x0D\xD0\x00\x80\x11\xD0"
-     "\x00\x80"
-     "\x15\xD0\x00\x80\x19\xD0\x00\x80\x1D\xD0\x00\x80\x21\xD0\x00\x80\x25\xD0"
-     "\x00\x80"
-     "\x29\xD0\x00\x80\x2D\xD0\x00\x80\x31\xD0\x00\x80\xEF\xD4\x01\x00\x35\xD0"
-     "\x00\x80"
-     "\x39\xD0\x00\x80\x3D\xD0\x00\x80\xB3\x6D\x02\x00\x41\xD0\x00\x80\xB5\x7D"
-     "\x02\x00"
-     "\xB6\x75\x02\x00\x45\xD0\x00\x80\x49\xD0\x00\x80\xBF\x00\x38\x00\xBA\x51"
-     "\x02\x00"
-     "\xBB\x51\x02\x00\xBC\xF1\x01\x00\xBD\xF1\x01\x00\xBE\xF1\x01\x00\xBF\xED"
-     "\x01\x00"
-     "\xA3\xA1\x05\x00\x82\xE5\x07\x00\x81\xD5\x07\x00\x80\xDD\x07\x00\x85\xF4"
-     "\x01\x00"
-     "\xA6\xB9\x05\x00\xA5\xB1\x05\x00\x51\xD0\x00\x80\xAB\x9D\x05\x00\xAA\x9D"
-     "\x05\x00"
-     "\x86\x88\x3E\x00\x87\x8C\x03\x00\xAF\x21\x06\x00\xAE\x3D\x06\x00\xAD\x3D"
-     "\x06\x00"
-     "\xAC\x3D\x06\x00\x55\xD0\x00\x80\xB3\xED\x06\x00\x59\xD0\x00\x80\x5D\xD0"
-     "\x00\x80"
-     "\xB6\xE5\x06\x00\x61\xD0\x00\x80\x65\xD0\x00\x80\xB5\xED\x06\x00\xBA\xA9"
-     "\x06\x00"
-     "\xBB\xA9\x06\x00\x69\xD0\x00\x80\x6D\xD0\x00\x80\xBE\xAD\x06\x00\xBF\x91"
-     "\x06\x00"
-     "\xBC\xAD\x06\x00\xBD\xA5\x06\x00\x6D\xCF\x00\x80\xDD\xCF\x00\x80\x71\xD0"
-     "\x00\x80"
-     "\x84\x80\x00\x00\x75\xD0\x00\x80\x79\xD0\x00\x80\x7D\xD0\x00\x80\x81\xD0"
-     "\x00\x80"
-     "\xA8\x01\x06\x00\xA9\x01\x06\x00\xAA\x01\x06\x00\xAB\x01\x06\x00\xAC\x01"
-     "\x06\x00"
-     "\xAD\x01\x06\x00\xAE\x79\x06\x00\xAF\x79\x06\x00\xB0\x09\x06\x00\xB1\x09"
-     "\x06\x00"
-     "\xB2\x1D\x06\x00\xB3\x15\x06\x00\xB4\x0D\x06\x00\xB5\x31\x06\x00\xB6\x35"
-     "\x06\x00"
-     "\xB7\x2D\x06\x00\xB8\x15\x06\x00\xB9\x1D\x06\x00\xBA\x15\x06\x00\xBB\xED"
-     "\x07\x00"
-     "\xBC\xF5\x07\x00\xBD\xFD\x07\x00\xBE\xF5\x07\x00\xBF\xE9\x07\x00\xA3\xA9"
-     "\x07\x00"
-     "\x85\xD0\x00\x80\x89\xD0\x00\x80\x8D\xD0\x00\x80\x4D\xD0\x00\x80\xA6\xA1"
-     "\x07\x00"
-     "\xA5\xA9\x07\x00\x91\xD0\x00\x80\xAB\xED\x07\x00\xAA\xED\x07\x00\x95\xD0"
-     "\x00\x80"
-     "\x99\xD0\x00\x80\xAF\xD5\x07\x00\xAE\xE9\x07\x00\xAD\xE1\x07\x00\xAC\xE9"
-     "\x07\x00"
-     "\x9D\xD0\x00\x80\xB3\x11\x06\x00\xA1\xD0\x00\x80\xA5\xD0\x00\x80\xB6\x31"
-     "\x06\x00"
-     "\xA9\xD0\x00\x80\xAD\xD0\x00\x80\xB5\x01\x06\x00\xBA\x1D\x06\x00\xBB\xE5"
-     "\x06\x00"
-     "\xB1\xD0\x00\x80\xB5\xD0\x00\x80\xBE\xE9\x06\x00\xBF\xE9\x06\x00\xBC\xF1"
-     "\x06\x00"
-     "\xBD\xE9\x06\x00\xA8\xBD\x06\x00\xA9\xC9\x06\x00\xAA\xDD\x06\x00\xAB\xD5"
-     "\x06\x00"
-     "\xAC\xCD\x06\x00\xAD\xF5\x06\x00\xAE\xFD\x06\x00\xAF\xED\x06\x00\x80\xC9"
-     "\x00\x00"
-     "\x81\xC9\x00\x00\x82\x05\x00\x00\xB9\xD0\x00\x80\xBE\xBC\x3F\x00\xC1\xD0"
-     "\x00\x80"
-     "\x87\x48\x00\x00\x86\x1C\x3E\x00\xB8\x69\x06\x00\xB9\x69\x06\x00\xBA\x79"
-     "\x06\x00"
-     "\xBB\x79\x06\x00\xBC\x69\x06\x00\xBD\x15\x06\x00\xBE\x1D\x06\x00\xBF\x15"
-     "\x06\x00"
-     "\xB0\x9D\x06\x00\xB1\x65\x06\x00\xB2\x61\x06\x00\xB3\x79\x06\x00\xB4\x69"
-     "\x06\x00"
-     "\xB5\x69\x06\x00\xB6\x59\x06\x00\xB7\x59\x06\x00\xA3\x55\x05\x00\x8F\x05"
-     "\x1A\x00"
-     "\xC5\xD0\x00\x80\xBE\x80\x3E\x00\xC9\xD0\x00\x80\xA6\x75\x05\x00\xA5\x45"
-     "\x05\x00"
-     "\xCD\xD0\x00\x80\xAB\xA1\x05\x00\xAA\x59\x05\x00\xD1\xD0\x00\x80\xD5\xD0"
-     "\x00\x80"
-     "\xAF\xAD\x05\x00\xAE\xAD\x05\x00\xAD\xAD\x05\x00\xAC\xB5\x05\x00\x9E\x85"
-     "\x06\x00"
-     "\x9F\x45\x06\x00\x9C\x69\x0A\x00\x9D\x79\x0A\x00\x9A\xC1\x0E\x00\x9B\xE5"
-     "\x0A\x00"
-     "\x98\xF5\x0E\x00\x99\x7D\x0E\x00\x96\x41\x12\x00\x97\xB9\x12\x00\x94\xC5"
-     "\x16\x00"
-     "\x95\xF5\x12\x00\x92\x75\x16\x00\x93\x7D\x16\x00\x90\x7D\x1A\x00\x91\x1D"
-     "\x1A\x00"
-     "\xD9\xD0\x00\x80\xDD\xD0\x00\x80\xBE\x04\x3F\x00\xBE\x4C\x3F\x00\xE1\x94"
-     "\x00\x00"
-     "\xE1\xA4\x00\x00\xE3\x2C\x3B\x00\xE3\x00\x00\x00\xE1\xD0\x00\x80\xE5\xD0"
-     "\x00\x80"
-     "\xEF\xFC\x33\x00\xEF\xE4\x01\x00\xA2\x31\x02\x00\xA3\x31\x02\x00\xA0\x2D"
-     "\x06\x00"
-     "\xA1\xD1\x07\x00\xB3\x89\x02\x00\xE9\xD0\x00\x80\xED\xD0\x00\x80\xF1\xD0"
-     "\x00\x80"
-     "\xF5\xD0\x00\x80\xB6\x8D\x02\x00\xB5\x8D\x02\x00\xF9\xD0\x00\x80\xBB\x71"
-     "\x02\x00"
-     "\xBA\x6D\x02\x00\xFD\xD0\x00\x80\x01\xD1\x00\x80\xBF\xF5\x03\x00\xBE\x69"
-     "\x02\x00"
-     "\xBD\x61\x02\x00\xBC\x69\x02\x00\xE1\x70\x06\x00\xE1\xA4\x07\x00\xE3\x6C"
-     "\x06\x00"
-     "\xE3\x58\x06\x00\x05\xD1\x00\x80\x09\xD1\x00\x80\x0D\xD1\x00\x80\x11\xD1"
-     "\x00\x80"
-     "\xB1\x01\x28\x00\xB0\x21\x2E\x00\xB3\xBD\x2B\x00\xB2\x0D\x2B\x00\xB5\x11"
-     "\x27\x00"
-     "\xB4\x01\x24\x00\xEF\x50\x06\x00\xEF\xFC\x06\x00\xA9\x65\x37\x00\xA8\x65"
-     "\x37\x00"
-     "\xAB\x49\x33\x00\xAA\x01\x30\x00\xAD\x0D\x32\x00\xAC\x45\x33\x00\xAF\x79"
-     "\x2F\x00"
-     "\xAE\x01\x2C\x00\xA1\x71\x02\x00\x15\xD1\x00\x80\xA3\x45\x3F\x00\xA2\x4D"
-     "\x3F\x00"
-     "\xA5\x51\x3B\x00\xA4\x01\x38\x00\xA7\x01\x34\x00\xA6\xED\x3B\x00\xBC\xD9"
-     "\x05\x00"
-     "\xBD\xD9\x05\x00\xBE\xF1\x05\x00\xBF\xF1\x05\x00\xB8\xB5\x05\x00\xB9\xD1"
-     "\x05\x00"
-     "\xBA\xD1\x05\x00\xBB\xCD\x05\x00\xB4\xA9\x05\x00\xB5\xB5\x05\x00\xB6\xB5"
-     "\x05\x00"
-     "\xB7\x8D\x05\x00\xB0\x09\x06\x00\xB1\x09\x06\x00\xB2\xB9\x05\x00\xB3\xB9"
-     "\x05\x00"
-     "\xAC\x69\x06\x00\xAD\x69\x06\x00\xAE\x79\x06\x00\xAF\x79\x06\x00\xA8\xA9"
-     "\x06\x00"
-     "\xA9\xA9\x06\x00\xAA\x79\x06\x00\xAB\x79\x06\x00\x19\xD1\x00\x80\x1D\xD1"
-     "\x00\x80"
-     "\x21\xD1\x00\x80\x25\xD1\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x29\xD1\x00\x80\xA8\x1D\x3C\x00\xA9\x45\x3C\x00\xAA\x41\x3C\x00\xAB\x41"
-     "\x3C\x00"
-     "\xAC\x45\x3C\x00\xAD\x4D\x3C\x00\xAE\x71\x3C\x00\xAF\x71\x3C\x00\xBD\xD0"
-     "\x00\x80"
-     "\x84\x2C\x02\x00\x31\xD1\x00\x80\x2D\xD1\x00\x80\x86\x14\x00\x00\x87\xAC"
-     "\x03\x00"
-     "\x35\xD1\x00\x80\x39\xD1\x00\x80\xB8\x5D\x3D\x00\xB9\x6D\x3D\x00\xBA\x65"
-     "\x3D\x00"
-     "\xBB\x0D\x3D\x00\xBC\x19\x3D\x00\xBD\x19\x3D\x00\xBE\x0D\x3D\x00\xBF\x05"
-     "\x3D\x00"
-     "\xB0\xE9\x3C\x00\xB1\xF5\x3C\x00\xB2\xFD\x3C\x00\xB3\xF5\x3C\x00\xB4\xED"
-     "\x3C\x00"
-     "\xB5\x71\x3D\x00\xB6\x71\x3D\x00\xB7\x71\x3D\x00\xA3\xC9\x3C\x00\x3D\xD1"
-     "\x00\x80"
-     "\x41\xD1\x00\x80\x84\x0C\x00\x00\x45\xD1\x00\x80\xA6\xCD\x3C\x00\xA5\xCD"
-     "\x3C\x00"
-     "\x49\xD1\x00\x80\xAB\x31\x3C\x00\xAA\x2D\x3C\x00\x4D\xD1\x00\x80\x51\xD1"
-     "\x00\x80"
-     "\xAF\xB5\x3D\x00\xAE\x29\x3C\x00\xAD\x21\x3C\x00\xAC\x29\x3C\x00\x55\xD1"
-     "\x00\x80"
-     "\xB3\xD5\x3E\x00\x59\xD1\x00\x80\x5D\xD1\x00\x80\xB6\xFD\x3E\x00\x61\xD1"
-     "\x00\x80"
-     "\x65\xD1\x00\x80\xB5\xF5\x3E\x00\xBA\x41\x3E\x00\xBB\x41\x3E\x00\x69\xD1"
-     "\x00\x80"
-     "\x6D\xD1\x00\x80\xBE\x3D\x3E\x00\xBF\x25\x3E\x00\xBC\x3D\x3E\x00\xBD\x35"
-     "\x3E\x00"
-     "\xA8\xB5\x3C\x00\xA9\x0D\x3F\x00\xAA\x19\x3F\x00\xAB\x11\x3F\x00\xAC\x39"
-     "\x3F\x00"
-     "\xAD\x39\x3F\x00\xAE\x99\x3F\x00\xAF\x99\x3F\x00\x71\xD1\x00\x80\x75\xD1"
-     "\x00\x80"
-     "\x79\xD1\x00\x80\x7D\xD1\x00\x80\x81\xD1\x00\x80\x85\xD1\x00\x80\x89\xD1"
-     "\x00\x80"
-     "\x8D\xD1\x00\x80\xB8\xAD\x3F\x00\xB9\x55\x3E\x00\xBA\x5D\x3E\x00\xBB\x55"
-     "\x3E\x00"
-     "\xBC\x4D\x3E\x00\xBD\x31\x3E\x00\xBE\xB9\x3E\x00\xBF\xB9\x3E\x00\xB0\xC1"
-     "\x3F\x00"
-     "\xB1\xC1\x3F\x00\xB2\xAD\x3F\x00\xB3\xA5\x3F\x00\xB4\xBD\x3F\x00\xB5\xA5"
-     "\x3F\x00"
-     "\xB6\xA5\x3F\x00\xB7\x95\x3F\x00\x91\xD1\x00\x80\x95\xD1\x00\x80\x9B\x09"
-     "\x3D\x00"
-     "\x9A\xE1\x38\x00\x82\x31\x00\x00\x9C\x11\x3D\x00\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x99\xD1\x00\x80\xA1\xD1\x00\x80\x93\x89\x35\x00\x92\x1D\x35\x00\x95\x75"
-     "\x39\x00"
-     "\x94\x81\x35\x00\x97\x2D\x39\x00\x96\xBD\x39\x00\x86\x60\x07\x00\x87\x84"
-     "\x00\x00"
-     "\x8B\xE9\x2D\x00\x8A\x55\x2D\x00\x8D\x7D\x31\x00\x8C\x09\x2D\x00\x8F\x49"
-     "\x31\x00"
-     "\x8E\xFD\x31\x00\xA5\xD1\x00\x80\xA9\xD1\x00\x80\x83\xED\x25\x00\x82\xED"
-     "\x25\x00"
-     "\x85\xF5\x29\x00\x84\x95\x25\x00\x87\x55\x29\x00\x86\xF5\x29\x00\xA3\x11"
-     "\x3C\x00"
-     "\xAD\xD1\x00\x80\x84\x04\x03\x00\x84\x40\x05\x00\xB1\xD1\x00\x80\xA6\x39"
-     "\x3C\x00"
-     "\xA5\x31\x3C\x00\xB5\xD1\x00\x80\xAB\x85\x3C\x00\xAA\x85\x3C\x00\xBE\x60"
-     "\x05\x00"
-     "\xB9\xD1\x00\x80\xAF\xE1\x3C\x00\xAE\xF9\x3C\x00\xAD\xF1\x3C\x00\xAC\xF9"
-     "\x3C\x00"
-     "\xBD\xD1\x00\x80\xC1\xD1\x00\x80\xC5\xD1\x00\x80\xB3\x71\x02\x00\xC9\xD1"
-     "\x00\x80"
-     "\xB5\xDD\x02\x00\xB6\xC9\x02\x00\xCD\xD1\x00\x80\xD1\xD1\x00\x80\xD5\xD1"
-     "\x00\x80"
-     "\xBA\xC5\x02\x00\xBB\xAD\x02\x00\xBC\xB9\x02\x00\xBD\xB9\x02\x00\xBE\xAD"
-     "\x02\x00"
-     "\xBF\xA5\x02\x00\xA9\xB9\x02\x00\xA8\xAD\x02\x00\xAB\x0D\x02\x00\xAA\xB1"
-     "\x02\x00"
-     "\xAD\x01\x02\x00\xAC\x1D\x02\x00\xAF\x31\x02\x00\xAE\x01\x02\x00\x84\xEC"
-     "\x06\x00"
-     "\xD9\xD1\x00\x80\xDD\xD1\x00\x80\xE1\xD1\x00\x80\xE5\xD1\x00\x80\xE9\xD1"
-     "\x00\x80"
-     "\xED\xD1\x00\x80\xF1\xD1\x00\x80\xB9\xE5\x03\x00\xB8\xE5\x03\x00\xBB\xE5"
-     "\x03\x00"
-     "\xBA\xE5\x03\x00\xBD\xE5\x03\x00\xBC\xE5\x03\x00\xBF\xE5\x03\x00\xBE\xE5"
-     "\x03\x00"
-     "\xB1\x21\x02\x00\xB0\x49\x02\x00\xB3\x25\x02\x00\xB2\x25\x02\x00\xB5\x29"
-     "\x02\x00"
-     "\xB4\x21\x02\x00\xB7\x15\x02\x00\xB6\x15\x02\x00\xA3\x3D\x03\x00\xF5\xD1"
-     "\x00\x80"
-     "\xF9\xD1\x00\x80\xFD\xD1\x00\x80\x01\xD2\x00\x80\xA6\x85\x03\x00\xA5\x91"
-     "\x03\x00"
-     "\x05\xD2\x00\x80\xAB\xE1\x03\x00\xAA\x89\x03\x00\x09\xD2\x00\x80\x11\xD2"
-     "\x00\x80"
-     "\xAF\xE9\x03\x00\xAE\xE1\x03\x00\xAD\xF5\x03\x00\xAC\xF5\x03\x00\x80\x51"
-     "\x00\x00"
-     "\x81\x59\x00\x00\x82\x61\x00\x00\xA3\xA9\x02\x00\x15\xD2\x00\x80\xA5\x45"
-     "\x01\x00"
-     "\xA6\x45\x01\x00\x19\xD2\x00\x80\x86\xC0\x04\x00\x87\xE4\x05\x00\xAA\x2D"
-     "\x01\x00"
-     "\xAB\x25\x01\x00\xAC\x21\x01\x00\xAD\x29\x01\x00\xAE\x19\x01\x00\xAF\x19"
-     "\x01\x00"
-     "\x1D\xD2\x00\x80\x21\xD2\x00\x80\x25\xD2\x00\x80\x29\xD2\x00\x80\xEF\xE0"
-     "\x3F\x00"
-     "\x2D\xD2\x00\x80\xBE\xD0\x07\x00\xEF\x68\x3F\x00\xE3\x3C\x01\x00\xE1\x00"
-     "\x3C\x00"
-     "\xE1\x80\x01\x00\xE3\x98\x01\x00\x31\xD2\x00\x80\x35\xD2\x00\x80\x39\xD2"
-     "\x00\x80"
-     "\x3D\xD2\x00\x80\xB3\x25\x02\x00\x41\xD2\x00\x80\x45\xD2\x00\x80\x49\xD2"
-     "\x00\x80"
-     "\x4D\xD2\x00\x80\xB6\xC9\x01\x00\xB5\xC9\x01\x00\x51\xD2\x00\x80\xBB\xA9"
-     "\x01\x00"
-     "\xBA\xA1\x01\x00\x55\xD2\x00\x80\x59\xD2\x00\x80\xBF\x95\x01\x00\xBE\x95"
-     "\x01\x00"
-     "\xBD\xA5\x01\x00\xBC\xAD\x01\x00\xA8\xA5\x05\x00\xA9\xE5\x05\x00\xAA\xED"
-     "\x05\x00"
-     "\xAB\x09\x06\x00\xAC\x01\x06\x00\xAD\x09\x06\x00\xAE\x35\x06\x00\xAF\x5D"
-     "\x06\x00"
-     "\x0D\xD2\x00\x80\x5D\xD2\x00\x80\x61\xD2\x00\x80\x65\xD2\x00\x80\x69\xD2"
-     "\x00\x80"
-     "\x6D\xD2\x00\x80\x71\xD2\x00\x80\x75\xD2\x00\x80\xB8\xED\x06\x00\xB9\xF5"
-     "\x06\x00"
-     "\xBA\xFD\x06\x00\xBB\xF5\x06\x00\xBC\xED\x06\x00\xBD\x51\x07\x00\xBE\x51"
-     "\x07\x00"
-     "\xBF\x51\x07\x00\xB0\x25\x06\x00\xB1\x2D\x06\x00\xB2\x39\x06\x00\xB3\x31"
-     "\x06\x00"
-     "\xB4\x11\x06\x00\xB5\x11\x06\x00\xB6\xDD\x06\x00\xB7\xD5\x06\x00\x80\xB9"
-     "\x00\x00"
-     "\x81\xB9\x00\x00\x82\x05\x00\x00\x79\xD2\x00\x80\x7D\xD2\x00\x80\x81\xD2"
-     "\x00\x80"
-     "\x87\xA8\x03\x00\x86\x3C\x00\x00\x84\x34\x01\x00\x85\xD2\x00\x80\x89\xD2"
-     "\x00\x80"
-     "\x8D\xD2\x00\x80\x91\xD2\x00\x80\x95\xD2\x00\x80\x99\xD2\x00\x80\x9D\xD2"
-     "\x00\x80"
-     "\xA1\xD2\x00\x80\xA5\xD2\x00\x80\xEF\xBC\x38\x00\xA9\xD2\x00\x80\xAD\xD2"
-     "\x00\x80"
-     "\xB1\xD2\x00\x80\xB5\xD2\x00\x80\xB9\xD2\x00\x80\xBD\xD2\x00\x80\xC1\xD2"
-     "\x00\x80"
-     "\xC5\xD2\x00\x80\xC9\xD2\x00\x80\xE1\x28\x07\x00\xCD\xD2\x00\x80\xE3\xE0"
-     "\x04\x00"
-     "\xD1\xD2\x00\x80\xA8\x2D\x06\x00\xA9\x55\x06\x00\xAA\x55\x06\x00\xAB\x8D"
-     "\x06\x00"
-     "\xAC\x95\x06\x00\xAD\x9D\x06\x00\xAE\x95\x06\x00\xAF\x89\x06\x00\xD5\xD2"
-     "\x00\x80"
-     "\xD9\xD2\x00\x80\xDD\xD2\x00\x80\xE1\xD2\x00\x80\x80\x1D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\xBD\x00\x00\xE5\xD2\x00\x80\xB8\x9D\x06\x00\xB9\xAD\x06\x00\xBA\xA5"
-     "\x06\x00"
-     "\xBB\x69\x01\x00\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x69\x01\x00\xBF\x69"
-     "\x01\x00"
-     "\xB0\xD1\x06\x00\xB1\xD1\x06\x00\xB2\xD1\x06\x00\xB3\xD1\x06\x00\xB4\xB5"
-     "\x06\x00"
-     "\xB5\xBD\x06\x00\xB6\xB5\x06\x00\xB7\xAD\x06\x00\xB3\xDD\x06\x00\xE9\xD2"
-     "\x00\x80"
-     "\x86\x28\x0F\x00\x87\x44\x01\x00\xF1\xD2\x00\x80\xB6\x1D\x06\x00\xB5\x1D"
-     "\x06\x00"
-     "\xF5\xD2\x00\x80\xBB\x15\x06\x00\xBA\x15\x06\x00\xF9\xD2\x00\x80\xFD\xD2"
-     "\x00\x80"
-     "\xBF\x79\x06\x00\xBE\x71\x06\x00\xBD\x05\x06\x00\xBC\x05\x06\x00\x01\xD3"
-     "\x00\x80"
-     "\xA3\x99\x06\x00\x05\xD3\x00\x80\x09\xD3\x00\x80\xA6\x59\x06\x00\x0D\xD3"
-     "\x00\x80"
-     "\x11\xD3\x00\x80\xA5\x59\x06\x00\xAA\x51\x06\x00\xAB\x51\x06\x00\x15\xD3"
-     "\x00\x80"
-     "\x19\xD3\x00\x80\xAE\x35\x06\x00\xAF\x3D\x06\x00\xAC\x41\x06\x00\xAD\x41"
-     "\x06\x00"
-     "\xA8\x95\x02\x00\xA9\xA1\x02\x00\xAA\xA1\x02\x00\xAB\xA1\x02\x00\xAC\xA1"
-     "\x02\x00"
-     "\xAD\xA1\x02\x00\xAE\xA1\x02\x00\xAF\xA1\x02\x00\x1D\xD3\x00\x80\x21\xD3"
-     "\x00\x80"
-     "\x25\xD3\x00\x80\x29\xD3\x00\x80\x2D\xD3\x00\x80\x31\xD3\x00\x80\x35\xD3"
-     "\x00\x80"
-     "\x39\xD3\x00\x80\xB8\xA5\x02\x00\xB9\xAD\x02\x00\xBA\xA5\x02\x00\xBB\xB9"
-     "\x02\x00"
-     "\xBC\xAD\x02\x00\xBD\x75\x03\x00\xBE\x7D\x03\x00\xBF\x75\x03\x00\xB0\xA1"
-     "\x02\x00"
-     "\xB1\xA1\x02\x00\xB2\xA5\x02\x00\xB3\xBD\x02\x00\xB4\xA5\x02\x00\xB5\xAD"
-     "\x02\x00"
-     "\xB6\xA5\x02\x00\xB7\x9D\x02\x00\xB3\x29\x02\x00\xBE\xC8\x03\x00\x3D\xD3"
-     "\x00\x80"
-     "\x41\xD3\x00\x80\x45\xD3\x00\x80\xB6\x21\x02\x00\xB5\x29\x02\x00\x49\xD3"
-     "\x00\x80"
-     "\xBB\x65\x02\x00\xBA\x65\x02\x00\x4D\xD3\x00\x80\x51\xD3\x00\x80\xBF\x69"
-     "\x02\x00"
-     "\xBE\x61\x02\x00\xBD\x69\x02\x00\xBC\x71\x02\x00\x82\x19\x00\x00\xA3\x6D"
-     "\x02\x00"
-     "\x80\x65\x00\x00\x81\x19\x00\x00\xA6\x65\x02\x00\x55\xD3\x00\x80\x59\xD3"
-     "\x00\x80"
-     "\xA5\x6D\x02\x00\xAA\x21\x02\x00\xAB\x21\x02\x00\x86\x48\x0D\x00\x87\x5C"
-     "\x03\x00"
-     "\xAE\x25\x02\x00\xAF\x2D\x02\x00\xAC\x35\x02\x00\xAD\x2D\x02\x00\x61\xD3"
-     "\x00\x80"
-     "\x65\xD3\x00\x80\x69\xD3\x00\x80\x6D\xD3\x00\x80\x71\xD3\x00\x80\xE1\x90"
-     "\x01\x00"
-     "\x75\xD3\x00\x80\xE3\xAC\x07\x00\x79\xD3\x00\x80\x7D\xD3\x00\x80\x81\xD3"
-     "\x00\x80"
-     "\xEF\x4C\x06\x00\xCC\x00\x00\x00\x85\xD3\x00\x80\x89\xD3\x00\x80\x8D\xD3"
-     "\x00\x80"
-     "\xB5\x81\x02\x00\x91\xD3\x00\x80\x84\x04\x0D\x00\xB6\x85\x02\x00\x95\xD3"
-     "\x00\x80"
-     "\x99\xD3\x00\x80\xB3\xF1\x02\x00\x9D\xD3\x00\x80\xBD\x41\x01\x00\xBC\x89"
-     "\x02\x00"
-     "\xBF\x41\x01\x00\xBE\x49\x01\x00\xA1\xD3\x00\x80\x9D\xD1\x00\x80\xBB\x91"
-     "\x02\x00"
-     "\xBA\xA5\x02\x00\xA5\xD3\x00\x80\xA9\xD3\x00\x80\xAD\xD3\x00\x80\xB1\xD3"
-     "\x00\x80"
-     "\xBE\x9C\x0C\x00\xB5\xD3\x00\x80\xB9\xD3\x00\x80\xEF\x4C\x07\x00\x85\x9C"
-     "\x1F\x00"
-     "\xE1\xDC\x06\x00\xBD\xD3\x00\x80\xE3\xB0\x07\x00\x80\x0D\x00\x00\x81\x71"
-     "\x00\x00"
-     "\x82\x71\x00\x00\xC1\xD3\x00\x80\xA3\x69\x02\x00\xC5\xD3\x00\x80\xC9\xD3"
-     "\x00\x80"
-     "\xD1\xD3\x00\x80\xD5\xD3\x00\x80\xA6\x1D\x02\x00\xA5\x19\x02\x00\xD9\xD3"
-     "\x00\x80"
-     "\xAB\x09\x02\x00\xAA\x3D\x02\x00\x86\x48\x0F\x00\x87\x00\x0D\x00\xAF\xD9"
-     "\x01\x00"
-     "\xAE\xD1\x01\x00\xAD\xD9\x01\x00\xAC\x11\x02\x00\xA8\x91\x0D\x00\xA9\x99"
-     "\x0D\x00"
-     "\xAA\xA1\x0D\x00\xAB\xA1\x0D\x00\xAC\xD1\x0D\x00\xAD\xDD\x0D\x00\xAE\xD5"
-     "\x0D\x00"
-     "\xAF\xC9\x0D\x00\x5D\xD3\x00\x80\xCD\xD3\x00\x80\xDD\xD3\x00\x80\xE1\xD3"
-     "\x00\x80"
-     "\xE5\xD3\x00\x80\xE9\xD3\x00\x80\xED\xD3\x00\x80\xF1\xD3\x00\x80\xB8\x75"
-     "\x0E\x00"
-     "\xB9\x7D\x0E\x00\xBA\x75\x0E\x00\xBB\xC5\x0E\x00\xBC\xDD\x0E\x00\xBD\xC5"
-     "\x0E\x00"
-     "\xBE\xC5\x0E\x00\xBF\xF5\x0E\x00\xB0\xB9\x0D\x00\xB1\x8D\x0D\x00\xB2\x85"
-     "\x0D\x00"
-     "\xB3\x4D\x0E\x00\xB4\x55\x0E\x00\xB5\x5D\x0E\x00\xB6\x55\x0E\x00\xB7\x4D"
-     "\x0E\x00"
-     "\xB3\x89\x0F\x00\xF5\xD3\x00\x80\xF9\xD3\x00\x80\xFD\xD3\x00\x80\x01\xD4"
-     "\x00\x80"
-     "\xB6\x8D\x0F\x00\xB5\x89\x0F\x00\x05\xD4\x00\x80\xBB\x95\x0F\x00\xBA\x8D"
-     "\x0F\x00"
-     "\x09\xD4\x00\x80\x0D\xD4\x00\x80\xBF\xE1\x0F\x00\xBE\xF1\x0F\x00\xBD\x85"
-     "\x0F\x00"
-     "\xBC\x85\x0F\x00\x11\xD4\x00\x80\xEF\xB8\x0C\x00\x15\xD4\x00\x80\x19\xD4"
-     "\x00\x80"
-     "\x1D\xD4\x00\x80\x21\xD4\x00\x80\x25\xD4\x00\x80\x29\xD4\x00\x80\x2D\xD4"
-     "\x00\x80"
-     "\x82\xC5\x01\x00\x81\x11\x00\x00\x80\x1D\x00\x00\x31\xD4\x00\x80\xE3\x44"
-     "\x0E\x00"
-     "\x35\xD4\x00\x80\xE1\x6C\x0F\x00\xA8\x01\x0E\x00\xA9\x01\x0E\x00\xAA\x01"
-     "\x0E\x00"
-     "\xAB\x01\x0E\x00\xAC\x01\x0E\x00\xAD\x0D\x0E\x00\xAE\x39\x0E\x00\xAF\x39"
-     "\x0E\x00"
-     "\x39\xD4\x00\x80\x84\x2C\x02\x00\xBE\x24\x01\x00\x41\xD4\x00\x80\x86\xD0"
-     "\x00\x00"
-     "\x87\xF0\x01\x00\x45\xD4\x00\x80\x49\xD4\x00\x80\xB8\x31\x0E\x00\xB9\xC5"
-     "\x0E\x00"
-     "\xBA\xCD\x0E\x00\xBB\xC5\x0E\x00\xBC\xDD\x0E\x00\xBD\xC1\x0E\x00\xBE\xC1"
-     "\x0E\x00"
-     "\xBF\xF5\x0E\x00\xB0\x49\x0E\x00\xB1\x51\x0E\x00\xB2\x51\x0E\x00\xB3\x51"
-     "\x0E\x00"
-     "\xB4\x35\x0E\x00\xB5\x25\x0E\x00\xB6\x2D\x0E\x00\xB7\x21\x0E\x00\xA3\x89"
-     "\x0E\x00"
-     "\x4D\xD4\x00\x80\x51\xD4\x00\x80\x55\xD4\x00\x80\x59\xD4\x00\x80\xA6\x8D"
-     "\x0E\x00"
-     "\xA5\x89\x0E\x00\x5D\xD4\x00\x80\xAB\x95\x0E\x00\xAA\x8D\x0E\x00\x61\xD4"
-     "\x00\x80"
-     "\x65\xD4\x00\x80\xAF\xE1\x0E\x00\xAE\xF1\x0E\x00\xAD\x85\x0E\x00\xAC\x85"
-     "\x0E\x00"
-     "\x69\xD4\x00\x80\xB3\x19\x0E\x00\x6D\xD4\x00\x80\x71\xD4\x00\x80\xB6\xE1"
-     "\x01\x00"
-     "\x75\xD4\x00\x80\x79\xD4\x00\x80\xB5\xF1\x01\x00\xBA\xDD\x01\x00\xBB\x61"
-     "\x01\x00"
-     "\x7D\xD4\x00\x80\x81\xD4\x00\x80\xBE\x61\x01\x00\xBF\x69\x01\x00\xBC\x71"
-     "\x01\x00"
-     "\xBD\x71\x01\x00\xA8\xCD\x0D\x00\xA9\xD5\x0D\x00\xAA\xD5\x0D\x00\xAB\x25"
-     "\x02\x00"
-     "\xAC\x3D\x02\x00\xAD\x25\x02\x00\xAE\x2D\x02\x00\xAF\x25\x02\x00\x85\xD4"
-     "\x00\x80"
-     "\x89\xD4\x00\x80\x8D\xD4\x00\x80\x91\xD4\x00\x80\x95\xD4\x00\x80\x99\xD4"
-     "\x00\x80"
-     "\x9D\xD4\x00\x80\xA1\xD4\x00\x80\xB8\x81\x02\x00\xB9\x8D\x02\x00\xBA\x85"
-     "\x02\x00"
-     "\xBB\x9D\x02\x00\xBC\x85\x02\x00\xBD\x8D\x02\x00\xBE\x85\x02\x00\xBF\xBD"
-     "\x02\x00"
-     "\xB0\x5D\x02\x00\xB1\xE1\x02\x00\xB2\xE1\x02\x00\xB3\xE1\x02\x00\xB4\xE1"
-     "\x02\x00"
-     "\xB5\xE9\x02\x00\xB6\xD1\x02\x00\xB7\xD1\x02\x00\xA3\x5D\x0D\x00\xA5\xD4"
-     "\x00\x80"
-     "\xA9\xD4\x00\x80\xBE\xC0\x03\x00\x84\x94\x02\x00\xA6\xA5\x02\x00\xA5\xB5"
-     "\x02\x00"
-     "\xB1\xD4\x00\x80\xAB\x25\x02\x00\xAA\x99\x02\x00\x86\x28\x04\x00\x87\xA0"
-     "\x04\x00"
-     "\xAF\x2D\x02\x00\xAE\x25\x02\x00\xAD\x35\x02\x00\xAC\x35\x02\x00\x81\x45"
-     "\x00\x00"
-     "\x80\xF9\x00\x00\xB3\xF5\x03\x00\x82\x45\x00\x00\xB5\x59\x03\x00\xB5\xD4"
-     "\x00\x80"
-     "\xB9\xD4\x00\x80\xB6\x51\x03\x00\xBD\xD4\x00\x80\xC1\xD4\x00\x80\xBB\x45"
-     "\x03\x00"
-     "\xBA\x45\x03\x00\xBD\x39\x03\x00\xBC\x35\x03\x00\xBF\x2D\x03\x00\xBE\x31"
-     "\x03\x00"
-     "\xB8\xED\x02\x00\xB9\x25\x01\x00\xBA\x2D\x01\x00\xBB\x25\x01\x00\xBC\x3D"
-     "\x01\x00"
-     "\xBD\x25\x01\x00\xBE\x2D\x01\x00\xBF\x25\x01\x00\xB0\x89\x02\x00\xB1\x89"
-     "\x02\x00"
-     "\xB2\xE9\x02\x00\xB3\xE5\x02\x00\xB4\xFD\x02\x00\xB5\xE1\x02\x00\xB6\xE1"
-     "\x02\x00"
-     "\xB7\xD5\x02\x00\xA8\x6D\x02\x00\xA9\x7D\x02\x00\xAA\x75\x02\x00\xAB\xC9"
-     "\x02\x00"
-     "\xAC\xD9\x02\x00\xAD\xD9\x02\x00\xAE\xC9\x02\x00\xAF\xC1\x02\x00\xC5\xD4"
-     "\x00\x80"
-     "\xC9\xD4\x00\x80\xCD\xD4\x00\x80\xD1\xD4\x00\x80\xD5\xD4\x00\x80\xD9\xD4"
-     "\x00\x80"
-     "\xED\xD2\x00\x80\xDD\xD4\x00\x80\xE1\xD4\x00\x80\xE5\xD4\x00\x80\xE9\xD4"
-     "\x00\x80"
-     "\xED\xD4\x00\x80\xF1\xD4\x00\x80\xF5\xD4\x00\x80\xEF\x48\x00\x00\xEF\xF4"
-     "\x0E\x00"
-     "\xE1\x50\x0E\x00\xE1\x54\x0F\x00\xE3\x9C\x00\x00\xE3\xF4\x0E\x00\xF9\xD4"
-     "\x00\x80"
-     "\xFD\xD4\x00\x80\x01\xD5\x00\x80\x05\xD5\x00\x80\xA3\xFD\x02\x00\x09\xD5"
-     "\x00\x80"
-     "\x0D\xD5\x00\x80\x11\xD5\x00\x80\xBE\x5C\x04\x00\xA6\x59\x02\x00\xA5\x51"
-     "\x02\x00"
-     "\x15\xD5\x00\x80\xAB\x4D\x02\x00\xAA\x4D\x02\x00\x19\xD5\x00\x80\x84\xA0"
-     "\x05\x00"
-     "\xAF\x25\x02\x00\xAE\x39\x02\x00\xAD\x31\x02\x00\xAC\x3D\x02\x00\xAD\xD4"
-     "\x00\x80"
-     "\x82\xED\x07\x00\x81\xD5\x07\x00\x80\xC1\x07\x00\x1D\xD5\x00\x80\x21\xD5"
-     "\x00\x80"
-     "\x86\x38\x03\x00\x87\x5C\x03\x00\xA8\xBD\x06\x00\xA9\xF1\x06\x00\xAA\xF1"
-     "\x06\x00"
-     "\xAB\xCD\x06\x00\xAC\xD5\x06\x00\xAD\x55\x07\x00\xAE\x5D\x07\x00\xAF\x55"
-     "\x07\x00"
-     "\xB0\x2D\x07\x00\xB1\x35\x07\x00\xB2\x3D\x07\x00\xB3\x31\x07\x00\xB4\xD1"
-     "\x07\x00"
-     "\xB5\xD1\x07\x00\xB6\xD1\x07\x00\xB7\xD1\x07\x00\xB8\xF1\x07\x00\xB9\xF1"
-     "\x07\x00"
-     "\xBA\xD9\x07\x00\xBB\xD9\x07\x00\xBC\xC9\x07\x00\xBD\xC9\x07\x00\xBE\xBD"
-     "\x07\x00"
-     "\xBF\xB5\x07\x00\xB3\x59\x06\x00\x25\xD5\x00\x80\x29\xD5\x00\x80\x2D\xD5"
-     "\x00\x80"
-     "\x31\xD5\x00\x80\xB6\xF5\x07\x00\xB5\x7D\x06\x00\x35\xD5\x00\x80\xBB\xD1"
-     "\x07\x00"
-     "\xBA\xC9\x07\x00\x39\xD5\x00\x80\x3D\xD5\x00\x80\xBF\x35\x07\x00\xBE\x35"
-     "\x07\x00"
-     "\xBD\xC5\x07\x00\xBC\xC9\x07\x00\xE3\x40\x04\x00\x41\xD5\x00\x80\xE1\xE0"
-     "\x07\x00"
-     "\x45\xD5\x00\x80\x49\xD5\x00\x80\x4D\xD5\x00\x80\x51\xD5\x00\x80\x55\xD5"
-     "\x00\x80"
-     "\x59\xD5\x00\x80\x5D\xD5\x00\x80\x61\xD5\x00\x80\x65\xD5\x00\x80\xEF\x84"
-     "\x04\x00"
-     "\x69\xD5\x00\x80\x6D\xD5\x00\x80\x71\xD5\x00\x80\xA3\xD9\x07\x00\x75\xD5"
-     "\x00\x80"
-     "\x79\xD5\x00\x80\x7D\xD5\x00\x80\x81\xD5\x00\x80\xA6\x75\x06\x00\xA5\xFD"
-     "\x07\x00"
-     "\x85\xD5\x00\x80\xAB\x51\x06\x00\xAA\x49\x06\x00\x89\xD5\x00\x80\x84\xAC"
-     "\x02\x00"
-     "\xAF\xB5\x06\x00\xAE\xB5\x06\x00\xAD\x45\x06\x00\xAC\x49\x06\x00\x80\x51"
-     "\x00\x00"
-     "\x81\x59\x00\x00\x82\x61\x00\x00\xB3\xC5\x01\x00\xBE\xD0\x01\x00\xB5\xC5"
-     "\x01\x00"
-     "\xB6\xC5\x01\x00\x8D\xD5\x00\x80\x86\xA0\x00\x00\x87\x14\x01\x00\xBA\xAD"
-     "\x01\x00"
-     "\xBB\xA1\x01\x00\xBC\xA1\x01\x00\xBD\xA1\x01\x00\xBE\xA1\x01\x00\xBF\xA1"
-     "\x01\x00"
-     "\xA8\x95\x01\x00\xA9\x01\x01\x00\xAA\x01\x01\x00\xAB\x01\x01\x00\xAC\x01"
-     "\x01\x00"
-     "\xAD\x09\x01\x00\xAE\x31\x01\x00\xAF\x31\x01\x00\x91\xD5\x00\x80\x95\xD5"
-     "\x00\x80"
-     "\x99\xD5\x00\x80\x9D\xD5\x00\x80\xA1\xD5\x00\x80\xA5\xD5\x00\x80\xA9\xD5"
-     "\x00\x80"
-     "\xAD\xD5\x00\x80\xB8\x29\x01\x00\xB9\x29\x01\x00\xBA\x3D\x01\x00\xBB\x35"
-     "\x01\x00"
-     "\xBC\x2D\x01\x00\xBD\xD1\x00\x00\xBE\xD1\x00\x00\xBF\xD1\x00\x00\xB0\x25"
-     "\x01\x00"
-     "\xB1\x29\x01\x00\xB2\x39\x01\x00\xB3\x39\x01\x00\xB4\x29\x01\x00\xB5\x29"
-     "\x01\x00"
-     "\xB6\x19\x01\x00\xB7\x19\x01\x00\x81\x00\x00\x00\xB1\xD5\x00\x80\xB5\xD5"
-     "\x00\x80"
-     "\xA3\x85\x02\x00\xB9\xD5\x00\x80\xA5\x85\x02\x00\xA6\x85\x02\x00\xBD\xD5"
-     "\x00\x80"
-     "\xC1\xD5\x00\x80\xC5\xD5\x00\x80\xAA\xED\x02\x00\xAB\xE1\x02\x00\xAC\xE1"
-     "\x02\x00"
-     "\xAD\xE1\x02\x00\xAE\xE1\x02\x00\xAF\xE1\x02\x00\xC9\xD5\x00\x80\xCD\xD5"
-     "\x00\x80"
-     "\xD1\xD5\x00\x80\x84\x00\x02\x00\xD5\xD5\x00\x80\xD9\xD5\x00\x80\xDD\xD5"
-     "\x00\x80"
-     "\xE1\xD5\x00\x80\xE5\xD5\x00\x80\xE9\xD5\x00\x80\xED\xD5\x00\x80\xF1\xD5"
-     "\x00\x80"
-     "\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xFD\x00\x00\xF5\xD5\x00\x80\xB3\x29"
-     "\x03\x00"
-     "\xF9\xD5\x00\x80\x01\xD6\x00\x80\xBE\xC0\x03\x00\x05\xD6\x00\x80\xB6\x21"
-     "\x03\x00"
-     "\xB5\x29\x03\x00\x09\xD6\x00\x80\xBB\xA9\x00\x00\xBA\xA1\x00\x00\x86\x08"
-     "\x1C\x00"
-     "\x87\x60\x03\x00\xBF\x95\x00\x00\xBE\x95\x00\x00\xBD\xA5\x00\x00\xBC\xAD"
-     "\x00\x00"
-     "\x0D\xD6\x00\x80\x11\xD6\x00\x80\x15\xD6\x00\x80\x19\xD6\x00\x80\xE3\x8C"
-     "\x02\x00"
-     "\x1D\xD6\x00\x80\xE1\xA0\x01\x00\x21\xD6\x00\x80\xEF\x94\x02\x00\x25\xD6"
-     "\x00\x80"
-     "\x29\xD6\x00\x80\x2D\xD6\x00\x80\x31\xD6\x00\x80\x35\xD6\x00\x80\x39\xD6"
-     "\x00\x80"
-     "\x3D\xD6\x00\x80\xB0\x05\x01\x00\xB1\x0D\x01\x00\xB2\x19\x01\x00\xB3\x11"
-     "\x01\x00"
-     "\xB4\x31\x01\x00\xB5\x31\x01\x00\xB6\x59\x01\x00\xB7\x59\x01\x00\xB8\x69"
-     "\x01\x00"
-     "\xB9\x75\x01\x00\xBA\x7D\x01\x00\xBB\x75\x01\x00\xBC\x6D\x01\x00\xBD\x55"
-     "\x01\x00"
-     "\xBE\x5D\x01\x00\xBF\x55\x01\x00\x84\x0C\x1D\x00\x41\xD6\x00\x80\x45\xD6"
-     "\x00\x80"
-     "\x3D\xD4\x00\x80\x49\xD6\x00\x80\x4D\xD6\x00\x80\x51\xD6\x00\x80\x55\xD6"
-     "\x00\x80"
-     "\xA8\xC1\x02\x00\xA9\xC1\x02\x00\xAA\xC1\x02\x00\xAB\xC1\x02\x00\xAC\xC1"
-     "\x02\x00"
-     "\xAD\xC1\x02\x00\xAE\xC1\x02\x00\xAF\xC1\x02\x00\xBE\x0C\x1C\x00\xE1\x34"
-     "\x07\x00"
-     "\x59\xD6\x00\x80\xE3\x70\x06\x00\x5D\xD6\x00\x80\xE1\x94\x01\x00\x61\xD6"
-     "\x00\x80"
-     "\xE3\xA0\x00\x00\x65\xD6\x00\x80\x69\xD6\x00\x80\xBF\xB8\x16\x00\xEF\x04"
-     "\x00\x00"
-     "\x71\xD6\x00\x80\x75\xD6\x00\x80\x79\xD6\x00\x80\xEF\x94\x06\x00\xA3\x69"
-     "\x02\x00"
-     "\x82\x61\x00\x00\x81\x69\x00\x00\x80\x55\x00\x00\x7D\xD6\x00\x80\xA6\x61"
-     "\x02\x00"
-     "\xA5\x69\x02\x00\x81\xD6\x00\x80\xAB\xE9\x01\x00\xAA\xE1\x01\x00\x86\xC8"
-     "\x1C\x00"
-     "\x87\x00\x1D\x00\xAF\xD5\x01\x00\xAE\xD5\x01\x00\xAD\xE5\x01\x00\xAC\xED"
-     "\x01\x00"
-     "\xA8\x89\x1E\x00\xA9\x95\x1E\x00\xAA\x95\x1E\x00\xAB\xCD\x1E\x00\xAC\xD5"
-     "\x1E\x00"
-     "\xAD\xDD\x1E\x00\xAE\xD5\x1E\x00\xAF\xC9\x1E\x00\x6D\xD6\x00\x80\x85\xD6"
-     "\x00\x80"
-     "\x89\xD6\x00\x80\x8D\xD6\x00\x80\x91\xD6\x00\x80\x95\xD6\x00\x80\x99\xD6"
-     "\x00\x80"
-     "\x9D\xD6\x00\x80\xB8\x05\x1F\x00\xB9\x0D\x1F\x00\xBA\x19\x1F\x00\xBB\x19"
-     "\x1F\x00"
-     "\xBC\x09\x1F\x00\xBD\x09\x1F\x00\xBE\xD1\x1F\x00\xBF\xCD\x1F\x00\xB0\xBD"
-     "\x1E\x00"
-     "\xB1\x81\x1E\x00\xB2\x81\x1E\x00\xB3\x81\x1E\x00\xB4\x81\x1E\x00\xB5\x81"
-     "\x1E\x00"
-     "\xB6\x81\x1E\x00\xB7\x81\x1E\x00\xB3\x1D\x1E\x00\xA1\xD6\x00\x80\xA5\xD6"
-     "\x00\x80"
-     "\xA9\xD6\x00\x80\xAD\xD6\x00\x80\xB6\x5D\x1E\x00\xB5\x5D\x1E\x00\xB1\xD6"
-     "\x00\x80"
-     "\xBB\x45\x1E\x00\xBA\x79\x1E\x00\xB5\xD6\x00\x80\xB9\xD6\x00\x80\xBF\x45"
-     "\x1E\x00"
-     "\xBE\x45\x1E\x00\xBD\x55\x1E\x00\xBC\x55\x1E\x00\xBD\xD6\x00\x80\xC1\xD6"
-     "\x00\x80"
-     "\xC5\xD6\x00\x80\xC9\xD6\x00\x80\xE3\x80\x1D\x00\xCD\xD6\x00\x80\xE1\x68"
-     "\x1C\x00"
-     "\xD1\xD6\x00\x80\xEF\x98\x1D\x00\x82\x11\x00\x00\x81\x19\x00\x00\x80\x31"
-     "\x00\x00"
-     "\xD5\xD6\x00\x80\xD9\xD6\x00\x80\x84\x70\x03\x00\xBE\x7C\x00\x00\xA3\x9D"
-     "\x1F\x00"
-     "\xDD\xD6\x00\x80\x86\xC8\x00\x00\x87\x4C\x01\x00\xE1\xD6\x00\x80\xA6\xDD"
-     "\x1F\x00"
-     "\xA5\xDD\x1F\x00\xE5\xD6\x00\x80\xAB\xC5\x1F\x00\xAA\xF9\x1F\x00\xE9\xD6"
-     "\x00\x80"
-     "\xED\xD6\x00\x80\xAF\xC5\x1F\x00\xAE\xC5\x1F\x00\xAD\xD5\x1F\x00\xAC\xD5"
-     "\x1F\x00"
-     "\xF1\xD6\x00\x80\xB3\xFD\x1E\x00\xF5\xD6\x00\x80\xF9\xD6\x00\x80\xB6\x75"
-     "\x01\x00"
-     "\xFD\xD6\x00\x80\x01\xD7\x00\x80\xB5\x81\x1E\x00\xBA\x51\x01\x00\xBB\x51"
-     "\x01\x00"
-     "\x05\xD7\x00\x80\x09\xD7\x00\x80\xBE\xF5\x01\x00\xBF\xFD\x01\x00\xBC\x41"
-     "\x01\x00"
-     "\xBD\x41\x01\x00\xA8\x31\x1E\x00\xA9\x31\x1E\x00\xAA\x31\x1E\x00\xAB\x5D"
-     "\x1E\x00"
-     "\xAC\x4D\x1E\x00\xAD\xB1\x01\x00\xAE\xB1\x01\x00\xAF\xB1\x01\x00\x0D\xD7"
-     "\x00\x80"
-     "\x11\xD7\x00\x80\x15\xD7\x00\x80\x19\xD7\x00\x80\x1D\xD7\x00\x80\x21\xD7"
-     "\x00\x80"
-     "\x25\xD7\x00\x80\x29\xD7\x00\x80\xB8\x5D\x01\x00\xB9\x6D\x01\x00\xBA\x65"
-     "\x01\x00"
-     "\xBB\x0D\x01\x00\xBC\x15\x01\x00\xBD\x1D\x01\x00\xBE\x15\x01\x00\xBF\x09"
-     "\x01\x00"
-     "\xB0\xD1\x01\x00\xB1\xD1\x01\x00\xB2\xD1\x01\x00\xB3\xD1\x01\x00\xB4\x75"
-     "\x01\x00"
-     "\xB5\x7D\x01\x00\xB6\x75\x01\x00\xB7\x6D\x01\x00\x2D\xD7\x00\x80\x84\x08"
-     "\x02\x00"
-     "\x31\xD7\x00\x80\xA3\xB5\x1D\x00\x35\xD7\x00\x80\xA5\xC9\x1D\x00\xA6\x3D"
-     "\x02\x00"
-     "\x39\xD7\x00\x80\x3D\xD7\x00\x80\x41\xD7\x00\x80\xAA\x19\x02\x00\xAB\x19"
-     "\x02\x00"
-     "\xAC\x09\x02\x00\xAD\x09\x02\x00\xAE\xBD\x02\x00\xAF\xB5\x02\x00\x80\x65"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x1D\x00\x00\xBE\x08\x03\x00\x45\xD7\x00\x80\x49\xD7"
-     "\x00\x80"
-     "\x87\x08\x03\x00\x86\xFC\x04\x00\x51\xD7\x00\x80\x55\xD7\x00\x80\x59\xD7"
-     "\x00\x80"
-     "\x5D\xD7\x00\x80\x61\xD7\x00\x80\x65\xD7\x00\x80\x69\xD7\x00\x80\x6D\xD7"
-     "\x00\x80"
-     "\xE1\x90\x01\x00\x71\xD7\x00\x80\xE3\x9C\x02\x00\x75\xD7\x00\x80\x79\xD7"
-     "\x00\x80"
-     "\x7D\xD7\x00\x80\x81\xD7\x00\x80\x85\xD7\x00\x80\x84\xAC\x04\x00\x89\xD7"
-     "\x00\x80"
-     "\x8D\xD7\x00\x80\x91\xD7\x00\x80\x95\xD7\x00\x80\x99\xD7\x00\x80\xEF\x0C"
-     "\x02\x00"
-     "\x9D\xD7\x00\x80\xB3\xE9\x03\x00\xA1\xD7\x00\x80\xA5\xD7\x00\x80\xA9\xD7"
-     "\x00\x80"
-     "\xAD\xD7\x00\x80\xB6\x9D\x03\x00\xB5\x99\x03\x00\xB1\xD7\x00\x80\xBB\x89"
-     "\x03\x00"
-     "\xBA\xBD\x03\x00\xBE\x64\x05\x00\x84\x64\x05\x00\xBF\x59\x03\x00\xBE\x51"
-     "\x03\x00"
-     "\xBD\x59\x03\x00\xBC\x91\x03\x00\xB0\x59\x02\x00\xB1\x6D\x02\x00\xB2\x65"
-     "\x02\x00"
-     "\xB3\xAD\x02\x00\xB4\xB5\x02\x00\xB5\xBD\x02\x00\xB6\xB5\x02\x00\xB7\xAD"
-     "\x02\x00"
-     "\xB8\x95\x02\x00\xB9\x9D\x02\x00\xBA\x95\x02\x00\xBB\x2D\x01\x00\xBC\x39"
-     "\x01\x00"
-     "\xBD\x25\x01\x00\xBE\x25\x01\x00\xBF\x1D\x01\x00\xB5\xD7\x00\x80\xB9\xD7"
-     "\x00\x80"
-     "\xC1\xD7\x00\x80\xC5\xD7\x00\x80\x80\x8D\x01\x00\x81\x15\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\xFD\xD5\x00\x80\xA8\x71\x02\x00\xA9\x79\x02\x00\xAA\x41\x02\x00\xAB\x41"
-     "\x02\x00"
-     "\xAC\x31\x02\x00\xAD\x3D\x02\x00\xAE\x35\x02\x00\xAF\x29\x02\x00\x86\xA4"
-     "\x04\x00"
-     "\x87\x04\x05\x00\xBE\x00\x07\x00\x84\x08\x07\x00\xC9\xD7\x00\x80\xCD\xD7"
-     "\x00\x80"
-     "\xD1\xD7\x00\x80\xEF\x98\x01\x00\xD5\xD7\x00\x80\xE1\x54\x1E\x00\xD9\xD7"
-     "\x00\x80"
-     "\xE3\x44\x01\x00\xDD\xD7\x00\x80\xE1\xD7\x00\x80\xE5\xD7\x00\x80\xE9\xD7"
-     "\x00\x80"
-     "\xED\xD7\x00\x80\xF1\xD7\x00\x80\xA3\xE1\x02\x00\xF5\xD7\x00\x80\xA5\x91"
-     "\x02\x00"
-     "\xF9\xD7\x00\x80\xFD\xD7\x00\x80\xA6\x95\x02\x00\x01\xD8\x00\x80\x05\xD8"
-     "\x00\x80"
-     "\xAB\x81\x02\x00\xAA\xB5\x02\x00\xAD\x51\x02\x00\xAC\x99\x02\x00\xAF\x51"
-     "\x02\x00"
-     "\xAE\x59\x02\x00\xA8\x61\x06\x00\xA9\x61\x06\x00\xAA\x61\x06\x00\xAB\x61"
-     "\x06\x00"
-     "\xAC\x61\x06\x00\xAD\x61\x06\x00\xAE\x61\x06\x00\xAF\x61\x06\x00\xBD\xD7"
-     "\x00\x80"
-     "\x09\xD8\x00\x80\x0D\xD8\x00\x80\x11\xD8\x00\x80\x15\xD8\x00\x80\x19\xD8"
-     "\x00\x80"
-     "\x1D\xD8\x00\x80\x21\xD8\x00\x80\xB8\xED\x06\x00\xB9\xF5\x06\x00\xBA\xFD"
-     "\x06\x00"
-     "\xBB\xF1\x06\x00\xBC\x95\x06\x00\xBD\x9D\x06\x00\xBE\x95\x06\x00\xBF\x8D"
-     "\x06\x00"
-     "\xB0\xE5\x06\x00\xB1\xE9\x06\x00\xB2\xF9\x06\x00\xB3\xF9\x06\x00\xB4\xE9"
-     "\x06\x00"
-     "\xB5\xE9\x06\x00\xB6\xDD\x06\x00\xB7\xD5\x06\x00\x80\xBD\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x25\xD8\x00\x80\x29\xD8\x00\x80\x2D\xD8\x00\x80\x31\xD8"
-     "\x00\x80"
-     "\xEF\x8C\x04\x00\x35\xD8\x00\x80\xE1\xE0\x04\x00\x39\xD8\x00\x80\xE3\x1C"
-     "\x04\x00"
-     "\x86\xB0\x00\x00\x87\x14\x03\x00\x3D\xD8\x00\x80\x41\xD8\x00\x80\xB3\xA9"
-     "\x06\x00"
-     "\x45\xD8\x00\x80\x49\xD8\x00\x80\x4D\xD8\x00\x80\x51\xD8\x00\x80\xB6\xA1"
-     "\x06\x00"
-     "\xB5\xA9\x06\x00\x55\xD8\x00\x80\xBB\x6D\x06\x00\xBA\x61\x06\x00\x59\xD8"
-     "\x00\x80"
-     "\x5D\xD8\x00\x80\xBF\x6D\x06\x00\xBE\x6D\x06\x00\xBD\x6D\x06\x00\xBC\x6D"
-     "\x06\x00"
-     "\xA3\x69\x07\x00\x61\xD8\x00\x80\x65\xD8\x00\x80\x69\xD8\x00\x80\x6D\xD8"
-     "\x00\x80"
-     "\xA6\x61\x07\x00\xA5\x69\x07\x00\x71\xD8\x00\x80\xAB\xAD\x07\x00\xAA\xA1"
-     "\x07\x00"
-     "\x75\xD8\x00\x80\x79\xD8\x00\x80\xAF\xAD\x07\x00\xAE\xAD\x07\x00\xAD\xAD"
-     "\x07\x00"
-     "\xAC\xAD\x07\x00\x7D\xD8\x00\x80\xB3\xD5\x06\x00\x81\xD8\x00\x80\x85\xD8"
-     "\x00\x80"
-     "\xB6\xF1\x06\x00\x89\xD8\x00\x80\x8D\xD8\x00\x80\xB5\xC5\x06\x00\xBA\xD5"
-     "\x06\x00"
-     "\xBB\xDD\x06\x00\x91\xD8\x00\x80\x95\xD8\x00\x80\xBE\xB5\x06\x00\xBF\xB9"
-     "\x06\x00"
-     "\xBC\xC5\x06\x00\xBD\xBD\x06\x00\xA9\xCD\x06\x00\xA8\xCD\x06\x00\xAB\x1D"
-     "\x06\x00"
-     "\xAA\x1D\x06\x00\xAD\x31\x06\x00\xAC\x0D\x06\x00\xAF\x29\x06\x00\xAE\x35"
-     "\x06\x00"
-     "\x80\xC9\x00\x00\x81\xC9\x00\x00\x82\x05\x00\x00\x99\xD8\x00\x80\xBE\x7C"
-     "\x00\x00"
-     "\xA1\xD8\x00\x80\x87\x88\x00\x00\x86\x1C\x0F\x00\xB9\x25\x06\x00\xB8\x25"
-     "\x06\x00"
-     "\xBB\xE1\x01\x00\xBA\xE9\x01\x00\xBD\xE1\x01\x00\xBC\xF9\x01\x00\xBF\xE1"
-     "\x01\x00"
-     "\xBE\xE9\x01\x00\xB1\x59\x06\x00\xB0\x51\x06\x00\xB3\x29\x06\x00\xB2\x51"
-     "\x06\x00"
-     "\xB5\x3D\x06\x00\xB4\x31\x06\x00\xB7\x21\x06\x00\xB6\x29\x06\x00\xA5\xD8"
-     "\x00\x80"
-     "\xA3\x95\x05\x00\xA9\xD8\x00\x80\x84\x0C\x03\x00\xA6\xB1\x05\x00\xAD\xD8"
-     "\x00\x80"
-     "\xB1\xD8\x00\x80\xA5\x85\x05\x00\xAA\x95\x05\x00\xAB\x9D\x05\x00\xB5\xD8"
-     "\x00\x80"
-     "\xB9\xD8\x00\x80\xAE\xF5\x05\x00\xAF\xF9\x05\x00\xAC\x85\x05\x00\xAD\xFD"
-     "\x05\x00"
-     "\xBD\xD8\x00\x80\xC1\xD8\x00\x80\xB3\xE1\x02\x00\xC5\xD8\x00\x80\xB5\xED"
-     "\x02\x00"
-     "\xC9\xD8\x00\x80\xCD\xD8\x00\x80\xB6\x99\x02\x00\xD1\xD8\x00\x80\xD5\xD8"
-     "\x00\x80"
-     "\xBB\xB5\x02\x00\xBA\xBD\x02\x00\xBD\x9D\x02\x00\xBC\x9D\x02\x00\xBF\x71"
-     "\x03\x00"
-     "\xBE\x79\x03\x00\xBE\xCC\x0C\x00\xD9\xD8\x00\x80\xDD\xD8\x00\x80\xE1\xD8"
-     "\x00\x80"
-     "\xE5\xD8\x00\x80\xE9\xD8\x00\x80\xED\xD8\x00\x80\xF1\xD8\x00\x80\xA8\x39"
-     "\x02\x00"
-     "\xA9\x05\x02\x00\xAA\x01\x02\x00\xAB\x15\x02\x00\xAC\x0D\x02\x00\xAD\x71"
-     "\x02\x00"
-     "\xAE\x71\x02\x00\xAF\x71\x02\x00\xB0\x11\x02\x00\xB1\x19\x02\x00\xB2\x21"
-     "\x02\x00"
-     "\xB3\x21\x02\x00\xB4\xD5\x03\x00\xB5\xDD\x03\x00\xB6\xD5\x03\x00\xB7\xCD"
-     "\x03\x00"
-     "\xB8\xFD\x03\x00\xB9\xCD\x03\x00\xBA\xC5\x03\x00\xBB\xAD\x03\x00\xBC\xB9"
-     "\x03\x00"
-     "\xBD\xB9\x03\x00\xBE\xAD\x03\x00\xBF\xA5\x03\x00\xF5\xD8\x00\x80\xF9\xD8"
-     "\x00\x80"
-     "\xFD\xD8\x00\x80\x01\xD9\x00\x80\xE1\xB8\x01\x00\x05\xD9\x00\x80\xE3\x84"
-     "\x06\x00"
-     "\x09\xD9\x00\x80\x4D\xD7\x00\x80\x84\x68\x0C\x00\xEF\x60\x00\x00\x11\xD9"
-     "\x00\x80"
-     "\x80\x3D\x00\x00\x81\x21\x00\x00\x82\x25\x00\x00\x15\xD9\x00\x80\xA3\xE9"
-     "\x03\x00"
-     "\x19\xD9\x00\x80\x86\xE8\x0C\x00\x87\x60\x0D\x00\x1D\xD9\x00\x80\xA6\x91"
-     "\x03\x00"
-     "\xA5\xE5\x03\x00\x21\xD9\x00\x80\xAB\xBD\x03\x00\xAA\xB5\x03\x00\x25\xD9"
-     "\x00\x80"
-     "\x29\xD9\x00\x80\xAF\x79\x02\x00\xAE\x71\x02\x00\xAD\x95\x03\x00\xAC\x95"
-     "\x03\x00"
-     "\x2D\xD9\x00\x80\x31\xD9\x00\x80\xBE\xC0\x0C\x00\x85\xC8\x0F\x00\x35\xD9"
-     "\x00\x80"
-     "\x39\xD9\x00\x80\x3D\xD9\x00\x80\xEF\xEC\x06\x00\x84\x60\x0F\x00\xE1\x3C"
-     "\x06\x00"
-     "\x41\xD9\x00\x80\xE3\x50\x07\x00\x45\xD9\x00\x80\x49\xD9\x00\x80\x4D\xD9"
-     "\x00\x80"
-     "\x51\xD9\x00\x80\x55\xD9\x00\x80\x59\xD9\x00\x80\x5D\xD9\x00\x80\xB3\xC5"
-     "\x01\x00"
-     "\x61\xD9\x00\x80\xB5\xCD\x01\x00\xB6\xF1\x01\x00\x65\xD9\x00\x80\x69\xD9"
-     "\x00\x80"
-     "\x6D\xD9\x00\x80\xBA\xAD\x01\x00\xBB\xA1\x01\x00\xBC\xA5\x01\x00\xBD\xAD"
-     "\x01\x00"
-     "\xBE\x99\x01\x00\xBF\x99\x01\x00\xA8\xB5\x0D\x00\xA9\x05\x0E\x00\xAA\x01"
-     "\x0E\x00"
-     "\xAB\x19\x0E\x00\xAC\x09\x0E\x00\xAD\x3D\x0E\x00\xAE\x35\x0E\x00\xAF\x5D"
-     "\x0E\x00"
-     "\x0D\xD9\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\x71\xD9"
-     "\x00\x80"
-     "\x75\xD9\x00\x80\x79\xD9\x00\x80\xBE\x70\x01\x00\xB8\xCD\x0E\x00\xB9\xD5"
-     "\x0E\x00"
-     "\xBA\xDD\x0E\x00\xBB\xD5\x0E\x00\xBC\xCD\x0E\x00\xBD\x55\x0F\x00\xBE\x5D"
-     "\x0F\x00"
-     "\xBF\x55\x0F\x00\xB0\x25\x0E\x00\xB1\x29\x0E\x00\xB2\x3D\x0E\x00\xB3\x35"
-     "\x0E\x00"
-     "\xB4\x11\x0E\x00\xB5\x11\x0E\x00\xB6\xFD\x0E\x00\xB7\xF5\x0E\x00\x81\xD9"
-     "\x00\x80"
-     "\xE1\x1C\x0D\x00\xEF\xF0\x0C\x00\xE3\x14\xE9\x03\x86\x00\x04\x00\x87\x5C"
-     "\x03\x00"
-     "\x85\xD9\x00\x80\x89\xD9\x00\x80\x8D\xD9\x00\x80\x91\xD9\x00\x80\x95\xD9"
-     "\x00\x80"
-     "\x99\xD9\x00\x80\xE1\x14\x0F\x00\x9D\xD9\x00\x80\xE3\xA4\x0C\x00\xEF\xE4"
-     "\xE9\x03"
-     "\xA3\xC9\x0E\x00\xA1\xD9\x00\x80\xA5\xD9\x00\x80\xA9\xD9\x00\x80\xAD\xD9"
-     "\x00\x80"
-     "\xA6\xFD\x0E\x00\xA5\xC1\x0E\x00\xB1\xD9\x00\x80\xAB\xAD\x0E\x00\xAA\xA1"
-     "\x0E\x00"
-     "\x7D\xD9\x00\x80\xB5\xD9\x00\x80\xAF\x95\x0E\x00\xAE\x95\x0E\x00\xAD\xA1"
-     "\x0E\x00"
-     "\xAC\xA9\x0E\x00\xA8\x2D\x0E\x00\xA9\x39\x0E\x00\xAA\x4D\x0E\x00\xAB\x5D"
-     "\x0E\x00"
-     "\xAC\x4D\x0E\x00\xAD\x7D\x0E\x00\xAE\x75\x0E\x00\xAF\x9D\x0E\x00\xB9\xD9"
-     "\x00\x80"
-     "\xBD\xD9\x00\x80\xC1\xD9\x00\x80\xC5\xD9\x00\x80\xC9\xD9\x00\x80\xCD\xD9"
-     "\x00\x80"
-     "\xD1\xD9\x00\x80\xD5\xD9\x00\x80\xB8\xFD\x0E\x00\xB9\x85\x0E\x00\xBA\x8D"
-     "\x0E\x00"
-     "\xBB\x85\x0E\x00\xBC\x81\x0E\x00\xBD\x89\x0E\x00\xBE\xB9\x0E\x00\xBF\xB9"
-     "\x0E\x00"
-     "\xB0\xE5\x0E\x00\xB1\xF5\x0E\x00\xB2\xFD\x0E\x00\xB3\xF5\x0E\x00\xB4\xD9"
-     "\x0E\x00"
-     "\xB5\xC5\x0E\x00\xB6\xCD\x0E\x00\xB7\xC5\x0E\x00\xB3\xC5\x0E\x00\xD9\xD9"
-     "\x00\x80"
-     "\xDD\xD9\x00\x80\xE1\xD9\x00\x80\xE5\xD9\x00\x80\xB6\xFD\x0E\x00\xB5\xC5"
-     "\x0E\x00"
-     "\xE9\xD9\x00\x80\xBB\x31\x0E\x00\xBA\x21\x0E\x00\x86\x28\x07\x00\x87\x1C"
-     "\x00\x00"
-     "\xBF\x01\x0E\x00\xBE\x1D\x0E\x00\xBD\x21\x0E\x00\xBC\x29\x0E\x00\x82\x5D"
-     "\x00\x00"
-     "\xA3\x81\x0E\x00\x80\x45\x00\x00\x81\x5D\x00\x00\xA6\xB9\x0E\x00\xF1\xD9"
-     "\x00\x80"
-     "\xF5\xD9\x00\x80\xA5\x81\x0E\x00\xAA\x65\x0E\x00\xAB\x75\x0E\x00\xF9\xD9"
-     "\x00\x80"
-     "\xFD\xD9\x00\x80\xAE\x59\x0E\x00\xAF\x45\x0E\x00\xAC\x6D\x0E\x00\xAD\x65"
-     "\x0E\x00"
-     "\x8F\x11\x27\x00\x01\xDA\x00\x80\x05\xDA\x00\x80\xB3\x91\x02\x00\x09\xDA"
-     "\x00\x80"
-     "\xB5\xB1\x02\x00\xB6\xB1\x02\x00\x0D\xDA\x00\x80\x11\xDA\x00\x80\x15\xDA"
-     "\x00\x80"
-     "\xBA\xED\x02\x00\xBB\xE5\x02\x00\xBC\xFD\x02\x00\xBD\xE5\x02\x00\xBE\xE5"
-     "\x02\x00"
-     "\xBF\xD1\x02\x00\x9F\x61\x0D\x00\x9E\x41\x0B\x00\x9D\x69\x0A\x00\x9C\x11"
-     "\x09\x00"
-     "\x9B\x0D\x17\x00\x9A\x45\x16\x00\x99\x59\x16\x00\x98\x11\x15\x00\x97\x45"
-     "\x12\x00"
-     "\x96\x45\x12\x00\x95\x11\x11\x00\x94\xFD\x1E\x00\x93\x51\x1E\x00\x92\x01"
-     "\x1D\x00"
-     "\x91\x55\x1A\x00\x90\x5D\x1A\x00\xB1\x01\xE8\x03\xB0\xB9\xF6\x03\xB3\x99"
-     "\xEA\x03"
-     "\xB2\x89\xEA\x03\xB5\x99\xEE\x03\xB4\xFD\xEB\x03\x19\xDA\x00\x80\x1D\xDA"
-     "\x00\x80"
-     "\x84\xEC\x04\x00\x21\xDA\x00\x80\x25\xDA\x00\x80\x29\xDA\x00\x80\xA3\x91"
-     "\x02\x00"
-     "\xA2\x81\x01\x00\xA1\x2D\x0E\x00\xA0\x9D\x0E\x00\xA1\x29\x02\x00\x84\x88"
-     "\x05\x00"
-     "\xA3\xCD\x06\x00\xA2\xCD\x06\x00\xA5\x01\xF8\x03\xA4\xF9\x07\x00\xA7\xFD"
-     "\xFB\x03"
-     "\xA6\xF1\xFA\x03\xA9\xC1\xFE\x03\xA8\x01\xFC\x03\xAB\x01\xF0\x03\xAA\x39"
-     "\xFE\x03"
-     "\xAD\x61\xF2\x03\xAC\xD5\xF2\x03\xAF\xD9\xF6\x03\xAE\x01\xF4\x03\x84\x2D"
-     "\x1B\x00"
-     "\x85\x71\x1C\x00\x86\x29\x1F\x00\x87\x9D\x1F\x00\x2D\xDA\x00\x80\x31\xDA"
-     "\x00\x80"
-     "\x82\x09\x26\x00\x83\x3D\x1B\x00\x8C\x05\x13\x00\x8D\x69\x14\x00\x8E\x29"
-     "\x17\x00"
-     "\x8F\x3D\x17\x00\x35\xDA\x00\x80\x9D\xD8\x00\x80\x8A\x41\x10\x00\x8B\x1D"
-     "\x13\x00"
-     "\x94\x61\x0B\x00\x95\x41\x0A\x00\x96\x69\x0C\x00\x97\x6D\x0F\x00\x39\xDA"
-     "\x00\x80"
-     "\xEF\x90\x01\x00\x92\x05\x16\x00\x93\x01\x08\x00\x9C\x41\x03\x00\xE3\xD4"
-     "\x0E\x00"
-     "\x3D\xDA\x00\x80\xE1\xE4\x01\x00\x41\xDA\x00\x80\x45\xDA\x00\x80\x9A\xED"
-     "\x0F\x00"
-     "\x9B\x49\x00\x00\x49\xDA\x00\x80\x4D\xDA\x00\x80\x51\xDA\x00\x80\x55\xDA"
-     "\x00\x80"
-     "\x59\xDA\x00\x80\xBE\xDC\x05\x00\x61\xDA\x00\x80\xEF\xEC\x0E\x00\xBE\xE8"
-     "\x04\x00"
-     "\xE1\xF4\x0E\x00\x65\xDA\x00\x80\xE3\x98\x0E\x00\x80\x39\x00\x00\x81\x39"
-     "\x00\x00"
-     "\x82\x29\x00\x00\x69\xDA\x00\x80\xA3\x5D\x03\x00\x6D\xDA\x00\x80\x86\x68"
-     "\x04\x00"
-     "\x87\xEC\x04\x00\x71\xDA\x00\x80\xA6\x7D\x03\x00\xA5\x7D\x03\x00\x75\xDA"
-     "\x00\x80"
-     "\xAB\x29\x03\x00\xAA\x21\x03\x00\x79\xDA\x00\x80\x7D\xDA\x00\x80\xAF\x1D"
-     "\x03\x00"
-     "\xAE\x29\x03\x00\xAD\x29\x03\x00\xAC\x31\x03\x00\xED\xD9\x00\x80\x5D\xDA"
-     "\x00\x80"
-     "\x81\xDA\x00\x80\x85\xDA\x00\x80\x89\xDA\x00\x80\x8D\xDA\x00\x80\x91\xDA"
-     "\x00\x80"
-     "\x95\xDA\x00\x80\xA8\xF5\x05\x00\xA9\xFD\x05\x00\xAA\xC1\x05\x00\xAB\xC1"
-     "\x05\x00"
-     "\xAC\x35\x06\x00\xAD\x3D\x06\x00\xAE\x29\x06\x00\xAF\x21\x06\x00\xB0\x61"
-     "\x06\x00"
-     "\xB1\x61\x06\x00\xB2\xD1\x06\x00\xB3\xD1\x06\x00\xB4\xF5\x06\x00\xB5\xFD"
-     "\x06\x00"
-     "\xB6\xF5\x06\x00\xB7\x4D\x07\x00\xB8\x75\x07\x00\xB9\x55\x07\x00\xBA\x39"
-     "\x07\x00"
-     "\xBB\x05\x07\x00\xBC\x01\x07\x00\xBD\x01\x07\x00\xBE\xF9\x07\x00\xBF\xF9"
-     "\x07\x00"
-     "\xB3\x69\x05\x00\x99\xDA\x00\x80\x9D\xDA\x00\x80\xA1\xDA\x00\x80\xBE\xDC"
-     "\x01\x00"
-     "\xB6\x9D\x06\x00\xB5\x9D\x06\x00\xA5\xDA\x00\x80\xBB\x85\x06\x00\xBA\x85"
-     "\x06\x00"
-     "\xA9\xDA\x00\x80\xAD\xDA\x00\x80\xBF\x19\x06\x00\xBE\x11\x06\x00\xBD\x25"
-     "\x06\x00"
-     "\xBC\x25\x06\x00\xB1\xDA\x00\x80\xA3\x2D\x05\x00\xB5\xDA\x00\x80\xB9\xDA"
-     "\x00\x80"
-     "\xA6\xD9\x06\x00\xBD\xDA\x00\x80\xC1\xDA\x00\x80\xA5\xD9\x06\x00\xAA\xC1"
-     "\x06\x00"
-     "\xAB\xC1\x06\x00\xC5\xDA\x00\x80\xC9\xDA\x00\x80\xAE\x55\x06\x00\xAF\x5D"
-     "\x06\x00"
-     "\xAC\x61\x06\x00\xAD\x61\x06\x00\xA8\x99\x06\x00\xA9\xAD\x06\x00\xAA\xB5"
-     "\x06\x00"
-     "\xAB\xC9\x06\x00\xAC\xC1\x06\x00\xAD\xC9\x06\x00\xAE\xF1\x06\x00\xAF\xF1"
-     "\x06\x00"
-     "\x80\xF9\x01\x00\x81\xCD\x01\x00\x82\xC5\x01\x00\xCD\xDA\x00\x80\x86\x90"
-     "\x00\x00"
-     "\x87\xB4\x01\x00\xD1\xDA\x00\x80\xD5\xDA\x00\x80\xB8\xCD\x01\x00\xB9\xD9"
-     "\x01\x00"
-     "\xBA\xE1\x01\x00\xBB\xE1\x01\x00\xBC\x91\x01\x00\xBD\x9D\x01\x00\xBE\x95"
-     "\x01\x00"
-     "\xBF\x89\x01\x00\xB0\x69\x01\x00\xB1\x69\x01\x00\xB2\x79\x01\x00\xB3\x79"
-     "\x01\x00"
-     "\xB4\x69\x01\x00\xB5\x69\x01\x00\xB6\xFD\x01\x00\xB7\xF5\x01\x00\xB3\x3D"
-     "\x06\x00"
-     "\xD9\xDA\x00\x80\xDD\xDA\x00\x80\xE1\xDA\x00\x80\xE5\xDA\x00\x80\xB6\x49"
-     "\x06\x00"
-     "\xB5\x49\x06\x00\xE9\xDA\x00\x80\xBB\xAD\x01\x00\xBA\xAD\x01\x00\xED\xDA"
-     "\x00\x80"
-     "\xF1\xDA\x00\x80\xBF\xAD\x01\x00\xBE\xAD\x01\x00\xBD\xAD\x01\x00\xBC\xAD"
-     "\x01\x00"
-     "\xF5\xDA\x00\x80\xA3\x79\x06\x00\xF9\xDA\x00\x80\xFD\xDA\x00\x80\xA6\x0D"
-     "\x06\x00"
-     "\x01\xDB\x00\x80\x05\xDB\x00\x80\xA5\x0D\x06\x00\xAA\xE9\x01\x00\xAB\xE9"
-     "\x01\x00"
-     "\x09\xDB\x00\x80\x0D\xDB\x00\x80\xAE\xE9\x01\x00\xAF\xE9\x01\x00\xAC\xE9"
-     "\x01\x00"
-     "\xAD\xE9\x01\x00\x11\xDB\x00\x80\x15\xDB\x00\x80\x19\xDB\x00\x80\x1D\xDB"
-     "\x00\x80"
-     "\x21\xDB\x00\x80\x25\xDB\x00\x80\x29\xDB\x00\x80\x2D\xDB\x00\x80\xAC\x85"
-     "\x06\x00"
-     "\xAD\x8D\x06\x00\xAE\x85\x06\x00\xAF\x99\x06\x00\xA8\xB1\x06\x00\xA9\xAD"
-     "\x06\x00"
-     "\xAA\xA5\x06\x00\xAB\x9D\x06\x00\xB4\x9D\x06\x00\xB5\xA9\x06\x00\xB6\x69"
-     "\x07\x00"
-     "\xB7\x69\x07\x00\xB0\x95\x06\x00\xB1\x9D\x06\x00\xB2\x95\x06\x00\xB3\x8D"
-     "\x06\x00"
-     "\xBC\x59\x07\x00\xBD\x59\x07\x00\xBE\xED\x07\x00\xBF\xE5\x07\x00\xB8\x59"
-     "\x07\x00"
-     "\xB9\x59\x07\x00\xBA\x49\x07\x00\xBB\x49\x07\x00\x80\x0D\x00\x00\x81\x11"
-     "\x00\x00"
-     "\x82\x11\x00\x00\x31\xDB\x00\x80\xEF\xF8\x00\x00\x35\xDB\x00\x80\x39\xDB"
-     "\x00\x80"
-     "\x41\xDB\x00\x80\xE3\x14\x02\x00\x84\x88\x02\x00\xE1\x80\x01\x00\x45\xDB"
-     "\x00\x80"
-     "\x49\xDB\x00\x80\x4D\xDB\x00\x80\x86\x18\xFC\x03\x87\x9C\xFC\x03\xBB\x01"
-     "\x03\x00"
-     "\xBA\x69\x03\x00\x51\xDB\x00\x80\x55\xDB\x00\x80\xBF\x09\x03\x00\xBE\x01"
-     "\x03\x00"
-     "\xBD\x15\x03\x00\xBC\x15\x03\x00\xB3\x65\x03\x00\x59\xDB\x00\x80\x5D\xDB"
-     "\x00\x80"
-     "\x61\xDB\x00\x80\x65\xDB\x00\x80\xB6\x65\x03\x00\xB5\x75\x03\x00\x69\xDB"
-     "\x00\x80"
-     "\xA8\x39\x02\x00\xA9\x39\x02\x00\xAA\x8D\x02\x00\xAB\x85\x02\x00\xAC\x9D"
-     "\x02\x00"
-     "\xAD\x85\x02\x00\xAE\x85\x02\x00\xAF\xB5\x02\x00\x84\x6C\xFD\x03\x6D\xDB"
-     "\x00\x80"
-     "\xBE\x60\xFD\x03\x71\xDB\x00\x80\x75\xDB\x00\x80\x79\xDB\x00\x80\x7D\xDB"
-     "\x00\x80"
-     "\x81\xDB\x00\x80\xB8\x41\x01\x00\xB9\x41\x01\x00\xBA\x41\x01\x00\xBB\x41"
-     "\x01\x00"
-     "\xBC\x41\x01\x00\xBD\x41\x01\x00\xBE\x41\x01\x00\xBF\x41\x01\x00\xB0\xCD"
-     "\x02\x00"
-     "\xB1\xA5\x02\x00\xB2\xA1\x02\x00\xB3\xA1\x02\x00\xB4\xA5\x02\x00\xB5\xAD"
-     "\x02\x00"
-     "\xB6\x91\x02\x00\xB7\x91\x02\x00\xE1\x50\x06\x00\xE1\x34\x04\x00\xE3\x5C"
-     "\x07\x00"
-     "\xE3\x3C\x07\x00\x85\xDB\x00\x80\x89\xDB\x00\x80\x8D\xDB\x00\x80\x91\xDB"
-     "\x00\x80"
-     "\x84\x64\xFF\x03\x95\xDB\x00\x80\x99\xDB\x00\x80\x9D\xDB\x00\x80\xA1\xDB"
-     "\x00\x80"
-     "\xA5\xDB\x00\x80\xEF\xE0\x06\x00\xEF\xD0\x06\x00\xA3\x45\x02\x00\xA9\xDB"
-     "\x00\x80"
-     "\xB1\xDB\x00\x80\xB5\xDB\x00\x80\xB9\xDB\x00\x80\xA6\x45\x02\x00\xA5\x55"
-     "\x02\x00"
-     "\xBD\xDB\x00\x80\xAB\x21\x02\x00\xAA\x49\x02\x00\x86\xC8\xFC\x03\x87\xFC"
-     "\xFC\x03"
-     "\xAF\x29\x02\x00\xAE\x21\x02\x00\xAD\x35\x02\x00\xAC\x35\x02\x00\xA8\x9D"
-     "\xFE\x03"
-     "\xA9\xA9\xFE\x03\xAA\xB9\xFE\x03\xAB\xB9\xFE\x03\xAC\xD9\xFE\x03\xAD\xD9"
-     "\xFE\x03"
-     "\xAE\xFD\xFE\x03\xAF\xF5\xFE\x03\x80\x91\xFF\x03\x81\x91\xFF\x03\x82\xED"
-     "\xFF\x03"
-     "\xAD\xDB\x00\x80\xC1\xDB\x00\x80\xC5\xDB\x00\x80\xC9\xDB\x00\x80\xCD\xDB"
-     "\x00\x80"
-     "\xB8\x7D\xFF\x03\xB9\xD5\xFF\x03\xBA\xDD\xFF\x03\xBB\xD5\xFF\x03\xBC\xCD"
-     "\xFF\x03"
-     "\xBD\xF1\xFF\x03\xBE\xF1\xFF\x03\xBF\xF1\xFF\x03\xB0\x8D\xFE\x03\xB1\x95"
-     "\xFE\x03"
-     "\xB2\x95\xFE\x03\xB3\x49\xFF\x03\xB4\x5D\xFF\x03\xB5\x45\xFF\x03\xB6\x4D"
-     "\xFF\x03"
-     "\xB7\x45\xFF\x03\xB3\x31\xFE\x03\xD1\xDB\x00\x80\xD5\xDB\x00\x80\xD9\xDB"
-     "\x00\x80"
-     "\xDD\xDB\x00\x80\xB6\x59\xFE\x03\xB5\x51\xFE\x03\xE1\xDB\x00\x80\xBB\x51"
-     "\xFE\x03"
-     "\xBA\x49\xFE\x03\xE5\xDB\x00\x80\xE9\xDB\x00\x80\xBF\x81\xFF\x03\xBE\x99"
-     "\xFF\x03"
-     "\xBD\x9D\xFF\x03\xBC\x41\xFE\x03\xED\xDB\x00\x80\xA3\x75\xFE\x03\xF1\xDB"
-     "\x00\x80"
-     "\xF5\xDB\x00\x80\xA6\x1D\xFE\x03\xF9\xDB\x00\x80\xFD\xDB\x00\x80\xA5\x15"
-     "\xFE\x03"
-     "\xAA\x0D\xFE\x03\xAB\x15\xFE\x03\x01\xDC\x00\x80\x05\xDC\x00\x80\xAE\xDD"
-     "\xFF\x03"
-     "\xAF\xC5\xFF\x03\xAC\x05\xFE\x03\xAD\xD9\xFF\x03\x09\xDC\x00\x80\x0D\xDC"
-     "\x00\x80"
-     "\x11\xDC\x00\x80\x15\xDC\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x19\xDC\x00\x80\xBE\xFC\x01\x00\x1D\xDC\x00\x80\x86\x88\x00\x00\x87\x28"
-     "\x01\x00"
-     "\x21\xDC\x00\x80\x25\xDC\x00\x80\x29\xDC\x00\x80\x2D\xDC\x00\x80\x31\xDC"
-     "\x00\x80"
-     "\x35\xDC\x00\x80\x39\xDC\x00\x80\x3D\xDC\x00\x80\x41\xDC\x00\x80\xE1\xF4"
-     "\xFD\x03"
-     "\x45\xDC\x00\x80\xE3\x6C\xFC\x03\x49\xDC\x00\x80\x4D\xDC\x00\x80\x51\xDC"
-     "\x00\x80"
-     "\xEF\xBC\xFC\x03\x55\xDC\x00\x80\x59\xDC\x00\x80\x5D\xDC\x00\x80\x61\xDC"
-     "\x00\x80"
-     "\xA8\x49\xFE\x03\xA9\x49\xFE\x03\xAA\x59\xFE\x03\xAB\x51\xFE\x03\xAC\xA1"
-     "\x01\x00"
-     "\xAD\xA1\x01\x00\xAE\xA1\x01\x00\xAF\xA1\x01\x00\x84\xAC\x01\x00\x65\xDC"
-     "\x00\x80"
-     "\x69\xDC\x00\x80\x6D\xDC\x00\x80\x71\xDC\x00\x80\x75\xDC\x00\x80\x79\xDC"
-     "\x00\x80"
-     "\x7D\xDC\x00\x80\xB8\x09\x01\x00\xB9\x09\x01\x00\xBA\x1D\x01\x00\xBB\x15"
-     "\x01\x00"
-     "\xBC\x39\x01\x00\xBD\x39\x01\x00\xBE\xD9\x00\x00\xBF\xD9\x00\x00\xB0\x65"
-     "\x01\x00"
-     "\xB1\x6D\x01\x00\xB2\x65\x01\x00\xB3\x7D\x01\x00\xB4\x6D\x01\x00\xB5\x5D"
-     "\x01\x00"
-     "\xB6\x55\x01\x00\xB7\x3D\x01\x00\x81\x19\x00\x00\x80\x11\x00\x00\xB3\xD9"
-     "\xFD\x03"
-     "\x82\x29\x00\x00\xB5\x21\x02\x00\x81\xDC\x00\x80\x85\xDC\x00\x80\xB6\x29"
-     "\x02\x00"
-     "\xB5\x4D\x2D\x00\x89\xDC\x00\x80\xBB\xA1\x02\x00\xBA\xA9\x02\x00\xBD\xA1"
-     "\x02\x00"
-     "\xBC\xB9\x02\x00\xBF\x91\x02\x00\xBE\xA1\x02\x00\xBE\x1C\x03\x00\xA6\x71"
-     "\x02\x00"
-     "\xA5\x79\x02\x00\x84\x00\x04\x00\xA3\x81\xFD\x03\x91\xDC\x00\x80\x86\xF8"
-     "\x04\x00"
-     "\x87\xDC\x03\x00\xAF\xC9\x02\x00\xAE\xF9\x02\x00\xAD\xF9\x02\x00\xAC\xE1"
-     "\x02\x00"
-     "\xAB\xF9\x02\x00\xAA\xF1\x02\x00\x95\xDC\x00\x80\x99\xDC\x00\x80\x9D\xDC"
-     "\x00\x80"
-     "\xA1\xDC\x00\x80\xA5\xDC\x00\x80\xA9\xDC\x00\x80\xAD\xDC\x00\x80\xB1\xDC"
-     "\x00\x80"
-     "\xEF\x20\x02\x00\xB5\xDC\x00\x80\xE1\x90\x01\x00\xB9\xDC\x00\x80\xE3\x18"
-     "\x01\x00"
-     "\xBD\xDC\x00\x80\xC1\xDC\x00\x80\xC5\xDC\x00\x80\xC9\xDC\x00\x80\xCD\xDC"
-     "\x00\x80"
-     "\xB3\xF9\x03\x00\xD1\xDC\x00\x80\xD5\xDC\x00\x80\xD9\xDC\x00\x80\x3D\xDB"
-     "\x00\x80"
-     "\xB6\x91\x03\x00\xB5\x99\x03\x00\xDD\xDC\x00\x80\xBB\xB9\x03\x00\xBA\xB5"
-     "\x03\x00"
-     "\xE1\xDC\x00\x80\xE5\xDC\x00\x80\xBF\x81\x03\x00\xBE\x9D\x03\x00\xBD\x9D"
-     "\x03\x00"
-     "\xBC\x9D\x03\x00\xA8\x31\x02\x00\xA9\x31\x02\x00\xAA\x31\x02\x00\xAB\x31"
-     "\x02\x00"
-     "\xAC\x51\x02\x00\xAD\x51\x02\x00\xAE\x51\x02\x00\xAF\x51\x02\x00\xBE\x68"
-     "\x05\x00"
-     "\xE9\xDC\x00\x80\xED\xDC\x00\x80\xF1\xDC\x00\x80\x84\xFC\x04\x00\xF5\xDC"
-     "\x00\x80"
-     "\xF9\xDC\x00\x80\x01\xDD\x00\x80\xB8\x3D\x02\x00\xB9\xE1\x02\x00\xBA\xE1"
-     "\x02\x00"
-     "\xBB\xE1\x02\x00\xBC\xE1\x02\x00\xBD\xE1\x02\x00\xBE\xE1\x02\x00\xBF\xE1"
-     "\x02\x00"
-     "\xB0\x31\x02\x00\xB1\x3D\x02\x00\xB2\x09\x02\x00\xB3\x09\x02\x00\xB4\x19"
-     "\x02\x00"
-     "\xB5\x05\x02\x00\xB6\x0D\x02\x00\xB7\x05\x02\x00\x80\xAD\x00\x00\x81\xB1"
-     "\x00\x00"
-     "\x82\xB1\x00\x00\x05\xDD\x00\x80\xEF\x48\x00\x00\x09\xDD\x00\x80\x87\x50"
-     "\x05\x00"
-     "\x86\xBC\x04\x00\xE3\x10\xFE\x03\x0D\xDD\x00\x80\xE1\x58\xFE\x03\x11\xDD"
-     "\x00\x80"
-     "\xEC\x00\x00\x00\x15\xDD\x00\x80\x19\xDD\x00\x80\x1D\xDD\x00\x80\x21\xDD"
-     "\x00\x80"
-     "\xA6\x81\x02\x00\xA5\x89\x02\x00\x25\xDD\x00\x80\xA3\xE9\x02\x00\x29\xDD"
-     "\x00\x80"
-     "\x84\x40\x07\x00\x2D\xDD\x00\x80\xAF\x91\x02\x00\xAE\x8D\x02\x00\xAD\x8D"
-     "\x02\x00"
-     "\xAC\x8D\x02\x00\xAB\xA9\x02\x00\xAA\xA5\x02\x00\x31\xDD\x00\x80\x35\xDD"
-     "\x00\x80"
-     "\xA8\x55\x06\x00\xA9\x7D\x06\x00\xAA\x75\x06\x00\xAB\x8D\x06\x00\xAC\x95"
-     "\x06\x00"
-     "\xAD\x9D\x06\x00\xAE\x95\x06\x00\xAF\x8D\x06\x00\xFD\xDC\x00\x80\x39\xDD"
-     "\x00\x80"
-     "\x3D\xDD\x00\x80\x41\xDD\x00\x80\x45\xDD\x00\x80\x49\xDD\x00\x80\x4D\xDD"
-     "\x00\x80"
-     "\x51\xDD\x00\x80\xB8\x85\x06\x00\xB9\x8D\x06\x00\xBA\x85\x06\x00\xBB\x91"
-     "\x06\x00"
-     "\xBC\xB9\x06\x00\xBD\xB9\x06\x00\xBE\x7D\x07\x00\xBF\x75\x07\x00\xB0\xF5"
-     "\x06\x00"
-     "\xB1\xF9\x06\x00\xB2\xCD\x06\x00\xB3\xC1\x06\x00\xB4\xC1\x06\x00\xB5\xC1"
-     "\x06\x00"
-     "\xB6\xC1\x06\x00\xB7\xC1\x06\x00\xB3\xFD\x06\x00\x55\xDD\x00\x80\x59\xDD"
-     "\x00\x80"
-     "\x5D\xDD\x00\x80\x61\xDD\x00\x80\xB6\x1D\x06\x00\xB5\x1D\x06\x00\x65\xDD"
-     "\x00\x80"
-     "\xBB\x3D\x06\x00\xBA\x31\x06\x00\x69\xDD\x00\x80\xBE\x2C\x00\x00\xBF\x05"
-     "\x06\x00"
-     "\xBE\x05\x06\x00\xBD\x15\x06\x00\xBC\x19\x06\x00\x82\x69\x00\x00\xA3\xB9"
-     "\x06\x00"
-     "\x80\x59\x00\x00\x81\x51\x00\x00\xA6\x59\x06\x00\x6D\xDD\x00\x80\x71\xDD"
-     "\x00\x80"
-     "\xA5\x59\x06\x00\xAA\x75\x06\x00\xAB\x79\x06\x00\x86\x88\x00\x00\x87\x5C"
-     "\x03\x00"
-     "\xAE\x41\x06\x00\xAF\x41\x06\x00\xAC\x5D\x06\x00\xAD\x51\x06\x00\xB3\x95"
-     "\x06\x00"
-     "\x75\xDD\x00\x80\x79\xDD\x00\x80\x7D\xDD\x00\x80\x81\xDD\x00\x80\xB6\xA1"
-     "\x06\x00"
-     "\xB5\xB5\x06\x00\x85\xDD\x00\x80\xBB\xED\x06\x00\xBA\xE5\x06\x00\x89\xDD"
-     "\x00\x80"
-     "\x8D\xDD\x00\x80\xBF\xCD\x06\x00\xBE\xCD\x06\x00\xBD\xE5\x06\x00\xBC\xE9"
-     "\x06\x00"
-     "\x91\xDD\x00\x80\x95\xDD\x00\x80\x99\xDD\x00\x80\x9D\xDD\x00\x80\xA1\xDD"
-     "\x00\x80"
-     "\xA5\xDD\x00\x80\xEF\x80\x04\x00\xA9\xDD\x00\x80\xE1\x20\x05\x00\xAD\xDD"
-     "\x00\x80"
-     "\xE3\x34\x05\x00\xB1\xDD\x00\x80\xB5\xDD\x00\x80\xB9\xDD\x00\x80\xBD\xDD"
-     "\x00\x80"
-     "\xC1\xDD\x00\x80\xA8\xBD\x06\x00\xA9\x09\x06\x00\xAA\x1D\x06\x00\xAB\x29"
-     "\x06\x00"
-     "\xAC\x3D\x06\x00\xAD\x21\x06\x00\xAE\x21\x06\x00\xAF\x21\x06\x00\xC5\xDD"
-     "\x00\x80"
-     "\xC9\xDD\x00\x80\xCD\xDD\x00\x80\xD1\xDD\x00\x80\x80\x1D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\xB9\x03\x00\xD5\xDD\x00\x80\xB8\xED\x01\x00\xB9\xF5\x01\x00\xBA\xFD"
-     "\x01\x00"
-     "\xBB\xF5\x01\x00\xBC\xED\x01\x00\xBD\x51\x01\x00\xBE\x51\x01\x00\xBF\x51"
-     "\x01\x00"
-     "\xB0\x21\x06\x00\xB1\x29\x06\x00\xB2\x3D\x06\x00\xB3\x31\x06\x00\xB4\x09"
-     "\x06\x00"
-     "\xB5\x09\x06\x00\xB6\xDD\x01\x00\xB7\xD5\x01\x00\xD9\xDD\x00\x80\xBE\x88"
-     "\x03\x00"
-     "\xE1\xDD\x00\x80\xA3\x99\x05\x00\xE5\xDD\x00\x80\xA5\xB9\x05\x00\xA6\xAD"
-     "\x05\x00"
-     "\xE9\xDD\x00\x80\xED\xDD\x00\x80\xF1\xDD\x00\x80\xAA\xE9\x05\x00\xAB\xE1"
-     "\x05\x00"
-     "\xAC\xE5\x05\x00\xAD\xE9\x05\x00\xAE\xC1\x05\x00\xAF\xC1\x05\x00\xB6\x5D"
-     "\x03\x00"
-     "\xF5\xDD\x00\x80\x84\x48\x03\x00\xB5\x5D\x03\x00\xF9\xDD\x00\x80\xB3\x5D"
-     "\x03\x00"
-     "\x86\xF8\x0C\x00\x87\xDC\x03\x00\xBE\x31\x03\x00\xBF\x31\x03\x00\xBC\x21"
-     "\x03\x00"
-     "\xBD\x21\x03\x00\xBA\x31\x03\x00\xBB\x31\x03\x00\xFD\xDD\x00\x80\x01\xDE"
-     "\x00\x80"
-     "\xB8\xED\x03\x00\xB9\xF5\x03\x00\xBA\xFD\x03\x00\xBB\xF5\x03\x00\xBC\xED"
-     "\x03\x00"
-     "\xBD\x35\x03\x00\xBE\x3D\x03\x00\xBF\x35\x03\x00\xB0\xE1\x03\x00\xB1\xE1"
-     "\x03\x00"
-     "\xB2\xE1\x03\x00\xB3\xE1\x03\x00\xB4\xE1\x03\x00\xB5\xE1\x03\x00\xB6\xE5"
-     "\x03\x00"
-     "\xB7\xD9\x03\x00\xA8\x19\x02\x00\xA9\xC1\x03\x00\xAA\xC1\x03\x00\xAB\xC1"
-     "\x03\x00"
-     "\xAC\xC1\x03\x00\xAD\xC9\x03\x00\xAE\xF1\x03\x00\xAF\xF1\x03\x00\x05\xDE"
-     "\x00\x80"
-     "\x8D\xDC\x00\x80\x09\xDE\x00\x80\x0D\xDE\x00\x80\xBE\xFC\x0C\x00\x11\xDE"
-     "\x00\x80"
-     "\x15\xDE\x00\x80\x19\xDE\x00\x80\xA3\x09\x02\x00\x1D\xDE\x00\x80\x21\xDE"
-     "\x00\x80"
-     "\x25\xDE\x00\x80\x29\xDE\x00\x80\xA6\x09\x02\x00\xA5\x09\x02\x00\x2D\xDE"
-     "\x00\x80"
-     "\xAB\x65\x02\x00\xAA\x65\x02\x00\x31\xDE\x00\x80\x35\xDE\x00\x80\xAF\x65"
-     "\x02\x00"
-     "\xAE\x65\x02\x00\xAD\x75\x02\x00\xAC\x75\x02\x00\x39\xDE\x00\x80\x3D\xDE"
-     "\x00\x80"
-     "\x41\xDE\x00\x80\x45\xDE\x00\x80\xE1\xA0\x01\x00\x49\xDE\x00\x80\xE3\x90"
-     "\x06\x00"
-     "\x51\xDE\x00\x80\x55\xDE\x00\x80\x59\xDE\x00\x80\xEF\x94\x03\x00\x5D\xDE"
-     "\x00\x80"
-     "\x86\x50\x0C\x00\x87\x44\x0D\x00\x61\xDE\x00\x80\x65\xDE\x00\x80\x80\xC9"
-     "\x00\x00"
-     "\x81\xC9\x00\x00\x82\xDD\x00\x00\x69\xDE\x00\x80\x6D\xDE\x00\x80\x71\xDE"
-     "\x00\x80"
-     "\x75\xDE\x00\x80\xEF\xC8\x01\x00\x79\xDE\x00\x80\xE1\x20\x06\x00\x7D\xDE"
-     "\x00\x80"
-     "\xE3\x8C\x06\x00\x82\x1C\x35\x00\x81\xDE\x00\x80\x85\xDE\x00\x80\x89\xDE"
-     "\x00\x80"
-     "\x8D\xDE\x00\x80\x91\xDE\x00\x80\x95\xDE\x00\x80\xB3\xD1\x01\x00\x99\xDE"
-     "\x00\x80"
-     "\xB5\xF1\x01\x00\xB6\x9D\x01\x00\x9D\xDE\x00\x80\xA1\xDE\x00\x80\xA5\xDE"
-     "\x00\x80"
-     "\xBA\x95\x01\x00\xBB\x0D\x01\x00\xBC\x19\x01\x00\xBD\x19\x01\x00\xBE\x0D"
-     "\x01\x00"
-     "\xBF\x05\x01\x00\xA8\x09\x0E\x00\xA9\x15\x0E\x00\xAA\x15\x0E\x00\xAB\x2D"
-     "\x0E\x00"
-     "\xAC\x35\x0E\x00\xAD\x51\x0E\x00\xAE\x49\x0E\x00\xAF\x65\x0E\x00\x4D\xDE"
-     "\x00\x80"
-     "\xA9\xDE\x00\x80\xAD\xDE\x00\x80\xB1\xDE\x00\x80\x80\x1D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\xBD\x00\x00\xB5\xDE\x00\x80\xB8\x49\x0F\x00\xB9\x49\x0F\x00\xBA\x59"
-     "\x0F\x00"
-     "\xBB\x51\x0F\x00\xBC\x79\x0F\x00\xBD\x79\x0F\x00\xBE\x19\x0F\x00\xBF\x19"
-     "\x0F\x00"
-     "\xB0\x1D\x0E\x00\xB1\x85\x0E\x00\xB2\x81\x0E\x00\xB3\x81\x0E\x00\xB4\x85"
-     "\x0E\x00"
-     "\xB5\x8D\x0E\x00\xB6\x85\x0E\x00\xB7\x79\x0F\x00\xA3\x9D\x0E\x00\xB9\xDE"
-     "\x00\x80"
-     "\x84\x80\x02\x00\xBE\x00\x03\x00\xBD\xDE\x00\x80\xA6\xD1\x0E\x00\xA5\xBD"
-     "\x0E\x00"
-     "\xC1\xDE\x00\x80\xAB\x41\x0E\x00\xAA\xD9\x0E\x00\x86\x08\x00\x00\x87\xDC"
-     "\x03\x00"
-     "\xAF\x49\x0E\x00\xAE\x41\x0E\x00\xAD\x55\x0E\x00\xAC\x55\x0E\x00\xC5\xDE"
-     "\x00\x80"
-     "\xB3\xA1\x0F\x00\xC9\xDE\x00\x80\xCD\xDE\x00\x80\xB6\xAD\x0F\x00\xD1\xDE"
-     "\x00\x80"
-     "\xD5\xDE\x00\x80\xB5\xA5\x0F\x00\xBA\x75\x0F\x00\xBB\x7D\x0F\x00\xD9\xDE"
-     "\x00\x80"
-     "\xDD\xDE\x00\x80\xBE\x45\x0F\x00\xBF\x4D\x0F\x00\xBC\x59\x0F\x00\xBD\x55"
-     "\x0F\x00"
-     "\xA3\x65\x0E\x00\xE1\xDE\x00\x80\xE5\xDE\x00\x80\xE9\xDE\x00\x80\xED\xDE"
-     "\x00\x80"
-     "\xA6\x69\x0E\x00\xA5\x61\x0E\x00\xF1\xDE\x00\x80\xAB\xB9\x0E\x00\xAA\xB1"
-     "\x0E\x00"
-     "\xF5\xDE\x00\x80\xF9\xDE\x00\x80\xAF\x89\x0E\x00\xAE\x81\x0E\x00\xAD\x91"
-     "\x0E\x00"
-     "\xAC\x9D\x0E\x00\xFD\xDE\x00\x80\x01\xDF\x00\x80\x05\xDF\x00\x80\x09\xDF"
-     "\x00\x80"
-     "\x0D\xDF\x00\x80\xEF\x5C\x0D\x00\x11\xDF\x00\x80\x15\xDF\x00\x80\x19\xDF"
-     "\x00\x80"
-     "\xE3\x3C\x0D\x00\x1D\xDF\x00\x80\xE1\x2C\x0D\x00\x80\x0D\x00\x00\x81\xD5"
-     "\x00\x00"
-     "\x82\xDD\x00\x00\x21\xDF\x00\x80\xA8\xDD\x0E\x00\xA9\xE5\x0E\x00\xAA\xED"
-     "\x0E\x00"
-     "\xAB\xE1\x0E\x00\xAC\xE9\x0E\x00\xAD\xE9\x0E\x00\xAE\x15\x0E\x00\xAF\x0D"
-     "\x0E\x00"
-     "\x25\xDF\x00\x80\x29\xDF\x00\x80\x84\xE4\x01\x00\xBE\xE4\x01\x00\x86\x10"
-     "\x07\x00"
-     "\x87\x94\x00\x00\x31\xDF\x00\x80\x35\xDF\x00\x80\xB8\x05\x0E\x00\xB9\x0D"
-     "\x0E\x00"
-     "\xBA\x19\x0E\x00\xBB\x11\x0E\x00\xBC\x31\x0E\x00\xBD\x31\x0E\x00\xBE\xFD"
-     "\x01\x00"
-     "\xBF\xF5\x01\x00\xB0\x75\x0E\x00\xB1\x7D\x0E\x00\xB2\x49\x0E\x00\xB3\x45"
-     "\x0E\x00"
-     "\xB4\x45\x0E\x00\xB5\x4D\x0E\x00\xB6\x45\x0E\x00\xB7\x3D\x0E\x00\xA8\x9D"
-     "\x02\x00"
-     "\xA9\xC5\x02\x00\xAA\xC1\x02\x00\xAB\xC1\x02\x00\xAC\xC5\x02\x00\xAD\xCD"
-     "\x02\x00"
-     "\xAE\xF1\x02\x00\xAF\xF1\x02\x00\x39\xDF\x00\x80\x3D\xDF\x00\x80\x41\xDF"
-     "\x00\x80"
-     "\x45\xDF\x00\x80\xE1\x28\x03\x00\x49\xDF\x00\x80\x4D\xDF\x00\x80\x51\xDF"
-     "\x00\x80"
-     "\xB8\x05\x03\x00\xB9\x0D\x03\x00\xBA\x05\x03\x00\xBB\x1D\x03\x00\xBC\x09"
-     "\x03\x00"
-     "\xBD\x09\x03\x00\xBE\x39\x03\x00\xBF\x35\x03\x00\xB0\x65\x03\x00\xB1\x6D"
-     "\x03\x00"
-     "\xB2\x65\x03\x00\xB3\x7D\x03\x00\xB4\x6D\x03\x00\xB5\x5D\x03\x00\xB6\x55"
-     "\x03\x00"
-     "\xB7\x3D\x03\x00\xB3\x49\x02\x00\x55\xDF\x00\x80\x59\xDF\x00\x80\x5D\xDF"
-     "\x00\x80"
-     "\x61\xDF\x00\x80\xB6\x4D\x02\x00\xB5\x4D\x02\x00\x65\xDF\x00\x80\xBB\xA9"
-     "\x03\x00"
-     "\xBA\xA1\x03\x00\x69\xDF\x00\x80\x6D\xDF\x00\x80\xBF\x99\x03\x00\xBE\xA9"
-     "\x03\x00"
-     "\xBD\xA9\x03\x00\xBC\xB1\x03\x00\x71\xDF\x00\x80\xA3\x0D\x02\x00\x75\xDF"
-     "\x00\x80"
-     "\x79\xDF\x00\x80\xA6\x09\x02\x00\x7D\xDF\x00\x80\x81\xDF\x00\x80\xA5\x09"
-     "\x02\x00"
-     "\xAA\xE5\x03\x00\xAB\xED\x03\x00\xDD\xDD\x00\x80\x85\xDF\x00\x80\xAE\xED"
-     "\x03\x00"
-     "\xAF\xDD\x03\x00\xAC\xF5\x03\x00\xAD\xED\x03\x00\x89\xDF\x00\x80\x8D\xDF"
-     "\x00\x80"
-     "\x91\xDF\x00\x80\x95\xDF\x00\x80\x99\xDF\x00\x80\xE1\xA4\x01\x00\xBE\x4C"
-     "\x02\x00"
-     "\xE3\xA4\x00\x00\x80\x29\x00\x00\x81\x29\x00\x00\x82\x39\x00\x00\xEF\x78"
-     "\x00\x00"
-     "\xA1\xDF\x00\x80\x84\xB8\x02\x00\xA5\xDF\x00\x80\xA9\xDF\x00\x80\xB3\x61"
-     "\x01\x00"
-     "\xAD\xDF\x00\x80\xB1\xDF\x00\x80\xB5\xDF\x00\x80\xB9\xDF\x00\x80\xB6\x79"
-     "\x01\x00"
-     "\xB5\x71\x01\x00\xBD\xDF\x00\x80\xBB\xE5\x01\x00\xBA\x59\x01\x00\x86\x48"
-     "\x05\x00"
-     "\x87\x40\x05\x00\xBF\xED\x01\x00\xBE\xE5\x01\x00\xBD\xF5\x01\x00\xBC\xF5"
-     "\x01\x00"
-     "\xC1\xDF\x00\x80\xC5\xDF\x00\x80\xC9\xDF\x00\x80\x84\x40\x04\x00\xCD\xDF"
-     "\x00\x80"
-     "\xD1\xDF\x00\x80\xD5\xDF\x00\x80\xEF\xE8\x01\x00\xD9\xDF\x00\x80\xE1\xCC"
-     "\x0E\x00"
-     "\xAC\x5C\x0B\x00\xE3\x34\x01\x00\xDD\xDF\x00\x80\xE1\xDF\x00\x80\x84\xB0"
-     "\x04\x00"
-     "\xE5\xDF\x00\x80\xA5\xE9\x01\x00\xE9\xDF\x00\x80\xED\xDF\x00\x80\xA6\xE1"
-     "\x01\x00"
-     "\xF1\xDF\x00\x80\xF5\xDF\x00\x80\xA3\xF9\x01\x00\xF9\xDF\x00\x80\xAD\x6D"
-     "\x01\x00"
-     "\xAC\x6D\x01\x00\xAF\x75\x01\x00\xAE\x7D\x01\x00\xFD\xDF\x00\x80\x01\xE0"
-     "\x00\x80"
-     "\xAB\x7D\x01\x00\xAA\xC1\x01\x00\xA8\xB9\x05\x00\xA9\xB9\x05\x00\xAA\x29"
-     "\x06\x00"
-     "\xAB\x29\x06\x00\xAC\x39\x06\x00\xAD\x39\x06\x00\xAE\x2D\x06\x00\xAF\x25"
-     "\x06\x00"
-     "\x9D\xDF\x00\x80\x82\x11\x00\x00\x81\x19\x00\x00\x80\xD9\x07\x00\x05\xE0"
-     "\x00\x80"
-     "\x09\xE0\x00\x80\xBE\x34\x03\x00\x0D\xE0\x00\x80\xB8\x81\x06\x00\xB9\x8D"
-     "\x06\x00"
-     "\xBA\x85\x06\x00\xBB\x99\x06\x00\xBC\x89\x06\x00\xBD\xB9\x06\x00\xBE\x59"
-     "\x07\x00"
-     "\xBF\x59\x07\x00\xB0\x5D\x06\x00\xB1\xE1\x06\x00\xB2\xE1\x06\x00\xB3\xE1"
-     "\x06\x00"
-     "\xB4\xE1\x06\x00\xB5\xE9\x06\x00\xB6\xD1\x06\x00\xB7\xD1\x06\x00\xA8\xB1"
-     "\x07\x00"
-     "\xA9\xBD\x07\x00\xAA\xB5\x07\x00\xAB\xC9\x07\x00\xAC\xD9\x07\x00\xAD\xD9"
-     "\x07\x00"
-     "\xAE\xC9\x07\x00\xAF\xC1\x07\x00\x86\x00\x00\x00\x87\x28\x03\x00\x11\xE0"
-     "\x00\x80"
-     "\x15\xE0\x00\x80\x19\xE0\x00\x80\x1D\xE0\x00\x80\x21\xE0\x00\x80\x25\xE0"
-     "\x00\x80"
-     "\xB8\x75\x07\x00\xB9\x79\x07\x00\xBA\x0D\x07\x00\xBB\x05\x07\x00\xBC\x01"
-     "\x07\x00"
-     "\xBD\x01\x07\x00\xBE\x05\x07\x00\xBF\x2D\x07\x00\xB0\x61\x07\x00\xB1\x61"
-     "\x07\x00"
-     "\xB2\x61\x07\x00\xB3\x7D\x07\x00\xB4\x65\x07\x00\xB5\x6D\x07\x00\xB6\x59"
-     "\x07\x00"
-     "\xB7\x4D\x07\x00\xB3\x3D\x06\x00\x29\xE0\x00\x80\x2D\xE0\x00\x80\x31\xE0"
-     "\x00\x80"
-     "\x35\xE0\x00\x80\xB6\x59\x06\x00\xB5\x51\x06\x00\x39\xE0\x00\x80\xBB\xA5"
-     "\x06\x00"
-     "\xBA\xA5\x06\x00\x3D\xE0\x00\x80\x41\xE0\x00\x80\xBF\xA9\x06\x00\xBE\xA1"
-     "\x06\x00"
-     "\xBD\xA9\x06\x00\xBC\xB5\x06\x00\xEF\xA0\x04\x00\x45\xE0\x00\x80\x49\xE0"
-     "\x00\x80"
-     "\x4D\xE0\x00\x80\x51\xE0\x00\x80\x55\xE0\x00\x80\x59\xE0\x00\x80\x5D\xE0"
-     "\x00\x80"
-     "\x61\xE0\x00\x80\x65\xE0\x00\x80\x69\xE0\x00\x80\x6D\xE0\x00\x80\xE3\x54"
-     "\x04\x00"
-     "\x71\xE0\x00\x80\xE1\xD0\x06\x00\x75\xE0\x00\x80\xA3\xBD\x06\x00\x82\x15"
-     "\x00\x00"
-     "\x81\xF5\x00\x00\x80\xF9\x00\x00\x79\xE0\x00\x80\xA6\xD9\x06\x00\xA5\xD1"
-     "\x06\x00"
-     "\xBE\x90\x01\x00\xAB\x25\x06\x00\xAA\x25\x06\x00\x86\x08\x00\x00\x87\x70"
-     "\x01\x00"
-     "\xAF\x29\x06\x00\xAE\x21\x06\x00\xAD\x29\x06\x00\xAC\x35\x06\x00\x81\xE0"
-     "\x00\x80"
-     "\xB3\x75\x01\x00\x85\xE0\x00\x80\x89\xE0\x00\x80\xB6\x6D\x01\x00\x8D\xE0"
-     "\x00\x80"
-     "\x91\xE0\x00\x80\xB5\x65\x01\x00\xBA\xED\x00\x00\xBB\xF5\x00\x00\x95\xE0"
-     "\x00\x80"
-     "\x99\xE0\x00\x80\xBE\xDD\x00\x00\xBF\xB1\x00\x00\xBC\xE5\x00\x00\xBD\xD5"
-     "\x00\x00"
-     "\xA8\xA9\x02\x00\xA9\xA9\x02\x00\xAA\xB9\x02\x00\xAB\xB9\x02\x00\xAC\xA9"
-     "\x02\x00"
-     "\xAD\xA9\x02\x00\xAE\x1D\x03\x00\xAF\x15\x03\x00\xBE\xAC\x1C\x00\x9D\xE0"
-     "\x00\x80"
-     "\xA1\xE0\x00\x80\xBE\xEC\x1C\x00\xA5\xE0\x00\x80\xA9\xE0\x00\x80\xAD\xE0"
-     "\x00\x80"
-     "\xB1\xE0\x00\x80\xB8\x11\x03\x00\xB9\x11\x03\x00\xBA\x11\x03\x00\xBB\x11"
-     "\x03\x00"
-     "\xBC\x31\x03\x00\xBD\x31\x03\x00\xBE\x31\x03\x00\xBF\x31\x03\x00\xB0\x6D"
-     "\x03\x00"
-     "\xB1\x75\x03\x00\xB2\x75\x03\x00\xB3\x45\x03\x00\xB4\x5D\x03\x00\xB5\x31"
-     "\x03\x00"
-     "\xB6\x31\x03\x00\xB7\x31\x03\x00\xA3\x31\x02\x00\xB5\xE0\x00\x80\xB9\xE0"
-     "\x00\x80"
-     "\xBD\xE0\x00\x80\xC1\xE0\x00\x80\xA6\x29\x02\x00\xA5\x21\x02\x00\xC5\xE0"
-     "\x00\x80"
-     "\xAB\xB1\x03\x00\xAA\xA9\x03\x00\xC9\xE0\x00\x80\xCD\xE0\x00\x80\xAF\xF5"
-     "\x03\x00"
-     "\xAE\x99\x03\x00\xAD\x91\x03\x00\xAC\xA1\x03\x00\x84\xC0\x03\x00\xD1\xE0"
-     "\x00\x80"
-     "\xD5\xE0\x00\x80\xD9\xE0\x00\x80\xDD\xE0\x00\x80\x82\x3D\x00\x00\x81\x3D"
-     "\x00\x00"
-     "\x80\x3D\x00\x00\xE1\xE0\x00\x80\xE5\xE0\x00\x80\xE9\xE0\x00\x80\xF1\xE0"
-     "\x00\x80"
-     "\x86\xD0\x1C\x00\x87\xB4\x1C\x00\xF5\xE0\x00\x80\x2D\xDF\x00\x80\xA8\xC1"
-     "\x02\x00"
-     "\xA9\xCD\x02\x00\xAA\xC5\x02\x00\xAB\xD9\x02\x00\xAC\xC9\x02\x00\xAD\xFD"
-     "\x02\x00"
-     "\xAE\xF5\x02\x00\xAF\x39\x01\x00\xF9\xE0\x00\x80\xFD\xE0\x00\x80\x01\xE1"
-     "\x00\x80"
-     "\x05\xE1\x00\x80\x09\xE1\x00\x80\x0D\xE1\x00\x80\x11\xE1\x00\x80\x15\xE1"
-     "\x00\x80"
-     "\xB8\xD1\x01\x00\xB9\xD9\x01\x00\xBA\xE1\x01\x00\xBB\xE1\x01\x00\xBC\x95"
-     "\x01\x00"
-     "\xBD\x9D\x01\x00\xBE\x95\x01\x00\xBF\x8D\x01\x00\xB0\x49\x01\x00\xB1\x55"
-     "\x01\x00"
-     "\xB2\x5D\x01\x00\xB3\x55\x01\x00\xB4\x4D\x01\x00\xB5\xF1\x01\x00\xB6\xF1"
-     "\x01\x00"
-     "\xB7\xF1\x01\x00\xEF\xE4\x00\x00\x19\xE1\x00\x80\x1D\xE1\x00\x80\x84\x00"
-     "\x1C\x00"
-     "\x21\xE1\x00\x80\x25\xE1\x00\x80\x29\xE1\x00\x80\xEF\x5C\x01\x00\x2D\xE1"
-     "\x00\x80"
-     "\xE1\xD4\x06\x00\x31\xE1\x00\x80\xE3\x30\x06\x00\xE3\x18\x06\x00\x35\xE1"
-     "\x00\x80"
-     "\xE1\x00\x01\x00\x39\xE1\x00\x80\x3D\xE1\x00\x80\x41\xE1\x00\x80\x45\xE1"
-     "\x00\x80"
-     "\xB3\x01\x02\x00\x49\xE1\x00\x80\xB5\x0D\x02\x00\xB6\x31\x02\x00\x4D\xE1"
-     "\x00\x80"
-     "\x51\xE1\x00\x80\x55\xE1\x00\x80\xBA\xC1\x01\x00\xBB\xDD\x01\x00\xBC\xC5"
-     "\x01\x00"
-     "\xBD\xCD\x01\x00\xBE\xC5\x01\x00\xBF\x79\x01\x00\xA3\xCD\x1D\x00\xED\xE0"
-     "\x00\x80"
-     "\x59\xE1\x00\x80\x61\xE1\x00\x80\x5D\xE1\x00\x80\xA6\xFD\x1D\x00\xA5\xC1"
-     "\x1D\x00"
-     "\x65\xE1\x00\x80\xAB\x11\x1E\x00\xAA\x0D\x1E\x00\x86\x08\x03\x00\x87\x1C"
-     "\x03\x00"
-     "\xAF\xB5\x1E\x00\xAE\x09\x1E\x00\xAD\x01\x1E\x00\xAC\x09\x1E\x00\x80\xA1"
-     "\x1F\x00"
-     "\x81\xA1\x1F\x00\x82\x41\x00\x00\xA3\x6D\x1F\x00\x69\xE1\x00\x80\xA5\x7D"
-     "\x1F\x00"
-     "\xA6\x75\x1F\x00\x6D\xE1\x00\x80\x71\xE1\x00\x80\x75\xE1\x00\x80\xAA\x51"
-     "\x1F\x00"
-     "\xAB\x51\x1F\x00\xAC\xF5\x1F\x00\xAD\xFD\x1F\x00\xAE\xE9\x1F\x00\xAF\xE9"
-     "\x1F\x00"
-     "\xA8\x6D\x1E\x00\xA9\x7D\x1E\x00\xAA\xA9\x1F\x00\xAB\xA9\x1F\x00\xAC\xB9"
-     "\x1F\x00"
-     "\xAD\xB9\x1F\x00\xAE\xA9\x1F\x00\xAF\xA9\x1F\x00\x79\xE1\x00\x80\x7D\xE1"
-     "\x00\x80"
-     "\x81\xE1\x00\x80\x85\xE1\x00\x80\x89\xE1\x00\x80\x8D\xE1\x00\x80\x91\xE1"
-     "\x00\x80"
-     "\x95\xE1\x00\x80\xB8\x69\x1F\x00\xB9\x71\x1F\x00\xBA\x75\x1F\x00\xBB\x0D"
-     "\x1F\x00"
-     "\xBC\x15\x1F\x00\xBD\x1D\x1F\x00\xBE\x09\x1F\x00\xBF\x01\x1F\x00\xB0\xD9"
-     "\x1F\x00"
-     "\xB1\xD9\x1F\x00\xB2\x6D\x1F\x00\xB3\x65\x1F\x00\xB4\x61\x1F\x00\xB5\x61"
-     "\x1F\x00"
-     "\xB6\x61\x1F\x00\xB7\x5D\x1F\x00\x99\xE1\x00\x80\xE1\xB4\x1E\x00\x9D\xE1"
-     "\x00\x80"
-     "\xE3\xDC\x1D\x00\xA1\xE1\x00\x80\xA5\xE1\x00\x80\xA9\xE1\x00\x80\xAD\xE1"
-     "\x00\x80"
-     "\xB1\xE1\x00\x80\xB5\xE1\x00\x80\xB9\xE1\x00\x80\xBD\xE1\x00\x80\x80\x75"
-     "\x00\x00"
-     "\x81\x7D\x00\x00\x82\x75\x00\x00\xEF\xE4\x1D\x00\xB3\x61\x1E\x00\xC1\xE1"
-     "\x00\x80"
-     "\xC5\xE1\x00\x80\xC9\xE1\x00\x80\x84\x9C\x01\x00\xB6\x79\x1E\x00\xB5\x71"
-     "\x1E\x00"
-     "\xBE\x94\x01\x00\xBB\x5D\x1E\x00\xBA\x5D\x1E\x00\xCD\xE1\x00\x80\xD1\xE1"
-     "\x00\x80"
-     "\xBF\xE5\x1E\x00\xBE\xE5\x1E\x00\xBD\xF1\x1E\x00\xBC\xF9\x1E\x00\xA8\x8D"
-     "\x01\x00"
-     "\xA9\x99\x01\x00\xAA\xD9\x01\x00\xAB\xD1\x01\x00\xAC\xF1\x01\x00\xAD\xF1"
-     "\x01\x00"
-     "\xAE\x39\x01\x00\xAF\x39\x01\x00\x86\x00\x00\x00\x87\x64\x01\x00\xD5\xE1"
-     "\x00\x80"
-     "\xD9\xE1\x00\x80\xDD\xE1\x00\x80\xE1\xE1\x00\x80\xE5\xE1\x00\x80\xE9\xE1"
-     "\x00\x80"
-     "\xB8\xD1\x00\x00\xB9\xD1\x00\x00\xBA\xD1\x00\x00\xBB\xD1\x00\x00\xBC\xF1"
-     "\x00\x00"
-     "\xBD\xF1\x00\x00\xBE\xF1\x00\x00\xBF\xF1\x00\x00\xB0\x49\x01\x00\xB1\x49"
-     "\x01\x00"
-     "\xB2\x59\x01\x00\xB3\x59\x01\x00\xB4\x49\x01\x00\xB5\x49\x01\x00\xB6\xF9"
-     "\x00\x00"
-     "\xB7\xF5\x00\x00\xB3\x51\x02\x00\xED\xE1\x00\x80\xF1\xE1\x00\x80\xF5\xE1"
-     "\x00\x80"
-     "\xF9\xE1\x00\x80\xB6\x71\x02\x00\xB5\x79\x02\x00\xFD\xE1\x00\x80\xBB\x8D"
-     "\x02\x00"
-     "\xBA\x8D\x02\x00\x01\xE2\x00\x80\x05\xE2\x00\x80\xBF\x8D\x02\x00\xBE\x8D"
-     "\x02\x00"
-     "\xBD\x8D\x02\x00\xBC\x8D\x02\x00\xAC\x24\x3B\x00\xA3\x15\x02\x00\x09\xE2"
-     "\x00\x80"
-     "\x0D\xE2\x00\x80\xA6\x35\x02\x00\x11\xE2\x00\x80\x15\xE2\x00\x80\xA5\x3D"
-     "\x02\x00"
-     "\xAA\xC9\x02\x00\xAB\xC9\x02\x00\x19\xE2\x00\x80\x1D\xE2\x00\x80\xAE\xC9"
-     "\x02\x00"
-     "\xAF\xC9\x02\x00\xAC\xC9\x02\x00\xAD\xC9\x02\x00\x7D\xE0\x00\x80\x21\xE2"
-     "\x00\x80"
-     "\x25\xE2\x00\x80\x84\x80\x03\x00\x29\xE2\x00\x80\x2D\xE2\x00\x80\x31\xE2"
-     "\x00\x80"
-     "\x35\xE2\x00\x80\x39\xE2\x00\x80\x41\xE2\x00\x80\x45\xE2\x00\x80\xBE\x00"
-     "\x04\x00"
-     "\x49\xE2\x00\x80\x4D\xE2\x00\x80\x51\xE2\x00\x80\x55\xE2\x00\x80\x80\x59"
-     "\x00\x00"
-     "\x81\x59\x00\x00\x82\x69\x00\x00\x59\xE2\x00\x80\x5D\xE2\x00\x80\xEF\x80"
-     "\x02\x00"
-     "\x87\x14\x03\x00\x86\xFC\x04\x00\x61\xE2\x00\x80\xE3\x78\x02\x00\x65\xE2"
-     "\x00\x80"
-     "\xE1\x54\x01\x00\x69\xE2\x00\x80\x6D\xE2\x00\x80\x71\xE2\x00\x80\x75\xE2"
-     "\x00\x80"
-     "\xA8\x6D\x02\x00\xA9\xAD\x02\x00\xAA\xA5\x02\x00\xAB\xBD\x02\x00\xAC\xA5"
-     "\x02\x00"
-     "\xAD\xAD\x02\x00\xAE\xA5\x02\x00\xAF\x19\x01\x00\x84\xAC\x05\x00\x79\xE2"
-     "\x00\x80"
-     "\x7D\xE2\x00\x80\x81\xE2\x00\x80\xBE\xBC\x04\x00\x85\xE2\x00\x80\x89\xE2"
-     "\x00\x80"
-     "\x8D\xE2\x00\x80\xB8\x15\x01\x00\xB9\x1D\x01\x00\xBA\x21\x01\x00\xBB\x21"
-     "\x01\x00"
-     "\xBC\xD5\x01\x00\xBD\xDD\x01\x00\xBE\xC9\x01\x00\xBF\xC9\x01\x00\xB0\x69"
-     "\x01\x00"
-     "\xB1\x69\x01\x00\xB2\x79\x01\x00\xB3\x71\x01\x00\xB4\x59\x01\x00\xB5\x59"
-     "\x01\x00"
-     "\xB6\x31\x01\x00\xB7\x31\x01\x00\xE1\x68\x1F\x00\x91\xE2\x00\x80\xE3\x14"
-     "\x1E\x00"
-     "\x85\xB4\x04\x00\x95\xE2\x00\x80\xE1\xA4\x01\x00\x99\xE2\x00\x80\xE3\xDC"
-     "\x00\x00"
-     "\x9D\xE2\x00\x80\xBF\xA8\x07\x00\xA1\xE2\x00\x80\xEF\x08\x00\x00\xA5\xE2"
-     "\x00\x80"
-     "\xA9\xE2\x00\x80\xEF\x6C\x1F\x00\xB1\xE2\x00\x80\xB3\x6D\x02\x00\x82\x71"
-     "\x00\x00"
-     "\x81\x79\x00\x00\x80\x65\x00\x00\x85\xD4\x07\x00\xB6\x6D\x02\x00\xB5\x6D"
-     "\x02\x00"
-     "\xB5\xE2\x00\x80\xBB\xED\x01\x00\xBA\xED\x01\x00\x86\x48\x06\x00\x87\x6C"
-     "\x04\x00"
-     "\xBF\xDD\x01\x00\xBE\xDD\x01\x00\xBD\xE5\x01\x00\xBC\xED\x01\x00\xAB\x0D"
-     "\x06\x00"
-     "\xAA\x0D\x06\x00\x3D\xE2\x00\x80\x84\x00\x03\x00\xAF\x3D\x06\x00\xAE\x3D"
-     "\x06\x00"
-     "\xAD\x05\x06\x00\xAC\x0D\x06\x00\xA3\x8D\x05\x00\xAD\xE2\x00\x80\xB9\xE2"
-     "\x00\x80"
-     "\xBD\xE2\x00\x80\xC1\xE2\x00\x80\xA6\x8D\x05\x00\xA5\x8D\x05\x00\xC5\xE2"
-     "\x00\x80"
-     "\xB3\xF1\x06\x00\xC9\xE2\x00\x80\xCD\xE2\x00\x80\xD1\xE2\x00\x80\xD5\xE2"
-     "\x00\x80"
-     "\xB6\xE9\x06\x00\xB5\xE1\x06\x00\xD9\xE2\x00\x80\xBB\x89\x06\x00\xBA\x85"
-     "\x06\x00"
-     "\xDD\xE2\x00\x80\xE1\xE2\x00\x80\xBF\x8D\x06\x00\xBE\x8D\x06\x00\xBD\x8D"
-     "\x06\x00"
-     "\xBC\x8D\x06\x00\xA8\x25\x06\x00\xA9\x2D\x06\x00\xAA\x39\x06\x00\xAB\x39"
-     "\x06\x00"
-     "\xAC\x29\x06\x00\xAD\x29\x06\x00\xAE\x55\x06\x00\xAF\x4D\x06\x00\xE5\xE2"
-     "\x00\x80"
-     "\xE9\xE2\x00\x80\xED\xE2\x00\x80\xF1\xE2\x00\x80\xF5\xE2\x00\x80\xF9\xE2"
-     "\x00\x80"
-     "\xFD\xE2\x00\x80\x01\xE3\x00\x80\xB8\xED\x07\x00\xB9\xF1\x07\x00\xBA\xF1"
-     "\x07\x00"
-     "\xBB\xF1\x07\x00\xBC\x55\x07\x00\xBD\x45\x07\x00\xBE\x41\x07\x00\xBF\x7D"
-     "\x07\x00"
-     "\xB0\x01\x06\x00\xB1\x0D\x06\x00\xB2\x19\x06\x00\xB3\x19\x06\x00\xB4\x09"
-     "\x06\x00"
-     "\xB5\x09\x06\x00\xB6\xDD\x07\x00\xB7\xD5\x07\x00\xA3\xB1\x07\x00\x8F\xED"
-     "\xED\x03"
-     "\x05\xE3\x00\x80\x09\xE3\x00\x80\x0D\xE3\x00\x80\xA6\xA9\x07\x00\xA5\xA1"
-     "\x07\x00"
-     "\x11\xE3\x00\x80\xAB\xC9\x07\x00\xAA\xC5\x07\x00\x15\xE3\x00\x80\x19\xE3"
-     "\x00\x80"
-     "\xAF\xCD\x07\x00\xAE\xCD\x07\x00\xAD\xCD\x07\x00\xAC\xCD\x07\x00\x9E\x7D"
-     "\x19\x00"
-     "\x9F\xC5\x07\x00\x9C\xB5\x1B\x00\x9D\x6D\x19\x00\x9A\x71\x1D\x00\x9B\xC5"
-     "\x1D\x00"
-     "\x98\x8D\x11\x00\x99\xB5\x1F\x00\x96\x85\x13\x00\x97\x45\x11\x00\x94\x45"
-     "\x15\x00"
-     "\x95\x49\x14\x00\x92\x5D\xE8\x03\x93\xA5\x17\x00\x90\x79\xE9\x03\x91\x79"
-     "\xE9\x03"
-     "\x80\x9D\x00\x00\x81\xAD\x00\x00\x82\xA5\x00\x00\xBE\xC0\x00\x00\xE1\x54"
-     "\x04\x00"
-     "\xEF\x40\x00\x00\xE3\xA4\x38\x00\x21\xE3\x00\x80\x84\x6C\x0F\x00\xE3\x14"
-     "\x02\x00"
-     "\xEF\x78\x38\x00\xE1\x7C\x02\x00\xA2\xD9\x04\x00\xA3\xBD\x01\x00\xA0\xBD"
-     "\x05\x00"
-     "\xA1\xAD\x05\x00\xB3\xE9\x01\x00\x25\xE3\x00\x80\x86\x00\x0C\x00\x87\x30"
-     "\x01\x00"
-     "\x29\xE3\x00\x80\xB6\xE9\x01\x00\xB5\xE9\x01\x00\x2D\xE3\x00\x80\xBB\xB1"
-     "\x01\x00"
-     "\xBA\xA9\x01\x00\x31\xE3\x00\x80\x35\xE3\x00\x80\xBF\x51\x01\x00\xBE\x99"
-     "\x01\x00"
-     "\xBD\x91\x01\x00\xBC\xA1\x01\x00\x83\x01\xEA\x03\x82\xED\xEE\x03\x39\xE3"
-     "\x00\x80"
-     "\x3D\xE3\x00\x80\x87\x3D\x16\x00\x86\xC5\x14\x00\x85\x35\xEB\x03\x84\x19"
-     "\xEA\x03"
-     "\x8B\xC5\x10\x00\x8A\x01\x17\x00\x1D\xE3\x00\x80\x41\xE3\x00\x80\x8F\x11"
-     "\x1E\x00"
-     "\x8E\xDD\x1C\x00\x8D\xED\x12\x00\x8C\x1D\x12\x00\x93\xBD\x18\x00\x92\xD5"
-     "\x1E\x00"
-     "\x45\xE3\x00\x80\x49\xE3\x00\x80\x97\x55\x06\x00\x96\xD5\x04\x00\x95\x65"
-     "\x1A\x00"
-     "\x94\x7D\x1A\x00\x9B\x09\x07\x00\x9A\x75\x06\x00\x4D\xE3\x00\x80\x51\xE3"
-     "\x00\x80"
-     "\x55\xE3\x00\x80\x59\xE3\x00\x80\x5D\xE3\x00\x80\x9C\x75\x02\x00\xA1\xA5"
-     "\x02\x00"
-     "\x61\xE3\x00\x80\xA3\x8D\x0E\x00\xA2\xA5\x00\x00\xA5\x01\x08\x00\xA4\xFD"
-     "\x0E\x00"
-     "\xA7\xBD\x0A\x00\xA6\xBD\x0A\x00\xA9\x81\x36\x00\xA8\x01\x34\x00\xAB\xE5"
-     "\x37\x00"
-     "\xAA\x89\x36\x00\xAD\xA1\x32\x00\xAC\xB5\x32\x00\xAF\x01\x3C\x00\xAE\xED"
-     "\x33\x00"
-     "\xB1\xD9\x3F\x00\xB0\xE5\x3E\x00\xB3\xE9\x3A\x00\xB2\x01\x38\x00\xB5\x01"
-     "\x24\x00"
-     "\xB4\x49\x3A\x00\x65\xE3\x00\x80\x69\xE3\x00\x80\x6D\xE3\x00\x80\x71\xE3"
-     "\x00\x80"
-     "\x75\xE3\x00\x80\x79\xE3\x00\x80\x7D\xE3\x00\x80\x82\xDD\x01\x00\x81\xDD"
-     "\x01\x00"
-     "\x80\xDD\x01\x00\xA8\xBD\x04\x00\xA9\xD9\x04\x00\xAA\x21\x03\x00\xAB\x21"
-     "\x03\x00"
-     "\xAC\x21\x03\x00\xAD\x29\x03\x00\xAE\x5D\x03\x00\xAF\x55\x03\x00\x81\xE3"
-     "\x00\x80"
-     "\x85\xE3\x00\x80\x89\xE3\x00\x80\xBE\xE4\x0D\x00\x91\xE3\x00\x80\x95\xE3"
-     "\x00\x80"
-     "\x86\x18\x0D\x00\x87\x9C\x0D\x00\xB8\xED\x03\x00\xB9\xF5\x03\x00\xBA\xFD"
-     "\x03\x00"
-     "\xBB\xF5\x03\x00\xBC\xED\x03\x00\xBD\x5D\x02\x00\xBE\x55\x02\x00\xBF\x4D"
-     "\x02\x00"
-     "\xB0\x2D\x03\x00\xB1\x35\x03\x00\xB2\x35\x03\x00\xB3\x05\x03\x00\xB4\x1D"
-     "\x03\x00"
-     "\xB5\xD5\x03\x00\xB6\xDD\x03\x00\xB7\xD5\x03\x00\x84\x4C\x0C\x00\xE3\xB4"
-     "\x06\x00"
-     "\x99\xE3\x00\x80\xE1\x48\x07\x00\x9D\xE3\x00\x80\xA1\xE3\x00\x80\xEF\xEC"
-     "\x00\x00"
-     "\xA5\xE3\x00\x80\xE1\x50\x06\x00\xA9\xE3\x00\x80\xE3\xB4\x01\x00\xAD\xE3"
-     "\x00\x80"
-     "\xBE\x78\x0C\x00\xEF\x7C\x06\x00\xB1\xE3\x00\x80\xB5\xE3\x00\x80\xA3\xE9"
-     "\x03\x00"
-     "\xB9\xE3\x00\x80\xBD\xE3\x00\x80\xC1\xE3\x00\x80\x84\x5C\x0F\x00\xA6\xE9"
-     "\x03\x00"
-     "\xA5\xE9\x03\x00\xC5\xE3\x00\x80\xAB\xB1\x03\x00\xAA\xA9\x03\x00\xC9\xE3"
-     "\x00\x80"
-     "\xCD\xE3\x00\x80\xAF\x51\x03\x00\xAE\x99\x03\x00\xAD\x91\x03\x00\xAC\xA1"
-     "\x03\x00"
-     "\xB3\xD1\x0D\x00\x8D\xE3\x00\x80\xD1\xE3\x00\x80\xD5\xE3\x00\x80\xD9\xE3"
-     "\x00\x80"
-     "\xB6\xF9\x0D\x00\xB5\xF1\x0D\x00\xDD\xE3\x00\x80\xBB\x69\x0E\x00\xBA\x09"
-     "\x0E\x00"
-     "\xE1\xE3\x00\x80\xE5\xE3\x00\x80\xBF\x41\x0E\x00\xBE\x59\x0E\x00\xBD\x51"
-     "\x0E\x00"
-     "\xBC\x71\x0E\x00\xE9\xE3\x00\x80\xA3\x95\x0D\x00\xED\xE3\x00\x80\xF1\xE3"
-     "\x00\x80"
-     "\xA6\xBD\x0D\x00\xF5\xE3\x00\x80\xF9\xE3\x00\x80\xA5\xB5\x0D\x00\xAA\x4D"
-     "\x0E\x00"
-     "\xAB\x2D\x0E\x00\xFD\xE3\x00\x80\x01\xE4\x00\x80\xAE\x1D\x0E\x00\xAF\x05"
-     "\x0E\x00"
-     "\xAC\x35\x0E\x00\xAD\x15\x0E\x00\xA8\xB1\x0F\x00\xA9\xB1\x0F\x00\xAA\x01"
-     "\x0F\x00"
-     "\xAB\x1D\x0F\x00\xAC\x05\x0F\x00\xAD\x09\x0F\x00\xAE\x3D\x0F\x00\xAF\x31"
-     "\x0F\x00"
-     "\x80\xED\x00\x00\x81\xF1\x00\x00\x82\xF1\x00\x00\x05\xE4\x00\x80\x86\x90"
-     "\x03\x00"
-     "\x87\xB4\x00\x00\x09\xE4\x00\x80\x0D\xE4\x00\x80\xB8\xC9\x00\x00\xB9\xC9"
-     "\x00\x00"
-     "\xBA\xD9\x00\x00\xBB\xD1\x00\x00\xBC\xF9\x00\x00\xBD\xF9\x00\x00\xBE\x99"
-     "\x00\x00"
-     "\xBF\x95\x00\x00\xB0\x59\x0F\x00\xB1\x59\x0F\x00\xB2\x2D\x0F\x00\xB3\x25"
-     "\x0F\x00"
-     "\xB4\x3D\x0F\x00\xB5\x29\x0F\x00\xB6\x11\x0F\x00\xB7\x11\x0F\x00\xB3\x89"
-     "\x0E\x00"
-     "\x11\xE4\x00\x80\x15\xE4\x00\x80\x19\xE4\x00\x80\x1D\xE4\x00\x80\xB6\x89"
-     "\x0E\x00"
-     "\xB5\x8D\x0E\x00\x21\xE4\x00\x80\xBB\x9D\x0E\x00\xBA\x9D\x0E\x00\x25\xE4"
-     "\x00\x80"
-     "\x29\xE4\x00\x80\xBF\xED\x0E\x00\xBE\xF9\x0E\x00\xBD\xF1\x0E\x00\xBC\xF9"
-     "\x0E\x00"
-     "\x2D\xE4\x00\x80\xA3\xCD\x0E\x00\x31\xE4\x00\x80\x35\xE4\x00\x80\xA6\xCD"
-     "\x0E\x00"
-     "\x39\xE4\x00\x80\x3D\xE4\x00\x80\xA5\xC9\x0E\x00\xAA\xD9\x0E\x00\xAB\xD9"
-     "\x0E\x00"
-     "\x41\xE4\x00\x80\x45\xE4\x00\x80\xAE\xBD\x0E\x00\xAF\xA9\x0E\x00\xAC\xBD"
-     "\x0E\x00"
-     "\xAD\xB5\x0E\x00\xA8\xED\x01\x00\xA9\x29\x01\x00\xAA\x31\x01\x00\xAB\x31"
-     "\x01\x00"
-     "\xAC\x95\x00\x00\xAD\x9D\x00\x00\xAE\x95\x00\x00\xAF\x8D\x00\x00\x49\xE4"
-     "\x00\x80"
-     "\x4D\xE4\x00\x80\x51\xE4\x00\x80\x55\xE4\x00\x80\x59\xE4\x00\x80\x5D\xE4"
-     "\x00\x80"
-     "\x61\xE4\x00\x80\x65\xE4\x00\x80\xB8\xA9\x00\x00\xB9\xA9\x00\x00\xBA\x69"
-     "\x00\x00"
-     "\xBB\x69\x00\x00\xBC\x79\x00\x00\xBD\x79\x00\x00\xBE\x69\x00\x00\xBF\x69"
-     "\x00\x00"
-     "\xB0\xFD\x00\x00\xB1\xCD\x00\x00\xB2\xC5\x00\x00\xB3\xAD\x00\x00\xB4\xB9"
-     "\x00\x00"
-     "\xB5\xB9\x00\x00\xB6\xAD\x00\x00\xB7\xA5\x00\x00\x80\xB9\x03\x00\x81\xB9"
-     "\x03\x00"
-     "\x82\x05\x00\x00\x69\xE4\x00\x80\xBE\x08\x02\x00\x71\xE4\x00\x80\x75\xE4"
-     "\x00\x80"
-     "\x79\xE4\x00\x80\x86\x80\x04\x00\x87\xE8\x02\x00\x7D\xE4\x00\x80\x81\xE4"
-     "\x00\x80"
-     "\x85\xE4\x00\x80\x89\xE4\x00\x80\x8D\xE4\x00\x80\x91\xE4\x00\x80\xB3\xF9"
-     "\x02\x00"
-     "\x95\xE4\x00\x80\x99\xE4\x00\x80\x9D\xE4\x00\x80\xA1\xE4\x00\x80\xB6\x5D"
-     "\x03\x00"
-     "\xB5\x5D\x03\x00\xA5\xE4\x00\x80\xBB\x49\x03\x00\xBA\x79\x03\x00\xA9\xE4"
-     "\x00\x80"
-     "\xAD\xE4\x00\x80\xBF\x3D\x03\x00\xBE\x3D\x03\x00\xBD\x39\x03\x00\xBC\x51"
-     "\x03\x00"
-     "\xA6\x89\x03\x00\xB1\xE4\x00\x80\xB5\xE4\x00\x80\xA5\x89\x03\x00\xB9\xE4"
-     "\x00\x80"
-     "\xA3\x2D\x02\x00\xBD\xE4\x00\x80\xC1\xE4\x00\x80\xAE\xE9\x03\x00\xAF\xE9"
-     "\x03\x00"
-     "\xAC\x85\x03\x00\xAD\xED\x03\x00\xAA\xAD\x03\x00\xAB\x9D\x03\x00\xC5\xE4"
-     "\x00\x80"
-     "\xC9\xE4\x00\x80\xCD\xE4\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xD1\xE4\x00\x80\xD5\xE4\x00\x80\xD9\xE4\x00\x80\xBE\x94\x04\x00\x84\xAC"
-     "\x04\x00"
-     "\xE1\xE4\x00\x80\x87\x28\x03\x00\x86\xCC\x05\x00\xE5\xE4\x00\x80\xE9\xE4"
-     "\x00\x80"
-     "\xED\xE4\x00\x80\xF1\xE4\x00\x80\xB2\x4D\x0D\x00\xB3\x6D\x0D\x00\xB0\x25"
-     "\x0D\x00"
-     "\xB1\x45\x0D\x00\xB6\x65\x0D\x00\xB7\x6D\x0D\x00\xB4\x75\x0D\x00\xB5\x6D"
-     "\x0D\x00"
-     "\xBA\xFD\x0D\x00\xBB\xC5\x0D\x00\xB8\xAD\x0D\x00\xB9\xF5\x0D\x00\xBE\xDD"
-     "\x0D\x00"
-     "\xBF\xC5\x0D\x00\xBC\xDD\x0D\x00\xBD\xD5\x0D\x00\xF5\xE4\x00\x80\xF9\xE4"
-     "\x00\x80"
-     "\xFD\xE4\x00\x80\x01\xE5\x00\x80\x05\xE5\x00\x80\x09\xE5\x00\x80\x0D\xE5"
-     "\x00\x80"
-     "\x11\xE5\x00\x80\xAA\xB1\x0E\x00\xAB\xB9\x0E\x00\xA8\xF5\x0E\x00\xA9\xA9"
-     "\x0E\x00"
-     "\xAE\x79\x0D\x00\xAF\x79\x0D\x00\xAC\x91\x0E\x00\xAD\x91\x0E\x00\xE1\xA0"
-     "\x0E\x00"
-     "\x15\xE5\x00\x80\xE3\x18\x0E\x00\x19\xE5\x00\x80\x1D\xE5\x00\x80\x21\xE5"
-     "\x00\x80"
-     "\x25\xE5\x00\x80\x29\xE5\x00\x80\x2D\xE5\x00\x80\x31\xE5\x00\x80\x35\xE5"
-     "\x00\x80"
-     "\x39\xE5\x00\x80\x3D\xE5\x00\x80\x41\xE5\x00\x80\xEF\xF0\x0E\x00\x45\xE5"
-     "\x00\x80"
-     "\x80\x75\x00\x00\x81\x79\x00\x00\x82\x19\x00\x00\x84\x40\x04\x00\x49\xE5"
-     "\x00\x80"
-     "\xEF\x3C\x03\x00\xBE\x54\x07\x00\x51\xE5\x00\x80\x55\xE5\x00\x80\xE3\x24"
-     "\x0E\x00"
-     "\x59\xE5\x00\x80\xE1\x44\x01\x00\x86\xF0\x07\x00\x87\xF4\x04\x00\x5D\xE5"
-     "\x00\x80"
-     "\x61\xE5\x00\x80\xA8\x15\x06\x00\xA9\x1D\x06\x00\xAA\x15\x06\x00\xAB\x25"
-     "\x06\x00"
-     "\xAC\x3D\x06\x00\xAD\x25\x06\x00\xAE\x2D\x06\x00\xAF\x25\x06\x00\xDD\xE4"
-     "\x00\x80"
-     "\x4D\xE5\x00\x80\x65\xE5\x00\x80\x69\xE5\x00\x80\x6D\xE5\x00\x80\x71\xE5"
-     "\x00\x80"
-     "\x75\xE5\x00\x80\x79\xE5\x00\x80\xB8\xC5\x06\x00\xB9\xCD\x06\x00\xBA\xD9"
-     "\x06\x00"
-     "\xBB\xD5\x06\x00\xBC\xCD\x06\x00\xBD\xED\x06\x00\xBE\xE5\x06\x00\xBF\x9D"
-     "\x06\x00"
-     "\xB0\x5D\x06\x00\xB1\x21\x06\x00\xB2\x21\x06\x00\xB3\x21\x06\x00\xB4\x21"
-     "\x06\x00"
-     "\xB5\x29\x06\x00\xB6\x11\x06\x00\xB7\x11\x06\x00\xB3\x9D\x06\x00\x7D\xE5"
-     "\x00\x80"
-     "\x81\xE5\x00\x80\x84\x0C\x00\x00\x85\xE5\x00\x80\xB6\xA5\x06\x00\xB5\xB5"
-     "\x06\x00"
-     "\x89\xE5\x00\x80\xBB\xE5\x06\x00\xBA\x99\x06\x00\x8D\xE5\x00\x80\x91\xE5"
-     "\x00\x80"
-     "\xBF\xED\x06\x00\xBE\xE5\x06\x00\xBD\xF5\x06\x00\xBC\xF5\x06\x00\x95\xE5"
-     "\x00\x80"
-     "\xA3\xD9\x06\x00\x99\xE5\x00\x80\x9D\xE5\x00\x80\xA6\xE1\x06\x00\xA1\xE5"
-     "\x00\x80"
-     "\xA5\xE5\x00\x80\xA5\xF1\x06\x00\xAA\xDD\x06\x00\xAB\xA1\x06\x00\xA9\xE5"
-     "\x00\x80"
-     "\xAD\xE5\x00\x80\xAE\xA1\x06\x00\xAF\xA9\x06\x00\xAC\xB1\x06\x00\xAD\xB1"
-     "\x06\x00"
-     "\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x11\x00\x00\xB1\xE5\x00\x80\xB5\xE5"
-     "\x00\x80"
-     "\xE1\x2C\x05\x00\xB9\xE5\x00\x80\xE3\xEC\x05\x00\xBE\xAC\x00\x00\xC1\xE5"
-     "\x00\x80"
-     "\xC5\xE5\x00\x80\xEF\x08\x05\x00\x86\x90\x3F\x00\x87\x74\x01\x00\xC9\xE5"
-     "\x00\x80"
-     "\xCD\xE5\x00\x80\xB3\xB5\x07\x00\xD1\xE5\x00\x80\xD5\xE5\x00\x80\xD9\xE5"
-     "\x00\x80"
-     "\xDD\xE5\x00\x80\xB6\x15\x06\x00\xB5\xA5\x07\x00\xE1\xE5\x00\x80\xBB\x31"
-     "\x06\x00"
-     "\xBA\x29\x06\x00\xE5\xE5\x00\x80\xE9\xE5\x00\x80\xBF\xF5\x06\x00\xBE\x29"
-     "\x06\x00"
-     "\xBD\x21\x06\x00\xBC\x29\x06\x00\xA8\x65\x07\x00\xA9\x75\x07\x00\xAA\x7D"
-     "\x07\x00"
-     "\xAB\x75\x07\x00\xAC\x6D\x07\x00\xAD\xD5\x06\x00\xAE\xDD\x06\x00\xAF\xD5"
-     "\x06\x00"
-     "\xED\xE5\x00\x80\xF1\xE5\x00\x80\xF5\xE5\x00\x80\xF9\xE5\x00\x80\xFD\xE5"
-     "\x00\x80"
-     "\x01\xE6\x00\x80\x05\xE6\x00\x80\x09\xE6\x00\x80\xB8\x55\x06\x00\xB9\x5D"
-     "\x06\x00"
-     "\xBA\x55\x06\x00\xBB\x6D\x06\x00\xBC\x75\x06\x00\xBD\x7D\x06\x00\xBE\x75"
-     "\x06\x00"
-     "\xBF\x6D\x06\x00\xB0\xAD\x06\x00\xB1\xB5\x06\x00\xB2\xBD\x06\x00\xB3\xB5"
-     "\x06\x00"
-     "\xB4\xAD\x06\x00\xB5\x71\x06\x00\xB6\x75\x06\x00\xB7\x6D\x06\x00\xA3\xF5"
-     "\x04\x00"
-     "\x0D\xE6\x00\x80\x11\xE6\x00\x80\x15\xE6\x00\x80\x19\xE6\x00\x80\xA6\x55"
-     "\x05\x00"
-     "\xA5\xE5\x04\x00\x1D\xE6\x00\x80\xAB\x71\x05\x00\xAA\x69\x05\x00\x21\xE6"
-     "\x00\x80"
-     "\x25\xE6\x00\x80\xAF\xB5\x05\x00\xAE\x69\x05\x00\xAD\x61\x05\x00\xAC\x69"
-     "\x05\x00"
-     "\x80\x69\x00\x00\x81\x69\x00\x00\x82\x19\x00\x00\xB3\xA5\x05\x00\x29\xE6"
-     "\x00\x80"
-     "\xB5\xA9\x05\x00\xB6\x71\x02\x00\x84\xD4\x03\x00\xBE\x20\x03\x00\x31\xE6"
-     "\x00\x80"
-     "\xBA\x5D\x02\x00\xBB\x55\x02\x00\xBC\x4D\x02\x00\xBD\xF1\x02\x00\xBE\xF1"
-     "\x02\x00"
-     "\xBF\xED\x02\x00\x35\xE6\x00\x80\x39\xE6\x00\x80\x87\x48\x03\x00\x86\xAC"
-     "\x3C\x00"
-     "\x6D\xE4\x00\x80\x3D\xE6\x00\x80\x41\xE6\x00\x80\xBE\x90\x3D\x00\xA8\x7D"
-     "\x05\x00"
-     "\xA9\x45\x05\x00\xAA\x41\x05\x00\xAB\x55\x05\x00\xAC\x4D\x05\x00\xAD\x9D"
-     "\x02\x00"
-     "\xAE\x95\x02\x00\xAF\x8D\x02\x00\xB0\xF5\x02\x00\xB1\xFD\x02\x00\xB2\xF5"
-     "\x02\x00"
-     "\xB3\x49\x02\x00\xB4\x59\x02\x00\xB5\x45\x02\x00\xB6\x45\x02\x00\xB7\x75"
-     "\x02\x00"
-     "\xB8\x4D\x02\x00\xB9\x25\x02\x00\xBA\x21\x02\x00\xBB\x21\x02\x00\xBC\x25"
-     "\x02\x00"
-     "\xBD\x2D\x02\x00\xBE\x11\x02\x00\xBF\x11\x02\x00\xA3\xE9\x04\x00\x45\xE6"
-     "\x00\x80"
-     "\x49\xE6\x00\x80\x4D\xE6\x00\x80\x51\xE6\x00\x80\xA6\x3D\x03\x00\xA5\xE5"
-     "\x04\x00"
-     "\x55\xE6\x00\x80\xAB\x19\x03\x00\xAA\x11\x03\x00\xBE\x2C\x3F\x00\x59\xE6"
-     "\x00\x80"
-     "\xAF\xA1\x03\x00\xAE\xBD\x03\x00\xAD\xBD\x03\x00\xAC\x01\x03\x00\x5D\xE6"
-     "\x00\x80"
-     "\xB3\x7D\x02\x00\x61\xE6\x00\x80\x65\xE6\x00\x80\xB6\x19\x02\x00\x69\xE6"
-     "\x00\x80"
-     "\x6D\xE6\x00\x80\xB5\x19\x02\x00\xBA\x3D\x02\x00\xBB\x0D\x02\x00\x71\xE6"
-     "\x00\x80"
-     "\x75\xE6\x00\x80\xBE\xD5\x02\x00\xBF\xDD\x02\x00\xBC\x15\x02\x00\xBD\xDD"
-     "\x02\x00"
-     "\x79\xE6\x00\x80\x7D\xE6\x00\x80\x81\xE6\x00\x80\x84\x00\x3C\x00\x85\xE6"
-     "\x00\x80"
-     "\x89\xE6\x00\x80\xEF\x40\x00\x00\xEF\xBC\x06\x00\xE1\x48\x06\x00\xE1\x50"
-     "\x07\x00"
-     "\xE3\x34\x01\x00\xE3\x70\x06\x00\x8D\xE6\x00\x80\x91\xE6\x00\x80\x95\xE6"
-     "\x00\x80"
-     "\x99\xE6\x00\x80\x80\x75\x00\x00\x81\x7D\x00\x00\x82\x75\x00\x00\xA3\xF5"
-     "\x02\x00"
-     "\xA1\xE6\x00\x80\xA5\x91\x02\x00\xA6\x91\x02\x00\xA5\xE6\x00\x80\x86\x40"
-     "\x3F\x00"
-     "\x87\xD4\x3C\x00\xAA\xB5\x02\x00\xAB\x85\x02\x00\xAC\x9D\x02\x00\xAD\x55"
-     "\x02\x00"
-     "\xAE\x5D\x02\x00\xAF\x55\x02\x00\xA8\x25\x3D\x00\xA9\x2D\x3D\x00\xAA\x31"
-     "\x3D\x00"
-     "\xAB\x31\x3D\x00\xAC\x55\x3D\x00\xAD\x5D\x3D\x00\xAE\x55\x3D\x00\xAF\x4D"
-     "\x3D\x00"
-     "\x2D\xE6\x00\x80\x9D\xE6\x00\x80\xA9\xE6\x00\x80\xAD\xE6\x00\x80\xB1\xE6"
-     "\x00\x80"
-     "\xB5\xE6\x00\x80\xB9\xE6\x00\x80\xBD\xE6\x00\x80\xB8\xF5\x3D\x00\xB9\xFD"
-     "\x3D\x00"
-     "\xBA\xF5\x3D\x00\xBB\x49\x3E\x00\xBC\x59\x3E\x00\xBD\x59\x3E\x00\xBE\x49"
-     "\x3E\x00"
-     "\xBF\x41\x3E\x00\xB0\x3D\x3D\x00\xB1\x0D\x3D\x00\xB2\x05\x3D\x00\xB3\xCD"
-     "\x3D\x00"
-     "\xB4\xD5\x3D\x00\xB5\xDD\x3D\x00\xB6\xD5\x3D\x00\xB7\xCD\x3D\x00\xA8\xE5"
-     "\x3E\x00"
-     "\xA9\x05\x3E\x00\xAA\x0D\x3E\x00\xAB\x19\x3E\x00\xAC\x09\x3E\x00\xAD\x91"
-     "\x3F\x00"
-     "\xAE\x91\x3F\x00\xAF\x91\x3F\x00\x84\xA8\x03\x00\xC1\xE6\x00\x80\xC5\xE6"
-     "\x00\x80"
-     "\xC9\xE6\x00\x80\xCD\xE6\x00\x80\xD1\xE6\x00\x80\xD5\xE6\x00\x80\xD9\xE6"
-     "\x00\x80"
-     "\xB8\xA9\x3F\x00\xB9\xA9\x3F\x00\xBA\x49\x3F\x00\xBB\x49\x3F\x00\xBC\x59"
-     "\x3F\x00"
-     "\xBD\x45\x3F\x00\xBE\x4D\x3F\x00\xBF\x6D\x3F\x00\xB0\xF1\x3F\x00\xB1\xF9"
-     "\x3F\x00"
-     "\xB2\xC5\x3F\x00\xB3\xAD\x3F\x00\xB4\xB9\x3F\x00\xB5\xB9\x3F\x00\xB6\xAD"
-     "\x3F\x00"
-     "\xB7\xA5\x3F\x00\xDD\xE6\x00\x80\xE1\xE6\x00\x80\xE5\xE6\x00\x80\xE9\xE6"
-     "\x00\x80"
-     "\xED\xE6\x00\x80\xF1\xE6\x00\x80\xEF\x80\x3D\x00\xF5\xE6\x00\x80\xE1\x08"
-     "\x3E\x00"
-     "\xF9\xE6\x00\x80\xE3\x64\x3D\x00\xFD\xE6\x00\x80\x80\x0D\x00\x00\x81\x71"
-     "\x00\x00"
-     "\x82\x71\x00\x00\x01\xE7\x00\x80\xB3\xC5\x3F\x00\x05\xE7\x00\x80\x09\xE7"
-     "\x00\x80"
-     "\x0D\xE7\x00\x80\x11\xE7\x00\x80\xB6\x51\x3E\x00\xB5\xC1\x3F\x00\x15\xE7"
-     "\x00\x80"
-     "\xBB\x7D\x3E\x00\xBA\x75\x3E\x00\x86\xC8\x00\x00\x87\xBC\x01\x00\xBF\x3D"
-     "\x3E\x00"
-     "\xBE\x3D\x3E\x00\xBD\x39\x3E\x00\xBC\x51\x3E\x00\xA3\x05\x3F\x00\x19\xE7"
-     "\x00\x80"
-     "\x1D\xE7\x00\x80\x21\xE7\x00\x80\x25\xE7\x00\x80\xA6\x91\x3E\x00\xA5\x01"
-     "\x3F\x00"
-     "\x29\xE7\x00\x80\xAB\xBD\x3E\x00\xAA\xB5\x3E\x00\x2D\xE7\x00\x80\x31\xE7"
-     "\x00\x80"
-     "\xAF\xFD\x3E\x00\xAE\xFD\x3E\x00\xAD\xF9\x3E\x00\xAC\x91\x3E\x00\x35\xE7"
-     "\x00\x80"
-     "\xB3\xED\x01\x00\x39\xE7\x00\x80\x3D\xE7\x00\x80\xB6\xE5\x01\x00\x41\xE7"
-     "\x00\x80"
-     "\x45\xE7\x00\x80\xB5\xED\x01\x00\xBA\x45\x01\x00\xBB\x4D\x01\x00\x49\xE7"
-     "\x00\x80"
-     "\x4D\xE7\x00\x80\xBE\x45\x01\x00\xBF\x4D\x01\x00\xBC\x55\x01\x00\xBD\x4D"
-     "\x01\x00"
-     "\xA8\xD1\x3D\x00\xA9\xD9\x3D\x00\xAA\x21\x02\x00\xAB\x21\x02\x00\xAC\x21"
-     "\x02\x00"
-     "\xAD\x21\x02\x00\xAE\x21\x02\x00\xAF\x21\x02\x00\xBE\x00\x04\x00\x51\xE7"
-     "\x00\x80"
-     "\x55\xE7\x00\x80\x59\xE7\x00\x80\x5D\xE7\x00\x80\x61\xE7\x00\x80\x65\xE7"
-     "\x00\x80"
-     "\x69\xE7\x00\x80\xB8\x45\x03\x00\xB9\x4D\x03\x00\xBA\x45\x03\x00\xBB\x5D"
-     "\x03\x00"
-     "\xBC\x45\x03\x00\xBD\x49\x03\x00\xBE\x7D\x03\x00\xBF\x71\x03\x00\xB0\xC5"
-     "\x02\x00"
-     "\xB1\xCD\x02\x00\xB2\xC5\x02\x00\xB3\xDD\x02\x00\xB4\xC5\x02\x00\xB5\xCD"
-     "\x02\x00"
-     "\xB6\xC5\x02\x00\xB7\x7D\x03\x00\x6D\xE7\x00\x80\x71\xE7\x00\x80\xA3\xA1"
-     "\x02\x00"
-     "\x75\xE7\x00\x80\xA5\xA1\x02\x00\x79\xE7\x00\x80\x84\x80\x03\x00\xA6\xA9"
-     "\x02\x00"
-     "\xBE\x2C\x05\x00\xBD\xE5\x00\x80\xAB\x01\x02\x00\xAA\x09\x02\x00\xAD\x01"
-     "\x02\x00"
-     "\xAC\x19\x02\x00\xAF\x01\x02\x00\xAE\x09\x02\x00\xE3\x48\x02\x00\x81\xE7"
-     "\x00\x80"
-     "\xE1\xB8\x01\x00\x85\xE7\x00\x80\x80\x55\x00\x00\x81\x59\x00\x00\x82\x4D"
-     "\x00\x00"
-     "\x89\xE7\x00\x80\x86\x60\x04\x00\x87\xE8\x04\x00\x8D\xE7\x00\x80\x91\xE7"
-     "\x00\x80"
-     "\xEF\x28\x02\x00\x95\xE7\x00\x80\x99\xE7\x00\x80\x9D\xE7\x00\x80\xA8\x51"
-     "\x02\x00"
-     "\xA9\x59\x02\x00\xAA\x8D\x02\x00\xAB\x85\x02\x00\xAC\x9D\x02\x00\xAD\x85"
-     "\x02\x00"
-     "\xAE\x85\x02\x00\xAF\xB5\x02\x00\xA1\xE7\x00\x80\xA5\xE7\x00\x80\xA9\xE7"
-     "\x00\x80"
-     "\xAD\xE7\x00\x80\x84\x38\x05\x00\xB1\xE7\x00\x80\xB5\xE7\x00\x80\xB9\xE7"
-     "\x00\x80"
-     "\xB8\x65\x01\x00\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x7D\x01\x00\xBC\x65"
-     "\x01\x00"
-     "\xBD\x6D\x01\x00\xBE\x65\x01\x00\xBF\xDD\x01\x00\xB0\xCD\x02\x00\xB1\xA5"
-     "\x02\x00"
-     "\xB2\xA1\x02\x00\xB3\xA1\x02\x00\xB4\xA5\x02\x00\xB5\xAD\x02\x00\xB6\x91"
-     "\x02\x00"
-     "\xB7\x91\x02\x00\xBD\xE7\x00\x80\xC1\xE7\x00\x80\xC5\xE7\x00\x80\xBE\x50"
-     "\x07\x00"
-     "\xC9\xE7\x00\x80\xCD\xE7\x00\x80\xEF\x48\x3E\x00\xEF\xF8\x3F\x00\xE1\x58"
-     "\x3E\x00"
-     "\xE1\x30\x3F\x00\xE3\x18\x01\x00\xE3\x30\x3E\x00\xD1\xE7\x00\x80\xD5\xE7"
-     "\x00\x80"
-     "\xD9\xE7\x00\x80\xDD\xE7\x00\x80\xB6\x51\x02\x00\xE1\xE7\x00\x80\xE5\xE7"
-     "\x00\x80"
-     "\xB5\x41\x02\x00\xE9\xE7\x00\x80\xB3\x51\x02\x00\x87\x98\x04\x00\x86\xFC"
-     "\x04\x00"
-     "\xBE\x35\x02\x00\xBF\x3D\x02\x00\xBC\x21\x02\x00\xBD\x21\x02\x00\xBA\x5D"
-     "\x02\x00"
-     "\xBB\x35\x02\x00\xF1\xE7\x00\x80\xF5\xE7\x00\x80\x7D\xE7\x00\x80\x82\xD5"
-     "\x07\x00"
-     "\x81\xC1\x07\x00\x80\xC9\x07\x00\xED\xE7\x00\x80\xF9\xE7\x00\x80\xFD\xE7"
-     "\x00\x80"
-     "\x01\xE8\x00\x80\xA8\xE5\x06\x00\xA9\xE9\x06\x00\xAA\xF5\x06\x00\xAB\xC5"
-     "\x06\x00"
-     "\xAC\xDD\x06\x00\xAD\xC1\x06\x00\xAE\xC1\x06\x00\xAF\xC1\x06\x00\xB0\x61"
-     "\x07\x00"
-     "\xB1\x61\x07\x00\xB2\x61\x07\x00\xB3\x61\x07\x00\xB4\x65\x07\x00\xB5\x6D"
-     "\x07\x00"
-     "\xB6\x65\x07\x00\xB7\xDD\x07\x00\xB8\xE5\x07\x00\xB9\xED\x07\x00\xBA\xF9"
-     "\x07\x00"
-     "\xBB\xF9\x07\x00\xBC\xC1\x07\x00\xBD\xC1\x07\x00\xBE\xC1\x07\x00\xBF\xC1"
-     "\x07\x00"
-     "\x05\xE8\x00\x80\x09\xE8\x00\x80\xAA\x25\x05\x00\xAB\x4D\x05\x00\xAC\x59"
-     "\x05\x00"
-     "\xAD\x59\x05\x00\xAE\x4D\x05\x00\xAF\x45\x05\x00\x0D\xE8\x00\x80\x11\xE8"
-     "\x00\x80"
-     "\x15\xE8\x00\x80\xA3\x29\x05\x00\x19\xE8\x00\x80\xA5\x39\x05\x00\xA6\x29"
-     "\x05\x00"
-     "\x1D\xE8\x00\x80\x21\xE8\x00\x80\xB3\x19\x06\x00\x25\xE8\x00\x80\x29\xE8"
-     "\x00\x80"
-     "\xB6\x25\x06\x00\x2D\xE8\x00\x80\x31\xE8\x00\x80\xB5\x31\x06\x00\xBA\xC1"
-     "\x07\x00"
-     "\xBB\xC1\x07\x00\x35\xE8\x00\x80\x39\xE8\x00\x80\xBE\xC5\x07\x00\xBF\xCD"
-     "\x07\x00"
-     "\xBC\xD1\x07\x00\xBD\xD1\x07\x00\xA3\xDD\x07\x00\x3D\xE8\x00\x80\x41\xE8"
-     "\x00\x80"
-     "\x45\xE8\x00\x80\x49\xE8\x00\x80\xA6\xE1\x07\x00\xA5\xF5\x07\x00\x4D\xE8"
-     "\x00\x80"
-     "\xAB\x05\x06\x00\xAA\x05\x06\x00\x51\xE8\x00\x80\x55\xE8\x00\x80\xAF\x09"
-     "\x06\x00"
-     "\xAE\x01\x06\x00\xAD\x15\x06\x00\xAC\x15\x06\x00\x80\x69\x00\x00\x81\x69"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x59\xE8\x00\x80\x5D\xE8\x00\x80\x84\x90\x01\x00\x61\xE8"
-     "\x00\x80"
-     "\x65\xE8\x00\x80\x86\xC0\x00\x00\x87\xEC\x01\x00\x69\xE8\x00\x80\x6D\xE8"
-     "\x00\x80"
-     "\x71\xE8\x00\x80\x75\xE8\x00\x80\x79\xE8\x00\x80\x7D\xE8\x00\x80\xA8\x79"
-     "\x06\x00"
-     "\xA9\x79\x06\x00\xAA\x9D\x01\x00\xAB\x95\x01\x00\xAC\x8D\x01\x00\xAD\x15"
-     "\x01\x00"
-     "\xAE\x1D\x01\x00\xAF\x3D\x01\x00\x84\xE8\x01\x00\x81\xE8\x00\x80\x85\xE8"
-     "\x00\x80"
-     "\x89\xE8\x00\x80\x8D\xE8\x00\x80\x91\xE8\x00\x80\x95\xE8\x00\x80\x99\xE8"
-     "\x00\x80"
-     "\xB8\x19\x01\x00\xB9\x2D\x01\x00\xBA\x25\x01\x00\xBB\xC9\x00\x00\xBC\xD9"
-     "\x00\x00"
-     "\xBD\xD9\x00\x00\xBE\xC9\x00\x00\xBF\xC1\x00\x00\xB0\x45\x01\x00\xB1\x4D"
-     "\x01\x00"
-     "\xB2\x59\x01\x00\xB3\x51\x01\x00\xB4\x31\x01\x00\xB5\x3D\x01\x00\xB6\x35"
-     "\x01\x00"
-     "\xB7\x29\x01\x00\x9D\xE8\x00\x80\xA1\xE8\x00\x80\xB3\x1D\x02\x00\xBE\x44"
-     "\x03\x00"
-     "\xB5\x0D\x02\x00\xA5\xE8\x00\x80\xA9\xE8\x00\x80\xB6\x9D\x02\x00\x84\x6C"
-     "\x0C\x00"
-     "\xAD\xE8\x00\x80\xBB\x81\x02\x00\xBA\x89\x02\x00\xBD\x8D\x02\x00\xBC\x85"
-     "\x02\x00"
-     "\xBF\xF1\x02\x00\xBE\xFD\x02\x00\xB1\xE8\x00\x80\xA6\xC5\x02\x00\xA5\x55"
-     "\x02\x00"
-     "\xB5\xE8\x00\x80\xA3\x45\x02\x00\xB9\xE8\x00\x80\xBD\xE8\x00\x80\xC1\xE8"
-     "\x00\x80"
-     "\xAF\xA9\x02\x00\xAE\xA5\x02\x00\xAD\xD5\x02\x00\xAC\xDD\x02\x00\xAB\xD9"
-     "\x02\x00"
-     "\xAA\xD1\x02\x00\xC5\xE8\x00\x80\xC9\xE8\x00\x80\x82\xFD\x00\x00\xB3\xED"
-     "\x03\x00"
-     "\x80\xD5\x00\x00\x81\xF5\x00\x00\xB6\xE5\x03\x00\xD1\xE8\x00\x80\xD5\xE8"
-     "\x00\x80"
-     "\xB5\xED\x03\x00\xBA\x05\x03\x00\xBB\x0D\x03\x00\x87\x60\x03\x00\x86\x0C"
-     "\x0C\x00"
-     "\xBE\x05\x03\x00\xBF\x0D\x03\x00\xBC\x15\x03\x00\xBD\x0D\x03\x00\xD9\xE8"
-     "\x00\x80"
-     "\xDD\xE8\x00\x80\xE1\xE8\x00\x80\xE5\xE8\x00\x80\xE9\xE8\x00\x80\xE1\xAC"
-     "\x01\x00"
-     "\xED\xE8\x00\x80\xE3\x6C\x01\x00\xF1\xE8\x00\x80\xF5\xE8\x00\x80\xF9\xE8"
-     "\x00\x80"
-     "\xEF\x18\x02\x00\xFD\xE8\x00\x80\x01\xE9\x00\x80\x05\xE9\x00\x80\x09\xE9"
-     "\x00\x80"
-     "\xA8\x59\x02\x00\xA9\x59\x02\x00\xAA\x61\x02\x00\xAB\x61\x02\x00\xAC\x61"
-     "\x02\x00"
-     "\xAD\x61\x02\x00\xAE\x61\x02\x00\xAF\x61\x02\x00\xBE\x68\x0D\x00\x0D\xE9"
-     "\x00\x80"
-     "\x11\xE9\x00\x80\x15\xE9\x00\x80\x19\xE9\x00\x80\x1D\xE9\x00\x80\x21\xE9"
-     "\x00\x80"
-     "\x25\xE9\x00\x80\xB8\x45\x01\x00\xB9\x51\x01\x00\xBA\x51\x01\x00\xBB\x69"
-     "\x01\x00"
-     "\xBC\x11\x01\x00\xBD\x19\x01\x00\xBE\x0D\x01\x00\xBF\x05\x01\x00\xB0\xC5"
-     "\x02\x00"
-     "\xB1\xCD\x02\x00\xB2\xC5\x02\x00\xB3\xDD\x02\x00\xB4\xC5\x02\x00\xB5\xCD"
-     "\x02\x00"
-     "\xB6\xC5\x02\x00\xB7\x7D\x01\x00\x29\xE9\x00\x80\xE1\x7C\x07\x00\x2D\xE9"
-     "\x00\x80"
-     "\xE3\xAC\x06\x00\x31\xE9\x00\x80\xE1\x50\x06\x00\x35\xE9\x00\x80\xE3\xD8"
-     "\x00\x00"
-     "\x39\xE9\x00\x80\x41\xE9\x00\x80\x45\xE9\x00\x80\xEF\xD8\x06\x00\xE1\x40"
-     "\x37\x00"
-     "\x49\xE9\x00\x80\x4D\xE9\x00\x80\xEF\xE4\x07\x00\x82\x61\x00\x00\xA3\xAD"
-     "\x02\x00"
-     "\x80\x59\x00\x00\x81\x55\x00\x00\xA6\xA5\x02\x00\x51\xE9\x00\x80\x55\xE9"
-     "\x00\x80"
-     "\xA5\xAD\x02\x00\xAA\x45\x02\x00\xAB\x4D\x02\x00\x86\xC8\x0C\x00\x87\xE4"
-     "\x0C\x00"
-     "\xAE\x45\x02\x00\xAF\x4D\x02\x00\xAC\x55\x02\x00\xAD\x4D\x02\x00\xA8\x95"
-     "\x0E\x00"
-     "\xA9\x9D\x0E\x00\xAA\xC9\x0E\x00\xAB\xC9\x0E\x00\xAC\xD9\x0E\x00\xAD\xD9"
-     "\x0E\x00"
-     "\xAE\xC9\x0E\x00\xAF\xC5\x0E\x00\x3D\xE9\x00\x80\x59\xE9\x00\x80\x5D\xE9"
-     "\x00\x80"
-     "\x61\xE9\x00\x80\x65\xE9\x00\x80\x69\xE9\x00\x80\x6D\xE9\x00\x80\x71\xE9"
-     "\x00\x80"
-     "\xB8\x69\x0F\x00\xB9\x69\x0F\x00\xBA\x79\x0F\x00\xBB\x79\x0F\x00\xBC\x69"
-     "\x0F\x00"
-     "\xBD\x69\x0F\x00\xBE\xD9\x0F\x00\xBF\xD5\x0F\x00\xB0\x81\x0E\x00\xB1\x81"
-     "\x0E\x00"
-     "\xB2\x81\x0E\x00\xB3\x9D\x0E\x00\xB4\x85\x0E\x00\xB5\x8D\x0E\x00\xB6\x85"
-     "\x0E\x00"
-     "\xB7\x59\x0F\x00\xB3\x41\x0E\x00\x75\xE9\x00\x80\x79\xE9\x00\x80\x7D\xE9"
-     "\x00\x80"
-     "\x81\xE9\x00\x80\xB6\x59\x0E\x00\xB5\x51\x0E\x00\x85\xE9\x00\x80\xBB\x45"
-     "\x0E\x00"
-     "\xBA\x45\x0E\x00\x89\xE9\x00\x80\x8D\xE9\x00\x80\xBF\x49\x0E\x00\xBE\x41"
-     "\x0E\x00"
-     "\xBD\x49\x0E\x00\xBC\x55\x0E\x00\x91\xE9\x00\x80\xA3\x05\x0E\x00\x95\xE9"
-     "\x00\x80"
-     "\x99\xE9\x00\x80\xA6\x1D\x0E\x00\x9D\xE9\x00\x80\xA1\xE9\x00\x80\xA5\x15"
-     "\x0E\x00"
-     "\xAA\x01\x0E\x00\xAB\x01\x0E\x00\xA5\xE9\x00\x80\xA9\xE9\x00\x80\xAE\x05"
-     "\x0E\x00"
-     "\xAF\x0D\x0E\x00\xAC\x11\x0E\x00\xAD\x0D\x0E\x00\x82\xED\x01\x00\xEF\xDC"
-     "\x0C\x00"
-     "\x80\xD1\x01\x00\x81\xE5\x01\x00\x84\x9C\x02\x00\xBE\x94\x01\x00\xAD\xE9"
-     "\x00\x80"
-     "\xB1\xE9\x00\x80\x86\x80\x00\x00\x87\x64\x01\x00\xB5\xE9\x00\x80\xB9\xE9"
-     "\x00\x80"
-     "\xBD\xE9\x00\x80\xE3\x04\x0D\x00\xC1\xE9\x00\x80\xE1\x88\x0D\x00\xB3\xF5"
-     "\x0E\x00"
-     "\xC5\xE9\x00\x80\xC9\xE9\x00\x80\xCD\xE9\x00\x80\xD1\xE9\x00\x80\xB6\x8D"
-     "\x0E\x00"
-     "\xB5\x8D\x0E\x00\xD5\xE9\x00\x80\xBB\x51\x01\x00\xBA\x49\x01\x00\xD9\xE9"
-     "\x00\x80"
-     "\xDD\xE9\x00\x80\xBF\xF5\x01\x00\xBE\x49\x01\x00\xBD\x41\x01\x00\xBC\x49"
-     "\x01\x00"
-     "\xA8\x25\x0E\x00\xA9\x2D\x0E\x00\xAA\x3D\x0E\x00\xAB\x35\x0E\x00\xAC\x45"
-     "\x0E\x00"
-     "\xAD\x4D\x0E\x00\xAE\xBD\x01\x00\xAF\xB5\x01\x00\x84\xAC\x01\x00\xE1\xE9"
-     "\x00\x80"
-     "\xE5\xE9\x00\x80\xE9\xE9\x00\x80\xED\xE9\x00\x80\xF1\xE9\x00\x80\xF5\xE9"
-     "\x00\x80"
-     "\xF9\xE9\x00\x80\xB8\x51\x01\x00\xB9\x59\x01\x00\xBA\x61\x01\x00\xBB\x61"
-     "\x01\x00"
-     "\xBC\x11\x01\x00\xBD\x1D\x01\x00\xBE\x15\x01\x00\xBF\x09\x01\x00\xB0\xCD"
-     "\x01\x00"
-     "\xB1\xD5\x01\x00\xB2\xDD\x01\x00\xB3\xD5\x01\x00\xB4\xCD\x01\x00\xB5\x71"
-     "\x01\x00"
-     "\xB6\x71\x01\x00\xB7\x71\x01\x00\xFD\xE9\x00\x80\x01\xEA\x00\x80\x05\xEA"
-     "\x00\x80"
-     "\xA3\xB9\x0D\x00\x09\xEA\x00\x80\xA5\xC1\x0D\x00\xA6\xC1\x0D\x00\x0D\xEA"
-     "\x00\x80"
-     "\x11\xEA\x00\x80\x15\xEA\x00\x80\xAA\x05\x02\x00\xAB\x1D\x02\x00\xAC\x05"
-     "\x02\x00"
-     "\xAD\x0D\x02\x00\xAE\x05\x02\x00\xAF\xB9\x02\x00\x80\x61\x00\x00\x81\x61"
-     "\x00\x00"
-     "\x82\x19\x00\x00\xBE\x0C\x03\x00\x19\xEA\x00\x80\x21\xEA\x00\x80\x87\x78"
-     "\x03\x00"
-     "\x86\xFC\x04\x00\x25\xEA\x00\x80\x29\xEA\x00\x80\x2D\xEA\x00\x80\x31\xEA"
-     "\x00\x80"
-     "\x35\xEA\x00\x80\x39\xEA\x00\x80\x3D\xEA\x00\x80\x41\xEA\x00\x80\x45\xEA"
-     "\x00\x80"
-     "\x49\xEA\x00\x80\x4D\xEA\x00\x80\x51\xEA\x00\x80\x55\xEA\x00\x80\xE3\x24"
-     "\x01\x00"
-     "\x59\xEA\x00\x80\xE1\xA4\x01\x00\x84\xEC\x04\x00\xEF\x10\x02\x00\x5D\xEA"
-     "\x00\x80"
-     "\x61\xEA\x00\x80\x65\xEA\x00\x80\x69\xEA\x00\x80\xCD\xE8\x00\x80\x6D\xEA"
-     "\x00\x80"
-     "\xB3\xED\x03\x00\x71\xEA\x00\x80\x75\xEA\x00\x80\x79\xEA\x00\x80\x7D\xEA"
-     "\x00\x80"
-     "\xB6\x91\x03\x00\xB5\xE1\x03\x00\x81\xEA\x00\x80\xBB\xBD\x03\x00\xBA\xB5"
-     "\x03\x00"
-     "\x85\xEA\x00\x80\x89\xEA\x00\x80\xBF\x59\x03\x00\xBE\x51\x03\x00\xBD\x95"
-     "\x03\x00"
-     "\xBC\x95\x03\x00\xA8\x1D\x02\x00\xA9\x2D\x02\x00\xAA\x25\x02\x00\xAB\x3D"
-     "\x02\x00"
-     "\xAC\x29\x02\x00\xAD\x51\x02\x00\xAE\x51\x02\x00\xAF\x51\x02\x00\x80\xFD"
-     "\x01\x00"
-     "\x81\xCD\x01\x00\x82\xC5\x01\x00\xBE\x64\x05\x00\x91\xEA\x00\x80\x95\xEA"
-     "\x00\x80"
-     "\x87\x04\x05\x00\x86\x1C\x04\x00\xB8\xF5\x02\x00\xB9\xFD\x02\x00\xBA\xF5"
-     "\x02\x00"
-     "\xBB\x4D\x01\x00\xBC\x55\x01\x00\xBD\x5D\x01\x00\xBE\x49\x01\x00\xBF\x41"
-     "\x01\x00"
-     "\xB0\x31\x02\x00\xB1\x39\x02\x00\xB2\x01\x02\x00\xB3\x01\x02\x00\xB4\xD5"
-     "\x02\x00"
-     "\xB5\xDD\x02\x00\xB6\xD5\x02\x00\xB7\xCD\x02\x00\x99\xEA\x00\x80\x9D\xEA"
-     "\x00\x80"
-     "\xA1\xEA\x00\x80\x84\x40\x04\x00\xA5\xEA\x00\x80\xA9\xEA\x00\x80\xAD\xEA"
-     "\x00\x80"
-     "\xEF\x70\x00\x00\xB1\xEA\x00\x80\xE1\x54\x0E\x00\xB5\xEA\x00\x80\xE3\xAC"
-     "\x00\x00"
-     "\xB5\x31\x04\x00\xB9\xEA\x00\x80\xBD\xEA\x00\x80\xC1\xEA\x00\x80\xC5\xEA"
-     "\x00\x80"
-     "\xC9\xEA\x00\x80\xCD\xEA\x00\x80\xA3\xE1\x02\x00\xD1\xEA\x00\x80\xA5\xED"
-     "\x02\x00"
-     "\xA6\x9D\x02\x00\xD5\xEA\x00\x80\xD9\xEA\x00\x80\xDD\xEA\x00\x80\xAA\xB9"
-     "\x02\x00"
-     "\xAB\xB1\x02\x00\xAC\x99\x02\x00\xAD\x99\x02\x00\xAE\x5D\x02\x00\xAF\x55"
-     "\x02\x00"
-     "\xA8\x55\x06\x00\xA9\x65\x06\x00\xAA\x6D\x06\x00\xAB\x65\x06\x00\xAC\x7D"
-     "\x06\x00"
-     "\xAD\x65\x06\x00\xAE\x6D\x06\x00\xAF\x65\x06\x00\x8D\xEA\x00\x80\xE1\xEA"
-     "\x00\x80"
-     "\xE5\xEA\x00\x80\xE9\xEA\x00\x80\xED\xEA\x00\x80\x82\xB9\x00\x00\x81\xB1"
-     "\x00\x00"
-     "\x80\xB9\x00\x00\xB8\xED\x06\x00\xB9\xF5\x06\x00\xBA\xFD\x06\x00\xBB\xF1"
-     "\x06\x00"
-     "\xBC\x95\x06\x00\xBD\x9D\x06\x00\xBE\x95\x06\x00\xBF\x8D\x06\x00\xB0\x1D"
-     "\x06\x00"
-     "\xB1\xE5\x06\x00\xB2\xED\x06\x00\xB3\xE5\x06\x00\xB4\xFD\x06\x00\xB5\xE5"
-     "\x06\x00"
-     "\xB6\xE1\x06\x00\xB7\xD5\x06\x00\xB3\xE5\x06\x00\xF1\xEA\x00\x80\xF5\xEA"
-     "\x00\x80"
-     "\xF9\xEA\x00\x80\xBE\x10\x00\x00\xB6\xE5\x06\x00\xB5\xF5\x06\x00\xFD\xEA"
-     "\x00\x80"
-     "\xBB\x21\x06\x00\xBA\xD9\x06\x00\x86\xE8\x00\x00\x87\xD0\x03\x00\xBF\x21"
-     "\x06\x00"
-     "\xBE\x39\x06\x00\xBD\x31\x06\x00\xBC\x39\x06\x00\x01\xEB\x00\x80\xA3\xA1"
-     "\x06\x00"
-     "\x05\xEB\x00\x80\x09\xEB\x00\x80\xA6\xA1\x06\x00\x0D\xEB\x00\x80\x11\xEB"
-     "\x00\x80"
-     "\xA5\xB1\x06\x00\xAA\x9D\x06\x00\xAB\x65\x06\x00\x15\xEB\x00\x80\x19\xEB"
-     "\x00\x80"
-     "\xAE\x7D\x06\x00\xAF\x65\x06\x00\xAC\x7D\x06\x00\xAD\x75\x06\x00\xB3\x81"
-     "\x06\x00"
-     "\x1D\xEB\x00\x80\x21\xEB\x00\x80\x25\xEB\x00\x80\x29\xEB\x00\x80\xB6\x85"
-     "\x06\x00"
-     "\xB5\x91\x06\x00\x2D\xEB\x00\x80\xBB\x89\x06\x00\xBA\x85\x06\x00\x31\xEB"
-     "\x00\x80"
-     "\x35\xEB\x00\x80\xBF\xF1\x06\x00\xBE\x89\x06\x00\xBD\x81\x06\x00\xBC\x89"
-     "\x06\x00"
-     "\xE3\xF0\x05\x00\x39\xEB\x00\x80\xE1\x0C\x05\x00\x3D\xEB\x00\x80\x41\xEB"
-     "\x00\x80"
-     "\x45\xEB\x00\x80\x49\xEB\x00\x80\x4D\xEB\x00\x80\x51\xEB\x00\x80\x55\xEB"
-     "\x00\x80"
-     "\x59\xEB\x00\x80\x5D\xEB\x00\x80\xEF\xA4\x04\x00\x82\xDD\x00\x00\x81\xDD"
-     "\x00\x00"
-     "\x80\xDD\x00\x00\xA8\xB1\x06\x00\xA9\xB1\x06\x00\xAA\x09\x06\x00\xAB\x09"
-     "\x06\x00"
-     "\xAC\x19\x06\x00\xAD\x05\x06\x00\xAE\x0D\x06\x00\xAF\x05\x06\x00\x61\xEB"
-     "\x00\x80"
-     "\x65\xEB\x00\x80\x69\xEB\x00\x80\x84\x34\x01\x00\xBE\x7C\x00\x00\x71\xEB"
-     "\x00\x80"
-     "\x86\x18\x1F\x00\x87\x8C\x00\x00\xB8\x11\x06\x00\xB9\x19\x06\x00\xBA\x29"
-     "\x06\x00"
-     "\xBB\x29\x06\x00\xBC\xF5\x01\x00\xBD\xFD\x01\x00\xBE\xF5\x01\x00\xBF\xED"
-     "\x01\x00"
-     "\xB0\x41\x06\x00\xB1\x4D\x06\x00\xB2\x5D\x06\x00\xB3\x55\x06\x00\xB4\x4D"
-     "\x06\x00"
-     "\xB5\x35\x06\x00\xB6\x3D\x06\x00\xB7\x35\x06\x00\x75\xEB\x00\x80\x79\xEB"
-     "\x00\x80"
-     "\x7D\xEB\x00\x80\xA3\x8D\x05\x00\x81\xEB\x00\x80\xA5\x9D\x05\x00\xA6\x89"
-     "\x05\x00"
-     "\x85\xEB\x00\x80\x89\xEB\x00\x80\x8D\xEB\x00\x80\xAA\x89\x05\x00\xAB\x85"
-     "\x05\x00"
-     "\xAC\x85\x05\x00\xAD\x8D\x05\x00\xAE\x85\x05\x00\xAF\xFD\x05\x00\xB6\xA9"
-     "\x02\x00"
-     "\x91\xEB\x00\x80\x84\x48\x03\x00\xB5\x95\x02\x00\x95\xEB\x00\x80\xB3\x99"
-     "\x02\x00"
-     "\x99\xEB\x00\x80\x9D\xEB\x00\x80\xBE\x49\x03\x00\xBF\x49\x03\x00\xBC\x79"
-     "\x03\x00"
-     "\xBD\x71\x03\x00\xBA\x79\x03\x00\xBB\x79\x03\x00\xA1\xEB\x00\x80\xA5\xEB"
-     "\x00\x80"
-     "\xB0\x89\x03\x00\xB1\x89\x03\x00\xB2\x99\x03\x00\xB3\x91\x03\x00\xB4\xB9"
-     "\x03\x00"
-     "\xB5\xB9\x03\x00\xB6\xDD\x03\x00\xB7\xD1\x03\x00\xB8\xF1\x03\x00\xB9\xF1"
-     "\x03\x00"
-     "\xBA\xF1\x03\x00\xBB\xF1\x03\x00\xBC\xD1\x03\x00\xBD\xD1\x03\x00\xBE\xD1"
-     "\x03\x00"
-     "\xBF\xD1\x03\x00\xA9\xEB\x00\x80\xAD\xEB\x00\x80\xB1\xEB\x00\x80\x1D\xEA"
-     "\x00\x80"
-     "\xB5\xEB\x00\x80\xB9\xEB\x00\x80\xBD\xEB\x00\x80\xC1\xEB\x00\x80\xA8\x21"
-     "\x02\x00"
-     "\xA9\x2D\x02\x00\xAA\x25\x02\x00\xAB\x39\x02\x00\xAC\x29\x02\x00\xAD\x1D"
-     "\x02\x00"
-     "\xAE\x15\x02\x00\xAF\xF9\x03\x00\xC5\xEB\x00\x80\xC9\xEB\x00\x80\xCD\xEB"
-     "\x00\x80"
-     "\xD1\xEB\x00\x80\xE3\xAC\x00\x00\xD5\xEB\x00\x80\xE1\xA8\x01\x00\xD9\xEB"
-     "\x00\x80"
-     "\xEF\x68\x00\x00\xBE\x28\x1D\x00\x85\x90\x1D\x00\xBF\x04\x1C\x00\x80\x3D"
-     "\x00\x00"
-     "\x81\x21\x00\x00\x82\x25\x00\x00\x4F\x1C\x00\x80\xA3\x8D\x03\x00\x53\x1C"
-     "\x00\x80"
-     "\xAC\x08\x02\x00\x57\x1C\x00\x80\x5B\x1C\x00\x80\xA6\xBD\x03\x00\xA5\x81"
-     "\x03\x00"
-     "\x5F\x1C\x00\x80\xAB\x6D\x02\x00\xAA\x6D\x02\x00\x63\x1C\x00\x80\x67\x1C"
-     "\x00\x80"
-     "\xAF\x5D\x02\x00\xAE\x5D\x02\x00\xAD\x65\x02\x00\xAC\x6D\x02\x00\x6B\x1C"
-     "\x00\x80"
-     "\xE1\x7C\x02\x00\x6F\x1C\x00\x80\xE1\xAC\x02\x00\xAC\xE4\x02\x00\xB5\x1D"
-     "\x00\x00"
-     "\x73\x1C\x00\x80\xEF\xF4\x01\x00\xB5\x39\x00\x00\xE1\x80\x06\x00\xAC\xBC"
-     "\x02\x00"
-     "\xE3\x30\x06\x00\xE1\xA8\x02\x00\x77\x1C\x00\x80\xE1\xC0\x02\x00\xE1\x2C"
-     "\x1D\x00"
-     "\xB5\x75\x00\x00\x7B\x1C\x00\x80\xAC\xFC\x02\x00\xB3\xC9\x01\x00\x7F\x1C"
-     "\x00\x80"
-     "\xB5\xC5\x01\x00\xB6\xC5\x01\x00\x83\x1C\x00\x80\xB5\x25\x00\x00\xE1\x7C"
-     "\x1D\x00"
-     "\xBA\xAD\x01\x00\xBB\xBD\x01\x00\xBC\xA9\x01\x00\xBD\xA9\x01\x00\xBE\x81"
-     "\x01\x00"
-     "\xBF\x81\x01\x00\xA8\xB5\x1D\x00\xA9\x05\x1E\x00\xAA\x0D\x1E\x00\xAB\x05"
-     "\x1E\x00"
-     "\xAC\x01\x1E\x00\xAD\x09\x1E\x00\xAE\x39\x1E\x00\xAF\x39\x1E\x00\xAC\xD0"
-     "\x02\x00"
-     "\xB5\xF5\x1F\x00\xAC\xD0\x02\x00\xB5\xF1\x1F\x00\xA2\xAD\x1C\x00\xA3\xE5"
-     "\x02\x00"
-     "\xA0\xF5\x1E\x00\xA1\x1D\x1C\x00\xB8\x89\x1E\x00\xB9\x95\x1E\x00\xBA\x9D"
-     "\x1E\x00"
-     "\xBB\x95\x1E\x00\xBC\x8D\x1E\x00\xBD\x55\x1F\x00\xBE\x5D\x1F\x00\xBF\x55"
-     "\x1F\x00"
-     "\xB0\x25\x1E\x00\xB1\x35\x1E\x00\xB2\x31\x1E\x00\xB3\x31\x1E\x00\xB4\x09"
-     "\x1E\x00"
-     "\xB5\x09\x1E\x00\xB6\xB9\x1E\x00\xB7\xB9\x1E\x00\xA3\x85\x1E\x00\x8F\xA1"
-     "\x24\x00"
-     "\xE1\xB4\x02\x00\xAC\xC8\x02\x00\xB5\x0D\x00\x00\xA6\x89\x1E\x00\xA5\x89"
-     "\x1E\x00"
-     "\x87\x1C\x00\x80\xAB\xF1\x1E\x00\xAA\xE1\x1E\x00\xE1\xB4\x02\x00\xAC\x70"
-     "\x03\x00"
-     "\xAF\xCD\x1E\x00\xAE\xCD\x1E\x00\xAD\xE5\x1E\x00\xAC\xE5\x1E\x00\x9E\xE1"
-     "\x18\x00"
-     "\x9F\xDD\x19\x00\x9C\x99\x15\x00\x9D\x75\x1A\x00\x9A\xD1\x14\x00\x9B\xC5"
-     "\x14\x00"
-     "\x98\xFD\x10\x00\x99\x91\x11\x00\x96\x45\x12\x00\x97\xC5\x10\x00\x94\xC9"
-     "\x2C\x00"
-     "\x95\xC9\x2C\x00\x92\x99\x28\x00\x93\x65\x2E\x00\x90\xD1\x26\x00\x91\xF9"
-     "\x28\x00"
-     "\x8A\x25\x2D\x00\x8B\xA1\x13\x00\xB5\x05\x00\x00\x8B\x1C\x00\x80\x8E\x5D"
-     "\x10\x00"
-     "\x8F\x15\x15\x00\x8C\x39\x11\x00\x8D\x39\x11\x00\x82\x05\x25\x00\x83\x0D"
-     "\x27\x00"
-     "\xE1\xB4\x02\x00\xAC\xB4\x02\x00\x86\xA1\x2F\x00\x87\x15\x2D\x00\x84\x3D"
-     "\x29\x00"
-     "\x85\x55\x29\x00\x9A\x61\x1D\x00\x9B\xD9\x1D\x00\x8F\x1C\x00\x80\xB5\x05"
-     "\x00\x00"
-     "\xC3\x0C\x00\x80\xE1\xF4\x1E\x00\x9C\x89\x03\x00\xE3\x04\x1D\x00\x92\x71"
-     "\x15\x00"
-     "\x93\x35\x14\x00\x83\x3B\x00\x80\xEF\xB0\x1C\x00\x96\x69\x18\x00\x97\xB9"
-     "\x1F\x00"
-     "\x94\xC1\x1B\x00\x95\x5D\x19\x00\xB8\xBD\x1E\x00\xB9\xC1\x1E\x00\xBA\xC5"
-     "\x1E\x00"
-     "\xBB\xDD\x1E\x00\xBC\xCD\x1E\x00\xBD\xFD\x1E\x00\xDF\x18\x00\x80\xE3\x18"
-     "\x00\x80"
-     "\xE7\x18\x00\x80\xEB\x18\x00\x80\xEF\x18\x00\x80\xF3\x18\x00\x80\xF7\x18"
-     "\x00\x80"
-     "\xFB\x18\x00\x80\xFF\x18\x00\x80\x03\x19\x00\x80\x07\x19\x00\x80\x0B\x19"
-     "\x00\x80"
-     "\x13\x19\x00\x80\x17\x19\x00\x80\x1B\x19\x00\x80\x1F\x19\x00\x80\x23\x19"
-     "\x00\x80"
-     "\x27\x19\x00\x80\x2B\x19\x00\x80\x2F\x19\x00\x80\x33\x19\x00\x80\x37\x19"
-     "\x00\x80"
-     "\x3B\x19\x00\x80\x3F\x19\x00\x80\x43\x19\x00\x80\x47\x19\x00\x80\x85\xB4"
-     "\x3B\x00"
-     "\x4B\x19\x00\x80\x87\x3C\x16\x00\xCC\xF0\x02\x00\x4F\x19\x00\x80\x53\x19"
-     "\x00\x80"
-     "\x57\x19\x00\x80\xB8\xC5\x08\x00\x8D\x14\x1E\x00\x5B\x19\x00\x80\x5F\x19"
-     "\x00\x80"
-     "\x63\x19\x00\x80\x89\x30\x1E\x00\x67\x19\x00\x80\x8B\x30\x1E\x00\xB8\xDD"
-     "\x08\x00"
-     "\x6B\x19\x00\x80\xCC\x98\x02\x00\x6F\x19\x00\x80\x73\x19\x00\x80\x91\x6C"
-     "\x1E\x00"
-     "\x77\x19\x00\x80\x7B\x19\x00\x80\x7F\x19\x00\x80\x83\x19\x00\x80\x87\x19"
-     "\x00\x80"
-     "\x8B\x19\x00\x80\x8F\x19\x00\x80\x93\x19\x00\x80\xC3\x0A\x00\x80\xC6\x0A"
-     "\x00\x80"
-     "\x93\x1C\x00\x80\xBC\x79\x37\x02\xBD\x55\x3C\x02\xBE\xA1\x3A\x02\xBF\xBD"
-     "\x3A\x02"
-     "\xB8\x7D\x36\x02\xB9\xA9\x37\x02\xBA\x09\x37\x02\xE4\x0A\x00\x80\xB4\x99"
-     "\x36\x02"
-     "\xB5\xF1\x36\x02\xB6\x1D\x36\x02\xB7\x69\x36\x02\xB0\xB9\x36\x02\xB1\xA5"
-     "\x36\x02"
-     "\xB2\x65\x3F\x02\xB3\x8D\x36\x02\xAC\xF5\x32\x02\xAD\x41\x3E\x02\xAE\x05"
-     "\x31\x02"
-     "\x97\x1C\x00\x80\x6D\xEB\x00\x80\x9B\x1C\x00\x80\xA3\x1C\x00\x80\xA7\x1C"
-     "\x00\x80"
-     "\xC9\x0A\x00\x80\xBC\x84\x1E\x00\xBF\xC4\x01\x00\xBE\xE0\x01\x00\xB8\x19"
-     "\x08\x00"
-     "\xD2\x0A\x00\x80\x8C\x69\x70\x00\xBA\xC4\x1E\x00\x80\xD9\x28\x02\x81\xD9"
-     "\x28\x02"
-     "\x82\x59\x02\x00\x83\xD1\x2E\x02\x84\xF1\x2E\x02\x85\xF5\x2E\x02\x86\x39"
-     "\x2D\x02"
-     "\x87\x81\x2D\x02\x88\x49\x02\x00\xEA\x0A\x00\x80\xCC\x38\x03\x00\xCC\xA4"
-     "\x07\x00"
-     "\xF0\x0A\x00\x80\xFC\x0A\x00\x80\x0B\x0B\x00\x80\xAF\x1C\x00\x80\x90\xD9"
-     "\x02\x00"
-     "\x91\xD9\x02\x00\x92\xE9\x02\x00\x93\xE9\x02\x00\x94\xF9\x02\x00\x95\xF9"
-     "\x02\x00"
-     "\x96\xE9\x02\x00\x97\xE9\x02\x00\x98\xD9\x02\x00\x99\xD9\x02\x00\x9A\xA9"
-     "\x02\x00"
-     "\x9B\xA9\x02\x00\x9C\x19\x1B\x02\x9D\xBD\x02\x00\x9E\xB5\x02\x00\x9F\xAD"
-     "\x02\x00"
-     "\xA0\x55\x02\x00\xA1\x5D\x02\x00\xA2\x55\x02\x00\xA3\x6D\x02\x00\xA4\x75"
-     "\x02\x00"
-     "\xA5\x7D\x02\x00\xA6\x75\x02\x00\xA7\x6D\x02\x00\xA8\x55\x02\x00\xA9\x5D"
-     "\x02\x00"
-     "\xAA\x55\x02\x00\xAB\xA9\x01\x00\xAC\xA1\x01\x00\xAD\xA1\x01\x00\xAE\xA1"
-     "\x01\x00"
-     "\xAF\x99\x01\x00\xB0\xE9\x01\x00\xB1\xE9\x01\x00\xB2\xFD\x01\x00\xB3\x8D"
-     "\x01\x00"
-     "\xB4\x99\x01\x00\xB5\x99\x01\x00\xB6\x8D\x01\x00\xB7\x85\x01\x00\xB8\xBD"
-     "\x01\x00"
-     "\xB9\x85\x01\x00\xBA\x8D\x01\x00\x89\x0B\x00\x80\x14\x0B\x00\x80\xBD\x9D"
-     "\x01\x00"
-     "\xBE\x95\x01\x00\xBF\x8D\x01\x00\x53\x08\x00\x80\x9B\xE5\x06\x00\x1D\x07"
-     "\x00\x80"
-     "\xB7\x1C\x00\x80\x9E\xA1\x38\x02\x9F\x89\x06\x00\x9C\xE9\x31\x02\x9D\xE9"
-     "\x06\x00"
-     "\x8F\x0B\x00\x80\x20\x0B\x00\x80\x1D\x0B\x00\x80\x26\x0B\x00\x80\x29\x0B"
-     "\x00\x80"
-     "\x35\x0B\x00\x80\x4D\x0B\x00\x80\xF1\x09\x00\x80\x9F\x1C\x00\x80\x71\x0B"
-     "\x00\x80"
-     "\x74\x0B\x00\x80\x2F\x1D\x00\x80\x77\x1D\x00\x80\x87\x1D\x00\x80\x74\x06"
-     "\x00\x80"
-     "\xB0\x08\x00\x80\xFF\x1D\x00\x80\x2F\x1E\x00\x80\x57\x1E\x00\x80\xB3\x1E"
-     "\x00\x80"
-     "\xB7\x1E\x00\x80\xBB\x1E\x00\x80\xBF\x1E\x00\x80\xC3\x1E\x00\x80\xBA\x01"
-     "\x3B\x02"
-     "\xBB\x11\x39\x02\xB8\xC9\x00\x00\xB9\x05\x3A\x02\xBE\x99\x24\x02\xBF\x15"
-     "\x26\x02"
-     "\xBC\xDD\x00\x00\xBD\xD5\x00\x00\xB2\x95\x00\x00\xB3\x9D\x00\x00\xB0\x99"
-     "\x00\x00"
-     "\xB1\xB1\x3A\x02\xB6\xCD\x3B\x02\xB7\x31\x3A\x02\xB4\xE5\x3A\x02\xB5\x81"
-     "\x00\x00"
-     "\xAA\xCD\x00\x00\xAB\xE5\x00\x00\xA8\x25\x07\x00\xA9\x3D\x07\x00\xAE\xC9"
-     "\x3A\x02"
-     "\xAF\xF9\x00\x00\xAC\xFD\x00\x00\xAD\xF5\x00\x00\xA2\xD1\x07\x00\xA3\xD1"
-     "\x07\x00"
-     "\xA0\xC5\x30\x02\xA1\xC9\x07\x00\xA6\xC9\x07\x00\xA7\xC9\x07\x00\xA4\x01"
-     "\x39\x02"
-     "\xA5\x09\x39\x02\x80\x8D\x34\x02\x80\x6D\x30\x02\x82\x05\x37\x02\x82\x75"
-     "\x30\x02"
-     "\x84\x01\x37\x02\x84\xAD\x30\x02\x86\x0D\x37\x02\x86\xDD\x30\x02\x88\x65"
-     "\x37\x02"
-     "\x88\xF1\x30\x02\x8A\x79\x37\x02\x8A\xE5\x30\x02\x8C\x51\x37\x02\x8C\x1D"
-     "\x33\x02"
-     "\x8E\x49\x37\x02\x8E\x09\x33\x02\x90\x29\x37\x02\x90\x7D\x33\x02\x92\x25"
-     "\x37\x02"
-     "\x92\x29\x33\x02\x94\xF9\x37\x02\x94\xCD\x32\x02\x96\xF5\x37\x02\x96\xC1"
-     "\x32\x02"
-     "\x98\x71\x36\x02\x98\xC9\x32\x02\x9A\x7D\x36\x02\x9A\x3D\x01\x00\x9C\x79"
-     "\x36\x02"
-     "\x9C\xBD\x32\x02\x9E\x5D\x36\x02\x9E\xA9\x32\x02\xA0\xB5\x36\x02\xA0\x6D"
-     "\x32\x02"
-     "\xA2\xD1\x36\x02\xA2\xD9\x01\x00\xA4\xFD\x36\x02\xA4\x75\x32\x02\xA6\xC5"
-     "\x36\x02"
-     "\xA6\x71\x32\x02\xA8\x35\x36\x02\xA8\xB9\x3D\x02\xAA\x09\x36\x02\xAA\xC9"
-     "\x3D\x02"
-     "\xAC\x1D\x31\x02\xAC\xC1\x3D\x02\xAE\xCD\x31\x02\xAE\x2D\x3D\x02\xB0\x09"
-     "\x31\x02"
-     "\xB0\x49\x3D\x02\xB2\x1D\x31\x02\xB2\x99\x3C\x02\xB4\x35\x31\x02\xB4\xB5"
-     "\x3C\x02"
-     "\xB6\x95\x01\x00\xB6\x81\x3C\x02\xB8\xED\x30\x02\xB8\x49\x3D\x02\xBA\xD9"
-     "\x30\x02"
-     "\xBA\x55\x3D\x02\xBC\xD1\x30\x02\xBC\x89\x3C\x02\xBE\xAD\x30\x02\xBE\x75"
-     "\x3C\x02"
-     "\xC7\x1E\x00\x80\xCB\x1E\x00\x80\xCF\x1E\x00\x80\xD3\x1E\x00\x80\xD7\x1E"
-     "\x00\x80"
-     "\xDB\x1E\x00\x80\xDF\x1E\x00\x80\xE3\x1E\x00\x80\xF7\x1E\x00\x80\xFB\x1E"
-     "\x00\x80"
-     "\xFF\x1E\x00\x80\x03\x1F\x00\x80\x07\x1F\x00\x80\xB8\x81\x2B\x02\x0B\x1F"
-     "\x00\x80"
-     "\x0F\x1F\x00\x80\xA5\x25\x3F\x02\xA4\xD1\x3E\x02\xA7\x55\x27\x02\xA6\x59"
-     "\x27\x02"
-     "\xA1\xA9\x22\x02\xA0\xFD\x22\x02\xA3\x69\x3E\x02\xA2\x55\x02\x00\x13\x1F"
-     "\x00\x80"
-     "\x17\x1F\x00\x80\xB8\x99\x0C\x00\x1B\x1F\x00\x80\xA9\xD9\x22\x02\xA8\x6D"
-     "\x2A\x02"
-     "\x37\x1F\x00\x80\xAA\x85\x22\x02\x95\x55\x04\x00\x94\x49\x04\x00\x97\x89"
-     "\x02\x00"
-     "\x96\xA5\x3E\x02\x91\x19\x3B\x02\x90\xA9\x33\x02\x93\x59\x3B\x02\x92\x65"
-     "\x33\x02"
-     "\x9D\xB5\x27\x02\x9C\x79\x04\x00\x9F\x91\x27\x02\x9E\xB1\x27\x02\x99\x69"
-     "\x04\x00"
-     "\x98\x21\x3E\x02\x9B\xDD\x3F\x02\x9A\xB5\x38\x02\x85\xB9\x04\x00\x84\xB1"
-     "\x04\x00"
-     "\x87\x05\x35\x02\x86\xB1\x04\x00\x81\x45\x3E\x02\x80\x8D\x3E\x02\x83\xED"
-     "\x36\x02"
-     "\x82\x51\x3F\x02\x8D\xD1\x32\x02\x8C\x79\x04\x00\x8F\x2D\x32\x02\x3B\x1F"
-     "\x00\x80"
-     "\x89\xE1\x32\x02\x88\xFD\x32\x02\x8B\x1D\x3B\x02\x8A\x1D\x3B\x02\xB8\xF9"
-     "\x0D\x00"
-     "\x92\x4D\x3D\x02\x53\x1F\x00\x80\x90\x3D\x3D\x02\x6F\x1F\x00\x80\x73\x1F"
-     "\x00\x80"
-     "\x77\x1F\x00\x80\x94\x51\x3D\x02\x9B\xC9\x30\x02\x9A\x7D\x03\x00\x7B\x1F"
-     "\x00\x80"
-     "\x7F\x1F\x00\x80\xCC\x2C\x38\x02\x9E\x69\x03\x00\xCC\x60\x39\x02\xCD\x84"
-     "\x31\x02"
-     "\x83\x1F\x00\x80\x82\xDD\x3E\x02\x87\x1F\x00\x80\x80\xD9\x3E\x02\x8B\x1F"
-     "\x00\x80"
-     "\x86\xF5\x3E\x02\x38\x00\x00\x80\x84\xF9\x3E\x02\x70\x00\x00\x80\x8A\xA1"
-     "\x3E\x02"
-     "\x2A\x00\x00\x80\x88\xC1\x3E\x02\x3C\x00\x00\x80\x8E\x9D\x3E\x02\x42\x00"
-     "\x00\x80"
-     "\x8C\xB5\x3E\x02\x48\x00\x00\x80\xB2\x3D\x37\x02\x52\x00\x00\x80\xB0\x25"
-     "\x37\x02"
-     "\x58\x00\x00\x80\xB6\x1D\x36\x02\x5E\x00\x00\x80\xB4\x29\x37\x02\x1E\x00"
-     "\x00\x80"
-     "\xBA\x71\x35\x02\x60\x06\x00\x80\xB8\xC9\x03\x00\x4E\x00\x00\x80\xBE\x25"
-     "\x35\x02"
-     "\x7F\x20\x00\x80\xBC\x1D\x35\x02\x6C\x00\x00\x80\xA2\x35\x37\x02\x41\x03"
-     "\x00\x80"
-     "\xA0\xAD\x03\x00\x43\x03\x00\x80\xA6\x99\x37\x02\x46\x03\x00\x80\xA4\xB1"
-     "\x37\x02"
-     "\x8B\x20\x00\x80\xAA\x6D\x37\x02\x92\x20\x00\x80\xA8\x7D\x37\x02\xA9\x20"
-     "\x00\x80"
-     "\xAE\x59\x37\x02\xB0\x20\x00\x80\xAC\x79\x36\x02\x88\x79\x0C\x00\x89\x79"
-     "\x0C\x00"
-     "\x8A\x09\x0E\x00\x8B\x09\x0E\x00\x8C\x19\x0E\x00\x8D\x19\x0E\x00\x8E\x09"
-     "\x0E\x00"
-     "\x8F\x09\x0E\x00\x88\xE5\x28\x02\x89\xDD\x28\x02\x8A\x85\x28\x02\x8B\xE1"
-     "\x28\x02"
-     "\x8C\x85\x28\x02\x8D\xE1\x28\x02\x8E\x85\x28\x02\x8F\xCD\x28\x02\x98\x59"
-     "\x0E\x00"
-     "\x99\x59\x0E\x00\x9A\x19\x29\x02\x9B\x25\x29\x02\x9C\x31\x29\x02\x9D\x25"
-     "\x29\x02"
-     "\x0A\x0C\x00\x80\xC8\x00\x00\x80\x98\xA9\x29\x02\x99\x35\x2A\x02\x9A\xC9"
-     "\x29\x02"
-     "\x9B\xB5\x29\x02\x9C\xC1\x29\x02\x9D\x8D\x29\x02\x9E\xB9\x29\x02\x9F\x6D"
-     "\x2A\x02"
-     "\xA8\xA1\x0E\x00\xA9\xA1\x0E\x00\xAA\xA1\x0E\x00\xAB\xA1\x0E\x00\xAC\xA1"
-     "\x0E\x00"
-     "\xAD\xAD\x0E\x00\xAE\xA5\x0E\x00\xAF\xFD\x0E\x00\xA8\xD9\x2D\x02\xA9\x6D"
-     "\x2D\x02"
-     "\xAA\xF1\x2D\x02\xAB\x3D\x2D\x02\xAC\xD1\x2D\x02\xAD\x1D\x2D\x02\xAE\xC9"
-     "\x2D\x02"
-     "\xAF\x05\x2D\x02\xB8\xA5\x0E\x00\xB9\xA9\x0E\x00\xBA\x31\x2A\x02\xBB\x8D"
-     "\x2B\x02"
-     "\xBC\xD1\x2B\x02\xBD\x9D\x2B\x02\xBE\xD1\x2B\x02\xBF\x8D\x2B\x02\xB8\x45"
-     "\x28\x02"
-     "\xB9\x79\x28\x02\xBA\x7D\x28\x02\xBB\xB9\x28\x02\xBC\xFD\x29\x02\xBD\xAD"
-     "\x0F\x00"
-     "\xBE\xA1\x00\x00\xBF\xA5\x0F\x00\x88\x79\x0F\x00\x89\xE9\x0F\x00\x8A\x85"
-     "\x2A\x02"
-     "\x8B\x91\x2A\x02\x8C\x85\x2A\x02\x8D\xB1\x2A\x02\x0E\x00\x00\x80\x51\x06"
-     "\x00\x80"
-     "\x4E\x06\x00\x80\xCB\x00\x00\x80\xB8\x0D\x02\x00\x9C\x20\x00\x80\x4C\x03"
-     "\x00\x80"
-     "\x74\x05\x00\x80\xAF\x05\x00\x80\xD5\x05\x00\x80\xEC\x05\x00\x80\x99\x4D"
-     "\x0E\x00"
-     "\x09\x06\x00\x80\x9B\x21\x2D\x02\x20\x06\x00\x80\x9D\x25\x2D\x02\x3B\x03"
-     "\x00\x80"
-     "\x9F\xF9\x2D\x02\x06\x21\x00\x80\xDA\x00\x00\x80\x41\x06\x00\x80\x37\x03"
-     "\x00\x80"
-     "\x3C\x08\x00\x80\x69\x06\x00\x80\x6D\x06\x00\x80\xA5\x08\x00\x80\xA8\x91"
-     "\x0E\x00"
-     "\xA9\x91\x0E\x00\xAA\x91\x0E\x00\xAB\xA5\x0E\x00\xAC\xBD\x0E\x00\xAD\xD5"
-     "\x0E\x00"
-     "\xAE\xD1\x0E\x00\xAF\xCD\x0E\x00\x89\x06\x00\x80\x8D\x06\x00\x80\xA9\x08"
-     "\x00\x80"
-     "\xE7\x07\x00\x80\x44\x08\x00\x80\x55\x08\x00\x80\x5D\x08\x00\x80\xBE\x20"
-     "\x00\x80"
-     "\xE5\x0E\x00\x80\xB9\xB9\x29\x02\xFC\x0E\x00\x80\xBB\x49\x29\x02\x09\x0F"
-     "\x00\x80"
-     "\xBD\x59\x29\x02\x11\x0F\x00\x80\xDC\x06\x00\x80\x02\x21\x00\x80\xB1\xB1"
-     "\x29\x02"
-     "\x61\x04\x00\x80\xB3\x89\x29\x02\x6D\x04\x00\x80\xB5\x99\x29\x02\x3C\x04"
-     "\x00\x80"
-     "\xB7\x89\x29\x02\xB8\x35\x01\x00\xBA\x69\x0E\x00\x49\x04\x00\x80\xB8\xED"
-     "\x33\x02"
-     "\x06\x05\x00\x80\xBE\x61\x0E\x00\x55\x04\x00\x80\xBC\x6D\x0E\x00\xCC\xC0"
-     "\x0F\x00"
-     "\xB2\xF5\x33\x02\x0C\x05\x00\x80\xB0\x69\x33\x02\x59\x05\x00\x80\xB6\xC5"
-     "\x33\x02"
-     "\x9A\x05\x00\x80\xB4\xC9\x33\x02\x5B\x04\x00\x80\xAA\x65\x33\x02\xBA\x05"
-     "\x00\x80"
-     "\xA8\x69\x33\x02\x67\x04\x00\x80\xAE\x7D\x33\x02\xCC\x98\x38\x02\xAC\x6D"
-     "\x33\x02"
-     "\x5F\x05\x00\x80\xA2\x11\x3D\x02\xC0\x05\x00\x80\xA0\x61\x3D\x02\xF4\x05"
-     "\x00\x80"
-     "\xA6\x71\x33\x02\x37\x04\x00\x80\xA4\x2D\x33\x02\x1F\x07\x00\x80\x9A\x79"
-     "\x3C\x02"
-     "\x22\x07\x00\x80\x98\x0D\x3D\x02\x2F\x07\x00\x80\x9E\xC5\x3D\x02\x64\x08"
-     "\x00\x80"
-     "\x9C\x65\x3C\x02\x67\x08\x00\x80\x92\x59\x3F\x02\x6B\x08\x00\x80\x90\xA9"
-     "\x3C\x02"
-     "\x32\x07\x00\x80\x96\x95\x3C\x02\x83\x08\x00\x80\x94\x99\x3C\x02\x86\x08"
-     "\x00\x80"
-     "\x8A\x61\x3E\x02\xC2\x20\x00\x80\x88\xB5\x3E\x02\x54\x03\x00\x80\x8E\x05"
-     "\x3C\x02"
-     "\xD6\x20\x00\x80\x8C\xB5\x0E\x00\xDA\x20\x00\x80\x82\x79\x38\x02\xDE\x20"
-     "\x00\x80"
-     "\x80\x65\x38\x02\xE2\x20\x00\x80\x86\x05\x38\x02\xEA\x20\x00\x80\x84\x69"
-     "\x38\x02"
-     "\x8A\x89\x2A\x02\x8B\x69\x2B\x02\x88\x49\x2A\x02\x89\x69\x2B\x02\x8E\x69"
-     "\x0C\x00"
-     "\x8F\x95\x0C\x00\x8C\x79\x29\x02\x8D\x6D\x0C\x00\xCC\xE8\x00\x00\xE6\x20"
-     "\x00\x80"
-     "\x80\x49\x0C\x00\x81\x45\x0C\x00\xCC\xD0\x03\x00\xF2\x20\x00\x80\xCC\xB8"
-     "\x02\x00"
-     "\xFA\x20\x00\x80\x9A\xC9\x29\x02\x9B\x31\x2B\x02\x98\xF9\x29\x02\x99\xDD"
-     "\x29\x02"
-     "\x9E\xC1\x0C\x00\x9F\xCD\x0C\x00\xCC\xC8\x02\x00\x9D\xD5\x0C\x00\xB8\x5D"
-     "\x00\x00"
-     "\x93\x95\x2C\x02\xCC\xC0\x20\x02\xCD\x2C\x23\x02\xB8\x05\x00\x00\xFE\x20"
-     "\x00\x80"
-     "\xCC\xE4\x02\x00\xB8\xDD\x00\x00\xAA\x75\x28\x02\xAB\x01\x2B\x02\xA8\xB9"
-     "\x2F\x02"
-     "\xA9\x71\x28\x02\xAE\xDD\x2B\x02\xAF\x3D\x0C\x00\xAC\x59\x28\x02\xAD\x3D"
-     "\x0C\x00"
-     "\x0A\x21\x00\x80\xA3\x59\x22\x02\xCC\xA8\x02\x00\x94\xBC\x03\x00\x0E\x21"
-     "\x00\x80"
-     "\x92\xA4\x02\x00\x12\x21\x00\x80\x90\xA0\x0B\x00\xBA\xED\x2E\x02\xBB\x49"
-     "\x2B\x02"
-     "\xB8\x4D\x28\x02\xB9\xB1\x2B\x02\xBE\x61\x0C\x00\x19\x21\x00\x80\xBC\x01"
-     "\x2E\x02"
-     "\xBD\x75\x30\x02\xB8\x51\x00\x00\x30\x21\x00\x80\x37\x21\x00\x80\xCC\xF8"
-     "\x02\x00"
-     "\x41\x21\x00\x80\xB8\x05\x00\x00\xCC\x28\x05\x00\x45\x21\x00\x80\x80\xF5"
-     "\x36\x02"
-     "\x81\xF1\x36\x02\x82\xFD\x36\x02\x83\xF9\x36\x02\x84\xE5\x36\x02\x85\xE1"
-     "\x36\x02"
-     "\x86\xED\x36\x02\x87\xE9\x36\x02\x88\xD5\x36\x02\x89\xD1\x36\x02\x8A\xDD"
-     "\x36\x02"
-     "\x94\x34\x0C\x00\xB8\xCD\x07\x00\x92\x40\x0D\x00\xCC\x98\x02\x00\x90\x00"
-     "\x0D\x00"
-     "\xB8\x09\x00\x00\x91\xFD\x0B\x00\x49\x21\x00\x80\xCC\xFC\x02\x00\xB8\x05"
-     "\x00\x00"
-     "\x4D\x21\x00\x80\x89\x89\x06\x00\x97\x9D\x0B\x00\xCC\xE4\x02\x00\xB8\x0D"
-     "\x00\x00"
-     "\x59\x21\x00\x80\xCC\xF8\x02\x00\x83\xB9\x06\x00\xB8\x0D\x00\x00\x55\x21"
-     "\x00\x80"
-     "\xCC\x1C\x02\x00\xB8\x05\x00\x00\x65\x21\x00\x80\xCC\xE4\x02\x00\xB8\x2D"
-     "\x00\x00"
-     "\xA4\x59\x0B\x00\xA5\x45\x0B\x00\xA6\x41\x0B\x00\xCC\xD8\x02\x00\x69\x21"
-     "\x00\x80"
-     "\xB8\x0D\x00\x00\x6D\x21\x00\x80\xCC\xFC\x02\x00\xB8\x05\x00\x00\x71\x21"
-     "\x00\x80"
-     "\xCC\x98\x02\x00\xAF\x49\x36\x02\xB8\x05\x00\x00\x75\x21\x00\x80\xCC\xFC"
-     "\x02\x00"
-     "\xB3\x2D\x0B\x00\xB4\x39\x0B\x00\xB8\x35\x00\x00\xB6\x29\x0B\x00\xB7\x25"
-     "\x0B\x00"
-     "\x79\x21\x00\x80\xCC\xEC\x02\x00\xB8\x05\x00\x00\x7D\x21\x00\x80\xBC\xAD"
-     "\x04\x00"
-     "\xCC\xE8\x02\x00\xBE\xE1\x05\x00\xB8\xF9\x01\x00\xCC\xE8\x02\x00\x81\x21"
-     "\x00\x80"
-     "\xB8\x05\x00\x00\xA5\x21\x00\x80\xB8\x09\x00\x00\xCC\xE4\x02\x00\xA9\x21"
-     "\x00\x80"
-     "\xCC\xDC\x02\x00\xB8\x05\x00\x00\x85\x21\x00\x80\xCC\xE4\x02\x00\xB8\x19"
-     "\x00\x00"
-     "\x91\xB0\x7D\x00\x89\x21\x00\x80\x93\x38\x7B\x00\x92\x00\x78\x00\xCC\xE8"
-     "\x02\x00"
-     "\x8C\x84\x01\x00\xB8\x05\x00\x00\x8D\x21\x00\x80\x89\x5C\x7E\x00\x88\xC8"
-     "\x01\x00"
-     "\xCC\xC0\x02\x00\x8A\x98\x7E\x00\x85\x60\x7F\x00\x84\x00\x7C\x00\x87\x94"
-     "\x04\x00"
-     "\x86\x10\x03\x00\x81\x6C\x01\x00\x80\x74\x03\x00\xB8\xE9\x00\x00\x82\x7C"
-     "\x00\x00"
-     "\xBD\x00\x70\x00\xBC\x84\x76\x00\xBF\x88\x71\x00\xBE\x8C\x73\x00\x91\x21"
-     "\x00\x80"
-     "\xCC\xD0\x02\x00\xBB\xC0\x01\x00\xBA\x9C\x01\x00\xB5\x6C\x01\x00\xB8\x0D"
-     "\x00\x00"
-     "\xAD\x21\x00\x80\xCC\x8C\x02\x00\xB1\x30\x79\x00\xB0\xB4\x7A\x00\xB3\xB8"
-     "\x77\x00"
-     "\xB2\x00\x74\x00\xB8\x05\x00\x00\xB1\x21\x00\x80\xCC\xE4\x02\x00\xB8\x11"
-     "\x00\x00"
-     "\xA9\x9C\x7D\x00\xA8\x1C\x0D\x00\xAB\xF8\x01\x00\x95\x21\x00\x80\xB8\x09"
-     "\x00\x00"
-     "\xCC\xE4\x02\x00\x99\x21\x00\x80\xCC\xFC\x02\x00\xB8\x05\x00\x00\x9D\x21"
-     "\x00\x80"
-     "\xCC\xE4\x02\x00\xB8\xF9\x03\x00\x87\x1D\x07\x00\xA1\x21\x00\x80\xA5\x90"
-     "\x06\x00"
-     "\xCC\xF8\x02\x00\xA3\x58\x06\x00\xB8\x5D\x00\x00\xCC\xE4\x02\x00\xB8\x51"
-     "\x00\x00"
-     "\xAF\xC5\x05\x00\xCC\xF4\x02\x00\xAD\xFD\x05\x00\xAC\x95\x05\x00\xB8\x65"
-     "\x00\x00"
-     "\xB8\x6D\x00\x00\x89\x1D\x07\x00\x88\x1D\x07\x00\x97\x6D\x07\x00\xCC\x90"
-     "\x02\x00"
-     "\xB5\x21\x00\x80\xB9\x21\x00\x80\x9C\xFD\x7C\x00\xBD\x21\x00\x80\xC1\x21"
-     "\x00\x80"
-     "\xB0\xB5\x05\x00\x9F\x95\x34\x02\xA9\x11\x00\x00\xAA\x11\x00\x00\xBC\xB0"
-     "\x06\x00"
-     "\xAF\x9D\x78\x00\x43\x24\x00\x80\x47\x24\x00\x80\x9F\x8D\x78\x00\x67\x03"
-     "\x00\x80"
-     "\x82\x7C\x02\x00\x80\x03\x00\x80\x99\x03\x00\x80\x83\x90\x06\x00\xB3\x61"
-     "\x78\x00"
-     "\xB2\x03\x00\x80\x99\x09\x70\x00\x9E\x35\x70\x00\xBE\x7C\x02\x00\xC5\x03"
-     "\x00\x80"
-     "\xD8\x03\x00\x80\x8B\xD8\x06\x00\xEB\x03\x00\x80\xFE\x03\x00\x80\x88\xB4"
-     "\x06\x00"
-     "\xB7\x81\x07\x00\xB6\xF9\x07\x00\xB5\xF1\x07\x00\xB4\xF9\x07\x00\xBE\x34"
-     "\x03\x00"
-     "\x82\xC4\x03\x00\xB1\xD1\x0A\x00\xB0\xED\x07\x00\xBF\x71\x3B\x02\xBE\x89"
-     "\x07\x00"
-     "\xBD\x8D\x07\x00\xBC\x85\x07\x00\xBB\x91\x07\x00\xBA\xAD\x07\x00\xB9\xA9"
-     "\x07\x00"
-     "\xB8\xB5\x07\x00\x80\x31\x06\x00\x81\xD1\x34\x02\x82\x81\x34\x02\x83\x81"
-     "\x34\x02"
-     "\x84\x25\x06\x00\x85\x2D\x06\x00\x86\x25\x06\x00\x87\x5D\x06\x00\x88\x69"
-     "\x06\x00"
-     "\x89\x75\x06\x00\x8A\x71\x06\x00\x8B\x4D\x06\x00\x8C\x59\x06\x00\x8D\x51"
-     "\x06\x00"
-     "\x8E\x55\x06\x00\x11\x04\x00\x80\x90\x75\x3F\x02\x91\x55\x3B\x02\x92\xB1"
-     "\x3A\x02"
-     "\x93\xFD\x36\x02\x94\xB1\x32\x02\x95\x61\x2F\x02\x96\x9D\x3B\x02\x97\xB1"
-     "\x3B\x02"
-     "\x98\x5D\x3A\x02\x99\xF5\x3A\x02\x9A\x41\x37\x02\x9B\x2D\x37\x02\x9C\x71"
-     "\x0B\x00"
-     "\x99\xED\x71\x00\x24\x04\x00\x80\xBE\xC8\x02\x00\x8A\x04\x00\x80\xDC\x00"
-     "\x00\x80"
-     "\xE3\x04\x03\x00\x12\x01\x00\x80\x1B\x01\x00\x80\x1F\x01\x00\x80\xE3\x44"
-     "\x03\x00"
-     "\x9E\x05\x71\x00\xA8\xD1\x06\x00\x23\x01\x00\x80\xEF\x80\x03\x00\x82\x0C"
-     "\x02\x00"
-     "\x27\x01\x00\x80\x2B\x01\x00\x80\xEF\xE4\x03\x00\x2F\x01\x00\x80\xE3\x64"
-     "\x02\x00"
-     "\x99\x59\x71\x00\x33\x01\x00\x80\x9E\x51\x71\x00\xE0\x00\x00\x80\x99\x75"
-     "\x71\x00"
-     "\xE5\x00\x00\x80\xEA\x00\x00\x80\xEF\x00\x00\x80\xF4\x00\x00\x80\x82\x58"
-     "\x02\x00"
-     "\xF9\x00\x00\x80\xEF\x74\x02\x00\xFE\x00\x00\x80\x9E\x59\x71\x00\x03\x01"
-     "\x00\x80"
-     "\x80\xE1\x79\x00\x08\x01\x00\x80\x0D\x01\x00\x80\x16\x01\x00\x80\x64\x03"
-     "\x00\x80"
-     "\x85\x1D\x44\x02\x86\x11\x44\x02\x87\x0D\x44\x02\x88\x29\x74\x00\x89\x19"
-     "\x50\x02"
-     "\x87\x04\x00\x80\x26\x05\x00\x80\x79\x05\x00\x80\xB4\x05\x00\x80\xDA\x05"
-     "\x00\x80"
-     "\xF1\x05\x00\x80\x90\xA9\x79\x00\x91\xB5\x79\x00\x92\xB1\x79\x00\x93\x45"
-     "\x7A\x00"
-     "\x94\x41\x7A\x00\x95\x4D\x7A\x00\x96\x71\x7A\x00\x97\x6D\x7A\x00\x98\x59"
-     "\x7A\x00"
-     "\x99\x25\x7A\x00\x9A\x21\x7A\x00\x9B\x3D\x7A\x00\x9C\x11\x7A\x00\x9D\x1D"
-     "\x7A\x00"
-     "\x9E\x09\x7A\x00\x9F\x05\x7A\x00\xA0\xD5\x44\x02\xA1\x05\x7A\x00\xA2\x01"
-     "\x7A\x00"
-     "\xA3\x1D\x7A\x00\xA4\xF5\x49\x02\xA5\x0D\x7A\x00\xA6\x39\x7A\x00\xA7\x35"
-     "\x7A\x00"
-     "\xA8\x11\x7A\x00\xA9\x15\x49\x02\xAA\x11\x7A\x00\xAB\x6D\x7A\x00\xAC\x5D"
-     "\x44\x02"
-     "\xAD\x45\x45\x02\xAE\xB1\x44\x02\xAF\x81\x45\x02\xB0\xC9\x74\x00\xB1\x45"
-     "\x7A\x00"
-     "\xB2\x41\x7A\x00\xB3\x5D\x7A\x00\xB4\xB5\x49\x02\xB5\x4D\x7A\x00\xB6\x79"
-     "\x7A\x00"
-     "\xB7\x75\x7A\x00\xB8\x51\x7A\x00\xB9\x55\x49\x02\xBA\x51\x7A\x00\xBB\x2D"
-     "\x7A\x00"
-     "\xBC\x1D\x44\x02\xBD\x05\x45\x02\xBE\xF1\x44\x02\xBF\xC1\x45\x02\x83\xC8"
-     "\x79\x00"
-     "\x0E\x06\x00\x80\xB3\x20\x7C\x00\xB2\x30\x7C\x00\xB5\x64\x7C\x00\xB4\x00"
-     "\x7C\x00"
-     "\x85\x10\x7E\x00\x25\x06\x00\x80\x6A\x03\x00\x80\x83\x03\x00\x80\x9C\x03"
-     "\x00\x80"
-     "\x88\x50\x7E\x00\xB5\x03\x00\x80\xC8\x03\x00\x80\x8D\x68\x7E\x00\xDB\x03"
-     "\x00\x80"
-     "\xEE\x03\x00\x80\x01\x04\x00\x80\x14\x04\x00\x80\xA2\xB4\x7F\x00\x27\x04"
-     "\x00\x80"
-     "\x8D\x04\x00\x80\x37\x01\x00\x80\x3B\x01\x00\x80\xA9\xBC\x7F\x00\xA8\xB0"
-     "\x7F\x00"
-     "\xAB\xA0\x7F\x00\x3F\x01\x00\x80\x43\x01\x00\x80\x47\x01\x00\x80\x4B\x01"
-     "\x00\x80"
-     "\x4F\x01\x00\x80\x91\x24\x7F\x00\x53\x01\x00\x80\xA1\xD4\x7E\x00\x92\x10"
-     "\x7F\x00"
-     "\x57\x01\x00\x80\x5B\x01\x00\x80\xA5\x20\x7E\x00\xA4\xD0\x7E\x00\x5F\x01"
-     "\x00\x80"
-     "\x63\x01\x00\x80\x67\x01\x00\x80\x6B\x01\x00\x80\x6F\x01\x00\x80\x73\x01"
-     "\x00\x80"
-     "\x77\x01\x00\x80\x7B\x01\x00\x80\xB3\x5C\x7E\x00\xB2\x68\x7E\x00\x83\xD0"
-     "\x7F\x00"
-     "\x82\xE0\x7F\x00\xB7\x28\x7E\x00\xB6\x5C\x7E\x00\x87\xB0\x7F\x00\x86\xAC"
-     "\x7F\x00"
-     "\xBB\x24\x7E\x00\xBA\x30\x7E\x00\x7F\x01\x00\x80\x83\x01\x00\x80\x87\x01"
-     "\x00\x80"
-     "\x8B\x01\x00\x80\xBD\xDC\x7F\x00\xBC\xCC\x7F\x00\xA0\xD5\x49\x02\xA1\x8D"
-     "\x49\x02"
-     "\xA2\x85\x49\x02\xA3\x39\x7B\x00\xA4\x29\x7B\x00\xA5\x29\x7B\x00\xA6\x59"
-     "\x7B\x00"
-     "\xA7\x55\x7B\x00\xA8\x71\x7B\x00\xA9\xA5\x7C\x00\xAA\xA1\x7C\x00\xAB\xA1"
-     "\x7C\x00"
-     "\xAC\xA9\x7C\x00\xAD\x9D\x7C\x00\xAE\x95\x7C\x00\xAF\x89\x7C\x00\xB0\xF9"
-     "\x7C\x00"
-     "\xB1\x79\x7D\x00\xB2\x31\x7D\x00\xB3\xCD\x7D\x00\xB4\xD5\x7D\x00\xB5\xD9"
-     "\x7D\x00"
-     "\xB6\xC9\x7D\x00\xB7\xC9\x7D\x00\xB8\xC5\x7D\x00\xB9\xD5\x7D\x00\xBA\xDD"
-     "\x7D\x00"
-     "\xBB\xD1\x7D\x00\xBC\xB9\x7D\x00\xBD\xA5\x7D\x00\xBE\xA1\x7D\x00\xBF\xA1"
-     "\x7D\x00"
-     "\x8F\x01\x00\x80\x93\x01\x00\x80\x97\x01\x00\x80\x9B\x01\x00\x80\x23\x21"
-     "\x00\x80"
-     "\xB6\x3D\x00\x00\xB5\x3D\x00\x00\xB4\x3D\x00\x00\x39\x06\x00\x80\x43\x06"
-     "\x00\x80"
-     "\x46\x06\x00\x80\xD9\x21\x00\x80\xDF\x21\x00\x80\xE3\x21\x00\x80\xE7\x21"
-     "\x00\x80"
-     "\xEB\x21\x00\x80\xEF\x21\x00\x80\xF3\x21\x00\x80\xF7\x21\x00\x80\xFB\x21"
-     "\x00\x80"
-     "\xFF\x21\x00\x80\x03\x22\x00\x80\x07\x22\x00\x80\x0B\x22\x00\x80\x0F\x22"
-     "\x00\x80"
-     "\x13\x22\x00\x80\x17\x22\x00\x80\x1B\x22\x00\x80\x1F\x22\x00\x80\x23\x22"
-     "\x00\x80"
-     "\x27\x22\x00\x80\x2B\x22\x00\x80\x99\x29\x4B\x02\x2F\x22\x00\x80\x9B\x39"
-     "\x4B\x02"
-     "\x9A\xA9\x44\x02\x9D\x21\x4B\x02\x9C\x25\x4B\x02\x33\x22\x00\x80\x37\x22"
-     "\x00\x80"
-     "\x91\x75\x47\x02\x90\x71\x47\x02\x93\xC5\x47\x02\x92\xC1\x47\x02\x95\xA1"
-     "\x46\x02"
-     "\x3B\x22\x00\x80\x3F\x22\x00\x80\x96\x91\x7A\x00\x89\xE9\x7A\x00\x43\x22"
-     "\x00\x80"
-     "\x8B\x71\x47\x02\x8A\x7D\x47\x02\x8D\x69\x47\x02\x8C\x6D\x47\x02\x8F\x09"
-     "\x4E\x02"
-     "\x8E\x69\x53\x02\x81\x19\x7A\x00\x80\x11\x7A\x00\x83\x15\x7A\x00\x82\x9D"
-     "\x46\x02"
-     "\x85\xD1\x7A\x00\x47\x22\x00\x80\x87\x9D\x4E\x02\x86\xD9\x7A\x00\xB9\x89"
-     "\x77\x00"
-     "\xB8\x5D\x7A\x00\xBB\x55\x7A\x00\x4B\x22\x00\x80\xBD\x8D\x52\x02\xBC\x21"
-     "\x50\x02"
-     "\xBF\x2D\x50\x02\xBE\x81\x52\x02\xB1\x8D\x47\x02\xB0\xB5\x47\x02\xB3\x95"
-     "\x46\x02"
-     "\xB2\x11\x7A\x00\xB5\x71\x7A\x00\xB4\x4D\x46\x02\xB7\x69\x7A\x00\xB6\x75"
-     "\x7A\x00"
-     "\x4F\x22\x00\x80\xA8\x51\x4E\x02\xAB\xF1\x4F\x02\xAA\x09\x47\x02\xAD\x21"
-     "\x46\x02"
-     "\xAC\x19\x40\x02\xAF\xC9\x47\x02\x53\x22\x00\x80\xA1\x55\x7A\x00\xA0\x55"
-     "\x7A\x00"
-     "\x57\x22\x00\x80\xA2\x55\x7A\x00\x5B\x22\x00\x80\xA4\x61\x4E\x02\x5F\x22"
-     "\x00\x80"
-     "\xA6\x89\x5B\x02\x80\xC5\x40\x02\x81\xE9\x40\x02\x82\x35\x41\x02\x83\x9D"
-     "\x41\x02"
-     "\x84\xE9\x41\x02\x85\x3D\x4C\x02\x86\xBD\x43\x02\x87\x2D\x4C\x02\x88\x61"
-     "\x4C\x02"
-     "\x89\x7D\x4C\x02\x8A\xBD\x4C\x02\x8B\x49\x4D\x02\x8C\xBD\x4D\x02\x8D\x85"
-     "\x4D\x02"
-     "\x8E\xF5\x4D\x02\x8F\xA9\x49\x02\x90\x75\x44\x02\x91\x89\x47\x02\x92\xB9"
-     "\x41\x02"
-     "\x93\x45\x40\x02\x94\x41\x40\x02\x95\x79\x40\x02\x96\x69\x40\x02\x97\x69"
-     "\x54\x02"
-     "\x98\x69\x70\x00\x99\x59\x54\x02\x9A\xAD\x40\x02\x9B\x81\x40\x02\x9C\x4D"
-     "\x41\x02"
-     "\x9D\xE5\x41\x02\x9E\x81\x41\x02\x9F\x55\x4C\x02\xA0\x25\x43\x02\xA1\xB5"
-     "\x4C\x02"
-     "\xA2\xC9\x4C\x02\xA3\x8D\x70\x00\xA4\x05\x4C\x02\xA5\xF1\x4D\x02\xA6\xDD"
-     "\x54\x02"
-     "\xA7\x2D\x4D\x02\xA8\xA5\x55\x02\xA9\x31\x49\x02\xAA\x99\x7D\x00\x63\x22"
-     "\x00\x80"
-     "\x67\x22\x00\x80\x6B\x22\x00\x80\x6F\x22\x00\x80\x73\x22\x00\x80\x77\x22"
-     "\x00\x80"
-     "\x7B\x22\x00\x80\x7F\x22\x00\x80\x83\x22\x00\x80\x87\x22\x00\x80\x8B\x22"
-     "\x00\x80"
-     "\x8F\x22\x00\x80\x93\x22\x00\x80\x97\x22\x00\x80\x9B\x22\x00\x80\x9F\x22"
-     "\x00\x80"
-     "\x0E\x0B\x00\x80\xB3\x1C\x00\x80\x41\x0B\x00\x80\xAB\x22\x00\x80\xAF\x22"
-     "\x00\x80"
-     "\x80\xB1\x07\x00\x81\xB1\x07\x00\x82\xB1\x07\x00\x83\xB1\x07\x00\x84\xD1"
-     "\x07\x00"
-     "\x85\xD1\x07\x00\x86\xD1\x07\x00\x87\xD1\x07\x00\x88\xF1\x07\x00\x89\xF1"
-     "\x07\x00"
-     "\x8A\xF1\x07\x00\x8B\xF1\x07\x00\x8C\xD1\x07\x00\x8D\xD1\x07\x00\x8E\xD1"
-     "\x07\x00"
-     "\x8F\xD1\x07\x00\x90\xB1\x07\x00\x91\xB1\x07\x00\x92\xB1\x07\x00\x93\xB1"
-     "\x07\x00"
-     "\x94\x55\x06\x00\x95\x4D\x06\x00\x96\x45\x06\x00\x97\x69\x06\x00\x98\x59"
-     "\x06\x00"
-     "\x99\x21\x06\x00\x9A\x1D\x06\x00\x9B\xE1\x06\x00\x9C\xB1\x06\x00\x9D\xA9"
-     "\x06\x00"
-     "\x9E\x91\x06\x00\x9F\x8D\x06\x00\xA0\x79\x06\x00\xA1\x29\x01\x00\xA2\x39"
-     "\x01\x00"
-     "\xA3\x39\x01\x00\xA4\x29\x01\x00\xA5\x29\x01\x00\xA6\x59\x01\x00\xA7\x59"
-     "\x01\x00"
-     "\xA8\x69\x01\x00\xA9\x69\x01\x00\xAA\x79\x01\x00\xAB\x79\x01\x00\xAC\x69"
-     "\x01\x00"
-     "\xAD\x69\x01\x00\xAE\x59\x01\x00\xAF\x59\x01\x00\xB3\x22\x00\x80\xBB\x22"
-     "\x00\x80"
-     "\xBF\x22\x00\x80\xAB\x0A\x00\x80\xAE\x0A\x00\x80\xC3\x22\x00\x80\xC7\x22"
-     "\x00\x80"
-     "\xCB\x22\x00\x80\xCF\x22\x00\x80\xD3\x22\x00\x80\xD7\x22\x00\x80\xDB\x22"
-     "\x00\x80"
-     "\xDF\x22\x00\x80\xE3\x22\x00\x80\xE7\x22\x00\x80\xEB\x22\x00\x80\xA0\xE9"
-     "\x02\x00"
-     "\xEF\x22\x00\x80\xA2\xE5\x02\x00\xA3\xFD\x02\x00\xA4\xE5\x02\x00\xF3\x22"
-     "\x00\x80"
-     "\xF7\x22\x00\x80\xA7\xE5\x02\x00\xFB\x22\x00\x80\xA9\xD9\x02\x00\xFF\x22"
-     "\x00\x80"
-     "\xAB\xED\x03\x00\x03\x23\x00\x80\xAD\x4D\x74\x00\xAE\x1D\x72\x00\xAF\x49"
-     "\x74\x00"
-     "\xB0\xA1\x74\x00\x07\x23\x00\x80\xB2\x8D\x03\x00\x0B\x23\x00\x80\x0F\x23"
-     "\x00\x80"
-     "\xB5\x9D\x03\x00\x13\x23\x00\x80\x17\x23\x00\x80\x1B\x23\x00\x80\x1F\x23"
-     "\x00\x80"
-     "\x23\x23\x00\x80\x27\x23\x00\x80\xBC\x4D\x56\x02\xBD\x11\x4F\x02\xBE\xB5"
-     "\x03\x00"
-     "\xBF\xAD\x03\x00\x2B\x23\x00\x80\x2F\x23\x00\x80\x33\x23\x00\x80\x37\x23"
-     "\x00\x80"
-     "\x3B\x23\x00\x80\x3F\x23\x00\x80\x43\x23\x00\x80\x47\x23\x00\x80\x4B\x23"
-     "\x00\x80"
-     "\x4F\x23\x00\x80\x53\x23\x00\x80\x57\x23\x00\x80\x5B\x23\x00\x80\x5F\x23"
-     "\x00\x80"
-     "\x63\x23\x00\x80\x67\x23\x00\x80\x6B\x23\x00\x80\x6F\x23\x00\x80\x73\x23"
-     "\x00\x80"
-     "\x77\x23\x00\x80\x7B\x23\x00\x80\x7F\x23\x00\x80\x83\x23\x00\x80\x87\x23"
-     "\x00\x80"
-     "\x8B\x23\x00\x80\x8F\x23\x00\x80\x93\x23\x00\x80\x37\x24\x00\x80\xFB\x32"
-     "\x00\x80"
-     "\x67\x3B\x00\x80\xC7\x2B\x00\x80\xEB\x2B\x00\x80\x8F\x05\x06\x00\x8E\x39"
-     "\x07\x00"
-     "\x8D\x31\x07\x00\x8C\x39\x07\x00\x8B\x21\x07\x00\x8A\x39\x07\x00\x89\x31"
-     "\x07\x00"
-     "\x88\x0D\x07\x00\x87\x2D\x07\x00\x86\x2D\x07\x00\x85\x2D\x07\x00\x84\xD1"
-     "\x78\x00"
-     "\x83\xC5\x78\x00\x82\xD9\x78\x00\x81\xDD\x78\x00\x80\xDD\x78\x00\x9F\x31"
-     "\x06\x00"
-     "\x9E\x29\x06\x00\x9D\x15\x06\x00\x9C\x15\x06\x00\x9B\x05\x06\x00\x9A\x05"
-     "\x06\x00"
-     "\x99\x11\x06\x00\x98\x19\x06\x00\x97\x21\x06\x00\x96\x59\x06\x00\x95\x5D"
-     "\x06\x00"
-     "\x94\x5D\x06\x00\x93\x59\x06\x00\x92\x51\x06\x00\x91\x59\x06\x00\x90\x61"
-     "\x06\x00"
-     "\xAF\x6D\x06\x00\xAE\x6D\x06\x00\xAD\x35\x06\x00\xAC\x35\x06\x00\xAB\x25"
-     "\x06\x00"
-     "\xAA\x25\x06\x00\xA9\x35\x06\x00\xA8\x35\x06\x00\xA7\x05\x06\x00\xA6\x05"
-     "\x06\x00"
-     "\xA5\x3D\x06\x00\xA4\x3D\x06\x00\xA3\x35\x06\x00\xA2\xC9\x06\x00\xA1\xC1"
-     "\x06\x00"
-     "\xA0\xC9\x06\x00\xBF\x75\x43\x02\xBE\x91\x43\x02\xBD\xF1\x43\x02\xBC\xFD"
-     "\x43\x02"
-     "\xBB\xFD\x43\x02\xBA\xC5\x43\x02\xB9\xD1\x43\x02\xB8\x2D\x42\x02\xB7\x2D"
-     "\x44\x02"
-     "\xB6\xD1\x47\x02\xB5\x65\x05\x00\xB4\x65\x05\x00\xB3\x75\x05\x00\xB2\x75"
-     "\x05\x00"
-     "\xB1\x0D\x06\x00\xB0\x0D\x06\x00\x80\x99\x0E\x00\xF7\x2B\x00\x80\x82\x95"
-     "\x0E\x00"
-     "\x07\x2C\x00\x80\x84\xB1\x0E\x00\x0F\x2C\x00\x80\x86\xBD\x0E\x00\x1B\x2C"
-     "\x00\x80"
-     "\x88\x89\x0E\x00\x23\x2C\x00\x80\x8A\x85\x0E\x00\x2B\x2C\x00\x80\x8C\x81"
-     "\x0E\x00"
-     "\x37\x2C\x00\x80\x8E\x8D\x0E\x00\x92\x2C\x00\x80\x90\xF9\x0E\x00\xAA\x2C"
-     "\x00\x80"
-     "\x92\xF5\x0E\x00\xBA\x2C\x00\x80\x94\x91\x0E\x00\xCA\x2C\x00\x80\x96\x9D"
-     "\x0E\x00"
-     "\xDA\x2C\x00\x80\x98\xA9\x0E\x00\xEA\x2C\x00\x80\x9A\xA5\x0E\x00\x0A\x2D"
-     "\x00\x80"
-     "\x9C\xA1\x0E\x00\x0E\x2D\x00\x80\x9E\xAD\x0E\x00\x12\x2D\x00\x80\xA0\x59"
-     "\x0E\x00"
-     "\x46\x2D\x00\x80\xA2\x55\x0E\x00\x6E\x2D\x00\x80\x7E\x2D\x00\x80\x86\x2D"
-     "\x00\x80"
-     "\x8A\x2D\x00\x80\x96\x2D\x00\x80\xB2\x2D\x00\x80\xB6\x2D\x00\x80\xCE\x2D"
-     "\x00\x80"
-     "\xAB\x4D\x0E\x00\xD2\x2D\x00\x80\xAD\x51\x0E\x00\xDA\x2D\x00\x80\xEA\x2D"
-     "\x00\x80"
-     "\x86\x2E\x00\x80\x9A\x2E\x00\x80\xB2\x21\x0E\x00\xDA\x2E\x00\x80\xEE\x2E"
-     "\x00\x80"
-     "\xF6\x2E\x00\x80\xFA\x2E\x00\x80\x0A\x2F\x00\x80\x2D\x2F\x00\x80\x59\x2F"
-     "\x00\x80"
-     "\x65\x2F\x00\x80\x85\x2F\x00\x80\x91\x2F\x00\x80\x95\x2F\x00\x80\x99\x2F"
-     "\x00\x80"
-     "\xB1\x2F\x00\x80\x80\xF9\x0F\x00\x81\xF9\x0F\x00\x82\x0D\x01\x00\x83\x01"
-     "\x01\x00"
-     "\x84\x05\x01\x00\x85\x09\x01\x00\x86\x3D\x01\x00\x87\x31\x01\x00\x88\x15"
-     "\x01\x00"
-     "\x89\x19\x01\x00\x8A\x6D\x01\x00\x8B\x61\x01\x00\x8C\x65\x01\x00\x8D\x69"
-     "\x01\x00"
-     "\x8E\x5D\x01\x00\x8F\x51\x01\x00\x90\x35\x01\x00\x91\x39\x01\x00\x92\xCD"
-     "\x01\x00"
-     "\x93\xC1\x01\x00\x94\xC1\x01\x00\x95\xC1\x01\x00\x96\xC1\x01\x00\x97\xC1"
-     "\x01\x00"
-     "\x98\xC1\x01\x00\x99\xC1\x01\x00\x9A\xC1\x01\x00\x9B\xDD\x01\x00\x9C\xC9"
-     "\x01\x00"
-     "\x9D\xC9\x01\x00\x9E\xB9\x01\x00\x9F\xB9\x01\x00\xA0\x4D\x01\x00\xA1\x55"
-     "\x01\x00"
-     "\xA2\x5D\x01\x00\xA3\x55\x01\x00\xA4\x4D\x01\x00\xA5\x75\x01\x00\xA6\x7D"
-     "\x01\x00"
-     "\xA7\x75\x01\x00\xA8\x4D\x01\x00\xA9\x55\x01\x00\xAA\x5D\x01\x00\xAB\x55"
-     "\x01\x00"
-     "\xAC\x4D\x01\x00\xAD\xB1\x03\x00\xAE\xB1\x03\x00\xAF\xB1\x03\x00\xB0\xD1"
-     "\x03\x00"
-     "\xB1\xD1\x03\x00\xB2\xD1\x03\x00\xB3\xD1\x03\x00\xB4\xF1\x03\x00\xB5\xF1"
-     "\x03\x00"
-     "\xB6\xF1\x03\x00\xB7\xF1\x03\x00\xB8\xD1\x03\x00\xB9\xD1\x03\x00\xBA\xD1"
-     "\x03\x00"
-     "\xBB\xD1\x03\x00\xBC\xB1\x03\x00\xBD\xB1\x03\x00\xBE\xB1\x03\x00\xBF\xB1"
-     "\x03\x00"
-     "\x82\xE4\x05\x00\xB5\x2F\x00\x80\xE1\x2F\x00\x80\xED\x2F\x00\x80\xF5\x2F"
-     "\x00\x80"
-     "\xFD\x2F\x00\x80\x0D\x30\x00\x80\x1C\x30\x00\x80\x24\x30\x00\x80\x50\x30"
-     "\x00\x80"
-     "\x54\x30\x00\x80\x5C\x30\x00\x80\x60\x30\x00\x80\x68\x30\x00\x80\x70\x30"
-     "\x00\x80"
-     "\x80\x30\x00\x80\x9C\x30\x00\x80\x14\x31\x00\x80\x28\x31\x00\x80\x2C\x31"
-     "\x00\x80"
-     "\xAC\x31\x00\x80\xB0\x31\x00\x80\xC4\x31\x00\x80\xC8\x31\x00\x80\xD0\x31"
-     "\x00\x80"
-     "\xD8\x31\x00\x80\xE0\x31\x00\x80\xEC\x31\x00\x80\x22\x32\x00\x80\x32\x32"
-     "\x00\x80"
-     "\x46\x32\x00\x80\xC3\x32\x00\x80\xCF\x32\x00\x80\xDF\x32\x00\x80\xEB\x32"
-     "\x00\x80"
-     "\xF7\x32\x00\x80\xFF\x32\x00\x80\x03\x33\x00\x80\x07\x33\x00\x80\x0B\x33"
-     "\x00\x80"
-     "\x0F\x33\x00\x80\xC7\x33\x00\x80\x0F\x34\x00\x80\x1B\x34\x00\x80\x1F\x34"
-     "\x00\x80"
-     "\x23\x34\x00\x80\x27\x34\x00\x80\x2F\x34\x00\x80\x33\x34\x00\x80\x47\x34"
-     "\x00\x80"
-     "\x63\x34\x00\x80\x6B\x34\x00\x80\xAB\x34\x00\x80\xAF\x34\x00\x80\xB7\x34"
-     "\x00\x80"
-     "\xBB\x34\x00\x80\xC3\x34\x00\x80\xC7\x34\x00\x80\xEB\x34\x00\x80\xEF\x34"
-     "\x00\x80"
-     "\x0B\x35\x00\x80\x0F\x35\x00\x80\x13\x35\x00\x80\x17\x35\x00\x80\xBE\x80"
-     "\x76\x00"
-     "\xAE\x61\x0D\x00\x2B\x35\x00\x80\xAC\x6D\x0D\x00\x4F\x35\x00\x80\xAA\x49"
-     "\x0D\x00"
-     "\x53\x35\x00\x80\xA8\x45\x0D\x00\x57\x35\x00\x80\xA6\x75\x0D\x00\x7B\x35"
-     "\x00\x80"
-     "\xA4\x61\x0D\x00\xBB\x35\x00\x80\xA2\x71\x0D\x00\xBF\x35\x00\x80\xA0\x0D"
-     "\x0D\x00"
-     "\xDF\x35\x00\x80\xBE\xD1\x0D\x00\xEF\x35\x00\x80\xBC\xDD\x0D\x00\xF7\x35"
-     "\x00\x80"
-     "\xBA\xD9\x0D\x00\x17\x36\x00\x80\xB8\xD5\x0D\x00\x3B\x36\x00\x80\xB6\xE1"
-     "\x0D\x00"
-     "\x7F\x36\x00\x80\xB4\xED\x0D\x00\x87\x36\x00\x80\xB2\x19\x0D\x00\x9B\x36"
-     "\x00\x80"
-     "\xB0\x15\x0D\x00\xA7\x36\x00\x80\x8E\x75\x0E\x00\xAF\x36\x00\x80\x8C\x61"
-     "\x0E\x00"
-     "\xB7\x36\x00\x80\x8A\x7D\x0E\x00\xBB\x36\x00\x80\x88\x79\x0E\x00\xBF\x36"
-     "\x00\x80"
-     "\x86\x45\x0E\x00\xD7\x36\x00\x80\x84\x71\x0E\x00\xDB\x36\x00\x80\x82\x6D"
-     "\x0E\x00"
-     "\xF7\x36\x00\x80\x80\x69\x0E\x00\xFF\x36\x00\x80\x9E\xED\x0D\x00\x07\x37"
-     "\x00\x80"
-     "\x9C\xED\x0D\x00\x0B\x37\x00\x80\x9A\xCD\x0D\x00\x17\x37\x00\x80\x98\xC9"
-     "\x0D\x00"
-     "\x1F\x37\x00\x80\x96\xE1\x0D\x00\x23\x37\x00\x80\x94\x11\x0E\x00\x27\x37"
-     "\x00\x80"
-     "\x92\x0D\x0E\x00\x2F\x37\x00\x80\x90\x09\x0E\x00\x80\xA9\x07\x00\x81\xB5"
-     "\x07\x00"
-     "\x82\xB1\x07\x00\x83\xCD\x07\x00\x84\xD9\x07\x00\x85\xC5\x07\x00\x86\xC1"
-     "\x07\x00"
-     "\x87\xFD\x07\x00\x88\xC9\x07\x00\x89\xD5\x07\x00\x8A\xD1\x07\x00\x8B\x2D"
-     "\x01\x00"
-     "\x8C\x35\x01\x00\x8D\x3D\x01\x00\x8E\x35\x01\x00\x8F\x2D\x01\x00\x90\x55"
-     "\x01\x00"
-     "\x91\x5D\x01\x00\x92\x55\x01\x00\x93\x6D\x01\x00\x94\x75\x01\x00\x95\x7D"
-     "\x01\x00"
-     "\x96\x75\x01\x00\x97\x6D\x01\x00\x98\x55\x01\x00\x99\x5D\x01\x00\x9A\x55"
-     "\x01\x00"
-     "\x9B\x2D\x01\x00\x9C\x35\x01\x00\x9D\x3D\x01\x00\x9E\x35\x01\x00\x9F\x2D"
-     "\x01\x00"
-     "\xA0\xD5\x01\x00\xA1\xDD\x01\x00\xA2\xD5\x01\x00\xA3\xED\x01\x00\xA4\xF5"
-     "\x01\x00"
-     "\xA5\xFD\x01\x00\xA6\xF5\x01\x00\xA7\xED\x01\x00\xA8\xD5\x01\x00\xA9\xDD"
-     "\x01\x00"
-     "\xAA\xD5\x01\x00\xAB\x29\x01\x00\xAC\x39\x01\x00\xAD\x39\x01\x00\xAE\x29"
-     "\x01\x00"
-     "\xAF\x29\x01\x00\xB0\x59\x01\x00\xB1\x59\x01\x00\xB2\x69\x01\x00\xB3\x69"
-     "\x01\x00"
-     "\xB4\x79\x01\x00\xB5\x79\x01\x00\xB6\x69\x01\x00\xB7\x69\x01\x00\xB8\x59"
-     "\x01\x00"
-     "\xB9\x59\x01\x00\xBA\x29\x01\x00\xBB\x29\x01\x00\xBC\x39\x01\x00\xBD\x39"
-     "\x01\x00"
-     "\xBE\x29\x01\x00\xBF\x29\x01\x00\x33\x37\x00\x80\xD3\x37\x00\x80\xE7\x37"
-     "\x00\x80"
-     "\x1B\x38\x00\x80\x1F\x38\x00\x80\x27\x38\x00\x80\x53\x38\x00\x80\x5B\x38"
-     "\x00\x80"
-     "\x63\x38\x00\x80\x6B\x38\x00\x80\xA7\x38\x00\x80\xAB\x38\x00\x80\xB3\x38"
-     "\x00\x80"
-     "\xB7\x38\x00\x80\xBB\x38\x00\x80\xE7\x38\x00\x80\xEB\x38\x00\x80\xF3\x38"
-     "\x00\x80"
-     "\x07\x39\x00\x80\x0B\x39\x00\x80\x27\x39\x00\x80\x2F\x39\x00\x80\x33\x39"
-     "\x00\x80"
-     "\x57\x39\x00\x80\x77\x39\x00\x80\x8B\x39\x00\x80\x8F\x39\x00\x80\x97\x39"
-     "\x00\x80"
-     "\xC3\x39\x00\x80\xC7\x39\x00\x80\xD7\x39\x00\x80\xFB\x39\x00\x80\x99\x25"
-     "\x03\x00"
-     "\x07\x3A\x00\x80\x17\x3A\x00\x80\x2F\x3A\x00\x80\x37\x3A\x00\x80\x3B\x3A"
-     "\x00\x80"
-     "\x3F\x3A\x00\x80\x43\x3A\x00\x80\x4F\x3A\x00\x80\x53\x3A\x00\x80\x5B\x3A"
-     "\x00\x80"
-     "\x77\x3A\x00\x80\x7B\x3A\x00\x80\x7F\x3A\x00\x80\x97\x3A\x00\x80\x9B\x3A"
-     "\x00\x80"
-     "\xA3\x3A\x00\x80\xB7\x3A\x00\x80\xBB\x3A\x00\x80\xBF\x3A\x00\x80\xC7\x3A"
-     "\x00\x80"
-     "\xCB\x3A\x00\x80\xCF\x3A\x00\x80\xD3\x3A\x00\x80\xD7\x3A\x00\x80\xE7\x3A"
-     "\x00\x80"
-     "\xFF\x3A\x00\x80\x03\x3B\x00\x80\x0F\x3B\x00\x80\x13\x3B\x00\x80\x17\x3B"
-     "\x00\x80"
-     "\x1B\x3B\x00\x80\x80\xD9\x03\x00\x81\xD9\x03\x00\x82\xE9\x03\x00\x83\xE9"
-     "\x03\x00"
-     "\x84\xF9\x03\x00\x85\xF9\x03\x00\x86\xE9\x03\x00\x87\xE9\x03\x00\x88\xD9"
-     "\x03\x00"
-     "\x89\xD9\x03\x00\x8A\x71\x00\x00\x8B\x71\x00\x00\x8C\x51\x00\x00\x8D\x51"
-     "\x00\x00"
-     "\x8E\x51\x00\x00\x8F\x51\x00\x00\x90\x31\x00\x00\x91\x31\x00\x00\x92\x31"
-     "\x00\x00"
-     "\x93\x31\x00\x00\x94\xD1\x00\x00\x95\xD1\x00\x00\x23\x3B\x00\x80\x27\x3B"
-     "\x00\x80"
-     "\x2B\x3B\x00\x80\x37\x3B\x00\x80\x3F\x3B\x00\x80\x47\x3B\x00\x80\x4B\x3B"
-     "\x00\x80"
-     "\x53\x3B\x00\x80\x57\x3B\x00\x80\x5B\x3B\x00\x80\x63\x3B\x00\x80\x6B\x3B"
-     "\x00\x80"
-     "\x6B\x24\x00\x80\x9A\x2D\x00\x80\x9E\x2D\x00\x80\x53\x25\x00\x80\xE1\x24"
-     "\x00\x80"
-     "\xE5\x24\x00\x80\xE9\x24\x00\x80\xED\x24\x00\x80\xBE\x7C\x02\x00\x9E\x09"
-     "\x00\x00"
-     "\x82\x60\x02\x00\x99\x0D\x00\x00\xF1\x24\x00\x80\xBE\x18\x02\x00\x80\x00"
-     "\x1C\x00"
-     "\x81\xEC\x1C\x00\x82\xA0\x1F\x00\x83\xBC\x18\x00\x84\x54\x1A\x00\x85\xD4"
-     "\x1B\x00"
-     "\x86\x98\x15\x00\x9E\x65\x00\x00\x88\x00\x10\x00\x89\xE4\x11\x00\x8A\xE8"
-     "\x13\x00"
-     "\x8B\x00\x2C\x00\x8C\x14\x2F\x00\x8D\xF4\x29\x00\x8E\x00\x24\x00\x8F\xFC"
-     "\x25\x00"
-     "\xA0\xD1\x05\x00\xA1\xD1\x05\x00\xA2\xD1\x05\x00\xA3\xD1\x05\x00\xA4\xF1"
-     "\x05\x00"
-     "\xA5\xF1\x05\x00\xA6\xF1\x05\x00\xA7\xF1\x05\x00\xA8\xD1\x05\x00\xA9\xD1"
-     "\x05\x00"
-     "\xAA\xD1\x05\x00\xAB\xD1\x05\x00\xAC\x39\x04\x00\xAD\x25\x04\x00\xAE\x21"
-     "\x04\x00"
-     "\xAF\x1D\x04\x00\xB0\x69\x04\x00\xB1\x75\x04\x00\xB2\x71\x04\x00\xB3\x0D"
-     "\x04\x00"
-     "\xB4\x19\x04\x00\xB5\x05\x04\x00\xB6\x01\x04\x00\xB7\x3D\x04\x00\xB8\x09"
-     "\x04\x00"
-     "\xB9\x15\x04\x00\xBA\x11\x04\x00\xBB\xED\x04\x00\xBC\xF9\x04\x00\xBD\xE5"
-     "\x04\x00"
-     "\xBE\xE1\x04\x00\xBF\xDD\x04\x00\x80\xD1\x05\x00\x81\xD1\x05\x00\x82\xD1"
-     "\x05\x00"
-     "\x83\xD1\x05\x00\x84\xF1\x05\x00\x85\xF1\x05\x00\x86\xF1\x05\x00\x87\xF1"
-     "\x05\x00"
-     "\x88\xD1\x05\x00\x89\xD1\x05\x00\x8A\xD1\x05\x00\x8B\xD1\x05\x00\x8C\x31"
-     "\x05\x00"
-     "\x8D\x31\x05\x00\x8E\x31\x05\x00\x8F\x31\x05\x00\x90\x51\x05\x00\x91\x51"
-     "\x05\x00"
-     "\x92\x51\x05\x00\x93\x51\x05\x00\x94\x71\x05\x00\x95\x71\x05\x00\x96\x71"
-     "\x05\x00"
-     "\x97\x71\x05\x00\x98\x51\x05\x00\x99\x51\x05\x00\x9A\x51\x05\x00\x9B\x51"
-     "\x05\x00"
-     "\x9C\x31\x05\x00\x9D\x31\x05\x00\x9E\x31\x05\x00\x9F\x31\x05\x00\x82\x74"
-     "\x02\x00"
-     "\x8B\x10\x6C\x00\xE3\xE8\x6D\x00\xE3\x04\x18\x00\x99\x3D\x00\x00\x8F\x90"
-     "\x03\x00"
-     "\xB6\x91\x1E\x00\x8D\x50\x6C\x00\xB8\xBD\x18\x00\xB9\xA9\x1E\x00\xBA\xB9"
-     "\x1E\x00"
-     "\xF5\x24\x00\x80\x86\x34\x6C\x00\xBE\x14\x02\x00\xEF\xC0\x19\x00\xEF\x6C"
-     "\x1C\x00"
-     "\xE3\xF8\x1C\x00\x9B\x1C\x03\x00\xE3\x40\x02\x00\x99\x58\x03\x00\xE3\x30"
-     "\x02\x00"
-     "\x9F\xD0\x02\x00\x9E\x75\x00\x00\x9D\xD0\x02\x00\xEF\x08\x02\x00\x93\xA8"
-     "\x03\x00"
-     "\x82\x8C\x02\x00\x91\xEC\x03\x00\xEF\x34\x1C\x00\x97\xA8\x03\x00\xEF\xB4"
-     "\x02\x00"
-     "\x95\xB8\x03\x00\x99\x05\x00\x00\xF9\x24\x00\x80\xA8\x8C\x02\x00\x9E\x09"
-     "\x00\x00"
-     "\xBE\x64\x02\x00\xAF\x8C\x02\x00\xE3\x0C\x02\x00\x82\x58\x02\x00\x99\x05"
-     "\x00\x00"
-     "\xFD\x24\x00\x80\xEF\xE4\x02\x00\xA1\x64\x02\x00\xA6\x64\x02\x00\xBE\x00"
-     "\x02\x00"
-     "\xA4\x64\x02\x00\x01\x25\x00\x80\x80\x35\x5A\x02\xBB\x0C\x01\x00\xB8\x40"
-     "\x01\x00"
-     "\x9E\x69\x00\x00\x82\x6C\x02\x00\xE3\x0C\x02\x00\x99\x65\x00\x00\xBE\x4C"
-     "\x02\x00"
-     "\xB2\xA4\x02\x00\xEF\xC0\x02\x00\x9E\x55\x00\x00\x99\x19\x00\x00\x82\x78"
-     "\x02\x00"
-     "\x05\x25\x00\x80\xBE\x84\x03\x00\xB5\xA0\x02\x00\xE3\xF8\x03\x00\x9E\xF1"
-     "\x01\x00"
-     "\xE3\x0C\x02\x00\x82\x78\x02\x00\x99\x05\x00\x00\x09\x25\x00\x80\xBE\x7C"
-     "\x02\x00"
-     "\x9E\x09\x00\x00\x82\x60\x02\x00\x99\x0D\x00\x00\x0D\x25\x00\x80\xBE\x60"
-     "\x02\x00"
-     "\xEF\x30\x03\x00\x9E\x1D\x00\x00\xEF\xD8\x02\x00\x82\x18\x02\x00\x99\x09"
-     "\x00\x00"
-     "\xE3\x6C\x02\x00\x11\x25\x00\x80\xBE\x60\x02\x00\xE3\x64\x02\x00\x9E\x1D"
-     "\x00\x00"
-     "\xE3\x0C\x02\x00\x82\x5C\x02\x00\xEF\xD4\x02\x00\x99\x09\x00\x00\xEF\xE0"
-     "\x02\x00"
-     "\x15\x25\x00\x80\xBE\x88\x02\x00\xEF\xA0\x02\x00\x82\x60\x02\x00\x99\xF9"
-     "\x00\x00"
-     "\x9E\xED\x00\x00\x19\x25\x00\x80\xE3\xF8\x02\x00\x9E\x09\x00\x00\xBE\x64"
-     "\x02\x00"
-     "\xE3\x04\x02\x00\x82\x50\x02\x00\x1D\x25\x00\x80\x9A\x09\x00\x00\xEF\xEC"
-     "\x02\x00"
-     "\x21\x25\x00\x80\x99\x31\x00\x00\xBE\x08\x02\x00\x82\x04\x02\x00\xEF\xD0"
-     "\x02\x00"
-     "\x25\x25\x00\x80\x9E\x5D\x00\x00\x9F\x6D\x00\x00\xE3\x74\x02\x00\x29\x25"
-     "\x00\x80"
-     "\xE3\x2C\x02\x00\x99\x69\x00\x00\x9A\x15\x00\x00\x2D\x25\x00\x80\xEF\x18"
-     "\x01\x00"
-     "\xBE\x70\x02\x00\x9F\x25\x00\x00\x9E\x51\x00\x00\x9A\xFD\x03\x00\x82\x6C"
-     "\x02\x00"
-     "\xEF\xE4\x02\x00\x99\x21\x00\x00\xBE\x68\x02\x00\xE3\x04\x02\x00\x82\x4C"
-     "\x02\x00"
-     "\x31\x25\x00\x80\x9E\xCD\x03\x00\x9F\x19\x00\x00\xBE\x58\x02\x00\x35\x25"
-     "\x00\x80"
-     "\x99\x3D\x00\x00\xBE\x3C\x02\x00\x82\x0C\x02\x00\x9A\x1D\x00\x00\x39\x25"
-     "\x00\x80"
-     "\xEF\xD4\x02\x00\x9E\x25\x00\x00\x9F\x0D\x00\x00\x99\x09\x00\x00\xE3\x64"
-     "\x02\x00"
-     "\x44\x25\x00\x80\x9A\x1D\x00\x00\x48\x25\x00\x80\xBC\x00\x00\x80\xC1\x00"
-     "\x00\x80"
-     "\xE3\x3C\x02\x00\x9E\x29\x00\x00\x9F\x35\x00\x00\x82\x7C\x02\x00\xEF\x14"
-     "\x02\x00"
-     "\x99\x09\x00\x00\xEF\xA8\x02\x00\x5F\x25\x00\x80\x9E\x05\x00\x00\xBE\x84"
-     "\x02\x00"
-     "\xE3\x04\x02\x00\x82\x60\x02\x00\x99\x11\x00\x00\x82\x50\x02\x00\xBE\x4C"
-     "\x02\x00"
-     "\x4C\x25\x00\x80\x9B\x2A\x00\x80\x9E\x3D\x00\x00\x99\x0D\x00\x00\x36\x26"
-     "\x00\x80"
-     "\xBE\x64\x02\x00\x9E\x19\x00\x00\xEF\xD8\x02\x00\x82\x18\x02\x00\x9E\x6D"
-     "\x00\x00"
-     "\xE3\x58\x02\x00\x99\x15\x00\x00\xE3\x78\x02\x00\xBE\x08\x02\x00\x99\x26"
-     "\x00\x80"
-     "\xE3\x04\x02\x00\x82\x60\x02\x00\x99\x3D\x00\x00\xDC\x26\x00\x80\xEF\xF4"
-     "\x02\x00"
-     "\xBE\x7C\x02\x00\x9E\x0D\x00\x00\xEF\xA4\x02\x00\x82\x80\x03\x00\xEF\xCC"
-     "\x02\x00"
-     "\x08\x27\x00\x80\xE3\xFC\x03\x00\x9B\x51\x01\x00\xBE\x54\x02\x00\x82\x5C"
-     "\x02\x00"
-     "\x99\xED\x01\x00\x9F\x8D\x01\x00\x9C\x41\x01\x00\x9D\xCC\x02\x00\x32\x27"
-     "\x00\x80"
-     "\x9E\xD9\x01\x00\x82\x04\x02\x00\xBE\x6C\x02\x00\xEF\xC0\x02\x00\x99\x15"
-     "\x00\x00"
-     "\x9E\x19\x00\x00\x57\x27\x00\x80\xE3\x44\x02\x00\x9E\x79\x00\x00\x99\x75"
-     "\x00\x00"
-     "\xBE\x70\x02\x00\xEF\xDC\x02\x00\x99\x39\x00\x00\xE3\x78\x02\x00\x82\x6C"
-     "\x02\x00"
-     "\xE3\x34\x02\x00\x9E\x09\x00\x00\xEF\xDC\x02\x00\x6C\x27\x00\x80\xEF\xA4"
-     "\x02\x00"
-     "\xBE\x68\x02\x00\x82\x60\x02\x00\x99\xDD\x00\x00\xE3\xF0\x02\x00\xBB\xD8"
-     "\x03\x00"
-     "\x82\x2C\x02\x00\xB9\x80\x03\x00\x9E\x31\x00\x00\xBF\x18\x04\x00\xBE\x70"
-     "\x02\x00"
-     "\xBD\xF8\x03\x00\x74\x27\x00\x80\xB3\x68\x03\x00\x90\x27\x00\x80\xB1\x20"
-     "\x03\x00"
-     "\xEF\xD0\x02\x00\xB7\x58\x03\x00\x9E\x11\x00\x00\xB5\x68\x03\x00\x99\x69"
-     "\x00\x00"
-     "\xAB\x88\x02\x00\x99\x25\x00\x00\xBE\x0C\x02\x00\xE3\x50\x02\x00\xAF\x8C"
-     "\x02\x00"
-     "\xEF\xD4\x02\x00\xAD\x98\x02\x00\xEF\xD4\x02\x00\x82\x48\x02\x00\xE3\x04"
-     "\x02\x00"
-     "\x98\x27\x00\x80\xA6\xF8\x02\x00\xBE\x00\x04\x00\xBE\x8C\x05\x00\x82\x9C"
-     "\x05\x00"
-     "\xE3\xF4\x05\x00\x99\x0D\x00\x00\xA0\x27\x00\x80\x9E\x05\x00\x00\x98\xD8"
-     "\x02\x00"
-     "\x82\x50\x02\x00\xE3\x04\x02\x00\x9B\xC4\x02\x00\x99\x0D\x00\x00\x95\xAC"
-     "\x02\x00"
-     "\xEF\xE4\x02\x00\xB2\x27\x00\x80\xEF\x30\x05\x00\xBE\x0C\x02\x00\x92\xB4"
-     "\x02\x00"
-     "\x82\x1C\x02\x00\x9E\x6D\x00\x00\x99\x11\x00\x00\xE3\x7C\x02\x00\x8F\x00"
-     "\x03\x00"
-     "\x88\x48\x03\x00\xBE\x27\x00\x80\xBE\x7C\x02\x00\x9E\x09\x00\x00\x84\xA4"
-     "\x03\x00"
-     "\x82\x6C\x02\x00\x86\x74\x03\x00\x99\x1D\x00\x00\xC6\x27\x00\x80\x81\xE0"
-     "\x03\x00"
-     "\xEF\xD8\x02\x00\xBE\x84\x02\x00\x9E\xF1\x00\x00\xBD\x08\x01\x00\x82\x74"
-     "\x02\x00"
-     "\xBF\xA5\x02\x00\xE3\xFC\x02\x00\xE3\x14\x02\x00\x99\x05\x00\x00\xDC\x27"
-     "\x00\x80"
-     "\xEF\x24\x02\x00\xEF\xF4\x02\x00\xBE\x7C\x02\x00\x9E\x31\x00\x00\xB0\xFC"
-     "\x02\x00"
-     "\xB1\xFC\x02\x00\xB2\xE4\x02\x00\x82\x10\x02\x00\xE3\x74\x02\x00\x9A\x11"
-     "\x00\x00"
-     "\x99\x19\x00\x00\xAF\xD8\x02\x00\x1A\x28\x00\x80\x2B\x28\x00\x80\x9F\x0D"
-     "\x00\x00"
-     "\x9E\x0D\x00\x00\xBE\x44\x02\x00\x82\x74\x02\x00\x4D\x28\x00\x80\x58\x28"
-     "\x00\x80"
-     "\xEF\xE8\x02\x00\x99\x1D\x00\x00\x9A\x15\x00\x00\xBE\x8C\x03\x00\xE3\xEC"
-     "\x03\x00"
-     "\xE3\x14\x02\x00\x9E\xE1\x01\x00\x9F\xE1\x01\x00\x99\x05\x00\x00\xA8\x28"
-     "\x00\x80"
-     "\x82\x6C\x02\x00\x9A\x3D\x00\x00\xBC\x28\x00\x80\xBE\x74\x02\x00\x9E\x3D"
-     "\x00\x00"
-     "\x9F\x0D\x00\x00\xEF\x34\x03\x00\xEF\xE8\x02\x00\x82\x1C\x02\x00\xDE\x28"
-     "\x00\x80"
-     "\x99\x7D\x00\x00\xE3\x74\x02\x00\xEC\x28\x00\x80\x9A\x05\x00\x00\xBE\x70"
-     "\x02\x00"
-     "\xE3\x3C\x02\x00\x9E\x65\x00\x00\x9F\x15\x00\x00\x82\x74\x02\x00\xEF\xE4"
-     "\x02\x00"
-     "\x34\x29\x00\x80\x42\x29\x00\x80\x99\x19\x00\x00\xEF\xDC\x02\x00\xBE\x80"
-     "\x02\x00"
-     "\x9A\x11\x00\x00\x9E\xE9\x00\x00\x9F\xE9\x00\x00\x99\x1D\x00\x00\x9E\x61"
-     "\x06\x00"
-     "\x82\x7C\x02\x00\x9F\x2A\x00\x80\xE3\x00\x02\x00\xBE\x44\x02\x00\x9E\x35"
-     "\x00\x00"
-     "\x82\x6C\x02\x00\xEF\xCC\x02\x00\x99\x1D\x00\x00\xA3\x2A\x00\x80\x9E\x05"
-     "\x00\x00"
-     "\xBE\x74\x02\x00\x82\x0C\x02\x00\xE3\x6C\x02\x00\xE3\x70\x02\x00\xA7\x2A"
-     "\x00\x80"
-     "\xBE\x6C\x02\x00\xE3\x04\x02\x00\x99\x1D\x00\x00\x9E\x11\x00\x00\x82\x54"
-     "\x02\x00"
-     "\xEF\xE0\x02\x00\xAB\x2A\x00\x80\x9E\x0D\x00\x00\x99\x09\x00\x00\xEF\xB0"
-     "\x02\x00"
-     "\xEF\xC0\x02\x00\xBE\x68\x02\x00\xB7\x2A\x00\x80\xE3\x0C\x02\x00\x82\x68"
-     "\x02\x00"
-     "\x99\xF5\x03\x00\xBE\x64\x02\x00\x9E\xED\x03\x00\x85\x00\x00\x00\x86\x00"
-     "\x00\x00"
-     "\x87\x00\x00\x00\x88\x00\x00\x00\x89\x00\x00\x00\x17\x27\x00\x80\xBF\x1A"
-     "\x00\x80"
-     "\xEF\xC4\x02\x00\x8D\x00\x00\x00\x8E\x00\x00\x00\xC3\x1A\x00\x80\xB1\x01"
-     "\x0C\x00"
-     "\x91\x00\x00\x00\xB3\x65\x00\x00\xB2\x4D\x0E\x00\xB5\x6D\x00\x00\xB4\x51"
-     "\x0E\x00"
-     "\xB7\x49\x0E\x00\xB6\xC9\x00\x00\xB9\x21\x0E\x00\xB8\x71\x0E\x00\xBB\xF5"
-     "\x00\x00"
-     "\xBA\xF5\x00\x00\xBD\xC1\x00\x00\xBC\xDD\x00\x00\xBF\xC5\x00\x00\xBE\xC5"
-     "\x00\x00"
-     "\xA0\x00\x00\x00\xA1\x00\x00\x00\xA2\x00\x00\x00\xA3\x00\x00\x00\xA4\x00"
-     "\x00\x00"
-     "\xC7\x1A\x00\x80\xCB\x1A\x00\x80\xCF\x1A\x00\x80\xA8\x00\x00\x00\xA9\x00"
-     "\x00\x00"
-     "\xD3\x1A\x00\x80\xAB\x00\x00\x00\xD7\x1A\x00\x80\xAD\x00\x00\x00\xDB\x1A"
-     "\x00\x80"
-     "\xDF\x1A\x00\x80\xEB\x19\x00\x80\xFF\x19\x00\x80\xB2\x00\x00\x00\xB3\x00"
-     "\x00\x00"
-     "\x4B\x1A\x00\x80\xB5\x00\x00\x00\xB6\x00\x00\x00\x4F\x1A\x00\x80\x53\x1A"
-     "\x00\x80"
-     "\x57\x1A\x00\x80\x5B\x1A\x00\x80\xBB\x00\x00\x00\x5F\x1A\x00\x80\x63\x1A"
-     "\x00\x80"
-     "\xBE\x00\x00\x00\xBF\x00\x00\x00\x80\xC1\x01\x00\x81\x41\x0F\x00\x82\x59"
-     "\x0F\x00"
-     "\x83\x59\x0F\x00\x84\xD5\x01\x00\x85\x95\x0F\x00\x86\x9D\x0F\x00\x87\x8D"
-     "\x0F\x00"
-     "\x88\xB5\x0F\x00\x89\xE5\x0F\x00\x8A\xED\x0F\x00\x8B\xFD\x0F\x00\x8C\xE5"
-     "\x0F\x00"
-     "\x8D\x35\x0E\x00\x8E\x3D\x0E\x00\x8F\xED\x01\x00\x90\x9D\x01\x00\x91\xA5"
-     "\x01\x00"
-     "\x92\xAD\x01\x00\x93\xA5\x01\x00\x94\xA1\x01\x00\x95\xA1\x01\x00\x96\x51"
-     "\x0F\x00"
-     "\x97\x51\x0F\x00\x98\x79\x0F\x00\x99\x79\x0F\x00\x9A\x49\x0F\x00\x9B\x49"
-     "\x0F\x00"
-     "\x9C\x59\x0F\x00\x9D\x59\x0F\x00\x9E\x49\x0F\x00\x9F\x49\x0F\x00\xA0\xB9"
-     "\x0F\x00"
-     "\xA1\xB9\x0F\x00\xA2\xC1\x0F\x00\xA3\xC1\x0F\x00\xA4\xC9\x0F\x00\xA5\xC9"
-     "\x0F\x00"
-     "\xA6\xF9\x0F\x00\xA7\xF9\x0F\x00\xA8\xC9\x0F\x00\xA9\xC9\x0F\x00\xAA\xD9"
-     "\x0F\x00"
-     "\xAB\xD9\x0F\x00\xAC\xC9\x0F\x00\xAD\xC9\x0F\x00\xAE\x31\x0F\x00\xAF\x31"
-     "\x0F\x00"
-     "\xB0\x59\x0F\x00\xB1\x59\x0F\x00\xB2\x69\x0F\x00\xB3\x69\x0F\x00\xB4\x79"
-     "\x0F\x00"
-     "\xB5\x79\x0F\x00\xB6\x69\x0F\x00\xB7\x69\x0F\x00\xB8\x59\x0F\x00\xB9\x59"
-     "\x0F\x00"
-     "\xBA\x21\x0F\x00\xBB\x21\x0F\x00\xBC\x29\x0F\x00\xBD\x29\x0F\x00\xBE\x19"
-     "\x0F\x00"
-     "\xBF\x19\x0F\x00\x9B\x19\x00\x80\x9F\x19\x00\x80\x67\x1A\x00\x80\x6B\x1A"
-     "\x00\x80"
-     "\xA3\x19\x00\x80\xA7\x19\x00\x80\x6F\x1A\x00\x80\x73\x1A\x00\x80\x77\x1A"
-     "\x00\x80"
-     "\x7B\x1A\x00\x80\x7F\x1A\x00\x80\x83\x1A\x00\x80\x87\x1A\x00\x80\x8B\x1A"
-     "\x00\x80"
-     "\x8F\x1A\x00\x80\x93\x1A\x00\x80\xAB\x19\x00\x80\xAF\x19\x00\x80\x97\x1A"
-     "\x00\x80"
-     "\x9B\x1A\x00\x80\xB3\x19\x00\x80\xB7\x19\x00\x80\x47\x1A\x00\x80\xE3\x19"
-     "\x00\x80"
-     "\xE7\x19\x00\x80\xE3\x1A\x00\x80\xE7\x1A\x00\x80\xEB\x1A\x00\x80\xEF\x1A"
-     "\x00\x80"
-     "\xF3\x1A\x00\x80\xF7\x1A\x00\x80\xFB\x1A\x00\x80\xBB\x19\x00\x80\xBF\x19"
-     "\x00\x80"
-     "\xEF\x19\x00\x80\xFF\x1A\x00\x80\xC3\x19\x00\x80\xC7\x19\x00\x80\x03\x1B"
-     "\x00\x80"
-     "\xF3\x19\x00\x80\xF7\x19\x00\x80\xFB\x19\x00\x80\x03\x1A\x00\x80\x07\x1A"
-     "\x00\x80"
-     "\x0B\x1A\x00\x80\x0F\x1A\x00\x80\x13\x1A\x00\x80\x17\x1A\x00\x80\xCB\x19"
-     "\x00\x80"
-     "\xCF\x19\x00\x80\x1B\x1A\x00\x80\x1F\x1A\x00\x80\xD3\x19\x00\x80\xD7\x19"
-     "\x00\x80"
-     "\x23\x1A\x00\x80\x27\x1A\x00\x80\x2B\x1A\x00\x80\x2F\x1A\x00\x80\x33\x1A"
-     "\x00\x80"
-     "\x37\x1A\x00\x80\x07\x1B\x00\x80\x0B\x1B\x00\x80\x3B\x1A\x00\x80\x3F\x1A"
-     "\x00\x80"
-     "\xDB\x19\x00\x80\xDF\x19\x00\x80\x92\xF1\x1F\x00\x93\x29\x11\x00\x94\xB1"
-     "\x00\x00"
-     "\x95\xB1\x00\x00\x96\xB1\x00\x00\x97\xB1\x00\x00\x98\x91\x00\x00\x99\x91"
-     "\x00\x00"
-     "\x9A\x3D\x11\x00\x9B\x95\x00\x00\x9C\x8D\x00\x00\x9D\xF5\x00\x00\x9E\xF1"
-     "\x00\x00"
-     "\x9F\x09\x11\x00\x80\xA9\x01\x00\x81\xA9\x01\x00\x82\xB9\x01\x00\x83\xB9"
-     "\x01\x00"
-     "\x84\xA9\x01\x00\x85\xA9\x01\x00\x86\x25\x00\x00\x87\xDD\x00\x00\x88\xE5"
-     "\x00\x00"
-     "\x89\xED\x00\x00\x8A\xE5\x00\x00\x8B\xFD\x00\x00\x8C\xE5\x00\x00\x8D\xED"
-     "\x00\x00"
-     "\x8E\xE5\x00\x00\x43\x1A\x00\x80\x9F\x1A\x00\x80\xA3\x1A\x00\x80\xA7\x1A"
-     "\x00\x80"
-     "\xAB\x1A\x00\x80\xAF\x1A\x00\x80\xB3\x1A\x00\x80\xB7\x1A\x00\x80\xBB\x1A"
-     "\x00\x80"
-     "\xD3\x2B\x00\x80\x8A\x2E\x00\x80\xD7\x2B\x00\x80\xEF\x2B\x00\x80\xDB\x2B"
-     "\x00\x80"
-     "\xCB\x34\x00\x80\xE3\x2B\x00\x80\xCB\x2B\x00\x80\x15\x2F\x00\x80\xE1\xAC"
-     "\x7D\x02"
-     "\x9E\x2E\x00\x80\xE3\x0C\x7D\x02\x9F\x01\x00\x80\xA5\x01\x00\x80\xAB\x01"
-     "\x00\x80"
-     "\xB1\x01\x00\x80\xB7\x01\x00\x80\xBD\x01\x00\x80\xC3\x01\x00\x80\xC9\x01"
-     "\x00\x80"
-     "\xCF\x01\x00\x80\xD5\x01\x00\x80\xDB\x01\x00\x80\xEF\xF0\x7D\x02\xE1\x01"
-     "\x00\x80"
-     "\xE7\x01\x00\x80\xED\x01\x00\x80\xCB\x02\x00\x80\xD1\x02\x00\x80\xD7\x02"
-     "\x00\x80"
-     "\xDD\x02\x00\x80\xE3\x02\x00\x80\xE9\x02\x00\x80\xEF\x02\x00\x80\xF5\x02"
-     "\x00\x80"
-     "\x0D\x03\x00\x80\x19\x03\x00\x80\x1F\x03\x00\x80\x25\x03\x00\x80\x2B\x03"
-     "\x00\x80"
-     "\x97\x19\x00\x80\x31\x03\x00\x80\x13\x03\x00\x80\xFB\x02\x00\x80\x04\x03"
-     "\x00\x80"
-     "\xF3\x01\x00\x80\x0B\x02\x00\x80\xFF\x01\x00\x80\x53\x02\x00\x80\x11\x02"
-     "\x00\x80"
-     "\x2F\x02\x00\x80\xF9\x01\x00\x80\x29\x02\x00\x80\x05\x02\x00\x80\x3B\x02"
-     "\x00\x80"
-     "\x6B\x02\x00\x80\xE1\x44\x4B\x00\xE1\x1C\xAF\x03\xE3\xB4\x77\x00\xE3\x14"
-     "\x4B\x00"
-     "\x89\x02\x00\x80\x83\x02\x00\x80\x77\x02\x00\x80\xC5\x02\x00\x80\x59\x02"
-     "\x00\x80"
-     "\x65\x02\x00\x80\x7D\x02\x00\x80\x71\x02\x00\x80\x9B\x02\x00\x80\x47\x02"
-     "\x00\x80"
-     "\xEF\x68\x77\x00\xEF\xE4\x4B\x00\x8F\x02\x00\x80\xB9\x02\x00\x80\xA1\x02"
-     "\x00\x80"
-     "\x35\x02\x00\x80\x17\x02\x00\x80\x4D\x02\x00\x80\x5F\x02\x00\x80\x95\x02"
-     "\x00\x80"
-     "\x1D\x02\x00\x80\xBF\x02\x00\x80\x41\x02\x00\x80\xA7\x02\x00\x80\x23\x02"
-     "\x00\x80"
-     "\xAD\x02\x00\x80\xB3\x02\x00\x80\x3E\x2E\x00\x80\xE1\x50\x6A\x03\xE1\xD8"
-     "\x44\x03"
-     "\xE3\x5C\x1F\x03\xE3\xC8\x76\x03\xE1\x34\x0F\x03\xE1\xD0\x0B\x03\xE3\xC4"
-     "\x2E\x03"
-     "\xE3\xC4\x0A\x03\x20\x30\x00\x80\xFF\x35\x00\x80\xEF\xF8\x3F\x03\xEF\xE8"
-     "\x02\x03"
-     "\x19\x08\x00\x80\x9D\x04\x00\x80\xEF\xE0\x0D\x03\xEF\x68\x73\x03\xE1\x84"
-     "\xC6\x03"
-     "\xE1\x98\xDF\x03\xE3\x1C\xEB\x03\xE3\x98\xC3\x03\xE1\x54\x85\x03\xE1\x9C"
-     "\x98\x03"
-     "\xE3\xEC\xBE\x03\xE3\x80\x9E\x03\xAC\x04\x00\x80\xBB\x04\x00\x80\xEF\x3C"
-     "\xB6\x03"
-     "\xEF\x74\x82\x03\xCA\x04\x00\x80\xD9\x04\x00\x80\xEF\xA4\xF7\x03\xEF\x28"
-     "\xD6\x03"
-     "\xE1\x4C\x17\x02\xE1\x30\x67\x02\xE3\x4C\x0A\x02\xE3\x38\x66\x02\xE1\xCC"
-     "\x3F\x02"
-     "\xE1\x48\x03\x02\xE3\x44\xDB\x02\xE3\x38\x3A\x02\xE2\x04\x00\x80\xEB\x04"
-     "\x00\x80"
-     "\xEF\x6C\xFA\x02\xEF\xA4\x2E\x02\xF4\x04\x00\x80\xFD\x04\x00\x80\xEF\x78"
-     "\x1F\x02"
-     "\xEF\x6C\x7E\x02\xE1\x04\xE7\x02\xE1\x5C\xD7\x02\xE3\x1C\x9A\x02\xE3\x78"
-     "\xF2\x02"
-     "\xE1\x28\x58\x03\xE1\x80\x9F\x02\xE3\x0C\x44\x03\xE3\x68\xAE\x02\x29\x05"
-     "\x00\x80"
-     "\x32\x05\x00\x80\xEF\x38\x43\x03\xEF\xF8\x58\x03\x3B\x05\x00\x80\x3E\x05"
-     "\x00\x80"
-     "\xEF\x24\xEE\x02\xEF\x2C\xF7\x02\xB8\xCD\x06\x00\xB9\xD5\x06\x00\xBA\xDD"
-     "\x06\x00"
-     "\xBB\xD5\x06\x00\xBC\xCD\x06\x00\xBD\xF5\x06\x00\xBE\xFD\x06\x00\xBF\xF5"
-     "\x06\x00"
-     "\xB0\x85\x06\x00\xB1\x8D\x06\x00\xB2\x85\x06\x00\xB3\x9D\x06\x00\xB4\x85"
-     "\x06\x00"
-     "\xB5\xF5\x06\x00\xB6\xFD\x06\x00\xB7\xF5\x06\x00\xA8\x35\x06\x00\xA9\x3D"
-     "\x06\x00"
-     "\xAA\x35\x06\x00\xAB\xDD\x06\x00\xAC\xC5\x06\x00\xAD\xCD\x06\x00\xAE\xC5"
-     "\x06\x00"
-     "\xAF\xFD\x06\x00\xA0\x35\x06\x00\xA1\x3D\x06\x00\xA2\x35\x06\x00\xA3\x0D"
-     "\x06\x00"
-     "\xA4\x15\x06\x00\xA5\x1D\x06\x00\xA6\x15\x06\x00\xA7\x0D\x06\x00\x98\xB5"
-     "\x06\x00"
-     "\x99\xBD\x06\x00\x9A\xB5\x06\x00\x9B\xC9\x06\x00\x9C\xD9\x06\x00\x9D\xD9"
-     "\x06\x00"
-     "\x9E\xC9\x06\x00\x41\x05\x00\x80\x90\xB5\x06\x00\x91\xBD\x06\x00\x92\xB5"
-     "\x06\x00"
-     "\x93\x8D\x06\x00\x94\x95\x06\x00\x95\x9D\x06\x00\x96\x95\x06\x00\x97\x8D"
-     "\x06\x00"
-     "\x88\x31\x05\x00\x89\x31\x05\x00\x8A\x31\x05\x00\x8B\xCD\x06\x00\x8C\xD5"
-     "\x06\x00"
-     "\x8D\xDD\x06\x00\x8E\xD5\x06\x00\x8F\xCD\x06\x00\x80\x31\x05\x00\x81\x31"
-     "\x05\x00"
-     "\x82\x31\x05\x00\x83\x31\x05\x00\x84\x11\x05\x00\x85\x11\x05\x00\x86\x11"
-     "\x05\x00"
-     "\x87\x11\x05\x00\x80\xED\x07\x00\x81\xF5\x07\x00\x82\xFD\x07\x00\x83\xF5"
-     "\x07\x00"
-     "\x84\xED\x07\x00\x85\x35\x06\x00\x86\x41\x17\x00\x87\x39\x06\x00\x44\x05"
-     "\x00\x80"
-     "\x3F\x3D\x00\x80\x9F\x4F\x00\x80\xCF\x58\x00\x80\x2F\x6B\x00\x80\x5F\x74"
-     "\x00\x80"
-     "\x8F\x7D\x00\x80\xEF\x8F\x00\x80\x90\x71\x06\x00\x91\x71\x06\x00\x92\x29"
-     "\x06\x00"
-     "\x93\x29\x06\x00\x94\x21\x06\x00\x95\x21\x06\x00\x96\xF9\x06\x00\x97\xF9"
-     "\x06\x00"
-     "\x98\xD1\x06\x00\x99\xD1\x06\x00\x9A\x89\x06\x00\x9B\x89\x06\x00\x9C\x81"
-     "\x06\x00"
-     "\x9D\x81\x06\x00\x9E\xE5\x01\x00\x9F\x5D\x00\x00\xA0\xC1\x07\x00\xA1\x81"
-     "\x05\x00"
-     "\xA2\x99\x05\x00\xA3\xCD\x05\x00\xA4\xD5\x05\x00\xA5\xC5\x05\x00\xA6\x19"
-     "\x05\x00"
-     "\xA7\x0D\x05\x00\xA8\x35\x05\x00\xA9\x61\x05\x00\xAA\x79\x05\x00\xAB\x79"
-     "\x05\x00"
-     "\xAC\xB1\x04\x00\xAD\xB1\x04\x00\xAE\x9D\x00\x00\xAF\x95\x00\x00\xB0\xED"
-     "\x00\x00"
-     "\xB1\xF5\x00\x00\xB2\xFD\x00\x00\xB3\xF5\x00\x00\xB4\xED\x00\x00\xB5\x29"
-     "\x00\x00"
-     "\xB6\xD9\x03\x00\xB7\xD9\x03\x00\xB8\xE9\x03\x00\xB9\xE9\x03\x00\xBA\xF9"
-     "\x03\x00"
-     "\xBB\xF9\x03\x00\xBC\xE9\x03\x00\xBD\xE9\x03\x00\xBE\xD9\x03\x00\x4F\xA2"
-     "\x00\x80"
-     "\x7F\xAB\x00\x80\xE6\xBD\x00\x80\x1D\xC7\x00\x80\x4D\xD0\x00\x80\x7D\xD9"
-     "\x00\x80"
-     "\xAD\xE2\x00\x80\x2A\xBE\x00\x80\x33\xB1\x00\x80\xFF\xA7\x00\x80\x27\x2C"
-     "\x00\x80"
-     "\xBE\x2C\x00\x80\xCF\x2B\x00\x80\x13\x2C\x00\x80\x76\x32\x00\x80\x36\x32"
-     "\x00\x80"
-     "\x83\x35\x00\x80\x0A\x2E\x00\x80\x3B\x34\x00\x80\xBF\x38\x00\x80\x97\x35"
-     "\x00\x80"
-     "\x52\x2D\x00\x80\xC7\x35\x00\x80\xD3\x34\x00\x80\x47\x39\x00\x80\x8E\x2C"
-     "\x00\x80"
-     "\xBA\x2D\x00\x80\x2F\x33\x00\x80\x5F\x3A\x00\x80\x52\x2C\x00\x80\xE2\x2C"
-     "\x00\x80"
-     "\xD3\x32\x00\x80\xF1\x2F\x00\x80\xFE\x2D\x00\x80\x8E\x2D\x00\x80\x6D\x2F"
-     "\x00\x80"
-     "\x61\x2F\x00\x80\x23\x35\x00\x80\x4E\x2C\x00\x80\xD3\x38\x00\x80\xAA\x2D"
-     "\x00\x80"
-     "\x02\x2F\x00\x80\x47\x05\x00\x80\x4A\x05\x00\x80\x4D\x05\x00\x80\x50\x05"
-     "\x00\x80"
-     "\x7C\x05\x00\x80\x7F\x05\x00\x80\x82\x05\x00\x80\x85\x05\x00\x80\x88\x05"
-     "\x00\x80"
-     "\x8B\x05\x00\x80\x8E\x05\x00\x80\x91\x05\x00\x80\x94\x05\x00\x80\x97\x05"
-     "\x00\x80"
-     "\xB7\x05\x00\x80\xB8\x00\x00\x00\x78\x04\x00\x80\x17\x05\x00\x80\x6A\x05"
-     "\x00\x80"
-     "\xA5\x05\x00\x80\xCB\x05\x00\x80\xE2\x05\x00\x80\xFF\x05\x00\x80\x80\xF9"
-     "\x19\x00"
-     "\x81\x21\x17\x00\x82\xA9\x06\x00\x83\xA9\x06\x00\x84\x35\x01\x00\x85\x3D"
-     "\x01\x00"
-     "\x86\x35\x01\x00\x87\x51\x17\x00\x88\x11\x01\x00\x89\xB9\x17\x00\x8A\x5D"
-     "\x15\x00"
-     "\x8B\x89\x15\x00\x8C\x91\x15\x00\x8D\x4D\x15\x00\x8E\x55\x12\x00\x8F\xF9"
-     "\x17\x00"
-     "\x90\x2D\x15\x00\x91\x71\x01\x00\x92\x71\x01\x00\x93\x71\x01\x00\x94\x11"
-     "\x01\x00"
-     "\x95\x11\x01\x00\x96\x11\x01\x00\x97\x11\x01\x00\x98\x31\x01\x00\x99\x31"
-     "\x01\x00"
-     "\x9A\x31\x01\x00\x9B\x8D\x17\x00\x9C\x2D\x01\x00\x9D\x15\x01\x00\x9E\x1D"
-     "\x01\x00"
-     "\x9F\x15\x01\x00\xA0\xED\x01\x00\xA1\xF5\x01\x00\xA2\xFD\x01\x00\xA3\xF5"
-     "\x01\x00"
-     "\xA4\x39\x06\x00\xA5\x39\x06\x00\xA6\x29\x06\x00\xA7\xE1\x01\x00\xA8\x21"
-     "\x01\x00"
-     "\xA9\x21\x01\x00\xAA\x21\x01\x00\xAB\x21\x01\x00\xAC\x21\x01\x00\xAD\x21"
-     "\x01\x00"
-     "\xAE\x21\x01\x00\xAF\x21\x01\x00\xB0\x61\x01\x00\xB1\x61\x01\x00\xB2\x61"
-     "\x01\x00"
-     "\xB3\x61\x01\x00\xB4\x61\x01\x00\xB5\x61\x01\x00\xB6\x61\x01\x00\xB7\x61"
-     "\x01\x00"
-     "\xB8\x21\x01\x00\xB9\x21\x01\x00\xBA\x21\x01\x00\xBB\x21\x01\x00\xBC\x21"
-     "\x01\x00"
-     "\xBD\x21\x01\x00\xBE\x21\x01\x00\xBF\x21\x01\x00\xBB\xD1\x01\x00\xBA\xC9"
-     "\x01\x00"
-     "\xB9\xC1\x01\x00\xB8\xB1\x01\x00\xBF\xF1\x01\x00\xBE\xC9\x01\x00\xBD\xC1"
-     "\x01\x00"
-     "\xBC\xC9\x01\x00\xB3\x89\x01\x00\xB2\x81\x01\x00\xB1\x89\x01\x00\xB0\x51"
-     "\x01\x00"
-     "\xB7\x89\x01\x00\xB6\x81\x01\x00\xB5\x89\x01\x00\xB4\x91\x01\x00\xAB\xDD"
-     "\x01\x00"
-     "\xAA\xDD\x01\x00\xA9\xDD\x01\x00\xA8\x3D\x01\x00\xAF\x29\x01\x00\xAE\x21"
-     "\x01\x00"
-     "\xAD\x9D\x01\x00\xAC\x9D\x01\x00\xA3\xF1\x01\x00\xA2\xE9\x01\x00\xA1\xE1"
-     "\x01\x00"
-     "\xA0\xE9\x01\x00\xA7\x1D\x01\x00\xA6\x1D\x01\x00\xA5\x1D\x01\x00\xA4\x1D"
-     "\x01\x00"
-     "\x9B\xD1\x01\x00\x9A\xC9\x01\x00\x99\xC1\x01\x00\x98\xC9\x01\x00\x9F\x11"
-     "\x01\x00"
-     "\x9E\xC9\x01\x00\x9D\xC1\x01\x00\x9C\xC9\x01\x00\x93\x91\x01\x00\x92\x89"
-     "\x01\x00"
-     "\x91\x81\x01\x00\x90\x89\x01\x00\x97\xF1\x01\x00\x96\x89\x01\x00\x95\x81"
-     "\x01\x00"
-     "\x94\x89\x01\x00\x8B\xD1\x01\x00\x8A\xC9\x01\x00\x89\xC1\x01\x00\x88\xC9"
-     "\x01\x00"
-     "\x8F\xF1\x01\x00\x8E\xC9\x01\x00\x8D\xC1\x01\x00\x8C\xC9\x01\x00\x83\x11"
-     "\x3E\x00"
-     "\x82\x09\x3E\x00\x81\x01\x3E\x00\x80\x09\x3E\x00\x87\xF1\x01\x00\x86\x09"
-     "\x3E\x00"
-     "\x85\x01\x3E\x00\x84\x09\x3E\x00\x16\x06\x00\x80\x2D\x06\x00\x80\x74\x03"
-     "\x00\x80"
-     "\x8D\x03\x00\x80\xA6\x03\x00\x80\x0D\x07\x00\x80\xD3\x06\x00\x80\xD7\x06"
-     "\x00\x80"
-     "\x7C\x07\x00\x80\x67\x25\x00\x80\xA0\x25\x00\x80\xBF\x25\x00\x80\xD1\x25"
-     "\x00\x80"
-     "\xF6\x25\x00\x80\x0E\x26\x00\x80\x4E\x26\x00\x80\xBB\x26\x00\x80\xFA\x26"
-     "\x00\x80"
-     "\x0C\x27\x00\x80\x36\x27\x00\x80\x5B\x27\x00\x80\x70\x27\x00\x80\x78\x27"
-     "\x00\x80"
-     "\x94\x27\x00\x80\xE7\x27\x00\x80\xF2\x27\x00\x80\xF6\x27\x00\x80\xFA\x27"
-     "\x00\x80"
-     "\xFE\x27\x00\x80\x5D\x29\x00\x80\x9C\x29\x00\x80\xC4\x29\x00\x80\x9C\x27"
-     "\x00\x80"
-     "\xAE\x27\x00\x80\xBA\x27\x00\x80\xC2\x27\x00\x80\xD1\x27\x00\x80\x0C\x28"
-     "\x00\x80"
-     "\x49\x28\x00\x80\x7D\x28\x00\x80\xC8\x29\x00\x80\xED\x29\x00\x80\xF5\x29"
-     "\x00\x80"
-     "\x11\x2A\x00\x80\x23\x2A\x00\x80\x27\x2A\x00\x80\x2B\x2A\x00\x80\x43\x2A"
-     "\x00\x80"
-     "\xC0\x28\x00\x80\x15\x29\x00\x80\x5E\x2A\x00\x80\x79\x2A\x00\x80\x7D\x2A"
-     "\x00\x80"
-     "\x8B\x2A\x00\x80\x8F\x2A\x00\x80\x93\x2A\x00\x80\x47\x2C\x00\x80\xF0\xA0"
-     "\x04\x00"
-     "\x6B\x25\x00\x80\x78\x25\x00\x80\x85\x25\x00\x80\x92\x25\x00\x80\xA4\x25"
-     "\x00\x80"
-     "\xB1\x25\x00\x80\xE3\x6C\x34\x00\xE3\x30\x34\x00\xE3\x14\x34\x00\xE3\xF8"
-     "\x33\x00"
-     "\xE3\x00\x35\x00\xE3\xF8\x34\x00\xE3\xAC\x34\x00\xE3\x9C\x34\x00\xEF\xFC"
-     "\x35\x00"
-     "\xEF\x20\x34\x00\xEF\x30\x34\x00\xEF\x7C\x34\x00\xEF\xB4\x34\x00\xEF\x94"
-     "\x34\x00"
-     "\xEF\xF4\x34\x00\xEF\x3C\x33\x00\xC3\x25\x00\x80\xD5\x25\x00\x80\xE3\x20"
-     "\x35\x00"
-     "\xE3\x7C\x35\x00\xE3\x84\x35\x00\xE3\xE0\x35\x00\xE3\xF4\x35\x00\xE3\xC4"
-     "\x35\x00"
-     "\xEF\x60\x35\x00\xEF\x50\x35\x00\xEF\x28\x35\x00\xEF\x08\x35\x00\xE5\x25"
-     "\x00\x80"
-     "\xFA\x25\x00\x80\xEF\xF8\x35\x00\xEF\xD4\x35\x00\xEF\xB4\x32\x00\xEF\xD0"
-     "\x32\x00"
-     "\x04\x26\x00\x80\x12\x26\x00\x80\xEF\xFC\x32\x00\x1C\x26\x00\x80\x29\x26"
-     "\x00\x80"
-     "\x3A\x26\x00\x80\xE3\x20\x32\x00\x44\x26\x00\x80\x9D\x26\x00\x80\xA4\x26"
-     "\x00\x80"
-     "\xE3\x00\x32\x00\xE3\x2C\x32\x00\x52\x26\x00\x80\xAE\x26\x00\x80\xE3\xE0"
-     "\x33\x00"
-     "\xE3\xC0\x33\x00\xE3\x34\x32\x00\xE3\x1C\x32\x00\xE3\x3C\x33\x00\xE3\x54"
-     "\x33\x00"
-     "\xE3\xA8\x33\x00\xE3\x88\x33\x00\xEF\xEC\x33\x00\xEF\xD0\x33\x00\xEF\x88"
-     "\x33\x00"
-     "\xEF\x60\x33\x00\xEF\x44\x33\x00\xEF\x20\x33\x00\xEF\x08\x33\x00\xEF\xEC"
-     "\x32\x00"
-     "\xB9\x91\x07\x00\xB8\x89\x07\x00\xBB\x91\x07\x00\xBA\x99\x07\x00\xBD\xB1"
-     "\x07\x00"
-     "\xBC\x89\x07\x00\xBF\xB1\x07\x00\xBE\xB9\x07\x00\xB1\x51\x07\x00\xB0\x49"
-     "\x07\x00"
-     "\xB3\x51\x07\x00\xB2\x59\x07\x00\xB5\xB1\x07\x00\xB4\x49\x07\x00\xB7\xB1"
-     "\x07\x00"
-     "\xB6\xB9\x07\x00\xA9\x11\x07\x00\xA8\x09\x07\x00\xAB\x11\x07\x00\xAA\x19"
-     "\x07\x00"
-     "\xAD\x31\x07\x00\xAC\x09\x07\x00\xAF\x31\x07\x00\xAE\x39\x07\x00\xA1\x51"
-     "\x07\x00"
-     "\xA0\x49\x07\x00\xA3\x51\x07\x00\xA2\x59\x07\x00\xA5\x31\x07\x00\xA4\x49"
-     "\x07\x00"
-     "\xA7\x31\x07\x00\xA6\x39\x07\x00\x99\x91\x07\x00\x98\x89\x07\x00\x9B\x91"
-     "\x07\x00"
-     "\x9A\x99\x07\x00\x9D\xB1\x07\x00\x9C\x89\x07\x00\x9F\xB1\x07\x00\x9E\xB9"
-     "\x07\x00"
-     "\x91\x0D\x01\x00\x90\x0D\x01\x00\x93\x15\x01\x00\x92\x15\x01\x00\x95\xB1"
-     "\x07\x00"
-     "\x94\x01\x04\x00\x97\xB1\x07\x00\x96\xB9\x07\x00\x89\xB5\x01\x00\x88\xB5"
-     "\x01\x00"
-     "\x8B\x45\x01\x00\x8A\x45\x01\x00\x8D\x49\x01\x00\x8C\x41\x01\x00\x8F\x7D"
-     "\x01\x00"
-     "\x8E\x41\x01\x00\x81\x0D\x02\x00\x80\x0D\x02\x00\x83\x1D\x02\x00\x82\x1D"
-     "\x02\x00"
-     "\x85\x0D\x02\x00\x84\x0D\x02\x00\x87\xBD\x01\x00\x86\xBD\x01\x00\x66\x26"
-     "\x00\x80"
-     "\x76\x26\x00\x80\x89\x26\x00\x80\xE0\x26\x00\x80\xEA\x26\x00\x80\xBF\x26"
-     "\x00\x80"
-     "\xCF\x26\x00\x80\xFE\x26\x00\x80\x1E\x27\x00\x80\x28\x27\x00\x80\x3A\x27"
-     "\x00\x80"
-     "\x47\x27\x00\x80\x5F\x27\x00\x80\x7C\x27\x00\x80\x86\x27\x00\x80\xA4\x27"
-     "\x00\x80"
-     "\xCA\x27\x00\x80\xE0\x27\x00\x80\xD5\x27\x00\x80\xEB\x27\x00\x80\x02\x28"
-     "\x00\x80"
-     "\x10\x28\x00\x80\x2F\x28\x00\x80\x3F\x28\x00\x80\x1E\x28\x00\x80\x5C\x28"
-     "\x00\x80"
-     "\x6C\x28\x00\x80\x76\x28\x00\x80\x51\x28\x00\x80\x81\x28\x00\x80\x91\x28"
-     "\x00\x80"
-     "\xAC\x28\x00\x80\x9E\x28\x00\x80\xC4\x28\x00\x80\xF0\x28\x00\x80\xF7\x28"
-     "\x00\x80"
-     "\xD4\x28\x00\x80\x01\x29\x00\x80\x0B\x29\x00\x80\xE2\x28\x00\x80\x46\x29"
-     "\x00\x80"
-     "\x38\x29\x00\x80\x19\x29\x00\x80\x53\x29\x00\x80\x20\x29\x00\x80\x2A\x29"
-     "\x00\x80"
-     "\x61\x29\x00\x80\x6E\x29\x00\x80\x78\x29\x00\x80\x82\x29\x00\x80\x8C\x29"
-     "\x00\x80"
-     "\xA0\x29\x00\x80\xAD\x29\x00\x80\xB4\x29\x00\x80\xCC\x29\x00\x80\xD3\x29"
-     "\x00\x80"
-     "\xE0\x29\x00\x80\xF9\x29\x00\x80\x03\x2A\x00\x80\x15\x2A\x00\x80\x2F\x2A"
-     "\x00\x80"
-     "\x3C\x2A\x00\x80\x47\x2A\x00\x80\x51\x2A\x00\x80\x80\xAD\x01\x00\x81\xB5"
-     "\x01\x00"
-     "\x82\xBD\x01\x00\x83\xB5\x01\x00\x84\xAD\x01\x00\x85\xD5\x01\x00\x86\xDD"
-     "\x01\x00"
-     "\x87\xD5\x01\x00\x88\xED\x01\x00\x89\xF5\x01\x00\x8A\xFD\x01\x00\x8B\xF5"
-     "\x01\x00"
-     "\x8C\xED\x01\x00\x8D\xD5\x01\x00\x8E\xDD\x01\x00\x8F\xD5\x01\x00\x90\xAD"
-     "\x01\x00"
-     "\x91\xB5\x01\x00\x92\xBD\x01\x00\x93\xB5\x01\x00\x94\xAD\x01\x00\x95\x55"
-     "\x03\x00"
-     "\x96\x5D\x03\x00\x97\x55\x03\x00\x98\x6D\x03\x00\x99\x75\x03\x00\x9A\x7D"
-     "\x03\x00"
-     "\x9B\x75\x03\x00\x9C\x6D\x03\x00\x9D\x55\x03\x00\x9E\x5D\x03\x00\x9F\x55"
-     "\x03\x00"
-     "\xA0\xAD\x03\x00\xA1\xB5\x03\x00\xA2\xBD\x03\x00\xA3\xB5\x03\x00\xA4\xAD"
-     "\x03\x00"
-     "\xA5\xD5\x03\x00\xA6\xDD\x03\x00\xA7\xD5\x03\x00\xA8\xED\x03\x00\xA9\xF5"
-     "\x03\x00"
-     "\xAA\xFD\x03\x00\xAB\xF5\x03\x00\xAC\xED\x03\x00\xAD\xD5\x03\x00\xAE\xDD"
-     "\x03\x00"
-     "\xAF\xD5\x03\x00\xB0\xAD\x03\x00\xB1\xB5\x03\x00\xB2\xBD\x03\x00\xB3\xB5"
-     "\x03\x00"
-     "\xB4\xAD\x03\x00\xB5\x55\x03\x00\xB6\x5D\x03\x00\xB7\x55\x03\x00\xB8\x6D"
-     "\x03\x00"
-     "\xB9\x75\x03\x00\xBA\x7D\x03\x00\xBB\x75\x03\x00\xBC\x6D\x03\x00\xBD\x55"
-     "\x03\x00"
-     "\xBE\x5D\x03\x00\xBF\x55\x03\x00\x62\x2A\x00\x80\x69\x2A\x00\x80\x81\x2A"
-     "\x00\x80"
-     "\x5A\x03\x00\x80\x7D\x04\x00\x80\x1C\x05\x00\x80\x6F\x05\x00\x80\xAA\x05"
-     "\x00\x80"
-     "\xD0\x05\x00\x80\xE7\x05\x00\x80\x04\x06\x00\x80\x1B\x06\x00\x80\x32\x06"
-     "\x00\x80"
-     "\x7A\x03\x00\x80\x93\x03\x00\x80\xAC\x03\x00\x80\xBF\x03\x00\x80\xD2\x03"
-     "\x00\x80"
-     "\xE5\x03\x00\x80\xF8\x03\x00\x80\x0B\x04\x00\x80\x1E\x04\x00\x80\x31\x04"
-     "\x00\x80"
-     "\x97\x04\x00\x80\xA6\x04\x00\x80\xB5\x04\x00\x80\xC4\x04\x00\x80\xD3\x04"
-     "\x00\x80"
-     "\x13\x07\x00\x80\xB3\x06\x00\x80\x71\x06\x00\x80\x80\x06\x00\x80\xF5\x07"
-     "\x00\x80"
-     "\x02\x08\x00\x80\xBF\x06\x00\x80\xC2\x06\x00\x80\xC6\x06\x00\x80\x2C\x07"
-     "\x00\x80"
-     "\x11\x30\x00\x80\xFF\x31\x00\x80\x0E\x2F\x00\x80\xF0\x31\x00\x80\x7A\x32"
-     "\x00\x80"
-     "\xFF\x07\x00\x80\xDB\x07\x00\x80\x16\x0C\x00\x80\x8F\x07\x00\x80\x39\x07"
-     "\x00\x80"
-     "\x3C\x07\x00\x80\x92\x07\x00\x80\xFA\x06\x00\x80\x91\x06\x00\x80\x3F\x07"
-     "\x00\x80"
-     "\x05\x08\x00\x80\xDE\x07\x00\x80\x1A\x0C\x00\x80\x1E\x0C\x00\x80\x9B\x07"
-     "\x00\x80"
-     "\x44\x07\x00\x80\x1A\x07\x00\x80\x47\x07\x00\x80\x9E\x07\x00\x80\xFD\x06"
-     "\x00\x80"
-     "\x59\x08\x00\x80\x91\xF0\x0F\x00\x22\x0C\x00\x80\xA6\x07\x00\x80\xBC\x06"
-     "\x00\x80"
-     "\x4E\x07\x00\x80\xF1\x06\x00\x80\xE4\x07\x00\x80\x26\x0C\x00\x80\xA9\x07"
-     "\x00\x80"
-     "\x9B\x06\x00\x80\x51\x07\x00\x80\xAC\x07\x00\x80\x9E\x06\x00\x80\x89\x07"
-     "\x00\x80"
-     "\x54\x07\x00\x80\xB0\x07\x00\x80\xA2\x06\x00\x80\x8C\x07\x00\x80\x58\x07"
-     "\x00\x80"
-     "\xCC\x07\x00\x80\xD2\x07\x00\x80\x5C\x07\x00\x80\xB8\x07\x00\x80\x01\x07"
-     "\x00\x80"
-     "\x27\x08\x00\x80\x2B\x08\x00\x80\x33\x08\x00\x80\x16\x08\x00\x80\xEA\x07"
-     "\x00\x80"
-     "\x2A\x0C\x00\x80\xBF\x07\x00\x80\x1D\x08\x00\x80\xED\x07\x00\x80\x2E\x0C"
-     "\x00\x80"
-     "\xC2\x07\x00\x80\x63\x07\x00\x80\x20\x08\x00\x80\xF0\x07\x00\x80\x32\x0C"
-     "\x00\x80"
-     "\xC5\x07\x00\x80\x66\x07\x00\x80\x69\x07\x00\x80\xC8\x07\x00\x80\x64\x06"
-     "\x00\x80"
-     "\x84\x06\x00\x80\x5F\x26\x00\x80\x95\x06\x00\x80\x98\x06\x00\x80\xAA\x06"
-     "\x00\x80"
-     "\xA6\x06\x00\x80\xB6\x06\x00\x80\x05\x07\x00\x80\x0A\x07\x00\x80\x10\x07"
-     "\x00\x80"
-     "\x29\x07\x00\x80\x4B\x07\x00\x80\x60\x07\x00\x80\x72\x07\x00\x80\x75\x07"
-     "\x00\x80"
-     "\x78\x07\x00\x80\x80\x07\x00\x80\xA2\x07\x00\x80\xB4\x07\x00\x80\x08\x08"
-     "\x00\x80"
-     "\xBE\xA1\x0F\x00\xBF\x99\x0F\x00\xBC\x91\x0F\x00\xBD\x91\x0F\x00\xBA\x81"
-     "\x0F\x00"
-     "\xBB\x81\x0F\x00\xB8\x95\x0F\x00\xB9\x95\x0F\x00\xB6\xA5\x0F\x00\xB7\xAD"
-     "\x0F\x00"
-     "\xB4\x55\x0F\x00\xB5\xAD\x0F\x00\xB2\x45\x0F\x00\xB3\x4D\x0F\x00\xB0\x55"
-     "\x0F\x00"
-     "\xB1\x4D\x0F\x00\xAE\x25\x0F\x00\xAF\x2D\x0F\x00\xAC\x15\x0F\x00\xAD\x2D"
-     "\x0F\x00"
-     "\xAA\x05\x0F\x00\xAB\x0D\x0F\x00\xA8\x09\x0F\x00\xA9\xF1\x0E\x00\xA6\x29"
-     "\x0F\x00"
-     "\xA7\x29\x0F\x00\xA4\x29\x0F\x00\xA5\x29\x0F\x00\xA2\x49\x0F\x00\xA3\x49"
-     "\x0F\x00"
-     "\xA0\x49\x0F\x00\xA1\x49\x0F\x00\x9E\xA9\x0F\x00\x9F\xA9\x0F\x00\x9C\xA9"
-     "\x0F\x00"
-     "\x9D\xA9\x0F\x00\x9A\x89\x0F\x00\x9B\x89\x0F\x00\x98\x89\x0F\x00\x99\x89"
-     "\x0F\x00"
-     "\x96\xA9\x0F\x00\x97\xA9\x0F\x00\x94\x55\x0E\x00\x95\xA9\x0F\x00\x92\x45"
-     "\x0E\x00"
-     "\x93\x4D\x0E\x00\x90\x55\x0E\x00\x91\x4D\x0E\x00\x8E\x25\x0E\x00\x8F\x2D"
-     "\x0E\x00"
-     "\x8C\x15\x0E\x00\x8D\x2D\x0E\x00\x8A\x05\x0E\x00\x8B\x0D\x0E\x00\x88\x15"
-     "\x0E\x00"
-     "\x89\x0D\x0E\x00\x86\x25\x0E\x00\x87\x2D\x0E\x00\x84\x55\x0E\x00\x85\x2D"
-     "\x0E\x00"
-     "\x82\x45\x0E\x00\x83\x4D\x0E\x00\x80\x55\x0E\x00\x81\x4D\x0E\x00\x80\xA5"
-     "\x0E\x00"
-     "\x81\xAD\x0E\x00\x82\xA5\x0E\x00\x83\xBD\x0E\x00\x84\xA9\x0E\x00\x85\xA9"
-     "\x0E\x00"
-     "\x86\xD9\x0E\x00\x87\xD9\x0E\x00\x88\xE9\x0E\x00\x89\xE9\x0E\x00\x8A\xF9"
-     "\x0E\x00"
-     "\x8B\xF9\x0E\x00\x8C\xE9\x0E\x00\x8D\xE9\x0E\x00\x8E\x11\x0E\x00\x8F\xDD"
-     "\x0E\x00"
-     "\x90\xA5\x0E\x00\x91\xAD\x0E\x00\x92\xA5\x0E\x00\x93\xBD\x0E\x00\x94\x95"
-     "\x0F\x00"
-     "\x95\xA1\x0E\x00\x96\xA1\x0E\x00\x97\xA1\x0E\x00\x98\x61\x03\x00\x99\x61"
-     "\x03\x00"
-     "\x9A\x61\x03\x00\x9B\x61\x03\x00\x9C\x61\x03\x00\x9D\x61\x03\x00\x9E\x61"
-     "\x03\x00"
-     "\x9F\x61\x03\x00\xA0\xA1\x03\x00\xA1\xA1\x03\x00\xA2\xA1\x03\x00\xA3\xA1"
-     "\x03\x00"
-     "\xA4\xA1\x03\x00\xA5\xA1\x03\x00\xA6\xA1\x03\x00\xA7\xA1\x03\x00\xA8\xE1"
-     "\x03\x00"
-     "\xA9\xE1\x03\x00\xAA\xE1\x03\x00\xAB\xE1\x03\x00\xAC\xE1\x03\x00\xAD\xE1"
-     "\x03\x00"
-     "\xAE\xE1\x03\x00\xAF\xE1\x03\x00\xB0\xA1\x03\x00\xB1\xA1\x03\x00\xB2\xA1"
-     "\x03\x00"
-     "\xB3\xA1\x03\x00\xB4\xA1\x03\x00\xB5\xA1\x03\x00\xB6\xA1\x03\x00\xB7\xA1"
-     "\x03\x00"
-     "\xB8\x61\x03\x00\xB9\x61\x03\x00\xBA\x61\x03\x00\xBB\x61\x03\x00\xBC\x61"
-     "\x03\x00"
-     "\xBD\x61\x03\x00\xBE\x61\x03\x00\xBF\x61\x03\x00\xFA\x07\x00\x80\x0B\x08"
-     "\x00\x80"
-     "\x13\x08\x00\x80\x47\x08\x00\x80\x50\x08\x00\x80\x78\x08\x00\x80\x70\x08"
-     "\x00\x80"
-     "\x78\x06\x00\x80\x73\x04\x00\x80\x12\x05\x00\x80\x65\x05\x00\x80\xA0\x05"
-     "\x00\x80"
-     "\xC6\x05\x00\x80\xDD\x05\x00\x80\xFA\x05\x00\x80\x11\x06\x00\x80\x28\x06"
-     "\x00\x80"
-     "\x6E\x03\x00\x80\x87\x03\x00\x80\xA0\x03\x00\x80\xB9\x03\x00\x80\xCC\x03"
-     "\x00\x80"
-     "\xDF\x03\x00\x80\xF2\x03\x00\x80\x05\x04\x00\x80\x18\x04\x00\x80\x2B\x04"
-     "\x00\x80"
-     "\x91\x04\x00\x80\xA0\x04\x00\x80\xAF\x04\x00\x80\xBE\x04\x00\x80\xCD\x04"
-     "\x00\x80"
-     "\xDC\x04\x00\x80\xE5\x04\x00\x80\xEE\x04\x00\x80\xF7\x04\x00\x80\x00\x05"
-     "\x00\x80"
-     "\x2C\x05\x00\x80\x35\x05\x00\x80\xF6\x06\x00\x80\x6F\x3B\x00\x80\x73\x3B"
-     "\x00\x80"
-     "\x77\x3B\x00\x80\x7B\x3B\x00\x80\x7F\x3B\x00\x80\x87\x3B\x00\x80\x8B\x3B"
-     "\x00\x80"
-     "\x8F\x3B\x00\x80\x93\x3B\x00\x80\x97\x3B\x00\x80\x9B\x3B\x00\x80\x9F\x3B"
-     "\x00\x80"
-     "\xA3\x3B\x00\x80\xA7\x3B\x00\x80\xAB\x3B\x00\x80\xAF\x3B\x00\x80\xB3\x3B"
-     "\x00\x80"
-     "\xB7\x3B\x00\x80\xBB\x3B\x00\x80\xBF\x3B\x00\x80\xC3\x3B\x00\x80\xC7\x3B"
-     "\x00\x80"
-     "\x82\x50\x18\x00\xCB\x3B\x00\x80\x99\x00\x04\x00\xCF\x3B\x00\x80\xD3\x3B"
-     "\x00\x80"
-     "\x9A\x0C\x05\x00\x9D\x14\x04\x00\x9C\x08\x02\x00\x9F\x18\x06\x00\x9E\x1C"
-     "\x07\x00"
-     "\xD7\x3B\x00\x80\xDB\x3B\x00\x80\xDF\x3B\x00\x80\xE3\x3B\x00\x80\xE7\x3B"
-     "\x00\x80"
-     "\xEB\x3B\x00\x80\xEF\x3B\x00\x80\xF3\x3B\x00\x80\xF7\x3B\x00\x80\xFB\x3B"
-     "\x00\x80"
-     "\xFF\x3B\x00\x80\x03\x3C\x00\x80\x07\x3C\x00\x80\x0B\x3C\x00\x80\x0F\x3C"
-     "\x00\x80"
-     "\x13\x3C\x00\x80\x17\x3C\x00\x80\x1B\x3C\x00\x80\x1F\x3C\x00\x80\x23\x3C"
-     "\x00\x80"
-     "\x27\x3C\x00\x80\x2B\x3C\x00\x80\x2F\x3C\x00\x80\x33\x3C\x00\x80\x37\x3C"
-     "\x00\x80"
-     "\xA6\xD1\x00\x00\x3B\x3C\x00\x80\xA4\xDD\x00\x00\x3F\x3C\x00\x80\xA2\xD9"
-     "\x00\x00"
-     "\x43\x3C\x00\x80\x47\x3C\x00\x80\x4B\x3C\x00\x80\xAE\xC1\x00\x00\x4F\x3C"
-     "\x00\x80"
-     "\xAC\xCD\x00\x00\x53\x3C\x00\x80\xAA\xE9\x00\x00\x57\x3C\x00\x80\xA8\xE5"
-     "\x00\x00"
-     "\x5B\x3C\x00\x80\xB6\xB5\x00\x00\x5F\x3C\x00\x80\xB4\xA1\x00\x00\xAD\x98"
-     "\x06\x00"
-     "\xB2\xBD\x00\x00\x63\x3C\x00\x80\xAE\x00\x18\x00\x67\x3C\x00\x80\xBE\x65"
-     "\x00\x00"
-     "\x6B\x3C\x00\x80\xBC\x91\x00\x00\x6F\x3C\x00\x80\xBA\x8D\x00\x00\x73\x3C"
-     "\x00\x80"
-     "\xB8\x89\x00\x00\x80\xFD\x06\x00\x77\x3C\x00\x80\x82\x0D\x07\x00\x7B\x3C"
-     "\x00\x80"
-     "\x84\x19\x07\x00\x7F\x3C\x00\x80\x86\x39\x07\x00\x83\x3C\x00\x80\x88\x05"
-     "\x07\x00"
-     "\x87\x3C\x00\x80\x8A\x01\x07\x00\x8B\x3C\x00\x80\x8C\x1D\x07\x00\x8F\x3C"
-     "\x00\x80"
-     "\x8E\x09\x07\x00\x93\x3C\x00\x80\x90\x75\x07\x00\x97\x3C\x00\x80\x92\x71"
-     "\x07\x00"
-     "\x9B\x3C\x00\x80\x94\x6D\x07\x00\x9F\x3C\x00\x80\x96\x19\x07\x00\xAB\x1C"
-     "\x00\x80"
-     "\x98\x25\x07\x00\xA3\x3C\x00\x80\x9A\x21\x07\x00\xA7\x3C\x00\x80\x9C\x51"
-     "\x72\x02"
-     "\x9D\x59\x72\x02\xAB\x3C\x00\x80\xAF\x3C\x00\x80\xB3\x3C\x00\x80\xB7\x3C"
-     "\x00\x80"
-     "\xBF\x3C\x00\x80\xC3\x3C\x00\x80\xCB\x3C\x00\x80\xCF\x3C\x00\x80\xD3\x3C"
-     "\x00\x80"
-     "\xD7\x3C\x00\x80\xDB\x3C\x00\x80\xDF\x3C\x00\x80\xE3\x3C\x00\x80\xE7\x3C"
-     "\x00\x80"
-     "\xEB\x3C\x00\x80\xEF\x3C\x00\x80\x11\x0B\x00\x80\x8C\x0B\x00\x80\x56\x0B"
-     "\x00\x80"
-     "\x33\x3D\x00\x80\xF3\x3C\x00\x80\xF7\x3C\x00\x80\xFB\x3C\x00\x80\xFF\x3C"
-     "\x00\x80"
-     "\x03\x3D\x00\x80\x07\x3D\x00\x80\x0B\x3D\x00\x80\x0F\x3D\x00\x80\xC7\x3C"
-     "\x00\x80"
-     "\xBB\x1C\x00\x80\x13\x3D\x00\x80\x17\x3D\x00\x80\x1B\x3D\x00\x80\x1F\x3D"
-     "\x00\x80"
-     "\x80\xA1\x07\x00\x80\x8D\x06\x00\x82\xAD\x07\x00\x82\x9D\x06\x00\x84\xB9"
-     "\x07\x00"
-     "\x84\x8D\x06\x00\x86\xB5\x07\x00\x86\x81\x06\x00\x88\x91\x07\x00\x88\x85"
-     "\x06\x00"
-     "\x8A\x41\x5C\x00\x8A\x85\x06\x00\x8C\x85\x07\x00\x8C\x85\x06\x00\x8E\x81"
-     "\x07\x00"
-     "\x8E\x85\x06\x00\x90\xFD\x07\x00\x90\x85\x06\x00\x92\x89\x07\x00\x92\x85"
-     "\x06\x00"
-     "\x94\x95\x07\x00\x94\x8D\x06\x00\x96\x91\x07\x00\x96\xBD\x06\x00\x98\xAD"
-     "\x07\x00"
-     "\x98\x8D\x06\x00\x9A\xB9\x07\x00\x9A\x9D\x06\x00\x9C\xA5\x07\x00\x9C\x8D"
-     "\x06\x00"
-     "\x9E\xA1\x07\x00\x9E\x79\x01\x00\xA0\x5D\x07\x00\xA0\x89\x01\x00\xA2\x69"
-     "\x07\x00"
-     "\xA2\x99\x01\x00\xA4\x75\x07\x00\xA4\x89\x01\x00\xA6\x71\x07\x00\xA6\xB9"
-     "\x01\x00"
-     "\xA8\x4D\x07\x00\xA8\x89\x01\x00\xAA\x59\x07\x00\xAA\x99\x01\x00\xAC\x45"
-     "\x07\x00"
-     "\xAC\x89\x01\x00\x23\x3D\x00\x80\xAE\xF9\x01\x00\x27\x3D\x00\x80\xB0\x81"
-     "\x01\x00"
-     "\x2B\x3D\x00\x80\x2F\x3D\x00\x80\x68\x0B\x00\x80\x0F\x1B\x00\x80\x13\x1B"
-     "\x00\x80"
-     "\x17\x1B\x00\x80\xB9\xAD\x01\x00\x1B\x1B\x00\x80\xBB\xBD\x01\x00\x1F\x1B"
-     "\x00\x80"
-     "\xBD\xAD\x01\x00\x23\x1B\x00\x80\x27\x1B\x00\x80\xBE\x99\x01\x00\x80\x6D"
-     "\x02\x00"
-     "\x2B\x1B\x00\x80\x82\x71\x02\x00\x2F\x1B\x00\x80\x84\x6D\x02\x00\x33\x1B"
-     "\x00\x80"
-     "\x86\x99\x02\x00\x37\x1B\x00\x80\x3B\x1B\x00\x80\x3F\x1B\x00\x80\x43\x1B"
-     "\x00\x80"
-     "\x8B\xA9\x02\x00\x47\x1B\x00\x80\x8D\x01\x5D\x00\x4B\x1B\x00\x80\x4F\x1B"
-     "\x00\x80"
-     "\x90\xC9\x02\x00\x53\x1B\x00\x80\x92\xC5\x02\x00\x57\x1B\x00\x80\x5B\x1B"
-     "\x00\x80"
-     "\x5F\x1B\x00\x80\x96\xC9\x02\x00\x63\x1B\x00\x80\x98\xF5\x02\x00\x67\x1B"
-     "\x00\x80"
-     "\x9A\xF1\x02\x00\x6B\x1B\x00\x80\x9C\xED\x02\x00\x6F\x1B\x00\x80\x9E\xD9"
-     "\x02\x00"
-     "\x73\x1B\x00\x80\xA0\x25\x02\x00\x77\x1B\x00\x80\xA2\x21\x02\x00\x7B\x1B"
-     "\x00\x80"
-     "\xA4\x3D\x02\x00\x7F\x1B\x00\x80\xA6\x29\x02\x00\x83\x1B\x00\x80\xA8\x15"
-     "\x02\x00"
-     "\x87\x1B\x00\x80\xAA\x0D\x61\x02\xAB\x31\x5C\x00\xAC\x81\x5D\x00\xAD\x0D"
-     "\x02\x00"
-     "\xAE\x81\x5D\x00\x8B\x1B\x00\x80\xB0\x7D\x02\x00\xB1\x05\x02\x00\xB2\x25"
-     "\x5D\x00"
-     "\xB3\x09\x02\x00\xB4\x19\x02\x00\x8F\x1B\x00\x80\xB6\x15\x02\x00\x93\x1B"
-     "\x00\x80"
-     "\xB8\x31\x02\x00\x97\x1B\x00\x80\xBA\x3D\x02\x00\x9B\x1B\x00\x80\xBC\x29"
-     "\x02\x00"
-     "\x9F\x1B\x00\x80\xBE\x25\x02\x00\xA3\x1B\x00\x80\x80\xE1\x03\x00\x91\xD0"
-     "\x1E\x00"
-     "\x82\xED\x03\x00\xA7\x1B\x00\x80\x84\xF9\x03\x00\x85\x2D\x5B\x00\x86\xF5"
-     "\x03\x00"
-     "\x87\xED\x03\x00\xAB\x1B\x00\x80\x89\xD1\x03\x00\xAF\x1B\x00\x80\xB3\x1B"
-     "\x00\x80"
-     "\x9E\xDD\x2B\x00\x9F\xF9\x02\x00\x9C\x6D\x04\x00\x9D\xF5\x02\x00\x90\xB9"
-     "\x03\x00"
-     "\xB7\x1B\x00\x80\xBB\x1B\x00\x80\xBF\x1B\x00\x80\xC3\x1B\x00\x80\xC7\x1B"
-     "\x00\x80"
-     "\x96\xA5\x03\x00\xCB\x1B\x00\x80\x98\xD9\x02\x00\xCF\x1B\x00\x80\xD3\x1B"
-     "\x00\x80"
-     "\xD7\x1B\x00\x80\xDB\x1B\x00\x80\xDF\x1B\x00\x80\xE3\x1B\x00\x80\xE7\x1B"
-     "\x00\x80"
-     "\xB2\x5D\x02\x00\xB3\x61\x02\x00\xB0\x5D\x02\x00\xB1\x55\x02\x00\xB6\x69"
-     "\x02\x00"
-     "\xB7\x91\x01\x00\xB4\x75\x02\x00\xB5\x69\x02\x00\xBA\xBD\x01\x00\xBB\x85"
-     "\x01\x00"
-     "\xB8\xA5\x01\x00\xB9\xB9\x01\x00\xBE\x81\x01\x00\xBF\x8D\x01\x00\xBC\x9D"
-     "\x01\x00"
-     "\xBD\x95\x01\x00\xEB\x1B\x00\x80\xEF\x1B\x00\x80\xB2\x39\x6A\x02\xB3\xE1"
-     "\x6B\x02"
-     "\xB4\x75\x68\x02\xB5\x15\x02\x00\xF3\x1B\x00\x80\xF7\x1B\x00\x80\xB8\xB1"
-     "\x62\x02"
-     "\xB9\x7D\x62\x02\xFB\x1B\x00\x80\xA9\x3D\x02\x00\xFF\x1B\x00\x80\x03\x1C"
-     "\x00\x80"
-     "\x07\x1C\x00\x80\x0B\x1C\x00\x80\x0F\x1C\x00\x80\x13\x1C\x00\x80\x17\x1C"
-     "\x00\x80"
-     "\x1B\x1C\x00\x80\x1F\x1C\x00\x80\x23\x1C\x00\x80\x27\x1C\x00\x80\x2B\x1C"
-     "\x00\x80"
-     "\x2F\x1C\x00\x80\x33\x1C\x00\x80\x37\x1C\x00\x80\x3B\x1C\x00\x80\x3F\x1C"
-     "\x00\x80"
-     "\x43\x1C\x00\x80\x47\x1C\x00\x80\x4B\x1C\x00\x80\xBA\x21\xEE\x00\xAF\x38"
-     "\x00\x80"
-     "\x26\x32\x00\x80\xD7\x38\x00\x80\x8F\x33\x00\x80\xF3\x2B\x00\x80\xEE\x2D"
-     "\x00\x80"
-     "\x21\x2F\x00\x80\x52\x2E\x00\x80\x1D\x2F\x00\x80\x0C\x31\x00\x80\x07\x35"
-     "\x00\x80"
-     "\x97\x36\x00\x80\xCF\x37\x00\x80\x0B\x38\x00\x80\x3F\x38\x00\x80\xA4\xC0"
-     "\x00\x00"
-     "\xA5\x00\x04\x00\xA6\x88\x07\x00\xA7\x00\x08\x00\x53\x39\x00\x80\xAB\x32"
-     "\x00\x80"
-     "\x37\x33\x00\x80\xE3\x33\x00\x80\xAC\x00\xF0\x00\xAD\xA4\xF2\x00\xAE\xFC"
-     "\xF1\x00"
-     "\xAF\x00\xF4\x00\xA8\xB0\x0B\x00\xA9\x00\x0C\x00\xAA\xB8\x0E\x00\xAB\x1C"
-     "\x0D\x00"
-     "\xB4\x94\xF9\x00\xB5\x00\xFC\x00\xB6\xC8\xFF\x00\xB7\xCC\xFD\x00\xB0\xD0"
-     "\xF6\x00"
-     "\xB1\x80\xF5\x00\xB2\x00\xF8\x00\xB3\xDC\xFA\x00\xBC\xE4\xE7\x00\xBD\x00"
-     "\xE8\x00"
-     "\xBE\xE8\xEA\x00\xBF\x34\xE1\x00\xB8\x00\xE0\x00\xB9\xF4\xE2\x00\xBA\xF8"
-     "\xE3\x00"
-     "\xBB\x00\xE4\x00\xB7\x1D\x01\x00\xB6\x1D\x01\x00\xB5\x1D\x01\x00\xB4\x1D"
-     "\x01\x00"
-     "\xB3\x1D\x01\x00\xB2\x1D\x01\x00\xB1\x1D\x01\x00\xB0\x1D\x01\x00\xBF\x1D"
-     "\x01\x00"
-     "\xBE\x1D\x01\x00\xBD\x1D\x01\x00\xBC\x1D\x01\x00\xBB\x1D\x01\x00\xBA\x1D"
-     "\x01\x00"
-     "\xB9\x1D\x01\x00\xB8\x1D\x01\x00\xA7\x81\x1E\x00\xA6\x39\x1E\x00\xA5\x31"
-     "\x1E\x00"
-     "\xA4\x39\x1E\x00\xA3\x21\x1E\x00\xA2\x39\x1E\x00\xA1\x31\x1E\x00\xA0\x39"
-     "\x1E\x00"
-     "\xAF\x9D\x1E\x00\xAE\x9D\x1E\x00\xAD\x9D\x1E\x00\xAC\x9D\x1E\x00\xAB\x99"
-     "\x1E\x00"
-     "\xAA\xA9\x1E\x00\xA9\xA1\x1E\x00\xA8\xB9\x1E\x00\x97\xA1\x1E\x00\x96\x89"
-     "\x1E\x00"
-     "\x95\x75\x1F\x00\x94\x79\x1F\x00\x93\x7D\x1F\x00\x92\x71\x1F\x00\x91\x45"
-     "\x1F\x00"
-     "\x90\x49\x1F\x00\x9F\xFD\x1E\x00\x9E\xF1\x1E\x00\x9D\xF9\x1E\x00\x9C\xE1"
-     "\x1E\x00"
-     "\x9B\xF9\x1E\x00\x9A\xF1\x1E\x00\x99\x95\x1E\x00\x98\x95\x1E\x00\x87\xAD"
-     "\x1F\x00"
-     "\x86\xA1\x1F\x00\x85\x95\x1F\x00\x84\x99\x1F\x00\x83\x81\x1F\x00\x82\x99"
-     "\x1F\x00"
-     "\x81\x9D\x1F\x00\x80\xE1\x1F\x00\x8F\x5D\x1F\x00\x8E\x41\x1F\x00\x8D\x75"
-     "\x1F\x00"
-     "\x8C\x79\x1F\x00\x8B\x7D\x1F\x00\x8A\x71\x1F\x00\x89\x85\x1F\x00\x88\x89"
-     "\x1F\x00"
-     "\x90\x75\x02\x00\x91\x7D\x02\x00\x92\x75\x02\x00\x93\x9D\x02\x00\x94\x85"
-     "\x02\x00"
-     "\x95\x8D\x02\x00\x96\x85\x02\x00\x97\x1D\x01\x00\x98\x25\x01\x00\x99\x2D"
-     "\x01\x00"
-     "\x9A\x25\x01\x00\x9B\x3D\x01\x00\x9C\x25\x01\x00\x9D\x2D\x01\x00\x9E\x25"
-     "\x01\x00"
-     "\x9F\x1D\x01\x00\x80\x05\x02\x00\x81\x0D\x02\x00\x82\x05\x3F\x00\x83\x01"
-     "\x02\x00"
-     "\x84\x01\x02\x00\x85\x01\x02\x00\x86\x01\x02\x00\x87\xDD\x3C\x00\x88\xE1"
-     "\x3C\x00"
-     "\x89\x39\x02\x00\x8A\x05\x3F\x00\x8B\x0D\x02\x00\x8C\x15\x02\x00\x8D\x1D"
-     "\x02\x00"
-     "\x8E\x15\x02\x00\x8F\x0D\x02\x00\xB0\xE5\x01\x00\xB1\xED\x01\x00\xB2\xE5"
-     "\x01\x00"
-     "\xB3\xFD\x01\x00\xB4\x19\x39\x00\xB5\xE1\x01\x00\xB6\xE1\x01\x00\xB7\xE1"
-     "\x01\x00"
-     "\xB8\x21\x01\x00\xB9\x21\x01\x00\xBA\x21\x01\x00\xBB\x21\x01\x00\xBC\x21"
-     "\x01\x00"
-     "\xBD\x21\x01\x00\xBE\x21\x01\x00\xBF\x21\x01\x00\xA0\xE5\x01\x00\xA1\xED"
-     "\x01\x00"
-     "\xA2\xE5\x01\x00\xA3\xFD\x01\x00\xA4\xE5\x01\x00\xA5\xED\x01\x00\xA6\xE5"
-     "\x01\x00"
-     "\xA7\x9D\x01\x00\xA8\xA5\x01\x00\xA9\xAD\x01\x00\xAA\xA5\x01\x00\xAB\xBD"
-     "\x01\x00"
-     "\xAC\xA5\x01\x00\xAD\xAD\x01\x00\xAE\xA5\x01\x00\xAF\x9D\x01\x00\x7B\x34"
-     "\x00\x80"
-     "\x97\x37\x00\x80\x7F\x39\x00\x80\xAB\x3A\x00\x80\x17\x2C\x00\x80\xC2\x2D"
-     "\x00\x80"
-     "\xBF\x32\x00\x80\x07\x34\x00\x80\xCB\x37\x00\x80\xFB\x3A\x00\x80\xC9\x2F"
-     "\x00\x80"
-     "\xAF\x33\x00\x80\xBB\x37\x00\x80\x4F\x38\x00\x80\x40\x31\x00\x80\xF3\x36"
-     "\x00\x80"
-     "\x17\x38\x00\x80\x3C\x30\x00\x80\x3A\x32\x00\x80\x4F\x33\x00\x80\x53\x34"
-     "\x00\x80"
-     "\x5F\x39\x00\x80\x5E\x2C\x00\x80\xF6\x2C\x00\x80\x62\x2D\x00\x80\xA8\x31"
-     "\x00\x80"
-     "\xB7\x32\x00\x80\x03\x34\x00\x80\x27\x35\x00\x80\xC3\x37\x00\x80\xDB\x37"
-     "\x00\x80"
-     "\xFF\x38\x00\x80\x27\x3A\x00\x80\xDB\x3A\x00\x80\xF7\x3A\x00\x80\x63\x35"
-     "\x00\x80"
-     "\xA3\x35\x00\x80\x5B\x36\x00\x80\x7B\x37\x00\x80\xB3\x39\x00\x80\x83\x38"
-     "\x00\x80"
-     "\xD6\x2E\x00\x80\x58\x30\x00\x80\x13\x36\x00\x80\xD3\x36\x00\x80\x37\x34"
-     "\x00\x80"
-     "\x6B\x35\x00\x80\xCF\x38\x00\x80\x1F\x3A\x00\x80\xD2\x2E\x00\x80\xA5\x2F"
-     "\x00\x80"
-     "\xAF\x32\x00\x80\x63\x33\x00\x80\x97\x33\x00\x80\x4B\x36\x00\x80\x82\xB0"
-     "\x05\x00"
-     "\x73\x36\x00\x80\xDF\x39\x00\x80\x5E\x2D\x00\x80\xDF\x36\x00\x80\x02\x2D"
-     "\x00\x80"
-     "\xFE\x2C\x00\x80\xCF\x35\x00\x80\x5F\x36\x00\x80\x93\x37\x00\x80\xE7\x39"
-     "\x00\x80"
-     "\x9F\x38\x00\x80\x44\x31\x00\x80\x8F\x38\x00\x80\xFF\x2B\x00\x80\x79\x2F"
-     "\x00\x80"
-     "\xB0\x30\x00\x80\x67\x34\x00\x80\xE7\x34\x00\x80\x82\x2D\x00\x80\x73\x35"
-     "\x00\x80"
-     "\x6E\x2C\x00\x80\x94\x30\x00\x80\xDF\x2B\x00\x80\x23\x33\x00\x80\xC0\x31"
-     "\x00\x80"
-     "\x47\x36\x00\x80\xD6\x2D\x00\x80\xC2\x2E\x00\x80\x33\x35\x00\x80\x9B\x37"
-     "\x00\x80"
-     "\x73\x38\x00\x80\xEF\x32\x00\x80\x1B\x33\x00\x80\x54\x31\x00\x80\x53\x37"
-     "\x00\x80"
-     "\x74\x31\x00\x80\xE3\x34\x00\x80\x33\x2C\x00\x80\xB6\x2C\x00\x80\xFA\x2C"
-     "\x00\x80"
-     "\x8B\x32\x00\x80\xBA\xFD\xEA\x00\x33\x36\x00\x80\x2B\x37\x00\x80\x7F\x38"
-     "\x00\x80"
-     "\x93\x39\x00\x80\x6A\x2D\x00\x80\x1E\x2E\x00\x80\x4C\x30\x00\x80\xDC\x31"
-     "\x00\x80"
-     "\xB3\x33\x00\x80\x77\x35\x00\x80\xCF\x39\x00\x80\xB3\x3A\x00\x80\x07\x3B"
-     "\x00\x80"
-     "\x1F\x3B\x00\x80\x1E\x32\x00\x80\xDB\x32\x00\x80\x23\x39\x00\x80\x18\x30"
-     "\x00\x80"
-     "\xF4\x30\x00\x80\x10\x31\x00\x80\xA0\x31\x00\x80\xA3\x33\x00\x80\xEF\x33"
-     "\x00\x80"
-     "\x9F\x34\x00\x80\xC3\x35\x00\x80\x67\x36\x00\x80\xCB\x36\x00\x80\x13\x39"
-     "\x00\x80"
-     "\xA3\x37\x00\x80\x3B\x39\x00\x80\x80\x9D\x01\x00\x81\xE5\x3D\x00\x82\xA9"
-     "\x01\x00"
-     "\x83\xA9\x01\x00\x84\xB9\x01\x00\x85\xB9\x01\x00\x86\xA9\x01\x00\x87\xA9"
-     "\x01\x00"
-     "\x88\x99\x01\x00\x89\x99\x01\x00\x8A\xAD\x3D\x00\x8B\xED\x01\x00\x8C\xF5"
-     "\x01\x00"
-     "\x8D\xFD\x01\x00\x8E\xF5\x01\x00\x8F\xED\x01\x00\x90\x95\x01\x00\x91\x9D"
-     "\x01\x00"
-     "\x92\x95\x01\x00\x93\xAD\x01\x00\x94\xB5\x01\x00\x95\xBD\x01\x00\x96\xB5"
-     "\x01\x00"
-     "\x97\xAD\x01\x00\x98\x95\x01\x00\x99\x9D\x01\x00\x9A\x95\x01\x00\x9B\x69"
-     "\x02\x00"
-     "\x9C\x79\x02\x00\x9D\x79\x02\x00\x9E\x69\x02\x00\x9F\x69\x02\x00\xA0\x99"
-     "\x02\x00"
-     "\xA1\xDD\x01\x00\xA2\x95\x02\x00\xA3\xAD\x02\x00\xA4\xB5\x02\x00\xA5\xBD"
-     "\x02\x00"
-     "\xA6\xB5\x02\x00\xA7\xAD\x02\x00\xA8\x95\x02\x00\xA9\x9D\x02\x00\xAA\xB1"
-     "\x01\x00"
-     "\xAB\x91\x02\x00\xAC\xF1\x02\x00\xAD\xF1\x02\x00\xAE\xF1\x02\x00\xAF\xF1"
-     "\x02\x00"
-     "\xB0\x91\x02\x00\xB1\x91\x02\x00\xB2\x91\x02\x00\xB3\x91\x02\x00\xB4\xB1"
-     "\x02\x00"
-     "\xB5\xB1\x02\x00\xB6\xB1\x02\x00\xB7\xB1\x02\x00\xB8\x91\x02\x00\xB9\x91"
-     "\x02\x00"
-     "\xBA\x91\x02\x00\xBB\x91\x02\x00\xBC\x71\x02\x00\xBD\x71\x02\x00\xBE\x71"
-     "\x02\x00"
-     "\xBF\x69\x05\x00\xA1\x21\x02\x00\xA0\x19\x02\x00\xA3\x21\x02\x00\xA2\x29"
-     "\x02\x00"
-     "\xA5\x21\x02\x00\xA4\x39\x02\x00\xA7\x21\x02\x00\xA6\x29\x02\x00\xA9\x61"
-     "\x02\x00"
-     "\xA8\x19\x02\x00\xAB\x61\x02\x00\xAA\x69\x02\x00\xAD\x61\x02\x00\xAC\x79"
-     "\x02\x00"
-     "\xAF\x61\x02\x00\xAE\x69\x02\x00\xB1\x15\x3B\x00\xB0\x19\x02\x00\xB3\x2D"
-     "\x02\x00"
-     "\xB2\x2D\x02\x00\xB5\x3D\x02\x00\xB4\x3D\x02\x00\xB7\x2D\x02\x00\xB6\x2D"
-     "\x02\x00"
-     "\xB9\x1D\x02\x00\xB8\x1D\x02\x00\xBB\xE1\x02\x00\xBA\xED\x02\x00\xBD\xE1"
-     "\x02\x00"
-     "\xBC\xF9\x02\x00\xBF\xE1\x02\x00\xBE\xE9\x02\x00\x81\x21\x05\x00\x80\x69"
-     "\x3B\x00"
-     "\x83\x21\x05\x00\x82\x29\x05\x00\x85\x21\x05\x00\x84\x39\x05\x00\x87\x21"
-     "\x05\x00"
-     "\x86\x29\x05\x00\x89\x61\x05\x00\x88\x19\x05\x00\x8B\x61\x05\x00\x8A\x69"
-     "\x05\x00"
-     "\x8D\x61\x05\x00\x8C\x79\x05\x00\x8F\x61\x05\x00\x8E\x69\x05\x00\x91\x25"
-     "\x05\x00"
-     "\x90\x25\x05\x00\x93\x25\x05\x00\x92\x25\x05\x00\x95\x25\x05\x00\x94\x25"
-     "\x05\x00"
-     "\x97\x25\x05\x00\x96\x25\x05\x00\x99\xE5\x02\x00\x98\xE5\x02\x00\x9B\xE5"
-     "\x02\x00"
-     "\x9A\xE5\x02\x00\x9D\xE1\x02\x00\x9C\x61\x06\x00\x9F\xE1\x02\x00\x9E\xE9"
-     "\x02\x00"
-     "\x91\x84\x0D\x00\xB7\x39\x00\x80\x1E\x2D\x00\x80\x4A\x2D\x00\x80\x32\x2E"
-     "\x00\x80"
-     "\xDF\x33\x00\x80\x2B\x38\x00\x80\x38\x30\x00\x80\xA8\x30\x00\x80\x6C\x31"
-     "\x00\x80"
-     "\xE7\x32\x00\x80\x0F\x36\x00\x80\x5B\x34\x00\x80\x43\x2C\x00\x80\x8E\x2E"
-     "\x00\x80"
-     "\xDD\x2F\x00\x80\xB4\x30\x00\x80\x77\x34\x00\x80\x93\x34\x00\x80\xA3\x36"
-     "\x00\x80"
-     "\xC3\x36\x00\x80\x9B\x39\x00\x80\x1B\x3A\x00\x80\x2B\x3A\x00\x80\x67\x3A"
-     "\x00\x80"
-     "\x62\x2C\x00\x80\xB7\x35\x00\x80\x87\x39\x00\x80\x03\x3A\x00\x80\xCC\x30"
-     "\x00\x80"
-     "\x1F\x2C\x00\x80\x86\x2C\x00\x80\x7D\x2F\x00\x80\x9D\x2F\x00\x80\xCC\x31"
-     "\x00\x80"
-     "\xFB\x33\x00\x80\x03\x35\x00\x80\xAF\x37\x00\x80\x4F\x39\x00\x80\x0E\x32"
-     "\x00\x80"
-     "\xDB\x39\x00\x80\x42\x2D\x00\x80\x5A\x32\x00\x80\x66\x32\x00\x80\x47\x33"
-     "\x00\x80"
-     "\x7F\x33\x00\x80\x83\x34\x00\x80\xDF\x34\x00\x80\x5F\x35\x00\x80\x43\x36"
-     "\x00\x80"
-     "\x73\x3A\x00\x80\x20\x31\x00\x80\xEB\x39\x00\x80\x76\x2C\x00\x80\xC1\x2F"
-     "\x00\x80"
-     "\x67\x33\x00\x80\x17\x39\x00\x80\x84\x30\x00\x80\xD4\x30\x00\x80\x6E\x32"
-     "\x00\x80"
-     "\xF3\x39\x00\x80\x26\x2D\x00\x80\x0E\x2E\x00\x80\xAD\x2F\x00\x80\xF7\x31"
-     "\x00\x80"
-     "\x4A\x32\x00\x80\x97\x32\x00\x80\x2B\x33\x00\x80\x7F\x34\x00\x80\xF3\x34"
-     "\x00\x80"
-     "\x8F\x36\x00\x80\x2F\x38\x00\x80\x3B\x38\x00\x80\x0F\x3A\x00\x80\x92\x2D"
-     "\x00\x80"
-     "\x83\x33\x00\x80\x12\x2E\x00\x80\xFF\x33\x00\x80\xA3\x34\x00\x80\xBF\x37"
-     "\x00\x80"
-     "\xF7\x39\x00\x80\xE3\x3A\x00\x80\x0B\x3B\x00\x80\x62\x32\x00\x80\x5F\x33"
-     "\x00\x80"
-     "\xFB\x36\x00\x80\xFB\x35\x00\x80\x1F\x36\x00\x80\x4F\x34\x00\x80\xD7\x33"
-     "\x00\x80"
-     "\x9B\x38\x00\x80\x3B\x2C\x00\x80\x63\x37\x00\x80\x2A\x2D\x00\x80\x1A\x2D"
-     "\x00\x80"
-     "\x28\x30\x00\x80\x48\x31\x00\x80\x2B\x36\x00\x80\x69\x2F\x00\x80\x82\x64"
-     "\x01\x00"
-     "\x3B\x33\x00\x80\x1A\x32\x00\x80\x1F\x39\x00\x80\xE3\x39\x00\x80\x44\x30"
-     "\x00\x80"
-     "\x96\x2C\x00\x80\x6A\x2E\x00\x80\xBE\x2E\x00\x80\x0A\x32\x00\x80\x06\x2D"
-     "\x00\x80"
-     "\x57\x34\x00\x80\x1B\x35\x00\x80\x7F\x35\x00\x80\x9B\x35\x00\x80\x9F\x35"
-     "\x00\x80"
-     "\xB3\x35\x00\x80\x33\x3A\x00\x80\x23\x36\x00\x80\xC7\x37\x00\x80\x8B\x38"
-     "\x00\x80"
-     "\x3F\x39\x00\x80\x6B\x39\x00\x80\x87\x3A\x00\x80\x8B\x3A\x00\x80\x8F\x3A"
-     "\x00\x80"
-     "\xF3\x3A\x00\x80\x63\x39\x00\x80\xFF\x39\x00\x80\x80\x79\x01\x00\x81\x79"
-     "\x01\x00"
-     "\x82\x89\x01\x00\x83\x89\x01\x00\x84\x99\x01\x00\x85\x85\x01\x00\x86\x8D"
-     "\x01\x00"
-     "\x87\x85\x01\x00\x88\x31\x34\x00\x89\xB9\x01\x00\x8A\x35\x34\x00\x8B\x8D"
-     "\x01\x00"
-     "\x8C\x95\x01\x00\x8D\x9D\x01\x00\x66\x2C\x00\x80\x8A\x2C\x00\x80\x90\xED"
-     "\x01\x00"
-     "\x9E\x2C\x00\x80\x92\xF9\x01\x00\x5A\x2D\x00\x80\x66\x2D\x00\x80\x95\xE1"
-     "\x01\x00"
-     "\x96\xE1\x01\x00\x97\xE1\x01\x00\x98\xA1\x01\x00\x99\xA1\x01\x00\x9A\xA1"
-     "\x01\x00"
-     "\x9B\xA1\x01\x00\x9C\xA1\x01\x00\x9D\xA1\x01\x00\x9E\xF9\x32\x00\xA6\x2D"
-     "\x00\x80"
-     "\xA0\x59\x01\x00\x5A\x2E\x00\x80\xA2\x55\x01\x00\x76\x2E\x00\x80\x7E\x2E"
-     "\x00\x80"
-     "\xA5\x75\x01\x00\xA6\x7D\x01\x00\xBA\x2E\x00\x80\xC6\x2E\x00\x80\xA9\x2F"
-     "\x00\x80"
-     "\xAA\x41\x01\x00\xAB\x41\x01\x00\xAC\x41\x01\x00\xAD\x41\x01\x00\xAE\x41"
-     "\x01\x00"
-     "\xAF\x41\x01\x00\xB0\xF9\x00\x00\xB1\xF9\x00\x00\xB2\x8D\x00\x00\xB3\x81"
-     "\x00\x00"
-     "\xB4\x81\x00\x00\xB5\xA9\x00\x00\xB6\x5D\x00\x00\xB7\x51\x00\x00\xB8\x71"
-     "\x00\x00"
-     "\xB9\x79\x00\x00\xBA\x49\x00\x00\xBB\x49\x00\x00\xBC\x0D\x31\x00\xBD\xF1"
-     "\x02\x00"
-     "\xBE\xF1\x02\x00\xBF\xF1\x02\x00\xA3\x0D\x02\x00\xA2\x0D\x02\x00\xA1\x0D"
-     "\x02\x00"
-     "\xA0\x0D\x02\x00\xA7\x0D\x02\x00\xA6\x0D\x02\x00\xA5\x0D\x02\x00\xA4\x0D"
-     "\x02\x00"
-     "\xAB\x09\x02\x00\xAA\x01\x02\x00\xA9\x41\x37\x00\xA8\x0D\x02\x00\xAF\x09"
-     "\x02\x00"
-     "\xAE\x01\x02\x00\xAD\x09\x02\x00\xAC\x11\x02\x00\xB3\x09\x02\x00\xB2\x01"
-     "\x02\x00"
-     "\xB1\x09\x02\x00\xB0\x71\x02\x00\xB7\x19\x31\x00\xB6\x01\x02\x00\xB5\x09"
-     "\x02\x00"
-     "\xB4\x11\x02\x00\xBB\x05\x02\x00\xBA\x05\x02\x00\xB9\x35\x02\x00\xB8\x35"
-     "\x02\x00"
-     "\xBF\x05\x02\x00\xBE\x05\x02\x00\xBD\x15\x02\x00\xBC\x15\x02\x00\x83\x19"
-     "\x0D\x00"
-     "\x82\x11\x0D\x00\x81\x19\x0D\x00\x80\x01\x0D\x00\x87\x35\x0D\x00\x86\x35"
-     "\x0D\x00"
-     "\x85\x05\x0D\x00\x84\xE1\x34\x00\x8B\x15\x0D\x00\x8A\x15\x0D\x00\x89\x05"
-     "\x0D\x00"
-     "\x88\x05\x0D\x00\x8F\x75\x0D\x00\x8E\x75\x0D\x00\x8D\x05\x0D\x00\x8C\x05"
-     "\x0D\x00"
-     "\x93\x11\x0D\x00\x92\x09\x0D\x00\x91\xE1\x24\x00\x90\x05\x0D\x00\x97\x31"
-     "\x0D\x00"
-     "\x96\x09\x0D\x00\x95\x01\x0D\x00\x94\x09\x0D\x00\x9B\xC1\x0E\x00\x9A\x09"
-     "\x0D\x00"
-     "\x99\x01\x0D\x00\x98\x09\x0D\x00\x9F\x0D\x0D\x00\x9E\x0D\x0D\x00\x9D\x0D"
-     "\x0D\x00"
-     "\x9C\x0D\x0D\x00\xBA\x81\xE5\x00\xC0\x30\x00\x80\xE8\x30\x00\x80\xF0\x30"
-     "\x00\x80"
-     "\x08\x31\x00\x80\xB4\x31\x00\x80\xE4\x31\x00\x80\x12\x32\x00\x80\x8F\x32"
-     "\x00\x80"
-     "\x53\x33\x00\x80\x6F\x33\x00\x80\x9F\x33\x00\x80\xF3\x33\x00\x80\x13\x34"
-     "\x00\x80"
-     "\x87\x34\x00\x80\x67\x35\x00\x80\x8B\x35\x00\x80\x87\x35\x00\x80\x8F\x35"
-     "\x00\x80"
-     "\x93\x35\x00\x80\xAB\x35\x00\x80\xAF\x35\x00\x80\xD3\x35\x00\x80\xE7\x35"
-     "\x00\x80"
-     "\x03\x36\x00\x80\x6F\x36\x00\x80\x77\x36\x00\x80\x8B\x36\x00\x80\xB3\x36"
-     "\x00\x80"
-     "\x03\x37\x00\x80\x37\x37\x00\x80\x9F\x37\x00\x80\x47\x38\x00\x80\x5F\x38"
-     "\x00\x80"
-     "\x93\x38\x00\x80\x97\x38\x00\x80\xDB\x38\x00\x80\xDF\x38\x00\x80\x37\x39"
-     "\x00\x80"
-     "\x13\x3A\x00\x80\x57\x3A\x00\x80\x6F\x3A\x00\x80\xB8\x30\x00\x80\xC0\xF2"
-     "\x00\x80"
-     "\x1B\x37\x00\x80\xE7\x2B\x00\x80\xF2\x2C\x00\x80\xB2\x2C\x00\x80\x5A\x2C"
-     "\x00\x80"
-     "\x9A\x2C\x00\x80\xC6\x2C\x00\x80\x56\x2D\x00\x80\x72\x2D\x00\x80\x46\x2E"
-     "\x00\x80"
-     "\x56\x2E\x00\x80\x72\x2E\x00\x80\xCE\x2E\x00\x80\x19\x2F\x00\x80\x25\x2F"
-     "\x00\x80"
-     "\x41\x2F\x00\x80\x51\x2F\x00\x80\x40\x30\x00\x80\x48\x30\x00\x80\x78\x30"
-     "\x00\x80"
-     "\x80\xED\x07\x00\x81\xF5\x07\x00\x82\xFD\x07\x00\x83\xF5\x07\x00\x84\xED"
-     "\x07\x00"
-     "\x85\x15\x01\x00\x86\x1D\x01\x00\x87\x15\x01\x00\x88\x2D\x01\x00\x89\x29"
-     "\x07\x00"
-     "\x8A\x39\x01\x00\x8B\x3D\x07\x00\x8C\x25\x01\x00\x8D\x2D\x01\x00\x8E\x25"
-     "\x01\x00"
-     "\x8F\x1D\x01\x00\x90\x65\x01\x00\x91\xF5\x04\x00\x92\x01\x0F\x00\x93\x65"
-     "\x01\x00"
-     "\x94\x7D\x01\x00\x95\xD1\x36\x00\x96\x05\x08\x00\x97\xED\x0B\x00\x98\x51"
-     "\x01\x00"
-     "\x99\x51\x01\x00\x9A\x45\x07\x00\x9B\x55\x01\x00\x9C\x41\x07\x00\x9D\x49"
-     "\x01\x00"
-     "\x9E\x39\x01\x00\x9F\x39\x01\x00\xA0\x39\x04\x00\xA1\xCD\x01\x00\xA2\xC5"
-     "\x01\x00"
-     "\xA3\xDD\x01\x00\xA4\xC5\x01\x00\xA5\xCD\x01\x00\xA6\x25\x04\x00\xA7\xC1"
-     "\x01\x00"
-     "\xA8\xC1\x01\x00\xA9\xC1\x01\x00\xAA\xC1\x01\x00\xAB\xC1\x01\x00\xAC\xC1"
-     "\x01\x00"
-     "\xAD\xD5\x07\x00\xAE\xCD\x01\x00\xAF\xC5\x01\x00\xB0\x2D\x08\x00\xB1\xB9"
-     "\x01\x00"
-     "\xB2\xB9\x07\x00\xB3\x4D\x01\x00\xB4\x55\x01\x00\xB5\x5D\x01\x00\xB6\x55"
-     "\x01\x00"
-     "\xB7\x4D\x01\x00\xB8\x65\x07\x00\xB9\x71\x01\x00\xBA\x35\x04\x00\xBB\x75"
-     "\x01\x00"
-     "\xBC\x79\x07\x00\xBD\xE5\x08\x00\xBE\x65\x01\x00\xBF\x71\x07\x00\x98\x30"
-     "\x00\x80"
-     "\xC8\x30\x00\x80\xE0\x30\x00\x80\xD0\x30\x00\x80\xE4\x30\x00\x80\x24\x31"
-     "\x00\x80"
-     "\x7C\x31\x00\x80\x88\x31\x00\x80\x90\x31\x00\x80\xB8\x31\x00\x80\x3E\x32"
-     "\x00\x80"
-     "\x52\x32\x00\x80\x93\x33\x00\x80\x8B\x33\x00\x80\xBB\x33\x00\x80\x4B\x35"
-     "\x00\x80"
-     "\x17\x34\x00\x80\x4B\x34\x00\x80\x97\x34\x00\x80\xB3\x34\x00\x80\xD7\x34"
-     "\x00\x80"
-     "\xFB\x34\x00\x80\xFF\x34\x00\x80\x1F\x35\x00\x80\x2F\x35\x00\x80\x43\x35"
-     "\x00\x80"
-     "\x3F\x35\x00\x80\x6F\x35\x00\x80\xEB\x35\x00\x80\x1B\x36\x00\x80\x53\x36"
-     "\x00\x80"
-     "\x83\x36\x00\x80\x67\x37\x00\x80\x8F\x37\x00\x80\x03\x38\x00\x80\x4B\x38"
-     "\x00\x80"
-     "\x57\x38\x00\x80\x77\x38\x00\x80\x7B\x38\x00\x80\x87\x38\x00\x80\xF0\xE0"
-     "\x10\x00"
-     "\xA3\x38\x00\x80\x1B\x39\x00\x80\x4B\x39\x00\x80\x83\x39\x00\x80\xA3\x39"
-     "\x00\x80"
-     "\xEF\x39\x00\x80\x47\x3A\x00\x80\x63\x3A\x00\x80\xC3\x3A\x00\x80\x57\xF2"
-     "\x00\x80"
-     "\x52\xF2\x00\x80\x84\xF2\x00\x80\x0F\x2B\x00\x80\x37\x2B\x00\x80\x3B\x2B"
-     "\x00\x80"
-     "\x06\xF3\x00\x80\x33\xF3\x00\x80\xAB\xF3\x00\x80\x4F\x3B\x00\x80\x5F\x3B"
-     "\x00\x80"
-     "\xE0\x06\x00\x80\xEB\x06\x00\x80\xEE\x06\x00\x80\xAD\xB1\x05\x00\xAC\xA9"
-     "\x05\x00"
-     "\xE3\x06\x00\x80\xE7\x06\x00\x80\xA9\xB1\x05\x00\xA8\x89\x05\x00\xAB\xB1"
-     "\x05\x00"
-     "\xAA\xB9\x05\x00\xA5\xB5\x05\x00\xA4\xB5\x05\x00\xA7\xC9\x06\x00\xA6\xB5"
-     "\x05\x00"
-     "\xA1\xB5\x05\x00\xA0\xB5\x05\x00\xA3\xB5\x05\x00\xA2\xB5\x05\x00\xBD\xA1"
-     "\x05\x00"
-     "\xBC\xE9\x06\x00\xBF\xA1\x05\x00\xBE\xA9\x05\x00\xB9\x85\x05\x00\xB8\xD1"
-     "\x07\x00"
-     "\xBB\x85\x05\x00\xBA\x85\x05\x00\xB5\xA9\x05\x00\xB4\xA1\x05\x00\xB7\xB9"
-     "\x05\x00"
-     "\xB6\xA1\x05\x00\xB1\xC9\x05\x00\xB0\xC1\x05\x00\xB3\xB9\x05\x00\xB2\xC1"
-     "\x05\x00"
-     "\x8D\xC1\x05\x00\x8C\xB9\x20\x00\x8F\xC1\x05\x00\x8E\xC9\x05\x00\x89\xA5"
-     "\x05\x00"
-     "\x88\xA5\x05\x00\x8B\xA5\x05\x00\x8A\xA5\x05\x00\x85\x85\x05\x00\x84\x85"
-     "\x05\x00"
-     "\x87\x85\x05\x00\x86\x85\x05\x00\x81\xA5\x05\x00\x80\xA5\x05\x00\x83\xA5"
-     "\x05\x00"
-     "\x82\xA5\x05\x00\x9D\xB9\x05\x00\x9C\xB1\x05\x00\x9F\xB5\x05\x00\x9E\xB5"
-     "\x05\x00"
-     "\x99\xB9\x05\x00\x98\xB1\x05\x00\x9B\xA9\x05\x00\x9A\xB1\x05\x00\x95\xBD"
-     "\x05\x00"
-     "\x94\xBD\x05\x00\x97\x05\x0A\x00\x96\x8D\x05\x00\x91\xBD\x20\x00\x90\xB9"
-     "\x05\x00"
-     "\x93\xAD\x05\x00\x92\xAD\x05\x00\x4D\x08\x00\x80\xA5\x0E\x00\x80\x96\x0E"
-     "\x00\x80"
-     "\x9B\x0E\x00\x80\xC8\x0E\x00\x80\xA0\x0E\x00\x80\x35\x0F\x00\x80\xB4\x0F"
-     "\x00\x80"
-     "\x6B\x0F\x00\x80\x19\x0F\x00\x80\x44\x0F\x00\x80\x51\x0F\x00\x80\x59\x0F"
-     "\x00\x80"
-     "\x87\x0F\x00\x80\xAC\x0F\x00\x80\xA2\x0F\x00\x80\xA7\x0F\x00\x80\x94\x0F"
-     "\x00\x80"
-     "\x9B\x0F\x00\x80\xE8\x0E\x00\x80\xED\x0E\x00\x80\xF2\x0E\x00\x80\xFF\x0E"
-     "\x00\x80"
-     "\x0C\x0F\x00\x80\x14\x0F\x00\x80\x1C\x0F\x00\x80\x26\x0F\x00\x80\x2B\x0F"
-     "\x00\x80"
-     "\x30\x0F\x00\x80\x3A\x0F\x00\x80\x3F\x0F\x00\x80\x47\x0F\x00\x80\x88\x51"
-     "\x07\x00"
-     "\x89\x39\x02\x00\x8A\x29\x04\x00\x8B\x4D\x02\x00\x8C\x35\x04\x00\x8D\x51"
-     "\x02\x00"
-     "\x8E\x19\x31\x00\x8F\x55\x02\x00\x80\x05\x02\x00\x81\x31\x04\x00\x82\x01"
-     "\x02\x00"
-     "\x83\x01\x02\x00\x84\x01\x02\x00\x85\x01\x02\x00\x86\x01\x02\x00\x87\x21"
-     "\x04\x00"
-     "\x98\x2D\x02\x00\x99\x35\x02\x00\x54\x0F\x00\x80\x5C\x0F\x00\x80\x61\x0F"
-     "\x00\x80"
-     "\x66\x0F\x00\x80\x6E\x0F\x00\x80\x73\x0F\x00\x80\x90\x2D\x02\x00\x91\x35"
-     "\x02\x00"
-     "\x92\x3D\x02\x00\x93\x35\x02\x00\x94\x2D\x02\x00\x95\x15\x02\x00\x96\x1D"
-     "\x02\x00"
-     "\x97\x15\x02\x00\x80\xD1\xFF\x00\x81\xD1\xFF\x00\x7D\x0F\x00\x80\x83\xD5"
-     "\xFF\x00"
-     "\x84\xCD\xFF\x00\x82\x0F\x00\x80\x86\x11\x00\x00\x87\x09\x00\x00\x88\x21"
-     "\x01\x00"
-     "\x89\x21\x01\x00\x8A\x21\x01\x00\x8B\x21\x01\x00\x8C\x21\x01\x00\x8D\x21"
-     "\x01\x00"
-     "\x8E\x21\x01\x00\x8F\x21\x01\x00\x90\x61\x01\x00\x91\x65\x01\x00\x92\x6D"
-     "\x01\x00"
-     "\x93\x69\x01\x00\x94\x75\x01\x00\x95\x71\x01\x00\x96\x71\x01\x00\x97\x75"
-     "\x01\x00"
-     "\x98\x49\x01\x00\x99\x4D\x01\x00\x9A\x45\x01\x00\x9B\x41\x01\x00\x9C\x5D"
-     "\x01\x00"
-     "\x9D\x59\x01\x00\x9E\x49\x01\x00\x9F\x4D\x01\x00\xA0\xB1\x01\x00\xA1\xB5"
-     "\x01\x00"
-     "\xA2\xBD\x01\x00\xA3\xB9\x01\x00\xA4\xA5\x01\x00\xA5\xA1\x01\x00\xA6\xA1"
-     "\x01\x00"
-     "\xA7\xA5\x01\x00\xA8\x99\x01\x00\xA9\x9D\x01\x00\xAA\x95\x01\x00\xAB\x91"
-     "\x01\x00"
-     "\xAC\x8D\x01\x00\xAD\x89\x01\x00\xAE\xF9\x01\x00\xAF\xFD\x01\x00\xB0\x81"
-     "\x01\x00"
-     "\xB1\x85\x01\x00\xB2\x8D\x01\x00\xB3\x89\x01\x00\xB4\x95\x01\x00\xB5\x91"
-     "\x01\x00"
-     "\xB6\x91\x01\x00\xB7\x95\x01\x00\xB8\xA9\x01\x00\xB9\xAD\x01\x00\xBA\xA5"
-     "\x01\x00"
-     "\xBB\xA1\x01\x00\xBC\xBD\x01\x00\xBD\xB9\x01\x00\xBE\xA9\x01\x00\xBF\xAD"
-     "\x01\x00"
-     "\x8A\x0F\x00\x80\x8F\x0F\x00\x80\xAF\x0F\x00\x80\x21\x0F\x00\x80\x04\x0F"
-     "\x00\x80"
-     "\x4C\x0F\x00\x80\x78\x0F\x00\x80\xF7\x0E\x00\x80\xC7\x16\x00\x80\xD0\x16"
-     "\x00\x80"
-     "\xD3\x16\x00\x80\xD9\x16\x00\x80\xCD\x16\x00\x80\xD6\x16\x00\x80\xCA\x16"
-     "\x00\x80"
-     "\xFD\x16\x00\x80\x00\x17\x00\x80\xDF\x16\x00\x80\xDC\x16\x00\x80\xE2\x16"
-     "\x00\x80"
-     "\xE5\x16\x00\x80\x91\x10\x06\x00\xEE\x16\x00\x80\xEB\x16\x00\x80\xF1\x16"
-     "\x00\x80"
-     "\xE8\x16\x00\x80\xF7\x16\x00\x80\xF4\x16\x00\x80\x03\x17\x00\x80\x09\x17"
-     "\x00\x80"
-     "\x0F\x17\x00\x80\x0C\x17\x00\x80\x12\x17\x00\x80\x15\x17\x00\x80\x18\x17"
-     "\x00\x80"
-     "\x1E\x17\x00\x80\x06\x17\x00\x80\x2A\x17\x00\x80\x27\x17\x00\x80\x32\x17"
-     "\x00\x80"
-     "\x41\x17\x00\x80\x38\x17\x00\x80\x24\x17\x00\x80\x35\x17\x00\x80\x3E\x17"
-     "\x00\x80"
-     "\x63\x16\x00\x80\xCB\x0F\x00\x80\x1B\x10\x00\x80\xF8\x0F\x00\x80\x0C\x10"
-     "\x00\x80"
-     "\x07\x10\x00\x80\x11\x10\x00\x80\x16\x10\x00\x80\xFD\x0F\x00\x80\x3B\x17"
-     "\x00\x80"
-     "\xD0\x0F\x00\x80\xD5\x0F\x00\x80\xE9\x0F\x00\x80\x02\x10\x00\x80\x42\x10"
-     "\x00\x80"
-     "\x47\x10\x00\x80\x53\x10\x00\x80\x69\x10\x00\x80\x78\x10\x00\x80\xA8\x39"
-     "\xFD\x00"
-     "\xA9\x5D\xBB\x00\xAA\x35\xFD\x00\xAB\x4D\xFD\x00\xAC\x55\xFD\x00\xAD\x5D"
-     "\xFD\x00"
-     "\xAE\x55\xFD\x00\xAF\x4D\xFD\x00\xA0\x01\xFD\x00\xA1\xE5\xBB\x00\xA2\xF1"
-     "\xBB\x00"
-     "\xA3\x09\xFD\x00\xA4\x19\xFD\x00\xA5\x19\xFD\x00\xA6\x09\xFD\x00\xA7\x09"
-     "\xFD\x00"
-     "\xB8\x31\xFD\x00\xB9\x31\xFD\x00\xBA\x31\xFD\x00\xBB\x31\xFD\x00\xBC\x99"
-     "\xFD\x00"
-     "\x7D\x10\x00\x80\xBE\x95\xFD\x00\x88\x10\x00\x80\xB0\x35\xFD\x00\xB1\x3D"
-     "\xFD\x00"
-     "\xB2\x35\xFD\x00\xB3\x0D\xFD\x00\xB4\x15\xFD\x00\xB5\x1D\xFD\x00\xB6\x15"
-     "\xFD\x00"
-     "\xA2\x10\x00\x80\xB5\x10\x00\x80\xD9\x10\x00\x80\x0B\x11\x00\x80\x10\x11"
-     "\x00\x80"
-     "\x18\x11\x00\x80\x27\x11\x00\x80\x36\x11\x00\x80\x3B\x11\x00\x80\x80\x55"
-     "\xFF\x00"
-     "\x81\x5D\xFF\x00\x82\x55\xFF\x00\x83\xA5\xFC\x00\x84\xBD\xFC\x00\x85\xB5"
-     "\xFD\x00"
-     "\x86\xB9\xFD\x00\x43\x11\x00\x80\x66\x11\x00\x80\xD9\xC0\xA3\x02\x8D\x11"
-     "\x00\x80"
-     "\x99\x11\x00\x80\xB9\x11\x00\x80\x9D\xCD\xFD\x00\xBE\x11\x00\x80\x9F\xC1"
-     "\xFD\x00"
-     "\xC3\x11\x00\x80\xFD\x11\x00\x80\xD9\xFC\xA3\x02\x93\xE9\xFD\x00\x94\xF9"
-     "\xFD\x00"
-     "\x95\xF9\xFD\x00\x96\xE9\xFD\x00\x97\xE9\xFD\x00\x95\x75\x02\x00\x94\x71"
-     "\x02\x00"
-     "\x97\x71\x02\x00\x96\x75\x02\x00\x91\x61\x02\x00\x90\x65\x02\x00\x93\x6D"
-     "\x02\x00"
-     "\x92\x69\x02\x00\x9D\x5D\x02\x00\x9C\x59\x02\x00\x9F\xA9\x02\x00\x9E\xAD"
-     "\x02\x00"
-     "\x99\x49\x02\x00\x98\x4D\x02\x00\x9B\x45\x02\x00\x9A\x41\x02\x00\x85\x1D"
-     "\x02\x00"
-     "\x84\x19\x02\x00\x87\x29\x02\x00\x86\x2D\x02\x00\x81\x01\x02\x00\x80\x05"
-     "\x02\x00"
-     "\x83\x01\x02\x00\x82\x05\x02\x00\x8D\x0D\x02\x00\x8C\x09\x02\x00\x8F\x19"
-     "\x02\x00"
-     "\x8E\x1D\x02\x00\x89\x15\x02\x00\x88\x11\x02\x00\x8B\x11\x02\x00\x8A\x15"
-     "\x02\x00"
-     "\x10\x12\x00\x80\x1C\x12\x00\x80\x34\x12\x00\x80\xAE\x12\x00\x80\xB1\x29"
-     "\xA1\x02"
-     "\xB0\x2D\xA1\x02\x04\x13\x00\x80\x1D\x13\x00\x80\x22\x13\x00\x80\x35\x13"
-     "\x00\x80"
-     "\x46\x13\x00\x80\x58\x13\x00\x80\x5D\x13\x00\x80\x84\x13\x00\x80\x8C\x13"
-     "\x00\x80"
-     "\xA1\x13\x00\x80\xA5\xCD\xA0\x02\xA4\xC9\xA0\x02\xA7\x45\x02\x00\xA6\x41"
-     "\x02\x00"
-     "\xA1\x55\x02\x00\xA0\x51\x02\x00\xA3\x5D\x02\x00\xA2\x59\x02\x00\xAD\x69"
-     "\x02\x00"
-     "\xAC\x6D\x02\x00\xAF\x49\x02\x00\xAE\x4D\x02\x00\xA9\x7D\x02\x00\xA8\x79"
-     "\x02\x00"
-     "\xAB\x71\x02\x00\xAA\x75\x02\x00\x80\xDD\x05\x00\x81\xE5\x05\x00\x82\xED"
-     "\x05\x00"
-     "\x83\xD9\x05\x00\x84\xF9\x05\x00\x85\xF9\x05\x00\x86\xE9\x05\x00\x87\xE9"
-     "\x05\x00"
-     "\x88\xD9\x05\x00\x89\xD9\x05\x00\x8A\x7D\x06\x00\x8B\x71\x06\x00\x8C\x4D"
-     "\x06\x00"
-     "\x8D\xB5\x06\x00\x8E\xBD\x06\x00\x8F\xB5\x06\x00\x90\xCD\x06\x00\x91\xD5"
-     "\x06\x00"
-     "\x92\xDD\x06\x00\x93\xD5\x06\x00\x94\xCD\x06\x00\x95\xE9\x06\x00\x96\x99"
-     "\x06\x00"
-     "\x97\x95\x06\x00\x98\xAD\x06\x00\x99\xB5\x06\x00\x9A\xB1\x06\x00\x9B\x8D"
-     "\x06\x00"
-     "\x9C\x95\x06\x00\x9D\xF5\x07\x00\x9E\xFD\x07\x00\x9F\xF5\x07\x00\xA0\x0D"
-     "\x07\x00"
-     "\xA1\x1D\x07\x00\xA2\x15\x07\x00\xA3\x2D\x07\x00\xA4\x35\x07\x00\xA5\x3D"
-     "\x07\x00"
-     "\xA6\x35\x07\x00\xA7\x2D\x07\x00\xA8\x15\x07\x00\xA9\x1D\x07\x00\xAA\x15"
-     "\x07\x00"
-     "\xAB\xAD\x01\x00\xAC\xB5\x01\x00\xAD\xBD\x01\x00\xAE\xB5\x01\x00\xAF\xAD"
-     "\x01\x00"
-     "\xB0\xD5\x01\x00\xB1\xDD\x01\x00\xB2\xD5\x01\x00\xB3\xED\x01\x00\xB4\xF5"
-     "\x01\x00"
-     "\xB5\xFD\x01\x00\xB6\xF5\x01\x00\xB7\xED\x01\x00\xB8\xD5\x01\x00\xB9\xDD"
-     "\x01\x00"
-     "\xBA\xD5\x01\x00\xBB\xAD\x01\x00\xBC\xB5\x01\x00\xBD\xBD\x01\x00\xBE\xB5"
-     "\x01\x00"
-     "\xBF\xAD\x01\x00\xC8\x13\x00\x80\xCD\x13\x00\x80\x27\x14\x00\x80\x2C\x14"
-     "\x00\x80"
-     "\x31\x14\x00\x80\x3D\x14\x00\x80\x49\x14\x00\x80\x4E\x14\x00\x80\x56\x14"
-     "\x00\x80"
-     "\x62\x14\x00\x80\x7C\x14\x00\x80\x81\x14\x00\x80\x89\x14\x00\x80\x8E\x14"
-     "\x00\x80"
-     "\x93\x14\x00\x80\x98\x14\x00\x80\x9D\x14\x00\x80\xA2\x14\x00\x80\xB5\x14"
-     "\x00\x80"
-     "\xBA\x14\x00\x80\xD6\x14\x00\x80\x84\xB0\x1B\x00\xF0\x14\x00\x80\x0A\x15"
-     "\x00\x80"
-     "\x16\x15\x00\x80\x2E\x15\x00\x80\x33\x15\x00\x80\x40\x15\x00\x80\x61\x15"
-     "\x00\x80"
-     "\x84\x15\x00\x80\x9E\x15\x00\x80\xAA\x15\x00\x80\xAF\x15\x00\x80\xB7\x15"
-     "\x00\x80"
-     "\xD8\x15\x00\x80\xF2\x15\x00\x80\x01\x16\x00\x80\x1E\x16\x00\x80\x23\x16"
-     "\x00\x80"
-     "\x2B\x16\x00\x80\x30\x16\x00\x80\x43\x16\x00\x80\x48\x16\x00\x80\x6E\x16"
-     "\x00\x80"
-     "\x7A\x16\x00\x80\x86\x16\x00\x80\x95\x16\x00\x80\xB2\x16\x00\x80\xB7\x16"
-     "\x00\x80"
-     "\xD8\x11\x00\x80\xE9\x11\x00\x80\x66\x16\x00\x80\x80\x00\x00\x80\x8A\x00"
-     "\x00\x80"
-     "\x94\x00\x00\x80\x9E\x00\x00\x80\xA8\x00\x00\x80\xB2\x00\x00\x80\xDF\x0F"
-     "\x00\x80"
-     "\xE4\x0F\x00\x80\xEE\x0F\x00\x80\x5F\x10\x00\x80\x64\x10\x00\x80\x6E\x10"
-     "\x00\x80"
-     "\xAB\xB9\x07\x00\xAA\xBD\x07\x00\xA9\xBD\x07\x00\xA8\xB9\x07\x00\xAF\xD1"
-     "\x07\x00"
-     "\xAE\xD5\x07\x00\xAD\xA5\x07\x00\xAC\xA1\x07\x00\xA3\x8D\x07\x00\xA2\x89"
-     "\x07\x00"
-     "\xA1\x81\x07\x00\xA0\x85\x07\x00\xA7\x81\x07\x00\xA6\x85\x07\x00\xA5\x89"
-     "\x07\x00"
-     "\xA4\x8D\x07\x00\xBB\x95\x07\x00\xBA\x91\x07\x00\xB9\x9D\x07\x00\xB8\x9D"
-     "\x07\x00"
-     "\xBF\x89\x07\x00\xBE\x8D\x07\x00\xBD\x81\x07\x00\xBC\x85\x07\x00\xB3\xA9"
-     "\x07\x00"
-     "\xB2\xAD\x07\x00\xB1\xAD\x07\x00\xB0\xA9\x07\x00\xB7\xA1\x07\x00\xB6\xA5"
-     "\x07\x00"
-     "\xB5\xB5\x07\x00\xB4\xB1\x07\x00\xCF\x10\x00\x80\xD4\x10\x00\x80\x01\x11"
-     "\x00\x80"
-     "\x1D\x11\x00\x80\x22\x11\x00\x80\x2C\x11\x00\x80\x3B\x15\x00\x80\xF7\x15"
-     "\x00\x80"
-     "\xFC\x15\x00\x80\x14\x16\x00\x80\x8B\x16\x00\x80\x90\x16\x00\x80\xA8\x16"
-     "\x00\x80"
-     "\xDA\x0F\x00\x80\xF3\x0F\x00\x80\x73\x10\x00\x80\x9B\x5D\x07\x00\x9A\x5D"
-     "\x07\x00"
-     "\x99\x51\x07\x00\x98\x59\x07\x00\x9F\x75\x07\x00\x9E\x71\x07\x00\x9D\x79"
-     "\xA4\x02"
-     "\x9C\x41\x07\x00\x93\x71\x07\x00\x06\x11\x00\x80\xC1\x12\x00\x80\x29\x15"
-     "\x00\x80"
-     "\x97\x65\x07\x00\x96\x65\x07\x00\x95\x69\x07\x00\x94\x6D\x07\x00\x95\xAD"
-     "\x02\x00"
-     "\x94\xAD\x02\x00\x97\xBD\x02\x00\x96\xBD\x02\x00\x91\x8D\x02\x00\x90\x8D"
-     "\x02\x00"
-     "\x93\xBD\x02\x00\x92\xBD\x02\x00\x9D\xED\x02\x00\x9C\xED\x02\x00\x9F\xFD"
-     "\x02\x00"
-     "\x9E\xFD\x02\x00\x99\x8D\x02\x00\x98\x8D\x02\x00\x9B\xFD\x02\x00\x9A\xFD"
-     "\x02\x00"
-     "\x85\x2D\x02\x00\x84\x2D\x02\x00\x87\x3D\x02\x00\x86\x3D\x02\x00\x81\x0D"
-     "\x02\x00"
-     "\x80\x0D\x02\x00\x83\x3D\x02\x00\x82\x3D\x02\x00\x8D\xED\x02\x00\x8C\xED"
-     "\x02\x00"
-     "\x8F\xFD\x02\x00\x8E\xFD\x02\x00\x89\x0D\x02\x00\x88\x0D\x02\x00\x8B\xFD"
-     "\x02\x00"
-     "\x8A\xFD\x02\x00\xB5\x0D\x05\x00\xB4\x0D\x05\x00\xB7\x19\x01\x00\xB6\x01"
-     "\x01\x00"
-     "\xB1\x15\x01\x00\xB0\x15\x01\x00\xB3\x11\x01\x00\xB2\x19\x05\x00\xBD\xBD"
-     "\x02\x00"
-     "\xBC\xBD\x02\x00\xBF\x85\x02\x00\xBE\x85\x02\x00\xB9\x25\x01\x00\xB8\x21"
-     "\x05\x00"
-     "\xBB\x25\x05\x00\xBA\x25\x05\x00\xA5\x2D\x02\x00\xA4\x2D\x02\x00\xA7\x39"
-     "\x02\x00"
-     "\xA6\x29\x06\x00\xA1\x0D\x02\x00\xA0\x0D\x02\x00\xA3\x3D\x02\x00\xA2\x3D"
-     "\x02\x00"
-     "\xAD\x1D\x02\x00\xAC\x1D\x06\x00\xAF\xBD\x05\x00\xAE\x11\x06\x00\xA9\x19"
-     "\x06\x00"
-     "\xA8\x21\x06\x00\xAB\x01\x02\x00\xAA\x09\x02\x00\x80\x11\x0D\x00\x81\x11"
-     "\x0D\x00"
-     "\x82\x29\x0D\x00\x83\x29\x0D\x00\x84\x25\x0D\x00\x85\x2D\x0D\x00\x86\x25"
-     "\x0D\x00"
-     "\x87\x5D\x0D\x00\x88\x65\x0D\x00\x89\x75\x0D\x00\x8A\x71\x0D\x00\x8B\x71"
-     "\x0D\x00"
-     "\x8C\x51\x0D\x00\x8D\xD9\x0E\x00\x8E\x55\x0D\x00\x8F\x4D\x0D\x00\x90\x35"
-     "\x0D\x00"
-     "\x91\x3D\x0D\x00\x92\xCD\x0D\x00\x93\xC5\x0D\x00\x94\xDD\x0D\x00\x95\xC5"
-     "\x0D\x00"
-     "\x96\x59\x01\x00\x97\xC1\x0D\x00\x98\xC1\x0D\x00\x99\x61\x01\x00\x9A\xC5"
-     "\x0D\x00"
-     "\x9B\xDD\x0D\x00\x9C\xC5\x0D\x00\x9D\xCD\x0D\x00\x9E\x71\x01\x00\x9F\x99"
-     "\x09\x00"
-     "\xA0\x39\x0E\x00\xA1\x7D\x09\x00\xA2\x35\x0E\x00\xA3\x09\x0A\x00\xA4\x75"
-     "\x0E\x00"
-     "\xA5\x01\x0A\x00\xA6\x95\x01\x00\xA7\x15\x0B\x00\xA8\xB1\x01\x00\xA9\xB1"
-     "\x01\x00"
-     "\xAA\xB1\x01\x00\xAB\xED\x0D\x00\xAC\xF5\x0D\x00\xAD\xD5\x0D\x00\xAE\x29"
-     "\x0D\x00"
-     "\xAF\xBD\x01\x00\xB0\x79\x0D\x00\xB1\xCD\x01\x00\xB2\xD5\x01\x00\xB3\xED"
-     "\x01\x00"
-     "\xB4\xF5\x01\x00\xB5\xFD\x01\x00\xB6\xF5\x01\x00\xB7\xED\x01\x00\xB8\xD5"
-     "\x01\x00"
-     "\xB9\xDD\x01\x00\xBA\xD5\x01\x00\xBB\xAD\x01\x00\xBC\xB5\x01\x00\xBD\xBD"
-     "\x01\x00"
-     "\xBE\xB5\x01\x00\xBF\xAD\x01\x00\x19\x16\x00\x80\x4D\x16\x00\x80\xAD\x16"
-     "\x00\x80"
-     "\x31\x11\x00\x80\x4E\x12\x00\x80\x89\x12\x00\x80\x9C\x12\x00\x80\xFB\x13"
-     "\x00\x80"
-     "\x07\x14\x00\x80\x13\x14\x00\x80\x77\x13\x00\x80\x7C\x13\x00\x80\xBB\x13"
-     "\x00\x80"
-     "\xC0\x13\x00\x80\xE7\x13\x00\x80\xEC\x13\x00\x80\x53\x12\x00\x80\x58\x12"
-     "\x00\x80"
-     "\xA1\x12\x00\x80\xA6\x12\x00\x80\xAC\x11\x00\x80\xF0\x10\xFC\x00\xB1\x11"
-     "\x00\x80"
-     "\x80\x11\x00\x80\x85\x11\x00\x80\xC8\x11\x00\x80\xCD\x11\x00\x80\x10\x13"
-     "\x00\x80"
-     "\x15\x13\x00\x80\x4B\x13\x00\x80\x50\x13\x00\x80\x60\x12\x00\x80\x6C\x12"
-     "\x00\x80"
-     "\x7F\x12\x00\x80\x84\x12\x00\x80\x2F\x12\x00\x80\xC6\x12\x00\x80\x41\x13"
-     "\x00\x80"
-     "\x35\x10\x00\x80\x8D\x10\x00\x80\xA7\x10\x00\x80\xAE\x10\x00\x80\xBA\x10"
-     "\x00\x80"
-     "\xDE\x10\x00\x80\xE5\x10\x00\x80\xEC\x10\x00\x80\x6B\x11\x00\x80\xA5\x11"
-     "\x00\x80"
-     "\x9E\x11\x00\x80\x15\x12\x00\x80\x02\x12\x00\x80\x09\x12\x00\x80\x40\x12"
-     "\x00\x80"
-     "\x39\x12\x00\x80\x47\x12\x00\x80\xB3\x12\x00\x80\x09\x13\x00\x80\x71\x12"
-     "\x00\x80"
-     "\x65\x12\x00\x80\x8E\x12\x00\x80\x95\x12\x00\x80\x27\x13\x00\x80\x3A\x13"
-     "\x00\x80"
-     "\x62\x13\x00\x80\xB6\xA5\x09\x00\xB7\xAD\x09\x00\xB4\x95\x09\x00\xB5\xAD"
-     "\x09\x00"
-     "\xB2\xF9\x0C\x00\xB3\x8D\x09\x00\xB0\x91\x09\x00\xB1\x91\x09\x00\xBE\x29"
-     "\x0F\x00"
-     "\xBF\x29\x0F\x00\xBC\x29\x0F\x00\xBD\x29\x0F\x00\xBA\x89\x09\x00\xBB\x89"
-     "\x09\x00"
-     "\xB8\x95\x09\x00\xB9\x89\x09\x00\xA6\xD5\x08\x00\xA7\xF9\x08\x00\xA4\xCD"
-     "\x08\x00"
-     "\xA5\xA9\x0C\x00\xA2\x2D\x08\x00\xA3\xD5\x08\x00\xA0\xE5\x0C\x00\xA1\x01"
-     "\x08\x00"
-     "\xAE\xE1\x09\x00\xAF\xE1\x09\x00\xAC\x89\x08\x00\xAD\x89\x08\x00\xAA\xBD"
-     "\x08\x00"
-     "\xAB\xA5\x08\x00\xA8\xB9\x08\x00\xA9\xB5\x08\x00\x96\x41\x0F\x00\x97\x59"
-     "\x0B\x00"
-     "\x94\x39\x0C\x00\x95\x31\x0F\x00\x92\x3D\x0C\x00\x93\x2D\x0F\x00\x90\x55"
-     "\x0F\x00"
-     "\x91\x35\x0C\x00\x9E\x55\x0B\x00\x9F\x5D\x0B\x00\x9C\x45\x0B\x00\x9D\x5D"
-     "\x0B\x00"
-     "\x9A\x55\x0B\x00\x9B\x0D\x0C\x00\x98\x59\x0B\x00\x99\x09\x0C\x00\x86\xA5"
-     "\x0F\x00"
-     "\x87\xAD\x0F\x00\x84\x85\x0F\x00\x85\xAD\x0F\x00\x82\x95\x0F\x00\x83\x9D"
-     "\x0F\x00"
-     "\x80\xE9\x0F\x00\x81\x9D\x0F\x00\x8E\x75\x0F\x00\x8F\x7D\x0F\x00\x8C\x79"
-     "\x0F\x00"
-     "\x8D\x49\x0C\x00\x8A\x6D\x0C\x00\x8B\x55\x0C\x00\x88\x61\x0C\x00\x89\x85"
-     "\x0F\x00"
-     "\x95\x39\x02\x00\x94\x31\x02\x00\x97\x49\x02\x00\x96\x31\x02\x00\x91\x39"
-     "\x02\x00"
-     "\x90\x81\x02\x00\x93\x29\x02\x00\x92\x31\x02\x00\x9D\x7D\x02\x00\x9C\x7D"
-     "\x02\x00"
-     "\x9F\x4D\x02\x00\x9E\x4D\x02\x00\x99\x79\x02\x00\x98\x71\x02\x00\x9B\x69"
-     "\x02\x00"
-     "\x9A\x71\x02\x00\x85\x2D\x02\x00\x84\x2D\x02\x00\x87\x3D\x02\x00\x86\x3D"
-     "\x02\x00"
-     "\x81\x0D\x02\x00\x80\x0D\x02\x00\x83\x3D\x02\x00\x82\x3D\x02\x00\x8D\xE9"
-     "\x02\x00"
-     "\x8C\x71\x02\x00\x8F\xF9\x02\x00\x8E\xE1\x02\x00\x89\x0D\x02\x00\x88\x0D"
-     "\x02\x00"
-     "\x8B\xFD\x02\x00\x8A\xFD\x02\x00\xB5\x01\x08\x00\xB4\x61\x09\x00\xB7\xF5"
-     "\x02\x00"
-     "\xB6\x09\x08\x00\xB1\x81\x02\x00\xB0\x81\x02\x00\xB3\x29\x08\x00\xB2\x85"
-     "\x02\x00"
-     "\xBD\x39\x02\x00\xBC\x3D\x02\x00\x69\x13\x00\x80\x70\x13\x00\x80\xB9\x35"
-     "\x02\x00"
-     "\xB8\x35\x02\x00\xBB\x0D\x08\x00\xBA\x0D\x08\x00\xA5\xBD\x02\x00\xA4\xBD"
-     "\x02\x00"
-     "\xA7\x4D\x02\x00\xA6\x4D\x02\x00\xA1\xBD\x02\x00\xA0\xBD\x02\x00\xA3\xAD"
-     "\x02\x00"
-     "\xA2\xAD\x02\x00\xAD\x9D\x02\x00\xAC\x71\x02\x00\xAF\x6D\x02\x00\xAE\x6D"
-     "\x02\x00"
-     "\xA9\x79\x02\x00\xA8\xE1\x02\x00\xAB\x69\x02\x00\xAA\x71\x02\x00\x91\x13"
-     "\x00\x80"
-     "\xA6\x13\x00\x80\xAD\x13\x00\x80\xD2\x13\x00\x80\xE0\x13\x00\x80\xD9\x13"
-     "\x00\x80"
-     "\x36\x14\x00\x80\x67\x14\x00\x80\x6E\x14\x00\x80\xF5\x14\x00\x80\x03\x15"
-     "\x00\x80"
-     "\xFC\x14\x00\x80\xDB\x14\x00\x80\x0F\x15\x00\x80\x1B\x15\x00\x80\x66\x15"
-     "\x00\x80"
-     "\x6D\x15\x00\x80\x7D\x15\x00\x80\x45\x15\x00\x80\x53\x15\x00\x80\x89\x15"
-     "\x00\x80"
-     "\x90\x15\x00\x80\x4C\x15\x00\x80\x35\x16\x00\x80\x3C\x16\x00\x80\xDD\x15"
-     "\x00\x80"
-     "\xE4\x15\x00\x80\xC3\x15\x00\x80\xCA\x15\x00\x80\x0D\x16\x00\x80\x06\x16"
-     "\x00\x80"
-     "\x9A\x16\x00\x80\x58\x10\x00\x80\x9B\x10\x00\x80\x94\x10\x00\x80\xC8\x10"
-     "\x00\x80"
-     "\xC1\x10\x00\x80\xFA\x10\x00\x80\xF3\x10\x00\x80\x79\x11\x00\x80\x48\x11"
-     "\x00\x80"
-     "\x72\x11\x00\x80\x21\x12\x00\x80\xBA\x12\x00\x80\x78\x12\x00\x80\x2E\x13"
-     "\x00\x80"
-     "\xE9\x14\x00\x80\x22\x15\x00\x80\x7F\x16\x00\x80\xCC\x30\x8E\x02\x73\x16"
-     "\x00\x80"
-     "\xA1\x16\x00\x80\xA3\x15\x00\x80\x75\x14\x00\x80\xEB\x15\x00\x80\xB4\x13"
-     "\x00\x80"
-     "\xAE\x14\x00\x80\xBC\x15\x00\x80\x97\x15\x00\x80\xE2\x14\x00\x80\xA7\x14"
-     "\x00\x80"
-     "\x4F\x11\x00\x80\x92\x11\x00\x80\x5A\x15\x00\x80\x80\x25\x01\x00\x81\x2D"
-     "\x01\x00"
-     "\x82\x25\x01\x00\x83\x3D\x01\x00\x84\x21\x01\x00\x85\x21\x01\x00\x86\x2D"
-     "\x01\x00"
-     "\x87\x25\x01\x00\x88\x19\x01\x00\x89\x19\x01\x00\x8A\x69\x01\x00\x8B\x69"
-     "\x01\x00"
-     "\x8C\x79\x01\x00\x8D\x79\x01\x00\x8E\x69\x01\x00\x8F\x69\x01\x00\x42\x14"
-     "\x00\x80"
-     "\x4C\x10\x00\x80\x92\x11\x01\x00\x93\x11\x01\x00\x94\x31\x01\x00\x95\x31"
-     "\x01\x00"
-     "\x96\x31\x01\x00\x97\x31\x01\x00\x98\x0D\x01\x00\x99\x15\x01\x00\x9A\x1D"
-     "\x01\x00"
-     "\x9B\x15\x01\x00\x9C\x0D\x01\x00\x9D\x09\x01\x00\x9E\xF9\x01\x00\x9F\xF9"
-     "\x01\x00"
-     "\xA0\x09\x01\x00\xA1\x09\x01\x00\xA2\x19\x01\x00\xA3\x19\x01\x00\xA4\x09"
-     "\x01\x00"
-     "\xA5\x09\x01\x00\xA6\x39\x01\x00\xA7\x39\x01\x00\xA8\x09\x01\x00\xA9\x09"
-     "\x01\x00"
-     "\xAA\x19\x01\x00\xAB\x19\x01\x00\xAC\x09\x01\x00\xAD\x09\x01\x00\xAE\x79"
-     "\x01\x00"
-     "\xAF\x75\x01\x00\xB0\x0D\x01\x00\xB1\x15\x01\x00\xB2\x1D\x01\x00\xB3\x15"
-     "\x01\x00"
-     "\xB4\xCD\x01\x00\xB5\xF1\x01\x00\xB6\x05\x01\x00\xB7\x3D\x01\x00\xB8\x05"
-     "\x01\x00"
-     "\xB9\x0D\x01\x00\xBA\x05\x01\x00\xBB\x1D\x01\x00\xBC\x1D\x01\x00\xBD\x11"
-     "\x01\x00"
-     "\xBE\x0D\x01\x00\xBF\x05\x01\x00\xB1\xFD\x07\x00\xB0\xFD\x07\x00\xB3\xA1"
-     "\x05\x00"
-     "\xB2\xF1\x07\x00\xB5\x11\x02\x00\xB4\xB9\x05\x00\xB7\x1D\x02\x00\xB6\x19"
-     "\x02\x00"
-     "\xB9\x2D\x02\x00\xB8\x2D\x02\x00\xBB\x3D\x02\x00\xBA\x3D\x02\x00\xBD\x29"
-     "\x02\x00"
-     "\xBC\x2D\x02\x00\xBF\x19\x02\x00\xBE\x21\x02\x00\xA1\x95\x07\x00\xA0\x95"
-     "\x07\x00"
-     "\xA3\x91\x07\x00\xA2\x95\x07\x00\xA5\x8D\x07\x00\xA4\x89\x07\x00\xA7\xBD"
-     "\x07\x00"
-     "\xA6\xBD\x07\x00\xA9\x89\x07\x00\xA8\x81\x07\x00\xAB\x85\x07\x00\xAA\x81"
-     "\x07\x00"
-     "\xAD\x81\x07\x00\xAC\x85\x07\x00\xAF\x81\x07\x00\xAE\x89\x07\x00\x91\x21"
-     "\x07\x00"
-     "\x90\x19\x07\x00\x93\x2D\x07\x00\x92\x2D\x07\x00\x95\x3D\x07\x00\x94\x3D"
-     "\x07\x00"
-     "\x97\x2D\x07\x00\x96\x2D\x07\x00\x99\x19\x07\x00\x98\x1D\x07\x00\x9B\x69"
-     "\x07\x00"
-     "\x9A\x11\x07\x00\x9D\x79\x07\x00\x9C\x71\x07\x00\x9F\x69\x07\x00\x9E\x71"
-     "\x07\x00"
-     "\x28\x12\x00\x80\xD1\x15\x00\x80\xFD\x12\x00\x80\x5B\x14\x00\x80\x2C\x10"
-     "\x00\x80"
-     "\x23\x10\x00\x80\x74\x15\x00\x80\xCB\x12\x00\x80\xE0\x11\x00\x80\x98\x13"
-     "\x00\x80"
-     "\x55\x16\x00\x80\xD4\x12\x00\x80\xDB\x12\x00\x80\x56\x11\x00\x80\xEE\x11"
-     "\x00\x80"
-     "\x3D\x03\x00\x80\x80\xFD\x03\x00\x81\x45\x02\x00\x82\x4D\x02\x00\x83\xFD"
-     "\x03\x00"
-     "\x84\x11\x03\x00\x85\xF5\x06\x00\x86\xD9\x01\x00\x87\xD9\x01\x00\x3B\x24"
-     "\x00\x80"
-     "\x3F\x24\x00\x80\x37\x06\x00\x80\xC6\x00\x00\x80\x4C\x06\x00\x80\xD9\x24"
-     "\x00\x80"
-     "\xDD\x24\x00\x80\x87\x20\x00\x80\x83\x20\x00\x80\x5E\x06\x00\x80\x8E\x08"
-     "\x00\x80"
-     "\x92\x08\x00\x80\x6F\x24\x00\x80\xD5\x24\x00\x80\x63\x24\x00\x80\x67\x24"
-     "\x00\x80"
-     "\x4B\x24\x00\x80\x4F\x24\x00\x80\x53\x24\x00\x80\x57\x24\x00\x80\x5B\x24"
-     "\x00\x80"
-     "\x5F\x24\x00\x80\x56\x06\x00\x80\x5A\x06\x00\x80\xD0\x00\x00\x80\xD6\x00"
-     "\x00\x80"
-     "\x39\x03\x00\x80\x3F\x03\x00\x80\x58\x06\x00\x80\xD2\x00\x00\x80\xD4\x00"
-     "\x00\x80"
-     "\x54\x06\x00\x80\x78\x00\x00\x80\xF1\x13\x00\x80\x7C\x00\x00\x80\x86\x00"
-     "\x00\x80"
-     "\x90\x00\x00\x80\xF6\x13\x00\x80\x9A\x00\x00\x80\x02\x14\x00\x80\xB0\x09"
-     "\x01\x00"
-     "\xB1\x09\x01\x00\xB2\x19\x01\x00\xB3\x19\x01\x00\xB4\x09\x01\x00\xB5\x09"
-     "\x01\x00"
-     "\xB6\x39\x01\x00\xB7\x39\x01\x00\xB8\x09\x01\x00\xB9\x09\x01\x00\xBA\x19"
-     "\x01\x00"
-     "\xBB\x19\x01\x00\xBC\x09\x01\x00\x3F\x06\x00\x80\xA4\x00\x00\x80\x0E\x14"
-     "\x00\x80"
-     "\x80\x79\xAE\x00\x81\x6D\x1F\x00\x82\x65\x1F\x00\x83\x7D\x1F\x00\x84\x65"
-     "\x1F\x00"
-     "\x4A\x06\x00\x80\xAE\x00\x00\x80\x87\x65\x1F\x00\x88\x5D\x1F\x00\x89\x3D"
-     "\xA9\x00"
-     "\x8A\x31\xA9\x00\x8B\x35\xA9\x00\x8C\x29\xA9\x00\x8D\x71\x1F\x00\x8E\x7D"
-     "\x1F\x00"
-     "\x8F\x79\x1F\x00\x90\xBD\x1E\x00\x91\x65\x1F\x00\x92\x81\xA7\x00\x1A\x14"
-     "\x00\x80"
-     "\x94\x71\x80\x02\x95\x7D\x1F\x00\x96\x69\x1F\x00\x97\x71\x1F\x00\x98\x5D"
-     "\x1F\x00"
-     "\x99\x01\xA9\x00\x9A\x05\xA9\x00\x9B\x25\x1F\x00\x9C\x3D\x1F\x00\x9D\x25"
-     "\x1F\x00"
-     "\x9E\x2D\x1F\x00\x9F\xFD\x1F\x00\xA0\x05\x1F\x00\xA1\x0D\x1F\x00\xA2\xD1"
-     "\xA9\x00"
-     "\xA3\x01\x1F\x00\xA4\x65\x1F\x00\xA5\x81\x00\x00\xA6\xF1\xA9\x00\xB8\x00"
-     "\x00\x80"
-     "\xA8\x31\x1F\x00\xA9\x31\x1F\x00\xAA\x31\x1F\x00\xAB\x31\x1F\x00\x1F\x14"
-     "\x00\x80"
-     "\xB9\x0F\x00\x80\x20\x10\x00\x80\x3F\x10\x00\x80\xB0\x61\x1F\x00\xB1\x61"
-     "\x1F\x00"
-     "\xB2\x61\x1F\x00\x82\x10\x00\x80\xB4\x7D\x1F\x00\x85\x10\x00\x80\xB6\x69"
-     "\x1F\x00"
-     "\xB7\x69\x1F\x00\xB8\x59\x1F\x00\xB9\x59\x1F\x00\xBA\x11\x1F\x00\xBB\x11"
-     "\x1F\x00"
-     "\xBC\xE9\x00\x00\xBD\xB9\x00\x00\xBE\x65\x00\x00\xBF\x4D\x00\x00\x80\xB5"
-     "\x01\x00"
-     "\x81\x65\xB3\x02\x82\x69\xB3\x02\x83\x65\xB3\x02\x84\x79\xB3\x02\x85\x95"
-     "\xB3\x02"
-     "\x86\x99\xB3\x02\x87\x6D\xB3\x02\x88\x51\xB3\x02\x89\xAD\xB3\x02\x8A\xA1"
-     "\xB3\x02"
-     "\x8B\xA5\xB3\x02\x8C\xB9\xB3\x02\x8D\x8D\x01\x00\x8E\x81\x01\x00\x8F\x81"
-     "\x01\x00"
-     "\x90\xFD\x01\x00\x91\xF9\x01\x00\x92\xF5\x01\x00\x93\x81\x01\x00\x94\x9D"
-     "\x01\x00"
-     "\x95\x81\x01\x00\x96\x8D\x01\x00\x97\x89\x01\x00\x98\xB5\x01\x00\x99\x25"
-     "\x03\x00"
-     "\x9A\x29\x03\x00\x9B\x2D\x03\x00\x9C\x31\x03\x00\x9D\x31\x03\x00\x9E\x3D"
-     "\x03\x00"
-     "\x9F\x39\x03\x00\xA0\xC5\x03\x00\xA1\xCD\x03\x00\xA2\xC1\x03\x00\xA3\xC5"
-     "\x03\x00"
-     "\xA4\xD9\x03\x00\xA5\xD9\x03\x00\xA6\xD5\x03\x00\xA7\xD1\x03\x00\xA8\xED"
-     "\x03\x00"
-     "\xA9\xF5\x03\x00\xAA\xF9\x03\x00\xAB\xF9\x03\x00\xAC\xE5\x03\x00\xAD\xED"
-     "\x03\x00"
-     "\xAE\xE1\x03\x00\xAF\xE1\x03\x00\xB0\x9D\x03\x00\xB1\xA5\x03\x00\xB2\xA9"
-     "\x03\x00"
-     "\xB3\xAD\x03\x00\xB4\xB1\x03\x00\xB5\xB1\x03\x00\xB6\xBD\x03\x00\xB7\xB9"
-     "\x03\x00"
-     "\xB8\x85\x03\x00\xB9\x8D\x03\x00\xBA\x81\x03\x00\xBB\x85\x03\x00\xBC\x99"
-     "\x03\x00"
-     "\xBD\x99\x03\x00\xBE\x95\x03\x00\xBF\x91\x03\x00\xBC\x5D\x1D\x00\xBD\x65"
-     "\x1D\x00"
-     "\xBE\x6D\x1D\x00\xBF\x79\xAC\x00\xB8\x5D\x1D\x00\xB9\x45\x1D\x00\xBA\x4D"
-     "\x1D\x00"
-     "\xBB\x45\x1D\x00\xB4\x65\x1D\x00\xB5\x41\xAB\x00\xB6\x45\xAB\x00\xB7\x65"
-     "\x1D\x00"
-     "\xB0\xE1\xA5\x00\xB1\x09\x1D\x00\xB2\x79\x1D\x00\xB3\x79\x1D\x00\x15\x11"
-     "\x00\x80"
-     "\x40\x11\x00\x80\x8A\x11\x00\x80\xB6\x11\x00\x80\xD2\x11\x00\x80\xD5\x11"
-     "\x00\x80"
-     "\xDD\x11\x00\x80\xF7\x11\x00\x80\xFA\x11\x00\x80\x5D\x12\x00\x80\xAB\x12"
-     "\x00\x80"
-     "\x1A\x13\x00\x80\x55\x13\x00\x80\x81\x13\x00\x80\x89\x13\x00\x80\xC5\x13"
-     "\x00\x80"
-     "\x24\x14\x00\x80\x53\x14\x00\x80\x86\x14\x00\x80\xBF\x14\x00\x80\x98\xA9"
-     "\x1D\x00"
-     "\x99\x4D\xA5\x00\x38\x15\x00\x80\xB4\x15\x00\x80\x94\x81\x82\x02\x95\x89"
-     "\x1D\x00"
-     "\x96\x99\x1D\x00\x97\x99\x1D\x00\x90\x4D\x1C\x00\x91\x95\x1D\x00\x92\x9D"
-     "\x1D\x00"
-     "\x93\x95\x1D\x00\x28\x16\x00\x80\xC2\x14\x00\x80\xC7\x14\x00\x80\xCC\x14"
-     "\x00\x80"
-     "\xD1\x14\x00\x80\xCE\x00\x00\x80\xD8\x00\x00\x80\x4A\x03\x00\x80\x5C\x06"
-     "\x00\x80"
-     "\x90\x08\x00\x80\x94\x08\x00\x80\xD1\x21\x00\x80\xD5\x21\x00\x80\xAF\x2A"
-     "\x00\x80"
-     "\x63\x25\x00\x80\x9C\x25\x00\x80\xF0\x10\x10\x00\xBB\x25\x00\x80\xCD\x25"
-     "\x00\x80"
-     "\xF2\x25\x00\x80\xF1\x29\x00\x80\x0D\x2A\x00\x80\x1F\x2A\x00\x80\xB6\x27"
-     "\x00\x80"
-     "\xB3\x2A\x00\x80\xCC\xE8\x91\x02\x97\x2A\x00\x80\x57\x25\x00\x80\x5B\x25"
-     "\x00\x80"
-     "\x96\x08\x00\x80\x99\x08\x00\x80\xA2\x08\x00\x80\xA6\xD9\x00\x00\x9F\x08"
-     "\x00\x80"
-     "\xA4\x0D\x00\x00\xA5\x29\x00\x00\xA2\x6D\x00\x00\xA3\x01\x98\x02\xA0\x6D"
-     "\x00\x00"
-     "\xA1\x65\x00\x00\xAE\xD9\x00\x00\x9C\x08\x00\x80\xAC\xCD\x00\x00\xAD\xC1"
-     "\x00\x00"
-     "\xAA\xD1\x00\x00\xAB\x79\x07\x00\xA8\xE5\x00\x00\xA9\x75\x07\x00\x96\xB5"
-     "\xCB\x00"
-     "\x97\xBD\xCB\x00\x94\xA5\xCB\x00\x95\xBD\xCB\x00\x92\xB5\xCB\x00\x93\xBD"
-     "\xCB\x00"
-     "\xBA\x20\x00\x80\xC5\x21\x00\x80\xCA\x20\x00\x80\xD2\x20\x00\x80\x9C\xE5"
-     "\xCB\x00"
-     "\xC9\x21\x00\x80\x9A\x8D\xCB\x00\x9B\xFD\xCB\x00\xCD\x21\x00\x80\xDD\xEB"
-     "\x00\x80"
-     "\x86\x29\xC5\x00\x87\x35\xC5\x00\x84\x29\xC5\x00\x85\x29\xC5\x00\x82\x01"
-     "\xC5\x00"
-     "\x83\x09\xC5\x00\xE2\xEB\x00\x80\xD9\x94\xB2\x02\x8E\xFD\xCB\x00\x8F\xC5"
-     "\xCB\x00"
-     "\x8C\xE5\xCB\x00\x8D\xF5\xCB\x00\x8A\x05\xC5\x00\x8B\xFD\xCB\x00\xE7\xEB"
-     "\x00\x80"
-     "\xEC\xEB\x00\x80\x52\x16\x00\x80\x6B\x16\x00\x80\xC6\x20\x00\x80\xCE\x20"
-     "\x00\x80"
-     "\xF1\xEB\x00\x80\x9D\x10\x34\x00\x9E\x28\x34\x00\x9F\x8C\x35\x00\x90\x20"
-     "\x02\x00"
-     "\x91\x2C\x09\x00\xF6\xEB\x00\x80\xFB\xEB\x00\x80\x00\xEC\x00\x80\x05\xEC"
-     "\x00\x80"
-     "\x96\xE0\x36\x00\x0A\xEC\x00\x80\x90\xC0\x01\x00\x0F\xEC\x00\x80\x92\x88"
-     "\x0F\x00"
-     "\x93\xCC\x0F\x00\x14\xEC\x00\x80\x95\x94\x0D\x00\x96\xD8\x0D\x00\x19\xEC"
-     "\x00\x80"
-     "\x1E\xEC\x00\x80\x23\xEC\x00\x80\x28\xEC\x00\x80\x2D\xEC\x00\x80\x32\xEC"
-     "\x00\x80"
-     "\x37\xEC\x00\x80\x9E\x00\x08\x00\x3C\xEC\x00\x80\x41\xEC\x00\x80\x46\xEC"
-     "\x00\x80"
-     "\x4B\xEC\x00\x80\x50\xEC\x00\x80\x55\xEC\x00\x80\x5A\xEC\x00\x80\x5F\xEC"
-     "\x00\x80"
-     "\x64\xEC\x00\x80\x69\xEC\x00\x80\x6E\xEC\x00\x80\x73\xEC\x00\x80\x78\xEC"
-     "\x00\x80"
-     "\x7D\xEC\x00\x80\x82\xEC\x00\x80\x87\xEC\x00\x80\x8C\xEC\x00\x80\x91\xEC"
-     "\x00\x80"
-     "\x96\xEC\x00\x80\xB2\xC8\x0B\x00\x9B\xEC\x00\x80\xA0\xEC\x00\x80\xA5\xEC"
-     "\x00\x80"
-     "\xAA\xEC\x00\x80\xAF\xD8\x2C\x00\xAF\xEC\x00\x80\xB4\xEC\x00\x80\xB9\xEC"
-     "\x00\x80"
-     "\xBE\xEC\x00\x80\xC3\xEC\x00\x80\xC8\xEC\x00\x80\xCD\xEC\x00\x80\xF0\xB0"
-     "\x09\x00"
-     "\x81\x2D\x1A\x00\xD2\xEC\x00\x80\x83\x89\x1A\x00\x82\xD9\x07\x00\x85\x89"
-     "\x1A\x00"
-     "\x84\x81\x1A\x00\x87\xC1\x07\x00\x86\x99\x1A\x00\x89\x4D\xAC\x00\x88\x41"
-     "\xAC\x00"
-     "\x8B\x71\xAC\x00\x8A\xA5\x1A\x00\x8D\xBD\x1A\x00\x8C\xC9\x1B\x00\x8F\xE5"
-     "\x07\x00"
-     "\x8E\xF5\xA2\x00\x91\x95\x9E\x02\x90\x75\xAC\x00\x93\xC5\x9E\x02\x92\xC5"
-     "\x9E\x02"
-     "\x95\x69\xAC\x00\x94\x61\xAC\x00\x97\x19\xAC\x00\x96\x61\xAC\x00\x99\x31"
-     "\xAC\x00"
-     "\x98\x2D\xAC\x00\x9B\x49\x85\x02\x9A\x45\x1A\x00\x9D\x19\xAC\x00\x9C\x81"
-     "\x1A\x00"
-     "\x9F\x49\x1A\x00\x9E\x69\x05\x00\xA1\xB5\x95\x02\xA0\x19\x1A\x00\xA3\x79"
-     "\x90\x02"
-     "\xA2\x41\x96\x02\xA5\x81\x91\x02\xA4\x9D\x91\x02\xA7\xA9\x91\x02\xA6\xB1"
-     "\x91\x02"
-     "\xA9\xF5\x91\x02\xA8\x9D\x91\x02\xAB\x6D\x91\x02\xAA\x19\x91\x02\xAD\x8D"
-     "\x90\x02"
-     "\xAC\x59\x91\x02\xAF\x41\x90\x02\xAE\x2D\x90\x02\xB1\x65\x92\x02\xB0\xED"
-     "\x9D\x02"
-     "\xB3\x89\x9D\x02\xB2\xFD\x9D\x02\xB5\x45\x9D\x02\xB4\x8D\x9D\x02\xB7\x55"
-     "\x9C\x02"
-     "\xB6\xB9\x9C\x02\xB9\x2D\x9C\x02\xB8\x55\x9C\x02\xBB\x91\x1A\x00\xBA\x79"
-     "\x98\x02"
-     "\xBD\x8D\x1A\x00\xBC\x65\x1A\x00\xBF\xBD\x1A\x00\xBE\x1D\x07\x00\x87\x54"
-     "\x0B\x00"
-     "\xD7\xEC\x00\x80\xDC\xEC\x00\x80\xBC\xDD\x04\x00\xBB\xC1\x04\x00\xBA\xD9"
-     "\x04\x00"
-     "\xB9\xD5\x04\x00\xB8\xD5\x04\x00\xB7\xE9\x04\x00\xB6\xE1\x04\x00\xB5\xED"
-     "\x04\x00"
-     "\xB4\x29\x02\x00\xB3\x35\x02\x00\xB2\x31\x02\x00\xB1\x3D\x02\x00\xB0\x89"
-     "\x17\x00"
-     "\xAF\xF5\x17\x00\xAE\x45\x02\x00\xAD\x49\x02\x00\xAC\x8D\x04\x00\xAB\x91"
-     "\x04\x00"
-     "\xAA\x95\x04\x00\xA9\x99\x04\x00\xA8\xC1\x04\x00\xA7\xFD\x04\x00\xA6\xF9"
-     "\x04\x00"
-     "\xA5\xF5\x04\x00\xA4\xF5\x04\x00\xA3\xE9\x04\x00\xA2\xED\x04\x00\xA1\xE1"
-     "\x04\x00"
-     "\xA0\xF1\x04\x00\x9F\x0D\x04\x00\x9E\x09\x04\x00\x9D\x05\x04\x00\x9C\x05"
-     "\x04\x00"
-     "\x9B\x19\x04\x00\x9A\x1D\x04\x00\x99\x11\x04\x00\x98\x19\x04\x00\x97\x25"
-     "\x04\x00"
-     "\x96\x21\x04\x00\x95\x2D\x04\x00\x94\x2D\x04\x00\x93\x31\x04\x00\x92\x35"
-     "\x04\x00"
-     "\x91\x39\x04\x00\x90\xC1\x04\x00\x8F\xBD\x04\x00\x8E\xB9\x04\x00\x8D\xB5"
-     "\x04\x00"
-     "\x8C\xB5\x04\x00\x8B\xA9\x04\x00\x8A\xAD\x04\x00\x89\xA1\x04\x00\x88\xA9"
-     "\x04\x00"
-     "\x87\x95\x04\x00\x86\x91\x04\x00\x85\x9D\x04\x00\x84\x9D\x04\x00\x83\x81"
-     "\x04\x00"
-     "\x82\x85\x04\x00\x81\x89\x04\x00\x80\x91\x04\x00\xA0\xB1\x03\x00\xA1\xB1"
-     "\x03\x00"
-     "\xA2\xB1\x03\x00\xA3\xB1\x03\x00\xA4\xD1\x03\x00\xA5\xD1\x03\x00\xA6\xD5"
-     "\x03\x00"
-     "\xA7\xCD\x03\x00\xE1\xEC\x00\x80\xE6\xEC\x00\x80\xEB\xEC\x00\x80\xF0\xEC"
-     "\x00\x80"
-     "\xF5\xEC\x00\x80\xFA\xEC\x00\x80\xCC\xC4\xA7\x00\xCC\xE0\xA7\x00\xFF\xEC"
-     "\x00\x80"
-     "\x04\xED\x00\x80\x09\xED\x00\x80\x0E\xED\x00\x80\x13\xED\x00\x80\x18\xED"
-     "\x00\x80"
-     "\x1D\xED\x00\x80\x22\xED\x00\x80\xD9\x38\xB6\x02\xD9\xF8\xB7\x02\x27\xED"
-     "\x00\x80"
-     "\x2C\xED\x00\x80\x31\xED\x00\x80\x36\xED\x00\x80\x3B\xED\x00\x80\x40\xED"
-     "\x00\x80"
-     "\x80\x3D\x04\x00\x81\x5D\x04\x00\x82\x71\x04\x00\x83\x71\x04\x00\x84\x81"
-     "\x03\x00"
-     "\x85\xBD\x03\x00\x86\xB5\x03\x00\x87\xAD\x03\x00\x88\x95\x03\x00\x89\x99"
-     "\x03\x00"
-     "\x8A\xED\x03\x00\x8B\xFD\x03\x00\x8C\xED\x03\x00\x8D\xD5\x03\x00\x8E\xDD"
-     "\x03\x00"
-     "\x8F\xD5\x03\x00\x90\xAD\x03\x00\x91\xB5\x03\x00\x92\xBD\x03\x00\x93\xB1"
-     "\x03\x00"
-     "\x94\x51\x03\x00\x95\x51\x03\x00\x96\x51\x03\x00\x97\x51\x03\x00\x98\x71"
-     "\x03\x00"
-     "\x99\x71\x03\x00\x9A\x71\x03\x00\x9B\x71\x03\x00\x9C\x51\x03\x00\x9D\x51"
-     "\x03\x00"
-     "\x9E\x51\x03\x00\x9F\x51\x03\x00\x80\x81\xF9\x00\x81\x81\xF9\x00\x82\x81"
-     "\xF9\x00"
-     "\x83\x81\xF9\x00\x84\x81\xF9\x00\x85\x75\xF9\x00\x86\x7D\xF9\x00\x87\x75"
-     "\xF9\x00"
-     "\x88\x4D\xF9\x00\x89\x55\xF9\x00\x8A\xBD\xF9\x00\x8B\xB5\xF9\x00\x8C\xAD"
-     "\xF9\x00"
-     "\x8D\xF5\xF9\x00\x8E\xFD\xF9\x00\x8F\x49\xF9\x00\x90\x39\xF9\x00\x91\x39"
-     "\xF9\x00"
-     "\x92\xE9\xF9\x00\x93\xE9\xF9\x00\x94\xF9\xF9\x00\x95\xF9\xF9\x00\x96\xE9"
-     "\xF9\x00"
-     "\x97\xE9\xF9\x00\x98\xD9\xF9\x00\x99\xD9\xF9\x00\x9A\xA1\xF9\x00\x9B\xA1"
-     "\xF9\x00"
-     "\x9C\xA1\xF9\x00\x9D\x5D\x0B\x00\x9E\x55\x0B\x00\x9F\x4D\x0B\x00\xA0\xD9"
-     "\xC0\x00"
-     "\xA1\xC5\xC2\x00\xA2\x89\xC0\x00\xA3\xA1\xCE\x00\xA4\x95\xC0\x00\xA5\xA9"
-     "\xCE\x00"
-     "\xA6\x0D\xCE\x00\xA7\x8D\xC0\x00\xA8\xB5\xC0\x00\xA9\xE5\xC0\x00\xAA\x31"
-     "\xCE\x00"
-     "\xAB\x31\xCE\x00\xAC\x19\xCE\x00\xAD\x05\xCE\x00\xAE\x01\xCE\x00\xAF\x01"
-     "\xCE\x00"
-     "\xB0\x01\xCE\x00\xB1\x81\xC0\x00\xB2\x99\xC0\x00\xB3\x99\xC0\x00\xB4\x15"
-     "\xCE\x00"
-     "\xB5\x55\xC0\x00\xB6\x5D\xC0\x00\xB7\x4D\xC0\x00\xB8\x75\xC0\x00\xB9\x25"
-     "\xC0\x00"
-     "\xBA\x2D\xC0\x00\xBB\x3D\xC0\x00\xBC\x25\xC0\x00\xBD\xF5\xC1\x00\xBE\xFD"
-     "\xC1\x00"
-     "\x45\xED\x00\x80\xB0\xD5\x0D\x00\xB1\xDD\x0D\x00\xB2\xE9\x0D\x00\xB3\xE9"
-     "\x0D\x00"
-     "\xB4\xF9\x0D\x00\xB5\xF9\x0D\x00\xB6\xE9\x0D\x00\xB7\xE9\x0D\x00\xB8\xD9"
-     "\x0D\x00"
-     "\xB9\x21\x0C\x00\xBA\x2D\x0F\x00\xBB\x25\x0F\x00\xBC\x11\x0E\x00\xBD\x11"
-     "\x0E\x00"
-     "\xBE\x11\x0E\x00\xBF\x11\x0E\x00\x4A\xED\x00\x80\x4F\xED\x00\x80\x54\xED"
-     "\x00\x80"
-     "\x59\xED\x00\x80\x5E\xED\x00\x80\x63\xED\x00\x80\x68\xED\x00\x80\x6D\xED"
-     "\x00\x80"
-     "\x72\xED\x00\x80\x77\xED\x00\x80\x7C\xED\x00\x80\x81\xED\x00\x80\x86\xED"
-     "\x00\x80"
-     "\x8B\xED\x00\x80\x90\xED\x00\x80\x95\xED\x00\x80\x80\xB1\x0E\x00\x81\xB1"
-     "\x0E\x00"
-     "\x82\xC9\x0E\x00\x83\xC9\x0E\x00\x84\xD9\x0E\x00\x85\xD9\x0E\x00\x86\xC9"
-     "\x0E\x00"
-     "\x87\xC9\x0E\x00\x88\xF9\x0E\x00\x89\xF9\x0E\x00\x8A\xC1\x0E\x00\x8B\xC1"
-     "\x0E\x00"
-     "\x8C\xC1\x0E\x00\x8D\xC1\x0E\x00\x8E\xC1\x0E\x00\x8F\xC1\x0E\x00\x90\x3D"
-     "\x01\x00"
-     "\x91\x85\x00\x00\x92\x8D\x00\x00\x93\x85\x00\x00\x9A\xED\x00\x80\x9F\xED"
-     "\x00\x80"
-     "\xA4\xED\x00\x80\xA9\xED\x00\x80\xAE\xED\x00\x80\xB3\xED\x00\x80\xB8\xED"
-     "\x00\x80"
-     "\xBD\xED\x00\x80\xC2\xED\x00\x80\xC7\xED\x00\x80\xCC\xED\x00\x80\xD1\xED"
-     "\x00\x80"
-     "\xB9\x45\xFB\x00\xB8\x45\xFB\x00\xBB\x45\xFB\x00\xBA\x45\xFB\x00\xBD\x45"
-     "\xFB\x00"
-     "\xBC\x45\xFB\x00\xBF\x45\xFB\x00\xBE\x45\xFB\x00\xB1\x05\xFB\x00\xB0\x05"
-     "\x09\x00"
-     "\xB3\x05\xFB\x00\xB2\x05\xFB\x00\xB5\x05\xFB\x00\xB4\x05\xFB\x00\xB7\x05"
-     "\xFB\x00"
-     "\xB6\x05\xFB\x00\xA9\x45\x09\x00\xA8\xB9\x08\x00\xAB\x45\x09\x00\xAA\x45"
-     "\x09\x00"
-     "\xAD\x45\x09\x00\xAC\x45\x09\x00\xAF\x45\x09\x00\xAE\x45\x09\x00\xA1\x45"
-     "\x15\x00"
-     "\xA0\x95\x08\x00\xA3\x05\x15\x00\xA2\x05\x15\x00\xA5\x85\x08\x00\xA4\x55"
-     "\x15\x00"
-     "\xA7\x81\x08\x00\xA6\xA1\xFB\x00\x99\x85\x8B\x02\x98\x35\x93\x02\x9B\xC1"
-     "\x15\x00"
-     "\x9A\x19\x97\x02\x9D\xDD\x15\x00\x9C\x71\x08\x00\x9F\x75\x08\x00\x9E\x75"
-     "\x08\x00"
-     "\x91\x05\x9D\x02\x90\x8D\x92\x02\x93\xE9\x92\x02\x92\x9D\x92\x02\x95\x25"
-     "\x92\x02"
-     "\x94\xB5\xAE\x00\x97\xFD\x8A\x02\x96\xD9\x93\x02\x89\xC9\xAE\x00\x88\xF1"
-     "\x8A\x02"
-     "\x8B\x0D\x9E\x02\x8A\xF1\x8A\x02\x8D\xED\x9F\x02\x8C\x39\x9E\x02\x8F\x21"
-     "\x9F\x02"
-     "\x8E\x4D\x9F\x02\x81\xD5\x9A\x02\x80\x4D\xAD\x00\x83\x19\x9F\x02\x82\x21"
-     "\x99\x02"
-     "\x85\xE1\x9E\x02\x84\xFD\x9E\x02\x87\xC9\x9E\x02\x86\xD1\x9E\x02\xB0\x4D"
-     "\x02\x00"
-     "\xB1\x55\x02\x00\xB2\x5D\x02\x00\xB3\x55\x02\x00\xB4\x4D\x02\x00\xB5\x75"
-     "\x02\x00"
-     "\xB6\x7D\x02\x00\xB7\x75\x02\x00\xB8\x4D\x02\x00\xB9\x55\x02\x00\xBA\x5D"
-     "\x02\x00"
-     "\xD6\xED\x00\x80\xBC\x49\x02\x00\xBD\x49\x02\x00\xBE\xE9\x02\x00\xBF\xE9"
-     "\x02\x00"
-     "\xDB\xED\x00\x80\xE0\xED\x00\x80\xE5\xED\x00\x80\xEA\xED\x00\x80\xEF\xED"
-     "\x00\x80"
-     "\xF4\xED\x00\x80\xF9\xED\x00\x80\xFE\xED\x00\x80\x03\xEE\x00\x80\xBA\x0B"
-     "\x00\x80"
-     "\xBD\x0B\x00\x80\x80\x0B\x00\x80\x98\x0B\x00\x80\x37\x3D\x00\x80\x9E\x0B"
-     "\x00\x80"
-     "\x9B\x0B\x00\x80\x80\x59\x02\x00\x81\x59\x02\x00\x82\x69\x02\x00\x83\x69"
-     "\x02\x00"
-     "\x84\x79\x02\x00\x85\x79\x02\x00\x86\x69\x02\x00\x87\x69\x02\x00\x88\x59"
-     "\x02\x00"
-     "\x89\x59\x02\x00\x8A\x85\x00\x00\xBF\x1C\x00\x80\x8C\xF1\x00\x00\x8D\xF1"
-     "\x00\x00"
-     "\x8E\xF1\x00\x00\x8F\xF1\x00\x00\x90\x91\x00\x00\x91\x91\x00\x00\x92\x91"
-     "\x00\x00"
-     "\xDB\x0A\x00\x80\x94\x8D\x00\x00\x95\xB5\x00\x00\xE7\x0A\x00\x80\xAA\x0B"
-     "\x00\x80"
-     "\xF9\x0A\x00\x80\xF3\x0A\x00\x80\x83\x0B\x00\x80\x86\x0B\x00\x80\x02\x0B"
-     "\x00\x80"
-     "\x53\x0B\x00\x80\xAD\x0B\x00\x80\xB0\x0B\x00\x80\x21\xF1\x00\x80\xBB\x3C"
-     "\x00\x80"
-     "\x17\x0B\x00\x80\x26\xF1\x00\x80\x6B\x0B\x00\x80\x2B\xF1\x00\x80\x2F\x0B"
-     "\x00\x80"
-     "\x32\x0B\x00\x80\x3B\x0B\x00\x80\x30\xF1\x00\x80\x44\x0B\x00\x80\xA7\x0B"
-     "\x00\x80"
-     "\xA1\x0B\x00\x80\x3B\x3D\x00\x80\xA4\x0B\x00\x80\xB7\x22\x00\x80\x6E\x0B"
-     "\x00\x80"
-     "\x92\x0B\x00\x80\x95\x0B\x00\x80\x7D\x0B\x00\x80\x09\x0A\x00\x80\x0C\x0A"
-     "\x00\x80"
-     "\x0F\x0A\x00\x80\x35\xF1\x00\x80\x3A\xF1\x00\x80\x08\xEE\x00\x80\x0D\xEE"
-     "\x00\x80"
-     "\x12\xEE\x00\x80\x17\xEE\x00\x80\x1C\xEE\x00\x80\xA7\xFD\x02\x00\x21\xEE"
-     "\x00\x80"
-     "\x26\xEE\x00\x80\x2B\xEE\x00\x80\x30\xEE\x00\x80\x35\xEE\x00\x80\x3A\xEE"
-     "\x00\x80"
-     "\x3F\xEE\x00\x80\x44\xEE\x00\x80\x49\xEE\x00\x80\x4E\xEE\x00\x80\x53\xEE"
-     "\x00\x80"
-     "\x58\xEE\x00\x80\x5D\xEE\x00\x80\x62\xEE\x00\x80\x67\xEE\x00\x80\x6C\xEE"
-     "\x00\x80"
-     "\x71\xEE\x00\x80\x76\xEE\x00\x80\x7B\xEE\x00\x80\x80\xEE\x00\x80\x85\xEE"
-     "\x00\x80"
-     "\x8A\xEE\x00\x80\x8F\xEE\x00\x80\x94\xEE\x00\x80\x99\xEE\x00\x80\x9E\xEE"
-     "\x00\x80"
-     "\xA3\xEE\x00\x80\xA8\xEE\x00\x80\xAD\xEE\x00\x80\xB2\xEE\x00\x80\xB7\xEE"
-     "\x00\x80"
-     "\xBC\xEE\x00\x80\xC1\xEE\x00\x80\x80\x65\x01\x00\x81\x6D\x01\x00\x82\x65"
-     "\x01\x00"
-     "\x83\x7D\x01\x00\x84\x65\x01\x00\x85\x69\x01\x00\x86\x99\x01\x00\x87\x99"
-     "\x01\x00"
-     "\x88\xA9\x01\x00\x89\xA9\x01\x00\x8A\xB9\x01\x00\x8B\xB9\x01\x00\x8C\xA9"
-     "\x01\x00"
-     "\x8D\xA9\x01\x00\x8E\x99\x01\x00\x8F\x99\x01\x00\x90\xE9\x01\x00\x91\xE9"
-     "\x01\x00"
-     "\x92\xF9\x01\x00\x93\xF9\x01\x00\x94\xE9\x01\x00\x95\xE9\x01\x00\x96\x99"
-     "\x01\x00"
-     "\x97\x99\x01\x00\x98\xA9\x01\x00\x99\xA9\x01\x00\x9A\xB9\x01\x00\x9B\xB9"
-     "\x01\x00"
-     "\x9C\xA9\x01\x00\x9D\xA9\x01\x00\x9E\x99\x01\x00\x9F\x99\x01\x00\xA0\x69"
-     "\x01\x00"
-     "\xA1\x69\x01\x00\xA2\x79\x01\x00\xA3\x79\x01\x00\xA4\x69\x01\x00\xA5\x69"
-     "\x01\x00"
-     "\xA6\x55\x00\x00\xA7\x4D\x00\x00\xA8\x75\x00\x00\xA9\x7D\x00\x00\xAA\x75"
-     "\x00\x00"
-     "\xAB\x4D\x00\x00\xAC\x55\x00\x00\xAD\x5D\x00\x00\xAE\x55\x00\x00\xAF\x4D"
-     "\x00\x00"
-     "\xB0\x35\x00\x00\xB1\x3D\x00\x00\xB2\x35\x00\x00\xC6\xEE\x00\x80\xCB\xEE"
-     "\x00\x80"
-     "\xD0\xEE\x00\x80\xD5\xEE\x00\x80\xDA\xEE\x00\x80\xDF\xEE\x00\x80\xE4\xEE"
-     "\x00\x80"
-     "\xE9\xEE\x00\x80\xEE\xEE\x00\x80\xF3\xEE\x00\x80\xF8\xEE\x00\x80\xFD\xEE"
-     "\x00\x80"
-     "\x02\xEF\x00\x80\x9E\x09\x02\x00\x9F\x09\x02\x00\x9C\x09\x02\x00\x9D\x09"
-     "\x02\x00"
-     "\x9A\xED\x05\x00\x9B\xAD\xF2\x00\x98\xED\x05\x00\x99\xE5\x05\x00\x96\xCD"
-     "\x05\x00"
-     "\x97\xD5\x05\x00\x94\xC1\x05\x00\x95\x7D\x93\x02\x92\xD5\x05\x00\x93\xD9"
-     "\x93\x02"
-     "\x90\xE5\x05\x00\x91\xDD\x05\x00\x8E\x8D\x05\x00\x8F\x9D\x05\x00\x8C\x99"
-     "\x93\x02"
-     "\x8D\xE1\x05\x00\x8A\x2D\x05\x00\x8B\x99\x93\x02\x88\x2D\x05\x00\x89\x25"
-     "\x05\x00"
-     "\x07\xEF\x00\x80\x87\x15\x05\x00\x84\x05\x05\x00\x85\xD9\x93\x02\x82\x19"
-     "\x05\x00"
-     "\x83\xF1\x92\x02\x0C\xEF\x00\x80\x81\x19\x05\x00\x11\xEF\x00\x80\x16\xEF"
-     "\x00\x80"
-     "\x1B\xEF\x00\x80\x20\xEF\x00\x80\xBA\xF1\x02\x00\xF0\xCC\x03\x00\xB8\xC1"
-     "\x02\x00"
-     "\xB9\xC1\x02\x00\xB6\xF1\x02\x00\xB7\xF1\x02\x00\xB4\xE1\x02\x00\xB5\xE1"
-     "\x02\x00"
-     "\xB2\xF1\x02\x00\xB3\xF1\x02\x00\xB0\x85\x02\x00\x25\xEF\x00\x80\xAE\xF9"
-     "\x02\x00"
-     "\xAF\xB9\x90\x02\xAC\xF9\x02\x00\xAD\xF9\x02\x00\xAA\xE5\x90\x02\xAB\xF9"
-     "\x02\x00"
-     "\xA8\x2D\xDA\x00\xA9\xF5\x02\x00\xA6\xC1\x02\x00\xA7\xC1\x02\x00\xA4\xF5"
-     "\x02\x00"
-     "\xA5\x49\x99\x02\xA2\xD1\x93\x02\xA3\xED\x02\x00\xA0\xE9\x02\x00\xA1\xE9"
-     "\x02\x00"
-     "\x82\x60\x02\x00\x99\x20\xFD\x00\x91\x08\x02\x00\x9B\x6C\xE1\x00\x84\x98"
-     "\x02\x00"
-     "\xA7\x9D\x01\x00\x84\x98\x02\x00\x91\x0C\x02\x00\x8C\xE8\x02\x00\x8D\x50"
-     "\x02\x00"
-     "\x97\x95\x01\x00\x91\x50\x02\x00\xBE\xBD\x01\x00\x92\xA0\x02\x00\x2F\xEF"
-     "\x00\x80"
-     "\x8D\x20\x0B\x00\x92\xE8\x02\x00\xB1\x84\x0F\x00\xB2\x74\x08\x00\xB9\x00"
-     "\x3C\x00"
-     "\x96\xDC\x02\x00\xB0\x6D\x00\x00\x2A\xEF\x00\x80\x34\xEF\x00\x80\xBD\x3D"
-     "\x00\x00"
-     "\x91\x4C\x02\x00\xB9\xAC\x3C\x00\xB8\xAC\x3C\x00\x96\xB4\x02\x00\x39\xEF"
-     "\x00\x80"
-     "\xAF\x0D\x00\x00\xBA\x25\x00\x00\xA2\x00\x3C\x00\x91\x1C\x02\x00\xAF\x05"
-     "\x00\x00"
-     "\x3E\xEF\x00\x80\x43\xEF\x00\x80\x48\xEF\x00\x80\xA4\x68\x3D\x00\x4D\xEF"
-     "\x00\x80"
-     "\x52\xEF\x00\x80\x57\xEF\x00\x80\x5C\xEF\x00\x80\x61\xEF\x00\x80\x66\xEF"
-     "\x00\x80"
-     "\x6B\xEF\x00\x80\x70\xEF\x00\x80\x75\xEF\x00\x80\x96\xF4\x02\x00\x7A\xEF"
-     "\x00\x80"
-     "\x7F\xEF\x00\x80\x84\xEF\x00\x80\x89\xEF\x00\x80\x8E\xEF\x00\x80\x93\xEF"
-     "\x00\x80"
-     "\x98\xEF\x00\x80\x9D\xEF\x00\x80\xA2\xEF\x00\x80\xA7\xEF\x00\x80\xAC\xEF"
-     "\x00\x80"
-     "\xB1\xEF\x00\x80\xA5\x60\x15\x00\xB6\xEF\x00\x80\xBB\xEF\x00\x80\xF0\x20"
-     "\x3D\x00"
-     "\xF0\xC4\x3D\x00\xC0\xEF\x00\x80\xC5\xEF\x00\x80\xCA\xEF\x00\x80\xCF\xEF"
-     "\x00\x80"
-     "\xD4\xEF\x00\x80\xD9\xEF\x00\x80\xDE\xEF\x00\x80\xB5\xCC\x03\x00\xF0\x40"
-     "\x02\x00"
-     "\x91\x8C\x02\x00\xA4\xC4\x02\x00\xB0\x0D\x00\x00\xE3\xEF\x00\x80\xB9\xB8"
-     "\x03\x00"
-     "\xE8\xEF\x00\x80\xED\xEF\x00\x80\xF2\xEF\x00\x80\xF7\xEF\x00\x80\xFC\xEF"
-     "\x00\x80"
-     "\x01\xF0\x00\x80\x06\xF0\x00\x80\x0B\xF0\x00\x80\x10\xF0\x00\x80\x15\xF0"
-     "\x00\x80"
-     "\x1A\xF0\x00\x80\x1F\xF0\x00\x80\x24\xF0\x00\x80\x29\xF0\x00\x80\x2E\xF0"
-     "\x00\x80"
-     "\x33\xF0\x00\x80\x38\xF0\x00\x80\x3D\xF0\x00\x80\x42\xF0\x00\x80\xF0\xE4"
-     "\x3D\x00"
-     "\x47\xF0\x00\x80\x4C\xF0\x00\x80\x51\xF0\x00\x80\x56\xF0\x00\x80\x5B\xF0"
-     "\x00\x80"
-     "\x60\xF0\x00\x80\x65\xF0\x00\x80\x6A\xF0\x00\x80\x6F\xF0\x00\x80\x74\xF0"
-     "\x00\x80"
-     "\x79\xF0\x00\x80\x7E\xF0\x00\x80\x83\xF0\x00\x80\x88\xF0\x00\x80\x91\xF0"
-     "\x00\x80"
-     "\x9A\xF0\x00\x80\xA7\xF0\x00\x80\xB4\xF0\x00\x80\xC1\xF0\x00\x80\xCE\xF0"
-     "\x00\x80"
-     "\xBB\x05\x00\x00\xDB\xF0\x00\x80\xFE\xF0\x00\x80\xE4\xF0\x00\x80\xBF\xF5"
-     "\x02\x00"
-     "\xBE\xF5\x02\x00\xBD\x15\x00\x00\xBC\x15\x00\x00\xA0\x91\x3E\x00\xA1\x91"
-     "\x3E\x00"
-     "\xA2\x95\x3E\x00\xA3\xAD\x3E\x00\xA4\xB5\x3E\x00\xA5\xBD\x3E\x00\xA6\xB5"
-     "\x3E\x00"
-     "\xA7\xAD\x3E\x00\xA8\x95\x3E\x00\xA9\x9D\x3E\x00\xAA\x95\x3E\x00\xAB\xE9"
-     "\x3E\x00"
-     "\xAC\xF9\x3E\x00\xAD\xF9\x3E\x00\xAE\xE9\x3E\x00\xAF\xE9\x3E\x00\xB0\x99"
-     "\x3E\x00"
-     "\xB1\x99\x3E\x00\xB2\xA9\x3E\x00\xB3\xA9\x3E\x00\xB4\xB9\x3E\x00\xB5\xB9"
-     "\x3E\x00"
-     "\xB6\xA9\x3E\x00\xB7\xA5\x3E\x00\xB8\x9D\x3E\x00\xB9\x6D\x01\x00\xBA\x65"
-     "\x01\x00"
-     "\xBB\x7D\x01\x00\xBC\x65\x01\x00\xBD\x6D\x01\x00\xBE\x65\x01\x00\xBF\x5D"
-     "\x01\x00"
-     "\x81\xC5\x01\x00\x80\xC5\x01\x00\x83\xC5\x01\x00\x82\xC5\x01\x00\x85\xC5"
-     "\x01\x00"
-     "\x84\xC5\x01\x00\x87\xC5\x01\x00\x86\xC5\x01\x00\x89\xC5\x01\x00\x88\xC5"
-     "\x01\x00"
-     "\x8B\xC5\x01\x00\x8A\xC5\x01\x00\x8D\xC5\x01\x00\x8C\xC5\x01\x00\x8F\xC5"
-     "\x01\x00"
-     "\x8E\xC5\x01\x00\x91\x45\x01\x00\x90\x45\x01\x00\x93\x59\x01\x00\x92\x45"
-     "\x01\x00"
-     "\x95\x49\x01\x00\x94\x41\x01\x00\x97\x79\x01\x00\x96\x41\x01\x00\x99\x49"
-     "\x01\x00"
-     "\x98\x41\x01\x00\x9B\x59\x01\x00\x9A\x41\x01\x00\x9D\x49\x01\x00\x9C\x41"
-     "\x01\x00"
-     "\x9F\x39\x01\x00\x9E\x41\x01\x00\x07\xF1\x00\x80\xF1\xF0\x00\x80\x14\xF1"
-     "\x00\x80"
-     "\x83\x00\x00\x00\x85\x68\x03\x00\x5B\x09\x00\x80\x87\x3C\x02\x00\x86\x10"
-     "\x00\x00"
-     "\xA5\x0A\x00\x80\x80\x2D\x00\x00\xF6\x20\x00\x80\xEE\x20\x00\x80\x80\xB0"
-     "\x03\x00"
-     "\x81\xF4\x10\x00\x3F\xF1\x00\x80\x44\xF1\x00\x80\x91\x40\x05\x00\x90\x00"
-     "\x04\x00"
-     "\x93\x48\x07\x00\x92\x4C\x06\x00\x95\x50\x09\x00\x94\x00\x08\x00\x97\x58"
-     "\x0B\x00"
-     "\x96\x5C\x0A\x00\x99\x60\x0D\x00\x98\x00\x0C\x00\x9B\x68\x0F\x00\x9A\x6C"
-     "\x0E\x00"
-     "\x9D\x74\x12\x00\x9C\x74\x11\x00\x9F\x78\x15\x00\x9E\x00\x14\x00\x49\xF1"
-     "\x00\x80"
-     "\x4E\xF1\x00\x80\x53\xF1\x00\x80\x58\xF1\x00\x80\x5D\xF1\x00\x80\x62\xF1"
-     "\x00\x80"
-     "\x67\xF1\x00\x80\x6C\xF1\x00\x80\xA4\x20\x13\x00\x71\xF1\x00\x80\x76\xF1"
-     "\x00\x80"
-     "\x7B\xF1\x00\x80\x80\xF1\x00\x80\x85\xF1\x00\x80\x8A\xF1\x00\x80\x8F\xF1"
-     "\x00\x80"
-     "\x94\xF1\x00\x80\x99\xF1\x00\x80\x9E\xF1\x00\x80\xA3\xF1\x00\x80\xB8\xD0"
-     "\x16\x00"
-     "\xB9\xDC\x17\x00\xBA\x00\x18\x00\xA8\xF1\x00\x80\xAD\xF1\x00\x80\xB2\xF1"
-     "\x00\x80"
-     "\xB7\xF1\x00\x80\xBC\xF1\x00\x80\xBC\x00\x00\x00\xBD\x00\x00\x00\xC1\xF1"
-     "\x00\x80"
-     "\xC6\xF1\x00\x80\x9E\xC5\x03\x00\x9F\xCD\x03\x00\xCB\xF1\x00\x80\xD0\xF1"
-     "\x00\x80"
-     "\xD5\xF1\x00\x80\xDA\xF1\x00\x80\xDF\xF1\x00\x80\xE4\xF1\x00\x80\xC1\x16"
-     "\x00\x80"
-     "\xFA\x16\x00\x80\xC4\x16\x00\x80\xAF\xAC\x1B\x00\x51\x03\x00\x80\x4E\x03"
-     "\x00\x80"
-     "\x61\x03\x00\x80\x84\x04\x00\x80\xB0\x75\xBF\x02\xB1\x91\xB8\x02\xB2\x91"
-     "\xB8\x02"
-     "\xB3\x75\xBF\x02\xB4\x8D\xB8\x02\xB5\x59\xBE\x02\xB6\x41\xBE\x02\xB7\x7D"
-     "\xBE\x02"
-     "\xB8\x4D\xBE\x02\xB9\x31\xBE\x02\xBA\x05\xBE\x02\xBB\x11\xBE\x02\xBC\xB1"
-     "\xBE\x02"
-     "\xBD\xA5\xBE\x02\xBE\xA1\xBE\x02\xBF\xA1\xBE\x02\x84\x80\x18\x00\x85\x84"
-     "\x1B\x00"
-     "\x86\x0C\x02\x00\x23\x05\x00\x80\x76\x05\x00\x80\xB1\x05\x00\x80\xD7\x05"
-     "\x00\x80"
-     "\xEE\x05\x00\x80\x0B\x06\x00\x80\x22\x06\x00\x80\x73\x24\x00\x80\x7E\x08"
-     "\x00\x80"
-     "\x88\xB0\x1A\x00\x89\xE4\x12\x00\x17\x07\x00\x80\x41\x08\x00\x80\xF0\x00"
-     "\x00\x00"
-     "\x90\x25\x00\x00\x4A\x08\x00\x80\x0F\x08\x00\x80\x7B\x08\x00\x80\x95\x07"
-     "\x00\x80"
-     "\x98\x07\x00\x80\xBC\x07\x00\x80\xB9\x06\x00\x80\x3D\x25\x00\x80\xA4\x35"
-     "\x03\x00"
-     "\x10\x27\x00\x80\xA2\x25\x03\x00\xA3\x2D\x03\x00\xA0\x35\x03\x00\xA1\x2D"
-     "\x03\x00"
-     "\x91\x31\xAC\x02\x90\xA1\xA3\x02\x93\x51\xAC\x02\x92\x4D\xAC\x02\x95\x75"
-     "\xAD\x02"
-     "\x94\x81\xAC\x02\x97\xA9\xAD\x02\x96\x91\xAD\x02\x99\xB5\xA9\x02\x98\xE9"
-     "\xAD\x02"
-     "\x9B\xE5\xA7\x02\x9A\x19\xA4\x02\x9D\x39\xA0\x02\x9C\xC5\xA1\x02\x9F\x15"
-     "\xA0\x02"
-     "\x9E\x2D\xA0\x02\x81\x8D\xA7\x02\x80\x71\xA4\x02\x83\x41\xA0\x02\x82\xBD"
-     "\xA1\x02"
-     "\x85\x7D\xA0\x02\x84\x45\xA0\x02\x87\x61\xA0\x02\x86\x6D\xA0\x02\x89\xD5"
-     "\xA0\x02"
-     "\x88\x31\xA0\x02\x8B\x85\xA0\x02\x8A\xA9\xA0\x02\x8D\xE1\xA1\x02\x8C\x49"
-     "\xA1\x02"
-     "\x8F\x51\xAC\x02\x8E\x85\xA1\x02\xB1\x31\xAD\x02\xB0\xC1\xB4\x02\xB3\x1D"
-     "\xA9\x02"
-     "\xB2\x89\xB5\x02\xB5\x5D\xA7\x02\xB4\xA1\xA4\x02\xB7\x91\xA0\x02\xB6\x6D"
-     "\xA1\x02"
-     "\xB9\x8D\xA0\x02\xB8\xB5\xA0\x02\xBB\x91\xA0\x02\xBA\x9D\xA0\x02\xBD\x05"
-     "\xA0\x02"
-     "\xBC\xE1\xA0\x02\xBF\x55\xA0\x02\xBE\x79\xA0\x02\xA1\xF5\xB4\x02\xA0\xF5"
-     "\xA0\x02"
-     "\xA3\xF5\xB4\x02\xA2\xC5\x90\x00\xA5\x3D\xA0\x02\xA4\x11\xA0\x02\xA7\x49"
-     "\xA1\x02"
-     "\xA6\xE1\xA1\x02\xA9\xC9\xAC\x02\xA8\x1D\xA1\x02\xAB\xD9\xAC\x02\xAA\x49"
-     "\xA3\x02"
-     "\xAD\xF1\x90\x00\xAC\xB5\xAC\x02\xAF\x9D\xAD\x02\xAE\x69\xAC\x02\x80\x3D"
-     "\xA0\x02"
-     "\x81\x95\xA0\x02\x82\xF1\xA0\x02\x83\x25\xAD\x02\x84\xB5\xA2\x02\x85\x25"
-     "\xAD\x02"
-     "\x86\x59\xAD\x02\x87\x45\xAD\x02\x88\xB5\xAD\x02\x89\x41\xAC\x02\x8A\xA5"
-     "\xAC\x02"
-     "\x8B\x9D\xAC\x02\x8C\xFD\xAC\x02\x8D\xA1\xA8\x02\x8E\x0D\xA5\x02\x8F\xF1"
-     "\xA6\x02"
-     "\x90\xB1\xA0\x02\x91\x4D\xA1\x02\x92\x59\xA1\x02\x93\x61\xA1\x02\x94\x61"
-     "\xA1\x02"
-     "\x5D\x2F\x00\x80\x96\x51\x91\x00\x97\x61\xB5\x02\x98\xA5\xA1\x02\x99\x89"
-     "\xA1\x02"
-     "\x9A\x55\xA0\x02\x9B\xFD\xA0\x02\x9C\x89\xA0\x02\x9D\x5D\xAD\x02\x9E\xDD"
-     "\xA2\x02"
-     "\x9F\x4D\xAD\x02\xA0\xC1\xAD\x02\xA1\x85\x91\x00\xA2\x1D\xAD\x02\xA3\xE9"
-     "\xAC\x02"
-     "\xA4\xD5\xB5\x02\xA5\x25\xAC\x02\xA6\x9D\xB4\x02\xA7\x09\xA8\x02\xA8\x95"
-     "\xA5\x02"
-     "\xA9\x69\xA6\x02\xAA\x59\xA0\x02\xAB\xA5\xA1\x02\xAC\xA1\xA1\x02\xAD\x99"
-     "\xA1\x02"
-     "\xAE\x89\xA1\x02\xAF\x85\xA1\x02\xB0\x95\xA1\x02\xB1\x71\xA1\x02\xB2\x0D"
-     "\xA1\x02"
-     "\xB3\x21\xA1\x02\xB4\xED\xA0\x02\xB5\x45\xA0\x02\xB6\x21\xA0\x02\xB7\xF5"
-     "\xAD\x02"
-     "\xB8\x45\xA2\x02\xB9\xD5\xAD\x02\xBA\xA9\xAD\x02\xBB\xB5\xAD\x02\xBC\x65"
-     "\xAD\x02"
-     "\xBD\x91\xAC\x02\xBE\x75\xAC\x02\xBF\x4D\xAC\x02\x80\xCD\xAF\x02\x81\x91"
-     "\xAB\x02"
-     "\x82\x3D\xA6\x02\x83\xC1\xA5\x02\x84\xE1\xA3\x02\x85\x1D\xA2\x02\x86\x09"
-     "\xA2\x02"
-     "\x87\x31\xA2\x02\x88\x11\xA2\x02\x89\x11\xB6\x02\x8A\x21\x92\x00\x8B\x11"
-     "\xB6\x02"
-     "\x8C\xF5\xA2\x02\x8D\xD9\xA2\x02\x8E\x05\xA3\x02\x8F\xAD\xA3\x02\x90\xB9"
-     "\xA3\x02"
-     "\x91\x6D\xAE\x02\x92\xED\xA1\x02\x93\x7D\xAE\x02\x94\x11\xAE\x02\x95\x55"
-     "\x92\x00"
-     "\x96\xCD\xAE\x02\x97\x39\xAF\x02\x98\x25\xB6\x02\x99\xD5\xAF\x02\x9A\x6D"
-     "\xB7\x02"
-     "\x9B\xF9\xAB\x02\x9C\x45\xA6\x02\xE2\x2D\x00\x80\x9E\x89\xA3\x02\x9F\x75"
-     "\xA2\x02"
-     "\xFE\x2E\x00\x80\x67\x38\x00\x80\xA2\xB9\xA2\x02\x2F\x2C\x00\x80\xFB\x31"
-     "\x00\x80"
-     "\xA5\x21\xA2\x02\xA6\x5D\xA2\x02\xE7\x33\x00\x80\x36\x2D\x00\x80\xA9\x35"
-     "\xA3\x02"
-     "\xAA\x51\xA3\x02\xAB\x85\xAE\x02\xAC\x15\xA1\x02\x88\x30\x00\x80\xAE\xF9"
-     "\xAE\x02"
-     "\xAF\xE5\xAE\x02\xB0\x55\xAE\x02\xB1\xA1\xAF\x02\xB2\x45\xAF\x02\xB3\x7D"
-     "\xAF\x02"
-     "\xB4\x1D\xAF\x02\xB5\x41\xAB\x02\xB6\xED\xA6\x02\xB7\x11\xA5\x02\xB8\x11"
-     "\xA3\x02"
-     "\xB9\xED\xA2\x02\xCE\x2C\x00\x80\xBB\xC1\xA2\x02\xD4\x31\x00\x80\xBD\xC1"
-     "\xB6\x02"
-     "\xBE\xF1\x92\x00\xBF\xC1\xB6\x02\x80\xC5\xA3\x02\x81\xE9\xA3\x02\x82\x35"
-     "\xA2\x02"
-     "\x83\x9D\xA2\x02\x22\x2D\x00\x80\x85\x3D\xAF\x02\x86\xBD\xA0\x02\x87\x2D"
-     "\xAF\x02"
-     "\x88\x61\xAF\x02\x89\x25\x93\x00\x8A\xBD\xAF\x02\x8B\x49\xAE\x02\x8C\x75"
-     "\xB7\x02"
-     "\x8D\x85\xAE\x02\x8E\x3D\xB6\x02\x8F\xA9\xAA\x02\x90\x75\xA7\x02\x91\x89"
-     "\xA4\x02"
-     "\x92\xB9\xA2\x02\x93\x45\xA3\x02\x94\x41\xA3\x02\x95\x79\xA3\x02\x96\x69"
-     "\xA3\x02"
-     "\x97\x65\xA3\x02\x98\x35\xA3\x02\x99\xD1\xA3\x02\x9A\xAD\xA3\x02\x9B\x81"
-     "\xA3\x02"
-     "\x9C\x4D\xA2\x02\x9D\xE5\xA2\x02\x9E\x81\xA2\x02\x9F\x55\xAF\x02\xA0\x25"
-     "\xA0\x02"
-     "\xA1\xB5\xAF\x02\xA2\xC9\xAF\x02\xA3\xD5\xAF\x02\xA4\x05\xAF\x02\xA5\xF1"
-     "\xAE\x02"
-     "\xA6\x15\xAE\x02\xA7\x2D\xAE\x02\xA8\x6D\xAE\x02\xA9\x31\xAA\x02\xAA\x9D"
-     "\xA7\x02"
-     "\xAB\x61\xA4\x02\xAC\x41\xA2\x02\xAD\xBD\xA3\x02\xAE\xA9\xA3\x02\xAF\x91"
-     "\xA3\x02"
-     "\xB0\xF1\xA3\x02\xB1\xF1\xB7\x02\xB2\xC1\x93\x00\xB3\xF1\xB7\x02\xB4\x15"
-     "\xA3\x02"
-     "\xB5\x39\xA3\x02\xB6\xE5\xA2\x02\xB7\x4D\xA2\x02\xB8\x19\xA2\x02\xB9\xCD"
-     "\xAF\x02"
-     "\xBA\x4D\xA0\x02\xBB\xDD\xAF\x02\xBC\xB1\xAF\x02\xBD\xF5\x93\x00\xBE\x6D"
-     "\xAF\x02"
-     "\xBF\x99\xAE\x02\x01\x00\x00\x00\x94\xA1\xAF\x02\x97\x51\xA0\x02\x96\x4D"
-     "\xA0\x02"
-     "\x91\x81\xAD\x02\x90\x29\xAD\x02\x93\x31\xA0\x02\x92\xE5\xAD\x02\x4F\x36"
-     "\x00\x80"
-     "\x9C\xE9\xA1\x02\x9F\xE5\xAB\x02\x9E\x19\xA8\x02\x99\x55\xA1\x02\x98\xA1"
-     "\xA0\x02"
-     "\x9B\x89\xA1\x02\x9A\xB1\xA1\x02\x85\x8D\xAB\x02\x84\x71\xA8\x02\x87\x41"
-     "\xAC\x02"
-     "\xC3\x38\x00\x80\x81\xE1\xA1\x02\x80\x11\xB8\x02\x83\xCD\xA5\x02\x82\x59"
-     "\xB9\x02"
-     "\x8D\xD5\xAC\x02\xE2\x2E\x00\x80\x8F\x85\xAC\x02\x8E\xA9\xAC\x02\x89\x5D"
-     "\xAC\x02"
-     "\x88\x65\xAC\x02\x1A\x2E\x00\x80\x8A\x4D\xAC\x02\xB5\x31\xA1\x02\xB4\xC1"
-     "\xB8\x02"
-     "\xB7\x1D\xA5\x02\xB6\x89\xB9\x02\xB1\x91\x9C\x00\xB0\xD5\xA0\x02\xB3\xFD"
-     "\xA1\x02"
-     "\xB2\x09\xA0\x02\xBD\x8D\xAC\x02\xBC\xB5\xAC\x02\x9B\x33\x00\x80\xBE\x9D"
-     "\xAC\x02"
-     "\xB9\x7D\xAB\x02\xB8\x81\xA8\x02\xBB\xB1\xAC\x02\x38\x31\x00\x80\xA5\xF5"
-     "\xB8\x02"
-     "\xA4\xF5\xAC\x02\xA7\xF5\xB8\x02\xA6\xC5\x9C\x00\xA1\xD9\xAC\x02\xA0\x25"
-     "\xAD\x02"
-     "\xA3\xF5\xAC\x02\xA2\xCD\xAC\x02\xAD\xC9\xA0\x02\xAC\x1D\xAD\x02\xAF\xD9"
-     "\xA0\x02"
-     "\xAE\x49\xAF\x02\xA9\x1D\xAC\x02\xA8\x31\xAC\x02\xAB\x69\xAD\x02\xAA\xC1"
-     "\xAD\x02"
-     "\x80\x55\xAD\x02\x81\xB1\xAD\x02\x82\xCD\xAD\x02\x83\xE1\xAD\x02\x84\x2D"
-     "\xAC\x02"
-     "\x64\x31\x00\x80\x86\xE1\xAC\x02\x43\x39\x00\x80\x58\x31\x00\x80\x30\x31"
-     "\x00\x80"
-     "\x8A\x69\xA1\x02\x8B\x75\xA1\x02\x8C\xA5\xA1\x02\x8D\x51\xA0\x02\x8E\xB5"
-     "\xA0\x02"
-     "\x8F\x8D\xA0\x02\x90\x8D\xA0\x02\xA7\x35\x00\x80\x92\x7D\xA9\x02\x93\x81"
-     "\xAA\x02"
-     "\x94\xA1\xAC\x02\x95\x5D\xAD\x02\x96\x49\xAD\x02\x97\x71\xAD\x02\x98\x51"
-     "\xAD\x02"
-     "\x99\x51\xB9\x02\x9A\x61\x9D\x00\x9B\x51\xB9\x02\x9C\xB5\xAD\x02\x9D\x99"
-     "\xAD\x02"
-     "\x9E\x45\xAC\x02\x9F\xED\xAC\x02\xA0\x79\xAC\x02\xA1\xAD\xA1\x02\xA2\x2D"
-     "\xAE\x02"
-     "\xA3\xBD\xA1\x02\xA4\xD1\xA1\x02\xA5\x95\x9D\x00\xA6\x0D\xA1\x02\xA7\xF9"
-     "\xA0\x02"
-     "\xA8\xE5\xB9\x02\xA9\x15\xA0\x02\xAA\xAD\xB8\x02\xAB\x39\xA4\x02\xAC\x85"
-     "\xA9\x02"
-     "\xAD\x79\xAA\x02\xAE\x49\xAC\x02\xAF\xB5\xAD\x02\xB0\xD1\xAD\x02\xB1\xE9"
-     "\xAD\x02"
-     "\xB2\xF9\xAD\x02\xB3\xF5\xAD\x02\xB4\x85\xAD\x02\xB5\x61\xAD\x02\xB6\x1D"
-     "\xAD\x02"
-     "\xB7\x31\xAD\x02\xB8\xDD\xAC\x02\xB9\x75\xAC\x02\xBA\x11\xAC\x02\xBB\xC5"
-     "\xA1\x02"
-     "\xBC\x55\xAE\x02\xBD\xC5\xA1\x02\xBE\xB9\xA1\x02\xBF\xA5\xA1\x02\x80\x95"
-     "\xA2\x02"
-     "\x81\x61\xA3\x02\x82\x85\xA3\x02\x83\xBD\xA3\x02\x84\xDD\xA3\x02\x85\x81"
-     "\xA7\x02"
-     "\x86\x2D\xAA\x02\x87\xD1\xA9\x02\x88\xD1\xAF\x02\x89\x2D\xAE\x02\x8A\x39"
-     "\xAE\x02"
-     "\x8B\x01\xAE\x02\x8C\x01\xAE\x02\x8D\x01\xBA\x02\x8E\x31\x9E\x00\x8F\x01"
-     "\xBA\x02"
-     "\x90\x85\xAE\x02\x91\xA9\xAE\x02\x92\x75\xAF\x02\x93\xDD\xAF\x02\x94\xA9"
-     "\xAF\x02"
-     "\x95\x7D\xA2\x02\x96\xFD\xAD\x02\x97\x6D\xA2\x02\x98\x21\xA2\x02\x99\x65"
-     "\x9E\x00"
-     "\x9A\xFD\xA2\x02\x9B\x09\xA3\x02\x9C\x35\xBA\x02\x9D\xC5\xA3\x02\x9E\x7D"
-     "\xBB\x02"
-     "\x9F\xE9\xA7\x02\xA0\xB5\xAA\x02\xA1\x49\xA9\x02\xA2\x79\xAF\x02\xA3\x85"
-     "\xAE\x02"
-     "\xA4\x81\xAE\x02\xA5\xB9\xAE\x02\xA6\xA9\xAE\x02\xA7\xA5\xAE\x02\xA8\xF5"
-     "\xAE\x02"
-     "\xA9\x11\xAE\x02\xAA\x6D\xAE\x02\xAB\x41\xAE\x02\xAC\x8D\xAF\x02\xAD\x25"
-     "\xAF\x02"
-     "\xAE\x41\xAF\x02\xAF\x95\xA2\x02\xB0\x65\xAD\x02\xB1\xF5\xA2\x02\xB2\x89"
-     "\xA2\x02"
-     "\xB3\x95\xA2\x02\xB4\x45\xA2\x02\xB5\xB1\xA3\x02\xB6\x55\xA3\x02\xB7\x6D"
-     "\xA3\x02"
-     "\xB8\x2D\xA3\x02\xB9\x71\xA7\x02\xBA\xDD\xAA\x02\xBB\x21\xA9\x02\xBC\x01"
-     "\xAF\x02"
-     "\xBD\xFD\xAE\x02\xBE\xE9\xAE\x02\xBF\xD1\xAE\x02\x80\x31\xAF\x02\x81\x31"
-     "\xBB\x02"
-     "\x82\x01\x9F\x00\x83\x31\xBB\x02\x84\xD5\xAF\x02\x85\xF9\xAF\x02\x86\x25"
-     "\xAE\x02"
-     "\x87\x8D\xAE\x02\x88\xD9\xAE\x02\x89\x0D\xA3\x02\x8A\x8D\xAC\x02\x8B\x1D"
-     "\xA3\x02"
-     "\x8C\x71\xA3\x02\x8D\x35\x9F\x00\x8E\xAD\xA3\x02\x8F\x59\xA2\x02\x90\x05"
-     "\xBB\x02"
-     "\x91\xF5\xA2\x02\x92\x4D\xBA\x02\x93\xD9\xA6\x02\x94\x65\xAB\x02\x95\x99"
-     "\xA8\x02"
-     "\x96\xA9\xAE\x02\x97\x55\xAF\x02\x98\x71\xAF\x02\x99\x49\xAF\x02\x9A\x59"
-     "\xAF\x02"
-     "\x9B\x55\xAF\x02\x9C\x25\xAF\x02\x9D\xC1\xAF\x02\x9E\xBD\xAF\x02\x9F\x91"
-     "\xAF\x02"
-     "\xA0\xBD\xAE\x02\xA1\x15\xAE\x02\xA2\x71\xAE\x02\xA3\xA5\xA3\x02\xA4\x35"
-     "\xAC\x02"
-     "\xA5\xA5\xA3\x02\xA6\xD9\xA3\x02\xA7\xC5\xA3\x02\xA8\x35\xA3\x02\xA9\xC1"
-     "\xA2\x02"
-     "\xAA\x25\xA2\x02\xAB\x1D\xA2\x02\xAC\x7D\xA2\x02\xAD\x21\xA6\x02\xAE\x8D"
-     "\xAB\x02"
-     "\xAF\x71\xA8\x02\xB0\x31\xAE\x02\xB1\xCD\xAF\x02\xB2\xD9\xAF\x02\xB3\xE1"
-     "\xAF\x02"
-     "\xB4\xE1\xAF\x02\xB5\xE1\xBB\x02\xB6\xD1\x9F\x00\xB7\xE1\xBB\x02\xB8\x25"
-     "\xAF\x02"
-     "\xB9\x09\xAF\x02\xBA\xD5\xAE\x02\xBB\x7D\xAE\x02\xBC\x09\xAE\x02\xBD\xDD"
-     "\xA3\x02"
-     "\xBE\x5D\xAC\x02\xBF\xCD\xA3\x02\x99\x31\xA4\x02\x98\xA1\xAB\x02\x9B\x51"
-     "\xA4\x02"
-     "\x9A\x4D\xA4\x02\x9D\x75\xA5\x02\x9C\x81\xA4\x02\x9F\xA9\xA5\x02\x9E\x91"
-     "\xA5\x02"
-     "\x91\x95\xA8\x02\x90\x71\xA8\x02\x93\xC5\xA8\x02\x92\xE9\xA8\x02\x95\xA1"
-     "\xA9\x02"
-     "\x94\x09\xA9\x02\x97\x11\xA4\x02\x96\xC5\xA9\x02\x89\x8D\xAF\x02\x88\x71"
-     "\xAC\x02"
-     "\x8B\x41\xA8\x02\x8A\xBD\xA9\x02\x8D\x7D\xA8\x02\x8C\x45\xA8\x02\x8F\x61"
-     "\xA8\x02"
-     "\x8E\x6D\xA8\x02\x81\x61\x98\x00\x80\x25\xA4\x02\x83\x0D\xA5\x02\x82\xF9"
-     "\xA4\x02"
-     "\x85\xC1\xA5\x02\x84\x31\xBC\x02\x87\xED\xA1\x02\x86\x79\xBD\x02\xB9\x31"
-     "\xA5\x02"
-     "\xB8\xC1\xBC\x02\xBB\x1D\xA1\x02\xBA\x89\xBD\x02\xBD\x5D\xAF\x02\xBC\xA1"
-     "\xAC\x02"
-     "\xBF\x91\xA8\x02\xBE\x6D\xA9\x02\xB1\x89\xA4\x02\xB0\x5D\xA9\x02\xB3\x99"
-     "\xA4\x02"
-     "\xB2\x09\xAB\x02\xB5\xB1\x98\x00\xB4\xF5\xA4\x02\xB7\xDD\xA5\x02\xB6\x29"
-     "\xA4\x02"
-     "\xA9\xF5\xBC\x02\xA8\xF5\xA8\x02\xAB\xF5\xBC\x02\xAA\xC5\x98\x00\xAD\x3D"
-     "\xA8\x02"
-     "\xAC\x11\xA8\x02\xAF\x49\xA9\x02\xAE\xE1\xA9\x02\xA1\x75\xA1\x02\xA0\x29"
-     "\xA5\x02"
-     "\xA3\x25\xAF\x02\xA2\xD9\xAC\x02\xA5\xF9\xA8\x02\xA4\x05\xA9\x02\xA7\xD5"
-     "\xA8\x02"
-     "\xA6\xED\xA8\x02\x80\x11\xA9\x02\x81\x29\xA9\x02\x82\x39\xA9\x02\x83\x35"
-     "\xA9\x02"
-     "\x84\x45\xA9\x02\x85\xA1\xA9\x02\x86\xDD\xA9\x02\x87\xF1\xA9\x02\x88\x1D"
-     "\xA8\x02"
-     "\x89\xB5\xA8\x02\x8A\xD1\xA8\x02\x8B\x05\xA5\x02\x8C\x95\xAA\x02\x8D\x05"
-     "\xA5\x02"
-     "\x8E\x79\xA5\x02\x8F\x65\xA5\x02\x90\xD5\xA5\x02\x91\x21\xA4\x02\x92\xC5"
-     "\xA4\x02"
-     "\x93\xFD\xA4\x02\x94\x9D\xA4\x02\x95\xC1\xA0\x02\x96\x6D\xAD\x02\x97\x91"
-     "\xAE\x02"
-     "\x98\x91\xA8\x02\x99\x6D\xA9\x02\x9A\x79\xA9\x02\x9B\x41\xA9\x02\x9C\x41"
-     "\xA9\x02"
-     "\x9D\x41\xBD\x02\x9E\x71\x99\x00\x9F\x41\xBD\x02\xA0\x45\xA9\x02\xA1\x69"
-     "\xA9\x02"
-     "\xA2\xB5\xA8\x02\xA3\x1D\xA8\x02\xA4\x69\xA8\x02\xA5\xBD\xA5\x02\xA6\x3D"
-     "\xAA\x02"
-     "\xA7\xAD\xA5\x02\xA8\xE1\xA5\x02\xA9\xA5\x99\x00\xAA\x3D\xA5\x02\xAB\xC9"
-     "\xA4\x02"
-     "\xAC\xF5\xBD\x02\xAD\x05\xA4\x02\xAE\xBD\xBC\x02\xAF\x29\xA0\x02\xB0\xF5"
-     "\xAD\x02"
-     "\xB1\x09\xAE\x02\xB2\x39\xA8\x02\xB3\xC5\xA9\x02\xB4\xC1\xA9\x02\xB5\xF9"
-     "\xA9\x02"
-     "\xB6\xE9\xA9\x02\xB7\xE5\xA9\x02\xB8\xB5\xA9\x02\xB9\x51\xA9\x02\xBA\x2D"
-     "\xA9\x02"
-     "\xBB\x01\xA9\x02\xBC\xCD\xA8\x02\xBD\x65\xA8\x02\xBE\x01\xA8\x02\xBF\xD5"
-     "\xA5\x02"
-     "\x80\xA5\xA9\x02\x81\x35\xA6\x02\x82\x49\xA6\x02\x83\x55\xA6\x02\x84\x85"
-     "\xA6\x02"
-     "\x85\x71\xA7\x02\x86\x95\xA7\x02\x87\xAD\xA7\x02\x88\xED\xA7\x02\x89\xB1"
-     "\xA3\x02"
-     "\x8A\x1D\xAE\x02\x8B\xE1\xAD\x02\x8C\xC1\xAB\x02\x8D\x3D\xAA\x02\x8E\x29"
-     "\xAA\x02"
-     "\x8F\x11\xAA\x02\x90\x71\xAA\x02\x91\x71\xBE\x02\x92\x41\x9A\x00\x93\x71"
-     "\xBE\x02"
-     "\x94\x95\xAA\x02\x95\xB9\xAA\x02\x96\x65\xAB\x02\x97\xCD\xAB\x02\x98\x99"
-     "\xAB\x02"
-     "\x99\x4D\xA6\x02\x9A\xCD\xA9\x02\x9B\x5D\xA6\x02\x9C\x31\xA6\x02\x9D\x75"
-     "\x9A\x00"
-     "\x9E\xED\xA6\x02\x9F\x19\xA7\x02\xA0\xC5\xBE\x02\xA1\x35\xA7\x02\xA2\x8D"
-     "\xBF\x02"
-     "\xA3\x19\xA3\x02\xA4\x85\x0C\x00\xA5\xB5\x0C\x00\xE3\x35\x00\x80\x06\x2E"
-     "\x00\x80"
-     "\xA8\xF9\xBF\x02\xA9\xB1\xBF\x02\xAA\xB5\xBF\x02\xAB\x95\xBF\x02\xAC\x79"
-     "\xBA\x02"
-     "\xAD\x35\xBF\x02\xAE\xA1\xBB\x02\xAF\x3D\xBC\x02\xB0\x49\xBB\x02\xB1\x9D"
-     "\xBD\x02"
-     "\xB2\xA9\xBD\x02\xB3\x61\xA5\x02\xB4\xB9\xBD\x02\xB5\xA9\xBD\x02\xB6\xD5"
-     "\xB6\x02"
-     "\xB7\x69\xBD\x02\xB8\xED\xBA\x02\xB9\x65\xBC\x02\xBA\x3D\xBD\x02\xBB\x09"
-     "\xBD\x02"
-     "\xBC\xB9\xB6\x02\xBD\x5D\xBA\x02\xBE\x01\xBA\x02\xBF\xDD\xBB\x02\x80\x75"
-     "\xB7\x02"
-     "\x81\x2D\x0D\x00\x82\x01\x1C\x00\x83\x19\xBE\x02\x84\x0D\xBE\x02\x85\x2D"
-     "\xBE\x02"
-     "\x86\xD1\xBB\x02\x87\x9D\xBE\x02\x88\x21\x1F\x00\x89\xA5\xBD\x02\x8A\x71"
-     "\x97\x00"
-     "\x8B\x75\xBC\x02\x8C\x51\xBC\x02\x8D\x99\xA4\x02\x8E\x51\xBC\x02\x8F\x41"
-     "\xBC\x02"
-     "\x90\x4D\xB7\x02\x91\xF1\xBC\x02\x92\x51\xB3\x02\x93\x99\xBC\x02\x94\x85"
-     "\xBC\x02"
-     "\x95\xB1\xBC\x02\x96\x51\x1F\x00\x97\xF5\xBB\x02\x98\x99\xBB\x02\x99\x45"
-     "\xBA\x02"
-     "\x9A\x79\x1F\x00\x9B\x41\x0D\x00\x9C\xB9\xBB\x02\x9D\xF5\xBD\x02\x9E\x21"
-     "\xBC\x02"
-     "\x9F\xD5\xBB\x02\xA0\x99\xB3\x02\xA1\x31\xBC\x02\xA2\xD1\xBE\x02\xA3\x99"
-     "\xBE\x02"
-     "\xA4\x8D\xBE\x02\xA5\xAD\xBE\x02\xA6\x51\xBB\x02\xA7\x1D\xBE\x02\xA8\xB9"
-     "\xBA\x02"
-     "\xA9\x25\xBD\x02\xAA\x21\xBA\x02\xAB\xF5\xBC\x02\xAC\xD1\xBC\x02\xAD\x19"
-     "\xA4\x02"
-     "\xAE\xD1\xBC\x02\xAF\xC1\xBC\x02\xB0\xCD\xB7\x02\xB1\x71\xBC\x02\xB2\xC5"
-     "\xBB\x02"
-     "\xB3\x4D\xBD\x02\xB4\x05\xBC\x02\xB5\x31\xBC\x02\xB6\x91\xB7\x02\xB7\x75"
-     "\xBB\x02"
-     "\xB8\x19\xBB\x02\xB9\xC5\xBA\x02\xBA\x9D\xB7\x02\xBB\xC5\x0D\x00\xBC\xF9"
-     "\x1C\x00"
-     "\xBD\xE1\xBE\x02\xBE\xE5\xBE\x02\xBF\xC5\xBE\x02\x04\x0C\x00\x80\x0F\x00"
-     "\x00\x00"
-     "\x4B\x0C\x00\x80\x7B\x33\x00\x80\x57\x0C\x00\x80\xCF\x34\x00\x80\x3A\x2D"
-     "\x00\x80"
-     "\x66\x2E\x00\x80\x7B\x39\x00\x80\xB7\x24\x00\x80\x7B\x24\x00\x80\x85\x24"
-     "\x00\x80"
-     "\x99\x24\x00\x80\xC1\x24\x00\x80\xCC\x18\xAB\x02\xCD\x7C\xAB\x02\xA0\xB9"
-     "\xAD\x02"
-     "\xA1\x45\xAE\x02\xA2\x4D\xAE\x02\xA3\x2D\x90\x02\xA4\x5D\xAE\x02\xA5\x75"
-     "\xAE\x02"
-     "\xA6\x59\x90\x02\xA7\xB1\xAE\x02\xA8\x4D\xAC\x02\xA9\xED\xAB\x02\xAA\xA1"
-     "\xAB\x02"
-     "\xAB\x01\xAF\x02\xAC\xCD\xDA\x00\xAD\x09\xAF\x02\xA3\x24\x00\x80\xCB\x24"
-     "\x00\x80"
-     "\x90\x49\xAF\x02\x91\xB1\xAC\x02\x92\x5D\xAB\x02\x93\x55\xAB\x02\x94\xA9"
-     "\xAC\x02"
-     "\x95\x49\xAB\x02\x96\x95\xAD\x02\x97\x85\xAD\x02\x98\x81\xAD\x02\x99\x89"
-     "\xAD\x02"
-     "\x9A\xFD\xAD\x02\x9B\xC1\xAD\x02\x9C\x71\xAD\x02\x9D\x65\xAD\x02\x9E\x65"
-     "\xAD\x02"
-     "\x9F\x55\xAD\x02\x80\x21\xAD\x02\x81\x2D\xAD\x02\x82\x31\xAD\x02\x83\xCD"
-     "\xAE\x02"
-     "\x84\xD5\xAE\x02\x85\xB5\x90\x02\x86\xD5\xAE\x02\x87\xDD\x90\x02\x88\xE5"
-     "\x90\x02"
-     "\x89\xD1\xAE\x02\x8A\xCD\xDB\x00\x8B\xF9\x90\x02\x8C\xDD\xAC\x02\x8D\x65"
-     "\xAB\x02"
-     "\x8E\xD9\x90\x02\x8F\x89\xAF\x02\x80\xC9\xA5\x02\x81\x85\xA0\x02\x82\x09"
-     "\x01\x00"
-     "\x83\x8D\xA3\x02\x84\x49\x89\x00\x85\x4D\xA2\x02\x86\x79\xA2\x02\x87\xB1"
-     "\xBA\x02"
-     "\x88\x49\xA2\x02\x89\x59\xA2\x02\x8A\x25\xA9\x02\x8B\x99\xA2\x02\x8C\x29"
-     "\xAD\x02"
-     "\x8D\xE1\xA2\x02\x8E\xED\xA2\x02\x8F\xD9\xA2\x02\x90\x49\x01\x00\x91\xED"
-     "\xA5\x02"
-     "\x92\xB1\xA5\x02\x93\x6D\xA4\x02\x94\x41\x01\x00\x95\x79\x13\x00\x96\x91"
-     "\xA5\x02"
-     "\x97\xDD\xA3\x02\x98\x39\xA2\x02\x99\xCD\xA5\x02\x9A\x71\xAD\x02\x9B\xD9"
-     "\xA2\x02"
-     "\x9C\x29\xA0\x02\x9D\x61\xA0\x02\x9E\x65\xA0\x02\x9F\x45\xA0\x02\xA0\x49"
-     "\xA5\x02"
-     "\xA1\x05\xA0\x02\xA2\x91\xA4\x02\xA3\x0D\xA3\x02\xA4\x19\xA4\x02\xA5\xCD"
-     "\xA2\x02"
-     "\xA6\xF9\xA2\x02\xA7\x31\xBA\x02\xA8\xC9\xA2\x02\xA9\xD9\xA2\x02\xAA\xA5"
-     "\xA9\x02"
-     "\xAB\x19\xA2\x02\xAC\xBD\xA5\x02\xAD\x35\xA3\x02\xAE\x6D\xA2\x02\xAF\x59"
-     "\xA2\x02"
-     "\xB0\x89\xA9\x02\xB1\x6D\xA5\x02\xB2\x31\xA5\x02\xB3\xED\xA4\x02\xB4\xA5"
-     "\xA9\x02"
-     "\xB5\xFD\x13\x00\xB6\xD1\x02\x00\xB7\xC9\xA0\x02\xB8\xFD\xA0\x02\xB9\xDD"
-     "\xA0\x02"
-     "\xBA\x21\xA5\x02\xBB\x6D\xA0\x02\xBC\xF1\x01\x00\xBD\x75\xA3\x02\xBE\xA1"
-     "\x89\x00"
-     "\xBF\xA5\xA2\x02\x81\xAD\xB9\x02\x80\x65\xA1\x02\x83\x75\xA1\x02\x82\x65"
-     "\xA1\x02"
-     "\x85\xA5\xA1\x02\x84\x19\xAA\x02\x87\xCD\xA1\x02\x86\x05\xAE\x02\x89\xC5"
-     "\xA1\x02"
-     "\x88\xF1\xA1\x02\x8B\x81\xA6\x02\x8A\x25\x02\x00\x8D\x11\xA7\x02\x8C\xCD"
-     "\xA6\x02"
-     "\x8F\x15\x10\x00\x8E\x2D\x02\x00\x91\xC1\xA0\x02\x90\x8D\xA6\x02\x93\xE1"
-     "\xA6\x02"
-     "\x92\x15\xA1\x02\x95\xE5\xA1\x02\x94\x4D\xAE\x02\x97\x4D\xA3\x02\x96\x05"
-     "\xA3\x02"
-     "\x99\x59\xA3\x02\x98\x79\xA3\x02\x9B\xE9\xA3\x02\x9A\xA5\xA6\x02\x9D\xF1"
-     "\xA0\x02"
-     "\x9C\x6D\xA7\x02\x9F\x21\xA1\x02\x9E\xF5\xA7\x02\xA1\x2D\xB9\x02\xA0\xE5"
-     "\xA1\x02"
-     "\xA3\xF5\xA1\x02\xA2\xE5\xA1\x02\xA5\x25\xA1\x02\xA4\x99\xAA\x02\xA7\x19"
-     "\xA0\x02"
-     "\xA6\x91\xA6\x02\xA9\x45\xA1\x02\xA8\x71\xA1\x02\xAB\x01\xA6\x02\xAA\xE5"
-     "\xAA\x02"
-     "\xAD\x91\xA7\x02\xAC\x4D\xA6\x02\xAF\x91\x10\x00\xAE\xC9\xAA\x02\xB1\xD5"
-     "\xA3\x02"
-     "\xB0\xCD\x01\x00\xB3\xF1\xA3\x02\xB2\xD1\xA3\x02\xB5\x51\xA3\x02\xB4\x1D"
-     "\xA6\x02"
-     "\xB7\x59\xA0\x02\xB6\xDD\x02\x00\xB9\xB9\xA1\x02\xB8\xBD\x8A\x00\xBB\x45"
-     "\xB9\x02"
-     "\xBA\x8D\xA1\x02\xBD\x8D\xA1\x02\xBC\x9D\xA1\x02\xBF\x4D\xA1\x02\xBE\xF1"
-     "\xAA\x02"
-     "\xA0\x19\x10\x00\xA1\x19\x10\x00\xF2\x0B\x00\x80\x8F\x24\x00\x80\x90\x0D"
-     "\x00\x00"
-     "\x91\x3D\x00\x00\xAD\x24\x00\x80\x8C\x30\x00\x80\xCC\xA8\xAB\x02\xCD\xDC"
-     "\xAB\x02"
-     "\xF6\x2D\x00\x80\x03\x2C\x00\x80\xCC\xB0\xA8\x02\xCD\xAC\xA8\x02\xCC\xB8"
-     "\xAB\x02"
-     "\xCD\x8C\xAB\x02\x84\x75\x03\x00\x85\x3D\x03\x00\x86\x35\x03\x00\x87\x41"
-     "\x00\x00"
-     "\x80\x75\x03\x00\x81\x7D\x03\x00\x82\x75\x03\x00\x83\x6D\x03\x00\xFB\x2B"
-     "\x00\x80"
-     "\x0B\x2C\x00\x80\xE9\xF1\x00\x80\x56\x2C\x00\x80\x88\x69\x00\x00\x6A\x2C"
-     "\x00\x80"
-     "\x72\x2C\x00\x80\x7A\x2C\x00\x80\x80\xB9\x11\x00\x81\xB9\x11\x00\x82\x09"
-     "\x11\x00"
-     "\x83\x09\x11\x00\x84\x19\x11\x00\x85\x19\x11\x00\x86\x09\x11\x00\x87\x09"
-     "\x11\x00"
-     "\x88\x39\x11\x00\x89\x39\x11\x00\x8A\x0D\x11\x00\x8B\x05\x11\x00\x8C\x1D"
-     "\x11\x00"
-     "\x8D\x05\x11\x00\x8E\x0D\x11\x00\x8F\x05\x11\x00\x90\x7D\x11\x00\x91\x05"
-     "\x11\x00"
-     "\x92\x0D\x11\x00\x93\x05\x11\x00\x94\x1D\x11\x00\x95\x09\x11\x00\x96\x39"
-     "\x11\x00"
-     "\x97\x39\x11\x00\x98\x09\x11\x00\x99\x09\x11\x00\x9A\x11\x11\x00\x9B\x11"
-     "\x11\x00"
-     "\x9C\xF9\x10\x00\x9D\xF9\x10\x00\x9E\xE9\x10\x00\x9F\xE9\x10\x00\x9F\xB5"
-     "\xA7\x02"
-     "\x9E\x09\xAC\x02\x9D\x75\xA7\x02\x9C\x65\xA7\x02\x9B\xBD\xBF\x02\x9A\x75"
-     "\xA7\x02"
-     "\x99\x41\xA7\x02\x98\x95\xA1\x02\x97\xA1\xA6\x02\x96\x3D\xA1\x02\x95\xA9"
-     "\xA5\x02"
-     "\x94\xE5\xA0\x02\x93\x09\xA5\x02\x92\x29\xA5\x02\x91\x2D\xA5\x02\x90\x65"
-     "\xA5\x02"
-     "\x8F\xF5\xA7\x02\x8E\x5D\xA8\x02\x8D\xE1\xA0\x02\x8C\x15\xA7\x02\x8B\xD1"
-     "\xA6\x02"
-     "\x8A\x9D\xA0\x02\x89\x75\x16\x00\x88\x4D\x04\x00\x87\x41\xA1\x02\x86\x9D"
-     "\xA0\x02"
-     "\x85\xC1\xA0\x02\x84\x65\x04\x00\x83\x95\xA7\x02\x82\xA1\xA7\x02\x81\xAD"
-     "\xA7\x02"
-     "\x80\x65\xA8\x02\xBF\x29\xA0\x02\xBE\x8D\x04\x00\xBD\x6D\xA7\x02\xBC\x59"
-     "\xA7\x02"
-     "\xBB\x45\xA7\x02\xBA\x8D\xA8\x02\xB9\x2D\xA7\x02\xB8\x91\xAC\x02\xB7\xDD"
-     "\xA7\x02"
-     "\xB6\xCD\xA7\x02\xB5\x05\xBF\x02\xB4\xCD\xA7\x02\xB3\xE9\xA7\x02\xB2\xED"
-     "\x8C\x00"
-     "\xB1\x39\xA6\x02\xB0\xBD\x04\x00\xAF\x41\xA5\x02\xAE\x0D\xA0\x02\xAD\xF1"
-     "\xA5\x02"
-     "\xAC\xD1\xA5\x02\xAB\xC5\xA5\x02\xAA\xDD\x07\x00\xA9\xF1\x16\x00\xA8\xA9"
-     "\xAC\x02"
-     "\xA7\xC1\xA1\x02\xA6\x1D\xA0\x02\xA5\x41\xA0\x02\xA4\xA5\xAC\x02\xA3\x15"
-     "\xA7\x02"
-     "\xA2\x21\xA7\x02\xA1\x79\xA6\x02\xA0\xF1\xA0\x02\x80\xF9\xA1\x02\x81\x25"
-     "\xA0\x02"
-     "\x82\x19\x05\x00\x83\x21\x17\x00\x84\xD9\xA1\x02\x85\x95\xA7\x02\x86\x41"
-     "\xA6\x02"
-     "\x87\xB5\xA1\x02\x88\x39\xA9\x02\x89\x91\xA6\x02\x8A\xF5\xAF\x02\x8B\xF1"
-     "\xAF\x02"
-     "\x7E\x2C\x00\x80\x82\x2C\x00\x80\x8E\x09\x8C\x00\x8F\xE9\xBE\x02\x90\xC9"
-     "\xBE\x02"
-     "\x91\xC9\xBE\x02\x92\x7D\x8C\x00\x93\x75\x8C\x00\x94\x6D\x8C\x00\x95\x15"
-     "\x8C\x00"
-     "\x96\x11\x8C\x00\x97\x0D\x8C\x00\x98\x51\x8C\x00\x99\xB1\xBE\x02\x9A\xE1"
-     "\xBE\x02"
-     "\x9B\xE1\xBE\x02\x9C\x45\x8C\x00\x9D\x4D\x8C\x00\x9E\x45\x8C\x00\x9F\x3D"
-     "\x8C\x00"
-     "\xA0\xC9\x8C\x00\xA1\xD5\x8C\x00\xA2\xB9\x8C\x00\xA3\x59\xBE\x02\xA4\x19"
-     "\xBE\x02"
-     "\xA5\x19\xBE\x02\xA6\xAD\x8C\x00\xA7\xA5\x8C\x00\xA8\x9D\x8C\x00\xA9\xE5"
-     "\x8C\x00"
-     "\xAA\xE1\x8C\x00\xAB\xFD\x8C\x00\xAC\x81\x8C\x00\xAD\x61\xBE\x02\xAE\x31"
-     "\xBE\x02"
-     "\xAF\x31\xBE\x02\xB0\xF5\x8C\x00\xB1\xFD\x8C\x00\xB2\xF5\x8C\x00\xB3\x8D"
-     "\x8C\x00"
-     "\xB4\x99\x8C\x00\xB5\x85\x8C\x00\xB6\xE9\x8C\x00\xB7\x09\xBE\x02\xB8\x69"
-     "\xBE\x02"
-     "\xB9\x69\xBE\x02\xBA\xDD\x8C\x00\xBB\xD5\x8C\x00\xBC\xCD\x8C\x00\xBD\xB5"
-     "\x8C\x00"
-     "\xBE\xB1\x8C\x00\xBF\xAD\x8C\x00\x80\xB9\x26\x00\x81\xB9\x26\x00\x82\x4D"
-     "\x24\x00"
-     "\x83\x5D\x24\x00\xBB\x2A\x00\x80\x85\x15\x22\x00\x86\x45\x24\x00\x87\x55"
-     "\x24\x00"
-     "\x88\x51\x24\x00\x89\x21\x22\x00\x8A\xA1\x24\x00\x8B\xA1\x24\x00\x8C\xA9"
-     "\x24\x00"
-     "\x8D\xA9\x24\x00\x8E\x59\x24\x00\x8F\x45\x24\x00\x90\xC1\x24\x00\x91\x2D"
-     "\x24\x00"
-     "\x92\xCD\x24\x00\x93\x15\x24\x00\x94\x35\x24\x00\x95\x81\x26\x00\x96\x19"
-     "\x24\x00"
-     "\x97\x11\x24\x00\x98\x35\x24\x00\x99\x09\x24\x00\x9A\x1D\x24\x00\x9B\x11"
-     "\x24\x00"
-     "\x9C\x51\x15\x00\x9D\xF5\x37\x00\x9E\x5D\x15\x00\x9F\x55\x15\x00\xFD\xF1"
-     "\x00\x80"
-     "\xA1\x39\x26\x00\xA2\xCD\x24\x00\xA2\x2C\x00\x80\xA4\x51\x24\x00\xA6\x2C"
-     "\x00\x80"
-     "\xC2\x2C\x00\x80\xA7\xD5\x24\x00\xEE\xF1\x00\x80\xA9\xA1\x22\x00\xAA\x21"
-     "\x24\x00"
-     "\xAB\x21\x24\x00\xAC\x29\x24\x00\xAD\x29\x24\x00\xAE\xD9\x24\x00\xAF\xC5"
-     "\x24\x00"
-     "\xB0\x41\x24\x00\xB1\xAD\x24\x00\xB2\x4D\x24\x00\xBF\x2A\x00\x80\xB4\xB5"
-     "\x24\x00"
-     "\xB5\x01\x26\x00\xB6\x99\x24\x00\xB7\x91\x24\x00\xAE\x2C\x00\x80\xB9\x89"
-     "\x24\x00"
-     "\xF8\xF1\x00\x80\xBB\x91\x24\x00\xDE\x2C\x00\x80\xE6\x2C\x00\x80\x3F\x2C"
-     "\x00\x80"
-     "\xEE\x2C\x00\x80\x82\x45\x25\x00\xD3\xF3\x00\x80\x16\x2D\x00\x80\xC7\x2A"
-     "\x00\x80"
-     "\x2E\x2D\x00\x80\x87\x5D\x25\x00\x32\x2D\x00\x80\x3E\x2D\x00\x80\xCB\x2A"
-     "\x00\x80"
-     "\x8B\xA9\x25\x00\x76\x2D\x00\x80\x89\x29\x23\x00\x8E\x51\x25\x00\x8F\x4D"
-     "\x25\x00"
-     "\x7A\x2D\x00\x80\x8D\xA1\x25\x00\x92\xC5\x25\x00\xA2\x2D\x00\x80\xAE\x2D"
-     "\x00\x80"
-     "\x91\x25\x25\x00\xBE\x2D\x00\x80\x97\x19\x25\x00\x94\x3D\x25\x00\xC6\x2D"
-     "\x00\x80"
-     "\xCA\x2D\x00\x80\x9B\x19\x25\x00\x0C\xF2\x00\x80\x99\x01\x25\x00\xCB\x33"
-     "\x00\x80"
-     "\x9F\x5D\x14\x00\xDE\x2D\x00\x80\x9D\xFD\x36\x00\xA2\xC5\x25\x00\xE6\x2D"
-     "\x00\x80"
-     "\x11\xF2\x00\x80\xA1\x31\x27\x00\xF2\x2D\x00\x80\xA7\xDD\x25\x00\xA4\x59"
-     "\x25\x00"
-     "\xFA\x2D\x00\x80\xAA\x29\x25\x00\x02\x2E\x00\x80\xA8\xD9\x25\x00\xA9\xA9"
-     "\x23\x00"
-     "\xAE\xD1\x25\x00\xAF\xCD\x25\x00\xAC\x21\x25\x00\xAD\x21\x25\x00\xB2\x45"
-     "\x25\x00"
-     "\x2A\x2E\x00\x80\xB0\x49\x25\x00\xB1\xA5\x25\x00\xB6\x91\x25\x00\xB7\x99"
-     "\x25\x00"
-     "\xB4\xBD\x25\x00\xB5\x09\x27\x00\xBA\x95\x25\x00\xBB\x99\x25\x00\x16\x2E"
-     "\x00\x80"
-     "\xB9\x81\x25\x00\xBE\xD5\x14\x00\x22\x2E\x00\x80\xBC\xD9\x14\x00\x26\x2E"
-     "\x00\x80"
-     "\xB6\x41\x2A\x00\xB7\x49\x2A\x00\xB4\x6D\x2A\x00\xB5\xD9\x28\x00\xB2\x95"
-     "\x2A\x00"
-     "\xB3\x4D\x2A\x00\xB0\x99\x2A\x00\xB1\x75\x2A\x00\x2E\x2E\x00\x80\x36\x2E"
-     "\x00\x80"
-     "\x3A\x2E\x00\x80\x42\x2E\x00\x80\xBA\x45\x2A\x00\xBB\x49\x2A\x00\xB8\x6D"
-     "\x2A\x00"
-     "\xB9\x51\x2A\x00\xA6\x1D\x2A\x00\xA7\x0D\x2A\x00\x4A\x2E\x00\x80\xA5\x4D"
-     "\x2C\x00"
-     "\xA2\x15\x2A\x00\xA3\x05\x2A\x00\x4E\x2E\x00\x80\xA1\xE1\x28\x00\xAE\x01"
-     "\x2A\x00"
-     "\xAF\x1D\x2A\x00\xAC\xF1\x2A\x00\xAD\xF1\x2A\x00\x5E\x2E\x00\x80\xAB\xF9"
-     "\x2A\x00"
-     "\xA8\x09\x2A\x00\xA9\x79\x2C\x00\x96\xC1\x2A\x00\x97\xC9\x2A\x00\x94\xED"
-     "\x2A\x00"
-     "\x95\x59\x28\x00\x92\x15\x2A\x00\x93\xCD\x2A\x00\x90\x19\x2A\x00\x91\xF5"
-     "\x2A\x00"
-     "\x62\x2E\x00\x80\x6E\x2E\x00\x80\x92\x2E\x00\x80\x96\x2E\x00\x80\x9A\xC5"
-     "\x2A\x00"
-     "\x9B\xC9\x2A\x00\x98\xED\x2A\x00\x99\xD1\x2A\x00\x86\x9D\x2A\x00\x87\x8D"
-     "\x2A\x00"
-     "\x84\x09\x2A\x00\x85\xCD\x2C\x00\x82\x95\x2A\x00\x83\x85\x2A\x00\x80\x61"
-     "\x28\x00"
-     "\x81\x61\x28\x00\x8E\x81\x2A\x00\x8F\x9D\x2A\x00\x8C\x71\x2A\x00\x8D\x71"
-     "\x2A\x00"
-     "\x7A\x2E\x00\x80\x8B\x79\x2A\x00\x88\x89\x2A\x00\x89\xF9\x2C\x00\x80\x31"
-     "\x1A\x00"
-     "\x81\x31\x1A\x00\x82\x31\x1A\x00\x83\x31\x1A\x00\x84\x9D\x1A\x00\x85\x85"
-     "\x1A\x00"
-     "\x86\x8D\x1A\x00\x87\x85\x1A\x00\x88\xBD\x1A\x00\x89\x85\x1A\x00\x8A\x8D"
-     "\x1A\x00"
-     "\x82\x2E\x00\x80\xDE\x2E\x00\x80\xA6\x2E\x00\x80\xAA\x2E\x00\x80\xAE\x2E"
-     "\x00\x80"
-     "\x90\x19\xF8\x00\x91\x19\xF8\x00\x92\x25\xF8\x00\x93\x3D\xF8\x00\x94\x25"
-     "\xF8\x00"
-     "\x95\x2D\xF8\x00\x96\x25\xF8\x00\x97\xD9\xF8\x00\x98\xF1\xF8\x00\x99\xF1"
-     "\xF8\x00"
-     "\x9A\xC9\xF8\x00\x9B\xC9\xF8\x00\x9C\xD9\xF8\x00\x9D\xD9\xF8\x00\x9E\xC9"
-     "\xF8\x00"
-     "\x9F\xC9\xF8\x00\xA0\x39\xF8\x00\xA1\x39\xF8\x00\xA2\x61\xF8\x00\xA3\x61"
-     "\xF8\x00"
-     "\xA4\x61\xF8\x00\xA5\x61\xF8\x00\xA6\x19\xFB\x00\xA7\x19\xFB\x00\xA8\x29"
-     "\xFB\x00"
-     "\xA9\x29\xFB\x00\xAA\x01\x1A\x00\xAB\x5D\xBC\x02\xAC\x81\xB1\x02\xAD\x09"
-     "\xDE\x00"
-     "\xAE\x15\x1A\x00\xA2\x2E\x00\x80\xB0\xF5\xB9\x02\xB1\x09\xBA\x02\xB2\x39"
-     "\xBC\x02"
-     "\xB3\xC5\xBD\x02\xB4\xC1\xBD\x02\xB5\xF9\xBD\x02\xB6\xE9\xBD\x02\xB7\xE5"
-     "\xBD\x02"
-     "\xB8\xB5\xBD\x02\xB9\x51\xBD\x02\xBA\x2D\xBD\x02\xBB\x01\xBD\x02\xBC\xCD"
-     "\xBC\x02"
-     "\xBD\x65\xBC\x02\xBE\x01\xBC\x02\xBF\xD5\xB1\x02\x80\xA5\xBD\x02\x81\x35"
-     "\xB2\x02"
-     "\x82\x49\xB2\x02\x83\x55\xB2\x02\x84\x85\xB2\x02\x85\x71\xB3\x02\x86\x95"
-     "\xB3\x02"
-     "\x87\xAD\xB3\x02\x88\xED\xB3\x02\x89\xB1\xB7\x02\x8A\x91\x19\x00\x8B\xA5"
-     "\xDD\x00"
-     "\x8C\x8D\x19\x00\x8D\xFD\x19\x00\x8E\xF5\x19\x00\x8F\xED\x19\x00\xB0\x71"
-     "\x83\x00"
-     "\xB1\x91\xB1\x02\xB2\xC1\xB1\x02\xB3\xC1\xB1\x02\xB4\x65\x83\x00\xB5\x6D"
-     "\x83\x00"
-     "\xB6\x65\x83\x00\xB7\x1D\x83\x00\xB8\x29\x83\x00\xB9\x35\x83\x00\xB6\x2E"
-     "\x00\x80"
-     "\xB2\x2E\x00\x80\xCA\x2E\x00\x80\x16\xF2\x00\x80\xE6\x2E\x00\x80\xEA\x2E"
-     "\x00\x80"
-     "\xA1\x80\x05\x00\xA0\x00\x04\x00\xA3\x00\x78\x00\xA2\x8C\x06\x00\xA5\x00"
-     "\x7C\x00"
-     "\xA4\x94\x7A\x00\xA7\x00\x70\x00\xA6\x9C\x7E\x00\xF2\x2E\x00\x80\xA8\x24"
-     "\x71\x00"
-     "\xAA\x7D\x19\x00\xAB\x75\x19\x00\xAC\x6D\x19\x00\x06\x2F\x00\x80\x29\x2F"
-     "\x00\x80"
-     "\x1B\xF2\x00\x80\x20\xF2\x00\x80\x35\x2F\x00\x80\x39\x2F\x00\x80\x3D\x2F"
-     "\x00\x80"
-     "\x31\x2F\x00\x80\x45\x2F\x00\x80\xB4\x00\x00\x00\xCF\x2A\x00\x80\xD3\x2A"
-     "\x00\x80"
-     "\x4D\x2F\x00\x80\x55\x2F\x00\x80\x25\xF2\x00\x80\x71\x2F\x00\x80\x75\x2F"
-     "\x00\x80"
-     "\x81\x2F\x00\x80\x2A\xF2\x00\x80\x80\xE5\x18\x00\x81\xE9\x18\x00\x82\x45"
-     "\xDA\x00"
-     "\x89\x2F\x00\x80\x8D\x2F\x00\x80\x6C\x30\x00\x80\xD7\x2A\x00\x80\xA1\x2F"
-     "\x00\x80"
-     "\xBD\x2F\x00\x80\xB9\x2F\x00\x80\x34\xF2\x00\x80\xC5\x2F\x00\x80\x39\xF2"
-     "\x00\x80"
-     "\xD1\x2F\x00\x80\xCD\x2F\x00\x80\xD5\x2F\x00\x80\x90\x0D\xF1\x00\x91\x11"
-     "\x1E\x00"
-     "\x92\x3D\x1D\x00\x93\x3D\xEF\x00\x94\x75\xF3\x00\x95\xD5\x1D\x00\x96\xDD"
-     "\x1D\x00"
-     "\x97\x9D\xE2\x00\x98\xED\x1D\x00\x99\xF5\x1D\x00\x9A\xF5\x1D\x00\x9B\xE5"
-     "\xCC\x00"
-     "\x9C\xD1\x1D\x00\x9D\xC1\x1D\x00\x9E\x91\x1D\x00\x9F\x8D\x1D\x00\xA0\x91"
-     "\x1C\x00"
-     "\xA1\xA5\x13\x00\xA2\x55\xF1\x00\xA3\xC1\x13\x00\xA4\xF5\x13\x00\xA5\xED"
-     "\x13\x00"
-     "\xA6\x31\x13\x00\xA7\x21\x13\x00\xA8\xB5\x12\x00\xA9\x5D\xFD\x00\xAA\x09"
-     "\xE2\x00"
-     "\xAB\xB1\x12\x00\xAC\xCD\xE5\x00\xAD\x19\xE2\x00\xAE\x39\xE5\x00\xAF\x9D"
-     "\x12\x00"
-     "\xB0\x81\xF6\x00\xB1\xE1\x12\x00\xB2\x8D\x12\x00\xB3\x55\x15\x00\xB4\x4D"
-     "\x15\x00"
-     "\xB5\xD9\x0B\x00\xB6\xE1\xE5\x00\xB7\xA9\xF1\x00\xB8\xF5\x0B\x00\xB9\xFD"
-     "\x0B\x00"
-     "\xBA\xF5\x0B\x00\xBB\xCD\x0B\x00\xE5\x2F\x00\x80\xE9\x2F\x00\x80\xDB\x2A"
-     "\x00\x80"
-     "\xF9\x2F\x00\x80\xA1\x19\x0B\x00\xA0\x15\x0B\x00\xA3\x15\x0B\x00\xA2\x15"
-     "\x0C\x00"
-     "\xA5\xDD\xC7\x00\xA4\x35\x0B\x00\xA7\x49\xC5\x00\xA6\x51\xC5\x00\xA9\x09"
-     "\x0B\x00"
-     "\xA8\xF5\xC7\x00\xAB\x01\xC8\x00\xAA\x15\x0B\x00\xAD\x4D\xC5\x00\xAC\x71"
-     "\x0B\x00"
-     "\xAF\x69\x0B\x00\xAE\x75\x0B\x00\xB1\x21\x0B\x00\xB0\x19\x0B\x00\xB3\x29"
-     "\x0B\x00"
-     "\xB2\x2D\x0B\x00\xB5\x21\x0B\x00\xB4\x3D\x0B\x00\xB7\xDD\x0B\x00\xB6\x25"
-     "\x0B\x00"
-     "\xB9\xF1\x0B\x00\xB8\xED\x0B\x00\xBB\xCD\x0B\x00\xBA\xF1\x0B\x00\xBD\x99"
-     "\x0B\x00"
-     "\xBC\xB1\x0B\x00\xBF\x85\x0B\x00\xBE\x89\x0B\x00\x81\xE1\x0F\x00\x80\xA9"
-     "\x0F\x00"
-     "\x83\xE1\x0F\x00\x82\xE9\x0F\x00\x85\xA9\xC5\x00\x84\xF9\x0F\x00\x87\xE1"
-     "\x0F\x00"
-     "\x86\xE9\x0F\x00\x89\x91\x09\x00\x88\xD9\x0F\x00\x8B\x9D\x09\x00\x8A\x91"
-     "\xC5\x00"
-     "\x8D\x31\x0A\x00\x8C\xA5\x0A\x00\x8F\x35\x0A\x00\x8E\xF9\xC5\x00\x91\x59"
-     "\x0A\x00"
-     "\x90\x51\x0A\x00\x93\x05\x0A\x00\x92\x6D\x0A\x00\x95\x39\x1A\x00\x94\x35"
-     "\x0A\x00"
-     "\x97\x29\x0A\x00\x96\x35\x0A\x00\x99\x19\x0A\x00\x98\x11\x0A\x00\x9B\x51"
-     "\xC7\x00"
-     "\x9A\x11\x0A\x00\x9D\x85\xF4\x00\x9C\xF1\x0B\x00\x9F\xF5\x0B\x00\x9E\xF5"
-     "\x0B\x00"
-     "\x80\x21\x05\x00\x81\x21\x05\x00\x82\x21\x05\x00\x83\x21\x05\x00\x84\x59"
-     "\x05\x00"
-     "\x85\x4D\x05\x00\x86\x41\x05\x00\x87\xC5\xC6\x00\x88\x51\x05\x00\x89\x85"
-     "\x05\x00"
-     "\x8A\x8D\x05\x00\x8B\x85\x05\x00\x8C\xBD\xC4\x00\x8D\x99\x05\x00\x8E\xC9"
-     "\x05\x00"
-     "\x8F\x11\x04\x00\x90\x39\xC5\x00\x91\x75\x04\x00\x92\x7D\x04\x00\x93\x75"
-     "\x04\x00"
-     "\x94\x6D\x04\x00\x95\xE9\x04\x00\x96\x31\x07\x00\x97\x31\x07\x00\x98\x11"
-     "\x07\x00"
-     "\x99\x11\x07\x00\x9A\x11\x07\x00\x9B\x11\x07\x00\x9C\xD1\x07\x00\x9D\xF5"
-     "\x1B\x00"
-     "\x9E\xCD\x07\x00\x9F\xC5\x07\x00\xA0\x3D\x07\x00\xA1\x45\x07\x00\xA2\x4D"
-     "\x07\x00"
-     "\xA3\x45\x07\x00\xA4\x5D\x07\x00\xA5\x45\x07\x00\xA6\x4D\x07\x00\xA7\x41"
-     "\x07\x00"
-     "\xA8\x41\x07\x00\xA9\x41\x07\x00\xAA\x41\x07\x00\xAB\x45\x07\x00\xAC\x5D"
-     "\x07\x00"
-     "\xAD\x45\x07\x00\xAE\x4D\x07\x00\xAF\xA5\xC9\x00\xB0\x39\x07\x00\xB1\x39"
-     "\x07\x00"
-     "\xB2\xC5\x06\x00\xB3\xDD\x06\x00\xB4\xC5\x06\x00\xB5\x21\xF5\x00\xB6\xC1"
-     "\x06\x00"
-     "\xB7\xC1\x06\x00\xB8\x1D\xF5\x00\xB9\xC5\x06\x00\xBA\xCD\x06\x00\xBB\xC5"
-     "\x06\x00"
-     "\xBC\xDD\x06\x00\xBD\xC5\x06\x00\xBE\xCD\x06\x00\xBF\xC5\x06\x00\x80\x3D"
-     "\x05\x00"
-     "\x81\xF5\x05\x00\x82\xFD\x05\x00\x83\xF5\x05\x00\x84\xED\x05\x00\x85\x11"
-     "\x01\x00"
-     "\x86\x11\x01\x00\x87\x11\x01\x00\x88\x31\x01\x00\x89\x31\x01\x00\x8A\x31"
-     "\x01\x00"
-     "\x8B\x31\x01\x00\x8C\x11\x01\x00\x8D\x11\x01\x00\x8E\x7D\xCD\x00\x8F\x15"
-     "\x01\x00"
-     "\x90\x69\xF4\x00\x91\x69\x01\x00\x92\x65\x01\x00\x93\x7D\x01\x00\x94\x65"
-     "\x01\x00"
-     "\x95\x61\x01\x00\x96\x61\x01\x00\x97\x61\x01\x00\x98\x21\x01\x00\x99\x21"
-     "\x01\x00"
-     "\x9A\x21\x01\x00\x9B\x21\x01\x00\x9C\x21\x01\x00\x9D\x21\x01\x00\x9E\x21"
-     "\x01\x00"
-     "\x9F\x21\x01\x00\xA0\xE1\x01\x00\xA1\xE1\x01\x00\xA2\xE1\x01\x00\xA3\x89"
-     "\xC5\x00"
-     "\xA4\xFD\x01\x00\xA5\xE5\x01\x00\xA6\xED\x01\x00\xA7\xE5\x01\x00\xA8\xA9"
-     "\xC3\x00"
-     "\xA9\xDD\x01\x00\xAA\xD5\x01\x00\xAB\xA1\xC5\x00\xAC\x31\x01\x00\xAD\x31"
-     "\x01\x00"
-     "\xAE\x31\x01\x00\xAF\x31\x01\x00\xB0\xD1\xC5\x00\xB1\xAD\xCE\x00\xB2\x59"
-     "\x01\x00"
-     "\xB3\x59\x01\x00\xB4\x49\x01\x00\xB5\x49\x01\x00\xB6\x79\x01\x00\xB7\x79"
-     "\x01\x00"
-     "\xB8\x49\x01\x00\xB9\x49\x01\x00\xBA\x59\x01\x00\xBB\x59\x01\x00\xBC\x49"
-     "\x01\x00"
-     "\xBD\x49\x01\x00\xBE\x39\x01\x00\xBF\x39\x01\x00\xC1\x00\x00\x00\x01\x30"
-     "\x00\x80"
-     "\x05\x30\x00\x80\x09\x30\x00\x80\xDF\x2A\x00\x80\x3E\xF2\x00\x80\xE3\x2A"
-     "\x00\x80"
-     "\x2C\x30\x00\x80\x30\x30\x00\x80\x34\x30\x00\x80\x05\xF4\x00\x80\x48\xF2"
-     "\x00\x80"
-     "\x0F\x37\x00\x80\x64\x30\x00\x80\xE7\x2A\x00\x80\x6B\xF2\x00\x80\x3D\xF3"
-     "\x00\x80"
-     "\x74\x30\x00\x80\x7C\x30\x00\x80\xEB\x2A\x00\x80\x90\x30\x00\x80\xA0\x30"
-     "\x00\x80"
-     "\xA4\x30\x00\x80\xAC\x30\x00\x80\xBC\x30\x00\x80\xF3\x2A\x00\x80\xEF\x2A"
-     "\x00\x80"
-     "\x4D\xF2\x00\x80\xC4\x30\x00\x80\xD8\x30\x00\x80\xDC\x30\x00\x80\xEC\x30"
-     "\x00\x80"
-     "\x00\x31\x00\x80\xF8\x30\x00\x80\xFC\x30\x00\x80\x04\x31\x00\x80\x18\x31"
-     "\x00\x80"
-     "\x1C\x31\x00\x80\x34\x31\x00\x80\x3C\x31\x00\x80\x4C\x31\x00\x80\x60\x31"
-     "\x00\x80"
-     "\x5C\xF2\x00\x80\x5C\x31\x00\x80\x50\x31\x00\x80\x68\x31\x00\x80\x70\x31"
-     "\x00\x80"
-     "\x84\x31\x00\x80\x61\xF2\x00\x80\x8C\x31\x00\x80\x80\x31\x00\x80\x78\x31"
-     "\x00\x80"
-     "\xF7\x2A\x00\x80\x94\x31\x00\x80\x9C\x31\x00\x80\xA4\x31\x00\x80\x98\x31"
-     "\x00\x80"
-     "\xFB\x2A\x00\x80\xBC\x31\x00\x80\x66\xF2\x00\x80\xE8\x31\x00\x80\x2A\x32"
-     "\x00\x80"
-     "\x06\x32\x00\x80\x07\x2B\x00\x80\xA2\x21\xBB\x00\xA3\x45\x01\x00\xA0\x31"
-     "\x01\x00"
-     "\xA1\x31\x01\x00\xA6\x5D\x01\x00\xA7\xC1\xBD\x00\xA4\x5D\x01\x00\xA5\x55"
-     "\x01\x00"
-     "\xAA\x79\x01\x00\xAB\x79\x01\x00\xA8\x79\x01\x00\xA9\x79\x01\x00\xAE\x59"
-     "\x01\x00"
-     "\xAF\x59\x01\x00\xAC\x59\x01\x00\xAD\x59\x01\x00\xB2\x39\x01\x00\xB3\x39"
-     "\x01\x00"
-     "\xB0\x39\x01\x00\xB1\x39\x01\x00\xB6\xD5\x01\x00\xB7\xDD\x01\x00\xB4\xD9"
-     "\x01\x00"
-     "\xB5\x01\xB4\x00\xBA\xF5\x01\x00\xBB\xFD\x01\x00\xB8\xE5\x01\x00\xB9\xFD"
-     "\x01\x00"
-     "\xBE\xD5\x01\x00\xBF\xDD\x01\x00\xBC\xE5\x01\x00\xBD\xDD\x01\x00\x82\x51"
-     "\x7F\x00"
-     "\x83\x51\x7F\x00\x80\x41\x7F\x00\x81\x41\x7F\x00\x86\x71\x7F\x00\x87\x71"
-     "\x7F\x00"
-     "\x84\x41\x7F\x00\x85\x41\x7F\x00\x8A\x4D\x7F\x00\x8B\x55\x7F\x00\x88\xBD"
-     "\xBB\x00"
-     "\x89\x45\x7F\x00\x8E\x4D\x7F\x00\x8F\xA9\xBB\x00\x8C\x4D\x7F\x00\x8D\x45"
-     "\x7F\x00"
-     "\x92\xC9\x01\x00\x93\xC9\x01\x00\x90\xC9\x01\x00\x91\xC9\x01\x00\x96\xC9"
-     "\x01\x00"
-     "\x97\xC9\x01\x00\x94\xC9\x01\x00\x95\xC9\x01\x00\x9A\xC1\x01\x00\x9B\xC1"
-     "\x01\x00"
-     "\x98\xA1\xB3\x00\x99\xC5\xBD\x00\x9E\xC1\x01\x00\x9F\xC1\x01\x00\x9C\xD1"
-     "\x01\x00"
-     "\x9D\xD1\x01\x00\x03\x2B\x00\x80\xFF\x2A\x00\x80\xD2\x2C\x00\x80\xD6\x2C"
-     "\x00\x80"
-     "\x2E\x32\x00\x80\x16\x32\x00\x80\xE3\x36\x00\x80\x6B\x2B\x00\x80\x42\x32"
-     "\x00\x80"
-     "\x56\x32\x00\x80\x4E\x32\x00\x80\x7F\xF2\x00\x80\x0B\x2B\x00\x80\x6A\x32"
-     "\x00\x80"
-     "\x5E\x32\x00\x80\x87\x32\x00\x80\x89\xF2\x00\x80\x93\x32\x00\x80\x72\x32"
-     "\x00\x80"
-     "\x9B\x32\x00\x80\x9F\x32\x00\x80\xA3\x32\x00\x80\xA7\x32\x00\x80\xB3\x32"
-     "\x00\x80"
-     "\x8E\xF2\x00\x80\xBB\x32\x00\x80\x13\x2B\x00\x80\xC7\x32\x00\x80\x93\xF2"
-     "\x00\x80"
-     "\xCB\x32\x00\x80\x17\x2B\x00\x80\xD7\x32\x00\x80\xE3\x32\x00\x80\xF3\x32"
-     "\x00\x80"
-     "\x98\xF2\x00\x80\x2F\xF2\x00\x80\x9D\xF2\x00\x80\x13\x33\x00\x80\xA2\xF2"
-     "\x00\x80"
-     "\x1F\x33\x00\x80\x27\x33\x00\x80\x17\x33\x00\x80\x33\x33\x00\x80\x43\x33"
-     "\x00\x80"
-     "\x4B\x33\x00\x80\x57\x33\x00\x80\x5B\x33\x00\x80\xA7\xF2\x00\x80\x3F\x33"
-     "\x00\x80"
-     "\x73\x33\x00\x80\x77\x33\x00\x80\x1B\x2B\x00\x80\x87\x33\x00\x80\xAC\xF2"
-     "\x00\x80"
-     "\x6B\x33\x00\x80\xA7\x33\x00\x80\xB1\xF2\x00\x80\xB6\xF2\x00\x80\xAB\x33"
-     "\x00\x80"
-     "\xBF\x33\x00\x80\xB7\x33\x00\x80\x1F\x2B\x00\x80\xC3\x33\x00\x80\xD3\x33"
-     "\x00\x80"
-     "\x80\xAD\x03\x00\x81\x21\xB9\x00\x82\xB9\xB5\x00\x83\xBD\x03\x00\x84\xA5"
-     "\x03\x00"
-     "\x85\xAD\x03\x00\x86\xA5\x03\x00\x87\xDD\x03\x00\x88\xE5\x03\x00\x89\xED"
-     "\x03\x00"
-     "\x8A\xE5\x03\x00\x8B\x19\xBF\x00\x8C\xE1\x03\x00\x8D\xE1\x03\x00\x8E\xE1"
-     "\x03\x00"
-     "\x8F\xE1\x03\x00\x90\xA1\x03\x00\x91\xA1\x03\x00\x92\xA1\x03\x00\x93\xA1"
-     "\x03\x00"
-     "\x94\x69\xBF\x00\x95\xA5\x03\x00\x96\xAD\x03\x00\x97\xA5\x03\x00\x98\x9D"
-     "\x03\x00"
-     "\x99\x65\x01\x00\x9A\x6D\x01\x00\x9B\x65\x01\x00\x9C\x7D\x01\x00\x9D\x65"
-     "\x01\x00"
-     "\x9E\x6D\x01\x00\x9F\x65\x01\x00\xA0\x9D\x01\x00\xA1\xC5\xBF\x00\xA2\xA9"
-     "\x01\x00"
-     "\xA3\xA9\x01\x00\xA4\xB9\x01\x00\xA5\xB9\x01\x00\xA6\xA9\x01\x00\xA7\xA9"
-     "\x01\x00"
-     "\xA8\x99\x01\x00\xA9\x99\x01\x00\xAA\xE9\x01\x00\xAB\xE9\x01\x00\xAC\xF9"
-     "\x01\x00"
-     "\xAD\xFD\x01\x00\xAE\xF5\x01\x00\xAF\xED\x01\x00\xB0\x89\xBF\x00\xB1\x91"
-     "\x01\x00"
-     "\xB2\x91\x01\x00\xB3\x91\x01\x00\xB4\xB1\x01\x00\xB5\xB1\x01\x00\xB6\xB1"
-     "\x01\x00"
-     "\xB7\xB1\x01\x00\xB8\xC5\xB6\x00\xB9\x95\x01\x00\xBA\x9D\x01\x00\xBB\x95"
-     "\x01\x00"
-     "\xBC\x8D\x01\x00\xBD\x75\x01\x00\xBE\x7D\x01\x00\xBF\x75\x01\x00\xCF\x33"
-     "\x00\x80"
-     "\xDB\x33\x00\x80\xF3\xF1\x00\x80\xEB\x33\x00\x80\xBB\xF2\x00\x80\xF7\x33"
-     "\x00\x80"
-     "\xC5\xF2\x00\x80\x0B\x34\x00\x80\x2B\x34\x00\x80\xCA\xF2\x00\x80\x3F\x34"
-     "\x00\x80"
-     "\x43\x34\x00\x80\xCF\xF2\x00\x80\xD4\xF2\x00\x80\x5F\x34\x00\x80\x6F\x34"
-     "\x00\x80"
-     "\x23\x2B\x00\x80\x73\x34\x00\x80\x27\x2B\x00\x80\x8B\x34\x00\x80\x8F\x34"
-     "\x00\x80"
-     "\x9B\x34\x00\x80\xA7\x34\x00\x80\x2B\x2B\x00\x80\xBF\x34\x00\x80\xD9\xF2"
-     "\x00\x80"
-     "\xDB\x34\x00\x80\xDE\xF2\x00\x80\x43\xF2\x00\x80\xF7\x34\x00\x80\xE3\xF2"
-     "\x00\x80"
-     "\xE8\xF2\x00\x80\xED\xF2\x00\x80\x2F\x2B\x00\x80\x33\x2B\x00\x80\xF7\xF2"
-     "\x00\x80"
-     "\xF2\xF2\x00\x80\xFC\xF2\x00\x80\x01\xF3\x00\x80\x37\x35\x00\x80\x3B\x35"
-     "\x00\x80"
-     "\x47\x35\x00\x80\x3F\x2B\x00\x80\x43\x2B\x00\x80\x0B\xF3\x00\x80\x5B\x35"
-     "\x00\x80"
-     "\x47\x2B\x00\x80\x10\xF3\x00\x80\x15\xF3\x00\x80\x1A\xF3\x00\x80\xCB\x35"
-     "\x00\x80"
-     "\x4B\x2B\x00\x80\xD7\x35\x00\x80\xDB\x35\x00\x80\x1F\xF3\x00\x80\x24\xF3"
-     "\x00\x80"
-     "\xF3\x35\x00\x80\x4F\x2B\x00\x80\x29\xF3\x00\x80\x07\x36\x00\x80\x0B\x36"
-     "\x00\x80"
-     "\x53\x2B\x00\x80\x2E\xF3\x00\x80\x27\x36\x00\x80\xA4\xF5\x07\x00\xA5\xFD"
-     "\x07\x00"
-     "\xA6\xF5\x07\x00\xA7\x0D\x01\x00\xA0\xF5\x07\x00\xA1\xFD\x07\x00\xA2\xF5"
-     "\x07\x00"
-     "\xA3\xED\x07\x00\xAC\x35\x01\x00\xAD\x3D\x01\x00\xAE\x35\x01\x00\xAF\x0D"
-     "\x01\x00"
-     "\xA8\x35\x01\x00\xA9\x3D\x01\x00\xAA\x35\x01\x00\xAB\x2D\x01\x00\xB4\x75"
-     "\x01\x00"
-     "\xB5\x7D\x01\x00\xB6\x75\x01\x00\xB7\x0D\x01\x00\xB0\x75\x01\x00\xB1\x7D"
-     "\x01\x00"
-     "\xB2\x75\x01\x00\xB3\x6D\x01\x00\xBC\x31\x01\x00\xBD\x31\x01\x00\xBE\x31"
-     "\x01\x00"
-     "\xBF\x31\x01\x00\xB8\x35\x01\x00\xB9\x3D\x01\x00\xBA\x09\xBF\x00\xBB\x31"
-     "\x01\x00"
-     "\x84\x19\x07\x00\x85\x19\x07\x00\x86\x29\x07\x00\x87\x2D\x07\x00\x80\xF1"
-     "\xB9\x00"
-     "\x81\x19\x07\x00\x82\x09\x07\x00\x83\x09\x07\x00\x8C\x0D\x07\x00\x8D\x15"
-     "\x07\x00"
-     "\x8E\x1D\x07\x00\x8F\x21\xB7\x00\x88\xD5\xB9\x00\x89\x69\xB9\x00\x8A\x1D"
-     "\x07\x00"
-     "\x8B\x15\x07\x00\x94\x01\x07\x00\x95\x01\x07\x00\x96\x05\xBC\x00\x97\x05"
-     "\x07\x00"
-     "\x90\xBD\xB9\x00\x91\x6D\x07\x00\x92\x65\x07\x00\x93\x91\xBF\x00\x9C\x39"
-     "\x07\x00"
-     "\x9D\x39\x07\x00\x9E\x35\x07\x00\x9F\x0D\x07\x00\x98\x3D\x07\x00\x99\xAD"
-     "\xBF\x00"
-     "\x9A\x29\x07\x00\x9B\x29\x07\x00\x57\x2B\x00\x80\x37\x36\x00\x80\x2F\x36"
-     "\x00\x80"
-     "\x3F\x36\x00\x80\x38\xF3\x00\x80\x57\x36\x00\x80\x5B\x2B\x00\x80\x63\x36"
-     "\x00\x80"
-     "\x6B\x36\x00\x80\x7B\x36\x00\x80\x5F\x2B\x00\x80\x42\xF3\x00\x80\x47\xF3"
-     "\x00\x80"
-     "\x63\x2B\x00\x80\x4C\xF3\x00\x80\x93\x36\x00\x80\x51\xF3\x00\x80\x9F\x36"
-     "\x00\x80"
-     "\xAB\x36\x00\x80\x56\xF3\x00\x80\x5B\xF3\x00\x80\xC7\x36\x00\x80\x60\xF3"
-     "\x00\x80"
-     "\xCF\x36\x00\x80\x70\xF2\x00\x80\x67\x2B\x00\x80\xE7\x36\x00\x80\xEB\x36"
-     "\x00\x80"
-     "\x6F\x2B\x00\x80\xEF\x36\x00\x80\x49\x2F\x00\x80\x65\xF3\x00\x80\x6A\xF3"
-     "\x00\x80"
-     "\x75\xF2\x00\x80\x7A\xF2\x00\x80\x13\x37\x00\x80\x2B\x39\x00\x80\x73\x2B"
-     "\x00\x80"
-     "\x3F\x37\x00\x80\x3B\x37\x00\x80\x43\x37\x00\x80\x4E\x2D\x00\x80\x47\x37"
-     "\x00\x80"
-     "\x4B\x37\x00\x80\x4F\x37\x00\x80\x57\x37\x00\x80\x6F\xF3\x00\x80\x5B\x37"
-     "\x00\x80"
-     "\x6F\x37\x00\x80\x73\x37\x00\x80\x5F\x37\x00\x80\x77\x37\x00\x80\x8B\x37"
-     "\x00\x80"
-     "\x6B\x37\x00\x80\x7F\x37\x00\x80\x83\x37\x00\x80\x87\x37\x00\x80\x74\xF3"
-     "\x00\x80"
-     "\x7E\xF3\x00\x80\x79\xF3\x00\x80\x77\x2B\x00\x80\xA7\x37\x00\x80\xAB\x37"
-     "\x00\x80"
-     "\xB3\x37\x00\x80\x80\x61\x03\x00\x81\x61\x03\x00\x82\x61\x03\x00\x83\x61"
-     "\x03\x00"
-     "\x84\x61\x03\x00\x85\x61\x03\x00\x86\x61\x03\x00\x87\x61\x03\x00\x88\xA1"
-     "\x03\x00"
-     "\x89\xA1\x03\x00\x8A\xA1\x03\x00\x8B\x1D\x79\x00\x8C\xBD\x03\x00\x8D\xA5"
-     "\x03\x00"
-     "\x8E\xAD\x03\x00\x8F\xA5\x03\x00\x90\xDD\x03\x00\x91\xE5\x03\x00\x92\xED"
-     "\x03\x00"
-     "\x93\xE5\x03\x00\x94\xFD\x03\x00\x95\xE5\x03\x00\x96\xED\x03\x00\x97\xE5"
-     "\x03\x00"
-     "\x98\x09\xB2\x00\x99\xD9\x03\x00\x9A\xA9\x03\x00\x9B\xA9\x03\x00\x9C\xB9"
-     "\x03\x00"
-     "\x9D\xB9\x03\x00\x9E\xA9\x03\x00\x9F\x01\xBD\x00\xA0\x55\x03\x00\xA1\x5D"
-     "\x03\x00"
-     "\xA2\x55\x03\x00\xA3\x6D\x03\x00\xA4\x75\x03\x00\xA5\x7D\x03\x00\xA6\x75"
-     "\x03\x00"
-     "\xA7\x6D\x03\x00\xA8\x55\x03\x00\xA9\x5D\x03\x00\xAA\x55\x03\x00\xAB\xAD"
-     "\x01\x00"
-     "\xAC\xB5\x01\x00\xAD\xBD\x01\x00\xAE\xB5\x01\x00\xAF\xAD\x01\x00\xB0\xD5"
-     "\x01\x00"
-     "\xB1\xDD\x01\x00\xB2\xD5\x01\x00\xB3\xED\x01\x00\xB4\xA9\xB5\x00\xB5\xF1"
-     "\x01\x00"
-     "\xB6\xF1\x01\x00\xB7\xF1\x01\x00\xB8\xD1\x01\x00\xB9\xD1\x01\x00\xBA\xD1"
-     "\x01\x00"
-     "\xBB\x65\xBB\x00\xBC\xCD\x01\x00\xBD\xB5\x01\x00\xBE\xBD\x01\x00\xBF\xB5"
-     "\x01\x00"
-     "\x92\xF3\x00\x80\xB7\x37\x00\x80\x83\xF3\x00\x80\x7B\x2B\x00\x80\x7F\x2B"
-     "\x00\x80"
-     "\x88\xF3\x00\x80\x8D\xF3\x00\x80\x83\x2B\x00\x80\xD7\x37\x00\x80\xDF\x37"
-     "\x00\x80"
-     "\xE3\x37\x00\x80\xEF\x37\x00\x80\xEB\x37\x00\x80\xF7\x37\x00\x80\xF3\x37"
-     "\x00\x80"
-     "\xFB\x37\x00\x80\xFF\x37\x00\x80\x07\x38\x00\x80\x87\x2B\x00\x80\x0F\x38"
-     "\x00\x80"
-     "\x13\x38\x00\x80\x8B\x2B\x00\x80\x23\x38\x00\x80\x97\xF3\x00\x80\x33\x38"
-     "\x00\x80"
-     "\x37\x38\x00\x80\x8F\x2B\x00\x80\x43\x38\x00\x80\xC3\x2A\x00\x80\x9C\xF3"
-     "\x00\x80"
-     "\xA1\xF3\x00\x80\x93\x2B\x00\x80\x97\x2B\x00\x80\x6F\x38\x00\x80\xA6\xF3"
-     "\x00\x80"
-     "\xC7\x38\x00\x80\xCB\x38\x00\x80\xE3\x38\x00\x80\xEF\x38\x00\x80\xB0\xF3"
-     "\x00\x80"
-     "\x02\xF2\x00\x80\xFB\x38\x00\x80\xF7\x38\x00\x80\x03\x39\x00\x80\x07\xF2"
-     "\x00\x80"
-     "\x0F\x39\x00\x80\xB5\xF3\x00\x80\xBA\xF3\x00\x80\x5B\x39\x00\x80\x67\x39"
-     "\x00\x80"
-     "\x6F\x39\x00\x80\xBF\xF3\x00\x80\x73\x39\x00\x80\x9F\x39\x00\x80\xAB\x39"
-     "\x00\x80"
-     "\xAF\x39\x00\x80\xA7\x39\x00\x80\xBB\x39\x00\x80\xBF\x39\x00\x80\xC4\xF3"
-     "\x00\x80"
-     "\xCB\x39\x00\x80\x9B\x2B\x00\x80\xD3\x39\x00\x80\xC9\xF3\x00\x80\xA6\xD1"
-     "\x0F\x00"
-     "\xA7\xD1\x0F\x00\xA4\xC1\x0F\x00\xA5\xC1\x0F\x00\xA2\xD1\x0F\x00\xA3\xD1"
-     "\x0F\x00"
-     "\xA0\xA1\x0F\x00\xA1\xA1\x0F\x00\xAE\xD1\x0F\x00\xAF\xD1\x0F\x00\xAC\xC1"
-     "\x0F\x00"
-     "\xAD\xC1\x0F\x00\xAA\xD1\x0F\x00\xAB\xD1\x0F\x00\xA8\xE1\x0F\x00\xA9\xE1"
-     "\x0F\x00"
-     "\xB6\x51\x01\x00\xB7\x51\x01\x00\xB4\x41\x01\x00\xB5\x41\x01\x00\xB2\x51"
-     "\x01\x00"
-     "\xB3\x51\x01\x00\xB0\xA1\x0F\x00\xB1\xA1\x0F\x00\xBE\xA1\xB5\x00\xBF\xA5"
-     "\xB5\x00"
-     "\xBC\x41\x01\x00\xBD\x41\x01\x00\xBA\x51\x01\x00\xBB\x51\x01\x00\xB8\x61"
-     "\x01\x00"
-     "\xB9\x61\x01\x00\x86\xE1\x0F\x00\x87\xE1\x0F\x00\x84\x9D\x85\x00\x85\xD1"
-     "\x0F\x00"
-     "\x82\xC5\x0F\x00\x83\xCD\x0F\x00\x80\xD5\x0F\x00\x81\xCD\x0F\x00\x8E\x21"
-     "\x0F\x00"
-     "\x8F\x21\x0F\x00\x8C\xD1\x0F\x00\x8D\xD1\x0F\x00\x8A\xC1\x0F\x00\x8B\xC1"
-     "\x0F\x00"
-     "\x88\xD1\x0F\x00\x89\xD1\x0F\x00\x96\x55\x0F\x00\x97\x5D\x0F\x00\x94\x45"
-     "\x0F\x00"
-     "\x95\x5D\x0F\x00\x92\xE1\x85\x00\x93\x5D\x0F\x00\x90\x45\xB5\x00\x91\x79"
-     "\xB5\x00"
-     "\x9E\x55\x0F\x00\x9F\x4D\xB5\x00\x9C\x45\x0F\x00\x9D\x5D\x0F\x00\x9A\x55"
-     "\x0F\x00"
-     "\x9B\x5D\x0F\x00\x98\x65\x0F\x00\x99\x5D\x0F\x00\x9F\x2B\x00\x80\x0B\x3A"
-     "\x00\x80"
-     "\xD9\x2F\x00\x80\x23\x3A\x00\x80\xCE\xF3\x00\x80\xD8\xF3\x00\x80\xA3\x2B"
-     "\x00\x80"
-     "\xA7\x2B\x00\x80\x4B\x3A\x00\x80\xDD\xF3\x00\x80\xAB\x2B\x00\x80\xE2\xF3"
-     "\x00\x80"
-     "\x6B\x3A\x00\x80\xE7\xF3\x00\x80\x83\x3A\x00\x80\xAF\x2B\x00\x80\x93\x3A"
-     "\x00\x80"
-     "\x9F\x3A\x00\x80\xA7\x3A\x00\x80\xAF\x3A\x00\x80\xB3\x2B\x00\x80\xEC\xF3"
-     "\x00\x80"
-     "\xDF\x3A\x00\x80\xEB\x3A\x00\x80\xB7\x2B\x00\x80\xBB\x2B\x00\x80\xEF\x3A"
-     "\x00\x80"
-     "\xF1\xF3\x00\x80\xBF\x2B\x00\x80\xF6\xF3\x00\x80\xFB\xF3\x00\x80\x00\xF4"
-     "\x00\x80"
-     "\x80\xB1\x00\x00\x81\xB1\x00\x00\x82\xB1\x00\x00\x83\xB1\x00\x00\x84\xD1"
-     "\x00\x00"
-     "\x85\xD1\x00\x00\x86\xD1\x00\x00\x87\xD1\x00\x00\x88\xF1\x00\x00\x89\xF1"
-     "\x00\x00"
-     "\x8A\x6D\xB4\x00\x8B\xF5\x00\x00\x8C\xED\x00\x00\x8D\xD5\x00\x00\x8E\xDD"
-     "\x00\x00"
-     "\x8F\xD5\x00\x00\x90\xAD\x00\x00\x91\xB5\x00\x00\x92\xBD\x00\x00\x93\xB5"
-     "\x00\x00"
-     "\x94\xAD\x00\x00\x95\x21\x84\x00\x96\x21\x00\x00\x97\x81\x87\x00\x98\x1D"
-     "\x00\x00"
-     "\x99\xE5\x03\x00\x9A\xED\x03\x00\x9B\xE5\x03\x00\x9C\x81\x87\x00\x9D\xF9"
-     "\x03\x00"
-     "\xC3\x2B\x00\x80\x2F\x3B\x00\x80\x33\x3B\x00\x80\x3B\x3B\x00\x80\x43\x3B"
-     "\x00\x80"
-     "\x0A\xF4\x00\x80\x84\x00\x00\x00\x85\x00\x00\x00\x86\x00\x00\x00\x87\x00"
-     "\x00\x00"
-     "\x88\x00\x00\x00\x89\x00\x00\x00\x8A\x00\x00\x00\x8B\x00\x00\x00\x8C\x00"
-     "\x00\x00"
-     "\x8D\x00\x00\x00\x8E\x00\x00\x00\x8F\x00\x00\x00\x90\x00\x00\x00\x91\x00"
-     "\x00\x00"
-     "\x92\x00\x00\x00\x93\x00\x00\x00\x94\x00\x00\x00\x95\x00\x00\x00\x96\x00"
-     "\x00\x00"
-     "\x97\x00\x00\x00\x98\x00\x00\x00\x99\x00\x00\x00\x9A\x00\x00\x00\x9B\x00"
-     "\x00\x00"
-     "\x9C\x00\x00\x00\x9D\x00\x00\x00\x9E\x00\x00\x00\x9F\x00\x00\x00\xA0\x00"
-     "\x00\x00"
-     "\xA1\x00\x00\x00\xA2\x00\x00\x00\xA3\x00\x00\x00\xA4\x00\x00\x00\xA5\x00"
-     "\x00\x00"
-     "\xA6\x00\x00\x00\xA7\x00\x00\x00\xA8\x00\x00\x00\xA9\x00\x00\x00\xAA\x00"
-     "\x00\x00"
-     "\xAB\x00\x00\x00\xAC\x00\x00\x00\xAD\x00\x00\x00\xAE\x00\x00\x00\xAF\x00"
-     "\x00\x00"
-     "\xB0\x00\x00\x00\xB1\x00\x00\x00\xB2\x00\x00\x00\xB3\x00\x00\x00\xB4\x00"
-     "\x00\x00"
-     "\xB5\x00\x00\x00\xB6\x00\x00\x00\xB7\x00\x00\x00\xB8\x00\x00\x00\xB9\x00"
-     "\x00\x00"
-     "\xBA\x00\x00\x00\xBB\x00\x00\x00\xBC\x00\x00\x00\xBD\x00\x00\x00\xBE\x00"
-     "\x00\x00"
-     "\xBF\x00\x00\x00\xC0\x00\x00\x00\xC1\x00\x00\x00\xC2\x00\x00\x00\xC3\x00"
-     "\x00\x00"
-     "\xC4\x00\x00\x00\xC5\x00\x00\x00\xC6\x00\x00\x00\xC7\x00\x00\x00\xC8\x00"
-     "\x00\x00"
-     "\xC9\x00\x00\x00\xCA\x00\x00\x00\xCB\x00\x00\x00\xCC\x00\x00\x00\xCD\x00"
-     "\x00\x00"
-     "\xCE\x00\x00\x00\xCF\x00\x00\x00\xD0\x00\x00\x00\xD1\x00\x00\x00\xD2\x00"
-     "\x00\x00"
-     "\xD3\x00\x00\x00\xD4\x00\x00\x00\xD5\x00\x00\x00\xD6\x00\x00\x00\xD7\x00"
-     "\x00\x00"
-     "\xD8\x00\x00\x00\xD9\x00\x00\x00\xDA\x00\x00\x00\xDB\x00\x00\x00\xDC\x00"
-     "\x00\x00"
-     "\xDD\x00\x00\x00\xDE\x00\x00\x00\xDF\x00\x00\x00\xE0\x00\x00\x00\xE1\x00"
-     "\x00\x00"
-     "\xE2\x00\x00\x00\xE3\x00\x00\x00\xE4\x00\x00\x00\xE5\x00\x00\x00\xE6\x00"
-     "\x00\x00"
-     "\xE7\x00\x00\x00\xE8\x00\x00\x00\xE9\x00\x00\x00\xEA\x00\x00\x00\xEB\x00"
-     "\x00\x00"
-     "\xEC\x00\x00\x00\xED\x00\x00\x00\xEE\x00\x00\x00\xEF\x00\x00\x00\xF0\x00"
-     "\x00\x00"
-     "\xF1\x00\x00\x00\xF2\x00\x00\x00\xF3\x00\x00\x00\xF4\x00\x00\x00\xF5\x00"
-     "\x00\x00"
-     "\xF6\x00\x00\x00\xF7\x00\x00\x00\xF8\x00\x00\x00\xF9\x00\x00\x00\xFA\x00"
-     "\x00\x00"
-     "\xFB\x00\x00\x00\xFC\x00\x00\x00\xFD\x00\x00\x00\xFE\x00\x00\x00\xFF\x00"
-     "\x00\x00"
-     "\x20\x00\x20\xCC\x81\x00\x20\xCC\x83\x00\x20\xCC\x84\x00\x20\xCC\x85\x00"
-     "\x20\xCC"
-     "\x86\x00\x20\xCC\x87\x00\x20\xCC\x88\x00\x20\xCC\x88\xCC\x80\x00\x20\xCC"
-     "\x88\xCC"
-     "\x81\x00\x20\xCC\x88\xCD\x82\x00\x20\xCC\x8A\x00\x20\xCC\x8B\x00\x20\xCC"
-     "\x93\x00"
-     "\x20\xCC\x93\xCC\x80\x00\x20\xCC\x93\xCC\x81\x00\x20\xCC\x93\xCD\x82\x00"
-     "\x20\xCC"
-     "\x94\x00\x20\xCC\x94\xCC\x80\x00\x20\xCC\x94\xCC\x81\x00\x20\xCC\x94\xCD"
-     "\x82\x00"
-     "\x20\xCC\xA7\x00\x20\xCC\xA8\x00\x20\xCC\xB3\x00\x20\xCD\x82\x00\x20\xCE"
-     "\xB9\x00"
-     "\x20\xD9\x8B\x00\x20\xD9\x8C\x00\x20\xD9\x8C\xD9\x91\x00\x20\xD9\x8D\x00"
-     "\x20\xD9"
-     "\x8D\xD9\x91\x00\x20\xD9\x8E\x00\x20\xD9\x8E\xD9\x91\x00\x20\xD9\x8F\x00"
-     "\x20\xD9"
-     "\x8F\xD9\x91\x00\x20\xD9\x90\x00\x20\xD9\x90\xD9\x91\x00\x20\xD9\x91\x00"
-     "\x20\xD9"
-     "\x91\xD9\xB0\x00\x20\xD9\x92\x00\x20\xE3\x82\x99\x00\x20\xE3\x82\x9A\x00"
-     "\x21\x00"
-     "\x21\x21\x00\x21\x3F\x00\x22\x00\x23\x00\x24\x00\x25\x00\x26\x00\x27\x00"
-     "\x28\x00"
-     "\x28\x31\x29\x00\x28\x31\x30\x29\x00\x28\x31\x31\x29\x00\x28\x31\x32\x29"
-     "\x00\x28"
-     "\x31\x33\x29\x00\x28\x31\x34\x29\x00\x28\x31\x35\x29\x00\x28\x31\x36\x29"
-     "\x00\x28"
-     "\x31\x37\x29\x00\x28\x31\x38\x29\x00\x28\x31\x39\x29\x00\x28\x32\x29\x00"
-     "\x28\x32"
-     "\x30\x29\x00\x28\x33\x29\x00\x28\x34\x29\x00\x28\x35\x29\x00\x28\x36\x29"
-     "\x00\x28"
-     "\x37\x29\x00\x28\x38\x29\x00\x28\x39\x29\x00\x28\x61\x29\x00\x28\x62\x29"
-     "\x00\x28"
-     "\x63\x29\x00\x28\x64\x29\x00\x28\x65\x29\x00\x28\x66\x29\x00\x28\x67\x29"
-     "\x00\x28"
-     "\x68\x29\x00\x28\x69\x29\x00\x28\x6A\x29\x00\x28\x6B\x29\x00\x28\x6C\x29"
-     "\x00\x28"
-     "\x6D\x29\x00\x28\x6E\x29\x00\x28\x6F\x29\x00\x28\x70\x29\x00\x28\x71\x29"
-     "\x00\x28"
-     "\x72\x29\x00\x28\x73\x29\x00\x28\x74\x29\x00\x28\x75\x29\x00\x28\x76\x29"
-     "\x00\x28"
-     "\x77\x29\x00\x28\x78\x29\x00\x28\x79\x29\x00\x28\x7A\x29\x00\x28\xE1\x84"
-     "\x80\x29"
-     "\x00\x28\xE1\x84\x82\x29\x00\x28\xE1\x84\x83\x29\x00\x28\xE1\x84\x85\x29"
-     "\x00\x28"
-     "\xE1\x84\x86\x29\x00\x28\xE1\x84\x87\x29\x00\x28\xE1\x84\x89\x29\x00\x28"
-     "\xE1\x84"
-     "\x8B\x29\x00\x28\xE1\x84\x8C\x29\x00\x28\xE1\x84\x8E\x29\x00\x28\xE1\x84"
-     "\x8F\x29"
-     "\x00\x28\xE1\x84\x90\x29\x00\x28\xE1\x84\x91\x29\x00\x28\xE1\x84\x92\x29"
-     "\x00\x28"
-     "\xE4\xB8\x80\x29\x00\x28\xE4\xB8\x83\x29\x00\x28\xE4\xB8\x89\x29\x00\x28"
-     "\xE4\xB9"
-     "\x9D\x29\x00\x28\xE4\xBA\x8C\x29\x00\x28\xE4\xBA\x94\x29\x00\x28\xE4\xBB"
-     "\xA3\x29"
-     "\x00\x28\xE4\xBC\x81\x29\x00\x28\xE4\xBC\x91\x29\x00\x28\xE5\x85\xAB\x29"
-     "\x00\x28"
-     "\xE5\x85\xAD\x29\x00\x28\xE5\x8A\xB4\x29\x00\x28\xE5\x8D\x81\x29\x00\x28"
-     "\xE5\x8D"
-     "\x94\x29\x00\x28\xE5\x90\x8D\x29\x00\x28\xE5\x91\xBC\x29\x00\x28\xE5\x9B"
-     "\x9B\x29"
-     "\x00\x28\xE5\x9C\x9F\x29\x00\x28\xE5\xAD\xA6\x29\x00\x28\xE6\x97\xA5\x29"
-     "\x00\x28"
-     "\xE6\x9C\x88\x29\x00\x28\xE6\x9C\x89\x29\x00\x28\xE6\x9C\xA8\x29\x00\x28"
-     "\xE6\xA0"
-     "\xAA\x29\x00\x28\xE6\xB0\xB4\x29\x00\x28\xE7\x81\xAB\x29\x00\x28\xE7\x89"
-     "\xB9\x29"
-     "\x00\x28\xE7\x9B\xA3\x29\x00\x28\xE7\xA4\xBE\x29\x00\x28\xE7\xA5\x9D\x29"
-     "\x00\x28"
-     "\xE7\xA5\xAD\x29\x00\x28\xE8\x87\xAA\x29\x00\x28\xE8\x87\xB3\x29\x00\x28"
-     "\xE8\xB2"
-     "\xA1\x29\x00\x28\xE8\xB3\x87\x29\x00\x28\xE9\x87\x91\x29\x00\x28\xEA\xB0"
-     "\x80\x29"
-     "\x00\x28\xEB\x82\x98\x29\x00\x28\xEB\x8B\xA4\x29\x00\x28\xEB\x9D\xBC\x29"
-     "\x00\x28"
-     "\xEB\xA7\x88\x29\x00\x28\xEB\xB0\x94\x29\x00\x28\xEC\x82\xAC\x29\x00\x28"
-     "\xEC\x95"
-     "\x84\x29\x00\x28\xEC\x98\xA4\xEC\xA0\x84\x29\x00\x28\xEC\x98\xA4\xED\x9B"
-     "\x84\x29"
-     "\x00\x28\xEC\x9E\x90\x29\x00\x28\xEC\xA3\xBC\x29\x00\x28\xEC\xB0\xA8\x29"
-     "\x00\x28"
-     "\xEC\xB9\xB4\x29\x00\x28\xED\x83\x80\x29\x00\x28\xED\x8C\x8C\x29\x00\x28"
-     "\xED\x95"
-     "\x98\x29\x00\x29\x00\x2A\x00\x2B\x00\x2C\x00\x2D\x00\x2E\x00\x2E\x2E\x00"
-     "\x2E\x2E"
-     "\x2E\x00\x2F\x00\x30\x00\x30\x2C\x00\x30\x2E\x00\x30\xE2\x81\x84\x33\x00"
-     "\x30\xE7"
-     "\x82\xB9\x00\x31\x00\x31\x2C\x00\x31\x2E\x00\x31\x30\x00\x31\x30\x2E\x00"
-     "\x31\x30"
-     "\xE6\x97\xA5\x00\x31\x30\xE6\x9C\x88\x00\x31\x30\xE7\x82\xB9\x00\x31\x31"
-     "\x00\x31"
-     "\x31\x2E\x00\x31\x31\xE6\x97\xA5\x00\x31\x31\xE6\x9C\x88\x00\x31\x31\xE7"
-     "\x82\xB9"
-     "\x00\x31\x32\x00\x31\x32\x2E\x00\x31\x32\xE6\x97\xA5\x00\x31\x32\xE6\x9C"
-     "\x88\x00"
-     "\x31\x32\xE7\x82\xB9\x00\x31\x33\x00\x31\x33\x2E\x00\x31\x33\xE6\x97\xA5"
-     "\x00\x31"
-     "\x33\xE7\x82\xB9\x00\x31\x34\x00\x31\x34\x2E\x00\x31\x34\xE6\x97\xA5\x00"
-     "\x31\x34"
-     "\xE7\x82\xB9\x00\x31\x35\x00\x31\x35\x2E\x00\x31\x35\xE6\x97\xA5\x00\x31"
-     "\x35\xE7"
-     "\x82\xB9\x00\x31\x36\x00\x31\x36\x2E\x00\x31\x36\xE6\x97\xA5\x00\x31\x36"
-     "\xE7\x82"
-     "\xB9\x00\x31\x37\x00\x31\x37\x2E\x00\x31\x37\xE6\x97\xA5\x00\x31\x37\xE7"
-     "\x82\xB9"
-     "\x00\x31\x38\x00\x31\x38\x2E\x00\x31\x38\xE6\x97\xA5\x00\x31\x38\xE7\x82"
-     "\xB9\x00"
-     "\x31\x39\x00\x31\x39\x2E\x00\x31\x39\xE6\x97\xA5\x00\x31\x39\xE7\x82\xB9"
-     "\x00\x31"
-     "\xE2\x81\x84\x00\x31\xE2\x81\x84\x31\x30\x00\x31\xE2\x81\x84\x32\x00\x31"
-     "\xE2\x81"
-     "\x84\x33\x00\x31\xE2\x81\x84\x34\x00\x31\xE2\x81\x84\x35\x00\x31\xE2\x81"
-     "\x84\x36"
-     "\x00\x31\xE2\x81\x84\x37\x00\x31\xE2\x81\x84\x38\x00\x31\xE2\x81\x84\x39"
-     "\x00\x31"
-     "\xE6\x97\xA5\x00\x31\xE6\x9C\x88\x00\x31\xE7\x82\xB9\x00\x32\x00\x32\x2C"
-     "\x00\x32"
-     "\x2E\x00\x32\x30\x00\x32\x30\x2E\x00\x32\x30\xE6\x97\xA5\x00\x32\x30\xE7"
-     "\x82\xB9"
-     "\x00\x32\x31\x00\x32\x31\xE6\x97\xA5\x00\x32\x31\xE7\x82\xB9\x00\x32\x32"
-     "\x00\x32"
-     "\x32\xE6\x97\xA5\x00\x32\x32\xE7\x82\xB9\x00\x32\x33\x00\x32\x33\xE6\x97"
-     "\xA5\x00"
-     "\x32\x33\xE7\x82\xB9\x00\x32\x34\x00\x32\x34\xE6\x97\xA5\x00\x32\x34\xE7"
-     "\x82\xB9"
-     "\x00\x32\x35\x00\x32\x35\xE6\x97\xA5\x00\x32\x36\x00\x32\x36\xE6\x97\xA5"
-     "\x00\x32"
-     "\x37\x00\x32\x37\xE6\x97\xA5\x00\x32\x38\x00\x32\x38\xE6\x97\xA5\x00\x32"
-     "\x39\x00"
-     "\x32\x39\xE6\x97\xA5\x00\x32\xE2\x81\x84\x33\x00\x32\xE2\x81\x84\x35\x00"
-     "\x32\xE6"
-     "\x97\xA5\x00\x32\xE6\x9C\x88\x00\x32\xE7\x82\xB9\x00\x33\x00\x33\x2C\x00"
-     "\x33\x2E"
-     "\x00\x33\x30\x00\x33\x30\xE6\x97\xA5\x00\x33\x31\x00\x33\x31\xE6\x97\xA5"
-     "\x00\x33"
-     "\x32\x00\x33\x33\x00\x33\x34\x00\x33\x35\x00\x33\x36\x00\x33\x37\x00\x33"
-     "\x38\x00"
-     "\x33\x39\x00\x33\xE2\x81\x84\x34\x00\x33\xE2\x81\x84\x35\x00\x33\xE2\x81"
-     "\x84\x38"
-     "\x00\x33\xE6\x97\xA5\x00\x33\xE6\x9C\x88\x00\x33\xE7\x82\xB9\x00\x34\x00"
-     "\x34\x2C"
-     "\x00\x34\x2E\x00\x34\x30\x00\x34\x31\x00\x34\x32\x00\x34\x33\x00\x34\x34"
-     "\x00\x34"
-     "\x35\x00\x34\x36\x00\x34\x37\x00\x34\x38\x00\x34\x39\x00\x34\xE2\x81\x84"
-     "\x35\x00"
-     "\x34\xE6\x97\xA5\x00\x34\xE6\x9C\x88\x00\x34\xE7\x82\xB9\x00\x35\x00\x35"
-     "\x2C\x00"
-     "\x35\x2E\x00\x35\x30\x00\x35\xE2\x81\x84\x36\x00\x35\xE2\x81\x84\x38\x00"
-     "\x35\xE6"
-     "\x97\xA5\x00\x35\xE6\x9C\x88\x00\x35\xE7\x82\xB9\x00\x36\x00\x36\x2C\x00"
-     "\x36\x2E"
-     "\x00\x36\xE6\x97\xA5\x00\x36\xE6\x9C\x88\x00\x36\xE7\x82\xB9\x00\x37\x00"
-     "\x37\x2C"
-     "\x00\x37\x2E\x00\x37\xE2\x81\x84\x38\x00\x37\xE6\x97\xA5\x00\x37\xE6\x9C"
-     "\x88\x00"
-     "\x37\xE7\x82\xB9\x00\x38\x00\x38\x2C\x00\x38\x2E\x00\x38\xE6\x97\xA5\x00"
-     "\x38\xE6"
-     "\x9C\x88\x00\x38\xE7\x82\xB9\x00\x39\x00\x39\x2C\x00\x39\x2E\x00\x39\xE6"
-     "\x97\xA5"
-     "\x00\x39\xE6\x9C\x88\x00\x39\xE7\x82\xB9\x00\x3A\x00\x3A\x3A\x3D\x00\x3B"
-     "\x00\x3C"
-     "\x00\x3D\x00\x3D\x3D\x00\x3D\x3D\x3D\x00\x3E\x00\x3F\x00\x3F\x21\x00\x3F"
-     "\x3F\x00"
-     "\x40\x00\x5B\x00\x5C\x00\x5D\x00\x5E\x00\x5F\x00\x60\x00\x61\x00\x61\x2E"
-     "\x6D\x2E"
-     "\x00\x61\x2F\x63\x00\x61\x2F\x73\x00\x61\x75\x00\x61\xCA\xBE\x00\x61\xE2"
-     "\x88\x95"
-     "\x6D\x00\x62\x00\x62\x61\x72\x00\x62\x71\x00\x63\x00\x63\x2F\x6F\x00\x63"
-     "\x2F\x75"
-     "\x00\x63\x61\x6C\x00\x63\x63\x00\x63\x64\x00\x63\x6D\x00\x63\x6D\x32\x00"
-     "\x63\x6D"
-     "\x33\x00\x63\x6F\x2E\x00\x63\xE2\x88\x95\x6B\x67\x00\x64\x00\x64\x61\x00"
-     "\x64\x62"
-     "\x00\x64\x6A\x00\x64\x6C\x00\x64\x6D\x00\x64\x6D\x32\x00\x64\x6D\x33\x00"
-     "\x64\x7A"
-     "\x00\x64\xC5\xBE\x00\x65\x00\x65\x72\x67\x00\x65\x76\x00\x66\x00\x66\x61"
-     "\x78\x00"
-     "\x66\x66\x00\x66\x66\x69\x00\x66\x66\x6C\x00\x66\x69\x00\x66\x6C\x00\x66"
-     "\x6D\x00"
-     "\x67\x00\x67\x61\x6C\x00\x67\x62\x00\x67\x68\x7A\x00\x67\x70\x61\x00\x67"
-     "\x79\x00"
-     "\x68\x00\x68\x61\x00\x68\x67\x00\x68\x70\x00\x68\x70\x61\x00\x68\x76\x00"
-     "\x68\x7A"
-     "\x00\x69\x00\x69\x69\x00\x69\x69\x69\x00\x69\x6A\x00\x69\x6E\x00\x69\x75"
-     "\x00\x69"
-     "\x76\x00\x69\x78\x00\x6A\x00\x6B\x00\x6B\x61\x00\x6B\x62\x00\x6B\x63\x61"
-     "\x6C\x00"
-     "\x6B\x67\x00\x6B\x68\x7A\x00\x6B\x6B\x00\x6B\x6C\x00\x6B\x6D\x00\x6B\x6D"
-     "\x32\x00"
-     "\x6B\x6D\x33\x00\x6B\x70\x61\x00\x6B\x74\x00\x6B\x76\x00\x6B\x77\x00\x6B"
-     "\xCF\x89"
-     "\x00\x6C\x00\x6C\x6A\x00\x6C\x6D\x00\x6C\x6E\x00\x6C\x6F\x67\x00\x6C\x74"
-     "\x64\x00"
-     "\x6C\x78\x00\x6C\xC2\xB7\x00\x6D\x00\x6D\x32\x00\x6D\x33\x00\x6D\x61\x00"
-     "\x6D\x62"
-     "\x00\x6D\x63\x00\x6D\x64\x00\x6D\x67\x00\x6D\x68\x7A\x00\x6D\x69\x6C\x00"
-     "\x6D\x6C"
-     "\x00\x6D\x6D\x00\x6D\x6D\x32\x00\x6D\x6D\x33\x00\x6D\x6F\x6C\x00\x6D\x70"
-     "\x61\x00"
-     "\x6D\x72\x00\x6D\x73\x00\x6D\x76\x00\x6D\x77\x00\x6D\xCF\x89\x00\x6D\xE2"
-     "\x88\x95"
-     "\x73\x00\x6D\xE2\x88\x95\x73\x32\x00\x6E\x00\x6E\x61\x00\x6E\x66\x00\x6E"
-     "\x6A\x00"
-     "\x6E\x6D\x00\x6E\x6F\x00\x6E\x73\x00\x6E\x76\x00\x6E\x77\x00\x6F\x00\x6F"
-     "\x76\x00"
-     "\x70\x00\x70\x2E\x6D\x2E\x00\x70\x61\x00\x70\x63\x00\x70\x66\x00\x70\x68"
-     "\x00\x70"
-     "\x70\x6D\x00\x70\x70\x76\x00\x70\x72\x00\x70\x73\x00\x70\x74\x65\x00\x70"
-     "\x76\x00"
-     "\x70\x77\x00\x71\x00\x72\x00\x72\x61\x64\x00\x72\x61\x64\xE2\x88\x95\x73"
-     "\x00\x72"
-     "\x61\x64\xE2\x88\x95\x73\x32\x00\x72\x73\x00\x73\x00\x73\x64\x00\x73\x6D"
-     "\x00\x73"
-     "\x72\x00\x73\x73\x00\x73\x74\x00\x73\x76\x00\x74\x00\x74\x65\x6C\x00\x74"
-     "\x68\x7A"
-     "\x00\x74\x6D\x00\x75\x00\x76\x00\x76\x69\x00\x76\x69\x69\x00\x76\x69\x69"
-     "\x69\x00"
-     "\x76\xE2\x88\x95\x6D\x00\x77\x00\x77\x62\x00\x77\x63\x00\x77\x7A\x00\x78"
-     "\x00\x78"
-     "\x69\x00\x78\x69\x69\x00\x79\x00\x7A\x00\x7B\x00\x7C\x00\x7D\x00\x7E\x00"
-     "\xC2\xA2"
-     "\x00\xC2\xA3\x00\xC2\xA5\x00\xC2\xA6\x00\xC2\xAC\x00\xC2\xB0\x63\x00\xC2"
-     "\xB0\x66"
-     "\x00\xC2\xB7\x00\xC3\x9F\x00\xC3\xA0\x00\xC3\xA1\x00\xC3\xA2\x00\xC3\xA3"
-     "\x00\xC3"
-     "\xA4\x00\xC3\xA5\x00\xC3\xA6\x00\xC3\xA7\x00\xC3\xA8\x00\xC3\xA9\x00\xC3"
-     "\xAA\x00"
-     "\xC3\xAB\x00\xC3\xAC\x00\xC3\xAD\x00\xC3\xAE\x00\xC3\xAF\x00\xC3\xB0\x00"
-     "\xC3\xB1"
-     "\x00\xC3\xB2\x00\xC3\xB3\x00\xC3\xB4\x00\xC3\xB5\x00\xC3\xB6\x00\xC3\xB8"
-     "\x00\xC3"
-     "\xB9\x00\xC3\xBA\x00\xC3\xBB\x00\xC3\xBC\x00\xC3\xBD\x00\xC3\xBE\x00\xC3"
-     "\xBF\x00"
-     "\xC4\x81\x00\xC4\x83\x00\xC4\x85\x00\xC4\x87\x00\xC4\x89\x00\xC4\x8B\x00"
-     "\xC4\x8D"
-     "\x00\xC4\x8F\x00\xC4\x91\x00\xC4\x93\x00\xC4\x95\x00\xC4\x97\x00\xC4\x99"
-     "\x00\xC4"
-     "\x9B\x00\xC4\x9D\x00\xC4\x9F\x00\xC4\xA1\x00\xC4\xA3\x00\xC4\xA5\x00\xC4"
-     "\xA7\x00"
-     "\xC4\xA9\x00\xC4\xAB\x00\xC4\xAD\x00\xC4\xAF\x00\xC4\xB0\x00\xC4\xB1\x00"
-     "\xC4\xB5"
-     "\x00\xC4\xB7\x00\xC4\xBA\x00\xC4\xBC\x00\xC4\xBE\x00\xC5\x82\x00\xC5\x84"
-     "\x00\xC5"
-     "\x86\x00\xC5\x88\x00\xC5\x8B\x00\xC5\x8D\x00\xC5\x8F\x00\xC5\x91\x00\xC5"
-     "\x93\x00"
-     "\xC5\x95\x00\xC5\x97\x00\xC5\x99\x00\xC5\x9B\x00\xC5\x9D\x00\xC5\x9F\x00"
-     "\xC5\xA1"
-     "\x00\xC5\xA3\x00\xC5\xA5\x00\xC5\xA7\x00\xC5\xA9\x00\xC5\xAB\x00\xC5\xAD"
-     "\x00\xC5"
-     "\xAF\x00\xC5\xB1\x00\xC5\xB3\x00\xC5\xB5\x00\xC5\xB7\x00\xC5\xBA\x00\xC5"
-     "\xBC\x00"
-     "\xC5\xBE\x00\xC6\x80\x00\xC6\x83\x00\xC6\x85\x00\xC6\x88\x00\xC6\x8C\x00"
-     "\xC6\x92"
-     "\x00\xC6\x95\x00\xC6\x99\x00\xC6\x9A\x00\xC6\x9E\x00\xC6\xA1\x00\xC6\xA3"
-     "\x00\xC6"
-     "\xA5\x00\xC6\xA8\x00\xC6\xAB\x00\xC6\xAD\x00\xC6\xB0\x00\xC6\xB4\x00\xC6"
-     "\xB6\x00"
-     "\xC6\xB9\x00\xC6\xBD\x00\xC6\xBF\x00\xC7\x80\x00\xC7\x81\x00\xC7\x82\x00"
-     "\xC7\x8E"
-     "\x00\xC7\x90\x00\xC7\x92\x00\xC7\x94\x00\xC7\x96\x00\xC7\x98\x00\xC7\x9A"
-     "\x00\xC7"
-     "\x9C\x00\xC7\x9D\x00\xC7\x9F\x00\xC7\xA1\x00\xC7\xA3\x00\xC7\xA5\x00\xC7"
-     "\xA7\x00"
-     "\xC7\xA9\x00\xC7\xAB\x00\xC7\xAD\x00\xC7\xAF\x00\xC7\xB0\x00\xC7\xB5\x00"
-     "\xC7\xB9"
-     "\x00\xC7\xBB\x00\xC7\xBD\x00\xC7\xBF\x00\xC8\x81\x00\xC8\x83\x00\xC8\x85"
-     "\x00\xC8"
-     "\x87\x00\xC8\x89\x00\xC8\x8B\x00\xC8\x8D\x00\xC8\x8F\x00\xC8\x91\x00\xC8"
-     "\x93\x00"
-     "\xC8\x95\x00\xC8\x97\x00\xC8\x99\x00\xC8\x9B\x00\xC8\x9D\x00\xC8\x9F\x00"
-     "\xC8\xA3"
-     "\x00\xC8\xA5\x00\xC8\xA7\x00\xC8\xA9\x00\xC8\xAB\x00\xC8\xAD\x00\xC8\xAF"
-     "\x00\xC8"
-     "\xB1\x00\xC8\xB3\x00\xC8\xB7\x00\xC8\xBC\x00\xC8\xBF\x00\xC9\x80\x00\xC9"
-     "\x82\x00"
-     "\xC9\x87\x00\xC9\x89\x00\xC9\x8B\x00\xC9\x8D\x00\xC9\x8F\x00\xC9\x90\x00"
-     "\xC9\x91"
-     "\x00\xC9\x92\x00\xC9\x93\x00\xC9\x94\x00\xC9\x95\x00\xC9\x96\x00\xC9\x97"
-     "\x00\xC9"
-     "\x98\x00\xC9\x99\x00\xC9\x9B\x00\xC9\x9C\x00\xC9\x9E\x00\xC9\x9F\x00\xC9"
-     "\xA0\x00"
-     "\xC9\xA1\x00\xC9\xA2\x00\xC9\xA3\x00\xC9\xA4\x00\xC9\xA5\x00\xC9\xA6\x00"
-     "\xC9\xA7"
-     "\x00\xC9\xA8\x00\xC9\xA9\x00\xC9\xAA\x00\xC9\xAB\x00\xC9\xAC\x00\xC9\xAD"
-     "\x00\xC9"
-     "\xAE\x00\xC9\xAF\x00\xC9\xB0\x00\xC9\xB1\x00\xC9\xB2\x00\xC9\xB3\x00\xC9"
-     "\xB4\x00"
-     "\xC9\xB5\x00\xC9\xB6\x00\xC9\xB7\x00\xC9\xB8\x00\xC9\xB9\x00\xC9\xBA\x00"
-     "\xC9\xBB"
-     "\x00\xC9\xBD\x00\xC9\xBE\x00\xCA\x80\x00\xCA\x81\x00\xCA\x82\x00\xCA\x83"
-     "\x00\xCA"
-     "\x84\x00\xCA\x87\x00\xCA\x88\x00\xCA\x89\x00\xCA\x8A\x00\xCA\x8B\x00\xCA"
-     "\x8C\x00"
-     "\xCA\x8D\x00\xCA\x8E\x00\xCA\x8F\x00\xCA\x90\x00\xCA\x91\x00\xCA\x92\x00"
-     "\xCA\x95"
-     "\x00\xCA\x98\x00\xCA\x99\x00\xCA\x9B\x00\xCA\x9C\x00\xCA\x9D\x00\xCA\x9E"
-     "\x00\xCA"
-     "\x9F\x00\xCA\xA1\x00\xCA\xA2\x00\xCA\xA3\x00\xCA\xA4\x00\xCA\xA5\x00\xCA"
-     "\xA6\x00"
-     "\xCA\xA7\x00\xCA\xA8\x00\xCA\xA9\x00\xCA\xAA\x00\xCA\xAB\x00\xCA\xB9\x00"
-     "\xCA\xBC"
-     "\x6E\x00\xCB\x90\x00\xCB\x91\x00\xCC\x80\x00\xCC\x81\x00\xCC\x88\xCC\x81"
-     "\x00\xCC"
-     "\x93\x00\xCD\xB1\x00\xCD\xB3\x00\xCD\xB7\x00\xCD\xBB\x00\xCD\xBC\x00\xCD"
-     "\xBD\x00"
-     "\xCE\x90\x00\xCE\xAC\x00\xCE\xAD\x00\xCE\xAE\x00\xCE\xAF\x00\xCE\xB0\x00"
-     "\xCE\xB1"
-     "\x00\xCE\xB2\x00\xCE\xB3\x00\xCE\xB4\x00\xCE\xB5\x00\xCE\xB6\x00\xCE\xB7"
-     "\x00\xCE"
-     "\xB8\x00\xCE\xB9\x00\xCE\xBA\x00\xCE\xBB\x00\xCE\xBC\x00\xCE\xBC\x61\x00"
-     "\xCE\xBC"
-     "\x66\x00\xCE\xBC\x67\x00\xCE\xBC\x6C\x00\xCE\xBC\x6D\x00\xCE\xBC\x73\x00"
-     "\xCE\xBC"
-     "\x76\x00\xCE\xBC\x77\x00\xCE\xBD\x00\xCE\xBE\x00\xCE\xBF\x00\xCF\x80\x00"
-     "\xCF\x81"
-     "\x00\xCF\x83\x00\xCF\x84\x00\xCF\x85\x00\xCF\x86\x00\xCF\x87\x00\xCF\x88"
-     "\x00\xCF"
-     "\x89\x00\xCF\x8A\x00\xCF\x8B\x00\xCF\x8C\x00\xCF\x8D\x00\xCF\x8E\x00\xCF"
-     "\x97\x00"
-     "\xCF\x99\x00\xCF\x9B\x00\xCF\x9D\x00\xCF\x9F\x00\xCF\xA1\x00\xCF\xA3\x00"
-     "\xCF\xA5"
-     "\x00\xCF\xA7\x00\xCF\xA9\x00\xCF\xAB\x00\xCF\xAD\x00\xCF\xAF\x00\xCF\xB3"
-     "\x00\xCF"
-     "\xB8\x00\xCF\xBB\x00\xD0\xB0\x00\xD0\xB1\x00\xD0\xB2\x00\xD0\xB3\x00\xD0"
-     "\xB4\x00"
-     "\xD0\xB5\x00\xD0\xB6\x00\xD0\xB7\x00\xD0\xB8\x00\xD0\xB9\x00\xD0\xBA\x00"
-     "\xD0\xBB"
-     "\x00\xD0\xBC\x00\xD0\xBD\x00\xD0\xBE\x00\xD0\xBF\x00\xD1\x80\x00\xD1\x81"
-     "\x00\xD1"
-     "\x82\x00\xD1\x83\x00\xD1\x84\x00\xD1\x85\x00\xD1\x86\x00\xD1\x87\x00\xD1"
-     "\x88\x00"
-     "\xD1\x89\x00\xD1\x8A\x00\xD1\x8B\x00\xD1\x8C\x00\xD1\x8D\x00\xD1\x8E\x00"
-     "\xD1\x8F"
-     "\x00\xD1\x90\x00\xD1\x91\x00\xD1\x92\x00\xD1\x93\x00\xD1\x94\x00\xD1\x95"
-     "\x00\xD1"
-     "\x96\x00\xD1\x97\x00\xD1\x98\x00\xD1\x99\x00\xD1\x9A\x00\xD1\x9B\x00\xD1"
-     "\x9C\x00"
-     "\xD1\x9D\x00\xD1\x9E\x00\xD1\x9F\x00\xD1\xA1\x00\xD1\xA3\x00\xD1\xA5\x00"
-     "\xD1\xA7"
-     "\x00\xD1\xA9\x00\xD1\xAB\x00\xD1\xAD\x00\xD1\xAF\x00\xD1\xB1\x00\xD1\xB3"
-     "\x00\xD1"
-     "\xB5\x00\xD1\xB7\x00\xD1\xB9\x00\xD1\xBB\x00\xD1\xBD\x00\xD1\xBF\x00\xD2"
-     "\x81\x00"
-     "\xD2\x8B\x00\xD2\x8D\x00\xD2\x8F\x00\xD2\x91\x00\xD2\x93\x00\xD2\x95\x00"
-     "\xD2\x97"
-     "\x00\xD2\x99\x00\xD2\x9B\x00\xD2\x9D\x00\xD2\x9F\x00\xD2\xA1\x00\xD2\xA3"
-     "\x00\xD2"
-     "\xA5\x00\xD2\xA7\x00\xD2\xA9\x00\xD2\xAB\x00\xD2\xAD\x00\xD2\xAF\x00\xD2"
-     "\xB1\x00"
-     "\xD2\xB3\x00\xD2\xB5\x00\xD2\xB7\x00\xD2\xB9\x00\xD2\xBB\x00\xD2\xBD\x00"
-     "\xD2\xBF"
-     "\x00\xD3\x82\x00\xD3\x84\x00\xD3\x86\x00\xD3\x88\x00\xD3\x8A\x00\xD3\x8C"
-     "\x00\xD3"
-     "\x8E\x00\xD3\x8F\x00\xD3\x91\x00\xD3\x93\x00\xD3\x95\x00\xD3\x97\x00\xD3"
-     "\x99\x00"
-     "\xD3\x9B\x00\xD3\x9D\x00\xD3\x9F\x00\xD3\xA1\x00\xD3\xA3\x00\xD3\xA5\x00"
-     "\xD3\xA7"
-     "\x00\xD3\xA9\x00\xD3\xAB\x00\xD3\xAD\x00\xD3\xAF\x00\xD3\xB1\x00\xD3\xB3"
-     "\x00\xD3"
-     "\xB5\x00\xD3\xB7\x00\xD3\xB9\x00\xD3\xBB\x00\xD3\xBD\x00\xD3\xBF\x00\xD4"
-     "\x81\x00"
-     "\xD4\x83\x00\xD4\x85\x00\xD4\x87\x00\xD4\x89\x00\xD4\x8B\x00\xD4\x8D\x00"
-     "\xD4\x8F"
-     "\x00\xD4\x91\x00\xD4\x93\x00\xD4\x95\x00\xD4\x97\x00\xD4\x99\x00\xD4\x9B"
-     "\x00\xD4"
-     "\x9D\x00\xD4\x9F\x00\xD4\xA1\x00\xD4\xA3\x00\xD4\xA5\x00\xD4\xA7\x00\xD4"
-     "\xA9\x00"
-     "\xD4\xAB\x00\xD4\xAD\x00\xD4\xAF\x00\xD5\xA1\x00\xD5\xA2\x00\xD5\xA3\x00"
-     "\xD5\xA4"
-     "\x00\xD5\xA5\x00\xD5\xA5\xD6\x82\x00\xD5\xA6\x00\xD5\xA7\x00\xD5\xA8\x00"
-     "\xD5\xA9"
-     "\x00\xD5\xAA\x00\xD5\xAB\x00\xD5\xAC\x00\xD5\xAD\x00\xD5\xAE\x00\xD5\xAF"
-     "\x00\xD5"
-     "\xB0\x00\xD5\xB1\x00\xD5\xB2\x00\xD5\xB3\x00\xD5\xB4\x00\xD5\xB4\xD5\xA5"
-     "\x00\xD5"
-     "\xB4\xD5\xAB\x00\xD5\xB4\xD5\xAD\x00\xD5\xB4\xD5\xB6\x00\xD5\xB5\x00\xD5"
-     "\xB6\x00"
-     "\xD5\xB7\x00\xD5\xB8\x00\xD5\xB9\x00\xD5\xBA\x00\xD5\xBB\x00\xD5\xBC\x00"
-     "\xD5\xBD"
-     "\x00\xD5\xBE\x00\xD5\xBE\xD5\xB6\x00\xD5\xBF\x00\xD6\x80\x00\xD6\x81\x00"
-     "\xD6\x82"
-     "\x00\xD6\x83\x00\xD6\x84\x00\xD6\x85\x00\xD6\x86\x00\xD7\x90\x00\xD7\x90"
-     "\xD6\xB7"
-     "\x00\xD7\x90\xD6\xB8\x00\xD7\x90\xD6\xBC\x00\xD7\x90\xD7\x9C\x00\xD7\x91"
-     "\x00\xD7"
-     "\x91\xD6\xBC\x00\xD7\x91\xD6\xBF\x00\xD7\x92\x00\xD7\x92\xD6\xBC\x00\xD7"
-     "\x93\x00"
-     "\xD7\x93\xD6\xBC\x00\xD7\x94\x00\xD7\x94\xD6\xBC\x00\xD7\x95\xD6\xB9\x00"
-     "\xD7\x95"
-     "\xD6\xBC\x00\xD7\x96\xD6\xBC\x00\xD7\x98\xD6\xBC\x00\xD7\x99\xD6\xB4\x00"
-     "\xD7\x99"
-     "\xD6\xBC\x00\xD7\x9A\xD6\xBC\x00\xD7\x9B\x00\xD7\x9B\xD6\xBC\x00\xD7\x9B"
-     "\xD6\xBF"
-     "\x00\xD7\x9C\x00\xD7\x9C\xD6\xBC\x00\xD7\x9D\x00\xD7\x9E\xD6\xBC\x00\xD7"
-     "\xA0\xD6"
-     "\xBC\x00\xD7\xA1\xD6\xBC\x00\xD7\xA2\x00\xD7\xA3\xD6\xBC\x00\xD7\xA4\xD6"
-     "\xBC\x00"
-     "\xD7\xA4\xD6\xBF\x00\xD7\xA6\xD6\xBC\x00\xD7\xA7\xD6\xBC\x00\xD7\xA8\x00"
-     "\xD7\xA8"
-     "\xD6\xBC\x00\xD7\xA9\xD6\xBC\x00\xD7\xA9\xD6\xBC\xD7\x81\x00\xD7\xA9\xD6"
-     "\xBC\xD7"
-     "\x82\x00\xD7\xA9\xD7\x81\x00\xD7\xA9\xD7\x82\x00\xD7\xAA\x00\xD7\xAA\xD6"
-     "\xBC\x00"
-     "\xD7\xB2\xD6\xB7\x00\xD8\xA1\x00\xD8\xA2\x00\xD8\xA3\x00\xD8\xA4\x00\xD8"
-     "\xA5\x00"
-     "\xD8\xA6\x00\xD8\xA6\xD8\xA7\x00\xD8\xA6\xD8\xAC\x00\xD8\xA6\xD8\xAD\x00"
-     "\xD8\xA6"
-     "\xD8\xAE\x00\xD8\xA6\xD8\xB1\x00\xD8\xA6\xD8\xB2\x00\xD8\xA6\xD9\x85\x00"
-     "\xD8\xA6"
-     "\xD9\x86\x00\xD8\xA6\xD9\x87\x00\xD8\xA6\xD9\x88\x00\xD8\xA6\xD9\x89\x00"
-     "\xD8\xA6"
-     "\xD9\x8A\x00\xD8\xA6\xDB\x86\x00\xD8\xA6\xDB\x87\x00\xD8\xA6\xDB\x88\x00"
-     "\xD8\xA6"
-     "\xDB\x90\x00\xD8\xA6\xDB\x95\x00\xD8\xA7\x00\xD8\xA7\xD9\x83\xD8\xA8\xD8"
-     "\xB1\x00"
-     "\xD8\xA7\xD9\x84\xD9\x84\xD9\x87\x00\xD8\xA7\xD9\x8B\x00\xD8\xA7\xD9\xB4"
-     "\x00\xD8"
-     "\xA8\x00\xD8\xA8\xD8\xAC\x00\xD8\xA8\xD8\xAD\x00\xD8\xA8\xD8\xAD\xD9\x8A"
-     "\x00\xD8"
-     "\xA8\xD8\xAE\x00\xD8\xA8\xD8\xAE\xD9\x8A\x00\xD8\xA8\xD8\xB1\x00\xD8\xA8"
-     "\xD8\xB2"
-     "\x00\xD8\xA8\xD9\x85\x00\xD8\xA8\xD9\x86\x00\xD8\xA8\xD9\x87\x00\xD8\xA8"
-     "\xD9\x89"
-     "\x00\xD8\xA8\xD9\x8A\x00\xD8\xA9\x00\xD8\xAA\x00\xD8\xAA\xD8\xAC\x00\xD8"
-     "\xAA\xD8"
-     "\xAC\xD9\x85\x00\xD8\xAA\xD8\xAC\xD9\x89\x00\xD8\xAA\xD8\xAC\xD9\x8A\x00"
-     "\xD8\xAA"
-     "\xD8\xAD\x00\xD8\xAA\xD8\xAD\xD8\xAC\x00\xD8\xAA\xD8\xAD\xD9\x85\x00\xD8"
-     "\xAA\xD8"
-     "\xAE\x00\xD8\xAA\xD8\xAE\xD9\x85\x00\xD8\xAA\xD8\xAE\xD9\x89\x00\xD8\xAA"
-     "\xD8\xAE"
-     "\xD9\x8A\x00\xD8\xAA\xD8\xB1\x00\xD8\xAA\xD8\xB2\x00\xD8\xAA\xD9\x85\x00"
-     "\xD8\xAA"
-     "\xD9\x85\xD8\xAC\x00\xD8\xAA\xD9\x85\xD8\xAD\x00\xD8\xAA\xD9\x85\xD8\xAE"
-     "\x00\xD8"
-     "\xAA\xD9\x85\xD9\x89\x00\xD8\xAA\xD9\x85\xD9\x8A\x00\xD8\xAA\xD9\x86\x00"
-     "\xD8\xAA"
-     "\xD9\x87\x00\xD8\xAA\xD9\x89\x00\xD8\xAA\xD9\x8A\x00\xD8\xAB\x00\xD8\xAB"
-     "\xD8\xAC"
-     "\x00\xD8\xAB\xD8\xB1\x00\xD8\xAB\xD8\xB2\x00\xD8\xAB\xD9\x85\x00\xD8\xAB"
-     "\xD9\x86"
-     "\x00\xD8\xAB\xD9\x87\x00\xD8\xAB\xD9\x89\x00\xD8\xAB\xD9\x8A\x00\xD8\xAC"
-     "\x00\xD8"
-     "\xAC\xD8\xAD\x00\xD8\xAC\xD8\xAD\xD9\x89\x00\xD8\xAC\xD8\xAD\xD9\x8A\x00"
-     "\xD8\xAC"
-     "\xD9\x84\x20\xD8\xAC\xD9\x84\xD8\xA7\xD9\x84\xD9\x87\x00\xD8\xAC\xD9\x85"
-     "\x00\xD8"
-     "\xAC\xD9\x85\xD8\xAD\x00\xD8\xAC\xD9\x85\xD9\x89\x00\xD8\xAC\xD9\x85\xD9"
-     "\x8A\x00"
-     "\xD8\xAC\xD9\x89\x00\xD8\xAC\xD9\x8A\x00\xD8\xAD\x00\xD8\xAD\xD8\xAC\x00"
-     "\xD8\xAD"
-     "\xD8\xAC\xD9\x8A\x00\xD8\xAD\xD9\x85\x00\xD8\xAD\xD9\x85\xD9\x89\x00\xD8"
-     "\xAD\xD9"
-     "\x85\xD9\x8A\x00\xD8\xAD\xD9\x89\x00\xD8\xAD\xD9\x8A\x00\xD8\xAE\x00\xD8"
-     "\xAE\xD8"
-     "\xAC\x00\xD8\xAE\xD8\xAD\x00\xD8\xAE\xD9\x85\x00\xD8\xAE\xD9\x89\x00\xD8"
-     "\xAE\xD9"
-     "\x8A\x00\xD8\xAF\x00\xD8\xB0\x00\xD8\xB0\xD9\xB0\x00\xD8\xB1\x00\xD8\xB1"
-     "\xD8\xB3"
-     "\xD9\x88\xD9\x84\x00\xD8\xB1\xD9\xB0\x00\xD8\xB1\xDB\x8C\xD8\xA7\xD9\x84"
-     "\x00\xD8"
-     "\xB2\x00\xD8\xB3\x00\xD8\xB3\xD8\xAC\x00\xD8\xB3\xD8\xAC\xD8\xAD\x00\xD8"
-     "\xB3\xD8"
-     "\xAC\xD9\x89\x00\xD8\xB3\xD8\xAD\x00\xD8\xB3\xD8\xAD\xD8\xAC\x00\xD8\xB3"
-     "\xD8\xAE"
-     "\x00\xD8\xB3\xD8\xAE\xD9\x89\x00\xD8\xB3\xD8\xAE\xD9\x8A\x00\xD8\xB3\xD8"
-     "\xB1\x00"
-     "\xD8\xB3\xD9\x85\x00\xD8\xB3\xD9\x85\xD8\xAC\x00\xD8\xB3\xD9\x85\xD8\xAD"
-     "\x00\xD8"
-     "\xB3\xD9\x85\xD9\x85\x00\xD8\xB3\xD9\x87\x00\xD8\xB3\xD9\x89\x00\xD8\xB3"
-     "\xD9\x8A"
-     "\x00\xD8\xB4\x00\xD8\xB4\xD8\xAC\x00\xD8\xB4\xD8\xAC\xD9\x8A\x00\xD8\xB4"
-     "\xD8\xAD"
-     "\x00\xD8\xB4\xD8\xAD\xD9\x85\x00\xD8\xB4\xD8\xAD\xD9\x8A\x00\xD8\xB4\xD8"
-     "\xAE\x00"
-     "\xD8\xB4\xD8\xB1\x00\xD8\xB4\xD9\x85\x00\xD8\xB4\xD9\x85\xD8\xAE\x00\xD8"
-     "\xB4\xD9"
-     "\x85\xD9\x85\x00\xD8\xB4\xD9\x87\x00\xD8\xB4\xD9\x89\x00\xD8\xB4\xD9\x8A"
-     "\x00\xD8"
-     "\xB5\x00\xD8\xB5\xD8\xAD\x00\xD8\xB5\xD8\xAD\xD8\xAD\x00\xD8\xB5\xD8\xAD"
-     "\xD9\x8A"
-     "\x00\xD8\xB5\xD8\xAE\x00\xD8\xB5\xD8\xB1\x00\xD8\xB5\xD9\x84\xD8\xB9\xD9"
-     "\x85\x00"
-     "\xD8\xB5\xD9\x84\xD9\x89\x00\xD8\xB5\xD9\x84\xD9\x89\x20\xD8\xA7\xD9\x84"
-     "\xD9\x84"
-     "\xD9\x87\x20\xD8\xB9\xD9\x84\xD9\x8A\xD9\x87\x20\xD9\x88\xD8\xB3\xD9\x84"
-     "\xD9\x85"
-     "\x00\xD8\xB5\xD9\x84\xDB\x92\x00\xD8\xB5\xD9\x85\x00\xD8\xB5\xD9\x85\xD9"
-     "\x85\x00"
-     "\xD8\xB5\xD9\x89\x00\xD8\xB5\xD9\x8A\x00\xD8\xB6\x00\xD8\xB6\xD8\xAC\x00"
-     "\xD8\xB6"
-     "\xD8\xAD\x00\xD8\xB6\xD8\xAD\xD9\x89\x00\xD8\xB6\xD8\xAD\xD9\x8A\x00\xD8"
-     "\xB6\xD8"
-     "\xAE\x00\xD8\xB6\xD8\xAE\xD9\x85\x00\xD8\xB6\xD8\xB1\x00\xD8\xB6\xD9\x85"
-     "\x00\xD8"
-     "\xB6\xD9\x89\x00\xD8\xB6\xD9\x8A\x00\xD8\xB7\x00\xD8\xB7\xD8\xAD\x00\xD8"
-     "\xB7\xD9"
-     "\x85\x00\xD8\xB7\xD9\x85\xD8\xAD\x00\xD8\xB7\xD9\x85\xD9\x85\x00\xD8\xB7"
-     "\xD9\x85"
-     "\xD9\x8A\x00\xD8\xB7\xD9\x89\x00\xD8\xB7\xD9\x8A\x00\xD8\xB8\x00\xD8\xB8"
-     "\xD9\x85"
-     "\x00\xD8\xB9\x00\xD8\xB9\xD8\xAC\x00\xD8\xB9\xD8\xAC\xD9\x85\x00\xD8\xB9"
-     "\xD9\x84"
-     "\xD9\x8A\xD9\x87\x00\xD8\xB9\xD9\x85\x00\xD8\xB9\xD9\x85\xD9\x85\x00\xD8"
-     "\xB9\xD9"
-     "\x85\xD9\x89\x00\xD8\xB9\xD9\x85\xD9\x8A\x00\xD8\xB9\xD9\x89\x00\xD8\xB9"
-     "\xD9\x8A"
-     "\x00\xD8\xBA\x00\xD8\xBA\xD8\xAC\x00\xD8\xBA\xD9\x85\x00\xD8\xBA\xD9\x85"
-     "\xD9\x85"
-     "\x00\xD8\xBA\xD9\x85\xD9\x89\x00\xD8\xBA\xD9\x85\xD9\x8A\x00\xD8\xBA\xD9"
-     "\x89\x00"
-     "\xD8\xBA\xD9\x8A\x00\xD9\x80\xD9\x8B\x00\xD9\x80\xD9\x8E\x00\xD9\x80\xD9"
-     "\x8E\xD9"
-     "\x91\x00\xD9\x80\xD9\x8F\x00\xD9\x80\xD9\x8F\xD9\x91\x00\xD9\x80\xD9\x90"
-     "\x00\xD9"
-     "\x80\xD9\x90\xD9\x91\x00\xD9\x80\xD9\x91\x00\xD9\x80\xD9\x92\x00\xD9\x81"
-     "\x00\xD9"
-     "\x81\xD8\xAC\x00\xD9\x81\xD8\xAD\x00\xD9\x81\xD8\xAE\x00\xD9\x81\xD8\xAE"
-     "\xD9\x85"
-     "\x00\xD9\x81\xD9\x85\x00\xD9\x81\xD9\x85\xD9\x8A\x00\xD9\x81\xD9\x89\x00"
-     "\xD9\x81"
-     "\xD9\x8A\x00\xD9\x82\x00\xD9\x82\xD8\xAD\x00\xD9\x82\xD9\x84\xDB\x92\x00"
-     "\xD9\x82"
-     "\xD9\x85\x00\xD9\x82\xD9\x85\xD8\xAD\x00\xD9\x82\xD9\x85\xD9\x85\x00\xD9"
-     "\x82\xD9"
-     "\x85\xD9\x8A\x00\xD9\x82\xD9\x89\x00\xD9\x82\xD9\x8A\x00\xD9\x83\x00\xD9"
-     "\x83\xD8"
-     "\xA7\x00\xD9\x83\xD8\xAC\x00\xD9\x83\xD8\xAD\x00\xD9\x83\xD8\xAE\x00\xD9"
-     "\x83\xD9"
-     "\x84\x00\xD9\x83\xD9\x85\x00\xD9\x83\xD9\x85\xD9\x85\x00\xD9\x83\xD9\x85"
-     "\xD9\x8A"
-     "\x00\xD9\x83\xD9\x89\x00\xD9\x83\xD9\x8A\x00\xD9\x84\x00\xD9\x84\xD8\xA2"
-     "\x00\xD9"
-     "\x84\xD8\xA3\x00\xD9\x84\xD8\xA5\x00\xD9\x84\xD8\xA7\x00\xD9\x84\xD8\xAC"
-     "\x00\xD9"
-     "\x84\xD8\xAC\xD8\xAC\x00\xD9\x84\xD8\xAC\xD9\x85\x00\xD9\x84\xD8\xAC\xD9"
-     "\x8A\x00"
-     "\xD9\x84\xD8\xAD\x00\xD9\x84\xD8\xAD\xD9\x85\x00\xD9\x84\xD8\xAD\xD9\x89"
-     "\x00\xD9"
-     "\x84\xD8\xAD\xD9\x8A\x00\xD9\x84\xD8\xAE\x00\xD9\x84\xD8\xAE\xD9\x85\x00"
-     "\xD9\x84"
-     "\xD9\x85\x00\xD9\x84\xD9\x85\xD8\xAD\x00\xD9\x84\xD9\x85\xD9\x8A\x00\xD9"
-     "\x84\xD9"
-     "\x87\x00\xD9\x84\xD9\x89\x00\xD9\x84\xD9\x8A\x00\xD9\x85\x00\xD9\x85\xD8"
-     "\xA7\x00"
-     "\xD9\x85\xD8\xAC\x00\xD9\x85\xD8\xAC\xD8\xAD\x00\xD9\x85\xD8\xAC\xD8\xAE"
-     "\x00\xD9"
-     "\x85\xD8\xAC\xD9\x85\x00\xD9\x85\xD8\xAC\xD9\x8A\x00\xD9\x85\xD8\xAD\x00"
-     "\xD9\x85"
-     "\xD8\xAD\xD8\xAC\x00\xD9\x85\xD8\xAD\xD9\x85\x00\xD9\x85\xD8\xAD\xD9\x85"
-     "\xD8\xAF"
-     "\x00\xD9\x85\xD8\xAD\xD9\x8A\x00\xD9\x85\xD8\xAE\x00\xD9\x85\xD8\xAE\xD8"
-     "\xAC\x00"
-     "\xD9\x85\xD8\xAE\xD9\x85\x00\xD9\x85\xD8\xAE\xD9\x8A\x00\xD9\x85\xD9\x85"
-     "\x00\xD9"
-     "\x85\xD9\x85\xD9\x8A\x00\xD9\x85\xD9\x89\x00\xD9\x85\xD9\x8A\x00\xD9\x86"
-     "\x00\xD9"
-     "\x86\xD8\xAC\x00\xD9\x86\xD8\xAC\xD8\xAD\x00\xD9\x86\xD8\xAC\xD9\x85\x00"
-     "\xD9\x86"
-     "\xD8\xAC\xD9\x89\x00\xD9\x86\xD8\xAC\xD9\x8A\x00\xD9\x86\xD8\xAD\x00\xD9"
-     "\x86\xD8"
-     "\xAD\xD9\x85\x00\xD9\x86\xD8\xAD\xD9\x89\x00\xD9\x86\xD8\xAD\xD9\x8A\x00"
-     "\xD9\x86"
-     "\xD8\xAE\x00\xD9\x86\xD8\xB1\x00\xD9\x86\xD8\xB2\x00\xD9\x86\xD9\x85\x00"
-     "\xD9\x86"
-     "\xD9\x85\xD9\x89\x00\xD9\x86\xD9\x85\xD9\x8A\x00\xD9\x86\xD9\x86\x00\xD9"
-     "\x86\xD9"
-     "\x87\x00\xD9\x86\xD9\x89\x00\xD9\x86\xD9\x8A\x00\xD9\x87\x00\xD9\x87\xD8"
-     "\xAC\x00"
-     "\xD9\x87\xD9\x85\x00\xD9\x87\xD9\x85\xD8\xAC\x00\xD9\x87\xD9\x85\xD9\x85"
-     "\x00\xD9"
-     "\x87\xD9\x89\x00\xD9\x87\xD9\x8A\x00\xD9\x87\xD9\xB0\x00\xD9\x88\x00\xD9"
-     "\x88\xD8"
-     "\xB3\xD9\x84\xD9\x85\x00\xD9\x88\xD9\xB4\x00\xD9\x89\x00\xD9\x89\xD9\xB0"
-     "\x00\xD9"
-     "\x8A\x00\xD9\x8A\xD8\xAC\x00\xD9\x8A\xD8\xAC\xD9\x8A\x00\xD9\x8A\xD8\xAD"
-     "\x00\xD9"
-     "\x8A\xD8\xAD\xD9\x8A\x00\xD9\x8A\xD8\xAE\x00\xD9\x8A\xD8\xB1\x00\xD9\x8A"
-     "\xD8\xB2"
-     "\x00\xD9\x8A\xD9\x85\x00\xD9\x8A\xD9\x85\xD9\x85\x00\xD9\x8A\xD9\x85\xD9"
-     "\x8A\x00"
-     "\xD9\x8A\xD9\x86\x00\xD9\x8A\xD9\x87\x00\xD9\x8A\xD9\x89\x00\xD9\x8A\xD9"
-     "\x8A\x00"
-     "\xD9\x8A\xD9\xB4\x00\xD9\xAE\x00\xD9\xAF\x00\xD9\xB1\x00\xD9\xB9\x00\xD9"
-     "\xBA\x00"
-     "\xD9\xBB\x00\xD9\xBE\x00\xD9\xBF\x00\xDA\x80\x00\xDA\x83\x00\xDA\x84\x00"
-     "\xDA\x86"
-     "\x00\xDA\x87\x00\xDA\x88\x00\xDA\x8C\x00\xDA\x8D\x00\xDA\x8E\x00\xDA\x91"
-     "\x00\xDA"
-     "\x98\x00\xDA\xA1\x00\xDA\xA4\x00\xDA\xA6\x00\xDA\xA9\x00\xDA\xAD\x00\xDA"
-     "\xAF\x00"
-     "\xDA\xB1\x00\xDA\xB3\x00\xDA\xBA\x00\xDA\xBB\x00\xDA\xBE\x00\xDB\x80\x00"
-     "\xDB\x81"
-     "\x00\xDB\x82\x00\xDB\x85\x00\xDB\x86\x00\xDB\x87\x00\xDB\x87\xD9\xB4\x00"
-     "\xDB\x88"
-     "\x00\xDB\x89\x00\xDB\x8B\x00\xDB\x8C\x00\xDB\x90\x00\xDB\x92\x00\xDB\x93"
-     "\x00\xE0"
-     "\xA4\x95\xE0\xA4\xBC\x00\xE0\xA4\x96\xE0\xA4\xBC\x00\xE0\xA4\x97\xE0\xA4"
-     "\xBC\x00"
-     "\xE0\xA4\x9C\xE0\xA4\xBC\x00\xE0\xA4\xA1\xE0\xA4\xBC\x00\xE0\xA4\xA2\xE0"
-     "\xA4\xBC"
-     "\x00\xE0\xA4\xA9\x00\xE0\xA4\xAB\xE0\xA4\xBC\x00\xE0\xA4\xAF\xE0\xA4\xBC"
-     "\x00\xE0"
-     "\xA4\xB1\x00\xE0\xA4\xB4\x00\xE0\xA6\xA1\xE0\xA6\xBC\x00\xE0\xA6\xA2\xE0"
-     "\xA6\xBC"
-     "\x00\xE0\xA6\xAF\xE0\xA6\xBC\x00\xE0\xA7\x8B\x00\xE0\xA7\x8C\x00\xE0\xA8"
-     "\x96\xE0"
-     "\xA8\xBC\x00\xE0\xA8\x97\xE0\xA8\xBC\x00\xE0\xA8\x9C\xE0\xA8\xBC\x00\xE0"
-     "\xA8\xAB"
-     "\xE0\xA8\xBC\x00\xE0\xA8\xB2\xE0\xA8\xBC\x00\xE0\xA8\xB8\xE0\xA8\xBC\x00"
-     "\xE0\xAC"
-     "\xA1\xE0\xAC\xBC\x00\xE0\xAC\xA2\xE0\xAC\xBC\x00\xE0\xAD\x88\x00\xE0\xAD"
-     "\x8B\x00"
-     "\xE0\xAD\x8C\x00\xE0\xAE\x94\x00\xE0\xAF\x8A\x00\xE0\xAF\x8B\x00\xE0\xAF"
-     "\x8C\x00"
-     "\xE0\xB1\x88\x00\xE0\xB3\x80\x00\xE0\xB3\x87\x00\xE0\xB3\x88\x00\xE0\xB3"
-     "\x8A\x00"
-     "\xE0\xB3\x8B\x00\xE0\xB5\x8A\x00\xE0\xB5\x8B\x00\xE0\xB5\x8C\x00\xE0\xB7"
-     "\x9A\x00"
-     "\xE0\xB7\x9C\x00\xE0\xB7\x9D\x00\xE0\xB7\x9E\x00\xE0\xB9\x8D\xE0\xB8\xB2"
-     "\x00\xE0"
-     "\xBA\xAB\xE0\xBA\x99\x00\xE0\xBA\xAB\xE0\xBA\xA1\x00\xE0\xBB\x8D\xE0\xBA"
-     "\xB2\x00"
-     "\xE0\xBC\x8B\x00\xE0\xBD\x80\xE0\xBE\xB5\x00\xE0\xBD\x82\xE0\xBE\xB7\x00"
-     "\xE0\xBD"
-     "\x8C\xE0\xBE\xB7\x00\xE0\xBD\x91\xE0\xBE\xB7\x00\xE0\xBD\x96\xE0\xBE\xB7"
-     "\x00\xE0"
-     "\xBD\x9B\xE0\xBE\xB7\x00\xE0\xBD\xB1\xE0\xBD\xB2\x00\xE0\xBD\xB1\xE0\xBD"
-     "\xB4\x00"
-     "\xE0\xBD\xB1\xE0\xBE\x80\x00\xE0\xBE\x90\xE0\xBE\xB5\x00\xE0\xBE\x92\xE0"
-     "\xBE\xB7"
-     "\x00\xE0\xBE\x9C\xE0\xBE\xB7\x00\xE0\xBE\xA1\xE0\xBE\xB7\x00\xE0\xBE\xA6"
-     "\xE0\xBE"
-     "\xB7\x00\xE0\xBE\xAB\xE0\xBE\xB7\x00\xE0\xBE\xB2\xE0\xBD\xB1\xE0\xBE\x80"
-     "\x00\xE0"
-     "\xBE\xB2\xE0\xBE\x80\x00\xE0\xBE\xB3\xE0\xBD\xB1\xE0\xBE\x80\x00\xE0\xBE"
-     "\xB3\xE0"
-     "\xBE\x80\x00\xE1\x80\xA6\x00\xE1\x83\x90\x00\xE1\x83\x91\x00\xE1\x83\x92"
-     "\x00\xE1"
-     "\x83\x93\x00\xE1\x83\x94\x00\xE1\x83\x95\x00\xE1\x83\x96\x00\xE1\x83\x97"
-     "\x00\xE1"
-     "\x83\x98\x00\xE1\x83\x99\x00\xE1\x83\x9A\x00\xE1\x83\x9B\x00\xE1\x83\x9C"
-     "\x00\xE1"
-     "\x83\x9D\x00\xE1\x83\x9E\x00\xE1\x83\x9F\x00\xE1\x83\xA0\x00\xE1\x83\xA1"
-     "\x00\xE1"
-     "\x83\xA2\x00\xE1\x83\xA3\x00\xE1\x83\xA4\x00\xE1\x83\xA5\x00\xE1\x83\xA6"
-     "\x00\xE1"
-     "\x83\xA7\x00\xE1\x83\xA8\x00\xE1\x83\xA9\x00\xE1\x83\xAA\x00\xE1\x83\xAB"
-     "\x00\xE1"
-     "\x83\xAC\x00\xE1\x83\xAD\x00\xE1\x83\xAE\x00\xE1\x83\xAF\x00\xE1\x83\xB0"
-     "\x00\xE1"
-     "\x83\xB1\x00\xE1\x83\xB2\x00\xE1\x83\xB3\x00\xE1\x83\xB4\x00\xE1\x83\xB5"
-     "\x00\xE1"
-     "\x83\xB6\x00\xE1\x83\xB7\x00\xE1\x83\xB8\x00\xE1\x83\xB9\x00\xE1\x83\xBA"
-     "\x00\xE1"
-     "\x83\xBD\x00\xE1\x83\xBE\x00\xE1\x83\xBF\x00\xE1\x84\x80\x00\xE1\x84\x81"
-     "\x00\xE1"
-     "\x84\x82\x00\xE1\x84\x83\x00\xE1\x84\x84\x00\xE1\x84\x85\x00\xE1\x84\x86"
-     "\x00\xE1"
-     "\x84\x87\x00\xE1\x84\x88\x00\xE1\x84\x89\x00\xE1\x84\x8A\x00\xE1\x84\x8B"
-     "\x00\xE1"
-     "\x84\x8C\x00\xE1\x84\x8D\x00\xE1\x84\x8E\x00\xE1\x84\x8F\x00\xE1\x84\x90"
-     "\x00\xE1"
-     "\x84\x91\x00\xE1\x84\x92\x00\xE1\x84\x94\x00\xE1\x84\x95\x00\xE1\x84\x9A"
-     "\x00\xE1"
-     "\x84\x9C\x00\xE1\x84\x9D\x00\xE1\x84\x9E\x00\xE1\x84\xA0\x00\xE1\x84\xA1"
-     "\x00\xE1"
-     "\x84\xA2\x00\xE1\x84\xA3\x00\xE1\x84\xA7\x00\xE1\x84\xA9\x00\xE1\x84\xAB"
-     "\x00\xE1"
-     "\x84\xAC\x00\xE1\x84\xAD\x00\xE1\x84\xAE\x00\xE1\x84\xAF\x00\xE1\x84\xB2"
-     "\x00\xE1"
-     "\x84\xB6\x00\xE1\x85\x80\x00\xE1\x85\x87\x00\xE1\x85\x8C\x00\xE1\x85\x97"
-     "\x00\xE1"
-     "\x85\x98\x00\xE1\x85\x99\x00\xE1\x85\xA0\x00\xE1\x85\xA1\x00\xE1\x85\xA2"
-     "\x00\xE1"
-     "\x85\xA3\x00\xE1\x85\xA4\x00\xE1\x85\xA5\x00\xE1\x85\xA6\x00\xE1\x85\xA7"
-     "\x00\xE1"
-     "\x85\xA8\x00\xE1\x85\xA9\x00\xE1\x85\xAA\x00\xE1\x85\xAB\x00\xE1\x85\xAC"
-     "\x00\xE1"
-     "\x85\xAD\x00\xE1\x85\xAE\x00\xE1\x85\xAF\x00\xE1\x85\xB0\x00\xE1\x85\xB1"
-     "\x00\xE1"
-     "\x85\xB2\x00\xE1\x85\xB3\x00\xE1\x85\xB4\x00\xE1\x85\xB5\x00\xE1\x86\x84"
-     "\x00\xE1"
-     "\x86\x85\x00\xE1\x86\x88\x00\xE1\x86\x91\x00\xE1\x86\x92\x00\xE1\x86\x94"
-     "\x00\xE1"
-     "\x86\x9E\x00\xE1\x86\xA1\x00\xE1\x86\xAA\x00\xE1\x86\xAC\x00\xE1\x86\xAD"
-     "\x00\xE1"
-     "\x86\xB0\x00\xE1\x86\xB1\x00\xE1\x86\xB2\x00\xE1\x86\xB3\x00\xE1\x86\xB4"
-     "\x00\xE1"
-     "\x86\xB5\x00\xE1\x87\x87\x00\xE1\x87\x88\x00\xE1\x87\x8C\x00\xE1\x87\x8E"
-     "\x00\xE1"
-     "\x87\x93\x00\xE1\x87\x97\x00\xE1\x87\x99\x00\xE1\x87\x9D\x00\xE1\x87\x9F"
-     "\x00\xE1"
-     "\x87\xB1\x00\xE1\x87\xB2\x00\xE1\x8E\xA0\x00\xE1\x8E\xA1\x00\xE1\x8E\xA2"
-     "\x00\xE1"
-     "\x8E\xA3\x00\xE1\x8E\xA4\x00\xE1\x8E\xA5\x00\xE1\x8E\xA6\x00\xE1\x8E\xA7"
-     "\x00\xE1"
-     "\x8E\xA8\x00\xE1\x8E\xA9\x00\xE1\x8E\xAA\x00\xE1\x8E\xAB\x00\xE1\x8E\xAC"
-     "\x00\xE1"
-     "\x8E\xAD\x00\xE1\x8E\xAE\x00\xE1\x8E\xAF\x00\xE1\x8E\xB0\x00\xE1\x8E\xB1"
-     "\x00\xE1"
-     "\x8E\xB2\x00\xE1\x8E\xB3\x00\xE1\x8E\xB4\x00\xE1\x8E\xB5\x00\xE1\x8E\xB6"
-     "\x00\xE1"
-     "\x8E\xB7\x00\xE1\x8E\xB8\x00\xE1\x8E\xB9\x00\xE1\x8E\xBA\x00\xE1\x8E\xBB"
-     "\x00\xE1"
-     "\x8E\xBC\x00\xE1\x8E\xBD\x00\xE1\x8E\xBE\x00\xE1\x8E\xBF\x00\xE1\x8F\x80"
-     "\x00\xE1"
-     "\x8F\x81\x00\xE1\x8F\x82\x00\xE1\x8F\x83\x00\xE1\x8F\x84\x00\xE1\x8F\x85"
-     "\x00\xE1"
-     "\x8F\x86\x00\xE1\x8F\x87\x00\xE1\x8F\x88\x00\xE1\x8F\x89\x00\xE1\x8F\x8A"
-     "\x00\xE1"
-     "\x8F\x8B\x00\xE1\x8F\x8C\x00\xE1\x8F\x8D\x00\xE1\x8F\x8E\x00\xE1\x8F\x8F"
-     "\x00\xE1"
-     "\x8F\x90\x00\xE1\x8F\x91\x00\xE1\x8F\x92\x00\xE1\x8F\x93\x00\xE1\x8F\x94"
-     "\x00\xE1"
-     "\x8F\x95\x00\xE1\x8F\x96\x00\xE1\x8F\x97\x00\xE1\x8F\x98\x00\xE1\x8F\x99"
-     "\x00\xE1"
-     "\x8F\x9A\x00\xE1\x8F\x9B\x00\xE1\x8F\x9C\x00\xE1\x8F\x9D\x00\xE1\x8F\x9E"
-     "\x00\xE1"
-     "\x8F\x9F\x00\xE1\x8F\xA0\x00\xE1\x8F\xA1\x00\xE1\x8F\xA2\x00\xE1\x8F\xA3"
-     "\x00\xE1"
-     "\x8F\xA4\x00\xE1\x8F\xA5\x00\xE1\x8F\xA6\x00\xE1\x8F\xA7\x00\xE1\x8F\xA8"
-     "\x00\xE1"
-     "\x8F\xA9\x00\xE1\x8F\xAA\x00\xE1\x8F\xAB\x00\xE1\x8F\xAC\x00\xE1\x8F\xAD"
-     "\x00\xE1"
-     "\x8F\xAE\x00\xE1\x8F\xAF\x00\xE1\x8F\xB0\x00\xE1\x8F\xB1\x00\xE1\x8F\xB2"
-     "\x00\xE1"
-     "\x8F\xB3\x00\xE1\x8F\xB4\x00\xE1\x8F\xB5\x00\xE1\xAC\x86\x00\xE1\xAC\x88"
-     "\x00\xE1"
-     "\xAC\x8A\x00\xE1\xAC\x8C\x00\xE1\xAC\x8E\x00\xE1\xAC\x92\x00\xE1\xAC\xBB"
-     "\x00\xE1"
-     "\xAC\xBD\x00\xE1\xAD\x80\x00\xE1\xAD\x81\x00\xE1\xAD\x83\x00\xE1\xB4\x82"
-     "\x00\xE1"
-     "\xB4\x96\x00\xE1\xB4\x97\x00\xE1\xB4\x9C\x00\xE1\xB4\x9D\x00\xE1\xB4\xA5"
-     "\x00\xE1"
-     "\xB5\xB9\x00\xE1\xB5\xBB\x00\xE1\xB5\xBD\x00\xE1\xB6\x85\x00\xE1\xB6\x8E"
-     "\x00\xE1"
-     "\xB6\x91\x00\xE1\xB8\x81\x00\xE1\xB8\x83\x00\xE1\xB8\x85\x00\xE1\xB8\x87"
-     "\x00\xE1"
-     "\xB8\x89\x00\xE1\xB8\x8B\x00\xE1\xB8\x8D\x00\xE1\xB8\x8F\x00\xE1\xB8\x91"
-     "\x00\xE1"
-     "\xB8\x93\x00\xE1\xB8\x95\x00\xE1\xB8\x97\x00\xE1\xB8\x99\x00\xE1\xB8\x9B"
-     "\x00\xE1"
-     "\xB8\x9D\x00\xE1\xB8\x9F\x00\xE1\xB8\xA1\x00\xE1\xB8\xA3\x00\xE1\xB8\xA5"
-     "\x00\xE1"
-     "\xB8\xA7\x00\xE1\xB8\xA9\x00\xE1\xB8\xAB\x00\xE1\xB8\xAD\x00\xE1\xB8\xAF"
-     "\x00\xE1"
-     "\xB8\xB1\x00\xE1\xB8\xB3\x00\xE1\xB8\xB5\x00\xE1\xB8\xB7\x00\xE1\xB8\xB9"
-     "\x00\xE1"
-     "\xB8\xBB\x00\xE1\xB8\xBD\x00\xE1\xB8\xBF\x00\xE1\xB9\x81\x00\xE1\xB9\x83"
-     "\x00\xE1"
-     "\xB9\x85\x00\xE1\xB9\x87\x00\xE1\xB9\x89\x00\xE1\xB9\x8B\x00\xE1\xB9\x8D"
-     "\x00\xE1"
-     "\xB9\x8F\x00\xE1\xB9\x91\x00\xE1\xB9\x93\x00\xE1\xB9\x95\x00\xE1\xB9\x97"
-     "\x00\xE1"
-     "\xB9\x99\x00\xE1\xB9\x9B\x00\xE1\xB9\x9D\x00\xE1\xB9\x9F\x00\xE1\xB9\xA1"
-     "\x00\xE1"
-     "\xB9\xA3\x00\xE1\xB9\xA5\x00\xE1\xB9\xA7\x00\xE1\xB9\xA9\x00\xE1\xB9\xAB"
-     "\x00\xE1"
-     "\xB9\xAD\x00\xE1\xB9\xAF\x00\xE1\xB9\xB1\x00\xE1\xB9\xB3\x00\xE1\xB9\xB5"
-     "\x00\xE1"
-     "\xB9\xB7\x00\xE1\xB9\xB9\x00\xE1\xB9\xBB\x00\xE1\xB9\xBD\x00\xE1\xB9\xBF"
-     "\x00\xE1"
-     "\xBA\x81\x00\xE1\xBA\x83\x00\xE1\xBA\x85\x00\xE1\xBA\x87\x00\xE1\xBA\x89"
-     "\x00\xE1"
-     "\xBA\x8B\x00\xE1\xBA\x8D\x00\xE1\xBA\x8F\x00\xE1\xBA\x91\x00\xE1\xBA\x93"
-     "\x00\xE1"
-     "\xBA\x95\x00\xE1\xBA\x96\x00\xE1\xBA\x97\x00\xE1\xBA\x98\x00\xE1\xBA\x99"
-     "\x00\xE1"
-     "\xBA\xA1\x00\xE1\xBA\xA3\x00\xE1\xBA\xA5\x00\xE1\xBA\xA7\x00\xE1\xBA\xA9"
-     "\x00\xE1"
-     "\xBA\xAB\x00\xE1\xBA\xAD\x00\xE1\xBA\xAF\x00\xE1\xBA\xB1\x00\xE1\xBA\xB3"
-     "\x00\xE1"
-     "\xBA\xB5\x00\xE1\xBA\xB7\x00\xE1\xBA\xB9\x00\xE1\xBA\xBB\x00\xE1\xBA\xBD"
-     "\x00\xE1"
-     "\xBA\xBF\x00\xE1\xBB\x81\x00\xE1\xBB\x83\x00\xE1\xBB\x85\x00\xE1\xBB\x87"
-     "\x00\xE1"
-     "\xBB\x89\x00\xE1\xBB\x8B\x00\xE1\xBB\x8D\x00\xE1\xBB\x8F\x00\xE1\xBB\x91"
-     "\x00\xE1"
-     "\xBB\x93\x00\xE1\xBB\x95\x00\xE1\xBB\x97\x00\xE1\xBB\x99\x00\xE1\xBB\x9B"
-     "\x00\xE1"
-     "\xBB\x9D\x00\xE1\xBB\x9F\x00\xE1\xBB\xA1\x00\xE1\xBB\xA3\x00\xE1\xBB\xA5"
-     "\x00\xE1"
-     "\xBB\xA7\x00\xE1\xBB\xA9\x00\xE1\xBB\xAB\x00\xE1\xBB\xAD\x00\xE1\xBB\xAF"
-     "\x00\xE1"
-     "\xBB\xB1\x00\xE1\xBB\xB3\x00\xE1\xBB\xB5\x00\xE1\xBB\xB7\x00\xE1\xBB\xB9"
-     "\x00\xE1"
-     "\xBB\xBB\x00\xE1\xBB\xBD\x00\xE1\xBB\xBF\x00\xE1\xBC\x80\x00\xE1\xBC\x81"
-     "\x00\xE1"
-     "\xBC\x82\x00\xE1\xBC\x83\x00\xE1\xBC\x84\x00\xE1\xBC\x85\x00\xE1\xBC\x86"
-     "\x00\xE1"
-     "\xBC\x87\x00\xE1\xBC\x90\x00\xE1\xBC\x91\x00\xE1\xBC\x92\x00\xE1\xBC\x93"
-     "\x00\xE1"
-     "\xBC\x94\x00\xE1\xBC\x95\x00\xE1\xBC\xA0\x00\xE1\xBC\xA1\x00\xE1\xBC\xA2"
-     "\x00\xE1"
-     "\xBC\xA3\x00\xE1\xBC\xA4\x00\xE1\xBC\xA5\x00\xE1\xBC\xA6\x00\xE1\xBC\xA7"
-     "\x00\xE1"
-     "\xBC\xB0\x00\xE1\xBC\xB1\x00\xE1\xBC\xB2\x00\xE1\xBC\xB3\x00\xE1\xBC\xB4"
-     "\x00\xE1"
-     "\xBC\xB5\x00\xE1\xBC\xB6\x00\xE1\xBC\xB7\x00\xE1\xBD\x80\x00\xE1\xBD\x81"
-     "\x00\xE1"
-     "\xBD\x82\x00\xE1\xBD\x83\x00\xE1\xBD\x84\x00\xE1\xBD\x85\x00\xE1\xBD\x90"
-     "\x00\xE1"
-     "\xBD\x91\x00\xE1\xBD\x92\x00\xE1\xBD\x93\x00\xE1\xBD\x94\x00\xE1\xBD\x95"
-     "\x00\xE1"
-     "\xBD\x96\x00\xE1\xBD\x97\x00\xE1\xBD\xA0\x00\xE1\xBD\xA1\x00\xE1\xBD\xA2"
-     "\x00\xE1"
-     "\xBD\xA3\x00\xE1\xBD\xA4\x00\xE1\xBD\xA5\x00\xE1\xBD\xA6\x00\xE1\xBD\xA7"
-     "\x00\xE1"
-     "\xBD\xB0\x00\xE1\xBD\xB2\x00\xE1\xBD\xB4\x00\xE1\xBD\xB6\x00\xE1\xBD\xB8"
-     "\x00\xE1"
-     "\xBD\xBA\x00\xE1\xBD\xBC\x00\xE1\xBE\x80\x00\xE1\xBE\x81\x00\xE1\xBE\x82"
-     "\x00\xE1"
-     "\xBE\x83\x00\xE1\xBE\x84\x00\xE1\xBE\x85\x00\xE1\xBE\x86\x00\xE1\xBE\x87"
-     "\x00\xE1"
-     "\xBE\x90\x00\xE1\xBE\x91\x00\xE1\xBE\x92\x00\xE1\xBE\x93\x00\xE1\xBE\x94"
-     "\x00\xE1"
-     "\xBE\x95\x00\xE1\xBE\x96\x00\xE1\xBE\x97\x00\xE1\xBE\xA0\x00\xE1\xBE\xA1"
-     "\x00\xE1"
-     "\xBE\xA2\x00\xE1\xBE\xA3\x00\xE1\xBE\xA4\x00\xE1\xBE\xA5\x00\xE1\xBE\xA6"
-     "\x00\xE1"
-     "\xBE\xA7\x00\xE1\xBE\xB0\x00\xE1\xBE\xB1\x00\xE1\xBE\xB2\x00\xE1\xBE\xB3"
-     "\x00\xE1"
-     "\xBE\xB4\x00\xE1\xBE\xB6\x00\xE1\xBE\xB7\x00\xE1\xBF\x82\x00\xE1\xBF\x83"
-     "\x00\xE1"
-     "\xBF\x84\x00\xE1\xBF\x86\x00\xE1\xBF\x87\x00\xE1\xBF\x90\x00\xE1\xBF\x91"
-     "\x00\xE1"
-     "\xBF\x92\x00\xE1\xBF\x96\x00\xE1\xBF\x97\x00\xE1\xBF\xA0\x00\xE1\xBF\xA1"
-     "\x00\xE1"
-     "\xBF\xA2\x00\xE1\xBF\xA4\x00\xE1\xBF\xA5\x00\xE1\xBF\xA6\x00\xE1\xBF\xA7"
-     "\x00\xE1"
-     "\xBF\xB2\x00\xE1\xBF\xB3\x00\xE1\xBF\xB4\x00\xE1\xBF\xB6\x00\xE1\xBF\xB7"
-     "\x00\xE2"
-     "\x80\x90\x00\xE2\x80\x93\x00\xE2\x80\x94\x00\xE2\x80\xB2\xE2\x80\xB2\x00"
-     "\xE2\x80"
-     "\xB2\xE2\x80\xB2\xE2\x80\xB2\x00\xE2\x80\xB2\xE2\x80\xB2\xE2\x80\xB2\xE2"
-     "\x80\xB2"
-     "\x00\xE2\x80\xB5\xE2\x80\xB5\x00\xE2\x80\xB5\xE2\x80\xB5\xE2\x80\xB5\x00"
-     "\xE2\x82"
-     "\xA9\x00\xE2\x85\x8E\x00\xE2\x86\x84\x00\xE2\x86\x90\x00\xE2\x86\x91\x00"
-     "\xE2\x86"
-     "\x92\x00\xE2\x86\x93\x00\xE2\x86\x9A\x00\xE2\x86\x9B\x00\xE2\x86\xAE\x00"
-     "\xE2\x87"
-     "\x8D\x00\xE2\x87\x8E\x00\xE2\x87\x8F\x00\xE2\x88\x82\x00\xE2\x88\x84\x00"
-     "\xE2\x88"
-     "\x87\x00\xE2\x88\x89\x00\xE2\x88\x8C\x00\xE2\x88\x91\x00\xE2\x88\x92\x00"
-     "\xE2\x88"
-     "\xA4\x00\xE2\x88\xA6\x00\xE2\x88\xAB\xE2\x88\xAB\x00\xE2\x88\xAB\xE2\x88"
-     "\xAB\xE2"
-     "\x88\xAB\x00\xE2\x88\xAB\xE2\x88\xAB\xE2\x88\xAB\xE2\x88\xAB\x00\xE2\x88"
-     "\xAE\xE2"
-     "\x88\xAE\x00\xE2\x88\xAE\xE2\x88\xAE\xE2\x88\xAE\x00\xE2\x89\x81\x00\xE2"
-     "\x89\x84"
-     "\x00\xE2\x89\x87\x00\xE2\x89\x89\x00\xE2\x89\xA0\x00\xE2\x89\xA2\x00\xE2"
-     "\x89\xAD"
-     "\x00\xE2\x89\xAE\x00\xE2\x89\xAF\x00\xE2\x89\xB0\x00\xE2\x89\xB1\x00\xE2"
-     "\x89\xB4"
-     "\x00\xE2\x89\xB5\x00\xE2\x89\xB8\x00\xE2\x89\xB9\x00\xE2\x8A\x80\x00\xE2"
-     "\x8A\x81"
-     "\x00\xE2\x8A\x84\x00\xE2\x8A\x85\x00\xE2\x8A\x88\x00\xE2\x8A\x89\x00\xE2"
-     "\x8A\xAC"
-     "\x00\xE2\x8A\xAD\x00\xE2\x8A\xAE\x00\xE2\x8A\xAF\x00\xE2\x8B\xA0\x00\xE2"
-     "\x8B\xA1"
-     "\x00\xE2\x8B\xA2\x00\xE2\x8B\xA3\x00\xE2\x8B\xAA\x00\xE2\x8B\xAB\x00\xE2"
-     "\x8B\xAC"
-     "\x00\xE2\x8B\xAD\x00\xE2\x94\x82\x00\xE2\x96\xA0\x00\xE2\x97\x8B\x00\xE2"
-     "\xA6\x85"
-     "\x00\xE2\xA6\x86\x00\xE2\xAB\x9D\xCC\xB8\x00\xE2\xB0\xB0\x00\xE2\xB0\xB1"
-     "\x00\xE2"
-     "\xB0\xB2\x00\xE2\xB0\xB3\x00\xE2\xB0\xB4\x00\xE2\xB0\xB5\x00\xE2\xB0\xB6"
-     "\x00\xE2"
-     "\xB0\xB7\x00\xE2\xB0\xB8\x00\xE2\xB0\xB9\x00\xE2\xB0\xBA\x00\xE2\xB0\xBB"
-     "\x00\xE2"
-     "\xB0\xBC\x00\xE2\xB0\xBD\x00\xE2\xB0\xBE\x00\xE2\xB0\xBF\x00\xE2\xB1\x80"
-     "\x00\xE2"
-     "\xB1\x81\x00\xE2\xB1\x82\x00\xE2\xB1\x83\x00\xE2\xB1\x84\x00\xE2\xB1\x85"
-     "\x00\xE2"
-     "\xB1\x86\x00\xE2\xB1\x87\x00\xE2\xB1\x88\x00\xE2\xB1\x89\x00\xE2\xB1\x8A"
-     "\x00\xE2"
-     "\xB1\x8B\x00\xE2\xB1\x8C\x00\xE2\xB1\x8D\x00\xE2\xB1\x8E\x00\xE2\xB1\x8F"
-     "\x00\xE2"
-     "\xB1\x90\x00\xE2\xB1\x91\x00\xE2\xB1\x92\x00\xE2\xB1\x93\x00\xE2\xB1\x94"
-     "\x00\xE2"
-     "\xB1\x95\x00\xE2\xB1\x96\x00\xE2\xB1\x97\x00\xE2\xB1\x98\x00\xE2\xB1\x99"
-     "\x00\xE2"
-     "\xB1\x9A\x00\xE2\xB1\x9B\x00\xE2\xB1\x9C\x00\xE2\xB1\x9D\x00\xE2\xB1\x9E"
-     "\x00\xE2"
-     "\xB1\x9F\x00\xE2\xB1\xA1\x00\xE2\xB1\xA5\x00\xE2\xB1\xA6\x00\xE2\xB1\xA8"
-     "\x00\xE2"
-     "\xB1\xAA\x00\xE2\xB1\xAC\x00\xE2\xB1\xB1\x00\xE2\xB1\xB3\x00\xE2\xB1\xB6"
-     "\x00\xE2"
-     "\xB2\x81\x00\xE2\xB2\x83\x00\xE2\xB2\x85\x00\xE2\xB2\x87\x00\xE2\xB2\x89"
-     "\x00\xE2"
-     "\xB2\x8B\x00\xE2\xB2\x8D\x00\xE2\xB2\x8F\x00\xE2\xB2\x91\x00\xE2\xB2\x93"
-     "\x00\xE2"
-     "\xB2\x95\x00\xE2\xB2\x97\x00\xE2\xB2\x99\x00\xE2\xB2\x9B\x00\xE2\xB2\x9D"
-     "\x00\xE2"
-     "\xB2\x9F\x00\xE2\xB2\xA1\x00\xE2\xB2\xA3\x00\xE2\xB2\xA5\x00\xE2\xB2\xA7"
-     "\x00\xE2"
-     "\xB2\xA9\x00\xE2\xB2\xAB\x00\xE2\xB2\xAD\x00\xE2\xB2\xAF\x00\xE2\xB2\xB1"
-     "\x00\xE2"
-     "\xB2\xB3\x00\xE2\xB2\xB5\x00\xE2\xB2\xB7\x00\xE2\xB2\xB9\x00\xE2\xB2\xBB"
-     "\x00\xE2"
-     "\xB2\xBD\x00\xE2\xB2\xBF\x00\xE2\xB3\x81\x00\xE2\xB3\x83\x00\xE2\xB3\x85"
-     "\x00\xE2"
-     "\xB3\x87\x00\xE2\xB3\x89\x00\xE2\xB3\x8B\x00\xE2\xB3\x8D\x00\xE2\xB3\x8F"
-     "\x00\xE2"
-     "\xB3\x91\x00\xE2\xB3\x93\x00\xE2\xB3\x95\x00\xE2\xB3\x97\x00\xE2\xB3\x99"
-     "\x00\xE2"
-     "\xB3\x9B\x00\xE2\xB3\x9D\x00\xE2\xB3\x9F\x00\xE2\xB3\xA1\x00\xE2\xB3\xA3"
-     "\x00\xE2"
-     "\xB3\xAC\x00\xE2\xB3\xAE\x00\xE2\xB3\xB3\x00\xE2\xB4\x80\x00\xE2\xB4\x81"
-     "\x00\xE2"
-     "\xB4\x82\x00\xE2\xB4\x83\x00\xE2\xB4\x84\x00\xE2\xB4\x85\x00\xE2\xB4\x86"
-     "\x00\xE2"
-     "\xB4\x87\x00\xE2\xB4\x88\x00\xE2\xB4\x89\x00\xE2\xB4\x8A\x00\xE2\xB4\x8B"
-     "\x00\xE2"
-     "\xB4\x8C\x00\xE2\xB4\x8D\x00\xE2\xB4\x8E\x00\xE2\xB4\x8F\x00\xE2\xB4\x90"
-     "\x00\xE2"
-     "\xB4\x91\x00\xE2\xB4\x92\x00\xE2\xB4\x93\x00\xE2\xB4\x94\x00\xE2\xB4\x95"
-     "\x00\xE2"
-     "\xB4\x96\x00\xE2\xB4\x97\x00\xE2\xB4\x98\x00\xE2\xB4\x99\x00\xE2\xB4\x9A"
-     "\x00\xE2"
-     "\xB4\x9B\x00\xE2\xB4\x9C\x00\xE2\xB4\x9D\x00\xE2\xB4\x9E\x00\xE2\xB4\x9F"
-     "\x00\xE2"
-     "\xB4\xA0\x00\xE2\xB4\xA1\x00\xE2\xB4\xA2\x00\xE2\xB4\xA3\x00\xE2\xB4\xA4"
-     "\x00\xE2"
-     "\xB4\xA5\x00\xE2\xB4\xA7\x00\xE2\xB4\xAD\x00\xE2\xB5\xA1\x00\xE3\x80\x81"
-     "\x00\xE3"
-     "\x80\x82\x00\xE3\x80\x88\x00\xE3\x80\x89\x00\xE3\x80\x8A\x00\xE3\x80\x8B"
-     "\x00\xE3"
-     "\x80\x8C\x00\xE3\x80\x8D\x00\xE3\x80\x8E\x00\xE3\x80\x8F\x00\xE3\x80\x90"
-     "\x00\xE3"
-     "\x80\x91\x00\xE3\x80\x92\x00\xE3\x80\x94\x00\xE3\x80\x94\x73\xE3\x80\x95"
-     "\x00\xE3"
-     "\x80\x94\xE4\xB8\x89\xE3\x80\x95\x00\xE3\x80\x94\xE4\xBA\x8C\xE3\x80\x95"
-     "\x00\xE3"
-     "\x80\x94\xE5\x8B\x9D\xE3\x80\x95\x00\xE3\x80\x94\xE5\xAE\x89\xE3\x80\x95"
-     "\x00\xE3"
-     "\x80\x94\xE6\x89\x93\xE3\x80\x95\x00\xE3\x80\x94\xE6\x95\x97\xE3\x80\x95"
-     "\x00\xE3"
-     "\x80\x94\xE6\x9C\xAC\xE3\x80\x95\x00\xE3\x80\x94\xE7\x82\xB9\xE3\x80\x95"
-     "\x00\xE3"
-     "\x80\x94\xE7\x9B\x97\xE3\x80\x95\x00\xE3\x80\x95\x00\xE3\x80\x96\x00\xE3"
-     "\x80\x97"
-     "\x00\xE3\x81\x8C\x00\xE3\x81\x8E\x00\xE3\x81\x90\x00\xE3\x81\x92\x00\xE3"
-     "\x81\x94"
-     "\x00\xE3\x81\x96\x00\xE3\x81\x98\x00\xE3\x81\x9A\x00\xE3\x81\x9C\x00\xE3"
-     "\x81\x9E"
-     "\x00\xE3\x81\xA0\x00\xE3\x81\xA2\x00\xE3\x81\xA5\x00\xE3\x81\xA7\x00\xE3"
-     "\x81\xA9"
-     "\x00\xE3\x81\xB0\x00\xE3\x81\xB1\x00\xE3\x81\xB3\x00\xE3\x81\xB4\x00\xE3"
-     "\x81\xB6"
-     "\x00\xE3\x81\xB7\x00\xE3\x81\xB9\x00\xE3\x81\xBA\x00\xE3\x81\xBB\xE3\x81"
-     "\x8B\x00"
-     "\xE3\x81\xBC\x00\xE3\x81\xBD\x00\xE3\x82\x88\xE3\x82\x8A\x00\xE3\x82\x94"
-     "\x00\xE3"
-     "\x82\x99\x00\xE3\x82\x9A\x00\xE3\x82\x9E\x00\xE3\x82\xA1\x00\xE3\x82\xA2"
-     "\x00\xE3"
-     "\x82\xA2\xE3\x83\x91\xE3\x83\xBC\xE3\x83\x88\x00\xE3\x82\xA2\xE3\x83\xAB"
-     "\xE3\x83"
-     "\x95\xE3\x82\xA1\x00\xE3\x82\xA2\xE3\x83\xB3\xE3\x83\x9A\xE3\x82\xA2\x00"
-     "\xE3\x82"
-     "\xA2\xE3\x83\xBC\xE3\x83\xAB\x00\xE3\x82\xA3\x00\xE3\x82\xA4\x00\xE3\x82"
-     "\xA4\xE3"
-     "\x83\x8B\xE3\x83\xB3\xE3\x82\xB0\x00\xE3\x82\xA4\xE3\x83\xB3\xE3\x83\x81"
-     "\x00\xE3"
-     "\x82\xA5\x00\xE3\x82\xA6\x00\xE3\x82\xA6\xE3\x82\xA9\xE3\x83\xB3\x00\xE3"
-     "\x82\xA7"
-     "\x00\xE3\x82\xA8\x00\xE3\x82\xA8\xE3\x82\xB9\xE3\x82\xAF\xE3\x83\xBC\xE3"
-     "\x83\x89"
-     "\x00\xE3\x82\xA8\xE3\x83\xBC\xE3\x82\xAB\xE3\x83\xBC\x00\xE3\x82\xA9\x00"
-     "\xE3\x82"
-     "\xAA\x00\xE3\x82\xAA\xE3\x83\xB3\xE3\x82\xB9\x00\xE3\x82\xAA\xE3\x83\xBC"
-     "\xE3\x83"
-     "\xA0\x00\xE3\x82\xAB\x00\xE3\x82\xAB\xE3\x82\xA4\xE3\x83\xAA\x00\xE3\x82"
-     "\xAB\xE3"
-     "\x83\xA9\xE3\x83\x83\xE3\x83\x88\x00\xE3\x82\xAB\xE3\x83\xAD\xE3\x83\xAA"
-     "\xE3\x83"
-     "\xBC\x00\xE3\x82\xAC\x00\xE3\x82\xAC\xE3\x83\xAD\xE3\x83\xB3\x00\xE3\x82"
-     "\xAC\xE3"
-     "\x83\xB3\xE3\x83\x9E\x00\xE3\x82\xAD\x00\xE3\x82\xAD\xE3\x83\xA5\xE3\x83"
-     "\xAA\xE3"
-     "\x83\xBC\x00\xE3\x82\xAD\xE3\x83\xAD\x00\xE3\x82\xAD\xE3\x83\xAD\xE3\x82"
-     "\xB0\xE3"
-     "\x83\xA9\xE3\x83\xA0\x00\xE3\x82\xAD\xE3\x83\xAD\xE3\x83\xA1\xE3\x83\xBC"
-     "\xE3\x83"
-     "\x88\xE3\x83\xAB\x00\xE3\x82\xAD\xE3\x83\xAD\xE3\x83\xAF\xE3\x83\x83\xE3"
-     "\x83\x88"
-     "\x00\xE3\x82\xAE\x00\xE3\x82\xAE\xE3\x82\xAC\x00\xE3\x82\xAE\xE3\x83\x8B"
-     "\xE3\x83"
-     "\xBC\x00\xE3\x82\xAE\xE3\x83\xAB\xE3\x83\x80\xE3\x83\xBC\x00\xE3\x82\xAF"
-     "\x00\xE3"
-     "\x82\xAF\xE3\x83\xAB\xE3\x82\xBC\xE3\x82\xA4\xE3\x83\xAD\x00\xE3\x82\xAF"
-     "\xE3\x83"
-     "\xAD\xE3\x83\xBC\xE3\x83\x8D\x00\xE3\x82\xB0\x00\xE3\x82\xB0\xE3\x83\xA9"
-     "\xE3\x83"
-     "\xA0\x00\xE3\x82\xB0\xE3\x83\xA9\xE3\x83\xA0\xE3\x83\x88\xE3\x83\xB3\x00"
-     "\xE3\x82"
-     "\xB1\x00\xE3\x82\xB1\xE3\x83\xBC\xE3\x82\xB9\x00\xE3\x82\xB2\x00\xE3\x82"
-     "\xB3\x00"
-     "\xE3\x82\xB3\xE3\x82\xB3\x00\xE3\x82\xB3\xE3\x83\x88\x00\xE3\x82\xB3\xE3"
-     "\x83\xAB"
-     "\xE3\x83\x8A\x00\xE3\x82\xB3\xE3\x83\xBC\xE3\x83\x9D\x00\xE3\x82\xB4\x00"
-     "\xE3\x82"
-     "\xB5\x00\xE3\x82\xB5\xE3\x82\xA4\xE3\x82\xAF\xE3\x83\xAB\x00\xE3\x82\xB5"
-     "\xE3\x83"
-     "\xB3\xE3\x83\x81\xE3\x83\xBC\xE3\x83\xA0\x00\xE3\x82\xB6\x00\xE3\x82\xB7"
-     "\x00\xE3"
-     "\x82\xB7\xE3\x83\xAA\xE3\x83\xB3\xE3\x82\xB0\x00\xE3\x82\xB8\x00\xE3\x82"
-     "\xB9\x00"
-     "\xE3\x82\xBA\x00\xE3\x82\xBB\x00\xE3\x82\xBB\xE3\x83\xB3\xE3\x83\x81\x00"
-     "\xE3\x82"
-     "\xBB\xE3\x83\xB3\xE3\x83\x88\x00\xE3\x82\xBC\x00\xE3\x82\xBD\x00\xE3\x82"
-     "\xBE\x00"
-     "\xE3\x82\xBF\x00\xE3\x83\x80\x00\xE3\x83\x80\xE3\x83\xBC\xE3\x82\xB9\x00"
-     "\xE3\x83"
-     "\x81\x00\xE3\x83\x82\x00\xE3\x83\x83\x00\xE3\x83\x84\x00\xE3\x83\x85\x00"
-     "\xE3\x83"
-     "\x86\x00\xE3\x83\x87\x00\xE3\x83\x87\xE3\x82\xB7\x00\xE3\x83\x88\x00\xE3"
-     "\x83\x88"
-     "\xE3\x83\xB3\x00\xE3\x83\x89\x00\xE3\x83\x89\xE3\x83\xAB\x00\xE3\x83\x8A"
-     "\x00\xE3"
-     "\x83\x8A\xE3\x83\x8E\x00\xE3\x83\x8B\x00\xE3\x83\x8C\x00\xE3\x83\x8D\x00"
-     "\xE3\x83"
-     "\x8E\x00\xE3\x83\x8E\xE3\x83\x83\xE3\x83\x88\x00\xE3\x83\x8F\x00\xE3\x83"
-     "\x8F\xE3"
-     "\x82\xA4\xE3\x83\x84\x00\xE3\x83\x90\x00\xE3\x83\x90\xE3\x83\xBC\xE3\x83"
-     "\xAC\xE3"
-     "\x83\xAB\x00\xE3\x83\x91\x00\xE3\x83\x91\xE3\x83\xBC\xE3\x82\xBB\xE3\x83"
-     "\xB3\xE3"
-     "\x83\x88\x00\xE3\x83\x91\xE3\x83\xBC\xE3\x83\x84\x00\xE3\x83\x92\x00\xE3"
-     "\x83\x93"
-     "\x00\xE3\x83\x93\xE3\x83\xAB\x00\xE3\x83\x94\x00\xE3\x83\x94\xE3\x82\xA2"
-     "\xE3\x82"
-     "\xB9\xE3\x83\x88\xE3\x83\xAB\x00\xE3\x83\x94\xE3\x82\xAF\xE3\x83\xAB\x00"
-     "\xE3\x83"
-     "\x94\xE3\x82\xB3\x00\xE3\x83\x95\x00\xE3\x83\x95\xE3\x82\xA1\xE3\x83\xA9"
-     "\xE3\x83"
-     "\x83\xE3\x83\x89\x00\xE3\x83\x95\xE3\x82\xA3\xE3\x83\xBC\xE3\x83\x88\x00"
-     "\xE3\x83"
-     "\x95\xE3\x83\xA9\xE3\x83\xB3\x00\xE3\x83\x96\x00\xE3\x83\x96\xE3\x83\x83"
-     "\xE3\x82"
-     "\xB7\xE3\x82\xA7\xE3\x83\xAB\x00\xE3\x83\x97\x00\xE3\x83\x98\x00\xE3\x83"
-     "\x98\xE3"
-     "\x82\xAF\xE3\x82\xBF\xE3\x83\xBC\xE3\x83\xAB\x00\xE3\x83\x98\xE3\x83\xAB"
-     "\xE3\x83"
-     "\x84\x00\xE3\x83\x99\x00\xE3\x83\x99\xE3\x83\xBC\xE3\x82\xBF\x00\xE3\x83"
-     "\x9A\x00"
-     "\xE3\x83\x9A\xE3\x82\xBD\x00\xE3\x83\x9A\xE3\x83\x8B\xE3\x83\x92\x00\xE3"
-     "\x83\x9A"
-     "\xE3\x83\xB3\xE3\x82\xB9\x00\xE3\x83\x9A\xE3\x83\xBC\xE3\x82\xB8\x00\xE3"
-     "\x83\x9B"
-     "\x00\xE3\x83\x9B\xE3\x83\xB3\x00\xE3\x83\x9B\xE3\x83\xBC\xE3\x83\xAB\x00"
-     "\xE3\x83"
-     "\x9B\xE3\x83\xBC\xE3\x83\xB3\x00\xE3\x83\x9C\x00\xE3\x83\x9C\xE3\x83\xAB"
-     "\xE3\x83"
-     "\x88\x00\xE3\x83\x9D\x00\xE3\x83\x9D\xE3\x82\xA4\xE3\x83\xB3\xE3\x83\x88"
-     "\x00\xE3"
-     "\x83\x9D\xE3\x83\xB3\xE3\x83\x89\x00\xE3\x83\x9E\x00\xE3\x83\x9E\xE3\x82"
-     "\xA4\xE3"
-     "\x82\xAF\xE3\x83\xAD\x00\xE3\x83\x9E\xE3\x82\xA4\xE3\x83\xAB\x00\xE3\x83"
-     "\x9E\xE3"
-     "\x83\x83\xE3\x83\x8F\x00\xE3\x83\x9E\xE3\x83\xAB\xE3\x82\xAF\x00\xE3\x83"
-     "\x9E\xE3"
-     "\x83\xB3\xE3\x82\xB7\xE3\x83\xA7\xE3\x83\xB3\x00\xE3\x83\x9F\x00\xE3\x83"
-     "\x9F\xE3"
-     "\x82\xAF\xE3\x83\xAD\xE3\x83\xB3\x00\xE3\x83\x9F\xE3\x83\xAA\x00\xE3\x83"
-     "\x9F\xE3"
-     "\x83\xAA\xE3\x83\x90\xE3\x83\xBC\xE3\x83\xAB\x00\xE3\x83\xA0\x00\xE3\x83"
-     "\xA1\x00"
-     "\xE3\x83\xA1\xE3\x82\xAC\x00\xE3\x83\xA1\xE3\x82\xAC\xE3\x83\x88\xE3\x83"
-     "\xB3\x00"
-     "\xE3\x83\xA1\xE3\x83\xBC\xE3\x83\x88\xE3\x83\xAB\x00\xE3\x83\xA2\x00\xE3"
-     "\x83\xA3"
-     "\x00\xE3\x83\xA4\x00\xE3\x83\xA4\xE3\x83\xBC\xE3\x83\x89\x00\xE3\x83\xA4"
-     "\xE3\x83"
-     "\xBC\xE3\x83\xAB\x00\xE3\x83\xA5\x00\xE3\x83\xA6\x00\xE3\x83\xA6\xE3\x82"
-     "\xA2\xE3"
-     "\x83\xB3\x00\xE3\x83\xA7\x00\xE3\x83\xA8\x00\xE3\x83\xA9\x00\xE3\x83\xAA"
-     "\x00\xE3"
-     "\x83\xAA\xE3\x83\x83\xE3\x83\x88\xE3\x83\xAB\x00\xE3\x83\xAA\xE3\x83\xA9"
-     "\x00\xE3"
-     "\x83\xAB\x00\xE3\x83\xAB\xE3\x83\x94\xE3\x83\xBC\x00\xE3\x83\xAB\xE3\x83"
-     "\xBC\xE3"
-     "\x83\x96\xE3\x83\xAB\x00\xE3\x83\xAC\x00\xE3\x83\xAC\xE3\x83\xA0\x00\xE3"
-     "\x83\xAC"
-     "\xE3\x83\xB3\xE3\x83\x88\xE3\x82\xB2\xE3\x83\xB3\x00\xE3\x83\xAD\x00\xE3"
-     "\x83\xAF"
-     "\x00\xE3\x83\xAF\xE3\x83\x83\xE3\x83\x88\x00\xE3\x83\xB0\x00\xE3\x83\xB1"
-     "\x00\xE3"
-     "\x83\xB2\x00\xE3\x83\xB3\x00\xE3\x83\xB4\x00\xE3\x83\xB7\x00\xE3\x83\xB8"
-     "\x00\xE3"
-     "\x83\xB9\x00\xE3\x83\xBA\x00\xE3\x83\xBB\x00\xE3\x83\xBC\x00\xE3\x83\xBE"
-     "\x00\xE3"
-     "\x92\x9E\x00\xE3\x92\xB9\x00\xE3\x92\xBB\x00\xE3\x93\x9F\x00\xE3\x94\x95"
-     "\x00\xE3"
-     "\x9B\xAE\x00\xE3\x9B\xBC\x00\xE3\x9E\x81\x00\xE3\xA0\xAF\x00\xE3\xA1\xA2"
-     "\x00\xE3"
-     "\xA1\xBC\x00\xE3\xA3\x87\x00\xE3\xA3\xA3\x00\xE3\xA4\x9C\x00\xE3\xA4\xBA"
-     "\x00\xE3"
-     "\xA8\xAE\x00\xE3\xA9\xAC\x00\xE3\xAB\xA4\x00\xE3\xAC\x88\x00\xE3\xAC\x99"
-     "\x00\xE3"
-     "\xAD\x89\x00\xE3\xAE\x9D\x00\xE3\xB0\x98\x00\xE3\xB1\x8E\x00\xE3\xB4\xB3"
-     "\x00\xE3"
-     "\xB6\x96\x00\xE3\xBA\xAC\x00\xE3\xBA\xB8\x00\xE3\xBC\x9B\x00\xE3\xBF\xBC"
-     "\x00\xE4"
-     "\x80\x88\x00\xE4\x80\x98\x00\xE4\x80\xB9\x00\xE4\x81\x86\x00\xE4\x82\x96"
-     "\x00\xE4"
-     "\x83\xA3\x00\xE4\x84\xAF\x00\xE4\x88\x82\x00\xE4\x88\xA7\x00\xE4\x8A\xA0"
-     "\x00\xE4"
-     "\x8C\x81\x00\xE4\x8C\xB4\x00\xE4\x8D\x99\x00\xE4\x8F\x95\x00\xE4\x8F\x99"
-     "\x00\xE4"
-     "\x90\x8B\x00\xE4\x91\xAB\x00\xE4\x94\xAB\x00\xE4\x95\x9D\x00\xE4\x95\xA1"
-     "\x00\xE4"
-     "\x95\xAB\x00\xE4\x97\x97\x00\xE4\x97\xB9\x00\xE4\x98\xB5\x00\xE4\x9A\xBE"
-     "\x00\xE4"
-     "\x9B\x87\x00\xE4\xA6\x95\x00\xE4\xA7\xA6\x00\xE4\xA9\xAE\x00\xE4\xA9\xB6"
-     "\x00\xE4"
-     "\xAA\xB2\x00\xE4\xAC\xB3\x00\xE4\xAF\x8E\x00\xE4\xB3\x8E\x00\xE4\xB3\xAD"
-     "\x00\xE4"
-     "\xB3\xB8\x00\xE4\xB5\x96\x00\xE4\xB8\x80\x00\xE4\xB8\x81\x00\xE4\xB8\x83"
-     "\x00\xE4"
-     "\xB8\x89\x00\xE4\xB8\x8A\x00\xE4\xB8\x8B\x00\xE4\xB8\x8D\x00\xE4\xB8\x99"
-     "\x00\xE4"
-     "\xB8\xA6\x00\xE4\xB8\xA8\x00\xE4\xB8\xAD\x00\xE4\xB8\xB2\x00\xE4\xB8\xB6"
-     "\x00\xE4"
-     "\xB8\xB8\x00\xE4\xB8\xB9\x00\xE4\xB8\xBD\x00\xE4\xB8\xBF\x00\xE4\xB9\x81"
-     "\x00\xE4"
-     "\xB9\x99\x00\xE4\xB9\x9D\x00\xE4\xBA\x82\x00\xE4\xBA\x85\x00\xE4\xBA\x86"
-     "\x00\xE4"
-     "\xBA\x8C\x00\xE4\xBA\x94\x00\xE4\xBA\xA0\x00\xE4\xBA\xA4\x00\xE4\xBA\xAE"
-     "\x00\xE4"
-     "\xBA\xBA\x00\xE4\xBB\x80\x00\xE4\xBB\x8C\x00\xE4\xBB\xA4\x00\xE4\xBB\xA4"
-     "\xE5\x92"
-     "\x8C\x00\xE4\xBC\x81\x00\xE4\xBC\x91\x00\xE4\xBD\xA0\x00\xE4\xBE\x80\x00"
-     "\xE4\xBE"
-     "\x86\x00\xE4\xBE\x8B\x00\xE4\xBE\xAE\x00\xE4\xBE\xBB\x00\xE4\xBE\xBF\x00"
-     "\xE5\x80"
-     "\x82\x00\xE5\x80\xAB\x00\xE5\x81\xBA\x00\xE5\x82\x99\x00\xE5\x83\x8F\x00"
-     "\xE5\x83"
-     "\x9A\x00\xE5\x83\xA7\x00\xE5\x84\xAA\x00\xE5\x84\xBF\x00\xE5\x85\x80\x00"
-     "\xE5\x85"
-     "\x85\x00\xE5\x85\x8D\x00\xE5\x85\x94\x00\xE5\x85\xA4\x00\xE5\x85\xA5\x00"
-     "\xE5\x85"
-     "\xA7\x00\xE5\x85\xA8\x00\xE5\x85\xA9\x00\xE5\x85\xAB\x00\xE5\x85\xAD\x00"
-     "\xE5\x85"
-     "\xB7\x00\xE5\x86\x80\x00\xE5\x86\x82\x00\xE5\x86\x8D\x00\xE5\x86\x92\x00"
-     "\xE5\x86"
-     "\x95\x00\xE5\x86\x96\x00\xE5\x86\x97\x00\xE5\x86\x99\x00\xE5\x86\xA4\x00"
-     "\xE5\x86"
-     "\xAB\x00\xE5\x86\xAC\x00\xE5\x86\xB5\x00\xE5\x86\xB7\x00\xE5\x87\x89\x00"
-     "\xE5\x87"
-     "\x8C\x00\xE5\x87\x9C\x00\xE5\x87\x9E\x00\xE5\x87\xA0\x00\xE5\x87\xB5\x00"
-     "\xE5\x88"
-     "\x80\x00\xE5\x88\x83\x00\xE5\x88\x87\x00\xE5\x88\x97\x00\xE5\x88\x9D\x00"
-     "\xE5\x88"
-     "\xA9\x00\xE5\x88\xBA\x00\xE5\x88\xBB\x00\xE5\x89\x86\x00\xE5\x89\x8D\x00"
-     "\xE5\x89"
-     "\xB2\x00\xE5\x89\xB7\x00\xE5\x8A\x89\x00\xE5\x8A\x9B\x00\xE5\x8A\xA3\x00"
-     "\xE5\x8A"
-     "\xB3\x00\xE5\x8A\xB4\x00\xE5\x8B\x87\x00\xE5\x8B\x89\x00\xE5\x8B\x92\x00"
-     "\xE5\x8B"
-     "\x9E\x00\xE5\x8B\xA4\x00\xE5\x8B\xB5\x00\xE5\x8B\xB9\x00\xE5\x8B\xBA\x00"
-     "\xE5\x8C"
-     "\x85\x00\xE5\x8C\x86\x00\xE5\x8C\x95\x00\xE5\x8C\x97\x00\xE5\x8C\x9A\x00"
-     "\xE5\x8C"
-     "\xB8\x00\xE5\x8C\xBB\x00\xE5\x8C\xBF\x00\xE5\x8D\x81\x00\xE5\x8D\x84\x00"
-     "\xE5\x8D"
-     "\x85\x00\xE5\x8D\x89\x00\xE5\x8D\x91\x00\xE5\x8D\x94\x00\xE5\x8D\x9A\x00"
-     "\xE5\x8D"
-     "\x9C\x00\xE5\x8D\xA9\x00\xE5\x8D\xB0\x00\xE5\x8D\xB3\x00\xE5\x8D\xB5\x00"
-     "\xE5\x8D"
-     "\xBD\x00\xE5\x8D\xBF\x00\xE5\x8E\x82\x00\xE5\x8E\xB6\x00\xE5\x8F\x83\x00"
-     "\xE5\x8F"
-     "\x88\x00\xE5\x8F\x8A\x00\xE5\x8F\x8C\x00\xE5\x8F\x9F\x00\xE5\x8F\xA3\x00"
-     "\xE5\x8F"
-     "\xA5\x00\xE5\x8F\xAB\x00\xE5\x8F\xAF\x00\xE5\x8F\xB1\x00\xE5\x8F\xB3\x00"
-     "\xE5\x90"
-     "\x86\x00\xE5\x90\x88\x00\xE5\x90\x8D\x00\xE5\x90\x8F\x00\xE5\x90\x9D\x00"
-     "\xE5\x90"
-     "\xB8\x00\xE5\x90\xB9\x00\xE5\x91\x82\x00\xE5\x91\x88\x00\xE5\x91\xA8\x00"
-     "\xE5\x92"
-     "\x9E\x00\xE5\x92\xA2\x00\xE5\x92\xBD\x00\xE5\x93\xB6\x00\xE5\x94\x90\x00"
-     "\xE5\x95"
-     "\x8F\x00\xE5\x95\x93\x00\xE5\x95\x95\x00\xE5\x95\xA3\x00\xE5\x96\x84\x00"
-     "\xE5\x96"
-     "\x87\x00\xE5\x96\x99\x00\xE5\x96\x9D\x00\xE5\x96\xAB\x00\xE5\x96\xB3\x00"
-     "\xE5\x96"
-     "\xB6\x00\xE5\x97\x80\x00\xE5\x97\x82\x00\xE5\x97\xA2\x00\xE5\x98\x86\x00"
-     "\xE5\x99"
-     "\x91\x00\xE5\x99\xA8\x00\xE5\x99\xB4\x00\xE5\x9B\x97\x00\xE5\x9B\x9B\x00"
-     "\xE5\x9B"
-     "\xB9\x00\xE5\x9C\x96\x00\xE5\x9C\x97\x00\xE5\x9C\x9F\x00\xE5\x9C\xB0\x00"
-     "\xE5\x9E"
-     "\x8B\x00\xE5\x9F\x8E\x00\xE5\x9F\xB4\x00\xE5\xA0\x8D\x00\xE5\xA0\xB1\x00"
-     "\xE5\xA0"
-     "\xB2\x00\xE5\xA1\x80\x00\xE5\xA1\x9A\x00\xE5\xA1\x9E\x00\xE5\xA2\xA8\x00"
-     "\xE5\xA2"
-     "\xAC\x00\xE5\xA2\xB3\x00\xE5\xA3\x98\x00\xE5\xA3\x9F\x00\xE5\xA3\xAB\x00"
-     "\xE5\xA3"
-     "\xAE\x00\xE5\xA3\xB0\x00\xE5\xA3\xB2\x00\xE5\xA3\xB7\x00\xE5\xA4\x82\x00"
-     "\xE5\xA4"
-     "\x86\x00\xE5\xA4\x8A\x00\xE5\xA4\x95\x00\xE5\xA4\x9A\x00\xE5\xA4\x9C\x00"
-     "\xE5\xA4"
-     "\xA2\x00\xE5\xA4\xA7\x00\xE5\xA4\xA7\xE6\xAD\xA3\x00\xE5\xA4\xA9\x00\xE5"
-     "\xA5\x84"
-     "\x00\xE5\xA5\x88\x00\xE5\xA5\x91\x00\xE5\xA5\x94\x00\xE5\xA5\xA2\x00\xE5"
-     "\xA5\xB3"
-     "\x00\xE5\xA7\x98\x00\xE5\xA7\xAC\x00\xE5\xA8\x9B\x00\xE5\xA8\xA7\x00\xE5"
-     "\xA9\xA2"
-     "\x00\xE5\xA9\xA6\x00\xE5\xAA\xB5\x00\xE5\xAC\x88\x00\xE5\xAC\xA8\x00\xE5"
-     "\xAC\xBE"
-     "\x00\xE5\xAD\x90\x00\xE5\xAD\x97\x00\xE5\xAD\xA6\x00\xE5\xAE\x80\x00\xE5"
-     "\xAE\x85"
-     "\x00\xE5\xAE\x97\x00\xE5\xAF\x83\x00\xE5\xAF\x98\x00\xE5\xAF\xA7\x00\xE5"
-     "\xAF\xAE"
-     "\x00\xE5\xAF\xB3\x00\xE5\xAF\xB8\x00\xE5\xAF\xBF\x00\xE5\xB0\x86\x00\xE5"
-     "\xB0\x8F"
-     "\x00\xE5\xB0\xA2\x00\xE5\xB0\xB8\x00\xE5\xB0\xBF\x00\xE5\xB1\xA0\x00\xE5"
-     "\xB1\xA2"
-     "\x00\xE5\xB1\xA4\x00\xE5\xB1\xA5\x00\xE5\xB1\xAE\x00\xE5\xB1\xB1\x00\xE5"
-     "\xB2\x8D"
-     "\x00\xE5\xB3\x80\x00\xE5\xB4\x99\x00\xE5\xB5\x83\x00\xE5\xB5\x90\x00\xE5"
-     "\xB5\xAB"
-     "\x00\xE5\xB5\xAE\x00\xE5\xB5\xBC\x00\xE5\xB6\xB2\x00\xE5\xB6\xBA\x00\xE5"
-     "\xB7\x9B"
-     "\x00\xE5\xB7\xA1\x00\xE5\xB7\xA2\x00\xE5\xB7\xA5\x00\xE5\xB7\xA6\x00\xE5"
-     "\xB7\xB1"
-     "\x00\xE5\xB7\xBD\x00\xE5\xB7\xBE\x00\xE5\xB8\xA8\x00\xE5\xB8\xBD\x00\xE5"
-     "\xB9\xA9"
-     "\x00\xE5\xB9\xB2\x00\xE5\xB9\xB3\xE6\x88\x90\x00\xE5\xB9\xB4\x00\xE5\xB9"
-     "\xBA\x00"
-     "\xE5\xB9\xBC\x00\xE5\xB9\xBF\x00\xE5\xBA\xA6\x00\xE5\xBA\xB0\x00\xE5\xBA"
-     "\xB3\x00"
-     "\xE5\xBA\xB6\x00\xE5\xBB\x89\x00\xE5\xBB\x8A\x00\xE5\xBB\x92\x00\xE5\xBB"
-     "\x93\x00"
-     "\xE5\xBB\x99\x00\xE5\xBB\xAC\x00\xE5\xBB\xB4\x00\xE5\xBB\xBE\x00\xE5\xBC"
-     "\x84\x00"
-     "\xE5\xBC\x8B\x00\xE5\xBC\x93\x00\xE5\xBC\xA2\x00\xE5\xBD\x90\x00\xE5\xBD"
-     "\x93\x00"
-     "\xE5\xBD\xA1\x00\xE5\xBD\xA2\x00\xE5\xBD\xA9\x00\xE5\xBD\xAB\x00\xE5\xBD"
-     "\xB3\x00"
-     "\xE5\xBE\x8B\x00\xE5\xBE\x8C\x00\xE5\xBE\x97\x00\xE5\xBE\x9A\x00\xE5\xBE"
-     "\xA9\x00"
-     "\xE5\xBE\xAD\x00\xE5\xBF\x83\x00\xE5\xBF\x8D\x00\xE5\xBF\x97\x00\xE5\xBF"
-     "\xB5\x00"
-     "\xE5\xBF\xB9\x00\xE6\x80\x92\x00\xE6\x80\x9C\x00\xE6\x81\xB5\x00\xE6\x82"
-     "\x81\x00"
-     "\xE6\x82\x94\x00\xE6\x83\x87\x00\xE6\x83\x98\x00\xE6\x83\xA1\x00\xE6\x84"
-     "\x88\x00"
-     "\xE6\x85\x84\x00\xE6\x85\x88\x00\xE6\x85\x8C\x00\xE6\x85\x8E\x00\xE6\x85"
-     "\xA0\x00"
-     "\xE6\x85\xA8\x00\xE6\x85\xBA\x00\xE6\x86\x8E\x00\xE6\x86\x90\x00\xE6\x86"
-     "\xA4\x00"
-     "\xE6\x86\xAF\x00\xE6\x86\xB2\x00\xE6\x87\x9E\x00\xE6\x87\xB2\x00\xE6\x87"
-     "\xB6\x00"
-     "\xE6\x88\x80\x00\xE6\x88\x88\x00\xE6\x88\x90\x00\xE6\x88\x9B\x00\xE6\x88"
-     "\xAE\x00"
-     "\xE6\x88\xB4\x00\xE6\x88\xB6\x00\xE6\x89\x8B\x00\xE6\x89\x93\x00\xE6\x89"
-     "\x9D\x00"
-     "\xE6\x8A\x95\x00\xE6\x8A\xB1\x00\xE6\x8B\x89\x00\xE6\x8B\x8F\x00\xE6\x8B"
-     "\x93\x00"
-     "\xE6\x8B\x94\x00\xE6\x8B\xBC\x00\xE6\x8B\xBE\x00\xE6\x8C\x87\x00\xE6\x8C"
-     "\xBD\x00"
-     "\xE6\x8D\x90\x00\xE6\x8D\x95\x00\xE6\x8D\xA8\x00\xE6\x8D\xBB\x00\xE6\x8E"
-     "\x83\x00"
-     "\xE6\x8E\xA0\x00\xE6\x8E\xA9\x00\xE6\x8F\x84\x00\xE6\x8F\x85\x00\xE6\x8F"
-     "\xA4\x00"
-     "\xE6\x90\x9C\x00\xE6\x90\xA2\x00\xE6\x91\x92\x00\xE6\x91\xA9\x00\xE6\x91"
-     "\xB7\x00"
-     "\xE6\x91\xBE\x00\xE6\x92\x9A\x00\xE6\x92\x9D\x00\xE6\x93\x84\x00\xE6\x94"
-     "\xAF\x00"
-     "\xE6\x94\xB4\x00\xE6\x95\x8F\x00\xE6\x95\x96\x00\xE6\x95\xAC\x00\xE6\x95"
-     "\xB8\x00"
-     "\xE6\x96\x87\x00\xE6\x96\x97\x00\xE6\x96\x99\x00\xE6\x96\xA4\x00\xE6\x96"
-     "\xB0\x00"
-     "\xE6\x96\xB9\x00\xE6\x97\x85\x00\xE6\x97\xA0\x00\xE6\x97\xA2\x00\xE6\x97"
-     "\xA3\x00"
-     "\xE6\x97\xA5\x00\xE6\x98\x8E\xE6\xB2\xBB\x00\xE6\x98\x93\x00\xE6\x98\xA0"
-     "\x00\xE6"
-     "\x98\xAD\xE5\x92\x8C\x00\xE6\x99\x89\x00\xE6\x99\xB4\x00\xE6\x9A\x88\x00"
-     "\xE6\x9A"
-     "\x91\x00\xE6\x9A\x9C\x00\xE6\x9A\xB4\x00\xE6\x9B\x86\x00\xE6\x9B\xB0\x00"
-     "\xE6\x9B"
-     "\xB4\x00\xE6\x9B\xB8\x00\xE6\x9C\x80\x00\xE6\x9C\x88\x00\xE6\x9C\x89\x00"
-     "\xE6\x9C"
-     "\x97\x00\xE6\x9C\x9B\x00\xE6\x9C\xA1\x00\xE6\x9C\xA8\x00\xE6\x9D\x8E\x00"
-     "\xE6\x9D"
-     "\x93\x00\xE6\x9D\x96\x00\xE6\x9D\x9E\x00\xE6\x9D\xBB\x00\xE6\x9E\x85\x00"
-     "\xE6\x9E"
-     "\x97\x00\xE6\x9F\xB3\x00\xE6\x9F\xBA\x00\xE6\xA0\x97\x00\xE6\xA0\x9F\x00"
-     "\xE6\xA0"
-     "\xAA\x00\xE6\xA0\xAA\xE5\xBC\x8F\xE4\xBC\x9A\xE7\xA4\xBE\x00\xE6\xA1\x92"
-     "\x00\xE6"
-     "\xA2\x81\x00\xE6\xA2\x85\x00\xE6\xA2\x8E\x00\xE6\xA2\xA8\x00\xE6\xA4\x94"
-     "\x00\xE6"
-     "\xA5\x82\x00\xE6\xA6\xA3\x00\xE6\xA7\xAA\x00\xE6\xA8\x82\x00\xE6\xA8\x93"
-     "\x00\xE6"
-     "\xAA\xA8\x00\xE6\xAB\x93\x00\xE6\xAB\x9B\x00\xE6\xAC\x84\x00\xE6\xAC\xA0"
-     "\x00\xE6"
-     "\xAC\xA1\x00\xE6\xAD\x94\x00\xE6\xAD\xA2\x00\xE6\xAD\xA3\x00\xE6\xAD\xB2"
-     "\x00\xE6"
-     "\xAD\xB7\x00\xE6\xAD\xB9\x00\xE6\xAE\x9F\x00\xE6\xAE\xAE\x00\xE6\xAE\xB3"
-     "\x00\xE6"
-     "\xAE\xBA\x00\xE6\xAE\xBB\x00\xE6\xAF\x8B\x00\xE6\xAF\x8D\x00\xE6\xAF\x94"
-     "\x00\xE6"
-     "\xAF\x9B\x00\xE6\xB0\x8F\x00\xE6\xB0\x94\x00\xE6\xB0\xB4\x00\xE6\xB1\x8E"
-     "\x00\xE6"
-     "\xB1\xA7\x00\xE6\xB2\x88\x00\xE6\xB2\xBF\x00\xE6\xB3\x8C\x00\xE6\xB3\x8D"
-     "\x00\xE6"
-     "\xB3\xA5\x00\xE6\xB3\xA8\x00\xE6\xB4\x96\x00\xE6\xB4\x9B\x00\xE6\xB4\x9E"
-     "\x00\xE6"
-     "\xB4\xB4\x00\xE6\xB4\xBE\x00\xE6\xB5\x81\x00\xE6\xB5\xA9\x00\xE6\xB5\xAA"
-     "\x00\xE6"
-     "\xB5\xB7\x00\xE6\xB5\xB8\x00\xE6\xB6\x85\x00\xE6\xB7\x8B\x00\xE6\xB7\x9A"
-     "\x00\xE6"
-     "\xB7\xAA\x00\xE6\xB7\xB9\x00\xE6\xB8\x9A\x00\xE6\xB8\xAF\x00\xE6\xB9\xAE"
-     "\x00\xE6"
-     "\xBA\x80\x00\xE6\xBA\x9C\x00\xE6\xBA\xBA\x00\xE6\xBB\x87\x00\xE6\xBB\x8B"
-     "\x00\xE6"
-     "\xBB\x91\x00\xE6\xBB\x9B\x00\xE6\xBC\x8F\x00\xE6\xBC\x94\x00\xE6\xBC\xA2"
-     "\x00\xE6"
-     "\xBC\xA3\x00\xE6\xBD\xAE\x00\xE6\xBF\x86\x00\xE6\xBF\xAB\x00\xE6\xBF\xBE"
-     "\x00\xE7"
-     "\x80\x9B\x00\xE7\x80\x9E\x00\xE7\x80\xB9\x00\xE7\x81\x8A\x00\xE7\x81\xAB"
-     "\x00\xE7"
-     "\x81\xB0\x00\xE7\x81\xB7\x00\xE7\x81\xBD\x00\xE7\x82\x99\x00\xE7\x82\xAD"
-     "\x00\xE7"
-     "\x83\x88\x00\xE7\x83\x99\x00\xE7\x84\xA1\x00\xE7\x85\x85\x00\xE7\x85\x89"
-     "\x00\xE7"
-     "\x85\xAE\x00\xE7\x86\x9C\x00\xE7\x87\x8E\x00\xE7\x87\x90\x00\xE7\x88\x90"
-     "\x00\xE7"
-     "\x88\x9B\x00\xE7\x88\xA8\x00\xE7\x88\xAA\x00\xE7\x88\xAB\x00\xE7\x88\xB5"
-     "\x00\xE7"
-     "\x88\xB6\x00\xE7\x88\xBB\x00\xE7\x88\xBF\x00\xE7\x89\x87\x00\xE7\x89\x90"
-     "\x00\xE7"
-     "\x89\x99\x00\xE7\x89\x9B\x00\xE7\x89\xA2\x00\xE7\x89\xB9\x00\xE7\x8A\x80"
-     "\x00\xE7"
-     "\x8A\x95\x00\xE7\x8A\xAC\x00\xE7\x8A\xAF\x00\xE7\x8B\x80\x00\xE7\x8B\xBC"
-     "\x00\xE7"
-     "\x8C\xAA\x00\xE7\x8D\xB5\x00\xE7\x8D\xBA\x00\xE7\x8E\x84\x00\xE7\x8E\x87"
-     "\x00\xE7"
-     "\x8E\x89\x00\xE7\x8E\x8B\x00\xE7\x8E\xA5\x00\xE7\x8E\xB2\x00\xE7\x8F\x9E"
-     "\x00\xE7"
-     "\x90\x86\x00\xE7\x90\x89\x00\xE7\x90\xA2\x00\xE7\x91\x87\x00\xE7\x91\x9C"
-     "\x00\xE7"
-     "\x91\xA9\x00\xE7\x91\xB1\x00\xE7\x92\x85\x00\xE7\x92\x89\x00\xE7\x92\x98"
-     "\x00\xE7"
-     "\x93\x8A\x00\xE7\x93\x9C\x00\xE7\x93\xA6\x00\xE7\x94\x86\x00\xE7\x94\x98"
-     "\x00\xE7"
-     "\x94\x9F\x00\xE7\x94\xA4\x00\xE7\x94\xA8\x00\xE7\x94\xB0\x00\xE7\x94\xB2"
-     "\x00\xE7"
-     "\x94\xB3\x00\xE7\x94\xB7\x00\xE7\x94\xBB\x00\xE7\x94\xBE\x00\xE7\x95\x99"
-     "\x00\xE7"
-     "\x95\xA5\x00\xE7\x95\xB0\x00\xE7\x96\x8B\x00\xE7\x96\x92\x00\xE7\x97\xA2"
-     "\x00\xE7"
-     "\x98\x90\x00\xE7\x98\x9D\x00\xE7\x98\x9F\x00\xE7\x99\x82\x00\xE7\x99\xA9"
-     "\x00\xE7"
-     "\x99\xB6\x00\xE7\x99\xBD\x00\xE7\x9A\xAE\x00\xE7\x9A\xBF\x00\xE7\x9B\x8A"
-     "\x00\xE7"
-     "\x9B\x9B\x00\xE7\x9B\xA3\x00\xE7\x9B\xA7\x00\xE7\x9B\xAE\x00\xE7\x9B\xB4"
-     "\x00\xE7"
-     "\x9C\x81\x00\xE7\x9C\x9E\x00\xE7\x9C\x9F\x00\xE7\x9D\x80\x00\xE7\x9D\x8A"
-     "\x00\xE7"
-     "\x9E\x8B\x00\xE7\x9E\xA7\x00\xE7\x9F\x9B\x00\xE7\x9F\xA2\x00\xE7\x9F\xB3"
-     "\x00\xE7"
-     "\xA1\x8E\x00\xE7\xA1\xAB\x00\xE7\xA2\x8C\x00\xE7\xA2\x91\x00\xE7\xA3\x8A"
-     "\x00\xE7"
-     "\xA3\x8C\x00\xE7\xA3\xBB\x00\xE7\xA4\xAA\x00\xE7\xA4\xBA\x00\xE7\xA4\xBC"
-     "\x00\xE7"
-     "\xA4\xBE\x00\xE7\xA5\x88\x00\xE7\xA5\x89\x00\xE7\xA5\x90\x00\xE7\xA5\x96"
-     "\x00\xE7"
-     "\xA5\x9D\x00\xE7\xA5\x9E\x00\xE7\xA5\xA5\x00\xE7\xA5\xBF\x00\xE7\xA6\x81"
-     "\x00\xE7"
-     "\xA6\x8D\x00\xE7\xA6\x8E\x00\xE7\xA6\x8F\x00\xE7\xA6\xAE\x00\xE7\xA6\xB8"
-     "\x00\xE7"
-     "\xA6\xBE\x00\xE7\xA7\x8A\x00\xE7\xA7\x98\x00\xE7\xA7\xAB\x00\xE7\xA8\x9C"
-     "\x00\xE7"
-     "\xA9\x80\x00\xE7\xA9\x8A\x00\xE7\xA9\x8F\x00\xE7\xA9\xB4\x00\xE7\xA9\xBA"
-     "\x00\xE7"
-     "\xAA\x81\x00\xE7\xAA\xB1\x00\xE7\xAB\x8B\x00\xE7\xAB\xAE\x00\xE7\xAB\xB9"
-     "\x00\xE7"
-     "\xAC\xA0\x00\xE7\xAE\x8F\x00\xE7\xAF\x80\x00\xE7\xAF\x86\x00\xE7\xAF\x89"
-     "\x00\xE7"
-     "\xB0\xBE\x00\xE7\xB1\xA0\x00\xE7\xB1\xB3\x00\xE7\xB1\xBB\x00\xE7\xB2\x92"
-     "\x00\xE7"
-     "\xB2\xBE\x00\xE7\xB3\x92\x00\xE7\xB3\x96\x00\xE7\xB3\xA3\x00\xE7\xB3\xA7"
-     "\x00\xE7"
-     "\xB3\xA8\x00\xE7\xB3\xB8\x00\xE7\xB4\x80\x00\xE7\xB4\x90\x00\xE7\xB4\xA2"
-     "\x00\xE7"
-     "\xB4\xAF\x00\xE7\xB5\x82\x00\xE7\xB5\x9B\x00\xE7\xB5\xA3\x00\xE7\xB6\xA0"
-     "\x00\xE7"
-     "\xB6\xBE\x00\xE7\xB7\x87\x00\xE7\xB7\xB4\x00\xE7\xB8\x82\x00\xE7\xB8\x89"
-     "\x00\xE7"
-     "\xB8\xB7\x00\xE7\xB9\x81\x00\xE7\xB9\x85\x00\xE7\xBC\xB6\x00\xE7\xBC\xBE"
-     "\x00\xE7"
-     "\xBD\x91\x00\xE7\xBD\xB2\x00\xE7\xBD\xB9\x00\xE7\xBD\xBA\x00\xE7\xBE\x85"
-     "\x00\xE7"
-     "\xBE\x8A\x00\xE7\xBE\x95\x00\xE7\xBE\x9A\x00\xE7\xBE\xBD\x00\xE7\xBF\xBA"
-     "\x00\xE8"
-     "\x80\x81\x00\xE8\x80\x85\x00\xE8\x80\x8C\x00\xE8\x80\x92\x00\xE8\x80\xB3"
-     "\x00\xE8"
-     "\x81\x86\x00\xE8\x81\xA0\x00\xE8\x81\xAF\x00\xE8\x81\xB0\x00\xE8\x81\xBE"
-     "\x00\xE8"
-     "\x81\xBF\x00\xE8\x82\x89\x00\xE8\x82\x8B\x00\xE8\x82\xAD\x00\xE8\x82\xB2"
-     "\x00\xE8"
-     "\x84\x83\x00\xE8\x84\xBE\x00\xE8\x87\x98\x00\xE8\x87\xA3\x00\xE8\x87\xA8"
-     "\x00\xE8"
-     "\x87\xAA\x00\xE8\x87\xAD\x00\xE8\x87\xB3\x00\xE8\x87\xBC\x00\xE8\x88\x81"
-     "\x00\xE8"
-     "\x88\x84\x00\xE8\x88\x8C\x00\xE8\x88\x98\x00\xE8\x88\x9B\x00\xE8\x88\x9F"
-     "\x00\xE8"
-     "\x89\xAE\x00\xE8\x89\xAF\x00\xE8\x89\xB2\x00\xE8\x89\xB8\x00\xE8\x89\xB9"
-     "\x00\xE8"
-     "\x8A\x8B\x00\xE8\x8A\x91\x00\xE8\x8A\x9D\x00\xE8\x8A\xB1\x00\xE8\x8A\xB3"
-     "\x00\xE8"
-     "\x8A\xBD\x00\xE8\x8B\xA5\x00\xE8\x8B\xA6\x00\xE8\x8C\x9D\x00\xE8\x8C\xA3"
-     "\x00\xE8"
-     "\x8C\xB6\x00\xE8\x8D\x92\x00\xE8\x8D\x93\x00\xE8\x8D\xA3\x00\xE8\x8E\xAD"
-     "\x00\xE8"
-     "\x8E\xBD\x00\xE8\x8F\x89\x00\xE8\x8F\x8A\x00\xE8\x8F\x8C\x00\xE8\x8F\x9C"
-     "\x00\xE8"
-     "\x8F\xA7\x00\xE8\x8F\xAF\x00\xE8\x8F\xB1\x00\xE8\x90\xBD\x00\xE8\x91\x89"
-     "\x00\xE8"
-     "\x91\x97\x00\xE8\x93\xAE\x00\xE8\x93\xB1\x00\xE8\x93\xB3\x00\xE8\x93\xBC"
-     "\x00\xE8"
-     "\x94\x96\x00\xE8\x95\xA4\x00\xE8\x97\x8D\x00\xE8\x97\xBA\x00\xE8\x98\x86"
-     "\x00\xE8"
-     "\x98\x92\x00\xE8\x98\xAD\x00\xE8\x98\xBF\x00\xE8\x99\x8D\x00\xE8\x99\x90"
-     "\x00\xE8"
-     "\x99\x9C\x00\xE8\x99\xA7\x00\xE8\x99\xA9\x00\xE8\x99\xAB\x00\xE8\x9A\x88"
-     "\x00\xE8"
-     "\x9A\xA9\x00\xE8\x9B\xA2\x00\xE8\x9C\x8E\x00\xE8\x9C\xA8\x00\xE8\x9D\xAB"
-     "\x00\xE8"
-     "\x9D\xB9\x00\xE8\x9E\x86\x00\xE8\x9E\xBA\x00\xE8\x9F\xA1\x00\xE8\xA0\x81"
-     "\x00\xE8"
-     "\xA0\x9F\x00\xE8\xA1\x80\x00\xE8\xA1\x8C\x00\xE8\xA1\xA0\x00\xE8\xA1\xA3"
-     "\x00\xE8"
-     "\xA3\x82\x00\xE8\xA3\x8F\x00\xE8\xA3\x97\x00\xE8\xA3\x9E\x00\xE8\xA3\xA1"
-     "\x00\xE8"
-     "\xA3\xB8\x00\xE8\xA3\xBA\x00\xE8\xA4\x90\x00\xE8\xA5\x81\x00\xE8\xA5\xA4"
-     "\x00\xE8"
-     "\xA5\xBE\x00\xE8\xA6\x86\x00\xE8\xA6\x8B\x00\xE8\xA6\x96\x00\xE8\xA7\x92"
-     "\x00\xE8"
-     "\xA7\xA3\x00\xE8\xA8\x80\x00\xE8\xAA\xA0\x00\xE8\xAA\xAA\x00\xE8\xAA\xBF"
-     "\x00\xE8"
-     "\xAB\x8B\x00\xE8\xAB\x92\x00\xE8\xAB\x96\x00\xE8\xAB\xAD\x00\xE8\xAB\xB8"
-     "\x00\xE8"
-     "\xAB\xBE\x00\xE8\xAC\x81\x00\xE8\xAC\xB9\x00\xE8\xAD\x98\x00\xE8\xAE\x80"
-     "\x00\xE8"
-     "\xAE\x8A\x00\xE8\xB0\xB7\x00\xE8\xB1\x86\x00\xE8\xB1\x88\x00\xE8\xB1\x95"
-     "\x00\xE8"
-     "\xB1\xB8\x00\xE8\xB2\x9D\x00\xE8\xB2\xA1\x00\xE8\xB2\xA9\x00\xE8\xB2\xAB"
-     "\x00\xE8"
-     "\xB3\x81\x00\xE8\xB3\x82\x00\xE8\xB3\x87\x00\xE8\xB3\x88\x00\xE8\xB3\x93"
-     "\x00\xE8"
-     "\xB4\x88\x00\xE8\xB4\x9B\x00\xE8\xB5\xA4\x00\xE8\xB5\xB0\x00\xE8\xB5\xB7"
-     "\x00\xE8"
-     "\xB6\xB3\x00\xE8\xB6\xBC\x00\xE8\xB7\x8B\x00\xE8\xB7\xAF\x00\xE8\xB7\xB0"
-     "\x00\xE8"
-     "\xBA\xAB\x00\xE8\xBB\x8A\x00\xE8\xBB\x94\x00\xE8\xBC\xA6\x00\xE8\xBC\xAA"
-     "\x00\xE8"
-     "\xBC\xB8\x00\xE8\xBC\xBB\x00\xE8\xBD\xA2\x00\xE8\xBE\x9B\x00\xE8\xBE\x9E"
-     "\x00\xE8"
-     "\xBE\xB0\x00\xE8\xBE\xB5\x00\xE8\xBE\xB6\x00\xE9\x80\xA3\x00\xE9\x80\xB8"
-     "\x00\xE9"
-     "\x81\x8A\x00\xE9\x81\xA9\x00\xE9\x81\xB2\x00\xE9\x81\xBC\x00\xE9\x82\x8F"
-     "\x00\xE9"
-     "\x82\x91\x00\xE9\x82\x94\x00\xE9\x83\x8E\x00\xE9\x83\x9E\x00\xE9\x83\xB1"
-     "\x00\xE9"
-     "\x83\xBD\x00\xE9\x84\x91\x00\xE9\x84\x9B\x00\xE9\x85\x89\x00\xE9\x85\x8D"
-     "\x00\xE9"
-     "\x85\xAA\x00\xE9\x86\x99\x00\xE9\x86\xB4\x00\xE9\x87\x86\x00\xE9\x87\x8C"
-     "\x00\xE9"
-     "\x87\x8F\x00\xE9\x87\x91\x00\xE9\x88\xB4\x00\xE9\x88\xB8\x00\xE9\x89\xB6"
-     "\x00\xE9"
-     "\x89\xBC\x00\xE9\x8B\x97\x00\xE9\x8B\x98\x00\xE9\x8C\x84\x00\xE9\x8D\x8A"
-     "\x00\xE9"
-     "\x8F\xB9\x00\xE9\x90\x95\x00\xE9\x95\xB7\x00\xE9\x96\x80\x00\xE9\x96\x8B"
-     "\x00\xE9"
-     "\x96\xAD\x00\xE9\x96\xB7\x00\xE9\x98\x9C\x00\xE9\x98\xAE\x00\xE9\x99\x8B"
-     "\x00\xE9"
-     "\x99\x8D\x00\xE9\x99\xB5\x00\xE9\x99\xB8\x00\xE9\x99\xBC\x00\xE9\x9A\x86"
-     "\x00\xE9"
-     "\x9A\xA3\x00\xE9\x9A\xB6\x00\xE9\x9A\xB7\x00\xE9\x9A\xB8\x00\xE9\x9A\xB9"
-     "\x00\xE9"
-     "\x9B\x83\x00\xE9\x9B\xA2\x00\xE9\x9B\xA3\x00\xE9\x9B\xA8\x00\xE9\x9B\xB6"
-     "\x00\xE9"
-     "\x9B\xB7\x00\xE9\x9C\xA3\x00\xE9\x9C\xB2\x00\xE9\x9D\x88\x00\xE9\x9D\x91"
-     "\x00\xE9"
-     "\x9D\x96\x00\xE9\x9D\x9E\x00\xE9\x9D\xA2\x00\xE9\x9D\xA9\x00\xE9\x9F\x8B"
-     "\x00\xE9"
-     "\x9F\x9B\x00\xE9\x9F\xA0\x00\xE9\x9F\xAD\x00\xE9\x9F\xB3\x00\xE9\x9F\xBF"
-     "\x00\xE9"
-     "\xA0\x81\x00\xE9\xA0\x85\x00\xE9\xA0\x8B\x00\xE9\xA0\x98\x00\xE9\xA0\xA9"
-     "\x00\xE9"
-     "\xA0\xBB\x00\xE9\xA1\x9E\x00\xE9\xA2\xA8\x00\xE9\xA3\x9B\x00\xE9\xA3\x9F"
-     "\x00\xE9"
-     "\xA3\xA2\x00\xE9\xA3\xAF\x00\xE9\xA3\xBC\x00\xE9\xA4\xA8\x00\xE9\xA4\xA9"
-     "\x00\xE9"
-     "\xA6\x96\x00\xE9\xA6\x99\x00\xE9\xA6\xA7\x00\xE9\xA6\xAC\x00\xE9\xA7\x82"
-     "\x00\xE9"
-     "\xA7\xB1\x00\xE9\xA7\xBE\x00\xE9\xA9\xAA\x00\xE9\xAA\xA8\x00\xE9\xAB\x98"
-     "\x00\xE9"
-     "\xAB\x9F\x00\xE9\xAC\x92\x00\xE9\xAC\xA5\x00\xE9\xAC\xAF\x00\xE9\xAC\xB2"
-     "\x00\xE9"
-     "\xAC\xBC\x00\xE9\xAD\x9A\x00\xE9\xAD\xAF\x00\xE9\xB1\x80\x00\xE9\xB1\x97"
-     "\x00\xE9"
-     "\xB3\xA5\x00\xE9\xB3\xBD\x00\xE9\xB5\xA7\x00\xE9\xB6\xB4\x00\xE9\xB7\xBA"
-     "\x00\xE9"
-     "\xB8\x9E\x00\xE9\xB9\xB5\x00\xE9\xB9\xBF\x00\xE9\xBA\x97\x00\xE9\xBA\x9F"
-     "\x00\xE9"
-     "\xBA\xA5\x00\xE9\xBA\xBB\x00\xE9\xBB\x83\x00\xE9\xBB\x8D\x00\xE9\xBB\x8E"
-     "\x00\xE9"
-     "\xBB\x91\x00\xE9\xBB\xB9\x00\xE9\xBB\xBD\x00\xE9\xBB\xBE\x00\xE9\xBC\x85"
-     "\x00\xE9"
-     "\xBC\x8E\x00\xE9\xBC\x8F\x00\xE9\xBC\x93\x00\xE9\xBC\x96\x00\xE9\xBC\xA0"
-     "\x00\xE9"
-     "\xBC\xBB\x00\xE9\xBD\x83\x00\xE9\xBD\x8A\x00\xE9\xBD\x92\x00\xE9\xBE\x8D"
-     "\x00\xE9"
-     "\xBE\x8E\x00\xE9\xBE\x9C\x00\xE9\xBE\x9F\x00\xE9\xBE\xA0\x00\xEA\x99\x81"
-     "\x00\xEA"
-     "\x99\x83\x00\xEA\x99\x85\x00\xEA\x99\x87\x00\xEA\x99\x89\x00\xEA\x99\x8B"
-     "\x00\xEA"
-     "\x99\x8D\x00\xEA\x99\x8F\x00\xEA\x99\x91\x00\xEA\x99\x93\x00\xEA\x99\x95"
-     "\x00\xEA"
-     "\x99\x97\x00\xEA\x99\x99\x00\xEA\x99\x9B\x00\xEA\x99\x9D\x00\xEA\x99\x9F"
-     "\x00\xEA"
-     "\x99\xA1\x00\xEA\x99\xA3\x00\xEA\x99\xA5\x00\xEA\x99\xA7\x00\xEA\x99\xA9"
-     "\x00\xEA"
-     "\x99\xAB\x00\xEA\x99\xAD\x00\xEA\x9A\x81\x00\xEA\x9A\x83\x00\xEA\x9A\x85"
-     "\x00\xEA"
-     "\x9A\x87\x00\xEA\x9A\x89\x00\xEA\x9A\x8B\x00\xEA\x9A\x8D\x00\xEA\x9A\x8F"
-     "\x00\xEA"
-     "\x9A\x91\x00\xEA\x9A\x93\x00\xEA\x9A\x95\x00\xEA\x9A\x97\x00\xEA\x9A\x99"
-     "\x00\xEA"
-     "\x9A\x9B\x00\xEA\x9C\xA3\x00\xEA\x9C\xA5\x00\xEA\x9C\xA7\x00\xEA\x9C\xA9"
-     "\x00\xEA"
-     "\x9C\xAB\x00\xEA\x9C\xAD\x00\xEA\x9C\xAF\x00\xEA\x9C\xB3\x00\xEA\x9C\xB5"
-     "\x00\xEA"
-     "\x9C\xB7\x00\xEA\x9C\xB9\x00\xEA\x9C\xBB\x00\xEA\x9C\xBD\x00\xEA\x9C\xBF"
-     "\x00\xEA"
-     "\x9D\x81\x00\xEA\x9D\x83\x00\xEA\x9D\x85\x00\xEA\x9D\x87\x00\xEA\x9D\x89"
-     "\x00\xEA"
-     "\x9D\x8B\x00\xEA\x9D\x8D\x00\xEA\x9D\x8F\x00\xEA\x9D\x91\x00\xEA\x9D\x93"
-     "\x00\xEA"
-     "\x9D\x95\x00\xEA\x9D\x97\x00\xEA\x9D\x99\x00\xEA\x9D\x9B\x00\xEA\x9D\x9D"
-     "\x00\xEA"
-     "\x9D\x9F\x00\xEA\x9D\xA1\x00\xEA\x9D\xA3\x00\xEA\x9D\xA5\x00\xEA\x9D\xA7"
-     "\x00\xEA"
-     "\x9D\xA9\x00\xEA\x9D\xAB\x00\xEA\x9D\xAD\x00\xEA\x9D\xAF\x00\xEA\x9D\xBA"
-     "\x00\xEA"
-     "\x9D\xBC\x00\xEA\x9D\xBF\x00\xEA\x9E\x81\x00\xEA\x9E\x83\x00\xEA\x9E\x85"
-     "\x00\xEA"
-     "\x9E\x87\x00\xEA\x9E\x8C\x00\xEA\x9E\x8E\x00\xEA\x9E\x91\x00\xEA\x9E\x93"
-     "\x00\xEA"
-     "\x9E\x94\x00\xEA\x9E\x97\x00\xEA\x9E\x99\x00\xEA\x9E\x9B\x00\xEA\x9E\x9D"
-     "\x00\xEA"
-     "\x9E\x9F\x00\xEA\x9E\xA1\x00\xEA\x9E\xA3\x00\xEA\x9E\xA5\x00\xEA\x9E\xA7"
-     "\x00\xEA"
-     "\x9E\xA9\x00\xEA\x9E\xB5\x00\xEA\x9E\xB7\x00\xEA\x9E\xB9\x00\xEA\x9E\xBB"
-     "\x00\xEA"
-     "\x9E\xBD\x00\xEA\x9E\xBF\x00\xEA\x9F\x81\x00\xEA\x9F\x83\x00\xEA\x9F\x88"
-     "\x00\xEA"
-     "\x9F\x8A\x00\xEA\x9F\x91\x00\xEA\x9F\x97\x00\xEA\x9F\x99\x00\xEA\x9F\xB6"
-     "\x00\xEA"
-     "\xAC\xB7\x00\xEA\xAD\x92\x00\xEA\xAD\x93\x00\xEA\xAD\xA6\x00\xEA\xAD\xA7"
-     "\x00\xEA"
-     "\xB0\x80\x00\xEA\xB0\x81\x00\xEA\xB0\x82\x00\xEA\xB0\x83\x00\xEA\xB0\x84"
-     "\x00\xEA"
-     "\xB0\x85\x00\xEA\xB0\x86\x00\xEA\xB0\x87\x00\xEA\xB0\x88\x00\xEA\xB0\x89"
-     "\x00\xEA"
-     "\xB0\x8A\x00\xEA\xB0\x8B\x00\xEA\xB0\x8C\x00\xEA\xB0\x8D\x00\xEA\xB0\x8E"
-     "\x00\xEA"
-     "\xB0\x8F\x00\xEA\xB0\x90\x00\xEA\xB0\x91\x00\xEA\xB0\x92\x00\xEA\xB0\x93"
-     "\x00\xEA"
-     "\xB0\x94\x00\xEA\xB0\x95\x00\xEA\xB0\x96\x00\xEA\xB0\x97\x00\xEA\xB0\x98"
-     "\x00\xEA"
-     "\xB0\x99\x00\xEA\xB0\x9A\x00\xEA\xB0\x9B\x00\xEA\xB0\x9C\x00\xEA\xB0\x9D"
-     "\x00\xEA"
-     "\xB0\x9E\x00\xEA\xB0\x9F\x00\xEA\xB0\xA0\x00\xEA\xB0\xA1\x00\xEA\xB0\xA2"
-     "\x00\xEA"
-     "\xB0\xA3\x00\xEA\xB0\xA4\x00\xEA\xB0\xA5\x00\xEA\xB0\xA6\x00\xEA\xB0\xA7"
-     "\x00\xEA"
-     "\xB0\xA8\x00\xEA\xB0\xA9\x00\xEA\xB0\xAA\x00\xEA\xB0\xAB\x00\xEA\xB0\xAC"
-     "\x00\xEA"
-     "\xB0\xAD\x00\xEA\xB0\xAE\x00\xEA\xB0\xAF\x00\xEA\xB0\xB0\x00\xEA\xB0\xB1"
-     "\x00\xEA"
-     "\xB0\xB2\x00\xEA\xB0\xB3\x00\xEA\xB0\xB4\x00\xEA\xB0\xB5\x00\xEA\xB0\xB6"
-     "\x00\xEA"
-     "\xB0\xB7\x00\xEA\xB0\xB8\x00\xEA\xB0\xB9\x00\xEA\xB0\xBA\x00\xEA\xB0\xBB"
-     "\x00\xEA"
-     "\xB0\xBC\x00\xEA\xB0\xBD\x00\xEA\xB0\xBE\x00\xEA\xB0\xBF\x00\xEA\xB1\x80"
-     "\x00\xEA"
-     "\xB1\x81\x00\xEA\xB1\x82\x00\xEA\xB1\x83\x00\xEA\xB1\x84\x00\xEA\xB1\x85"
-     "\x00\xEA"
-     "\xB1\x86\x00\xEA\xB1\x87\x00\xEA\xB1\x88\x00\xEA\xB1\x89\x00\xEA\xB1\x8A"
-     "\x00\xEA"
-     "\xB1\x8B\x00\xEA\xB1\x8C\x00\xEA\xB1\x8D\x00\xEA\xB1\x8E\x00\xEA\xB1\x8F"
-     "\x00\xEA"
-     "\xB1\x90\x00\xEA\xB1\x91\x00\xEA\xB1\x92\x00\xEA\xB1\x93\x00\xEA\xB1\x94"
-     "\x00\xEA"
-     "\xB1\x95\x00\xEA\xB1\x96\x00\xEA\xB1\x97\x00\xEA\xB1\x98\x00\xEA\xB1\x99"
-     "\x00\xEA"
-     "\xB1\x9A\x00\xEA\xB1\x9B\x00\xEA\xB1\x9C\x00\xEA\xB1\x9D\x00\xEA\xB1\x9E"
-     "\x00\xEA"
-     "\xB1\x9F\x00\xEA\xB1\xA0\x00\xEA\xB1\xA1\x00\xEA\xB1\xA2\x00\xEA\xB1\xA3"
-     "\x00\xEA"
-     "\xB1\xA4\x00\xEA\xB1\xA5\x00\xEA\xB1\xA6\x00\xEA\xB1\xA7\x00\xEA\xB1\xA8"
-     "\x00\xEA"
-     "\xB1\xA9\x00\xEA\xB1\xAA\x00\xEA\xB1\xAB\x00\xEA\xB1\xAC\x00\xEA\xB1\xAD"
-     "\x00\xEA"
-     "\xB1\xAE\x00\xEA\xB1\xAF\x00\xEA\xB1\xB0\x00\xEA\xB1\xB1\x00\xEA\xB1\xB2"
-     "\x00\xEA"
-     "\xB1\xB3\x00\xEA\xB1\xB4\x00\xEA\xB1\xB5\x00\xEA\xB1\xB6\x00\xEA\xB1\xB7"
-     "\x00\xEA"
-     "\xB1\xB8\x00\xEA\xB1\xB9\x00\xEA\xB1\xBA\x00\xEA\xB1\xBB\x00\xEA\xB1\xBC"
-     "\x00\xEA"
-     "\xB1\xBD\x00\xEA\xB1\xBE\x00\xEA\xB1\xBF\x00\xEA\xB2\x80\x00\xEA\xB2\x81"
-     "\x00\xEA"
-     "\xB2\x82\x00\xEA\xB2\x83\x00\xEA\xB2\x84\x00\xEA\xB2\x85\x00\xEA\xB2\x86"
-     "\x00\xEA"
-     "\xB2\x87\x00\xEA\xB2\x88\x00\xEA\xB2\x89\x00\xEA\xB2\x8A\x00\xEA\xB2\x8B"
-     "\x00\xEA"
-     "\xB2\x8C\x00\xEA\xB2\x8D\x00\xEA\xB2\x8E\x00\xEA\xB2\x8F\x00\xEA\xB2\x90"
-     "\x00\xEA"
-     "\xB2\x91\x00\xEA\xB2\x92\x00\xEA\xB2\x93\x00\xEA\xB2\x94\x00\xEA\xB2\x95"
-     "\x00\xEA"
-     "\xB2\x96\x00\xEA\xB2\x97\x00\xEA\xB2\x98\x00\xEA\xB2\x99\x00\xEA\xB2\x9A"
-     "\x00\xEA"
-     "\xB2\x9B\x00\xEA\xB2\x9C\x00\xEA\xB2\x9D\x00\xEA\xB2\x9E\x00\xEA\xB2\x9F"
-     "\x00\xEA"
-     "\xB2\xA0\x00\xEA\xB2\xA1\x00\xEA\xB2\xA2\x00\xEA\xB2\xA3\x00\xEA\xB2\xA4"
-     "\x00\xEA"
-     "\xB2\xA5\x00\xEA\xB2\xA6\x00\xEA\xB2\xA7\x00\xEA\xB2\xA8\x00\xEA\xB2\xA9"
-     "\x00\xEA"
-     "\xB2\xAA\x00\xEA\xB2\xAB\x00\xEA\xB2\xAC\x00\xEA\xB2\xAD\x00\xEA\xB2\xAE"
-     "\x00\xEA"
-     "\xB2\xAF\x00\xEA\xB2\xB0\x00\xEA\xB2\xB1\x00\xEA\xB2\xB2\x00\xEA\xB2\xB3"
-     "\x00\xEA"
-     "\xB2\xB4\x00\xEA\xB2\xB5\x00\xEA\xB2\xB6\x00\xEA\xB2\xB7\x00\xEA\xB2\xB8"
-     "\x00\xEA"
-     "\xB2\xB9\x00\xEA\xB2\xBA\x00\xEA\xB2\xBB\x00\xEA\xB2\xBC\x00\xEA\xB2\xBD"
-     "\x00\xEA"
-     "\xB2\xBE\x00\xEA\xB2\xBF\x00\xEA\xB3\x80\x00\xEA\xB3\x81\x00\xEA\xB3\x82"
-     "\x00\xEA"
-     "\xB3\x83\x00\xEA\xB3\x84\x00\xEA\xB3\x85\x00\xEA\xB3\x86\x00\xEA\xB3\x87"
-     "\x00\xEA"
-     "\xB3\x88\x00\xEA\xB3\x89\x00\xEA\xB3\x8A\x00\xEA\xB3\x8B\x00\xEA\xB3\x8C"
-     "\x00\xEA"
-     "\xB3\x8D\x00\xEA\xB3\x8E\x00\xEA\xB3\x8F\x00\xEA\xB3\x90\x00\xEA\xB3\x91"
-     "\x00\xEA"
-     "\xB3\x92\x00\xEA\xB3\x93\x00\xEA\xB3\x94\x00\xEA\xB3\x95\x00\xEA\xB3\x96"
-     "\x00\xEA"
-     "\xB3\x97\x00\xEA\xB3\x98\x00\xEA\xB3\x99\x00\xEA\xB3\x9A\x00\xEA\xB3\x9B"
-     "\x00\xEA"
-     "\xB3\x9C\x00\xEA\xB3\x9D\x00\xEA\xB3\x9E\x00\xEA\xB3\x9F\x00\xEA\xB3\xA0"
-     "\x00\xEA"
-     "\xB3\xA1\x00\xEA\xB3\xA2\x00\xEA\xB3\xA3\x00\xEA\xB3\xA4\x00\xEA\xB3\xA5"
-     "\x00\xEA"
-     "\xB3\xA6\x00\xEA\xB3\xA7\x00\xEA\xB3\xA8\x00\xEA\xB3\xA9\x00\xEA\xB3\xAA"
-     "\x00\xEA"
-     "\xB3\xAB\x00\xEA\xB3\xAC\x00\xEA\xB3\xAD\x00\xEA\xB3\xAE\x00\xEA\xB3\xAF"
-     "\x00\xEA"
-     "\xB3\xB0\x00\xEA\xB3\xB1\x00\xEA\xB3\xB2\x00\xEA\xB3\xB3\x00\xEA\xB3\xB4"
-     "\x00\xEA"
-     "\xB3\xB5\x00\xEA\xB3\xB6\x00\xEA\xB3\xB7\x00\xEA\xB3\xB8\x00\xEA\xB3\xB9"
-     "\x00\xEA"
-     "\xB3\xBA\x00\xEA\xB3\xBB\x00\xEA\xB3\xBC\x00\xEA\xB3\xBD\x00\xEA\xB3\xBE"
-     "\x00\xEA"
-     "\xB3\xBF\x00\xEA\xB4\x80\x00\xEA\xB4\x81\x00\xEA\xB4\x82\x00\xEA\xB4\x83"
-     "\x00\xEA"
-     "\xB4\x84\x00\xEA\xB4\x85\x00\xEA\xB4\x86\x00\xEA\xB4\x87\x00\xEA\xB4\x88"
-     "\x00\xEA"
-     "\xB4\x89\x00\xEA\xB4\x8A\x00\xEA\xB4\x8B\x00\xEA\xB4\x8C\x00\xEA\xB4\x8D"
-     "\x00\xEA"
-     "\xB4\x8E\x00\xEA\xB4\x8F\x00\xEA\xB4\x90\x00\xEA\xB4\x91\x00\xEA\xB4\x92"
-     "\x00\xEA"
-     "\xB4\x93\x00\xEA\xB4\x94\x00\xEA\xB4\x95\x00\xEA\xB4\x96\x00\xEA\xB4\x97"
-     "\x00\xEA"
-     "\xB4\x98\x00\xEA\xB4\x99\x00\xEA\xB4\x9A\x00\xEA\xB4\x9B\x00\xEA\xB4\x9C"
-     "\x00\xEA"
-     "\xB4\x9D\x00\xEA\xB4\x9E\x00\xEA\xB4\x9F\x00\xEA\xB4\xA0\x00\xEA\xB4\xA1"
-     "\x00\xEA"
-     "\xB4\xA2\x00\xEA\xB4\xA3\x00\xEA\xB4\xA4\x00\xEA\xB4\xA5\x00\xEA\xB4\xA6"
-     "\x00\xEA"
-     "\xB4\xA7\x00\xEA\xB4\xA8\x00\xEA\xB4\xA9\x00\xEA\xB4\xAA\x00\xEA\xB4\xAB"
-     "\x00\xEA"
-     "\xB4\xAC\x00\xEA\xB4\xAD\x00\xEA\xB4\xAE\x00\xEA\xB4\xAF\x00\xEA\xB4\xB0"
-     "\x00\xEA"
-     "\xB4\xB1\x00\xEA\xB4\xB2\x00\xEA\xB4\xB3\x00\xEA\xB4\xB4\x00\xEA\xB4\xB5"
-     "\x00\xEA"
-     "\xB4\xB6\x00\xEA\xB4\xB7\x00\xEA\xB4\xB8\x00\xEA\xB4\xB9\x00\xEA\xB4\xBA"
-     "\x00\xEA"
-     "\xB4\xBB\x00\xEA\xB4\xBC\x00\xEA\xB4\xBD\x00\xEA\xB4\xBE\x00\xEA\xB4\xBF"
-     "\x00\xEA"
-     "\xB5\x80\x00\xEA\xB5\x81\x00\xEA\xB5\x82\x00\xEA\xB5\x83\x00\xEA\xB5\x84"
-     "\x00\xEA"
-     "\xB5\x85\x00\xEA\xB5\x86\x00\xEA\xB5\x87\x00\xEA\xB5\x88\x00\xEA\xB5\x89"
-     "\x00\xEA"
-     "\xB5\x8A\x00\xEA\xB5\x8B\x00\xEA\xB5\x8C\x00\xEA\xB5\x8D\x00\xEA\xB5\x8E"
-     "\x00\xEA"
-     "\xB5\x8F\x00\xEA\xB5\x90\x00\xEA\xB5\x91\x00\xEA\xB5\x92\x00\xEA\xB5\x93"
-     "\x00\xEA"
-     "\xB5\x94\x00\xEA\xB5\x95\x00\xEA\xB5\x96\x00\xEA\xB5\x97\x00\xEA\xB5\x98"
-     "\x00\xEA"
-     "\xB5\x99\x00\xEA\xB5\x9A\x00\xEA\xB5\x9B\x00\xEA\xB5\x9C\x00\xEA\xB5\x9D"
-     "\x00\xEA"
-     "\xB5\x9E\x00\xEA\xB5\x9F\x00\xEA\xB5\xA0\x00\xEA\xB5\xA1\x00\xEA\xB5\xA2"
-     "\x00\xEA"
-     "\xB5\xA3\x00\xEA\xB5\xA4\x00\xEA\xB5\xA5\x00\xEA\xB5\xA6\x00\xEA\xB5\xA7"
-     "\x00\xEA"
-     "\xB5\xA8\x00\xEA\xB5\xA9\x00\xEA\xB5\xAA\x00\xEA\xB5\xAB\x00\xEA\xB5\xAC"
-     "\x00\xEA"
-     "\xB5\xAD\x00\xEA\xB5\xAE\x00\xEA\xB5\xAF\x00\xEA\xB5\xB0\x00\xEA\xB5\xB1"
-     "\x00\xEA"
-     "\xB5\xB2\x00\xEA\xB5\xB3\x00\xEA\xB5\xB4\x00\xEA\xB5\xB5\x00\xEA\xB5\xB6"
-     "\x00\xEA"
-     "\xB5\xB7\x00\xEA\xB5\xB8\x00\xEA\xB5\xB9\x00\xEA\xB5\xBA\x00\xEA\xB5\xBB"
-     "\x00\xEA"
-     "\xB5\xBC\x00\xEA\xB5\xBD\x00\xEA\xB5\xBE\x00\xEA\xB5\xBF\x00\xEA\xB6\x80"
-     "\x00\xEA"
-     "\xB6\x81\x00\xEA\xB6\x82\x00\xEA\xB6\x83\x00\xEA\xB6\x84\x00\xEA\xB6\x85"
-     "\x00\xEA"
-     "\xB6\x86\x00\xEA\xB6\x87\x00\xEA\xB6\x88\x00\xEA\xB6\x89\x00\xEA\xB6\x8A"
-     "\x00\xEA"
-     "\xB6\x8B\x00\xEA\xB6\x8C\x00\xEA\xB6\x8D\x00\xEA\xB6\x8E\x00\xEA\xB6\x8F"
-     "\x00\xEA"
-     "\xB6\x90\x00\xEA\xB6\x91\x00\xEA\xB6\x92\x00\xEA\xB6\x93\x00\xEA\xB6\x94"
-     "\x00\xEA"
-     "\xB6\x95\x00\xEA\xB6\x96\x00\xEA\xB6\x97\x00\xEA\xB6\x98\x00\xEA\xB6\x99"
-     "\x00\xEA"
-     "\xB6\x9A\x00\xEA\xB6\x9B\x00\xEA\xB6\x9C\x00\xEA\xB6\x9D\x00\xEA\xB6\x9E"
-     "\x00\xEA"
-     "\xB6\x9F\x00\xEA\xB6\xA0\x00\xEA\xB6\xA1\x00\xEA\xB6\xA2\x00\xEA\xB6\xA3"
-     "\x00\xEA"
-     "\xB6\xA4\x00\xEA\xB6\xA5\x00\xEA\xB6\xA6\x00\xEA\xB6\xA7\x00\xEA\xB6\xA8"
-     "\x00\xEA"
-     "\xB6\xA9\x00\xEA\xB6\xAA\x00\xEA\xB6\xAB\x00\xEA\xB6\xAC\x00\xEA\xB6\xAD"
-     "\x00\xEA"
-     "\xB6\xAE\x00\xEA\xB6\xAF\x00\xEA\xB6\xB0\x00\xEA\xB6\xB1\x00\xEA\xB6\xB2"
-     "\x00\xEA"
-     "\xB6\xB3\x00\xEA\xB6\xB4\x00\xEA\xB6\xB5\x00\xEA\xB6\xB6\x00\xEA\xB6\xB7"
-     "\x00\xEA"
-     "\xB6\xB8\x00\xEA\xB6\xB9\x00\xEA\xB6\xBA\x00\xEA\xB6\xBB\x00\xEA\xB6\xBC"
-     "\x00\xEA"
-     "\xB6\xBD\x00\xEA\xB6\xBE\x00\xEA\xB6\xBF\x00\xEA\xB7\x80\x00\xEA\xB7\x81"
-     "\x00\xEA"
-     "\xB7\x82\x00\xEA\xB7\x83\x00\xEA\xB7\x84\x00\xEA\xB7\x85\x00\xEA\xB7\x86"
-     "\x00\xEA"
-     "\xB7\x87\x00\xEA\xB7\x88\x00\xEA\xB7\x89\x00\xEA\xB7\x8A\x00\xEA\xB7\x8B"
-     "\x00\xEA"
-     "\xB7\x8C\x00\xEA\xB7\x8D\x00\xEA\xB7\x8E\x00\xEA\xB7\x8F\x00\xEA\xB7\x90"
-     "\x00\xEA"
-     "\xB7\x91\x00\xEA\xB7\x92\x00\xEA\xB7\x93\x00\xEA\xB7\x94\x00\xEA\xB7\x95"
-     "\x00\xEA"
-     "\xB7\x96\x00\xEA\xB7\x97\x00\xEA\xB7\x98\x00\xEA\xB7\x99\x00\xEA\xB7\x9A"
-     "\x00\xEA"
-     "\xB7\x9B\x00\xEA\xB7\x9C\x00\xEA\xB7\x9D\x00\xEA\xB7\x9E\x00\xEA\xB7\x9F"
-     "\x00\xEA"
-     "\xB7\xA0\x00\xEA\xB7\xA1\x00\xEA\xB7\xA2\x00\xEA\xB7\xA3\x00\xEA\xB7\xA4"
-     "\x00\xEA"
-     "\xB7\xA5\x00\xEA\xB7\xA6\x00\xEA\xB7\xA7\x00\xEA\xB7\xA8\x00\xEA\xB7\xA9"
-     "\x00\xEA"
-     "\xB7\xAA\x00\xEA\xB7\xAB\x00\xEA\xB7\xAC\x00\xEA\xB7\xAD\x00\xEA\xB7\xAE"
-     "\x00\xEA"
-     "\xB7\xAF\x00\xEA\xB7\xB0\x00\xEA\xB7\xB1\x00\xEA\xB7\xB2\x00\xEA\xB7\xB3"
-     "\x00\xEA"
-     "\xB7\xB4\x00\xEA\xB7\xB5\x00\xEA\xB7\xB6\x00\xEA\xB7\xB7\x00\xEA\xB7\xB8"
-     "\x00\xEA"
-     "\xB7\xB9\x00\xEA\xB7\xBA\x00\xEA\xB7\xBB\x00\xEA\xB7\xBC\x00\xEA\xB7\xBD"
-     "\x00\xEA"
-     "\xB7\xBE\x00\xEA\xB7\xBF\x00\xEA\xB8\x80\x00\xEA\xB8\x81\x00\xEA\xB8\x82"
-     "\x00\xEA"
-     "\xB8\x83\x00\xEA\xB8\x84\x00\xEA\xB8\x85\x00\xEA\xB8\x86\x00\xEA\xB8\x87"
-     "\x00\xEA"
-     "\xB8\x88\x00\xEA\xB8\x89\x00\xEA\xB8\x8A\x00\xEA\xB8\x8B\x00\xEA\xB8\x8C"
-     "\x00\xEA"
-     "\xB8\x8D\x00\xEA\xB8\x8E\x00\xEA\xB8\x8F\x00\xEA\xB8\x90\x00\xEA\xB8\x91"
-     "\x00\xEA"
-     "\xB8\x92\x00\xEA\xB8\x93\x00\xEA\xB8\x94\x00\xEA\xB8\x95\x00\xEA\xB8\x96"
-     "\x00\xEA"
-     "\xB8\x97\x00\xEA\xB8\x98\x00\xEA\xB8\x99\x00\xEA\xB8\x9A\x00\xEA\xB8\x9B"
-     "\x00\xEA"
-     "\xB8\x9C\x00\xEA\xB8\x9D\x00\xEA\xB8\x9E\x00\xEA\xB8\x9F\x00\xEA\xB8\xA0"
-     "\x00\xEA"
-     "\xB8\xA1\x00\xEA\xB8\xA2\x00\xEA\xB8\xA3\x00\xEA\xB8\xA4\x00\xEA\xB8\xA5"
-     "\x00\xEA"
-     "\xB8\xA6\x00\xEA\xB8\xA7\x00\xEA\xB8\xA8\x00\xEA\xB8\xA9\x00\xEA\xB8\xAA"
-     "\x00\xEA"
-     "\xB8\xAB\x00\xEA\xB8\xAC\x00\xEA\xB8\xAD\x00\xEA\xB8\xAE\x00\xEA\xB8\xAF"
-     "\x00\xEA"
-     "\xB8\xB0\x00\xEA\xB8\xB1\x00\xEA\xB8\xB2\x00\xEA\xB8\xB3\x00\xEA\xB8\xB4"
-     "\x00\xEA"
-     "\xB8\xB5\x00\xEA\xB8\xB6\x00\xEA\xB8\xB7\x00\xEA\xB8\xB8\x00\xEA\xB8\xB9"
-     "\x00\xEA"
-     "\xB8\xBA\x00\xEA\xB8\xBB\x00\xEA\xB8\xBC\x00\xEA\xB8\xBD\x00\xEA\xB8\xBE"
-     "\x00\xEA"
-     "\xB8\xBF\x00\xEA\xB9\x80\x00\xEA\xB9\x81\x00\xEA\xB9\x82\x00\xEA\xB9\x83"
-     "\x00\xEA"
-     "\xB9\x84\x00\xEA\xB9\x85\x00\xEA\xB9\x86\x00\xEA\xB9\x87\x00\xEA\xB9\x88"
-     "\x00\xEA"
-     "\xB9\x89\x00\xEA\xB9\x8A\x00\xEA\xB9\x8B\x00\xEA\xB9\x8C\x00\xEA\xB9\x8D"
-     "\x00\xEA"
-     "\xB9\x8E\x00\xEA\xB9\x8F\x00\xEA\xB9\x90\x00\xEA\xB9\x91\x00\xEA\xB9\x92"
-     "\x00\xEA"
-     "\xB9\x93\x00\xEA\xB9\x94\x00\xEA\xB9\x95\x00\xEA\xB9\x96\x00\xEA\xB9\x97"
-     "\x00\xEA"
-     "\xB9\x98\x00\xEA\xB9\x99\x00\xEA\xB9\x9A\x00\xEA\xB9\x9B\x00\xEA\xB9\x9C"
-     "\x00\xEA"
-     "\xB9\x9D\x00\xEA\xB9\x9E\x00\xEA\xB9\x9F\x00\xEA\xB9\xA0\x00\xEA\xB9\xA1"
-     "\x00\xEA"
-     "\xB9\xA2\x00\xEA\xB9\xA3\x00\xEA\xB9\xA4\x00\xEA\xB9\xA5\x00\xEA\xB9\xA6"
-     "\x00\xEA"
-     "\xB9\xA7\x00\xEA\xB9\xA8\x00\xEA\xB9\xA9\x00\xEA\xB9\xAA\x00\xEA\xB9\xAB"
-     "\x00\xEA"
-     "\xB9\xAC\x00\xEA\xB9\xAD\x00\xEA\xB9\xAE\x00\xEA\xB9\xAF\x00\xEA\xB9\xB0"
-     "\x00\xEA"
-     "\xB9\xB1\x00\xEA\xB9\xB2\x00\xEA\xB9\xB3\x00\xEA\xB9\xB4\x00\xEA\xB9\xB5"
-     "\x00\xEA"
-     "\xB9\xB6\x00\xEA\xB9\xB7\x00\xEA\xB9\xB8\x00\xEA\xB9\xB9\x00\xEA\xB9\xBA"
-     "\x00\xEA"
-     "\xB9\xBB\x00\xEA\xB9\xBC\x00\xEA\xB9\xBD\x00\xEA\xB9\xBE\x00\xEA\xB9\xBF"
-     "\x00\xEA"
-     "\xBA\x80\x00\xEA\xBA\x81\x00\xEA\xBA\x82\x00\xEA\xBA\x83\x00\xEA\xBA\x84"
-     "\x00\xEA"
-     "\xBA\x85\x00\xEA\xBA\x86\x00\xEA\xBA\x87\x00\xEA\xBA\x88\x00\xEA\xBA\x89"
-     "\x00\xEA"
-     "\xBA\x8A\x00\xEA\xBA\x8B\x00\xEA\xBA\x8C\x00\xEA\xBA\x8D\x00\xEA\xBA\x8E"
-     "\x00\xEA"
-     "\xBA\x8F\x00\xEA\xBA\x90\x00\xEA\xBA\x91\x00\xEA\xBA\x92\x00\xEA\xBA\x93"
-     "\x00\xEA"
-     "\xBA\x94\x00\xEA\xBA\x95\x00\xEA\xBA\x96\x00\xEA\xBA\x97\x00\xEA\xBA\x98"
-     "\x00\xEA"
-     "\xBA\x99\x00\xEA\xBA\x9A\x00\xEA\xBA\x9B\x00\xEA\xBA\x9C\x00\xEA\xBA\x9D"
-     "\x00\xEA"
-     "\xBA\x9E\x00\xEA\xBA\x9F\x00\xEA\xBA\xA0\x00\xEA\xBA\xA1\x00\xEA\xBA\xA2"
-     "\x00\xEA"
-     "\xBA\xA3\x00\xEA\xBA\xA4\x00\xEA\xBA\xA5\x00\xEA\xBA\xA6\x00\xEA\xBA\xA7"
-     "\x00\xEA"
-     "\xBA\xA8\x00\xEA\xBA\xA9\x00\xEA\xBA\xAA\x00\xEA\xBA\xAB\x00\xEA\xBA\xAC"
-     "\x00\xEA"
-     "\xBA\xAD\x00\xEA\xBA\xAE\x00\xEA\xBA\xAF\x00\xEA\xBA\xB0\x00\xEA\xBA\xB1"
-     "\x00\xEA"
-     "\xBA\xB2\x00\xEA\xBA\xB3\x00\xEA\xBA\xB4\x00\xEA\xBA\xB5\x00\xEA\xBA\xB6"
-     "\x00\xEA"
-     "\xBA\xB7\x00\xEA\xBA\xB8\x00\xEA\xBA\xB9\x00\xEA\xBA\xBA\x00\xEA\xBA\xBB"
-     "\x00\xEA"
-     "\xBA\xBC\x00\xEA\xBA\xBD\x00\xEA\xBA\xBE\x00\xEA\xBA\xBF\x00\xEA\xBB\x80"
-     "\x00\xEA"
-     "\xBB\x81\x00\xEA\xBB\x82\x00\xEA\xBB\x83\x00\xEA\xBB\x84\x00\xEA\xBB\x85"
-     "\x00\xEA"
-     "\xBB\x86\x00\xEA\xBB\x87\x00\xEA\xBB\x88\x00\xEA\xBB\x89\x00\xEA\xBB\x8A"
-     "\x00\xEA"
-     "\xBB\x8B\x00\xEA\xBB\x8C\x00\xEA\xBB\x8D\x00\xEA\xBB\x8E\x00\xEA\xBB\x8F"
-     "\x00\xEA"
-     "\xBB\x90\x00\xEA\xBB\x91\x00\xEA\xBB\x92\x00\xEA\xBB\x93\x00\xEA\xBB\x94"
-     "\x00\xEA"
-     "\xBB\x95\x00\xEA\xBB\x96\x00\xEA\xBB\x97\x00\xEA\xBB\x98\x00\xEA\xBB\x99"
-     "\x00\xEA"
-     "\xBB\x9A\x00\xEA\xBB\x9B\x00\xEA\xBB\x9C\x00\xEA\xBB\x9D\x00\xEA\xBB\x9E"
-     "\x00\xEA"
-     "\xBB\x9F\x00\xEA\xBB\xA0\x00\xEA\xBB\xA1\x00\xEA\xBB\xA2\x00\xEA\xBB\xA3"
-     "\x00\xEA"
-     "\xBB\xA4\x00\xEA\xBB\xA5\x00\xEA\xBB\xA6\x00\xEA\xBB\xA7\x00\xEA\xBB\xA8"
-     "\x00\xEA"
-     "\xBB\xA9\x00\xEA\xBB\xAA\x00\xEA\xBB\xAB\x00\xEA\xBB\xAC\x00\xEA\xBB\xAD"
-     "\x00\xEA"
-     "\xBB\xAE\x00\xEA\xBB\xAF\x00\xEA\xBB\xB0\x00\xEA\xBB\xB1\x00\xEA\xBB\xB2"
-     "\x00\xEA"
-     "\xBB\xB3\x00\xEA\xBB\xB4\x00\xEA\xBB\xB5\x00\xEA\xBB\xB6\x00\xEA\xBB\xB7"
-     "\x00\xEA"
-     "\xBB\xB8\x00\xEA\xBB\xB9\x00\xEA\xBB\xBA\x00\xEA\xBB\xBB\x00\xEA\xBB\xBC"
-     "\x00\xEA"
-     "\xBB\xBD\x00\xEA\xBB\xBE\x00\xEA\xBB\xBF\x00\xEA\xBC\x80\x00\xEA\xBC\x81"
-     "\x00\xEA"
-     "\xBC\x82\x00\xEA\xBC\x83\x00\xEA\xBC\x84\x00\xEA\xBC\x85\x00\xEA\xBC\x86"
-     "\x00\xEA"
-     "\xBC\x87\x00\xEA\xBC\x88\x00\xEA\xBC\x89\x00\xEA\xBC\x8A\x00\xEA\xBC\x8B"
-     "\x00\xEA"
-     "\xBC\x8C\x00\xEA\xBC\x8D\x00\xEA\xBC\x8E\x00\xEA\xBC\x8F\x00\xEA\xBC\x90"
-     "\x00\xEA"
-     "\xBC\x91\x00\xEA\xBC\x92\x00\xEA\xBC\x93\x00\xEA\xBC\x94\x00\xEA\xBC\x95"
-     "\x00\xEA"
-     "\xBC\x96\x00\xEA\xBC\x97\x00\xEA\xBC\x98\x00\xEA\xBC\x99\x00\xEA\xBC\x9A"
-     "\x00\xEA"
-     "\xBC\x9B\x00\xEA\xBC\x9C\x00\xEA\xBC\x9D\x00\xEA\xBC\x9E\x00\xEA\xBC\x9F"
-     "\x00\xEA"
-     "\xBC\xA0\x00\xEA\xBC\xA1\x00\xEA\xBC\xA2\x00\xEA\xBC\xA3\x00\xEA\xBC\xA4"
-     "\x00\xEA"
-     "\xBC\xA5\x00\xEA\xBC\xA6\x00\xEA\xBC\xA7\x00\xEA\xBC\xA8\x00\xEA\xBC\xA9"
-     "\x00\xEA"
-     "\xBC\xAA\x00\xEA\xBC\xAB\x00\xEA\xBC\xAC\x00\xEA\xBC\xAD\x00\xEA\xBC\xAE"
-     "\x00\xEA"
-     "\xBC\xAF\x00\xEA\xBC\xB0\x00\xEA\xBC\xB1\x00\xEA\xBC\xB2\x00\xEA\xBC\xB3"
-     "\x00\xEA"
-     "\xBC\xB4\x00\xEA\xBC\xB5\x00\xEA\xBC\xB6\x00\xEA\xBC\xB7\x00\xEA\xBC\xB8"
-     "\x00\xEA"
-     "\xBC\xB9\x00\xEA\xBC\xBA\x00\xEA\xBC\xBB\x00\xEA\xBC\xBC\x00\xEA\xBC\xBD"
-     "\x00\xEA"
-     "\xBC\xBE\x00\xEA\xBC\xBF\x00\xEA\xBD\x80\x00\xEA\xBD\x81\x00\xEA\xBD\x82"
-     "\x00\xEA"
-     "\xBD\x83\x00\xEA\xBD\x84\x00\xEA\xBD\x85\x00\xEA\xBD\x86\x00\xEA\xBD\x87"
-     "\x00\xEA"
-     "\xBD\x88\x00\xEA\xBD\x89\x00\xEA\xBD\x8A\x00\xEA\xBD\x8B\x00\xEA\xBD\x8C"
-     "\x00\xEA"
-     "\xBD\x8D\x00\xEA\xBD\x8E\x00\xEA\xBD\x8F\x00\xEA\xBD\x90\x00\xEA\xBD\x91"
-     "\x00\xEA"
-     "\xBD\x92\x00\xEA\xBD\x93\x00\xEA\xBD\x94\x00\xEA\xBD\x95\x00\xEA\xBD\x96"
-     "\x00\xEA"
-     "\xBD\x97\x00\xEA\xBD\x98\x00\xEA\xBD\x99\x00\xEA\xBD\x9A\x00\xEA\xBD\x9B"
-     "\x00\xEA"
-     "\xBD\x9C\x00\xEA\xBD\x9D\x00\xEA\xBD\x9E\x00\xEA\xBD\x9F\x00\xEA\xBD\xA0"
-     "\x00\xEA"
-     "\xBD\xA1\x00\xEA\xBD\xA2\x00\xEA\xBD\xA3\x00\xEA\xBD\xA4\x00\xEA\xBD\xA5"
-     "\x00\xEA"
-     "\xBD\xA6\x00\xEA\xBD\xA7\x00\xEA\xBD\xA8\x00\xEA\xBD\xA9\x00\xEA\xBD\xAA"
-     "\x00\xEA"
-     "\xBD\xAB\x00\xEA\xBD\xAC\x00\xEA\xBD\xAD\x00\xEA\xBD\xAE\x00\xEA\xBD\xAF"
-     "\x00\xEA"
-     "\xBD\xB0\x00\xEA\xBD\xB1\x00\xEA\xBD\xB2\x00\xEA\xBD\xB3\x00\xEA\xBD\xB4"
-     "\x00\xEA"
-     "\xBD\xB5\x00\xEA\xBD\xB6\x00\xEA\xBD\xB7\x00\xEA\xBD\xB8\x00\xEA\xBD\xB9"
-     "\x00\xEA"
-     "\xBD\xBA\x00\xEA\xBD\xBB\x00\xEA\xBD\xBC\x00\xEA\xBD\xBD\x00\xEA\xBD\xBE"
-     "\x00\xEA"
-     "\xBD\xBF\x00\xEA\xBE\x80\x00\xEA\xBE\x81\x00\xEA\xBE\x82\x00\xEA\xBE\x83"
-     "\x00\xEA"
-     "\xBE\x84\x00\xEA\xBE\x85\x00\xEA\xBE\x86\x00\xEA\xBE\x87\x00\xEA\xBE\x88"
-     "\x00\xEA"
-     "\xBE\x89\x00\xEA\xBE\x8A\x00\xEA\xBE\x8B\x00\xEA\xBE\x8C\x00\xEA\xBE\x8D"
-     "\x00\xEA"
-     "\xBE\x8E\x00\xEA\xBE\x8F\x00\xEA\xBE\x90\x00\xEA\xBE\x91\x00\xEA\xBE\x92"
-     "\x00\xEA"
-     "\xBE\x93\x00\xEA\xBE\x94\x00\xEA\xBE\x95\x00\xEA\xBE\x96\x00\xEA\xBE\x97"
-     "\x00\xEA"
-     "\xBE\x98\x00\xEA\xBE\x99\x00\xEA\xBE\x9A\x00\xEA\xBE\x9B\x00\xEA\xBE\x9C"
-     "\x00\xEA"
-     "\xBE\x9D\x00\xEA\xBE\x9E\x00\xEA\xBE\x9F\x00\xEA\xBE\xA0\x00\xEA\xBE\xA1"
-     "\x00\xEA"
-     "\xBE\xA2\x00\xEA\xBE\xA3\x00\xEA\xBE\xA4\x00\xEA\xBE\xA5\x00\xEA\xBE\xA6"
-     "\x00\xEA"
-     "\xBE\xA7\x00\xEA\xBE\xA8\x00\xEA\xBE\xA9\x00\xEA\xBE\xAA\x00\xEA\xBE\xAB"
-     "\x00\xEA"
-     "\xBE\xAC\x00\xEA\xBE\xAD\x00\xEA\xBE\xAE\x00\xEA\xBE\xAF\x00\xEA\xBE\xB0"
-     "\x00\xEA"
-     "\xBE\xB1\x00\xEA\xBE\xB2\x00\xEA\xBE\xB3\x00\xEA\xBE\xB4\x00\xEA\xBE\xB5"
-     "\x00\xEA"
-     "\xBE\xB6\x00\xEA\xBE\xB7\x00\xEA\xBE\xB8\x00\xEA\xBE\xB9\x00\xEA\xBE\xBA"
-     "\x00\xEA"
-     "\xBE\xBB\x00\xEA\xBE\xBC\x00\xEA\xBE\xBD\x00\xEA\xBE\xBE\x00\xEA\xBE\xBF"
-     "\x00\xEA"
-     "\xBF\x80\x00\xEA\xBF\x81\x00\xEA\xBF\x82\x00\xEA\xBF\x83\x00\xEA\xBF\x84"
-     "\x00\xEA"
-     "\xBF\x85\x00\xEA\xBF\x86\x00\xEA\xBF\x87\x00\xEA\xBF\x88\x00\xEA\xBF\x89"
-     "\x00\xEA"
-     "\xBF\x8A\x00\xEA\xBF\x8B\x00\xEA\xBF\x8C\x00\xEA\xBF\x8D\x00\xEA\xBF\x8E"
-     "\x00\xEA"
-     "\xBF\x8F\x00\xEA\xBF\x90\x00\xEA\xBF\x91\x00\xEA\xBF\x92\x00\xEA\xBF\x93"
-     "\x00\xEA"
-     "\xBF\x94\x00\xEA\xBF\x95\x00\xEA\xBF\x96\x00\xEA\xBF\x97\x00\xEA\xBF\x98"
-     "\x00\xEA"
-     "\xBF\x99\x00\xEA\xBF\x9A\x00\xEA\xBF\x9B\x00\xEA\xBF\x9C\x00\xEA\xBF\x9D"
-     "\x00\xEA"
-     "\xBF\x9E\x00\xEA\xBF\x9F\x00\xEA\xBF\xA0\x00\xEA\xBF\xA1\x00\xEA\xBF\xA2"
-     "\x00\xEA"
-     "\xBF\xA3\x00\xEA\xBF\xA4\x00\xEA\xBF\xA5\x00\xEA\xBF\xA6\x00\xEA\xBF\xA7"
-     "\x00\xEA"
-     "\xBF\xA8\x00\xEA\xBF\xA9\x00\xEA\xBF\xAA\x00\xEA\xBF\xAB\x00\xEA\xBF\xAC"
-     "\x00\xEA"
-     "\xBF\xAD\x00\xEA\xBF\xAE\x00\xEA\xBF\xAF\x00\xEA\xBF\xB0\x00\xEA\xBF\xB1"
-     "\x00\xEA"
-     "\xBF\xB2\x00\xEA\xBF\xB3\x00\xEA\xBF\xB4\x00\xEA\xBF\xB5\x00\xEA\xBF\xB6"
-     "\x00\xEA"
-     "\xBF\xB7\x00\xEA\xBF\xB8\x00\xEA\xBF\xB9\x00\xEA\xBF\xBA\x00\xEA\xBF\xBB"
-     "\x00\xEA"
-     "\xBF\xBC\x00\xEA\xBF\xBD\x00\xEA\xBF\xBE\x00\xEA\xBF\xBF\x00\xEB\x80\x80"
-     "\x00\xEB"
-     "\x80\x81\x00\xEB\x80\x82\x00\xEB\x80\x83\x00\xEB\x80\x84\x00\xEB\x80\x85"
-     "\x00\xEB"
-     "\x80\x86\x00\xEB\x80\x87\x00\xEB\x80\x88\x00\xEB\x80\x89\x00\xEB\x80\x8A"
-     "\x00\xEB"
-     "\x80\x8B\x00\xEB\x80\x8C\x00\xEB\x80\x8D\x00\xEB\x80\x8E\x00\xEB\x80\x8F"
-     "\x00\xEB"
-     "\x80\x90\x00\xEB\x80\x91\x00\xEB\x80\x92\x00\xEB\x80\x93\x00\xEB\x80\x94"
-     "\x00\xEB"
-     "\x80\x95\x00\xEB\x80\x96\x00\xEB\x80\x97\x00\xEB\x80\x98\x00\xEB\x80\x99"
-     "\x00\xEB"
-     "\x80\x9A\x00\xEB\x80\x9B\x00\xEB\x80\x9C\x00\xEB\x80\x9D\x00\xEB\x80\x9E"
-     "\x00\xEB"
-     "\x80\x9F\x00\xEB\x80\xA0\x00\xEB\x80\xA1\x00\xEB\x80\xA2\x00\xEB\x80\xA3"
-     "\x00\xEB"
-     "\x80\xA4\x00\xEB\x80\xA5\x00\xEB\x80\xA6\x00\xEB\x80\xA7\x00\xEB\x80\xA8"
-     "\x00\xEB"
-     "\x80\xA9\x00\xEB\x80\xAA\x00\xEB\x80\xAB\x00\xEB\x80\xAC\x00\xEB\x80\xAD"
-     "\x00\xEB"
-     "\x80\xAE\x00\xEB\x80\xAF\x00\xEB\x80\xB0\x00\xEB\x80\xB1\x00\xEB\x80\xB2"
-     "\x00\xEB"
-     "\x80\xB3\x00\xEB\x80\xB4\x00\xEB\x80\xB5\x00\xEB\x80\xB6\x00\xEB\x80\xB7"
-     "\x00\xEB"
-     "\x80\xB8\x00\xEB\x80\xB9\x00\xEB\x80\xBA\x00\xEB\x80\xBB\x00\xEB\x80\xBC"
-     "\x00\xEB"
-     "\x80\xBD\x00\xEB\x80\xBE\x00\xEB\x80\xBF\x00\xEB\x81\x80\x00\xEB\x81\x81"
-     "\x00\xEB"
-     "\x81\x82\x00\xEB\x81\x83\x00\xEB\x81\x84\x00\xEB\x81\x85\x00\xEB\x81\x86"
-     "\x00\xEB"
-     "\x81\x87\x00\xEB\x81\x88\x00\xEB\x81\x89\x00\xEB\x81\x8A\x00\xEB\x81\x8B"
-     "\x00\xEB"
-     "\x81\x8C\x00\xEB\x81\x8D\x00\xEB\x81\x8E\x00\xEB\x81\x8F\x00\xEB\x81\x90"
-     "\x00\xEB"
-     "\x81\x91\x00\xEB\x81\x92\x00\xEB\x81\x93\x00\xEB\x81\x94\x00\xEB\x81\x95"
-     "\x00\xEB"
-     "\x81\x96\x00\xEB\x81\x97\x00\xEB\x81\x98\x00\xEB\x81\x99\x00\xEB\x81\x9A"
-     "\x00\xEB"
-     "\x81\x9B\x00\xEB\x81\x9C\x00\xEB\x81\x9D\x00\xEB\x81\x9E\x00\xEB\x81\x9F"
-     "\x00\xEB"
-     "\x81\xA0\x00\xEB\x81\xA1\x00\xEB\x81\xA2\x00\xEB\x81\xA3\x00\xEB\x81\xA4"
-     "\x00\xEB"
-     "\x81\xA5\x00\xEB\x81\xA6\x00\xEB\x81\xA7\x00\xEB\x81\xA8\x00\xEB\x81\xA9"
-     "\x00\xEB"
-     "\x81\xAA\x00\xEB\x81\xAB\x00\xEB\x81\xAC\x00\xEB\x81\xAD\x00\xEB\x81\xAE"
-     "\x00\xEB"
-     "\x81\xAF\x00\xEB\x81\xB0\x00\xEB\x81\xB1\x00\xEB\x81\xB2\x00\xEB\x81\xB3"
-     "\x00\xEB"
-     "\x81\xB4\x00\xEB\x81\xB5\x00\xEB\x81\xB6\x00\xEB\x81\xB7\x00\xEB\x81\xB8"
-     "\x00\xEB"
-     "\x81\xB9\x00\xEB\x81\xBA\x00\xEB\x81\xBB\x00\xEB\x81\xBC\x00\xEB\x81\xBD"
-     "\x00\xEB"
-     "\x81\xBE\x00\xEB\x81\xBF\x00\xEB\x82\x80\x00\xEB\x82\x81\x00\xEB\x82\x82"
-     "\x00\xEB"
-     "\x82\x83\x00\xEB\x82\x84\x00\xEB\x82\x85\x00\xEB\x82\x86\x00\xEB\x82\x87"
-     "\x00\xEB"
-     "\x82\x88\x00\xEB\x82\x89\x00\xEB\x82\x8A\x00\xEB\x82\x8B\x00\xEB\x82\x8C"
-     "\x00\xEB"
-     "\x82\x8D\x00\xEB\x82\x8E\x00\xEB\x82\x8F\x00\xEB\x82\x90\x00\xEB\x82\x91"
-     "\x00\xEB"
-     "\x82\x92\x00\xEB\x82\x93\x00\xEB\x82\x94\x00\xEB\x82\x95\x00\xEB\x82\x96"
-     "\x00\xEB"
-     "\x82\x97\x00\xEB\x82\x98\x00\xEB\x82\x99\x00\xEB\x82\x9A\x00\xEB\x82\x9B"
-     "\x00\xEB"
-     "\x82\x9C\x00\xEB\x82\x9D\x00\xEB\x82\x9E\x00\xEB\x82\x9F\x00\xEB\x82\xA0"
-     "\x00\xEB"
-     "\x82\xA1\x00\xEB\x82\xA2\x00\xEB\x82\xA3\x00\xEB\x82\xA4\x00\xEB\x82\xA5"
-     "\x00\xEB"
-     "\x82\xA6\x00\xEB\x82\xA7\x00\xEB\x82\xA8\x00\xEB\x82\xA9\x00\xEB\x82\xAA"
-     "\x00\xEB"
-     "\x82\xAB\x00\xEB\x82\xAC\x00\xEB\x82\xAD\x00\xEB\x82\xAE\x00\xEB\x82\xAF"
-     "\x00\xEB"
-     "\x82\xB0\x00\xEB\x82\xB1\x00\xEB\x82\xB2\x00\xEB\x82\xB3\x00\xEB\x82\xB4"
-     "\x00\xEB"
-     "\x82\xB5\x00\xEB\x82\xB6\x00\xEB\x82\xB7\x00\xEB\x82\xB8\x00\xEB\x82\xB9"
-     "\x00\xEB"
-     "\x82\xBA\x00\xEB\x82\xBB\x00\xEB\x82\xBC\x00\xEB\x82\xBD\x00\xEB\x82\xBE"
-     "\x00\xEB"
-     "\x82\xBF\x00\xEB\x83\x80\x00\xEB\x83\x81\x00\xEB\x83\x82\x00\xEB\x83\x83"
-     "\x00\xEB"
-     "\x83\x84\x00\xEB\x83\x85\x00\xEB\x83\x86\x00\xEB\x83\x87\x00\xEB\x83\x88"
-     "\x00\xEB"
-     "\x83\x89\x00\xEB\x83\x8A\x00\xEB\x83\x8B\x00\xEB\x83\x8C\x00\xEB\x83\x8D"
-     "\x00\xEB"
-     "\x83\x8E\x00\xEB\x83\x8F\x00\xEB\x83\x90\x00\xEB\x83\x91\x00\xEB\x83\x92"
-     "\x00\xEB"
-     "\x83\x93\x00\xEB\x83\x94\x00\xEB\x83\x95\x00\xEB\x83\x96\x00\xEB\x83\x97"
-     "\x00\xEB"
-     "\x83\x98\x00\xEB\x83\x99\x00\xEB\x83\x9A\x00\xEB\x83\x9B\x00\xEB\x83\x9C"
-     "\x00\xEB"
-     "\x83\x9D\x00\xEB\x83\x9E\x00\xEB\x83\x9F\x00\xEB\x83\xA0\x00\xEB\x83\xA1"
-     "\x00\xEB"
-     "\x83\xA2\x00\xEB\x83\xA3\x00\xEB\x83\xA4\x00\xEB\x83\xA5\x00\xEB\x83\xA6"
-     "\x00\xEB"
-     "\x83\xA7\x00\xEB\x83\xA8\x00\xEB\x83\xA9\x00\xEB\x83\xAA\x00\xEB\x83\xAB"
-     "\x00\xEB"
-     "\x83\xAC\x00\xEB\x83\xAD\x00\xEB\x83\xAE\x00\xEB\x83\xAF\x00\xEB\x83\xB0"
-     "\x00\xEB"
-     "\x83\xB1\x00\xEB\x83\xB2\x00\xEB\x83\xB3\x00\xEB\x83\xB4\x00\xEB\x83\xB5"
-     "\x00\xEB"
-     "\x83\xB6\x00\xEB\x83\xB7\x00\xEB\x83\xB8\x00\xEB\x83\xB9\x00\xEB\x83\xBA"
-     "\x00\xEB"
-     "\x83\xBB\x00\xEB\x83\xBC\x00\xEB\x83\xBD\x00\xEB\x83\xBE\x00\xEB\x83\xBF"
-     "\x00\xEB"
-     "\x84\x80\x00\xEB\x84\x81\x00\xEB\x84\x82\x00\xEB\x84\x83\x00\xEB\x84\x84"
-     "\x00\xEB"
-     "\x84\x85\x00\xEB\x84\x86\x00\xEB\x84\x87\x00\xEB\x84\x88\x00\xEB\x84\x89"
-     "\x00\xEB"
-     "\x84\x8A\x00\xEB\x84\x8B\x00\xEB\x84\x8C\x00\xEB\x84\x8D\x00\xEB\x84\x8E"
-     "\x00\xEB"
-     "\x84\x8F\x00\xEB\x84\x90\x00\xEB\x84\x91\x00\xEB\x84\x92\x00\xEB\x84\x93"
-     "\x00\xEB"
-     "\x84\x94\x00\xEB\x84\x95\x00\xEB\x84\x96\x00\xEB\x84\x97\x00\xEB\x84\x98"
-     "\x00\xEB"
-     "\x84\x99\x00\xEB\x84\x9A\x00\xEB\x84\x9B\x00\xEB\x84\x9C\x00\xEB\x84\x9D"
-     "\x00\xEB"
-     "\x84\x9E\x00\xEB\x84\x9F\x00\xEB\x84\xA0\x00\xEB\x84\xA1\x00\xEB\x84\xA2"
-     "\x00\xEB"
-     "\x84\xA3\x00\xEB\x84\xA4\x00\xEB\x84\xA5\x00\xEB\x84\xA6\x00\xEB\x84\xA7"
-     "\x00\xEB"
-     "\x84\xA8\x00\xEB\x84\xA9\x00\xEB\x84\xAA\x00\xEB\x84\xAB\x00\xEB\x84\xAC"
-     "\x00\xEB"
-     "\x84\xAD\x00\xEB\x84\xAE\x00\xEB\x84\xAF\x00\xEB\x84\xB0\x00\xEB\x84\xB1"
-     "\x00\xEB"
-     "\x84\xB2\x00\xEB\x84\xB3\x00\xEB\x84\xB4\x00\xEB\x84\xB5\x00\xEB\x84\xB6"
-     "\x00\xEB"
-     "\x84\xB7\x00\xEB\x84\xB8\x00\xEB\x84\xB9\x00\xEB\x84\xBA\x00\xEB\x84\xBB"
-     "\x00\xEB"
-     "\x84\xBC\x00\xEB\x84\xBD\x00\xEB\x84\xBE\x00\xEB\x84\xBF\x00\xEB\x85\x80"
-     "\x00\xEB"
-     "\x85\x81\x00\xEB\x85\x82\x00\xEB\x85\x83\x00\xEB\x85\x84\x00\xEB\x85\x85"
-     "\x00\xEB"
-     "\x85\x86\x00\xEB\x85\x87\x00\xEB\x85\x88\x00\xEB\x85\x89\x00\xEB\x85\x8A"
-     "\x00\xEB"
-     "\x85\x8B\x00\xEB\x85\x8C\x00\xEB\x85\x8D\x00\xEB\x85\x8E\x00\xEB\x85\x8F"
-     "\x00\xEB"
-     "\x85\x90\x00\xEB\x85\x91\x00\xEB\x85\x92\x00\xEB\x85\x93\x00\xEB\x85\x94"
-     "\x00\xEB"
-     "\x85\x95\x00\xEB\x85\x96\x00\xEB\x85\x97\x00\xEB\x85\x98\x00\xEB\x85\x99"
-     "\x00\xEB"
-     "\x85\x9A\x00\xEB\x85\x9B\x00\xEB\x85\x9C\x00\xEB\x85\x9D\x00\xEB\x85\x9E"
-     "\x00\xEB"
-     "\x85\x9F\x00\xEB\x85\xA0\x00\xEB\x85\xA1\x00\xEB\x85\xA2\x00\xEB\x85\xA3"
-     "\x00\xEB"
-     "\x85\xA4\x00\xEB\x85\xA5\x00\xEB\x85\xA6\x00\xEB\x85\xA7\x00\xEB\x85\xA8"
-     "\x00\xEB"
-     "\x85\xA9\x00\xEB\x85\xAA\x00\xEB\x85\xAB\x00\xEB\x85\xAC\x00\xEB\x85\xAD"
-     "\x00\xEB"
-     "\x85\xAE\x00\xEB\x85\xAF\x00\xEB\x85\xB0\x00\xEB\x85\xB1\x00\xEB\x85\xB2"
-     "\x00\xEB"
-     "\x85\xB3\x00\xEB\x85\xB4\x00\xEB\x85\xB5\x00\xEB\x85\xB6\x00\xEB\x85\xB7"
-     "\x00\xEB"
-     "\x85\xB8\x00\xEB\x85\xB9\x00\xEB\x85\xBA\x00\xEB\x85\xBB\x00\xEB\x85\xBC"
-     "\x00\xEB"
-     "\x85\xBD\x00\xEB\x85\xBE\x00\xEB\x85\xBF\x00\xEB\x86\x80\x00\xEB\x86\x81"
-     "\x00\xEB"
-     "\x86\x82\x00\xEB\x86\x83\x00\xEB\x86\x84\x00\xEB\x86\x85\x00\xEB\x86\x86"
-     "\x00\xEB"
-     "\x86\x87\x00\xEB\x86\x88\x00\xEB\x86\x89\x00\xEB\x86\x8A\x00\xEB\x86\x8B"
-     "\x00\xEB"
-     "\x86\x8C\x00\xEB\x86\x8D\x00\xEB\x86\x8E\x00\xEB\x86\x8F\x00\xEB\x86\x90"
-     "\x00\xEB"
-     "\x86\x91\x00\xEB\x86\x92\x00\xEB\x86\x93\x00\xEB\x86\x94\x00\xEB\x86\x95"
-     "\x00\xEB"
-     "\x86\x96\x00\xEB\x86\x97\x00\xEB\x86\x98\x00\xEB\x86\x99\x00\xEB\x86\x9A"
-     "\x00\xEB"
-     "\x86\x9B\x00\xEB\x86\x9C\x00\xEB\x86\x9D\x00\xEB\x86\x9E\x00\xEB\x86\x9F"
-     "\x00\xEB"
-     "\x86\xA0\x00\xEB\x86\xA1\x00\xEB\x86\xA2\x00\xEB\x86\xA3\x00\xEB\x86\xA4"
-     "\x00\xEB"
-     "\x86\xA5\x00\xEB\x86\xA6\x00\xEB\x86\xA7\x00\xEB\x86\xA8\x00\xEB\x86\xA9"
-     "\x00\xEB"
-     "\x86\xAA\x00\xEB\x86\xAB\x00\xEB\x86\xAC\x00\xEB\x86\xAD\x00\xEB\x86\xAE"
-     "\x00\xEB"
-     "\x86\xAF\x00\xEB\x86\xB0\x00\xEB\x86\xB1\x00\xEB\x86\xB2\x00\xEB\x86\xB3"
-     "\x00\xEB"
-     "\x86\xB4\x00\xEB\x86\xB5\x00\xEB\x86\xB6\x00\xEB\x86\xB7\x00\xEB\x86\xB8"
-     "\x00\xEB"
-     "\x86\xB9\x00\xEB\x86\xBA\x00\xEB\x86\xBB\x00\xEB\x86\xBC\x00\xEB\x86\xBD"
-     "\x00\xEB"
-     "\x86\xBE\x00\xEB\x86\xBF\x00\xEB\x87\x80\x00\xEB\x87\x81\x00\xEB\x87\x82"
-     "\x00\xEB"
-     "\x87\x83\x00\xEB\x87\x84\x00\xEB\x87\x85\x00\xEB\x87\x86\x00\xEB\x87\x87"
-     "\x00\xEB"
-     "\x87\x88\x00\xEB\x87\x89\x00\xEB\x87\x8A\x00\xEB\x87\x8B\x00\xEB\x87\x8C"
-     "\x00\xEB"
-     "\x87\x8D\x00\xEB\x87\x8E\x00\xEB\x87\x8F\x00\xEB\x87\x90\x00\xEB\x87\x91"
-     "\x00\xEB"
-     "\x87\x92\x00\xEB\x87\x93\x00\xEB\x87\x94\x00\xEB\x87\x95\x00\xEB\x87\x96"
-     "\x00\xEB"
-     "\x87\x97\x00\xEB\x87\x98\x00\xEB\x87\x99\x00\xEB\x87\x9A\x00\xEB\x87\x9B"
-     "\x00\xEB"
-     "\x87\x9C\x00\xEB\x87\x9D\x00\xEB\x87\x9E\x00\xEB\x87\x9F\x00\xEB\x87\xA0"
-     "\x00\xEB"
-     "\x87\xA1\x00\xEB\x87\xA2\x00\xEB\x87\xA3\x00\xEB\x87\xA4\x00\xEB\x87\xA5"
-     "\x00\xEB"
-     "\x87\xA6\x00\xEB\x87\xA7\x00\xEB\x87\xA8\x00\xEB\x87\xA9\x00\xEB\x87\xAA"
-     "\x00\xEB"
-     "\x87\xAB\x00\xEB\x87\xAC\x00\xEB\x87\xAD\x00\xEB\x87\xAE\x00\xEB\x87\xAF"
-     "\x00\xEB"
-     "\x87\xB0\x00\xEB\x87\xB1\x00\xEB\x87\xB2\x00\xEB\x87\xB3\x00\xEB\x87\xB4"
-     "\x00\xEB"
-     "\x87\xB5\x00\xEB\x87\xB6\x00\xEB\x87\xB7\x00\xEB\x87\xB8\x00\xEB\x87\xB9"
-     "\x00\xEB"
-     "\x87\xBA\x00\xEB\x87\xBB\x00\xEB\x87\xBC\x00\xEB\x87\xBD\x00\xEB\x87\xBE"
-     "\x00\xEB"
-     "\x87\xBF\x00\xEB\x88\x80\x00\xEB\x88\x81\x00\xEB\x88\x82\x00\xEB\x88\x83"
-     "\x00\xEB"
-     "\x88\x84\x00\xEB\x88\x85\x00\xEB\x88\x86\x00\xEB\x88\x87\x00\xEB\x88\x88"
-     "\x00\xEB"
-     "\x88\x89\x00\xEB\x88\x8A\x00\xEB\x88\x8B\x00\xEB\x88\x8C\x00\xEB\x88\x8D"
-     "\x00\xEB"
-     "\x88\x8E\x00\xEB\x88\x8F\x00\xEB\x88\x90\x00\xEB\x88\x91\x00\xEB\x88\x92"
-     "\x00\xEB"
-     "\x88\x93\x00\xEB\x88\x94\x00\xEB\x88\x95\x00\xEB\x88\x96\x00\xEB\x88\x97"
-     "\x00\xEB"
-     "\x88\x98\x00\xEB\x88\x99\x00\xEB\x88\x9A\x00\xEB\x88\x9B\x00\xEB\x88\x9C"
-     "\x00\xEB"
-     "\x88\x9D\x00\xEB\x88\x9E\x00\xEB\x88\x9F\x00\xEB\x88\xA0\x00\xEB\x88\xA1"
-     "\x00\xEB"
-     "\x88\xA2\x00\xEB\x88\xA3\x00\xEB\x88\xA4\x00\xEB\x88\xA5\x00\xEB\x88\xA6"
-     "\x00\xEB"
-     "\x88\xA7\x00\xEB\x88\xA8\x00\xEB\x88\xA9\x00\xEB\x88\xAA\x00\xEB\x88\xAB"
-     "\x00\xEB"
-     "\x88\xAC\x00\xEB\x88\xAD\x00\xEB\x88\xAE\x00\xEB\x88\xAF\x00\xEB\x88\xB0"
-     "\x00\xEB"
-     "\x88\xB1\x00\xEB\x88\xB2\x00\xEB\x88\xB3\x00\xEB\x88\xB4\x00\xEB\x88\xB5"
-     "\x00\xEB"
-     "\x88\xB6\x00\xEB\x88\xB7\x00\xEB\x88\xB8\x00\xEB\x88\xB9\x00\xEB\x88\xBA"
-     "\x00\xEB"
-     "\x88\xBB\x00\xEB\x88\xBC\x00\xEB\x88\xBD\x00\xEB\x88\xBE\x00\xEB\x88\xBF"
-     "\x00\xEB"
-     "\x89\x80\x00\xEB\x89\x81\x00\xEB\x89\x82\x00\xEB\x89\x83\x00\xEB\x89\x84"
-     "\x00\xEB"
-     "\x89\x85\x00\xEB\x89\x86\x00\xEB\x89\x87\x00\xEB\x89\x88\x00\xEB\x89\x89"
-     "\x00\xEB"
-     "\x89\x8A\x00\xEB\x89\x8B\x00\xEB\x89\x8C\x00\xEB\x89\x8D\x00\xEB\x89\x8E"
-     "\x00\xEB"
-     "\x89\x8F\x00\xEB\x89\x90\x00\xEB\x89\x91\x00\xEB\x89\x92\x00\xEB\x89\x93"
-     "\x00\xEB"
-     "\x89\x94\x00\xEB\x89\x95\x00\xEB\x89\x96\x00\xEB\x89\x97\x00\xEB\x89\x98"
-     "\x00\xEB"
-     "\x89\x99\x00\xEB\x89\x9A\x00\xEB\x89\x9B\x00\xEB\x89\x9C\x00\xEB\x89\x9D"
-     "\x00\xEB"
-     "\x89\x9E\x00\xEB\x89\x9F\x00\xEB\x89\xA0\x00\xEB\x89\xA1\x00\xEB\x89\xA2"
-     "\x00\xEB"
-     "\x89\xA3\x00\xEB\x89\xA4\x00\xEB\x89\xA5\x00\xEB\x89\xA6\x00\xEB\x89\xA7"
-     "\x00\xEB"
-     "\x89\xA8\x00\xEB\x89\xA9\x00\xEB\x89\xAA\x00\xEB\x89\xAB\x00\xEB\x89\xAC"
-     "\x00\xEB"
-     "\x89\xAD\x00\xEB\x89\xAE\x00\xEB\x89\xAF\x00\xEB\x89\xB0\x00\xEB\x89\xB1"
-     "\x00\xEB"
-     "\x89\xB2\x00\xEB\x89\xB3\x00\xEB\x89\xB4\x00\xEB\x89\xB5\x00\xEB\x89\xB6"
-     "\x00\xEB"
-     "\x89\xB7\x00\xEB\x89\xB8\x00\xEB\x89\xB9\x00\xEB\x89\xBA\x00\xEB\x89\xBB"
-     "\x00\xEB"
-     "\x89\xBC\x00\xEB\x89\xBD\x00\xEB\x89\xBE\x00\xEB\x89\xBF\x00\xEB\x8A\x80"
-     "\x00\xEB"
-     "\x8A\x81\x00\xEB\x8A\x82\x00\xEB\x8A\x83\x00\xEB\x8A\x84\x00\xEB\x8A\x85"
-     "\x00\xEB"
-     "\x8A\x86\x00\xEB\x8A\x87\x00\xEB\x8A\x88\x00\xEB\x8A\x89\x00\xEB\x8A\x8A"
-     "\x00\xEB"
-     "\x8A\x8B\x00\xEB\x8A\x8C\x00\xEB\x8A\x8D\x00\xEB\x8A\x8E\x00\xEB\x8A\x8F"
-     "\x00\xEB"
-     "\x8A\x90\x00\xEB\x8A\x91\x00\xEB\x8A\x92\x00\xEB\x8A\x93\x00\xEB\x8A\x94"
-     "\x00\xEB"
-     "\x8A\x95\x00\xEB\x8A\x96\x00\xEB\x8A\x97\x00\xEB\x8A\x98\x00\xEB\x8A\x99"
-     "\x00\xEB"
-     "\x8A\x9A\x00\xEB\x8A\x9B\x00\xEB\x8A\x9C\x00\xEB\x8A\x9D\x00\xEB\x8A\x9E"
-     "\x00\xEB"
-     "\x8A\x9F\x00\xEB\x8A\xA0\x00\xEB\x8A\xA1\x00\xEB\x8A\xA2\x00\xEB\x8A\xA3"
-     "\x00\xEB"
-     "\x8A\xA4\x00\xEB\x8A\xA5\x00\xEB\x8A\xA6\x00\xEB\x8A\xA7\x00\xEB\x8A\xA8"
-     "\x00\xEB"
-     "\x8A\xA9\x00\xEB\x8A\xAA\x00\xEB\x8A\xAB\x00\xEB\x8A\xAC\x00\xEB\x8A\xAD"
-     "\x00\xEB"
-     "\x8A\xAE\x00\xEB\x8A\xAF\x00\xEB\x8A\xB0\x00\xEB\x8A\xB1\x00\xEB\x8A\xB2"
-     "\x00\xEB"
-     "\x8A\xB3\x00\xEB\x8A\xB4\x00\xEB\x8A\xB5\x00\xEB\x8A\xB6\x00\xEB\x8A\xB7"
-     "\x00\xEB"
-     "\x8A\xB8\x00\xEB\x8A\xB9\x00\xEB\x8A\xBA\x00\xEB\x8A\xBB\x00\xEB\x8A\xBC"
-     "\x00\xEB"
-     "\x8A\xBD\x00\xEB\x8A\xBE\x00\xEB\x8A\xBF\x00\xEB\x8B\x80\x00\xEB\x8B\x81"
-     "\x00\xEB"
-     "\x8B\x82\x00\xEB\x8B\x83\x00\xEB\x8B\x84\x00\xEB\x8B\x85\x00\xEB\x8B\x86"
-     "\x00\xEB"
-     "\x8B\x87\x00\xEB\x8B\x88\x00\xEB\x8B\x89\x00\xEB\x8B\x8A\x00\xEB\x8B\x8B"
-     "\x00\xEB"
-     "\x8B\x8C\x00\xEB\x8B\x8D\x00\xEB\x8B\x8E\x00\xEB\x8B\x8F\x00\xEB\x8B\x90"
-     "\x00\xEB"
-     "\x8B\x91\x00\xEB\x8B\x92\x00\xEB\x8B\x93\x00\xEB\x8B\x94\x00\xEB\x8B\x95"
-     "\x00\xEB"
-     "\x8B\x96\x00\xEB\x8B\x97\x00\xEB\x8B\x98\x00\xEB\x8B\x99\x00\xEB\x8B\x9A"
-     "\x00\xEB"
-     "\x8B\x9B\x00\xEB\x8B\x9C\x00\xEB\x8B\x9D\x00\xEB\x8B\x9E\x00\xEB\x8B\x9F"
-     "\x00\xEB"
-     "\x8B\xA0\x00\xEB\x8B\xA1\x00\xEB\x8B\xA2\x00\xEB\x8B\xA3\x00\xEB\x8B\xA4"
-     "\x00\xEB"
-     "\x8B\xA5\x00\xEB\x8B\xA6\x00\xEB\x8B\xA7\x00\xEB\x8B\xA8\x00\xEB\x8B\xA9"
-     "\x00\xEB"
-     "\x8B\xAA\x00\xEB\x8B\xAB\x00\xEB\x8B\xAC\x00\xEB\x8B\xAD\x00\xEB\x8B\xAE"
-     "\x00\xEB"
-     "\x8B\xAF\x00\xEB\x8B\xB0\x00\xEB\x8B\xB1\x00\xEB\x8B\xB2\x00\xEB\x8B\xB3"
-     "\x00\xEB"
-     "\x8B\xB4\x00\xEB\x8B\xB5\x00\xEB\x8B\xB6\x00\xEB\x8B\xB7\x00\xEB\x8B\xB8"
-     "\x00\xEB"
-     "\x8B\xB9\x00\xEB\x8B\xBA\x00\xEB\x8B\xBB\x00\xEB\x8B\xBC\x00\xEB\x8B\xBD"
-     "\x00\xEB"
-     "\x8B\xBE\x00\xEB\x8B\xBF\x00\xEB\x8C\x80\x00\xEB\x8C\x81\x00\xEB\x8C\x82"
-     "\x00\xEB"
-     "\x8C\x83\x00\xEB\x8C\x84\x00\xEB\x8C\x85\x00\xEB\x8C\x86\x00\xEB\x8C\x87"
-     "\x00\xEB"
-     "\x8C\x88\x00\xEB\x8C\x89\x00\xEB\x8C\x8A\x00\xEB\x8C\x8B\x00\xEB\x8C\x8C"
-     "\x00\xEB"
-     "\x8C\x8D\x00\xEB\x8C\x8E\x00\xEB\x8C\x8F\x00\xEB\x8C\x90\x00\xEB\x8C\x91"
-     "\x00\xEB"
-     "\x8C\x92\x00\xEB\x8C\x93\x00\xEB\x8C\x94\x00\xEB\x8C\x95\x00\xEB\x8C\x96"
-     "\x00\xEB"
-     "\x8C\x97\x00\xEB\x8C\x98\x00\xEB\x8C\x99\x00\xEB\x8C\x9A\x00\xEB\x8C\x9B"
-     "\x00\xEB"
-     "\x8C\x9C\x00\xEB\x8C\x9D\x00\xEB\x8C\x9E\x00\xEB\x8C\x9F\x00\xEB\x8C\xA0"
-     "\x00\xEB"
-     "\x8C\xA1\x00\xEB\x8C\xA2\x00\xEB\x8C\xA3\x00\xEB\x8C\xA4\x00\xEB\x8C\xA5"
-     "\x00\xEB"
-     "\x8C\xA6\x00\xEB\x8C\xA7\x00\xEB\x8C\xA8\x00\xEB\x8C\xA9\x00\xEB\x8C\xAA"
-     "\x00\xEB"
-     "\x8C\xAB\x00\xEB\x8C\xAC\x00\xEB\x8C\xAD\x00\xEB\x8C\xAE\x00\xEB\x8C\xAF"
-     "\x00\xEB"
-     "\x8C\xB0\x00\xEB\x8C\xB1\x00\xEB\x8C\xB2\x00\xEB\x8C\xB3\x00\xEB\x8C\xB4"
-     "\x00\xEB"
-     "\x8C\xB5\x00\xEB\x8C\xB6\x00\xEB\x8C\xB7\x00\xEB\x8C\xB8\x00\xEB\x8C\xB9"
-     "\x00\xEB"
-     "\x8C\xBA\x00\xEB\x8C\xBB\x00\xEB\x8C\xBC\x00\xEB\x8C\xBD\x00\xEB\x8C\xBE"
-     "\x00\xEB"
-     "\x8C\xBF\x00\xEB\x8D\x80\x00\xEB\x8D\x81\x00\xEB\x8D\x82\x00\xEB\x8D\x83"
-     "\x00\xEB"
-     "\x8D\x84\x00\xEB\x8D\x85\x00\xEB\x8D\x86\x00\xEB\x8D\x87\x00\xEB\x8D\x88"
-     "\x00\xEB"
-     "\x8D\x89\x00\xEB\x8D\x8A\x00\xEB\x8D\x8B\x00\xEB\x8D\x8C\x00\xEB\x8D\x8D"
-     "\x00\xEB"
-     "\x8D\x8E\x00\xEB\x8D\x8F\x00\xEB\x8D\x90\x00\xEB\x8D\x91\x00\xEB\x8D\x92"
-     "\x00\xEB"
-     "\x8D\x93\x00\xEB\x8D\x94\x00\xEB\x8D\x95\x00\xEB\x8D\x96\x00\xEB\x8D\x97"
-     "\x00\xEB"
-     "\x8D\x98\x00\xEB\x8D\x99\x00\xEB\x8D\x9A\x00\xEB\x8D\x9B\x00\xEB\x8D\x9C"
-     "\x00\xEB"
-     "\x8D\x9D\x00\xEB\x8D\x9E\x00\xEB\x8D\x9F\x00\xEB\x8D\xA0\x00\xEB\x8D\xA1"
-     "\x00\xEB"
-     "\x8D\xA2\x00\xEB\x8D\xA3\x00\xEB\x8D\xA4\x00\xEB\x8D\xA5\x00\xEB\x8D\xA6"
-     "\x00\xEB"
-     "\x8D\xA7\x00\xEB\x8D\xA8\x00\xEB\x8D\xA9\x00\xEB\x8D\xAA\x00\xEB\x8D\xAB"
-     "\x00\xEB"
-     "\x8D\xAC\x00\xEB\x8D\xAD\x00\xEB\x8D\xAE\x00\xEB\x8D\xAF\x00\xEB\x8D\xB0"
-     "\x00\xEB"
-     "\x8D\xB1\x00\xEB\x8D\xB2\x00\xEB\x8D\xB3\x00\xEB\x8D\xB4\x00\xEB\x8D\xB5"
-     "\x00\xEB"
-     "\x8D\xB6\x00\xEB\x8D\xB7\x00\xEB\x8D\xB8\x00\xEB\x8D\xB9\x00\xEB\x8D\xBA"
-     "\x00\xEB"
-     "\x8D\xBB\x00\xEB\x8D\xBC\x00\xEB\x8D\xBD\x00\xEB\x8D\xBE\x00\xEB\x8D\xBF"
-     "\x00\xEB"
-     "\x8E\x80\x00\xEB\x8E\x81\x00\xEB\x8E\x82\x00\xEB\x8E\x83\x00\xEB\x8E\x84"
-     "\x00\xEB"
-     "\x8E\x85\x00\xEB\x8E\x86\x00\xEB\x8E\x87\x00\xEB\x8E\x88\x00\xEB\x8E\x89"
-     "\x00\xEB"
-     "\x8E\x8A\x00\xEB\x8E\x8B\x00\xEB\x8E\x8C\x00\xEB\x8E\x8D\x00\xEB\x8E\x8E"
-     "\x00\xEB"
-     "\x8E\x8F\x00\xEB\x8E\x90\x00\xEB\x8E\x91\x00\xEB\x8E\x92\x00\xEB\x8E\x93"
-     "\x00\xEB"
-     "\x8E\x94\x00\xEB\x8E\x95\x00\xEB\x8E\x96\x00\xEB\x8E\x97\x00\xEB\x8E\x98"
-     "\x00\xEB"
-     "\x8E\x99\x00\xEB\x8E\x9A\x00\xEB\x8E\x9B\x00\xEB\x8E\x9C\x00\xEB\x8E\x9D"
-     "\x00\xEB"
-     "\x8E\x9E\x00\xEB\x8E\x9F\x00\xEB\x8E\xA0\x00\xEB\x8E\xA1\x00\xEB\x8E\xA2"
-     "\x00\xEB"
-     "\x8E\xA3\x00\xEB\x8E\xA4\x00\xEB\x8E\xA5\x00\xEB\x8E\xA6\x00\xEB\x8E\xA7"
-     "\x00\xEB"
-     "\x8E\xA8\x00\xEB\x8E\xA9\x00\xEB\x8E\xAA\x00\xEB\x8E\xAB\x00\xEB\x8E\xAC"
-     "\x00\xEB"
-     "\x8E\xAD\x00\xEB\x8E\xAE\x00\xEB\x8E\xAF\x00\xEB\x8E\xB0\x00\xEB\x8E\xB1"
-     "\x00\xEB"
-     "\x8E\xB2\x00\xEB\x8E\xB3\x00\xEB\x8E\xB4\x00\xEB\x8E\xB5\x00\xEB\x8E\xB6"
-     "\x00\xEB"
-     "\x8E\xB7\x00\xEB\x8E\xB8\x00\xEB\x8E\xB9\x00\xEB\x8E\xBA\x00\xEB\x8E\xBB"
-     "\x00\xEB"
-     "\x8E\xBC\x00\xEB\x8E\xBD\x00\xEB\x8E\xBE\x00\xEB\x8E\xBF\x00\xEB\x8F\x80"
-     "\x00\xEB"
-     "\x8F\x81\x00\xEB\x8F\x82\x00\xEB\x8F\x83\x00\xEB\x8F\x84\x00\xEB\x8F\x85"
-     "\x00\xEB"
-     "\x8F\x86\x00\xEB\x8F\x87\x00\xEB\x8F\x88\x00\xEB\x8F\x89\x00\xEB\x8F\x8A"
-     "\x00\xEB"
-     "\x8F\x8B\x00\xEB\x8F\x8C\x00\xEB\x8F\x8D\x00\xEB\x8F\x8E\x00\xEB\x8F\x8F"
-     "\x00\xEB"
-     "\x8F\x90\x00\xEB\x8F\x91\x00\xEB\x8F\x92\x00\xEB\x8F\x93\x00\xEB\x8F\x94"
-     "\x00\xEB"
-     "\x8F\x95\x00\xEB\x8F\x96\x00\xEB\x8F\x97\x00\xEB\x8F\x98\x00\xEB\x8F\x99"
-     "\x00\xEB"
-     "\x8F\x9A\x00\xEB\x8F\x9B\x00\xEB\x8F\x9C\x00\xEB\x8F\x9D\x00\xEB\x8F\x9E"
-     "\x00\xEB"
-     "\x8F\x9F\x00\xEB\x8F\xA0\x00\xEB\x8F\xA1\x00\xEB\x8F\xA2\x00\xEB\x8F\xA3"
-     "\x00\xEB"
-     "\x8F\xA4\x00\xEB\x8F\xA5\x00\xEB\x8F\xA6\x00\xEB\x8F\xA7\x00\xEB\x8F\xA8"
-     "\x00\xEB"
-     "\x8F\xA9\x00\xEB\x8F\xAA\x00\xEB\x8F\xAB\x00\xEB\x8F\xAC\x00\xEB\x8F\xAD"
-     "\x00\xEB"
-     "\x8F\xAE\x00\xEB\x8F\xAF\x00\xEB\x8F\xB0\x00\xEB\x8F\xB1\x00\xEB\x8F\xB2"
-     "\x00\xEB"
-     "\x8F\xB3\x00\xEB\x8F\xB4\x00\xEB\x8F\xB5\x00\xEB\x8F\xB6\x00\xEB\x8F\xB7"
-     "\x00\xEB"
-     "\x8F\xB8\x00\xEB\x8F\xB9\x00\xEB\x8F\xBA\x00\xEB\x8F\xBB\x00\xEB\x8F\xBC"
-     "\x00\xEB"
-     "\x8F\xBD\x00\xEB\x8F\xBE\x00\xEB\x8F\xBF\x00\xEB\x90\x80\x00\xEB\x90\x81"
-     "\x00\xEB"
-     "\x90\x82\x00\xEB\x90\x83\x00\xEB\x90\x84\x00\xEB\x90\x85\x00\xEB\x90\x86"
-     "\x00\xEB"
-     "\x90\x87\x00\xEB\x90\x88\x00\xEB\x90\x89\x00\xEB\x90\x8A\x00\xEB\x90\x8B"
-     "\x00\xEB"
-     "\x90\x8C\x00\xEB\x90\x8D\x00\xEB\x90\x8E\x00\xEB\x90\x8F\x00\xEB\x90\x90"
-     "\x00\xEB"
-     "\x90\x91\x00\xEB\x90\x92\x00\xEB\x90\x93\x00\xEB\x90\x94\x00\xEB\x90\x95"
-     "\x00\xEB"
-     "\x90\x96\x00\xEB\x90\x97\x00\xEB\x90\x98\x00\xEB\x90\x99\x00\xEB\x90\x9A"
-     "\x00\xEB"
-     "\x90\x9B\x00\xEB\x90\x9C\x00\xEB\x90\x9D\x00\xEB\x90\x9E\x00\xEB\x90\x9F"
-     "\x00\xEB"
-     "\x90\xA0\x00\xEB\x90\xA1\x00\xEB\x90\xA2\x00\xEB\x90\xA3\x00\xEB\x90\xA4"
-     "\x00\xEB"
-     "\x90\xA5\x00\xEB\x90\xA6\x00\xEB\x90\xA7\x00\xEB\x90\xA8\x00\xEB\x90\xA9"
-     "\x00\xEB"
-     "\x90\xAA\x00\xEB\x90\xAB\x00\xEB\x90\xAC\x00\xEB\x90\xAD\x00\xEB\x90\xAE"
-     "\x00\xEB"
-     "\x90\xAF\x00\xEB\x90\xB0\x00\xEB\x90\xB1\x00\xEB\x90\xB2\x00\xEB\x90\xB3"
-     "\x00\xEB"
-     "\x90\xB4\x00\xEB\x90\xB5\x00\xEB\x90\xB6\x00\xEB\x90\xB7\x00\xEB\x90\xB8"
-     "\x00\xEB"
-     "\x90\xB9\x00\xEB\x90\xBA\x00\xEB\x90\xBB\x00\xEB\x90\xBC\x00\xEB\x90\xBD"
-     "\x00\xEB"
-     "\x90\xBE\x00\xEB\x90\xBF\x00\xEB\x91\x80\x00\xEB\x91\x81\x00\xEB\x91\x82"
-     "\x00\xEB"
-     "\x91\x83\x00\xEB\x91\x84\x00\xEB\x91\x85\x00\xEB\x91\x86\x00\xEB\x91\x87"
-     "\x00\xEB"
-     "\x91\x88\x00\xEB\x91\x89\x00\xEB\x91\x8A\x00\xEB\x91\x8B\x00\xEB\x91\x8C"
-     "\x00\xEB"
-     "\x91\x8D\x00\xEB\x91\x8E\x00\xEB\x91\x8F\x00\xEB\x91\x90\x00\xEB\x91\x91"
-     "\x00\xEB"
-     "\x91\x92\x00\xEB\x91\x93\x00\xEB\x91\x94\x00\xEB\x91\x95\x00\xEB\x91\x96"
-     "\x00\xEB"
-     "\x91\x97\x00\xEB\x91\x98\x00\xEB\x91\x99\x00\xEB\x91\x9A\x00\xEB\x91\x9B"
-     "\x00\xEB"
-     "\x91\x9C\x00\xEB\x91\x9D\x00\xEB\x91\x9E\x00\xEB\x91\x9F\x00\xEB\x91\xA0"
-     "\x00\xEB"
-     "\x91\xA1\x00\xEB\x91\xA2\x00\xEB\x91\xA3\x00\xEB\x91\xA4\x00\xEB\x91\xA5"
-     "\x00\xEB"
-     "\x91\xA6\x00\xEB\x91\xA7\x00\xEB\x91\xA8\x00\xEB\x91\xA9\x00\xEB\x91\xAA"
-     "\x00\xEB"
-     "\x91\xAB\x00\xEB\x91\xAC\x00\xEB\x91\xAD\x00\xEB\x91\xAE\x00\xEB\x91\xAF"
-     "\x00\xEB"
-     "\x91\xB0\x00\xEB\x91\xB1\x00\xEB\x91\xB2\x00\xEB\x91\xB3\x00\xEB\x91\xB4"
-     "\x00\xEB"
-     "\x91\xB5\x00\xEB\x91\xB6\x00\xEB\x91\xB7\x00\xEB\x91\xB8\x00\xEB\x91\xB9"
-     "\x00\xEB"
-     "\x91\xBA\x00\xEB\x91\xBB\x00\xEB\x91\xBC\x00\xEB\x91\xBD\x00\xEB\x91\xBE"
-     "\x00\xEB"
-     "\x91\xBF\x00\xEB\x92\x80\x00\xEB\x92\x81\x00\xEB\x92\x82\x00\xEB\x92\x83"
-     "\x00\xEB"
-     "\x92\x84\x00\xEB\x92\x85\x00\xEB\x92\x86\x00\xEB\x92\x87\x00\xEB\x92\x88"
-     "\x00\xEB"
-     "\x92\x89\x00\xEB\x92\x8A\x00\xEB\x92\x8B\x00\xEB\x92\x8C\x00\xEB\x92\x8D"
-     "\x00\xEB"
-     "\x92\x8E\x00\xEB\x92\x8F\x00\xEB\x92\x90\x00\xEB\x92\x91\x00\xEB\x92\x92"
-     "\x00\xEB"
-     "\x92\x93\x00\xEB\x92\x94\x00\xEB\x92\x95\x00\xEB\x92\x96\x00\xEB\x92\x97"
-     "\x00\xEB"
-     "\x92\x98\x00\xEB\x92\x99\x00\xEB\x92\x9A\x00\xEB\x92\x9B\x00\xEB\x92\x9C"
-     "\x00\xEB"
-     "\x92\x9D\x00\xEB\x92\x9E\x00\xEB\x92\x9F\x00\xEB\x92\xA0\x00\xEB\x92\xA1"
-     "\x00\xEB"
-     "\x92\xA2\x00\xEB\x92\xA3\x00\xEB\x92\xA4\x00\xEB\x92\xA5\x00\xEB\x92\xA6"
-     "\x00\xEB"
-     "\x92\xA7\x00\xEB\x92\xA8\x00\xEB\x92\xA9\x00\xEB\x92\xAA\x00\xEB\x92\xAB"
-     "\x00\xEB"
-     "\x92\xAC\x00\xEB\x92\xAD\x00\xEB\x92\xAE\x00\xEB\x92\xAF\x00\xEB\x92\xB0"
-     "\x00\xEB"
-     "\x92\xB1\x00\xEB\x92\xB2\x00\xEB\x92\xB3\x00\xEB\x92\xB4\x00\xEB\x92\xB5"
-     "\x00\xEB"
-     "\x92\xB6\x00\xEB\x92\xB7\x00\xEB\x92\xB8\x00\xEB\x92\xB9\x00\xEB\x92\xBA"
-     "\x00\xEB"
-     "\x92\xBB\x00\xEB\x92\xBC\x00\xEB\x92\xBD\x00\xEB\x92\xBE\x00\xEB\x92\xBF"
-     "\x00\xEB"
-     "\x93\x80\x00\xEB\x93\x81\x00\xEB\x93\x82\x00\xEB\x93\x83\x00\xEB\x93\x84"
-     "\x00\xEB"
-     "\x93\x85\x00\xEB\x93\x86\x00\xEB\x93\x87\x00\xEB\x93\x88\x00\xEB\x93\x89"
-     "\x00\xEB"
-     "\x93\x8A\x00\xEB\x93\x8B\x00\xEB\x93\x8C\x00\xEB\x93\x8D\x00\xEB\x93\x8E"
-     "\x00\xEB"
-     "\x93\x8F\x00\xEB\x93\x90\x00\xEB\x93\x91\x00\xEB\x93\x92\x00\xEB\x93\x93"
-     "\x00\xEB"
-     "\x93\x94\x00\xEB\x93\x95\x00\xEB\x93\x96\x00\xEB\x93\x97\x00\xEB\x93\x98"
-     "\x00\xEB"
-     "\x93\x99\x00\xEB\x93\x9A\x00\xEB\x93\x9B\x00\xEB\x93\x9C\x00\xEB\x93\x9D"
-     "\x00\xEB"
-     "\x93\x9E\x00\xEB\x93\x9F\x00\xEB\x93\xA0\x00\xEB\x93\xA1\x00\xEB\x93\xA2"
-     "\x00\xEB"
-     "\x93\xA3\x00\xEB\x93\xA4\x00\xEB\x93\xA5\x00\xEB\x93\xA6\x00\xEB\x93\xA7"
-     "\x00\xEB"
-     "\x93\xA8\x00\xEB\x93\xA9\x00\xEB\x93\xAA\x00\xEB\x93\xAB\x00\xEB\x93\xAC"
-     "\x00\xEB"
-     "\x93\xAD\x00\xEB\x93\xAE\x00\xEB\x93\xAF\x00\xEB\x93\xB0\x00\xEB\x93\xB1"
-     "\x00\xEB"
-     "\x93\xB2\x00\xEB\x93\xB3\x00\xEB\x93\xB4\x00\xEB\x93\xB5\x00\xEB\x93\xB6"
-     "\x00\xEB"
-     "\x93\xB7\x00\xEB\x93\xB8\x00\xEB\x93\xB9\x00\xEB\x93\xBA\x00\xEB\x93\xBB"
-     "\x00\xEB"
-     "\x93\xBC\x00\xEB\x93\xBD\x00\xEB\x93\xBE\x00\xEB\x93\xBF\x00\xEB\x94\x80"
-     "\x00\xEB"
-     "\x94\x81\x00\xEB\x94\x82\x00\xEB\x94\x83\x00\xEB\x94\x84\x00\xEB\x94\x85"
-     "\x00\xEB"
-     "\x94\x86\x00\xEB\x94\x87\x00\xEB\x94\x88\x00\xEB\x94\x89\x00\xEB\x94\x8A"
-     "\x00\xEB"
-     "\x94\x8B\x00\xEB\x94\x8C\x00\xEB\x94\x8D\x00\xEB\x94\x8E\x00\xEB\x94\x8F"
-     "\x00\xEB"
-     "\x94\x90\x00\xEB\x94\x91\x00\xEB\x94\x92\x00\xEB\x94\x93\x00\xEB\x94\x94"
-     "\x00\xEB"
-     "\x94\x95\x00\xEB\x94\x96\x00\xEB\x94\x97\x00\xEB\x94\x98\x00\xEB\x94\x99"
-     "\x00\xEB"
-     "\x94\x9A\x00\xEB\x94\x9B\x00\xEB\x94\x9C\x00\xEB\x94\x9D\x00\xEB\x94\x9E"
-     "\x00\xEB"
-     "\x94\x9F\x00\xEB\x94\xA0\x00\xEB\x94\xA1\x00\xEB\x94\xA2\x00\xEB\x94\xA3"
-     "\x00\xEB"
-     "\x94\xA4\x00\xEB\x94\xA5\x00\xEB\x94\xA6\x00\xEB\x94\xA7\x00\xEB\x94\xA8"
-     "\x00\xEB"
-     "\x94\xA9\x00\xEB\x94\xAA\x00\xEB\x94\xAB\x00\xEB\x94\xAC\x00\xEB\x94\xAD"
-     "\x00\xEB"
-     "\x94\xAE\x00\xEB\x94\xAF\x00\xEB\x94\xB0\x00\xEB\x94\xB1\x00\xEB\x94\xB2"
-     "\x00\xEB"
-     "\x94\xB3\x00\xEB\x94\xB4\x00\xEB\x94\xB5\x00\xEB\x94\xB6\x00\xEB\x94\xB7"
-     "\x00\xEB"
-     "\x94\xB8\x00\xEB\x94\xB9\x00\xEB\x94\xBA\x00\xEB\x94\xBB\x00\xEB\x94\xBC"
-     "\x00\xEB"
-     "\x94\xBD\x00\xEB\x94\xBE\x00\xEB\x94\xBF\x00\xEB\x95\x80\x00\xEB\x95\x81"
-     "\x00\xEB"
-     "\x95\x82\x00\xEB\x95\x83\x00\xEB\x95\x84\x00\xEB\x95\x85\x00\xEB\x95\x86"
-     "\x00\xEB"
-     "\x95\x87\x00\xEB\x95\x88\x00\xEB\x95\x89\x00\xEB\x95\x8A\x00\xEB\x95\x8B"
-     "\x00\xEB"
-     "\x95\x8C\x00\xEB\x95\x8D\x00\xEB\x95\x8E\x00\xEB\x95\x8F\x00\xEB\x95\x90"
-     "\x00\xEB"
-     "\x95\x91\x00\xEB\x95\x92\x00\xEB\x95\x93\x00\xEB\x95\x94\x00\xEB\x95\x95"
-     "\x00\xEB"
-     "\x95\x96\x00\xEB\x95\x97\x00\xEB\x95\x98\x00\xEB\x95\x99\x00\xEB\x95\x9A"
-     "\x00\xEB"
-     "\x95\x9B\x00\xEB\x95\x9C\x00\xEB\x95\x9D\x00\xEB\x95\x9E\x00\xEB\x95\x9F"
-     "\x00\xEB"
-     "\x95\xA0\x00\xEB\x95\xA1\x00\xEB\x95\xA2\x00\xEB\x95\xA3\x00\xEB\x95\xA4"
-     "\x00\xEB"
-     "\x95\xA5\x00\xEB\x95\xA6\x00\xEB\x95\xA7\x00\xEB\x95\xA8\x00\xEB\x95\xA9"
-     "\x00\xEB"
-     "\x95\xAA\x00\xEB\x95\xAB\x00\xEB\x95\xAC\x00\xEB\x95\xAD\x00\xEB\x95\xAE"
-     "\x00\xEB"
-     "\x95\xAF\x00\xEB\x95\xB0\x00\xEB\x95\xB1\x00\xEB\x95\xB2\x00\xEB\x95\xB3"
-     "\x00\xEB"
-     "\x95\xB4\x00\xEB\x95\xB5\x00\xEB\x95\xB6\x00\xEB\x95\xB7\x00\xEB\x95\xB8"
-     "\x00\xEB"
-     "\x95\xB9\x00\xEB\x95\xBA\x00\xEB\x95\xBB\x00\xEB\x95\xBC\x00\xEB\x95\xBD"
-     "\x00\xEB"
-     "\x95\xBE\x00\xEB\x95\xBF\x00\xEB\x96\x80\x00\xEB\x96\x81\x00\xEB\x96\x82"
-     "\x00\xEB"
-     "\x96\x83\x00\xEB\x96\x84\x00\xEB\x96\x85\x00\xEB\x96\x86\x00\xEB\x96\x87"
-     "\x00\xEB"
-     "\x96\x88\x00\xEB\x96\x89\x00\xEB\x96\x8A\x00\xEB\x96\x8B\x00\xEB\x96\x8C"
-     "\x00\xEB"
-     "\x96\x8D\x00\xEB\x96\x8E\x00\xEB\x96\x8F\x00\xEB\x96\x90\x00\xEB\x96\x91"
-     "\x00\xEB"
-     "\x96\x92\x00\xEB\x96\x93\x00\xEB\x96\x94\x00\xEB\x96\x95\x00\xEB\x96\x96"
-     "\x00\xEB"
-     "\x96\x97\x00\xEB\x96\x98\x00\xEB\x96\x99\x00\xEB\x96\x9A\x00\xEB\x96\x9B"
-     "\x00\xEB"
-     "\x96\x9C\x00\xEB\x96\x9D\x00\xEB\x96\x9E\x00\xEB\x96\x9F\x00\xEB\x96\xA0"
-     "\x00\xEB"
-     "\x96\xA1\x00\xEB\x96\xA2\x00\xEB\x96\xA3\x00\xEB\x96\xA4\x00\xEB\x96\xA5"
-     "\x00\xEB"
-     "\x96\xA6\x00\xEB\x96\xA7\x00\xEB\x96\xA8\x00\xEB\x96\xA9\x00\xEB\x96\xAA"
-     "\x00\xEB"
-     "\x96\xAB\x00\xEB\x96\xAC\x00\xEB\x96\xAD\x00\xEB\x96\xAE\x00\xEB\x96\xAF"
-     "\x00\xEB"
-     "\x96\xB0\x00\xEB\x96\xB1\x00\xEB\x96\xB2\x00\xEB\x96\xB3\x00\xEB\x96\xB4"
-     "\x00\xEB"
-     "\x96\xB5\x00\xEB\x96\xB6\x00\xEB\x96\xB7\x00\xEB\x96\xB8\x00\xEB\x96\xB9"
-     "\x00\xEB"
-     "\x96\xBA\x00\xEB\x96\xBB\x00\xEB\x96\xBC\x00\xEB\x96\xBD\x00\xEB\x96\xBE"
-     "\x00\xEB"
-     "\x96\xBF\x00\xEB\x97\x80\x00\xEB\x97\x81\x00\xEB\x97\x82\x00\xEB\x97\x83"
-     "\x00\xEB"
-     "\x97\x84\x00\xEB\x97\x85\x00\xEB\x97\x86\x00\xEB\x97\x87\x00\xEB\x97\x88"
-     "\x00\xEB"
-     "\x97\x89\x00\xEB\x97\x8A\x00\xEB\x97\x8B\x00\xEB\x97\x8C\x00\xEB\x97\x8D"
-     "\x00\xEB"
-     "\x97\x8E\x00\xEB\x97\x8F\x00\xEB\x97\x90\x00\xEB\x97\x91\x00\xEB\x97\x92"
-     "\x00\xEB"
-     "\x97\x93\x00\xEB\x97\x94\x00\xEB\x97\x95\x00\xEB\x97\x96\x00\xEB\x97\x97"
-     "\x00\xEB"
-     "\x97\x98\x00\xEB\x97\x99\x00\xEB\x97\x9A\x00\xEB\x97\x9B\x00\xEB\x97\x9C"
-     "\x00\xEB"
-     "\x97\x9D\x00\xEB\x97\x9E\x00\xEB\x97\x9F\x00\xEB\x97\xA0\x00\xEB\x97\xA1"
-     "\x00\xEB"
-     "\x97\xA2\x00\xEB\x97\xA3\x00\xEB\x97\xA4\x00\xEB\x97\xA5\x00\xEB\x97\xA6"
-     "\x00\xEB"
-     "\x97\xA7\x00\xEB\x97\xA8\x00\xEB\x97\xA9\x00\xEB\x97\xAA\x00\xEB\x97\xAB"
-     "\x00\xEB"
-     "\x97\xAC\x00\xEB\x97\xAD\x00\xEB\x97\xAE\x00\xEB\x97\xAF\x00\xEB\x97\xB0"
-     "\x00\xEB"
-     "\x97\xB1\x00\xEB\x97\xB2\x00\xEB\x97\xB3\x00\xEB\x97\xB4\x00\xEB\x97\xB5"
-     "\x00\xEB"
-     "\x97\xB6\x00\xEB\x97\xB7\x00\xEB\x97\xB8\x00\xEB\x97\xB9\x00\xEB\x97\xBA"
-     "\x00\xEB"
-     "\x97\xBB\x00\xEB\x97\xBC\x00\xEB\x97\xBD\x00\xEB\x97\xBE\x00\xEB\x97\xBF"
-     "\x00\xEB"
-     "\x98\x80\x00\xEB\x98\x81\x00\xEB\x98\x82\x00\xEB\x98\x83\x00\xEB\x98\x84"
-     "\x00\xEB"
-     "\x98\x85\x00\xEB\x98\x86\x00\xEB\x98\x87\x00\xEB\x98\x88\x00\xEB\x98\x89"
-     "\x00\xEB"
-     "\x98\x8A\x00\xEB\x98\x8B\x00\xEB\x98\x8C\x00\xEB\x98\x8D\x00\xEB\x98\x8E"
-     "\x00\xEB"
-     "\x98\x8F\x00\xEB\x98\x90\x00\xEB\x98\x91\x00\xEB\x98\x92\x00\xEB\x98\x93"
-     "\x00\xEB"
-     "\x98\x94\x00\xEB\x98\x95\x00\xEB\x98\x96\x00\xEB\x98\x97\x00\xEB\x98\x98"
-     "\x00\xEB"
-     "\x98\x99\x00\xEB\x98\x9A\x00\xEB\x98\x9B\x00\xEB\x98\x9C\x00\xEB\x98\x9D"
-     "\x00\xEB"
-     "\x98\x9E\x00\xEB\x98\x9F\x00\xEB\x98\xA0\x00\xEB\x98\xA1\x00\xEB\x98\xA2"
-     "\x00\xEB"
-     "\x98\xA3\x00\xEB\x98\xA4\x00\xEB\x98\xA5\x00\xEB\x98\xA6\x00\xEB\x98\xA7"
-     "\x00\xEB"
-     "\x98\xA8\x00\xEB\x98\xA9\x00\xEB\x98\xAA\x00\xEB\x98\xAB\x00\xEB\x98\xAC"
-     "\x00\xEB"
-     "\x98\xAD\x00\xEB\x98\xAE\x00\xEB\x98\xAF\x00\xEB\x98\xB0\x00\xEB\x98\xB1"
-     "\x00\xEB"
-     "\x98\xB2\x00\xEB\x98\xB3\x00\xEB\x98\xB4\x00\xEB\x98\xB5\x00\xEB\x98\xB6"
-     "\x00\xEB"
-     "\x98\xB7\x00\xEB\x98\xB8\x00\xEB\x98\xB9\x00\xEB\x98\xBA\x00\xEB\x98\xBB"
-     "\x00\xEB"
-     "\x98\xBC\x00\xEB\x98\xBD\x00\xEB\x98\xBE\x00\xEB\x98\xBF\x00\xEB\x99\x80"
-     "\x00\xEB"
-     "\x99\x81\x00\xEB\x99\x82\x00\xEB\x99\x83\x00\xEB\x99\x84\x00\xEB\x99\x85"
-     "\x00\xEB"
-     "\x99\x86\x00\xEB\x99\x87\x00\xEB\x99\x88\x00\xEB\x99\x89\x00\xEB\x99\x8A"
-     "\x00\xEB"
-     "\x99\x8B\x00\xEB\x99\x8C\x00\xEB\x99\x8D\x00\xEB\x99\x8E\x00\xEB\x99\x8F"
-     "\x00\xEB"
-     "\x99\x90\x00\xEB\x99\x91\x00\xEB\x99\x92\x00\xEB\x99\x93\x00\xEB\x99\x94"
-     "\x00\xEB"
-     "\x99\x95\x00\xEB\x99\x96\x00\xEB\x99\x97\x00\xEB\x99\x98\x00\xEB\x99\x99"
-     "\x00\xEB"
-     "\x99\x9A\x00\xEB\x99\x9B\x00\xEB\x99\x9C\x00\xEB\x99\x9D\x00\xEB\x99\x9E"
-     "\x00\xEB"
-     "\x99\x9F\x00\xEB\x99\xA0\x00\xEB\x99\xA1\x00\xEB\x99\xA2\x00\xEB\x99\xA3"
-     "\x00\xEB"
-     "\x99\xA4\x00\xEB\x99\xA5\x00\xEB\x99\xA6\x00\xEB\x99\xA7\x00\xEB\x99\xA8"
-     "\x00\xEB"
-     "\x99\xA9\x00\xEB\x99\xAA\x00\xEB\x99\xAB\x00\xEB\x99\xAC\x00\xEB\x99\xAD"
-     "\x00\xEB"
-     "\x99\xAE\x00\xEB\x99\xAF\x00\xEB\x99\xB0\x00\xEB\x99\xB1\x00\xEB\x99\xB2"
-     "\x00\xEB"
-     "\x99\xB3\x00\xEB\x99\xB4\x00\xEB\x99\xB5\x00\xEB\x99\xB6\x00\xEB\x99\xB7"
-     "\x00\xEB"
-     "\x99\xB8\x00\xEB\x99\xB9\x00\xEB\x99\xBA\x00\xEB\x99\xBB\x00\xEB\x99\xBC"
-     "\x00\xEB"
-     "\x99\xBD\x00\xEB\x99\xBE\x00\xEB\x99\xBF\x00\xEB\x9A\x80\x00\xEB\x9A\x81"
-     "\x00\xEB"
-     "\x9A\x82\x00\xEB\x9A\x83\x00\xEB\x9A\x84\x00\xEB\x9A\x85\x00\xEB\x9A\x86"
-     "\x00\xEB"
-     "\x9A\x87\x00\xEB\x9A\x88\x00\xEB\x9A\x89\x00\xEB\x9A\x8A\x00\xEB\x9A\x8B"
-     "\x00\xEB"
-     "\x9A\x8C\x00\xEB\x9A\x8D\x00\xEB\x9A\x8E\x00\xEB\x9A\x8F\x00\xEB\x9A\x90"
-     "\x00\xEB"
-     "\x9A\x91\x00\xEB\x9A\x92\x00\xEB\x9A\x93\x00\xEB\x9A\x94\x00\xEB\x9A\x95"
-     "\x00\xEB"
-     "\x9A\x96\x00\xEB\x9A\x97\x00\xEB\x9A\x98\x00\xEB\x9A\x99\x00\xEB\x9A\x9A"
-     "\x00\xEB"
-     "\x9A\x9B\x00\xEB\x9A\x9C\x00\xEB\x9A\x9D\x00\xEB\x9A\x9E\x00\xEB\x9A\x9F"
-     "\x00\xEB"
-     "\x9A\xA0\x00\xEB\x9A\xA1\x00\xEB\x9A\xA2\x00\xEB\x9A\xA3\x00\xEB\x9A\xA4"
-     "\x00\xEB"
-     "\x9A\xA5\x00\xEB\x9A\xA6\x00\xEB\x9A\xA7\x00\xEB\x9A\xA8\x00\xEB\x9A\xA9"
-     "\x00\xEB"
-     "\x9A\xAA\x00\xEB\x9A\xAB\x00\xEB\x9A\xAC\x00\xEB\x9A\xAD\x00\xEB\x9A\xAE"
-     "\x00\xEB"
-     "\x9A\xAF\x00\xEB\x9A\xB0\x00\xEB\x9A\xB1\x00\xEB\x9A\xB2\x00\xEB\x9A\xB3"
-     "\x00\xEB"
-     "\x9A\xB4\x00\xEB\x9A\xB5\x00\xEB\x9A\xB6\x00\xEB\x9A\xB7\x00\xEB\x9A\xB8"
-     "\x00\xEB"
-     "\x9A\xB9\x00\xEB\x9A\xBA\x00\xEB\x9A\xBB\x00\xEB\x9A\xBC\x00\xEB\x9A\xBD"
-     "\x00\xEB"
-     "\x9A\xBE\x00\xEB\x9A\xBF\x00\xEB\x9B\x80\x00\xEB\x9B\x81\x00\xEB\x9B\x82"
-     "\x00\xEB"
-     "\x9B\x83\x00\xEB\x9B\x84\x00\xEB\x9B\x85\x00\xEB\x9B\x86\x00\xEB\x9B\x87"
-     "\x00\xEB"
-     "\x9B\x88\x00\xEB\x9B\x89\x00\xEB\x9B\x8A\x00\xEB\x9B\x8B\x00\xEB\x9B\x8C"
-     "\x00\xEB"
-     "\x9B\x8D\x00\xEB\x9B\x8E\x00\xEB\x9B\x8F\x00\xEB\x9B\x90\x00\xEB\x9B\x91"
-     "\x00\xEB"
-     "\x9B\x92\x00\xEB\x9B\x93\x00\xEB\x9B\x94\x00\xEB\x9B\x95\x00\xEB\x9B\x96"
-     "\x00\xEB"
-     "\x9B\x97\x00\xEB\x9B\x98\x00\xEB\x9B\x99\x00\xEB\x9B\x9A\x00\xEB\x9B\x9B"
-     "\x00\xEB"
-     "\x9B\x9C\x00\xEB\x9B\x9D\x00\xEB\x9B\x9E\x00\xEB\x9B\x9F\x00\xEB\x9B\xA0"
-     "\x00\xEB"
-     "\x9B\xA1\x00\xEB\x9B\xA2\x00\xEB\x9B\xA3\x00\xEB\x9B\xA4\x00\xEB\x9B\xA5"
-     "\x00\xEB"
-     "\x9B\xA6\x00\xEB\x9B\xA7\x00\xEB\x9B\xA8\x00\xEB\x9B\xA9\x00\xEB\x9B\xAA"
-     "\x00\xEB"
-     "\x9B\xAB\x00\xEB\x9B\xAC\x00\xEB\x9B\xAD\x00\xEB\x9B\xAE\x00\xEB\x9B\xAF"
-     "\x00\xEB"
-     "\x9B\xB0\x00\xEB\x9B\xB1\x00\xEB\x9B\xB2\x00\xEB\x9B\xB3\x00\xEB\x9B\xB4"
-     "\x00\xEB"
-     "\x9B\xB5\x00\xEB\x9B\xB6\x00\xEB\x9B\xB7\x00\xEB\x9B\xB8\x00\xEB\x9B\xB9"
-     "\x00\xEB"
-     "\x9B\xBA\x00\xEB\x9B\xBB\x00\xEB\x9B\xBC\x00\xEB\x9B\xBD\x00\xEB\x9B\xBE"
-     "\x00\xEB"
-     "\x9B\xBF\x00\xEB\x9C\x80\x00\xEB\x9C\x81\x00\xEB\x9C\x82\x00\xEB\x9C\x83"
-     "\x00\xEB"
-     "\x9C\x84\x00\xEB\x9C\x85\x00\xEB\x9C\x86\x00\xEB\x9C\x87\x00\xEB\x9C\x88"
-     "\x00\xEB"
-     "\x9C\x89\x00\xEB\x9C\x8A\x00\xEB\x9C\x8B\x00\xEB\x9C\x8C\x00\xEB\x9C\x8D"
-     "\x00\xEB"
-     "\x9C\x8E\x00\xEB\x9C\x8F\x00\xEB\x9C\x90\x00\xEB\x9C\x91\x00\xEB\x9C\x92"
-     "\x00\xEB"
-     "\x9C\x93\x00\xEB\x9C\x94\x00\xEB\x9C\x95\x00\xEB\x9C\x96\x00\xEB\x9C\x97"
-     "\x00\xEB"
-     "\x9C\x98\x00\xEB\x9C\x99\x00\xEB\x9C\x9A\x00\xEB\x9C\x9B\x00\xEB\x9C\x9C"
-     "\x00\xEB"
-     "\x9C\x9D\x00\xEB\x9C\x9E\x00\xEB\x9C\x9F\x00\xEB\x9C\xA0\x00\xEB\x9C\xA1"
-     "\x00\xEB"
-     "\x9C\xA2\x00\xEB\x9C\xA3\x00\xEB\x9C\xA4\x00\xEB\x9C\xA5\x00\xEB\x9C\xA6"
-     "\x00\xEB"
-     "\x9C\xA7\x00\xEB\x9C\xA8\x00\xEB\x9C\xA9\x00\xEB\x9C\xAA\x00\xEB\x9C\xAB"
-     "\x00\xEB"
-     "\x9C\xAC\x00\xEB\x9C\xAD\x00\xEB\x9C\xAE\x00\xEB\x9C\xAF\x00\xEB\x9C\xB0"
-     "\x00\xEB"
-     "\x9C\xB1\x00\xEB\x9C\xB2\x00\xEB\x9C\xB3\x00\xEB\x9C\xB4\x00\xEB\x9C\xB5"
-     "\x00\xEB"
-     "\x9C\xB6\x00\xEB\x9C\xB7\x00\xEB\x9C\xB8\x00\xEB\x9C\xB9\x00\xEB\x9C\xBA"
-     "\x00\xEB"
-     "\x9C\xBB\x00\xEB\x9C\xBC\x00\xEB\x9C\xBD\x00\xEB\x9C\xBE\x00\xEB\x9C\xBF"
-     "\x00\xEB"
-     "\x9D\x80\x00\xEB\x9D\x81\x00\xEB\x9D\x82\x00\xEB\x9D\x83\x00\xEB\x9D\x84"
-     "\x00\xEB"
-     "\x9D\x85\x00\xEB\x9D\x86\x00\xEB\x9D\x87\x00\xEB\x9D\x88\x00\xEB\x9D\x89"
-     "\x00\xEB"
-     "\x9D\x8A\x00\xEB\x9D\x8B\x00\xEB\x9D\x8C\x00\xEB\x9D\x8D\x00\xEB\x9D\x8E"
-     "\x00\xEB"
-     "\x9D\x8F\x00\xEB\x9D\x90\x00\xEB\x9D\x91\x00\xEB\x9D\x92\x00\xEB\x9D\x93"
-     "\x00\xEB"
-     "\x9D\x94\x00\xEB\x9D\x95\x00\xEB\x9D\x96\x00\xEB\x9D\x97\x00\xEB\x9D\x98"
-     "\x00\xEB"
-     "\x9D\x99\x00\xEB\x9D\x9A\x00\xEB\x9D\x9B\x00\xEB\x9D\x9C\x00\xEB\x9D\x9D"
-     "\x00\xEB"
-     "\x9D\x9E\x00\xEB\x9D\x9F\x00\xEB\x9D\xA0\x00\xEB\x9D\xA1\x00\xEB\x9D\xA2"
-     "\x00\xEB"
-     "\x9D\xA3\x00\xEB\x9D\xA4\x00\xEB\x9D\xA5\x00\xEB\x9D\xA6\x00\xEB\x9D\xA7"
-     "\x00\xEB"
-     "\x9D\xA8\x00\xEB\x9D\xA9\x00\xEB\x9D\xAA\x00\xEB\x9D\xAB\x00\xEB\x9D\xAC"
-     "\x00\xEB"
-     "\x9D\xAD\x00\xEB\x9D\xAE\x00\xEB\x9D\xAF\x00\xEB\x9D\xB0\x00\xEB\x9D\xB1"
-     "\x00\xEB"
-     "\x9D\xB2\x00\xEB\x9D\xB3\x00\xEB\x9D\xB4\x00\xEB\x9D\xB5\x00\xEB\x9D\xB6"
-     "\x00\xEB"
-     "\x9D\xB7\x00\xEB\x9D\xB8\x00\xEB\x9D\xB9\x00\xEB\x9D\xBA\x00\xEB\x9D\xBB"
-     "\x00\xEB"
-     "\x9D\xBC\x00\xEB\x9D\xBD\x00\xEB\x9D\xBE\x00\xEB\x9D\xBF\x00\xEB\x9E\x80"
-     "\x00\xEB"
-     "\x9E\x81\x00\xEB\x9E\x82\x00\xEB\x9E\x83\x00\xEB\x9E\x84\x00\xEB\x9E\x85"
-     "\x00\xEB"
-     "\x9E\x86\x00\xEB\x9E\x87\x00\xEB\x9E\x88\x00\xEB\x9E\x89\x00\xEB\x9E\x8A"
-     "\x00\xEB"
-     "\x9E\x8B\x00\xEB\x9E\x8C\x00\xEB\x9E\x8D\x00\xEB\x9E\x8E\x00\xEB\x9E\x8F"
-     "\x00\xEB"
-     "\x9E\x90\x00\xEB\x9E\x91\x00\xEB\x9E\x92\x00\xEB\x9E\x93\x00\xEB\x9E\x94"
-     "\x00\xEB"
-     "\x9E\x95\x00\xEB\x9E\x96\x00\xEB\x9E\x97\x00\xEB\x9E\x98\x00\xEB\x9E\x99"
-     "\x00\xEB"
-     "\x9E\x9A\x00\xEB\x9E\x9B\x00\xEB\x9E\x9C\x00\xEB\x9E\x9D\x00\xEB\x9E\x9E"
-     "\x00\xEB"
-     "\x9E\x9F\x00\xEB\x9E\xA0\x00\xEB\x9E\xA1\x00\xEB\x9E\xA2\x00\xEB\x9E\xA3"
-     "\x00\xEB"
-     "\x9E\xA4\x00\xEB\x9E\xA5\x00\xEB\x9E\xA6\x00\xEB\x9E\xA7\x00\xEB\x9E\xA8"
-     "\x00\xEB"
-     "\x9E\xA9\x00\xEB\x9E\xAA\x00\xEB\x9E\xAB\x00\xEB\x9E\xAC\x00\xEB\x9E\xAD"
-     "\x00\xEB"
-     "\x9E\xAE\x00\xEB\x9E\xAF\x00\xEB\x9E\xB0\x00\xEB\x9E\xB1\x00\xEB\x9E\xB2"
-     "\x00\xEB"
-     "\x9E\xB3\x00\xEB\x9E\xB4\x00\xEB\x9E\xB5\x00\xEB\x9E\xB6\x00\xEB\x9E\xB7"
-     "\x00\xEB"
-     "\x9E\xB8\x00\xEB\x9E\xB9\x00\xEB\x9E\xBA\x00\xEB\x9E\xBB\x00\xEB\x9E\xBC"
-     "\x00\xEB"
-     "\x9E\xBD\x00\xEB\x9E\xBE\x00\xEB\x9E\xBF\x00\xEB\x9F\x80\x00\xEB\x9F\x81"
-     "\x00\xEB"
-     "\x9F\x82\x00\xEB\x9F\x83\x00\xEB\x9F\x84\x00\xEB\x9F\x85\x00\xEB\x9F\x86"
-     "\x00\xEB"
-     "\x9F\x87\x00\xEB\x9F\x88\x00\xEB\x9F\x89\x00\xEB\x9F\x8A\x00\xEB\x9F\x8B"
-     "\x00\xEB"
-     "\x9F\x8C\x00\xEB\x9F\x8D\x00\xEB\x9F\x8E\x00\xEB\x9F\x8F\x00\xEB\x9F\x90"
-     "\x00\xEB"
-     "\x9F\x91\x00\xEB\x9F\x92\x00\xEB\x9F\x93\x00\xEB\x9F\x94\x00\xEB\x9F\x95"
-     "\x00\xEB"
-     "\x9F\x96\x00\xEB\x9F\x97\x00\xEB\x9F\x98\x00\xEB\x9F\x99\x00\xEB\x9F\x9A"
-     "\x00\xEB"
-     "\x9F\x9B\x00\xEB\x9F\x9C\x00\xEB\x9F\x9D\x00\xEB\x9F\x9E\x00\xEB\x9F\x9F"
-     "\x00\xEB"
-     "\x9F\xA0\x00\xEB\x9F\xA1\x00\xEB\x9F\xA2\x00\xEB\x9F\xA3\x00\xEB\x9F\xA4"
-     "\x00\xEB"
-     "\x9F\xA5\x00\xEB\x9F\xA6\x00\xEB\x9F\xA7\x00\xEB\x9F\xA8\x00\xEB\x9F\xA9"
-     "\x00\xEB"
-     "\x9F\xAA\x00\xEB\x9F\xAB\x00\xEB\x9F\xAC\x00\xEB\x9F\xAD\x00\xEB\x9F\xAE"
-     "\x00\xEB"
-     "\x9F\xAF\x00\xEB\x9F\xB0\x00\xEB\x9F\xB1\x00\xEB\x9F\xB2\x00\xEB\x9F\xB3"
-     "\x00\xEB"
-     "\x9F\xB4\x00\xEB\x9F\xB5\x00\xEB\x9F\xB6\x00\xEB\x9F\xB7\x00\xEB\x9F\xB8"
-     "\x00\xEB"
-     "\x9F\xB9\x00\xEB\x9F\xBA\x00\xEB\x9F\xBB\x00\xEB\x9F\xBC\x00\xEB\x9F\xBD"
-     "\x00\xEB"
-     "\x9F\xBE\x00\xEB\x9F\xBF\x00\xEB\xA0\x80\x00\xEB\xA0\x81\x00\xEB\xA0\x82"
-     "\x00\xEB"
-     "\xA0\x83\x00\xEB\xA0\x84\x00\xEB\xA0\x85\x00\xEB\xA0\x86\x00\xEB\xA0\x87"
-     "\x00\xEB"
-     "\xA0\x88\x00\xEB\xA0\x89\x00\xEB\xA0\x8A\x00\xEB\xA0\x8B\x00\xEB\xA0\x8C"
-     "\x00\xEB"
-     "\xA0\x8D\x00\xEB\xA0\x8E\x00\xEB\xA0\x8F\x00\xEB\xA0\x90\x00\xEB\xA0\x91"
-     "\x00\xEB"
-     "\xA0\x92\x00\xEB\xA0\x93\x00\xEB\xA0\x94\x00\xEB\xA0\x95\x00\xEB\xA0\x96"
-     "\x00\xEB"
-     "\xA0\x97\x00\xEB\xA0\x98\x00\xEB\xA0\x99\x00\xEB\xA0\x9A\x00\xEB\xA0\x9B"
-     "\x00\xEB"
-     "\xA0\x9C\x00\xEB\xA0\x9D\x00\xEB\xA0\x9E\x00\xEB\xA0\x9F\x00\xEB\xA0\xA0"
-     "\x00\xEB"
-     "\xA0\xA1\x00\xEB\xA0\xA2\x00\xEB\xA0\xA3\x00\xEB\xA0\xA4\x00\xEB\xA0\xA5"
-     "\x00\xEB"
-     "\xA0\xA6\x00\xEB\xA0\xA7\x00\xEB\xA0\xA8\x00\xEB\xA0\xA9\x00\xEB\xA0\xAA"
-     "\x00\xEB"
-     "\xA0\xAB\x00\xEB\xA0\xAC\x00\xEB\xA0\xAD\x00\xEB\xA0\xAE\x00\xEB\xA0\xAF"
-     "\x00\xEB"
-     "\xA0\xB0\x00\xEB\xA0\xB1\x00\xEB\xA0\xB2\x00\xEB\xA0\xB3\x00\xEB\xA0\xB4"
-     "\x00\xEB"
-     "\xA0\xB5\x00\xEB\xA0\xB6\x00\xEB\xA0\xB7\x00\xEB\xA0\xB8\x00\xEB\xA0\xB9"
-     "\x00\xEB"
-     "\xA0\xBA\x00\xEB\xA0\xBB\x00\xEB\xA0\xBC\x00\xEB\xA0\xBD\x00\xEB\xA0\xBE"
-     "\x00\xEB"
-     "\xA0\xBF\x00\xEB\xA1\x80\x00\xEB\xA1\x81\x00\xEB\xA1\x82\x00\xEB\xA1\x83"
-     "\x00\xEB"
-     "\xA1\x84\x00\xEB\xA1\x85\x00\xEB\xA1\x86\x00\xEB\xA1\x87\x00\xEB\xA1\x88"
-     "\x00\xEB"
-     "\xA1\x89\x00\xEB\xA1\x8A\x00\xEB\xA1\x8B\x00\xEB\xA1\x8C\x00\xEB\xA1\x8D"
-     "\x00\xEB"
-     "\xA1\x8E\x00\xEB\xA1\x8F\x00\xEB\xA1\x90\x00\xEB\xA1\x91\x00\xEB\xA1\x92"
-     "\x00\xEB"
-     "\xA1\x93\x00\xEB\xA1\x94\x00\xEB\xA1\x95\x00\xEB\xA1\x96\x00\xEB\xA1\x97"
-     "\x00\xEB"
-     "\xA1\x98\x00\xEB\xA1\x99\x00\xEB\xA1\x9A\x00\xEB\xA1\x9B\x00\xEB\xA1\x9C"
-     "\x00\xEB"
-     "\xA1\x9D\x00\xEB\xA1\x9E\x00\xEB\xA1\x9F\x00\xEB\xA1\xA0\x00\xEB\xA1\xA1"
-     "\x00\xEB"
-     "\xA1\xA2\x00\xEB\xA1\xA3\x00\xEB\xA1\xA4\x00\xEB\xA1\xA5\x00\xEB\xA1\xA6"
-     "\x00\xEB"
-     "\xA1\xA7\x00\xEB\xA1\xA8\x00\xEB\xA1\xA9\x00\xEB\xA1\xAA\x00\xEB\xA1\xAB"
-     "\x00\xEB"
-     "\xA1\xAC\x00\xEB\xA1\xAD\x00\xEB\xA1\xAE\x00\xEB\xA1\xAF\x00\xEB\xA1\xB0"
-     "\x00\xEB"
-     "\xA1\xB1\x00\xEB\xA1\xB2\x00\xEB\xA1\xB3\x00\xEB\xA1\xB4\x00\xEB\xA1\xB5"
-     "\x00\xEB"
-     "\xA1\xB6\x00\xEB\xA1\xB7\x00\xEB\xA1\xB8\x00\xEB\xA1\xB9\x00\xEB\xA1\xBA"
-     "\x00\xEB"
-     "\xA1\xBB\x00\xEB\xA1\xBC\x00\xEB\xA1\xBD\x00\xEB\xA1\xBE\x00\xEB\xA1\xBF"
-     "\x00\xEB"
-     "\xA2\x80\x00\xEB\xA2\x81\x00\xEB\xA2\x82\x00\xEB\xA2\x83\x00\xEB\xA2\x84"
-     "\x00\xEB"
-     "\xA2\x85\x00\xEB\xA2\x86\x00\xEB\xA2\x87\x00\xEB\xA2\x88\x00\xEB\xA2\x89"
-     "\x00\xEB"
-     "\xA2\x8A\x00\xEB\xA2\x8B\x00\xEB\xA2\x8C\x00\xEB\xA2\x8D\x00\xEB\xA2\x8E"
-     "\x00\xEB"
-     "\xA2\x8F\x00\xEB\xA2\x90\x00\xEB\xA2\x91\x00\xEB\xA2\x92\x00\xEB\xA2\x93"
-     "\x00\xEB"
-     "\xA2\x94\x00\xEB\xA2\x95\x00\xEB\xA2\x96\x00\xEB\xA2\x97\x00\xEB\xA2\x98"
-     "\x00\xEB"
-     "\xA2\x99\x00\xEB\xA2\x9A\x00\xEB\xA2\x9B\x00\xEB\xA2\x9C\x00\xEB\xA2\x9D"
-     "\x00\xEB"
-     "\xA2\x9E\x00\xEB\xA2\x9F\x00\xEB\xA2\xA0\x00\xEB\xA2\xA1\x00\xEB\xA2\xA2"
-     "\x00\xEB"
-     "\xA2\xA3\x00\xEB\xA2\xA4\x00\xEB\xA2\xA5\x00\xEB\xA2\xA6\x00\xEB\xA2\xA7"
-     "\x00\xEB"
-     "\xA2\xA8\x00\xEB\xA2\xA9\x00\xEB\xA2\xAA\x00\xEB\xA2\xAB\x00\xEB\xA2\xAC"
-     "\x00\xEB"
-     "\xA2\xAD\x00\xEB\xA2\xAE\x00\xEB\xA2\xAF\x00\xEB\xA2\xB0\x00\xEB\xA2\xB1"
-     "\x00\xEB"
-     "\xA2\xB2\x00\xEB\xA2\xB3\x00\xEB\xA2\xB4\x00\xEB\xA2\xB5\x00\xEB\xA2\xB6"
-     "\x00\xEB"
-     "\xA2\xB7\x00\xEB\xA2\xB8\x00\xEB\xA2\xB9\x00\xEB\xA2\xBA\x00\xEB\xA2\xBB"
-     "\x00\xEB"
-     "\xA2\xBC\x00\xEB\xA2\xBD\x00\xEB\xA2\xBE\x00\xEB\xA2\xBF\x00\xEB\xA3\x80"
-     "\x00\xEB"
-     "\xA3\x81\x00\xEB\xA3\x82\x00\xEB\xA3\x83\x00\xEB\xA3\x84\x00\xEB\xA3\x85"
-     "\x00\xEB"
-     "\xA3\x86\x00\xEB\xA3\x87\x00\xEB\xA3\x88\x00\xEB\xA3\x89\x00\xEB\xA3\x8A"
-     "\x00\xEB"
-     "\xA3\x8B\x00\xEB\xA3\x8C\x00\xEB\xA3\x8D\x00\xEB\xA3\x8E\x00\xEB\xA3\x8F"
-     "\x00\xEB"
-     "\xA3\x90\x00\xEB\xA3\x91\x00\xEB\xA3\x92\x00\xEB\xA3\x93\x00\xEB\xA3\x94"
-     "\x00\xEB"
-     "\xA3\x95\x00\xEB\xA3\x96\x00\xEB\xA3\x97\x00\xEB\xA3\x98\x00\xEB\xA3\x99"
-     "\x00\xEB"
-     "\xA3\x9A\x00\xEB\xA3\x9B\x00\xEB\xA3\x9C\x00\xEB\xA3\x9D\x00\xEB\xA3\x9E"
-     "\x00\xEB"
-     "\xA3\x9F\x00\xEB\xA3\xA0\x00\xEB\xA3\xA1\x00\xEB\xA3\xA2\x00\xEB\xA3\xA3"
-     "\x00\xEB"
-     "\xA3\xA4\x00\xEB\xA3\xA5\x00\xEB\xA3\xA6\x00\xEB\xA3\xA7\x00\xEB\xA3\xA8"
-     "\x00\xEB"
-     "\xA3\xA9\x00\xEB\xA3\xAA\x00\xEB\xA3\xAB\x00\xEB\xA3\xAC\x00\xEB\xA3\xAD"
-     "\x00\xEB"
-     "\xA3\xAE\x00\xEB\xA3\xAF\x00\xEB\xA3\xB0\x00\xEB\xA3\xB1\x00\xEB\xA3\xB2"
-     "\x00\xEB"
-     "\xA3\xB3\x00\xEB\xA3\xB4\x00\xEB\xA3\xB5\x00\xEB\xA3\xB6\x00\xEB\xA3\xB7"
-     "\x00\xEB"
-     "\xA3\xB8\x00\xEB\xA3\xB9\x00\xEB\xA3\xBA\x00\xEB\xA3\xBB\x00\xEB\xA3\xBC"
-     "\x00\xEB"
-     "\xA3\xBD\x00\xEB\xA3\xBE\x00\xEB\xA3\xBF\x00\xEB\xA4\x80\x00\xEB\xA4\x81"
-     "\x00\xEB"
-     "\xA4\x82\x00\xEB\xA4\x83\x00\xEB\xA4\x84\x00\xEB\xA4\x85\x00\xEB\xA4\x86"
-     "\x00\xEB"
-     "\xA4\x87\x00\xEB\xA4\x88\x00\xEB\xA4\x89\x00\xEB\xA4\x8A\x00\xEB\xA4\x8B"
-     "\x00\xEB"
-     "\xA4\x8C\x00\xEB\xA4\x8D\x00\xEB\xA4\x8E\x00\xEB\xA4\x8F\x00\xEB\xA4\x90"
-     "\x00\xEB"
-     "\xA4\x91\x00\xEB\xA4\x92\x00\xEB\xA4\x93\x00\xEB\xA4\x94\x00\xEB\xA4\x95"
-     "\x00\xEB"
-     "\xA4\x96\x00\xEB\xA4\x97\x00\xEB\xA4\x98\x00\xEB\xA4\x99\x00\xEB\xA4\x9A"
-     "\x00\xEB"
-     "\xA4\x9B\x00\xEB\xA4\x9C\x00\xEB\xA4\x9D\x00\xEB\xA4\x9E\x00\xEB\xA4\x9F"
-     "\x00\xEB"
-     "\xA4\xA0\x00\xEB\xA4\xA1\x00\xEB\xA4\xA2\x00\xEB\xA4\xA3\x00\xEB\xA4\xA4"
-     "\x00\xEB"
-     "\xA4\xA5\x00\xEB\xA4\xA6\x00\xEB\xA4\xA7\x00\xEB\xA4\xA8\x00\xEB\xA4\xA9"
-     "\x00\xEB"
-     "\xA4\xAA\x00\xEB\xA4\xAB\x00\xEB\xA4\xAC\x00\xEB\xA4\xAD\x00\xEB\xA4\xAE"
-     "\x00\xEB"
-     "\xA4\xAF\x00\xEB\xA4\xB0\x00\xEB\xA4\xB1\x00\xEB\xA4\xB2\x00\xEB\xA4\xB3"
-     "\x00\xEB"
-     "\xA4\xB4\x00\xEB\xA4\xB5\x00\xEB\xA4\xB6\x00\xEB\xA4\xB7\x00\xEB\xA4\xB8"
-     "\x00\xEB"
-     "\xA4\xB9\x00\xEB\xA4\xBA\x00\xEB\xA4\xBB\x00\xEB\xA4\xBC\x00\xEB\xA4\xBD"
-     "\x00\xEB"
-     "\xA4\xBE\x00\xEB\xA4\xBF\x00\xEB\xA5\x80\x00\xEB\xA5\x81\x00\xEB\xA5\x82"
-     "\x00\xEB"
-     "\xA5\x83\x00\xEB\xA5\x84\x00\xEB\xA5\x85\x00\xEB\xA5\x86\x00\xEB\xA5\x87"
-     "\x00\xEB"
-     "\xA5\x88\x00\xEB\xA5\x89\x00\xEB\xA5\x8A\x00\xEB\xA5\x8B\x00\xEB\xA5\x8C"
-     "\x00\xEB"
-     "\xA5\x8D\x00\xEB\xA5\x8E\x00\xEB\xA5\x8F\x00\xEB\xA5\x90\x00\xEB\xA5\x91"
-     "\x00\xEB"
-     "\xA5\x92\x00\xEB\xA5\x93\x00\xEB\xA5\x94\x00\xEB\xA5\x95\x00\xEB\xA5\x96"
-     "\x00\xEB"
-     "\xA5\x97\x00\xEB\xA5\x98\x00\xEB\xA5\x99\x00\xEB\xA5\x9A\x00\xEB\xA5\x9B"
-     "\x00\xEB"
-     "\xA5\x9C\x00\xEB\xA5\x9D\x00\xEB\xA5\x9E\x00\xEB\xA5\x9F\x00\xEB\xA5\xA0"
-     "\x00\xEB"
-     "\xA5\xA1\x00\xEB\xA5\xA2\x00\xEB\xA5\xA3\x00\xEB\xA5\xA4\x00\xEB\xA5\xA5"
-     "\x00\xEB"
-     "\xA5\xA6\x00\xEB\xA5\xA7\x00\xEB\xA5\xA8\x00\xEB\xA5\xA9\x00\xEB\xA5\xAA"
-     "\x00\xEB"
-     "\xA5\xAB\x00\xEB\xA5\xAC\x00\xEB\xA5\xAD\x00\xEB\xA5\xAE\x00\xEB\xA5\xAF"
-     "\x00\xEB"
-     "\xA5\xB0\x00\xEB\xA5\xB1\x00\xEB\xA5\xB2\x00\xEB\xA5\xB3\x00\xEB\xA5\xB4"
-     "\x00\xEB"
-     "\xA5\xB5\x00\xEB\xA5\xB6\x00\xEB\xA5\xB7\x00\xEB\xA5\xB8\x00\xEB\xA5\xB9"
-     "\x00\xEB"
-     "\xA5\xBA\x00\xEB\xA5\xBB\x00\xEB\xA5\xBC\x00\xEB\xA5\xBD\x00\xEB\xA5\xBE"
-     "\x00\xEB"
-     "\xA5\xBF\x00\xEB\xA6\x80\x00\xEB\xA6\x81\x00\xEB\xA6\x82\x00\xEB\xA6\x83"
-     "\x00\xEB"
-     "\xA6\x84\x00\xEB\xA6\x85\x00\xEB\xA6\x86\x00\xEB\xA6\x87\x00\xEB\xA6\x88"
-     "\x00\xEB"
-     "\xA6\x89\x00\xEB\xA6\x8A\x00\xEB\xA6\x8B\x00\xEB\xA6\x8C\x00\xEB\xA6\x8D"
-     "\x00\xEB"
-     "\xA6\x8E\x00\xEB\xA6\x8F\x00\xEB\xA6\x90\x00\xEB\xA6\x91\x00\xEB\xA6\x92"
-     "\x00\xEB"
-     "\xA6\x93\x00\xEB\xA6\x94\x00\xEB\xA6\x95\x00\xEB\xA6\x96\x00\xEB\xA6\x97"
-     "\x00\xEB"
-     "\xA6\x98\x00\xEB\xA6\x99\x00\xEB\xA6\x9A\x00\xEB\xA6\x9B\x00\xEB\xA6\x9C"
-     "\x00\xEB"
-     "\xA6\x9D\x00\xEB\xA6\x9E\x00\xEB\xA6\x9F\x00\xEB\xA6\xA0\x00\xEB\xA6\xA1"
-     "\x00\xEB"
-     "\xA6\xA2\x00\xEB\xA6\xA3\x00\xEB\xA6\xA4\x00\xEB\xA6\xA5\x00\xEB\xA6\xA6"
-     "\x00\xEB"
-     "\xA6\xA7\x00\xEB\xA6\xA8\x00\xEB\xA6\xA9\x00\xEB\xA6\xAA\x00\xEB\xA6\xAB"
-     "\x00\xEB"
-     "\xA6\xAC\x00\xEB\xA6\xAD\x00\xEB\xA6\xAE\x00\xEB\xA6\xAF\x00\xEB\xA6\xB0"
-     "\x00\xEB"
-     "\xA6\xB1\x00\xEB\xA6\xB2\x00\xEB\xA6\xB3\x00\xEB\xA6\xB4\x00\xEB\xA6\xB5"
-     "\x00\xEB"
-     "\xA6\xB6\x00\xEB\xA6\xB7\x00\xEB\xA6\xB8\x00\xEB\xA6\xB9\x00\xEB\xA6\xBA"
-     "\x00\xEB"
-     "\xA6\xBB\x00\xEB\xA6\xBC\x00\xEB\xA6\xBD\x00\xEB\xA6\xBE\x00\xEB\xA6\xBF"
-     "\x00\xEB"
-     "\xA7\x80\x00\xEB\xA7\x81\x00\xEB\xA7\x82\x00\xEB\xA7\x83\x00\xEB\xA7\x84"
-     "\x00\xEB"
-     "\xA7\x85\x00\xEB\xA7\x86\x00\xEB\xA7\x87\x00\xEB\xA7\x88\x00\xEB\xA7\x89"
-     "\x00\xEB"
-     "\xA7\x8A\x00\xEB\xA7\x8B\x00\xEB\xA7\x8C\x00\xEB\xA7\x8D\x00\xEB\xA7\x8E"
-     "\x00\xEB"
-     "\xA7\x8F\x00\xEB\xA7\x90\x00\xEB\xA7\x91\x00\xEB\xA7\x92\x00\xEB\xA7\x93"
-     "\x00\xEB"
-     "\xA7\x94\x00\xEB\xA7\x95\x00\xEB\xA7\x96\x00\xEB\xA7\x97\x00\xEB\xA7\x98"
-     "\x00\xEB"
-     "\xA7\x99\x00\xEB\xA7\x9A\x00\xEB\xA7\x9B\x00\xEB\xA7\x9C\x00\xEB\xA7\x9D"
-     "\x00\xEB"
-     "\xA7\x9E\x00\xEB\xA7\x9F\x00\xEB\xA7\xA0\x00\xEB\xA7\xA1\x00\xEB\xA7\xA2"
-     "\x00\xEB"
-     "\xA7\xA3\x00\xEB\xA7\xA4\x00\xEB\xA7\xA5\x00\xEB\xA7\xA6\x00\xEB\xA7\xA7"
-     "\x00\xEB"
-     "\xA7\xA8\x00\xEB\xA7\xA9\x00\xEB\xA7\xAA\x00\xEB\xA7\xAB\x00\xEB\xA7\xAC"
-     "\x00\xEB"
-     "\xA7\xAD\x00\xEB\xA7\xAE\x00\xEB\xA7\xAF\x00\xEB\xA7\xB0\x00\xEB\xA7\xB1"
-     "\x00\xEB"
-     "\xA7\xB2\x00\xEB\xA7\xB3\x00\xEB\xA7\xB4\x00\xEB\xA7\xB5\x00\xEB\xA7\xB6"
-     "\x00\xEB"
-     "\xA7\xB7\x00\xEB\xA7\xB8\x00\xEB\xA7\xB9\x00\xEB\xA7\xBA\x00\xEB\xA7\xBB"
-     "\x00\xEB"
-     "\xA7\xBC\x00\xEB\xA7\xBD\x00\xEB\xA7\xBE\x00\xEB\xA7\xBF\x00\xEB\xA8\x80"
-     "\x00\xEB"
-     "\xA8\x81\x00\xEB\xA8\x82\x00\xEB\xA8\x83\x00\xEB\xA8\x84\x00\xEB\xA8\x85"
-     "\x00\xEB"
-     "\xA8\x86\x00\xEB\xA8\x87\x00\xEB\xA8\x88\x00\xEB\xA8\x89\x00\xEB\xA8\x8A"
-     "\x00\xEB"
-     "\xA8\x8B\x00\xEB\xA8\x8C\x00\xEB\xA8\x8D\x00\xEB\xA8\x8E\x00\xEB\xA8\x8F"
-     "\x00\xEB"
-     "\xA8\x90\x00\xEB\xA8\x91\x00\xEB\xA8\x92\x00\xEB\xA8\x93\x00\xEB\xA8\x94"
-     "\x00\xEB"
-     "\xA8\x95\x00\xEB\xA8\x96\x00\xEB\xA8\x97\x00\xEB\xA8\x98\x00\xEB\xA8\x99"
-     "\x00\xEB"
-     "\xA8\x9A\x00\xEB\xA8\x9B\x00\xEB\xA8\x9C\x00\xEB\xA8\x9D\x00\xEB\xA8\x9E"
-     "\x00\xEB"
-     "\xA8\x9F\x00\xEB\xA8\xA0\x00\xEB\xA8\xA1\x00\xEB\xA8\xA2\x00\xEB\xA8\xA3"
-     "\x00\xEB"
-     "\xA8\xA4\x00\xEB\xA8\xA5\x00\xEB\xA8\xA6\x00\xEB\xA8\xA7\x00\xEB\xA8\xA8"
-     "\x00\xEB"
-     "\xA8\xA9\x00\xEB\xA8\xAA\x00\xEB\xA8\xAB\x00\xEB\xA8\xAC\x00\xEB\xA8\xAD"
-     "\x00\xEB"
-     "\xA8\xAE\x00\xEB\xA8\xAF\x00\xEB\xA8\xB0\x00\xEB\xA8\xB1\x00\xEB\xA8\xB2"
-     "\x00\xEB"
-     "\xA8\xB3\x00\xEB\xA8\xB4\x00\xEB\xA8\xB5\x00\xEB\xA8\xB6\x00\xEB\xA8\xB7"
-     "\x00\xEB"
-     "\xA8\xB8\x00\xEB\xA8\xB9\x00\xEB\xA8\xBA\x00\xEB\xA8\xBB\x00\xEB\xA8\xBC"
-     "\x00\xEB"
-     "\xA8\xBD\x00\xEB\xA8\xBE\x00\xEB\xA8\xBF\x00\xEB\xA9\x80\x00\xEB\xA9\x81"
-     "\x00\xEB"
-     "\xA9\x82\x00\xEB\xA9\x83\x00\xEB\xA9\x84\x00\xEB\xA9\x85\x00\xEB\xA9\x86"
-     "\x00\xEB"
-     "\xA9\x87\x00\xEB\xA9\x88\x00\xEB\xA9\x89\x00\xEB\xA9\x8A\x00\xEB\xA9\x8B"
-     "\x00\xEB"
-     "\xA9\x8C\x00\xEB\xA9\x8D\x00\xEB\xA9\x8E\x00\xEB\xA9\x8F\x00\xEB\xA9\x90"
-     "\x00\xEB"
-     "\xA9\x91\x00\xEB\xA9\x92\x00\xEB\xA9\x93\x00\xEB\xA9\x94\x00\xEB\xA9\x95"
-     "\x00\xEB"
-     "\xA9\x96\x00\xEB\xA9\x97\x00\xEB\xA9\x98\x00\xEB\xA9\x99\x00\xEB\xA9\x9A"
-     "\x00\xEB"
-     "\xA9\x9B\x00\xEB\xA9\x9C\x00\xEB\xA9\x9D\x00\xEB\xA9\x9E\x00\xEB\xA9\x9F"
-     "\x00\xEB"
-     "\xA9\xA0\x00\xEB\xA9\xA1\x00\xEB\xA9\xA2\x00\xEB\xA9\xA3\x00\xEB\xA9\xA4"
-     "\x00\xEB"
-     "\xA9\xA5\x00\xEB\xA9\xA6\x00\xEB\xA9\xA7\x00\xEB\xA9\xA8\x00\xEB\xA9\xA9"
-     "\x00\xEB"
-     "\xA9\xAA\x00\xEB\xA9\xAB\x00\xEB\xA9\xAC\x00\xEB\xA9\xAD\x00\xEB\xA9\xAE"
-     "\x00\xEB"
-     "\xA9\xAF\x00\xEB\xA9\xB0\x00\xEB\xA9\xB1\x00\xEB\xA9\xB2\x00\xEB\xA9\xB3"
-     "\x00\xEB"
-     "\xA9\xB4\x00\xEB\xA9\xB5\x00\xEB\xA9\xB6\x00\xEB\xA9\xB7\x00\xEB\xA9\xB8"
-     "\x00\xEB"
-     "\xA9\xB9\x00\xEB\xA9\xBA\x00\xEB\xA9\xBB\x00\xEB\xA9\xBC\x00\xEB\xA9\xBD"
-     "\x00\xEB"
-     "\xA9\xBE\x00\xEB\xA9\xBF\x00\xEB\xAA\x80\x00\xEB\xAA\x81\x00\xEB\xAA\x82"
-     "\x00\xEB"
-     "\xAA\x83\x00\xEB\xAA\x84\x00\xEB\xAA\x85\x00\xEB\xAA\x86\x00\xEB\xAA\x87"
-     "\x00\xEB"
-     "\xAA\x88\x00\xEB\xAA\x89\x00\xEB\xAA\x8A\x00\xEB\xAA\x8B\x00\xEB\xAA\x8C"
-     "\x00\xEB"
-     "\xAA\x8D\x00\xEB\xAA\x8E\x00\xEB\xAA\x8F\x00\xEB\xAA\x90\x00\xEB\xAA\x91"
-     "\x00\xEB"
-     "\xAA\x92\x00\xEB\xAA\x93\x00\xEB\xAA\x94\x00\xEB\xAA\x95\x00\xEB\xAA\x96"
-     "\x00\xEB"
-     "\xAA\x97\x00\xEB\xAA\x98\x00\xEB\xAA\x99\x00\xEB\xAA\x9A\x00\xEB\xAA\x9B"
-     "\x00\xEB"
-     "\xAA\x9C\x00\xEB\xAA\x9D\x00\xEB\xAA\x9E\x00\xEB\xAA\x9F\x00\xEB\xAA\xA0"
-     "\x00\xEB"
-     "\xAA\xA1\x00\xEB\xAA\xA2\x00\xEB\xAA\xA3\x00\xEB\xAA\xA4\x00\xEB\xAA\xA5"
-     "\x00\xEB"
-     "\xAA\xA6\x00\xEB\xAA\xA7\x00\xEB\xAA\xA8\x00\xEB\xAA\xA9\x00\xEB\xAA\xAA"
-     "\x00\xEB"
-     "\xAA\xAB\x00\xEB\xAA\xAC\x00\xEB\xAA\xAD\x00\xEB\xAA\xAE\x00\xEB\xAA\xAF"
-     "\x00\xEB"
-     "\xAA\xB0\x00\xEB\xAA\xB1\x00\xEB\xAA\xB2\x00\xEB\xAA\xB3\x00\xEB\xAA\xB4"
-     "\x00\xEB"
-     "\xAA\xB5\x00\xEB\xAA\xB6\x00\xEB\xAA\xB7\x00\xEB\xAA\xB8\x00\xEB\xAA\xB9"
-     "\x00\xEB"
-     "\xAA\xBA\x00\xEB\xAA\xBB\x00\xEB\xAA\xBC\x00\xEB\xAA\xBD\x00\xEB\xAA\xBE"
-     "\x00\xEB"
-     "\xAA\xBF\x00\xEB\xAB\x80\x00\xEB\xAB\x81\x00\xEB\xAB\x82\x00\xEB\xAB\x83"
-     "\x00\xEB"
-     "\xAB\x84\x00\xEB\xAB\x85\x00\xEB\xAB\x86\x00\xEB\xAB\x87\x00\xEB\xAB\x88"
-     "\x00\xEB"
-     "\xAB\x89\x00\xEB\xAB\x8A\x00\xEB\xAB\x8B\x00\xEB\xAB\x8C\x00\xEB\xAB\x8D"
-     "\x00\xEB"
-     "\xAB\x8E\x00\xEB\xAB\x8F\x00\xEB\xAB\x90\x00\xEB\xAB\x91\x00\xEB\xAB\x92"
-     "\x00\xEB"
-     "\xAB\x93\x00\xEB\xAB\x94\x00\xEB\xAB\x95\x00\xEB\xAB\x96\x00\xEB\xAB\x97"
-     "\x00\xEB"
-     "\xAB\x98\x00\xEB\xAB\x99\x00\xEB\xAB\x9A\x00\xEB\xAB\x9B\x00\xEB\xAB\x9C"
-     "\x00\xEB"
-     "\xAB\x9D\x00\xEB\xAB\x9E\x00\xEB\xAB\x9F\x00\xEB\xAB\xA0\x00\xEB\xAB\xA1"
-     "\x00\xEB"
-     "\xAB\xA2\x00\xEB\xAB\xA3\x00\xEB\xAB\xA4\x00\xEB\xAB\xA5\x00\xEB\xAB\xA6"
-     "\x00\xEB"
-     "\xAB\xA7\x00\xEB\xAB\xA8\x00\xEB\xAB\xA9\x00\xEB\xAB\xAA\x00\xEB\xAB\xAB"
-     "\x00\xEB"
-     "\xAB\xAC\x00\xEB\xAB\xAD\x00\xEB\xAB\xAE\x00\xEB\xAB\xAF\x00\xEB\xAB\xB0"
-     "\x00\xEB"
-     "\xAB\xB1\x00\xEB\xAB\xB2\x00\xEB\xAB\xB3\x00\xEB\xAB\xB4\x00\xEB\xAB\xB5"
-     "\x00\xEB"
-     "\xAB\xB6\x00\xEB\xAB\xB7\x00\xEB\xAB\xB8\x00\xEB\xAB\xB9\x00\xEB\xAB\xBA"
-     "\x00\xEB"
-     "\xAB\xBB\x00\xEB\xAB\xBC\x00\xEB\xAB\xBD\x00\xEB\xAB\xBE\x00\xEB\xAB\xBF"
-     "\x00\xEB"
-     "\xAC\x80\x00\xEB\xAC\x81\x00\xEB\xAC\x82\x00\xEB\xAC\x83\x00\xEB\xAC\x84"
-     "\x00\xEB"
-     "\xAC\x85\x00\xEB\xAC\x86\x00\xEB\xAC\x87\x00\xEB\xAC\x88\x00\xEB\xAC\x89"
-     "\x00\xEB"
-     "\xAC\x8A\x00\xEB\xAC\x8B\x00\xEB\xAC\x8C\x00\xEB\xAC\x8D\x00\xEB\xAC\x8E"
-     "\x00\xEB"
-     "\xAC\x8F\x00\xEB\xAC\x90\x00\xEB\xAC\x91\x00\xEB\xAC\x92\x00\xEB\xAC\x93"
-     "\x00\xEB"
-     "\xAC\x94\x00\xEB\xAC\x95\x00\xEB\xAC\x96\x00\xEB\xAC\x97\x00\xEB\xAC\x98"
-     "\x00\xEB"
-     "\xAC\x99\x00\xEB\xAC\x9A\x00\xEB\xAC\x9B\x00\xEB\xAC\x9C\x00\xEB\xAC\x9D"
-     "\x00\xEB"
-     "\xAC\x9E\x00\xEB\xAC\x9F\x00\xEB\xAC\xA0\x00\xEB\xAC\xA1\x00\xEB\xAC\xA2"
-     "\x00\xEB"
-     "\xAC\xA3\x00\xEB\xAC\xA4\x00\xEB\xAC\xA5\x00\xEB\xAC\xA6\x00\xEB\xAC\xA7"
-     "\x00\xEB"
-     "\xAC\xA8\x00\xEB\xAC\xA9\x00\xEB\xAC\xAA\x00\xEB\xAC\xAB\x00\xEB\xAC\xAC"
-     "\x00\xEB"
-     "\xAC\xAD\x00\xEB\xAC\xAE\x00\xEB\xAC\xAF\x00\xEB\xAC\xB0\x00\xEB\xAC\xB1"
-     "\x00\xEB"
-     "\xAC\xB2\x00\xEB\xAC\xB3\x00\xEB\xAC\xB4\x00\xEB\xAC\xB5\x00\xEB\xAC\xB6"
-     "\x00\xEB"
-     "\xAC\xB7\x00\xEB\xAC\xB8\x00\xEB\xAC\xB9\x00\xEB\xAC\xBA\x00\xEB\xAC\xBB"
-     "\x00\xEB"
-     "\xAC\xBC\x00\xEB\xAC\xBD\x00\xEB\xAC\xBE\x00\xEB\xAC\xBF\x00\xEB\xAD\x80"
-     "\x00\xEB"
-     "\xAD\x81\x00\xEB\xAD\x82\x00\xEB\xAD\x83\x00\xEB\xAD\x84\x00\xEB\xAD\x85"
-     "\x00\xEB"
-     "\xAD\x86\x00\xEB\xAD\x87\x00\xEB\xAD\x88\x00\xEB\xAD\x89\x00\xEB\xAD\x8A"
-     "\x00\xEB"
-     "\xAD\x8B\x00\xEB\xAD\x8C\x00\xEB\xAD\x8D\x00\xEB\xAD\x8E\x00\xEB\xAD\x8F"
-     "\x00\xEB"
-     "\xAD\x90\x00\xEB\xAD\x91\x00\xEB\xAD\x92\x00\xEB\xAD\x93\x00\xEB\xAD\x94"
-     "\x00\xEB"
-     "\xAD\x95\x00\xEB\xAD\x96\x00\xEB\xAD\x97\x00\xEB\xAD\x98\x00\xEB\xAD\x99"
-     "\x00\xEB"
-     "\xAD\x9A\x00\xEB\xAD\x9B\x00\xEB\xAD\x9C\x00\xEB\xAD\x9D\x00\xEB\xAD\x9E"
-     "\x00\xEB"
-     "\xAD\x9F\x00\xEB\xAD\xA0\x00\xEB\xAD\xA1\x00\xEB\xAD\xA2\x00\xEB\xAD\xA3"
-     "\x00\xEB"
-     "\xAD\xA4\x00\xEB\xAD\xA5\x00\xEB\xAD\xA6\x00\xEB\xAD\xA7\x00\xEB\xAD\xA8"
-     "\x00\xEB"
-     "\xAD\xA9\x00\xEB\xAD\xAA\x00\xEB\xAD\xAB\x00\xEB\xAD\xAC\x00\xEB\xAD\xAD"
-     "\x00\xEB"
-     "\xAD\xAE\x00\xEB\xAD\xAF\x00\xEB\xAD\xB0\x00\xEB\xAD\xB1\x00\xEB\xAD\xB2"
-     "\x00\xEB"
-     "\xAD\xB3\x00\xEB\xAD\xB4\x00\xEB\xAD\xB5\x00\xEB\xAD\xB6\x00\xEB\xAD\xB7"
-     "\x00\xEB"
-     "\xAD\xB8\x00\xEB\xAD\xB9\x00\xEB\xAD\xBA\x00\xEB\xAD\xBB\x00\xEB\xAD\xBC"
-     "\x00\xEB"
-     "\xAD\xBD\x00\xEB\xAD\xBE\x00\xEB\xAD\xBF\x00\xEB\xAE\x80\x00\xEB\xAE\x81"
-     "\x00\xEB"
-     "\xAE\x82\x00\xEB\xAE\x83\x00\xEB\xAE\x84\x00\xEB\xAE\x85\x00\xEB\xAE\x86"
-     "\x00\xEB"
-     "\xAE\x87\x00\xEB\xAE\x88\x00\xEB\xAE\x89\x00\xEB\xAE\x8A\x00\xEB\xAE\x8B"
-     "\x00\xEB"
-     "\xAE\x8C\x00\xEB\xAE\x8D\x00\xEB\xAE\x8E\x00\xEB\xAE\x8F\x00\xEB\xAE\x90"
-     "\x00\xEB"
-     "\xAE\x91\x00\xEB\xAE\x92\x00\xEB\xAE\x93\x00\xEB\xAE\x94\x00\xEB\xAE\x95"
-     "\x00\xEB"
-     "\xAE\x96\x00\xEB\xAE\x97\x00\xEB\xAE\x98\x00\xEB\xAE\x99\x00\xEB\xAE\x9A"
-     "\x00\xEB"
-     "\xAE\x9B\x00\xEB\xAE\x9C\x00\xEB\xAE\x9D\x00\xEB\xAE\x9E\x00\xEB\xAE\x9F"
-     "\x00\xEB"
-     "\xAE\xA0\x00\xEB\xAE\xA1\x00\xEB\xAE\xA2\x00\xEB\xAE\xA3\x00\xEB\xAE\xA4"
-     "\x00\xEB"
-     "\xAE\xA5\x00\xEB\xAE\xA6\x00\xEB\xAE\xA7\x00\xEB\xAE\xA8\x00\xEB\xAE\xA9"
-     "\x00\xEB"
-     "\xAE\xAA\x00\xEB\xAE\xAB\x00\xEB\xAE\xAC\x00\xEB\xAE\xAD\x00\xEB\xAE\xAE"
-     "\x00\xEB"
-     "\xAE\xAF\x00\xEB\xAE\xB0\x00\xEB\xAE\xB1\x00\xEB\xAE\xB2\x00\xEB\xAE\xB3"
-     "\x00\xEB"
-     "\xAE\xB4\x00\xEB\xAE\xB5\x00\xEB\xAE\xB6\x00\xEB\xAE\xB7\x00\xEB\xAE\xB8"
-     "\x00\xEB"
-     "\xAE\xB9\x00\xEB\xAE\xBA\x00\xEB\xAE\xBB\x00\xEB\xAE\xBC\x00\xEB\xAE\xBD"
-     "\x00\xEB"
-     "\xAE\xBE\x00\xEB\xAE\xBF\x00\xEB\xAF\x80\x00\xEB\xAF\x81\x00\xEB\xAF\x82"
-     "\x00\xEB"
-     "\xAF\x83\x00\xEB\xAF\x84\x00\xEB\xAF\x85\x00\xEB\xAF\x86\x00\xEB\xAF\x87"
-     "\x00\xEB"
-     "\xAF\x88\x00\xEB\xAF\x89\x00\xEB\xAF\x8A\x00\xEB\xAF\x8B\x00\xEB\xAF\x8C"
-     "\x00\xEB"
-     "\xAF\x8D\x00\xEB\xAF\x8E\x00\xEB\xAF\x8F\x00\xEB\xAF\x90\x00\xEB\xAF\x91"
-     "\x00\xEB"
-     "\xAF\x92\x00\xEB\xAF\x93\x00\xEB\xAF\x94\x00\xEB\xAF\x95\x00\xEB\xAF\x96"
-     "\x00\xEB"
-     "\xAF\x97\x00\xEB\xAF\x98\x00\xEB\xAF\x99\x00\xEB\xAF\x9A\x00\xEB\xAF\x9B"
-     "\x00\xEB"
-     "\xAF\x9C\x00\xEB\xAF\x9D\x00\xEB\xAF\x9E\x00\xEB\xAF\x9F\x00\xEB\xAF\xA0"
-     "\x00\xEB"
-     "\xAF\xA1\x00\xEB\xAF\xA2\x00\xEB\xAF\xA3\x00\xEB\xAF\xA4\x00\xEB\xAF\xA5"
-     "\x00\xEB"
-     "\xAF\xA6\x00\xEB\xAF\xA7\x00\xEB\xAF\xA8\x00\xEB\xAF\xA9\x00\xEB\xAF\xAA"
-     "\x00\xEB"
-     "\xAF\xAB\x00\xEB\xAF\xAC\x00\xEB\xAF\xAD\x00\xEB\xAF\xAE\x00\xEB\xAF\xAF"
-     "\x00\xEB"
-     "\xAF\xB0\x00\xEB\xAF\xB1\x00\xEB\xAF\xB2\x00\xEB\xAF\xB3\x00\xEB\xAF\xB4"
-     "\x00\xEB"
-     "\xAF\xB5\x00\xEB\xAF\xB6\x00\xEB\xAF\xB7\x00\xEB\xAF\xB8\x00\xEB\xAF\xB9"
-     "\x00\xEB"
-     "\xAF\xBA\x00\xEB\xAF\xBB\x00\xEB\xAF\xBC\x00\xEB\xAF\xBD\x00\xEB\xAF\xBE"
-     "\x00\xEB"
-     "\xAF\xBF\x00\xEB\xB0\x80\x00\xEB\xB0\x81\x00\xEB\xB0\x82\x00\xEB\xB0\x83"
-     "\x00\xEB"
-     "\xB0\x84\x00\xEB\xB0\x85\x00\xEB\xB0\x86\x00\xEB\xB0\x87\x00\xEB\xB0\x88"
-     "\x00\xEB"
-     "\xB0\x89\x00\xEB\xB0\x8A\x00\xEB\xB0\x8B\x00\xEB\xB0\x8C\x00\xEB\xB0\x8D"
-     "\x00\xEB"
-     "\xB0\x8E\x00\xEB\xB0\x8F\x00\xEB\xB0\x90\x00\xEB\xB0\x91\x00\xEB\xB0\x92"
-     "\x00\xEB"
-     "\xB0\x93\x00\xEB\xB0\x94\x00\xEB\xB0\x95\x00\xEB\xB0\x96\x00\xEB\xB0\x97"
-     "\x00\xEB"
-     "\xB0\x98\x00\xEB\xB0\x99\x00\xEB\xB0\x9A\x00\xEB\xB0\x9B\x00\xEB\xB0\x9C"
-     "\x00\xEB"
-     "\xB0\x9D\x00\xEB\xB0\x9E\x00\xEB\xB0\x9F\x00\xEB\xB0\xA0\x00\xEB\xB0\xA1"
-     "\x00\xEB"
-     "\xB0\xA2\x00\xEB\xB0\xA3\x00\xEB\xB0\xA4\x00\xEB\xB0\xA5\x00\xEB\xB0\xA6"
-     "\x00\xEB"
-     "\xB0\xA7\x00\xEB\xB0\xA8\x00\xEB\xB0\xA9\x00\xEB\xB0\xAA\x00\xEB\xB0\xAB"
-     "\x00\xEB"
-     "\xB0\xAC\x00\xEB\xB0\xAD\x00\xEB\xB0\xAE\x00\xEB\xB0\xAF\x00\xEB\xB0\xB0"
-     "\x00\xEB"
-     "\xB0\xB1\x00\xEB\xB0\xB2\x00\xEB\xB0\xB3\x00\xEB\xB0\xB4\x00\xEB\xB0\xB5"
-     "\x00\xEB"
-     "\xB0\xB6\x00\xEB\xB0\xB7\x00\xEB\xB0\xB8\x00\xEB\xB0\xB9\x00\xEB\xB0\xBA"
-     "\x00\xEB"
-     "\xB0\xBB\x00\xEB\xB0\xBC\x00\xEB\xB0\xBD\x00\xEB\xB0\xBE\x00\xEB\xB0\xBF"
-     "\x00\xEB"
-     "\xB1\x80\x00\xEB\xB1\x81\x00\xEB\xB1\x82\x00\xEB\xB1\x83\x00\xEB\xB1\x84"
-     "\x00\xEB"
-     "\xB1\x85\x00\xEB\xB1\x86\x00\xEB\xB1\x87\x00\xEB\xB1\x88\x00\xEB\xB1\x89"
-     "\x00\xEB"
-     "\xB1\x8A\x00\xEB\xB1\x8B\x00\xEB\xB1\x8C\x00\xEB\xB1\x8D\x00\xEB\xB1\x8E"
-     "\x00\xEB"
-     "\xB1\x8F\x00\xEB\xB1\x90\x00\xEB\xB1\x91\x00\xEB\xB1\x92\x00\xEB\xB1\x93"
-     "\x00\xEB"
-     "\xB1\x94\x00\xEB\xB1\x95\x00\xEB\xB1\x96\x00\xEB\xB1\x97\x00\xEB\xB1\x98"
-     "\x00\xEB"
-     "\xB1\x99\x00\xEB\xB1\x9A\x00\xEB\xB1\x9B\x00\xEB\xB1\x9C\x00\xEB\xB1\x9D"
-     "\x00\xEB"
-     "\xB1\x9E\x00\xEB\xB1\x9F\x00\xEB\xB1\xA0\x00\xEB\xB1\xA1\x00\xEB\xB1\xA2"
-     "\x00\xEB"
-     "\xB1\xA3\x00\xEB\xB1\xA4\x00\xEB\xB1\xA5\x00\xEB\xB1\xA6\x00\xEB\xB1\xA7"
-     "\x00\xEB"
-     "\xB1\xA8\x00\xEB\xB1\xA9\x00\xEB\xB1\xAA\x00\xEB\xB1\xAB\x00\xEB\xB1\xAC"
-     "\x00\xEB"
-     "\xB1\xAD\x00\xEB\xB1\xAE\x00\xEB\xB1\xAF\x00\xEB\xB1\xB0\x00\xEB\xB1\xB1"
-     "\x00\xEB"
-     "\xB1\xB2\x00\xEB\xB1\xB3\x00\xEB\xB1\xB4\x00\xEB\xB1\xB5\x00\xEB\xB1\xB6"
-     "\x00\xEB"
-     "\xB1\xB7\x00\xEB\xB1\xB8\x00\xEB\xB1\xB9\x00\xEB\xB1\xBA\x00\xEB\xB1\xBB"
-     "\x00\xEB"
-     "\xB1\xBC\x00\xEB\xB1\xBD\x00\xEB\xB1\xBE\x00\xEB\xB1\xBF\x00\xEB\xB2\x80"
-     "\x00\xEB"
-     "\xB2\x81\x00\xEB\xB2\x82\x00\xEB\xB2\x83\x00\xEB\xB2\x84\x00\xEB\xB2\x85"
-     "\x00\xEB"
-     "\xB2\x86\x00\xEB\xB2\x87\x00\xEB\xB2\x88\x00\xEB\xB2\x89\x00\xEB\xB2\x8A"
-     "\x00\xEB"
-     "\xB2\x8B\x00\xEB\xB2\x8C\x00\xEB\xB2\x8D\x00\xEB\xB2\x8E\x00\xEB\xB2\x8F"
-     "\x00\xEB"
-     "\xB2\x90\x00\xEB\xB2\x91\x00\xEB\xB2\x92\x00\xEB\xB2\x93\x00\xEB\xB2\x94"
-     "\x00\xEB"
-     "\xB2\x95\x00\xEB\xB2\x96\x00\xEB\xB2\x97\x00\xEB\xB2\x98\x00\xEB\xB2\x99"
-     "\x00\xEB"
-     "\xB2\x9A\x00\xEB\xB2\x9B\x00\xEB\xB2\x9C\x00\xEB\xB2\x9D\x00\xEB\xB2\x9E"
-     "\x00\xEB"
-     "\xB2\x9F\x00\xEB\xB2\xA0\x00\xEB\xB2\xA1\x00\xEB\xB2\xA2\x00\xEB\xB2\xA3"
-     "\x00\xEB"
-     "\xB2\xA4\x00\xEB\xB2\xA5\x00\xEB\xB2\xA6\x00\xEB\xB2\xA7\x00\xEB\xB2\xA8"
-     "\x00\xEB"
-     "\xB2\xA9\x00\xEB\xB2\xAA\x00\xEB\xB2\xAB\x00\xEB\xB2\xAC\x00\xEB\xB2\xAD"
-     "\x00\xEB"
-     "\xB2\xAE\x00\xEB\xB2\xAF\x00\xEB\xB2\xB0\x00\xEB\xB2\xB1\x00\xEB\xB2\xB2"
-     "\x00\xEB"
-     "\xB2\xB3\x00\xEB\xB2\xB4\x00\xEB\xB2\xB5\x00\xEB\xB2\xB6\x00\xEB\xB2\xB7"
-     "\x00\xEB"
-     "\xB2\xB8\x00\xEB\xB2\xB9\x00\xEB\xB2\xBA\x00\xEB\xB2\xBB\x00\xEB\xB2\xBC"
-     "\x00\xEB"
-     "\xB2\xBD\x00\xEB\xB2\xBE\x00\xEB\xB2\xBF\x00\xEB\xB3\x80\x00\xEB\xB3\x81"
-     "\x00\xEB"
-     "\xB3\x82\x00\xEB\xB3\x83\x00\xEB\xB3\x84\x00\xEB\xB3\x85\x00\xEB\xB3\x86"
-     "\x00\xEB"
-     "\xB3\x87\x00\xEB\xB3\x88\x00\xEB\xB3\x89\x00\xEB\xB3\x8A\x00\xEB\xB3\x8B"
-     "\x00\xEB"
-     "\xB3\x8C\x00\xEB\xB3\x8D\x00\xEB\xB3\x8E\x00\xEB\xB3\x8F\x00\xEB\xB3\x90"
-     "\x00\xEB"
-     "\xB3\x91\x00\xEB\xB3\x92\x00\xEB\xB3\x93\x00\xEB\xB3\x94\x00\xEB\xB3\x95"
-     "\x00\xEB"
-     "\xB3\x96\x00\xEB\xB3\x97\x00\xEB\xB3\x98\x00\xEB\xB3\x99\x00\xEB\xB3\x9A"
-     "\x00\xEB"
-     "\xB3\x9B\x00\xEB\xB3\x9C\x00\xEB\xB3\x9D\x00\xEB\xB3\x9E\x00\xEB\xB3\x9F"
-     "\x00\xEB"
-     "\xB3\xA0\x00\xEB\xB3\xA1\x00\xEB\xB3\xA2\x00\xEB\xB3\xA3\x00\xEB\xB3\xA4"
-     "\x00\xEB"
-     "\xB3\xA5\x00\xEB\xB3\xA6\x00\xEB\xB3\xA7\x00\xEB\xB3\xA8\x00\xEB\xB3\xA9"
-     "\x00\xEB"
-     "\xB3\xAA\x00\xEB\xB3\xAB\x00\xEB\xB3\xAC\x00\xEB\xB3\xAD\x00\xEB\xB3\xAE"
-     "\x00\xEB"
-     "\xB3\xAF\x00\xEB\xB3\xB0\x00\xEB\xB3\xB1\x00\xEB\xB3\xB2\x00\xEB\xB3\xB3"
-     "\x00\xEB"
-     "\xB3\xB4\x00\xEB\xB3\xB5\x00\xEB\xB3\xB6\x00\xEB\xB3\xB7\x00\xEB\xB3\xB8"
-     "\x00\xEB"
-     "\xB3\xB9\x00\xEB\xB3\xBA\x00\xEB\xB3\xBB\x00\xEB\xB3\xBC\x00\xEB\xB3\xBD"
-     "\x00\xEB"
-     "\xB3\xBE\x00\xEB\xB3\xBF\x00\xEB\xB4\x80\x00\xEB\xB4\x81\x00\xEB\xB4\x82"
-     "\x00\xEB"
-     "\xB4\x83\x00\xEB\xB4\x84\x00\xEB\xB4\x85\x00\xEB\xB4\x86\x00\xEB\xB4\x87"
-     "\x00\xEB"
-     "\xB4\x88\x00\xEB\xB4\x89\x00\xEB\xB4\x8A\x00\xEB\xB4\x8B\x00\xEB\xB4\x8C"
-     "\x00\xEB"
-     "\xB4\x8D\x00\xEB\xB4\x8E\x00\xEB\xB4\x8F\x00\xEB\xB4\x90\x00\xEB\xB4\x91"
-     "\x00\xEB"
-     "\xB4\x92\x00\xEB\xB4\x93\x00\xEB\xB4\x94\x00\xEB\xB4\x95\x00\xEB\xB4\x96"
-     "\x00\xEB"
-     "\xB4\x97\x00\xEB\xB4\x98\x00\xEB\xB4\x99\x00\xEB\xB4\x9A\x00\xEB\xB4\x9B"
-     "\x00\xEB"
-     "\xB4\x9C\x00\xEB\xB4\x9D\x00\xEB\xB4\x9E\x00\xEB\xB4\x9F\x00\xEB\xB4\xA0"
-     "\x00\xEB"
-     "\xB4\xA1\x00\xEB\xB4\xA2\x00\xEB\xB4\xA3\x00\xEB\xB4\xA4\x00\xEB\xB4\xA5"
-     "\x00\xEB"
-     "\xB4\xA6\x00\xEB\xB4\xA7\x00\xEB\xB4\xA8\x00\xEB\xB4\xA9\x00\xEB\xB4\xAA"
-     "\x00\xEB"
-     "\xB4\xAB\x00\xEB\xB4\xAC\x00\xEB\xB4\xAD\x00\xEB\xB4\xAE\x00\xEB\xB4\xAF"
-     "\x00\xEB"
-     "\xB4\xB0\x00\xEB\xB4\xB1\x00\xEB\xB4\xB2\x00\xEB\xB4\xB3\x00\xEB\xB4\xB4"
-     "\x00\xEB"
-     "\xB4\xB5\x00\xEB\xB4\xB6\x00\xEB\xB4\xB7\x00\xEB\xB4\xB8\x00\xEB\xB4\xB9"
-     "\x00\xEB"
-     "\xB4\xBA\x00\xEB\xB4\xBB\x00\xEB\xB4\xBC\x00\xEB\xB4\xBD\x00\xEB\xB4\xBE"
-     "\x00\xEB"
-     "\xB4\xBF\x00\xEB\xB5\x80\x00\xEB\xB5\x81\x00\xEB\xB5\x82\x00\xEB\xB5\x83"
-     "\x00\xEB"
-     "\xB5\x84\x00\xEB\xB5\x85\x00\xEB\xB5\x86\x00\xEB\xB5\x87\x00\xEB\xB5\x88"
-     "\x00\xEB"
-     "\xB5\x89\x00\xEB\xB5\x8A\x00\xEB\xB5\x8B\x00\xEB\xB5\x8C\x00\xEB\xB5\x8D"
-     "\x00\xEB"
-     "\xB5\x8E\x00\xEB\xB5\x8F\x00\xEB\xB5\x90\x00\xEB\xB5\x91\x00\xEB\xB5\x92"
-     "\x00\xEB"
-     "\xB5\x93\x00\xEB\xB5\x94\x00\xEB\xB5\x95\x00\xEB\xB5\x96\x00\xEB\xB5\x97"
-     "\x00\xEB"
-     "\xB5\x98\x00\xEB\xB5\x99\x00\xEB\xB5\x9A\x00\xEB\xB5\x9B\x00\xEB\xB5\x9C"
-     "\x00\xEB"
-     "\xB5\x9D\x00\xEB\xB5\x9E\x00\xEB\xB5\x9F\x00\xEB\xB5\xA0\x00\xEB\xB5\xA1"
-     "\x00\xEB"
-     "\xB5\xA2\x00\xEB\xB5\xA3\x00\xEB\xB5\xA4\x00\xEB\xB5\xA5\x00\xEB\xB5\xA6"
-     "\x00\xEB"
-     "\xB5\xA7\x00\xEB\xB5\xA8\x00\xEB\xB5\xA9\x00\xEB\xB5\xAA\x00\xEB\xB5\xAB"
-     "\x00\xEB"
-     "\xB5\xAC\x00\xEB\xB5\xAD\x00\xEB\xB5\xAE\x00\xEB\xB5\xAF\x00\xEB\xB5\xB0"
-     "\x00\xEB"
-     "\xB5\xB1\x00\xEB\xB5\xB2\x00\xEB\xB5\xB3\x00\xEB\xB5\xB4\x00\xEB\xB5\xB5"
-     "\x00\xEB"
-     "\xB5\xB6\x00\xEB\xB5\xB7\x00\xEB\xB5\xB8\x00\xEB\xB5\xB9\x00\xEB\xB5\xBA"
-     "\x00\xEB"
-     "\xB5\xBB\x00\xEB\xB5\xBC\x00\xEB\xB5\xBD\x00\xEB\xB5\xBE\x00\xEB\xB5\xBF"
-     "\x00\xEB"
-     "\xB6\x80\x00\xEB\xB6\x81\x00\xEB\xB6\x82\x00\xEB\xB6\x83\x00\xEB\xB6\x84"
-     "\x00\xEB"
-     "\xB6\x85\x00\xEB\xB6\x86\x00\xEB\xB6\x87\x00\xEB\xB6\x88\x00\xEB\xB6\x89"
-     "\x00\xEB"
-     "\xB6\x8A\x00\xEB\xB6\x8B\x00\xEB\xB6\x8C\x00\xEB\xB6\x8D\x00\xEB\xB6\x8E"
-     "\x00\xEB"
-     "\xB6\x8F\x00\xEB\xB6\x90\x00\xEB\xB6\x91\x00\xEB\xB6\x92\x00\xEB\xB6\x93"
-     "\x00\xEB"
-     "\xB6\x94\x00\xEB\xB6\x95\x00\xEB\xB6\x96\x00\xEB\xB6\x97\x00\xEB\xB6\x98"
-     "\x00\xEB"
-     "\xB6\x99\x00\xEB\xB6\x9A\x00\xEB\xB6\x9B\x00\xEB\xB6\x9C\x00\xEB\xB6\x9D"
-     "\x00\xEB"
-     "\xB6\x9E\x00\xEB\xB6\x9F\x00\xEB\xB6\xA0\x00\xEB\xB6\xA1\x00\xEB\xB6\xA2"
-     "\x00\xEB"
-     "\xB6\xA3\x00\xEB\xB6\xA4\x00\xEB\xB6\xA5\x00\xEB\xB6\xA6\x00\xEB\xB6\xA7"
-     "\x00\xEB"
-     "\xB6\xA8\x00\xEB\xB6\xA9\x00\xEB\xB6\xAA\x00\xEB\xB6\xAB\x00\xEB\xB6\xAC"
-     "\x00\xEB"
-     "\xB6\xAD\x00\xEB\xB6\xAE\x00\xEB\xB6\xAF\x00\xEB\xB6\xB0\x00\xEB\xB6\xB1"
-     "\x00\xEB"
-     "\xB6\xB2\x00\xEB\xB6\xB3\x00\xEB\xB6\xB4\x00\xEB\xB6\xB5\x00\xEB\xB6\xB6"
-     "\x00\xEB"
-     "\xB6\xB7\x00\xEB\xB6\xB8\x00\xEB\xB6\xB9\x00\xEB\xB6\xBA\x00\xEB\xB6\xBB"
-     "\x00\xEB"
-     "\xB6\xBC\x00\xEB\xB6\xBD\x00\xEB\xB6\xBE\x00\xEB\xB6\xBF\x00\xEB\xB7\x80"
-     "\x00\xEB"
-     "\xB7\x81\x00\xEB\xB7\x82\x00\xEB\xB7\x83\x00\xEB\xB7\x84\x00\xEB\xB7\x85"
-     "\x00\xEB"
-     "\xB7\x86\x00\xEB\xB7\x87\x00\xEB\xB7\x88\x00\xEB\xB7\x89\x00\xEB\xB7\x8A"
-     "\x00\xEB"
-     "\xB7\x8B\x00\xEB\xB7\x8C\x00\xEB\xB7\x8D\x00\xEB\xB7\x8E\x00\xEB\xB7\x8F"
-     "\x00\xEB"
-     "\xB7\x90\x00\xEB\xB7\x91\x00\xEB\xB7\x92\x00\xEB\xB7\x93\x00\xEB\xB7\x94"
-     "\x00\xEB"
-     "\xB7\x95\x00\xEB\xB7\x96\x00\xEB\xB7\x97\x00\xEB\xB7\x98\x00\xEB\xB7\x99"
-     "\x00\xEB"
-     "\xB7\x9A\x00\xEB\xB7\x9B\x00\xEB\xB7\x9C\x00\xEB\xB7\x9D\x00\xEB\xB7\x9E"
-     "\x00\xEB"
-     "\xB7\x9F\x00\xEB\xB7\xA0\x00\xEB\xB7\xA1\x00\xEB\xB7\xA2\x00\xEB\xB7\xA3"
-     "\x00\xEB"
-     "\xB7\xA4\x00\xEB\xB7\xA5\x00\xEB\xB7\xA6\x00\xEB\xB7\xA7\x00\xEB\xB7\xA8"
-     "\x00\xEB"
-     "\xB7\xA9\x00\xEB\xB7\xAA\x00\xEB\xB7\xAB\x00\xEB\xB7\xAC\x00\xEB\xB7\xAD"
-     "\x00\xEB"
-     "\xB7\xAE\x00\xEB\xB7\xAF\x00\xEB\xB7\xB0\x00\xEB\xB7\xB1\x00\xEB\xB7\xB2"
-     "\x00\xEB"
-     "\xB7\xB3\x00\xEB\xB7\xB4\x00\xEB\xB7\xB5\x00\xEB\xB7\xB6\x00\xEB\xB7\xB7"
-     "\x00\xEB"
-     "\xB7\xB8\x00\xEB\xB7\xB9\x00\xEB\xB7\xBA\x00\xEB\xB7\xBB\x00\xEB\xB7\xBC"
-     "\x00\xEB"
-     "\xB7\xBD\x00\xEB\xB7\xBE\x00\xEB\xB7\xBF\x00\xEB\xB8\x80\x00\xEB\xB8\x81"
-     "\x00\xEB"
-     "\xB8\x82\x00\xEB\xB8\x83\x00\xEB\xB8\x84\x00\xEB\xB8\x85\x00\xEB\xB8\x86"
-     "\x00\xEB"
-     "\xB8\x87\x00\xEB\xB8\x88\x00\xEB\xB8\x89\x00\xEB\xB8\x8A\x00\xEB\xB8\x8B"
-     "\x00\xEB"
-     "\xB8\x8C\x00\xEB\xB8\x8D\x00\xEB\xB8\x8E\x00\xEB\xB8\x8F\x00\xEB\xB8\x90"
-     "\x00\xEB"
-     "\xB8\x91\x00\xEB\xB8\x92\x00\xEB\xB8\x93\x00\xEB\xB8\x94\x00\xEB\xB8\x95"
-     "\x00\xEB"
-     "\xB8\x96\x00\xEB\xB8\x97\x00\xEB\xB8\x98\x00\xEB\xB8\x99\x00\xEB\xB8\x9A"
-     "\x00\xEB"
-     "\xB8\x9B\x00\xEB\xB8\x9C\x00\xEB\xB8\x9D\x00\xEB\xB8\x9E\x00\xEB\xB8\x9F"
-     "\x00\xEB"
-     "\xB8\xA0\x00\xEB\xB8\xA1\x00\xEB\xB8\xA2\x00\xEB\xB8\xA3\x00\xEB\xB8\xA4"
-     "\x00\xEB"
-     "\xB8\xA5\x00\xEB\xB8\xA6\x00\xEB\xB8\xA7\x00\xEB\xB8\xA8\x00\xEB\xB8\xA9"
-     "\x00\xEB"
-     "\xB8\xAA\x00\xEB\xB8\xAB\x00\xEB\xB8\xAC\x00\xEB\xB8\xAD\x00\xEB\xB8\xAE"
-     "\x00\xEB"
-     "\xB8\xAF\x00\xEB\xB8\xB0\x00\xEB\xB8\xB1\x00\xEB\xB8\xB2\x00\xEB\xB8\xB3"
-     "\x00\xEB"
-     "\xB8\xB4\x00\xEB\xB8\xB5\x00\xEB\xB8\xB6\x00\xEB\xB8\xB7\x00\xEB\xB8\xB8"
-     "\x00\xEB"
-     "\xB8\xB9\x00\xEB\xB8\xBA\x00\xEB\xB8\xBB\x00\xEB\xB8\xBC\x00\xEB\xB8\xBD"
-     "\x00\xEB"
-     "\xB8\xBE\x00\xEB\xB8\xBF\x00\xEB\xB9\x80\x00\xEB\xB9\x81\x00\xEB\xB9\x82"
-     "\x00\xEB"
-     "\xB9\x83\x00\xEB\xB9\x84\x00\xEB\xB9\x85\x00\xEB\xB9\x86\x00\xEB\xB9\x87"
-     "\x00\xEB"
-     "\xB9\x88\x00\xEB\xB9\x89\x00\xEB\xB9\x8A\x00\xEB\xB9\x8B\x00\xEB\xB9\x8C"
-     "\x00\xEB"
-     "\xB9\x8D\x00\xEB\xB9\x8E\x00\xEB\xB9\x8F\x00\xEB\xB9\x90\x00\xEB\xB9\x91"
-     "\x00\xEB"
-     "\xB9\x92\x00\xEB\xB9\x93\x00\xEB\xB9\x94\x00\xEB\xB9\x95\x00\xEB\xB9\x96"
-     "\x00\xEB"
-     "\xB9\x97\x00\xEB\xB9\x98\x00\xEB\xB9\x99\x00\xEB\xB9\x9A\x00\xEB\xB9\x9B"
-     "\x00\xEB"
-     "\xB9\x9C\x00\xEB\xB9\x9D\x00\xEB\xB9\x9E\x00\xEB\xB9\x9F\x00\xEB\xB9\xA0"
-     "\x00\xEB"
-     "\xB9\xA1\x00\xEB\xB9\xA2\x00\xEB\xB9\xA3\x00\xEB\xB9\xA4\x00\xEB\xB9\xA5"
-     "\x00\xEB"
-     "\xB9\xA6\x00\xEB\xB9\xA7\x00\xEB\xB9\xA8\x00\xEB\xB9\xA9\x00\xEB\xB9\xAA"
-     "\x00\xEB"
-     "\xB9\xAB\x00\xEB\xB9\xAC\x00\xEB\xB9\xAD\x00\xEB\xB9\xAE\x00\xEB\xB9\xAF"
-     "\x00\xEB"
-     "\xB9\xB0\x00\xEB\xB9\xB1\x00\xEB\xB9\xB2\x00\xEB\xB9\xB3\x00\xEB\xB9\xB4"
-     "\x00\xEB"
-     "\xB9\xB5\x00\xEB\xB9\xB6\x00\xEB\xB9\xB7\x00\xEB\xB9\xB8\x00\xEB\xB9\xB9"
-     "\x00\xEB"
-     "\xB9\xBA\x00\xEB\xB9\xBB\x00\xEB\xB9\xBC\x00\xEB\xB9\xBD\x00\xEB\xB9\xBE"
-     "\x00\xEB"
-     "\xB9\xBF\x00\xEB\xBA\x80\x00\xEB\xBA\x81\x00\xEB\xBA\x82\x00\xEB\xBA\x83"
-     "\x00\xEB"
-     "\xBA\x84\x00\xEB\xBA\x85\x00\xEB\xBA\x86\x00\xEB\xBA\x87\x00\xEB\xBA\x88"
-     "\x00\xEB"
-     "\xBA\x89\x00\xEB\xBA\x8A\x00\xEB\xBA\x8B\x00\xEB\xBA\x8C\x00\xEB\xBA\x8D"
-     "\x00\xEB"
-     "\xBA\x8E\x00\xEB\xBA\x8F\x00\xEB\xBA\x90\x00\xEB\xBA\x91\x00\xEB\xBA\x92"
-     "\x00\xEB"
-     "\xBA\x93\x00\xEB\xBA\x94\x00\xEB\xBA\x95\x00\xEB\xBA\x96\x00\xEB\xBA\x97"
-     "\x00\xEB"
-     "\xBA\x98\x00\xEB\xBA\x99\x00\xEB\xBA\x9A\x00\xEB\xBA\x9B\x00\xEB\xBA\x9C"
-     "\x00\xEB"
-     "\xBA\x9D\x00\xEB\xBA\x9E\x00\xEB\xBA\x9F\x00\xEB\xBA\xA0\x00\xEB\xBA\xA1"
-     "\x00\xEB"
-     "\xBA\xA2\x00\xEB\xBA\xA3\x00\xEB\xBA\xA4\x00\xEB\xBA\xA5\x00\xEB\xBA\xA6"
-     "\x00\xEB"
-     "\xBA\xA7\x00\xEB\xBA\xA8\x00\xEB\xBA\xA9\x00\xEB\xBA\xAA\x00\xEB\xBA\xAB"
-     "\x00\xEB"
-     "\xBA\xAC\x00\xEB\xBA\xAD\x00\xEB\xBA\xAE\x00\xEB\xBA\xAF\x00\xEB\xBA\xB0"
-     "\x00\xEB"
-     "\xBA\xB1\x00\xEB\xBA\xB2\x00\xEB\xBA\xB3\x00\xEB\xBA\xB4\x00\xEB\xBA\xB5"
-     "\x00\xEB"
-     "\xBA\xB6\x00\xEB\xBA\xB7\x00\xEB\xBA\xB8\x00\xEB\xBA\xB9\x00\xEB\xBA\xBA"
-     "\x00\xEB"
-     "\xBA\xBB\x00\xEB\xBA\xBC\x00\xEB\xBA\xBD\x00\xEB\xBA\xBE\x00\xEB\xBA\xBF"
-     "\x00\xEB"
-     "\xBB\x80\x00\xEB\xBB\x81\x00\xEB\xBB\x82\x00\xEB\xBB\x83\x00\xEB\xBB\x84"
-     "\x00\xEB"
-     "\xBB\x85\x00\xEB\xBB\x86\x00\xEB\xBB\x87\x00\xEB\xBB\x88\x00\xEB\xBB\x89"
-     "\x00\xEB"
-     "\xBB\x8A\x00\xEB\xBB\x8B\x00\xEB\xBB\x8C\x00\xEB\xBB\x8D\x00\xEB\xBB\x8E"
-     "\x00\xEB"
-     "\xBB\x8F\x00\xEB\xBB\x90\x00\xEB\xBB\x91\x00\xEB\xBB\x92\x00\xEB\xBB\x93"
-     "\x00\xEB"
-     "\xBB\x94\x00\xEB\xBB\x95\x00\xEB\xBB\x96\x00\xEB\xBB\x97\x00\xEB\xBB\x98"
-     "\x00\xEB"
-     "\xBB\x99\x00\xEB\xBB\x9A\x00\xEB\xBB\x9B\x00\xEB\xBB\x9C\x00\xEB\xBB\x9D"
-     "\x00\xEB"
-     "\xBB\x9E\x00\xEB\xBB\x9F\x00\xEB\xBB\xA0\x00\xEB\xBB\xA1\x00\xEB\xBB\xA2"
-     "\x00\xEB"
-     "\xBB\xA3\x00\xEB\xBB\xA4\x00\xEB\xBB\xA5\x00\xEB\xBB\xA6\x00\xEB\xBB\xA7"
-     "\x00\xEB"
-     "\xBB\xA8\x00\xEB\xBB\xA9\x00\xEB\xBB\xAA\x00\xEB\xBB\xAB\x00\xEB\xBB\xAC"
-     "\x00\xEB"
-     "\xBB\xAD\x00\xEB\xBB\xAE\x00\xEB\xBB\xAF\x00\xEB\xBB\xB0\x00\xEB\xBB\xB1"
-     "\x00\xEB"
-     "\xBB\xB2\x00\xEB\xBB\xB3\x00\xEB\xBB\xB4\x00\xEB\xBB\xB5\x00\xEB\xBB\xB6"
-     "\x00\xEB"
-     "\xBB\xB7\x00\xEB\xBB\xB8\x00\xEB\xBB\xB9\x00\xEB\xBB\xBA\x00\xEB\xBB\xBB"
-     "\x00\xEB"
-     "\xBB\xBC\x00\xEB\xBB\xBD\x00\xEB\xBB\xBE\x00\xEB\xBB\xBF\x00\xEB\xBC\x80"
-     "\x00\xEB"
-     "\xBC\x81\x00\xEB\xBC\x82\x00\xEB\xBC\x83\x00\xEB\xBC\x84\x00\xEB\xBC\x85"
-     "\x00\xEB"
-     "\xBC\x86\x00\xEB\xBC\x87\x00\xEB\xBC\x88\x00\xEB\xBC\x89\x00\xEB\xBC\x8A"
-     "\x00\xEB"
-     "\xBC\x8B\x00\xEB\xBC\x8C\x00\xEB\xBC\x8D\x00\xEB\xBC\x8E\x00\xEB\xBC\x8F"
-     "\x00\xEB"
-     "\xBC\x90\x00\xEB\xBC\x91\x00\xEB\xBC\x92\x00\xEB\xBC\x93\x00\xEB\xBC\x94"
-     "\x00\xEB"
-     "\xBC\x95\x00\xEB\xBC\x96\x00\xEB\xBC\x97\x00\xEB\xBC\x98\x00\xEB\xBC\x99"
-     "\x00\xEB"
-     "\xBC\x9A\x00\xEB\xBC\x9B\x00\xEB\xBC\x9C\x00\xEB\xBC\x9D\x00\xEB\xBC\x9E"
-     "\x00\xEB"
-     "\xBC\x9F\x00\xEB\xBC\xA0\x00\xEB\xBC\xA1\x00\xEB\xBC\xA2\x00\xEB\xBC\xA3"
-     "\x00\xEB"
-     "\xBC\xA4\x00\xEB\xBC\xA5\x00\xEB\xBC\xA6\x00\xEB\xBC\xA7\x00\xEB\xBC\xA8"
-     "\x00\xEB"
-     "\xBC\xA9\x00\xEB\xBC\xAA\x00\xEB\xBC\xAB\x00\xEB\xBC\xAC\x00\xEB\xBC\xAD"
-     "\x00\xEB"
-     "\xBC\xAE\x00\xEB\xBC\xAF\x00\xEB\xBC\xB0\x00\xEB\xBC\xB1\x00\xEB\xBC\xB2"
-     "\x00\xEB"
-     "\xBC\xB3\x00\xEB\xBC\xB4\x00\xEB\xBC\xB5\x00\xEB\xBC\xB6\x00\xEB\xBC\xB7"
-     "\x00\xEB"
-     "\xBC\xB8\x00\xEB\xBC\xB9\x00\xEB\xBC\xBA\x00\xEB\xBC\xBB\x00\xEB\xBC\xBC"
-     "\x00\xEB"
-     "\xBC\xBD\x00\xEB\xBC\xBE\x00\xEB\xBC\xBF\x00\xEB\xBD\x80\x00\xEB\xBD\x81"
-     "\x00\xEB"
-     "\xBD\x82\x00\xEB\xBD\x83\x00\xEB\xBD\x84\x00\xEB\xBD\x85\x00\xEB\xBD\x86"
-     "\x00\xEB"
-     "\xBD\x87\x00\xEB\xBD\x88\x00\xEB\xBD\x89\x00\xEB\xBD\x8A\x00\xEB\xBD\x8B"
-     "\x00\xEB"
-     "\xBD\x8C\x00\xEB\xBD\x8D\x00\xEB\xBD\x8E\x00\xEB\xBD\x8F\x00\xEB\xBD\x90"
-     "\x00\xEB"
-     "\xBD\x91\x00\xEB\xBD\x92\x00\xEB\xBD\x93\x00\xEB\xBD\x94\x00\xEB\xBD\x95"
-     "\x00\xEB"
-     "\xBD\x96\x00\xEB\xBD\x97\x00\xEB\xBD\x98\x00\xEB\xBD\x99\x00\xEB\xBD\x9A"
-     "\x00\xEB"
-     "\xBD\x9B\x00\xEB\xBD\x9C\x00\xEB\xBD\x9D\x00\xEB\xBD\x9E\x00\xEB\xBD\x9F"
-     "\x00\xEB"
-     "\xBD\xA0\x00\xEB\xBD\xA1\x00\xEB\xBD\xA2\x00\xEB\xBD\xA3\x00\xEB\xBD\xA4"
-     "\x00\xEB"
-     "\xBD\xA5\x00\xEB\xBD\xA6\x00\xEB\xBD\xA7\x00\xEB\xBD\xA8\x00\xEB\xBD\xA9"
-     "\x00\xEB"
-     "\xBD\xAA\x00\xEB\xBD\xAB\x00\xEB\xBD\xAC\x00\xEB\xBD\xAD\x00\xEB\xBD\xAE"
-     "\x00\xEB"
-     "\xBD\xAF\x00\xEB\xBD\xB0\x00\xEB\xBD\xB1\x00\xEB\xBD\xB2\x00\xEB\xBD\xB3"
-     "\x00\xEB"
-     "\xBD\xB4\x00\xEB\xBD\xB5\x00\xEB\xBD\xB6\x00\xEB\xBD\xB7\x00\xEB\xBD\xB8"
-     "\x00\xEB"
-     "\xBD\xB9\x00\xEB\xBD\xBA\x00\xEB\xBD\xBB\x00\xEB\xBD\xBC\x00\xEB\xBD\xBD"
-     "\x00\xEB"
-     "\xBD\xBE\x00\xEB\xBD\xBF\x00\xEB\xBE\x80\x00\xEB\xBE\x81\x00\xEB\xBE\x82"
-     "\x00\xEB"
-     "\xBE\x83\x00\xEB\xBE\x84\x00\xEB\xBE\x85\x00\xEB\xBE\x86\x00\xEB\xBE\x87"
-     "\x00\xEB"
-     "\xBE\x88\x00\xEB\xBE\x89\x00\xEB\xBE\x8A\x00\xEB\xBE\x8B\x00\xEB\xBE\x8C"
-     "\x00\xEB"
-     "\xBE\x8D\x00\xEB\xBE\x8E\x00\xEB\xBE\x8F\x00\xEB\xBE\x90\x00\xEB\xBE\x91"
-     "\x00\xEB"
-     "\xBE\x92\x00\xEB\xBE\x93\x00\xEB\xBE\x94\x00\xEB\xBE\x95\x00\xEB\xBE\x96"
-     "\x00\xEB"
-     "\xBE\x97\x00\xEB\xBE\x98\x00\xEB\xBE\x99\x00\xEB\xBE\x9A\x00\xEB\xBE\x9B"
-     "\x00\xEB"
-     "\xBE\x9C\x00\xEB\xBE\x9D\x00\xEB\xBE\x9E\x00\xEB\xBE\x9F\x00\xEB\xBE\xA0"
-     "\x00\xEB"
-     "\xBE\xA1\x00\xEB\xBE\xA2\x00\xEB\xBE\xA3\x00\xEB\xBE\xA4\x00\xEB\xBE\xA5"
-     "\x00\xEB"
-     "\xBE\xA6\x00\xEB\xBE\xA7\x00\xEB\xBE\xA8\x00\xEB\xBE\xA9\x00\xEB\xBE\xAA"
-     "\x00\xEB"
-     "\xBE\xAB\x00\xEB\xBE\xAC\x00\xEB\xBE\xAD\x00\xEB\xBE\xAE\x00\xEB\xBE\xAF"
-     "\x00\xEB"
-     "\xBE\xB0\x00\xEB\xBE\xB1\x00\xEB\xBE\xB2\x00\xEB\xBE\xB3\x00\xEB\xBE\xB4"
-     "\x00\xEB"
-     "\xBE\xB5\x00\xEB\xBE\xB6\x00\xEB\xBE\xB7\x00\xEB\xBE\xB8\x00\xEB\xBE\xB9"
-     "\x00\xEB"
-     "\xBE\xBA\x00\xEB\xBE\xBB\x00\xEB\xBE\xBC\x00\xEB\xBE\xBD\x00\xEB\xBE\xBE"
-     "\x00\xEB"
-     "\xBE\xBF\x00\xEB\xBF\x80\x00\xEB\xBF\x81\x00\xEB\xBF\x82\x00\xEB\xBF\x83"
-     "\x00\xEB"
-     "\xBF\x84\x00\xEB\xBF\x85\x00\xEB\xBF\x86\x00\xEB\xBF\x87\x00\xEB\xBF\x88"
-     "\x00\xEB"
-     "\xBF\x89\x00\xEB\xBF\x8A\x00\xEB\xBF\x8B\x00\xEB\xBF\x8C\x00\xEB\xBF\x8D"
-     "\x00\xEB"
-     "\xBF\x8E\x00\xEB\xBF\x8F\x00\xEB\xBF\x90\x00\xEB\xBF\x91\x00\xEB\xBF\x92"
-     "\x00\xEB"
-     "\xBF\x93\x00\xEB\xBF\x94\x00\xEB\xBF\x95\x00\xEB\xBF\x96\x00\xEB\xBF\x97"
-     "\x00\xEB"
-     "\xBF\x98\x00\xEB\xBF\x99\x00\xEB\xBF\x9A\x00\xEB\xBF\x9B\x00\xEB\xBF\x9C"
-     "\x00\xEB"
-     "\xBF\x9D\x00\xEB\xBF\x9E\x00\xEB\xBF\x9F\x00\xEB\xBF\xA0\x00\xEB\xBF\xA1"
-     "\x00\xEB"
-     "\xBF\xA2\x00\xEB\xBF\xA3\x00\xEB\xBF\xA4\x00\xEB\xBF\xA5\x00\xEB\xBF\xA6"
-     "\x00\xEB"
-     "\xBF\xA7\x00\xEB\xBF\xA8\x00\xEB\xBF\xA9\x00\xEB\xBF\xAA\x00\xEB\xBF\xAB"
-     "\x00\xEB"
-     "\xBF\xAC\x00\xEB\xBF\xAD\x00\xEB\xBF\xAE\x00\xEB\xBF\xAF\x00\xEB\xBF\xB0"
-     "\x00\xEB"
-     "\xBF\xB1\x00\xEB\xBF\xB2\x00\xEB\xBF\xB3\x00\xEB\xBF\xB4\x00\xEB\xBF\xB5"
-     "\x00\xEB"
-     "\xBF\xB6\x00\xEB\xBF\xB7\x00\xEB\xBF\xB8\x00\xEB\xBF\xB9\x00\xEB\xBF\xBA"
-     "\x00\xEB"
-     "\xBF\xBB\x00\xEB\xBF\xBC\x00\xEB\xBF\xBD\x00\xEB\xBF\xBE\x00\xEB\xBF\xBF"
-     "\x00\xEC"
-     "\x80\x80\x00\xEC\x80\x81\x00\xEC\x80\x82\x00\xEC\x80\x83\x00\xEC\x80\x84"
-     "\x00\xEC"
-     "\x80\x85\x00\xEC\x80\x86\x00\xEC\x80\x87\x00\xEC\x80\x88\x00\xEC\x80\x89"
-     "\x00\xEC"
-     "\x80\x8A\x00\xEC\x80\x8B\x00\xEC\x80\x8C\x00\xEC\x80\x8D\x00\xEC\x80\x8E"
-     "\x00\xEC"
-     "\x80\x8F\x00\xEC\x80\x90\x00\xEC\x80\x91\x00\xEC\x80\x92\x00\xEC\x80\x93"
-     "\x00\xEC"
-     "\x80\x94\x00\xEC\x80\x95\x00\xEC\x80\x96\x00\xEC\x80\x97\x00\xEC\x80\x98"
-     "\x00\xEC"
-     "\x80\x99\x00\xEC\x80\x9A\x00\xEC\x80\x9B\x00\xEC\x80\x9C\x00\xEC\x80\x9D"
-     "\x00\xEC"
-     "\x80\x9E\x00\xEC\x80\x9F\x00\xEC\x80\xA0\x00\xEC\x80\xA1\x00\xEC\x80\xA2"
-     "\x00\xEC"
-     "\x80\xA3\x00\xEC\x80\xA4\x00\xEC\x80\xA5\x00\xEC\x80\xA6\x00\xEC\x80\xA7"
-     "\x00\xEC"
-     "\x80\xA8\x00\xEC\x80\xA9\x00\xEC\x80\xAA\x00\xEC\x80\xAB\x00\xEC\x80\xAC"
-     "\x00\xEC"
-     "\x80\xAD\x00\xEC\x80\xAE\x00\xEC\x80\xAF\x00\xEC\x80\xB0\x00\xEC\x80\xB1"
-     "\x00\xEC"
-     "\x80\xB2\x00\xEC\x80\xB3\x00\xEC\x80\xB4\x00\xEC\x80\xB5\x00\xEC\x80\xB6"
-     "\x00\xEC"
-     "\x80\xB7\x00\xEC\x80\xB8\x00\xEC\x80\xB9\x00\xEC\x80\xBA\x00\xEC\x80\xBB"
-     "\x00\xEC"
-     "\x80\xBC\x00\xEC\x80\xBD\x00\xEC\x80\xBE\x00\xEC\x80\xBF\x00\xEC\x81\x80"
-     "\x00\xEC"
-     "\x81\x81\x00\xEC\x81\x82\x00\xEC\x81\x83\x00\xEC\x81\x84\x00\xEC\x81\x85"
-     "\x00\xEC"
-     "\x81\x86\x00\xEC\x81\x87\x00\xEC\x81\x88\x00\xEC\x81\x89\x00\xEC\x81\x8A"
-     "\x00\xEC"
-     "\x81\x8B\x00\xEC\x81\x8C\x00\xEC\x81\x8D\x00\xEC\x81\x8E\x00\xEC\x81\x8F"
-     "\x00\xEC"
-     "\x81\x90\x00\xEC\x81\x91\x00\xEC\x81\x92\x00\xEC\x81\x93\x00\xEC\x81\x94"
-     "\x00\xEC"
-     "\x81\x95\x00\xEC\x81\x96\x00\xEC\x81\x97\x00\xEC\x81\x98\x00\xEC\x81\x99"
-     "\x00\xEC"
-     "\x81\x9A\x00\xEC\x81\x9B\x00\xEC\x81\x9C\x00\xEC\x81\x9D\x00\xEC\x81\x9E"
-     "\x00\xEC"
-     "\x81\x9F\x00\xEC\x81\xA0\x00\xEC\x81\xA1\x00\xEC\x81\xA2\x00\xEC\x81\xA3"
-     "\x00\xEC"
-     "\x81\xA4\x00\xEC\x81\xA5\x00\xEC\x81\xA6\x00\xEC\x81\xA7\x00\xEC\x81\xA8"
-     "\x00\xEC"
-     "\x81\xA9\x00\xEC\x81\xAA\x00\xEC\x81\xAB\x00\xEC\x81\xAC\x00\xEC\x81\xAD"
-     "\x00\xEC"
-     "\x81\xAE\x00\xEC\x81\xAF\x00\xEC\x81\xB0\x00\xEC\x81\xB1\x00\xEC\x81\xB2"
-     "\x00\xEC"
-     "\x81\xB3\x00\xEC\x81\xB4\x00\xEC\x81\xB5\x00\xEC\x81\xB6\x00\xEC\x81\xB7"
-     "\x00\xEC"
-     "\x81\xB8\x00\xEC\x81\xB9\x00\xEC\x81\xBA\x00\xEC\x81\xBB\x00\xEC\x81\xBC"
-     "\x00\xEC"
-     "\x81\xBD\x00\xEC\x81\xBE\x00\xEC\x81\xBF\x00\xEC\x82\x80\x00\xEC\x82\x81"
-     "\x00\xEC"
-     "\x82\x82\x00\xEC\x82\x83\x00\xEC\x82\x84\x00\xEC\x82\x85\x00\xEC\x82\x86"
-     "\x00\xEC"
-     "\x82\x87\x00\xEC\x82\x88\x00\xEC\x82\x89\x00\xEC\x82\x8A\x00\xEC\x82\x8B"
-     "\x00\xEC"
-     "\x82\x8C\x00\xEC\x82\x8D\x00\xEC\x82\x8E\x00\xEC\x82\x8F\x00\xEC\x82\x90"
-     "\x00\xEC"
-     "\x82\x91\x00\xEC\x82\x92\x00\xEC\x82\x93\x00\xEC\x82\x94\x00\xEC\x82\x95"
-     "\x00\xEC"
-     "\x82\x96\x00\xEC\x82\x97\x00\xEC\x82\x98\x00\xEC\x82\x99\x00\xEC\x82\x9A"
-     "\x00\xEC"
-     "\x82\x9B\x00\xEC\x82\x9C\x00\xEC\x82\x9D\x00\xEC\x82\x9E\x00\xEC\x82\x9F"
-     "\x00\xEC"
-     "\x82\xA0\x00\xEC\x82\xA1\x00\xEC\x82\xA2\x00\xEC\x82\xA3\x00\xEC\x82\xA4"
-     "\x00\xEC"
-     "\x82\xA5\x00\xEC\x82\xA6\x00\xEC\x82\xA7\x00\xEC\x82\xA8\x00\xEC\x82\xA9"
-     "\x00\xEC"
-     "\x82\xAA\x00\xEC\x82\xAB\x00\xEC\x82\xAC\x00\xEC\x82\xAD\x00\xEC\x82\xAE"
-     "\x00\xEC"
-     "\x82\xAF\x00\xEC\x82\xB0\x00\xEC\x82\xB1\x00\xEC\x82\xB2\x00\xEC\x82\xB3"
-     "\x00\xEC"
-     "\x82\xB4\x00\xEC\x82\xB5\x00\xEC\x82\xB6\x00\xEC\x82\xB7\x00\xEC\x82\xB8"
-     "\x00\xEC"
-     "\x82\xB9\x00\xEC\x82\xBA\x00\xEC\x82\xBB\x00\xEC\x82\xBC\x00\xEC\x82\xBD"
-     "\x00\xEC"
-     "\x82\xBE\x00\xEC\x82\xBF\x00\xEC\x83\x80\x00\xEC\x83\x81\x00\xEC\x83\x82"
-     "\x00\xEC"
-     "\x83\x83\x00\xEC\x83\x84\x00\xEC\x83\x85\x00\xEC\x83\x86\x00\xEC\x83\x87"
-     "\x00\xEC"
-     "\x83\x88\x00\xEC\x83\x89\x00\xEC\x83\x8A\x00\xEC\x83\x8B\x00\xEC\x83\x8C"
-     "\x00\xEC"
-     "\x83\x8D\x00\xEC\x83\x8E\x00\xEC\x83\x8F\x00\xEC\x83\x90\x00\xEC\x83\x91"
-     "\x00\xEC"
-     "\x83\x92\x00\xEC\x83\x93\x00\xEC\x83\x94\x00\xEC\x83\x95\x00\xEC\x83\x96"
-     "\x00\xEC"
-     "\x83\x97\x00\xEC\x83\x98\x00\xEC\x83\x99\x00\xEC\x83\x9A\x00\xEC\x83\x9B"
-     "\x00\xEC"
-     "\x83\x9C\x00\xEC\x83\x9D\x00\xEC\x83\x9E\x00\xEC\x83\x9F\x00\xEC\x83\xA0"
-     "\x00\xEC"
-     "\x83\xA1\x00\xEC\x83\xA2\x00\xEC\x83\xA3\x00\xEC\x83\xA4\x00\xEC\x83\xA5"
-     "\x00\xEC"
-     "\x83\xA6\x00\xEC\x83\xA7\x00\xEC\x83\xA8\x00\xEC\x83\xA9\x00\xEC\x83\xAA"
-     "\x00\xEC"
-     "\x83\xAB\x00\xEC\x83\xAC\x00\xEC\x83\xAD\x00\xEC\x83\xAE\x00\xEC\x83\xAF"
-     "\x00\xEC"
-     "\x83\xB0\x00\xEC\x83\xB1\x00\xEC\x83\xB2\x00\xEC\x83\xB3\x00\xEC\x83\xB4"
-     "\x00\xEC"
-     "\x83\xB5\x00\xEC\x83\xB6\x00\xEC\x83\xB7\x00\xEC\x83\xB8\x00\xEC\x83\xB9"
-     "\x00\xEC"
-     "\x83\xBA\x00\xEC\x83\xBB\x00\xEC\x83\xBC\x00\xEC\x83\xBD\x00\xEC\x83\xBE"
-     "\x00\xEC"
-     "\x83\xBF\x00\xEC\x84\x80\x00\xEC\x84\x81\x00\xEC\x84\x82\x00\xEC\x84\x83"
-     "\x00\xEC"
-     "\x84\x84\x00\xEC\x84\x85\x00\xEC\x84\x86\x00\xEC\x84\x87\x00\xEC\x84\x88"
-     "\x00\xEC"
-     "\x84\x89\x00\xEC\x84\x8A\x00\xEC\x84\x8B\x00\xEC\x84\x8C\x00\xEC\x84\x8D"
-     "\x00\xEC"
-     "\x84\x8E\x00\xEC\x84\x8F\x00\xEC\x84\x90\x00\xEC\x84\x91\x00\xEC\x84\x92"
-     "\x00\xEC"
-     "\x84\x93\x00\xEC\x84\x94\x00\xEC\x84\x95\x00\xEC\x84\x96\x00\xEC\x84\x97"
-     "\x00\xEC"
-     "\x84\x98\x00\xEC\x84\x99\x00\xEC\x84\x9A\x00\xEC\x84\x9B\x00\xEC\x84\x9C"
-     "\x00\xEC"
-     "\x84\x9D\x00\xEC\x84\x9E\x00\xEC\x84\x9F\x00\xEC\x84\xA0\x00\xEC\x84\xA1"
-     "\x00\xEC"
-     "\x84\xA2\x00\xEC\x84\xA3\x00\xEC\x84\xA4\x00\xEC\x84\xA5\x00\xEC\x84\xA6"
-     "\x00\xEC"
-     "\x84\xA7\x00\xEC\x84\xA8\x00\xEC\x84\xA9\x00\xEC\x84\xAA\x00\xEC\x84\xAB"
-     "\x00\xEC"
-     "\x84\xAC\x00\xEC\x84\xAD\x00\xEC\x84\xAE\x00\xEC\x84\xAF\x00\xEC\x84\xB0"
-     "\x00\xEC"
-     "\x84\xB1\x00\xEC\x84\xB2\x00\xEC\x84\xB3\x00\xEC\x84\xB4\x00\xEC\x84\xB5"
-     "\x00\xEC"
-     "\x84\xB6\x00\xEC\x84\xB7\x00\xEC\x84\xB8\x00\xEC\x84\xB9\x00\xEC\x84\xBA"
-     "\x00\xEC"
-     "\x84\xBB\x00\xEC\x84\xBC\x00\xEC\x84\xBD\x00\xEC\x84\xBE\x00\xEC\x84\xBF"
-     "\x00\xEC"
-     "\x85\x80\x00\xEC\x85\x81\x00\xEC\x85\x82\x00\xEC\x85\x83\x00\xEC\x85\x84"
-     "\x00\xEC"
-     "\x85\x85\x00\xEC\x85\x86\x00\xEC\x85\x87\x00\xEC\x85\x88\x00\xEC\x85\x89"
-     "\x00\xEC"
-     "\x85\x8A\x00\xEC\x85\x8B\x00\xEC\x85\x8C\x00\xEC\x85\x8D\x00\xEC\x85\x8E"
-     "\x00\xEC"
-     "\x85\x8F\x00\xEC\x85\x90\x00\xEC\x85\x91\x00\xEC\x85\x92\x00\xEC\x85\x93"
-     "\x00\xEC"
-     "\x85\x94\x00\xEC\x85\x95\x00\xEC\x85\x96\x00\xEC\x85\x97\x00\xEC\x85\x98"
-     "\x00\xEC"
-     "\x85\x99\x00\xEC\x85\x9A\x00\xEC\x85\x9B\x00\xEC\x85\x9C\x00\xEC\x85\x9D"
-     "\x00\xEC"
-     "\x85\x9E\x00\xEC\x85\x9F\x00\xEC\x85\xA0\x00\xEC\x85\xA1\x00\xEC\x85\xA2"
-     "\x00\xEC"
-     "\x85\xA3\x00\xEC\x85\xA4\x00\xEC\x85\xA5\x00\xEC\x85\xA6\x00\xEC\x85\xA7"
-     "\x00\xEC"
-     "\x85\xA8\x00\xEC\x85\xA9\x00\xEC\x85\xAA\x00\xEC\x85\xAB\x00\xEC\x85\xAC"
-     "\x00\xEC"
-     "\x85\xAD\x00\xEC\x85\xAE\x00\xEC\x85\xAF\x00\xEC\x85\xB0\x00\xEC\x85\xB1"
-     "\x00\xEC"
-     "\x85\xB2\x00\xEC\x85\xB3\x00\xEC\x85\xB4\x00\xEC\x85\xB5\x00\xEC\x85\xB6"
-     "\x00\xEC"
-     "\x85\xB7\x00\xEC\x85\xB8\x00\xEC\x85\xB9\x00\xEC\x85\xBA\x00\xEC\x85\xBB"
-     "\x00\xEC"
-     "\x85\xBC\x00\xEC\x85\xBD\x00\xEC\x85\xBE\x00\xEC\x85\xBF\x00\xEC\x86\x80"
-     "\x00\xEC"
-     "\x86\x81\x00\xEC\x86\x82\x00\xEC\x86\x83\x00\xEC\x86\x84\x00\xEC\x86\x85"
-     "\x00\xEC"
-     "\x86\x86\x00\xEC\x86\x87\x00\xEC\x86\x88\x00\xEC\x86\x89\x00\xEC\x86\x8A"
-     "\x00\xEC"
-     "\x86\x8B\x00\xEC\x86\x8C\x00\xEC\x86\x8D\x00\xEC\x86\x8E\x00\xEC\x86\x8F"
-     "\x00\xEC"
-     "\x86\x90\x00\xEC\x86\x91\x00\xEC\x86\x92\x00\xEC\x86\x93\x00\xEC\x86\x94"
-     "\x00\xEC"
-     "\x86\x95\x00\xEC\x86\x96\x00\xEC\x86\x97\x00\xEC\x86\x98\x00\xEC\x86\x99"
-     "\x00\xEC"
-     "\x86\x9A\x00\xEC\x86\x9B\x00\xEC\x86\x9C\x00\xEC\x86\x9D\x00\xEC\x86\x9E"
-     "\x00\xEC"
-     "\x86\x9F\x00\xEC\x86\xA0\x00\xEC\x86\xA1\x00\xEC\x86\xA2\x00\xEC\x86\xA3"
-     "\x00\xEC"
-     "\x86\xA4\x00\xEC\x86\xA5\x00\xEC\x86\xA6\x00\xEC\x86\xA7\x00\xEC\x86\xA8"
-     "\x00\xEC"
-     "\x86\xA9\x00\xEC\x86\xAA\x00\xEC\x86\xAB\x00\xEC\x86\xAC\x00\xEC\x86\xAD"
-     "\x00\xEC"
-     "\x86\xAE\x00\xEC\x86\xAF\x00\xEC\x86\xB0\x00\xEC\x86\xB1\x00\xEC\x86\xB2"
-     "\x00\xEC"
-     "\x86\xB3\x00\xEC\x86\xB4\x00\xEC\x86\xB5\x00\xEC\x86\xB6\x00\xEC\x86\xB7"
-     "\x00\xEC"
-     "\x86\xB8\x00\xEC\x86\xB9\x00\xEC\x86\xBA\x00\xEC\x86\xBB\x00\xEC\x86\xBC"
-     "\x00\xEC"
-     "\x86\xBD\x00\xEC\x86\xBE\x00\xEC\x86\xBF\x00\xEC\x87\x80\x00\xEC\x87\x81"
-     "\x00\xEC"
-     "\x87\x82\x00\xEC\x87\x83\x00\xEC\x87\x84\x00\xEC\x87\x85\x00\xEC\x87\x86"
-     "\x00\xEC"
-     "\x87\x87\x00\xEC\x87\x88\x00\xEC\x87\x89\x00\xEC\x87\x8A\x00\xEC\x87\x8B"
-     "\x00\xEC"
-     "\x87\x8C\x00\xEC\x87\x8D\x00\xEC\x87\x8E\x00\xEC\x87\x8F\x00\xEC\x87\x90"
-     "\x00\xEC"
-     "\x87\x91\x00\xEC\x87\x92\x00\xEC\x87\x93\x00\xEC\x87\x94\x00\xEC\x87\x95"
-     "\x00\xEC"
-     "\x87\x96\x00\xEC\x87\x97\x00\xEC\x87\x98\x00\xEC\x87\x99\x00\xEC\x87\x9A"
-     "\x00\xEC"
-     "\x87\x9B\x00\xEC\x87\x9C\x00\xEC\x87\x9D\x00\xEC\x87\x9E\x00\xEC\x87\x9F"
-     "\x00\xEC"
-     "\x87\xA0\x00\xEC\x87\xA1\x00\xEC\x87\xA2\x00\xEC\x87\xA3\x00\xEC\x87\xA4"
-     "\x00\xEC"
-     "\x87\xA5\x00\xEC\x87\xA6\x00\xEC\x87\xA7\x00\xEC\x87\xA8\x00\xEC\x87\xA9"
-     "\x00\xEC"
-     "\x87\xAA\x00\xEC\x87\xAB\x00\xEC\x87\xAC\x00\xEC\x87\xAD\x00\xEC\x87\xAE"
-     "\x00\xEC"
-     "\x87\xAF\x00\xEC\x87\xB0\x00\xEC\x87\xB1\x00\xEC\x87\xB2\x00\xEC\x87\xB3"
-     "\x00\xEC"
-     "\x87\xB4\x00\xEC\x87\xB5\x00\xEC\x87\xB6\x00\xEC\x87\xB7\x00\xEC\x87\xB8"
-     "\x00\xEC"
-     "\x87\xB9\x00\xEC\x87\xBA\x00\xEC\x87\xBB\x00\xEC\x87\xBC\x00\xEC\x87\xBD"
-     "\x00\xEC"
-     "\x87\xBE\x00\xEC\x87\xBF\x00\xEC\x88\x80\x00\xEC\x88\x81\x00\xEC\x88\x82"
-     "\x00\xEC"
-     "\x88\x83\x00\xEC\x88\x84\x00\xEC\x88\x85\x00\xEC\x88\x86\x00\xEC\x88\x87"
-     "\x00\xEC"
-     "\x88\x88\x00\xEC\x88\x89\x00\xEC\x88\x8A\x00\xEC\x88\x8B\x00\xEC\x88\x8C"
-     "\x00\xEC"
-     "\x88\x8D\x00\xEC\x88\x8E\x00\xEC\x88\x8F\x00\xEC\x88\x90\x00\xEC\x88\x91"
-     "\x00\xEC"
-     "\x88\x92\x00\xEC\x88\x93\x00\xEC\x88\x94\x00\xEC\x88\x95\x00\xEC\x88\x96"
-     "\x00\xEC"
-     "\x88\x97\x00\xEC\x88\x98\x00\xEC\x88\x99\x00\xEC\x88\x9A\x00\xEC\x88\x9B"
-     "\x00\xEC"
-     "\x88\x9C\x00\xEC\x88\x9D\x00\xEC\x88\x9E\x00\xEC\x88\x9F\x00\xEC\x88\xA0"
-     "\x00\xEC"
-     "\x88\xA1\x00\xEC\x88\xA2\x00\xEC\x88\xA3\x00\xEC\x88\xA4\x00\xEC\x88\xA5"
-     "\x00\xEC"
-     "\x88\xA6\x00\xEC\x88\xA7\x00\xEC\x88\xA8\x00\xEC\x88\xA9\x00\xEC\x88\xAA"
-     "\x00\xEC"
-     "\x88\xAB\x00\xEC\x88\xAC\x00\xEC\x88\xAD\x00\xEC\x88\xAE\x00\xEC\x88\xAF"
-     "\x00\xEC"
-     "\x88\xB0\x00\xEC\x88\xB1\x00\xEC\x88\xB2\x00\xEC\x88\xB3\x00\xEC\x88\xB4"
-     "\x00\xEC"
-     "\x88\xB5\x00\xEC\x88\xB6\x00\xEC\x88\xB7\x00\xEC\x88\xB8\x00\xEC\x88\xB9"
-     "\x00\xEC"
-     "\x88\xBA\x00\xEC\x88\xBB\x00\xEC\x88\xBC\x00\xEC\x88\xBD\x00\xEC\x88\xBE"
-     "\x00\xEC"
-     "\x88\xBF\x00\xEC\x89\x80\x00\xEC\x89\x81\x00\xEC\x89\x82\x00\xEC\x89\x83"
-     "\x00\xEC"
-     "\x89\x84\x00\xEC\x89\x85\x00\xEC\x89\x86\x00\xEC\x89\x87\x00\xEC\x89\x88"
-     "\x00\xEC"
-     "\x89\x89\x00\xEC\x89\x8A\x00\xEC\x89\x8B\x00\xEC\x89\x8C\x00\xEC\x89\x8D"
-     "\x00\xEC"
-     "\x89\x8E\x00\xEC\x89\x8F\x00\xEC\x89\x90\x00\xEC\x89\x91\x00\xEC\x89\x92"
-     "\x00\xEC"
-     "\x89\x93\x00\xEC\x89\x94\x00\xEC\x89\x95\x00\xEC\x89\x96\x00\xEC\x89\x97"
-     "\x00\xEC"
-     "\x89\x98\x00\xEC\x89\x99\x00\xEC\x89\x9A\x00\xEC\x89\x9B\x00\xEC\x89\x9C"
-     "\x00\xEC"
-     "\x89\x9D\x00\xEC\x89\x9E\x00\xEC\x89\x9F\x00\xEC\x89\xA0\x00\xEC\x89\xA1"
-     "\x00\xEC"
-     "\x89\xA2\x00\xEC\x89\xA3\x00\xEC\x89\xA4\x00\xEC\x89\xA5\x00\xEC\x89\xA6"
-     "\x00\xEC"
-     "\x89\xA7\x00\xEC\x89\xA8\x00\xEC\x89\xA9\x00\xEC\x89\xAA\x00\xEC\x89\xAB"
-     "\x00\xEC"
-     "\x89\xAC\x00\xEC\x89\xAD\x00\xEC\x89\xAE\x00\xEC\x89\xAF\x00\xEC\x89\xB0"
-     "\x00\xEC"
-     "\x89\xB1\x00\xEC\x89\xB2\x00\xEC\x89\xB3\x00\xEC\x89\xB4\x00\xEC\x89\xB5"
-     "\x00\xEC"
-     "\x89\xB6\x00\xEC\x89\xB7\x00\xEC\x89\xB8\x00\xEC\x89\xB9\x00\xEC\x89\xBA"
-     "\x00\xEC"
-     "\x89\xBB\x00\xEC\x89\xBC\x00\xEC\x89\xBD\x00\xEC\x89\xBE\x00\xEC\x89\xBF"
-     "\x00\xEC"
-     "\x8A\x80\x00\xEC\x8A\x81\x00\xEC\x8A\x82\x00\xEC\x8A\x83\x00\xEC\x8A\x84"
-     "\x00\xEC"
-     "\x8A\x85\x00\xEC\x8A\x86\x00\xEC\x8A\x87\x00\xEC\x8A\x88\x00\xEC\x8A\x89"
-     "\x00\xEC"
-     "\x8A\x8A\x00\xEC\x8A\x8B\x00\xEC\x8A\x8C\x00\xEC\x8A\x8D\x00\xEC\x8A\x8E"
-     "\x00\xEC"
-     "\x8A\x8F\x00\xEC\x8A\x90\x00\xEC\x8A\x91\x00\xEC\x8A\x92\x00\xEC\x8A\x93"
-     "\x00\xEC"
-     "\x8A\x94\x00\xEC\x8A\x95\x00\xEC\x8A\x96\x00\xEC\x8A\x97\x00\xEC\x8A\x98"
-     "\x00\xEC"
-     "\x8A\x99\x00\xEC\x8A\x9A\x00\xEC\x8A\x9B\x00\xEC\x8A\x9C\x00\xEC\x8A\x9D"
-     "\x00\xEC"
-     "\x8A\x9E\x00\xEC\x8A\x9F\x00\xEC\x8A\xA0\x00\xEC\x8A\xA1\x00\xEC\x8A\xA2"
-     "\x00\xEC"
-     "\x8A\xA3\x00\xEC\x8A\xA4\x00\xEC\x8A\xA5\x00\xEC\x8A\xA6\x00\xEC\x8A\xA7"
-     "\x00\xEC"
-     "\x8A\xA8\x00\xEC\x8A\xA9\x00\xEC\x8A\xAA\x00\xEC\x8A\xAB\x00\xEC\x8A\xAC"
-     "\x00\xEC"
-     "\x8A\xAD\x00\xEC\x8A\xAE\x00\xEC\x8A\xAF\x00\xEC\x8A\xB0\x00\xEC\x8A\xB1"
-     "\x00\xEC"
-     "\x8A\xB2\x00\xEC\x8A\xB3\x00\xEC\x8A\xB4\x00\xEC\x8A\xB5\x00\xEC\x8A\xB6"
-     "\x00\xEC"
-     "\x8A\xB7\x00\xEC\x8A\xB8\x00\xEC\x8A\xB9\x00\xEC\x8A\xBA\x00\xEC\x8A\xBB"
-     "\x00\xEC"
-     "\x8A\xBC\x00\xEC\x8A\xBD\x00\xEC\x8A\xBE\x00\xEC\x8A\xBF\x00\xEC\x8B\x80"
-     "\x00\xEC"
-     "\x8B\x81\x00\xEC\x8B\x82\x00\xEC\x8B\x83\x00\xEC\x8B\x84\x00\xEC\x8B\x85"
-     "\x00\xEC"
-     "\x8B\x86\x00\xEC\x8B\x87\x00\xEC\x8B\x88\x00\xEC\x8B\x89\x00\xEC\x8B\x8A"
-     "\x00\xEC"
-     "\x8B\x8B\x00\xEC\x8B\x8C\x00\xEC\x8B\x8D\x00\xEC\x8B\x8E\x00\xEC\x8B\x8F"
-     "\x00\xEC"
-     "\x8B\x90\x00\xEC\x8B\x91\x00\xEC\x8B\x92\x00\xEC\x8B\x93\x00\xEC\x8B\x94"
-     "\x00\xEC"
-     "\x8B\x95\x00\xEC\x8B\x96\x00\xEC\x8B\x97\x00\xEC\x8B\x98\x00\xEC\x8B\x99"
-     "\x00\xEC"
-     "\x8B\x9A\x00\xEC\x8B\x9B\x00\xEC\x8B\x9C\x00\xEC\x8B\x9D\x00\xEC\x8B\x9E"
-     "\x00\xEC"
-     "\x8B\x9F\x00\xEC\x8B\xA0\x00\xEC\x8B\xA1\x00\xEC\x8B\xA2\x00\xEC\x8B\xA3"
-     "\x00\xEC"
-     "\x8B\xA4\x00\xEC\x8B\xA5\x00\xEC\x8B\xA6\x00\xEC\x8B\xA7\x00\xEC\x8B\xA8"
-     "\x00\xEC"
-     "\x8B\xA9\x00\xEC\x8B\xAA\x00\xEC\x8B\xAB\x00\xEC\x8B\xAC\x00\xEC\x8B\xAD"
-     "\x00\xEC"
-     "\x8B\xAE\x00\xEC\x8B\xAF\x00\xEC\x8B\xB0\x00\xEC\x8B\xB1\x00\xEC\x8B\xB2"
-     "\x00\xEC"
-     "\x8B\xB3\x00\xEC\x8B\xB4\x00\xEC\x8B\xB5\x00\xEC\x8B\xB6\x00\xEC\x8B\xB7"
-     "\x00\xEC"
-     "\x8B\xB8\x00\xEC\x8B\xB9\x00\xEC\x8B\xBA\x00\xEC\x8B\xBB\x00\xEC\x8B\xBC"
-     "\x00\xEC"
-     "\x8B\xBD\x00\xEC\x8B\xBE\x00\xEC\x8B\xBF\x00\xEC\x8C\x80\x00\xEC\x8C\x81"
-     "\x00\xEC"
-     "\x8C\x82\x00\xEC\x8C\x83\x00\xEC\x8C\x84\x00\xEC\x8C\x85\x00\xEC\x8C\x86"
-     "\x00\xEC"
-     "\x8C\x87\x00\xEC\x8C\x88\x00\xEC\x8C\x89\x00\xEC\x8C\x8A\x00\xEC\x8C\x8B"
-     "\x00\xEC"
-     "\x8C\x8C\x00\xEC\x8C\x8D\x00\xEC\x8C\x8E\x00\xEC\x8C\x8F\x00\xEC\x8C\x90"
-     "\x00\xEC"
-     "\x8C\x91\x00\xEC\x8C\x92\x00\xEC\x8C\x93\x00\xEC\x8C\x94\x00\xEC\x8C\x95"
-     "\x00\xEC"
-     "\x8C\x96\x00\xEC\x8C\x97\x00\xEC\x8C\x98\x00\xEC\x8C\x99\x00\xEC\x8C\x9A"
-     "\x00\xEC"
-     "\x8C\x9B\x00\xEC\x8C\x9C\x00\xEC\x8C\x9D\x00\xEC\x8C\x9E\x00\xEC\x8C\x9F"
-     "\x00\xEC"
-     "\x8C\xA0\x00\xEC\x8C\xA1\x00\xEC\x8C\xA2\x00\xEC\x8C\xA3\x00\xEC\x8C\xA4"
-     "\x00\xEC"
-     "\x8C\xA5\x00\xEC\x8C\xA6\x00\xEC\x8C\xA7\x00\xEC\x8C\xA8\x00\xEC\x8C\xA9"
-     "\x00\xEC"
-     "\x8C\xAA\x00\xEC\x8C\xAB\x00\xEC\x8C\xAC\x00\xEC\x8C\xAD\x00\xEC\x8C\xAE"
-     "\x00\xEC"
-     "\x8C\xAF\x00\xEC\x8C\xB0\x00\xEC\x8C\xB1\x00\xEC\x8C\xB2\x00\xEC\x8C\xB3"
-     "\x00\xEC"
-     "\x8C\xB4\x00\xEC\x8C\xB5\x00\xEC\x8C\xB6\x00\xEC\x8C\xB7\x00\xEC\x8C\xB8"
-     "\x00\xEC"
-     "\x8C\xB9\x00\xEC\x8C\xBA\x00\xEC\x8C\xBB\x00\xEC\x8C\xBC\x00\xEC\x8C\xBD"
-     "\x00\xEC"
-     "\x8C\xBE\x00\xEC\x8C\xBF\x00\xEC\x8D\x80\x00\xEC\x8D\x81\x00\xEC\x8D\x82"
-     "\x00\xEC"
-     "\x8D\x83\x00\xEC\x8D\x84\x00\xEC\x8D\x85\x00\xEC\x8D\x86\x00\xEC\x8D\x87"
-     "\x00\xEC"
-     "\x8D\x88\x00\xEC\x8D\x89\x00\xEC\x8D\x8A\x00\xEC\x8D\x8B\x00\xEC\x8D\x8C"
-     "\x00\xEC"
-     "\x8D\x8D\x00\xEC\x8D\x8E\x00\xEC\x8D\x8F\x00\xEC\x8D\x90\x00\xEC\x8D\x91"
-     "\x00\xEC"
-     "\x8D\x92\x00\xEC\x8D\x93\x00\xEC\x8D\x94\x00\xEC\x8D\x95\x00\xEC\x8D\x96"
-     "\x00\xEC"
-     "\x8D\x97\x00\xEC\x8D\x98\x00\xEC\x8D\x99\x00\xEC\x8D\x9A\x00\xEC\x8D\x9B"
-     "\x00\xEC"
-     "\x8D\x9C\x00\xEC\x8D\x9D\x00\xEC\x8D\x9E\x00\xEC\x8D\x9F\x00\xEC\x8D\xA0"
-     "\x00\xEC"
-     "\x8D\xA1\x00\xEC\x8D\xA2\x00\xEC\x8D\xA3\x00\xEC\x8D\xA4\x00\xEC\x8D\xA5"
-     "\x00\xEC"
-     "\x8D\xA6\x00\xEC\x8D\xA7\x00\xEC\x8D\xA8\x00\xEC\x8D\xA9\x00\xEC\x8D\xAA"
-     "\x00\xEC"
-     "\x8D\xAB\x00\xEC\x8D\xAC\x00\xEC\x8D\xAD\x00\xEC\x8D\xAE\x00\xEC\x8D\xAF"
-     "\x00\xEC"
-     "\x8D\xB0\x00\xEC\x8D\xB1\x00\xEC\x8D\xB2\x00\xEC\x8D\xB3\x00\xEC\x8D\xB4"
-     "\x00\xEC"
-     "\x8D\xB5\x00\xEC\x8D\xB6\x00\xEC\x8D\xB7\x00\xEC\x8D\xB8\x00\xEC\x8D\xB9"
-     "\x00\xEC"
-     "\x8D\xBA\x00\xEC\x8D\xBB\x00\xEC\x8D\xBC\x00\xEC\x8D\xBD\x00\xEC\x8D\xBE"
-     "\x00\xEC"
-     "\x8D\xBF\x00\xEC\x8E\x80\x00\xEC\x8E\x81\x00\xEC\x8E\x82\x00\xEC\x8E\x83"
-     "\x00\xEC"
-     "\x8E\x84\x00\xEC\x8E\x85\x00\xEC\x8E\x86\x00\xEC\x8E\x87\x00\xEC\x8E\x88"
-     "\x00\xEC"
-     "\x8E\x89\x00\xEC\x8E\x8A\x00\xEC\x8E\x8B\x00\xEC\x8E\x8C\x00\xEC\x8E\x8D"
-     "\x00\xEC"
-     "\x8E\x8E\x00\xEC\x8E\x8F\x00\xEC\x8E\x90\x00\xEC\x8E\x91\x00\xEC\x8E\x92"
-     "\x00\xEC"
-     "\x8E\x93\x00\xEC\x8E\x94\x00\xEC\x8E\x95\x00\xEC\x8E\x96\x00\xEC\x8E\x97"
-     "\x00\xEC"
-     "\x8E\x98\x00\xEC\x8E\x99\x00\xEC\x8E\x9A\x00\xEC\x8E\x9B\x00\xEC\x8E\x9C"
-     "\x00\xEC"
-     "\x8E\x9D\x00\xEC\x8E\x9E\x00\xEC\x8E\x9F\x00\xEC\x8E\xA0\x00\xEC\x8E\xA1"
-     "\x00\xEC"
-     "\x8E\xA2\x00\xEC\x8E\xA3\x00\xEC\x8E\xA4\x00\xEC\x8E\xA5\x00\xEC\x8E\xA6"
-     "\x00\xEC"
-     "\x8E\xA7\x00\xEC\x8E\xA8\x00\xEC\x8E\xA9\x00\xEC\x8E\xAA\x00\xEC\x8E\xAB"
-     "\x00\xEC"
-     "\x8E\xAC\x00\xEC\x8E\xAD\x00\xEC\x8E\xAE\x00\xEC\x8E\xAF\x00\xEC\x8E\xB0"
-     "\x00\xEC"
-     "\x8E\xB1\x00\xEC\x8E\xB2\x00\xEC\x8E\xB3\x00\xEC\x8E\xB4\x00\xEC\x8E\xB5"
-     "\x00\xEC"
-     "\x8E\xB6\x00\xEC\x8E\xB7\x00\xEC\x8E\xB8\x00\xEC\x8E\xB9\x00\xEC\x8E\xBA"
-     "\x00\xEC"
-     "\x8E\xBB\x00\xEC\x8E\xBC\x00\xEC\x8E\xBD\x00\xEC\x8E\xBE\x00\xEC\x8E\xBF"
-     "\x00\xEC"
-     "\x8F\x80\x00\xEC\x8F\x81\x00\xEC\x8F\x82\x00\xEC\x8F\x83\x00\xEC\x8F\x84"
-     "\x00\xEC"
-     "\x8F\x85\x00\xEC\x8F\x86\x00\xEC\x8F\x87\x00\xEC\x8F\x88\x00\xEC\x8F\x89"
-     "\x00\xEC"
-     "\x8F\x8A\x00\xEC\x8F\x8B\x00\xEC\x8F\x8C\x00\xEC\x8F\x8D\x00\xEC\x8F\x8E"
-     "\x00\xEC"
-     "\x8F\x8F\x00\xEC\x8F\x90\x00\xEC\x8F\x91\x00\xEC\x8F\x92\x00\xEC\x8F\x93"
-     "\x00\xEC"
-     "\x8F\x94\x00\xEC\x8F\x95\x00\xEC\x8F\x96\x00\xEC\x8F\x97\x00\xEC\x8F\x98"
-     "\x00\xEC"
-     "\x8F\x99\x00\xEC\x8F\x9A\x00\xEC\x8F\x9B\x00\xEC\x8F\x9C\x00\xEC\x8F\x9D"
-     "\x00\xEC"
-     "\x8F\x9E\x00\xEC\x8F\x9F\x00\xEC\x8F\xA0\x00\xEC\x8F\xA1\x00\xEC\x8F\xA2"
-     "\x00\xEC"
-     "\x8F\xA3\x00\xEC\x8F\xA4\x00\xEC\x8F\xA5\x00\xEC\x8F\xA6\x00\xEC\x8F\xA7"
-     "\x00\xEC"
-     "\x8F\xA8\x00\xEC\x8F\xA9\x00\xEC\x8F\xAA\x00\xEC\x8F\xAB\x00\xEC\x8F\xAC"
-     "\x00\xEC"
-     "\x8F\xAD\x00\xEC\x8F\xAE\x00\xEC\x8F\xAF\x00\xEC\x8F\xB0\x00\xEC\x8F\xB1"
-     "\x00\xEC"
-     "\x8F\xB2\x00\xEC\x8F\xB3\x00\xEC\x8F\xB4\x00\xEC\x8F\xB5\x00\xEC\x8F\xB6"
-     "\x00\xEC"
-     "\x8F\xB7\x00\xEC\x8F\xB8\x00\xEC\x8F\xB9\x00\xEC\x8F\xBA\x00\xEC\x8F\xBB"
-     "\x00\xEC"
-     "\x8F\xBC\x00\xEC\x8F\xBD\x00\xEC\x8F\xBE\x00\xEC\x8F\xBF\x00\xEC\x90\x80"
-     "\x00\xEC"
-     "\x90\x81\x00\xEC\x90\x82\x00\xEC\x90\x83\x00\xEC\x90\x84\x00\xEC\x90\x85"
-     "\x00\xEC"
-     "\x90\x86\x00\xEC\x90\x87\x00\xEC\x90\x88\x00\xEC\x90\x89\x00\xEC\x90\x8A"
-     "\x00\xEC"
-     "\x90\x8B\x00\xEC\x90\x8C\x00\xEC\x90\x8D\x00\xEC\x90\x8E\x00\xEC\x90\x8F"
-     "\x00\xEC"
-     "\x90\x90\x00\xEC\x90\x91\x00\xEC\x90\x92\x00\xEC\x90\x93\x00\xEC\x90\x94"
-     "\x00\xEC"
-     "\x90\x95\x00\xEC\x90\x96\x00\xEC\x90\x97\x00\xEC\x90\x98\x00\xEC\x90\x99"
-     "\x00\xEC"
-     "\x90\x9A\x00\xEC\x90\x9B\x00\xEC\x90\x9C\x00\xEC\x90\x9D\x00\xEC\x90\x9E"
-     "\x00\xEC"
-     "\x90\x9F\x00\xEC\x90\xA0\x00\xEC\x90\xA1\x00\xEC\x90\xA2\x00\xEC\x90\xA3"
-     "\x00\xEC"
-     "\x90\xA4\x00\xEC\x90\xA5\x00\xEC\x90\xA6\x00\xEC\x90\xA7\x00\xEC\x90\xA8"
-     "\x00\xEC"
-     "\x90\xA9\x00\xEC\x90\xAA\x00\xEC\x90\xAB\x00\xEC\x90\xAC\x00\xEC\x90\xAD"
-     "\x00\xEC"
-     "\x90\xAE\x00\xEC\x90\xAF\x00\xEC\x90\xB0\x00\xEC\x90\xB1\x00\xEC\x90\xB2"
-     "\x00\xEC"
-     "\x90\xB3\x00\xEC\x90\xB4\x00\xEC\x90\xB5\x00\xEC\x90\xB6\x00\xEC\x90\xB7"
-     "\x00\xEC"
-     "\x90\xB8\x00\xEC\x90\xB9\x00\xEC\x90\xBA\x00\xEC\x90\xBB\x00\xEC\x90\xBC"
-     "\x00\xEC"
-     "\x90\xBD\x00\xEC\x90\xBE\x00\xEC\x90\xBF\x00\xEC\x91\x80\x00\xEC\x91\x81"
-     "\x00\xEC"
-     "\x91\x82\x00\xEC\x91\x83\x00\xEC\x91\x84\x00\xEC\x91\x85\x00\xEC\x91\x86"
-     "\x00\xEC"
-     "\x91\x87\x00\xEC\x91\x88\x00\xEC\x91\x89\x00\xEC\x91\x8A\x00\xEC\x91\x8B"
-     "\x00\xEC"
-     "\x91\x8C\x00\xEC\x91\x8D\x00\xEC\x91\x8E\x00\xEC\x91\x8F\x00\xEC\x91\x90"
-     "\x00\xEC"
-     "\x91\x91\x00\xEC\x91\x92\x00\xEC\x91\x93\x00\xEC\x91\x94\x00\xEC\x91\x95"
-     "\x00\xEC"
-     "\x91\x96\x00\xEC\x91\x97\x00\xEC\x91\x98\x00\xEC\x91\x99\x00\xEC\x91\x9A"
-     "\x00\xEC"
-     "\x91\x9B\x00\xEC\x91\x9C\x00\xEC\x91\x9D\x00\xEC\x91\x9E\x00\xEC\x91\x9F"
-     "\x00\xEC"
-     "\x91\xA0\x00\xEC\x91\xA1\x00\xEC\x91\xA2\x00\xEC\x91\xA3\x00\xEC\x91\xA4"
-     "\x00\xEC"
-     "\x91\xA5\x00\xEC\x91\xA6\x00\xEC\x91\xA7\x00\xEC\x91\xA8\x00\xEC\x91\xA9"
-     "\x00\xEC"
-     "\x91\xAA\x00\xEC\x91\xAB\x00\xEC\x91\xAC\x00\xEC\x91\xAD\x00\xEC\x91\xAE"
-     "\x00\xEC"
-     "\x91\xAF\x00\xEC\x91\xB0\x00\xEC\x91\xB1\x00\xEC\x91\xB2\x00\xEC\x91\xB3"
-     "\x00\xEC"
-     "\x91\xB4\x00\xEC\x91\xB5\x00\xEC\x91\xB6\x00\xEC\x91\xB7\x00\xEC\x91\xB8"
-     "\x00\xEC"
-     "\x91\xB9\x00\xEC\x91\xBA\x00\xEC\x91\xBB\x00\xEC\x91\xBC\x00\xEC\x91\xBD"
-     "\x00\xEC"
-     "\x91\xBE\x00\xEC\x91\xBF\x00\xEC\x92\x80\x00\xEC\x92\x81\x00\xEC\x92\x82"
-     "\x00\xEC"
-     "\x92\x83\x00\xEC\x92\x84\x00\xEC\x92\x85\x00\xEC\x92\x86\x00\xEC\x92\x87"
-     "\x00\xEC"
-     "\x92\x88\x00\xEC\x92\x89\x00\xEC\x92\x8A\x00\xEC\x92\x8B\x00\xEC\x92\x8C"
-     "\x00\xEC"
-     "\x92\x8D\x00\xEC\x92\x8E\x00\xEC\x92\x8F\x00\xEC\x92\x90\x00\xEC\x92\x91"
-     "\x00\xEC"
-     "\x92\x92\x00\xEC\x92\x93\x00\xEC\x92\x94\x00\xEC\x92\x95\x00\xEC\x92\x96"
-     "\x00\xEC"
-     "\x92\x97\x00\xEC\x92\x98\x00\xEC\x92\x99\x00\xEC\x92\x9A\x00\xEC\x92\x9B"
-     "\x00\xEC"
-     "\x92\x9C\x00\xEC\x92\x9D\x00\xEC\x92\x9E\x00\xEC\x92\x9F\x00\xEC\x92\xA0"
-     "\x00\xEC"
-     "\x92\xA1\x00\xEC\x92\xA2\x00\xEC\x92\xA3\x00\xEC\x92\xA4\x00\xEC\x92\xA5"
-     "\x00\xEC"
-     "\x92\xA6\x00\xEC\x92\xA7\x00\xEC\x92\xA8\x00\xEC\x92\xA9\x00\xEC\x92\xAA"
-     "\x00\xEC"
-     "\x92\xAB\x00\xEC\x92\xAC\x00\xEC\x92\xAD\x00\xEC\x92\xAE\x00\xEC\x92\xAF"
-     "\x00\xEC"
-     "\x92\xB0\x00\xEC\x92\xB1\x00\xEC\x92\xB2\x00\xEC\x92\xB3\x00\xEC\x92\xB4"
-     "\x00\xEC"
-     "\x92\xB5\x00\xEC\x92\xB6\x00\xEC\x92\xB7\x00\xEC\x92\xB8\x00\xEC\x92\xB9"
-     "\x00\xEC"
-     "\x92\xBA\x00\xEC\x92\xBB\x00\xEC\x92\xBC\x00\xEC\x92\xBD\x00\xEC\x92\xBE"
-     "\x00\xEC"
-     "\x92\xBF\x00\xEC\x93\x80\x00\xEC\x93\x81\x00\xEC\x93\x82\x00\xEC\x93\x83"
-     "\x00\xEC"
-     "\x93\x84\x00\xEC\x93\x85\x00\xEC\x93\x86\x00\xEC\x93\x87\x00\xEC\x93\x88"
-     "\x00\xEC"
-     "\x93\x89\x00\xEC\x93\x8A\x00\xEC\x93\x8B\x00\xEC\x93\x8C\x00\xEC\x93\x8D"
-     "\x00\xEC"
-     "\x93\x8E\x00\xEC\x93\x8F\x00\xEC\x93\x90\x00\xEC\x93\x91\x00\xEC\x93\x92"
-     "\x00\xEC"
-     "\x93\x93\x00\xEC\x93\x94\x00\xEC\x93\x95\x00\xEC\x93\x96\x00\xEC\x93\x97"
-     "\x00\xEC"
-     "\x93\x98\x00\xEC\x93\x99\x00\xEC\x93\x9A\x00\xEC\x93\x9B\x00\xEC\x93\x9C"
-     "\x00\xEC"
-     "\x93\x9D\x00\xEC\x93\x9E\x00\xEC\x93\x9F\x00\xEC\x93\xA0\x00\xEC\x93\xA1"
-     "\x00\xEC"
-     "\x93\xA2\x00\xEC\x93\xA3\x00\xEC\x93\xA4\x00\xEC\x93\xA5\x00\xEC\x93\xA6"
-     "\x00\xEC"
-     "\x93\xA7\x00\xEC\x93\xA8\x00\xEC\x93\xA9\x00\xEC\x93\xAA\x00\xEC\x93\xAB"
-     "\x00\xEC"
-     "\x93\xAC\x00\xEC\x93\xAD\x00\xEC\x93\xAE\x00\xEC\x93\xAF\x00\xEC\x93\xB0"
-     "\x00\xEC"
-     "\x93\xB1\x00\xEC\x93\xB2\x00\xEC\x93\xB3\x00\xEC\x93\xB4\x00\xEC\x93\xB5"
-     "\x00\xEC"
-     "\x93\xB6\x00\xEC\x93\xB7\x00\xEC\x93\xB8\x00\xEC\x93\xB9\x00\xEC\x93\xBA"
-     "\x00\xEC"
-     "\x93\xBB\x00\xEC\x93\xBC\x00\xEC\x93\xBD\x00\xEC\x93\xBE\x00\xEC\x93\xBF"
-     "\x00\xEC"
-     "\x94\x80\x00\xEC\x94\x81\x00\xEC\x94\x82\x00\xEC\x94\x83\x00\xEC\x94\x84"
-     "\x00\xEC"
-     "\x94\x85\x00\xEC\x94\x86\x00\xEC\x94\x87\x00\xEC\x94\x88\x00\xEC\x94\x89"
-     "\x00\xEC"
-     "\x94\x8A\x00\xEC\x94\x8B\x00\xEC\x94\x8C\x00\xEC\x94\x8D\x00\xEC\x94\x8E"
-     "\x00\xEC"
-     "\x94\x8F\x00\xEC\x94\x90\x00\xEC\x94\x91\x00\xEC\x94\x92\x00\xEC\x94\x93"
-     "\x00\xEC"
-     "\x94\x94\x00\xEC\x94\x95\x00\xEC\x94\x96\x00\xEC\x94\x97\x00\xEC\x94\x98"
-     "\x00\xEC"
-     "\x94\x99\x00\xEC\x94\x9A\x00\xEC\x94\x9B\x00\xEC\x94\x9C\x00\xEC\x94\x9D"
-     "\x00\xEC"
-     "\x94\x9E\x00\xEC\x94\x9F\x00\xEC\x94\xA0\x00\xEC\x94\xA1\x00\xEC\x94\xA2"
-     "\x00\xEC"
-     "\x94\xA3\x00\xEC\x94\xA4\x00\xEC\x94\xA5\x00\xEC\x94\xA6\x00\xEC\x94\xA7"
-     "\x00\xEC"
-     "\x94\xA8\x00\xEC\x94\xA9\x00\xEC\x94\xAA\x00\xEC\x94\xAB\x00\xEC\x94\xAC"
-     "\x00\xEC"
-     "\x94\xAD\x00\xEC\x94\xAE\x00\xEC\x94\xAF\x00\xEC\x94\xB0\x00\xEC\x94\xB1"
-     "\x00\xEC"
-     "\x94\xB2\x00\xEC\x94\xB3\x00\xEC\x94\xB4\x00\xEC\x94\xB5\x00\xEC\x94\xB6"
-     "\x00\xEC"
-     "\x94\xB7\x00\xEC\x94\xB8\x00\xEC\x94\xB9\x00\xEC\x94\xBA\x00\xEC\x94\xBB"
-     "\x00\xEC"
-     "\x94\xBC\x00\xEC\x94\xBD\x00\xEC\x94\xBE\x00\xEC\x94\xBF\x00\xEC\x95\x80"
-     "\x00\xEC"
-     "\x95\x81\x00\xEC\x95\x82\x00\xEC\x95\x83\x00\xEC\x95\x84\x00\xEC\x95\x85"
-     "\x00\xEC"
-     "\x95\x86\x00\xEC\x95\x87\x00\xEC\x95\x88\x00\xEC\x95\x89\x00\xEC\x95\x8A"
-     "\x00\xEC"
-     "\x95\x8B\x00\xEC\x95\x8C\x00\xEC\x95\x8D\x00\xEC\x95\x8E\x00\xEC\x95\x8F"
-     "\x00\xEC"
-     "\x95\x90\x00\xEC\x95\x91\x00\xEC\x95\x92\x00\xEC\x95\x93\x00\xEC\x95\x94"
-     "\x00\xEC"
-     "\x95\x95\x00\xEC\x95\x96\x00\xEC\x95\x97\x00\xEC\x95\x98\x00\xEC\x95\x99"
-     "\x00\xEC"
-     "\x95\x9A\x00\xEC\x95\x9B\x00\xEC\x95\x9C\x00\xEC\x95\x9D\x00\xEC\x95\x9E"
-     "\x00\xEC"
-     "\x95\x9F\x00\xEC\x95\xA0\x00\xEC\x95\xA1\x00\xEC\x95\xA2\x00\xEC\x95\xA3"
-     "\x00\xEC"
-     "\x95\xA4\x00\xEC\x95\xA5\x00\xEC\x95\xA6\x00\xEC\x95\xA7\x00\xEC\x95\xA8"
-     "\x00\xEC"
-     "\x95\xA9\x00\xEC\x95\xAA\x00\xEC\x95\xAB\x00\xEC\x95\xAC\x00\xEC\x95\xAD"
-     "\x00\xEC"
-     "\x95\xAE\x00\xEC\x95\xAF\x00\xEC\x95\xB0\x00\xEC\x95\xB1\x00\xEC\x95\xB2"
-     "\x00\xEC"
-     "\x95\xB3\x00\xEC\x95\xB4\x00\xEC\x95\xB5\x00\xEC\x95\xB6\x00\xEC\x95\xB7"
-     "\x00\xEC"
-     "\x95\xB8\x00\xEC\x95\xB9\x00\xEC\x95\xBA\x00\xEC\x95\xBB\x00\xEC\x95\xBC"
-     "\x00\xEC"
-     "\x95\xBD\x00\xEC\x95\xBE\x00\xEC\x95\xBF\x00\xEC\x96\x80\x00\xEC\x96\x81"
-     "\x00\xEC"
-     "\x96\x82\x00\xEC\x96\x83\x00\xEC\x96\x84\x00\xEC\x96\x85\x00\xEC\x96\x86"
-     "\x00\xEC"
-     "\x96\x87\x00\xEC\x96\x88\x00\xEC\x96\x89\x00\xEC\x96\x8A\x00\xEC\x96\x8B"
-     "\x00\xEC"
-     "\x96\x8C\x00\xEC\x96\x8D\x00\xEC\x96\x8E\x00\xEC\x96\x8F\x00\xEC\x96\x90"
-     "\x00\xEC"
-     "\x96\x91\x00\xEC\x96\x92\x00\xEC\x96\x93\x00\xEC\x96\x94\x00\xEC\x96\x95"
-     "\x00\xEC"
-     "\x96\x96\x00\xEC\x96\x97\x00\xEC\x96\x98\x00\xEC\x96\x99\x00\xEC\x96\x9A"
-     "\x00\xEC"
-     "\x96\x9B\x00\xEC\x96\x9C\x00\xEC\x96\x9D\x00\xEC\x96\x9E\x00\xEC\x96\x9F"
-     "\x00\xEC"
-     "\x96\xA0\x00\xEC\x96\xA1\x00\xEC\x96\xA2\x00\xEC\x96\xA3\x00\xEC\x96\xA4"
-     "\x00\xEC"
-     "\x96\xA5\x00\xEC\x96\xA6\x00\xEC\x96\xA7\x00\xEC\x96\xA8\x00\xEC\x96\xA9"
-     "\x00\xEC"
-     "\x96\xAA\x00\xEC\x96\xAB\x00\xEC\x96\xAC\x00\xEC\x96\xAD\x00\xEC\x96\xAE"
-     "\x00\xEC"
-     "\x96\xAF\x00\xEC\x96\xB0\x00\xEC\x96\xB1\x00\xEC\x96\xB2\x00\xEC\x96\xB3"
-     "\x00\xEC"
-     "\x96\xB4\x00\xEC\x96\xB5\x00\xEC\x96\xB6\x00\xEC\x96\xB7\x00\xEC\x96\xB8"
-     "\x00\xEC"
-     "\x96\xB9\x00\xEC\x96\xBA\x00\xEC\x96\xBB\x00\xEC\x96\xBC\x00\xEC\x96\xBD"
-     "\x00\xEC"
-     "\x96\xBE\x00\xEC\x96\xBF\x00\xEC\x97\x80\x00\xEC\x97\x81\x00\xEC\x97\x82"
-     "\x00\xEC"
-     "\x97\x83\x00\xEC\x97\x84\x00\xEC\x97\x85\x00\xEC\x97\x86\x00\xEC\x97\x87"
-     "\x00\xEC"
-     "\x97\x88\x00\xEC\x97\x89\x00\xEC\x97\x8A\x00\xEC\x97\x8B\x00\xEC\x97\x8C"
-     "\x00\xEC"
-     "\x97\x8D\x00\xEC\x97\x8E\x00\xEC\x97\x8F\x00\xEC\x97\x90\x00\xEC\x97\x91"
-     "\x00\xEC"
-     "\x97\x92\x00\xEC\x97\x93\x00\xEC\x97\x94\x00\xEC\x97\x95\x00\xEC\x97\x96"
-     "\x00\xEC"
-     "\x97\x97\x00\xEC\x97\x98\x00\xEC\x97\x99\x00\xEC\x97\x9A\x00\xEC\x97\x9B"
-     "\x00\xEC"
-     "\x97\x9C\x00\xEC\x97\x9D\x00\xEC\x97\x9E\x00\xEC\x97\x9F\x00\xEC\x97\xA0"
-     "\x00\xEC"
-     "\x97\xA1\x00\xEC\x97\xA2\x00\xEC\x97\xA3\x00\xEC\x97\xA4\x00\xEC\x97\xA5"
-     "\x00\xEC"
-     "\x97\xA6\x00\xEC\x97\xA7\x00\xEC\x97\xA8\x00\xEC\x97\xA9\x00\xEC\x97\xAA"
-     "\x00\xEC"
-     "\x97\xAB\x00\xEC\x97\xAC\x00\xEC\x97\xAD\x00\xEC\x97\xAE\x00\xEC\x97\xAF"
-     "\x00\xEC"
-     "\x97\xB0\x00\xEC\x97\xB1\x00\xEC\x97\xB2\x00\xEC\x97\xB3\x00\xEC\x97\xB4"
-     "\x00\xEC"
-     "\x97\xB5\x00\xEC\x97\xB6\x00\xEC\x97\xB7\x00\xEC\x97\xB8\x00\xEC\x97\xB9"
-     "\x00\xEC"
-     "\x97\xBA\x00\xEC\x97\xBB\x00\xEC\x97\xBC\x00\xEC\x97\xBD\x00\xEC\x97\xBE"
-     "\x00\xEC"
-     "\x97\xBF\x00\xEC\x98\x80\x00\xEC\x98\x81\x00\xEC\x98\x82\x00\xEC\x98\x83"
-     "\x00\xEC"
-     "\x98\x84\x00\xEC\x98\x85\x00\xEC\x98\x86\x00\xEC\x98\x87\x00\xEC\x98\x88"
-     "\x00\xEC"
-     "\x98\x89\x00\xEC\x98\x8A\x00\xEC\x98\x8B\x00\xEC\x98\x8C\x00\xEC\x98\x8D"
-     "\x00\xEC"
-     "\x98\x8E\x00\xEC\x98\x8F\x00\xEC\x98\x90\x00\xEC\x98\x91\x00\xEC\x98\x92"
-     "\x00\xEC"
-     "\x98\x93\x00\xEC\x98\x94\x00\xEC\x98\x95\x00\xEC\x98\x96\x00\xEC\x98\x97"
-     "\x00\xEC"
-     "\x98\x98\x00\xEC\x98\x99\x00\xEC\x98\x9A\x00\xEC\x98\x9B\x00\xEC\x98\x9C"
-     "\x00\xEC"
-     "\x98\x9D\x00\xEC\x98\x9E\x00\xEC\x98\x9F\x00\xEC\x98\xA0\x00\xEC\x98\xA1"
-     "\x00\xEC"
-     "\x98\xA2\x00\xEC\x98\xA3\x00\xEC\x98\xA4\x00\xEC\x98\xA5\x00\xEC\x98\xA6"
-     "\x00\xEC"
-     "\x98\xA7\x00\xEC\x98\xA8\x00\xEC\x98\xA9\x00\xEC\x98\xAA\x00\xEC\x98\xAB"
-     "\x00\xEC"
-     "\x98\xAC\x00\xEC\x98\xAD\x00\xEC\x98\xAE\x00\xEC\x98\xAF\x00\xEC\x98\xB0"
-     "\x00\xEC"
-     "\x98\xB1\x00\xEC\x98\xB2\x00\xEC\x98\xB3\x00\xEC\x98\xB4\x00\xEC\x98\xB5"
-     "\x00\xEC"
-     "\x98\xB6\x00\xEC\x98\xB7\x00\xEC\x98\xB8\x00\xEC\x98\xB9\x00\xEC\x98\xBA"
-     "\x00\xEC"
-     "\x98\xBB\x00\xEC\x98\xBC\x00\xEC\x98\xBD\x00\xEC\x98\xBE\x00\xEC\x98\xBF"
-     "\x00\xEC"
-     "\x99\x80\x00\xEC\x99\x81\x00\xEC\x99\x82\x00\xEC\x99\x83\x00\xEC\x99\x84"
-     "\x00\xEC"
-     "\x99\x85\x00\xEC\x99\x86\x00\xEC\x99\x87\x00\xEC\x99\x88\x00\xEC\x99\x89"
-     "\x00\xEC"
-     "\x99\x8A\x00\xEC\x99\x8B\x00\xEC\x99\x8C\x00\xEC\x99\x8D\x00\xEC\x99\x8E"
-     "\x00\xEC"
-     "\x99\x8F\x00\xEC\x99\x90\x00\xEC\x99\x91\x00\xEC\x99\x92\x00\xEC\x99\x93"
-     "\x00\xEC"
-     "\x99\x94\x00\xEC\x99\x95\x00\xEC\x99\x96\x00\xEC\x99\x97\x00\xEC\x99\x98"
-     "\x00\xEC"
-     "\x99\x99\x00\xEC\x99\x9A\x00\xEC\x99\x9B\x00\xEC\x99\x9C\x00\xEC\x99\x9D"
-     "\x00\xEC"
-     "\x99\x9E\x00\xEC\x99\x9F\x00\xEC\x99\xA0\x00\xEC\x99\xA1\x00\xEC\x99\xA2"
-     "\x00\xEC"
-     "\x99\xA3\x00\xEC\x99\xA4\x00\xEC\x99\xA5\x00\xEC\x99\xA6\x00\xEC\x99\xA7"
-     "\x00\xEC"
-     "\x99\xA8\x00\xEC\x99\xA9\x00\xEC\x99\xAA\x00\xEC\x99\xAB\x00\xEC\x99\xAC"
-     "\x00\xEC"
-     "\x99\xAD\x00\xEC\x99\xAE\x00\xEC\x99\xAF\x00\xEC\x99\xB0\x00\xEC\x99\xB1"
-     "\x00\xEC"
-     "\x99\xB2\x00\xEC\x99\xB3\x00\xEC\x99\xB4\x00\xEC\x99\xB5\x00\xEC\x99\xB6"
-     "\x00\xEC"
-     "\x99\xB7\x00\xEC\x99\xB8\x00\xEC\x99\xB9\x00\xEC\x99\xBA\x00\xEC\x99\xBB"
-     "\x00\xEC"
-     "\x99\xBC\x00\xEC\x99\xBD\x00\xEC\x99\xBE\x00\xEC\x99\xBF\x00\xEC\x9A\x80"
-     "\x00\xEC"
-     "\x9A\x81\x00\xEC\x9A\x82\x00\xEC\x9A\x83\x00\xEC\x9A\x84\x00\xEC\x9A\x85"
-     "\x00\xEC"
-     "\x9A\x86\x00\xEC\x9A\x87\x00\xEC\x9A\x88\x00\xEC\x9A\x89\x00\xEC\x9A\x8A"
-     "\x00\xEC"
-     "\x9A\x8B\x00\xEC\x9A\x8C\x00\xEC\x9A\x8D\x00\xEC\x9A\x8E\x00\xEC\x9A\x8F"
-     "\x00\xEC"
-     "\x9A\x90\x00\xEC\x9A\x91\x00\xEC\x9A\x92\x00\xEC\x9A\x93\x00\xEC\x9A\x94"
-     "\x00\xEC"
-     "\x9A\x95\x00\xEC\x9A\x96\x00\xEC\x9A\x97\x00\xEC\x9A\x98\x00\xEC\x9A\x99"
-     "\x00\xEC"
-     "\x9A\x9A\x00\xEC\x9A\x9B\x00\xEC\x9A\x9C\x00\xEC\x9A\x9D\x00\xEC\x9A\x9E"
-     "\x00\xEC"
-     "\x9A\x9F\x00\xEC\x9A\xA0\x00\xEC\x9A\xA1\x00\xEC\x9A\xA2\x00\xEC\x9A\xA3"
-     "\x00\xEC"
-     "\x9A\xA4\x00\xEC\x9A\xA5\x00\xEC\x9A\xA6\x00\xEC\x9A\xA7\x00\xEC\x9A\xA8"
-     "\x00\xEC"
-     "\x9A\xA9\x00\xEC\x9A\xAA\x00\xEC\x9A\xAB\x00\xEC\x9A\xAC\x00\xEC\x9A\xAD"
-     "\x00\xEC"
-     "\x9A\xAE\x00\xEC\x9A\xAF\x00\xEC\x9A\xB0\x00\xEC\x9A\xB1\x00\xEC\x9A\xB2"
-     "\x00\xEC"
-     "\x9A\xB3\x00\xEC\x9A\xB4\x00\xEC\x9A\xB5\x00\xEC\x9A\xB6\x00\xEC\x9A\xB7"
-     "\x00\xEC"
-     "\x9A\xB8\x00\xEC\x9A\xB9\x00\xEC\x9A\xBA\x00\xEC\x9A\xBB\x00\xEC\x9A\xBC"
-     "\x00\xEC"
-     "\x9A\xBD\x00\xEC\x9A\xBE\x00\xEC\x9A\xBF\x00\xEC\x9B\x80\x00\xEC\x9B\x81"
-     "\x00\xEC"
-     "\x9B\x82\x00\xEC\x9B\x83\x00\xEC\x9B\x84\x00\xEC\x9B\x85\x00\xEC\x9B\x86"
-     "\x00\xEC"
-     "\x9B\x87\x00\xEC\x9B\x88\x00\xEC\x9B\x89\x00\xEC\x9B\x8A\x00\xEC\x9B\x8B"
-     "\x00\xEC"
-     "\x9B\x8C\x00\xEC\x9B\x8D\x00\xEC\x9B\x8E\x00\xEC\x9B\x8F\x00\xEC\x9B\x90"
-     "\x00\xEC"
-     "\x9B\x91\x00\xEC\x9B\x92\x00\xEC\x9B\x93\x00\xEC\x9B\x94\x00\xEC\x9B\x95"
-     "\x00\xEC"
-     "\x9B\x96\x00\xEC\x9B\x97\x00\xEC\x9B\x98\x00\xEC\x9B\x99\x00\xEC\x9B\x9A"
-     "\x00\xEC"
-     "\x9B\x9B\x00\xEC\x9B\x9C\x00\xEC\x9B\x9D\x00\xEC\x9B\x9E\x00\xEC\x9B\x9F"
-     "\x00\xEC"
-     "\x9B\xA0\x00\xEC\x9B\xA1\x00\xEC\x9B\xA2\x00\xEC\x9B\xA3\x00\xEC\x9B\xA4"
-     "\x00\xEC"
-     "\x9B\xA5\x00\xEC\x9B\xA6\x00\xEC\x9B\xA7\x00\xEC\x9B\xA8\x00\xEC\x9B\xA9"
-     "\x00\xEC"
-     "\x9B\xAA\x00\xEC\x9B\xAB\x00\xEC\x9B\xAC\x00\xEC\x9B\xAD\x00\xEC\x9B\xAE"
-     "\x00\xEC"
-     "\x9B\xAF\x00\xEC\x9B\xB0\x00\xEC\x9B\xB1\x00\xEC\x9B\xB2\x00\xEC\x9B\xB3"
-     "\x00\xEC"
-     "\x9B\xB4\x00\xEC\x9B\xB5\x00\xEC\x9B\xB6\x00\xEC\x9B\xB7\x00\xEC\x9B\xB8"
-     "\x00\xEC"
-     "\x9B\xB9\x00\xEC\x9B\xBA\x00\xEC\x9B\xBB\x00\xEC\x9B\xBC\x00\xEC\x9B\xBD"
-     "\x00\xEC"
-     "\x9B\xBE\x00\xEC\x9B\xBF\x00\xEC\x9C\x80\x00\xEC\x9C\x81\x00\xEC\x9C\x82"
-     "\x00\xEC"
-     "\x9C\x83\x00\xEC\x9C\x84\x00\xEC\x9C\x85\x00\xEC\x9C\x86\x00\xEC\x9C\x87"
-     "\x00\xEC"
-     "\x9C\x88\x00\xEC\x9C\x89\x00\xEC\x9C\x8A\x00\xEC\x9C\x8B\x00\xEC\x9C\x8C"
-     "\x00\xEC"
-     "\x9C\x8D\x00\xEC\x9C\x8E\x00\xEC\x9C\x8F\x00\xEC\x9C\x90\x00\xEC\x9C\x91"
-     "\x00\xEC"
-     "\x9C\x92\x00\xEC\x9C\x93\x00\xEC\x9C\x94\x00\xEC\x9C\x95\x00\xEC\x9C\x96"
-     "\x00\xEC"
-     "\x9C\x97\x00\xEC\x9C\x98\x00\xEC\x9C\x99\x00\xEC\x9C\x9A\x00\xEC\x9C\x9B"
-     "\x00\xEC"
-     "\x9C\x9C\x00\xEC\x9C\x9D\x00\xEC\x9C\x9E\x00\xEC\x9C\x9F\x00\xEC\x9C\xA0"
-     "\x00\xEC"
-     "\x9C\xA1\x00\xEC\x9C\xA2\x00\xEC\x9C\xA3\x00\xEC\x9C\xA4\x00\xEC\x9C\xA5"
-     "\x00\xEC"
-     "\x9C\xA6\x00\xEC\x9C\xA7\x00\xEC\x9C\xA8\x00\xEC\x9C\xA9\x00\xEC\x9C\xAA"
-     "\x00\xEC"
-     "\x9C\xAB\x00\xEC\x9C\xAC\x00\xEC\x9C\xAD\x00\xEC\x9C\xAE\x00\xEC\x9C\xAF"
-     "\x00\xEC"
-     "\x9C\xB0\x00\xEC\x9C\xB1\x00\xEC\x9C\xB2\x00\xEC\x9C\xB3\x00\xEC\x9C\xB4"
-     "\x00\xEC"
-     "\x9C\xB5\x00\xEC\x9C\xB6\x00\xEC\x9C\xB7\x00\xEC\x9C\xB8\x00\xEC\x9C\xB9"
-     "\x00\xEC"
-     "\x9C\xBA\x00\xEC\x9C\xBB\x00\xEC\x9C\xBC\x00\xEC\x9C\xBD\x00\xEC\x9C\xBE"
-     "\x00\xEC"
-     "\x9C\xBF\x00\xEC\x9D\x80\x00\xEC\x9D\x81\x00\xEC\x9D\x82\x00\xEC\x9D\x83"
-     "\x00\xEC"
-     "\x9D\x84\x00\xEC\x9D\x85\x00\xEC\x9D\x86\x00\xEC\x9D\x87\x00\xEC\x9D\x88"
-     "\x00\xEC"
-     "\x9D\x89\x00\xEC\x9D\x8A\x00\xEC\x9D\x8B\x00\xEC\x9D\x8C\x00\xEC\x9D\x8D"
-     "\x00\xEC"
-     "\x9D\x8E\x00\xEC\x9D\x8F\x00\xEC\x9D\x90\x00\xEC\x9D\x91\x00\xEC\x9D\x92"
-     "\x00\xEC"
-     "\x9D\x93\x00\xEC\x9D\x94\x00\xEC\x9D\x95\x00\xEC\x9D\x96\x00\xEC\x9D\x97"
-     "\x00\xEC"
-     "\x9D\x98\x00\xEC\x9D\x99\x00\xEC\x9D\x9A\x00\xEC\x9D\x9B\x00\xEC\x9D\x9C"
-     "\x00\xEC"
-     "\x9D\x9D\x00\xEC\x9D\x9E\x00\xEC\x9D\x9F\x00\xEC\x9D\xA0\x00\xEC\x9D\xA1"
-     "\x00\xEC"
-     "\x9D\xA2\x00\xEC\x9D\xA3\x00\xEC\x9D\xA4\x00\xEC\x9D\xA5\x00\xEC\x9D\xA6"
-     "\x00\xEC"
-     "\x9D\xA7\x00\xEC\x9D\xA8\x00\xEC\x9D\xA9\x00\xEC\x9D\xAA\x00\xEC\x9D\xAB"
-     "\x00\xEC"
-     "\x9D\xAC\x00\xEC\x9D\xAD\x00\xEC\x9D\xAE\x00\xEC\x9D\xAF\x00\xEC\x9D\xB0"
-     "\x00\xEC"
-     "\x9D\xB1\x00\xEC\x9D\xB2\x00\xEC\x9D\xB3\x00\xEC\x9D\xB4\x00\xEC\x9D\xB5"
-     "\x00\xEC"
-     "\x9D\xB6\x00\xEC\x9D\xB7\x00\xEC\x9D\xB8\x00\xEC\x9D\xB9\x00\xEC\x9D\xBA"
-     "\x00\xEC"
-     "\x9D\xBB\x00\xEC\x9D\xBC\x00\xEC\x9D\xBD\x00\xEC\x9D\xBE\x00\xEC\x9D\xBF"
-     "\x00\xEC"
-     "\x9E\x80\x00\xEC\x9E\x81\x00\xEC\x9E\x82\x00\xEC\x9E\x83\x00\xEC\x9E\x84"
-     "\x00\xEC"
-     "\x9E\x85\x00\xEC\x9E\x86\x00\xEC\x9E\x87\x00\xEC\x9E\x88\x00\xEC\x9E\x89"
-     "\x00\xEC"
-     "\x9E\x8A\x00\xEC\x9E\x8B\x00\xEC\x9E\x8C\x00\xEC\x9E\x8D\x00\xEC\x9E\x8E"
-     "\x00\xEC"
-     "\x9E\x8F\x00\xEC\x9E\x90\x00\xEC\x9E\x91\x00\xEC\x9E\x92\x00\xEC\x9E\x93"
-     "\x00\xEC"
-     "\x9E\x94\x00\xEC\x9E\x95\x00\xEC\x9E\x96\x00\xEC\x9E\x97\x00\xEC\x9E\x98"
-     "\x00\xEC"
-     "\x9E\x99\x00\xEC\x9E\x9A\x00\xEC\x9E\x9B\x00\xEC\x9E\x9C\x00\xEC\x9E\x9D"
-     "\x00\xEC"
-     "\x9E\x9E\x00\xEC\x9E\x9F\x00\xEC\x9E\xA0\x00\xEC\x9E\xA1\x00\xEC\x9E\xA2"
-     "\x00\xEC"
-     "\x9E\xA3\x00\xEC\x9E\xA4\x00\xEC\x9E\xA5\x00\xEC\x9E\xA6\x00\xEC\x9E\xA7"
-     "\x00\xEC"
-     "\x9E\xA8\x00\xEC\x9E\xA9\x00\xEC\x9E\xAA\x00\xEC\x9E\xAB\x00\xEC\x9E\xAC"
-     "\x00\xEC"
-     "\x9E\xAD\x00\xEC\x9E\xAE\x00\xEC\x9E\xAF\x00\xEC\x9E\xB0\x00\xEC\x9E\xB1"
-     "\x00\xEC"
-     "\x9E\xB2\x00\xEC\x9E\xB3\x00\xEC\x9E\xB4\x00\xEC\x9E\xB5\x00\xEC\x9E\xB6"
-     "\x00\xEC"
-     "\x9E\xB7\x00\xEC\x9E\xB8\x00\xEC\x9E\xB9\x00\xEC\x9E\xBA\x00\xEC\x9E\xBB"
-     "\x00\xEC"
-     "\x9E\xBC\x00\xEC\x9E\xBD\x00\xEC\x9E\xBE\x00\xEC\x9E\xBF\x00\xEC\x9F\x80"
-     "\x00\xEC"
-     "\x9F\x81\x00\xEC\x9F\x82\x00\xEC\x9F\x83\x00\xEC\x9F\x84\x00\xEC\x9F\x85"
-     "\x00\xEC"
-     "\x9F\x86\x00\xEC\x9F\x87\x00\xEC\x9F\x88\x00\xEC\x9F\x89\x00\xEC\x9F\x8A"
-     "\x00\xEC"
-     "\x9F\x8B\x00\xEC\x9F\x8C\x00\xEC\x9F\x8D\x00\xEC\x9F\x8E\x00\xEC\x9F\x8F"
-     "\x00\xEC"
-     "\x9F\x90\x00\xEC\x9F\x91\x00\xEC\x9F\x92\x00\xEC\x9F\x93\x00\xEC\x9F\x94"
-     "\x00\xEC"
-     "\x9F\x95\x00\xEC\x9F\x96\x00\xEC\x9F\x97\x00\xEC\x9F\x98\x00\xEC\x9F\x99"
-     "\x00\xEC"
-     "\x9F\x9A\x00\xEC\x9F\x9B\x00\xEC\x9F\x9C\x00\xEC\x9F\x9D\x00\xEC\x9F\x9E"
-     "\x00\xEC"
-     "\x9F\x9F\x00\xEC\x9F\xA0\x00\xEC\x9F\xA1\x00\xEC\x9F\xA2\x00\xEC\x9F\xA3"
-     "\x00\xEC"
-     "\x9F\xA4\x00\xEC\x9F\xA5\x00\xEC\x9F\xA6\x00\xEC\x9F\xA7\x00\xEC\x9F\xA8"
-     "\x00\xEC"
-     "\x9F\xA9\x00\xEC\x9F\xAA\x00\xEC\x9F\xAB\x00\xEC\x9F\xAC\x00\xEC\x9F\xAD"
-     "\x00\xEC"
-     "\x9F\xAE\x00\xEC\x9F\xAF\x00\xEC\x9F\xB0\x00\xEC\x9F\xB1\x00\xEC\x9F\xB2"
-     "\x00\xEC"
-     "\x9F\xB3\x00\xEC\x9F\xB4\x00\xEC\x9F\xB5\x00\xEC\x9F\xB6\x00\xEC\x9F\xB7"
-     "\x00\xEC"
-     "\x9F\xB8\x00\xEC\x9F\xB9\x00\xEC\x9F\xBA\x00\xEC\x9F\xBB\x00\xEC\x9F\xBC"
-     "\x00\xEC"
-     "\x9F\xBD\x00\xEC\x9F\xBE\x00\xEC\x9F\xBF\x00\xEC\xA0\x80\x00\xEC\xA0\x81"
-     "\x00\xEC"
-     "\xA0\x82\x00\xEC\xA0\x83\x00\xEC\xA0\x84\x00\xEC\xA0\x85\x00\xEC\xA0\x86"
-     "\x00\xEC"
-     "\xA0\x87\x00\xEC\xA0\x88\x00\xEC\xA0\x89\x00\xEC\xA0\x8A\x00\xEC\xA0\x8B"
-     "\x00\xEC"
-     "\xA0\x8C\x00\xEC\xA0\x8D\x00\xEC\xA0\x8E\x00\xEC\xA0\x8F\x00\xEC\xA0\x90"
-     "\x00\xEC"
-     "\xA0\x91\x00\xEC\xA0\x92\x00\xEC\xA0\x93\x00\xEC\xA0\x94\x00\xEC\xA0\x95"
-     "\x00\xEC"
-     "\xA0\x96\x00\xEC\xA0\x97\x00\xEC\xA0\x98\x00\xEC\xA0\x99\x00\xEC\xA0\x9A"
-     "\x00\xEC"
-     "\xA0\x9B\x00\xEC\xA0\x9C\x00\xEC\xA0\x9D\x00\xEC\xA0\x9E\x00\xEC\xA0\x9F"
-     "\x00\xEC"
-     "\xA0\xA0\x00\xEC\xA0\xA1\x00\xEC\xA0\xA2\x00\xEC\xA0\xA3\x00\xEC\xA0\xA4"
-     "\x00\xEC"
-     "\xA0\xA5\x00\xEC\xA0\xA6\x00\xEC\xA0\xA7\x00\xEC\xA0\xA8\x00\xEC\xA0\xA9"
-     "\x00\xEC"
-     "\xA0\xAA\x00\xEC\xA0\xAB\x00\xEC\xA0\xAC\x00\xEC\xA0\xAD\x00\xEC\xA0\xAE"
-     "\x00\xEC"
-     "\xA0\xAF\x00\xEC\xA0\xB0\x00\xEC\xA0\xB1\x00\xEC\xA0\xB2\x00\xEC\xA0\xB3"
-     "\x00\xEC"
-     "\xA0\xB4\x00\xEC\xA0\xB5\x00\xEC\xA0\xB6\x00\xEC\xA0\xB7\x00\xEC\xA0\xB8"
-     "\x00\xEC"
-     "\xA0\xB9\x00\xEC\xA0\xBA\x00\xEC\xA0\xBB\x00\xEC\xA0\xBC\x00\xEC\xA0\xBD"
-     "\x00\xEC"
-     "\xA0\xBE\x00\xEC\xA0\xBF\x00\xEC\xA1\x80\x00\xEC\xA1\x81\x00\xEC\xA1\x82"
-     "\x00\xEC"
-     "\xA1\x83\x00\xEC\xA1\x84\x00\xEC\xA1\x85\x00\xEC\xA1\x86\x00\xEC\xA1\x87"
-     "\x00\xEC"
-     "\xA1\x88\x00\xEC\xA1\x89\x00\xEC\xA1\x8A\x00\xEC\xA1\x8B\x00\xEC\xA1\x8C"
-     "\x00\xEC"
-     "\xA1\x8D\x00\xEC\xA1\x8E\x00\xEC\xA1\x8F\x00\xEC\xA1\x90\x00\xEC\xA1\x91"
-     "\x00\xEC"
-     "\xA1\x92\x00\xEC\xA1\x93\x00\xEC\xA1\x94\x00\xEC\xA1\x95\x00\xEC\xA1\x96"
-     "\x00\xEC"
-     "\xA1\x97\x00\xEC\xA1\x98\x00\xEC\xA1\x99\x00\xEC\xA1\x9A\x00\xEC\xA1\x9B"
-     "\x00\xEC"
-     "\xA1\x9C\x00\xEC\xA1\x9D\x00\xEC\xA1\x9E\x00\xEC\xA1\x9F\x00\xEC\xA1\xA0"
-     "\x00\xEC"
-     "\xA1\xA1\x00\xEC\xA1\xA2\x00\xEC\xA1\xA3\x00\xEC\xA1\xA4\x00\xEC\xA1\xA5"
-     "\x00\xEC"
-     "\xA1\xA6\x00\xEC\xA1\xA7\x00\xEC\xA1\xA8\x00\xEC\xA1\xA9\x00\xEC\xA1\xAA"
-     "\x00\xEC"
-     "\xA1\xAB\x00\xEC\xA1\xAC\x00\xEC\xA1\xAD\x00\xEC\xA1\xAE\x00\xEC\xA1\xAF"
-     "\x00\xEC"
-     "\xA1\xB0\x00\xEC\xA1\xB1\x00\xEC\xA1\xB2\x00\xEC\xA1\xB3\x00\xEC\xA1\xB4"
-     "\x00\xEC"
-     "\xA1\xB5\x00\xEC\xA1\xB6\x00\xEC\xA1\xB7\x00\xEC\xA1\xB8\x00\xEC\xA1\xB9"
-     "\x00\xEC"
-     "\xA1\xBA\x00\xEC\xA1\xBB\x00\xEC\xA1\xBC\x00\xEC\xA1\xBD\x00\xEC\xA1\xBE"
-     "\x00\xEC"
-     "\xA1\xBF\x00\xEC\xA2\x80\x00\xEC\xA2\x81\x00\xEC\xA2\x82\x00\xEC\xA2\x83"
-     "\x00\xEC"
-     "\xA2\x84\x00\xEC\xA2\x85\x00\xEC\xA2\x86\x00\xEC\xA2\x87\x00\xEC\xA2\x88"
-     "\x00\xEC"
-     "\xA2\x89\x00\xEC\xA2\x8A\x00\xEC\xA2\x8B\x00\xEC\xA2\x8C\x00\xEC\xA2\x8D"
-     "\x00\xEC"
-     "\xA2\x8E\x00\xEC\xA2\x8F\x00\xEC\xA2\x90\x00\xEC\xA2\x91\x00\xEC\xA2\x92"
-     "\x00\xEC"
-     "\xA2\x93\x00\xEC\xA2\x94\x00\xEC\xA2\x95\x00\xEC\xA2\x96\x00\xEC\xA2\x97"
-     "\x00\xEC"
-     "\xA2\x98\x00\xEC\xA2\x99\x00\xEC\xA2\x9A\x00\xEC\xA2\x9B\x00\xEC\xA2\x9C"
-     "\x00\xEC"
-     "\xA2\x9D\x00\xEC\xA2\x9E\x00\xEC\xA2\x9F\x00\xEC\xA2\xA0\x00\xEC\xA2\xA1"
-     "\x00\xEC"
-     "\xA2\xA2\x00\xEC\xA2\xA3\x00\xEC\xA2\xA4\x00\xEC\xA2\xA5\x00\xEC\xA2\xA6"
-     "\x00\xEC"
-     "\xA2\xA7\x00\xEC\xA2\xA8\x00\xEC\xA2\xA9\x00\xEC\xA2\xAA\x00\xEC\xA2\xAB"
-     "\x00\xEC"
-     "\xA2\xAC\x00\xEC\xA2\xAD\x00\xEC\xA2\xAE\x00\xEC\xA2\xAF\x00\xEC\xA2\xB0"
-     "\x00\xEC"
-     "\xA2\xB1\x00\xEC\xA2\xB2\x00\xEC\xA2\xB3\x00\xEC\xA2\xB4\x00\xEC\xA2\xB5"
-     "\x00\xEC"
-     "\xA2\xB6\x00\xEC\xA2\xB7\x00\xEC\xA2\xB8\x00\xEC\xA2\xB9\x00\xEC\xA2\xBA"
-     "\x00\xEC"
-     "\xA2\xBB\x00\xEC\xA2\xBC\x00\xEC\xA2\xBD\x00\xEC\xA2\xBE\x00\xEC\xA2\xBF"
-     "\x00\xEC"
-     "\xA3\x80\x00\xEC\xA3\x81\x00\xEC\xA3\x82\x00\xEC\xA3\x83\x00\xEC\xA3\x84"
-     "\x00\xEC"
-     "\xA3\x85\x00\xEC\xA3\x86\x00\xEC\xA3\x87\x00\xEC\xA3\x88\x00\xEC\xA3\x89"
-     "\x00\xEC"
-     "\xA3\x8A\x00\xEC\xA3\x8B\x00\xEC\xA3\x8C\x00\xEC\xA3\x8D\x00\xEC\xA3\x8E"
-     "\x00\xEC"
-     "\xA3\x8F\x00\xEC\xA3\x90\x00\xEC\xA3\x91\x00\xEC\xA3\x92\x00\xEC\xA3\x93"
-     "\x00\xEC"
-     "\xA3\x94\x00\xEC\xA3\x95\x00\xEC\xA3\x96\x00\xEC\xA3\x97\x00\xEC\xA3\x98"
-     "\x00\xEC"
-     "\xA3\x99\x00\xEC\xA3\x9A\x00\xEC\xA3\x9B\x00\xEC\xA3\x9C\x00\xEC\xA3\x9D"
-     "\x00\xEC"
-     "\xA3\x9E\x00\xEC\xA3\x9F\x00\xEC\xA3\xA0\x00\xEC\xA3\xA1\x00\xEC\xA3\xA2"
-     "\x00\xEC"
-     "\xA3\xA3\x00\xEC\xA3\xA4\x00\xEC\xA3\xA5\x00\xEC\xA3\xA6\x00\xEC\xA3\xA7"
-     "\x00\xEC"
-     "\xA3\xA8\x00\xEC\xA3\xA9\x00\xEC\xA3\xAA\x00\xEC\xA3\xAB\x00\xEC\xA3\xAC"
-     "\x00\xEC"
-     "\xA3\xAD\x00\xEC\xA3\xAE\x00\xEC\xA3\xAF\x00\xEC\xA3\xB0\x00\xEC\xA3\xB1"
-     "\x00\xEC"
-     "\xA3\xB2\x00\xEC\xA3\xB3\x00\xEC\xA3\xB4\x00\xEC\xA3\xB5\x00\xEC\xA3\xB6"
-     "\x00\xEC"
-     "\xA3\xB7\x00\xEC\xA3\xB8\x00\xEC\xA3\xB9\x00\xEC\xA3\xBA\x00\xEC\xA3\xBB"
-     "\x00\xEC"
-     "\xA3\xBC\x00\xEC\xA3\xBC\xEC\x9D\x98\x00\xEC\xA3\xBD\x00\xEC\xA3\xBE\x00"
-     "\xEC\xA3"
-     "\xBF\x00\xEC\xA4\x80\x00\xEC\xA4\x81\x00\xEC\xA4\x82\x00\xEC\xA4\x83\x00"
-     "\xEC\xA4"
-     "\x84\x00\xEC\xA4\x85\x00\xEC\xA4\x86\x00\xEC\xA4\x87\x00\xEC\xA4\x88\x00"
-     "\xEC\xA4"
-     "\x89\x00\xEC\xA4\x8A\x00\xEC\xA4\x8B\x00\xEC\xA4\x8C\x00\xEC\xA4\x8D\x00"
-     "\xEC\xA4"
-     "\x8E\x00\xEC\xA4\x8F\x00\xEC\xA4\x90\x00\xEC\xA4\x91\x00\xEC\xA4\x92\x00"
-     "\xEC\xA4"
-     "\x93\x00\xEC\xA4\x94\x00\xEC\xA4\x95\x00\xEC\xA4\x96\x00\xEC\xA4\x97\x00"
-     "\xEC\xA4"
-     "\x98\x00\xEC\xA4\x99\x00\xEC\xA4\x9A\x00\xEC\xA4\x9B\x00\xEC\xA4\x9C\x00"
-     "\xEC\xA4"
-     "\x9D\x00\xEC\xA4\x9E\x00\xEC\xA4\x9F\x00\xEC\xA4\xA0\x00\xEC\xA4\xA1\x00"
-     "\xEC\xA4"
-     "\xA2\x00\xEC\xA4\xA3\x00\xEC\xA4\xA4\x00\xEC\xA4\xA5\x00\xEC\xA4\xA6\x00"
-     "\xEC\xA4"
-     "\xA7\x00\xEC\xA4\xA8\x00\xEC\xA4\xA9\x00\xEC\xA4\xAA\x00\xEC\xA4\xAB\x00"
-     "\xEC\xA4"
-     "\xAC\x00\xEC\xA4\xAD\x00\xEC\xA4\xAE\x00\xEC\xA4\xAF\x00\xEC\xA4\xB0\x00"
-     "\xEC\xA4"
-     "\xB1\x00\xEC\xA4\xB2\x00\xEC\xA4\xB3\x00\xEC\xA4\xB4\x00\xEC\xA4\xB5\x00"
-     "\xEC\xA4"
-     "\xB6\x00\xEC\xA4\xB7\x00\xEC\xA4\xB8\x00\xEC\xA4\xB9\x00\xEC\xA4\xBA\x00"
-     "\xEC\xA4"
-     "\xBB\x00\xEC\xA4\xBC\x00\xEC\xA4\xBD\x00\xEC\xA4\xBE\x00\xEC\xA4\xBF\x00"
-     "\xEC\xA5"
-     "\x80\x00\xEC\xA5\x81\x00\xEC\xA5\x82\x00\xEC\xA5\x83\x00\xEC\xA5\x84\x00"
-     "\xEC\xA5"
-     "\x85\x00\xEC\xA5\x86\x00\xEC\xA5\x87\x00\xEC\xA5\x88\x00\xEC\xA5\x89\x00"
-     "\xEC\xA5"
-     "\x8A\x00\xEC\xA5\x8B\x00\xEC\xA5\x8C\x00\xEC\xA5\x8D\x00\xEC\xA5\x8E\x00"
-     "\xEC\xA5"
-     "\x8F\x00\xEC\xA5\x90\x00\xEC\xA5\x91\x00\xEC\xA5\x92\x00\xEC\xA5\x93\x00"
-     "\xEC\xA5"
-     "\x94\x00\xEC\xA5\x95\x00\xEC\xA5\x96\x00\xEC\xA5\x97\x00\xEC\xA5\x98\x00"
-     "\xEC\xA5"
-     "\x99\x00\xEC\xA5\x9A\x00\xEC\xA5\x9B\x00\xEC\xA5\x9C\x00\xEC\xA5\x9D\x00"
-     "\xEC\xA5"
-     "\x9E\x00\xEC\xA5\x9F\x00\xEC\xA5\xA0\x00\xEC\xA5\xA1\x00\xEC\xA5\xA2\x00"
-     "\xEC\xA5"
-     "\xA3\x00\xEC\xA5\xA4\x00\xEC\xA5\xA5\x00\xEC\xA5\xA6\x00\xEC\xA5\xA7\x00"
-     "\xEC\xA5"
-     "\xA8\x00\xEC\xA5\xA9\x00\xEC\xA5\xAA\x00\xEC\xA5\xAB\x00\xEC\xA5\xAC\x00"
-     "\xEC\xA5"
-     "\xAD\x00\xEC\xA5\xAE\x00\xEC\xA5\xAF\x00\xEC\xA5\xB0\x00\xEC\xA5\xB1\x00"
-     "\xEC\xA5"
-     "\xB2\x00\xEC\xA5\xB3\x00\xEC\xA5\xB4\x00\xEC\xA5\xB5\x00\xEC\xA5\xB6\x00"
-     "\xEC\xA5"
-     "\xB7\x00\xEC\xA5\xB8\x00\xEC\xA5\xB9\x00\xEC\xA5\xBA\x00\xEC\xA5\xBB\x00"
-     "\xEC\xA5"
-     "\xBC\x00\xEC\xA5\xBD\x00\xEC\xA5\xBE\x00\xEC\xA5\xBF\x00\xEC\xA6\x80\x00"
-     "\xEC\xA6"
-     "\x81\x00\xEC\xA6\x82\x00\xEC\xA6\x83\x00\xEC\xA6\x84\x00\xEC\xA6\x85\x00"
-     "\xEC\xA6"
-     "\x86\x00\xEC\xA6\x87\x00\xEC\xA6\x88\x00\xEC\xA6\x89\x00\xEC\xA6\x8A\x00"
-     "\xEC\xA6"
-     "\x8B\x00\xEC\xA6\x8C\x00\xEC\xA6\x8D\x00\xEC\xA6\x8E\x00\xEC\xA6\x8F\x00"
-     "\xEC\xA6"
-     "\x90\x00\xEC\xA6\x91\x00\xEC\xA6\x92\x00\xEC\xA6\x93\x00\xEC\xA6\x94\x00"
-     "\xEC\xA6"
-     "\x95\x00\xEC\xA6\x96\x00\xEC\xA6\x97\x00\xEC\xA6\x98\x00\xEC\xA6\x99\x00"
-     "\xEC\xA6"
-     "\x9A\x00\xEC\xA6\x9B\x00\xEC\xA6\x9C\x00\xEC\xA6\x9D\x00\xEC\xA6\x9E\x00"
-     "\xEC\xA6"
-     "\x9F\x00\xEC\xA6\xA0\x00\xEC\xA6\xA1\x00\xEC\xA6\xA2\x00\xEC\xA6\xA3\x00"
-     "\xEC\xA6"
-     "\xA4\x00\xEC\xA6\xA5\x00\xEC\xA6\xA6\x00\xEC\xA6\xA7\x00\xEC\xA6\xA8\x00"
-     "\xEC\xA6"
-     "\xA9\x00\xEC\xA6\xAA\x00\xEC\xA6\xAB\x00\xEC\xA6\xAC\x00\xEC\xA6\xAD\x00"
-     "\xEC\xA6"
-     "\xAE\x00\xEC\xA6\xAF\x00\xEC\xA6\xB0\x00\xEC\xA6\xB1\x00\xEC\xA6\xB2\x00"
-     "\xEC\xA6"
-     "\xB3\x00\xEC\xA6\xB4\x00\xEC\xA6\xB5\x00\xEC\xA6\xB6\x00\xEC\xA6\xB7\x00"
-     "\xEC\xA6"
-     "\xB8\x00\xEC\xA6\xB9\x00\xEC\xA6\xBA\x00\xEC\xA6\xBB\x00\xEC\xA6\xBC\x00"
-     "\xEC\xA6"
-     "\xBD\x00\xEC\xA6\xBE\x00\xEC\xA6\xBF\x00\xEC\xA7\x80\x00\xEC\xA7\x81\x00"
-     "\xEC\xA7"
-     "\x82\x00\xEC\xA7\x83\x00\xEC\xA7\x84\x00\xEC\xA7\x85\x00\xEC\xA7\x86\x00"
-     "\xEC\xA7"
-     "\x87\x00\xEC\xA7\x88\x00\xEC\xA7\x89\x00\xEC\xA7\x8A\x00\xEC\xA7\x8B\x00"
-     "\xEC\xA7"
-     "\x8C\x00\xEC\xA7\x8D\x00\xEC\xA7\x8E\x00\xEC\xA7\x8F\x00\xEC\xA7\x90\x00"
-     "\xEC\xA7"
-     "\x91\x00\xEC\xA7\x92\x00\xEC\xA7\x93\x00\xEC\xA7\x94\x00\xEC\xA7\x95\x00"
-     "\xEC\xA7"
-     "\x96\x00\xEC\xA7\x97\x00\xEC\xA7\x98\x00\xEC\xA7\x99\x00\xEC\xA7\x9A\x00"
-     "\xEC\xA7"
-     "\x9B\x00\xEC\xA7\x9C\x00\xEC\xA7\x9D\x00\xEC\xA7\x9E\x00\xEC\xA7\x9F\x00"
-     "\xEC\xA7"
-     "\xA0\x00\xEC\xA7\xA1\x00\xEC\xA7\xA2\x00\xEC\xA7\xA3\x00\xEC\xA7\xA4\x00"
-     "\xEC\xA7"
-     "\xA5\x00\xEC\xA7\xA6\x00\xEC\xA7\xA7\x00\xEC\xA7\xA8\x00\xEC\xA7\xA9\x00"
-     "\xEC\xA7"
-     "\xAA\x00\xEC\xA7\xAB\x00\xEC\xA7\xAC\x00\xEC\xA7\xAD\x00\xEC\xA7\xAE\x00"
-     "\xEC\xA7"
-     "\xAF\x00\xEC\xA7\xB0\x00\xEC\xA7\xB1\x00\xEC\xA7\xB2\x00\xEC\xA7\xB3\x00"
-     "\xEC\xA7"
-     "\xB4\x00\xEC\xA7\xB5\x00\xEC\xA7\xB6\x00\xEC\xA7\xB7\x00\xEC\xA7\xB8\x00"
-     "\xEC\xA7"
-     "\xB9\x00\xEC\xA7\xBA\x00\xEC\xA7\xBB\x00\xEC\xA7\xBC\x00\xEC\xA7\xBD\x00"
-     "\xEC\xA7"
-     "\xBE\x00\xEC\xA7\xBF\x00\xEC\xA8\x80\x00\xEC\xA8\x81\x00\xEC\xA8\x82\x00"
-     "\xEC\xA8"
-     "\x83\x00\xEC\xA8\x84\x00\xEC\xA8\x85\x00\xEC\xA8\x86\x00\xEC\xA8\x87\x00"
-     "\xEC\xA8"
-     "\x88\x00\xEC\xA8\x89\x00\xEC\xA8\x8A\x00\xEC\xA8\x8B\x00\xEC\xA8\x8C\x00"
-     "\xEC\xA8"
-     "\x8D\x00\xEC\xA8\x8E\x00\xEC\xA8\x8F\x00\xEC\xA8\x90\x00\xEC\xA8\x91\x00"
-     "\xEC\xA8"
-     "\x92\x00\xEC\xA8\x93\x00\xEC\xA8\x94\x00\xEC\xA8\x95\x00\xEC\xA8\x96\x00"
-     "\xEC\xA8"
-     "\x97\x00\xEC\xA8\x98\x00\xEC\xA8\x99\x00\xEC\xA8\x9A\x00\xEC\xA8\x9B\x00"
-     "\xEC\xA8"
-     "\x9C\x00\xEC\xA8\x9D\x00\xEC\xA8\x9E\x00\xEC\xA8\x9F\x00\xEC\xA8\xA0\x00"
-     "\xEC\xA8"
-     "\xA1\x00\xEC\xA8\xA2\x00\xEC\xA8\xA3\x00\xEC\xA8\xA4\x00\xEC\xA8\xA5\x00"
-     "\xEC\xA8"
-     "\xA6\x00\xEC\xA8\xA7\x00\xEC\xA8\xA8\x00\xEC\xA8\xA9\x00\xEC\xA8\xAA\x00"
-     "\xEC\xA8"
-     "\xAB\x00\xEC\xA8\xAC\x00\xEC\xA8\xAD\x00\xEC\xA8\xAE\x00\xEC\xA8\xAF\x00"
-     "\xEC\xA8"
-     "\xB0\x00\xEC\xA8\xB1\x00\xEC\xA8\xB2\x00\xEC\xA8\xB3\x00\xEC\xA8\xB4\x00"
-     "\xEC\xA8"
-     "\xB5\x00\xEC\xA8\xB6\x00\xEC\xA8\xB7\x00\xEC\xA8\xB8\x00\xEC\xA8\xB9\x00"
-     "\xEC\xA8"
-     "\xBA\x00\xEC\xA8\xBB\x00\xEC\xA8\xBC\x00\xEC\xA8\xBD\x00\xEC\xA8\xBE\x00"
-     "\xEC\xA8"
-     "\xBF\x00\xEC\xA9\x80\x00\xEC\xA9\x81\x00\xEC\xA9\x82\x00\xEC\xA9\x83\x00"
-     "\xEC\xA9"
-     "\x84\x00\xEC\xA9\x85\x00\xEC\xA9\x86\x00\xEC\xA9\x87\x00\xEC\xA9\x88\x00"
-     "\xEC\xA9"
-     "\x89\x00\xEC\xA9\x8A\x00\xEC\xA9\x8B\x00\xEC\xA9\x8C\x00\xEC\xA9\x8D\x00"
-     "\xEC\xA9"
-     "\x8E\x00\xEC\xA9\x8F\x00\xEC\xA9\x90\x00\xEC\xA9\x91\x00\xEC\xA9\x92\x00"
-     "\xEC\xA9"
-     "\x93\x00\xEC\xA9\x94\x00\xEC\xA9\x95\x00\xEC\xA9\x96\x00\xEC\xA9\x97\x00"
-     "\xEC\xA9"
-     "\x98\x00\xEC\xA9\x99\x00\xEC\xA9\x9A\x00\xEC\xA9\x9B\x00\xEC\xA9\x9C\x00"
-     "\xEC\xA9"
-     "\x9D\x00\xEC\xA9\x9E\x00\xEC\xA9\x9F\x00\xEC\xA9\xA0\x00\xEC\xA9\xA1\x00"
-     "\xEC\xA9"
-     "\xA2\x00\xEC\xA9\xA3\x00\xEC\xA9\xA4\x00\xEC\xA9\xA5\x00\xEC\xA9\xA6\x00"
-     "\xEC\xA9"
-     "\xA7\x00\xEC\xA9\xA8\x00\xEC\xA9\xA9\x00\xEC\xA9\xAA\x00\xEC\xA9\xAB\x00"
-     "\xEC\xA9"
-     "\xAC\x00\xEC\xA9\xAD\x00\xEC\xA9\xAE\x00\xEC\xA9\xAF\x00\xEC\xA9\xB0\x00"
-     "\xEC\xA9"
-     "\xB1\x00\xEC\xA9\xB2\x00\xEC\xA9\xB3\x00\xEC\xA9\xB4\x00\xEC\xA9\xB5\x00"
-     "\xEC\xA9"
-     "\xB6\x00\xEC\xA9\xB7\x00\xEC\xA9\xB8\x00\xEC\xA9\xB9\x00\xEC\xA9\xBA\x00"
-     "\xEC\xA9"
-     "\xBB\x00\xEC\xA9\xBC\x00\xEC\xA9\xBD\x00\xEC\xA9\xBE\x00\xEC\xA9\xBF\x00"
-     "\xEC\xAA"
-     "\x80\x00\xEC\xAA\x81\x00\xEC\xAA\x82\x00\xEC\xAA\x83\x00\xEC\xAA\x84\x00"
-     "\xEC\xAA"
-     "\x85\x00\xEC\xAA\x86\x00\xEC\xAA\x87\x00\xEC\xAA\x88\x00\xEC\xAA\x89\x00"
-     "\xEC\xAA"
-     "\x8A\x00\xEC\xAA\x8B\x00\xEC\xAA\x8C\x00\xEC\xAA\x8D\x00\xEC\xAA\x8E\x00"
-     "\xEC\xAA"
-     "\x8F\x00\xEC\xAA\x90\x00\xEC\xAA\x91\x00\xEC\xAA\x92\x00\xEC\xAA\x93\x00"
-     "\xEC\xAA"
-     "\x94\x00\xEC\xAA\x95\x00\xEC\xAA\x96\x00\xEC\xAA\x97\x00\xEC\xAA\x98\x00"
-     "\xEC\xAA"
-     "\x99\x00\xEC\xAA\x9A\x00\xEC\xAA\x9B\x00\xEC\xAA\x9C\x00\xEC\xAA\x9D\x00"
-     "\xEC\xAA"
-     "\x9E\x00\xEC\xAA\x9F\x00\xEC\xAA\xA0\x00\xEC\xAA\xA1\x00\xEC\xAA\xA2\x00"
-     "\xEC\xAA"
-     "\xA3\x00\xEC\xAA\xA4\x00\xEC\xAA\xA5\x00\xEC\xAA\xA6\x00\xEC\xAA\xA7\x00"
-     "\xEC\xAA"
-     "\xA8\x00\xEC\xAA\xA9\x00\xEC\xAA\xAA\x00\xEC\xAA\xAB\x00\xEC\xAA\xAC\x00"
-     "\xEC\xAA"
-     "\xAD\x00\xEC\xAA\xAE\x00\xEC\xAA\xAF\x00\xEC\xAA\xB0\x00\xEC\xAA\xB1\x00"
-     "\xEC\xAA"
-     "\xB2\x00\xEC\xAA\xB3\x00\xEC\xAA\xB4\x00\xEC\xAA\xB5\x00\xEC\xAA\xB6\x00"
-     "\xEC\xAA"
-     "\xB7\x00\xEC\xAA\xB8\x00\xEC\xAA\xB9\x00\xEC\xAA\xBA\x00\xEC\xAA\xBB\x00"
-     "\xEC\xAA"
-     "\xBC\x00\xEC\xAA\xBD\x00\xEC\xAA\xBE\x00\xEC\xAA\xBF\x00\xEC\xAB\x80\x00"
-     "\xEC\xAB"
-     "\x81\x00\xEC\xAB\x82\x00\xEC\xAB\x83\x00\xEC\xAB\x84\x00\xEC\xAB\x85\x00"
-     "\xEC\xAB"
-     "\x86\x00\xEC\xAB\x87\x00\xEC\xAB\x88\x00\xEC\xAB\x89\x00\xEC\xAB\x8A\x00"
-     "\xEC\xAB"
-     "\x8B\x00\xEC\xAB\x8C\x00\xEC\xAB\x8D\x00\xEC\xAB\x8E\x00\xEC\xAB\x8F\x00"
-     "\xEC\xAB"
-     "\x90\x00\xEC\xAB\x91\x00\xEC\xAB\x92\x00\xEC\xAB\x93\x00\xEC\xAB\x94\x00"
-     "\xEC\xAB"
-     "\x95\x00\xEC\xAB\x96\x00\xEC\xAB\x97\x00\xEC\xAB\x98\x00\xEC\xAB\x99\x00"
-     "\xEC\xAB"
-     "\x9A\x00\xEC\xAB\x9B\x00\xEC\xAB\x9C\x00\xEC\xAB\x9D\x00\xEC\xAB\x9E\x00"
-     "\xEC\xAB"
-     "\x9F\x00\xEC\xAB\xA0\x00\xEC\xAB\xA1\x00\xEC\xAB\xA2\x00\xEC\xAB\xA3\x00"
-     "\xEC\xAB"
-     "\xA4\x00\xEC\xAB\xA5\x00\xEC\xAB\xA6\x00\xEC\xAB\xA7\x00\xEC\xAB\xA8\x00"
-     "\xEC\xAB"
-     "\xA9\x00\xEC\xAB\xAA\x00\xEC\xAB\xAB\x00\xEC\xAB\xAC\x00\xEC\xAB\xAD\x00"
-     "\xEC\xAB"
-     "\xAE\x00\xEC\xAB\xAF\x00\xEC\xAB\xB0\x00\xEC\xAB\xB1\x00\xEC\xAB\xB2\x00"
-     "\xEC\xAB"
-     "\xB3\x00\xEC\xAB\xB4\x00\xEC\xAB\xB5\x00\xEC\xAB\xB6\x00\xEC\xAB\xB7\x00"
-     "\xEC\xAB"
-     "\xB8\x00\xEC\xAB\xB9\x00\xEC\xAB\xBA\x00\xEC\xAB\xBB\x00\xEC\xAB\xBC\x00"
-     "\xEC\xAB"
-     "\xBD\x00\xEC\xAB\xBE\x00\xEC\xAB\xBF\x00\xEC\xAC\x80\x00\xEC\xAC\x81\x00"
-     "\xEC\xAC"
-     "\x82\x00\xEC\xAC\x83\x00\xEC\xAC\x84\x00\xEC\xAC\x85\x00\xEC\xAC\x86\x00"
-     "\xEC\xAC"
-     "\x87\x00\xEC\xAC\x88\x00\xEC\xAC\x89\x00\xEC\xAC\x8A\x00\xEC\xAC\x8B\x00"
-     "\xEC\xAC"
-     "\x8C\x00\xEC\xAC\x8D\x00\xEC\xAC\x8E\x00\xEC\xAC\x8F\x00\xEC\xAC\x90\x00"
-     "\xEC\xAC"
-     "\x91\x00\xEC\xAC\x92\x00\xEC\xAC\x93\x00\xEC\xAC\x94\x00\xEC\xAC\x95\x00"
-     "\xEC\xAC"
-     "\x96\x00\xEC\xAC\x97\x00\xEC\xAC\x98\x00\xEC\xAC\x99\x00\xEC\xAC\x9A\x00"
-     "\xEC\xAC"
-     "\x9B\x00\xEC\xAC\x9C\x00\xEC\xAC\x9D\x00\xEC\xAC\x9E\x00\xEC\xAC\x9F\x00"
-     "\xEC\xAC"
-     "\xA0\x00\xEC\xAC\xA1\x00\xEC\xAC\xA2\x00\xEC\xAC\xA3\x00\xEC\xAC\xA4\x00"
-     "\xEC\xAC"
-     "\xA5\x00\xEC\xAC\xA6\x00\xEC\xAC\xA7\x00\xEC\xAC\xA8\x00\xEC\xAC\xA9\x00"
-     "\xEC\xAC"
-     "\xAA\x00\xEC\xAC\xAB\x00\xEC\xAC\xAC\x00\xEC\xAC\xAD\x00\xEC\xAC\xAE\x00"
-     "\xEC\xAC"
-     "\xAF\x00\xEC\xAC\xB0\x00\xEC\xAC\xB1\x00\xEC\xAC\xB2\x00\xEC\xAC\xB3\x00"
-     "\xEC\xAC"
-     "\xB4\x00\xEC\xAC\xB5\x00\xEC\xAC\xB6\x00\xEC\xAC\xB7\x00\xEC\xAC\xB8\x00"
-     "\xEC\xAC"
-     "\xB9\x00\xEC\xAC\xBA\x00\xEC\xAC\xBB\x00\xEC\xAC\xBC\x00\xEC\xAC\xBD\x00"
-     "\xEC\xAC"
-     "\xBE\x00\xEC\xAC\xBF\x00\xEC\xAD\x80\x00\xEC\xAD\x81\x00\xEC\xAD\x82\x00"
-     "\xEC\xAD"
-     "\x83\x00\xEC\xAD\x84\x00\xEC\xAD\x85\x00\xEC\xAD\x86\x00\xEC\xAD\x87\x00"
-     "\xEC\xAD"
-     "\x88\x00\xEC\xAD\x89\x00\xEC\xAD\x8A\x00\xEC\xAD\x8B\x00\xEC\xAD\x8C\x00"
-     "\xEC\xAD"
-     "\x8D\x00\xEC\xAD\x8E\x00\xEC\xAD\x8F\x00\xEC\xAD\x90\x00\xEC\xAD\x91\x00"
-     "\xEC\xAD"
-     "\x92\x00\xEC\xAD\x93\x00\xEC\xAD\x94\x00\xEC\xAD\x95\x00\xEC\xAD\x96\x00"
-     "\xEC\xAD"
-     "\x97\x00\xEC\xAD\x98\x00\xEC\xAD\x99\x00\xEC\xAD\x9A\x00\xEC\xAD\x9B\x00"
-     "\xEC\xAD"
-     "\x9C\x00\xEC\xAD\x9D\x00\xEC\xAD\x9E\x00\xEC\xAD\x9F\x00\xEC\xAD\xA0\x00"
-     "\xEC\xAD"
-     "\xA1\x00\xEC\xAD\xA2\x00\xEC\xAD\xA3\x00\xEC\xAD\xA4\x00\xEC\xAD\xA5\x00"
-     "\xEC\xAD"
-     "\xA6\x00\xEC\xAD\xA7\x00\xEC\xAD\xA8\x00\xEC\xAD\xA9\x00\xEC\xAD\xAA\x00"
-     "\xEC\xAD"
-     "\xAB\x00\xEC\xAD\xAC\x00\xEC\xAD\xAD\x00\xEC\xAD\xAE\x00\xEC\xAD\xAF\x00"
-     "\xEC\xAD"
-     "\xB0\x00\xEC\xAD\xB1\x00\xEC\xAD\xB2\x00\xEC\xAD\xB3\x00\xEC\xAD\xB4\x00"
-     "\xEC\xAD"
-     "\xB5\x00\xEC\xAD\xB6\x00\xEC\xAD\xB7\x00\xEC\xAD\xB8\x00\xEC\xAD\xB9\x00"
-     "\xEC\xAD"
-     "\xBA\x00\xEC\xAD\xBB\x00\xEC\xAD\xBC\x00\xEC\xAD\xBD\x00\xEC\xAD\xBE\x00"
-     "\xEC\xAD"
-     "\xBF\x00\xEC\xAE\x80\x00\xEC\xAE\x81\x00\xEC\xAE\x82\x00\xEC\xAE\x83\x00"
-     "\xEC\xAE"
-     "\x84\x00\xEC\xAE\x85\x00\xEC\xAE\x86\x00\xEC\xAE\x87\x00\xEC\xAE\x88\x00"
-     "\xEC\xAE"
-     "\x89\x00\xEC\xAE\x8A\x00\xEC\xAE\x8B\x00\xEC\xAE\x8C\x00\xEC\xAE\x8D\x00"
-     "\xEC\xAE"
-     "\x8E\x00\xEC\xAE\x8F\x00\xEC\xAE\x90\x00\xEC\xAE\x91\x00\xEC\xAE\x92\x00"
-     "\xEC\xAE"
-     "\x93\x00\xEC\xAE\x94\x00\xEC\xAE\x95\x00\xEC\xAE\x96\x00\xEC\xAE\x97\x00"
-     "\xEC\xAE"
-     "\x98\x00\xEC\xAE\x99\x00\xEC\xAE\x9A\x00\xEC\xAE\x9B\x00\xEC\xAE\x9C\x00"
-     "\xEC\xAE"
-     "\x9D\x00\xEC\xAE\x9E\x00\xEC\xAE\x9F\x00\xEC\xAE\xA0\x00\xEC\xAE\xA1\x00"
-     "\xEC\xAE"
-     "\xA2\x00\xEC\xAE\xA3\x00\xEC\xAE\xA4\x00\xEC\xAE\xA5\x00\xEC\xAE\xA6\x00"
-     "\xEC\xAE"
-     "\xA7\x00\xEC\xAE\xA8\x00\xEC\xAE\xA9\x00\xEC\xAE\xAA\x00\xEC\xAE\xAB\x00"
-     "\xEC\xAE"
-     "\xAC\x00\xEC\xAE\xAD\x00\xEC\xAE\xAE\x00\xEC\xAE\xAF\x00\xEC\xAE\xB0\x00"
-     "\xEC\xAE"
-     "\xB1\x00\xEC\xAE\xB2\x00\xEC\xAE\xB3\x00\xEC\xAE\xB4\x00\xEC\xAE\xB5\x00"
-     "\xEC\xAE"
-     "\xB6\x00\xEC\xAE\xB7\x00\xEC\xAE\xB8\x00\xEC\xAE\xB9\x00\xEC\xAE\xBA\x00"
-     "\xEC\xAE"
-     "\xBB\x00\xEC\xAE\xBC\x00\xEC\xAE\xBD\x00\xEC\xAE\xBE\x00\xEC\xAE\xBF\x00"
-     "\xEC\xAF"
-     "\x80\x00\xEC\xAF\x81\x00\xEC\xAF\x82\x00\xEC\xAF\x83\x00\xEC\xAF\x84\x00"
-     "\xEC\xAF"
-     "\x85\x00\xEC\xAF\x86\x00\xEC\xAF\x87\x00\xEC\xAF\x88\x00\xEC\xAF\x89\x00"
-     "\xEC\xAF"
-     "\x8A\x00\xEC\xAF\x8B\x00\xEC\xAF\x8C\x00\xEC\xAF\x8D\x00\xEC\xAF\x8E\x00"
-     "\xEC\xAF"
-     "\x8F\x00\xEC\xAF\x90\x00\xEC\xAF\x91\x00\xEC\xAF\x92\x00\xEC\xAF\x93\x00"
-     "\xEC\xAF"
-     "\x94\x00\xEC\xAF\x95\x00\xEC\xAF\x96\x00\xEC\xAF\x97\x00\xEC\xAF\x98\x00"
-     "\xEC\xAF"
-     "\x99\x00\xEC\xAF\x9A\x00\xEC\xAF\x9B\x00\xEC\xAF\x9C\x00\xEC\xAF\x9D\x00"
-     "\xEC\xAF"
-     "\x9E\x00\xEC\xAF\x9F\x00\xEC\xAF\xA0\x00\xEC\xAF\xA1\x00\xEC\xAF\xA2\x00"
-     "\xEC\xAF"
-     "\xA3\x00\xEC\xAF\xA4\x00\xEC\xAF\xA5\x00\xEC\xAF\xA6\x00\xEC\xAF\xA7\x00"
-     "\xEC\xAF"
-     "\xA8\x00\xEC\xAF\xA9\x00\xEC\xAF\xAA\x00\xEC\xAF\xAB\x00\xEC\xAF\xAC\x00"
-     "\xEC\xAF"
-     "\xAD\x00\xEC\xAF\xAE\x00\xEC\xAF\xAF\x00\xEC\xAF\xB0\x00\xEC\xAF\xB1\x00"
-     "\xEC\xAF"
-     "\xB2\x00\xEC\xAF\xB3\x00\xEC\xAF\xB4\x00\xEC\xAF\xB5\x00\xEC\xAF\xB6\x00"
-     "\xEC\xAF"
-     "\xB7\x00\xEC\xAF\xB8\x00\xEC\xAF\xB9\x00\xEC\xAF\xBA\x00\xEC\xAF\xBB\x00"
-     "\xEC\xAF"
-     "\xBC\x00\xEC\xAF\xBD\x00\xEC\xAF\xBE\x00\xEC\xAF\xBF\x00\xEC\xB0\x80\x00"
-     "\xEC\xB0"
-     "\x81\x00\xEC\xB0\x82\x00\xEC\xB0\x83\x00\xEC\xB0\x84\x00\xEC\xB0\x85\x00"
-     "\xEC\xB0"
-     "\x86\x00\xEC\xB0\x87\x00\xEC\xB0\x88\x00\xEC\xB0\x89\x00\xEC\xB0\x8A\x00"
-     "\xEC\xB0"
-     "\x8B\x00\xEC\xB0\x8C\x00\xEC\xB0\x8D\x00\xEC\xB0\x8E\x00\xEC\xB0\x8F\x00"
-     "\xEC\xB0"
-     "\x90\x00\xEC\xB0\x91\x00\xEC\xB0\x92\x00\xEC\xB0\x93\x00\xEC\xB0\x94\x00"
-     "\xEC\xB0"
-     "\x95\x00\xEC\xB0\x96\x00\xEC\xB0\x97\x00\xEC\xB0\x98\x00\xEC\xB0\x99\x00"
-     "\xEC\xB0"
-     "\x9A\x00\xEC\xB0\x9B\x00\xEC\xB0\x9C\x00\xEC\xB0\x9D\x00\xEC\xB0\x9E\x00"
-     "\xEC\xB0"
-     "\x9F\x00\xEC\xB0\xA0\x00\xEC\xB0\xA1\x00\xEC\xB0\xA2\x00\xEC\xB0\xA3\x00"
-     "\xEC\xB0"
-     "\xA4\x00\xEC\xB0\xA5\x00\xEC\xB0\xA6\x00\xEC\xB0\xA7\x00\xEC\xB0\xA8\x00"
-     "\xEC\xB0"
-     "\xA9\x00\xEC\xB0\xAA\x00\xEC\xB0\xAB\x00\xEC\xB0\xAC\x00\xEC\xB0\xAD\x00"
-     "\xEC\xB0"
-     "\xAE\x00\xEC\xB0\xAF\x00\xEC\xB0\xB0\x00\xEC\xB0\xB1\x00\xEC\xB0\xB2\x00"
-     "\xEC\xB0"
-     "\xB3\x00\xEC\xB0\xB4\x00\xEC\xB0\xB5\x00\xEC\xB0\xB6\x00\xEC\xB0\xB7\x00"
-     "\xEC\xB0"
-     "\xB8\x00\xEC\xB0\xB8\xEA\xB3\xA0\x00\xEC\xB0\xB9\x00\xEC\xB0\xBA\x00\xEC"
-     "\xB0\xBB"
-     "\x00\xEC\xB0\xBC\x00\xEC\xB0\xBD\x00\xEC\xB0\xBE\x00\xEC\xB0\xBF\x00\xEC"
-     "\xB1\x80"
-     "\x00\xEC\xB1\x81\x00\xEC\xB1\x82\x00\xEC\xB1\x83\x00\xEC\xB1\x84\x00\xEC"
-     "\xB1\x85"
-     "\x00\xEC\xB1\x86\x00\xEC\xB1\x87\x00\xEC\xB1\x88\x00\xEC\xB1\x89\x00\xEC"
-     "\xB1\x8A"
-     "\x00\xEC\xB1\x8B\x00\xEC\xB1\x8C\x00\xEC\xB1\x8D\x00\xEC\xB1\x8E\x00\xEC"
-     "\xB1\x8F"
-     "\x00\xEC\xB1\x90\x00\xEC\xB1\x91\x00\xEC\xB1\x92\x00\xEC\xB1\x93\x00\xEC"
-     "\xB1\x94"
-     "\x00\xEC\xB1\x95\x00\xEC\xB1\x96\x00\xEC\xB1\x97\x00\xEC\xB1\x98\x00\xEC"
-     "\xB1\x99"
-     "\x00\xEC\xB1\x9A\x00\xEC\xB1\x9B\x00\xEC\xB1\x9C\x00\xEC\xB1\x9D\x00\xEC"
-     "\xB1\x9E"
-     "\x00\xEC\xB1\x9F\x00\xEC\xB1\xA0\x00\xEC\xB1\xA1\x00\xEC\xB1\xA2\x00\xEC"
-     "\xB1\xA3"
-     "\x00\xEC\xB1\xA4\x00\xEC\xB1\xA5\x00\xEC\xB1\xA6\x00\xEC\xB1\xA7\x00\xEC"
-     "\xB1\xA8"
-     "\x00\xEC\xB1\xA9\x00\xEC\xB1\xAA\x00\xEC\xB1\xAB\x00\xEC\xB1\xAC\x00\xEC"
-     "\xB1\xAD"
-     "\x00\xEC\xB1\xAE\x00\xEC\xB1\xAF\x00\xEC\xB1\xB0\x00\xEC\xB1\xB1\x00\xEC"
-     "\xB1\xB2"
-     "\x00\xEC\xB1\xB3\x00\xEC\xB1\xB4\x00\xEC\xB1\xB5\x00\xEC\xB1\xB6\x00\xEC"
-     "\xB1\xB7"
-     "\x00\xEC\xB1\xB8\x00\xEC\xB1\xB9\x00\xEC\xB1\xBA\x00\xEC\xB1\xBB\x00\xEC"
-     "\xB1\xBC"
-     "\x00\xEC\xB1\xBD\x00\xEC\xB1\xBE\x00\xEC\xB1\xBF\x00\xEC\xB2\x80\x00\xEC"
-     "\xB2\x81"
-     "\x00\xEC\xB2\x82\x00\xEC\xB2\x83\x00\xEC\xB2\x84\x00\xEC\xB2\x85\x00\xEC"
-     "\xB2\x86"
-     "\x00\xEC\xB2\x87\x00\xEC\xB2\x88\x00\xEC\xB2\x89\x00\xEC\xB2\x8A\x00\xEC"
-     "\xB2\x8B"
-     "\x00\xEC\xB2\x8C\x00\xEC\xB2\x8D\x00\xEC\xB2\x8E\x00\xEC\xB2\x8F\x00\xEC"
-     "\xB2\x90"
-     "\x00\xEC\xB2\x91\x00\xEC\xB2\x92\x00\xEC\xB2\x93\x00\xEC\xB2\x94\x00\xEC"
-     "\xB2\x95"
-     "\x00\xEC\xB2\x96\x00\xEC\xB2\x97\x00\xEC\xB2\x98\x00\xEC\xB2\x99\x00\xEC"
-     "\xB2\x9A"
-     "\x00\xEC\xB2\x9B\x00\xEC\xB2\x9C\x00\xEC\xB2\x9D\x00\xEC\xB2\x9E\x00\xEC"
-     "\xB2\x9F"
-     "\x00\xEC\xB2\xA0\x00\xEC\xB2\xA1\x00\xEC\xB2\xA2\x00\xEC\xB2\xA3\x00\xEC"
-     "\xB2\xA4"
-     "\x00\xEC\xB2\xA5\x00\xEC\xB2\xA6\x00\xEC\xB2\xA7\x00\xEC\xB2\xA8\x00\xEC"
-     "\xB2\xA9"
-     "\x00\xEC\xB2\xAA\x00\xEC\xB2\xAB\x00\xEC\xB2\xAC\x00\xEC\xB2\xAD\x00\xEC"
-     "\xB2\xAE"
-     "\x00\xEC\xB2\xAF\x00\xEC\xB2\xB0\x00\xEC\xB2\xB1\x00\xEC\xB2\xB2\x00\xEC"
-     "\xB2\xB3"
-     "\x00\xEC\xB2\xB4\x00\xEC\xB2\xB5\x00\xEC\xB2\xB6\x00\xEC\xB2\xB7\x00\xEC"
-     "\xB2\xB8"
-     "\x00\xEC\xB2\xB9\x00\xEC\xB2\xBA\x00\xEC\xB2\xBB\x00\xEC\xB2\xBC\x00\xEC"
-     "\xB2\xBD"
-     "\x00\xEC\xB2\xBE\x00\xEC\xB2\xBF\x00\xEC\xB3\x80\x00\xEC\xB3\x81\x00\xEC"
-     "\xB3\x82"
-     "\x00\xEC\xB3\x83\x00\xEC\xB3\x84\x00\xEC\xB3\x85\x00\xEC\xB3\x86\x00\xEC"
-     "\xB3\x87"
-     "\x00\xEC\xB3\x88\x00\xEC\xB3\x89\x00\xEC\xB3\x8A\x00\xEC\xB3\x8B\x00\xEC"
-     "\xB3\x8C"
-     "\x00\xEC\xB3\x8D\x00\xEC\xB3\x8E\x00\xEC\xB3\x8F\x00\xEC\xB3\x90\x00\xEC"
-     "\xB3\x91"
-     "\x00\xEC\xB3\x92\x00\xEC\xB3\x93\x00\xEC\xB3\x94\x00\xEC\xB3\x95\x00\xEC"
-     "\xB3\x96"
-     "\x00\xEC\xB3\x97\x00\xEC\xB3\x98\x00\xEC\xB3\x99\x00\xEC\xB3\x9A\x00\xEC"
-     "\xB3\x9B"
-     "\x00\xEC\xB3\x9C\x00\xEC\xB3\x9D\x00\xEC\xB3\x9E\x00\xEC\xB3\x9F\x00\xEC"
-     "\xB3\xA0"
-     "\x00\xEC\xB3\xA1\x00\xEC\xB3\xA2\x00\xEC\xB3\xA3\x00\xEC\xB3\xA4\x00\xEC"
-     "\xB3\xA5"
-     "\x00\xEC\xB3\xA6\x00\xEC\xB3\xA7\x00\xEC\xB3\xA8\x00\xEC\xB3\xA9\x00\xEC"
-     "\xB3\xAA"
-     "\x00\xEC\xB3\xAB\x00\xEC\xB3\xAC\x00\xEC\xB3\xAD\x00\xEC\xB3\xAE\x00\xEC"
-     "\xB3\xAF"
-     "\x00\xEC\xB3\xB0\x00\xEC\xB3\xB1\x00\xEC\xB3\xB2\x00\xEC\xB3\xB3\x00\xEC"
-     "\xB3\xB4"
-     "\x00\xEC\xB3\xB5\x00\xEC\xB3\xB6\x00\xEC\xB3\xB7\x00\xEC\xB3\xB8\x00\xEC"
-     "\xB3\xB9"
-     "\x00\xEC\xB3\xBA\x00\xEC\xB3\xBB\x00\xEC\xB3\xBC\x00\xEC\xB3\xBD\x00\xEC"
-     "\xB3\xBE"
-     "\x00\xEC\xB3\xBF\x00\xEC\xB4\x80\x00\xEC\xB4\x81\x00\xEC\xB4\x82\x00\xEC"
-     "\xB4\x83"
-     "\x00\xEC\xB4\x84\x00\xEC\xB4\x85\x00\xEC\xB4\x86\x00\xEC\xB4\x87\x00\xEC"
-     "\xB4\x88"
-     "\x00\xEC\xB4\x89\x00\xEC\xB4\x8A\x00\xEC\xB4\x8B\x00\xEC\xB4\x8C\x00\xEC"
-     "\xB4\x8D"
-     "\x00\xEC\xB4\x8E\x00\xEC\xB4\x8F\x00\xEC\xB4\x90\x00\xEC\xB4\x91\x00\xEC"
-     "\xB4\x92"
-     "\x00\xEC\xB4\x93\x00\xEC\xB4\x94\x00\xEC\xB4\x95\x00\xEC\xB4\x96\x00\xEC"
-     "\xB4\x97"
-     "\x00\xEC\xB4\x98\x00\xEC\xB4\x99\x00\xEC\xB4\x9A\x00\xEC\xB4\x9B\x00\xEC"
-     "\xB4\x9C"
-     "\x00\xEC\xB4\x9D\x00\xEC\xB4\x9E\x00\xEC\xB4\x9F\x00\xEC\xB4\xA0\x00\xEC"
-     "\xB4\xA1"
-     "\x00\xEC\xB4\xA2\x00\xEC\xB4\xA3\x00\xEC\xB4\xA4\x00\xEC\xB4\xA5\x00\xEC"
-     "\xB4\xA6"
-     "\x00\xEC\xB4\xA7\x00\xEC\xB4\xA8\x00\xEC\xB4\xA9\x00\xEC\xB4\xAA\x00\xEC"
-     "\xB4\xAB"
-     "\x00\xEC\xB4\xAC\x00\xEC\xB4\xAD\x00\xEC\xB4\xAE\x00\xEC\xB4\xAF\x00\xEC"
-     "\xB4\xB0"
-     "\x00\xEC\xB4\xB1\x00\xEC\xB4\xB2\x00\xEC\xB4\xB3\x00\xEC\xB4\xB4\x00\xEC"
-     "\xB4\xB5"
-     "\x00\xEC\xB4\xB6\x00\xEC\xB4\xB7\x00\xEC\xB4\xB8\x00\xEC\xB4\xB9\x00\xEC"
-     "\xB4\xBA"
-     "\x00\xEC\xB4\xBB\x00\xEC\xB4\xBC\x00\xEC\xB4\xBD\x00\xEC\xB4\xBE\x00\xEC"
-     "\xB4\xBF"
-     "\x00\xEC\xB5\x80\x00\xEC\xB5\x81\x00\xEC\xB5\x82\x00\xEC\xB5\x83\x00\xEC"
-     "\xB5\x84"
-     "\x00\xEC\xB5\x85\x00\xEC\xB5\x86\x00\xEC\xB5\x87\x00\xEC\xB5\x88\x00\xEC"
-     "\xB5\x89"
-     "\x00\xEC\xB5\x8A\x00\xEC\xB5\x8B\x00\xEC\xB5\x8C\x00\xEC\xB5\x8D\x00\xEC"
-     "\xB5\x8E"
-     "\x00\xEC\xB5\x8F\x00\xEC\xB5\x90\x00\xEC\xB5\x91\x00\xEC\xB5\x92\x00\xEC"
-     "\xB5\x93"
-     "\x00\xEC\xB5\x94\x00\xEC\xB5\x95\x00\xEC\xB5\x96\x00\xEC\xB5\x97\x00\xEC"
-     "\xB5\x98"
-     "\x00\xEC\xB5\x99\x00\xEC\xB5\x9A\x00\xEC\xB5\x9B\x00\xEC\xB5\x9C\x00\xEC"
-     "\xB5\x9D"
-     "\x00\xEC\xB5\x9E\x00\xEC\xB5\x9F\x00\xEC\xB5\xA0\x00\xEC\xB5\xA1\x00\xEC"
-     "\xB5\xA2"
-     "\x00\xEC\xB5\xA3\x00\xEC\xB5\xA4\x00\xEC\xB5\xA5\x00\xEC\xB5\xA6\x00\xEC"
-     "\xB5\xA7"
-     "\x00\xEC\xB5\xA8\x00\xEC\xB5\xA9\x00\xEC\xB5\xAA\x00\xEC\xB5\xAB\x00\xEC"
-     "\xB5\xAC"
-     "\x00\xEC\xB5\xAD\x00\xEC\xB5\xAE\x00\xEC\xB5\xAF\x00\xEC\xB5\xB0\x00\xEC"
-     "\xB5\xB1"
-     "\x00\xEC\xB5\xB2\x00\xEC\xB5\xB3\x00\xEC\xB5\xB4\x00\xEC\xB5\xB5\x00\xEC"
-     "\xB5\xB6"
-     "\x00\xEC\xB5\xB7\x00\xEC\xB5\xB8\x00\xEC\xB5\xB9\x00\xEC\xB5\xBA\x00\xEC"
-     "\xB5\xBB"
-     "\x00\xEC\xB5\xBC\x00\xEC\xB5\xBD\x00\xEC\xB5\xBE\x00\xEC\xB5\xBF\x00\xEC"
-     "\xB6\x80"
-     "\x00\xEC\xB6\x81\x00\xEC\xB6\x82\x00\xEC\xB6\x83\x00\xEC\xB6\x84\x00\xEC"
-     "\xB6\x85"
-     "\x00\xEC\xB6\x86\x00\xEC\xB6\x87\x00\xEC\xB6\x88\x00\xEC\xB6\x89\x00\xEC"
-     "\xB6\x8A"
-     "\x00\xEC\xB6\x8B\x00\xEC\xB6\x8C\x00\xEC\xB6\x8D\x00\xEC\xB6\x8E\x00\xEC"
-     "\xB6\x8F"
-     "\x00\xEC\xB6\x90\x00\xEC\xB6\x91\x00\xEC\xB6\x92\x00\xEC\xB6\x93\x00\xEC"
-     "\xB6\x94"
-     "\x00\xEC\xB6\x95\x00\xEC\xB6\x96\x00\xEC\xB6\x97\x00\xEC\xB6\x98\x00\xEC"
-     "\xB6\x99"
-     "\x00\xEC\xB6\x9A\x00\xEC\xB6\x9B\x00\xEC\xB6\x9C\x00\xEC\xB6\x9D\x00\xEC"
-     "\xB6\x9E"
-     "\x00\xEC\xB6\x9F\x00\xEC\xB6\xA0\x00\xEC\xB6\xA1\x00\xEC\xB6\xA2\x00\xEC"
-     "\xB6\xA3"
-     "\x00\xEC\xB6\xA4\x00\xEC\xB6\xA5\x00\xEC\xB6\xA6\x00\xEC\xB6\xA7\x00\xEC"
-     "\xB6\xA8"
-     "\x00\xEC\xB6\xA9\x00\xEC\xB6\xAA\x00\xEC\xB6\xAB\x00\xEC\xB6\xAC\x00\xEC"
-     "\xB6\xAD"
-     "\x00\xEC\xB6\xAE\x00\xEC\xB6\xAF\x00\xEC\xB6\xB0\x00\xEC\xB6\xB1\x00\xEC"
-     "\xB6\xB2"
-     "\x00\xEC\xB6\xB3\x00\xEC\xB6\xB4\x00\xEC\xB6\xB5\x00\xEC\xB6\xB6\x00\xEC"
-     "\xB6\xB7"
-     "\x00\xEC\xB6\xB8\x00\xEC\xB6\xB9\x00\xEC\xB6\xBA\x00\xEC\xB6\xBB\x00\xEC"
-     "\xB6\xBC"
-     "\x00\xEC\xB6\xBD\x00\xEC\xB6\xBE\x00\xEC\xB6\xBF\x00\xEC\xB7\x80\x00\xEC"
-     "\xB7\x81"
-     "\x00\xEC\xB7\x82\x00\xEC\xB7\x83\x00\xEC\xB7\x84\x00\xEC\xB7\x85\x00\xEC"
-     "\xB7\x86"
-     "\x00\xEC\xB7\x87\x00\xEC\xB7\x88\x00\xEC\xB7\x89\x00\xEC\xB7\x8A\x00\xEC"
-     "\xB7\x8B"
-     "\x00\xEC\xB7\x8C\x00\xEC\xB7\x8D\x00\xEC\xB7\x8E\x00\xEC\xB7\x8F\x00\xEC"
-     "\xB7\x90"
-     "\x00\xEC\xB7\x91\x00\xEC\xB7\x92\x00\xEC\xB7\x93\x00\xEC\xB7\x94\x00\xEC"
-     "\xB7\x95"
-     "\x00\xEC\xB7\x96\x00\xEC\xB7\x97\x00\xEC\xB7\x98\x00\xEC\xB7\x99\x00\xEC"
-     "\xB7\x9A"
-     "\x00\xEC\xB7\x9B\x00\xEC\xB7\x9C\x00\xEC\xB7\x9D\x00\xEC\xB7\x9E\x00\xEC"
-     "\xB7\x9F"
-     "\x00\xEC\xB7\xA0\x00\xEC\xB7\xA1\x00\xEC\xB7\xA2\x00\xEC\xB7\xA3\x00\xEC"
-     "\xB7\xA4"
-     "\x00\xEC\xB7\xA5\x00\xEC\xB7\xA6\x00\xEC\xB7\xA7\x00\xEC\xB7\xA8\x00\xEC"
-     "\xB7\xA9"
-     "\x00\xEC\xB7\xAA\x00\xEC\xB7\xAB\x00\xEC\xB7\xAC\x00\xEC\xB7\xAD\x00\xEC"
-     "\xB7\xAE"
-     "\x00\xEC\xB7\xAF\x00\xEC\xB7\xB0\x00\xEC\xB7\xB1\x00\xEC\xB7\xB2\x00\xEC"
-     "\xB7\xB3"
-     "\x00\xEC\xB7\xB4\x00\xEC\xB7\xB5\x00\xEC\xB7\xB6\x00\xEC\xB7\xB7\x00\xEC"
-     "\xB7\xB8"
-     "\x00\xEC\xB7\xB9\x00\xEC\xB7\xBA\x00\xEC\xB7\xBB\x00\xEC\xB7\xBC\x00\xEC"
-     "\xB7\xBD"
-     "\x00\xEC\xB7\xBE\x00\xEC\xB7\xBF\x00\xEC\xB8\x80\x00\xEC\xB8\x81\x00\xEC"
-     "\xB8\x82"
-     "\x00\xEC\xB8\x83\x00\xEC\xB8\x84\x00\xEC\xB8\x85\x00\xEC\xB8\x86\x00\xEC"
-     "\xB8\x87"
-     "\x00\xEC\xB8\x88\x00\xEC\xB8\x89\x00\xEC\xB8\x8A\x00\xEC\xB8\x8B\x00\xEC"
-     "\xB8\x8C"
-     "\x00\xEC\xB8\x8D\x00\xEC\xB8\x8E\x00\xEC\xB8\x8F\x00\xEC\xB8\x90\x00\xEC"
-     "\xB8\x91"
-     "\x00\xEC\xB8\x92\x00\xEC\xB8\x93\x00\xEC\xB8\x94\x00\xEC\xB8\x95\x00\xEC"
-     "\xB8\x96"
-     "\x00\xEC\xB8\x97\x00\xEC\xB8\x98\x00\xEC\xB8\x99\x00\xEC\xB8\x9A\x00\xEC"
-     "\xB8\x9B"
-     "\x00\xEC\xB8\x9C\x00\xEC\xB8\x9D\x00\xEC\xB8\x9E\x00\xEC\xB8\x9F\x00\xEC"
-     "\xB8\xA0"
-     "\x00\xEC\xB8\xA1\x00\xEC\xB8\xA2\x00\xEC\xB8\xA3\x00\xEC\xB8\xA4\x00\xEC"
-     "\xB8\xA5"
-     "\x00\xEC\xB8\xA6\x00\xEC\xB8\xA7\x00\xEC\xB8\xA8\x00\xEC\xB8\xA9\x00\xEC"
-     "\xB8\xAA"
-     "\x00\xEC\xB8\xAB\x00\xEC\xB8\xAC\x00\xEC\xB8\xAD\x00\xEC\xB8\xAE\x00\xEC"
-     "\xB8\xAF"
-     "\x00\xEC\xB8\xB0\x00\xEC\xB8\xB1\x00\xEC\xB8\xB2\x00\xEC\xB8\xB3\x00\xEC"
-     "\xB8\xB4"
-     "\x00\xEC\xB8\xB5\x00\xEC\xB8\xB6\x00\xEC\xB8\xB7\x00\xEC\xB8\xB8\x00\xEC"
-     "\xB8\xB9"
-     "\x00\xEC\xB8\xBA\x00\xEC\xB8\xBB\x00\xEC\xB8\xBC\x00\xEC\xB8\xBD\x00\xEC"
-     "\xB8\xBE"
-     "\x00\xEC\xB8\xBF\x00\xEC\xB9\x80\x00\xEC\xB9\x81\x00\xEC\xB9\x82\x00\xEC"
-     "\xB9\x83"
-     "\x00\xEC\xB9\x84\x00\xEC\xB9\x85\x00\xEC\xB9\x86\x00\xEC\xB9\x87\x00\xEC"
-     "\xB9\x88"
-     "\x00\xEC\xB9\x89\x00\xEC\xB9\x8A\x00\xEC\xB9\x8B\x00\xEC\xB9\x8C\x00\xEC"
-     "\xB9\x8D"
-     "\x00\xEC\xB9\x8E\x00\xEC\xB9\x8F\x00\xEC\xB9\x90\x00\xEC\xB9\x91\x00\xEC"
-     "\xB9\x92"
-     "\x00\xEC\xB9\x93\x00\xEC\xB9\x94\x00\xEC\xB9\x95\x00\xEC\xB9\x96\x00\xEC"
-     "\xB9\x97"
-     "\x00\xEC\xB9\x98\x00\xEC\xB9\x99\x00\xEC\xB9\x9A\x00\xEC\xB9\x9B\x00\xEC"
-     "\xB9\x9C"
-     "\x00\xEC\xB9\x9D\x00\xEC\xB9\x9E\x00\xEC\xB9\x9F\x00\xEC\xB9\xA0\x00\xEC"
-     "\xB9\xA1"
-     "\x00\xEC\xB9\xA2\x00\xEC\xB9\xA3\x00\xEC\xB9\xA4\x00\xEC\xB9\xA5\x00\xEC"
-     "\xB9\xA6"
-     "\x00\xEC\xB9\xA7\x00\xEC\xB9\xA8\x00\xEC\xB9\xA9\x00\xEC\xB9\xAA\x00\xEC"
-     "\xB9\xAB"
-     "\x00\xEC\xB9\xAC\x00\xEC\xB9\xAD\x00\xEC\xB9\xAE\x00\xEC\xB9\xAF\x00\xEC"
-     "\xB9\xB0"
-     "\x00\xEC\xB9\xB1\x00\xEC\xB9\xB2\x00\xEC\xB9\xB3\x00\xEC\xB9\xB4\x00\xEC"
-     "\xB9\xB5"
-     "\x00\xEC\xB9\xB6\x00\xEC\xB9\xB7\x00\xEC\xB9\xB8\x00\xEC\xB9\xB9\x00\xEC"
-     "\xB9\xBA"
-     "\x00\xEC\xB9\xBB\x00\xEC\xB9\xBC\x00\xEC\xB9\xBD\x00\xEC\xB9\xBE\x00\xEC"
-     "\xB9\xBF"
-     "\x00\xEC\xBA\x80\x00\xEC\xBA\x81\x00\xEC\xBA\x82\x00\xEC\xBA\x83\x00\xEC"
-     "\xBA\x84"
-     "\x00\xEC\xBA\x85\x00\xEC\xBA\x86\x00\xEC\xBA\x87\x00\xEC\xBA\x88\x00\xEC"
-     "\xBA\x89"
-     "\x00\xEC\xBA\x8A\x00\xEC\xBA\x8B\x00\xEC\xBA\x8C\x00\xEC\xBA\x8D\x00\xEC"
-     "\xBA\x8E"
-     "\x00\xEC\xBA\x8F\x00\xEC\xBA\x90\x00\xEC\xBA\x91\x00\xEC\xBA\x92\x00\xEC"
-     "\xBA\x93"
-     "\x00\xEC\xBA\x94\x00\xEC\xBA\x95\x00\xEC\xBA\x96\x00\xEC\xBA\x97\x00\xEC"
-     "\xBA\x98"
-     "\x00\xEC\xBA\x99\x00\xEC\xBA\x9A\x00\xEC\xBA\x9B\x00\xEC\xBA\x9C\x00\xEC"
-     "\xBA\x9D"
-     "\x00\xEC\xBA\x9E\x00\xEC\xBA\x9F\x00\xEC\xBA\xA0\x00\xEC\xBA\xA1\x00\xEC"
-     "\xBA\xA2"
-     "\x00\xEC\xBA\xA3\x00\xEC\xBA\xA4\x00\xEC\xBA\xA5\x00\xEC\xBA\xA6\x00\xEC"
-     "\xBA\xA7"
-     "\x00\xEC\xBA\xA8\x00\xEC\xBA\xA9\x00\xEC\xBA\xAA\x00\xEC\xBA\xAB\x00\xEC"
-     "\xBA\xAC"
-     "\x00\xEC\xBA\xAD\x00\xEC\xBA\xAE\x00\xEC\xBA\xAF\x00\xEC\xBA\xB0\x00\xEC"
-     "\xBA\xB1"
-     "\x00\xEC\xBA\xB2\x00\xEC\xBA\xB3\x00\xEC\xBA\xB4\x00\xEC\xBA\xB5\x00\xEC"
-     "\xBA\xB6"
-     "\x00\xEC\xBA\xB7\x00\xEC\xBA\xB8\x00\xEC\xBA\xB9\x00\xEC\xBA\xBA\x00\xEC"
-     "\xBA\xBB"
-     "\x00\xEC\xBA\xBC\x00\xEC\xBA\xBD\x00\xEC\xBA\xBE\x00\xEC\xBA\xBF\x00\xEC"
-     "\xBB\x80"
-     "\x00\xEC\xBB\x81\x00\xEC\xBB\x82\x00\xEC\xBB\x83\x00\xEC\xBB\x84\x00\xEC"
-     "\xBB\x85"
-     "\x00\xEC\xBB\x86\x00\xEC\xBB\x87\x00\xEC\xBB\x88\x00\xEC\xBB\x89\x00\xEC"
-     "\xBB\x8A"
-     "\x00\xEC\xBB\x8B\x00\xEC\xBB\x8C\x00\xEC\xBB\x8D\x00\xEC\xBB\x8E\x00\xEC"
-     "\xBB\x8F"
-     "\x00\xEC\xBB\x90\x00\xEC\xBB\x91\x00\xEC\xBB\x92\x00\xEC\xBB\x93\x00\xEC"
-     "\xBB\x94"
-     "\x00\xEC\xBB\x95\x00\xEC\xBB\x96\x00\xEC\xBB\x97\x00\xEC\xBB\x98\x00\xEC"
-     "\xBB\x99"
-     "\x00\xEC\xBB\x9A\x00\xEC\xBB\x9B\x00\xEC\xBB\x9C\x00\xEC\xBB\x9D\x00\xEC"
-     "\xBB\x9E"
-     "\x00\xEC\xBB\x9F\x00\xEC\xBB\xA0\x00\xEC\xBB\xA1\x00\xEC\xBB\xA2\x00\xEC"
-     "\xBB\xA3"
-     "\x00\xEC\xBB\xA4\x00\xEC\xBB\xA5\x00\xEC\xBB\xA6\x00\xEC\xBB\xA7\x00\xEC"
-     "\xBB\xA8"
-     "\x00\xEC\xBB\xA9\x00\xEC\xBB\xAA\x00\xEC\xBB\xAB\x00\xEC\xBB\xAC\x00\xEC"
-     "\xBB\xAD"
-     "\x00\xEC\xBB\xAE\x00\xEC\xBB\xAF\x00\xEC\xBB\xB0\x00\xEC\xBB\xB1\x00\xEC"
-     "\xBB\xB2"
-     "\x00\xEC\xBB\xB3\x00\xEC\xBB\xB4\x00\xEC\xBB\xB5\x00\xEC\xBB\xB6\x00\xEC"
-     "\xBB\xB7"
-     "\x00\xEC\xBB\xB8\x00\xEC\xBB\xB9\x00\xEC\xBB\xBA\x00\xEC\xBB\xBB\x00\xEC"
-     "\xBB\xBC"
-     "\x00\xEC\xBB\xBD\x00\xEC\xBB\xBE\x00\xEC\xBB\xBF\x00\xEC\xBC\x80\x00\xEC"
-     "\xBC\x81"
-     "\x00\xEC\xBC\x82\x00\xEC\xBC\x83\x00\xEC\xBC\x84\x00\xEC\xBC\x85\x00\xEC"
-     "\xBC\x86"
-     "\x00\xEC\xBC\x87\x00\xEC\xBC\x88\x00\xEC\xBC\x89\x00\xEC\xBC\x8A\x00\xEC"
-     "\xBC\x8B"
-     "\x00\xEC\xBC\x8C\x00\xEC\xBC\x8D\x00\xEC\xBC\x8E\x00\xEC\xBC\x8F\x00\xEC"
-     "\xBC\x90"
-     "\x00\xEC\xBC\x91\x00\xEC\xBC\x92\x00\xEC\xBC\x93\x00\xEC\xBC\x94\x00\xEC"
-     "\xBC\x95"
-     "\x00\xEC\xBC\x96\x00\xEC\xBC\x97\x00\xEC\xBC\x98\x00\xEC\xBC\x99\x00\xEC"
-     "\xBC\x9A"
-     "\x00\xEC\xBC\x9B\x00\xEC\xBC\x9C\x00\xEC\xBC\x9D\x00\xEC\xBC\x9E\x00\xEC"
-     "\xBC\x9F"
-     "\x00\xEC\xBC\xA0\x00\xEC\xBC\xA1\x00\xEC\xBC\xA2\x00\xEC\xBC\xA3\x00\xEC"
-     "\xBC\xA4"
-     "\x00\xEC\xBC\xA5\x00\xEC\xBC\xA6\x00\xEC\xBC\xA7\x00\xEC\xBC\xA8\x00\xEC"
-     "\xBC\xA9"
-     "\x00\xEC\xBC\xAA\x00\xEC\xBC\xAB\x00\xEC\xBC\xAC\x00\xEC\xBC\xAD\x00\xEC"
-     "\xBC\xAE"
-     "\x00\xEC\xBC\xAF\x00\xEC\xBC\xB0\x00\xEC\xBC\xB1\x00\xEC\xBC\xB2\x00\xEC"
-     "\xBC\xB3"
-     "\x00\xEC\xBC\xB4\x00\xEC\xBC\xB5\x00\xEC\xBC\xB6\x00\xEC\xBC\xB7\x00\xEC"
-     "\xBC\xB8"
-     "\x00\xEC\xBC\xB9\x00\xEC\xBC\xBA\x00\xEC\xBC\xBB\x00\xEC\xBC\xBC\x00\xEC"
-     "\xBC\xBD"
-     "\x00\xEC\xBC\xBE\x00\xEC\xBC\xBF\x00\xEC\xBD\x80\x00\xEC\xBD\x81\x00\xEC"
-     "\xBD\x82"
-     "\x00\xEC\xBD\x83\x00\xEC\xBD\x84\x00\xEC\xBD\x85\x00\xEC\xBD\x86\x00\xEC"
-     "\xBD\x87"
-     "\x00\xEC\xBD\x88\x00\xEC\xBD\x89\x00\xEC\xBD\x8A\x00\xEC\xBD\x8B\x00\xEC"
-     "\xBD\x8C"
-     "\x00\xEC\xBD\x8D\x00\xEC\xBD\x8E\x00\xEC\xBD\x8F\x00\xEC\xBD\x90\x00\xEC"
-     "\xBD\x91"
-     "\x00\xEC\xBD\x92\x00\xEC\xBD\x93\x00\xEC\xBD\x94\x00\xEC\xBD\x95\x00\xEC"
-     "\xBD\x96"
-     "\x00\xEC\xBD\x97\x00\xEC\xBD\x98\x00\xEC\xBD\x99\x00\xEC\xBD\x9A\x00\xEC"
-     "\xBD\x9B"
-     "\x00\xEC\xBD\x9C\x00\xEC\xBD\x9D\x00\xEC\xBD\x9E\x00\xEC\xBD\x9F\x00\xEC"
-     "\xBD\xA0"
-     "\x00\xEC\xBD\xA1\x00\xEC\xBD\xA2\x00\xEC\xBD\xA3\x00\xEC\xBD\xA4\x00\xEC"
-     "\xBD\xA5"
-     "\x00\xEC\xBD\xA6\x00\xEC\xBD\xA7\x00\xEC\xBD\xA8\x00\xEC\xBD\xA9\x00\xEC"
-     "\xBD\xAA"
-     "\x00\xEC\xBD\xAB\x00\xEC\xBD\xAC\x00\xEC\xBD\xAD\x00\xEC\xBD\xAE\x00\xEC"
-     "\xBD\xAF"
-     "\x00\xEC\xBD\xB0\x00\xEC\xBD\xB1\x00\xEC\xBD\xB2\x00\xEC\xBD\xB3\x00\xEC"
-     "\xBD\xB4"
-     "\x00\xEC\xBD\xB5\x00\xEC\xBD\xB6\x00\xEC\xBD\xB7\x00\xEC\xBD\xB8\x00\xEC"
-     "\xBD\xB9"
-     "\x00\xEC\xBD\xBA\x00\xEC\xBD\xBB\x00\xEC\xBD\xBC\x00\xEC\xBD\xBD\x00\xEC"
-     "\xBD\xBE"
-     "\x00\xEC\xBD\xBF\x00\xEC\xBE\x80\x00\xEC\xBE\x81\x00\xEC\xBE\x82\x00\xEC"
-     "\xBE\x83"
-     "\x00\xEC\xBE\x84\x00\xEC\xBE\x85\x00\xEC\xBE\x86\x00\xEC\xBE\x87\x00\xEC"
-     "\xBE\x88"
-     "\x00\xEC\xBE\x89\x00\xEC\xBE\x8A\x00\xEC\xBE\x8B\x00\xEC\xBE\x8C\x00\xEC"
-     "\xBE\x8D"
-     "\x00\xEC\xBE\x8E\x00\xEC\xBE\x8F\x00\xEC\xBE\x90\x00\xEC\xBE\x91\x00\xEC"
-     "\xBE\x92"
-     "\x00\xEC\xBE\x93\x00\xEC\xBE\x94\x00\xEC\xBE\x95\x00\xEC\xBE\x96\x00\xEC"
-     "\xBE\x97"
-     "\x00\xEC\xBE\x98\x00\xEC\xBE\x99\x00\xEC\xBE\x9A\x00\xEC\xBE\x9B\x00\xEC"
-     "\xBE\x9C"
-     "\x00\xEC\xBE\x9D\x00\xEC\xBE\x9E\x00\xEC\xBE\x9F\x00\xEC\xBE\xA0\x00\xEC"
-     "\xBE\xA1"
-     "\x00\xEC\xBE\xA2\x00\xEC\xBE\xA3\x00\xEC\xBE\xA4\x00\xEC\xBE\xA5\x00\xEC"
-     "\xBE\xA6"
-     "\x00\xEC\xBE\xA7\x00\xEC\xBE\xA8\x00\xEC\xBE\xA9\x00\xEC\xBE\xAA\x00\xEC"
-     "\xBE\xAB"
-     "\x00\xEC\xBE\xAC\x00\xEC\xBE\xAD\x00\xEC\xBE\xAE\x00\xEC\xBE\xAF\x00\xEC"
-     "\xBE\xB0"
-     "\x00\xEC\xBE\xB1\x00\xEC\xBE\xB2\x00\xEC\xBE\xB3\x00\xEC\xBE\xB4\x00\xEC"
-     "\xBE\xB5"
-     "\x00\xEC\xBE\xB6\x00\xEC\xBE\xB7\x00\xEC\xBE\xB8\x00\xEC\xBE\xB9\x00\xEC"
-     "\xBE\xBA"
-     "\x00\xEC\xBE\xBB\x00\xEC\xBE\xBC\x00\xEC\xBE\xBD\x00\xEC\xBE\xBE\x00\xEC"
-     "\xBE\xBF"
-     "\x00\xEC\xBF\x80\x00\xEC\xBF\x81\x00\xEC\xBF\x82\x00\xEC\xBF\x83\x00\xEC"
-     "\xBF\x84"
-     "\x00\xEC\xBF\x85\x00\xEC\xBF\x86\x00\xEC\xBF\x87\x00\xEC\xBF\x88\x00\xEC"
-     "\xBF\x89"
-     "\x00\xEC\xBF\x8A\x00\xEC\xBF\x8B\x00\xEC\xBF\x8C\x00\xEC\xBF\x8D\x00\xEC"
-     "\xBF\x8E"
-     "\x00\xEC\xBF\x8F\x00\xEC\xBF\x90\x00\xEC\xBF\x91\x00\xEC\xBF\x92\x00\xEC"
-     "\xBF\x93"
-     "\x00\xEC\xBF\x94\x00\xEC\xBF\x95\x00\xEC\xBF\x96\x00\xEC\xBF\x97\x00\xEC"
-     "\xBF\x98"
-     "\x00\xEC\xBF\x99\x00\xEC\xBF\x9A\x00\xEC\xBF\x9B\x00\xEC\xBF\x9C\x00\xEC"
-     "\xBF\x9D"
-     "\x00\xEC\xBF\x9E\x00\xEC\xBF\x9F\x00\xEC\xBF\xA0\x00\xEC\xBF\xA1\x00\xEC"
-     "\xBF\xA2"
-     "\x00\xEC\xBF\xA3\x00\xEC\xBF\xA4\x00\xEC\xBF\xA5\x00\xEC\xBF\xA6\x00\xEC"
-     "\xBF\xA7"
-     "\x00\xEC\xBF\xA8\x00\xEC\xBF\xA9\x00\xEC\xBF\xAA\x00\xEC\xBF\xAB\x00\xEC"
-     "\xBF\xAC"
-     "\x00\xEC\xBF\xAD\x00\xEC\xBF\xAE\x00\xEC\xBF\xAF\x00\xEC\xBF\xB0\x00\xEC"
-     "\xBF\xB1"
-     "\x00\xEC\xBF\xB2\x00\xEC\xBF\xB3\x00\xEC\xBF\xB4\x00\xEC\xBF\xB5\x00\xEC"
-     "\xBF\xB6"
-     "\x00\xEC\xBF\xB7\x00\xEC\xBF\xB8\x00\xEC\xBF\xB9\x00\xEC\xBF\xBA\x00\xEC"
-     "\xBF\xBB"
-     "\x00\xEC\xBF\xBC\x00\xEC\xBF\xBD\x00\xEC\xBF\xBE\x00\xEC\xBF\xBF\x00\xED"
-     "\x80\x80"
-     "\x00\xED\x80\x81\x00\xED\x80\x82\x00\xED\x80\x83\x00\xED\x80\x84\x00\xED"
-     "\x80\x85"
-     "\x00\xED\x80\x86\x00\xED\x80\x87\x00\xED\x80\x88\x00\xED\x80\x89\x00\xED"
-     "\x80\x8A"
-     "\x00\xED\x80\x8B\x00\xED\x80\x8C\x00\xED\x80\x8D\x00\xED\x80\x8E\x00\xED"
-     "\x80\x8F"
-     "\x00\xED\x80\x90\x00\xED\x80\x91\x00\xED\x80\x92\x00\xED\x80\x93\x00\xED"
-     "\x80\x94"
-     "\x00\xED\x80\x95\x00\xED\x80\x96\x00\xED\x80\x97\x00\xED\x80\x98\x00\xED"
-     "\x80\x99"
-     "\x00\xED\x80\x9A\x00\xED\x80\x9B\x00\xED\x80\x9C\x00\xED\x80\x9D\x00\xED"
-     "\x80\x9E"
-     "\x00\xED\x80\x9F\x00\xED\x80\xA0\x00\xED\x80\xA1\x00\xED\x80\xA2\x00\xED"
-     "\x80\xA3"
-     "\x00\xED\x80\xA4\x00\xED\x80\xA5\x00\xED\x80\xA6\x00\xED\x80\xA7\x00\xED"
-     "\x80\xA8"
-     "\x00\xED\x80\xA9\x00\xED\x80\xAA\x00\xED\x80\xAB\x00\xED\x80\xAC\x00\xED"
-     "\x80\xAD"
-     "\x00\xED\x80\xAE\x00\xED\x80\xAF\x00\xED\x80\xB0\x00\xED\x80\xB1\x00\xED"
-     "\x80\xB2"
-     "\x00\xED\x80\xB3\x00\xED\x80\xB4\x00\xED\x80\xB5\x00\xED\x80\xB6\x00\xED"
-     "\x80\xB7"
-     "\x00\xED\x80\xB8\x00\xED\x80\xB9\x00\xED\x80\xBA\x00\xED\x80\xBB\x00\xED"
-     "\x80\xBC"
-     "\x00\xED\x80\xBD\x00\xED\x80\xBE\x00\xED\x80\xBF\x00\xED\x81\x80\x00\xED"
-     "\x81\x81"
-     "\x00\xED\x81\x82\x00\xED\x81\x83\x00\xED\x81\x84\x00\xED\x81\x85\x00\xED"
-     "\x81\x86"
-     "\x00\xED\x81\x87\x00\xED\x81\x88\x00\xED\x81\x89\x00\xED\x81\x8A\x00\xED"
-     "\x81\x8B"
-     "\x00\xED\x81\x8C\x00\xED\x81\x8D\x00\xED\x81\x8E\x00\xED\x81\x8F\x00\xED"
-     "\x81\x90"
-     "\x00\xED\x81\x91\x00\xED\x81\x92\x00\xED\x81\x93\x00\xED\x81\x94\x00\xED"
-     "\x81\x95"
-     "\x00\xED\x81\x96\x00\xED\x81\x97\x00\xED\x81\x98\x00\xED\x81\x99\x00\xED"
-     "\x81\x9A"
-     "\x00\xED\x81\x9B\x00\xED\x81\x9C\x00\xED\x81\x9D\x00\xED\x81\x9E\x00\xED"
-     "\x81\x9F"
-     "\x00\xED\x81\xA0\x00\xED\x81\xA1\x00\xED\x81\xA2\x00\xED\x81\xA3\x00\xED"
-     "\x81\xA4"
-     "\x00\xED\x81\xA5\x00\xED\x81\xA6\x00\xED\x81\xA7\x00\xED\x81\xA8\x00\xED"
-     "\x81\xA9"
-     "\x00\xED\x81\xAA\x00\xED\x81\xAB\x00\xED\x81\xAC\x00\xED\x81\xAD\x00\xED"
-     "\x81\xAE"
-     "\x00\xED\x81\xAF\x00\xED\x81\xB0\x00\xED\x81\xB1\x00\xED\x81\xB2\x00\xED"
-     "\x81\xB3"
-     "\x00\xED\x81\xB4\x00\xED\x81\xB5\x00\xED\x81\xB6\x00\xED\x81\xB7\x00\xED"
-     "\x81\xB8"
-     "\x00\xED\x81\xB9\x00\xED\x81\xBA\x00\xED\x81\xBB\x00\xED\x81\xBC\x00\xED"
-     "\x81\xBD"
-     "\x00\xED\x81\xBE\x00\xED\x81\xBF\x00\xED\x82\x80\x00\xED\x82\x81\x00\xED"
-     "\x82\x82"
-     "\x00\xED\x82\x83\x00\xED\x82\x84\x00\xED\x82\x85\x00\xED\x82\x86\x00\xED"
-     "\x82\x87"
-     "\x00\xED\x82\x88\x00\xED\x82\x89\x00\xED\x82\x8A\x00\xED\x82\x8B\x00\xED"
-     "\x82\x8C"
-     "\x00\xED\x82\x8D\x00\xED\x82\x8E\x00\xED\x82\x8F\x00\xED\x82\x90\x00\xED"
-     "\x82\x91"
-     "\x00\xED\x82\x92\x00\xED\x82\x93\x00\xED\x82\x94\x00\xED\x82\x95\x00\xED"
-     "\x82\x96"
-     "\x00\xED\x82\x97\x00\xED\x82\x98\x00\xED\x82\x99\x00\xED\x82\x9A\x00\xED"
-     "\x82\x9B"
-     "\x00\xED\x82\x9C\x00\xED\x82\x9D\x00\xED\x82\x9E\x00\xED\x82\x9F\x00\xED"
-     "\x82\xA0"
-     "\x00\xED\x82\xA1\x00\xED\x82\xA2\x00\xED\x82\xA3\x00\xED\x82\xA4\x00\xED"
-     "\x82\xA5"
-     "\x00\xED\x82\xA6\x00\xED\x82\xA7\x00\xED\x82\xA8\x00\xED\x82\xA9\x00\xED"
-     "\x82\xAA"
-     "\x00\xED\x82\xAB\x00\xED\x82\xAC\x00\xED\x82\xAD\x00\xED\x82\xAE\x00\xED"
-     "\x82\xAF"
-     "\x00\xED\x82\xB0\x00\xED\x82\xB1\x00\xED\x82\xB2\x00\xED\x82\xB3\x00\xED"
-     "\x82\xB4"
-     "\x00\xED\x82\xB5\x00\xED\x82\xB6\x00\xED\x82\xB7\x00\xED\x82\xB8\x00\xED"
-     "\x82\xB9"
-     "\x00\xED\x82\xBA\x00\xED\x82\xBB\x00\xED\x82\xBC\x00\xED\x82\xBD\x00\xED"
-     "\x82\xBE"
-     "\x00\xED\x82\xBF\x00\xED\x83\x80\x00\xED\x83\x81\x00\xED\x83\x82\x00\xED"
-     "\x83\x83"
-     "\x00\xED\x83\x84\x00\xED\x83\x85\x00\xED\x83\x86\x00\xED\x83\x87\x00\xED"
-     "\x83\x88"
-     "\x00\xED\x83\x89\x00\xED\x83\x8A\x00\xED\x83\x8B\x00\xED\x83\x8C\x00\xED"
-     "\x83\x8D"
-     "\x00\xED\x83\x8E\x00\xED\x83\x8F\x00\xED\x83\x90\x00\xED\x83\x91\x00\xED"
-     "\x83\x92"
-     "\x00\xED\x83\x93\x00\xED\x83\x94\x00\xED\x83\x95\x00\xED\x83\x96\x00\xED"
-     "\x83\x97"
-     "\x00\xED\x83\x98\x00\xED\x83\x99\x00\xED\x83\x9A\x00\xED\x83\x9B\x00\xED"
-     "\x83\x9C"
-     "\x00\xED\x83\x9D\x00\xED\x83\x9E\x00\xED\x83\x9F\x00\xED\x83\xA0\x00\xED"
-     "\x83\xA1"
-     "\x00\xED\x83\xA2\x00\xED\x83\xA3\x00\xED\x83\xA4\x00\xED\x83\xA5\x00\xED"
-     "\x83\xA6"
-     "\x00\xED\x83\xA7\x00\xED\x83\xA8\x00\xED\x83\xA9\x00\xED\x83\xAA\x00\xED"
-     "\x83\xAB"
-     "\x00\xED\x83\xAC\x00\xED\x83\xAD\x00\xED\x83\xAE\x00\xED\x83\xAF\x00\xED"
-     "\x83\xB0"
-     "\x00\xED\x83\xB1\x00\xED\x83\xB2\x00\xED\x83\xB3\x00\xED\x83\xB4\x00\xED"
-     "\x83\xB5"
-     "\x00\xED\x83\xB6\x00\xED\x83\xB7\x00\xED\x83\xB8\x00\xED\x83\xB9\x00\xED"
-     "\x83\xBA"
-     "\x00\xED\x83\xBB\x00\xED\x83\xBC\x00\xED\x83\xBD\x00\xED\x83\xBE\x00\xED"
-     "\x83\xBF"
-     "\x00\xED\x84\x80\x00\xED\x84\x81\x00\xED\x84\x82\x00\xED\x84\x83\x00\xED"
-     "\x84\x84"
-     "\x00\xED\x84\x85\x00\xED\x84\x86\x00\xED\x84\x87\x00\xED\x84\x88\x00\xED"
-     "\x84\x89"
-     "\x00\xED\x84\x8A\x00\xED\x84\x8B\x00\xED\x84\x8C\x00\xED\x84\x8D\x00\xED"
-     "\x84\x8E"
-     "\x00\xED\x84\x8F\x00\xED\x84\x90\x00\xED\x84\x91\x00\xED\x84\x92\x00\xED"
-     "\x84\x93"
-     "\x00\xED\x84\x94\x00\xED\x84\x95\x00\xED\x84\x96\x00\xED\x84\x97\x00\xED"
-     "\x84\x98"
-     "\x00\xED\x84\x99\x00\xED\x84\x9A\x00\xED\x84\x9B\x00\xED\x84\x9C\x00\xED"
-     "\x84\x9D"
-     "\x00\xED\x84\x9E\x00\xED\x84\x9F\x00\xED\x84\xA0\x00\xED\x84\xA1\x00\xED"
-     "\x84\xA2"
-     "\x00\xED\x84\xA3\x00\xED\x84\xA4\x00\xED\x84\xA5\x00\xED\x84\xA6\x00\xED"
-     "\x84\xA7"
-     "\x00\xED\x84\xA8\x00\xED\x84\xA9\x00\xED\x84\xAA\x00\xED\x84\xAB\x00\xED"
-     "\x84\xAC"
-     "\x00\xED\x84\xAD\x00\xED\x84\xAE\x00\xED\x84\xAF\x00\xED\x84\xB0\x00\xED"
-     "\x84\xB1"
-     "\x00\xED\x84\xB2\x00\xED\x84\xB3\x00\xED\x84\xB4\x00\xED\x84\xB5\x00\xED"
-     "\x84\xB6"
-     "\x00\xED\x84\xB7\x00\xED\x84\xB8\x00\xED\x84\xB9\x00\xED\x84\xBA\x00\xED"
-     "\x84\xBB"
-     "\x00\xED\x84\xBC\x00\xED\x84\xBD\x00\xED\x84\xBE\x00\xED\x84\xBF\x00\xED"
-     "\x85\x80"
-     "\x00\xED\x85\x81\x00\xED\x85\x82\x00\xED\x85\x83\x00\xED\x85\x84\x00\xED"
-     "\x85\x85"
-     "\x00\xED\x85\x86\x00\xED\x85\x87\x00\xED\x85\x88\x00\xED\x85\x89\x00\xED"
-     "\x85\x8A"
-     "\x00\xED\x85\x8B\x00\xED\x85\x8C\x00\xED\x85\x8D\x00\xED\x85\x8E\x00\xED"
-     "\x85\x8F"
-     "\x00\xED\x85\x90\x00\xED\x85\x91\x00\xED\x85\x92\x00\xED\x85\x93\x00\xED"
-     "\x85\x94"
-     "\x00\xED\x85\x95\x00\xED\x85\x96\x00\xED\x85\x97\x00\xED\x85\x98\x00\xED"
-     "\x85\x99"
-     "\x00\xED\x85\x9A\x00\xED\x85\x9B\x00\xED\x85\x9C\x00\xED\x85\x9D\x00\xED"
-     "\x85\x9E"
-     "\x00\xED\x85\x9F\x00\xED\x85\xA0\x00\xED\x85\xA1\x00\xED\x85\xA2\x00\xED"
-     "\x85\xA3"
-     "\x00\xED\x85\xA4\x00\xED\x85\xA5\x00\xED\x85\xA6\x00\xED\x85\xA7\x00\xED"
-     "\x85\xA8"
-     "\x00\xED\x85\xA9\x00\xED\x85\xAA\x00\xED\x85\xAB\x00\xED\x85\xAC\x00\xED"
-     "\x85\xAD"
-     "\x00\xED\x85\xAE\x00\xED\x85\xAF\x00\xED\x85\xB0\x00\xED\x85\xB1\x00\xED"
-     "\x85\xB2"
-     "\x00\xED\x85\xB3\x00\xED\x85\xB4\x00\xED\x85\xB5\x00\xED\x85\xB6\x00\xED"
-     "\x85\xB7"
-     "\x00\xED\x85\xB8\x00\xED\x85\xB9\x00\xED\x85\xBA\x00\xED\x85\xBB\x00\xED"
-     "\x85\xBC"
-     "\x00\xED\x85\xBD\x00\xED\x85\xBE\x00\xED\x85\xBF\x00\xED\x86\x80\x00\xED"
-     "\x86\x81"
-     "\x00\xED\x86\x82\x00\xED\x86\x83\x00\xED\x86\x84\x00\xED\x86\x85\x00\xED"
-     "\x86\x86"
-     "\x00\xED\x86\x87\x00\xED\x86\x88\x00\xED\x86\x89\x00\xED\x86\x8A\x00\xED"
-     "\x86\x8B"
-     "\x00\xED\x86\x8C\x00\xED\x86\x8D\x00\xED\x86\x8E\x00\xED\x86\x8F\x00\xED"
-     "\x86\x90"
-     "\x00\xED\x86\x91\x00\xED\x86\x92\x00\xED\x86\x93\x00\xED\x86\x94\x00\xED"
-     "\x86\x95"
-     "\x00\xED\x86\x96\x00\xED\x86\x97\x00\xED\x86\x98\x00\xED\x86\x99\x00\xED"
-     "\x86\x9A"
-     "\x00\xED\x86\x9B\x00\xED\x86\x9C\x00\xED\x86\x9D\x00\xED\x86\x9E\x00\xED"
-     "\x86\x9F"
-     "\x00\xED\x86\xA0\x00\xED\x86\xA1\x00\xED\x86\xA2\x00\xED\x86\xA3\x00\xED"
-     "\x86\xA4"
-     "\x00\xED\x86\xA5\x00\xED\x86\xA6\x00\xED\x86\xA7\x00\xED\x86\xA8\x00\xED"
-     "\x86\xA9"
-     "\x00\xED\x86\xAA\x00\xED\x86\xAB\x00\xED\x86\xAC\x00\xED\x86\xAD\x00\xED"
-     "\x86\xAE"
-     "\x00\xED\x86\xAF\x00\xED\x86\xB0\x00\xED\x86\xB1\x00\xED\x86\xB2\x00\xED"
-     "\x86\xB3"
-     "\x00\xED\x86\xB4\x00\xED\x86\xB5\x00\xED\x86\xB6\x00\xED\x86\xB7\x00\xED"
-     "\x86\xB8"
-     "\x00\xED\x86\xB9\x00\xED\x86\xBA\x00\xED\x86\xBB\x00\xED\x86\xBC\x00\xED"
-     "\x86\xBD"
-     "\x00\xED\x86\xBE\x00\xED\x86\xBF\x00\xED\x87\x80\x00\xED\x87\x81\x00\xED"
-     "\x87\x82"
-     "\x00\xED\x87\x83\x00\xED\x87\x84\x00\xED\x87\x85\x00\xED\x87\x86\x00\xED"
-     "\x87\x87"
-     "\x00\xED\x87\x88\x00\xED\x87\x89\x00\xED\x87\x8A\x00\xED\x87\x8B\x00\xED"
-     "\x87\x8C"
-     "\x00\xED\x87\x8D\x00\xED\x87\x8E\x00\xED\x87\x8F\x00\xED\x87\x90\x00\xED"
-     "\x87\x91"
-     "\x00\xED\x87\x92\x00\xED\x87\x93\x00\xED\x87\x94\x00\xED\x87\x95\x00\xED"
-     "\x87\x96"
-     "\x00\xED\x87\x97\x00\xED\x87\x98\x00\xED\x87\x99\x00\xED\x87\x9A\x00\xED"
-     "\x87\x9B"
-     "\x00\xED\x87\x9C\x00\xED\x87\x9D\x00\xED\x87\x9E\x00\xED\x87\x9F\x00\xED"
-     "\x87\xA0"
-     "\x00\xED\x87\xA1\x00\xED\x87\xA2\x00\xED\x87\xA3\x00\xED\x87\xA4\x00\xED"
-     "\x87\xA5"
-     "\x00\xED\x87\xA6\x00\xED\x87\xA7\x00\xED\x87\xA8\x00\xED\x87\xA9\x00\xED"
-     "\x87\xAA"
-     "\x00\xED\x87\xAB\x00\xED\x87\xAC\x00\xED\x87\xAD\x00\xED\x87\xAE\x00\xED"
-     "\x87\xAF"
-     "\x00\xED\x87\xB0\x00\xED\x87\xB1\x00\xED\x87\xB2\x00\xED\x87\xB3\x00\xED"
-     "\x87\xB4"
-     "\x00\xED\x87\xB5\x00\xED\x87\xB6\x00\xED\x87\xB7\x00\xED\x87\xB8\x00\xED"
-     "\x87\xB9"
-     "\x00\xED\x87\xBA\x00\xED\x87\xBB\x00\xED\x87\xBC\x00\xED\x87\xBD\x00\xED"
-     "\x87\xBE"
-     "\x00\xED\x87\xBF\x00\xED\x88\x80\x00\xED\x88\x81\x00\xED\x88\x82\x00\xED"
-     "\x88\x83"
-     "\x00\xED\x88\x84\x00\xED\x88\x85\x00\xED\x88\x86\x00\xED\x88\x87\x00\xED"
-     "\x88\x88"
-     "\x00\xED\x88\x89\x00\xED\x88\x8A\x00\xED\x88\x8B\x00\xED\x88\x8C\x00\xED"
-     "\x88\x8D"
-     "\x00\xED\x88\x8E\x00\xED\x88\x8F\x00\xED\x88\x90\x00\xED\x88\x91\x00\xED"
-     "\x88\x92"
-     "\x00\xED\x88\x93\x00\xED\x88\x94\x00\xED\x88\x95\x00\xED\x88\x96\x00\xED"
-     "\x88\x97"
-     "\x00\xED\x88\x98\x00\xED\x88\x99\x00\xED\x88\x9A\x00\xED\x88\x9B\x00\xED"
-     "\x88\x9C"
-     "\x00\xED\x88\x9D\x00\xED\x88\x9E\x00\xED\x88\x9F\x00\xED\x88\xA0\x00\xED"
-     "\x88\xA1"
-     "\x00\xED\x88\xA2\x00\xED\x88\xA3\x00\xED\x88\xA4\x00\xED\x88\xA5\x00\xED"
-     "\x88\xA6"
-     "\x00\xED\x88\xA7\x00\xED\x88\xA8\x00\xED\x88\xA9\x00\xED\x88\xAA\x00\xED"
-     "\x88\xAB"
-     "\x00\xED\x88\xAC\x00\xED\x88\xAD\x00\xED\x88\xAE\x00\xED\x88\xAF\x00\xED"
-     "\x88\xB0"
-     "\x00\xED\x88\xB1\x00\xED\x88\xB2\x00\xED\x88\xB3\x00\xED\x88\xB4\x00\xED"
-     "\x88\xB5"
-     "\x00\xED\x88\xB6\x00\xED\x88\xB7\x00\xED\x88\xB8\x00\xED\x88\xB9\x00\xED"
-     "\x88\xBA"
-     "\x00\xED\x88\xBB\x00\xED\x88\xBC\x00\xED\x88\xBD\x00\xED\x88\xBE\x00\xED"
-     "\x88\xBF"
-     "\x00\xED\x89\x80\x00\xED\x89\x81\x00\xED\x89\x82\x00\xED\x89\x83\x00\xED"
-     "\x89\x84"
-     "\x00\xED\x89\x85\x00\xED\x89\x86\x00\xED\x89\x87\x00\xED\x89\x88\x00\xED"
-     "\x89\x89"
-     "\x00\xED\x89\x8A\x00\xED\x89\x8B\x00\xED\x89\x8C\x00\xED\x89\x8D\x00\xED"
-     "\x89\x8E"
-     "\x00\xED\x89\x8F\x00\xED\x89\x90\x00\xED\x89\x91\x00\xED\x89\x92\x00\xED"
-     "\x89\x93"
-     "\x00\xED\x89\x94\x00\xED\x89\x95\x00\xED\x89\x96\x00\xED\x89\x97\x00\xED"
-     "\x89\x98"
-     "\x00\xED\x89\x99\x00\xED\x89\x9A\x00\xED\x89\x9B\x00\xED\x89\x9C\x00\xED"
-     "\x89\x9D"
-     "\x00\xED\x89\x9E\x00\xED\x89\x9F\x00\xED\x89\xA0\x00\xED\x89\xA1\x00\xED"
-     "\x89\xA2"
-     "\x00\xED\x89\xA3\x00\xED\x89\xA4\x00\xED\x89\xA5\x00\xED\x89\xA6\x00\xED"
-     "\x89\xA7"
-     "\x00\xED\x89\xA8\x00\xED\x89\xA9\x00\xED\x89\xAA\x00\xED\x89\xAB\x00\xED"
-     "\x89\xAC"
-     "\x00\xED\x89\xAD\x00\xED\x89\xAE\x00\xED\x89\xAF\x00\xED\x89\xB0\x00\xED"
-     "\x89\xB1"
-     "\x00\xED\x89\xB2\x00\xED\x89\xB3\x00\xED\x89\xB4\x00\xED\x89\xB5\x00\xED"
-     "\x89\xB6"
-     "\x00\xED\x89\xB7\x00\xED\x89\xB8\x00\xED\x89\xB9\x00\xED\x89\xBA\x00\xED"
-     "\x89\xBB"
-     "\x00\xED\x89\xBC\x00\xED\x89\xBD\x00\xED\x89\xBE\x00\xED\x89\xBF\x00\xED"
-     "\x8A\x80"
-     "\x00\xED\x8A\x81\x00\xED\x8A\x82\x00\xED\x8A\x83\x00\xED\x8A\x84\x00\xED"
-     "\x8A\x85"
-     "\x00\xED\x8A\x86\x00\xED\x8A\x87\x00\xED\x8A\x88\x00\xED\x8A\x89\x00\xED"
-     "\x8A\x8A"
-     "\x00\xED\x8A\x8B\x00\xED\x8A\x8C\x00\xED\x8A\x8D\x00\xED\x8A\x8E\x00\xED"
-     "\x8A\x8F"
-     "\x00\xED\x8A\x90\x00\xED\x8A\x91\x00\xED\x8A\x92\x00\xED\x8A\x93\x00\xED"
-     "\x8A\x94"
-     "\x00\xED\x8A\x95\x00\xED\x8A\x96\x00\xED\x8A\x97\x00\xED\x8A\x98\x00\xED"
-     "\x8A\x99"
-     "\x00\xED\x8A\x9A\x00\xED\x8A\x9B\x00\xED\x8A\x9C\x00\xED\x8A\x9D\x00\xED"
-     "\x8A\x9E"
-     "\x00\xED\x8A\x9F\x00\xED\x8A\xA0\x00\xED\x8A\xA1\x00\xED\x8A\xA2\x00\xED"
-     "\x8A\xA3"
-     "\x00\xED\x8A\xA4\x00\xED\x8A\xA5\x00\xED\x8A\xA6\x00\xED\x8A\xA7\x00\xED"
-     "\x8A\xA8"
-     "\x00\xED\x8A\xA9\x00\xED\x8A\xAA\x00\xED\x8A\xAB\x00\xED\x8A\xAC\x00\xED"
-     "\x8A\xAD"
-     "\x00\xED\x8A\xAE\x00\xED\x8A\xAF\x00\xED\x8A\xB0\x00\xED\x8A\xB1\x00\xED"
-     "\x8A\xB2"
-     "\x00\xED\x8A\xB3\x00\xED\x8A\xB4\x00\xED\x8A\xB5\x00\xED\x8A\xB6\x00\xED"
-     "\x8A\xB7"
-     "\x00\xED\x8A\xB8\x00\xED\x8A\xB9\x00\xED\x8A\xBA\x00\xED\x8A\xBB\x00\xED"
-     "\x8A\xBC"
-     "\x00\xED\x8A\xBD\x00\xED\x8A\xBE\x00\xED\x8A\xBF\x00\xED\x8B\x80\x00\xED"
-     "\x8B\x81"
-     "\x00\xED\x8B\x82\x00\xED\x8B\x83\x00\xED\x8B\x84\x00\xED\x8B\x85\x00\xED"
-     "\x8B\x86"
-     "\x00\xED\x8B\x87\x00\xED\x8B\x88\x00\xED\x8B\x89\x00\xED\x8B\x8A\x00\xED"
-     "\x8B\x8B"
-     "\x00\xED\x8B\x8C\x00\xED\x8B\x8D\x00\xED\x8B\x8E\x00\xED\x8B\x8F\x00\xED"
-     "\x8B\x90"
-     "\x00\xED\x8B\x91\x00\xED\x8B\x92\x00\xED\x8B\x93\x00\xED\x8B\x94\x00\xED"
-     "\x8B\x95"
-     "\x00\xED\x8B\x96\x00\xED\x8B\x97\x00\xED\x8B\x98\x00\xED\x8B\x99\x00\xED"
-     "\x8B\x9A"
-     "\x00\xED\x8B\x9B\x00\xED\x8B\x9C\x00\xED\x8B\x9D\x00\xED\x8B\x9E\x00\xED"
-     "\x8B\x9F"
-     "\x00\xED\x8B\xA0\x00\xED\x8B\xA1\x00\xED\x8B\xA2\x00\xED\x8B\xA3\x00\xED"
-     "\x8B\xA4"
-     "\x00\xED\x8B\xA5\x00\xED\x8B\xA6\x00\xED\x8B\xA7\x00\xED\x8B\xA8\x00\xED"
-     "\x8B\xA9"
-     "\x00\xED\x8B\xAA\x00\xED\x8B\xAB\x00\xED\x8B\xAC\x00\xED\x8B\xAD\x00\xED"
-     "\x8B\xAE"
-     "\x00\xED\x8B\xAF\x00\xED\x8B\xB0\x00\xED\x8B\xB1\x00\xED\x8B\xB2\x00\xED"
-     "\x8B\xB3"
-     "\x00\xED\x8B\xB4\x00\xED\x8B\xB5\x00\xED\x8B\xB6\x00\xED\x8B\xB7\x00\xED"
-     "\x8B\xB8"
-     "\x00\xED\x8B\xB9\x00\xED\x8B\xBA\x00\xED\x8B\xBB\x00\xED\x8B\xBC\x00\xED"
-     "\x8B\xBD"
-     "\x00\xED\x8B\xBE\x00\xED\x8B\xBF\x00\xED\x8C\x80\x00\xED\x8C\x81\x00\xED"
-     "\x8C\x82"
-     "\x00\xED\x8C\x83\x00\xED\x8C\x84\x00\xED\x8C\x85\x00\xED\x8C\x86\x00\xED"
-     "\x8C\x87"
-     "\x00\xED\x8C\x88\x00\xED\x8C\x89\x00\xED\x8C\x8A\x00\xED\x8C\x8B\x00\xED"
-     "\x8C\x8C"
-     "\x00\xED\x8C\x8D\x00\xED\x8C\x8E\x00\xED\x8C\x8F\x00\xED\x8C\x90\x00\xED"
-     "\x8C\x91"
-     "\x00\xED\x8C\x92\x00\xED\x8C\x93\x00\xED\x8C\x94\x00\xED\x8C\x95\x00\xED"
-     "\x8C\x96"
-     "\x00\xED\x8C\x97\x00\xED\x8C\x98\x00\xED\x8C\x99\x00\xED\x8C\x9A\x00\xED"
-     "\x8C\x9B"
-     "\x00\xED\x8C\x9C\x00\xED\x8C\x9D\x00\xED\x8C\x9E\x00\xED\x8C\x9F\x00\xED"
-     "\x8C\xA0"
-     "\x00\xED\x8C\xA1\x00\xED\x8C\xA2\x00\xED\x8C\xA3\x00\xED\x8C\xA4\x00\xED"
-     "\x8C\xA5"
-     "\x00\xED\x8C\xA6\x00\xED\x8C\xA7\x00\xED\x8C\xA8\x00\xED\x8C\xA9\x00\xED"
-     "\x8C\xAA"
-     "\x00\xED\x8C\xAB\x00\xED\x8C\xAC\x00\xED\x8C\xAD\x00\xED\x8C\xAE\x00\xED"
-     "\x8C\xAF"
-     "\x00\xED\x8C\xB0\x00\xED\x8C\xB1\x00\xED\x8C\xB2\x00\xED\x8C\xB3\x00\xED"
-     "\x8C\xB4"
-     "\x00\xED\x8C\xB5\x00\xED\x8C\xB6\x00\xED\x8C\xB7\x00\xED\x8C\xB8\x00\xED"
-     "\x8C\xB9"
-     "\x00\xED\x8C\xBA\x00\xED\x8C\xBB\x00\xED\x8C\xBC\x00\xED\x8C\xBD\x00\xED"
-     "\x8C\xBE"
-     "\x00\xED\x8C\xBF\x00\xED\x8D\x80\x00\xED\x8D\x81\x00\xED\x8D\x82\x00\xED"
-     "\x8D\x83"
-     "\x00\xED\x8D\x84\x00\xED\x8D\x85\x00\xED\x8D\x86\x00\xED\x8D\x87\x00\xED"
-     "\x8D\x88"
-     "\x00\xED\x8D\x89\x00\xED\x8D\x8A\x00\xED\x8D\x8B\x00\xED\x8D\x8C\x00\xED"
-     "\x8D\x8D"
-     "\x00\xED\x8D\x8E\x00\xED\x8D\x8F\x00\xED\x8D\x90\x00\xED\x8D\x91\x00\xED"
-     "\x8D\x92"
-     "\x00\xED\x8D\x93\x00\xED\x8D\x94\x00\xED\x8D\x95\x00\xED\x8D\x96\x00\xED"
-     "\x8D\x97"
-     "\x00\xED\x8D\x98\x00\xED\x8D\x99\x00\xED\x8D\x9A\x00\xED\x8D\x9B\x00\xED"
-     "\x8D\x9C"
-     "\x00\xED\x8D\x9D\x00\xED\x8D\x9E\x00\xED\x8D\x9F\x00\xED\x8D\xA0\x00\xED"
-     "\x8D\xA1"
-     "\x00\xED\x8D\xA2\x00\xED\x8D\xA3\x00\xED\x8D\xA4\x00\xED\x8D\xA5\x00\xED"
-     "\x8D\xA6"
-     "\x00\xED\x8D\xA7\x00\xED\x8D\xA8\x00\xED\x8D\xA9\x00\xED\x8D\xAA\x00\xED"
-     "\x8D\xAB"
-     "\x00\xED\x8D\xAC\x00\xED\x8D\xAD\x00\xED\x8D\xAE\x00\xED\x8D\xAF\x00\xED"
-     "\x8D\xB0"
-     "\x00\xED\x8D\xB1\x00\xED\x8D\xB2\x00\xED\x8D\xB3\x00\xED\x8D\xB4\x00\xED"
-     "\x8D\xB5"
-     "\x00\xED\x8D\xB6\x00\xED\x8D\xB7\x00\xED\x8D\xB8\x00\xED\x8D\xB9\x00\xED"
-     "\x8D\xBA"
-     "\x00\xED\x8D\xBB\x00\xED\x8D\xBC\x00\xED\x8D\xBD\x00\xED\x8D\xBE\x00\xED"
-     "\x8D\xBF"
-     "\x00\xED\x8E\x80\x00\xED\x8E\x81\x00\xED\x8E\x82\x00\xED\x8E\x83\x00\xED"
-     "\x8E\x84"
-     "\x00\xED\x8E\x85\x00\xED\x8E\x86\x00\xED\x8E\x87\x00\xED\x8E\x88\x00\xED"
-     "\x8E\x89"
-     "\x00\xED\x8E\x8A\x00\xED\x8E\x8B\x00\xED\x8E\x8C\x00\xED\x8E\x8D\x00\xED"
-     "\x8E\x8E"
-     "\x00\xED\x8E\x8F\x00\xED\x8E\x90\x00\xED\x8E\x91\x00\xED\x8E\x92\x00\xED"
-     "\x8E\x93"
-     "\x00\xED\x8E\x94\x00\xED\x8E\x95\x00\xED\x8E\x96\x00\xED\x8E\x97\x00\xED"
-     "\x8E\x98"
-     "\x00\xED\x8E\x99\x00\xED\x8E\x9A\x00\xED\x8E\x9B\x00\xED\x8E\x9C\x00\xED"
-     "\x8E\x9D"
-     "\x00\xED\x8E\x9E\x00\xED\x8E\x9F\x00\xED\x8E\xA0\x00\xED\x8E\xA1\x00\xED"
-     "\x8E\xA2"
-     "\x00\xED\x8E\xA3\x00\xED\x8E\xA4\x00\xED\x8E\xA5\x00\xED\x8E\xA6\x00\xED"
-     "\x8E\xA7"
-     "\x00\xED\x8E\xA8\x00\xED\x8E\xA9\x00\xED\x8E\xAA\x00\xED\x8E\xAB\x00\xED"
-     "\x8E\xAC"
-     "\x00\xED\x8E\xAD\x00\xED\x8E\xAE\x00\xED\x8E\xAF\x00\xED\x8E\xB0\x00\xED"
-     "\x8E\xB1"
-     "\x00\xED\x8E\xB2\x00\xED\x8E\xB3\x00\xED\x8E\xB4\x00\xED\x8E\xB5\x00\xED"
-     "\x8E\xB6"
-     "\x00\xED\x8E\xB7\x00\xED\x8E\xB8\x00\xED\x8E\xB9\x00\xED\x8E\xBA\x00\xED"
-     "\x8E\xBB"
-     "\x00\xED\x8E\xBC\x00\xED\x8E\xBD\x00\xED\x8E\xBE\x00\xED\x8E\xBF\x00\xED"
-     "\x8F\x80"
-     "\x00\xED\x8F\x81\x00\xED\x8F\x82\x00\xED\x8F\x83\x00\xED\x8F\x84\x00\xED"
-     "\x8F\x85"
-     "\x00\xED\x8F\x86\x00\xED\x8F\x87\x00\xED\x8F\x88\x00\xED\x8F\x89\x00\xED"
-     "\x8F\x8A"
-     "\x00\xED\x8F\x8B\x00\xED\x8F\x8C\x00\xED\x8F\x8D\x00\xED\x8F\x8E\x00\xED"
-     "\x8F\x8F"
-     "\x00\xED\x8F\x90\x00\xED\x8F\x91\x00\xED\x8F\x92\x00\xED\x8F\x93\x00\xED"
-     "\x8F\x94"
-     "\x00\xED\x8F\x95\x00\xED\x8F\x96\x00\xED\x8F\x97\x00\xED\x8F\x98\x00\xED"
-     "\x8F\x99"
-     "\x00\xED\x8F\x9A\x00\xED\x8F\x9B\x00\xED\x8F\x9C\x00\xED\x8F\x9D\x00\xED"
-     "\x8F\x9E"
-     "\x00\xED\x8F\x9F\x00\xED\x8F\xA0\x00\xED\x8F\xA1\x00\xED\x8F\xA2\x00\xED"
-     "\x8F\xA3"
-     "\x00\xED\x8F\xA4\x00\xED\x8F\xA5\x00\xED\x8F\xA6\x00\xED\x8F\xA7\x00\xED"
-     "\x8F\xA8"
-     "\x00\xED\x8F\xA9\x00\xED\x8F\xAA\x00\xED\x8F\xAB\x00\xED\x8F\xAC\x00\xED"
-     "\x8F\xAD"
-     "\x00\xED\x8F\xAE\x00\xED\x8F\xAF\x00\xED\x8F\xB0\x00\xED\x8F\xB1\x00\xED"
-     "\x8F\xB2"
-     "\x00\xED\x8F\xB3\x00\xED\x8F\xB4\x00\xED\x8F\xB5\x00\xED\x8F\xB6\x00\xED"
-     "\x8F\xB7"
-     "\x00\xED\x8F\xB8\x00\xED\x8F\xB9\x00\xED\x8F\xBA\x00\xED\x8F\xBB\x00\xED"
-     "\x8F\xBC"
-     "\x00\xED\x8F\xBD\x00\xED\x8F\xBE\x00\xED\x8F\xBF\x00\xED\x90\x80\x00\xED"
-     "\x90\x81"
-     "\x00\xED\x90\x82\x00\xED\x90\x83\x00\xED\x90\x84\x00\xED\x90\x85\x00\xED"
-     "\x90\x86"
-     "\x00\xED\x90\x87\x00\xED\x90\x88\x00\xED\x90\x89\x00\xED\x90\x8A\x00\xED"
-     "\x90\x8B"
-     "\x00\xED\x90\x8C\x00\xED\x90\x8D\x00\xED\x90\x8E\x00\xED\x90\x8F\x00\xED"
-     "\x90\x90"
-     "\x00\xED\x90\x91\x00\xED\x90\x92\x00\xED\x90\x93\x00\xED\x90\x94\x00\xED"
-     "\x90\x95"
-     "\x00\xED\x90\x96\x00\xED\x90\x97\x00\xED\x90\x98\x00\xED\x90\x99\x00\xED"
-     "\x90\x9A"
-     "\x00\xED\x90\x9B\x00\xED\x90\x9C\x00\xED\x90\x9D\x00\xED\x90\x9E\x00\xED"
-     "\x90\x9F"
-     "\x00\xED\x90\xA0\x00\xED\x90\xA1\x00\xED\x90\xA2\x00\xED\x90\xA3\x00\xED"
-     "\x90\xA4"
-     "\x00\xED\x90\xA5\x00\xED\x90\xA6\x00\xED\x90\xA7\x00\xED\x90\xA8\x00\xED"
-     "\x90\xA9"
-     "\x00\xED\x90\xAA\x00\xED\x90\xAB\x00\xED\x90\xAC\x00\xED\x90\xAD\x00\xED"
-     "\x90\xAE"
-     "\x00\xED\x90\xAF\x00\xED\x90\xB0\x00\xED\x90\xB1\x00\xED\x90\xB2\x00\xED"
-     "\x90\xB3"
-     "\x00\xED\x90\xB4\x00\xED\x90\xB5\x00\xED\x90\xB6\x00\xED\x90\xB7\x00\xED"
-     "\x90\xB8"
-     "\x00\xED\x90\xB9\x00\xED\x90\xBA\x00\xED\x90\xBB\x00\xED\x90\xBC\x00\xED"
-     "\x90\xBD"
-     "\x00\xED\x90\xBE\x00\xED\x90\xBF\x00\xED\x91\x80\x00\xED\x91\x81\x00\xED"
-     "\x91\x82"
-     "\x00\xED\x91\x83\x00\xED\x91\x84\x00\xED\x91\x85\x00\xED\x91\x86\x00\xED"
-     "\x91\x87"
-     "\x00\xED\x91\x88\x00\xED\x91\x89\x00\xED\x91\x8A\x00\xED\x91\x8B\x00\xED"
-     "\x91\x8C"
-     "\x00\xED\x91\x8D\x00\xED\x91\x8E\x00\xED\x91\x8F\x00\xED\x91\x90\x00\xED"
-     "\x91\x91"
-     "\x00\xED\x91\x92\x00\xED\x91\x93\x00\xED\x91\x94\x00\xED\x91\x95\x00\xED"
-     "\x91\x96"
-     "\x00\xED\x91\x97\x00\xED\x91\x98\x00\xED\x91\x99\x00\xED\x91\x9A\x00\xED"
-     "\x91\x9B"
-     "\x00\xED\x91\x9C\x00\xED\x91\x9D\x00\xED\x91\x9E\x00\xED\x91\x9F\x00\xED"
-     "\x91\xA0"
-     "\x00\xED\x91\xA1\x00\xED\x91\xA2\x00\xED\x91\xA3\x00\xED\x91\xA4\x00\xED"
-     "\x91\xA5"
-     "\x00\xED\x91\xA6\x00\xED\x91\xA7\x00\xED\x91\xA8\x00\xED\x91\xA9\x00\xED"
-     "\x91\xAA"
-     "\x00\xED\x91\xAB\x00\xED\x91\xAC\x00\xED\x91\xAD\x00\xED\x91\xAE\x00\xED"
-     "\x91\xAF"
-     "\x00\xED\x91\xB0\x00\xED\x91\xB1\x00\xED\x91\xB2\x00\xED\x91\xB3\x00\xED"
-     "\x91\xB4"
-     "\x00\xED\x91\xB5\x00\xED\x91\xB6\x00\xED\x91\xB7\x00\xED\x91\xB8\x00\xED"
-     "\x91\xB9"
-     "\x00\xED\x91\xBA\x00\xED\x91\xBB\x00\xED\x91\xBC\x00\xED\x91\xBD\x00\xED"
-     "\x91\xBE"
-     "\x00\xED\x91\xBF\x00\xED\x92\x80\x00\xED\x92\x81\x00\xED\x92\x82\x00\xED"
-     "\x92\x83"
-     "\x00\xED\x92\x84\x00\xED\x92\x85\x00\xED\x92\x86\x00\xED\x92\x87\x00\xED"
-     "\x92\x88"
-     "\x00\xED\x92\x89\x00\xED\x92\x8A\x00\xED\x92\x8B\x00\xED\x92\x8C\x00\xED"
-     "\x92\x8D"
-     "\x00\xED\x92\x8E\x00\xED\x92\x8F\x00\xED\x92\x90\x00\xED\x92\x91\x00\xED"
-     "\x92\x92"
-     "\x00\xED\x92\x93\x00\xED\x92\x94\x00\xED\x92\x95\x00\xED\x92\x96\x00\xED"
-     "\x92\x97"
-     "\x00\xED\x92\x98\x00\xED\x92\x99\x00\xED\x92\x9A\x00\xED\x92\x9B\x00\xED"
-     "\x92\x9C"
-     "\x00\xED\x92\x9D\x00\xED\x92\x9E\x00\xED\x92\x9F\x00\xED\x92\xA0\x00\xED"
-     "\x92\xA1"
-     "\x00\xED\x92\xA2\x00\xED\x92\xA3\x00\xED\x92\xA4\x00\xED\x92\xA5\x00\xED"
-     "\x92\xA6"
-     "\x00\xED\x92\xA7\x00\xED\x92\xA8\x00\xED\x92\xA9\x00\xED\x92\xAA\x00\xED"
-     "\x92\xAB"
-     "\x00\xED\x92\xAC\x00\xED\x92\xAD\x00\xED\x92\xAE\x00\xED\x92\xAF\x00\xED"
-     "\x92\xB0"
-     "\x00\xED\x92\xB1\x00\xED\x92\xB2\x00\xED\x92\xB3\x00\xED\x92\xB4\x00\xED"
-     "\x92\xB5"
-     "\x00\xED\x92\xB6\x00\xED\x92\xB7\x00\xED\x92\xB8\x00\xED\x92\xB9\x00\xED"
-     "\x92\xBA"
-     "\x00\xED\x92\xBB\x00\xED\x92\xBC\x00\xED\x92\xBD\x00\xED\x92\xBE\x00\xED"
-     "\x92\xBF"
-     "\x00\xED\x93\x80\x00\xED\x93\x81\x00\xED\x93\x82\x00\xED\x93\x83\x00\xED"
-     "\x93\x84"
-     "\x00\xED\x93\x85\x00\xED\x93\x86\x00\xED\x93\x87\x00\xED\x93\x88\x00\xED"
-     "\x93\x89"
-     "\x00\xED\x93\x8A\x00\xED\x93\x8B\x00\xED\x93\x8C\x00\xED\x93\x8D\x00\xED"
-     "\x93\x8E"
-     "\x00\xED\x93\x8F\x00\xED\x93\x90\x00\xED\x93\x91\x00\xED\x93\x92\x00\xED"
-     "\x93\x93"
-     "\x00\xED\x93\x94\x00\xED\x93\x95\x00\xED\x93\x96\x00\xED\x93\x97\x00\xED"
-     "\x93\x98"
-     "\x00\xED\x93\x99\x00\xED\x93\x9A\x00\xED\x93\x9B\x00\xED\x93\x9C\x00\xED"
-     "\x93\x9D"
-     "\x00\xED\x93\x9E\x00\xED\x93\x9F\x00\xED\x93\xA0\x00\xED\x93\xA1\x00\xED"
-     "\x93\xA2"
-     "\x00\xED\x93\xA3\x00\xED\x93\xA4\x00\xED\x93\xA5\x00\xED\x93\xA6\x00\xED"
-     "\x93\xA7"
-     "\x00\xED\x93\xA8\x00\xED\x93\xA9\x00\xED\x93\xAA\x00\xED\x93\xAB\x00\xED"
-     "\x93\xAC"
-     "\x00\xED\x93\xAD\x00\xED\x93\xAE\x00\xED\x93\xAF\x00\xED\x93\xB0\x00\xED"
-     "\x93\xB1"
-     "\x00\xED\x93\xB2\x00\xED\x93\xB3\x00\xED\x93\xB4\x00\xED\x93\xB5\x00\xED"
-     "\x93\xB6"
-     "\x00\xED\x93\xB7\x00\xED\x93\xB8\x00\xED\x93\xB9\x00\xED\x93\xBA\x00\xED"
-     "\x93\xBB"
-     "\x00\xED\x93\xBC\x00\xED\x93\xBD\x00\xED\x93\xBE\x00\xED\x93\xBF\x00\xED"
-     "\x94\x80"
-     "\x00\xED\x94\x81\x00\xED\x94\x82\x00\xED\x94\x83\x00\xED\x94\x84\x00\xED"
-     "\x94\x85"
-     "\x00\xED\x94\x86\x00\xED\x94\x87\x00\xED\x94\x88\x00\xED\x94\x89\x00\xED"
-     "\x94\x8A"
-     "\x00\xED\x94\x8B\x00\xED\x94\x8C\x00\xED\x94\x8D\x00\xED\x94\x8E\x00\xED"
-     "\x94\x8F"
-     "\x00\xED\x94\x90\x00\xED\x94\x91\x00\xED\x94\x92\x00\xED\x94\x93\x00\xED"
-     "\x94\x94"
-     "\x00\xED\x94\x95\x00\xED\x94\x96\x00\xED\x94\x97\x00\xED\x94\x98\x00\xED"
-     "\x94\x99"
-     "\x00\xED\x94\x9A\x00\xED\x94\x9B\x00\xED\x94\x9C\x00\xED\x94\x9D\x00\xED"
-     "\x94\x9E"
-     "\x00\xED\x94\x9F\x00\xED\x94\xA0\x00\xED\x94\xA1\x00\xED\x94\xA2\x00\xED"
-     "\x94\xA3"
-     "\x00\xED\x94\xA4\x00\xED\x94\xA5\x00\xED\x94\xA6\x00\xED\x94\xA7\x00\xED"
-     "\x94\xA8"
-     "\x00\xED\x94\xA9\x00\xED\x94\xAA\x00\xED\x94\xAB\x00\xED\x94\xAC\x00\xED"
-     "\x94\xAD"
-     "\x00\xED\x94\xAE\x00\xED\x94\xAF\x00\xED\x94\xB0\x00\xED\x94\xB1\x00\xED"
-     "\x94\xB2"
-     "\x00\xED\x94\xB3\x00\xED\x94\xB4\x00\xED\x94\xB5\x00\xED\x94\xB6\x00\xED"
-     "\x94\xB7"
-     "\x00\xED\x94\xB8\x00\xED\x94\xB9\x00\xED\x94\xBA\x00\xED\x94\xBB\x00\xED"
-     "\x94\xBC"
-     "\x00\xED\x94\xBD\x00\xED\x94\xBE\x00\xED\x94\xBF\x00\xED\x95\x80\x00\xED"
-     "\x95\x81"
-     "\x00\xED\x95\x82\x00\xED\x95\x83\x00\xED\x95\x84\x00\xED\x95\x85\x00\xED"
-     "\x95\x86"
-     "\x00\xED\x95\x87\x00\xED\x95\x88\x00\xED\x95\x89\x00\xED\x95\x8A\x00\xED"
-     "\x95\x8B"
-     "\x00\xED\x95\x8C\x00\xED\x95\x8D\x00\xED\x95\x8E\x00\xED\x95\x8F\x00\xED"
-     "\x95\x90"
-     "\x00\xED\x95\x91\x00\xED\x95\x92\x00\xED\x95\x93\x00\xED\x95\x94\x00\xED"
-     "\x95\x95"
-     "\x00\xED\x95\x96\x00\xED\x95\x97\x00\xED\x95\x98\x00\xED\x95\x99\x00\xED"
-     "\x95\x9A"
-     "\x00\xED\x95\x9B\x00\xED\x95\x9C\x00\xED\x95\x9D\x00\xED\x95\x9E\x00\xED"
-     "\x95\x9F"
-     "\x00\xED\x95\xA0\x00\xED\x95\xA1\x00\xED\x95\xA2\x00\xED\x95\xA3\x00\xED"
-     "\x95\xA4"
-     "\x00\xED\x95\xA5\x00\xED\x95\xA6\x00\xED\x95\xA7\x00\xED\x95\xA8\x00\xED"
-     "\x95\xA9"
-     "\x00\xED\x95\xAA\x00\xED\x95\xAB\x00\xED\x95\xAC\x00\xED\x95\xAD\x00\xED"
-     "\x95\xAE"
-     "\x00\xED\x95\xAF\x00\xED\x95\xB0\x00\xED\x95\xB1\x00\xED\x95\xB2\x00\xED"
-     "\x95\xB3"
-     "\x00\xED\x95\xB4\x00\xED\x95\xB5\x00\xED\x95\xB6\x00\xED\x95\xB7\x00\xED"
-     "\x95\xB8"
-     "\x00\xED\x95\xB9\x00\xED\x95\xBA\x00\xED\x95\xBB\x00\xED\x95\xBC\x00\xED"
-     "\x95\xBD"
-     "\x00\xED\x95\xBE\x00\xED\x95\xBF\x00\xED\x96\x80\x00\xED\x96\x81\x00\xED"
-     "\x96\x82"
-     "\x00\xED\x96\x83\x00\xED\x96\x84\x00\xED\x96\x85\x00\xED\x96\x86\x00\xED"
-     "\x96\x87"
-     "\x00\xED\x96\x88\x00\xED\x96\x89\x00\xED\x96\x8A\x00\xED\x96\x8B\x00\xED"
-     "\x96\x8C"
-     "\x00\xED\x96\x8D\x00\xED\x96\x8E\x00\xED\x96\x8F\x00\xED\x96\x90\x00\xED"
-     "\x96\x91"
-     "\x00\xED\x96\x92\x00\xED\x96\x93\x00\xED\x96\x94\x00\xED\x96\x95\x00\xED"
-     "\x96\x96"
-     "\x00\xED\x96\x97\x00\xED\x96\x98\x00\xED\x96\x99\x00\xED\x96\x9A\x00\xED"
-     "\x96\x9B"
-     "\x00\xED\x96\x9C\x00\xED\x96\x9D\x00\xED\x96\x9E\x00\xED\x96\x9F\x00\xED"
-     "\x96\xA0"
-     "\x00\xED\x96\xA1\x00\xED\x96\xA2\x00\xED\x96\xA3\x00\xED\x96\xA4\x00\xED"
-     "\x96\xA5"
-     "\x00\xED\x96\xA6\x00\xED\x96\xA7\x00\xED\x96\xA8\x00\xED\x96\xA9\x00\xED"
-     "\x96\xAA"
-     "\x00\xED\x96\xAB\x00\xED\x96\xAC\x00\xED\x96\xAD\x00\xED\x96\xAE\x00\xED"
-     "\x96\xAF"
-     "\x00\xED\x96\xB0\x00\xED\x96\xB1\x00\xED\x96\xB2\x00\xED\x96\xB3\x00\xED"
-     "\x96\xB4"
-     "\x00\xED\x96\xB5\x00\xED\x96\xB6\x00\xED\x96\xB7\x00\xED\x96\xB8\x00\xED"
-     "\x96\xB9"
-     "\x00\xED\x96\xBA\x00\xED\x96\xBB\x00\xED\x96\xBC\x00\xED\x96\xBD\x00\xED"
-     "\x96\xBE"
-     "\x00\xED\x96\xBF\x00\xED\x97\x80\x00\xED\x97\x81\x00\xED\x97\x82\x00\xED"
-     "\x97\x83"
-     "\x00\xED\x97\x84\x00\xED\x97\x85\x00\xED\x97\x86\x00\xED\x97\x87\x00\xED"
-     "\x97\x88"
-     "\x00\xED\x97\x89\x00\xED\x97\x8A\x00\xED\x97\x8B\x00\xED\x97\x8C\x00\xED"
-     "\x97\x8D"
-     "\x00\xED\x97\x8E\x00\xED\x97\x8F\x00\xED\x97\x90\x00\xED\x97\x91\x00\xED"
-     "\x97\x92"
-     "\x00\xED\x97\x93\x00\xED\x97\x94\x00\xED\x97\x95\x00\xED\x97\x96\x00\xED"
-     "\x97\x97"
-     "\x00\xED\x97\x98\x00\xED\x97\x99\x00\xED\x97\x9A\x00\xED\x97\x9B\x00\xED"
-     "\x97\x9C"
-     "\x00\xED\x97\x9D\x00\xED\x97\x9E\x00\xED\x97\x9F\x00\xED\x97\xA0\x00\xED"
-     "\x97\xA1"
-     "\x00\xED\x97\xA2\x00\xED\x97\xA3\x00\xED\x97\xA4\x00\xED\x97\xA5\x00\xED"
-     "\x97\xA6"
-     "\x00\xED\x97\xA7\x00\xED\x97\xA8\x00\xED\x97\xA9\x00\xED\x97\xAA\x00\xED"
-     "\x97\xAB"
-     "\x00\xED\x97\xAC\x00\xED\x97\xAD\x00\xED\x97\xAE\x00\xED\x97\xAF\x00\xED"
-     "\x97\xB0"
-     "\x00\xED\x97\xB1\x00\xED\x97\xB2\x00\xED\x97\xB3\x00\xED\x97\xB4\x00\xED"
-     "\x97\xB5"
-     "\x00\xED\x97\xB6\x00\xED\x97\xB7\x00\xED\x97\xB8\x00\xED\x97\xB9\x00\xED"
-     "\x97\xBA"
-     "\x00\xED\x97\xBB\x00\xED\x97\xBC\x00\xED\x97\xBD\x00\xED\x97\xBE\x00\xED"
-     "\x97\xBF"
-     "\x00\xED\x98\x80\x00\xED\x98\x81\x00\xED\x98\x82\x00\xED\x98\x83\x00\xED"
-     "\x98\x84"
-     "\x00\xED\x98\x85\x00\xED\x98\x86\x00\xED\x98\x87\x00\xED\x98\x88\x00\xED"
-     "\x98\x89"
-     "\x00\xED\x98\x8A\x00\xED\x98\x8B\x00\xED\x98\x8C\x00\xED\x98\x8D\x00\xED"
-     "\x98\x8E"
-     "\x00\xED\x98\x8F\x00\xED\x98\x90\x00\xED\x98\x91\x00\xED\x98\x92\x00\xED"
-     "\x98\x93"
-     "\x00\xED\x98\x94\x00\xED\x98\x95\x00\xED\x98\x96\x00\xED\x98\x97\x00\xED"
-     "\x98\x98"
-     "\x00\xED\x98\x99\x00\xED\x98\x9A\x00\xED\x98\x9B\x00\xED\x98\x9C\x00\xED"
-     "\x98\x9D"
-     "\x00\xED\x98\x9E\x00\xED\x98\x9F\x00\xED\x98\xA0\x00\xED\x98\xA1\x00\xED"
-     "\x98\xA2"
-     "\x00\xED\x98\xA3\x00\xED\x98\xA4\x00\xED\x98\xA5\x00\xED\x98\xA6\x00\xED"
-     "\x98\xA7"
-     "\x00\xED\x98\xA8\x00\xED\x98\xA9\x00\xED\x98\xAA\x00\xED\x98\xAB\x00\xED"
-     "\x98\xAC"
-     "\x00\xED\x98\xAD\x00\xED\x98\xAE\x00\xED\x98\xAF\x00\xED\x98\xB0\x00\xED"
-     "\x98\xB1"
-     "\x00\xED\x98\xB2\x00\xED\x98\xB3\x00\xED\x98\xB4\x00\xED\x98\xB5\x00\xED"
-     "\x98\xB6"
-     "\x00\xED\x98\xB7\x00\xED\x98\xB8\x00\xED\x98\xB9\x00\xED\x98\xBA\x00\xED"
-     "\x98\xBB"
-     "\x00\xED\x98\xBC\x00\xED\x98\xBD\x00\xED\x98\xBE\x00\xED\x98\xBF\x00\xED"
-     "\x99\x80"
-     "\x00\xED\x99\x81\x00\xED\x99\x82\x00\xED\x99\x83\x00\xED\x99\x84\x00\xED"
-     "\x99\x85"
-     "\x00\xED\x99\x86\x00\xED\x99\x87\x00\xED\x99\x88\x00\xED\x99\x89\x00\xED"
-     "\x99\x8A"
-     "\x00\xED\x99\x8B\x00\xED\x99\x8C\x00\xED\x99\x8D\x00\xED\x99\x8E\x00\xED"
-     "\x99\x8F"
-     "\x00\xED\x99\x90\x00\xED\x99\x91\x00\xED\x99\x92\x00\xED\x99\x93\x00\xED"
-     "\x99\x94"
-     "\x00\xED\x99\x95\x00\xED\x99\x96\x00\xED\x99\x97\x00\xED\x99\x98\x00\xED"
-     "\x99\x99"
-     "\x00\xED\x99\x9A\x00\xED\x99\x9B\x00\xED\x99\x9C\x00\xED\x99\x9D\x00\xED"
-     "\x99\x9E"
-     "\x00\xED\x99\x9F\x00\xED\x99\xA0\x00\xED\x99\xA1\x00\xED\x99\xA2\x00\xED"
-     "\x99\xA3"
-     "\x00\xED\x99\xA4\x00\xED\x99\xA5\x00\xED\x99\xA6\x00\xED\x99\xA7\x00\xED"
-     "\x99\xA8"
-     "\x00\xED\x99\xA9\x00\xED\x99\xAA\x00\xED\x99\xAB\x00\xED\x99\xAC\x00\xED"
-     "\x99\xAD"
-     "\x00\xED\x99\xAE\x00\xED\x99\xAF\x00\xED\x99\xB0\x00\xED\x99\xB1\x00\xED"
-     "\x99\xB2"
-     "\x00\xED\x99\xB3\x00\xED\x99\xB4\x00\xED\x99\xB5\x00\xED\x99\xB6\x00\xED"
-     "\x99\xB7"
-     "\x00\xED\x99\xB8\x00\xED\x99\xB9\x00\xED\x99\xBA\x00\xED\x99\xBB\x00\xED"
-     "\x99\xBC"
-     "\x00\xED\x99\xBD\x00\xED\x99\xBE\x00\xED\x99\xBF\x00\xED\x9A\x80\x00\xED"
-     "\x9A\x81"
-     "\x00\xED\x9A\x82\x00\xED\x9A\x83\x00\xED\x9A\x84\x00\xED\x9A\x85\x00\xED"
-     "\x9A\x86"
-     "\x00\xED\x9A\x87\x00\xED\x9A\x88\x00\xED\x9A\x89\x00\xED\x9A\x8A\x00\xED"
-     "\x9A\x8B"
-     "\x00\xED\x9A\x8C\x00\xED\x9A\x8D\x00\xED\x9A\x8E\x00\xED\x9A\x8F\x00\xED"
-     "\x9A\x90"
-     "\x00\xED\x9A\x91\x00\xED\x9A\x92\x00\xED\x9A\x93\x00\xED\x9A\x94\x00\xED"
-     "\x9A\x95"
-     "\x00\xED\x9A\x96\x00\xED\x9A\x97\x00\xED\x9A\x98\x00\xED\x9A\x99\x00\xED"
-     "\x9A\x9A"
-     "\x00\xED\x9A\x9B\x00\xED\x9A\x9C\x00\xED\x9A\x9D\x00\xED\x9A\x9E\x00\xED"
-     "\x9A\x9F"
-     "\x00\xED\x9A\xA0\x00\xED\x9A\xA1\x00\xED\x9A\xA2\x00\xED\x9A\xA3\x00\xED"
-     "\x9A\xA4"
-     "\x00\xED\x9A\xA5\x00\xED\x9A\xA6\x00\xED\x9A\xA7\x00\xED\x9A\xA8\x00\xED"
-     "\x9A\xA9"
-     "\x00\xED\x9A\xAA\x00\xED\x9A\xAB\x00\xED\x9A\xAC\x00\xED\x9A\xAD\x00\xED"
-     "\x9A\xAE"
-     "\x00\xED\x9A\xAF\x00\xED\x9A\xB0\x00\xED\x9A\xB1\x00\xED\x9A\xB2\x00\xED"
-     "\x9A\xB3"
-     "\x00\xED\x9A\xB4\x00\xED\x9A\xB5\x00\xED\x9A\xB6\x00\xED\x9A\xB7\x00\xED"
-     "\x9A\xB8"
-     "\x00\xED\x9A\xB9\x00\xED\x9A\xBA\x00\xED\x9A\xBB\x00\xED\x9A\xBC\x00\xED"
-     "\x9A\xBD"
-     "\x00\xED\x9A\xBE\x00\xED\x9A\xBF\x00\xED\x9B\x80\x00\xED\x9B\x81\x00\xED"
-     "\x9B\x82"
-     "\x00\xED\x9B\x83\x00\xED\x9B\x84\x00\xED\x9B\x85\x00\xED\x9B\x86\x00\xED"
-     "\x9B\x87"
-     "\x00\xED\x9B\x88\x00\xED\x9B\x89\x00\xED\x9B\x8A\x00\xED\x9B\x8B\x00\xED"
-     "\x9B\x8C"
-     "\x00\xED\x9B\x8D\x00\xED\x9B\x8E\x00\xED\x9B\x8F\x00\xED\x9B\x90\x00\xED"
-     "\x9B\x91"
-     "\x00\xED\x9B\x92\x00\xED\x9B\x93\x00\xED\x9B\x94\x00\xED\x9B\x95\x00\xED"
-     "\x9B\x96"
-     "\x00\xED\x9B\x97\x00\xED\x9B\x98\x00\xED\x9B\x99\x00\xED\x9B\x9A\x00\xED"
-     "\x9B\x9B"
-     "\x00\xED\x9B\x9C\x00\xED\x9B\x9D\x00\xED\x9B\x9E\x00\xED\x9B\x9F\x00\xED"
-     "\x9B\xA0"
-     "\x00\xED\x9B\xA1\x00\xED\x9B\xA2\x00\xED\x9B\xA3\x00\xED\x9B\xA4\x00\xED"
-     "\x9B\xA5"
-     "\x00\xED\x9B\xA6\x00\xED\x9B\xA7\x00\xED\x9B\xA8\x00\xED\x9B\xA9\x00\xED"
-     "\x9B\xAA"
-     "\x00\xED\x9B\xAB\x00\xED\x9B\xAC\x00\xED\x9B\xAD\x00\xED\x9B\xAE\x00\xED"
-     "\x9B\xAF"
-     "\x00\xED\x9B\xB0\x00\xED\x9B\xB1\x00\xED\x9B\xB2\x00\xED\x9B\xB3\x00\xED"
-     "\x9B\xB4"
-     "\x00\xED\x9B\xB5\x00\xED\x9B\xB6\x00\xED\x9B\xB7\x00\xED\x9B\xB8\x00\xED"
-     "\x9B\xB9"
-     "\x00\xED\x9B\xBA\x00\xED\x9B\xBB\x00\xED\x9B\xBC\x00\xED\x9B\xBD\x00\xED"
-     "\x9B\xBE"
-     "\x00\xED\x9B\xBF\x00\xED\x9C\x80\x00\xED\x9C\x81\x00\xED\x9C\x82\x00\xED"
-     "\x9C\x83"
-     "\x00\xED\x9C\x84\x00\xED\x9C\x85\x00\xED\x9C\x86\x00\xED\x9C\x87\x00\xED"
-     "\x9C\x88"
-     "\x00\xED\x9C\x89\x00\xED\x9C\x8A\x00\xED\x9C\x8B\x00\xED\x9C\x8C\x00\xED"
-     "\x9C\x8D"
-     "\x00\xED\x9C\x8E\x00\xED\x9C\x8F\x00\xED\x9C\x90\x00\xED\x9C\x91\x00\xED"
-     "\x9C\x92"
-     "\x00\xED\x9C\x93\x00\xED\x9C\x94\x00\xED\x9C\x95\x00\xED\x9C\x96\x00\xED"
-     "\x9C\x97"
-     "\x00\xED\x9C\x98\x00\xED\x9C\x99\x00\xED\x9C\x9A\x00\xED\x9C\x9B\x00\xED"
-     "\x9C\x9C"
-     "\x00\xED\x9C\x9D\x00\xED\x9C\x9E\x00\xED\x9C\x9F\x00\xED\x9C\xA0\x00\xED"
-     "\x9C\xA1"
-     "\x00\xED\x9C\xA2\x00\xED\x9C\xA3\x00\xED\x9C\xA4\x00\xED\x9C\xA5\x00\xED"
-     "\x9C\xA6"
-     "\x00\xED\x9C\xA7\x00\xED\x9C\xA8\x00\xED\x9C\xA9\x00\xED\x9C\xAA\x00\xED"
-     "\x9C\xAB"
-     "\x00\xED\x9C\xAC\x00\xED\x9C\xAD\x00\xED\x9C\xAE\x00\xED\x9C\xAF\x00\xED"
-     "\x9C\xB0"
-     "\x00\xED\x9C\xB1\x00\xED\x9C\xB2\x00\xED\x9C\xB3\x00\xED\x9C\xB4\x00\xED"
-     "\x9C\xB5"
-     "\x00\xED\x9C\xB6\x00\xED\x9C\xB7\x00\xED\x9C\xB8\x00\xED\x9C\xB9\x00\xED"
-     "\x9C\xBA"
-     "\x00\xED\x9C\xBB\x00\xED\x9C\xBC\x00\xED\x9C\xBD\x00\xED\x9C\xBE\x00\xED"
-     "\x9C\xBF"
-     "\x00\xED\x9D\x80\x00\xED\x9D\x81\x00\xED\x9D\x82\x00\xED\x9D\x83\x00\xED"
-     "\x9D\x84"
-     "\x00\xED\x9D\x85\x00\xED\x9D\x86\x00\xED\x9D\x87\x00\xED\x9D\x88\x00\xED"
-     "\x9D\x89"
-     "\x00\xED\x9D\x8A\x00\xED\x9D\x8B\x00\xED\x9D\x8C\x00\xED\x9D\x8D\x00\xED"
-     "\x9D\x8E"
-     "\x00\xED\x9D\x8F\x00\xED\x9D\x90\x00\xED\x9D\x91\x00\xED\x9D\x92\x00\xED"
-     "\x9D\x93"
-     "\x00\xED\x9D\x94\x00\xED\x9D\x95\x00\xED\x9D\x96\x00\xED\x9D\x97\x00\xED"
-     "\x9D\x98"
-     "\x00\xED\x9D\x99\x00\xED\x9D\x9A\x00\xED\x9D\x9B\x00\xED\x9D\x9C\x00\xED"
-     "\x9D\x9D"
-     "\x00\xED\x9D\x9E\x00\xED\x9D\x9F\x00\xED\x9D\xA0\x00\xED\x9D\xA1\x00\xED"
-     "\x9D\xA2"
-     "\x00\xED\x9D\xA3\x00\xED\x9D\xA4\x00\xED\x9D\xA5\x00\xED\x9D\xA6\x00\xED"
-     "\x9D\xA7"
-     "\x00\xED\x9D\xA8\x00\xED\x9D\xA9\x00\xED\x9D\xAA\x00\xED\x9D\xAB\x00\xED"
-     "\x9D\xAC"
-     "\x00\xED\x9D\xAD\x00\xED\x9D\xAE\x00\xED\x9D\xAF\x00\xED\x9D\xB0\x00\xED"
-     "\x9D\xB1"
-     "\x00\xED\x9D\xB2\x00\xED\x9D\xB3\x00\xED\x9D\xB4\x00\xED\x9D\xB5\x00\xED"
-     "\x9D\xB6"
-     "\x00\xED\x9D\xB7\x00\xED\x9D\xB8\x00\xED\x9D\xB9\x00\xED\x9D\xBA\x00\xED"
-     "\x9D\xBB"
-     "\x00\xED\x9D\xBC\x00\xED\x9D\xBD\x00\xED\x9D\xBE\x00\xED\x9D\xBF\x00\xED"
-     "\x9E\x80"
-     "\x00\xED\x9E\x81\x00\xED\x9E\x82\x00\xED\x9E\x83\x00\xED\x9E\x84\x00\xED"
-     "\x9E\x85"
-     "\x00\xED\x9E\x86\x00\xED\x9E\x87\x00\xED\x9E\x88\x00\xED\x9E\x89\x00\xED"
-     "\x9E\x8A"
-     "\x00\xED\x9E\x8B\x00\xED\x9E\x8C\x00\xED\x9E\x8D\x00\xED\x9E\x8E\x00\xED"
-     "\x9E\x8F"
-     "\x00\xED\x9E\x90\x00\xED\x9E\x91\x00\xED\x9E\x92\x00\xED\x9E\x93\x00\xED"
-     "\x9E\x94"
-     "\x00\xED\x9E\x95\x00\xED\x9E\x96\x00\xED\x9E\x97\x00\xED\x9E\x98\x00\xED"
-     "\x9E\x99"
-     "\x00\xED\x9E\x9A\x00\xED\x9E\x9B\x00\xED\x9E\x9C\x00\xED\x9E\x9D\x00\xED"
-     "\x9E\x9E"
-     "\x00\xED\x9E\x9F\x00\xED\x9E\xA0\x00\xED\x9E\xA1\x00\xED\x9E\xA2\x00\xED"
-     "\x9E\xA3"
-     "\x00\xF0\x90\x90\xA8\x00\xF0\x90\x90\xA9\x00\xF0\x90\x90\xAA\x00\xF0\x90"
-     "\x90\xAB"
-     "\x00\xF0\x90\x90\xAC\x00\xF0\x90\x90\xAD\x00\xF0\x90\x90\xAE\x00\xF0\x90"
-     "\x90\xAF"
-     "\x00\xF0\x90\x90\xB0\x00\xF0\x90\x90\xB1\x00\xF0\x90\x90\xB2\x00\xF0\x90"
-     "\x90\xB3"
-     "\x00\xF0\x90\x90\xB4\x00\xF0\x90\x90\xB5\x00\xF0\x90\x90\xB6\x00\xF0\x90"
-     "\x90\xB7"
-     "\x00\xF0\x90\x90\xB8\x00\xF0\x90\x90\xB9\x00\xF0\x90\x90\xBA\x00\xF0\x90"
-     "\x90\xBB"
-     "\x00\xF0\x90\x90\xBC\x00\xF0\x90\x90\xBD\x00\xF0\x90\x90\xBE\x00\xF0\x90"
-     "\x90\xBF"
-     "\x00\xF0\x90\x91\x80\x00\xF0\x90\x91\x81\x00\xF0\x90\x91\x82\x00\xF0\x90"
-     "\x91\x83"
-     "\x00\xF0\x90\x91\x84\x00\xF0\x90\x91\x85\x00\xF0\x90\x91\x86\x00\xF0\x90"
-     "\x91\x87"
-     "\x00\xF0\x90\x91\x88\x00\xF0\x90\x91\x89\x00\xF0\x90\x91\x8A\x00\xF0\x90"
-     "\x91\x8B"
-     "\x00\xF0\x90\x91\x8C\x00\xF0\x90\x91\x8D\x00\xF0\x90\x91\x8E\x00\xF0\x90"
-     "\x91\x8F"
-     "\x00\xF0\x90\x93\x98\x00\xF0\x90\x93\x99\x00\xF0\x90\x93\x9A\x00\xF0\x90"
-     "\x93\x9B"
-     "\x00\xF0\x90\x93\x9C\x00\xF0\x90\x93\x9D\x00\xF0\x90\x93\x9E\x00\xF0\x90"
-     "\x93\x9F"
-     "\x00\xF0\x90\x93\xA0\x00\xF0\x90\x93\xA1\x00\xF0\x90\x93\xA2\x00\xF0\x90"
-     "\x93\xA3"
-     "\x00\xF0\x90\x93\xA4\x00\xF0\x90\x93\xA5\x00\xF0\x90\x93\xA6\x00\xF0\x90"
-     "\x93\xA7"
-     "\x00\xF0\x90\x93\xA8\x00\xF0\x90\x93\xA9\x00\xF0\x90\x93\xAA\x00\xF0\x90"
-     "\x93\xAB"
-     "\x00\xF0\x90\x93\xAC\x00\xF0\x90\x93\xAD\x00\xF0\x90\x93\xAE\x00\xF0\x90"
-     "\x93\xAF"
-     "\x00\xF0\x90\x93\xB0\x00\xF0\x90\x93\xB1\x00\xF0\x90\x93\xB2\x00\xF0\x90"
-     "\x93\xB3"
-     "\x00\xF0\x90\x93\xB4\x00\xF0\x90\x93\xB5\x00\xF0\x90\x93\xB6\x00\xF0\x90"
-     "\x93\xB7"
-     "\x00\xF0\x90\x93\xB8\x00\xF0\x90\x93\xB9\x00\xF0\x90\x93\xBA\x00\xF0\x90"
-     "\x93\xBB"
-     "\x00\xF0\x90\x96\x97\x00\xF0\x90\x96\x98\x00\xF0\x90\x96\x99\x00\xF0\x90"
-     "\x96\x9A"
-     "\x00\xF0\x90\x96\x9B\x00\xF0\x90\x96\x9C\x00\xF0\x90\x96\x9D\x00\xF0\x90"
-     "\x96\x9E"
-     "\x00\xF0\x90\x96\x9F\x00\xF0\x90\x96\xA0\x00\xF0\x90\x96\xA1\x00\xF0\x90"
-     "\x96\xA3"
-     "\x00\xF0\x90\x96\xA4\x00\xF0\x90\x96\xA5\x00\xF0\x90\x96\xA6\x00\xF0\x90"
-     "\x96\xA7"
-     "\x00\xF0\x90\x96\xA8\x00\xF0\x90\x96\xA9\x00\xF0\x90\x96\xAA\x00\xF0\x90"
-     "\x96\xAB"
-     "\x00\xF0\x90\x96\xAC\x00\xF0\x90\x96\xAD\x00\xF0\x90\x96\xAE\x00\xF0\x90"
-     "\x96\xAF"
-     "\x00\xF0\x90\x96\xB0\x00\xF0\x90\x96\xB1\x00\xF0\x90\x96\xB3\x00\xF0\x90"
-     "\x96\xB4"
-     "\x00\xF0\x90\x96\xB5\x00\xF0\x90\x96\xB6\x00\xF0\x90\x96\xB7\x00\xF0\x90"
-     "\x96\xB8"
-     "\x00\xF0\x90\x96\xB9\x00\xF0\x90\x96\xBB\x00\xF0\x90\x96\xBC\x00\xF0\x90"
-     "\xB3\x80"
-     "\x00\xF0\x90\xB3\x81\x00\xF0\x90\xB3\x82\x00\xF0\x90\xB3\x83\x00\xF0\x90"
-     "\xB3\x84"
-     "\x00\xF0\x90\xB3\x85\x00\xF0\x90\xB3\x86\x00\xF0\x90\xB3\x87\x00\xF0\x90"
-     "\xB3\x88"
-     "\x00\xF0\x90\xB3\x89\x00\xF0\x90\xB3\x8A\x00\xF0\x90\xB3\x8B\x00\xF0\x90"
-     "\xB3\x8C"
-     "\x00\xF0\x90\xB3\x8D\x00\xF0\x90\xB3\x8E\x00\xF0\x90\xB3\x8F\x00\xF0\x90"
-     "\xB3\x90"
-     "\x00\xF0\x90\xB3\x91\x00\xF0\x90\xB3\x92\x00\xF0\x90\xB3\x93\x00\xF0\x90"
-     "\xB3\x94"
-     "\x00\xF0\x90\xB3\x95\x00\xF0\x90\xB3\x96\x00\xF0\x90\xB3\x97\x00\xF0\x90"
-     "\xB3\x98"
-     "\x00\xF0\x90\xB3\x99\x00\xF0\x90\xB3\x9A\x00\xF0\x90\xB3\x9B\x00\xF0\x90"
-     "\xB3\x9C"
-     "\x00\xF0\x90\xB3\x9D\x00\xF0\x90\xB3\x9E\x00\xF0\x90\xB3\x9F\x00\xF0\x90"
-     "\xB3\xA0"
-     "\x00\xF0\x90\xB3\xA1\x00\xF0\x90\xB3\xA2\x00\xF0\x90\xB3\xA3\x00\xF0\x90"
-     "\xB3\xA4"
-     "\x00\xF0\x90\xB3\xA5\x00\xF0\x90\xB3\xA6\x00\xF0\x90\xB3\xA7\x00\xF0\x90"
-     "\xB3\xA8"
-     "\x00\xF0\x90\xB3\xA9\x00\xF0\x90\xB3\xAA\x00\xF0\x90\xB3\xAB\x00\xF0\x90"
-     "\xB3\xAC"
-     "\x00\xF0\x90\xB3\xAD\x00\xF0\x90\xB3\xAE\x00\xF0\x90\xB3\xAF\x00\xF0\x90"
-     "\xB3\xB0"
-     "\x00\xF0\x90\xB3\xB1\x00\xF0\x90\xB3\xB2\x00\xF0\x91\x82\x9A\x00\xF0\x91"
-     "\x82\x9C"
-     "\x00\xF0\x91\x82\xAB\x00\xF0\x91\x84\xAE\x00\xF0\x91\x84\xAF\x00\xF0\x91"
-     "\x8D\x8B"
-     "\x00\xF0\x91\x8D\x8C\x00\xF0\x91\x92\xBB\x00\xF0\x91\x92\xBC\x00\xF0\x91"
-     "\x92\xBE"
-     "\x00\xF0\x91\x96\xBA\x00\xF0\x91\x96\xBB\x00\xF0\x91\xA3\x80\x00\xF0\x91"
-     "\xA3\x81"
-     "\x00\xF0\x91\xA3\x82\x00\xF0\x91\xA3\x83\x00\xF0\x91\xA3\x84\x00\xF0\x91"
-     "\xA3\x85"
-     "\x00\xF0\x91\xA3\x86\x00\xF0\x91\xA3\x87\x00\xF0\x91\xA3\x88\x00\xF0\x91"
-     "\xA3\x89"
-     "\x00\xF0\x91\xA3\x8A\x00\xF0\x91\xA3\x8B\x00\xF0\x91\xA3\x8C\x00\xF0\x91"
-     "\xA3\x8D"
-     "\x00\xF0\x91\xA3\x8E\x00\xF0\x91\xA3\x8F\x00\xF0\x91\xA3\x90\x00\xF0\x91"
-     "\xA3\x91"
-     "\x00\xF0\x91\xA3\x92\x00\xF0\x91\xA3\x93\x00\xF0\x91\xA3\x94\x00\xF0\x91"
-     "\xA3\x95"
-     "\x00\xF0\x91\xA3\x96\x00\xF0\x91\xA3\x97\x00\xF0\x91\xA3\x98\x00\xF0\x91"
-     "\xA3\x99"
-     "\x00\xF0\x91\xA3\x9A\x00\xF0\x91\xA3\x9B\x00\xF0\x91\xA3\x9C\x00\xF0\x91"
-     "\xA3\x9D"
-     "\x00\xF0\x91\xA3\x9E\x00\xF0\x91\xA3\x9F\x00\xF0\x91\xA4\xB8\x00\xF0\x96"
-     "\xB9\xA0"
-     "\x00\xF0\x96\xB9\xA1\x00\xF0\x96\xB9\xA2\x00\xF0\x96\xB9\xA3\x00\xF0\x96"
-     "\xB9\xA4"
-     "\x00\xF0\x96\xB9\xA5\x00\xF0\x96\xB9\xA6\x00\xF0\x96\xB9\xA7\x00\xF0\x96"
-     "\xB9\xA8"
-     "\x00\xF0\x96\xB9\xA9\x00\xF0\x96\xB9\xAA\x00\xF0\x96\xB9\xAB\x00\xF0\x96"
-     "\xB9\xAC"
-     "\x00\xF0\x96\xB9\xAD\x00\xF0\x96\xB9\xAE\x00\xF0\x96\xB9\xAF\x00\xF0\x96"
-     "\xB9\xB0"
-     "\x00\xF0\x96\xB9\xB1\x00\xF0\x96\xB9\xB2\x00\xF0\x96\xB9\xB3\x00\xF0\x96"
-     "\xB9\xB4"
-     "\x00\xF0\x96\xB9\xB5\x00\xF0\x96\xB9\xB6\x00\xF0\x96\xB9\xB7\x00\xF0\x96"
-     "\xB9\xB8"
-     "\x00\xF0\x96\xB9\xB9\x00\xF0\x96\xB9\xBA\x00\xF0\x96\xB9\xBB\x00\xF0\x96"
-     "\xB9\xBC"
-     "\x00\xF0\x96\xB9\xBD\x00\xF0\x96\xB9\xBE\x00\xF0\x96\xB9\xBF\x00\xF0\x9D"
-     "\x85\x97"
-     "\xF0\x9D\x85\xA5\x00\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\x00\xF0\x9D\x85\x98"
-     "\xF0\x9D"
-     "\x85\xA5\xF0\x9D\x85\xAE\x00\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\xF0\x9D\x85"
-     "\xAF\x00"
-     "\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\xF0\x9D\x85\xB0\x00\xF0\x9D\x85\x98\xF0"
-     "\x9D\x85"
-     "\xA5\xF0\x9D\x85\xB1\x00\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\xF0\x9D\x85\xB2"
-     "\x00\xF0"
-     "\x9D\x86\xB9\xF0\x9D\x85\xA5\x00\xF0\x9D\x86\xB9\xF0\x9D\x85\xA5\xF0\x9D"
-     "\x85\xAE"
-     "\x00\xF0\x9D\x86\xB9\xF0\x9D\x85\xA5\xF0\x9D\x85\xAF\x00\xF0\x9D\x86\xBA"
-     "\xF0\x9D"
-     "\x85\xA5\x00\xF0\x9D\x86\xBA\xF0\x9D\x85\xA5\xF0\x9D\x85\xAE\x00\xF0\x9D"
-     "\x86\xBA"
-     "\xF0\x9D\x85\xA5\xF0\x9D\x85\xAF\x00\xF0\x9D\xBC\x84\x00\xF0\x9D\xBC\x85"
-     "\x00\xF0"
-     "\x9D\xBC\x86\x00\xF0\x9D\xBC\x88\x00\xF0\x9D\xBC\x8A\x00\xF0\x9D\xBC\x9E"
-     "\x00\xF0"
-     "\x9E\xA4\xA2\x00\xF0\x9E\xA4\xA3\x00\xF0\x9E\xA4\xA4\x00\xF0\x9E\xA4\xA5"
-     "\x00\xF0"
-     "\x9E\xA4\xA6\x00\xF0\x9E\xA4\xA7\x00\xF0\x9E\xA4\xA8\x00\xF0\x9E\xA4\xA9"
-     "\x00\xF0"
-     "\x9E\xA4\xAA\x00\xF0\x9E\xA4\xAB\x00\xF0\x9E\xA4\xAC\x00\xF0\x9E\xA4\xAD"
-     "\x00\xF0"
-     "\x9E\xA4\xAE\x00\xF0\x9E\xA4\xAF\x00\xF0\x9E\xA4\xB0\x00\xF0\x9E\xA4\xB1"
-     "\x00\xF0"
-     "\x9E\xA4\xB2\x00\xF0\x9E\xA4\xB3\x00\xF0\x9E\xA4\xB4\x00\xF0\x9E\xA4\xB5"
-     "\x00\xF0"
-     "\x9E\xA4\xB6\x00\xF0\x9E\xA4\xB7\x00\xF0\x9E\xA4\xB8\x00\xF0\x9E\xA4\xB9"
-     "\x00\xF0"
-     "\x9E\xA4\xBA\x00\xF0\x9E\xA4\xBB\x00\xF0\x9E\xA4\xBC\x00\xF0\x9E\xA4\xBD"
-     "\x00\xF0"
-     "\x9E\xA4\xBE\x00\xF0\x9E\xA4\xBF\x00\xF0\x9E\xA5\x80\x00\xF0\x9E\xA5\x81"
-     "\x00\xF0"
-     "\x9E\xA5\x82\x00\xF0\x9E\xA5\x83\x00\xF0\xA0\x84\xA2\x00\xF0\xA0\x94\x9C"
-     "\x00\xF0"
-     "\xA0\x94\xA5\x00\xF0\xA0\x95\x8B\x00\xF0\xA0\x98\xBA\x00\xF0\xA0\xA0\x84"
-     "\x00\xF0"
-     "\xA0\xA3\x9E\x00\xF0\xA0\xA8\xAC\x00\xF0\xA0\xAD\xA3\x00\xF0\xA1\x93\xA4"
-     "\x00\xF0"
-     "\xA1\x9A\xA8\x00\xF0\xA1\x9B\xAA\x00\xF0\xA1\xA7\x88\x00\xF0\xA1\xAC\x98"
-     "\x00\xF0"
-     "\xA1\xB4\x8B\x00\xF0\xA1\xB7\xA4\x00\xF0\xA1\xB7\xA6\x00\xF0\xA2\x86\x83"
-     "\x00\xF0"
-     "\xA2\x86\x9F\x00\xF0\xA2\x8C\xB1\x00\xF0\xA2\x9B\x94\x00\xF0\xA2\xA1\x84"
-     "\x00\xF0"
-     "\xA2\xA1\x8A\x00\xF0\xA2\xAC\x8C\x00\xF0\xA2\xAF\xB1\x00\xF0\xA3\x80\x8A"
-     "\x00\xF0"
-     "\xA3\x8A\xB8\x00\xF0\xA3\x8D\x9F\x00\xF0\xA3\x8E\x93\x00\xF0\xA3\x8E\x9C"
-     "\x00\xF0"
-     "\xA3\x8F\x83\x00\xF0\xA3\x8F\x95\x00\xF0\xA3\x91\xAD\x00\xF0\xA3\x9A\xA3"
-     "\x00\xF0"
-     "\xA3\xA2\xA7\x00\xF0\xA3\xAA\x8D\x00\xF0\xA3\xAB\xBA\x00\xF0\xA3\xB2\xBC"
-     "\x00\xF0"
-     "\xA3\xB4\x9E\x00\xF0\xA3\xBB\x91\x00\xF0\xA3\xBD\x9E\x00\xF0\xA3\xBE\x8E"
-     "\x00\xF0"
-     "\xA4\x89\xA3\x00\xF0\xA4\x8B\xAE\x00\xF0\xA4\x8E\xAB\x00\xF0\xA4\x98\x88"
-     "\x00\xF0"
-     "\xA4\x9C\xB5\x00\xF0\xA4\xA0\x94\x00\xF0\xA4\xB0\xB6\x00\xF0\xA4\xB2\x92"
-     "\x00\xF0"
-     "\xA4\xBE\xA1\x00\xF0\xA4\xBE\xB8\x00\xF0\xA5\x81\x84\x00\xF0\xA5\x83\xB2"
-     "\x00\xF0"
-     "\xA5\x83\xB3\x00\xF0\xA5\x84\x99\x00\xF0\xA5\x84\xB3\x00\xF0\xA5\x89\x89"
-     "\x00\xF0"
-     "\xA5\x90\x9D\x00\xF0\xA5\x98\xA6\x00\xF0\xA5\x9A\x9A\x00\xF0\xA5\x9B\x85"
-     "\x00\xF0"
-     "\xA5\xA5\xBC\x00\xF0\xA5\xAA\xA7\x00\xF0\xA5\xAE\xAB\x00\xF0\xA5\xB2\x80"
-     "\x00\xF0"
-     "\xA5\xB3\x90\x00\xF0\xA5\xBE\x86\x00\xF0\xA6\x87\x9A\x00\xF0\xA6\x88\xA8"
-     "\x00\xF0"
-     "\xA6\x89\x87\x00\xF0\xA6\x8B\x99\x00\xF0\xA6\x8C\xBE\x00\xF0\xA6\x93\x9A"
-     "\x00\xF0"
-     "\xA6\x94\xA3\x00\xF0\xA6\x96\xA8\x00\xF0\xA6\x9E\xA7\x00\xF0\xA6\x9E\xB5"
-     "\x00\xF0"
-     "\xA6\xAC\xBC\x00\xF0\xA6\xB0\xB6\x00\xF0\xA6\xB3\x95\x00\xF0\xA6\xB5\xAB"
-     "\x00\xF0"
-     "\xA6\xBC\xAC\x00\xF0\xA6\xBE\xB1\x00\xF0\xA7\x83\x92\x00\xF0\xA7\x8F\x8A"
-     "\x00\xF0"
-     "\xA7\x99\xA7\x00\xF0\xA7\xA2\xAE\x00\xF0\xA7\xA5\xA6\x00\xF0\xA7\xB2\xA8"
-     "\x00\xF0"
-     "\xA7\xBB\x93\x00\xF0\xA7\xBC\xAF\x00\xF0\xA8\x97\x92\x00\xF0\xA8\x97\xAD"
-     "\x00\xF0"
-     "\xA8\x9C\xAE\x00\xF0\xA8\xAF\xBA\x00\xF0\xA8\xB5\xB7\x00\xF0\xA9\x85\x85"
-     "\x00\xF0"
-     "\xA9\x87\x9F\x00\xF0\xA9\x88\x9A\x00\xF0\xA9\x90\x8A\x00\xF0\xA9\x92\x96"
-     "\x00\xF0"
-     "\xA9\x96\xB6\x00\xF0\xA9\xAC\xB0\x00\xF0\xAA\x83\x8E\x00\xF0\xAA\x84\x85"
-     "\x00\xF0"
-     "\xAA\x88\x8E\x00\xF0\xAA\x8A\x91\x00\xF0\xAA\x8E\x92\x00\xF0\xAA\x98\x80"
-     "\x00"},
-    {"nmt_nfkc_cf", 245778,
-     "\x00\xCC\x02\x00\x00\x84\x00\x00\x00\x00\x00\x80\x01\x00\x00\x80\xCC\xFC"
-     "\x02\x00"
-     "\xB8\x05\x00\x00\x5C\x21\x00\x80\xCC\xE4\x02\x00\xB8\x3D\x00\x00\x50\x21"
-     "\x00\x80"
-     "\xCC\xEC\x02\x00\xB8\x05\x00\x00\x60\x21\x00\x80\xB2\x08\x00\x80\xB5\x08"
-     "\x00\x80"
-     "\x0A\x1E\x00\x80\x06\x1E\x00\x80\x63\x06\x00\x80\x12\x1E\x00\x80\xB8\x08"
-     "\x00\x80"
-     "\x0E\x1E\x00\x80\x83\x15\x00\x00\xBB\x08\x00\x80\x81\x65\x00\x00\x80\x65"
-     "\x00\x00"
-     "\x80\x59\x00\x00\x81\x59\x00\x00\x0F\x09\x00\x80\x12\x09\x00\x80\x3D\x58"
-     "\x03\x00"
-     "\x3C\x48\x03\x00\x89\x35\x00\x00\x3E\x68\x03\x00\x01\x85\x00\x00\x84\xF9"
-     "\x01\x00"
-     "\x03\x8D\x00\x00\x02\x89\x00\x00\x05\x95\x00\x00\x04\x91\x00\x00\x07\x9D"
-     "\x00\x00"
-     "\x06\x99\x00\x00\x09\xA9\x00\x00\x08\xA1\x00\x00\x0B\xAD\x00\x00\x0A\xA5"
-     "\x00\x00"
-     "\x0D\xB9\x00\x00\x0C\xBD\x00\x00\x0F\xBD\x00\x00\x0E\xB9\x00\x00\x11\xC5"
-     "\x00\x00"
-     "\x10\xC1\x00\x00\x13\xCD\x00\x00\x12\xC9\x00\x00\x15\xD5\x00\x00\x14\xD1"
-     "\x00\x00"
-     "\x17\xDD\x00\x00\x16\xD9\x00\x00\x19\xE5\x00\x00\x18\xE1\x00\x00\x1B\xED"
-     "\x00\x00"
-     "\x1A\xE9\x00\x00\x1D\xF5\x00\x00\x1C\xF1\x00\x00\x1F\xFD\x00\x00\x1E\xF9"
-     "\x00\x00"
-     "\x61\xB0\x09\x00\x1E\x1E\x00\x80\x63\x90\x09\x00\x62\x24\x09\x00\x65\xE0"
-     "\x09\x00"
-     "\x64\x9C\x09\x00\x67\xCC\x09\x00\x66\xE0\x09\x00\x69\xE0\x09\x00\x68\xFC"
-     "\x09\x00"
-     "\x6B\xE4\x09\x00\x6A\xF0\x09\x00\x6D\x90\x0E\x00\x6C\x64\x0E\x00\x6F\xA8"
-     "\x0E\x00"
-     "\x6E\x94\x0E\x00\x1A\x1E\x00\x80\x70\xDC\x0E\x00\x73\xA0\x0E\x00\x72\xCC"
-     "\x0E\x00"
-     "\x75\x80\x0E\x00\x74\x60\x0E\x00\x77\x80\x0E\x00\x76\x60\x0E\x00\x79\xA8"
-     "\x0E\x00"
-     "\x78\x48\x0E\x00\x26\x1E\x00\x80\x7A\x04\x0F\x00\x22\x1E\x00\x80\x8F\x0A"
-     "\x00\x80"
-     "\x7F\x7D\x01\x00\x2F\x0A\x00\x80\x41\xC1\x01\x00\xBE\x08\x00\x80\x43\x81"
-     "\x05\x00"
-     "\x42\x89\x05\x00\x45\xB5\x05\x00\x44\x89\x05\x00\x47\x19\x05\x00\x46\xCD"
-     "\x05\x00"
-     "\x49\x1D\x05\x00\x48\x31\x05\x00\x4B\xF5\x04\x00\x4A\xE9\x04\x00\x4D\x15"
-     "\x04\x00"
-     "\x4C\x35\x04\x00\x4F\x55\x04\x00\x4E\x7D\x04\x00\x51\x05\x06\x00\x50\xFD"
-     "\x09\x00"
-     "\x53\x95\x09\x00\x52\x8D\x09\x00\x55\x39\x09\x00\x54\x75\x09\x00\x57\x01"
-     "\x0C\x00"
-     "\x56\x25\x08\x00\x59\xE9\x0D\x00\x58\xE5\x0D\x00\x84\x11\x00\x00\x5A\xF5"
-     "\x0D\x00"
-     "\x80\xF5\x00\x00\x81\xB5\x00\x00\x2C\x0A\x00\x80\x02\x1E\x00\x80\x83\x55"
-     "\x02\x00"
-     "\x82\x4D\x02\x00\x81\x3D\x02\x00\x80\x3D\x02\x00\x87\x65\x03\x00\x86\x79"
-     "\x02\x00"
-     "\x81\x39\x00\x00\x84\x75\x02\x00\x50\x0A\x00\x80\x8A\x01\x04\x00\x89\xD5"
-     "\x03\x00"
-     "\x88\xA9\x03\x00\x8F\x09\x00\x00\x11\x0A\x00\x80\x59\x0A\x00\x80\x8C\x09"
-     "\x00\x00"
-     "\x5C\x0A\x00\x80\xCC\xD0\x00\x00\x91\x09\x00\x00\x16\x1E\x00\x80\x80\x55"
-     "\x03\x00"
-     "\x81\x15\x03\x00\x2A\x1E\x00\x80\x83\x35\x03\x00\x87\x11\x00\x00\xC2\x1C"
-     "\x00\x80"
-     "\x82\x25\x00\x00\x15\x09\x00\x80\xC6\x1C\x00\x80\x89\x05\x03\x00\x86\x21"
-     "\x00\x00"
-     "\xCA\x1C\x00\x80\xA3\x81\x06\x00\xCE\x1C\x00\x80\x80\xB9\x02\x00\x81\xB9"
-     "\x02\x00"
-     "\x18\x09\x00\x80\x81\x3D\x00\x00\xA5\xFD\x00\x00\x1B\x09\x00\x80\x1E\x09"
-     "\x00\x80"
-     "\x21\x09\x00\x80\xD2\x1C\x00\x80\xA8\xC1\x00\x00\x81\x21\x00\x00\xCC\x4C"
-     "\x01\x00"
-     "\xB1\x3D\x00\x00\xD6\x1C\x00\x80\x24\x09\x00\x80\xDA\x1C\x00\x80\x89\xD9"
-     "\x01\x00"
-     "\xE2\x1C\x00\x80\xE6\x1C\x00\x80\xDE\x1C\x00\x80\xCA\x08\x00\x80\xCD\x08"
-     "\x00\x80"
-     "\x83\x19\x00\x00\x3E\x1E\x00\x80\x81\x1D\x00\x00\x80\x09\x00\x00\xA3\x8D"
-     "\x00\x00"
-     "\x3A\x1E\x00\x80\x46\x1E\x00\x80\x82\x91\x01\x00\xE1\x04\x29\x00\xE0\x14"
-     "\x25\x00"
-     "\xE3\x48\x51\x02\xE2\x0C\x3D\x02\x42\x1E\x00\x80\x86\x95\x00\x00\x80\xFD"
-     "\x01\x00"
-     "\x81\xE9\x01\x00\x36\x1E\x00\x80\xEA\x1C\x00\x80\x87\xC5\x00\x00\xEA\x00"
-     "\x6C\x02"
-     "\x80\x15\x00\x00\x81\x75\x00\x00\xEF\xA8\x73\x02\x81\xC1\x00\x00\xEE\x1C"
-     "\x00\x80"
-     "\xF0\x44\x92\x02\x2D\x09\x00\x80\x30\x09\x00\x80\xD3\x08\x00\x80\x32\x1E"
-     "\x00\x80"
-     "\xCD\xA8\x00\x00\xCC\x0C\x03\x00\x36\x09\x00\x80\x5F\x0A\x00\x80\x62\x0A"
-     "\x00\x80"
-     "\x4A\x1E\x00\x80\xCC\x7C\x03\x00\xCD\xFC\x03\x00\xCC\x24\x01\x00\xCD\x1C"
-     "\x01\x00"
-     "\xFA\x1C\x00\x80\x33\x09\x00\x80\xC3\x08\x0E\x00\xC2\xB8\x0F\x00\xC5\x90"
-     "\x11\x00"
-     "\xC4\x00\x10\x00\xC7\x98\x12\x00\xC6\x98\x13\x00\xC9\xA4\x15\x00\xC8\x00"
-     "\x14\x00"
-     "\xCB\x98\x15\x00\xCA\xE0\x15\x00\xCD\xB0\x17\x00\xF2\x1C\x00\x80\xCF\x00"
-     "\x1C\x00"
-     "\xCE\xAC\x16\x00\xD1\x84\x1E\x00\xD0\x20\x1D\x00\xD3\xC8\x20\x00\xD2\xD4"
-     "\x21\x00"
-     "\xD5\x90\x22\x00\xD4\x2C\x23\x00\xF6\x1C\x00\x80\xD6\x04\x23\x00\xD9\x00"
-     "\x24\x00"
-     "\xD8\xBC\x22\x00\xDB\xE0\x25\x00\xFE\x1C\x00\x80\x80\x29\x01\x00\x81\x29"
-     "\x01\x00"
-     "\x84\x69\x06\x00\x4A\x0A\x00\x80\xFE\x1D\x00\x80\x7F\x06\x00\x80\x88\x06"
-     "\x00\x80"
-     "\x87\xAD\x06\x00\xC7\x08\x00\x80\x81\x85\x06\x00\x82\x85\x06\x00\xB2\x06"
-     "\x00\x80"
-     "\x8C\x85\x06\x00\xD2\x06\x00\x80\xD0\x08\x00\x80\x39\x09\x00\x80\x2A\x09"
-     "\x00\x80"
-     "\x02\x1D\x00\x80\x2E\x1E\x00\x80\x91\x55\x07\x00\x87\xFD\x06\x00\x8F\x21"
-     "\x07\x00"
-     "\x8C\x29\x07\x00\x92\x0A\x00\x80\xDB\x06\x00\x80\x3C\x09\x00\x80\x88\x09"
-     "\x07\x00"
-     "\x89\x09\x07\x00\x86\x29\x07\x00\x87\x29\x07\x00\x84\x59\x00\x00\x8C\xAD"
-     "\x06\x00"
-     "\x82\x59\x00\x00\x83\x55\x07\x00\x80\xA1\x06\x00\x81\xA1\x06\x00\xF5\x06"
-     "\x00\x80"
-     "\x3F\x09\x00\x80\x38\x0A\x00\x80\xA7\x9D\x00\x00\x42\x09\x00\x80\x09\x07"
-     "\x00\x80"
-     "\xB1\x4D\x06\x00\x45\x09\x00\x80\x06\x1D\x00\x80\x0E\x1D\x00\x80\x86\x0A"
-     "\x00\x80"
-     "\x0A\x1D\x00\x80\x12\x1D\x00\x80\x16\x1D\x00\x80\xB0\x61\x07\x00\x1E\x07"
-     "\x00\x80"
-     "\xA7\x0D\x06\x00\xD6\x08\x00\x80\xD9\x08\x00\x80\xAD\x79\x07\x00\xA3\x15"
-     "\x06\x00"
-     "\xDC\x08\x00\x80\xA8\x5D\x07\x00\x4B\x09\x00\x80\x4E\x09\x00\x80\xA7\xA9"
-     "\x00\x00"
-     "\xAD\x39\x06\x00\x51\x09\x00\x80\x57\x09\x00\x80\xA3\xCD\x00\x00\xDF\x08"
-     "\x00\x80"
-     "\x81\xFD\x01\x00\x1E\x1D\x00\x80\x4E\x1E\x00\x80\x14\x0A\x00\x80\x65\x0A"
-     "\x00\x80"
-     "\xCD\x0C\x05\x00\xCC\x08\x05\x00\x68\x0A\x00\x80\x8C\x95\x01\x00\x52\x1E"
-     "\x00\x80"
-     "\x82\x09\x01\x00\x81\xD5\x06\x00\x54\x09\x00\x80\x87\xB5\x01\x00\x86\x61"
-     "\x01\x00"
-     "\x1A\x1D\x00\x80\x84\x09\x01\x00\x82\xDD\x01\x00\x36\x07\x00\x80\x81\x81"
-     "\x01\x00"
-     "\x80\x89\x01\x00\x82\x09\x00\x00\x87\xF5\x01\x00\x5D\x09\x00\x80\x38\x07"
-     "\x00\x80"
-     "\x22\x1D\x00\x80\x3B\x0A\x00\x80\x88\xCD\x01\x00\x26\x1D\x00\x80\x60\x09"
-     "\x00\x80"
-     "\x2A\x1D\x00\x80\x8C\xD1\x01\x00\x81\x1D\x00\x00\x6E\x07\x00\x80\x63\x09"
-     "\x00\x80"
-     "\x69\x09\x00\x80\x2E\x1D\x00\x80\x32\x1D\x00\x80\x66\x09\x00\x80\x3A\x1D"
-     "\x00\x80"
-     "\x36\x1D\x00\x80\x81\x25\x00\x00\x88\x07\x00\x80\x3E\x1D\x00\x80\x42\x1D"
-     "\x00\x80"
-     "\xA7\x21\x01\x00\x46\x1D\x00\x80\x81\x11\x00\x00\x4D\x0A\x00\x80\xDA\x07"
-     "\x00\x80"
-     "\xA3\x69\x01\x00\x6F\x09\x00\x80\xE5\x08\x00\x80\x4A\x1D\x00\x80\xA7\x65"
-     "\x01\x00"
-     "\x75\x09\x00\x80\x4E\x1D\x00\x80\x72\x09\x00\x80\x56\x1D\x00\x80\x52\x1D"
-     "\x00\x80"
-     "\xF4\x07\x00\x80\xAE\x45\x01\x00\xE8\x08\x00\x80\x80\x45\x00\x00\x81\x35"
-     "\x00\x00"
-     "\x83\xDD\x02\x00\x82\xD1\x02\x00\x81\xC1\x02\x00\x80\xC9\x02\x00\x87\xE1"
-     "\x02\x00"
-     "\x86\xF9\x02\x00\xEB\x08\x00\x80\x84\xFD\x02\x00\x62\x1E\x00\x80\xC1\x08"
-     "\x00\x80"
-     "\x89\x2D\x03\x00\x88\xD5\x02\x00\x8F\x3D\x03\x00\x5E\x1E\x00\x80\x6A\x1E"
-     "\x00\x80"
-     "\x8C\x35\x03\x00\x81\xE9\x03\x00\x66\x1E\x00\x80\x91\x51\x03\x00\x81\x2D"
-     "\x03\x00"
-     "\x5A\x1D\x00\x80\x8C\xDD\x03\x00\x87\xF5\x03\x00\x81\x0D\x00\x00\x81\xE5"
-     "\x03\x00"
-     "\x98\x0A\x00\x80\xCD\x10\x01\x00\xCC\x24\x02\x00\x5E\x1D\x00\x80\xCC\x2C"
-     "\x01\x00"
-     "\x8C\x1D\x03\x00\x62\x1D\x00\x80\xA3\xA1\x03\x00\x66\x1D\x00\x80\x7E\x09"
-     "\x00\x80"
-     "\xB1\xF9\x03\x00\xCD\x2C\x01\x00\xCC\xD0\x02\x00\x80\xB9\x00\x00\x81\xA9"
-     "\x00\x00"
-     "\xB1\x3D\x03\x00\x9E\x0A\x00\x80\x84\x0D\x00\x00\xA8\x81\x03\x00\xCC\xD4"
-     "\x02\x00"
-     "\xCD\x10\x01\x00\x95\x0A\x00\x80\xCC\x54\x07\x00\x84\x79\x00\x00\xA3\xA9"
-     "\x03\x00"
-     "\xA3\x7D\x03\x00\xB0\xF5\x03\x00\xAD\x49\x03\x00\xA7\xA5\x03\x00\xCD\x60"
-     "\x03\x00"
-     "\xCC\x00\x0C\x00\x81\x65\x00\x00\x80\x99\x00\x00\xA3\x65\x03\x00\x5A\x1E"
-     "\x00\x80"
-     "\xCC\x98\x03\x00\xCD\x0C\x01\x00\xA7\x6D\x03\x00\x81\x09\x00\x80\xCC\x30"
-     "\x06\x00"
-     "\xCD\x34\x06\x00\xCC\xFC\x07\x00\x17\x0A\x00\x80\xCD\xE0\x06\x00\xCC\x6C"
-     "\x01\x00"
-     "\xCC\xF4\x07\x00\xCD\x04\x06\x00\xCC\xB0\x07\x00\xCD\xB4\x07\x00\x6B\x0A"
-     "\x00\x80"
-     "\xCC\x70\x01\x00\xCC\x50\x07\x00\xCC\x48\x07\x00\xCC\x28\x01\x00\xCD\x88"
-     "\x07\x00"
-     "\x84\x25\x01\x00\x6E\x0A\x00\x80\x80\x19\x00\x00\x81\x11\x00\x00\x76\x1E"
-     "\x00\x80"
-     "\x81\x05\x01\x00\x8C\x81\x02\x00\x72\x1E\x00\x80\xCC\x54\x06\x00\x83\xA9"
-     "\x02\x00"
-     "\x80\xD5\x02\x00\x81\xA5\x02\x00\x88\x01\x01\x00\x87\xA5\x02\x00\x81\xE5"
-     "\x00\x00"
-     "\xCC\x70\x06\x00\x7E\x1E\x00\x80\x7A\x1E\x00\x80\xCD\xC4\x00\x00\xCC\x80"
-     "\x03\x00"
-     "\x82\x1E\x00\x80\x6E\x1E\x00\x80\x41\x0A\x00\x80\x6A\x1D\x00\x80\x84\x39"
-     "\x00\x00"
-     "\xCC\xEC\x00\x00\x82\x3D\x00\x00\xB1\x45\x02\x00\xCC\x9C\x00\x00\xCD\x8C"
-     "\x06\x00"
-     "\x80\x4D\x02\x00\x81\xA5\x01\x00\x6E\x1D\x00\x80\x24\x08\x00\x80\xCC\x30"
-     "\x01\x00"
-     "\xCD\x34\x01\x00\x81\xE5\x01\x00\x80\xF5\x01\x00\x83\xE1\x01\x00\xAD\x39"
-     "\x02\x00"
-     "\x87\x21\x00\x00\xA3\x39\x02\x00\x87\x09\x00\x80\x72\x1D\x00\x80\x89\xB5"
-     "\x01\x00"
-     "\xA7\x15\x02\x00\x81\x65\x00\x00\x8D\x09\x00\x80\xEE\x08\x00\x80\x71\x0A"
-     "\x00\x80"
-     "\xF1\x08\x00\x80\x89\x89\x0F\x00\x7B\x09\x00\x80\x74\x0A\x00\x80\x9B\x0A"
-     "\x00\x80"
-     "\xF4\x08\x00\x80\xE4\x09\x00\x80\x83\xA5\x0F\x00\x80\x79\x0F\x00\x81\x71"
-     "\x0F\x00"
-     "\x56\x1E\x00\x80\x3E\x0A\x00\x80\xF9\x07\x00\x80\x84\x79\x00\x00\x26\x08"
-     "\x00\x80"
-     "\x7A\x1D\x00\x80\x76\x1D\x00\x80\x8A\x09\x00\x80\xA3\x61\x00\x00\x7E\x1D"
-     "\x00\x80"
-     "\x81\xB1\x0F\x00\x40\x08\x00\x80\x90\x09\x00\x80\x87\x0D\x00\x00\x8A\x1D"
-     "\x00\x80"
-     "\x93\x09\x00\x80\x82\x1D\x00\x80\x99\x09\x00\x80\x87\xF9\x00\x00\xA8\x49"
-     "\x00\x00"
-     "\x8E\x1D\x00\x80\x86\x1D\x00\x80\x87\x05\x00\x00\x92\x1D\x00\x80\x7D\x0A"
-     "\x00\x80"
-     "\x96\x09\x00\x80\x91\xDD\x0F\x00\x81\xFD\x00\x00\x54\x08\x00\x80\xA3\x35"
-     "\x0F\x00"
-     "\x96\x1D\x00\x80\x9F\x09\x00\x80\x9B\x91\x00\x00\x9A\x1D\x00\x80\x80\x0A"
-     "\x00\x80"
-     "\x9C\x09\x00\x80\x87\xD9\x00\x00\x86\x4D\x0E\x00\xA2\x1D\x00\x80\x84\xDD"
-     "\x00\x00"
-     "\x83\xD9\x00\x00\x82\xD5\x00\x00\x81\xC1\x0E\x00\x80\x29\x0D\x00\x8F\xC9"
-     "\x0F\x00"
-     "\x9E\x1D\x00\x80\x61\x08\x00\x80\x8C\xE1\x0F\x00\x8B\x0D\x0E\x00\xFA\x08"
-     "\x00\x80"
-     "\x89\x15\x0E\x00\x88\xE1\x00\x00\x81\xE9\x0E\x00\x82\x69\x01\x00\x81\x69"
-     "\x01\x00"
-     "\xFD\x08\x00\x80\x87\x61\x01\x00\x00\x09\x00\x80\x87\xF5\x0E\x00\x81\x3D"
-     "\x00\x00"
-     "\xB2\x1D\x00\x80\xB6\x1D\x00\x80\xA8\x09\x00\x80\xAB\x09\x00\x80\x88\x15"
-     "\x00\x00"
-     "\x8C\xC9\x0E\x00\x8F\x3D\x01\x00\x8C\x49\x01\x00\x91\x55\x01\x00\xA5\x09"
-     "\x00\x80"
-     "\x03\x09\x00\x80\x84\x29\x00\x00\x26\x0A\x00\x80\x87\xFD\x01\x00\x81\x39"
-     "\x00\x00"
-     "\x80\x0D\x00\x00\x20\x0A\x00\x80\x1D\x0A\x00\x80\x23\x0A\x00\x80\x8C\x05"
-     "\x00\x00"
-     "\x80\x21\x00\x00\x81\x15\x00\x00\x8C\xD5\x01\x00\x8A\x1E\x00\x80\xA3\x05"
-     "\x01\x00"
-     "\xAE\x09\x00\x80\xA3\xC1\x01\x00\xB1\x69\x01\x00\xA7\x05\x01\x00\xA6\x05"
-     "\x01\x00"
-     "\xA7\xD5\x01\x00\xB1\x09\x00\x80\x1A\x0A\x00\x80\x77\x0A\x00\x80\x7A\x0A"
-     "\x00\x80"
-     "\x92\x1E\x00\x80\x8E\x1E\x00\x80\x9A\x1E\x00\x80\xF6\x09\x00\x80\x96\x1E"
-     "\x00\x80"
-     "\xB1\x95\x01\x00\xA3\x71\x01\x00\x9E\x1E\x00\x80\x86\x1E\x00\x80\xA6\x69"
-     "\x01\x00"
-     "\xA7\x69\x01\x00\x80\x75\x00\x00\x81\x6D\x00\x00\xA6\x1D\x00\x80\xB4\x09"
-     "\x00\x80"
-     "\xAE\x1D\x00\x80\xAA\x1D\x00\x80\x80\x05\x01\x00\x81\xF9\x00\x00\x63\x08"
-     "\x00\x80"
-     "\xAD\x35\x01\x00\x83\x45\x03\x00\x82\x11\x03\x00\x81\x05\x03\x00\x80\xF9"
-     "\x02\x00"
-     "\xBA\x1D\x00\x80\x86\x39\x03\x00\xCC\x08\x01\x00\x84\x35\x03\x00\x8B\xBD"
-     "\x03\x00"
-     "\x8A\xA1\x03\x00\x89\x55\x03\x00\x88\x65\x03\x00\x8F\x95\x03\x00\xBE\x1D"
-     "\x00\x80"
-     "\xC2\x1D\x00\x80\x8C\x9D\x03\x00\xC6\x1D\x00\x80\xB7\x09\x00\x80\x91\xE1"
-     "\x03\x00"
-     "\xCE\x1D\x00\x80\xCA\x1D\x00\x80\xD2\x1D\x00\x80\x88\xDD\x00\x00\xA3\x95"
-     "\x03\x00"
-     "\x9B\xD9\x03\x00\x87\x09\x00\x00\x83\x79\x00\x00\xD6\x1D\x00\x80\xCD\x68"
-     "\x01\x00"
-     "\xCC\x6C\x01\x00\xCC\x24\x01\x00\xCD\x0C\x01\x00\xA3\x4D\x03\x00\xDA\x1D"
-     "\x00\x80"
-     "\xCC\x58\x02\x00\xCD\x54\x01\x00\xA2\x1E\x00\x80\x94\xE1\x1F\x00\x81\x7D"
-     "\x00\x00"
-     "\xA4\x7D\x03\x00\x81\xE1\x00\x00\x80\x9D\x00\x00\x06\x09\x00\x80\xA8\x49"
-     "\x03\x00"
-     "\xBA\x09\x00\x80\xAE\x1E\x00\x80\xAD\x51\x03\x00\xA1\x0A\x00\x80\xDE\x1D"
-     "\x00\x80"
-     "\x0C\x09\x00\x80\xCC\xAC\x02\x00\xB0\x21\x03\x00\x80\x7D\x03\x00\x81\x65"
-     "\x03\x00"
-     "\xAA\x1E\x00\x80\x83\x69\x03\x00\xA6\x1E\x00\x80\xBD\x09\x00\x80\xA3\xDD"
-     "\x00\x00"
-     "\xE2\x1D\x00\x80\xC0\x09\x00\x80\x89\x49\x03\x00\x80\x69\x00\x00\x81\x55"
-     "\x00\x00"
-     "\xCC\xA0\x0E\x00\xCC\xB4\x0E\x00\xCC\xF4\x0E\x00\xCD\x70\x0E\x00\xCC\x38"
-     "\x01\x00"
-     "\xCD\x5C\x0E\x00\xCC\x64\x0B\x00\xC3\x09\x00\x80\xCC\xC4\x0F\x00\xCD\xC0"
-     "\x0F\x00"
-     "\xCC\x90\x0F\x00\xCC\xFC\x0F\x00\xCC\xE4\x0E\x00\xCD\xAC\x0F\x00\xCC\x68"
-     "\x0E\x00"
-     "\xCD\x64\x0F\x00\x87\x0D\x01\x00\xCC\x20\x01\x00\x81\xC5\x07\x00\xE6\x1D"
-     "\x00\x80"
-     "\xCC\x28\x01\x00\x82\x11\x01\x00\x81\x19\x01\x00\x80\x65\x01\x00\xEA\x1D"
-     "\x00\x80"
-     "\xEE\x1D\x00\x80\xCD\xF0\x01\x00\xCC\x60\x00\x00\xCC\x74\x00\x00\xCD\x28"
-     "\x01\x00"
-     "\xCC\x54\x01\x00\x88\x2D\x01\x00\xCC\x8C\x0B\x00\xCD\x0C\x0B\x00\x82\xD1"
-     "\x0E\x00"
-     "\x47\x0A\x00\x80\xCC\xD8\x00\x00\xF2\x1D\x00\x80\xF6\x1D\x00\x80\xFA\x1D"
-     "\x00\x80"
-     "\x1B\x00\x00\x80\x0B\x00\x00\x80\xCC\xB8\x0D\x00\xCD\x18\x0B\x00\x8C\x3D"
-     "\x00\x00"
-     "\xCC\x20\x01\x00\xCC\xE4\x0D\x00\x83\x04\x00\x80\xCC\xB8\x0D\x00\xCD\x34"
-     "\x0A\x00"
-     "\x22\x05\x00\x80\xCC\xF8\x0A\x00\xA3\x85\x01\x00\x03\x00\x00\x80\xCC\xD4"
-     "\x03\x00"
-     "\xCD\x2C\x00\x00\xCD\xD8\x0A\x00\xCC\xC8\x0C\x00\xCC\xA4\x05\x00\xCC\x6C"
-     "\x02\x00"
-     "\xCC\xF0\x04\x00\xCD\xB8\x0C\x00\xCD\x80\x0C\x00\xCC\xB4\x0F\x00\x82\x08"
-     "\x00\x80"
-     "\xCC\x0C\x03\x00\x8B\x08\x00\x80\xCC\x68\x04\x00\x88\xD5\x06\x00\x89\xCD"
-     "\x06\x00"
-     "\x8D\x08\x00\x80\xCC\x64\x04\x00\x84\x9D\x06\x00\x12\x0C\x00\x80\x65\x00"
-     "\x00\x80"
-     "\x87\xED\x06\x00\x80\xA1\x06\x00\x81\x9D\x06\x00\x82\x89\x06\x00\x83\x89"
-     "\x06\x00"
-     "\x81\xA5\x06\x00\x82\xD9\x08\x00\x60\x03\x00\x80\x82\xA1\x06\x00\x87\xF1"
-     "\x08\x00"
-     "\x50\x04\x00\x80\x87\xA9\x06\x00\x44\x04\x00\x80\xCC\x04\x08\x00\xCD\xC0"
-     "\x0B\x00"
-     "\x54\x05\x00\x80\x88\xC9\x08\x00\xB8\x08\x00\x80\x8C\x69\x07\x00\xBE\x08"
-     "\x00\x80"
-     "\x8C\xD5\x08\x00\xB1\xE1\x07\x00\xC1\x08\x00\x80\xB1\xED\x07\x00\x53\x0A"
-     "\x00\x80"
-     "\xCC\xD4\x08\x00\xCD\x28\x0B\x00\xCC\xF0\x09\x00\xCD\x64\x0B\x00\xCC\x7C"
-     "\x06\x00"
-     "\xCD\x1C\x09\x00\xCC\x54\x09\x00\xCD\xD0\x09\x00\xCC\xB4\x07\x00\xCD\xE8"
-     "\x06\x00"
-     "\x32\x0A\x00\x80\xA3\x5D\x06\x00\xA3\x6D\x08\x00\xC7\x08\x00\x80\xA3\x85"
-     "\x07\x00"
-     "\xD0\x08\x00\x80\xA7\x61\x08\x00\xDF\x08\x00\x80\xCC\xD4\x07\x00\xCD\x44"
-     "\x06\x00"
-     "\xCC\xEC\x04\x00\xCD\x14\x07\x00\xCC\xA0\x03\x00\xCD\x48\x04\x00\xE2\x08"
-     "\x00\x80"
-     "\xAE\x41\x08\x00\xCC\x0C\x03\x00\xCD\x04\x04\x00\x80\xC5\x09\x00\x81\xCD"
-     "\x09\x00"
-     "\x82\xDD\x09\x00\x83\xD1\x09\x00\x84\xF1\x09\x00\x87\xBD\x05\x00\x86\xF5"
-     "\x09\x00"
-     "\xC4\x08\x00\x80\x88\xD9\x09\x00\x89\x21\x08\x00\x88\xBD\x06\x00\xEE\x08"
-     "\x00\x80"
-     "\x8C\x39\x08\x00\xF1\x08\x00\x80\x8C\x95\x05\x00\x8F\x31\x08\x00\x82\x05"
-     "\x07\x00"
-     "\x91\x5D\x08\x00\x80\x71\x07\x00\x81\x0D\x07\x00\xF4\x08\x00\x80\x87\x19"
-     "\x07\x00"
-     "\xCC\x1C\x03\x00\xCD\xC0\x05\x00\x8A\xF9\x06\x00\x81\xE9\x01\x00\x88\x39"
-     "\x07\x00"
-     "\xF7\x08\x00\x80\x84\xC9\x01\x00\x81\x09\x00\x00\xA3\x99\x07\x00\x56\x0A"
-     "\x00\x80"
-     "\x8C\xC1\x02\x00\xA3\x31\x05\x00\x03\x09\x00\x80\xA3\xAD\x08\x00\xA6\x29"
-     "\x05\x00"
-     "\xA7\x29\x05\x00\x09\x09\x00\x80\x12\x09\x00\x80\xA8\x8D\x08\x00\x27\x09"
-     "\x00\x80"
-     "\xCC\xC8\x02\x00\xCD\xE0\x04\x00\xCC\xC0\x02\x00\xCD\xC4\x02\x00\x2A\x09"
-     "\x00\x80"
-     "\xAD\x75\x05\x00\xB0\xF9\x08\x00\xA3\x91\x07\x00\x48\x09\x00\x80\xB1\x29"
-     "\x05\x00"
-     "\x89\x09\x07\x00\x88\x11\x07\x00\x77\x08\x00\x80\x8A\x41\x06\x00\x27\x07"
-     "\x00\x80"
-     "\x84\x59\x07\x00\x87\x29\x07\x00\x84\x07\x00\x80\x81\x59\x07\x00\x80\x65"
-     "\x07\x00"
-     "\x83\x4D\x07\x00\x82\x4D\x07\x00\x6C\x09\x00\x80\xB5\x0B\x00\x80\x8F\x0D"
-     "\x0E\x00"
-     "\x78\x09\x00\x80\x7B\x09\x00\x80\x84\x09\x00\x80\x90\x09\x00\x80\x99\x09"
-     "\x00\x80"
-     "\xA2\x09\x00\x80\xA5\x09\x00\x80\xCD\x64\x07\x00\xCC\x4C\x01\x00\xCC\x80"
-     "\x01\x00"
-     "\xA8\x09\x00\x80\xCB\x0A\x00\x80\xC9\x09\x00\x80\xCC\x09\x00\x80\x85\xB5"
-     "\x1F\x00"
-     "\x9F\x4D\x0E\x00\x85\xB5\x1F\x00\xCE\x0A\x00\x80\xCF\x09\x00\x80\xCD\x28"
-     "\x01\x00"
-     "\xCC\x2C\x01\x00\xD4\x0A\x00\x80\xD7\x0A\x00\x80\x85\x6D\x1C\x00\xD2\x09"
-     "\x00\x80"
-     "\x29\x0A\x00\x80\xDD\x0A\x00\x80\xE0\x0A\x00\x80\xD5\x09\x00\x80\x85\x91"
-     "\x1C\x00"
-     "\xEC\x0A\x00\x80\xAF\x2D\x05\x00\xF5\x0A\x00\x80\x07\x0B\x00\x80\xA8\x35"
-     "\x05\x00"
-     "\x81\x6D\x15\x00\xAA\xDD\x0E\x00\x04\x0B\x00\x80\xDB\x09\x00\x80\x19\x0B"
-     "\x00\x80"
-     "\x22\x0B\x00\x80\x2B\x0B\x00\x80\xA0\xBD\x0E\x00\xE4\x09\x00\x80\xE7\x09"
-     "\x00\x80"
-     "\xBD\x9D\x02\x00\xBC\x91\x02\x00\xEA\x09\x00\x80\xBE\xA5\x02\x00\xB9\x99"
-     "\x02\x00"
-     "\xB8\xFD\x02\x00\x46\x0B\x00\x80\xBA\x31\x0B\x00\xB5\xC5\x02\x00\xB4\x31"
-     "\x05\x00"
-     "\xCC\xB4\x07\x00\xCD\x48\x06\x00\xED\x09\x00\x80\x4F\x0B\x00\x80\xB3\x31"
-     "\x05\x00"
-     "\xB2\x41\x05\x00\x85\x25\x1D\x00\xF3\x09\x00\x80\x83\xF5\x16\x00\x58\x0B"
-     "\x00\x80"
-     "\xF6\x09\x00\x80\x5E\x0B\x00\x80\xA6\x80\x01\x00\x61\x0B\x00\x80\xF9\x09"
-     "\x00\x80"
-     "\xFC\x09\x00\x80\xCC\x6C\x01\x00\xCD\xB0\x07\x00\xCC\x48\x07\x00\x85\x79"
-     "\x1D\x00"
-     "\xCC\x08\x01\x00\xCD\xC4\x07\x00\x44\x0A\x00\x80\xFF\x09\x00\x80\x85\x19"
-     "\x1C\x00"
-     "\x02\x0A\x00\x80\x85\x09\x1C\x00\xCE\x06\x00\x80\x70\x07\x00\x80\xE2\x07"
-     "\x00\x80"
-     "\xB8\xE0\x01\x00\x35\x0A\x00\x80\x3E\x0A\x00\x80\xCB\x06\x00\x80\x85\x11"
-     "\x1C\x00"
-     "\xD8\x09\x00\x80\x05\x0A\x00\x80\x83\x0A\x00\x80\x80\xB1\x0F\x00\x81\xB1"
-     "\x0F\x00"
-     "\x82\xF9\x03\x00\x83\xD9\x0F\x00\x84\xE9\x03\x00\x85\x11\x03\x00\x86\x85"
-     "\x02\x00"
-     "\x87\x59\x03\x00\x88\x79\x0D\x00\x89\x79\x0D\x00\x8A\x65\x03\x00\x8B\xFD"
-     "\x0C\x00"
-     "\x8C\x75\x0B\x00\x8D\x7D\x0B\x00\x8E\x6D\x0B\x00\x8F\x69\x03\x00\x90\x31"
-     "\x03\x00"
-     "\x91\x09\x0A\x00\x92\x3D\x0A\x00\x93\xD5\x09\x00\x94\xFD\x02\x00\x95\xE1"
-     "\x02\x00"
-     "\x96\x89\x02\x00\xE1\x09\x00\x80\x98\x8D\x02\x00\x99\x11\x07\x00\x9A\xE5"
-     "\x06\x00"
-     "\x9B\xF9\x06\x00\x9C\x79\x02\x00\x9D\x5D\x05\x00\x9E\x61\x02\x00\x89\x0A"
-     "\x00\x80"
-     "\x80\xED\x1D\x00\x81\x05\x1E\x00\x76\x0B\x00\x80\x83\xC5\x15\x00\x8C\x0A"
-     "\x00\x80"
-     "\x85\xF5\x15\x00\x8F\x0A\x00\x80\x87\xC5\x15\x00\x92\x0A\x00\x80\x89\x21"
-     "\x1E\x00"
-     "\x8A\x25\x1E\x00\x9B\x0A\x00\x80\x8C\x21\x1E\x00\xA2\x22\x00\x80\x8E\xB1"
-     "\x16\x00"
-     "\xA7\x0A\x00\x80\x90\xBD\x16\x00\xDE\x09\x00\x80\x92\x5D\x1E\x00\xA6\x22"
-     "\x00\x80"
-     "\x94\xB9\x17\x00\xB0\x0A\x00\x80\x96\x15\x18\x00\xC6\x09\x00\x80\x98\x9D"
-     "\x17\x00"
-     "\x5B\x0B\x00\x80\x9A\x71\x1E\x00\x64\x0B\x00\x80\x9C\x1D\x18\x00\xB3\x0A"
-     "\x00\x80"
-     "\x9E\xED\x18\x00\xB6\x0A\x00\x80\xA0\x9D\x1E\x00\xB9\x0A\x00\x80\xA2\x35"
-     "\x18\x00"
-     "\xBC\x0A\x00\x80\xA4\x3D\x18\x00\xBF\x0A\x00\x80\xA6\xB9\x1E\x00\x09\x07"
-     "\x00\x80"
-     "\xA8\x85\x1E\x00\xFE\x0A\x00\x80\xAA\x9D\x1E\x00\x36\x07\x00\x80\xAC\x9D"
-     "\x19\x00"
-     "\x37\x0B\x00\x80\xAE\x3D\x19\x00\x3D\x0B\x00\x80\xB0\x5D\x19\x00\x49\x0B"
-     "\x00\x80"
-     "\xB2\x2D\x19\x00\x13\x00\x00\x80\xB4\x95\x1A\x00\x17\x00\x00\x80\xB6\x69"
-     "\x1A\x00"
-     "\x31\x00\x00\x80\xB8\x25\x1A\x00\xB9\x01\x1A\x00\x69\x00\x00\x80\xBB\x1D"
-     "\x1A\x00"
-     "\x07\x00\x00\x80\xBD\xE9\x1B\x00\x35\x00\x00\x80\xBF\xA1\x18\x00\x85\x7D"
-     "\x02\x00"
-     "\xA4\x99\x15\x00\x79\x0B\x00\x80\xA6\xC5\x1C\x00\xBF\x0B\x00\x80\xA0\x91"
-     "\x15\x00"
-     "\xC2\x0B\x00\x80\xA2\x8D\x15\x00\xCA\x0B\x00\x80\xAC\xC9\x15\x00\xC5\x0B"
-     "\x00\x80"
-     "\xAE\x01\x14\x00\x09\x0C\x00\x80\xA8\xE9\x15\x00\xCD\x0B\x00\x80\xAA\xDD"
-     "\x15\x00"
-     "\xD0\x0B\x00\x80\xB4\x1D\x14\x00\xB2\x0B\x00\x80\xB6\x3D\x14\x00\xD3\x0B"
-     "\x00\x80"
-     "\x75\x00\x00\x80\xB3\xB9\x1C\x00\xB2\xBD\x1C\x00\xBD\xE9\x14\x00\x3E\x06"
-     "\x00\x80"
-     "\xBF\x85\x1C\x00\x8F\x0C\x00\x80\xB9\xF5\x14\x00\x25\x00\x00\x80\xBB\xED"
-     "\x14\x00"
-     "\xE2\x0B\x00\x80\xAC\x08\x00\x80\x84\xE9\x13\x00\xE5\x0B\x00\x80\x86\x1D"
-     "\x13\x00"
-     "\xE8\x0B\x00\x80\x80\xFD\x11\x00\x77\x08\x00\x80\x82\x01\x1C\x00\xEB\x0B"
-     "\x00\x80"
-     "\x8C\x01\x13\x00\x5F\x0C\x00\x80\x8E\x6D\x13\x00\x62\x0C\x00\x80\x88\x29"
-     "\x13\x00"
-     "\x65\x0C\x00\x80\x8A\x1D\x13\x00\x8E\x1F\x00\x80\x94\x8D\x12\x00\x8B\x08"
-     "\x00\x80"
-     "\x96\x81\x12\x00\x0E\x20\x00\x80\x90\x71\x1C\x00\xF1\x0B\x00\x80\x92\x65"
-     "\x1C\x00"
-     "\x0A\x20\x00\x80\x9C\xC9\x15\x00\xC6\x1E\x00\x80\x9E\xB9\x15\x00\xB2\x1F"
-     "\x00\x80"
-     "\x98\x71\x12\x00\xBA\x1F\x00\x80\x9A\x9D\x12\x00\x81\x39\x1C\x00\xC2\x1F"
-     "\x00\x80"
-     "\xAA\x1F\x00\x80\x82\x31\x1C\x00\xB6\x1F\x00\x80\x84\x55\x1C\x00\x87\x4D"
-     "\x1C\x00"
-     "\x86\x4D\x1C\x00\x89\x41\x1C\x00\xBE\x1F\x00\x80\x8B\x45\x1C\x00\x8A\x49"
-     "\x1C\x00"
-     "\x80\xD9\x03\x00\x81\x51\x06\x00\x8F\x4D\x1C\x00\x8E\x4D\x1C\x00\x91\x3D"
-     "\x1C\x00"
-     "\x90\x3D\x1C\x00\x93\xCD\x1C\x00\xC6\x1F\x00\x80\xAE\x1F\x00\x80\x94\xD9"
-     "\x1C\x00"
-     "\x97\xF9\x1C\x00\x96\xCD\x1C\x00\x82\xA5\x06\x00\x98\xC1\x1C\x00\x80\x81"
-     "\x03\x00"
-     "\x81\x09\x06\x00\x9D\xDD\x1C\x00\x9C\xDD\x1C\x00\x9F\xC9\x1C\x00\x85\x75"
-     "\x00\x00"
-     "\x16\x20\x00\x80\xA0\x3D\x1C\x00\xF4\x0B\x00\x80\xA2\x31\x1C\x00\xF7\x0B"
-     "\x00\x80"
-     "\xA4\x5D\x1C\x00\xA7\x69\x1C\x00\xA6\x45\x1C\x00\xA9\x55\x1C\x00\xFA\x0B"
-     "\x00\x80"
-     "\xCC\xE0\x1D\x00\xCD\x08\x1A\x00\x92\x1F\x00\x80\xAC\xB1\x1D\x00\xAF\xA9"
-     "\x1D\x00"
-     "\xAE\xB1\x1D\x00\xB1\xD5\x1D\x00\xE6\x1E\x00\x80\xB3\xE9\x1D\x00\xB2\xD1"
-     "\x1D\x00"
-     "\xB5\xF5\x1D\x00\xEE\x1E\x00\x80\xB7\xD1\x02\x00\xCD\x8C\x1E\x00\xEA\x1E"
-     "\x00\x80"
-     "\xB8\xA1\x1D\x00\xCC\xC8\x06\x00\xCD\x0C\x04\x00\xF2\x1E\x00\x80\xBC\xB9"
-     "\x1D\x00"
-     "\xCC\x0C\x1D\x00\xCD\x04\x1A\x00\x81\xD5\x07\x00\x80\xD5\x07\x00\x81\xDD"
-     "\x01\x00"
-     "\x80\xC9\x01\x00\x84\x45\x1C\x00\x85\x41\x1C\x00\x86\x4D\x1C\x00\x87\x45"
-     "\x1C\x00"
-     "\x88\x79\x1C\x00\x89\x7D\x1C\x00\x8A\x75\x1C\x00\x8B\x71\x1C\x00\x8C\x6D"
-     "\x1C\x00"
-     "\x8D\x01\x11\x00\xCC\x88\x1C\x00\x8F\x35\x16\x00\x00\x0C\x00\x80\x91\x49"
-     "\x14\x00"
-     "\x96\x1F\x00\x80\x93\xED\x1A\x00\xD2\x1F\x00\x80\x95\x31\x1A\x00\xDA\x1F"
-     "\x00\x80"
-     "\x97\x3D\x1A\x00\xE2\x1F\x00\x80\x99\x0D\x1A\x00\xCA\x1F\x00\x80\x9B\x3D"
-     "\x1A\x00"
-     "\x80\x91\x01\x00\x81\x85\x01\x00\x9E\x81\x12\x00\xD6\x1F\x00\x80\xA0\xFD"
-     "\x12\x00"
-     "\xDE\x1F\x00\x80\xA2\x31\x1F\x00\xE6\x1F\x00\x80\xA4\xF5\x1C\x00\x88\x99"
-     "\x0E\x00"
-     "\xA6\x11\x17\x00\xCC\x1C\x1A\x00\xA8\xC5\x16\x00\xCE\x1F\x00\x80\xAA\xC1"
-     "\x1C\x00"
-     "\xCC\xE0\x1D\x00\xAC\x29\x14\x00\x2A\x20\x00\x80\xAE\xF9\x1C\x00\x06\x0C"
-     "\x00\x80"
-     "\x85\x45\x00\x00\xB1\xA9\x1C\x00\xB2\xA5\x1C\x00\xB3\xA1\x1C\x00\xB4\x2D"
-     "\x10\x00"
-     "\x9A\x1F\x00\x80\xB6\xAD\x1C\x00\xB7\xA5\x1C\x00\xB8\x5D\x16\x00\x3E\x20"
-     "\x00\x80"
-     "\xBA\xC9\x11\x00\x3A\x20\x00\x80\xBC\x7D\x1F\x00\x59\x0C\x00\x80\xBE\x85"
-     "\x1E\x00"
-     "\x1E\x1F\x00\x80\x81\x55\x04\x00\x92\x04\x19\x00\x83\x55\x04\x00\x85\x81"
-     "\x07\x00"
-     "\x85\x7D\x04\x00\x84\x71\x04\x00\x85\x95\x02\x00\x86\x69\x04\x00\x85\x89"
-     "\x02\x00"
-     "\x88\x59\x04\x00\xBE\x1E\x00\x80\x8A\xA9\x04\x00\x26\x1F\x00\x80\x8C\xB9"
-     "\x04\x00"
-     "\x2E\x1F\x00\x80\x8E\xA9\x04\x00\x9C\xB1\x04\x00\x9D\xBD\x04\x00\x80\xB5"
-     "\x00\x00"
-     "\x81\x4D\x01\x00\x98\x9D\x04\x00\x99\x81\x04\x00\x9A\x85\x04\x00\x9B\xB5"
-     "\x04\x00"
-     "\x82\xBD\x00\x00\xCE\x1E\x00\x80\x80\xB9\x00\x00\x81\xB9\x00\x00\x81\xF9"
-     "\x05\x00"
-     "\x80\xC5\x06\x00\x83\x61\x01\x00\x85\x2D\x06\x00\xB3\xC1\x1C\x00\xB2\x29"
-     "\x04\x00"
-     "\xB1\x2D\x04\x00\xB0\x11\x04\x00\xB7\x09\x05\x00\xB6\x51\x04\x00\xB5\x25"
-     "\x04\x00"
-     "\xB4\x29\x04\x00\xA4\xA9\x05\x00\x22\x1F\x00\x80\x2A\x1F\x00\x80\xB8\x65"
-     "\x05\x00"
-     "\xA0\x3D\x1A\x00\xA1\x35\x11\x00\xA2\xE5\x1C\x00\xA3\xBD\x18\x00\x32\x1F"
-     "\x00\x80"
-     "\xD6\x1E\x00\x80\xC2\x1E\x00\x80\x0C\x0C\x00\x80\xCA\x1E\x00\x80\xD2\x1E"
-     "\x00\x80"
-     "\x0F\x0C\x00\x80\xDA\x1E\x00\x80\x80\x21\x05\x00\x81\x99\x05\x00\xFD\x0B"
-     "\x00\x80"
-     "\x83\xC5\x00\x00\x35\x0C\x00\x80\x85\x75\x06\x00\x38\x0C\x00\x80\xFE\x1E"
-     "\x00\x80"
-     "\x06\x1F\x00\x80\xA8\x25\x05\x00\x0E\x1F\x00\x80\xAA\xB5\x13\x00\x85\xA9"
-     "\x03\x00"
-     "\xAC\x71\x13\x00\xDE\x1E\x00\x80\xAE\x31\x05\x00\x9E\x1F\x00\x80\xA0\xF9"
-     "\x1A\x00"
-     "\x3E\x1F\x00\x80\xA2\xD1\x1A\x00\x46\x1F\x00\x80\xA4\x25\x05\x00\x42\x1F"
-     "\x00\x80"
-     "\xA6\x25\x05\x00\x80\x39\x00\x00\x81\xCD\x00\x00\xBB\x75\x05\x00\xBA\x79"
-     "\x05\x00"
-     "\xBD\x15\x05\x00\xE2\x1E\x00\x80\x4A\x1F\x00\x80\xBE\x11\x05\x00\x3E\x0C"
-     "\x00\x80"
-     "\xB0\xC1\x13\x00\x03\x0C\x00\x80\xB2\xD1\x1F\x00\xF6\x1E\x00\x80\x5E\x20"
-     "\x00\x80"
-     "\x02\x1F\x00\x80\x44\x0C\x00\x80\xFA\x1E\x00\x80\x88\x89\x10\x00\x0A\x1F"
-     "\x00\x80"
-     "\x8A\x95\x10\x00\x47\x0C\x00\x80\x8C\xBD\x12\x00\x12\x1F\x00\x80\x8E\xD9"
-     "\x12\x00"
-     "\xCC\x64\x1D\x00\x80\x9D\x17\x00\x09\x0C\x00\x80\x82\xED\x17\x00\xA2\x1F"
-     "\x00\x80"
-     "\x84\xD1\x16\x00\x52\x20\x00\x80\x86\xD5\x16\x00\x4E\x20\x00\x80\x98\x55"
-     "\x1D\x00"
-     "\x5C\x0C\x00\x80\x9A\x75\x1D\x00\x16\x1F\x00\x80\x9C\xA9\x1A\x00\x5A\x1F"
-     "\x00\x80"
-     "\x9E\x75\x11\x00\x62\x1F\x00\x80\x90\xE1\x1D\x00\x6A\x1F\x00\x80\x92\xF9"
-     "\x1D\x00"
-     "\x4D\x0C\x00\x80\x94\x51\x1C\x00\x50\x0C\x00\x80\x96\x55\x1C\x00\x80\x11"
-     "\x07\x00"
-     "\x81\x1D\x07\x00\x53\x0C\x00\x80\x83\x2D\x07\x00\x84\x39\x07\x00\x85\x25"
-     "\x07\x00"
-     "\x80\xA9\x04\x00\x81\x95\x07\x00\x80\xE1\x07\x00\x81\x59\x07\x00\xCC\x24"
-     "\x07\x00"
-     "\xCD\x20\x07\x00\x84\xE1\x07\x00\xCD\x70\x1B\x00\x86\xD9\x07\x00\xCD\x58"
-     "\x1B\x00"
-     "\xCC\x08\x00\x00\xCD\x38\x02\x00\x94\x1D\x03\x00\xCD\x58\x05\x00\xA6\x1F"
-     "\x00\x80"
-     "\x93\x4D\x03\x00\xCC\xF8\x02\x00\xCD\x00\x0C\x00\xCD\x10\x07\x00\xCC\x14"
-     "\x07\x00"
-     "\x85\x01\x05\x00\x93\x45\x02\x00\x94\xB9\x02\x00\x82\x7D\x03\x00\x81\x79"
-     "\x03\x00"
-     "\x80\x81\x02\x00\x80\xF9\x03\x00\x81\x0D\x03\x00\x82\x11\x03\x00\xF2\x1F"
-     "\x00\x80"
-     "\x41\x0C\x00\x80\x85\x31\x05\x00\xCC\xB8\x0C\x00\xCD\xDC\x0C\x00\x80\xE9"
-     "\x05\x00"
-     "\x81\x35\x07\x00\xFA\x1F\x00\x80\xCD\xB4\x00\x00\xCC\xF0\x01\x00\xCD\x34"
-     "\x02\x00"
-     "\x02\x20\x00\x80\xCD\x68\x07\x00\xB0\xF9\x07\x00\xEA\x1F\x00\x80\xB2\x89"
-     "\x07\x00"
-     "\xF6\x1F\x00\x80\xB4\x99\x07\x00\xFE\x1F\x00\x80\xB6\x89\x07\x00\x06\x20"
-     "\x00\x80"
-     "\xCC\xF4\x03\x00\xCD\x44\x03\x00\xBA\xBD\x07\x00\x93\x9D\x03\x00\x94\x71"
-     "\x03\x00"
-     "\x85\x39\x00\x00\xBE\x9D\x07\x00\xBF\x91\x07\x00\x84\xD5\x1C\x00\x85\x71"
-     "\x06\x00"
-     "\x86\x75\x06\x00\x87\x8D\x06\x00\x81\x21\x03\x00\x80\xE9\x04\x00\xCC\x28"
-     "\x07\x00"
-     "\xCD\x74\x07\x00\x8C\x89\x06\x00\xEE\x1F\x00\x80\x8E\x99\x06\x00\x8F\x95"
-     "\x06\x00"
-     "\x88\xB9\x06\x00\x89\xA5\x06\x00\x8A\x99\x06\x00\x6E\x20\x00\x80\x94\xE5"
-     "\x05\x00"
-     "\x95\xAD\x03\x00\x96\x09\x04\x00\x97\x25\x02\x00\xCD\x08\x1B\x00\x91\x8D"
-     "\x06\x00"
-     "\x92\xD1\x05\x00\x93\xCD\x05\x00\x9C\x45\x1B\x00\x9D\x95\x03\x00\x9E\x91"
-     "\x03\x00"
-     "\x9F\x71\x0F\x00\x98\xCD\x04\x00\x99\xDD\x02\x00\x9A\xB5\x03\x00\x9B\xA5"
-     "\x03\x00"
-     "\xA4\x11\x02\x00\xA5\xD1\x0F\x00\xA6\x79\x02\x00\xA7\x75\x02\x00\xA0\xF1"
-     "\x02\x00"
-     "\xA1\x05\x01\x00\x85\x4D\x01\x00\xA3\xE1\x02\x00\x80\x5D\x04\x00\x81\x95"
-     "\x03\x00"
-     "\x82\x91\x03\x00\x12\x20\x00\x80\xA8\xB9\x01\x00\xA9\xDD\x0F\x00\xAA\x4D"
-     "\x04\x00"
-     "\xAB\x75\x02\x00\x85\x1D\x01\x00\xB5\x8C\x0D\x00\xCD\x94\x1A\x00\xB7\x70"
-     "\x0D\x00"
-     "\x85\x7D\x00\x00\xB1\xD8\x0D\x00\xCC\xC0\x03\x00\xCD\xA8\x01\x00\xCD\xF4"
-     "\x1A\x00"
-     "\xCD\x70\x1B\x00\xCD\xC8\x1A\x00\xBF\xCC\x0E\x00\xCD\x8C\x1B\x00\xB9\x04"
-     "\x0D\x00"
-     "\xCC\xFC\x06\x00\xCD\x94\x01\x00\x88\xF5\x0B\x00\x1A\x1F\x00\x80\x80\xAD"
-     "\x0C\x00"
-     "\x81\xA5\x0C\x00\x80\xB5\x0C\x00\x81\xBD\x0C\x00\x82\xD9\x0C\x00\x3B\x0C"
-     "\x00\x80"
-     "\x80\xF5\x0B\x00\x81\x85\x09\x00\x80\x09\x0D\x00\x81\x85\x09\x00\x84\xD5"
-     "\x0B\x00"
-     "\x36\x1F\x00\x80\x86\xD5\x0B\x00\x3A\x1F\x00\x80\x80\x95\x0B\x00\x81\xE5"
-     "\x09\x00"
-     "\x85\xFD\x0B\x00\x83\x9D\x0D\x00\x83\x21\x0D\x00\x4A\x0C\x00\x80\x81\xC9"
-     "\x09\x00"
-     "\x80\x15\x0B\x00\x52\x1F\x00\x80\x93\x51\x00\x00\x56\x0C\x00\x80\x93\xBD"
-     "\x0D\x00"
-     "\x94\x75\x00\x00\x6E\x1F\x00\x80\x94\x45\x00\x00\x76\x1F\x00\x80\x80\xA9"
-     "\x0D\x00"
-     "\x81\xB5\x0D\x00\x80\xB1\x0D\x00\x81\xD5\x0D\x00\x80\xB1\x0D\x00\x81\x3D"
-     "\x09\x00"
-     "\x88\x51\x0D\x00\x83\xA9\x00\x00\x86\x61\x0D\x00\x7E\x1F\x00\x80\x84\x11"
-     "\x0D\x00"
-     "\x86\x1F\x00\x80\x80\x69\x0D\x00\x81\x55\x0D\x00\x80\x09\x0D\x00\x81\x0D"
-     "\x09\x00"
-     "\x80\x19\x0D\x00\x81\x25\x0D\x00\x82\x21\x0D\x00\x72\x1F\x00\x80\x80\x61"
-     "\x0D\x00"
-     "\x81\x65\x09\x00\x80\xA5\x00\x00\x81\x79\x00\x00\x81\x45\x00\x00\x80\x99"
-     "\x00\x00"
-     "\x94\x89\x00\x00\x82\x41\x00\x00\x85\x35\x0E\x00\x7A\x1F\x00\x80\x80\x0D"
-     "\x00\x00"
-     "\x81\xF5\x01\x00\x80\x51\x0F\x00\x81\xBD\x08\x00\x82\x1F\x00\x80\x85\x29"
-     "\x0E\x00"
-     "\x82\x05\x00\x00\x8A\x1F\x00\x80\x80\xED\x01\x00\x81\x15\x00\x00\x80\x71"
-     "\x0F\x00"
-     "\x81\x9D\x08\x00\x8E\x1F\x00\x80\x83\x59\x01\x00\xE2\x0B\x00\x80\x85\x09"
-     "\x0E\x00"
-     "\xCD\x58\x02\x00\xCC\xBC\x02\x00\x85\x05\x00\x00\x1A\x20\x00\x80\x1E\x20"
-     "\x00\x80"
-     "\x93\x39\x01\x00\x94\x9D\x01\x00\x85\x0D\x00\x00\x22\x20\x00\x80\x96\x1F"
-     "\x00\x80"
-     "\x80\x49\x00\x00\x81\x55\x00\x00\x82\x21\x00\x00\x83\x45\x0D\x00\x85\xE5"
-     "\x00\x00"
-     "\x85\xF5\x0B\x00\xCC\x60\x0D\x00\xCD\x08\x0F\x00\x80\xA9\x00\x00\x81\xB5"
-     "\x00\x00"
-     "\xCD\xC8\x01\x00\xCC\x2C\x02\x00\x84\x41\x08\x00\x26\x20\x00\x80\x86\x79"
-     "\x08\x00"
-     "\xE8\x0B\x00\x80\x80\xFD\x00\x00\x81\x39\x00\x00\x85\x05\x00\x00\x2E\x20"
-     "\x00\x80"
-     "\x32\x20\x00\x80\x36\x20\x00\x80\x85\x0D\x00\x00\x59\x0C\x00\x80\xCC\xE8"
-     "\x01\x00"
-     "\xCD\x54\x01\x00\x42\x20\x00\x80\x93\xE5\x0D\x00\x94\x19\x0D\x00\xDF\x0B"
-     "\x00\x80"
-     "\x80\x11\x00\x00\x81\x09\x00\x00\x80\x29\x00\x00\x81\x31\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\x93\x9D\x0C\x00\x94\x71\x0C\x00\x4A\x20\x00\x80\x46\x20\x00\x80\x5A\x20"
-     "\x00\x80"
-     "\x56\x20\x00\x80\xCD\x50\x01\x00\x88\xB5\x03\x00\x85\xB9\x00\x00\xCC\x38"
-     "\x00\x00"
-     "\xCD\x84\x00\x00\xEE\x0B\x00\x80\xCD\xD8\x0F\x00\xCD\xF4\x0F\x00\x66\x20"
-     "\x00\x80"
-     "\x80\xFD\x09\x00\x81\x8D\x0B\x00\x86\xDD\x09\x00\x56\x1F\x00\x80\x84\xDD"
-     "\x09\x00"
-     "\xCD\xB0\x0F\x00\x80\x1D\x03\x00\x81\xD9\x03\x00\x82\xF9\x03\x00\x83\x3D"
-     "\x0F\x00"
-     "\x5E\x1F\x00\x80\x85\xE1\x09\x00\xCC\x78\x02\x00\xCD\x58\x17\x00\x66\x1F"
-     "\x00\x80"
-     "\x93\xB5\x0F\x00\xCD\x0C\x01\x00\x93\x91\x03\x00\x94\x05\x0F\x00\x62\x20"
-     "\x00\x80"
-     "\x94\x7D\x02\x00\xCD\x3C\x01\x00\x85\x45\x00\x00\xCD\x28\x01\x00\xCD\x68"
-     "\x01\x00"
-     "\xCC\x54\x01\x00\x80\x45\x09\x00\x81\x35\x0B\x00\x82\x61\x03\x00\x83\x4D"
-     "\x0F\x00"
-     "\xCC\x40\x02\x00\xCD\xD4\x03\x00\xCD\x14\x01\x00\x93\xA5\x06\x00\x94\xD1"
-     "\x02\x00"
-     "\x6A\x20\x00\x80\x80\xB9\x00\x00\x81\xA5\x00\x00\x88\xC1\x06\x00\x72\x20"
-     "\x00\x80"
-     "\x80\xFD\x00\x00\x81\x9D\x00\x00\x80\xD1\x00\x00\x81\xCD\x00\x00\x82\xFD"
-     "\x00\x00"
-     "\x7A\x20\x00\x80\x80\x51\x0F\x00\x81\x55\x0B\x00\xCC\x84\x03\x00\xCD\x6C"
-     "\x03\x00"
-     "\x84\x49\x0F\x00\x68\x0C\x00\x80\x86\x39\x0F\x00\x85\x21\x00\x00\xCD\x80"
-     "\x01\x00"
-     "\xCC\x84\x01\x00\xCD\x80\x01\x00\xCC\x84\x01\x00\x80\x25\x01\x00\x81\x45"
-     "\x01\x00"
-     "\x82\x79\x01\x00\x6B\x0C\x00\x80\x93\x55\x03\x00\x6E\x0C\x00\x80\xCD\xB8"
-     "\x01\x00"
-     "\xCC\x04\x01\x00\xCD\x20\x01\x00\xCC\x3C\x01\x00\x71\x0C\x00\x80\x94\xF1"
-     "\x03\x00"
-     "\xCD\xA4\x01\x00\xCC\x9C\x01\x00\xCD\x44\x0D\x00\x74\x0C\x00\x80\xCC\xE0"
-     "\x01\x00"
-     "\xCD\x94\x01\x00\xCC\x58\x00\x00\xCD\x7C\x00\x00\xCD\xB0\x01\x00\xCC\xDC"
-     "\x01\x00"
-     "\x81\x7D\x07\x00\x80\x7D\x07\x00\x80\xC1\x0E\x00\x81\xC5\x0A\x00\x82\x0D"
-     "\x01\x00"
-     "\x83\x75\x07\x00\x80\x91\x07\x00\x81\x91\x07\x00\x82\x91\x07\x00\x83\xF9"
-     "\x03\x00"
-     "\xCD\xA4\x0D\x00\x85\xA9\x0C\x00\xCD\x9C\x0C\x00\x77\x0C\x00\x80\x80\x49"
-     "\x00\x00"
-     "\x7A\x0C\x00\x80\x7D\x0C\x00\x80\x88\x15\x06\x00\x80\x0C\x00\x80\x83\x0C"
-     "\x00\x80"
-     "\x86\x0C\x00\x80\x81\x89\x06\x00\xCD\x20\x0F\x00\x89\x0C\x00\x80\x8C\x0C"
-     "\x00\x80"
-     "\x92\x0C\x00\x80\x95\x0C\x00\x80\xD6\x0B\x00\x80\x86\xB9\x06\x00\xD9\x0B"
-     "\x00\x80"
-     "\x88\x95\x06\x00\xDC\x0B\x00\x80\xF8\x0C\x00\x80\xFB\x0C\x00\x80\xFE\x0C"
-     "\x00\x80"
-     "\x01\x0D\x00\x80\xCD\xE0\x00\x00\xCC\x04\x01\x00\x5C\x0C\x00\x80\xB8\xD4"
-     "\x01\x00"
-     "\x4E\x1F\x00\x80\xBA\xD4\x01\x00\xA6\x1F\x00\x80\x65\x0C\x00\x80\x76\x20"
-     "\x00\x80"
-     "\xBE\xF0\x01\x00\x41\x0C\x00\x80\xB0\xD4\x01\x00\xB3\xD0\x01\x00\x04\x0D"
-     "\x00\x80"
-     "\xB5\xD0\x01\x00\x07\x0D\x00\x80\xB7\xD0\x01\x00\xB6\xC8\x01\x00\xA9\xC9"
-     "\x02\x00"
-     "\xA8\xC5\x02\x00\xAB\x45\x3C\x00\xAA\xC1\x02\x00\xAD\x41\x3C\x00\xAC\xC9"
-     "\x02\x00"
-     "\xAF\xED\x02\x00\xAE\xED\x02\x00\xA1\x25\x02\x00\xA0\x5D\x02\x00\xA3\x6D"
-     "\x3C\x00"
-     "\xA2\x25\x02\x00\xA5\x01\x02\x00\xA4\x11\x02\x00\xA7\x71\x3C\x00\xA6\x09"
-     "\x02\x00"
-     "\x99\x81\x02\x00\x98\x49\x01\x00\x9B\x99\x02\x00\x9A\x01\x3C\x00\x9D\xB1"
-     "\x02\x00"
-     "\x9C\x8D\x02\x00\x9F\xA9\x02\x00\x9E\x9D\x3C\x00\x91\x21\x01\x00\x90\xFD"
-     "\x01\x00"
-     "\x93\x25\x01\x00\x92\x25\x01\x00\x95\x25\x01\x00\x94\x25\x01\x00\x97\x0D"
-     "\x01\x00"
-     "\x96\x1D\x01\x00\x89\xDD\x01\x00\x88\xD1\x01\x00\x8B\xE5\x01\x00\x8A\xE9"
-     "\x01\x00"
-     "\x8D\xED\x01\x00\x8C\xE1\x01\x00\x8F\x8D\x01\x00\x8E\x95\x01\x00\x81\x0D"
-     "\x07\x00"
-     "\x80\x0D\x07\x00\x83\x1D\x07\x00\x82\x1D\x07\x00\x85\xC5\x00\x00\x84\xD9"
-     "\x00\x00"
-     "\x87\xF5\x01\x00\x86\xF9\x01\x00\xAE\xC9\x06\x00\x0A\x0D\x00\x80\xAC\xCD"
-     "\x06\x00"
-     "\x0D\x0D\x00\x80\xAA\xA9\x06\x00\x10\x0D\x00\x80\xA8\xAD\x06\x00\x13\x0D"
-     "\x00\x80"
-     "\xA6\x91\x06\x00\x16\x0D\x00\x80\xA4\x81\x06\x00\x19\x0D\x00\x80\xA2\x95"
-     "\x06\x00"
-     "\x1C\x0D\x00\x80\xA0\xA5\x06\x00\x1F\x0D\x00\x80\xBE\x9D\x06\x00\xBF\xA1"
-     "\x06\x00"
-     "\x22\x0D\x00\x80\xBD\x99\x06\x00\xBA\x81\x06\x00\x25\x0D\x00\x80\x98\x0C"
-     "\x00\x80"
-     "\xB9\x21\x08\x00\xC4\x0D\x00\x80\xB7\xA9\x06\x00\xB4\xD5\x0E\x00\xB5\x85"
-     "\x09\x00"
-     "\xB2\x0D\x08\x00\xC7\x0D\x00\x80\xB0\xB5\x09\x00\xB1\x5D\x01\x00\x9B\x0C"
-     "\x00\x80"
-     "\x8F\x71\x07\x00\x9E\x0C\x00\x80\xA1\x0C\x00\x80\xA4\x0C\x00\x80\xA7\x0C"
-     "\x00\x80"
-     "\xCD\x0D\x00\x80\x89\x44\x04\x00\xAA\x0C\x00\x80\xAC\x0D\x00\x80\xD6\x0D"
-     "\x00\x80"
-     "\x85\x3C\x04\x00\x82\xCD\x08\x00\xAD\x0C\x00\x80\xD9\x0D\x00\x80\x81\xD4"
-     "\x04\x00"
-     "\x9E\x8D\x06\x00\x80\x2D\x06\x00\x9C\xF1\x06\x00\xB0\x0C\x00\x80\x9A\xF5"
-     "\x06\x00"
-     "\xDF\x0D\x00\x80\x98\xC5\x06\x00\x86\x45\x00\x00\x96\x81\x08\x00\x88\x09"
-     "\x06\x00"
-     "\x94\x21\x08\x00\x95\x1D\x08\x00\x92\xA5\x05\x00\x93\x45\x0C\x00\x90\x71"
-     "\x0F\x00"
-     "\x91\x41\x0E\x00\xB3\x0C\x00\x80\x86\xA1\x03\x00\xE2\x0D\x00\x80\x80\x31"
-     "\x03\x00"
-     "\xB9\x0C\x00\x80\x86\x15\x00\x00\xBC\x0C\x00\x80\x84\xC9\x03\x00\xBF\x0C"
-     "\x00\x80"
-     "\x80\x19\x03\x00\xE5\x0D\x00\x80\x88\x25\x00\x00\xC5\x0C\x00\x80\xCC\x80"
-     "\x0F\x00"
-     "\xC8\x0C\x00\x80\x88\x95\x03\x00\x88\x69\x00\x00\xCB\x0C\x00\x80\xCE\x0C"
-     "\x00\x80"
-     "\xF1\x0D\x00\x80\x84\x1D\x00\x00\xF4\x0D\x00\x80\x86\x11\x03\x00\xF7\x0D"
-     "\x00\x80"
-     "\x88\x35\x00\x00\xD4\x0C\x00\x80\xCC\xA8\x07\x00\x8B\x31\x00\x00\xD7\x0C"
-     "\x00\x80"
-     "\xDA\x0C\x00\x80\x88\x05\x00\x00\xFA\x0D\x00\x80\xE0\x0C\x00\x80\xCC\x08"
-     "\x07\x00"
-     "\xE3\x0C\x00\x80\xE6\x0C\x00\x80\xCC\x8C\x02\x00\x88\x0D\x00\x00\x00\x0E"
-     "\x00\x80"
-     "\xEC\x0C\x00\x80\xCD\x00\x07\x00\xCC\x60\x01\x00\x88\x05\x00\x00\xEE\x0D"
-     "\x00\x80"
-     "\xF2\x0C\x00\x80\xF5\x0C\x00\x80\xCD\x24\x12\x00\xCC\x20\x12\x00\xCC\x00"
-     "\x07\x00"
-     "\xCC\xD8\x02\x00\xCC\x54\x12\x00\xCD\x50\x12\x00\xCC\x14\x01\x00\xCD\x74"
-     "\x07\x00"
-     "\xCC\x54\x07\x00\xCC\xBC\x02\x00\xCC\xEC\x02\x00\xCD\xC0\x02\x00\xCC\x50"
-     "\x07\x00"
-     "\xCD\x54\x07\x00\xCC\xB0\x02\x00\xCC\xAC\x02\x00\xCD\xDC\x02\x00\xCC\xF0"
-     "\x02\x00"
-     "\xCC\xB4\x03\x00\x28\x0D\x00\x80\xCC\x80\x03\x00\xCC\xB8\x0E\x00\xCC\x30"
-     "\x07\x00"
-     "\xCD\x10\x12\x00\x2B\x0D\x00\x80\x96\x20\x03\x00\xCD\xB4\x07\x00\xCD\x3C"
-     "\x07\x00"
-     "\x2E\x0D\x00\x80\xCD\xC4\x07\x00\xCC\x88\x07\x00\xCD\xE4\x07\x00\xCC\xF0"
-     "\x07\x00"
-     "\xCD\x2C\x06\x00\x31\x0D\x00\x80\x34\x0D\x00\x80\x83\x88\x02\x00\x37\x0D"
-     "\x00\x80"
-     "\x3A\x0D\x00\x80\x3D\x0D\x00\x80\x87\x9C\x02\x00\x40\x0D\x00\x80\x43\x0D"
-     "\x00\x80"
-     "\x49\x0D\x00\x80\x8B\x58\x03\x00\x8F\x09\x00\x00\x8D\x48\x03\x00\xCC\x24"
-     "\x02\x00"
-     "\x4C\x0D\x00\x80\x4F\x0D\x00\x80\x52\x0D\x00\x80\xB0\xD9\x00\x00\x55\x0D"
-     "\x00\x80"
-     "\xB2\xED\x00\x00\xB5\xD4\x03\x00\xB4\x81\x3F\x00\x58\x0D\x00\x80\xB6\xF9"
-     "\x00\x00"
-     "\x5B\x0D\x00\x80\xB8\xDD\x00\x00\x5E\x0D\x00\x80\xBA\xD1\x00\x00\x61\x0D"
-     "\x00\x80"
-     "\xBC\x35\x00\x00\xD9\x38\x3E\x00\xBE\x35\x00\x00\x64\x0D\x00\x80\xA0\xC1"
-     "\x00\x00"
-     "\x67\x0D\x00\x80\xA2\xD5\x00\x00\x6A\x0D\x00\x80\xA4\xFD\x00\x00\x6D\x0D"
-     "\x00\x80"
-     "\xA6\x9D\x00\x00\x70\x0D\x00\x80\xA8\xA1\x00\x00\x73\x0D\x00\x80\xAA\xA1"
-     "\x00\x00"
-     "\x76\x0D\x00\x80\xAC\xA5\x00\x00\x79\x0D\x00\x80\xAE\xA9\x00\x00\xC2\x0C"
-     "\x00\x80"
-     "\xD1\x0C\x00\x80\xDD\x0C\x00\x80\xE9\x0C\x00\x80\xEF\x0C\x00\x80\x46\x0D"
-     "\x00\x80"
-     "\x80\x81\x3F\x00\x7C\x0D\x00\x80\x8E\x91\x3F\x00\x7F\x0D\x00\x80\x8C\x81"
-     "\x3F\x00"
-     "\x82\x0D\x00\x80\x8A\x91\x3F\x00\x85\x0D\x00\x80\x88\x0D\x00\x80\x8B\x0D"
-     "\x00\x80"
-     "\x96\x99\x3F\x00\x8E\x0D\x00\x80\x94\x99\x3F\x00\x91\x0D\x00\x80\x92\x99"
-     "\x3F\x00"
-     "\x94\x0D\x00\x80\x90\x99\x3F\x00\x97\x0D\x00\x80\x9E\x99\x3F\x00\x9A\x0D"
-     "\x00\x80"
-     "\x9C\x99\x3F\x00\x9D\x0D\x00\x80\x9A\x99\x3F\x00\xA0\x0D\x00\x80\x98\x99"
-     "\x3F\x00"
-     "\xA3\x0D\x00\x80\xA6\xB5\x00\x00\xA6\x0D\x00\x80\xA4\xA5\x00\x00\xB6\x0C"
-     "\x00\x80"
-     "\xA2\xB5\x00\x00\xA9\x0D\x00\x80\xA0\x85\x00\x00\xC1\x0D\x00\x80\xAE\xED"
-     "\x00\x00"
-     "\xAF\x0D\x00\x80\xAC\xED\x00\x00\xB2\x0D\x00\x80\xAA\x8D\x00\x00\xB5\x0D"
-     "\x00\x80"
-     "\xA8\x81\x00\x00\xB8\x0D\x00\x80\xB6\xAD\x00\x00\xBB\x0D\x00\x80\xB4\xAD"
-     "\x00\x00"
-     "\xBE\x0D\x00\x80\xB2\x8D\x00\x00\xCA\x0D\x00\x80\xB0\x8D\x00\x00\xD0\x0D"
-     "\x00\x80"
-     "\xBE\x75\x00\x00\x88\x09\x00\x00\xBC\x6D\x00\x00\xD3\x0D\x00\x80\xBA\x8D"
-     "\x00\x00"
-     "\xCC\x30\x02\x00\xB8\x8D\x00\x00\xDC\x0D\x00\x80\x80\x9D\x01\x00\x81\xA1"
-     "\x3C\x00"
-     "\xE8\x0D\x00\x80\x83\xA9\x01\x00\x88\x09\x00\x00\x85\xB9\x01\x00\xEB\x0D"
-     "\x00\x80"
-     "\x87\xA9\x01\x00\xCC\x10\x02\x00\x89\x99\x01\x00\xFD\x0D\x00\x80\x8B\xE9"
-     "\x01\x00"
-     "\x03\x0E\x00\x80\x8D\xF9\x01\x00\x06\x0E\x00\x80\x09\x0E\x00\x80\x90\x21"
-     "\x3C\x00"
-     "\x0C\x0E\x00\x80\x92\x3D\x3C\x00\x0F\x0E\x00\x80\x94\x85\x01\x00\x12\x0E"
-     "\x00\x80"
-     "\x96\xC1\x3C\x00\x15\x0E\x00\x80\x98\xBD\x01\x00\x99\xA0\x02\x00\x9A\x85"
-     "\x01\x00"
-     "\x18\x0E\x00\x80\x9C\xD9\x3C\x00\x1B\x0E\x00\x80\x9E\xC1\x3C\x00\x1E\x0E"
-     "\x00\x80"
-     "\xA0\x7D\x01\x00\x21\x0E\x00\x80\xA2\x5D\x3C\x00\x24\x0E\x00\x80\xA4\x99"
-     "\x3F\x00"
-     "\x27\x0E\x00\x80\xA6\xB1\x3F\x00\x2A\x0E\x00\x80\xA8\xA9\x00\x00\xA9\xB4"
-     "\x03\x00"
-     "\xAA\xB1\x00\x00\x2D\x0E\x00\x80\xAC\x1D\x3F\x00\x30\x0E\x00\x80\xAE\xF5"
-     "\x3F\x00"
-     "\x33\x0E\x00\x80\xB0\x95\x3F\x00\x36\x0E\x00\x80\xB2\x95\x3F\x00\x39\x0E"
-     "\x00\x80"
-     "\xB4\xAD\x3F\x00\x3C\x0E\x00\x80\xB6\xF1\x00\x00\x3F\x0E\x00\x80\xB8\x79"
-     "\x3F\x00"
-     "\x42\x0E\x00\x80\xBA\xD9\x00\x00\x45\x0E\x00\x80\xBC\xC9\x00\x00\x48\x0E"
-     "\x00\x80"
-     "\xBE\xC5\x00\x00\x4B\x0E\x00\x80\xBE\xBD\x00\x00\xBF\xC9\x00\x00\xBC\xA5"
-     "\x00\x00"
-     "\xBD\xB9\x00\x00\xBA\x8D\x00\x00\xBB\xB1\x00\x00\xB8\x95\x00\x00\xB9\x89"
-     "\x00\x00"
-     "\xB6\x5D\x00\x00\xB7\xA1\x00\x00\xB4\x45\x00\x00\xB5\x59\x00\x00\xB2\x6D"
-     "\x00\x00"
-     "\xB3\x51\x00\x00\xCC\x08\x01\x00\xB1\x69\x00\x00\xAE\x0D\x00\x00\x4E\x0E"
-     "\x00\x80"
-     "\xAC\x0D\x00\x00\x51\x0E\x00\x80\xAA\xAD\x03\x00\x54\x0E\x00\x80\xA8\xAD"
-     "\x03\x00"
-     "\x57\x0E\x00\x80\xA6\x8D\x03\x00\x5A\x0E\x00\x80\xA4\x8D\x03\x00\x5D\x0E"
-     "\x00\x80"
-     "\xA2\xAD\x03\x00\x60\x0E\x00\x80\xA0\xAD\x03\x00\x68\x0E\x00\x80\x9E\x4D"
-     "\x03\x00"
-     "\x6B\x0E\x00\x80\x9C\x4D\x03\x00\x6E\x0E\x00\x80\x9A\x2D\x03\x00\x71\x0E"
-     "\x00\x80"
-     "\x98\x2D\x03\x00\x74\x0E\x00\x80\x96\x0D\x03\x00\x77\x0E\x00\x80\x94\x35"
-     "\x03\x00"
-     "\x7A\x0E\x00\x80\x92\x25\x03\x00\x7D\x0E\x00\x80\x90\x35\x03\x00\x80\x0E"
-     "\x00\x80"
-     "\x8E\x45\x03\x00\x87\x05\x01\x00\x8C\xB5\x03\x00\x83\x0E\x00\x80\x8A\xA5"
-     "\x03\x00"
-     "\x86\x0E\x00\x80\x88\xB5\x03\x00\x89\x0E\x00\x80\x86\xBD\x03\x00\x8C\x0E"
-     "\x00\x80"
-     "\x84\xBD\x03\x00\x8F\x0E\x00\x80\x82\xBD\x03\x00\x92\x0E\x00\x80\x80\xBD"
-     "\x03\x00"
-     "\xA9\x0E\x00\x80\x90\x9D\x00\x00\x91\xA5\x00\x00\x92\xAD\x00\x00\x93\xA5"
-     "\x00\x00"
-     "\x94\xBD\x00\x00\x95\xA5\x00\x00\x96\xAD\x00\x00\xAC\x0E\x00\x80\xCC\xA8"
-     "\x3F\x00"
-     "\xCC\x74\x3E\x00\xAF\x0E\x00\x80\xB2\x0E\x00\x80\xCC\x50\x3F\x00\xCC\x74"
-     "\x3F\x00"
-     "\xCC\x60\x3F\x00\xCC\x88\x3F\x00\x80\x95\x01\x00\x81\x99\x01\x00\x82\xAD"
-     "\x01\x00"
-     "\x83\xA1\x01\x00\x84\xA5\x01\x00\x85\xA9\x01\x00\x86\x45\x00\x00\x87\x75"
-     "\x00\x00"
-     "\x88\x4D\x00\x00\x89\xE5\x00\x00\x8A\xED\x00\x00\x8B\xE5\x00\x00\x8C\xFD"
-     "\x00\x00"
-     "\x8D\xE5\x00\x00\x8E\xED\x00\x00\x8F\xE5\x00\x00\xB5\x0E\x00\x80\xB8\x0E"
-     "\x00\x80"
-     "\xBB\x0E\x00\x80\xBE\x0E\x00\x80\xC1\x0E\x00\x80\xC4\x0E\x00\x80\xCC\x0E"
-     "\x00\x80"
-     "\xCF\x0E\x00\x80\xD2\x0E\x00\x80\xD5\x0E\x00\x80\xD8\x0E\x00\x80\xDB\x0E"
-     "\x00\x80"
-     "\xDE\x0E\x00\x80\xE1\x0E\x00\x80\xCD\x00\x3A\x00\xCC\x04\x3A\x00\x63\x0E"
-     "\x00\x80"
-     "\xA7\x18\x3F\x00\xBB\x0F\x00\x80\x93\x05\x00\x00\x94\x09\x00\x00\x95\x09"
-     "\x00\x00"
-     "\xBE\x0F\x00\x80\xC4\x0F\x00\x80\x94\x09\x00\x00\xD9\x54\x02\x00\xC1\x0F"
-     "\x00\x80"
-     "\xCC\x28\x01\x00\x94\x09\x00\x00\xD9\x54\x02\x00\xC7\x0F\x00\x80\x39\x10"
-     "\x00\x80"
-     "\x5D\x16\x00\x80\x2C\x17\x00\x80\xBB\x16\x00\x80\x81\x78\x03\x00\xD9\x54"
-     "\x02\x00"
-     "\x94\x0D\x00\x00\x20\x17\x00\x80\xD9\x6C\x02\x00\x94\x05\x00\x00\x43\x17"
-     "\x00\x80"
-     "\xD9\x54\x02\x00\x94\x1D\x00\x00\x1A\x17\x00\x80\xA8\x8C\x03\x00\xE0\x94"
-     "\x02\x00"
-     "\xA4\x88\x02\x00\x92\x38\x03\x00\xBC\x0D\x00\x00\x70\x17\x00\x80\xE0\x8C"
-     "\x02\x00"
-     "\xA4\xC0\x02\x00\xB0\x98\x03\x00\xB3\xB0\x03\x00\x95\x10\x03\x00\xBC\x05"
-     "\x00\x00"
-     "\x82\x17\x00\x80\xE0\x94\x02\x00\xA4\xEC\x02\x00\xBC\x05\x00\x00\x86\x17"
-     "\x00\x80"
-     "\x46\x17\x00\x80\x4D\x17\x00\x80\xA4\x14\x02\x00\xA5\xA4\x03\x00\x54\x17"
-     "\x00\x80"
-     "\xA7\x94\x03\x00\x5B\x17\x00\x80\x62\x17\x00\x80\x69\x17\x00\x80\x74\x17"
-     "\x00\x80"
-     "\x7B\x17\x00\x80\xAD\x3C\x03\x00\xAE\x50\x03\x00\xAF\x68\x03\x00\xA8\x4C"
-     "\x03\x00"
-     "\xA9\x50\x03\x00\xE0\x28\x03\x00\x9F\x17\x00\x80\xBE\x7D\x00\x00\xB5\x84"
-     "\x00\x00"
-     "\xA3\x17\x00\x80\xB7\x30\x00\x00\x97\x19\x00\x00\xB1\xC8\x00\x00\xB2\x3C"
-     "\x00\x00"
-     "\xB3\xE8\x00\x00\xBC\x30\x00\x00\xBD\xEC\x01\x00\xBE\xD4\x01\x00\x8A\x17"
-     "\x00\x80"
-     "\xB8\xA8\x00\x00\x91\x17\x00\x80\xBA\xF4\x00\x00\xBB\x44\x00\x00\xB8\x09"
-     "\x01\x00"
-     "\x87\x3C\x02\x00\xA6\x30\x03\x00\xA7\x80\x03\x00\x98\x17\x00\x80\xC3\x17"
-     "\x00\x80"
-     "\xCA\x17\x00\x80\xA7\x17\x00\x80\xAE\x17\x00\x80\xB6\x3D\x00\x00\xB5\x17"
-     "\x00\x80"
-     "\xBC\x17\x00\x80\xB3\x25\x00\x00\xB5\xC9\x06\x00\xB6\x39\x01\x00\xB7\x39"
-     "\x01\x00"
-     "\x98\x39\x01\x00\x99\x39\x01\x00\x9A\xC1\x01\x00\x9B\xDD\x01\x00\x9C\xC5"
-     "\x01\x00"
-     "\x9D\xCD\x01\x00\x9E\xC5\x01\x00\x9F\xFD\x01\x00\xE0\xA0\x02\x00\x9F\x75"
-     "\x00\x00"
-     "\x9C\x85\x01\x00\x9D\x99\x01\x00\xAC\xF0\x02\x00\xAD\xBC\x02\x00\xBE\x05"
-     "\x00\x00"
-     "\xE3\x17\x00\x80\x99\x9D\x00\x00\x87\x64\x03\x00\x9B\xA1\x00\x00\x9A\xAD"
-     "\x00\x00"
-     "\x96\x09\x00\x00\x97\x35\x00\x00\xDF\x17\x00\x80\x9E\xB1\x00\x00\xE7\x17"
-     "\x00\x80"
-     "\xD1\x17\x00\x80\xD8\x17\x00\x80\xE0\xA4\x02\x00\x92\x9C\x03\x00\xAF\x38"
-     "\x02\x00"
-     "\xEB\x17\x00\x80\xE0\xE4\x02\x00\x88\x40\x05\x00\xBE\x09\x00\x00\x8A\x58"
-     "\x05\x00"
-     "\xEF\x17\x00\x80\x97\x69\x00\x00\x97\x25\x00\x00\x87\xAC\x03\x00\x86\xE0"
-     "\x03\x00"
-     "\xAE\xDC\x02\x00\xAF\x6C\x02\x00\x9C\x4D\x00\x00\x9D\x45\x00\x00\xF7\x17"
-     "\x00\x80"
-     "\xE0\xB4\x02\x00\xAE\xFC\x02\x00\xBE\xFD\x03\x00\xF3\x17\x00\x80\x86\x8C"
-     "\x03\x00"
-     "\xE0\xDC\x02\x00\x96\x19\x00\x00\xB1\x44\x02\x00\xFB\x17\x00\x80\xBF\x84"
-     "\x03\x00"
-     "\xE0\xF4\x02\x00\xFF\x17\x00\x80\xB3\x44\x02\x00\x82\x01\x0C\x00\x86\x9C"
-     "\x03\x00"
-     "\xE0\xC0\x02\x00\x95\x15\x00\x00\x95\x71\x00\x00\xB3\x1C\x02\x00\xB3\x2C"
-     "\x02\x00"
-     "\xE0\xF4\x02\x00\x0F\x18\x00\x80\x03\x18\x00\x80\x07\x18\x00\x80\xBE\x3D"
-     "\x00\x00"
-     "\x86\x9C\x03\x00\x87\xA0\x03\x00\xB4\xCC\x02\x00\xB5\x48\x02\x00\x13\x18"
-     "\x00\x80"
-     "\x1B\x18\x00\x80\x97\x1D\x00\x00\x95\x39\x00\x00\x96\x29\x00\x00\xE0\x2C"
-     "\x02\x00"
-     "\xB4\xF0\x02\x00\x17\x18\x00\x80\xBE\x0D\x00\x00\xE0\xC0\x02\x00\xB7\x10"
-     "\x02\x00"
-     "\xB3\x25\x00\x00\x99\x94\x03\x00\x1F\x18\x00\x80\x27\x18\x00\x80\x2B\x18"
-     "\x00\x80"
-     "\x8A\x35\x00\x00\x8A\x0D\x00\x00\x2F\x18\x00\x80\xB7\x30\x02\x00\x44\x18"
-     "\x00\x80"
-     "\x8F\xBD\x0E\x00\xB3\x79\x00\x00\x36\x18\x00\x80\x9C\x0D\x00\x00\x9D\x1D"
-     "\x00\x00"
-     "\x3D\x18\x00\x80\x4B\x18\x00\x80\x56\x18\x00\x80\x5D\x18\x00\x80\x8C\x35"
-     "\x00\x00"
-     "\x64\x18\x00\x80\x6B\x18\x00\x80\x72\x18\x00\x80\x4F\x18\x00\x80\x79\x18"
-     "\x00\x80"
-     "\x80\x18\x00\x80\x9F\x59\x00\x00\xC2\x18\x00\x80\x83\xDD\x01\x00\xBF\x00"
-     "\x60\x00"
-     "\xB8\x18\x00\x80\xD3\x18\x00\x80\xC9\x18\x00\x80\xA5\xBC\x03\x00\x87\x18"
-     "\x00\x80"
-     "\xE1\x6C\x02\x00\x95\x18\x00\x80\x9C\x18\x00\x80\xA3\x18\x00\x80\xAA\x18"
-     "\x00\x80"
-     "\xB1\x18\x00\x80\x81\x25\x00\x00\x8D\xE1\x01\x00\x92\xA5\x01\x00\x8E\x18"
-     "\x00\x80"
-     "\x9D\x61\x00\x00\x80\xA4\x02\x00\xAE\x0D\x00\x00\x97\xBD\x01\x00\x84\xD0"
-     "\x0E\x00"
-     "\xDA\x18\x00\x80\x96\x23\x00\x80\x9A\x23\x00\x80\x9E\x23\x00\x80\xA2\x23"
-     "\x00\x80"
-     "\x93\x55\x00\x00\x85\xDC\x0F\x00\x9C\x95\x01\x00\xA6\x23\x00\x80\xBE\x88"
-     "\x0E\x00"
-     "\xAA\x23\x00\x80\xBE\x84\x09\x00\xAC\xB9\x00\x00\x84\x90\x0E\x00\xAE\x23"
-     "\x00\x80"
-     "\xBF\x10\x44\x00\xB2\x23\x00\x80\xA7\x91\x00\x00\xB6\x23\x00\x80\xBA\x23"
-     "\x00\x80"
-     "\xA9\x5D\x01\x00\xBE\x23\x00\x80\xA2\x99\x00\x00\xBE\x38\x0B\x00\xC2\x23"
-     "\x00\x80"
-     "\xE1\xA0\x02\x00\xB3\x31\x01\x00\xE3\x14\x16\x00\x84\xC0\x08\x00\xB6\xD1"
-     "\x00\x00"
-     "\xB7\xD9\x00\x00\xB9\x9D\x00\x00\xB5\x25\x01\x00\x86\xD8\x0F\x00\x87\x54"
-     "\x0F\x00"
-     "\xB8\xF9\x00\x00\xB9\xF9\x00\x00\xC6\x23\x00\x80\xCA\x23\x00\x80\xEF\x0C"
-     "\x02\x00"
-     "\x84\x3C\x0A\x00\x23\x18\x00\x80\x80\x88\x0D\x00\x83\x80\x02\x00\x82\x8C"
-     "\x03\x00"
-     "\xCE\x23\x00\x80\x84\xD4\x03\x00\xD2\x23\x00\x80\xD6\x23\x00\x80\xDA\x23"
-     "\x00\x80"
-     "\xDE\x23\x00\x80\xE2\x23\x00\x80\xE6\x23\x00\x80\xEA\x23\x00\x80\xEE\x23"
-     "\x00\x80"
-     "\x8F\xC0\x1A\x02\xF2\x23\x00\x80\xF6\x23\x00\x80\xFA\x23\x00\x80\xFE\x23"
-     "\x00\x80"
-     "\x02\x24\x00\x80\x06\x24\x00\x80\x0A\x24\x00\x80\x0E\x24\x00\x80\x12\x24"
-     "\x00\x80"
-     "\x16\x24\x00\x80\x1A\x24\x00\x80\x1E\x24\x00\x80\x9A\x5C\xD4\x01\x22\x24"
-     "\x00\x80"
-     "\x26\x24\x00\x80\xBC\xDD\x00\x00\x2A\x24\x00\x80\x82\xE9\x00\x00\x83\xF5"
-     "\x00\x00"
-     "\x80\xE9\x00\x00\x81\xE9\x00\x00\x2E\x24\x00\x80\x87\x05\x00\x00\x84\xED"
-     "\x00\x00"
-     "\x85\xE1\x00\x00\x32\x24\x00\x80\x0E\x19\x00\x80\x84\x64\x1A\x00\x3E\x3D"
-     "\x00\x80"
-     "\xAD\x24\x1B\x02\xAC\xA0\x19\x02\x42\x3D\x00\x80\x8D\x1D\x00\x00\x46\x3D"
-     "\x00\x80"
-     "\x4A\x3D\x00\x80\x4E\x3D\x00\x80\xB2\x00\x1C\x02\xB5\x6C\x1E\x02\xB4\x64"
-     "\x1E\x02"
-     "\x52\x3D\x00\x80\xB6\xB0\x1F\x02\xB9\xE0\x12\x02\xB8\x00\x10\x02\xBB\xEC"
-     "\x12\x02"
-     "\xBA\xAC\x1D\x02\xBD\xD0\x11\x02\xBC\xD4\x13\x02\xBF\x00\x14\x02\xBE\xFC"
-     "\x13\x02"
-     "\xA0\x61\x0E\x00\xA1\x61\x0E\x00\xA2\x61\x0E\x00\xA3\x61\x0E\x00\xA4\x6D"
-     "\x0E\x00"
-     "\xA5\x91\x0E\x00\xA6\x8D\x0E\x00\xA7\x81\x0E\x00\xA8\x85\x0E\x00\xA9\x8D"
-     "\x0E\x00"
-     "\xAA\x99\x0E\x00\xAB\x91\x0E\x00\xAC\xC1\x0E\x00\xAD\xC1\x0E\x00\xAE\x29"
-     "\x01\x00"
-     "\xAF\x25\x01\x00\xB0\x5D\x01\x00\xB1\x65\x01\x00\xB2\x6D\x01\x00\xB3\x65"
-     "\x01\x00"
-     "\xB4\x7D\x01\x00\xB5\x65\x01\x00\xB6\x6D\x01\x00\xB7\x61\x01\x00\xB8\x21"
-     "\x01\x00"
-     "\xB9\x21\x01\x00\xBA\x21\x01\x00\xBB\x21\x01\x00\xBC\x21\x01\x00\xBD\x21"
-     "\x01\x00"
-     "\xBE\x21\x01\x00\xBF\x21\x01\x00\x90\x00\xF5\x01\x91\x80\xD0\x01\x92\x08"
-     "\x3F\x02"
-     "\x56\x3D\x00\x80\x5A\x3D\x00\x80\x5E\x3D\x00\x80\x62\x3D\x00\x80\x66\x3D"
-     "\x00\x80"
-     "\x6A\x3D\x00\x80\x6E\x3D\x00\x80\xE0\x24\x0D\x00\x72\x3D\x00\x80\x76\x3D"
-     "\x00\x80"
-     "\x7A\x3D\x00\x80\x7E\x3D\x00\x80\x82\x3D\x00\x80\x80\x84\x0D\x00\x81\xCC"
-     "\x69\x00"
-     "\x82\xA8\x74\x00\x83\x78\x51\x00\x84\xA0\xBC\x00\x85\x54\x98\x00\x86\x4C"
-     "\xE7\x00"
-     "\x87\xDC\xCD\x00\x88\xE4\x29\x01\x89\x04\x34\x01\x8A\x98\x10\x01\x8B\x68"
-     "\x7C\x01"
-     "\x8C\x70\x5B\x01\x8D\xD0\xA1\x01\x8E\xFC\x8D\x01\x8F\x6C\xE9\x01\xAA\x6D"
-     "\x1E\x00"
-     "\xAB\x01\x10\x00\xA8\x01\x1C\x00\xA9\x09\x1E\x00\xAE\x59\x13\x00\xAF\x15"
-     "\x16\x00"
-     "\xAC\x1D\x12\x00\xAD\x15\x12\x00\xA2\x01\x04\x00\xA3\x05\x06\x00\x0B\x18"
-     "\x00\x80"
-     "\xA1\x81\x02\x00\xA6\x35\x1A\x00\xA7\x3D\x1A\x00\xA4\xC9\x06\x00\xA5\xD1"
-     "\x07\x00"
-     "\x86\x3D\x00\x80\x8A\x3D\x00\x80\x8E\x3D\x00\x80\x92\x3D\x00\x80\x96\x3D"
-     "\x00\x80"
-     "\x9A\x3D\x00\x80\x9E\x3D\x00\x80\xA2\x3D\x00\x80\xB2\x01\x68\x00\xB3\x45"
-     "\x6A\x00"
-     "\xB0\x6D\x16\x00\xB1\x41\x17\x00\xA6\x3D\x00\x80\xAA\x3D\x00\x80\xB4\x95"
-     "\x6A\x00"
-     "\xB5\x01\x6C\x00\xAE\x11\x03\x00\xAF\x11\x03\x00\xAC\x51\x02\x00\xAD\x01"
-     "\x03\x00"
-     "\xAA\x55\x02\x00\xAB\x5D\x02\x00\xA8\x21\x02\x00\xA9\x5D\x02\x00\x87\x60"
-     "\x03\x00"
-     "\x86\x04\x04\x00\xB2\x3D\x00\x80\xB6\x3D\x00\x80\x80\xED\x00\x00\x81\xC5"
-     "\x00\x00"
-     "\x82\xCD\x00\x00\xBA\x3D\x00\x80\xBE\x95\x00\x00\xBF\x9D\x00\x00\xBC\x85"
-     "\x00\x00"
-     "\xBD\x9D\x00\x00\xBA\x95\x00\x00\xBB\x9D\x00\x00\xB8\x65\x03\x00\xB9\x9D"
-     "\x00\x00"
-     "\xB6\x55\x03\x00\xB7\x5D\x03\x00\xB4\x45\x03\x00\xB5\x5D\x03\x00\xB2\x51"
-     "\x03\x00"
-     "\xB3\x51\x03\x00\xB0\x61\x03\x00\xB1\x61\x03\x00\xBE\x3D\x00\x80\xC2\x3D"
-     "\x00\x80"
-     "\xC6\x3D\x00\x80\xCA\x3D\x00\x80\xEF\x68\x0E\x00\xCE\x3D\x00\x80\xD2\x3D"
-     "\x00\x80"
-     "\xD6\x3D\x00\x80\xE3\x68\x0E\x00\xDA\x3D\x00\x80\xE1\xD0\x0E\x00\xDE\x3D"
-     "\x00\x80"
-     "\xE2\x3D\x00\x80\xE6\x3D\x00\x80\xEA\x3D\x00\x80\xEE\x3D\x00\x80\xF2\x3D"
-     "\x00\x80"
-     "\xB3\x81\x03\x00\xF6\x3D\x00\x80\xFA\x3D\x00\x80\xB6\xDD\x02\x00\xFE\x3D"
-     "\x00\x80"
-     "\x02\x3E\x00\x80\xB5\x85\x03\x00\xBA\xF9\x02\x00\xBB\xF9\x02\x00\x06\x3E"
-     "\x00\x80"
-     "\x0A\x3E\x00\x80\xBE\xDD\x02\x00\xBF\xC5\x02\x00\xBC\xD9\x02\x00\xBD\xD9"
-     "\x02\x00"
-     "\xE0\xE8\x0F\x00\xA3\x41\x03\x00\x0E\x3E\x00\x80\x12\x3E\x00\x80\xA6\x1D"
-     "\x02\x00"
-     "\x16\x3E\x00\x80\x1A\x3E\x00\x80\xA5\x45\x03\x00\xAA\x39\x02\x00\xAB\x39"
-     "\x02\x00"
-     "\x22\x3E\x00\x80\x26\x3E\x00\x80\xAE\x1D\x02\x00\xAF\x05\x02\x00\xAC\x19"
-     "\x02\x00"
-     "\xAD\x19\x02\x00\x80\x4D\x00\x00\x81\x51\x00\x00\x82\x51\x00\x00\xB3\x71"
-     "\x01\x00"
-     "\x2A\x3E\x00\x80\xB5\xD1\x00\x00\xB6\xD1\x00\x00\x2E\x3E\x00\x80\x86\x40"
-     "\x04\x00"
-     "\x87\x04\x05\x00\xBA\xF1\x00\x00\xBB\xF1\x00\x00\xBC\xD5\x00\x00\xBD\xD9"
-     "\x00\x00"
-     "\xBE\xC9\x00\x00\xBF\xC9\x00\x00\xA8\xA9\x06\x00\xA9\xA9\x06\x00\xAA\xB5"
-     "\x06\x00"
-     "\xAB\x0D\x07\x00\xAC\x15\x07\x00\xAD\x1D\x07\x00\xAE\x15\x07\x00\xAF\x09"
-     "\x07\x00"
-     "\x1E\x3E\x00\x80\x32\x3E\x00\x80\x36\x3E\x00\x80\x3A\x3E\x00\x80\x3E\x3E"
-     "\x00\x80"
-     "\x42\x3E\x00\x80\x46\x3E\x00\x80\x4A\x3E\x00\x80\xB8\x01\x07\x00\xB9\x0D"
-     "\x07\x00"
-     "\xBA\x05\x07\x00\xBB\x19\x07\x00\xBC\x09\x07\x00\xBD\x3D\x07\x00\xBE\x35"
-     "\x07\x00"
-     "\xBF\xD5\x07\x00\xB0\x79\x07\x00\xB1\x79\x07\x00\xB2\x4D\x07\x00\xB3\x41"
-     "\x07\x00"
-     "\xB4\x41\x07\x00\xB5\x41\x07\x00\xB6\x41\x07\x00\xB7\x41\x07\x00\xA3\x3D"
-     "\x06\x00"
-     "\x4E\x3E\x00\x80\x52\x3E\x00\x80\x56\x3E\x00\x80\x5A\x3E\x00\x80\xA6\x9D"
-     "\x07\x00"
-     "\xA5\x9D\x07\x00\x5E\x3E\x00\x80\xAB\xBD\x07\x00\xAA\xBD\x07\x00\x62\x3E"
-     "\x00\x80"
-     "\x66\x3E\x00\x80\xAF\x85\x07\x00\xAE\x85\x07\x00\xAD\x95\x07\x00\xAC\x99"
-     "\x07\x00"
-     "\x6A\x3E\x00\x80\x6E\x3E\x00\x80\x72\x3E\x00\x80\x76\x3E\x00\x80\x7A\x3E"
-     "\x00\x80"
-     "\x7E\x3E\x00\x80\x82\x3E\x00\x80\x86\x3E\x00\x80\x80\x39\x00\x00\x81\x39"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x8A\x3E\x00\x80\x8E\x3E\x00\x80\x92\x3E\x00\x80\x87\xB8"
-     "\x00\x00"
-     "\x86\xFC\x00\x00\xA8\xA9\x06\x00\xA9\xA9\x06\x00\xAA\xD9\x06\x00\xAB\xD1"
-     "\x06\x00"
-     "\xAC\x35\x01\x00\xAD\x3D\x01\x00\xAE\x35\x01\x00\xAF\x2D\x01\x00\x96\x3E"
-     "\x00\x80"
-     "\x9A\x3E\x00\x80\x9E\x3E\x00\x80\xA2\x3E\x00\x80\xA6\x3E\x00\x80\xAA\x3E"
-     "\x00\x80"
-     "\xAE\x3E\x00\x80\xB2\x3E\x00\x80\xB8\xE9\x01\x00\xB9\xE9\x01\x00\xBA\x89"
-     "\x01\x00"
-     "\xBB\x89\x01\x00\xBC\x99\x01\x00\xBD\x99\x01\x00\xBE\x89\x01\x00\xBF\x89"
-     "\x01\x00"
-     "\xB0\x55\x01\x00\xB1\x5D\x01\x00\xB2\x55\x01\x00\xB3\xE9\x01\x00\xB4\xF9"
-     "\x01\x00"
-     "\xB5\xF9\x01\x00\xB6\xE9\x01\x00\xB7\xE1\x01\x00\xB6\x3E\x00\x80\xBA\x3E"
-     "\x00\x80"
-     "\xBE\x3E\x00\x80\xC2\x3E\x00\x80\xC6\x3E\x00\x80\xCA\x3E\x00\x80\xCE\x3E"
-     "\x00\x80"
-     "\xEF\xEC\x08\x00\xD2\x3E\x00\x80\xE1\x18\x05\x00\xD6\x3E\x00\x80\xE3\x2C"
-     "\x08\x00"
-     "\xDA\x3E\x00\x80\xDE\x3E\x00\x80\xE2\x3E\x00\x80\xE6\x3E\x00\x80\xB3\x11"
-     "\x06\x00"
-     "\xEA\x3E\x00\x80\xEE\x3E\x00\x80\xF2\x3E\x00\x80\xF6\x3E\x00\x80\xB6\xFD"
-     "\x01\x00"
-     "\xB5\xFD\x01\x00\xFA\x3E\x00\x80\xBB\xD9\x01\x00\xBA\xD1\x01\x00\xBE\xA4"
-     "\x00\x00"
-     "\xFE\x3E\x00\x80\xBF\x7D\x01\x00\xBE\x7D\x01\x00\xBD\x7D\x01\x00\xBC\xC1"
-     "\x01\x00"
-     "\x86\x00\x1C\x00\x87\x44\x03\x00\x02\x3F\x00\x80\xA3\xDD\x05\x00\x06\x3F"
-     "\x00\x80"
-     "\xA5\x31\x02\x00\xA6\x31\x02\x00\x0A\x3F\x00\x80\xAE\x3D\x00\x80\x0E\x3F"
-     "\x00\x80"
-     "\xAA\x1D\x02\x00\xAB\x15\x02\x00\xAC\x0D\x02\x00\xAD\xB1\x02\x00\xAE\xB1"
-     "\x02\x00"
-     "\xAF\xB1\x02\x00\x80\x8D\x03\x00\x81\x95\x03\x00\x82\x95\x03\x00\x12\x3F"
-     "\x00\x80"
-     "\xEF\xB8\x03\x00\x16\x3F\x00\x80\x1A\x3F\x00\x80\x1E\x3F\x00\x80\xE3\xD8"
-     "\x04\x00"
-     "\x22\x3F\x00\x80\xE1\x70\x02\x00\x26\x3F\x00\x80\x2A\x3F\x00\x80\x2E\x3F"
-     "\x00\x80"
-     "\x32\x3F\x00\x80\x36\x3F\x00\x80\x3A\x3F\x00\x80\x3E\x3F\x00\x80\x42\x3F"
-     "\x00\x80"
-     "\x46\x3F\x00\x80\x4A\x3F\x00\x80\x4E\x3F\x00\x80\x52\x3F\x00\x80\x56\x3F"
-     "\x00\x80"
-     "\xA8\x55\x02\x00\xA9\x85\x03\x00\xAA\x8D\x03\x00\xAB\x85\x03\x00\xAC\x9D"
-     "\x03\x00"
-     "\xAD\x85\x03\x00\xAE\x8D\x03\x00\xAF\x85\x03\x00\xB0\xFD\x03\x00\xB1\x45"
-     "\x03\x00"
-     "\xB2\x4D\x03\x00\xB3\x45\x03\x00\xB4\x5D\x03\x00\xB5\x45\x03\x00\xB6\x4D"
-     "\x03\x00"
-     "\xB7\x45\x03\x00\xB8\x41\x03\x00\xB9\x4D\x03\x00\xBA\x59\x03\x00\xBB\x59"
-     "\x03\x00"
-     "\xBC\x49\x03\x00\xBD\x49\x03\x00\xBE\x3D\x03\x00\xBF\x35\x03\x00\x5A\x3F"
-     "\x00\x80"
-     "\x5E\x3F\x00\x80\x62\x3F\x00\x80\xB3\x29\x02\x00\xBE\xD8\x1D\x00\xB5\x39"
-     "\x02\x00"
-     "\xB6\x29\x02\x00\x66\x3F\x00\x80\x6A\x3F\x00\x80\x72\x3F\x00\x80\xBA\x15"
-     "\x02\x00"
-     "\xBB\xAD\x02\x00\xBC\xB5\x02\x00\xBD\xBD\x02\x00\xBE\xB5\x02\x00\xBF\xAD"
-     "\x02\x00"
-     "\x80\x49\x00\x00\x81\x59\x00\x00\x82\x69\x00\x00\xEF\x5C\x00\x00\x86\xF0"
-     "\x1C\x00"
-     "\x87\x1C\x1D\x00\x76\x3F\x00\x80\x7A\x3F\x00\x80\x7E\x3F\x00\x80\x82\x3F"
-     "\x00\x80"
-     "\x86\x3F\x00\x80\x8A\x3F\x00\x80\x8E\x3F\x00\x80\xE1\x90\x01\x00\x92\x3F"
-     "\x00\x80"
-     "\xE3\x90\x0A\x00\x96\x3F\x00\x80\x9A\x3F\x00\x80\x9E\x3F\x00\x80\xA2\x3F"
-     "\x00\x80"
-     "\xA6\x3F\x00\x80\xAA\x3F\x00\x80\xAE\x3F\x00\x80\xEF\xDC\x0A\x00\xBE\xE8"
-     "\x1C\x00"
-     "\xE1\x5C\x06\x00\xB2\x3F\x00\x80\xE3\x70\x0A\x00\xB6\x3F\x00\x80\xBA\x3F"
-     "\x00\x80"
-     "\xBE\x3F\x00\x80\xC2\x3F\x00\x80\xA3\xE5\x02\x00\xC6\x3F\x00\x80\xCA\x3F"
-     "\x00\x80"
-     "\xCE\x3F\x00\x80\xD2\x3F\x00\x80\xA6\xE5\x02\x00\xA5\xF5\x02\x00\xD6\x3F"
-     "\x00\x80"
-     "\xAB\x61\x02\x00\xAA\xD9\x02\x00\xDA\x3F\x00\x80\x84\xA0\x1C\x00\xAF\x61"
-     "\x02\x00"
-     "\xAE\x79\x02\x00\xAD\x71\x02\x00\xAC\x79\x02\x00\xB3\x1D\x1E\x00\x6E\x3F"
-     "\x00\x80"
-     "\xDE\x3F\x00\x80\xE2\x3F\x00\x80\xE6\x3F\x00\x80\xB6\x41\x1E\x00\xB5\x55"
-     "\x1E\x00"
-     "\xEA\x3F\x00\x80\xBB\x4D\x1E\x00\xBA\x41\x1E\x00\x86\x88\x00\x00\x87\xE4"
-     "\x03\x00"
-     "\xBF\x4D\x1E\x00\xBE\x4D\x1E\x00\xBD\x4D\x1E\x00\xBC\x4D\x1E\x00\x82\xA9"
-     "\x1F\x00"
-     "\xA3\x59\x1E\x00\x80\x99\x1F\x00\x81\x91\x1F\x00\xA6\x05\x1E\x00\xEE\x3F"
-     "\x00\x80"
-     "\xF2\x3F\x00\x80\xA5\x11\x1E\x00\xAA\x05\x1E\x00\xAB\x09\x1E\x00\xF6\x3F"
-     "\x00\x80"
-     "\xFA\x3F\x00\x80\xAE\x09\x1E\x00\xAF\x09\x1E\x00\xAC\x09\x1E\x00\xAD\x09"
-     "\x1E\x00"
-     "\xFE\x3F\x00\x80\x02\x40\x00\x80\x06\x40\x00\x80\x0A\x40\x00\x80\x0E\x40"
-     "\x00\x80"
-     "\x12\x40\x00\x80\x16\x40\x00\x80\x1A\x40\x00\x80\xA8\xA9\x1E\x00\xA9\xB5"
-     "\x1E\x00"
-     "\xAA\xB5\x1E\x00\xAB\x89\x1E\x00\xAC\xFD\x1E\x00\xAD\xE1\x1E\x00\xAE\xE1"
-     "\x1E\x00"
-     "\xAF\xE1\x1E\x00\xB0\xA5\x1E\x00\xB1\xA9\x1E\x00\xB2\xB9\x1E\x00\xB3\xB9"
-     "\x1E\x00"
-     "\xB4\xA9\x1E\x00\xB5\xA9\x1E\x00\xB6\x59\x1E\x00\xB7\x59\x1E\x00\xB8\x69"
-     "\x1E\x00"
-     "\xB9\x69\x1E\x00\xBA\x79\x1E\x00\xBB\x79\x1E\x00\xBC\x69\x1E\x00\xBD\x69"
-     "\x1E\x00"
-     "\xBE\xF9\x1F\x00\xBF\xF9\x1F\x00\xA8\x25\x1E\x00\xA9\x5D\x1E\x00\xAA\x55"
-     "\x1E\x00"
-     "\xAB\x6D\x1E\x00\xAC\x75\x1E\x00\xAD\x7D\x1E\x00\xAE\x75\x1E\x00\xAF\x6D"
-     "\x1E\x00"
-     "\x1E\x40\x00\x80\x22\x40\x00\x80\x26\x40\x00\x80\x2A\x40\x00\x80\x2E\x40"
-     "\x00\x80"
-     "\x32\x40\x00\x80\x36\x40\x00\x80\x3A\x40\x00\x80\xB8\xE9\x1E\x00\xB9\xF5"
-     "\x1E\x00"
-     "\xBA\xFD\x1E\x00\xBB\xF5\x1E\x00\xBC\xED\x1E\x00\xBD\x91\x1E\x00\xBE\x91"
-     "\x1E\x00"
-     "\xBF\x91\x1E\x00\xB0\x19\x1E\x00\xB1\xE5\x1E\x00\xB2\xED\x1E\x00\xB3\xE5"
-     "\x1E\x00"
-     "\xB4\xFD\x1E\x00\xB5\xE5\x1E\x00\xB6\xED\x1E\x00\xB7\xE5\x1E\x00\xB3\xDD"
-     "\x1E\x00"
-     "\x3E\x40\x00\x80\x42\x40\x00\x80\x46\x40\x00\x80\x4A\x40\x00\x80\xB6\xFD"
-     "\x1E\x00"
-     "\xB5\xFD\x1E\x00\x84\xD0\x01\x00\xBB\x21\x1E\x00\xBA\xDD\x1E\x00\xBE\xE0"
-     "\x01\x00"
-     "\x52\x40\x00\x80\xBF\x21\x1E\x00\xBE\x39\x1E\x00\xBD\x31\x1E\x00\xBC\x39"
-     "\x1E\x00"
-     "\x82\x51\x00\x00\xA3\x99\x1E\x00\x80\x41\x00\x00\x81\x41\x00\x00\xA6\xB9"
-     "\x1E\x00"
-     "\x56\x40\x00\x80\x5A\x40\x00\x80\xA5\xB9\x1E\x00\xAA\x99\x1E\x00\xAB\x65"
-     "\x1E\x00"
-     "\x86\x48\x00\x00\x87\x78\x00\x00\xAE\x7D\x1E\x00\xAF\x65\x1E\x00\xAC\x7D"
-     "\x1E\x00"
-     "\xAD\x75\x1E\x00\xA8\x0D\x02\x00\xA9\x15\x02\x00\xAA\x15\x02\x00\xAB\x59"
-     "\x02\x00"
-     "\xAC\x49\x02\x00\xAD\x7D\x02\x00\xAE\x75\x02\x00\xAF\x9D\x02\x00\x5E\x40"
-     "\x00\x80"
-     "\x62\x40\x00\x80\x66\x40\x00\x80\x6A\x40\x00\x80\x84\x3C\x03\x00\x6E\x40"
-     "\x00\x80"
-     "\x72\x40\x00\x80\x76\x40\x00\x80\xB8\x65\x03\x00\xB9\x6D\x03\x00\xBA\x65"
-     "\x03\x00"
-     "\xBB\x7D\x03\x00\xBC\x65\x03\x00\xBD\x6D\x03\x00\xBE\x65\x03\x00\xBF\xD9"
-     "\x03\x00"
-     "\xB0\xE5\x02\x00\xB1\xED\x02\x00\xB2\xE5\x02\x00\xB3\xFD\x02\x00\xB4\xE5"
-     "\x02\x00"
-     "\xB5\xED\x02\x00\xB6\xE5\x02\x00\xB7\x5D\x03\x00\xB3\x9D\x02\x00\x7A\x40"
-     "\x00\x80"
-     "\x7E\x40\x00\x80\x82\x40\x00\x80\x86\x40\x00\x80\xB6\xFD\x02\x00\xB5\xC1"
-     "\x02\x00"
-     "\x8A\x40\x00\x80\xBB\x29\x02\x00\xBA\x21\x02\x00\x8E\x40\x00\x80\x92\x40"
-     "\x00\x80"
-     "\xBF\x29\x02\x00\xBE\x21\x02\x00\xBD\x29\x02\x00\xBC\x31\x02\x00\x96\x40"
-     "\x00\x80"
-     "\xA3\xD9\x02\x00\xBE\x44\x04\x00\x84\x00\x05\x00\xA6\xB9\x02\x00\x9A\x40"
-     "\x00\x80"
-     "\x9E\x40\x00\x80\xA5\x85\x02\x00\xAA\x65\x02\x00\xAB\x6D\x02\x00\xA2\x40"
-     "\x00\x80"
-     "\xA6\x40\x00\x80\xAE\x65\x02\x00\xAF\x6D\x02\x00\xAC\x75\x02\x00\xAD\x6D"
-     "\x02\x00"
-     "\xA8\xAD\x02\x00\xA9\xD1\x02\x00\xAA\xD1\x02\x00\xAB\x2D\x01\x00\xAC\x35"
-     "\x01\x00"
-     "\xAD\x3D\x01\x00\xAE\x35\x01\x00\xAF\x29\x01\x00\xAA\x40\x00\x80\xAE\x40"
-     "\x00\x80"
-     "\xB2\x40\x00\x80\xB6\x40\x00\x80\xBA\x40\x00\x80\xC2\x40\x00\x80\xC6\x40"
-     "\x00\x80"
-     "\xCA\x40\x00\x80\xB8\xED\x01\x00\xB9\x85\x01\x00\xBA\x9D\x01\x00\xBB\x95"
-     "\x01\x00"
-     "\xBC\xB9\x01\x00\xBD\xB9\x01\x00\xBE\x79\x00\x00\xBF\x79\x00\x00\xB0\x59"
-     "\x01\x00"
-     "\xB1\x59\x01\x00\xB2\xED\x01\x00\xB3\xE5\x01\x00\xB4\xFD\x01\x00\xB5\xE5"
-     "\x01\x00"
-     "\xB6\xE5\x01\x00\xB7\xD5\x01\x00\xE1\xC8\x1E\x00\xE3\x08\x00\x00\xE3\x74"
-     "\x1E\x00"
-     "\xE1\xBC\x01\x00\x81\xBD\x00\x00\x80\xBD\x00\x00\xCE\x40\x00\x80\x82\xAD"
-     "\x00\x00"
-     "\x86\x80\x05\x00\x87\xF4\x05\x00\xD2\x40\x00\x80\xD6\x40\x00\x80\xDA\x40"
-     "\x00\x80"
-     "\xEF\xE4\x00\x00\xEF\xE0\x1E\x00\xDE\x40\x00\x80\xB3\x19\x02\x00\xE2\x40"
-     "\x00\x80"
-     "\x85\x70\x75\x00\xE6\x40\x00\x80\xEA\x40\x00\x80\xB6\xFD\x01\x00\xB5\xFD"
-     "\x01\x00"
-     "\xEE\x40\x00\x80\xBB\xDD\x01\x00\xBA\xDD\x01\x00\xF2\x40\x00\x80\xF6\x40"
-     "\x00\x80"
-     "\xBF\x61\x01\x00\xBE\x79\x01\x00\xBD\x71\x01\x00\xBC\x79\x01\x00\xA3\xD9"
-     "\x05\x00"
-     "\xBE\x40\x00\x80\xFA\x40\x00\x80\xFE\x40\x00\x80\x02\x41\x00\x80\xA6\x3D"
-     "\x06\x00"
-     "\xA5\x3D\x06\x00\x06\x41\x00\x80\xAB\x1D\x06\x00\xAA\x1D\x06\x00\x0A\x41"
-     "\x00\x80"
-     "\x0E\x41\x00\x80\xAF\xA1\x06\x00\xAE\xB9\x06\x00\xAD\xB1\x06\x00\xAC\xB9"
-     "\x06\x00"
-     "\x12\x41\x00\x80\x16\x41\x00\x80\x1A\x41\x00\x80\x1E\x41\x00\x80\x80\x0D"
-     "\x00\x00"
-     "\x81\x35\x00\x00\x82\x3D\x00\x00\x22\x41\x00\x80\x26\x41\x00\x80\x2A\x41"
-     "\x00\x80"
-     "\xBE\xA0\x03\x00\x2E\x41\x00\x80\x86\xF0\x00\x00\x87\xE4\x03\x00\x32\x41"
-     "\x00\x80"
-     "\x36\x41\x00\x80\xA8\xB5\x07\x00\xA9\xBD\x07\x00\xAA\xB5\x07\x00\xAB\x15"
-     "\x07\x00"
-     "\xAC\x39\x07\x00\xAD\x39\x07\x00\xAE\x29\x07\x00\xAF\x21\x07\x00\x3A\x41"
-     "\x00\x80"
-     "\x3E\x41\x00\x80\x42\x41\x00\x80\x46\x41\x00\x80\x4A\x41\x00\x80\x4E\x41"
-     "\x00\x80"
-     "\x52\x41\x00\x80\x56\x41\x00\x80\xB8\xD1\x00\x00\xB9\xD9\x00\x00\xBA\xE1"
-     "\x00\x00"
-     "\xBB\xE1\x00\x00\xBC\x91\x00\x00\xBD\x91\x00\x00\xBE\x91\x00\x00\xBF\x91"
-     "\x00\x00"
-     "\xB0\x25\x07\x00\xB1\x29\x07\x00\xB2\x39\x07\x00\xB3\x31\x07\x00\xB4\x19"
-     "\x07\x00"
-     "\xB5\x19\x07\x00\xB6\xF1\x00\x00\xB7\xF1\x00\x00\xB3\x3D\x06\x00\x5A\x41"
-     "\x00\x80"
-     "\x5E\x41\x00\x80\x62\x41\x00\x80\x66\x41\x00\x80\xB6\xB9\x06\x00\xB5\xB1"
-     "\x06\x00"
-     "\x6A\x41\x00\x80\xBB\xED\x06\x00\xBA\xE1\x06\x00\x6E\x41\x00\x80\x72\x41"
-     "\x00\x80"
-     "\xBF\xDD\x06\x00\xBE\xDD\x06\x00\xBD\xE5\x06\x00\xBC\xED\x06\x00\x76\x41"
-     "\x00\x80"
-     "\xA3\x79\x06\x00\x7A\x41\x00\x80\x7E\x41\x00\x80\xA6\xFD\x06\x00\x82\x41"
-     "\x00\x80"
-     "\x86\x41\x00\x80\xA5\xF5\x06\x00\xAA\xA5\x06\x00\xAB\xA9\x06\x00\x8A\x41"
-     "\x00\x80"
-     "\x8E\x41\x00\x80\xAE\x99\x06\x00\xAF\x99\x06\x00\xAC\xA9\x06\x00\xAD\xA1"
-     "\x06\x00"
-     "\x92\x41\x00\x80\xE3\xA4\x05\x00\x96\x41\x00\x80\xE1\x7C\x06\x00\x80\x11"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x82\x05\x00\x00\x9A\x41\x00\x80\x84\xAC\x02\x00\xBE\xA4"
-     "\x01\x00"
-     "\x87\x64\x01\x00\x86\x0C\x00\x00\xA2\x41\x00\x80\xEF\x84\x05\x00\xA6\x41"
-     "\x00\x80"
-     "\xAA\x41\x00\x80\xB3\x61\x01\x00\xAE\x41\x00\x80\xB2\x41\x00\x80\xB6\x41"
-     "\x00\x80"
-     "\xBA\x41\x00\x80\xB6\x79\x01\x00\xB5\x71\x01\x00\xBE\x41\x00\x80\xBB\x5D"
-     "\x01\x00"
-     "\xBA\x5D\x01\x00\xC2\x41\x00\x80\xC6\x41\x00\x80\xBF\xE1\x00\x00\xBE\xF9"
-     "\x00\x00"
-     "\xBD\xF1\x00\x00\xBC\xF9\x00\x00\xCA\x41\x00\x80\xCE\x41\x00\x80\x4E\x40"
-     "\x00\x80"
-     "\xD2\x41\x00\x80\xD6\x41\x00\x80\xDA\x41\x00\x80\xDE\x41\x00\x80\xE2\x41"
-     "\x00\x80"
-     "\xA8\x59\x02\x00\xA9\x59\x02\x00\xAA\x89\x02\x00\xAB\x89\x02\x00\xAC\x99"
-     "\x02\x00"
-     "\xAD\x99\x02\x00\xAE\x89\x02\x00\xAF\x89\x02\x00\xB0\xF9\x02\x00\xB1\xF9"
-     "\x02\x00"
-     "\xB2\x4D\x03\x00\xB3\x45\x03\x00\xB4\x5D\x03\x00\xB5\x45\x03\x00\xB6\x45"
-     "\x03\x00"
-     "\xB7\x75\x03\x00\xB8\x4D\x03\x00\xB9\x25\x03\x00\xBA\x21\x03\x00\xBB\x21"
-     "\x03\x00"
-     "\xBC\x25\x03\x00\xBD\x2D\x03\x00\xBE\x11\x03\x00\xBF\x11\x03\x00\xE6\x41"
-     "\x00\x80"
-     "\xEA\x41\x00\x80\xEE\x41\x00\x80\xA3\x2D\x02\x00\xBE\x58\x0C\x00\xA5\x3D"
-     "\x02\x00"
-     "\xA6\x35\x02\x00\xF2\x41\x00\x80\xF6\x41\x00\x80\xFA\x41\x00\x80\xAA\x11"
-     "\x02\x00"
-     "\xAB\x11\x02\x00\xAC\xB5\x03\x00\xAD\xBD\x03\x00\xAE\xB5\x03\x00\xAF\xAD"
-     "\x03\x00"
-     "\xFE\x41\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\x02\x42"
-     "\x00\x80"
-     "\x06\x42\x00\x80\x0A\x42\x00\x80\x12\x42\x00\x80\x84\x2C\x0C\x00\x16\x42"
-     "\x00\x80"
-     "\x87\x28\x03\x00\x86\xCC\x0C\x00\x1A\x42\x00\x80\x1E\x42\x00\x80\x22\x42"
-     "\x00\x80"
-     "\x26\x42\x00\x80\xA8\xB1\x02\x00\xA9\xBD\x02\x00\xAA\xB5\x02\x00\xAB\xC9"
-     "\x02\x00"
-     "\xAC\xD9\x02\x00\xAD\xD9\x02\x00\xAE\xC9\x02\x00\xAF\xC1\x02\x00\x2A\x42"
-     "\x00\x80"
-     "\x2E\x42\x00\x80\x32\x42\x00\x80\x36\x42\x00\x80\x3A\x42\x00\x80\x3E\x42"
-     "\x00\x80"
-     "\x42\x42\x00\x80\x46\x42\x00\x80\xB8\x5D\x01\x00\xB9\xC5\x01\x00\xBA\xCD"
-     "\x01\x00"
-     "\xBB\xC5\x01\x00\xBC\xCD\x01\x00\xBD\xF5\x01\x00\xBE\xFD\x01\x00\xBF\x9D"
-     "\x01\x00"
-     "\xB0\x89\x02\x00\xB1\x89\x02\x00\xB2\x69\x01\x00\xB3\x69\x01\x00\xB4\x7D"
-     "\x01\x00"
-     "\xB5\x65\x01\x00\xB6\x6D\x01\x00\xB7\x65\x01\x00\xE1\x38\x06\x00\x4A\x42"
-     "\x00\x80"
-     "\xE3\x54\x07\x00\x4E\x42\x00\x80\x85\x8C\x70\x00\x52\x42\x00\x80\x56\x42"
-     "\x00\x80"
-     "\x5A\x42\x00\x80\x5E\x42\x00\x80\x62\x42\x00\x80\x66\x42\x00\x80\x6A\x42"
-     "\x00\x80"
-     "\xE9\x00\x00\x00\x84\x78\x0C\x00\xEF\x28\x01\x00\x6E\x42\x00\x80\x80\x15"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x82\x15\x00\x00\xBE\x04\x0C\x00\xEF\x8C\x00\x00\x72\x42"
-     "\x00\x80"
-     "\x76\x42\x00\x80\x7A\x42\x00\x80\xE3\x50\x00\x00\x82\x42\x00\x80\xE1\x80"
-     "\x01\x00"
-     "\x86\x42\x00\x80\x8A\x42\x00\x80\x8E\x42\x00\x80\x86\x78\x0F\x00\x87\xF0"
-     "\x0C\x00"
-     "\xB3\xBD\x0D\x00\x0E\x42\x00\x80\x7E\x42\x00\x80\x92\x42\x00\x80\x96\x42"
-     "\x00\x80"
-     "\xB6\xD9\x0D\x00\xB5\xD1\x0D\x00\x9A\x42\x00\x80\xBB\xCD\x0D\x00\xBA\xCD"
-     "\x0D\x00"
-     "\x9E\x42\x00\x80\xA2\x42\x00\x80\xBF\x21\x0E\x00\xBE\x39\x0E\x00\xBD\x3D"
-     "\x0E\x00"
-     "\xBC\x3D\x0E\x00\xA6\x42\x00\x80\xA3\xF9\x0D\x00\xAA\x42\x00\x80\xAE\x42"
-     "\x00\x80"
-     "\xA6\x9D\x0D\x00\xB2\x42\x00\x80\xB6\x42\x00\x80\xA5\x95\x0D\x00\xAA\x89"
-     "\x0D\x00"
-     "\xAB\x89\x0D\x00\xBA\x42\x00\x80\xBE\x42\x00\x80\xAE\x7D\x0E\x00\xAF\x65"
-     "\x0E\x00"
-     "\xAC\x79\x0E\x00\xAD\x79\x0E\x00\xA8\x5D\x0E\x00\xA9\x61\x0E\x00\xAA\x79"
-     "\x0E\x00"
-     "\xAB\x79\x0E\x00\xAC\x69\x0E\x00\xAD\x69\x0E\x00\xAE\xB9\x0F\x00\xAF\xB9"
-     "\x0F\x00"
-     "\xC2\x42\x00\x80\xC6\x42\x00\x80\xCA\x42\x00\x80\xCE\x42\x00\x80\xD2\x42"
-     "\x00\x80"
-     "\xD6\x42\x00\x80\xDA\x42\x00\x80\xDE\x42\x00\x80\xB8\x55\x0F\x00\xB9\x5D"
-     "\x0F\x00"
-     "\xBA\x55\x0F\x00\xBB\x6D\x0F\x00\xBC\x75\x0F\x00\xBD\x7D\x0F\x00\xBE\x75"
-     "\x0F\x00"
-     "\xBF\x61\x0F\x00\xB0\xC9\x0F\x00\xB1\xC9\x0F\x00\xB2\xD9\x0F\x00\xB3\xD9"
-     "\x0F\x00"
-     "\xB4\xC9\x0F\x00\xB5\xC9\x0F\x00\xB6\x7D\x0F\x00\xB7\x71\x0F\x00\xB3\xF1"
-     "\x0F\x00"
-     "\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D\x00\x00\xE2\x42\x00\x80\xB6\xE9"
-     "\x0F\x00"
-     "\xB5\xE1\x0F\x00\xE6\x42\x00\x80\xBB\x0D\x0E\x00\xBA\x0D\x0E\x00\xEA\x42"
-     "\x00\x80"
-     "\x84\x24\x02\x00\xBF\x0D\x0E\x00\xBE\x0D\x0E\x00\xBD\x0D\x0E\x00\xBC\x0D"
-     "\x0E\x00"
-     "\xBE\x40\x01\x00\xA3\xB5\x0F\x00\xEE\x42\x00\x80\xF2\x42\x00\x80\xA6\xAD"
-     "\x0F\x00"
-     "\xF6\x42\x00\x80\xFA\x42\x00\x80\xA5\xA5\x0F\x00\xAA\x49\x0E\x00\xAB\x49"
-     "\x0E\x00"
-     "\x86\xC8\x00\x00\x87\x78\x01\x00\xAE\x49\x0E\x00\xAF\x49\x0E\x00\xAC\x49"
-     "\x0E\x00"
-     "\xAD\x49\x0E\x00\xE1\x34\x0E\x00\xFE\x42\x00\x80\xE3\x78\x0D\x00\x02\x43"
-     "\x00\x80"
-     "\x06\x43\x00\x80\x0A\x43\x00\x80\x0E\x43\x00\x80\x12\x43\x00\x80\x84\xAC"
-     "\x00\x00"
-     "\x16\x43\x00\x80\x1A\x43\x00\x80\x1E\x43\x00\x80\x22\x43\x00\x80\x26\x43"
-     "\x00\x80"
-     "\xEF\xD0\x0D\x00\x2A\x43\x00\x80\x2E\x43\x00\x80\xB3\x35\x0E\x00\x32\x43"
-     "\x00\x80"
-     "\x36\x43\x00\x80\xB6\xC1\x01\x00\x3A\x43\x00\x80\x3E\x43\x00\x80\xB5\xD1"
-     "\x01\x00"
-     "\xBA\xCD\x01\x00\xBB\xA5\x01\x00\x42\x43\x00\x80\x46\x43\x00\x80\xBE\xA5"
-     "\x01\x00"
-     "\xBF\xAD\x01\x00\xBC\xB1\x01\x00\xBD\xB1\x01\x00\x4A\x43\x00\x80\x9E\x41"
-     "\x00\x80"
-     "\x4E\x43\x00\x80\x52\x43\x00\x80\x81\x09\x00\x00\x80\x19\x00\x00\x56\x43"
-     "\x00\x80"
-     "\x82\x9D\x00\x00\xA8\xC5\x0D\x00\xA9\xC9\x0D\x00\xAA\xD9\x0D\x00\xAB\xD9"
-     "\x0D\x00"
-     "\xAC\x3D\x02\x00\xAD\x25\x02\x00\xAE\x25\x02\x00\xAF\x15\x02\x00\xB0\x6D"
-     "\x02\x00"
-     "\xB1\x05\x02\x00\xB2\x01\x02\x00\xB3\x01\x02\x00\xB4\x05\x02\x00\xB5\x0D"
-     "\x02\x00"
-     "\xB6\x31\x02\x00\xB7\x31\x02\x00\xB8\xC1\x02\x00\xB9\xC1\x02\x00\xBA\xC1"
-     "\x02\x00"
-     "\xBB\xC1\x02\x00\xBC\xC1\x02\x00\xBD\xC1\x02\x00\xBE\xC1\x02\x00\xBF\xC1"
-     "\x02\x00"
-     "\x5A\x43\x00\x80\xBE\x00\x04\x00\x62\x43\x00\x80\xA3\x7D\x0D\x00\x66\x43"
-     "\x00\x80"
-     "\xA5\x99\x02\x00\xA6\x89\x02\x00\x6A\x43\x00\x80\x6E\x43\x00\x80\x72\x43"
-     "\x00\x80"
-     "\xAA\x85\x02\x00\xAB\xED\x02\x00\xAC\xF9\x02\x00\xAD\xF9\x02\x00\xAE\xED"
-     "\x02\x00"
-     "\xAF\xE5\x02\x00\x76\x43\x00\x80\xE3\x74\x01\x00\x84\x48\x04\x00\xE1\xBC"
-     "\x01\x00"
-     "\x86\xF0\x04\x00\x87\x94\x04\x00\x7A\x43\x00\x80\x7E\x43\x00\x80\x82\x43"
-     "\x00\x80"
-     "\x86\x43\x00\x80\x8A\x43\x00\x80\x8E\x43\x00\x80\x92\x43\x00\x80\xEF\x88"
-     "\x02\x00"
-     "\x96\x43\x00\x80\x9A\x43\x00\x80\xA8\x4D\x02\x00\xA9\x51\x02\x00\xAA\x51"
-     "\x02\x00"
-     "\xAB\x89\x02\x00\xAC\x9D\x02\x00\xAD\x85\x02\x00\xAE\x8D\x02\x00\xAF\x85"
-     "\x02\x00"
-     "\x9E\x43\x00\x80\xA2\x43\x00\x80\xA6\x43\x00\x80\xAA\x43\x00\x80\xAE\x43"
-     "\x00\x80"
-     "\xB2\x43\x00\x80\xB6\x43\x00\x80\xBE\xB0\x04\x00\xB8\x89\x02\x00\xB9\x89"
-     "\x02\x00"
-     "\xBA\x9D\x02\x00\xBB\x95\x02\x00\xBC\xB9\x02\x00\xBD\xB9\x02\x00\xBE\x59"
-     "\x01\x00"
-     "\xBF\x55\x01\x00\xB0\xFD\x02\x00\xB1\xC1\x02\x00\xB2\xC1\x02\x00\xB3\xC1"
-     "\x02\x00"
-     "\xB4\xC1\x02\x00\xB5\xC1\x02\x00\xB6\xC5\x02\x00\xB7\xBD\x02\x00\xE1\x74"
-     "\x0F\x00"
-     "\xE3\xB8\x00\x00\xE3\x2C\x0E\x00\xE1\x6C\x0E\x00\xBA\x43\x00\x80\xBE\x43"
-     "\x00\x80"
-     "\xC2\x43\x00\x80\xC6\x43\x00\x80\xCA\x43\x00\x80\xD2\x43\x00\x80\xD6\x43"
-     "\x00\x80"
-     "\xDA\x43\x00\x80\xEA\x00\x00\x00\xEF\x0C\x00\x00\xEF\x8C\x0E\x00\xDE\x43"
-     "\x00\x80"
-     "\xB3\x99\x02\x00\xE2\x43\x00\x80\x86\xE8\x04\x00\x87\xC0\x04\x00\xE6\x43"
-     "\x00\x80"
-     "\xB6\x45\x02\x00\xB5\x55\x02\x00\xEA\x43\x00\x80\xBB\x45\x02\x00\xBA\x79"
-     "\x02\x00"
-     "\xEE\x43\x00\x80\xF2\x43\x00\x80\xBF\x45\x02\x00\xBE\x45\x02\x00\xBD\x55"
-     "\x02\x00"
-     "\xBC\x55\x02\x00\x5E\x43\x00\x80\xCE\x43\x00\x80\xA3\x51\x05\x00\xF6\x43"
-     "\x00\x80"
-     "\xA5\x9D\x05\x00\xFA\x43\x00\x80\xFE\x43\x00\x80\xA6\x8D\x05\x00\x02\x44"
-     "\x00\x80"
-     "\x06\x44\x00\x80\xAB\x8D\x05\x00\xAA\xB1\x05\x00\xAD\x9D\x05\x00\xAC\x9D"
-     "\x05\x00"
-     "\xAF\x8D\x05\x00\xAE\x8D\x05\x00\x80\xD9\x07\x00\x81\xD9\x07\x00\x82\xE9"
-     "\x07\x00"
-     "\xB3\xD1\x06\x00\x0A\x44\x00\x80\xB5\xF1\x06\x00\xB6\x99\x06\x00\x0E\x44"
-     "\x00\x80"
-     "\x12\x44\x00\x80\x16\x44\x00\x80\xBA\x89\x06\x00\xBB\x89\x06\x00\xBC\x99"
-     "\x06\x00"
-     "\xBD\x99\x06\x00\xBE\x89\x06\x00\xBF\x85\x06\x00\xA8\x09\x06\x00\xA9\x15"
-     "\x06\x00"
-     "\xAA\x15\x06\x00\xAB\x2D\x06\x00\xAC\x35\x06\x00\xAD\x55\x06\x00\xAE\x41"
-     "\x06\x00"
-     "\xAF\x41\x06\x00\x1A\x44\x00\x80\x1E\x44\x00\x80\x22\x44\x00\x80\x26\x44"
-     "\x00\x80"
-     "\x2A\x44\x00\x80\x2E\x44\x00\x80\x32\x44\x00\x80\x36\x44\x00\x80\xB8\xE5"
-     "\x07\x00"
-     "\xB9\xED\x07\x00\xBA\xE5\x07\x00\xBB\xFD\x07\x00\xBC\xE5\x07\x00\xBD\xED"
-     "\x07\x00"
-     "\xBE\x49\x07\x00\xBF\x49\x07\x00\xB0\x01\x06\x00\xB1\x01\x06\x00\xB2\x01"
-     "\x06\x00"
-     "\xB3\x01\x06\x00\xB4\x01\x06\x00\xB5\x0D\x06\x00\xB6\x05\x06\x00\xB7\xDD"
-     "\x07\x00"
-     "\xA3\x9D\x07\x00\x82\x15\x00\x00\x81\xB5\x01\x00\x80\xB5\x01\x00\x3A\x44"
-     "\x00\x80"
-     "\xA6\xD5\x07\x00\xA5\xBD\x07\x00\xBE\x90\x01\x00\xAB\xC5\x07\x00\xAA\xC5"
-     "\x07\x00"
-     "\x84\xA0\x02\x00\x3E\x44\x00\x80\xAF\xC9\x07\x00\xAE\xC5\x07\x00\xAD\xD5"
-     "\x07\x00"
-     "\xAC\xD5\x07\x00\x42\x44\x00\x80\xB3\xE9\x06\x00\x86\xE8\x00\x00\x87\x24"
-     "\x01\x00"
-     "\xB6\x81\x06\x00\x46\x44\x00\x80\x4A\x44\x00\x80\xB5\x99\x06\x00\xBA\x8D"
-     "\x06\x00"
-     "\xBB\x69\x01\x00\x4E\x44\x00\x80\x52\x44\x00\x80\xBE\x51\x01\x00\xBF\x51"
-     "\x01\x00"
-     "\xBC\x6D\x01\x00\xBD\x65\x01\x00\xA8\x25\x06\x00\xA9\x2D\x06\x00\xAA\x25"
-     "\x06\x00"
-     "\xAB\x31\x06\x00\xAC\x55\x06\x00\xAD\x45\x06\x00\xAE\x45\x06\x00\xAF\x75"
-     "\x06\x00"
-     "\x84\xA8\x01\x00\x56\x44\x00\x80\x5A\x44\x00\x80\x5E\x44\x00\x80\x62\x44"
-     "\x00\x80"
-     "\x66\x44\x00\x80\x6A\x44\x00\x80\x6E\x44\x00\x80\xB8\xA1\x01\x00\xB9\xA1"
-     "\x01\x00"
-     "\xBA\x69\x01\x00\xBB\x69\x01\x00\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x69"
-     "\x01\x00"
-     "\xBF\x69\x01\x00\xB0\x0D\x06\x00\xB1\xE9\x01\x00\xB2\xFD\x01\x00\xB3\xF5"
-     "\x01\x00"
-     "\xB4\xD1\x01\x00\xB5\xD1\x01\x00\xB6\xA9\x01\x00\xB7\xA1\x01\x00\xA3\xAD"
-     "\x05\x00"
-     "\x72\x44\x00\x80\x76\x44\x00\x80\x7A\x44\x00\x80\x7E\x44\x00\x80\xA6\xC5"
-     "\x05\x00"
-     "\xA5\xDD\x05\x00\x82\x44\x00\x80\xAB\x2D\x02\x00\xAA\xC9\x05\x00\x86\x44"
-     "\x00\x80"
-     "\x8A\x44\x00\x80\xAF\x15\x02\x00\xAE\x15\x02\x00\xAD\x21\x02\x00\xAC\x29"
-     "\x02\x00"
-     "\x85\x78\x5C\x00\x8E\x44\x00\x80\x92\x44\x00\x80\x96\x44\x00\x80\x9A\x44"
-     "\x00\x80"
-     "\x9E\x44\x00\x80\xA2\x44\x00\x80\xA6\x44\x00\x80\x82\x05\x00\x00\xAA\x44"
-     "\x00\x80"
-     "\x80\x31\x00\x00\x81\x31\x00\x00\xBE\x3C\x03\x00\xB2\x44\x00\x80\xB6\x44"
-     "\x00\x80"
-     "\xBA\x44\x00\x80\x86\x20\x7C\x00\x87\xB4\x03\x00\xBE\x44\x00\x80\xC2\x44"
-     "\x00\x80"
-     "\xEF\x18\x02\x00\xC6\x44\x00\x80\xCA\x44\x00\x80\xCE\x44\x00\x80\xE3\x10"
-     "\x01\x00"
-     "\xD2\x44\x00\x80\xE1\xF8\x01\x00\xD6\x44\x00\x80\xDA\x44\x00\x80\xDE\x44"
-     "\x00\x80"
-     "\xE2\x44\x00\x80\xE6\x44\x00\x80\xEA\x44\x00\x80\xB3\x9D\x03\x00\xEE\x44"
-     "\x00\x80"
-     "\xF2\x44\x00\x80\xB6\x75\x03\x00\xF6\x44\x00\x80\xFA\x44\x00\x80\xB5\xB1"
-     "\x03\x00"
-     "\xBA\x51\x03\x00\xBB\x51\x03\x00\xFE\x44\x00\x80\x02\x45\x00\x80\xBE\xF5"
-     "\x00\x00"
-     "\xBF\xFD\x00\x00\xBC\x41\x03\x00\xBD\x41\x03\x00\xA8\x41\x02\x00\xA9\x51"
-     "\x02\x00"
-     "\xAA\x51\x02\x00\xAB\x65\x02\x00\xAC\x7D\x02\x00\xAD\xB1\x02\x00\xAE\xB1"
-     "\x02\x00"
-     "\xAF\xB1\x02\x00\xBE\x68\x7D\x00\x06\x45\x00\x80\x0A\x45\x00\x80\x0E\x45"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\xA5\x00\x00\x82\xAD\x00\x00\x12\x45\x00\x80\xB8\x5D"
-     "\x01\x00"
-     "\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x09\x01\x00\xBC\x19\x01\x00\xBD\x19"
-     "\x01\x00"
-     "\xBE\x09\x01\x00\xBF\x09\x01\x00\xB0\xD1\x02\x00\xB1\xD1\x02\x00\xB2\xD1"
-     "\x02\x00"
-     "\xB3\xD1\x02\x00\xB4\x75\x01\x00\xB5\x7D\x01\x00\xB6\x75\x01\x00\xB7\x6D"
-     "\x01\x00"
-     "\xE1\x50\x06\x00\xE1\x4C\x07\x00\xE3\xB0\x00\x00\xE3\x54\x07\x00\x16\x45"
-     "\x00\x80"
-     "\x1A\x45\x00\x80\x84\xD4\x7D\x00\x22\x45\x00\x80\x26\x45\x00\x80\x2A\x45"
-     "\x00\x80"
-     "\x2E\x45\x00\x80\x32\x45\x00\x80\xBE\x30\x7F\x00\x36\x45\x00\x80\xEF\x5C"
-     "\x00\x00"
-     "\xEF\x78\x06\x00\xA3\x99\x02\x00\x3A\x45\x00\x80\x86\x68\x7C\x00\x87\x04"
-     "\x7D\x00"
-     "\x3E\x45\x00\x80\xA6\x71\x02\x00\xA5\xB5\x02\x00\x42\x45\x00\x80\xAB\x55"
-     "\x02\x00"
-     "\xAA\x55\x02\x00\x46\x45\x00\x80\x4A\x45\x00\x80\xAF\xF9\x01\x00\xAE\xF1"
-     "\x01\x00"
-     "\xAD\x45\x02\x00\xAC\x45\x02\x00\xA8\x75\x7E\x00\xA9\x7D\x7E\x00\xAA\x75"
-     "\x7E\x00"
-     "\xAB\x85\x7E\x00\xAC\x9D\x7E\x00\xAD\x81\x7E\x00\xAE\x81\x7E\x00\xAF\x81"
-     "\x7E\x00"
-     "\x1E\x45\x00\x80\x4E\x45\x00\x80\x52\x45\x00\x80\x56\x45\x00\x80\x5A\x45"
-     "\x00\x80"
-     "\x5E\x45\x00\x80\x62\x45\x00\x80\x66\x45\x00\x80\xB8\x8D\x7E\x00\xB9\x91"
-     "\x7E\x00"
-     "\xBA\x91\x7E\x00\xBB\xA5\x7E\x00\xBC\xBD\x7E\x00\xBD\x71\x7F\x00\xBE\x71"
-     "\x7F\x00"
-     "\xBF\x71\x7F\x00\xB0\xC5\x7E\x00\xB1\xC9\x7E\x00\xB2\xD9\x7E\x00\xB3\xD9"
-     "\x7E\x00"
-     "\xB4\xC9\x7E\x00\xB5\xC9\x7E\x00\xB6\xB9\x7E\x00\xB7\xB5\x7E\x00\xB3\xFD"
-     "\x7E\x00"
-     "\x6A\x45\x00\x80\x6E\x45\x00\x80\x72\x45\x00\x80\x76\x45\x00\x80\xB6\x01"
-     "\x7E\x00"
-     "\xB5\x15\x7E\x00\x7A\x45\x00\x80\xBB\x0D\x7E\x00\xBA\x01\x7E\x00\x7E\x45"
-     "\x00\x80"
-     "\x82\x45\x00\x80\xBF\x0D\x7E\x00\xBE\x0D\x7E\x00\xBD\x0D\x7E\x00\xBC\x0D"
-     "\x7E\x00"
-     "\x82\x19\x00\x00\xA3\xB9\x7E\x00\x80\x65\x00\x00\x81\x19\x00\x00\xA6\x45"
-     "\x7E\x00"
-     "\x86\x45\x00\x80\x8A\x45\x00\x80\xA5\x51\x7E\x00\xAA\x45\x7E\x00\xAB\x49"
-     "\x7E\x00"
-     "\x84\xA4\x03\x00\x8E\x45\x00\x80\xAE\x49\x7E\x00\xAF\x49\x7E\x00\xAC\x49"
-     "\x7E\x00"
-     "\xAD\x49\x7E\x00\xA8\x9D\x7E\x00\xA9\xA5\x7E\x00\xAA\xAD\x7E\x00\xAB\xA5"
-     "\x7E\x00"
-     "\xAC\xBD\x7E\x00\xAD\xA1\x7E\x00\xAE\xDD\x7E\x00\xAF\xCD\x7E\x00\x86\x20"
-     "\x01\x00"
-     "\x87\x64\x01\x00\x92\x45\x00\x80\x96\x45\x00\x80\x9A\x45\x00\x80\x9E\x45"
-     "\x00\x80"
-     "\xA2\x45\x00\x80\xA6\x45\x00\x80\xB8\x79\x01\x00\xB9\x79\x01\x00\xBA\xCD"
-     "\x01\x00"
-     "\xBB\xC5\x01\x00\xBC\xDD\x01\x00\xBD\xC5\x01\x00\xBE\xC5\x01\x00\xBF\xF5"
-     "\x01\x00"
-     "\xB0\xBD\x7E\x00\xB1\x8D\x7E\x00\xB2\x85\x7E\x00\xB3\x49\x01\x00\xB4\x59"
-     "\x01\x00"
-     "\xB5\x59\x01\x00\xB6\x49\x01\x00\xB7\x49\x01\x00\xB3\x25\x7E\x00\xAA\x45"
-     "\x00\x80"
-     "\xAE\x45\x00\x80\xB2\x45\x00\x80\xB6\x45\x00\x80\xB6\x21\x7E\x00\xB5\x35"
-     "\x7E\x00"
-     "\xBA\x45\x00\x80\xBB\x49\x7E\x00\xBA\x79\x7E\x00\xBE\x45\x00\x80\xC2\x45"
-     "\x00\x80"
-     "\xBF\x9D\x01\x00\xBE\x9D\x01\x00\xBD\x9D\x01\x00\xBC\x51\x7E\x00\xC6\x45"
-     "\x00\x80"
-     "\xCA\x45\x00\x80\xCE\x45\x00\x80\xD2\x45\x00\x80\xD6\x45\x00\x80\xDA\x45"
-     "\x00\x80"
-     "\xDE\x45\x00\x80\xEF\x04\x7C\x00\xE2\x45\x00\x80\xE1\x7C\x7D\x00\xE6\x45"
-     "\x00\x80"
-     "\xE3\x64\x7C\x00\xEA\x45\x00\x80\xEE\x45\x00\x80\xF2\x45\x00\x80\xF6\x45"
-     "\x00\x80"
-     "\xFA\x45\x00\x80\x84\xD8\x03\x00\xBE\xC8\x02\x00\xA3\xA9\x7D\x00\x02\x46"
-     "\x00\x80"
-     "\xA5\xB9\x7D\x00\xA6\xAD\x7D\x00\x06\x46\x00\x80\x0A\x46\x00\x80\x0E\x46"
-     "\x00\x80"
-     "\xAA\xF5\x7D\x00\xAB\xC5\x7D\x00\xAC\xDD\x7D\x00\xAD\x11\x02\x00\xAE\x11"
-     "\x02\x00"
-     "\xAF\x11\x02\x00\x80\xB9\x03\x00\x81\xB9\x03\x00\x82\x49\x00\x00\xA3\x7D"
-     "\x03\x00"
-     "\x12\x46\x00\x80\xA5\x6D\x03\x00\xA6\x65\x03\x00\x16\x46\x00\x80\xAE\x44"
-     "\x00\x80"
-     "\x1A\x46\x00\x80\xAA\xE5\x03\x00\xAB\xFD\x03\x00\xAC\xED\x03\x00\xAD\xDD"
-     "\x03\x00"
-     "\xAE\xD5\x03\x00\xAF\xBD\x03\x00\x86\x20\x04\x00\x87\xF4\x03\x00\x1E\x46"
-     "\x00\x80"
-     "\x22\x46\x00\x80\x26\x46\x00\x80\x2A\x46\x00\x80\x2E\x46\x00\x80\x32\x46"
-     "\x00\x80"
-     "\xA8\x89\x03\x00\xA9\x89\x03\x00\xAA\x99\x03\x00\xAB\x99\x03\x00\xAC\x89"
-     "\x03\x00"
-     "\xAD\x89\x03\x00\xAE\x3D\x03\x00\xAF\x35\x03\x00\xB0\x4D\x03\x00\xB1\x55"
-     "\x03\x00"
-     "\xB2\x55\x03\x00\xB3\x65\x03\x00\xB4\x7D\x03\x00\xB5\x15\x03\x00\xB6\x1D"
-     "\x03\x00"
-     "\xB7\x15\x03\x00\xB8\x2D\x03\x00\xB9\x35\x03\x00\xBA\x3D\x03\x00\xBB\x35"
-     "\x03\x00"
-     "\xBC\x11\x03\x00\xBD\x1D\x03\x00\xBE\x09\x03\x00\xBF\x09\x03\x00\x36\x46"
-     "\x00\x80"
-     "\xB6\x39\x02\x00\xB5\x31\x02\x00\x3A\x46\x00\x80\xB3\x21\x02\x00\x3E\x46"
-     "\x00\x80"
-     "\x42\x46\x00\x80\x46\x46\x00\x80\xBF\xE1\x02\x00\xBE\x89\x02\x00\xBD\x81"
-     "\x02\x00"
-     "\xBC\xB1\x02\x00\xBB\xA1\x02\x00\xBA\xB9\x02\x00\xCF\x00\x00\x00\x4A\x46"
-     "\x00\x80"
-     "\x4E\x46\x00\x80\x52\x46\x00\x80\x56\x46\x00\x80\x84\x40\x02\x00\xEF\xA0"
-     "\x03\x00"
-     "\x5A\x46\x00\x80\x5E\x46\x00\x80\x62\x46\x00\x80\xE3\x74\x03\x00\x66\x46"
-     "\x00\x80"
-     "\xE1\x50\x00\x00\x6A\x46\x00\x80\x80\x2D\x00\x00\x81\x11\x00\x00\x82\x15"
-     "\x00\x00"
-     "\x72\x46\x00\x80\xBE\x8C\x05\x00\xE3\xD0\x7F\x00\x76\x46\x00\x80\xE1\x08"
-     "\x76\x00"
-     "\x7A\x46\x00\x80\x7E\x46\x00\x80\x82\x46\x00\x80\xEF\xD4\x7E\x00\x84\xEC"
-     "\x07\x00"
-     "\xE1\x74\x7E\x00\x86\x46\x00\x80\xE3\x54\x01\x00\x8A\x46\x00\x80\xEF\xD0"
-     "\x66\x00"
-     "\x86\x58\x07\x00\x87\xC8\x04\x00\xB3\xE1\x01\x00\x8E\x46\x00\x80\x92\x46"
-     "\x00\x80"
-     "\x96\x46\x00\x80\x9A\x46\x00\x80\xB6\x89\x01\x00\xB5\x89\x01\x00\x9E\x46"
-     "\x00\x80"
-     "\xBB\x71\x01\x00\xBA\x69\x01\x00\xA2\x46\x00\x80\xA6\x46\x00\x80\xBF\xF1"
-     "\x00\x00"
-     "\xBE\x69\x01\x00\xBD\x61\x01\x00\xBC\x69\x01\x00\xA8\x11\x06\x00\xA9\x19"
-     "\x06\x00"
-     "\xAA\x29\x06\x00\xAB\x29\x06\x00\xAC\x41\x06\x00\xAD\x49\x06\x00\xAE\x7D"
-     "\x06\x00"
-     "\xAF\x95\x06\x00\x6E\x46\x00\x80\xAA\x46\x00\x80\xAE\x46\x00\x80\xB2\x46"
-     "\x00\x80"
-     "\xA2\xDD\x04\x00\xA3\x49\x02\x00\xA0\x59\x06\x00\xA1\x1D\x04\x00\xB8\x59"
-     "\x07\x00"
-     "\xB9\x59\x07\x00\xBA\x75\x07\x00\xBB\x0D\x07\x00\xBC\x15\x07\x00\xBD\x1D"
-     "\x07\x00"
-     "\xBE\x0D\x07\x00\xBF\x05\x07\x00\xB0\xED\x06\x00\xB1\xF5\x06\x00\xB2\xFD"
-     "\x06\x00"
-     "\xB3\xF5\x06\x00\xB4\xED\x06\x00\xB5\x75\x07\x00\xB6\x75\x07\x00\xB7\x69"
-     "\x07\x00"
-     "\xA3\xA1\x06\x00\x8F\x29\x6C\x00\xB6\x46\x00\x80\xBA\x46\x00\x80\xBE\x46"
-     "\x00\x80"
-     "\xA6\xC9\x06\x00\xA5\xC9\x06\x00\xC2\x46\x00\x80\xAB\x31\x06\x00\xAA\x29"
-     "\x06\x00"
-     "\xC6\x46\x00\x80\xCA\x46\x00\x80\xAF\xB1\x07\x00\xAE\x29\x06\x00\xAD\x21"
-     "\x06\x00"
-     "\xAC\x29\x06\x00\x9E\xC5\x78\x00\x9F\xE9\x79\x00\x9C\x81\x7D\x00\x9D\xCD"
-     "\x78\x00"
-     "\x9A\xD5\x7C\x00\x9B\xDD\x7C\x00\x98\xB5\x70\x00\x99\xD9\x7E\x00\x96\xE9"
-     "\x72\x00"
-     "\x97\xE1\x70\x00\x94\xED\x74\x00\x95\xE5\x74\x00\x92\x01\x68\x00\x93\x19"
-     "\x69\x00"
-     "\x90\xD9\x6A\x00\x91\xDD\x68\x00\x82\xAD\x6D\x00\x83\x25\x6B\x00\xCE\x46"
-     "\x00\x80"
-     "\xD2\x46\x00\x80\x86\xFD\x68\x00\x87\x11\x75\x00\x84\x39\x69\x00\x85\xED"
-     "\x69\x00"
-     "\x8A\x29\x75\x00\x8B\x2D\x73\x00\xD6\x46\x00\x80\xDA\x46\x00\x80\x8E\x35"
-     "\x7F\x00"
-     "\x8F\x31\x7D\x00\x8C\x3D\x71\x00\x8D\x51\x71\x00\x92\x49\x7D\x00\x93\x0D"
-     "\x7C\x00"
-     "\xBE\x00\x01\x00\xE2\x46\x00\x80\x96\x7D\x78\x00\x97\x35\x07\x00\x94\x59"
-     "\x79\x00"
-     "\x95\x59\x79\x00\x9A\x41\x05\x00\x9B\x89\x05\x00\x86\xC8\x00\x00\x87\x4C"
-     "\x00\x00"
-     "\x81\x59\x00\x00\x80\x79\x00\x00\x9C\x05\x03\x00\x82\x51\x00\x00\xA7\x0D"
-     "\x09\x00"
-     "\xA6\x05\x09\x00\xA5\xE1\x0C\x00\xA4\x1D\x0D\x00\xA3\x91\x0D\x00\xA2\xED"
-     "\x03\x00"
-     "\xA1\xB9\x01\x00\xE6\x46\x00\x80\xAF\x01\x1C\x00\xAE\x79\x11\x00\xAD\x89"
-     "\x11\x00"
-     "\xAC\x01\x10\x00\xAB\x9D\x14\x00\xAA\x95\x15\x00\xA9\x01\x14\x00\xA8\x6D"
-     "\x09\x00"
-     "\xEF\xB8\x04\x00\xEF\x0C\x00\x00\xB5\x05\x18\x00\xB4\x91\x19\x00\xB3\xD5"
-     "\x19\x00"
-     "\xB2\x01\x18\x00\xB1\xE9\x1D\x00\xB0\xFD\x1D\x00\xEA\x46\x00\x80\xEE\x46"
-     "\x00\x80"
-     "\xF2\x46\x00\x80\xF6\x46\x00\x80\xE3\xAC\x04\x00\xE3\xD0\x00\x00\xE1\x60"
-     "\x05\x00"
-     "\xE1\x8C\x02\x00\xB8\x2D\x02\x00\xB9\x8D\x02\x00\xBA\x85\x02\x00\xBB\x9D"
-     "\x02\x00"
-     "\xBC\x8D\x02\x00\xBD\xBD\x02\x00\xBE\xB5\x02\x00\xBF\x15\x03\x00\xB0\xE9"
-     "\x05\x00"
-     "\xB1\xE9\x05\x00\xB2\xF9\x05\x00\xB3\xF1\x05\x00\xB4\xD9\x05\x00\xB5\xD9"
-     "\x05\x00"
-     "\xB6\x1D\x02\x00\xB7\x15\x02\x00\xA8\x3D\x05\x00\xA9\xCD\x05\x00\xAA\xD9"
-     "\x05\x00"
-     "\xAB\xD9\x05\x00\xAC\xC9\x05\x00\xAD\xFD\x05\x00\xAE\xE9\x05\x00\xAF\x99"
-     "\x05\x00"
-     "\xFA\x46\x00\x80\xFE\x46\x00\x80\x02\x47\x00\x80\x06\x47\x00\x80\x0A\x47"
-     "\x00\x80"
-     "\x0E\x47\x00\x80\xDE\x46\x00\x80\xFE\x45\x00\x80\xB3\x11\x05\x00\x12\x47"
-     "\x00\x80"
-     "\x16\x47\x00\x80\x1A\x47\x00\x80\x1E\x47\x00\x80\xB6\x3D\x05\x00\xB5\x01"
-     "\x05\x00"
-     "\x22\x47\x00\x80\xBB\x6D\x05\x00\xBA\x6D\x05\x00\x26\x47\x00\x80\x2A\x47"
-     "\x00\x80"
-     "\xBF\x5D\x05\x00\xBE\x5D\x05\x00\xBD\x65\x05\x00\xBC\x6D\x05\x00\x2E\x47"
-     "\x00\x80"
-     "\xA3\x55\x05\x00\x32\x47\x00\x80\x36\x47\x00\x80\xA6\x79\x05\x00\x3A\x47"
-     "\x00\x80"
-     "\x3E\x47\x00\x80\xA5\x45\x05\x00\xAA\x29\x05\x00\xAB\x29\x05\x00\x42\x47"
-     "\x00\x80"
-     "\x46\x47\x00\x80\xAE\x19\x05\x00\xAF\x19\x05\x00\xAC\x29\x05\x00\xAD\x21"
-     "\x05\x00"
-     "\xA8\x4D\x02\x00\xA9\x99\x02\x00\xAA\xE9\x02\x00\xAB\xE9\x02\x00\xAC\xF9"
-     "\x02\x00"
-     "\xAD\xF9\x02\x00\xAE\x79\x01\x00\xAF\x79\x01\x00\x80\xC9\x01\x00\x81\xC9"
-     "\x01\x00"
-     "\x82\x05\x00\x00\x4A\x47\x00\x80\xBE\x00\x0C\x00\x84\x30\x02\x00\x52\x47"
-     "\x00\x80"
-     "\x56\x47\x00\x80\xB8\xDD\x01\x00\xB9\xED\x01\x00\xBA\xE5\x01\x00\xBB\x8D"
-     "\x01\x00"
-     "\xBC\x99\x01\x00\xBD\x99\x01\x00\xBE\x8D\x01\x00\xBF\x85\x01\x00\xB0\x09"
-     "\x01\x00"
-     "\xB1\x09\x01\x00\xB2\x19\x01\x00\xB3\x19\x01\x00\xB4\xF5\x01\x00\xB5\xFD"
-     "\x01\x00"
-     "\xB6\xF5\x01\x00\xB7\xED\x01\x00\x86\x00\x0D\x00\x87\x00\x0D\x00\x5A\x47"
-     "\x00\x80"
-     "\x5E\x47\x00\x80\x62\x47\x00\x80\x66\x47\x00\x80\x6A\x47\x00\x80\xEF\x2C"
-     "\x00\x00"
-     "\x6E\x47\x00\x80\xE1\x04\x06\x00\x72\x47\x00\x80\xE3\x1C\x06\x00\xBE\x3C"
-     "\x0C\x00"
-     "\x76\x47\x00\x80\x7A\x47\x00\x80\x7E\x47\x00\x80\x82\x47\x00\x80\x86\x47"
-     "\x00\x80"
-     "\xB3\x29\x02\x00\x8A\x47\x00\x80\xB5\x39\x02\x00\x8E\x47\x00\x80\x92\x47"
-     "\x00\x80"
-     "\xB6\x31\x02\x00\x96\x47\x00\x80\x9A\x47\x00\x80\xBB\x85\x01\x00\xBA\x85"
-     "\x01\x00"
-     "\xBD\x85\x01\x00\xBC\x85\x01\x00\xBF\x71\x01\x00\xBE\x79\x01\x00\x4E\x47"
-     "\x00\x80"
-     "\x9E\x47\x00\x80\xA2\x47\x00\x80\xA6\x47\x00\x80\xAA\x47\x00\x80\xAE\x47"
-     "\x00\x80"
-     "\xB2\x47\x00\x80\xB6\x47\x00\x80\xA8\x19\x0F\x00\xA9\x19\x0F\x00\xAA\xA1"
-     "\x0F\x00"
-     "\xAB\xA1\x0F\x00\xAC\xA1\x0F\x00\xAD\xA1\x0F\x00\xAE\xA1\x0F\x00\xAF\x9D"
-     "\x0F\x00"
-     "\xB0\xE9\x0F\x00\xB1\xF1\x0F\x00\xB2\xF1\x0F\x00\xB3\xF1\x0F\x00\xB4\x91"
-     "\x0F\x00"
-     "\xB5\x91\x0F\x00\xB6\x95\x0F\x00\xB7\x89\x0F\x00\xB8\xB9\x0F\x00\xB9\x85"
-     "\x0F\x00"
-     "\xBA\x8D\x0F\x00\xBB\x69\x00\x00\xBC\x79\x00\x00\xBD\x79\x00\x00\xBE\x69"
-     "\x00\x00"
-     "\xBF\x69\x00\x00\x81\x9D\x00\x00\x80\x9D\x00\x00\xBA\x47\x00\x80\x82\x05"
-     "\x00\x00"
-     "\xBE\x47\x00\x80\xC2\x47\x00\x80\x84\x94\x02\x00\xC6\x47\x00\x80\x86\x00"
-     "\x00\x00"
-     "\x87\xF4\x03\x00\xCA\x47\x00\x80\xCE\x47\x00\x80\xD2\x47\x00\x80\xD6\x47"
-     "\x00\x80"
-     "\xDA\x47\x00\x80\xDE\x47\x00\x80\xB3\x49\x0F\x00\xE2\x47\x00\x80\xE6\x47"
-     "\x00\x80"
-     "\xEA\x47\x00\x80\xEE\x47\x00\x80\xB6\x41\x0F\x00\xB5\x49\x0F\x00\xF2\x47"
-     "\x00\x80"
-     "\xBB\x55\x0F\x00\xBA\x4D\x0F\x00\xF6\x47\x00\x80\xFA\x47\x00\x80\xBF\x35"
-     "\x0F\x00"
-     "\xBE\x35\x0F\x00\xBD\x45\x0F\x00\xBC\x45\x0F\x00\xA3\x89\x0E\x00\xFE\x47"
-     "\x00\x80"
-     "\x02\x48\x00\x80\x06\x48\x00\x80\x0A\x48\x00\x80\xA6\x81\x0E\x00\xA5\x89"
-     "\x0E\x00"
-     "\x0E\x48\x00\x80\xAB\x95\x0E\x00\xAA\x8D\x0E\x00\x12\x48\x00\x80\x16\x48"
-     "\x00\x80"
-     "\xAF\xF5\x0E\x00\xAE\xF5\x0E\x00\xAD\x85\x0E\x00\xAC\x85\x0E\x00\x1A\x48"
-     "\x00\x80"
-     "\x1E\x48\x00\x80\x22\x48\x00\x80\x26\x48\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x2A\x48\x00\x80\x84\x6C\x02\x00\xBE\x48\x01\x00\x2E\x48"
-     "\x00\x80"
-     "\x32\x48\x00\x80\x86\x00\x04\x00\x87\xB4\x01\x00\x36\x48\x00\x80\x3A\x48"
-     "\x00\x80"
-     "\xAC\xF1\x0F\x00\xAD\xF1\x0F\x00\xAE\x0D\x0F\x00\xAF\x05\x0F\x00\x3E\x48"
-     "\x00\x80"
-     "\x42\x48\x00\x80\xAA\xF1\x0F\x00\xAB\xF1\x0F\x00\x46\x48\x00\x80\xA5\x4D"
-     "\x0C\x00"
-     "\xA6\x45\x0C\x00\x4A\x48\x00\x80\x4E\x48\x00\x80\x52\x48\x00\x80\x56\x48"
-     "\x00\x80"
-     "\xA3\x5D\x0C\x00\xB3\x69\x01\x00\x5A\x48\x00\x80\x5E\x48\x00\x80\x62\x48"
-     "\x00\x80"
-     "\x66\x48\x00\x80\xB6\x61\x01\x00\xB5\x69\x01\x00\x6A\x48\x00\x80\xBB\x29"
-     "\x01\x00"
-     "\xBA\x21\x01\x00\x6E\x48\x00\x80\x72\x48\x00\x80\xBF\x19\x01\x00\xBE\x29"
-     "\x01\x00"
-     "\xBD\x29\x01\x00\xBC\x31\x01\x00\xA8\x95\x02\x00\xA9\x99\x02\x00\xAA\xA1"
-     "\x02\x00"
-     "\xAB\xA1\x02\x00\xAC\xA1\x02\x00\xAD\xA1\x02\x00\xAE\xA1\x02\x00\xAF\xA1"
-     "\x02\x00"
-     "\x76\x48\x00\x80\x7A\x48\x00\x80\x7E\x48\x00\x80\x82\x48\x00\x80\x86\x48"
-     "\x00\x80"
-     "\x8A\x48\x00\x80\x8E\x48\x00\x80\x92\x48\x00\x80\xB8\x49\x03\x00\xB9\x49"
-     "\x03\x00"
-     "\xBA\x59\x03\x00\xBB\x51\x03\x00\xBC\x79\x03\x00\xBD\x79\x03\x00\xBE\x19"
-     "\x03\x00"
-     "\xBF\x19\x03\x00\xB0\xA5\x02\x00\xB1\xAD\x02\x00\xB2\xA5\x02\x00\xB3\xBD"
-     "\x02\x00"
-     "\xB4\xAD\x02\x00\xB5\x9D\x02\x00\xB6\x95\x02\x00\xB7\x79\x03\x00\xAB\x49"
-     "\x02\x00"
-     "\xAA\x41\x02\x00\x84\x04\x03\x00\x96\x48\x00\x80\xAF\x79\x02\x00\xAE\x49"
-     "\x02\x00"
-     "\xAD\x49\x02\x00\xAC\x51\x02\x00\xA3\x09\x02\x00\x9A\x48\x00\x80\xBE\xF0"
-     "\x03\x00"
-     "\xA2\x48\x00\x80\xA6\x48\x00\x80\xA6\x01\x02\x00\xA5\x09\x02\x00\xAA\x48"
-     "\x00\x80"
-     "\x80\xFD\x00\x00\x81\x49\x00\x00\x82\x6D\x00\x00\xAE\x48\x00\x80\xB2\x48"
-     "\x00\x80"
-     "\xEF\x84\x02\x00\xB6\x48\x00\x80\xBA\x48\x00\x80\xBE\x48\x00\x80\xE3\x7C"
-     "\x02\x00"
-     "\xC2\x48\x00\x80\xE1\x84\x01\x00\xC6\x48\x00\x80\xCA\x48\x00\x80\x86\x58"
-     "\x04\x00"
-     "\x87\x64\x03\x00\xCE\x48\x00\x80\xD2\x48\x00\x80\xD6\x48\x00\x80\xBE\x40"
-     "\x05\x00"
-     "\xDA\x48\x00\x80\xE1\xAC\x01\x00\xDE\x48\x00\x80\xE3\x6C\x00\x00\xE2\x48"
-     "\x00\x80"
-     "\xE6\x48\x00\x80\xEA\x48\x00\x80\xEF\xB0\x00\x00\x84\x78\x05\x00\xEE\x48"
-     "\x00\x80"
-     "\xF2\x48\x00\x80\xF6\x48\x00\x80\xB3\x3D\x02\x00\xFA\x48\x00\x80\xFE\x48"
-     "\x00\x80"
-     "\x02\x49\x00\x80\xBE\x18\x04\x00\xB6\xF5\x02\x00\xB5\xCD\x02\x00\x06\x49"
-     "\x00\x80"
-     "\xBB\xD9\x02\x00\xBA\xD1\x02\x00\xD2\x00\x00\x00\x0A\x49\x00\x80\xBF\xA1"
-     "\x02\x00"
-     "\xBE\xB9\x02\x00\xBD\xB1\x02\x00\xBC\xB9\x02\x00\x12\x49\x00\x80\xE1\x1C"
-     "\x0E\x00"
-     "\x16\x49\x00\x80\xE3\x04\x01\x00\x82\xFD\x00\x00\x1A\x49\x00\x80\x80\xD5"
-     "\x00\x00"
-     "\x81\xC1\x00\x00\x87\x6C\x05\x00\x86\x04\x04\x00\x1E\x49\x00\x80\x22\x49"
-     "\x00\x80"
-     "\x26\x49\x00\x80\x2A\x49\x00\x80\x2E\x49\x00\x80\xEF\xF4\x0E\x00\xA6\x61"
-     "\x02\x00"
-     "\x32\x49\x00\x80\x36\x49\x00\x80\xA5\x59\x02\x00\x3A\x49\x00\x80\xA3\xA9"
-     "\x02\x00"
-     "\x3E\x49\x00\x80\x42\x49\x00\x80\xAE\x2D\x02\x00\xAF\x35\x02\x00\xAC\x2D"
-     "\x02\x00"
-     "\xAD\x25\x02\x00\xAA\x45\x02\x00\xAB\x4D\x02\x00\x46\x49\x00\x80\x4A\x49"
-     "\x00\x80"
-     "\xA8\xAD\x06\x00\xA9\xB5\x06\x00\xAA\xB5\x06\x00\xAB\xC1\x06\x00\xAC\xC1"
-     "\x06\x00"
-     "\xAD\xCD\x06\x00\xAE\xC5\x06\x00\xAF\xFD\x06\x00\x4E\x49\x00\x80\x52\x49"
-     "\x00\x80"
-     "\x56\x49\x00\x80\x5A\x49\x00\x80\x5E\x49\x00\x80\x62\x49\x00\x80\x66\x49"
-     "\x00\x80"
-     "\x6A\x49\x00\x80\xB8\x61\x07\x00\xB9\x61\x07\x00\xBA\x61\x07\x00\xBB\x79"
-     "\x07\x00"
-     "\xBC\x69\x07\x00\xBD\x69\x07\x00\xBE\x1D\x07\x00\xBF\x15\x07\x00\xB0\x89"
-     "\x06\x00"
-     "\xB1\x95\x06\x00\xB2\x9D\x06\x00\xB3\x6D\x07\x00\xB4\x75\x07\x00\xB5\x7D"
-     "\x07\x00"
-     "\xB6\x69\x07\x00\xB7\x65\x07\x00\x9E\x48\x00\x80\xB3\x39\x06\x00\x6E\x49"
-     "\x00\x80"
-     "\x0E\x49\x00\x80\xB6\x49\x06\x00\x72\x49\x00\x80\x76\x49\x00\x80\xB5\x4D"
-     "\x06\x00"
-     "\xBA\x49\x06\x00\xBB\x55\x06\x00\x7A\x49\x00\x80\x84\xE0\x03\x00\xBE\xB5"
-     "\x07\x00"
-     "\xBF\xBD\x07\x00\xBC\x4D\x06\x00\xBD\xBD\x07\x00\xA3\x7D\x06\x00\x82\x6D"
-     "\x00\x00"
-     "\x81\x51\x00\x00\x80\x59\x00\x00\x7E\x49\x00\x80\xA6\x0D\x06\x00\xA5\x09"
-     "\x06\x00"
-     "\x82\x49\x00\x80\xAB\x11\x06\x00\xAA\x0D\x06\x00\x86\x48\x00\x00\x87\x68"
-     "\x00\x00"
-     "\xAF\xF9\x07\x00\xAE\xF1\x07\x00\xAD\xF9\x07\x00\xAC\x09\x06\x00\xA8\x81"
-     "\x06\x00"
-     "\xA9\x8D\x06\x00\xAA\x9D\x06\x00\xAB\xA5\x06\x00\xAC\xBD\x06\x00\xAD\xA5"
-     "\x06\x00"
-     "\xAE\xAD\x06\x00\xAF\xA5\x06\x00\x86\x49\x00\x80\x8A\x49\x00\x80\x8E\x49"
-     "\x00\x80"
-     "\x92\x49\x00\x80\x96\x49\x00\x80\x9A\x49\x00\x80\x9E\x49\x00\x80\xA2\x49"
-     "\x00\x80"
-     "\xB8\x65\x01\x00\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x7D\x01\x00\xBC\x65"
-     "\x01\x00"
-     "\xBD\x6D\x01\x00\xBE\x65\x01\x00\xBF\xD5\x01\x00\xB0\xA5\x06\x00\xB1\xAD"
-     "\x06\x00"
-     "\xB2\xB9\x06\x00\xB3\xB5\x06\x00\xB4\xAD\x06\x00\xB5\x8D\x06\x00\xB6\x85"
-     "\x06\x00"
-     "\xB7\x5D\x01\x00\xB3\x15\x06\x00\xA6\x49\x00\x80\xAA\x49\x00\x80\xAE\x49"
-     "\x00\x80"
-     "\xB2\x49\x00\x80\xB6\x25\x06\x00\xB5\x35\x06\x00\xB6\x49\x00\x80\xBB\x69"
-     "\x06\x00"
-     "\xBA\x61\x06\x00\xBA\x49\x00\x80\xBE\x49\x00\x80\xBF\x49\x06\x00\xBE\x69"
-     "\x06\x00"
-     "\xBD\x61\x06\x00\xBC\x6D\x06\x00\xC2\x49\x00\x80\xA3\x51\x06\x00\xC6\x49"
-     "\x00\x80"
-     "\xCA\x49\x00\x80\xA6\x61\x06\x00\xCE\x49\x00\x80\xD2\x49\x00\x80\xA5\x71"
-     "\x06\x00"
-     "\xAA\x25\x06\x00\xAB\x2D\x06\x00\xD6\x49\x00\x80\xDA\x49\x00\x80\xAE\x2D"
-     "\x06\x00"
-     "\xAF\x0D\x06\x00\xAC\x29\x06\x00\xAD\x25\x06\x00\xA8\xDD\x02\x00\xA9\x25"
-     "\x03\x00"
-     "\xAA\x2D\x03\x00\xAB\x25\x03\x00\xAC\x3D\x03\x00\xAD\x25\x03\x00\xAE\x2D"
-     "\x03\x00"
-     "\xAF\x25\x03\x00\xDE\x49\x00\x80\xE2\x49\x00\x80\xE6\x49\x00\x80\xEA\x49"
-     "\x00\x80"
-     "\xBE\x7C\x02\x00\xF2\x49\x00\x80\xF6\x49\x00\x80\xFA\x49\x00\x80\xB8\x81"
-     "\x03\x00"
-     "\xB9\x8D\x03\x00\xBA\x85\x03\x00\xBB\x99\x03\x00\xBC\x89\x03\x00\xBD\xBD"
-     "\x03\x00"
-     "\xBE\xB5\x03\x00\xBF\x7D\x00\x00\xB0\x5D\x03\x00\xB1\xE1\x03\x00\xB2\xE1"
-     "\x03\x00"
-     "\xB3\xE1\x03\x00\xB4\xE1\x03\x00\xB5\xE9\x03\x00\xB6\xD1\x03\x00\xB7\xD1"
-     "\x03\x00"
-     "\x81\xA9\x00\x00\x80\xA1\x00\x00\xB3\xAD\x03\x00\x82\xA1\x00\x00\xB5\xBD"
-     "\x03\x00"
-     "\xFE\x49\x00\x80\x02\x4A\x00\x80\xB6\xAD\x03\x00\x84\xE4\x02\x00\x06\x4A"
-     "\x00\x80"
-     "\xBB\x2D\x03\x00\xBA\x91\x03\x00\xBD\x3D\x03\x00\xBC\x3D\x03\x00\xBF\x25"
-     "\x03\x00"
-     "\xBE\x2D\x03\x00\xA3\xE5\x03\x00\x0A\x4A\x00\x80\x87\xA8\x03\x00\x86\x6C"
-     "\x1C\x00"
-     "\x0E\x4A\x00\x80\xA6\xE5\x03\x00\xA5\xF5\x03\x00\x12\x4A\x00\x80\xAB\x65"
-     "\x03\x00"
-     "\xAA\xD9\x03\x00\x16\x4A\x00\x80\x1A\x4A\x00\x80\xAF\x6D\x03\x00\xAE\x65"
-     "\x03\x00"
-     "\xAD\x75\x03\x00\xAC\x75\x03\x00\xE3\xBC\x01\x00\xE1\xE4\x07\x00\xE1\xB4"
-     "\x06\x00"
-     "\xE3\x2C\x06\x00\x1E\x4A\x00\x80\x22\x4A\x00\x80\x26\x4A\x00\x80\x2A\x4A"
-     "\x00\x80"
-     "\x2E\x4A\x00\x80\x32\x4A\x00\x80\x36\x4A\x00\x80\x3A\x4A\x00\x80\xEF\x98"
-     "\x06\x00"
-     "\x3E\x4A\x00\x80\x42\x4A\x00\x80\xEF\xC8\x06\x00\x46\x4A\x00\x80\x4A\x4A"
-     "\x00\x80"
-     "\x84\x04\x1D\x00\xBE\x04\x1D\x00\x4E\x4A\x00\x80\xE1\xA4\x01\x00\x52\x4A"
-     "\x00\x80"
-     "\xE3\xEC\x03\x00\x80\x39\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\xEF\xA4"
-     "\x03\x00"
-     "\x56\x4A\x00\x80\x5A\x4A\x00\x80\x87\x58\x1C\x00\x86\xDC\x1C\x00\xBC\xAD"
-     "\x05\x00"
-     "\xBD\xB5\x05\x00\xBE\xBD\x05\x00\xBF\xB5\x05\x00\xB8\x7D\x06\x00\xB9\xB5"
-     "\x05\x00"
-     "\xBA\xBD\x05\x00\xBB\xB5\x05\x00\xB4\x6D\x06\x00\xB5\x75\x06\x00\xB6\x75"
-     "\x06\x00"
-     "\xB7\x45\x06\x00\xB0\x05\x06\x00\xB1\x0D\x06\x00\xB2\x7D\x06\x00\xB3\x75"
-     "\x06\x00"
-     "\xAC\x55\x06\x00\xAD\x65\x06\x00\xAE\x65\x06\x00\xAF\x7D\x06\x00\xA8\xBD"
-     "\x06\x00"
-     "\xA9\xAD\x06\x00\xAA\xA5\x06\x00\xAB\x4D\x06\x00\x62\x4A\x00\x80\x66\x4A"
-     "\x00\x80"
-     "\x6A\x4A\x00\x80\x6E\x4A\x00\x80\x72\x4A\x00\x80\x76\x4A\x00\x80\x7A\x4A"
-     "\x00\x80"
-     "\x7E\x4A\x00\x80\xA8\x35\x1E\x00\xA9\x3D\x1E\x00\xAA\x35\x1E\x00\xAB\x99"
-     "\x1E\x00"
-     "\xAC\x89\x1E\x00\xAD\xB1\x1E\x00\xAE\xDD\x1E\x00\xAF\xCD\x1E\x00\x5E\x4A"
-     "\x00\x80"
-     "\x82\x4A\x00\x80\x86\x4A\x00\x80\x8A\x4A\x00\x80\x8E\x4A\x00\x80\x92\x4A"
-     "\x00\x80"
-     "\x96\x4A\x00\x80\x9A\x4A\x00\x80\xB8\x65\x1F\x00\xB9\x6D\x1F\x00\xBA\x65"
-     "\x1F\x00"
-     "\xBB\x79\x1F\x00\xBC\x69\x1F\x00\xBD\x15\x1F\x00\xBE\x1D\x1F\x00\xBF\x0D"
-     "\x1F\x00"
-     "\xB0\xBD\x1E\x00\xB1\x8D\x1E\x00\xB2\x85\x1E\x00\xB3\x7D\x1F\x00\xB4\x65"
-     "\x1F\x00"
-     "\xB5\x6D\x1F\x00\xB6\x65\x1F\x00\xB7\x5D\x1F\x00\xB3\xBD\x1E\x00\x9E\x4A"
-     "\x00\x80"
-     "\xA2\x4A\x00\x80\xA6\x4A\x00\x80\xAA\x4A\x00\x80\xB6\x31\x1E\x00\xB5\x01"
-     "\x1E\x00"
-     "\xAE\x4A\x00\x80\xBB\x49\x1E\x00\xBA\x79\x1E\x00\xB2\x4A\x00\x80\xB6\x4A"
-     "\x00\x80"
-     "\xBF\xA9\x1F\x00\xBE\xA1\x1F\x00\xBD\xA9\x1F\x00\xBC\x51\x1E\x00\xBA\x4A"
-     "\x00\x80"
-     "\xA3\xF9\x1E\x00\xBE\x4A\x00\x80\xC2\x4A\x00\x80\xA6\x75\x1E\x00\xC6\x4A"
-     "\x00\x80"
-     "\xCA\x4A\x00\x80\xA5\x45\x1E\x00\xAA\x3D\x1E\x00\xAB\x0D\x1E\x00\x84\xB4"
-     "\x03\x00"
-     "\xBE\xA0\x03\x00\xAE\xE5\x1F\x00\xAF\xED\x1F\x00\xAC\x15\x1E\x00\xAD\xED"
-     "\x1F\x00"
-     "\xA8\xE1\x1E\x00\xA9\xE1\x1E\x00\xAA\xE1\x1E\x00\xAB\xE1\x1E\x00\xAC\xE1"
-     "\x1E\x00"
-     "\xAD\xE1\x1E\x00\xAE\xE1\x1E\x00\xAF\xE1\x1E\x00\x80\xED\x01\x00\x81\xF1"
-     "\x01\x00"
-     "\x82\xF1\x01\x00\xCE\x4A\x00\x80\x86\x90\x00\x00\x87\x74\x01\x00\xD2\x4A"
-     "\x00\x80"
-     "\xD6\x4A\x00\x80\xB8\xC9\x01\x00\xB9\xC9\x01\x00\xBA\xD9\x01\x00\xBB\xD1"
-     "\x01\x00"
-     "\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\x99\x01\x00\xBF\x95\x01\x00\xB0\x45"
-     "\x01\x00"
-     "\xB1\x4D\x01\x00\xB2\x45\x01\x00\xB3\x5D\x01\x00\xB4\x45\x01\x00\xB5\x4D"
-     "\x01\x00"
-     "\xB6\x45\x01\x00\xB7\xF9\x01\x00\xDA\x4A\x00\x80\xB6\x71\x1E\x00\xB5\x79"
-     "\x1E\x00"
-     "\xDE\x4A\x00\x80\xB3\x79\x1E\x00\xE2\x4A\x00\x80\xE6\x4A\x00\x80\xEA\x4A"
-     "\x00\x80"
-     "\xBF\x99\x01\x00\xBE\x91\x01\x00\xBD\x99\x01\x00\xBC\x81\x01\x00\xBB\x99"
-     "\x01\x00"
-     "\xBA\x91\x01\x00\xEE\x49\x00\x80\xEE\x4A\x00\x80\xA3\x29\x1E\x00\xF2\x4A"
-     "\x00\x80"
-     "\xF6\x4A\x00\x80\xFA\x4A\x00\x80\xFE\x4A\x00\x80\xA6\x21\x1E\x00\xA5\x29"
-     "\x1E\x00"
-     "\x02\x4B\x00\x80\xAB\xC9\x01\x00\xAA\xC1\x01\x00\x06\x4B\x00\x80\x0A\x4B"
-     "\x00\x80"
-     "\xAF\xC9\x01\x00\xAE\xC1\x01\x00\xAD\xC9\x01\x00\xAC\xD1\x01\x00\x0E\x4B"
-     "\x00\x80"
-     "\xEF\xB0\x1F\x00\x12\x4B\x00\x80\x16\x4B\x00\x80\x1A\x4B\x00\x80\x84\x94"
-     "\x02\x00"
-     "\x1E\x4B\x00\x80\x22\x4B\x00\x80\x26\x4B\x00\x80\x2A\x4B\x00\x80\x2E\x4B"
-     "\x00\x80"
-     "\x32\x4B\x00\x80\x36\x4B\x00\x80\xE3\x6C\x1F\x00\x3A\x4B\x00\x80\xE1\xE8"
-     "\x1E\x00"
-     "\x80\x6D\x00\x00\x81\x75\x00\x00\x82\x71\x00\x00\xB3\xCD\x03\x00\x42\x4B"
-     "\x00\x80"
-     "\xB5\xC1\x03\x00\xB6\xC1\x03\x00\xBE\x10\x02\x00\x86\xC0\x04\x00\x87\xE0"
-     "\x03\x00"
-     "\xBA\xAD\x03\x00\xBB\xA5\x03\x00\xBC\xBD\x03\x00\xBD\xA5\x03\x00\xBE\xA5"
-     "\x03\x00"
-     "\xBF\x95\x03\x00\x46\x4B\x00\x80\x4A\x4B\x00\x80\x4E\x4B\x00\x80\x52\x4B"
-     "\x00\x80"
-     "\xEF\x34\x02\x00\x56\x4B\x00\x80\x5A\x4B\x00\x80\x5E\x4B\x00\x80\xE3\xA4"
-     "\x02\x00"
-     "\x62\x4B\x00\x80\xE1\x80\x01\x00\x66\x4B\x00\x80\x6A\x4B\x00\x80\x6E\x4B"
-     "\x00\x80"
-     "\x72\x4B\x00\x80\x76\x4B\x00\x80\xA3\x41\x03\x00\x7A\x4B\x00\x80\x7E\x4B"
-     "\x00\x80"
-     "\x82\x4B\x00\x80\x86\x4B\x00\x80\xA6\x4D\x03\x00\xA5\x4D\x03\x00\x8A\x4B"
-     "\x00\x80"
-     "\xAB\x29\x03\x00\xAA\x21\x03\x00\x8E\x4B\x00\x80\x92\x4B\x00\x80\xAF\x19"
-     "\x03\x00"
-     "\xAE\x29\x03\x00\xAD\x29\x03\x00\xAC\x31\x03\x00\xA8\x39\x02\x00\xA9\x39"
-     "\x02\x00"
-     "\xAA\x89\x02\x00\xAB\x81\x02\x00\xAC\x85\x02\x00\xAD\x8D\x02\x00\xAE\x85"
-     "\x02\x00"
-     "\xAF\xA1\x02\x00\x84\x2C\x02\x00\x96\x4B\x00\x80\xBF\x00\x2C\x00\x9A\x4B"
-     "\x00\x80"
-     "\x9E\x4B\x00\x80\xA2\x4B\x00\x80\xA6\x4B\x00\x80\xAA\x4B\x00\x80\xB8\x49"
-     "\x01\x00"
-     "\xB9\x49\x01\x00\xBA\x59\x01\x00\xBB\x55\x01\x00\xBC\x4D\x01\x00\xBD\x75"
-     "\x01\x00"
-     "\xBE\x7D\x01\x00\xBF\x75\x01\x00\xB0\xA5\x02\x00\xB1\xAD\x02\x00\xB2\xA5"
-     "\x02\x00"
-     "\xB3\xBD\x02\x00\xB4\xAD\x02\x00\xB5\x9D\x02\x00\xB6\x95\x02\x00\xB7\x7D"
-     "\x01\x00"
-     "\x81\xB9\x00\x00\x80\xB1\x00\x00\xBE\x84\x05\x00\x82\xB1\x00\x00\x86\x30"
-     "\x04\x00"
-     "\x87\x54\x05\x00\xB2\x4B\x00\x80\xEF\x88\x01\x00\x84\xA0\x05\x00\xE1\xC0"
-     "\x1E\x00"
-     "\xB6\x4B\x00\x80\xE3\x54\x01\x00\xBA\x4B\x00\x80\xBE\x4B\x00\x80\xC2\x4B"
-     "\x00\x80"
-     "\xC6\x4B\x00\x80\xB3\x45\x01\x00\xCA\x4B\x00\x80\xCE\x4B\x00\x80\xD2\x4B"
-     "\x00\x80"
-     "\xD6\x4B\x00\x80\xB6\x41\x01\x00\xB5\x55\x01\x00\xDA\x4B\x00\x80\xBB\x4D"
-     "\x01\x00"
-     "\xBA\x41\x01\x00\xDE\x4B\x00\x80\xE2\x4B\x00\x80\xBF\x4D\x01\x00\xBE\x4D"
-     "\x01\x00"
-     "\xBD\x4D\x01\x00\xBC\x4D\x01\x00\xA8\x51\x06\x00\xA9\x85\x06\x00\xAA\x8D"
-     "\x06\x00"
-     "\xAB\x85\x06\x00\xAC\x9D\x06\x00\xAD\x81\x06\x00\xAE\x81\x06\x00\xAF\x81"
-     "\x06\x00"
-     "\xAE\x4B\x00\x80\xE6\x4B\x00\x80\xEA\x4B\x00\x80\xEE\x4B\x00\x80\xF2\x4B"
-     "\x00\x80"
-     "\xF6\x4B\x00\x80\xFA\x4B\x00\x80\xFE\x4B\x00\x80\xB8\x8D\x06\x00\xB9\x95"
-     "\x06\x00"
-     "\xBA\x95\x06\x00\xBB\xA5\x06\x00\xBC\xBD\x06\x00\xBD\x71\x07\x00\xBE\x75"
-     "\x07\x00"
-     "\xBF\x6D\x07\x00\xB0\xC5\x06\x00\xB1\xC9\x06\x00\xB2\xD9\x06\x00\xB3\xD9"
-     "\x06\x00"
-     "\xB4\xC9\x06\x00\xB5\xC9\x06\x00\xB6\xBD\x06\x00\xB7\xB5\x06\x00\xA3\x05"
-     "\x06\x00"
-     "\x02\x4C\x00\x80\x06\x4C\x00\x80\x0A\x4C\x00\x80\x0E\x4C\x00\x80\xA6\x01"
-     "\x06\x00"
-     "\xA5\x15\x06\x00\x12\x4C\x00\x80\xAB\x0D\x06\x00\xAA\x01\x06\x00\x16\x4C"
-     "\x00\x80"
-     "\x1A\x4C\x00\x80\xAF\x0D\x06\x00\xAE\x0D\x06\x00\xAD\x0D\x06\x00\xAC\x0D"
-     "\x06\x00"
-     "\x80\x5D\x00\x00\x81\x6D\x00\x00\x82\x65\x00\x00\xB3\x7D\x07\x00\xBE\xDC"
-     "\x03\x00"
-     "\xB5\x6D\x07\x00\xB6\x65\x07\x00\x1E\x4C\x00\x80\x86\xE0\x00\x00\x87\x14"
-     "\x03\x00"
-     "\xBA\x25\x07\x00\xBB\x3D\x07\x00\xBC\x2D\x07\x00\xBD\x1D\x07\x00\xBE\x15"
-     "\x07\x00"
-     "\xBF\xD9\x00\x00\xA8\x99\x06\x00\xA9\xAD\x06\x00\xAA\xB9\x06\x00\xAB\xB9"
-     "\x06\x00"
-     "\xAC\xA9\x06\x00\xAD\xA9\x06\x00\xAE\xDD\x06\x00\xAF\xD5\x06\x00\x22\x4C"
-     "\x00\x80"
-     "\x26\x4C\x00\x80\x2A\x4C\x00\x80\x2E\x4C\x00\x80\x32\x4C\x00\x80\x36\x4C"
-     "\x00\x80"
-     "\x3A\x4C\x00\x80\x3E\x4C\x00\x80\xB8\x71\x01\x00\xB9\x71\x01\x00\xBA\x71"
-     "\x01\x00"
-     "\xBB\x71\x01\x00\xBC\xD5\x01\x00\xBD\xDD\x01\x00\xBE\xD5\x01\x00\xBF\xCD"
-     "\x01\x00"
-     "\xB0\xAD\x06\x00\xB1\xB5\x06\x00\xB2\xB5\x06\x00\xB3\x85\x06\x00\xB4\x9D"
-     "\x06\x00"
-     "\xB5\x51\x01\x00\xB6\x51\x01\x00\xB7\x51\x01\x00\xA3\x31\x06\x00\x42\x4C"
-     "\x00\x80"
-     "\x46\x4C\x00\x80\x4A\x4C\x00\x80\x4E\x4C\x00\x80\xA6\x29\x06\x00\xA5\x21"
-     "\x06\x00"
-     "\x52\x4C\x00\x80\xAB\x71\x06\x00\xAA\x69\x06\x00\x56\x4C\x00\x80\x3E\x4B"
-     "\x00\x80"
-     "\xAF\x95\x01\x00\xAE\x59\x06\x00\xAD\x51\x06\x00\xAC\x61\x06\x00\x5A\x4C"
-     "\x00\x80"
-     "\x5E\x4C\x00\x80\x62\x4C\x00\x80\x66\x4C\x00\x80\x6A\x4C\x00\x80\x6E\x4C"
-     "\x00\x80"
-     "\x72\x4C\x00\x80\x76\x4C\x00\x80\x7A\x4C\x00\x80\x7E\x4C\x00\x80\x82\x4C"
-     "\x00\x80"
-     "\x86\x4C\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x8A\x4C"
-     "\x00\x80"
-     "\x84\x80\x02\x00\xBE\x84\x03\x00\x87\x08\x03\x00\x86\xAC\x0C\x00\xEF\x60"
-     "\x07\x00"
-     "\x92\x4C\x00\x80\x96\x4C\x00\x80\x9A\x4C\x00\x80\xE3\x10\x07\x00\x9E\x4C"
-     "\x00\x80"
-     "\xE1\x50\x06\x00\xA2\x4C\x00\x80\xA6\x4C\x00\x80\xAA\x4C\x00\x80\xAE\x4C"
-     "\x00\x80"
-     "\xB2\x4C\x00\x80\xB6\x4C\x00\x80\xBA\x4C\x00\x80\xBE\x4C\x00\x80\xB3\x65"
-     "\x03\x00"
-     "\xC2\x4C\x00\x80\xB5\x65\x03\x00\xB6\x6D\x03\x00\xC6\x4C\x00\x80\xCA\x4C"
-     "\x00\x80"
-     "\xCE\x4C\x00\x80\xBA\xED\x03\x00\xBB\xE5\x03\x00\xBC\xFD\x03\x00\xBD\xE5"
-     "\x03\x00"
-     "\xBE\xE5\x03\x00\xBF\xD5\x03\x00\xD2\x4C\x00\x80\xD6\x4C\x00\x80\xDA\x4C"
-     "\x00\x80"
-     "\xDE\x4C\x00\x80\xE2\x4C\x00\x80\xE6\x4C\x00\x80\xEA\x4C\x00\x80\xEE\x4C"
-     "\x00\x80"
-     "\xA8\x81\x03\x00\xA9\x81\x03\x00\xAA\x81\x03\x00\xAB\x81\x03\x00\xAC\x81"
-     "\x03\x00"
-     "\xAD\x81\x03\x00\xAE\x81\x03\x00\xAF\x81\x03\x00\xB0\x45\x03\x00\xB1\x4D"
-     "\x03\x00"
-     "\xB2\x45\x03\x00\xB3\x5D\x03\x00\xB4\x4D\x03\x00\xB5\x7D\x03\x00\xB6\x19"
-     "\x03\x00"
-     "\xB7\x19\x03\x00\xB8\x29\x03\x00\xB9\x29\x03\x00\xBA\x39\x03\x00\xBB\x39"
-     "\x03\x00"
-     "\xBC\x29\x03\x00\xBD\x29\x03\x00\xBE\x19\x03\x00\xBF\x19\x03\x00\x80\x11"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x82\x29\x00\x00\xA3\x25\x02\x00\xF2\x4C\x00\x80\xA5\x25"
-     "\x02\x00"
-     "\xA6\x2D\x02\x00\xF6\x4C\x00\x80\xFA\x4C\x00\x80\x02\x4D\x00\x80\xAA\xAD"
-     "\x02\x00"
-     "\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xA5\x02\x00\xAE\xA5\x02\x00\xAF\x95"
-     "\x02\x00"
-     "\x84\x4C\x02\x00\xBE\x48\x0D\x00\x87\x64\x0D\x00\x86\xEC\x0C\x00\xE3\x40"
-     "\x03\x00"
-     "\x06\x4D\x00\x80\xE1\xA0\x01\x00\x0A\x4D\x00\x80\xEF\x9C\x03\x00\x0E\x4D"
-     "\x00\x80"
-     "\x12\x4D\x00\x80\x16\x4D\x00\x80\x1A\x4D\x00\x80\x1E\x4D\x00\x80\x22\x4D"
-     "\x00\x80"
-     "\x26\x4D\x00\x80\x2A\x4D\x00\x80\x2E\x4D\x00\x80\x32\x4D\x00\x80\x36\x4D"
-     "\x00\x80"
-     "\x3A\x4D\x00\x80\x3E\x4D\x00\x80\x42\x4D\x00\x80\xEF\xB4\x01\x00\x84\xA8"
-     "\x0D\x00"
-     "\xE1\x2C\x06\x00\x46\x4D\x00\x80\xE3\x1C\x06\x00\x4A\x4D\x00\x80\x4E\x4D"
-     "\x00\x80"
-     "\x52\x4D\x00\x80\x56\x4D\x00\x80\xB3\xDD\x01\x00\x5A\x4D\x00\x80\x5E\x4D"
-     "\x00\x80"
-     "\x62\x4D\x00\x80\x66\x4D\x00\x80\xB6\xE5\x01\x00\xB5\xF5\x01\x00\x6A\x4D"
-     "\x00\x80"
-     "\xBB\xA1\x01\x00\xBA\xD9\x01\x00\xBE\xEC\x0C\x00\x72\x4D\x00\x80\xBF\xAD"
-     "\x01\x00"
-     "\xBE\xA5\x01\x00\xBD\xB1\x01\x00\xBC\xB9\x01\x00\xA8\x15\x0E\x00\xA9\x1D"
-     "\x0E\x00"
-     "\xAA\x15\x0E\x00\xAB\x29\x0E\x00\xAC\x3D\x0E\x00\xAD\x25\x0E\x00\xAE\x2D"
-     "\x0E\x00"
-     "\xAF\x25\x0E\x00\xFE\x4C\x00\x80\x82\xF9\x0F\x00\x81\xF9\x0F\x00\x80\xE1"
-     "\x0F\x00"
-     "\x6E\x4D\x00\x80\x76\x4D\x00\x80\x86\x1C\x00\x00\x87\xB8\x03\x00\xB8\xC1"
-     "\x0E\x00"
-     "\xB9\xC1\x0E\x00\xBA\xC1\x0E\x00\xBB\xC1\x0E\x00\xBC\xC1\x0E\x00\xBD\xC1"
-     "\x0E\x00"
-     "\xBE\xC1\x0E\x00\xBF\xC1\x0E\x00\xB0\x5D\x0E\x00\xB1\x25\x0E\x00\xB2\x2D"
-     "\x0E\x00"
-     "\xB3\x25\x0E\x00\xB4\x21\x0E\x00\xB5\x29\x0E\x00\xB6\x19\x0E\x00\xB7\x19"
-     "\x0E\x00"
-     "\xA3\x9D\x0E\x00\x7A\x4D\x00\x80\x7E\x4D\x00\x80\x82\x4D\x00\x80\x86\x4D"
-     "\x00\x80"
-     "\xA6\xA5\x0E\x00\xA5\xB5\x0E\x00\x8A\x4D\x00\x80\xAB\xE1\x0E\x00\xAA\x99"
-     "\x0E\x00"
-     "\x8E\x4D\x00\x80\x92\x4D\x00\x80\xAF\xED\x0E\x00\xAE\xE5\x0E\x00\xAD\xF1"
-     "\x0E\x00"
-     "\xAC\xF9\x0E\x00\x96\x4D\x00\x80\xB3\x99\x0F\x00\x9A\x4D\x00\x80\x9E\x4D"
-     "\x00\x80"
-     "\xB6\xBD\x0F\x00\xA2\x4D\x00\x80\xA6\x4D\x00\x80\xB5\xB5\x0F\x00\xBA\x99"
-     "\x0F\x00"
-     "\xBB\x99\x0F\x00\xAA\x4D\x00\x80\xAE\x4D\x00\x80\xBE\x79\x0F\x00\xBF\x79"
-     "\x0F\x00"
-     "\xBC\x79\x0F\x00\xBD\x79\x0F\x00\xA8\x25\x0E\x00\xA9\x51\x0E\x00\xAA\x55"
-     "\x0E\x00"
-     "\xAB\x69\x0E\x00\xAC\x79\x0E\x00\xAD\x79\x0E\x00\xAE\x69\x0E\x00\xAF\x69"
-     "\x0E\x00"
-     "\xB2\x4D\x00\x80\xB6\x4D\x00\x80\xBA\x4D\x00\x80\xBE\x4D\x00\x80\xC2\x4D"
-     "\x00\x80"
-     "\xC6\x4D\x00\x80\xCA\x4D\x00\x80\xCE\x4D\x00\x80\xB8\xD9\x0E\x00\xB9\xED"
-     "\x0E\x00"
-     "\xBA\xF9\x0E\x00\xBB\xF9\x0E\x00\xBC\xE9\x0E\x00\xBD\xE9\x0E\x00\xBE\x9D"
-     "\x0E\x00"
-     "\xBF\x95\x0E\x00\xB0\x19\x0E\x00\xB1\x19\x0E\x00\xB2\xE9\x0E\x00\xB3\xE9"
-     "\x0E\x00"
-     "\xB4\xF9\x0E\x00\xB5\xF9\x0E\x00\xB6\xE9\x0E\x00\xB7\xE9\x0E\x00\x81\x19"
-     "\x00\x00"
-     "\x80\x11\x00\x00\xA3\xD5\x0E\x00\x82\x29\x00\x00\xA5\xF9\x0E\x00\xD2\x4D"
-     "\x00\x80"
-     "\xD6\x4D\x00\x80\xA6\xF1\x0E\x00\x8E\x4C\x00\x80\xDA\x4D\x00\x80\xAB\xD5"
-     "\x0E\x00"
-     "\xAA\xD5\x0E\x00\xAD\x35\x0E\x00\xAC\x35\x0E\x00\xAF\x35\x0E\x00\xAE\x35"
-     "\x0E\x00"
-     "\x84\xC8\x02\x00\xB3\x41\x01\x00\xBE\x0C\x01\x00\xE2\x4D\x00\x80\xB6\x41"
-     "\x01\x00"
-     "\xE6\x4D\x00\x80\xEA\x4D\x00\x80\xB5\x51\x01\x00\xBA\x4D\x01\x00\xBB\x25"
-     "\x01\x00"
-     "\x86\x40\x00\x00\x87\x50\x01\x00\xBE\x25\x01\x00\xBF\x2D\x01\x00\xBC\x31"
-     "\x01\x00"
-     "\xBD\x31\x01\x00\xEE\x4D\x00\x80\xF2\x4D\x00\x80\x84\x04\x03\x00\xBE\x80"
-     "\x04\x00"
-     "\xF6\x4D\x00\x80\xFA\x4D\x00\x80\xFE\x4D\x00\x80\xEF\xA0\x0F\x00\xBF\xA8"
-     "\xCE\x00"
-     "\xE1\x30\x0E\x00\x02\x4E\x00\x80\xE3\x7C\x0F\x00\x06\x4E\x00\x80\x0A\x4E"
-     "\x00\x80"
-     "\x0E\x4E\x00\x80\x12\x4E\x00\x80\x16\x4E\x00\x80\x1A\x4E\x00\x80\xA3\xCD"
-     "\x02\x00"
-     "\x1E\x4E\x00\x80\xA5\xDD\x02\x00\x22\x4E\x00\x80\x26\x4E\x00\x80\xA6\xCD"
-     "\x02\x00"
-     "\x2A\x4E\x00\x80\x2E\x4E\x00\x80\xAB\xA9\x02\x00\xAA\xC1\x02\x00\xAD\xBD"
-     "\x02\x00"
-     "\xAC\xBD\x02\x00\xAF\xA1\x02\x00\xAE\xA9\x02\x00\xA9\xB5\x02\x00\xA8\x69"
-     "\x0D\x00"
-     "\xAB\x01\x02\x00\xAA\x09\x02\x00\xAD\x01\x02\x00\xAC\x19\x02\x00\xAF\x31"
-     "\x02\x00"
-     "\xAE\x01\x02\x00\x84\x6C\x05\x00\x32\x4E\x00\x80\x36\x4E\x00\x80\x3A\x4E"
-     "\x00\x80"
-     "\x3E\x4E\x00\x80\x82\x9D\x01\x00\x81\x9D\x01\x00\x80\x9D\x01\x00\xB9\xE5"
-     "\x03\x00"
-     "\xB8\xE5\x03\x00\xBB\xE5\x03\x00\xBA\xE5\x03\x00\xBD\xE5\x03\x00\xBC\xE5"
-     "\x03\x00"
-     "\xBF\xE5\x03\x00\xBE\xE5\x03\x00\xB1\x21\x02\x00\xB0\x49\x02\x00\xB3\x25"
-     "\x02\x00"
-     "\xB2\x25\x02\x00\xB5\x29\x02\x00\xB4\x21\x02\x00\xB7\x15\x02\x00\xB6\x15"
-     "\x02\x00"
-     "\xAA\xD9\x02\x00\xAB\x05\x01\x00\xA8\xC5\x02\x00\xA9\xD9\x02\x00\xAE\x05"
-     "\x01\x00"
-     "\xAF\x0D\x01\x00\xAC\x1D\x01\x00\xAD\x0D\x01\x00\x42\x4E\x00\x80\x46\x4E"
-     "\x00\x80"
-     "\x4A\x4E\x00\x80\x52\x4E\x00\x80\xBE\xF8\x04\x00\x56\x4E\x00\x80\x5A\x4E"
-     "\x00\x80"
-     "\x5E\x4E\x00\x80\xBA\x21\x01\x00\xBB\x21\x01\x00\xB8\x1D\x01\x00\xB9\x15"
-     "\x01\x00"
-     "\xBE\xFD\x01\x00\xBF\xE5\x01\x00\xBC\xFD\x01\x00\xBD\xF5\x01\x00\xB2\x55"
-     "\x01\x00"
-     "\xB3\x5D\x01\x00\xB0\x41\x01\x00\xB1\x5D\x01\x00\xB6\x35\x01\x00\xB7\x39"
-     "\x01\x00"
-     "\xB4\x45\x01\x00\xB5\x3D\x01\x00\x87\xCC\x05\x00\x86\xA4\x05\x00\x62\x4E"
-     "\x00\x80"
-     "\xEF\xBC\x00\x00\x66\x4E\x00\x80\x6A\x4E\x00\x80\x6E\x4E\x00\x80\xEF\x8C"
-     "\x0E\x00"
-     "\x84\x6C\x07\x00\xE1\xF0\x0E\x00\x72\x4E\x00\x80\xE3\x1C\x0E\x00\x76\x4E"
-     "\x00\x80"
-     "\xE1\x94\x01\x00\x7A\x4E\x00\x80\xE3\x6C\x00\x00\xB3\x19\x02\x00\x7E\x4E"
-     "\x00\x80"
-     "\x82\x4E\x00\x80\x86\x4E\x00\x80\x84\x00\x38\x00\xB6\xC5\x01\x00\xB5\xDD"
-     "\x01\x00"
-     "\x8A\x4E\x00\x80\xBB\xD1\x01\x00\xBA\xCD\x01\x00\x8E\x4E\x00\x80\x92\x4E"
-     "\x00\x80"
-     "\xBF\xB1\x01\x00\xBE\xC9\x01\x00\xBD\xC1\x01\x00\xBC\xC9\x01\x00\xA3\xD9"
-     "\x05\x00"
-     "\x4E\x4E\x00\x80\x96\x4E\x00\x80\x9A\x4E\x00\x80\x9E\x4E\x00\x80\xA6\x05"
-     "\x06\x00"
-     "\xA5\x1D\x06\x00\xA2\x4E\x00\x80\xAB\x11\x06\x00\xAA\x0D\x06\x00\xA6\x4E"
-     "\x00\x80"
-     "\xAA\x4E\x00\x80\xAF\x71\x06\x00\xAE\x09\x06\x00\xAD\x01\x06\x00\xAC\x09"
-     "\x06\x00"
-     "\xAE\x4E\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\xB2\x4E"
-     "\x00\x80"
-     "\xB6\x4E\x00\x80\xBA\x4E\x00\x80\xBE\x4E\x00\x80\x86\x20\x03\x00\x87\xA4"
-     "\x03\x00"
-     "\xC2\x4E\x00\x80\xC6\x4E\x00\x80\xCA\x4E\x00\x80\xCE\x4E\x00\x80\xD2\x4E"
-     "\x00\x80"
-     "\xD6\x4E\x00\x80\xA8\xAD\x07\x00\xA9\xB1\x07\x00\xAA\xB1\x07\x00\xAB\xB1"
-     "\x07\x00"
-     "\xAC\x19\x07\x00\xAD\x05\x07\x00\xAE\x0D\x07\x00\xAF\x05\x07\x00\xDA\x4E"
-     "\x00\x80"
-     "\xDE\x4E\x00\x80\xE2\x4E\x00\x80\xE6\x4E\x00\x80\xEA\x4E\x00\x80\xEE\x4E"
-     "\x00\x80"
-     "\xF2\x4E\x00\x80\xF6\x4E\x00\x80\xB8\x2D\x07\x00\xB9\xCD\x00\x00\xBA\xC5"
-     "\x00\x00"
-     "\xBB\xDD\x00\x00\xBC\xCD\x00\x00\xBD\xFD\x00\x00\xBE\xF5\x00\x00\xBF\x9D"
-     "\x00\x00"
-     "\xB0\x49\x07\x00\xB1\x55\x07\x00\xB2\x51\x07\x00\xB3\x29\x07\x00\xB4\x39"
-     "\x07\x00"
-     "\xB5\x39\x07\x00\xB6\x25\x07\x00\xB7\x15\x07\x00\xB3\x39\x06\x00\xFA\x4E"
-     "\x00\x80"
-     "\xFE\x4E\x00\x80\x02\x4F\x00\x80\x06\x4F\x00\x80\xB6\x85\x06\x00\xB5\x91"
-     "\x06\x00"
-     "\x0A\x4F\x00\x80\xBB\x91\x06\x00\xBA\x8D\x06\x00\x0E\x4F\x00\x80\x12\x4F"
-     "\x00\x80"
-     "\xBF\xFD\x06\x00\xBE\xFD\x06\x00\xBD\xFD\x06\x00\xBC\x85\x06\x00\x16\x4F"
-     "\x00\x80"
-     "\xA3\x7D\x06\x00\x1A\x4F\x00\x80\x1E\x4F\x00\x80\xA6\xC1\x06\x00\x22\x4F"
-     "\x00\x80"
-     "\x26\x4F\x00\x80\xA5\xD5\x06\x00\xAA\xC9\x06\x00\xAB\xD5\x06\x00\x2A\x4F"
-     "\x00\x80"
-     "\xBE\x6C\x01\x00\xAE\xB9\x06\x00\xAF\xB9\x06\x00\xAC\xC1\x06\x00\xAD\xB9"
-     "\x06\x00"
-     "\xB0\x29\x01\x00\xB1\x29\x01\x00\xB2\x39\x01\x00\xB3\x39\x01\x00\xB4\x29"
-     "\x01\x00"
-     "\xB5\x29\x01\x00\xB6\x9D\x00\x00\xB7\x95\x00\x00\xB8\xAD\x00\x00\xB9\xB5"
-     "\x00\x00"
-     "\xBA\xB5\x00\x00\xBB\x85\x00\x00\xBC\x9D\x00\x00\xBD\xF5\x00\x00\xBE\xF1"
-     "\x00\x00"
-     "\xBF\xF1\x00\x00\x80\x95\x00\x00\x81\x9D\x00\x00\x82\xA1\x00\x00\x2E\x4F"
-     "\x00\x80"
-     "\x32\x4F\x00\x80\xDE\x4D\x00\x80\x87\x08\x01\x00\x86\x7C\x00\x00\xA8\x89"
-     "\x01\x00"
-     "\xA9\x89\x01\x00\xAA\x99\x01\x00\xAB\x99\x01\x00\xAC\x89\x01\x00\xAD\x89"
-     "\x01\x00"
-     "\xAE\x59\x01\x00\xAF\x59\x01\x00\xB3\x71\x02\x00\x36\x4F\x00\x80\x3A\x4F"
-     "\x00\x80"
-     "\x3E\x4F\x00\x80\x42\x4F\x00\x80\xB6\x69\x02\x00\xB5\x61\x02\x00\x46\x4F"
-     "\x00\x80"
-     "\xBB\x8D\x02\x00\xBA\x8D\x02\x00\x4A\x4F\x00\x80\x4E\x4F\x00\x80\xBF\x8D"
-     "\x02\x00"
-     "\xBE\x8D\x02\x00\xBD\x8D\x02\x00\xBC\x8D\x02\x00\xE3\x80\x07\x00\x52\x4F"
-     "\x00\x80"
-     "\xE1\x8C\x01\x00\x96\x00\x00\x00\xE1\xA0\x05\x00\x56\x4F\x00\x80\xE3\x40"
-     "\x06\x00"
-     "\x5A\x4F\x00\x80\x5E\x4F\x00\x80\x62\x4F\x00\x80\xEF\x9C\x06\x00\x66\x4F"
-     "\x00\x80"
-     "\xEF\xF0\x02\x00\x6A\x4F\x00\x80\xBE\xF4\x3C\x00\x6E\x4F\x00\x80\x72\x4F"
-     "\x00\x80"
-     "\x76\x4F\x00\x80\x7A\x4F\x00\x80\xA3\xFD\x02\x00\x7E\x4F\x00\x80\xA5\xED"
-     "\x02\x00"
-     "\xA6\xE5\x02\x00\x82\x4F\x00\x80\x86\x4F\x00\x80\x8A\x4F\x00\x80\xAA\x01"
-     "\x02\x00"
-     "\xAB\x01\x02\x00\xAC\x01\x02\x00\xAD\x01\x02\x00\xAE\x01\x02\x00\xAF\x01"
-     "\x02\x00"
-     "\x8E\x4F\x00\x80\x92\x4F\x00\x80\x96\x4F\x00\x80\x9A\x4F\x00\x80\x80\x15"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x82\x15\x00\x00\xEF\x08\x0F\x00\x85\xA4\x3C\x00\xE1\x10"
-     "\x02\x00"
-     "\xA2\x4F\x00\x80\xE3\x48\x3F\x00\x86\x50\x3E\x00\x87\x34\x03\x00\xA6\x4F"
-     "\x00\x80"
-     "\xAA\x4F\x00\x80\xA8\xD1\x02\x00\xA9\x01\x01\x00\xAA\x01\x01\x00\xAB\x01"
-     "\x01\x00"
-     "\xAC\x01\x01\x00\xAD\x09\x01\x00\xAE\x31\x01\x00\xAF\x31\x01\x00\xAE\x4F"
-     "\x00\x80"
-     "\xB2\x4F\x00\x80\xB6\x4F\x00\x80\xBA\x4F\x00\x80\xBE\x4F\x00\x80\xC2\x4F"
-     "\x00\x80"
-     "\xC6\x4F\x00\x80\xCA\x4F\x00\x80\xB8\xC1\x01\x00\xB9\xC1\x01\x00\xBA\xC1"
-     "\x01\x00"
-     "\xBB\xDD\x01\x00\xBC\xCD\x01\x00\xBD\xF5\x01\x00\xBE\xFD\x01\x00\xBF\x9D"
-     "\x01\x00"
-     "\xB0\x25\x01\x00\xB1\x31\x01\x00\xB2\x35\x01\x00\xB3\x0D\x01\x00\xB4\x15"
-     "\x01\x00"
-     "\xB5\x19\x01\x00\xB6\x0D\x01\x00\xB7\x01\x01\x00\xE1\x1C\x06\x00\xCE\x4F"
-     "\x00\x80"
-     "\xE3\x38\x07\x00\xD2\x4F\x00\x80\xD6\x4F\x00\x80\x84\x18\x3C\x00\xDA\x4F"
-     "\x00\x80"
-     "\xDE\x4F\x00\x80\xE2\x4F\x00\x80\xE6\x4F\x00\x80\xEA\x4F\x00\x80\xEE\x4F"
-     "\x00\x80"
-     "\xF2\x4F\x00\x80\xF6\x4F\x00\x80\xEF\x70\x06\x00\xFA\x4F\x00\x80\x80\x1D"
-     "\x00\x00"
-     "\x81\x21\x00\x00\x82\x21\x00\x00\xBE\xC0\x3F\x00\xE3\x60\x06\x00\xFE\x4F"
-     "\x00\x80"
-     "\xE1\xA0\x01\x00\x02\x50\x00\x80\xEF\x74\x00\x00\x06\x50\x00\x80\x0A\x50"
-     "\x00\x80"
-     "\x12\x50\x00\x80\x86\x50\x3E\x00\x87\x00\x3C\x00\x16\x50\x00\x80\x1A\x50"
-     "\x00\x80"
-     "\xA3\x09\x3E\x00\x8F\x25\x24\x00\x9E\x4F\x00\x80\x84\x08\x00\x00\x0E\x50"
-     "\x00\x80"
-     "\xA6\x09\x3E\x00\xA5\x09\x3E\x00\x1E\x50\x00\x80\xAB\x75\x3E\x00\xAA\x61"
-     "\x3E\x00"
-     "\x22\x50\x00\x80\x26\x50\x00\x80\xAF\x5D\x3E\x00\xAE\x51\x3E\x00\xAD\x59"
-     "\x3E\x00"
-     "\xAC\x61\x3E\x00\x9E\x61\x38\x00\x9F\x75\x38\x00\x9C\x95\x34\x00\x9D\xED"
-     "\x39\x00"
-     "\x9A\xFD\x35\x00\x9B\x75\x34\x00\x98\x79\x30\x00\x99\x71\x31\x00\x96\xB1"
-     "\x2C\x00"
-     "\x97\xDD\x31\x00\x94\xE9\x2C\x00\x95\xE1\x2C\x00\x92\xE1\x28\x00\x93\x1D"
-     "\x29\x00"
-     "\x90\x01\x26\x00\x91\x7D\x28\x00\xB1\x0D\x14\x00\xB0\x05\x14\x00\xB3\x01"
-     "\x18\x00"
-     "\xB2\x55\x14\x00\xB5\x79\x18\x00\xB4\xB5\x18\x00\x2A\x50\x00\x80\x2E\x50"
-     "\x00\x80"
-     "\xBE\xE0\x01\x00\x32\x50\x00\x80\x36\x50\x00\x80\x3A\x50\x00\x80\xA2\xED"
-     "\x3C\x00"
-     "\xA3\x79\x3D\x00\xA0\x6D\x3D\x00\xA1\xB9\x3C\x00\xA1\xC9\x00\x00\x3E\x50"
-     "\x00\x80"
-     "\xA3\x01\x04\x00\xA2\xDD\x00\x00\xA5\x7D\x04\x00\xA4\xF5\x04\x00\xA7\xF9"
-     "\x08\x00"
-     "\xA6\x01\x08\x00\xA9\x01\x0C\x00\xA8\x6D\x08\x00\xAB\xCD\x0C\x00\xAA\xCD"
-     "\x0C\x00"
-     "\xAD\xD1\x10\x00\xAC\x01\x10\x00\xAF\xD5\x11\x00\xAE\xD9\x10\x00\x42\x50"
-     "\x00\x80"
-     "\xB5\x15\x3F\x00\xB6\x15\x3F\x00\x46\x50\x00\x80\x4A\x50\x00\x80\x4E\x50"
-     "\x00\x80"
-     "\x52\x50\x00\x80\xB3\x15\x3F\x00\xBC\x7D\x3F\x00\xBD\x45\x3F\x00\xBE\x4D"
-     "\x3F\x00"
-     "\xBF\x41\x3F\x00\x56\x50\x00\x80\x5A\x50\x00\x80\xBA\x7D\x3F\x00\xBB\x69"
-     "\x3F\x00"
-     "\xB3\xBD\x3F\x00\x5E\x50\x00\x80\x62\x50\x00\x80\x66\x50\x00\x80\x6A\x50"
-     "\x00\x80"
-     "\xB6\x7D\x3E\x00\xB5\x7D\x3E\x00\x6E\x50\x00\x80\xBB\x59\x3E\x00\xBA\x51"
-     "\x3E\x00"
-     "\x72\x50\x00\x80\x76\x50\x00\x80\xBF\xE1\x3E\x00\xBE\xF9\x3E\x00\xBD\xFD"
-     "\x3E\x00"
-     "\xBC\x41\x3E\x00\x82\xA1\x25\x00\x83\xFD\x27\x00\x7A\x50\x00\x80\x82\x50"
-     "\x00\x80"
-     "\x86\xF9\x28\x00\x87\x15\x2D\x00\x84\x99\x29\x00\x85\x0D\x29\x00\x8A\x2D"
-     "\x2D\x00"
-     "\x8B\x75\x2D\x00\x86\x08\x07\x00\x87\x8C\x00\x00\x8E\x9D\x30\x00\x8F\x19"
-     "\x34\x00"
-     "\x8C\x01\x30\x00\x8D\x9D\x31\x00\x92\xE1\x35\x00\x93\x19\x35\x00\x86\x50"
-     "\x00\x80"
-     "\x8A\x50\x00\x80\x96\xE1\x39\x00\x97\x01\x3C\x00\x94\x79\x38\x00\x95\xFD"
-     "\x39\x00"
-     "\x9A\xE5\x3D\x00\x9B\xB9\x3D\x00\x8E\x50\x00\x80\x92\x50\x00\x80\x81\x9D"
-     "\x00\x00"
-     "\x80\x9D\x00\x00\x9C\x35\x3C\x00\x82\x75\x00\x00\xA8\xE9\x3C\x00\xA9\xFD"
-     "\x3C\x00"
-     "\xAA\xF5\x3C\x00\xAB\x2D\x3D\x00\xAC\x35\x3D\x00\xAD\x3D\x3D\x00\xAE\x35"
-     "\x3D\x00"
-     "\xAF\x2D\x3D\x00\x96\x50\x00\x80\x9A\x50\x00\x80\x9E\x50\x00\x80\xA2\x50"
-     "\x00\x80"
-     "\xA6\x50\x00\x80\xAA\x50\x00\x80\xBE\x70\x04\x00\xAE\x50\x00\x80\xB8\xE1"
-     "\x3D\x00"
-     "\xB9\xE1\x3D\x00\xBA\xE1\x3D\x00\xBB\xE1\x3D\x00\xBC\xE1\x3D\x00\xBD\xE1"
-     "\x3D\x00"
-     "\xBE\xE5\x3D\x00\xBF\x89\x3D\x00\xB0\x55\x3D\x00\xB1\x5D\x3D\x00\xB2\x55"
-     "\x3D\x00"
-     "\xB3\xE9\x3D\x00\xB4\xFD\x3D\x00\xB5\xE5\x3D\x00\xB6\xE1\x3D\x00\xB7\xE1"
-     "\x3D\x00"
-     "\xA3\x7D\x3C\x00\xB2\x50\x00\x80\xB6\x50\x00\x80\xBE\xC4\x05\x00\xBA\x50"
-     "\x00\x80"
-     "\xA6\xBD\x3D\x00\xA5\xBD\x3D\x00\xBE\x50\x00\x80\xAB\x99\x3D\x00\xAA\x91"
-     "\x3D\x00"
-     "\xC2\x50\x00\x80\xC6\x50\x00\x80\xAF\x21\x3D\x00\xAE\x39\x3D\x00\xAD\x3D"
-     "\x3D\x00"
-     "\xAC\x81\x3D\x00\xCA\x50\x00\x80\xCE\x50\x00\x80\xB3\x35\x3D\x00\xD2\x50"
-     "\x00\x80"
-     "\xB5\xF9\x02\x00\xD6\x50\x00\x80\xDA\x50\x00\x80\xB6\xF1\x02\x00\xDE\x50"
-     "\x00\x80"
-     "\xE2\x50\x00\x80\xBB\x41\x02\x00\xBA\x49\x02\x00\xBD\x35\x02\x00\xBC\x59"
-     "\x02\x00"
-     "\xBF\xF5\x03\x00\xBE\xFD\x03\x00\xA8\xED\x3C\x00\xA9\xFD\x3C\x00\xAA\xF5"
-     "\x3C\x00"
-     "\xAB\x11\x03\x00\xAC\x39\x03\x00\xAD\x39\x03\x00\xAE\x81\x03\x00\xAF\x81"
-     "\x03\x00"
-     "\x80\xC5\x01\x00\x81\x09\x00\x00\x82\x19\x00\x00\xE6\x50\x00\x80\xEA\x50"
-     "\x00\x80"
-     "\x84\x78\x05\x00\xF2\x50\x00\x80\xF6\x50\x00\x80\xB8\x59\x02\x00\xB9\x75"
-     "\x02\x00"
-     "\xBA\x7D\x02\x00\xBB\x09\x02\x00\xBC\x19\x02\x00\xBD\x19\x02\x00\xBE\x09"
-     "\x02\x00"
-     "\xBF\x01\x02\x00\xB0\xC5\x03\x00\xB1\xCD\x03\x00\xB2\xC5\x03\x00\xB3\xA9"
-     "\x03\x00"
-     "\xB4\x71\x02\x00\xB5\x79\x02\x00\xB6\x69\x02\x00\xB7\x69\x02\x00\xFA\x50"
-     "\x00\x80"
-     "\xE3\x00\x3E\x00\x84\xC4\x04\x00\xE1\x64\x3F\x00\xFE\x50\x00\x80\xE1\x7C"
-     "\x3E\x00"
-     "\x02\x51\x00\x80\xE3\x80\x01\x00\x86\x00\x04\x00\x87\x84\x04\x00\x06\x51"
-     "\x00\x80"
-     "\xEF\xF4\x01\x00\x85\x14\x1A\x00\xEF\x80\x3E\x00\x0A\x51\x00\x80\x0E\x51"
-     "\x00\x80"
-     "\x12\x51\x00\x80\x16\x51\x00\x80\x1A\x51\x00\x80\xA3\x31\x3C\x00\x84\x90"
-     "\x07\x00"
-     "\xA5\xFD\x03\x00\xA6\xF5\x03\x00\x1E\x51\x00\x80\x22\x51\x00\x80\x26\x51"
-     "\x00\x80"
-     "\xAA\x4D\x03\x00\xAB\x45\x03\x00\xAC\x5D\x03\x00\xAD\x31\x03\x00\xAE\xF9"
-     "\x02\x00"
-     "\xAF\xF1\x02\x00\xA8\x8D\x05\x00\xA9\xCD\x05\x00\xAA\xC5\x05\x00\xAB\xE5"
-     "\x05\x00"
-     "\xAC\xFD\x05\x00\xAD\x35\x06\x00\xAE\x3D\x06\x00\xAF\x35\x06\x00\xEE\x50"
-     "\x00\x80"
-     "\x2A\x51\x00\x80\x2E\x51\x00\x80\x32\x51\x00\x80\x36\x51\x00\x80\x3A\x51"
-     "\x00\x80"
-     "\x3E\x51\x00\x80\x42\x51\x00\x80\xB8\xE9\x06\x00\xB9\xE9\x06\x00\xBA\x89"
-     "\x06\x00"
-     "\xBB\x89\x06\x00\xBC\x9D\x06\x00\xBD\x81\x06\x00\xBE\x81\x06\x00\xBF\xBD"
-     "\x06\x00"
-     "\xB0\x4D\x06\x00\xB1\x55\x06\x00\xB2\x5D\x06\x00\xB3\x51\x06\x00\xB4\xF5"
-     "\x06\x00"
-     "\xB5\xFD\x06\x00\xB6\xE9\x06\x00\xB7\xE1\x06\x00\xB3\x4D\x05\x00\x46\x51"
-     "\x00\x80"
-     "\x4A\x51\x00\x80\x4E\x51\x00\x80\x52\x51\x00\x80\xB6\xB5\x06\x00\xB5\x75"
-     "\x05\x00"
-     "\x56\x51\x00\x80\xBB\x91\x06\x00\xBA\x89\x06\x00\x5A\x51\x00\x80\x5E\x51"
-     "\x00\x80"
-     "\xBF\x39\x06\x00\xBE\x31\x06\x00\xBD\x85\x06\x00\xBC\x89\x06\x00\x62\x51"
-     "\x00\x80"
-     "\xA3\x09\x05\x00\x7E\x50\x00\x80\x66\x51\x00\x80\xA6\xF1\x06\x00\x6A\x51"
-     "\x00\x80"
-     "\x6E\x51\x00\x80\xA5\x31\x05\x00\xAA\xCD\x06\x00\xAB\xD5\x06\x00\x86\x48"
-     "\x00\x00"
-     "\x87\x6C\x00\x00\xAE\x75\x06\x00\xAF\x7D\x06\x00\xAC\xCD\x06\x00\xAD\xC1"
-     "\x06\x00"
-     "\xA8\x39\x06\x00\xA9\x39\x06\x00\xAA\x89\x06\x00\xAB\x9D\x06\x00\xAC\x89"
-     "\x06\x00"
-     "\xAD\xBD\x06\x00\xAE\xA1\x06\x00\xAF\xA1\x06\x00\x80\xB1\x01\x00\x81\xB9"
-     "\x01\x00"
-     "\x82\x81\x01\x00\x72\x51\x00\x80\x76\x51\x00\x80\x7A\x51\x00\x80\x7E\x51"
-     "\x00\x80"
-     "\x82\x51\x00\x80\xB8\x6D\x01\x00\xB9\x75\x01\x00\xBA\x7D\x01\x00\xBB\x75"
-     "\x01\x00"
-     "\xBC\x6D\x01\x00\xBD\xD1\x01\x00\xBE\xD1\x01\x00\xBF\xD1\x01\x00\xB0\xA1"
-     "\x06\x00"
-     "\xB1\xA1\x06\x00\xB2\xA1\x06\x00\xB3\xB1\x06\x00\xB4\x91\x06\x00\xB5\x91"
-     "\x06\x00"
-     "\xB6\x5D\x01\x00\xB7\x55\x01\x00\xB3\x01\x06\x00\x86\x51\x00\x80\x8A\x51"
-     "\x00\x80"
-     "\x8E\x51\x00\x80\x92\x51\x00\x80\xB6\x3D\x06\x00\xB5\x01\x06\x00\x96\x51"
-     "\x00\x80"
-     "\xBB\x65\x06\x00\xBA\x65\x06\x00\x9A\x51\x00\x80\x9E\x51\x00\x80\xBF\x55"
-     "\x06\x00"
-     "\xBE\x55\x06\x00\xBD\x65\x06\x00\xBC\x75\x06\x00\xA2\x51\x00\x80\xA3\x45"
-     "\x06\x00"
-     "\xA6\x51\x00\x80\xAA\x51\x00\x80\xA6\x79\x06\x00\xAE\x51\x00\x80\xB2\x51"
-     "\x00\x80"
-     "\xA5\x45\x06\x00\xAA\x21\x06\x00\xAB\x21\x06\x00\xB6\x51\x00\x80\xBA\x51"
-     "\x00\x80"
-     "\xAE\x11\x06\x00\xAF\x11\x06\x00\xAC\x31\x06\x00\xAD\x21\x06\x00\xA8\xC1"
-     "\x02\x00"
-     "\xA9\xC9\x02\x00\xAA\xDD\x02\x00\xAB\xD5\x02\x00\xAC\x3D\x03\x00\xAD\x25"
-     "\x03\x00"
-     "\xAE\x2D\x03\x00\xAF\x25\x03\x00\xBE\x51\x00\x80\xC2\x51\x00\x80\xC6\x51"
-     "\x00\x80"
-     "\xCA\x51\x00\x80\xD2\x51\x00\x80\xD6\x51\x00\x80\xDA\x51\x00\x80\xBE\x70"
-     "\x02\x00"
-     "\xB8\x81\x03\x00\xB9\x8D\x03\x00\xBA\x85\x03\x00\xBB\x99\x03\x00\xBC\x89"
-     "\x03\x00"
-     "\xBD\xBD\x03\x00\xBE\xB5\x03\x00\xBF\x7D\x00\x00\xB0\x5D\x03\x00\xB1\xE1"
-     "\x03\x00"
-     "\xB2\xE1\x03\x00\xB3\xE1\x03\x00\xB4\xE1\x03\x00\xB5\xE9\x03\x00\xB6\xD1"
-     "\x03\x00"
-     "\xB7\xD1\x03\x00\x80\xA5\x00\x00\x81\xAD\x00\x00\x82\xA5\x00\x00\xB3\x59"
-     "\x02\x00"
-     "\xDE\x51\x00\x80\xB5\xB9\x03\x00\xB6\xA9\x03\x00\xE2\x51\x00\x80\x86\x00"
-     "\x0C\x00"
-     "\x87\xF4\x03\x00\xBA\x95\x03\x00\xBB\x29\x03\x00\xBC\x39\x03\x00\xBD\x39"
-     "\x03\x00"
-     "\xBE\x29\x03\x00\xBF\x21\x03\x00\xA3\x15\x02\x00\xE6\x51\x00\x80\xEA\x51"
-     "\x00\x80"
-     "\xEE\x51\x00\x80\xF2\x51\x00\x80\xA6\xE5\x03\x00\xA5\xF5\x03\x00\xF6\x51"
-     "\x00\x80"
-     "\xAB\x65\x03\x00\xAA\xD9\x03\x00\xFA\x51\x00\x80\xFE\x51\x00\x80\xAF\x6D"
-     "\x03\x00"
-     "\xAE\x65\x03\x00\xAD\x75\x03\x00\xAC\x75\x03\x00\x02\x52\x00\x80\x06\x52"
-     "\x00\x80"
-     "\x0A\x52\x00\x80\x0E\x52\x00\x80\xEF\xD4\x03\x00\x12\x52\x00\x80\x16\x52"
-     "\x00\x80"
-     "\x1A\x52\x00\x80\xE3\xE0\x06\x00\x1E\x52\x00\x80\xE1\xD8\x01\x00\x22\x52"
-     "\x00\x80"
-     "\x26\x52\x00\x80\x2A\x52\x00\x80\x2E\x52\x00\x80\x32\x52\x00\x80\x36\x52"
-     "\x00\x80"
-     "\xE3\x60\x06\x00\x84\x44\x02\x00\xE1\xFC\x04\x00\x80\x6D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x3A\x52\x00\x80\xBE\x6C\x0D\x00\x42\x52\x00\x80\x86\xC8"
-     "\x0C\x00"
-     "\x87\x5C\x0D\x00\x46\x52\x00\x80\xEF\xB4\x07\x00\x4A\x52\x00\x80\x4E\x52"
-     "\x00\x80"
-     "\x52\x52\x00\x80\x56\x52\x00\x80\x5A\x52\x00\x80\x5E\x52\x00\x80\x62\x52"
-     "\x00\x80"
-     "\x66\x52\x00\x80\x6A\x52\x00\x80\xEF\xE8\x07\x00\x84\xA0\x0D\x00\xE1\x1C"
-     "\x06\x00"
-     "\x6E\x52\x00\x80\xE3\x34\x06\x00\x72\x52\x00\x80\x76\x52\x00\x80\x7A\x52"
-     "\x00\x80"
-     "\x7E\x52\x00\x80\xB3\xF9\x01\x00\x82\x52\x00\x80\x86\x52\x00\x80\x8A\x52"
-     "\x00\x80"
-     "\x8E\x52\x00\x80\xB6\x11\x01\x00\xB5\x45\x01\x00\x92\x52\x00\x80\xBB\x3D"
-     "\x01\x00"
-     "\xBA\x35\x01\x00\x96\x52\x00\x80\x9A\x52\x00\x80\xBF\xF5\x00\x00\xBE\xF5"
-     "\x00\x00"
-     "\xBD\x15\x01\x00\xBC\x15\x01\x00\xA8\x31\x0E\x00\xA9\x31\x0E\x00\xAA\x31"
-     "\x0E\x00"
-     "\xAB\x9D\x0E\x00\xAC\x8D\x0E\x00\xAD\xD1\x0E\x00\xAE\xD1\x0E\x00\xAF\xD1"
-     "\x0E\x00"
-     "\x3E\x52\x00\x80\x9E\x52\x00\x80\xA2\x52\x00\x80\xA6\x52\x00\x80\x80\x19"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x82\x05\x00\x00\xAA\x52\x00\x80\xB8\x55\x0F\x00\xB9\x5D"
-     "\x0F\x00"
-     "\xBA\x55\x0F\x00\xBB\x69\x0F\x00\xBC\x7D\x0F\x00\xBD\x65\x0F\x00\xBE\x6D"
-     "\x0F\x00"
-     "\xBF\x65\x0F\x00\xB0\xB1\x0E\x00\xB1\xB9\x0E\x00\xB2\x81\x0E\x00\xB3\x81"
-     "\x0E\x00"
-     "\xB4\x71\x0F\x00\xB5\x71\x0F\x00\xB6\x71\x0F\x00\xB7\x71\x0F\x00\xA3\xB9"
-     "\x0E\x00"
-     "\xBE\x88\x03\x00\xAE\x52\x00\x80\xB2\x52\x00\x80\xB6\x52\x00\x80\xA6\x51"
-     "\x0E\x00"
-     "\xA5\x05\x0E\x00\xBA\x52\x00\x80\xAB\x7D\x0E\x00\xAA\x75\x0E\x00\x86\x08"
-     "\x00\x00"
-     "\x87\x6C\x03\x00\xAF\xB5\x0F\x00\xAE\xB5\x0F\x00\xAD\x55\x0E\x00\xAC\x55"
-     "\x0E\x00"
-     "\xBE\x52\x00\x80\xC2\x52\x00\x80\xC6\x52\x00\x80\xB3\xB5\x0F\x00\xCA\x52"
-     "\x00\x80"
-     "\xB5\x55\x0F\x00\xB6\x5D\x0F\x00\xCE\x51\x00\x80\xCE\x52\x00\x80\xD2\x52"
-     "\x00\x80"
-     "\xBA\x79\x0F\x00\xBB\x79\x0F\x00\xBC\x69\x0F\x00\xBD\x5D\x0F\x00\xBE\x49"
-     "\x0F\x00"
-     "\xBF\x49\x0F\x00\xA8\x65\x0E\x00\xA9\x75\x0E\x00\xAA\x71\x0E\x00\xAB\x71"
-     "\x0E\x00"
-     "\xAC\x91\x0E\x00\xAD\x91\x0E\x00\xAE\x91\x0E\x00\xAF\x91\x0E\x00\xD6\x52"
-     "\x00\x80"
-     "\xDA\x52\x00\x80\xDE\x52\x00\x80\xE2\x52\x00\x80\xE6\x52\x00\x80\xEA\x52"
-     "\x00\x80"
-     "\xEE\x52\x00\x80\xF2\x52\x00\x80\xB8\x85\x0E\x00\xB9\x8D\x0E\x00\xBA\x85"
-     "\x0E\x00"
-     "\xBB\x9D\x0E\x00\xBC\x8D\x0E\x00\xBD\xBD\x0E\x00\xBE\xB5\x0E\x00\xBF\x79"
-     "\x01\x00"
-     "\xB0\xF1\x0E\x00\xB1\xF1\x0E\x00\xB2\xF1\x0E\x00\xB3\xC5\x0E\x00\xB4\xC1"
-     "\x0E\x00"
-     "\xB5\xC1\x0E\x00\xB6\xC1\x0E\x00\xB7\xC1\x0E\x00\xA3\xF9\x0E\x00\xF6\x52"
-     "\x00\x80"
-     "\xFA\x52\x00\x80\xFE\x52\x00\x80\x02\x53\x00\x80\xA6\x11\x0E\x00\xA5\x19"
-     "\x0E\x00"
-     "\x06\x53\x00\x80\xAB\x35\x0E\x00\xAA\x35\x0E\x00\x0A\x53\x00\x80\x0E\x53"
-     "\x00\x80"
-     "\xAF\x05\x0E\x00\xAE\x05\x0E\x00\xAD\x11\x0E\x00\xAC\x25\x0E\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x82\x1D\x00\x00\x12\x53\x00\x80\x16\x53\x00\x80\x1A\x53"
-     "\x00\x80"
-     "\x84\x94\x01\x00\xBE\x94\x01\x00\x86\x40\x07\x00\x87\xF4\x00\x00\x22\x53"
-     "\x00\x80"
-     "\x26\x53\x00\x80\x2A\x53\x00\x80\x2E\x53\x00\x80\x32\x53\x00\x80\x36\x53"
-     "\x00\x80"
-     "\xA8\x8D\x02\x00\xA9\x95\x02\x00\xAA\x95\x02\x00\xAB\xCD\x02\x00\xAC\xD5"
-     "\x02\x00"
-     "\xAD\xDD\x02\x00\xAE\xC9\x02\x00\xAF\xFD\x02\x00\x3A\x53\x00\x80\x3E\x53"
-     "\x00\x80"
-     "\x42\x53\x00\x80\x46\x53\x00\x80\x8C\x00\x00\x00\x4A\x53\x00\x80\x4E\x53"
-     "\x00\x80"
-     "\x52\x53\x00\x80\xB8\x7D\x03\x00\xB9\xC1\x03\x00\xBA\xC1\x03\x00\xBB\xC1"
-     "\x03\x00"
-     "\xBC\xC1\x03\x00\xBD\xC9\x03\x00\xBE\xF1\x03\x00\xBF\xF1\x03\x00\xB0\x85"
-     "\x02\x00"
-     "\xB1\x45\x03\x00\xB2\x4D\x03\x00\xB3\x45\x03\x00\xB4\x5D\x03\x00\xB5\x45"
-     "\x03\x00"
-     "\xB6\x4D\x03\x00\xB7\x45\x03\x00\xB3\x1D\x02\x00\x56\x53\x00\x80\x5A\x53"
-     "\x00\x80"
-     "\x5E\x53\x00\x80\x62\x53\x00\x80\xB6\x5D\x02\x00\xB5\x5D\x02\x00\x66\x53"
-     "\x00\x80"
-     "\xBB\x81\x03\x00\xBA\x41\x02\x00\x6A\x53\x00\x80\x6E\x53\x00\x80\xBF\x81"
-     "\x03\x00"
-     "\xBE\x99\x03\x00\xBD\x91\x03\x00\xBC\x99\x03\x00\x72\x53\x00\x80\xA3\x59"
-     "\x02\x00"
-     "\x76\x53\x00\x80\x7A\x53\x00\x80\xA6\x19\x02\x00\x7E\x53\x00\x80\x82\x53"
-     "\x00\x80"
-     "\xA5\x19\x02\x00\xAA\x05\x02\x00\xAB\xC5\x03\x00\x86\x53\x00\x80\x8A\x53"
-     "\x00\x80"
-     "\xAE\xDD\x03\x00\xAF\xC5\x03\x00\xAC\xDD\x03\x00\xAD\xD5\x03\x00\x92\x53"
-     "\x00\x80"
-     "\xE3\xCC\x00\x00\x84\x04\x02\x00\xE1\xBC\x01\x00\x80\xC9\x01\x00\x81\xFD"
-     "\x01\x00"
-     "\x82\xF5\x01\x00\xBE\x10\x05\x00\x96\x53\x00\x80\xBE\x28\x04\x00\x9A\x53"
-     "\x00\x80"
-     "\x9E\x53\x00\x80\xA2\x53\x00\x80\xEF\x10\x00\x00\xA6\x53\x00\x80\xAA\x53"
-     "\x00\x80"
-     "\x86\xE0\x04\x00\x87\xF4\x02\x00\xAE\x53\x00\x80\xB2\x53\x00\x80\xE3\x08"
-     "\x0E\x00"
-     "\xB6\x53\x00\x80\xE1\xE0\x0F\x00\xBA\x53\x00\x80\xEF\xF8\x0E\x00\xBE\x53"
-     "\x00\x80"
-     "\xC2\x53\x00\x80\xC6\x53\x00\x80\xCA\x53\x00\x80\xCE\x53\x00\x80\xD2\x53"
-     "\x00\x80"
-     "\xD6\x53\x00\x80\xDA\x53\x00\x80\xDE\x53\x00\x80\xE2\x53\x00\x80\xE6\x53"
-     "\x00\x80"
-     "\xEA\x53\x00\x80\xEE\x53\x00\x80\xF2\x53\x00\x80\xEF\xE0\x01\x00\x85\x24"
-     "\x15\x00"
-     "\xE1\xD4\x0E\x00\xF6\x53\x00\x80\xE3\x1C\x0E\x00\x80\x29\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x82\x05\x00\x00\xFA\x53\x00\x80\xB3\x31\x02\x00\x02\x54\x00\x80\x84\xCC"
-     "\x05\x00"
-     "\x06\x54\x00\x80\x0A\x54\x00\x80\xB6\x29\x02\x00\xB5\x21\x02\x00\x0E\x54"
-     "\x00\x80"
-     "\xBB\xCD\x01\x00\xBA\xCD\x01\x00\x12\x54\x00\x80\x16\x54\x00\x80\xBF\x75"
-     "\x01\x00"
-     "\xBE\xC9\x01\x00\xBD\xC1\x01\x00\xBC\xC9\x01\x00\xA8\xE9\x05\x00\xA9\xE9"
-     "\x05\x00"
-     "\xAA\xF9\x05\x00\xAB\xF9\x05\x00\xAC\xE9\x05\x00\xAD\xE9\x05\x00\xAE\x39"
-     "\x06\x00"
-     "\xAF\x39\x06\x00\x8E\x53\x00\x80\xFE\x53\x00\x80\x86\x88\x00\x00\x87\x00"
-     "\x03\x00"
-     "\x1A\x54\x00\x80\x1E\x54\x00\x80\x22\x54\x00\x80\x26\x54\x00\x80\xB8\xD1"
-     "\x06\x00"
-     "\xB9\xD9\x06\x00\xBA\xE1\x06\x00\xBB\xE1\x06\x00\xBC\x91\x06\x00\xBD\x9D"
-     "\x06\x00"
-     "\xBE\x95\x06\x00\xBF\x89\x06\x00\xB0\x49\x06\x00\xB1\x49\x06\x00\xB2\x5D"
-     "\x06\x00"
-     "\xB3\x55\x06\x00\xB4\x4D\x06\x00\xB5\xF1\x06\x00\xB6\xF1\x06\x00\xB7\xF1"
-     "\x06\x00"
-     "\xA3\x71\x05\x00\x2A\x54\x00\x80\x2E\x54\x00\x80\x32\x54\x00\x80\x36\x54"
-     "\x00\x80"
-     "\xA6\x69\x05\x00\xA5\x61\x05\x00\x3A\x54\x00\x80\xAB\x8D\x06\x00\xAA\x8D"
-     "\x06\x00"
-     "\x3E\x54\x00\x80\x42\x54\x00\x80\xAF\x35\x06\x00\xAE\x89\x06\x00\xAD\x81"
-     "\x06\x00"
-     "\xAC\x89\x06\x00\x46\x54\x00\x80\x4A\x54\x00\x80\xB3\xE1\x07\x00\x4E\x54"
-     "\x00\x80"
-     "\xB5\xE1\x07\x00\x52\x54\x00\x80\x56\x54\x00\x80\xB6\xE5\x07\x00\x1E\x53"
-     "\x00\x80"
-     "\x5A\x54\x00\x80\xBB\xBD\x07\x00\xBA\xA1\x07\x00\xBD\x95\x07\x00\xBC\xA9"
-     "\x07\x00"
-     "\xBF\x95\x07\x00\xBE\x95\x07\x00\xA8\x01\x06\x00\xA9\x25\x06\x00\xAA\x21"
-     "\x06\x00"
-     "\xAB\x21\x06\x00\xAC\x21\x06\x00\xAD\x2D\x06\x00\xAE\x25\x06\x00\xAF\x55"
-     "\x06\x00"
-     "\x5E\x54\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\x62\x54"
-     "\x00\x80"
-     "\x66\x54\x00\x80\x6A\x54\x00\x80\xBE\x30\x01\x00\xB8\x39\x06\x00\xB9\x39"
-     "\x06\x00"
-     "\xBA\xC9\x06\x00\xBB\xC9\x06\x00\xBC\xD9\x06\x00\xBD\xD9\x06\x00\xBE\xC9"
-     "\x06\x00"
-     "\xBF\xC9\x06\x00\xB0\x2D\x06\x00\xB1\x31\x06\x00\xB2\x35\x06\x00\xB3\x09"
-     "\x06\x00"
-     "\xB4\x19\x06\x00\xB5\x19\x06\x00\xB6\x09\x06\x00\xB7\x09\x06\x00\xA3\xA9"
-     "\x06\x00"
-     "\x84\x8C\x02\x00\x86\x28\x1F\x00\x87\x44\x01\x00\x72\x54\x00\x80\xA6\xAD"
-     "\x06\x00"
-     "\xA5\xA9\x06\x00\x76\x54\x00\x80\xAB\xF5\x06\x00\xAA\xE9\x06\x00\x7A\x54"
-     "\x00\x80"
-     "\x7E\x54\x00\x80\xAF\xDD\x06\x00\xAE\xDD\x06\x00\xAD\xDD\x06\x00\xAC\xE1"
-     "\x06\x00"
-     "\x82\x54\x00\x80\xB3\x15\x06\x00\x86\x54\x00\x80\x8A\x54\x00\x80\xB6\x3D"
-     "\x06\x00"
-     "\x8E\x54\x00\x80\x92\x54\x00\x80\xB5\x35\x06\x00\xBA\xD9\x01\x00\xBB\xD9"
-     "\x01\x00"
-     "\x96\x54\x00\x80\x9A\x54\x00\x80\xBE\x7D\x01\x00\xBF\x65\x01\x00\xBC\x7D"
-     "\x01\x00"
-     "\xBD\x75\x01\x00\xA8\xC5\x05\x00\xA9\xC9\x05\x00\xAA\xD9\x05\x00\xAB\xD1"
-     "\x05\x00"
-     "\xAC\xF9\x05\x00\xAD\xF9\x05\x00\xAE\x29\x02\x00\xAF\x29\x02\x00\x9E\x54"
-     "\x00\x80"
-     "\xA2\x54\x00\x80\xA6\x54\x00\x80\xAA\x54\x00\x80\x8C\x00\x00\x00\xAE\x54"
-     "\x00\x80"
-     "\xB2\x54\x00\x80\xB6\x54\x00\x80\xB8\xED\x02\x00\xB9\x85\x02\x00\xBA\x8D"
-     "\x02\x00"
-     "\xBB\x81\x02\x00\xBC\x85\x02\x00\xBD\x8D\x02\x00\xBE\xBD\x02\x00\xBF\x7D"
-     "\x03\x00"
-     "\xB0\x59\x02\x00\xB1\x59\x02\x00\xB2\xED\x02\x00\xB3\xE5\x02\x00\xB4\xFD"
-     "\x02\x00"
-     "\xB5\xE5\x02\x00\xB6\xE5\x02\x00\xB7\xD5\x02\x00\xA3\x51\x05\x00\xBA\x54"
-     "\x00\x80"
-     "\xBE\x54\x00\x80\xC2\x54\x00\x80\xC6\x54\x00\x80\xA6\x79\x05\x00\xA5\x71"
-     "\x05\x00"
-     "\xCA\x54\x00\x80\xAB\x9D\x02\x00\xAA\x9D\x02\x00\xCE\x54\x00\x80\xD2\x54"
-     "\x00\x80"
-     "\xAF\x21\x02\x00\xAE\x39\x02\x00\xAD\x31\x02\x00\xAC\x39\x02\x00\x82\x11"
-     "\x00\x00"
-     "\xD6\x54\x00\x80\x80\x65\x00\x00\x81\x09\x00\x00\xDA\x54\x00\x80\xBE\x98"
-     "\x03\x00"
-     "\xE2\x54\x00\x80\xE6\x54\x00\x80\x84\x24\x03\x00\xEA\x54\x00\x80\x87\x68"
-     "\x03\x00"
-     "\x86\x8C\x1C\x00\xEE\x54\x00\x80\xF2\x54\x00\x80\xF6\x54\x00\x80\xFA\x54"
-     "\x00\x80"
-     "\xFE\x54\x00\x80\x02\x55\x00\x80\xB3\xA1\x02\x00\x84\xC0\x1C\x00\xB5\xD1"
-     "\x02\x00"
-     "\x06\x55\x00\x80\x0A\x55\x00\x80\xB6\xD5\x02\x00\x0E\x55\x00\x80\x12\x55"
-     "\x00\x80"
-     "\xBB\xC1\x02\x00\xBA\xF5\x02\x00\xBD\x11\x01\x00\xBC\xD9\x02\x00\xBF\x11"
-     "\x01\x00"
-     "\xBE\x19\x01\x00\x16\x55\x00\x80\x1A\x55\x00\x80\x1E\x55\x00\x80\x22\x55"
-     "\x00\x80"
-     "\x26\x55\x00\x80\x2A\x55\x00\x80\x2E\x55\x00\x80\xEF\xB8\x06\x00\x32\x55"
-     "\x00\x80"
-     "\xE1\x9C\x04\x00\x36\x55\x00\x80\xE3\x74\x06\x00\x3A\x55\x00\x80\x3E\x55"
-     "\x00\x80"
-     "\x42\x55\x00\x80\x46\x55\x00\x80\x80\xF9\x00\x00\x81\xF9\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x4A\x55\x00\x80\xBE\x58\x1C\x00\x84\x58\x1F\x00\xEF\x5C\x00\x00\xEF\x40"
-     "\x06\x00"
-     "\xE1\x90\x01\x00\xE1\xFC\x06\x00\xE3\x3C\x00\x00\xE3\x74\x06\x00\x52\x55"
-     "\x00\x80"
-     "\x56\x55\x00\x80\x86\x98\x1C\x00\x87\xF4\x1C\x00\xA3\x69\x02\x00\xBE\x0C"
-     "\x1F\x00"
-     "\x5A\x55\x00\x80\x5E\x55\x00\x80\x62\x55\x00\x80\xA6\x1D\x02\x00\xA5\x19"
-     "\x02\x00"
-     "\x66\x55\x00\x80\xAB\x09\x02\x00\xAA\x3D\x02\x00\x6A\x55\x00\x80\x6E\x55"
-     "\x00\x80"
-     "\xAF\xD9\x01\x00\xAE\xD1\x01\x00\xAD\xD9\x01\x00\xAC\x11\x02\x00\xA8\x91"
-     "\x1D\x00"
-     "\xA9\x99\x1D\x00\xAA\xA1\x1D\x00\xAB\xA1\x1D\x00\xAC\xD1\x1D\x00\xAD\xDD"
-     "\x1D\x00"
-     "\xAE\xD5\x1D\x00\xAF\xC9\x1D\x00\xDE\x54\x00\x80\x4E\x55\x00\x80\x72\x55"
-     "\x00\x80"
-     "\x76\x55\x00\x80\x7A\x55\x00\x80\x7E\x55\x00\x80\x82\x55\x00\x80\x86\x55"
-     "\x00\x80"
-     "\xB8\x79\x1E\x00\xB9\x79\x1E\x00\xBA\xCD\x1E\x00\xBB\xC5\x1E\x00\xBC\xDD"
-     "\x1E\x00"
-     "\xBD\xC5\x1E\x00\xBE\xC5\x1E\x00\xBF\xF5\x1E\x00\xB0\xB9\x1D\x00\xB1\x8D"
-     "\x1D\x00"
-     "\xB2\x85\x1D\x00\xB3\x4D\x1E\x00\xB4\x55\x1E\x00\xB5\x5D\x1E\x00\xB6\x55"
-     "\x1E\x00"
-     "\xB7\x49\x1E\x00\xB8\xCD\x1F\x00\xB9\xD5\x1F\x00\xBA\xDD\x1F\x00\xBB\xD5"
-     "\x1F\x00"
-     "\xBC\xF1\x1F\x00\xBD\xFD\x1F\x00\xBE\xE9\x1F\x00\xBF\xE9\x1F\x00\xB0\xA5"
-     "\x1F\x00"
-     "\xB1\xB1\x1F\x00\xB2\xB1\x1F\x00\xB3\x8D\x1F\x00\xB4\x95\x1F\x00\xB5\xF5"
-     "\x1F\x00"
-     "\xB6\xFD\x1F\x00\xB7\xF5\x1F\x00\xA8\x19\x1E\x00\xA9\x19\x1E\x00\xAA\x2D"
-     "\x1E\x00"
-     "\xAB\x3D\x1E\x00\xAC\x25\x1E\x00\xAD\x2D\x1E\x00\xAE\x25\x1E\x00\xAF\xDD"
-     "\x1F\x00"
-     "\x8A\x55\x00\x80\x8E\x55\x00\x80\x92\x55\x00\x80\x96\x55\x00\x80\x9A\x55"
-     "\x00\x80"
-     "\x6E\x54\x00\x80\x9E\x55\x00\x80\xA2\x55\x00\x80\xB3\xE5\x1F\x00\xA6\x55"
-     "\x00\x80"
-     "\xAA\x55\x00\x80\xAE\x55\x00\x80\xB2\x55\x00\x80\xB6\xED\x1F\x00\xB5\xED"
-     "\x1F\x00"
-     "\xB6\x55\x00\x80\xBB\x35\x1E\x00\xBA\x21\x1E\x00\xBA\x55\x00\x80\xBE\x55"
-     "\x00\x80"
-     "\xBF\x71\x1E\x00\xBE\x11\x1E\x00\xBD\x19\x1E\x00\xBC\x25\x1E\x00\x82\x69"
-     "\x00\x00"
-     "\xA3\xA1\x1F\x00\x80\x59\x00\x00\x81\x51\x00\x00\xA6\xA9\x1F\x00\xC2\x55"
-     "\x00\x80"
-     "\xC6\x55\x00\x80\xA5\xA9\x1F\x00\xAA\x65\x1E\x00\xAB\x71\x1E\x00\x86\x00"
-     "\x04\x00"
-     "\x87\xEC\x01\x00\xAE\x55\x1E\x00\xAF\x35\x1E\x00\xAC\x61\x1E\x00\xAD\x5D"
-     "\x1E\x00"
-     "\xA8\x31\x1E\x00\xA9\x31\x1E\x00\xAA\x41\x1E\x00\xAB\x41\x1E\x00\xAC\x41"
-     "\x1E\x00"
-     "\xAD\x49\x1E\x00\xAE\x71\x1E\x00\xAF\x71\x1E\x00\xCA\x55\x00\x80\xCE\x55"
-     "\x00\x80"
-     "\xD2\x55\x00\x80\xD6\x55\x00\x80\xDA\x55\x00\x80\xDE\x55\x00\x80\xE2\x55"
-     "\x00\x80"
-     "\xE6\x55\x00\x80\xB8\x29\x01\x00\xB9\x29\x01\x00\xBA\x39\x01\x00\xBB\x35"
-     "\x01\x00"
-     "\xBC\x2D\x01\x00\xBD\xD1\x00\x00\xBE\xD1\x00\x00\xBF\xD1\x00\x00\xB0\xC9"
-     "\x01\x00"
-     "\xB1\xC9\x01\x00\xB2\xD9\x01\x00\xB3\xD9\x01\x00\xB4\xC9\x01\x00\xB5\xC9"
-     "\x01\x00"
-     "\xB6\x19\x01\x00\xB7\x19\x01\x00\xB3\xC9\x1D\x00\xEA\x55\x00\x80\xEE\x55"
-     "\x00\x80"
-     "\xF2\x55\x00\x80\xF6\x55\x00\x80\xB6\xC9\x1D\x00\xB5\xC9\x1D\x00\xFA\x55"
-     "\x00\x80"
-     "\xBB\x0D\x02\x00\xBA\x0D\x02\x00\xFE\x55\x00\x80\x02\x56\x00\x80\xBF\x0D"
-     "\x02\x00"
-     "\xBE\x0D\x02\x00\xBD\x0D\x02\x00\xBC\x0D\x02\x00\x06\x56\x00\x80\xA3\x8D"
-     "\x1D\x00"
-     "\x0A\x56\x00\x80\x0E\x56\x00\x80\xA6\x8D\x1D\x00\x12\x56\x00\x80\x16\x56"
-     "\x00\x80"
-     "\xA5\x8D\x1D\x00\xAA\x49\x02\x00\xAB\x49\x02\x00\x1A\x56\x00\x80\x1E\x56"
-     "\x00\x80"
-     "\xAE\x49\x02\x00\xAF\x49\x02\x00\xAC\x49\x02\x00\xAD\x49\x02\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x11\x00\x00\x82\x11\x00\x00\x22\x56\x00\x80\xEF\xCC\x02\x00\x26\x56"
-     "\x00\x80"
-     "\x2A\x56\x00\x80\x84\x90\x02\x00\xE3\x2C\x02\x00\xBE\x28\x03\x00\xE1\xD8"
-     "\x01\x00"
-     "\x32\x56\x00\x80\x86\x10\x04\x00\x87\xD4\x03\x00\x36\x56\x00\x80\x3A\x56"
-     "\x00\x80"
-     "\xB3\x61\x03\x00\x3E\x56\x00\x80\x42\x56\x00\x80\x46\x56\x00\x80\x4A\x56"
-     "\x00\x80"
-     "\xB6\x79\x03\x00\xB5\x71\x03\x00\x4E\x56\x00\x80\xBB\x5D\x03\x00\xBA\x5D"
-     "\x03\x00"
-     "\x52\x56\x00\x80\x56\x56\x00\x80\xBF\xE1\x00\x00\xBE\xF9\x00\x00\xBD\xF1"
-     "\x00\x00"
-     "\xBC\xF9\x00\x00\xA3\xA1\x02\x00\x5A\x56\x00\x80\x5E\x56\x00\x80\x62\x56"
-     "\x00\x80"
-     "\x66\x56\x00\x80\xA6\xB9\x02\x00\xA5\xB1\x02\x00\x6A\x56\x00\x80\xAB\x9D"
-     "\x02\x00"
-     "\xAA\x9D\x02\x00\x6E\x56\x00\x80\x72\x56\x00\x80\xAF\x21\x01\x00\xAE\x39"
-     "\x01\x00"
-     "\xAD\x31\x01\x00\xAC\x39\x01\x00\x76\x56\x00\x80\x7A\x56\x00\x80\xEF\x64"
-     "\x1F\x00"
-     "\x7E\x56\x00\x80\x82\x56\x00\x80\x86\x56\x00\x80\x8A\x56\x00\x80\x8E\x56"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\x11\x00\x00\x82\x15\x00\x00\x92\x56\x00\x80\xE1\xE0"
-     "\x1C\x00"
-     "\x96\x56\x00\x80\xE3\x88\x1F\x00\x9A\x56\x00\x80\x84\x80\x02\x00\xBE\x8C"
-     "\x05\x00"
-     "\x87\x48\x05\x00\x86\x2C\x04\x00\xA2\x56\x00\x80\xA6\x56\x00\x80\xEF\xA4"
-     "\x1E\x00"
-     "\xEF\xF4\x1E\x00\xE1\x00\x1E\x00\xE1\x50\x1F\x00\xE3\x2C\x1E\x00\xE3\xB0"
-     "\x1E\x00"
-     "\xAA\x56\x00\x80\xAE\x56\x00\x80\xB2\x56\x00\x80\xB6\x56\x00\x80\xBA\x56"
-     "\x00\x80"
-     "\xBE\x56\x00\x80\x84\x84\x04\x00\xB3\x71\x01\x00\xC2\x56\x00\x80\xB5\x1D"
-     "\x01\x00"
-     "\xB6\x15\x01\x00\xC6\x56\x00\x80\xCA\x56\x00\x80\xCE\x56\x00\x80\xBA\x01"
-     "\x01\x00"
-     "\xBB\x01\x01\x00\xBC\xF5\x00\x00\xBD\xFD\x00\x00\xBE\xF5\x00\x00\xBF\xED"
-     "\x00\x00"
-     "\xA8\xAD\x06\x00\xA9\xBD\x06\x00\xAA\xB5\x06\x00\xAB\xC9\x06\x00\xAC\xD9"
-     "\x06\x00"
-     "\xAD\xD9\x06\x00\xAE\xC9\x06\x00\xAF\xC1\x06\x00\xD2\x56\x00\x80\xD6\x56"
-     "\x00\x80"
-     "\xDA\x56\x00\x80\xDE\x56\x00\x80\xE2\x56\x00\x80\xE6\x56\x00\x80\xEA\x56"
-     "\x00\x80"
-     "\xEE\x56\x00\x80\xB8\x6D\x07\x00\xB9\x05\x07\x00\xBA\x0D\x07\x00\xBB\x01"
-     "\x07\x00"
-     "\xBC\x01\x07\x00\xBD\x01\x07\x00\xBE\x01\x07\x00\xBF\x01\x07\x00\xB0\x89"
-     "\x06\x00"
-     "\xB1\x89\x06\x00\xB2\x6D\x07\x00\xB3\x65\x07\x00\xB4\x7D\x07\x00\xB5\x65"
-     "\x07\x00"
-     "\xB6\x65\x07\x00\xB7\x55\x07\x00\x2E\x56\x00\x80\xA3\x39\x06\x00\xF2\x56"
-     "\x00\x80"
-     "\x9E\x56\x00\x80\xA6\x5D\x06\x00\xF6\x56\x00\x80\xFA\x56\x00\x80\xA5\x55"
-     "\x06\x00"
-     "\xAA\x49\x06\x00\xAB\x49\x06\x00\xFE\x56\x00\x80\x02\x57\x00\x80\xAE\xBD"
-     "\x07\x00"
-     "\xAF\xA5\x07\x00\xAC\xBD\x07\x00\xAD\xB5\x07\x00\x80\x6D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x06\x57\x00\x80\x0A\x57\x00\x80\xBE\x9C\x03\x00\x0E\x57"
-     "\x00\x80"
-     "\x12\x57\x00\x80\x86\x40\x00\x00\x87\x60\x00\x00\x16\x57\x00\x80\x1A\x57"
-     "\x00\x80"
-     "\x1E\x57\x00\x80\x22\x57\x00\x80\x26\x57\x00\x80\x2A\x57\x00\x80\xA8\x91"
-     "\x06\x00"
-     "\xA9\x91\x06\x00\xAA\xB9\x06\x00\xAB\xB9\x06\x00\xAC\xA9\x06\x00\xAD\xA9"
-     "\x06\x00"
-     "\xAE\xD9\x06\x00\xAF\xD9\x06\x00\x2E\x57\x00\x80\x32\x57\x00\x80\x36\x57"
-     "\x00\x80"
-     "\x3A\x57\x00\x80\x3E\x57\x00\x80\x42\x57\x00\x80\x46\x57\x00\x80\x4A\x57"
-     "\x00\x80"
-     "\xB8\x71\x01\x00\xB9\x71\x01\x00\xBA\x71\x01\x00\xBB\x71\x01\x00\xBC\xD9"
-     "\x01\x00"
-     "\xBD\xC1\x01\x00\xBE\xC1\x01\x00\xBF\xF5\x01\x00\xB0\xB1\x06\x00\xB1\xB9"
-     "\x06\x00"
-     "\xB2\x89\x06\x00\xB3\x89\x06\x00\xB4\x51\x01\x00\xB5\x51\x01\x00\xB6\x51"
-     "\x01\x00"
-     "\xB7\x51\x01\x00\xB3\x31\x06\x00\x4E\x57\x00\x80\x52\x57\x00\x80\x56\x57"
-     "\x00\x80"
-     "\x5A\x57\x00\x80\xB6\x29\x06\x00\xB5\x21\x06\x00\x5E\x57\x00\x80\xBB\x7D"
-     "\x06\x00"
-     "\xBA\x75\x06\x00\x62\x57\x00\x80\x66\x57\x00\x80\xBF\x95\x01\x00\xBE\x95"
-     "\x01\x00"
-     "\xBD\x5D\x06\x00\xBC\x5D\x06\x00\x6A\x57\x00\x80\xA3\x75\x06\x00\x6E\x57"
-     "\x00\x80"
-     "\x72\x57\x00\x80\xA6\x6D\x06\x00\x76\x57\x00\x80\x7A\x57\x00\x80\xA5\x65"
-     "\x06\x00"
-     "\xAA\x31\x06\x00\xAB\x39\x06\x00\x84\xAC\x01\x00\xBE\xA0\x01\x00\xAE\xD1"
-     "\x01\x00"
-     "\xAF\xD1\x01\x00\xAC\x19\x06\x00\xAD\x19\x06\x00\xA8\xDD\x02\x00\xA9\xED"
-     "\x02\x00"
-     "\xAA\xE5\x02\x00\xAB\xFD\x02\x00\xAC\xE5\x02\x00\xAD\xED\x02\x00\xAE\xE5"
-     "\x02\x00"
-     "\xAF\x3D\x03\x00\x82\x57\x00\x80\x86\x57\x00\x80\x8A\x57\x00\x80\xBE\x64"
-     "\x0C\x00"
-     "\x8E\x57\x00\x80\x92\x57\x00\x80\x96\x57\x00\x80\x9A\x57\x00\x80\xB8\xC9"
-     "\x03\x00"
-     "\xB9\xC9\x03\x00\xBA\xD9\x03\x00\xBB\xD1\x03\x00\xBC\xF9\x03\x00\xBD\xF9"
-     "\x03\x00"
-     "\xBE\x99\x03\x00\xBF\x95\x03\x00\xB0\x45\x03\x00\xB1\x4D\x03\x00\xB2\x45"
-     "\x03\x00"
-     "\xB3\x5D\x03\x00\xB4\x45\x03\x00\xB5\x4D\x03\x00\xB6\x45\x03\x00\xB7\xF9"
-     "\x03\x00"
-     "\x81\x55\x03\x00\x80\x49\x03\x00\xB3\x65\x02\x00\x82\x55\x03\x00\xB5\x65"
-     "\x02\x00"
-     "\x9E\x57\x00\x80\xA2\x57\x00\x80\xB6\x65\x02\x00\x86\x00\x0C\x00\x87\xE4"
-     "\x03\x00"
-     "\xBB\x81\x03\x00\xBA\x89\x03\x00\xBD\x81\x03\x00\xBC\x99\x03\x00\xBF\x81"
-     "\x03\x00"
-     "\xBE\x89\x03\x00\xA3\x2D\x02\x00\xA6\x57\x00\x80\xAA\x57\x00\x80\xAE\x57"
-     "\x00\x80"
-     "\xB2\x57\x00\x80\xA6\x2D\x02\x00\xA5\x2D\x02\x00\xB6\x57\x00\x80\xAB\xC9"
-     "\x03\x00"
-     "\xAA\xC1\x03\x00\xBA\x57\x00\x80\xBE\x57\x00\x80\xAF\xC9\x03\x00\xAE\xC1"
-     "\x03\x00"
-     "\xAD\xC9\x03\x00\xAC\xD1\x03\x00\xE3\xD8\x06\x00\xE1\xAC\x07\x00\xE1\x9C"
-     "\x06\x00"
-     "\xE3\x9C\x06\x00\xC2\x57\x00\x80\x84\x58\x0D\x00\xC6\x57\x00\x80\xCA\x57"
-     "\x00\x80"
-     "\xCE\x57\x00\x80\xD2\x57\x00\x80\xD6\x57\x00\x80\xDA\x57\x00\x80\xEF\x1C"
-     "\x01\x00"
-     "\xDE\x57\x00\x80\xE2\x57\x00\x80\xEF\x40\x06\x00\x80\x79\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x11\x00\x00\x84\x00\x0C\x00\xE6\x57\x00\x80\xE3\xAC\x00\x00\xEA\x57"
-     "\x00\x80"
-     "\xE1\xA4\x01\x00\xF2\x57\x00\x80\xEF\x70\x00\x00\x86\xC8\x0C\x00\x87\xA4"
-     "\x0D\x00"
-     "\xF6\x57\x00\x80\xFA\x57\x00\x80\xFE\x57\x00\x80\x02\x58\x00\x80\xBA\xC9"
-     "\x05\x00"
-     "\xBB\xC9\x05\x00\xB8\xA5\x05\x00\xB9\xCD\x05\x00\xBE\xF9\x05\x00\xBF\xF9"
-     "\x05\x00"
-     "\xBC\xCD\x05\x00\xBD\xC5\x05\x00\xB2\xA5\x05\x00\xB3\xAD\x05\x00\xB0\x11"
-     "\x06\x00"
-     "\xB1\x11\x06\x00\xB6\xAD\x05\x00\xB7\x9D\x05\x00\xB4\xB5\x05\x00\xB5\xAD"
-     "\x05\x00"
-     "\xAA\x61\x06\x00\xAB\x61\x06\x00\xA8\x9D\x06\x00\xA9\x95\x06\x00\xAE\x61"
-     "\x06\x00"
-     "\xAF\x61\x06\x00\xAC\x71\x06\x00\xAD\x71\x06\x00\x06\x58\x00\x80\x0A\x58"
-     "\x00\x80"
-     "\x0E\x58\x00\x80\x12\x58\x00\x80\x16\x58\x00\x80\x1A\x58\x00\x80\xBE\xB0"
-     "\x0C\x00"
-     "\x1E\x58\x00\x80\xA8\x21\x0E\x00\xA9\x21\x0E\x00\xAA\x21\x0E\x00\xAB\x3D"
-     "\x0E\x00"
-     "\xAC\x25\x0E\x00\xAD\x2D\x0E\x00\xAE\x25\x0E\x00\xAF\x89\x0E\x00\xEE\x57"
-     "\x00\x80"
-     "\x22\x58\x00\x80\x26\x58\x00\x80\x2A\x58\x00\x80\x2E\x58\x00\x80\x32\x58"
-     "\x00\x80"
-     "\x36\x58\x00\x80\x3A\x58\x00\x80\xB8\x51\x0F\x00\xB9\x5D\x0F\x00\xBA\x55"
-     "\x0F\x00"
-     "\xBB\x6D\x0F\x00\xBC\x75\x0F\x00\xBD\x7D\x0F\x00\xBE\x75\x0F\x00\xBF\x69"
-     "\x0F\x00"
-     "\xB0\xF9\x0E\x00\xB1\xA1\x0E\x00\xB2\xA1\x0E\x00\xB3\xA1\x0E\x00\xB4\xA1"
-     "\x0E\x00"
-     "\xB5\xA9\x0E\x00\xB6\x91\x0E\x00\xB7\x91\x0E\x00\xB3\xA9\x0E\x00\x3E\x58"
-     "\x00\x80"
-     "\x7E\x57\x00\x80\x42\x58\x00\x80\x46\x58\x00\x80\xB6\xAD\x0E\x00\xB5\xAD"
-     "\x0E\x00"
-     "\x4A\x58\x00\x80\xBB\x65\x0E\x00\xBA\x3D\x0E\x00\x4E\x58\x00\x80\x52\x58"
-     "\x00\x80"
-     "\xBF\x6D\x0E\x00\xBE\x65\x0E\x00\xBD\x75\x0E\x00\xBC\x75\x0E\x00\x82\x19"
-     "\x00\x00"
-     "\xA3\xED\x0E\x00\x80\x65\x00\x00\x81\x19\x00\x00\xA6\xE9\x0E\x00\x56\x58"
-     "\x00\x80"
-     "\x5A\x58\x00\x80\xA5\xE9\x0E\x00\xAA\x79\x0E\x00\xAB\x21\x0E\x00\x5E\x58"
-     "\x00\x80"
-     "\x62\x58\x00\x80\xAE\x21\x0E\x00\xAF\x29\x0E\x00\xAC\x31\x0E\x00\xAD\x31"
-     "\x0E\x00"
-     "\xA8\x61\x0E\x00\xA9\x75\x0E\x00\xAA\x7D\x0E\x00\xAB\x75\x0E\x00\xAC\x6D"
-     "\x0E\x00"
-     "\xAD\xF5\x0E\x00\xAE\xFD\x0E\x00\xAF\xF5\x0E\x00\x86\x80\x01\x00\x87\xA4"
-     "\x01\x00"
-     "\x66\x58\x00\x80\x6A\x58\x00\x80\x6E\x58\x00\x80\x72\x58\x00\x80\x76\x58"
-     "\x00\x80"
-     "\x7A\x58\x00\x80\xB8\x75\x01\x00\xB9\x7D\x01\x00\xBA\x75\x01\x00\xBB\xC9"
-     "\x01\x00"
-     "\xBC\xDD\x01\x00\xBD\xC5\x01\x00\xBE\xC5\x01\x00\xBF\xF5\x01\x00\xB0\x8D"
-     "\x0E\x00"
-     "\xB1\x95\x0E\x00\xB2\x9D\x0E\x00\xB3\x91\x0E\x00\xB4\x55\x01\x00\xB5\x5D"
-     "\x01\x00"
-     "\xB6\x55\x01\x00\xB7\x4D\x01\x00\xB3\xF5\x0E\x00\x7E\x58\x00\x80\x82\x58"
-     "\x00\x80"
-     "\x86\x58\x00\x80\x8A\x58\x00\x80\xB6\x75\x0E\x00\xB5\xE5\x0E\x00\x8E\x58"
-     "\x00\x80"
-     "\xBB\x51\x0E\x00\xBA\x49\x0E\x00\x92\x58\x00\x80\x96\x58\x00\x80\xBF\x99"
-     "\x01\x00"
-     "\xBE\x91\x01\x00\xBD\x45\x0E\x00\xBC\x49\x0E\x00\x9A\x58\x00\x80\xA3\xB1"
-     "\x0E\x00"
-     "\x9E\x58\x00\x80\xA2\x58\x00\x80\xA6\x31\x0E\x00\xA6\x58\x00\x80\xAA\x58"
-     "\x00\x80"
-     "\xA5\xA1\x0E\x00\xAA\x0D\x0E\x00\xAB\x15\x0E\x00\xAE\x58\x00\x80\xB2\x58"
-     "\x00\x80"
-     "\xAE\xD5\x01\x00\xAF\xDD\x01\x00\xAC\x0D\x0E\x00\xAD\x01\x0E\x00\xA8\xED"
-     "\x02\x00"
-     "\xA9\x2D\x03\x00\xAA\x25\x03\x00\xAB\x3D\x03\x00\xAC\x25\x03\x00\xAD\x2D"
-     "\x03\x00"
-     "\xAE\x25\x03\x00\xAF\x99\x03\x00\xB6\x58\x00\x80\xBA\x58\x00\x80\xBE\x58"
-     "\x00\x80"
-     "\xC2\x58\x00\x80\xC6\x58\x00\x80\xCA\x58\x00\x80\xBE\x74\x02\x00\xD2\x58"
-     "\x00\x80"
-     "\xB8\x8D\x03\x00\xB9\x91\x03\x00\xBA\x91\x03\x00\xBB\xA5\x03\x00\xBC\xBD"
-     "\x03\x00"
-     "\xBD\x75\x00\x00\xBE\x7D\x00\x00\xBF\x75\x00\x00\xB0\xE9\x03\x00\xB1\xE9"
-     "\x03\x00"
-     "\xB2\xF9\x03\x00\xB3\xF1\x03\x00\xB4\xD9\x03\x00\xB5\xD9\x03\x00\xB6\xB9"
-     "\x03\x00"
-     "\xB7\xB5\x03\x00\x80\xAD\x00\x00\x81\xB5\x00\x00\x82\xBD\x00\x00\xB3\xA1"
-     "\x03\x00"
-     "\xD6\x58\x00\x80\xB5\xA1\x03\x00\xB6\xA1\x03\x00\xDA\x58\x00\x80\x84\xE0"
-     "\x02\x00"
-     "\xDE\x58\x00\x80\xBA\x21\x03\x00\xBB\x21\x03\x00\xBC\x21\x03\x00\xBD\x29"
-     "\x03\x00"
-     "\xBE\x11\x03\x00\xBF\x11\x03\x00\xA3\xED\x03\x00\xE2\x58\x00\x80\x86\xE8"
-     "\x04\x00"
-     "\x87\x4C\x03\x00\xE6\x58\x00\x80\xA6\xED\x03\x00\xA5\xED\x03\x00\xEA\x58"
-     "\x00\x80"
-     "\xAB\x6D\x03\x00\xAA\x6D\x03\x00\xEE\x58\x00\x80\xF2\x58\x00\x80\xAF\x5D"
-     "\x03\x00"
-     "\xAE\x5D\x03\x00\xAD\x65\x03\x00\xAC\x6D\x03\x00\xF6\x58\x00\x80\xE3\x00"
-     "\x0E\x00"
-     "\xEF\xD4\x23\x00\xE1\xB4\x0F\x00\xFA\x58\x00\x80\xFE\x58\x00\x80\x02\x59"
-     "\x00\x80"
-     "\x06\x59\x00\x80\xA1\xA9\x03\x00\xA0\xFD\x0F\x00\xA3\xC1\x03\x00\xA2\x0D"
-     "\x03\x00"
-     "\xE1\x84\x12\x00\xEF\xE0\x0F\x00\xE3\xE4\xF3\x00\x0A\x59\x00\x80\x0E\x59"
-     "\x00\x80"
-     "\x12\x59\x00\x80\x84\x04\x04\x00\xBE\x04\x04\x00\x16\x59\x00\x80\xEF\x94"
-     "\x03\x00"
-     "\x1A\x59\x00\x80\x1E\x59\x00\x80\x22\x59\x00\x80\xE3\xD4\x03\x00\x26\x59"
-     "\x00\x80"
-     "\xE1\x54\x00\x00\x2A\x59\x00\x80\x2E\x59\x00\x80\x32\x59\x00\x80\x36\x59"
-     "\x00\x80"
-     "\x80\xF9\x00\x00\x81\xF9\x00\x00\x82\x05\x00\x00\x3A\x59\x00\x80\x42\x59"
-     "\x00\x80"
-     "\x46\x59\x00\x80\x8F\x91\x1B\x00\xEF\x9C\x0E\x00\x84\xEC\x07\x00\xE1\xD4"
-     "\x0E\x00"
-     "\x4A\x59\x00\x80\xE3\xF0\x0E\x00\x4E\x59\x00\x80\x52\x59\x00\x80\x86\x18"
-     "\x07\x00"
-     "\x87\xE8\x04\x00\x99\xE5\x17\x00\x98\xE5\x17\x00\x9B\xE9\x0B\x00\x9A\x39"
-     "\x0B\x00"
-     "\x9D\xFD\x0A\x00\x9C\xF1\x0B\x00\x9F\x55\x0F\x00\x9E\x5D\x0F\x00\x91\x29"
-     "\x1F\x00"
-     "\x90\xF5\x1B\x00\x93\x45\x1F\x00\x92\xCD\x1F\x00\x95\xD1\x13\x00\x94\x29"
-     "\x13\x00"
-     "\x97\x19\x17\x00\x96\x75\x13\x00\xA1\x05\x00\x00\xCE\x58\x00\x80\xA3\x01"
-     "\xFC\x00"
-     "\xA2\x55\x00\x00\xA5\x01\xFC\x00\xA4\x11\xFC\x00\xA7\x19\xF8\x00\xA6\x45"
-     "\xFD\x00"
-     "\xA9\x09\xF9\x00\xA8\x21\xF8\x00\xAB\x09\xF4\x00\xAA\x01\xF4\x00\xAD\x01"
-     "\xF0\x00"
-     "\xAC\x15\xF5\x00\xAF\xF1\xF0\x00\xAE\x1D\xF0\x00\xB1\x01\xEC\x00\xB0\x01"
-     "\xEC\x00"
-     "\xB3\x01\xE8\x00\xB2\x11\xEC\x00\xB5\x4D\xE8\x00\xB4\x51\xE8\x00\x3E\x59"
-     "\x00\x80"
-     "\x56\x59\x00\x80\x5A\x59\x00\x80\x5E\x59\x00\x80\x62\x59\x00\x80\x66\x59"
-     "\x00\x80"
-     "\x6A\x59\x00\x80\x6E\x59\x00\x80\x72\x59\x00\x80\x76\x59\x00\x80\xA8\xBD"
-     "\x06\x00"
-     "\xA9\x15\x06\x00\xAA\x15\x06\x00\xAB\x29\x06\x00\xAC\x55\x06\x00\xAD\x45"
-     "\x06\x00"
-     "\xAE\x45\x06\x00\xAF\x75\x06\x00\xBE\xE8\x00\x00\x7A\x59\x00\x80\x7E\x59"
-     "\x00\x80"
-     "\x82\x59\x00\x80\x86\x59\x00\x80\x8A\x59\x00\x80\x8E\x59\x00\x80\x92\x59"
-     "\x00\x80"
-     "\xB8\x85\x07\x00\xB9\x95\x07\x00\xBA\x91\x07\x00\xBB\x91\x07\x00\xBC\xB1"
-     "\x07\x00"
-     "\xBD\xBD\x07\x00\xBE\xA9\x07\x00\xBF\xA9\x07\x00\xB0\x0D\x06\x00\xB1\xE5"
-     "\x07\x00"
-     "\xB2\xF9\x07\x00\xB3\xF9\x07\x00\xB4\xE9\x07\x00\xB5\xE9\x07\x00\xB6\xC5"
-     "\x07\x00"
-     "\xB7\xBD\x07\x00\x8A\x21\x15\x00\x8B\x8D\x15\x00\x96\x59\x00\x80\x9A\x59"
-     "\x00\x80"
-     "\x8E\x35\x11\x00\x8F\xE1\x0D\x00\x8C\xF9\x11\x00\x8D\x3D\x11\x00\x82\x29"
-     "\x1D\x00"
-     "\x83\x35\x1D\x00\x9E\x59\x00\x80\xA2\x59\x00\x80\x86\x9D\x19\x00\x87\xE9"
-     "\x15\x00"
-     "\x84\xF1\x19\x00\x85\x1D\x19\x00\x9A\xC9\x05\x00\x9B\x65\x05\x00\xA6\x59"
-     "\x00\x80"
-     "\xAA\x59\x00\x80\x81\x19\x00\x00\x80\x79\x00\x00\x9C\xB1\x05\x00\x82\x11"
-     "\x00\x00"
-     "\x92\x41\x0D\x00\x93\x41\x0D\x00\x86\xC8\x00\x00\x87\xA4\x01\x00\x96\xFD"
-     "\x09\x00"
-     "\x97\xAD\x09\x00\x94\x55\x0C\x00\x95\xFD\x09\x00\xB3\x1D\x07\x00\xB2\x59"
-     "\x00\x80"
-     "\xB6\x59\x00\x80\xBA\x59\x00\x80\xBE\x59\x00\x80\xB6\x45\x07\x00\xB5\x5D"
-     "\x07\x00"
-     "\xC2\x59\x00\x80\xBB\xA1\x06\x00\xBA\x49\x07\x00\xC6\x59\x00\x80\xCA\x59"
-     "\x00\x80"
-     "\xBF\xAD\x06\x00\xBE\xAD\x06\x00\xBD\xAD\x06\x00\xBC\xAD\x06\x00\xCE\x59"
-     "\x00\x80"
-     "\xA3\x59\x07\x00\xD2\x59\x00\x80\xD6\x59\x00\x80\xA6\x01\x07\x00\xDA\x59"
-     "\x00\x80"
-     "\xDE\x59\x00\x80\xA5\x19\x07\x00\xAA\x0D\x07\x00\xAB\xE5\x06\x00\xE2\x59"
-     "\x00\x80"
-     "\xE6\x59\x00\x80\xAE\xE9\x06\x00\xAF\xE9\x06\x00\xAC\xE9\x06\x00\xAD\xE9"
-     "\x06\x00"
-     "\xA8\x91\x05\x00\xA9\x91\x05\x00\xAA\xA1\x05\x00\xAB\xBD\x05\x00\xAC\xA5"
-     "\x05\x00"
-     "\xAD\x49\x02\x00\xAE\x79\x02\x00\xAF\x79\x02\x00\x84\xEC\xFC\x00\xEA\x59"
-     "\x00\x80"
-     "\xEE\x59\x00\x80\xF2\x59\x00\x80\xF6\x59\x00\x80\xFA\x59\x00\x80\xFE\x59"
-     "\x00\x80"
-     "\x02\x5A\x00\x80\xB8\xED\x02\x00\xB9\xF5\x02\x00\xBA\xFD\x02\x00\xBB\xF5"
-     "\x02\x00"
-     "\xBC\xED\x02\x00\xBD\x5D\x03\x00\xBE\x55\x03\x00\xBF\x75\x03\x00\xB0\x09"
-     "\x02\x00"
-     "\xB1\x09\x02\x00\xB2\x19\x02\x00\xB3\x19\x02\x00\xB4\x09\x02\x00\xB5\x09"
-     "\x02\x00"
-     "\xB6\xDD\x02\x00\xB7\xD5\x02\x00\xB3\x29\x05\x00\x84\x48\xFD\x00\x06\x5A"
-     "\x00\x80"
-     "\x0A\x5A\x00\x80\x0E\x5A\x00\x80\xB6\xC9\x02\x00\xB5\x2D\x05\x00\x12\x5A"
-     "\x00\x80"
-     "\xBB\xCD\x02\x00\xBA\xCD\x02\x00\x16\x5A\x00\x80\x1A\x5A\x00\x80\xBF\xCD"
-     "\x02\x00"
-     "\xBE\xCD\x02\x00\xBD\xCD\x02\x00\xBC\xCD\x02\x00\x82\x6D\x00\x00\xA3\x6D"
-     "\x05\x00"
-     "\x80\x55\x00\x00\x81\x65\x00\x00\xA6\x8D\x02\x00\xBE\x1C\xFC\x00\x22\x5A"
-     "\x00\x80"
-     "\xA5\x69\x05\x00\xAA\x89\x02\x00\xAB\x89\x02\x00\x86\xC8\xFC\x00\x87\x1C"
-     "\x03\x00"
-     "\xAE\x89\x02\x00\xAF\x89\x02\x00\xAC\x89\x02\x00\xAD\x89\x02\x00\xA8\x85"
-     "\x02\x00"
-     "\xA9\x8D\x02\x00\xAA\x85\x02\x00\xAB\x9D\x02\x00\xAC\x8D\x02\x00\xAD\xBD"
-     "\x02\x00"
-     "\xAE\xB5\x02\x00\xAF\xDD\x02\x00\x26\x5A\x00\x80\x2A\x5A\x00\x80\xBE\xE0"
-     "\xFC\x00"
-     "\x2E\x5A\x00\x80\x32\x5A\x00\x80\x36\x5A\x00\x80\x3A\x5A\x00\x80\x3E\x5A"
-     "\x00\x80"
-     "\xB8\x6D\x01\x00\xB9\x75\x01\x00\xBA\x7D\x01\x00\xBB\x75\x01\x00\xBC\x6D"
-     "\x01\x00"
-     "\xBD\xDD\x01\x00\xBE\xD5\x01\x00\xBF\xCD\x01\x00\xB0\xA9\x02\x00\xB1\xA9"
-     "\x02\x00"
-     "\xB2\xBD\x02\x00\xB3\xB5\x02\x00\xB4\x99\x02\x00\xB5\x99\x02\x00\xB6\x5D"
-     "\x01\x00"
-     "\xB7\x55\x01\x00\xE1\x50\x07\x00\x42\x5A\x00\x80\xE3\xD0\x06\x00\x46\x5A"
-     "\x00\x80"
-     "\x4A\x5A\x00\x80\x4E\x5A\x00\x80\x52\x5A\x00\x80\x56\x5A\x00\x80\x5A\x5A"
-     "\x00\x80"
-     "\x5E\x5A\x00\x80\x62\x5A\x00\x80\x66\x5A\x00\x80\x6A\x5A\x00\x80\x6E\x5A"
-     "\x00\x80"
-     "\xEF\x0C\x06\x00\x72\x5A\x00\x80\x76\x5A\x00\x80\x7A\x5A\x00\x80\x7E\x5A"
-     "\x00\x80"
-     "\x82\x5A\x00\x80\x86\x5A\x00\x80\xE3\x64\x06\x00\x8A\x5A\x00\x80\xE1\x7C"
-     "\x06\x00"
-     "\x92\x5A\x00\x80\xEF\xBC\x01\x00\x87\xA0\xFC\x00\x86\xCC\xFC\x00\x80\x3D"
-     "\x00\x00"
-     "\x81\x25\x00\x00\x82\x21\x00\x00\x96\x5A\x00\x80\xB3\x8D\xFD\x00\x1E\x5A"
-     "\x00\x80"
-     "\xAE\x59\x00\x80\x9A\x5A\x00\x80\x8E\x5A\x00\x80\xB6\xBD\xFD\x00\xB5\x85"
-     "\xFD\x00"
-     "\x9E\x5A\x00\x80\xBB\xED\xFD\x00\xBA\xED\xFD\x00\xA2\x5A\x00\x80\xA6\x5A"
-     "\x00\x80"
-     "\xBF\xDD\xFD\x00\xBE\xDD\xFD\x00\xBD\xE1\xFD\x00\xBC\xE9\xFD\x00\xAA\x5A"
-     "\x00\x80"
-     "\xA3\xC9\xFD\x00\xAE\x5A\x00\x80\xB2\x5A\x00\x80\xA6\xF9\xFD\x00\xB6\x5A"
-     "\x00\x80"
-     "\xBA\x5A\x00\x80\xA5\xC1\xFD\x00\xAA\xA9\xFD\x00\xAB\xA9\xFD\x00\xBE\x5A"
-     "\x00\x80"
-     "\xC2\x5A\x00\x80\xAE\x99\xFD\x00\xAF\x99\xFD\x00\xAC\xAD\xFD\x00\xAD\xA5"
-     "\xFD\x00"
-     "\xA8\x59\xFE\x00\xA9\xA5\xFF\x00\xAA\xAD\xFF\x00\xAB\xA1\xFF\x00\xAC\xA1"
-     "\xFF\x00"
-     "\xAD\xA1\xFF\x00\xAE\xA1\xFF\x00\xAF\xA1\xFF\x00\xC6\x5A\x00\x80\xCA\x5A"
-     "\x00\x80"
-     "\xCE\x5A\x00\x80\xD2\x5A\x00\x80\xD6\x5A\x00\x80\xDA\x5A\x00\x80\xDE\x5A"
-     "\x00\x80"
-     "\xE2\x5A\x00\x80\xB8\x69\xFF\x00\xB9\x69\xFF\x00\xBA\x79\xFF\x00\xBB\x79"
-     "\xFF\x00"
-     "\xBC\x6D\xFF\x00\xBD\x15\xFF\x00\xBE\x1D\xFF\x00\xBF\x15\xFF\x00\xB0\x65"
-     "\xFF\x00"
-     "\xB1\x69\xFF\x00\xB2\x79\xFF\x00\xB3\x79\xFF\x00\xB4\x69\xFF\x00\xB5\x69"
-     "\xFF\x00"
-     "\xB6\x59\xFF\x00\xB7\x59\xFF\x00\xB3\x25\xFE\x00\xE6\x5A\x00\x80\xEA\x5A"
-     "\x00\x80"
-     "\xEE\x5A\x00\x80\xF2\x5A\x00\x80\xB6\x21\xFE\x00\xB5\x29\xFE\x00\xF6\x5A"
-     "\x00\x80"
-     "\xBB\xAD\xFE\x00\xBA\xA1\xFE\x00\xFA\x5A\x00\x80\x84\xE0\x01\x00\xBF\xAD"
-     "\xFE\x00"
-     "\xBE\xAD\xFE\x00\xBD\xAD\xFE\x00\xBC\xAD\xFE\x00\x82\x69\x00\x00\xA3\x61"
-     "\xFE\x00"
-     "\x80\x59\x00\x00\x81\x51\x00\x00\xA6\x65\xFE\x00\xBE\x9C\x01\x00\xFE\x5A"
-     "\x00\x80"
-     "\xA5\x6D\xFE\x00\xAA\xE5\xFE\x00\xAB\xE9\xFE\x00\x86\xC8\x00\x00\x87\xFC"
-     "\x01\x00"
-     "\xAE\xE9\xFE\x00\xAF\xE9\xFE\x00\xAC\xE9\xFE\x00\xAD\xE9\xFE\x00\xA8\x61"
-     "\xFE\x00"
-     "\xA9\x79\xFE\x00\xAA\x85\x01\x00\xAB\x91\x01\x00\xAC\xB9\x01\x00\xAD\xB9"
-     "\x01\x00"
-     "\xAE\xD9\x01\x00\xAF\xD5\x01\x00\x84\xA0\x00\x00\x02\x5B\x00\x80\x06\x5B"
-     "\x00\x80"
-     "\x0A\x5B\x00\x80\x0E\x5B\x00\x80\x12\x5B\x00\x80\x16\x5B\x00\x80\x1A\x5B"
-     "\x00\x80"
-     "\xB8\x6D\x01\x00\xB9\x75\x01\x00\xBA\x7D\x01\x00\xBB\x75\x01\x00\xBC\x6D"
-     "\x01\x00"
-     "\xBD\xD1\x00\x00\xBE\xD1\x00\x00\xBF\xD1\x00\x00\xB0\xAD\x01\x00\xB1\xB1"
-     "\x01\x00"
-     "\xB2\xB1\x01\x00\xB3\x85\x01\x00\xB4\x9D\x01\x00\xB5\x55\x01\x00\xB6\x5D"
-     "\x01\x00"
-     "\xB7\x55\x01\x00\xB3\x0D\x02\x00\x1E\x5B\x00\x80\x22\x5B\x00\x80\x26\x5B"
-     "\x00\x80"
-     "\x2A\x5B\x00\x80\xB6\x39\x02\x00\xB5\x31\x02\x00\x2E\x5B\x00\x80\xBB\x75"
-     "\x02\x00"
-     "\xBA\x69\x02\x00\x32\x5B\x00\x80\x36\x5B\x00\x80\xBF\x91\x02\x00\xBE\x59"
-     "\x02\x00"
-     "\xBD\x51\x02\x00\xBC\x65\x02\x00\x3A\x5B\x00\x80\xA3\x49\x02\x00\x3E\x5B"
-     "\x00\x80"
-     "\x42\x5B\x00\x80\xA6\x7D\x02\x00\x46\x5B\x00\x80\x4A\x5B\x00\x80\xA5\x75"
-     "\x02\x00"
-     "\xAA\x2D\x02\x00\xAB\x31\x02\x00\x4E\x5B\x00\x80\x52\x5B\x00\x80\xAE\x1D"
-     "\x02\x00"
-     "\xAF\xD5\x02\x00\xAC\x21\x02\x00\xAD\x15\x02\x00\x56\x5B\x00\x80\x5A\x5B"
-     "\x00\x80"
-     "\x5E\x5B\x00\x80\x62\x5B\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39"
-     "\x00\x00"
-     "\x66\x5B\x00\x80\x6A\x5B\x00\x80\xBE\x00\x04\x00\x72\x5B\x00\x80\x76\x5B"
-     "\x00\x80"
-     "\x7A\x5B\x00\x80\x7E\x5B\x00\x80\x82\x5B\x00\x80\x86\x5B\x00\x80\x86\xE0"
-     "\x04\x00"
-     "\x87\x04\x03\x00\xBE\x44\x05\x00\x84\x40\x04\x00\x8A\x5B\x00\x80\xE3\x64"
-     "\x01\x00"
-     "\x8E\x5B\x00\x80\xE1\xAC\x01\x00\x92\x5B\x00\x80\xEF\xB8\x02\x00\x96\x5B"
-     "\x00\x80"
-     "\x9A\x5B\x00\x80\x9E\x5B\x00\x80\xA2\x5B\x00\x80\xA6\x5B\x00\x80\xAA\x5B"
-     "\x00\x80"
-     "\xA8\xA5\x02\x00\xA9\xAD\x02\x00\xAA\xA5\x02\x00\xAB\xBD\x02\x00\xAC\xA5"
-     "\x02\x00"
-     "\xAD\xAD\x02\x00\xAE\xA5\x02\x00\xAF\x19\x01\x00\x84\xA0\x04\x00\xAE\x5B"
-     "\x00\x80"
-     "\xB2\x5B\x00\x80\xB6\x5B\x00\x80\xBA\x5B\x00\x80\xBE\x5B\x00\x80\xC2\x5B"
-     "\x00\x80"
-     "\xBE\xB0\x07\x00\xB8\x0D\x01\x00\xB9\x11\x01\x00\xBA\x11\x01\x00\xBB\x25"
-     "\x01\x00"
-     "\xBC\x3D\x01\x00\xBD\xD5\x01\x00\xBE\xDD\x01\x00\xBF\xD5\x01\x00\xB0\x69"
-     "\x01\x00"
-     "\xB1\x69\x01\x00\xB2\x79\x01\x00\xB3\x71\x01\x00\xB4\x59\x01\x00\xB5\x59"
-     "\x01\x00"
-     "\xB6\x39\x01\x00\xB7\x35\x01\x00\xE1\x70\xFF\x00\xE3\xB4\xFF\x00\xE3\x1C"
-     "\xFF\x00"
-     "\xE1\xFC\xFC\x00\xC6\x5B\x00\x80\xE1\x60\xFE\x00\xCA\x5B\x00\x80\xE3\xA0"
-     "\x00\x00"
-     "\xCE\x5B\x00\x80\xD2\x5B\x00\x80\xD6\x5B\x00\x80\xEF\x00\x00\x00\xBE\xFC"
-     "\x06\x00"
-     "\xEF\x98\xFF\x00\xEF\x68\xFE\x00\xDA\x5B\x00\x80\x80\x65\x00\x00\x81\x6D"
-     "\x00\x00"
-     "\x82\x75\x00\x00\xB3\x61\x02\x00\xE2\x5B\x00\x80\xB5\x61\x02\x00\xB6\x61"
-     "\x02\x00"
-     "\xE6\x5B\x00\x80\x86\xC0\x04\x00\x87\xA4\x04\x00\xBA\xE1\x01\x00\xBB\xE1"
-     "\x01\x00"
-     "\xBC\xE1\x01\x00\xBD\xE9\x01\x00\xBE\xD1\x01\x00\xBF\xD1\x01\x00\x6E\x5B"
-     "\x00\x80"
-     "\xDE\x5B\x00\x80\xAB\x05\x06\x00\xAA\x05\x06\x00\xAD\x0D\x06\x00\xAC\x05"
-     "\x06\x00"
-     "\xAF\x35\x06\x00\xAE\x35\x06\x00\xEA\x5B\x00\x80\xEE\x5B\x00\x80\xA3\x85"
-     "\x05\x00"
-     "\xF2\x5B\x00\x80\xA5\x85\x05\x00\xF6\x5B\x00\x80\xFA\x5B\x00\x80\xA6\x85"
-     "\x05\x00"
-     "\xB3\xF1\x06\x00\xFE\x5B\x00\x80\x02\x5C\x00\x80\x06\x5C\x00\x80\x0A\x5C"
-     "\x00\x80"
-     "\xB6\xE9\x06\x00\xB5\xE1\x06\x00\x0E\x5C\x00\x80\xBB\x89\x06\x00\xBA\x85"
-     "\x06\x00"
-     "\x12\x5C\x00\x80\x16\x5C\x00\x80\xBF\x89\x06\x00\xBE\x81\x06\x00\xBD\x89"
-     "\x06\x00"
-     "\xBC\x91\x06\x00\xA8\x29\x06\x00\xA9\x29\x06\x00\xAA\x39\x06\x00\xAB\x39"
-     "\x06\x00"
-     "\xAC\x29\x06\x00\xAD\x29\x06\x00\xAE\x49\x06\x00\xAF\x45\x06\x00\x1A\x5C"
-     "\x00\x80"
-     "\x1E\x5C\x00\x80\x22\x5C\x00\x80\x26\x5C\x00\x80\x2A\x5C\x00\x80\x2E\x5C"
-     "\x00\x80"
-     "\x32\x5C\x00\x80\x36\x5C\x00\x80\xB8\xE9\x07\x00\xB9\xE9\x07\x00\xBA\xF9"
-     "\x07\x00"
-     "\xBB\xF9\x07\x00\xBC\xE9\x07\x00\xBD\x45\x07\x00\xBE\x41\x07\x00\xBF\x7D"
-     "\x07\x00"
-     "\xB0\x01\x06\x00\xB1\x0D\x06\x00\xB2\x05\x06\x00\xB3\x1D\x06\x00\xB4\x05"
-     "\x06\x00"
-     "\xB5\x0D\x06\x00\xB6\x05\x06\x00\xB7\xD9\x07\x00\xA3\xB1\x07\x00\x3A\x5C"
-     "\x00\x80"
-     "\x3E\x5C\x00\x80\x42\x5C\x00\x80\x46\x5C\x00\x80\xA6\xA9\x07\x00\xA5\xA1"
-     "\x07\x00"
-     "\x4A\x5C\x00\x80\xAB\xC9\x07\x00\xAA\xC5\x07\x00\x4E\x5C\x00\x80\x52\x5C"
-     "\x00\x80"
-     "\xAF\xC9\x07\x00\xAE\xC1\x07\x00\xAD\xC9\x07\x00\xAC\xD1\x07\x00\x80\x4D"
-     "\x00\x00"
-     "\x81\x55\x00\x00\x82\x55\x00\x00\xB3\x6D\x01\x00\x56\x5C\x00\x80\xB5\x79"
-     "\x01\x00"
-     "\xB6\x6D\x01\x00\x5A\x5C\x00\x80\x86\xC0\x00\x00\x87\x04\x01\x00\xBA\x21"
-     "\x01\x00"
-     "\xBB\x21\x01\x00\xBC\x21\x01\x00\xBD\x29\x01\x00\xBE\x11\x01\x00\xBF\x11"
-     "\x01\x00"
-     "\xA8\x51\x06\x00\xA9\x59\x06\x00\xAA\xA9\x01\x00\xAB\xA9\x01\x00\xAC\xBD"
-     "\x01\x00"
-     "\xAD\xA1\x01\x00\xAE\xA1\x01\x00\xAF\xD9\x01\x00\x84\xAC\x01\x00\x5E\x5C"
-     "\x00\x80"
-     "\x62\x5C\x00\x80\x66\x5C\x00\x80\x6A\x5C\x00\x80\x6E\x5C\x00\x80\x72\x5C"
-     "\x00\x80"
-     "\x76\x5C\x00\x80\xB8\x69\x01\x00\xB9\x69\x01\x00\xBA\x79\x01\x00\xBB\x79"
-     "\x01\x00"
-     "\xBC\x69\x01\x00\xBD\x69\x01\x00\xBE\xDD\x00\x00\xBF\xD5\x00\x00\xB0\xA9"
-     "\x01\x00"
-     "\xB1\xA9\x01\x00\xB2\xB9\x01\x00\xB3\xB1\x01\x00\xB4\x99\x01\x00\xB5\x99"
-     "\x01\x00"
-     "\xB6\x59\x01\x00\xB7\x59\x01\x00\xA3\x21\x02\x00\x7A\x5C\x00\x80\x7E\x5C"
-     "\x00\x80"
-     "\x82\x5C\x00\x80\x86\x5C\x00\x80\xA6\x21\x02\x00\xA5\x35\x02\x00\x8A\x5C"
-     "\x00\x80"
-     "\xAB\x6D\x02\x00\xAA\x6D\x02\x00\x8E\x5C\x00\x80\x92\x5C\x00\x80\xAF\x5D"
-     "\x02\x00"
-     "\xAE\x5D\x02\x00\xAD\x65\x02\x00\xAC\x6D\x02\x00\x96\x5C\x00\x80\x9A\x5C"
-     "\x00\x80"
-     "\x9E\x5C\x00\x80\xA2\x5C\x00\x80\xA6\x5C\x00\x80\xAA\x5C\x00\x80\xAE\x5C"
-     "\x00\x80"
-     "\xB2\x5C\x00\x80\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\xB6\x5C"
-     "\x00\x80"
-     "\xBA\x5C\x00\x80\xC2\x5C\x00\x80\x87\x44\x03\x00\x86\xDC\x0C\x00\xC6\x5C"
-     "\x00\x80"
-     "\xCA\x5C\x00\x80\xCE\x5C\x00\x80\x84\xC0\x0C\x00\xD2\x5C\x00\x80\xEF\xDC"
-     "\x07\x00"
-     "\xD6\x5C\x00\x80\xDA\x5C\x00\x80\xDE\x5C\x00\x80\xE3\x14\x01\x00\xE2\x5C"
-     "\x00\x80"
-     "\xE1\xD4\x01\x00\xE6\x5C\x00\x80\xEA\x5C\x00\x80\xEE\x5C\x00\x80\xF2\x5C"
-     "\x00\x80"
-     "\xB3\x9D\x03\x00\xF6\x5C\x00\x80\xFA\x5C\x00\x80\xFE\x5C\x00\x80\x02\x5D"
-     "\x00\x80"
-     "\xB6\x79\x03\x00\xB5\x71\x03\x00\x06\x5D\x00\x80\xBB\x5D\x03\x00\xBA\x5D"
-     "\x03\x00"
-     "\x0A\x5D\x00\x80\x0E\x5D\x00\x80\xBF\xE1\x00\x00\xBE\xF9\x00\x00\xBD\xF1"
-     "\x00\x00"
-     "\xBC\xF9\x00\x00\xA8\x2D\x02\x00\xA9\x55\x02\x00\xAA\x55\x02\x00\xAB\xA9"
-     "\x02\x00"
-     "\xAC\xB9\x02\x00\xAD\xB9\x02\x00\xAE\xA9\x02\x00\xAF\xA9\x02\x00\xBE\x68"
-     "\x0D\x00"
-     "\x12\x5D\x00\x80\x16\x5D\x00\x80\x1A\x5D\x00\x80\x1E\x5D\x00\x80\x82\xAD"
-     "\x00\x00"
-     "\x81\xAD\x00\x00\x80\xAD\x00\x00\xB8\x6D\x01\x00\xB9\x05\x01\x00\xBA\x0D"
-     "\x01\x00"
-     "\xBB\x05\x01\x00\xBC\x1D\x01\x00\xBD\x05\x01\x00\xBE\x0D\x01\x00\xBF\x05"
-     "\x01\x00"
-     "\xB0\xD9\x02\x00\xB1\xD9\x02\x00\xB2\x6D\x01\x00\xB3\x65\x01\x00\xB4\x7D"
-     "\x01\x00"
-     "\xB5\x65\x01\x00\xB6\x65\x01\x00\xB7\x55\x01\x00\xE1\x78\x06\x00\xE1\x5C"
-     "\x07\x00"
-     "\xE3\xB8\x06\x00\xE3\x50\x07\x00\x22\x5D\x00\x80\x26\x5D\x00\x80\x2A\x5D"
-     "\x00\x80"
-     "\x32\x5D\x00\x80\x36\x5D\x00\x80\x3A\x5D\x00\x80\x3E\x5D\x00\x80\x42\x5D"
-     "\x00\x80"
-     "\xBE\xF8\x0C\x00\x46\x5D\x00\x80\xEF\x7C\x06\x00\xEF\x8C\x07\x00\xA3\x9D"
-     "\x02\x00"
-     "\x4A\x5D\x00\x80\x86\x68\x0C\x00\x87\x04\x0D\x00\x4E\x5D\x00\x80\xA6\x79"
-     "\x02\x00"
-     "\xA5\x71\x02\x00\x52\x5D\x00\x80\xAB\x5D\x02\x00\xAA\x5D\x02\x00\x56\x5D"
-     "\x00\x80"
-     "\x5A\x5D\x00\x80\xAF\xE1\x01\x00\xAE\xF9\x01\x00\xAD\xF1\x01\x00\xAC\xF9"
-     "\x01\x00"
-     "\xA8\x75\x0E\x00\xA9\x85\x0E\x00\xAA\x8D\x0E\x00\xAB\x85\x0E\x00\xAC\x81"
-     "\x0E\x00"
-     "\xAD\x8D\x0E\x00\xAE\x85\x0E\x00\xAF\xBD\x0E\x00\x2E\x5D\x00\x80\x5E\x5D"
-     "\x00\x80"
-     "\x62\x5D\x00\x80\xBE\x5C\x00\x80\x66\x5D\x00\x80\x6A\x5D\x00\x80\x6E\x5D"
-     "\x00\x80"
-     "\x72\x5D\x00\x80\xB8\x91\x0E\x00\xB9\x99\x0E\x00\xBA\xA1\x0E\x00\xBB\xA1"
-     "\x0E\x00"
-     "\xBC\x75\x0F\x00\xBD\x7D\x0F\x00\xBE\x75\x0F\x00\xBF\x6D\x0F\x00\xB0\xC9"
-     "\x0E\x00"
-     "\xB1\xD5\x0E\x00\xB2\xDD\x0E\x00\xB3\xD5\x0E\x00\xB4\xCD\x0E\x00\xB5\xB1"
-     "\x0E\x00"
-     "\xB6\xB1\x0E\x00\xB7\xB1\x0E\x00\xB3\x05\x0E\x00\x76\x5D\x00\x80\x7A\x5D"
-     "\x00\x80"
-     "\x7E\x5D\x00\x80\x82\x5D\x00\x80\xB6\x0D\x0E\x00\xB5\x09\x0E\x00\x86\x5D"
-     "\x00\x80"
-     "\xBB\x11\x0E\x00\xBA\x0D\x0E\x00\x8A\x5D\x00\x80\x8E\x5D\x00\x80\xBF\x75"
-     "\x0E\x00"
-     "\xBE\x09\x0E\x00\xBD\x01\x0E\x00\xBC\x09\x0E\x00\x82\x29\x00\x00\xA3\x41"
-     "\x0E\x00"
-     "\x80\x1D\x00\x00\x81\x15\x00\x00\xA6\x49\x0E\x00\x92\x5D\x00\x80\x96\x5D"
-     "\x00\x80"
-     "\xA5\x4D\x0E\x00\xAA\x49\x0E\x00\xAB\x55\x0E\x00\x9A\x5D\x00\x80\x84\xA0"
-     "\x03\x00"
-     "\xAE\x4D\x0E\x00\xAF\x31\x0E\x00\xAC\x4D\x0E\x00\xAD\x45\x0E\x00\xA8\x9D"
-     "\x0E\x00"
-     "\xA9\xA5\x0E\x00\xAA\xAD\x0E\x00\xAB\xA5\x0E\x00\xAC\xBD\x0E\x00\xAD\xA1"
-     "\x0E\x00"
-     "\xAE\xDD\x0E\x00\xAF\xCD\x0E\x00\x9E\x5D\x00\x80\xA2\x5D\x00\x80\x86\x88"
-     "\x01\x00"
-     "\x87\xAC\x01\x00\xA6\x5D\x00\x80\xAA\x5D\x00\x80\xAE\x5D\x00\x80\xB2\x5D"
-     "\x00\x80"
-     "\xB8\x7D\x01\x00\xB9\xC1\x01\x00\xBA\xC1\x01\x00\xBB\xC1\x01\x00\xBC\xC1"
-     "\x01\x00"
-     "\xBD\xC9\x01\x00\xBE\xF1\x01\x00\xBF\xF1\x01\x00\xB0\xBD\x0E\x00\xB1\x8D"
-     "\x0E\x00"
-     "\xB2\x85\x0E\x00\xB3\x49\x01\x00\xB4\x59\x01\x00\xB5\x45\x01\x00\xB6\x4D"
-     "\x01\x00"
-     "\xB7\x45\x01\x00\xB3\x25\x0E\x00\xB6\x5D\x00\x80\xBA\x5D\x00\x80\xBE\x5D"
-     "\x00\x80"
-     "\xC2\x5D\x00\x80\xB6\x21\x0E\x00\xB5\x35\x0E\x00\xC6\x5D\x00\x80\xBB\x49"
-     "\x0E\x00"
-     "\xBA\x79\x0E\x00\xCA\x5D\x00\x80\xCE\x5D\x00\x80\xBF\x81\x01\x00\xBE\x9D"
-     "\x01\x00"
-     "\xBD\x9D\x01\x00\xBC\x51\x0E\x00\xD2\x5D\x00\x80\xA3\x61\x0E\x00\xD6\x5D"
-     "\x00\x80"
-     "\xDA\x5D\x00\x80\xA6\x65\x0E\x00\xDE\x5D\x00\x80\xE2\x5D\x00\x80\xA5\x71"
-     "\x0E\x00"
-     "\xAA\x3D\x0E\x00\xAB\x0D\x0E\x00\xE6\x5D\x00\x80\xEA\x5D\x00\x80\xAE\xD9"
-     "\x01\x00"
-     "\xAF\xC5\x01\x00\xAC\x15\x0E\x00\xAD\xD9\x01\x00\xA8\x25\x03\x00\xA9\x35"
-     "\x03\x00"
-     "\xAA\x3D\x03\x00\xAB\x35\x03\x00\xAC\x2D\x03\x00\xAD\x91\x03\x00\xAE\x91"
-     "\x03\x00"
-     "\xAF\x91\x03\x00\xEE\x5D\x00\x80\xF2\x5D\x00\x80\xF6\x5D\x00\x80\xFA\x5D"
-     "\x00\x80"
-     "\xFE\x5D\x00\x80\x02\x5E\x00\x80\x06\x5E\x00\x80\x0A\x5E\x00\x80\xB8\x99"
-     "\x03\x00"
-     "\xB9\xAD\x03\x00\xBA\xA5\x03\x00\xBB\x6D\x00\x00\xBC\x75\x00\x00\xBD\x7D"
-     "\x00\x00"
-     "\xBE\x75\x00\x00\xBF\x6D\x00\x00\xB0\xF1\x03\x00\xB1\xF9\x03\x00\xB2\xC1"
-     "\x03\x00"
-     "\xB3\xC1\x03\x00\xB4\xB1\x03\x00\xB5\xBD\x03\x00\xB6\xB5\x03\x00\xB7\xA9"
-     "\x03\x00"
-     "\x80\xB5\x00\x00\x81\xBD\x00\x00\x82\xB5\x00\x00\xBE\x0C\x03\x00\x12\x5E"
-     "\x00\x80"
-     "\x16\x5E\x00\x80\x1A\x5E\x00\x80\xEF\xE8\x02\x00\x84\xAC\x02\x00\xE1\x94"
-     "\x01\x00"
-     "\x1E\x5E\x00\x80\xE3\x1C\x02\x00\x22\x5E\x00\x80\x26\x5E\x00\x80\x87\xD8"
-     "\x03\x00"
-     "\x86\x1C\x04\x00\xB3\xF5\x03\x00\x2A\x5E\x00\x80\x85\xA4\x1F\x00\x2E\x5E"
-     "\x00\x80"
-     "\x32\x5E\x00\x80\xB6\x51\x03\x00\xB5\xE5\x03\x00\x36\x5E\x00\x80\xBB\x7D"
-     "\x03\x00"
-     "\xBA\x75\x03\x00\x3A\x5E\x00\x80\x3E\x5E\x00\x80\xBF\x39\x03\x00\xBE\x35"
-     "\x03\x00"
-     "\xBD\x55\x03\x00\xBC\x55\x03\x00\x42\x5E\x00\x80\x46\x5E\x00\x80\x4A\x5E"
-     "\x00\x80"
-     "\xA3\x39\x02\x00\x4E\x5E\x00\x80\xA5\x29\x02\x00\xA6\x9D\x02\x00\x52\x5E"
-     "\x00\x80"
-     "\x56\x5E\x00\x80\x5A\x5E\x00\x80\xAA\xB9\x02\x00\xAB\xB1\x02\x00\xAC\x99"
-     "\x02\x00"
-     "\xAD\x99\x02\x00\xAE\xF9\x02\x00\xAF\xF5\x02\x00\x5E\x5E\x00\x80\x62\x5E"
-     "\x00\x80"
-     "\x84\x44\x02\x00\xBE\x40\x05\x00\x66\x5E\x00\x80\x6A\x5E\x00\x80\x6E\x5E"
-     "\x00\x80"
-     "\x72\x5E\x00\x80\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x76\x5E"
-     "\x00\x80"
-     "\x7A\x5E\x00\x80\x82\x5E\x00\x80\x87\xF4\x05\x00\x86\xDC\x04\x00\x86\x5E"
-     "\x00\x80"
-     "\x8A\x5E\x00\x80\x8E\x5E\x00\x80\x92\x5E\x00\x80\xEF\x48\x0E\x00\x96\x5E"
-     "\x00\x80"
-     "\x9A\x5E\x00\x80\xEF\xD4\x0E\x00\xE3\x3C\x0F\x00\xE1\x74\x0E\x00\xE1\xCC"
-     "\x0F\x00"
-     "\xE3\x50\x01\x00\x9E\x5E\x00\x80\xA2\x5E\x00\x80\xA6\x5E\x00\x80\xAA\x5E"
-     "\x00\x80"
-     "\xB3\x59\x01\x00\xAE\x5E\x00\x80\xB2\x5E\x00\x80\xB6\x5E\x00\x80\xBA\x5E"
-     "\x00\x80"
-     "\xB6\x61\x01\x00\xB5\x69\x01\x00\xBE\x5E\x00\x80\xBB\x3D\x01\x00\xBA\x35"
-     "\x01\x00"
-     "\xC2\x5E\x00\x80\xC6\x5E\x00\x80\xBF\xF5\x00\x00\xBE\xF5\x00\x00\xBD\x15"
-     "\x01\x00"
-     "\xBC\x15\x01\x00\xA8\x91\x06\x00\xA9\x91\x06\x00\xAA\x91\x06\x00\xAB\xA5"
-     "\x06\x00"
-     "\xAC\xA1\x06\x00\xAD\xA1\x06\x00\xAE\xDD\x06\x00\xAF\xD1\x06\x00\x7E\x5E"
-     "\x00\x80"
-     "\xCA\x5E\x00\x80\xCE\x5E\x00\x80\xD2\x5E\x00\x80\x0E\x5E\x00\x80\xD6\x5E"
-     "\x00\x80"
-     "\xDA\x5E\x00\x80\xDE\x5E\x00\x80\xB8\x5D\x07\x00\xB9\x65\x07\x00\xBA\x6D"
-     "\x07\x00"
-     "\xBB\x0D\x07\x00\xBC\x15\x07\x00\xBD\x05\x07\x00\xBE\x0D\x07\x00\xBF\x05"
-     "\x07\x00"
-     "\xB0\xB1\x06\x00\xB1\xB9\x06\x00\xB2\x81\x06\x00\xB3\x81\x06\x00\xB4\x71"
-     "\x07\x00"
-     "\xB5\x71\x07\x00\xB6\x71\x07\x00\xB7\x6D\x07\x00\xA3\x19\x06\x00\x82\x2D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x80\x1D\x00\x00\xE2\x5E\x00\x80\xA6\x21\x06\x00\xA5\x29"
-     "\x06\x00"
-     "\xE6\x5E\x00\x80\xAB\x7D\x06\x00\xAA\x75\x06\x00\xEA\x5E\x00\x80\x84\xA4"
-     "\x02\x00"
-     "\xAF\xB5\x07\x00\xAE\xB5\x07\x00\xAD\x55\x06\x00\xAC\x55\x06\x00\xBE\xCC"
-     "\x03\x00"
-     "\xB3\x41\x07\x00\xEE\x5E\x00\x80\xF2\x5E\x00\x80\xB6\x45\x07\x00\xF6\x5E"
-     "\x00\x80"
-     "\xFA\x5E\x00\x80\xB5\x51\x07\x00\xBA\x55\x07\x00\xBB\x5D\x07\x00\x86\x48"
-     "\x00\x00"
-     "\x87\x0C\x03\x00\xBE\x39\x07\x00\xBF\x39\x07\x00\xBC\x45\x07\x00\xBD\x39"
-     "\x07\x00"
-     "\xA8\x75\x06\x00\xA9\x85\x06\x00\xAA\x8D\x06\x00\xAB\x85\x06\x00\xAC\x9D"
-     "\x06\x00"
-     "\xAD\x81\x06\x00\xAE\x81\x06\x00\xAF\xAD\x06\x00\xFE\x5E\x00\x80\x02\x5F"
-     "\x00\x80"
-     "\x06\x5F\x00\x80\x0A\x5F\x00\x80\x0E\x5F\x00\x80\x12\x5F\x00\x80\x16\x5F"
-     "\x00\x80"
-     "\x1A\x5F\x00\x80\xB8\xA9\x06\x00\xB9\xA9\x06\x00\xBA\x6D\x01\x00\xBB\x65"
-     "\x01\x00"
-     "\xBC\x7D\x01\x00\xBD\x61\x01\x00\xBE\x61\x01\x00\xBF\x61\x01\x00\xB0\xD5"
-     "\x06\x00"
-     "\xB1\xDD\x06\x00\xB2\xD5\x06\x00\xB3\xA9\x06\x00\xB4\xB9\x06\x00\xB5\xB9"
-     "\x06\x00"
-     "\xB6\xA9\x06\x00\xB7\xA1\x06\x00\xA3\x05\x06\x00\x1E\x5F\x00\x80\x22\x5F"
-     "\x00\x80"
-     "\x26\x5F\x00\x80\x2A\x5F\x00\x80\xA6\x01\x06\x00\xA5\x15\x06\x00\x2E\x5F"
-     "\x00\x80"
-     "\xAB\x19\x06\x00\xAA\x11\x06\x00\x32\x5F\x00\x80\x36\x5F\x00\x80\xAF\x7D"
-     "\x06\x00"
-     "\xAE\x7D\x06\x00\xAD\x7D\x06\x00\xAC\x01\x06\x00\x3A\x5F\x00\x80\x3E\x5F"
-     "\x00\x80"
-     "\x42\x5F\x00\x80\x46\x5F\x00\x80\x4A\x5F\x00\x80\x4E\x5F\x00\x80\x52\x5F"
-     "\x00\x80"
-     "\x56\x5F\x00\x80\x80\x39\x00\x00\x81\x39\x00\x00\x82\x05\x00\x00\x5A\x5F"
-     "\x00\x80"
-     "\xBE\xB8\x01\x00\x84\xB8\x01\x00\x62\x5F\x00\x80\x66\x5F\x00\x80\xA8\xCD"
-     "\x02\x00"
-     "\xA9\xD1\x02\x00\xAA\xD1\x02\x00\xAB\x2D\x03\x00\xAC\x35\x03\x00\xAD\x3D"
-     "\x03\x00"
-     "\xAE\x35\x03\x00\xAF\x2D\x03\x00\x6A\x5F\x00\x80\x6E\x5F\x00\x80\x72\x5F"
-     "\x00\x80"
-     "\x76\x5F\x00\x80\x7A\x5F\x00\x80\x7E\x5F\x00\x80\xBE\x34\x1C\x00\x82\x5F"
-     "\x00\x80"
-     "\xB8\xE9\x03\x00\xB9\xE9\x03\x00\xBA\x89\x03\x00\xBB\x89\x03\x00\xBC\x99"
-     "\x03\x00"
-     "\xBD\x99\x03\x00\xBE\x89\x03\x00\xBF\x89\x03\x00\xB0\x55\x03\x00\xB1\x5D"
-     "\x03\x00"
-     "\xB2\x55\x03\x00\xB3\xE9\x03\x00\xB4\xF9\x03\x00\xB5\xF9\x03\x00\xB6\xE9"
-     "\x03\x00"
-     "\xB7\xE1\x03\x00\xB3\x59\x02\x00\x86\x5F\x00\x80\x87\x94\x03\x00\x86\x4C"
-     "\x1C\x00"
-     "\x8A\x5F\x00\x80\xB6\xBD\x03\x00\xB5\xBD\x03\x00\x8E\x5F\x00\x80\xBB\x99"
-     "\x03\x00"
-     "\xBA\x91\x03\x00\x92\x5F\x00\x80\x96\x5F\x00\x80\xBF\x3D\x03\x00\xBE\x3D"
-     "\x03\x00"
-     "\xBD\x3D\x03\x00\xBC\x81\x03\x00\x9A\x5F\x00\x80\xA3\x1D\x02\x00\x9E\x5F"
-     "\x00\x80"
-     "\xA2\x5F\x00\x80\xA6\xF9\x03\x00\xA6\x5F\x00\x80\xAA\x5F\x00\x80\xA5\xF9"
-     "\x03\x00"
-     "\xAA\xD5\x03\x00\xAB\xDD\x03\x00\xAE\x5F\x00\x80\xB2\x5F\x00\x80\xAE\x79"
-     "\x03\x00"
-     "\xAF\x79\x03\x00\xAC\xC5\x03\x00\xAD\x79\x03\x00\xB6\x5F\x00\x80\xBA\x5F"
-     "\x00\x80"
-     "\x84\x04\x02\x00\xBE\x00\x1C\x00\xBE\x5F\x00\x80\x82\x3D\x00\x00\x81\x3D"
-     "\x00\x00"
-     "\x80\x3D\x00\x00\xC2\x5F\x00\x80\xC6\x5F\x00\x80\xCA\x5F\x00\x80\xD2\x5F"
-     "\x00\x80"
-     "\xD6\x5F\x00\x80\xDA\x5F\x00\x80\xDE\x5F\x00\x80\xE2\x5F\x00\x80\xE6\x5F"
-     "\x00\x80"
-     "\xE3\xA0\x00\x00\xEA\x5F\x00\x80\xE1\xBC\x01\x00\x86\xF0\x1C\x00\x87\x04"
-     "\x1D\x00"
-     "\xEE\x5F\x00\x80\xF2\x5F\x00\x80\xF6\x5F\x00\x80\xFA\x5F\x00\x80\xFE\x5F"
-     "\x00\x80"
-     "\x02\x60\x00\x80\x06\x60\x00\x80\xEF\x7C\x00\x00\x0A\x60\x00\x80\x0E\x60"
-     "\x00\x80"
-     "\x12\x60\x00\x80\x16\x60\x00\x80\x1A\x60\x00\x80\xEF\x38\x06\x00\x1E\x60"
-     "\x00\x80"
-     "\x22\x60\x00\x80\x26\x60\x00\x80\xEF\xA4\x01\x00\x84\xA8\x1D\x00\xE1\xC0"
-     "\x06\x00"
-     "\x2A\x60\x00\x80\xE3\x10\x01\x00\x2E\x60\x00\x80\xE1\x10\x07\x00\x32\x60"
-     "\x00\x80"
-     "\xE3\xE4\x06\x00\xB3\xE1\x01\x00\x36\x60\x00\x80\x3A\x60\x00\x80\x42\x60"
-     "\x00\x80"
-     "\x46\x60\x00\x80\xB6\xF9\x01\x00\xB5\xF1\x01\x00\x4A\x60\x00\x80\xBB\x65"
-     "\x01\x00"
-     "\xBA\xD9\x01\x00\x86\x48\x1F\x00\x87\xD0\x1C\x00\xBF\x5D\x01\x00\xBE\x5D"
-     "\x01\x00"
-     "\xBD\x65\x01\x00\xBC\x6D\x01\x00\xB0\x3D\x1E\x00\xB1\x81\x1E\x00\xB2\x99"
-     "\x1E\x00"
-     "\xB3\x91\x1E\x00\xB4\xB9\x1E\x00\xB5\xB9\x1E\x00\xB6\xD9\x1E\x00\xB7\xD5"
-     "\x1E\x00"
-     "\xB8\xED\x1E\x00\xB9\xF1\x1E\x00\xBA\xF1\x1E\x00\xBB\xC5\x1E\x00\xBC\xDD"
-     "\x1E\x00"
-     "\xBD\x35\x1F\x00\xBE\x3D\x1F\x00\xBF\x2D\x1F\x00\xCE\x5F\x00\x80\x5E\x5F"
-     "\x00\x80"
-     "\x4E\x60\x00\x80\x52\x60\x00\x80\x80\xE9\x1F\x00\x81\x91\x1F\x00\x82\x91"
-     "\x1F\x00"
-     "\x3E\x60\x00\x80\xA8\x99\x1D\x00\xA9\x99\x1D\x00\xAA\x49\x1E\x00\xAB\x49"
-     "\x1E\x00"
-     "\xAC\x59\x1E\x00\xAD\x59\x1E\x00\xAE\x49\x1E\x00\xAF\x49\x1E\x00\xA3\xA1"
-     "\x1E\x00"
-     "\x56\x60\x00\x80\x5A\x60\x00\x80\x5E\x60\x00\x80\x62\x60\x00\x80\xA6\xB9"
-     "\x1E\x00"
-     "\xA5\xB1\x1E\x00\x66\x60\x00\x80\xAB\x25\x1E\x00\xAA\x99\x1E\x00\x6A\x60"
-     "\x00\x80"
-     "\x6E\x60\x00\x80\xAF\x1D\x1E\x00\xAE\x1D\x1E\x00\xAD\x25\x1E\x00\xAC\x2D"
-     "\x1E\x00"
-     "\x72\x60\x00\x80\xB3\xF1\x1F\x00\x76\x60\x00\x80\x7A\x60\x00\x80\xB6\x91"
-     "\x1F\x00"
-     "\x7E\x60\x00\x80\x82\x60\x00\x80\xB5\xE1\x1F\x00\xBA\xBD\x1F\x00\xBB\x85"
-     "\x1F\x00"
-     "\x86\x60\x00\x80\x8A\x60\x00\x80\xBE\x9D\x1F\x00\xBF\x85\x1F\x00\xBC\x9D"
-     "\x1F\x00"
-     "\xBD\x95\x1F\x00\xA8\x2D\x1E\x00\xA9\x35\x1E\x00\xAA\x3D\x1E\x00\xAB\x35"
-     "\x1E\x00"
-     "\xAC\x2D\x1E\x00\xAD\x55\x1E\x00\xAE\x51\x1E\x00\xAF\x45\x1E\x00\x8E\x60"
-     "\x00\x80"
-     "\x92\x60\x00\x80\x96\x60\x00\x80\x9A\x60\x00\x80\x9E\x60\x00\x80\x82\xB9"
-     "\x00\x00"
-     "\x81\xB1\x00\x00\x80\xB9\x00\x00\xB8\xC1\x1E\x00\xB9\xC1\x1E\x00\xBA\xC1"
-     "\x1E\x00"
-     "\xBB\xDD\x1E\x00\xBC\xC5\x1E\x00\xBD\xCD\x1E\x00\xBE\xF1\x1E\x00\xBF\xED"
-     "\x1E\x00"
-     "\xB0\x3D\x1E\x00\xB1\x05\x1E\x00\xB2\x0D\x1E\x00\xB3\x05\x1E\x00\xB4\x1D"
-     "\x1E\x00"
-     "\xB5\x05\x1E\x00\xB6\x01\x1E\x00\xB7\x01\x1E\x00\xA3\xB5\x1E\x00\xA2\x60"
-     "\x00\x80"
-     "\xA6\x60\x00\x80\xAA\x60\x00\x80\x84\x98\x02\x00\xA6\xD5\x1E\x00\xA5\xA5"
-     "\x1E\x00"
-     "\xBE\x90\x01\x00\xAB\xC1\x1E\x00\xAA\xF9\x1E\x00\x86\x08\x00\x00\x87\x7C"
-     "\x01\x00"
-     "\xAF\xC1\x1E\x00\xAE\xD9\x1E\x00\xAD\xD1\x1E\x00\xAC\xD9\x1E\x00\xB2\x60"
-     "\x00\x80"
-     "\xB3\x19\x1E\x00\xB6\x60\x00\x80\xBA\x60\x00\x80\xB6\x55\x01\x00\xBE\x60"
-     "\x00\x80"
-     "\xC2\x60\x00\x80\xB5\x91\x01\x00\xBA\x71\x01\x00\xBB\x79\x01\x00\xC6\x60"
-     "\x00\x80"
-     "\xCA\x60\x00\x80\xBE\x31\x01\x00\xBF\x3D\x01\x00\xBC\x51\x01\x00\xBD\x51"
-     "\x01\x00"
-     "\xA8\xCD\x1D\x00\xA9\xD5\x1D\x00\xAA\xD5\x1D\x00\xAB\x45\x02\x00\xAC\x5D"
-     "\x02\x00"
-     "\xAD\x91\x02\x00\xAE\x91\x02\x00\xAF\x91\x02\x00\xCE\x60\x00\x80\xD2\x60"
-     "\x00\x80"
-     "\xD6\x60\x00\x80\xDA\x60\x00\x80\xDE\x60\x00\x80\xE2\x60\x00\x80\xE6\x60"
-     "\x00\x80"
-     "\xEA\x60\x00\x80\xB8\x99\x02\x00\xB9\xAD\x02\x00\xBA\xA5\x02\x00\xBB\x6D"
-     "\x03\x00"
-     "\xBC\x75\x03\x00\xBD\x7D\x03\x00\xBE\x75\x03\x00\xBF\x6D\x03\x00\xB0\xF1"
-     "\x02\x00"
-     "\xB1\xF9\x02\x00\xB2\xC1\x02\x00\xB3\xC1\x02\x00\xB4\xB1\x02\x00\xB5\xBD"
-     "\x02\x00"
-     "\xB6\xB5\x02\x00\xB7\xA9\x02\x00\xA3\x5D\x1D\x00\xEE\x60\x00\x80\xF2\x60"
-     "\x00\x80"
-     "\xF6\x60\x00\x80\xFA\x60\x00\x80\xA6\x11\x02\x00\xA5\xD5\x02\x00\xFE\x60"
-     "\x00\x80"
-     "\xAB\x3D\x02\x00\xAA\x35\x02\x00\x02\x61\x00\x80\x06\x61\x00\x80\xAF\x79"
-     "\x02\x00"
-     "\xAE\x75\x02\x00\xAD\x15\x02\x00\xAC\x15\x02\x00\xBE\x4C\x04\x00\x0A\x61"
-     "\x00\x80"
-     "\x0E\x61\x00\x80\x12\x61\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39"
-     "\x00\x00"
-     "\x16\x61\x00\x80\x1A\x61\x00\x80\x84\xA8\x03\x00\x86\xC8\x04\x00\x87\xA8"
-     "\x04\x00"
-     "\x22\x61\x00\x80\x26\x61\x00\x80\x2A\x61\x00\x80\x2E\x61\x00\x80\xA8\xC1"
-     "\x02\x00"
-     "\xA9\xCD\x02\x00\xAA\xC5\x02\x00\xAB\xD9\x02\x00\xAC\xC9\x02\x00\xAD\xFD"
-     "\x02\x00"
-     "\xAE\xF5\x02\x00\xAF\x3D\x01\x00\x32\x61\x00\x80\x36\x61\x00\x80\x3A\x61"
-     "\x00\x80"
-     "\x3E\x61\x00\x80\x42\x61\x00\x80\x46\x61\x00\x80\x4A\x61\x00\x80\xBE\xF4"
-     "\x04\x00"
-     "\xB8\xC9\x01\x00\xB9\xC9\x01\x00\xBA\xD9\x01\x00\xBB\xD1\x01\x00\xBC\xF9"
-     "\x01\x00"
-     "\xBD\xF9\x01\x00\xBE\x9D\x01\x00\xBF\x95\x01\x00\xB0\x45\x01\x00\xB1\x4D"
-     "\x01\x00"
-     "\xB2\x45\x01\x00\xB3\x5D\x01\x00\xB4\x45\x01\x00\xB5\x4D\x01\x00\xB6\x45"
-     "\x01\x00"
-     "\xB7\xF9\x01\x00\xB3\x4D\x02\x00\x4E\x61\x00\x80\x52\x61\x00\x80\x56\x61"
-     "\x00\x80"
-     "\x5A\x61\x00\x80\xB6\x7D\x02\x00\xB5\x41\x02\x00\x5E\x61\x00\x80\xBB\x89"
-     "\x01\x00"
-     "\xBA\x81\x01\x00\x62\x61\x00\x80\x66\x61\x00\x80\xBF\x89\x01\x00\xBE\x81"
-     "\x01\x00"
-     "\xBD\x89\x01\x00\xBC\x91\x01\x00\xE1\x54\x1F\x00\xE1\x8C\x1E\x00\xE3\xD0"
-     "\x1E\x00"
-     "\xE3\x7C\x1E\x00\xE3\x50\x00\x00\x6A\x61\x00\x80\xE1\x28\x1E\x00\x6E\x61"
-     "\x00\x80"
-     "\xEF\x84\x00\x00\x72\x61\x00\x80\x76\x61\x00\x80\x7A\x61\x00\x80\x7E\x61"
-     "\x00\x80"
-     "\x82\x61\x00\x80\xEF\x38\x1F\x00\xEF\x08\x1E\x00\xA3\xCD\x05\x00\x1E\x61"
-     "\x00\x80"
-     "\x86\x61\x00\x80\x8A\x61\x00\x80\x84\xDC\x03\x00\xA6\xFD\x05\x00\xA5\xC1"
-     "\x05\x00"
-     "\x8E\x61\x00\x80\xAB\x09\x06\x00\xAA\x01\x06\x00\x86\x08\x03\x00\x87\x64"
-     "\x03\x00"
-     "\xAF\x09\x06\x00\xAE\x01\x06\x00\xAD\x09\x06\x00\xAC\x11\x06\x00\x81\x49"
-     "\x00\x00"
-     "\x80\xB1\x07\x00\xAE\x60\x00\x80\x82\x41\x00\x00\x92\x61\x00\x80\x96\x61"
-     "\x00\x80"
-     "\x9A\x61\x00\x80\x9E\x61\x00\x80\xA2\x61\x00\x80\xA6\x61\x00\x80\xAA\x61"
-     "\x00\x80"
-     "\xAE\x61\x00\x80\xB2\x61\x00\x80\xB6\x61\x00\x80\xBA\x61\x00\x80\xBE\x61"
-     "\x00\x80"
-     "\xA8\x71\x06\x00\xA9\x71\x06\x00\xAA\x71\x06\x00\xAB\x71\x06\x00\xAC\xB1"
-     "\x07\x00"
-     "\xAD\xB1\x07\x00\xAE\xB1\x07\x00\xAF\xB1\x07\x00\xC2\x61\x00\x80\xC6\x61"
-     "\x00\x80"
-     "\xCA\x61\x00\x80\xCE\x61\x00\x80\xD2\x61\x00\x80\xD6\x61\x00\x80\xDA\x61"
-     "\x00\x80"
-     "\xDE\x61\x00\x80\xB8\x69\x07\x00\xB9\x69\x07\x00\xBA\x1D\x07\x00\xBB\x11"
-     "\x07\x00"
-     "\xBC\x35\x07\x00\xBD\x3D\x07\x00\xBE\x35\x07\x00\xBF\x2D\x07\x00\xB0\xD1"
-     "\x07\x00"
-     "\xB1\xD1\x07\x00\xB2\xD1\x07\x00\xB3\x69\x07\x00\xB4\x79\x07\x00\xB5\x79"
-     "\x07\x00"
-     "\xB6\x69\x07\x00\xB7\x61\x07\x00\xB3\xF9\x07\x00\xE2\x61\x00\x80\xE6\x61"
-     "\x00\x80"
-     "\xEA\x61\x00\x80\xEE\x61\x00\x80\xB6\x31\x06\x00\xB5\x39\x06\x00\xF2\x61"
-     "\x00\x80"
-     "\xBB\x15\x06\x00\xBA\x15\x06\x00\xF6\x61\x00\x80\xFA\x61\x00\x80\xBF\xBD"
-     "\x06\x00"
-     "\xBE\xBD\x06\x00\xBD\xBD\x06\x00\xBC\x05\x06\x00\x82\x6D\x00\x00\xA3\xBD"
-     "\x07\x00"
-     "\x80\x55\x00\x00\x81\x65\x00\x00\xA6\x75\x06\x00\x84\x98\x01\x00\xBE\x90"
-     "\x01\x00"
-     "\xA5\x7D\x06\x00\xAA\x51\x06\x00\xAB\x51\x06\x00\x85\x00\x0C\x00\xBE\x60"
-     "\x0F\x00"
-     "\xAE\xF9\x06\x00\xAF\xF9\x06\x00\xAC\x41\x06\x00\xAD\xF9\x06\x00\xA8\xD1"
-     "\x01\x00"
-     "\xA9\xD1\x01\x00\xAA\xD1\x01\x00\xAB\xD1\x01\x00\xAC\xF1\x01\x00\xAD\xF1"
-     "\x01\x00"
-     "\xAE\xF1\x01\x00\xAF\xF1\x01\x00\x86\x20\x0F\x00\x87\xE4\x00\x00\x02\x62"
-     "\x00\x80"
-     "\x06\x62\x00\x80\x0A\x62\x00\x80\x0E\x62\x00\x80\x12\x62\x00\x80\x16\x62"
-     "\x00\x80"
-     "\xB8\x71\x01\x00\xB9\x71\x01\x00\xBA\x71\x01\x00\xBB\x71\x01\x00\xBC\xD5"
-     "\x00\x00"
-     "\xBD\xDD\x00\x00\xBE\xD5\x00\x00\xBF\xCD\x00\x00\xB0\x91\x01\x00\xB1\x91"
-     "\x01\x00"
-     "\xB2\x91\x01\x00\xB3\x91\x01\x00\xB4\x51\x01\x00\xB5\x51\x01\x00\xB6\x51"
-     "\x01\x00"
-     "\xB7\x51\x01\x00\xB3\x59\x02\x00\x1A\x62\x00\x80\x1E\x62\x00\x80\x22\x62"
-     "\x00\x80"
-     "\x26\x62\x00\x80\xB6\x71\x02\x00\xB5\x79\x02\x00\x2A\x62\x00\x80\xBB\x55"
-     "\x02\x00"
-     "\xBA\x55\x02\x00\x2E\x62\x00\x80\x32\x62\x00\x80\xBF\x95\x02\x00\xBE\x95"
-     "\x02\x00"
-     "\xBD\x45\x02\x00\xBC\x45\x02\x00\x36\x62\x00\x80\xA3\x1D\x02\x00\x3A\x62"
-     "\x00\x80"
-     "\x3E\x62\x00\x80\xA6\x35\x02\x00\x42\x62\x00\x80\x46\x62\x00\x80\xA5\x3D"
-     "\x02\x00"
-     "\xAA\x11\x02\x00\xAB\x11\x02\x00\x4A\x62\x00\x80\x4E\x62\x00\x80\xAE\xD1"
-     "\x02\x00"
-     "\xAF\xD1\x02\x00\xAC\x01\x02\x00\xAD\x01\x02\x00\x82\x85\xEA\x00\x83\xDD"
-     "\xEA\x00"
-     "\x52\x62\x00\x80\x84\x0C\x03\x00\x86\x81\x16\x00\x87\xCD\x17\x00\x84\x91"
-     "\x16\x00"
-     "\x85\x99\x16\x00\x8A\xA1\x12\x00\x8B\xA1\x12\x00\x56\x62\x00\x80\x5A\x62"
-     "\x00\x80"
-     "\x8E\x91\x1E\x00\x8F\x95\x1F\x00\x8C\x91\x13\x00\x8D\x91\x1E\x00\x92\xF1"
-     "\x1A\x00"
-     "\x93\xE5\x1A\x00\x5E\x62\x00\x80\xEF\x58\x24\x00\x96\x91\x06\x00\x97\x89"
-     "\x06\x00"
-     "\x94\x11\x1A\x00\x95\x99\x06\x00\x9A\xA9\x02\x00\x9B\xE1\x02\x00\x62\x62"
-     "\x00\x80"
-     "\x66\x62\x00\x80\x6A\x62\x00\x80\xE1\x88\x01\x00\x9C\xE5\x02\x00\xE3\x28"
-     "\x37\x00"
-     "\x9B\x31\x1B\x00\x9A\x25\x1E\x00\x99\x29\x1F\x00\x98\x21\x1F\x00\x9F\x51"
-     "\x07\x00"
-     "\x9E\x51\x07\x00\x9D\xD1\x1B\x00\x9C\x3D\x1B\x00\x93\x31\x17\x00\x92\x21"
-     "\x17\x00"
-     "\x91\x21\x17\x00\x90\x75\xEB\x00\x97\x19\x12\x00\x96\x31\x13\x00\x95\x39"
-     "\x13\x00"
-     "\x94\x65\x16\x00\x80\xA9\x01\x00\x81\xA9\x01\x00\x82\xB9\x01\x00\xB3\xFD"
-     "\x02\x00"
-     "\x8F\x55\xEB\x00\xB5\xED\x02\x00\xB6\xE5\x02\x00\x72\x62\x00\x80\x86\xC0"
-     "\x0C\x00"
-     "\x87\xE8\x0D\x00\xBA\x65\x01\x00\xBB\x7D\x01\x00\xBC\x6D\x01\x00\xBD\x5D"
-     "\x01\x00"
-     "\xBE\x55\x01\x00\xBF\x3D\x01\x00\xE1\x70\x07\x00\x76\x62\x00\x80\xE3\x88"
-     "\x01\x00"
-     "\x7A\x62\x00\x80\xE3\xD4\x00\x00\x84\x9C\x0D\x00\xE1\xA8\x01\x00\x7E\x62"
-     "\x00\x80"
-     "\xEF\x00\x00\x00\x82\x62\x00\x80\x86\x62\x00\x80\x8A\x62\x00\x80\x8E\x62"
-     "\x00\x80"
-     "\x92\x62\x00\x80\xEF\x7C\x06\x00\x96\x62\x00\x80\x9A\x62\x00\x80\xE3\x30"
-     "\x07\x00"
-     "\xBE\xC4\x0C\x00\xE1\x7C\x04\x00\x9E\x62\x00\x80\xA2\x62\x00\x80\xA6\x62"
-     "\x00\x80"
-     "\xAA\x62\x00\x80\xA3\xFD\x03\x00\xA2\xE1\x03\x00\xA1\xA1\x03\x00\xA0\xB1"
-     "\x07\x00"
-     "\xAE\x62\x00\x80\xEF\xCC\x07\x00\xB2\x62\x00\x80\xB6\x62\x00\x80\xBA\x62"
-     "\x00\x80"
-     "\xBE\x62\x00\x80\xFE\x61\x00\x80\x6E\x62\x00\x80\xC2\x62\x00\x80\xC6\x62"
-     "\x00\x80"
-     "\xCA\x62\x00\x80\xCE\x62\x00\x80\xA8\x89\x0D\x00\xA9\x89\x0D\x00\xAA\x99"
-     "\x0D\x00"
-     "\xAB\x99\x0D\x00\xAC\x89\x0D\x00\xAD\x89\x0D\x00\xAE\x3D\x0E\x00\xAF\x35"
-     "\x0E\x00"
-     "\xB0\x4D\x0E\x00\xB1\x55\x0E\x00\xB2\x55\x0E\x00\xB3\x65\x0E\x00\xB4\x7D"
-     "\x0E\x00"
-     "\xB5\x15\x0E\x00\xB6\x11\x0E\x00\xB7\x11\x0E\x00\xB8\x35\x0E\x00\xB9\x3D"
-     "\x0E\x00"
-     "\xBA\x01\x0E\x00\xBB\x01\x0E\x00\xBC\xF9\x0E\x00\xBD\xDD\x0E\x00\xBE\xD5"
-     "\x0E\x00"
-     "\xBF\xB5\x0E\x00\xA3\x31\x0D\x00\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xD2\x62\x00\x80\xA6\x29\x0D\x00\xA5\x21\x0D\x00\xD6\x62\x00\x80\xAB\xB1"
-     "\x0E\x00"
-     "\xAA\xA9\x0E\x00\xDA\x62\x00\x80\xDE\x62\x00\x80\xAF\xF1\x0E\x00\xAE\x99"
-     "\x0E\x00"
-     "\xAD\x91\x0E\x00\xAC\xA1\x0E\x00\xE2\x62\x00\x80\xB3\x49\x0F\x00\x86\x68"
-     "\x00\x00"
-     "\x87\x1C\x03\x00\xB6\x79\x0F\x00\xE6\x62\x00\x80\xEA\x62\x00\x80\xB5\x79"
-     "\x0F\x00"
-     "\xBA\x55\x0F\x00\xBB\x59\x0F\x00\xEE\x62\x00\x80\xF2\x62\x00\x80\xBE\x25"
-     "\x0F\x00"
-     "\xBF\x2D\x0F\x00\xBC\x39\x0F\x00\xBD\x2D\x0F\x00\xA8\x5D\x0E\x00\xA9\x81"
-     "\x0E\x00"
-     "\xAA\x85\x0E\x00\xAB\x91\x0E\x00\xAC\xB5\x0E\x00\xAD\xBD\x0E\x00\xAE\xB5"
-     "\x0E\x00"
-     "\xAF\xAD\x0E\x00\xF6\x62\x00\x80\xFA\x62\x00\x80\xFE\x62\x00\x80\x02\x63"
-     "\x00\x80"
-     "\x06\x63\x00\x80\x0A\x63\x00\x80\x0E\x63\x00\x80\x12\x63\x00\x80\xB8\xB9"
-     "\x0E\x00"
-     "\xB9\xB9\x0E\x00\xBA\x49\x01\x00\xBB\x49\x01\x00\xBC\x41\x01\x00\xBD\x41"
-     "\x01\x00"
-     "\xBE\x41\x01\x00\xBF\x41\x01\x00\xB0\xD5\x0E\x00\xB1\xD9\x0E\x00\xB2\xA9"
-     "\x0E\x00"
-     "\xB3\xBD\x0E\x00\xB4\xA5\x0E\x00\xB5\xAD\x0E\x00\xB6\xA5\x0E\x00\xB7\x8D"
-     "\x0E\x00"
-     "\xA3\x0D\x0E\x00\x16\x63\x00\x80\x1A\x63\x00\x80\x1E\x63\x00\x80\x22\x63"
-     "\x00\x80"
-     "\xA6\x3D\x0E\x00\xA5\x3D\x0E\x00\x26\x63\x00\x80\xAB\x1D\x0E\x00\xAA\x11"
-     "\x0E\x00"
-     "\x2A\x63\x00\x80\x2E\x63\x00\x80\xAF\x69\x0E\x00\xAE\x61\x0E\x00\xAD\x69"
-     "\x0E\x00"
-     "\xAC\x7D\x0E\x00\x32\x63\x00\x80\x36\x63\x00\x80\x3A\x63\x00\x80\x3E\x63"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\x42\x63\x00\x80\x46\x63"
-     "\x00\x80"
-     "\x4A\x63\x00\x80\x84\xA4\x01\x00\xBE\xA4\x01\x00\x86\xD0\x07\x00\x87\xF8"
-     "\x00\x00"
-     "\x52\x63\x00\x80\x56\x63\x00\x80\xA8\xC1\x02\x00\xA9\xD1\x02\x00\xAA\xD1"
-     "\x02\x00"
-     "\xAB\xE5\x02\x00\xAC\xFD\x02\x00\xAD\x35\x03\x00\xAE\x3D\x03\x00\xAF\x35"
-     "\x03\x00"
-     "\x5A\x63\x00\x80\x5E\x63\x00\x80\x62\x63\x00\x80\x66\x63\x00\x80\x6A\x63"
-     "\x00\x80"
-     "\x6E\x63\x00\x80\x72\x63\x00\x80\x76\x63\x00\x80\xB8\xD1\x03\x00\xB9\xD9"
-     "\x03\x00"
-     "\xBA\xE1\x03\x00\xBB\xE1\x03\x00\xBC\x91\x03\x00\xBD\x91\x03\x00\xBE\x91"
-     "\x03\x00"
-     "\xBF\x91\x03\x00\xB0\x4D\x03\x00\xB1\x55\x03\x00\xB2\x5D\x03\x00\xB3\x55"
-     "\x03\x00"
-     "\xB4\x4D\x03\x00\xB5\xF1\x03\x00\xB6\xF1\x03\x00\xB7\xF1\x03\x00\xB3\x59"
-     "\x02\x00"
-     "\x7A\x63\x00\x80\x7E\x63\x00\x80\xBE\x00\x04\x00\x82\x63\x00\x80\xB6\xB5"
-     "\x03\x00"
-     "\xB5\x75\x02\x00\x86\x63\x00\x80\xBB\x91\x03\x00\xBA\x89\x03\x00\x8A\x63"
-     "\x00\x80"
-     "\x8E\x63\x00\x80\xBF\x35\x03\x00\xBE\x89\x03\x00\xBD\x81\x03\x00\xBC\x89"
-     "\x03\x00"
-     "\x92\x63\x00\x80\xA3\x1D\x02\x00\x96\x63\x00\x80\x9A\x63\x00\x80\xA6\xF1"
-     "\x03\x00"
-     "\x9E\x63\x00\x80\xA2\x63\x00\x80\xA5\x31\x02\x00\xAA\xCD\x03\x00\xAB\xD5"
-     "\x03\x00"
-     "\xA6\x63\x00\x80\xAA\x63\x00\x80\xAE\xCD\x03\x00\xAF\x71\x03\x00\xAC\xCD"
-     "\x03\x00"
-     "\xAD\xC5\x03\x00\xAE\x63\x00\x80\x82\x21\x00\x00\x81\x15\x00\x00\x80\x15"
-     "\x00\x00"
-     "\xEF\x04\x00\x00\x84\x94\x02\x00\xB2\x63\x00\x80\xB6\x63\x00\x80\xE3\xD8"
-     "\x00\x00"
-     "\xBA\x63\x00\x80\xE1\xF8\x01\x00\xC2\x63\x00\x80\xC6\x63\x00\x80\xCA\x63"
-     "\x00\x80"
-     "\xCE\x63\x00\x80\xD2\x63\x00\x80\x86\xE0\x04\x00\x87\x18\x05\x00\xD6\x63"
-     "\x00\x80"
-     "\xB3\x6D\x01\x00\xDA\x63\x00\x80\xB5\x7D\x01\x00\xB6\x75\x01\x00\xDE\x63"
-     "\x00\x80"
-     "\xE2\x63\x00\x80\xE6\x63\x00\x80\xBA\x51\x01\x00\xBB\x51\x01\x00\xBC\xF5"
-     "\x01\x00"
-     "\xBD\xFD\x01\x00\xBE\xE9\x01\x00\xBF\xE1\x01\x00\xE3\x04\x0E\x00\xE3\x18"
-     "\x0F\x00"
-     "\xE1\xF0\x0E\x00\xE1\x34\x0C\x00\xEA\x63\x00\x80\xEE\x63\x00\x80\xF2\x63"
-     "\x00\x80"
-     "\xF6\x63\x00\x80\xB2\xAD\x37\x00\xB3\x19\x37\x00\xB0\xB9\x32\x00\xB1\x01"
-     "\x34\x00"
-     "\xEF\x80\x0E\x00\xEF\x20\x0C\x00\xB4\x01\x28\x00\xB5\xBD\x2B\x00\xAA\x01"
-     "\x3C\x00"
-     "\xAB\xC1\x3F\x00\xA8\xCD\x3B\x00\xA9\xA1\x3B\x00\xAE\xD5\x33\x00\xAF\xDD"
-     "\x33\x00"
-     "\xAC\xC9\x3F\x00\xAD\x01\x30\x00\xA2\xED\x03\x00\xA3\x49\x03\x00\xBE\xA0"
-     "\x04\x00"
-     "\xA1\xE5\x03\x00\xA6\x81\x07\x00\xA7\x01\x38\x00\xA4\x01\x04\x00\xA5\xF9"
-     "\x07\x00"
-     "\xA3\x21\x06\x00\xBE\x63\x00\x80\xFA\x63\x00\x80\xFE\x63\x00\x80\x02\x64"
-     "\x00\x80"
-     "\xA6\x39\x06\x00\xA5\x31\x06\x00\x06\x64\x00\x80\xAB\x1D\x06\x00\xAA\x1D"
-     "\x06\x00"
-     "\x0A\x64\x00\x80\x0E\x64\x00\x80\xAF\xAD\x06\x00\xAE\xA5\x06\x00\xAD\xB1"
-     "\x06\x00"
-     "\xAC\xB9\x06\x00\x12\x64\x00\x80\xEF\x30\x04\x00\x16\x64\x00\x80\x1A\x64"
-     "\x00\x80"
-     "\x1E\x64\x00\x80\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D\x00\x00\x22\x64"
-     "\x00\x80"
-     "\x26\x64\x00\x80\x2A\x64\x00\x80\x84\x20\x03\x00\x2E\x64\x00\x80\xE3\x70"
-     "\x05\x00"
-     "\x32\x64\x00\x80\xE1\x94\x07\x00\xA8\x79\x05\x00\xA9\x79\x05\x00\xAA\xA9"
-     "\x06\x00"
-     "\xAB\xA9\x06\x00\xAC\xB9\x06\x00\xAD\xB9\x06\x00\xAE\xA9\x06\x00\xAF\xA9"
-     "\x06\x00"
-     "\x86\x00\x00\x00\x87\x84\x00\x00\x36\x64\x00\x80\x3A\x64\x00\x80\x3E\x64"
-     "\x00\x80"
-     "\x42\x64\x00\x80\x46\x64\x00\x80\x4A\x64\x00\x80\xB8\x01\x06\x00\xB9\x0D"
-     "\x06\x00"
-     "\xBA\x05\x06\x00\xBB\x19\x06\x00\xBC\x09\x06\x00\xBD\x3D\x06\x00\xBE\x35"
-     "\x06\x00"
-     "\xBF\xDD\x07\x00\xB0\xD9\x06\x00\xB1\xD9\x06\x00\xB2\x6D\x06\x00\xB3\x65"
-     "\x06\x00"
-     "\xB4\x61\x06\x00\xB5\x69\x06\x00\xB6\x51\x06\x00\xB7\x51\x06\x00\xA8\x2D"
-     "\x06\x00"
-     "\xA9\x35\x06\x00\xAA\x3D\x06\x00\xAB\x35\x06\x00\xAC\x2D\x06\x00\xAD\x5D"
-     "\x06\x00"
-     "\xAE\x49\x06\x00\xAF\x41\x06\x00\x4E\x64\x00\x80\x52\x64\x00\x80\x56\x64"
-     "\x00\x80"
-     "\x5A\x64\x00\x80\x5E\x64\x00\x80\x62\x64\x00\x80\x66\x64\x00\x80\x6A\x64"
-     "\x00\x80"
-     "\xB8\x51\x01\x00\xB9\x59\x01\x00\xBA\x61\x01\x00\xBB\x61\x01\x00\xBC\x11"
-     "\x01\x00"
-     "\xBD\x1D\x01\x00\xBE\x15\x01\x00\xBF\x0D\x01\x00\xB0\x01\x06\x00\xB1\x01"
-     "\x06\x00"
-     "\xB2\xF9\x06\x00\xB3\xF9\x06\x00\xB4\xE9\x06\x00\xB5\xE9\x06\x00\xB6\x71"
-     "\x01\x00"
-     "\xB7\x71\x01\x00\xB3\xB5\x06\x00\x6E\x64\x00\x80\x72\x64\x00\x80\x76\x64"
-     "\x00\x80"
-     "\x7A\x64\x00\x80\xB6\xDD\x06\x00\xB5\xA5\x06\x00\x7E\x64\x00\x80\xBB\xC5"
-     "\x06\x00"
-     "\xBA\xC5\x06\x00\x82\x64\x00\x80\x86\x64\x00\x80\xBF\x2D\x06\x00\xBE\x2D"
-     "\x06\x00"
-     "\xBD\x2D\x06\x00\xBC\x2D\x06\x00\x8A\x64\x00\x80\xA3\xF1\x06\x00\x8E\x64"
-     "\x00\x80"
-     "\x92\x64\x00\x80\xA6\x99\x06\x00\x96\x64\x00\x80\x9A\x64\x00\x80\xA5\xE1"
-     "\x06\x00"
-     "\xAA\x81\x06\x00\xAB\x81\x06\x00\xBE\xB4\x01\x00\x84\xA8\x01\x00\xAE\x69"
-     "\x06\x00"
-     "\xAF\x69\x06\x00\xAC\x69\x06\x00\xAD\x69\x06\x00\xA8\x79\x02\x00\xA9\xAD"
-     "\x02\x00"
-     "\xAA\xA5\x02\x00\xAB\xBD\x02\x00\xAC\xA5\x02\x00\xAD\xAD\x02\x00\xAE\xA5"
-     "\x02\x00"
-     "\xAF\x3D\x03\x00\x80\xED\x03\x00\x81\xF1\x03\x00\x82\xF1\x03\x00\xA2\x64"
-     "\x00\x80"
-     "\xA6\x64\x00\x80\xAA\x64\x00\x80\x87\xAC\x03\x00\x86\x9C\x03\x00\xB8\xC9"
-     "\x03\x00"
-     "\xB9\xC9\x03\x00\xBA\xD9\x03\x00\xBB\xD1\x03\x00\xBC\xF9\x03\x00\xBD\xF9"
-     "\x03\x00"
-     "\xBE\x99\x03\x00\xBF\x95\x03\x00\xB0\x45\x03\x00\xB1\x4D\x03\x00\xB2\x45"
-     "\x03\x00"
-     "\xB3\x5D\x03\x00\xB4\x45\x03\x00\xB5\x4D\x03\x00\xB6\x45\x03\x00\xB7\xF9"
-     "\x03\x00"
-     "\xAE\x64\x00\x80\xB2\x64\x00\x80\xB3\x25\x02\x00\x84\x40\x3C\x00\xB5\x25"
-     "\x02\x00"
-     "\xB6\x64\x00\x80\xBA\x64\x00\x80\xB6\x25\x02\x00\xBE\x64\x00\x80\xC2\x64"
-     "\x00\x80"
-     "\xBB\x81\x03\x00\xBA\x89\x03\x00\xBD\x81\x03\x00\xBC\x99\x03\x00\xBF\x81"
-     "\x03\x00"
-     "\xBE\x89\x03\x00\xC6\x64\x00\x80\xA6\x7D\x02\x00\xA5\x7D\x02\x00\xCA\x64"
-     "\x00\x80"
-     "\xA3\x7D\x02\x00\xCE\x64\x00\x80\xD2\x64\x00\x80\xD6\x64\x00\x80\xAF\xD9"
-     "\x03\x00"
-     "\xAE\xD1\x03\x00\xAD\xD9\x03\x00\xAC\xC1\x03\x00\xAB\xD9\x03\x00\xAA\xD1"
-     "\x03\x00"
-     "\xDA\x64\x00\x80\xDE\x64\x00\x80\xA8\x2D\x01\x00\xA9\x35\x01\x00\xAA\x3D"
-     "\x01\x00"
-     "\xAB\x8D\x01\x00\xAC\x95\x01\x00\xAD\x81\x01\x00\xAE\x81\x01\x00\xAF\xBD"
-     "\x01\x00"
-     "\xE2\x64\x00\x80\xE6\x64\x00\x80\xEA\x64\x00\x80\xEE\x64\x00\x80\xF2\x64"
-     "\x00\x80"
-     "\xF6\x64\x00\x80\xFA\x64\x00\x80\xFE\x64\x00\x80\xB8\x61\x00\x00\xB9\x61"
-     "\x00\x00"
-     "\xBA\x61\x00\x00\xBB\x61\x00\x00\xBC\x61\x00\x00\xBD\x61\x00\x00\xBE\x61"
-     "\x00\x00"
-     "\xBF\x61\x00\x00\xB0\xC5\x01\x00\xB1\xA5\x01\x00\xB2\xA5\x01\x00\xB3\xB9"
-     "\x01\x00"
-     "\xB4\xA9\x01\x00\xB5\xA9\x01\x00\xB6\x81\x01\x00\xB7\x81\x01\x00\xE1\xD8"
-     "\x06\x00"
-     "\x02\x65\x00\x80\xE3\xF4\x01\x00\x06\x65\x00\x80\x81\x1D\x00\x00\x80\x11"
-     "\x00\x00"
-     "\x0A\x65\x00\x80\x82\x05\x00\x00\xBE\xAC\x3D\x00\x12\x65\x00\x80\x86\x08"
-     "\x3C\x00"
-     "\x87\xFC\x3D\x00\x16\x65\x00\x80\x1A\x65\x00\x80\xEF\xA8\x06\x00\x1E\x65"
-     "\x00\x80"
-     "\x22\x65\x00\x80\x26\x65\x00\x80\x2A\x65\x00\x80\xEF\xDC\x00\x00\x2E\x65"
-     "\x00\x80"
-     "\x32\x65\x00\x80\x36\x65\x00\x80\x3A\x65\x00\x80\x3E\x65\x00\x80\x42\x65"
-     "\x00\x80"
-     "\x46\x65\x00\x80\x4A\x65\x00\x80\x4E\x65\x00\x80\xE1\x4C\x01\x00\x4E\x63"
-     "\x00\x80"
-     "\xE3\x00\x00\x00\xB3\x35\x3E\x00\x0E\x65\x00\x80\x52\x65\x00\x80\x56\x65"
-     "\x00\x80"
-     "\x5A\x65\x00\x80\xB6\x81\x3E\x00\xB5\x9D\x3E\x00\x5E\x65\x00\x80\xBB\xE1"
-     "\x3E\x00"
-     "\xBA\x81\x3E\x00\x62\x65\x00\x80\x66\x65\x00\x80\xBF\xED\x3E\x00\xBE\xED"
-     "\x3E\x00"
-     "\xBD\xED\x3E\x00\xBC\xF1\x3E\x00\x6A\x65\x00\x80\x6E\x65\x00\x80\xEF\x10"
-     "\x3D\x00"
-     "\x72\x65\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39\x00\x00\x76\x65"
-     "\x00\x80"
-     "\x7A\x65\x00\x80\x84\x6C\x02\x00\xBE\x64\x03\x00\x7E\x65\x00\x80\xE1\x54"
-     "\x3C\x00"
-     "\x82\x65\x00\x80\xE3\xE4\x3C\x00\x86\x65\x00\x80\xA3\xB5\x3E\x00\x8A\x65"
-     "\x00\x80"
-     "\x87\x48\x03\x00\x86\x2C\x00\x00\x8E\x65\x00\x80\xA6\x01\x3E\x00\xA5\x1D"
-     "\x3E\x00"
-     "\x92\x65\x00\x80\xAB\x61\x3E\x00\xAA\x01\x3E\x00\x96\x65\x00\x80\x9A\x65"
-     "\x00\x80"
-     "\xAF\x6D\x3E\x00\xAE\x6D\x3E\x00\xAD\x6D\x3E\x00\xAC\x71\x3E\x00\x9E\x65"
-     "\x00\x80"
-     "\xB3\x71\x3F\x00\xA2\x65\x00\x80\xA6\x65\x00\x80\xB6\x15\x3F\x00\xAA\x65"
-     "\x00\x80"
-     "\xAE\x65\x00\x80\xB5\x1D\x3F\x00\xBA\x35\x3F\x00\xBB\x3D\x3F\x00\xB2\x65"
-     "\x00\x80"
-     "\xB6\x65\x00\x80\xBE\x15\x3F\x00\xBF\x1D\x3F\x00\xBC\x25\x3F\x00\xBD\x1D"
-     "\x3F\x00"
-     "\xA8\xA5\x3E\x00\xA9\xB5\x3E\x00\xAA\xBD\x3E\x00\xAB\xB1\x3E\x00\xAC\xD1"
-     "\x3E\x00"
-     "\xAD\xD1\x3E\x00\xAE\xD1\x3E\x00\xAF\xCD\x3E\x00\xBA\x65\x00\x80\xBE\x65"
-     "\x00\x80"
-     "\xC2\x65\x00\x80\xC6\x65\x00\x80\xCA\x65\x00\x80\xCE\x65\x00\x80\xD2\x65"
-     "\x00\x80"
-     "\xD6\x65\x00\x80\xB8\x49\x01\x00\xB9\x49\x01\x00\xBA\x59\x01\x00\xBB\x51"
-     "\x01\x00"
-     "\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x19\x01\x00\xBF\x19\x01\x00\xB0\xB5"
-     "\x3E\x00"
-     "\xB1\xBD\x3E\x00\xB2\xB5\x3E\x00\xB3\x8D\x3E\x00\xB4\x95\x3E\x00\xB5\x9D"
-     "\x3E\x00"
-     "\xB6\x95\x3E\x00\xB7\x8D\x3E\x00\xA3\x35\x3E\x00\xDA\x65\x00\x80\xDE\x65"
-     "\x00\x80"
-     "\xE2\x65\x00\x80\xE6\x65\x00\x80\xA6\x51\x3E\x00\xA5\x59\x3E\x00\xEA\x65"
-     "\x00\x80"
-     "\xAB\x79\x3E\x00\xAA\x71\x3E\x00\x84\xA8\x02\x00\xBE\xAC\x01\x00\xAF\x59"
-     "\x3E\x00"
-     "\xAE\x51\x3E\x00\xAD\x59\x3E\x00\xAC\x61\x3E\x00\x80\x4D\x00\x00\x81\x55"
-     "\x00\x00"
-     "\x82\x55\x00\x00\xB3\x49\x01\x00\xF2\x65\x00\x80\xB5\x49\x01\x00\xB6\x71"
-     "\x01\x00"
-     "\xF6\x65\x00\x80\x86\x40\x00\x00\x87\x14\x01\x00\xBA\x21\x01\x00\xBB\x21"
-     "\x01\x00"
-     "\xBC\x25\x01\x00\xBD\x2D\x01\x00\xBE\x11\x01\x00\xBF\x11\x01\x00\xA8\x75"
-     "\x02\x00"
-     "\xA9\x79\x02\x00\xAA\x8D\x02\x00\xAB\x9D\x02\x00\xAC\x8D\x02\x00\xAD\xBD"
-     "\x02\x00"
-     "\xAE\xB5\x02\x00\xAF\xDD\x02\x00\xFA\x65\x00\x80\xFE\x65\x00\x80\x02\x66"
-     "\x00\x80"
-     "\x06\x66\x00\x80\x0A\x66\x00\x80\x0E\x66\x00\x80\x12\x66\x00\x80\x16\x66"
-     "\x00\x80"
-     "\xB8\x69\x03\x00\xB9\x69\x03\x00\xBA\x79\x03\x00\xBB\x79\x03\x00\xBC\x69"
-     "\x03\x00"
-     "\xBD\x69\x03\x00\xBE\xDD\x03\x00\xBF\xD5\x03\x00\xB0\xA9\x02\x00\xB1\xA9"
-     "\x02\x00"
-     "\xB2\xBD\x02\x00\xB3\xB5\x02\x00\xB4\x99\x02\x00\xB5\x99\x02\x00\xB6\x59"
-     "\x03\x00"
-     "\xB7\x59\x03\x00\x1A\x66\x00\x80\x1E\x66\x00\x80\x22\x66\x00\x80\xA3\x09"
-     "\x02\x00"
-     "\x84\x5C\x04\x00\xA5\x09\x02\x00\xA6\x31\x02\x00\x26\x66\x00\x80\x2A\x66"
-     "\x00\x80"
-     "\x2E\x66\x00\x80\xAA\x61\x02\x00\xAB\x61\x02\x00\xAC\x65\x02\x00\xAD\x6D"
-     "\x02\x00"
-     "\xAE\x51\x02\x00\xAF\x51\x02\x00\x32\x66\x00\x80\x36\x66\x00\x80\x3A\x66"
-     "\x00\x80"
-     "\xBE\xC0\x04\x00\x3E\x66\x00\x80\x42\x66\x00\x80\x46\x66\x00\x80\x4A\x66"
-     "\x00\x80"
-     "\x4E\x66\x00\x80\x52\x66\x00\x80\x56\x66\x00\x80\x5A\x66\x00\x80\x80\x15"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x82\x15\x00\x00\x62\x66\x00\x80\xA8\xD1\x02\x00\xA9\xD9"
-     "\x02\x00"
-     "\xAA\x29\x01\x00\xAB\x29\x01\x00\xAC\x39\x01\x00\xAD\x39\x01\x00\xAE\x29"
-     "\x01\x00"
-     "\xAF\x29\x01\x00\x66\x66\x00\x80\x6A\x66\x00\x80\x6E\x66\x00\x80\x72\x66"
-     "\x00\x80"
-     "\x76\x66\x00\x80\x7A\x66\x00\x80\x7E\x66\x00\x80\x82\x66\x00\x80\xB8\x81"
-     "\x01\x00"
-     "\xB9\x81\x01\x00\xBA\x81\x01\x00\xBB\x9D\x01\x00\xBC\x85\x01\x00\xBD\x8D"
-     "\x01\x00"
-     "\xBE\x85\x01\x00\xBF\xBD\x01\x00\xB0\x59\x01\x00\xB1\x59\x01\x00\xB2\xE9"
-     "\x01\x00"
-     "\xB3\xE9\x01\x00\xB4\xF9\x01\x00\xB5\xE9\x01\x00\xB6\xD9\x01\x00\xB7\xD9"
-     "\x01\x00"
-     "\xE1\x30\x3E\x00\x86\x66\x00\x80\xE3\xFC\x01\x00\x8A\x66\x00\x80\x8E\x66"
-     "\x00\x80"
-     "\x92\x66\x00\x80\x86\x38\x04\x00\x87\x64\x02\x00\x84\xA8\x05\x00\x96\x66"
-     "\x00\x80"
-     "\x9A\x66\x00\x80\x9E\x66\x00\x80\xA2\x66\x00\x80\xA6\x66\x00\x80\xEF\x28"
-     "\x01\x00"
-     "\xAA\x66\x00\x80\x85\xFC\x13\x00\xAE\x66\x00\x80\xB2\x66\x00\x80\xB6\x66"
-     "\x00\x80"
-     "\xEF\x84\x00\x00\xBA\x66\x00\x80\xBE\x66\x00\x80\xC2\x66\x00\x80\xE3\x58"
-     "\x00\x00"
-     "\xC6\x66\x00\x80\xE1\x7C\x01\x00\xCA\x66\x00\x80\x82\x1D\x00\x00\x9E\x64"
-     "\x00\x80"
-     "\x80\x25\x00\x00\x81\x19\x00\x00\xB3\x21\x06\x00\x5E\x66\x00\x80\xBE\x04"
-     "\x03\x00"
-     "\xD2\x66\x00\x80\xCE\x66\x00\x80\xB6\x39\x06\x00\xB5\x31\x06\x00\xD6\x66"
-     "\x00\x80"
-     "\xBB\x1D\x06\x00\xBA\x1D\x06\x00\x86\x88\x00\x00\x87\x0C\x03\x00\xBF\xAD"
-     "\x06\x00"
-     "\xBE\xBD\x06\x00\xBD\xBD\x06\x00\xBC\xBD\x06\x00\xDA\x66\x00\x80\xE1\xBC"
-     "\x07\x00"
-     "\xDE\x66\x00\x80\xE3\xCC\x04\x00\xE2\x66\x00\x80\xE6\x66\x00\x80\xEA\x66"
-     "\x00\x80"
-     "\xEE\x66\x00\x80\xF2\x66\x00\x80\xF6\x66\x00\x80\xFA\x66\x00\x80\xFE\x66"
-     "\x00\x80"
-     "\x02\x67\x00\x80\x06\x67\x00\x80\x0A\x67\x00\x80\xEF\x48\x04\x00\xA3\xA1"
-     "\x06\x00"
-     "\x0E\x67\x00\x80\x12\x67\x00\x80\x16\x67\x00\x80\x1A\x67\x00\x80\xA6\xB9"
-     "\x06\x00"
-     "\xA5\xB1\x06\x00\x1E\x67\x00\x80\xAB\x9D\x06\x00\xAA\x9D\x06\x00\x22\x67"
-     "\x00\x80"
-     "\x26\x67\x00\x80\xAF\x2D\x06\x00\xAE\x3D\x06\x00\xAD\x3D\x06\x00\xAC\x3D"
-     "\x06\x00"
-     "\x2A\x67\x00\x80\xB3\xE1\x07\x00\x2E\x67\x00\x80\x32\x67\x00\x80\xB6\xE1"
-     "\x07\x00"
-     "\x36\x67\x00\x80\x3A\x67\x00\x80\xB5\xF1\x07\x00\xBA\xDD\x07\x00\xBB\x61"
-     "\x07\x00"
-     "\x84\xE0\x03\x00\xBE\xE4\x00\x00\xBE\x65\x07\x00\xBF\x51\x07\x00\xBC\x75"
-     "\x07\x00"
-     "\xBD\x6D\x07\x00\xA8\xFD\x07\x00\xA9\x25\x06\x00\xAA\x2D\x06\x00\xAB\x25"
-     "\x06\x00"
-     "\xAC\x3D\x06\x00\xAD\x25\x06\x00\xAE\x2D\x06\x00\xAF\x25\x06\x00\x80\xED"
-     "\x01\x00"
-     "\x81\xF1\x01\x00\x82\xF1\x01\x00\x42\x67\x00\x80\x86\x90\x00\x00\x87\xF0"
-     "\x01\x00"
-     "\x46\x67\x00\x80\x4A\x67\x00\x80\xB8\xF1\x06\x00\xB9\xF1\x06\x00\xBA\xF5"
-     "\x06\x00"
-     "\xBB\x89\x06\x00\xBC\x99\x06\x00\xBD\x99\x06\x00\xBE\x8D\x06\x00\xBF\x85"
-     "\x06\x00"
-     "\xB0\x5D\x06\x00\xB1\xE1\x06\x00\xB2\xE5\x06\x00\xB3\xFD\x06\x00\xB4\xE5"
-     "\x06\x00"
-     "\xB5\xD1\x06\x00\xB6\xD1\x06\x00\xB7\xD1\x06\x00\xA3\xA5\x06\x00\x4E\x67"
-     "\x00\x80"
-     "\x52\x67\x00\x80\x56\x67\x00\x80\x5A\x67\x00\x80\xA6\xA5\x06\x00\xA5\xB5"
-     "\x06\x00"
-     "\x5E\x67\x00\x80\xAB\x25\x06\x00\xAA\x99\x06\x00\x62\x67\x00\x80\x66\x67"
-     "\x00\x80"
-     "\xAF\x15\x06\x00\xAE\x21\x06\x00\xAD\x29\x06\x00\xAC\x31\x06\x00\x6A\x67"
-     "\x00\x80"
-     "\xB3\x85\x01\x00\x6E\x67\x00\x80\x72\x67\x00\x80\xB6\x85\x01\x00\x76\x67"
-     "\x00\x80"
-     "\x7A\x67\x00\x80\xB5\x8D\x01\x00\xBA\x81\x01\x00\xBB\x81\x01\x00\x7E\x67"
-     "\x00\x80"
-     "\x82\x67\x00\x80\xBE\x81\x01\x00\xBF\x81\x01\x00\xBC\x91\x01\x00\xBD\x91"
-     "\x01\x00"
-     "\xA8\x0D\x02\x00\xA9\x19\x02\x00\xAA\x49\x02\x00\xAB\x45\x02\x00\xAC\x41"
-     "\x02\x00"
-     "\xAD\x41\x02\x00\xAE\x41\x02\x00\xAF\x41\x02\x00\x86\x67\x00\x80\x8A\x67"
-     "\x00\x80"
-     "\x8E\x67\x00\x80\x92\x67\x00\x80\x84\x38\x03\x00\x96\x67\x00\x80\x9A\x67"
-     "\x00\x80"
-     "\x9E\x67\x00\x80\xB8\xC9\x02\x00\xB9\xC9\x02\x00\xBA\xD9\x02\x00\xBB\xD1"
-     "\x02\x00"
-     "\xBC\xF9\x02\x00\xBD\xF9\x02\x00\xBE\x99\x02\x00\xBF\x95\x02\x00\xB0\x01"
-     "\x02\x00"
-     "\xB1\x01\x02\x00\xB2\x01\x02\x00\xB3\x01\x02\x00\xB4\x01\x02\x00\xB5\x01"
-     "\x02\x00"
-     "\xB6\x01\x02\x00\xB7\xF9\x02\x00\xA2\x67\x00\x80\xA3\xC5\x02\x00\xA6\x67"
-     "\x00\x80"
-     "\xAA\x67\x00\x80\xA6\xC5\x02\x00\xB2\x67\x00\x80\xB6\x67\x00\x80\xA5\xCD"
-     "\x02\x00"
-     "\xAA\xC1\x02\x00\xAB\xC1\x02\x00\x86\x88\x0C\x00\x87\x18\x03\x00\xAE\xC1"
-     "\x02\x00"
-     "\xAF\xC1\x02\x00\xAC\xD1\x02\x00\xAD\xD1\x02\x00\x81\x49\x00\x00\x80\x45"
-     "\x00\x00"
-     "\xBE\x44\x0C\x00\x82\x41\x00\x00\xBA\x67\x00\x80\xBE\x67\x00\x80\xC2\x67"
-     "\x00\x80"
-     "\xC6\x67\x00\x80\xCA\x67\x00\x80\xCE\x67\x00\x80\xD2\x67\x00\x80\xD6\x67"
-     "\x00\x80"
-     "\x84\x3C\x0D\x00\xDA\x67\x00\x80\xBE\x30\x0D\x00\xDE\x67\x00\x80\xA8\x2D"
-     "\x02\x00"
-     "\xA9\x3D\x02\x00\xAA\x35\x02\x00\xAB\x89\x02\x00\xAC\x99\x02\x00\xAD\x99"
-     "\x02\x00"
-     "\xAE\x89\x02\x00\xAF\x81\x02\x00\x84\xEC\x0F\x00\xE2\x67\x00\x80\xBE\xE0"
-     "\x0F\x00"
-     "\xE6\x67\x00\x80\xEA\x67\x00\x80\xEE\x67\x00\x80\xF2\x67\x00\x80\xF6\x67"
-     "\x00\x80"
-     "\xB8\xAD\x02\x00\xB9\x65\x01\x00\xBA\x6D\x01\x00\xBB\x65\x01\x00\xBC\x7D"
-     "\x01\x00"
-     "\xBD\x61\x01\x00\xBE\x61\x01\x00\xBF\x61\x01\x00\xB0\xC9\x02\x00\xB1\xC9"
-     "\x02\x00"
-     "\xB2\xA9\x02\x00\xB3\xA5\x02\x00\xB4\xBD\x02\x00\xB5\xA1\x02\x00\xB6\xA1"
-     "\x02\x00"
-     "\xB7\x95\x02\x00\xE1\x00\x06\x00\xFA\x67\x00\x80\xE3\x38\x06\x00\xFE\x67"
-     "\x00\x80"
-     "\x02\x68\x00\x80\xE1\xA4\x01\x00\x06\x68\x00\x80\xE3\x64\x01\x00\x84\xA0"
-     "\x0F\x00"
-     "\x0A\x68\x00\x80\x0E\x68\x00\x80\xEF\xB8\x01\x00\x81\x7D\x00\x00\x80\x7D"
-     "\x00\x00"
-     "\xEF\x48\x01\x00\x82\x69\x00\x00\xEE\x65\x00\x80\xB3\xF9\x02\x00\x12\x68"
-     "\x00\x80"
-     "\x16\x68\x00\x80\xB6\x5D\x02\x00\x1A\x68\x00\x80\x22\x68\x00\x80\xB5\x55"
-     "\x02\x00"
-     "\xBA\x49\x02\x00\xBB\x49\x02\x00\x86\xC8\x0C\x00\x87\x50\x0D\x00\xBE\x3D"
-     "\x02\x00"
-     "\xBF\x21\x02\x00\xBC\x39\x02\x00\xBD\x35\x02\x00\xA3\x3D\x0D\x00\xAE\x67"
-     "\x00\x80"
-     "\x1E\x68\x00\x80\x26\x68\x00\x80\x2A\x68\x00\x80\xA6\x99\x0D\x00\xA5\x91"
-     "\x0D\x00"
-     "\x2E\x68\x00\x80\xAB\x8D\x0D\x00\xAA\x8D\x0D\x00\x32\x68\x00\x80\x36\x68"
-     "\x00\x80"
-     "\xAF\xE5\x0D\x00\xAE\xF9\x0D\x00\xAD\xF1\x0D\x00\xAC\xFD\x0D\x00\x3A\x68"
-     "\x00\x80"
-     "\x3E\x68\x00\x80\x42\x68\x00\x80\x46\x68\x00\x80\xE3\x70\x0C\x00\x4A\x68"
-     "\x00\x80"
-     "\xE1\x70\x0F\x00\x4E\x68\x00\x80\xEF\x50\x0C\x00\x52\x68\x00\x80\x56\x68"
-     "\x00\x80"
-     "\x5A\x68\x00\x80\x5E\x68\x00\x80\x62\x68\x00\x80\x66\x68\x00\x80\x6A\x68"
-     "\x00\x80"
-     "\xA8\x15\x0E\x00\xA9\x1D\x0E\x00\xAA\x59\x0E\x00\xAB\x59\x0E\x00\xAC\x49"
-     "\x0E\x00"
-     "\xAD\x49\x0E\x00\xAE\x79\x0E\x00\xAF\x79\x0E\x00\x6E\x68\x00\x80\x72\x68"
-     "\x00\x80"
-     "\x76\x68\x00\x80\x7A\x68\x00\x80\x7E\x68\x00\x80\x82\xB9\x01\x00\x81\xB1"
-     "\x01\x00"
-     "\x80\xB9\x01\x00\xB8\xF9\x0F\x00\xB9\xF9\x0F\x00\xBA\x4D\x0F\x00\xBB\x41"
-     "\x0F\x00"
-     "\xBC\x41\x0F\x00\xBD\x4D\x0F\x00\xBE\x7D\x0F\x00\xBF\x75\x0F\x00\xB0\x09"
-     "\x0E\x00"
-     "\xB1\x09\x0E\x00\xB2\x1D\x0E\x00\xB3\x11\x0E\x00\xB4\xD5\x0F\x00\xB5\xD9"
-     "\x0F\x00"
-     "\xB6\xC9\x0F\x00\xB7\xC9\x0F\x00\xB3\xD1\x0F\x00\x82\x68\x00\x80\x86\x68"
-     "\x00\x80"
-     "\x8A\x68\x00\x80\x8E\x68\x00\x80\xB6\xC9\x0F\x00\xB5\xC1\x0F\x00\x92\x68"
-     "\x00\x80"
-     "\xBB\xCD\x0F\x00\xBA\xCD\x0F\x00\x86\x88\x00\x00\x87\xAC\x01\x00\xBF\x1D"
-     "\x0E\x00"
-     "\xBE\x11\x0E\x00\xBD\xC5\x0F\x00\xBC\xC9\x0F\x00\x96\x68\x00\x80\xA3\x95"
-     "\x0F\x00"
-     "\x9A\x68\x00\x80\x9E\x68\x00\x80\xA6\x8D\x0F\x00\xA2\x68\x00\x80\xA6\x68"
-     "\x00\x80"
-     "\xA5\x85\x0F\x00\xAA\x89\x0F\x00\xAB\x89\x0F\x00\xAA\x68\x00\x80\xAE\x68"
-     "\x00\x80"
-     "\xAE\x55\x0E\x00\xAF\x59\x0E\x00\xAC\x8D\x0F\x00\xAD\x81\x0F\x00\xA8\x59"
-     "\x0E\x00"
-     "\xA9\x89\x01\x00\xAA\x99\x01\x00\xAB\x91\x01\x00\xAC\xB9\x01\x00\xAD\xB9"
-     "\x01\x00"
-     "\xAE\xD9\x01\x00\xAF\xD9\x01\x00\xBE\xFC\x00\x00\xB2\x68\x00\x80\xB6\x68"
-     "\x00\x80"
-     "\xBA\x68\x00\x80\xBE\x68\x00\x80\xC2\x68\x00\x80\xC6\x68\x00\x80\xCA\x68"
-     "\x00\x80"
-     "\xB8\xB5\x01\x00\xB9\xBD\x01\x00\xBA\xB5\x01\x00\xBB\x6D\x01\x00\xBC\x75"
-     "\x01\x00"
-     "\xBD\x7D\x01\x00\xBE\x75\x01\x00\xBF\x6D\x01\x00\xB0\xA9\x01\x00\xB1\xA9"
-     "\x01\x00"
-     "\xB2\xBD\x01\x00\xB3\xB1\x01\x00\xB4\x95\x01\x00\xB5\x9D\x01\x00\xB6\x95"
-     "\x01\x00"
-     "\xB7\x8D\x01\x00\xB3\x11\x02\x00\xCE\x68\x00\x80\xD2\x68\x00\x80\xD6\x68"
-     "\x00\x80"
-     "\xDA\x68\x00\x80\xB6\x39\x02\x00\xB5\x31\x02\x00\xDE\x68\x00\x80\xBB\x6D"
-     "\x02\x00"
-     "\xBA\x6D\x02\x00\xE2\x68\x00\x80\xE6\x68\x00\x80\xBF\x59\x02\x00\xBE\x51"
-     "\x02\x00"
-     "\xBD\x65\x02\x00\xBC\x69\x02\x00\xEA\x68\x00\x80\xA3\x55\x02\x00\xEE\x68"
-     "\x00\x80"
-     "\xF2\x68\x00\x80\xA6\x7D\x02\x00\xF6\x68\x00\x80\xFA\x68\x00\x80\xA5\x75"
-     "\x02\x00"
-     "\xAA\x29\x02\x00\xAB\x29\x02\x00\x02\x69\x00\x80\x06\x69\x00\x80\xAE\x15"
-     "\x02\x00"
-     "\xAF\x1D\x02\x00\xAC\x2D\x02\x00\xAD\x21\x02\x00\x80\xCD\x00\x00\x81\xD1"
-     "\x00\x00"
-     "\x82\xD1\x00\x00\x84\x00\x04\x00\x0A\x69\x00\x80\xE1\xA4\x01\x00\x0E\x69"
-     "\x00\x80"
-     "\xE3\x2C\x0F\x00\x86\x80\x04\x00\x87\xA4\x04\x00\x12\x69\x00\x80\xEF\xF4"
-     "\x01\x00"
-     "\x16\x69\x00\x80\x1A\x69\x00\x80\x1E\x69\x00\x80\xAA\x00\x00\x00\x22\x69"
-     "\x00\x80"
-     "\x26\x69\x00\x80\x2A\x69\x00\x80\xB3\x85\x03\x00\x2E\x69\x00\x80\xB5\x81"
-     "\x03\x00"
-     "\xB6\x81\x03\x00\x32\x69\x00\x80\x36\x69\x00\x80\x3A\x69\x00\x80\xBA\x4D"
-     "\x03\x00"
-     "\xBB\x45\x03\x00\xBC\x5D\x03\x00\xBD\x45\x03\x00\xBE\x4D\x03\x00\xBF\x45"
-     "\x03\x00"
-     "\xA8\x1D\x02\x00\xA9\x45\x02\x00\xAA\x41\x02\x00\xAB\x41\x02\x00\xAC\x45"
-     "\x02\x00"
-     "\xAD\x4D\x02\x00\xAE\x71\x02\x00\xAF\x71\x02\x00\xBE\xAC\x04\x00\x3E\x69"
-     "\x00\x80"
-     "\x42\x69\x00\x80\x46\x69\x00\x80\x84\xB8\x04\x00\x4A\x69\x00\x80\x4E\x69"
-     "\x00\x80"
-     "\x52\x69\x00\x80\xB8\x49\x01\x00\xB9\x49\x01\x00\xBA\x59\x01\x00\xBB\x51"
-     "\x01\x00"
-     "\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x19\x01\x00\xBF\x15\x01\x00\xB0\xC5"
-     "\x02\x00"
-     "\xB1\xCD\x02\x00\xB2\xC5\x02\x00\xB3\xDD\x02\x00\xB4\xC5\x02\x00\xB5\xCD"
-     "\x02\x00"
-     "\xB6\xC5\x02\x00\xB7\x79\x01\x00\xE1\x30\x0F\x00\xE3\xB4\x0E\x00\xE3\x68"
-     "\x0F\x00"
-     "\xE1\x7C\x0E\x00\x56\x69\x00\x80\x5A\x69\x00\x80\x5E\x69\x00\x80\x62\x69"
-     "\x00\x80"
-     "\x66\x69\x00\x80\x6A\x69\x00\x80\x72\x69\x00\x80\x76\x69\x00\x80\x7A\x69"
-     "\x00\x80"
-     "\xEF\x64\x0E\x00\xEF\xB8\x0F\x00\x7E\x69\x00\x80\x82\x69\x00\x80\xA6\x9D"
-     "\x02\x00"
-     "\xA5\x9D\x02\x00\xBE\x40\x07\x00\xA3\x99\x02\x00\x86\x69\x00\x80\x8A\x69"
-     "\x00\x80"
-     "\x3E\x67\x00\x80\xAF\x59\x02\x00\xAE\x51\x02\x00\xAD\x59\x02\x00\xAC\x41"
-     "\x02\x00"
-     "\xAB\x59\x02\x00\xAA\x51\x02\x00\x86\xD8\x04\x00\x87\x50\x05\x00\x80\x9D"
-     "\x07\x00"
-     "\x81\xA5\x07\x00\x82\xAD\x07\x00\x6E\x69\x00\x80\xEF\xC8\x05\x00\x8E\x69"
-     "\x00\x80"
-     "\x92\x69\x00\x80\x96\x69\x00\x80\xE3\xBC\x04\x00\x9A\x69\x00\x80\xE1\x90"
-     "\x04\x00"
-     "\x9E\x69\x00\x80\xA2\x69\x00\x80\xA6\x69\x00\x80\xAA\x69\x00\x80\xAE\x69"
-     "\x00\x80"
-     "\xB3\xD1\x06\x00\xB2\x69\x00\x80\xB6\x69\x00\x80\xBA\x69\x00\x80\xBE\x69"
-     "\x00\x80"
-     "\xB6\xE5\x06\x00\xB5\xF1\x06\x00\xC2\x69\x00\x80\xBB\xA5\x06\x00\xBA\xA5"
-     "\x06\x00"
-     "\xC6\x69\x00\x80\xCA\x69\x00\x80\xBF\x91\x06\x00\xBE\xA9\x06\x00\xBD\xA9"
-     "\x06\x00"
-     "\xBC\xB5\x06\x00\xA8\xE9\x06\x00\xA9\xE9\x06\x00\xAA\x19\x06\x00\xAB\x15"
-     "\x06\x00"
-     "\xAC\x39\x06\x00\xAD\x25\x06\x00\xAE\x2D\x06\x00\xAF\x25\x06\x00\xCE\x69"
-     "\x00\x80"
-     "\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\xD2\x69\x00\x80\xD6\x69"
-     "\x00\x80"
-     "\xDA\x69\x00\x80\xDE\x69\x00\x80\xB8\xE9\x07\x00\xB9\xE9\x07\x00\xBA\xFD"
-     "\x07\x00"
-     "\xBB\xF5\x07\x00\xBC\xED\x07\x00\xBD\x45\x07\x00\xBE\x4D\x07\x00\xBF\x45"
-     "\x07\x00"
-     "\xB0\x21\x06\x00\xB1\x21\x06\x00\xB2\x21\x06\x00\xB3\x3D\x06\x00\xB4\x2D"
-     "\x06\x00"
-     "\xB5\x15\x06\x00\xB6\x1D\x06\x00\xB7\xD9\x07\x00\xA3\x91\x07\x00\xE2\x69"
-     "\x00\x80"
-     "\x86\x28\x01\x00\x87\x4C\x01\x00\xE6\x69\x00\x80\xA6\xA5\x07\x00\xA5\xB1"
-     "\x07\x00"
-     "\xEA\x69\x00\x80\xAB\xE5\x07\x00\xAA\xE5\x07\x00\xEE\x69\x00\x80\xF2\x69"
-     "\x00\x80"
-     "\xAF\xD1\x07\x00\xAE\xE9\x07\x00\xAD\xE9\x07\x00\xAC\xF5\x07\x00\xF6\x69"
-     "\x00\x80"
-     "\xB3\xF5\x06\x00\xFA\x69\x00\x80\xFE\x69\x00\x80\xB6\x85\x06\x00\x02\x6A"
-     "\x00\x80"
-     "\x06\x6A\x00\x80\xB5\x9D\x06\x00\xBA\x79\x01\x00\xBB\x79\x01\x00\x0A\x6A"
-     "\x00\x80"
-     "\x0E\x6A\x00\x80\xBE\x41\x01\x00\xBF\x41\x01\x00\xBC\x5D\x01\x00\xBD\x51"
-     "\x01\x00"
-     "\xA8\x29\x06\x00\xA9\x29\x06\x00\xAA\x39\x06\x00\xAB\x31\x06\x00\xAC\x51"
-     "\x06\x00"
-     "\xAD\x41\x06\x00\xAE\x41\x06\x00\xAF\x89\x01\x00\x12\x6A\x00\x80\x16\x6A"
-     "\x00\x80"
-     "\x1A\x6A\x00\x80\x1E\x6A\x00\x80\x22\x6A\x00\x80\x26\x6A\x00\x80\x2A\x6A"
-     "\x00\x80"
-     "\x2E\x6A\x00\x80\xB8\x85\x01\x00\xB9\x8D\x01\x00\xBA\x85\x01\x00\xBB\x9D"
-     "\x01\x00"
-     "\xBC\x8D\x01\x00\xBD\xBD\x01\x00\xBE\xB5\x01\x00\xBF\x7D\x01\x00\xB0\xF9"
-     "\x01\x00"
-     "\xB1\xF9\x01\x00\xB2\xCD\x01\x00\xB3\xC1\x01\x00\xB4\xC1\x01\x00\xB5\xC1"
-     "\x01\x00"
-     "\xB6\xC1\x01\x00\xB7\xC1\x01\x00\xA3\xB1\x05\x00\x32\x6A\x00\x80\x36\x6A"
-     "\x00\x80"
-     "\x3A\x6A\x00\x80\x3E\x6A\x00\x80\xA6\xC1\x05\x00\xA5\xD9\x05\x00\x42\x6A"
-     "\x00\x80"
-     "\xAB\x3D\x02\x00\xAA\x3D\x02\x00\x46\x6A\x00\x80\x4A\x6A\x00\x80\xAF\x05"
-     "\x02\x00"
-     "\xAE\x05\x02\x00\xAD\x15\x02\x00\xAC\x19\x02\x00\x80\x5D\x00\x00\x81\x6D"
-     "\x00\x00"
-     "\x82\x65\x00\x00\xB3\x51\x03\x00\xBE\x9C\x03\x00\xB5\x79\x03\x00\xB6\x19"
-     "\x03\x00"
-     "\x52\x6A\x00\x80\x84\xE0\x02\x00\x56\x6A\x00\x80\xBA\x3D\x03\x00\xBB\x35"
-     "\x03\x00"
-     "\xBC\x19\x03\x00\xBD\x19\x03\x00\xBE\xD9\x03\x00\xBF\xD9\x03\x00\xA8\x85"
-     "\x03\x00"
-     "\xA9\x95\x03\x00\xAA\x95\x03\x00\xAB\xA5\x03\x00\xAC\xBD\x03\x00\xAD\xD5"
-     "\x03\x00"
-     "\xAE\xD1\x03\x00\xAF\xD1\x03\x00\x5A\x6A\x00\x80\x5E\x6A\x00\x80\x62\x6A"
-     "\x00\x80"
-     "\xBE\x60\x1C\x00\x87\x20\x03\x00\x86\x94\x1C\x00\x66\x6A\x00\x80\x6A\x6A"
-     "\x00\x80"
-     "\xB8\x71\x03\x00\xB9\x71\x03\x00\xBA\x71\x03\x00\xBB\x71\x03\x00\xBC\xD5"
-     "\x00\x00"
-     "\xBD\xDD\x00\x00\xBE\xD5\x00\x00\xBF\xCD\x00\x00\xB0\xB5\x03\x00\xB1\xBD"
-     "\x03\x00"
-     "\xB2\x81\x03\x00\xB3\x81\x03\x00\xB4\x51\x03\x00\xB5\x51\x03\x00\xB6\x51"
-     "\x03\x00"
-     "\xB7\x51\x03\x00\xEF\xA8\x03\x00\x6E\x6A\x00\x80\x72\x6A\x00\x80\x76\x6A"
-     "\x00\x80"
-     "\x84\x1C\x02\x00\x7A\x6A\x00\x80\x7E\x6A\x00\x80\x82\x6A\x00\x80\x84\x6C"
-     "\x1C\x00"
-     "\x86\x6A\x00\x80\x8A\x6A\x00\x80\x8E\x6A\x00\x80\xE3\x40\x03\x00\x92\x6A"
-     "\x00\x80"
-     "\xE1\x28\x00\x00\x96\x6A\x00\x80\xA3\x5D\x02\x00\x9A\x6A\x00\x80\x9E\x6A"
-     "\x00\x80"
-     "\xA2\x6A\x00\x80\xA6\x6A\x00\x80\xA6\x15\x02\x00\xA5\x75\x02\x00\xAA\x6A"
-     "\x00\x80"
-     "\xAB\x39\x02\x00\xAA\x31\x02\x00\xAE\x6A\x00\x80\xB2\x6A\x00\x80\xAF\xD5"
-     "\x02\x00"
-     "\xAE\xD5\x02\x00\xAD\x15\x02\x00\xAC\x15\x02\x00\xE3\xA0\x06\x00\xB6\x6A"
-     "\x00\x80"
-     "\xE1\x68\x06\x00\xBE\xC4\x1C\x00\x80\xFD\x00\x00\x81\x11\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xBA\x6A\x00\x80\xC2\x6A\x00\x80\xC6\x6A\x00\x80\xCA\x6A\x00\x80\xFE\x68"
-     "\x00\x80"
-     "\xEF\x84\x06\x00\xCE\x6A\x00\x80\xD2\x6A\x00\x80\xD6\x6A\x00\x80\xB3\x61"
-     "\x01\x00"
-     "\xDA\x6A\x00\x80\x86\x68\x1C\x00\x87\x6C\x1D\x00\xDE\x6A\x00\x80\xB6\x6D"
-     "\x01\x00"
-     "\xB5\x71\x01\x00\xE2\x6A\x00\x80\xBB\xED\x00\x00\xBA\xED\x00\x00\xE6\x6A"
-     "\x00\x80"
-     "\xEA\x6A\x00\x80\xBF\xD5\x00\x00\xBE\xE9\x00\x00\xBD\xE1\x00\x00\xBC\xE9"
-     "\x00\x00"
-     "\xBE\x6A\x00\x80\xEF\x50\x1C\x00\xEE\x6A\x00\x80\xF2\x6A\x00\x80\xF6\x6A"
-     "\x00\x80"
-     "\xFA\x6A\x00\x80\xFE\x6A\x00\x80\x02\x6B\x00\x80\x06\x6B\x00\x80\x0A\x6B"
-     "\x00\x80"
-     "\x0E\x6B\x00\x80\x12\x6B\x00\x80\x16\x6B\x00\x80\xE3\x34\x1D\x00\x1A\x6B"
-     "\x00\x80"
-     "\xE1\x90\x1C\x00\xA3\xE1\x1E\x00\x1E\x6B\x00\x80\x22\x6B\x00\x80\x26\x6B"
-     "\x00\x80"
-     "\x2A\x6B\x00\x80\xA6\xED\x1E\x00\xA5\xF1\x1E\x00\x85\x1C\x00\x00\xAB\x6D"
-     "\x1F\x00"
-     "\xAA\x6D\x1F\x00\xBF\x60\x01\x00\x2E\x6B\x00\x80\xAF\x55\x1F\x00\xAE\x69"
-     "\x1F\x00"
-     "\xAD\x61\x1F\x00\xAC\x69\x1F\x00\xA8\xDD\x1E\x00\xA9\xE1\x1E\x00\xAA\x29"
-     "\x1E\x00"
-     "\xAB\x29\x1E\x00\xAC\x39\x1E\x00\xAD\x2D\x1E\x00\xAE\x25\x1E\x00\xAF\x99"
-     "\x1F\x00"
-     "\x80\xE9\x00\x00\x81\xE9\x00\x00\x82\xF9\x00\x00\x32\x6B\x00\x80\x86\x00"
-     "\x04\x00"
-     "\x87\x14\x00\x00\x36\x6B\x00\x80\x3A\x6B\x00\x80\xB8\xF5\x1F\x00\xB9\xF9"
-     "\x1F\x00"
-     "\xBA\x8D\x1F\x00\xBB\x85\x1F\x00\xBC\x9D\x1F\x00\xBD\x85\x1F\x00\xBE\x85"
-     "\x1F\x00"
-     "\xBF\xB5\x1F\x00\xB0\xE9\x1F\x00\xB1\xE9\x1F\x00\xB2\xFD\x1F\x00\xB3\xF5"
-     "\x1F\x00"
-     "\xB4\xED\x1F\x00\xB5\xD1\x1F\x00\xB6\xD1\x1F\x00\xB7\xD1\x1F\x00\xA8\xA1"
-     "\x1E\x00"
-     "\xA9\xA1\x1E\x00\xAA\xA1\x1E\x00\xAB\xB9\x1E\x00\xAC\xA9\x1E\x00\xAD\xA9"
-     "\x1E\x00"
-     "\xAE\xDD\x1E\x00\xAF\xCD\x1E\x00\x3E\x6B\x00\x80\x42\x6B\x00\x80\x46\x6B"
-     "\x00\x80"
-     "\x4A\x6B\x00\x80\x4E\x6B\x00\x80\x52\x6B\x00\x80\x56\x6B\x00\x80\x5A\x6B"
-     "\x00\x80"
-     "\xB8\x5D\x01\x00\xB9\x65\x01\x00\xBA\x6D\x01\x00\xBB\x65\x01\x00\xBC\x7D"
-     "\x01\x00"
-     "\xBD\x65\x01\x00\xBE\x61\x01\x00\xBF\x19\x01\x00\xB0\xBD\x1E\x00\xB1\x8D"
-     "\x1E\x00"
-     "\xB2\x85\x1E\x00\xB3\x65\x01\x00\xB4\x7D\x01\x00\xB5\x65\x01\x00\xB6\x6D"
-     "\x01\x00"
-     "\xB7\x65\x01\x00\xB3\x29\x1E\x00\x5E\x6B\x00\x80\x62\x6B\x00\x80\x66\x6B"
-     "\x00\x80"
-     "\x6A\x6B\x00\x80\xB6\x29\x1E\x00\xB5\x21\x1E\x00\x6E\x6B\x00\x80\xBB\x49"
-     "\x1E\x00"
-     "\xBA\x79\x1E\x00\x72\x6B\x00\x80\x76\x6B\x00\x80\xBF\xA1\x01\x00\xBE\xB9"
-     "\x01\x00"
-     "\xBD\xB1\x01\x00\xBC\x51\x1E\x00\x7A\x6B\x00\x80\xA3\x6D\x1E\x00\x7E\x6B"
-     "\x00\x80"
-     "\x82\x6B\x00\x80\xA6\x6D\x1E\x00\x86\x6B\x00\x80\x8A\x6B\x00\x80\xA5\x65"
-     "\x1E\x00"
-     "\xAA\x3D\x1E\x00\xAB\x0D\x1E\x00\x8E\x6B\x00\x80\x92\x6B\x00\x80\xAE\xFD"
-     "\x01\x00"
-     "\xAF\xE5\x01\x00\xAC\x15\x1E\x00\xAD\xF5\x01\x00\x8F\x09\x2A\x00\xE1\x4C"
-     "\x00\x00"
-     "\x96\x6B\x00\x80\xE3\x58\x0A\x00\x9A\x6B\x00\x80\x84\xD8\x04\x00\xBE\xD0"
-     "\x04\x00"
-     "\xA2\x6B\x00\x80\xBE\xE8\x05\x00\xA6\x6B\x00\x80\x87\x88\x03\x00\x86\x0C"
-     "\x04\x00"
-     "\xAA\x6B\x00\x80\xAE\x6B\x00\x80\xB2\x6B\x00\x80\xEF\x4C\x7B\x00\x9F\xED"
-     "\x1D\x00"
-     "\x9E\x55\x1B\x00\x9D\x41\x1A\x00\x9C\x49\x1A\x00\x9B\x8D\x19\x00\x9A\x45"
-     "\x16\x00"
-     "\x99\x5D\x16\x00\x98\x9D\x15\x00\x97\x0D\x12\x00\x96\x61\x12\x00\x95\x9D"
-     "\x11\x00"
-     "\x94\x1D\x2E\x00\x93\x75\x2E\x00\x92\x8D\x2D\x00\x91\xD5\x2A\x00\x90\x71"
-     "\x2A\x00"
-     "\xB0\x29\x76\x00\xB1\x01\x78\x00\xB2\xA9\x7A\x00\xB3\xB1\x7A\x00\xB4\x01"
-     "\x7C\x00"
-     "\xB5\x91\x7E\x00\xB6\x6B\x00\x80\xBA\x6B\x00\x80\x80\x4D\x03\x00\x81\xAD"
-     "\x00\x00"
-     "\x82\xB9\x00\x00\xBE\x6B\x00\x80\xA3\xB1\x02\x00\xA2\x0D\x01\x00\xA1\x05"
-     "\x1E\x00"
-     "\xA0\xB1\x1E\x00\xC2\x6B\x00\x80\xA1\xA9\x02\x00\xA2\xCD\x06\x00\xA3\xC5"
-     "\x06\x00"
-     "\xA4\x01\x08\x00\xA5\xD1\x0A\x00\xA6\xD1\x0A\x00\xA7\xB9\x0B\x00\xA8\xE1"
-     "\x0E\x00"
-     "\xA9\xE1\x0E\x00\xAA\xF9\x0F\x00\xAB\x01\x70\x00\xAC\xD1\x72\x00\xAD\x2D"
-     "\x72\x00"
-     "\xAE\x01\x74\x00\xAF\xDD\x76\x00\x82\x09\x2B\x00\x83\x09\x2B\x00\xC6\x6B"
-     "\x00\x80"
-     "\xCA\x6B\x00\x80\x86\x15\x2F\x00\x87\x65\x2F\x00\x84\xB5\x2B\x00\x85\xE5"
-     "\x2C\x00"
-     "\x8A\xD5\x10\x00\x8B\x21\x13\x00\xCE\x6B\x00\x80\xD2\x6B\x00\x80\x8E\x35"
-     "\x17\x00"
-     "\x8F\x25\x17\x00\x8C\x55\x13\x00\x8D\xFD\x14\x00\x92\x9D\x18\x00\x93\x41"
-     "\x1B\x00"
-     "\x84\x84\x04\x00\x84\x08\x07\x00\x96\xFD\x1C\x00\x97\x59\x1F\x00\x94\x51"
-     "\x1B\x00"
-     "\x95\x4D\x1A\x00\x9A\xDD\x1F\x00\x9B\xDD\x00\x00\xD6\x6B\x00\x80\xDA\x6B"
-     "\x00\x80"
-     "\xDE\x6B\x00\x80\xE2\x6B\x00\x80\x9C\x79\x03\x00\xE6\x6B\x00\x80\xEA\x6B"
-     "\x00\x80"
-     "\xE1\x54\x1E\x00\xEE\x6B\x00\x80\xE3\xBC\x1E\x00\xF2\x6B\x00\x80\xF6\x6B"
-     "\x00\x80"
-     "\xFA\x6B\x00\x80\xFE\x6B\x00\x80\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\x02\x6C\x00\x80\x06\x6C\x00\x80\x0A\x6C\x00\x80\x4E\x6A\x00\x80\xEF\xC8"
-     "\x1E\x00"
-     "\x87\x6C\x05\x00\x86\xE4\x04\x00\x12\x6C\x00\x80\x16\x6C\x00\x80\xEF\x30"
-     "\x03\x00"
-     "\x1A\x6C\x00\x80\x1E\x6C\x00\x80\x22\x6C\x00\x80\xE3\xE4\x03\x00\x26\x6C"
-     "\x00\x80"
-     "\xE1\x40\x01\x00\x2A\x6C\x00\x80\x2E\x6C\x00\x80\x32\x6C\x00\x80\x36\x6C"
-     "\x00\x80"
-     "\x3A\x6C\x00\x80\xB3\xBD\x04\x00\x9E\x6B\x00\x80\x0E\x6C\x00\x80\x3E\x6C"
-     "\x00\x80"
-     "\x42\x6C\x00\x80\xB6\x3D\x05\x00\xB5\x3D\x05\x00\x46\x6C\x00\x80\xBB\x19"
-     "\x05\x00"
-     "\xBA\x11\x05\x00\x4A\x6C\x00\x80\x4E\x6C\x00\x80\xBF\xBD\x05\x00\xBE\xBD"
-     "\x05\x00"
-     "\xBD\xBD\x05\x00\xBC\x01\x05\x00\x52\x6C\x00\x80\xA3\xF9\x04\x00\x56\x6C"
-     "\x00\x80"
-     "\x5A\x6C\x00\x80\xA6\x79\x05\x00\x5E\x6C\x00\x80\x62\x6C\x00\x80\xA5\x79"
-     "\x05\x00"
-     "\xAA\x55\x05\x00\xAB\x5D\x05\x00\x66\x6C\x00\x80\x6A\x6C\x00\x80\xAE\xF9"
-     "\x05\x00"
-     "\xAF\xF9\x05\x00\xAC\x45\x05\x00\xAD\xF9\x05\x00\xA8\x9D\x06\x00\xA9\xA1"
-     "\x06\x00"
-     "\xAA\xB9\x06\x00\xAB\xB9\x06\x00\xAC\xA9\x06\x00\xAD\x0D\x06\x00\xAE\x05"
-     "\x06\x00"
-     "\xAF\x31\x06\x00\x6E\x6C\x00\x80\x82\x19\x00\x00\x81\x11\x00\x00\x80\x19"
-     "\x00\x00"
-     "\xBE\xEC\x00\x00\x72\x6C\x00\x80\x76\x6C\x00\x80\x7A\x6C\x00\x80\xB8\xB5"
-     "\x07\x00"
-     "\xB9\xB9\x07\x00\xBA\x49\x07\x00\xBB\x45\x07\x00\xBC\x41\x07\x00\xBD\x41"
-     "\x07\x00"
-     "\xBE\x41\x07\x00\xBF\x41\x07\x00\xB0\x21\x06\x00\xB1\xED\x07\x00\xB2\xE5"
-     "\x07\x00"
-     "\xB3\xC5\x07\x00\xB4\xDD\x07\x00\xB5\x9D\x07\x00\xB6\x95\x07\x00\xB7\x8D"
-     "\x07\x00"
-     "\xB3\x31\x07\x00\x82\x6C\x00\x80\x86\xA8\x00\x00\x87\x88\x01\x00\x86\x6C"
-     "\x00\x80"
-     "\xB6\x8D\x07\x00\xB5\x21\x07\x00\x8A\x6C\x00\x80\xBB\x29\x06\x00\xBA\xE5"
-     "\x07\x00"
-     "\x8E\x6C\x00\x80\x92\x6C\x00\x80\xBF\x59\x06\x00\xBE\x19\x06\x00\xBD\x11"
-     "\x06\x00"
-     "\xBC\x31\x06\x00\x96\x6C\x00\x80\xA3\x75\x07\x00\x9A\x6C\x00\x80\x9E\x6C"
-     "\x00\x80"
-     "\xA6\xC9\x07\x00\xA2\x6C\x00\x80\xA6\x6C\x00\x80\xA5\x65\x07\x00\xAA\xA1"
-     "\x07\x00"
-     "\xAB\x6D\x06\x00\xAA\x6C\x00\x80\xAE\x6C\x00\x80\xAE\x5D\x06\x00\xAF\x1D"
-     "\x06\x00"
-     "\xAC\x75\x06\x00\xAD\x55\x06\x00\xA8\x49\x06\x00\xA9\x49\x06\x00\xAA\x5D"
-     "\x06\x00"
-     "\xAB\x55\x06\x00\xAC\x4D\x06\x00\xAD\x71\x06\x00\xAE\x75\x06\x00\xAF\x6D"
-     "\x06\x00"
-     "\x84\xAC\x00\x00\xB2\x6C\x00\x80\xB6\x6C\x00\x80\xBA\x6C\x00\x80\xBE\x6C"
-     "\x00\x80"
-     "\xC2\x6C\x00\x80\xC6\x6C\x00\x80\xCA\x6C\x00\x80\xB8\x81\x01\x00\xB9\x8D"
-     "\x01\x00"
-     "\xBA\x85\x01\x00\xBB\x99\x01\x00\xBC\x89\x01\x00\xBD\xBD\x01\x00\xBE\xB5"
-     "\x01\x00"
-     "\xBF\x7D\x01\x00\xB0\x15\x06\x00\xB1\x1D\x06\x00\xB2\x15\x06\x00\xB3\xE1"
-     "\x01\x00"
-     "\xB4\xE1\x01\x00\xB5\xE9\x01\x00\xB6\xD1\x01\x00\xB7\xD1\x01\x00\xAC\x25"
-     "\x07\x00"
-     "\xAD\x2D\x07\x00\xAE\x25\x07\x00\xAF\x3D\x07\x00\xA8\xBD\x06\x00\xA9\xAD"
-     "\x06\x00"
-     "\xAA\xA5\x06\x00\xAB\x3D\x07\x00\xBE\xAC\x0C\x00\xCE\x6C\x00\x80\xD2\x6C"
-     "\x00\x80"
-     "\xD6\x6C\x00\x80\xDA\x6C\x00\x80\xDE\x6C\x00\x80\xE2\x6C\x00\x80\xE6\x6C"
-     "\x00\x80"
-     "\xBC\x8D\x07\x00\xBD\x91\x07\x00\xBE\x91\x07\x00\xBF\xA5\x07\x00\xB8\xF9"
-     "\x07\x00"
-     "\xB9\xF9\x07\x00\xBA\x99\x07\x00\xBB\x95\x07\x00\xB4\xE9\x07\x00\xB5\xE9"
-     "\x07\x00"
-     "\xB6\xF9\x07\x00\xB7\xF1\x07\x00\xB0\x45\x07\x00\xB1\x4D\x07\x00\xB2\x45"
-     "\x07\x00"
-     "\xB3\xF9\x07\x00\x81\xBD\x00\x00\x80\xBD\x00\x00\xB3\x5D\x05\x00\x82\x19"
-     "\x00\x00"
-     "\xB5\x4D\x05\x00\xEA\x6C\x00\x80\xF2\x6C\x00\x80\xB6\x79\x05\x00\xF6\x6C"
-     "\x00\x80"
-     "\xFA\x6C\x00\x80\xBB\x51\x05\x00\xBA\x59\x05\x00\xBD\xBD\x02\x00\xBC\x49"
-     "\x05\x00"
-     "\xBF\xA5\x02\x00\xBE\xAD\x02\x00\xFE\x6C\x00\x80\xA6\x21\x05\x00\xA5\x15"
-     "\x05\x00"
-     "\x84\x4C\x0C\x00\xA3\x05\x05\x00\x02\x6D\x00\x80\x86\xF8\x0C\x00\x87\x9C"
-     "\x0C\x00"
-     "\xAF\xFD\x02\x00\xAE\xF5\x02\x00\xAD\xE5\x02\x00\xAC\x11\x05\x00\xAB\x09"
-     "\x05\x00"
-     "\xAA\x01\x05\x00\x06\x6D\x00\x80\x0A\x6D\x00\x80\xA8\x05\x02\x00\xA9\x15"
-     "\x02\x00"
-     "\xAA\x15\x02\x00\xAB\x25\x02\x00\xAC\x3D\x02\x00\xAD\x55\x02\x00\xAE\x51"
-     "\x02\x00"
-     "\xAF\x51\x02\x00\xBE\xAC\x0C\x00\x0E\x6D\x00\x80\x12\x6D\x00\x80\x16\x6D"
-     "\x00\x80"
-     "\x1A\x6D\x00\x80\x1E\x6D\x00\x80\x22\x6D\x00\x80\x26\x6D\x00\x80\xB8\xF5"
-     "\x02\x00"
-     "\xB9\xFD\x02\x00\xBA\xF5\x02\x00\xBB\x69\x01\x00\xBC\x79\x01\x00\xBD\x79"
-     "\x01\x00"
-     "\xBE\x69\x01\x00\xBF\x69\x01\x00\xB0\x35\x02\x00\xB1\x3D\x02\x00\xB2\x01"
-     "\x02\x00"
-     "\xB3\x01\x02\x00\xB4\xD5\x02\x00\xB5\xDD\x02\x00\xB6\xD5\x02\x00\xB7\xCD"
-     "\x02\x00"
-     "\xE1\xB4\x07\x00\xE3\xD8\x06\x00\xE3\xD0\x06\x00\xE1\x50\x06\x00\x2A\x6D"
-     "\x00\x80"
-     "\x2E\x6D\x00\x80\x32\x6D\x00\x80\x36\x6D\x00\x80\x3A\x6D\x00\x80\x3E\x6D"
-     "\x00\x80"
-     "\x42\x6D\x00\x80\x46\x6D\x00\x80\xEC\x00\x00\x00\xEF\xFC\x01\x00\xEF\x60"
-     "\x07\x00"
-     "\x4A\x6D\x00\x80\xB3\xDD\x02\x00\x4E\x6D\x00\x80\x86\xE8\x0C\x00\x87\x8C"
-     "\x0C\x00"
-     "\x52\x6D\x00\x80\xB6\x95\x02\x00\xB5\xF5\x02\x00\x56\x6D\x00\x80\xBB\xB9"
-     "\x02\x00"
-     "\xBA\xB1\x02\x00\x5A\x6D\x00\x80\x62\x6D\x00\x80\xBF\x59\x02\x00\xBE\x51"
-     "\x02\x00"
-     "\xBD\x95\x02\x00\xBC\x95\x02\x00\xA3\x1D\x0D\x00\xEE\x6C\x00\x80\x5E\x6D"
-     "\x00\x80"
-     "\x66\x6D\x00\x80\x6A\x6D\x00\x80\xA6\x55\x0D\x00\xA5\x35\x0D\x00\x6E\x6D"
-     "\x00\x80"
-     "\xAB\x79\x0D\x00\xAA\x71\x0D\x00\x72\x6D\x00\x80\x76\x6D\x00\x80\xAF\x99"
-     "\x0D\x00"
-     "\xAE\x91\x0D\x00\xAD\x55\x0D\x00\xAC\x55\x0D\x00\x80\x91\x0F\x00\x81\x99"
-     "\x0F\x00"
-     "\x82\xA1\x0F\x00\xB3\xE9\x0E\x00\x7A\x6D\x00\x80\xB5\x81\x0E\x00\xB6\xA1"
-     "\x0E\x00"
-     "\x7E\x6D\x00\x80\x82\x6D\x00\x80\x86\x6D\x00\x80\xBA\x41\x0F\x00\xBB\x41"
-     "\x0F\x00"
-     "\xBC\x41\x0F\x00\xBD\x41\x0F\x00\xBE\x41\x0F\x00\xBF\x41\x0F\x00\xA8\x19"
-     "\x0E\x00"
-     "\xA9\x25\x0E\x00\xAA\x2D\x0E\x00\xAB\x4D\x0E\x00\xAC\x45\x0E\x00\xAD\x6D"
-     "\x0E\x00"
-     "\xAE\x65\x0E\x00\xAF\xB9\x0F\x00\x8A\x6D\x00\x80\x8E\x6D\x00\x80\x92\x6D"
-     "\x00\x80"
-     "\x96\x6D\x00\x80\x9A\x6D\x00\x80\x9E\x6D\x00\x80\xA2\x6D\x00\x80\xA6\x6D"
-     "\x00\x80"
-     "\xB8\x59\x0F\x00\xB9\x59\x0F\x00\xBA\x69\x0F\x00\xBB\x69\x0F\x00\xBC\x79"
-     "\x0F\x00"
-     "\xBD\x79\x0F\x00\xBE\x65\x0F\x00\xBF\x19\x0F\x00\xB0\xC9\x0F\x00\xB1\xC9"
-     "\x0F\x00"
-     "\xB2\xD9\x0F\x00\xB3\xD9\x0F\x00\xB4\xC9\x0F\x00\xB5\xC9\x0F\x00\xB6\x69"
-     "\x0F\x00"
-     "\xB7\x69\x0F\x00\xA3\xA5\x0F\x00\xAA\x6D\x00\x80\xAE\x6D\x00\x80\xB2\x6D"
-     "\x00\x80"
-     "\xB6\x6D\x00\x80\xA6\xED\x0F\x00\xA5\xCD\x0F\x00\xBA\x6D\x00\x80\xAB\x0D"
-     "\x0E\x00"
-     "\xAA\x0D\x0E\x00\xBE\x6D\x00\x80\xC2\x6D\x00\x80\xAF\x0D\x0E\x00\xAE\x0D"
-     "\x0E\x00"
-     "\xAD\x0D\x0E\x00\xAC\x0D\x0E\x00\x80\x6D\x00\x00\x81\x09\x00\x00\x82\x19"
-     "\x00\x00"
-     "\xC6\x6D\x00\x80\xCA\x6D\x00\x80\xE1\x44\x0E\x00\x84\x94\x01\x00\xE3\xEC"
-     "\x0E\x00"
-     "\xBE\xA8\x01\x00\xCE\x6D\x00\x80\xD2\x6D\x00\x80\xEF\xB4\x0D\x00\x86\xD0"
-     "\x00\x00"
-     "\x87\xFC\x01\x00\xD6\x6D\x00\x80\xDA\x6D\x00\x80\xA8\x6D\x0E\x00\xA9\x89"
-     "\x01\x00"
-     "\xAA\x99\x01\x00\xAB\x91\x01\x00\xAC\xB9\x01\x00\xAD\xB9\x01\x00\xAE\xC9"
-     "\x01\x00"
-     "\xAF\xC1\x01\x00\x84\xA0\x00\x00\xDE\x6D\x00\x80\xE2\x6D\x00\x80\xE6\x6D"
-     "\x00\x80"
-     "\xEA\x6D\x00\x80\xEE\x6D\x00\x80\xF2\x6D\x00\x80\xF6\x6D\x00\x80\xB8\x79"
-     "\x01\x00"
-     "\xB9\x79\x01\x00\xBA\xCD\x00\x00\xBB\xC5\x00\x00\xBC\xDD\x00\x00\xBD\xC5"
-     "\x00\x00"
-     "\xBE\xC5\x00\x00\xBF\xF5\x00\x00\xB0\x81\x01\x00\xB1\x81\x01\x00\xB2\x49"
-     "\x01\x00"
-     "\xB3\x49\x01\x00\xB4\x59\x01\x00\xB5\x59\x01\x00\xB6\x49\x01\x00\xB7\x49"
-     "\x01\x00"
-     "\xB3\x11\x02\x00\xFA\x6D\x00\x80\xFE\x6D\x00\x80\x02\x6E\x00\x80\x06\x6E"
-     "\x00\x80"
-     "\xB6\x39\x02\x00\xB5\x31\x02\x00\x0A\x6E\x00\x80\xBB\x45\x02\x00\xBA\x45"
-     "\x02\x00"
-     "\x0E\x6E\x00\x80\x12\x6E\x00\x80\xBF\x9D\x02\x00\xBE\x9D\x02\x00\xBD\x9D"
-     "\x02\x00"
-     "\xBC\x9D\x02\x00\x85\xFC\x75\x00\xA3\x55\x02\x00\x16\x6E\x00\x80\x1A\x6E"
-     "\x00\x80"
-     "\xA6\x7D\x02\x00\x1E\x6E\x00\x80\x22\x6E\x00\x80\xA5\x75\x02\x00\xAA\x01"
-     "\x02\x00"
-     "\xAB\x01\x02\x00\x26\x6E\x00\x80\x2A\x6E\x00\x80\xAE\xD9\x02\x00\xAF\xD9"
-     "\x02\x00"
-     "\xAC\xD9\x02\x00\xAD\xD9\x02\x00\x82\x19\x00\x00\x2E\x6E\x00\x80\x80\x19"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x32\x6E\x00\x80\x36\x6E\x00\x80\x3A\x6E\x00\x80\x42\x6E"
-     "\x00\x80"
-     "\xBE\x00\x04\x00\x46\x6E\x00\x80\x4A\x6E\x00\x80\x4E\x6E\x00\x80\x52\x6E"
-     "\x00\x80"
-     "\x56\x6E\x00\x80\x5A\x6E\x00\x80\x5E\x6E\x00\x80\x86\xE0\x04\x00\x87\x54"
-     "\x03\x00"
-     "\x62\x6E\x00\x80\x66\x6E\x00\x80\x7E\x6C\x00\x80\x6A\x6E\x00\x80\xEF\x80"
-     "\x02\x00"
-     "\x6E\x6E\x00\x80\xE1\x88\x01\x00\x72\x6E\x00\x80\xE3\x58\x01\x00\x76\x6E"
-     "\x00\x80"
-     "\x7A\x6E\x00\x80\x7E\x6E\x00\x80\x82\x6E\x00\x80\x86\x6E\x00\x80\xA8\x69"
-     "\x02\x00"
-     "\xA9\x7D\x02\x00\xAA\x75\x02\x00\xAB\xA9\x02\x00\xAC\xB9\x02\x00\xAD\xB9"
-     "\x02\x00"
-     "\xAE\xA9\x02\x00\xAF\xA9\x02\x00\x84\xAC\x05\x00\x8A\x6E\x00\x80\x8E\x6E"
-     "\x00\x80"
-     "\x92\x6E\x00\x80\x96\x6E\x00\x80\x9A\x6E\x00\x80\x9E\x6E\x00\x80\xA2\x6E"
-     "\x00\x80"
-     "\xB8\x6D\x01\x00\xB9\x0D\x01\x00\xBA\x05\x01\x00\xBB\x19\x01\x00\xBC\x09"
-     "\x01\x00"
-     "\xBD\x3D\x01\x00\xBE\x35\x01\x00\xBF\xD9\x01\x00\xB0\xD9\x02\x00\xB1\xD9"
-     "\x02\x00"
-     "\xB2\x6D\x01\x00\xB3\x65\x01\x00\xB4\x7D\x01\x00\xB5\x65\x01\x00\xB6\x65"
-     "\x01\x00"
-     "\xB7\x55\x01\x00\xE1\x40\x0F\x00\xE3\xCC\x00\x00\xE3\x54\x0F\x00\xE1\x5C"
-     "\x0E\x00"
-     "\x80\xAD\x00\x00\x81\x09\x00\x00\x82\x39\x00\x00\xA6\x6E\x00\x80\xAA\x6E"
-     "\x00\x80"
-     "\xBE\x28\x04\x00\xB2\x6E\x00\x80\xB6\x6E\x00\x80\xBE\x3C\x07\x00\xEF\x2C"
-     "\x00\x00"
-     "\xEF\x60\x0E\x00\xBA\x6E\x00\x80\x86\xE0\x04\x00\x87\x74\x05\x00\xBE\x6E"
-     "\x00\x80"
-     "\xB3\xB1\x02\x00\xC2\x6E\x00\x80\xB5\x7D\x02\x00\xB6\x75\x02\x00\xC6\x6E"
-     "\x00\x80"
-     "\xCA\x6E\x00\x80\xCE\x6E\x00\x80\xBA\x51\x02\x00\xBB\x51\x02\x00\xBC\xF5"
-     "\x01\x00"
-     "\xBD\xFD\x01\x00\xBE\xF5\x01\x00\xBF\xED\x01\x00\x3E\x6E\x00\x80\xAE\x6E"
-     "\x00\x80"
-     "\xAA\xB1\x05\x00\xAB\xB1\x05\x00\xAC\x15\x06\x00\xAD\x1D\x06\x00\xAE\x15"
-     "\x06\x00"
-     "\xAF\x0D\x06\x00\xD2\x6E\x00\x80\xD6\x6E\x00\x80\xDA\x6E\x00\x80\xA3\x51"
-     "\x05\x00"
-     "\xDE\x6E\x00\x80\xA5\x9D\x05\x00\xA6\x95\x05\x00\xE2\x6E\x00\x80\xB3\xE9"
-     "\x06\x00"
-     "\xE6\x6E\x00\x80\xEA\x6E\x00\x80\xEE\x6E\x00\x80\xF2\x6E\x00\x80\xB6\xE1"
-     "\x06\x00"
-     "\xB5\xE9\x06\x00\xF6\x6E\x00\x80\xBB\xAD\x06\x00\xBA\xAD\x06\x00\xFA\x6E"
-     "\x00\x80"
-     "\xFE\x6E\x00\x80\xBF\x9D\x06\x00\xBE\x91\x06\x00\xBD\xA5\x06\x00\xBC\xA9"
-     "\x06\x00"
-     "\xA8\x1D\x06\x00\xA9\x21\x06\x00\xAA\x21\x06\x00\xAB\x21\x06\x00\xAC\x21"
-     "\x06\x00"
-     "\xAD\x21\x06\x00\xAE\x21\x06\x00\xAF\x21\x06\x00\x02\x6F\x00\x80\x06\x6F"
-     "\x00\x80"
-     "\x0A\x6F\x00\x80\x0E\x6F\x00\x80\x80\x0D\x00\x00\x81\xB1\x01\x00\x82\xB1"
-     "\x01\x00"
-     "\x12\x6F\x00\x80\xB8\x39\x06\x00\xB9\x39\x06\x00\xBA\xED\x07\x00\xBB\xE5"
-     "\x07\x00"
-     "\xBC\xFD\x07\x00\xBD\xE5\x07\x00\xBE\xED\x07\x00\xBF\xE5\x07\x00\xB0\x29"
-     "\x06\x00"
-     "\xB1\x29\x06\x00\xB2\x3D\x06\x00\xB3\x31\x06\x00\xB4\x15\x06\x00\xB5\x19"
-     "\x06\x00"
-     "\xB6\x09\x06\x00\xB7\x09\x06\x00\xA3\xA9\x07\x00\x16\x6F\x00\x80\x1A\x6F"
-     "\x00\x80"
-     "\x84\x84\x02\x00\x1E\x6F\x00\x80\xA6\xA1\x07\x00\xA5\xA9\x07\x00\x22\x6F"
-     "\x00\x80"
-     "\xAB\xED\x07\x00\xAA\xED\x07\x00\x86\x88\x00\x00\x87\x78\x01\x00\xAF\xDD"
-     "\x07\x00"
-     "\xAE\xD1\x07\x00\xAD\xE5\x07\x00\xAC\xE9\x07\x00\x26\x6F\x00\x80\xB3\xF9"
-     "\x06\x00"
-     "\x2A\x6F\x00\x80\x2E\x6F\x00\x80\xB6\x8D\x06\x00\x32\x6F\x00\x80\x36\x6F"
-     "\x00\x80"
-     "\xB5\x8D\x06\x00\xBA\x41\x01\x00\xBB\x41\x01\x00\x3A\x6F\x00\x80\x3E\x6F"
-     "\x00\x80"
-     "\xBE\x4D\x01\x00\xBF\x31\x01\x00\xBC\x51\x01\x00\xBD\x4D\x01\x00\xA8\x29"
-     "\x06\x00"
-     "\xA9\x29\x06\x00\xAA\x35\x06\x00\xAB\x45\x06\x00\xAC\x41\x06\x00\xAD\x49"
-     "\x06\x00"
-     "\xAE\x79\x06\x00\xAF\x79\x06\x00\x84\xAC\x01\x00\x42\x6F\x00\x80\x46\x6F"
-     "\x00\x80"
-     "\x4A\x6F\x00\x80\x4E\x6F\x00\x80\x52\x6F\x00\x80\x56\x6F\x00\x80\x5A\x6F"
-     "\x00\x80"
-     "\xB8\x91\x01\x00\xB9\x99\x01\x00\xBA\xA1\x01\x00\xBB\xA1\x01\x00\xBC\x71"
-     "\x01\x00"
-     "\xBD\x71\x01\x00\xBE\x71\x01\x00\xBF\x71\x01\x00\xB0\xC1\x01\x00\xB1\xC1"
-     "\x01\x00"
-     "\xB2\xC1\x01\x00\xB3\xDD\x01\x00\xB4\xCD\x01\x00\xB5\xB1\x01\x00\xB6\xB1"
-     "\x01\x00"
-     "\xB7\xB1\x01\x00\xA3\xBD\x05\x00\x5E\x6F\x00\x80\x62\x6F\x00\x80\x66\x6F"
-     "\x00\x80"
-     "\x6A\x6F\x00\x80\xA6\xC9\x05\x00\xA5\xC9\x05\x00\x6E\x6F\x00\x80\xAB\x05"
-     "\x02\x00"
-     "\xAA\x05\x02\x00\x72\x6F\x00\x80\x76\x6F\x00\x80\xAF\x75\x02\x00\xAE\x09"
-     "\x02\x00"
-     "\xAD\x09\x02\x00\xAC\x15\x02\x00\x7A\x6F\x00\x80\x7E\x6F\x00\x80\x82\x6F"
-     "\x00\x80"
-     "\x86\x6F\x00\x80\x81\x1D\x00\x00\x80\x1D\x00\x00\x8A\x6F\x00\x80\x82\x05"
-     "\x00\x00"
-     "\xBE\x24\x03\x00\x92\x6F\x00\x80\x86\xC8\x7C\x00\x87\x5C\x03\x00\x96\x6F"
-     "\x00\x80"
-     "\x9A\x6F\x00\x80\x9E\x6F\x00\x80\xA2\x6F\x00\x80\xA6\x6F\x00\x80\xAA\x6F"
-     "\x00\x80"
-     "\xAE\x6F\x00\x80\x84\xC0\x7C\x00\xB2\x6F\x00\x80\xE1\xA4\x01\x00\xB6\x6F"
-     "\x00\x80"
-     "\xE3\x2C\x01\x00\xBE\xE8\x7C\x00\xBA\x6F\x00\x80\xBE\x6F\x00\x80\xEF\x34"
-     "\x02\x00"
-     "\xAE\x00\x00\x00\xC2\x6F\x00\x80\xC6\x6F\x00\x80\xCA\x6F\x00\x80\xB3\x89"
-     "\x03\x00"
-     "\xCE\x6F\x00\x80\xD2\x6F\x00\x80\xD6\x6F\x00\x80\xDA\x6F\x00\x80\xB6\x8D"
-     "\x03\x00"
-     "\xB5\x8D\x03\x00\xDE\x6F\x00\x80\xBB\x49\x03\x00\xBA\x41\x03\x00\xE2\x6F"
-     "\x00\x80"
-     "\xE6\x6F\x00\x80\xBF\x49\x03\x00\xBE\x41\x03\x00\xBD\x49\x03\x00\xBC\x51"
-     "\x03\x00"
-     "\xA8\x1D\x02\x00\xA9\x45\x02\x00\xAA\x41\x02\x00\xAB\x41\x02\x00\xAC\x45"
-     "\x02\x00"
-     "\xAD\x4D\x02\x00\xAE\x71\x02\x00\xAF\x71\x02\x00\xBE\xA0\x7F\x00\xEA\x6F"
-     "\x00\x80"
-     "\xEE\x6F\x00\x80\xF2\x6F\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xA9"
-     "\x00\x00"
-     "\xF6\x6F\x00\x80\xB8\x49\x01\x00\xB9\x49\x01\x00\xBA\x59\x01\x00\xBB\x51"
-     "\x01\x00"
-     "\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x19\x01\x00\xBF\x19\x01\x00\xB0\xC5"
-     "\x02\x00"
-     "\xB1\xCD\x02\x00\xB2\xC5\x02\x00\xB3\xDD\x02\x00\xB4\xC5\x02\x00\xB5\xCD"
-     "\x02\x00"
-     "\xB6\xC5\x02\x00\xB7\x79\x01\x00\xE1\x58\x06\x00\xE1\x5C\x07\x00\xE3\x94"
-     "\x00\x00"
-     "\xE3\xBC\x06\x00\xFA\x6F\x00\x80\x02\x70\x00\x80\x86\xB8\x7C\x00\x87\x54"
-     "\x7D\x00"
-     "\x06\x70\x00\x80\x0A\x70\x00\x80\x0E\x70\x00\x80\x12\x70\x00\x80\xEE\x00"
-     "\x00\x00"
-     "\x16\x70\x00\x80\xEF\xF0\x06\x00\xEF\xE0\x06\x00\x1A\x70\x00\x80\x1E\x70"
-     "\x00\x80"
-     "\xA3\x81\x02\x00\x22\x70\x00\x80\xA5\x85\x02\x00\x26\x70\x00\x80\x2A\x70"
-     "\x00\x80"
-     "\xA6\x85\x02\x00\x2E\x70\x00\x80\x32\x70\x00\x80\xAB\x41\x02\x00\xAA\x49"
-     "\x02\x00"
-     "\xAD\x41\x02\x00\xAC\x59\x02\x00\xAF\x41\x02\x00\xAE\x49\x02\x00\xA8\x65"
-     "\x7E\x00"
-     "\xA9\x75\x7E\x00\xAA\x7D\x7E\x00\xAB\x75\x7E\x00\xAC\x6D\x7E\x00\xAD\x95"
-     "\x7E\x00"
-     "\xAE\x9D\x7E\x00\xAF\x95\x7E\x00\xFE\x6F\x00\x80\x36\x70\x00\x80\x3A\x70"
-     "\x00\x80"
-     "\x3E\x70\x00\x80\x42\x70\x00\x80\x46\x70\x00\x80\x4A\x70\x00\x80\x4E\x70"
-     "\x00\x80"
-     "\xB8\x89\x7E\x00\xB9\x89\x7E\x00\xBA\x99\x7E\x00\xBB\x91\x7E\x00\xBC\xB9"
-     "\x7E\x00"
-     "\xBD\xB9\x7E\x00\xBE\x7D\x7F\x00\xBF\x75\x7F\x00\xB0\xF1\x7E\x00\xB1\xFD"
-     "\x7E\x00"
-     "\xB2\xCD\x7E\x00\xB3\xC5\x7E\x00\xB4\xC5\x7E\x00\xB5\xCD\x7E\x00\xB6\xC5"
-     "\x7E\x00"
-     "\xB7\xB9\x7E\x00\xB3\xF5\x7E\x00\x52\x70\x00\x80\x56\x70\x00\x80\x5A\x70"
-     "\x00\x80"
-     "\x5E\x70\x00\x80\xB6\x15\x7E\x00\xB5\xE5\x7E\x00\x62\x70\x00\x80\xBB\x39"
-     "\x7E\x00"
-     "\xBA\x35\x7E\x00\x66\x70\x00\x80\x6A\x70\x00\x80\xBF\x09\x7E\x00\xBE\x01"
-     "\x7E\x00"
-     "\xBD\x11\x7E\x00\xBC\x19\x7E\x00\x82\x6D\x00\x00\xA3\xB1\x7E\x00\x80\x55"
-     "\x00\x00"
-     "\x81\x65\x00\x00\xA6\x51\x7E\x00\x84\x18\x03\x00\x6E\x70\x00\x80\xA5\xA1"
-     "\x7E\x00"
-     "\xAA\x71\x7E\x00\xAB\x7D\x7E\x00\x86\x88\x00\x00\x87\x5C\x03\x00\xAE\x45"
-     "\x7E\x00"
-     "\xAF\x4D\x7E\x00\xAC\x5D\x7E\x00\xAD\x55\x7E\x00\xB3\xA5\x7E\x00\x72\x70"
-     "\x00\x80"
-     "\x76\x70\x00\x80\x7A\x70\x00\x80\x7E\x70\x00\x80\xB6\xA1\x7E\x00\xB5\xB5"
-     "\x7E\x00"
-     "\x82\x70\x00\x80\xBB\xF5\x7E\x00\xBA\xE9\x7E\x00\x86\x70\x00\x80\x8A\x70"
-     "\x00\x80"
-     "\xBF\x15\x01\x00\xBE\xD1\x7E\x00\xBD\xD9\x7E\x00\xBC\xE5\x7E\x00\x8E\x70"
-     "\x00\x80"
-     "\x92\x70\x00\x80\x96\x70\x00\x80\x9A\x70\x00\x80\xEF\x08\x7D\x00\x9E\x70"
-     "\x00\x80"
-     "\xA2\x70\x00\x80\xA6\x70\x00\x80\xE3\xFC\x7D\x00\xAA\x70\x00\x80\xE1\x18"
-     "\x7D\x00"
-     "\xAE\x70\x00\x80\xB2\x70\x00\x80\xB6\x70\x00\x80\xBA\x70\x00\x80\xBE\x70"
-     "\x00\x80"
-     "\xA8\x15\x7E\x00\xA9\x25\x7E\x00\xAA\x2D\x7E\x00\xAB\x25\x7E\x00\xAC\x3D"
-     "\x7E\x00"
-     "\xAD\x21\x7E\x00\xAE\x59\x7E\x00\xAF\x59\x7E\x00\xC2\x70\x00\x80\xC6\x70"
-     "\x00\x80"
-     "\xCA\x70\x00\x80\xCE\x70\x00\x80\x80\x0D\x00\x00\x81\xB5\x03\x00\x82\xBD"
-     "\x03\x00"
-     "\xD2\x70\x00\x80\xB8\xF1\x01\x00\xB9\xF1\x01\x00\xBA\xF1\x01\x00\xBB\xF1"
-     "\x01\x00"
-     "\xBC\x55\x01\x00\xBD\x5D\x01\x00\xBE\x55\x01\x00\xBF\x4D\x01\x00\xB0\x2D"
-     "\x7E\x00"
-     "\xB1\x31\x7E\x00\xB2\x31\x7E\x00\xB3\x0D\x7E\x00\xB4\x15\x7E\x00\xB5\xD1"
-     "\x01\x00"
-     "\xB6\xD1\x01\x00\xB7\xD1\x01\x00\xD6\x70\x00\x80\xDA\x70\x00\x80\x84\x90"
-     "\x02\x00"
-     "\xA3\xA9\x7D\x00\xBE\x9C\x03\x00\xA5\xB9\x7D\x00\xA6\xAD\x7D\x00\xE2\x70"
-     "\x00\x80"
-     "\xE6\x70\x00\x80\xEA\x70\x00\x80\xAA\xE5\x7D\x00\xAB\xF9\x7D\x00\xAC\xE9"
-     "\x7D\x00"
-     "\xAD\xD5\x7D\x00\xAE\xDD\x7D\x00\xAF\x19\x02\x00\xB6\x69\x03\x00\xEE\x70"
-     "\x00\x80"
-     "\x84\x48\x03\x00\xB5\x51\x03\x00\xF2\x70\x00\x80\xB3\x59\x03\x00\x86\xF8"
-     "\x04\x00"
-     "\x87\x34\x03\x00\xBE\x39\x03\x00\xBF\x39\x03\x00\xBC\x25\x03\x00\xBD\x39"
-     "\x03\x00"
-     "\xBA\x35\x03\x00\xBB\x3D\x03\x00\xF6\x70\x00\x80\xFA\x70\x00\x80\xB8\xF1"
-     "\x03\x00"
-     "\xB9\xF1\x03\x00\xBA\xF1\x03\x00\xBB\xF1\x03\x00\xBC\x31\x03\x00\xBD\x31"
-     "\x03\x00"
-     "\xBE\x31\x03\x00\xBF\x31\x03\x00\xB0\xE5\x03\x00\xB1\xED\x03\x00\xB2\xE5"
-     "\x03\x00"
-     "\xB3\xF9\x03\x00\xB4\xE9\x03\x00\xB5\xE9\x03\x00\xB6\xDD\x03\x00\xB7\xD1"
-     "\x03\x00"
-     "\xA8\xC5\x03\x00\xA9\xCD\x03\x00\xAA\xC5\x03\x00\xAB\xDD\x03\x00\xAC\xCD"
-     "\x03\x00"
-     "\xAD\xFD\x03\x00\xAE\xF5\x03\x00\xAF\x9D\x03\x00\xFE\x70\x00\x80\x02\x71"
-     "\x00\x80"
-     "\x06\x71\x00\x80\x8E\x6F\x00\x80\xBE\xFC\x04\x00\x0A\x71\x00\x80\x0E\x71"
-     "\x00\x80"
-     "\x12\x71\x00\x80\xA3\x0D\x02\x00\x16\x71\x00\x80\x1A\x71\x00\x80\x1E\x71"
-     "\x00\x80"
-     "\x22\x71\x00\x80\xA6\x3D\x02\x00\xA5\x05\x02\x00\x26\x71\x00\x80\xAB\x69"
-     "\x02\x00"
-     "\xAA\x61\x02\x00\x2A\x71\x00\x80\x2E\x71\x00\x80\xAF\x6D\x02\x00\xAE\x6D"
-     "\x02\x00"
-     "\xAD\x6D\x02\x00\xAC\x71\x02\x00\x32\x71\x00\x80\x36\x71\x00\x80\x3A\x71"
-     "\x00\x80"
-     "\x3E\x71\x00\x80\x42\x71\x00\x80\xE3\x4C\x03\x00\x46\x71\x00\x80\xE1\xAC"
-     "\x01\x00"
-     "\x82\x11\x00\x00\xEF\x8C\x03\x00\x80\x39\x00\x00\x81\x35\x00\x00\x4A\x71"
-     "\x00\x80"
-     "\x52\x71\x00\x80\x86\x58\x04\x00\x87\x40\x05\x00\x56\x71\x00\x80\x5A\x71"
-     "\x00\x80"
-     "\x5E\x71\x00\x80\x62\x71\x00\x80\x66\x71\x00\x80\x6A\x71\x00\x80\x6E\x71"
-     "\x00\x80"
-     "\xEF\x44\x7E\x00\x72\x71\x00\x80\xE1\x24\x7E\x00\x76\x71\x00\x80\xE3\x68"
-     "\x01\x00"
-     "\xEC\x00\x00\x00\x7A\x71\x00\x80\x7E\x71\x00\x80\x82\x71\x00\x80\x86\x71"
-     "\x00\x80"
-     "\x8A\x71\x00\x80\x8E\x71\x00\x80\xB3\xE9\x01\x00\x92\x71\x00\x80\xB5\x85"
-     "\x01\x00"
-     "\xB6\xA1\x01\x00\x96\x71\x00\x80\x9A\x71\x00\x80\x9E\x71\x00\x80\xBA\x0D"
-     "\x01\x00"
-     "\xBB\x05\x01\x00\xBC\x05\x01\x00\xBD\x0D\x01\x00\xBE\x05\x01\x00\xBF\xF9"
-     "\x00\x00"
-     "\xA8\x1D\x06\x00\xA9\x25\x06\x00\xAA\x2D\x06\x00\xAB\x49\x06\x00\xAC\x41"
-     "\x06\x00"
-     "\xAD\x6D\x06\x00\xAE\x65\x06\x00\xAF\xFD\x06\x00\x4E\x71\x00\x80\xA2\x71"
-     "\x00\x80"
-     "\xA6\x71\x00\x80\xAA\x71\x00\x80\xAE\x71\x00\x80\x82\xB9\x00\x00\x81\xB1"
-     "\x00\x00"
-     "\x80\xB9\x00\x00\xB8\x51\x07\x00\xB9\x59\x07\x00\xBA\x61\x07\x00\xBB\x61"
-     "\x07\x00"
-     "\xBC\x11\x07\x00\xBD\x11\x07\x00\xBE\x11\x07\x00\xBF\x11\x07\x00\xB0\x85"
-     "\x06\x00"
-     "\xB1\x8D\x06\x00\xB2\x9D\x06\x00\xB3\x95\x06\x00\xB4\x8D\x06\x00\xB5\x71"
-     "\x07\x00"
-     "\xB6\x71\x07\x00\xB7\x71\x07\x00\xA3\xA5\x06\x00\xB2\x71\x00\x80\xB6\x71"
-     "\x00\x80"
-     "\xBA\x71\x00\x80\x84\x98\x02\x00\xA6\xED\x06\x00\xA5\xC9\x06\x00\xBE\x10"
-     "\x03\x00"
-     "\xAB\x49\x06\x00\xAA\x41\x06\x00\xBE\x71\x00\x80\xC2\x71\x00\x80\xAF\xB5"
-     "\x07\x00"
-     "\xAE\x49\x06\x00\xAD\x41\x06\x00\xAC\x49\x06\x00\xC6\x71\x00\x80\xB3\x41"
-     "\x07\x00"
-     "\x86\x68\x00\x00\x87\x6C\x03\x00\xB6\x41\x07\x00\xCA\x71\x00\x80\xCE\x71"
-     "\x00\x80"
-     "\xB5\x51\x07\x00\xBA\x41\x07\x00\xBB\x4D\x07\x00\xD2\x71\x00\x80\xD6\x71"
-     "\x00\x80"
-     "\xBE\x4D\x07\x00\xBF\x35\x07\x00\xBC\x4D\x07\x00\xBD\x45\x07\x00\xA3\x85"
-     "\x06\x00"
-     "\xDA\x71\x00\x80\xDE\x71\x00\x80\xE2\x71\x00\x80\xE6\x71\x00\x80\xA6\x85"
-     "\x06\x00"
-     "\xA5\x95\x06\x00\xEA\x71\x00\x80\xAB\x89\x06\x00\xAA\x85\x06\x00\xEE\x71"
-     "\x00\x80"
-     "\xF2\x71\x00\x80\xAF\xF1\x06\x00\xAE\x89\x06\x00\xAD\x81\x06\x00\xAC\x89"
-     "\x06\x00"
-     "\xF6\x71\x00\x80\xEF\x4C\x05\x00\xFA\x71\x00\x80\xFE\x71\x00\x80\x02\x72"
-     "\x00\x80"
-     "\x06\x72\x00\x80\x0A\x72\x00\x80\x0E\x72\x00\x80\x12\x72\x00\x80\x16\x72"
-     "\x00\x80"
-     "\x1A\x72\x00\x80\x1E\x72\x00\x80\x22\x72\x00\x80\xE3\x2C\x05\x00\x26\x72"
-     "\x00\x80"
-     "\xE1\x3C\x05\x00\xA8\xF5\x06\x00\xA9\x05\x06\x00\xAA\x0D\x06\x00\xAB\x05"
-     "\x06\x00"
-     "\xAC\x1D\x06\x00\xAD\x05\x06\x00\xAE\x0D\x06\x00\xAF\x05\x06\x00\x80\xD1"
-     "\x00\x00"
-     "\x81\xDD\x00\x00\x82\x05\x00\x00\x2A\x72\x00\x80\xBE\xEC\x01\x00\x84\xF8"
-     "\x01\x00"
-     "\x32\x72\x00\x80\x36\x72\x00\x80\xB8\x11\x06\x00\xB9\x19\x06\x00\xBA\x29"
-     "\x06\x00"
-     "\xBB\x29\x06\x00\xBC\xF5\x01\x00\xBD\xFD\x01\x00\xBE\xF5\x01\x00\xBF\xED"
-     "\x01\x00"
-     "\xB0\x41\x06\x00\xB1\x4D\x06\x00\xB2\x5D\x06\x00\xB3\x55\x06\x00\xB4\x4D"
-     "\x06\x00"
-     "\xB5\x35\x06\x00\xB6\x3D\x06\x00\xB7\x35\x06\x00\xA8\xAD\x02\x00\xA9\xC5"
-     "\x02\x00"
-     "\xAA\xDD\x02\x00\xAB\xD5\x02\x00\xAC\xF9\x02\x00\xAD\xF9\x02\x00\xAE\x1D"
-     "\x03\x00"
-     "\xAF\x15\x03\x00\x86\x80\x0C\x00\x87\x3C\x03\x00\x3A\x72\x00\x80\x3E\x72"
-     "\x00\x80"
-     "\x42\x72\x00\x80\x46\x72\x00\x80\x4A\x72\x00\x80\xBE\x78\x03\x00\xB8\x11"
-     "\x03\x00"
-     "\xB9\x11\x03\x00\xBA\x11\x03\x00\xBB\x11\x03\x00\xBC\x31\x03\x00\xBD\x31"
-     "\x03\x00"
-     "\xBE\x31\x03\x00\xBF\x31\x03\x00\xB0\x6D\x03\x00\xB1\x75\x03\x00\xB2\x75"
-     "\x03\x00"
-     "\xB3\x45\x03\x00\xB4\x5D\x03\x00\xB5\x35\x03\x00\xB6\x31\x03\x00\xB7\x31"
-     "\x03\x00"
-     "\xB3\x55\x02\x00\x4E\x72\x00\x80\x52\x72\x00\x80\x56\x72\x00\x80\x5A\x72"
-     "\x00\x80"
-     "\xB6\x79\x02\x00\xB5\x71\x02\x00\x5E\x72\x00\x80\xBB\xB1\x03\x00\xBA\xA9"
-     "\x03\x00"
-     "\x62\x72\x00\x80\x66\x72\x00\x80\xBF\xF1\x03\x00\xBE\x99\x03\x00\xBD\x91"
-     "\x03\x00"
-     "\xBC\xA1\x03\x00\x6A\x72\x00\x80\xA6\x29\x02\x00\xA5\x21\x02\x00\x6E\x72"
-     "\x00\x80"
-     "\xA3\x05\x02\x00\x72\x72\x00\x80\x76\x72\x00\x80\x7A\x72\x00\x80\xAF\xA1"
-     "\x03\x00"
-     "\xAE\xC9\x03\x00\xAD\xC1\x03\x00\xAC\xF1\x03\x00\xAB\xE1\x03\x00\xAA\xF9"
-     "\x03\x00"
-     "\xDE\x70\x00\x80\x7E\x72\x00\x80\x82\x72\x00\x80\x86\x72\x00\x80\x8A\x72"
-     "\x00\x80"
-     "\x8E\x72\x00\x80\x82\x25\x00\x00\xEF\x5C\x00\x00\x80\x31\x00\x00\x81\x2D"
-     "\x00\x00"
-     "\x84\xAC\x02\x00\xE3\x80\x00\x00\x92\x72\x00\x80\xE1\xFC\x01\x00\x96\x72"
-     "\x00\x80"
-     "\x9A\x72\x00\x80\x86\x18\x0D\x00\x87\x10\x0D\x00\xB3\x69\x01\x00\xA2\x72"
-     "\x00\x80"
-     "\xBE\x44\x0C\x00\xA6\x72\x00\x80\xAA\x72\x00\x80\xB6\x61\x01\x00\xB5\x69"
-     "\x01\x00"
-     "\xAE\x72\x00\x80\xBB\xED\x01\x00\xBA\xED\x01\x00\xB2\x72\x00\x80\xB6\x72"
-     "\x00\x80"
-     "\xBF\xDD\x01\x00\xBE\xDD\x01\x00\xBD\xE5\x01\x00\xBC\xED\x01\x00\xBA\x72"
-     "\x00\x80"
-     "\xBE\x72\x00\x80\xC2\x72\x00\x80\x84\x00\x0C\x00\xC6\x72\x00\x80\xCA\x72"
-     "\x00\x80"
-     "\xCE\x72\x00\x80\xEF\xF8\x01\x00\x85\xEC\x0B\x00\xE1\xC4\x06\x00\xD2\x72"
-     "\x00\x80"
-     "\xE3\x2C\x01\x00\xD6\x72\x00\x80\xDA\x72\x00\x80\xDE\x72\x00\x80\xE2\x72"
-     "\x00\x80"
-     "\xE6\x72\x00\x80\xEA\x72\x00\x80\xA3\xE1\x01\x00\xEE\x72\x00\x80\xA5\xE1"
-     "\x01\x00"
-     "\xF2\x72\x00\x80\xF6\x72\x00\x80\xA6\xE9\x01\x00\xFA\x72\x00\x80\xFE\x72"
-     "\x00\x80"
-     "\xAB\x65\x01\x00\xAA\x65\x01\x00\xAD\x6D\x01\x00\xAC\x65\x01\x00\xAF\x55"
-     "\x01\x00"
-     "\xAE\x55\x01\x00\xA8\xB9\x0D\x00\xA9\xB9\x0D\x00\xAA\x21\x0E\x00\xAB\x21"
-     "\x0E\x00"
-     "\xAC\x21\x0E\x00\xAD\x21\x0E\x00\xAE\x21\x0E\x00\xAF\x99\x0E\x00\x9E\x72"
-     "\x00\x80"
-     "\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\x02\x73\x00\x80\x06\x73"
-     "\x00\x80"
-     "\x0A\x73\x00\x80\x0E\x73\x00\x80\xB8\x95\x0E\x00\xB9\x9D\x0E\x00\xBA\x95"
-     "\x0E\x00"
-     "\xBB\xB1\x0E\x00\xBC\x51\x0F\x00\xBD\x51\x0F\x00\xBE\x51\x0F\x00\xBF\x51"
-     "\x0F\x00"
-     "\xB0\xE9\x0E\x00\xB1\xE9\x0E\x00\xB2\xF9\x0E\x00\xB3\xF1\x0E\x00\xB4\xD9"
-     "\x0E\x00"
-     "\xB5\xD9\x0E\x00\xB6\xB9\x0E\x00\xB7\xB1\x0E\x00\xA8\xC5\x0F\x00\xA9\xC9"
-     "\x0F\x00"
-     "\xAA\xD9\x0F\x00\xAB\xD1\x0F\x00\xAC\xF9\x0F\x00\xAD\xF9\x0F\x00\xAE\x19"
-     "\x0F\x00"
-     "\xAF\x19\x0F\x00\x86\x00\x00\x00\x87\x64\x03\x00\x12\x73\x00\x80\x16\x73"
-     "\x00\x80"
-     "\x1A\x73\x00\x80\x1E\x73\x00\x80\x22\x73\x00\x80\x26\x73\x00\x80\xB8\x7D"
-     "\x0F\x00"
-     "\xB9\x05\x0F\x00\xBA\x0D\x0F\x00\xBB\x01\x0F\x00\xBC\x05\x0F\x00\xBD\x0D"
-     "\x0F\x00"
-     "\xBE\x39\x0F\x00\xBF\x39\x0F\x00\xB0\x69\x0F\x00\xB1\x75\x0F\x00\xB2\x7D"
-     "\x0F\x00"
-     "\xB3\x75\x0F\x00\xB4\x59\x0F\x00\xB5\x45\x0F\x00\xB6\x4D\x0F\x00\xB7\x45"
-     "\x0F\x00"
-     "\xE1\x38\x0D\x00\x2A\x73\x00\x80\xE3\x38\x0C\x00\x2E\x73\x00\x80\x32\x73"
-     "\x00\x80"
-     "\x36\x73\x00\x80\x3A\x73\x00\x80\x3E\x73\x00\x80\x42\x73\x00\x80\x46\x73"
-     "\x00\x80"
-     "\x4A\x73\x00\x80\x4E\x73\x00\x80\x52\x73\x00\x80\x56\x73\x00\x80\xEF\xFC"
-     "\x0D\x00"
-     "\x5A\x73\x00\x80\xB3\x11\x0E\x00\x5E\x73\x00\x80\x62\x73\x00\x80\x66\x73"
-     "\x00\x80"
-     "\x6A\x73\x00\x80\xB6\x39\x0E\x00\xB5\x31\x0E\x00\x6E\x73\x00\x80\xBB\xF1"
-     "\x0E\x00"
-     "\xBA\xED\x0E\x00\x72\x73\x00\x80\x76\x73\x00\x80\xBF\xC1\x0E\x00\xBE\xDD"
-     "\x0E\x00"
-     "\xBD\xE1\x0E\x00\xBC\xE9\x0E\x00\xA3\xD1\x0E\x00\x7A\x73\x00\x80\x84\x88"
-     "\x02\x00"
-     "\xBE\x80\x01\x00\x82\x73\x00\x80\xA6\xF9\x0E\x00\xA5\xF1\x0E\x00\x86\x73"
-     "\x00\x80"
-     "\xAB\x31\x0E\x00\xAA\x2D\x0E\x00\x86\x08\x00\x00\x87\x4C\x01\x00\xAF\x01"
-     "\x0E\x00"
-     "\xAE\x1D\x0E\x00\xAD\x21\x0E\x00\xAC\x29\x0E\x00\x80\xA9\x00\x00\x81\xA9"
-     "\x00\x00"
-     "\x82\x4D\x00\x00\xB3\x7D\x01\x00\x8A\x73\x00\x80\xB5\x6D\x01\x00\xB6\x65"
-     "\x01\x00"
-     "\x8E\x73\x00\x80\x92\x73\x00\x80\x96\x73\x00\x80\xBA\xF9\x00\x00\xBB\xF9"
-     "\x00\x00"
-     "\xBC\xE9\x00\x00\xBD\xE9\x00\x00\xBE\xD9\x00\x00\xBF\xD9\x00\x00\xA8\xA9"
-     "\x02\x00"
-     "\xA9\xA9\x02\x00\xAA\xB9\x02\x00\xAB\xB9\x02\x00\xAC\xA9\x02\x00\xAD\xA9"
-     "\x02\x00"
-     "\xAE\x1D\x03\x00\xAF\x11\x03\x00\xBE\xAC\x04\x00\x9A\x73\x00\x80\x9E\x73"
-     "\x00\x80"
-     "\xBE\xE0\x05\x00\xA2\x73\x00\x80\xA6\x73\x00\x80\xAA\x73\x00\x80\xAE\x73"
-     "\x00\x80"
-     "\xB8\x71\x03\x00\xB9\x71\x03\x00\xBA\x75\x03\x00\xBB\x09\x03\x00\xBC\x19"
-     "\x03\x00"
-     "\xBD\x19\x03\x00\xBE\x09\x03\x00\xBF\x01\x03\x00\xB0\x71\x03\x00\xB1\x71"
-     "\x03\x00"
-     "\xB2\x71\x03\x00\xB3\x71\x03\x00\xB4\x51\x03\x00\xB5\x51\x03\x00\xB6\x51"
-     "\x03\x00"
-     "\xB7\x51\x03\x00\xA3\x31\x02\x00\x2E\x72\x00\x80\xB2\x73\x00\x80\xB6\x73"
-     "\x00\x80"
-     "\xBA\x73\x00\x80\xA6\x29\x02\x00\xA5\x21\x02\x00\xBE\x73\x00\x80\xAB\xB5"
-     "\x03\x00"
-     "\xAA\xB5\x03\x00\xC2\x73\x00\x80\xC6\x73\x00\x80\xAF\x95\x03\x00\xAE\x95"
-     "\x03\x00"
-     "\xAD\xA5\x03\x00\xAC\xA5\x03\x00\xCA\x73\x00\x80\xCE\x73\x00\x80\xD2\x73"
-     "\x00\x80"
-     "\xD6\x73\x00\x80\xDA\x73\x00\x80\xDE\x73\x00\x80\xE2\x73\x00\x80\xE6\x73"
-     "\x00\x80"
-     "\x80\x39\x00\x00\x81\x39\x00\x00\x82\x05\x00\x00\xEA\x73\x00\x80\xF2\x73"
-     "\x00\x80"
-     "\x84\xFC\x03\x00\xF6\x73\x00\x80\xFA\x73\x00\x80\xA8\x91\x02\x00\xA9\x99"
-     "\x02\x00"
-     "\xAA\xD9\x02\x00\xAB\xD1\x02\x00\xAC\xF1\x02\x00\xAD\xF1\x02\x00\xAE\x3D"
-     "\x01\x00"
-     "\xAF\x35\x01\x00\x86\x00\x05\x00\x87\x64\x05\x00\xFE\x73\x00\x80\x02\x74"
-     "\x00\x80"
-     "\x06\x74\x00\x80\x0A\x74\x00\x80\x0E\x74\x00\x80\x12\x74\x00\x80\xB8\xE9"
-     "\x01\x00"
-     "\xB9\xE9\x01\x00\xBA\x89\x01\x00\xBB\x89\x01\x00\xBC\x99\x01\x00\xBD\x99"
-     "\x01\x00"
-     "\xBE\x89\x01\x00\xBF\x89\x01\x00\xB0\x51\x01\x00\xB1\x51\x01\x00\xB2\x51"
-     "\x01\x00"
-     "\xB3\x51\x01\x00\xB4\xF9\x01\x00\xB5\xF9\x01\x00\xB6\xE9\x01\x00\xB7\xE1"
-     "\x01\x00"
-     "\xEF\xF4\x00\x00\x16\x74\x00\x80\x1A\x74\x00\x80\xBF\x8C\x07\x00\x1E\x74"
-     "\x00\x80"
-     "\x22\x74\x00\x80\x84\x18\x05\x00\xEF\x2C\x0E\x00\x26\x74\x00\x80\xE1\x84"
-     "\x0E\x00"
-     "\x2A\x74\x00\x80\xE3\x30\x0E\x00\xE3\x28\x00\x00\x2E\x74\x00\x80\xE1\x00"
-     "\x01\x00"
-     "\x32\x74\x00\x80\x85\xFC\x20\x00\xB3\x15\x02\x00\x36\x74\x00\x80\x3A\x74"
-     "\x00\x80"
-     "\xB6\x35\x02\x00\x3E\x74\x00\x80\x42\x74\x00\x80\xB5\x3D\x02\x00\xBA\xD1"
-     "\x01\x00"
-     "\xBB\xD1\x01\x00\x46\x74\x00\x80\x4A\x74\x00\x80\xBE\x79\x01\x00\xBF\x79"
-     "\x01\x00"
-     "\xBC\xC1\x01\x00\xBD\xC1\x01\x00\xA3\xD1\x05\x00\xEE\x73\x00\x80\x4E\x74"
-     "\x00\x80"
-     "\x52\x74\x00\x80\x56\x74\x00\x80\xA6\xF1\x05\x00\xA5\xF9\x05\x00\x5A\x74"
-     "\x00\x80"
-     "\xAB\x15\x06\x00\xAA\x15\x06\x00\x5E\x74\x00\x80\x62\x74\x00\x80\xAF\xBD"
-     "\x06\x00"
-     "\xAE\xBD\x06\x00\xAD\x05\x06\x00\xAC\x05\x06\x00\x80\x4D\x00\x00\x81\x55"
-     "\x00\x00"
-     "\x82\x55\x00\x00\xB3\xB1\x06\x00\x66\x74\x00\x80\xB5\x79\x07\x00\xB6\x69"
-     "\x07\x00"
-     "\x6A\x74\x00\x80\x86\xC0\x01\x00\x87\xE8\x03\x00\xBA\x55\x07\x00\xBB\xE9"
-     "\x07\x00"
-     "\xBC\xF9\x07\x00\xBD\xE5\x07\x00\xBE\xED\x07\x00\xBF\xDD\x07\x00\xA8\x71"
-     "\x06\x00"
-     "\xA9\x7D\x06\x00\xAA\x75\x06\x00\xAB\xA5\x07\x00\xAC\xBD\x07\x00\xAD\xA5"
-     "\x07\x00"
-     "\xAE\xAD\x07\x00\xAF\xA5\x07\x00\x6E\x74\x00\x80\x72\x74\x00\x80\x76\x74"
-     "\x00\x80"
-     "\x7A\x74\x00\x80\x7E\x74\x00\x80\x82\x74\x00\x80\x86\x74\x00\x80\x8A\x74"
-     "\x00\x80"
-     "\xB8\x7D\x07\x00\xB9\x0D\x07\x00\xBA\x05\x07\x00\xBB\x19\x07\x00\xBC\x09"
-     "\x07\x00"
-     "\xBD\x3D\x07\x00\xBE\x35\x07\x00\xBF\xD5\x00\x00\xB0\xDD\x07\x00\xB1\x61"
-     "\x07\x00"
-     "\xB2\x61\x07\x00\xB3\x7D\x07\x00\xB4\x65\x07\x00\xB5\x55\x07\x00\xB6\x51"
-     "\x07\x00"
-     "\xB7\x49\x07\x00\x8E\x74\x00\x80\x92\x74\x00\x80\x96\x74\x00\x80\x9A\x74"
-     "\x00\x80"
-     "\xE1\xA4\x08\x00\xE1\xA4\x06\x00\xE3\xCC\x04\x00\xE3\xF8\x05\x00\x9E\x74"
-     "\x00\x80"
-     "\xA2\x74\x00\x80\xEF\x7C\x18\x00\xEF\xF4\x05\x00\xA3\x35\x01\x00\xA2\xA5"
-     "\x03\x00"
-     "\xA1\x05\x05\x00\xA0\x1D\x05\x00\xA3\xBD\x07\x00\xA6\x74\x00\x80\xAA\x74"
-     "\x00\x80"
-     "\xAE\x74\x00\x80\xB2\x74\x00\x80\xA6\x65\x06\x00\xA5\x75\x06\x00\xB6\x74"
-     "\x00\x80"
-     "\xAB\xE5\x06\x00\xAA\x59\x06\x00\xBA\x74\x00\x80\xBE\x74\x00\x80\xAF\xD1"
-     "\x06\x00"
-     "\xAE\xE1\x06\x00\xAD\xE9\x06\x00\xAC\xF5\x06\x00\x8F\xED\x7D\x00\x82\x1D"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x80\x1D\x00\x00\xC2\x74\x00\x80\xC6\x74\x00\x80\xCA\x74"
-     "\x00\x80"
-     "\xBE\x00\x1C\x00\x84\xBC\x00\x00\xD2\x74\x00\x80\x87\x98\x00\x00\x86\x2C"
-     "\x1F\x00"
-     "\xD6\x74\x00\x80\xDA\x74\x00\x80\xDE\x74\x00\x80\xE2\x74\x00\x80\x9F\x45"
-     "\x07\x00"
-     "\x9E\x6D\x09\x00\x9D\xE9\x09\x00\x9C\x35\x0B\x00\x9B\x09\x0D\x00\x9A\xF5"
-     "\x0D\x00"
-     "\x99\x35\x0F\x00\x98\xCD\x70\x00\x97\xE5\x71\x00\x96\xE5\x71\x00\x95\x8D"
-     "\x74\x00"
-     "\x94\xE5\x75\x00\x93\xF5\x75\x00\x92\x25\x77\x00\x91\xF1\x79\x00\x90\xF9"
-     "\x79\x00"
-     "\xE6\x74\x00\x80\xEA\x74\x00\x80\x82\x65\x7E\x00\x83\x09\x7A\x00\x84\x19"
-     "\x7A\x00"
-     "\x85\xC5\x74\x00\x86\x1D\x76\x00\x87\x15\x76\x00\x84\x6C\x1C\x00\xEE\x74"
-     "\x00\x80"
-     "\x8A\x4D\x77\x00\x8B\x2D\x72\x00\x8C\x35\x72\x00\x8D\x25\x73\x00\x8E\xD5"
-     "\x0C\x00"
-     "\x8F\x1D\x0E\x00\x85\x04\x25\x00\xF2\x74\x00\x80\x92\x91\x0E\x00\x93\xB5"
-     "\x08\x00"
-     "\x94\x71\x0A\x00\x95\xFD\x0A\x00\x96\xDD\x04\x00\x97\x7D\x06\x00\xF6\x74"
-     "\x00\x80"
-     "\xFA\x74\x00\x80\x9A\x55\x06\x00\x9B\xFD\x00\x00\x9C\x75\x02\x00\xFE\x74"
-     "\x00\x80"
-     "\xBE\x30\x1C\x00\x02\x75\x00\x80\xB3\xA1\x2A\x00\xB2\x01\x28\x00\xB1\xA9"
-     "\x2F\x00"
-     "\xB0\x85\x2E\x00\x7E\x73\x00\x80\x84\xD8\x1D\x00\xB5\x01\x24\x00\xB4\xA9"
-     "\x2B\x00"
-     "\x06\x75\x00\x80\x0A\x75\x00\x80\x0E\x75\x00\x80\x12\x75\x00\x80\x16\x75"
-     "\x00\x80"
-     "\x1A\x75\x00\x80\x1E\x75\x00\x80\x22\x75\x00\x80\xA3\xC5\x1E\x00\xA2\xC5"
-     "\x1E\x00"
-     "\xA1\xE1\x02\x00\x26\x75\x00\x80\xA7\x01\x1A\x00\xA6\xF5\x1A\x00\xA5\x01"
-     "\x18\x00"
-     "\xA4\x05\x1E\x00\xAB\x01\x10\x00\xAA\xD5\x16\x00\xA9\xC9\x16\x00\xA8\x01"
-     "\x14\x00"
-     "\xAF\xF5\x2E\x00\xAE\x8D\x13\x00\xAD\xA5\x12\x00\xAC\xF9\x12\x00\xA8\x2D"
-     "\x03\x00"
-     "\xA9\x55\x03\x00\xAA\x55\x03\x00\xAB\xAD\x03\x00\xAC\xB5\x03\x00\xAD\xBD"
-     "\x03\x00"
-     "\xAE\xB5\x03\x00\xAF\xAD\x03\x00\x2A\x75\x00\x80\x2E\x75\x00\x80\x32\x75"
-     "\x00\x80"
-     "\x36\x75\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x3A\x75"
-     "\x00\x80"
-     "\xB8\x65\x02\x00\xB9\x01\x02\x00\xBA\x01\x02\x00\xBB\x01\x02\x00\xBC\x01"
-     "\x02\x00"
-     "\xBD\x09\x02\x00\xBE\x31\x02\x00\xBF\x31\x02\x00\xB0\xD5\x03\x00\xB1\xDD"
-     "\x03\x00"
-     "\xB2\xD5\x03\x00\xB3\x6D\x02\x00\xB4\x75\x02\x00\xB5\x7D\x02\x00\xB6\x75"
-     "\x02\x00"
-     "\xB7\x5D\x02\x00\xE1\xB8\x06\x00\x42\x75\x00\x80\xE3\x00\x06\x00\xBE\xC4"
-     "\x1C\x00"
-     "\x46\x75\x00\x80\xEF\xF0\x00\x00\x86\x38\x1C\x00\x87\xBC\x1C\x00\x4A\x75"
-     "\x00\x80"
-     "\xE3\x14\x00\x00\x4E\x75\x00\x80\xE1\xD4\x01\x00\x86\x10\x1F\x00\x87\x84"
-     "\x1F\x00"
-     "\xEF\x6C\x07\x00\x84\xF0\x1F\x00\x52\x75\x00\x80\xB6\x6D\x03\x00\xB5\x6D"
-     "\x03\x00"
-     "\xBE\x84\x1F\x00\xB3\x8D\x03\x00\x56\x75\x00\x80\x5A\x75\x00\x80\x5E\x75"
-     "\x00\x80"
-     "\xBF\xE9\x02\x00\xBE\xE1\x02\x00\xBD\xE9\x02\x00\xBC\x51\x03\x00\xBB\x49"
-     "\x03\x00"
-     "\xBA\x41\x03\x00\x62\x75\x00\x80\x66\x75\x00\x80\xCE\x74\x00\x80\x3E\x75"
-     "\x00\x80"
-     "\x6A\x75\x00\x80\x6E\x75\x00\x80\x72\x75\x00\x80\x76\x75\x00\x80\x7A\x75"
-     "\x00\x80"
-     "\x7E\x75\x00\x80\xA8\x85\x1D\x00\xA9\x95\x1D\x00\xAA\x9D\x1D\x00\xAB\x95"
-     "\x1D\x00"
-     "\xAC\x8D\x1D\x00\xAD\x35\x1E\x00\xAE\x3D\x1E\x00\xAF\x19\x1E\x00\xB0\x05"
-     "\x1E\x00"
-     "\xB1\x25\x1E\x00\xB2\x2D\x1E\x00\xB3\x39\x1E\x00\xB4\x2D\x1E\x00\xB5\xF5"
-     "\x1E\x00"
-     "\xB6\xFD\x1E\x00\xB7\xF5\x1E\x00\xB8\xCD\x1E\x00\xB9\xD5\x1E\x00\xBA\xDD"
-     "\x1E\x00"
-     "\xBB\xD5\x1E\x00\xBC\xCD\x1E\x00\xBD\xB9\x1E\x00\xBE\x59\x1F\x00\xBF\x59"
-     "\x1F\x00"
-     "\xB3\x35\x1D\x00\x82\x75\x00\x80\x86\x75\x00\x80\x8A\x75\x00\x80\x8E\x75"
-     "\x00\x80"
-     "\xB6\x95\x1E\x00\xB5\x25\x1D\x00\x92\x75\x00\x80\xBB\xC1\x1E\x00\xBA\xE1"
-     "\x1E\x00"
-     "\x96\x75\x00\x80\x9A\x75\x00\x80\xBF\x11\x1E\x00\xBE\xC9\x1E\x00\xBD\xCD"
-     "\x1E\x00"
-     "\xBC\xD9\x1E\x00\x9E\x75\x00\x80\xA3\x71\x1D\x00\xA2\x75\x00\x80\xA6\x75"
-     "\x00\x80"
-     "\xA6\xD1\x1E\x00\xAA\x75\x00\x80\xAE\x75\x00\x80\xA5\x61\x1D\x00\xAA\xA5"
-     "\x1E\x00"
-     "\xAB\x85\x1E\x00\xB2\x75\x00\x80\xB6\x75\x00\x80\xAE\x8D\x1E\x00\xAF\x55"
-     "\x1E\x00"
-     "\xAC\x9D\x1E\x00\xAD\x89\x1E\x00\xA8\x85\x1E\x00\xA9\x95\x1E\x00\xAA\xA9"
-     "\x1E\x00"
-     "\xAB\xA5\x1E\x00\xAC\xD1\x1E\x00\xAD\xC1\x1E\x00\xAE\xC1\x1E\x00\xAF\xC1"
-     "\x1E\x00"
-     "\x80\xE9\x01\x00\x81\xE9\x01\x00\x82\xFD\x01\x00\xBA\x75\x00\x80\xBE\x75"
-     "\x00\x80"
-     "\xC2\x75\x00\x80\xC6\x75\x00\x80\xCA\x75\x00\x80\xB8\xC5\x01\x00\xB9\xCD"
-     "\x01\x00"
-     "\xBA\xC5\x01\x00\xBB\xDD\x01\x00\xBC\xCD\x01\x00\xBD\xFD\x01\x00\xBE\xF5"
-     "\x01\x00"
-     "\xBF\x9D\x01\x00\xB0\x99\x1E\x00\xB1\x99\x1E\x00\xB2\x61\x01\x00\xB3\x61"
-     "\x01\x00"
-     "\xB4\x61\x01\x00\xB5\x61\x01\x00\xB6\x61\x01\x00\xB7\x61\x01\x00\xB3\x21"
-     "\x1E\x00"
-     "\xCE\x75\x00\x80\xD2\x75\x00\x80\xD6\x75\x00\x80\xDA\x75\x00\x80\xB6\x41"
-     "\x1E\x00"
-     "\xB5\x59\x1E\x00\x84\x90\x00\x00\xBB\x5D\x1E\x00\xBA\x5D\x1E\x00\xDE\x75"
-     "\x00\x80"
-     "\xE2\x75\x00\x80\xBF\xA5\x01\x00\xBE\xA5\x01\x00\xBD\xB5\x01\x00\xBC\xB5"
-     "\x01\x00"
-     "\xE6\x75\x00\x80\xA3\x65\x1E\x00\xEA\x75\x00\x80\xEE\x75\x00\x80\xA6\x05"
-     "\x1E\x00"
-     "\xF2\x75\x00\x80\xF6\x75\x00\x80\xA5\x1D\x1E\x00\xAA\x19\x1E\x00\xAB\x19"
-     "\x1E\x00"
-     "\xFA\x75\x00\x80\xFE\x75\x00\x80\xAE\xE1\x01\x00\xAF\xE1\x01\x00\xAC\xF1"
-     "\x01\x00"
-     "\xAD\xF1\x01\x00\xA8\xE9\x02\x00\xA9\xE9\x02\x00\xAA\x25\x03\x00\xAB\x3D"
-     "\x03\x00"
-     "\xAC\x25\x03\x00\xAD\x2D\x03\x00\xAE\x25\x03\x00\xAF\x99\x03\x00\x02\x76"
-     "\x00\x80"
-     "\x06\x76\x00\x80\x0A\x76\x00\x80\x0E\x76\x00\x80\x80\x0D\x00\x00\x81\xB1"
-     "\x00\x00"
-     "\x82\xB1\x00\x00\x12\x76\x00\x80\xB8\x95\x03\x00\xB9\x9D\x03\x00\xBA\xA1"
-     "\x03\x00"
-     "\xBB\xA1\x03\x00\xBC\x71\x00\x00\xBD\x71\x00\x00\xBE\x71\x00\x00\xBF\x71"
-     "\x00\x00"
-     "\xB0\xE9\x03\x00\xB1\xE9\x03\x00\xB2\xF9\x03\x00\xB3\xC5\x03\x00\xB4\xDD"
-     "\x03\x00"
-     "\xB5\xB5\x03\x00\xB6\xB1\x03\x00\xB7\xB1\x03\x00\xB3\xAD\x03\x00\x16\x76"
-     "\x00\x80"
-     "\x1A\x76\x00\x80\x22\x76\x00\x80\x26\x76\x00\x80\xB6\xAD\x03\x00\xB5\xAD"
-     "\x03\x00"
-     "\x2A\x76\x00\x80\xBB\x2D\x03\x00\xBA\x2D\x03\x00\x86\x08\x04\x00\x87\xDC"
-     "\x03\x00"
-     "\xBF\x71\x03\x00\xBE\x19\x03\x00\xBD\x11\x03\x00\xBC\x2D\x03\x00\x2E\x76"
-     "\x00\x80"
-     "\xA3\xE9\x03\x00\x32\x76\x00\x80\x36\x76\x00\x80\xA6\xE9\x03\x00\x3A\x76"
-     "\x00\x80"
-     "\x3E\x76\x00\x80\xA5\xE9\x03\x00\xAA\x69\x03\x00\xAB\x69\x03\x00\x42\x76"
-     "\x00\x80"
-     "\x46\x76\x00\x80\xAE\x5D\x03\x00\xAF\x35\x03\x00\xAC\x69\x03\x00\xAD\x55"
-     "\x03\x00"
-     "\x4A\x76\x00\x80\xA6\xAD\x1F\x00\xA5\xAD\x1F\x00\x4E\x76\x00\x80\xA3\x4D"
-     "\x1F\x00"
-     "\x52\x76\x00\x80\x56\x76\x00\x80\x5A\x76\x00\x80\xAF\x29\x1E\x00\xAE\x21"
-     "\x1E\x00"
-     "\xAD\x29\x1E\x00\xAC\x91\x1F\x00\xAB\x89\x1F\x00\xAA\x81\x1F\x00\x5E\x76"
-     "\x00\x80"
-     "\x62\x76\x00\x80\x66\x76\x00\x80\x6A\x76\x00\x80\xBE\x44\x05\x00\x84\x80"
-     "\x04\x00"
-     "\x6E\x76\x00\x80\xE3\xF8\x1E\x00\x72\x76\x00\x80\xE1\xF4\x1E\x00\x76\x76"
-     "\x00\x80"
-     "\xEF\x50\x1E\x00\x7A\x76\x00\x80\x7E\x76\x00\x80\x80\x0D\x00\x00\x81\xE5"
-     "\x00\x00"
-     "\x82\xED\x00\x00\x82\x76\x00\x80\xE1\x00\x1E\x00\xE1\x30\x1C\x00\xE3\x0C"
-     "\x1F\x00"
-     "\xE3\xBC\x1F\x00\x86\x76\x00\x80\x8A\x76\x00\x80\x86\xB8\x04\x00\x87\x5C"
-     "\x05\x00"
-     "\xBE\x20\x07\x00\x92\x76\x00\x80\x96\x76\x00\x80\x9A\x76\x00\x80\x9E\x76"
-     "\x00\x80"
-     "\xA2\x76\x00\x80\xEF\xCC\x1E\x00\xEF\x3C\x1F\x00\xB3\x5D\x01\x00\xA6\x76"
-     "\x00\x80"
-     "\xAA\x76\x00\x80\xAE\x76\x00\x80\xB2\x76\x00\x80\xB6\x19\x01\x00\xB5\x11"
-     "\x01\x00"
-     "\xB6\x76\x00\x80\xBB\x0D\x01\x00\xBA\x0D\x01\x00\xBA\x76\x00\x80\xBE\x76"
-     "\x00\x80"
-     "\xBF\xE9\x00\x00\xBE\xE1\x00\x00\xBD\xF5\x00\x00\xBC\xFD\x00\x00\xA8\x8D"
-     "\x06\x00"
-     "\xA9\x95\x06\x00\xAA\x95\x06\x00\xAB\xCD\x06\x00\xAC\xD9\x06\x00\xAD\xD9"
-     "\x06\x00"
-     "\xAE\xCD\x06\x00\xAF\xC5\x06\x00\x8E\x76\x00\x80\xC2\x76\x00\x80\xC6\x76"
-     "\x00\x80"
-     "\xCA\x76\x00\x80\xCE\x76\x00\x80\xD2\x76\x00\x80\xD6\x76\x00\x80\xDA\x76"
-     "\x00\x80"
-     "\xB8\x7D\x07\x00\xB9\x05\x07\x00\xBA\x0D\x07\x00\xBB\x1D\x07\x00\xBC\x09"
-     "\x07\x00"
-     "\xBD\x35\x07\x00\xBE\x31\x07\x00\xBF\x31\x07\x00\xB0\x89\x06\x00\xB1\x89"
-     "\x06\x00"
-     "\xB2\x69\x07\x00\xB3\x61\x07\x00\xB4\x65\x07\x00\xB5\x6D\x07\x00\xB6\x65"
-     "\x07\x00"
-     "\xB7\x45\x07\x00\xA3\x1D\x06\x00\xDE\x76\x00\x80\xE2\x76\x00\x80\xE6\x76"
-     "\x00\x80"
-     "\xEA\x76\x00\x80\xA6\x59\x06\x00\xA5\x51\x06\x00\xEE\x76\x00\x80\xAB\x4D"
-     "\x06\x00"
-     "\xAA\x4D\x06\x00\xF2\x76\x00\x80\xF6\x76\x00\x80\xAF\xA9\x07\x00\xAE\xA1"
-     "\x07\x00"
-     "\xAD\xB5\x07\x00\xAC\xBD\x07\x00\x80\x69\x00\x00\x81\x69\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xFA\x76\x00\x80\xFE\x76\x00\x80\x02\x77\x00\x80\x87\xA8\x03\x00\x86\x9C"
-     "\x00\x00"
-     "\x06\x77\x00\x80\x0A\x77\x00\x80\x0E\x77\x00\x80\x12\x77\x00\x80\x16\x77"
-     "\x00\x80"
-     "\x1A\x77\x00\x80\x1E\x77\x00\x80\x22\x77\x00\x80\xB3\xAD\x06\x00\x26\x77"
-     "\x00\x80"
-     "\x2A\x77\x00\x80\x2E\x77\x00\x80\x32\x77\x00\x80\xB6\xD1\x06\x00\xB5\xAD"
-     "\x06\x00"
-     "\x36\x77\x00\x80\xBB\xFD\x06\x00\xBA\xF5\x06\x00\x3A\x77\x00\x80\x3E\x77"
-     "\x00\x80"
-     "\xBF\x19\x01\x00\xBE\x11\x01\x00\xBD\xD5\x06\x00\xBC\xD5\x06\x00\x42\x77"
-     "\x00\x80"
-     "\x46\x77\x00\x80\x4A\x77\x00\x80\x4E\x77\x00\x80\x52\x77\x00\x80\xE3\xD8"
-     "\x06\x00"
-     "\x56\x77\x00\x80\xE1\xEC\x05\x00\x5A\x77\x00\x80\xEF\xD4\x04\x00\x5E\x77"
-     "\x00\x80"
-     "\x62\x77\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xD9\x00\x00\x66\x77"
-     "\x00\x80"
-     "\xA8\x25\x06\x00\xA9\x2D\x06\x00\xAA\x25\x06\x00\xAB\x3D\x06\x00\xAC\x25"
-     "\x06\x00"
-     "\xAD\x51\x06\x00\xAE\x51\x06\x00\xAF\x51\x06\x00\x6A\x77\x00\x80\x84\x3C"
-     "\x01\x00"
-     "\xBE\x24\x00\x00\x72\x77\x00\x80\x86\x10\x0F\x00\x87\xC4\x00\x00\x76\x77"
-     "\x00\x80"
-     "\x7A\x77\x00\x80\xB8\xF5\x01\x00\xB9\xFD\x01\x00\xBA\xF5\x01\x00\xBB\x49"
-     "\x01\x00"
-     "\xBC\x59\x01\x00\xBD\x59\x01\x00\xBE\x49\x01\x00\xBF\x41\x01\x00\xB0\x31"
-     "\x06\x00"
-     "\xB1\x39\x06\x00\xB2\x01\x06\x00\xB3\x01\x06\x00\xB4\xD5\x01\x00\xB5\xDD"
-     "\x01\x00"
-     "\xB6\xD5\x01\x00\xB7\xCD\x01\x00\x7E\x77\x00\x80\x82\x77\x00\x80\x86\x77"
-     "\x00\x80"
-     "\xA3\xA1\x05\x00\x8A\x77\x00\x80\xA5\xA1\x05\x00\xA6\xDD\x05\x00\x8E\x77"
-     "\x00\x80"
-     "\x92\x77\x00\x80\x96\x77\x00\x80\xAA\xF9\x05\x00\xAB\xF1\x05\x00\xAC\xD9"
-     "\x05\x00"
-     "\xAD\xD9\x05\x00\xAE\x1D\x02\x00\xAF\x15\x02\x00\xB6\xB1\x02\x00\x9A\x77"
-     "\x00\x80"
-     "\x9E\x77\x00\x80\xB5\xA1\x02\x00\xA2\x77\x00\x80\xB3\xB1\x02\x00\xA6\x77"
-     "\x00\x80"
-     "\xAA\x77\x00\x80\xBE\x71\x03\x00\xBF\x79\x03\x00\xBC\x61\x03\x00\xBD\x61"
-     "\x03\x00"
-     "\xBA\x8D\x02\x00\xBB\x71\x03\x00\xAE\x77\x00\x80\xB2\x77\x00\x80\xB0\x7D"
-     "\x02\x00"
-     "\xB1\x81\x03\x00\xB2\x81\x03\x00\xB3\x81\x03\x00\xB4\x81\x03\x00\xB5\x89"
-     "\x03\x00"
-     "\xB6\xB1\x03\x00\xB7\xB1\x03\x00\xB8\xE1\x03\x00\xB9\xE1\x03\x00\xBA\xE1"
-     "\x03\x00"
-     "\xBB\xE1\x03\x00\xBC\xE1\x03\x00\xBD\xED\x03\x00\xBE\xD9\x03\x00\xBF\xD5"
-     "\x03\x00"
-     "\x86\xE0\x0D\x00\x87\xE8\x0C\x00\xB6\x77\x00\x80\xBA\x77\x00\x80\xBE\x77"
-     "\x00\x80"
-     "\xC2\x77\x00\x80\x1E\x76\x00\x80\xC6\x77\x00\x80\xA8\x35\x02\x00\xA9\x05"
-     "\x02\x00"
-     "\xAA\x0D\x02\x00\xAB\x05\x02\x00\xAC\x1D\x02\x00\xAD\x05\x02\x00\xAE\x0D"
-     "\x02\x00"
-     "\xAF\x05\x02\x00\xA3\xE5\x03\x00\xCA\x77\x00\x80\xCE\x77\x00\x80\xD2\x77"
-     "\x00\x80"
-     "\xD6\x77\x00\x80\xA6\xE5\x03\x00\xA5\xF5\x03\x00\xDA\x77\x00\x80\xAB\x25"
-     "\x02\x00"
-     "\xAA\xD9\x03\x00\x84\x28\x02\x00\xBE\x20\x0D\x00\xAF\x2D\x02\x00\xAE\x25"
-     "\x02\x00"
-     "\xAD\x35\x02\x00\xAC\x35\x02\x00\xE2\x77\x00\x80\xE6\x77\x00\x80\xEA\x77"
-     "\x00\x80"
-     "\xEE\x77\x00\x80\xBE\xDC\x0C\x00\xE1\x90\x01\x00\xF2\x77\x00\x80\xE3\xE8"
-     "\x00\x00"
-     "\xF6\x77\x00\x80\xFA\x77\x00\x80\xFE\x77\x00\x80\xEF\x3C\x00\x00\x80\x7D"
-     "\x00\x00"
-     "\x81\x65\x00\x00\x82\x65\x00\x00\x02\x78\x00\x80\x06\x78\x00\x80\x0A\x78"
-     "\x00\x80"
-     "\x0E\x78\x00\x80\x12\x78\x00\x80\x16\x78\x00\x80\x1A\x78\x00\x80\x1E\x78"
-     "\x00\x80"
-     "\xEF\xC4\x06\x00\x22\x78\x00\x80\xE1\x24\x06\x00\xBF\xA0\x3F\x00\xE3\xF0"
-     "\x06\x00"
-     "\x26\x78\x00\x80\x84\xBC\x0D\x00\x2A\x78\x00\x80\x2E\x78\x00\x80\x32\x78"
-     "\x00\x80"
-     "\x36\x78\x00\x80\x3A\x78\x00\x80\xB3\x71\x02\x00\x3E\x78\x00\x80\xB5\xD1"
-     "\x01\x00"
-     "\xB6\xD1\x01\x00\x42\x78\x00\x80\x46\x78\x00\x80\x4A\x78\x00\x80\xBA\xC9"
-     "\x01\x00"
-     "\xBB\xC9\x01\x00\xBC\xB5\x01\x00\xBD\xBD\x01\x00\xBE\xA9\x01\x00\xBF\xA5"
-     "\x01\x00"
-     "\xA8\xAD\x0D\x00\xA9\xBD\x0D\x00\xAA\xB5\x0D\x00\xAB\x0D\x0E\x00\xAC\x15"
-     "\x0E\x00"
-     "\xAD\x1D\x0E\x00\xAE\x09\x0E\x00\xAF\x01\x0E\x00\xDE\x77\x00\x80\x82\xC1"
-     "\x0F\x00"
-     "\x81\xC9\x0F\x00\x80\xF1\x0F\x00\x4E\x78\x00\x80\x52\x78\x00\x80\x86\x98"
-     "\x00\x00"
-     "\x87\x60\x03\x00\xB8\x3D\x0E\x00\xB9\x8D\x0E\x00\xBA\x85\x0E\x00\xBB\x9D"
-     "\x0E\x00"
-     "\xBC\x85\x0E\x00\xBD\x89\x0E\x00\xBE\x5D\x0F\x00\xBF\x55\x0F\x00\xB0\x41"
-     "\x0E\x00"
-     "\xB1\x41\x0E\x00\xB2\x2D\x0E\x00\xB3\x25\x0E\x00\xB4\x21\x0E\x00\xB5\x2D"
-     "\x0E\x00"
-     "\xB6\x25\x0E\x00\xB7\x05\x0E\x00\xA3\x3D\x0D\x00\x56\x78\x00\x80\x5A\x78"
-     "\x00\x80"
-     "\x5E\x78\x00\x80\x62\x78\x00\x80\xA6\x9D\x0E\x00\xA5\x9D\x0E\x00\x66\x78"
-     "\x00\x80"
-     "\xAB\x85\x0E\x00\xAA\x85\x0E\x00\x6A\x78\x00\x80\x6E\x78\x00\x80\xAF\xE9"
-     "\x0E\x00"
-     "\xAE\xE5\x0E\x00\xAD\xF1\x0E\x00\xAC\xF9\x0E\x00\x72\x78\x00\x80\xB3\x8D"
-     "\x0F\x00"
-     "\x76\x78\x00\x80\x7A\x78\x00\x80\xB6\xB1\x0F\x00\x7E\x78\x00\x80\x82\x78"
-     "\x00\x80"
-     "\xB5\x85\x0F\x00\xBA\x9D\x0F\x00\xBB\x65\x0F\x00\x86\x78\x00\x80\x8A\x78"
-     "\x00\x80"
-     "\xBE\x7D\x0F\x00\xBF\x65\x0F\x00\xBC\x7D\x0F\x00\xBD\x75\x0F\x00\xA3\x49"
-     "\x0E\x00"
-     "\x8E\x78\x00\x80\x92\x78\x00\x80\x96\x78\x00\x80\x9A\x78\x00\x80\xA6\x75"
-     "\x0E\x00"
-     "\xA5\x41\x0E\x00\x9E\x78\x00\x80\xAB\xA1\x0E\x00\xAA\x59\x0E\x00\xA2\x78"
-     "\x00\x80"
-     "\xA6\x78\x00\x80\xAF\xA1\x0E\x00\xAE\xB9\x0E\x00\xAD\xB1\x0E\x00\xAC\xB9"
-     "\x0E\x00"
-     "\xAA\x78\x00\x80\xAE\x78\x00\x80\xB2\x78\x00\x80\xB6\x78\x00\x80\xEF\x84"
-     "\x0C\x00"
-     "\x82\x31\x00\x00\x81\x15\x00\x00\x80\x15\x00\x00\xE3\x58\x0C\x00\xBA\x78"
-     "\x00\x80"
-     "\xE1\xC0\x0E\x00\x84\x20\x01\x00\xBE\x38\x01\x00\xC2\x78\x00\x80\x86\xD8"
-     "\x00\x00"
-     "\x87\x5C\x00\x00\xA8\xC1\x0E\x00\xA9\xC1\x0E\x00\xAA\xC1\x0E\x00\xAB\xC1"
-     "\x0E\x00"
-     "\xAC\xC9\x0E\x00\xAD\xF5\x0E\x00\xAE\xFD\x0E\x00\xAF\xF5\x0E\x00\x84\xA0"
-     "\x00\x00"
-     "\xC6\x78\x00\x80\xCA\x78\x00\x80\xCE\x78\x00\x80\xD2\x78\x00\x80\xD6\x78"
-     "\x00\x80"
-     "\xDA\x78\x00\x80\xDE\x78\x00\x80\xB8\x61\x0E\x00\xB9\x75\x0E\x00\xBA\x71"
-     "\x0E\x00"
-     "\xBB\x71\x0E\x00\xBC\x11\x0E\x00\xBD\x11\x0E\x00\xBE\x11\x0E\x00\xBF\x0D"
-     "\x0E\x00"
-     "\xB0\x9D\x0E\x00\xB1\x65\x0E\x00\xB2\x6D\x0E\x00\xB3\x65\x0E\x00\xB4\x7D"
-     "\x0E\x00"
-     "\xB5\x65\x0E\x00\xB6\x6D\x0E\x00\xB7\x65\x0E\x00\xA8\x35\x02\x00\xA9\x81"
-     "\x02\x00"
-     "\xAA\x81\x02\x00\xAB\x81\x02\x00\xAC\x81\x02\x00\xAD\x89\x02\x00\xAE\xB1"
-     "\x02\x00"
-     "\xAF\xB1\x02\x00\xBF\x28\xDD\x01\xE2\x78\x00\x80\xE6\x78\x00\x80\x84\xE0"
-     "\x05\x00"
-     "\xEA\x78\x00\x80\xEE\x78\x00\x80\xF2\x78\x00\x80\xF6\x78\x00\x80\xB8\x49"
-     "\x03\x00"
-     "\xB9\x49\x03\x00\xBA\x59\x03\x00\xBB\x51\x03\x00\xBC\x79\x03\x00\xBD\x79"
-     "\x03\x00"
-     "\xBE\x19\x03\x00\xBF\x19\x03\x00\xB0\xA1\x02\x00\xB1\xAD\x02\x00\xB2\xA5"
-     "\x02\x00"
-     "\xB3\xB9\x02\x00\xB4\xA9\x02\x00\xB5\x9D\x02\x00\xB6\x95\x02\x00\xB7\x79"
-     "\x03\x00"
-     "\xB3\x09\x02\x00\xFA\x78\x00\x80\xFE\x78\x00\x80\x02\x79\x00\x80\x06\x79"
-     "\x00\x80"
-     "\xB6\x09\x02\x00\xB5\x09\x02\x00\x0A\x79\x00\x80\xBB\x69\x02\x00\xBA\x65"
-     "\x02\x00"
-     "\x0E\x79\x00\x80\x12\x79\x00\x80\xBF\x59\x02\x00\xBE\x6D\x02\x00\xBD\x6D"
-     "\x02\x00"
-     "\xBC\x71\x02\x00\x16\x79\x00\x80\xA3\x4D\x02\x00\x1A\x79\x00\x80\x1E\x79"
-     "\x00\x80"
-     "\xA6\x4D\x02\x00\x22\x79\x00\x80\x26\x79\x00\x80\xA5\x4D\x02\x00\xAA\x21"
-     "\x02\x00"
-     "\xAB\x2D\x02\x00\x2A\x79\x00\x80\x6E\x77\x00\x80\xAE\x29\x02\x00\xAF\x1D"
-     "\x02\x00"
-     "\xAC\x35\x02\x00\xAD\x29\x02\x00\xA8\x75\x02\x00\xA9\xA1\x02\x00\xAA\xA1"
-     "\x02\x00"
-     "\xAB\xA1\x02\x00\xAC\xA1\x02\x00\xAD\xA1\x02\x00\xAE\xA1\x02\x00\xAF\xA1"
-     "\x02\x00"
-     "\x80\xF5\x01\x00\x81\xFD\x01\x00\x82\xC1\x01\x00\xBE\xA0\x05\x00\x32\x79"
-     "\x00\x80"
-     "\x36\x79\x00\x80\x87\x74\x05\x00\x86\x1C\x05\x00\xB8\xCD\x01\x00\xB9\xD5"
-     "\x01\x00"
-     "\xBA\xD5\x01\x00\xBB\xE5\x01\x00\xBC\xFD\x01\x00\xBD\x95\x01\x00\xBE\x91"
-     "\x01\x00"
-     "\xBF\x91\x01\x00\xB0\x45\x01\x00\xB1\x4D\x01\x00\xB2\x59\x01\x00\xB3\x59"
-     "\x01\x00"
-     "\xB4\x49\x01\x00\xB5\x49\x01\x00\xB6\xFD\x01\x00\xB7\xF5\x01\x00\x3A\x79"
-     "\x00\x80"
-     "\x3E\x79\x00\x80\xBE\x84\x04\x00\xEF\x04\x00\x00\x42\x79\x00\x80\x46\x79"
-     "\x00\x80"
-     "\x4A\x79\x00\x80\xEF\x34\x0E\x00\x4E\x79\x00\x80\xE1\x7C\x0F\x00\x52\x79"
-     "\x00\x80"
-     "\xE3\x08\x0E\x00\xEC\x00\x00\x00\xE1\x94\x01\x00\x56\x79\x00\x80\xE3\xC0"
-     "\x00\x00"
-     "\x5A\x79\x00\x80\x5E\x79\x00\x80\x62\x79\x00\x80\xB3\x65\x02\x00\x66\x79"
-     "\x00\x80"
-     "\xB5\x65\x02\x00\xB6\x6D\x02\x00\x6A\x79\x00\x80\x6E\x79\x00\x80\x72\x79"
-     "\x00\x80"
-     "\xBA\xCD\x01\x00\xBB\xC5\x01\x00\xBC\xC1\x01\x00\xBD\xC1\x01\x00\xBE\xC1"
-     "\x01\x00"
-     "\xBF\xC1\x01\x00\xA3\xA9\x05\x00\x2E\x79\x00\x80\x76\x79\x00\x80\x7A\x79"
-     "\x00\x80"
-     "\x7E\x79\x00\x80\xA6\xA1\x05\x00\xA5\xA9\x05\x00\x82\x79\x00\x80\xAB\x09"
-     "\x06\x00"
-     "\xAA\x01\x06\x00\x86\x79\x00\x80\x8A\x79\x00\x80\xAF\x0D\x06\x00\xAE\x0D"
-     "\x06\x00"
-     "\xAD\x0D\x06\x00\xAC\x0D\x06\x00\x8E\x79\x00\x80\x82\x1D\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x92\x79\x00\x80\x96\x79\x00\x80\x9A\x79\x00\x80\x9E\x79"
-     "\x00\x80"
-     "\x86\xC0\x00\x00\x87\x34\x03\x00\xA2\x79\x00\x80\xA6\x79\x00\x80\xAA\x79"
-     "\x00\x80"
-     "\xAE\x79\x00\x80\xB2\x79\x00\x80\xB6\x79\x00\x80\xA8\xB1\x07\x00\xA9\xB1"
-     "\x07\x00"
-     "\xAA\x09\x07\x00\xAB\x09\x07\x00\xAC\x01\x07\x00\xAD\x01\x07\x00\xAE\x01"
-     "\x07\x00"
-     "\xAF\x3D\x07\x00\x84\xA8\x03\x00\xBA\x79\x00\x80\xBE\x79\x00\x80\xC2\x79"
-     "\x00\x80"
-     "\xC6\x79\x00\x80\xCA\x79\x00\x80\xCE\x79\x00\x80\xD2\x79\x00\x80\xB8\x25"
-     "\x07\x00"
-     "\xB9\xCD\x00\x00\xBA\xC5\x00\x00\xBB\xDD\x00\x00\xBC\xCD\x00\x00\xBD\xFD"
-     "\x00\x00"
-     "\xBE\xF5\x00\x00\xBF\x9D\x00\x00\xB0\x49\x07\x00\xB1\x5D\x07\x00\xB2\x29"
-     "\x07\x00"
-     "\xB3\x29\x07\x00\xB4\x39\x07\x00\xB5\x25\x07\x00\xB6\x25\x07\x00\xB7\x1D"
-     "\x07\x00"
-     "\xB3\x81\x06\x00\xD6\x79\x00\x80\xDA\x79\x00\x80\xDE\x79\x00\x80\xE2\x79"
-     "\x00\x80"
-     "\xB6\x81\x06\x00\xB5\x89\x06\x00\xE6\x79\x00\x80\xBB\x99\x06\x00\xBA\x8D"
-     "\x06\x00"
-     "\xEA\x79\x00\x80\xEE\x79\x00\x80\xBF\xE1\x06\x00\xBE\xFD\x06\x00\xBD\xFD"
-     "\x06\x00"
-     "\xBC\xFD\x06\x00\xF2\x79\x00\x80\xF6\x79\x00\x80\xFA\x79\x00\x80\xFE\x79"
-     "\x00\x80"
-     "\x80\x31\x00\x00\x81\x3D\x00\x00\x82\x29\x00\x00\xEF\x88\x04\x00\x02\x7A"
-     "\x00\x80"
-     "\xE1\x04\x06\x00\x06\x7A\x00\x80\xE3\xDC\x06\x00\x0A\x7A\x00\x80\xBE\x74"
-     "\x01\x00"
-     "\x0E\x7A\x00\x80\x12\x7A\x00\x80\xA3\x01\x06\x00\x16\x7A\x00\x80\x86\x28"
-     "\x00\x00"
-     "\x87\x44\x01\x00\x1A\x7A\x00\x80\xA6\x01\x06\x00\xA5\x09\x06\x00\x1E\x7A"
-     "\x00\x80"
-     "\xAB\x19\x06\x00\xAA\x0D\x06\x00\x22\x7A\x00\x80\x26\x7A\x00\x80\xAF\x61"
-     "\x06\x00"
-     "\xAE\x7D\x06\x00\xAD\x7D\x06\x00\xAC\x7D\x06\x00\x2A\x7A\x00\x80\xB3\xB5"
-     "\x01\x00"
-     "\x2E\x7A\x00\x80\x32\x7A\x00\x80\xB6\xAD\x01\x00\x36\x7A\x00\x80\x3A\x7A"
-     "\x00\x80"
-     "\xB5\xA5\x01\x00\xBA\x4D\x01\x00\xBB\x55\x01\x00\x3E\x7A\x00\x80\x42\x7A"
-     "\x00\x80"
-     "\xBE\x4D\x01\x00\xBF\xF1\x00\x00\xBC\x4D\x01\x00\xBD\x45\x01\x00\xA8\x69"
-     "\x02\x00"
-     "\xA9\x69\x02\x00\xAA\x79\x02\x00\xAB\x79\x02\x00\xAC\x69\x02\x00\xAD\x69"
-     "\x02\x00"
-     "\xAE\xBD\x02\x00\xAF\xB5\x02\x00\x84\x00\x3C\x00\x46\x7A\x00\x80\x4A\x7A"
-     "\x00\x80"
-     "\x4E\x7A\x00\x80\x52\x7A\x00\x80\x56\x7A\x00\x80\x5A\x7A\x00\x80\x5E\x7A"
-     "\x00\x80"
-     "\xB8\x51\x03\x00\xB9\x59\x03\x00\xBA\x61\x03\x00\xBB\x61\x03\x00\xBC\x11"
-     "\x03\x00"
-     "\xBD\x11\x03\x00\xBE\x11\x03\x00\xBF\x11\x03\x00\xB0\xCD\x02\x00\xB1\xD5"
-     "\x02\x00"
-     "\xB2\xDD\x02\x00\xB3\xD5\x02\x00\xB4\xCD\x02\x00\xB5\x71\x03\x00\xB6\x71"
-     "\x03\x00"
-     "\xB7\x71\x03\x00\x62\x7A\x00\x80\x66\x7A\x00\x80\x6A\x7A\x00\x80\xA3\xFD"
-     "\x02\x00"
-     "\x6E\x7A\x00\x80\xA5\xED\x02\x00\xA6\xE5\x02\x00\x72\x7A\x00\x80\x76\x7A"
-     "\x00\x80"
-     "\x7A\x7A\x00\x80\xAA\x05\x02\x00\xAB\x1D\x02\x00\xAC\x05\x02\x00\xAD\x0D"
-     "\x02\x00"
-     "\xAE\x05\x02\x00\xAF\xB9\x03\x00\x82\x65\x00\x00\xEF\xE4\x02\x00\x80\x55"
-     "\x00\x00"
-     "\x81\x6D\x00\x00\x87\x14\x03\x00\x86\x74\x3D\x00\xBE\x54\x3C\x00\x82\x7A"
-     "\x00\x80"
-     "\xBE\x78\x00\x80\x86\x7A\x00\x80\x8A\x7A\x00\x80\x8E\x7A\x00\x80\x92\x7A"
-     "\x00\x80"
-     "\xE3\xC4\x02\x00\x96\x7A\x00\x80\xE1\x6C\x01\x00\xA8\x7D\x02\x00\xA9\x81"
-     "\x02\x00"
-     "\xAA\x99\x02\x00\xAB\x91\x02\x00\xAC\xB9\x02\x00\xAD\xB9\x02\x00\xAE\xD9"
-     "\x02\x00"
-     "\xAF\xD5\x02\x00\x9A\x7A\x00\x80\x9E\x7A\x00\x80\xA2\x7A\x00\x80\xA6\x7A"
-     "\x00\x80"
-     "\xAA\x7A\x00\x80\xAE\x7A\x00\x80\xB2\x7A\x00\x80\xB6\x7A\x00\x80\xB8\x71"
-     "\x01\x00"
-     "\xB9\x71\x01\x00\xBA\x71\x01\x00\xBB\x71\x01\x00\xBC\xD1\x01\x00\xBD\xD1"
-     "\x01\x00"
-     "\xBE\xD1\x01\x00\xBF\xCD\x01\x00\xB0\xAD\x02\x00\xB1\xB1\x02\x00\xB2\xB1"
-     "\x02\x00"
-     "\xB3\x85\x02\x00\xB4\x9D\x02\x00\xB5\x51\x01\x00\xB6\x51\x01\x00\xB7\x51"
-     "\x01\x00"
-     "\xE1\xA4\x06\x00\xBA\x7A\x00\x80\xE3\xB4\x07\x00\xBE\x80\x3D\x00\x84\x18"
-     "\x3D\x00"
-     "\xBE\x7A\x00\x80\xC2\x7A\x00\x80\xC6\x7A\x00\x80\xCA\x7A\x00\x80\xCE\x7A"
-     "\x00\x80"
-     "\xD2\x7A\x00\x80\xD6\x7A\x00\x80\xDA\x7A\x00\x80\xDE\x7A\x00\x80\xEF\xC8"
-     "\x06\x00"
-     "\xE2\x7A\x00\x80\x81\x09\x00\x00\x80\x79\x00\x00\xE6\x7A\x00\x80\x82\x19"
-     "\x00\x00"
-     "\xE3\x60\x01\x00\xEA\x7A\x00\x80\xE1\xA0\x01\x00\xF2\x7A\x00\x80\xEF\x6C"
-     "\x00\x00"
-     "\xF6\x7A\x00\x80\x86\xC8\x3C\x00\x87\x0C\x3D\x00\xFA\x7A\x00\x80\xFE\x7A"
-     "\x00\x80"
-     "\x02\x7B\x00\x80\x06\x7B\x00\x80\xA3\x91\x3D\x00\x7E\x7A\x00\x80\xEE\x7A"
-     "\x00\x80"
-     "\x0A\x7B\x00\x80\x0E\x7B\x00\x80\xA6\xB9\x3D\x00\xA5\xB1\x3D\x00\x12\x7B"
-     "\x00\x80"
-     "\xAB\xF5\x3D\x00\xAA\xE9\x3D\x00\x16\x7B\x00\x80\x1A\x7B\x00\x80\xAF\x15"
-     "\x3E\x00"
-     "\xAE\xD9\x3D\x00\xAD\xD1\x3D\x00\xAC\xE5\x3D\x00\x1E\x7B\x00\x80\xB3\xAD"
-     "\x3E\x00"
-     "\x22\x7B\x00\x80\x26\x7B\x00\x80\xB6\xA9\x3E\x00\x2A\x7B\x00\x80\x2E\x7B"
-     "\x00\x80"
-     "\xB5\xAD\x3E\x00\xBA\x49\x3F\x00\xBB\x49\x3F\x00\x32\x7B\x00\x80\x36\x7B"
-     "\x00\x80"
-     "\xBE\x49\x3F\x00\xBF\x49\x3F\x00\xBC\x49\x3F\x00\xBD\x49\x3F\x00\xA3\x69"
-     "\x3E\x00"
-     "\x3A\x7B\x00\x80\x3E\x7B\x00\x80\x42\x7B\x00\x80\x46\x7B\x00\x80\xA6\x6D"
-     "\x3E\x00"
-     "\xA5\x69\x3E\x00\x4A\x7B\x00\x80\xAB\x8D\x3F\x00\xAA\x8D\x3F\x00\x4E\x7B"
-     "\x00\x80"
-     "\x52\x7B\x00\x80\xAF\x8D\x3F\x00\xAE\x8D\x3F\x00\xAD\x8D\x3F\x00\xAC\x8D"
-     "\x3F\x00"
-     "\x56\x7B\x00\x80\x5A\x7B\x00\x80\xBA\x05\x3D\x00\xBB\x19\x3D\x00\xBC\x09"
-     "\x3D\x00"
-     "\xBD\x3D\x3D\x00\xBE\x35\x3D\x00\xBF\xF9\x3E\x00\x80\x4D\x00\x00\x81\x25"
-     "\x00\x00"
-     "\x82\x2D\x00\x00\xB3\x7D\x3D\x00\xBE\x38\x00\x00\xB5\x5D\x3D\x00\xB6\x55"
-     "\x3D\x00"
-     "\x84\x30\x00\x00\xA8\xDD\x3F\x00\xA9\xE1\x3F\x00\xAA\xE1\x3F\x00\xAB\xE1"
-     "\x3F\x00"
-     "\xAC\xE1\x3F\x00\xAD\xED\x3F\x00\xAE\x39\x3E\x00\xAF\x39\x3E\x00\x5E\x7B"
-     "\x00\x80"
-     "\x62\x7B\x00\x80\x86\x88\x00\x00\x87\xAC\x01\x00\x66\x7B\x00\x80\x6A\x7B"
-     "\x00\x80"
-     "\x6E\x7B\x00\x80\x72\x7B\x00\x80\xB8\xD9\x3E\x00\xB9\xD9\x3E\x00\xBA\xE9"
-     "\x3E\x00"
-     "\xBB\xE9\x3E\x00\xBC\xF9\x3E\x00\xBD\xF9\x3E\x00\xBE\xE9\x3E\x00\xBF\xE9"
-     "\x3E\x00"
-     "\xB0\x49\x3E\x00\xB1\x49\x3E\x00\xB2\x59\x3E\x00\xB3\x59\x3E\x00\xB4\x49"
-     "\x3E\x00"
-     "\xB5\x49\x3E\x00\xB6\xFD\x3E\x00\xB7\xE9\x3E\x00\xA8\x5D\x3E\x00\xA9\x61"
-     "\x3E\x00"
-     "\xAA\x79\x3E\x00\xAB\x79\x3E\x00\xAC\x69\x3E\x00\xAD\x69\x3E\x00\xAE\x95"
-     "\x01\x00"
-     "\xAF\x8D\x01\x00\x84\xA0\x01\x00\x76\x7B\x00\x80\x7A\x7B\x00\x80\x7E\x7B"
-     "\x00\x80"
-     "\x82\x7B\x00\x80\x86\x7B\x00\x80\x8A\x7B\x00\x80\x8E\x7B\x00\x80\xB8\xA9"
-     "\x01\x00"
-     "\xB9\xA9\x01\x00\xBA\x4D\x01\x00\xBB\x45\x01\x00\xBC\x5D\x01\x00\xBD\x45"
-     "\x01\x00"
-     "\xBE\x45\x01\x00\xBF\x75\x01\x00\xB0\xFD\x01\x00\xB1\xCD\x01\x00\xB2\xC5"
-     "\x01\x00"
-     "\xB3\xAD\x01\x00\xB4\xB9\x01\x00\xB5\xB9\x01\x00\xB6\xAD\x01\x00\xB7\xA5"
-     "\x01\x00"
-     "\xB3\xF1\x3D\x00\x92\x7B\x00\x80\xBE\x44\x03\x00\xBE\x44\x04\x00\x96\x7B"
-     "\x00\x80"
-     "\xB6\xE9\x3D\x00\xB5\xE1\x3D\x00\x9A\x7B\x00\x80\xBB\x09\x02\x00\xBA\x39"
-     "\x02\x00"
-     "\x9E\x7B\x00\x80\xA2\x7B\x00\x80\xBF\x7D\x02\x00\xBE\x7D\x02\x00\xBD\x79"
-     "\x02\x00"
-     "\xBC\x11\x02\x00\xA6\x7B\x00\x80\xA3\xB5\x3D\x00\xAA\x7B\x00\x80\xAE\x7B"
-     "\x00\x80"
-     "\xA6\xAD\x3D\x00\xB2\x7B\x00\x80\xB6\x7B\x00\x80\xA5\xA5\x3D\x00\xAA\x7D"
-     "\x02\x00"
-     "\xAB\x4D\x02\x00\xBA\x7B\x00\x80\xBE\x7B\x00\x80\xAE\x39\x02\x00\xAF\x39"
-     "\x02\x00"
-     "\xAC\x55\x02\x00\xAD\x3D\x02\x00\x80\x11\x00\x00\x81\x11\x00\x00\x82\x11"
-     "\x00\x00"
-     "\xBE\x0C\x06\x00\xC2\x7B\x00\x80\xC6\x7B\x00\x80\xCA\x7B\x00\x80\xEF\x6C"
-     "\x02\x00"
-     "\xD2\x7B\x00\x80\xE1\x94\x01\x00\xD6\x7B\x00\x80\xE3\x1C\x01\x00\xDA\x7B"
-     "\x00\x80"
-     "\xDE\x7B\x00\x80\x86\x18\x04\x00\x87\x70\x03\x00\xE2\x7B\x00\x80\xE6\x7B"
-     "\x00\x80"
-     "\xEA\x7B\x00\x80\xB3\xF9\x03\x00\xEE\x7B\x00\x80\xB5\xE9\x03\x00\xB6\x7D"
-     "\x03\x00"
-     "\xF2\x7B\x00\x80\xBE\xAC\x07\x00\xF6\x7B\x00\x80\xBA\x59\x03\x00\xBB\x59"
-     "\x03\x00"
-     "\xBC\x49\x03\x00\xBD\x49\x03\x00\xBE\xFD\x00\x00\xBF\xED\x00\x00\xA8\x25"
-     "\x02\x00"
-     "\xA9\x35\x02\x00\xAA\x3D\x02\x00\xAB\x35\x02\x00\xAC\x2D\x02\x00\xAD\xB1"
-     "\x02\x00"
-     "\xAE\xB1\x02\x00\xAF\xB1\x02\x00\xBE\x68\x09\x00\xFA\x7B\x00\x80\xFE\x7B"
-     "\x00\x80"
-     "\x02\x7C\x00\x80\xBE\xF8\x0A\x00\x06\x7C\x00\x80\x0A\x7C\x00\x80\x0E\x7C"
-     "\x00\x80"
-     "\xB8\x6D\x01\x00\xB9\x05\x01\x00\xBA\x01\x01\x00\xBB\x01\x01\x00\xBC\x05"
-     "\x01\x00"
-     "\xBD\x0D\x01\x00\xBE\x31\x01\x00\xBF\x31\x01\x00\xB0\xD1\x02\x00\xB1\xD1"
-     "\x02\x00"
-     "\xB2\xD1\x02\x00\xB3\xD1\x02\x00\xB4\x75\x01\x00\xB5\x65\x01\x00\xB6\x65"
-     "\x01\x00"
-     "\xB7\x55\x01\x00\xE1\x4C\x3F\x00\xE3\x68\x01\x00\xE3\x08\x3E\x00\xE1\x7C"
-     "\x3E\x00"
-     "\x12\x7C\x00\x80\x16\x7C\x00\x80\x1A\x7C\x00\x80\x1E\x7C\x00\x80\xBE\x2C"
-     "\x0D\x00"
-     "\x22\x7C\x00\x80\x26\x7C\x00\x80\x2A\x7C\x00\x80\xBE\x3C\x0C\x00\xEF\xBC"
-     "\x01\x00"
-     "\xEF\xCC\x3E\x00\x2E\x7C\x00\x80\xA3\xF5\x02\x00\x82\x2D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x32\x7C\x00\x80\xA6\x71\x02\x00\xA5\xE5\x02\x00\x36\x7C"
-     "\x00\x80"
-     "\xAB\x55\x02\x00\xAA\x55\x02\x00\x3A\x7C\x00\x80\x84\xA0\x05\x00\xAF\xE1"
-     "\x01\x00"
-     "\xAE\xF1\x01\x00\xAD\x45\x02\x00\xAC\x45\x02\x00\xA8\xA1\x06\x00\xA9\xAD"
-     "\x06\x00"
-     "\xAA\xB9\x06\x00\xAB\xB9\x06\x00\xAC\xD1\x06\x00\xAD\xD1\x06\x00\xAE\xD1"
-     "\x06\x00"
-     "\xAF\xCD\x06\x00\x3E\x7C\x00\x80\x42\x7C\x00\x80\x86\x08\x03\x00\x87\x08"
-     "\x03\x00"
-     "\x46\x7C\x00\x80\x4A\x7C\x00\x80\x4E\x7C\x00\x80\x52\x7C\x00\x80\xB8\xD1"
-     "\x07\x00"
-     "\xB9\xD1\x07\x00\xBA\xD1\x07\x00\xBB\xD1\x07\x00\xBC\xF5\x07\x00\xBD\xFD"
-     "\x07\x00"
-     "\xBE\xF5\x07\x00\xBF\xE1\x07\x00\xB0\xBD\x06\x00\xB1\x81\x06\x00\xB2\x4D"
-     "\x07\x00"
-     "\xB3\x45\x07\x00\xB4\x5D\x07\x00\xB5\x41\x07\x00\xB6\x41\x07\x00\xB7\x41"
-     "\x07\x00"
-     "\xB3\x31\x06\x00\x56\x7C\x00\x80\x5A\x7C\x00\x80\x5E\x7C\x00\x80\x62\x7C"
-     "\x00\x80"
-     "\xB6\x51\x06\x00\xB5\x59\x06\x00\x66\x7C\x00\x80\xBB\x45\x06\x00\xBA\x79"
-     "\x06\x00"
-     "\x6A\x7C\x00\x80\x6E\x7C\x00\x80\xBF\x85\x07\x00\xBE\x99\x07\x00\xBD\x91"
-     "\x07\x00"
-     "\xBC\x99\x07\x00\x72\x7C\x00\x80\xA3\x75\x06\x00\x76\x7C\x00\x80\x7A\x7C"
-     "\x00\x80"
-     "\xA6\x15\x06\x00\x7E\x7C\x00\x80\x82\x7C\x00\x80\xA5\x1D\x06\x00\xAA\x3D"
-     "\x06\x00"
-     "\xAB\x01\x06\x00\x86\x7C\x00\x80\x8A\x7C\x00\x80\xAE\xDD\x07\x00\xAF\xC1"
-     "\x07\x00"
-     "\xAC\xDD\x07\x00\xAD\xD5\x07\x00\xE1\xA0\x05\x00\x8E\x7C\x00\x80\xE3\xF4"
-     "\x04\x00"
-     "\x92\x7C\x00\x80\x96\x7C\x00\x80\x9A\x7C\x00\x80\x9E\x7C\x00\x80\xA2\x7C"
-     "\x00\x80"
-     "\x81\x09\x00\x00\x80\x39\x00\x00\xA6\x7C\x00\x80\x82\x1D\x00\x00\xAA\x7C"
-     "\x00\x80"
-     "\xAE\x7C\x00\x80\xEF\xCC\x05\x00\x84\xB0\x02\x00\xB3\x75\x01\x00\xB2\x7C"
-     "\x00\x80"
-     "\x86\xE8\x00\x00\x87\x2C\x01\x00\xB6\x7C\x00\x80\xB6\xD1\x01\x00\xB5\x65"
-     "\x01\x00"
-     "\xBA\x7C\x00\x80\xBB\xFD\x01\x00\xBA\xF5\x01\x00\xBE\x7C\x00\x80\xC2\x7C"
-     "\x00\x80"
-     "\xBF\xB9\x01\x00\xBE\xB5\x01\x00\xBD\xD5\x01\x00\xBC\xD5\x01\x00\xA8\xA5"
-     "\x01\x00"
-     "\xA9\xB5\x01\x00\xAA\xBD\x01\x00\xAB\xB5\x01\x00\xAC\xAD\x01\x00\xAD\x11"
-     "\x01\x00"
-     "\xAE\x11\x01\x00\xAF\x11\x01\x00\x84\xAC\x01\x00\xC6\x7C\x00\x80\xCA\x7C"
-     "\x00\x80"
-     "\xCE\x7C\x00\x80\xD2\x7C\x00\x80\xD6\x7C\x00\x80\xDA\x7C\x00\x80\xDE\x7C"
-     "\x00\x80"
-     "\xB8\x19\x01\x00\xB9\x2D\x01\x00\xBA\x25\x01\x00\xBB\xC9\x00\x00\xBC\xDD"
-     "\x00\x00"
-     "\xBD\xC5\x00\x00\xBE\xCD\x00\x00\xBF\xC5\x00\x00\xB0\x71\x01\x00\xB1\x79"
-     "\x01\x00"
-     "\xB2\x41\x01\x00\xB3\x41\x01\x00\xB4\x31\x01\x00\xB5\x3D\x01\x00\xB6\x35"
-     "\x01\x00"
-     "\xB7\x29\x01\x00\xE2\x7C\x00\x80\xE6\x7C\x00\x80\xEA\x7C\x00\x80\xA3\x39"
-     "\x02\x00"
-     "\xEE\x7C\x00\x80\xA5\x29\x02\x00\xA6\x9D\x02\x00\xF2\x7C\x00\x80\xF6\x7C"
-     "\x00\x80"
-     "\xFA\x7C\x00\x80\xAA\xB9\x02\x00\xAB\xB1\x02\x00\xAC\x99\x02\x00\xAD\x99"
-     "\x02\x00"
-     "\xAE\xF9\x02\x00\xAF\xF5\x02\x00\xFE\x7C\x00\x80\x02\x7D\x00\x80\x86\xE8"
-     "\x0C\x00"
-     "\x87\x68\x03\x00\x06\x7D\x00\x80\x0A\x7D\x00\x80\x0E\x7D\x00\x80\x12\x7D"
-     "\x00\x80"
-     "\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x16\x7D\x00\x80\x1A\x7D"
-     "\x00\x80"
-     "\x22\x7D\x00\x80\x26\x7D\x00\x80\x2A\x7D\x00\x80\x2E\x7D\x00\x80\xCE\x7B"
-     "\x00\x80"
-     "\xEF\x90\x06\x00\x32\x7D\x00\x80\x36\x7D\x00\x80\x3A\x7D\x00\x80\x3E\x7D"
-     "\x00\x80"
-     "\x42\x7D\x00\x80\x84\xEC\x0C\x00\x46\x7D\x00\x80\x4A\x7D\x00\x80\x4E\x7D"
-     "\x00\x80"
-     "\xE1\xE0\x01\x00\x52\x7D\x00\x80\xE3\x08\x01\x00\x56\x7D\x00\x80\xB3\x9D"
-     "\x03\x00"
-     "\x5A\x7D\x00\x80\x5E\x7D\x00\x80\x62\x7D\x00\x80\x66\x7D\x00\x80\xB6\x71"
-     "\x03\x00"
-     "\xB5\xB5\x03\x00\x6A\x7D\x00\x80\xBB\x55\x03\x00\xBA\x55\x03\x00\x6E\x7D"
-     "\x00\x80"
-     "\x72\x7D\x00\x80\xBF\xF5\x00\x00\xBE\xF5\x00\x00\xBD\x45\x03\x00\xBC\x45"
-     "\x03\x00"
-     "\xA8\x45\x02\x00\xA9\x55\x02\x00\xAA\x55\x02\x00\xAB\x65\x02\x00\xAC\x7D"
-     "\x02\x00"
-     "\xAD\xB1\x02\x00\xAE\xB1\x02\x00\xAF\xB1\x02\x00\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\x76\x7D\x00\x80\x82\xAD\x00\x00\x7A\x7D\x00\x80\x7E\x7D\x00\x80\x82\x7D"
-     "\x00\x80"
-     "\x86\x7D\x00\x80\xB8\x5D\x01\x00\xB9\x65\x01\x00\xBA\x6D\x01\x00\xBB\x0D"
-     "\x01\x00"
-     "\xBC\x15\x01\x00\xBD\x05\x01\x00\xBE\x0D\x01\x00\xBF\x05\x01\x00\xB0\xD1"
-     "\x02\x00"
-     "\xB1\xD1\x02\x00\xB2\xD1\x02\x00\xB3\xD1\x02\x00\xB4\x71\x01\x00\xB5\x71"
-     "\x01\x00"
-     "\xB6\x71\x01\x00\xB7\x6D\x01\x00\x8A\x7D\x00\x80\xE1\xD0\x2B\x00\x92\x7D"
-     "\x00\x80"
-     "\xE3\xD8\x17\x00\x86\x90\x0F\x00\x87\x5C\x0D\x00\x96\x7D\x00\x80\xEF\xE8"
-     "\x07\x00"
-     "\x9A\x7D\x00\x80\xE1\x94\x01\x00\x9E\x7D\x00\x80\xE3\x80\x00\x00\xBE\xB0"
-     "\x0B\x00"
-     "\xA2\x7D\x00\x80\xA6\x7D\x00\x80\xEF\x60\x3F\x00\xA3\x9D\x02\x00\xAA\x7D"
-     "\x00\x80"
-     "\xAE\x7D\x00\x80\xB2\x7D\x00\x80\xB6\x7D\x00\x80\xA6\x71\x02\x00\xA5\xB5"
-     "\x02\x00"
-     "\xBA\x7D\x00\x80\xAB\x55\x02\x00\xAA\x55\x02\x00\xBE\x7D\x00\x80\xC2\x7D"
-     "\x00\x80"
-     "\xAF\xF5\x01\x00\xAE\xF5\x01\x00\xAD\x45\x02\x00\xAC\x45\x02\x00\xA8\x75"
-     "\x0E\x00"
-     "\xA9\x7D\x0E\x00\xAA\x75\x0E\x00\xAB\x8D\x0E\x00\xAC\x9D\x0E\x00\xAD\x85"
-     "\x0E\x00"
-     "\xAE\x8D\x0E\x00\xAF\x85\x0E\x00\x8E\x7D\x00\x80\xC6\x7D\x00\x80\xCA\x7D"
-     "\x00\x80"
-     "\xCE\x7D\x00\x80\xD2\x7D\x00\x80\xD6\x7D\x00\x80\xDA\x7D\x00\x80\xDE\x7D"
-     "\x00\x80"
-     "\xB8\x8D\x0E\x00\xB9\x95\x0E\x00\xBA\x95\x0E\x00\xBB\xA5\x0E\x00\xBC\xBD"
-     "\x0E\x00"
-     "\xBD\x7D\x0F\x00\xBE\x69\x0F\x00\xBF\x69\x0F\x00\xB0\xFD\x0E\x00\xB1\xC1"
-     "\x0E\x00"
-     "\xB2\xC1\x0E\x00\xB3\xC1\x0E\x00\xB4\xC5\x0E\x00\xB5\xC9\x0E\x00\xB6\xBD"
-     "\x0E\x00"
-     "\xB7\xB5\x0E\x00\xB3\xFD\x0E\x00\xE2\x7D\x00\x80\xE6\x7D\x00\x80\xEA\x7D"
-     "\x00\x80"
-     "\xEE\x7D\x00\x80\xB6\x05\x0E\x00\xB5\x15\x0E\x00\xF2\x7D\x00\x80\xBB\x05"
-     "\x0E\x00"
-     "\xBA\x39\x0E\x00\xF6\x7D\x00\x80\xFA\x7D\x00\x80\xBF\x0D\x0E\x00\xBE\x01"
-     "\x0E\x00"
-     "\xBD\x15\x0E\x00\xBC\x15\x0E\x00\x82\x6D\x00\x00\xA3\xB9\x0E\x00\x80\x55"
-     "\x00\x00"
-     "\x81\x65\x00\x00\xA6\x41\x0E\x00\x84\x18\x01\x00\x02\x7E\x00\x80\xA5\x51"
-     "\x0E\x00"
-     "\xAA\x7D\x0E\x00\xAB\x41\x0E\x00\x86\x00\x04\x00\x87\x78\x00\x00\xAE\x45"
-     "\x0E\x00"
-     "\xAF\x49\x0E\x00\xAC\x51\x0E\x00\xAD\x51\x0E\x00\x8F\x29\x25\x00\xE1\xF4"
-     "\x03\x00"
-     "\x06\x7E\x00\x80\xE3\x54\x01\x00\x0A\x7E\x00\x80\xE3\xA4\x0D\x00\x0E\x7E"
-     "\x00\x80"
-     "\xE1\x4C\x0D\x00\x12\x7E\x00\x80\xEF\xFC\x08\x00\x16\x7E\x00\x80\x1A\x7E"
-     "\x00\x80"
-     "\x1E\x7E\x00\x80\x22\x7E\x00\x80\x26\x7E\x00\x80\xEF\x64\x08\x00\x9F\x61"
-     "\x08\x00"
-     "\x9E\x4D\x09\x00\x9D\x4D\x09\x00\x9C\x35\x34\x00\x9B\x0D\x35\x00\x9A\x51"
-     "\x35\x00"
-     "\x99\xB9\x37\x00\x98\x6D\x31\x00\x97\x41\x31\x00\x96\x89\x33\x00\x95\xB9"
-     "\x3D\x00"
-     "\x94\x4D\x3D\x00\x93\xA9\x3F\x00\x92\xAD\x39\x00\x91\x7D\x39\x00\x90\x7D"
-     "\x39\x00"
-     "\xB1\xA5\x15\x00\xB0\xA5\x15\x00\xB3\x89\xE9\x01\xB2\x41\x14\x00\xB5\x55"
-     "\xE8\x01"
-     "\xB4\x89\xE9\x01\x2A\x7E\x00\x80\x2E\x7E\x00\x80\x32\x7E\x00\x80\x36\x7E"
-     "\x00\x80"
-     "\x3A\x7E\x00\x80\x3E\x7E\x00\x80\xA3\x29\x03\x00\xA2\x81\x0C\x00\xA1\x99"
-     "\x0D\x00"
-     "\xA0\x39\x0F\x00\xA1\xE1\x01\x00\x42\x7E\x00\x80\xA3\xCD\x05\x00\xA2\x4D"
-     "\x03\x00"
-     "\xA5\x81\x04\x00\xA4\xDD\x05\x00\xA7\xD5\x19\x00\xA6\xDD\x19\x00\xA9\xE1"
-     "\x1D\x00"
-     "\xA8\x8D\x18\x00\xAB\xE5\x1C\x00\xAA\xE9\x1D\x00\xAD\xD1\x11\x00\xAC\x01"
-     "\x10\x00"
-     "\xAF\x01\x14\x00\xAE\x51\x11\x00\x82\x29\x26\x00\x83\x05\x3A\x00\x46\x7E"
-     "\x00\x80"
-     "\x4A\x7E\x00\x80\x86\xC9\x3C\x00\x87\x35\x3E\x00\x84\x1D\x3A\x00\x85\xC5"
-     "\x3A\x00"
-     "\x8A\xF1\x3E\x00\x8B\xC9\x30\x00\x4E\x7E\x00\x80\x52\x7E\x00\x80\x8E\xD1"
-     "\x34\x00"
-     "\x8F\x31\x36\x00\x8C\x19\x32\x00\x8D\x0D\x32\x00\x92\x1D\x36\x00\x93\x3D"
-     "\x37\x00"
-     "\x84\x44\x03\x00\x84\xC4\x04\x00\x96\x71\x0B\x00\x97\xD1\x0C\x00\x94\x5D"
-     "\x0A\x00"
-     "\x95\x55\x0A\x00\x9A\x41\x0E\x00\x9B\x51\x0F\x00\x56\x7E\x00\x80\x5A\x7E"
-     "\x00\x80"
-     "\x5E\x7E\x00\x80\x62\x7E\x00\x80\x9C\xE1\x00\x00\x66\x7E\x00\x80\xFE\x7D"
-     "\x00\x80"
-     "\x6A\x7E\x00\x80\x72\x7E\x00\x80\x76\x7E\x00\x80\x81\xED\x00\x00\x80\xE1"
-     "\x00\x00"
-     "\x1E\x7D\x00\x80\x82\x19\x00\x00\xA8\x29\x0D\x00\xA9\x3D\x0D\x00\xAA\x09"
-     "\x0D\x00"
-     "\xAB\x05\x0D\x00\xAC\x05\x0D\x00\xAD\x0D\x0D\x00\xAE\x7D\x0D\x00\xAF\x75"
-     "\x0D\x00"
-     "\xB0\x0D\x0D\x00\xB1\x15\x0D\x00\xB2\x15\x0D\x00\xB3\x25\x0D\x00\xB4\x3D"
-     "\x0D\x00"
-     "\xB5\xF1\x02\x00\xB6\xF1\x02\x00\xB7\xF1\x02\x00\xB8\xD1\x02\x00\xB9\xD1"
-     "\x02\x00"
-     "\xBA\xD1\x02\x00\xBB\xD1\x02\x00\xBC\x75\x02\x00\xBD\x7D\x02\x00\xBE\x75"
-     "\x02\x00"
-     "\xBF\x6D\x02\x00\xB3\xA1\x0C\x00\x7A\x7E\x00\x80\x87\xB4\x02\x00\x86\x2C"
-     "\x05\x00"
-     "\x7E\x7E\x00\x80\xB6\xAD\x0C\x00\xB5\xAD\x0C\x00\x82\x7E\x00\x80\xBB\xF1"
-     "\x0C\x00"
-     "\xBA\xE9\x0C\x00\x86\x7E\x00\x80\x8A\x7E\x00\x80\xBF\x15\x03\x00\xBE\xD9"
-     "\x0C\x00"
-     "\xBD\xD1\x0C\x00\xBC\xE1\x0C\x00\x8E\x7E\x00\x80\xA3\xE5\x0C\x00\x92\x7E"
-     "\x00\x80"
-     "\x96\x7E\x00\x80\xA6\xE9\x0C\x00\x9A\x7E\x00\x80\x9E\x7E\x00\x80\xA5\xE9"
-     "\x0C\x00"
-     "\xAA\xAD\x0C\x00\xAB\xB5\x0C\x00\xA2\x7E\x00\x80\xA6\x7E\x00\x80\xAE\x9D"
-     "\x0C\x00"
-     "\xAF\x51\x03\x00\xAC\xA5\x0C\x00\xAD\x95\x0C\x00\xAA\x7E\x00\x80\xAE\x7E"
-     "\x00\x80"
-     "\xB2\x7E\x00\x80\xB6\x7E\x00\x80\xBA\x7E\x00\x80\xBE\x7E\x00\x80\xC2\x7E"
-     "\x00\x80"
-     "\xEF\x54\x0A\x00\xC6\x7E\x00\x80\xE1\x2C\x0E\x00\xCA\x7E\x00\x80\xE3\x68"
-     "\x01\x00"
-     "\xCE\x7E\x00\x80\x82\x79\x00\x00\x81\x71\x00\x00\x80\x79\x00\x00\xB3\xD1"
-     "\x02\x00"
-     "\xD2\x7E\x00\x80\xD6\x7E\x00\x80\xDA\x7E\x00\x80\xE2\x7E\x00\x80\xB6\xFD"
-     "\x02\x00"
-     "\xB5\xF9\x02\x00\xE6\x7E\x00\x80\xBB\xD9\x02\x00\xBA\xD1\x02\x00\x86\xC8"
-     "\x04\x00"
-     "\x87\xAC\x04\x00\xBF\x55\x02\x00\xBE\x55\x02\x00\xBD\x65\x02\x00\xBC\x65"
-     "\x02\x00"
-     "\xA3\x11\x05\x00\x6E\x7E\x00\x80\xDE\x7E\x00\x80\xEA\x7E\x00\x80\xEE\x7E"
-     "\x00\x80"
-     "\xA6\x3D\x05\x00\xA5\x39\x05\x00\xF2\x7E\x00\x80\xAB\x19\x05\x00\xAA\x11"
-     "\x05\x00"
-     "\xF6\x7E\x00\x80\xFA\x7E\x00\x80\xAF\x95\x05\x00\xAE\x95\x05\x00\xAD\xA5"
-     "\x05\x00"
-     "\xAC\xA5\x05\x00\xFE\x7E\x00\x80\xB3\x41\x07\x00\x02\x7F\x00\x80\x06\x7F"
-     "\x00\x80"
-     "\xB6\x79\x07\x00\x0A\x7F\x00\x80\x0E\x7F\x00\x80\xB5\x41\x07\x00\xBA\x29"
-     "\x07\x00"
-     "\xBB\x29\x07\x00\x12\x7F\x00\x80\x16\x7F\x00\x80\xBE\x19\x07\x00\xBF\x19"
-     "\x07\x00"
-     "\xBC\x2D\x07\x00\xBD\x25\x07\x00\xA8\x09\x06\x00\xA9\x09\x06\x00\xAA\x8D"
-     "\x07\x00"
-     "\xAB\x9D\x07\x00\xAC\x8D\x07\x00\xAD\xBD\x07\x00\xAE\xB5\x07\x00\xAF\xDD"
-     "\x07\x00"
-     "\x1A\x7F\x00\x80\x1E\x7F\x00\x80\x22\x7F\x00\x80\x26\x7F\x00\x80\x2A\x7F"
-     "\x00\x80"
-     "\x2E\x7F\x00\x80\x32\x7F\x00\x80\x36\x7F\x00\x80\xB8\x49\x07\x00\xB9\x49"
-     "\x07\x00"
-     "\xBA\x59\x07\x00\xBB\x59\x07\x00\xBC\x49\x07\x00\xBD\x75\x07\x00\xBE\x71"
-     "\x07\x00"
-     "\xBF\x6D\x07\x00\xB0\xA9\x07\x00\xB1\xA9\x07\x00\xB2\xBD\x07\x00\xB3\xB5"
-     "\x07\x00"
-     "\xB4\x99\x07\x00\xB5\x99\x07\x00\xB6\x79\x07\x00\xB7\x79\x07\x00\xA3\x05"
-     "\x06\x00"
-     "\x3A\x7F\x00\x80\x3E\x7F\x00\x80\x42\x7F\x00\x80\x46\x7F\x00\x80\xA6\x3D"
-     "\x06\x00"
-     "\xA5\x05\x06\x00\x4A\x7F\x00\x80\xAB\x6D\x06\x00\xAA\x6D\x06\x00\x84\xA8"
-     "\x02\x00"
-     "\x52\x7F\x00\x80\xAF\x5D\x06\x00\xAE\x5D\x06\x00\xAD\x61\x06\x00\xAC\x69"
-     "\x06\x00"
-     "\x80\x4D\x00\x00\x81\x55\x00\x00\x82\x55\x00\x00\xB3\x41\x01\x00\x4E\x7F"
-     "\x00\x80"
-     "\xB5\x49\x01\x00\xB6\x71\x01\x00\x56\x7F\x00\x80\x86\xC0\x00\x00\x87\xA8"
-     "\x01\x00"
-     "\xBA\x21\x01\x00\xBB\x21\x01\x00\xBC\x25\x01\x00\xBD\x2D\x01\x00\xBE\x11"
-     "\x01\x00"
-     "\xBF\x11\x01\x00\xA8\x51\x06\x00\xA9\x59\x06\x00\xAA\x85\x01\x00\xAB\x9D"
-     "\x01\x00"
-     "\xAC\x8D\x01\x00\xAD\xBD\x01\x00\xAE\xB5\x01\x00\xAF\xDD\x01\x00\x84\xA0"
-     "\x00\x00"
-     "\x5A\x7F\x00\x80\x5E\x7F\x00\x80\x62\x7F\x00\x80\x66\x7F\x00\x80\x6A\x7F"
-     "\x00\x80"
-     "\x6E\x7F\x00\x80\x72\x7F\x00\x80\xB8\x69\x01\x00\xB9\x69\x01\x00\xBA\x79"
-     "\x01\x00"
-     "\xBB\x79\x01\x00\xBC\x69\x01\x00\xBD\x69\x01\x00\xBE\xDD\x00\x00\xBF\xD5"
-     "\x00\x00"
-     "\xB0\xA9\x01\x00\xB1\xA9\x01\x00\xB2\xBD\x01\x00\xB3\xB5\x01\x00\xB4\x99"
-     "\x01\x00"
-     "\xB5\x99\x01\x00\xB6\x59\x01\x00\xB7\x59\x01\x00\xBA\x0D\x04\x00\xBB\x2D"
-     "\x04\x00"
-     "\xB8\xCD\x07\x00\xB9\x05\x04\x00\xBE\x7D\x04\x00\xBF\x65\x04\x00\xBC\x35"
-     "\x04\x00"
-     "\xBD\x75\x04\x00\xB2\xB9\x07\x00\xB3\xB9\x07\x00\xB0\x1D\x07\x00\xB1\x15"
-     "\x07\x00"
-     "\xB6\x99\x07\x00\xB7\x99\x07\x00\xB4\x99\x07\x00\xB5\x99\x07\x00\xAA\x51"
-     "\x07\x00"
-     "\xAB\x5D\x07\x00\xA8\x91\x06\x00\xA9\x45\x07\x00\xAE\x7D\x07\x00\xAF\x65"
-     "\x07\x00"
-     "\xAC\x79\x07\x00\xAD\x75\x07\x00\x76\x7F\x00\x80\x7A\x7F\x00\x80\x7E\x7F"
-     "\x00\x80"
-     "\x82\x7F\x00\x80\x86\x7F\x00\x80\x8A\x7F\x00\x80\x8E\x7F\x00\x80\x92\x7F"
-     "\x00\x80"
-     "\x96\x7F\x00\x80\x9A\x7F\x00\x80\xBE\x00\x1C\x00\xA3\x81\x02\x00\x9E\x7F"
-     "\x00\x80"
-     "\xA5\x89\x02\x00\xA6\xB1\x02\x00\xA2\x7F\x00\x80\xA6\x7F\x00\x80\xAA\x7F"
-     "\x00\x80"
-     "\xAA\xE1\x02\x00\xAB\xE1\x02\x00\xAC\xE5\x02\x00\xAD\xED\x02\x00\xAE\xD1"
-     "\x02\x00"
-     "\xAF\xD1\x02\x00\xAE\x7F\x00\x80\x82\x25\x00\x00\x81\x11\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xB2\x7F\x00\x80\xE1\xAC\x01\x00\xB6\x7F\x00\x80\xE3\x6C\x01\x00\xBA\x7F"
-     "\x00\x80"
-     "\xC2\x7F\x00\x80\xC6\x7F\x00\x80\xEF\x9C\x02\x00\x87\xB0\x1C\x00\x86\xD4"
-     "\x1C\x00"
-     "\xCA\x7F\x00\x80\xCE\x7F\x00\x80\xA8\x2D\x02\x00\xA9\x59\x02\x00\xAA\xAD"
-     "\x02\x00"
-     "\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xA5\x02\x00\xAE\xAD\x02\x00\xAF\xA5"
-     "\x02\x00"
-     "\x84\xA0\x1D\x00\xD2\x7F\x00\x80\xD6\x7F\x00\x80\xDA\x7F\x00\x80\xDE\x7F"
-     "\x00\x80"
-     "\xE2\x7F\x00\x80\xBE\xB4\x1C\x00\xE6\x7F\x00\x80\xB8\x3D\x01\x00\xB9\xC5"
-     "\x01\x00"
-     "\xBA\xCD\x01\x00\xBB\xDD\x01\x00\xBC\xCD\x01\x00\xBD\xF5\x01\x00\xBE\xFD"
-     "\x01\x00"
-     "\xBF\x9D\x01\x00\xB0\xDD\x02\x00\xB1\x05\x01\x00\xB2\x0D\x01\x00\xB3\x05"
-     "\x01\x00"
-     "\xB4\x1D\x01\x00\xB5\x05\x01\x00\xB6\x0D\x01\x00\xB7\x05\x01\x00\xE1\x94"
-     "\x06\x00"
-     "\xE3\xBC\x00\x00\xE3\xD4\x07\x00\xE1\x7C\x06\x00\xEA\x7F\x00\x80\xEE\x7F"
-     "\x00\x80"
-     "\xF2\x7F\x00\x80\xF6\x7F\x00\x80\xFA\x7F\x00\x80\xFE\x7F\x00\x80\x02\x80"
-     "\x00\x80"
-     "\x06\x80\x00\x80\xBE\x78\x1F\x00\xEF\xFC\x0A\x00\xEF\xF0\x0E\x00\x0A\x80"
-     "\x00\x80"
-     "\x0E\x80\x00\x80\x12\x80\x00\x80\x16\x80\x00\x80\xB3\x69\x02\x00\x1A\x80"
-     "\x00\x80"
-     "\xB5\x79\x02\x00\xB6\x69\x02\x00\x1E\x80\x00\x80\x22\x80\x00\x80\x26\x80"
-     "\x00\x80"
-     "\xBA\x55\x02\x00\xBB\x8D\x01\x00\xBC\x95\x01\x00\xBD\x9D\x01\x00\xBE\x95"
-     "\x01\x00"
-     "\xBF\x8D\x01\x00\xBE\x7F\x00\x80\xA3\xA1\x1D\x00\x2A\x80\x00\x80\x2E\x80"
-     "\x00\x80"
-     "\xA6\xA1\x1D\x00\x32\x80\x00\x80\x36\x80\x00\x80\xA5\xB1\x1D\x00\xAA\x9D"
-     "\x1D\x00"
-     "\xAB\x45\x1E\x00\x86\x08\x03\x00\x87\x7C\x03\x00\xAE\x5D\x1E\x00\xAF\x45"
-     "\x1E\x00"
-     "\xAC\x5D\x1E\x00\xAD\x55\x1E\x00\x80\xA1\x1F\x00\x81\xA1\x1F\x00\x82\x41"
-     "\x00\x00"
-     "\xB3\xB5\x1E\x00\x3A\x80\x00\x80\xB5\x71\x1F\x00\xB6\x71\x1F\x00\x3E\x80"
-     "\x00\x80"
-     "\x42\x80\x00\x80\x46\x80\x00\x80\xBA\x5D\x1F\x00\xBB\x51\x1F\x00\xBC\xE1"
-     "\x1F\x00"
-     "\xBD\xE1\x1F\x00\xBE\xE1\x1F\x00\xBF\xE1\x1F\x00\xA8\x65\x1E\x00\xA9\x6D"
-     "\x1E\x00"
-     "\xAA\x71\x1E\x00\xAB\x71\x1E\x00\xAC\xB5\x1F\x00\xAD\xBD\x1F\x00\xAE\xB5"
-     "\x1F\x00"
-     "\xAF\xAD\x1F\x00\x4A\x80\x00\x80\x4E\x80\x00\x80\x52\x80\x00\x80\x56\x80"
-     "\x00\x80"
-     "\x5A\x80\x00\x80\x5E\x80\x00\x80\x62\x80\x00\x80\x66\x80\x00\x80\xB8\x69"
-     "\x1F\x00"
-     "\xB9\x69\x1F\x00\xBA\x75\x1F\x00\xBB\x0D\x1F\x00\xBC\x15\x1F\x00\xBD\x1D"
-     "\x1F\x00"
-     "\xBE\x09\x1F\x00\xBF\x01\x1F\x00\xB0\xD5\x1F\x00\xB1\xD9\x1F\x00\xB2\x79"
-     "\x1F\x00"
-     "\xB3\x79\x1F\x00\xB4\x69\x1F\x00\xB5\x69\x1F\x00\xB6\x59\x1F\x00\xB7\x59"
-     "\x1F\x00"
-     "\xA3\xF9\x1F\x00\x6A\x80\x00\x80\x6E\x80\x00\x80\x72\x80\x00\x80\x76\x80"
-     "\x00\x80"
-     "\xA6\x3D\x1E\x00\xA5\x3D\x1E\x00\x7A\x80\x00\x80\xAB\x1D\x1E\x00\xAA\x11"
-     "\x1E\x00"
-     "\x7E\x80\x00\x80\x82\x80\x00\x80\xAF\xAD\x1E\x00\xAE\xAD\x1E\x00\xAD\xAD"
-     "\x1E\x00"
-     "\xAC\xAD\x1E\x00\x86\x80\x00\x80\x8A\x80\x00\x80\x8E\x80\x00\x80\x92\x80"
-     "\x00\x80"
-     "\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39\x00\x00\x96\x80\x00\x80\x9A\x80"
-     "\x00\x80"
-     "\x84\xA8\x01\x00\x9E\x80\x00\x80\xA2\x80\x00\x80\x86\xD0\x00\x00\x87\xFC"
-     "\x01\x00"
-     "\xA6\x80\x00\x80\xAA\x80\x00\x80\xA8\x95\x01\x00\xA9\x9D\x01\x00\xAA\xD9"
-     "\x01\x00"
-     "\xAB\xD1\x01\x00\xAC\xF1\x01\x00\xAD\xF1\x01\x00\xAE\x39\x01\x00\xAF\x39"
-     "\x01\x00"
-     "\x84\xA0\x00\x00\xAE\x80\x00\x80\xB2\x80\x00\x80\xB6\x80\x00\x80\xBA\x80"
-     "\x00\x80"
-     "\xBE\x80\x00\x80\xC2\x80\x00\x80\xC6\x80\x00\x80\xB8\xCD\x00\x00\xB9\xD5"
-     "\x00\x00"
-     "\xBA\xD5\x00\x00\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x9D"
-     "\x00\x00"
-     "\xBF\x95\x00\x00\xB0\x49\x01\x00\xB1\x49\x01\x00\xB2\x59\x01\x00\xB3\x59"
-     "\x01\x00"
-     "\xB4\x49\x01\x00\xB5\x49\x01\x00\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xB3\x51"
-     "\x02\x00"
-     "\xCA\x80\x00\x80\xCE\x80\x00\x80\xD2\x80\x00\x80\xD6\x80\x00\x80\xB6\x71"
-     "\x02\x00"
-     "\xB5\x79\x02\x00\xDA\x80\x00\x80\xBB\x8D\x02\x00\xBA\x8D\x02\x00\xDE\x80"
-     "\x00\x80"
-     "\xE2\x80\x00\x80\xBF\x8D\x02\x00\xBE\x8D\x02\x00\xBD\x8D\x02\x00\xBC\x8D"
-     "\x02\x00"
-     "\x85\x3C\xF4\x01\xA3\x15\x02\x00\xE6\x80\x00\x80\xEA\x80\x00\x80\xA6\x35"
-     "\x02\x00"
-     "\xEE\x80\x00\x80\xF2\x80\x00\x80\xA5\x3D\x02\x00\xAA\xC9\x02\x00\xAB\xC9"
-     "\x02\x00"
-     "\xF6\x80\x00\x80\xFA\x80\x00\x80\xAE\xC9\x02\x00\xAF\xC9\x02\x00\xAC\xC9"
-     "\x02\x00"
-     "\xAD\xC9\x02\x00\x82\x19\x00\x00\xFE\x80\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x02\x81\x00\x80\x06\x81\x00\x80\x0A\x81\x00\x80\x12\x81\x00\x80\xBE\x00"
-     "\x04\x00"
-     "\x16\x81\x00\x80\x1A\x81\x00\x80\x1E\x81\x00\x80\x22\x81\x00\x80\x26\x81"
-     "\x00\x80"
-     "\x2A\x81\x00\x80\x2E\x81\x00\x80\x86\xE0\x04\x00\x87\x54\x03\x00\x32\x81"
-     "\x00\x80"
-     "\x36\x81\x00\x80\x3A\x81\x00\x80\xEF\x8C\x02\x00\x3E\x81\x00\x80\x42\x81"
-     "\x00\x80"
-     "\x46\x81\x00\x80\xE3\x54\x01\x00\x4A\x81\x00\x80\xE1\x84\x01\x00\x4E\x81"
-     "\x00\x80"
-     "\x52\x81\x00\x80\x56\x81\x00\x80\x5A\x81\x00\x80\xA8\x6D\x02\x00\xA9\x7D"
-     "\x02\x00"
-     "\xAA\x75\x02\x00\xAB\xA9\x02\x00\xAC\xB9\x02\x00\xAD\xB9\x02\x00\xAE\xA9"
-     "\x02\x00"
-     "\xAF\xA9\x02\x00\x84\xAC\x05\x00\x5E\x81\x00\x80\x62\x81\x00\x80\x66\x81"
-     "\x00\x80"
-     "\x6A\x81\x00\x80\x6E\x81\x00\x80\x72\x81\x00\x80\x76\x81\x00\x80\xB8\x6D"
-     "\x01\x00"
-     "\xB9\x0D\x01\x00\xBA\x05\x01\x00\xBB\x19\x01\x00\xBC\x09\x01\x00\xBD\x3D"
-     "\x01\x00"
-     "\xBE\x35\x01\x00\xBF\xD9\x01\x00\xB0\xD9\x02\x00\xB1\xD9\x02\x00\xB2\x6D"
-     "\x01\x00"
-     "\xB3\x65\x01\x00\xB4\x7D\x01\x00\xB5\x65\x01\x00\xB6\x65\x01\x00\xB7\x55"
-     "\x01\x00"
-     "\xE3\xF4\x00\x00\xE1\x68\x1F\x00\xE1\xB8\x01\x00\xE3\xBC\x1E\x00\x80\xA9"
-     "\x00\x00"
-     "\x81\xA9\x00\x00\x82\x05\x00\x00\x7A\x81\x00\x80\xBE\x2C\x04\x00\x82\x81"
-     "\x00\x80"
-     "\x86\x81\x00\x80\x8A\x81\x00\x80\xEF\xF8\x1E\x00\x8E\x81\x00\x80\x92\x81"
-     "\x00\x80"
-     "\xEF\x7C\x1E\x00\x86\xE0\x04\x00\x87\x74\x05\x00\x96\x81\x00\x80\xB3\xB1"
-     "\x02\x00"
-     "\x9A\x81\x00\x80\xB5\x7D\x02\x00\xB6\x75\x02\x00\x9E\x81\x00\x80\xA2\x81"
-     "\x00\x80"
-     "\xA6\x81\x00\x80\xBA\x51\x02\x00\xBB\x51\x02\x00\xBC\xF5\x01\x00\xBD\xFD"
-     "\x01\x00"
-     "\xBE\xF5\x01\x00\xBF\xED\x01\x00\x0E\x81\x00\x80\x7E\x81\x00\x80\xAA\xB1"
-     "\x05\x00"
-     "\xAB\xB1\x05\x00\xAC\x15\x06\x00\xAD\x1D\x06\x00\xAE\x15\x06\x00\xAF\x0D"
-     "\x06\x00"
-     "\xAA\x81\x00\x80\xAE\x81\x00\x80\xB2\x81\x00\x80\xA3\x51\x05\x00\xB6\x81"
-     "\x00\x80"
-     "\xA5\x9D\x05\x00\xA6\x95\x05\x00\xBA\x81\x00\x80\xB3\xE9\x06\x00\xBE\x81"
-     "\x00\x80"
-     "\xC2\x81\x00\x80\xC6\x81\x00\x80\xCA\x81\x00\x80\xB6\xE1\x06\x00\xB5\xE9"
-     "\x06\x00"
-     "\xCE\x81\x00\x80\xBB\xAD\x06\x00\xBA\xAD\x06\x00\xD2\x81\x00\x80\xD6\x81"
-     "\x00\x80"
-     "\xBF\x91\x06\x00\xBE\xA9\x06\x00\xBD\xA1\x06\x00\xBC\xAD\x06\x00\xA8\x1D"
-     "\x06\x00"
-     "\xA9\x21\x06\x00\xAA\x21\x06\x00\xAB\x21\x06\x00\xAC\x21\x06\x00\xAD\x21"
-     "\x06\x00"
-     "\xAE\x21\x06\x00\xAF\x21\x06\x00\xDA\x81\x00\x80\xDE\x81\x00\x80\xE2\x81"
-     "\x00\x80"
-     "\xE6\x81\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\xEA\x81"
-     "\x00\x80"
-     "\xB8\x35\x06\x00\xB9\x3D\x06\x00\xBA\x35\x06\x00\xBB\xE9\x07\x00\xBC\xF9"
-     "\x07\x00"
-     "\xBD\xF9\x07\x00\xBE\xE9\x07\x00\xBF\xE9\x07\x00\xB0\x29\x06\x00\xB1\x29"
-     "\x06\x00"
-     "\xB2\x39\x06\x00\xB3\x35\x06\x00\xB4\x2D\x06\x00\xB5\x15\x06\x00\xB6\x11"
-     "\x06\x00"
-     "\xB7\x0D\x06\x00\xA3\xA9\x07\x00\x84\x8C\x02\x00\xBE\x84\x01\x00\xEE\x81"
-     "\x00\x80"
-     "\xF2\x81\x00\x80\xA6\xA1\x07\x00\xA5\xA9\x07\x00\xF6\x81\x00\x80\xAB\xED"
-     "\x07\x00"
-     "\xAA\xED\x07\x00\x86\x88\x00\x00\x87\x78\x01\x00\xAF\xD1\x07\x00\xAE\xE9"
-     "\x07\x00"
-     "\xAD\xE1\x07\x00\xAC\xED\x07\x00\xFA\x81\x00\x80\xB3\xE9\x06\x00\xFE\x81"
-     "\x00\x80"
-     "\x02\x82\x00\x80\xB6\x99\x06\x00\x06\x82\x00\x80\x0A\x82\x00\x80\xB5\xE5"
-     "\x06\x00"
-     "\xBA\x81\x06\x00\xBB\x81\x06\x00\x0E\x82\x00\x80\x12\x82\x00\x80\xBE\x59"
-     "\x01\x00"
-     "\xBF\x45\x01\x00\xBC\x59\x01\x00\xBD\x59\x01\x00\xA8\x25\x06\x00\xA9\x2D"
-     "\x06\x00"
-     "\xAA\x25\x06\x00\xAB\x39\x06\x00\xAC\x29\x06\x00\xAD\x5D\x06\x00\xAE\x49"
-     "\x06\x00"
-     "\xAF\x41\x06\x00\x84\xAC\x01\x00\x16\x82\x00\x80\x1A\x82\x00\x80\x1E\x82"
-     "\x00\x80"
-     "\x22\x82\x00\x80\x26\x82\x00\x80\x2A\x82\x00\x80\x2E\x82\x00\x80\xB8\x89"
-     "\x01\x00"
-     "\xB9\x89\x01\x00\xBA\x99\x01\x00\xBB\x91\x01\x00\xBC\xB9\x01\x00\xBD\xB9"
-     "\x01\x00"
-     "\xBE\x79\x01\x00\xBF\x79\x01\x00\xB0\x01\x06\x00\xB1\x01\x06\x00\xB2\xC9"
-     "\x01\x00"
-     "\xB3\xC9\x01\x00\xB4\xD9\x01\x00\xB5\xC5\x01\x00\xB6\xC5\x01\x00\xB7\xB9"
-     "\x01\x00"
-     "\xA3\xAD\x05\x00\x32\x82\x00\x80\x36\x82\x00\x80\x3A\x82\x00\x80\x3E\x82"
-     "\x00\x80"
-     "\xA6\xDD\x05\x00\xA5\xA1\x05\x00\x42\x82\x00\x80\xAB\xC5\x05\x00\xAA\xC5"
-     "\x05\x00"
-     "\x46\x82\x00\x80\x4A\x82\x00\x80\xAF\x01\x02\x00\xAE\x1D\x02\x00\xAD\x1D"
-     "\x02\x00"
-     "\xAC\x1D\x02\x00\x4E\x82\x00\x80\x52\x82\x00\x80\x56\x82\x00\x80\x5A\x82"
-     "\x00\x80"
-     "\x81\x19\x00\x00\x80\x11\x00\x00\x62\x82\x00\x80\x82\x11\x00\x00\xBE\x24"
-     "\x03\x00"
-     "\x66\x82\x00\x80\x86\xC8\x0C\x00\x87\x4C\x03\x00\x6A\x82\x00\x80\x6E\x82"
-     "\x00\x80"
-     "\x72\x82\x00\x80\x76\x82\x00\x80\xE3\x30\x01\x00\x7A\x82\x00\x80\xE1\xB8"
-     "\x01\x00"
-     "\x84\xC0\x0C\x00\x7E\x82\x00\x80\x82\x82\x00\x80\x86\x82\x00\x80\x8A\x82"
-     "\x00\x80"
-     "\xBE\xE8\x0C\x00\x8E\x82\x00\x80\x92\x82\x00\x80\x96\x82\x00\x80\xEF\x28"
-     "\x02\x00"
-     "\x9A\x82\x00\x80\x9E\x82\x00\x80\xA2\x82\x00\x80\xB3\x81\x03\x00\xA6\x82"
-     "\x00\x80"
-     "\xAA\x82\x00\x80\xAE\x82\x00\x80\xB2\x82\x00\x80\xB6\x81\x03\x00\xB5\x95"
-     "\x03\x00"
-     "\xB6\x82\x00\x80\xBB\x45\x03\x00\xBA\x89\x03\x00\xBA\x82\x00\x80\xBE\x82"
-     "\x00\x80"
-     "\xBF\x45\x03\x00\xBE\x45\x03\x00\xBD\x55\x03\x00\xBC\x55\x03\x00\xA8\x11"
-     "\x02\x00"
-     "\xA9\x19\x02\x00\xAA\x49\x02\x00\xAB\x45\x02\x00\xAC\x5D\x02\x00\xAD\x41"
-     "\x02\x00"
-     "\xAE\x41\x02\x00\xAF\x75\x02\x00\x80\x0D\x00\x00\x81\x11\x00\x00\x82\x11"
-     "\x00\x00"
-     "\xC2\x82\x00\x80\xBE\x64\x0C\x00\xC6\x82\x00\x80\xCA\x82\x00\x80\xD2\x82"
-     "\x00\x80"
-     "\xB8\x45\x01\x00\xB9\x4D\x01\x00\xBA\x45\x01\x00\xBB\x5D\x01\x00\xBC\x4D"
-     "\x01\x00"
-     "\xBD\x7D\x01\x00\xBE\x75\x01\x00\xBF\x1D\x01\x00\xB0\x0D\x02\x00\xB1\xC1"
-     "\x02\x00"
-     "\xB2\xC1\x02\x00\xB3\xC1\x02\x00\xB4\xC1\x02\x00\xB5\xC1\x02\x00\xB6\xC1"
-     "\x02\x00"
-     "\xB7\xC1\x02\x00\xE1\x58\x06\x00\xE1\x5C\x07\x00\xE3\x9C\x00\x00\xE3\xBC"
-     "\x06\x00"
-     "\x86\x30\x0C\x00\x87\x5C\x0D\x00\xD6\x82\x00\x80\xDA\x82\x00\x80\xDE\x82"
-     "\x00\x80"
-     "\xE2\x82\x00\x80\xE6\x82\x00\x80\xEA\x82\x00\x80\xEE\x00\x00\x00\xEE\x82"
-     "\x00\x80"
-     "\xEF\x48\x00\x00\xEF\xE4\x06\x00\xF2\x82\x00\x80\xF6\x82\x00\x80\xA3\x89"
-     "\x02\x00"
-     "\xFA\x82\x00\x80\xA5\x9D\x02\x00\xFE\x82\x00\x80\x02\x83\x00\x80\xA6\x89"
-     "\x02\x00"
-     "\x06\x83\x00\x80\x0A\x83\x00\x80\xAB\x4D\x02\x00\xAA\x81\x02\x00\xAD\x5D"
-     "\x02\x00"
-     "\xAC\x5D\x02\x00\xAF\x4D\x02\x00\xAE\x4D\x02\x00\xA8\x59\x0E\x00\xA9\x61"
-     "\x0E\x00"
-     "\xAA\x79\x0E\x00\xAB\x79\x0E\x00\xAC\x69\x0E\x00\xAD\x69\x0E\x00\xAE\x9D"
-     "\x0E\x00"
-     "\xAF\x8D\x0E\x00\xCE\x82\x00\x80\x0E\x83\x00\x80\x12\x83\x00\x80\x16\x83"
-     "\x00\x80"
-     "\x1A\x83\x00\x80\x1E\x83\x00\x80\x22\x83\x00\x80\x26\x83\x00\x80\xB8\xFD"
-     "\x0E\x00"
-     "\xB9\x81\x0E\x00\xBA\x81\x0E\x00\xBB\x81\x0E\x00\xBC\x81\x0E\x00\xBD\x89"
-     "\x0E\x00"
-     "\xBE\xB1\x0E\x00\xBF\xB1\x0E\x00\xB0\xF5\x0E\x00\xB1\xFD\x0E\x00\xB2\xF5"
-     "\x0E\x00"
-     "\xB3\xC9\x0E\x00\xB4\xD9\x0E\x00\xB5\xD9\x0E\x00\xB6\xCD\x0E\x00\xB7\xC5"
-     "\x0E\x00"
-     "\xB3\xF1\x0E\x00\x2A\x83\x00\x80\x2E\x83\x00\x80\x32\x83\x00\x80\x36\x83"
-     "\x00\x80"
-     "\xB6\xE9\x0E\x00\xB5\xE1\x0E\x00\x3A\x83\x00\x80\xBB\x39\x0E\x00\xBA\x31"
-     "\x0E\x00"
-     "\x84\xE4\x03\x00\x3E\x83\x00\x80\xBF\x1D\x0E\x00\xBE\x1D\x0E\x00\xBD\x1D"
-     "\x0E\x00"
-     "\xBC\x21\x0E\x00\x82\x5D\x00\x00\xA3\xB5\x0E\x00\x80\x45\x00\x00\x81\x5D"
-     "\x00\x00"
-     "\xA6\xAD\x0E\x00\x42\x83\x00\x80\x46\x83\x00\x80\xA5\xA5\x0E\x00\xAA\x75"
-     "\x0E\x00"
-     "\xAB\x7D\x0E\x00\x86\xE8\x00\x00\x87\xE0\x00\x00\xAE\x59\x0E\x00\xAF\x59"
-     "\x0E\x00"
-     "\xAC\x65\x0E\x00\xAD\x59\x0E\x00\xA8\x3D\x0E\x00\xA9\x9D\x0E\x00\xAA\x95"
-     "\x0E\x00"
-     "\xAB\xAD\x0E\x00\xAC\xB5\x0E\x00\xAD\xBD\x0E\x00\xAE\xB5\x0E\x00\xAF\xA9"
-     "\x0E\x00"
-     "\x4A\x83\x00\x80\x4E\x83\x00\x80\x52\x83\x00\x80\x56\x83\x00\x80\x5A\x83"
-     "\x00\x80"
-     "\x5E\x83\x00\x80\x62\x83\x00\x80\x66\x83\x00\x80\xB8\x69\x01\x00\xB9\x69"
-     "\x01\x00"
-     "\xBA\x79\x01\x00\xBB\x79\x01\x00\xBC\x69\x01\x00\xBD\x69\x01\x00\xBE\xDD"
-     "\x01\x00"
-     "\xBF\xD5\x01\x00\xB0\xA1\x0E\x00\xB1\xA1\x0E\x00\xB2\xA5\x0E\x00\xB3\xB9"
-     "\x0E\x00"
-     "\xB4\xA9\x0E\x00\xB5\x95\x0E\x00\xB6\x9D\x0E\x00\xB7\x59\x01\x00\xB3\x1D"
-     "\x0E\x00"
-     "\x6A\x83\x00\x80\x6E\x83\x00\x80\x72\x83\x00\x80\x76\x83\x00\x80\xB6\x3D"
-     "\x0E\x00"
-     "\xB5\x3D\x0E\x00\x7A\x83\x00\x80\xBB\x65\x0E\x00\xBA\x65\x0E\x00\x7E\x83"
-     "\x00\x80"
-     "\x82\x83\x00\x80\xBF\x51\x0E\x00\xBE\x6D\x0E\x00\xBD\x6D\x0E\x00\xBC\x71"
-     "\x0E\x00"
-     "\xA5\x65\x0E\x00\x86\x83\x00\x80\x8A\x83\x00\x80\xA6\x65\x0E\x00\x8E\x83"
-     "\x00\x80"
-     "\x92\x83\x00\x80\xA3\x45\x0E\x00\x96\x83\x00\x80\xAD\x35\x0E\x00\xAC\x29"
-     "\x0E\x00"
-     "\xAF\x09\x0E\x00\xAE\x35\x0E\x00\x9A\x83\x00\x80\x5E\x82\x00\x80\xAB\x3D"
-     "\x0E\x00"
-     "\xAA\x3D\x0E\x00\x9E\x83\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xA2\x83\x00\x80\xA6\x83\x00\x80\xAA\x83\x00\x80\xBE\x50\x03\x00\xB2\x83"
-     "\x00\x80"
-     "\xB6\x83\x00\x80\xBA\x83\x00\x80\xBE\x83\x00\x80\xC2\x83\x00\x80\xC6\x83"
-     "\x00\x80"
-     "\xCA\x83\x00\x80\xCE\x83\x00\x80\x84\xC0\x03\x00\xD2\x83\x00\x80\xEF\x10"
-     "\x0E\x00"
-     "\x84\x00\x04\x00\x86\xF0\x04\x00\x87\x54\x03\x00\xD6\x83\x00\x80\xDA\x83"
-     "\x00\x80"
-     "\xDE\x83\x00\x80\xE2\x83\x00\x80\xE6\x83\x00\x80\xEA\x83\x00\x80\xE1\x00"
-     "\x0E\x00"
-     "\xEE\x83\x00\x80\xE3\xE4\x0E\x00\xF2\x83\x00\x80\xF6\x83\x00\x80\xFA\x83"
-     "\x00\x80"
-     "\xFE\x83\x00\x80\x02\x84\x00\x80\x06\x84\x00\x80\x0A\x84\x00\x80\x0E\x84"
-     "\x00\x80"
-     "\x12\x84\x00\x80\xA8\x75\x02\x00\xA9\x7D\x02\x00\xAA\x89\x03\x00\xAB\x89"
-     "\x03\x00"
-     "\xAC\x99\x03\x00\xAD\x99\x03\x00\xAE\x89\x03\x00\xAF\x89\x03\x00\xB0\xF9"
-     "\x03\x00"
-     "\xB1\xF9\x03\x00\xB2\x4D\x03\x00\xB3\x45\x03\x00\xB4\x5D\x03\x00\xB5\x45"
-     "\x03\x00"
-     "\xB6\x45\x03\x00\xB7\x75\x03\x00\xB8\x4D\x03\x00\xB9\x21\x03\x00\xBA\x21"
-     "\x03\x00"
-     "\xBB\x3D\x03\x00\xBC\x25\x03\x00\xBD\x29\x03\x00\xBE\x09\x03\x00\xBF\xFD"
-     "\x00\x00"
-     "\x80\x9D\x01\x00\x81\x09\x00\x00\x82\x1D\x00\x00\x16\x84\x00\x80\xEF\xF4"
-     "\x03\x00"
-     "\x1A\x84\x00\x80\xBE\x50\x05\x00\x22\x84\x00\x80\xE3\x70\x03\x00\x26\x84"
-     "\x00\x80"
-     "\xE1\xD8\x01\x00\x2A\x84\x00\x80\x2E\x84\x00\x80\x32\x84\x00\x80\x36\x84"
-     "\x00\x80"
-     "\x3A\x84\x00\x80\x86\xE0\x04\x00\x87\x88\x02\x00\x3E\x84\x00\x80\xB3\x6D"
-     "\x02\x00"
-     "\x42\x84\x00\x80\xB5\x7D\x02\x00\xB6\x75\x02\x00\x46\x84\x00\x80\x4A\x84"
-     "\x00\x80"
-     "\x4E\x84\x00\x80\xBA\x51\x02\x00\xBB\x51\x02\x00\xBC\xF5\x02\x00\xBD\xFD"
-     "\x02\x00"
-     "\xBE\xF5\x02\x00\xBF\xED\x02\x00\x52\x84\x00\x80\x56\x84\x00\x80\x5A\x84"
-     "\x00\x80"
-     "\x5E\x84\x00\x80\x62\x84\x00\x80\x66\x84\x00\x80\x6A\x84\x00\x80\xEF\x54"
-     "\x0E\x00"
-     "\x84\xEC\x04\x00\xE1\x2C\x0E\x00\x6E\x84\x00\x80\xE3\x14\x01\x00\x72\x84"
-     "\x00\x80"
-     "\x76\x84\x00\x80\x7A\x84\x00\x80\x7E\x84\x00\x80\xA3\xE1\x02\x00\x82\x2D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x80\x1D\x00\x00\x82\x84\x00\x80\xA6\xF9\x02\x00\xA5\xF1"
-     "\x02\x00"
-     "\x86\x84\x00\x80\xAB\xDD\x02\x00\xAA\xDD\x02\x00\x8A\x84\x00\x80\xBE\xB0"
-     "\x04\x00"
-     "\xAF\x61\x02\x00\xAE\x79\x02\x00\xAD\x71\x02\x00\xAC\x79\x02\x00\xA8\x21"
-     "\x06\x00"
-     "\xA9\x21\x06\x00\xAA\x21\x06\x00\xAB\x21\x06\x00\xAC\x21\x06\x00\xAD\x21"
-     "\x06\x00"
-     "\xAE\x21\x06\x00\xAF\x21\x06\x00\x1E\x84\x00\x80\x8E\x84\x00\x80\x86\x00"
-     "\x00\x00"
-     "\x87\xA0\x03\x00\x92\x84\x00\x80\x96\x84\x00\x80\x9A\x84\x00\x80\x9E\x84"
-     "\x00\x80"
-     "\xB8\x3D\x06\x00\xB9\xC5\x06\x00\xBA\xCD\x06\x00\xBB\xC5\x06\x00\xBC\xDD"
-     "\x06\x00"
-     "\xBD\xC5\x06\x00\xBE\xCD\x06\x00\xBF\xC5\x06\x00\xB0\x25\x06\x00\xB1\x3D"
-     "\x06\x00"
-     "\xB2\x35\x06\x00\xB3\x0D\x06\x00\xB4\x15\x06\x00\xB5\x1D\x06\x00\xB6\x15"
-     "\x06\x00"
-     "\xB7\x09\x06\x00\xB3\xA9\x06\x00\xA2\x84\x00\x80\xA6\x84\x00\x80\xAA\x84"
-     "\x00\x80"
-     "\xAE\x84\x00\x80\xB6\xA1\x06\x00\xB5\xA9\x06\x00\xB2\x84\x00\x80\xBB\xF9"
-     "\x06\x00"
-     "\xBA\xE1\x06\x00\xB6\x84\x00\x80\xBA\x84\x00\x80\xBF\xD9\x06\x00\xBE\xD1"
-     "\x06\x00"
-     "\xBD\xD9\x06\x00\xBC\xE1\x06\x00\xBE\x84\x00\x80\xA3\xED\x06\x00\xC2\x84"
-     "\x00\x80"
-     "\xC6\x84\x00\x80\xA6\xE5\x06\x00\xCA\x84\x00\x80\xCE\x84\x00\x80\xA5\xED"
-     "\x06\x00"
-     "\xAA\xA5\x06\x00\xAB\xBD\x06\x00\xD2\x84\x00\x80\xD6\x84\x00\x80\xAE\x95"
-     "\x06\x00"
-     "\xAF\x9D\x06\x00\xAC\xA5\x06\x00\xAD\x9D\x06\x00\xA8\x1D\x06\x00\xA9\x21"
-     "\x06\x00"
-     "\xAA\x25\x06\x00\xAB\x3D\x06\x00\xAC\x25\x06\x00\xAD\x2D\x06\x00\xAE\x25"
-     "\x06\x00"
-     "\xAF\x55\x06\x00\xDA\x84\x00\x80\xDE\x84\x00\x80\xE2\x84\x00\x80\xE6\x84"
-     "\x00\x80"
-     "\xEA\x84\x00\x80\xEE\x84\x00\x80\xF2\x84\x00\x80\xF6\x84\x00\x80\xB8\xD1"
-     "\x06\x00"
-     "\xB9\xD1\x06\x00\xBA\xD1\x06\x00\xBB\xE5\x06\x00\xBC\xE1\x06\x00\xBD\xE1"
-     "\x06\x00"
-     "\xBE\xE1\x06\x00\xBF\xE1\x06\x00\xB0\x31\x06\x00\xB1\x39\x06\x00\xB2\x09"
-     "\x06\x00"
-     "\xB3\x09\x06\x00\xB4\xF1\x06\x00\xB5\xF1\x06\x00\xB6\xF1\x06\x00\xB7\xF1"
-     "\x06\x00"
-     "\xB3\xAD\x06\x00\x82\x15\x00\x00\x81\xB5\x00\x00\x80\xB5\x00\x00\xFA\x84"
-     "\x00\x80"
-     "\xB6\xAD\x06\x00\xB5\xAD\x06\x00\xBE\x10\x00\x00\xBB\xFD\x06\x00\xBA\xF5"
-     "\x06\x00"
-     "\x84\x6C\x01\x00\x02\x85\x00\x80\xBF\x35\x06\x00\xBE\x35\x06\x00\xBD\xDD"
-     "\x06\x00"
-     "\xBC\xDD\x06\x00\x06\x85\x00\x80\xA6\xFD\x06\x00\xA5\xFD\x06\x00\xFE\x84"
-     "\x00\x80"
-     "\xA3\xFD\x06\x00\x0A\x85\x00\x80\x86\x00\xFC\x01\x87\x60\x01\x00\xAF\x65"
-     "\x06\x00"
-     "\xAE\x65\x06\x00\xAD\x8D\x06\x00\xAC\x8D\x06\x00\xAB\xAD\x06\x00\xAA\xA5"
-     "\x06\x00"
-     "\xAE\x83\x00\x80\x0E\x85\x00\x80\x12\x85\x00\x80\x16\x85\x00\x80\xB3\x31"
-     "\x02\x00"
-     "\x1A\x85\x00\x80\xB5\x95\x02\x00\x1E\x85\x00\x80\x22\x85\x00\x80\xB6\x95"
-     "\x02\x00"
-     "\x26\x85\x00\x80\x2A\x85\x00\x80\xBB\x81\x02\x00\xBA\xB1\x02\x00\xBD\xF1"
-     "\x02\x00"
-     "\xBC\x99\x02\x00\xBF\xF5\x02\x00\xBE\xF5\x02\x00\xE3\x1C\x02\x00\xEF\xA4"
-     "\x07\x00"
-     "\xE1\x88\x01\x00\x84\xC4\x03\x00\x2E\x85\x00\x80\x32\x85\x00\x80\x36\x85"
-     "\x00\x80"
-     "\x3A\x85\x00\x80\x3E\x85\x00\x80\x42\x85\x00\x80\x46\x85\x00\x80\x4A\x85"
-     "\x00\x80"
-     "\xEF\x28\x02\x00\xE3\x28\x06\x00\x4E\x85\x00\x80\xE1\x34\x06\x00\xAA\xB9"
-     "\x02\x00"
-     "\xAB\xB9\x02\x00\xA8\xB9\x02\x00\xA9\xB9\x02\x00\xAE\x1D\x02\x00\xAF\x05"
-     "\x02\x00"
-     "\xAC\x1D\x02\x00\xAD\x15\x02\x00\xBE\xEC\x03\x00\x52\x85\x00\x80\x56\x85"
-     "\x00\x80"
-     "\x5A\x85\x00\x80\x5E\x85\x00\x80\x82\xB9\x01\x00\x81\xB1\x01\x00\x80\xB9"
-     "\x01\x00"
-     "\xBA\xE1\x03\x00\xBB\xE1\x03\x00\xB8\x21\x02\x00\xB9\x21\x02\x00\xBE\xE1"
-     "\x03\x00"
-     "\xBF\xE1\x03\x00\xBC\xF1\x03\x00\xBD\xF1\x03\x00\xB2\x4D\x02\x00\xB3\x25"
-     "\x02\x00"
-     "\xB0\x75\x02\x00\xB1\x45\x02\x00\xB6\x25\x02\x00\xB7\x2D\x02\x00\xB4\x31"
-     "\x02\x00"
-     "\xB5\x31\x02\x00\xA3\x39\x03\x00\x62\x85\x00\x80\x66\x85\x00\x80\x6A\x85"
-     "\x00\x80"
-     "\x72\x85\x00\x80\xA6\x9D\x03\x00\xA5\x9D\x03\x00\x76\x85\x00\x80\xAB\x89"
-     "\x03\x00"
-     "\xAA\xB9\x03\x00\x7A\x85\x00\x80\x7E\x85\x00\x80\xAF\xFD\x03\x00\xAE\xFD"
-     "\x03\x00"
-     "\xAD\xF9\x03\x00\xAC\x91\x03\x00\x82\x85\x00\x80\xB3\x91\x02\x00\x86\xC8"
-     "\xFD\x01"
-     "\x87\x40\xFD\x01\xB6\x55\x01\x00\x86\x85\x00\x80\x8A\x85\x00\x80\xB5\x5D"
-     "\x01\x00"
-     "\xBA\x79\x01\x00\xBB\x45\x01\x00\x8E\x85\x00\x80\x92\x85\x00\x80\xBE\x25"
-     "\x01\x00"
-     "\xBF\x2D\x01\x00\xBC\x5D\x01\x00\xBD\x2D\x01\x00\x96\x85\x00\x80\x9A\x85"
-     "\x00\x80"
-     "\x9E\x85\x00\x80\x84\xC0\xFD\x01\xA2\x85\x00\x80\xA6\x85\x00\x80\xEF\xFC"
-     "\x00\x00"
-     "\xEF\x88\x06\x00\xE1\x90\x01\x00\xE1\xFC\x06\x00\xE3\x10\x00\x00\xE3\x1C"
-     "\x06\x00"
-     "\xAA\x85\x00\x80\xAE\x85\x00\x80\xBE\xF0\xFD\x01\xB2\x85\x00\x80\xA3\x15"
-     "\x02\x00"
-     "\xB6\x85\x00\x80\xBA\x85\x00\x80\xBE\x85\x00\x80\xC2\x85\x00\x80\xA6\xD1"
-     "\x01\x00"
-     "\xA5\xD9\x01\x00\xC6\x85\x00\x80\xAB\xC1\x01\x00\xAA\xFD\x01\x00\xCA\x85"
-     "\x00\x80"
-     "\xCE\x85\x00\x80\xAF\xA9\x01\x00\xAE\xA1\x01\x00\xAD\xA9\x01\x00\xAC\xD9"
-     "\x01\x00"
-     "\x6E\x85\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\xD2\x85"
-     "\x00\x80"
-     "\xD6\x85\x00\x80\xDA\x85\x00\x80\xDE\x85\x00\x80\xA8\x31\xFF\x01\xA9\x39"
-     "\xFF\x01"
-     "\xAA\x01\xFF\x01\xAB\x01\xFF\x01\xAC\x71\xFF\x01\xAD\x61\xFF\x01\xAE\x61"
-     "\xFF\x01"
-     "\xAF\x55\xFF\x01\xB0\x2D\xFF\x01\xB1\xC5\xFF\x01\xB2\xCD\xFF\x01\xB3\xC5"
-     "\xFF\x01"
-     "\xB4\xC1\xFF\x01\xB5\xC1\xFF\x01\xB6\xE9\xFF\x01\xB7\xE9\xFF\x01\xB8\xDD"
-     "\xFF\x01"
-     "\xB9\xA1\xFF\x01\xBA\xA1\xFF\x01\xBB\xA1\xFF\x01\xBC\xA1\xFF\x01\xBD\xA9"
-     "\xFF\x01"
-     "\xBE\x91\xFF\x01\xBF\x91\xFF\x01\xB3\xA9\xFF\x01\xE2\x85\x00\x80\x86\x28"
-     "\x03\x00"
-     "\x87\x1C\x03\x00\xE6\x85\x00\x80\xB6\xC1\xFF\x01\xB5\xD9\xFF\x01\xEA\x85"
-     "\x00\x80"
-     "\xBB\x21\xFF\x01\xBA\xC9\xFF\x01\xEE\x85\x00\x80\xF2\x85\x00\x80\xBF\x25"
-     "\xFF\x01"
-     "\xBE\x25\xFF\x01\xBD\x31\xFF\x01\xBC\x39\xFF\x01\xF6\x85\x00\x80\xA3\xED"
-     "\xFF\x01"
-     "\xFA\x85\x00\x80\xFE\x85\x00\x80\xA6\x85\xFF\x01\x02\x86\x00\x80\x06\x86"
-     "\x00\x80"
-     "\xA5\x9D\xFF\x01\xAA\x8D\xFF\x01\xAB\x65\xFF\x01\x0A\x86\x00\x80\x0E\x86"
-     "\x00\x80"
-     "\xAE\x61\xFF\x01\xAF\x61\xFF\x01\xAC\x7D\xFF\x01\xAD\x75\xFF\x01\xA8\x85"
-     "\x01\x00"
-     "\xA9\x95\x01\x00\xAA\x95\x01\x00\xAB\xA5\x01\x00\xAC\xBD\x01\x00\xAD\xD5"
-     "\x01\x00"
-     "\xAE\xD1\x01\x00\xAF\xD1\x01\x00\x12\x86\x00\x80\x16\x86\x00\x80\x1A\x86"
-     "\x00\x80"
-     "\x1E\x86\x00\x80\x22\x86\x00\x80\x26\x86\x00\x80\x2A\x86\x00\x80\x2E\x86"
-     "\x00\x80"
-     "\xB8\x71\x00\x00\xB9\x71\x00\x00\xBA\x71\x00\x00\xBB\x71\x00\x00\xBC\xD1"
-     "\x00\x00"
-     "\xBD\xD1\x00\x00\xBE\xD1\x00\x00\xBF\xD1\x00\x00\xB0\xB5\x01\x00\xB1\xBD"
-     "\x01\x00"
-     "\xB2\x81\x01\x00\xB3\x81\x01\x00\xB4\x51\x00\x00\xB5\x51\x00\x00\xB6\x51"
-     "\x00\x00"
-     "\xB7\x51\x00\x00\x32\x86\x00\x80\x36\x86\x00\x80\x3A\x86\x00\x80\x3E\x86"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\x42\x86\x00\x80\x46\x86"
-     "\x00\x80"
-     "\x4A\x86\x00\x80\x84\xA4\x02\x00\xBE\x20\x01\x00\x86\x90\x01\x00\x87\x14"
-     "\x01\x00"
-     "\x52\x86\x00\x80\x56\x86\x00\x80\xB3\x5D\x01\x00\x4E\x86\x00\x80\x5A\x86"
-     "\x00\x80"
-     "\x5E\x86\x00\x80\x62\x86\x00\x80\xB6\x15\x01\x00\xB5\x75\x01\x00\x66\x86"
-     "\x00\x80"
-     "\xBB\x39\x01\x00\xBA\x31\x01\x00\x6A\x86\x00\x80\x6E\x86\x00\x80\xBF\xD5"
-     "\x00\x00"
-     "\xBE\xD5\x00\x00\xBD\x15\x01\x00\xBC\x15\x01\x00\xA3\x9D\x02\x00\x72\x86"
-     "\x00\x80"
-     "\x76\x86\x00\x80\xBE\x00\x04\x00\x7A\x86\x00\x80\xA6\xD5\x02\x00\xA5\xB5"
-     "\x02\x00"
-     "\x7E\x86\x00\x80\xAB\xF9\x02\x00\xAA\xF1\x02\x00\x82\x86\x00\x80\x86\x86"
-     "\x00\x80"
-     "\xAF\x15\x03\x00\xAE\x15\x03\x00\xAD\xD5\x02\x00\xAC\xD5\x02\x00\xE1\xE8"
-     "\x01\x00"
-     "\x8A\x86\x00\x80\xE3\x14\x02\x00\x84\x40\x03\x00\x8E\x86\x00\x80\x92\x86"
-     "\x00\x80"
-     "\x96\x86\x00\x80\x9A\x86\x00\x80\x9E\x86\x00\x80\xA2\x86\x00\x80\xA6\x86"
-     "\x00\x80"
-     "\xAA\x86\x00\x80\xAE\x86\x00\x80\xB2\x86\x00\x80\xEF\x8C\x02\x00\xB6\x86"
-     "\x00\x80"
-     "\xA8\x39\x03\x00\xA9\x39\x03\x00\xAA\x41\x03\x00\xAB\x41\x03\x00\xAC\x41"
-     "\x03\x00"
-     "\xAD\x49\x03\x00\xAE\x71\x03\x00\xAF\x71\x03\x00\x80\xD5\x00\x00\x81\xD9"
-     "\x00\x00"
-     "\x82\x05\x00\x00\xBA\x86\x00\x80\x85\xB4\x05\x00\xC2\x86\x00\x80\xC6\x86"
-     "\x00\x80"
-     "\xCA\x86\x00\x80\xB8\x95\x00\x00\xB9\x9D\x00\x00\xBA\x95\x00\x00\xBB\xAD"
-     "\x00\x00"
-     "\xBC\xB5\x00\x00\xBD\xBD\x00\x00\xBE\xB5\x00\x00\xBF\xAD\x00\x00\xB0\xE5"
-     "\x00\x00"
-     "\xB1\xED\x00\x00\xB2\xF9\x00\x00\xB3\xF1\x00\x00\xB4\xD9\x00\x00\xB5\xD9"
-     "\x00\x00"
-     "\xB6\xBD\x00\x00\xB7\xAD\x00\x00\xB6\x5D\xFD\x01\xB7\x65\xFD\x01\xB4\x5D"
-     "\xFD\x01"
-     "\xB5\x55\xFD\x01\xB2\x35\xFD\x01\xB3\x45\xFD\x01\xB0\x01\xFD\x01\xB1\x3D"
-     "\xFD\x01"
-     "\xBE\x8D\xFD\x01\xBF\x95\xFD\x01\xBC\x8D\xFD\x01\xBD\x85\xFD\x01\xBA\x5D"
-     "\xFD\x01"
-     "\xBB\x4D\xFD\x01\xB8\x5D\xFD\x01\xB9\x55\xFD\x01\x86\xE0\x07\x00\x87\x18"
-     "\x05\x00"
-     "\xCE\x86\x00\x80\xD2\x86\x00\x80\xD6\x86\x00\x80\x84\x14\x07\x00\xDA\x86"
-     "\x00\x80"
-     "\xDE\x86\x00\x80\xAE\x6D\xFD\x01\xAF\x71\xFD\x01\xAC\x69\xFD\x01\xAD\x69"
-     "\xFD\x01"
-     "\xAA\xA5\xFE\x01\xAB\x89\xFE\x01\xA8\xD5\xFE\x01\xA9\xBD\xFE\x01\xE1\x24"
-     "\x03\x00"
-     "\xE1\xD4\xFE\x01\xE3\xDC\x1F\x00\xE3\x10\xFE\x01\xE3\x28\x00\x00\xE2\x86"
-     "\x00\x80"
-     "\xE1\x08\x01\x00\xE6\x86\x00\x80\xEF\xF4\x00\x00\xEA\x86\x00\x80\xBE\xF4"
-     "\x07\x00"
-     "\xEE\x86\x00\x80\xF2\x86\x00\x80\xF6\x86\x00\x80\xEF\x7C\x1F\x00\xEF\xFC"
-     "\xFE\x01"
-     "\xB3\x09\x02\x00\xFA\x86\x00\x80\xFE\x86\x00\x80\x02\x87\x00\x80\x06\x87"
-     "\x00\x80"
-     "\xB6\x09\x02\x00\xB5\x09\x02\x00\x0A\x87\x00\x80\xBB\xE9\x01\x00\xBA\xE1"
-     "\x01\x00"
-     "\x0E\x87\x00\x80\x12\x87\x00\x80\xBF\xDD\x01\x00\xBE\xDD\x01\x00\xBD\xE5"
-     "\x01\x00"
-     "\xBC\xED\x01\x00\x9F\x25\xF8\x01\x9E\x25\xF8\x01\x9D\xF1\xF9\x01\x9C\xA1"
-     "\xFC\x01"
-     "\x9B\x31\xFC\x01\x9A\xE1\xFD\x01\x99\x15\xF1\x01\x98\x19\xF0\x01\x97\x21"
-     "\xF0\x01"
-     "\x96\x4D\xF5\x01\x95\x25\xF4\x01\x94\x2D\xF4\x01\x93\x61\xE9\x01\x92\x3D"
-     "\xE8\x01"
-     "\x91\x3D\xE8\x01\x90\xBD\xEE\x01\x8F\x61\xEC\x01\xBE\x86\x00\x80\x16\x87"
-     "\x00\x80"
-     "\x1A\x87\x00\x80\x1E\x87\x00\x80\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D"
-     "\x00\x00"
-     "\x22\x87\x00\x80\x26\x87\x00\x80\x2A\x87\x00\x80\x2E\x87\x00\x80\x86\xD0"
-     "\x01\x00"
-     "\x87\x00\x03\x00\x32\x87\x00\x80\x36\x87\x00\x80\xA3\x49\x05\x00\x3A\x87"
-     "\x00\x80"
-     "\x3E\x87\x00\x80\x42\x87\x00\x80\x46\x87\x00\x80\xA6\x49\x05\x00\xA5\x49"
-     "\x05\x00"
-     "\x4A\x87\x00\x80\xAB\xA9\x06\x00\xAA\xA1\x06\x00\x4E\x87\x00\x80\x52\x87"
-     "\x00\x80"
-     "\xAF\x9D\x06\x00\xAE\x9D\x06\x00\xAD\xA5\x06\x00\xAC\xAD\x06\x00\x56\x87"
-     "\x00\x80"
-     "\x5A\x87\x00\x80\x5E\x87\x00\x80\x62\x87\x00\x80\x66\x87\x00\x80\x6A\x87"
-     "\x00\x80"
-     "\x6E\x87\x00\x80\x72\x87\x00\x80\x76\x87\x00\x80\x7A\x87\x00\x80\x7E\x87"
-     "\x00\x80"
-     "\x82\x87\x00\x80\xA3\x35\x05\x00\xA2\xE9\x04\x00\xA1\xF9\x04\x00\xA0\x31"
-     "\xF9\x01"
-     "\x82\x21\xED\x01\x83\x85\xEF\x01\x86\x87\x00\x80\x8A\x87\x00\x80\x86\x41"
-     "\xE8\x01"
-     "\x87\x15\xF5\x01\x84\x1D\xE9\x01\x85\x15\xE9\x01\x8A\x2D\xF5\x01\x8B\x4D"
-     "\xF4\x01"
-     "\x8E\x87\x00\x80\x92\x87\x00\x80\x8E\x3D\xF0\x01\x8F\xC1\xFC\x01\x8C\x39"
-     "\xF1\x01"
-     "\x8D\x39\xF1\x01\x92\x61\xFD\x01\x93\xE9\xFD\x01\x96\x87\x00\x80\x9A\x87"
-     "\x00\x80"
-     "\x96\x75\xF9\x01\x97\x99\xF8\x01\x94\xA1\xF8\x01\x95\x7D\xF9\x01\x9A\x61"
-     "\x05\x00"
-     "\x9B\x79\x05\x00\x84\x2C\x00\x00\xBE\x00\x0C\x00\x81\x39\x00\x00\x80\x59"
-     "\x00\x00"
-     "\x9C\xBD\x04\x00\x82\x31\x00\x00\xA1\xC5\x01\x00\xA2\x87\x00\x80\xA3\x89"
-     "\x0D\x00"
-     "\xA2\xE1\x01\x00\xA5\x01\x0D\x00\xA4\x91\x0D\x00\xA7\xB9\x09\x00\xA6\x01"
-     "\x08\x00"
-     "\xA9\x01\x14\x00\xA8\x0D\x09\x00\xAB\x99\x15\x00\xAA\x8D\x15\x00\xAD\xB1"
-     "\x11\x00"
-     "\xAC\x21\x14\x00\xAF\x01\x1C\x00\xAE\x15\x11\x00\xB1\xCD\x1D\x00\xB0\xC5"
-     "\x1D\x00"
-     "\xB3\x01\x18\x00\xB2\x55\x1C\x00\xB5\x39\x19\x00\xB4\xF5\x19\x00\x86\x58"
-     "\x0F\x00"
-     "\x87\xFC\x00\x00\xA6\x87\x00\x80\xAA\x87\x00\x80\xAE\x87\x00\x80\xB2\x87"
-     "\x00\x80"
-     "\xB6\x87\x00\x80\xBA\x87\x00\x80\xBE\x87\x00\x80\xC2\x87\x00\x80\xA8\xB5"
-     "\x04\x00"
-     "\xA9\xBD\x04\x00\xAA\xB5\x04\x00\xAB\x45\x05\x00\xAC\x5D\x05\x00\xAD\x45"
-     "\x05\x00"
-     "\xAE\x41\x05\x00\xAF\x41\x05\x00\xC6\x87\x00\x80\xCA\x87\x00\x80\xCE\x87"
-     "\x00\x80"
-     "\xD2\x87\x00\x80\xD6\x87\x00\x80\xDA\x87\x00\x80\xDE\x87\x00\x80\xE2\x87"
-     "\x00\x80"
-     "\xB8\x8D\x05\x00\xB9\x95\x05\x00\xBA\x95\x05\x00\xBB\xA5\x05\x00\xBC\xBD"
-     "\x05\x00"
-     "\xBD\x1D\x02\x00\xBE\x15\x02\x00\xBF\x0D\x02\x00\xB0\xF5\x05\x00\xB1\xF9"
-     "\x05\x00"
-     "\xB2\xCD\x05\x00\xB3\xC1\x05\x00\xB4\xC1\x05\x00\xB5\xC1\x05\x00\xB6\xBD"
-     "\x05\x00"
-     "\xB7\xB5\x05\x00\xE6\x87\x00\x80\xEA\x87\x00\x80\xEE\x87\x00\x80\xB3\x31"
-     "\x04\x00"
-     "\xF2\x87\x00\x80\xB5\xD9\x05\x00\xB6\xC9\x05\x00\xF6\x87\x00\x80\xFA\x87"
-     "\x00\x80"
-     "\xFE\x87\x00\x80\xBA\x3D\x05\x00\xBB\x31\x05\x00\xBC\x15\x05\x00\xBD\x19"
-     "\x05\x00"
-     "\xBE\x09\x05\x00\xBF\x09\x05\x00\x02\x88\x00\x80\xA6\x95\x05\x00\xA5\x85"
-     "\x05\x00"
-     "\x06\x88\x00\x80\xA3\x6D\x04\x00\x0A\x88\x00\x80\xBE\x90\x0C\x00\x12\x88"
-     "\x00\x80"
-     "\xAF\x55\x05\x00\xAE\x55\x05\x00\xAD\x45\x05\x00\xAC\x49\x05\x00\xAB\x6D"
-     "\x05\x00"
-     "\xAA\x61\x05\x00\x16\x88\x00\x80\x1A\x88\x00\x80\xA8\x79\x03\x00\xA9\x79"
-     "\x03\x00"
-     "\xAA\x8D\x03\x00\xAB\x85\x03\x00\xAC\x9D\x03\x00\xAD\x85\x03\x00\xAE\x85"
-     "\x03\x00"
-     "\xAF\xB5\x03\x00\x80\xE1\x01\x00\x81\xE9\x01\x00\x82\xFD\x01\x00\x84\x60"
-     "\x0D\x00"
-     "\x1E\x88\x00\x80\x22\x88\x00\x80\x87\x24\x0D\x00\x86\x9C\x0C\x00\xB8\x9D"
-     "\x03\x00"
-     "\xB9\xA5\x03\x00\xBA\xAD\x03\x00\xBB\xA5\x03\x00\xBC\xBD\x03\x00\xBD\x5D"
-     "\x02\x00"
-     "\xBE\x55\x02\x00\xBF\x75\x02\x00\xB0\xCD\x03\x00\xB1\xA5\x03\x00\xB2\xAD"
-     "\x03\x00"
-     "\xB3\xA5\x03\x00\xB4\xBD\x03\x00\xB5\xA5\x03\x00\xB6\xAD\x03\x00\xB7\xA5"
-     "\x03\x00"
-     "\x26\x88\x00\x80\x2A\x88\x00\x80\xB3\x0D\x03\x00\x2E\x88\x00\x80\xB5\x1D"
-     "\x03\x00"
-     "\xBE\xDC\x0C\x00\x32\x88\x00\x80\xB6\x0D\x03\x00\x36\x88\x00\x80\x3A\x88"
-     "\x00\x80"
-     "\xBB\x69\x03\x00\xBA\x01\x03\x00\xBD\x79\x03\x00\xBC\x71\x03\x00\xBF\x69"
-     "\x03\x00"
-     "\xBE\x71\x03\x00\xE1\x98\x06\x00\x3E\x88\x00\x80\xE3\x8C\x06\x00\x42\x88"
-     "\x00\x80"
-     "\xEF\x44\x00\x00\x46\x88\x00\x80\x4A\x88\x00\x80\x4E\x88\x00\x80\xE3\x98"
-     "\x00\x00"
-     "\x52\x88\x00\x80\xE1\x80\x01\x00\x56\x88\x00\x80\x5A\x88\x00\x80\x5E\x88"
-     "\x00\x80"
-     "\xEF\xA8\x06\x00\x62\x88\x00\x80\x9E\x87\x00\x80\x0E\x88\x00\x80\x66\x88"
-     "\x00\x80"
-     "\x6A\x88\x00\x80\x6E\x88\x00\x80\x82\x99\x00\x00\x81\x91\x00\x00\x80\x99"
-     "\x00\x00"
-     "\xA8\xA5\x0D\x00\xA9\xC5\x0D\x00\xAA\xCD\x0D\x00\xAB\xC5\x0D\x00\xAC\xDD"
-     "\x0D\x00"
-     "\xAD\xC5\x0D\x00\xAE\xCD\x0D\x00\xAF\xC5\x0D\x00\xB0\xBD\x0D\x00\xB1\x65"
-     "\x0E\x00"
-     "\xB2\x6D\x0E\x00\xB3\x65\x0E\x00\xB4\x7D\x0E\x00\xB5\x65\x0E\x00\xB6\x6D"
-     "\x0E\x00"
-     "\xB7\x65\x0E\x00\xB8\x5D\x0E\x00\xB9\xE5\x0E\x00\xBA\xED\x0E\x00\xBB\xE5"
-     "\x0E\x00"
-     "\xBC\xE1\x0E\x00\xBD\xE9\x0E\x00\xBE\xD9\x0E\x00\xBF\xD9\x0E\x00\xB3\x65"
-     "\x0D\x00"
-     "\x72\x88\x00\x80\x76\x88\x00\x80\x7A\x88\x00\x80\x7E\x88\x00\x80\xB6\x65"
-     "\x0D\x00"
-     "\xB5\x75\x0D\x00\x84\xD0\x03\x00\xBB\x81\x0E\x00\xBA\x59\x0D\x00\x86\x08"
-     "\x00\x00"
-     "\x87\x8C\x00\x00\xBF\x81\x0E\x00\xBE\x99\x0E\x00\xBD\x91\x0E\x00\xBC\x99"
-     "\x0E\x00"
-     "\x82\x88\x00\x80\xA3\x21\x0D\x00\x86\x88\x00\x80\x8A\x88\x00\x80\xA6\x21"
-     "\x0D\x00"
-     "\x8E\x88\x00\x80\x92\x88\x00\x80\xA5\x31\x0D\x00\xAA\x1D\x0D\x00\xAB\xC5"
-     "\x0E\x00"
-     "\x96\x88\x00\x80\x9A\x88\x00\x80\xAE\xDD\x0E\x00\xAF\xC5\x0E\x00\xAC\xDD"
-     "\x0E\x00"
-     "\xAD\xD5\x0E\x00\xA8\xDD\x0F\x00\xA9\xFD\x0F\x00\xAA\xF5\x0F\x00\xAB\x3D"
-     "\x0E\x00"
-     "\xAC\x25\x0E\x00\xAD\x95\x0E\x00\xAE\x9D\x0E\x00\xAF\x91\x0E\x00\x9E\x88"
-     "\x00\x80"
-     "\xA2\x88\x00\x80\xA6\x88\x00\x80\xAA\x88\x00\x80\xAE\x88\x00\x80\xB2\x88"
-     "\x00\x80"
-     "\xB6\x88\x00\x80\xBA\x88\x00\x80\xB8\xA5\x0E\x00\xB9\xA9\x0E\x00\xBA\xB9"
-     "\x0E\x00"
-     "\xBB\xB9\x0E\x00\xBC\xAD\x0E\x00\xBD\x5D\x01\x00\xBE\x55\x01\x00\xBF\x4D"
-     "\x01\x00"
-     "\xB0\xF9\x0E\x00\xB1\xC5\x0E\x00\xB2\xCD\x0E\x00\xB3\xA9\x0E\x00\xB4\xBD"
-     "\x0E\x00"
-     "\xB5\xA1\x0E\x00\xB6\xA1\x0E\x00\xB7\xA1\x0E\x00\xB3\x7D\x0F\x00\xBE\x88"
-     "\x00\x80"
-     "\xC2\x88\x00\x80\xC6\x88\x00\x80\xCA\x88\x00\x80\xB6\x15\x0E\x00\xB5\xAD"
-     "\x0E\x00"
-     "\xCE\x88\x00\x80\xBB\x01\x0E\x00\xBA\x3D\x0E\x00\xD2\x88\x00\x80\xD6\x88"
-     "\x00\x80"
-     "\xBF\x65\x0E\x00\xBE\x79\x0E\x00\xBD\x7D\x0E\x00\xBC\x19\x0E\x00\xDA\x88"
-     "\x00\x80"
-     "\xA3\x39\x0F\x00\xDE\x88\x00\x80\xE2\x88\x00\x80\xA6\x51\x0E\x00\xE6\x88"
-     "\x00\x80"
-     "\xEA\x88\x00\x80\xA5\xE9\x0E\x00\xAA\x79\x0E\x00\xAB\x45\x0E\x00\x84\xA4"
-     "\x01\x00"
-     "\xBE\xA4\x01\x00\xAE\x3D\x0E\x00\xAF\x21\x0E\x00\xAC\x5D\x0E\x00\xAD\x39"
-     "\x0E\x00"
-     "\xA8\xC1\x02\x00\xA9\xCD\x02\x00\xAA\xC5\x02\x00\xAB\xD9\x02\x00\xAC\xC9"
-     "\x02\x00"
-     "\xAD\xFD\x02\x00\xAE\xF5\x02\x00\xAF\x3D\x03\x00\x80\xED\x03\x00\x81\xF1"
-     "\x03\x00"
-     "\x82\xF1\x03\x00\xF2\x88\x00\x80\x86\x90\x04\x00\x87\x74\x03\x00\xF6\x88"
-     "\x00\x80"
-     "\xFA\x88\x00\x80\xB8\xC9\x03\x00\xB9\xC9\x03\x00\xBA\xD9\x03\x00\xBB\xD1"
-     "\x03\x00"
-     "\xBC\xF9\x03\x00\xBD\xF9\x03\x00\xBE\x99\x03\x00\xBF\x95\x03\x00\xB0\x45"
-     "\x03\x00"
-     "\xB1\x4D\x03\x00\xB2\x45\x03\x00\xB3\x5D\x03\x00\xB4\x45\x03\x00\xB5\x4D"
-     "\x03\x00"
-     "\xB6\x45\x03\x00\xB7\xF9\x03\x00\xB3\x4D\x02\x00\xFE\x88\x00\x80\x02\x89"
-     "\x00\x80"
-     "\x84\x0C\x03\x00\x06\x89\x00\x80\xB6\x7D\x02\x00\xB5\x41\x02\x00\x0A\x89"
-     "\x00\x80"
-     "\xBB\x89\x03\x00\xBA\x81\x03\x00\x0E\x89\x00\x80\x12\x89\x00\x80\xBF\x89"
-     "\x03\x00"
-     "\xBE\x81\x03\x00\xBD\x89\x03\x00\xBC\x91\x03\x00\x16\x89\x00\x80\xA3\x09"
-     "\x02\x00"
-     "\x1A\x89\x00\x80\x1E\x89\x00\x80\xA6\x39\x02\x00\x22\x89\x00\x80\x26\x89"
-     "\x00\x80"
-     "\xA5\x05\x02\x00\xAA\xC5\x03\x00\xAB\xCD\x03\x00\x2A\x89\x00\x80\x2E\x89"
-     "\x00\x80"
-     "\xAE\xC5\x03\x00\xAF\xCD\x03\x00\xAC\xD5\x03\x00\xAD\xCD\x03\x00\x32\x89"
-     "\x00\x80"
-     "\x36\x89\x00\x80\x3A\x89\x00\x80\x3E\x89\x00\x80\x42\x89\x00\x80\xE1\xA4"
-     "\x01\x00"
-     "\x46\x89\x00\x80\xE3\xE4\x00\x00\x4A\x89\x00\x80\x4E\x89\x00\x80\x52\x89"
-     "\x00\x80"
-     "\xEF\x38\x00\x00\x56\x89\x00\x80\x5A\x89\x00\x80\x62\x89\x00\x80\xBE\x30"
-     "\x05\x00"
-     "\x80\x69\x00\x00\x81\x75\x00\x00\x82\x7D\x00\x00\xB3\x71\x01\x00\x66\x89"
-     "\x00\x80"
-     "\xB5\xDD\x01\x00\xB6\xC9\x01\x00\x6A\x89\x00\x80\x86\xC0\x04\x00\x87\xE8"
-     "\x05\x00"
-     "\xBA\xC5\x01\x00\xBB\xAD\x01\x00\xBC\xB9\x01\x00\xBD\xB9\x01\x00\xBE\xAD"
-     "\x01\x00"
-     "\xBF\xA5\x01\x00\xE1\xA4\x0E\x00\xE1\x44\x0C\x00\xE3\x04\x01\x00\xE3\xB0"
-     "\x0E\x00"
-     "\x6E\x89\x00\x80\x72\x89\x00\x80\x76\x89\x00\x80\x7A\x89\x00\x80\x84\xA8"
-     "\x05\x00"
-     "\x7E\x89\x00\x80\x82\x89\x00\x80\x86\x89\x00\x80\x8A\x89\x00\x80\x8E\x89"
-     "\x00\x80"
-     "\xEF\xD4\x0E\x00\xEF\x9C\x0F\x00\xAB\x01\x0F\x00\xAA\x69\x0F\x00\x92\x89"
-     "\x00\x80"
-     "\x96\x89\x00\x80\xAF\x01\x0F\x00\xAE\x19\x0F\x00\xAD\x11\x0F\x00\xAC\x19"
-     "\x0F\x00"
-     "\xA3\x65\x0F\x00\x9A\x89\x00\x80\x9E\x89\x00\x80\xA2\x89\x00\x80\xA6\x89"
-     "\x00\x80"
-     "\xA6\x65\x0F\x00\xA5\x75\x0F\x00\xAA\x89\x00\x80\xA8\x2D\x06\x00\xA9\x3D"
-     "\x06\x00"
-     "\xAA\x35\x06\x00\xAB\x89\x06\x00\xAC\x99\x06\x00\xAD\x85\x06\x00\xAE\x85"
-     "\x06\x00"
-     "\xAF\xB5\x06\x00\x5E\x89\x00\x80\xAE\x89\x00\x80\xB2\x89\x00\x80\xB6\x89"
-     "\x00\x80"
-     "\xBA\x89\x00\x80\xBE\x89\x00\x80\xC2\x89\x00\x80\xC6\x89\x00\x80\xB8\x41"
-     "\x07\x00"
-     "\xB9\x41\x07\x00\xBA\x41\x07\x00\xBB\x41\x07\x00\xBC\x41\x07\x00\xBD\x4D"
-     "\x07\x00"
-     "\xBE\x79\x07\x00\xBF\x79\x07\x00\xB0\xCD\x06\x00\xB1\xA5\x06\x00\xB2\xA1"
-     "\x06\x00"
-     "\xB3\xA1\x06\x00\xB4\xA5\x06\x00\xB5\xAD\x06\x00\xB6\x91\x06\x00\xB7\x91"
-     "\x06\x00"
-     "\xA3\xBD\x06\x00\x82\x15\x00\x00\x81\xB5\x00\x00\x80\xB5\x00\x00\xCA\x89"
-     "\x00\x80"
-     "\xA6\x05\x06\x00\xA5\x11\x06\x00\xBE\x90\x03\x00\xAB\x61\x06\x00\xAA\x09"
-     "\x06\x00"
-     "\xCE\x89\x00\x80\xD2\x89\x00\x80\xAF\x69\x06\x00\xAE\x61\x06\x00\xAD\x75"
-     "\x06\x00"
-     "\xAC\x75\x06\x00\xD6\x89\x00\x80\xB3\x99\x07\x00\x86\xC8\x00\x00\x87\x1C"
-     "\x03\x00"
-     "\xB6\x49\x07\x00\xDA\x89\x00\x80\xDE\x89\x00\x80\xB5\x49\x07\x00\xBA\x49"
-     "\x07\x00"
-     "\xBB\x49\x07\x00\xE2\x89\x00\x80\xE6\x89\x00\x80\xBE\x39\x07\x00\xBF\x39"
-     "\x07\x00"
-     "\xBC\x49\x07\x00\xBD\x49\x07\x00\xA8\x39\x06\x00\xA9\x55\x06\x00\xAA\x55"
-     "\x06\x00"
-     "\xAB\x9D\x06\x00\xAC\x85\x06\x00\xAD\x8D\x06\x00\xAE\x85\x06\x00\xAF\xB9"
-     "\x06\x00"
-     "\xEA\x89\x00\x80\xEE\x89\x00\x80\xF2\x89\x00\x80\xF6\x89\x00\x80\xFA\x89"
-     "\x00\x80"
-     "\xFE\x89\x00\x80\x02\x8A\x00\x80\x06\x8A\x00\x80\xB8\xAD\x06\x00\xB9\xB5"
-     "\x06\x00"
-     "\xBA\xB5\x06\x00\xBB\x69\x01\x00\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x69"
-     "\x01\x00"
-     "\xBF\x69\x01\x00\xB0\xC9\x06\x00\xB1\xC9\x06\x00\xB2\xD9\x06\x00\xB3\xD9"
-     "\x06\x00"
-     "\xB4\xB9\x06\x00\xB5\xB9\x06\x00\xB6\x9D\x06\x00\xB7\x95\x06\x00\xA3\xDD"
-     "\x06\x00"
-     "\x0A\x8A\x00\x80\x0E\x8A\x00\x80\x12\x8A\x00\x80\x16\x8A\x00\x80\xA6\x0D"
-     "\x06\x00"
-     "\xA5\x0D\x06\x00\xEE\x88\x00\x80\xAB\x0D\x06\x00\xAA\x0D\x06\x00\x1A\x8A"
-     "\x00\x80"
-     "\x1E\x8A\x00\x80\xAF\x7D\x06\x00\xAE\x7D\x06\x00\xAD\x0D\x06\x00\xAC\x0D"
-     "\x06\x00"
-     "\x22\x8A\x00\x80\x26\x8A\x00\x80\x2A\x8A\x00\x80\x2E\x8A\x00\x80\x80\x0D"
-     "\x00\x00"
-     "\x81\x35\x00\x00\x82\x3D\x00\x00\x32\x8A\x00\x80\x36\x8A\x00\x80\x3A\x8A"
-     "\x00\x80"
-     "\x84\x64\x01\x00\xBE\x34\x00\x00\x86\x50\x1F\x00\x87\x68\x01\x00\x42\x8A"
-     "\x00\x80"
-     "\x46\x8A\x00\x80\x4A\x8A\x00\x80\x4E\x8A\x00\x80\x84\x04\x03\x00\xBE\x00"
-     "\x1C\x00"
-     "\xE3\x8C\x06\x00\x52\x8A\x00\x80\xE1\x30\x06\x00\x56\x8A\x00\x80\xEF\xE4"
-     "\x07\x00"
-     "\x5A\x8A\x00\x80\x5E\x8A\x00\x80\x62\x8A\x00\x80\x66\x8A\x00\x80\x6A\x8A"
-     "\x00\x80"
-     "\x6E\x8A\x00\x80\x72\x8A\x00\x80\xB3\x95\x02\x00\x76\x8A\x00\x80\x7A\x8A"
-     "\x00\x80"
-     "\x7E\x8A\x00\x80\x82\x8A\x00\x80\xB6\xB9\x02\x00\xB5\xB1\x02\x00\x86\x8A"
-     "\x00\x80"
-     "\xBB\x4D\x03\x00\xBA\x4D\x03\x00\x8A\x8A\x00\x80\x8E\x8A\x00\x80\xBF\x4D"
-     "\x03\x00"
-     "\xBE\x4D\x03\x00\xBD\x4D\x03\x00\xBC\x4D\x03\x00\x92\x8A\x00\x80\x96\x8A"
-     "\x00\x80"
-     "\x9A\x8A\x00\x80\xBE\x88\x1D\x00\x9E\x8A\x00\x80\x82\x9D\x01\x00\x81\x9D"
-     "\x01\x00"
-     "\x80\x9D\x01\x00\xA8\x0D\x02\x00\xA9\x65\x02\x00\xAA\x7D\x02\x00\xAB\x75"
-     "\x02\x00"
-     "\xAC\x59\x02\x00\xAD\x59\x02\x00\xAE\x99\x03\x00\xAF\x99\x03\x00\xB0\xE9"
-     "\x03\x00"
-     "\xB1\xE9\x03\x00\xB2\xF9\x03\x00\xB3\xF9\x03\x00\xB4\xE9\x03\x00\xB5\xE9"
-     "\x03\x00"
-     "\xB6\x5D\x03\x00\xB7\x55\x03\x00\xB8\x6D\x03\x00\xB9\x75\x03\x00\xBA\x41"
-     "\x03\x00"
-     "\xBB\x41\x03\x00\xBC\x31\x03\x00\xBD\x31\x03\x00\xBE\x31\x03\x00\xBF\x31"
-     "\x03\x00"
-     "\xA2\x8A\x00\x80\xA6\x8A\x00\x80\xAA\x8A\x00\x80\x84\x04\x1D\x00\xB2\x8A"
-     "\x00\x80"
-     "\xEF\x00\x00\x00\x86\xB8\x1C\x00\x87\x30\x1D\x00\xB6\x8A\x00\x80\xE3\xDC"
-     "\x00\x00"
-     "\xBA\x8A\x00\x80\xE1\xFC\x01\x00\xBE\x8A\x00\x80\xC2\x8A\x00\x80\xC6\x8A"
-     "\x00\x80"
-     "\xCA\x8A\x00\x80\xA6\xAD\x03\x00\xCE\x8A\x00\x80\xD2\x8A\x00\x80\xA5\xA5"
-     "\x03\x00"
-     "\xD6\x8A\x00\x80\xA3\x81\x03\x00\xDA\x8A\x00\x80\xDE\x8A\x00\x80\xAE\x59"
-     "\x02\x00"
-     "\xAF\x59\x02\x00\xAC\x59\x02\x00\xAD\x59\x02\x00\xAA\x59\x02\x00\xAB\x59"
-     "\x02\x00"
-     "\xE2\x8A\x00\x80\xE6\x8A\x00\x80\xEA\x8A\x00\x80\xEE\x8A\x00\x80\xF2\x8A"
-     "\x00\x80"
-     "\xF6\x8A\x00\x80\xFA\x8A\x00\x80\xFE\x8A\x00\x80\x02\x8B\x00\x80\xEF\xF8"
-     "\x01\x00"
-     "\x84\xAC\x1C\x00\xE1\xDC\x06\x00\x06\x8B\x00\x80\xE3\x1C\x06\x00\x0A\x8B"
-     "\x00\x80"
-     "\x0E\x8B\x00\x80\x12\x8B\x00\x80\x16\x8B\x00\x80\xB3\xC1\x01\x00\x1A\x8B"
-     "\x00\x80"
-     "\x86\xEC\x1F\x00\x87\xF0\x1C\x00\xF5\x00\x00\x00\xB6\xC9\x01\x00\xB5\xCD"
-     "\x01\x00"
-     "\x22\x8B\x00\x80\xBB\xD1\x01\x00\xBA\xC9\x01\x00\xBE\x2C\x1F\x00\x26\x8B"
-     "\x00\x80"
-     "\xBF\xB5\x01\x00\xBE\xC9\x01\x00\xBD\xC1\x01\x00\xBC\xC9\x01\x00\xA8\xF9"
-     "\x1D\x00"
-     "\xA9\xF9\x1D\x00\xAA\x09\x1E\x00\xAB\x09\x1E\x00\xAC\x05\x1E\x00\xAD\x09"
-     "\x1E\x00"
-     "\xAE\x3D\x1E\x00\xAF\x35\x1E\x00\xAE\x8A\x00\x80\x82\xE1\x1F\x00\x81\x95"
-     "\x1F\x00"
-     "\x80\x95\x1F\x00\x1E\x8B\x00\x80\x2A\x8B\x00\x80\x2E\x8B\x00\x80\x32\x8B"
-     "\x00\x80"
-     "\xB8\x11\x1E\x00\xB9\x19\x1E\x00\xBA\x21\x1E\x00\xBB\x21\x1E\x00\xBC\xD1"
-     "\x1E\x00"
-     "\xBD\xD1\x1E\x00\xBE\xD1\x1E\x00\xBF\xCD\x1E\x00\xB0\x4D\x1E\x00\xB1\x55"
-     "\x1E\x00"
-     "\xB2\x5D\x1E\x00\xB3\x55\x1E\x00\xB4\x4D\x1E\x00\xB5\x31\x1E\x00\xB6\x31"
-     "\x1E\x00"
-     "\xB7\x31\x1E\x00\xA3\x81\x1E\x00\x36\x8B\x00\x80\x3A\x8B\x00\x80\x3E\x8B"
-     "\x00\x80"
-     "\x42\x8B\x00\x80\xA6\x89\x1E\x00\xA5\x8D\x1E\x00\x46\x8B\x00\x80\xAB\x91"
-     "\x1E\x00"
-     "\xAA\x89\x1E\x00\x4A\x8B\x00\x80\x4E\x8B\x00\x80\xAF\xF5\x1E\x00\xAE\x89"
-     "\x1E\x00"
-     "\xAD\x81\x1E\x00\xAC\x89\x1E\x00\x52\x8B\x00\x80\xB3\xED\x1F\x00\x56\x8B"
-     "\x00\x80"
-     "\x5A\x8B\x00\x80\xB6\xE9\x1F\x00\x5E\x8B\x00\x80\x62\x8B\x00\x80\xB5\xE9"
-     "\x1F\x00"
-     "\xBA\xA9\x1F\x00\xBB\xB5\x1F\x00\x66\x8B\x00\x80\x6A\x8B\x00\x80\xBE\x89"
-     "\x1F\x00"
-     "\xBF\x89\x1F\x00\xBC\xA1\x1F\x00\xBD\xA1\x1F\x00\xA8\x11\x1E\x00\xA9\x25"
-     "\x1E\x00"
-     "\xAA\x21\x1E\x00\xAB\x3D\x1E\x00\xAC\x25\x1E\x00\xAD\x2D\x1E\x00\xAE\x25"
-     "\x1E\x00"
-     "\xAF\x59\x1E\x00\x6E\x8B\x00\x80\x72\x8B\x00\x80\x76\x8B\x00\x80\x7A\x8B"
-     "\x00\x80"
-     "\x7E\x8B\x00\x80\x82\xB9\x00\x00\x81\xBD\x00\x00\x80\xBD\x00\x00\xB8\xC9"
-     "\x1E\x00"
-     "\xB9\xC9\x1E\x00\xBA\xD9\x1E\x00\xBB\xD9\x1E\x00\xBC\xC9\x1E\x00\xBD\xFD"
-     "\x1E\x00"
-     "\xBE\xE9\x1E\x00\xBF\xE9\x1E\x00\xB0\x29\x1E\x00\xB1\x35\x1E\x00\xB2\x31"
-     "\x1E\x00"
-     "\xB3\x31\x1E\x00\xB4\x09\x1E\x00\xB5\x09\x1E\x00\xB6\xF9\x1E\x00\xB7\xF9"
-     "\x1E\x00"
-     "\x82\x8B\x00\x80\x86\x8B\x00\x80\xA3\xA1\x1E\x00\x8A\x8B\x00\x80\xA5\xA5"
-     "\x1E\x00"
-     "\x8E\x8B\x00\x80\x92\x8B\x00\x80\xA6\xA5\x1E\x00\x3E\x8A\x00\x80\x84\xA8"
-     "\x00\x00"
-     "\xAB\xF9\x1E\x00\xAA\xE5\x1E\x00\xAD\xED\x1E\x00\xAC\xED\x1E\x00\xAF\xC5"
-     "\x1E\x00"
-     "\xAE\xC5\x1E\x00\xB3\xF5\x01\x00\x96\x8B\x00\x80\x86\x6C\x00\x00\x87\x7C"
-     "\x01\x00"
-     "\x9A\x8B\x00\x80\xB6\x51\x01\x00\xB5\xE5\x01\x00\x9E\x8B\x00\x80\xBB\x7D"
-     "\x01\x00"
-     "\xBA\x75\x01\x00\xA2\x8B\x00\x80\xA6\x8B\x00\x80\xBF\x39\x01\x00\xBE\x35"
-     "\x01\x00"
-     "\xBD\x55\x01\x00\xBC\x55\x01\x00\xAA\x8B\x00\x80\xAE\x8B\x00\x80\xBE\x00"
-     "\x04\x00"
-     "\x84\x00\x04\x00\xE1\xC0\x01\x00\xB2\x8B\x00\x80\xE3\xAC\x01\x00\xEF\xE8"
-     "\x1E\x00"
-     "\xB6\x8B\x00\x80\xE1\xCC\x1D\x00\xEF\xD8\x02\x00\xE3\x34\x1E\x00\xBA\x8B"
-     "\x00\x80"
-     "\xBE\x8B\x00\x80\xC2\x8B\x00\x80\xC6\x8B\x00\x80\xCA\x8B\x00\x80\xCE\x8B"
-     "\x00\x80"
-     "\xA3\x7D\x02\x00\xD2\x8B\x00\x80\xA5\x6D\x02\x00\xD6\x8B\x00\x80\xDA\x8B"
-     "\x00\x80"
-     "\xA6\xD9\x02\x00\xDE\x8B\x00\x80\xE2\x8B\x00\x80\xAB\xF5\x02\x00\xAA\xFD"
-     "\x02\x00"
-     "\xAD\xDD\x02\x00\xAC\xDD\x02\x00\xAF\xB1\x02\x00\xAE\xBD\x02\x00\xA9\x6D"
-     "\x1D\x00"
-     "\xA8\x59\x1D\x00\xAB\xB1\x02\x00\xAA\xB9\x02\x00\xAD\x15\x02\x00\xAC\xA9"
-     "\x02\x00"
-     "\xAF\x15\x02\x00\xAE\x15\x02\x00\xBE\x2C\x05\x00\xE6\x8B\x00\x80\xEA\x8B"
-     "\x00\x80"
-     "\xEE\x8B\x00\x80\x80\x0D\x00\x00\x81\x95\x01\x00\x82\x9D\x01\x00\xF2\x8B"
-     "\x00\x80"
-     "\xB9\x29\x02\x00\xB8\x1D\x02\x00\xBB\xE9\x03\x00\xBA\x21\x02\x00\xBD\xF9"
-     "\x03\x00"
-     "\xBC\xF1\x03\x00\xBF\xE9\x03\x00\xBE\xF1\x03\x00\xB1\x7D\x02\x00\xB0\x75"
-     "\x02\x00"
-     "\xB3\x45\x02\x00\xB2\x45\x02\x00\xB5\x39\x02\x00\xB4\x35\x02\x00\xB7\x2D"
-     "\x02\x00"
-     "\xB6\x31\x02\x00\xA9\xC5\x02\x00\xA8\x9D\x02\x00\xAB\xD5\x02\x00\xAA\xDD"
-     "\x02\x00"
-     "\xAD\xFD\x02\x00\xAC\xFD\x02\x00\xAF\x1D\x01\x00\xAE\x1D\x01\x00\xF6\x8B"
-     "\x00\x80"
-     "\xFA\x8B\x00\x80\x87\xA0\x05\x00\x86\xAC\x05\x00\x02\x8C\x00\x80\x06\x8C"
-     "\x00\x80"
-     "\x0A\x8C\x00\x80\xCF\x00\x00\x00\xB9\x15\x01\x00\xB8\x09\x01\x00\xBB\x29"
-     "\x01\x00"
-     "\xBA\x15\x01\x00\xBD\xF5\x01\x00\xBC\x31\x01\x00\xBF\xF5\x01\x00\xBE\xF5"
-     "\x01\x00"
-     "\xB1\x41\x01\x00\xB0\x71\x01\x00\xB3\x5D\x01\x00\xB2\x5D\x01\x00\xB5\x4D"
-     "\x01\x00"
-     "\xB4\x4D\x01\x00\xB7\x3D\x01\x00\xB6\x3D\x01\x00\x0E\x8C\x00\x80\x12\x8C"
-     "\x00\x80"
-     "\x16\x8C\x00\x80\xEF\x9C\x1E\x00\x1A\x8C\x00\x80\x1E\x8C\x00\x80\x22\x8C"
-     "\x00\x80"
-     "\xEF\xE8\x1F\x00\x84\x6C\x07\x00\xE1\xA4\x1E\x00\x26\x8C\x00\x80\xE3\xB0"
-     "\x1E\x00"
-     "\x2A\x8C\x00\x80\xE1\x64\x1E\x00\x2E\x8C\x00\x80\xE3\x0C\x1E\x00\xB3\x11"
-     "\x02\x00"
-     "\x32\x8C\x00\x80\x36\x8C\x00\x80\x3A\x8C\x00\x80\x84\x5C\x06\x00\xB6\x39"
-     "\x02\x00"
-     "\xB5\x31\x02\x00\x3E\x8C\x00\x80\xBB\xC1\x01\x00\xBA\xF1\x01\x00\x42\x8C"
-     "\x00\x80"
-     "\x46\x8C\x00\x80\xBF\xCD\x01\x00\xBE\xCD\x01\x00\xBD\xCD\x01\x00\xBC\xCD"
-     "\x01\x00"
-     "\xFE\x8B\x00\x80\x4A\x8C\x00\x80\x4E\x8C\x00\x80\x52\x8C\x00\x80\x56\x8C"
-     "\x00\x80"
-     "\x5A\x8C\x00\x80\x5E\x8C\x00\x80\x62\x8C\x00\x80\xA8\x11\x07\x00\xA9\x11"
-     "\x07\x00"
-     "\xAA\x11\x07\x00\xAB\xAD\x07\x00\xAC\xB5\x07\x00\xAD\xBD\x07\x00\xAE\xA9"
-     "\x07\x00"
-     "\xAF\xA9\x07\x00\xB0\xD9\x07\x00\xB1\xED\x07\x00\xB2\xF9\x07\x00\xB3\xF5"
-     "\x07\x00"
-     "\xB4\x95\x07\x00\xB5\x9D\x07\x00\xB6\x95\x07\x00\xB7\x81\x07\x00\xB8\x89"
-     "\x07\x00"
-     "\xB9\x89\x07\x00\xBA\x6D\x00\x00\xBB\x65\x00\x00\xBC\x7D\x00\x00\xBD\x65"
-     "\x00\x00"
-     "\xBE\x6D\x00\x00\xBF\x65\x00\x00\x82\x19\x00\x00\xA3\x55\x05\x00\x80\x95"
-     "\x00\x00"
-     "\x81\x19\x00\x00\xA6\x7D\x05\x00\x66\x8C\x00\x80\x6A\x8C\x00\x80\xA5\x75"
-     "\x05\x00"
-     "\xAA\xB5\x06\x00\xAB\x85\x06\x00\x6E\x8C\x00\x80\x72\x8C\x00\x80\xAE\x89"
-     "\x06\x00"
-     "\xAF\x89\x06\x00\xAC\x89\x06\x00\xAD\x89\x06\x00\xB3\xF9\x07\x00\x76\x8C"
-     "\x00\x80"
-     "\x86\xE8\x03\x00\x87\xBC\x03\x00\x7A\x8C\x00\x80\xB6\x5D\x07\x00\xB5\x5D"
-     "\x07\x00"
-     "\x7E\x8C\x00\x80\xBB\x49\x07\x00\xBA\x7D\x07\x00\x82\x8C\x00\x80\x86\x8C"
-     "\x00\x80"
-     "\xBF\x39\x07\x00\xBE\x31\x07\x00\xBD\x41\x07\x00\xBC\x4D\x07\x00\xA3\x39"
-     "\x06\x00"
-     "\x8A\x8C\x00\x80\x8E\x8C\x00\x80\x92\x8C\x00\x80\x96\x8C\x00\x80\xA6\x9D"
-     "\x06\x00"
-     "\xA5\x9D\x06\x00\x9A\x8C\x00\x80\xAB\x89\x06\x00\xAA\xBD\x06\x00\x9E\x8C"
-     "\x00\x80"
-     "\xA2\x8C\x00\x80\xAF\xF9\x06\x00\xAE\xF1\x06\x00\xAD\x81\x06\x00\xAC\x8D"
-     "\x06\x00"
-     "\xA6\x8C\x00\x80\xAA\x8C\x00\x80\xAE\x8C\x00\x80\xB2\x8C\x00\x80\xB6\x8C"
-     "\x00\x80"
-     "\xBA\x8C\x00\x80\xBE\x8C\x00\x80\xC2\x8C\x00\x80\xC6\x8C\x00\x80\xCA\x8C"
-     "\x00\x80"
-     "\xCE\x8C\x00\x80\xD2\x8C\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xD9"
-     "\x00\x00"
-     "\xD6\x8C\x00\x80\xA8\x2D\x01\x00\xA9\x41\x01\x00\xAA\x59\x01\x00\xAB\x51"
-     "\x01\x00"
-     "\xAC\x79\x01\x00\xAD\x79\x01\x00\xAE\x9D\x00\x00\xAF\x95\x00\x00\xDA\x8C"
-     "\x00\x80"
-     "\xBE\x6C\x01\x00\x86\x88\x01\x00\x87\xAC\x00\x00\xE2\x8C\x00\x80\xE6\x8C"
-     "\x00\x80"
-     "\xEA\x8C\x00\x80\xEE\x8C\x00\x80\xB8\x8D\x00\x00\xB9\x95\x00\x00\xBA\x9D"
-     "\x00\x00"
-     "\xBB\x95\x00\x00\xBC\x8D\x00\x00\xBD\xB5\x00\x00\xBE\xBD\x00\x00\xBF\xB5"
-     "\x00\x00"
-     "\xB0\xED\x00\x00\xB1\xF5\x00\x00\xB2\xF5\x00\x00\xB3\xC5\x00\x00\xB4\xDD"
-     "\x00\x00"
-     "\xB5\xB5\x00\x00\xB6\xBD\x00\x00\xB7\xB5\x00\x00\xB3\xD1\x02\x00\xF2\x8C"
-     "\x00\x80"
-     "\xBE\x70\x03\x00\xBE\x44\x0C\x00\xF6\x8C\x00\x80\xB6\xF9\x02\x00\xB5\xF1"
-     "\x02\x00"
-     "\xFA\x8C\x00\x80\xBB\x31\x03\x00\xBA\x29\x03\x00\xFE\x8C\x00\x80\x02\x8D"
-     "\x00\x80"
-     "\xBF\x71\x03\x00\xBE\x19\x03\x00\xBD\x11\x03\x00\xBC\x21\x03\x00\x06\x8D"
-     "\x00\x80"
-     "\xA3\x95\x02\x00\x0A\x8D\x00\x80\x0E\x8D\x00\x80\xA6\xBD\x02\x00\x12\x8D"
-     "\x00\x80"
-     "\x16\x8D\x00\x80\xA5\xB5\x02\x00\xAA\x6D\x03\x00\xAB\x75\x03\x00\x1A\x8D"
-     "\x00\x80"
-     "\x1E\x8D\x00\x80\xAE\x5D\x03\x00\xAF\x35\x03\x00\xAC\x65\x03\x00\xAD\x55"
-     "\x03\x00"
-     "\xA8\x31\x03\x00\xA9\x31\x03\x00\xAA\x31\x03\x00\xAB\x31\x03\x00\xAC\x95"
-     "\x00\x00"
-     "\xAD\x85\x00\x00\xAE\x85\x00\x00\xAF\xB5\x00\x00\x22\x8D\x00\x80\x26\x8D"
-     "\x00\x80"
-     "\x2A\x8D\x00\x80\x2E\x8D\x00\x80\x32\x8D\x00\x80\x36\x8D\x00\x80\x3A\x8D"
-     "\x00\x80"
-     "\x3E\x8D\x00\x80\xB8\x61\x00\x00\xB9\x61\x00\x00\xBA\x61\x00\x00\xBB\x61"
-     "\x00\x00"
-     "\xBC\x61\x00\x00\xBD\x61\x00\x00\xBE\x61\x00\x00\xBF\x61\x00\x00\xB0\xCD"
-     "\x00\x00"
-     "\xB1\xA5\x00\x00\xB2\xA1\x00\x00\xB3\xA1\x00\x00\xB4\xA5\x00\x00\xB5\xAD"
-     "\x00\x00"
-     "\xB6\x91\x00\x00\xB7\x91\x00\x00\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x11"
-     "\x00\x00"
-     "\x42\x8D\x00\x80\x46\x8D\x00\x80\xEF\x48\x00\x00\x4A\x8D\x00\x80\x52\x8D"
-     "\x00\x80"
-     "\x84\x6C\x0D\x00\xE3\x94\x06\x00\x56\x8D\x00\x80\xE1\xFC\x01\x00\x5A\x8D"
-     "\x00\x80"
-     "\x5E\x8D\x00\x80\x86\xB8\x0D\x00\x87\xC4\x02\x00\xB3\xF9\x02\x00\x62\x8D"
-     "\x00\x80"
-     "\x66\x8D\x00\x80\x6A\x8D\x00\x80\x6E\x8D\x00\x80\xB6\x45\x01\x00\xB5\x5D"
-     "\x01\x00"
-     "\xBE\x98\x0F\x00\xBB\x21\x01\x00\xBA\x49\x01\x00\x72\x8D\x00\x80\x76\x8D"
-     "\x00\x80"
-     "\xBF\x29\x01\x00\xBE\x21\x01\x00\xBD\x35\x01\x00\xBC\x35\x01\x00\x7A\x8D"
-     "\x00\x80"
-     "\xE1\xB0\x06\x00\x7E\x8D\x00\x80\xE3\x90\x07\x00\x82\x8D\x00\x80\x86\x8D"
-     "\x00\x80"
-     "\x8A\x8D\x00\x80\x8E\x8D\x00\x80\x92\x8D\x00\x80\x96\x8D\x00\x80\x9A\x8D"
-     "\x00\x80"
-     "\x9E\x8D\x00\x80\xE5\x00\x00\x00\xA2\x8D\x00\x80\xA6\x8D\x00\x80\xEF\xE4"
-     "\x07\x00"
-     "\xAE\x95\x01\x00\xAF\x9D\x01\x00\xAC\x81\x01\x00\xAD\x81\x01\x00\xAA\xFD"
-     "\x01\x00"
-     "\xAB\x95\x01\x00\xAA\x8D\x00\x80\xAE\x8D\x00\x80\xA6\xF1\x01\x00\xB2\x8D"
-     "\x00\x80"
-     "\xB6\x8D\x00\x80\xA5\xE9\x01\x00\x82\xF5\x0F\x00\xA3\x4D\x02\x00\x80\x1D"
-     "\x00\x00"
-     "\x81\x15\x00\x00\x4E\x8D\x00\x80\xBA\x8D\x00\x80\xC2\x8D\x00\x80\xBE\x8D"
-     "\x00\x80"
-     "\x86\xB0\x03\x00\x87\x94\x03\x00\xC6\x8D\x00\x80\xCA\x8D\x00\x80\xA8\x3D"
-     "\x0F\x00"
-     "\xA9\x0D\x0F\x00\xAA\x19\x0F\x00\xAB\x19\x0F\x00\xAC\x75\x0F\x00\xAD\x7D"
-     "\x0F\x00"
-     "\xAE\x75\x0F\x00\xAF\x6D\x0F\x00\xB0\x29\x0F\x00\xB1\x29\x0F\x00\xB2\xC9"
-     "\x0F\x00"
-     "\xB3\xC5\x0F\x00\xB4\xC1\x0F\x00\xB5\xC1\x0F\x00\xB6\xC1\x0F\x00\xB7\xC1"
-     "\x0F\x00"
-     "\xB8\xC1\x0F\x00\xB9\xC1\x0F\x00\xBA\xC1\x0F\x00\xBB\xC1\x0F\x00\xBC\xC5"
-     "\x0F\x00"
-     "\xBD\xCD\x0F\x00\xBE\xA1\x0F\x00\xBF\xA1\x0F\x00\xB3\xB1\x0F\x00\xCE\x8D"
-     "\x00\x80"
-     "\xD2\x8D\x00\x80\xD6\x8D\x00\x80\xDA\x8D\x00\x80\xB6\xDD\x0F\x00\xB5\xDD"
-     "\x0F\x00"
-     "\xDE\x8D\x00\x80\xBB\xCD\x0F\x00\xBA\xCD\x0F\x00\xE2\x8D\x00\x80\xE6\x8D"
-     "\x00\x80"
-     "\xBF\x25\x0F\x00\xBE\x25\x0F\x00\xBD\x31\x0F\x00\xBC\x3D\x0F\x00\xEA\x8D"
-     "\x00\x80"
-     "\xA3\xF5\x0F\x00\xEE\x8D\x00\x80\xF2\x8D\x00\x80\xA6\x99\x0F\x00\xF6\x8D"
-     "\x00\x80"
-     "\xFA\x8D\x00\x80\xA5\x99\x0F\x00\xAA\x89\x0F\x00\xAB\x89\x0F\x00\xFE\x8D"
-     "\x00\x80"
-     "\x02\x8E\x00\x80\xAE\x61\x0F\x00\xAF\x61\x0F\x00\xAC\x79\x0F\x00\xAD\x75"
-     "\x0F\x00"
-     "\xA8\x09\x01\x00\xA9\x1D\x01\x00\xAA\x15\x01\x00\xAB\x89\x01\x00\xAC\x99"
-     "\x01\x00"
-     "\xAD\x99\x01\x00\xAE\x89\x01\x00\xAF\x81\x01\x00\x06\x8E\x00\x80\x0A\x8E"
-     "\x00\x80"
-     "\x0E\x8E\x00\x80\x12\x8E\x00\x80\x16\x8E\x00\x80\x1A\x8E\x00\x80\x1E\x8E"
-     "\x00\x80"
-     "\x22\x8E\x00\x80\xB8\xAD\x01\x00\xB9\x65\x00\x00\xBA\x6D\x00\x00\xBB\x65"
-     "\x00\x00"
-     "\xBC\x7D\x00\x00\xBD\x65\x00\x00\xBE\x6D\x00\x00\xBF\x65\x00\x00\xB0\xC9"
-     "\x01\x00"
-     "\xB1\xC9\x01\x00\xB2\xA9\x01\x00\xB3\xA5\x01\x00\xB4\xBD\x01\x00\xB5\xA1"
-     "\x01\x00"
-     "\xB6\xA1\x01\x00\xB7\x95\x01\x00\x80\xBD\x00\x00\x81\x09\x00\x00\x82\x19"
-     "\x00\x00"
-     "\x26\x8E\x00\x80\x2A\x8E\x00\x80\x84\x98\x02\x00\x2E\x8E\x00\x80\x32\x8E"
-     "\x00\x80"
-     "\x86\x80\x01\x00\x87\x04\x01\x00\x36\x8E\x00\x80\x3A\x8E\x00\x80\x3E\x8E"
-     "\x00\x80"
-     "\x42\x8E\x00\x80\xDE\x8C\x00\x80\x46\x8E\x00\x80\xB3\xDD\x01\x00\x4A\x8E"
-     "\x00\x80"
-     "\x4E\x8E\x00\x80\x52\x8E\x00\x80\x56\x8E\x00\x80\xB6\x59\x01\x00\xB5\x51"
-     "\x01\x00"
-     "\x5A\x8E\x00\x80\xBB\x4D\x01\x00\xBA\x4D\x01\x00\x5E\x8E\x00\x80\x62\x8E"
-     "\x00\x80"
-     "\xBF\x25\x01\x00\xBE\x39\x01\x00\xBD\x31\x01\x00\xBC\x3D\x01\x00\x66\x8E"
-     "\x00\x80"
-     "\xEF\xF4\x0E\x00\xBE\xD0\x03\x00\x6A\x8E\x00\x80\x6E\x8E\x00\x80\xE3\x48"
-     "\x02\x00"
-     "\x72\x8E\x00\x80\xE1\xA4\x01\x00\x76\x8E\x00\x80\xEF\xD4\x02\x00\x7A\x8E"
-     "\x00\x80"
-     "\x7E\x8E\x00\x80\x82\x8E\x00\x80\xE3\xF4\x0B\x00\x86\x8E\x00\x80\xE1\x8C"
-     "\x0D\x00"
-     "\x8A\x8E\x00\x80\x8E\x8E\x00\x80\xA3\x55\x02\x00\x84\x8C\x03\x00\xA5\xD9"
-     "\x02\x00"
-     "\x92\x8E\x00\x80\x96\x8E\x00\x80\xA6\xD1\x02\x00\x9A\x8E\x00\x80\xA2\x8E"
-     "\x00\x80"
-     "\xAB\xC5\x02\x00\xAA\xC5\x02\x00\xAD\xB9\x02\x00\xAC\xB5\x02\x00\xAF\xAD"
-     "\x02\x00"
-     "\xAE\xB1\x02\x00\xA8\x1D\x03\x00\xA9\x2D\x03\x00\xAA\x25\x03\x00\xAB\x3D"
-     "\x03\x00"
-     "\xAC\x25\x03\x00\xAD\x29\x03\x00\xAE\x5D\x03\x00\xAF\x55\x03\x00\x80\xF5"
-     "\x00\x00"
-     "\x81\xFD\x00\x00\x82\xC9\x00\x00\xA6\x8E\x00\x80\xAA\x8E\x00\x80\xAE\x8E"
-     "\x00\x80"
-     "\x87\x6C\x03\x00\x86\x1C\x04\x00\xB8\xD9\x00\x00\xB9\xE5\x00\x00\xBA\xE1"
-     "\x00\x00"
-     "\xBB\xE1\x00\x00\xBC\xE1\x00\x00\xBD\xED\x00\x00\xBE\x99\x00\x00\xBF\x99"
-     "\x00\x00"
-     "\xB0\x2D\x03\x00\xB1\x35\x03\x00\xB2\x35\x03\x00\xB3\x05\x03\x00\xB4\x1D"
-     "\x03\x00"
-     "\xB5\xF5\x00\x00\xB6\xF5\x00\x00\xB7\xED\x00\x00\xB2\x8E\x00\x80\xB6\x8E"
-     "\x00\x80"
-     "\xBA\x8E\x00\x80\xB3\xA1\x02\x00\xBE\x8E\x00\x80\xB5\xA1\x02\x00\xB6\xA5"
-     "\x02\x00"
-     "\xC2\x8E\x00\x80\xC6\x8E\x00\x80\xCA\x8E\x00\x80\xBA\xE5\x02\x00\xBB\xFD"
-     "\x02\x00"
-     "\xBC\xED\x02\x00\xBD\xDD\x02\x00\xBE\xD5\x02\x00\xBF\x3D\x01\x00\xA6\xFD"
-     "\x02\x00"
-     "\xCE\x8E\x00\x80\xD2\x8E\x00\x80\xA5\xF9\x02\x00\xD6\x8E\x00\x80\xA3\xF9"
-     "\x02\x00"
-     "\xDA\x8E\x00\x80\xDE\x8E\x00\x80\xAE\x8D\x02\x00\xAF\x65\x01\x00\xAC\xB5"
-     "\x02\x00"
-     "\xAD\x85\x02\x00\xAA\xBD\x02\x00\xAB\xA5\x02\x00\xE2\x8E\x00\x80\xE6\x8E"
-     "\x00\x80"
-     "\xE3\x04\x00\x00\xE1\xA4\x0E\x00\xE1\x80\x01\x00\xE3\xBC\x0E\x00\xEA\x8E"
-     "\x00\x80"
-     "\xEE\x8E\x00\x80\xF2\x8E\x00\x80\xF6\x8E\x00\x80\xBE\xE8\x05\x00\xFA\x8E"
-     "\x00\x80"
-     "\xFE\x8E\x00\x80\x02\x8F\x00\x80\xEF\x14\x0E\x00\x06\x8F\x00\x80\xBE\xF0"
-     "\x04\x00"
-     "\xEF\x48\x0F\x00\x0A\x8F\x00\x80\x12\x8F\x00\x80\x16\x8F\x00\x80\x1A\x8F"
-     "\x00\x80"
-     "\x81\x65\x00\x00\x80\x79\x00\x00\x1E\x8F\x00\x80\x82\x1D\x00\x00\x84\xAC"
-     "\x05\x00"
-     "\x22\x8F\x00\x80\x87\xC8\x04\x00\x86\x4C\x04\x00\x26\x8F\x00\x80\x2A\x8F"
-     "\x00\x80"
-     "\x2E\x8F\x00\x80\x32\x8F\x00\x80\xA8\x65\x06\x00\xA9\xA9\x06\x00\xAA\xB9"
-     "\x06\x00"
-     "\xAB\xB9\x06\x00\xAC\x11\x07\x00\xAD\x11\x07\x00\xAE\x11\x07\x00\xAF\x0D"
-     "\x07\x00"
-     "\x0E\x8F\x00\x80\x36\x8F\x00\x80\x3A\x8F\x00\x80\x3E\x8F\x00\x80\x42\x8F"
-     "\x00\x80"
-     "\x46\x8F\x00\x80\x4A\x8F\x00\x80\x4E\x8F\x00\x80\xB8\x39\x07\x00\xB9\x39"
-     "\x07\x00"
-     "\xBA\xD9\x07\x00\xBB\xD9\x07\x00\xBC\xC9\x07\x00\xBD\xC9\x07\x00\xBE\xFD"
-     "\x07\x00"
-     "\xBF\xF5\x07\x00\xB0\x7D\x07\x00\xB1\x45\x07\x00\xB2\x4D\x07\x00\xB3\x29"
-     "\x07\x00"
-     "\xB4\x39\x07\x00\xB5\x25\x07\x00\xB6\x21\x07\x00\xB7\x21\x07\x00\xB3\x31"
-     "\x06\x00"
-     "\x52\x8F\x00\x80\x56\x8F\x00\x80\x5A\x8F\x00\x80\x5E\x8F\x00\x80\xB6\x91"
-     "\x07\x00"
-     "\xB5\x99\x07\x00\x62\x8F\x00\x80\xBB\x81\x07\x00\xBA\xB9\x07\x00\x66\x8F"
-     "\x00\x80"
-     "\x6A\x8F\x00\x80\xBF\xE1\x07\x00\xBE\xFD\x07\x00\xBD\xFD\x07\x00\xBC\x99"
-     "\x07\x00"
-     "\x6E\x8F\x00\x80\xA3\x75\x06\x00\x72\x8F\x00\x80\x76\x8F\x00\x80\xA6\xD5"
-     "\x07\x00"
-     "\x7A\x8F\x00\x80\x84\x94\x03\x00\xA5\xDD\x07\x00\xAA\xFD\x07\x00\xAB\xC5"
-     "\x07\x00"
-     "\x7E\x8F\x00\x80\x82\x8F\x00\x80\xAE\xB9\x07\x00\xAF\xA5\x07\x00\xAC\xDD"
-     "\x07\x00"
-     "\xAD\xB9\x07\x00\xA8\xC5\x06\x00\xA9\xCD\x06\x00\xAA\xC5\x06\x00\xAB\xD5"
-     "\x06\x00"
-     "\xAC\xF5\x06\x00\xAD\xE5\x06\x00\xAE\xED\x06\x00\xAF\xE5\x06\x00\x80\xE9"
-     "\x01\x00"
-     "\x81\xE9\x01\x00\x82\xFD\x01\x00\x86\x8F\x00\x80\x86\x30\x01\x00\x87\x64"
-     "\x01\x00"
-     "\x8A\x8F\x00\x80\x8E\x8F\x00\x80\xB8\xC5\x01\x00\xB9\xCD\x01\x00\xBA\xC5"
-     "\x01\x00"
-     "\xBB\xDD\x01\x00\xBC\xCD\x01\x00\xBD\xFD\x01\x00\xBE\xF5\x01\x00\xBF\x9D"
-     "\x01\x00"
-     "\xB0\x9D\x06\x00\xB1\x41\x01\x00\xB2\x41\x01\x00\xB3\x41\x01\x00\xB4\x41"
-     "\x01\x00"
-     "\xB5\x41\x01\x00\xB6\x41\x01\x00\xB7\x41\x01\x00\xB3\x4D\x06\x00\x92\x8F"
-     "\x00\x80"
-     "\x96\x8F\x00\x80\x9A\x8F\x00\x80\x9E\x8F\x00\x80\xB6\x65\x06\x00\xB5\x7D"
-     "\x06\x00"
-     "\xA2\x8F\x00\x80\xBB\x85\x01\x00\xBA\x59\x06\x00\xA6\x8F\x00\x80\xAA\x8F"
-     "\x00\x80"
-     "\xBF\x85\x01\x00\xBE\x85\x01\x00\xBD\x95\x01\x00\xBC\x95\x01\x00\xAE\x8F"
-     "\x00\x80"
-     "\xA3\x09\x06\x00\x9E\x8E\x00\x80\xB2\x8F\x00\x80\xA6\x21\x06\x00\xB6\x8F"
-     "\x00\x80"
-     "\xBA\x8F\x00\x80\xA5\x39\x06\x00\xAA\x1D\x06\x00\xAB\xC1\x01\x00\xBE\x8F"
-     "\x00\x80"
-     "\xC2\x8F\x00\x80\xAE\xC1\x01\x00\xAF\xC1\x01\x00\xAC\xD1\x01\x00\xAD\xD1"
-     "\x01\x00"
-     "\xC6\x8F\x00\x80\xCA\x8F\x00\x80\xCE\x8F\x00\x80\xD2\x8F\x00\x80\xD6\x8F"
-     "\x00\x80"
-     "\xDA\x8F\x00\x80\xDE\x8F\x00\x80\xE2\x8F\x00\x80\xE6\x8F\x00\x80\xEA\x8F"
-     "\x00\x80"
-     "\xBE\x64\x03\x00\xF2\x8F\x00\x80\xF6\x8F\x00\x80\xFA\x8F\x00\x80\xBF\x30"
-     "\x3D\x00"
-     "\xFE\x8F\x00\x80\x85\x00\x3C\x00\xE3\xB4\x07\x00\x84\x8C\x03\x00\xE1\xE8"
-     "\x06\x00"
-     "\x80\x4D\x00\x00\x81\x75\x00\x00\x82\x7D\x00\x00\xEF\x8C\x2A\x00\x85\x44"
-     "\x3C\x00"
-     "\xE1\x10\x02\x00\x02\x90\x00\x80\xE3\xA8\x2E\x00\x06\x90\x00\x80\xEF\x2C"
-     "\x07\x00"
-     "\x86\x58\x3D\x00\x87\x0C\x03\x00\x0A\x90\x00\x80\x0E\x90\x00\x80\x12\x90"
-     "\x00\x80"
-     "\x16\x90\x00\x80\x1A\x90\x00\x80\x1E\x90\x00\x80\x22\x90\x00\x80\x26\x90"
-     "\x00\x80"
-     "\xA8\x4D\x02\x00\xA9\x89\x03\x00\xAA\x91\x03\x00\xAB\x91\x03\x00\xAC\x35"
-     "\x03\x00"
-     "\xAD\x3D\x03\x00\xAE\x35\x03\x00\xAF\x2D\x03\x00\xB0\x5D\x03\x00\xB1\x6D"
-     "\x03\x00"
-     "\xB2\x65\x03\x00\xB3\x0D\x03\x00\xB4\x1D\x03\x00\xB5\x01\x03\x00\xB6\x01"
-     "\x03\x00"
-     "\xB7\x35\x03\x00\xB8\x0D\x03\x00\xB9\xE5\x00\x00\xBA\xED\x00\x00\xBB\xE5"
-     "\x00\x00"
-     "\xBC\xFD\x00\x00\xBD\xE5\x00\x00\xBE\xED\x00\x00\xBF\xE5\x00\x00\x2A\x90"
-     "\x00\x80"
-     "\x2E\x90\x00\x80\xB3\x31\x02\x00\x32\x90\x00\x80\xB5\x95\x02\x00\x36\x90"
-     "\x00\x80"
-     "\x3A\x90\x00\x80\xB6\x95\x02\x00\x84\xBC\x3D\x00\x3E\x90\x00\x80\xBB\x81"
-     "\x02\x00"
-     "\xBA\xB1\x02\x00\xBD\xF1\x02\x00\xBC\x99\x02\x00\xBF\xED\x02\x00\xBE\xF1"
-     "\x02\x00"
-     "\x42\x90\x00\x80\x46\x90\x00\x80\xBE\xC4\x3F\x00\x84\xC4\x3F\x00\x4A\x90"
-     "\x00\x80"
-     "\xEF\x84\x03\x00\x4E\x90\x00\x80\x52\x90\x00\x80\x56\x90\x00\x80\xE3\x3C"
-     "\x03\x00"
-     "\x5A\x90\x00\x80\xE1\x44\x00\x00\x80\x2D\x00\x00\x81\x15\x00\x00\x82\x11"
-     "\x00\x00"
-     "\x62\x90\x00\x80\x86\xA0\x3F\x00\x87\x58\x3D\x00\x66\x90\x00\x80\x6A\x90"
-     "\x00\x80"
-     "\x6E\x90\x00\x80\x72\x90\x00\x80\x76\x90\x00\x80\xEF\xDC\x06\x00\xBE\x2C"
-     "\x3E\x00"
-     "\xE1\x50\x06\x00\x7A\x90\x00\x80\xE3\x64\x01\x00\x7E\x90\x00\x80\x82\x90"
-     "\x00\x80"
-     "\x86\x90\x00\x80\x8A\x90\x00\x80\xA3\xF9\x02\x00\x8E\x90\x00\x80\x92\x90"
-     "\x00\x80"
-     "\x96\x90\x00\x80\x9A\x90\x00\x80\xA6\x5D\x02\x00\xA5\x5D\x02\x00\x9E\x90"
-     "\x00\x80"
-     "\xAB\x49\x02\x00\xAA\x79\x02\x00\xA2\x90\x00\x80\xA6\x90\x00\x80\xAF\x25"
-     "\x02\x00"
-     "\xAE\x39\x02\x00\xAD\x39\x02\x00\xAC\x51\x02\x00\x90\x6D\x24\x00\x91\x05"
-     "\x28\x00"
-     "\x92\x0D\x28\x00\x93\x9D\x28\x00\x94\x8D\x2D\x00\x95\x35\x2C\x00\x96\xB1"
-     "\x2C\x00"
-     "\x97\xBD\x31\x00\x98\x01\x30\x00\x99\x15\x30\x00\x9A\xBD\x31\x00\x9B\x2D"
-     "\x34\x00"
-     "\x9C\x99\x34\x00\x9D\x8D\x39\x00\x9E\x39\x38\x00\x9F\x31\x38\x00\xEE\x8F"
-     "\x00\x80"
-     "\xB6\x35\x3E\x00\xB5\x21\x3E\x00\x5E\x90\x00\x80\xB3\x65\x3E\x00\xAA\x90"
-     "\x00\x80"
-     "\xAE\x90\x00\x80\xB2\x90\x00\x80\xBF\xFD\x3E\x00\xBE\xFD\x3E\x00\xBD\xFD"
-     "\x3E\x00"
-     "\xBC\xFD\x3E\x00\xBB\xFD\x3E\x00\xBA\xFD\x3E\x00\xB6\x90\x00\x80\x8F\x39"
-     "\x24\x00"
-     "\xA9\x45\x09\x00\xA8\x9D\x08\x00\xAB\xA9\x0C\x00\xAA\x01\x0C\x00\xAD\x01"
-     "\x10\x00"
-     "\xAC\x1D\x0D\x00\xAF\x35\x10\x00\xAE\xBD\x10\x00\xA1\xE1\x00\x00\xBA\x90"
-     "\x00\x80"
-     "\xA3\x01\x04\x00\xA2\xE1\x00\x00\xA5\x95\x04\x00\xA4\x95\x04\x00\xA7\xB9"
-     "\x08\x00"
-     "\xA6\x01\x08\x00\xA0\x59\x39\x00\xA1\x0D\x3D\x00\xA2\xE9\x3C\x00\xA3\x25"
-     "\x3C\x00"
-     "\xBE\x90\x00\x80\x82\xDD\x01\x00\x81\x25\x00\x00\x80\x25\x00\x00\xB1\xC1"
-     "\x14\x00"
-     "\xB0\x01\x14\x00\xB3\x01\x18\x00\xB2\xD5\x14\x00\xB5\xD5\x18\x00\xB4\xD5"
-     "\x18\x00"
-     "\xC2\x90\x00\x80\xC6\x90\x00\x80\x82\x45\x25\x00\x83\x69\x25\x00\xCA\x90"
-     "\x00\x80"
-     "\xD2\x90\x00\x80\x86\x81\x29\x00\x87\x89\x2C\x00\x84\x19\x29\x00\x85\x19"
-     "\x29\x00"
-     "\x8A\x01\x2D\x00\x8B\x8D\x2D\x00\x86\x88\x07\x00\x87\x8C\x01\x00\x8E\x01"
-     "\x31\x00"
-     "\x8F\xA1\x30\x00\x8C\x99\x30\x00\x8D\x1D\x31\x00\x92\x41\x35\x00\x93\xED"
-     "\x35\x00"
-     "\xD6\x90\x00\x80\xDA\x90\x00\x80\x96\x5D\x39\x00\x97\xCD\x38\x00\x94\xE1"
-     "\x38\x00"
-     "\x95\x5D\x39\x00\x9A\xA9\x3C\x00\x9B\x45\x3D\x00\xDE\x90\x00\x80\xE2\x90"
-     "\x00\x80"
-     "\xE6\x90\x00\x80\xEA\x90\x00\x80\x9C\x91\x3D\x00\xEE\x90\x00\x80\xA8\xAD"
-     "\x3F\x00"
-     "\xA9\xB5\x3F\x00\xAA\xBD\x3F\x00\xAB\xB1\x3F\x00\xAC\xF9\x3F\x00\xAD\xE5"
-     "\x3F\x00"
-     "\xAE\x19\x3F\x00\xAF\x19\x3F\x00\xF2\x90\x00\x80\xF6\x90\x00\x80\xFA\x90"
-     "\x00\x80"
-     "\xFE\x90\x00\x80\x02\x91\x00\x80\x06\x91\x00\x80\x0A\x91\x00\x80\x0E\x91"
-     "\x00\x80"
-     "\xB8\xCD\x3E\x00\xB9\xD1\x3E\x00\xBA\xD1\x3E\x00\xBB\xE9\x3E\x00\xBC\x91"
-     "\x3E\x00"
-     "\xBD\x91\x3E\x00\xBE\x89\x3E\x00\xBF\x85\x3E\x00\xB0\x69\x3F\x00\xB1\x69"
-     "\x3F\x00"
-     "\xB2\x79\x3F\x00\xB3\x79\x3F\x00\xB4\x69\x3F\x00\xB5\x69\x3F\x00\xB6\xF9"
-     "\x3E\x00"
-     "\xB7\xF9\x3E\x00\xA3\x35\x3C\x00\x12\x91\x00\x80\x16\x91\x00\x80\x1A\x91"
-     "\x00\x80"
-     "\x1E\x91\x00\x80\xA6\x65\x3C\x00\xA5\x71\x3C\x00\x22\x91\x00\x80\xAB\xAD"
-     "\x3C\x00"
-     "\xAA\xAD\x3C\x00\x26\x91\x00\x80\x2A\x91\x00\x80\xAF\xAD\x3C\x00\xAE\xAD"
-     "\x3C\x00"
-     "\xAD\xAD\x3C\x00\xAC\xAD\x3C\x00\x84\xC8\x03\x00\xB3\xC9\x3D\x00\x2E\x91"
-     "\x00\x80"
-     "\x32\x91\x00\x80\xB6\xC9\x3D\x00\x36\x91\x00\x80\x3A\x91\x00\x80\xB5\xC9"
-     "\x3D\x00"
-     "\xBA\xD5\x3D\x00\xBB\xDD\x3D\x00\xBE\x6C\x03\x00\x42\x91\x00\x80\xBE\xB9"
-     "\x3D\x00"
-     "\xBF\xB9\x3D\x00\xBC\xC5\x3D\x00\xBD\xB9\x3D\x00\xA8\xFD\x3D\x00\xA9\x0D"
-     "\x3D\x00"
-     "\xAA\x05\x3D\x00\xAB\x1D\x3D\x00\xAC\x05\x3D\x00\xAD\x0D\x3D\x00\xAE\x31"
-     "\x3D\x00"
-     "\xAF\x2D\x3D\x00\x80\xED\x00\x00\x81\xF1\x00\x00\x82\xF1\x00\x00\x46\x91"
-     "\x00\x80"
-     "\x86\x90\x04\x00\x87\x64\x03\x00\x4A\x91\x00\x80\x4E\x91\x00\x80\xB8\x29"
-     "\x3D\x00"
-     "\xB9\x29\x3D\x00\xBA\x9D\x02\x00\xBB\x95\x02\x00\xBC\x8D\x02\x00\xBD\xAD"
-     "\x02\x00"
-     "\xBE\x7D\x02\x00\xBF\x0D\x02\x00\xB0\x55\x3D\x00\xB1\x5D\x3D\x00\xB2\x55"
-     "\x3D\x00"
-     "\xB3\x29\x3D\x00\xB4\x39\x3D\x00\xB5\x39\x3D\x00\xB6\x29\x3D\x00\xB7\x21"
-     "\x3D\x00"
-     "\xA3\x8D\x3C\x00\x52\x91\x00\x80\x84\xC4\x04\x00\x56\x91\x00\x80\x5A\x91"
-     "\x00\x80"
-     "\xA6\x8D\x3C\x00\xA5\x8D\x3C\x00\x5E\x91\x00\x80\xAB\x99\x3C\x00\xAA\x91"
-     "\x3C\x00"
-     "\x62\x91\x00\x80\x66\x91\x00\x80\xAF\xFD\x3C\x00\xAE\xFD\x3C\x00\xAD\xFD"
-     "\x3C\x00"
-     "\xAC\x81\x3C\x00\x6A\x91\x00\x80\xB3\x09\x02\x00\x6E\x91\x00\x80\x72\x91"
-     "\x00\x80"
-     "\xB6\x21\x02\x00\x76\x91\x00\x80\x7A\x91\x00\x80\xB5\x31\x02\x00\xBA\xC1"
-     "\x02\x00"
-     "\xBB\xC1\x02\x00\x7E\x91\x00\x80\x82\x91\x00\x80\xBE\xC1\x02\x00\xBF\xC1"
-     "\x02\x00"
-     "\xBC\xD1\x02\x00\xBD\xD1\x02\x00\x86\x91\x00\x80\x8A\x91\x00\x80\xA3\xC5"
-     "\x02\x00"
-     "\x8E\x91\x00\x80\xA5\xFD\x02\x00\x92\x91\x00\x80\x96\x91\x00\x80\xA6\xED"
-     "\x02\x00"
-     "\x9A\x91\x00\x80\x9E\x91\x00\x80\xAB\x0D\x02\x00\xAA\x0D\x02\x00\xAD\x1D"
-     "\x02\x00"
-     "\xAC\x1D\x02\x00\xAF\x0D\x02\x00\xAE\x0D\x02\x00\xE3\x9C\x3E\x00\xE1\xA4"
-     "\x3F\x00"
-     "\xE1\x50\x3E\x00\xE3\xFC\x3E\x00\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D"
-     "\x00\x00"
-     "\xA2\x91\x00\x80\xA6\x91\x00\x80\xAA\x91\x00\x80\xBE\x20\x04\x00\xB2\x91"
-     "\x00\x80"
-     "\xEF\xA8\x3E\x00\xB6\x91\x00\x80\xBA\x91\x00\x80\xEF\x4C\x3E\x00\xA8\x0D"
-     "\x05\x00"
-     "\xA9\x4D\x05\x00\xAA\x45\x05\x00\xAB\x65\x05\x00\xAC\x7D\x05\x00\xAD\x65"
-     "\x05\x00"
-     "\xAE\x6D\x05\x00\xAF\x61\x05\x00\x3E\x91\x00\x80\xAE\x91\x00\x80\x86\x08"
-     "\x03\x00"
-     "\x87\x8C\x03\x00\xBE\x91\x00\x80\xC2\x91\x00\x80\xC6\x91\x00\x80\xCA\x91"
-     "\x00\x80"
-     "\xB8\x45\x06\x00\xB9\x4D\x06\x00\xBA\x45\x06\x00\xBB\x5D\x06\x00\xBC\x4D"
-     "\x06\x00"
-     "\xBD\x7D\x06\x00\xBE\x75\x06\x00\xBF\x15\x06\x00\xB0\xC1\x05\x00\xB1\xC1"
-     "\x05\x00"
-     "\xB2\xC1\x05\x00\xB3\xC1\x05\x00\xB4\xC1\x05\x00\xB5\xC1\x05\x00\xB6\xC1"
-     "\x05\x00"
-     "\xB7\xC1\x05\x00\xA8\xED\x06\x00\xA9\x29\x06\x00\xAA\x31\x06\x00\xAB\x31"
-     "\x06\x00"
-     "\xAC\x95\x07\x00\xAD\x99\x07\x00\xAE\x89\x07\x00\xAF\x81\x07\x00\xCE\x91"
-     "\x00\x80"
-     "\xD2\x91\x00\x80\xCE\x90\x00\x80\xD6\x91\x00\x80\xDA\x91\x00\x80\xDE\x91"
-     "\x00\x80"
-     "\xE2\x91\x00\x80\xE6\x91\x00\x80\xB8\xAD\x07\x00\xB9\x45\x07\x00\xBA\x4D"
-     "\x07\x00"
-     "\xBB\x41\x07\x00\xBC\x45\x07\x00\xBD\x4D\x07\x00\xBE\x79\x07\x00\xBF\x79"
-     "\x07\x00"
-     "\xB0\xC9\x07\x00\xB1\xC9\x07\x00\xB2\xA9\x07\x00\xB3\xA5\x07\x00\xB4\xBD"
-     "\x07\x00"
-     "\xB5\xA1\x07\x00\xB6\xA1\x07\x00\xB7\x95\x07\x00\xB3\xB9\x07\x00\xEA\x91"
-     "\x00\x80"
-     "\xEE\x91\x00\x80\xF2\x91\x00\x80\xF6\x91\x00\x80\xB6\x19\x06\x00\xB5\x1D"
-     "\x06\x00"
-     "\xFA\x91\x00\x80\xBB\x0D\x06\x00\xBA\x0D\x06\x00\xFE\x91\x00\x80\x02\x92"
-     "\x00\x80"
-     "\xBF\x65\x06\x00\xBE\x79\x06\x00\xBD\x71\x06\x00\xBC\x7D\x06\x00\x06\x92"
-     "\x00\x80"
-     "\xA3\xFD\x07\x00\x0A\x92\x00\x80\x0E\x92\x00\x80\xA6\x5D\x06\x00\x12\x92"
-     "\x00\x80"
-     "\x16\x92\x00\x80\xA5\x59\x06\x00\xAA\x49\x06\x00\xAB\x49\x06\x00\x1A\x92"
-     "\x00\x80"
-     "\x84\xA0\x01\x00\xAE\x3D\x06\x00\xAF\x21\x06\x00\xAC\x39\x06\x00\xAD\x35"
-     "\x06\x00"
-     "\xA8\x6D\x06\x00\xA9\x71\x06\x00\xAA\x71\x06\x00\xAB\xBD\x01\x00\xAC\xA5"
-     "\x01\x00"
-     "\xAD\xAD\x01\x00\xAE\xA5\x01\x00\xAF\x3D\x01\x00\x80\xF5\x00\x00\x81\xFD"
-     "\x00\x00"
-     "\x82\xC1\x00\x00\xBE\x64\x01\x00\x86\x98\x01\x00\x87\xAC\x00\x00\x22\x92"
-     "\x00\x80"
-     "\x26\x92\x00\x80\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5\x00\x00\xBB\xE5"
-     "\x00\x00"
-     "\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x91\x00\x00\xBF\x91\x00\x00\xB0\x45"
-     "\x01\x00"
-     "\xB1\x49\x01\x00\xB2\x59\x01\x00\xB3\x59\x01\x00\xB4\x49\x01\x00\xB5\x49"
-     "\x01\x00"
-     "\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xB3\xF9\x05\x00\x2A\x92\x00\x80\x2E\x92"
-     "\x00\x80"
-     "\x32\x92\x00\x80\x36\x92\x00\x80\xB6\x2D\x02\x00\xB5\x2D\x02\x00\x3A\x92"
-     "\x00\x80"
-     "\xBB\x8D\x02\x00\xBA\x81\x02\x00\x3E\x92\x00\x80\x42\x92\x00\x80\xBF\x8D"
-     "\x02\x00"
-     "\xBE\x8D\x02\x00\xBD\x8D\x02\x00\xBC\x8D\x02\x00\x46\x92\x00\x80\xA3\xBD"
-     "\x05\x00"
-     "\x4A\x92\x00\x80\x4E\x92\x00\x80\xA6\x69\x02\x00\x52\x92\x00\x80\x56\x92"
-     "\x00\x80"
-     "\xA5\x69\x02\x00\xAA\xC5\x02\x00\xAB\xC9\x02\x00\x5A\x92\x00\x80\x5E\x92"
-     "\x00\x80"
-     "\xAE\xC9\x02\x00\xAF\xC9\x02\x00\xAC\xC9\x02\x00\xAD\xC9\x02\x00\xAA\x85"
-     "\x00\x00"
-     "\xAB\x8D\x00\x00\xA8\x31\x03\x00\xA9\x31\x03\x00\xAE\x8D\x00\x00\xAF\xBD"
-     "\x00\x00"
-     "\xAC\x95\x00\x00\xAD\x8D\x00\x00\xBE\xE4\x03\x00\x62\x92\x00\x80\x66\x92"
-     "\x00\x80"
-     "\x6A\x92\x00\x80\x6E\x92\x00\x80\x72\x92\x00\x80\x76\x92\x00\x80\x7A\x92"
-     "\x00\x80"
-     "\xBA\x6D\x00\x00\xBB\x75\x00\x00\xB8\x6D\x00\x00\xB9\x65\x00\x00\xBE\x6D"
-     "\x00\x00"
-     "\xBF\xD5\x01\x00\xBC\x6D\x00\x00\xBD\x65\x00\x00\xB2\xA9\x00\x00\xB3\xA9"
-     "\x00\x00"
-     "\xB0\xC5\x00\x00\xB1\xAD\x00\x00\xB6\x99\x00\x00\xB7\x99\x00\x00\xB4\xAD"
-     "\x00\x00"
-     "\xB5\xA5\x00\x00\x7E\x92\x00\x80\x82\x92\x00\x80\x86\x92\x00\x80\x8A\x92"
-     "\x00\x80"
-     "\x82\x1D\x00\x00\xEF\x84\x07\x00\x80\x1D\x00\x00\x81\x15\x00\x00\x84\xAC"
-     "\x02\x00"
-     "\xE3\x74\x07\x00\x92\x92\x00\x80\xE1\x1C\x04\x00\x96\x92\x00\x80\x9A\x92"
-     "\x00\x80"
-     "\x86\xB8\x0D\x00\x87\x30\x0D\x00\xB3\x45\x01\x00\x9E\x92\x00\x80\xA2\x92"
-     "\x00\x80"
-     "\xA6\x92\x00\x80\xAA\x92\x00\x80\xB6\x45\x01\x00\xB5\x55\x01\x00\xAE\x92"
-     "\x00\x80"
-     "\xBB\x21\x01\x00\xBA\x49\x01\x00\xB2\x92\x00\x80\xB6\x92\x00\x80\xBF\x29"
-     "\x01\x00"
-     "\xBE\x21\x01\x00\xBD\x35\x01\x00\xBC\x35\x01\x00\xBA\x92\x00\x80\xBE\x92"
-     "\x00\x80"
-     "\xC2\x92\x00\x80\xC6\x92\x00\x80\xCA\x92\x00\x80\xCE\x92\x00\x80\xEF\xCC"
-     "\x00\x00"
-     "\xEF\x8C\x06\x00\xE1\x90\x01\x00\xE1\x0C\x07\x00\xE3\xD4\x00\x00\xE3\xB4"
-     "\x06\x00"
-     "\xD2\x92\x00\x80\xD6\x92\x00\x80\xDA\x92\x00\x80\xDE\x92\x00\x80\xE2\x92"
-     "\x00\x80"
-     "\xE6\x92\x00\x80\xA3\xCD\x01\x00\x84\x90\x0D\x00\xA5\xDD\x01\x00\xEA\x92"
-     "\x00\x80"
-     "\xEE\x92\x00\x80\xA6\xCD\x01\x00\xF2\x92\x00\x80\xF6\x92\x00\x80\xAB\xA9"
-     "\x01\x00"
-     "\xAA\xC1\x01\x00\xAD\xBD\x01\x00\xAC\xBD\x01\x00\xAF\xA1\x01\x00\xAE\xA9"
-     "\x01\x00"
-     "\x8E\x92\x00\x80\x82\x15\x00\x00\x81\xED\x0F\x00\x80\xED\x0F\x00\xFA\x92"
-     "\x00\x80"
-     "\xBE\x58\x02\x00\xFE\x92\x00\x80\x02\x93\x00\x80\xA8\x09\x0F\x00\xA9\x15"
-     "\x0F\x00"
-     "\xAA\x1D\x0F\x00\xAB\x69\x0F\x00\xAC\x79\x0F\x00\xAD\x79\x0F\x00\xAE\x69"
-     "\x0F\x00"
-     "\xAF\x61\x0F\x00\xB0\x29\x0F\x00\xB1\x29\x0F\x00\xB2\xC9\x0F\x00\xB3\xC9"
-     "\x0F\x00"
-     "\xB4\xD9\x0F\x00\xB5\xD9\x0F\x00\xB6\xC9\x0F\x00\xB7\xC9\x0F\x00\xB8\xD1"
-     "\x0F\x00"
-     "\xB9\xD1\x0F\x00\xBA\xD1\x0F\x00\xBB\xD1\x0F\x00\xBC\xB1\x0F\x00\xBD\xB1"
-     "\x0F\x00"
-     "\xBE\xAD\x0F\x00\xBF\xA5\x0F\x00\x06\x93\x00\x80\xB6\xC9\x0F\x00\xB5\xC1"
-     "\x0F\x00"
-     "\x0A\x93\x00\x80\xB3\xA5\x0F\x00\x0E\x93\x00\x80\x87\x08\x03\x00\x86\x3C"
-     "\x03\x00"
-     "\xBF\x2D\x0F\x00\xBE\x2D\x0F\x00\xBD\x2D\x0F\x00\xBC\x2D\x0F\x00\xBB\xDD"
-     "\x0F\x00"
-     "\xBA\xDD\x0F\x00\x12\x93\x00\x80\x1E\x92\x00\x80\xA3\xF5\x0F\x00\x16\x93"
-     "\x00\x80"
-     "\x1A\x93\x00\x80\x1E\x93\x00\x80\x22\x93\x00\x80\xA6\x99\x0F\x00\xA5\x91"
-     "\x0F\x00"
-     "\x26\x93\x00\x80\xAB\x8D\x0F\x00\xAA\x8D\x0F\x00\x2A\x93\x00\x80\x2E\x93"
-     "\x00\x80"
-     "\xAF\x7D\x0F\x00\xAE\x7D\x0F\x00\xAD\x7D\x0F\x00\xAC\x7D\x0F\x00\xA8\x1D"
-     "\x01\x00"
-     "\xA9\x85\x01\x00\xAA\x85\x01\x00\xAB\x99\x01\x00\xAC\xA9\x01\x00\xAD\xD1"
-     "\x01\x00"
-     "\xAE\xD1\x01\x00\xAF\xD1\x01\x00\x32\x93\x00\x80\x36\x93\x00\x80\x3A\x93"
-     "\x00\x80"
-     "\x3E\x93\x00\x80\x42\x93\x00\x80\x46\x93\x00\x80\x4A\x93\x00\x80\x4E\x93"
-     "\x00\x80"
-     "\xB8\x71\x00\x00\xB9\x71\x00\x00\xBA\x71\x00\x00\xBB\x71\x00\x00\xBC\xD1"
-     "\x00\x00"
-     "\xBD\xD1\x00\x00\xBE\xD1\x00\x00\xBF\xD1\x00\x00\xB0\xB1\x01\x00\xB1\xB9"
-     "\x01\x00"
-     "\xB2\x81\x01\x00\xB3\x81\x01\x00\xB4\x51\x00\x00\xB5\x51\x00\x00\xB6\x51"
-     "\x00\x00"
-     "\xB7\x51\x00\x00\x52\x93\x00\x80\x56\x93\x00\x80\x5A\x93\x00\x80\x5E\x93"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\x62\x93\x00\x80\x66\x93"
-     "\x00\x80"
-     "\x6A\x93\x00\x80\x84\xA4\x02\x00\xBE\x24\x01\x00\x86\x90\x01\x00\x87\x14"
-     "\x01\x00"
-     "\x72\x93\x00\x80\x76\x93\x00\x80\xB3\x4D\x01\x00\x6E\x93\x00\x80\x7A\x93"
-     "\x00\x80"
-     "\x7E\x93\x00\x80\x82\x93\x00\x80\xB6\x11\x01\x00\xB5\x61\x01\x00\x86\x93"
-     "\x00\x80"
-     "\xBB\x3D\x01\x00\xBA\x35\x01\x00\x8A\x93\x00\x80\x8E\x93\x00\x80\xBF\xD5"
-     "\x00\x00"
-     "\xBE\xD5\x00\x00\xBD\x15\x01\x00\xBC\x15\x01\x00\x92\x93\x00\x80\xA3\x89"
-     "\x02\x00"
-     "\x96\x93\x00\x80\xBE\x40\x03\x00\xA6\xD5\x02\x00\x9A\x93\x00\x80\x9E\x93"
-     "\x00\x80"
-     "\xA5\xA5\x02\x00\xAA\xF1\x02\x00\xAB\xF9\x02\x00\xA2\x93\x00\x80\xA6\x93"
-     "\x00\x80"
-     "\xAE\x11\x03\x00\xAF\x11\x03\x00\xAC\xD1\x02\x00\xAD\xD1\x02\x00\xE1\xE8"
-     "\x01\x00"
-     "\xAA\x93\x00\x80\xE3\x14\x02\x00\x84\xD4\x03\x00\xAE\x93\x00\x80\xB2\x93"
-     "\x00\x80"
-     "\xB6\x93\x00\x80\x84\x10\x04\x00\xBA\x93\x00\x80\xBE\x93\x00\x80\xC2\x93"
-     "\x00\x80"
-     "\xC6\x93\x00\x80\xCA\x93\x00\x80\xCE\x93\x00\x80\xEF\x8C\x02\x00\xD2\x93"
-     "\x00\x80"
-     "\xA8\x39\x03\x00\xA9\x39\x03\x00\xAA\x41\x03\x00\xAB\x41\x03\x00\xAC\x41"
-     "\x03\x00"
-     "\xAD\x49\x03\x00\xAE\x71\x03\x00\xAF\x71\x03\x00\x80\xDD\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\xD6\x93\x00\x80\xDA\x93\x00\x80\xE2\x93\x00\x80\x87\x08"
-     "\x03\x00"
-     "\x86\x1C\x04\x00\xB8\x95\x00\x00\xB9\x9D\x00\x00\xBA\xA9\x00\x00\xBB\xA9"
-     "\x00\x00"
-     "\xBC\xB9\x00\x00\xBD\xB9\x00\x00\xBE\xA9\x00\x00\xBF\xA9\x00\x00\xB0\xE1"
-     "\x00\x00"
-     "\xB1\xED\x00\x00\xB2\xFD\x00\x00\xB3\xF5\x00\x00\xB4\xD9\x00\x00\xB5\xD9"
-     "\x00\x00"
-     "\xB6\xBD\x00\x00\xB7\xAD\x00\x00\xEF\x58\x0F\x00\xE6\x93\x00\x80\xEA\x93"
-     "\x00\x80"
-     "\xBE\x44\x05\x00\xEE\x93\x00\x80\xF2\x93\x00\x80\xF6\x93\x00\x80\xFA\x93"
-     "\x00\x80"
-     "\xFE\x93\x00\x80\x02\x94\x00\x80\x06\x94\x00\x80\x0A\x94\x00\x80\xE3\x00"
-     "\x0E\x00"
-     "\x0E\x94\x00\x80\xE1\x90\x0C\x00\x12\x94\x00\x80\xB6\x9D\x02\x00\x16\x94"
-     "\x00\x80"
-     "\x1A\x94\x00\x80\xB5\x9D\x02\x00\x1E\x94\x00\x80\xB3\x9D\x02\x00\x22\x94"
-     "\x00\x80"
-     "\x26\x94\x00\x80\xBE\x49\x01\x00\xBF\x49\x01\x00\xBC\x7D\x01\x00\xBD\x75"
-     "\x01\x00"
-     "\xBA\x71\x01\x00\xBB\x7D\x01\x00\x2A\x94\x00\x80\x2E\x94\x00\x80\x32\x94"
-     "\x00\x80"
-     "\xE1\xA4\x0E\x00\x36\x94\x00\x80\xE3\x50\x0E\x00\xBE\xDC\x04\x00\xE3\x2C"
-     "\x00\x00"
-     "\x3A\x94\x00\x80\xE1\x18\x01\x00\xBE\xEC\x06\x00\xEF\xF0\x00\x00\xBE\xE0"
-     "\x19\x00"
-     "\x3E\x94\x00\x80\x80\x75\x00\x00\x81\x7D\x00\x00\x82\x75\x00\x00\xEF\x50"
-     "\x01\x00"
-     "\xA3\x09\x02\x00\x42\x94\x00\x80\x46\x94\x00\x80\x4A\x94\x00\x80\x52\x94"
-     "\x00\x80"
-     "\xA6\x09\x02\x00\xA5\x09\x02\x00\x56\x94\x00\x80\xAB\xE9\x01\x00\xAA\xE5"
-     "\x01\x00"
-     "\x86\x48\x04\x00\x87\xDC\x04\x00\xAF\xDD\x01\x00\xAE\xDD\x01\x00\xAD\xE1"
-     "\x01\x00"
-     "\xAC\xE9\x01\x00\xA8\xB5\x06\x00\xA9\x01\x07\x00\xAA\x01\x07\x00\xAB\x1D"
-     "\x07\x00"
-     "\xAC\x05\x07\x00\xAD\x0D\x07\x00\xAE\x05\x07\x00\xAF\x3D\x07\x00\x4E\x94"
-     "\x00\x80"
-     "\x5A\x94\x00\x80\x5E\x94\x00\x80\x62\x94\x00\x80\x66\x94\x00\x80\x6A\x94"
-     "\x00\x80"
-     "\x6E\x94\x00\x80\x72\x94\x00\x80\xB8\x19\x07\x00\xB9\x19\x07\x00\xBA\x31"
-     "\x07\x00"
-     "\xBB\x31\x07\x00\xBC\xD1\x07\x00\xBD\xDD\x07\x00\xBE\xC1\x07\x00\xBF\xF1"
-     "\x07\x00"
-     "\xB0\x45\x07\x00\xB1\x4D\x07\x00\xB2\x45\x07\x00\xB3\x59\x07\x00\xB4\x4D"
-     "\x07\x00"
-     "\xB5\x31\x07\x00\xB6\x31\x07\x00\xB7\x2D\x07\x00\xB3\x89\x07\x00\x76\x94"
-     "\x00\x80"
-     "\x7A\x94\x00\x80\xDE\x93\x00\x80\x7E\x94\x00\x80\xB6\x8D\x07\x00\xB5\x8D"
-     "\x07\x00"
-     "\x82\x94\x00\x80\xBB\x89\x07\x00\xBA\x81\x07\x00\x86\x94\x00\x80\x8A\x94"
-     "\x00\x80"
-     "\xBF\xF5\x07\x00\xBE\x89\x07\x00\xBD\x8D\x07\x00\xBC\x91\x07\x00\x8E\x94"
-     "\x00\x80"
-     "\xA3\xCD\x07\x00\x92\x94\x00\x80\x96\x94\x00\x80\xA6\xC9\x07\x00\x9A\x94"
-     "\x00\x80"
-     "\x9E\x94\x00\x80\xA5\xC9\x07\x00\xAA\xC5\x07\x00\xAB\xCD\x07\x00\xA2\x94"
-     "\x00\x80"
-     "\xA6\x94\x00\x80\xAE\xCD\x07\x00\xAF\xB1\x07\x00\xAC\xD5\x07\x00\xAD\xC9"
-     "\x07\x00"
-     "\xA8\xD1\x06\x00\xA9\xD9\x06\x00\xAA\x2D\x01\x00\xAB\x25\x01\x00\xAC\x3D"
-     "\x01\x00"
-     "\xAD\x25\x01\x00\xAE\x2D\x01\x00\xAF\x25\x01\x00\xAA\x94\x00\x80\xAE\x94"
-     "\x00\x80"
-     "\xB2\x94\x00\x80\xB6\x94\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xBA\x94\x00\x80\xB8\x89\x01\x00\xB9\x89\x01\x00\xBA\x9D\x01\x00\xBB\x95"
-     "\x01\x00"
-     "\xBC\xB9\x01\x00\xBD\xB9\x01\x00\xBE\x79\x00\x00\xBF\x79\x00\x00\xB0\x5D"
-     "\x01\x00"
-     "\xB1\xE1\x01\x00\xB2\xE1\x01\x00\xB3\xFD\x01\x00\xB4\xED\x01\x00\xB5\xDD"
-     "\x01\x00"
-     "\xB6\xD5\x01\x00\xB7\xBD\x01\x00\x84\xC8\x02\x00\xB3\xA1\x01\x00\xBE\x94"
-     "\x00\x80"
-     "\xC2\x94\x00\x80\xB6\xA1\x01\x00\xC6\x94\x00\x80\xCA\x94\x00\x80\xB5\xB1"
-     "\x01\x00"
-     "\xBA\x9D\x01\x00\xBB\x21\x01\x00\x86\xA8\x01\x00\x87\xE0\x01\x00\xBE\x2D"
-     "\x01\x00"
-     "\xBF\x1D\x01\x00\xBC\x31\x01\x00\xBD\x2D\x01\x00\xA3\xE5\x01\x00\xCE\x94"
-     "\x00\x80"
-     "\xD2\x94\x00\x80\xD6\x94\x00\x80\xDA\x94\x00\x80\xA6\xE5\x01\x00\xA5\xF5"
-     "\x01\x00"
-     "\xDE\x94\x00\x80\xAB\x65\x01\x00\xAA\xD9\x01\x00\xE2\x94\x00\x80\xE6\x94"
-     "\x00\x80"
-     "\xAF\x59\x01\x00\xAE\x69\x01\x00\xAD\x69\x01\x00\xAC\x75\x01\x00\xEA\x94"
-     "\x00\x80"
-     "\xE3\x74\x02\x00\xEE\x94\x00\x80\xE1\xBC\x01\x00\xF2\x94\x00\x80\xF6\x94"
-     "\x00\x80"
-     "\xFA\x94\x00\x80\xFE\x94\x00\x80\x02\x95\x00\x80\x06\x95\x00\x80\x0A\x95"
-     "\x00\x80"
-     "\x0E\x95\x00\x80\x12\x95\x00\x80\xEF\x6C\x07\x00\x16\x95\x00\x80\x1A\x95"
-     "\x00\x80"
-     "\x1E\x95\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\x22\x95"
-     "\x00\x80"
-     "\x26\x95\x00\x80\x2A\x95\x00\x80\x32\x95\x00\x80\x84\xE4\x03\x00\x36\x95"
-     "\x00\x80"
-     "\x87\x2C\x03\x00\x86\xCC\x1C\x00\x3A\x95\x00\x80\x3E\x95\x00\x80\x42\x95"
-     "\x00\x80"
-     "\x46\x95\x00\x80\x4A\x95\x00\x80\x4E\x95\x00\x80\x52\x95\x00\x80\x56\x95"
-     "\x00\x80"
-     "\x5A\x95\x00\x80\x5E\x95\x00\x80\x62\x95\x00\x80\x66\x95\x00\x80\xA8\x2D"
-     "\x03\x00"
-     "\xA9\x31\x03\x00\xAA\x31\x03\x00\xAB\x69\x03\x00\xAC\x79\x03\x00\xAD\x79"
-     "\x03\x00"
-     "\xAE\x69\x03\x00\xAF\x61\x03\x00\xB0\x29\x03\x00\xB1\x29\x03\x00\xB2\xC9"
-     "\x00\x00"
-     "\xB3\xC5\x00\x00\xB4\xC1\x00\x00\xB5\xC1\x00\x00\xB6\xC1\x00\x00\xB7\xC1"
-     "\x00\x00"
-     "\xB8\xC1\x00\x00\xB9\xC1\x00\x00\xBA\xC1\x00\x00\xBB\xC1\x00\x00\xBC\xC1"
-     "\x00\x00"
-     "\xBD\xCD\x00\x00\xBE\xC5\x00\x00\xBF\xBD\x00\x00\xE3\x4C\x07\x00\x6A\x95"
-     "\x00\x80"
-     "\xE1\xAC\x07\x00\x6E\x95\x00\x80\x72\x95\x00\x80\x76\x95\x00\x80\x7A\x95"
-     "\x00\x80"
-     "\x7E\x95\x00\x80\x82\x95\x00\x80\x86\x95\x00\x80\x8A\x95\x00\x80\x8E\x95"
-     "\x00\x80"
-     "\xEF\xC4\x06\x00\x92\x95\x00\x80\x96\x95\x00\x80\x9A\x95\x00\x80\x81\x79"
-     "\x00\x00"
-     "\x80\x71\x00\x00\xB3\xD1\x02\x00\x82\x71\x00\x00\xB5\x99\x02\x00\xA2\x95"
-     "\x00\x80"
-     "\x84\x14\x1D\x00\xB6\x91\x02\x00\x86\xC0\x1C\x00\x87\xE8\x1D\x00\xBB\x85"
-     "\x02\x00"
-     "\xBA\x85\x02\x00\xBD\x79\x01\x00\xBC\x75\x01\x00\xBF\x6D\x01\x00\xBE\x6D"
-     "\x01\x00"
-     "\xA6\x95\x00\x80\xAA\x95\x00\x80\xAE\x95\x00\x80\xB2\x95\x00\x80\x84\xDC"
-     "\x1C\x00"
-     "\xB6\x95\x00\x80\xBA\x95\x00\x80\xEF\xF4\x06\x00\xBE\xE8\x1C\x00\xE1\x54"
-     "\x06\x00"
-     "\xBE\x95\x00\x80\xE3\x3C\x00\x00\x84\xF0\x1F\x00\xC2\x95\x00\x80\xC6\x95"
-     "\x00\x80"
-     "\xCA\x95\x00\x80\xA3\x59\x02\x00\xCE\x95\x00\x80\xD2\x95\x00\x80\xD6\x95"
-     "\x00\x80"
-     "\xDA\x95\x00\x80\xA6\x19\x02\x00\xA5\x11\x02\x00\xDE\x95\x00\x80\xAB\x0D"
-     "\x02\x00"
-     "\xAA\x0D\x02\x00\xE2\x95\x00\x80\xE6\x95\x00\x80\xAF\xE5\x01\x00\xAE\xE5"
-     "\x01\x00"
-     "\xAD\xF1\x01\x00\xAC\xFD\x01\x00\xA8\x5D\x1E\x00\xA9\x61\x1E\x00\xAA\x79"
-     "\x1E\x00"
-     "\xAB\x79\x1E\x00\xAC\x69\x1E\x00\xAD\x69\x1E\x00\xAE\x99\x1E\x00\xAF\x99"
-     "\x1E\x00"
-     "\x9E\x95\x00\x80\xEA\x95\x00\x80\xEE\x95\x00\x80\xF2\x95\x00\x80\xF6\x95"
-     "\x00\x80"
-     "\xFA\x95\x00\x80\xFE\x95\x00\x80\x02\x96\x00\x80\xB8\x51\x1F\x00\xB9\x51"
-     "\x1F\x00"
-     "\xBA\x51\x1F\x00\xBB\x51\x1F\x00\xBC\x71\x1F\x00\xBD\x71\x1F\x00\xBE\x71"
-     "\x1F\x00"
-     "\xBF\x71\x1F\x00\xB0\xE9\x1E\x00\xB1\xE9\x1E\x00\xB2\xF9\x1E\x00\xB3\xF9"
-     "\x1E\x00"
-     "\xB4\xE9\x1E\x00\xB5\xE9\x1E\x00\xB6\x7D\x1F\x00\xB7\x71\x1F\x00\xB3\xF1"
-     "\x1E\x00"
-     "\x82\x2D\x00\x00\x81\x15\x00\x00\x80\xB1\x00\x00\x2E\x95\x00\x80\xB6\xE9"
-     "\x1E\x00"
-     "\xB5\xE1\x1E\x00\x06\x96\x00\x80\xBB\x2D\x1E\x00\xBA\x2D\x1E\x00\x0A\x96"
-     "\x00\x80"
-     "\x0E\x96\x00\x80\xBF\x2D\x1E\x00\xBE\x2D\x1E\x00\xBD\x2D\x1E\x00\xBC\x2D"
-     "\x1E\x00"
-     "\x12\x96\x00\x80\xA3\xB5\x1E\x00\x86\x68\x00\x00\x87\x4C\x00\x00\xA6\xAD"
-     "\x1E\x00"
-     "\x16\x96\x00\x80\x1A\x96\x00\x80\xA5\xA5\x1E\x00\xAA\x69\x1E\x00\xAB\x69"
-     "\x1E\x00"
-     "\x1E\x96\x00\x80\x22\x96\x00\x80\xAE\x69\x1E\x00\xAF\x69\x1E\x00\xAC\x69"
-     "\x1E\x00"
-     "\xAD\x69\x1E\x00\xA8\x55\x1E\x00\xA9\x85\x1E\x00\xAA\x8D\x1E\x00\xAB\x85"
-     "\x1E\x00"
-     "\xAC\x9D\x1E\x00\xAD\x81\x1E\x00\xAE\x81\x1E\x00\xAF\xB5\x1E\x00\x26\x96"
-     "\x00\x80"
-     "\x2A\x96\x00\x80\x2E\x96\x00\x80\x32\x96\x00\x80\x36\x96\x00\x80\x3A\x96"
-     "\x00\x80"
-     "\x3E\x96\x00\x80\x42\x96\x00\x80\xB8\xA9\x1E\x00\xB9\xA9\x1E\x00\xBA\x6D"
-     "\x01\x00"
-     "\xBB\x65\x01\x00\xBC\x7D\x01\x00\xBD\x65\x01\x00\xBE\x6D\x01\x00\xBF\x65"
-     "\x01\x00"
-     "\xB0\xD5\x1E\x00\xB1\xDD\x1E\x00\xB2\xD5\x1E\x00\xB3\xA9\x1E\x00\xB4\xB9"
-     "\x1E\x00"
-     "\xB5\xB9\x1E\x00\xB6\xA9\x1E\x00\xB7\xA1\x1E\x00\xB3\x05\x1E\x00\x46\x96"
-     "\x00\x80"
-     "\x4A\x96\x00\x80\x4E\x96\x00\x80\x52\x96\x00\x80\xB6\x01\x1E\x00\xB5\x15"
-     "\x1E\x00"
-     "\x56\x96\x00\x80\xBB\x19\x1E\x00\xBA\x11\x1E\x00\x5A\x96\x00\x80\x5E\x96"
-     "\x00\x80"
-     "\xBF\x7D\x1E\x00\xBE\x7D\x1E\x00\xBD\x7D\x1E\x00\xBC\x01\x1E\x00\x62\x96"
-     "\x00\x80"
-     "\xA3\x41\x1E\x00\x66\x96\x00\x80\x6A\x96\x00\x80\xA6\x45\x1E\x00\x6E\x96"
-     "\x00\x80"
-     "\x72\x96\x00\x80\xA5\x51\x1E\x00\xAA\x55\x1E\x00\xAB\x5D\x1E\x00\x76\x96"
-     "\x00\x80"
-     "\x7A\x96\x00\x80\xAE\x39\x1E\x00\xAF\x39\x1E\x00\xAC\x45\x1E\x00\xAD\x39"
-     "\x1E\x00"
-     "\xA8\xC1\x02\x00\xA9\xD1\x02\x00\xAA\xD1\x02\x00\xAB\xE5\x02\x00\xAC\xFD"
-     "\x02\x00"
-     "\xAD\x35\x03\x00\xAE\x3D\x03\x00\xAF\x35\x03\x00\x82\x96\x00\x80\xBE\x68"
-     "\x02\x00"
-     "\x86\x96\x00\x80\x8A\x96\x00\x80\x8E\x96\x00\x80\x92\x96\x00\x80\x96\x96"
-     "\x00\x80"
-     "\x9A\x96\x00\x80\xB8\xD1\x03\x00\xB9\xD9\x03\x00\xBA\xE1\x03\x00\xBB\xE1"
-     "\x03\x00"
-     "\xBC\x91\x03\x00\xBD\x9D\x03\x00\xBE\x95\x03\x00\xBF\x89\x03\x00\xB0\x4D"
-     "\x03\x00"
-     "\xB1\x55\x03\x00\xB2\x5D\x03\x00\xB3\x55\x03\x00\xB4\x4D\x03\x00\xB5\xF1"
-     "\x03\x00"
-     "\xB6\xF1\x03\x00\xB7\xF1\x03\x00\x80\x59\x03\x00\x81\x6D\x03\x00\x82\x65"
-     "\x03\x00"
-     "\xB3\x55\x02\x00\x9E\x96\x00\x80\xB5\x79\x02\x00\xB6\xB9\x03\x00\xA2\x96"
-     "\x00\x80"
-     "\x86\x00\x04\x00\x87\xE4\x03\x00\xBA\x85\x03\x00\xBB\x9D\x03\x00\xBC\x85"
-     "\x03\x00"
-     "\xBD\x8D\x03\x00\xBE\x85\x03\x00\xBF\x39\x03\x00\xA3\x19\x02\x00\xA6\x96"
-     "\x00\x80"
-     "\xAA\x96\x00\x80\xAE\x96\x00\x80\x85\xAC\x0B\x00\xA6\xF5\x03\x00\xA5\x35"
-     "\x02\x00"
-     "\xB2\x96\x00\x80\xAB\xD1\x03\x00\xAA\xC9\x03\x00\xB6\x96\x00\x80\xBA\x96"
-     "\x00\x80"
-     "\xAF\x75\x03\x00\xAE\xC9\x03\x00\xAD\xC1\x03\x00\xAC\xC9\x03\x00\xBE\x96"
-     "\x00\x80"
-     "\xC2\x96\x00\x80\xC6\x96\x00\x80\xCA\x96\x00\x80\xEF\xCC\x03\x00\xE1\x60"
-     "\x1F\x00"
-     "\xCE\x96\x00\x80\xE3\x54\x1E\x00\xE3\x80\x1E\x00\xD2\x96\x00\x80\xE1\xF8"
-     "\x01\x00"
-     "\xEF\x7C\x1A\x00\xD6\x96\x00\x80\xDA\x96\x00\x80\xDE\x96\x00\x80\xE2\x96"
-     "\x00\x80"
-     "\x80\x7D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\x84\x40\x02\x00\xE6\x96"
-     "\x00\x80"
-     "\xEA\x96\x00\x80\xBE\x54\x05\x00\xF2\x96\x00\x80\x86\xC0\x04\x00\x87\x48"
-     "\x05\x00"
-     "\xF6\x96\x00\x80\xFA\x96\x00\x80\xFE\x96\x00\x80\x02\x97\x00\x80\x06\x97"
-     "\x00\x80"
-     "\x0A\x97\x00\x80\x0E\x97\x00\x80\x12\x97\x00\x80\x16\x97\x00\x80\x1A\x97"
-     "\x00\x80"
-     "\x1E\x97\x00\x80\x22\x97\x00\x80\x26\x97\x00\x80\xEF\xC4\x1E\x00\x84\xA4"
-     "\x05\x00"
-     "\xE1\xDC\x1E\x00\x2A\x97\x00\x80\xE3\x2C\x1E\x00\x2E\x97\x00\x80\x32\x97"
-     "\x00\x80"
-     "\x36\x97\x00\x80\x3A\x97\x00\x80\xB3\xED\x01\x00\x3E\x97\x00\x80\x42\x97"
-     "\x00\x80"
-     "\x46\x97\x00\x80\x4A\x97\x00\x80\xB6\xED\x01\x00\xB5\xED\x01\x00\x4E\x97"
-     "\x00\x80"
-     "\xBB\x25\x01\x00\xBA\x79\x01\x00\x52\x97\x00\x80\x56\x97\x00\x80\xBF\x19"
-     "\x01\x00"
-     "\xBE\x29\x01\x00\xBD\x35\x01\x00\xBC\x35\x01\x00\xA8\x21\x06\x00\xA9\x2D"
-     "\x06\x00"
-     "\xAA\x25\x06\x00\xAB\x3D\x06\x00\xAC\x25\x06\x00\xAD\x2D\x06\x00\xAE\x25"
-     "\x06\x00"
-     "\xAF\x8D\x06\x00\xEE\x96\x00\x80\x82\x15\x00\x00\x81\xC5\x07\x00\x80\xC5"
-     "\x07\x00"
-     "\x5A\x97\x00\x80\xBE\x38\x03\x00\x5E\x97\x00\x80\x62\x97\x00\x80\xB8\x45"
-     "\x07\x00"
-     "\xB9\x4D\x07\x00\xBA\x45\x07\x00\xBB\x55\x07\x00\xBC\x75\x07\x00\xBD\x65"
-     "\x07\x00"
-     "\xBE\x6D\x07\x00\xBF\x65\x07\x00\xB0\xFD\x06\x00\xB1\xA1\x06\x00\xB2\xA1"
-     "\x06\x00"
-     "\xB3\xA1\x06\x00\xB4\xAD\x06\x00\xB5\x9D\x06\x00\xB6\x95\x06\x00\xB7\x7D"
-     "\x07\x00"
-     "\xA3\xAD\x06\x00\x66\x97\x00\x80\x86\x28\x00\x00\x87\x4C\x03\x00\x6A\x97"
-     "\x00\x80"
-     "\xA6\xAD\x06\x00\xA5\xAD\x06\x00\x6E\x97\x00\x80\xAB\x65\x06\x00\xAA\x39"
-     "\x06\x00"
-     "\x72\x97\x00\x80\x76\x97\x00\x80\xAF\x59\x06\x00\xAE\x69\x06\x00\xAD\x75"
-     "\x06\x00"
-     "\xAC\x75\x06\x00\x7A\x97\x00\x80\x7E\x97\x00\x80\xB3\xA5\x07\x00\x82\x97"
-     "\x00\x80"
-     "\xB5\xA5\x07\x00\x86\x97\x00\x80\x8A\x97\x00\x80\xB6\xA5\x07\x00\x8E\x97"
-     "\x00\x80"
-     "\x7E\x96\x00\x80\xBB\x61\x07\x00\xBA\x69\x07\x00\xBD\x61\x07\x00\xBC\x79"
-     "\x07\x00"
-     "\xBF\x61\x07\x00\xBE\x69\x07\x00\xA8\x5D\x06\x00\xA9\x6D\x06\x00\xAA\x65"
-     "\x06\x00"
-     "\xAB\x7D\x06\x00\xAC\x65\x06\x00\xAD\x6D\x06\x00\xAE\x65\x06\x00\xAF\x9D"
-     "\x06\x00"
-     "\x92\x97\x00\x80\x96\x97\x00\x80\x9A\x97\x00\x80\x9E\x97\x00\x80\xA2\x97"
-     "\x00\x80"
-     "\xA6\x97\x00\x80\xAA\x97\x00\x80\xAE\x97\x00\x80\xB8\xF5\x06\x00\xB9\xFD"
-     "\x06\x00"
-     "\xBA\xF5\x06\x00\xBB\x89\x06\x00\xBC\x99\x06\x00\xBD\x99\x06\x00\xBE\x89"
-     "\x06\x00"
-     "\xBF\x81\x06\x00\xB0\xE5\x06\x00\xB1\xED\x06\x00\xB2\xE5\x06\x00\xB3\xFD"
-     "\x06\x00"
-     "\xB4\xE5\x06\x00\xB5\xED\x06\x00\xB6\xD1\x06\x00\xB7\xCD\x06\x00\xA3\xED"
-     "\x06\x00"
-     "\xB2\x97\x00\x80\xB6\x97\x00\x80\xBA\x97\x00\x80\xBE\x97\x00\x80\xA6\xED"
-     "\x06\x00"
-     "\xA5\xED\x06\x00\xC2\x97\x00\x80\xAB\x29\x06\x00\xAA\x21\x06\x00\xC6\x97"
-     "\x00\x80"
-     "\xCA\x97\x00\x80\xAF\x29\x06\x00\xAE\x21\x06\x00\xAD\x29\x06\x00\xAC\x31"
-     "\x06\x00"
-     "\x80\x5D\x00\x00\x81\x6D\x00\x00\x82\x65\x00\x00\x84\x80\x01\x00\xBE\x98"
-     "\x01\x00"
-     "\xD2\x97\x00\x80\x87\xD8\x00\x00\x86\x7C\x0F\x00\xD6\x97\x00\x80\xDA\x97"
-     "\x00\x80"
-     "\xDE\x97\x00\x80\xE2\x97\x00\x80\xE6\x97\x00\x80\xEA\x97\x00\x80\xEE\x97"
-     "\x00\x80"
-     "\xF2\x97\x00\x80\xA8\x3D\x02\x00\xA9\x81\x02\x00\xAA\x99\x02\x00\xAB\x91"
-     "\x02\x00"
-     "\xAC\xB9\x02\x00\xAD\xB9\x02\x00\xAE\xD9\x02\x00\xAF\xD9\x02\x00\xF6\x97"
-     "\x00\x80"
-     "\xFA\x97\x00\x80\xFE\x97\x00\x80\xBF\xA4\x0E\x00\x02\x98\x00\x80\x06\x98"
-     "\x00\x80"
-     "\x0A\x98\x00\x80\x0E\x98\x00\x80\xB8\x71\x03\x00\xB9\x71\x03\x00\xBA\x71"
-     "\x03\x00"
-     "\xBB\x71\x03\x00\xBC\xD5\x03\x00\xBD\xDD\x03\x00\xBE\xD5\x03\x00\xBF\xCD"
-     "\x03\x00"
-     "\xB0\xAD\x02\x00\xB1\xB1\x02\x00\xB2\xB1\x02\x00\xB3\x8D\x02\x00\xB4\x51"
-     "\x03\x00"
-     "\xB5\x51\x03\x00\xB6\x51\x03\x00\xB7\x51\x03\x00\xB3\x11\x02\x00\x12\x98"
-     "\x00\x80"
-     "\x16\x98\x00\x80\x1A\x98\x00\x80\x84\x50\x03\x00\xB6\x39\x02\x00\xB5\x31"
-     "\x02\x00"
-     "\x1E\x98\x00\x80\xBB\x75\x02\x00\xBA\x69\x02\x00\x22\x98\x00\x80\x26\x98"
-     "\x00\x80"
-     "\xBF\x95\x03\x00\xBE\x95\x03\x00\xBD\x59\x02\x00\xBC\x65\x02\x00\x2A\x98"
-     "\x00\x80"
-     "\xA3\x55\x02\x00\x2E\x98\x00\x80\xBE\x80\x03\x00\xA6\x7D\x02\x00\x32\x98"
-     "\x00\x80"
-     "\x36\x98\x00\x80\xA5\x75\x02\x00\xAA\x2D\x02\x00\xAB\x31\x02\x00\x3A\x98"
-     "\x00\x80"
-     "\x42\x98\x00\x80\xAE\xD1\x03\x00\xAF\xD1\x03\x00\xAC\x21\x02\x00\xAD\x1D"
-     "\x02\x00"
-     "\x80\xD5\x01\x00\x81\xDD\x01\x00\x82\xE1\x01\x00\xB3\xE9\x02\x00\x46\x98"
-     "\x00\x80"
-     "\xB5\xF9\x02\x00\xB6\xE9\x02\x00\x4A\x98\x00\x80\x4E\x98\x00\x80\x52\x98"
-     "\x00\x80"
-     "\xBA\xD5\x02\x00\xBB\x0D\x01\x00\xBC\x15\x01\x00\xBD\x1D\x01\x00\xBE\x09"
-     "\x01\x00"
-     "\xBF\x09\x01\x00\x86\x20\x0D\x00\x87\x88\x02\x00\x56\x98\x00\x80\x5A\x98"
-     "\x00\x80"
-     "\x5E\x98\x00\x80\x62\x98\x00\x80\xEF\xF8\x06\x00\x66\x98\x00\x80\xE1\xD8"
-     "\x07\x00"
-     "\x6A\x98\x00\x80\xE3\x34\x06\x00\x6E\x98\x00\x80\x72\x98\x00\x80\x76\x98"
-     "\x00\x80"
-     "\x7A\x98\x00\x80\x7E\x98\x00\x80\x82\x98\x00\x80\xEF\xAC\x00\x00\x86\x98"
-     "\x00\x80"
-     "\x8A\x98\x00\x80\x8E\x98\x00\x80\x92\x98\x00\x80\x96\x98\x00\x80\xEF\x50"
-     "\x01\x00"
-     "\x84\x6C\x0C\x00\xE1\xDC\x06\x00\x9A\x98\x00\x80\xE3\x1C\x06\x00\x9E\x98"
-     "\x00\x80"
-     "\xE3\x68\x00\x00\xA2\x98\x00\x80\xE1\x2C\x01\x00\xA3\x25\x02\x00\x82\x2D"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x80\x75\x00\x00\xA6\x98\x00\x80\xA6\x25\x02\x00\xA5\x35"
-     "\x02\x00"
-     "\x85\x00\x08\x00\xAB\xC1\x01\x00\xAA\x19\x02\x00\xAA\x98\x00\x80\xBE\x24"
-     "\x0C\x00"
-     "\xAF\xC5\x01\x00\xAE\xC5\x01\x00\xAD\xD1\x01\x00\xAC\xD9\x01\x00\xA8\xD5"
-     "\x0D\x00"
-     "\xA9\xE5\x0D\x00\xAA\xED\x0D\x00\xAB\xE5\x0D\x00\xAC\xFD\x0D\x00\xAD\xE5"
-     "\x0D\x00"
-     "\xAE\xED\x0D\x00\xAF\xE5\x0D\x00\x3E\x98\x00\x80\xAE\x98\x00\x80\x86\x08"
-     "\x03\x00"
-     "\x87\x8C\x03\x00\xB2\x98\x00\x80\xB6\x98\x00\x80\xBA\x98\x00\x80\xBE\x98"
-     "\x00\x80"
-     "\xB8\xC5\x0E\x00\xB9\xCD\x0E\x00\xBA\xC5\x0E\x00\xBB\xD1\x0E\x00\xBC\xF9"
-     "\x0E\x00"
-     "\xBD\xF9\x0E\x00\xBE\x99\x0E\x00\xBF\x95\x0E\x00\xB0\x9D\x0D\x00\xB1\x45"
-     "\x0E\x00"
-     "\xB2\x4D\x0E\x00\xB3\x45\x0E\x00\xB4\x41\x0E\x00\xB5\x41\x0E\x00\xB6\x41"
-     "\x0E\x00"
-     "\xB7\x41\x0E\x00\xB8\x01\x0F\x00\xB9\x0D\x0F\x00\xBA\x05\x0F\x00\xBB\x1D"
-     "\x0F\x00"
-     "\xBC\x05\x0F\x00\xBD\x0D\x0F\x00\xBE\x31\x0F\x00\xBF\x2D\x0F\x00\xB0\xF5"
-     "\x0F\x00"
-     "\xB1\xF9\x0F\x00\xB2\xCD\x0F\x00\xB3\xC1\x0F\x00\xB4\xC1\x0F\x00\xB5\xC1"
-     "\x0F\x00"
-     "\xB6\xC1\x0F\x00\xB7\xC1\x0F\x00\xA8\x2D\x0E\x00\xA9\xD5\x0F\x00\xAA\xD5"
-     "\x0F\x00"
-     "\xAB\xED\x0F\x00\xAC\xF5\x0F\x00\xAD\x9D\x0F\x00\xAE\x95\x0F\x00\xAF\x8D"
-     "\x0F\x00"
-     "\xC2\x98\x00\x80\xC6\x98\x00\x80\xCE\x97\x00\x80\xCA\x98\x00\x80\xCE\x98"
-     "\x00\x80"
-     "\xD2\x98\x00\x80\xD6\x98\x00\x80\xDA\x98\x00\x80\xB3\x1D\x0E\x00\xDE\x98"
-     "\x00\x80"
-     "\xE2\x98\x00\x80\xE6\x98\x00\x80\xEA\x98\x00\x80\xB6\x5D\x0E\x00\xB5\x3D"
-     "\x0E\x00"
-     "\xEE\x98\x00\x80\xBB\x7D\x0E\x00\xBA\x71\x0E\x00\xF2\x98\x00\x80\xF6\x98"
-     "\x00\x80"
-     "\xBF\x45\x0E\x00\xBE\x45\x0E\x00\xBD\x55\x0E\x00\xBC\x59\x0E\x00\xFA\x98"
-     "\x00\x80"
-     "\xA3\x59\x0E\x00\xFE\x98\x00\x80\x02\x99\x00\x80\xA6\x19\x0E\x00\x06\x99"
-     "\x00\x80"
-     "\x0A\x99\x00\x80\xA5\x79\x0E\x00\xAA\x35\x0E\x00\xAB\x39\x0E\x00\x0E\x99"
-     "\x00\x80"
-     "\x12\x99\x00\x80\xAE\x01\x0E\x00\xAF\x01\x0E\x00\xAC\x1D\x0E\x00\xAD\x11"
-     "\x0E\x00"
-     "\xA8\xB1\x01\x00\xA9\xB1\x01\x00\xAA\xB1\x01\x00\xAB\xB1\x01\x00\xAC\x71"
-     "\x01\x00"
-     "\xAD\x71\x01\x00\xAE\x75\x01\x00\xAF\x6D\x01\x00\x80\xCD\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x16\x99\x00\x80\x1A\x99\x00\x80\x84\x78\x01\x00\xBE\x70"
-     "\x01\x00"
-     "\x22\x99\x00\x80\xB8\xE9\x00\x00\xB9\xE9\x00\x00\xBA\x89\x00\x00\xBB\x85"
-     "\x00\x00"
-     "\xBC\x9D\x00\x00\xBD\x81\x00\x00\xBE\x81\x00\x00\xBF\xB5\x00\x00\xB0\x15"
-     "\x01\x00"
-     "\xB1\x1D\x01\x00\xB2\x15\x01\x00\xB3\xE9\x00\x00\xB4\xF9\x00\x00\xB5\xF9"
-     "\x00\x00"
-     "\xB6\xE9\x00\x00\xB7\xE1\x00\x00\xB3\x39\x02\x00\x26\x99\x00\x80\x86\x88"
-     "\x04\x00"
-     "\x87\x78\x03\x00\x2A\x99\x00\x80\xB6\xF1\x02\x00\xB5\xF9\x02\x00\x2E\x99"
-     "\x00\x80"
-     "\xBB\xD9\x02\x00\xBA\xD1\x02\x00\x32\x99\x00\x80\x36\x99\x00\x80\xBF\x3D"
-     "\x03\x00"
-     "\xBE\x3D\x03\x00\xBD\x3D\x03\x00\xBC\xC1\x02\x00\x3A\x99\x00\x80\xA3\x7D"
-     "\x02\x00"
-     "\x3E\x99\x00\x80\x42\x99\x00\x80\xA6\xB5\x02\x00\x46\x99\x00\x80\x4A\x99"
-     "\x00\x80"
-     "\xA5\xBD\x02\x00\xAA\x95\x02\x00\xAB\x9D\x02\x00\x4E\x99\x00\x80\x52\x99"
-     "\x00\x80"
-     "\xAE\x79\x03\x00\xAF\x79\x03\x00\xAC\x85\x02\x00\xAD\x79\x03\x00\x82\xC1"
-     "\x1A\x00"
-     "\x83\xC9\x1A\x00\x56\x99\x00\x80\x5A\x99\x00\x80\x86\xB5\x1E\x00\x87\x29"
-     "\x12\x00"
-     "\x84\x31\x1E\x00\x85\xBD\x1E\x00\x8A\xA1\x12\x00\x8B\xBD\x12\x00\x5E\x99"
-     "\x00\x80"
-     "\x62\x99\x00\x80\x8E\x91\x16\x00\x8F\x3D\x17\x00\x8C\x7D\x13\x00\x8D\x31"
-     "\x16\x00"
-     "\x92\x51\x0A\x00\x93\xE5\x0A\x00\x66\x99\x00\x80\xEF\xDC\x6C\x00\x96\xF1"
-     "\x0E\x00"
-     "\x97\xED\x0E\x00\x94\x75\x0A\x00\x95\x39\x0E\x00\x9A\x09\x02\x00\x9B\xC5"
-     "\x02\x00"
-     "\x6A\x99\x00\x80\x6E\x99\x00\x80\x72\x99\x00\x80\xE1\x3C\x0F\x00\x9C\xDD"
-     "\x02\x00"
-     "\xE3\xA8\x76\x00\x9F\x21\x0F\x00\x9E\xE1\x0F\x00\x9D\xA5\x0B\x00\x9C\x2D"
-     "\x0B\x00"
-     "\x9B\x81\x0B\x00\x9A\x9D\x16\x00\x99\x39\x17\x00\x98\x91\x17\x00\x97\xE5"
-     "\x12\x00"
-     "\x96\x3D\x13\x00\x95\x29\x13\x00\x94\x91\x13\x00\x93\x15\x1F\x00\x92\x15"
-     "\x1F\x00"
-     "\x91\x91\x1F\x00\x90\x71\x1B\x00\x8F\x01\x1B\x00\x76\x99\x00\x80\x7A\x99"
-     "\x00\x80"
-     "\x7E\x99\x00\x80\x82\x99\x00\x80\xEF\x60\x0F\x00\x86\x99\x00\x80\x8A\x99"
-     "\x00\x80"
-     "\xBE\x28\x05\x00\xE3\xE4\x0F\x00\x84\x24\x05\x00\xE1\xA8\x0C\x00\x81\x29"
-     "\x00\x00"
-     "\x80\x25\x00\x00\x92\x99\x00\x80\x82\x21\x00\x00\x87\x0C\x05\x00\x86\xA4"
-     "\x04\x00"
-     "\x96\x99\x00\x80\x9A\x99\x00\x80\x9E\x99\x00\x80\xA2\x99\x00\x80\xEF\xD8"
-     "\x00\x00"
-     "\xEF\x88\x0E\x00\xE1\x90\x01\x00\xE1\xFC\x0E\x00\xE3\x0C\x00\x00\xE3\x1C"
-     "\x0E\x00"
-     "\xA6\x99\x00\x80\xAA\x99\x00\x80\xAE\x99\x00\x80\xB2\x99\x00\x80\xB6\x99"
-     "\x00\x80"
-     "\xBA\x99\x00\x80\x84\x04\x07\x00\xBE\x04\x07\x00\xBE\x99\x00\x80\xC2\x99"
-     "\x00\x80"
-     "\xC6\x99\x00\x80\xCA\x99\x00\x80\xCE\x99\x00\x80\xD2\x99\x00\x80\xD6\x99"
-     "\x00\x80"
-     "\xDA\x99\x00\x80\xA3\xD5\x03\x00\xA2\xD5\x03\x00\xA1\x11\x03\x00\xA0\xC5"
-     "\x0F\x00"
-     "\x1E\x99\x00\x80\x8E\x99\x00\x80\xDE\x99\x00\x80\xE2\x99\x00\x80\xE6\x99"
-     "\x00\x80"
-     "\xEA\x99\x00\x80\xEE\x99\x00\x80\xF2\x99\x00\x80\xA8\x9D\x05\x00\xA9\x21"
-     "\x06\x00"
-     "\xAA\x39\x06\x00\xAB\x31\x06\x00\xAC\x19\x06\x00\xAD\x19\x06\x00\xAE\x79"
-     "\x06\x00"
-     "\xAF\x75\x06\x00\xB0\x0D\x06\x00\xB1\x11\x06\x00\xB2\x11\x06\x00\xB3\x25"
-     "\x06\x00"
-     "\xB4\x3D\x06\x00\xB5\xDD\x06\x00\xB6\xD5\x06\x00\xB7\xF5\x06\x00\xB8\xCD"
-     "\x06\x00"
-     "\xB9\xAD\x06\x00\xBA\x81\x06\x00\xBB\x81\x06\x00\xBC\x81\x06\x00\xBD\x31"
-     "\x07\x00"
-     "\xBE\x31\x07\x00\xBF\x31\x07\x00\xB3\x91\x06\x00\x82\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x80\x91\x00\x00\xF6\x99\x00\x80\xB6\xB9\x06\x00\xB5\xB1\x06\x00\xFA\x99"
-     "\x00\x80"
-     "\xBB\xF5\x06\x00\xBA\xE9\x06\x00\xFE\x99\x00\x80\x02\x9A\x00\x80\xBF\x39"
-     "\x06\x00"
-     "\xBE\xD9\x06\x00\xBD\xD1\x06\x00\xBC\xE5\x06\x00\x06\x9A\x00\x80\xA3\xD5"
-     "\x06\x00"
-     "\x86\xE8\x00\x00\x87\x84\x00\x00\xA6\xFD\x06\x00\x0A\x9A\x00\x80\x0E\x9A"
-     "\x00\x80"
-     "\xA5\xF5\x06\x00\xAA\xAD\x06\x00\xAB\xB1\x06\x00\x12\x9A\x00\x80\x16\x9A"
-     "\x00\x80"
-     "\xAE\x9D\x06\x00\xAF\x7D\x06\x00\xAC\xA1\x06\x00\xAD\x95\x06\x00\x1A\x9A"
-     "\x00\x80"
-     "\x1E\x9A\x00\x80\xB5\x29\x69\x00\xB4\x01\x68\x00\xB3\x51\x6C\x00\xB2\x11"
-     "\x6D\x00"
-     "\xB1\x19\x6D\x00\xB0\x71\x70\x00\x22\x9A\x00\x80\x26\x9A\x00\x80\x2A\x9A"
-     "\x00\x80"
-     "\x2E\x9A\x00\x80\x32\x9A\x00\x80\x36\x9A\x00\x80\x3A\x9A\x00\x80\x3E\x9A"
-     "\x00\x80"
-     "\xA7\x01\x78\x00\xA6\xE9\x7D\x00\xA5\x69\x7D\x00\xA4\x01\x7C\x00\xA3\xF9"
-     "\x01\x00"
-     "\xA2\x51\x01\x00\xA1\x59\x01\x00\x42\x9A\x00\x80\xAF\x69\x71\x00\xAE\x61"
-     "\x71\x00"
-     "\xAD\x01\x70\x00\xAC\x49\x75\x00\xAB\x51\x75\x00\xAA\x01\x74\x00\xA9\xF5"
-     "\x79\x00"
-     "\xA8\x5D\x79\x00\xA8\xF9\x06\x00\xA9\x0D\x06\x00\xAA\x05\x06\x00\xAB\x1D"
-     "\x06\x00"
-     "\xAC\x05\x06\x00\xAD\x25\x06\x00\xAE\x2D\x06\x00\xAF\x25\x06\x00\x46\x9A"
-     "\x00\x80"
-     "\x4A\x9A\x00\x80\x4E\x9A\x00\x80\x52\x9A\x00\x80\x56\x9A\x00\x80\x5A\x9A"
-     "\x00\x80"
-     "\x5E\x9A\x00\x80\x62\x9A\x00\x80\xB8\x25\x06\x00\xB9\x2D\x06\x00\xBA\x25"
-     "\x06\x00"
-     "\xBB\xC5\x01\x00\xBC\xDD\x01\x00\xBD\xC5\x01\x00\xBE\xCD\x01\x00\xBF\xC5"
-     "\x01\x00"
-     "\xB0\x5D\x06\x00\xB1\x25\x06\x00\xB2\x2D\x06\x00\xB3\x3D\x06\x00\xB4\x25"
-     "\x06\x00"
-     "\xB5\x2D\x06\x00\xB6\x25\x06\x00\xB7\x1D\x06\x00\xB3\x8D\x05\x00\x66\x9A"
-     "\x00\x80"
-     "\x6A\x9A\x00\x80\xBE\x40\x03\x00\x72\x9A\x00\x80\xB6\xA5\x05\x00\xB5\x8D"
-     "\x05\x00"
-     "\x76\x9A\x00\x80\xBB\xE1\x05\x00\xBA\x99\x05\x00\x84\xA4\x03\x00\x7A\x9A"
-     "\x00\x80"
-     "\xBF\xE9\x05\x00\xBE\xE1\x05\x00\xBD\xE9\x05\x00\xBC\xF9\x05\x00\x82\x49"
-     "\x00\x00"
-     "\xA3\xC9\x05\x00\x80\xF5\x03\x00\x81\x49\x00\x00\xA6\xE1\x05\x00\x7E\x9A"
-     "\x00\x80"
-     "\x82\x9A\x00\x80\xA5\xC9\x05\x00\xAA\xDD\x05\x00\xAB\xA5\x05\x00\x86\xC8"
-     "\x7C\x00"
-     "\x87\xFC\x03\x00\xAE\xA5\x05\x00\xAF\xAD\x05\x00\xAC\xBD\x05\x00\xAD\xAD"
-     "\x05\x00"
-     "\xA8\x2D\x02\x00\xA9\x45\x02\x00\xAA\x5D\x02\x00\xAB\x55\x02\x00\xAC\x79"
-     "\x02\x00"
-     "\xAD\x79\x02\x00\xAE\x99\x03\x00\xAF\x99\x03\x00\x86\x9A\x00\x80\x8A\x9A"
-     "\x00\x80"
-     "\x8E\x9A\x00\x80\x92\x9A\x00\x80\x96\x9A\x00\x80\x9A\x9A\x00\x80\x9E\x9A"
-     "\x00\x80"
-     "\xA2\x9A\x00\x80\xB8\x9D\x03\x00\xB9\x65\x03\x00\xBA\x6D\x03\x00\xBB\x65"
-     "\x03\x00"
-     "\xBC\x7D\x03\x00\xBD\x65\x03\x00\xBE\x6D\x03\x00\xBF\x65\x03\x00\xB0\xC1"
-     "\x03\x00"
-     "\xB1\xC1\x03\x00\xB2\xC1\x03\x00\xB3\xC1\x03\x00\xB4\xC1\x03\x00\xB5\xC1"
-     "\x03\x00"
-     "\xB6\xC1\x03\x00\xB7\xC1\x03\x00\xA6\x9A\x00\x80\xAA\x9A\x00\x80\xAE\x9A"
-     "\x00\x80"
-     "\xB2\x9A\x00\x80\xB6\x9A\x00\x80\xEF\xE0\x03\x00\xBA\x9A\x00\x80\xBE\x9A"
-     "\x00\x80"
-     "\xBE\x68\x7D\x00\xE3\x1C\x03\x00\xC2\x9A\x00\x80\xE1\xFC\x01\x00\xC6\x9A"
-     "\x00\x80"
-     "\xCA\x9A\x00\x80\xCE\x9A\x00\x80\xD2\x9A\x00\x80\xB3\x95\x03\x00\xD6\x9A"
-     "\x00\x80"
-     "\xDA\x9A\x00\x80\xE2\x9A\x00\x80\x84\x1C\x7D\x00\xB6\xB9\x03\x00\xB5\xB1"
-     "\x03\x00"
-     "\xE6\x9A\x00\x80\xBB\x45\x02\x00\xBA\x45\x02\x00\x86\xA8\x7C\x00\x87\xA4"
-     "\x7D\x00"
-     "\xBF\x45\x02\x00\xBE\x45\x02\x00\xBD\x55\x02\x00\xBC\x55\x02\x00\xE1\x1C"
-     "\x06\x00"
-     "\xE1\x94\x07\x00\xE3\x50\x06\x00\xE3\xF4\x06\x00\x80\xAD\x00\x00\x81\xD1"
-     "\x00\x00"
-     "\x82\xD1\x00\x00\xEA\x9A\x00\x80\xEE\x9A\x00\x80\xF2\x9A\x00\x80\xF6\x9A"
-     "\x00\x80"
-     "\xFA\x9A\x00\x80\xFE\x9A\x00\x80\x02\x9B\x00\x80\xEF\x8C\x06\x00\xEF\x24"
-     "\x06\x00"
-     "\xA3\x15\x03\x00\x06\x9B\x00\x80\x0A\x9B\x00\x80\x0E\x9B\x00\x80\x12\x9B"
-     "\x00\x80"
-     "\xA6\x39\x03\x00\xA5\x31\x03\x00\x84\x90\x7C\x00\xAB\xC5\x02\x00\xAA\xC5"
-     "\x02\x00"
-     "\x16\x9B\x00\x80\x1A\x9B\x00\x80\xAF\xC5\x02\x00\xAE\xC5\x02\x00\xAD\xD5"
-     "\x02\x00"
-     "\xAC\xD5\x02\x00\xA8\x11\x7E\x00\xA9\x19\x7E\x00\xAA\x25\x7E\x00\xAB\x59"
-     "\x7E\x00"
-     "\xAC\x49\x7E\x00\xAD\xB5\x7E\x00\xAE\xBD\x7E\x00\xAF\xB5\x7E\x00\xDE\x9A"
-     "\x00\x80"
-     "\x1E\x9B\x00\x80\x22\x9B\x00\x80\x26\x9B\x00\x80\x2A\x9B\x00\x80\x2E\x9B"
-     "\x00\x80"
-     "\x32\x9B\x00\x80\x36\x9B\x00\x80\xB8\x4D\x7F\x00\xB9\x55\x7F\x00\xBA\x51"
-     "\x7F\x00"
-     "\xBB\x51\x7F\x00\xBC\x79\x7F\x00\xBD\x65\x7F\x00\xBE\x6D\x7F\x00\xBF\x65"
-     "\x7F\x00"
-     "\xB0\xCD\x7E\x00\xB1\xD5\x7E\x00\xB2\xDD\x7E\x00\xB3\xD5\x7E\x00\xB4\xCD"
-     "\x7E\x00"
-     "\xB5\x75\x7F\x00\xB6\x7D\x7F\x00\xB7\x75\x7F\x00\xB3\xAD\x7E\x00\x3A\x9B"
-     "\x00\x80"
-     "\x3E\x9B\x00\x80\x42\x9B\x00\x80\x46\x9B\x00\x80\xB6\x35\x7E\x00\xB5\xC1"
-     "\x7E\x00"
-     "\x4A\x9B\x00\x80\xBB\x11\x7E\x00\xBA\x09\x7E\x00\xBE\xE0\x03\x00\x4E\x9B"
-     "\x00\x80"
-     "\xBF\xB1\x7F\x00\xBE\x09\x7E\x00\xBD\x01\x7E\x00\xBC\x09\x7E\x00\x82\x5D"
-     "\x00\x00"
-     "\xA3\xE9\x7E\x00\x80\x45\x00\x00\x81\x5D\x00\x00\xA6\x71\x7E\x00\x52\x9B"
-     "\x00\x80"
-     "\x56\x9B\x00\x80\xA5\x85\x7E\x00\xAA\x4D\x7E\x00\xAB\x55\x7E\x00\x86\x48"
-     "\x00\x00"
-     "\x87\x6C\x00\x00\xAE\x4D\x7E\x00\xAF\xF5\x7F\x00\xAC\x4D\x7E\x00\xAD\x45"
-     "\x7E\x00"
-     "\xB0\xB5\x7E\x00\xB1\xC1\x7E\x00\xB2\xC1\x7E\x00\xB3\xC1\x7E\x00\xB4\xC1"
-     "\x7E\x00"
-     "\xB5\xFD\x7E\x00\xB6\xF5\x7E\x00\xB7\x39\x01\x00\xB8\x09\x01\x00\xB9\x09"
-     "\x01\x00"
-     "\xBA\x19\x01\x00\xBB\x19\x01\x00\xBC\x09\x01\x00\xBD\x09\x01\x00\xBE\xBD"
-     "\x01\x00"
-     "\xBF\xB5\x01\x00\x5A\x9B\x00\x80\x5E\x9B\x00\x80\x62\x9B\x00\x80\x6E\x9A"
-     "\x00\x80"
-     "\x66\x9B\x00\x80\x6A\x9B\x00\x80\x6E\x9B\x00\x80\x72\x9B\x00\x80\xA8\x2D"
-     "\x7E\x00"
-     "\xA9\x39\x7E\x00\xAA\xF5\x7E\x00\xAB\xCD\x7E\x00\xAC\xD5\x7E\x00\xAD\xDD"
-     "\x7E\x00"
-     "\xAE\xD5\x7E\x00\xAF\xCD\x7E\x00\xB3\x1D\x7E\x00\x76\x9B\x00\x80\x7A\x9B"
-     "\x00\x80"
-     "\x7E\x9B\x00\x80\x82\x9B\x00\x80\xB6\x3D\x7E\x00\xB5\x3D\x7E\x00\x86\x9B"
-     "\x00\x80"
-     "\xBB\x65\x7E\x00\xBA\x11\x7E\x00\x8A\x9B\x00\x80\x8E\x9B\x00\x80\xBF\x59"
-     "\x7E\x00"
-     "\xBE\x65\x7E\x00\xBD\x75\x7E\x00\xBC\x75\x7E\x00\x92\x9B\x00\x80\xA3\x59"
-     "\x7E\x00"
-     "\x96\x9B\x00\x80\x9A\x9B\x00\x80\xA6\x79\x7E\x00\x9E\x9B\x00\x80\xA2\x9B"
-     "\x00\x80"
-     "\xA5\x79\x7E\x00\xAA\x55\x7E\x00\xAB\x21\x7E\x00\xA6\x9B\x00\x80\xAA\x9B"
-     "\x00\x80"
-     "\xAE\x21\x7E\x00\xAF\x1D\x7E\x00\xAC\x31\x7E\x00\xAD\x31\x7E\x00\xA8\xD5"
-     "\x02\x00"
-     "\xA9\xDD\x02\x00\xAA\x49\x03\x00\xAB\x49\x03\x00\xAC\x59\x03\x00\xAD\x45"
-     "\x03\x00"
-     "\xAE\x4D\x03\x00\xAF\x45\x03\x00\xAE\x9B\x00\x80\xB2\x9B\x00\x80\xB6\x9B"
-     "\x00\x80"
-     "\xBA\x9B\x00\x80\xBE\xF4\x03\x00\xC2\x9B\x00\x80\xC6\x9B\x00\x80\xCA\x9B"
-     "\x00\x80"
-     "\xB8\x81\x03\x00\xB9\x8D\x03\x00\xBA\x85\x03\x00\xBB\x99\x03\x00\xBC\x89"
-     "\x03\x00"
-     "\xBD\xBD\x03\x00\xBE\xB5\x03\x00\xBF\x7D\x00\x00\xB0\x3D\x03\x00\xB1\xE1"
-     "\x03\x00"
-     "\xB2\xE1\x03\x00\xB3\xE1\x03\x00\xB4\xE1\x03\x00\xB5\xE9\x03\x00\xB6\xD1"
-     "\x03\x00"
-     "\xB7\xD1\x03\x00\x80\xA5\x00\x00\x81\xAD\x00\x00\x82\xA5\x00\x00\xB3\xCD"
-     "\x03\x00"
-     "\xCE\x9B\x00\x80\xB5\xDD\x03\x00\xB6\xC9\x03\x00\xD2\x9B\x00\x80\x84\xE0"
-     "\x02\x00"
-     "\xD6\x9B\x00\x80\xBA\xF5\x03\x00\xBB\x29\x03\x00\xBC\x39\x03\x00\xBD\x39"
-     "\x03\x00"
-     "\xBE\x29\x03\x00\xBF\x21\x03\x00\xE3\x80\x7F\x00\xDA\x9B\x00\x80\xE1\xB8"
-     "\x7E\x00"
-     "\xDE\x9B\x00\x80\x87\xE0\x03\x00\x86\x74\x04\x00\xE2\x9B\x00\x80\xE6\x9B"
-     "\x00\x80"
-     "\xEA\x9B\x00\x80\xEE\x9B\x00\x80\xF2\x9B\x00\x80\xF6\x9B\x00\x80\xEF\x28"
-     "\x7F\x00"
-     "\xFA\x9B\x00\x80\xFE\x9B\x00\x80\x02\x9C\x00\x80\x06\x9C\x00\x80\x0A\x9C"
-     "\x00\x80"
-     "\x0E\x9C\x00\x80\x12\x9C\x00\x80\xEF\xD8\x03\x00\x16\x9C\x00\x80\x1A\x9C"
-     "\x00\x80"
-     "\x1E\x9C\x00\x80\xE3\x90\x03\x00\x22\x9C\x00\x80\xE1\xD8\x01\x00\x26\x9C"
-     "\x00\x80"
-     "\x80\x15\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x2A\x9C\x00\x80\xBE\x4C"
-     "\x05\x00"
-     "\x32\x9C\x00\x80\xA3\x09\x02\x00\xBE\x40\x04\x00\xA5\x19\x02\x00\x36\x9C"
-     "\x00\x80"
-     "\x3A\x9C\x00\x80\xA6\x0D\x02\x00\x84\x6C\x02\x00\x3E\x9C\x00\x80\xAB\xED"
-     "\x02\x00"
-     "\xAA\x31\x02\x00\xAD\xFD\x02\x00\xAC\xFD\x02\x00\xAF\xE5\x02\x00\xAE\xED"
-     "\x02\x00"
-     "\x86\x40\x04\x00\x87\xB4\x05\x00\x42\x9C\x00\x80\x46\x9C\x00\x80\x4A\x9C"
-     "\x00\x80"
-     "\x4E\x9C\x00\x80\x52\x9C\x00\x80\xEF\xCC\x01\x00\x56\x9C\x00\x80\xE1\xD4"
-     "\x7E\x00"
-     "\x5A\x9C\x00\x80\xE3\x60\x7E\x00\x5E\x9C\x00\x80\x62\x9C\x00\x80\x66\x9C"
-     "\x00\x80"
-     "\x6A\x9C\x00\x80\xB3\xFD\x01\x00\x6E\x9C\x00\x80\x72\x9C\x00\x80\x76\x9C"
-     "\x00\x80"
-     "\x7A\x9C\x00\x80\xB6\x71\x01\x00\xB5\x45\x01\x00\x7E\x9C\x00\x80\xBB\x3D"
-     "\x01\x00"
-     "\xBA\x3D\x01\x00\x82\x9C\x00\x80\x86\x9C\x00\x80\xBF\x1D\x01\x00\xBE\x1D"
-     "\x01\x00"
-     "\xBD\x1D\x01\x00\xBC\x1D\x01\x00\xA8\x35\x06\x00\xA9\x3D\x06\x00\xAA\x35"
-     "\x06\x00"
-     "\xAB\x9D\x06\x00\xAC\x8D\x06\x00\xAD\xB1\x06\x00\xAE\xDD\x06\x00\xAF\xD1"
-     "\x06\x00"
-     "\x2E\x9C\x00\x80\x8A\x9C\x00\x80\x8E\x9C\x00\x80\x92\x9C\x00\x80\x80\x1D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\xBD\x00\x00\x96\x9C\x00\x80\xB8\xBD\x06\x00\xB9\x45"
-     "\x07\x00"
-     "\xBA\x4D\x07\x00\xBB\x45\x07\x00\xBC\x5D\x07\x00\xBD\x41\x07\x00\xBE\x41"
-     "\x07\x00"
-     "\xBF\x41\x07\x00\xB0\xB9\x06\x00\xB1\xB9\x06\x00\xB2\x89\x06\x00\xB3\x89"
-     "\x06\x00"
-     "\xB4\x99\x06\x00\xB5\x99\x06\x00\xB6\x8D\x06\x00\xB7\x85\x06\x00\xA3\xBD"
-     "\x06\x00"
-     "\x9A\x9C\x00\x80\x84\x80\x02\x00\x9E\x9C\x00\x80\xA2\x9C\x00\x80\xA6\x31"
-     "\x06\x00"
-     "\xA5\x05\x06\x00\xA6\x9C\x00\x80\xAB\x7D\x06\x00\xAA\x7D\x06\x00\x86\x08"
-     "\x00\x00"
-     "\x87\x60\x03\x00\xAF\x5D\x06\x00\xAE\x5D\x06\x00\xAD\x5D\x06\x00\xAC\x5D"
-     "\x06\x00"
-     "\xAA\x9C\x00\x80\xB3\x9D\x07\x00\xAE\x9C\x00\x80\xB2\x9C\x00\x80\xB6\x5D"
-     "\x07\x00"
-     "\xB6\x9C\x00\x80\xBA\x9C\x00\x80\xB5\x55\x07\x00\xBA\x79\x07\x00\xBB\x45"
-     "\x07\x00"
-     "\xBE\x9C\x00\x80\xC2\x9C\x00\x80\xBE\x45\x07\x00\xBF\x4D\x07\x00\xBC\x51"
-     "\x07\x00"
-     "\xBD\x4D\x07\x00\xB0\x99\x06\x00\xB1\xA5\x06\x00\xB2\xA1\x06\x00\xB3\xBD"
-     "\x06\x00"
-     "\xB4\xA5\x06\x00\xB5\xAD\x06\x00\xB6\xA5\x06\x00\xB7\xD9\x06\x00\xB8\xE9"
-     "\x06\x00"
-     "\xB9\xE9\x06\x00\xBA\xF9\x06\x00\xBB\xF1\x06\x00\xBC\xD9\x06\x00\xBD\xD9"
-     "\x06\x00"
-     "\xBE\x1D\x01\x00\xBF\x15\x01\x00\xBE\x9B\x00\x80\xC6\x9C\x00\x80\xCA\x9C"
-     "\x00\x80"
-     "\xCE\x9C\x00\x80\xD2\x9C\x00\x80\xD6\x9C\x00\x80\xDA\x9C\x00\x80\xDE\x9C"
-     "\x00\x80"
-     "\xA8\x25\x06\x00\xA9\x31\x06\x00\xAA\x31\x06\x00\xAB\x09\x06\x00\xAC\xF9"
-     "\x06\x00"
-     "\xAD\xF9\x06\x00\xAE\xE9\x06\x00\xAF\xE9\x06\x00\xA3\xD9\x06\x00\xE2\x9C"
-     "\x00\x80"
-     "\xE6\x9C\x00\x80\xEA\x9C\x00\x80\xEE\x9C\x00\x80\xA6\x19\x06\x00\xA5\x11"
-     "\x06\x00"
-     "\xF2\x9C\x00\x80\xAB\x01\x06\x00\xAA\x3D\x06\x00\xF6\x9C\x00\x80\xFA\x9C"
-     "\x00\x80"
-     "\xAF\x09\x06\x00\xAE\x01\x06\x00\xAD\x09\x06\x00\xAC\x15\x06\x00\xFE\x9C"
-     "\x00\x80"
-     "\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\x02\x9D\x00\x80\x06\x9D"
-     "\x00\x80"
-     "\x0A\x9D\x00\x80\xBE\x10\x00\x00\x84\xAC\x01\x00\x12\x9D\x00\x80\x87\x08"
-     "\x01\x00"
-     "\x86\x4C\x0F\x00\x16\x9D\x00\x80\x1A\x9D\x00\x80\x1E\x9D\x00\x80\x22\x9D"
-     "\x00\x80"
-     "\xA8\x91\x02\x00\xA9\x99\x02\x00\xAA\xC9\x02\x00\xAB\xC5\x02\x00\xAC\xDD"
-     "\x02\x00"
-     "\xAD\xC1\x02\x00\xAE\xC1\x02\x00\xAF\xF5\x02\x00\x26\x9D\x00\x80\x2A\x9D"
-     "\x00\x80"
-     "\x2E\x9D\x00\x80\x32\x9D\x00\x80\x8D\x00\x00\x00\x36\x9D\x00\x80\x3A\x9D"
-     "\x00\x80"
-     "\x3E\x9D\x00\x80\xB8\xC5\x03\x00\xB9\xCD\x03\x00\xBA\xC5\x03\x00\xBB\xDD"
-     "\x03\x00"
-     "\xBC\xCD\x03\x00\xBD\xFD\x03\x00\xBE\xF5\x03\x00\xBF\x9D\x03\x00\xB0\x8D"
-     "\x02\x00"
-     "\xB1\x21\x03\x00\xB2\x21\x03\x00\xB3\x21\x03\x00\xB4\x21\x03\x00\xB5\x21"
-     "\x03\x00"
-     "\xB6\x21\x03\x00\xB7\x21\x03\x00\xB3\x41\x02\x00\x42\x9D\x00\x80\x46\x9D"
-     "\x00\x80"
-     "\x4A\x9D\x00\x80\x4E\x9D\x00\x80\xB6\x41\x02\x00\xB5\x55\x02\x00\x52\x9D"
-     "\x00\x80"
-     "\xBB\xE5\x03\x00\xBA\x49\x02\x00\x56\x9D\x00\x80\x5A\x9D\x00\x80\xBF\xE5"
-     "\x03\x00"
-     "\xBE\xE5\x03\x00\xBD\xF5\x03\x00\xBC\xF5\x03\x00\x5E\x9D\x00\x80\xE1\x08"
-     "\x06\x00"
-     "\x62\x9D\x00\x80\xE3\xBC\x07\x00\x66\x9D\x00\x80\x6A\x9D\x00\x80\x6E\x9D"
-     "\x00\x80"
-     "\x72\x9D\x00\x80\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x76\x9D"
-     "\x00\x80"
-     "\x7A\x9D\x00\x80\x84\xF8\x03\x00\xBE\xF4\x0C\x00\xEF\x44\x07\x00\xE3\xF0"
-     "\x00\x00"
-     "\x82\x9D\x00\x80\xE1\xB8\x01\x00\x84\x04\x0D\x00\x86\xF0\x0C\x00\x87\xA4"
-     "\x02\x00"
-     "\x86\x9D\x00\x80\x8A\x9D\x00\x80\x8E\x9D\x00\x80\xBE\x28\x0C\x00\x92\x9D"
-     "\x00\x80"
-     "\x96\x9D\x00\x80\xEF\x14\x00\x00\x9A\x9D\x00\x80\x9E\x9D\x00\x80\xA2\x9D"
-     "\x00\x80"
-     "\xA3\x81\x03\x00\xA6\x9D\x00\x80\xAA\x9D\x00\x80\xAE\x9D\x00\x80\xB2\x9D"
-     "\x00\x80"
-     "\xA6\x81\x03\x00\xA5\x95\x03\x00\xB6\x9D\x00\x80\xAB\x25\x02\x00\xAA\x89"
-     "\x03\x00"
-     "\xBA\x9D\x00\x80\xBE\x9D\x00\x80\xAF\x25\x02\x00\xAE\x25\x02\x00\xAD\x35"
-     "\x02\x00"
-     "\xAC\x35\x02\x00\xC2\x9D\x00\x80\xC6\x9D\x00\x80\xCA\x9D\x00\x80\xCE\x9D"
-     "\x00\x80"
-     "\xD2\x9D\x00\x80\xD6\x9D\x00\x80\xDA\x9D\x00\x80\xEF\x28\x06\x00\xDE\x9D"
-     "\x00\x80"
-     "\xE1\x3C\x06\x00\xE2\x9D\x00\x80\xE3\x34\x01\x00\x80\x19\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x7D\x00\x00\xE6\x9D\x00\x80\xB3\x35\x02\x00\xEA\x9D\x00\x80\xF2\x9D"
-     "\x00\x80"
-     "\xBE\xC0\x0F\x00\xF6\x9D\x00\x80\xB6\xF5\x01\x00\xB5\x25\x02\x00\xFA\x9D"
-     "\x00\x80"
-     "\xBB\xD5\x01\x00\xBA\xC9\x01\x00\xFE\x9D\x00\x80\x02\x9E\x00\x80\xBF\x79"
-     "\x01\x00"
-     "\xBE\x71\x01\x00\xBD\xC5\x01\x00\xBC\xC5\x01\x00\xA8\xE5\x0D\x00\xA9\xF5"
-     "\x0D\x00"
-     "\xAA\xFD\x0D\x00\xAB\xF5\x0D\x00\xAC\xED\x0D\x00\xAD\x35\x0E\x00\xAE\x3D"
-     "\x0E\x00"
-     "\xAF\x35\x0E\x00\x7E\x9D\x00\x80\xEE\x9D\x00\x80\x86\x88\x00\x00\x87\x40"
-     "\x03\x00"
-     "\x06\x9E\x00\x80\x0A\x9E\x00\x80\x0E\x9E\x00\x80\x12\x9E\x00\x80\xB8\xDD"
-     "\x0E\x00"
-     "\xB9\xED\x0E\x00\xBA\xE5\x0E\x00\xBB\x8D\x0E\x00\xBC\x99\x0E\x00\xBD\x85"
-     "\x0E\x00"
-     "\xBE\x8D\x0E\x00\xBF\x85\x0E\x00\xB0\x4D\x0E\x00\xB1\x51\x0E\x00\xB2\x51"
-     "\x0E\x00"
-     "\xB3\x51\x0E\x00\xB4\xF5\x0E\x00\xB5\xFD\x0E\x00\xB6\xF5\x0E\x00\xB7\xED"
-     "\x0E\x00"
-     "\xA3\x75\x0D\x00\x16\x9E\x00\x80\x1A\x9E\x00\x80\x1E\x9E\x00\x80\x22\x9E"
-     "\x00\x80"
-     "\xA6\xB5\x0E\x00\xA5\x65\x0D\x00\x26\x9E\x00\x80\xAB\x95\x0E\x00\xAA\x89"
-     "\x0E\x00"
-     "\x2A\x9E\x00\x80\x2E\x9E\x00\x80\xAF\x39\x0E\x00\xAE\x31\x0E\x00\xAD\x85"
-     "\x0E\x00"
-     "\xAC\x85\x0E\x00\x32\x9E\x00\x80\xB3\xD9\x0F\x00\x36\x9E\x00\x80\x3A\x9E"
-     "\x00\x80"
-     "\xB6\xFD\x0F\x00\x3E\x9E\x00\x80\x42\x9E\x00\x80\xB5\xF5\x0F\x00\xBA\xDD"
-     "\x0F\x00"
-     "\xBB\xA1\x0F\x00\x46\x9E\x00\x80\x4A\x9E\x00\x80\xBE\xA1\x0F\x00\xBF\xA9"
-     "\x0F\x00"
-     "\xBC\xB1\x0F\x00\xBD\xB1\x0F\x00\xB8\x81\x0E\x00\xB9\x81\x0E\x00\xBA\x81"
-     "\x0E\x00"
-     "\xBB\x81\x0E\x00\xBC\x81\x0E\x00\xBD\x81\x0E\x00\xBE\x81\x0E\x00\xBF\xBD"
-     "\x0E\x00"
-     "\xB0\x1D\x0E\x00\xB1\x61\x0E\x00\xB2\x61\x0E\x00\xB3\x61\x0E\x00\xB4\x61"
-     "\x0E\x00"
-     "\xB5\x69\x0E\x00\xB6\x51\x0E\x00\xB7\x51\x0E\x00\xA8\x45\x0E\x00\xA9\x5D"
-     "\x0E\x00"
-     "\xAA\x55\x0E\x00\xAB\x6D\x0E\x00\xAC\x79\x0E\x00\xAD\x79\x0E\x00\xAE\x6D"
-     "\x0E\x00"
-     "\xAF\x65\x0E\x00\x4E\x9E\x00\x80\x82\x19\x00\x00\x81\x11\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\x52\x9E\x00\x80\x0E\x9D\x00\x80\x56\x9E\x00\x80\x5A\x9E\x00\x80\xA3\x9D"
-     "\x0E\x00"
-     "\x84\x8C\x02\x00\x86\x28\x00\x00\x87\x48\x01\x00\x62\x9E\x00\x80\xA6\xB9"
-     "\x0E\x00"
-     "\xA5\xB1\x0E\x00\x66\x9E\x00\x80\xAB\xE5\x0E\x00\xAA\x99\x0E\x00\x6A\x9E"
-     "\x00\x80"
-     "\x6E\x9E\x00\x80\xAF\xED\x0E\x00\xAE\xE5\x0E\x00\xAD\xF5\x0E\x00\xAC\xF5"
-     "\x0E\x00"
-     "\x72\x9E\x00\x80\xB3\x19\x0E\x00\x76\x9E\x00\x80\x7A\x9E\x00\x80\xB6\xE1"
-     "\x01\x00"
-     "\x7E\x9E\x00\x80\x82\x9E\x00\x80\xB5\x31\x0E\x00\xBA\xDD\x01\x00\xBB\x61"
-     "\x01\x00"
-     "\x86\x9E\x00\x80\x8A\x9E\x00\x80\xBE\x61\x01\x00\xBF\x69\x01\x00\xBC\x71"
-     "\x01\x00"
-     "\xBD\x71\x01\x00\xA8\xC9\x0D\x00\xA9\xD5\x0D\x00\xAA\xD5\x0D\x00\xAB\xE5"
-     "\x0D\x00"
-     "\xAC\xFD\x0D\x00\xAD\x25\x02\x00\xAE\x2D\x02\x00\xAF\x25\x02\x00\x8E\x9E"
-     "\x00\x80"
-     "\x92\x9E\x00\x80\x96\x9E\x00\x80\x9A\x9E\x00\x80\x8C\x00\x00\x00\x9E\x9E"
-     "\x00\x80"
-     "\xA2\x9E\x00\x80\xA6\x9E\x00\x80\xB8\x81\x02\x00\xB9\x8D\x02\x00\xBA\x85"
-     "\x02\x00"
-     "\xBB\x99\x02\x00\xBC\x89\x02\x00\xBD\xBD\x02\x00\xBE\xB5\x02\x00\xBF\x3D"
-     "\x03\x00"
-     "\xB0\x5D\x02\x00\xB1\xE1\x02\x00\xB2\xE1\x02\x00\xB3\xE1\x02\x00\xB4\xE1"
-     "\x02\x00"
-     "\xB5\xE9\x02\x00\xB6\xD1\x02\x00\xB7\xD1\x02\x00\xA3\x5D\x0D\x00\xAA\x9E"
-     "\x00\x80"
-     "\xAE\x9E\x00\x80\xB2\x9E\x00\x80\xB6\x9E\x00\x80\xA6\xA5\x02\x00\xA5\x75"
-     "\x0D\x00"
-     "\xBA\x9E\x00\x80\xAB\x25\x02\x00\xAA\x99\x02\x00\xBE\x9E\x00\x80\xC2\x9E"
-     "\x00\x80"
-     "\xAF\x2D\x02\x00\xAE\x25\x02\x00\xAD\x35\x02\x00\xAC\x35\x02\x00\x81\x11"
-     "\x00\x00"
-     "\x80\x69\x00\x00\xEF\xC8\x0E\x00\x82\x19\x00\x00\xC6\x9E\x00\x80\xCA\x9E"
-     "\x00\x80"
-     "\xD2\x9E\x00\x80\xD6\x9E\x00\x80\x84\xEC\x03\x00\xBE\x68\x04\x00\x87\xA8"
-     "\x04\x00"
-     "\x86\xCC\x04\x00\xE1\x04\x0D\x00\xDA\x9E\x00\x80\xE3\x1C\x0E\x00\xDE\x9E"
-     "\x00\x80"
-     "\xA8\x91\x02\x00\xA9\x99\x02\x00\xAA\xA1\x02\x00\xAB\xA1\x02\x00\xAC\xD1"
-     "\x02\x00"
-     "\xAD\xDD\x02\x00\xAE\xD5\x02\x00\xAF\xC9\x02\x00\xE2\x9E\x00\x80\xE6\x9E"
-     "\x00\x80"
-     "\xEA\x9E\x00\x80\xEE\x9E\x00\x80\xF2\x9E\x00\x80\xF6\x9E\x00\x80\xFA\x9E"
-     "\x00\x80"
-     "\xFE\x9E\x00\x80\xB8\x79\x01\x00\xB9\x79\x01\x00\xBA\xCD\x01\x00\xBB\xC5"
-     "\x01\x00"
-     "\xBC\xDD\x01\x00\xBD\xC5\x01\x00\xBE\xC5\x01\x00\xBF\xF5\x01\x00\xB0\xB9"
-     "\x02\x00"
-     "\xB1\x8D\x02\x00\xB2\x85\x02\x00\xB3\x4D\x01\x00\xB4\x55\x01\x00\xB5\x5D"
-     "\x01\x00"
-     "\xB6\x55\x01\x00\xB7\x49\x01\x00\x02\x9F\x00\x80\x06\x9F\x00\x80\x0A\x9F"
-     "\x00\x80"
-     "\x0E\x9F\x00\x80\x84\x10\x05\x00\xBE\x18\x04\x00\xEF\xAC\x0E\x00\xEF\x40"
-     "\x0E\x00"
-     "\xE1\x90\x01\x00\xE1\xAC\x0E\x00\xE3\x3C\x00\x00\xE3\xB0\x0E\x00\x12\x9F"
-     "\x00\x80"
-     "\x16\x9F\x00\x80\x1A\x9F\x00\x80\x1E\x9F\x00\x80\xB3\x69\x02\x00\x22\x9F"
-     "\x00\x80"
-     "\x26\x9F\x00\x80\x2A\x9F\x00\x80\x2E\x9F\x00\x80\xB6\x1D\x02\x00\xB5\x19"
-     "\x02\x00"
-     "\x32\x9F\x00\x80\xBB\x09\x02\x00\xBA\x3D\x02\x00\x36\x9F\x00\x80\x3A\x9F"
-     "\x00\x80"
-     "\xBF\xD9\x01\x00\xBE\xD1\x01\x00\xBD\xD9\x01\x00\xBC\x11\x02\x00\xA3\xA9"
-     "\x05\x00"
-     "\xCE\x9E\x00\x80\x3E\x9F\x00\x80\x42\x9F\x00\x80\x46\x9F\x00\x80\xA6\xDD"
-     "\x05\x00"
-     "\xA5\xD9\x05\x00\x4A\x9F\x00\x80\xAB\xC9\x05\x00\xAA\xFD\x05\x00\x86\x08"
-     "\x03\x00"
-     "\x87\x7C\x03\x00\xAF\x19\x06\x00\xAE\x11\x06\x00\xAD\x19\x06\x00\xAC\xD1"
-     "\x05\x00"
-     "\x80\x9D\x07\x00\x81\xAD\x07\x00\x82\xA5\x07\x00\xB3\xB9\x06\x00\x4E\x9F"
-     "\x00\x80"
-     "\xB5\x75\x07\x00\xB6\x7D\x07\x00\x52\x9F\x00\x80\x56\x9F\x00\x80\x5A\x9F"
-     "\x00\x80"
-     "\xBA\x59\x07\x00\xBB\x59\x07\x00\xBC\x49\x07\x00\xBD\x49\x07\x00\xBE\xF9"
-     "\x07\x00"
-     "\xBF\xF9\x07\x00\xA8\x59\x06\x00\xA9\x59\x06\x00\xAA\x7D\x06\x00\xAB\x71"
-     "\x06\x00"
-     "\xAC\xB1\x07\x00\xAD\xB1\x07\x00\xAE\xB1\x07\x00\xAF\xB1\x07\x00\x5E\x9F"
-     "\x00\x80"
-     "\x62\x9F\x00\x80\x66\x9F\x00\x80\x6A\x9F\x00\x80\x6E\x9F\x00\x80\x72\x9F"
-     "\x00\x80"
-     "\x76\x9F\x00\x80\x7A\x9F\x00\x80\xB8\x71\x07\x00\xB9\x71\x07\x00\xBA\x71"
-     "\x07\x00"
-     "\xBB\x71\x07\x00\xBC\x15\x07\x00\xBD\x1D\x07\x00\xBE\x15\x07\x00\xBF\x0D"
-     "\x07\x00"
-     "\xB0\xD1\x07\x00\xB1\xD1\x07\x00\xB2\xD1\x07\x00\xB3\xD1\x07\x00\xB4\x71"
-     "\x07\x00"
-     "\xB5\x71\x07\x00\xB6\x71\x07\x00\xB7\x71\x07\x00\xA3\xF5\x07\x00\x7E\x9F"
-     "\x00\x80"
-     "\x82\x9F\x00\x80\x86\x9F\x00\x80\x8A\x9F\x00\x80\xA6\x31\x06\x00\xA5\x39"
-     "\x06\x00"
-     "\x8E\x9F\x00\x80\xAB\x15\x06\x00\xAA\x15\x06\x00\x92\x9F\x00\x80\x96\x9F"
-     "\x00\x80"
-     "\xAF\xB5\x06\x00\xAE\xB5\x06\x00\xAD\x05\x06\x00\xAC\x05\x06\x00\x9A\x9F"
-     "\x00\x80"
-     "\x9E\x9F\x00\x80\xA2\x9F\x00\x80\xA6\x9F\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x05\x00\x00\xAA\x9F\x00\x80\x84\xAC\x01\x00\xBE\xAC\x01\x00\x5E\x9E"
-     "\x00\x80"
-     "\xAE\x9F\x00\x80\x86\xD0\x00\x00\x87\xE4\x01\x00\xB2\x9F\x00\x80\xB6\x9F"
-     "\x00\x80"
-     "\xA8\x8D\x01\x00\xA9\x95\x01\x00\xAA\x95\x01\x00\xAB\xDD\x01\x00\xAC\xCD"
-     "\x01\x00"
-     "\xAD\xF5\x01\x00\xAE\xFD\x01\x00\xAF\x3D\x01\x00\x84\xA0\x00\x00\xBA\x9F"
-     "\x00\x80"
-     "\xBE\x9F\x00\x80\xC2\x9F\x00\x80\xC6\x9F\x00\x80\xCA\x9F\x00\x80\xCE\x9F"
-     "\x00\x80"
-     "\xD2\x9F\x00\x80\xB8\xC9\x00\x00\xB9\xC9\x00\x00\xBA\xD9\x00\x00\xBB\xD1"
-     "\x00\x00"
-     "\xBC\xF9\x00\x00\xBD\xF9\x00\x00\xBE\x99\x00\x00\xBF\x99\x00\x00\xB0\x45"
-     "\x01\x00"
-     "\xB1\x4D\x01\x00\xB2\x45\x01\x00\xB3\x5D\x01\x00\xB4\x45\x01\x00\xB5\x4D"
-     "\x01\x00"
-     "\xB6\x45\x01\x00\xB7\xF9\x00\x00\xB3\x1D\x02\x00\xD6\x9F\x00\x80\xDA\x9F"
-     "\x00\x80"
-     "\xDE\x9F\x00\x80\xE2\x9F\x00\x80\xB6\x75\x02\x00\xB5\x45\x02\x00\xE6\x9F"
-     "\x00\x80"
-     "\xBB\x89\x02\x00\xBA\x81\x02\x00\xEA\x9F\x00\x80\xEE\x9F\x00\x80\xBF\x89"
-     "\x02\x00"
-     "\xBE\x81\x02\x00\xBD\x89\x02\x00\xBC\x91\x02\x00\x85\x5C\x1A\x00\xA3\x59"
-     "\x02\x00"
-     "\xF2\x9F\x00\x80\xF6\x9F\x00\x80\xA6\x31\x02\x00\xFA\x9F\x00\x80\xFE\x9F"
-     "\x00\x80"
-     "\xA5\x01\x02\x00\xAA\xC5\x02\x00\xAB\xCD\x02\x00\x02\xA0\x00\x80\x06\xA0"
-     "\x00\x80"
-     "\xAE\xC5\x02\x00\xAF\xCD\x02\x00\xAC\xD5\x02\x00\xAD\xCD\x02\x00\xBE\x00"
-     "\x1C\x00"
-     "\x0A\xA0\x00\x80\x0E\xA0\x00\x80\x84\x00\x1C\x00\x80\x0D\x00\x00\x81\x35"
-     "\x00\x00"
-     "\x82\x3D\x00\x00\x12\xA0\x00\x80\x16\xA0\x00\x80\x1A\xA0\x00\x80\x22\xA0"
-     "\x00\x80"
-     "\x26\xA0\x00\x80\x2A\xA0\x00\x80\x2E\xA0\x00\x80\x32\xA0\x00\x80\x36\xA0"
-     "\x00\x80"
-     "\x86\xE0\x1C\x00\x87\x04\x03\x00\x3A\xA0\x00\x80\x3E\xA0\x00\x80\xEF\xA8"
-     "\x02\x00"
-     "\xEF\xC8\x06\x00\x42\xA0\x00\x80\x46\xA0\x00\x80\xE3\x50\x01\x00\xE3\x14"
-     "\x06\x00"
-     "\xE1\x80\x01\x00\xE1\x7C\x05\x00\x4A\xA0\x00\x80\x4E\xA0\x00\x80\x52\xA0"
-     "\x00\x80"
-     "\x56\xA0\x00\x80\xA8\x79\x02\x00\xA9\x79\x02\x00\xAA\xAD\x02\x00\xAB\xA5"
-     "\x02\x00"
-     "\xAC\xBD\x02\x00\xAD\xA5\x02\x00\xAE\xAD\x02\x00\xAF\xA5\x02\x00\xBE\xAC"
-     "\x1C\x00"
-     "\x5A\xA0\x00\x80\x5E\xA0\x00\x80\x62\xA0\x00\x80\x66\xA0\x00\x80\x6A\xA0"
-     "\x00\x80"
-     "\x6E\xA0\x00\x80\x72\xA0\x00\x80\xB8\x09\x01\x00\xB9\x09\x01\x00\xBA\x1D"
-     "\x01\x00"
-     "\xBB\x15\x01\x00\xBC\x39\x01\x00\xBD\x39\x01\x00\xBE\xDD\x01\x00\xBF\xD5"
-     "\x01\x00"
-     "\xB0\xDD\x02\x00\xB1\x61\x01\x00\xB2\x61\x01\x00\xB3\x61\x01\x00\xB4\x61"
-     "\x01\x00"
-     "\xB5\x69\x01\x00\xB6\x51\x01\x00\xB7\x51\x01\x00\xE1\x74\x07\x00\xE3\xF4"
-     "\x00\x00"
-     "\xE3\x08\x06\x00\xE1\xBC\x01\x00\x76\xA0\x00\x80\x7A\xA0\x00\x80\x7E\xA0"
-     "\x00\x80"
-     "\x84\x14\x1C\x00\x82\xA0\x00\x80\x86\xA0\x00\x80\x8A\xA0\x00\x80\x92\xA0"
-     "\x00\x80"
-     "\xEC\x00\x00\x00\xEF\x28\x00\x00\xEF\xD4\x06\x00\x96\xA0\x00\x80\x80\x61"
-     "\x00\x00"
-     "\x81\x61\x00\x00\x82\x61\x00\x00\xB3\x69\x02\x00\x9A\xA0\x00\x80\xB5\x79"
-     "\x02\x00"
-     "\xB6\x69\x02\x00\x9E\xA0\x00\x80\x86\xC0\x1C\x00\x87\xE8\x1C\x00\xBA\x55"
-     "\x02\x00"
-     "\xBB\xE9\x01\x00\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\xE9\x01\x00\xBF\xE1"
-     "\x01\x00"
-     "\x1E\xA0\x00\x80\x8E\xA0\x00\x80\xA2\xA0\x00\x80\xA3\xA9\x1D\x00\xA6\xA0"
-     "\x00\x80"
-     "\xA5\xB9\x1D\x00\xA6\xA9\x1D\x00\xAA\xA0\x00\x80\xAE\xA0\x00\x80\xB2\xA0"
-     "\x00\x80"
-     "\xAA\x95\x1D\x00\xAB\x29\x1E\x00\xAC\x39\x1E\x00\xAD\x39\x1E\x00\xAE\x29"
-     "\x1E\x00"
-     "\xAF\x21\x1E\x00\xB3\xF1\x1E\x00\xB6\xA0\x00\x80\xBA\xA0\x00\x80\xBE\xA0"
-     "\x00\x80"
-     "\xC2\xA0\x00\x80\xB6\xE9\x1E\x00\xB5\xE1\x1E\x00\xC6\xA0\x00\x80\xBB\xBD"
-     "\x1E\x00"
-     "\xBA\xBD\x1E\x00\xCA\xA0\x00\x80\xCE\xA0\x00\x80\xBF\x8D\x1E\x00\xBE\x8D"
-     "\x1E\x00"
-     "\xBD\x8D\x1E\x00\xBC\x8D\x1E\x00\xA8\x29\x1E\x00\xA9\x29\x1E\x00\xAA\x39"
-     "\x1E\x00"
-     "\xAB\x39\x1E\x00\xAC\x29\x1E\x00\xAD\x29\x1E\x00\xAE\x51\x1E\x00\xAF\x51"
-     "\x1E\x00"
-     "\xD2\xA0\x00\x80\xD6\xA0\x00\x80\xDA\xA0\x00\x80\xDE\xA0\x00\x80\xE2\xA0"
-     "\x00\x80"
-     "\xE6\xA0\x00\x80\xEA\xA0\x00\x80\xEE\xA0\x00\x80\xB8\xED\x1F\x00\xB9\xF5"
-     "\x1F\x00"
-     "\xBA\xFD\x1F\x00\xBB\xF5\x1F\x00\xBC\xED\x1F\x00\xBD\x45\x1F\x00\xBE\x4D"
-     "\x1F\x00"
-     "\xBF\x45\x1F\x00\xB0\x39\x1E\x00\xB1\x39\x1E\x00\xB2\x19\x1E\x00\xB3\x19"
-     "\x1E\x00"
-     "\xB4\x09\x1E\x00\xB5\x09\x1E\x00\xB6\xDD\x1F\x00\xB7\xD5\x1F\x00\xA3\xB1"
-     "\x1F\x00"
-     "\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D\x00\x00\xF2\xA0\x00\x80\xA6\xA9"
-     "\x1F\x00"
-     "\xA5\xA1\x1F\x00\xF6\xA0\x00\x80\xAB\xFD\x1F\x00\xAA\xFD\x1F\x00\xFA\xA0"
-     "\x00\x80"
-     "\xBE\xA0\x01\x00\xAF\xCD\x1F\x00\xAE\xCD\x1F\x00\xAD\xCD\x1F\x00\xAC\xCD"
-     "\x1F\x00"
-     "\xFE\xA0\x00\x80\xB3\x95\x1E\x00\x86\xE8\x00\x00\x87\x2C\x01\x00\xB6\x41"
-     "\x01\x00"
-     "\x02\xA1\x00\x80\x06\xA1\x00\x80\xB5\x55\x01\x00\xBA\x45\x01\x00\xBB\x21"
-     "\x01\x00"
-     "\x0A\xA1\x00\x80\x0E\xA1\x00\x80\xBE\x21\x01\x00\xBF\x29\x01\x00\xBC\x31"
-     "\x01\x00"
-     "\xBD\x31\x01\x00\xA8\x2D\x1E\x00\xA9\x39\x1E\x00\xAA\x59\x1E\x00\xAB\x51"
-     "\x1E\x00"
-     "\xAC\x99\x01\x00\xAD\x85\x01\x00\xAE\x85\x01\x00\xAF\xBD\x01\x00\x84\xE0"
-     "\x01\x00"
-     "\x12\xA1\x00\x80\x16\xA1\x00\x80\x1A\xA1\x00\x80\x1E\xA1\x00\x80\x22\xA1"
-     "\x00\x80"
-     "\x26\xA1\x00\x80\x2A\xA1\x00\x80\xB8\x61\x01\x00\xB9\x61\x01\x00\xBA\x61"
-     "\x01\x00"
-     "\xBB\x61\x01\x00\xBC\x61\x01\x00\xBD\x61\x01\x00\xBE\x61\x01\x00\xBF\x61"
-     "\x01\x00"
-     "\xB0\xC5\x01\x00\xB1\xA1\x01\x00\xB2\xA1\x01\x00\xB3\xA1\x01\x00\xB4\xA1"
-     "\x01\x00"
-     "\xB5\xA9\x01\x00\xB6\x91\x01\x00\xB7\x91\x01\x00\xA3\xD1\x1D\x00\x2E\xA1"
-     "\x00\x80"
-     "\x32\xA1\x00\x80\x36\xA1\x00\x80\x3A\xA1\x00\x80\xA6\x05\x02\x00\xA5\x11"
-     "\x02\x00"
-     "\x3E\xA1\x00\x80\xAB\x65\x02\x00\xAA\x01\x02\x00\x42\xA1\x00\x80\x46\xA1"
-     "\x00\x80"
-     "\xAF\x6D\x02\x00\xAE\x65\x02\x00\xAD\x75\x02\x00\xAC\x75\x02\x00\x4A\xA1"
-     "\x00\x80"
-     "\x4E\xA1\x00\x80\x52\xA1\x00\x80\x56\xA1\x00\x80\x5A\xA1\x00\x80\x5E\xA1"
-     "\x00\x80"
-     "\x62\xA1\x00\x80\x66\xA1\x00\x80\x80\x39\x00\x00\x81\x39\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x6A\xA1\x00\x80\xBE\x3C\x03\x00\x72\xA1\x00\x80\x87\x44\x03\x00\x86\xDC"
-     "\x04\x00"
-     "\x84\xCC\x04\x00\x76\xA1\x00\x80\x7A\xA1\x00\x80\xB3\xE5\x03\x00\x7E\xA1"
-     "\x00\x80"
-     "\xB5\x31\x03\x00\xB6\x31\x03\x00\x82\xA1\x00\x80\x86\xA1\x00\x80\x8A\xA1"
-     "\x00\x80"
-     "\xBA\x1D\x03\x00\xBB\x15\x03\x00\xBC\x0D\x03\x00\xBD\xB1\x00\x00\xBE\xB1"
-     "\x00\x00"
-     "\xBF\xB1\x00\x00\x8E\xA1\x00\x80\xE1\x94\x01\x00\x92\xA1\x00\x80\xE3\x7C"
-     "\x01\x00"
-     "\x96\xA1\x00\x80\x9A\xA1\x00\x80\x9E\xA1\x00\x80\xA2\xA1\x00\x80\xA6\xA1"
-     "\x00\x80"
-     "\xAA\xA1\x00\x80\xAE\xA1\x00\x80\xB2\xA1\x00\x80\xB6\xA1\x00\x80\xBA\xA1"
-     "\x00\x80"
-     "\xBE\xA1\x00\x80\xEF\x74\x02\x00\xA8\x55\x02\x00\xA9\x5D\x02\x00\xAA\x61"
-     "\x02\x00"
-     "\xAB\x61\x02\x00\xAC\xB5\x02\x00\xAD\xBD\x02\x00\xAE\xB5\x02\x00\xAF\xAD"
-     "\x02\x00"
-     "\xBE\x68\x05\x00\xC2\xA1\x00\x80\xC6\xA1\x00\x80\xCA\xA1\x00\x80\xCE\xA1"
-     "\x00\x80"
-     "\x82\xAD\x00\x00\x81\xAD\x00\x00\x80\xAD\x00\x00\xB8\x69\x01\x00\xB9\x69"
-     "\x01\x00"
-     "\xBA\x09\x01\x00\xBB\x09\x01\x00\xBC\x19\x01\x00\xBD\x19\x01\x00\xBE\x09"
-     "\x01\x00"
-     "\xBF\x09\x01\x00\xB0\xD5\x02\x00\xB1\xDD\x02\x00\xB2\xD5\x02\x00\xB3\x69"
-     "\x01\x00"
-     "\xB4\x79\x01\x00\xB5\x79\x01\x00\xB6\x69\x01\x00\xB7\x61\x01\x00\xE1\x78"
-     "\x1E\x00"
-     "\xE1\x5C\x1F\x00\xE3\x04\x1E\x00\xE3\x10\x1F\x00\xD2\xA1\x00\x80\xD6\xA1"
-     "\x00\x80"
-     "\xDA\xA1\x00\x80\xE2\xA1\x00\x80\xE6\xA1\x00\x80\xEA\xA1\x00\x80\xEE\xA1"
-     "\x00\x80"
-     "\xF2\xA1\x00\x80\xBE\xF8\x04\x00\xF6\xA1\x00\x80\xEF\x6C\x00\x00\xEF\x64"
-     "\x1E\x00"
-     "\xA3\xA9\x02\x00\xFA\xA1\x00\x80\x86\x48\x04\x00\x87\x04\x05\x00\xFE\xA1"
-     "\x00\x80"
-     "\xA6\x7D\x02\x00\xA5\x7D\x02\x00\x02\xA2\x00\x80\xAB\x59\x02\x00\xAA\x51"
-     "\x02\x00"
-     "\x06\xA2\x00\x80\x0A\xA2\x00\x80\xAF\xFD\x01\x00\xAE\xFD\x01\x00\xAD\xFD"
-     "\x01\x00"
-     "\xAC\x41\x02\x00\xA8\x75\x06\x00\xA9\x81\x06\x00\xAA\x81\x06\x00\xAB\x9D"
-     "\x06\x00"
-     "\xAC\x8D\x06\x00\xAD\xB5\x06\x00\xAE\xBD\x06\x00\xAF\xDD\x06\x00\xDE\xA1"
-     "\x00\x80"
-     "\x0E\xA2\x00\x80\x12\xA2\x00\x80\x16\xA2\x00\x80\x1A\xA2\x00\x80\x1E\xA2"
-     "\x00\x80"
-     "\x22\xA2\x00\x80\x26\xA2\x00\x80\xB8\xAD\x06\x00\xB9\xB5\x06\x00\xBA\xBD"
-     "\x06\x00"
-     "\xBB\xB5\x06\x00\xBC\x75\x07\x00\xBD\x7D\x07\x00\xBE\x75\x07\x00\xBF\x6D"
-     "\x07\x00"
-     "\xB0\xA9\x06\x00\xB1\xB5\x06\x00\xB2\xBD\x06\x00\xB3\xB5\x06\x00\xB4\xAD"
-     "\x06\x00"
-     "\xB5\x95\x06\x00\xB6\x9D\x06\x00\xB7\x95\x06\x00\xB3\x09\x06\x00\x2A\xA2"
-     "\x00\x80"
-     "\x2E\xA2\x00\x80\x32\xA2\x00\x80\x36\xA2\x00\x80\xB6\x35\x06\x00\xB5\x05"
-     "\x06\x00"
-     "\x3A\xA2\x00\x80\xBB\x71\x06\x00\xBA\x6D\x06\x00\x3E\xA2\x00\x80\x42\xA2"
-     "\x00\x80"
-     "\xBF\x51\x06\x00\xBE\x69\x06\x00\xBD\x61\x06\x00\xBC\x69\x06\x00\x82\x19"
-     "\x00\x00"
-     "\xA3\x4D\x06\x00\x80\x65\x00\x00\x81\x19\x00\x00\xA6\x71\x06\x00\x46\xA2"
-     "\x00\x80"
-     "\x4A\xA2\x00\x80\xA5\x41\x06\x00\xAA\x29\x06\x00\xAB\x35\x06\x00\x84\xA4"
-     "\x03\x00"
-     "\xBF\xE4\x01\x00\xAE\x2D\x06\x00\xAF\x15\x06\x00\xAC\x2D\x06\x00\xAD\x25"
-     "\x06\x00"
-     "\x85\x6C\x01\x00\x52\xA2\x00\x80\x84\x08\x00\x00\x6E\xA1\x00\x80\x86\x00"
-     "\x0C\x00"
-     "\x87\x60\x00\x00\x56\xA2\x00\x80\x5A\xA2\x00\x80\xA8\xBD\x06\x00\xA9\x85"
-     "\x06\x00"
-     "\xAA\x8D\x06\x00\xAB\x85\x06\x00\xAC\x9D\x06\x00\xAD\x81\x06\x00\xAE\xFD"
-     "\x06\x00"
-     "\xAF\xED\x06\x00\xB0\x9D\x06\x00\xB1\xAD\x06\x00\xB2\xA5\x06\x00\xB3\x69"
-     "\x01\x00"
-     "\xB4\x79\x01\x00\xB5\x79\x01\x00\xB6\x69\x01\x00\xB7\x69\x01\x00\xB8\x59"
-     "\x01\x00"
-     "\xB9\x59\x01\x00\xBA\xED\x01\x00\xBB\xE5\x01\x00\xBC\xFD\x01\x00\xBD\xE5"
-     "\x01\x00"
-     "\xBE\xE5\x01\x00\xBF\xD5\x01\x00\xB3\x25\x06\x00\xBE\x08\x0F\x00\x4E\xA2"
-     "\x00\x80"
-     "\x5E\xA2\x00\x80\x62\xA2\x00\x80\xB6\x21\x06\x00\xB5\x35\x06\x00\x66\xA2"
-     "\x00\x80"
-     "\xBB\x49\x06\x00\xBA\x79\x06\x00\x6A\xA2\x00\x80\x6E\xA2\x00\x80\xBF\x9D"
-     "\x01\x00"
-     "\xBE\x9D\x01\x00\xBD\x9D\x01\x00\xBC\x51\x06\x00\x72\xA2\x00\x80\xA3\x61"
-     "\x06\x00"
-     "\x76\xA2\x00\x80\x7A\xA2\x00\x80\xA6\x65\x06\x00\x7E\xA2\x00\x80\x82\xA2"
-     "\x00\x80"
-     "\xA5\x71\x06\x00\xAA\x3D\x06\x00\xAB\x0D\x06\x00\x86\xA2\x00\x80\x8A\xA2"
-     "\x00\x80"
-     "\xAE\xD9\x01\x00\xAF\xD9\x01\x00\xAC\x15\x06\x00\xAD\xD9\x01\x00\xE1\x14"
-     "\x01\x00"
-     "\xE3\x1C\x01\x00\xE3\xD0\x36\x00\xE1\x04\x01\x00\x84\x90\x0C\x00\x8E\xA2"
-     "\x00\x80"
-     "\x92\xA2\x00\x80\x96\xA2\x00\x80\x8F\x21\x6A\x00\x9A\xA2\x00\x80\x9E\xA2"
-     "\x00\x80"
-     "\xA2\xA2\x00\x80\xA6\xA2\x00\x80\xEF\x48\x01\x00\xEF\xFC\x37\x00\xAA\xA2"
-     "\x00\x80"
-     "\x97\xCD\x12\x00\x96\x65\x12\x00\x95\x09\x13\x00\x94\xE1\x16\x00\x93\x71"
-     "\x16\x00"
-     "\x92\x19\x17\x00\x91\xF1\x6A\x00\x90\x59\x6A\x00\x9F\x61\x06\x00\x9E\x09"
-     "\x1B\x00"
-     "\x9D\x61\x1A\x00\x9C\x69\x1A\x00\x9B\x19\x1B\x00\x9A\x51\x1E\x00\x99\x59"
-     "\x1E\x00"
-     "\x98\x09\x1F\x00\xE3\x10\x06\x00\xAE\xA2\x00\x80\xE1\xEC\x06\x00\xB2\xA2"
-     "\x00\x80"
-     "\xA3\xB1\x02\x00\xA2\x99\x03\x00\xA1\xD9\x07\x00\xA0\x99\x06\x00\xB3\x9D"
-     "\x3A\x00"
-     "\xB2\x85\x3A\x00\xB1\x01\x38\x00\xB0\x01\x3E\x00\xEF\x68\x07\x00\xB6\xA2"
-     "\x00\x80"
-     "\xB5\xBD\x26\x00\xB4\x01\x24\x00\xAB\x01\x30\x00\xAA\x19\x36\x00\xA9\x45"
-     "\x36\x00"
-     "\xA8\x4D\x36\x00\xAF\xF1\x3E\x00\xAE\x01\x3C\x00\xAD\xC5\x33\x00\xAC\xFD"
-     "\x32\x00"
-     "\xA3\x65\x0E\x00\xA2\x65\x0E\x00\xA1\x61\x03\x00\xBA\xA2\x00\x80\xA7\xE1"
-     "\x0B\x00"
-     "\xA6\x71\x0A\x00\xA5\x79\x0A\x00\xA4\x89\x0F\x00\x82\x01\x6B\x00\x83\x01"
-     "\x6B\x00"
-     "\x86\x08\x0D\x00\x87\x4C\x0D\x00\x86\x31\x17\x00\x87\xB9\x17\x00\x84\xB1"
-     "\x6B\x00"
-     "\x85\x51\x16\x00\x8A\x61\x12\x00\x8B\x05\x13\x00\xC2\xA2\x00\x80\xC6\xA2"
-     "\x00\x80"
-     "\x8E\x11\x1F\x00\x8F\x11\x1F\x00\x8C\xB5\x13\x00\x8D\x49\x1E\x00\x92\x29"
-     "\x1A\x00"
-     "\x93\x41\x1B\x00\xBE\xC4\x0C\x00\x84\x80\x0F\x00\x96\x51\x07\x00\x97\x51"
-     "\x07\x00"
-     "\x94\x51\x1B\x00\x95\x31\x1A\x00\x9A\x21\x06\x00\x9B\x69\x02\x00\xCA\xA2"
-     "\x00\x80"
-     "\xCE\xA2\x00\x80\x81\xC5\x01\x00\x80\xF9\x01\x00\x9C\x59\x03\x00\x82\x91"
-     "\x01\x00"
-     "\xBE\x48\x0F\x00\xD2\xA2\x00\x80\xD6\xA2\x00\x80\xDA\xA2\x00\x80\xDE\xA2"
-     "\x00\x80"
-     "\xE2\xA2\x00\x80\xE6\xA2\x00\x80\xEF\xD4\x06\x00\xEA\xA2\x00\x80\xE1\x74"
-     "\x06\x00"
-     "\xEE\xA2\x00\x80\xE3\x54\x01\x00\xF2\xA2\x00\x80\xF6\xA2\x00\x80\xFA\xA2"
-     "\x00\x80"
-     "\xFE\xA2\x00\x80\x02\xA3\x00\x80\x06\xA3\x00\x80\x0A\xA3\x00\x80\xB3\xD1"
-     "\x02\x00"
-     "\x0E\xA3\x00\x80\xB5\x45\x02\x00\xB6\x45\x02\x00\x12\xA3\x00\x80\x16\xA3"
-     "\x00\x80"
-     "\x1A\xA3\x00\x80\xBA\x29\x02\x00\xBB\x25\x02\x00\xBC\x3D\x02\x00\xBD\x21"
-     "\x02\x00"
-     "\xBE\x21\x02\x00\xBF\x15\x02\x00\xA3\x1D\x0D\x00\x1E\xA3\x00\x80\x22\xA3"
-     "\x00\x80"
-     "\x26\xA3\x00\x80\x2A\xA3\x00\x80\xA6\x89\x0D\x00\xA5\x89\x0D\x00\x32\xA3"
-     "\x00\x80"
-     "\xAB\xE9\x0D\x00\xAA\xE5\x0D\x00\x86\x08\x03\x00\x87\x4C\x03\x00\xAF\xD9"
-     "\x0D\x00"
-     "\xAE\xED\x0D\x00\xAD\xED\x0D\x00\xAC\xF1\x0D\x00\x80\x49\x00\x00\x81\x49"
-     "\x00\x00"
-     "\x82\x59\x00\x00\xB3\x25\x0F\x00\x36\xA3\x00\x80\xB5\xD1\x0F\x00\xB6\xD1"
-     "\x0F\x00"
-     "\x3A\xA3\x00\x80\x3E\xA3\x00\x80\x42\xA3\x00\x80\xBA\xFD\x0F\x00\xBB\xF5"
-     "\x0F\x00"
-     "\xBC\xD1\x0F\x00\xBD\xDD\x0F\x00\xBE\xC9\x0F\x00\xBF\xC9\x0F\x00\xBE\xA2"
-     "\x00\x80"
-     "\x2E\xA3\x00\x80\x46\xA3\x00\x80\x4A\xA3\x00\x80\x4E\xA3\x00\x80\x52\xA3"
-     "\x00\x80"
-     "\x56\xA3\x00\x80\x5A\xA3\x00\x80\xA8\x89\x0F\x00\xA9\x89\x0F\x00\xAA\xC1"
-     "\x0F\x00"
-     "\xAB\xC1\x0F\x00\xAC\x35\x0F\x00\xAD\x3D\x0F\x00\xAE\x35\x0F\x00\xAF\x2D"
-     "\x0F\x00"
-     "\xB0\x55\x0F\x00\xB1\x5D\x0F\x00\xB2\x69\x0F\x00\xB3\x65\x0F\x00\xB4\x61"
-     "\x0F\x00"
-     "\xB5\x61\x0F\x00\xB6\x61\x0F\x00\xB7\x61\x0F\x00\xB8\x21\x0F\x00\xB9\x21"
-     "\x0F\x00"
-     "\xBA\x21\x0F\x00\xBB\x3D\x0F\x00\xBC\x2D\x0F\x00\xBD\x15\x0F\x00\xBE\x1D"
-     "\x0F\x00"
-     "\xBF\xF9\x00\x00\xA3\x69\x0E\x00\x5E\xA3\x00\x80\x62\xA3\x00\x80\x66\xA3"
-     "\x00\x80"
-     "\x84\x9C\x00\x00\xA6\x9D\x0E\x00\xA5\x9D\x0E\x00\x6A\xA3\x00\x80\xAB\xB9"
-     "\x0E\x00"
-     "\xAA\xB1\x0E\x00\x6E\xA3\x00\x80\x72\xA3\x00\x80\xAF\x85\x0E\x00\xAE\x85"
-     "\x0E\x00"
-     "\xAD\x91\x0E\x00\xAC\x9D\x0E\x00\x76\xA3\x00\x80\x7A\xA3\x00\x80\x7E\xA3"
-     "\x00\x80"
-     "\x82\xA3\x00\x80\x86\xA3\x00\x80\x8A\xA3\x00\x80\x8E\xA3\x00\x80\x92\xA3"
-     "\x00\x80"
-     "\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x96\xA3\x00\x80\x9A\xA3"
-     "\x00\x80"
-     "\x9E\xA3\x00\x80\x87\xD8\x01\x00\x86\x00\x04\x00\xA8\x9D\x01\x00\xA9\xD5"
-     "\x01\x00"
-     "\xAA\xD5\x01\x00\xAB\xED\x01\x00\xAC\xF5\x01\x00\xAD\x1D\x01\x00\xAE\x15"
-     "\x01\x00"
-     "\xAF\x0D\x01\x00\xA2\xA3\x00\x80\xA6\xA3\x00\x80\xAA\xA3\x00\x80\xAE\xA3"
-     "\x00\x80"
-     "\xB2\xA3\x00\x80\xB6\xA3\x00\x80\xBA\xA3\x00\x80\xBE\xA3\x00\x80\xB8\xED"
-     "\x00\x00"
-     "\xB9\x85\x00\x00\xBA\x8D\x00\x00\xBB\x85\x00\x00\xBC\x9D\x00\x00\xBD\x85"
-     "\x00\x00"
-     "\xBE\x8D\x00\x00\xBF\x85\x00\x00\xB0\x75\x01\x00\xB1\x7D\x01\x00\xB2\x75"
-     "\x01\x00"
-     "\xB3\xE9\x00\x00\xB4\xF9\x00\x00\xB5\xF9\x00\x00\xB6\xE5\x00\x00\xB7\xD5"
-     "\x00\x00"
-     "\xB3\x5D\x02\x00\xC2\xA3\x00\x80\xC6\xA3\x00\x80\xCA\xA3\x00\x80\xCE\xA3"
-     "\x00\x80"
-     "\xB6\x9D\x02\x00\xB5\x7D\x02\x00\xD2\xA3\x00\x80\xBB\xB9\x02\x00\xBA\xB1"
-     "\x02\x00"
-     "\xD6\xA3\x00\x80\xDA\xA3\x00\x80\xBF\x3D\x03\x00\xBE\x3D\x03\x00\xBD\x3D"
-     "\x03\x00"
-     "\xBC\xA1\x02\x00\xBE\xCC\x04\x00\xA3\x19\x02\x00\xDE\xA3\x00\x80\xE2\xA3"
-     "\x00\x80"
-     "\xA6\xD9\x02\x00\xE6\xA3\x00\x80\xEA\xA3\x00\x80\xA5\x39\x02\x00\xAA\xF5"
-     "\x02\x00"
-     "\xAB\xFD\x02\x00\xEE\xA3\x00\x80\xF2\xA3\x00\x80\xAE\x79\x03\x00\xAF\x79"
-     "\x03\x00"
-     "\xAC\xE5\x02\x00\xAD\x79\x03\x00\xBE\x0C\x05\x00\xF6\xA3\x00\x80\xFA\xA3"
-     "\x00\x80"
-     "\xBE\x10\x06\x00\xFE\xA3\x00\x80\x02\xA4\x00\x80\x06\xA4\x00\x80\x0A\xA4"
-     "\x00\x80"
-     "\x80\x35\x00\x00\x81\x3D\x00\x00\x82\x35\x00\x00\x12\xA4\x00\x80\x86\xB0"
-     "\x04\x00"
-     "\x87\x34\x03\x00\x16\xA4\x00\x80\x1A\xA4\x00\x80\x1E\xA4\x00\x80\x22\xA4"
-     "\x00\x80"
-     "\xEF\x70\x02\x00\x26\xA4\x00\x80\x2A\xA4\x00\x80\x2E\xA4\x00\x80\x32\xA4"
-     "\x00\x80"
-     "\x36\xA4\x00\x80\x3A\xA4\x00\x80\x3E\xA4\x00\x80\x42\xA4\x00\x80\x46\xA4"
-     "\x00\x80"
-     "\xE1\x94\x01\x00\x4A\xA4\x00\x80\xE3\xF8\x01\x00\x4E\xA4\x00\x80\xE1\x30"
-     "\x01\x00"
-     "\xE1\x14\x0E\x00\xE3\x54\x0E\x00\xE3\x0C\x0F\x00\x52\xA4\x00\x80\x56\xA4"
-     "\x00\x80"
-     "\x5A\xA4\x00\x80\x5E\xA4\x00\x80\x84\xA8\x05\x00\x62\xA4\x00\x80\x66\xA4"
-     "\x00\x80"
-     "\x6A\xA4\x00\x80\x6E\xA4\x00\x80\x72\xA4\x00\x80\xEF\x40\x0E\x00\xEF\x8C"
-     "\x0E\x00"
-     "\xB3\xE9\x02\x00\x76\xA4\x00\x80\x7A\xA4\x00\x80\x82\xA4\x00\x80\x86\xA4"
-     "\x00\x80"
-     "\xB6\xE1\x02\x00\xB5\xE9\x02\x00\x8A\xA4\x00\x80\xBB\x69\x01\x00\xBA\x61"
-     "\x01\x00"
-     "\x8E\xA4\x00\x80\x92\xA4\x00\x80\xBF\x59\x01\x00\xBE\x69\x01\x00\xBD\x69"
-     "\x01\x00"
-     "\xBC\x71\x01\x00\x80\xB5\x00\x00\x81\xC1\x00\x00\x82\xC1\x00\x00\x84\x04"
-     "\x04\x00"
-     "\x86\xB0\x04\x00\x87\x08\x05\x00\x96\xA4\x00\x80\x9A\xA4\x00\x80\xAC\x31"
-     "\x0E\x00"
-     "\xAD\x39\x0E\x00\xAE\x01\x0E\x00\xAF\x01\x0E\x00\xA8\xB5\x0E\x00\xA9\xAD"
-     "\x0E\x00"
-     "\xAA\xA5\x0E\x00\xAB\x31\x0E\x00\xB4\x19\x0E\x00\xB5\x2D\x0E\x00\xB6\x25"
-     "\x0E\x00"
-     "\xB7\xD9\x0D\x00\xB0\x11\x0E\x00\xB1\x1D\x0E\x00\xB2\x15\x0E\x00\xB3\x09"
-     "\x0E\x00"
-     "\xBC\xCD\x0D\x00\xBD\xD5\x0D\x00\xBE\xD5\x0D\x00\xBF\x69\x0D\x00\xB8\xE9"
-     "\x0D\x00"
-     "\xB9\xE9\x0D\x00\xBA\xDD\x0D\x00\xBB\xD5\x0D\x00\xA3\xA9\x05\x00\x9E\xA4"
-     "\x00\x80"
-     "\xA2\xA4\x00\x80\xA6\xA4\x00\x80\xAA\xA4\x00\x80\xA6\xA1\x05\x00\xA5\xA9"
-     "\x05\x00"
-     "\xAE\xA4\x00\x80\xAB\x29\x06\x00\xAA\x21\x06\x00\xB2\xA4\x00\x80\xB6\xA4"
-     "\x00\x80"
-     "\xAF\x19\x06\x00\xAE\x29\x06\x00\xAD\x29\x06\x00\xAC\x31\x06\x00\xBA\xA4"
-     "\x00\x80"
-     "\xB3\xF9\x06\x00\xBE\xA4\x00\x80\xC2\xA4\x00\x80\xB6\x9D\x06\x00\xC6\xA4"
-     "\x00\x80"
-     "\xCA\xA4\x00\x80\xB5\x99\x06\x00\xBA\xB9\x06\x00\xBB\xB9\x06\x00\xCE\xA4"
-     "\x00\x80"
-     "\xD2\xA4\x00\x80\xBE\x99\x06\x00\xBF\x99\x06\x00\xBC\x99\x06\x00\xBD\x99"
-     "\x06\x00"
-     "\x0E\xA4\x00\x80\x7E\xA4\x00\x80\xD6\xA4\x00\x80\xDA\xA4\x00\x80\xDE\xA4"
-     "\x00\x80"
-     "\x82\x99\x01\x00\x81\x91\x01\x00\x80\x99\x01\x00\xA8\x0D\x06\x00\xA9\x1D"
-     "\x06\x00"
-     "\xAA\x15\x06\x00\xAB\x6D\x06\x00\xAC\x75\x06\x00\xAD\x79\x06\x00\xAE\x69"
-     "\x06\x00"
-     "\xAF\x69\x06\x00\xB0\x19\x06\x00\xB1\x19\x06\x00\xB2\x29\x06\x00\xB3\x29"
-     "\x06\x00"
-     "\xB4\x39\x06\x00\xB5\x39\x06\x00\xB6\x2D\x06\x00\xB7\x21\x06\x00\xB8\xF1"
-     "\x07\x00"
-     "\xB9\xF1\x07\x00\xBA\xF1\x07\x00\xBB\xF1\x07\x00\xBC\xD5\x07\x00\xBD\xDD"
-     "\x07\x00"
-     "\xBE\xD5\x07\x00\xBF\xCD\x07\x00\xA3\xBD\x07\x00\xE2\xA4\x00\x80\xE6\xA4"
-     "\x00\x80"
-     "\xEA\xA4\x00\x80\xEE\xA4\x00\x80\xA6\xD9\x07\x00\xA5\xDD\x07\x00\xF2\xA4"
-     "\x00\x80"
-     "\xAB\xFD\x07\x00\xAA\xFD\x07\x00\x86\x00\x3C\x00\x87\x7C\x01\x00\xAF\xDD"
-     "\x07\x00"
-     "\xAE\xDD\x07\x00\xAD\xDD\x07\x00\xAC\xDD\x07\x00\xF6\xA4\x00\x80\xB3\x19"
-     "\x06\x00"
-     "\xFA\xA4\x00\x80\xFE\xA4\x00\x80\xB6\xC9\x06\x00\x02\xA5\x00\x80\x06\xA5"
-     "\x00\x80"
-     "\xB5\x31\x06\x00\xBA\x69\x01\x00\xBB\x69\x01\x00\x0A\xA5\x00\x80\x0E\xA5"
-     "\x00\x80"
-     "\xBE\x59\x01\x00\xBF\x59\x01\x00\xBC\x69\x01\x00\xBD\x61\x01\x00\xA8\xCD"
-     "\x06\x00"
-     "\xA9\xD5\x06\x00\xAA\xD5\x06\x00\xAB\xE5\x06\x00\xAC\xFD\x06\x00\xAD\x0D"
-     "\x06\x00"
-     "\xAE\x05\x06\x00\xAF\x99\x01\x00\x12\xA5\x00\x80\x16\xA5\x00\x80\x1A\xA5"
-     "\x00\x80"
-     "\x1E\xA5\x00\x80\x22\xA5\x00\x80\x26\xA5\x00\x80\x2A\xA5\x00\x80\x2E\xA5"
-     "\x00\x80"
-     "\xB8\x8D\x01\x00\xB9\x91\x01\x00\xBA\x91\x01\x00\xBB\xA5\x01\x00\xBC\xBD"
-     "\x01\x00"
-     "\xBD\x5D\x01\x00\xBE\x55\x01\x00\xBF\x4D\x01\x00\xB0\xE9\x01\x00\xB1\xE9"
-     "\x01\x00"
-     "\xB2\xF9\x01\x00\xB3\xF1\x01\x00\xB4\xD9\x01\x00\xB5\xD9\x01\x00\xB6\xB9"
-     "\x01\x00"
-     "\xB7\xB5\x01\x00\xBE\xCC\x03\x00\xA6\x9D\x05\x00\xA5\x65\x05\x00\x32\xA5"
-     "\x00\x80"
-     "\xA3\x4D\x05\x00\x36\xA5\x00\x80\x3A\xA5\x00\x80\x3E\xA5\x00\x80\xAF\x0D"
-     "\x02\x00"
-     "\xAE\x0D\x02\x00\xAD\x35\x02\x00\xAC\x3D\x02\x00\xAB\x3D\x02\x00\xAA\x3D"
-     "\x02\x00"
-     "\x42\xA5\x00\x80\x46\xA5\x00\x80\x4A\xA5\x00\x80\x4E\xA5\x00\x80\x52\xA5"
-     "\x00\x80"
-     "\x56\xA5\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x5A\xA5"
-     "\x00\x80"
-     "\x84\x60\x02\x00\x62\xA5\x00\x80\x87\x48\x03\x00\x86\xCC\x3C\x00\x66\xA5"
-     "\x00\x80"
-     "\x6A\xA5\x00\x80\x6E\xA5\x00\x80\x72\xA5\x00\x80\x76\xA5\x00\x80\x7A\xA5"
-     "\x00\x80"
-     "\x84\xC0\x3C\x00\xA6\x00\x00\x00\x7E\xA5\x00\x80\xE1\xA4\x01\x00\x82\xA5"
-     "\x00\x80"
-     "\xE3\xEC\x02\x00\x86\xA5\x00\x80\x8A\xA5\x00\x80\x8E\xA5\x00\x80\xEF\x54"
-     "\x02\x00"
-     "\x92\xA5\x00\x80\x96\xA5\x00\x80\x9A\xA5\x00\x80\x9E\xA5\x00\x80\xA2\xA5"
-     "\x00\x80"
-     "\xA6\xA5\x00\x80\xAA\xA5\x00\x80\xB3\x85\x03\x00\xAE\xA5\x00\x80\xB5\x81"
-     "\x03\x00"
-     "\xB6\x81\x03\x00\xB2\xA5\x00\x80\xB6\xA5\x00\x80\xBA\xA5\x00\x80\xBA\x49"
-     "\x03\x00"
-     "\xBB\x49\x03\x00\xBC\x59\x03\x00\xBD\x59\x03\x00\xBE\x49\x03\x00\xBF\x49"
-     "\x03\x00"
-     "\xE1\x88\x06\x00\xE1\xAC\x07\x00\xE3\x94\x06\x00\xE3\x3C\x06\x00\xBE\xA5"
-     "\x00\x80"
-     "\xC2\xA5\x00\x80\xC6\xA5\x00\x80\xCA\xA5\x00\x80\x80\x35\x00\x00\x81\x3D"
-     "\x00\x00"
-     "\x82\x35\x00\x00\xBE\x60\x3D\x00\x86\x90\x3C\x00\x87\xF8\x3D\x00\xEF\x4C"
-     "\x07\x00"
-     "\xEF\x44\x07\x00\xA3\x09\x02\x00\xD2\xA5\x00\x80\xD6\xA5\x00\x80\xDA\xA5"
-     "\x00\x80"
-     "\xDE\xA5\x00\x80\xA6\x0D\x02\x00\xA5\x0D\x02\x00\xE2\xA5\x00\x80\xAB\xC5"
-     "\x02\x00"
-     "\xAA\xC5\x02\x00\xE6\xA5\x00\x80\xEA\xA5\x00\x80\xAF\xC5\x02\x00\xAE\xC5"
-     "\x02\x00"
-     "\xAD\xD5\x02\x00\xAC\xD5\x02\x00\xBF\x11\x05\x00\xBE\x0D\x05\x00\xBD\x0D"
-     "\x05\x00"
-     "\xBC\xE1\x05\x00\xBB\xF9\x05\x00\xBA\xCD\x05\x00\xB9\xCD\x05\x00\xB8\xD1"
-     "\x05\x00"
-     "\xB7\xE9\x05\x00\xB6\x89\x05\x00\xB5\x81\x05\x00\xB4\xB1\x05\x00\xB3\xA1"
-     "\x05\x00"
-     "\xB2\xB9\x05\x00\xB1\xB1\x05\x00\xB0\xB9\x05\x00\xAF\x7D\x06\x00\xAE\x7D"
-     "\x06\x00"
-     "\xAD\x7D\x06\x00\xAC\x7D\x06\x00\xAB\x7D\x06\x00\xAA\x7D\x06\x00\xA9\x7D"
-     "\x06\x00"
-     "\xA8\x7D\x06\x00\xEE\xA5\x00\x80\xF2\xA5\x00\x80\xF6\xA5\x00\x80\xFA\xA5"
-     "\x00\x80"
-     "\xFE\xA5\x00\x80\x02\xA6\x00\x80\x06\xA6\x00\x80\x0A\xA6\x00\x80\xA8\x65"
-     "\x3E\x00"
-     "\xA9\x75\x3E\x00\xAA\x7D\x3E\x00\xAB\x75\x3E\x00\xAC\x6D\x3E\x00\xAD\x95"
-     "\x3E\x00"
-     "\xAE\x9D\x3E\x00\xAF\x8D\x3E\x00\x0E\xA6\x00\x80\x12\xA6\x00\x80\x16\xA6"
-     "\x00\x80"
-     "\x1A\xA6\x00\x80\x1E\xA6\x00\x80\x22\xA6\x00\x80\x26\xA6\x00\x80\x2A\xA6"
-     "\x00\x80"
-     "\xB8\x81\x3E\x00\xB9\x81\x3E\x00\xBA\x81\x3E\x00\xBB\x9D\x3E\x00\xBC\x8D"
-     "\x3E\x00"
-     "\xBD\xB5\x3E\x00\xBE\xBD\x3E\x00\xBF\x4D\x3F\x00\xB0\xF9\x3E\x00\xB1\xC5"
-     "\x3E\x00"
-     "\xB2\xCD\x3E\x00\xB3\xC5\x3E\x00\xB4\xC1\x3E\x00\xB5\xC1\x3E\x00\xB6\xC1"
-     "\x3E\x00"
-     "\xB7\xC1\x3E\x00\x5E\xA5\x00\x80\xB3\xF1\x3E\x00\x2E\xA6\x00\x80\xCE\xA5"
-     "\x00\x80"
-     "\xB6\x11\x3E\x00\x32\xA6\x00\x80\x36\xA6\x00\x80\xB5\xE1\x3E\x00\xBA\x39"
-     "\x3E\x00"
-     "\xBB\x05\x3E\x00\x3A\xA6\x00\x80\x84\xE0\x03\x00\xBE\x01\x3E\x00\xBF\x01"
-     "\x3E\x00"
-     "\xBC\x1D\x3E\x00\xBD\x15\x3E\x00\xA3\xB5\x3E\x00\x82\x6D\x00\x00\x81\x51"
-     "\x00\x00"
-     "\x80\x59\x00\x00\xBE\x98\x03\x00\xA6\x55\x3E\x00\xA5\xA5\x3E\x00\x3E\xA6"
-     "\x00\x80"
-     "\xAB\x41\x3E\x00\xAA\x7D\x3E\x00\x86\x48\x00\x00\x87\x68\x00\x00\xAF\x45"
-     "\x3E\x00"
-     "\xAE\x45\x3E\x00\xAD\x51\x3E\x00\xAC\x59\x3E\x00\xA8\x45\x3E\x00\xA9\x4D"
-     "\x3E\x00"
-     "\xAA\x45\x3E\x00\xAB\x5D\x3E\x00\xAC\x4D\x3E\x00\xAD\x7D\x3E\x00\xAE\x75"
-     "\x3E\x00"
-     "\xAF\xFD\x3E\x00\x42\xA6\x00\x80\x46\xA6\x00\x80\x84\x6C\x01\x00\x4A\xA6"
-     "\x00\x80"
-     "\x4E\xA6\x00\x80\x52\xA6\x00\x80\x56\xA6\x00\x80\x5A\xA6\x00\x80\xB8\x65"
-     "\x01\x00"
-     "\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x7D\x01\x00\xBC\x65\x01\x00\xBD\x6D"
-     "\x01\x00"
-     "\xBE\x65\x01\x00\xBF\xD5\x01\x00\xB0\x85\x3E\x00\xB1\x8D\x3E\x00\xB2\x85"
-     "\x3E\x00"
-     "\xB3\x9D\x3E\x00\xB4\x85\x3E\x00\xB5\x8D\x3E\x00\xB6\x85\x3E\x00\xB7\x5D"
-     "\x01\x00"
-     "\xB3\xCD\x3E\x00\x5E\xA6\x00\x80\x62\xA6\x00\x80\x66\xA6\x00\x80\x6A\xA6"
-     "\x00\x80"
-     "\xB6\xFD\x3E\x00\xB5\xC5\x3E\x00\x6E\xA6\x00\x80\xBB\x49\x3E\x00\xBA\x41"
-     "\x3E\x00"
-     "\x72\xA6\x00\x80\x76\xA6\x00\x80\xBF\x49\x3E\x00\xBE\x41\x3E\x00\xBD\x49"
-     "\x3E\x00"
-     "\xBC\x51\x3E\x00\x7A\xA6\x00\x80\xA3\x89\x3E\x00\x7E\xA6\x00\x80\x82\xA6"
-     "\x00\x80"
-     "\xA6\xB9\x3E\x00\x86\xA6\x00\x80\x8A\xA6\x00\x80\xA5\x81\x3E\x00\xAA\x05"
-     "\x3E\x00"
-     "\xAB\x0D\x3E\x00\x8E\xA6\x00\x80\x92\xA6\x00\x80\xAE\x05\x3E\x00\xAF\x0D"
-     "\x3E\x00"
-     "\xAC\x15\x3E\x00\xAD\x0D\x3E\x00\xA8\x21\x03\x00\xA9\x21\x03\x00\xAA\x25"
-     "\x03\x00"
-     "\xAB\x3D\x03\x00\xAC\x25\x03\x00\xAD\x2D\x03\x00\xAE\x25\x03\x00\xAF\x99"
-     "\x03\x00"
-     "\x96\xA6\x00\x80\x9A\xA6\x00\x80\xBE\x64\x02\x00\x9E\xA6\x00\x80\xA2\xA6"
-     "\x00\x80"
-     "\xA6\xA6\x00\x80\xAA\xA6\x00\x80\xB2\xA6\x00\x80\xB8\x8D\x03\x00\xB9\x91"
-     "\x03\x00"
-     "\xBA\x91\x03\x00\xBB\xA5\x03\x00\xBC\xBD\x03\x00\xBD\x75\x00\x00\xBE\x7D"
-     "\x00\x00"
-     "\xBF\x71\x00\x00\xB0\xE9\x03\x00\xB1\xE9\x03\x00\xB2\xF9\x03\x00\xB3\xF1"
-     "\x03\x00"
-     "\xB4\xD9\x03\x00\xB5\xD9\x03\x00\xB6\xB9\x03\x00\xB7\xB5\x03\x00\x81\xB5"
-     "\x00\x00"
-     "\x80\xB5\x00\x00\xB3\xA5\x03\x00\x82\xB5\x00\x00\xB5\xA5\x03\x00\xB6\xA6"
-     "\x00\x80"
-     "\xBA\xA6\x00\x80\xB6\xA5\x03\x00\x86\x00\x04\x00\x87\xD4\x03\x00\xBB\x25"
-     "\x03\x00"
-     "\xBA\x25\x03\x00\xBD\x2D\x03\x00\xBC\x25\x03\x00\xBF\x15\x03\x00\xBE\x15"
-     "\x03\x00"
-     "\xA3\xED\x03\x00\xBE\xA6\x00\x80\xC2\xA6\x00\x80\x85\x60\x11\x00\xC6\xA6"
-     "\x00\x80"
-     "\xA6\xED\x03\x00\xA5\xED\x03\x00\xCA\xA6\x00\x80\xAB\x6D\x03\x00\xAA\x6D"
-     "\x03\x00"
-     "\xCE\xA6\x00\x80\xD2\xA6\x00\x80\xAF\x5D\x03\x00\xAE\x5D\x03\x00\xAD\x65"
-     "\x03\x00"
-     "\xAC\x6D\x03\x00\xE3\xBC\x01\x00\xE1\x70\x3F\x00\xE1\x20\x3E\x00\xE3\x94"
-     "\x3F\x00"
-     "\xD6\xA6\x00\x80\xDA\xA6\x00\x80\xDE\xA6\x00\x80\xE2\xA6\x00\x80\xE6\xA6"
-     "\x00\x80"
-     "\xEA\xA6\x00\x80\xEE\xA6\x00\x80\xF2\xA6\x00\x80\xEF\x18\x01\x00\xF6\xA6"
-     "\x00\x80"
-     "\xBF\x38\x13\x00\xEF\xE8\x3E\x00\xFA\xA6\x00\x80\xFE\xA6\x00\x80\x84\x04"
-     "\x05\x00"
-     "\xBE\x04\x05\x00\xE3\x68\x00\x00\x02\xA7\x00\x80\xE1\xA8\x01\x00\x06\xA7"
-     "\x00\x80"
-     "\xEF\xB0\x03\x00\x0A\xA7\x00\x80\x0E\xA7\x00\x80\x12\xA7\x00\x80\x80\x1D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\xDD\x00\x00\x16\xA7\x00\x80\xAA\xC9\x3E\x00\xAB\xC9"
-     "\x3E\x00"
-     "\xA8\x95\x3E\x00\xA9\xCD\x3E\x00\xAE\xF9\x3E\x00\xAF\xF9\x3E\x00\xAC\xCD"
-     "\x3E\x00"
-     "\xAD\xC5\x3E\x00\xBF\xA8\xEF\x00\x1A\xA7\x00\x80\x86\x88\x04\x00\x87\xAC"
-     "\x04\x00"
-     "\x22\xA7\x00\x80\x26\xA7\x00\x80\x2A\xA7\x00\x80\x2E\xA7\x00\x80\xBA\xD9"
-     "\x3D\x00"
-     "\xBB\xD9\x3D\x00\xB8\xC5\x3D\x00\xB9\xD9\x3D\x00\xBE\x91\x3D\x00\xBF\x91"
-     "\x3D\x00"
-     "\xBC\xE1\x3D\x00\xBD\xE1\x3D\x00\xB2\x49\x3E\x00\xB3\x49\x3E\x00\xB0\x49"
-     "\x3E\x00"
-     "\xB1\x49\x3E\x00\xB6\x49\x3E\x00\xB7\x49\x3E\x00\xB4\x49\x3E\x00\xB5\x49"
-     "\x3E\x00"
-     "\xA8\x3D\x06\x00\xA9\x91\x06\x00\xAA\x91\x06\x00\xAB\xC9\x06\x00\xAC\xD9"
-     "\x06\x00"
-     "\xAD\xC5\x06\x00\xAE\xC5\x06\x00\xAF\xF5\x06\x00\x1E\xA7\x00\x80\x32\xA7"
-     "\x00\x80"
-     "\x36\xA7\x00\x80\x3A\xA7\x00\x80\x3E\xA7\x00\x80\x42\xA7\x00\x80\x46\xA7"
-     "\x00\x80"
-     "\x4A\xA7\x00\x80\xB8\x6D\x07\x00\xB9\x71\x07\x00\xBA\x09\x07\x00\xBB\x09"
-     "\x07\x00"
-     "\xBC\x05\x07\x00\xBD\x09\x07\x00\xBE\x3D\x07\x00\xBF\x35\x07\x00\xB0\x8D"
-     "\x06\x00"
-     "\xB1\x75\x07\x00\xB2\x7D\x07\x00\xB3\x75\x07\x00\xB4\x6D\x07\x00\xB5\x55"
-     "\x07\x00"
-     "\xB6\x5D\x07\x00\xB7\x55\x07\x00\xB3\x19\x06\x00\x4E\xA7\x00\x80\x52\xA7"
-     "\x00\x80"
-     "\x56\xA7\x00\x80\x5A\xA7\x00\x80\xB6\x45\x06\x00\xB5\x51\x06\x00\x5E\xA7"
-     "\x00\x80"
-     "\xBB\xB1\x07\x00\xBA\x49\x06\x00\x62\xA7\x00\x80\x66\xA7\x00\x80\xBF\x91"
-     "\x07\x00"
-     "\xBE\xA9\x07\x00\xBD\xA1\x07\x00\xBC\xA9\x07\x00\x6A\xA7\x00\x80\xA3\x5D"
-     "\x06\x00"
-     "\x6E\xA7\x00\x80\x72\xA7\x00\x80\xA6\x01\x06\x00\x76\xA7\x00\x80\x7A\xA7"
-     "\x00\x80"
-     "\xA5\x15\x06\x00\xAA\x0D\x06\x00\xAB\xF5\x07\x00\x7E\xA7\x00\x80\x82\xA7"
-     "\x00\x80"
-     "\xAE\xED\x07\x00\xAF\xD5\x07\x00\xAC\xED\x07\x00\xAD\xE5\x07\x00\x80\xED"
-     "\x01\x00"
-     "\x81\x09\x00\x00\x82\x19\x00\x00\x86\xA7\x00\x80\x8A\xA7\x00\x80\x84\xD0"
-     "\x01\x00"
-     "\xBE\xC4\x01\x00\x8E\xA7\x00\x80\x86\x80\x00\x00\x87\x20\x01\x00\x92\xA7"
-     "\x00\x80"
-     "\x96\xA7\x00\x80\x9A\xA7\x00\x80\x9E\xA7\x00\x80\xA2\xA7\x00\x80\xA6\xA7"
-     "\x00\x80"
-     "\xAA\xA7\x00\x80\xE1\xFC\x05\x00\xAE\xA7\x00\x80\xE3\x14\x04\x00\xB2\xA7"
-     "\x00\x80"
-     "\xB6\xA7\x00\x80\xBA\xA7\x00\x80\xBE\xA7\x00\x80\xC2\xA7\x00\x80\xC6\xA7"
-     "\x00\x80"
-     "\xCA\xA7\x00\x80\xCE\xA7\x00\x80\xD2\xA7\x00\x80\xD6\xA7\x00\x80\xDA\xA7"
-     "\x00\x80"
-     "\xEF\xC8\x04\x00\xA8\x31\x06\x00\xA9\x31\x06\x00\xAA\x31\x06\x00\xAB\x31"
-     "\x06\x00"
-     "\xAC\xB5\x01\x00\xAD\xBD\x01\x00\xAE\xB5\x01\x00\xAF\xAD\x01\x00\x84\xA8"
-     "\x01\x00"
-     "\xDE\xA7\x00\x80\xE2\xA7\x00\x80\xE6\xA7\x00\x80\xEA\xA7\x00\x80\xEE\xA7"
-     "\x00\x80"
-     "\xF2\xA7\x00\x80\xF6\xA7\x00\x80\xB8\x69\x01\x00\xB9\x69\x01\x00\xBA\x09"
-     "\x01\x00"
-     "\xBB\x05\x01\x00\xBC\x1D\x01\x00\xBD\x01\x01\x00\xBE\x01\x01\x00\xBF\x35"
-     "\x01\x00"
-     "\xB0\xD5\x01\x00\xB1\xDD\x01\x00\xB2\xD5\x01\x00\xB3\x69\x01\x00\xB4\x79"
-     "\x01\x00"
-     "\xB5\x79\x01\x00\xB6\x69\x01\x00\xB7\x61\x01\x00\x81\xBD\x03\x00\x80\xBD"
-     "\x03\x00"
-     "\xB3\xB1\x05\x00\x82\x19\x00\x00\xB5\x35\x02\x00\xFA\xA7\x00\x80\xBE\x54"
-     "\x03\x00"
-     "\xB6\x35\x02\x00\x02\xA8\x00\x80\x06\xA8\x00\x80\xBB\x11\x02\x00\xBA\x19"
-     "\x02\x00"
-     "\xBD\xB5\x02\x00\xBC\x09\x02\x00\xBF\xB5\x02\x00\xBE\xB5\x02\x00\x0A\xA8"
-     "\x00\x80"
-     "\xA6\x6D\x02\x00\xA5\x6D\x02\x00\x84\x00\x0C\x00\xA3\xE9\x05\x00\x0E\xA8"
-     "\x00\x80"
-     "\x86\xF8\x0C\x00\x87\x7C\x03\x00\xAF\xED\x02\x00\xAE\xED\x02\x00\xAD\xED"
-     "\x02\x00"
-     "\xAC\x51\x02\x00\xAB\x49\x02\x00\xAA\x41\x02\x00\x12\xA8\x00\x80\x16\xA8"
-     "\x00\x80"
-     "\x1A\xA8\x00\x80\x1E\xA8\x00\x80\x22\xA8\x00\x80\x26\xA8\x00\x80\xE3\x20"
-     "\x01\x00"
-     "\x2A\xA8\x00\x80\xE1\xA0\x01\x00\x2E\xA8\x00\x80\xEF\x40\x02\x00\x32\xA8"
-     "\x00\x80"
-     "\x36\xA8\x00\x80\x3A\xA8\x00\x80\x3E\xA8\x00\x80\x42\xA8\x00\x80\x46\xA8"
-     "\x00\x80"
-     "\x4A\xA8\x00\x80\xAE\xA6\x00\x80\xB3\xF5\x03\x00\x4E\xA8\x00\x80\x52\xA8"
-     "\x00\x80"
-     "\xB6\xED\x03\x00\x56\xA8\x00\x80\x5A\xA8\x00\x80\xB5\xE5\x03\x00\xBA\xB1"
-     "\x03\x00"
-     "\xBB\xB1\x03\x00\x5E\xA8\x00\x80\x62\xA8\x00\x80\xBE\x91\x03\x00\xBF\x91"
-     "\x03\x00"
-     "\xBC\xA1\x03\x00\xBD\xA1\x03\x00\xA8\x29\x02\x00\xA9\x29\x02\x00\xAA\x39"
-     "\x02\x00"
-     "\xAB\x39\x02\x00\xAC\x29\x02\x00\xAD\x29\x02\x00\xAE\x59\x02\x00\xAF\x55"
-     "\x02\x00"
-     "\x80\xCD\x01\x00\x81\x09\x00\x00\x82\x19\x00\x00\x66\xA8\x00\x80\x6A\xA8"
-     "\x00\x80"
-     "\xBE\x7C\x0D\x00\x87\xB4\x0D\x00\x86\x1C\x0C\x00\xB8\x31\x02\x00\xB9\x31"
-     "\x02\x00"
-     "\xBA\x31\x02\x00\xBB\xE9\x02\x00\xBC\xF9\x02\x00\xBD\xF9\x02\x00\xBE\xE9"
-     "\x02\x00"
-     "\xBF\xE9\x02\x00\xB0\x31\x02\x00\xB1\x31\x02\x00\xB2\x31\x02\x00\xB3\x31"
-     "\x02\x00"
-     "\xB4\x11\x02\x00\xB5\x11\x02\x00\xB6\x11\x02\x00\xB7\x11\x02\x00\x72\xA8"
-     "\x00\x80"
-     "\x76\xA8\x00\x80\x7A\xA8\x00\x80\x7E\xA8\x00\x80\x82\xA8\x00\x80\x86\xA8"
-     "\x00\x80"
-     "\x8A\xA8\x00\x80\xEF\x7C\x00\x00\x8E\xA8\x00\x80\xE1\x14\x06\x00\x92\xA8"
-     "\x00\x80"
-     "\xE3\xA8\x00\x00\xEC\x00\x00\x00\x96\xA8\x00\x80\x9A\xA8\x00\x80\x9E\xA8"
-     "\x00\x80"
-     "\xA3\xF1\x02\x00\xA2\xA8\x00\x80\xA6\xA8\x00\x80\xAA\xA8\x00\x80\xAE\xA8"
-     "\x00\x80"
-     "\xA6\xE9\x02\x00\xA5\xE1\x02\x00\xB2\xA8\x00\x80\xAB\xB5\x02\x00\xAA\xB5"
-     "\x02\x00"
-     "\xB6\xA8\x00\x80\xBA\xA8\x00\x80\xAF\x95\x02\x00\xAE\x95\x02\x00\xAD\xA5"
-     "\x02\x00"
-     "\xAC\xA5\x02\x00\xA8\x21\x0E\x00\xA9\x21\x0E\x00\xAA\x49\x0E\x00\xAB\x59"
-     "\x0E\x00"
-     "\xAC\x69\x0E\x00\xAD\x69\x0E\x00\xAE\x99\x0E\x00\xAF\x99\x0E\x00\x6E\xA8"
-     "\x00\x80"
-     "\xBE\xA8\x00\x80\xC2\xA8\x00\x80\xC6\xA8\x00\x80\xCA\xA8\x00\x80\xCE\xA8"
-     "\x00\x80"
-     "\xD2\xA8\x00\x80\xD6\xA8\x00\x80\xB8\xF5\x0E\x00\xB9\xFD\x0E\x00\xBA\xF5"
-     "\x0E\x00"
-     "\xBB\x85\x0E\x00\xBC\x9D\x0E\x00\xBD\x81\x0E\x00\xBE\x81\x0E\x00\xBF\xB5"
-     "\x0E\x00"
-     "\xB0\xE9\x0E\x00\xB1\xE9\x0E\x00\xB2\xFD\x0E\x00\xB3\xF1\x0E\x00\xB4\xD5"
-     "\x0E\x00"
-     "\xB5\xDD\x0E\x00\xB6\xD5\x0E\x00\xB7\xCD\x0E\x00\xB3\xC1\x0E\x00\x82\x15"
-     "\x00\x00"
-     "\x81\xB5\x00\x00\x80\xB5\x00\x00\xDA\xA8\x00\x80\xB6\xE9\x0E\x00\xB5\xE1"
-     "\x0E\x00"
-     "\xBE\x10\x00\x00\xBB\x2D\x0E\x00\xBA\x2D\x0E\x00\x84\x64\x03\x00\xDE\xA8"
-     "\x00\x80"
-     "\xBF\x19\x0E\x00\xBE\x11\x0E\x00\xBD\x25\x0E\x00\xBC\x29\x0E\x00\xE2\xA8"
-     "\x00\x80"
-     "\xA3\x85\x0E\x00\x86\x88\x00\x00\x87\x6C\x03\x00\xA6\xAD\x0E\x00\xE6\xA8"
-     "\x00\x80"
-     "\xEA\xA8\x00\x80\xA5\xA5\x0E\x00\xAA\x69\x0E\x00\xAB\x69\x0E\x00\xEE\xA8"
-     "\x00\x80"
-     "\xF2\xA8\x00\x80\xAE\x55\x0E\x00\xAF\x5D\x0E\x00\xAC\x6D\x0E\x00\xAD\x61"
-     "\x0E\x00"
-     "\xB3\x89\x0E\x00\xF6\xA8\x00\x80\xFA\xA8\x00\x80\xFE\xA8\x00\x80\x02\xA9"
-     "\x00\x80"
-     "\xB6\x81\x0E\x00\xB5\x89\x0E\x00\x06\xA9\x00\x80\xBB\x95\x0E\x00\xBA\x8D"
-     "\x0E\x00"
-     "\x0A\xA9\x00\x80\x0E\xA9\x00\x80\xBF\xF9\x0E\x00\xBE\xF1\x0E\x00\xBD\x85"
-     "\x0E\x00"
-     "\xBC\x85\x0E\x00\x12\xA9\x00\x80\x16\xA9\x00\x80\x1A\xA9\x00\x80\x1E\xA9"
-     "\x00\x80"
-     "\xE1\x10\x0D\x00\x22\xA9\x00\x80\xE3\x04\x0D\x00\x26\xA9\x00\x80\x2A\xA9"
-     "\x00\x80"
-     "\x2E\xA9\x00\x80\xEF\xA4\x0C\x00\x32\xA9\x00\x80\x36\xA9\x00\x80\x3A\xA9"
-     "\x00\x80"
-     "\x3E\xA9\x00\x80\x42\xA9\x00\x80\xA8\x01\x0E\x00\xA9\x01\x0E\x00\xAA\x01"
-     "\x0E\x00"
-     "\xAB\x01\x0E\x00\xAC\x01\x0E\x00\xAD\x01\x0E\x00\xAE\x01\x0E\x00\xAF\x3D"
-     "\x0E\x00"
-     "\x80\xDD\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x46\xA9\x00\x80\x4A\xA9"
-     "\x00\x80"
-     "\x84\x3C\x01\x00\xBE\x74\x00\x00\x52\xA9\x00\x80\xB8\x1D\x0E\x00\xB9\x2D"
-     "\x0E\x00"
-     "\xBA\x25\x0E\x00\xBB\xE9\x01\x00\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\xE9"
-     "\x01\x00"
-     "\xBF\xE9\x01\x00\xB0\x49\x0E\x00\xB1\x51\x0E\x00\xB2\x51\x0E\x00\xB3\x51"
-     "\x0E\x00"
-     "\xB4\x35\x0E\x00\xB5\x3D\x0E\x00\xB6\x35\x0E\x00\xB7\x2D\x0E\x00\xA3\x89"
-     "\x0D\x00"
-     "\x56\xA9\x00\x80\x86\xA8\x04\x00\x87\x3C\x03\x00\x5A\xA9\x00\x80\xA6\x81"
-     "\x0D\x00"
-     "\xA5\x89\x0D\x00\x5E\xA9\x00\x80\xAB\x95\x0D\x00\xAA\x8D\x0D\x00\x62\xA9"
-     "\x00\x80"
-     "\x66\xA9\x00\x80\xAF\xF9\x0D\x00\xAE\xF1\x0D\x00\xAD\x85\x0D\x00\xAC\x85"
-     "\x0D\x00"
-     "\x6A\xA9\x00\x80\xB3\x9D\x02\x00\x84\x48\x03\x00\xBE\x40\x04\x00\xB6\x55"
-     "\x03\x00"
-     "\x6E\xA9\x00\x80\x72\xA9\x00\x80\xB5\xB1\x02\x00\xBA\x71\x03\x00\xBB\x79"
-     "\x03\x00"
-     "\x76\xA9\x00\x80\x7A\xA9\x00\x80\xBE\x31\x03\x00\xBF\x31\x03\x00\xBC\x51"
-     "\x03\x00"
-     "\xBD\x51\x03\x00\xB0\x91\x03\x00\xB1\x99\x03\x00\xB2\xA1\x03\x00\xB3\xA1"
-     "\x03\x00"
-     "\xB4\xD1\x03\x00\xB5\xD1\x03\x00\xB6\xD1\x03\x00\xB7\xD1\x03\x00\xB8\xF5"
-     "\x03\x00"
-     "\xB9\xF9\x03\x00\xBA\xC9\x03\x00\xBB\xC9\x03\x00\xBC\xDD\x03\x00\xBD\xC5"
-     "\x03\x00"
-     "\xBE\xCD\x03\x00\xBF\xC5\x03\x00\x7E\xA9\x00\x80\x82\xA9\x00\x80\x86\xA9"
-     "\x00\x80"
-     "\x8A\xA9\x00\x80\x85\xEC\x1A\x00\xFE\xA7\x00\x80\x8E\xA9\x00\x80\x92\xA9"
-     "\x00\x80"
-     "\xA8\x21\x02\x00\xA9\x31\x02\x00\xAA\x31\x02\x00\xAB\x05\x02\x00\xAC\x1D"
-     "\x02\x00"
-     "\xAD\xF1\x03\x00\xAE\xF1\x03\x00\xAF\xF1\x03\x00\x96\xA9\x00\x80\x9A\xA9"
-     "\x00\x80"
-     "\x9E\xA9\x00\x80\xA2\xA9\x00\x80\xEF\x50\x00\x00\xA6\xA9\x00\x80\xAA\xA9"
-     "\x00\x80"
-     "\xAE\xA9\x00\x80\xE3\x84\x00\x00\xB2\xA9\x00\x80\xE1\xF8\x01\x00\xB6\xA9"
-     "\x00\x80"
-     "\x80\x15\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\xBA\xA9\x00\x80\xA3\x99"
-     "\x03\x00"
-     "\xC2\xA9\x00\x80\x86\x68\x04\x00\x87\x60\x05\x00\xC6\xA9\x00\x80\xA6\x51"
-     "\x02\x00"
-     "\xA5\xB5\x03\x00\xCA\xA9\x00\x80\xAB\x7D\x02\x00\xAA\x75\x02\x00\xCE\xA9"
-     "\x00\x80"
-     "\xD2\xA9\x00\x80\xAF\x35\x02\x00\xAE\x35\x02\x00\xAD\x55\x02\x00\xAC\x55"
-     "\x02\x00"
-     "\xD6\xA9\x00\x80\xDA\xA9\x00\x80\xDE\xA9\x00\x80\xE2\xA9\x00\x80\xE6\xA9"
-     "\x00\x80"
-     "\xEA\xA9\x00\x80\xEE\xA9\x00\x80\xEF\xF8\x01\x00\xBE\x6C\x04\x00\xE1\x80"
-     "\x0E\x00"
-     "\xF2\xA9\x00\x80\xE3\x14\x01\x00\xF6\xA9\x00\x80\xFA\xA9\x00\x80\xFE\xA9"
-     "\x00\x80"
-     "\x02\xAA\x00\x80\x06\xAA\x00\x80\x0A\xAA\x00\x80\x0E\xAA\x00\x80\xB3\xDD"
-     "\x01\x00"
-     "\x12\xAA\x00\x80\xB5\xFD\x01\x00\xB6\xF5\x01\x00\x16\xAA\x00\x80\x1A\xAA"
-     "\x00\x80"
-     "\x1E\xAA\x00\x80\xBA\xB1\x01\x00\xBB\x85\x01\x00\xBC\x9D\x01\x00\xBD\x35"
-     "\x01\x00"
-     "\xBE\x3D\x01\x00\xBF\x35\x01\x00\xA8\xBD\x05\x00\xA9\x01\x06\x00\xAA\x19"
-     "\x06\x00"
-     "\xAB\x11\x06\x00\xAC\x39\x06\x00\xAD\x39\x06\x00\xAE\x59\x06\x00\xAF\x49"
-     "\x06\x00"
-     "\xBE\xA9\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D\x00\x00\x22\xAA"
-     "\x00\x80"
-     "\x26\xAA\x00\x80\x2A\xAA\x00\x80\x2E\xAA\x00\x80\xB8\x9D\x06\x00\xB9\x45"
-     "\x07\x00"
-     "\xBA\x4D\x07\x00\xBB\x45\x07\x00\xBC\x5D\x07\x00\xBD\x41\x07\x00\xBE\x41"
-     "\x07\x00"
-     "\xBF\x41\x07\x00\xB0\x39\x06\x00\xB1\x0D\x06\x00\xB2\x05\x06\x00\xB3\xAD"
-     "\x06\x00"
-     "\xB4\xB5\x06\x00\xB5\xBD\x06\x00\xB6\xB5\x06\x00\xB7\xA5\x06\x00\xA3\x91"
-     "\x06\x00"
-     "\x84\x8C\x02\x00\x86\x28\x00\x00\x87\xC0\x03\x00\x32\xAA\x00\x80\xA6\xB9"
-     "\x06\x00"
-     "\xA5\xB1\x06\x00\x36\xAA\x00\x80\xAB\xC9\x06\x00\xAA\xFD\x06\x00\x3A\xAA"
-     "\x00\x80"
-     "\x3E\xAA\x00\x80\xAF\x79\x06\x00\xAE\x71\x06\x00\xAD\x79\x06\x00\xAC\xD1"
-     "\x06\x00"
-     "\x42\xAA\x00\x80\xB3\x99\x07\x00\x46\xAA\x00\x80\x4A\xAA\x00\x80\xB6\x51"
-     "\x07\x00"
-     "\x4E\xAA\x00\x80\x52\xAA\x00\x80\xB5\xB1\x07\x00\xBA\x6D\x07\x00\xBB\x75"
-     "\x07\x00"
-     "\x56\xAA\x00\x80\x5A\xAA\x00\x80\xBE\x59\x07\x00\xBF\x45\x07\x00\xBC\x6D"
-     "\x07\x00"
-     "\xBD\x65\x07\x00\x5E\xAA\x00\x80\x62\xAA\x00\x80\x66\xAA\x00\x80\x6A\xAA"
-     "\x00\x80"
-     "\x6E\xAA\x00\x80\xEF\xA0\x05\x00\x72\xAA\x00\x80\x76\xAA\x00\x80\x7A\xAA"
-     "\x00\x80"
-     "\xE3\x7C\x05\x00\x7E\xAA\x00\x80\xE1\x3C\x05\x00\x82\xAA\x00\x80\x86\xAA"
-     "\x00\x80"
-     "\x8A\xAA\x00\x80\x8E\xAA\x00\x80\xA3\x1D\x06\x00\x82\x2D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x92\xAA\x00\x80\xA6\xD5\x06\x00\xA5\x35\x06\x00\x96\xAA"
-     "\x00\x80"
-     "\xAB\xF1\x06\x00\xAA\xE9\x06\x00\x9A\xAA\x00\x80\x84\x20\x01\x00\xAF\xC1"
-     "\x06\x00"
-     "\xAE\xDD\x06\x00\xAD\xE1\x06\x00\xAC\xE9\x06\x00\xA8\xC5\x06\x00\xA9\xD5"
-     "\x06\x00"
-     "\xAA\xD5\x06\x00\xAB\xE5\x06\x00\xAC\xFD\x06\x00\xAD\x15\x06\x00\xAE\x1D"
-     "\x06\x00"
-     "\xAF\x15\x06\x00\xBE\xE8\x01\x00\xA2\xAA\x00\x80\x86\x08\x00\x00\x87\x88"
-     "\x00\x00"
-     "\xA6\xAA\x00\x80\xAA\xAA\x00\x80\xAE\xAA\x00\x80\xB2\xAA\x00\x80\xB8\x7D"
-     "\x06\x00"
-     "\xB9\x05\x06\x00\xBA\x0D\x06\x00\xBB\x05\x06\x00\xBC\x01\x06\x00\xBD\x09"
-     "\x06\x00"
-     "\xBE\x39\x06\x00\xBF\x39\x06\x00\xB0\x6D\x06\x00\xB1\x75\x06\x00\xB2\x7D"
-     "\x06\x00"
-     "\xB3\x75\x06\x00\xB4\x59\x06\x00\xB5\x45\x06\x00\xB6\x4D\x06\x00\xB7\x45"
-     "\x06\x00"
-     "\xA8\x91\x02\x00\xA9\x99\x02\x00\xAA\xA1\x02\x00\xAB\xA1\x02\x00\xAC\xD1"
-     "\x02\x00"
-     "\xAD\xDD\x02\x00\xAE\xD5\x02\x00\xAF\xC9\x02\x00\xB6\xAA\x00\x80\xBA\xAA"
-     "\x00\x80"
-     "\xBE\xAA\x00\x80\xBF\x24\xC5\x00\xC2\xAA\x00\x80\xC6\xAA\x00\x80\xCA\xAA"
-     "\x00\x80"
-     "\xCE\xAA\x00\x80\xB8\x59\x03\x00\xB9\x59\x03\x00\xBA\x69\x03\x00\xBB\x65"
-     "\x03\x00"
-     "\xBC\x61\x03\x00\xBD\x61\x03\x00\xBE\x61\x03\x00\xBF\x61\x03\x00\xB0\xB9"
-     "\x02\x00"
-     "\xB1\x8D\x02\x00\xB2\x85\x02\x00\xB3\x6D\x03\x00\xB4\x75\x03\x00\xB5\x7D"
-     "\x03\x00"
-     "\xB6\x75\x03\x00\xB7\x6D\x03\x00\xB3\x29\x02\x00\xD2\xAA\x00\x80\xD6\xAA"
-     "\x00\x80"
-     "\xDA\xAA\x00\x80\x84\x84\x03\x00\xB6\x5D\x02\x00\xB5\x59\x02\x00\xDE\xAA"
-     "\x00\x80"
-     "\xBB\x49\x02\x00\xBA\x7D\x02\x00\x4E\xA9\x00\x80\xE2\xAA\x00\x80\xBF\xB9"
-     "\x03\x00"
-     "\xBE\xB1\x03\x00\xBD\xB9\x03\x00\xBC\x51\x02\x00\xE6\xAA\x00\x80\xA3\x6D"
-     "\x02\x00"
-     "\xEA\xAA\x00\x80\xEE\xAA\x00\x80\xA6\x19\x02\x00\xF2\xAA\x00\x80\xF6\xAA"
-     "\x00\x80"
-     "\xA5\x1D\x02\x00\xAA\x39\x02\x00\xAB\x0D\x02\x00\xFA\xAA\x00\x80\xFE\xAA"
-     "\x00\x80"
-     "\xAE\xF5\x03\x00\xAF\xFD\x03\x00\xAC\x15\x02\x00\xAD\xFD\x03\x00\xA8\xA5"
-     "\x02\x00"
-     "\xA9\xB5\x02\x00\xAA\xBD\x02\x00\xAB\xB5\x02\x00\xAC\xAD\x02\x00\xAD\x35"
-     "\x01\x00"
-     "\xAE\x3D\x01\x00\xAF\x35\x01\x00\x81\x11\x00\x00\x80\x0D\x00\x00\x02\xAB"
-     "\x00\x80"
-     "\x82\x19\x00\x00\x06\xAB\x00\x80\x0A\xAB\x00\x80\xBE\xB4\x1D\x00\x12\xAB"
-     "\x00\x80"
-     "\xB8\xED\x01\x00\xB9\x85\x01\x00\xBA\x81\x01\x00\xBB\x81\x01\x00\xBC\x85"
-     "\x01\x00"
-     "\xBD\x8D\x01\x00\xBE\xB1\x01\x00\xBF\xB1\x01\x00\xB0\x51\x01\x00\xB1\x51"
-     "\x01\x00"
-     "\xB2\x51\x01\x00\xB3\x51\x01\x00\xB4\xF5\x01\x00\xB5\xFD\x01\x00\xB6\xF5"
-     "\x01\x00"
-     "\xB7\xE1\x01\x00\xEF\xD8\x00\x00\x16\xAB\x00\x80\x86\xA8\x1D\x00\x87\xDC"
-     "\x1D\x00"
-     "\x1A\xAB\x00\x80\x1E\xAB\x00\x80\x22\xAB\x00\x80\xEF\xC4\x06\x00\x26\xAB"
-     "\x00\x80"
-     "\xE1\xF4\x06\x00\x2A\xAB\x00\x80\xE3\x38\x01\x00\xE3\xCC\x00\x00\x2E\xAB"
-     "\x00\x80"
-     "\xE1\x28\x01\x00\x32\xAB\x00\x80\xB5\x7D\x02\x00\x36\xAB\x00\x80\x3A\xAB"
-     "\x00\x80"
-     "\xB6\xED\x01\x00\x3E\xAB\x00\x80\x42\xAB\x00\x80\xB3\x6D\x02\x00\x46\xAB"
-     "\x00\x80"
-     "\xBD\xDD\x01\x00\xBC\xDD\x01\x00\xBF\x61\x01\x00\xBE\x69\x01\x00\x4A\xAB"
-     "\x00\x80"
-     "\x4E\xAB\x00\x80\xBB\xCD\x01\x00\xBA\xCD\x01\x00\xA3\xB5\x1D\x00\x0E\xAB"
-     "\x00\x80"
-     "\x52\xAB\x00\x80\x56\xAB\x00\x80\x5A\xAB\x00\x80\xA6\x35\x1E\x00\xA5\xA5"
-     "\x1D\x00"
-     "\x5E\xAB\x00\x80\xAB\x15\x1E\x00\xAA\x15\x1E\x00\x62\xAB\x00\x80\x66\xAB"
-     "\x00\x80"
-     "\xAF\xB9\x1E\x00\xAE\xB1\x1E\x00\xAD\x05\x1E\x00\xAC\x05\x1E\x00\x6A\xAB"
-     "\x00\x80"
-     "\x6E\xAB\x00\x80\x72\xAB\x00\x80\x76\xAB\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x7A\xAB\x00\x80\x84\xEC\x03\x00\xBE\x68\x03\x00\x7E\xAB"
-     "\x00\x80"
-     "\x82\xAB\x00\x80\x86\xD0\x01\x00\x87\xB4\x00\x00\x86\xAB\x00\x80\x8A\xAB"
-     "\x00\x80"
-     "\xA8\xBD\x1F\x00\xA9\x01\x1F\x00\xAA\x19\x1F\x00\xAB\x19\x1F\x00\xAC\x09"
-     "\x1F\x00"
-     "\xAD\x35\x1F\x00\xAE\x31\x1F\x00\xAF\x29\x1F\x00\x8E\xAB\x00\x80\x92\xAB"
-     "\x00\x80"
-     "\x96\xAB\x00\x80\x9A\xAB\x00\x80\x9E\xAB\x00\x80\xA2\xAB\x00\x80\xA6\xAB"
-     "\x00\x80"
-     "\xAA\xAB\x00\x80\xB8\xC9\x00\x00\xB9\xC9\x00\x00\xBA\xD9\x00\x00\xBB\xD1"
-     "\x00\x00"
-     "\xBC\xF9\x00\x00\xBD\xF9\x00\x00\xBE\x99\x00\x00\xBF\x99\x00\x00\xB0\x5D"
-     "\x1F\x00"
-     "\xB1\x21\x1F\x00\xB2\x21\x1F\x00\xB3\x3D\x1F\x00\xB4\x25\x1F\x00\xB5\x29"
-     "\x1F\x00"
-     "\xB6\x09\x1F\x00\xB7\x09\x1F\x00\xEF\xC0\x1C\x00\xE1\xC0\x19\x00\xAE\xAB"
-     "\x00\x80"
-     "\xE3\xD0\x09\x00\xB2\xAB\x00\x80\xB6\xAB\x00\x80\xBA\xAB\x00\x80\xBE\xAB"
-     "\x00\x80"
-     "\xA3\x25\x01\x00\xA2\x99\x03\x00\xA1\x15\x1D\x00\xA0\x0D\x1D\x00\xE3\xB0"
-     "\x1D\x00"
-     "\xC2\xAB\x00\x80\xE1\x88\x1E\x00\xEF\x60\x10\x00\xB3\xD1\x1E\x00\xC6\xAB"
-     "\x00\x80"
-     "\xCA\xAB\x00\x80\xCE\xAB\x00\x80\xD2\xAB\x00\x80\xB6\xF5\x1E\x00\xB5\xC1"
-     "\x1E\x00"
-     "\xD6\xAB\x00\x80\xBB\xA5\x1E\x00\xBA\xD9\x1E\x00\xDA\xAB\x00\x80\xDE\xAB"
-     "\x00\x80"
-     "\xBF\xAD\x1E\x00\xBE\xA1\x1E\x00\xBD\xA9\x1E\x00\xBC\xB5\x1E\x00\xA3\x11"
-     "\x1E\x00"
-     "\xE2\xAB\x00\x80\xE6\xAB\x00\x80\xEA\xAB\x00\x80\x8F\x19\x24\x00\xA6\x35"
-     "\x1E\x00"
-     "\xA5\x01\x1E\x00\xBE\x10\x07\x00\xAB\x65\x1E\x00\xAA\x19\x1E\x00\x84\x00"
-     "\x04\x00"
-     "\xF2\xAB\x00\x80\xAF\x6D\x1E\x00\xAE\x61\x1E\x00\xAD\x69\x1E\x00\xAC\x75"
-     "\x1E\x00"
-     "\x9B\xDD\x14\x00\x9A\xE5\x15\x00\x99\x09\x17\x00\x98\x11\x11\x00\x9F\x79"
-     "\x1F\x00"
-     "\x9E\x79\x19\x00\x9D\xF9\x19\x00\x9C\x09\x1B\x00\x93\x61\x2D\x00\x92\x91"
-     "\x28\x00"
-     "\x91\x6D\x29\x00\x90\x6D\x29\x00\x97\x7D\x11\x00\x96\x75\x11\x00\x95\xE9"
-     "\x2C\x00"
-     "\x94\x79\x2D\x00\x86\x20\x04\x00\x87\x94\x03\x00\x82\x81\x27\x00\x83\x8D"
-     "\x2A\x00"
-     "\x84\x95\x2A\x00\x85\x61\x2B\x00\x86\x99\x2E\x00\x87\x99\x2E\x00\x84\xA8"
-     "\x04\x00"
-     "\xF6\xAB\x00\x80\x8A\x39\x2F\x00\x8B\xAD\x12\x00\x8C\xBD\x12\x00\x8D\xE9"
-     "\x12\x00"
-     "\x8E\xF9\x14\x00\x8F\x1D\x16\x00\xFA\xAB\x00\x80\xFE\xAB\x00\x80\x92\x55"
-     "\x17\x00"
-     "\x93\x99\x18\x00\x94\x71\x1A\x00\x95\xF9\x1A\x00\x96\xF1\x1C\x00\x97\x7D"
-     "\x1E\x00"
-     "\x02\xAC\x00\x80\x06\xAC\x00\x80\x9A\x55\x1E\x00\x9B\xD1\x00\x00\x9C\x75"
-     "\x02\x00"
-     "\x0A\xAC\x00\x80\x0E\xAC\x00\x80\x12\xAC\x00\x80\xAF\x85\xF6\x00\xAE\x01"
-     "\xF4\x00"
-     "\xAD\xF1\xF2\x00\xAC\x89\xF2\x00\xAB\x01\xF0\x00\xAA\x85\x0E\x00\xA9\x9D"
-     "\x0E\x00"
-     "\xA8\x9D\x0E\x00\xA7\x75\x0A\x00\xA6\x85\x0A\x00\xA5\x01\x08\x00\xA4\x01"
-     "\x06\x00"
-     "\xA3\xB9\x06\x00\xA2\xB1\x06\x00\xA1\xD1\x02\x00\x9E\xAA\x00\x80\x80\x45"
-     "\x03\x00"
-     "\x81\x4D\x03\x00\x82\x45\x03\x00\x16\xAC\x00\x80\x1A\xAC\x00\x80\x1E\xAC"
-     "\x00\x80"
-     "\x22\xAC\x00\x80\x26\xAC\x00\x80\x2A\xAC\x00\x80\x2E\xAC\x00\x80\xB5\x01"
-     "\xFC\x00"
-     "\xB4\xC1\xFB\x00\xB3\xF5\xFA\x00\xB2\xF5\xFA\x00\xB1\xBD\xF7\x00\xB0\xF5"
-     "\xF6\x00"
-     "\xA8\x6D\x03\x00\xA9\x7D\x03\x00\xAA\x75\x03\x00\xAB\xAD\x03\x00\xAC\xB5"
-     "\x03\x00"
-     "\xAD\xBD\x03\x00\xAE\xB5\x03\x00\xAF\xAD\x03\x00\x32\xAC\x00\x80\x36\xAC"
-     "\x00\x80"
-     "\x3A\xAC\x00\x80\x3E\xAC\x00\x80\x42\xAC\x00\x80\x46\xAC\x00\x80\x4A\xAC"
-     "\x00\x80"
-     "\x4E\xAC\x00\x80\xB8\x75\x02\x00\xB9\x01\x02\x00\xBA\x01\x02\x00\xBB\x01"
-     "\x02\x00"
-     "\xBC\x01\x02\x00\xBD\x09\x02\x00\xBE\x31\x02\x00\xBF\x31\x02\x00\xB0\xD5"
-     "\x03\x00"
-     "\xB1\xDD\x03\x00\xB2\xD5\x03\x00\xB3\x65\x02\x00\xB4\x61\x02\x00\xB5\x61"
-     "\x02\x00"
-     "\xB6\x61\x02\x00\xB7\x61\x02\x00\xE1\x10\x1E\x00\x52\xAC\x00\x80\xE3\xF8"
-     "\x01\x00"
-     "\xBE\x00\x04\x00\x80\x15\x00\x00\x81\x09\x00\x00\x82\x39\x00\x00\x56\xAC"
-     "\x00\x80"
-     "\x5A\xAC\x00\x80\x62\xAC\x00\x80\x86\x88\x04\x00\x87\x0C\x05\x00\x66\xAC"
-     "\x00\x80"
-     "\x6A\xAC\x00\x80\xEF\x24\x01\x00\x6E\xAC\x00\x80\x72\xAC\x00\x80\x76\xAC"
-     "\x00\x80"
-     "\x7A\xAC\x00\x80\x7E\xAC\x00\x80\x82\xAC\x00\x80\xEF\xB0\x00\x00\x86\xAC"
-     "\x00\x80"
-     "\x8A\xAC\x00\x80\x8E\xAC\x00\x80\xE3\x5C\x00\x00\x92\xAC\x00\x80\xE1\x84"
-     "\x01\x00"
-     "\x96\xAC\x00\x80\x9A\xAC\x00\x80\x9E\xAC\x00\x80\xA2\xAC\x00\x80\xA3\x7D"
-     "\x04\x00"
-     "\xEE\xAB\x00\x80\x5E\xAC\x00\x80\xA6\xAC\x00\x80\xAA\xAC\x00\x80\xA6\xBD"
-     "\x04\x00"
-     "\xA5\xBD\x04\x00\xAE\xAC\x00\x80\xAB\x99\x04\x00\xAA\x91\x04\x00\xB2\xAC"
-     "\x00\x80"
-     "\xB6\xAC\x00\x80\xAF\x25\x05\x00\xAE\x25\x05\x00\xAD\x31\x05\x00\xAC\x81"
-     "\x04\x00"
-     "\xBA\xAC\x00\x80\xB3\x05\x06\x00\xBE\xAC\x00\x80\xC2\xAC\x00\x80\xB6\x35"
-     "\x06\x00"
-     "\xC6\xAC\x00\x80\xCA\xAC\x00\x80\xB5\x3D\x06\x00\xBA\x1D\x06\x00\xBB\xA9"
-     "\x06\x00"
-     "\x84\xE0\x03\x00\xBE\xE0\x00\x00\xBE\xA9\x06\x00\xBF\xA9\x06\x00\xBC\xA9"
-     "\x06\x00"
-     "\xBD\xA9\x06\x00\xA8\x2D\x05\x00\xA9\xA1\x06\x00\xAA\xC9\x06\x00\xAB\xC9"
-     "\x06\x00"
-     "\xAC\xF1\x06\x00\xAD\xF1\x06\x00\xAE\xED\x06\x00\xAF\xE5\x06\x00\x80\xED"
-     "\x00\x00"
-     "\x81\xF1\x00\x00\x82\xF1\x00\x00\xCE\xAC\x00\x80\x86\x90\x00\x00\x87\xF0"
-     "\x00\x00"
-     "\xD2\xAC\x00\x80\xD6\xAC\x00\x80\xB8\xE1\x07\x00\xB9\xE1\x07\x00\xBA\xE1"
-     "\x07\x00"
-     "\xBB\xE1\x07\x00\xBC\xE1\x07\x00\xBD\xE1\x07\x00\xBE\xE1\x07\x00\xBF\xE1"
-     "\x07\x00"
-     "\xB0\x9D\x06\x00\xB1\x29\x06\x00\xB2\x39\x06\x00\xB3\x39\x06\x00\xB4\x29"
-     "\x06\x00"
-     "\xB5\x29\x06\x00\xB6\x19\x06\x00\xB7\x19\x06\x00\xA3\x41\x07\x00\xDA\xAC"
-     "\x00\x80"
-     "\xDE\xAC\x00\x80\xE2\xAC\x00\x80\xE6\xAC\x00\x80\xA6\x71\x07\x00\xA5\x79"
-     "\x07\x00"
-     "\xEA\xAC\x00\x80\xAB\xED\x07\x00\xAA\x59\x07\x00\xEE\xAC\x00\x80\xF2\xAC"
-     "\x00\x80"
-     "\xAF\xED\x07\x00\xAE\xED\x07\x00\xAD\xED\x07\x00\xAC\xED\x07\x00\xF6\xAC"
-     "\x00\x80"
-     "\xB3\xF1\x06\x00\xFA\xAC\x00\x80\xFE\xAC\x00\x80\xB6\x95\x06\x00\x02\xAD"
-     "\x00\x80"
-     "\x06\xAD\x00\x80\xB5\x9D\x06\x00\xBA\xB1\x06\x00\xBB\xBD\x06\x00\x0A\xAD"
-     "\x00\x80"
-     "\x0E\xAD\x00\x80\xBE\x99\x06\x00\xBF\x85\x06\x00\xBC\xA5\x06\x00\xBD\x9D"
-     "\x06\x00"
-     "\xA8\x25\x06\x00\xA9\x35\x06\x00\xAA\x3D\x06\x00\xAB\x31\x06\x00\xAC\x51"
-     "\x06\x00"
-     "\xAD\x51\x06\x00\xAE\x51\x06\x00\xAF\x51\x06\x00\x12\xAD\x00\x80\x16\xAD"
-     "\x00\x80"
-     "\x84\x2C\x01\x00\x1A\xAD\x00\x80\x1E\xAD\x00\x80\x22\xAD\x00\x80\x26\xAD"
-     "\x00\x80"
-     "\x2A\xAD\x00\x80\xB8\x3D\x06\x00\xB9\xC9\x01\x00\xBA\xD9\x01\x00\xBB\xD1"
-     "\x01\x00"
-     "\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\x99\x01\x00\xBF\x95\x01\x00\xB0\x31"
-     "\x06\x00"
-     "\xB1\x3D\x06\x00\xB2\x35\x06\x00\xB3\x0D\x06\x00\xB4\x19\x06\x00\xB5\x05"
-     "\x06\x00"
-     "\xB6\x0D\x06\x00\xB7\x05\x06\x00\xA3\xB5\x05\x00\x2E\xAD\x00\x80\x32\xAD"
-     "\x00\x80"
-     "\x36\xAD\x00\x80\x3A\xAD\x00\x80\xA6\xD1\x05\x00\xA5\xD9\x05\x00\xBE\x58"
-     "\x03\x00"
-     "\xAB\xF9\x05\x00\xAA\xF5\x05\x00\x86\x88\x0C\x00\x87\xBC\x03\x00\xAF\xC1"
-     "\x05\x00"
-     "\xAE\xDD\x05\x00\xAD\xD9\x05\x00\xAC\xE1\x05\x00\x81\x4D\x00\x00\x80\x4D"
-     "\x00\x00"
-     "\xB3\x4D\x03\x00\x82\x5D\x00\x00\xB5\x5D\x03\x00\x42\xAD\x00\x80\x46\xAD"
-     "\x00\x80"
-     "\xB6\x4D\x03\x00\x4A\xAD\x00\x80\x4E\xAD\x00\x80\xBB\x29\x03\x00\xBA\x41"
-     "\x03\x00"
-     "\xBD\x3D\x03\x00\xBC\x3D\x03\x00\xBF\x21\x03\x00\xBE\x29\x03\x00\xA8\x39"
-     "\x02\x00"
-     "\xA9\x39\x02\x00\xAA\x8D\x03\x00\xAB\x85\x03\x00\xAC\x9D\x03\x00\xAD\x85"
-     "\x03\x00"
-     "\xAE\x85\x03\x00\xAF\xB5\x03\x00\x84\xEC\x0C\x00\x52\xAD\x00\x80\xBE\xE0"
-     "\x0C\x00"
-     "\x56\xAD\x00\x80\x5A\xAD\x00\x80\x5E\xAD\x00\x80\x62\xAD\x00\x80\x66\xAD"
-     "\x00\x80"
-     "\xB8\x61\x03\x00\xB9\x61\x03\x00\xBA\x65\x03\x00\xBB\x7D\x03\x00\xBC\x65"
-     "\x03\x00"
-     "\xBD\x6D\x03\x00\xBE\x65\x03\x00\xBF\xD9\x00\x00\xB0\xCD\x03\x00\xB1\xA5"
-     "\x03\x00"
-     "\xB2\xA1\x03\x00\xB3\xA1\x03\x00\xB4\xA5\x03\x00\xB5\xAD\x03\x00\xB6\x91"
-     "\x03\x00"
-     "\xB7\x91\x03\x00\x6A\xAD\x00\x80\x6E\xAD\x00\x80\x72\xAD\x00\x80\x76\xAD"
-     "\x00\x80"
-     "\xEF\xF4\x03\x00\x7A\xAD\x00\x80\x7E\xAD\x00\x80\x82\xAD\x00\x80\xE3\x98"
-     "\x00\x00"
-     "\x86\xAD\x00\x80\xE1\x80\x01\x00\x8A\xAD\x00\x80\x8E\xAD\x00\x80\x92\xAD"
-     "\x00\x80"
-     "\x96\xAD\x00\x80\x9A\xAD\x00\x80\xBA\xF5\x07\x00\xBB\xFD\x07\x00\x9E\xAD"
-     "\x00\x80"
-     "\xA2\xAD\x00\x80\xBE\x41\x06\x00\xBF\x41\x06\x00\xBC\xE5\x07\x00\xBD\x55"
-     "\x06\x00"
-     "\xA6\xAD\x00\x80\xB3\x19\x07\x00\xAA\xAD\x00\x80\xB2\xAD\x00\x80\xB6\xD9"
-     "\x07\x00"
-     "\xB6\xAD\x00\x80\xBA\xAD\x00\x80\xB5\xD9\x07\x00\xE1\x98\x07\x00\xE3\xFC"
-     "\x01\x00"
-     "\xE3\xB8\x07\x00\xE1\x3C\x06\x00\x80\xDD\x00\x00\x81\xF5\x00\x00\x82\xF1"
-     "\x00\x00"
-     "\xBE\xAD\x00\x80\xC2\xAD\x00\x80\xC6\xAD\x00\x80\xCA\xAD\x00\x80\xCE\xAD"
-     "\x00\x80"
-     "\xD2\xAD\x00\x80\xEF\x20\x01\x00\xEF\xCC\x06\x00\xD6\xAD\x00\x80\x86\xE0"
-     "\x0C\x00"
-     "\x87\x0C\x0D\x00\xA3\xCD\x02\x00\xDA\xAD\x00\x80\xA5\xDD\x02\x00\xDE\xAD"
-     "\x00\x80"
-     "\xE2\xAD\x00\x80\xA6\xCD\x02\x00\xE6\xAD\x00\x80\xEA\xAD\x00\x80\xAB\xA9"
-     "\x02\x00"
-     "\xAA\xC1\x02\x00\xAD\xBD\x02\x00\xAC\xBD\x02\x00\xAF\xA1\x02\x00\xAE\xA9"
-     "\x02\x00"
-     "\xA8\x55\x0E\x00\xA9\x5D\x0E\x00\xAA\x69\x0E\x00\xAB\x69\x0E\x00\xAC\xB5"
-     "\x0E\x00"
-     "\xAD\xB9\x0E\x00\xAE\xA9\x0E\x00\xAF\xA9\x0E\x00\xAE\xAD\x00\x80\xEE\xAD"
-     "\x00\x80"
-     "\xF2\xAD\x00\x80\xF6\xAD\x00\x80\xFA\xAD\x00\x80\xFE\xAD\x00\x80\x02\xAE"
-     "\x00\x80"
-     "\x06\xAE\x00\x80\xB8\x5D\x0F\x00\xB9\x61\x0F\x00\xBA\x65\x0F\x00\xBB\x7D"
-     "\x0F\x00"
-     "\xBC\x65\x0F\x00\xBD\x6D\x0F\x00\xBE\x65\x0F\x00\xBF\x15\x0F\x00\xB0\xD9"
-     "\x0E\x00"
-     "\xB1\xD9\x0E\x00\xB2\x69\x0F\x00\xB3\x69\x0F\x00\xB4\x79\x0F\x00\xB5\x79"
-     "\x0F\x00"
-     "\xB6\x6D\x0F\x00\xB7\x65\x0F\x00\xB3\xE1\x0E\x00\x0A\xAE\x00\x80\x0E\xAE"
-     "\x00\x80"
-     "\x12\xAE\x00\x80\x16\xAE\x00\x80\xB6\x39\x0E\x00\xB5\x3D\x0E\x00\x1A\xAE"
-     "\x00\x80"
-     "\xBB\x1D\x0E\x00\xBA\x1D\x0E\x00\x84\x24\x03\x00\xBE\x2C\x00\x00\xBF\xBD"
-     "\x0F\x00"
-     "\xBE\xBD\x0F\x00\xBD\xBD\x0F\x00\xBC\xBD\x0F\x00\x82\x5D\x00\x00\xA3\xA5"
-     "\x0E\x00"
-     "\x80\x45\x00\x00\x81\x5D\x00\x00\xA6\x7D\x0E\x00\x1E\xAE\x00\x80\x22\xAE"
-     "\x00\x80"
-     "\xA5\x79\x0E\x00\xAA\x59\x0E\x00\xAB\x59\x0E\x00\x86\x00\x04\x00\x87\x5C"
-     "\x03\x00"
-     "\xAE\xF9\x0F\x00\xAF\xF9\x0F\x00\xAC\xF9\x0F\x00\xAD\xF9\x0F\x00\x26\xAE"
-     "\x00\x80"
-     "\x2A\xAE\x00\x80\x2E\xAE\x00\x80\x32\xAE\x00\x80\x36\xAE\x00\x80\xEF\x44"
-     "\x0D\x00"
-     "\x3A\xAE\x00\x80\x3E\xAE\x00\x80\x42\xAE\x00\x80\xE3\x94\x0D\x00\x46\xAE"
-     "\x00\x80"
-     "\xE1\x1C\x0D\x00\x4A\xAE\x00\x80\x4E\xAE\x00\x80\x52\xAE\x00\x80\x56\xAE"
-     "\x00\x80"
-     "\xB3\x39\x0E\x00\x5A\xAE\x00\x80\x5E\xAE\x00\x80\x62\xAE\x00\x80\x66\xAE"
-     "\x00\x80"
-     "\xB6\xF1\x0E\x00\xB5\xF9\x0E\x00\x6A\xAE\x00\x80\xBB\xDD\x0E\x00\xBA\xD5"
-     "\x0E\x00"
-     "\x6E\xAE\x00\x80\x72\xAE\x00\x80\xBF\x91\x0E\x00\xBE\xAD\x0E\x00\xBD\xAD"
-     "\x0E\x00"
-     "\xBC\xB1\x0E\x00\xA8\xED\x0E\x00\xA9\xF1\x0E\x00\xAA\xF1\x0E\x00\xAB\x11"
-     "\x0E\x00"
-     "\xAC\x31\x0E\x00\xAD\x31\x0E\x00\xAE\x31\x0E\x00\xAF\x31\x0E\x00\x76\xAE"
-     "\x00\x80"
-     "\x7A\xAE\x00\x80\x7E\xAE\x00\x80\x82\xAE\x00\x80\x80\x1D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\xB9\x03\x00\x86\xAE\x00\x80\xB8\xE9\x01\x00\xB9\xE9\x01\x00\xBA\xF9"
-     "\x01\x00"
-     "\xBB\xF9\x01\x00\xBC\xE9\x01\x00\xBD\xE9\x01\x00\xBE\x5D\x01\x00\xBF\x55"
-     "\x01\x00"
-     "\xB0\x51\x0E\x00\xB1\x59\x0E\x00\xB2\x25\x0E\x00\xB3\x39\x0E\x00\xB4\x29"
-     "\x0E\x00"
-     "\xB5\x15\x0E\x00\xB6\x1D\x0E\x00\xB7\xD9\x01\x00\x8A\xAE\x00\x80\x84\x9C"
-     "\x02\x00"
-     "\xBE\x04\x02\x00\xA3\x75\x0D\x00\x92\xAE\x00\x80\xA5\xB5\x0D\x00\xA6\xBD"
-     "\x0D\x00"
-     "\x96\xAE\x00\x80\x9A\xAE\x00\x80\x9E\xAE\x00\x80\xAA\x99\x0D\x00\xAB\x91"
-     "\x0D\x00"
-     "\xAC\xFD\x0D\x00\xAD\xE1\x0D\x00\xAE\xE1\x0D\x00\xAF\xDD\x0D\x00\xB6\x51"
-     "\x03\x00"
-     "\xA2\xAE\x00\x80\x84\x80\x03\x00\xB5\x41\x03\x00\xA6\xAE\x00\x80\xB3\x51"
-     "\x03\x00"
-     "\x86\xD8\x04\x00\x87\xA8\x02\x00\xBE\x59\x03\x00\xBF\x59\x03\x00\xBC\x59"
-     "\x03\x00"
-     "\xBD\x59\x03\x00\xBA\x51\x03\x00\xBB\x5D\x03\x00\xAA\xAE\x00\x80\xAE\xAE"
-     "\x00\x80"
-     "\xA3\x85\x03\x00\xB2\xAE\x00\x80\xB6\xAE\x00\x80\xBA\xAE\x00\x80\xBE\xAE"
-     "\x00\x80"
-     "\xA6\x85\x03\x00\xA5\x95\x03\x00\xC2\xAE\x00\x80\xAB\x89\x03\x00\xAA\x85"
-     "\x03\x00"
-     "\x3E\xAD\x00\x80\xC6\xAE\x00\x80\xAF\x8D\x03\x00\xAE\x8D\x03\x00\xAD\x8D"
-     "\x03\x00"
-     "\xAC\x8D\x03\x00\xCA\xAE\x00\x80\xCE\xAE\x00\x80\x85\xB4\xF5\x00\xD2\xAE"
-     "\x00\x80"
-     "\xD6\xAE\x00\x80\xDA\xAE\x00\x80\xDE\xAE\x00\x80\xE2\xAE\x00\x80\xE6\xAE"
-     "\x00\x80"
-     "\xEA\xAE\x00\x80\xEE\xAE\x00\x80\xF2\xAE\x00\x80\x81\x09\x00\x00\x80\x19"
-     "\x00\x00"
-     "\xF6\xAE\x00\x80\x82\xFD\x01\x00\xFA\xAE\x00\x80\xB3\x25\x01\x00\xBE\x48"
-     "\x02\x00"
-     "\x02\xAF\x00\x80\xB6\x41\x01\x00\x06\xAF\x00\x80\x0A\xAF\x00\x80\xB5\x59"
-     "\x01\x00"
-     "\xBA\xA5\x01\x00\xBB\xAD\x01\x00\x0E\xAF\x00\x80\x12\xAF\x00\x80\xBE\xAD"
-     "\x01\x00"
-     "\xBF\x99\x01\x00\xBC\xB5\x01\x00\xBD\xAD\x01\x00\x86\x60\x05\x00\x87\xF4"
-     "\x02\x00"
-     "\x16\xAF\x00\x80\x1A\xAF\x00\x80\xE1\xA0\x01\x00\x1E\xAF\x00\x80\xE3\x4C"
-     "\x03\x00"
-     "\x22\xAF\x00\x80\x26\xAF\x00\x80\x2A\xAF\x00\x80\xEF\x98\x03\x00\x2E\xAF"
-     "\x00\x80"
-     "\x32\xAF\x00\x80\x36\xAF\x00\x80\x3A\xAF\x00\x80\x3E\xAF\x00\x80\x42\xAF"
-     "\x00\x80"
-     "\x46\xAF\x00\x80\x4A\xAF\x00\x80\x84\x40\x04\x00\x4E\xAF\x00\x80\x52\xAF"
-     "\x00\x80"
-     "\x56\xAF\x00\x80\xEF\xC8\x01\x00\x5A\xAF\x00\x80\xE1\x7C\x0E\x00\x5E\xAF"
-     "\x00\x80"
-     "\xE3\x14\x01\x00\x80\x0D\x00\x00\x81\x75\x00\x00\x82\x7D\x00\x00\x62\xAF"
-     "\x00\x80"
-     "\x66\xAF\x00\x80\x6A\xAF\x00\x80\x72\xAF\x00\x80\xA3\xED\x01\x00\x76\xAF"
-     "\x00\x80"
-     "\xA5\x91\x01\x00\xA6\x89\x01\x00\xBE\x14\x04\x00\x86\xC0\x04\x00\x87\x44"
-     "\x04\x00"
-     "\xAA\x6D\x01\x00\xAB\x65\x01\x00\xAC\x7D\x01\x00\xAD\x65\x01\x00\xAE\x65"
-     "\x01\x00"
-     "\xAF\x51\x01\x00\x8E\xAE\x00\x80\xFE\xAE\x00\x80\x7A\xAF\x00\x80\x7E\xAF"
-     "\x00\x80"
-     "\x6E\xAF\x00\x80\x82\xAF\x00\x80\x86\xAF\x00\x80\x8A\xAF\x00\x80\xA8\x31"
-     "\x06\x00"
-     "\xA9\x39\x06\x00\xAA\x09\x06\x00\xAB\x09\x06\x00\xAC\x75\x06\x00\xAD\x65"
-     "\x06\x00"
-     "\xAE\x41\x06\x00\xAF\x41\x06\x00\xB0\xC5\x06\x00\xB1\xCD\x06\x00\xB2\xC5"
-     "\x06\x00"
-     "\xB3\xDD\x06\x00\xB4\xCD\x06\x00\xB5\xF9\x06\x00\xB6\x99\x06\x00\xB7\x99"
-     "\x06\x00"
-     "\xB8\xAD\x06\x00\xB9\xB5\x06\x00\xBA\xBD\x06\x00\xBB\xB5\x06\x00\xBC\xAD"
-     "\x06\x00"
-     "\xBD\x95\x06\x00\xBE\x9D\x06\x00\xBF\x95\x06\x00\xA8\x8D\x07\x00\xA9\x91"
-     "\x07\x00"
-     "\xAA\x91\x07\x00\xAB\xA5\x07\x00\xAC\xBD\x07\x00\xAD\xDD\x07\x00\xAE\xD5"
-     "\x07\x00"
-     "\xAF\xC9\x07\x00\x8E\xAF\x00\x80\x92\xAF\x00\x80\x96\xAF\x00\x80\x9A\xAF"
-     "\x00\x80"
-     "\x9E\xAF\x00\x80\xA2\xAF\x00\x80\xA6\xAF\x00\x80\xAA\xAF\x00\x80\xB8\x41"
-     "\x07\x00"
-     "\xB9\x41\x07\x00\xBA\x41\x07\x00\xBB\x5D\x07\x00\xBC\x45\x07\x00\xBD\x4D"
-     "\x07\x00"
-     "\xBE\x79\x07\x00\xBF\x75\x07\x00\xB0\xBD\x07\x00\xB1\x85\x07\x00\xB2\x8D"
-     "\x07\x00"
-     "\xB3\x81\x07\x00\xB4\x81\x07\x00\xB5\x81\x07\x00\xB6\x81\x07\x00\xB7\x81"
-     "\x07\x00"
-     "\xB3\x19\x06\x00\xAE\xAF\x00\x80\xB2\xAF\x00\x80\xB6\xAF\x00\x80\xBA\xAF"
-     "\x00\x80"
-     "\xB6\x5D\x06\x00\xB5\x35\x06\x00\xBE\xAF\x00\x80\xBB\x41\x06\x00\xBA\x79"
-     "\x06\x00"
-     "\xC2\xAF\x00\x80\xC6\xAF\x00\x80\xBF\x45\x06\x00\xBE\x45\x06\x00\xBD\x55"
-     "\x06\x00"
-     "\xBC\x59\x06\x00\xCA\xAF\x00\x80\xA3\x5D\x06\x00\xCE\xAF\x00\x80\xD2\xAF"
-     "\x00\x80"
-     "\xA6\x19\x06\x00\xD6\xAF\x00\x80\xDA\xAF\x00\x80\xA5\x71\x06\x00\xAA\x3D"
-     "\x06\x00"
-     "\xAB\x05\x06\x00\x84\x24\x01\x00\xBE\x24\x01\x00\xAE\x01\x06\x00\xAF\x01"
-     "\x06\x00"
-     "\xAC\x1D\x06\x00\xAD\x11\x06\x00\xA8\x49\x06\x00\xA9\x55\x06\x00\xAA\x81"
-     "\x01\x00"
-     "\xAB\x81\x01\x00\xAC\x85\x01\x00\xAD\x8D\x01\x00\xAE\x85\x01\x00\xAF\x3D"
-     "\x01\x00"
-     "\x80\xED\x00\x00\x81\xF1\x00\x00\x82\xF1\x00\x00\xE2\xAF\x00\x80\x86\x10"
-     "\x00\x00"
-     "\x87\x94\x00\x00\xE6\xAF\x00\x80\xEA\xAF\x00\x80\xB8\xC9\x00\x00\xB9\xC9"
-     "\x00\x00"
-     "\xBA\xD9\x00\x00\xBB\xD1\x00\x00\xBC\xF9\x00\x00\xBD\xF9\x00\x00\xBE\x99"
-     "\x00\x00"
-     "\xBF\x95\x00\x00\xB0\x45\x01\x00\xB1\x4D\x01\x00\xB2\x45\x01\x00\xB3\x5D"
-     "\x01\x00"
-     "\xB4\x45\x01\x00\xB5\x4D\x01\x00\xB6\x45\x01\x00\xB7\xF9\x00\x00\xBA\x5D"
-     "\x07\x00"
-     "\xBB\x6D\x07\x00\xB8\x45\x07\x00\xB9\x5D\x07\x00\xBE\xB5\x07\x00\xBF\xBD"
-     "\x07\x00"
-     "\xBC\x75\x07\x00\xBD\xBD\x07\x00\xB2\x11\x07\x00\xB3\x11\x07\x00\xB0\x09"
-     "\x07\x00"
-     "\xB1\x05\x07\x00\xB6\x75\x07\x00\xB7\x7D\x07\x00\xB4\x01\x07\x00\xB5\x01"
-     "\x07\x00"
-     "\xAA\x45\x07\x00\xAB\x4D\x07\x00\xA8\x91\x06\x00\xA9\x4D\x07\x00\xAE\x49"
-     "\x07\x00"
-     "\xAF\x49\x07\x00\xAC\x55\x07\x00\xAD\x4D\x07\x00\xEE\xAF\x00\x80\xF2\xAF"
-     "\x00\x80"
-     "\xF6\xAF\x00\x80\xFA\xAF\x00\x80\xFE\xAF\x00\x80\x02\xB0\x00\x80\x06\xB0"
-     "\x00\x80"
-     "\x0A\xB0\x00\x80\xB3\x89\x02\x00\x0E\xB0\x00\x80\x12\xB0\x00\x80\x16\xB0"
-     "\x00\x80"
-     "\x1A\xB0\x00\x80\xB6\x8D\x02\x00\xB5\x8D\x02\x00\x1E\xB0\x00\x80\xBB\x09"
-     "\x02\x00"
-     "\xBA\x01\x02\x00\x22\xB0\x00\x80\x26\xB0\x00\x80\xBF\x09\x02\x00\xBE\x01"
-     "\x02\x00"
-     "\xBD\x09\x02\x00\xBC\x11\x02\x00\x2A\xB0\x00\x80\xA3\xCD\x02\x00\x2E\xB0"
-     "\x00\x80"
-     "\x32\xB0\x00\x80\xA6\xC9\x02\x00\x36\xB0\x00\x80\x3A\xB0\x00\x80\xA5\xC9"
-     "\x02\x00"
-     "\xAA\x45\x02\x00\xAB\x4D\x02\x00\x3E\xB0\x00\x80\x42\xB0\x00\x80\xAE\x45"
-     "\x02\x00"
-     "\xAF\x4D\x02\x00\xAC\x55\x02\x00\xAD\x4D\x02\x00\xA8\xAD\x02\x00\xA9\xBD"
-     "\x02\x00"
-     "\xAA\xB5\x02\x00\xAB\x6D\x01\x00\xAC\x75\x01\x00\xAD\x7D\x01\x00\xAE\x75"
-     "\x01\x00"
-     "\xAF\x6D\x01\x00\x80\xCD\x01\x00\x81\x09\x00\x00\x82\x19\x00\x00\x46\xB0"
-     "\x00\x80"
-     "\x4A\xB0\x00\x80\x84\x34\x02\x00\xBE\xB4\xFD\x00\x52\xB0\x00\x80\xB8\xE9"
-     "\x01\x00"
-     "\xB9\xE9\x01\x00\xBA\x89\x01\x00\xBB\x85\x01\x00\xBC\x9D\x01\x00\xBD\x81"
-     "\x01\x00"
-     "\xBE\x81\x01\x00\xBF\xB5\x01\x00\xB0\x15\x01\x00\xB1\x1D\x01\x00\xB2\x15"
-     "\x01\x00"
-     "\xB3\xE9\x01\x00\xB4\xF9\x01\x00\xB5\xF9\x01\x00\xB6\xE9\x01\x00\xB7\xE1"
-     "\x01\x00"
-     "\xE1\xD8\x06\x00\xE1\xC4\x07\x00\xE3\x14\x06\x00\xE3\xD4\x07\x00\x56\xB0"
-     "\x00\x80"
-     "\xE1\x6C\x06\x00\x5A\xB0\x00\x80\xE3\xE4\x00\x00\x84\x2C\xFC\x00\x5E\xB0"
-     "\x00\x80"
-     "\xBE\xA4\xFC\x00\xEF\x38\x00\x00\x86\x90\xFD\x00\x87\xD4\xFD\x00\xEF\x9C"
-     "\x06\x00"
-     "\xEF\x4C\x06\x00\x62\xB0\x00\x80\x66\xB0\x00\x80\x84\xC8\xFC\x00\xB3\x71"
-     "\x02\x00"
-     "\x6A\xB0\x00\x80\xB5\xB1\x01\x00\xB6\xB1\x01\x00\x6E\xB0\x00\x80\x72\xB0"
-     "\x00\x80"
-     "\x76\xB0\x00\x80\xBA\x9D\x01\x00\xBB\x95\x01\x00\xBC\x8D\x01\x00\xBD\x71"
-     "\x01\x00"
-     "\xBE\x71\x01\x00\xBF\x71\x01\x00\xA3\xBD\xFD\x00\x4E\xB0\x00\x80\x7A\xB0"
-     "\x00\x80"
-     "\x7E\xB0\x00\x80\x82\xB0\x00\x80\xA6\x7D\xFE\x00\xA5\x7D\xFE\x00\x86\xB0"
-     "\x00\x80"
-     "\xAB\x59\xFE\x00\xAA\x51\xFE\x00\x8A\xB0\x00\x80\x8E\xB0\x00\x80\xAF\xBD"
-     "\xFE\x00"
-     "\xAE\xBD\xFE\x00\xAD\xBD\xFE\x00\xAC\x41\xFE\x00\x92\xB0\x00\x80\x96\xB0"
-     "\x00\x80"
-     "\x9A\xB0\x00\x80\x9E\xB0\x00\x80\xA2\xB0\x00\x80\xA6\xB0\x00\x80\xAA\xB0"
-     "\x00\x80"
-     "\xAE\xB0\x00\x80\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\xB2\xB0"
-     "\x00\x80"
-     "\xB6\xB0\x00\x80\xBA\xB0\x00\x80\xBE\xB0\x00\x80\xC2\xB0\x00\x80\xA8\xBD"
-     "\xFF\x00"
-     "\xA9\x15\xFF\x00\xAA\x11\xFF\x00\xAB\x29\xFF\x00\xAC\x51\xFF\x00\xAD\x51"
-     "\xFF\x00"
-     "\xAE\x49\xFF\x00\xAF\x41\xFF\x00\x86\x00\x00\x00\x87\x64\x03\x00\xC6\xB0"
-     "\x00\x80"
-     "\xCA\xB0\x00\x80\xCE\xB0\x00\x80\xD2\xB0\x00\x80\xD6\xB0\x00\x80\xDA\xB0"
-     "\x00\x80"
-     "\xB8\x81\x00\x00\xB9\x81\x00\x00\xBA\x81\x00\x00\xBB\x81\x00\x00\xBC\x81"
-     "\x00\x00"
-     "\xBD\x81\x00\x00\xBE\x81\x00\x00\xBF\x81\x00\x00\xB0\x01\xFF\x00\xB1\x01"
-     "\xFF\x00"
-     "\xB2\xE1\x00\x00\xB3\xE1\x00\x00\xB4\xE1\x00\x00\xB5\xE9\x00\x00\xB6\xD1"
-     "\x00\x00"
-     "\xB7\xD1\x00\x00\xDE\xB0\x00\x80\xE2\xB0\x00\x80\xE6\xB0\x00\x80\xEA\xB0"
-     "\x00\x80"
-     "\xEE\xB0\x00\x80\xE1\xBC\xFE\x00\xF2\xB0\x00\x80\xE3\x38\xFC\x00\xF6\xB0"
-     "\x00\x80"
-     "\xFA\xB0\x00\x80\xFE\xB0\x00\x80\xEF\xEC\xFC\x00\x02\xB1\x00\x80\x06\xB1"
-     "\x00\x80"
-     "\x0A\xB1\x00\x80\x0E\xB1\x00\x80\xB3\xD9\xFE\x00\x12\xB1\x00\x80\x16\xB1"
-     "\x00\x80"
-     "\x1A\xB1\x00\x80\x1E\xB1\x00\x80\xB6\x91\xFE\x00\xB5\x99\xFE\x00\x22\xB1"
-     "\x00\x80"
-     "\xBB\x85\xFE\x00\xBA\x85\xFE\x00\x26\xB1\x00\x80\x2A\xB1\x00\x80\xBF\x6D"
-     "\x01\x00"
-     "\xBE\x65\x01\x00\xBD\x75\x01\x00\xBC\x75\x01\x00\xA3\x19\xFE\x00\xBE\x84"
-     "\x01\x00"
-     "\x86\x28\x00\x00\x87\x5C\x01\x00\x39\xB1\x00\x80\xA6\x51\xFE\x00\xA5\x59"
-     "\xFE\x00"
-     "\x3D\xB1\x00\x80\xAB\x45\xFE\x00\xAA\x45\xFE\x00\x41\xB1\x00\x80\x45\xB1"
-     "\x00\x80"
-     "\xAF\xAD\x01\x00\xAE\xA5\x01\x00\xAD\xB5\x01\x00\xAC\xB5\x01\x00\x80\x9D"
-     "\x00\x00"
-     "\x81\xAD\x00\x00\x82\xA5\x00\x00\xB3\x61\x01\x00\x49\xB1\x00\x80\xB5\x61"
-     "\x01\x00"
-     "\xB6\x61\x01\x00\x4D\xB1\x00\x80\x51\xB1\x00\x80\x55\xB1\x00\x80\xBA\xED"
-     "\x00\x00"
-     "\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\xE5\x00\x00\xBE\xE1\x00\x00\xBF\xDD"
-     "\x00\x00"
-     "\xA8\x7D\x02\x00\xA9\xAD\x02\x00\xAA\xA5\x02\x00\xAB\xBD\x02\x00\xAC\xA5"
-     "\x02\x00"
-     "\xAD\xAD\x02\x00\xAE\xA5\x02\x00\xAF\x1D\x03\x00\x84\xEC\x04\x00\x59\xB1"
-     "\x00\x80"
-     "\x5D\xB1\x00\x80\x61\xB1\x00\x80\x65\xB1\x00\x80\x69\xB1\x00\x80\x6D\xB1"
-     "\x00\x80"
-     "\x71\xB1\x00\x80\xB8\x71\x03\x00\xB9\x71\x03\x00\xBA\x71\x03\x00\xBB\x71"
-     "\x03\x00"
-     "\xBC\x15\x03\x00\xBD\x1D\x03\x00\xBE\x15\x03\x00\xBF\x0D\x03\x00\xB0\x65"
-     "\x03\x00"
-     "\xB1\x6D\x03\x00\xB2\x65\x03\x00\xB3\x7D\x03\x00\xB4\x69\x03\x00\xB5\x55"
-     "\x03\x00"
-     "\xB6\x5D\x03\x00\xB7\x55\x03\x00\x75\xB1\x00\x80\x79\xB1\x00\x80\x7D\xB1"
-     "\x00\x80"
-     "\xA3\x21\x02\x00\xDE\xAF\x00\x80\xA5\x21\x02\x00\xA6\x21\x02\x00\x81\xB1"
-     "\x00\x80"
-     "\x85\xB1\x00\x80\x89\xB1\x00\x80\xAA\xAD\x03\x00\xAB\xA5\x03\x00\xAC\xBD"
-     "\x03\x00"
-     "\xAD\xA5\x03\x00\xAE\xA1\x03\x00\xAF\x9D\x03\x00\x8D\xB1\x00\x80\x91\xB1"
-     "\x00\x80"
-     "\x95\xB1\x00\x80\x99\xB1\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39"
-     "\x00\x00"
-     "\x9D\xB1\x00\x80\xA1\xB1\x00\x80\xA9\xB1\x00\x80\xAD\xB1\x00\x80\xBE\x24"
-     "\x04\x00"
-     "\x86\xD0\x04\x00\x87\x20\x03\x00\xB1\xB1\x00\x80\xB5\xB1\x00\x80\xA8\x91"
-     "\x02\x00"
-     "\xA9\x99\x02\x00\xAA\xA1\x02\x00\xAB\xA1\x02\x00\xAC\xC1\x02\x00\xAD\xC9"
-     "\x02\x00"
-     "\xAE\xF9\x02\x00\xAF\xF9\x02\x00\xB9\xB1\x00\x80\xBD\xB1\x00\x80\xC1\xB1"
-     "\x00\x80"
-     "\xC5\xB1\x00\x80\xC9\xB1\x00\x80\xCD\xB1\x00\x80\xD1\xB1\x00\x80\xD5\xB1"
-     "\x00\x80"
-     "\xB8\xC5\x01\x00\xB9\xCD\x01\x00\xBA\xD5\x01\x00\xBB\xED\x01\x00\xBC\xF5"
-     "\x01\x00"
-     "\xBD\x95\x01\x00\xBE\x9D\x01\x00\xBF\x95\x01\x00\xB0\x45\x01\x00\xB1\x4D"
-     "\x01\x00"
-     "\xB2\x45\x01\x00\xB3\x5D\x01\x00\xB4\x45\x01\x00\xB5\x4D\x01\x00\xB6\x45"
-     "\x01\x00"
-     "\xB7\xFD\x01\x00\xE1\x28\xFE\x00\xD9\xB1\x00\x80\xE3\x38\xFE\x00\xBE\x80"
-     "\x04\x00"
-     "\x84\x1C\x04\x00\xDD\xB1\x00\x80\xE1\xB1\x00\x80\xE5\xB1\x00\x80\xE9\xB1"
-     "\x00\x80"
-     "\xED\xB1\x00\x80\xF1\xB1\x00\x80\xF5\xB1\x00\x80\xF9\xB1\x00\x80\xFD\xB1"
-     "\x00\x80"
-     "\xEF\xE8\xFE\x00\x01\xB2\x00\x80\xF6\x00\x00\x00\xEF\xB8\x00\x00\x05\xB2"
-     "\x00\x80"
-     "\x09\xB2\x00\x80\x81\x09\x00\x00\x80\x19\x00\x00\x0D\xB2\x00\x80\x82\x3D"
-     "\x00\x00"
-     "\x11\xB2\x00\x80\x19\xB2\x00\x80\x86\x48\x07\x00\x87\xFC\x04\x00\x1D\xB2"
-     "\x00\x80"
-     "\xE3\x2C\xFE\x00\x21\xB2\x00\x80\xE1\x44\x01\x00\xB3\xA9\x05\x00\xA5\xB1"
-     "\x00\x80"
-     "\x15\xB2\x00\x80\x25\xB2\x00\x80\x29\xB2\x00\x80\xB6\xC9\x05\x00\xB5\xC9"
-     "\x05\x00"
-     "\x2D\xB2\x00\x80\xBB\x09\x06\x00\xBA\x01\x06\x00\x31\xB2\x00\x80\x35\xB2"
-     "\x00\x80"
-     "\xBF\x09\x06\x00\xBE\x01\x06\x00\xBD\x09\x06\x00\xBC\x11\x06\x00\x39\xB2"
-     "\x00\x80"
-     "\xA3\xED\x05\x00\x3D\xB2\x00\x80\x41\xB2\x00\x80\xA6\x8D\x05\x00\x45\xB2"
-     "\x00\x80"
-     "\x49\xB2\x00\x80\xA5\x8D\x05\x00\xAA\x45\x06\x00\xAB\x4D\x06\x00\x4D\xB2"
-     "\x00\x80"
-     "\x51\xB2\x00\x80\xAE\x45\x06\x00\xAF\x4D\x06\x00\xAC\x55\x06\x00\xAD\x4D"
-     "\x06\x00"
-     "\xA8\x65\x06\x00\xA9\x6D\x06\x00\xAA\x71\x06\x00\xAB\x71\x06\x00\xAC\xB1"
-     "\x07\x00"
-     "\xAD\xB1\x07\x00\xAE\xB1\x07\x00\xAF\xB1\x07\x00\x84\xA8\x03\x00\x55\xB2"
-     "\x00\x80"
-     "\x59\xB2\x00\x80\x5D\xB2\x00\x80\x61\xB2\x00\x80\x65\xB2\x00\x80\x69\xB2"
-     "\x00\x80"
-     "\x6D\xB2\x00\x80\xB8\x61\x07\x00\xB9\x61\x07\x00\xBA\x61\x07\x00\xBB\x61"
-     "\x07\x00"
-     "\xBC\x61\x07\x00\xBD\x61\x07\x00\xBE\x65\x07\x00\xBF\x15\x07\x00\xB0\xD1"
-     "\x07\x00"
-     "\xB1\xD1\x07\x00\xB2\xD1\x07\x00\xB3\xD1\x07\x00\xB4\x75\x07\x00\xB5\x61"
-     "\x07\x00"
-     "\xB6\x61\x07\x00\xB7\x61\x07\x00\xE1\x90\x06\x00\x71\xB2\x00\x80\xE3\x38"
-     "\x05\x00"
-     "\x75\xB2\x00\x80\x80\x0D\x00\x00\x81\x35\x00\x00\x82\x3D\x00\x00\x79\xB2"
-     "\x00\x80"
-     "\x7D\xB2\x00\x80\x81\xB2\x00\x80\xBE\x28\x01\x00\x89\xB2\x00\x80\x86\x90"
-     "\x00\x00"
-     "\x87\x04\x01\x00\xEF\x2C\x05\x00\x8D\xB2\x00\x80\xB3\xB9\x07\x00\x85\xB2"
-     "\x00\x80"
-     "\x91\xB2\x00\x80\x95\xB2\x00\x80\x99\xB2\x00\x80\xB6\x71\x06\x00\xB5\x79"
-     "\x06\x00"
-     "\x9D\xB2\x00\x80\xBB\x55\x06\x00\xBA\x55\x06\x00\xA1\xB2\x00\x80\xA5\xB2"
-     "\x00\x80"
-     "\xBF\xE5\x06\x00\xBE\xF1\x06\x00\xBD\x45\x06\x00\xBC\x45\x06\x00\xA3\x79"
-     "\x07\x00"
-     "\xA9\xB2\x00\x80\xAD\xB2\x00\x80\xB1\xB2\x00\x80\xB5\xB2\x00\x80\xA6\xB1"
-     "\x06\x00"
-     "\xA5\xB9\x06\x00\xB9\xB2\x00\x80\xAB\x95\x06\x00\xAA\x95\x06\x00\xBD\xB2"
-     "\x00\x80"
-     "\xC1\xB2\x00\x80\xAF\x25\x06\x00\xAE\x31\x06\x00\xAD\x85\x06\x00\xAC\x85"
-     "\x06\x00"
-     "\xC5\xB2\x00\x80\xB3\x3D\x06\x00\xC9\xB2\x00\x80\xCD\xB2\x00\x80\xB6\xCD"
-     "\x01\x00"
-     "\xD1\xB2\x00\x80\xD5\xB2\x00\x80\xB5\xCD\x01\x00\xBA\xA1\x01\x00\xBB\xAD"
-     "\x01\x00"
-     "\xD9\xB2\x00\x80\xDD\xB2\x00\x80\xBE\xA9\x01\x00\xBF\x9D\x01\x00\xBC\xB5"
-     "\x01\x00"
-     "\xBD\xA9\x01\x00\xA8\xE5\x05\x00\xA9\xF5\x05\x00\xAA\xF1\x05\x00\xAB\x01"
-     "\x02\x00"
-     "\xAC\x01\x02\x00\xAD\x09\x02\x00\xAE\x31\x02\x00\xAF\x31\x02\x00\xBE\x00"
-     "\x0C\x00"
-     "\xE1\xB2\x00\x80\xE5\xB2\x00\x80\xE9\xB2\x00\x80\x80\x1D\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\xB9\x00\x00\xED\xB2\x00\x80\xB8\xE9\x02\x00\xB9\xE9\x02\x00\xBA\xF9"
-     "\x02\x00"
-     "\xBB\xF9\x02\x00\xBC\xE9\x02\x00\xBD\xE9\x02\x00\xBE\x5D\x03\x00\xBF\x51"
-     "\x03\x00"
-     "\xB0\x21\x02\x00\xB1\x2D\x02\x00\xB2\x25\x02\x00\xB3\x39\x02\x00\xB4\x29"
-     "\x02\x00"
-     "\xB5\x1D\x02\x00\xB6\x15\x02\x00\xB7\xD9\x02\x00\xF1\xB2\x00\x80\x2E\xB1"
-     "\x00\x80"
-     "\x84\x80\x03\x00\xA3\x75\x05\x00\xF9\xB2\x00\x80\xA5\x85\x02\x00\xA6\x85"
-     "\x02\x00"
-     "\xFD\xB2\x00\x80\x01\xB3\x00\x80\x05\xB3\x00\x80\xAA\xE9\x02\x00\xAB\xE5"
-     "\x02\x00"
-     "\xAC\xFD\x02\x00\xAD\xE1\x02\x00\xAE\xE1\x02\x00\xAF\xD5\x02\x00\xE1\xE8"
-     "\x01\x00"
-     "\x09\xB3\x00\x80\xE3\x50\x02\x00\x84\xC0\x0C\x00\x86\x50\x0C\x00\x87\x70"
-     "\x03\x00"
-     "\x0D\xB3\x00\x80\x11\xB3\x00\x80\x15\xB3\x00\x80\x19\xB3\x00\x80\x1D\xB3"
-     "\x00\x80"
-     "\x21\xB3\x00\x80\x25\xB3\x00\x80\x29\xB3\x00\x80\xEF\x20\x02\x00\x2D\xB3"
-     "\x00\x80"
-     "\xA8\x2D\x02\x00\xA9\x39\x02\x00\xAA\x41\x02\x00\xAB\x41\x02\x00\xAC\x41"
-     "\x02\x00"
-     "\xAD\x49\x02\x00\xAE\x71\x02\x00\xAF\x71\x02\x00\x31\xB3\x00\x80\x35\xB3"
-     "\x00\x80"
-     "\xBE\xA4\x0C\x00\x39\xB3\x00\x80\x3D\xB3\x00\x80\x41\xB3\x00\x80\x45\xB3"
-     "\x00\x80"
-     "\x49\xB3\x00\x80\xB8\x89\x02\x00\xB9\x89\x02\x00\xBA\x9D\x02\x00\xBB\x95"
-     "\x02\x00"
-     "\xBC\xB9\x02\x00\xBD\xB9\x02\x00\xBE\x7D\x01\x00\xBF\x75\x01\x00\xB0\xE5"
-     "\x02\x00"
-     "\xB1\xED\x02\x00\xB2\xE5\x02\x00\xB3\xFD\x02\x00\xB4\xED\x02\x00\xB5\xDD"
-     "\x02\x00"
-     "\xB6\xD5\x02\x00\xB7\xBD\x02\x00\xE1\x68\x07\x00\xE3\x34\x01\x00\xE3\x08"
-     "\x06\x00"
-     "\xE1\x7C\x06\x00\x4D\xB3\x00\x80\x51\xB3\x00\x80\x55\xB3\x00\x80\x59\xB3"
-     "\x00\x80"
-     "\x81\x09\x00\x00\x80\x39\x00\x00\x5D\xB3\x00\x80\x82\x1D\x00\x00\x61\xB3"
-     "\x00\x80"
-     "\xEF\x64\x00\x00\xEF\xCC\x06\x00\x69\xB3\x00\x80\x6D\xB3\x00\x80\xB3\x8D"
-     "\x02\x00"
-     "\x71\xB3\x00\x80\x75\xB3\x00\x80\xB6\x8D\x02\x00\x79\xB3\x00\x80\x7D\xB3"
-     "\x00\x80"
-     "\xB5\x8D\x02\x00\xBA\x65\x02\x00\xBB\x6D\x02\x00\x86\xC8\x0C\x00\x87\xE0"
-     "\x0C\x00"
-     "\xBE\x6D\x02\x00\xBF\x5D\x02\x00\xBC\x75\x02\x00\xBD\x6D\x02\x00\xF5\xB2"
-     "\x00\x80"
-     "\x65\xB3\x00\x80\x81\xB3\x00\x80\x85\xB3\x00\x80\x89\xB3\x00\x80\x8D\xB3"
-     "\x00\x80"
-     "\x91\xB3\x00\x80\x95\xB3\x00\x80\xA8\xB1\x0E\x00\xA9\xB9\x0E\x00\xAA\x89"
-     "\x0E\x00"
-     "\xAB\x89\x0E\x00\xAC\xF5\x0E\x00\xAD\xED\x0E\x00\xAE\xE5\x0E\x00\xAF\xDD"
-     "\x0E\x00"
-     "\xB0\xA5\x0E\x00\xB1\xAD\x0E\x00\xB2\xA5\x0E\x00\xB3\xBD\x0E\x00\xB4\xA5"
-     "\x0E\x00"
-     "\xB5\xA9\x0E\x00\xB6\x79\x0F\x00\xB7\x79\x0F\x00\xB8\x4D\x0F\x00\xB9\x55"
-     "\x0F\x00"
-     "\xBA\x5D\x0F\x00\xBB\x55\x0F\x00\xBC\x4D\x0F\x00\xBD\xE5\x0F\x00\xBE\xED"
-     "\x0F\x00"
-     "\xBF\xE5\x0F\x00\x99\xB3\x00\x80\x9D\xB3\x00\x80\xAA\x0D\x0D\x00\xAB\x05"
-     "\x0D\x00"
-     "\xAC\x1D\x0D\x00\xAD\x05\x0D\x00\xAE\x05\x0D\x00\xAF\x35\x0D\x00\xA1\xB3"
-     "\x00\x80"
-     "\xA5\xB3\x00\x80\xA9\xB3\x00\x80\xA3\xE5\x0D\x00\xAD\xB3\x00\x80\xA5\xE5"
-     "\x0D\x00"
-     "\xA6\xE5\x0D\x00\xB1\xB3\x00\x80\xB3\x61\x0E\x00\xB5\xB3\x00\x80\xB9\xB3"
-     "\x00\x80"
-     "\xBD\xB3\x00\x80\xC1\xB3\x00\x80\xB6\x0D\x0E\x00\xB5\x1D\x0E\x00\xC5\xB3"
-     "\x00\x80"
-     "\xBB\x09\x0E\x00\xBA\x01\x0E\x00\xC9\xB3\x00\x80\xCD\xB3\x00\x80\xBF\x0D"
-     "\x0E\x00"
-     "\xBE\x01\x0E\x00\xBD\x09\x0E\x00\xBC\x11\x0E\x00\xA3\xA1\x0F\x00\x82\x15"
-     "\x00\x00"
-     "\x81\xE5\x01\x00\x80\xE5\x01\x00\xD1\xB3\x00\x80\xA6\xCD\x0F\x00\xA5\xDD"
-     "\x0F\x00"
-     "\xBE\x90\x01\x00\xAB\xC9\x0F\x00\xAA\xC1\x0F\x00\x84\xA0\x02\x00\xD5\xB3"
-     "\x00\x80"
-     "\xAF\xCD\x0F\x00\xAE\xC1\x0F\x00\xAD\xC9\x0F\x00\xAC\xD1\x0F\x00\xD9\xB3"
-     "\x00\x80"
-     "\xB3\x8D\x0E\x00\x86\xE8\x00\x00\x87\x2C\x01\x00\xB6\x89\x0E\x00\xDD\xB3"
-     "\x00\x80"
-     "\xE1\xB3\x00\x80\xB5\x89\x0E\x00\xBA\x71\x01\x00\xBB\x71\x01\x00\xE5\xB3"
-     "\x00\x80"
-     "\xE9\xB3\x00\x80\xBE\xF1\x01\x00\xBF\xF1\x01\x00\xBC\x61\x01\x00\xBD\x61"
-     "\x01\x00"
-     "\xA8\x3D\x0E\x00\xA9\x45\x0E\x00\xAA\x41\x0E\x00\xAB\x41\x0E\x00\xAC\x45"
-     "\x0E\x00"
-     "\xAD\x4D\x0E\x00\xAE\x91\x01\x00\xAF\x91\x01\x00\x84\x8C\x01\x00\xED\xB3"
-     "\x00\x80"
-     "\xF1\xB3\x00\x80\xF5\xB3\x00\x80\xF9\xB3\x00\x80\xFD\xB3\x00\x80\x01\xB4"
-     "\x00\x80"
-     "\x05\xB4\x00\x80\xB8\x69\x01\x00\xB9\x75\x01\x00\xBA\x75\x01\x00\xBB\x09"
-     "\x01\x00"
-     "\xBC\x19\x01\x00\xBD\x19\x01\x00\xBE\x09\x01\x00\xBF\x01\x01\x00\xB0\xF1"
-     "\x01\x00"
-     "\xB1\xF1\x01\x00\xB2\xF1\x01\x00\xB3\xF1\x01\x00\xB4\x71\x01\x00\xB5\x71"
-     "\x01\x00"
-     "\xB6\x59\x01\x00\xB7\x59\x01\x00\xA3\xC9\x0D\x00\x09\xB4\x00\x80\x0D\xB4"
-     "\x00\x80"
-     "\xBE\x40\x03\x00\x11\xB4\x00\x80\xA6\xCD\x0D\x00\xA5\xCD\x0D\x00\x15\xB4"
-     "\x00\x80"
-     "\xAB\x35\x02\x00\xAA\x35\x02\x00\x19\xB4\x00\x80\x1D\xB4\x00\x80\xAF\xB5"
-     "\x02\x00"
-     "\xAE\xB5\x02\x00\xAD\x25\x02\x00\xAC\x25\x02\x00\x21\xB4\x00\x80\x25\xB4"
-     "\x00\x80"
-     "\x29\xB4\x00\x80\x2D\xB4\x00\x80\x31\xB4\x00\x80\x35\xB4\x00\x80\x39\xB4"
-     "\x00\x80"
-     "\x3D\xB4\x00\x80\x80\x39\x00\x00\x81\x39\x00\x00\x82\x05\x00\x00\x41\xB4"
-     "\x00\x80"
-     "\x86\xD0\x04\x00\x87\x44\x03\x00\x49\xB4\x00\x80\x4D\xB4\x00\x80\x84\xCC"
-     "\x04\x00"
-     "\x51\xB4\x00\x80\x55\xB4\x00\x80\x59\xB4\x00\x80\xEF\x64\x02\x00\x5D\xB4"
-     "\x00\x80"
-     "\x61\xB4\x00\x80\x65\xB4\x00\x80\xE3\x10\x01\x00\x69\xB4\x00\x80\xE1\xF8"
-     "\x01\x00"
-     "\x6D\xB4\x00\x80\x71\xB4\x00\x80\x75\xB4\x00\x80\x79\xB4\x00\x80\x7D\xB4"
-     "\x00\x80"
-     "\xBB\x75\x03\x00\xBA\x75\x03\x00\x81\xB4\x00\x80\x85\xB4\x00\x80\xBF\xD9"
-     "\x00\x00"
-     "\xBE\xD1\x00\x00\xBD\x65\x03\x00\xBC\x65\x03\x00\xB3\xB9\x03\x00\x89\xB4"
-     "\x00\x80"
-     "\x8D\xB4\x00\x80\x91\xB4\x00\x80\x95\xB4\x00\x80\xB6\x51\x03\x00\xB5\x95"
-     "\x03\x00"
-     "\x99\xB4\x00\x80\xA8\x41\x02\x00\xA9\x51\x02\x00\xAA\x51\x02\x00\xAB\x65"
-     "\x02\x00"
-     "\xAC\x7D\x02\x00\xAD\xB1\x02\x00\xAE\xB1\x02\x00\xAF\xB1\x02\x00\xBE\x68"
-     "\x05\x00"
-     "\x9D\xB4\x00\x80\xA1\xB4\x00\x80\xA5\xB4\x00\x80\x80\x0D\x00\x00\x81\xA5"
-     "\x00\x00"
-     "\x82\xAD\x00\x00\xA9\xB4\x00\x80\xB8\x69\x01\x00\xB9\x69\x01\x00\xBA\x09"
-     "\x01\x00"
-     "\xBB\x09\x01\x00\xBC\x19\x01\x00\xBD\x19\x01\x00\xBE\x09\x01\x00\xBF\x09"
-     "\x01\x00"
-     "\xB0\xD1\x02\x00\xB1\xD1\x02\x00\xB2\xD1\x02\x00\xB3\xD1\x02\x00\xB4\x75"
-     "\x01\x00"
-     "\xB5\x7D\x01\x00\xB6\x69\x01\x00\xB7\x61\x01\x00\xE1\x70\x0E\x00\xE1\x4C"
-     "\x0F\x00"
-     "\xE3\xB0\x00\x00\xE3\x50\x0F\x00\xAD\xB4\x00\x80\xB1\xB4\x00\x80\xEF\x48"
-     "\xE7\x00"
-     "\x85\x98\x04\x00\xE1\x28\x02\x00\xB9\xB4\x00\x80\xE3\x70\x13\x00\xBD\xB4"
-     "\x00\x80"
-     "\xBE\x78\x06\x00\xC1\xB4\x00\x80\xEF\x5C\x00\x00\xEF\x6C\x0E\x00\xA3\x99"
-     "\x02\x00"
-     "\xC5\xB4\x00\x80\x86\xA8\x07\x00\x87\x80\x04\x00\xC9\xB4\x00\x80\xA6\x71"
-     "\x02\x00"
-     "\xA5\xB5\x02\x00\xCD\xB4\x00\x80\xAB\x55\x02\x00\xAA\x55\x02\x00\xD1\xB4"
-     "\x00\x80"
-     "\xD5\xB4\x00\x80\xAF\xF9\x01\x00\xAE\xF1\x01\x00\xAD\x45\x02\x00\xAC\x45"
-     "\x02\x00"
-     "\xA8\x79\x06\x00\xA9\x79\x06\x00\xAA\x8D\x06\x00\xAB\x85\x06\x00\xAC\x9D"
-     "\x06\x00"
-     "\xAD\x81\x06\x00\xAE\x81\x06\x00\xAF\x81\x06\x00\xB5\xB4\x00\x80\xD9\xB4"
-     "\x00\x80"
-     "\xDD\xB4\x00\x80\xE1\xB4\x00\x80\xE5\xB4\x00\x80\xE9\xB4\x00\x80\xED\xB4"
-     "\x00\x80"
-     "\xF1\xB4\x00\x80\xB8\xA5\x06\x00\xB9\xAD\x06\x00\xBA\xA5\x06\x00\xBB\xBD"
-     "\x06\x00"
-     "\xBC\xAD\x06\x00\xBD\x71\x07\x00\xBE\x71\x07\x00\xBF\x71\x07\x00\xB0\xC5"
-     "\x06\x00"
-     "\xB1\xC9\x06\x00\xB2\xD9\x06\x00\xB3\xD9\x06\x00\xB4\xC9\x06\x00\xB5\xC9"
-     "\x06\x00"
-     "\xB6\xB1\x06\x00\xB7\xB1\x06\x00\xB3\x05\x06\x00\xF5\xB4\x00\x80\xF9\xB4"
-     "\x00\x80"
-     "\xFD\xB4\x00\x80\x01\xB5\x00\x80\xB6\x01\x06\x00\xB5\x15\x06\x00\x05\xB5"
-     "\x00\x80"
-     "\xBB\x0D\x06\x00\xBA\x01\x06\x00\x09\xB5\x00\x80\x0D\xB5\x00\x80\xBF\x0D"
-     "\x06\x00"
-     "\xBE\x0D\x06\x00\xBD\x0D\x06\x00\xBC\x0D\x06\x00\x11\xB5\x00\x80\xA3\x41"
-     "\x06\x00"
-     "\x15\xB5\x00\x80\x19\xB5\x00\x80\xA6\x45\x06\x00\x1D\xB5\x00\x80\x21\xB5"
-     "\x00\x80"
-     "\xA5\x51\x06\x00\xAA\x45\x06\x00\xAB\x49\x06\x00\x84\x24\x01\x00\x29\xB5"
-     "\x00\x80"
-     "\xAE\x49\x06\x00\xAF\x49\x06\x00\xAC\x49\x06\x00\xAD\x49\x06\x00\x9F\x45"
-     "\x08\x00"
-     "\x9E\x0D\x09\x00\x9D\x0D\x09\x00\x9C\x89\x0B\x00\x9B\x69\x0D\x00\x9A\x11"
-     "\x0D\x00"
-     "\x99\x89\x0F\x00\x98\x0D\xF1\x00\x97\x25\xF1\x00\x96\x25\xF1\x00\x95\xFD"
-     "\xF5\x00"
-     "\x94\x0D\xF5\x00\x93\x99\xF7\x00\x92\x99\xF9\x00\x91\x31\xF9\x00\x90\x39"
-     "\xF9\x00"
-     "\x8F\x29\xFD\x00\x82\x9D\x01\x00\x81\x9D\x01\x00\x80\x81\x01\x00\x2D\xB5"
-     "\x00\x80"
-     "\xEF\x8C\x04\x00\xEF\xEC\x01\x00\x31\xB5\x00\x80\xE1\x08\x02\x00\xE3\xF4"
-     "\x05\x00"
-     "\xE3\x30\x01\x00\xE1\x04\x05\x00\x86\x00\x1C\x00\x87\x88\x01\x00\x35\xB5"
-     "\x00\x80"
-     "\x39\xB5\x00\x80\xA1\xA1\x01\x00\x3D\xB5\x00\x80\xA3\x09\x1D\x00\xA2\x69"
-     "\x03\x00"
-     "\xA5\xF9\x1C\x00\xA4\x11\x1D\x00\xA7\x1D\x19\x00\xA6\x1D\x19\x00\xA9\x21"
-     "\x15\x00"
-     "\xA8\xB5\x18\x00\xAB\x79\x15\x00\xAA\x21\x15\x00\xAD\x91\x11\x00\xAC\x01"
-     "\x10\x00"
-     "\xAF\x01\x2C\x00\xAE\x89\x10\x00\xB1\x5D\x2D\x00\xB0\xE5\x2D\x00\xB3\xE9"
-     "\x29\x00"
-     "\xB2\x01\x28\x00\xB5\x01\x28\x00\xB4\xED\x29\x00\x41\xB5\x00\x80\x45\xB5"
-     "\x00\x80"
-     "\x49\xB5\x00\x80\x4D\xB5\x00\x80\x51\xB5\x00\x80\x55\xB5\x00\x80\xA3\x19"
-     "\x03\x00"
-     "\xA2\xD9\x04\x00\xA1\xFD\x05\x00\xA0\xFD\x05\x00\x82\x69\xFE\x00\x83\x01"
-     "\xFA\x00"
-     "\x59\xB5\x00\x80\x5D\xB5\x00\x80\x86\xB9\xF4\x00\x87\x35\xF6\x00\x84\x11"
-     "\xFA\x00"
-     "\x85\xB1\xFA\x00\x8A\xF5\xF6\x00\x8B\x25\xF2\x00\x61\xB5\x00\x80\x65\xB5"
-     "\x00\x80"
-     "\x8E\xA1\x0C\x00\x8F\x31\x0E\x00\x8C\x3D\xF2\x00\x8D\x2D\xF2\x00\x92\x39"
-     "\x0E\x00"
-     "\x93\xC1\x08\x00\x84\x44\x03\x00\xBE\xC0\x1C\x00\x96\x15\x0B\x00\x97\x55"
-     "\x06\x00"
-     "\x94\x5D\x0A\x00\x95\x55\x0A\x00\x9A\x65\x06\x00\x9B\x49\x07\x00\x69\xB5"
-     "\x00\x80"
-     "\x6D\xB5\x00\x80\x71\xB5\x00\x80\x75\xB5\x00\x80\x9C\x91\x00\x00\x79\xB5"
-     "\x00\x80"
-     "\xB8\x09\x02\x00\xB9\x09\x02\x00\xBA\x19\x02\x00\xBB\x11\x02\x00\xBC\x39"
-     "\x02\x00"
-     "\xBD\x39\x02\x00\xBE\x59\x02\x00\xBF\x55\x02\x00\xB0\x85\x02\x00\xB1\x8D"
-     "\x02\x00"
-     "\xB2\x85\x02\x00\xB3\x9D\x02\x00\xB4\x85\x02\x00\xB5\x8D\x02\x00\xB6\x85"
-     "\x02\x00"
-     "\xB7\x39\x02\x00\xA8\x65\x05\x00\xA9\x69\x05\x00\xAA\x7D\x05\x00\xAB\x09"
-     "\x05\x00"
-     "\xAC\x01\x05\x00\xAD\x09\x05\x00\xAE\x31\x05\x00\xAF\x31\x05\x00\x7D\xB5"
-     "\x00\x80"
-     "\x25\xB5\x00\x80\x45\xB4\x00\x80\x81\xB5\x00\x80\x85\xB5\x00\x80\x82\xE5"
-     "\x01\x00"
-     "\x81\xF5\x01\x00\x80\xF5\x01\x00\xB3\xB5\x04\x00\x84\x8C\x1D\x00\x89\xB5"
-     "\x00\x80"
-     "\x8D\xB5\x00\x80\x91\xB5\x00\x80\xB6\xC9\x04\x00\xB5\xC9\x04\x00\x99\xB5"
-     "\x00\x80"
-     "\xBB\x09\x03\x00\xBA\x01\x03\x00\x86\x28\x1D\x00\x87\x20\x1D\x00\xBF\x09"
-     "\x03\x00"
-     "\xBE\x01\x03\x00\xBD\x09\x03\x00\xBC\x11\x03\x00\x9D\xB5\x00\x80\xA3\xF1"
-     "\x04\x00"
-     "\xA1\xB5\x00\x80\xBE\xC0\x1C\x00\xA6\x8D\x04\x00\xA5\xB5\x00\x80\xA9\xB5"
-     "\x00\x80"
-     "\xA5\x8D\x04\x00\xAA\x45\x03\x00\xAB\x4D\x03\x00\xBE\xA0\x1F\x00\xAD\xB5"
-     "\x00\x80"
-     "\xAE\x45\x03\x00\xAF\x4D\x03\x00\xAC\x55\x03\x00\xAD\x4D\x03\x00\xB1\xB5"
-     "\x00\x80"
-     "\xB5\xB5\x00\x80\xB9\xB5\x00\x80\xBD\xB5\x00\x80\xC1\xB5\x00\x80\xC5\xB5"
-     "\x00\x80"
-     "\xC9\xB5\x00\x80\xEF\xD4\x06\x00\xCD\xB5\x00\x80\xE1\x74\x06\x00\xD1\xB5"
-     "\x00\x80"
-     "\xE3\x54\x01\x00\xD5\xB5\x00\x80\xD9\xB5\x00\x80\xDD\xB5\x00\x80\xE1\xB5"
-     "\x00\x80"
-     "\xB3\x9D\x02\x00\xE5\xB5\x00\x80\xE9\xB5\x00\x80\xED\xB5\x00\x80\xF1\xB5"
-     "\x00\x80"
-     "\xB6\x51\x02\x00\xB5\xA5\x02\x00\xF5\xB5\x00\x80\xBB\x05\x02\x00\xBA\x05"
-     "\x02\x00"
-     "\xF9\xB5\x00\x80\xFD\xB5\x00\x80\xBF\xFD\x01\x00\xBE\xFD\x01\x00\xBD\x15"
-     "\x02\x00"
-     "\xBC\x15\x02\x00\xA3\x5D\x1D\x00\x82\x15\x00\x00\x81\xE5\x1F\x00\x80\xE5"
-     "\x1F\x00"
-     "\x01\xB6\x00\x80\xA6\x91\x1D\x00\xA5\x65\x1D\x00\x09\xB6\x00\x80\xAB\xC5"
-     "\x1D\x00"
-     "\xAA\xC5\x1D\x00\x86\x08\x03\x00\x87\x7C\x03\x00\xAF\x3D\x1E\x00\xAE\x3D"
-     "\x1E\x00"
-     "\xAD\xD5\x1D\x00\xAC\xD5\x1D\x00\x0D\xB6\x00\x80\xB3\x0D\x1F\x00\x11\xB6"
-     "\x00\x80"
-     "\x15\xB6\x00\x80\xB6\x3D\x1F\x00\x19\xB6\x00\x80\x1D\xB6\x00\x80\xB5\x05"
-     "\x1F\x00"
-     "\xBA\xE9\x1F\x00\xBB\xE9\x1F\x00\x21\xB6\x00\x80\x25\xB6\x00\x80\xBE\xE9"
-     "\x1F\x00"
-     "\xBF\xD5\x1F\x00\xBC\xE9\x1F\x00\xBD\xE9\x1F\x00\x95\xB5\x00\x80\x05\xB6"
-     "\x00\x80"
-     "\x29\xB6\x00\x80\x2D\xB6\x00\x80\x31\xB6\x00\x80\x35\xB6\x00\x80\x39\xB6"
-     "\x00\x80"
-     "\x3D\xB6\x00\x80\xA8\xBD\x1F\x00\xA9\xE5\x1F\x00\xAA\xE1\x1F\x00\xAB\xF9"
-     "\x1F\x00"
-     "\xAC\xE9\x1F\x00\xAD\xD9\x1F\x00\xAE\x39\x1F\x00\xAF\x39\x1F\x00\xB0\x49"
-     "\x1F\x00"
-     "\xB1\x49\x1F\x00\xB2\x59\x1F\x00\xB3\x59\x1F\x00\xB4\x49\x1F\x00\xB5\x75"
-     "\x1F\x00"
-     "\xB6\x71\x1F\x00\xB7\x6D\x1F\x00\xB8\x55\x1F\x00\xB9\x5D\x1F\x00\xBA\x55"
-     "\x1F\x00"
-     "\xBB\x29\x1F\x00\xBC\x39\x1F\x00\xBD\x39\x1F\x00\xBE\x29\x1F\x00\xBF\x21"
-     "\x1F\x00"
-     "\xA3\x49\x1E\x00\x41\xB6\x00\x80\x45\xB6\x00\x80\x49\xB6\x00\x80\x4D\xB6"
-     "\x00\x80"
-     "\xA6\x79\x1E\x00\xA5\x41\x1E\x00\x51\xB6\x00\x80\xAB\xAD\x1E\x00\xAA\xAD"
-     "\x1E\x00"
-     "\x55\xB6\x00\x80\x59\xB6\x00\x80\xAF\x91\x1E\x00\xAE\xAD\x1E\x00\xAD\xAD"
-     "\x1E\x00"
-     "\xAC\xAD\x1E\x00\x5D\xB6\x00\x80\x61\xB6\x00\x80\x65\xB6\x00\x80\x69\xB6"
-     "\x00\x80"
-     "\x80\x1D\x00\x00\x81\x09\x00\x00\x82\x39\x00\x00\x6D\xB6\x00\x80\x71\xB6"
-     "\x00\x80"
-     "\x84\xA8\x01\x00\x75\xB6\x00\x80\x79\xB6\x00\x80\x86\x00\x04\x00\x87\xD4"
-     "\x01\x00"
-     "\x7D\xB6\x00\x80\x81\xB6\x00\x80\xA8\x9D\x01\x00\xA9\xC5\x01\x00\xAA\xC1"
-     "\x01\x00"
-     "\xAB\xC1\x01\x00\xAC\xC5\x01\x00\xAD\xCD\x01\x00\xAE\xF1\x01\x00\xAF\xF1"
-     "\x01\x00"
-     "\x85\xB6\x00\x80\x89\xB6\x00\x80\x8D\xB6\x00\x80\x91\xB6\x00\x80\x95\xB6"
-     "\x00\x80"
-     "\x99\xB6\x00\x80\x9D\xB6\x00\x80\xA1\xB6\x00\x80\xB8\xCD\x00\x00\xB9\xD5"
-     "\x00\x00"
-     "\xBA\xD5\x00\x00\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x9D"
-     "\x00\x00"
-     "\xBF\x95\x00\x00\xB0\x69\x01\x00\xB1\x69\x01\x00\xB2\x79\x01\x00\xB3\x79"
-     "\x01\x00"
-     "\xB4\x69\x01\x00\xB5\x69\x01\x00\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xB3\x49"
-     "\x02\x00"
-     "\xA5\xB6\x00\x80\xA9\xB6\x00\x80\xBE\x80\x04\x00\xAD\xB6\x00\x80\xB6\x4D"
-     "\x02\x00"
-     "\xB5\x4D\x02\x00\xB1\xB6\x00\x80\xBB\xAD\x02\x00\xBA\xAD\x02\x00\xB5\xB6"
-     "\x00\x80"
-     "\xB9\xB6\x00\x80\xBF\xAD\x02\x00\xBE\xAD\x02\x00\xBD\xAD\x02\x00\xBC\xAD"
-     "\x02\x00"
-     "\xBD\xB6\x00\x80\xA3\x0D\x02\x00\xC1\xB6\x00\x80\xC5\xB6\x00\x80\xA6\x09"
-     "\x02\x00"
-     "\xC9\xB6\x00\x80\xCD\xB6\x00\x80\xA5\x09\x02\x00\xAA\xE9\x02\x00\xAB\xE9"
-     "\x02\x00"
-     "\xD1\xB6\x00\x80\xD5\xB6\x00\x80\xAE\xE9\x02\x00\xAF\xE9\x02\x00\xAC\xE9"
-     "\x02\x00"
-     "\xAD\xE9\x02\x00\xD9\xB6\x00\x80\xDD\xB6\x00\x80\xBE\x04\x05\x00\xBE\x14"
-     "\x06\x00"
-     "\xE1\xB6\x00\x80\xE9\xB6\x00\x80\xED\xB6\x00\x80\xF1\xB6\x00\x80\x80\x21"
-     "\x00\x00"
-     "\x81\x21\x00\x00\x82\x39\x00\x00\x84\x20\x03\x00\x86\xB0\x04\x00\x87\x54"
-     "\x03\x00"
-     "\xF5\xB6\x00\x80\xF9\xB6\x00\x80\xFD\xB6\x00\x80\x01\xB7\x00\x80\x05\xB7"
-     "\x00\x80"
-     "\x09\xB7\x00\x80\x0D\xB7\x00\x80\xEF\x20\x02\x00\x11\xB7\x00\x80\x15\xB7"
-     "\x00\x80"
-     "\x19\xB7\x00\x80\xE3\x58\x02\x00\x1D\xB7\x00\x80\xE1\x84\x01\x00\x21\xB7"
-     "\x00\x80"
-     "\x25\xB7\x00\x80\x29\xB7\x00\x80\x2D\xB7\x00\x80\xE1\x30\x01\x00\xE1\x34"
-     "\x1E\x00"
-     "\xE3\x7C\x00\x00\xE3\x18\x1E\x00\x31\xB7\x00\x80\x35\xB7\x00\x80\x39\xB7"
-     "\x00\x80"
-     "\x3D\xB7\x00\x80\x84\xA8\x05\x00\x41\xB7\x00\x80\x45\xB7\x00\x80\x49\xB7"
-     "\x00\x80"
-     "\x4D\xB7\x00\x80\x51\xB7\x00\x80\xEF\xA8\x1E\x00\xEF\x88\x1E\x00\xB3\x3D"
-     "\x02\x00"
-     "\x59\xB7\x00\x80\x86\x68\x04\x00\x87\xDC\x05\x00\xD4\x00\x00\x00\xB6\xFD"
-     "\x02\x00"
-     "\xB5\xFD\x02\x00\x5D\xB7\x00\x80\xBB\xD9\x02\x00\xBA\xD1\x02\x00\xBE\xA4"
-     "\x06\x00"
-     "\x61\xB7\x00\x80\xBF\x61\x01\x00\xBE\x7D\x01\x00\xBD\x7D\x01\x00\xBC\xC1"
-     "\x02\x00"
-     "\xBC\xFD\x1D\x00\xBD\x65\x1D\x00\xBE\x65\x1D\x00\xBF\x7D\x1D\x00\xB8\xF9"
-     "\x1D\x00"
-     "\xB9\xF9\x1D\x00\xBA\xE9\x1D\x00\xBB\xE9\x1D\x00\xB4\xA9\x1D\x00\xB5\xA9"
-     "\x1D\x00"
-     "\xB6\x81\x1D\x00\xB7\x81\x1D\x00\xB0\x51\x1E\x00\xB1\x51\x1E\x00\xB2\x51"
-     "\x1E\x00"
-     "\xB3\x51\x1E\x00\xAC\xBD\x1E\x00\xAD\x41\x1E\x00\xAE\x11\x1E\x00\xAF\x11"
-     "\x1E\x00"
-     "\xA8\x3D\x1F\x00\xA9\x85\x1E\x00\xAA\x85\x1E\x00\xAB\xA5\x1E\x00\x80\xCD"
-     "\x00\x00"
-     "\x81\xD5\x00\x00\x82\xDD\x00\x00\x65\xB7\x00\x80\x69\xB7\x00\x80\x6D\xB7"
-     "\x00\x80"
-     "\x71\xB7\x00\x80\x75\xB7\x00\x80\xA3\x7D\x05\x00\x79\xB7\x00\x80\x7D\xB7"
-     "\x00\x80"
-     "\x81\xB7\x00\x80\x85\xB7\x00\x80\xA6\xBD\x05\x00\xA5\xBD\x05\x00\x89\xB7"
-     "\x00\x80"
-     "\xAB\x99\x05\x00\xAA\x91\x05\x00\x8D\xB7\x00\x80\x91\xB7\x00\x80\xAF\x21"
-     "\x06\x00"
-     "\xAE\x3D\x06\x00\xAD\x3D\x06\x00\xAC\x81\x05\x00\x95\xB7\x00\x80\xB3\xD9"
-     "\x06\x00"
-     "\x99\xB7\x00\x80\x9D\xB7\x00\x80\xB6\x91\x06\x00\xA1\xB7\x00\x80\xA5\xB7"
-     "\x00\x80"
-     "\xB5\xE1\x06\x00\xBA\xAD\x06\x00\xBB\xB5\x06\x00\xA9\xB7\x00\x80\xAD\xB7"
-     "\x00\x80"
-     "\xBE\x91\x06\x00\xBF\x9D\x06\x00\xBC\xA1\x06\x00\xBD\xA1\x06\x00\xE5\xB6"
-     "\x00\x80"
-     "\x55\xB7\x00\x80\xB1\xB7\x00\x80\xB5\xB7\x00\x80\x80\x0D\x00\x00\x81\x91"
-     "\x01\x00"
-     "\x82\x91\x01\x00\xB9\xB7\x00\x80\xA8\x35\x06\x00\xA9\x3D\x06\x00\xAA\x35"
-     "\x06\x00"
-     "\xAB\x15\x06\x00\xAC\x0D\x06\x00\xAD\x75\x06\x00\xAE\x7D\x06\x00\xAF\x75"
-     "\x06\x00"
-     "\xB0\x0D\x06\x00\xB1\x15\x06\x00\xB2\x11\x06\x00\xB3\x11\x06\x00\xB4\x31"
-     "\x06\x00"
-     "\xB5\x3D\x06\x00\xB6\x29\x06\x00\xB7\x29\x06\x00\xB8\x19\x06\x00\xB9\x19"
-     "\x06\x00"
-     "\xBA\xF9\x07\x00\xBB\xF9\x07\x00\xBC\xE9\x07\x00\xBD\xE9\x07\x00\xBE\xDD"
-     "\x07\x00"
-     "\xBF\xD5\x07\x00\xA3\x9D\x07\x00\xBD\xB7\x00\x80\xC1\xB7\x00\x80\x84\x84"
-     "\x02\x00"
-     "\xC5\xB7\x00\x80\xA6\xD5\x07\x00\xA5\xA5\x07\x00\xC9\xB7\x00\x80\xAB\xF1"
-     "\x07\x00"
-     "\xAA\xE9\x07\x00\x86\x00\x0C\x00\x87\x7C\x01\x00\xAF\xD9\x07\x00\xAE\xD5"
-     "\x07\x00"
-     "\xAD\xE5\x07\x00\xAC\xE5\x07\x00\xCD\xB7\x00\x80\xB3\x29\x06\x00\xD1\xB7"
-     "\x00\x80"
-     "\xD5\xB7\x00\x80\xB6\xA5\x06\x00\xD9\xB7\x00\x80\xDD\xB7\x00\x80\xB5\xAD"
-     "\x06\x00"
-     "\xBA\x65\x01\x00\xBB\x6D\x01\x00\xE1\xB7\x00\x80\xE5\xB7\x00\x80\xBE\x6D"
-     "\x01\x00"
-     "\xBF\x5D\x01\x00\xBC\x75\x01\x00\xBD\x6D\x01\x00\xA8\xC5\x06\x00\xA9\xD9"
-     "\x06\x00"
-     "\xAA\xE5\x06\x00\xAB\x61\x06\x00\xAC\x61\x06\x00\xAD\x61\x06\x00\xAE\x61"
-     "\x06\x00"
-     "\xAF\x61\x06\x00\xE9\xB7\x00\x80\xED\xB7\x00\x80\xF1\xB7\x00\x80\xF5\xB7"
-     "\x00\x80"
-     "\xF9\xB7\x00\x80\xFD\xB7\x00\x80\x01\xB8\x00\x80\x05\xB8\x00\x80\xB8\x89"
-     "\x01\x00"
-     "\xB9\x89\x01\x00\xBA\x9D\x01\x00\xBB\x95\x01\x00\xBC\xB9\x01\x00\xBD\xB9"
-     "\x01\x00"
-     "\xBE\x51\x01\x00\xBF\x51\x01\x00\xB0\xE5\x01\x00\xB1\xED\x01\x00\xB2\xE5"
-     "\x01\x00"
-     "\xB3\xFD\x01\x00\xB4\xED\x01\x00\xB5\xDD\x01\x00\xB6\xD5\x01\x00\xB7\xBD"
-     "\x01\x00"
-     "\xA6\xF5\x05\x00\x09\xB8\x00\x80\x0D\xB8\x00\x80\xA5\xFD\x05\x00\x11\xB8"
-     "\x00\x80"
-     "\xA3\x79\x05\x00\x15\xB8\x00\x80\x19\xB8\x00\x80\xAE\x3D\x02\x00\xAF\x0D"
-     "\x02\x00"
-     "\xAC\x25\x02\x00\xAD\x3D\x02\x00\xAA\x35\x02\x00\xAB\x3D\x02\x00\x1D\xB8"
-     "\x00\x80"
-     "\x21\xB8\x00\x80\x25\xB8\x00\x80\x29\xB8\x00\x80\x2D\xB8\x00\x80\x31\xB8"
-     "\x00\x80"
-     "\x80\x15\x00\x00\x81\x1D\x00\x00\x82\x15\x00\x00\x39\xB8\x00\x80\x84\x60"
-     "\x02\x00"
-     "\x3D\xB8\x00\x80\x87\x48\x03\x00\x86\xCC\x0C\x00\x41\xB8\x00\x80\x45\xB8"
-     "\x00\x80"
-     "\x49\xB8\x00\x80\x4D\xB8\x00\x80\x51\xB8\x00\x80\x55\xB8\x00\x80\x86\x28"
-     "\x0D\x00"
-     "\x87\x40\x0C\x00\xA1\x00\x00\x00\xE1\xA4\x01\x00\x59\xB8\x00\x80\xE3\xEC"
-     "\x02\x00"
-     "\x5D\xB8\x00\x80\x61\xB8\x00\x80\x65\xB8\x00\x80\xEF\x3C\x07\x00\x69\xB8"
-     "\x00\x80"
-     "\x6D\xB8\x00\x80\x71\xB8\x00\x80\x75\xB8\x00\x80\x79\xB8\x00\x80\x7D\xB8"
-     "\x00\x80"
-     "\x81\xB8\x00\x80\xB3\x85\x03\x00\x85\xB8\x00\x80\xB5\x81\x03\x00\xB6\x81"
-     "\x03\x00"
-     "\x89\xB8\x00\x80\x8D\xB8\x00\x80\x91\xB8\x00\x80\xBA\x49\x03\x00\xBB\x49"
-     "\x03\x00"
-     "\xBC\x59\x03\x00\xBD\x59\x03\x00\xBE\x49\x03\x00\xBF\x49\x03\x00\xE1\x50"
-     "\x06\x00"
-     "\xE1\xAC\x07\x00\xE3\x14\x00\x00\xE3\x3C\x06\x00\x95\xB8\x00\x80\x99\xB8"
-     "\x00\x80"
-     "\x9D\xB8\x00\x80\x84\x54\x0D\x00\x80\x35\x00\x00\x81\x3D\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xA1\xB8\x00\x80\xBE\x7C\x0D\x00\xA9\xB8\x00\x80\xEF\x4C\x07\x00\xEF\x4C"
-     "\x07\x00"
-     "\xA3\x09\x02\x00\xAD\xB8\x00\x80\xB1\xB8\x00\x80\xB5\xB8\x00\x80\xB9\xB8"
-     "\x00\x80"
-     "\xA6\x0D\x02\x00\xA5\x0D\x02\x00\xBD\xB8\x00\x80\xAB\xC5\x02\x00\xAA\xC5"
-     "\x02\x00"
-     "\xC1\xB8\x00\x80\xC5\xB8\x00\x80\xAF\xC5\x02\x00\xAE\xC5\x02\x00\xAD\xD5"
-     "\x02\x00"
-     "\xAC\xD5\x02\x00\xBC\xD5\x05\x00\xBD\xDD\x05\x00\xBE\xE1\x05\x00\xBF\xE1"
-     "\x05\x00"
-     "\xB8\xBD\x05\x00\xB9\xD5\x05\x00\xBA\xD1\x05\x00\xBB\xD1\x05\x00\xB4\xA9"
-     "\x05\x00"
-     "\xB5\xB5\x05\x00\xB6\xB5\x05\x00\xB7\x85\x05\x00\xB0\x05\x06\x00\xB1\x0D"
-     "\x06\x00"
-     "\xB2\x05\x06\x00\xB3\xB9\x05\x00\xAC\x65\x06\x00\xAD\x6D\x06\x00\xAE\x65"
-     "\x06\x00"
-     "\xAF\x7D\x06\x00\xA8\xBD\x06\x00\xA9\xAD\x06\x00\xAA\xA5\x06\x00\xAB\x7D"
-     "\x06\x00"
-     "\xC9\xB8\x00\x80\xCD\xB8\x00\x80\xD1\xB8\x00\x80\xD5\xB8\x00\x80\xD9\xB8"
-     "\x00\x80"
-     "\xDD\xB8\x00\x80\xE1\xB8\x00\x80\xE5\xB8\x00\x80\xA8\x5D\x0E\x00\xA9\x61"
-     "\x0E\x00"
-     "\xAA\x79\x0E\x00\xAB\x79\x0E\x00\xAC\x69\x0E\x00\xAD\x69\x0E\x00\xAE\x99"
-     "\x0E\x00"
-     "\xAF\x99\x0E\x00\xE9\xB8\x00\x80\xED\xB8\x00\x80\xF1\xB8\x00\x80\xF5\xB8"
-     "\x00\x80"
-     "\xF9\xB8\x00\x80\xFD\xB8\x00\x80\x01\xB9\x00\x80\x05\xB9\x00\x80\xB8\x81"
-     "\x0E\x00"
-     "\xB9\x81\x0E\x00\xBA\x81\x0E\x00\xBB\x9D\x0E\x00\xBC\x8D\x0E\x00\xBD\xB5"
-     "\x0E\x00"
-     "\xBE\xBD\x0E\x00\xBF\x4D\x0F\x00\xB0\xF9\x0E\x00\xB1\xC5\x0E\x00\xB2\xCD"
-     "\x0E\x00"
-     "\xB3\xC5\x0E\x00\xB4\xC1\x0E\x00\xB5\xC1\x0E\x00\xB6\xC1\x0E\x00\xB7\xC1"
-     "\x0E\x00"
-     "\x35\xB8\x00\x80\xB3\xF5\x0E\x00\x09\xB9\x00\x80\xA5\xB8\x00\x80\xB6\xED"
-     "\x0E\x00"
-     "\x0D\xB9\x00\x80\x11\xB9\x00\x80\xB5\xE5\x0E\x00\xBA\x39\x0E\x00\xBB\x05"
-     "\x0E\x00"
-     "\x84\xE4\x03\x00\xBE\xE4\x03\x00\xBE\x01\x0E\x00\xBF\x01\x0E\x00\xBC\x1D"
-     "\x0E\x00"
-     "\xBD\x15\x0E\x00\xA3\xB1\x0E\x00\x82\x5D\x00\x00\x81\x5D\x00\x00\x80\x45"
-     "\x00\x00"
-     "\x15\xB9\x00\x80\xA6\xA9\x0E\x00\xA5\xA1\x0E\x00\x19\xB9\x00\x80\xAB\x41"
-     "\x0E\x00"
-     "\xAA\x7D\x0E\x00\x86\x48\x00\x00\x87\x68\x00\x00\xAF\x45\x0E\x00\xAE\x45"
-     "\x0E\x00"
-     "\xAD\x51\x0E\x00\xAC\x59\x0E\x00\xA8\x35\x0E\x00\xA9\x41\x0E\x00\xAA\x41"
-     "\x0E\x00"
-     "\xAB\x41\x0E\x00\xAC\x41\x0E\x00\xAD\x49\x0E\x00\xAE\x71\x0E\x00\xAF\x71"
-     "\x0E\x00"
-     "\x1D\xB9\x00\x80\x21\xB9\x00\x80\x84\x6C\x01\x00\x25\xB9\x00\x80\x29\xB9"
-     "\x00\x80"
-     "\x2D\xB9\x00\x80\x31\xB9\x00\x80\x35\xB9\x00\x80\xB8\x61\x01\x00\xB9\x61"
-     "\x01\x00"
-     "\xBA\x61\x01\x00\xBB\x61\x01\x00\xBC\x61\x01\x00\xBD\x61\x01\x00\xBE\x61"
-     "\x01\x00"
-     "\xBF\x61\x01\x00\xB0\x81\x0E\x00\xB1\x81\x0E\x00\xB2\x81\x0E\x00\xB3\x81"
-     "\x0E\x00"
-     "\xB4\x81\x0E\x00\xB5\x81\x0E\x00\xB6\x81\x0E\x00\xB7\x81\x0E\x00\xB3\xC9"
-     "\x0E\x00"
-     "\x39\xB9\x00\x80\x3D\xB9\x00\x80\x41\xB9\x00\x80\x45\xB9\x00\x80\xB6\xC9"
-     "\x0E\x00"
-     "\xB5\xC9\x0E\x00\x49\xB9\x00\x80\xBB\x45\x0E\x00\xBA\x45\x0E\x00\x4D\xB9"
-     "\x00\x80"
-     "\x51\xB9\x00\x80\xBF\x45\x0E\x00\xBE\x45\x0E\x00\xBD\x55\x0E\x00\xBC\x55"
-     "\x0E\x00"
-     "\x55\xB9\x00\x80\xA3\x8D\x0E\x00\x59\xB9\x00\x80\x5D\xB9\x00\x80\xA6\x8D"
-     "\x0E\x00"
-     "\x61\xB9\x00\x80\x65\xB9\x00\x80\xA5\x8D\x0E\x00\xAA\x01\x0E\x00\xAB\x01"
-     "\x0E\x00"
-     "\x69\xB9\x00\x80\x6D\xB9\x00\x80\xAE\x01\x0E\x00\xAF\x01\x0E\x00\xAC\x11"
-     "\x0E\x00"
-     "\xAD\x11\x0E\x00\xA8\xDD\x02\x00\xA9\x25\x03\x00\xAA\x2D\x03\x00\xAB\x21"
-     "\x03\x00"
-     "\xAC\x21\x03\x00\xAD\x21\x03\x00\xAE\x21\x03\x00\xAF\x21\x03\x00\x71\xB9"
-     "\x00\x80"
-     "\x75\xB9\x00\x80\x79\xB9\x00\x80\x7D\xB9\x00\x80\x81\xB9\x00\x80\xBE\x78"
-     "\x02\x00"
-     "\x89\xB9\x00\x80\x8D\xB9\x00\x80\xB8\x89\x03\x00\xB9\x89\x03\x00\xBA\x9D"
-     "\x03\x00"
-     "\xBB\x95\x03\x00\xBC\xB9\x03\x00\xBD\xB9\x03\x00\xBE\x79\x00\x00\xBF\x79"
-     "\x00\x00"
-     "\xB0\xE5\x03\x00\xB1\xED\x03\x00\xB2\xE5\x03\x00\xB3\xFD\x03\x00\xB4\xED"
-     "\x03\x00"
-     "\xB5\xDD\x03\x00\xB6\xD5\x03\x00\xB7\xBD\x03\x00\x81\xAD\x00\x00\x80\xAD"
-     "\x00\x00"
-     "\xB3\xAD\x03\x00\x82\xBD\x00\x00\xB5\xA1\x03\x00\x91\xB9\x00\x80\x95\xB9"
-     "\x00\x80"
-     "\xB6\xA9\x03\x00\x86\x00\x04\x00\x87\xD4\x03\x00\xBB\x21\x03\x00\xBA\x29"
-     "\x03\x00"
-     "\xBD\x21\x03\x00\xBC\x39\x03\x00\xBF\x11\x03\x00\xBE\x21\x03\x00\xA3\xE5"
-     "\x03\x00"
-     "\x99\xB9\x00\x80\x9D\xB9\x00\x80\x85\xF0\x30\x00\xA1\xB9\x00\x80\xA6\xE1"
-     "\x03\x00"
-     "\xA5\xE9\x03\x00\xA5\xB9\x00\x80\xAB\x69\x03\x00\xAA\x61\x03\x00\xA9\xB9"
-     "\x00\x80"
-     "\xAD\xB9\x00\x80\xAF\x59\x03\x00\xAE\x69\x03\x00\xAD\x69\x03\x00\xAC\x71"
-     "\x03\x00"
-     "\xE3\xB8\x01\x00\xE1\x70\x0F\x00\xE1\x98\x0F\x00\xE3\x00\x0E\x00\xB1\xB9"
-     "\x00\x80"
-     "\xB5\xB9\x00\x80\xB9\xB9\x00\x80\xBD\xB9\x00\x80\xC1\xB9\x00\x80\xC5\xB9"
-     "\x00\x80"
-     "\xC9\xB9\x00\x80\xCD\xB9\x00\x80\xEF\x64\x01\x00\xD1\xB9\x00\x80\xBF\x30"
-     "\x5D\x00"
-     "\xEF\xF4\x0E\x00\xD5\xB9\x00\x80\xD9\xB9\x00\x80\x84\x04\x05\x00\xBE\x04"
-     "\x05\x00"
-     "\xDD\xB9\x00\x80\xE1\xA4\x01\x00\xE1\xB9\x00\x80\xE3\x64\x00\x00\xE5\xB9"
-     "\x00\x80"
-     "\xE9\xB9\x00\x80\xED\xB9\x00\x80\xEF\xA0\x03\x00\x80\x15\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x82\x05\x00\x00\xF1\xB9\x00\x80\xAA\xC9\x0E\x00\xAB\xC9\x0E\x00\xA8\x95"
-     "\x0E\x00"
-     "\xA9\xCD\x0E\x00\xAE\xF9\x0E\x00\xAF\xF9\x0E\x00\xAC\xCD\x0E\x00\xAD\xC5"
-     "\x0E\x00"
-     "\x86\x00\x04\x00\x87\x84\x04\x00\xF9\xB9\x00\x80\xFD\xB9\x00\x80\x01\xBA"
-     "\x00\x80"
-     "\x05\xBA\x00\x80\x09\xBA\x00\x80\x0D\xBA\x00\x80\xBA\xD1\x0D\x00\xBB\xDD"
-     "\x0D\x00"
-     "\xB8\xCD\x0D\x00\xB9\xC5\x0D\x00\xBE\xED\x0D\x00\xBF\x95\x0D\x00\xBC\xC5"
-     "\x0D\x00"
-     "\xBD\xE5\x0D\x00\xB2\x49\x0E\x00\xB3\x49\x0E\x00\xB0\x49\x0E\x00\xB1\x49"
-     "\x0E\x00"
-     "\xB6\x49\x0E\x00\xB7\x49\x0E\x00\xB4\x49\x0E\x00\xB5\x49\x0E\x00\xA8\x39"
-     "\x06\x00"
-     "\xA9\x39\x06\x00\xAA\x9D\x06\x00\xAB\x95\x06\x00\xAC\xD5\x06\x00\xAD\xDD"
-     "\x06\x00"
-     "\xAE\xC9\x06\x00\xAF\xC1\x06\x00\xF5\xB9\x00\x80\x11\xBA\x00\x80\x15\xBA"
-     "\x00\x80"
-     "\x19\xBA\x00\x80\x1D\xBA\x00\x80\x21\xBA\x00\x80\x25\xBA\x00\x80\x29\xBA"
-     "\x00\x80"
-     "\xB8\x69\x07\x00\xB9\x69\x07\x00\xBA\x7D\x07\x00\xBB\x0D\x07\x00\xBC\x15"
-     "\x07\x00"
-     "\xBD\x05\x07\x00\xBE\x01\x07\x00\xBF\x3D\x07\x00\xB0\x89\x06\x00\xB1\x89"
-     "\x06\x00"
-     "\xB2\x79\x07\x00\xB3\x79\x07\x00\xB4\x69\x07\x00\xB5\x69\x07\x00\xB6\x59"
-     "\x07\x00"
-     "\xB7\x59\x07\x00\xB3\x15\x06\x00\x2D\xBA\x00\x80\x31\xBA\x00\x80\x35\xBA"
-     "\x00\x80"
-     "\x39\xBA\x00\x80\xB6\x5D\x06\x00\xB5\x5D\x06\x00\x3D\xBA\x00\x80\xBB\x4D"
-     "\x06\x00"
-     "\xBA\x4D\x06\x00\x41\xBA\x00\x80\x45\xBA\x00\x80\xBF\xAD\x07\x00\xBE\xAD"
-     "\x07\x00"
-     "\xBD\xAD\x07\x00\xBC\xAD\x07\x00\x49\xBA\x00\x80\xA3\x51\x06\x00\x4D\xBA"
-     "\x00\x80"
-     "\x51\xBA\x00\x80\xA6\x19\x06\x00\x55\xBA\x00\x80\x59\xBA\x00\x80\xA5\x19"
-     "\x06\x00"
-     "\xAA\x09\x06\x00\xAB\x09\x06\x00\x5D\xBA\x00\x80\x61\xBA\x00\x80\xAE\xE9"
-     "\x07\x00"
-     "\xAF\xE9\x07\x00\xAC\xE9\x07\x00\xAD\xE9\x07\x00\xA8\xA9\x06\x00\xA9\xA9"
-     "\x06\x00"
-     "\xAA\xB9\x06\x00\xAB\xB9\x06\x00\xAC\xA9\x06\x00\xAD\xA9\x06\x00\xAE\x3D"
-     "\x01\x00"
-     "\xAF\x35\x01\x00\x80\xF9\x01\x00\x81\xCD\x01\x00\x82\xC5\x01\x00\x84\x78"
-     "\x02\x00"
-     "\xBE\x7C\x02\x00\x65\xBA\x00\x80\x87\x78\x01\x00\x86\x9C\x00\x00\xB8\xD1"
-     "\x01\x00"
-     "\xB9\xD9\x01\x00\xBA\xE1\x01\x00\xBB\xE1\x01\x00\xBC\x91\x01\x00\xBD\x9D"
-     "\x01\x00"
-     "\xBE\x95\x01\x00\xBF\x89\x01\x00\xB0\x4D\x01\x00\xB1\x55\x01\x00\xB2\x5D"
-     "\x01\x00"
-     "\xB3\x55\x01\x00\xB4\x4D\x01\x00\xB5\xF1\x01\x00\xB6\xF1\x01\x00\xB7\xF1"
-     "\x01\x00"
-     "\xB5\x39\x06\x00\x69\xBA\x00\x80\x6D\xBA\x00\x80\xB6\x31\x06\x00\x71\xBA"
-     "\x00\x80"
-     "\x75\xBA\x00\x80\xB3\x29\x06\x00\x79\xBA\x00\x80\xBD\x99\x01\x00\xBC\x91"
-     "\x01\x00"
-     "\xBF\x2D\x01\x00\xBE\x91\x01\x00\x7D\xBA\x00\x80\x85\xB9\x00\x80\xBB\x89"
-     "\x01\x00"
-     "\xBA\x91\x01\x00\xA3\x71\x06\x00\x81\xBA\x00\x80\x85\xBA\x00\x80\x89\xBA"
-     "\x00\x80"
-     "\x8D\xBA\x00\x80\xA6\x69\x06\x00\xA5\x61\x06\x00\x91\xBA\x00\x80\xAB\xD1"
-     "\x01\x00"
-     "\xAA\xC9\x01\x00\x95\xBA\x00\x80\x99\xBA\x00\x80\xAF\x75\x01\x00\xAE\xC9"
-     "\x01\x00"
-     "\xAD\xC1\x01\x00\xAC\xC9\x01\x00\x9D\xBA\x00\x80\xA1\xBA\x00\x80\xEF\xBC"
-     "\x07\x00"
-     "\xA5\xBA\x00\x80\xA9\xBA\x00\x80\xAD\xBA\x00\x80\xB1\xBA\x00\x80\xB5\xBA"
-     "\x00\x80"
-     "\xB9\xBA\x00\x80\xBD\xBA\x00\x80\xC1\xBA\x00\x80\xC5\xBA\x00\x80\xE1\x88"
-     "\x06\x00"
-     "\xC9\xBA\x00\x80\xE3\x60\x07\x00\xCD\xBA\x00\x80\x80\x75\x00\x00\x81\x79"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x84\x40\x02\x00\xD1\xBA\x00\x80\xD9\xBA\x00\x80\xBE\x14"
-     "\x03\x00"
-     "\xDD\xBA\x00\x80\x86\xC0\x3C\x00\x87\x44\x03\x00\xE1\xBA\x00\x80\xE5\xBA"
-     "\x00\x80"
-     "\xE9\xBA\x00\x80\xED\xBA\x00\x80\xF1\xBA\x00\x80\xF5\xBA\x00\x80\xF9\xBA"
-     "\x00\x80"
-     "\xFD\xBA\x00\x80\xEF\x28\x02\x00\x01\xBB\x00\x80\x05\xBB\x00\x80\x09\xBB"
-     "\x00\x80"
-     "\x0D\xBB\x00\x80\x11\xBB\x00\x80\x84\xEC\x3C\x00\x15\xBB\x00\x80\x19\xBB"
-     "\x00\x80"
-     "\x1D\xBB\x00\x80\xE1\x94\x01\x00\x21\xBB\x00\x80\xE3\xE4\x02\x00\x25\xBB"
-     "\x00\x80"
-     "\xB3\x51\x03\x00\x29\xBB\x00\x80\x2D\xBB\x00\x80\x31\xBB\x00\x80\x35\xBB"
-     "\x00\x80"
-     "\xB6\x65\x03\x00\xB5\x45\x03\x00\x39\xBB\x00\x80\xBB\x31\x03\x00\xBA\x29"
-     "\x03\x00"
-     "\x3D\xBB\x00\x80\x41\xBB\x00\x80\xBF\xF1\x00\x00\xBE\x19\x03\x00\xBD\x11"
-     "\x03\x00"
-     "\xBC\x21\x03\x00\xA8\x85\x02\x00\xA9\x8D\x02\x00\xAA\x99\x02\x00\xAB\x99"
-     "\x02\x00"
-     "\xAC\x8D\x02\x00\xAD\xA5\x02\x00\xAE\xDD\x02\x00\xAF\xD5\x02\x00\x80\xFD"
-     "\x01\x00"
-     "\x81\xCD\x01\x00\x82\xC5\x01\x00\xBE\x64\x3D\x00\x86\x90\x3C\x00\x87\xB8"
-     "\x3D\x00"
-     "\x49\xBB\x00\x80\x84\x60\x3C\x00\xB8\x51\x01\x00\xB9\x51\x01\x00\xBA\x51"
-     "\x01\x00"
-     "\xBB\x51\x01\x00\xBC\x71\x01\x00\xBD\x71\x01\x00\xBE\x71\x01\x00\xBF\x71"
-     "\x01\x00"
-     "\xB0\xAD\x02\x00\xB1\xB5\x02\x00\xB2\xB5\x02\x00\xB3\x85\x02\x00\xB4\x9D"
-     "\x02\x00"
-     "\xB5\x75\x01\x00\xB6\x7D\x01\x00\xB7\x71\x01\x00\x4D\xBB\x00\x80\x51\xBB"
-     "\x00\x80"
-     "\x55\xBB\x00\x80\x59\xBB\x00\x80\x5D\xBB\x00\x80\x61\xBB\x00\x80\x65\xBB"
-     "\x00\x80"
-     "\xEF\x54\x00\x00\xBE\xEC\x3C\x00\xE1\x1C\x06\x00\x69\xBB\x00\x80\xE3\x90"
-     "\x00\x00"
-     "\x6D\xBB\x00\x80\x71\xBB\x00\x80\x75\xBB\x00\x80\x79\xBB\x00\x80\xA3\x51"
-     "\x02\x00"
-     "\x7D\xBB\x00\x80\x84\x44\x3F\x00\x81\xBB\x00\x80\x85\xBB\x00\x80\xA6\x65"
-     "\x02\x00"
-     "\xA5\x45\x02\x00\x89\xBB\x00\x80\xAB\x31\x02\x00\xAA\x29\x02\x00\x8D\xBB"
-     "\x00\x80"
-     "\x91\xBB\x00\x80\xAF\xF1\x01\x00\xAE\x19\x02\x00\xAD\x11\x02\x00\xAC\x21"
-     "\x02\x00"
-     "\xA8\x55\x3E\x00\xA9\x59\x3E\x00\xAA\x61\x3E\x00\xAB\x61\x3E\x00\xAC\x61"
-     "\x3E\x00"
-     "\xAD\x61\x3E\x00\xAE\x61\x3E\x00\xAF\x61\x3E\x00\x45\xBB\x00\x80\x95\xBB"
-     "\x00\x80"
-     "\x99\xBB\x00\x80\x9D\xBB\x00\x80\xA1\xBB\x00\x80\xA5\xBB\x00\x80\xA9\xBB"
-     "\x00\x80"
-     "\xAD\xBB\x00\x80\xB8\xE9\x3E\x00\xB9\xE9\x3E\x00\xBA\xFD\x3E\x00\xBB\xF1"
-     "\x3E\x00"
-     "\xBC\x95\x3E\x00\xBD\x9D\x3E\x00\xBE\x95\x3E\x00\xBF\x8D\x3E\x00\xB0\xE1"
-     "\x3E\x00"
-     "\xB1\xE1\x3E\x00\xB2\xE5\x3E\x00\xB3\xFD\x3E\x00\xB4\xE5\x3E\x00\xB5\xED"
-     "\x3E\x00"
-     "\xB6\xE5\x3E\x00\xB7\xD9\x3E\x00\xB3\xE9\x3E\x00\x82\x15\x00\x00\x81\xB5"
-     "\x00\x00"
-     "\x80\xB5\x00\x00\xB1\xBB\x00\x80\xB6\xE1\x3E\x00\xB5\xE9\x3E\x00\xB5\xBB"
-     "\x00\x80"
-     "\xBB\x25\x3E\x00\xBA\x25\x3E\x00\x86\xA8\x00\x00\x87\xA4\x00\x00\xBF\x29"
-     "\x3E\x00"
-     "\xBE\x21\x3E\x00\xBD\x29\x3E\x00\xBC\x31\x3E\x00\xB9\xBB\x00\x80\xA3\xAD"
-     "\x3E\x00"
-     "\xBD\xBB\x00\x80\xC1\xBB\x00\x80\xA6\xA5\x3E\x00\xC5\xBB\x00\x80\xC9\xBB"
-     "\x00\x80"
-     "\xA5\xAD\x3E\x00\xAA\x61\x3E\x00\xAB\x61\x3E\x00\xCD\xBB\x00\x80\xD1\xBB"
-     "\x00\x80"
-     "\xAE\x65\x3E\x00\xAF\x6D\x3E\x00\xAC\x75\x3E\x00\xAD\x6D\x3E\x00\xA8\x39"
-     "\x3E\x00"
-     "\xA9\x85\x3E\x00\xAA\x8D\x3E\x00\xAB\x85\x3E\x00\xAC\x9D\x3E\x00\xAD\x85"
-     "\x3E\x00"
-     "\xAE\x8D\x3E\x00\xAF\x85\x3E\x00\xD5\xBB\x00\x80\xD9\xBB\x00\x80\xDD\xBB"
-     "\x00\x80"
-     "\xE1\xBB\x00\x80\xE5\xBB\x00\x80\xE9\xBB\x00\x80\xED\xBB\x00\x80\xF1\xBB"
-     "\x00\x80"
-     "\xB8\x9D\x3E\x00\xB9\xAD\x3E\x00\xBA\xA5\x3E\x00\xBB\x69\x01\x00\xBC\x79"
-     "\x01\x00"
-     "\xBD\x79\x01\x00\xBE\x69\x01\x00\xBF\x69\x01\x00\xB0\xC9\x3E\x00\xB1\xD5"
-     "\x3E\x00"
-     "\xB2\xDD\x3E\x00\xB3\xD5\x3E\x00\xB4\xCD\x3E\x00\xB5\xB1\x3E\x00\xB6\xB5"
-     "\x3E\x00"
-     "\xB7\xAD\x3E\x00\xF5\xBB\x00\x80\xB6\x15\x3E\x00\xB5\x05\x3E\x00\xF9\xBB"
-     "\x00\x80"
-     "\xB3\x15\x3E\x00\xFD\xBB\x00\x80\x01\xBC\x00\x80\x05\xBC\x00\x80\xBF\x65"
-     "\x3E\x00"
-     "\xBE\x19\x3E\x00\xBD\x11\x3E\x00\xBC\x19\x3E\x00\xBB\x01\x3E\x00\xBA\x1D"
-     "\x3E\x00"
-     "\xD5\xBA\x00\x80\x09\xBC\x00\x80\xA3\x45\x3E\x00\x0D\xBC\x00\x80\x11\xBC"
-     "\x00\x80"
-     "\x15\xBC\x00\x80\x19\xBC\x00\x80\xA6\x45\x3E\x00\xA5\x55\x3E\x00\x1D\xBC"
-     "\x00\x80"
-     "\xAB\x51\x3E\x00\xAA\x4D\x3E\x00\x21\xBC\x00\x80\xBE\x2C\x00\x00\xAF\x35"
-     "\x3E\x00"
-     "\xAE\x49\x3E\x00\xAD\x41\x3E\x00\xAC\x49\x3E\x00\x82\xE9\x03\x00\xEF\xD0"
-     "\x3F\x00"
-     "\x80\xD9\x03\x00\x81\xD1\x03\x00\x87\x50\x03\x00\x86\xB4\x04\x00\x29\xBC"
-     "\x00\x80"
-     "\x2D\xBC\x00\x80\x31\xBC\x00\x80\x35\xBC\x00\x80\x39\xBC\x00\x80\x3D\xBC"
-     "\x00\x80"
-     "\x41\xBC\x00\x80\xE3\x84\x3E\x00\x45\xBC\x00\x80\xE1\x44\x3E\x00\xB3\x8D"
-     "\x02\x00"
-     "\x49\xBC\x00\x80\x84\x44\x03\x00\xBE\x40\x04\x00\x4D\xBC\x00\x80\xB6\xBD"
-     "\x02\x00"
-     "\xB5\x81\x02\x00\x51\xBC\x00\x80\xBB\x49\x03\x00\xBA\x41\x03\x00\x55\xBC"
-     "\x00\x80"
-     "\x59\xBC\x00\x80\xBF\x49\x03\x00\xBE\x41\x03\x00\xBD\x49\x03\x00\xBC\x51"
-     "\x03\x00"
-     "\x5D\xBC\x00\x80\x61\xBC\x00\x80\x65\xBC\x00\x80\xBE\x80\x05\x00\x69\xBC"
-     "\x00\x80"
-     "\x6D\xBC\x00\x80\x71\xBC\x00\x80\x75\xBC\x00\x80\xA8\x61\x02\x00\xA9\x6D"
-     "\x02\x00"
-     "\xAA\x65\x02\x00\xAB\x79\x02\x00\xAC\x69\x02\x00\xAD\x5D\x02\x00\xAE\x55"
-     "\x02\x00"
-     "\xAF\x9D\x03\x00\xB0\xE5\x03\x00\xB1\xED\x03\x00\xB2\xE5\x03\x00\xB3\xFD"
-     "\x03\x00"
-     "\xB4\xE5\x03\x00\xB5\xED\x03\x00\xB6\xE5\x03\x00\xB7\x5D\x03\x00\xB8\x6D"
-     "\x03\x00"
-     "\xB9\x71\x03\x00\xBA\x71\x03\x00\xBB\x71\x03\x00\xBC\x4D\x03\x00\xBD\x31"
-     "\x03\x00"
-     "\xBE\x31\x03\x00\xBF\x31\x03\x00\x79\xBC\x00\x80\x7D\xBC\x00\x80\x81\xBC"
-     "\x00\x80"
-     "\x85\xBC\x00\x80\xE3\xC8\x3E\x00\x84\x1C\x05\x00\xE1\xA8\x01\x00\x89\xBC"
-     "\x00\x80"
-     "\xEF\x34\x00\x00\x8D\xBC\x00\x80\x91\xBC\x00\x80\x99\xBC\x00\x80\x9D\xBC"
-     "\x00\x80"
-     "\xA1\xBC\x00\x80\xA5\xBC\x00\x80\xA9\xBC\x00\x80\x80\x5D\x00\x00\x81\x61"
-     "\x00\x00"
-     "\x82\x61\x00\x00\xA3\x81\x03\x00\xAD\xBC\x00\x80\xA5\x8D\x03\x00\xA6\xB1"
-     "\x03\x00"
-     "\x84\x98\x07\x00\x86\xC0\x04\x00\x87\x48\x05\x00\xAA\x4D\x02\x00\xAB\x45"
-     "\x02\x00"
-     "\xAC\x5D\x02\x00\xAD\x45\x02\x00\xAE\x4D\x02\x00\xAF\x45\x02\x00\xB1\xBC"
-     "\x00\x80"
-     "\xB5\xBC\x00\x80\xB9\xBC\x00\x80\xBD\xBC\x00\x80\xC1\xBC\x00\x80\xC5\xBC"
-     "\x00\x80"
-     "\xC9\xBC\x00\x80\xEF\xC4\x3E\x00\xBE\x20\x06\x00\xE1\x8C\x3E\x00\xCD\xBC"
-     "\x00\x80"
-     "\xE3\x80\x3E\x00\xD1\xBC\x00\x80\xD5\xBC\x00\x80\xD9\xBC\x00\x80\xDD\xBC"
-     "\x00\x80"
-     "\xB3\xE1\x01\x00\xE1\xBC\x00\x80\xE5\xBC\x00\x80\xE9\xBC\x00\x80\xED\xBC"
-     "\x00\x80"
-     "\xB6\xF9\x01\x00\xB5\xF1\x01\x00\xF1\xBC\x00\x80\xBB\xB1\x01\x00\xBA\xA1"
-     "\x01\x00"
-     "\xF5\xBC\x00\x80\xF9\xBC\x00\x80\xBF\x71\x01\x00\xBE\x99\x01\x00\xBD\x91"
-     "\x01\x00"
-     "\xBC\xA1\x01\x00\xA8\x19\x06\x00\xA9\x19\x06\x00\xAA\x29\x06\x00\xAB\x29"
-     "\x06\x00"
-     "\xAC\x39\x06\x00\xAD\x39\x06\x00\xAE\x29\x06\x00\xAF\x25\x06\x00\x95\xBC"
-     "\x00\x80"
-     "\x82\x11\x00\x00\x81\x19\x00\x00\x80\xC1\x07\x00\xFD\xBC\x00\x80\x01\xBD"
-     "\x00\x80"
-     "\x05\xBD\x00\x80\x09\xBD\x00\x80\xB8\xD1\x06\x00\xB9\xD1\x06\x00\xBA\xD1"
-     "\x06\x00"
-     "\xBB\xD1\x06\x00\xBC\x51\x07\x00\xBD\x51\x07\x00\xBE\x51\x07\x00\xBF\x51"
-     "\x07\x00"
-     "\xB0\x25\x06\x00\xB1\x35\x06\x00\xB2\x35\x06\x00\xB3\x05\x06\x00\xB4\x1D"
-     "\x06\x00"
-     "\xB5\xF5\x06\x00\xB6\xFD\x06\x00\xB7\xF5\x06\x00\xA3\xA1\x06\x00\x0D\xBD"
-     "\x00\x80"
-     "\x86\xAC\x00\x00\x87\x4C\x03\x00\x11\xBD\x00\x80\xA6\xB9\x06\x00\xA5\xB1"
-     "\x06\x00"
-     "\x15\xBD\x00\x80\xAB\xF1\x06\x00\xAA\xE1\x06\x00\x19\xBD\x00\x80\x1D\xBD"
-     "\x00\x80"
-     "\xAF\x31\x06\x00\xAE\xD9\x06\x00\xAD\xD1\x06\x00\xAC\xE1\x06\x00\x21\xBD"
-     "\x00\x80"
-     "\xB3\xF1\x07\x00\x25\xBD\x00\x80\x29\xBD\x00\x80\xB6\x81\x07\x00\x2D\xBD"
-     "\x00\x80"
-     "\x31\xBD\x00\x80\xB5\xE1\x07\x00\xBA\x69\x07\x00\xBB\x69\x07\x00\x35\xBD"
-     "\x00\x80"
-     "\x39\xBD\x00\x80\xBE\x69\x07\x00\xBF\x5D\x07\x00\xBC\x69\x07\x00\xBD\x69"
-     "\x07\x00"
-     "\xA8\x2D\x06\x00\xA9\x35\x06\x00\xAA\x3D\x06\x00\xAB\x35\x06\x00\xAC\x2D"
-     "\x06\x00"
-     "\xAD\x45\x06\x00\xAE\x99\x06\x00\xAF\x99\x06\x00\x3D\xBD\x00\x80\x41\xBD"
-     "\x00\x80"
-     "\x45\xBD\x00\x80\x49\xBD\x00\x80\x4D\xBD\x00\x80\x51\xBD\x00\x80\x55\xBD"
-     "\x00\x80"
-     "\x59\xBD\x00\x80\xB8\xF9\x06\x00\xB9\xF9\x06\x00\xBA\x8D\x06\x00\xBB\x85"
-     "\x06\x00"
-     "\xBC\x9D\x06\x00\xBD\x85\x06\x00\xBE\x85\x06\x00\xBF\xB5\x06\x00\xB0\xE9"
-     "\x06\x00"
-     "\xB1\xE9\x06\x00\xB2\xF9\x06\x00\xB3\xF9\x06\x00\xB4\xE9\x06\x00\xB5\xDD"
-     "\x06\x00"
-     "\xB6\xC9\x06\x00\xB7\xC9\x06\x00\xA3\xB5\x06\x00\x5D\xBD\x00\x80\x61\xBD"
-     "\x00\x80"
-     "\x65\xBD\x00\x80\x69\xBD\x00\x80\xA6\xC5\x06\x00\xA5\xA5\x06\x00\x6D\xBD"
-     "\x00\x80"
-     "\xAB\x2D\x06\x00\xAA\x2D\x06\x00\x71\xBD\x00\x80\x79\xBD\x00\x80\xAF\x19"
-     "\x06\x00"
-     "\xAE\x2D\x06\x00\xAD\x2D\x06\x00\xAC\x2D\x06\x00\x80\x51\x00\x00\x81\x59"
-     "\x00\x00"
-     "\x82\x61\x00\x00\x25\xBC\x00\x80\x75\xBD\x00\x80\x7D\xBD\x00\x80\x87\x78"
-     "\x01\x00"
-     "\x86\x00\x0C\x00\x81\xBD\x00\x80\x85\xBD\x00\x80\x89\xBD\x00\x80\x8D\xBD"
-     "\x00\x80"
-     "\x91\xBD\x00\x80\x95\xBD\x00\x80\x99\xBD\x00\x80\x9D\xBD\x00\x80\xA1\xBD"
-     "\x00\x80"
-     "\xA5\xBD\x00\x80\xB3\x99\x02\x00\x84\x94\x03\x00\xB5\xB9\x02\x00\xA9\xBD"
-     "\x00\x80"
-     "\xAD\xBD\x00\x80\xB6\xB1\x02\x00\xB1\xBD\x00\x80\xB5\xBD\x00\x80\xBB\xFD"
-     "\x02\x00"
-     "\xBA\xE1\x02\x00\xBD\xD9\x02\x00\xBC\xED\x02\x00\xBF\x19\x03\x00\xBE\xD1"
-     "\x02\x00"
-     "\xE3\x18\x06\x00\xB9\xBD\x00\x80\xE1\x44\x06\x00\xBD\xBD\x00\x80\xC1\xBD"
-     "\x00\x80"
-     "\xE1\x90\x01\x00\xC5\xBD\x00\x80\xE3\x40\x02\x00\xC9\xBD\x00\x80\xCD\xBD"
-     "\x00\x80"
-     "\xD1\xBD\x00\x80\xEF\x24\x02\x00\xEF\xC4\x06\x00\xD5\xBD\x00\x80\xD9\xBD"
-     "\x00\x80"
-     "\xDD\xBD\x00\x80\xA9\x05\x02\x00\xA8\xB9\x02\x00\xAB\x15\x02\x00\xAA\x1D"
-     "\x02\x00"
-     "\xAD\x3D\x02\x00\xAC\x3D\x02\x00\xAF\x51\x02\x00\xAE\x5D\x02\x00\x80\xD9"
-     "\x00\x00"
-     "\x81\xD9\x00\x00\x82\x05\x00\x00\xE1\xBD\x00\x80\xBE\xFC\x03\x00\xE9\xBD"
-     "\x00\x80"
-     "\x87\xA8\x0C\x00\x86\x9C\x0D\x00\xB9\xF1\x03\x00\xB8\xE9\x03\x00\xBB\xF1"
-     "\x03\x00"
-     "\xBA\xF9\x03\x00\xBD\x55\x03\x00\xBC\xE9\x03\x00\xBF\x55\x03\x00\xBE\x55"
-     "\x03\x00"
-     "\xB1\x35\x02\x00\xB0\x29\x02\x00\xB3\x01\x02\x00\xB2\x35\x02\x00\xB5\xD1"
-     "\x03\x00"
-     "\xB4\x19\x02\x00\xB7\xD1\x03\x00\xB6\xD9\x03\x00\x85\x44\x0D\x00\xED\xBD"
-     "\x00\x80"
-     "\xF1\xBD\x00\x80\xA3\x9D\x03\x00\xF5\xBD\x00\x80\xA5\xBD\x03\x00\xA6\xB5"
-     "\x03\x00"
-     "\xF9\xBD\x00\x80\xFD\xBD\x00\x80\x01\xBE\x00\x80\xAA\xE5\x03\x00\xAB\xF9"
-     "\x03\x00"
-     "\xAC\xE9\x03\x00\xAD\xDD\x03\x00\xAE\xD5\x03\x00\xAF\x1D\x02\x00\xB6\xAD"
-     "\x02\x00"
-     "\x05\xBE\x00\x80\x09\xBE\x00\x80\xB5\xA5\x02\x00\x0D\xBE\x00\x80\xB3\xB5"
-     "\x02\x00"
-     "\x11\xBE\x00\x80\x15\xBE\x00\x80\xBE\x75\x01\x00\xBF\x7D\x01\x00\xBC\x69"
-     "\x01\x00"
-     "\xBD\x69\x01\x00\xBA\x89\x02\x00\xBB\x89\x02\x00\x19\xBE\x00\x80\x1D\xBE"
-     "\x00\x80"
-     "\xE1\x94\x02\x00\xEF\xDC\x06\x00\xE3\x3C\x12\x00\x21\xBE\x00\x80\x25\xBE"
-     "\x00\x80"
-     "\x30\xBE\x00\x80\x34\xBE\x00\x80\xEF\xE8\x07\x00\x84\x6C\x0F\x00\xE1\xDC"
-     "\x06\x00"
-     "\x38\xBE\x00\x80\xE3\xA8\x06\x00\x3C\xBE\x00\x80\xE3\xF8\x06\x00\xEF\x7C"
-     "\x3E\x00"
-     "\xE1\x2C\x01\x00\xA3\x21\x02\x00\xF1\x00\x00\x00\x40\xBE\x00\x80\x44\xBE"
-     "\x00\x80"
-     "\x48\xBE\x00\x80\xA6\x39\x02\x00\xA5\x31\x02\x00\x4C\xBE\x00\x80\xAB\x1D"
-     "\x02\x00"
-     "\xAA\x1D\x02\x00\x50\xBE\x00\x80\x54\xBE\x00\x80\xAF\xE9\x01\x00\xAE\xE1"
-     "\x01\x00"
-     "\xAD\xFD\x01\x00\xAC\xFD\x01\x00\x8F\x21\x24\x00\xE5\xBD\x00\x80\x58\xBE"
-     "\x00\x80"
-     "\xBE\xC4\x00\x00\x80\xF9\x0F\x00\x81\xF9\x0F\x00\x82\x11\x00\x00\x5C\xBE"
-     "\x00\x80"
-     "\x86\x00\x00\x00\x87\x84\x03\x00\x60\xBE\x00\x80\x64\xBE\x00\x80\x68\xBE"
-     "\x00\x80"
-     "\x6C\xBE\x00\x80\x70\xBE\x00\x80\x74\xBE\x00\x80\x9F\xDD\x08\x00\x9E\x65"
-     "\x08\x00"
-     "\x9D\xF1\x09\x00\x9C\x79\x35\x00\x9B\x71\x34\x00\x9A\xE1\x35\x00\x99\x89"
-     "\x30\x00"
-     "\x98\xD1\x30\x00\x97\xE1\x30\x00\x96\x75\x3D\x00\x95\xED\x3C\x00\x94\xED"
-     "\x3C\x00"
-     "\x93\x19\x39\x00\x92\xF1\x38\x00\x91\xF9\x38\x00\x90\x99\x26\x00\xB1\xA5"
-     "\x14\x00"
-     "\xB0\xA5\x14\x00\xB3\x09\x68\x00\xB2\xE9\x15\x00\xB5\x05\x68\x00\xB4\x19"
-     "\x68\x00"
-     "\x78\xBE\x00\x80\x7C\xBE\x00\x80\x80\xBE\x00\x80\x84\xBE\x00\x80\x88\xBE"
-     "\x00\x80"
-     "\x8C\xBE\x00\x80\xA3\x61\x0D\x00\xA2\x8D\x0C\x00\xA1\x99\x0C\x00\xA0\x71"
-     "\x0D\x00"
-     "\xA1\xC5\x00\x00\x90\xBE\x00\x80\xA3\xC9\x04\x00\xA2\xD1\x00\x00\xA5\x39"
-     "\x04\x00"
-     "\xA4\xD1\x04\x00\xA7\xF9\x18\x00\xA6\x01\x18\x00\xA9\x01\x1C\x00\xA8\x79"
-     "\x18\x00"
-     "\xAB\x6D\x1C\x00\xAA\xCD\x1C\x00\xAD\xD1\x10\x00\xAC\x01\x10\x00\xAF\x01"
-     "\x14\x00"
-     "\xAE\xD9\x10\x00\x82\x21\x25\x00\x83\xE1\x27\x00\x94\xBE\x00\x80\x98\xBE"
-     "\x00\x80"
-     "\x86\x79\x38\x00\x87\x95\x3D\x00\x84\x99\x39\x00\x85\x99\x39\x00\x8A\xA5"
-     "\x3D\x00"
-     "\x8B\x35\x3C\x00\x9C\xBE\x00\x80\xA0\xBE\x00\x80\x8E\xE1\x31\x00\x8F\x81"
-     "\x34\x00"
-     "\x8C\xB9\x31\x00\x8D\xB1\x31\x00\x92\x61\x35\x00\x93\x71\x34\x00\xA4\xBE"
-     "\x00\x80"
-     "\xA8\xBE\x00\x80\x96\xCD\x09\x00\x97\x99\x0C\x00\x94\xE1\x08\x00\x95\x7D"
-     "\x09\x00"
-     "\x9A\x41\x0D\x00\x9B\x5D\x0D\x00\xAC\xBE\x00\x80\xB0\xBE\x00\x80\xB4\xBE"
-     "\x00\x80"
-     "\xB8\xBE\x00\x80\x9C\xA9\x0C\x00\xBC\xBE\x00\x80\xA8\x35\x0F\x00\xA9\x81"
-     "\x0E\x00"
-     "\xAA\x81\x0E\x00\xAB\x9D\x0E\x00\xAC\x8D\x0E\x00\xAD\xB5\x0E\x00\xAE\xBD"
-     "\x0E\x00"
-     "\xAF\xD9\x0E\x00\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\xC0\xBE"
-     "\x00\x80"
-     "\xC4\xBE\x00\x80\xC8\xBE\x00\x80\x84\xBC\x00\x00\xBE\x00\x04\x00\xB8\x55"
-     "\x0E\x00"
-     "\xB9\x5D\x0E\x00\xBA\x61\x0E\x00\xBB\x7D\x0E\x00\xBC\x19\x0E\x00\xBD\x19"
-     "\x0E\x00"
-     "\xBE\x09\x0E\x00\xBF\x01\x0E\x00\xB0\xA9\x0E\x00\xB1\xB5\x0E\x00\xB2\xB1"
-     "\x0E\x00"
-     "\xB3\xB1\x0E\x00\xB4\x89\x0E\x00\xB5\x89\x0E\x00\xB6\x75\x0E\x00\xB7\x6D"
-     "\x0E\x00"
-     "\xB3\x09\x0D\x00\xD0\xBE\x00\x80\x86\x88\x04\x00\x87\xC8\x03\x00\x84\x5C"
-     "\x04\x00"
-     "\xB6\x35\x0D\x00\xB5\x05\x0D\x00\xD4\xBE\x00\x80\xBB\x71\x0D\x00\xBA\x6D"
-     "\x0D\x00"
-     "\xD8\xBE\x00\x80\xDC\xBE\x00\x80\xBF\x4D\x0D\x00\xBE\x4D\x0D\x00\xBD\x65"
-     "\x0D\x00"
-     "\xBC\x65\x0D\x00\xE0\xBE\x00\x80\xA3\x4D\x0D\x00\xE4\xBE\x00\x80\xE8\xBE"
-     "\x00\x80"
-     "\xA6\x71\x0D\x00\xEC\xBE\x00\x80\xF0\xBE\x00\x80\xA5\x41\x0D\x00\xAA\x29"
-     "\x0D\x00"
-     "\xAB\x35\x0D\x00\xF4\xBE\x00\x80\xF8\xBE\x00\x80\xAE\x09\x0D\x00\xAF\x09"
-     "\x0D\x00"
-     "\xAC\x21\x0D\x00\xAD\x21\x0D\x00\xA8\xA9\x02\x00\xA9\xA9\x02\x00\xAA\xB9"
-     "\x02\x00"
-     "\xAB\xB9\x02\x00\xAC\xA9\x02\x00\xAD\xA9\x02\x00\xAE\x01\x02\x00\xAF\x01"
-     "\x02\x00"
-     "\xFC\xBE\x00\x80\x00\xBF\x00\x80\x04\xBF\x00\x80\x08\xBF\x00\x80\x0C\xBF"
-     "\x00\x80"
-     "\x10\xBF\x00\x80\x14\xBF\x00\x80\x18\xBF\x00\x80\xB8\xED\x03\x00\xB9\x8D"
-     "\x03\x00"
-     "\xBA\x85\x03\x00\xBB\xA5\x03\x00\xBC\xBD\x03\x00\xBD\xA5\x03\x00\xBE\xAD"
-     "\x03\x00"
-     "\xBF\xA1\x03\x00\xB0\x41\x02\x00\xB1\x41\x02\x00\xB2\x41\x02\x00\xB3\x41"
-     "\x02\x00"
-     "\xB4\x35\x02\x00\xB5\xFD\x03\x00\xB6\xF5\x03\x00\xB7\xD5\x03\x00\xB7\x25"
-     "\x0D\x00"
-     "\xB6\x25\x0D\x00\xB5\x11\x0D\x00\xB4\x19\x0D\x00\xB3\x15\x0D\x00\xB2\x15"
-     "\x0D\x00"
-     "\xB1\xE5\x0E\x00\xB0\xE5\x0E\x00\xBF\x55\x0D\x00\xBE\x4D\x0D\x00\xBD\x4D"
-     "\x0D\x00"
-     "\xBC\x4D\x0D\x00\xBB\x69\x0D\x00\xBA\x65\x0D\x00\xB9\x75\x0D\x00\xB8\x79"
-     "\x0D\x00"
-     "\x1C\xBF\x00\x80\xBE\x0C\x05\x00\x20\xBF\x00\x80\x24\xBF\x00\x80\x28\xBF"
-     "\x00\x80"
-     "\x2C\xBF\x00\x80\x30\xBF\x00\x80\x34\xBF\x00\x80\xAF\x95\x0E\x00\xAE\x95"
-     "\x0E\x00"
-     "\xAD\x85\x0E\x00\xAC\x85\x0E\x00\xAB\x95\x0E\x00\xAA\x95\x0E\x00\xA9\xAD"
-     "\x0E\x00"
-     "\xA8\xD1\x0E\x00\xE1\x24\x0F\x00\x38\xBF\x00\x80\xE3\x38\x0E\x00\x40\xBF"
-     "\x00\x80"
-     "\xE3\x38\x00\x00\x44\xBF\x00\x80\xE1\xA0\x01\x00\x48\xBF\x00\x80\xEF\xE4"
-     "\x00\x00"
-     "\x4C\xBF\x00\x80\x84\xE4\x04\x00\x50\xBF\x00\x80\x86\x10\x04\x00\x87\x94"
-     "\x04\x00"
-     "\xEF\x4C\x01\x00\x54\xBF\x00\x80\x81\x9D\x00\x00\x80\x9D\x00\x00\xB3\xF9"
-     "\x03\x00"
-     "\x82\x49\x00\x00\xB5\xE9\x03\x00\x58\xBF\x00\x80\x5C\xBF\x00\x80\xB6\xE1"
-     "\x03\x00"
-     "\x60\xBF\x00\x80\x64\xBF\x00\x80\xBB\x4D\x03\x00\xBA\x4D\x03\x00\xBD\x5D"
-     "\x03\x00"
-     "\xBC\x5D\x03\x00\xBF\xF1\x02\x00\xBE\x39\x03\x00\xCC\xBE\x00\x80\x3C\xBF"
-     "\x00\x80"
-     "\x68\xBF\x00\x80\x6C\xBF\x00\x80\x70\xBF\x00\x80\x74\xBF\x00\x80\x78\xBF"
-     "\x00\x80"
-     "\x7C\xBF\x00\x80\xA8\x25\x06\x00\xA9\x39\x06\x00\xAA\x01\x06\x00\xAB\x01"
-     "\x06\x00"
-     "\xAC\x71\x06\x00\xAD\x7D\x06\x00\xAE\x75\x06\x00\xAF\x69\x06\x00\xB0\x19"
-     "\x06\x00"
-     "\xB1\x2D\x06\x00\xB2\x25\x06\x00\xB3\xED\x06\x00\xB4\xF5\x06\x00\xB5\xFD"
-     "\x06\x00"
-     "\xB6\xF5\x06\x00\xB7\xED\x06\x00\xB8\xD5\x06\x00\xB9\xDD\x06\x00\xBA\xD5"
-     "\x06\x00"
-     "\xBB\x2D\x07\x00\xBC\x39\x07\x00\xBD\x39\x07\x00\xBE\x29\x07\x00\xBF\x29"
-     "\x07\x00"
-     "\xB3\xA9\x06\x00\x80\xBF\x00\x80\x84\xBF\x00\x80\x88\xBF\x00\x80\x8C\xBF"
-     "\x00\x80"
-     "\xB6\xDD\x06\x00\xB5\xD9\x06\x00\x90\xBF\x00\x80\xBB\xC9\x06\x00\xBA\xFD"
-     "\x06\x00"
-     "\x94\xBF\x00\x80\x98\xBF\x00\x80\xBF\x19\x06\x00\xBE\x11\x06\x00\xBD\x19"
-     "\x06\x00"
-     "\xBC\xD1\x06\x00\x9C\xBF\x00\x80\xA3\xED\x06\x00\xA0\xBF\x00\x80\xA4\xBF"
-     "\x00\x80"
-     "\xA6\x99\x06\x00\xA8\xBF\x00\x80\xBE\x90\x03\x00\xA5\x9D\x06\x00\xAA\xB9"
-     "\x06\x00"
-     "\xAB\x8D\x06\x00\xAC\xBF\x00\x80\xB0\xBF\x00\x80\xAE\x55\x06\x00\xAF\x5D"
-     "\x06\x00"
-     "\xAC\x95\x06\x00\xAD\x5D\x06\x00\xA8\x8D\x06\x00\xA9\x91\x06\x00\xAA\x95"
-     "\x06\x00"
-     "\xAB\xA9\x06\x00\xAC\xBD\x06\x00\xAD\xA9\x06\x00\xAE\xCD\x06\x00\xAF\xC5"
-     "\x06\x00"
-     "\x80\xE9\x01\x00\x81\xE9\x01\x00\x82\xFD\x01\x00\xB4\xBF\x00\x80\x86\x90"
-     "\x01\x00"
-     "\x87\xB4\x01\x00\xB8\xBF\x00\x80\xBC\xBF\x00\x80\xB8\xC5\x01\x00\xB9\xCD"
-     "\x01\x00"
-     "\xBA\xC5\x01\x00\xBB\xDD\x01\x00\xBC\xCD\x01\x00\xBD\xFD\x01\x00\xBE\xF5"
-     "\x01\x00"
-     "\xBF\x9D\x01\x00\xB0\x81\x06\x00\xB1\x81\x06\x00\xB2\x61\x01\x00\xB3\x61"
-     "\x01\x00"
-     "\xB4\x61\x01\x00\xB5\x61\x01\x00\xB6\x61\x01\x00\xB7\x61\x01\x00\xB3\x1D"
-     "\x06\x00"
-     "\xC0\xBF\x00\x80\xC4\xBF\x00\x80\xC8\xBF\x00\x80\xCC\xBF\x00\x80\xB6\x29"
-     "\x06\x00"
-     "\xB5\x35\x06\x00\xD0\xBF\x00\x80\xBB\x45\x06\x00\xBA\x45\x06\x00\xD4\xBF"
-     "\x00\x80"
-     "\xD8\xBF\x00\x80\xBF\xA5\x01\x00\xBE\xA5\x01\x00\xBD\xB5\x01\x00\xBC\xB5"
-     "\x01\x00"
-     "\xDC\xBF\x00\x80\xA3\x59\x06\x00\xE0\xBF\x00\x80\xE4\xBF\x00\x80\xA6\x6D"
-     "\x06\x00"
-     "\xE8\xBF\x00\x80\xEC\xBF\x00\x80\xA5\x71\x06\x00\xAA\x01\x06\x00\xAB\x01"
-     "\x06\x00"
-     "\xF0\xBF\x00\x80\xF4\xBF\x00\x80\xAE\xE1\x01\x00\xAF\xE1\x01\x00\xAC\xF1"
-     "\x01\x00"
-     "\xAD\xF1\x01\x00\xA8\xED\x02\x00\xA9\x29\x03\x00\xAA\x31\x03\x00\xAB\x31"
-     "\x03\x00"
-     "\xAC\x95\x03\x00\xAD\x9D\x03\x00\xAE\x95\x03\x00\xAF\x8D\x03\x00\x84\x0C"
-     "\x02\x00"
-     "\xF8\xBF\x00\x80\xFC\xBF\x00\x80\x00\xC0\x00\x80\x04\xC0\x00\x80\x08\xC0"
-     "\x00\x80"
-     "\x0C\xC0\x00\x80\x10\xC0\x00\x80\xB8\xA9\x03\x00\xB9\xA9\x03\x00\xBA\x6D"
-     "\x00\x00"
-     "\xBB\x65\x00\x00\xBC\x7D\x00\x00\xBD\x65\x00\x00\xBE\x6D\x00\x00\xBF\x65"
-     "\x00\x00"
-     "\xB0\xFD\x03\x00\xB1\xCD\x03\x00\xB2\xC5\x03\x00\xB3\xAD\x03\x00\xB4\xB9"
-     "\x03\x00"
-     "\xB5\xB9\x03\x00\xB6\xAD\x03\x00\xB7\xA5\x03\x00\x80\xBD\x00\x00\x81\x09"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x14\xC0\x00\x80\x18\xC0\x00\x80\xBE\xD8\x03\x00\x20\xC0"
-     "\x00\x80"
-     "\x24\xC0\x00\x80\x86\x00\x1C\x00\x87\xD4\x03\x00\x28\xC0\x00\x80\x2C\xC0"
-     "\x00\x80"
-     "\x30\xC0\x00\x80\x34\xC0\x00\x80\x38\xC0\x00\x80\x3C\xC0\x00\x80\xBB\x69"
-     "\x03\x00"
-     "\xBA\x59\x03\x00\x40\xC0\x00\x80\x44\xC0\x00\x80\xBF\x1D\x03\x00\xBE\x1D"
-     "\x03\x00"
-     "\xBD\x19\x03\x00\xBC\x71\x03\x00\xB3\xD9\x03\x00\x48\xC0\x00\x80\x4C\xC0"
-     "\x00\x80"
-     "\x50\xC0\x00\x80\x54\xC0\x00\x80\xB6\x7D\x03\x00\xB5\x7D\x03\x00\x58\xC0"
-     "\x00\x80"
-     "\xAF\x09\x06\x00\xAE\xC1\x07\x00\xAD\xA5\x07\x00\xAC\xA5\x07\x00\xAB\xB5"
-     "\x07\x00"
-     "\xAA\xB5\x07\x00\x5C\xC0\x00\x80\x60\xC0\x00\x80\x84\x6C\x1D\x00\xA6\x19"
-     "\x07\x00"
-     "\xA5\x11\x07\x00\x64\xC0\x00\x80\xA3\x01\x07\x00\x68\xC0\x00\x80\x6C\xC0"
-     "\x00\x80"
-     "\x70\xC0\x00\x80\x74\xC0\x00\x80\x78\xC0\x00\x80\xBE\x08\x1D\x00\x84\x00"
-     "\x1C\x00"
-     "\x7C\xC0\x00\x80\xEF\x38\x03\x00\x80\xC0\x00\x80\x84\xC0\x00\x80\x88\xC0"
-     "\x00\x80"
-     "\xE3\x54\x03\x00\x90\xC0\x00\x80\xE1\x44\x00\x00\x80\x29\x00\x00\x81\x29"
-     "\x00\x00"
-     "\x82\x19\x00\x00\x94\xC0\x00\x80\xE1\x28\x06\x00\xE1\x94\x07\x00\xE3\x30"
-     "\x06\x00"
-     "\xE3\x8C\x07\x00\x98\xC0\x00\x80\x9C\xC0\x00\x80\x86\x38\x1C\x00\x87\x5C"
-     "\x1D\x00"
-     "\xBE\xE8\x1C\x00\xA0\xC0\x00\x80\xA4\xC0\x00\x80\xA8\xC0\x00\x80\xAC\xC0"
-     "\x00\x80"
-     "\xB0\xC0\x00\x80\xEF\x98\x01\x00\xEF\x00\x06\x00\xA3\xF9\x02\x00\xB4\xC0"
-     "\x00\x80"
-     "\xB8\xC0\x00\x80\xBC\xC0\x00\x80\xC0\xC0\x00\x80\xA6\x5D\x02\x00\xA5\x5D"
-     "\x02\x00"
-     "\xC4\xC0\x00\x80\xAB\x49\x02\x00\xAA\x79\x02\x00\xC8\xC0\x00\x80\xCC\xC0"
-     "\x00\x80"
-     "\xAF\x3D\x02\x00\xAE\x3D\x02\x00\xAD\x39\x02\x00\xAC\x51\x02\x00\xA8\x99"
-     "\x1E\x00"
-     "\xA9\x99\x1E\x00\xAA\xA1\x1E\x00\xAB\xA1\x1E\x00\xAC\xA1\x1E\x00\xAD\xA1"
-     "\x1E\x00"
-     "\xAE\xA1\x1E\x00\xAF\xA1\x1E\x00\x8C\xC0\x00\x80\xD0\xC0\x00\x80\xD4\xC0"
-     "\x00\x80"
-     "\xD8\xC0\x00\x80\xDC\xC0\x00\x80\xE0\xC0\x00\x80\xE4\xC0\x00\x80\xE8\xC0"
-     "\x00\x80"
-     "\xB8\x5D\x1F\x00\xB9\x69\x1F\x00\xBA\x7D\x1F\x00\xBB\x75\x1F\x00\xBC\x6D"
-     "\x1F\x00"
-     "\xBD\x15\x1F\x00\xBE\x1D\x1F\x00\xBF\x0D\x1F\x00\xB0\xA9\x1E\x00\xB1\xA9"
-     "\x1E\x00"
-     "\xB2\x8D\x1E\x00\xB3\x85\x1E\x00\xB4\x9D\x1E\x00\xB5\x85\x1E\x00\xB6\x85"
-     "\x1E\x00"
-     "\xB7\x65\x1F\x00\xB3\x29\x1E\x00\xEC\xC0\x00\x80\xF0\xC0\x00\x80\xF4\xC0"
-     "\x00\x80"
-     "\xF8\xC0\x00\x80\xB6\x21\x1E\x00\xB5\x29\x1E\x00\xFC\xC0\x00\x80\xBB\x6D"
-     "\x1E\x00"
-     "\xBA\x6D\x1E\x00\x86\x08\x00\x00\x87\x2C\x00\x00\xBF\x41\x1E\x00\xBE\x59"
-     "\x1E\x00"
-     "\xBD\x51\x1E\x00\xBC\x59\x1E\x00\x80\x49\x00\x00\x81\x49\x00\x00\x82\x59"
-     "\x00\x00"
-     "\xA3\x65\x1E\x00\x00\xC1\x00\x80\xA5\x65\x1E\x00\xA6\x6D\x1E\x00\x1C\xC0"
-     "\x00\x80"
-     "\x04\xC1\x00\x80\x08\xC1\x00\x80\xAA\x21\x1E\x00\xAB\x21\x1E\x00\xAC\x15"
-     "\x1E\x00"
-     "\xAD\x1D\x1E\x00\xAE\x15\x1E\x00\xAF\x0D\x1E\x00\xA8\x69\x1E\x00\xA9\x79"
-     "\x1E\x00"
-     "\xAA\x99\x1E\x00\xAB\x99\x1E\x00\xAC\xB5\x1E\x00\xAD\xBD\x1E\x00\xAE\xB5"
-     "\x1E\x00"
-     "\xAF\xAD\x1E\x00\x0C\xC1\x00\x80\x10\xC1\x00\x80\x14\xC1\x00\x80\x18\xC1"
-     "\x00\x80"
-     "\x1C\xC1\x00\x80\x20\xC1\x00\x80\x24\xC1\x00\x80\x28\xC1\x00\x80\xB8\x65"
-     "\x01\x00"
-     "\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x7D\x01\x00\xBC\x65\x01\x00\xBD\x6D"
-     "\x01\x00"
-     "\xBE\x65\x01\x00\xBF\xD9\x01\x00\xB0\xD5\x1E\x00\xB1\xA1\x1E\x00\xB2\xA1"
-     "\x1E\x00"
-     "\xB3\xA1\x1E\x00\xB4\xA1\x1E\x00\xB5\xA9\x1E\x00\xB6\x91\x1E\x00\xB7\x91"
-     "\x1E\x00"
-     "\xB3\x11\x1E\x00\x2C\xC1\x00\x80\x30\xC1\x00\x80\x34\xC1\x00\x80\x38\xC1"
-     "\x00\x80"
-     "\xB6\x3D\x1E\x00\xB5\x3D\x1E\x00\x3C\xC1\x00\x80\xBB\x65\x1E\x00\xBA\x11"
-     "\x1E\x00"
-     "\x40\xC1\x00\x80\x44\xC1\x00\x80\xBF\x6D\x1E\x00\xBE\x65\x1E\x00\xBD\x75"
-     "\x1E\x00"
-     "\xBC\x75\x1E\x00\x48\xC1\x00\x80\xA3\x55\x1E\x00\x4C\xC1\x00\x80\x50\xC1"
-     "\x00\x80"
-     "\xA6\x79\x1E\x00\x54\xC1\x00\x80\x58\xC1\x00\x80\xA5\x79\x1E\x00\xAA\x55"
-     "\x1E\x00"
-     "\xAB\x21\x1E\x00\x5C\xC1\x00\x80\x60\xC1\x00\x80\xAE\x21\x1E\x00\xAF\x29"
-     "\x1E\x00"
-     "\xAC\x31\x1E\x00\xAD\x31\x1E\x00\xA8\xD1\x02\x00\xA9\xE5\x02\x00\xAA\xED"
-     "\x02\x00"
-     "\xAB\x2D\x03\x00\xAC\x35\x03\x00\xAD\x3D\x03\x00\xAE\x35\x03\x00\xAF\x2D"
-     "\x03\x00"
-     "\x64\xC1\x00\x80\x68\xC1\x00\x80\x70\xC1\x00\x80\xBE\x60\x02\x00\x74\xC1"
-     "\x00\x80"
-     "\x78\xC1\x00\x80\x7C\xC1\x00\x80\x80\xC1\x00\x80\xB8\xE9\x03\x00\xB9\xE9"
-     "\x03\x00"
-     "\xBA\x89\x03\x00\xBB\x85\x03\x00\xBC\x9D\x03\x00\xBD\x81\x03\x00\xBE\x81"
-     "\x03\x00"
-     "\xBF\xB5\x03\x00\xB0\x55\x03\x00\xB1\x5D\x03\x00\xB2\x55\x03\x00\xB3\xE9"
-     "\x03\x00"
-     "\xB4\xF9\x03\x00\xB5\xF9\x03\x00\xB6\xE9\x03\x00\xB7\xE1\x03\x00\x80\x6D"
-     "\x03\x00"
-     "\x81\xA5\x00\x00\x82\xAD\x00\x00\xB3\x69\x02\x00\x84\xC1\x00\x80\xB5\xB1"
-     "\x03\x00"
-     "\xB6\xB1\x03\x00\x88\xC1\x00\x80\x84\xE0\x02\x00\x8C\xC1\x00\x80\xBA\x9D"
-     "\x03\x00"
-     "\xBB\x95\x03\x00\xBC\x8D\x03\x00\xBD\x31\x03\x00\xBE\x31\x03\x00\xBF\x31"
-     "\x03\x00"
-     "\xA3\x25\x02\x00\x90\xC1\x00\x80\x94\xC1\x00\x80\x98\xC1\x00\x80\x9C\xC1"
-     "\x00\x80"
-     "\xA6\xFD\x03\x00\xA5\xFD\x03\x00\xA0\xC1\x00\x80\xAB\xD9\x03\x00\xAA\xD1"
-     "\x03\x00"
-     "\x86\x48\x04\x00\x87\xAC\x03\x00\xAF\x7D\x03\x00\xAE\x7D\x03\x00\xAD\x7D"
-     "\x03\x00"
-     "\xAC\xC1\x03\x00\xA4\xC1\x00\x80\xA8\xC1\x00\x80\xAC\xC1\x00\x80\xB0\xC1"
-     "\x00\x80"
-     "\xE3\xA0\x03\x00\xB4\xC1\x00\x80\xE1\xA8\x01\x00\xB8\xC1\x00\x80\xEF\xF4"
-     "\x03\x00"
-     "\xBC\xC1\x00\x80\xC0\xC1\x00\x80\xC4\xC1\x00\x80\xC8\xC1\x00\x80\xCC\xC1"
-     "\x00\x80"
-     "\xD0\xC1\x00\x80\xD4\xC1\x00\x80\x80\x79\x00\x00\x81\x79\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\x84\x40\x02\x00\xE3\x60\x1E\x00\xD8\xC1\x00\x80\xE1\xE0\x1F\x00\xBE\x50"
-     "\x05\x00"
-     "\xEF\xD0\x1E\x00\xE0\xC1\x00\x80\x86\xC8\x04\x00\x87\x40\x05\x00\xE4\xC1"
-     "\x00\x80"
-     "\xE8\xC1\x00\x80\xEC\xC1\x00\x80\xF0\xC1\x00\x80\xF4\xC1\x00\x80\xF8\xC1"
-     "\x00\x80"
-     "\xFC\xC1\x00\x80\x00\xC2\x00\x80\x04\xC2\x00\x80\x08\xC2\x00\x80\x0C\xC2"
-     "\x00\x80"
-     "\xEF\xC4\x1E\x00\x84\xA4\x05\x00\xE1\x24\x1E\x00\x10\xC2\x00\x80\xE3\xF0"
-     "\x1E\x00"
-     "\x14\xC2\x00\x80\x18\xC2\x00\x80\x1C\xC2\x00\x80\x20\xC2\x00\x80\xB3\xF5"
-     "\x01\x00"
-     "\x24\xC2\x00\x80\x28\xC2\x00\x80\x2C\xC2\x00\x80\x30\xC2\x00\x80\xB6\x45"
-     "\x01\x00"
-     "\xB5\xE5\x01\x00\x34\xC2\x00\x80\xBB\x29\x01\x00\xBA\x21\x01\x00\x38\xC2"
-     "\x00\x80"
-     "\x3C\xC2\x00\x80\xBF\xF1\x00\x00\xBE\x29\x01\x00\xBD\x29\x01\x00\xBC\x31"
-     "\x01\x00"
-     "\xA8\x29\x06\x00\xA9\x35\x06\x00\xAA\x3D\x06\x00\xAB\x35\x06\x00\xAC\x2D"
-     "\x06\x00"
-     "\xAD\x85\x06\x00\xAE\x85\x06\x00\xAF\xB9\x06\x00\xDC\xC1\x00\x80\x82\x1D"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\x80\x1D\x00\x00\x40\xC2\x00\x80\x44\xC2\x00\x80\x48\xC2"
-     "\x00\x80"
-     "\xBE\x30\x03\x00\xB8\x4D\x07\x00\xB9\x51\x07\x00\xBA\x55\x07\x00\xBB\x69"
-     "\x07\x00"
-     "\xBC\x79\x07\x00\xBD\x79\x07\x00\xBE\x69\x07\x00\xBF\x69\x07\x00\xB0\xA5"
-     "\x06\x00"
-     "\xB1\xAD\x06\x00\xB2\xA5\x06\x00\xB3\xBD\x06\x00\xB4\xAD\x06\x00\xB5\x75"
-     "\x07\x00"
-     "\xB6\x7D\x07\x00\xB7\x75\x07\x00\xA3\xB5\x06\x00\x4C\xC2\x00\x80\x86\x28"
-     "\x00\x00"
-     "\x87\x4C\x03\x00\x50\xC2\x00\x80\xA6\x05\x06\x00\xA5\xA5\x06\x00\x54\xC2"
-     "\x00\x80"
-     "\xAB\x69\x06\x00\xAA\x61\x06\x00\x58\xC2\x00\x80\x5C\xC2\x00\x80\xAF\xB1"
-     "\x07\x00"
-     "\xAE\x69\x06\x00\xAD\x69\x06\x00\xAC\x71\x06\x00\x60\xC2\x00\x80\x64\xC2"
-     "\x00\x80"
-     "\xB3\xBD\x07\x00\x68\xC2\x00\x80\xB5\xAD\x07\x00\x6C\xC2\x00\x80\x70\xC2"
-     "\x00\x80"
-     "\xB6\x5D\x07\x00\x6C\xC1\x00\x80\x74\xC2\x00\x80\xBB\x79\x07\x00\xBA\x61"
-     "\x07\x00"
-     "\xBD\x69\x07\x00\xBC\x61\x07\x00\xBF\x49\x07\x00\xBE\x55\x07\x00\xA8\x65"
-     "\x06\x00"
-     "\xA9\x75\x06\x00\xAA\x7D\x06\x00\xAB\x75\x06\x00\xAC\x6D\x06\x00\xAD\x95"
-     "\x06\x00"
-     "\xAE\x9D\x06\x00\xAF\x95\x06\x00\x78\xC2\x00\x80\x7C\xC2\x00\x80\x80\xC2"
-     "\x00\x80"
-     "\x84\xC2\x00\x80\x88\xC2\x00\x80\x8C\xC2\x00\x80\x90\xC2\x00\x80\x94\xC2"
-     "\x00\x80"
-     "\xB8\xFD\x06\x00\xB9\x81\x06\x00\xBA\x81\x06\x00\xBB\x81\x06\x00\xBC\x81"
-     "\x06\x00"
-     "\xBD\x89\x06\x00\xBE\xB1\x06\x00\xBF\xB1\x06\x00\xB0\xED\x06\x00\xB1\xF5"
-     "\x06\x00"
-     "\xB2\xFD\x06\x00\xB3\xF5\x06\x00\xB4\xD9\x06\x00\xB5\xC5\x06\x00\xB6\xCD"
-     "\x06\x00"
-     "\xB7\xC5\x06\x00\xA3\xF5\x06\x00\x98\xC2\x00\x80\x9C\xC2\x00\x80\xA0\xC2"
-     "\x00\x80"
-     "\xA4\xC2\x00\x80\xA6\x15\x06\x00\xA5\xE5\x06\x00\xA8\xC2\x00\x80\xAB\x31"
-     "\x06\x00"
-     "\xAA\x29\x06\x00\xAC\xC2\x00\x80\xB0\xC2\x00\x80\xAF\x01\x06\x00\xAE\x1D"
-     "\x06\x00"
-     "\xAD\x21\x06\x00\xAC\x29\x06\x00\x80\x6D\x00\x00\x81\x09\x00\x00\x82\x19"
-     "\x00\x00"
-     "\xB4\xC2\x00\x80\xB8\xC2\x00\x80\x84\x98\x01\x00\xBE\x90\x01\x00\xC0\xC2"
-     "\x00\x80"
-     "\x86\x40\x0F\x00\x87\xF4\x00\x00\xC4\xC2\x00\x80\xC8\xC2\x00\x80\xCC\xC2"
-     "\x00\x80"
-     "\xD0\xC2\x00\x80\xD4\xC2\x00\x80\xD8\xC2\x00\x80\xA8\x91\x02\x00\xA9\x99"
-     "\x02\x00"
-     "\xAA\xA1\x02\x00\xAB\xA1\x02\x00\xAC\xD1\x02\x00\xAD\xD1\x02\x00\xAE\xD5"
-     "\x02\x00"
-     "\xAF\xC9\x02\x00\xDC\xC2\x00\x80\xE0\xC2\x00\x80\xE4\xC2\x00\x80\x84\x20"
-     "\x03\x00"
-     "\xE8\xC2\x00\x80\xEC\xC2\x00\x80\xF0\xC2\x00\x80\xF4\xC2\x00\x80\xB8\x79"
-     "\x03\x00"
-     "\xB9\x79\x03\x00\xBA\xCD\x03\x00\xBB\xC5\x03\x00\xBC\xDD\x03\x00\xBD\xC5"
-     "\x03\x00"
-     "\xBE\xC5\x03\x00\xBF\xF5\x03\x00\xB0\xB9\x02\x00\xB1\x81\x02\x00\xB2\x49"
-     "\x03\x00"
-     "\xB3\x49\x03\x00\xB4\x59\x03\x00\xB5\x59\x03\x00\xB6\x49\x03\x00\xB7\x49"
-     "\x03\x00"
-     "\xB3\x29\x02\x00\xF8\xC2\x00\x80\xFC\xC2\x00\x80\x00\xC3\x00\x80\xBE\x9C"
-     "\x0C\x00"
-     "\xB6\x51\x02\x00\xB5\x59\x02\x00\x04\xC3\x00\x80\xBB\x45\x02\x00\xBA\x7D"
-     "\x02\x00"
-     "\x08\xC3\x00\x80\x0C\xC3\x00\x80\xBF\x9D\x03\x00\xBE\x9D\x03\x00\xBD\x9D"
-     "\x03\x00"
-     "\xBC\x9D\x03\x00\x10\xC3\x00\x80\xA3\x6D\x02\x00\x14\xC3\x00\x80\x18\xC3"
-     "\x00\x80"
-     "\xA6\x15\x02\x00\x1C\xC3\x00\x80\x20\xC3\x00\x80\xA5\x1D\x02\x00\xAA\x39"
-     "\x02\x00"
-     "\xAB\x01\x02\x00\x24\xC3\x00\x80\x28\xC3\x00\x80\xAE\xD9\x03\x00\xAF\xD9"
-     "\x03\x00"
-     "\xAC\xD9\x03\x00\xAD\xD9\x03\x00\xE1\x90\x01\x00\x30\xC3\x00\x80\xE3\xCC"
-     "\x00\x00"
-     "\x84\x18\x0D\x00\x80\xC9\x01\x00\x81\xFD\x01\x00\x82\xF5\x01\x00\x34\xC3"
-     "\x00\x80"
-     "\xBE\x2C\x0C\x00\x38\xC3\x00\x80\x3C\xC3\x00\x80\x40\xC3\x00\x80\x44\xC3"
-     "\x00\x80"
-     "\x48\xC3\x00\x80\xEF\x18\x00\x00\x4C\xC3\x00\x80\x86\xE0\x0C\x00\x87\xF4"
-     "\x02\x00"
-     "\x50\xC3\x00\x80\x54\xC3\x00\x80\x58\xC3\x00\x80\x5C\xC3\x00\x80\x60\xC3"
-     "\x00\x80"
-     "\xEF\xF8\x06\x00\x64\xC3\x00\x80\xE1\x94\x07\x00\x68\xC3\x00\x80\xE3\x30"
-     "\x06\x00"
-     "\x6C\xC3\x00\x80\x70\xC3\x00\x80\x74\xC3\x00\x80\x78\xC3\x00\x80\x7C\xC3"
-     "\x00\x80"
-     "\x80\xC3\x00\x80\x84\xC3\x00\x80\x88\xC3\x00\x80\x8C\xC3\x00\x80\x90\xC3"
-     "\x00\x80"
-     "\x94\xC3\x00\x80\xEF\x58\x01\x00\x85\x00\x70\x00\xE1\xD4\x06\x00\x98\xC3"
-     "\x00\x80"
-     "\xE3\x1C\x06\x00\x80\x25\x00\x00\x81\x2D\x00\x00\x82\x11\x00\x00\xA0\xC3"
-     "\x00\x80"
-     "\xB3\x2D\x02\x00\xF1\x00\x00\x00\x87\xE8\x0C\x00\x86\x6C\x0F\x00\xA4\xC3"
-     "\x00\x80"
-     "\xB6\x2D\x02\x00\xB5\x2D\x02\x00\xA8\xC3\x00\x80\xBB\xC9\x01\x00\xBA\xC1"
-     "\x01\x00"
-     "\xAC\xC3\x00\x80\xB0\xC3\x00\x80\xBF\xCD\x01\x00\xBE\xCD\x01\x00\xBD\xCD"
-     "\x01\x00"
-     "\xBC\xCD\x01\x00\xA8\xDD\x0D\x00\xA9\xED\x0D\x00\xAA\xE5\x0D\x00\xAB\xFD"
-     "\x0D\x00"
-     "\xAC\xE5\x0D\x00\xAD\xED\x0D\x00\xAE\xE5\x0D\x00\xAF\x3D\x0E\x00\x2C\xC3"
-     "\x00\x80"
-     "\x9C\xC3\x00\x80\xB4\xC3\x00\x80\xB8\xC3\x00\x80\xBC\xC3\x00\x80\xC0\xC3"
-     "\x00\x80"
-     "\xC4\xC3\x00\x80\xC8\xC3\x00\x80\xB8\xCD\x0E\x00\xB9\xD9\x0E\x00\xBA\xE1"
-     "\x0E\x00"
-     "\xBB\xE1\x0E\x00\xBC\x91\x0E\x00\xBD\x9D\x0E\x00\xBE\x95\x0E\x00\xBF\x89"
-     "\x0E\x00"
-     "\xB0\x45\x0E\x00\xB1\x4D\x0E\x00\xB2\x59\x0E\x00\xB3\x59\x0E\x00\xB4\x49"
-     "\x0E\x00"
-     "\xB5\x49\x0E\x00\xB6\xFD\x0E\x00\xB7\xF5\x0E\x00\xA3\x6D\x0D\x00\xCC\xC3"
-     "\x00\x80"
-     "\xD0\xC3\x00\x80\xD4\xC3\x00\x80\xD8\xC3\x00\x80\xA6\x6D\x0D\x00\xA5\x6D"
-     "\x0D\x00"
-     "\xDC\xC3\x00\x80\xAB\x89\x0E\x00\xAA\x81\x0E\x00\xE0\xC3\x00\x80\xE4\xC3"
-     "\x00\x80"
-     "\xAF\x8D\x0E\x00\xAE\x8D\x0E\x00\xAD\x8D\x0E\x00\xAC\x8D\x0E\x00\xB5\xED"
-     "\x0F\x00"
-     "\xE8\xC3\x00\x80\xEC\xC3\x00\x80\xB6\xFD\x0F\x00\xF0\xC3\x00\x80\xF4\xC3"
-     "\x00\x80"
-     "\xB3\xFD\x0F\x00\xF8\xC3\x00\x80\xBD\xB5\x0F\x00\xBC\xB5\x0F\x00\xBF\x89"
-     "\x0F\x00"
-     "\xBE\xB5\x0F\x00\xFC\xC3\x00\x80\xBC\xC2\x00\x80\xBB\xB5\x0F\x00\xBA\xB5"
-     "\x0F\x00"
-     "\xA8\x05\x0E\x00\xA9\x19\x0E\x00\xAA\x2D\x0E\x00\xAB\x25\x0E\x00\xAC\x3D"
-     "\x0E\x00"
-     "\xAD\x25\x0E\x00\xAE\x2D\x0E\x00\xAF\x21\x0E\x00\x80\x0D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x1D\x00\x00\x00\xC4\x00\x80\x04\xC4\x00\x80\x08\xC4\x00\x80\x84\x30"
-     "\x02\x00"
-     "\xBE\x30\x01\x00\xB8\x35\x0E\x00\xB9\x3D\x0E\x00\xBA\x35\x0E\x00\xBB\xCD"
-     "\x0E\x00"
-     "\xBC\xD5\x0E\x00\xBD\xDD\x0E\x00\xBE\xD5\x0E\x00\xBF\xCD\x0E\x00\xB0\x29"
-     "\x0E\x00"
-     "\xB1\x29\x0E\x00\xB2\x39\x0E\x00\xB3\x39\x0E\x00\xB4\x29\x0E\x00\xB5\x15"
-     "\x0E\x00"
-     "\xB6\x11\x0E\x00\xB7\x0D\x0E\x00\xA3\xA5\x0E\x00\x10\xC4\x00\x80\x86\x28"
-     "\x07\x00"
-     "\x87\x4C\x01\x00\x14\xC4\x00\x80\xA6\xA5\x0E\x00\xA5\xB5\x0E\x00\x18\xC4"
-     "\x00\x80"
-     "\xAB\xED\x0E\x00\xAA\xED\x0E\x00\x1C\xC4\x00\x80\x20\xC4\x00\x80\xAF\xD1"
-     "\x0E\x00"
-     "\xAE\xED\x0E\x00\xAD\xED\x0E\x00\xAC\xED\x0E\x00\x24\xC4\x00\x80\xB3\x19"
-     "\x0E\x00"
-     "\x28\xC4\x00\x80\x2C\xC4\x00\x80\xB6\xF9\x01\x00\x30\xC4\x00\x80\x34\xC4"
-     "\x00\x80"
-     "\xB5\x31\x0E\x00\xBA\xD5\x01\x00\xBB\xDD\x01\x00\x38\xC4\x00\x80\x3C\xC4"
-     "\x00\x80"
-     "\xBE\x79\x01\x00\xBF\x79\x01\x00\xBC\xC5\x01\x00\xBD\x79\x01\x00\xA8\xBD"
-     "\x0D\x00"
-     "\xA9\xD5\x0D\x00\xAA\xD5\x0D\x00\xAB\xE5\x0D\x00\xAC\xFD\x0D\x00\xAD\x3D"
-     "\x02\x00"
-     "\xAE\x35\x02\x00\xAF\x2D\x02\x00\x40\xC4\x00\x80\x44\xC4\x00\x80\x48\xC4"
-     "\x00\x80"
-     "\x4C\xC4\x00\x80\x50\xC4\x00\x80\x54\xC4\x00\x80\x58\xC4\x00\x80\x5C\xC4"
-     "\x00\x80"
-     "\xB8\xE9\x02\x00\xB9\xE9\x02\x00\xBA\x8D\x02\x00\xBB\x85\x02\x00\xBC\x81"
-     "\x02\x00"
-     "\xBD\x81\x02\x00\xBE\x85\x02\x00\xBF\xAD\x02\x00\xB0\x55\x02\x00\xB1\x5D"
-     "\x02\x00"
-     "\xB2\x55\x02\x00\xB3\xE9\x02\x00\xB4\xF9\x02\x00\xB5\xF9\x02\x00\xB6\xE9"
-     "\x02\x00"
-     "\xB7\xE1\x02\x00\xA3\x5D\x0D\x00\x60\xC4\x00\x80\x64\xC4\x00\x80\x68\xC4"
-     "\x00\x80"
-     "\x6C\xC4\x00\x80\xA6\xBD\x02\x00\xA5\x75\x0D\x00\x70\xC4\x00\x80\xAB\x99"
-     "\x02\x00"
-     "\xAA\x91\x02\x00\x74\xC4\x00\x80\x78\xC4\x00\x80\xAF\x3D\x02\x00\xAE\x3D"
-     "\x02\x00"
-     "\xAD\x3D\x02\x00\xAC\x81\x02\x00\x80\x5D\x00\x00\x81\x6D\x00\x00\x82\x65"
-     "\x00\x00"
-     "\xB3\x51\x03\x00\x80\xC4\x00\x80\xB5\x79\x03\x00\xB6\x19\x03\x00\x84\xDC"
-     "\x02\x00"
-     "\xBE\x6C\x03\x00\x84\xC4\x00\x80\xBA\x3D\x03\x00\xBB\x35\x03\x00\xBC\x19"
-     "\x03\x00"
-     "\xBD\x19\x03\x00\xBE\xD9\x00\x00\xBF\xD9\x00\x00\xA3\x9D\x02\x00\x88\xC4"
-     "\x00\x80"
-     "\x86\x68\x05\x00\x87\xBC\x02\x00\x8C\xC4\x00\x80\xA6\xD5\x02\x00\xA5\xB5"
-     "\x02\x00"
-     "\x90\xC4\x00\x80\xAB\xF9\x02\x00\xAA\xF1\x02\x00\x84\x20\x02\x00\x94\xC4"
-     "\x00\x80"
-     "\xAF\x15\x01\x00\xAE\x15\x01\x00\xAD\xD5\x02\x00\xAC\xD5\x02\x00\x98\xC4"
-     "\x00\x80"
-     "\x9C\xC4\x00\x80\xEF\x20\x0F\x00\xA0\xC4\x00\x80\xA4\xC4\x00\x80\xA8\xC4"
-     "\x00\x80"
-     "\xAC\xC4\x00\x80\xB0\xC4\x00\x80\xB4\xC4\x00\x80\xB8\xC4\x00\x80\xBC\xC4"
-     "\x00\x80"
-     "\xC0\xC4\x00\x80\xE1\x88\x0C\x00\xC4\xC4\x00\x80\xE3\x44\x0F\x00\xC8\xC4"
-     "\x00\x80"
-     "\xCC\xC4\x00\x80\xD0\xC4\x00\x80\xD4\xC4\x00\x80\xD8\xC4\x00\x80\xBE\x1C"
-     "\x04\x00"
-     "\xDC\xC4\x00\x80\xEF\x5C\x00\x00\xEF\x44\x0E\x00\xE1\x74\x0E\x00\xE1\xFC"
-     "\x0E\x00"
-     "\xE3\x80\x00\x00\xE3\xB0\x0E\x00\x80\x0D\x00\x00\x81\x75\x00\x00\x82\x71"
-     "\x00\x00"
-     "\xE0\xC4\x00\x80\xE4\xC4\x00\x80\xB3\xC1\x00\x00\xE8\xC4\x00\x80\xF0\xC4"
-     "\x00\x80"
-     "\xB6\xFD\x00\x00\xF4\xC4\x00\x80\xF8\xC4\x00\x80\xB5\xC1\x00\x00\xBA\xAD"
-     "\x00\x00"
-     "\xBB\xB5\x00\x00\x86\xC8\x04\x00\x87\xA0\x04\x00\xBE\xAD\x00\x00\xBF\x95"
-     "\x00\x00"
-     "\xBC\xAD\x00\x00\xBD\xA5\x00\x00\xA8\x85\x05\x00\xA9\x95\x05\x00\xAA\x95"
-     "\x05\x00"
-     "\xAB\xA5\x05\x00\xAC\xBD\x05\x00\xAD\xD5\x05\x00\xAE\xD1\x05\x00\xAF\xD1"
-     "\x05\x00"
-     "\x7C\xC4\x00\x80\xEC\xC4\x00\x80\xFC\xC4\x00\x80\x00\xC5\x00\x80\x04\xC5"
-     "\x00\x80"
-     "\x08\xC5\x00\x80\x0C\xC5\x00\x80\x10\xC5\x00\x80\xB8\x71\x06\x00\xB9\x71"
-     "\x06\x00"
-     "\xBA\x71\x06\x00\xBB\x71\x06\x00\xBC\xD5\x06\x00\xBD\xDD\x06\x00\xBE\xD5"
-     "\x06\x00"
-     "\xBF\xCD\x06\x00\xB0\xB5\x05\x00\xB1\xBD\x05\x00\xB2\x81\x05\x00\xB3\x81"
-     "\x05\x00"
-     "\xB4\x51\x06\x00\xB5\x51\x06\x00\xB6\x51\x06\x00\xB7\x51\x06\x00\xB8\xED"
-     "\x07\x00"
-     "\xB9\xF1\x07\x00\xBA\xF5\x07\x00\xBB\x0D\x07\x00\xBC\x15\x07\x00\xBD\x1D"
-     "\x07\x00"
-     "\xBE\x15\x07\x00\xBF\x09\x07\x00\xB0\xED\x07\x00\xB1\xF5\x07\x00\xB2\xFD"
-     "\x07\x00"
-     "\xB3\xF5\x07\x00\xB4\xED\x07\x00\xB5\xD5\x07\x00\xB6\xDD\x07\x00\xB7\xD5"
-     "\x07\x00"
-     "\xA8\x11\x06\x00\xA9\x21\x06\x00\xAA\xCD\x07\x00\xAB\xDD\x07\x00\xAC\xCD"
-     "\x07\x00"
-     "\xAD\xF9\x07\x00\xAE\x99\x07\x00\xAF\x99\x07\x00\x14\xC5\x00\x80\x18\xC5"
-     "\x00\x80"
-     "\x1C\xC5\x00\x80\x20\xC5\x00\x80\x24\xC5\x00\x80\x28\xC5\x00\x80\x0C\xC4"
-     "\x00\x80"
-     "\x2C\xC5\x00\x80\xA3\x05\x06\x00\x30\xC5\x00\x80\x34\xC5\x00\x80\x38\xC5"
-     "\x00\x80"
-     "\x3C\xC5\x00\x80\xA6\x39\x06\x00\xA5\x05\x06\x00\x40\xC5\x00\x80\xAB\x71"
-     "\x06\x00"
-     "\xAA\x69\x06\x00\x44\xC5\x00\x80\x48\xC5\x00\x80\xAF\x51\x06\x00\xAE\x69"
-     "\x06\x00"
-     "\xAD\x61\x06\x00\xAC\x69\x06\x00\x4C\xC5\x00\x80\x82\x1D\x00\x00\x81\x1D"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x50\xC5\x00\x80\x54\xC5\x00\x80\x58\xC5\x00\x80\xBE\x94"
-     "\x01\x00"
-     "\x84\xAC\x01\x00\x5C\xC5\x00\x80\x87\xE8\x01\x00\x86\x00\x7C\x00\x60\xC5"
-     "\x00\x80"
-     "\x64\xC5\x00\x80\x68\xC5\x00\x80\x6C\xC5\x00\x80\xA8\x4D\x06\x00\xA9\x51"
-     "\x06\x00"
-     "\xAA\x51\x06\x00\xAB\xA5\x01\x00\xAC\xBD\x01\x00\xAD\xA5\x01\x00\xAE\xAD"
-     "\x01\x00"
-     "\xAF\xA5\x01\x00\x70\xC5\x00\x80\x74\xC5\x00\x80\x78\xC5\x00\x80\x7C\xC5"
-     "\x00\x80"
-     "\x80\xC5\x00\x80\x84\xC5\x00\x80\x88\xC5\x00\x80\x8C\xC5\x00\x80\xB8\xC5"
-     "\x00\x00"
-     "\xB9\xCD\x00\x00\xBA\xC5\x00\x00\xBB\xDD\x00\x00\xBC\xCD\x00\x00\xBD\xFD"
-     "\x00\x00"
-     "\xBE\xF5\x00\x00\xBF\x9D\x00\x00\xB0\xDD\x01\x00\xB1\x05\x01\x00\xB2\x0D"
-     "\x01\x00"
-     "\xB3\x05\x01\x00\xB4\x1D\x01\x00\xB5\x05\x01\x00\xB6\x0D\x01\x00\xB7\x01"
-     "\x01\x00"
-     "\xB3\xD9\x05\x00\x90\xC5\x00\x80\x94\xC5\x00\x80\x98\xC5\x00\x80\x9C\xC5"
-     "\x00\x80"
-     "\xB6\x25\x02\x00\xB5\x35\x02\x00\xA0\xC5\x00\x80\xBB\xC1\x02\x00\xBA\x19"
-     "\x02\x00"
-     "\xA4\xC5\x00\x80\xA8\xC5\x00\x80\xBF\xC1\x02\x00\xBE\xD9\x02\x00\xBD\xD1"
-     "\x02\x00"
-     "\xBC\xD9\x02\x00\xAC\xC5\x00\x80\xA3\x9D\x05\x00\xB0\xC5\x00\x80\xB4\xC5"
-     "\x00\x80"
-     "\xA6\x61\x02\x00\xB8\xC5\x00\x80\xBC\xC5\x00\x80\xA5\x71\x02\x00\xAA\x5D"
-     "\x02\x00"
-     "\xAB\x85\x02\x00\xC0\xC5\x00\x80\xC4\xC5\x00\x80\xAE\x9D\x02\x00\xAF\x85"
-     "\x02\x00"
-     "\xAC\x9D\x02\x00\xAD\x95\x02\x00\x80\xE9\x00\x00\x81\xE9\x00\x00\x82\x05"
-     "\x00\x00"
-     "\xC8\xC5\x00\x80\xBE\xDC\x03\x00\x84\x58\x03\x00\xD0\xC5\x00\x80\xD4\xC5"
-     "\x00\x80"
-     "\x86\x00\x7C\x00\x87\xEC\x03\x00\xD8\xC5\x00\x80\xDC\xC5\x00\x80\xE0\xC5"
-     "\x00\x80"
-     "\xE4\xC5\x00\x80\xE8\xC5\x00\x80\xEC\xC5\x00\x80\xF0\xC5\x00\x80\xF4\xC5"
-     "\x00\x80"
-     "\xF8\xC5\x00\x80\xFC\xC5\x00\x80\x00\xC6\x00\x80\x04\xC6\x00\x80\xEF\xB0"
-     "\x02\x00"
-     "\x08\xC6\x00\x80\xE1\x94\x01\x00\x0C\xC6\x00\x80\xE3\x6C\x02\x00\x10\xC6"
-     "\x00\x80"
-     "\x14\xC6\x00\x80\x18\xC6\x00\x80\x1C\xC6\x00\x80\x20\xC6\x00\x80\xB3\x79"
-     "\x02\x00"
-     "\x24\xC6\x00\x80\x28\xC6\x00\x80\x2C\xC6\x00\x80\x30\xC6\x00\x80\xB6\xB1"
-     "\x02\x00"
-     "\xB5\xB9\x02\x00\x34\xC6\x00\x80\xBB\x95\x02\x00\xBA\x95\x02\x00\x38\xC6"
-     "\x00\x80"
-     "\x84\x64\x7D\x00\xBF\x39\x01\x00\xBE\x31\x01\x00\xBD\x85\x02\x00\xBC\x85"
-     "\x02\x00"
-     "\x82\x69\x00\x00\xEF\x5C\x07\x00\x80\x59\x00\x00\x81\x51\x00\x00\x40\xC6"
-     "\x00\x80"
-     "\x44\xC6\x00\x80\x87\x68\x7D\x00\x86\xFC\x7C\x00\xBE\x2C\x7D\x00\x48\xC6"
-     "\x00\x80"
-     "\x4C\xC6\x00\x80\x50\xC6\x00\x80\x84\x3C\x7C\x00\xE3\xB8\x07\x00\x54\xC6"
-     "\x00\x80"
-     "\xE1\xEC\x04\x00\x58\xC6\x00\x80\x5C\xC6\x00\x80\x60\xC6\x00\x80\x84\xC0"
-     "\x7F\x00"
-     "\x64\xC6\x00\x80\x68\xC6\x00\x80\xEF\x3C\x00\x00\xEF\xF4\x06\x00\xE1\x50"
-     "\x06\x00"
-     "\xE1\xB4\x06\x00\xE3\xD8\x00\x00\xE3\x94\x06\x00\x6C\xC6\x00\x80\x70\xC6"
-     "\x00\x80"
-     "\x74\xC6\x00\x80\x78\xC6\x00\x80\xA3\xB9\x02\x00\x7C\xC6\x00\x80\x80\xC6"
-     "\x00\x80"
-     "\x84\xC6\x00\x80\x88\xC6\x00\x80\xA6\x71\x02\x00\xA5\x79\x02\x00\x8C\xC6"
-     "\x00\x80"
-     "\xAB\x55\x02\x00\xAA\x55\x02\x00\x90\xC6\x00\x80\x94\xC6\x00\x80\xAF\xF9"
-     "\x01\x00"
-     "\xAE\xF1\x01\x00\xAD\x45\x02\x00\xAC\x45\x02\x00\xA8\x99\x7E\x00\xA9\x99"
-     "\x7E\x00"
-     "\xAA\xA1\x7E\x00\xAB\xA1\x7E\x00\xAC\xA1\x7E\x00\xAD\xA1\x7E\x00\xAE\xA1"
-     "\x7E\x00"
-     "\xAF\xA1\x7E\x00\x98\xC6\x00\x80\x9C\xC6\x00\x80\xA0\xC6\x00\x80\xA4\xC6"
-     "\x00\x80"
-     "\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\xA8\xC6\x00\x80\xB8\xAD"
-     "\x7E\x00"
-     "\xB9\xB5\x7E\x00\xBA\xBD\x7E\x00\xBB\xB5\x7E\x00\xBC\xAD\x7E\x00\xBD\x51"
-     "\x7F\x00"
-     "\xBE\x51\x7F\x00\xBF\x51\x7F\x00\xB0\xA1\x7E\x00\xB1\xA1\x7E\x00\xB2\xA1"
-     "\x7E\x00"
-     "\xB3\xA1\x7E\x00\xB4\xA1\x7E\x00\xB5\xA1\x7E\x00\xB6\xA5\x7E\x00\xB7\x99"
-     "\x7E\x00"
-     "\xCC\xC5\x00\x80\xB3\x2D\x7E\x00\xBE\xC0\x03\x00\x3C\xC6\x00\x80\xB6\x25"
-     "\x7E\x00"
-     "\xAC\xC6\x00\x80\xB0\xC6\x00\x80\xB5\x2D\x7E\x00\xBA\x61\x7E\x00\xBB\x61"
-     "\x7E\x00"
-     "\x86\x08\x00\x00\x87\x9C\x00\x00\xBE\x61\x7E\x00\xBF\x61\x7E\x00\xBC\x71"
-     "\x7E\x00"
-     "\xBD\x71\x7E\x00\xA3\x69\x7E\x00\xB4\xC6\x00\x80\xB8\xC6\x00\x80\xBC\xC6"
-     "\x00\x80"
-     "\xC0\xC6\x00\x80\xA6\x61\x7E\x00\xA5\x69\x7E\x00\xC4\xC6\x00\x80\xAB\x25"
-     "\x7E\x00"
-     "\xAA\x25\x7E\x00\xC8\xC6\x00\x80\xCC\xC6\x00\x80\xAF\x25\x7E\x00\xAE\x25"
-     "\x7E\x00"
-     "\xAD\x35\x7E\x00\xAC\x35\x7E\x00\xA8\x51\x7E\x00\xA9\x51\x7E\x00\xAA\x6D"
-     "\x7E\x00"
-     "\xAB\x65\x7E\x00\xAC\x7D\x7E\x00\xAD\x6D\x7E\x00\xAE\x99\x7E\x00\xAF\x99"
-     "\x7E\x00"
-     "\xD0\xC6\x00\x80\xD4\xC6\x00\x80\xD8\xC6\x00\x80\xDC\xC6\x00\x80\xE0\xC6"
-     "\x00\x80"
-     "\xE4\xC6\x00\x80\xE8\xC6\x00\x80\xEC\xC6\x00\x80\xB8\x89\x7E\x00\xB9\x89"
-     "\x7E\x00"
-     "\xBA\x99\x7E\x00\xBB\x91\x7E\x00\xBC\xB9\x7E\x00\xBD\xB9\x7E\x00\xBE\x79"
-     "\x01\x00"
-     "\xBF\x79\x01\x00\xB0\xE9\x7E\x00\xB1\xF5\x7E\x00\xB2\xFD\x7E\x00\xB3\xDD"
-     "\x7E\x00"
-     "\xB4\xC5\x7E\x00\xB5\xCD\x7E\x00\xB6\xC5\x7E\x00\xB7\xB9\x7E\x00\xB3\xE5"
-     "\x7E\x00"
-     "\xF0\xC6\x00\x80\xF4\xC6\x00\x80\xF8\xC6\x00\x80\xFC\xC6\x00\x80\xB6\xED"
-     "\x7E\x00"
-     "\xB5\xF5\x7E\x00\x00\xC7\x00\x80\xBB\x31\x7E\x00\xBA\x2D\x7E\x00\x04\xC7"
-     "\x00\x80"
-     "\x08\xC7\x00\x80\xBF\x09\x7E\x00\xBE\x01\x7E\x00\xBD\x09\x7E\x00\xBC\x29"
-     "\x7E\x00"
-     "\x0C\xC7\x00\x80\xA3\xA1\x7E\x00\x10\xC7\x00\x80\x14\xC7\x00\x80\xA6\xA9"
-     "\x7E\x00"
-     "\x18\xC7\x00\x80\xBE\x90\x01\x00\xA5\xB1\x7E\x00\xAA\x69\x7E\x00\xAB\x75"
-     "\x7E\x00"
-     "\x85\xEC\x07\x00\x20\xC7\x00\x80\xAE\x45\x7E\x00\xAF\x4D\x7E\x00\xAC\x6D"
-     "\x7E\x00"
-     "\xAD\x4D\x7E\x00\xA8\x99\x02\x00\xA9\xC1\x02\x00\xAA\xC1\x02\x00\xAB\xC1"
-     "\x02\x00"
-     "\xAC\xC1\x02\x00\xAD\xC9\x02\x00\xAE\xF1\x02\x00\xAF\xF1\x02\x00\x86\x84"
-     "\x05\x00"
-     "\x87\x78\x03\x00\x24\xC7\x00\x80\x28\xC7\x00\x80\x80\xF9\x03\x00\x81\xF9"
-     "\x03\x00"
-     "\x82\xED\x03\x00\x2C\xC7\x00\x80\xB8\xC5\x03\x00\xB9\xCD\x03\x00\xBA\xC5"
-     "\x03\x00"
-     "\xBB\xDD\x03\x00\xBC\xCD\x03\x00\xBD\xFD\x03\x00\xBE\xF5\x03\x00\xBF\x9D"
-     "\x03\x00"
-     "\xB0\x41\x03\x00\xB1\x41\x03\x00\xB2\x41\x03\x00\xB3\x41\x03\x00\xB4\x41"
-     "\x03\x00"
-     "\xB5\x41\x03\x00\xB6\x41\x03\x00\xB7\x41\x03\x00\xB3\x49\x02\x00\x30\xC7"
-     "\x00\x80"
-     "\x34\xC7\x00\x80\x38\xC7\x00\x80\x3C\xC7\x00\x80\xB6\x49\x02\x00\xB5\x49"
-     "\x02\x00"
-     "\x40\xC7\x00\x80\xBB\x85\x03\x00\xBA\x85\x03\x00\x44\xC7\x00\x80\x48\xC7"
-     "\x00\x80"
-     "\xBF\x85\x03\x00\xBE\x85\x03\x00\xBD\x95\x03\x00\xBC\x95\x03\x00\x4C\xC7"
-     "\x00\x80"
-     "\xA3\x0D\x02\x00\x50\xC7\x00\x80\x54\xC7\x00\x80\xA6\x0D\x02\x00\x58\xC7"
-     "\x00\x80"
-     "\x5C\xC7\x00\x80\xA5\x0D\x02\x00\xAA\xC1\x03\x00\xAB\xC1\x03\x00\x60\xC7"
-     "\x00\x80"
-     "\x64\xC7\x00\x80\xAE\xC1\x03\x00\xAF\xC1\x03\x00\xAC\xD1\x03\x00\xAD\xD1"
-     "\x03\x00"
-     "\xE3\x60\x7E\x00\xE1\xBC\x7F\x00\xE1\x4C\x7E\x00\xE3\x30\x7E\x00\xE1\xA0"
-     "\x01\x00"
-     "\xE1\xE8\x00\x00\xE3\x84\x7E\x00\xE3\xC4\x17\x00\xBE\xAC\x07\x00\x68\xC7"
-     "\x00\x80"
-     "\xEF\x08\x00\x00\xEF\x7C\x1C\x00\xEF\x40\x01\x00\x6C\xC7\x00\x80\x70\xC7"
-     "\x00\x80"
-     "\xEF\xA4\x7E\x00\x74\xC7\x00\x80\x78\xC7\x00\x80\x84\x04\x04\x00\xBE\xC0"
-     "\x06\x00"
-     "\x7C\xC7\x00\x80\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D\x00\x00\x80\xC7"
-     "\x00\x80"
-     "\x84\xC7\x00\x80\x88\xC7\x00\x80\x90\xC7\x00\x80\x86\x50\x07\x00\x87\x84"
-     "\x05\x00"
-     "\x94\xC7\x00\x80\x98\xC7\x00\x80\xBB\xC1\x7D\x00\xBA\xA9\x7D\x00\xB9\xA1"
-     "\x7D\x00"
-     "\xB8\x91\x7D\x00\xBF\xC9\x7D\x00\xBE\xC1\x7D\x00\xBD\xD5\x7D\x00\xBC\xD5"
-     "\x7D\x00"
-     "\xB3\x1D\x7E\x00\xB2\x1D\x7E\x00\xB1\x1D\x7E\x00\xB0\x1D\x7E\x00\xB7\xA1"
-     "\x7D\x00"
-     "\xB6\xB9\x7D\x00\xB5\xB1\x7D\x00\xB4\xB9\x7D\x00\xAB\xBD\x7E\x00\xAA\xBD"
-     "\x7E\x00"
-     "\xA9\xBD\x7E\x00\xA8\xBD\x7E\x00\xAF\x7D\x7E\x00\xAE\x7D\x7E\x00\xAD\x7D"
-     "\x7E\x00"
-     "\xAC\x7D\x7E\x00\x9C\xC7\x00\x80\xA0\xC7\x00\x80\xA4\xC7\x00\x80\xA8\xC7"
-     "\x00\x80"
-     "\xAC\xC7\x00\x80\xB0\xC7\x00\x80\xB4\xC7\x00\x80\xBE\xB0\x06\x00\x82\x05"
-     "\x6C\x00"
-     "\x83\x19\x6C\x00\x1C\xC7\x00\x80\x8C\xC7\x00\x80\x86\xF9\x68\x00\x87\xC9"
-     "\x74\x00"
-     "\x84\x19\x68\x00\x85\x19\x68\x00\x8A\x01\x74\x00\x8B\xB9\x74\x00\xB8\xC7"
-     "\x00\x80"
-     "\xBC\xC7\x00\x80\x8E\xB5\x70\x00\x8F\xC1\x7C\x00\x8C\xD9\x70\x00\x8D\x1D"
-     "\x70\x00"
-     "\x92\x61\x7C\x00\x93\x61\x7C\x00\xC0\xC7\x00\x80\xC4\xC7\x00\x80\x96\x75"
-     "\x78\x00"
-     "\x97\x31\x79\x00\x94\xA1\x78\x00\x95\x7D\x78\x00\x9A\xE1\x04\x00\x9B\xE9"
-     "\x04\x00"
-     "\xC8\xC7\x00\x80\xCC\xC7\x00\x80\xD0\xC7\x00\x80\xD4\xC7\x00\x80\x9C\xED"
-     "\x04\x00"
-     "\xD8\xC7\x00\x80\xA1\x89\x00\x00\xDC\xC7\x00\x80\xA3\x01\x0C\x00\xA2\x81"
-     "\x00\x00"
-     "\xA5\x15\x0C\x00\xA4\xB5\x0C\x00\xA7\xB9\x08\x00\xA6\x01\x08\x00\xA9\xF9"
-     "\x09\x00"
-     "\xA8\x81\x08\x00\xAB\x25\x14\x00\xAA\x2D\x14\x00\xAD\x01\x10\x00\xAC\x21"
-     "\x14\x00"
-     "\xAF\x99\x11\x00\xAE\x9D\x10\x00\xB1\xE1\x1C\x00\xB0\x01\x1C\x00\xB3\x01"
-     "\x18\x00"
-     "\xB2\x65\x1C\x00\xB5\xF5\x18\x00\xB4\xF5\x18\x00\xE0\xC7\x00\x80\xE4\xC7"
-     "\x00\x80"
-     "\xE8\xC7\x00\x80\xEC\xC7\x00\x80\xF0\xC7\x00\x80\xF4\xC7\x00\x80\x80\x19"
-     "\x00\x00"
-     "\x81\x19\x00\x00\x82\x05\x00\x00\xF8\xC7\x00\x80\xA9\x29\x07\x00\xA8\xE9"
-     "\x04\x00"
-     "\xAB\x35\x07\x00\xAA\x21\x07\x00\xAD\x99\x07\x00\xAC\x91\x07\x00\xAF\x89"
-     "\x07\x00"
-     "\xAE\x91\x07\x00\xA0\x45\x78\x00\xA1\xA5\x05\x00\xA2\xA5\x05\x00\xA3\xB9"
-     "\x05\x00"
-     "\x84\xB8\x01\x00\x00\xC8\x00\x80\x87\x34\x00\x00\x86\x9C\x0F\x00\xB9\xAD"
-     "\x07\x00"
-     "\xB8\xAD\x07\x00\xBB\x7D\x06\x00\xBA\x69\x06\x00\xBD\x11\x06\x00\xBC\x6D"
-     "\x06\x00"
-     "\xBF\x09\x06\x00\xBE\x19\x06\x00\xB1\xC9\x07\x00\xB0\xF9\x07\x00\xB3\xA9"
-     "\x07\x00"
-     "\xB2\xC1\x07\x00\xB5\xBD\x07\x00\xB4\xBD\x07\x00\xB7\xA1\x07\x00\xB6\xA9"
-     "\x07\x00"
-     "\xB3\xAD\x07\x00\x04\xC8\x00\x80\x08\xC8\x00\x80\x84\x0C\x00\x00\x0C\xC8"
-     "\x00\x80"
-     "\xB6\x1D\x07\x00\xB5\x1D\x07\x00\x8F\x99\x6D\x00\xBB\x09\x07\x00\xBA\x39"
-     "\x07\x00"
-     "\x10\xC8\x00\x80\x14\xC8\x00\x80\xBF\x7D\x07\x00\xBE\x7D\x07\x00\xBD\x79"
-     "\x07\x00"
-     "\xBC\x11\x07\x00\x98\xE1\x71\x00\x99\x41\x71\x00\x9A\x3D\x7D\x00\x9B\xED"
-     "\x7D\x00"
-     "\x9C\xF5\x7D\x00\x9D\x2D\x79\x00\x9E\xF9\x79\x00\x9F\xF9\x79\x00\x90\xFD"
-     "\x6D\x00"
-     "\x91\xE5\x69\x00\x92\xED\x69\x00\x93\x05\x69\x00\x94\x2D\x75\x00\x95\xD5"
-     "\x75\x00"
-     "\x96\x65\x75\x00\x97\x1D\x71\x00\xA3\x2D\x04\x00\x84\xC8\x0D\x00\xBE\xC0"
-     "\x0D\x00"
-     "\x18\xC8\x00\x80\x1C\xC8\x00\x80\xA6\x9D\x04\x00\xA5\x9D\x04\x00\x20\xC8"
-     "\x00\x80"
-     "\xAB\x89\x04\x00\xAA\xB9\x04\x00\x24\xC8\x00\x80\x28\xC8\x00\x80\xAF\xFD"
-     "\x04\x00"
-     "\xAE\xFD\x04\x00\xAD\xF9\x04\x00\xAC\x91\x04\x00\x2C\xC8\x00\x80\xB3\x39"
-     "\x05\x00"
-     "\x30\xC8\x00\x80\x34\xC8\x00\x80\xB6\xB9\x05\x00\x38\xC8\x00\x80\x3C\xC8"
-     "\x00\x80"
-     "\xB5\xB9\x05\x00\xBA\x95\x05\x00\xBB\x9D\x05\x00\x40\xC8\x00\x80\x44\xC8"
-     "\x00\x80"
-     "\xBE\x39\x02\x00\xBF\x39\x02\x00\xBC\x51\x02\x00\xBD\x51\x02\x00\xA8\xED"
-     "\x05\x00"
-     "\xA9\xFD\x05\x00\xAA\xF5\x05\x00\xAB\x6D\x05\x00\xAC\x75\x05\x00\xAD\x7D"
-     "\x05\x00"
-     "\xAE\x75\x05\x00\xAF\x6D\x05\x00\x48\xC8\x00\x80\x4C\xC8\x00\x80\x50\xC8"
-     "\x00\x80"
-     "\x54\xC8\x00\x80\x58\xC8\x00\x80\x5C\xC8\x00\x80\x60\xC8\x00\x80\x64\xC8"
-     "\x00\x80"
-     "\xB8\xB1\x02\x00\xB9\xB1\x02\x00\xBA\xB5\x02\x00\xBB\x49\x02\x00\xBC\x09"
-     "\x02\x00"
-     "\xBD\x09\x02\x00\xBE\x39\x02\x00\xBF\x39\x02\x00\xB0\x15\x05\x00\xB1\x1D"
-     "\x05\x00"
-     "\xB2\xC1\x02\x00\xB3\xC1\x02\x00\xB4\xB9\x02\x00\xB5\xB9\x02\x00\xB6\x91"
-     "\x02\x00"
-     "\xB7\x91\x02\x00\xA3\x7D\x04\x00\x82\x15\x00\x00\x81\xB5\x01\x00\x80\xB5"
-     "\x01\x00"
-     "\x68\xC8\x00\x80\xA6\xFD\x04\x00\xA5\xFD\x04\x00\x70\xC8\x00\x80\xAB\xD9"
-     "\x04\x00"
-     "\xAA\xD1\x04\x00\x86\x28\x0D\x00\x87\x1C\x0D\x00\xAF\x7D\x03\x00\xAE\x7D"
-     "\x03\x00"
-     "\xAD\x15\x03\x00\xAC\x15\x03\x00\x74\xC8\x00\x80\x78\xC8\x00\x80\x7C\xC8"
-     "\x00\x80"
-     "\xB3\xC5\x02\x00\x80\xC8\x00\x80\xB5\xFD\x02\x00\xB6\xF5\x02\x00\x84\xC8"
-     "\x00\x80"
-     "\x88\xC8\x00\x80\x8C\xC8\x00\x80\xBA\x75\x01\x00\xBB\x45\x01\x00\xBC\x5D"
-     "\x01\x00"
-     "\xBD\x35\x01\x00\xBE\x31\x01\x00\xBF\x31\x01\x00\xE3\xA0\x00\x00\xE3\xDC"
-     "\x06\x00"
-     "\xE1\xE0\x06\x00\xE1\xB4\x07\x00\x90\xC8\x00\x80\x94\xC8\x00\x80\x98\xC8"
-     "\x00\x80"
-     "\xBE\x90\x0F\x00\x86\x80\x0C\x00\x87\x18\x0C\x00\x9C\xC8\x00\x80\xA0\xC8"
-     "\x00\x80"
-     "\xEF\x04\x06\x00\xEF\x6C\x06\x00\xA4\xC8\x00\x80\xA8\xC8\x00\x80\xA3\x49"
-     "\x02\x00"
-     "\xAC\xC8\x00\x80\xB0\xC8\x00\x80\xB4\xC8\x00\x80\xB8\xC8\x00\x80\xA6\x79"
-     "\x02\x00"
-     "\xA5\x71\x02\x00\xBC\xC8\x00\x80\xAB\xC9\x01\x00\xAA\xF9\x01\x00\xC0\xC8"
-     "\x00\x80"
-     "\xC4\xC8\x00\x80\xAF\xBD\x01\x00\xAE\xBD\x01\x00\xAD\xB9\x01\x00\xAC\xD1"
-     "\x01\x00"
-     "\xA8\x35\x0D\x00\xA9\x81\x0D\x00\xAA\x81\x0D\x00\xAB\x9D\x0D\x00\xAC\xB9"
-     "\x0D\x00"
-     "\xAD\xB9\x0D\x00\xAE\x15\x0E\x00\xAF\x0D\x0E\x00\x6C\xC8\x00\x80\xC8\xC8"
-     "\x00\x80"
-     "\xCC\xC8\x00\x80\xD0\xC8\x00\x80\x80\x1D\x00\x00\x81\x09\x00\x00\x82\xB9"
-     "\x00\x00"
-     "\xD4\xC8\x00\x80\xB8\x29\x0E\x00\xB9\x29\x0E\x00\xBA\xE9\x0E\x00\xBB\xE9"
-     "\x0E\x00"
-     "\xBC\xF9\x0E\x00\xBD\xF9\x0E\x00\xBE\xE9\x0E\x00\xBF\xE9\x0E\x00\xB0\x7D"
-     "\x0E\x00"
-     "\xB1\x4D\x0E\x00\xB2\x45\x0E\x00\xB3\x2D\x0E\x00\xB4\x39\x0E\x00\xB5\x39"
-     "\x0E\x00"
-     "\xB6\x2D\x0E\x00\xB7\x25\x0E\x00\xD8\xC8\x00\x80\xDC\xC8\x00\x80\xFC\xC7"
-     "\x00\x80"
-     "\xE0\xC8\x00\x80\xE4\xC8\x00\x80\xE8\xC8\x00\x80\xEC\xC8\x00\x80\xF0\xC8"
-     "\x00\x80"
-     "\xA8\xBD\x0F\x00\xA9\xE5\x0F\x00\xAA\xED\x0F\x00\xAB\xE5\x0F\x00\xAC\xE1"
-     "\x0F\x00"
-     "\xAD\xE9\x0F\x00\xAE\xD9\x0F\x00\xAF\xD9\x0F\x00\xB0\x51\x0F\x00\xB1\x51"
-     "\x0F\x00"
-     "\xB2\x51\x0F\x00\xB3\x65\x0F\x00\xB4\x7D\x0F\x00\xB5\x6D\x0F\x00\xB6\x65"
-     "\x0F\x00"
-     "\xB7\x19\x0F\x00\xB8\x29\x0F\x00\xB9\x29\x0F\x00\xBA\x39\x0F\x00\xBB\x31"
-     "\x0F\x00"
-     "\xBC\x19\x0F\x00\xBD\x19\x0F\x00\xBE\xDD\x00\x00\xBF\xD5\x00\x00\xB3\x45"
-     "\x0E\x00"
-     "\xF4\xC8\x00\x80\xF8\xC8\x00\x80\xFC\xC8\x00\x80\x00\xC9\x00\x80\xB6\x49"
-     "\x0E\x00"
-     "\xB5\x49\x0E\x00\x04\xC9\x00\x80\xBB\xB5\x0E\x00\xBA\xB5\x0E\x00\x08\xC9"
-     "\x00\x80"
-     "\x0C\xC9\x00\x80\xBF\x89\x0E\x00\xBE\x99\x0E\x00\xBD\x91\x0E\x00\xBC\xA5"
-     "\x0E\x00"
-     "\x10\xC9\x00\x80\xA3\x01\x0E\x00\x14\xC9\x00\x80\x18\xC9\x00\x80\xA6\x0D"
-     "\x0E\x00"
-     "\x1C\xC9\x00\x80\x20\xC9\x00\x80\xA5\x0D\x0E\x00\xAA\xF1\x0E\x00\xAB\xF1"
-     "\x0E\x00"
-     "\x24\xC9\x00\x80\x28\xC9\x00\x80\xAE\xDD\x0E\x00\xAF\xCD\x0E\x00\xAC\xE1"
-     "\x0E\x00"
-     "\xAD\xD5\x0E\x00\xA8\x0D\x01\x00\xA9\x15\x01\x00\xAA\x1D\x01\x00\xAB\x15"
-     "\x01\x00"
-     "\xAC\x0D\x01\x00\xAD\x91\x00\x00\xAE\x91\x00\x00\xAF\x91\x00\x00\x2C\xC9"
-     "\x00\x80"
-     "\x30\xC9\x00\x80\x34\xC9\x00\x80\x38\xC9\x00\x80\x3C\xC9\x00\x80\x82\xB9"
-     "\x03\x00"
-     "\x81\xB1\x03\x00\x80\xB9\x03\x00\xB8\x99\x00\x00\xB9\xAD\x00\x00\xBA\xA5"
-     "\x00\x00"
-     "\xBB\x6D\x00\x00\xBC\x75\x00\x00\xBD\x7D\x00\x00\xBE\x75\x00\x00\xBF\x6D"
-     "\x00\x00"
-     "\xB0\xF1\x00\x00\xB1\xF9\x00\x00\xB2\xC1\x00\x00\xB3\xC1\x00\x00\xB4\xB1"
-     "\x00\x00"
-     "\xB5\xBD\x00\x00\xB6\xB5\x00\x00\xB7\xA9\x00\x00\x84\xAC\x02\x00\x40\xC9"
-     "\x00\x80"
-     "\x44\xC9\x00\x80\x48\xC9\x00\x80\x86\xB0\x04\x00\x87\xD8\x02\x00\x50\xC9"
-     "\x00\x80"
-     "\x54\xC9\x00\x80\x58\xC9\x00\x80\x5C\xC9\x00\x80\x60\xC9\x00\x80\x64\xC9"
-     "\x00\x80"
-     "\xBE\x74\x03\x00\x68\xC9\x00\x80\x6C\xC9\x00\x80\x70\xC9\x00\x80\xBB\x5D"
-     "\x03\x00"
-     "\xBA\x55\x03\x00\x74\xC9\x00\x80\x78\xC9\x00\x80\xBF\x19\x03\x00\xBE\x15"
-     "\x03\x00"
-     "\xBD\x75\x03\x00\xBC\x75\x03\x00\xB3\xF5\x02\x00\x7C\xC9\x00\x80\x80\xC9"
-     "\x00\x80"
-     "\x84\xC9\x00\x80\x88\xC9\x00\x80\xB6\x71\x03\x00\xB5\xE5\x02\x00\x8C\xC9"
-     "\x00\x80"
-     "\xA3\x15\x02\x00\x90\xC9\x00\x80\x94\xC9\x00\x80\x98\xC9\x00\x80\x9C\xC9"
-     "\x00\x80"
-     "\xA6\x91\x03\x00\xA5\x05\x02\x00\xA0\xC9\x00\x80\xAB\xBD\x03\x00\xAA\xB5"
-     "\x03\x00"
-     "\xA4\xC9\x00\x80\xA8\xC9\x00\x80\xAF\xF9\x03\x00\xAE\xF5\x03\x00\xAD\x95"
-     "\x03\x00"
-     "\xAC\x95\x03\x00\xAC\xC9\x00\x80\x82\x1D\x00\x00\x81\x1D\x00\x00\x80\x1D"
-     "\x00\x00"
-     "\xB0\xC9\x00\x80\xB4\xC9\x00\x80\xB8\xC9\x00\x80\xC0\xC9\x00\x80\x84\x6C"
-     "\x03\x00"
-     "\xC4\xC9\x00\x80\x87\x78\x03\x00\x86\xEC\x04\x00\xC8\xC9\x00\x80\xCC\xC9"
-     "\x00\x80"
-     "\xD0\xC9\x00\x80\xD4\xC9\x00\x80\xE1\xB8\x0E\x00\xE3\x98\x0C\x00\xE3\x18"
-     "\x03\x00"
-     "\xE1\x28\x0C\x00\xD8\xC9\x00\x80\xDC\xC9\x00\x80\xE0\xC9\x00\x80\xE4\xC9"
-     "\x00\x80"
-     "\x84\xEC\x04\x00\xE8\xC9\x00\x80\xEC\xC9\x00\x80\xF0\xC9\x00\x80\xF4\xC9"
-     "\x00\x80"
-     "\xEF\x30\x0A\x00\xEF\x5C\x0E\x00\xF8\xC9\x00\x80\xB3\x61\x01\x00\xFC\xC9"
-     "\x00\x80"
-     "\x00\xCA\x00\x80\x04\xCA\x00\x80\x08\xCA\x00\x80\xB6\x79\x01\x00\xB5\x71"
-     "\x01\x00"
-     "\x0C\xCA\x00\x80\xBB\x29\x01\x00\xBA\x59\x01\x00\x10\xCA\x00\x80\x14\xCA"
-     "\x00\x80"
-     "\xBF\x1D\x01\x00\xBE\x1D\x01\x00\xBD\x1D\x01\x00\xBC\x31\x01\x00\x18\xCA"
-     "\x00\x80"
-     "\x1C\xCA\x00\x80\x20\xCA\x00\x80\x24\xCA\x00\x80\x80\x19\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x82\x39\x00\x00\xEF\x6C\x0F\x00\x28\xCA\x00\x80\xE1\x94\x0E\x00\xBE\x60"
-     "\x04\x00"
-     "\xE3\xB8\x0F\x00\x30\xCA\x00\x80\x34\xCA\x00\x80\x38\xCA\x00\x80\xEF\x00"
-     "\x00\x00"
-     "\x86\xE0\x04\x00\x87\x8C\x04\x00\xA3\xE9\x01\x00\x84\x00\x07\x00\xA5\xF9"
-     "\x01\x00"
-     "\x3C\xCA\x00\x80\x40\xCA\x00\x80\xA6\xF1\x01\x00\x44\xCA\x00\x80\x48\xCA"
-     "\x00\x80"
-     "\xAB\xA1\x01\x00\xAA\xD1\x01\x00\xAD\x95\x01\x00\xAC\xB9\x01\x00\xAF\x95"
-     "\x01\x00"
-     "\xAE\x95\x01\x00\xA8\x19\x06\x00\xA9\x19\x06\x00\xAA\x29\x06\x00\xAB\x29"
-     "\x06\x00"
-     "\xAC\x39\x06\x00\xAD\x39\x06\x00\xAE\x2D\x06\x00\xAF\x25\x06\x00\xBC\xC9"
-     "\x00\x80"
-     "\x2C\xCA\x00\x80\x4C\xCA\x00\x80\x50\xCA\x00\x80\x54\xCA\x00\x80\x58\xCA"
-     "\x00\x80"
-     "\x5C\xCA\x00\x80\x60\xCA\x00\x80\xB8\xC5\x06\x00\xB9\xCD\x06\x00\xBA\xC5"
-     "\x06\x00"
-     "\xBB\xDD\x06\x00\xBC\xCD\x06\x00\xBD\xFD\x06\x00\xBE\xF5\x06\x00\xBF\x9D"
-     "\x06\x00"
-     "\xB0\x5D\x06\x00\xB1\x2D\x06\x00\xB2\x25\x06\x00\xB3\x09\x06\x00\xB4\x19"
-     "\x06\x00"
-     "\xB5\x19\x06\x00\xB6\x09\x06\x00\xB7\x01\x06\x00\x64\xCA\x00\x80\x68\xCA"
-     "\x00\x80"
-     "\x6C\xCA\x00\x80\x70\xCA\x00\x80\x74\xCA\x00\x80\x78\xCA\x00\x80\x7C\xCA"
-     "\x00\x80"
-     "\x4C\xC9\x00\x80\xA8\x7D\x06\x00\xA9\x41\x06\x00\xAA\x59\x06\x00\xAB\x59"
-     "\x06\x00"
-     "\xAC\x49\x06\x00\xAD\x49\x06\x00\xAE\xA9\x07\x00\xAF\xA9\x07\x00\xB0\xD9"
-     "\x07\x00"
-     "\xB1\xD9\x07\x00\xB2\xED\x07\x00\xB3\xE5\x07\x00\xB4\xFD\x07\x00\xB5\xE5"
-     "\x07\x00"
-     "\xB6\xE5\x07\x00\xB7\x99\x07\x00\xB8\xA9\x07\x00\xB9\xA9\x07\x00\xBA\xB9"
-     "\x07\x00"
-     "\xBB\xB1\x07\x00\xBC\x99\x07\x00\xBD\x99\x07\x00\xBE\x79\x07\x00\xBF\x79"
-     "\x07\x00"
-     "\xB3\xF1\x07\x00\x80\xCA\x00\x80\x84\xCA\x00\x80\x88\xCA\x00\x80\x8C\xCA"
-     "\x00\x80"
-     "\xB6\xE9\x07\x00\xB5\xE1\x07\x00\x90\xCA\x00\x80\xBB\x3D\x06\x00\xBA\x3D"
-     "\x06\x00"
-     "\x94\xCA\x00\x80\x98\xCA\x00\x80\xBF\x01\x06\x00\xBE\x19\x06\x00\xBD\x11"
-     "\x06\x00"
-     "\xBC\x19\x06\x00\x82\x6D\x00\x00\xA3\xB5\x07\x00\x80\x55\x00\x00\x81\x65"
-     "\x00\x00"
-     "\xA6\xAD\x07\x00\xBE\x9C\x01\x00\x9C\xCA\x00\x80\xA5\xA5\x07\x00\xAA\x79"
-     "\x06\x00"
-     "\xAB\x79\x06\x00\x86\x00\x1C\x00\x87\xCC\x01\x00\xAE\x5D\x06\x00\xAF\x45"
-     "\x06\x00"
-     "\xAC\x5D\x06\x00\xAD\x55\x06\x00\xA8\x31\x06\x00\xA9\x31\x06\x00\xAA\x31"
-     "\x06\x00"
-     "\xAB\x59\x06\x00\xAC\x49\x06\x00\xAD\x75\x06\x00\xAE\x7D\x06\x00\xAF\xB5"
-     "\x01\x00"
-     "\xA0\xCA\x00\x80\xA4\xCA\x00\x80\xA8\xCA\x00\x80\xAC\xCA\x00\x80\xB0\xCA"
-     "\x00\x80"
-     "\xB4\xCA\x00\x80\xB8\xCA\x00\x80\xBC\xCA\x00\x80\xB8\x4D\x01\x00\xB9\x55"
-     "\x01\x00"
-     "\xBA\x5D\x01\x00\xBB\x55\x01\x00\xBC\xD5\x00\x00\xBD\xDD\x00\x00\xBE\xD5"
-     "\x00\x00"
-     "\xBF\xCD\x00\x00\xB0\xCD\x01\x00\xB1\xD5\x01\x00\xB2\xDD\x01\x00\xB3\xD5"
-     "\x01\x00"
-     "\xB4\xCD\x01\x00\xB5\x75\x01\x00\xB6\x7D\x01\x00\xB7\x75\x01\x00\xB3\xB9"
-     "\x05\x00"
-     "\xC0\xCA\x00\x80\xC4\xCA\x00\x80\xC8\xCA\x00\x80\xCC\xCA\x00\x80\xB6\xF5"
-     "\x05\x00"
-     "\xB5\xC1\x05\x00\xD0\xCA\x00\x80\xBB\x11\x02\x00\xBA\x09\x02\x00\xD4\xCA"
-     "\x00\x80"
-     "\xD8\xCA\x00\x80\xBF\xB1\x02\x00\xBE\x09\x02\x00\xBD\x01\x02\x00\xBC\x09"
-     "\x02\x00"
-     "\xDC\xCA\x00\x80\xA3\xFD\x05\x00\xE0\xCA\x00\x80\xE4\xCA\x00\x80\xA6\xB1"
-     "\x05\x00"
-     "\xE8\xCA\x00\x80\xEC\xCA\x00\x80\xA5\x85\x05\x00\xAA\x4D\x02\x00\xAB\x55"
-     "\x02\x00"
-     "\xF0\xCA\x00\x80\xF4\xCA\x00\x80\xAE\x4D\x02\x00\xAF\xF5\x02\x00\xAC\x4D"
-     "\x02\x00"
-     "\xAD\x45\x02\x00\xF8\xCA\x00\x80\xFC\xCA\x00\x80\x00\xCB\x00\x80\x04\xCB"
-     "\x00\x80"
-     "\x08\xCB\x00\x80\xBE\xD8\x03\x00\x10\xCB\x00\x80\x14\xCB\x00\x80\x80\x29"
-     "\x00\x00"
-     "\x81\x29\x00\x00\x82\x39\x00\x00\x84\x20\x03\x00\x86\x10\x1C\x00\x87\x84"
-     "\x03\x00"
-     "\x18\xCB\x00\x80\x1C\xCB\x00\x80\x20\xCB\x00\x80\x24\xCB\x00\x80\x28\xCB"
-     "\x00\x80"
-     "\x2C\xCB\x00\x80\x30\xCB\x00\x80\xE3\x68\x02\x00\x34\xCB\x00\x80\xE1\xAC"
-     "\x01\x00"
-     "\x38\xCB\x00\x80\xEF\x88\x02\x00\x3C\xCB\x00\x80\x40\xCB\x00\x80\x44\xCB"
-     "\x00\x80"
-     "\x48\xCB\x00\x80\x4C\xCB\x00\x80\x50\xCB\x00\x80\x54\xCB\x00\x80\x58\xCB"
-     "\x00\x80"
-     "\x5C\xCB\x00\x80\x60\xCB\x00\x80\xE3\xE0\x07\x00\x64\xCB\x00\x80\xE1\x10"
-     "\x07\x00"
-     "\x68\xCB\x00\x80\xEF\xE8\x07\x00\x6C\xCB\x00\x80\x70\xCB\x00\x80\x74\xCB"
-     "\x00\x80"
-     "\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x78\xCB\x00\x80\xB3\xE9"
-     "\x02\x00"
-     "\x84\x0C\x1D\x00\x86\xE8\x1C\x00\x87\x80\x1D\x00\x80\xCB\x00\x80\xB6\xE1"
-     "\x02\x00"
-     "\xB5\xE9\x02\x00\x84\xCB\x00\x80\xBB\x69\x01\x00\xBA\x61\x01\x00\xBE\x24"
-     "\x1D\x00"
-     "\x88\xCB\x00\x80\xBF\x59\x01\x00\xBE\x69\x01\x00\xBD\x69\x01\x00\xBC\x71"
-     "\x01\x00"
-     "\x8C\xCB\x00\x80\x90\xCB\x00\x80\x94\xCB\x00\x80\x98\xCB\x00\x80\x9C\xCB"
-     "\x00\x80"
-     "\xA0\xCB\x00\x80\xEF\x34\x00\x00\xEF\xCC\x06\x00\xE1\x78\x06\x00\xE1\x50"
-     "\x07\x00"
-     "\xE3\x74\x06\x00\xE3\x9C\x06\x00\xA4\xCB\x00\x80\xA8\xCB\x00\x80\xAC\xCB"
-     "\x00\x80"
-     "\xB0\xCB\x00\x80\xA3\x69\x02\x00\xB4\xCB\x00\x80\x84\x84\x1C\x00\xBE\x84"
-     "\x1C\x00"
-     "\xB8\xCB\x00\x80\xA6\x61\x02\x00\xA5\x69\x02\x00\xBC\xCB\x00\x80\xAB\xE9"
-     "\x01\x00"
-     "\xAA\xE1\x01\x00\xC0\xCB\x00\x80\xC4\xCB\x00\x80\xAF\xD9\x01\x00\xAE\xE9"
-     "\x01\x00"
-     "\xAD\xE9\x01\x00\xAC\xF1\x01\x00\xA8\x99\x1E\x00\xA9\x99\x1E\x00\xAA\xB1"
-     "\x1E\x00"
-     "\xAB\xB1\x1E\x00\xAC\xD1\x1E\x00\xAD\xD1\x1E\x00\xAE\xD1\x1E\x00\xAF\xD1"
-     "\x1E\x00"
-     "\xC8\xCB\x00\x80\xCC\xCB\x00\x80\xD0\xCB\x00\x80\xD4\xCB\x00\x80\xD8\xCB"
-     "\x00\x80"
-     "\xDC\xCB\x00\x80\xE0\xCB\x00\x80\xE4\xCB\x00\x80\xB8\xBD\x1E\x00\xB9\x45"
-     "\x1F\x00"
-     "\xBA\x4D\x1F\x00\xBB\x45\x1F\x00\xBC\x5D\x1F\x00\xBD\x41\x1F\x00\xBE\x45"
-     "\x1F\x00"
-     "\xBF\x79\x1F\x00\xB0\xB1\x1E\x00\xB1\xBD\x1E\x00\xB2\x89\x1E\x00\xB3\x85"
-     "\x1E\x00"
-     "\xB4\x9D\x1E\x00\xB5\x85\x1E\x00\xB6\x8D\x1E\x00\xB7\x85\x1E\x00\x0C\xCB"
-     "\x00\x80"
-     "\xB3\x3D\x1E\x00\xE8\xCB\x00\x80\x7C\xCB\x00\x80\xB6\x55\x1E\x00\xEC\xCB"
-     "\x00\x80"
-     "\xF0\xCB\x00\x80\xB5\x5D\x1E\x00\xBA\x71\x1E\x00\xBB\x7D\x1E\x00\x86\x08"
-     "\x00\x00"
-     "\x87\x28\x00\x00\xBE\x5D\x1E\x00\xBF\x45\x1E\x00\xBC\x59\x1E\x00\xBD\x55"
-     "\x1E\x00"
-     "\xA3\x79\x1E\x00\x82\x4D\x00\x00\x81\xF5\x00\x00\x80\xF5\x00\x00\xF4\xCB"
-     "\x00\x80"
-     "\xA6\x11\x1E\x00\xA5\x19\x1E\x00\xF8\xCB\x00\x80\xAB\x39\x1E\x00\xAA\x35"
-     "\x1E\x00"
-     "\xFC\xCB\x00\x80\x00\xCC\x00\x80\xAF\x01\x1E\x00\xAE\x19\x1E\x00\xAD\x11"
-     "\x1E\x00"
-     "\xAC\x1D\x1E\x00\xA8\x51\x1E\x00\xA9\x59\x1E\x00\xAA\x65\x1E\x00\xAB\x8D"
-     "\x1E\x00"
-     "\xAC\x95\x1E\x00\xAD\x9D\x1E\x00\xAE\x95\x1E\x00\xAF\x8D\x1E\x00\x04\xCC"
-     "\x00\x80"
-     "\x08\xCC\x00\x80\x0C\xCC\x00\x80\x10\xCC\x00\x80\x14\xCC\x00\x80\x18\xCC"
-     "\x00\x80"
-     "\x1C\xCC\x00\x80\x20\xCC\x00\x80\xB8\xA9\x1E\x00\xB9\xA9\x1E\x00\xBA\x69"
-     "\x01\x00"
-     "\xBB\x69\x01\x00\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x69\x01\x00\xBF\x69"
-     "\x01\x00"
-     "\xB0\xF5\x1E\x00\xB1\xD5\x1E\x00\xB2\xDD\x1E\x00\xB3\xD5\x1E\x00\xB4\xCD"
-     "\x1E\x00"
-     "\xB5\xB1\x1E\x00\xB6\xA9\x1E\x00\xB7\xA1\x1E\x00\xB3\xED\x1E\x00\x24\xCC"
-     "\x00\x80"
-     "\x28\xCC\x00\x80\x2C\xCC\x00\x80\x30\xCC\x00\x80\xB6\x1D\x1E\x00\xB5\x1D"
-     "\x1E\x00"
-     "\x34\xCC\x00\x80\xBB\x11\x1E\x00\xBA\x31\x1E\x00\x38\xCC\x00\x80\x3C\xCC"
-     "\x00\x80"
-     "\xBF\x75\x1E\x00\xBE\x09\x1E\x00\xBD\x01\x1E\x00\xBC\x09\x1E\x00\x40\xCC"
-     "\x00\x80"
-     "\xA3\xA9\x1E\x00\x44\xCC\x00\x80\x48\xCC\x00\x80\xA6\x59\x1E\x00\x4C\xCC"
-     "\x00\x80"
-     "\x50\xCC\x00\x80\xA5\x59\x1E\x00\xAA\x75\x1E\x00\xAB\x55\x1E\x00\x54\xCC"
-     "\x00\x80"
-     "\x58\xCC\x00\x80\xAE\x4D\x1E\x00\xAF\x31\x1E\x00\xAC\x4D\x1E\x00\xAD\x45"
-     "\x1E\x00"
-     "\xA8\x99\x02\x00\xA9\xD5\x02\x00\xAA\xD5\x02\x00\xAB\xE5\x02\x00\xAC\xFD"
-     "\x02\x00"
-     "\xAD\x35\x03\x00\xAE\x3D\x03\x00\xAF\x35\x03\x00\xBE\x6C\x02\x00\x60\xCC"
-     "\x00\x80"
-     "\x64\xCC\x00\x80\x68\xCC\x00\x80\x6C\xCC\x00\x80\x70\xCC\x00\x80\x74\xCC"
-     "\x00\x80"
-     "\x78\xCC\x00\x80\xB8\xD1\x03\x00\xB9\xD9\x03\x00\xBA\xE1\x03\x00\xBB\xE1"
-     "\x03\x00"
-     "\xBC\x91\x03\x00\xBD\x9D\x03\x00\xBE\x95\x03\x00\xBF\x89\x03\x00\xB0\x4D"
-     "\x03\x00"
-     "\xB1\x55\x03\x00\xB2\x5D\x03\x00\xB3\x55\x03\x00\xB4\x4D\x03\x00\xB5\xF1"
-     "\x03\x00"
-     "\xB6\xF1\x03\x00\xB7\xF1\x03\x00\x81\x69\x03\x00\x80\x5D\x03\x00\xB3\x55"
-     "\x02\x00"
-     "\x82\x61\x03\x00\xB5\x7D\x02\x00\x7C\xCC\x00\x80\x80\xCC\x00\x80\xB6\xBD"
-     "\x03\x00"
-     "\x84\xE4\x02\x00\x84\xCC\x00\x80\xBB\x99\x03\x00\xBA\x81\x03\x00\xBD\x89"
-     "\x03\x00"
-     "\xBC\x81\x03\x00\xBF\x3D\x03\x00\xBE\x81\x03\x00\xA3\x1D\x02\x00\x88\xCC"
-     "\x00\x80"
-     "\x87\x98\x03\x00\x86\x6C\x04\x00\x8C\xCC\x00\x80\xA6\xF5\x03\x00\xA5\x35"
-     "\x02\x00"
-     "\x90\xCC\x00\x80\xAB\xD1\x03\x00\xAA\xC9\x03\x00\x94\xCC\x00\x80\x98\xCC"
-     "\x00\x80"
-     "\xAF\x75\x03\x00\xAE\xC9\x03\x00\xAD\xC1\x03\x00\xAC\xC9\x03\x00\xE3\xD8"
-     "\x1E\x00"
-     "\xE1\xAC\x1F\x00\xE1\x58\x1E\x00\xE3\x58\x1E\x00\xE1\xA0\x01\x00\x9C\xCC"
-     "\x00\x80"
-     "\xE3\xA8\x03\x00\xA0\xCC\x00\x80\xA4\xCC\x00\x80\xA8\xCC\x00\x80\xEF\xF0"
-     "\x03\x00"
-     "\xAC\xCC\x00\x80\xEF\x04\x1E\x00\xB0\xCC\x00\x80\xB4\xCC\x00\x80\xEF\xAC"
-     "\x1E\x00"
-     "\xB8\xCC\x00\x80\xBC\xCC\x00\x80\x84\x1C\x05\x00\xBE\x00\x05\x00\x80\x0D"
-     "\x00\x00"
-     "\x81\x35\x00\x00\x82\x3D\x00\x00\xC0\xCC\x00\x80\xC4\xCC\x00\x80\xC8\xCC"
-     "\x00\x80"
-     "\x86\xC8\x04\x00\x87\xA0\x05\x00\xD0\xCC\x00\x80\xD4\xCC\x00\x80\xD8\xCC"
-     "\x00\x80"
-     "\xDC\xCC\x00\x80\xBA\xD5\x1D\x00\xBB\xDD\x1D\x00\xB8\xC1\x1D\x00\xB9\xC1"
-     "\x1D\x00"
-     "\xBE\x11\x1D\x00\xBF\x11\x1D\x00\xBC\xF1\x1D\x00\xBD\xF1\x1D\x00\xB2\xAD"
-     "\x1D\x00"
-     "\xB3\xB5\x1D\x00\xB0\xAD\x1D\x00\xB1\xA5\x1D\x00\xB6\x9D\x1D\x00\xB7\xF5"
-     "\x1D\x00"
-     "\xB4\xA5\x1D\x00\xB5\x95\x1D\x00\xAA\x69\x1E\x00\xAB\x69\x1E\x00\xA8\x69"
-     "\x1E\x00"
-     "\xA9\x69\x1E\x00\xAE\x69\x1E\x00\xAF\x69\x1E\x00\xAC\x69\x1E\x00\xAD\x69"
-     "\x1E\x00"
-     "\xE0\xCC\x00\x80\xE4\xCC\x00\x80\xE8\xCC\x00\x80\xEC\xCC\x00\x80\xF0\xCC"
-     "\x00\x80"
-     "\xF4\xCC\x00\x80\xF8\xCC\x00\x80\xFC\xCC\x00\x80\xA8\x25\x06\x00\xA9\x2D"
-     "\x06\x00"
-     "\xAA\x25\x06\x00\xAB\x3D\x06\x00\xAC\x25\x06\x00\xAD\x2D\x06\x00\xAE\x25"
-     "\x06\x00"
-     "\xAF\x89\x06\x00\xCC\xCC\x00\x80\x00\xCD\x00\x80\x04\xCD\x00\x80\x08\xCD"
-     "\x00\x80"
-     "\x0C\xCD\x00\x80\x10\xCD\x00\x80\x14\xCD\x00\x80\x18\xCD\x00\x80\xB8\x5D"
-     "\x07\x00"
-     "\xB9\x65\x07\x00\xBA\x6D\x07\x00\xBB\x61\x07\x00\xBC\x65\x07\x00\xBD\x6D"
-     "\x07\x00"
-     "\xBE\x19\x07\x00\xBF\x19\x07\x00\xB0\xF9\x06\x00\xB1\xC1\x06\x00\xB2\xAD"
-     "\x06\x00"
-     "\xB3\xBD\x06\x00\xB4\xAD\x06\x00\xB5\x9D\x06\x00\xB6\x95\x06\x00\xB7\x69"
-     "\x07\x00"
-     "\xB5\xB5\x06\x00\x1C\xCD\x00\x80\x20\xCD\x00\x80\xB6\xB5\x06\x00\x24\xCD"
-     "\x00\x80"
-     "\x28\xCD\x00\x80\xB3\xB5\x06\x00\x2C\xCD\x00\x80\xBD\x71\x06\x00\xBC\x61"
-     "\x06\x00"
-     "\xBF\x41\x06\x00\xBE\x71\x06\x00\x5C\xCC\x00\x80\x30\xCD\x00\x80\xBB\x1D"
-     "\x06\x00"
-     "\xBA\x25\x06\x00\xA3\xED\x06\x00\x82\x19\x00\x00\x81\x19\x00\x00\x80\x79"
-     "\x00\x00"
-     "\x34\xCD\x00\x80\xA6\xED\x06\x00\xA5\xED\x06\x00\x38\xCD\x00\x80\xAB\x45"
-     "\x06\x00"
-     "\xAA\x7D\x06\x00\x84\xA4\x03\x00\xBE\xA4\x03\x00\xAF\x19\x06\x00\xAE\x29"
-     "\x06\x00"
-     "\xAD\x29\x06\x00\xAC\x39\x06\x00\xA8\x71\x06\x00\xA9\x71\x06\x00\xAA\x71"
-     "\x06\x00"
-     "\xAB\x71\x06\x00\xAC\xF1\x06\x00\xAD\xF1\x06\x00\xAE\xF1\x06\x00\xAF\xF1"
-     "\x06\x00"
-     "\x3C\xCD\x00\x80\x40\xCD\x00\x80\x86\x88\x01\x00\x87\xAC\x01\x00\x44\xCD"
-     "\x00\x80"
-     "\x48\xCD\x00\x80\x4C\xCD\x00\x80\x50\xCD\x00\x80\xB8\x75\x01\x00\xB9\x7D"
-     "\x01\x00"
-     "\xBA\x75\x01\x00\xBB\xC9\x01\x00\xBC\xD9\x01\x00\xBD\xC5\x01\x00\xBE\xCD"
-     "\x01\x00"
-     "\xBF\xC1\x01\x00\xB0\x91\x06\x00\xB1\x91\x06\x00\xB2\x91\x06\x00\xB3\x91"
-     "\x06\x00"
-     "\xB4\x55\x01\x00\xB5\x5D\x01\x00\xB6\x55\x01\x00\xB7\x4D\x01\x00\xB3\xF9"
-     "\x06\x00"
-     "\x54\xCD\x00\x80\x58\xCD\x00\x80\x5C\xCD\x00\x80\x60\xCD\x00\x80\xB6\x71"
-     "\x06\x00"
-     "\xB5\x79\x06\x00\x64\xCD\x00\x80\xBB\x55\x06\x00\xBA\x55\x06\x00\x68\xCD"
-     "\x00\x80"
-     "\x6C\xCD\x00\x80\xBF\x99\x01\x00\xBE\x91\x01\x00\xBD\x45\x06\x00\xBC\x45"
-     "\x06\x00"
-     "\x70\xCD\x00\x80\xA3\xBD\x06\x00\x74\xCD\x00\x80\x78\xCD\x00\x80\xA6\x35"
-     "\x06\x00"
-     "\x7C\xCD\x00\x80\x80\xCD\x00\x80\xA5\x3D\x06\x00\xAA\x11\x06\x00\xAB\x11"
-     "\x06\x00"
-     "\x84\xCD\x00\x80\x88\xCD\x00\x80\xAE\xD5\x01\x00\xAF\xDD\x01\x00\xAC\x01"
-     "\x06\x00"
-     "\xAD\x01\x06\x00\xA8\x25\x03\x00\xA9\x35\x03\x00\xAA\x3D\x03\x00\xAB\x35"
-     "\x03\x00"
-     "\xAC\x2D\x03\x00\xAD\x91\x03\x00\xAE\x91\x03\x00\xAF\x91\x03\x00\x8C\xCD"
-     "\x00\x80"
-     "\x90\xCD\x00\x80\x94\xCD\x00\x80\x98\xCD\x00\x80\x9C\xCD\x00\x80\xA0\xCD"
-     "\x00\x80"
-     "\xA4\xCD\x00\x80\xA8\xCD\x00\x80\xB8\x99\x03\x00\xB9\xAD\x03\x00\xBA\xA5"
-     "\x03\x00"
-     "\xBB\x6D\x00\x00\xBC\x75\x00\x00\xBD\x7D\x00\x00\xBE\x75\x00\x00\xBF\x6D"
-     "\x00\x00"
-     "\xB0\xF1\x03\x00\xB1\xF9\x03\x00\xB2\xC1\x03\x00\xB3\xC1\x03\x00\xB4\xB1"
-     "\x03\x00"
-     "\xB5\xBD\x03\x00\xB6\xB5\x03\x00\xB7\xA9\x03\x00\x80\xB5\x00\x00\x81\xBD"
-     "\x00\x00"
-     "\x82\xB5\x00\x00\xBE\x48\x03\x00\x87\x10\x03\x00\x86\x34\x0C\x00\xB0\xCD"
-     "\x00\x80"
-     "\xEF\xE8\x02\x00\xBE\xAC\x0C\x00\xE1\x94\x01\x00\x84\xA4\x02\x00\xE3\x14"
-     "\x02\x00"
-     "\xB4\xCD\x00\x80\xB8\xCD\x00\x80\xBC\xCD\x00\x80\xC0\xCD\x00\x80\xB3\xF5"
-     "\x03\x00"
-     "\xC4\xCD\x00\x80\xC8\xCD\x00\x80\x85\x74\x0A\x00\xCC\xCD\x00\x80\xB6\x51"
-     "\x03\x00"
-     "\xB5\xE5\x03\x00\xD0\xCD\x00\x80\xBB\x7D\x03\x00\xBA\x75\x03\x00\xD4\xCD"
-     "\x00\x80"
-     "\xD8\xCD\x00\x80\xBF\x39\x03\x00\xBE\x35\x03\x00\xBD\x55\x03\x00\xBC\x55"
-     "\x03\x00"
-     "\xDC\xCD\x00\x80\xE0\xCD\x00\x80\xE4\xCD\x00\x80\xEF\xB8\x06\x00\xE8\xCD"
-     "\x00\x80"
-     "\xEC\xCD\x00\x80\xF0\xCD\x00\x80\xF4\xCD\x00\x80\xF8\xCD\x00\x80\xFC\xCD"
-     "\x00\x80"
-     "\x00\xCE\x00\x80\x04\xCE\x00\x80\xCC\x00\x00\x00\xE1\xE0\x07\x00\x08\xCE"
-     "\x00\x80"
-     "\xE3\x64\x06\x00\x0C\xCE\x00\x80\x10\xCE\x00\x80\xA3\x7D\x02\x00\x84\x40"
-     "\x02\x00"
-     "\xA5\x6D\x02\x00\x14\xCE\x00\x80\x18\xCE\x00\x80\xA6\xD9\x02\x00\x86\xC0"
-     "\x0C\x00"
-     "\x87\xAC\x0D\x00\xAB\xF5\x02\x00\xAA\xFD\x02\x00\xAD\xDD\x02\x00\xAC\xDD"
-     "\x02\x00"
-     "\xAF\xB1\x02\x00\xAE\xBD\x02\x00\x80\xC5\x00\x00\x81\xD1\x00\x00\x82\xD1"
-     "\x00\x00"
-     "\x20\xCE\x00\x80\x24\xCE\x00\x80\x28\xCE\x00\x80\x2C\xCE\x00\x80\xEF\x88"
-     "\x01\x00"
-     "\x84\xEC\x0C\x00\xE1\xD4\x06\x00\x30\xCE\x00\x80\xE3\x54\x01\x00\x34\xCE"
-     "\x00\x80"
-     "\x38\xCE\x00\x80\x3C\xCE\x00\x80\x40\xCE\x00\x80\xB3\x75\x01\x00\x44\xCE"
-     "\x00\x80"
-     "\x48\xCE\x00\x80\x4C\xCE\x00\x80\x50\xCE\x00\x80\xB6\x11\x01\x00\xB5\x65"
-     "\x01\x00"
-     "\x54\xCE\x00\x80\xBB\x09\x01\x00\xBA\x39\x01\x00\x58\xCE\x00\x80\x5C\xCE"
-     "\x00\x80"
-     "\xBF\xF9\x00\x00\xBE\xF1\x00\x00\xBD\xF9\x00\x00\xBC\x11\x01\x00\xA8\x99"
-     "\x0E\x00"
-     "\xA9\xB5\x0E\x00\xAA\xBD\x0E\x00\xAB\xB5\x0E\x00\xAC\xAD\x0E\x00\xAD\xD1"
-     "\x0E\x00"
-     "\xAE\xD1\x0E\x00\xAF\xCD\x0E\x00\x1C\xCE\x00\x80\x60\xCE\x00\x80\x64\xCE"
-     "\x00\x80"
-     "\x68\xCE\x00\x80\x6C\xCE\x00\x80\x70\xCE\x00\x80\x74\xCE\x00\x80\x78\xCE"
-     "\x00\x80"
-     "\xB8\x59\x0F\x00\xB9\x59\x0F\x00\xBA\x69\x0F\x00\xBB\x69\x0F\x00\xBC\x79"
-     "\x0F\x00"
-     "\xBD\x79\x0F\x00\xBE\x61\x0F\x00\xBF\x1D\x0F\x00\xB0\xBD\x0E\x00\xB1\x8D"
-     "\x0E\x00"
-     "\xB2\x85\x0E\x00\xB3\x6D\x0F\x00\xB4\x75\x0F\x00\xB5\x7D\x0F\x00\xB6\x69"
-     "\x0F\x00"
-     "\xB7\x69\x0F\x00\xA3\x35\x0E\x00\x7C\xCE\x00\x80\x80\xCE\x00\x80\x84\xCE"
-     "\x00\x80"
-     "\x88\xCE\x00\x80\xA6\x51\x0E\x00\xA5\x25\x0E\x00\xBE\xD4\x03\x00\xAB\x49"
-     "\x0E\x00"
-     "\xAA\x79\x0E\x00\x8C\xCE\x00\x80\x90\xCE\x00\x80\xAF\xB9\x0F\x00\xAE\xB1"
-     "\x0F\x00"
-     "\xAD\xB9\x0F\x00\xAC\x51\x0E\x00\x82\x51\x00\x00\xAC\xCD\x00\x80\x80\x41"
-     "\x00\x00"
-     "\x81\x41\x00\x00\x87\x18\x03\x00\x86\x74\x00\x00\x94\xCE\x00\x80\x98\xCE"
-     "\x00\x80"
-     "\x9C\xCE\x00\x80\xA0\xCE\x00\x80\xA4\xCE\x00\x80\xA8\xCE\x00\x80\xAC\xCE"
-     "\x00\x80"
-     "\xB0\xCE\x00\x80\xB4\xCE\x00\x80\xB8\xCE\x00\x80\xA8\x8D\x0E\x00\xA9\x95"
-     "\x0E\x00"
-     "\xAA\x9D\x0E\x00\xAB\x95\x0E\x00\xAC\xB9\x0E\x00\xAD\xA5\x0E\x00\xAE\xAD"
-     "\x0E\x00"
-     "\xAF\xA5\x0E\x00\xBC\xCE\x00\x80\xC0\xCE\x00\x80\xC4\xCE\x00\x80\xC8\xCE"
-     "\x00\x80"
-     "\xCC\xCE\x00\x80\xD0\xCE\x00\x80\xD4\xCE\x00\x80\xD8\xCE\x00\x80\xB8\x65"
-     "\x01\x00"
-     "\xB9\x6D\x01\x00\xBA\x65\x01\x00\xBB\x7D\x01\x00\xBC\x65\x01\x00\xBD\x6D"
-     "\x01\x00"
-     "\xBE\x65\x01\x00\xBF\xD9\x01\x00\xB0\xDD\x0E\x00\xB1\xA1\x0E\x00\xB2\xA1"
-     "\x0E\x00"
-     "\xB3\xA1\x0E\x00\xB4\xA1\x0E\x00\xB5\xA9\x0E\x00\xB6\x91\x0E\x00\xB7\x91"
-     "\x0E\x00"
-     "\xB3\x15\x0E\x00\xDC\xCE\x00\x80\xE0\xCE\x00\x80\xE4\xCE\x00\x80\xE8\xCE"
-     "\x00\x80"
-     "\xB6\x25\x0E\x00\xB5\x31\x0E\x00\xEC\xCE\x00\x80\xBB\x65\x0E\x00\xBA\x19"
-     "\x0E\x00"
-     "\xF0\xCE\x00\x80\xF4\xCE\x00\x80\xBF\x6D\x0E\x00\xBE\x65\x0E\x00\xBD\x75"
-     "\x0E\x00"
-     "\xBC\x75\x0E\x00\x82\x15\x00\x00\xA3\x51\x0E\x00\x80\x61\x00\x00\x81\x61"
-     "\x00\x00"
-     "\xA6\x61\x0E\x00\xF8\xCE\x00\x80\xBE\x90\x01\x00\xA5\x75\x0E\x00\xAA\x5D"
-     "\x0E\x00"
-     "\xAB\x21\x0E\x00\x84\xA4\x01\x00\x00\xCF\x00\x80\xAE\x21\x0E\x00\xAF\x29"
-     "\x0E\x00"
-     "\xAC\x31\x0E\x00\xAD\x31\x0E\x00\xA8\xAD\x02\x00\xA9\xDD\x02\x00\xAA\xD5"
-     "\x02\x00"
-     "\xAB\xED\x02\x00\xAC\xF5\x02\x00\xAD\xFD\x02\x00\xAE\xF5\x02\x00\xAF\xED"
-     "\x02\x00"
-     "\x04\xCF\x00\x80\x08\xCF\x00\x80\x0C\xCF\x00\x80\x10\xCF\x00\x80\x14\xCF"
-     "\x00\x80"
-     "\x18\xCF\x00\x80\x1C\xCF\x00\x80\x20\xCF\x00\x80\xB8\x75\x03\x00\xB9\x7D"
-     "\x03\x00"
-     "\xBA\x75\x03\x00\xBB\xC9\x03\x00\xBC\xD9\x03\x00\xBD\xD9\x03\x00\xBE\xC9"
-     "\x03\x00"
-     "\xBF\xC1\x03\x00\xB0\x95\x02\x00\xB1\x9D\x02\x00\xB2\x95\x02\x00\xB3\x4D"
-     "\x03\x00"
-     "\xB4\x55\x03\x00\xB5\x5D\x03\x00\xB6\x55\x03\x00\xB7\x4D\x03\x00\xB3\x5D"
-     "\x02\x00"
-     "\x24\xCF\x00\x80\x87\xD8\x03\x00\x86\x2C\x04\x00\x28\xCF\x00\x80\xB6\x7D"
-     "\x02\x00"
-     "\xB5\x7D\x02\x00\x2C\xCF\x00\x80\xBB\x59\x02\x00\xBA\x51\x02\x00\x30\xCF"
-     "\x00\x80"
-     "\x34\xCF\x00\x80\xBF\x99\x03\x00\xBE\x91\x03\x00\xBD\x99\x03\x00\xBC\x41"
-     "\x02\x00"
-     "\x38\xCF\x00\x80\xA3\x19\x02\x00\x86\x68\x05\x00\x87\x00\x03\x00\xA6\x39"
-     "\x02\x00"
-     "\x3C\xCF\x00\x80\x40\xCF\x00\x80\xA5\x39\x02\x00\xAA\x15\x02\x00\xAB\x1D"
-     "\x02\x00"
-     "\x44\xCF\x00\x80\x48\xCF\x00\x80\xAE\xD5\x03\x00\xAF\xDD\x03\x00\xAC\x05"
-     "\x02\x00"
-     "\xAD\xDD\x03\x00\x4C\xCF\x00\x80\x50\xCF\x00\x80\x54\xCF\x00\x80\x58\xCF"
-     "\x00\x80"
-     "\x5C\xCF\x00\x80\xEF\x34\x00\x00\x60\xCF\x00\x80\x64\xCF\x00\x80\xBE\x2C"
-     "\x05\x00"
-     "\xE3\xDC\x00\x00\x68\xCF\x00\x80\xE1\xE0\x01\x00\x70\xCF\x00\x80\x74\xCF"
-     "\x00\x80"
-     "\x78\xCF\x00\x80\x7C\xCF\x00\x80\xE1\xCC\x0F\x00\x84\x48\x02\x00\xE3\x14"
-     "\x0E\x00"
-     "\x84\x44\x05\x00\x80\x49\x00\x00\x81\x49\x00\x00\x82\x71\x00\x00\x80\xCF"
-     "\x00\x80"
-     "\x84\xCF\x00\x80\x88\xCF\x00\x80\x8C\xCF\x00\x80\x90\xCF\x00\x80\x94\xCF"
-     "\x00\x80"
-     "\x98\xCF\x00\x80\xEF\x20\x0F\x00\x9C\xCF\x00\x80\xA0\xCF\x00\x80\xA4\xCF"
-     "\x00\x80"
-     "\xA8\xCF\x00\x80\xAC\xCF\x00\x80\xB0\xCF\x00\x80\xB4\xCF\x00\x80\xB8\xCF"
-     "\x00\x80"
-     "\xEF\xF8\x01\x00\xBE\xAC\x04\x00\xE1\xDC\x0E\x00\xBC\xCF\x00\x80\xE3\x1C"
-     "\x0E\x00"
-     "\xC0\xCF\x00\x80\xC4\xCF\x00\x80\xC8\xCF\x00\x80\xCC\xCF\x00\x80\xB3\xE5"
-     "\x01\x00"
-     "\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D\x00\x00\xD0\xCF\x00\x80\xB6\xE5"
-     "\x01\x00"
-     "\xB5\xF5\x01\x00\xD4\xCF\x00\x80\xBB\x65\x01\x00\xBA\xD9\x01\x00\xD8\xCF"
-     "\x00\x80"
-     "\xE0\xCF\x00\x80\xBF\x5D\x01\x00\xBE\x5D\x01\x00\xBD\x65\x01\x00\xBC\x6D"
-     "\x01\x00"
-     "\xA8\xED\x05\x00\xA9\x25\x06\x00\xAA\x2D\x06\x00\xAB\x25\x06\x00\xAC\x3D"
-     "\x06\x00"
-     "\xAD\x25\x06\x00\xAE\x2D\x06\x00\xAF\x25\x06\x00\x6C\xCF\x00\x80\xDC\xCF"
-     "\x00\x80"
-     "\x86\x88\x00\x00\x87\xA0\x03\x00\xE4\xCF\x00\x80\xE8\xCF\x00\x80\xEC\xCF"
-     "\x00\x80"
-     "\xF0\xCF\x00\x80\xB8\x99\x06\x00\xB9\xAD\x06\x00\xBA\xA5\x06\x00\xBB\x4D"
-     "\x07\x00"
-     "\xBC\x55\x07\x00\xBD\x5D\x07\x00\xBE\x55\x07\x00\xBF\x4D\x07\x00\xB0\x5D"
-     "\x06\x00"
-     "\xB1\xE1\x06\x00\xB2\xF9\x06\x00\xB3\xF1\x06\x00\xB4\xD9\x06\x00\xB5\xD9"
-     "\x06\x00"
-     "\xB6\xB9\x06\x00\xB7\xA9\x06\x00\xA3\xA5\x06\x00\xF4\xCF\x00\x80\xF8\xCF"
-     "\x00\x80"
-     "\xFC\xCF\x00\x80\x00\xD0\x00\x80\xA6\xA5\x06\x00\xA5\xB5\x06\x00\x04\xD0"
-     "\x00\x80"
-     "\xAB\x25\x06\x00\xAA\x99\x06\x00\x08\xD0\x00\x80\x0C\xD0\x00\x80\xAF\x1D"
-     "\x06\x00"
-     "\xAE\x1D\x06\x00\xAD\x25\x06\x00\xAC\x2D\x06\x00\x10\xD0\x00\x80\xB3\xFD"
-     "\x07\x00"
-     "\xFC\xCE\x00\x80\x14\xD0\x00\x80\xB6\xB1\x07\x00\x18\xD0\x00\x80\x1C\xD0"
-     "\x00\x80"
-     "\xB5\x81\x07\x00\xBA\x8D\x07\x00\xBB\x95\x07\x00\x20\xD0\x00\x80\x24\xD0"
-     "\x00\x80"
-     "\xBE\x71\x07\x00\xBF\x71\x07\x00\xBC\x8D\x07\x00\xBD\x81\x07\x00\xA8\x31"
-     "\x06\x00"
-     "\xA9\x31\x06\x00\xAA\x31\x06\x00\xAB\x31\x06\x00\xAC\x4D\x06\x00\xAD\x75"
-     "\x06\x00"
-     "\xAE\x7D\x06\x00\xAF\x75\x06\x00\x28\xD0\x00\x80\x2C\xD0\x00\x80\x30\xD0"
-     "\x00\x80"
-     "\x34\xD0\x00\x80\x38\xD0\x00\x80\x3C\xD0\x00\x80\x40\xD0\x00\x80\x44\xD0"
-     "\x00\x80"
-     "\xB8\xD1\x06\x00\xB9\xD1\x06\x00\xBA\xD1\x06\x00\xBB\xE5\x06\x00\xBC\xE1"
-     "\x06\x00"
-     "\xBD\xE1\x06\x00\xBE\xE1\x06\x00\xBF\xE1\x06\x00\xB0\x0D\x06\x00\xB1\x15"
-     "\x06\x00"
-     "\xB2\x1D\x06\x00\xB3\x11\x06\x00\xB4\xF1\x06\x00\xB5\xF1\x06\x00\xB6\xF1"
-     "\x06\x00"
-     "\xB7\xF1\x06\x00\xA3\xB9\x06\x00\x82\x15\x00\x00\x81\xB5\x00\x00\x80\xB5"
-     "\x00\x00"
-     "\x48\xD0\x00\x80\xA6\xF5\x06\x00\xA5\xC5\x06\x00\xBE\x90\x01\x00\xAB\xD1"
-     "\x06\x00"
-     "\xAA\xC9\x06\x00\x84\xA0\x02\x00\xBF\x00\x3C\x00\xAF\x35\x06\x00\xAE\x35"
-     "\x06\x00"
-     "\xAD\xC5\x06\x00\xAC\xC9\x06\x00\x85\x84\x00\x00\xB3\xF1\x01\x00\x86\x08"
-     "\x3F\x00"
-     "\x87\x9C\x00\x00\xB6\x55\x01\x00\x50\xD0\x00\x80\x54\xD0\x00\x80\xB5\xE1"
-     "\x01\x00"
-     "\xBA\x71\x01\x00\xBB\x79\x01\x00\x58\xD0\x00\x80\x5C\xD0\x00\x80\xBE\x31"
-     "\x01\x00"
-     "\xBF\x39\x01\x00\xBC\x51\x01\x00\xBD\x51\x01\x00\xA8\xED\x05\x00\xA9\x35"
-     "\x02\x00"
-     "\xAA\x3D\x02\x00\xAB\x35\x02\x00\xAC\x2D\x02\x00\xAD\x91\x02\x00\xAE\x91"
-     "\x02\x00"
-     "\xAF\x91\x02\x00\x60\xD0\x00\x80\x64\xD0\x00\x80\x68\xD0\x00\x80\x6C\xD0"
-     "\x00\x80"
-     "\xA2\x11\x02\x00\xA3\x11\x02\x00\xA0\x91\x06\x00\xA1\xF5\x02\x00\xB8\xA9"
-     "\x02\x00"
-     "\xB9\xA9\x02\x00\xBA\x69\x03\x00\xBB\x69\x03\x00\xBC\x79\x03\x00\xBD\x79"
-     "\x03\x00"
-     "\xBE\x69\x03\x00\xBF\x69\x03\x00\xB0\xF1\x02\x00\xB1\xF9\x02\x00\xB2\xC1"
-     "\x02\x00"
-     "\xB3\xC1\x02\x00\xB4\xB1\x02\x00\xB5\xB9\x02\x00\xB6\xAD\x02\x00\xB7\xA5"
-     "\x02\x00"
-     "\xA3\xB5\x02\x00\x8F\x8D\x1A\x00\x70\xD0\x00\x80\x74\xD0\x00\x80\x78\xD0"
-     "\x00\x80"
-     "\xA6\x11\x02\x00\xA5\xA5\x02\x00\x7C\xD0\x00\x80\xAB\x3D\x02\x00\xAA\x35"
-     "\x02\x00"
-     "\x80\xD0\x00\x80\x84\xD0\x00\x80\xAF\x7D\x02\x00\xAE\x75\x02\x00\xAD\x15"
-     "\x02\x00"
-     "\xAC\x15\x02\x00\x9E\x05\x06\x00\x9F\xE5\x06\x00\x9C\xE9\x0A\x00\x9D\xED"
-     "\x0B\x00"
-     "\x9A\x45\x0E\x00\x9B\x65\x0A\x00\x98\x59\x0E\x00\x99\x51\x0E\x00\x96\xC5"
-     "\x12\x00"
-     "\x97\xBD\x13\x00\x94\x75\x12\x00\x95\xCD\x12\x00\x92\xD1\x16\x00\x93\x71"
-     "\x16\x00"
-     "\x90\xF5\x1A\x00\x91\x75\x16\x00\x88\xD0\x00\x80\xE1\xD0\x00\x00\x84\x44"
-     "\x3D\x00"
-     "\xE3\x10\x24\x00\x8C\xD0\x00\x80\x90\xD0\x00\x80\x94\xD0\x00\x80\x98\xD0"
-     "\x00\x80"
-     "\x9C\xD0\x00\x80\xA0\xD0\x00\x80\xA4\xD0\x00\x80\xA8\xD0\x00\x80\xAC\xD0"
-     "\x00\x80"
-     "\xB0\xD0\x00\x80\xB4\xD0\x00\x80\xEF\xCC\x24\x00\xB8\xD0\x00\x80\xE3\x10"
-     "\x06\x00"
-     "\xBE\x00\x3D\x00\xE1\xD8\x07\x00\x80\x65\x00\x00\x81\x6D\x00\x00\x82\x19"
-     "\x00\x00"
-     "\xC0\xD0\x00\x80\x86\xC0\x3C\x00\x87\x44\x3C\x00\xC4\xD0\x00\x80\xC8\xD0"
-     "\x00\x80"
-     "\xCC\xD0\x00\x80\xEF\xAC\x06\x00\xD0\xD0\x00\x80\xD4\xD0\x00\x80\xB1\x85"
-     "\x2B\x00"
-     "\xB0\x01\x28\x00\xB3\xA1\x2A\x00\xB2\x9D\x2B\x00\xB5\x91\x27\x00\xB4\x01"
-     "\x24\x00"
-     "\xEF\xFC\x00\x00\xEF\xC4\x06\x00\xE1\x90\x01\x00\xE1\x74\x07\x00\xE3\x10"
-     "\x00\x00"
-     "\xE3\xF0\x06\x00\xD8\xD0\x00\x80\xDC\xD0\x00\x80\x84\xF4\x3C\x00\xE0\xD0"
-     "\x00\x80"
-     "\xA1\x95\x3F\x00\xE4\xD0\x00\x80\xA3\x21\x3F\x00\xA2\x49\x3F\x00\xA5\xD5"
-     "\x3B\x00"
-     "\xA4\xD5\x3B\x00\xA7\x01\x34\x00\xA6\x01\x3B\x00\xA9\x41\x37\x00\xA8\xE1"
-     "\x37\x00"
-     "\xAB\xE9\x33\x00\xAA\x01\x30\x00\xAD\x01\x2C\x00\xAC\xC1\x33\x00\xAF\xA1"
-     "\x2F\x00"
-     "\xAE\xF9\x2F\x00\xB0\x89\x3E\x00\xB1\x89\x3E\x00\xB2\x99\x3E\x00\xB3\x91"
-     "\x3E\x00"
-     "\xB4\xB9\x3E\x00\xB5\xB9\x3E\x00\xB6\x19\x3F\x00\xB7\x0D\x3F\x00\xB8\x35"
-     "\x3F\x00"
-     "\xB9\x3D\x3F\x00\xBA\x35\x3F\x00\xBB\x0D\x3F\x00\xBC\x15\x3F\x00\xBD\x1D"
-     "\x3F\x00"
-     "\xBE\x15\x3F\x00\xBF\x0D\x3F\x00\xE8\xD0\x00\x80\xEC\xD0\x00\x80\xF0\xD0"
-     "\x00\x80"
-     "\xF4\xD0\x00\x80\xF8\xD0\x00\x80\x4C\xD0\x00\x80\xFC\xD0\x00\x80\x00\xD1"
-     "\x00\x80"
-     "\xA8\xB5\x3D\x00\xA9\xBD\x3D\x00\xAA\x81\x3D\x00\xAB\x81\x3D\x00\xAC\x51"
-     "\x3E\x00"
-     "\xAD\x51\x3E\x00\xAE\x51\x3E\x00\xAF\x51\x3E\x00\xB3\x69\x3D\x00\xBC\xD0"
-     "\x00\x80"
-     "\x04\xD1\x00\x80\x08\xD1\x00\x80\x0C\xD1\x00\x80\xB6\xB1\x3E\x00\xB5\xB9"
-     "\x3E\x00"
-     "\x10\xD1\x00\x80\xBB\x2D\x3E\x00\xBA\x2D\x3E\x00\x14\xD1\x00\x80\x18\xD1"
-     "\x00\x80"
-     "\xBF\x1D\x3E\x00\xBE\x1D\x3E\x00\xBD\x25\x3E\x00\xBC\x2D\x3E\x00\x1C\xD1"
-     "\x00\x80"
-     "\xA3\x2D\x3D\x00\x20\xD1\x00\x80\x24\xD1\x00\x80\xA6\xF5\x3E\x00\x28\xD1"
-     "\x00\x80"
-     "\xBE\x90\x03\x00\xA5\xFD\x3E\x00\xAA\x69\x3E\x00\xAB\x69\x3E\x00\x2C\xD1"
-     "\x00\x80"
-     "\x30\xD1\x00\x80\xAE\x59\x3E\x00\xAF\x59\x3E\x00\xAC\x69\x3E\x00\xAD\x61"
-     "\x3E\x00"
-     "\xA8\x5D\x3E\x00\xA9\x6D\x3E\x00\xAA\x65\x3E\x00\xAB\x7D\x3E\x00\xAC\x69"
-     "\x3E\x00"
-     "\xAD\x69\x3E\x00\xAE\xA9\x3E\x00\xAF\xA9\x3E\x00\x80\xE9\x01\x00\x81\xE9"
-     "\x01\x00"
-     "\x82\xFD\x01\x00\x34\xD1\x00\x80\x86\x90\x01\x00\x87\xB4\x01\x00\x38\xD1"
-     "\x00\x80"
-     "\x3C\xD1\x00\x80\xB8\x1D\x01\x00\xB9\xCD\x01\x00\xBA\xC5\x01\x00\xBB\xDD"
-     "\x01\x00"
-     "\xBC\xCD\x01\x00\xBD\xFD\x01\x00\xBE\xF5\x01\x00\xBF\x9D\x01\x00\xB0\xDD"
-     "\x3E\x00"
-     "\xB1\xA1\x3E\x00\xB2\x29\x01\x00\xB3\x29\x01\x00\xB4\x39\x01\x00\xB5\x39"
-     "\x01\x00"
-     "\xB6\x29\x01\x00\xB7\x25\x01\x00\xB3\xED\x3E\x00\x40\xD1\x00\x80\x44\xD1"
-     "\x00\x80"
-     "\x48\xD1\x00\x80\x4C\xD1\x00\x80\xB6\xE9\x3E\x00\xB5\xE1\x3E\x00\x50\xD1"
-     "\x00\x80"
-     "\xBB\x65\x3E\x00\xBA\x19\x3E\x00\x54\xD1\x00\x80\x84\xA0\x01\x00\xBF\xFD"
-     "\x01\x00"
-     "\xBE\xFD\x01\x00\xBD\xFD\x01\x00\xBC\xFD\x01\x00\x58\xD1\x00\x80\xA3\xA9"
-     "\x3E\x00"
-     "\x5C\xD1\x00\x80\x60\xD1\x00\x80\xA6\xAD\x3E\x00\x64\xD1\x00\x80\x68\xD1"
-     "\x00\x80"
-     "\xA5\xA5\x3E\x00\xAA\x5D\x3E\x00\xAB\x21\x3E\x00\x6C\xD1\x00\x80\x70\xD1"
-     "\x00\x80"
-     "\xAE\xB9\x01\x00\xAF\xB9\x01\x00\xAC\xB9\x01\x00\xAD\xB9\x01\x00\xA8\xED"
-     "\x02\x00"
-     "\xA9\x29\x03\x00\xAA\x31\x03\x00\xAB\x31\x03\x00\xAC\x95\x03\x00\xAD\x9D"
-     "\x03\x00"
-     "\xAE\x95\x03\x00\xAF\x8D\x03\x00\xBE\x00\x04\x00\x74\xD1\x00\x80\x78\xD1"
-     "\x00\x80"
-     "\x7C\xD1\x00\x80\x80\xD1\x00\x80\x84\xD1\x00\x80\x88\xD1\x00\x80\x8C\xD1"
-     "\x00\x80"
-     "\xB8\xAD\x03\x00\xB9\x61\x00\x00\xBA\x61\x00\x00\xBB\x61\x00\x00\xBC\x61"
-     "\x00\x00"
-     "\xBD\x61\x00\x00\xBE\x61\x00\x00\xBF\x61\x00\x00\xB0\xFD\x03\x00\xB1\xCD"
-     "\x03\x00"
-     "\xB2\xA9\x03\x00\xB3\xA5\x03\x00\xB4\xBD\x03\x00\xB5\xA1\x03\x00\xB6\xA1"
-     "\x03\x00"
-     "\xB7\x95\x03\x00\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\x90\xD1"
-     "\x00\x80"
-     "\x94\xD1\x00\x80\x98\xD1\x00\x80\xA0\xD1\x00\x80\xA4\xD1\x00\x80\x86\x20"
-     "\x04\x00"
-     "\x87\xE4\x03\x00\xA8\xD1\x00\x80\xAC\xD1\x00\x80\xB0\xD1\x00\x80\xB4\xD1"
-     "\x00\x80"
-     "\xB8\xD1\x00\x80\xBC\xD1\x00\x80\xB3\xF9\x03\x00\xC0\xD1\x00\x80\xC4\xD1"
-     "\x00\x80"
-     "\xC8\xD1\x00\x80\xCC\xD1\x00\x80\xB6\x5D\x03\x00\xB5\x5D\x03\x00\xD0\xD1"
-     "\x00\x80"
-     "\xBB\x49\x03\x00\xBA\x79\x03\x00\xD4\xD1\x00\x80\xD8\xD1\x00\x80\xBF\x25"
-     "\x03\x00"
-     "\xBE\x39\x03\x00\xBD\x31\x03\x00\xBC\x3D\x03\x00\xE1\x78\x3E\x00\xDC\xD1"
-     "\x00\x80"
-     "\xE3\xA0\x3E\x00\xE0\xD1\x00\x80\xE4\xD1\x00\x80\xE8\xD1\x00\x80\xEC\xD1"
-     "\x00\x80"
-     "\xF0\xD1\x00\x80\x84\x2C\x05\x00\xF4\xD1\x00\x80\xBE\xA4\x04\x00\xF8\xD1"
-     "\x00\x80"
-     "\xFC\xD1\x00\x80\x00\xD2\x00\x80\xEF\x18\x01\x00\x04\xD2\x00\x80\x80\x75"
-     "\x00\x00"
-     "\x81\x79\x00\x00\x82\x19\x00\x00\x84\xC4\x04\x00\x08\xD2\x00\x80\xEF\x8C"
-     "\x03\x00"
-     "\x87\x94\x05\x00\x86\x7C\x04\x00\x10\xD2\x00\x80\xE3\xD8\x00\x00\x14\xD2"
-     "\x00\x80"
-     "\xE1\x44\x00\x00\x18\xD2\x00\x80\x1C\xD2\x00\x80\x20\xD2\x00\x80\x24\xD2"
-     "\x00\x80"
-     "\x8B\xA5\x2F\x00\x8A\xA5\x2F\x00\x28\xD2\x00\x80\x2C\xD2\x00\x80\x8F\x05"
-     "\x33\x00"
-     "\x8E\x19\x33\x00\x8D\x19\x33\x00\x8C\xC5\x2E\x00\x83\xAD\x27\x00\x82\xAD"
-     "\x27\x00"
-     "\x30\xD2\x00\x80\x34\xD2\x00\x80\x87\x2D\x2F\x00\x86\x31\x2B\x00\x85\x99"
-     "\x2B\x00"
-     "\x84\x35\x2B\x00\x9B\xE1\x3F\x00\x9A\x0D\x3F\x00\x38\xD2\x00\x80\x3C\xD2"
-     "\x00\x80"
-     "\xE1\x00\x3C\x00\x40\xD2\x00\x80\xE3\xC0\x3E\x00\x9C\xF9\x3F\x00\x93\xC1"
-     "\x37\x00"
-     "\x92\x55\x37\x00\xEF\x58\x3E\x00\x44\xD2\x00\x80\x97\x59\x3B\x00\x96\xD5"
-     "\x3B\x00"
-     "\x95\x3D\x3B\x00\x94\xDD\x36\x00\x9C\xD1\x00\x80\x0C\xD2\x00\x80\x48\xD2"
-     "\x00\x80"
-     "\x4C\xD2\x00\x80\x50\xD2\x00\x80\x54\xD2\x00\x80\x58\xD2\x00\x80\x5C\xD2"
-     "\x00\x80"
-     "\xA8\xB9\x06\x00\xA9\xB9\x06\x00\xAA\x81\x06\x00\xAB\x81\x06\x00\xAC\x81"
-     "\x06\x00"
-     "\xAD\x81\x06\x00\xAE\x81\x06\x00\xAF\x81\x06\x00\xB0\x85\x06\x00\xB1\x8D"
-     "\x06\x00"
-     "\xB2\x85\x06\x00\xB3\x9D\x06\x00\xB4\x8D\x06\x00\xB5\xBD\x06\x00\xB6\xB5"
-     "\x06\x00"
-     "\xB7\x59\x07\x00\xB8\x6D\x07\x00\xB9\x75\x07\x00\xBA\x7D\x07\x00\xBB\x75"
-     "\x07\x00"
-     "\xBC\x6D\x07\x00\xBD\x51\x07\x00\xBE\x55\x07\x00\xBF\x4D\x07\x00\x60\xD2"
-     "\x00\x80"
-     "\x64\xD2\x00\x80\xAA\x15\x05\x00\xAB\x25\x05\x00\xAC\x51\x05\x00\xAD\x5D"
-     "\x05\x00"
-     "\xAE\x55\x05\x00\xAF\x49\x05\x00\x68\xD2\x00\x80\x6C\xD2\x00\x80\x70\xD2"
-     "\x00\x80"
-     "\xA3\x95\x05\x00\x74\xD2\x00\x80\xA5\x31\x05\x00\xA6\x31\x05\x00\x78\xD2"
-     "\x00\x80"
-     "\xB3\x69\x06\x00\x82\x79\x00\x00\x81\x79\x00\x00\x80\x65\x00\x00\xBE\x5C"
-     "\x00\x00"
-     "\xB6\x61\x06\x00\xB5\x69\x06\x00\x7C\xD2\x00\x80\xBB\x29\x06\x00\xBA\x21"
-     "\x06\x00"
-     "\x86\xC8\x00\x00\x87\xAC\x00\x00\xBF\x19\x06\x00\xBE\x29\x06\x00\xBD\x29"
-     "\x06\x00"
-     "\xBC\x31\x06\x00\xA3\xA9\x07\x00\x80\xD2\x00\x80\x84\xD2\x00\x80\x88\xD2"
-     "\x00\x80"
-     "\x8C\xD2\x00\x80\xA6\xA1\x07\x00\xA5\xA9\x07\x00\x90\xD2\x00\x80\xAB\xE9"
-     "\x07\x00"
-     "\xAA\xE1\x07\x00\x94\xD2\x00\x80\x98\xD2\x00\x80\xAF\xD9\x07\x00\xAE\xE9"
-     "\x07\x00"
-     "\xAD\xE9\x07\x00\xAC\xF1\x07\x00\x9C\xD2\x00\x80\xB3\x35\x06\x00\xA0\xD2"
-     "\x00\x80"
-     "\xA4\xD2\x00\x80\xB6\x2D\x06\x00\xA8\xD2\x00\x80\xAC\xD2\x00\x80\xB5\x25"
-     "\x06\x00"
-     "\xBA\xC9\x06\x00\xBB\xC9\x06\x00\xB0\xD2\x00\x80\xB4\xD2\x00\x80\xBE\xA5"
-     "\x06\x00"
-     "\xBF\x99\x06\x00\xBC\xB1\x06\x00\xBD\xB1\x06\x00\xA8\xE1\x06\x00\xA9\xED"
-     "\x06\x00"
-     "\xAA\xF9\x06\x00\xAB\xF9\x06\x00\xAC\xE9\x06\x00\xAD\xE9\x06\x00\xAE\x11"
-     "\x06\x00"
-     "\xAF\x11\x06\x00\xB8\xD2\x00\x80\xBC\xD2\x00\x80\xC0\xD2\x00\x80\xC4\xD2"
-     "\x00\x80"
-     "\xC8\xD2\x00\x80\xCC\xD2\x00\x80\xD0\xD2\x00\x80\xD4\xD2\x00\x80\xB8\xD1"
-     "\x01\x00"
-     "\xB9\xD1\x01\x00\xBA\xD1\x01\x00\xBB\xD1\x01\x00\xBC\x51\x01\x00\xBD\x51"
-     "\x01\x00"
-     "\xBE\x79\x01\x00\xBF\x79\x01\x00\xB0\x49\x06\x00\xB1\x49\x06\x00\xB2\x21"
-     "\x06\x00"
-     "\xB3\x21\x06\x00\xB4\x25\x06\x00\xB5\x19\x06\x00\xB6\xF1\x01\x00\xB7\xF1"
-     "\x01\x00"
-     "\xA3\x71\x05\x00\xD8\xD2\x00\x80\xDC\xD2\x00\x80\xBE\xCC\x02\x00\xE0\xD2"
-     "\x00\x80"
-     "\xA6\x69\x05\x00\xA5\x61\x05\x00\xE4\xD2\x00\x80\xAB\x8D\x05\x00\xAA\x8D"
-     "\x05\x00"
-     "\x84\xE4\x03\x00\xE8\xD2\x00\x80\xAF\xDD\x05\x00\xAE\xE1\x05\x00\xAD\xF5"
-     "\x05\x00"
-     "\xAC\xF5\x05\x00\x80\x51\x00\x00\x81\x59\x00\x00\x82\x65\x00\x00\xA3\x41"
-     "\x03\x00"
-     "\xF0\xD2\x00\x80\xA5\x49\x03\x00\xA6\x71\x03\x00\xF4\xD2\x00\x80\x86\x40"
-     "\x0C\x00"
-     "\x87\x14\x03\x00\xAA\x21\x03\x00\xAB\x21\x03\x00\xAC\x25\x03\x00\xAD\x2D"
-     "\x03\x00"
-     "\xAE\x11\x03\x00\xAF\x11\x03\x00\xA8\x85\x03\x00\xA9\x8D\x03\x00\xAA\x85"
-     "\x03\x00"
-     "\xAB\x9D\x03\x00\xAC\x8D\x03\x00\xAD\xBD\x03\x00\xAE\xB5\x03\x00\xAF\xDD"
-     "\x03\x00"
-     "\xF8\xD2\x00\x80\xFC\xD2\x00\x80\x00\xD3\x00\x80\x04\xD3\x00\x80\x08\xD3"
-     "\x00\x80"
-     "\x0C\xD3\x00\x80\x10\xD3\x00\x80\x14\xD3\x00\x80\xB8\x69\x03\x00\xB9\x69"
-     "\x03\x00"
-     "\xBA\x79\x03\x00\xBB\x79\x03\x00\xBC\x69\x03\x00\xBD\x69\x03\x00\xBE\xDD"
-     "\x00\x00"
-     "\xBF\xD5\x00\x00\xB0\xA9\x03\x00\xB1\xA9\x03\x00\xB2\xBD\x03\x00\xB3\xB5"
-     "\x03\x00"
-     "\xB4\x99\x03\x00\xB5\x99\x03\x00\xB6\x59\x03\x00\xB7\x59\x03\x00\xB3\x0D"
-     "\x02\x00"
-     "\x18\xD3\x00\x80\x1C\xD3\x00\x80\x20\xD3\x00\x80\x24\xD3\x00\x80\xB6\x3D"
-     "\x02\x00"
-     "\xB5\x05\x02\x00\x28\xD3\x00\x80\xBB\x6D\x02\x00\xBA\x6D\x02\x00\x2C\xD3"
-     "\x00\x80"
-     "\x30\xD3\x00\x80\xBF\x5D\x02\x00\xBE\x5D\x02\x00\xBD\x61\x02\x00\xBC\x69"
-     "\x02\x00"
-     "\x34\xD3\x00\x80\x38\xD3\x00\x80\x3C\xD3\x00\x80\x40\xD3\x00\x80\xE3\x68"
-     "\x03\x00"
-     "\x44\xD3\x00\x80\xE1\xA8\x01\x00\x48\xD3\x00\x80\xEF\x94\x03\x00\x4C\xD3"
-     "\x00\x80"
-     "\x50\xD3\x00\x80\x54\xD3\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05"
-     "\x00\x00"
-     "\x58\xD3\x00\x80\xE1\xC4\x06\x00\xE1\x3C\x07\x00\xE3\xD4\x06\x00\xE3\xBC"
-     "\x06\x00"
-     "\x84\x90\x02\x00\x60\xD3\x00\x80\xBE\x14\x0D\x00\x64\xD3\x00\x80\x68\xD3"
-     "\x00\x80"
-     "\x6C\xD3\x00\x80\x70\xD3\x00\x80\x74\xD3\x00\x80\x78\xD3\x00\x80\x7C\xD3"
-     "\x00\x80"
-     "\xEF\x90\x01\x00\xEF\x6C\x06\x00\x86\x60\x0C\x00\x87\xAC\x0D\x00\x80\xD3"
-     "\x00\x80"
-     "\xB3\x6D\x01\x00\x84\xD3\x00\x80\xB5\x7D\x01\x00\xB6\x75\x01\x00\x88\xD3"
-     "\x00\x80"
-     "\x8C\xD3\x00\x80\x90\xD3\x00\x80\xBA\x51\x01\x00\xBB\x51\x01\x00\xBC\xF5"
-     "\x00\x00"
-     "\xBD\xFD\x00\x00\xBE\xF5\x00\x00\xBF\xED\x00\x00\xA3\xA1\x0E\x00\x5C\xD3"
-     "\x00\x80"
-     "\x94\xD3\x00\x80\x98\xD3\x00\x80\x9C\xD3\x00\x80\xA6\xB9\x0E\x00\xA5\xB1"
-     "\x0E\x00"
-     "\xA0\xD3\x00\x80\xAB\x9D\x0E\x00\xAA\x9D\x0E\x00\xA4\xD3\x00\x80\xA8\xD3"
-     "\x00\x80"
-     "\xAF\x21\x0F\x00\xAE\x39\x0F\x00\xAD\x31\x0F\x00\xAC\x39\x0F\x00\xAC\xD3"
-     "\x00\x80"
-     "\xB0\xD3\x00\x80\xB4\xD3\x00\x80\xB8\xD3\x00\x80\xBC\xD3\x00\x80\xE3\xD4"
-     "\x0D\x00"
-     "\xC0\xD3\x00\x80\xE1\x24\x0C\x00\x82\x15\x00\x00\xEF\x04\x0D\x00\x80\x31"
-     "\x00\x00"
-     "\x81\x1D\x00\x00\xC4\xD3\x00\x80\xC8\xD3\x00\x80\xBE\x30\x03\x00\x84\x30"
-     "\x03\x00"
-     "\xA8\xD1\x0E\x00\xA9\xD9\x0E\x00\xAA\x29\x0E\x00\xAB\x29\x0E\x00\xAC\x39"
-     "\x0E\x00"
-     "\xAD\x39\x0E\x00\xAE\x29\x0E\x00\xAF\x29\x0E\x00\xCC\xD3\x00\x80\x84\xF8"
-     "\x00\x00"
-     "\x86\x0C\x00\x00\x87\x8C\x00\x00\xD0\xD3\x00\x80\xD4\xD3\x00\x80\xD8\xD3"
-     "\x00\x80"
-     "\xDC\xD3\x00\x80\xB8\xED\x0F\x00\xB9\x85\x0F\x00\xBA\x8D\x0F\x00\xBB\x85"
-     "\x0F\x00"
-     "\xBC\x9D\x0F\x00\xBD\x81\x0F\x00\xBE\x85\x0F\x00\xBF\xB9\x0F\x00\xB0\x59"
-     "\x0E\x00"
-     "\xB1\x59\x0E\x00\xB2\xED\x0F\x00\xB3\xE5\x0F\x00\xB4\xFD\x0F\x00\xB5\xE5"
-     "\x0F\x00"
-     "\xB6\xE5\x0F\x00\xB7\xD5\x0F\x00\xA8\x95\x0E\x00\xA9\x99\x0E\x00\xAA\xA9"
-     "\x0E\x00"
-     "\xAB\xA9\x0E\x00\xAC\xB9\x0E\x00\xAD\xB9\x0E\x00\xAE\xA9\x0E\x00\xAF\xA9"
-     "\x0E\x00"
-     "\xE0\xD3\x00\x80\xE4\xD3\x00\x80\xE8\xD3\x00\x80\xEC\xD3\x00\x80\xF0\xD3"
-     "\x00\x80"
-     "\xF4\xD3\x00\x80\xF8\xD3\x00\x80\xFC\xD3\x00\x80\xB8\x5D\x01\x00\xB9\x6D"
-     "\x01\x00"
-     "\xBA\x65\x01\x00\xBB\x0D\x01\x00\xBC\x15\x01\x00\xBD\x1D\x01\x00\xBE\x15"
-     "\x01\x00"
-     "\xBF\x0D\x01\x00\xB0\xA9\x0E\x00\xB1\xB5\x0E\x00\xB2\xB5\x0E\x00\xB3\x8D"
-     "\x0E\x00"
-     "\xB4\x95\x0E\x00\xB5\x7D\x01\x00\xB6\x75\x01\x00\xB7\x6D\x01\x00\xB3\x21"
-     "\x0E\x00"
-     "\x00\xD4\x00\x80\x04\xD4\x00\x80\x08\xD4\x00\x80\x0C\xD4\x00\x80\xB6\x39"
-     "\x0E\x00"
-     "\xB5\x31\x0E\x00\x10\xD4\x00\x80\xBB\x71\x0E\x00\xBA\x6D\x0E\x00\x14\xD4"
-     "\x00\x80"
-     "\x18\xD4\x00\x80\xBF\xB9\x01\x00\xBE\x51\x0E\x00\xBD\x59\x0E\x00\xBC\x61"
-     "\x0E\x00"
-     "\x1C\xD4\x00\x80\xA3\x65\x0E\x00\x20\xD4\x00\x80\x24\xD4\x00\x80\xA6\x7D"
-     "\x0E\x00"
-     "\x28\xD4\x00\x80\x2C\xD4\x00\x80\xA5\x75\x0E\x00\xAA\x29\x0E\x00\xAB\x35"
-     "\x0E\x00"
-     "\x30\xD4\x00\x80\x34\xD4\x00\x80\xAE\x15\x0E\x00\xAF\xFD\x01\x00\xAC\x25"
-     "\x0E\x00"
-     "\xAD\x1D\x0E\x00\xA9\xB5\x02\x00\xA8\xB5\x02\x00\xAB\xB5\x02\x00\xAA\xB5"
-     "\x02\x00"
-     "\xAD\x35\x03\x00\xAC\x35\x03\x00\xAF\x35\x03\x00\xAE\x35\x03\x00\x80\xC9"
-     "\x03\x00"
-     "\x81\xC9\x03\x00\x82\x05\x00\x00\x38\xD4\x00\x80\xBE\x64\x02\x00\x40\xD4"
-     "\x00\x80"
-     "\x87\x78\x03\x00\x86\x1C\x04\x00\xB9\xE9\x03\x00\xB8\xD9\x03\x00\xBB\x89"
-     "\x03\x00"
-     "\xBA\xE1\x03\x00\xBD\x9D\x03\x00\xBC\x9D\x03\x00\xBF\x81\x03\x00\xBE\x89"
-     "\x03\x00"
-     "\xB1\x55\x03\x00\xB0\x55\x03\x00\xB3\x55\x03\x00\xB2\x55\x03\x00\xB5\xF9"
-     "\x03\x00"
-     "\xB4\xF1\x03\x00\xB7\xE9\x03\x00\xB6\xF1\x03\x00\x84\x00\x03\x00\xB6\xA1"
-     "\x03\x00"
-     "\xB5\xA9\x03\x00\x84\xC0\x05\x00\xB3\x29\x02\x00\x44\xD4\x00\x80\x48\xD4"
-     "\x00\x80"
-     "\x4C\xD4\x00\x80\xBF\x29\x03\x00\xBE\x21\x03\x00\xBD\x95\x03\x00\xBC\x95"
-     "\x03\x00"
-     "\xBB\x85\x03\x00\xBA\x85\x03\x00\x50\xD4\x00\x80\x54\xD4\x00\x80\xA5\xE1"
-     "\x03\x00"
-     "\x58\xD4\x00\x80\x5C\xD4\x00\x80\xA6\xE9\x03\x00\x60\xD4\x00\x80\x64\xD4"
-     "\x00\x80"
-     "\xA3\x61\x02\x00\x68\xD4\x00\x80\xAD\xDD\x03\x00\xAC\xDD\x03\x00\xAF\x61"
-     "\x03\x00"
-     "\xAE\x69\x03\x00\x6C\xD4\x00\x80\x70\xD4\x00\x80\xAB\xCD\x03\x00\xAA\xCD"
-     "\x03\x00"
-     "\x74\xD4\x00\x80\x78\xD4\x00\x80\x7C\xD4\x00\x80\x80\xD4\x00\x80\x84\xD4"
-     "\x00\x80"
-     "\x88\xD4\x00\x80\x8C\xD4\x00\x80\xEF\xD4\x03\x00\x90\xD4\x00\x80\xE1\x94"
-     "\x01\x00"
-     "\x94\xD4\x00\x80\xE3\x98\x01\x00\x98\xD4\x00\x80\x9C\xD4\x00\x80\xA0\xD4"
-     "\x00\x80"
-     "\xA4\xD4\x00\x80\x80\x79\x00\x00\x81\x79\x00\x00\x82\x19\x00\x00\xB3\xD5"
-     "\x01\x00"
-     "\xA8\xD4\x00\x80\xB5\xED\x01\x00\xB6\xE5\x01\x00\xBE\x14\x02\x00\x86\xC0"
-     "\x04\x00"
-     "\x87\x40\x05\x00\xBA\xB9\x01\x00\xBB\xB9\x01\x00\xBC\xAD\x01\x00\xBD\x85"
-     "\x01\x00"
-     "\xBE\x8D\x01\x00\xBF\x7D\x00\x00\xE1\x58\x0E\x00\xB0\xD4\x00\x80\xE3\xE4"
-     "\x01\x00"
-     "\xB4\xD4\x00\x80\xB8\xD4\x00\x80\xBC\xD4\x00\x80\xC0\xD4\x00\x80\xC4\xD4"
-     "\x00\x80"
-     "\xBE\xE8\x04\x00\xC8\xD4\x00\x80\xCC\xD4\x00\x80\xD0\xD4\x00\x80\xD4\xD4"
-     "\x00\x80"
-     "\xD8\xD4\x00\x80\xEF\x38\x01\x00\xDC\xD4\x00\x80\xE0\xD4\x00\x80\xA3\x5D"
-     "\x01\x00"
-     "\xE4\xD4\x00\x80\xE8\xD4\x00\x80\xA6\x6D\x01\x00\xEC\xD4\x00\x80\xEC\xD2"
-     "\x00\x80"
-     "\xA5\x65\x01\x00\xAA\x31\x01\x00\xAB\x31\x01\x00\xF0\xD4\x00\x80\xF4\xD4"
-     "\x00\x80"
-     "\xAE\x05\x01\x00\xAF\xF5\x00\x00\xAC\x25\x01\x00\xAD\x0D\x01\x00\xAC\xD4"
-     "\x00\x80"
-     "\xF8\xD4\x00\x80\xFC\xD4\x00\x80\x00\xD5\x00\x80\x04\xD5\x00\x80\x08\xD5"
-     "\x00\x80"
-     "\x0C\xD5\x00\x80\x10\xD5\x00\x80\x81\x09\x00\x00\x80\x39\x00\x00\x14\xD5"
-     "\x00\x80"
-     "\x82\x1D\x00\x00\x18\xD5\x00\x80\x1C\xD5\x00\x80\x20\xD5\x00\x80\x24\xD5"
-     "\x00\x80"
-     "\x86\xE0\x00\x00\x87\x74\x03\x00\x28\xD5\x00\x80\x2C\xD5\x00\x80\xEF\xF4"
-     "\x05\x00"
-     "\x30\xD5\x00\x80\x34\xD5\x00\x80\x38\xD5\x00\x80\xE3\xF0\x04\x00\x3C\xD5"
-     "\x00\x80"
-     "\xE1\x4C\x04\x00\x40\xD5\x00\x80\x44\xD5\x00\x80\x48\xD5\x00\x80\x4C\xD5"
-     "\x00\x80"
-     "\x50\xD5\x00\x80\xA8\xB5\x06\x00\xA9\x11\x06\x00\xAA\x11\x06\x00\xAB\x11"
-     "\x06\x00"
-     "\xAC\x29\x06\x00\xAD\x29\x06\x00\xAE\x5D\x06\x00\xAF\x55\x06\x00\x84\xE4"
-     "\x00\x00"
-     "\x54\xD5\x00\x80\x58\xD5\x00\x80\x5C\xD5\x00\x80\x60\xD5\x00\x80\x64\xD5"
-     "\x00\x80"
-     "\x68\xD5\x00\x80\x6C\xD5\x00\x80\xB8\xCD\x07\x00\xB9\xD5\x07\x00\xBA\xDD"
-     "\x07\x00"
-     "\xBB\xD5\x07\x00\xBC\xF1\x07\x00\xBD\xFD\x07\x00\xBE\xE9\x07\x00\xBF\xE9"
-     "\x07\x00"
-     "\xB0\x31\x06\x00\xB1\x31\x06\x00\xB2\x35\x06\x00\xB3\x1D\x06\x00\xB4\x05"
-     "\x06\x00"
-     "\xB5\xF5\x07\x00\xB6\xFD\x07\x00\xB7\xF5\x07\x00\xA8\x85\x06\x00\xA9\x89"
-     "\x06\x00"
-     "\xAA\x99\x06\x00\xAB\x99\x06\x00\xAC\x89\x06\x00\xAD\x89\x06\x00\xAE\xBD"
-     "\x06\x00"
-     "\xAF\xB5\x06\x00\x70\xD5\x00\x80\x74\xD5\x00\x80\x78\xD5\x00\x80\x7C\xD5"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\xB1\x00\x00\x82\xB1\x00\x00\x80\xD5\x00\x80\xB8\xA9"
-     "\x06\x00"
-     "\xB9\xA9\x06\x00\xBA\x4D\x01\x00\xBB\x45\x01\x00\xBC\x5D\x01\x00\xBD\x45"
-     "\x01\x00"
-     "\xBE\x4D\x01\x00\xBF\x45\x01\x00\xB0\xCD\x06\x00\xB1\xD5\x06\x00\xB2\xDD"
-     "\x06\x00"
-     "\xB3\xD1\x06\x00\xB4\xB1\x06\x00\xB5\xBD\x06\x00\xB6\xB5\x06\x00\xB7\xA9"
-     "\x06\x00"
-     "\xB3\x11\x06\x00\x84\xD5\x00\x80\x88\xD5\x00\x80\xBE\xCC\x01\x00\x90\xD5"
-     "\x00\x80"
-     "\xB6\x09\x06\x00\xB5\x01\x06\x00\x94\xD5\x00\x80\xBB\x11\x06\x00\xBA\x09"
-     "\x06\x00"
-     "\x86\x08\x00\x00\x87\x9C\x00\x00\xBF\x79\x06\x00\xBE\x75\x06\x00\xBD\x05"
-     "\x06\x00"
-     "\xBC\x09\x06\x00\x98\xD5\x00\x80\xA3\x55\x06\x00\x9C\xD5\x00\x80\xA0\xD5"
-     "\x00\x80"
-     "\xA6\x4D\x06\x00\xA4\xD5\x00\x80\xA8\xD5\x00\x80\xA5\x45\x06\x00\xAA\x4D"
-     "\x06\x00"
-     "\xAB\x55\x06\x00\xAC\xD5\x00\x80\xB0\xD5\x00\x80\xAE\x31\x06\x00\xAF\x3D"
-     "\x06\x00"
-     "\xAC\x4D\x06\x00\xAD\x41\x06\x00\xA8\x39\x02\x00\xA9\x39\x02\x00\xAA\x41"
-     "\x02\x00"
-     "\xAB\x41\x02\x00\xAC\x45\x02\x00\xAD\x4D\x02\x00\xAE\x45\x02\x00\xAF\x79"
-     "\x02\x00"
-     "\xBE\xAC\x1C\x00\xBE\xEC\x1D\x00\xB4\xD5\x00\x80\xB8\xD5\x00\x80\xBC\xD5"
-     "\x00\x80"
-     "\xC0\xD5\x00\x80\xC4\xD5\x00\x80\xC8\xD5\x00\x80\xB8\xF1\x02\x00\xB9\xF1"
-     "\x02\x00"
-     "\xBA\xF1\x02\x00\xBB\xF1\x02\x00\xBC\x71\x03\x00\xBD\x71\x03\x00\xBE\x71"
-     "\x03\x00"
-     "\xBF\x71\x03\x00\xB0\x0D\x02\x00\xB1\x15\x02\x00\xB2\x1D\x02\x00\xB3\x15"
-     "\x02\x00"
-     "\xB4\x0D\x02\x00\xB5\xD1\x02\x00\xB6\xD1\x02\x00\xB7\xD1\x02\x00\xCC\xD5"
-     "\x00\x80"
-     "\xD0\xD5\x00\x80\xB3\xC1\x02\x00\xD4\xD5\x00\x80\xB5\xC5\x02\x00\xD8\xD5"
-     "\x00\x80"
-     "\xDC\xD5\x00\x80\xB6\xC5\x02\x00\xE0\xD5\x00\x80\xE4\xD5\x00\x80\xBB\xD9"
-     "\x02\x00"
-     "\xBA\xC1\x02\x00\xBD\xC9\x02\x00\xBC\xC1\x02\x00\xBF\x1D\x02\x00\xBE\xC1"
-     "\x02\x00"
-     "\xE8\xD5\x00\x80\xEC\xD5\x00\x80\xF0\xD5\x00\x80\xA3\x85\x02\x00\xF4\xD5"
-     "\x00\x80"
-     "\xA5\x81\x02\x00\xA6\x81\x02\x00\xF8\xD5\x00\x80\x84\xEC\x03\x00\x00\xD6"
-     "\x00\x80"
-     "\xAA\x85\x02\x00\xAB\x9D\x02\x00\xAC\x85\x02\x00\xAD\x8D\x02\x00\xAE\x85"
-     "\x02\x00"
-     "\xAF\x59\x02\x00\xA8\x91\x02\x00\xA9\x99\x02\x00\xAA\xC9\x02\x00\xAB\xC5"
-     "\x02\x00"
-     "\xAC\xDD\x02\x00\xAD\xC1\x02\x00\xAE\xC1\x02\x00\xAF\xF5\x02\x00\x86\x00"
-     "\x1D\x00"
-     "\x87\x64\x1D\x00\x04\xD6\x00\x80\x08\xD6\x00\x80\x80\xF9\x01\x00\x81\xE5"
-     "\x01\x00"
-     "\x82\xE5\x01\x00\x0C\xD6\x00\x80\xB8\xC5\x01\x00\xB9\xC9\x01\x00\xBA\xDD"
-     "\x01\x00"
-     "\xBB\xD5\x01\x00\xBC\xF1\x01\x00\xBD\xF1\x01\x00\xBE\x99\x01\x00\xBF\x99"
-     "\x01\x00"
-     "\xB0\x8D\x02\x00\xB1\x4D\x01\x00\xB2\x45\x01\x00\xB3\x5D\x01\x00\xB4\x45"
-     "\x01\x00"
-     "\xB5\x4D\x01\x00\xB6\x45\x01\x00\xB7\xFD\x01\x00\x10\xD6\x00\x80\xE1\xFC"
-     "\x06\x00"
-     "\xEF\xFC\x00\x00\xE3\x10\x07\x00\x14\xD6\x00\x80\x18\xD6\x00\x80\x1C\xD6"
-     "\x00\x80"
-     "\x20\xD6\x00\x80\x84\x6C\x1C\x00\x24\xD6\x00\x80\x28\xD6\x00\x80\x2C\xD6"
-     "\x00\x80"
-     "\xE1\xE8\x01\x00\x30\xD6\x00\x80\xE3\x08\x01\x00\xEF\xE4\x06\x00\xB3\x01"
-     "\x02\x00"
-     "\x34\xD6\x00\x80\x38\xD6\x00\x80\x3C\xD6\x00\x80\x40\xD6\x00\x80\xB6\x01"
-     "\x02\x00"
-     "\xB5\x15\x02\x00\x44\xD6\x00\x80\xBB\xC9\x01\x00\xBA\x09\x02\x00\x48\xD6"
-     "\x00\x80"
-     "\x3C\xD4\x00\x80\xBF\xC9\x01\x00\xBE\xC1\x01\x00\xBD\xC9\x01\x00\xBC\xD1"
-     "\x01\x00"
-     "\xA3\xC1\x1D\x00\xFC\xD5\x00\x80\x4C\xD6\x00\x80\x50\xD6\x00\x80\x54\xD6"
-     "\x00\x80"
-     "\xA6\xC1\x1D\x00\xA5\xD5\x1D\x00\x58\xD6\x00\x80\xAB\x09\x1E\x00\xAA\xC9"
-     "\x1D\x00"
-     "\x5C\xD6\x00\x80\x60\xD6\x00\x80\xAF\x09\x1E\x00\xAE\x01\x1E\x00\xAD\x09"
-     "\x1E\x00"
-     "\xAC\x11\x1E\x00\x80\x6D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\xEF\x94"
-     "\x1C\x00"
-     "\x64\xD6\x00\x80\x68\xD6\x00\x80\x6C\xD6\x00\x80\x70\xD6\x00\x80\x86\x40"
-     "\x03\x00"
-     "\x87\xC4\x03\x00\x74\xD6\x00\x80\x78\xD6\x00\x80\x7C\xD6\x00\x80\xE1\xC4"
-     "\x1F\x00"
-     "\x80\xD6\x00\x80\xE3\x8C\x1C\x00\xA8\x65\x1E\x00\xA9\xAD\x1F\x00\xAA\xA5"
-     "\x1F\x00"
-     "\xAB\xB1\x1F\x00\xAC\x11\x1F\x00\xAD\x05\x1F\x00\xAE\x0D\x1F\x00\xAF\x05"
-     "\x1F\x00"
-     "\x84\xD6\x00\x80\x88\xD6\x00\x80\x8C\xD6\x00\x80\x90\xD6\x00\x80\x94\xD6"
-     "\x00\x80"
-     "\x98\xD6\x00\x80\x9C\xD6\x00\x80\xA0\xD6\x00\x80\xB8\x11\x1F\x00\xB9\x19"
-     "\x1F\x00"
-     "\xBA\x21\x1F\x00\xBB\xC5\x00\x00\xBC\xDD\x00\x00\xBD\xC5\x00\x00\xBE\xC5"
-     "\x00\x00"
-     "\xBF\xF5\x00\x00\xB0\x7D\x1F\x00\xB1\x41\x1F\x00\xB2\x41\x1F\x00\xB3\x41"
-     "\x1F\x00"
-     "\xB4\x45\x1F\x00\xB5\x31\x1F\x00\xB6\x31\x1F\x00\xB7\x31\x1F\x00\xB3\x2D"
-     "\x1E\x00"
-     "\xA4\xD6\x00\x80\xA8\xD6\x00\x80\xAC\xD6\x00\x80\xB0\xD6\x00\x80\xB6\x85"
-     "\x1E\x00"
-     "\xB5\x99\x1E\x00\xB4\xD6\x00\x80\xBB\x85\x1E\x00\xBA\xB9\x1E\x00\xB8\xD6"
-     "\x00\x80"
-     "\xBC\xD6\x00\x80\xBF\xF5\x1E\x00\xBE\x81\x1E\x00\xBD\x95\x1E\x00\xBC\x95"
-     "\x1E\x00"
-     "\xC0\xD6\x00\x80\xA3\x69\x1E\x00\xC4\xD6\x00\x80\xC8\xD6\x00\x80\xA6\xC1"
-     "\x1E\x00"
-     "\xCC\xD6\x00\x80\xD0\xD6\x00\x80\xA5\xDD\x1E\x00\xAA\xFD\x1E\x00\xAB\xC1"
-     "\x1E\x00"
-     "\xD4\xD6\x00\x80\xD8\xD6\x00\x80\xAE\xC5\x1E\x00\xAF\xB1\x1E\x00\xAC\xD1"
-     "\x1E\x00"
-     "\xAD\xD1\x1E\x00\xA8\xDD\x01\x00\xA9\xED\x01\x00\xAA\xE5\x01\x00\xAB\xFD"
-     "\x01\x00"
-     "\xAC\xE9\x01\x00\xAD\x35\x01\x00\xAE\x3D\x01\x00\xAF\x35\x01\x00\x80\xF9"
-     "\x00\x00"
-     "\x81\xCD\x00\x00\x82\xC5\x00\x00\x84\x60\x01\x00\xBE\x78\x01\x00\xE0\xD6"
-     "\x00\x80"
-     "\x87\xAC\x00\x00\x86\x9C\x01\x00\xB8\xD1\x00\x00\xB9\xD9\x00\x00\xBA\xE1"
-     "\x00\x00"
-     "\xBB\xE1\x00\x00\xBC\x91\x00\x00\xBD\x9D\x00\x00\xBE\x95\x00\x00\xBF\x89"
-     "\x00\x00"
-     "\xB0\x4D\x01\x00\xB1\x55\x01\x00\xB2\x5D\x01\x00\xB3\x55\x01\x00\xB4\x4D"
-     "\x01\x00"
-     "\xB5\xF1\x00\x00\xB6\xF1\x00\x00\xB7\xF1\x00\x00\xB3\x6D\x02\x00\xE4\xD6"
-     "\x00\x80"
-     "\xE8\xD6\x00\x80\xEC\xD6\x00\x80\xF0\xD6\x00\x80\xB6\xB5\x02\x00\xB5\x61"
-     "\x02\x00"
-     "\xF4\xD6\x00\x80\xBB\x91\x02\x00\xBA\x89\x02\x00\xF8\xD6\x00\x80\xFC\xD6"
-     "\x00\x80"
-     "\xBF\x35\x03\x00\xBE\x89\x02\x00\xBD\x81\x02\x00\xBC\x89\x02\x00\x90\x00"
-     "\x00\x00"
-     "\xA3\x29\x02\x00\x00\xD7\x00\x80\x04\xD7\x00\x80\xA6\xF1\x02\x00\x08\xD7"
-     "\x00\x80"
-     "\x0C\xD7\x00\x80\xA5\x25\x02\x00\xAA\xCD\x02\x00\xAB\xD5\x02\x00\x10\xD7"
-     "\x00\x80"
-     "\x14\xD7\x00\x80\xAE\xCD\x02\x00\xAF\x71\x03\x00\xAC\xCD\x02\x00\xAD\xC5"
-     "\x02\x00"
-     "\xA8\x35\x03\x00\xA9\x81\x00\x00\xAA\x81\x00\x00\xAB\x81\x00\x00\xAC\x81"
-     "\x00\x00"
-     "\xAD\x89\x00\x00\xAE\xB1\x00\x00\xAF\xB1\x00\x00\x18\xD7\x00\x80\x1C\xD7"
-     "\x00\x80"
-     "\x20\xD7\x00\x80\x24\xD7\x00\x80\x28\xD7\x00\x80\x2C\xD7\x00\x80\x30\xD7"
-     "\x00\x80"
-     "\x34\xD7\x00\x80\xB8\x69\x00\x00\xB9\x69\x00\x00\xBA\x79\x00\x00\xBB\x79"
-     "\x00\x00"
-     "\xBC\x69\x00\x00\xBD\x69\x00\x00\xBE\xD9\x01\x00\xBF\xD9\x01\x00\xB0\xA9"
-     "\x00\x00"
-     "\xB1\xA9\x00\x00\xB2\xBD\x00\x00\xB3\xB5\x00\x00\xB4\x99\x00\x00\xB5\x99"
-     "\x00\x00"
-     "\xB6\x59\x00\x00\xB7\x59\x00\x00\x38\xD7\x00\x80\x3C\xD7\x00\x80\x40\xD7"
-     "\x00\x80"
-     "\x44\xD7\x00\x80\x81\x1D\x00\x00\x80\x1D\x00\x00\x48\xD7\x00\x80\x82\x05"
-     "\x00\x00"
-     "\x84\xAC\x02\x00\xBE\xE8\x05\x00\x87\xC0\x02\x00\x86\x0C\x05\x00\x50\xD7"
-     "\x00\x80"
-     "\x54\xD7\x00\x80\x58\xD7\x00\x80\x5C\xD7\x00\x80\xB3\x49\x01\x00\x60\xD7"
-     "\x00\x80"
-     "\xD2\x00\x00\x00\x64\xD7\x00\x80\x68\xD7\x00\x80\xB6\x49\x01\x00\xB5\x49"
-     "\x01\x00"
-     "\x6C\xD7\x00\x80\xBB\x2D\x01\x00\xBA\x2D\x01\x00\x70\xD7\x00\x80\x74\xD7"
-     "\x00\x80"
-     "\xBF\x1D\x01\x00\xBE\x1D\x01\x00\xBD\x21\x01\x00\xBC\x29\x01\x00\xE1\xB8"
-     "\x1E\x00"
-     "\x78\xD7\x00\x80\xE3\x38\x1E\x00\x7C\xD7\x00\x80\x80\xD7\x00\x80\x84\xD7"
-     "\x00\x80"
-     "\x88\xD7\x00\x80\x8C\xD7\x00\x80\x84\x6C\x04\x00\xBE\x6C\x04\x00\x90\xD7"
-     "\x00\x80"
-     "\x94\xD7\x00\x80\x98\xD7\x00\x80\x9C\xD7\x00\x80\xEF\x64\x1F\x00\x84\x70"
-     "\x07\x00"
-     "\xA0\xD7\x00\x80\xA4\xD7\x00\x80\x86\xE8\x04\x00\x87\xC0\x04\x00\xA8\xD7"
-     "\x00\x80"
-     "\xE3\x68\x00\x00\xAC\xD7\x00\x80\xE1\xA4\x01\x00\x82\x1D\x00\x00\xEF\xAC"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\x81\x15\x00\x00\x8C\xD5\x00\x80\xB0\xD7\x00\x80\xB4\xD7"
-     "\x00\x80"
-     "\xB8\xD7\x00\x80\xE1\xA4\x07\x00\x4C\xD7\x00\x80\xE3\x38\x05\x00\xC0\xD7"
-     "\x00\x80"
-     "\xBC\xD7\x00\x80\xC4\xD7\x00\x80\xC8\xD7\x00\x80\xCC\xD7\x00\x80\xD0\xD7"
-     "\x00\x80"
-     "\xD4\xD7\x00\x80\xD8\xD7\x00\x80\xDC\xD7\x00\x80\xE0\xD7\x00\x80\xE4\xD7"
-     "\x00\x80"
-     "\xEF\xE4\x05\x00\xE8\xD7\x00\x80\xA3\x49\x06\x00\xEC\xD7\x00\x80\xF0\xD7"
-     "\x00\x80"
-     "\xF4\xD7\x00\x80\xF8\xD7\x00\x80\xA6\x49\x06\x00\xA5\x49\x06\x00\xFC\xD7"
-     "\x00\x80"
-     "\xAB\x2D\x06\x00\xAA\x2D\x06\x00\x00\xD8\x00\x80\x04\xD8\x00\x80\xAF\x1D"
-     "\x06\x00"
-     "\xAE\x1D\x06\x00\xAD\x21\x06\x00\xAC\x29\x06\x00\xA8\xB1\x07\x00\xA9\xB1"
-     "\x07\x00"
-     "\xAA\xB1\x07\x00\xAB\xB1\x07\x00\xAC\xD5\x07\x00\xAD\xD9\x07\x00\xAE\xC9"
-     "\x07\x00"
-     "\xAF\xC1\x07\x00\x08\xD8\x00\x80\x0C\xD8\x00\x80\x10\xD8\x00\x80\x14\xD8"
-     "\x00\x80"
-     "\x18\xD8\x00\x80\x1C\xD8\x00\x80\x20\xD8\x00\x80\x24\xD8\x00\x80\xB8\x69"
-     "\x07\x00"
-     "\xB9\x69\x07\x00\xBA\x79\x07\x00\xBB\x79\x07\x00\xBC\x11\x07\x00\xBD\x11"
-     "\x07\x00"
-     "\xBE\x09\x07\x00\xBF\x05\x07\x00\xB0\x89\x07\x00\xB1\x89\x07\x00\xB2\x6D"
-     "\x07\x00"
-     "\xB3\x61\x07\x00\xB4\x61\x07\x00\xB5\x61\x07\x00\xB6\x61\x07\x00\xB7\x61"
-     "\x07\x00"
-     "\xB3\x39\x06\x00\x82\x15\x00\x00\x81\xB5\x01\x00\x80\xB5\x01\x00\x28\xD8"
-     "\x00\x80"
-     "\xB6\x59\x06\x00\xB5\x5D\x06\x00\xBE\xD4\x01\x00\xBB\x4D\x06\x00\xBA\x4D"
-     "\x06\x00"
-     "\x2C\xD8\x00\x80\x30\xD8\x00\x80\xBF\xA5\x06\x00\xBE\xA5\x06\x00\xBD\xB5"
-     "\x06\x00"
-     "\xBC\xB9\x06\x00\x34\xD8\x00\x80\xA3\x7D\x06\x00\x86\xE8\x00\x00\x87\xCC"
-     "\x01\x00"
-     "\xA6\x1D\x06\x00\x38\xD8\x00\x80\x3C\xD8\x00\x80\xA5\x19\x06\x00\xAA\x09"
-     "\x06\x00"
-     "\xAB\x09\x06\x00\x40\xD8\x00\x80\x44\xD8\x00\x80\xAE\xE1\x06\x00\xAF\xE1"
-     "\x06\x00"
-     "\xAC\xFD\x06\x00\xAD\xF1\x06\x00\xA8\x8D\x01\x00\xA9\x91\x01\x00\xAA\x91"
-     "\x01\x00"
-     "\xAB\x91\x01\x00\xAC\xB1\x01\x00\xAD\xB1\x01\x00\xAE\xB1\x01\x00\xAF\xB1"
-     "\x01\x00"
-     "\x84\xA8\x01\x00\x48\xD8\x00\x80\x4C\xD8\x00\x80\x50\xD8\x00\x80\x54\xD8"
-     "\x00\x80"
-     "\x58\xD8\x00\x80\x5C\xD8\x00\x80\x60\xD8\x00\x80\xB8\xA9\x01\x00\xB9\xA9"
-     "\x01\x00"
-     "\xBA\x69\x01\x00\xBB\x69\x01\x00\xBC\x79\x01\x00\xBD\x79\x01\x00\xBE\x69"
-     "\x01\x00"
-     "\xBF\x69\x01\x00\xB0\xD1\x01\x00\xB1\xD1\x01\x00\xB2\xD5\x01\x00\xB3\xA9"
-     "\x01\x00"
-     "\xB4\xB9\x01\x00\xB5\xB9\x01\x00\xB6\xA9\x01\x00\xB7\xA1\x01\x00\xB3\x19"
-     "\x02\x00"
-     "\x64\xD8\x00\x80\x68\xD8\x00\x80\x6C\xD8\x00\x80\x70\xD8\x00\x80\xB6\x31"
-     "\x02\x00"
-     "\xB5\x39\x02\x00\x74\xD8\x00\x80\xBB\x15\x02\x00\xBA\x15\x02\x00\x78\xD8"
-     "\x00\x80"
-     "\x7C\xD8\x00\x80\xBF\x7D\x02\x00\xBE\x7D\x02\x00\xBD\x7D\x02\x00\xBC\x01"
-     "\x02\x00"
-     "\x80\xD8\x00\x80\xA3\x5D\x02\x00\x84\xD8\x00\x80\x88\xD8\x00\x80\xA6\x75"
-     "\x02\x00"
-     "\x8C\xD8\x00\x80\x90\xD8\x00\x80\xA5\x7D\x02\x00\xAA\x51\x02\x00\xAB\x51"
-     "\x02\x00"
-     "\x94\xD8\x00\x80\x98\xD8\x00\x80\xAE\x39\x02\x00\xAF\x39\x02\x00\xAC\x45"
-     "\x02\x00"
-     "\xAD\x39\x02\x00\x80\xD9\x00\x00\x81\xED\x00\x00\x82\xE5\x00\x00\xBE\xC0"
-     "\x03\x00"
-     "\xA0\xD8\x00\x80\xEF\xE0\x02\x00\x87\x54\x03\x00\x86\xBC\x0C\x00\xA4\xD8"
-     "\x00\x80"
-     "\xE3\x14\x01\x00\xA8\xD8\x00\x80\xE1\xFC\x01\x00\xAC\xD8\x00\x80\xB0\xD8"
-     "\x00\x80"
-     "\xB4\xD8\x00\x80\xB8\xD8\x00\x80\xBC\xD8\x00\x80\xB3\x91\x03\x00\xC0\xD8"
-     "\x00\x80"
-     "\x84\x40\x0C\x00\xB6\xBD\x03\x00\xC4\xD8\x00\x80\xC8\xD8\x00\x80\xB5\xB5"
-     "\x03\x00"
-     "\xBA\x4D\x03\x00\xBB\x55\x03\x00\xCC\xD8\x00\x80\xD0\xD8\x00\x80\xBE\x4D"
-     "\x03\x00"
-     "\xBF\xF1\x00\x00\xBC\x4D\x03\x00\xBD\x45\x03\x00\xA8\x2D\x02\x00\xA9\x45"
-     "\x02\x00"
-     "\xAA\x5D\x02\x00\xAB\x55\x02\x00\xAC\x79\x02\x00\xAD\x79\x02\x00\xAE\xBD"
-     "\x02\x00"
-     "\xAF\xB5\x02\x00\xBE\xAC\x0C\x00\xD4\xD8\x00\x80\xD8\xD8\x00\x80\xDC\xD8"
-     "\x00\x80"
-     "\xE0\xD8\x00\x80\xE4\xD8\x00\x80\xE8\xD8\x00\x80\xEC\xD8\x00\x80\xB8\x51"
-     "\x01\x00"
-     "\xB9\x59\x01\x00\xBA\x61\x01\x00\xBB\x61\x01\x00\xBC\x11\x01\x00\xBD\x1D"
-     "\x01\x00"
-     "\xBE\x15\x01\x00\xBF\x09\x01\x00\xB0\xCD\x02\x00\xB1\xD5\x02\x00\xB2\xDD"
-     "\x02\x00"
-     "\xB3\xD5\x02\x00\xB4\xCD\x02\x00\xB5\x71\x01\x00\xB6\x71\x01\x00\xB7\x71"
-     "\x01\x00"
-     "\xE1\xB4\x06\x00\xDC\xD6\x00\x80\xE3\x38\x06\x00\xF0\xD8\x00\x80\xF4\xD8"
-     "\x00\x80"
-     "\xE1\x44\x06\x00\xF8\xD8\x00\x80\xE3\x30\x06\x00\xFC\xD8\x00\x80\x00\xD9"
-     "\x00\x80"
-     "\x04\xD9\x00\x80\xEF\x48\x00\x00\xBE\x3C\x0F\x00\x08\xD9\x00\x80\xEF\xE4"
-     "\x06\x00"
-     "\x10\xD9\x00\x80\xA3\x95\x02\x00\x82\x71\x00\x00\x81\x61\x00\x00\x80\x69"
-     "\x00\x00"
-     "\x14\xD9\x00\x80\xA6\xB9\x02\x00\xA5\xB1\x02\x00\x18\xD9\x00\x80\xAB\x51"
-     "\x02\x00"
-     "\xAA\x49\x02\x00\x86\xC8\x0C\x00\x87\x5C\x0D\x00\xAF\xF5\x01\x00\xAE\x49"
-     "\x02\x00"
-     "\xAD\x41\x02\x00\xAC\x49\x02\x00\x0C\xD9\x00\x80\x1C\xD9\x00\x80\x20\xD9"
-     "\x00\x80"
-     "\x24\xD9\x00\x80\xE1\x94\x0C\x00\x28\xD9\x00\x80\xE3\xC4\x0C\x00\x2C\xD9"
-     "\x00\x80"
-     "\x30\xD9\x00\x80\x34\xD9\x00\x80\xEF\xF0\x0D\x00\x38\xD9\x00\x80\x3C\xD9"
-     "\x00\x80"
-     "\x40\xD9\x00\x80\x44\xD9\x00\x80\x48\xD9\x00\x80\xB3\xE9\x0E\x00\x4C\xD9"
-     "\x00\x80"
-     "\x50\xD9\x00\x80\x54\xD9\x00\x80\x58\xD9\x00\x80\xB6\xED\x0E\x00\xB5\xED"
-     "\x0E\x00"
-     "\x5C\xD9\x00\x80\xBB\xB5\x0E\x00\xBA\xAD\x0E\x00\x60\xD9\x00\x80\x64\xD9"
-     "\x00\x80"
-     "\xBF\x55\x0F\x00\xBE\x99\x0E\x00\xBD\x91\x0E\x00\xBC\xA5\x0E\x00\xA8\x11"
-     "\x0E\x00"
-     "\xA9\x25\x0E\x00\xAA\x21\x0E\x00\xAB\x3D\x0E\x00\xAC\x25\x0E\x00\xAD\x2D"
-     "\x0E\x00"
-     "\xAE\x25\x0E\x00\xAF\x5D\x0E\x00\x68\xD9\x00\x80\x6C\xD9\x00\x80\x70\xD9"
-     "\x00\x80"
-     "\x74\xD9\x00\x80\x80\x19\x00\x00\x81\x19\x00\x00\x82\x05\x00\x00\x78\xD9"
-     "\x00\x80"
-     "\xB8\xF1\x0F\x00\xB9\xF1\x0F\x00\xBA\xF1\x0F\x00\xBB\xF1\x0F\x00\xBC\x5D"
-     "\x0F\x00"
-     "\xBD\x45\x0F\x00\xBE\x41\x0F\x00\xBF\x7D\x0F\x00\xB0\x29\x0E\x00\xB1\x31"
-     "\x0E\x00"
-     "\xB2\x31\x0E\x00\xB3\x05\x0E\x00\xB4\x1D\x0E\x00\xB5\xD1\x0F\x00\xB6\xD1"
-     "\x0F\x00"
-     "\xB7\xD1\x0F\x00\xA3\xA9\x0F\x00\x84\x8C\x02\x00\x85\x18\x01\x00\xBF\x04"
-     "\x00\x00"
-     "\x85\x74\x01\x00\xA6\xAD\x0F\x00\xA5\xAD\x0F\x00\x80\xD9\x00\x80\xAB\xF5"
-     "\x0F\x00"
-     "\xAA\xED\x0F\x00\x86\x88\x07\x00\x87\x58\x00\x00\xAF\x15\x0E\x00\xAE\xD9"
-     "\x0F\x00"
-     "\xAD\xD1\x0F\x00\xAC\xE5\x0F\x00\x84\xD9\x00\x80\xB3\x99\x0E\x00\x88\xD9"
-     "\x00\x80"
-     "\x8C\xD9\x00\x80\xB6\x5D\x01\x00\x90\xD9\x00\x80\x94\xD9\x00\x80\xB5\x55"
-     "\x01\x00"
-     "\xBA\x41\x01\x00\xBB\x41\x01\x00\x98\xD9\x00\x80\x9C\xD9\x00\x80\xBE\x45"
-     "\x01\x00"
-     "\xBF\x4D\x01\x00\xBC\x55\x01\x00\xBD\x4D\x01\x00\xA8\x3D\x0E\x00\xA9\x55"
-     "\x0E\x00"
-     "\xAA\x55\x0E\x00\xAB\x89\x01\x00\xAC\x99\x01\x00\xAD\x99\x01\x00\xAE\x8D"
-     "\x01\x00"
-     "\xAF\x81\x01\x00\xA0\xD9\x00\x80\xA4\xD9\x00\x80\xA8\xD9\x00\x80\xAC\xD9"
-     "\x00\x80"
-     "\xB0\xD9\x00\x80\xB4\xD9\x00\x80\xB8\xD9\x00\x80\xBC\xD9\x00\x80\xB8\x89"
-     "\x01\x00"
-     "\xB9\x89\x01\x00\xBA\x99\x01\x00\xBB\x91\x01\x00\xBC\xB9\x01\x00\xBD\xB9"
-     "\x01\x00"
-     "\xBE\x79\x01\x00\xBF\x79\x01\x00\xB0\xC1\x01\x00\xB1\xC1\x01\x00\xB2\xC5"
-     "\x01\x00"
-     "\xB3\xDD\x01\x00\xB4\xC5\x01\x00\xB5\xCD\x01\x00\xB6\xC5\x01\x00\xB7\xB9"
-     "\x01\x00"
-     "\xA3\xDD\x0D\x00\xC0\xD9\x00\x80\xC4\xD9\x00\x80\xC8\xD9\x00\x80\xCC\xD9"
-     "\x00\x80"
-     "\xA6\x19\x02\x00\xA5\x11\x02\x00\xD0\xD9\x00\x80\xAB\x05\x02\x00\xAA\x05"
-     "\x02\x00"
-     "\xD4\xD9\x00\x80\x8F\x79\x1A\x00\xAF\x09\x02\x00\xAE\x01\x02\x00\xAD\x09"
-     "\x02\x00"
-     "\xAC\x11\x02\x00\x94\x9D\x1E\x00\x95\x9D\x11\x00\x96\x4D\x12\x00\x97\xED"
-     "\x12\x00"
-     "\x90\xD5\x1A\x00\x91\xBD\x1A\x00\x92\x59\x1E\x00\x93\x59\x1E\x00\x9C\x65"
-     "\x0A\x00"
-     "\x9D\x3D\x0A\x00\x9E\xED\x0D\x00\x9F\x69\x0E\x00\x98\x9D\x15\x00\x99\x75"
-     "\x16\x00"
-     "\x9A\x4D\x16\x00\x9B\x8D\x09\x00\xE1\x84\x01\x00\xD8\xD9\x00\x80\xE3\xDC"
-     "\xFE\x03"
-     "\x84\x80\x03\x00\xA0\x81\x0E\x00\xA1\xBD\x0F\x00\xA2\x0D\x01\x00\xA3\x9D"
-     "\x02\x00"
-     "\xB1\x01\xE8\x03\xB0\x49\xF6\x03\xB3\x69\xEA\x03\xB2\x8D\xEA\x03\xB5\x91"
-     "\xEE\x03"
-     "\xB4\x01\xEC\x03\xEF\x60\xEC\x03\xDC\xD9\x00\x80\xA9\xD1\xFE\x03\xA8\xC1"
-     "\xFE\x03"
-     "\xAB\xCD\xF2\x03\xAA\xBD\xFF\x03\xAD\x91\xF3\x03\xAC\xDD\xF2\x03\xAF\xF9"
-     "\xF6\x03"
-     "\xAE\x01\xF4\x03\xA1\x01\x04\x00\xE0\xD9\x00\x80\xA3\x79\x06\x00\xA2\xCD"
-     "\x06\x00"
-     "\xA5\xD1\xFA\x03\xA4\x01\xF8\x03\xA7\x01\xFC\x03\xA6\xC1\xFA\x03\x82\x55"
-     "\x1B\x00"
-     "\x83\x81\x1B\x00\xE4\xD9\x00\x80\xE8\xD9\x00\x80\x86\x15\x1F\x00\x87\xC9"
-     "\x1F\x00"
-     "\x84\xF1\x1B\x00\x85\x1D\x1F\x00\x8A\xF9\x10\x00\x8B\x21\x13\x00\xBE\x24"
-     "\x05\x00"
-     "\xF0\xD9\x00\x80\x8E\x31\x17\x00\x8F\x01\x17\x00\x8C\x99\x13\x00\x8D\xD1"
-     "\x14\x00"
-     "\x92\xB1\x08\x00\x93\x41\x0B\x00\x86\x68\x04\x00\x87\x7C\x05\x00\x96\x55"
-     "\x0F\x00"
-     "\x97\x45\x0F\x00\x94\x01\x0B\x00\x95\xD9\x0C\x00\x9A\x49\x0E\x00\x9B\xF1"
-     "\x00\x00"
-     "\xF4\xD9\x00\x80\xF8\xD9\x00\x80\x81\x79\x00\x00\x80\x91\x01\x00\x9C\x79"
-     "\x03\x00"
-     "\x82\x71\x00\x00\xFC\xD9\x00\x80\xE1\x74\x0E\x00\x84\x80\x05\x00\xE3\x98"
-     "\x00\x00"
-     "\x00\xDA\x00\x80\x9C\xD8\x00\x80\x04\xDA\x00\x80\x08\xDA\x00\x80\x0C\xDA"
-     "\x00\x80"
-     "\x10\xDA\x00\x80\x14\xDA\x00\x80\x18\xDA\x00\x80\x1C\xDA\x00\x80\x20\xDA"
-     "\x00\x80"
-     "\x24\xDA\x00\x80\xEF\xF4\x0E\x00\xB3\x89\x02\x00\x28\xDA\x00\x80\x2C\xDA"
-     "\x00\x80"
-     "\x30\xDA\x00\x80\x34\xDA\x00\x80\xB6\x8D\x02\x00\xB5\x8D\x02\x00\x38\xDA"
-     "\x00\x80"
-     "\xBB\x45\x02\x00\xBA\x45\x02\x00\x3C\xDA\x00\x80\x40\xDA\x00\x80\xBF\x45"
-     "\x02\x00"
-     "\xBE\x45\x02\x00\xBD\x55\x02\x00\xBC\x55\x02\x00\xA8\x1D\x05\x00\xA9\x45"
-     "\x05\x00"
-     "\xAA\x41\x05\x00\xAB\x41\x05\x00\xAC\x45\x05\x00\xAD\x4D\x05\x00\xAE\x71"
-     "\x05\x00"
-     "\xAF\x71\x05\x00\xEC\xD9\x00\x80\x44\xDA\x00\x80\x48\xDA\x00\x80\x4C\xDA"
-     "\x00\x80"
-     "\x80\x0D\x00\x00\x81\xB1\x00\x00\x82\xB1\x00\x00\x50\xDA\x00\x80\xB8\x45"
-     "\x06\x00"
-     "\xB9\x4D\x06\x00\xBA\x45\x06\x00\xBB\x5D\x06\x00\xBC\x4D\x06\x00\xBD\x7D"
-     "\x06\x00"
-     "\xBE\xDD\x06\x00\xBF\xE1\x06\x00\xB0\xC1\x05\x00\xB1\xC1\x05\x00\xB2\xC1"
-     "\x05\x00"
-     "\xB3\xC1\x05\x00\xB4\xC1\x05\x00\xB5\xC1\x05\x00\xB6\xC1\x05\x00\xB7\xC1"
-     "\x05\x00"
-     "\xA3\xC9\x05\x00\x54\xDA\x00\x80\x58\xDA\x00\x80\xBE\x00\x03\x00\x5C\xDA"
-     "\x00\x80"
-     "\xA6\xCD\x05\x00\xA5\xCD\x05\x00\x60\xDA\x00\x80\xAB\x05\x05\x00\xAA\x05"
-     "\x05\x00"
-     "\x86\x08\x00\x00\x87\xDC\x03\x00\xAF\x05\x05\x00\xAE\x05\x05\x00\xAD\x15"
-     "\x05\x00"
-     "\xAC\x15\x05\x00\x64\xDA\x00\x80\xB3\xA9\x07\x00\x68\xDA\x00\x80\x6C\xDA"
-     "\x00\x80"
-     "\xB6\x49\x07\x00\x70\xDA\x00\x80\x74\xDA\x00\x80\xB5\x4D\x07\x00\xBA\x45"
-     "\x07\x00"
-     "\xBB\x4D\x07\x00\x78\xDA\x00\x80\x7C\xDA\x00\x80\xBE\x45\x07\x00\xBF\x4D"
-     "\x07\x00"
-     "\xBC\x55\x07\x00\xBD\x4D\x07\x00\xA3\x6D\x06\x00\x80\xDA\x00\x80\x84\xDA"
-     "\x00\x80"
-     "\x88\xDA\x00\x80\x8C\xDA\x00\x80\xA6\x8D\x06\x00\xA5\x89\x06\x00\x90\xDA"
-     "\x00\x80"
-     "\xAB\x89\x06\x00\xAA\x81\x06\x00\x94\xDA\x00\x80\x98\xDA\x00\x80\xAF\x89"
-     "\x06\x00"
-     "\xAE\x81\x06\x00\xAD\x89\x06\x00\xAC\x91\x06\x00\x9C\xDA\x00\x80\xA0\xDA"
-     "\x00\x80"
-     "\xA4\xDA\x00\x80\xA8\xDA\x00\x80\xAC\xDA\x00\x80\xB0\xDA\x00\x80\xB4\xDA"
-     "\x00\x80"
-     "\xEF\x8C\x04\x00\xB8\xDA\x00\x80\xE1\x54\x06\x00\xBC\xDA\x00\x80\xE3\xF0"
-     "\x05\x00"
-     "\x80\x0D\x00\x00\x81\xD5\x00\x00\x82\xDD\x00\x00\xC0\xDA\x00\x80\xA8\xAD"
-     "\x06\x00"
-     "\xA9\xED\x06\x00\xAA\xE5\x06\x00\xAB\x0D\x06\x00\xAC\x01\x06\x00\xAD\x0D"
-     "\x06\x00"
-     "\xAE\x05\x06\x00\xAF\x3D\x06\x00\xC4\xDA\x00\x80\xC8\xDA\x00\x80\x84\xE4"
-     "\x01\x00"
-     "\xBE\xE8\x01\x00\x86\x10\xFF\x03\x87\x94\x00\x00\xD0\xDA\x00\x80\xD4\xDA"
-     "\x00\x80"
-     "\xB8\x0D\x06\x00\xB9\x15\x06\x00\xBA\x15\x06\x00\xBB\x25\x06\x00\xBC\x3D"
-     "\x06\x00"
-     "\xBD\xF1\x01\x00\xBE\xF1\x01\x00\xBF\xF1\x01\x00\xB0\x45\x06\x00\xB1\x4D"
-     "\x06\x00"
-     "\xB2\x45\x06\x00\xB3\x5D\x06\x00\xB4\x45\x06\x00\xB5\x4D\x06\x00\xB6\x3D"
-     "\x06\x00"
-     "\xB7\x35\x06\x00\xA8\x9D\x02\x00\xA9\xC5\x02\x00\xAA\xC1\x02\x00\xAB\xC1"
-     "\x02\x00"
-     "\xAC\xC5\x02\x00\xAD\xCD\x02\x00\xAE\xF1\x02\x00\xAF\xF1\x02\x00\xD8\xDA"
-     "\x00\x80"
-     "\xDC\xDA\x00\x80\xE0\xDA\x00\x80\xE4\xDA\x00\x80\x80\x39\xFE\x01\xE8\xDA"
-     "\x00\x80"
-     "\xEC\xDA\x00\x80\xF0\xDA\x00\x80\xB8\x05\x03\x00\xB9\x0D\x03\x00\xBA\x05"
-     "\x03\x00"
-     "\xBB\x1D\x03\x00\xBC\x05\x03\x00\xBD\x0D\x03\x00\xBE\x39\x03\x00\xBF\x35"
-     "\x03\x00"
-     "\xB0\x65\x03\x00\xB1\x6D\x03\x00\xB2\x65\x03\x00\xB3\x7D\x03\x00\xB4\x6D"
-     "\x03\x00"
-     "\xB5\x5D\x03\x00\xB6\x55\x03\x00\xB7\x3D\x03\x00\xB3\x49\x02\x00\xF4\xDA"
-     "\x00\x80"
-     "\xF8\xDA\x00\x80\xFC\xDA\x00\x80\x00\xDB\x00\x80\xB6\x4D\x02\x00\xB5\x4D"
-     "\x02\x00"
-     "\x04\xDB\x00\x80\xBB\xA9\x03\x00\xBA\xA1\x03\x00\x08\xDB\x00\x80\x0C\xDB"
-     "\x00\x80"
-     "\xBF\x99\x03\x00\xBE\xA9\x03\x00\xBD\xA9\x03\x00\xBC\xB1\x03\x00\x10\xDB"
-     "\x00\x80"
-     "\x14\xDB\x00\x80\xA3\x01\x02\x00\x18\xDB\x00\x80\xA5\x05\x02\x00\x1C\xDB"
-     "\x00\x80"
-     "\x20\xDB\x00\x80\xA6\x05\x02\x00\x7C\xD9\x00\x80\x24\xDB\x00\x80\xAB\xE1"
-     "\x03\x00"
-     "\xAA\xE9\x03\x00\xAD\xE1\x03\x00\xAC\xF9\x03\x00\xAF\xD1\x03\x00\xAE\xE1"
-     "\x03\x00"
-     "\x28\xDB\x00\x80\x2C\xDB\x00\x80\x30\xDB\x00\x80\x34\xDB\x00\x80\x38\xDB"
-     "\x00\x80"
-     "\xE1\xA4\x01\x00\xBE\x44\x02\x00\xE3\xA4\x00\x00\x80\x29\x00\x00\x81\x29"
-     "\x00\x00"
-     "\x82\x39\x00\x00\xEF\x78\x00\x00\x40\xDB\x00\x80\x84\xB8\x02\x00\x44\xDB"
-     "\x00\x80"
-     "\x48\xDB\x00\x80\xB3\x61\x01\x00\x4C\xDB\x00\x80\x50\xDB\x00\x80\x54\xDB"
-     "\x00\x80"
-     "\x58\xDB\x00\x80\xB6\x79\x01\x00\xB5\x71\x01\x00\x5C\xDB\x00\x80\xBB\xE5"
-     "\x01\x00"
-     "\xBA\x59\x01\x00\x86\x48\xFD\x03\x87\x40\xFD\x03\xBF\xED\x01\x00\xBE\xE5"
-     "\x01\x00"
-     "\xBD\xF5\x01\x00\xBC\xF5\x01\x00\x60\xDB\x00\x80\x64\xDB\x00\x80\x68\xDB"
-     "\x00\x80"
-     "\x84\x40\xFC\x03\x6C\xDB\x00\x80\x70\xDB\x00\x80\x74\xDB\x00\x80\xEF\xE8"
-     "\x01\x00"
-     "\x78\xDB\x00\x80\xE1\xF4\x06\x00\xAC\xB4\xF3\x03\xE3\x34\x06\x00\x7C\xDB"
-     "\x00\x80"
-     "\x80\xDB\x00\x80\x84\xB0\xFC\x03\x84\xDB\x00\x80\x88\xDB\x00\x80\xA6\xE9"
-     "\x01\x00"
-     "\xA5\xE1\x01\x00\x8C\xDB\x00\x80\xA3\xF1\x01\x00\x90\xDB\x00\x80\x94\xDB"
-     "\x00\x80"
-     "\x98\xDB\x00\x80\xAF\x7D\x01\x00\xAE\x75\x01\x00\xAD\x65\x01\x00\xAC\x65"
-     "\x01\x00"
-     "\xAB\x75\x01\x00\xAA\xC9\x01\x00\x9C\xDB\x00\x80\xA0\xDB\x00\x80\xA8\xB9"
-     "\xFD\x03"
-     "\xA9\xB9\xFD\x03\xAA\x29\xFE\x03\xAB\x29\xFE\x03\xAC\x39\xFE\x03\xAD\x39"
-     "\xFE\x03"
-     "\xAE\x2D\xFE\x03\xAF\x25\xFE\x03\x3C\xDB\x00\x80\x82\x11\x00\x00\x81\x19"
-     "\x00\x00"
-     "\x80\xD1\xFF\x03\xA4\xDB\x00\x80\xA8\xDB\x00\x80\xBE\x34\x03\x00\xAC\xDB"
-     "\x00\x80"
-     "\xB8\x81\xFE\x03\xB9\x81\xFE\x03\xBA\x85\xFE\x03\xBB\x99\xFE\x03\xBC\x89"
-     "\xFE\x03"
-     "\xBD\xB1\xFE\x03\xBE\x59\xFF\x03\xBF\x59\xFF\x03\xB0\x5D\xFE\x03\xB1\xE1"
-     "\xFE\x03"
-     "\xB2\xE1\xFE\x03\xB3\xE1\xFE\x03\xB4\xE1\xFE\x03\xB5\xE9\xFE\x03\xB6\xD1"
-     "\xFE\x03"
-     "\xB7\xD1\xFE\x03\xA8\xB1\xFF\x03\xA9\xBD\xFF\x03\xAA\xB5\xFF\x03\xAB\xC9"
-     "\xFF\x03"
-     "\xAC\xD9\xFF\x03\xAD\xD9\xFF\x03\xAE\xC9\xFF\x03\xAF\xC1\xFF\x03\x86\x00"
-     "\x00\x00"
-     "\x87\x28\x03\x00\xB0\xDB\x00\x80\xB4\xDB\x00\x80\xB8\xDB\x00\x80\xBC\xDB"
-     "\x00\x80"
-     "\xC0\xDB\x00\x80\xC4\xDB\x00\x80\xB8\x75\xFF\x03\xB9\x79\xFF\x03\xBA\x09"
-     "\xFF\x03"
-     "\xBB\x01\xFF\x03\xBC\x05\xFF\x03\xBD\x0D\xFF\x03\xBE\x05\xFF\x03\xBF\x25"
-     "\xFF\x03"
-     "\xB0\x61\xFF\x03\xB1\x61\xFF\x03\xB2\x61\xFF\x03\xB3\x7D\xFF\x03\xB4\x65"
-     "\xFF\x03"
-     "\xB5\x6D\xFF\x03\xB6\x59\xFF\x03\xB7\x4D\xFF\x03\xB3\x3D\xFE\x03\xC8\xDB"
-     "\x00\x80"
-     "\xCC\xDB\x00\x80\xD0\xDB\x00\x80\xD4\xDB\x00\x80\xB6\x59\xFE\x03\xB5\x51"
-     "\xFE\x03"
-     "\xD8\xDB\x00\x80\xBB\xA5\xFE\x03\xBA\xA5\xFE\x03\xDC\xDB\x00\x80\xE0\xDB"
-     "\x00\x80"
-     "\xBF\xA9\xFE\x03\xBE\xA1\xFE\x03\xBD\xA9\xFE\x03\xBC\xB5\xFE\x03\xE4\xDB"
-     "\x00\x80"
-     "\xE8\xDB\x00\x80\xEF\xA8\xFC\x03\xEC\xDB\x00\x80\xF0\xDB\x00\x80\xF4\xDB"
-     "\x00\x80"
-     "\xF8\xDB\x00\x80\xFC\xDB\x00\x80\x00\xDC\x00\x80\x04\xDC\x00\x80\x08\xDC"
-     "\x00\x80"
-     "\x0C\xDC\x00\x80\xE1\xD8\xFE\x03\x10\xDC\x00\x80\xE3\x5C\xFC\x03\x14\xDC"
-     "\x00\x80"
-     "\xA3\xBD\xFE\x03\x82\x15\x00\x00\x81\xF5\x00\x00\x80\xF9\x00\x00\x18\xDC"
-     "\x00\x80"
-     "\xA6\xD9\xFE\x03\xA5\xD1\xFE\x03\xBE\x90\x01\x00\xAB\x25\xFE\x03\xAA\x25"
-     "\xFE\x03"
-     "\x86\x08\x00\x00\x87\x70\x01\x00\xAF\x29\xFE\x03\xAE\x21\xFE\x03\xAD\x29"
-     "\xFE\x03"
-     "\xAC\x35\xFE\x03\x20\xDC\x00\x80\xB3\x75\x01\x00\x24\xDC\x00\x80\x28\xDC"
-     "\x00\x80"
-     "\xB6\x6D\x01\x00\x2C\xDC\x00\x80\x30\xDC\x00\x80\xB5\x65\x01\x00\xBA\xED"
-     "\x00\x00"
-     "\xBB\xF5\x00\x00\x34\xDC\x00\x80\x38\xDC\x00\x80\xBE\xDD\x00\x00\xBF\xB5"
-     "\x00\x00"
-     "\xBC\xE5\x00\x00\xBD\xD5\x00\x00\xA8\xA9\x02\x00\xA9\xA9\x02\x00\xAA\xB9"
-     "\x02\x00"
-     "\xAB\xB9\x02\x00\xAC\xA9\x02\x00\xAD\xA9\x02\x00\xAE\x1D\x03\x00\xAF\x15"
-     "\x03\x00"
-     "\xBE\xAC\x04\x00\x3C\xDC\x00\x80\x40\xDC\x00\x80\xBE\xA0\x05\x00\x44\xDC"
-     "\x00\x80"
-     "\x48\xDC\x00\x80\x4C\xDC\x00\x80\x50\xDC\x00\x80\xB8\x11\x03\x00\xB9\x11"
-     "\x03\x00"
-     "\xBA\x11\x03\x00\xBB\x11\x03\x00\xBC\x31\x03\x00\xBD\x31\x03\x00\xBE\x31"
-     "\x03\x00"
-     "\xBF\x31\x03\x00\xB0\x6D\x03\x00\xB1\x75\x03\x00\xB2\x75\x03\x00\xB3\x45"
-     "\x03\x00"
-     "\xB4\x5D\x03\x00\xB5\x35\x03\x00\xB6\x3D\x03\x00\xB7\x31\x03\x00\xA3\x31"
-     "\x02\x00"
-     "\x54\xDC\x00\x80\x58\xDC\x00\x80\x5C\xDC\x00\x80\x60\xDC\x00\x80\xA6\x29"
-     "\x02\x00"
-     "\xA5\x21\x02\x00\x64\xDC\x00\x80\xAB\xB1\x03\x00\xAA\xA9\x03\x00\x68\xDC"
-     "\x00\x80"
-     "\x6C\xDC\x00\x80\xAF\xF1\x03\x00\xAE\x99\x03\x00\xAD\x91\x03\x00\xAC\xA1"
-     "\x03\x00"
-     "\x84\xC0\x03\x00\x70\xDC\x00\x80\x74\xDC\x00\x80\x78\xDC\x00\x80\x7C\xDC"
-     "\x00\x80"
-     "\x82\x3D\x00\x00\x81\x3D\x00\x00\x80\x3D\x00\x00\x80\xDC\x00\x80\x84\xDC"
-     "\x00\x80"
-     "\x88\xDC\x00\x80\x90\xDC\x00\x80\x94\xDC\x00\x80\xCC\xDA\x00\x80\x86\xD8"
-     "\x04\x00"
-     "\x87\xBC\x04\x00\xA8\xC1\x02\x00\xA9\xCD\x02\x00\xAA\xC5\x02\x00\xAB\xD9"
-     "\x02\x00"
-     "\xAC\xC9\x02\x00\xAD\xFD\x02\x00\xAE\xF5\x02\x00\xAF\x39\x01\x00\x98\xDC"
-     "\x00\x80"
-     "\x9C\xDC\x00\x80\xA0\xDC\x00\x80\xA4\xDC\x00\x80\xA8\xDC\x00\x80\xAC\xDC"
-     "\x00\x80"
-     "\xB0\xDC\x00\x80\xB4\xDC\x00\x80\xB8\xD1\x01\x00\xB9\xD9\x01\x00\xBA\xE1"
-     "\x01\x00"
-     "\xBB\xE1\x01\x00\xBC\x95\x01\x00\xBD\x9D\x01\x00\xBE\x95\x01\x00\xBF\x8D"
-     "\x01\x00"
-     "\xB0\x49\x01\x00\xB1\x55\x01\x00\xB2\x5D\x01\x00\xB3\x55\x01\x00\xB4\x4D"
-     "\x01\x00"
-     "\xB5\xF1\x01\x00\xB6\xF1\x01\x00\xB7\xF1\x01\x00\xEF\xE4\x00\x00\xB8\xDC"
-     "\x00\x80"
-     "\xBC\xDC\x00\x80\x84\x0C\x05\x00\xC0\xDC\x00\x80\xC4\xDC\x00\x80\xC8\xDC"
-     "\x00\x80"
-     "\xEF\x5C\x01\x00\xCC\xDC\x00\x80\xE1\xD4\xFE\x03\xD0\xDC\x00\x80\xE3\x30"
-     "\xFE\x03"
-     "\xE3\x18\xFE\x03\xD4\xDC\x00\x80\xE1\x00\x01\x00\xD8\xDC\x00\x80\xDC\xDC"
-     "\x00\x80"
-     "\xE0\xDC\x00\x80\xE4\xDC\x00\x80\xB3\x01\x02\x00\xE8\xDC\x00\x80\xB5\x0D"
-     "\x02\x00"
-     "\xB6\x31\x02\x00\xEC\xDC\x00\x80\xF0\xDC\x00\x80\xF4\xDC\x00\x80\xBA\xC1"
-     "\x01\x00"
-     "\xBB\xDD\x01\x00\xBC\xC5\x01\x00\xBD\xCD\x01\x00\xBE\xC5\x01\x00\xBF\x79"
-     "\x01\x00"
-     "\xA3\xCD\x05\x00\x8C\xDC\x00\x80\xF8\xDC\x00\x80\x00\xDD\x00\x80\xFC\xDC"
-     "\x00\x80"
-     "\xA6\xFD\x05\x00\xA5\xC1\x05\x00\x04\xDD\x00\x80\xAB\x11\x06\x00\xAA\x0D"
-     "\x06\x00"
-     "\x86\x08\x03\x00\x87\x7C\x03\x00\xAF\xB5\x06\x00\xAE\x09\x06\x00\xAD\x01"
-     "\x06\x00"
-     "\xAC\x09\x06\x00\x80\xA1\x07\x00\x81\xA1\x07\x00\x82\x41\x00\x00\xB3\x6D"
-     "\x07\x00"
-     "\x08\xDD\x00\x80\xB5\x7D\x07\x00\xB6\x75\x07\x00\x0C\xDD\x00\x80\x10\xDD"
-     "\x00\x80"
-     "\x14\xDD\x00\x80\xBA\x51\x07\x00\xBB\x51\x07\x00\xBC\xF5\x07\x00\xBD\xFD"
-     "\x07\x00"
-     "\xBE\xE9\x07\x00\xBF\xE9\x07\x00\xA8\x6D\x06\x00\xA9\x75\x06\x00\xAA\xA9"
-     "\x07\x00"
-     "\xAB\xA9\x07\x00\xAC\xB9\x07\x00\xAD\xB9\x07\x00\xAE\xA9\x07\x00\xAF\xA9"
-     "\x07\x00"
-     "\x18\xDD\x00\x80\x1C\xDD\x00\x80\x20\xDD\x00\x80\x24\xDD\x00\x80\x28\xDD"
-     "\x00\x80"
-     "\x2C\xDD\x00\x80\x30\xDD\x00\x80\x34\xDD\x00\x80\xB8\x69\x07\x00\xB9\x71"
-     "\x07\x00"
-     "\xBA\x75\x07\x00\xBB\x0D\x07\x00\xBC\x15\x07\x00\xBD\x1D\x07\x00\xBE\x09"
-     "\x07\x00"
-     "\xBF\x01\x07\x00\xB0\xD9\x07\x00\xB1\xD9\x07\x00\xB2\x6D\x07\x00\xB3\x65"
-     "\x07\x00"
-     "\xB4\x61\x07\x00\xB5\x61\x07\x00\xB6\x61\x07\x00\xB7\x5D\x07\x00\xA3\x21"
-     "\x06\x00"
-     "\x38\xDD\x00\x80\x3C\xDD\x00\x80\x40\xDD\x00\x80\x44\xDD\x00\x80\xA6\x39"
-     "\x06\x00"
-     "\xA5\x31\x06\x00\x48\xDD\x00\x80\xAB\x1D\x06\x00\xAA\x1D\x06\x00\x4C\xDD"
-     "\x00\x80"
-     "\x50\xDD\x00\x80\xAF\xA5\x06\x00\xAE\xA5\x06\x00\xAD\xB1\x06\x00\xAC\xB9"
-     "\x06\x00"
-     "\x54\xDD\x00\x80\xEF\x9C\x05\x00\x58\xDD\x00\x80\x5C\xDD\x00\x80\x80\x0D"
-     "\x00\x00"
-     "\x81\x35\x00\x00\x82\x3D\x00\x00\x60\xDD\x00\x80\x64\xDD\x00\x80\x68\xDD"
-     "\x00\x80"
-     "\x84\xA4\x01\x00\xBE\xA4\x01\x00\x6C\xDD\x00\x80\xE3\xA4\x05\x00\x70\xDD"
-     "\x00\x80"
-     "\xE1\xCC\x06\x00\xA8\x8D\x01\x00\xA9\x99\x01\x00\xAA\xD9\x01\x00\xAB\xD1"
-     "\x01\x00"
-     "\xAC\xF1\x01\x00\xAD\xF1\x01\x00\xAE\x39\x01\x00\xAF\x39\x01\x00\x86\x00"
-     "\x00\x00"
-     "\x87\x64\x01\x00\x74\xDD\x00\x80\x78\xDD\x00\x80\x7C\xDD\x00\x80\x80\xDD"
-     "\x00\x80"
-     "\x84\xDD\x00\x80\x88\xDD\x00\x80\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5"
-     "\x00\x00"
-     "\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x91\x00\x00\xBF\x91"
-     "\x00\x00"
-     "\xB0\x49\x01\x00\xB1\x49\x01\x00\xB2\x59\x01\x00\xB3\x59\x01\x00\xB4\x49"
-     "\x01\x00"
-     "\xB5\x49\x01\x00\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xB3\x51\x02\x00\x8C\xDD"
-     "\x00\x80"
-     "\x90\xDD\x00\x80\x94\xDD\x00\x80\x98\xDD\x00\x80\xB6\x71\x02\x00\xB5\x79"
-     "\x02\x00"
-     "\x9C\xDD\x00\x80\xBB\x8D\x02\x00\xBA\x8D\x02\x00\xA0\xDD\x00\x80\xA4\xDD"
-     "\x00\x80"
-     "\xBF\x8D\x02\x00\xBE\x8D\x02\x00\xBD\x8D\x02\x00\xBC\x8D\x02\x00\x85\x6C"
-     "\x12\x00"
-     "\xA3\x15\x02\x00\xA8\xDD\x00\x80\xAC\xDD\x00\x80\xA6\x35\x02\x00\xB0\xDD"
-     "\x00\x80"
-     "\xB4\xDD\x00\x80\xA5\x3D\x02\x00\xAA\xC9\x02\x00\xAB\xC9\x02\x00\xB8\xDD"
-     "\x00\x80"
-     "\xBC\xDD\x00\x80\xAE\xC9\x02\x00\xAF\xC9\x02\x00\xAC\xC9\x02\x00\xAD\xC9"
-     "\x02\x00"
-     "\x84\x80\x03\x00\xC0\xDD\x00\x80\xC4\xDD\x00\x80\xBE\x00\x0C\x00\xC8\xDD"
-     "\x00\x80"
-     "\xCC\xDD\x00\x80\xD0\xDD\x00\x80\xD4\xDD\x00\x80\xD8\xDD\x00\x80\xE0\xDD"
-     "\x00\x80"
-     "\xE4\xDD\x00\x80\xE8\xDD\x00\x80\xEC\xDD\x00\x80\xF0\xDD\x00\x80\xF4\xDD"
-     "\x00\x80"
-     "\xF8\xDD\x00\x80\x80\x59\x00\x00\x81\x59\x00\x00\x82\x69\x00\x00\xEF\xB8"
-     "\x02\x00"
-     "\xFC\xDD\x00\x80\x00\xDE\x00\x80\x87\x18\x03\x00\x86\xFC\x0C\x00\x04\xDE"
-     "\x00\x80"
-     "\x08\xDE\x00\x80\x0C\xDE\x00\x80\x10\xDE\x00\x80\x14\xDE\x00\x80\xE1\x4C"
-     "\x01\x00"
-     "\x1C\xDC\x00\x80\xE3\x6C\x02\x00\xA8\x79\x02\x00\xA9\xAD\x02\x00\xAA\xA5"
-     "\x02\x00"
-     "\xAB\xBD\x02\x00\xAC\xA5\x02\x00\xAD\xAD\x02\x00\xAE\xA5\x02\x00\xAF\x19"
-     "\x01\x00"
-     "\x84\xA8\x0D\x00\x18\xDE\x00\x80\x1C\xDE\x00\x80\x20\xDE\x00\x80\x24\xDE"
-     "\x00\x80"
-     "\x28\xDE\x00\x80\x2C\xDE\x00\x80\x30\xDE\x00\x80\xB8\x15\x01\x00\xB9\x1D"
-     "\x01\x00"
-     "\xBA\x21\x01\x00\xBB\x21\x01\x00\xBC\xD5\x01\x00\xBD\xDD\x01\x00\xBE\xC9"
-     "\x01\x00"
-     "\xBF\xC9\x01\x00\xB0\x69\x01\x00\xB1\x69\x01\x00\xB2\x79\x01\x00\xB3\x71"
-     "\x01\x00"
-     "\xB4\x59\x01\x00\xB5\x59\x01\x00\xB6\x31\x01\x00\xB7\x31\x01\x00\xE1\x60"
-     "\x07\x00"
-     "\x34\xDE\x00\x80\xE3\x14\x06\x00\xBE\x00\x0C\x00\x38\xDE\x00\x80\xE1\xA4"
-     "\x01\x00"
-     "\x3C\xDE\x00\x80\xE3\xE4\x00\x00\x40\xDE\x00\x80\x44\xDE\x00\x80\x48\xDE"
-     "\x00\x80"
-     "\xEF\x38\x00\x00\x50\xDE\x00\x80\x54\xDE\x00\x80\xEF\x6C\x07\x00\x58\xDE"
-     "\x00\x80"
-     "\x80\x61\x00\x00\x81\x61\x00\x00\x82\x61\x00\x00\xB3\x61\x02\x00\x5C\xDE"
-     "\x00\x80"
-     "\xB5\x61\x02\x00\xB6\x61\x02\x00\x60\xDE\x00\x80\x86\xC0\x0C\x00\x87\xE8"
-     "\x0C\x00"
-     "\xBA\xE1\x01\x00\xBB\xE1\x01\x00\xBC\xE1\x01\x00\xBD\xE9\x01\x00\xBE\xD1"
-     "\x01\x00"
-     "\xBF\xD1\x01\x00\xA3\xAD\x0D\x00\xDC\xDD\x00\x80\x4C\xDE\x00\x80\x64\xDE"
-     "\x00\x80"
-     "\x68\xDE\x00\x80\xA6\xAD\x0D\x00\xA5\xAD\x0D\x00\x6C\xDE\x00\x80\xAB\x2D"
-     "\x0E\x00"
-     "\xAA\x2D\x0E\x00\x70\xDE\x00\x80\x74\xDE\x00\x80\xAF\x1D\x0E\x00\xAE\x1D"
-     "\x0E\x00"
-     "\xAD\x25\x0E\x00\xAC\x2D\x0E\x00\x78\xDE\x00\x80\xB3\xE9\x0E\x00\x7C\xDE"
-     "\x00\x80"
-     "\x80\xDE\x00\x80\xB6\xE9\x0E\x00\x84\xDE\x00\x80\x88\xDE\x00\x80\xB5\xE9"
-     "\x0E\x00"
-     "\xBA\xB5\x0E\x00\xBB\x85\x0E\x00\x8C\xDE\x00\x80\x90\xDE\x00\x80\xBE\x9D"
-     "\x0E\x00"
-     "\xBF\x85\x0E\x00\xBC\x9D\x0E\x00\xBD\x95\x0E\x00\xA8\x25\x0E\x00\xA9\x2D"
-     "\x0E\x00"
-     "\xAA\x25\x0E\x00\xAB\x3D\x0E\x00\xAC\x25\x0E\x00\xAD\x2D\x0E\x00\xAE\x25"
-     "\x0E\x00"
-     "\xAF\x4D\x0E\x00\x94\xDE\x00\x80\x98\xDE\x00\x80\x9C\xDE\x00\x80\xA0\xDE"
-     "\x00\x80"
-     "\xA4\xDE\x00\x80\xA8\xDE\x00\x80\xAC\xDE\x00\x80\xB0\xDE\x00\x80\xB8\xED"
-     "\x0F\x00"
-     "\xB9\xF5\x0F\x00\xBA\xFD\x0F\x00\xBB\xF5\x0F\x00\xBC\xED\x0F\x00\xBD\x51"
-     "\x0F\x00"
-     "\xBE\x49\x0F\x00\xBF\x45\x0F\x00\xB0\x35\x0E\x00\xB1\x05\x0E\x00\xB2\x0D"
-     "\x0E\x00"
-     "\xB3\x05\x0E\x00\xB4\x1D\x0E\x00\xB5\x05\x0E\x00\xB6\xDD\x0F\x00\xB7\xD5"
-     "\x0F\x00"
-     "\xA3\xAD\x0F\x00\x82\x19\x00\x00\x81\x19\x00\x00\x80\xB1\x01\x00\xB4\xDE"
-     "\x00\x80"
-     "\xA6\xAD\x0F\x00\xA5\xAD\x0F\x00\xB8\xDE\x00\x80\xAB\xC1\x0F\x00\xAA\xF1"
-     "\x0F\x00"
-     "\x84\xA0\x02\x00\xBE\xA0\x01\x00\xAF\xC1\x0F\x00\xAE\xD9\x0F\x00\xAD\xD1"
-     "\x0F\x00"
-     "\xAC\xD9\x0F\x00\xBC\xDE\x00\x80\xB3\xF9\x0E\x00\x86\xE8\x00\x00\x87\x2C"
-     "\x01\x00"
-     "\xB6\x5D\x01\x00\xC0\xDE\x00\x80\xC4\xDE\x00\x80\xB5\x81\x0E\x00\xBA\x49"
-     "\x01\x00"
-     "\xBB\x49\x01\x00\xC8\xDE\x00\x80\xCC\xDE\x00\x80\xBE\x3D\x01\x00\xBF\x21"
-     "\x01\x00"
-     "\xBC\x39\x01\x00\xBD\x35\x01\x00\xA8\x31\x0E\x00\xA9\x31\x0E\x00\xAA\x35"
-     "\x0E\x00"
-     "\xAB\x5D\x0E\x00\xAC\x4D\x0E\x00\xAD\x99\x01\x00\xAE\x89\x01\x00\xAF\x81"
-     "\x01\x00"
-     "\x84\xE4\x01\x00\xD0\xDE\x00\x80\xD4\xDE\x00\x80\xD8\xDE\x00\x80\xDC\xDE"
-     "\x00\x80"
-     "\xE0\xDE\x00\x80\xE4\xDE\x00\x80\xE8\xDE\x00\x80\xB8\xAD\x01\x00\xB9\x65"
-     "\x01\x00"
-     "\xBA\x6D\x01\x00\xBB\x65\x01\x00\xBC\x7D\x01\x00\xBD\x65\x01\x00\xBE\x6D"
-     "\x01\x00"
-     "\xBF\x65\x01\x00\xB0\xC9\x01\x00\xB1\xC9\x01\x00\xB2\xA9\x01\x00\xB3\xA5"
-     "\x01\x00"
-     "\xB4\xBD\x01\x00\xB5\xA1\x01\x00\xB6\xA1\x01\x00\xB7\x95\x01\x00\xA3\xBD"
-     "\x0D\x00"
-     "\xEC\xDE\x00\x80\xF0\xDE\x00\x80\xF4\xDE\x00\x80\xF8\xDE\x00\x80\xA6\x19"
-     "\x02\x00"
-     "\xA5\xC5\x0D\x00\xFC\xDE\x00\x80\xAB\x0D\x02\x00\xAA\x0D\x02\x00\x00\xDF"
-     "\x00\x80"
-     "\x04\xDF\x00\x80\xAF\x65\x02\x00\xAE\x79\x02\x00\xAD\x71\x02\x00\xAC\x7D"
-     "\x02\x00"
-     "\x08\xDF\x00\x80\x0C\xDF\x00\x80\x10\xDF\x00\x80\x14\xDF\x00\x80\x18\xDF"
-     "\x00\x80"
-     "\x1C\xDF\x00\x80\x20\xDF\x00\x80\x24\xDF\x00\x80\x80\x39\x00\x00\x81\x39"
-     "\x00\x00"
-     "\x82\x05\x00\x00\x28\xDF\x00\x80\xBE\x3C\x03\x00\x30\xDF\x00\x80\x87\x44"
-     "\x03\x00"
-     "\x86\xDC\x04\x00\x34\xDF\x00\x80\x38\xDF\x00\x80\xB3\xE1\x03\x00\x84\xC0"
-     "\x04\x00"
-     "\xB5\x35\x03\x00\x3C\xDF\x00\x80\x40\xDF\x00\x80\xB6\x35\x03\x00\x44\xDF"
-     "\x00\x80"
-     "\x48\xDF\x00\x80\xBB\x11\x03\x00\xBA\x19\x03\x00\xBD\xB5\x00\x00\xBC\x09"
-     "\x03\x00"
-     "\xBF\xB5\x00\x00\xBE\xB5\x00\x00\x4C\xDF\x00\x80\xE1\x94\x01\x00\x50\xDF"
-     "\x00\x80"
-     "\xE3\x7C\x01\x00\x54\xDF\x00\x80\x58\xDF\x00\x80\x5C\xDF\x00\x80\x60\xDF"
-     "\x00\x80"
-     "\x64\xDF\x00\x80\x68\xDF\x00\x80\x6C\xDF\x00\x80\x70\xDF\x00\x80\x74\xDF"
-     "\x00\x80"
-     "\x78\xDF\x00\x80\x7C\xDF\x00\x80\xEF\x74\x02\x00\xA8\x55\x02\x00\xA9\x5D"
-     "\x02\x00"
-     "\xAA\x61\x02\x00\xAB\x61\x02\x00\xAC\xB5\x02\x00\xAD\xBD\x02\x00\xAE\xB5"
-     "\x02\x00"
-     "\xAF\xAD\x02\x00\xBE\x68\x05\x00\x80\xDF\x00\x80\x84\xDF\x00\x80\x88\xDF"
-     "\x00\x80"
-     "\x8C\xDF\x00\x80\x82\xAD\x00\x00\x81\xAD\x00\x00\x80\xAD\x00\x00\xB8\x69"
-     "\x01\x00"
-     "\xB9\x69\x01\x00\xBA\x09\x01\x00\xBB\x09\x01\x00\xBC\x19\x01\x00\xBD\x19"
-     "\x01\x00"
-     "\xBE\x09\x01\x00\xBF\x09\x01\x00\xB0\xD5\x02\x00\xB1\xDD\x02\x00\xB2\xD5"
-     "\x02\x00"
-     "\xB3\x69\x01\x00\xB4\x79\x01\x00\xB5\x79\x01\x00\xB6\x69\x01\x00\xB7\x61"
-     "\x01\x00"
-     "\xE1\x44\x0F\x00\xE3\xBC\x00\x00\xE3\xB8\x0E\x00\xE1\x74\x0E\x00\x90\xDF"
-     "\x00\x80"
-     "\x94\xDF\x00\x80\x98\xDF\x00\x80\xA0\xDF\x00\x80\xA4\xDF\x00\x80\xA8\xDF"
-     "\x00\x80"
-     "\xAC\xDF\x00\x80\xB0\xDF\x00\x80\xBE\x30\x07\x00\xEF\x60\x00\x00\xEF\xCC"
-     "\x0E\x00"
-     "\xB4\xDF\x00\x80\xA3\xA9\x02\x00\xB8\xDF\x00\x80\x86\xE8\x04\x00\x87\x04"
-     "\x05\x00"
-     "\xBC\xDF\x00\x80\xA6\x7D\x02\x00\xA5\x7D\x02\x00\xC0\xDF\x00\x80\xAB\x59"
-     "\x02\x00"
-     "\xAA\x51\x02\x00\xC4\xDF\x00\x80\xC8\xDF\x00\x80\xAF\xFD\x01\x00\xAE\xFD"
-     "\x01\x00"
-     "\xAD\xFD\x01\x00\xAC\x41\x02\x00\xA8\x75\x06\x00\xA9\x85\x06\x00\xAA\x8D"
-     "\x06\x00"
-     "\xAB\x99\x06\x00\xAC\x89\x06\x00\xAD\xB5\x06\x00\xAE\xBD\x06\x00\xAF\xDD"
-     "\x06\x00"
-     "\x9C\xDF\x00\x80\xCC\xDF\x00\x80\xD0\xDF\x00\x80\xD4\xDF\x00\x80\xD8\xDF"
-     "\x00\x80"
-     "\xDC\xDF\x00\x80\xE0\xDF\x00\x80\xE4\xDF\x00\x80\xB8\xAD\x06\x00\xB9\xB5"
-     "\x06\x00"
-     "\xBA\xBD\x06\x00\xBB\xB5\x06\x00\xBC\x75\x07\x00\xBD\x7D\x07\x00\xBE\x75"
-     "\x07\x00"
-     "\xBF\x6D\x07\x00\xB0\xA9\x06\x00\xB1\xB5\x06\x00\xB2\xBD\x06\x00\xB3\xB5"
-     "\x06\x00"
-     "\xB4\xAD\x06\x00\xB5\x95\x06\x00\xB6\x9D\x06\x00\xB7\x95\x06\x00\xB3\x05"
-     "\x06\x00"
-     "\xE8\xDF\x00\x80\xEC\xDF\x00\x80\xF0\xDF\x00\x80\xF4\xDF\x00\x80\xB6\x35"
-     "\x06\x00"
-     "\xB5\x01\x06\x00\xF8\xDF\x00\x80\xBB\x71\x06\x00\xBA\x6D\x06\x00\xFC\xDF"
-     "\x00\x80"
-     "\x00\xE0\x00\x80\xBF\x51\x06\x00\xBE\x69\x06\x00\xBD\x61\x06\x00\xBC\x69"
-     "\x06\x00"
-     "\x82\x19\x00\x00\xA3\x41\x06\x00\x80\x65\x00\x00\x81\x19\x00\x00\xA6\x71"
-     "\x06\x00"
-     "\x04\xE0\x00\x80\x08\xE0\x00\x80\xA5\x45\x06\x00\xAA\x29\x06\x00\xAB\x35"
-     "\x06\x00"
-     "\x84\x24\x03\x00\x0C\xE0\x00\x80\xAE\x2D\x06\x00\xAF\x15\x06\x00\xAC\x2D"
-     "\x06\x00"
-     "\xAD\x25\x06\x00\xB3\xA5\x06\x00\x10\xE0\x00\x80\x86\x48\x01\x00\x87\x0C"
-     "\x01\x00"
-     "\x14\xE0\x00\x80\xB6\xA9\x06\x00\xB5\xB5\x06\x00\x18\xE0\x00\x80\xBB\xC9"
-     "\x06\x00"
-     "\xBA\xF9\x06\x00\x1C\xE0\x00\x80\x20\xE0\x00\x80\xBF\x1D\x01\x00\xBE\x1D"
-     "\x01\x00"
-     "\xBD\x1D\x01\x00\xBC\xD1\x06\x00\x24\xE0\x00\x80\x28\xE0\x00\x80\x2C\xE0"
-     "\x00\x80"
-     "\x30\xE0\x00\x80\xE3\xCC\x05\x00\x34\xE0\x00\x80\xE1\x3C\x05\x00\x38\xE0"
-     "\x00\x80"
-     "\xEF\xB8\x04\x00\x3C\xE0\x00\x80\x40\xE0\x00\x80\x44\xE0\x00\x80\x48\xE0"
-     "\x00\x80"
-     "\x4C\xE0\x00\x80\x50\xE0\x00\x80\x54\xE0\x00\x80\xA8\x1D\x06\x00\xA9\x25"
-     "\x06\x00"
-     "\xAA\x2D\x06\x00\xAB\x25\x06\x00\xAC\x3D\x06\x00\xAD\x29\x06\x00\xAE\x5D"
-     "\x06\x00"
-     "\xAF\x4D\x06\x00\x58\xE0\x00\x80\x5C\xE0\x00\x80\x60\xE0\x00\x80\x64\xE0"
-     "\x00\x80"
-     "\x68\xE0\x00\x80\x6C\xE0\x00\x80\x70\xE0\x00\x80\x74\xE0\x00\x80\xB8\xF9"
-     "\x01\x00"
-     "\xB9\xF9\x01\x00\xBA\x4D\x01\x00\xBB\x45\x01\x00\xBC\x5D\x01\x00\xBD\x45"
-     "\x01\x00"
-     "\xBE\x45\x01\x00\xBF\x75\x01\x00\xB0\x3D\x06\x00\xB1\x0D\x06\x00\xB2\x05"
-     "\x06\x00"
-     "\xB3\xC9\x01\x00\xB4\xD9\x01\x00\xB5\xD9\x01\x00\xB6\xC9\x01\x00\xB7\xC9"
-     "\x01\x00"
-     "\x80\xB9\x03\x00\x81\xB9\x03\x00\x82\x19\x00\x00\xA3\xA9\x05\x00\x78\xE0"
-     "\x00\x80"
-     "\xA5\xB9\x05\x00\xA6\xA5\x05\x00\x84\x84\x02\x00\x80\xE0\x00\x80\xBE\xA8"
-     "\x03\x00"
-     "\xAA\xF5\x05\x00\xAB\xC5\x05\x00\xAC\xDD\x05\x00\xAD\x11\x02\x00\xAE\x11"
-     "\x02\x00"
-     "\xAF\x11\x02\x00\xB6\x01\x03\x00\x84\xE0\x00\x80\x84\x44\x03\x00\xB5\x61"
-     "\x03\x00"
-     "\x88\xE0\x00\x80\xB3\x49\x03\x00\x87\xD8\x03\x00\x86\xF8\x1C\x00\xBE\x09"
-     "\x03\x00"
-     "\xBF\x09\x03\x00\xBC\x35\x03\x00\xBD\x09\x03\x00\xBA\x3D\x03\x00\xBB\x21"
-     "\x03\x00"
-     "\x8C\xE0\x00\x80\x90\xE0\x00\x80\xB0\xCD\x03\x00\xB1\xD1\x03\x00\xB2\xD5"
-     "\x03\x00"
-     "\xB3\xE9\x03\x00\xB4\xF9\x03\x00\xB5\xF9\x03\x00\xB6\xE9\x03\x00\xB7\xE9"
-     "\x03\x00"
-     "\xB8\xD9\x03\x00\xB9\xD9\x03\x00\xBA\x09\x03\x00\xBB\x09\x03\x00\xBC\x19"
-     "\x03\x00"
-     "\xBD\x19\x03\x00\xBE\x09\x03\x00\xBF\x09\x03\x00\x94\xE0\x00\x80\x98\xE0"
-     "\x00\x80"
-     "\xB5\x0D\x21\x00\x9C\xE0\x00\x80\xA0\xE0\x00\x80\xA4\xE0\x00\x80\xA8\xE0"
-     "\x00\x80"
-     "\x2C\xDF\x00\x80\xA8\xE5\x03\x00\xA9\xF5\x03\x00\xAA\xF5\x03\x00\xAB\xC5"
-     "\x03\x00"
-     "\xAC\xDD\x03\x00\xAD\xB5\x03\x00\xAE\xBD\x03\x00\xAF\xB5\x03\x00\xA3\x1D"
-     "\x02\x00"
-     "\xAC\xE0\x00\x80\xB0\xE0\x00\x80\xB4\xE0\x00\x80\xBE\x1C\x1D\x00\xA6\x55"
-     "\x02\x00"
-     "\xA5\x35\x02\x00\xB8\xE0\x00\x80\xAB\x75\x02\x00\xAA\x69\x02\x00\xBC\xE0"
-     "\x00\x80"
-     "\xC0\xE0\x00\x80\xAF\x5D\x02\x00\xAE\x5D\x02\x00\xAD\x5D\x02\x00\xAC\x61"
-     "\x02\x00"
-     "\xC4\xE0\x00\x80\xC8\xE0\x00\x80\xCC\xE0\x00\x80\xD0\xE0\x00\x80\xEF\x8C"
-     "\x03\x00"
-     "\xD4\xE0\x00\x80\xD8\xE0\x00\x80\xDC\xE0\x00\x80\xE3\x6C\x03\x00\xE0\xE0"
-     "\x00\x80"
-     "\xE1\x78\x00\x00\xE4\xE0\x00\x80\xE8\xE0\x00\x80\xF0\xE0\x00\x80\x86\x58"
-     "\x1C\x00"
-     "\x87\xF0\x1D\x00\x80\xE5\x00\x00\x81\xE9\x00\x00\x82\xF9\x00\x00\xF4\xE0"
-     "\x00\x80"
-     "\xF8\xE0\x00\x80\xFC\xE0\x00\x80\x00\xE1\x00\x80\xEF\xD4\x06\x00\x04\xE1"
-     "\x00\x80"
-     "\xE1\x74\x06\x00\x08\xE1\x00\x80\xE3\x54\x01\x00\x0C\xE1\x00\x80\x10\xE1"
-     "\x00\x80"
-     "\x14\xE1\x00\x80\x18\xE1\x00\x80\xB3\x91\x01\x00\x1C\xE1\x00\x80\x20\xE1"
-     "\x00\x80"
-     "\x84\xC4\x1D\x00\x24\xE1\x00\x80\xB6\x15\x01\x00\xB5\xA5\x01\x00\x28\xE1"
-     "\x00\x80"
-     "\xBB\x35\x01\x00\xBA\x35\x01\x00\x2C\xE1\x00\x80\x30\xE1\x00\x80\xBF\xFD"
-     "\x00\x00"
-     "\xBE\xFD\x00\x00\xBD\xFD\x00\x00\xBC\x25\x01\x00\xA9\x45\x1E\x00\xA8\x25"
-     "\x1E\x00"
-     "\xAB\x71\x1E\x00\xAA\x5D\x1E\x00\xAD\xD1\x1E\x00\xAC\x69\x1E\x00\xAF\xD5"
-     "\x1E\x00"
-     "\xAE\xD9\x1E\x00\xEC\xE0\x00\x80\x34\xE1\x00\x80\x38\xE1\x00\x80\x3C\xE1"
-     "\x00\x80"
-     "\x40\xE1\x00\x80\x44\xE1\x00\x80\x48\xE1\x00\x80\x4C\xE1\x00\x80\xB9\x01"
-     "\x1F\x00"
-     "\xB8\x69\x1F\x00\xBB\x01\x1F\x00\xBA\x09\x1F\x00\xBD\x05\x1F\x00\xBC\x19"
-     "\x1F\x00"
-     "\xBF\x05\x1F\x00\xBE\x05\x1F\x00\xB1\xB5\x1E\x00\xB0\xB5\x1E\x00\xB3\x6D"
-     "\x1F\x00"
-     "\xB2\xB5\x1E\x00\xB5\x7D\x1F\x00\xB4\x7D\x1F\x00\xB7\x51\x1F\x00\xB6\x61"
-     "\x1F\x00"
-     "\xA3\xD1\x1E\x00\x82\x2D\x00\x00\x81\x15\x00\x00\x80\x1D\x00\x00\x50\xE1"
-     "\x00\x80"
-     "\xA6\x55\x1E\x00\xA5\xE5\x1E\x00\x54\xE1\x00\x80\xAB\x75\x1E\x00\xAA\x75"
-     "\x1E\x00"
-     "\x58\xE1\x00\x80\xBE\x60\x03\x00\xAF\xBD\x1F\x00\xAE\xBD\x1F\x00\xAD\xBD"
-     "\x1F\x00"
-     "\xAC\x65\x1E\x00\x5C\xE1\x00\x80\xB3\x55\x1F\x00\x86\x68\x00\x00\x87\x6C"
-     "\x03\x00"
-     "\xB6\x71\x1F\x00\x60\xE1\x00\x80\x64\xE1\x00\x80\xB5\x45\x1F\x00\xBA\x55"
-     "\x1F\x00"
-     "\xBB\x5D\x1F\x00\x68\xE1\x00\x80\x6C\xE1\x00\x80\xBE\x39\x1F\x00\xBF\x25"
-     "\x1F\x00"
-     "\xBC\x45\x1F\x00\xBD\x39\x1F\x00\x70\xE1\x00\x80\x74\xE1\x00\x80\x78\xE1"
-     "\x00\x80"
-     "\x7C\xE1\x00\x80\x80\xE1\x00\x80\x84\xE1\x00\x80\x88\xE1\x00\x80\xEF\xF4"
-     "\x1C\x00"
-     "\x8C\xE1\x00\x80\xE1\x60\x1D\x00\x90\xE1\x00\x80\xE3\x74\x1D\x00\x94\xE1"
-     "\x00\x80"
-     "\x98\xE1\x00\x80\x9C\xE1\x00\x80\xA0\xE1\x00\x80\xA3\xD1\x1E\x00\xA4\xE1"
-     "\x00\x80"
-     "\xA8\xE1\x00\x80\xAC\xE1\x00\x80\xB0\xE1\x00\x80\xA6\xF5\x1E\x00\xA5\xC1"
-     "\x1E\x00"
-     "\xB4\xE1\x00\x80\xAB\xD9\x1E\x00\xAA\xD1\x1E\x00\xB8\xE1\x00\x80\xBC\xE1"
-     "\x00\x80"
-     "\xAF\xA1\x1E\x00\xAE\xBD\x1E\x00\xAD\xBD\x1E\x00\xAC\xC1\x1E\x00\xA8\xF5"
-     "\x1E\x00"
-     "\xA9\x09\x1E\x00\xAA\x19\x1E\x00\xAB\x19\x1E\x00\xAC\x09\x1E\x00\xAD\x35"
-     "\x1E\x00"
-     "\xAE\x31\x1E\x00\xAF\x2D\x1E\x00\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\xC0\xE1\x00\x80\xC4\xE1\x00\x80\xC8\xE1\x00\x80\x84\xFC\x01\x00\xBE\xF0"
-     "\x01\x00"
-     "\xB8\x2D\x1E\x00\xB9\xE1\x01\x00\xBA\xE1\x01\x00\xBB\xE1\x01\x00\xBC\xE1"
-     "\x01\x00"
-     "\xBD\xE1\x01\x00\xBE\xE1\x01\x00\xBF\xE1\x01\x00\xB0\x55\x1E\x00\xB1\x5D"
-     "\x1E\x00"
-     "\xB2\x55\x1E\x00\xB3\x29\x1E\x00\xB4\x39\x1E\x00\xB5\x25\x1E\x00\xB6\x25"
-     "\x1E\x00"
-     "\xB7\x15\x1E\x00\xA8\xD5\x02\x00\xA9\xDD\x02\x00\xAA\xE1\x02\x00\xAB\xE1"
-     "\x02\x00"
-     "\xAC\x11\x03\x00\xAD\x11\x03\x00\xAE\x11\x03\x00\xAF\x11\x03\x00\xD0\xE1"
-     "\x00\x80"
-     "\xD4\xE1\x00\x80\xD8\xE1\x00\x80\xDC\xE1\x00\x80\xE0\xE1\x00\x80\xE4\xE1"
-     "\x00\x80"
-     "\xE8\xE1\x00\x80\xEC\xE1\x00\x80\xB8\x7D\x03\x00\xB9\x01\x03\x00\xBA\x01"
-     "\x03\x00"
-     "\xBB\x01\x03\x00\xBC\x01\x03\x00\xBD\x09\x03\x00\xBE\x31\x03\x00\xBF\x31"
-     "\x03\x00"
-     "\xB0\x71\x03\x00\xB1\x71\x03\x00\xB2\x71\x03\x00\xB3\x4D\x03\x00\xB4\x59"
-     "\x03\x00"
-     "\xB5\x45\x03\x00\xB6\x4D\x03\x00\xB7\x45\x03\x00\x86\x20\x04\x00\x87\xA0"
-     "\x03\x00"
-     "\xB3\x61\x02\x00\xBE\x80\x04\x00\xB5\x91\x03\x00\xF0\xE1\x00\x80\xF4\xE1"
-     "\x00\x80"
-     "\xB6\x99\x03\x00\x7C\xE0\x00\x80\x84\xB8\x03\x00\xBB\xBD\x03\x00\xBA\xBD"
-     "\x03\x00"
-     "\xBD\x91\x03\x00\xBC\xAD\x03\x00\xBF\x89\x03\x00\xBE\x95\x03\x00\xA3\x29"
-     "\x02\x00"
-     "\xF8\xE1\x00\x80\xFC\xE1\x00\x80\x00\xE2\x00\x80\x85\x94\x06\x00\xA6\xD1"
-     "\x03\x00"
-     "\xA5\xD9\x03\x00\x04\xE2\x00\x80\xAB\xF5\x03\x00\xAA\xF5\x03\x00\x08\xE2"
-     "\x00\x80"
-     "\x0C\xE2\x00\x80\xAF\xC1\x03\x00\xAE\xDD\x03\x00\xAD\xD9\x03\x00\xAC\xE5"
-     "\x03\x00"
-     "\x10\xE2\x00\x80\x14\xE2\x00\x80\x18\xE2\x00\x80\x1C\xE2\x00\x80\x20\xE2"
-     "\x00\x80"
-     "\x24\xE2\x00\x80\x28\xE2\x00\x80\x2C\xE2\x00\x80\x81\x11\x00\x00\x80\x0D"
-     "\x00\x00"
-     "\x30\xE2\x00\x80\x82\x19\x00\x00\x34\xE2\x00\x80\x38\xE2\x00\x80\x40\xE2"
-     "\x00\x80"
-     "\x44\xE2\x00\x80\x48\xE2\x00\x80\xE1\x94\x01\x00\x4C\xE2\x00\x80\xE3\xE4"
-     "\x00\x00"
-     "\x50\xE2\x00\x80\x54\xE2\x00\x80\x86\x78\x04\x00\x87\x80\x02\x00\x58\xE2"
-     "\x00\x80"
-     "\x5C\xE2\x00\x80\x60\xE2\x00\x80\x64\xE2\x00\x80\x68\xE2\x00\x80\x6C\xE2"
-     "\x00\x80"
-     "\x70\xE2\x00\x80\xEF\x38\x00\x00\x74\xE2\x00\x80\x78\xE2\x00\x80\x7C\xE2"
-     "\x00\x80"
-     "\x80\xE2\x00\x80\x84\xE2\x00\x80\x88\xE2\x00\x80\x8C\xE2\x00\x80\xEF\xC8"
-     "\x01\x00"
-     "\xBE\x6C\x04\x00\xE1\xC4\x1E\x00\x84\xA0\x05\x00\xE3\x70\x1E\x00\x90\xE2"
-     "\x00\x80"
-     "\x94\xE2\x00\x80\x98\xE2\x00\x80\x9C\xE2\x00\x80\xA0\xE2\x00\x80\xB6\xF1"
-     "\x01\x00"
-     "\xB5\xF9\x01\x00\x85\xF4\x20\x00\xB3\xF9\x01\x00\x82\x29\x00\x00\x81\x21"
-     "\x00\x00"
-     "\x80\x1D\x00\x00\xBF\x3D\x01\x00\xBE\x29\x01\x00\xBD\x21\x01\x00\xBC\x29"
-     "\x01\x00"
-     "\xBB\x69\x01\x00\xBA\x61\x01\x00\xA4\xE2\x00\x80\xA8\xE2\x00\x80\xA8\x21"
-     "\x06\x00"
-     "\xA9\x21\x06\x00\xAA\x21\x06\x00\xAB\x21\x06\x00\xAC\x21\x06\x00\xAD\x21"
-     "\x06\x00"
-     "\xAE\x21\x06\x00\xAF\x21\x06\x00\x3C\xE2\x00\x80\xBF\x2C\x01\x00\x86\x00"
-     "\x0C\x00"
-     "\x87\x60\x03\x00\xAC\xE2\x00\x80\xB0\xE2\x00\x80\xB4\xE2\x00\x80\xB8\xE2"
-     "\x00\x80"
-     "\xB8\x41\x07\x00\xB9\x41\x07\x00\xBA\x41\x07\x00\xBB\x41\x07\x00\xBC\x41"
-     "\x07\x00"
-     "\xBD\x41\x07\x00\xBE\x41\x07\x00\xBF\x41\x07\x00\xB0\xF5\x06\x00\xB1\xFD"
-     "\x06\x00"
-     "\xB2\xAD\x06\x00\xB3\xA5\x06\x00\xB4\xBD\x06\x00\xB5\xA9\x06\x00\xB6\x91"
-     "\x06\x00"
-     "\xB7\x91\x06\x00\xA3\xA9\x06\x00\xBC\xE2\x00\x80\xC0\xE2\x00\x80\xC4\xE2"
-     "\x00\x80"
-     "\xC8\xE2\x00\x80\xA6\xA1\x06\x00\xA5\xA9\x06\x00\xCC\xE2\x00\x80\xAB\x39"
-     "\x06\x00"
-     "\xAA\x31\x06\x00\xD0\xE2\x00\x80\xD4\xE2\x00\x80\xAF\x6D\x06\x00\xAE\x79"
-     "\x06\x00"
-     "\xAD\x71\x06\x00\xAC\x79\x06\x00\xD8\xE2\x00\x80\xB3\x85\x07\x00\xDC\xE2"
-     "\x00\x80"
-     "\xE0\xE2\x00\x80\xB6\x85\x07\x00\xE4\xE2\x00\x80\xE8\xE2\x00\x80\xB5\x8D"
-     "\x07\x00"
-     "\xBA\x89\x07\x00\xBB\x89\x07\x00\xEC\xE2\x00\x80\xF0\xE2\x00\x80\xBE\x89"
-     "\x07\x00"
-     "\xBF\x75\x07\x00\xBC\x89\x07\x00\xBD\x89\x07\x00\xF4\xE2\x00\x80\xE1\xC0"
-     "\x05\x00"
-     "\xEF\x7C\x05\x00\xE3\xD4\x04\x00\xF8\xE2\x00\x80\xFC\xE2\x00\x80\x00\xE3"
-     "\x00\x80"
-     "\x04\xE3\x00\x80\x08\xE3\x00\x80\x0C\xE3\x00\x80\x10\xE3\x00\x80\x14\xE3"
-     "\x00\x80"
-     "\xE1\xA8\x05\x00\x18\xE3\x00\x80\xE3\x8C\x05\x00\xEF\xF8\x0C\x00\xA3\x01"
-     "\x06\x00"
-     "\x82\x71\x00\x00\x81\x65\x00\x00\x80\x65\x00\x00\x84\x80\x01\x00\xA6\x01"
-     "\x06\x00"
-     "\xA5\x09\x06\x00\xBE\x10\x00\x00\xAB\x0D\x06\x00\xAA\x0D\x06\x00\x86\xE8"
-     "\x0F\x00"
-     "\x87\x18\x00\x00\xAF\xF1\x06\x00\xAE\x0D\x06\x00\xAD\x0D\x06\x00\xAC\x0D"
-     "\x06\x00"
-     "\xA8\xB9\x06\x00\xA9\xB9\x06\x00\xAA\xC9\x06\x00\xAB\xC5\x06\x00\xAC\xC1"
-     "\x06\x00"
-     "\xAD\xC1\x06\x00\xAE\xF9\x06\x00\xAF\xF9\x06\x00\x20\xE3\x00\x80\x24\xE3"
-     "\x00\x80"
-     "\x28\xE3\x00\x80\x2C\xE3\x00\x80\x30\xE3\x00\x80\x34\xE3\x00\x80\x38\xE3"
-     "\x00\x80"
-     "\x3C\xE3\x00\x80\xB8\x4D\x06\x00\xB9\x55\x06\x00\xBA\x5D\x06\x00\xBB\x55"
-     "\x06\x00"
-     "\xBC\x61\x06\x00\xBD\x61\x06\x00\xBE\x61\x06\x00\xBF\x61\x06\x00\xB0\x89"
-     "\x06\x00"
-     "\xB1\x89\x06\x00\xB2\x99\x06\x00\xB3\x99\x06\x00\xB4\x89\x06\x00\xB5\x75"
-     "\x06\x00"
-     "\xB6\x7D\x06\x00\xB7\x75\x06\x00\x40\xE3\x00\x80\xB5\x8D\x02\x00\xB6\x85"
-     "\x02\x00"
-     "\x84\x40\x0C\x00\x8F\xF1\xE8\x03\x44\xE3\x00\x80\x48\xE3\x00\x80\xB3\x21"
-     "\x02\x00"
-     "\xBC\xE1\x02\x00\xBD\xED\x02\x00\xBE\xE5\x02\x00\xBF\xF9\x02\x00\xAC\xC8"
-     "\x32\x00"
-     "\x4C\xE3\x00\x80\xBA\x91\x02\x00\xBB\x91\x02\x00\x9B\x01\x1F\x00\x9A\x4D"
-     "\x1E\x00"
-     "\x99\x4D\x1E\x00\x98\x3D\x13\x00\x9F\xF1\x04\x00\x9E\xB9\x1A\x00\x9D\x79"
-     "\x1A\x00"
-     "\x9C\x81\x18\x00\x93\x41\x16\x00\x92\x91\x14\x00\x91\xF9\xEA\x03\x90\x4D"
-     "\xEA\x03"
-     "\x97\x61\x12\x00\x96\x71\x12\x00\x95\x81\x10\x00\x94\x41\x16\x00\xB1\x01"
-     "\x38\x00"
-     "\xB0\xE1\x3E\x00\xB3\x81\x3A\x00\xB2\x89\x3A\x00\xB5\x91\x26\x00\xB4\xD1"
-     "\x3B\x00"
-     "\x50\xE3\x00\x80\x54\xE3\x00\x80\xA3\x81\x02\x00\xA2\x11\x00\x00\xA1\x69"
-     "\x06\x00"
-     "\xA0\x8D\x06\x00\x58\xE3\x00\x80\xCC\xE1\x00\x80\x5C\xE3\x00\x80\x60\xE3"
-     "\x00\x80"
-     "\xA1\xF1\x02\x00\x64\xE3\x00\x80\xA3\xCD\x0E\x00\xA2\xCD\x0E\x00\xA5\x01"
-     "\x08\x00"
-     "\xA4\x51\x0F\x00\xA7\xC5\x0B\x00\xA6\xFD\x0A\x00\xA9\xC1\x36\x00\xA8\x01"
-     "\x34\x00"
-     "\xAB\x01\x30\x00\xAA\x41\x36\x00\xAD\xCD\x32\x00\xAC\xD5\x32\x00\xAF\xFD"
-     "\x3E\x00"
-     "\xAE\x01\x3C\x00\x82\xE1\xE9\x03\x83\x25\xEB\x03\x68\xE3\x00\x80\x6C\xE3"
-     "\x00\x80"
-     "\x86\x31\x17\x00\x87\x29\x17\x00\x84\x89\xEB\x03\x85\xE9\x15\x00\x8A\xD9"
-     "\x11\x00"
-     "\x8B\x21\x13\x00\x70\xE3\x00\x80\x74\xE3\x00\x80\x8E\x35\x1F\x00\x8F\x3D"
-     "\x1F\x00"
-     "\x8C\x29\x13\x00\x8D\x4D\x12\x00\x92\x01\x1E\x00\x93\x99\x19\x00\x78\xE3"
-     "\x00\x80"
-     "\x84\x00\x0C\x00\x96\xF1\x05\x00\x97\x75\x07\x00\x94\x79\x1B\x00\x95\xB1"
-     "\x1B\x00"
-     "\x9A\xA1\x07\x00\x9B\xD1\x01\x00\x7C\xE3\x00\x80\x80\xE3\x00\x80\x81\xF1"
-     "\x00\x00"
-     "\x80\x19\x00\x00\x9C\x59\x03\x00\x82\xF9\x00\x00\x84\xE3\x00\x80\x88\xE3"
-     "\x00\x80"
-     "\xBE\xC0\x0D\x00\xEF\x28\x00\x00\x86\x30\x0C\x00\x87\x5C\x0D\x00\x90\xE3"
-     "\x00\x80"
-     "\xEF\xC4\x06\x00\x94\xE3\x00\x80\xE1\xE4\x06\x00\x98\xE3\x00\x80\xE3\x14"
-     "\x06\x00"
-     "\x9C\xE3\x00\x80\xE1\x84\x01\x00\xA0\xE3\x00\x80\xE3\xFC\x00\x00\xA3\xFD"
-     "\x03\x00"
-     "\xA4\xE3\x00\x80\xA8\xE3\x00\x80\x84\x4C\x0F\x00\xAC\xE3\x00\x80\xA6\x59"
-     "\x03\x00"
-     "\xA5\x51\x03\x00\xB0\xE3\x00\x80\xAB\x4D\x03\x00\xAA\x4D\x03\x00\xB4\xE3"
-     "\x00\x80"
-     "\xB8\xE3\x00\x80\xAF\x25\x03\x00\xAE\x39\x03\x00\xAD\x31\x03\x00\xAC\x3D"
-     "\x03\x00"
-     "\x1C\xE3\x00\x80\x8C\xE3\x00\x80\xBC\xE3\x00\x80\xC0\xE3\x00\x80\xC4\xE3"
-     "\x00\x80"
-     "\xC8\xE3\x00\x80\xCC\xE3\x00\x80\xD0\xE3\x00\x80\xA8\x81\x0D\x00\xA9\x81"
-     "\x0D\x00"
-     "\xAA\x81\x0D\x00\xAB\x81\x0D\x00\xAC\x81\x0D\x00\xAD\x81\x0D\x00\xAE\x81"
-     "\x0D\x00"
-     "\xAF\x81\x0D\x00\xB0\x41\x0E\x00\xB1\x51\x0E\x00\xB2\x51\x0E\x00\xB3\x79"
-     "\x0E\x00"
-     "\xB4\xC9\x0E\x00\xB5\xC9\x0E\x00\xB6\xE9\x0E\x00\xB7\xE5\x0E\x00\xB8\xDD"
-     "\x0E\x00"
-     "\xB9\xA1\x0E\x00\xBA\x61\x0F\x00\xBB\x61\x0F\x00\xBC\x59\x0F\x00\xBD\x59"
-     "\x0F\x00"
-     "\xBE\x31\x0F\x00\xBF\x15\x0F\x00\xB3\x29\x0D\x00\xD4\xE3\x00\x80\xD8\xE3"
-     "\x00\x80"
-     "\xDC\xE3\x00\x80\xE0\xE3\x00\x80\xB6\x21\x0D\x00\xB5\x29\x0D\x00\xE4\xE3"
-     "\x00\x80"
-     "\xBB\xB5\x0E\x00\xBA\xA5\x0E\x00\xE8\xE3\x00\x80\xEC\xE3\x00\x80\xBF\x2D"
-     "\x0E\x00"
-     "\xBE\x2D\x0E\x00\xBD\x8D\x0E\x00\xBC\xA5\x0E\x00\x82\x29\x00\x00\xA3\x6D"
-     "\x0D\x00"
-     "\x80\x1D\x00\x00\x81\x15\x00\x00\xA6\x65\x0D\x00\xF0\xE3\x00\x80\xF4\xE3"
-     "\x00\x80"
-     "\xA5\x6D\x0D\x00\xAA\xE1\x0E\x00\xAB\xF1\x0E\x00\xF8\xE3\x00\x80\xBE\xA4"
-     "\x03\x00"
-     "\xAE\x69\x0E\x00\xAF\x69\x0E\x00\xAC\xE1\x0E\x00\xAD\xC9\x0E\x00\xA8\x99"
-     "\x0E\x00"
-     "\xA9\xA5\x0E\x00\xAA\xA1\x0E\x00\xAB\xBD\x0E\x00\xAC\xA9\x0E\x00\xAD\xD1"
-     "\x0E\x00"
-     "\xAE\xD1\x0E\x00\xAF\xCD\x0E\x00\x86\x80\x00\x00\x87\x04\x01\x00\x00\xE4"
-     "\x00\x80"
-     "\x04\xE4\x00\x80\x08\xE4\x00\x80\x0C\xE4\x00\x80\x10\xE4\x00\x80\x14\xE4"
-     "\x00\x80"
-     "\xB8\x5D\x01\x00\xB9\xC1\x01\x00\xBA\xC1\x01\x00\xBB\xC1\x01\x00\xBC\xC1"
-     "\x01\x00"
-     "\xBD\xC9\x01\x00\xBE\xF1\x01\x00\xBF\xF1\x01\x00\xB0\xBD\x0E\x00\xB1\x8D"
-     "\x0E\x00"
-     "\xB2\x85\x0E\x00\xB3\x65\x01\x00\xB4\x7D\x01\x00\xB5\x65\x01\x00\xB6\x6D"
-     "\x01\x00"
-     "\xB7\x65\x01\x00\xB3\x29\x0E\x00\xFC\xE3\x00\x80\x18\xE4\x00\x80\x1C\xE4"
-     "\x00\x80"
-     "\x20\xE4\x00\x80\xB6\x51\x0E\x00\xB5\x21\x0E\x00\x24\xE4\x00\x80\xBB\x49"
-     "\x0E\x00"
-     "\xBA\x79\x0E\x00\x28\xE4\x00\x80\x2C\xE4\x00\x80\xBF\xA1\x01\x00\xBE\xB9"
-     "\x01\x00"
-     "\xBD\xB1\x01\x00\xBC\x51\x0E\x00\x30\xE4\x00\x80\xA3\x6D\x0E\x00\x34\xE4"
-     "\x00\x80"
-     "\x38\xE4\x00\x80\xA6\x15\x0E\x00\x3C\xE4\x00\x80\x40\xE4\x00\x80\xA5\x65"
-     "\x0E\x00"
-     "\xAA\x3D\x0E\x00\xAB\x0D\x0E\x00\x44\xE4\x00\x80\x48\xE4\x00\x80\xAE\xFD"
-     "\x01\x00"
-     "\xAF\xE5\x01\x00\xAC\x15\x0E\x00\xAD\xF5\x01\x00\xE1\x38\x01\x00\x4C\xE4"
-     "\x00\x80"
-     "\xE3\xD4\x0F\x00\x50\xE4\x00\x80\x54\xE4\x00\x80\xBE\x58\x03\x00\x58\xE4"
-     "\x00\x80"
-     "\x5C\xE4\x00\x80\x80\x0D\x00\x00\x81\x15\x00\x00\x82\x1D\x00\x00\x60\xE4"
-     "\x00\x80"
-     "\x64\xE4\x00\x80\x68\xE4\x00\x80\xEF\xD4\x02\x00\x70\xE4\x00\x80\x86\xE0"
-     "\x04\x00"
-     "\x87\xE0\x03\x00\xBE\x00\x04\x00\xB3\x7D\x03\x00\x74\xE4\x00\x80\xB5\x6D"
-     "\x03\x00"
-     "\xB6\x65\x03\x00\x78\xE4\x00\x80\x7C\xE4\x00\x80\x80\xE4\x00\x80\xBA\xF9"
-     "\x03\x00"
-     "\xBB\xF1\x03\x00\xBC\xD9\x03\x00\xBD\xD9\x03\x00\xBE\xB9\x03\x00\xBF\xB5"
-     "\x03\x00"
-     "\xAA\xBD\x0E\x00\xAB\x01\x0E\x00\xA8\xBD\x0E\x00\xA9\xB5\x0E\x00\xAE\x01"
-     "\x0E\x00"
-     "\xAF\x09\x0E\x00\xAC\x11\x0E\x00\xAD\x11\x0E\x00\x84\xE4\x00\x80\x88\xE4"
-     "\x00\x80"
-     "\x8C\xE4\x00\x80\x90\xE4\x00\x80\x94\xE4\x00\x80\x98\xE4\x00\x80\x9C\xE4"
-     "\x00\x80"
-     "\xA0\xE4\x00\x80\xBA\xD1\x0F\x00\xBB\xD1\x0F\x00\xB8\x25\x0E\x00\xB9\xCD"
-     "\x0F\x00"
-     "\xBE\xF1\x0F\x00\xBF\xF1\x0F\x00\xBC\xC1\x0F\x00\xBD\xC1\x0F\x00\xB2\x21"
-     "\x0E\x00"
-     "\xB3\x2D\x0E\x00\xB0\x41\x0E\x00\xB1\x41\x0E\x00\xB6\x29\x0E\x00\xB7\x1D"
-     "\x0E\x00"
-     "\xB4\x35\x0E\x00\xB5\x29\x0E\x00\xA8\x29\x02\x00\xA9\x29\x02\x00\xAA\x39"
-     "\x02\x00"
-     "\xAB\x39\x02\x00\xAC\x29\x02\x00\xAD\x29\x02\x00\xAE\x91\x02\x00\xAF\x91"
-     "\x02\x00"
-     "\xA4\xE4\x00\x80\xA8\xE4\x00\x80\xAC\xE4\x00\x80\xB0\xE4\x00\x80\xB4\xE4"
-     "\x00\x80"
-     "\xB8\xE4\x00\x80\xBF\x30\x2B\x00\xBC\xE4\x00\x80\xB8\x99\x02\x00\xB9\xAD"
-     "\x02\x00"
-     "\xBA\xA5\x02\x00\xBB\x49\x01\x00\xBC\x5D\x01\x00\xBD\x45\x01\x00\xBE\x41"
-     "\x01\x00"
-     "\xBF\x41\x01\x00\xB0\xF1\x02\x00\xB1\xF9\x02\x00\xB2\xC1\x02\x00\xB3\xC1"
-     "\x02\x00"
-     "\xB4\xB1\x02\x00\xB5\xBD\x02\x00\xB6\xB5\x02\x00\xB7\xA9\x02\x00\xE1\x08"
-     "\x0E\x00"
-     "\xE1\xF4\x0C\x00\xE3\xD4\x0E\x00\xE3\x90\x0F\x00\xC0\xE4\x00\x80\xC4\xE4"
-     "\x00\x80"
-     "\xC8\xE4\x00\x80\xCC\xE4\x00\x80\x81\x11\x00\x00\x80\x0D\x00\x00\xD0\xE4"
-     "\x00\x80"
-     "\x82\x19\x00\x00\xD4\xE4\x00\x80\xD8\xE4\x00\x80\xEF\xAC\x01\x00\xEF\xCC"
-     "\x0E\x00"
-     "\xA3\xF1\x02\x00\xE0\xE4\x00\x80\xE4\xE4\x00\x80\x84\x80\x05\x00\xE8\xE4"
-     "\x00\x80"
-     "\xA6\xE9\x02\x00\xA5\xE1\x02\x00\xEC\xE4\x00\x80\xAB\x7D\x02\x00\xAA\x75"
-     "\x02\x00"
-     "\x86\xC8\x04\x00\x87\xEC\x04\x00\xAF\x39\x02\x00\xAE\x35\x02\x00\xAD\x55"
-     "\x02\x00"
-     "\xAC\x55\x02\x00\xA8\x3D\x06\x00\xA9\x55\x06\x00\xAA\x55\x06\x00\xAB\x6D"
-     "\x06\x00"
-     "\xAC\x75\x06\x00\xAD\xB5\x06\x00\xAE\xBD\x06\x00\xAF\xB5\x06\x00\xDC\xE4"
-     "\x00\x80"
-     "\xF0\xE4\x00\x80\xF4\xE4\x00\x80\xF8\xE4\x00\x80\xFC\xE4\x00\x80\x00\xE5"
-     "\x00\x80"
-     "\x04\xE5\x00\x80\x08\xE5\x00\x80\xB8\x6D\x07\x00\xB9\x75\x07\x00\xBA\x7D"
-     "\x07\x00"
-     "\xBB\x71\x07\x00\xBC\xC1\x07\x00\xBD\xC1\x07\x00\xBE\xC1\x07\x00\xBF\xC1"
-     "\x07\x00"
-     "\xB0\xCD\x06\x00\xB1\xD5\x06\x00\xB2\xDD\x06\x00\xB3\xD5\x06\x00\xB4\xCD"
-     "\x06\x00"
-     "\xB5\x55\x07\x00\xB6\x5D\x07\x00\xB7\x55\x07\x00\xB3\xDD\x06\x00\x0C\xE5"
-     "\x00\x80"
-     "\x10\xE5\x00\x80\x14\xE5\x00\x80\x18\xE5\x00\x80\xB6\x35\x06\x00\xB5\xFD"
-     "\x06\x00"
-     "\x1C\xE5\x00\x80\xBB\x11\x06\x00\xBA\x09\x06\x00\x20\xE5\x00\x80\x24\xE5"
-     "\x00\x80"
-     "\xBF\x91\x07\x00\xBE\x09\x06\x00\xBD\x01\x06\x00\xBC\x09\x06\x00\x28\xE5"
-     "\x00\x80"
-     "\xA3\x99\x06\x00\x2C\xE5\x00\x80\x30\xE5\x00\x80\xA6\x71\x06\x00\x34\xE5"
-     "\x00\x80"
-     "\x38\xE5\x00\x80\xA5\xB9\x06\x00\xAA\x4D\x06\x00\xAB\x55\x06\x00\x3C\xE5"
-     "\x00\x80"
-     "\x40\xE5\x00\x80\xAE\x4D\x06\x00\xAF\xD5\x07\x00\xAC\x4D\x06\x00\xAD\x45"
-     "\x06\x00"
-     "\x80\xED\x01\x00\x81\x09\x00\x00\x82\x19\x00\x00\x44\xE5\x00\x80\x48\xE5"
-     "\x00\x80"
-     "\x84\xD0\x01\x00\xBE\xC4\x01\x00\x4C\xE5\x00\x80\x86\x80\x00\x00\x87\x20"
-     "\x01\x00"
-     "\x50\xE5\x00\x80\x54\xE5\x00\x80\x58\xE5\x00\x80\x5C\xE5\x00\x80\x60\xE5"
-     "\x00\x80"
-     "\x64\xE5\x00\x80\x68\xE5\x00\x80\x6C\xE5\x00\x80\x70\xE5\x00\x80\x74\xE5"
-     "\x00\x80"
-     "\x78\xE5\x00\x80\xE1\x0C\x05\x00\x7C\xE5\x00\x80\xE3\x80\x04\x00\x80\xE5"
-     "\x00\x80"
-     "\x84\xE5\x00\x80\x88\xE5\x00\x80\xEF\xDC\x05\x00\x8C\xE5\x00\x80\x90\xE5"
-     "\x00\x80"
-     "\x94\xE5\x00\x80\x98\xE5\x00\x80\xA8\x45\x06\x00\xA9\x4D\x06\x00\xAA\x59"
-     "\x06\x00"
-     "\xAB\x51\x06\x00\xAC\xB5\x01\x00\xAD\xBD\x01\x00\xAE\xB5\x01\x00\xAF\xAD"
-     "\x01\x00"
-     "\x84\xAC\x01\x00\x9C\xE5\x00\x80\xA0\xE5\x00\x80\xA4\xE5\x00\x80\xA8\xE5"
-     "\x00\x80"
-     "\xAC\xE5\x00\x80\xB0\xE5\x00\x80\xB4\xE5\x00\x80\xB8\x69\x01\x00\xB9\x69"
-     "\x01\x00"
-     "\xBA\x09\x01\x00\xBB\x05\x01\x00\xBC\x1D\x01\x00\xBD\x01\x01\x00\xBE\x01"
-     "\x01\x00"
-     "\xBF\x35\x01\x00\xB0\xD5\x01\x00\xB1\xDD\x01\x00\xB2\xD5\x01\x00\xB3\x69"
-     "\x01\x00"
-     "\xB4\x79\x01\x00\xB5\x79\x01\x00\xB6\x69\x01\x00\xB7\x61\x01\x00\x81\xBD"
-     "\x03\x00"
-     "\x80\xBD\x03\x00\xB3\xD9\x05\x00\x82\x19\x00\x00\xB5\x35\x02\x00\xB8\xE5"
-     "\x00\x80"
-     "\xBE\x44\x03\x00\xB6\x35\x02\x00\xC0\xE5\x00\x80\xC4\xE5\x00\x80\xBB\x11"
-     "\x02\x00"
-     "\xBA\x19\x02\x00\xBD\xB5\x02\x00\xBC\x09\x02\x00\xBF\xB5\x02\x00\xBE\xB5"
-     "\x02\x00"
-     "\xC8\xE5\x00\x80\xA6\x6D\x02\x00\xA5\x6D\x02\x00\x84\x00\x3C\x00\xA3\x81"
-     "\x05\x00"
-     "\xCC\xE5\x00\x80\x86\xF8\x3C\x00\x87\x7C\x03\x00\xAF\xED\x02\x00\xAE\xED"
-     "\x02\x00"
-     "\xAD\xED\x02\x00\xAC\x51\x02\x00\xAB\x49\x02\x00\xAA\x41\x02\x00\xD0\xE5"
-     "\x00\x80"
-     "\xD4\xE5\x00\x80\xD8\xE5\x00\x80\xDC\xE5\x00\x80\xE0\xE5\x00\x80\xE4\xE5"
-     "\x00\x80"
-     "\xE3\x20\x01\x00\xE8\xE5\x00\x80\xE1\xA0\x01\x00\xEC\xE5\x00\x80\xEF\x40"
-     "\x02\x00"
-     "\xF0\xE5\x00\x80\xF4\xE5\x00\x80\xF8\xE5\x00\x80\xFC\xE5\x00\x80\x00\xE6"
-     "\x00\x80"
-     "\x04\xE6\x00\x80\x08\xE6\x00\x80\xB3\xF1\x03\x00\x0C\xE6\x00\x80\x10\xE6"
-     "\x00\x80"
-     "\x14\xE6\x00\x80\x6C\xE4\x00\x80\xB6\xE9\x03\x00\xB5\xE1\x03\x00\x18\xE6"
-     "\x00\x80"
-     "\xBB\xAD\x03\x00\xBA\xAD\x03\x00\x1C\xE6\x00\x80\x20\xE6\x00\x80\xBF\x95"
-     "\x03\x00"
-     "\xBE\x95\x03\x00\xBD\xA1\x03\x00\xBC\xAD\x03\x00\xA8\x29\x02\x00\xA9\x29"
-     "\x02\x00"
-     "\xAA\x39\x02\x00\xAB\x39\x02\x00\xAC\x29\x02\x00\xAD\x29\x02\x00\xAE\x59"
-     "\x02\x00"
-     "\xAF\x59\x02\x00\x80\xCD\x01\x00\x81\x09\x00\x00\x82\x19\x00\x00\x24\xE6"
-     "\x00\x80"
-     "\x28\xE6\x00\x80\xBE\x7C\x3D\x00\x87\xB4\x3D\x00\x86\x1C\x3C\x00\xB8\x35"
-     "\x02\x00"
-     "\xB9\x3D\x02\x00\xBA\x35\x02\x00\xBB\xE9\x02\x00\xBC\xF9\x02\x00\xBD\xF9"
-     "\x02\x00"
-     "\xBE\xE9\x02\x00\xBF\xE9\x02\x00\xB0\x29\x02\x00\xB1\x29\x02\x00\xB2\x39"
-     "\x02\x00"
-     "\xB3\x35\x02\x00\xB4\x11\x02\x00\xB5\x11\x02\x00\xB6\x11\x02\x00\xB7\x0D"
-     "\x02\x00"
-     "\x30\xE6\x00\x80\x34\xE6\x00\x80\x38\xE6\x00\x80\x3C\xE6\x00\x80\x40\xE6"
-     "\x00\x80"
-     "\x44\xE6\x00\x80\x48\xE6\x00\x80\xEF\x7C\x00\x00\x4C\xE6\x00\x80\xE1\x54"
-     "\x06\x00"
-     "\x50\xE6\x00\x80\xE3\xA8\x00\x00\xAC\xA8\x32\x00\x54\xE6\x00\x80\x58\xE6"
-     "\x00\x80"
-     "\x5C\xE6\x00\x80\x60\xE6\x00\x80\xA6\xF9\x02\x00\xA5\xF1\x02\x00\x64\xE6"
-     "\x00\x80"
-     "\xA3\xE1\x02\x00\x68\xE6\x00\x80\x6C\xE6\x00\x80\x70\xE6\x00\x80\xAF\x85"
-     "\x02\x00"
-     "\xAE\x85\x02\x00\xAD\xB1\x02\x00\xAC\xBD\x02\x00\xAB\xBD\x02\x00\xAA\xBD"
-     "\x02\x00"
-     "\x74\xE6\x00\x80\x78\xE6\x00\x80\xA8\x21\x3E\x00\xA9\x21\x3E\x00\xAA\x49"
-     "\x3E\x00"
-     "\xAB\x59\x3E\x00\xAC\x69\x3E\x00\xAD\x69\x3E\x00\xAE\x99\x3E\x00\xAF\x99"
-     "\x3E\x00"
-     "\x2C\xE6\x00\x80\x7C\xE6\x00\x80\x80\xE6\x00\x80\x84\xE6\x00\x80\x88\xE6"
-     "\x00\x80"
-     "\x8C\xE6\x00\x80\x90\xE6\x00\x80\x94\xE6\x00\x80\xB8\xF5\x3E\x00\xB9\xFD"
-     "\x3E\x00"
-     "\xBA\xF5\x3E\x00\xBB\x89\x3E\x00\xBC\x99\x3E\x00\xBD\x99\x3E\x00\xBE\x85"
-     "\x3E\x00"
-     "\xBF\xB5\x3E\x00\xB0\xE9\x3E\x00\xB1\xE9\x3E\x00\xB2\xFD\x3E\x00\xB3\xF1"
-     "\x3E\x00"
-     "\xB4\xD5\x3E\x00\xB5\xDD\x3E\x00\xB6\xD5\x3E\x00\xB7\xCD\x3E\x00\xB3\xC1"
-     "\x3E\x00"
-     "\x82\x15\x00\x00\x81\xB5\x00\x00\x80\xB5\x00\x00\x98\xE6\x00\x80\xB6\xE9"
-     "\x3E\x00"
-     "\xB5\xE1\x3E\x00\xBE\x10\x00\x00\xBB\x2D\x3E\x00\xBA\x2D\x3E\x00\x84\x64"
-     "\x03\x00"
-     "\x9C\xE6\x00\x80\xBF\x19\x3E\x00\xBE\x11\x3E\x00\xBD\x25\x3E\x00\xBC\x29"
-     "\x3E\x00"
-     "\xA0\xE6\x00\x80\xA3\x85\x3E\x00\x86\xA8\x00\x00\x87\x6C\x03\x00\xA6\xAD"
-     "\x3E\x00"
-     "\xA4\xE6\x00\x80\xA8\xE6\x00\x80\xA5\xA5\x3E\x00\xAA\x69\x3E\x00\xAB\x69"
-     "\x3E\x00"
-     "\xAC\xE6\x00\x80\xB0\xE6\x00\x80\xAE\x55\x3E\x00\xAF\x5D\x3E\x00\xAC\x6D"
-     "\x3E\x00"
-     "\xAD\x61\x3E\x00\xB3\x89\x3E\x00\xB4\xE6\x00\x80\xB8\xE6\x00\x80\xBC\xE6"
-     "\x00\x80"
-     "\xC0\xE6\x00\x80\xB6\x81\x3E\x00\xB5\x89\x3E\x00\xC4\xE6\x00\x80\xBB\x95"
-     "\x3E\x00"
-     "\xBA\x8D\x3E\x00\xC8\xE6\x00\x80\xCC\xE6\x00\x80\xBF\xFD\x3E\x00\xBE\xF5"
-     "\x3E\x00"
-     "\xBD\x85\x3E\x00\xBC\x85\x3E\x00\xD0\xE6\x00\x80\xD4\xE6\x00\x80\xD8\xE6"
-     "\x00\x80"
-     "\xDC\xE6\x00\x80\xE0\xE6\x00\x80\xE4\xE6\x00\x80\xEF\x94\x3C\x00\xE8\xE6"
-     "\x00\x80"
-     "\xE1\x20\x3D\x00\xEC\xE6\x00\x80\xE3\x34\x3D\x00\xF0\xE6\x00\x80\xF4\xE6"
-     "\x00\x80"
-     "\xF8\xE6\x00\x80\xFC\xE6\x00\x80\x00\xE7\x00\x80\xA8\x01\x3E\x00\xA9\x01"
-     "\x3E\x00"
-     "\xAA\x01\x3E\x00\xAB\x01\x3E\x00\xAC\x01\x3E\x00\xAD\x01\x3E\x00\xAE\x01"
-     "\x3E\x00"
-     "\xAF\x3D\x3E\x00\x80\xDD\x00\x00\x81\x09\x00\x00\x82\x19\x00\x00\x04\xE7"
-     "\x00\x80"
-     "\x08\xE7\x00\x80\x84\x3C\x01\x00\xBE\x74\x00\x00\x10\xE7\x00\x80\xB8\x1D"
-     "\x3E\x00"
-     "\xB9\x3D\x3E\x00\xBA\x35\x3E\x00\xBB\xE9\x01\x00\xBC\xF9\x01\x00\xBD\xF9"
-     "\x01\x00"
-     "\xBE\xE9\x01\x00\xBF\xE9\x01\x00\xB0\x49\x3E\x00\xB1\x51\x3E\x00\xB2\x51"
-     "\x3E\x00"
-     "\xB3\x51\x3E\x00\xB4\x31\x3E\x00\xB5\x39\x3E\x00\xB6\x2D\x3E\x00\xB7\x25"
-     "\x3E\x00"
-     "\xA3\x89\x3D\x00\x14\xE7\x00\x80\x86\xA8\x04\x00\x87\x3C\x03\x00\x18\xE7"
-     "\x00\x80"
-     "\xA6\x81\x3D\x00\xA5\x89\x3D\x00\x1C\xE7\x00\x80\xAB\x95\x3D\x00\xAA\x8D"
-     "\x3D\x00"
-     "\x20\xE7\x00\x80\x24\xE7\x00\x80\xAF\xFD\x3D\x00\xAE\xF5\x3D\x00\xAD\x85"
-     "\x3D\x00"
-     "\xAC\x85\x3D\x00\x28\xE7\x00\x80\xB3\x9D\x02\x00\x84\x48\x03\x00\xBE\x40"
-     "\x04\x00"
-     "\xB6\x55\x03\x00\x2C\xE7\x00\x80\x30\xE7\x00\x80\xB5\xB1\x02\x00\xBA\x71"
-     "\x03\x00"
-     "\xBB\x79\x03\x00\x34\xE7\x00\x80\x38\xE7\x00\x80\xBE\x31\x03\x00\xBF\x31"
-     "\x03\x00"
-     "\xBC\x51\x03\x00\xBD\x51\x03\x00\xB8\xD1\x03\x00\xB9\xD1\x03\x00\xBA\xD5"
-     "\x03\x00"
-     "\xBB\xE9\x03\x00\xBC\xFD\x03\x00\xBD\xE5\x03\x00\xBE\xED\x03\x00\xBF\xE5"
-     "\x03\x00"
-     "\xB0\xB1\x03\x00\xB1\xB9\x03\x00\xB2\x81\x03\x00\xB3\x81\x03\x00\xB4\xF1"
-     "\x03\x00"
-     "\xB5\xF1\x03\x00\xB6\xF1\x03\x00\xB7\xF1\x03\x00\xA8\x01\x02\x00\xA9\x11"
-     "\x02\x00"
-     "\xAA\x11\x02\x00\xAB\x25\x02\x00\xAC\x3D\x02\x00\xAD\xD1\x03\x00\xAE\xD1"
-     "\x03\x00"
-     "\xAF\xD1\x03\x00\x3C\xE7\x00\x80\xBC\xE5\x00\x80\x40\xE7\x00\x80\x44\xE7"
-     "\x00\x80"
-     "\x48\xE7\x00\x80\x4C\xE7\x00\x80\x50\xE7\x00\x80\x54\xE7\x00\x80\x58\xE7"
-     "\x00\x80"
-     "\x5C\xE7\x00\x80\x60\xE7\x00\x80\x64\xE7\x00\x80\xEF\x50\x00\x00\x68\xE7"
-     "\x00\x80"
-     "\x6C\xE7\x00\x80\x70\xE7\x00\x80\xE3\x84\x00\x00\x74\xE7\x00\x80\xE1\xF8"
-     "\x01\x00"
-     "\x78\xE7\x00\x80\x80\x2D\x00\x00\x81\x11\x00\x00\x82\x15\x00\x00\x80\xE7"
-     "\x00\x80"
-     "\xA3\x99\x03\x00\x84\xE7\x00\x80\x88\xE7\x00\x80\x8C\xE7\x00\x80\x90\xE7"
-     "\x00\x80"
-     "\xA6\x51\x02\x00\xA5\xB5\x03\x00\x94\xE7\x00\x80\xAB\x7D\x02\x00\xAA\x75"
-     "\x02\x00"
-     "\x98\xE7\x00\x80\x9C\xE7\x00\x80\xAF\x35\x02\x00\xAE\x35\x02\x00\xAD\x55"
-     "\x02\x00"
-     "\xAC\x55\x02\x00\x86\xA0\x04\x00\x87\xA8\x05\x00\xA0\xE7\x00\x80\xA4\xE7"
-     "\x00\x80"
-     "\xA8\xE7\x00\x80\xAC\xE7\x00\x80\xB0\xE7\x00\x80\xEF\xF8\x01\x00\xBE\x2C"
-     "\x04\x00"
-     "\xE1\x80\x3E\x00\xB4\xE7\x00\x80\xE3\x14\x01\x00\xB8\xE7\x00\x80\xBC\xE7"
-     "\x00\x80"
-     "\xC0\xE7\x00\x80\xC4\xE7\x00\x80\xC8\xE7\x00\x80\xCC\xE7\x00\x80\xD0\xE7"
-     "\x00\x80"
-     "\xB3\xDD\x01\x00\xD4\xE7\x00\x80\xB5\xFD\x01\x00\xB6\xF5\x01\x00\xD8\xE7"
-     "\x00\x80"
-     "\xDC\xE7\x00\x80\xE0\xE7\x00\x80\xBA\xB1\x01\x00\xBB\x85\x01\x00\xBC\x9D"
-     "\x01\x00"
-     "\xBD\x15\x01\x00\xBE\x11\x01\x00\xBF\x11\x01\x00\xA8\xBD\x05\x00\xA9\x01"
-     "\x06\x00"
-     "\xAA\x19\x06\x00\xAB\x11\x06\x00\xAC\x39\x06\x00\xAD\x39\x06\x00\xAE\x59"
-     "\x06\x00"
-     "\xAF\x49\x06\x00\x7C\xE7\x00\x80\x82\x11\x00\x00\x81\x19\x00\x00\x80\xC9"
-     "\x07\x00"
-     "\xE4\xE7\x00\x80\xE8\xE7\x00\x80\x84\x30\x02\x00\xEC\xE7\x00\x80\xB8\xB9"
-     "\x06\x00"
-     "\xB9\xB9\x06\x00\xBA\x49\x07\x00\xBB\x49\x07\x00\xBC\x59\x07\x00\xBD\x59"
-     "\x07\x00"
-     "\xBE\x4D\x07\x00\xBF\x45\x07\x00\xB0\x39\x06\x00\xB1\x0D\x06\x00\xB2\x05"
-     "\x06\x00"
-     "\xB3\x8D\x06\x00\xB4\x99\x06\x00\xB5\x99\x06\x00\xB6\x89\x06\x00\xB7\x89"
-     "\x06\x00"
-     "\xA3\x91\x06\x00\xF0\xE7\x00\x80\x86\x28\x00\x00\x87\xFC\x03\x00\xF4\xE7"
-     "\x00\x80"
-     "\xA6\xB9\x06\x00\xA5\xB1\x06\x00\xF8\xE7\x00\x80\xAB\xC9\x06\x00\xAA\xFD"
-     "\x06\x00"
-     "\xFC\xE7\x00\x80\x00\xE8\x00\x80\xAF\x5D\x06\x00\xAE\x5D\x06\x00\xAD\x59"
-     "\x06\x00"
-     "\xAC\xD1\x06\x00\x04\xE8\x00\x80\xB3\x8D\x07\x00\x08\xE8\x00\x80\x0C\xE8"
-     "\x00\x80"
-     "\xB6\x8D\x07\x00\x10\xE8\x00\x80\x14\xE8\x00\x80\xB5\x8D\x07\x00\xBA\x69"
-     "\x07\x00"
-     "\xBB\x69\x07\x00\x18\xE8\x00\x80\x1C\xE8\x00\x80\xBE\x69\x07\x00\xBF\x5D"
-     "\x07\x00"
-     "\xBC\x69\x07\x00\xBD\x69\x07\x00\xA3\x49\x06\x00\x20\xE8\x00\x80\x24\xE8"
-     "\x00\x80"
-     "\x28\xE8\x00\x80\x2C\xE8\x00\x80\xA6\x49\x06\x00\xA5\x49\x06\x00\x30\xE8"
-     "\x00\x80"
-     "\xAB\xAD\x06\x00\xAA\xAD\x06\x00\x34\xE8\x00\x80\x38\xE8\x00\x80\xAF\x99"
-     "\x06\x00"
-     "\xAE\xAD\x06\x00\xAD\xAD\x06\x00\xAC\xAD\x06\x00\x3C\xE8\x00\x80\x40\xE8"
-     "\x00\x80"
-     "\x44\xE8\x00\x80\x48\xE8\x00\x80\x4C\xE8\x00\x80\xEF\xE0\x05\x00\x50\xE8"
-     "\x00\x80"
-     "\x54\xE8\x00\x80\x58\xE8\x00\x80\xE3\x3C\x05\x00\x84\x24\x01\x00\xE1\x7C"
-     "\x05\x00"
-     "\x80\x29\x00\x00\x81\x29\x00\x00\x82\x19\x00\x00\xBE\x34\x01\x00\xA8\xBD"
-     "\x06\x00"
-     "\xA9\xC1\x06\x00\xAA\xC1\x06\x00\xAB\xC1\x06\x00\xAC\xC1\x06\x00\xAD\xC9"
-     "\x06\x00"
-     "\xAE\xF1\x06\x00\xAF\xF1\x06\x00\x86\x00\x00\x00\x87\x84\x00\x00\x60\xE8"
-     "\x00\x80"
-     "\x64\xE8\x00\x80\x68\xE8\x00\x80\x6C\xE8\x00\x80\x70\xE8\x00\x80\x74\xE8"
-     "\x00\x80"
-     "\xB8\x79\x06\x00\xB9\x79\x06\x00\xBA\x09\x06\x00\xBB\x09\x06\x00\xBC\x19"
-     "\x06\x00"
-     "\xBD\x05\x06\x00\xBE\x05\x06\x00\xBF\x3D\x06\x00\xB0\x69\x06\x00\xB1\x69"
-     "\x06\x00"
-     "\xB2\x79\x06\x00\xB3\x79\x06\x00\xB4\x69\x06\x00\xB5\x5D\x06\x00\xB6\x49"
-     "\x06\x00"
-     "\xB7\x49\x06\x00\xA8\x85\x02\x00\xA9\x95\x02\x00\xAA\x95\x02\x00\xAB\xA5"
-     "\x02\x00"
-     "\xAC\xBD\x02\x00\xAD\xD5\x02\x00\xAE\xD1\x02\x00\xAF\xD1\x02\x00\x78\xE8"
-     "\x00\x80"
-     "\x7C\xE8\x00\x80\x84\x28\x03\x00\xB5\xE5\x09\x00\x80\xE8\x00\x80\x84\xE8"
-     "\x00\x80"
-     "\x88\xE8\x00\x80\x8C\xE8\x00\x80\xB8\x5D\x03\x00\xB9\x6D\x03\x00\xBA\x65"
-     "\x03\x00"
-     "\xBB\x0D\x03\x00\xBC\x15\x03\x00\xBD\x1D\x03\x00\xBE\x15\x03\x00\xBF\x0D"
-     "\x03\x00"
-     "\xB0\xB5\x02\x00\xB1\xBD\x02\x00\xB2\x81\x02\x00\xB3\x81\x02\x00\xB4\x75"
-     "\x03\x00"
-     "\xB5\x7D\x03\x00\xB6\x75\x03\x00\xB7\x6D\x03\x00\xB3\x1D\x02\x00\x90\xE8"
-     "\x00\x80"
-     "\x94\xE8\x00\x80\x98\xE8\x00\x80\xBE\x1C\x0C\x00\xB6\x55\x02\x00\xB5\x35"
-     "\x02\x00"
-     "\x9C\xE8\x00\x80\xBB\x79\x02\x00\xBA\x71\x02\x00\xA0\xE8\x00\x80\xA4\xE8"
-     "\x00\x80"
-     "\xBF\xB9\x03\x00\xBE\xB1\x03\x00\xBD\x55\x02\x00\xBC\x55\x02\x00\xA8\xE8"
-     "\x00\x80"
-     "\xA3\x59\x02\x00\xAC\xE8\x00\x80\xB0\xE8\x00\x80\xA6\x11\x02\x00\xB4\xE8"
-     "\x00\x80"
-     "\xB8\xE8\x00\x80\xA5\x71\x02\x00\xAA\x35\x02\x00\xAB\x3D\x02\x00\x0C\xE7"
-     "\x00\x80"
-     "\xBC\xE8\x00\x80\xAE\xF5\x03\x00\xAF\xFD\x03\x00\xAC\x11\x02\x00\xAD\x11"
-     "\x02\x00"
-     "\xA8\xA9\x02\x00\xA9\xA9\x02\x00\xAA\xB9\x02\x00\xAB\xB9\x02\x00\xAC\xA9"
-     "\x02\x00"
-     "\xAD\xA9\x02\x00\xAE\x39\x01\x00\xAF\x39\x01\x00\x80\x0D\x00\x00\x81\x15"
-     "\x00\x00"
-     "\x82\x1D\x00\x00\xC0\xE8\x00\x80\xC4\xE8\x00\x80\xC8\xE8\x00\x80\x87\x68"
-     "\x0D\x00"
-     "\x86\x1C\x0D\x00\xB8\xE9\x01\x00\xB9\xE9\x01\x00\xBA\x89\x01\x00\xBB\x85"
-     "\x01\x00"
-     "\xBC\x9D\x01\x00\xBD\x81\x01\x00\xBE\x81\x01\x00\xBF\xB1\x01\x00\xB0\x51"
-     "\x01\x00"
-     "\xB1\x51\x01\x00\xB2\x51\x01\x00\xB3\x51\x01\x00\xB4\xF5\x01\x00\xB5\xFD"
-     "\x01\x00"
-     "\xB6\xF5\x01\x00\xB7\xE1\x01\x00\xEF\x04\x00\x00\x84\x48\x0C\x00\xD0\xE8"
-     "\x00\x80"
-     "\xBE\x80\x0C\x00\xD4\xE8\x00\x80\xD8\xE8\x00\x80\xDC\xE8\x00\x80\xEF\xC4"
-     "\x06\x00"
-     "\xE0\xE8\x00\x80\xE1\x3C\x06\x00\xE4\xE8\x00\x80\xE3\x84\x01\x00\xE3\xC8"
-     "\x00\x00"
-     "\xE8\xE8\x00\x80\xE1\x28\x01\x00\xEC\xE8\x00\x80\xF0\xE8\x00\x80\xF4\xE8"
-     "\x00\x80"
-     "\xF8\xE8\x00\x80\xB3\x7D\x02\x00\xFC\xE8\x00\x80\xB5\x6D\x02\x00\xB6\x65"
-     "\x02\x00"
-     "\x00\xE9\x00\x80\x04\xE9\x00\x80\x08\xE9\x00\x80\xBA\xD9\x01\x00\xBB\xD9"
-     "\x01\x00"
-     "\xBC\xC9\x01\x00\xBD\xC9\x01\x00\xBE\x7D\x01\x00\xBF\x75\x01\x00\xA3\xB1"
-     "\x0D\x00"
-     "\xCC\xE8\x00\x80\x0C\xE9\x00\x80\x10\xE9\x00\x80\x14\xE9\x00\x80\xA6\xA9"
-     "\x0D\x00"
-     "\xA5\xA1\x0D\x00\x18\xE9\x00\x80\xAB\x15\x0E\x00\xAA\x15\x0E\x00\x1C\xE9"
-     "\x00\x80"
-     "\x20\xE9\x00\x80\xAF\xB9\x0E\x00\xAE\xB1\x0E\x00\xAD\x05\x0E\x00\xAC\x05"
-     "\x0E\x00"
-     "\x24\xE9\x00\x80\x28\xE9\x00\x80\x2C\xE9\x00\x80\x30\xE9\x00\x80\x80\x1D"
-     "\x00\x00"
-     "\x81\x09\x00\x00\x82\x39\x00\x00\x34\xE9\x00\x80\x38\xE9\x00\x80\x3C\xE9"
-     "\x00\x80"
-     "\x86\xC8\x00\x00\x87\x0C\x03\x00\x40\xE9\x00\x80\x44\xE9\x00\x80\x48\xE9"
-     "\x00\x80"
-     "\x4C\xE9\x00\x80\xA8\x09\x0F\x00\xA9\x15\x0F\x00\xAA\x1D\x0F\x00\xAB\x15"
-     "\x0F\x00"
-     "\xAC\x31\x0F\x00\xAD\x3D\x0F\x00\xAE\x2D\x0F\x00\xAF\x21\x0F\x00\x84\xA8"
-     "\x03\x00"
-     "\x50\xE9\x00\x80\x54\xE9\x00\x80\x58\xE9\x00\x80\x5C\xE9\x00\x80\x60\xE9"
-     "\x00\x80"
-     "\x64\xE9\x00\x80\x68\xE9\x00\x80\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xD5"
-     "\x00\x00"
-     "\xBB\xE5\x00\x00\xBC\xFD\x00\x00\xBD\x95\x00\x00\xBE\x9D\x00\x00\xBF\x95"
-     "\x00\x00"
-     "\xB0\x21\x0F\x00\xB1\x21\x0F\x00\xB2\x21\x0F\x00\xB3\x3D\x0F\x00\xB4\x2D"
-     "\x0F\x00"
-     "\xB5\x15\x0F\x00\xB6\x1D\x0F\x00\xB7\xF5\x00\x00\xB3\x95\x0E\x00\x6C\xE9"
-     "\x00\x80"
-     "\x70\xE9\x00\x80\x74\xE9\x00\x80\x78\xE9\x00\x80\xB6\xBD\x0E\x00\xB5\xB9"
-     "\x0E\x00"
-     "\x7C\xE9\x00\x80\xBB\xE5\x0E\x00\xBA\xE5\x0E\x00\x80\xE9\x00\x80\x84\xE9"
-     "\x00\x80"
-     "\xBF\xD1\x0E\x00\xBE\xE9\x0E\x00\xBD\xE9\x0E\x00\xBC\xF5\x0E\x00\xEF\x28"
-     "\x0D\x00"
-     "\x88\xE9\x00\x80\x8C\xE9\x00\x80\x90\xE9\x00\x80\x94\xE9\x00\x80\x98\xE9"
-     "\x00\x80"
-     "\x9C\xE9\x00\x80\xA0\xE9\x00\x80\x80\x3D\x00\x00\x81\x09\x00\x00\x82\x1D"
-     "\x00\x00"
-     "\xA4\xE9\x00\x80\xE3\x48\x0D\x00\xA8\xE9\x00\x80\xE1\xC0\x0E\x00\xBE\x7C"
-     "\x01\x00"
-     "\xA3\x15\x0E\x00\xB0\xE9\x00\x80\x86\x28\x00\x00\x87\x4C\x01\x00\xB4\xE9"
-     "\x00\x80"
-     "\xA6\x3D\x0E\x00\xA5\x39\x0E\x00\xB8\xE9\x00\x80\xAB\x65\x0E\x00\xAA\x65"
-     "\x0E\x00"
-     "\xBC\xE9\x00\x80\xC0\xE9\x00\x80\xAF\x51\x0E\x00\xAE\x69\x0E\x00\xAD\x69"
-     "\x0E\x00"
-     "\xAC\x75\x0E\x00\xC4\xE9\x00\x80\xB3\xBD\x01\x00\xC8\xE9\x00\x80\xCC\xE9"
-     "\x00\x80"
-     "\xB6\x79\x01\x00\xD0\xE9\x00\x80\xD4\xE9\x00\x80\xB5\x79\x01\x00\xBA\x55"
-     "\x01\x00"
-     "\xBB\x5D\x01\x00\xD8\xE9\x00\x80\xDC\xE9\x00\x80\xBE\xF9\x00\x00\xBF\xF9"
-     "\x00\x00"
-     "\xBC\x45\x01\x00\xBD\xF9\x00\x00\xA8\x71\x02\x00\xA9\x71\x02\x00\xAA\x71"
-     "\x02\x00"
-     "\xAB\x71\x02\x00\xAC\xB5\x02\x00\xAD\xBD\x02\x00\xAE\xB5\x02\x00\xAF\xAD"
-     "\x02\x00"
-     "\x84\xEC\x04\x00\xE0\xE9\x00\x80\xE4\xE9\x00\x80\xE8\xE9\x00\x80\xEC\xE9"
-     "\x00\x80"
-     "\xF0\xE9\x00\x80\xF4\xE9\x00\x80\xF8\xE9\x00\x80\xB8\x69\x03\x00\xB9\x69"
-     "\x03\x00"
-     "\xBA\x0D\x03\x00\xBB\x05\x03\x00\xBC\x1D\x03\x00\xBD\x05\x03\x00\xBE\x0D"
-     "\x03\x00"
-     "\xBF\x05\x03\x00\xB0\xD5\x02\x00\xB1\xDD\x02\x00\xB2\xD5\x02\x00\xB3\x69"
-     "\x03\x00"
-     "\xB4\x79\x03\x00\xB5\x79\x03\x00\xB6\x69\x03\x00\xB7\x61\x03\x00\xFC\xE9"
-     "\x00\x80"
-     "\x00\xEA\x00\x80\x04\xEA\x00\x80\xA3\xF5\x02\x00\x08\xEA\x00\x80\xA5\x31"
-     "\x02\x00"
-     "\xA6\x31\x02\x00\x0C\xEA\x00\x80\x10\xEA\x00\x80\x14\xEA\x00\x80\xAA\x1D"
-     "\x02\x00"
-     "\xAB\x15\x02\x00\xAC\x0D\x02\x00\xAD\xB1\x03\x00\xAE\xB1\x03\x00\xAF\xB1"
-     "\x03\x00"
-     "\x80\x61\x00\x00\x81\x61\x00\x00\x82\x05\x00\x00\x18\xEA\x00\x80\x86\xF0"
-     "\x04\x00"
-     "\x87\x60\x03\x00\xBE\x10\x04\x00\x20\xEA\x00\x80\x24\xEA\x00\x80\x28\xEA"
-     "\x00\x80"
-     "\x2C\xEA\x00\x80\x30\xEA\x00\x80\x34\xEA\x00\x80\x38\xEA\x00\x80\x3C\xEA"
-     "\x00\x80"
-     "\x5C\xE8\x00\x80\xA8\x85\x02\x00\xA9\x95\x02\x00\xAA\x95\x02\x00\xAB\xA5"
-     "\x02\x00"
-     "\xAC\xBD\x02\x00\xAD\xD5\x02\x00\xAE\xD1\x02\x00\xAF\xD1\x02\x00\x40\xEA"
-     "\x00\x80"
-     "\x44\xEA\x00\x80\x48\xEA\x00\x80\x4C\xEA\x00\x80\x50\xEA\x00\x80\x54\xEA"
-     "\x00\x80"
-     "\x58\xEA\x00\x80\x5C\xEA\x00\x80\xB8\x75\x01\x00\xB9\x7D\x01\x00\xBA\x75"
-     "\x01\x00"
-     "\xBB\xCD\x01\x00\xBC\xD5\x01\x00\xBD\xDD\x01\x00\xBE\xC9\x01\x00\xBF\xC1"
-     "\x01\x00"
-     "\xB0\xB5\x02\x00\xB1\xBD\x02\x00\xB2\x81\x02\x00\xB3\x81\x02\x00\xB4\x55"
-     "\x01\x00"
-     "\xB5\x5D\x01\x00\xB6\x55\x01\x00\xB7\x4D\x01\x00\xE1\xA4\x0E\x00\x84\x08"
-     "\x04\x00"
-     "\xE3\x1C\x0E\x00\x60\xEA\x00\x80\xBE\x9C\x04\x00\x64\xEA\x00\x80\x68\xEA"
-     "\x00\x80"
-     "\x6C\xEA\x00\x80\x70\xEA\x00\x80\x74\xEA\x00\x80\x78\xEA\x00\x80\x7C\xEA"
-     "\x00\x80"
-     "\x81\x75\x00\x00\x80\x75\x00\x00\xEF\xCC\x0E\x00\x82\x75\x00\x00\x80\xEA"
-     "\x00\x80"
-     "\x84\xEA\x00\x80\x88\xEA\x00\x80\x84\xC4\x04\x00\x90\xEA\x00\x80\x94\xEA"
-     "\x00\x80"
-     "\xEF\x5C\x00\x00\x98\xEA\x00\x80\xE1\x90\x01\x00\x9C\xEA\x00\x80\xE3\x58"
-     "\x01\x00"
-     "\xA0\xEA\x00\x80\xA4\xEA\x00\x80\xA8\xEA\x00\x80\x86\x58\x07\x00\x87\xEC"
-     "\x04\x00"
-     "\xB3\x9D\x05\x00\x1C\xEA\x00\x80\x8C\xEA\x00\x80\xAC\xEA\x00\x80\xB0\xEA"
-     "\x00\x80"
-     "\xB6\xD5\x05\x00\xB5\xB5\x05\x00\xB4\xEA\x00\x80\xBB\xF9\x05\x00\xBA\xF1"
-     "\x05\x00"
-     "\xB8\xEA\x00\x80\xBC\xEA\x00\x80\xBF\x19\x06\x00\xBE\x11\x06\x00\xBD\xD5"
-     "\x05\x00"
-     "\xBC\xD5\x05\x00\xC0\xEA\x00\x80\xA3\xD9\x05\x00\xC4\xEA\x00\x80\xC8\xEA"
-     "\x00\x80"
-     "\xA6\x91\x05\x00\xCC\xEA\x00\x80\xD0\xEA\x00\x80\xA5\xF1\x05\x00\xAA\xB5"
-     "\x05\x00"
-     "\xAB\xBD\x05\x00\xD4\xEA\x00\x80\xD8\xEA\x00\x80\xAE\x55\x06\x00\xAF\x5D"
-     "\x06\x00"
-     "\xAC\x91\x05\x00\xAD\x91\x05\x00\xA8\x5D\x06\x00\xA9\x65\x06\x00\xAA\x6D"
-     "\x06\x00"
-     "\xAB\x65\x06\x00\xAC\x7D\x06\x00\xAD\x65\x06\x00\xAE\x6D\x06\x00\xAF\x65"
-     "\x06\x00"
-     "\xBE\xA0\x00\x00\xDC\xEA\x00\x80\xE0\xEA\x00\x80\xE4\xEA\x00\x80\xE8\xEA"
-     "\x00\x80"
-     "\xEC\xEA\x00\x80\xF0\xEA\x00\x80\xF4\xEA\x00\x80\xB8\x4D\x07\x00\xB9\x51"
-     "\x07\x00"
-     "\xBA\x51\x07\x00\xBB\x51\x07\x00\xBC\x71\x07\x00\xBD\x71\x07\x00\xBE\x71"
-     "\x07\x00"
-     "\xBF\x71\x07\x00\xB0\xC1\x07\x00\xB1\xC1\x07\x00\xB2\xC1\x07\x00\xB3\xC1"
-     "\x07\x00"
-     "\xB4\xC1\x07\x00\xB5\xC1\x07\x00\xB6\xC1\x07\x00\xB7\xC1\x07\x00\xF8\xEA"
-     "\x00\x80"
-     "\xE1\x94\x06\x00\x00\xEB\x00\x80\xE3\x3C\x05\x00\x80\xA9\x01\x00\x81\xA9"
-     "\x01\x00"
-     "\x82\x19\x00\x00\x04\xEB\x00\x80\x86\x80\x00\x00\x87\x04\x01\x00\x08\xEB"
-     "\x00\x80"
-     "\x0C\xEB\x00\x80\x10\xEB\x00\x80\x14\xEB\x00\x80\x18\xEB\x00\x80\xEF\x1C"
-     "\x05\x00"
-     "\xB3\xA5\x07\x00\xFC\xEA\x00\x80\x1C\xEB\x00\x80\x20\xEB\x00\x80\x24\xEB"
-     "\x00\x80"
-     "\xB6\xA5\x07\x00\xB5\xB5\x07\x00\x28\xEB\x00\x80\xBB\x45\x06\x00\xBA\x45"
-     "\x06\x00"
-     "\x2C\xEB\x00\x80\x30\xEB\x00\x80\xBF\x45\x06\x00\xBE\x45\x06\x00\xBD\x55"
-     "\x06\x00"
-     "\xBC\x55\x06\x00\xA3\x65\x07\x00\x34\xEB\x00\x80\x38\xEB\x00\x80\x3C\xEB"
-     "\x00\x80"
-     "\x40\xEB\x00\x80\xA6\x65\x07\x00\xA5\x75\x07\x00\x44\xEB\x00\x80\xAB\x85"
-     "\x06\x00"
-     "\xAA\x85\x06\x00\x48\xEB\x00\x80\x4C\xEB\x00\x80\xAF\x85\x06\x00\xAE\x85"
-     "\x06\x00"
-     "\xAD\x95\x06\x00\xAC\x95\x06\x00\x50\xEB\x00\x80\xB3\x19\x06\x00\x54\xEB"
-     "\x00\x80"
-     "\x58\xEB\x00\x80\xB6\x21\x06\x00\x5C\xEB\x00\x80\x60\xEB\x00\x80\xB5\x31"
-     "\x06\x00"
-     "\xBA\xF1\x01\x00\xBB\xF9\x01\x00\x64\xEB\x00\x80\x68\xEB\x00\x80\xBE\xB1"
-     "\x01\x00"
-     "\xBF\xBD\x01\x00\xBC\xD1\x01\x00\xBD\xD1\x01\x00\xA8\xD1\x05\x00\xA9\xD1"
-     "\x05\x00"
-     "\xAA\xD5\x05\x00\xAB\xE9\x05\x00\xAC\xFD\x05\x00\xAD\xE5\x05\x00\xAE\xED"
-     "\x05\x00"
-     "\xAF\x11\x02\x00\x80\xF9\x03\x00\x81\xCD\x03\x00\x82\xC5\x03\x00\x84\x20"
-     "\x03\x00"
-     "\xBE\x00\x1C\x00\x70\xEB\x00\x80\x87\x04\x03\x00\x86\xBC\x1C\x00\xB8\x19"
-     "\x02\x00"
-     "\xB9\x2D\x02\x00\xBA\x25\x02\x00\xBB\xE9\x02\x00\xBC\xF9\x02\x00\xBD\xF9"
-     "\x02\x00"
-     "\xBE\xE9\x02\x00\xBF\xE9\x02\x00\xB0\x71\x02\x00\xB1\x79\x02\x00\xB2\x41"
-     "\x02\x00"
-     "\xB3\x41\x02\x00\xB4\x31\x02\x00\xB5\x3D\x02\x00\xB6\x35\x02\x00\xB7\x29"
-     "\x02\x00"
-     "\xAC\xE9\x00\x80\x74\xEB\x00\x80\x78\xEB\x00\x80\xA3\x51\x05\x00\x7C\xEB"
-     "\x00\x80"
-     "\xA5\x79\x05\x00\xA6\x69\x05\x00\x80\xEB\x00\x80\x84\xEB\x00\x80\x88\xEB"
-     "\x00\x80"
-     "\xAA\xB9\x02\x00\xAB\xB1\x02\x00\xAC\x99\x02\x00\xAD\x99\x02\x00\xAE\xF9"
-     "\x02\x00"
-     "\xAF\xF5\x02\x00\xE1\xF8\x01\x00\x8C\xEB\x00\x80\xE3\xF4\x02\x00\x84\xC0"
-     "\x1C\x00"
-     "\x90\xEB\x00\x80\x94\xEB\x00\x80\x98\xEB\x00\x80\x9C\xEB\x00\x80\xA0\xEB"
-     "\x00\x80"
-     "\xA4\xEB\x00\x80\xA8\xEB\x00\x80\xAC\xEB\x00\x80\xB0\xEB\x00\x80\xB4\xEB"
-     "\x00\x80"
-     "\xEF\x30\x02\x00\xB8\xEB\x00\x80\xA8\x09\x02\x00\xA9\x19\x02\x00\xAA\x21"
-     "\x02\x00"
-     "\xAB\x21\x02\x00\xAC\x21\x02\x00\xAD\x21\x02\x00\xAE\x21\x02\x00\xAF\x55"
-     "\x02\x00"
-     "\xBC\xEB\x00\x80\xC0\xEB\x00\x80\xC4\xEB\x00\x80\xC8\xEB\x00\x80\xCC\xEB"
-     "\x00\x80"
-     "\x82\xAD\x00\x00\x81\xAD\x00\x00\x80\xAD\x00\x00\xB8\xD1\x02\x00\xB9\xD9"
-     "\x02\x00"
-     "\xBA\xE1\x02\x00\xBB\xE1\x02\x00\xBC\x91\x02\x00\xBD\x9D\x02\x00\xBE\x95"
-     "\x02\x00"
-     "\xBF\x89\x02\x00\xB0\x2D\x02\x00\xB1\x35\x02\x00\xB2\x35\x02\x00\xB3\x05"
-     "\x02\x00"
-     "\xB4\x1D\x02\x00\xB5\xF1\x02\x00\xB6\xF1\x02\x00\xB7\xF1\x02\x00\xE1\x48"
-     "\x07\x00"
-     "\xE3\x58\x06\x00\xE3\x1C\x06\x00\xE1\x6C\x06\x00\xD0\xEB\x00\x80\xD4\xEB"
-     "\x00\x80"
-     "\xD8\xEB\x00\x80\xBE\x10\x1C\x00\x4E\x1C\x00\x80\x52\x1C\x00\x80\x56\x1C"
-     "\x00\x80"
-     "\x5A\x1C\x00\x80\x5E\x1C\x00\x80\xEF\xEC\x01\x00\xEF\xD0\x06\x00\x62\x1C"
-     "\x00\x80"
-     "\xB3\xE9\x02\x00\x66\x1C\x00\x80\x6A\x1C\x00\x80\xB5\x1D\x00\x00\x6E\x1C"
-     "\x00\x80"
-     "\xB6\xE1\x02\x00\xB5\xE9\x02\x00\x72\x1C\x00\x80\xBB\xB1\x02\x00\xBA\xA9"
-     "\x02\x00"
-     "\xAC\x90\x1D\x00\x76\x1C\x00\x80\xBF\x75\x02\x00\xBE\x99\x02\x00\xBD\x91"
-     "\x02\x00"
-     "\xBC\xA1\x02\x00\x6C\xEB\x00\x80\xE1\x1C\xCD\x03\xBC\xB9\x1F\x00\xBD\xB1"
-     "\x1F\x00"
-     "\xBA\xB9\x1F\x00\xBB\xB9\x1F\x00\xB8\xB9\x1F\x00\xB9\xB9\x1F\x00\xA8\x15"
-     "\x1E\x00"
-     "\xA9\xA1\x1E\x00\xAA\xA1\x1E\x00\xAB\xBD\x1E\x00\xAC\xAD\x1E\x00\xAD\x95"
-     "\x1E\x00"
-     "\xAE\x9D\x1E\x00\xAF\xF9\x1E\x00\xB0\x91\x1E\x00\xB1\x91\x1E\x00\xB2\x91"
-     "\x1E\x00"
-     "\xB3\x91\x1E\x00\xB4\xB1\x1E\x00\xB5\xB1\x1E\x00\xB6\xB1\x1E\x00\xB7\xB1"
-     "\x1E\x00"
-     "\xB8\x91\x1E\x00\xB9\x91\x1E\x00\xBA\x95\x1E\x00\xBB\x4D\x1F\x00\xBC\x55"
-     "\x1F\x00"
-     "\xBD\x5D\x1F\x00\xBE\x55\x1F\x00\xBF\x4D\x1F\x00\xAC\xDD\x1D\x00\xAD\xED"
-     "\x1D\x00"
-     "\xAE\xE5\x1D\x00\xAF\x09\x1D\x00\xE1\x10\x24\x00\xE1\x94\x2D\x00\xAA\xD5"
-     "\x1D\x00"
-     "\xAB\xCD\x1D\x00\xE1\xF8\x1D\x00\xA5\x95\x1D\x00\xA6\x9D\x1D\x00\xB5\x71"
-     "\x1F\x00"
-     "\xE1\xB4\x02\x00\xAC\xD8\x02\x00\xB5\x55\x1F\x00\xA3\x95\x1D\x00\xB3\x49"
-     "\x1E\x00"
-     "\xE1\xB8\x02\x00\xB5\x19\x00\x00\xAC\xD0\x02\x00\x7A\x1C\x00\x80\xB6\x75"
-     "\x1E\x00"
-     "\xB5\x45\x1E\x00\xE1\x84\x02\x00\xBB\x35\x1E\x00\xBA\x35\x1E\x00\xAC\xD0"
-     "\x02\x00"
-     "\xB5\xE5\x01\x00\xBF\x15\x1E\x00\xBE\x15\x1E\x00\xBD\x25\x1E\x00\xBC\x25"
-     "\x1E\x00"
-     "\xA3\x89\x1F\x00\x8F\x35\x25\x00\x7E\x1C\x00\x80\xE1\xAC\x02\x00\xAC\xD8"
-     "\x02\x00"
-     "\xA6\xB5\x1F\x00\xA5\x85\x1F\x00\xB5\x35\x00\x00\xAB\xF5\x1F\x00\xAA\xF5"
-     "\x1F\x00"
-     "\x82\x1C\x00\x80\xE1\x1C\x02\x00\xAF\xD5\x1F\x00\xAE\xD5\x1F\x00\xAD\xE5"
-     "\x1F\x00"
-     "\xAC\xE5\x1F\x00\x9E\x25\x19\x00\x9F\xC5\x1F\x00\x9C\xB5\x1B\x00\x9D\x49"
-     "\x19\x00"
-     "\x9A\x71\x15\x00\x9B\x69\x15\x00\x98\xF5\x11\x00\x99\xB5\x17\x00\x96\x85"
-     "\x13\x00"
-     "\x97\x45\x11\x00\x94\x49\x2D\x00\x95\x71\x2C\x00\x92\xF5\x28\x00\x93\xA5"
-     "\x2F\x00"
-     "\x90\x79\x29\x00\x91\x79\x29\x00\xAC\xD0\x02\x00\xB5\x0D\x00\x00\x86\x1C"
-     "\x00\x80"
-     "\xE1\xAC\x02\x00\xAC\xDC\x02\x00\x8A\x1C\x00\x80\xB5\x0D\x00\x00\x82\xB8"
-     "\x03\x00"
-     "\xE1\xB8\x02\x00\xB5\x09\x00\x00\xAC\xD8\x02\x00\x8E\x1C\x00\x80\xA2\xF5"
-     "\x1C\x00"
-     "\xA3\xB5\x01\x00\xA0\xBD\x1D\x00\xA1\xB5\x1D\x00\xC2\x0C\x00\x80\x82\x3B"
-     "\x00\x80"
-     "\xDE\x18\x00\x80\xE2\x18\x00\x80\xE6\x18\x00\x80\xEA\x18\x00\x80\xEE\x18"
-     "\x00\x80"
-     "\xF2\x18\x00\x80\xF6\x18\x00\x80\xFA\x18\x00\x80\xFE\x18\x00\x80\x02\x19"
-     "\x00\x80"
-     "\x06\x19\x00\x80\x0A\x19\x00\x80\x12\x19\x00\x80\x16\x19\x00\x80\x82\x31"
-     "\x26\x00"
-     "\x83\x05\x2A\x00\xCC\xE4\x02\x00\xB8\x61\x09\x00\x86\xC1\x2C\x00\x87\x35"
-     "\x2E\x00"
-     "\x84\x1D\x2A\x00\x85\x99\x2B\x00\x8A\x3D\x2F\x00\x8B\xC1\x10\x00\x1A\x19"
-     "\x00\x80"
-     "\x1E\x19\x00\x80\x8E\xD9\x14\x00\x8F\x15\x16\x00\x8C\x19\x12\x00\x8D\x91"
-     "\x12\x00"
-     "\x92\x7D\x16\x00\x93\xB9\x18\x00\x22\x19\x00\x80\x26\x19\x00\x80\x96\xD1"
-     "\x1C\x00"
-     "\x97\x51\x1E\x00\x94\x5D\x1A\x00\x95\x39\x1A\x00\x9A\x69\x1E\x00\x9B\x21"
-     "\x1F\x00"
-     "\x2A\x19\x00\x80\xCC\xF0\x02\x00\x2E\x19\x00\x80\x32\x19\x00\x80\x9C\x79"
-     "\x02\x00"
-     "\xB8\x0D\x09\x00\xCC\xE8\x02\x00\x36\x19\x00\x80\xB8\xFD\x09\x00\x3A\x19"
-     "\x00\x80"
-     "\x3E\x19\x00\x80\x42\x19\x00\x80\x46\x19\x00\x80\x4A\x19\x00\x80\x4E\x19"
-     "\x00\x80"
-     "\x52\x19\x00\x80\xCC\xB0\x02\x00\x56\x19\x00\x80\x5A\x19\x00\x80\x5E\x19"
-     "\x00\x80"
-     "\x62\x19\x00\x80\x66\x19\x00\x80\x6A\x19\x00\x80\x6E\x19\x00\x80\x72\x19"
-     "\x00\x80"
-     "\x76\x19\x00\x80\x7A\x19\x00\x80\x7E\x19\x00\x80\x82\x19\x00\x80\x86\x19"
-     "\x00\x80"
-     "\x8A\x19\x00\x80\x8E\x19\x00\x80\x92\x19\x00\x80\xC2\x0A\x00\x80\xC5\x0A"
-     "\x00\x80"
-     "\x92\x1C\x00\x80\xB8\x81\x09\x00\xE3\x0A\x00\x80\x85\x80\x00\x00\x96\x1C"
-     "\x00\x80"
-     "\x87\x1C\x00\x00\x9A\x1C\x00\x80\xCC\xB0\x02\x00\xCC\xAC\x02\x00\xCC\x88"
-     "\x02\x00"
-     "\xA2\x1C\x00\x80\x8D\x14\x00\x00\xCC\x44\x02\x00\xA6\x1C\x00\x80\xC8\x0A"
-     "\x00\x80"
-     "\x89\x20\x00\x00\xCC\x40\x02\x00\x8B\x1C\x00\x00\xD1\x0A\x00\x80\xB8\x3D"
-     "\x08\x00"
-     "\xE9\x0A\x00\x80\xEF\x0A\x00\x80\xCC\x0C\x02\x00\x91\x10\x00\x00\xFB\x0A"
-     "\x00\x80"
-     "\xB8\xED\x08\x00\x0A\x0B\x00\x80\xAE\x1C\x00\x80\x88\x0B\x00\x80\x13\x0B"
-     "\x00\x80"
-     "\xB6\x1C\x00\x80\xB8\xC9\x08\x00\x8E\x0B\x00\x80\x1F\x0B\x00\x80\x1C\x0B"
-     "\x00\x80"
-     "\xB8\x41\x08\x00\xE1\xA0\x1F\x00\x25\x0B\x00\x80\xE3\x50\x1F\x00\x28\x0B"
-     "\x00\x80"
-     "\xB8\x45\x08\x00\x34\x0B\x00\x80\x4C\x0B\x00\x80\xB8\x79\x08\x00\xF0\x09"
-     "\x00\x80"
-     "\x9E\x1C\x00\x80\x70\x0B\x00\x80\x73\x0B\x00\x80\x2E\x1D\x00\x80\x76\x1D"
-     "\x00\x80"
-     "\xEF\xD8\x1E\x00\x86\x1D\x00\x80\x75\x06\x00\x80\xCC\x68\x04\x00\xAF\x08"
-     "\x00\x80"
-     "\xFE\x1D\x00\x80\x2E\x1E\x00\x80\x56\x1E\x00\x80\xB2\x1E\x00\x80\xB6\x1E"
-     "\x00\x80"
-     "\xBC\x6C\x01\x00\xBF\xE0\x01\x00\xBE\x44\x01\x00\xBA\x1E\x00\x80\xBE\x1E"
-     "\x00\x80"
-     "\xC2\x1E\x00\x80\xBA\xA4\x00\x00\xAC\x41\x34\x02\xAD\x19\x39\x02\xAE\x0D"
-     "\x30\x02"
-     "\x54\x08\x00\x80\x1E\x07\x00\x80\xAA\xED\x21\x02\xA9\xE9\x21\x02\xA8\x3D"
-     "\x28\x02"
-     "\xA7\xB1\x26\x02\xA6\xAD\x26\x02\xA5\xD1\x3D\x02\xA4\xC5\x3C\x02\xA3\x71"
-     "\x3C\x02"
-     "\xA2\x25\x00\x00\xA1\xC1\x21\x02\xA0\xCD\x21\x02\xBC\xAD\x31\x02\xBD\xB9"
-     "\x3B\x02"
-     "\xBE\x71\x3C\x02\xBF\x0D\x3C\x02\xB8\xD1\x31\x02\xB9\xF1\x31\x02\xBA\x99"
-     "\x31\x02"
-     "\xB8\x7D\x2A\x02\xB4\xF5\x30\x02\xB5\xD9\x30\x02\xB6\x2D\x31\x02\xB7\x31"
-     "\x31\x02"
-     "\xB0\x6D\x30\x02\xB1\x51\x30\x02\xB2\x09\x3A\x02\xB3\xFD\x30\x02\x8F\xDD"
-     "\x31\x02"
-     "\xC6\x1E\x00\x80\x8D\x09\x30\x02\x8C\x71\x06\x00\x8B\xE9\x3A\x02\x8A\xE1"
-     "\x3A\x02"
-     "\x89\xBD\x30\x02\x88\x81\x30\x02\x87\xA5\x30\x02\x86\x51\x06\x00\x85\x59"
-     "\x06\x00"
-     "\x84\x41\x06\x00\x83\xF1\x34\x02\x82\x6D\x39\x02\x81\x51\x3C\x02\x80\x1D"
-     "\x3C\x02"
-     "\x9F\x65\x26\x02\x9E\x55\x26\x02\x9D\x41\x26\x02\x9C\xE5\x07\x00\x9B\x29"
-     "\x3D\x02"
-     "\x9A\x7D\x3A\x02\x99\xC5\x07\x00\x98\x35\x3C\x02\x97\xED\x00\x00\x96\xDD"
-     "\x3C\x02"
-     "\x95\xE5\x07\x00\x94\xE9\x07\x00\x93\xA1\x3A\x02\x92\x19\x31\x02\x91\xF5"
-     "\x3A\x02"
-     "\x90\x59\x31\x02\x9B\x2D\x06\x00\xCA\x1E\x00\x80\xCE\x1E\x00\x80\xD2\x1E"
-     "\x00\x80"
-     "\x9F\xED\x07\x00\x9E\xE5\x39\x02\x9D\x25\x06\x00\x9C\x15\x31\x02\xD6\x1E"
-     "\x00\x80"
-     "\xDA\x1E\x00\x80\xDE\x1E\x00\x80\xE2\x1E\x00\x80\xF6\x1E\x00\x80\xFA\x1E"
-     "\x00\x80"
-     "\xFE\x1E\x00\x80\x02\x1F\x00\x80\xB8\x9D\x0E\x00\x06\x1F\x00\x80\x0A\x1F"
-     "\x00\x80"
-     "\x0E\x1F\x00\x80\x12\x1F\x00\x80\x16\x1F\x00\x80\x1A\x1F\x00\x80\x36\x1F"
-     "\x00\x80"
-     "\x3A\x1F\x00\x80\x52\x1F\x00\x80\x6E\x1F\x00\x80\x72\x1F\x00\x80\x76\x1F"
-     "\x00\x80"
-     "\x7A\x1F\x00\x80\x7E\x1F\x00\x80\x82\x1F\x00\x80\xBB\x99\x39\x02\xBA\xE9"
-     "\x25\x02"
-     "\xB9\x95\x3A\x02\xB8\x25\x06\x00\xBF\x2D\x26\x02\xBE\x9D\x25\x02\xBD\x29"
-     "\x06\x00"
-     "\xBC\x31\x06\x00\xB3\x3D\x06\x00\xB2\x3D\x06\x00\xB1\x19\x3B\x02\xB0\x3D"
-     "\x06\x00"
-     "\xB7\xA5\x3A\x02\xB6\xD1\x25\x02\xB5\x1D\x06\x00\xB4\x49\x3B\x02\xAB\xB5"
-     "\x06\x00"
-     "\xAA\xA9\x06\x00\xA9\xA1\x06\x00\xA8\xA9\x06\x00\xAF\x59\x06\x00\xAE\x79"
-     "\x3B\x02"
-     "\xAD\xA5\x06\x00\xAC\xA5\x06\x00\xA3\xB5\x06\x00\xA2\xAD\x06\x00\xA1\xAD"
-     "\x06\x00"
-     "\xA0\xBD\x31\x02\xA7\x91\x06\x00\xA6\xA9\x06\x00\xA5\x69\x3B\x02\xA4\x5D"
-     "\x3B\x02"
-     "\xB3\xC1\x04\x00\xB2\xB9\x04\x00\xB1\xB1\x04\x00\xB0\xB9\x04\x00\xB7\xC1"
-     "\x04\x00"
-     "\xB6\xD9\x04\x00\xB5\xD1\x04\x00\xB4\xD9\x04\x00\x86\x1F\x00\x80\xBA\xF9"
-     "\x04\x00"
-     "\xB9\xF1\x04\x00\xB8\xF9\x04\x00\xBF\xD9\x04\x00\xBE\xD1\x04\x00\xBD\xD9"
-     "\x04\x00"
-     "\x8A\x1F\x00\x80\xA3\x29\x04\x00\xA2\x09\x04\x00\xA1\x01\x04\x00\xA0\x61"
-     "\x04\x00"
-     "\xA7\xDD\x04\x00\xA6\xD1\x04\x00\xA5\x2D\x04\x00\xA4\x2D\x04\x00\xAB\xFD"
-     "\x04\x00"
-     "\xAA\xFD\x04\x00\xA9\xFD\x04\x00\xA8\xFD\x04\x00\xAF\xC1\x04\x00\xAE\xD9"
-     "\x04\x00"
-     "\xAD\xDD\x04\x00\xAC\xDD\x04\x00\x93\x55\x07\x00\x92\x55\x07\x00\x91\x45"
-     "\x07\x00"
-     "\x90\x45\x07\x00\x97\x75\x07\x00\x96\x75\x07\x00\x95\x45\x07\x00\x94\x45"
-     "\x07\x00"
-     "\x9B\x55\x07\x00\x9A\x55\x07\x00\x99\x45\x07\x00\x98\x45\x07\x00\x9F\x99"
-     "\x04\x00"
-     "\x9E\x49\x07\x00\x9D\x41\x07\x00\x9C\x19\x1E\x02\x83\x85\x28\x02\x82\x05"
-     "\x07\x00"
-     "\x81\x59\x2B\x02\x80\x45\x2B\x02\x87\xC9\x29\x02\x86\x59\x28\x02\x85\x91"
-     "\x28\x02"
-     "\x84\x95\x28\x02\x39\x00\x00\x80\x71\x00\x00\x80\x2B\x00\x00\x80\x88\x29"
-     "\x07\x00"
-     "\x3D\x00\x00\x80\x43\x00\x00\x80\x49\x00\x00\x80\x53\x00\x00\x80\x90\xED"
-     "\x3D\x02"
-     "\x59\x00\x00\x80\x92\x45\x3C\x02\x5F\x00\x00\x80\x94\x71\x3C\x02\x1F\x00"
-     "\x00\x80"
-     "\x61\x06\x00\x80\x4F\x00\x00\x80\xCC\x20\x38\x02\xCD\x68\x30\x02\x9A\xED"
-     "\x04\x00"
-     "\x9B\x5D\x3F\x02\x7E\x20\x00\x80\x6D\x00\x00\x80\x9E\xF5\x04\x00\xCC\x34"
-     "\x38\x02"
-     "\x80\x79\x3D\x02\x42\x03\x00\x80\x82\x09\x3D\x02\x44\x03\x00\x80\x84\x01"
-     "\x3D\x02"
-     "\x47\x03\x00\x80\x86\x15\x3D\x02\x8A\x20\x00\x80\x88\x35\x3D\x02\x91\x20"
-     "\x00\x80"
-     "\x8A\x05\x3D\x02\xCC\x8C\x0F\x00\x8C\xF5\x3D\x02\xA8\x20\x00\x80\x8E\xB9"
-     "\x3D\x02"
-     "\xAF\x20\x00\x80\xB0\x85\x36\x02\xC9\x00\x00\x80\xB2\xF9\x36\x02\x0F\x00"
-     "\x00\x80"
-     "\xB4\xF9\x36\x02\x52\x06\x00\x80\xB6\xC1\x35\x02\x4F\x06\x00\x80\xB8\x75"
-     "\x04\x00"
-     "\xCC\x00\x00\x80\xBA\xFD\x34\x02\x9B\x20\x00\x80\xBC\x91\x34\x02\x4D\x03"
-     "\x00\x80"
-     "\xBE\x4D\x35\x02\x75\x05\x00\x80\xA0\x11\x04\x00\xB0\x05\x00\x80\xA2\x35"
-     "\x36\x02"
-     "\xD6\x05\x00\x80\xA4\xED\x37\x02\xED\x05\x00\x80\xA6\xE1\x37\x02\x0A\x06"
-     "\x00\x80"
-     "\xA8\xAD\x37\x02\x21\x06\x00\x80\xAA\xAD\x37\x02\x3C\x03\x00\x80\xAC\xBD"
-     "\x35\x02"
-     "\x05\x21\x00\x80\xAE\xB9\x36\x02\xDB\x00\x00\x80\x80\xAD\x3D\x02\x80\xE1"
-     "\x3A\x02"
-     "\x82\xBD\x3D\x02\x82\x1D\x3B\x02\x84\xC1\x3D\x02\x84\xED\x3A\x02\x86\xC5"
-     "\x3D\x02"
-     "\x86\x71\x3B\x02\x88\xC9\x3D\x02\x88\x21\x3D\x02\x8A\xCD\x3D\x02\x8A\x0D"
-     "\x3D\x02"
-     "\x8C\x61\x3E\x02\x8C\xED\x0B\x00\x8E\x49\x3E\x02\x8E\xD1\x3E\x02\x90\x3D"
-     "\x3E\x02"
-     "\x90\x71\x3E\x02\x92\xCD\x3E\x02\x92\x6D\x3E\x02\x94\xCD\x3F\x02\x94\x11"
-     "\x3E\x02"
-     "\x96\x99\x3F\x02\x96\x65\x3E\x02\x98\x8D\x3F\x02\x98\xF1\x3F\x02\x9A\xD1"
-     "\x0B\x00"
-     "\x9A\x31\x3F\x02\x9C\x35\x30\x02\x9C\x25\x3F\x02\x9E\x2D\x30\x02\x9E\xF9"
-     "\x3F\x02"
-     "\xA0\xF1\x30\x02\xA0\x05\x3F\x02\xA2\x49\x0B\x00\xA2\x1D\x3F\x02\xA4\xF9"
-     "\x30\x02"
-     "\xA4\x59\x31\x02\xA6\x19\x30\x02\xA6\xE1\x31\x02\xA8\x2D\x30\x02\xA8\x15"
-     "\x31\x02"
-     "\xAA\x01\x30\x02\xAA\x01\x31\x02\xAC\x05\x30\x02\xAC\x09\x31\x02\xAE\x5D"
-     "\x30\x02"
-     "\xAE\x09\x31\x02\xB0\x09\x30\x02\xB0\x0D\x31\x02\xB2\x29\x31\x02\xB2\x5D"
-     "\x32\x02"
-     "\xB4\x3D\x31\x02\xB4\x35\x32\x02\xB6\x31\x31\x02\xB6\x05\x32\x02\xB8\xC1"
-     "\x31\x02"
-     "\xB8\x51\x32\x02\xBA\xCD\x31\x02\xBA\x31\x0B\x00\xBC\xE1\x32\x02\xBC\x15"
-     "\x0B\x00"
-     "\xBE\xCD\x32\x02\xBE\x0D\x0B\x00\x88\xF5\x0A\x00\x89\xFD\x0A\x00\x8A\x8D"
-     "\x0D\x00"
-     "\x8B\x09\x0C\x00\x8C\x19\x0C\x00\x8D\x19\x0C\x00\x8E\x39\x0C\x00\x8F\x39"
-     "\x0C\x00"
-     "\x88\x29\x2B\x02\x89\x75\x2B\x02\x8A\xD9\x28\x02\x8B\x3D\x2B\x02\x8C\xC9"
-     "\x28\x02"
-     "\x8D\x11\x2B\x02\x8E\x3D\x2B\x02\x8F\x71\x2B\x02\x98\x69\x0C\x00\x99\x69"
-     "\x0C\x00"
-     "\x9A\x8D\x2B\x02\x9B\xAD\x2B\x02\x9C\x85\x2B\x02\x9D\xA5\x2B\x02\xCC\x70"
-     "\x21\x02"
-     "\xCD\xCC\x21\x02\x98\x09\x2A\x02\x99\xC1\x2A\x02\x9A\x39\x2A\x02\x9B\x11"
-     "\x2A\x02"
-     "\x9C\x29\x2A\x02\x9D\xF1\x2A\x02\x9E\x19\x2A\x02\x9F\xF1\x2A\x02\xA8\xB1"
-     "\x0C\x00"
-     "\xA9\xB1\x0C\x00\xAA\xB1\x0C\x00\xAB\xB1\x0C\x00\xAC\xD5\x0C\x00\xAD\xDD"
-     "\x0C\x00"
-     "\xAE\xD5\x0C\x00\xAF\xCD\x0C\x00\xA8\x65\x2F\x02\xA9\x81\x2E\x02\xAA\x25"
-     "\x2F\x02"
-     "\xAB\x61\x2F\x02\xAC\x19\x2F\x02\xAD\x61\x2F\x02\xAE\x01\x2F\x02\xAF\x61"
-     "\x2F\x02"
-     "\xB8\x95\x0C\x00\xB9\x9D\x0C\x00\xBA\x35\x2A\x02\xBB\x69\x2D\x02\xBC\xE1"
-     "\x2D\x02"
-     "\xBD\x7D\x2D\x02\xBE\xE1\x2D\x02\xBF\x1D\x2D\x02\xB8\x01\x28\x02\xB9\x35"
-     "\x28\x02"
-     "\xBA\x29\x28\x02\xBB\x91\x28\x02\xBC\x71\x2B\x02\xBD\x15\x0F\x00\xBE\x91"
-     "\x03\x00"
-     "\xBF\x1D\x0F\x00\xB8\x35\x06\x00\xB8\x95\x3F\x02\x38\x03\x00\x80\xBA\x91"
-     "\x3F\x02"
-     "\x3D\x08\x00\x80\xBC\x8D\x3F\x02\x6A\x06\x00\x80\xBE\xB5\x3F\x02\x6E\x06"
-     "\x00\x80"
-     "\xB0\x45\x3F\x02\xA4\x08\x00\x80\xB2\x41\x3F\x02\x8A\x06\x00\x80\xB4\x41"
-     "\x3F\x02"
-     "\x8E\x06\x00\x80\xB6\x89\x09\x00\xA8\x08\x00\x80\xA8\xF5\x3E\x02\xE8\x07"
-     "\x00\x80"
-     "\xAA\xF9\x3E\x02\x45\x08\x00\x80\xAC\x6D\x3F\x02\x56\x08\x00\x80\xAE\x5D"
-     "\x3F\x02"
-     "\x5E\x08\x00\x80\xA0\x51\x3E\x02\x09\x0C\x00\x80\xA2\xCD\x3E\x02\xBD\x20"
-     "\x00\x80"
-     "\xA4\xD9\x3E\x02\xE4\x0E\x00\x80\xA6\xD9\x3E\x02\xFB\x0E\x00\x80\x98\x31"
-     "\x39\x02"
-     "\x08\x0F\x00\x80\x9A\x55\x39\x02\x10\x0F\x00\x80\x9C\x15\x39\x02\xDD\x06"
-     "\x00\x80"
-     "\x9E\x71\x39\x02\x01\x21\x00\x80\x90\x69\x38\x02\x62\x04\x00\x80\x92\x7D"
-     "\x38\x02"
-     "\x6E\x04\x00\x80\x94\x91\x39\x02\x3D\x04\x00\x80\x96\xFD\x39\x02\x4A\x04"
-     "\x00\x80"
-     "\x88\x6D\x38\x02\x07\x05\x00\x80\x8A\x71\x38\x02\x56\x04\x00\x80\x8C\x11"
-     "\x38\x02"
-     "\x0D\x05\x00\x80\x8E\x09\x38\x02\x5A\x05\x00\x80\x80\x81\x38\x02\x9B\x05"
-     "\x00\x80"
-     "\x82\x9D\x38\x02\x5C\x04\x00\x80\x84\x89\x38\x02\xBB\x05\x00\x80\x86\x79"
-     "\x38\x02"
-     "\x88\x5D\x0E\x00\x89\x65\x0E\x00\x8A\x21\x2E\x02\x8B\x35\x2E\x02\x8C\x21"
-     "\x2E\x02"
-     "\x8D\x4D\x2E\x02\x68\x04\x00\x80\x60\x05\x00\x80\xCC\xF8\x05\x00\xC1\x05"
-     "\x00\x80"
-     "\xF5\x05\x00\x80\x38\x04\x00\x80\x20\x07\x00\x80\x23\x07\x00\x80\x30\x07"
-     "\x00\x80"
-     "\x65\x08\x00\x80\x68\x08\x00\x80\x99\x21\x0E\x00\x6C\x08\x00\x80\x9B\x95"
-     "\x2E\x02"
-     "\x33\x07\x00\x80\x9D\xB5\x2E\x02\x84\x08\x00\x80\x9F\xB5\x2E\x02\x87\x08"
-     "\x00\x80"
-     "\xC1\x20\x00\x80\x55\x03\x00\x80\xD5\x20\x00\x80\xD9\x20\x00\x80\xDD\x20"
-     "\x00\x80"
-     "\xE1\x20\x00\x80\xE9\x20\x00\x80\xA8\xE9\x0E\x00\xA9\xE9\x0E\x00\xAA\xF9"
-     "\x0E\x00"
-     "\xAB\xF9\x0E\x00\xAC\xE9\x0E\x00\xAD\xE9\x0E\x00\xAE\xB9\x0D\x00\xAF\xA1"
-     "\x0D\x00"
-     "\x90\x38\x0A\x00\x92\xF8\x0A\x00\x92\x88\x0B\x00\x90\x94\x0A\x00\x94\x94"
-     "\x0B\x00"
-     "\xE5\x20\x00\x80\xF1\x20\x00\x80\x94\x0C\x0A\x00\xF9\x20\x00\x80\xB9\x6D"
-     "\x2A\x02"
-     "\x89\xA1\x00\x00\xBB\x7D\x2A\x02\xFD\x20\x00\x80\xBD\x6D\x2A\x02\x09\x21"
-     "\x00\x80"
-     "\x0D\x21\x00\x80\x83\x85\x00\x00\xB1\x99\x2A\x02\x11\x21\x00\x80\xB3\x65"
-     "\x2A\x02"
-     "\x18\x21\x00\x80\xB5\x65\x2A\x02\x2F\x21\x00\x80\xB7\x5D\x2A\x02\x80\x00"
-     "\x03\x00"
-     "\x81\x18\x01\x00\x82\x00\x7C\x00\x36\x21\x00\x80\x84\x10\x7E\x00\x85\x14"
-     "\x7F\x00"
-     "\x86\xF4\x05\x00\x87\xFC\x05\x00\x88\xBC\x01\x00\x89\x28\x7D\x00\x8A\xEC"
-     "\x7D\x00"
-     "\x42\x06\x00\x80\x8C\x24\x00\x00\x40\x21\x00\x80\xB8\x05\x00\x00\x44\x21"
-     "\x00\x80"
-     "\xCC\xF0\x02\x00\x91\x00\x78\x00\x92\x48\x7A\x00\x93\x4C\x7B\x00\xB8\x05"
-     "\x00\x00"
-     "\x48\x21\x00\x80\x96\xD4\x00\x00\xCC\xDC\x02\x00\xB8\x05\x00\x00\x4C\x21"
-     "\x00\x80"
-     "\xCC\xE4\x02\x00\xB8\x1D\x00\x00\x58\x21\x00\x80\xCC\xEC\x02\x00\xB8\x05"
-     "\x00\x00"
-     "\x54\x21\x00\x80\xCC\xE4\x02\x00\xB8\x0D\x00\x00\x64\x21\x00\x80\xCC\xFC"
-     "\x02\x00"
-     "\xB8\x05\x00\x00\x68\x21\x00\x80\xCC\xE4\x02\x00\xB8\x35\x00\x00\xA8\xA4"
-     "\x01\x00"
-     "\xA9\x80\x7C\x00\x6C\x21\x00\x80\xAB\x8C\x01\x00\xCC\xE4\x02\x00\xB8\x0D"
-     "\x00\x00"
-     "\x70\x21\x00\x80\xCC\x8C\x02\x00\xB0\xC0\x79\x00\xB1\x44\x78\x00\xB2\x00"
-     "\x74\x00"
-     "\xB3\x58\x76\x00\xB8\x09\x00\x00\xB5\x2C\x01\x00\x74\x21\x00\x80\xCC\xD8"
-     "\x02\x00"
-     "\x78\x21\x00\x80\xB8\x05\x00\x00\xBA\xCC\x01\x00\xBB\x50\x01\x00\xBC\xF0"
-     "\x77\x00"
-     "\xBD\x00\x70\x00\xBE\xF8\x73\x00\xBF\x7C\x71\x00\x80\x09\x3D\x02\x81\x0D"
-     "\x3D\x02"
-     "\x82\x01\x3D\x02\x83\x05\x3D\x02\x84\x19\x3D\x02\x85\x1D\x3D\x02\x86\x11"
-     "\x3D\x02"
-     "\x87\x15\x3D\x02\x88\x29\x3D\x02\x89\x2D\x3D\x02\x8A\x21\x3D\x02\x8B\x25"
-     "\x3D\x02"
-     "\x8C\x39\x3D\x02\xB8\x75\x00\x00\x8E\x31\x3D\x02\x8F\x35\x3D\x02\x7C\x21"
-     "\x00\x80"
-     "\x91\x75\x0A\x00\xCC\x9C\x02\x00\xB8\x19\x00\x00\xCC\xFC\x02\x00\x80\x21"
-     "\x00\x80"
-     "\xCC\xDC\x02\x00\x97\x69\x0A\x00\xA4\x21\x00\x80\xB8\x05\x00\x00\xCC\xF8"
-     "\x02\x00"
-     "\xA8\x21\x00\x80\xB8\x1D\x00\x00\xCC\xEC\x02\x00\xB8\x05\x00\x00\x84\x21"
-     "\x00\x80"
-     "\xB8\x09\x00\x00\xCC\xE4\x02\x00\x88\x21\x00\x80\xCC\xF0\x02\x00\xA4\xD5"
-     "\x0A\x00"
-     "\xA5\xD9\x0A\x00\xA6\xCD\x0A\x00\xB8\x35\x00\x00\xA8\xA9\x3D\x02\xA9\xAD"
-     "\x3D\x02"
-     "\x8C\x21\x00\x80\xCC\xFC\x02\x00\xB8\x05\x00\x00\x90\x21\x00\x80\xCC\x98"
-     "\x02\x00"
-     "\xAF\xB5\x3D\x02\xB8\x05\x00\x00\xAC\x21\x00\x80\xCC\xFC\x02\x00\xB3\x91"
-     "\x0A\x00"
-     "\xB4\xB5\x0A\x00\xB8\x35\x00\x00\xB6\xAD\x0A\x00\xB7\xA1\x0A\x00\xB0\x21"
-     "\x00\x80"
-     "\xCC\xEC\x02\x00\xB8\x05\x00\x00\x94\x21\x00\x80\xBC\x75\x0A\x00\xCC\xE8"
-     "\x02\x00"
-     "\xBE\x75\x0A\x00\xB8\xDD\x03\x00\xBE\x1D\x09\x00\xCC\xC0\x02\x00\xBC\x35"
-     "\x29\x02"
-     "\xBD\xD9\x35\x02\xBA\x41\x28\x02\xBB\xAD\x2F\x02\xB8\x39\x2B\x02\xB9\xB5"
-     "\x2F\x02"
-     "\x98\x21\x00\x80\xB8\x0D\x00\x00\x9C\x21\x00\x80\xB8\x19\x00\x00\xCC\xFC"
-     "\x02\x00"
-     "\xA0\x21\x00\x80\xCC\xE4\x02\x00\xB8\x51\x00\x00\xAE\x35\x2F\x02\xAF\x5D"
-     "\x09\x00"
-     "\xAC\x3D\x2B\x02\xAD\x59\x09\x00\xAA\xE1\x2B\x02\xAB\xE5\x2F\x02\xA8\x99"
-     "\x2B\x02"
-     "\xA9\xE5\x2B\x02\xCC\xE4\x02\x00\xB8\xC5\x00\x00\xCC\xF8\x02\x00\xB4\x21"
-     "\x00\x80"
-     "\xB8\xD5\x00\x00\xA3\x6D\x24\x02\xCC\xE4\x02\x00\xB8\xD5\x00\x00\x9E\x85"
-     "\x09\x00"
-     "\x9F\x89\x09\x00\x81\x81\x3E\x02\x9D\x71\x08\x00\x9A\x45\x2D\x02\x9B\x35"
-     "\x2F\x02"
-     "\x98\x75\x2D\x02\x99\x79\x2D\x02\xB8\x21\x00\x80\xBC\x21\x00\x80\xC0\x21"
-     "\x00\x80"
-     "\x42\x24\x00\x80\xA9\x1D\x00\x00\x93\x61\x27\x02\x46\x24\x00\x80\xAA\x05"
-     "\x00\x00"
-     "\x8E\x35\x08\x00\x8F\x3D\x08\x00\x8C\x7D\x2D\x02\x8D\x3D\x08\x00\x8A\x99"
-     "\x2D\x02"
-     "\x8B\x45\x2F\x02\x88\x69\x2C\x02\x89\x55\x2F\x02\x68\x03\x00\x80\x81\x03"
-     "\x00\x80"
-     "\x9A\x03\x00\x80\xB3\x03\x00\x80\xC6\x03\x00\x80\xD9\x03\x00\x80\x80\x1D"
-     "\x08\x00"
-     "\x81\x15\x08\x00\x87\x95\x08\x00\xEC\x03\x00\x80\xA5\xF0\x0F\x00\xFF\x03"
-     "\x00\x80"
-     "\xA3\x6C\x0F\x00\x12\x04\x00\x80\x25\x04\x00\x80\xCC\xE0\x3D\x02\xAF\xD9"
-     "\x04\x00"
-     "\x8B\x04\x00\x80\xAD\xD9\x04\x00\xAC\xC5\x04\x00\xDD\x00\x00\x80\x8C\xD1"
-     "\x7D\x00"
-     "\x89\x9D\x08\x00\x88\xA1\x08\x00\x97\xED\x08\x00\xAF\x39\x77\x00\x13\x01"
-     "\x00\x80"
-     "\x1C\x01\x00\x80\x9C\x55\x7C\x00\x20\x01\x00\x80\x9F\xD9\x77\x00\xB0\x51"
-     "\x03\x00"
-     "\x9F\x69\x3F\x02\x24\x01\x00\x80\x28\x01\x00\x80\xBC\x30\x0B\x00\xB3\xF9"
-     "\x77\x00"
-     "\x2C\x01\x00\x80\x30\x01\x00\x80\x82\x98\x02\x00\x99\x99\x6F\x00\x9E\x9D"
-     "\x6F\x00"
-     "\x34\x01\x00\x80\x82\x44\x02\x00\x83\xB8\x0F\x00\xE1\x00\x00\x80\xEF\xB4"
-     "\x02\x00"
-     "\xE6\x00\x00\x80\xEB\x00\x00\x80\xF0\x00\x00\x80\xE3\xDC\x02\x00\x99\x8D"
-     "\x6F\x00"
-     "\x8B\xB4\x0F\x00\xF5\x00\x00\x80\xFA\x00\x00\x80\x88\xA8\x0F\x00\xB7\x15"
-     "\x08\x00"
-     "\xB6\x09\x08\x00\xB5\x0D\x08\x00\xB4\x71\x08\x00\xFF\x00\x00\x80\xBE\x28"
-     "\x02\x00"
-     "\xB1\xC9\x0B\x00\xB0\x69\x08\x00\xBF\x21\x3A\x02\xBE\xED\x05\x00\xBD\x15"
-     "\x08\x00"
-     "\xBC\xC1\x03\x00\xBB\x05\x08\x00\xBA\x19\x08\x00\xB9\x1D\x08\x00\xB8\x21"
-     "\x08\x00"
-     "\x80\x19\x7A\x00\x81\x25\x7A\x00\x82\x01\x44\x02\x83\x25\x7A\x00\x04\x01"
-     "\x00\x80"
-     "\x85\x25\x7A\x00\x86\x21\x7A\x00\x87\x65\x4E\x02\x09\x01\x00\x80\x89\x65"
-     "\x7A\x00"
-     "\x8A\xA9\x44\x02\x8B\xB9\x44\x02\x8C\xA5\x44\x02\x8D\xA1\x44\x02\x8E\xA5"
-     "\x50\x02"
-     "\x8F\xF5\x4D\x02\x90\xFD\x44\x02\x91\xF9\x44\x02\x92\xC9\x45\x02\x93\xCD"
-     "\x45\x02"
-     "\x0E\x01\x00\x80\x95\x95\x45\x02\x96\x11\x7A\x00\x17\x01\x00\x80\x65\x03"
-     "\x00\x80"
-     "\x99\x5D\x48\x02\x9A\xAD\x47\x02\x9B\x2D\x48\x02\x9C\x31\x48\x02\x9D\x35"
-     "\x48\x02"
-     "\x88\x04\x00\x80\x27\x05\x00\x80\xA0\xD1\x7A\x00\xA1\xDD\x7A\x00\xA2\xE9"
-     "\x7A\x00"
-     "\x7A\x05\x00\x80\xA4\x89\x4C\x02\xB5\x05\x00\x80\xA6\xF1\x58\x02\xDB\x05"
-     "\x00\x80"
-     "\xA8\xB9\x4C\x02\xF2\x05\x00\x80\xAA\xF5\x45\x02\xAB\x29\x4C\x02\xAC\xB5"
-     "\x44\x02"
-     "\xAD\xBD\x44\x02\x0F\x06\x00\x80\xAF\x99\x44\x02\xB0\xE5\x44\x02\xB1\x95"
-     "\x44\x02"
-     "\xB2\xB9\x7A\x00\xB3\x69\x45\x02\xB4\x3D\x45\x02\xB5\xAD\x7A\x00\xB6\x59"
-     "\x7A\x00"
-     "\xB7\x55\x7A\x00\xB8\x71\x7A\x00\xB9\xF5\x74\x00\x26\x06\x00\x80\xBB\x75"
-     "\x7A\x00"
-     "\xBC\x91\x55\x02\xBD\x65\x52\x02\xBE\x69\x52\x02\xBF\x9D\x55\x02\x80\xA1"
-     "\x7B\x00"
-     "\x6B\x03\x00\x80\x84\x03\x00\x80\x99\x29\x11\x00\x9D\x03\x00\x80\x85\x09"
-     "\x45\x02"
-     "\x86\x05\x45\x02\x87\x39\x45\x02\x88\x31\x75\x00\x89\x89\x51\x02\xB6\x03"
-     "\x00\x80"
-     "\xC9\x03\x00\x80\xDC\x03\x00\x80\xEF\x03\x00\x80\x02\x04\x00\x80\x15\x04"
-     "\x00\x80"
-     "\x90\xE9\x7B\x00\x91\xF5\x7B\x00\x92\xF1\x7B\x00\x93\x8D\x7B\x00\x94\x99"
-     "\x7B\x00"
-     "\x95\x85\x7B\x00\x96\x81\x7B\x00\x97\xBD\x7B\x00\x98\x89\x7B\x00\x99\x95"
-     "\x7B\x00"
-     "\x9A\x91\x7B\x00\x9B\x75\x7A\x00\x9C\x6D\x7A\x00\x9D\x51\x7A\x00\x9E\x51"
-     "\x7A\x00"
-     "\x9F\x51\x7A\x00\xA0\x3D\x45\x02\xA1\xB5\x7A\x00\xA2\xBD\x7A\x00\xA3\xB5"
-     "\x7A\x00"
-     "\xA4\x69\x48\x02\xA5\xA9\x7A\x00\xA6\xD9\x7A\x00\xA7\xD5\x7A\x00\xA8\xF1"
-     "\x7A\x00"
-     "\xA9\xA5\x4D\x02\xAA\xF1\x7A\x00\xAB\xCD\x7A\x00\xAC\x31\x44\x02\xAD\xBD"
-     "\x45\x02"
-     "\xAE\xA5\x45\x02\xAF\x19\x44\x02\xB0\xD1\x75\x00\xB1\xF5\x7A\x00\xB2\xFD"
-     "\x7A\x00"
-     "\xB3\xF5\x7A\x00\xB4\x29\x48\x02\xB5\xE9\x7A\x00\xB6\x99\x7A\x00\xB7\x95"
-     "\x7A\x00"
-     "\xB8\xB1\x7A\x00\xB9\xE5\x4D\x02\xBA\xB1\x7A\x00\xBB\x8D\x7A\x00\xBC\x71"
-     "\x44\x02"
-     "\xBD\xFD\x45\x02\xBE\xE5\x45\x02\xBF\x59\x44\x02\x82\xBD\x49\x02\x83\xCD"
-     "\x49\x02"
-     "\x80\xBD\x75\x00\x81\x45\x4E\x02\x86\xDD\x75\x00\x87\xC1\x75\x00\x84\xA5"
-     "\x75\x00"
-     "\x85\xD9\x75\x00\x8A\xCD\x75\x00\x8B\xD1\x75\x00\x88\xF5\x75\x00\x89\xC9"
-     "\x75\x00"
-     "\x8E\x39\x78\x00\x28\x04\x00\x80\x8C\x15\x7E\x00\x8D\xC1\x75\x00\x92\xAD"
-     "\x47\x02"
-     "\x93\x45\x4E\x02\x90\x09\x42\x02\x91\x69\x46\x02\x96\x0D\x47\x02\x97\xD5"
-     "\x47\x02"
-     "\x94\x2D\x4C\x02\x95\xC1\x52\x02\x9A\x59\x4A\x02\x9B\x39\x4A\x02\x98\xCD"
-     "\x47\x02"
-     "\x99\xAD\x47\x02\x8E\x04\x00\x80\x38\x01\x00\x80\x9C\x75\x76\x00\x3C\x01"
-     "\x00\x80"
-     "\x40\x01\x00\x80\x44\x01\x00\x80\x48\x01\x00\x80\x4C\x01\x00\x80\x50\x01"
-     "\x00\x80"
-     "\x54\x01\x00\x80\x58\x01\x00\x80\x5C\x01\x00\x80\x60\x01\x00\x80\x64\x01"
-     "\x00\x80"
-     "\xA8\xB9\x78\x00\x68\x01\x00\x80\x6C\x01\x00\x80\x70\x01\x00\x80\x74\x01"
-     "\x00\x80"
-     "\x78\x01\x00\x80\x7C\x01\x00\x80\x80\x01\x00\x80\x84\x01\x00\x80\x88\x01"
-     "\x00\x80"
-     "\x8C\x01\x00\x80\x90\x01\x00\x80\x94\x01\x00\x80\x98\x01\x00\x80\x9C\x01"
-     "\x00\x80"
-     "\x22\x21\x00\x80\x3A\x06\x00\x80\x44\x06\x00\x80\xB5\x1D\x00\x00\xB4\x1D"
-     "\x00\x00"
-     "\x47\x06\x00\x80\xB6\x05\x00\x00\x83\x10\x7B\x00\xD8\x21\x00\x80\xB3\x58"
-     "\x7E\x00"
-     "\xB2\x04\x7E\x00\xB5\x58\x7E\x00\xB4\x4C\x7E\x00\x85\x68\x7C\x00\xDE\x21"
-     "\x00\x80"
-     "\xE2\x21\x00\x80\xE6\x21\x00\x80\xEA\x21\x00\x80\x88\x40\x7C\x00\xEE\x21"
-     "\x00\x80"
-     "\xF2\x21\x00\x80\x8D\x60\x7C\x00\xF6\x21\x00\x80\xFA\x21\x00\x80\xFE\x21"
-     "\x00\x80"
-     "\x02\x22\x00\x80\xA2\x98\x7D\x00\x06\x22\x00\x80\x0A\x22\x00\x80\x0E\x22"
-     "\x00\x80"
-     "\x12\x22\x00\x80\xA9\x54\x7E\x00\xA8\xA0\x7D\x00\xAB\x68\x7E\x00\x16\x22"
-     "\x00\x80"
-     "\x1A\x22\x00\x80\x1E\x22\x00\x80\x22\x22\x00\x80\x26\x22\x00\x80\x91\x08"
-     "\x7D\x00"
-     "\x2A\x22\x00\x80\xA1\xCC\x7C\x00\x92\x74\x7D\x00\x2E\x22\x00\x80\x32\x22"
-     "\x00\x80"
-     "\xA5\x24\x7C\x00\xA4\x2C\x7C\x00\x36\x22\x00\x80\x3A\x22\x00\x80\x3E\x22"
-     "\x00\x80"
-     "\x42\x22\x00\x80\x46\x22\x00\x80\x4A\x22\x00\x80\x4E\x22\x00\x80\x52\x22"
-     "\x00\x80"
-     "\xB3\x40\x7C\x00\xB2\x6C\x7C\x00\x83\x7C\x7D\x00\x82\x88\x7D\x00\xB7\x3C"
-     "\x7C\x00"
-     "\xB6\x58\x7C\x00\x87\x44\x7D\x00\x86\x60\x7D\x00\xBB\x80\x7D\x00\xBA\x9C"
-     "\x7D\x00"
-     "\x56\x22\x00\x80\x5A\x22\x00\x80\x5E\x22\x00\x80\x62\x22\x00\x80\xBD\xA0"
-     "\x7D\x00"
-     "\xBC\x94\x7D\x00\x80\xA5\x7B\x00\x81\xB1\x7B\x00\x82\xB1\x7B\x00\x83\xDD"
-     "\x7B\x00"
-     "\x84\x01\x04\x00\x85\x35\x04\x00\x86\x49\x04\x00\x87\x41\x04\x00\x88\x41"
-     "\x04\x00"
-     "\x89\x5D\x04\x00\x8A\x55\x04\x00\x8B\xA1\x04\x00\x8C\xA5\x04\x00\x8D\xA9"
-     "\x04\x00"
-     "\x8E\x9D\x04\x00\x8F\x85\x04\x00\x90\xA9\x04\x00\x91\x41\x05\x00\x92\x41"
-     "\x05\x00"
-     "\x93\x5D\x05\x00\x94\x79\x05\x00\x95\x79\x05\x00\x96\x69\x05\x00\x97\x69"
-     "\x05\x00"
-     "\x98\x59\x05\x00\x99\x59\x05\x00\x9A\x5D\x06\x00\x9B\x1D\x06\x00\x9C\x05"
-     "\x06\x00"
-     "\x9D\x09\x06\x00\x9E\xF9\x06\x00\x9F\xF9\x06\x00\xA0\x09\x06\x00\xA1\x09"
-     "\x06\x00"
-     "\xA2\x19\x06\x00\xA3\x19\x06\x00\xA4\x09\x06\x00\xA5\x09\x06\x00\xA6\x39"
-     "\x06\x00"
-     "\xA7\x39\x06\x00\xA8\x0D\x06\x00\xA9\x15\x06\x00\xAA\x1D\x06\x00\xAB\x15"
-     "\x06\x00"
-     "\xAC\x0D\x06\x00\xAD\x75\x06\x00\xAE\x7D\x06\x00\xAF\x75\x06\x00\xB0\x0D"
-     "\x06\x00"
-     "\xB1\x15\x06\x00\xB2\x1D\x06\x00\xB3\x15\x06\x00\xB4\x0D\x06\x00\xB5\x35"
-     "\x06\x00"
-     "\xB6\x99\x44\x02\xB7\xD9\x40\x02\xB8\xE9\x40\x02\xB9\xF9\x40\x02\xBA\xCD"
-     "\x40\x02"
-     "\xBB\xBD\x40\x02\xBC\x71\x40\x02\xBD\x61\x40\x02\xBE\x45\x40\x02\xBF\xB9"
-     "\x41\x02"
-     "\x80\x5D\x40\x02\x81\x85\x40\x02\x82\xAD\x40\x02\x83\xCD\x40\x02\x84\xFD"
-     "\x40\x02"
-     "\x85\x2D\x4D\x02\x86\xDD\x42\x02\x87\x5D\x4D\x02\x88\x7D\x4D\x02\x89\x85"
-     "\x4D\x02"
-     "\x8A\xC1\x4D\x02\x8B\xD5\x4C\x02\x8C\xE9\x48\x02\x8D\x35\x49\x02\x8E\x31"
-     "\x49\x02"
-     "\x8F\x25\x49\x02\x90\x01\x45\x02\x91\x41\x41\x02\x92\x41\x41\x02\x93\x51"
-     "\x41\x02"
-     "\x94\x75\x41\x02\x95\x05\x41\x02\x96\xD9\x41\x02\x97\xC1\x55\x02\x98\x71"
-     "\x71\x00"
-     "\x99\xC9\x55\x02\x9A\x35\x40\x02\x9B\xED\x40\x02\x9C\xD5\x40\x02\x9D\xB5"
-     "\x40\x02"
-     "\x9E\x95\x40\x02\x9F\x45\x4D\x02\xA0\x45\x42\x02\xA1\xC5\x4D\x02\xA2\xD5"
-     "\x4D\x02"
-     "\xA3\x81\x71\x00\xA4\x79\x4D\x02\xA5\x6D\x4C\x02\xA6\x01\x54\x02\xA7\x9D"
-     "\x49\x02"
-     "\xA8\x69\x54\x02\xA9\xBD\x49\x02\xAA\x1D\x72\x00\x66\x22\x00\x80\x6A\x22"
-     "\x00\x80"
-     "\x6E\x22\x00\x80\x72\x22\x00\x80\x76\x22\x00\x80\x7A\x22\x00\x80\x7E\x22"
-     "\x00\x80"
-     "\x82\x22\x00\x80\x86\x22\x00\x80\x8A\x22\x00\x80\x8E\x22\x00\x80\x92\x22"
-     "\x00\x80"
-     "\x96\x22\x00\x80\x9A\x22\x00\x80\x9E\x22\x00\x80\x0D\x0B\x00\x80\xB2\x1C"
-     "\x00\x80"
-     "\x40\x0B\x00\x80\xAA\x22\x00\x80\xAE\x22\x00\x80\xB2\x22\x00\x80\xA0\xE5"
-     "\x03\x00"
-     "\xBA\x22\x00\x80\xA2\xE1\x03\x00\xA3\xE1\x03\x00\xA4\xE1\x03\x00\xBE\x22"
-     "\x00\x80"
-     "\xAA\x0A\x00\x80\xA7\xE9\x03\x00\xAD\x0A\x00\x80\xA9\xDD\x03\x00\xC2\x22"
-     "\x00\x80"
-     "\xAB\xD1\x03\x00\xC6\x22\x00\x80\xAD\xC5\x74\x00\xAE\x41\x75\x00\xAF\xD1"
-     "\x74\x00"
-     "\xB0\x6D\x75\x00\xCA\x22\x00\x80\xB2\x4D\x00\x00\xCE\x22\x00\x80\xD2\x22"
-     "\x00\x80"
-     "\xB5\x41\x00\x00\xD6\x22\x00\x80\xDA\x22\x00\x80\xDE\x22\x00\x80\xE2\x22"
-     "\x00\x80"
-     "\xE6\x22\x00\x80\xEA\x22\x00\x80\xBC\xDD\x56\x02\xBD\x8D\x4F\x02\xBE\x61"
-     "\x00\x00"
-     "\xBF\x5D\x00\x00\xB1\x95\x79\x00\xB0\x89\x79\x00\xB3\xAD\x79\x00\xB2\x95"
-     "\x79\x00"
-     "\xB5\xD9\x79\x00\xB4\xA5\x79\x00\xB7\xCD\x79\x00\xB6\xD1\x79\x00\xB9\xC9"
-     "\x79\x00"
-     "\xB8\xC1\x79\x00\xBB\xD1\x79\x00\xBA\xD9\x79\x00\xBD\x25\x79\x00\xBC\x3D"
-     "\x79\x00"
-     "\xBF\x19\x79\x00\xBE\x21\x79\x00\xA1\x7D\x4D\x02\xA0\x8D\x4A\x02\xA3\x75"
-     "\x71\x00"
-     "\xA2\x05\x4D\x02\xA5\x1D\x71\x00\xA4\x11\x71\x00\xA7\x05\x71\x00\xA6\x09"
-     "\x71\x00"
-     "\xA9\x01\x78\x00\xA8\x01\x71\x00\xAB\x01\x78\x00\xAA\x09\x78\x00\xAD\x01"
-     "\x78\x00"
-     "\xAC\x19\x78\x00\xAF\xFD\x79\x00\xAE\x09\x78\x00\x80\x1D\x04\x00\x81\x25"
-     "\x04\x00"
-     "\x82\x2D\x04\x00\x83\x25\x04\x00\x84\x21\x04\x00\x85\x21\x04\x00\x86\x25"
-     "\x04\x00"
-     "\x87\x5D\x04\x00\x88\x65\x04\x00\x89\x6D\x04\x00\x8A\x79\x04\x00\x8B\x79"
-     "\x04\x00"
-     "\x8C\x69\x04\x00\x8D\x69\x04\x00\x8E\x55\x04\x00\x8F\x4D\x04\x00\x90\x35"
-     "\x04\x00"
-     "\x91\x3D\x04\x00\x92\x35\x04\x00\x93\xC9\x04\x00\x94\xC1\x04\x00\x95\xC1"
-     "\x04\x00"
-     "\x96\xF9\x04\x00\x97\xF9\x04\x00\x98\xC9\x04\x00\x99\xC9\x04\x00\x9A\xD9"
-     "\x04\x00"
-     "\x9B\xD9\x04\x00\x9C\xC9\x04\x00\x9D\xC9\x04\x00\x9E\x91\x04\x00\x9F\x91"
-     "\x04\x00"
-     "\xA0\x71\x04\x00\xA1\x71\x04\x00\xA2\x25\x02\x00\xA3\x3D\x02\x00\xA4\x25"
-     "\x02\x00"
-     "\xA5\x2D\x02\x00\xA6\x25\x02\x00\xA7\x5D\x02\x00\xA8\x65\x02\x00\xA9\x6D"
-     "\x02\x00"
-     "\xAA\x65\x02\x00\xAB\x7D\x02\x00\xAC\x65\x02\x00\xAD\x6D\x02\x00\xAE\x65"
-     "\x02\x00"
-     "\xAF\x5D\x02\x00\xEE\x22\x00\x80\xF2\x22\x00\x80\xF6\x22\x00\x80\xFA\x22"
-     "\x00\x80"
-     "\xFE\x22\x00\x80\x02\x23\x00\x80\x06\x23\x00\x80\x0A\x23\x00\x80\x0E\x23"
-     "\x00\x80"
-     "\x12\x23\x00\x80\x16\x23\x00\x80\x1A\x23\x00\x80\x1E\x23\x00\x80\x22\x23"
-     "\x00\x80"
-     "\x26\x23\x00\x80\x2A\x23\x00\x80\x2E\x23\x00\x80\x32\x23\x00\x80\x36\x23"
-     "\x00\x80"
-     "\x3A\x23\x00\x80\x3E\x23\x00\x80\xA0\x7D\x00\x00\x42\x23\x00\x80\xA2\x71"
-     "\x00\x00"
-     "\xAD\x55\x00\x00\x46\x23\x00\x80\x4A\x23\x00\x80\x4E\x23\x00\x80\x52\x23"
-     "\x00\x80"
-     "\x56\x23\x00\x80\xAB\x49\x00\x00\x5A\x23\x00\x80\x5E\x23\x00\x80\x62\x23"
-     "\x00\x80"
-     "\x66\x23\x00\x80\x6A\x23\x00\x80\x6E\x23\x00\x80\x72\x23\x00\x80\x76\x23"
-     "\x00\x80"
-     "\xB2\x25\x00\x00\x7A\x23\x00\x80\x7E\x23\x00\x80\x82\x23\x00\x80\x86\x23"
-     "\x00\x80"
-     "\x8A\x23\x00\x80\x8E\x23\x00\x80\x92\x23\x00\x80\x36\x24\x00\x80\xFA\x32"
-     "\x00\x80"
-     "\x84\xAD\x00\x00\x66\x3B\x00\x80\x86\xA1\x00\x00\xC6\x2B\x00\x80\x80\x8D"
-     "\x00\x00"
-     "\xEA\x2B\x00\x80\x82\xB9\x00\x00\xF6\x2B\x00\x80\x8C\x99\x00\x00\x06\x2C"
-     "\x00\x80"
-     "\x8E\xED\x00\x00\x0E\x2C\x00\x80\x88\x85\x00\x00\x1A\x2C\x00\x80\x8A\x89"
-     "\x00\x00"
-     "\x22\x2C\x00\x80\x94\x89\x00\x00\xBE\x24\x00\x00\x96\x9D\x00\x00\x2A\x2C"
-     "\x00\x80"
-     "\x90\x91\x00\x00\x36\x2C\x00\x80\x92\x95\x00\x00\x91\x2C\x00\x80\x9C\x85"
-     "\x00\x00"
-     "\xA9\x2C\x00\x80\x9E\x89\x00\x00\xB9\x2C\x00\x80\x98\xA1\x00\x00\xC9\x2C"
-     "\x00\x80"
-     "\x9A\xA5\x00\x00\x80\x91\x01\x00\x81\x9D\x01\x00\x82\xA9\x01\x00\x83\xA5"
-     "\x01\x00"
-     "\x84\xA1\x01\x00\x85\xAD\x01\x00\x86\xD9\x01\x00\x87\xCD\x01\x00\x88\xF9"
-     "\x01\x00"
-     "\x89\xC5\x01\x00\x8A\xC1\x01\x00\x8B\xDD\x01\x00\x8C\xC9\x01\x00\x8D\x31"
-     "\x03\x00"
-     "\x8E\x35\x03\x00\x8F\x29\x03\x00\x90\x5D\x03\x00\x91\x61\x03\x00\x92\x65"
-     "\x03\x00"
-     "\x93\x79\x03\x00\x94\x6D\x03\x00\x95\x11\x03\x00\x96\x15\x03\x00\x97\x09"
-     "\x03\x00"
-     "\x98\x3D\x03\x00\x99\x01\x03\x00\x9A\x05\x03\x00\x9B\x19\x03\x00\x9C\x0D"
-     "\x03\x00"
-     "\x9D\xF1\x03\x00\x9E\xF5\x03\x00\x9F\xE1\x03\x00\xA0\x25\x03\x00\xA1\x29"
-     "\x03\x00"
-     "\xA2\x3D\x03\x00\xA3\x31\x03\x00\xA4\x55\x03\x00\xA5\x59\x03\x00\xA6\x4D"
-     "\x03\x00"
-     "\xA7\x41\x03\x00\xA8\x45\x03\x00\xA9\x49\x03\x00\xAA\x5D\x03\x00\xAB\x51"
-     "\x03\x00"
-     "\xAC\xB5\x02\x00\xAD\xBD\x02\x00\xAE\xB5\x02\x00\xAF\xAD\x02\x00\xB0\xD5"
-     "\x02\x00"
-     "\xB1\xDD\x02\x00\xB2\xD5\x02\x00\xB3\xED\x02\x00\xB4\xF5\x02\x00\xB5\xFD"
-     "\x02\x00"
-     "\xB6\xF5\x02\x00\xB7\xED\x02\x00\xB8\xD5\x02\x00\xB9\xDD\x02\x00\xBA\xD5"
-     "\x02\x00"
-     "\xBB\xAD\x02\x00\xBC\xB5\x02\x00\xBD\xBD\x02\x00\xBE\xB5\x02\x00\xBF\xAD"
-     "\x02\x00"
-     "\xB2\xF9\x0D\x00\xD9\x2C\x00\x80\xB0\xFD\x0D\x00\xE9\x2C\x00\x80\xB6\x11"
-     "\x02\x00"
-     "\x09\x2D\x00\x80\xB4\xE5\x0D\x00\x0D\x2D\x00\x80\xBA\x29\x02\x00\x11\x2D"
-     "\x00\x80"
-     "\xB8\x2D\x02\x00\x45\x2D\x00\x80\xBE\x21\x02\x00\x6D\x2D\x00\x80\xBC\x35"
-     "\x02\x00"
-     "\x7D\x2D\x00\x80\xA2\x99\x0D\x00\x85\x2D\x00\x80\xA0\x9D\x0D\x00\x89\x2D"
-     "\x00\x80"
-     "\xA6\xB1\x0D\x00\x95\x2D\x00\x80\xA4\x85\x0D\x00\xB1\x2D\x00\x80\xAA\x89"
-     "\x0D\x00"
-     "\xB5\x2D\x00\x80\xA8\x8D\x0D\x00\xCD\x2D\x00\x80\xAE\x81\x0D\x00\xD1\x2D"
-     "\x00\x80"
-     "\xAC\x95\x0D\x00\xD9\x2D\x00\x80\x92\xE9\x0C\x00\xE9\x2D\x00\x80\x90\xED"
-     "\x0C\x00"
-     "\x85\x2E\x00\x80\x96\x51\x0D\x00\x9E\xB5\x1A\x00\x94\xF5\x0C\x00\x99\x2E"
-     "\x00\x80"
-     "\x9A\x69\x0D\x00\xD9\x2E\x00\x80\x98\x6D\x0D\x00\xED\x2E\x00\x80\x9E\x61"
-     "\x0D\x00"
-     "\xF5\x2E\x00\x80\x9C\x75\x0D\x00\xF9\x2E\x00\x80\x82\xF1\x0C\x00\x09\x2F"
-     "\x00\x80"
-     "\x80\xE1\x0C\x00\x2C\x2F\x00\x80\x86\x9D\x0C\x00\x58\x2F\x00\x80\x84\x9D"
-     "\x0C\x00"
-     "\x64\x2F\x00\x80\x8A\xB9\x0C\x00\x84\x2F\x00\x80\x88\xB9\x0C\x00\x90\x2F"
-     "\x00\x80"
-     "\x8E\x91\x0C\x00\x94\x2F\x00\x80\x8C\xA5\x0C\x00\x98\x2F\x00\x80\x82\x68"
-     "\x08\x00"
-     "\xB0\x2F\x00\x80\xB4\x2F\x00\x80\xE0\x2F\x00\x80\xEC\x2F\x00\x80\xF4\x2F"
-     "\x00\x80"
-     "\xFC\x2F\x00\x80\x0C\x30\x00\x80\x1B\x30\x00\x80\x23\x30\x00\x80\x4F\x30"
-     "\x00\x80"
-     "\x53\x30\x00\x80\x5B\x30\x00\x80\x5F\x30\x00\x80\x67\x30\x00\x80\x6F\x30"
-     "\x00\x80"
-     "\x7F\x30\x00\x80\x9B\x30\x00\x80\x13\x31\x00\x80\x27\x31\x00\x80\x2B\x31"
-     "\x00\x80"
-     "\xAB\x31\x00\x80\xAF\x31\x00\x80\xC3\x31\x00\x80\xC7\x31\x00\x80\xCF\x31"
-     "\x00\x80"
-     "\xD7\x31\x00\x80\xDF\x31\x00\x80\xEB\x31\x00\x80\x21\x32\x00\x80\x31\x32"
-     "\x00\x80"
-     "\x45\x32\x00\x80\xC2\x32\x00\x80\xCE\x32\x00\x80\xDE\x32\x00\x80\xEA\x32"
-     "\x00\x80"
-     "\xF6\x32\x00\x80\xFE\x32\x00\x80\x02\x33\x00\x80\x06\x33\x00\x80\x0A\x33"
-     "\x00\x80"
-     "\x0E\x33\x00\x80\xC6\x33\x00\x80\x0E\x34\x00\x80\x1A\x34\x00\x80\x1E\x34"
-     "\x00\x80"
-     "\x22\x34\x00\x80\x26\x34\x00\x80\x2E\x34\x00\x80\x32\x34\x00\x80\x46\x34"
-     "\x00\x80"
-     "\x62\x34\x00\x80\x6A\x34\x00\x80\xAA\x34\x00\x80\xAE\x34\x00\x80\xB6\x34"
-     "\x00\x80"
-     "\xBA\x34\x00\x80\xC2\x34\x00\x80\xC6\x34\x00\x80\xEA\x34\x00\x80\xEE\x34"
-     "\x00\x80"
-     "\x0A\x35\x00\x80\x0E\x35\x00\x80\x12\x35\x00\x80\x16\x35\x00\x80\x2A\x35"
-     "\x00\x80"
-     "\x4E\x35\x00\x80\x52\x35\x00\x80\x56\x35\x00\x80\x7A\x35\x00\x80\xBA\x35"
-     "\x00\x80"
-     "\xBE\x35\x00\x80\xDE\x35\x00\x80\xEE\x35\x00\x80\xF6\x35\x00\x80\x16\x36"
-     "\x00\x80"
-     "\x3A\x36\x00\x80\x7E\x36\x00\x80\x86\x36\x00\x80\x9A\x36\x00\x80\xA6\x36"
-     "\x00\x80"
-     "\xAE\x36\x00\x80\xB6\x36\x00\x80\xBA\x36\x00\x80\xBE\x36\x00\x80\xD6\x36"
-     "\x00\x80"
-     "\xDA\x36\x00\x80\xF6\x36\x00\x80\xFE\x36\x00\x80\x06\x37\x00\x80\x0A\x37"
-     "\x00\x80"
-     "\x16\x37\x00\x80\x1E\x37\x00\x80\x22\x37\x00\x80\x26\x37\x00\x80\x2E\x37"
-     "\x00\x80"
-     "\x32\x37\x00\x80\xD2\x37\x00\x80\xE6\x37\x00\x80\x1A\x38\x00\x80\x1E\x38"
-     "\x00\x80"
-     "\x26\x38\x00\x80\x52\x38\x00\x80\x5A\x38\x00\x80\x62\x38\x00\x80\x6A\x38"
-     "\x00\x80"
-     "\xA6\x38\x00\x80\xAA\x38\x00\x80\xB2\x38\x00\x80\xB6\x38\x00\x80\xBA\x38"
-     "\x00\x80"
-     "\xE6\x38\x00\x80\xEA\x38\x00\x80\xF2\x38\x00\x80\x06\x39\x00\x80\x0A\x39"
-     "\x00\x80"
-     "\x26\x39\x00\x80\x2E\x39\x00\x80\x32\x39\x00\x80\x56\x39\x00\x80\x76\x39"
-     "\x00\x80"
-     "\x8A\x39\x00\x80\x8E\x39\x00\x80\x96\x39\x00\x80\xC2\x39\x00\x80\xBE\x8C"
-     "\x01\x00"
-     "\xC6\x39\x00\x80\xD6\x39\x00\x80\x80\x55\x01\x00\x81\x5D\x01\x00\x82\x55"
-     "\x01\x00"
-     "\x83\x6D\x01\x00\x84\x75\x01\x00\x85\x7D\x01\x00\x86\x75\x01\x00\x87\x6D"
-     "\x01\x00"
-     "\x88\x55\x01\x00\x89\x5D\x01\x00\x8A\x55\x01\x00\x8B\xAD\x01\x00\x8C\xB5"
-     "\x01\x00"
-     "\x8D\xBD\x01\x00\x8E\xB5\x01\x00\x8F\xAD\x01\x00\x90\xD5\x01\x00\x91\xDD"
-     "\x01\x00"
-     "\x92\xD5\x01\x00\x93\xED\x01\x00\x94\xF5\x01\x00\x95\xFD\x01\x00\x96\xF5"
-     "\x01\x00"
-     "\x97\xED\x01\x00\x98\xD5\x01\x00\x99\xDD\x01\x00\x9A\xD5\x01\x00\x9B\xAD"
-     "\x01\x00"
-     "\x9C\xB5\x01\x00\x9D\xBD\x01\x00\x9E\xB5\x01\x00\x9F\xAD\x01\x00\xA0\x55"
-     "\x01\x00"
-     "\xA1\x5D\x01\x00\xA2\x55\x01\x00\xA3\x6D\x01\x00\xA4\x75\x01\x00\xA5\x7D"
-     "\x01\x00"
-     "\xA6\x75\x01\x00\xA7\x6D\x01\x00\xA8\x59\x01\x00\xA9\x59\x01\x00\xAA\xAD"
-     "\x02\x00"
-     "\xAB\xA5\x02\x00\xAC\xBD\x02\x00\xAD\xA5\x02\x00\xAE\xAD\x02\x00\xAF\xA5"
-     "\x02\x00"
-     "\xB0\xDD\x02\x00\xB1\xE5\x02\x00\xB2\xED\x02\x00\xB3\xE5\x02\x00\xB4\xFD"
-     "\x02\x00"
-     "\xB5\xE5\x02\x00\xB6\xED\x02\x00\xB7\xE5\x02\x00\xB8\xDD\x02\x00\xB9\xA5"
-     "\x02\x00"
-     "\xBA\xAD\x02\x00\xBB\xA5\x02\x00\xBC\xBD\x02\x00\xBD\xA5\x02\x00\xBE\xAD"
-     "\x02\x00"
-     "\xBF\xA5\x02\x00\xBD\x49\x02\x00\xBC\x41\x02\x00\xBF\x59\x02\x00\xBE\x41"
-     "\x02\x00"
-     "\xB9\x29\x02\x00\xB8\x21\x02\x00\xBB\x59\x02\x00\xBA\x21\x02\x00\xB5\x09"
-     "\x02\x00"
-     "\xB4\x01\x02\x00\xB7\x19\x02\x00\xB6\x01\x02\x00\xB1\x29\x02\x00\xB0\x21"
-     "\x02\x00"
-     "\xB3\x19\x02\x00\xB2\x21\x02\x00\xAD\x49\x02\x00\xAC\x41\x02\x00\xAF\x59"
-     "\x02\x00"
-     "\xAE\x41\x02\x00\xA9\xA9\x05\x00\xA8\xA1\x05\x00\xAB\x59\x02\x00\xAA\xA1"
-     "\x05\x00"
-     "\xA5\x89\x05\x00\xA4\x81\x05\x00\xA7\x99\x05\x00\xA6\x81\x05\x00\xA1\xA9"
-     "\x05\x00"
-     "\xA0\xA1\x05\x00\xA3\x99\x05\x00\xA2\xA1\x05\x00\x9D\x49\x05\x00\x9C\x41"
-     "\x05\x00"
-     "\x9F\x59\x05\x00\x9E\x41\x05\x00\x99\x29\x05\x00\x98\x21\x05\x00\x9B\x59"
-     "\x05\x00"
-     "\x9A\x21\x05\x00\x95\x09\x05\x00\x94\x01\x05\x00\x97\x19\x05\x00\x96\x01"
-     "\x05\x00"
-     "\x91\x29\x05\x00\x90\x21\x05\x00\x93\x19\x05\x00\x92\x21\x05\x00\x8D\x49"
-     "\x05\x00"
-     "\x8C\x41\x05\x00\x8F\x59\x05\x00\x8E\x41\x05\x00\x89\xA9\x05\x00\x88\xA1"
-     "\x05\x00"
-     "\x8B\x59\x05\x00\x8A\xA1\x05\x00\x85\x89\x05\x00\x84\x81\x05\x00\x87\x99"
-     "\x05\x00"
-     "\x86\x81\x05\x00\x81\xA9\x05\x00\x80\xA1\x05\x00\x83\x99\x05\x00\x82\xA1"
-     "\x05\x00"
-     "\x9E\x25\x1F\x00\xFA\x39\x00\x80\x06\x3A\x00\x80\x16\x3A\x00\x80\x2E\x3A"
-     "\x00\x80"
-     "\x36\x3A\x00\x80\x3A\x3A\x00\x80\x3E\x3A\x00\x80\x42\x3A\x00\x80\x4E\x3A"
-     "\x00\x80"
-     "\x52\x3A\x00\x80\x5A\x3A\x00\x80\x76\x3A\x00\x80\x7A\x3A\x00\x80\x7E\x3A"
-     "\x00\x80"
-     "\x96\x3A\x00\x80\x9A\x3A\x00\x80\xA2\x3A\x00\x80\xB6\x3A\x00\x80\xBA\x3A"
-     "\x00\x80"
-     "\xBE\x3A\x00\x80\xC6\x3A\x00\x80\xCA\x3A\x00\x80\xCE\x3A\x00\x80\xD2\x3A"
-     "\x00\x80"
-     "\xD6\x3A\x00\x80\xE6\x3A\x00\x80\xFE\x3A\x00\x80\x02\x3B\x00\x80\x0E\x3B"
-     "\x00\x80"
-     "\x12\x3B\x00\x80\x16\x3B\x00\x80\x80\xDD\x00\x00\x81\xE5\x00\x00\x82\xED"
-     "\x00\x00"
-     "\x83\xE5\x00\x00\x84\xFD\x00\x00\x85\xE5\x00\x00\x86\xED\x00\x00\x87\xE5"
-     "\x00\x00"
-     "\x88\xDD\x00\x00\x89\x7D\x00\x00\x8A\x75\x00\x00\x8B\x4D\x00\x00\x8C\x55"
-     "\x00\x00"
-     "\x8D\x5D\x00\x00\x8E\x55\x00\x00\x8F\x4D\x00\x00\x90\x35\x00\x00\x91\x3D"
-     "\x00\x00"
-     "\x92\x35\x00\x00\x93\x01\x1C\x00\x94\xD1\x1F\x00\x95\xD1\x1F\x00\x1A\x3B"
-     "\x00\x80"
-     "\x22\x3B\x00\x80\x26\x3B\x00\x80\x2A\x3B\x00\x80\x36\x3B\x00\x80\x3E\x3B"
-     "\x00\x80"
-     "\x46\x3B\x00\x80\x4A\x3B\x00\x80\x52\x3B\x00\x80\x56\x3B\x00\x80\x62\x3B"
-     "\x00\x80"
-     "\x6A\x3B\x00\x80\x80\x01\x50\x02\x6A\x24\x00\x80\x99\x2D\x00\x80\x9D\x2D"
-     "\x00\x80"
-     "\x52\x25\x00\x80\xE3\x98\x6D\x00\xE0\x24\x00\x80\xE4\x24\x00\x80\x82\x7C"
-     "\x02\x00"
-     "\xEF\x1C\x1A\x00\x99\x05\x00\x00\xE8\x24\x00\x80\xBE\x7C\x02\x00\x9E\x09"
-     "\x00\x00"
-     "\x80\x48\x02\x00\x81\x5C\x03\x00\x82\x04\x01\x00\x83\x4C\x07\x00\x84\x90"
-     "\x05\x00"
-     "\x85\x54\x0B\x00\x86\xD8\x08\x00\x99\xCD\x00\x00\x88\x00\x0C\x00\x89\x64"
-     "\x0F\x00"
-     "\x8A\x68\x0D\x00\x8B\x6C\x33\x00\x8C\x00\x38\x00\x8D\x74\x3A\x00\x8E\x78"
-     "\x3B\x00"
-     "\x8F\x7C\x39\x00\xE3\x88\x19\x00\xE3\xA4\x02\x00\xE3\x0C\x02\x00\x82\xB4"
-     "\x02\x00"
-     "\xEC\x24\x00\x80\xBE\x74\x02\x00\x9E\x09\x00\x00\xE3\x34\x02\x00\x82\x6C"
-     "\x02\x00"
-     "\xF0\x24\x00\x80\x99\x0D\x00\x00\xEF\x90\x02\x00\xEF\xBC\x1E\x00\xEF\x74"
-     "\x02\x00"
-     "\xEF\xD4\x02\x00\x9E\x19\x00\x00\xE3\x1C\x02\x00\xBE\x00\x02\x00\x99\x05"
-     "\x00\x00"
-     "\xF4\x24\x00\x80\xB6\xDD\x00\x00\x82\x78\x02\x00\x9E\x15\x00\x00\xBE\x7C"
-     "\x02\x00"
-     "\xBA\xF5\x00\x00\x82\x74\x02\x00\xB8\xBD\x1A\x00\xB9\xFD\x00\x00\xEF\xD4"
-     "\x02\x00"
-     "\x99\x0D\x00\x00\xF8\x24\x00\x80\xBE\x80\x03\x00\x86\x24\x6D\x00\x9E\xFD"
-     "\x01\x00"
-     "\x82\x78\x02\x00\xE3\xE0\x03\x00\xE3\x10\x02\x00\x99\x0D\x00\x00\xFC\x24"
-     "\x00\x80"
-     "\xBE\x4C\x02\x00\xEF\xC4\x02\x00\x8F\x2C\x02\x00\x9E\x31\x00\x00\x8D\x20"
-     "\x02\x00"
-     "\x82\x6C\x02\x00\x8B\xA4\x02\x00\x99\x79\x00\x00\xEF\x38\x03\x00\x00\x25"
-     "\x00\x80"
-     "\x97\xC4\x03\x00\xBE\x60\x02\x00\x95\x00\x02\x00\x9E\x11\x00\x00\x93\x44"
-     "\x02\x00"
-     "\xE3\x60\x02\x00\x91\x58\x02\x00\x82\x74\x02\x00\x9F\x78\x03\x00\xEF\xD8"
-     "\x02\x00"
-     "\x9D\x6C\x03\x00\x99\x09\x00\x00\x9B\xA0\x03\x00\x04\x25\x00\x80\x99\xE0"
-     "\x03\x00"
-     "\xA6\x04\x00\x00\x9E\xFD\x00\x00\xA4\x04\x00\x00\xE3\xE4\x02\x00\xE3\x30"
-     "\x02\x00"
-     "\xBE\x68\x02\x00\x08\x25\x00\x80\xA1\xC4\x03\x00\xEF\xE8\x02\x00\xAF\x68"
-     "\x00\x00"
-     "\x82\x60\x02\x00\x99\x35\x00\x00\xBE\x08\x02\x00\x82\x18\x02\x00\xA8\x28"
-     "\x00\x00"
-     "\xEF\xD0\x02\x00\x9E\x59\x00\x00\x0C\x25\x00\x80\x99\x0D\x00\x00\xB5\xC4"
-     "\x00\x00"
-     "\xB2\x04\x00\x00\xE3\x68\x02\x00\xBE\x7C\x02\x00\x9E\x19\x00\x00\x82\x6C"
-     "\x02\x00"
-     "\xEF\xC4\x02\x00\x99\x05\x00\x00\x10\x25\x00\x80\x9E\x1D\x00\x00\xBB\xA0"
-     "\x00\x00"
-     "\xB8\xE8\x00\x00\x82\x94\x01\x00\xE3\xE8\x01\x00\xBE\x8C\x01\x00\xE3\xFC"
-     "\x01\x00"
-     "\x99\x1D\x00\x00\x14\x25\x00\x80\xBE\x74\x02\x00\x9E\x09\x00\x00\xE3\x34"
-     "\x02\x00"
-     "\x82\x60\x02\x00\x99\x0D\x00\x00\x18\x25\x00\x80\xEF\xE8\x02\x00\xEF\xCC"
-     "\x02\x00"
-     "\x9E\x1D\x00\x00\xEF\xD4\x02\x00\xBE\x70\x02\x00\xE3\x04\x02\x00\x9A\x19"
-     "\x00\x00"
-     "\x99\x19\x00\x00\x82\x60\x02\x00\x1C\x25\x00\x80\xBE\x48\x02\x00\xE3\x3C"
-     "\x02\x00"
-     "\x20\x25\x00\x80\x9F\x3D\x00\x00\x9E\x35\x00\x00\xEF\xEC\x02\x00\x82\x88"
-     "\x02\x00"
-     "\xEF\xDC\x02\x00\x24\x25\x00\x80\x28\x25\x00\x80\xBE\x94\x02\x00\x99\xF5"
-     "\x00\x00"
-     "\xE3\x04\x02\x00\x82\x48\x02\x00\x9A\xF5\x00\x00\x9E\xE5\x00\x00\x9F\xED"
-     "\x00\x00"
-     "\x2C\x25\x00\x80\xE3\x2C\x02\x00\xBE\x18\x02\x00\x99\x3D\x00\x00\x9A\x19"
-     "\x00\x00"
-     "\xEF\xA4\x02\x00\x30\x25\x00\x80\xEF\xEC\x02\x00\x82\x0C\x02\x00\x34\x25"
-     "\x00\x80"
-     "\x9E\x59\x00\x00\x9F\x75\x00\x00\x38\x25\x00\x80\x99\x8D\x01\x00\x99\x6D"
-     "\x00\x00"
-     "\x82\x48\x02\x00\xE3\x04\x02\x00\x9A\x15\x00\x00\x9E\x5D\x00\x00\x9F\x2D"
-     "\x00\x00"
-     "\xEF\x10\x03\x00\x9E\xAD\x01\x00\xBE\x68\x02\x00\x9A\xE9\x01\x00\x99\x05"
-     "\x00\x00"
-     "\x43\x25\x00\x80\xBE\x70\x02\x00\xE3\x18\x02\x00\x9E\xF5\x01\x00\x9F\x11"
-     "\x00\x00"
-     "\xBE\x84\x03\x00\x82\xBC\x03\x00\x99\x41\x00\x00\x47\x25\x00\x80\x9B\x05"
-     "\x00\x00"
-     "\xBD\x00\x00\x80\x82\x54\x02\x00\xC2\x00\x00\x80\x9F\x75\x00\x00\xEF\xDC"
-     "\x02\x00"
-     "\x9D\xB0\x03\x00\x9C\x11\x00\x00\x5E\x25\x00\x80\x4B\x25\x00\x80\xBE\x7C"
-     "\x02\x00"
-     "\x9E\x15\x00\x00\x82\x54\x02\x00\xE3\x74\x02\x00\x9A\x2A\x00\x80\xE3\x04"
-     "\x02\x00"
-     "\x99\x15\x00\x00\xEF\xD4\x02\x00\x82\x98\x02\x00\xEF\xE4\x02\x00\xBE\x70"
-     "\x02\x00"
-     "\x35\x26\x00\x80\x9E\x0D\x00\x00\xD3\x00\x00\x00\xB3\x08\x04\x00\xEF\xF0"
-     "\x02\x00"
-     "\xB1\xDC\x03\x00\xBE\x50\x02\x00\xB7\x5C\x04\x00\x99\x09\x00\x00\xB5\x00"
-     "\x04\x00"
-     "\x98\x26\x00\x80\xBB\xAC\x04\x00\x9E\x39\x00\x00\xB9\x70\x04\x00\x82\x74"
-     "\x02\x00"
-     "\xBF\xA0\x04\x00\xE3\xD4\x02\x00\xBD\xBC\x04\x00\x99\x7D\x00\x00\xDB\x26"
-     "\x00\x80"
-     "\xE3\x60\x02\x00\xBE\x60\x02\x00\x5A\x3B\x00\x80\x9E\x11\x00\x00\xA6\x0C"
-     "\x03\x00"
-     "\xE3\x3C\x02\x00\x99\x01\x04\x00\xAB\x30\x03\x00\x9E\x39\x04\x00\xEF\x10"
-     "\x05\x00"
-     "\x82\x54\x02\x00\xAF\xB8\x03\x00\xEF\xC4\x02\x00\xAD\xC0\x03\x00\x82\x90"
-     "\x05\x00"
-     "\xBE\x9C\x05\x00\xE3\xF0\x05\x00\x99\x05\x00\x00\x31\x27\x00\x80\xBE\x70"
-     "\x02\x00"
-     "\xE3\x00\x02\x00\x9E\x15\x00\x00\x82\x58\x02\x00\x99\x09\x00\x00\xEF\xF0"
-     "\x02\x00"
-     "\x56\x27\x00\x80\xBE\x64\x02\x00\x9E\x19\x00\x00\xEF\xDC\x02\x00\x82\x60"
-     "\x02\x00"
-     "\x99\x79\x00\x00\xE3\x70\x02\x00\x6B\x27\x00\x80\xBE\x7C\x02\x00\x9E\x09"
-     "\x00\x00"
-     "\x73\x27\x00\x80\xE3\x04\x02\x00\x82\x60\x02\x00\x99\x0D\x00\x00\xBE\x70"
-     "\x02\x00"
-     "\xEF\xFC\x02\x00\x9E\x39\x00\x00\x82\x70\x02\x00\xEF\xC0\x02\x00\x8F\x27"
-     "\x00\x80"
-     "\x99\x0D\x00\x00\xBE\x80\x02\x00\xE3\xE4\x02\x00\x9E\xF1\x00\x00\xE3\x14"
-     "\x02\x00"
-     "\x97\x27\x00\x80\x99\x1D\x00\x00\x82\x60\x02\x00\x9E\x15\x00\x00\xE3\x34"
-     "\x02\x00"
-     "\x82\x6C\x02\x00\xBE\x44\x02\x00\x99\x1D\x00\x00\xEF\x90\x02\x00\xEF\x34"
-     "\x02\x00"
-     "\x9F\x27\x00\x80\xEF\xE4\x02\x00\x9E\x09\x00\x00\xE3\x00\x02\x00\xBE\x00"
-     "\x02\x00"
-     "\x82\x7C\x02\x00\xB1\x27\x00\x80\x99\x1D\x00\x00\xBE\x74\x02\x00\x9E\x15"
-     "\x00\x00"
-     "\xE3\x30\x02\x00\xBD\x27\x00\x80\x82\x68\x02\x00\x99\x09\x00\x00\xEF\xE0"
-     "\x02\x00"
-     "\xEF\xDC\x02\x00\xBE\x74\x02\x00\x9E\x19\x00\x00\xC5\x27\x00\x80\x82\x6C"
-     "\x02\x00"
-     "\x81\x48\x02\x00\x99\xF5\x01\x00\xBE\x50\x02\x00\x84\x40\x02\x00\x82\x40"
-     "\x02\x00"
-     "\x86\x88\x03\x00\xE3\x14\x02\x00\x88\xE8\x03\x00\x9E\xD9\x01\x00\xDB\x27"
-     "\x00\x80"
-     "\xEF\xD8\x02\x00\x99\x19\x00\x00\x82\x00\x02\x00\xBE\x00\x02\x00\x8F\xD8"
-     "\x03\x00"
-     "\x9E\x69\x00\x00\xE3\x58\x02\x00\x92\x4C\x03\x00\x19\x28\x00\x80\x99\x1D"
-     "\x00\x00"
-     "\x95\x00\x03\x00\xEF\xD8\x02\x00\x9A\x39\x00\x00\x98\xE8\x00\x00\x2A\x28"
-     "\x00\x80"
-     "\xE3\x54\x02\x00\x9B\xE8\x00\x00\x82\x80\x02\x00\xEF\xB4\x02\x00\xBE\x80"
-     "\x02\x00"
-     "\x4C\x28\x00\x80\x9E\xCD\x00\x00\x9F\xE1\x00\x00\xE3\xF0\x02\x00\x57\x28"
-     "\x00\x80"
-     "\xBE\x40\x02\x00\x99\xE9\x00\x00\x9A\x15\x00\x00\x82\x28\x02\x00\xA7\x28"
-     "\x00\x80"
-     "\xBB\x28\x00\x80\x9E\xD5\x00\x00\x9F\x21\x00\x00\xBE\x2C\x02\x00\x82\xD0"
-     "\x01\x00"
-     "\xEF\xD0\x02\x00\xAF\x78\x00\x00\xB0\x1C\x00\x00\xB1\xC8\x00\x00\xB2\xC0"
-     "\x00\x00"
-     "\xDD\x28\x00\x80\x99\x71\x00\x00\xEB\x28\x00\x80\xE3\x4C\x02\x00\x9A\x79"
-     "\x00\x00"
-     "\x9F\x45\x00\x00\x9E\x45\x00\x00\xEF\xA0\x02\x00\xBE\x84\x01\x00\x33\x29"
-     "\x00\x80"
-     "\xBD\xE4\x00\x00\x41\x29\x00\x80\xBF\x2D\x03\x00\x99\xCD\x03\x00\xE3\xB8"
-     "\x01\x00"
-     "\xE3\x18\x02\x00\x9A\xD9\x03\x00\x9E\xDD\x03\x00\x9F\xC1\x03\x00\x82\x58"
-     "\x02\x00"
-     "\xBE\x24\x02\x00\x9E\x2A\x00\x80\x99\xD5\x03\x00\x9A\xD1\x03\x00\x82\x78"
-     "\x02\x00"
-     "\x99\x11\x00\x00\xEF\x20\x01\x00\xEF\xDC\x02\x00\xBE\x08\x02\x00\x9E\xB1"
-     "\x03\x00"
-     "\x9F\xBD\x03\x00\xE3\x6C\x02\x00\x9E\x6D\x00\x00\xE3\x0C\x02\x00\x82\x68"
-     "\x02\x00"
-     "\x99\x31\x00\x00\xBE\x40\x02\x00\xEF\xC4\x02\x00\x9E\x0D\x00\x00\xA2\x2A"
-     "\x00\x80"
-     "\x99\x19\x00\x00\x82\x78\x02\x00\xA6\x2A\x00\x80\xEF\xBC\x02\x00\xBE\x8C"
-     "\x02\x00"
-     "\xE3\xF8\x02\x00\xE3\x1C\x02\x00\x9E\xFD\x00\x00\x99\x19\x00\x00\x82\x78"
-     "\x02\x00"
-     "\xAA\x2A\x00\x80\xBE\x40\x02\x00\xE3\x30\x02\x00\x9E\x35\x00\x00\x82\x78"
-     "\x02\x00"
-     "\x9E\x15\x00\x00\xEF\x9C\x02\x00\xEF\x34\x02\x00\xEF\xD4\x02\x00\x99\x05"
-     "\x00\x00"
-     "\xB6\x2A\x00\x80\xB0\x00\x00\x00\xB1\x00\x00\x00\xBE\x18\x02\x00\xB3\x00"
-     "\x00\x00"
-     "\x16\x27\x00\x80\xB5\x00\x00\x00\xBE\x1A\x00\x80\xC2\x1A\x00\x80\xB8\x00"
-     "\x00\x00"
-     "\xB9\x00\x00\x00\xBA\x00\x00\x00\xBB\x00\x00\x00\xC6\x1A\x00\x80\xBD\x00"
-     "\x00\x00"
-     "\xBE\x00\x00\x00\xBF\x00\x00\x00\xC0\x00\x00\x00\xB1\x01\x0C\x00\xB2\x09"
-     "\x0F\x00"
-     "\xB3\xD5\x01\x00\xB4\x15\x0F\x00\xB5\xC9\x01\x00\xB6\xE9\x01\x00\xB7\x0D"
-     "\x0F\x00"
-     "\xB8\x35\x0F\x00\xB9\x65\x0F\x00\xBA\x69\x00\x00\xBB\x69\x00\x00\xBC\x69"
-     "\x00\x00"
-     "\xBD\x55\x00\x00\xBE\x5D\x00\x00\xBF\x4D\x00\x00\xCA\x1A\x00\x80\xCE\x1A"
-     "\x00\x80"
-     "\xD2\x00\x00\x00\xD3\x00\x00\x00\xD4\x00\x00\x00\xD5\x00\x00\x00\xD2\x1A"
-     "\x00\x80"
-     "\xD7\x00\x00\x00\xD6\x1A\x00\x80\xDA\x1A\x00\x80\xDA\x00\x00\x00\xDB\x00"
-     "\x00\x00"
-     "\xDE\x1A\x00\x80\xEA\x19\x00\x80\xDE\x00\x00\x00\xFE\x19\x00\x80\xE0\x00"
-     "\x00\x00"
-     "\x4A\x1A\x00\x80\xE2\x00\x00\x00\x4E\x1A\x00\x80\x52\x1A\x00\x80\xE5\x00"
-     "\x00\x00"
-     "\xE6\x00\x00\x00\xE7\x00\x00\x00\x56\x1A\x00\x80\x5A\x1A\x00\x80\x5E\x1A"
-     "\x00\x80"
-     "\x62\x1A\x00\x80\x66\x1A\x00\x80\xED\x00\x00\x00\x6A\x1A\x00\x80\xEF\x00"
-     "\x00\x00"
-     "\xF0\x00\x00\x00\x6E\x1A\x00\x80\x72\x1A\x00\x80\xF3\x00\x00\x00\xF4\x00"
-     "\x00\x00"
-     "\x76\x1A\x00\x80\xF6\x00\x00\x00\x07\x27\x00\x80\x7A\x1A\x00\x80\x7E\x1A"
-     "\x00\x80"
-     "\x82\x1A\x00\x80\x86\x1A\x00\x80\xFC\x00\x00\x00\xFD\x00\x00\x00\xFE\x00"
-     "\x00\x00"
-     "\x8A\x1A\x00\x80\x9A\x19\x00\x80\x9E\x19\x00\x80\x8E\x1A\x00\x80\x92\x1A"
-     "\x00\x80"
-     "\xA2\x19\x00\x80\xA6\x19\x00\x80\x96\x1A\x00\x80\x9A\x1A\x00\x80\x46\x1A"
-     "\x00\x80"
-     "\xE2\x19\x00\x80\xE6\x19\x00\x80\xE2\x1A\x00\x80\xE6\x1A\x00\x80\xEA\x1A"
-     "\x00\x80"
-     "\xEE\x1A\x00\x80\xF2\x1A\x00\x80\xAA\x19\x00\x80\xAE\x19\x00\x80\xF6\x1A"
-     "\x00\x80"
-     "\xFA\x1A\x00\x80\xB2\x19\x00\x80\xB6\x19\x00\x80\xEE\x19\x00\x80\xFE\x1A"
-     "\x00\x80"
-     "\x02\x1B\x00\x80\xF2\x19\x00\x80\xF6\x19\x00\x80\xFA\x19\x00\x80\x02\x1A"
-     "\x00\x80"
-     "\x06\x1A\x00\x80\x0A\x1A\x00\x80\x0E\x1A\x00\x80\xE1\xE0\x7D\x02\xBA\x19"
-     "\x00\x80"
-     "\xE3\x54\x69\x02\x12\x1A\x00\x80\xBE\x19\x00\x80\xC2\x19\x00\x80\xEF\x5C"
-     "\x41\x00"
-     "\xEF\x44\x4E\x00\xE1\x64\x48\x00\xE1\xA0\xA3\x03\xE3\x70\x49\x00\xE3\xF8"
-     "\x5C\x00"
-     "\xC6\x19\x00\x80\x16\x1A\x00\x80\xEF\x4C\x7D\x02\x1A\x1A\x00\x80\xCA\x19"
-     "\x00\x80"
-     "\xCE\x19\x00\x80\x1E\x1A\x00\x80\x22\x1A\x00\x80\xD2\x19\x00\x80\xD6\x19"
-     "\x00\x80"
-     "\x26\x1A\x00\x80\x2A\x1A\x00\x80\x2E\x1A\x00\x80\x32\x1A\x00\x80\x36\x1A"
-     "\x00\x80"
-     "\x06\x1B\x00\x80\x0A\x1B\x00\x80\x3A\x1A\x00\x80\x3E\x1A\x00\x80\x42\x1A"
-     "\x00\x80"
-     "\x80\x75\x0E\x00\x81\x41\x01\x00\x82\x59\x01\x00\x83\x59\x01\x00\x84\x6D"
-     "\x0E\x00"
-     "\x85\x91\x01\x00\x86\x89\x01\x00\x87\x89\x01\x00\x88\x91\x01\x00\x89\xE5"
-     "\x01\x00"
-     "\x8A\xED\x01\x00\x8B\xFD\x01\x00\x8C\xE5\x01\x00\x8D\x15\x02\x00\x8E\x1D"
-     "\x02\x00"
-     "\x8F\xB9\x0E\x00\x90\xCD\x0E\x00\x91\xD5\x0E\x00\x92\xE9\x0E\x00\x93\xE9"
-     "\x0E\x00"
-     "\x94\xF9\x0E\x00\x95\xF9\x0E\x00\x96\xE9\x0E\x00\x97\xE5\x0E\x00\x98\xA5"
-     "\x0E\x00"
-     "\x99\xAD\x0E\x00\x9A\xBD\x0E\x00\x9B\x8D\x0E\x00\x9C\x95\x0E\x00\x9D\x9D"
-     "\x0E\x00"
-     "\x9E\x95\x0E\x00\x9F\x81\x0E\x00\xA0\x89\x01\x00\xA1\x89\x01\x00\xA2\x91"
-     "\x01\x00"
-     "\xA3\x91\x01\x00\xA4\xB1\x01\x00\xA5\xB1\x01\x00\xA6\xB1\x01\x00\xA7\xB1"
-     "\x01\x00"
-     "\xA8\x91\x01\x00\xA9\x91\x01\x00\xAA\x91\x01\x00\xAB\x91\x01\x00\xAC\xF9"
-     "\x01\x00"
-     "\xAD\xF9\x01\x00\xAE\xE1\x01\x00\xAF\xE1\x01\x00\xB0\xA1\x01\x00\xB1\xA1"
-     "\x01\x00"
-     "\xB2\xA1\x01\x00\xB3\xA1\x01\x00\xB4\xA1\x01\x00\xB5\xA1\x01\x00\xB6\xA1"
-     "\x01\x00"
-     "\xB7\xA1\x01\x00\xB8\x6D\x01\x00\xB9\x51\x01\x00\xBA\x55\x01\x00\xBB\x25"
-     "\x01\x00"
-     "\xBC\x3D\x01\x00\xBD\x2D\x01\x00\xBE\x25\x01\x00\xBF\x1D\x01\x00\x9E\x1A"
-     "\x00\x80"
-     "\xA2\x1A\x00\x80\xA6\x1A\x00\x80\xAA\x1A\x00\x80\xAE\x1A\x00\x80\xB2\x1A"
-     "\x00\x80"
-     "\xB6\x1A\x00\x80\xBA\x1A\x00\x80\xD2\x2B\x00\x80\x89\x2E\x00\x80\xD6\x2B"
-     "\x00\x80"
-     "\xEE\x2B\x00\x80\xDA\x2B\x00\x80\xCA\x34\x00\x80\xE2\x2B\x00\x80\xCA\x2B"
-     "\x00\x80"
-     "\x14\x2F\x00\x80\x9D\x2E\x00\x80\xA0\x01\x00\x80\xA6\x01\x00\x80\xAC\x01"
-     "\x00\x80"
-     "\xB2\x01\x00\x80\xB8\x01\x00\x80\xBE\x01\x00\x80\xC4\x01\x00\x80\xCA\x01"
-     "\x00\x80"
-     "\xD0\x01\x00\x80\xD6\x01\x00\x80\xDC\x01\x00\x80\xE2\x01\x00\x80\xE8\x01"
-     "\x00\x80"
-     "\xEE\x01\x00\x80\x80\x65\x02\x00\x81\x6D\x02\x00\x82\x65\x02\x00\x83\x7D"
-     "\x02\x00"
-     "\x84\x65\x02\x00\x85\x6D\x02\x00\x86\x65\x02\x00\x87\x9D\x00\x00\x88\xA5"
-     "\x00\x00"
-     "\x89\xAD\x00\x00\x8A\xA5\x00\x00\x8B\xBD\x00\x00\x8C\xA5\x00\x00\x8D\xAD"
-     "\x00\x00"
-     "\x8E\xA5\x00\x00\xCC\x02\x00\x80\xD2\x02\x00\x80\xD8\x02\x00\x80\x92\x59"
-     "\x12\x00"
-     "\x93\x0D\x12\x00\x94\xF1\x00\x00\x95\xF1\x00\x00\x96\xF1\x00\x00\x97\xF1"
-     "\x00\x00"
-     "\x98\xD1\x00\x00\x99\xD1\x00\x00\x9A\x59\x12\x00\x9B\xD5\x00\x00\x9C\xCD"
-     "\x00\x00"
-     "\x9D\xB5\x00\x00\x9E\xBD\x00\x00\x9F\x25\x12\x00\xE1\x74\x47\x03\x96\x19"
-     "\x00\x80"
-     "\xE3\xCC\x6B\x03\xEF\x50\x3A\x03\xE1\x04\x0C\x03\xE1\xDC\x6E\x03\xE3\xC0"
-     "\x3B\x03"
-     "\xE3\x48\x17\x03\xDA\x19\x00\x80\xDE\x19\x00\x80\xEF\xEC\x3A\x03\xEF\x3C"
-     "\x0D\x03"
-     "\xDE\x02\x00\x80\xE1\xF0\x33\x03\xEF\xE8\x6F\x03\xE3\xE0\x23\x03\xE1\x54"
-     "\xFE\x03"
-     "\xE1\xCC\xD2\x03\xE3\xFC\xEE\x03\xE3\x94\xC3\x03\xE1\x84\x84\x03\xE1\xA8"
-     "\x98\x03"
-     "\xE3\x44\xA0\x03\xE3\x84\x8F\x03\xE4\x02\x00\x80\xEA\x02\x00\x80\xEF\x98"
-     "\xA0\x03"
-     "\xEF\x3C\x84\x03\xF0\x02\x00\x80\xF6\x02\x00\x80\xEF\xBC\xC2\x03\xEF\xAC"
-     "\x37\x03"
-     "\xE1\x74\x07\x02\xE1\x00\x6E\x02\xE3\x2C\x3A\x02\xE3\xD8\x67\x02\xE1\xEC"
-     "\xCC\x02"
-     "\xE1\x20\x32\x02\xE3\x4C\xFA\x02\xE3\x40\xD3\x02\x0E\x03\x00\x80\x1A\x03"
-     "\x00\x80"
-     "\xEF\x3C\xE9\x02\xEF\x70\x03\x02\x20\x03\x00\x80\x26\x03\x00\x80\xEF\x6C"
-     "\x0A\x02"
-     "\xEF\xD8\x7E\x02\xE1\x74\x9E\x02\xE1\x8C\xE9\x02\xE3\x1C\x92\x02\xE3\x98"
-     "\xE3\x02"
-     "\xE1\x74\x5B\x03\xE1\x80\x97\x02\xE3\x8C\x47\x03\xE3\x88\xAB\x02\x2C\x03"
-     "\x00\x80"
-     "\x32\x03\x00\x80\xEF\x2C\x47\x03\xEF\x30\x83\x02\x14\x03\x00\x80\xFC\x02"
-     "\x00\x80"
-     "\xEF\xAC\x8E\x02\xEF\x8C\x92\x02\x05\x03\x00\x80\xF4\x01\x00\x80\x0C\x02"
-     "\x00\x80"
-     "\x00\x02\x00\x80\x54\x02\x00\x80\x12\x02\x00\x80\x30\x02\x00\x80\xFA\x01"
-     "\x00\x80"
-     "\x2A\x02\x00\x80\x06\x02\x00\x80\x3C\x02\x00\x80\x6C\x02\x00\x80\x8A\x02"
-     "\x00\x80"
-     "\x84\x02\x00\x80\x78\x02\x00\x80\xC6\x02\x00\x80\x5A\x02\x00\x80\x66\x02"
-     "\x00\x80"
-     "\x7E\x02\x00\x80\x72\x02\x00\x80\x9C\x02\x00\x80\x48\x02\x00\x80\x90\x02"
-     "\x00\x80"
-     "\xBA\x02\x00\x80\x59\x00\x00\x00\xA2\x02\x00\x80\x36\x02\x00\x80\x18\x02"
-     "\x00\x80"
-     "\x4E\x02\x00\x80\x60\x02\x00\x80\x96\x02\x00\x80\x1E\x02\x00\x80\xC0\x02"
-     "\x00\x80"
-     "\x42\x02\x00\x80\xA8\x02\x00\x80\x24\x02\x00\x80\xAE\x02\x00\x80\xB4\x02"
-     "\x00\x80"
-     "\x3D\x2E\x00\x80\x1F\x30\x00\x80\xFE\x35\x00\x80\x1A\x08\x00\x80\x9E\x04"
-     "\x00\x80"
-     "\xAD\x04\x00\x80\xBC\x04\x00\x80\xCB\x04\x00\x80\xDA\x04\x00\x80\xE3\x04"
-     "\x00\x80"
-     "\xEC\x04\x00\x80\xF5\x04\x00\x80\xFE\x04\x00\x80\x2A\x05\x00\x80\x33\x05"
-     "\x00\x80"
-     "\x3C\x05\x00\x80\x3F\x05\x00\x80\x42\x05\x00\x80\x45\x05\x00\x80\x3E\x3D"
-     "\x00\x80"
-     "\x9E\x4F\x00\x80\xCE\x58\x00\x80\x2E\x6B\x00\x80\x5E\x74\x00\x80\x8E\x7D"
-     "\x00\x80"
-     "\xEE\x8F\x00\x80\x80\x89\x01\x00\x81\x89\x01\x00\x82\x99\x01\x00\x83\x99"
-     "\x01\x00"
-     "\x84\x89\x01\x00\x85\x89\x01\x00\x86\x45\x16\x00\x87\xBD\x01\x00\x4E\xA2"
-     "\x00\x80"
-     "\x7E\xAB\x00\x80\xE5\xBD\x00\x80\x1C\xC7\x00\x80\x4C\xD0\x00\x80\x7C\xD9"
-     "\x00\x80"
-     "\xAC\xE2\x00\x80\x29\xBE\x00\x80\x90\xE5\x01\x00\x91\xED\x01\x00\x92\xE5"
-     "\x01\x00"
-     "\x93\xFD\x01\x00\x94\xE5\x01\x00\x95\xED\x01\x00\x96\xE5\x01\x00\x97\x9D"
-     "\x01\x00"
-     "\x98\xA5\x01\x00\x99\xAD\x01\x00\x9A\xA5\x01\x00\x9B\xBD\x01\x00\x9C\xA5"
-     "\x01\x00"
-     "\x9D\xAD\x01\x00\x9E\xA5\x01\x00\x9F\x9D\x01\x00\xA0\x85\x06\x00\xA1\x81"
-     "\x05\x00"
-     "\xA2\x99\x05\x00\xA3\xCD\x05\x00\xA4\xD5\x05\x00\xA5\xC5\x05\x00\xA6\x1D"
-     "\x05\x00"
-     "\xA7\x09\x05\x00\xA8\x11\x05\x00\xA9\x61\x05\x00\xAA\x79\x05\x00\xAB\x79"
-     "\x05\x00"
-     "\xAC\x91\x06\x00\xAD\x91\x06\x00\xAE\x5D\x01\x00\xAF\x55\x01\x00\xB0\x2D"
-     "\x01\x00"
-     "\xB1\x35\x01\x00\xB2\x3D\x01\x00\xB3\x35\x01\x00\xB4\x2D\x01\x00\xB5\xF5"
-     "\x00\x00"
-     "\xB6\xFD\x00\x00\xB7\xF5\x00\x00\xB8\xCD\x00\x00\xB9\xD5\x00\x00\xBA\xDD"
-     "\x00\x00"
-     "\xBB\xD5\x00\x00\xBC\xCD\x00\x00\xBD\x09\x00\x00\xBE\xE5\x03\x00\x32\xB1"
-     "\x00\x80"
-     "\x98\xB1\x05\x00\x99\xB1\x05\x00\x9A\xE9\x05\x00\x9B\xE9\x05\x00\x9C\xE1"
-     "\x05\x00"
-     "\x9D\xE1\x05\x00\x9E\x19\x02\x00\xFE\xA7\x00\x80\x90\xE5\x04\x00\x91\x15"
-     "\x05\x00"
-     "\x92\x49\x05\x00\x93\x49\x05\x00\x94\x41\x05\x00\x95\x41\x05\x00\x96\x99"
-     "\x05\x00"
-     "\x97\x99\x05\x00\x88\x29\x04\x00\x89\x29\x04\x00\x8A\x39\x04\x00\x8B\x39"
-     "\x04\x00"
-     "\x8C\x29\x04\x00\x8D\x29\x04\x00\x8E\xE5\x04\x00\x8F\x9D\x04\x00\x80\x29"
-     "\x04\x00"
-     "\x81\x29\x04\x00\x82\x39\x04\x00\x83\x39\x04\x00\x84\x29\x04\x00\x85\x29"
-     "\x04\x00"
-     "\x86\x19\x04\x00\x87\x19\x04\x00\xB8\xE9\x02\x00\xB9\xE9\x02\x00\xBA\xF9"
-     "\x02\x00"
-     "\xBB\xF9\x02\x00\xBC\xE9\x02\x00\xBD\xE9\x02\x00\xBE\x19\x02\x00\xBF\x19"
-     "\x02\x00"
-     "\xB0\xE5\x02\x00\xB1\xED\x02\x00\xB2\xE5\x02\x00\xB3\xFD\x02\x00\xB4\xE5"
-     "\x02\x00"
-     "\xB5\xED\x02\x00\xB6\xE5\x02\x00\xB7\xD9\x02\x00\xA8\xE5\x02\x00\xA9\xED"
-     "\x02\x00"
-     "\xAA\xE5\x02\x00\xAB\xFD\x02\x00\xAC\xE5\x02\x00\xAD\xED\x02\x00\xAE\xE5"
-     "\x02\x00"
-     "\xAF\x9D\x02\x00\xA0\xE5\x02\x00\xA1\xED\x02\x00\xA2\xE5\x02\x00\xA3\xFD"
-     "\x02\x00"
-     "\xA4\xE5\x02\x00\xA5\xED\x02\x00\xA6\xE5\x02\x00\xA7\xDD\x02\x00\x80\x91"
-     "\x17\x00"
-     "\x81\xC5\x17\x00\x82\x29\x05\x00\x83\x29\x05\x00\x84\x01\x3C\x00\x85\x15"
-     "\x3F\x00"
-     "\x86\x1D\x3F\x00\x87\xED\x17\x00\x88\x29\x3F\x00\x89\x71\x15\x00\x8A\x51"
-     "\x14\x00"
-     "\x8B\x85\x14\x00\x8C\x95\x14\x00\x8D\x49\x14\x00\x8E\xD5\x13\x00\x8F\xA1"
-     "\x15\x00"
-     "\x90\x31\x14\x00\x91\x49\x3F\x00\x92\x59\x3F\x00\x93\x59\x3F\x00\x94\x49"
-     "\x3F\x00"
-     "\x95\x49\x3F\x00\x96\x79\x3F\x00\x97\x79\x3F\x00\x98\x49\x3F\x00\x99\x49"
-     "\x3F\x00"
-     "\x9A\x59\x3F\x00\x9B\xA1\x15\x00\x9C\x45\x3F\x00\x9D\x4D\x3F\x00\x9E\x45"
-     "\x3F\x00"
-     "\x9F\x3D\x3F\x00\xA0\xC5\x3F\x00\xA1\xCD\x3F\x00\xA2\xC5\x3F\x00\xA3\xDD"
-     "\x3F\x00"
-     "\xA4\xB9\x05\x00\xA5\xB9\x05\x00\xA6\xA9\x05\x00\xA7\xC9\x3F\x00\xA8\xF9"
-     "\x3F\x00"
-     "\xA9\xF9\x3F\x00\xAA\xC9\x3F\x00\xAB\xC9\x3F\x00\xAC\xD9\x3F\x00\xAD\xD9"
-     "\x3F\x00"
-     "\xAE\xC9\x3F\x00\xAF\xC9\x3F\x00\xB0\xB9\x3F\x00\xB1\xB9\x3F\x00\xB2\x49"
-     "\x3F\x00"
-     "\xB3\x49\x3F\x00\xB4\x59\x3F\x00\xB5\x59\x3F\x00\xB6\x49\x3F\x00\xB7\x49"
-     "\x3F\x00"
-     "\xB8\x79\x3F\x00\xB9\x79\x3F\x00\xBA\x49\x3F\x00\xBB\x49\x3F\x00\xBC\x59"
-     "\x3F\x00"
-     "\xBD\x59\x3F\x00\xBE\x49\x3F\x00\xBF\x49\x3F\x00\xBD\x2C\x00\x80\xCE\x2B"
-     "\x00\x80"
-     "\x12\x2C\x00\x80\x75\x32\x00\x80\x35\x32\x00\x80\x82\x35\x00\x80\x09\x2E"
-     "\x00\x80"
-     "\x3A\x34\x00\x80\xBE\x38\x00\x80\x96\x35\x00\x80\x51\x2D\x00\x80\xC6\x35"
-     "\x00\x80"
-     "\xD2\x34\x00\x80\x46\x39\x00\x80\x8D\x2C\x00\x80\xB9\x2D\x00\x80\x2E\x33"
-     "\x00\x80"
-     "\x5E\x3A\x00\x80\x51\x2C\x00\x80\xE1\x2C\x00\x80\xD2\x32\x00\x80\xF0\x2F"
-     "\x00\x80"
-     "\xFD\x2D\x00\x80\x8D\x2D\x00\x80\x6C\x2F\x00\x80\x60\x2F\x00\x80\x22\x35"
-     "\x00\x80"
-     "\x4D\x2C\x00\x80\xD2\x38\x00\x80\xA9\x2D\x00\x80\x01\x2F\x00\x80\x48\x05"
-     "\x00\x80"
-     "\x4B\x05\x00\x80\x4E\x05\x00\x80\x51\x05\x00\x80\x7D\x05\x00\x80\x80\x05"
-     "\x00\x80"
-     "\x83\x05\x00\x80\x86\x05\x00\x80\x89\x05\x00\x80\x8C\x05\x00\x80\x8F\x05"
-     "\x00\x80"
-     "\x92\x05\x00\x80\x95\x05\x00\x80\x98\x05\x00\x80\xB8\x05\x00\x80\x79\x04"
-     "\x00\x80"
-     "\x18\x05\x00\x80\x6B\x05\x00\x80\xA6\x05\x00\x80\xCC\x05\x00\x80\xE3\x05"
-     "\x00\x80"
-     "\x00\x06\x00\x80\x17\x06\x00\x80\x2E\x06\x00\x80\x75\x03\x00\x80\x8E\x03"
-     "\x00\x80"
-     "\xA7\x03\x00\x80\x0E\x07\x00\x80\xD4\x06\x00\x80\xD8\x06\x00\x80\x7D\x07"
-     "\x00\x80"
-     "\x66\x25\x00\x80\x9F\x25\x00\x80\x80\xB9\x01\x00\x81\xB9\x01\x00\x82\xC9"
-     "\x01\x00"
-     "\x83\xC9\x01\x00\x84\xD9\x01\x00\x85\xD9\x01\x00\x86\xC9\x01\x00\x87\xC9"
-     "\x01\x00"
-     "\x88\xF9\x01\x00\x89\xF9\x01\x00\x8A\xC9\x01\x00\x8B\xC9\x01\x00\x8C\xD9"
-     "\x01\x00"
-     "\x8D\xD9\x01\x00\x8E\xC9\x01\x00\x8F\xC9\x01\x00\x90\xB9\x01\x00\x91\xB9"
-     "\x01\x00"
-     "\x92\x01\x04\x00\x93\x4D\x03\x00\x94\x55\x03\x00\x95\x55\x05\x00\x96\x5D"
-     "\x05\x00"
-     "\x97\x55\x05\x00\x98\x6D\x05\x00\x99\x75\x05\x00\x9A\x7D\x05\x00\x9B\x75"
-     "\x05\x00"
-     "\x9C\x6D\x05\x00\x9D\x35\x05\x00\x9E\x3D\x05\x00\x9F\x35\x05\x00\xA0\xCD"
-     "\x05\x00"
-     "\xA1\xD5\x05\x00\xA2\xDD\x05\x00\xA3\xD5\x05\x00\xA4\x99\x03\x00\xA5\x99"
-     "\x03\x00"
-     "\xA6\x89\x03\x00\xA7\x89\x03\x00\xA8\xB9\x03\x00\xA9\xF9\x05\x00\xAA\x29"
-     "\x05\x00"
-     "\xAB\x29\x05\x00\xAC\x39\x05\x00\xAD\x39\x05\x00\xAE\xA5\x03\x00\xAF\x9D"
-     "\x03\x00"
-     "\xB0\xE5\x03\x00\xB1\xED\x03\x00\xB2\xE5\x03\x00\xB3\xFD\x03\x00\xB4\xE5"
-     "\x03\x00"
-     "\xB5\xED\x03\x00\xB6\xE5\x03\x00\xB7\x9D\x03\x00\xB8\xA5\x03\x00\xB9\xAD"
-     "\x03\x00"
-     "\xBA\xA5\x03\x00\xBB\x7D\x05\x00\xBC\x65\x05\x00\xBD\x6D\x05\x00\xBE\x65"
-     "\x05\x00"
-     "\xBF\xAD\x03\x00\xD0\x25\x00\x80\xF5\x25\x00\x80\xE6\x27\x00\x80\xF1\x27"
-     "\x00\x80"
-     "\xF5\x27\x00\x80\xF9\x27\x00\x80\xFD\x27\x00\x80\x5C\x29\x00\x80\x9B\x29"
-     "\x00\x80"
-     "\xC3\x29\x00\x80\xC7\x29\x00\x80\xEC\x29\x00\x80\xF4\x29\x00\x80\x10\x2A"
-     "\x00\x80"
-     "\x22\x2A\x00\x80\x26\x2A\x00\x80\x2A\x2A\x00\x80\x42\x2A\x00\x80\x5D\x2A"
-     "\x00\x80"
-     "\x78\x2A\x00\x80\x46\x2C\x00\x80\x6A\x25\x00\x80\x77\x25\x00\x80\x84\x25"
-     "\x00\x80"
-     "\x91\x25\x00\x80\xA3\x25\x00\x80\xB0\x25\x00\x80\xC2\x25\x00\x80\xD4\x25"
-     "\x00\x80"
-     "\xE4\x25\x00\x80\xF9\x25\x00\x80\x03\x26\x00\x80\x11\x26\x00\x80\x1B\x26"
-     "\x00\x80"
-     "\x28\x26\x00\x80\x39\x26\x00\x80\x43\x26\x00\x80\x9C\x26\x00\x80\xA3\x26"
-     "\x00\x80"
-     "\x51\x26\x00\x80\xAD\x26\x00\x80\x5E\x26\x00\x80\x65\x26\x00\x80\x75\x26"
-     "\x00\x80"
-     "\x88\x26\x00\x80\xDF\x26\x00\x80\xE9\x26\x00\x80\xBE\x26\x00\x80\xCE\x26"
-     "\x00\x80"
-     "\xFD\x26\x00\x80\x1D\x27\x00\x80\x27\x27\x00\x80\x39\x27\x00\x80\x46\x27"
-     "\x00\x80"
-     "\x5E\x27\x00\x80\x7B\x27\x00\x80\x85\x27\x00\x80\xA3\x27\x00\x80\xC9\x27"
-     "\x00\x80"
-     "\xDF\x27\x00\x80\xD4\x27\x00\x80\xEA\x27\x00\x80\x01\x28\x00\x80\x0F\x28"
-     "\x00\x80"
-     "\x80\x00\x00\x00\x2E\x28\x00\x80\x3E\x28\x00\x80\x1D\x28\x00\x80\x26\x2C"
-     "\x00\x80"
-     "\x5B\x28\x00\x80\x6B\x28\x00\x80\x75\x28\x00\x80\x50\x28\x00\x80\x80\x28"
-     "\x00\x80"
-     "\x90\x28\x00\x80\xAB\x28\x00\x80\x9D\x28\x00\x80\xC3\x28\x00\x80\xEF\x28"
-     "\x00\x80"
-     "\xF6\x28\x00\x80\xD3\x28\x00\x80\x00\x29\x00\x80\x0A\x29\x00\x80\xE1\x28"
-     "\x00\x80"
-     "\x45\x29\x00\x80\x37\x29\x00\x80\x18\x29\x00\x80\x52\x29\x00\x80\x1F\x29"
-     "\x00\x80"
-     "\x29\x29\x00\x80\x60\x29\x00\x80\x6D\x29\x00\x80\x77\x29\x00\x80\x81\x29"
-     "\x00\x80"
-     "\x8B\x29\x00\x80\x9F\x29\x00\x80\xAC\x29\x00\x80\xB3\x29\x00\x80\xCB\x29"
-     "\x00\x80"
-     "\xD2\x29\x00\x80\xDF\x29\x00\x80\xF8\x29\x00\x80\x02\x2A\x00\x80\x14\x2A"
-     "\x00\x80"
-     "\x2E\x2A\x00\x80\x3B\x2A\x00\x80\x46\x2A\x00\x80\x50\x2A\x00\x80\x61\x2A"
-     "\x00\x80"
-     "\x68\x2A\x00\x80\x80\x2A\x00\x80\x5B\x03\x00\x80\x7E\x04\x00\x80\x1D\x05"
-     "\x00\x80"
-     "\x70\x05\x00\x80\xAB\x05\x00\x80\xD1\x05\x00\x80\xE8\x05\x00\x80\x05\x06"
-     "\x00\x80"
-     "\x1C\x06\x00\x80\x33\x06\x00\x80\x7B\x03\x00\x80\x94\x03\x00\x80\xAD\x03"
-     "\x00\x80"
-     "\xC0\x03\x00\x80\xD3\x03\x00\x80\xE6\x03\x00\x80\xF9\x03\x00\x80\x0D\x26"
-     "\x00\x80"
-     "\x4D\x26\x00\x80\xBA\x26\x00\x80\xF9\x26\x00\x80\x0B\x27\x00\x80\x35\x27"
-     "\x00\x80"
-     "\x5A\x27\x00\x80\x6F\x27\x00\x80\x0C\x04\x00\x80\x1F\x04\x00\x80\x32\x04"
-     "\x00\x80"
-     "\x98\x04\x00\x80\xA7\x04\x00\x80\xB6\x04\x00\x80\xC5\x04\x00\x80\xD4\x04"
-     "\x00\x80"
-     "\x77\x27\x00\x80\x93\x27\x00\x80\x9B\x27\x00\x80\xAD\x27\x00\x80\xB9\x27"
-     "\x00\x80"
-     "\xC1\x27\x00\x80\xD0\x27\x00\x80\x0B\x28\x00\x80\x14\x07\x00\x80\xB4\x06"
-     "\x00\x80"
-     "\x72\x06\x00\x80\x81\x06\x00\x80\xF6\x07\x00\x80\x03\x08\x00\x80\xC0\x06"
-     "\x00\x80"
-     "\xC3\x06\x00\x80\x48\x28\x00\x80\x7C\x28\x00\x80\xBF\x28\x00\x80\x14\x29"
-     "\x00\x80"
-     "\x7C\x2A\x00\x80\x8A\x2A\x00\x80\x8E\x2A\x00\x80\x92\x2A\x00\x80\xC7\x06"
-     "\x00\x80"
-     "\x2D\x07\x00\x80\x10\x30\x00\x80\xFE\x31\x00\x80\x0D\x2F\x00\x80\xEF\x31"
-     "\x00\x80"
-     "\x79\x32\x00\x80\x00\x08\x00\x80\xDC\x07\x00\x80\x15\x0C\x00\x80\x90\x07"
-     "\x00\x80"
-     "\x3A\x07\x00\x80\x3D\x07\x00\x80\x93\x07\x00\x80\xFB\x06\x00\x80\x92\x06"
-     "\x00\x80"
-     "\x40\x07\x00\x80\x06\x08\x00\x80\xDF\x07\x00\x80\x19\x0C\x00\x80\x1D\x0C"
-     "\x00\x80"
-     "\x9C\x07\x00\x80\x45\x07\x00\x80\x1B\x07\x00\x80\x48\x07\x00\x80\x9F\x07"
-     "\x00\x80"
-     "\xFE\x06\x00\x80\x5A\x08\x00\x80\x21\x0C\x00\x80\xA7\x07\x00\x80\xBD\x06"
-     "\x00\x80"
-     "\x4F\x07\x00\x80\xF2\x06\x00\x80\xE5\x07\x00\x80\x25\x0C\x00\x80\xAA\x07"
-     "\x00\x80"
-     "\x9C\x06\x00\x80\x52\x07\x00\x80\xAD\x07\x00\x80\x9F\x06\x00\x80\x8A\x07"
-     "\x00\x80"
-     "\x55\x07\x00\x80\xBE\x25\x00\x80\xB1\x07\x00\x80\xA3\x06\x00\x80\x8D\x07"
-     "\x00\x80"
-     "\x59\x07\x00\x80\xCD\x07\x00\x80\xD3\x07\x00\x80\x5D\x07\x00\x80\xB9\x07"
-     "\x00\x80"
-     "\x02\x07\x00\x80\x28\x08\x00\x80\x2C\x08\x00\x80\x34\x08\x00\x80\x17\x08"
-     "\x00\x80"
-     "\xEB\x07\x00\x80\x29\x0C\x00\x80\xC0\x07\x00\x80\x1E\x08\x00\x80\xEE\x07"
-     "\x00\x80"
-     "\x2D\x0C\x00\x80\xC3\x07\x00\x80\x64\x07\x00\x80\x21\x08\x00\x80\xF1\x07"
-     "\x00\x80"
-     "\x31\x0C\x00\x80\xC6\x07\x00\x80\x67\x07\x00\x80\x6A\x07\x00\x80\xC9\x07"
-     "\x00\x80"
-     "\x65\x06\x00\x80\x85\x06\x00\x80\x96\x06\x00\x80\x99\x06\x00\x80\xAB\x06"
-     "\x00\x80"
-     "\xA7\x06\x00\x80\xB7\x06\x00\x80\x06\x07\x00\x80\x0B\x07\x00\x80\x11\x07"
-     "\x00\x80"
-     "\x2A\x07\x00\x80\x4C\x07\x00\x80\x61\x07\x00\x80\x73\x07\x00\x80\x76\x07"
-     "\x00\x80"
-     "\x79\x07\x00\x80\x81\x07\x00\x80\xA3\x07\x00\x80\xB5\x07\x00\x80\x09\x08"
-     "\x00\x80"
-     "\xFB\x07\x00\x80\x0C\x08\x00\x80\x14\x08\x00\x80\x48\x08\x00\x80\x51\x08"
-     "\x00\x80"
-     "\x79\x08\x00\x80\x71\x08\x00\x80\x79\x06\x00\x80\x74\x04\x00\x80\x13\x05"
-     "\x00\x80"
-     "\x66\x05\x00\x80\xA1\x05\x00\x80\xC7\x05\x00\x80\xDE\x05\x00\x80\xFB\x05"
-     "\x00\x80"
-     "\x12\x06\x00\x80\x29\x06\x00\x80\x6F\x03\x00\x80\x88\x03\x00\x80\xA1\x03"
-     "\x00\x80"
-     "\xBA\x03\x00\x80\xCD\x03\x00\x80\xE0\x03\x00\x80\xF3\x03\x00\x80\x06\x04"
-     "\x00\x80"
-     "\x19\x04\x00\x80\x2C\x04\x00\x80\x92\x04\x00\x80\xA1\x04\x00\x80\xB0\x04"
-     "\x00\x80"
-     "\xBF\x04\x00\x80\xCE\x04\x00\x80\xDD\x04\x00\x80\xE6\x04\x00\x80\xEF\x04"
-     "\x00\x80"
-     "\xF8\x04\x00\x80\x01\x05\x00\x80\x2D\x05\x00\x80\x36\x05\x00\x80\xF7\x06"
-     "\x00\x80"
-     "\x6E\x3B\x00\x80\x72\x3B\x00\x80\x76\x3B\x00\x80\x7A\x3B\x00\x80\x7E\x3B"
-     "\x00\x80"
-     "\x86\x3B\x00\x80\xE3\xE4\x37\x00\x8A\x3B\x00\x80\x8E\x3B\x00\x80\x92\x3B"
-     "\x00\x80"
-     "\x96\x3B\x00\x80\x9A\x3B\x00\x80\x9E\x3B\x00\x80\xA2\x3B\x00\x80\xA6\x3B"
-     "\x00\x80"
-     "\xAA\x3B\x00\x80\xAE\x3B\x00\x80\xB2\x3B\x00\x80\xEF\x44\x37\x00\xB6\x3B"
-     "\x00\x80"
-     "\xBA\x3B\x00\x80\xE3\x10\x33\x00\xE3\xB8\x30\x00\xE3\x9C\x30\x00\xE3\x74"
-     "\x30\x00"
-     "\xE3\x88\x33\x00\xE3\x6C\x33\x00\xE3\x44\x33\x00\xE3\x38\x33\x00\xEF\x40"
-     "\x33\x00"
-     "\xEF\xA0\x33\x00\xEF\x8C\x33\x00\xEF\xE8\x33\x00\xEF\xD4\x33\x00\xEF\x20"
-     "\x30\x00"
-     "\xEF\x70\x30\x00\xEF\x54\x30\x00\xBE\x3B\x00\x80\xC2\x3B\x00\x80\xC6\x3B"
-     "\x00\x80"
-     "\xCA\x3B\x00\x80\xCE\x3B\x00\x80\xD2\x3B\x00\x80\xD6\x3B\x00\x80\xDA\x3B"
-     "\x00\x80"
-     "\xDE\x3B\x00\x80\xE2\x3B\x00\x80\xE6\x3B\x00\x80\xEA\x3B\x00\x80\xEE\x3B"
-     "\x00\x80"
-     "\xF2\x3B\x00\x80\xF6\x3B\x00\x80\xFA\x3B\x00\x80\xE3\x64\x36\x00\xE3\x20"
-     "\x36\x00"
-     "\xE3\xE8\x36\x00\xE3\xD4\x36\x00\xE3\xC0\x31\x00\xE3\xA4\x31\x00\xE3\x8C"
-     "\x31\x00"
-     "\xE3\x68\x36\x00\xEF\x38\x31\x00\xEF\x14\x31\x00\xEF\x7C\x31\x00\xEF\x44"
-     "\x31\x00"
-     "\xEF\x48\x31\x00\xEF\x84\x36\x00\xEF\xCC\x36\x00\xEF\x34\x36\x00\xE3\x00"
-     "\x31\x00"
-     "\xE3\x6C\x31\x00\xE3\x54\x31\x00\xE3\xA8\x31\x00\xE3\xEC\x30\x00\xE3\xC8"
-     "\x30\x00"
-     "\xE3\xD0\x30\x00\xE3\x30\x31\x00\xEF\x60\x30\x00\xEF\x24\x30\x00\xEF\x1C"
-     "\x30\x00"
-     "\xEF\xE0\x31\x00\xEF\xDC\x31\x00\xEF\xC8\x31\x00\xEF\xA4\x31\x00\xEF\x78"
-     "\x31\x00"
-     "\x80\x5D\x0B\x00\x81\x65\x0B\x00\x82\x6D\x0B\x00\x83\x65\x0B\x00\x84\x7D"
-     "\x0B\x00"
-     "\x85\x65\x0B\x00\x86\x6D\x0B\x00\x87\x65\x0B\x00\x88\x5D\x0B\x00\x89\xA5"
-     "\x0B\x00"
-     "\x8A\xAD\x0B\x00\x8B\xA5\x0B\x00\x8C\xBD\x0B\x00\x8D\xA5\x0B\x00\x8E\xAD"
-     "\x0B\x00"
-     "\x8F\xA5\x0B\x00\x90\xDD\x0B\x00\x91\xE5\x0B\x00\x92\xED\x0B\x00\x93\xE5"
-     "\x0B\x00"
-     "\x94\xFD\x0B\x00\x95\xE5\x0B\x00\x96\xED\x0B\x00\x97\xE5\x0B\x00\x98\xDD"
-     "\x0B\x00"
-     "\x99\xA5\x0B\x00\x9A\xAD\x0B\x00\x9B\xA5\x0B\x00\x9C\xBD\x0B\x00\x9D\xA5"
-     "\x0B\x00"
-     "\x9E\xAD\x0B\x00\x9F\xA5\x0B\x00\xA0\x5D\x0B\x00\xA1\x65\x0B\x00\xA2\x6D"
-     "\x0B\x00"
-     "\xA3\x65\x0B\x00\xA4\x7D\x0B\x00\xA5\x65\x0B\x00\xA6\x6D\x0B\x00\xA7\x65"
-     "\x0B\x00"
-     "\xA8\x5D\x0B\x00\xA9\x85\x0C\x00\xAA\x8D\x0C\x00\xAB\x85\x0C\x00\xAC\x9D"
-     "\x0C\x00"
-     "\xAD\x85\x0C\x00\xAE\x8D\x0C\x00\xAF\x85\x0C\x00\xB0\xFD\x0C\x00\xB1\xA5"
-     "\x0C\x00"
-     "\xB2\xAD\x0C\x00\xB3\xA5\x0C\x00\xB4\xBD\x0C\x00\xB5\xA5\x0C\x00\xB6\xAD"
-     "\x0C\x00"
-     "\xB7\xA5\x0C\x00\xB8\x9D\x0C\x00\xB9\x45\x0C\x00\xBA\x4D\x0C\x00\xBB\x45"
-     "\x0C\x00"
-     "\xBC\x5D\x0C\x00\xBD\x45\x0C\x00\xBE\x4D\x0C\x00\xBF\x45\x0C\x00\x80\xBD"
-     "\x0D\x00"
-     "\x81\xC5\x0D\x00\x82\xCD\x0D\x00\x83\xC5\x0D\x00\x84\xDD\x0D\x00\x85\xC5"
-     "\x0D\x00"
-     "\x86\xCD\x0D\x00\x87\xC5\x0D\x00\x88\xFD\x0D\x00\x89\xC5\x0D\x00\x8A\xCD"
-     "\x0D\x00"
-     "\x8B\xC5\x0D\x00\x8C\xDD\x0D\x00\x8D\xC5\x0D\x00\x8E\xCD\x0D\x00\x8F\xC5"
-     "\x0D\x00"
-     "\x90\xBD\x0D\x00\x91\x45\x0C\x00\x92\x4D\x0C\x00\x93\x45\x0C\x00\x94\x5D"
-     "\x0C\x00"
-     "\x95\x45\x0C\x00\x96\x4D\x0C\x00\x97\x45\x0C\x00\x98\x7D\x0C\x00\x99\x45"
-     "\x0C\x00"
-     "\x9A\x4D\x0C\x00\x9B\x45\x0C\x00\x9C\x5D\x0C\x00\x9D\x45\x0C\x00\x9E\x4D"
-     "\x0C\x00"
-     "\x9F\x45\x0C\x00\xA0\xBD\x0C\x00\xA1\xC5\x0C\x00\xA2\xCD\x0C\x00\xA3\xC1"
-     "\x0C\x00"
-     "\xA4\xC1\x0C\x00\xA5\xC1\x0C\x00\xA6\xC1\x0C\x00\xA7\xC1\x0C\x00\xA8\xC1"
-     "\x0C\x00"
-     "\xA9\x19\x0D\x00\xAA\xCD\x0C\x00\xAB\xC5\x0C\x00\xAC\xDD\x0C\x00\xAD\xC5"
-     "\x0C\x00"
-     "\xAE\xCD\x0C\x00\xAF\xC5\x0C\x00\xB0\xBD\x0C\x00\xB1\x45\x0C\x00\xB2\x4D"
-     "\x0C\x00"
-     "\xB3\x45\x0C\x00\xB4\x5D\x0C\x00\xB5\x45\x0C\x00\xB6\x4D\x0C\x00\xB7\x45"
-     "\x0C\x00"
-     "\xB8\x7D\x0C\x00\xB9\x7D\x0C\x00\xBA\x49\x0C\x00\xBB\x49\x0C\x00\xBC\x59"
-     "\x0C\x00"
-     "\xBD\x59\x0C\x00\xBE\x49\x0C\x00\xBF\x51\x0C\x00\x9C\xC5\x08\x00\x9D\xCD"
-     "\x08\x00"
-     "\x9E\xC5\x08\x00\x9F\xDD\x08\x00\x98\xA5\x08\x00\x99\xAD\x08\x00\x9A\xA5"
-     "\x08\x00"
-     "\x9B\xDD\x08\x00\x94\x85\x08\x00\x95\x8D\x08\x00\x96\x85\x08\x00\x97\x9D"
-     "\x08\x00"
-     "\x90\xA5\x08\x00\x91\xAD\x08\x00\x92\xA5\x08\x00\x93\x9D\x08\x00\x8C\xC5"
-     "\x08\x00"
-     "\x8D\xCD\x08\x00\x8E\xC5\x08\x00\x8F\xDD\x08\x00\x88\x25\x08\x00\x89\x2D"
-     "\x08\x00"
-     "\x8A\x25\x08\x00\x8B\xDD\x08\x00\x84\x05\x08\x00\x85\x0D\x08\x00\x86\x05"
-     "\x08\x00"
-     "\x87\x1D\x08\x00\x80\x25\x08\x00\x81\x2D\x08\x00\x82\x25\x08\x00\x83\x1D"
-     "\x08\x00"
-     "\xBC\xCD\x09\x00\xBD\xD5\x09\x00\xBE\xDD\x09\x00\xBF\xD5\x09\x00\xB8\xAD"
-     "\x09\x00"
-     "\xB9\xD5\x09\x00\xBA\xDD\x09\x00\xBB\xD5\x09\x00\xB4\x8D\x09\x00\xB5\x95"
-     "\x09\x00"
-     "\xB6\x9D\x09\x00\xB7\x95\x09\x00\xB0\xAD\x09\x00\xB1\x95\x09\x00\xB2\x9D"
-     "\x09\x00"
-     "\xB3\x95\x09\x00\xAC\xC9\x09\x00\xAD\xC9\x09\x00\xAE\xDD\x09\x00\xAF\xD5"
-     "\x09\x00"
-     "\xA8\x25\x08\x00\xA9\x2D\x08\x00\xAA\x25\x08\x00\xAB\xD9\x09\x00\xA4\x05"
-     "\x08\x00"
-     "\xA5\x0D\x08\x00\xA6\x05\x08\x00\xA7\x1D\x08\x00\xA0\x25\x08\x00\xA1\x2D"
-     "\x08\x00"
-     "\xA2\x25\x08\x00\xA3\x1D\x08\x00\x80\xB5\x0E\x00\x81\xBD\x0E\x00\x82\xB5"
-     "\x0E\x00"
-     "\x83\xCD\x0E\x00\x84\xD5\x0E\x00\x85\xDD\x0E\x00\x86\xD5\x0E\x00\x87\xCD"
-     "\x0E\x00"
-     "\x88\xF5\x0E\x00\x89\xFD\x0E\x00\x8A\xF5\x0E\x00\x8B\xCD\x0E\x00\x8C\xD5"
-     "\x0E\x00"
-     "\x8D\xDD\x0E\x00\x8E\x0D\x0E\x00\x8F\xD1\x0E\x00\x90\xB1\x0E\x00\x91\xB1"
-     "\x0E\x00"
-     "\x92\xB1\x0E\x00\x93\xB1\x0E\x00\x94\x85\x0F\x00\x95\x55\x0D\x00\x96\x5D"
-     "\x0D\x00"
-     "\x97\x55\x0D\x00\x98\x6D\x0D\x00\x99\x75\x0D\x00\x9A\x7D\x0D\x00\x9B\x75"
-     "\x0D\x00"
-     "\x9C\x6D\x0D\x00\x9D\x55\x0D\x00\x9E\x5D\x0D\x00\x9F\x55\x0D\x00\xA0\xAD"
-     "\x0D\x00"
-     "\xA1\xB5\x0D\x00\xA2\xBD\x0D\x00\xA3\xB5\x0D\x00\xA4\xAD\x0D\x00\xA5\xD5"
-     "\x0D\x00"
-     "\xA6\xDD\x0D\x00\xA7\xD5\x0D\x00\xA8\xED\x0D\x00\xA9\xF5\x0D\x00\xAA\xFD"
-     "\x0D\x00"
-     "\xAB\xF5\x0D\x00\xAC\xED\x0D\x00\xAD\xD5\x0D\x00\xAE\xDD\x0D\x00\xAF\xD5"
-     "\x0D\x00"
-     "\xB0\xAD\x0D\x00\xB1\xB5\x0D\x00\xB2\xBD\x0D\x00\xB3\xB5\x0D\x00\xB4\xAD"
-     "\x0D\x00"
-     "\xB5\x55\x0D\x00\xB6\x5D\x0D\x00\xB7\x55\x0D\x00\xB8\x6D\x0D\x00\xB9\x75"
-     "\x0D\x00"
-     "\xBA\x7D\x0D\x00\xBB\x75\x0D\x00\xBC\x6D\x0D\x00\xBD\x55\x0D\x00\xBE\x5D"
-     "\x0D\x00"
-     "\xBF\x55\x0D\x00\x80\xAD\x0A\x00\xFE\x3B\x00\x80\x82\xB9\x0A\x00\xB2\xE5"
-     "\x00\x00"
-     "\x84\xA5\x0A\x00\xB4\xED\x00\x00\x86\xA1\x0A\x00\xB6\xE1\x00\x00\x88\x9D"
-     "\x0A\x00"
-     "\xB8\x25\x01\x00\x8A\xED\x5D\x00\xBA\x29\x01\x00\x8C\xF1\x0A\x00\xBC\x3D"
-     "\x01\x00"
-     "\x8E\xF1\x0A\x00\xBE\x31\x01\x00\x90\x8D\x0A\x00\x02\x3C\x00\x80\x92\x99"
-     "\x0A\x00"
-     "\xA2\x09\x00\x00\x94\x85\x0A\x00\xA4\xED\x00\x00\x96\x81\x0A\x00\xA6\x9D"
-     "\x00\x00"
-     "\x98\xBD\x0A\x00\xA8\xAD\x00\x00\x9A\x89\x0A\x00\xAA\xBD\x00\x00\x9C\x95"
-     "\x0A\x00"
-     "\xAC\xAD\x00\x00\x9E\x91\x0A\x00\xAE\x9D\x00\x00\xA0\x6D\x0A\x00\x80\xC9"
-     "\x0B\x00"
-     "\xA2\x79\x0A\x00\x82\xDD\x0B\x00\xA4\x69\x0A\x00\x84\xC1\x0B\x00\xA6\x65"
-     "\x0A\x00"
-     "\x86\xC5\x0B\x00\xA8\xE1\x0B\x00\x88\xF9\x0B\x00\xAA\xED\x0B\x00\x8A\xCD"
-     "\x0B\x00"
-     "\xAC\xF9\x0B\x00\x8C\xD1\x0B\x00\x06\x3C\x00\x80\x8E\xD5\x0B\x00\x0A\x3C"
-     "\x00\x80"
-     "\x90\xA9\x0B\x00\x0E\x3C\x00\x80\x92\xBD\x0B\x00\x12\x3C\x00\x80\x94\xA1"
-     "\x0B\x00"
-     "\x16\x3C\x00\x80\x96\xA5\x0B\x00\x1A\x3C\x00\x80\x98\x99\x0B\x00\x1E\x3C"
-     "\x00\x80"
-     "\x9A\xE9\x00\x00\x9D\x6D\x72\x02\x9C\x79\x72\x02\x22\x3C\x00\x80\x26\x3C"
-     "\x00\x80"
-     "\x2A\x3C\x00\x80\x2E\x3C\x00\x80\x32\x3C\x00\x80\x36\x3C\x00\x80\x3A\x3C"
-     "\x00\x80"
-     "\x3E\x3C\x00\x80\x42\x3C\x00\x80\x46\x3C\x00\x80\x4A\x3C\x00\x80\x4E\x3C"
-     "\x00\x80"
-     "\x52\x3C\x00\x80\x56\x3C\x00\x80\x5A\x3C\x00\x80\x5E\x3C\x00\x80\x62\x3C"
-     "\x00\x80"
-     "\x66\x3C\x00\x80\x6A\x3C\x00\x80\x9A\xC0\x03\x00\x99\x48\x03\x00\x6E\x3C"
-     "\x00\x80"
-     "\x9F\x00\x1C\x00\x9E\x10\x00\x00\x9D\x58\x01\x00\x9C\x98\x03\x00\x72\x3C"
-     "\x00\x80"
-     "\x76\x3C\x00\x80\x7A\x3C\x00\x80\x7E\x3C\x00\x80\x82\x3C\x00\x80\x86\x3C"
-     "\x00\x80"
-     "\x8A\x3C\x00\x80\x8E\x3C\x00\x80\x92\x3C\x00\x80\x96\x3C\x00\x80\x9A\x3C"
-     "\x00\x80"
-     "\x9E\x3C\x00\x80\xAA\x1C\x00\x80\xAE\x94\x1E\x00\xAD\xE8\x1F\x00\xA2\x3C"
-     "\x00\x80"
-     "\xA6\x3C\x00\x80\xAA\x3C\x00\x80\xAE\x3C\x00\x80\xB2\x3C\x00\x80\xB6\x3C"
-     "\x00\x80"
-     "\xBE\x3C\x00\x80\xC2\x3C\x00\x80\xCA\x3C\x00\x80\xCE\x3C\x00\x80\xD2\x3C"
-     "\x00\x80"
-     "\xD6\x3C\x00\x80\xDA\x3C\x00\x80\xDE\x3C\x00\x80\xE2\x3C\x00\x80\xE6\x3C"
-     "\x00\x80"
-     "\xEA\x3C\x00\x80\xEE\x3C\x00\x80\x10\x0B\x00\x80\x8B\x0B\x00\x80\x55\x0B"
-     "\x00\x80"
-     "\x32\x3D\x00\x80\xF2\x3C\x00\x80\xF6\x3C\x00\x80\xFA\x3C\x00\x80\x80\x11"
-     "\x03\x00"
-     "\xFE\x3C\x00\x80\x82\x1D\x03\x00\x02\x3D\x00\x80\x84\x09\x03\x00\x06\x3D"
-     "\x00\x80"
-     "\x86\x05\x03\x00\x0A\x3D\x00\x80\x88\x01\x03\x00\x0E\x3D\x00\x80\x8A\x0D"
-     "\x03\x00"
-     "\xC6\x3C\x00\x80\x8C\x19\x03\x00\xBA\x1C\x00\x80\x8E\x15\x03\x00\x12\x3D"
-     "\x00\x80"
-     "\x90\x71\x03\x00\x16\x3D\x00\x80\x92\x7D\x03\x00\x1A\x3D\x00\x80\x94\x69"
-     "\x03\x00"
-     "\x1E\x3D\x00\x80\x96\x65\x03\x00\x22\x3D\x00\x80\x98\x21\x03\x00\x26\x3D"
-     "\x00\x80"
-     "\x9A\x25\x03\x00\x2A\x3D\x00\x80\x9C\x11\x03\x00\x2E\x3D\x00\x80\x9E\x1D"
-     "\x03\x00"
-     "\x67\x0B\x00\x80\xA0\xE9\x03\x00\x82\x6C\x1F\x00\xA2\xE5\x03\x00\x0E\x1B"
-     "\x00\x80"
-     "\xA4\xE1\x03\x00\x12\x1B\x00\x80\xA6\xED\x03\x00\x16\x1B\x00\x80\xA8\xD9"
-     "\x03\x00"
-     "\x1A\x1B\x00\x80\xAA\xD5\x03\x00\x1E\x1B\x00\x80\xAC\x31\x03\x00\x22\x1B"
-     "\x00\x80"
-     "\xAE\x3D\x03\x00\x26\x1B\x00\x80\xB0\x45\x03\x00\x91\x48\x02\x00\x2A\x1B"
-     "\x00\x80"
-     "\x2E\x1B\x00\x80\x32\x1B\x00\x80\x36\x1B\x00\x80\x3A\x1B\x00\x80\x3E\x1B"
-     "\x00\x80"
-     "\x42\x1B\x00\x80\xB9\x6D\x03\x00\x46\x1B\x00\x80\xBB\x61\x03\x00\x4A\x1B"
-     "\x00\x80"
-     "\xBD\x65\x03\x00\xBE\x65\x03\x00\x4E\x1B\x00\x80\x52\x1B\x00\x80\x90\xB1"
-     "\x02\x00"
-     "\x56\x1B\x00\x80\x92\xB5\x02\x00\x5A\x1B\x00\x80\x5E\x1B\x00\x80\x62\x1B"
-     "\x00\x80"
-     "\x96\xA1\x02\x00\x66\x1B\x00\x80\x98\xE5\x02\x00\x6A\x1B\x00\x80\x9A\xE9"
-     "\x02\x00"
-     "\x6E\x1B\x00\x80\x9C\xFD\x02\x00\x72\x1B\x00\x80\x9E\xF1\x02\x00\x76\x1B"
-     "\x00\x80"
-     "\x80\xE5\x02\x00\x7A\x1B\x00\x80\x82\xE9\x02\x00\x7E\x1B\x00\x80\x84\xFD"
-     "\x02\x00"
-     "\x82\x1B\x00\x80\x86\xF1\x02\x00\x86\x1B\x00\x80\x8A\x1B\x00\x80\x8B\xD9"
-     "\x02\x00"
-     "\x8E\x1B\x00\x80\x8D\x25\x5C\x00\x92\x1B\x00\x80\x96\x1B\x00\x80\x9A\x1B"
-     "\x00\x80"
-     "\xB1\x6D\x02\x00\xB0\x6D\x02\x00\xB3\x79\x02\x00\xB2\xE9\x5C\x00\x9E\x1B"
-     "\x00\x80"
-     "\xB4\x61\x02\x00\xA2\x1B\x00\x80\xB6\x65\x02\x00\xA6\x1B\x00\x80\xB8\x59"
-     "\x02\x00"
-     "\xAA\x1B\x00\x80\xBA\xA9\x01\x00\xAE\x1B\x00\x80\xBC\xB9\x01\x00\xB2\x1B"
-     "\x00\x80"
-     "\xBE\xA9\x01\x00\xB6\x1B\x00\x80\xA0\x15\x02\x00\xBA\x1B\x00\x80\xA2\x19"
-     "\x02\x00"
-     "\xBE\x1B\x00\x80\xA4\x0D\x02\x00\xC2\x1B\x00\x80\xA6\x01\x02\x00\xC6\x1B"
-     "\x00\x80"
-     "\xA8\x05\x02\x00\xAB\x15\x5D\x00\xAA\x49\x7F\x02\xAD\x15\x02\x00\xAC\xBD"
-     "\x5C\x00"
-     "\xCA\x1B\x00\x80\xAE\xA1\x5C\x00\x9C\xC1\x1C\x00\x9D\x69\x1E\x00\x9E\xE1"
-     "\x35\x00"
-     "\x9F\x79\x1E\x00\xCE\x1B\x00\x80\xD2\x1B\x00\x80\xD6\x1B\x00\x80\xDA\x1B"
-     "\x00\x80"
-     "\xDE\x1B\x00\x80\xE2\x1B\x00\x80\xE6\x1B\x00\x80\xEA\x1B\x00\x80\xEE\x1B"
-     "\x00\x80"
-     "\xF2\x1B\x00\x80\xF6\x1B\x00\x80\xFA\x1B\x00\x80\xFE\x1B\x00\x80\x02\x1C"
-     "\x00\x80"
-     "\x06\x1C\x00\x80\x0A\x1C\x00\x80\x0E\x1C\x00\x80\x12\x1C\x00\x80\x16\x1C"
-     "\x00\x80"
-     "\x1A\x1C\x00\x80\x1E\x1C\x00\x80\x22\x1C\x00\x80\x26\x1C\x00\x80\x2A\x1C"
-     "\x00\x80"
-     "\x2E\x1C\x00\x80\x32\x1C\x00\x80\x36\x1C\x00\x80\x3A\x1C\x00\x80\xBC\x5D"
-     "\x1E\x00"
-     "\xBD\x65\x1E\x00\xBE\x61\x1E\x00\xBF\x7D\x1E\x00\xB8\x5D\x1E\x00\xB9\x45"
-     "\x1E\x00"
-     "\xBA\x4D\x1E\x00\xBB\x45\x1E\x00\xB4\x0D\x1E\x00\xB5\x11\x1E\x00\xB6\x15"
-     "\x1E\x00"
-     "\xB7\x65\x1E\x00\xB0\x3D\x1E\x00\xB1\x21\x1E\x00\xB2\x25\x1E\x00\xB3\x19"
-     "\x1E\x00"
-     "\x3E\x1C\x00\x80\x42\x1C\x00\x80\x46\x1C\x00\x80\x4A\x1C\x00\x80\xAE\x38"
-     "\x00\x80"
-     "\xA9\xA9\x1E\x00\x25\x32\x00\x80\xD6\x38\x00\x80\x8E\x33\x00\x80\xF2\x2B"
-     "\x00\x80"
-     "\xED\x2D\x00\x80\x20\x2F\x00\x80\x51\x2E\x00\x80\x1C\x2F\x00\x80\x0B\x31"
-     "\x00\x80"
-     "\x06\x35\x00\x80\x80\xFD\x1F\x00\x81\x05\x1E\x00\x82\x01\x1E\x00\x83\x1D"
-     "\x1E\x00"
-     "\x84\x05\x1E\x00\x85\x0D\x1E\x00\x86\x39\x1E\x00\x87\x35\x1E\x00\x88\x11"
-     "\x1E\x00"
-     "\x89\x1D\x1E\x00\x8A\x69\x1E\x00\x8B\x65\x1E\x00\x8C\x61\x1E\x00\x8D\x6D"
-     "\x1E\x00"
-     "\x8E\x59\x1E\x00\x8F\x59\x1E\x00\x90\x2D\x1E\x00\x91\x31\x1E\x00\x92\x31"
-     "\x1E\x00"
-     "\x93\x31\x1E\x00\x94\xC1\x1E\x00\x95\xCD\x1E\x00\x96\xF9\x1E\x00\x97\xF5"
-     "\x1E\x00"
-     "\x98\xD1\x1E\x00\x99\xDD\x1E\x00\x9A\xA9\x1E\x00\x9B\xA5\x1E\x00\x9C\xA1"
-     "\x1E\x00"
-     "\x9D\xAD\x1E\x00\x9E\x99\x1E\x00\x9F\x85\x1E\x00\xA0\x91\x01\x00\xA1\x91"
-     "\x01\x00"
-     "\xA2\x91\x01\x00\xA3\x91\x01\x00\xA4\xB1\x01\x00\xA5\xB1\x01\x00\xA6\xB1"
-     "\x01\x00"
-     "\xA7\xB1\x01\x00\xA8\x91\x01\x00\xA9\x91\x01\x00\xAA\x91\x01\x00\xAB\x91"
-     "\x01\x00"
-     "\xAC\xF1\x01\x00\xAD\xF1\x01\x00\xAE\xF1\x01\x00\xAF\xF1\x01\x00\xB0\x91"
-     "\x01\x00"
-     "\xB1\x91\x01\x00\xB2\x91\x01\x00\xB3\x91\x01\x00\xB4\xB1\x01\x00\xB5\xB1"
-     "\x01\x00"
-     "\xB6\xB1\x01\x00\xB7\xB1\x01\x00\xB8\x91\x01\x00\xB9\x91\x01\x00\xBA\x91"
-     "\x01\x00"
-     "\xBB\x91\x01\x00\xBC\x31\x01\x00\xBD\x31\x01\x00\xBE\x31\x01\x00\xBF\x31"
-     "\x01\x00"
-     "\xA4\x00\x04\x00\xA5\x04\x05\x00\xA6\x08\x07\x00\xA7\x0C\x09\x00\x96\x36"
-     "\x00\x80"
-     "\xCE\x37\x00\x80\x0A\x38\x00\x80\x3E\x38\x00\x80\xAC\x20\xF1\x00\xAD\x24"
-     "\xF2\x00"
-     "\xAE\x28\xF3\x00\xAF\x60\xF5\x00\xA8\x30\x0B\x00\xA9\x00\x0C\x00\xAA\x38"
-     "\x0E\x00"
-     "\xAB\x00\xF0\x00\xB4\x00\xFC\x00\xB5\x04\xFD\x00\xB6\x48\xFE\x00\xB7\x00"
-     "\xE0\x00"
-     "\xB0\x58\xF6\x00\xB1\x00\xF8\x00\xB2\x58\xF9\x00\xB3\x5C\xFA\x00\xBC\x64"
-     "\xE6\x00"
-     "\xBD\x64\xE7\x00\xBE\x00\xE8\x00\xBF\x64\xE9\x00\xB8\x30\xE1\x00\xB9\x78"
-     "\xE2\x00"
-     "\xBA\x00\xE4\x00\xBB\x7C\xE5\x00\x52\x39\x00\x80\xAA\x32\x00\x80\x36\x33"
-     "\x00\x80"
-     "\xE2\x33\x00\x80\x7A\x34\x00\x80\x96\x37\x00\x80\x7E\x39\x00\x80\xAA\x3A"
-     "\x00\x80"
-     "\x16\x2C\x00\x80\xC1\x2D\x00\x80\xBE\x32\x00\x80\x06\x34\x00\x80\xCA\x37"
-     "\x00\x80"
-     "\xFA\x3A\x00\x80\xC8\x2F\x00\x80\xAE\x33\x00\x80\xBA\x37\x00\x80\x4E\x38"
-     "\x00\x80"
-     "\x3F\x31\x00\x80\xF2\x36\x00\x80\x16\x38\x00\x80\x3B\x30\x00\x80\x39\x32"
-     "\x00\x80"
-     "\x4E\x33\x00\x80\x52\x34\x00\x80\x5E\x39\x00\x80\x5D\x2C\x00\x80\xF5\x2C"
-     "\x00\x80"
-     "\x61\x2D\x00\x80\xA7\x31\x00\x80\xB6\x32\x00\x80\x02\x34\x00\x80\xBA\xB9"
-     "\xEE\x00"
-     "\x26\x35\x00\x80\x96\x5D\x1D\x00\xC2\x37\x00\x80\x90\x5D\x1D\x00\xDA\x37"
-     "\x00\x80"
-     "\xFE\x38\x00\x80\x26\x3A\x00\x80\xDA\x3A\x00\x80\xF6\x3A\x00\x80\x62\x35"
-     "\x00\x80"
-     "\xA2\x35\x00\x80\x98\x6D\x1D\x00\x5A\x36\x00\x80\x7A\x37\x00\x80\xB2\x39"
-     "\x00\x80"
-     "\x84\x6D\x1D\x00\x85\xD9\x40\x00\x86\x7D\x1D\x00\x87\x71\x1D\x00\x80\x4D"
-     "\x1D\x00"
-     "\x82\x38\x00\x80\x82\x7D\x1D\x00\xD5\x2E\x00\x80\x57\x30\x00\x80\x12\x36"
-     "\x00\x80"
-     "\xD2\x36\x00\x80\x36\x34\x00\x80\x6A\x35\x00\x80\x89\x31\x1D\x00\xCE\x38"
-     "\x00\x80"
-     "\x1E\x3A\x00\x80\xB4\x45\x74\x02\xB5\xE1\x1D\x00\xD1\x2E\x00\x80\xA4\x2F"
-     "\x00\x80"
-     "\xAE\x32\x00\x80\x62\x33\x00\x80\xB2\x91\x77\x02\xB3\xCD\x77\x02\x96\x33"
-     "\x00\x80"
-     "\x4A\x36\x00\x80\x72\x36\x00\x80\xDE\x39\x00\x80\xB8\x79\x7E\x02\xB9\xA1"
-     "\x7D\x02"
-     "\x5D\x2D\x00\x80\xDE\x36\x00\x80\x01\x2D\x00\x80\xFD\x2C\x00\x80\xCE\x35"
-     "\x00\x80"
-     "\x5E\x36\x00\x80\x92\x37\x00\x80\xE6\x39\x00\x80\x9E\x38\x00\x80\x43\x31"
-     "\x00\x80"
-     "\x8E\x38\x00\x80\xFE\x2B\x00\x80\x78\x2F\x00\x80\xAF\x30\x00\x80\x66\x34"
-     "\x00\x80"
-     "\xE6\x34\x00\x80\x81\x2D\x00\x80\x72\x35\x00\x80\x80\xD1\x04\x00\x81\xDD"
-     "\x04\x00"
-     "\x82\xC5\x38\x00\x83\xD1\x04\x00\x84\xF1\x04\x00\x85\xF1\x04\x00\x86\xF1"
-     "\x04\x00"
-     "\x87\x1D\x24\x00\x88\x21\x24\x00\x89\xC9\x04\x00\x8A\xC5\x38\x00\x8B\xDD"
-     "\x04\x00"
-     "\x8C\xC5\x04\x00\x8D\xCD\x04\x00\x8E\xC5\x04\x00\x8F\x2D\x06\x00\x90\x55"
-     "\x06\x00"
-     "\x91\x5D\x06\x00\x92\x55\x06\x00\x93\xCD\x06\x00\x94\xD5\x06\x00\x95\xDD"
-     "\x06\x00"
-     "\x96\xD5\x06\x00\x97\xCD\x06\x00\x98\xF5\x06\x00\x99\xFD\x06\x00\x9A\xF5"
-     "\x06\x00"
-     "\x9B\xCD\x06\x00\x9C\xD5\x06\x00\x9D\xDD\x06\x00\x9E\xD5\x06\x00\x9F\xCD"
-     "\x06\x00"
-     "\xA0\x35\x06\x00\xA1\x3D\x06\x00\xA2\x35\x06\x00\xA3\x4D\x06\x00\xA4\x55"
-     "\x06\x00"
-     "\xA5\x5D\x06\x00\xA6\x55\x06\x00\xA7\x4D\x06\x00\xA8\x75\x06\x00\xA9\x7D"
-     "\x06\x00"
-     "\xAA\x75\x06\x00\xAB\x4D\x06\x00\xAC\x55\x06\x00\xAD\x5D\x06\x00\xAE\x55"
-     "\x06\x00"
-     "\xAF\x4D\x06\x00\xB0\x35\x06\x00\xB1\x3D\x06\x00\xB2\x35\x06\x00\xB3\xC9"
-     "\x07\x00"
-     "\xB4\x95\x38\x00\xB5\xD9\x07\x00\xB6\xCD\x07\x00\xB7\xC5\x07\x00\xB8\xFD"
-     "\x07\x00"
-     "\xB9\xC5\x07\x00\xBA\xCD\x07\x00\xBB\xC5\x07\x00\xBC\xDD\x07\x00\xBD\xC1"
-     "\x07\x00"
-     "\xBE\xC1\x07\x00\xBF\xC1\x07\x00\x80\x71\x3A\x00\x81\x51\x06\x00\x82\x55"
-     "\x06\x00"
-     "\x83\x6D\x06\x00\x84\x75\x06\x00\x85\x7D\x06\x00\x86\x75\x06\x00\x87\x6D"
-     "\x06\x00"
-     "\x88\x59\x06\x00\x89\x59\x06\x00\x8A\xA1\x06\x00\x8B\xA1\x06\x00\x8C\xA1"
-     "\x06\x00"
-     "\x8D\xA1\x06\x00\x8E\x99\x06\x00\x8F\x99\x06\x00\x90\xE9\x06\x00\x91\xE9"
-     "\x06\x00"
-     "\x92\xF1\x06\x00\x93\xF1\x06\x00\x94\x91\x06\x00\x95\x91\x06\x00\x96\x91"
-     "\x06\x00"
-     "\x97\x91\x06\x00\x98\xB1\x06\x00\x99\xB1\x06\x00\x9A\xB1\x06\x00\x9B\xB1"
-     "\x06\x00"
-     "\x9C\xF5\x07\x00\x9D\x95\x06\x00\x9E\x9D\x06\x00\x9F\x95\x06\x00\xA0\x6D"
-     "\x06\x00"
-     "\xA1\x75\x06\x00\xA2\x7D\x06\x00\xA3\x75\x06\x00\xA4\x6D\x06\x00\xA5\x91"
-     "\x03\x00"
-     "\xA6\x91\x03\x00\xA7\x91\x03\x00\xA8\xB1\x03\x00\xA9\xB1\x03\x00\xAA\xB1"
-     "\x03\x00"
-     "\xAB\xB1\x03\x00\xAC\x91\x03\x00\xAD\x91\x03\x00\xAE\x91\x03\x00\xAF\x91"
-     "\x03\x00"
-     "\xB0\xF1\x03\x00\xB1\x11\x39\x00\xB2\xFD\x03\x00\xB3\xF5\x03\x00\xB4\xED"
-     "\x03\x00"
-     "\xB5\x95\x03\x00\xB6\x9D\x03\x00\xB7\x95\x03\x00\xB8\xAD\x03\x00\xB9\xB5"
-     "\x03\x00"
-     "\xBA\xBD\x03\x00\xBB\xB5\x03\x00\xBC\xAD\x03\x00\xBD\x95\x03\x00\xBE\x9D"
-     "\x03\x00"
-     "\xBF\x95\x03\x00\xF0\x80\x1A\x00\x6D\x2C\x00\x80\x93\x30\x00\x80\xDE\x2B"
-     "\x00\x80"
-     "\x22\x33\x00\x80\xBF\x31\x00\x80\x46\x36\x00\x80\xD5\x2D\x00\x80\xC1\x2E"
-     "\x00\x80"
-     "\x32\x35\x00\x80\x9A\x37\x00\x80\x72\x38\x00\x80\xEE\x32\x00\x80\x1A\x33"
-     "\x00\x80"
-     "\x53\x31\x00\x80\x52\x37\x00\x80\x73\x31\x00\x80\xE2\x34\x00\x80\x32\x2C"
-     "\x00\x80"
-     "\xB5\x2C\x00\x80\xF9\x2C\x00\x80\x8A\x32\x00\x80\x32\x36\x00\x80\x2A\x37"
-     "\x00\x80"
-     "\x7E\x38\x00\x80\x92\x39\x00\x80\x69\x2D\x00\x80\x1D\x2E\x00\x80\x4B\x30"
-     "\x00\x80"
-     "\xDB\x31\x00\x80\xB2\x33\x00\x80\x76\x35\x00\x80\xCE\x39\x00\x80\xB2\x3A"
-     "\x00\x80"
-     "\x06\x3B\x00\x80\x1E\x3B\x00\x80\x1D\x32\x00\x80\xDA\x32\x00\x80\x22\x39"
-     "\x00\x80"
-     "\x17\x30\x00\x80\xF3\x30\x00\x80\x0F\x31\x00\x80\x9F\x31\x00\x80\xA2\x33"
-     "\x00\x80"
-     "\xEE\x33\x00\x80\x9E\x34\x00\x80\xC2\x35\x00\x80\x66\x36\x00\x80\xCA\x36"
-     "\x00\x80"
-     "\x12\x39\x00\x80\xA2\x37\x00\x80\x3A\x39\x00\x80\xB6\x39\x00\x80\x1D\x2D"
-     "\x00\x80"
-     "\x49\x2D\x00\x80\x31\x2E\x00\x80\xDE\x33\x00\x80\x2A\x38\x00\x80\x37\x30"
-     "\x00\x80"
-     "\xA7\x30\x00\x80\x6B\x31\x00\x80\xE6\x32\x00\x80\x0E\x36\x00\x80\x5A\x34"
-     "\x00\x80"
-     "\x80\x6D\x01\x00\x81\xC9\x3D\x00\x82\x79\x01\x00\x83\x79\x01\x00\x84\x69"
-     "\x01\x00"
-     "\x85\x69\x01\x00\x86\x99\x01\x00\x87\x99\x01\x00\x88\xA9\x01\x00\x89\xA9"
-     "\x01\x00"
-     "\x8A\x05\x3D\x00\x8B\xBD\x01\x00\x8C\xA5\x01\x00\x8D\xAD\x01\x00\x8E\xA5"
-     "\x01\x00"
-     "\x8F\x9D\x01\x00\x90\xE5\x01\x00\x91\xED\x01\x00\x92\xE5\x01\x00\x93\xFD"
-     "\x01\x00"
-     "\x94\xE5\x01\x00\x95\xED\x01\x00\x96\xE5\x01\x00\x97\x9D\x01\x00\x98\xA5"
-     "\x01\x00"
-     "\x99\xAD\x01\x00\x9A\xA5\x01\x00\x9B\xBD\x01\x00\x9C\xA5\x01\x00\x9D\xAD"
-     "\x01\x00"
-     "\x9E\xA5\x01\x00\x9F\x9D\x01\x00\xA0\x65\x01\x00\xA1\xA9\x01\x00\xA2\x61"
-     "\x01\x00"
-     "\xA3\x61\x01\x00\xA4\x61\x01\x00\xA5\x61\x01\x00\xA6\x61\x01\x00\xA7\x61"
-     "\x01\x00"
-     "\xA8\x01\x0C\x00\xA9\xA5\x0F\x00\xAA\x41\x04\x00\xAB\xA9\x0F\x00\xAC\xB9"
-     "\x0F\x00"
-     "\xAD\xB9\x0F\x00\xAE\xA9\x0F\x00\xAF\xA9\x0F\x00\xB0\xD9\x0F\x00\xB1\xD9"
-     "\x0F\x00"
-     "\xB2\xE9\x0F\x00\xB3\xE9\x0F\x00\xB4\xF9\x0F\x00\xB5\xF9\x0F\x00\xB6\xE9"
-     "\x0F\x00"
-     "\xB7\xE9\x0F\x00\xB8\xD9\x0F\x00\xB9\xD9\x0F\x00\xBA\xA9\x0F\x00\xBB\xA9"
-     "\x0F\x00"
-     "\xBC\xB9\x0F\x00\xBD\xB9\x0F\x00\xBE\xA9\x0F\x00\xBF\x79\x05\x00\x8D\x2E"
-     "\x00\x80"
-     "\xDC\x2F\x00\x80\xB3\x30\x00\x80\x76\x34\x00\x80\x92\x34\x00\x80\xA2\x36"
-     "\x00\x80"
-     "\xC2\x36\x00\x80\x9A\x39\x00\x80\x1A\x3A\x00\x80\x2A\x3A\x00\x80\x66\x3A"
-     "\x00\x80"
-     "\x61\x2C\x00\x80\xB6\x35\x00\x80\x86\x39\x00\x80\x02\x3A\x00\x80\xCB\x30"
-     "\x00\x80"
-     "\x1E\x2C\x00\x80\x85\x2C\x00\x80\x7C\x2F\x00\x80\x9C\x2F\x00\x80\xCB\x31"
-     "\x00\x80"
-     "\xFA\x33\x00\x80\x02\x35\x00\x80\xAE\x37\x00\x80\x4E\x39\x00\x80\x0D\x32"
-     "\x00\x80"
-     "\xDA\x39\x00\x80\x41\x2D\x00\x80\x59\x32\x00\x80\x65\x32\x00\x80\x46\x33"
-     "\x00\x80"
-     "\x7E\x33\x00\x80\x82\x34\x00\x80\xDE\x34\x00\x80\x5E\x35\x00\x80\x42\x36"
-     "\x00\x80"
-     "\x72\x3A\x00\x80\x1F\x31\x00\x80\xEA\x39\x00\x80\x75\x2C\x00\x80\xC0\x2F"
-     "\x00\x80"
-     "\x66\x33\x00\x80\x16\x39\x00\x80\x83\x30\x00\x80\xD3\x30\x00\x80\x6D\x32"
-     "\x00\x80"
-     "\xF2\x39\x00\x80\x25\x2D\x00\x80\x0D\x2E\x00\x80\xAC\x2F\x00\x80\xF6\x31"
-     "\x00\x80"
-     "\x49\x32\x00\x80\x96\x32\x00\x80\x2A\x33\x00\x80\x7E\x34\x00\x80\xF2\x34"
-     "\x00\x80"
-     "\x8E\x36\x00\x80\x2E\x38\x00\x80\x3A\x38\x00\x80\x0E\x3A\x00\x80\x91\x2D"
-     "\x00\x80"
-     "\x82\x33\x00\x80\x11\x2E\x00\x80\xFE\x33\x00\x80\x80\x55\x01\x00\x81\x5D"
-     "\x01\x00"
-     "\x82\x55\x01\x00\x83\x6D\x01\x00\x84\xDD\x2A\x00\x85\x71\x01\x00\x86\x71"
-     "\x01\x00"
-     "\x87\x71\x01\x00\x88\x51\x01\x00\x89\x51\x01\x00\x8A\x51\x01\x00\x8B\x51"
-     "\x01\x00"
-     "\x8C\xB1\x01\x00\x8D\xB1\x01\x00\x8E\xB1\x01\x00\x8F\xB1\x01\x00\x90\xD1"
-     "\x01\x00"
-     "\x91\x45\x19\x00\x92\xDD\x01\x00\x93\xD5\x01\x00\x94\xCD\x01\x00\x95\xF5"
-     "\x01\x00"
-     "\x96\xFD\x01\x00\x97\xF5\x01\x00\x98\xCD\x01\x00\x99\xD5\x01\x00\x9A\xDD"
-     "\x01\x00"
-     "\x9B\x9D\x0A\x00\x9C\xC9\x01\x00\x9D\xC9\x01\x00\x9E\xB9\x01\x00\x9F\xB9"
-     "\x01\x00"
-     "\xA0\x49\x01\x00\xA1\x49\x01\x00\xA2\x59\x01\x00\xA3\x59\x01\x00\xA4\x49"
-     "\x01\x00"
-     "\xA5\x49\x01\x00\xA6\x79\x01\x00\xA7\x79\x01\x00\xA8\x49\x01\x00\xA9\x4D"
-     "\x35\x00"
-     "\xAA\x45\x01\x00\xAB\x5D\x01\x00\xAC\x45\x01\x00\xAD\x4D\x01\x00\xAE\x45"
-     "\x01\x00"
-     "\xAF\xBD\x03\x00\xB0\xC5\x03\x00\xB1\xCD\x03\x00\xB2\xC5\x03\x00\xB3\xDD"
-     "\x03\x00"
-     "\xB4\xC5\x03\x00\xB5\xCD\x03\x00\xB6\xC5\x03\x00\xB7\xF9\x35\x00\xB8\xC1"
-     "\x03\x00"
-     "\xB9\xC1\x03\x00\xBA\xC1\x03\x00\xBB\xC1\x03\x00\xBC\xC1\x03\x00\xBD\xC1"
-     "\x03\x00"
-     "\xBE\xC1\x03\x00\xBF\xC1\x03\x00\xA2\x34\x00\x80\xBE\x37\x00\x80\xF6\x39"
-     "\x00\x80"
-     "\xE2\x3A\x00\x80\x0A\x3B\x00\x80\x61\x32\x00\x80\x5E\x33\x00\x80\xFA\x36"
-     "\x00\x80"
-     "\xFA\x35\x00\x80\x1E\x36\x00\x80\x4E\x34\x00\x80\xD6\x33\x00\x80\x9A\x38"
-     "\x00\x80"
-     "\x3A\x2C\x00\x80\x62\x37\x00\x80\x29\x2D\x00\x80\x19\x2D\x00\x80\x27\x30"
-     "\x00\x80"
-     "\x47\x31\x00\x80\x2A\x36\x00\x80\x68\x2F\x00\x80\x3A\x33\x00\x80\x19\x32"
-     "\x00\x80"
-     "\x1E\x39\x00\x80\xE2\x39\x00\x80\x43\x30\x00\x80\x95\x2C\x00\x80\x69\x2E"
-     "\x00\x80"
-     "\xBD\x2E\x00\x80\x09\x32\x00\x80\x05\x2D\x00\x80\x56\x34\x00\x80\x1A\x35"
-     "\x00\x80"
-     "\x7E\x35\x00\x80\x9A\x35\x00\x80\x9E\x35\x00\x80\xB2\x35\x00\x80\x32\x3A"
-     "\x00\x80"
-     "\x22\x36\x00\x80\xC6\x37\x00\x80\x8A\x38\x00\x80\x3E\x39\x00\x80\x6A\x39"
-     "\x00\x80"
-     "\x86\x3A\x00\x80\x8A\x3A\x00\x80\x8E\x3A\x00\x80\xF2\x3A\x00\x80\x62\x39"
-     "\x00\x80"
-     "\xFE\x39\x00\x80\x65\x2C\x00\x80\x89\x2C\x00\x80\x9D\x2C\x00\x80\x59\x2D"
-     "\x00\x80"
-     "\x65\x2D\x00\x80\xA5\x2D\x00\x80\x59\x2E\x00\x80\x75\x2E\x00\x80\x7D\x2E"
-     "\x00\x80"
-     "\xB9\x2E\x00\x80\xC5\x2E\x00\x80\xA8\x2F\x00\x80\xBF\x30\x00\x80\xE7\x30"
-     "\x00\x80"
-     "\xEF\x30\x00\x80\x80\x41\x01\x00\x81\x41\x01\x00\x82\x41\x01\x00\x83\x41"
-     "\x01\x00"
-     "\x84\x41\x01\x00\x85\x41\x01\x00\x86\x41\x01\x00\x87\x41\x01\x00\x88\xB1"
-     "\x37\x00"
-     "\x89\x45\x01\x00\x8A\xB5\x37\x00\x8B\x49\x01\x00\x8C\x59\x01\x00\x8D\x59"
-     "\x01\x00"
-     "\x07\x31\x00\x80\xB3\x31\x00\x80\x90\x31\x01\x00\xE3\x31\x00\x80\x92\x3D"
-     "\x01\x00"
-     "\x11\x32\x00\x80\x8E\x32\x00\x80\x95\x2D\x01\x00\x96\x25\x01\x00\x97\xDD"
-     "\x01\x00"
-     "\x98\xE5\x01\x00\x99\xED\x01\x00\x9A\xE5\x01\x00\x9B\xFD\x01\x00\x9C\xE5"
-     "\x01\x00"
-     "\x9D\xED\x01\x00\x9E\x39\x37\x00\x52\x33\x00\x80\xA0\x1D\x01\x00\x6E\x33"
-     "\x00\x80"
-     "\xA2\x29\x01\x00\x9E\x33\x00\x80\xF2\x33\x00\x80\xA5\x31\x01\x00\xA6\x31"
-     "\x01\x00"
-     "\x12\x34\x00\x80\x42\x2C\x00\x80\x86\x34\x00\x80\xAA\x05\x01\x00\xAB\x1D"
-     "\x01\x00"
-     "\xAC\x05\x01\x00\xAD\x0D\x01\x00\xAE\x05\x01\x00\xAF\x7D\x01\x00\xB0\x05"
-     "\x01\x00"
-     "\xB1\x0D\x01\x00\xB2\x05\x01\x00\xB3\x1D\x01\x00\xB4\x05\x01\x00\xB5\x0D"
-     "\x01\x00"
-     "\xB6\x05\x01\x00\xB7\x3D\x01\x00\xB8\x05\x01\x00\xB9\x0D\x01\x00\xBA\x05"
-     "\x01\x00"
-     "\xBB\x1D\x01\x00\xBC\xF5\x30\x00\xBD\x01\x01\x00\xBE\x01\x01\x00\xBF\x01"
-     "\x01\x00"
-     "\x8D\x8D\x00\x00\x8C\x11\x1C\x00\x8F\x0D\x01\x00\x8E\x0D\x01\x00\x89\x8D"
-     "\x07\x00"
-     "\x88\x85\x07\x00\x8B\x91\x00\x00\x8A\xB9\x07\x00\x85\x5D\x07\x00\x84\x75"
-     "\x07\x00"
-     "\x87\xA5\x07\x00\x86\xA9\x07\x00\x81\x0D\x07\x00\x80\x0D\x07\x00\x83\x75"
-     "\x07\x00"
-     "\x82\x79\x07\x00\x9D\x75\x01\x00\x9C\x75\x01\x00\x9F\x71\x01\x00\x9E\x79"
-     "\x01\x00"
-     "\x99\x75\x01\x00\x98\x75\x01\x00\x9B\x75\x01\x00\x9A\x75\x01\x00\x95\x79"
-     "\x01\x00"
-     "\x94\x71\x01\x00\x97\xD5\x0A\x00\x96\x71\x01\x00\x91\x55\x1C\x00\x90\x7D"
-     "\x01\x00"
-     "\x93\x69\x01\x00\x92\x71\x01\x00\xAD\xED\x01\x00\xAC\xED\x01\x00\x66\x35"
-     "\x00\x80"
-     "\x8A\x35\x00\x80\xA9\xCD\x01\x00\xA8\xCD\x01\x00\xAB\xFD\x01\x00\xAA\xFD"
-     "\x01\x00"
-     "\xA5\xF1\x01\x00\xA4\xE9\x01\x00\xA7\x0D\x06\x00\xA6\xF9\x01\x00\xA1\xF1"
-     "\x01\x00"
-     "\xA0\x89\x01\x00\xA3\xF1\x01\x00\xA2\xF9\x01\x00\xBD\x5D\x01\x00\xBC\xB5"
-     "\x06\x00"
-     "\xBF\x6D\x01\x00\xBE\x6D\x01\x00\xB9\x41\x01\x00\xB8\x09\x06\x00\xBB\x41"
-     "\x01\x00"
-     "\xBA\x49\x01\x00\xB5\x65\x01\x00\xB4\x65\x01\x00\xB7\x65\x01\x00\xB6\x65"
-     "\x01\x00"
-     "\xB1\x85\x01\x00\xB0\x85\x01\x00\xB3\x85\x01\x00\xB2\x85\x01\x00\x86\x35"
-     "\x00\x80"
-     "\x8E\x35\x00\x80\x92\x35\x00\x80\xAA\x35\x00\x80\xAE\x35\x00\x80\xD2\x35"
-     "\x00\x80"
-     "\xE6\x35\x00\x80\x02\x36\x00\x80\x6E\x36\x00\x80\x76\x36\x00\x80\x8A\x36"
-     "\x00\x80"
-     "\xB2\x36\x00\x80\x02\x37\x00\x80\x36\x37\x00\x80\x9E\x37\x00\x80\x46\x38"
-     "\x00\x80"
-     "\x5E\x38\x00\x80\x92\x38\x00\x80\x96\x38\x00\x80\xDA\x38\x00\x80\xDE\x38"
-     "\x00\x80"
-     "\x36\x39\x00\x80\x12\x3A\x00\x80\x56\x3A\x00\x80\x6E\x3A\x00\x80\xB7\x30"
-     "\x00\x80"
-     "\xBF\xF2\x00\x80\x1A\x37\x00\x80\xE6\x2B\x00\x80\xF1\x2C\x00\x80\xB1\x2C"
-     "\x00\x80"
-     "\x59\x2C\x00\x80\x99\x2C\x00\x80\xC5\x2C\x00\x80\x55\x2D\x00\x80\x71\x2D"
-     "\x00\x80"
-     "\x45\x2E\x00\x80\x55\x2E\x00\x80\x71\x2E\x00\x80\xCD\x2E\x00\x80\x18\x2F"
-     "\x00\x80"
-     "\x24\x2F\x00\x80\x40\x2F\x00\x80\x50\x2F\x00\x80\x3F\x30\x00\x80\x47\x30"
-     "\x00\x80"
-     "\x77\x30\x00\x80\x97\x30\x00\x80\xC7\x30\x00\x80\xDF\x30\x00\x80\xCF\x30"
-     "\x00\x80"
-     "\xE3\x30\x00\x80\x23\x31\x00\x80\x7B\x31\x00\x80\x87\x31\x00\x80\x8F\x31"
-     "\x00\x80"
-     "\xB7\x31\x00\x80\x3D\x32\x00\x80\x51\x32\x00\x80\x92\x33\x00\x80\x8A\x33"
-     "\x00\x80"
-     "\xBA\x33\x00\x80\x4A\x35\x00\x80\x16\x34\x00\x80\x80\xA9\x03\x00\x81\xA9"
-     "\x03\x00"
-     "\x82\xB9\x03\x00\x83\xB9\x03\x00\x84\xA9\x03\x00\x85\xA9\x03\x00\x86\xD9"
-     "\x03\x00"
-     "\x87\xD9\x03\x00\x88\xE9\x03\x00\x89\xD9\x04\x00\x8A\xE5\x03\x00\x8B\x15"
-     "\x05\x00"
-     "\x8C\xE1\x03\x00\x8D\xE1\x03\x00\x8E\xE1\x03\x00\x8F\xE1\x03\x00\x90\xA1"
-     "\x03\x00"
-     "\x91\x31\x04\x00\x92\x91\x0C\x00\x93\xA9\x03\x00\x94\xB9\x03\x00\x95\xDD"
-     "\x35\x00"
-     "\x96\x69\x08\x00\x97\x25\x06\x00\x98\x8D\x03\x00\x99\x95\x03\x00\x9A\xE5"
-     "\x05\x00"
-     "\x9B\x99\x03\x00\x9C\xE1\x05\x00\x9D\x8D\x03\x00\x9E\x85\x03\x00\x9F\x7D"
-     "\x01\x00"
-     "\xA0\xFD\x04\x00\xA1\x81\x01\x00\xA2\x81\x01\x00\xA3\x81\x01\x00\xA4\x81"
-     "\x01\x00"
-     "\xA5\x81\x01\x00\xA6\x19\x04\x00\xA7\x85\x01\x00\xA8\xBD\x01\x00\xA9\x85"
-     "\x01\x00"
-     "\xAA\x8D\x01\x00\xAB\x85\x01\x00\xAC\x9D\x01\x00\xAD\xA9\x03\x00\xAE\x89"
-     "\x01\x00"
-     "\xAF\x89\x01\x00\xB0\x7D\x08\x00\xB1\xFD\x01\x00\xB2\xE5\x03\x00\xB3\xF1"
-     "\x01\x00"
-     "\xB4\x91\x01\x00\xB5\x91\x01\x00\xB6\x91\x01\x00\xB7\x91\x01\x00\xB8\xA1"
-     "\x03\x00"
-     "\xB9\xB5\x01\x00\xBA\x49\x04\x00\xBB\xB9\x01\x00\xBC\xB5\x03\x00\xBD\x15"
-     "\x0D\x00"
-     "\xBE\xA1\x01\x00\xBF\xB5\x03\x00\x4A\x34\x00\x80\x96\x34\x00\x80\xB2\x34"
-     "\x00\x80"
-     "\xD6\x34\x00\x80\xFA\x34\x00\x80\xFE\x34\x00\x80\x1E\x35\x00\x80\x2E\x35"
-     "\x00\x80"
-     "\x42\x35\x00\x80\x3E\x35\x00\x80\x6E\x35\x00\x80\xEA\x35\x00\x80\x1A\x36"
-     "\x00\x80"
-     "\x52\x36\x00\x80\x82\x36\x00\x80\x66\x37\x00\x80\x8E\x37\x00\x80\x02\x38"
-     "\x00\x80"
-     "\x4A\x38\x00\x80\x56\x38\x00\x80\x76\x38\x00\x80\x7A\x38\x00\x80\x86\x38"
-     "\x00\x80"
-     "\xA2\x38\x00\x80\x1A\x39\x00\x80\x4A\x39\x00\x80\x82\x39\x00\x80\xA2\x39"
-     "\x00\x80"
-     "\xEE\x39\x00\x80\x46\x3A\x00\x80\x62\x3A\x00\x80\xC2\x3A\x00\x80\x56\xF2"
-     "\x00\x80"
-     "\x51\xF2\x00\x80\x83\xF2\x00\x80\x0E\x2B\x00\x80\x36\x2B\x00\x80\x3A\x2B"
-     "\x00\x80"
-     "\x05\xF3\x00\x80\x32\xF3\x00\x80\xAA\xF3\x00\x80\x4E\x3B\x00\x80\x5E\x3B"
-     "\x00\x80"
-     "\xE1\x06\x00\x80\xEC\x06\x00\x80\xEF\x06\x00\x80\xE4\x06\x00\x80\xE8\x06"
-     "\x00\x80"
-     "\x4E\x08\x00\x80\xA4\x0E\x00\x80\x95\x0E\x00\x80\x9A\x0E\x00\x80\xC7\x0E"
-     "\x00\x80"
-     "\x9F\x0E\x00\x80\x34\x0F\x00\x80\xB3\x0F\x00\x80\x6A\x0F\x00\x80\x18\x0F"
-     "\x00\x80"
-     "\x43\x0F\x00\x80\x50\x0F\x00\x80\x58\x0F\x00\x80\x86\x0F\x00\x80\xAB\x0F"
-     "\x00\x80"
-     "\xA1\x0F\x00\x80\x8F\x81\xFF\x00\x8E\x05\xD0\x00\x8D\x75\xFF\x00\x8C\x6D"
-     "\xFD\x00"
-     "\x8B\x69\xFF\x00\x8A\x49\xFD\x00\x89\x6D\xFF\x00\x88\x89\xFA\x00\x87\x79"
-     "\xFD\x00"
-     "\x86\x55\xFF\x00\x85\x45\xFF\x00\x84\x45\xFF\x00\x83\x55\xFF\x00\x82\x55"
-     "\xFF\x00"
-     "\x81\x69\xFD\x00\x80\x61\xFF\x00\xA6\x0F\x00\x80\x93\x0F\x00\x80\x9A\x0F"
-     "\x00\x80"
-     "\xE7\x0E\x00\x80\xEC\x0E\x00\x80\xF1\x0E\x00\x80\x99\xF1\xFF\x00\x98\xF9"
-     "\xFF\x00"
-     "\x97\xC1\xFF\x00\x96\xF9\xFF\x00\x95\xF1\xFF\x00\x94\xF9\xFF\x00\x93\xE1"
-     "\xFF\x00"
-     "\x92\xF9\xFF\x00\x91\xF1\xFF\x00\x90\xF9\xFF\x00\xFE\x0E\x00\x80\x0B\x0F"
-     "\x00\x80"
-     "\x13\x0F\x00\x80\x1B\x0F\x00\x80\x25\x0F\x00\x80\x2A\x0F\x00\x80\x2F\x0F"
-     "\x00\x80"
-     "\x39\x0F\x00\x80\x3E\x0F\x00\x80\x46\x0F\x00\x80\x53\x0F\x00\x80\x5B\x0F"
-     "\x00\x80"
-     "\x60\x0F\x00\x80\x65\x0F\x00\x80\x6D\x0F\x00\x80\x72\x0F\x00\x80\x7C\x0F"
-     "\x00\x80"
-     "\x81\x0F\x00\x80\x89\x0F\x00\x80\x8E\x0F\x00\x80\xAE\x0F\x00\x80\x20\x0F"
-     "\x00\x80"
-     "\x03\x0F\x00\x80\x4B\x0F\x00\x80\x77\x0F\x00\x80\xF6\x0E\x00\x80\xC6\x16"
-     "\x00\x80"
-     "\xCF\x16\x00\x80\xD2\x16\x00\x80\xD8\x16\x00\x80\xCC\x16\x00\x80\xD5\x16"
-     "\x00\x80"
-     "\x80\xAD\xFE\x00\x81\xB5\xFE\x00\x82\xBD\xFE\x00\x83\xB5\xFE\x00\x84\xAD"
-     "\xFE\x00"
-     "\x85\xD5\xFE\x00\x86\xD9\xFE\x00\xC9\x16\x00\x80\xFC\x16\x00\x80\xFF\x16"
-     "\x00\x80"
-     "\xDE\x16\x00\x80\xDB\x16\x00\x80\xE1\x16\x00\x80\xE4\x16\x00\x80\xED\x16"
-     "\x00\x80"
-     "\xEA\x16\x00\x80\xF0\x16\x00\x80\xE7\x16\x00\x80\xF6\x16\x00\x80\x93\x89"
-     "\xFE\x00"
-     "\x94\x99\xFE\x00\x95\x99\xFE\x00\x96\x89\xFE\x00\x97\x89\xFE\x00\xF3\x16"
-     "\x00\x80"
-     "\x02\x17\x00\x80\x08\x17\x00\x80\x0E\x17\x00\x80\x0B\x17\x00\x80\x9D\xAD"
-     "\xFE\x00"
-     "\x11\x17\x00\x80\x9F\xA1\xFE\x00\xA0\x61\xFE\x00\xA1\xFD\xBB\x00\xA2\x19"
-     "\xBB\x00"
-     "\xA3\x69\xFE\x00\xA4\x79\xFE\x00\xA5\x79\xFE\x00\xA6\x69\xFE\x00\xA7\x69"
-     "\xFE\x00"
-     "\xA8\x59\xFE\x00\xA9\x49\xB6\x00\xAA\x55\xFE\x00\xAB\xED\x01\x00\xAC\xF5"
-     "\x01\x00"
-     "\xAD\xFD\x01\x00\xAE\xF5\x01\x00\xAF\xED\x01\x00\xB0\x95\x01\x00\xB1\x45"
-     "\x01\x00"
-     "\xB2\x4D\x01\x00\xB3\x45\x01\x00\xB4\x5D\x01\x00\xB5\x45\x01\x00\xB6\x4D"
-     "\x01\x00"
-     "\x14\x17\x00\x80\xB8\x79\x01\x00\xB9\x79\x01\x00\xBA\x49\x01\x00\xBB\x49"
-     "\x01\x00"
-     "\xBC\x59\x01\x00\x17\x17\x00\x80\xBE\x55\x01\x00\x05\x17\x00\x80\x80\xB1"
-     "\x02\x00"
-     "\x81\xB1\x02\x00\x29\x17\x00\x80\x83\xB5\x02\x00\x84\xAD\x02\x00\x26\x17"
-     "\x00\x80"
-     "\x86\xD9\x02\x00\x87\xD9\x02\x00\x88\xE9\x02\x00\x89\xE9\x02\x00\x8A\xF9"
-     "\x02\x00"
-     "\x8B\xF9\x02\x00\x8C\xE9\x02\x00\x8D\xE9\x02\x00\x8E\xD9\x02\x00\x8F\xD9"
-     "\x02\x00"
-     "\x90\xA9\x02\x00\x91\xAD\x02\x00\x92\xA5\x02\x00\x93\xA1\x02\x00\x94\xBD"
-     "\x02\x00"
-     "\x95\xB9\x02\x00\x96\xA9\x02\x00\x97\xAD\x02\x00\x98\x91\x02\x00\x99\x95"
-     "\x02\x00"
-     "\x9A\x9D\x02\x00\x9B\x99\x02\x00\x9C\x85\x02\x00\x9D\x81\x02\x00\x9E\x81"
-     "\x02\x00"
-     "\x9F\x85\x02\x00\xA0\x79\x02\x00\xA1\x7D\x02\x00\xA2\x75\x02\x00\xA3\x71"
-     "\x02\x00"
-     "\xA4\x6D\x02\x00\xA5\x69\x02\x00\xA6\x85\x03\x00\xA7\x81\x03\x00\xA8\xBD"
-     "\x03\x00"
-     "\xA9\xB9\x03\x00\xAA\x89\x03\x00\xAB\x8D\x03\x00\xAC\x91\x03\x00\xAD\x95"
-     "\x03\x00"
-     "\xAE\x9D\x03\x00\xAF\x99\x03\x00\xB0\xE5\x03\x00\xB1\xE1\x03\x00\xB2\xE1"
-     "\x03\x00"
-     "\xB3\xE5\x03\x00\xB4\xF9\x03\x00\xB5\xFD\x03\x00\xB6\xF5\x03\x00\xB7\xF1"
-     "\x03\x00"
-     "\xB8\xCD\x03\x00\xB9\xC9\x03\x00\xBA\xD9\x03\x00\xBB\xDD\x03\x00\xBC\xC1"
-     "\x03\x00"
-     "\xBD\xC5\x03\x00\xBE\xCD\x03\x00\xBF\xC9\x03\x00\x80\x35\x02\x00\x81\x31"
-     "\x02\x00"
-     "\x82\x31\x02\x00\x83\x35\x02\x00\x84\x2D\x02\x00\x85\x29\x02\x00\x86\x59"
-     "\x02\x00"
-     "\x87\x5D\x02\x00\x88\x65\x02\x00\x89\x61\x02\x00\x8A\x61\x02\x00\x8B\x65"
-     "\x02\x00"
-     "\x8C\x51\x02\x00\x8D\x55\x02\x00\x8E\x5D\x02\x00\x8F\x59\x02\x00\x90\x25"
-     "\x02\x00"
-     "\x91\x21\x02\x00\x92\x21\x02\x00\x93\x25\x02\x00\x94\x39\x02\x00\x95\x3D"
-     "\x02\x00"
-     "\x96\x35\x02\x00\x97\x31\x02\x00\x98\x0D\x02\x00\x99\x09\x02\x00\x9A\x19"
-     "\x02\x00"
-     "\x9B\x1D\x02\x00\x9C\x01\x02\x00\x9D\x05\x02\x00\x9E\x01\x02\x00\x9F\x05"
-     "\x02\x00"
-     "\xA0\x5D\x02\x00\xA1\x59\x02\x00\xA2\x55\x02\x00\xA3\x51\x02\x00\xA4\xA1"
-     "\xA7\x02"
-     "\xA5\xA5\xA7\x02\xA6\x01\x04\x00\xA7\x05\x04\x00\xA8\x39\x04\x00\xA9\x3D"
-     "\x04\x00"
-     "\xAA\x5D\x02\x00\xAB\x59\x02\x00\xAC\x45\x02\x00\xAD\x41\x02\x00\xAE\xB9"
-     "\x07\x00"
-     "\xAF\xBD\x07\x00\xB0\xE5\xA7\x02\xB1\xE1\xA7\x02\x31\x17\x00\x80\x37\x17"
-     "\x00\x80"
-     "\x23\x17\x00\x80\x34\x17\x00\x80\x3D\x17\x00\x80\x62\x16\x00\x80\xCA\x0F"
-     "\x00\x80"
-     "\x1A\x10\x00\x80\xF7\x0F\x00\x80\x0B\x10\x00\x80\x06\x10\x00\x80\x10\x10"
-     "\x00\x80"
-     "\x15\x10\x00\x80\xFC\x0F\x00\x80\x40\x17\x00\x80\x3A\x17\x00\x80\x91\x40"
-     "\x08\x00"
-     "\xCF\x0F\x00\x80\xD4\x0F\x00\x80\xE8\x0F\x00\x80\x01\x10\x00\x80\x41\x10"
-     "\x00\x80"
-     "\x46\x10\x00\x80\x52\x10\x00\x80\x68\x10\x00\x80\x77\x10\x00\x80\x7C\x10"
-     "\x00\x80"
-     "\x87\x10\x00\x80\xA1\x10\x00\x80\xB4\x10\x00\x80\xD8\x10\x00\x80\x0A\x11"
-     "\x00\x80"
-     "\x0F\x11\x00\x80\x17\x11\x00\x80\x26\x11\x00\x80\x35\x11\x00\x80\x3A\x11"
-     "\x00\x80"
-     "\x42\x11\x00\x80\x65\x11\x00\x80\x8C\x11\x00\x80\x98\x11\x00\x80\xB8\x11"
-     "\x00\x80"
-     "\xBD\x11\x00\x80\xC2\x11\x00\x80\xFC\x11\x00\x80\x0F\x12\x00\x80\x1B\x12"
-     "\x00\x80"
-     "\x33\x12\x00\x80\xAD\x12\x00\x80\x03\x13\x00\x80\x1C\x13\x00\x80\x21\x13"
-     "\x00\x80"
-     "\x34\x13\x00\x80\x45\x13\x00\x80\x57\x13\x00\x80\x5C\x13\x00\x80\x83\x13"
-     "\x00\x80"
-     "\x8B\x13\x00\x80\xA0\x13\x00\x80\xC7\x13\x00\x80\xCC\x13\x00\x80\x26\x14"
-     "\x00\x80"
-     "\x2B\x14\x00\x80\x30\x14\x00\x80\x3C\x14\x00\x80\x48\x14\x00\x80\x4D\x14"
-     "\x00\x80"
-     "\x55\x14\x00\x80\x61\x14\x00\x80\x7B\x14\x00\x80\x80\x14\x00\x80\x88\x14"
-     "\x00\x80"
-     "\x8D\x14\x00\x80\x92\x14\x00\x80\x97\x14\x00\x80\x9C\x14\x00\x80\xA1\x14"
-     "\x00\x80"
-     "\xD9\xB8\xA2\x02\x93\xFD\x04\x00\xB4\x14\x00\x80\xB9\x14\x00\x80\xD5\x14"
-     "\x00\x80"
-     "\x97\x19\x07\x00\x96\xE9\x04\x00\x95\xE5\x04\x00\x94\xE1\x04\x00\x9B\xE9"
-     "\x06\x00"
-     "\x9A\x31\x07\x00\x99\x3D\x07\x00\x98\x25\x07\x00\x9F\xFD\x06\x00\x9E\xF9"
-     "\x06\x00"
-     "\x9D\x3D\xA1\x02\x9C\xF5\x06\x00\xEF\x14\x00\x80\x09\x15\x00\x80\x15\x15"
-     "\x00\x80"
-     "\x2D\x15\x00\x80\x32\x15\x00\x80\x3F\x15\x00\x80\x60\x15\x00\x80\x83\x15"
-     "\x00\x80"
-     "\x9D\x15\x00\x80\xA9\x15\x00\x80\xAE\x15\x00\x80\xB6\x15\x00\x80\xD7\x15"
-     "\x00\x80"
-     "\xF1\x15\x00\x80\x00\x16\x00\x80\x1D\x16\x00\x80\xB3\x71\x06\x00\xB2\x75"
-     "\x06\x00"
-     "\xB1\x65\x06\x00\xB0\x61\x06\x00\xB7\x69\x06\x00\xB6\x6D\x06\x00\xB5\x6D"
-     "\x06\x00"
-     "\xB4\x69\x06\x00\xBB\x5D\x06\x00\xBA\x59\x06\x00\xB9\x55\x06\x00\xB8\x55"
-     "\x06\x00"
-     "\xBF\xB5\x01\x00\xBE\xB1\x01\x00\xBD\xBD\x01\x00\xBC\xB9\x01\x00\xA3\x15"
-     "\x06\x00"
-     "\xA2\x11\x06\x00\xA1\x19\x06\x00\xA0\x1D\x06\x00\xA7\x09\x06\x00\xA6\x0D"
-     "\x06\x00"
-     "\xA5\x01\x06\x00\xA4\x05\x06\x00\xAB\x01\x06\x00\xAA\x05\x06\x00\xA9\x35"
-     "\x06\x00"
-     "\xA8\x31\x06\x00\xAF\x19\x06\x00\xAE\x1D\x06\x00\xAD\x1D\x06\x00\xAC\x19"
-     "\x06\x00"
-     "\x82\x15\x02\x00\x83\xF9\x05\x00\x80\x05\x02\x00\x81\x1D\x02\x00\x86\x31"
-     "\x02\x00"
-     "\x87\x31\x02\x00\x84\x01\x02\x00\x85\x01\x02\x00\x8A\x11\x02\x00\x8B\x11"
-     "\x02\x00"
-     "\x88\x01\x02\x00\x89\x01\x02\x00\x8E\x71\x02\x00\x8F\x71\x02\x00\x8C\x01"
-     "\x02\x00"
-     "\x8D\x01\x02\x00\x92\x11\x02\x00\x93\x11\x02\x00\x90\x01\x02\x00\x91\x01"
-     "\x02\x00"
-     "\x96\x31\x02\x00\x97\x31\x02\x00\x94\x01\x02\x00\x95\x01\x02\x00\x9A\x11"
-     "\x02\x00"
-     "\x9B\x11\x02\x00\x98\x01\x02\x00\x99\x01\x02\x00\x9E\xF1\x02\x00\x9F\xF1"
-     "\x02\x00"
-     "\x9C\x01\x02\x00\x9D\x01\x02\x00\xA2\x11\x02\x00\xA3\x11\x02\x00\xA0\x01"
-     "\x02\x00"
-     "\xA1\x01\x02\x00\xA6\x31\x02\x00\xA7\x31\x02\x00\xA4\x01\x02\x00\xA5\x01"
-     "\x02\x00"
-     "\xAA\x11\x02\x00\xAB\x11\x02\x00\xA8\x01\x02\x00\xA9\x01\x02\x00\xAE\x71"
-     "\x02\x00"
-     "\xAF\x71\x02\x00\xAC\x01\x02\x00\xAD\x01\x02\x00\xB2\x11\x02\x00\xB3\x11"
-     "\x02\x00"
-     "\xB0\x01\x02\x00\xB1\x01\x02\x00\xB6\x31\x02\x00\xB7\x31\x02\x00\xB4\x01"
-     "\x02\x00"
-     "\xB5\x01\x02\x00\xBA\x11\x02\x00\xBB\x11\x02\x00\xB8\x01\x02\x00\xB9\x01"
-     "\x02\x00"
-     "\xBE\xF9\x03\x00\xBF\xF9\x03\x00\xBC\x01\x02\x00\xBD\xF9\x03\x00\x22\x16"
-     "\x00\x80"
-     "\x2A\x16\x00\x80\x2F\x16\x00\x80\x42\x16\x00\x80\x47\x16\x00\x80\x6D\x16"
-     "\x00\x80"
-     "\x79\x16\x00\x80\x85\x16\x00\x80\x94\x16\x00\x80\xB1\x16\x00\x80\xB6\x16"
-     "\x00\x80"
-     "\xD7\x11\x00\x80\xE8\x11\x00\x80\x65\x16\x00\x80\x81\x00\x00\x80\x8B\x00"
-     "\x00\x80"
-     "\x95\x00\x00\x80\x9F\x00\x00\x80\xA9\x00\x00\x80\xBA\xF1\x1A\x00\xB3\x00"
-     "\x00\x80"
-     "\xDE\x0F\x00\x80\xE3\x0F\x00\x80\xED\x0F\x00\x80\x5E\x10\x00\x80\xD9\x14"
-     "\xA1\x02"
-     "\x63\x10\x00\x80\x6D\x10\x00\x80\xCE\x10\x00\x80\xD3\x10\x00\x80\x00\x11"
-     "\x00\x80"
-     "\x1C\x11\x00\x80\x21\x11\x00\x80\x2B\x11\x00\x80\x3A\x15\x00\x80\xF6\x15"
-     "\x00\x80"
-     "\xFB\x15\x00\x80\x13\x16\x00\x80\x1D\x17\x00\x80\x8A\x16\x00\x80\x8F\x16"
-     "\x00\x80"
-     "\xA7\x16\x00\x80\xD9\x0F\x00\x80\xF2\x0F\x00\x80\x72\x10\x00\x80\x05\x11"
-     "\x00\x80"
-     "\xC0\x12\x00\x80\x28\x15\x00\x80\x18\x16\x00\x80\x4C\x16\x00\x80\xAC\x16"
-     "\x00\x80"
-     "\x30\x11\x00\x80\x4D\x12\x00\x80\x88\x12\x00\x80\x9B\x12\x00\x80\xFA\x13"
-     "\x00\x80"
-     "\x06\x14\x00\x80\x12\x14\x00\x80\x76\x13\x00\x80\x7B\x13\x00\x80\xBA\x13"
-     "\x00\x80"
-     "\xBF\x13\x00\x80\xE6\x13\x00\x80\xEB\x13\x00\x80\x95\x15\x0F\x00\x94\x15"
-     "\x0F\x00"
-     "\x97\x15\x0F\x00\x96\x15\x0F\x00\x91\x15\x0F\x00\x90\x15\x0F\x00\x93\x15"
-     "\x0F\x00"
-     "\x92\x15\x0F\x00\x9D\x15\x0F\x00\x9C\x15\x0F\x00\x9F\x15\x0F\x00\x9E\x15"
-     "\x0F\x00"
-     "\x99\x15\x0F\x00\x98\x15\x0F\x00\x9B\x15\x0F\x00\x9A\x15\x0F\x00\x85\x15"
-     "\x0D\x00"
-     "\x84\x15\x0D\x00\x87\x15\x0D\x00\x86\x15\x0D\x00\x81\x15\x0D\x00\x80\x15"
-     "\x0D\x00"
-     "\x83\x15\x0D\x00\x82\x15\x0D\x00\x8D\x15\x0D\x00\x8C\x15\x0D\x00\x8F\x15"
-     "\x0D\x00"
-     "\x8E\x15\x0D\x00\x89\x15\x0D\x00\x88\x15\x0D\x00\x8B\x15\x0D\x00\x8A\x15"
-     "\x0D\x00"
-     "\xB5\xC9\x0C\x00\xB4\xC1\x0C\x00\xB7\x09\x0F\x00\xB6\xF1\x0F\x00\xB1\xE5"
-     "\x0F\x00"
-     "\xB0\xE5\x0F\x00\xB3\xE1\x0F\x00\xB2\xDD\x0C\x00\xBD\x71\x0C\x00\xBC\x69"
-     "\x0C\x00"
-     "\xBF\x49\x0C\x00\xBE\x71\x0C\x00\xB9\x35\x0F\x00\xB8\xE5\x0C\x00\xBB\xE1"
-     "\x0C\x00"
-     "\xBA\xE9\x0C\x00\xA5\x99\x0F\x00\xA4\x91\x0F\x00\xA7\x95\x0F\x00\xA6\xD9"
-     "\x0C\x00"
-     "\xA1\x95\x0F\x00\xA0\x95\x0F\x00\xA3\x89\x0F\x00\xA2\x95\x0F\x00\xAD\x8D"
-     "\x0F\x00"
-     "\xAC\xCD\x0C\x00\xAF\xD9\x0C\x00\xAE\xC1\x0C\x00\xA9\xE9\x0C\x00\xA8\x09"
-     "\x0B\x00"
-     "\xAB\x91\x0F\x00\xAA\x9D\x0F\x00\x80\xE5\x0D\x00\x81\xF1\x0D\x00\x82\xF1"
-     "\x0D\x00"
-     "\x83\x09\x0C\x00\x84\x19\x0C\x00\x85\x5D\x0C\x00\x86\x55\x0C\x00\x87\x4D"
-     "\x0C\x00"
-     "\x88\xA9\x0E\x00\x89\x45\x0C\x00\x8A\xA5\x0E\x00\x8B\xBD\x0E\x00\x8C\x49"
-     "\x0C\x00"
-     "\x8D\xA1\x0E\x00\x8E\x45\x0C\x00\x8F\x3D\x0E\x00\x90\x75\x0E\x00\x91\xD9"
-     "\x0E\x00"
-     "\x92\xE9\x0E\x00\x93\x6D\x0E\x00\x94\xF5\x0E\x00\x95\x71\x0E\x00\x96\x71"
-     "\x0E\x00"
-     "\x97\x51\x0D\x00\x98\x71\x0D\x00\x99\xCD\x0E\x00\x9A\x7D\x0D\x00\x9B\xC1"
-     "\x0E\x00"
-     "\x9C\x6D\x0D\x00\x9D\x55\x0D\x00\x9E\x5D\x0D\x00\x9F\x55\x0D\x00\xA0\x31"
-     "\x0E\x00"
-     "\xA1\xB1\x0D\x00\xA2\xB1\x0D\x00\xA3\xB1\x0D\x00\xA4\xD1\x0D\x00\xA5\x21"
-     "\x0E\x00"
-     "\xA6\xC9\x0D\x00\xA7\xC1\x0D\x00\xA8\xC1\x0D\x00\xA9\xC1\x0D\x00\xAA\xC1"
-     "\x0D\x00"
-     "\xAB\xC1\x0D\x00\xAC\xC5\x0D\x00\xAD\xCD\x0D\x00\xAE\xC5\x0D\x00\xAF\x3D"
-     "\x0D\x00"
-     "\xB0\x45\x0D\x00\xB1\x4D\x0D\x00\xB2\x71\x0E\x00\xB3\x41\x0D\x00\xB4\x41"
-     "\x0D\x00"
-     "\xB5\x41\x0D\x00\xB6\x41\x0D\x00\xB7\x41\x0D\x00\xB8\x41\x0D\x00\xB9\x4D"
-     "\x0D\x00"
-     "\xBA\x45\x0D\x00\xBB\x5D\x0D\x00\xBC\x45\x0D\x00\xBD\x4D\x0D\x00\xBE\x45"
-     "\x0D\x00"
-     "\xBF\x3D\x0D\x00\x80\xC5\x0E\x00\x81\xCD\x0E\x00\x82\xDD\x0E\x00\x83\xD5"
-     "\x0E\x00"
-     "\x84\xF9\x0E\x00\x85\xF9\x0E\x00\x86\xE9\x0E\x00\x87\xE9\x0E\x00\x88\xD9"
-     "\x0E\x00"
-     "\x89\x29\x0F\x00\x8A\x69\x0F\x00\x8B\x69\x0F\x00\x8C\x79\x0F\x00\x8D\x89"
-     "\x0D\x00"
-     "\x8E\x6D\x0F\x00\x8F\x65\x0F\x00\x90\x1D\x0F\x00\x91\x25\x0F\x00\x92\x25"
-     "\x0F\x00"
-     "\x93\x3D\x0F\x00\x94\x25\x0F\x00\x95\x2D\x0F\x00\x96\x99\x0D\x00\x97\x59"
-     "\x0D\x00"
-     "\x98\x69\x0D\x00\x99\xA1\x0D\x00\x9A\x7D\x0D\x00\x9B\x75\x0D\x00\x9C\x6D"
-     "\x0D\x00"
-     "\x9D\x55\x0D\x00\x9E\xB1\x0D\x00\x9F\x69\x0E\x00\xA0\x2D\x0D\x00\xA1\xAD"
-     "\x0E\x00"
-     "\xA2\x39\x0D\x00\xA3\xCD\x0E\x00\xA4\x25\x0D\x00\xA5\xC5\x0E\x00\xA6\x55"
-     "\x0D\x00"
-     "\xA7\x19\x0E\x00\xA8\x71\x0D\x00\xA9\x71\x0D\x00\xAA\x71\x0D\x00\xAB\x59"
-     "\x0E\x00"
-     "\xAC\x49\x0E\x00\xAD\xFD\x0F\x00\xAE\xC1\x0F\x00\xAF\x7D\x0D\x00\xB0\xE1"
-     "\x0D\x00"
-     "\xB1\x0D\x0D\x00\xB2\x15\x0D\x00\xB3\x2D\x0D\x00\xB4\x35\x0D\x00\xB5\x3D"
-     "\x0D\x00"
-     "\xB6\x35\x0D\x00\xB7\x2D\x0D\x00\xB8\x15\x0D\x00\xB9\x1D\x0D\x00\xBA\x15"
-     "\x0D\x00"
-     "\xBB\xE9\x01\x00\xBC\xF9\x01\x00\xBD\xF9\x01\x00\xBE\xE9\x01\x00\xBF\xE9"
-     "\x01\x00"
-     "\xF0\x4C\x0D\x00\x52\x12\x00\x80\x57\x12\x00\x80\xA0\x12\x00\x80\xA5\x12"
-     "\x00\x80"
-     "\xAB\x11\x00\x80\xB0\x11\x00\x80\x7F\x11\x00\x80\x84\x11\x00\x80\xC7\x11"
-     "\x00\x80"
-     "\xCC\x11\x00\x80\x0F\x13\x00\x80\x14\x13\x00\x80\x4A\x13\x00\x80\x4F\x13"
-     "\x00\x80"
-     "\x5F\x12\x00\x80\x6B\x12\x00\x80\x7E\x12\x00\x80\x83\x12\x00\x80\x2E\x12"
-     "\x00\x80"
-     "\xC5\x12\x00\x80\x40\x13\x00\x80\x34\x10\x00\x80\x8C\x10\x00\x80\xA6\x10"
-     "\x00\x80"
-     "\xAD\x10\x00\x80\xB9\x10\x00\x80\xDD\x10\x00\x80\xE4\x10\x00\x80\xEB\x10"
-     "\x00\x80"
-     "\x6A\x11\x00\x80\xA4\x11\x00\x80\x9D\x11\x00\x80\x14\x12\x00\x80\x01\x12"
-     "\x00\x80"
-     "\x08\x12\x00\x80\x3F\x12\x00\x80\x38\x12\x00\x80\x46\x12\x00\x80\xB2\x12"
-     "\x00\x80"
-     "\x08\x13\x00\x80\x70\x12\x00\x80\x64\x12\x00\x80\x8D\x12\x00\x80\x94\x12"
-     "\x00\x80"
-     "\x26\x13\x00\x80\x39\x13\x00\x80\x61\x13\x00\x80\x68\x13\x00\x80\x6F\x13"
-     "\x00\x80"
-     "\x90\x13\x00\x80\xA5\x13\x00\x80\xAC\x13\x00\x80\xD1\x13\x00\x80\xDF\x13"
-     "\x00\x80"
-     "\xD8\x13\x00\x80\x35\x14\x00\x80\x66\x14\x00\x80\x6D\x14\x00\x80\xF4\x14"
-     "\x00\x80"
-     "\x02\x15\x00\x80\xFB\x14\x00\x80\xDA\x14\x00\x80\x0E\x15\x00\x80\x90\x55"
-     "\x07\x00"
-     "\x91\xED\x0B\x00\x92\xE5\x0B\x00\x93\xFD\x0B\x00\x94\xE5\x0B\x00\x95\xED"
-     "\x0B\x00"
-     "\x96\xE5\x0B\x00\x97\x19\x07\x00\x98\x29\x07\x00\x99\x29\x07\x00\x9A\x39"
-     "\x07\x00"
-     "\x9B\x39\x07\x00\x9C\x29\x07\x00\x9D\x29\x07\x00\x9E\x19\x07\x00\x9F\x19"
-     "\x07\x00"
-     "\x80\x59\x07\x00\x81\x59\x07\x00\x82\x69\x07\x00\x83\x69\x07\x00\x84\x79"
-     "\x07\x00"
-     "\x85\x79\x07\x00\x86\x69\x07\x00\x87\x69\x07\x00\x88\x59\x07\x00\x89\x59"
-     "\x07\x00"
-     "\x8A\x29\x07\x00\x8B\x29\x07\x00\x8C\xA5\x0B\x00\x8D\x3D\x07\x00\x8E\x35"
-     "\x07\x00"
-     "\x8F\x2D\x07\x00\xB0\x55\x0B\x00\xB1\x55\x0B\x00\xB2\x51\x0B\x00\xB3\x29"
-     "\x08\x00"
-     "\xB4\xE9\x08\x00\xB5\xE9\x08\x00\xB6\x19\x08\x00\xB7\xA1\x07\x00\xB8\xE1"
-     "\x07\x00"
-     "\xB9\xE1\x07\x00\xBA\x0D\x08\x00\xBB\x05\x08\x00\xBC\xE9\x07\x00\xBD\xED"
-     "\x07\x00"
-     "\x1A\x15\x00\x80\x65\x15\x00\x80\xA0\xE9\x07\x00\xA1\xE9\x07\x00\xA2\xF9"
-     "\x07\x00"
-     "\xA3\xF9\x07\x00\xA4\xE9\x07\x00\xA5\xE9\x07\x00\xA6\x99\x07\x00\xA7\x99"
-     "\x07\x00"
-     "\xA8\x35\x0B\x00\xA9\xAD\x07\x00\xAA\xA5\x07\x00\xAB\xBD\x07\x00\xAC\xA5"
-     "\x07\x00"
-     "\xAD\xC9\x07\x00\xAE\xB9\x07\x00\xAF\xB9\x07\x00\x90\x5D\x06\x00\x91\x65"
-     "\x06\x00"
-     "\x92\x69\x06\x00\x93\x69\x06\x00\x94\x79\x06\x00\x95\x79\x06\x00\x96\x69"
-     "\x06\x00"
-     "\x97\x69\x06\x00\x98\x59\x06\x00\x99\x5D\x06\x00\x9A\x55\x06\x00\x9B\xAD"
-     "\x06\x00"
-     "\x9C\xB5\x06\x00\x9D\xBD\x06\x00\x9E\xB5\x06\x00\x9F\xAD\x06\x00\x6C\x15"
-     "\x00\x80"
-     "\x7C\x15\x00\x80\x44\x15\x00\x80\x52\x15\x00\x80\x88\x15\x00\x80\x8F\x15"
-     "\x00\x80"
-     "\x4B\x15\x00\x80\x34\x16\x00\x80\x3B\x16\x00\x80\xDC\x15\x00\x80\xE3\x15"
-     "\x00\x80"
-     "\xC2\x15\x00\x80\xC9\x15\x00\x80\x0C\x16\x00\x80\x05\x16\x00\x80\x99\x16"
-     "\x00\x80"
-     "\xB0\x39\x06\x00\xB1\x39\x06\x00\xB2\x35\x06\x00\xB3\x4D\x06\x00\xB4\x55"
-     "\x06\x00"
-     "\xB5\x5D\x06\x00\xB6\x55\x06\x00\xB7\x51\x06\x00\xB8\x71\x06\x00\xB9\x71"
-     "\x06\x00"
-     "\xBA\x71\x06\x00\xBB\x71\x06\x00\xBC\x51\x06\x00\xBD\x55\x06\x00\xBE\x5D"
-     "\x06\x00"
-     "\xBF\x55\x06\x00\xA0\x51\x06\x00\xA1\x51\x06\x00\xA2\x51\x06\x00\xA3\x55"
-     "\x06\x00"
-     "\xA4\x4D\x06\x00\xA5\x49\x06\x00\xA6\x79\x06\x00\xA7\x79\x06\x00\xA8\x45"
-     "\x06\x00"
-     "\xA9\x4D\x06\x00\xAA\x45\x06\x00\xAB\x41\x06\x00\xAC\x41\x06\x00\xAD\x45"
-     "\x06\x00"
-     "\xAE\x4D\x06\x00\xAF\x45\x06\x00\x80\xED\x05\x00\x81\xF5\x05\x00\x82\xFD"
-     "\x05\x00"
-     "\x83\xF5\x05\x00\x84\xE9\x05\x00\x85\xE9\x05\x00\x86\xE5\x05\x00\x87\xA5"
-     "\x02\x00"
-     "\x88\x99\x02\x00\x89\x99\x02\x00\x8A\x69\x03\x00\x8B\x69\x03\x00\x8C\x79"
-     "\x03\x00"
-     "\x8D\x79\x03\x00\x8E\x69\x03\x00\x8F\x69\x03\x00\x82\xE4\x0B\x00\x57\x10"
-     "\x00\x80"
-     "\x92\x11\x03\x00\x93\x11\x03\x00\x94\x31\x03\x00\x95\x31\x03\x00\x96\x31"
-     "\x03\x00"
-     "\x97\x31\x03\x00\x98\x0D\x03\x00\x99\x15\x03\x00\x9A\x1D\x03\x00\x9B\x15"
-     "\x03\x00"
-     "\x9C\x0D\x03\x00\x9D\x09\x03\x00\x9E\x3D\x00\x00\x9F\x7D\x01\x00\xA0\x85"
-     "\x01\x00"
-     "\xA1\x8D\x01\x00\xA2\x85\x01\x00\xA3\x9D\x01\x00\xA4\x85\x01\x00\xA5\x8D"
-     "\x01\x00"
-     "\xA6\x85\x01\x00\xA7\xBD\x01\x00\xA8\x85\x01\x00\xA9\x8D\x01\x00\xAA\x85"
-     "\x01\x00"
-     "\xAB\x9D\x01\x00\xAC\x85\x01\x00\xAD\x8D\x01\x00\xAE\x85\x01\x00\xAF\xF9"
-     "\x01\x00"
-     "\xB0\x89\x01\x00\xB1\x89\x01\x00\xB2\x99\x01\x00\xB3\x99\x01\x00\xB4\x35"
-     "\x05\x00"
-     "\xB5\x39\x05\x00\xB6\x81\x01\x00\xB7\x81\x01\x00\xB8\x81\x01\x00\xB9\x81"
-     "\x01\x00"
-     "\xBA\x81\x01\x00\xBB\x81\x01\x00\xBC\x99\x01\x00\xBD\x95\x01\x00\xBE\x89"
-     "\x01\x00"
-     "\xBF\x89\x01\x00\x9A\x10\x00\x80\x93\x10\x00\x80\xC7\x10\x00\x80\xC0\x10"
-     "\x00\x80"
-     "\xF9\x10\x00\x80\xF2\x10\x00\x80\x78\x11\x00\x80\x47\x11\x00\x80\x71\x11"
-     "\x00\x80"
-     "\x20\x12\x00\x80\xB9\x12\x00\x80\x77\x12\x00\x80\x2D\x13\x00\x80\xE8\x14"
-     "\x00\x80"
-     "\x21\x15\x00\x80\x7E\x16\x00\x80\xF0\x20\x11\x00\x72\x16\x00\x80\xA0\x16"
-     "\x00\x80"
-     "\xA2\x15\x00\x80\x74\x14\x00\x80\xEA\x15\x00\x80\xB3\x13\x00\x80\xAD\x14"
-     "\x00\x80"
-     "\xBB\x15\x00\x80\x96\x15\x00\x80\xE1\x14\x00\x80\xA6\x14\x00\x80\x4E\x11"
-     "\x00\x80"
-     "\x91\x11\x00\x80\x59\x15\x00\x80\x41\x14\x00\x80\x4B\x10\x00\x80\x27\x12"
-     "\x00\x80"
-     "\xD0\x15\x00\x80\xFC\x12\x00\x80\x5A\x14\x00\x80\x2B\x10\x00\x80\x22\x10"
-     "\x00\x80"
-     "\x73\x15\x00\x80\xCA\x12\x00\x80\xDF\x11\x00\x80\x97\x13\x00\x80\x54\x16"
-     "\x00\x80"
-     "\xD3\x12\x00\x80\xDA\x12\x00\x80\x55\x11\x00\x80\xED\x11\x00\x80\x3E\x03"
-     "\x00\x80"
-     "\x3A\x24\x00\x80\x3E\x24\x00\x80\x38\x06\x00\x80\xC7\x00\x00\x80\x4D\x06"
-     "\x00\x80"
-     "\xD8\x24\x00\x80\xDC\x24\x00\x80\x86\x20\x00\x80\x82\x20\x00\x80\x5F\x06"
-     "\x00\x80"
-     "\x8F\x08\x00\x80\x93\x08\x00\x80\x6E\x24\x00\x80\xD4\x24\x00\x80\x62\x24"
-     "\x00\x80"
-     "\x66\x24\x00\x80\x4A\x24\x00\x80\x4E\x24\x00\x80\xCC\xE0\xA0\x00\x52\x24"
-     "\x00\x80"
-     "\x56\x24\x00\x80\x5A\x24\x00\x80\x5E\x24\x00\x80\x57\x06\x00\x80\x5B\x06"
-     "\x00\x80"
-     "\xD1\x00\x00\x80\xD7\x00\x00\x80\x3A\x03\x00\x80\x40\x03\x00\x80\x40\x06"
-     "\x00\x80"
-     "\x4B\x06\x00\x80\x83\x2D\x1F\x00\x82\xC5\x1F\x00\x81\x35\x1F\x00\x80\x35"
-     "\x1F\x00"
-     "\x87\xD9\x1F\x00\x86\xD1\x1F\x00\x85\xF9\x1B\x00\x84\x95\x1B\x00\x59\x06"
-     "\x00\x80"
-     "\xD3\x00\x00\x80\xD5\x00\x00\x80\x55\x06\x00\x80\x79\x00\x00\x80\xF0\x13"
-     "\x00\x80"
-     "\x7D\x00\x00\x80\x87\x00\x00\x80\xB3\x19\x1F\x00\xB2\x11\x1F\x00\xB1\x19"
-     "\x1F\x00"
-     "\xB0\x01\x1F\x00\xB7\x39\x1F\x00\xB6\x31\x1F\x00\xB5\x39\x1F\x00\xB4\x01"
-     "\x1F\x00"
-     "\xBB\x19\x1F\x00\xBA\x11\x1F\x00\xB9\x19\x1F\x00\xB8\x01\x1F\x00\x91\x00"
-     "\x00\x80"
-     "\xF5\x13\x00\x80\x9B\x00\x00\x80\xBC\x01\x1F\x00\x01\x14\x00\x80\xA5\x00"
-     "\x00\x80"
-     "\x0D\x14\x00\x80\xAF\x00\x00\x80\x19\x14\x00\x80\xB9\x00\x00\x80\x1E\x14"
-     "\x00\x80"
-     "\xB8\x0F\x00\x80\x1F\x10\x00\x80\x3E\x10\x00\x80\x81\x10\x00\x80\x84\x10"
-     "\x00\x80"
-     "\x14\x11\x00\x80\x3F\x11\x00\x80\x89\x11\x00\x80\xB5\x11\x00\x80\x90\xC1"
-     "\x1E\x00"
-     "\x91\xC1\x1E\x00\x92\xC1\x1E\x00\x93\xC1\x1E\x00\x94\x75\x80\x02\x95\xC5"
-     "\x1E\x00"
-     "\x96\xCD\x1E\x00\x97\xC5\x1E\x00\x98\xFD\x1E\x00\x99\x71\xA6\x00\xD1\x11"
-     "\x00\x80"
-     "\xD4\x11\x00\x80\xDC\x11\x00\x80\xF6\x11\x00\x80\xF9\x11\x00\x80\x5C\x12"
-     "\x00\x80"
-     "\xAA\x12\x00\x80\x19\x13\x00\x80\x54\x13\x00\x80\x80\x13\x00\x80\x88\x13"
-     "\x00\x80"
-     "\xC4\x13\x00\x80\x23\x14\x00\x80\x52\x14\x00\x80\x85\x14\x00\x80\xBE\x14"
-     "\x00\x80"
-     "\x37\x15\x00\x80\xB3\x15\x00\x80\x27\x16\x00\x80\xC1\x14\x00\x80\xC6\x14"
-     "\x00\x80"
-     "\x51\x16\x00\x80\xB0\xCD\xA6\x00\xB1\x65\x1E\x00\xB2\x6D\x1E\x00\xB3\x65"
-     "\x1E\x00"
-     "\xB4\x79\x1E\x00\xB5\x69\xA6\x00\xB6\x91\xAB\x00\xB7\x71\x1E\x00\xB8\x51"
-     "\x1E\x00"
-     "\xB9\x51\x1E\x00\xBA\x51\x1E\x00\xBB\x51\x1E\x00\xBC\xB1\x01\x00\xBD\xB1"
-     "\x01\x00"
-     "\xBE\xB1\x01\x00\xBF\x11\xAF\x00\xCB\x14\x00\x80\xD0\x14\x00\x80\xCF\x00"
-     "\x00\x80"
-     "\xD9\x00\x00\x80\x6A\x16\x00\x80\x4B\x03\x00\x80\x5D\x06\x00\x80\x91\x08"
-     "\x00\x80"
-     "\xD0\x21\x00\x80\xD4\x21\x00\x80\xAE\x2A\x00\x80\x62\x25\x00\x80\x9B\x25"
-     "\x00\x80"
-     "\xBA\x25\x00\x80\xCC\x25\x00\x80\xF1\x25\x00\x80\x83\x19\x02\x00\x82\x11"
-     "\x02\x00"
-     "\x81\x19\x02\x00\x80\xB5\xAC\x00\x87\x31\x02\x00\xF0\x29\x00\x80\xD9\xA8"
-     "\xB3\x02"
-     "\x84\x01\x02\x00\x8B\xAD\xA5\x00\x8A\xA9\xA5\x00\x89\xA5\xA5\x00\x88\x09"
-     "\x02\x00"
-     "\x8F\xD9\x1D\x00\x8E\xDD\x1D\x00\x8D\xD1\x1D\x00\x8C\xB1\xA5\x00\x0C\x2A"
-     "\x00\x80"
-     "\x92\x01\xA5\x00\x91\x8D\x1D\x00\x90\x8D\x1D\x00\x97\x81\x1D\x00\x96\x81"
-     "\x1D\x00"
-     "\x95\x95\x1D\x00\x94\x39\x83\x02\x9B\x8D\x1D\x00\x9A\x6D\xA8\x00\x99\x95"
-     "\xA5\x00"
-     "\x98\x8D\x1D\x00\x9F\x59\x02\x00\x9E\x8D\x1D\x00\x9D\x8D\x1D\x00\x9C\x8D"
-     "\x1D\x00"
-     "\xA3\xB5\x02\x00\xA2\x69\xA5\x00\xA1\xB9\x02\x00\xA0\xA1\x02\x00\x1E\x2A"
-     "\x00\x80"
-     "\xA6\xB9\xAE\x00\xA5\xA5\x02\x00\xA4\xA5\x02\x00\xAB\xCD\x02\x00\xAA\xCD"
-     "\x02\x00"
-     "\xA9\xCD\x02\x00\xA8\xCD\x02\x00\xB5\x27\x00\x80\xD9\x14\xB3\x02\xB2\x2A"
-     "\x00\x80"
-     "\x96\x2A\x00\x80\x56\x25\x00\x80\xB2\xBD\x02\x00\xB1\xBD\x02\x00\xB0\xBD"
-     "\x02\x00"
-     "\xB7\x65\x02\x00\xB6\x65\x02\x00\x5A\x25\x00\x80\xB4\xA1\x02\x00\xBB\x29"
-     "\x02\x00"
-     "\xBA\x21\x02\x00\xB9\x29\x02\x00\xB8\x55\x02\x00\xBF\x29\x02\x00\xBE\x21"
-     "\x02\x00"
-     "\xBD\x29\x02\x00\xBC\x31\x02\x00\x95\x08\x00\x80\x98\x08\x00\x80\xCC\xF8"
-     "\x91\x02"
-     "\xCC\xC4\x91\x02\xA1\x08\x00\x80\x9E\x08\x00\x80\x9B\x08\x00\x80\xB9\x20"
-     "\x00\x80"
-     "\xC4\x21\x00\x80\xC9\x20\x00\x80\xD1\x20\x00\x80\xC8\x21\x00\x80\xCC\x21"
-     "\x00\x80"
-     "\xDC\xEB\x00\x80\xE1\xEB\x00\x80\xC5\x20\x00\x80\x90\xC0\x02\x00\x91\x00"
-     "\x30\x00"
-     "\xB2\x48\x0E\x00\xE6\xEB\x00\x80\xEB\xEB\x00\x80\xF0\xEB\x00\x80\x96\xBC"
-     "\x30\x00"
-     "\xF5\xEB\x00\x80\xFA\xEB\x00\x80\xFF\xEB\x00\x80\x04\xEC\x00\x80\x09\xEC"
-     "\x00\x80"
-     "\x0E\xEC\x00\x80\x9D\x68\x32\x00\x9E\x60\x21\x00\x9F\x2C\x21\x00\x13\xEC"
-     "\x00\x80"
-     "\xD9\x04\xB2\x02\x18\xEC\x00\x80\x1D\xEC\x00\x80\x22\xEC\x00\x80\x27\xEC"
-     "\x00\x80"
-     "\x2C\xEC\x00\x80\x31\xEC\x00\x80\x36\xEC\x00\x80\x3B\xEC\x00\x80\x40\xEC"
-     "\x00\x80"
-     "\x45\xEC\x00\x80\x4A\xEC\x00\x80\x4F\xEC\x00\x80\x54\xEC\x00\x80\xAF\x7C"
-     "\x2A\x00"
-     "\x90\xC0\x0B\x00\x59\xEC\x00\x80\x92\xCC\x0B\x00\x93\xCC\x0A\x00\x5E\xEC"
-     "\x00\x80"
-     "\x95\xD0\x0A\x00\x96\x00\x0C\x00\x63\xEC\x00\x80\x68\xEC\x00\x80\x6D\xEC"
-     "\x00\x80"
-     "\x72\xEC\x00\x80\x77\xEC\x00\x80\x7C\xEC\x00\x80\x81\xEC\x00\x80\x9E\xDC"
-     "\x0F\x00"
-     "\x86\xEC\x00\x80\x9E\xF5\x04\x00\x9F\xF1\x04\x00\x9C\xF9\x04\x00\x9D\xF9"
-     "\x04\x00"
-     "\x9A\xE1\x04\x00\x9B\xE5\x04\x00\x98\xF5\x04\x00\x99\xED\x04\x00\x96\xCD"
-     "\x04\x00"
-     "\x97\xC9\x04\x00\x94\xC1\x04\x00\x95\xC1\x04\x00\x92\xD5\x04\x00\x93\xDD"
-     "\x04\x00"
-     "\x90\xDD\x04\x00\x91\xD9\x04\x00\x8E\xA1\x04\x00\x8F\xA1\x04\x00\x8C\xB1"
-     "\xB7\x02"
-     "\x8D\xAD\x04\x00\x8A\xA9\xB7\x02\x8B\xAD\xB7\x02\x88\x85\xB7\x02\x89\xA5"
-     "\xB7\x02"
-     "\x86\x89\xB7\x02\x87\xB9\xB7\x02\x84\xB1\xB7\x02\x85\x85\xB7\x02\x82\xB9"
-     "\xB7\x02"
-     "\x83\xAD\xB7\x02\x80\xA5\x04\x00\x81\xB5\xB7\x02\xBE\xC5\x05\x00\xBF\xC1"
-     "\x05\x00"
-     "\xBC\xC9\x05\x00\xBD\xC9\x05\x00\xBA\xD1\x05\x00\xBB\xD5\x05\x00\xB8\xA5"
-     "\x05\x00"
-     "\xB9\xDD\x05\x00\xB6\x9D\x05\x00\xB7\x99\x05\x00\xB4\x91\x05\x00\xB5\x91"
-     "\x05\x00"
-     "\xB2\x89\x05\x00\xB3\x8D\x05\x00\xB0\x8D\x05\x00\xB1\x85\x05\x00\xAE\xF1"
-     "\x05\x00"
-     "\xAF\xF1\x05\x00\xAC\xE5\x05\x00\xAD\xFD\x05\x00\xAA\xF9\x05\x00\xAB\xF9"
-     "\x05\x00"
-     "\xA8\x25\x04\x00\xA9\xF5\x05\x00\xA6\x1D\x04\x00\xA7\x19\x04\x00\xA4\x11"
-     "\x04\x00"
-     "\xA5\x11\x04\x00\xA2\x09\x04\x00\xA3\x0D\x04\x00\xA0\x0D\x04\x00\xA1\x05"
-     "\x04\x00"
-     "\x80\x45\x04\x00\x81\x4D\x04\x00\x82\x41\x04\x00\x83\x45\x04\x00\x84\x59"
-     "\x04\x00"
-     "\x85\x59\x04\x00\x86\x55\x04\x00\x87\x51\x04\x00\x88\x6D\x04\x00\x89\x75"
-     "\x04\x00"
-     "\x8A\x79\x04\x00\x8B\x7D\x04\x00\x8C\x61\x04\x00\x8D\x61\x04\x00\x8E\x6D"
-     "\x04\x00"
-     "\x8F\x69\x04\x00\x90\x15\x04\x00\x91\x1D\x04\x00\x92\x11\x04\x00\x93\x15"
-     "\x04\x00"
-     "\x94\x09\x04\x00\x95\x09\x04\x00\x96\x05\x04\x00\x97\x01\x04\x00\x98\x3D"
-     "\x04\x00"
-     "\x99\x05\x04\x00\x9A\x09\x04\x00\x9B\x0D\x04\x00\x9C\x11\x04\x00\x9D\x11"
-     "\x04\x00"
-     "\x9E\x1D\x04\x00\x9F\x19\x04\x00\xA0\xE5\x04\x00\xA1\xED\x04\x00\xA2\xE1"
-     "\x04\x00"
-     "\xA3\xE5\x04\x00\xA4\xF9\x04\x00\xA5\xF9\x04\x00\xA6\xF5\x04\x00\xA7\xF1"
-     "\x04\x00"
-     "\xA8\xCD\x04\x00\xA9\xD5\x04\x00\xAA\xD9\x04\x00\xAB\xDD\x04\x00\xAC\xC1"
-     "\x04\x00"
-     "\xAD\xC9\x04\x00\xAE\xC5\x04\x00\xAF\x61\x16\x00\xB0\x1D\x16\x00\xB1\x15"
-     "\x04\x00"
-     "\xB2\x19\x04\x00\xB3\x1D\x04\x00\xB4\x01\x04\x00\xB5\xA1\x04\x00\xB6\xAD"
-     "\x04\x00"
-     "\xB7\xA5\x04\x00\xB8\x99\x04\x00\xB9\x25\x04\x00\xBA\x29\x04\x00\xBB\x29"
-     "\x04\x00"
-     "\xBC\x35\x04\x00\x8B\xEC\x00\x80\x90\xEC\x00\x80\xBF\xF5\x95\x02\x81\xD1"
-     "\x19\x00"
-     "\x95\xEC\x00\x80\x83\x21\x06\x00\x82\xC5\x07\x00\x85\x79\x06\x00\x84\x71"
-     "\x06\x00"
-     "\x87\xD5\x07\x00\x86\x31\x06\x00\x89\x29\xAC\x00\x88\xD9\xA1\x00\x8B\xC5"
-     "\xA1\x00"
-     "\x8A\x1D\x06\x00\x8D\x05\x06\x00\x8C\xF5\x19\x00\x8F\xED\x07\x00\x8E\x79"
-     "\xA1\x00"
-     "\x91\x09\x9A\x02\x90\xF1\xA1\x00\x93\x81\x9D\x02\x92\xF1\x9D\x02\x95\x95"
-     "\xA1\x00"
-     "\x94\xE9\xA1\x00\x97\x8D\xA1\x00\x96\x91\xA1\x00\x99\x85\xA1\x00\x98\xB9"
-     "\xA1\x00"
-     "\x9B\x0D\x87\x02\x9A\xA1\x19\x00\x9D\x5D\xAA\x00\x9C\x4D\x06\x00\x9F\xAD"
-     "\x19\x00"
-     "\x9E\x41\x06\x00\xA1\xC1\x96\x02\xA0\xE9\x06\x00\xA3\x81\x92\x02\xA2\x89"
-     "\x92\x02"
-     "\xA5\xB5\x92\x02\xA4\x89\x92\x02\xA7\x19\x92\x02\xA6\xCD\x92\x02\xA9\x1D"
-     "\x92\x02"
-     "\xA8\x31\x92\x02\xAB\xF5\x93\x02\xAA\xE9\x93\x02\xAD\x15\x93\x02\xAC\x35"
-     "\x93\x02"
-     "\xAF\x55\x93\x02\xAE\x7D\x93\x02\xB1\x05\x91\x02\xB0\xFD\x9E\x02\xB3\x95"
-     "\x9E\x02"
-     "\xB2\x8D\x9E\x02\xB5\x39\x9E\x02\xB4\x75\x9E\x02\xB7\x01\x9B\x02\xB6\x25"
-     "\x9F\x02"
-     "\xB9\xE9\x9A\x02\xB8\xE5\x9A\x02\xBB\xC9\x06\x00\xBA\xF5\x9A\x02\xBD\xD5"
-     "\x06\x00"
-     "\xBC\x95\x06\x00\xBF\x11\x19\x00\xBE\x25\x07\x00\x80\x19\xA0\x00\x81\x45"
-     "\x97\x02"
-     "\x82\x0D\x93\x02\x83\x05\x93\x02\x84\x0D\x93\x02\x85\x31\x93\x02\x86\x49"
-     "\x93\x02"
-     "\x87\x9D\x93\x02\x88\xBD\x87\x02\x89\x35\xA3\x00\x8A\x85\x87\x02\x8B\x71"
-     "\x92\x02"
-     "\x8C\xB1\x92\x02\x8D\x91\x92\x02\x8E\xF9\x92\x02\x8F\xD1\x92\x02\x90\x79"
-     "\x9F\x02"
-     "\x91\x81\x90\x02\x92\x09\x9F\x02\x93\x11\x9F\x02\x94\x5D\xA3\x00\x95\xBD"
-     "\x9F\x02"
-     "\x96\xA1\x9E\x02\x97\xC5\x86\x02\x98\x61\x9B\x02\x99\xAD\x86\x02\x9A\x71"
-     "\x9B\x02"
-     "\x9B\x81\x18\x00\x9C\xAD\x06\x00\x9D\x85\x18\x00\x9A\xEC\x00\x80\x9F\x9D"
-     "\x06\x00"
-     "\xA0\x65\x06\x00\xA1\x4D\x18\x00\xA2\x99\x07\x00\xA3\x99\x07\x00\xA4\x55"
-     "\x18\x00"
-     "\xA5\x7D\x06\x00\xA6\x05\xF3\x00\xA7\x71\x06\x00\xA8\x51\x06\x00\xA9\x51"
-     "\x06\x00"
-     "\xAA\x51\x06\x00\xAB\x51\x06\x00\xAC\xA5\x05\x00\xAD\xF5\x05\x00\xAE\x29"
-     "\x05\x00"
-     "\xAF\x0D\x05\x00\xB0\x79\x05\x00\xB1\x3D\xF7\x00\xB2\x35\xF7\x00\xB3\x85"
-     "\xF2\x00"
-     "\xB4\xD1\xF5\x00\xB5\xD1\xF5\x00\xB6\xD9\xF3\x00\xB7\xD9\xF3\x00\xB8\xE9"
-     "\xF3\x00"
-     "\xB9\xE9\xF3\x00\xBA\xF9\xF3\x00\xBB\xF9\xF3\x00\xBC\xE9\xF3\x00\xBD\xE9"
-     "\xF3\x00"
-     "\xBE\xB9\xF3\x00\xBF\xB9\xF3\x00\x9A\x49\xFA\x00\x9B\x49\xFA\x00\x98\x49"
-     "\xFA\x00"
-     "\x99\x49\xFA\x00\x9E\xD1\x0A\x00\x9F\xCD\x0A\x00\x9C\x89\xFC\x00\x9D\xA1"
-     "\x0A\x00"
-     "\x92\x09\xFA\x00\x93\x09\xFA\x00\x90\x09\xFA\x00\x91\x09\xFA\x00\x96\x09"
-     "\xFA\x00"
-     "\x97\x09\xFA\x00\x94\x09\xFA\x00\x95\x09\xFA\x00\x8A\x1D\xFC\x00\x8B\xC5"
-     "\xFC\x00"
-     "\x88\x5D\xFA\x00\x89\x55\xFA\x00\x8E\xDD\xFC\x00\x8F\x49\xFA\x00\x8C\xDD"
-     "\xFC\x00"
-     "\x8D\xD5\xFC\x00\x82\x31\xFC\x00\x83\x31\xFC\x00\x80\x21\xFC\x00\x81\x21"
-     "\xFC\x00"
-     "\x86\x7D\xFA\x00\x87\x65\xFA\x00\x84\x21\xFC\x00\x85\x75\xFA\x00\xBA\x45"
-     "\xC0\x00"
-     "\xBB\x55\xC0\x00\xB8\x39\xC0\x00\xB9\x4D\xC0\x00\xBE\xB5\xC3\x00\xCD\x20"
-     "\x00\x80"
-     "\xBC\x4D\xC0\x00\xBD\xBD\xC3\x00\xB2\xF1\xC0\x00\xB3\xF1\xC0\x00\xB0\xDD"
-     "\xCF\x00"
-     "\xB1\xE9\xC0\x00\xB6\x21\xC0\x00\xB7\x21\xC0\x00\xB4\xC5\xCF\x00\xB5\x39"
-     "\xC0\x00"
-     "\xAA\x81\xCF\x00\xAB\x81\xCF\x00\xA8\xDD\xC0\x00\xA9\x8D\xC0\x00\xAE\xB5"
-     "\xCF\x00"
-     "\xAF\xA5\xCF\x00\xAC\x81\xCF\x00\xAD\xBD\xCF\x00\xA2\xE1\xC0\x00\xA3\x3D"
-     "\xCE\x00"
-     "\xA0\xC9\xC0\x00\xA1\xE9\xC3\x00\xA6\x01\xCE\x00\xA7\xE5\xC0\x00\xA4\xFD"
-     "\xC0\x00"
-     "\xA5\x21\xCE\x00\x82\x85\xCE\x00\x83\x89\xCE\x00\x9F\xEC\x00\x80\xA4\xEC"
-     "\x00\x80"
-     "\x86\xB5\xCE\x00\x87\xBD\xCE\x00\x84\x89\xCE\x00\x85\xBD\xCE\x00\x8A\x8D"
-     "\xCE\x00"
-     "\x8B\x95\xCE\x00\xA9\xEC\x00\x80\xAE\xEC\x00\x80\x8E\xF9\xCE\x00\x8F\xE1"
-     "\xCE\x00"
-     "\x8C\x81\xCE\x00\x8D\xF9\xCE\x00\x92\xA1\xCE\x00\x93\xA1\xCE\x00\xB3\xEC"
-     "\x00\x80"
-     "\xB8\xEC\x00\x80\x96\xAD\xCE\x00\x97\x5D\xC1\x00\x94\xB1\xCE\x00\x95\xB1"
-     "\xCE\x00"
-     "\x9A\x6D\xC1\x00\x9B\x7D\xC1\x00\xBD\xEC\x00\x80\xC2\xEC\x00\x80\xC7\xEC"
-     "\x00\x80"
-     "\xCC\xEC\x00\x80\x9C\x65\xC1\x00\xD1\xEC\x00\x80\xA2\x91\x0A\x00\xA3\x21"
-     "\x92\x02"
-     "\xA0\x89\x0A\x00\xA1\x89\x0A\x00\xA6\x8D\x0A\x00\xD6\xEC\x00\x80\xA4\x8D"
-     "\x0A\x00"
-     "\xA5\x85\x0A\x00\xAA\x85\x0A\x00\xAB\x31\x01\x00\xA8\x89\x0A\x00\xA9\x91"
-     "\x0A\x00"
-     "\xAE\x81\x0A\x00\xDB\xEC\x00\x80\xAC\x91\x0A\x00\xAD\x91\x0A\x00\xE0\xEC"
-     "\x00\x80"
-     "\xE5\xEC\x00\x80\xEA\xEC\x00\x80\xEF\xEC\x00\x80\xF4\xEC\x00\x80\xF9\xEC"
-     "\x00\x80"
-     "\xFE\xEC\x00\x80\x03\xED\x00\x80\x08\xED\x00\x80\x0D\xED\x00\x80\x12\xED"
-     "\x00\x80"
-     "\x17\xED\x00\x80\x1C\xED\x00\x80\x21\xED\x00\x80\x26\xED\x00\x80\xBD\xF5"
-     "\x99\x02"
-     "\x80\x35\x09\x00\x81\x3D\x09\x00\x82\x45\x09\x00\x83\x5D\x09\x00\x84\x45"
-     "\x09\x00"
-     "\x85\x49\x09\x00\x86\x79\x09\x00\x87\x79\x09\x00\x88\x49\x09\x00\x89\x49"
-     "\x09\x00"
-     "\x8A\x59\x09\x00\x8B\xAD\x09\x00\x8C\xB9\x09\x00\x8D\xB9\x09\x00\x8E\xA9"
-     "\x09\x00"
-     "\x8F\xA9\x09\x00\x90\xD9\x09\x00\x91\xD9\x09\x00\x92\xE9\x09\x00\x93\xE9"
-     "\x09\x00"
-     "\x94\xF9\x09\x00\x95\xF9\x09\x00\x96\xE9\x09\x00\x97\xE9\x09\x00\x98\xD9"
-     "\x09\x00"
-     "\x99\xA1\x09\x00\x9A\xB9\x09\x00\x9B\xB1\x09\x00\x9C\x91\x09\x00\x9D\x91"
-     "\x09\x00"
-     "\x9E\x91\x09\x00\x9F\x91\x09\x00\xA0\x71\x09\x00\xA1\x71\x09\x00\xA2\x75"
-     "\x09\x00"
-     "\xA3\x79\x0F\x00\xA4\x69\x0F\x00\xA5\x91\x0C\x00\xA6\xE1\x0D\x00\xA7\x95"
-     "\x03\x00"
-     "\x2B\xED\x00\x80\x30\xED\x00\x80\x35\xED\x00\x80\x3A\xED\x00\x80\x3F\xED"
-     "\x00\x80"
-     "\x44\xED\x00\x80\x49\xED\x00\x80\x4E\xED\x00\x80\xB1\xE9\x03\x00\xB0\xC9"
-     "\x03\x00"
-     "\xB3\x81\x03\x00\xB2\xE1\x03\x00\xB5\xB9\x03\x00\xB4\x99\x03\x00\xB7\xA9"
-     "\x03\x00"
-     "\xB6\xB1\x03\x00\xB9\x9D\x03\x00\xB8\x91\x03\x00\xBB\x5D\x03\x00\xBA\x79"
-     "\x03\x00"
-     "\xBD\x35\x03\x00\xBC\x35\x03\x00\xBF\x35\x03\x00\xBE\x35\x03\x00\x80\xD1"
-     "\x02\x00"
-     "\x81\xD1\x02\x00\x82\xD1\x02\x00\x83\xD1\x02\x00\x84\xF1\x02\x00\x85\xF1"
-     "\x02\x00"
-     "\x86\xF1\x02\x00\x87\xF1\x02\x00\x88\xD1\x02\x00\x89\xD1\x02\x00\x8A\xD1"
-     "\x02\x00"
-     "\x8B\x8D\x01\x00\x8C\x95\x01\x00\x8D\x9D\x01\x00\x8E\x95\x01\x00\x8F\x8D"
-     "\x01\x00"
-     "\x90\xF5\x01\x00\x91\xFD\x01\x00\x92\xF5\x01\x00\x93\x1D\x00\x00\x53\xED"
-     "\x00\x80"
-     "\x58\xED\x00\x80\x5D\xED\x00\x80\x62\xED\x00\x80\x67\xED\x00\x80\x6C\xED"
-     "\x00\x80"
-     "\x71\xED\x00\x80\x76\xED\x00\x80\x7B\xED\x00\x80\x80\xED\x00\x80\x85\xED"
-     "\x00\x80"
-     "\x8A\xED\x00\x80\x8F\xED\x00\x80\x94\xED\x00\x80\x99\xED\x00\x80\x9E\xED"
-     "\x00\x80"
-     "\xA3\xED\x00\x80\xA8\xED\x00\x80\xAD\xED\x00\x80\xB2\xED\x00\x80\xB7\xED"
-     "\x00\x80"
-     "\xCC\x38\xAB\x00\xBC\xED\x00\x80\xC1\xED\x00\x80\xC6\xED\x00\x80\xCB\xED"
-     "\x00\x80"
-     "\xD0\xED\x00\x80\xD5\xED\x00\x80\xB1\x99\x00\x00\xB0\x91\x00\x00\xB3\xA9"
-     "\x00\x00"
-     "\xB2\x91\x00\x00\xB5\xB9\x00\x00\xB4\xB1\x00\x00\xB7\xA9\x00\x00\xB6\xB1"
-     "\x00\x00"
-     "\xB9\x99\x00\x00\xB8\x91\x00\x00\xDA\xED\x00\x80\xBA\x91\x00\x00\xBD\x75"
-     "\x00\x00"
-     "\xBC\x75\x00\x00\xBF\x75\x00\x00\xBE\x75\x00\x00\xDF\xED\x00\x80\xE4\xED"
-     "\x00\x80"
-     "\xE9\xED\x00\x80\xEE\xED\x00\x80\xF3\xED\x00\x80\xF8\xED\x00\x80\xFD\xED"
-     "\x00\x80"
-     "\x02\xEE\x00\x80\xB9\x0B\x00\x80\x91\x2C\x03\x00\xBC\x0B\x00\x80\x7F\x0B"
-     "\x00\x80"
-     "\x97\x0B\x00\x80\x36\x3D\x00\x80\x9D\x0B\x00\x80\x9A\x0B\x00\x80\xBE\x1C"
-     "\x00\x80"
-     "\xDA\x0A\x00\x80\xE6\x0A\x00\x80\xA9\x0B\x00\x80\xF8\x0A\x00\x80\xF2\x0A"
-     "\x00\x80"
-     "\x82\x0B\x00\x80\x85\x0B\x00\x80\x01\x0B\x00\x80\x52\x0B\x00\x80\xAC\x0B"
-     "\x00\x80"
-     "\xAF\x0B\x00\x80\x20\xF1\x00\x80\xBA\x3C\x00\x80\x16\x0B\x00\x80\x25\xF1"
-     "\x00\x80"
-     "\x6A\x0B\x00\x80\x2A\xF1\x00\x80\x2E\x0B\x00\x80\x31\x0B\x00\x80\x3A\x0B"
-     "\x00\x80"
-     "\x2F\xF1\x00\x80\x43\x0B\x00\x80\xA6\x0B\x00\x80\xA0\x0B\x00\x80\x3A\x3D"
-     "\x00\x80"
-     "\xA3\x0B\x00\x80\xB6\x22\x00\x80\x6D\x0B\x00\x80\x91\x0B\x00\x80\x94\x0B"
-     "\x00\x80"
-     "\x7C\x0B\x00\x80\x08\x0A\x00\x80\x0B\x0A\x00\x80\x0E\x0A\x00\x80\x34\xF1"
-     "\x00\x80"
-     "\x39\xF1\x00\x80\x07\xEE\x00\x80\x0C\xEE\x00\x80\x11\xEE\x00\x80\x16\xEE"
-     "\x00\x80"
-     "\x1B\xEE\x00\x80\x20\xEE\x00\x80\x25\xEE\x00\x80\x2A\xEE\x00\x80\x2F\xEE"
-     "\x00\x80"
-     "\x34\xEE\x00\x80\x39\xEE\x00\x80\x82\x1C\x02\x00\x3E\xEE\x00\x80\x43\xEE"
-     "\x00\x80"
-     "\x48\xEE\x00\x80\x4D\xEE\x00\x80\x52\xEE\x00\x80\x57\xEE\x00\x80\x5C\xEE"
-     "\x00\x80"
-     "\x61\xEE\x00\x80\x66\xEE\x00\x80\x6B\xEE\x00\x80\x70\xEE\x00\x80\x75\xEE"
-     "\x00\x80"
-     "\x7A\xEE\x00\x80\x7F\xEE\x00\x80\x84\xEE\x00\x80\x89\xEE\x00\x80\x8E\xEE"
-     "\x00\x80"
-     "\x93\xEE\x00\x80\x98\xEE\x00\x80\x9D\xEE\x00\x80\xA2\xEE\x00\x80\xA7\xEE"
-     "\x00\x80"
-     "\xAC\xEE\x00\x80\xB1\xEE\x00\x80\xB6\xEE\x00\x80\xBB\xEE\x00\x80\xC0\xEE"
-     "\x00\x80"
-     "\xC5\xEE\x00\x80\xCA\xEE\x00\x80\xCF\xEE\x00\x80\xD4\xEE\x00\x80\xD9\xEE"
-     "\x00\x80"
-     "\xDE\xEE\x00\x80\x94\x91\x01\x00\x95\x91\x01\x00\x92\x85\x01\x00\xE3\xEE"
-     "\x00\x80"
-     "\x90\x95\x01\x00\x91\x8D\x01\x00\xE8\xEE\x00\x80\xED\xEE\x00\x80\xF2\xEE"
-     "\x00\x80"
-     "\xF7\xEE\x00\x80\xFC\xEE\x00\x80\x01\xEF\x00\x80\x06\xEF\x00\x80\x0B\xEF"
-     "\x00\x80"
-     "\x86\x6D\x06\x00\x87\x75\x06\x00\x84\x69\x06\x00\x85\x65\x06\x00\x82\x49"
-     "\x06\x00"
-     "\x83\x49\x06\x00\x80\x49\x06\x00\x81\x49\x06\x00\x8E\xE5\x01\x00\x8F\xED"
-     "\x01\x00"
-     "\x8C\x01\x06\x00\x8D\xED\x01\x00\x8A\x4D\x06\x00\xBA\xD1\x03\x00\x88\x4D"
-     "\x06\x00"
-     "\x89\x45\x06\x00\x89\x39\x02\x00\x88\x31\x02\x00\x8B\x69\x90\x02\x8A\x31"
-     "\x02\x00"
-     "\x8D\x51\x02\x00\x8C\x71\x90\x02\x8F\x51\x02\x00\x8E\x59\x02\x00\x81\x01"
-     "\x02\x00"
-     "\x10\xEF\x00\x80\x83\x35\x93\x02\x82\x05\x02\x00\x85\x09\x90\x02\x84\x19"
-     "\x02\x00"
-     "\x87\x09\x02\x00\xF0\xB4\x03\x00\x99\x29\x02\x00\x98\x21\x02\x00\x9B\xAD"
-     "\xF3\x00"
-     "\x9A\x21\x02\x00\x9D\x25\x02\x00\x9C\x25\x02\x00\x9F\x25\x02\x00\x9E\x25"
-     "\x02\x00"
-     "\x91\x31\x02\x00\x90\x29\x02\x00\x93\xED\x90\x02\x92\x39\x02\x00\x95\xB9"
-     "\x93\x02"
-     "\x94\x2D\x02\x00\x97\x19\x02\x00\x96\x21\x02\x00\xA9\x19\x02\x00\xA8\x75"
-     "\xDB\x00"
-     "\xAB\x15\x02\x00\xAA\x55\x90\x02\xAD\x35\x02\x00\xAC\x35\x02\x00\xAF\x95"
-     "\x91\x02"
-     "\xAE\x35\x02\x00\xA1\x25\x02\x00\xA0\x25\x02\x00\xA3\x21\x02\x00\xA2\xC1"
-     "\x93\x02"
-     "\xA5\x75\x99\x02\xA4\x39\x02\x00\xA7\x2D\x02\x00\xA6\x2D\x02\x00\xB9\x0D"
-     "\x02\x00"
-     "\xB8\x0D\x02\x00\x15\xEF\x00\x80\xBA\x1D\x02\x00\x1A\xEF\x00\x80\x1F\xEF"
-     "\x00\x80"
-     "\x24\xEF\x00\x80\x2E\xEF\x00\x80\x29\xEF\x00\x80\xB0\x49\x02\x00\xB3\x5D"
-     "\x02\x00"
-     "\xB2\x5D\x02\x00\xB5\x4D\x02\x00\xB4\x4D\x02\x00\xB7\x3D\x02\x00\xB6\x3D"
-     "\x02\x00"
-     "\x80\xD5\x03\x00\x81\xDD\x03\x00\x82\xD5\x03\x00\x83\xED\x03\x00\x84\xF5"
-     "\x03\x00"
-     "\x85\xFD\x03\x00\x86\xF5\x03\x00\x87\xED\x03\x00\x88\xD5\x03\x00\x89\xDD"
-     "\x03\x00"
-     "\x8A\xD5\x03\x00\x8B\x29\x02\x00\x8C\x39\x02\x00\x8D\x39\x02\x00\x8E\x29"
-     "\x02\x00"
-     "\x8F\x29\x02\x00\x90\x59\x02\x00\x91\x59\x02\x00\x92\x69\x02\x00\x93\x69"
-     "\x02\x00"
-     "\x94\x79\x02\x00\x95\x79\x02\x00\x96\x69\x02\x00\x97\x69\x02\x00\x98\x59"
-     "\x02\x00"
-     "\x99\x59\x02\x00\x9A\x29\x02\x00\x9B\x29\x02\x00\x9C\x39\x02\x00\x9D\x39"
-     "\x02\x00"
-     "\x9E\x29\x02\x00\x9F\x29\x02\x00\xA0\xD9\x02\x00\xA1\xD9\x02\x00\xA2\xE9"
-     "\x02\x00"
-     "\xA3\xE9\x02\x00\xA4\xF9\x02\x00\xA5\xF9\x02\x00\xA6\xE9\x02\x00\xA7\xE9"
-     "\x02\x00"
-     "\xA8\xD9\x02\x00\xA9\xD9\x02\x00\xAA\x29\x02\x00\xAB\x29\x02\x00\xAC\x25"
-     "\x02\x00"
-     "\xAD\x15\x02\x00\xAE\x1D\x02\x00\xAF\x15\x02\x00\xB0\x6D\x02\x00\xB1\x75"
-     "\x02\x00"
-     "\xB2\x7D\x02\x00\x91\x0C\x02\x00\x84\x48\x3D\x00\x91\x1C\x02\x00\x84\x50"
-     "\x3D\x00"
-     "\x91\x04\x02\x00\x33\xEF\x00\x80\x87\x3C\x3C\x00\x8C\x30\x3D\x00\x8D\x94"
-     "\x3D\x00"
-     "\x38\xEF\x00\x80\x3D\xEF\x00\x80\x42\xEF\x00\x80\x47\xEF\x00\x80\x99\xC0"
-     "\xD5\x00"
-     "\xA7\xCD\x3E\x00\x9B\xC8\x28\x00\xF0\x9C\x3D\x00\xB1\xEC\x3D\x00\xA7\xC5"
-     "\x3E\x00"
-     "\xF0\x80\x3D\x00\xB2\xD0\x03\x00\xBE\xF5\x3E\x00\x97\xFD\x3E\x00\xB9\xC4"
-     "\x03\x00"
-     "\xF0\x58\x02\x00\x91\x4C\x02\x00\x92\xB4\x02\x00\xB9\xAC\x03\x00\xB8\xB8"
-     "\x03\x00"
-     "\xB0\x9D\x3E\x00\xF0\x48\x02\x00\x91\x5C\x02\x00\x96\xA0\x02\x00\xAF\xA1"
-     "\x3F\x00"
-     "\xF0\x48\x02\x00\x91\x5C\x02\x00\x96\x8C\x02\x00\x4C\xEF\x00\x80\x51\xEF"
-     "\x00\x80"
-     "\xBA\x89\x3E\x00\xAF\x99\x3F\x00\x56\xEF\x00\x80\xBD\x95\x3F\x00\x5B\xEF"
-     "\x00\x80"
-     "\x60\xEF\x00\x80\xB5\xA8\x03\x00\x65\xEF\x00\x80\x6A\xEF\x00\x80\x6F\xEF"
-     "\x00\x80"
-     "\x74\xEF\x00\x80\x79\xEF\x00\x80\x7E\xEF\x00\x80\x83\xEF\x00\x80\x88\xEF"
-     "\x00\x80"
-     "\x8D\xEF\x00\x80\x92\xEF\x00\x80\x97\xEF\x00\x80\x9C\xEF\x00\x80\xA1\xEF"
-     "\x00\x80"
-     "\xA6\xEF\x00\x80\xAB\xEF\x00\x80\xB0\xEF\x00\x80\xB5\xEF\x00\x80\xBA\xEF"
-     "\x00\x80"
-     "\xBF\xEF\x00\x80\xC4\xEF\x00\x80\xC9\xEF\x00\x80\xCE\xEF\x00\x80\xD3\xEF"
-     "\x00\x80"
-     "\xD8\xEF\x00\x80\xDD\xEF\x00\x80\xF0\x40\x02\x00\x91\x84\x02\x00\xA5\x70"
-     "\x2C\x00"
-     "\xB0\x09\x00\x00\xA4\xCC\x02\x00\xE2\xEF\x00\x80\xB9\xE8\x01\x00\xE7\xEF"
-     "\x00\x80"
-     "\xEC\xEF\x00\x80\x92\xC0\x03\x00\xF1\xEF\x00\x80\xF6\xEF\x00\x80\xFB\xEF"
-     "\x00\x80"
-     "\x96\xC0\x03\x00\x00\xF0\x00\x80\x05\xF0\x00\x80\x0A\xF0\x00\x80\x0F\xF0"
-     "\x00\x80"
-     "\x14\xF0\x00\x80\x19\xF0\x00\x80\x1E\xF0\x00\x80\x23\xF0\x00\x80\x28\xF0"
-     "\x00\x80"
-     "\x2D\xF0\x00\x80\x32\xF0\x00\x80\x82\x28\x03\x00\x37\xF0\x00\x80\x84\x80"
-     "\x03\x00"
-     "\x3C\xF0\x00\x80\x41\xF0\x00\x80\x46\xF0\x00\x80\x4B\xF0\x00\x80\x50\xF0"
-     "\x00\x80"
-     "\x55\xF0\x00\x80\x8D\x88\x3C\x00\x5A\xF0\x00\x80\x5F\xF0\x00\x80\x64\xF0"
-     "\x00\x80"
-     "\x69\xF0\x00\x80\x6E\xF0\x00\x80\x73\xF0\x00\x80\x78\xF0\x00\x80\x7D\xF0"
-     "\x00\x80"
-     "\x82\xF0\x00\x80\x87\xF0\x00\x80\x90\xF0\x00\x80\x99\xF0\x00\x80\xA6\xF0"
-     "\x00\x80"
-     "\xB3\xF0\x00\x80\xC0\xF0\x00\x80\xCD\xF0\x00\x80\xDA\xF0\x00\x80\xFD\xF0"
-     "\x00\x80"
-     "\xE3\xF0\x00\x80\x06\xF1\x00\x80\xF0\xF0\x00\x80\x80\x19\x00\x00\xA2\x4C"
-     "\x01\x00"
-     "\x13\xF1\x00\x80\xA4\x80\x03\x00\x5A\x09\x00\x80\xA4\x0A\x00\x80\xBB\x55"
-     "\x00\x00"
-     "\xF5\x20\x00\x80\xED\x20\x00\x80\x3E\xF1\x00\x80\xBF\x35\x00\x00\xBE\x35"
-     "\x00\x00"
-     "\xBD\x45\x00\x00\xBC\x45\x00\x00\xA0\xF9\x3D\x00\xA1\xF9\x3D\x00\xA2\x69"
-     "\x02\x00"
-     "\xA3\x69\x02\x00\xA4\x61\x02\x00\xA5\x6D\x02\x00\xA6\x65\x02\x00\xA7\x99"
-     "\x02\x00"
-     "\xA8\xA9\x02\x00\xA9\xA9\x02\x00\xAA\xB9\x02\x00\xAB\xB9\x02\x00\xAC\xA9"
-     "\x02\x00"
-     "\xAD\xA9\x02\x00\xAE\x99\x02\x00\xAF\x99\x02\x00\xB0\xE9\x02\x00\xB1\xE9"
-     "\x02\x00"
-     "\xB2\xF9\x02\x00\xB3\xF9\x02\x00\xB4\xE9\x02\x00\xB5\xE9\x02\x00\xB6\x99"
-     "\x02\x00"
-     "\xB7\x99\x02\x00\xB8\xA9\x02\x00\xB9\xA9\x02\x00\xBA\xB9\x02\x00\xBB\xB9"
-     "\x02\x00"
-     "\xBC\xA9\x02\x00\xBD\xA9\x02\x00\xBE\x99\x02\x00\xBF\x99\x02\x00\x9A\x01"
-     "\x03\x00"
-     "\x9B\x01\x03\x00\x98\xF1\x03\x00\x99\xF1\x03\x00\x9E\x01\x03\x00\x9F\x01"
-     "\x03\x00"
-     "\x9C\x11\x03\x00\x9D\x11\x03\x00\x92\xFD\x03\x00\x93\xC5\x03\x00\x90\xF9"
-     "\x03\x00"
-     "\x91\xF5\x03\x00\x96\xDD\x03\x00\x97\xC1\x03\x00\x94\xDD\x03\x00\x95\xD5"
-     "\x03\x00"
-     "\x8A\xF5\x03\x00\x8B\xFD\x03\x00\x88\xE5\x03\x00\x89\xFD\x03\x00\x8E\x95"
-     "\x03\x00"
-     "\x8F\x9D\x03\x00\x8C\xE5\x03\x00\x8D\x9D\x03\x00\x82\xF1\x03\x00\x83\xF1"
-     "\x03\x00"
-     "\x80\xED\x03\x00\x81\xE5\x03\x00\x86\xD5\x03\x00\x87\xDD\x03\x00\x84\xE1"
-     "\x03\x00"
-     "\x85\xDD\x03\x00\x87\xC8\x00\x00\x86\x08\x00\x00\x85\x5C\x02\x00\xC3\x00"
-     "\x00\x00"
-     "\x43\xF1\x00\x80\x48\xF1\x00\x80\x4D\xF1\x00\x80\x52\xF1\x00\x80\x57\xF1"
-     "\x00\x80"
-     "\x5C\xF1\x00\x80\x9F\xB5\x02\x00\x9E\xB5\x02\x00\x61\xF1\x00\x80\x66\xF1"
-     "\x00\x80"
-     "\x6B\xF1\x00\x80\x70\xF1\x00\x80\x97\x40\x0A\x00\x96\x44\x0B\x00\x95\x48"
-     "\x09\x00"
-     "\x94\x00\x08\x00\x93\x50\x06\x00\x92\x54\x07\x00\x91\x58\x05\x00\x90\x00"
-     "\x04\x00"
-     "\x9F\x60\x15\x00\x9E\x00\x14\x00\x9D\x6C\x13\x00\x9C\x6C\x10\x00\x9B\x70"
-     "\x0E\x00"
-     "\x9A\x74\x0F\x00\x99\x78\x0D\x00\x98\x00\x0C\x00\xB3\x0D\xBE\x02\xB2\x0D"
-     "\xBE\x02"
-     "\xB1\x09\xBE\x02\xB0\xD5\xBE\x02\xB7\x25\xBE\x02\xB6\x35\xBE\x02\xB5\x0D"
-     "\xBE\x02"
-     "\xB4\x0D\xBE\x02\xBB\xB9\xBD\x02\xBA\xB5\xBD\x02\xB9\x25\x83\x02\xB8\x61"
-     "\xBE\x02"
-     "\xBF\xF5\xBD\x02\xBE\x8D\xBD\x02\xBD\x89\xBD\x02\xBC\xBD\x83\x02\x75\xF1"
-     "\x00\x80"
-     "\xA4\x75\x02\x00\xF2\x00\x00\x00\xF3\x00\x00\x00\xA1\x75\x02\x00\xA0\x75"
-     "\x02\x00"
-     "\xA3\x75\x02\x00\xA2\x75\x02\x00\x7A\xF1\x00\x80\xF9\x00\x00\x00\x7F\xF1"
-     "\x00\x80"
-     "\xFB\x00\x00\x00\x84\xF1\x00\x80\x89\xF1\x00\x80\x8E\xF1\x00\x80\x93\xF1"
-     "\x00\x80"
-     "\x80\xA5\xA1\x02\x81\xC5\xA1\x02\x82\xE5\xA1\x02\x83\x35\xAC\x02\x84\xD5"
-     "\xA3\x02"
-     "\x85\x55\xAC\x02\x86\x45\xAC\x02\x87\xBD\xAC\x02\x88\xC9\xAC\x02\x89\xDD"
-     "\xAD\x02"
-     "\x8A\xF1\xA9\x02\x8B\x2D\xA8\x02\x8C\x39\xA8\x02\x8D\x2D\xA8\x02\x8E\x79"
-     "\xA4\x02"
-     "\x8F\x39\xA0\x02\x90\x49\xA0\x02\x91\x59\xA0\x02\x92\x6D\xA0\x02\x93\x1D"
-     "\xA0\x02"
-     "\x94\xD1\xA0\x02\x98\xF1\x00\x80\x96\x49\x90\x00\x97\xF1\xB4\x02\x98\x3D"
-     "\xA1\x02"
-     "\x99\xE5\xA1\x02\x9A\xCD\xA1\x02\x9B\xAD\xA1\x02\x9C\x9D\xA1\x02\x9D\x4D"
-     "\xAC\x02"
-     "\x9E\xBD\xA3\x02\x9F\x3D\xAC\x02\xA0\xDD\xAC\x02\xA1\x89\x90\x00\xA2\x61"
-     "\xAC\x02"
-     "\xA3\x75\xAD\x02\xA4\x09\xB5\x02\xA5\x95\xA8\x02\xA6\x51\xB5\x02\xA7\x85"
-     "\xA8\x02"
-     "\xA8\xE1\xA4\x02\xA9\xA1\xA0\x02\xAA\xA1\xA0\x02\xAB\xB1\xA0\x02\xAC\x95"
-     "\xA0\x02"
-     "\xAD\xE5\xA0\x02\xAE\x39\xA0\x02\xAF\x29\xA0\x02\xB0\x7D\xA0\x02\xB1\x81"
-     "\xA1\x02"
-     "\xB2\x95\xA1\x02\xB3\x4D\xA1\x02\xB4\x75\xA1\x02\xB5\x15\xA1\x02\xB6\x35"
-     "\xA1\x02"
-     "\xB7\xE5\xAC\x02\xB8\x25\xA3\x02\xB9\xA5\xAC\x02\xBA\xB5\xAC\x02\xBB\x4D"
-     "\xAC\x02"
-     "\xBC\x19\xAC\x02\xBD\x0D\xAD\x02\xBE\x21\xA9\x02\xBF\xFD\xA8\x02\x97\x01"
-     "\xA9\x02"
-     "\x96\xDD\xA8\x02\x95\xF1\xAC\x02\x94\xE5\xAD\x02\x93\xB1\xAD\x02\x92\x49"
-     "\xAD\x02"
-     "\x91\x59\xAD\x02\x90\xD9\xA2\x02\x9F\x71\xA1\x02\x9E\x01\xA1\x02\x9D\x35"
-     "\xA1\x02"
-     "\x9C\x25\xA1\x02\x9B\x35\xA1\x02\x9A\x75\xA5\x02\x99\x21\xA9\x02\x98\x35"
-     "\xA9\x02"
-     "\x87\xD5\xA1\x02\x86\xC5\xA1\x02\x85\x19\xA1\x02\x84\x69\xA1\x02\x83\x4D"
-     "\xA1\x02"
-     "\x82\x5D\xA1\x02\x81\x5D\xA1\x02\x80\x1D\xA5\x02\x8F\x59\xAD\x02\x8E\x89"
-     "\xA0\x02"
-     "\x8D\xA9\xA0\x02\x8C\xC9\xA0\x02\x8B\xF1\xA0\x02\x8A\x29\xA0\x02\x89\x3D"
-     "\xA0\x02"
-     "\x88\xC1\xA1\x02\xB7\x9D\xA1\x02\xB6\x8D\xA1\x02\xB5\x8D\xA1\x02\xB4\xCD"
-     "\xA5\x02"
-     "\xB3\x89\xA9\x02\xB2\x5D\xB4\x02\xB1\x99\xA9\x02\xB0\x05\xB4\x02\xBF\x21"
-     "\xA0\x02"
-     "\xBE\xF9\xA0\x02\xBD\xED\xA0\x02\xBC\x11\xA1\x02\xBB\x25\xA1\x02\xBA\x35"
-     "\xA1\x02"
-     "\xB9\xE9\xA1\x02\xB8\x99\xA1\x02\xA7\x01\xA0\x02\xA6\x61\xA0\x02\xA5\x49"
-     "\xA0\x02"
-     "\xA4\x91\xA0\x02\xA3\x7D\xB5\x02\xA2\xC5\x91\x00\xA1\x45\xB5\x02\xA0\x5D"
-     "\xA1\x02"
-     "\xAF\x19\xAC\x02\xAE\x0D\xAD\x02\xAD\xE5\x91\x00\xAC\xB1\xAD\x02\xAB\xB1"
-     "\xAD\x02"
-     "\xAA\x31\xA2\x02\xA9\xC1\xAD\x02\xA8\x11\xA0\x02\x80\x09\xAA\x02\x81\x1D"
-     "\xAA\x02"
-     "\x82\x49\xA6\x02\x83\x09\xA2\x02\x84\x19\xA2\x02\x85\x09\xA2\x02\x86\x3D"
-     "\xA2\x02"
-     "\x87\x4D\xA2\x02\x88\xA1\xA2\x02\x89\xB9\xB6\x02\x8A\x39\x92\x00\x8B\x81"
-     "\xB6\x02"
-     "\x8C\x6D\xA3\x02\x8D\xB5\xA3\x02\x8E\x9D\xA3\x02\x8F\xFD\xA3\x02\x90\xAD"
-     "\xA3\x02"
-     "\x91\x7D\xAE\x02\x92\x8D\xA1\x02\x93\x0D\xAE\x02\x94\x0D\xAE\x02\x95\x59"
-     "\x92\x00"
-     "\x96\xB1\xAE\x02\x97\xA5\xAF\x02\x98\xF9\xB7\x02\x99\x65\xAA\x02\x9A\xA1"
-     "\xB7\x02"
-     "\x9B\x75\xAA\x02\x9C\x31\xA6\x02\x9D\xF1\x00\x80\x9E\x71\xA2\x02\x9F\x61"
-     "\xA2\x02"
-     "\xA2\xF1\x00\x80\xA7\xF1\x00\x80\xA2\x09\xA2\x02\xAC\xF1\x00\x80\xB1\xF1"
-     "\x00\x80"
-     "\xA5\xD1\xA3\x02\xA6\xC5\xA3\x02\xB6\xF1\x00\x80\xBB\xF1\x00\x80\xA9\x65"
-     "\xA3\x02"
-     "\xAA\x45\xA3\x02\xAB\x95\xAE\x02\xAC\x75\xA1\x02\xC0\xF1\x00\x80\xAE\xE5"
-     "\xAE\x02"
-     "\xAF\x1D\xAE\x02\xB0\x29\xAE\x02\xB1\x3D\xAF\x02\xB2\x11\xAB\x02\xB3\xCD"
-     "\xAA\x02"
-     "\xB4\xD9\xAA\x02\xB5\xCD\xAA\x02\xB6\x99\xA6\x02\xB7\xD9\xA2\x02\xB8\xE9"
-     "\xA2\x02"
-     "\xB9\xF9\xA2\x02\xC5\xF1\x00\x80\xBB\xBD\xA2\x02\xCA\xF1\x00\x80\xBD\x69"
-     "\xB6\x02"
-     "\xBE\xE9\x92\x00\xBF\x51\xB6\x02\x80\x5D\xA2\x02\x81\x85\xA2\x02\x82\xAD"
-     "\xA2\x02"
-     "\x83\xCD\xA2\x02\xCF\xF1\x00\x80\x85\x2D\xAF\x02\x86\xDD\xA0\x02\x87\x5D"
-     "\xAF\x02"
-     "\x88\x7D\xAF\x02\x89\x29\x93\x00\x8A\xC1\xAF\x02\x8B\xD5\xAE\x02\x8C\xA9"
-     "\xB6\x02"
-     "\x8D\x35\xAB\x02\x8E\xF1\xB6\x02\x8F\x25\xAB\x02\x90\x01\xA7\x02\x91\x41"
-     "\xA3\x02"
-     "\x92\x41\xA3\x02\x93\x51\xA3\x02\x94\x75\xA3\x02\x95\x05\xA3\x02\x96\xD9"
-     "\xA3\x02"
-     "\x97\xC9\xA3\x02\x98\xDD\xA3\x02\x99\x21\xA2\x02\x9A\x35\xA2\x02\x9B\xED"
-     "\xA2\x02"
-     "\x9C\xD5\xA2\x02\x9D\xB5\xA2\x02\x9E\x95\xA2\x02\x9F\x45\xAF\x02\xA0\x45"
-     "\xA0\x02"
-     "\xA1\xC5\xAF\x02\xA2\xD5\xAF\x02\xA3\x2D\xAF\x02\xA4\x79\xAF\x02\xA5\x6D"
-     "\xAE\x02"
-     "\xA6\x41\xAA\x02\xA7\x9D\xAB\x02\xA8\xA9\xAB\x02\xA9\xBD\xAB\x02\xAA\xE9"
-     "\xA7\x02"
-     "\xAB\xA9\xA3\x02\xAC\xB9\xA3\x02\xAD\xA9\xA3\x02\xAE\x9D\xA3\x02\xAF\xED"
-     "\xA3\x02"
-     "\xB0\x41\xA3\x02\xB1\x59\xB7\x02\xB2\xD9\x93\x00\xB3\x61\xB7\x02\xB4\x8D"
-     "\xA2\x02"
-     "\xB5\x55\xA2\x02\xB6\x7D\xA2\x02\xB7\x1D\xA2\x02\xB8\x0D\xA2\x02\xB9\xDD"
-     "\xAF\x02"
-     "\xBA\x2D\xA0\x02\xBB\xAD\xAF\x02\xBC\xAD\xAF\x02\xBD\xF9\x93\x00\xBE\x11"
-     "\xAF\x02"
-     "\xBF\x05\xAE\x02\x80\xBD\xAC\x02\x81\x41\xAD\x02\x82\x55\xAD\x02\x83\x8D"
-     "\xAD\x02"
-     "\x84\xB5\xAD\x02\xD4\xF1\x00\x80\x86\xF5\xAD\x02\xD9\xF1\x00\x80\xDE\xF1"
-     "\x00\x80"
-     "\xE3\xF1\x00\x80\x8A\x75\xA0\x02\x8B\x8D\xA0\x02\x8C\xD9\xA0\x02\x8D\xCD"
-     "\xA1\x02"
-     "\x8E\xE1\xA5\x02\x8F\x3D\xA4\x02\x90\x49\xA4\x02\xC0\x16\x00\x80\x92\x09"
-     "\xA8\x02"
-     "\x93\x49\xAC\x02\x94\x59\xAC\x02\x95\x49\xAC\x02\x96\x7D\xAC\x02\x97\x0D"
-     "\xAC\x02"
-     "\x98\xE1\xAC\x02\x99\xF9\xB8\x02\x9A\x79\x9C\x00\x9B\xC1\xB8\x02\x9C\x2D"
-     "\xAD\x02"
-     "\x9D\xF5\xAD\x02\x9E\xDD\xAD\x02\x9F\xBD\xAD\x02\xA0\x6D\xAD\x02\xA1\xBD"
-     "\xA0\x02"
-     "\xA2\x4D\xAF\x02\xA3\xCD\xA0\x02\xA4\xCD\xA0\x02\xA5\x99\x9C\x00\xA6\x71"
-     "\xA0\x02"
-     "\xA7\x65\xA1\x02\xA8\x39\xB9\x02\xA9\xA5\xA4\x02\xAA\x61\xB9\x02\xAB\xB5"
-     "\xA4\x02"
-     "\xAC\xF1\xA8\x02\xAD\xB1\xAC\x02\xAE\xB1\xAC\x02\xAF\xA1\xAC\x02\xB0\xE5"
-     "\xAC\x02"
-     "\xB1\x95\xAC\x02\xB2\x49\xAC\x02\xB3\x59\xAC\x02\xB4\x6D\xAC\x02\xB5\x91"
-     "\xAD\x02"
-     "\xB6\x85\xAD\x02\xB7\x5D\xAD\x02\xB8\x45\xAD\x02\xB9\x25\xAD\x02\xBA\x05"
-     "\xAD\x02"
-     "\xBB\xD5\xA0\x02\xBC\x35\xAF\x02\xBD\xB5\xA0\x02\xBE\xA5\xA0\x02\xBF\x5D"
-     "\xA0\x02"
-     "\x93\x39\xA1\x02\x92\xE9\xAC\x02\x91\xC9\xAC\x02\x90\xA9\xAC\x02\x97\xB1"
-     "\xA1\x02"
-     "\x96\x49\xA1\x02\xF9\x16\x00\x80\x94\xD9\xAE\x02\x9B\x21\xA5\x02\x9A\xFD"
-     "\xA4\x02"
-     "\x99\xD1\xA0\x02\x98\xC5\xA1\x02\x9F\x35\xAD\x02\x9E\x75\xA9\x02\xC3\x16"
-     "\x00\x80"
-     "\x9C\x35\xA5\x02\x83\x59\xA5\x02\x82\x8D\xB8\x02\x81\x49\xA5\x02\x80\xD5"
-     "\xB8\x02"
-     "\x87\x4D\xAD\x02\x52\x03\x00\x80\x85\x5D\xAD\x02\x84\x1D\xA9\x02\x4F\x03"
-     "\x00\x80"
-     "\x8A\xE5\xAD\x02\x89\x39\xAD\x02\x88\x49\xAD\x02\x8F\xF1\xAC\x02\x8E\x29"
-     "\xAC\x02"
-     "\x8D\x3D\xAC\x02\x62\x03\x00\x80\xB3\x79\xA0\x02\xB2\x6D\xA1\x02\xB1\x85"
-     "\x9D\x00"
-     "\xB0\xD1\xA1\x02\xB7\x89\xA5\x02\xB6\x5D\xB8\x02\xB5\x99\xA5\x02\xB4\x05"
-     "\xB8\x02"
-     "\xBB\xBD\xAD\x02\x85\x04\x00\x80\xB9\xAD\xAD\x02\xB8\xED\xA9\x02\x24\x05"
-     "\x00\x80"
-     "\xBE\x35\xAD\x02\xBD\xE9\xAD\x02\xBC\x99\xAD\x02\xA3\x91\xAD\x02\xA2\xE1"
-     "\xAD\x02"
-     "\xA1\xD5\xAD\x02\xA0\xC5\xAD\x02\xA7\x7D\xB9\x02\xA6\xC5\x9D\x00\xA5\x45"
-     "\xB9\x02"
-     "\xA4\x5D\xAD\x02\xAB\x21\xAC\x02\xAA\x41\xAC\x02\xA9\x69\xAC\x02\xA8\xB1"
-     "\xAC\x02"
-     "\xAF\xB1\xA1\x02\xAE\x31\xAE\x02\xAD\xC1\xA1\x02\xAC\x11\xAC\x02\x80\xE9"
-     "\xA2\x02"
-     "\x81\xFD\xA3\x02\x82\xD1\xA7\x02\x83\x0D\xA6\x02\x84\x19\xA6\x02\x85\x0D"
-     "\xA6\x02"
-     "\x86\x59\xAA\x02\x87\x19\xAE\x02\x88\x29\xAE\x02\x89\x39\xAE\x02\x8A\x0D"
-     "\xAE\x02"
-     "\x8B\x7D\xAE\x02\x8C\xB1\xAE\x02\x8D\xA9\xBA\x02\x8E\x29\x9E\x00\x8F\x91"
-     "\xBA\x02"
-     "\x90\x1D\xAF\x02\x91\xC5\xAF\x02\x92\xED\xAF\x02\x93\x8D\xAF\x02\x94\xBD"
-     "\xAF\x02"
-     "\x95\x6D\xA2\x02\x96\x9D\xAD\x02\x97\x1D\xA2\x02\x98\x3D\xA2\x02\x99\x69"
-     "\x9E\x00"
-     "\x9A\x81\xA2\x02\x9B\x95\xA3\x02\x9C\xE9\xBB\x02\x9D\x75\xA6\x02\x9E\xB1"
-     "\xBB\x02"
-     "\x9F\x65\xA6\x02\xA0\xC1\xAA\x02\xA1\x81\xAE\x02\xA2\x81\xAE\x02\xA3\x91"
-     "\xAE\x02"
-     "\xA4\xB5\xAE\x02\xA5\xC5\xAE\x02\xA6\x19\xAE\x02\xA7\x09\xAE\x02\xA8\x1D"
-     "\xAE\x02"
-     "\xA9\xE1\xAF\x02\xAA\xF5\xAF\x02\xAB\x2D\xAF\x02\xAC\x15\xAF\x02\xAD\x75"
-     "\xAF\x02"
-     "\xAE\x55\xAF\x02\xAF\x85\xA2\x02\xB0\x05\xAD\x02\xB1\x85\xA2\x02\xB2\x95"
-     "\xA2\x02"
-     "\xB3\x6D\xA2\x02\xB4\x39\xA2\x02\xB5\x2D\xA3\x02\xB6\x01\xA7\x02\xB7\xDD"
-     "\xA6\x02"
-     "\xB8\xE9\xA6\x02\xB9\xFD\xA6\x02\xBA\xA9\xAA\x02\xBB\xE9\xAE\x02\xBC\xF9"
-     "\xAE\x02"
-     "\xBD\xE9\xAE\x02\xBE\xDD\xAE\x02\xBF\xAD\xAE\x02\x80\x81\xAF\x02\x81\x99"
-     "\xBB\x02"
-     "\x82\x19\x9F\x00\x83\xA1\xBB\x02\x84\x4D\xAE\x02\x85\x95\xAE\x02\x86\xBD"
-     "\xAE\x02"
-     "\x87\xDD\xAE\x02\x88\xCD\xAE\x02\x89\x1D\xA3\x02\x8A\xED\xAC\x02\x8B\x6D"
-     "\xA3\x02"
-     "\x8C\x6D\xA3\x02\x8D\x39\x9F\x00\x8E\xD1\xA3\x02\x8F\xC5\xA2\x02\x90\xD9"
-     "\xBA\x02"
-     "\x91\x45\xA7\x02\x92\x81\xBA\x02\x93\x55\xA7\x02\x94\x11\xAB\x02\x95\x51"
-     "\xAF\x02"
-     "\x96\x51\xAF\x02\x97\x41\xAF\x02\x98\x45\xAF\x02\x99\x35\xAF\x02\x9A\xE9"
-     "\xAF\x02"
-     "\x9B\xF9\xAF\x02\x9C\xCD\xAF\x02\x9D\x31\xAE\x02\x9E\x25\xAE\x02\x9F\xFD"
-     "\xAE\x02"
-     "\xA0\x25\xAE\x02\xA1\x45\xAE\x02\xA2\x65\xAE\x02\xA3\xB5\xA3\x02\xA4\x55"
-     "\xAC\x02"
-     "\xA5\xD5\xA3\x02\xA6\xC5\xA3\x02\xA7\x3D\xA3\x02\xA8\x49\xA3\x02\xA9\x5D"
-     "\xA2\x02"
-     "\xAA\x71\xA6\x02\xAB\xAD\xA7\x02\xAC\xB9\xA7\x02\xAD\xAD\xA7\x02\xAE\xF9"
-     "\xAB\x02"
-     "\xAF\xB9\xAF\x02\xB0\xC9\xAF\x02\xB1\xD9\xAF\x02\xB2\xED\xAF\x02\xB3\x9D"
-     "\xAF\x02"
-     "\xB4\x51\xAF\x02\xB5\x49\xBB\x02\xB6\xC9\x9F\x00\xB7\x71\xBB\x02\xB8\xBD"
-     "\xAE\x02"
-     "\xB9\x65\xAE\x02\xBA\x4D\xAE\x02\xBB\x2D\xAE\x02\xBC\x1D\xAE\x02\xBD\xCD"
-     "\xA3\x02"
-     "\xBE\x3D\xAC\x02\xBF\xBD\xA3\x02\x80\x25\xA8\x02\x81\x55\xA8\x02\x82\x89"
-     "\xA8\x02"
-     "\x83\x99\xA8\x02\x84\xAD\xA8\x02\x85\x51\xA9\x02\x86\x45\xA9\x02\x87\x9D"
-     "\xA9\x02"
-     "\x88\x85\xA9\x02\x89\xE5\xA9\x02\x8A\xC5\xA9\x02\x8B\x15\xA4\x02\x8C\xF5"
-     "\xAB\x02"
-     "\x8D\x75\xA4\x02\x8E\x65\xA4\x02\x8F\x9D\xA4\x02\x90\xA9\xA4\x02\x91\xBD"
-     "\xA5\x02"
-     "\x92\x91\xA1\x02\x93\x4D\xA0\x02\x94\x59\xA0\x02\x95\x4D\xA0\x02\x96\x19"
-     "\xAC\x02"
-     "\x97\x59\xA8\x02\x98\x69\xA8\x02\x99\x79\xA8\x02\x9A\x4D\xA8\x02\x9B\x3D"
-     "\xA8\x02"
-     "\x9C\xF1\xA8\x02\x9D\xE9\xBC\x02\x9E\x69\x98\x00\x9F\xD1\xBC\x02\xA0\xDD"
-     "\xA9\x02"
-     "\xA1\x05\xA9\x02\xA2\x2D\xA9\x02\xA3\x4D\xA9\x02\xA4\x7D\xA9\x02\xA5\xAD"
-     "\xA4\x02"
-     "\xA6\x5D\xAB\x02\xA7\xDD\xA4\x02\xA8\xFD\xA4\x02\xA9\xA9\x98\x00\xAA\x41"
-     "\xA4\x02"
-     "\xAB\x55\xA5\x02\xAC\x29\xBD\x02\xAD\xB5\xA0\x02\xAE\x71\xBD\x02\xAF\xA5"
-     "\xA0\x02"
-     "\xB0\x81\xAC\x02\xB1\xC1\xA8\x02\xB2\xC1\xA8\x02\xB3\xD1\xA8\x02\xB4\xF5"
-     "\xA8\x02"
-     "\xB5\x85\xA8\x02\xB6\x59\xA8\x02\xB7\x49\xA8\x02\xB8\x5D\xA8\x02\xB9\xA1"
-     "\xA9\x02"
-     "\xBA\xB5\xA9\x02\xBB\x6D\xA9\x02\xBC\x55\xA9\x02\xBD\x35\xA9\x02\xBE\x15"
-     "\xA9\x02"
-     "\xBF\xC5\xA4\x02\x9F\x01\xA1\x02\x9E\xDD\xA0\x02\x9D\xF1\xA4\x02\x9C\xE5"
-     "\xA5\x02"
-     "\x9B\xB1\xA5\x02\x9A\x49\xA5\x02\x99\x59\xA5\x02\x98\xD9\xAA\x02\x97\x19"
-     "\xA5\x02"
-     "\x96\xC9\xA8\x02\x95\xE9\xA8\x02\x94\x89\xA8\x02\x93\xB1\xA8\x02\x92\x69"
-     "\xA8\x02"
-     "\x91\x7D\xA8\x02\x90\x81\xA9\x02\x8F\xD5\xA9\x02\x8E\xC5\xA9\x02\x8D\x19"
-     "\xA9\x02"
-     "\x8C\x69\xA9\x02\x8B\x4D\xA9\x02\x8A\x5D\xA9\x02\x89\x5D\xA9\x02\x88\x1D"
-     "\xAD\x02"
-     "\x87\x79\xA1\x02\x86\xAD\xBC\x02\x85\x69\xA1\x02\x84\xF5\xBC\x02\x83\x89"
-     "\xA4\x02"
-     "\x82\x9D\xA5\x02\x81\x75\x99\x00\x80\x21\xA5\x02\xBF\x9D\xA9\x02\xBE\x8D"
-     "\xA9\x02"
-     "\xBD\x8D\xA9\x02\xBC\xCD\xAD\x02\xBB\x89\xA1\x02\xBA\x5D\xBC\x02\xB9\x99"
-     "\xA1\x02"
-     "\xB8\x05\xBC\x02\xB7\x59\xA4\x02\xB6\x4D\xA5\x02\xB5\xA5\x99\x00\xB4\xF1"
-     "\xA5\x02"
-     "\xB3\xF1\xA5\x02\xB2\x71\xAA\x02\xB1\x81\xA5\x02\xB0\x51\xA8\x02\xAF\x01"
-     "\xA8\x02"
-     "\xAE\x61\xA8\x02\xAD\x49\xA8\x02\xAC\x91\xA8\x02\xAB\x7D\xBD\x02\xAA\xC5"
-     "\x99\x00"
-     "\xA9\x45\xBD\x02\xA8\x5D\xA9\x02\xA7\xB1\xA9\x02\xA6\xC1\xA9\x02\xA5\xF5"
-     "\xA9\x02"
-     "\xA4\xE5\xA9\x02\xA3\xF5\xA9\x02\xA2\xB5\xAD\x02\xA1\xE1\xA1\x02\xA0\xF5"
-     "\xA1\x02"
-     "\x80\xC5\xA9\x02\x81\x45\xA6\x02\x82\x55\xA6\x02\x83\xAD\xA6\x02\x84\xF9"
-     "\xA6\x02"
-     "\x85\xED\xA7\x02\x86\xC1\xA3\x02\x87\x1D\xA2\x02\x88\x29\xA2\x02\x89\x3D"
-     "\xA2\x02"
-     "\x8A\x69\xAE\x02\x8B\x29\xAA\x02\x8C\x39\xAA\x02\x8D\x29\xAA\x02\x8E\x1D"
-     "\xAA\x02"
-     "\x8F\x6D\xAA\x02\x90\xC1\xAA\x02\x91\xD9\xBE\x02\x92\x59\x9A\x00\x93\xE1"
-     "\xBE\x02"
-     "\x94\x0D\xAB\x02\x95\xD5\xAB\x02\x96\xFD\xAB\x02\x97\x9D\xAB\x02\x98\x8D"
-     "\xAB\x02"
-     "\x99\x5D\xA6\x02\x9A\xAD\xA9\x02\x9B\x2D\xA6\x02\x9C\x2D\xA6\x02\x9D\x79"
-     "\x9A\x00"
-     "\x9E\x91\xA6\x02\x9F\x85\xA7\x02\xA0\x19\xBF\x02\xA1\x85\xA2\x02\xA2\x41"
-     "\xBF\x02"
-     "\xA3\x95\xA2\x02\xA4\x49\x0F\x00\xA5\x49\x0F\x00\x77\x05\x00\x80\xB2\x05"
-     "\x00\x80"
-     "\xA8\x79\xBF\x02\xA9\x2D\xBC\x02\xAA\x39\xBC\x02\xAB\x09\xBC\x02\xAC\x59"
-     "\xBA\x02"
-     "\xAD\xF5\xBD\x02\xAE\xD1\xBB\x02\xAF\x01\xBD\x02\xB0\x69\xBB\x02\xB1\x09"
-     "\xBA\x02"
-     "\xB2\x11\xBA\x02\xB3\xE9\xA2\x02\xB4\x21\xBA\x02\xB5\x2D\xBA\x02\xB6\xC5"
-     "\xB6\x02"
-     "\xB7\xBD\xBB\x02\xB8\x71\xB8\x02\xB9\x59\xBD\x02\xBA\x95\xBB\x02\xBB\x7D"
-     "\xBB\x02"
-     "\xBC\xA5\xB6\x02\xBD\x05\xBB\x02\xBE\x01\xBB\x02\xBF\xF5\xB8\x02\x80\x69"
-     "\xB7\x02"
-     "\x81\xE1\x0E\x00\x82\x01\x1C\x00\x83\x85\xBD\x02\x84\x81\xBD\x02\x85\xB1"
-     "\xBD\x02"
-     "\x86\xF1\xBB\x02\x87\x5D\xBC\x02\x88\x21\x1F\x00\x89\x99\xBC\x02\x8A\x41"
-     "\x95\x00"
-     "\x8B\xE1\xBB\x02\x8C\xE9\xBB\x02\x8D\x11\xA3\x02\x8E\xC9\xBB\x02\x8F\xC5"
-     "\xBB\x02"
-     "\x90\x5D\xB7\x02\x91\x25\xBA\x02\x92\x69\xB3\x02\x93\x31\xBA\x02\x94\x2D"
-     "\xBA\x02"
-     "\x95\xC5\xBA\x02\x96\x51\x1F\x00\x97\xAD\xBA\x02\x98\x99\xBA\x02\x99\x6D"
-     "\xB9\x02"
-     "\x9A\x79\x1F\x00\x9B\x85\x0E\x00\x9C\x99\xBB\x02\x9D\xC9\xBC\x02\x9E\xB5"
-     "\xBB\x02"
-     "\x9F\x8D\xBA\x02\xA0\xA1\xB3\x02\xA1\xE5\xBA\x02\xA2\x51\xBE\x02\xA3\x05"
-     "\xBD\x02"
-     "\xA4\x01\xBD\x02\xA5\x31\xBD\x02\xA6\x71\xBB\x02\xA7\xDD\xBC\x02\xA8\xC9"
-     "\xBA\x02"
-     "\xA9\x19\xBC\x02\xAA\x01\xBA\x02\xAB\x61\xBB\x02\xAC\x69\xBB\x02\xAD\x91"
-     "\xA3\x02"
-     "\xAE\x49\xBB\x02\xAF\x45\xBB\x02\xB0\xDD\xB7\x02\xB1\xA5\xBA\x02\xB2\x59"
-     "\xB9\x02"
-     "\xB3\x71\xBC\x02\xB4\xAD\xBA\x02\xB5\x45\xBA\x02\xB6\x8D\xB7\x02\xB7\x2D"
-     "\xBA\x02"
-     "\xB8\x19\xBA\x02\xB9\xED\xB9\x02\xBA\x81\xB7\x02\xBB\x09\x0E\x00\xBC\xF9"
-     "\x1C\x00"
-     "\xBD\x7D\xBD\x02\xBE\x69\xBD\x02\xBF\x59\xBD\x02\x03\x0C\x00\x80\x90\x61"
-     "\x00\x00"
-     "\x4A\x0C\x00\x80\xD8\x05\x00\x80\x56\x0C\x00\x80\xEF\x05\x00\x80\x80\x54"
-     "\x11\x00"
-     "\x81\xDC\x01\x00\x0C\x06\x00\x80\x23\x06\x00\x80\x72\x24\x00\x80\x7F\x08"
-     "\x00\x80"
-     "\x18\x07\x00\x80\x42\x08\x00\x80\xCC\xB8\xAB\x02\xCD\x5C\xAB\x02\x84\x40"
-     "\x0B\x00"
-     "\x85\x44\x09\x00\x86\x0C\x02\x00\x4B\x08\x00\x80\x10\x08\x00\x80\x7C\x08"
-     "\x00\x80"
-     "\x96\x07\x00\x80\x99\x07\x00\x80\xBD\x07\x00\x80\xBA\x06\x00\x80\x3C\x25"
-     "\x00\x80"
-     "\x0F\x27\x00\x80\x88\x00\x0C\x00\x89\xDC\x04\x00\x5C\x2F\x00\x80\xE1\x2D"
-     "\x00\x80"
-     "\xFD\x2E\x00\x80\x66\x38\x00\x80\xA4\x88\x04\x00\x2E\x2C\x00\x80\xFA\x31"
-     "\x00\x80"
-     "\xE6\x33\x00\x80\x35\x2D\x00\x80\x87\x30\x00\x80\xCD\x2C\x00\x80\xD3\x31"
-     "\x00\x80"
-     "\x21\x2D\x00\x80\x4E\x36\x00\x80\xC2\x38\x00\x80\xE1\x2E\x00\x80\x19\x2E"
-     "\x00\x80"
-     "\x9A\x33\x00\x80\x37\x31\x00\x80\x63\x31\x00\x80\x42\x39\x00\x80\x57\x31"
-     "\x00\x80"
-     "\x2F\x31\x00\x80\xA6\x35\x00\x80\xE2\x35\x00\x80\x05\x2E\x00\x80\x7A\x33"
-     "\x00\x80"
-     "\xCE\x34\x00\x80\x39\x2D\x00\x80\xAF\xE8\x04\x00\xBA\xF0\x0A\x00\x65\x2E"
-     "\x00\x80"
-     "\xB8\xF8\x05\x00\xB9\x00\x08\x00\x80\xE9\xA5\x02\x81\x45\xA2\x02\x82\x09"
-     "\x01\x00"
-     "\x83\xB1\xA2\x02\x84\x79\x8B\x00\x85\xD9\xA5\x02\x86\xC1\xA5\x02\x87\x39"
-     "\xBD\x02"
-     "\x88\xD1\xA5\x02\x89\xDD\xA5\x02\x8A\x35\xA9\x02\x8B\x4D\xA4\x02\x8C\x11"
-     "\xAD\x02"
-     "\x8D\x49\xA4\x02\x8E\x45\xA4\x02\x8F\xAD\xA4\x02\x90\x49\x01\x00\x91\xB5"
-     "\xA4\x02"
-     "\x92\xB1\xA4\x02\x93\x45\xA7\x02\x94\x41\x01\x00\x95\xBD\x10\x00\x96\xB1"
-     "\xA5\x02"
-     "\x97\xE1\xA2\x02\x98\xAD\xA5\x02\x99\x95\xA4\x02\x9A\x49\xAD\x02\x9B\x0D"
-     "\xA4\x02"
-     "\x9C\xA9\xA0\x02\x9D\xFD\xA3\x02\x9E\xE9\xA3\x02\x9F\xD9\xA3\x02\xA0\x69"
-     "\xA5\x02"
-     "\xA1\xC5\xA2\x02\xA2\xE1\xA4\x02\xA3\x31\xA2\x02\xA4\x39\xA4\x02\xA5\x59"
-     "\xA5\x02"
-     "\xA6\x41\xA5\x02\xA7\xB9\xBD\x02\xA8\x51\xA5\x02\xA9\x5D\xA5\x02\xAA\xB5"
-     "\xA9\x02"
-     "\xAB\xCD\xA4\x02\xAC\x21\xA7\x02\xAD\x09\xA2\x02\xAE\xC5\xA4\x02\xAF\x2D"
-     "\xA4\x02"
-     "\xB0\x95\xA9\x02\xB1\x35\xA4\x02\xB2\x31\xA4\x02\xB3\xC5\xA7\x02\xB4\xB9"
-     "\xA9\x02"
-     "\xB5\x31\x10\x00\xB6\xD1\x02\x00\xB7\x55\xA3\x02\xB8\x71\xA3\x02\xB9\x41"
-     "\xA3\x02"
-     "\xBA\x01\xA5\x02\xBB\xAD\xA2\x02\xBC\xF1\x01\x00\xBD\x49\xA2\x02\xBE\x91"
-     "\x8B\x00"
-     "\xBF\x31\xA5\x02\x81\x25\xBE\x02\x80\xDD\xA6\x02\x83\xF1\xA6\x02\x82\xFD"
-     "\xA6\x02"
-     "\x85\x71\xA7\x02\x84\x09\xAA\x02\x87\x65\xA7\x02\x86\x3D\xAE\x02\x89\xB1"
-     "\xA7\x02"
-     "\x88\x59\xA7\x02\x8B\xD9\xA7\x02\x8A\x25\x02\x00\x8D\x39\xA4\x02\x8C\xCD"
-     "\xA7\x02"
-     "\x8F\xD1\x13\x00\x8E\x2D\x02\x00\x91\xFD\xA1\x02\x90\xAD\xA6\x02\x93\xB9"
-     "\xA7\x02"
-     "\x92\x81\xA6\x02\x95\x31\xA7\x02\x94\x75\xAE\x02\x97\xD1\xA0\x02\x96\x85"
-     "\xA3\x02"
-     "\x99\xC5\xA0\x02\x98\xF5\xA0\x02\x9B\x29\xA1\x02\x9A\x85\xA6\x02\x9D\xCD"
-     "\xA1\x02"
-     "\x9C\x1D\xA7\x02\x9F\xB5\xA6\x02\x9E\xD5\xA7\x02\xA1\xA5\xBE\x02\xA0\x5D"
-     "\xA6\x02"
-     "\xA3\x71\xA6\x02\xA2\x7D\xA6\x02\xA5\xF1\xA7\x02\xA4\x89\xAA\x02\xA7\x25"
-     "\xA1\x02"
-     "\xA6\x0D\xA4\x02\xA9\x31\xA7\x02\xA8\xD9\xA7\x02\xAB\x59\xA7\x02\xAA\xF9"
-     "\xAA\x02"
-     "\xAD\xB9\xA4\x02\xAC\x4D\xA7\x02\xAF\x5D\x13\x00\xAE\xD5\xAA\x02\xB1\x49"
-     "\xA0\x02"
-     "\xB0\xCD\x01\x00\xB3\x6D\xA0\x02\xB2\x5D\xA0\x02\xB5\x91\xA1\x02\xB4\x3D"
-     "\xA6\x02"
-     "\xB7\x65\xA1\x02\xB6\xDD\x02\x00\xB9\x2D\xA6\x02\xB8\x8D\x88\x00\xBB\xCD"
-     "\xBE\x02"
-     "\xBA\x35\xA6\x02\xBD\x09\xA6\x02\xBC\x05\xA6\x02\xBF\x99\xA7\x02\xBE\xE1"
-     "\xAA\x02"
-     "\xC6\x00\x00\x00\x7A\x39\x00\x80\xF1\x0B\x00\x80\xB6\x24\x00\x80\x7A\x24"
-     "\x00\x80"
-     "\x84\x24\x00\x80\x98\x24\x00\x80\xC0\x24\x00\x80\xCC\x4C\xAA\x02\xCD\xA4"
-     "\xAA\x02"
-     "\xA2\x24\x00\x80\xCA\x24\x00\x80\xCC\x90\xA8\x02\xCD\x74\xAB\x02\xCC\xD8"
-     "\xAB\x02"
-     "\xCD\x4C\xAA\x02\xA0\x25\xAD\x02\xA1\x35\xAD\x02\xA2\x3D\xAD\x02\xA3\x45"
-     "\x93\x02"
-     "\xA4\xD1\xAD\x02\xA5\xD9\xAD\x02\xA6\x55\x93\x02\xA7\x9D\xAC\x02\xA8\x65"
-     "\xAE\x02"
-     "\xA9\x51\xAF\x02\xAA\x3D\xAE\x02\xAB\x51\x93\x02\xAC\xBD\xD5\x00\xAD\x21"
-     "\x93\x02"
-     "\x8E\x24\x00\x80\xAC\x24\x00\x80\x90\x1D\x93\x02\x91\xDD\xAE\x02\x92\x09"
-     "\xAE\x02"
-     "\x93\x05\xAE\x02\x94\x1D\xAE\x02\x95\x05\xAE\x02\x96\x0D\xAE\x02\x97\x3D"
-     "\xAE\x02"
-     "\x98\x15\xAE\x02\x99\x69\xAE\x02\x9A\x25\x93\x02\x9B\xBD\xAD\x02\x9C\xB1"
-     "\x93\x02"
-     "\x9D\x85\xAD\x02\x9E\xFD\xAD\x02\x9F\xB9\x93\x02\x80\x89\xAD\x02\x81\xC1"
-     "\x93\x02"
-     "\x82\xAD\xAD\x02\x83\xBD\xAD\x02\x84\xA5\xAD\x02\x85\xDD\x93\x02\x86\x59"
-     "\xAD\x02"
-     "\x87\xD1\x93\x02\x88\xF1\x93\x02\x89\x55\xAD\x02\x8A\xB5\xD4\x00\x8B\x31"
-     "\x93\x02"
-     "\x8C\xF5\xAE\x02\x8D\xE1\xAE\x02\x8E\xF1\x92\x02\x8F\xB1\x93\x02\x80\xF9"
-     "\xA1\x02"
-     "\x81\x0D\xA2\x02\x82\x19\x04\x00\x83\xE5\x15\x00\x84\xF9\xA0\x02\x85\xA9"
-     "\xA7\x02"
-     "\x86\xD5\xA0\x02\x87\xED\xA1\x02\x88\x01\xA8\x02\x89\x45\xA1\x02\x8A\x11"
-     "\xAF\x02"
-     "\x8B\x15\xAF\x02\x8B\x30\x00\x80\xF5\x2D\x00\x80\x8E\x8D\x83\x00\x8F\x75"
-     "\xB8\x02"
-     "\x90\xFD\xBF\x02\x91\x8D\xBF\x02\x92\xF5\x83\x00\x93\x89\x83\x00\x94\x9D"
-     "\x83\x00"
-     "\x95\x81\x83\x00\x96\x85\x83\x00\x97\xB9\x83\x00\x98\xD5\x83\x00\x99\x2D"
-     "\xB8\x02"
-     "\x9A\xD5\xBF\x02\x9B\xA5\xBF\x02\x9C\xCD\x83\x00\x9D\xB1\x83\x00\x9E\xB5"
-     "\x83\x00"
-     "\x9F\xA9\x83\x00\xA0\x5D\x83\x00\xA1\x61\x83\x00\xA2\x3D\x83\x00\xA3\xC5"
-     "\xB8\x02"
-     "\xA4\x2D\xBF\x02\xA5\x5D\xBF\x02\xA6\x25\x83\x00\xA7\x59\x83\x00\xA8\x6D"
-     "\x83\x00"
-     "\xA9\x71\x83\x00\xAA\x75\x83\x00\xAB\x49\x83\x00\xAC\x05\x83\x00\xAD\xFD"
-     "\xB8\x02"
-     "\xAE\x05\xBF\x02\xAF\x75\xBF\x02\xB0\x7D\x83\x00\xB1\x01\x83\x00\xB2\x05"
-     "\x83\x00"
-     "\xB3\x19\x83\x00\xB4\x0D\x83\x00\xB5\x31\x83\x00\xB6\x6D\x83\x00\xB7\x95"
-     "\xB8\x02"
-     "\xB8\x5D\xBF\x02\xB9\x2D\xBF\x02\xBA\x55\x83\x00\xBB\x29\x83\x00\xBC\x3D"
-     "\x83\x00"
-     "\xBD\x21\x83\x00\xBE\x25\x83\x00\xBF\x19\x83\x00\x99\xCD\xA1\x02\x98\xAD"
-     "\xA0\x02"
-     "\x9B\x2D\xB9\x02\x9A\xD5\xA1\x02\x9D\xE9\xA1\x02\x9C\xE5\xA1\x02\x9F\x79"
-     "\xA0\x02"
-     "\x9E\x01\xAD\x02\x91\xA9\xA7\x02\x90\xFD\xA4\x02\x93\x8D\xA7\x02\x92\xBD"
-     "\xA7\x02"
-     "\x95\x71\xA6\x02\x94\xDD\xA1\x02\x97\x85\xA6\x02\x96\x55\xA0\x02\x89\xA9"
-     "\x14\x00"
-     "\x88\x55\x05\x00\x8B\xF5\xA6\x02\x8A\xA5\xA1\x02\x8D\xA1\xA0\x02\x8C\x99"
-     "\xA1\x02"
-     "\x8F\x39\xA0\x02\x8E\x7D\xA9\x02\x81\x1D\xA0\x02\x80\x45\xA9\x02\x83\xF9"
-     "\xA0\x02"
-     "\x82\x11\xA0\x02\x85\x81\xA0\x02\x84\x7D\x05\x00\x87\x71\xA3\x02\x86\x85"
-     "\xA0\x02"
-     "\xB9\xE1\xA0\x02\xB8\x99\xAD\x02\xBB\xF5\xA0\x02\xBA\xAD\xA9\x02\xBD\x01"
-     "\xA0\x02"
-     "\xBC\xE9\xA0\x02\xBF\x69\xA0\x02\xBE\x95\x05\x00\xB1\x1D\xA6\x02\xB0\xA5"
-     "\x05\x00"
-     "\xB3\x65\xA1\x02\xB2\xC5\x8F\x00\xB5\x95\xB9\x02\xB4\x6D\xA1\x02\xB7\x41"
-     "\xA1\x02"
-     "\xB6\x4D\xA1\x02\xA9\x25\x14\x00\xA8\xAD\xAD\x02\xAB\x41\xA7\x02\xAA\xC5"
-     "\x06\x00"
-     "\xAD\x75\xA7\x02\xAC\x45\xA7\x02\xAF\x99\xA6\x02\xAE\x35\xA1\x02\xA1\x5D"
-     "\xA6\x02"
-     "\xA0\x75\xA3\x02\xA3\x79\xA0\x02\xA2\x91\xA0\x02\xA5\x01\xA0\x02\xA4\xA1"
-     "\xAD\x02"
-     "\xA7\xF1\xA3\x02\xA6\x05\xA0\x02\x80\xED\x17\x00\x81\x15\x15\x00\x82\x1D"
-     "\x15\x00"
-     "\x83\x15\x15\x00\x84\x0D\x15\x00\x85\x35\x15\x00\x86\x3D\x15\x00\x87\x2D"
-     "\x15\x00"
-     "\x88\x15\x15\x00\x89\x1D\x15\x00\x8A\x15\x15\x00\x8B\xED\x15\x00\x8C\xD1"
-     "\x15\x00"
-     "\x8D\xDD\x15\x00\x8E\xC9\x15\x00\x8F\xC9\x15\x00\x90\xB9\x15\x00\x91\xB9"
-     "\x15\x00"
-     "\x92\x1D\x12\x00\x93\x39\x10\x00\x94\xD1\x10\x00\x95\xD1\x10\x00\x96\xD5"
-     "\x10\x00"
-     "\x97\xCD\x10\x00\x98\xFD\x10\x00\x99\xC5\x10\x00\x9A\xDD\x10\x00\x9B\x85"
-     "\x10\x00"
-     "\x9C\x81\x10\x00\x9D\x65\x11\x00\x9E\x6D\x1E\x00\x9F\x61\x1E\x00\xA0\xA1"
-     "\x1E\x00"
-     "\xA1\xA1\x1E\x00\x88\xF5\x05\x00\x02\x2C\x00\x80\xFA\x2B\x00\x80\x0A\x2C"
-     "\x00\x80"
-     "\xE8\xF1\x00\x80\x55\x2C\x00\x80\x82\xB5\x05\x00\x83\xBD\x05\x00\x80\xA5"
-     "\x05\x00"
-     "\x81\xBD\x05\x00\x86\x9D\x05\x00\x87\xCD\x05\x00\x84\xA5\x05\x00\x85\x95"
-     "\x05\x00"
-     "\xB1\x89\xBA\x02\xB0\x71\x81\x00\xB3\x01\xBD\x02\xB2\x71\xBD\x02\xB5\x15"
-     "\x81\x00"
-     "\xB4\x69\x81\x00\xB7\x0D\x81\x00\xB6\x11\x81\x00\xB9\x05\x81\x00\xB8\x39"
-     "\x81\x00"
-     "\x90\xD9\x02\x00\x91\xD9\x02\x00\x69\x2C\x00\x80\x71\x2C\x00\x80\x79\x2C"
-     "\x00\x80"
-     "\x7D\x2C\x00\x80\x80\xE1\x27\x00\x81\xE1\x27\x00\x82\xFD\x27\x00\x83\x25"
-     "\x26\x00"
-     "\x81\x2C\x00\x80\x85\x69\x26\x00\x86\x2D\x26\x00\x87\xE5\x27\x00\x88\x69"
-     "\x26\x00"
-     "\x89\xF5\x26\x00\x8A\x49\x26\x00\x8B\x49\x26\x00\x8C\x59\x26\x00\x8D\x59"
-     "\x26\x00"
-     "\x8E\x01\x26\x00\x8F\x6D\x26\x00\x90\x19\x26\x00\x91\x05\x26\x00\x92\x15"
-     "\x26\x00"
-     "\x93\x7D\x26\x00\x94\x6D\x26\x00\x95\xB9\x27\x00\x96\xA9\x27\x00\x97\xA1"
-     "\x27\x00"
-     "\x98\x4D\x26\x00\x99\x21\x26\x00\x9A\x25\x26\x00\x9B\x39\x26\x00\x9C\x35"
-     "\x1F\x00"
-     "\x9D\x0D\x36\x00\x9E\x61\x1E\x00\x9F\x45\x1E\x00\xBA\x2A\x00\x80\xA1\x61"
-     "\x27\x00"
-     "\xA2\x7D\x27\x00\xFC\xF1\x00\x80\xA4\xE1\x26\x00\xA1\x2C\x00\x80\xA5\x2C"
-     "\x00\x80"
-     "\xA7\x65\x27\x00\xC1\x2C\x00\x80\xA9\x75\x26\x00\xAA\xC9\x26\x00\xAB\xC9"
-     "\x26\x00"
-     "\xAC\xD9\x26\x00\xAD\xD9\x26\x00\xAE\x81\x26\x00\xAF\xED\x26\x00\xB0\x99"
-     "\x26\x00"
-     "\xB1\x85\x26\x00\xB2\x95\x26\x00\xED\xF1\x00\x80\xB4\xED\x26\x00\xB5\x39"
-     "\x27\x00"
-     "\xB6\x29\x27\x00\xB7\x21\x27\x00\xBE\x2A\x00\x80\xB9\xA1\x26\x00\xAD\x2C"
-     "\x00\x80"
-     "\xBB\xB9\x26\x00\xF7\xF1\x00\x80\xDD\x2C\x00\x80\xE5\x2C\x00\x80\x3E\x2C"
-     "\x00\x80"
-     "\x80\xE1\x28\x00\x81\xE1\x28\x00\x82\xFD\x28\x00\x83\x25\x29\x00\x84\x61"
-     "\x29\x00"
-     "\x85\x69\x29\x00\x86\x2D\x29\x00\x87\xE5\x28\x00\x88\x69\x29\x00\x89\xF5"
-     "\x29\x00"
-     "\xED\x2C\x00\x80\x8B\x49\x29\x00\x8C\x59\x29\x00\x8D\x59\x29\x00\x8E\x01"
-     "\x29\x00"
-     "\x8F\x6D\x29\x00\x90\x19\x29\x00\x91\x05\x29\x00\x92\x15\x29\x00\x93\x7D"
-     "\x29\x00"
-     "\x94\x6D\x29\x00\x95\xB9\x28\x00\x96\xA9\x28\x00\x97\xA1\x28\x00\x98\x4D"
-     "\x29\x00"
-     "\x99\x21\x29\x00\x9A\x25\x29\x00\x9B\x39\x29\x00\xD2\xF3\x00\x80\x15\x2D"
-     "\x00\x80"
-     "\xC6\x2A\x00\x80\x2D\x2D\x00\x80\x31\x2D\x00\x80\xA1\x61\x28\x00\xA2\x7D"
-     "\x28\x00"
-     "\xA3\xA5\x29\x00\x3D\x2D\x00\x80\xA5\xE9\x29\x00\xA6\xAD\x29\x00\xA7\x65"
-     "\x28\x00"
-     "\xA8\xE9\x29\x00\xA9\x75\x29\x00\xCA\x2A\x00\x80\xAB\xC9\x29\x00\xAC\xD9"
-     "\x29\x00"
-     "\xAD\xD9\x29\x00\xAE\x81\x29\x00\xAF\xED\x29\x00\xB0\x99\x29\x00\xB1\x85"
-     "\x29\x00"
-     "\xB2\x95\x29\x00\xB3\xFD\x29\x00\xB4\xED\x29\x00\xB5\x39\x28\x00\xB6\x29"
-     "\x28\x00"
-     "\xB7\x21\x28\x00\xB8\xCD\x29\x00\xB9\xA1\x29\x00\xBA\xA5\x29\x00\xBB\xB9"
-     "\x29\x00"
-     "\x75\x2D\x00\x80\x79\x2D\x00\x80\xA1\x2D\x00\x80\x3F\x00\x00\x00\x80\x55"
-     "\x10\x00"
-     "\x81\x65\x10\x00\x82\x75\x10\x00\x83\xA9\x10\x00\x84\xA1\x10\x00\x85\x8D"
-     "\x17\x00"
-     "\x86\x85\x17\x00\x87\x11\x09\x00\x88\x35\x09\x00\x89\x05\x09\x00\x8A\x0D"
-     "\x09\x00"
-     "\xAD\x2D\x00\x80\xBD\x2D\x00\x80\xC5\x2D\x00\x80\xC9\x2D\x00\x80\x0B\xF2"
-     "\x00\x80"
-     "\x90\x35\xFB\x00\x91\x39\xFB\x00\x92\xC9\xFB\x00\x93\xC9\xFB\x00\x94\xD9"
-     "\xFB\x00"
-     "\x95\xD9\xFB\x00\x96\xC9\xFB\x00\x97\xC9\xFB\x00\x98\xF9\xFB\x00\x99\xF9"
-     "\xFB\x00"
-     "\x9A\xC9\xFB\x00\x9B\xC9\xFB\x00\x9C\xDD\xFB\x00\x9D\xC5\xFB\x00\x9E\xCD"
-     "\xFB\x00"
-     "\x9F\xC5\xFB\x00\xA0\x3D\xFB\x00\xA1\x45\xFB\x00\xA2\x4D\xFB\x00\xA3\x45"
-     "\xFB\x00"
-     "\xA4\x5D\xFB\x00\xA5\x45\xFB\x00\xA6\x4D\xFB\x00\xA7\x45\xFB\x00\xA8\x7D"
-     "\xFB\x00"
-     "\xA9\x45\xFB\x00\xAA\x81\x09\x00\xAB\xA5\xBD\x02\xAC\xF1\xB1\x02\xAD\x7D"
-     "\xDC\x00"
-     "\xAE\x95\x09\x00\xCA\x33\x00\x80\xB0\x81\xB9\x02\xB1\xC1\xBD\x02\xB2\xC1"
-     "\xBD\x02"
-     "\xB3\xD1\xBD\x02\xB4\xF5\xBD\x02\xB5\x85\xBD\x02\xB6\x59\xBD\x02\xB7\x49"
-     "\xBD\x02"
-     "\xB8\x5D\xBD\x02\xB9\xA1\xBC\x02\xBA\xB5\xBC\x02\xBB\x6D\xBC\x02\xBC\x55"
-     "\xBC\x02"
-     "\xBD\x35\xBC\x02\xBE\x15\xBC\x02\xBF\xC5\xB1\x02\x80\x00\x00\x00\xBE\x1D"
-     "\x13\x00"
-     "\xDD\x2D\x00\x80\xBC\x49\x12\x00\xBB\x45\x2B\x00\xBA\x59\x2B\x00\xB9\x5D"
-     "\x2B\x00"
-     "\xE5\x2D\x00\x80\xB7\xDD\x2A\x00\xB6\xD5\x2A\x00\xB5\xC5\x2A\x00\xB4\x11"
-     "\x2B\x00"
-     "\x10\xF2\x00\x80\xB2\x69\x2B\x00\xB1\x79\x2B\x00\xB0\x65\x2B\x00\xAF\x11"
-     "\x2B\x00"
-     "\xAE\x7D\x2B\x00\xAD\x25\x2B\x00\xAC\x25\x2B\x00\xF1\x2D\x00\x80\xAA\x35"
-     "\x2B\x00"
-     "\xA9\x89\x2B\x00\xA8\x15\x2B\x00\xA7\x99\x2A\x00\xF9\x2D\x00\x80\x01\x2E"
-     "\x00\x80"
-     "\xA4\x1D\x2B\x00\x29\x2E\x00\x80\xA2\x81\x2A\x00\xA1\x9D\x2A\x00\x15\x2E"
-     "\x00\x80"
-     "\x9F\xB9\x13\x00\x21\x2E\x00\x80\x9D\xF1\x3B\x00\x25\x2E\x00\x80\x9B\xC5"
-     "\x2B\x00"
-     "\x2D\x2E\x00\x80\x99\xDD\x2B\x00\x35\x2E\x00\x80\x97\x5D\x2A\x00\x39\x2E"
-     "\x00\x80"
-     "\x41\x2E\x00\x80\x94\x91\x2B\x00\x49\x2E\x00\x80\x92\xE9\x2B\x00\x91\xF9"
-     "\x2B\x00"
-     "\x4D\x2E\x00\x80\x8F\x91\x2B\x00\x8E\xFD\x2B\x00\x8D\xA5\x2B\x00\x5D\x2E"
-     "\x00\x80"
-     "\x8B\xB5\x2B\x00\x61\x2E\x00\x80\x89\x09\x2B\x00\x6D\x2E\x00\x80\x87\x19"
-     "\x2A\x00"
-     "\x91\x2E\x00\x80\x95\x2E\x00\x80\x79\x2E\x00\x80\x81\x2E\x00\x80\x82\x01"
-     "\x2A\x00"
-     "\xDD\x2E\x00\x80\xA5\x2E\x00\x80\x80\xC5\xBC\x02\x81\x45\xB3\x02\x82\x55"
-     "\xB3\x02"
-     "\x83\xAD\xB3\x02\x84\xF9\xB3\x02\x85\xED\xB2\x02\x86\xC1\xB6\x02\x87\x1D"
-     "\xB7\x02"
-     "\x88\x29\xB7\x02\x89\x3D\xB7\x02\x8A\x19\x0B\x00\x8B\xB5\xDE\x00\x8C\x05"
-     "\x0B\x00"
-     "\x8D\x79\x0B\x00\x8E\x69\x0B\x00\x8F\x69\x0B\x00\xA0\x40\x05\x00\xA1\x44"
-     "\x04\x00"
-     "\xA2\x48\x06\x00\xA3\x4C\x78\x00\xA4\x00\x7C\x00\xA5\x54\x7D\x00\xA6\x58"
-     "\x7E\x00"
-     "\xA7\x5C\x70\x00\xA8\x60\x72\x00\xA9\x2E\x00\x80\xAD\x2E\x00\x80\xA1\x2E"
-     "\x00\x80"
-     "\xB5\x2E\x00\x80\xB1\x2E\x00\x80\xC9\x2E\x00\x80\x15\xF2\x00\x80\xE5\x2E"
-     "\x00\x80"
-     "\xE9\x2E\x00\x80\xF1\x2E\x00\x80\x05\x2F\x00\x80\x28\x2F\x00\x80\x1A\xF2"
-     "\x00\x80"
-     "\x1F\xF2\x00\x80\x34\x2F\x00\x80\x38\x2F\x00\x80\x3C\x2F\x00\x80\xAA\xF1"
-     "\x0B\x00"
-     "\xAB\xF1\x0B\x00\xAC\xD1\x0B\x00\x30\x2F\x00\x80\x44\x2F\x00\x80\xCE\x2A"
-     "\x00\x80"
-     "\xD2\x2A\x00\x80\x4C\x2F\x00\x80\x54\x2F\x00\x80\x24\xF2\x00\x80\x70\x2F"
-     "\x00\x80"
-     "\x74\x2F\x00\x80\x80\x2F\x00\x80\x29\xF2\x00\x80\x88\x2F\x00\x80\x8C\x2F"
-     "\x00\x80"
-     "\x6B\x30\x00\x80\xD6\x2A\x00\x80\xA0\x2F\x00\x80\xBC\x2F\x00\x80\xB8\x2F"
-     "\x00\x80"
-     "\x33\xF2\x00\x80\x80\x8D\x0A\x00\x81\x95\x0A\x00\x82\x9D\x0A\x00\x83\x95"
-     "\x0A\x00"
-     "\x84\x8D\x0A\x00\x85\xD1\xC6\x00\x86\xE9\x0A\x00\x87\xE9\x0A\x00\x88\xD9"
-     "\x0A\x00"
-     "\x89\xD9\x0A\x00\x8A\xE1\xC6\x00\x8B\x3D\x0B\x00\x8C\xB1\x0B\x00\x8D\xB9"
-     "\x0B\x00"
-     "\x8E\xF5\xC6\x00\x8F\xA9\x0B\x00\x90\xD9\x0B\x00\x91\xE5\x0B\x00\x92\x85"
-     "\x0B\x00"
-     "\x93\xAD\x0B\x00\x94\xB9\x0B\x00\x95\xF5\x0C\x00\x96\xA9\x0B\x00\x97\xA9"
-     "\x0B\x00"
-     "\x98\x99\x0B\x00\x99\x99\x0B\x00\x9A\x41\x04\x00\x9B\x19\xC1\x00\x9C\x01"
-     "\x04\x00"
-     "\x9D\x69\xF6\x00\x9E\x0D\x04\x00\x9F\x05\x04\x00\xA0\xFD\x04\x00\xA1\x05"
-     "\x04\x00"
-     "\xA2\x61\x0C\x00\xA3\x1D\x04\x00\xA4\x39\x04\x00\xA5\x1D\xC1\x00\xA6\x49"
-     "\xC6\x00"
-     "\xA7\x49\xC6\x00\xA8\x2D\xC1\x00\xA9\x55\x04\x00\xAA\x5D\x04\x00\xAB\x55"
-     "\xCF\x00"
-     "\xAC\x49\x04\x00\xAD\x6D\xC6\x00\xAE\x95\x05\x00\xAF\x8D\x05\x00\xB0\xF5"
-     "\x05\x00"
-     "\xB1\xFD\x05\x00\xB2\xF1\x05\x00\xB3\xF5\x05\x00\xB4\xED\x05\x00\xB5\x69"
-     "\x05\x00"
-     "\xB6\xD1\x06\x00\xB7\xC1\x06\x00\xB8\xD1\x06\x00\xB9\xB5\x06\x00\xBA\x8D"
-     "\x06\x00"
-     "\xBB\x85\x06\x00\xBC\x81\x06\x00\xBD\x89\x06\x00\xBE\x7D\x06\x00\xBF\x71"
-     "\x06\x00"
-     "\x80\x95\x07\x00\x81\x99\x07\x00\x82\xAD\x07\x00\x83\xA5\x07\x00\x84\xA1"
-     "\x07\x00"
-     "\x85\xA9\x07\x00\x86\xA5\x07\x00\x87\x89\xC0\x00\x88\xED\x07\x00\x89\xF1"
-     "\x07\x00"
-     "\x8A\xF5\x07\x00\x8B\xC9\x07\x00\x8C\xD1\xC7\x00\x8D\xDD\x07\x00\x8E\xD5"
-     "\x07\x00"
-     "\x8F\xCD\x07\x00\x90\x01\xC7\x00\x91\xBD\x07\x00\x92\xB5\x07\x00\x93\x29"
-     "\x06\x00"
-     "\x94\x39\x06\x00\x95\x39\x06\x00\x96\x29\x06\x00\x97\x29\x06\x00\x98\x19"
-     "\x06\x00"
-     "\x99\x19\x06\x00\x9A\xE9\x06\x00\x9B\xE9\x06\x00\x9C\xF9\x06\x00\x9D\xF5"
-     "\x0D\x00"
-     "\x9E\xF5\x06\x00\x9F\xED\x06\x00\xA0\x15\x06\x00\xA1\x1D\x06\x00\xA2\x15"
-     "\x06\x00"
-     "\xA3\x2D\x06\x00\xA4\x35\x06\x00\xA5\x21\x06\x00\xA6\x21\x06\x00\xA7\x21"
-     "\x06\x00"
-     "\xA8\x61\x06\x00\xA9\x61\x06\x00\xAA\x61\x06\x00\xAB\x65\x06\x00\xAC\x7D"
-     "\x06\x00"
-     "\xAD\x65\x06\x00\xAE\x6D\x06\x00\xAF\x55\xCE\x00\xB0\x19\x06\x00\xB1\x19"
-     "\x06\x00"
-     "\xB2\x29\x06\x00\xB3\x29\x06\x00\xB4\x39\x06\x00\xB5\x21\xF7\x00\xB6\x35"
-     "\x06\x00"
-     "\xB7\x2D\x06\x00\xB8\xE5\xF6\x00\xB9\x11\x06\x00\xBA\x11\x06\x00\xBB\x11"
-     "\x06\x00"
-     "\xBC\xB1\x03\x00\xBD\xB1\x03\x00\xBE\xB1\x03\x00\xBF\xB1\x03\x00\x9C\x99"
-     "\x0E\x00"
-     "\x9D\x99\x0E\x00\x9E\xA9\x0E\x00\x9F\xA9\x0E\x00\x98\x9D\x0E\x00\x99\x85"
-     "\x0E\x00"
-     "\x9A\x8D\x0E\x00\x9B\x4D\xC6\x00\x94\xE1\xF6\x00\x95\xA5\x0E\x00\x96\xAD"
-     "\x0E\x00"
-     "\x97\x6D\xE4\x00\x90\x61\xF5\x00\x91\x4D\x0E\x00\x92\x45\x0E\x00\x93\xC1"
-     "\xEA\x00"
-     "\xC4\x2F\x00\x80\x38\xF2\x00\x80\xD0\x2F\x00\x80\xCC\x2F\x00\x80\xD4\x2F"
-     "\x00\x80"
-     "\xE4\x2F\x00\x80\xE8\x2F\x00\x80\xDA\x2A\x00\x80\xF8\x2F\x00\x80\x00\x30"
-     "\x00\x80"
-     "\x04\x30\x00\x80\x08\x30\x00\x80\x80\x19\x0E\x00\x81\x19\x0E\x00\x82\x6D"
-     "\xDA\x00"
-     "\xDE\x2A\x00\x80\x3D\xF2\x00\x80\xE2\x2A\x00\x80\x2B\x30\x00\x80\x2F\x30"
-     "\x00\x80"
-     "\xB8\x75\x0E\x00\xB9\x7D\x0E\x00\xBA\x6D\x0E\x00\xBB\x99\x0D\x00\xB4\x7D"
-     "\x0E\x00"
-     "\xB5\x45\x0E\x00\xB6\xB9\xDE\x00\xB7\xD5\xF5\x00\xB0\x71\xF2\x00\xB1\x65"
-     "\x0E\x00"
-     "\xB2\x6D\x0E\x00\xB3\x65\x0E\x00\xAC\x95\xDE\x00\xAD\xE9\xE4\x00\xAE\x91"
-     "\xDE\x00"
-     "\xAF\x01\x0E\x00\xA8\x15\x0E\x00\xA9\x45\xF6\x00\xAA\xF9\xE4\x00\xAB\x15"
-     "\x0E\x00"
-     "\xA4\x55\x0E\x00\xA5\x5D\x0E\x00\xA6\x55\x0E\x00\xA7\x2D\x0E\x00\xA0\x59"
-     "\x0E\x00"
-     "\xA1\x59\x0E\x00\xA2\x19\xF5\x00\xA3\x4D\x0E\x00\x80\x51\x01\x00\x81\x51"
-     "\x01\x00"
-     "\x82\x51\x01\x00\x83\x51\x01\x00\x84\x71\x01\x00\x85\x71\x01\x00\x86\x71"
-     "\x01\x00"
-     "\x87\x71\x01\x00\x88\x5D\x01\x00\x89\xA5\x01\x00\x8A\xAD\x01\x00\x8B\xA5"
-     "\x01\x00"
-     "\x8C\xBD\x01\x00\x8D\x01\x7C\x00\x8E\xED\xCD\x00\x8F\x3D\x7F\x00\x90\x6D"
-     "\xF4\x00"
-     "\x91\x41\x7F\x00\x92\x4D\x7F\x00\x93\x45\x7F\x00\x94\x5D\x7F\x00\x95\x59"
-     "\x7F\x00"
-     "\x96\x49\x7F\x00\x97\x49\x7F\x00\x98\x79\x7F\x00\x99\x79\x7F\x00\x9A\x49"
-     "\x7F\x00"
-     "\x9B\x49\x7F\x00\x9C\x59\x7F\x00\x9D\x59\x7F\x00\x9E\x49\x7F\x00\x9F\x49"
-     "\x7F\x00"
-     "\xA0\xB9\x7F\x00\xA1\xB9\x7F\x00\xA2\xC9\x7F\x00\xA3\x79\xC5\x00\xA4\xD5"
-     "\x7F\x00"
-     "\xA5\xDD\x7F\x00\xA6\xD5\x7F\x00\xA7\xCD\x7F\x00\xA8\x65\xC2\x00\xA9\xF5"
-     "\x7F\x00"
-     "\xAA\xFD\x7F\x00\xAB\x51\xC5\x00\xAC\xE9\x7F\x00\xAD\xE9\x7F\x00\xAE\xD9"
-     "\x7F\x00"
-     "\xAF\xD9\x7F\x00\xB0\xF9\xC4\x00\xB1\x3D\xCF\x00\xB2\xA1\x7F\x00\xB3\xA1"
-     "\x7F\x00"
-     "\xB4\xA1\x7F\x00\xB5\xA1\x7F\x00\xB6\xA1\x7F\x00\xB7\xA1\x7F\x00\xB8\x61"
-     "\x7F\x00"
-     "\xB9\x61\x7F\x00\xBA\x61\x7F\x00\xBB\x61\x7F\x00\xBC\x61\x7F\x00\xBD\x61"
-     "\x7F\x00"
-     "\xBE\x61\x7F\x00\xBF\x61\x7F\x00\x04\xF4\x00\x80\x47\xF2\x00\x80\x0E\x37"
-     "\x00\x80"
-     "\x63\x30\x00\x80\xE6\x2A\x00\x80\x6A\xF2\x00\x80\x3C\xF3\x00\x80\x73\x30"
-     "\x00\x80"
-     "\x7B\x30\x00\x80\xEA\x2A\x00\x80\x8F\x30\x00\x80\x9F\x30\x00\x80\xA3\x30"
-     "\x00\x80"
-     "\xAB\x30\x00\x80\xBB\x30\x00\x80\xF2\x2A\x00\x80\xEE\x2A\x00\x80\x4C\xF2"
-     "\x00\x80"
-     "\xC3\x30\x00\x80\xD7\x30\x00\x80\xDB\x30\x00\x80\xEB\x30\x00\x80\xFF\x30"
-     "\x00\x80"
-     "\xF7\x30\x00\x80\xFB\x30\x00\x80\x03\x31\x00\x80\x17\x31\x00\x80\x1B\x31"
-     "\x00\x80"
-     "\x33\x31\x00\x80\x3B\x31\x00\x80\x4B\x31\x00\x80\x5F\x31\x00\x80\x5B\xF2"
-     "\x00\x80"
-     "\x5B\x31\x00\x80\x4F\x31\x00\x80\x67\x31\x00\x80\x6F\x31\x00\x80\x83\x31"
-     "\x00\x80"
-     "\x60\xF2\x00\x80\x8B\x31\x00\x80\x7F\x31\x00\x80\x77\x31\x00\x80\xF6\x2A"
-     "\x00\x80"
-     "\x93\x31\x00\x80\x9B\x31\x00\x80\xA3\x31\x00\x80\x97\x31\x00\x80\xFA\x2A"
-     "\x00\x80"
-     "\xBB\x31\x00\x80\x65\xF2\x00\x80\xE7\x31\x00\x80\x29\x32\x00\x80\x05\x32"
-     "\x00\x80"
-     "\x06\x2B\x00\x80\x02\x2B\x00\x80\xFE\x2A\x00\x80\xD1\x2C\x00\x80\xD5\x2C"
-     "\x00\x80"
-     "\x2D\x32\x00\x80\x15\x32\x00\x80\xE2\x36\x00\x80\x6A\x2B\x00\x80\x41\x32"
-     "\x00\x80"
-     "\x55\x32\x00\x80\x80\xA1\x01\x00\x81\xA1\x01\x00\x82\xA1\x01\x00\x83\xA1"
-     "\x01\x00"
-     "\x84\xA1\x01\x00\x85\xA1\x01\x00\x86\xA1\x01\x00\x87\xA1\x01\x00\x88\x1D"
-     "\xBA\x00"
-     "\x89\xE5\x01\x00\x8A\xED\x01\x00\x8B\xE5\x01\x00\x8C\xFD\x01\x00\x8D\xE5"
-     "\x01\x00"
-     "\x8E\xED\x01\x00\x8F\x71\xBA\x00\x90\x99\x01\x00\x91\x99\x01\x00\x92\xA9"
-     "\x01\x00"
-     "\x93\xA9\x01\x00\x94\xB9\x01\x00\x95\xB9\x01\x00\x96\xA9\x01\x00\x97\xA9"
-     "\x01\x00"
-     "\x98\xB9\xB3\x00\x99\x81\xBC\x00\x9A\x91\x01\x00\x9B\x91\x01\x00\x9C\x71"
-     "\x03\x00"
-     "\x9D\x71\x03\x00\x9E\x71\x03\x00\x9F\x71\x03\x00\xA0\x91\x03\x00\xA1\x91"
-     "\x03\x00"
-     "\xA2\xD9\xBA\x00\xA3\x95\x03\x00\xA4\x8D\x03\x00\xA5\xB5\x03\x00\xA6\xBD"
-     "\x03\x00"
-     "\xA7\x11\xBE\x00\xA8\x89\x03\x00\xA9\x89\x03\x00\xAA\x99\x03\x00\xAB\x99"
-     "\x03\x00"
-     "\xAC\x89\x03\x00\xAD\x89\x03\x00\xAE\xF9\x03\x00\xAF\xF9\x03\x00\xB0\x89"
-     "\x03\x00"
-     "\xB1\x89\x03\x00\xB2\x99\x03\x00\xB3\x99\x03\x00\xB4\x89\x03\x00\xB5\xE5"
-     "\xB7\x00"
-     "\xB6\x85\x03\x00\xB7\xBD\x03\x00\xB8\x85\x03\x00\xB9\x8D\x03\x00\xBA\x85"
-     "\x03\x00"
-     "\xBB\x9D\x03\x00\xBC\x85\x03\x00\xBD\x8D\x03\x00\xBE\x85\x03\x00\xBF\x7D"
-     "\x03\x00"
-     "\x4D\x32\x00\x80\x7E\xF2\x00\x80\x0A\x2B\x00\x80\x69\x32\x00\x80\x5D\x32"
-     "\x00\x80"
-     "\x86\x32\x00\x80\x88\xF2\x00\x80\x92\x32\x00\x80\x71\x32\x00\x80\x9A\x32"
-     "\x00\x80"
-     "\x9E\x32\x00\x80\xA2\x32\x00\x80\xA6\x32\x00\x80\xB2\x32\x00\x80\x8D\xF2"
-     "\x00\x80"
-     "\xBA\x32\x00\x80\x12\x2B\x00\x80\xC6\x32\x00\x80\x92\xF2\x00\x80\xCA\x32"
-     "\x00\x80"
-     "\x16\x2B\x00\x80\xD6\x32\x00\x80\xE2\x32\x00\x80\xF2\x32\x00\x80\x97\xF2"
-     "\x00\x80"
-     "\x2E\xF2\x00\x80\x9C\xF2\x00\x80\x12\x33\x00\x80\xA1\xF2\x00\x80\x1E\x33"
-     "\x00\x80"
-     "\x26\x33\x00\x80\x16\x33\x00\x80\x32\x33\x00\x80\x42\x33\x00\x80\x4A\x33"
-     "\x00\x80"
-     "\x56\x33\x00\x80\x5A\x33\x00\x80\xA6\xF2\x00\x80\x3E\x33\x00\x80\x72\x33"
-     "\x00\x80"
-     "\x76\x33\x00\x80\x1A\x2B\x00\x80\x86\x33\x00\x80\xAB\xF2\x00\x80\x6A\x33"
-     "\x00\x80"
-     "\xA6\x33\x00\x80\xB0\xF2\x00\x80\xB5\xF2\x00\x80\xAA\x33\x00\x80\xBE\x33"
-     "\x00\x80"
-     "\xB6\x33\x00\x80\x1E\x2B\x00\x80\xC2\x33\x00\x80\xD2\x33\x00\x80\xCE\x33"
-     "\x00\x80"
-     "\xDA\x33\x00\x80\xF2\xF1\x00\x80\xEA\x33\x00\x80\xBA\xF2\x00\x80\xF6\x33"
-     "\x00\x80"
-     "\xC4\xF2\x00\x80\x0A\x34\x00\x80\x2A\x34\x00\x80\xC9\xF2\x00\x80\x80\x00"
-     "\x00\x00"
-     "\x3E\x34\x00\x80\x42\x34\x00\x80\xCE\xF2\x00\x80\xD3\xF2\x00\x80\x5E\x34"
-     "\x00\x80"
-     "\x6E\x34\x00\x80\x22\x2B\x00\x80\x72\x34\x00\x80\x26\x2B\x00\x80\x8A\x34"
-     "\x00\x80"
-     "\x8E\x34\x00\x80\x9A\x34\x00\x80\x33\x30\x00\x80\xA6\x34\x00\x80\x2A\x2B"
-     "\x00\x80"
-     "\xBE\x34\x00\x80\xD8\xF2\x00\x80\xDA\x34\x00\x80\xDD\xF2\x00\x80\x42\xF2"
-     "\x00\x80"
-     "\xF6\x34\x00\x80\xE2\xF2\x00\x80\xE7\xF2\x00\x80\xEC\xF2\x00\x80\x2E\x2B"
-     "\x00\x80"
-     "\x32\x2B\x00\x80\xF6\xF2\x00\x80\xF1\xF2\x00\x80\xFB\xF2\x00\x80\x00\xF3"
-     "\x00\x80"
-     "\x36\x35\x00\x80\x3A\x35\x00\x80\x46\x35\x00\x80\x3E\x2B\x00\x80\x42\x2B"
-     "\x00\x80"
-     "\x0A\xF3\x00\x80\x5A\x35\x00\x80\x46\x2B\x00\x80\x0F\xF3\x00\x80\x14\xF3"
-     "\x00\x80"
-     "\x19\xF3\x00\x80\xCA\x35\x00\x80\x4A\x2B\x00\x80\xD6\x35\x00\x80\xDA\x35"
-     "\x00\x80"
-     "\x1E\xF3\x00\x80\x23\xF3\x00\x80\xF2\x35\x00\x80\x4E\x2B\x00\x80\x28\xF3"
-     "\x00\x80"
-     "\x06\x36\x00\x80\x0A\x36\x00\x80\x52\x2B\x00\x80\x2D\xF3\x00\x80\x26\x36"
-     "\x00\x80"
-     "\x56\x2B\x00\x80\x36\x36\x00\x80\x2E\x36\x00\x80\x3E\x36\x00\x80\x37\xF3"
-     "\x00\x80"
-     "\x56\x36\x00\x80\x5A\x2B\x00\x80\x62\x36\x00\x80\x80\x1D\xBB\x00\x81\x69"
-     "\x07\x00"
-     "\x82\x79\x07\x00\x83\x79\x07\x00\x84\x69\x07\x00\x85\x69\x07\x00\x86\x99"
-     "\x07\x00"
-     "\x87\x9D\x07\x00\x88\x01\xBB\x00\x89\xB1\xBB\x00\x8A\xAD\x07\x00\x8B\xA5"
-     "\x07\x00"
-     "\x8C\xBD\x07\x00\x8D\xA5\x07\x00\x8E\xAD\x07\x00\x8F\x15\xB7\x00\x90\x69"
-     "\xBB\x00"
-     "\x91\xDD\x07\x00\x92\xD5\x07\x00\x93\x45\xB9\x00\x94\xF1\x07\x00\x95\xF1"
-     "\x07\x00"
-     "\x96\xC9\xBE\x00\x97\xF5\x07\x00\x98\xCD\x07\x00\x99\x71\xB9\x00\x9A\xD9"
-     "\x07\x00"
-     "\x9B\xD9\x07\x00\x9C\xC9\x07\x00\x9D\xC9\x07\x00\x9E\xC5\x07\x00\x9F\xBD"
-     "\x07\x00"
-     "\xA0\x45\x07\x00\xA1\x4D\x07\x00\xA2\x45\x07\x00\xA3\x5D\x07\x00\xA4\x45"
-     "\x07\x00"
-     "\xA5\x4D\x07\x00\xA6\x45\x07\x00\xA7\x7D\x07\x00\xA8\x45\x07\x00\xA9\x4D"
-     "\x07\x00"
-     "\xAA\x45\x07\x00\xAB\x5D\x07\x00\xAC\x45\x07\x00\xAD\x4D\x07\x00\xAE\x45"
-     "\x07\x00"
-     "\xAF\xB9\x02\x00\xB0\xC9\x02\x00\xB1\xC9\x02\x00\xB2\xD9\x02\x00\xB3\xD9"
-     "\x02\x00"
-     "\xB4\xC9\x02\x00\xB5\xC9\x02\x00\xB6\xF9\x02\x00\xB7\xF9\x02\x00\xB8\xC9"
-     "\x02\x00"
-     "\xB9\xC9\x02\x00\xBA\xC5\xB9\x00\xBB\xDD\x02\x00\xBC\xC5\x02\x00\xBD\xCD"
-     "\x02\x00"
-     "\xBE\xC5\x02\x00\xBF\xBD\x02\x00\x94\xF5\xB8\x00\x95\xCD\x07\x00\x96\xC5"
-     "\x07\x00"
-     "\x97\xDD\x07\x00\x90\xA9\x07\x00\x91\xA9\x07\x00\x92\xD9\x07\x00\x93\xD9"
-     "\x07\x00"
-     "\x9C\xC5\x07\x00\x9D\xCD\x07\x00\x9E\xC5\x07\x00\x9F\xDD\x07\x00\x98\xE5"
-     "\x07\x00"
-     "\x99\xED\x07\x00\x9A\xE5\x07\x00\x9B\xDD\x07\x00\x84\xCD\x07\x00\x85\xD5"
-     "\x07\x00"
-     "\x86\xDD\x07\x00\x87\xD5\x07\x00\x80\xD5\x07\x00\x81\x29\xBE\x00\x82\x21"
-     "\xBD\x00"
-     "\x83\xD5\x07\x00\x8C\xC9\x07\x00\x8D\xC9\x07\x00\x8E\xD9\x07\x00\x8F\xD9"
-     "\x07\x00"
-     "\x88\xED\x07\x00\x89\xD5\x07\x00\x8A\xDD\x07\x00\x8B\x8D\xB8\x00\xB4\xC1"
-     "\x06\x00"
-     "\xB5\xC1\x06\x00\xB6\xC1\x06\x00\xB7\xC1\x06\x00\xB0\x95\xBA\x00\xB1\xA5"
-     "\x06\x00"
-     "\xB2\xA1\x06\x00\xB3\xA1\x06\x00\xBC\xFD\x06\x00\xBD\xC5\x06\x00\xBE\xCD"
-     "\x06\x00"
-     "\xBF\xC5\x06\x00\xB8\x45\xB6\x00\xB9\xE5\x06\x00\xBA\xED\x06\x00\xBB\xE5"
-     "\x06\x00"
-     "\xA4\xC5\x06\x00\xA5\xCD\x06\x00\xA6\xC5\x06\x00\xA7\xDD\x06\x00\xA0\x25"
-     "\x07\x00"
-     "\xA1\x29\xB8\x00\xA2\x21\x07\x00\xA3\x21\x07\x00\xAC\xC5\x06\x00\xAD\xC1"
-     "\x06\x00"
-     "\xAE\xC1\x06\x00\xAF\xC1\x06\x00\xA8\xE5\x06\x00\xA9\xED\x06\x00\xAA\xE5"
-     "\x06\x00"
-     "\xAB\xDD\x06\x00\x80\x00\x00\x00\x6A\x36\x00\x80\x7A\x36\x00\x80\x5E\x2B"
-     "\x00\x80"
-     "\x41\xF3\x00\x80\x46\xF3\x00\x80\x62\x2B\x00\x80\x4B\xF3\x00\x80\x92\x36"
-     "\x00\x80"
-     "\x50\xF3\x00\x80\x9E\x36\x00\x80\xAA\x36\x00\x80\x55\xF3\x00\x80\x5A\xF3"
-     "\x00\x80"
-     "\xC6\x36\x00\x80\x5F\xF3\x00\x80\xCE\x36\x00\x80\x6F\xF2\x00\x80\x66\x2B"
-     "\x00\x80"
-     "\xE6\x36\x00\x80\xEA\x36\x00\x80\x6E\x2B\x00\x80\xEE\x36\x00\x80\x48\x2F"
-     "\x00\x80"
-     "\x64\xF3\x00\x80\x69\xF3\x00\x80\x74\xF2\x00\x80\x79\xF2\x00\x80\x12\x37"
-     "\x00\x80"
-     "\x2A\x39\x00\x80\x72\x2B\x00\x80\x3E\x37\x00\x80\x3A\x37\x00\x80\x42\x37"
-     "\x00\x80"
-     "\x4D\x2D\x00\x80\x46\x37\x00\x80\x4A\x37\x00\x80\x4E\x37\x00\x80\x56\x37"
-     "\x00\x80"
-     "\x6E\xF3\x00\x80\x5A\x37\x00\x80\x6E\x37\x00\x80\x72\x37\x00\x80\x5E\x37"
-     "\x00\x80"
-     "\x76\x37\x00\x80\x8A\x37\x00\x80\x6A\x37\x00\x80\x7E\x37\x00\x80\x82\x37"
-     "\x00\x80"
-     "\x86\x37\x00\x80\x73\xF3\x00\x80\x7D\xF3\x00\x80\x78\xF3\x00\x80\x76\x2B"
-     "\x00\x80"
-     "\xA6\x37\x00\x80\xAA\x37\x00\x80\xB2\x37\x00\x80\x91\xF3\x00\x80\xB6\x37"
-     "\x00\x80"
-     "\x82\xF3\x00\x80\x7A\x2B\x00\x80\x7E\x2B\x00\x80\x87\xF3\x00\x80\x8C\xF3"
-     "\x00\x80"
-     "\x80\x45\x01\x00\x81\x4D\x01\x00\x82\x45\x01\x00\x83\x5D\x01\x00\x84\x45"
-     "\x01\x00"
-     "\x85\x4D\x01\x00\x86\x45\x01\x00\x87\x7D\x01\x00\x88\x45\x01\x00\x89\x4D"
-     "\x01\x00"
-     "\x8A\x45\x01\x00\x8B\x25\x07\x00\x8C\x41\x01\x00\x8D\x41\x01\x00\x8E\x41"
-     "\x01\x00"
-     "\x8F\x41\x01\x00\x90\xC1\x01\x00\x91\xC1\x01\x00\x92\xC1\x01\x00\x93\xC1"
-     "\x01\x00"
-     "\x94\xC1\x01\x00\x95\xC1\x01\x00\x96\xC1\x01\x00\x97\xC1\x01\x00\x98\x5D"
-     "\xB1\x00"
-     "\x99\xC5\x01\x00\x9A\xCD\x01\x00\x9B\xC5\x01\x00\x9C\xDD\x01\x00\x9D\xC5"
-     "\x01\x00"
-     "\x9E\xCD\x01\x00\x9F\x45\xBA\x00\xA0\x39\x01\x00\xA1\x39\x01\x00\xA2\x49"
-     "\x01\x00"
-     "\xA3\x49\x01\x00\xA4\x59\x01\x00\xA5\x59\x01\x00\xA6\x49\x01\x00\xA7\x49"
-     "\x01\x00"
-     "\xA8\x79\x01\x00\xA9\x79\x01\x00\xAA\x49\x01\x00\xAB\x49\x01\x00\xAC\x59"
-     "\x01\x00"
-     "\xAD\x59\x01\x00\xAE\x49\x01\x00\xAF\x49\x01\x00\xB0\x39\x01\x00\xB1\x39"
-     "\x01\x00"
-     "\xB2\x01\x0C\x00\xB3\xCD\x0F\x00\xB4\xC5\xB4\x00\xB5\xD1\x0F\x00\xB6\xD1"
-     "\x0F\x00"
-     "\xB7\xD1\x0F\x00\xB8\xF1\x0F\x00\xB9\xF1\x0F\x00\xBA\xF1\x0F\x00\xBB\xA9"
-     "\xB8\x00"
-     "\xBC\xED\x0F\x00\xBD\xD5\x0F\x00\xBE\xDD\x0F\x00\xBF\xD5\x0F\x00\xD6\x37"
-     "\x00\x80"
-     "\xDE\x37\x00\x80\xE2\x37\x00\x80\xEE\x37\x00\x80\xEA\x37\x00\x80\xF6\x37"
-     "\x00\x80"
-     "\xF2\x37\x00\x80\xFA\x37\x00\x80\xFE\x37\x00\x80\x06\x38\x00\x80\x86\x2B"
-     "\x00\x80"
-     "\x0E\x38\x00\x80\x12\x38\x00\x80\x8A\x2B\x00\x80\x22\x38\x00\x80\x96\xF3"
-     "\x00\x80"
-     "\x32\x38\x00\x80\x36\x38\x00\x80\x8E\x2B\x00\x80\x42\x38\x00\x80\xC2\x2A"
-     "\x00\x80"
-     "\x9B\xF3\x00\x80\xA0\xF3\x00\x80\x92\x2B\x00\x80\x96\x2B\x00\x80\x6E\x38"
-     "\x00\x80"
-     "\xA5\xF3\x00\x80\xC6\x38\x00\x80\xCA\x38\x00\x80\xE2\x38\x00\x80\xEE\x38"
-     "\x00\x80"
-     "\xAF\xF3\x00\x80\x01\xF2\x00\x80\xFA\x38\x00\x80\xF6\x38\x00\x80\x02\x39"
-     "\x00\x80"
-     "\x06\xF2\x00\x80\x0E\x39\x00\x80\xB4\xF3\x00\x80\xB9\xF3\x00\x80\x5A\x39"
-     "\x00\x80"
-     "\x66\x39\x00\x80\x6E\x39\x00\x80\xBE\xF3\x00\x80\x72\x39\x00\x80\x9E\x39"
-     "\x00\x80"
-     "\xAA\x39\x00\x80\xAE\x39\x00\x80\xA6\x39\x00\x80\xBA\x39\x00\x80\xBE\x39"
-     "\x00\x80"
-     "\xC3\xF3\x00\x80\xCA\x39\x00\x80\x9A\x2B\x00\x80\xD2\x39\x00\x80\xC8\xF3"
-     "\x00\x80"
-     "\x9E\x2B\x00\x80\x0A\x3A\x00\x80\xD8\x2F\x00\x80\x22\x3A\x00\x80\xCD\xF3"
-     "\x00\x80"
-     "\xD7\xF3\x00\x80\xA2\x2B\x00\x80\xA6\x2B\x00\x80\x80\x2D\x01\x00\x81\x35"
-     "\x01\x00"
-     "\x82\x3D\x01\x00\x83\x35\x01\x00\x84\x85\x85\x00\x85\x29\x01\x00\x86\x59"
-     "\x01\x00"
-     "\x87\x59\x01\x00\x88\x69\x01\x00\x89\x69\x01\x00\x8A\x79\x01\x00\x8B\x79"
-     "\x01\x00"
-     "\x8C\x69\x01\x00\x8D\x69\x01\x00\x8E\x59\x01\x00\x8F\x59\x01\x00\x90\x19"
-     "\xB6\x00"
-     "\x91\x19\xB6\x00\x92\xF9\x85\x00\x93\x25\x01\x00\x94\x3D\x01\x00\x95\x25"
-     "\x01\x00"
-     "\x96\x2D\x01\x00\x97\x25\x01\x00\x98\x1D\x01\x00\x99\xE5\x01\x00\x9A\xED"
-     "\x01\x00"
-     "\x9B\xE5\x01\x00\x9C\xFD\x01\x00\x9D\xE5\x01\x00\x9E\xED\x01\x00\x9F\x1D"
-     "\xB6\x00"
-     "\xA0\x19\x01\x00\xA1\x19\x01\x00\xA2\x29\x01\x00\xA3\x29\x01\x00\xA4\x39"
-     "\x01\x00"
-     "\xA5\x39\x01\x00\xA6\x29\x01\x00\xA7\x29\x01\x00\xA8\x19\x01\x00\xA9\x19"
-     "\x01\x00"
-     "\xAA\x69\x01\x00\xAB\x69\x01\x00\xAC\x79\x01\x00\xAD\x79\x01\x00\xAE\x69"
-     "\x01\x00"
-     "\xAF\x69\x01\x00\xB0\x19\x01\x00\xB1\x19\x01\x00\xB2\x29\x01\x00\xB3\x29"
-     "\x01\x00"
-     "\xB4\x39\x01\x00\xB5\x39\x01\x00\xB6\x29\x01\x00\xB7\x29\x01\x00\xB8\x19"
-     "\x01\x00"
-     "\xB9\x19\x01\x00\xBA\xE9\x03\x00\xBB\xE9\x03\x00\xBC\xF9\x03\x00\xBD\xF9"
-     "\x03\x00"
-     "\xBE\x81\xB6\x00\xBF\x85\xB6\x00\x4A\x3A\x00\x80\xDC\xF3\x00\x80\xAA\x2B"
-     "\x00\x80"
-     "\xE1\xF3\x00\x80\x6A\x3A\x00\x80\xE6\xF3\x00\x80\x82\x3A\x00\x80\xAE\x2B"
-     "\x00\x80"
-     "\x92\x3A\x00\x80\x9E\x3A\x00\x80\xA6\x3A\x00\x80\xAE\x3A\x00\x80\xB2\x2B"
-     "\x00\x80"
-     "\xEB\xF3\x00\x80\xDE\x3A\x00\x80\xEA\x3A\x00\x80\xB6\x2B\x00\x80\xBA\x2B"
-     "\x00\x80"
-     "\xEE\x3A\x00\x80\xF0\xF3\x00\x80\xBE\x2B\x00\x80\xF5\xF3\x00\x80\xFA\xF3"
-     "\x00\x80"
-     "\xFF\xF3\x00\x80\xC2\x2B\x00\x80\x2E\x3B\x00\x80\x32\x3B\x00\x80\x3A\x3B"
-     "\x00\x80"
-     "\x42\x3B\x00\x80\x09\xF4\x00\x80\xDF\x00\x00\x00\xDE\x00\x00\x00\xE1\x00"
-     "\x00\x00"
-     "\xE0\x00\x00\x00\xE3\x00\x00\x00\xE2\x00\x00\x00\xE5\x00\x00\x00\xE4\x00"
-     "\x00\x00"
-     "\xE7\x00\x00\x00\xE6\x00\x00\x00\xE9\x00\x00\x00\xE8\x00\x00\x00\xEB\x00"
-     "\x00\x00"
-     "\xEA\x00\x00\x00\xED\x00\x00\x00\xEC\x00\x00\x00\xEF\x00\x00\x00\xEE\x00"
-     "\x00\x00"
-     "\xF1\x00\x00\x00\xF0\x00\x00\x00\xF3\x00\x00\x00\xF2\x00\x00\x00\xF5\x00"
-     "\x00\x00"
-     "\xF4\x00\x00\x00\xF7\x00\x00\x00\xF6\x00\x00\x00\xF9\x00\x00\x00\xF8\x00"
-     "\x00\x00"
-     "\xFB\x00\x00\x00\xFA\x00\x00\x00\xFD\x00\x00\x00\xFC\x00\x00\x00\xFF\x00"
-     "\x00\x00"
-     "\xFE\x00\x00\x00\x80\x11\x01\x00\x81\x11\x01\x00\x82\x11\x01\x00\x83\x11"
-     "\x01\x00"
-     "\x84\x31\x01\x00\x85\x31\x01\x00\x86\x31\x01\x00\x87\x31\x01\x00\x88\x11"
-     "\x01\x00"
-     "\x89\x11\x01\x00\x8A\x55\xB4\x00\x8B\x15\x01\x00\x8C\x0D\x01\x00\x8D\x75"
-     "\x01\x00"
-     "\x8E\x7D\x01\x00\x8F\x75\x01\x00\x90\x0D\x01\x00\x91\x15\x01\x00\x92\x1D"
-     "\x01\x00"
-     "\x93\x15\x01\x00\x94\x0D\x01\x00\x95\x2D\x84\x00\x96\x39\x01\x00\x97\xBD"
-     "\x84\x00"
-     "\x98\x05\x01\x00\x99\x0D\x01\x00\x9A\x05\x01\x00\x9B\x1D\x01\x00\x9C\x85"
-     "\x84\x00"
-     "\x9D\x01\x01\x00\x9F\x00\x00\x00\x9E\x00\x00\x00\xA1\x00\x00\x00\xA0\x00"
-     "\x00\x00"
-     "\xA3\x00\x00\x00\xA2\x00\x00\x00\xA5\x00\x00\x00\xA4\x00\x00\x00\xA7\x00"
-     "\x00\x00"
-     "\xA6\x00\x00\x00\xA9\x00\x00\x00\xA8\x00\x00\x00\xAB\x00\x00\x00\xAA\x00"
-     "\x00\x00"
-     "\xAD\x00\x00\x00\xAC\x00\x00\x00\xAF\x00\x00\x00\xAE\x00\x00\x00\xB1\x00"
-     "\x00\x00"
-     "\xB0\x00\x00\x00\x82\x2B\x00\x80\xB2\x00\x00\x00\xB5\x00\x00\x00\xB4\x00"
-     "\x00\x00"
-     "\xB7\x00\x00\x00\xB6\x00\x00\x00\xB9\x00\x00\x00\xB8\x00\x00\x00\xBB\x00"
-     "\x00\x00"
-     "\xBA\x00\x00\x00\xBD\x00\x00\x00\xBC\x00\x00\x00\xBF\x00\x00\x00\xBE\x00"
-     "\x00\x00"
-     "\x00\x20\x00\x20\xCC\x81\x00\x20\xCC\x83\x00\x20\xCC\x84\x00\x20\xCC\x85"
-     "\x00\x20"
-     "\xCC\x86\x00\x20\xCC\x87\x00\x20\xCC\x88\x00\x20\xCC\x88\xCC\x80\x00\x20"
-     "\xCC\x88"
-     "\xCC\x81\x00\x20\xCC\x88\xCD\x82\x00\x20\xCC\x8A\x00\x20\xCC\x8B\x00\x20"
-     "\xCC\x93"
-     "\x00\x20\xCC\x93\xCC\x80\x00\x20\xCC\x93\xCC\x81\x00\x20\xCC\x93\xCD\x82"
-     "\x00\x20"
-     "\xCC\x94\x00\x20\xCC\x94\xCC\x80\x00\x20\xCC\x94\xCC\x81\x00\x20\xCC\x94"
-     "\xCD\x82"
-     "\x00\x20\xCC\xA7\x00\x20\xCC\xA8\x00\x20\xCC\xB3\x00\x20\xCD\x82\x00\x20"
-     "\xCE\xB9"
-     "\x00\x20\xD9\x8B\x00\x20\xD9\x8C\x00\x20\xD9\x8C\xD9\x91\x00\x20\xD9\x8D"
-     "\x00\x20"
-     "\xD9\x8D\xD9\x91\x00\x20\xD9\x8E\x00\x20\xD9\x8E\xD9\x91\x00\x20\xD9\x8F"
-     "\x00\x20"
-     "\xD9\x8F\xD9\x91\x00\x20\xD9\x90\x00\x20\xD9\x90\xD9\x91\x00\x20\xD9\x91"
-     "\x00\x20"
-     "\xD9\x91\xD9\xB0\x00\x20\xD9\x92\x00\x20\xE3\x82\x99\x00\x20\xE3\x82\x9A"
-     "\x00\x21"
-     "\x00\x21\x21\x00\x21\x3F\x00\x22\x00\x23\x00\x24\x00\x25\x00\x26\x00\x27"
-     "\x00\x28"
-     "\x00\x28\x31\x29\x00\x28\x31\x30\x29\x00\x28\x31\x31\x29\x00\x28\x31\x32"
-     "\x29\x00"
-     "\x28\x31\x33\x29\x00\x28\x31\x34\x29\x00\x28\x31\x35\x29\x00\x28\x31\x36"
-     "\x29\x00"
-     "\x28\x31\x37\x29\x00\x28\x31\x38\x29\x00\x28\x31\x39\x29\x00\x28\x32\x29"
-     "\x00\x28"
-     "\x32\x30\x29\x00\x28\x33\x29\x00\x28\x34\x29\x00\x28\x35\x29\x00\x28\x36"
-     "\x29\x00"
-     "\x28\x37\x29\x00\x28\x38\x29\x00\x28\x39\x29\x00\x28\x61\x29\x00\x28\x62"
-     "\x29\x00"
-     "\x28\x63\x29\x00\x28\x64\x29\x00\x28\x65\x29\x00\x28\x66\x29\x00\x28\x67"
-     "\x29\x00"
-     "\x28\x68\x29\x00\x28\x69\x29\x00\x28\x6A\x29\x00\x28\x6B\x29\x00\x28\x6C"
-     "\x29\x00"
-     "\x28\x6D\x29\x00\x28\x6E\x29\x00\x28\x6F\x29\x00\x28\x70\x29\x00\x28\x71"
-     "\x29\x00"
-     "\x28\x72\x29\x00\x28\x73\x29\x00\x28\x74\x29\x00\x28\x75\x29\x00\x28\x76"
-     "\x29\x00"
-     "\x28\x77\x29\x00\x28\x78\x29\x00\x28\x79\x29\x00\x28\x7A\x29\x00\x28\xE1"
-     "\x84\x80"
-     "\x29\x00\x28\xE1\x84\x82\x29\x00\x28\xE1\x84\x83\x29\x00\x28\xE1\x84\x85"
-     "\x29\x00"
-     "\x28\xE1\x84\x86\x29\x00\x28\xE1\x84\x87\x29\x00\x28\xE1\x84\x89\x29\x00"
-     "\x28\xE1"
-     "\x84\x8B\x29\x00\x28\xE1\x84\x8C\x29\x00\x28\xE1\x84\x8E\x29\x00\x28\xE1"
-     "\x84\x8F"
-     "\x29\x00\x28\xE1\x84\x90\x29\x00\x28\xE1\x84\x91\x29\x00\x28\xE1\x84\x92"
-     "\x29\x00"
-     "\x28\xE4\xB8\x80\x29\x00\x28\xE4\xB8\x83\x29\x00\x28\xE4\xB8\x89\x29\x00"
-     "\x28\xE4"
-     "\xB9\x9D\x29\x00\x28\xE4\xBA\x8C\x29\x00\x28\xE4\xBA\x94\x29\x00\x28\xE4"
-     "\xBB\xA3"
-     "\x29\x00\x28\xE4\xBC\x81\x29\x00\x28\xE4\xBC\x91\x29\x00\x28\xE5\x85\xAB"
-     "\x29\x00"
-     "\x28\xE5\x85\xAD\x29\x00\x28\xE5\x8A\xB4\x29\x00\x28\xE5\x8D\x81\x29\x00"
-     "\x28\xE5"
-     "\x8D\x94\x29\x00\x28\xE5\x90\x8D\x29\x00\x28\xE5\x91\xBC\x29\x00\x28\xE5"
-     "\x9B\x9B"
-     "\x29\x00\x28\xE5\x9C\x9F\x29\x00\x28\xE5\xAD\xA6\x29\x00\x28\xE6\x97\xA5"
-     "\x29\x00"
-     "\x28\xE6\x9C\x88\x29\x00\x28\xE6\x9C\x89\x29\x00\x28\xE6\x9C\xA8\x29\x00"
-     "\x28\xE6"
-     "\xA0\xAA\x29\x00\x28\xE6\xB0\xB4\x29\x00\x28\xE7\x81\xAB\x29\x00\x28\xE7"
-     "\x89\xB9"
-     "\x29\x00\x28\xE7\x9B\xA3\x29\x00\x28\xE7\xA4\xBE\x29\x00\x28\xE7\xA5\x9D"
-     "\x29\x00"
-     "\x28\xE7\xA5\xAD\x29\x00\x28\xE8\x87\xAA\x29\x00\x28\xE8\x87\xB3\x29\x00"
-     "\x28\xE8"
-     "\xB2\xA1\x29\x00\x28\xE8\xB3\x87\x29\x00\x28\xE9\x87\x91\x29\x00\x28\xEA"
-     "\xB0\x80"
-     "\x29\x00\x28\xEB\x82\x98\x29\x00\x28\xEB\x8B\xA4\x29\x00\x28\xEB\x9D\xBC"
-     "\x29\x00"
-     "\x28\xEB\xA7\x88\x29\x00\x28\xEB\xB0\x94\x29\x00\x28\xEC\x82\xAC\x29\x00"
-     "\x28\xEC"
-     "\x95\x84\x29\x00\x28\xEC\x98\xA4\xEC\xA0\x84\x29\x00\x28\xEC\x98\xA4\xED"
-     "\x9B\x84"
-     "\x29\x00\x28\xEC\x9E\x90\x29\x00\x28\xEC\xA3\xBC\x29\x00\x28\xEC\xB0\xA8"
-     "\x29\x00"
-     "\x28\xEC\xB9\xB4\x29\x00\x28\xED\x83\x80\x29\x00\x28\xED\x8C\x8C\x29\x00"
-     "\x28\xED"
-     "\x95\x98\x29\x00\x29\x00\x2A\x00\x2B\x00\x2C\x00\x2D\x00\x2E\x00\x2E\x2E"
-     "\x00\x2E"
-     "\x2E\x2E\x00\x2F\x00\x30\x00\x30\x2C\x00\x30\x2E\x00\x30\xE2\x81\x84\x33"
-     "\x00\x30"
-     "\xE7\x82\xB9\x00\x31\x00\x31\x2C\x00\x31\x2E\x00\x31\x30\x00\x31\x30\x2E"
-     "\x00\x31"
-     "\x30\xE6\x97\xA5\x00\x31\x30\xE6\x9C\x88\x00\x31\x30\xE7\x82\xB9\x00\x31"
-     "\x31\x00"
-     "\x31\x31\x2E\x00\x31\x31\xE6\x97\xA5\x00\x31\x31\xE6\x9C\x88\x00\x31\x31"
-     "\xE7\x82"
-     "\xB9\x00\x31\x32\x00\x31\x32\x2E\x00\x31\x32\xE6\x97\xA5\x00\x31\x32\xE6"
-     "\x9C\x88"
-     "\x00\x31\x32\xE7\x82\xB9\x00\x31\x33\x00\x31\x33\x2E\x00\x31\x33\xE6\x97"
-     "\xA5\x00"
-     "\x31\x33\xE7\x82\xB9\x00\x31\x34\x00\x31\x34\x2E\x00\x31\x34\xE6\x97\xA5"
-     "\x00\x31"
-     "\x34\xE7\x82\xB9\x00\x31\x35\x00\x31\x35\x2E\x00\x31\x35\xE6\x97\xA5\x00"
-     "\x31\x35"
-     "\xE7\x82\xB9\x00\x31\x36\x00\x31\x36\x2E\x00\x31\x36\xE6\x97\xA5\x00\x31"
-     "\x36\xE7"
-     "\x82\xB9\x00\x31\x37\x00\x31\x37\x2E\x00\x31\x37\xE6\x97\xA5\x00\x31\x37"
-     "\xE7\x82"
-     "\xB9\x00\x31\x38\x00\x31\x38\x2E\x00\x31\x38\xE6\x97\xA5\x00\x31\x38\xE7"
-     "\x82\xB9"
-     "\x00\x31\x39\x00\x31\x39\x2E\x00\x31\x39\xE6\x97\xA5\x00\x31\x39\xE7\x82"
-     "\xB9\x00"
-     "\x31\xE2\x81\x84\x00\x31\xE2\x81\x84\x31\x30\x00\x31\xE2\x81\x84\x32\x00"
-     "\x31\xE2"
-     "\x81\x84\x33\x00\x31\xE2\x81\x84\x34\x00\x31\xE2\x81\x84\x35\x00\x31\xE2"
-     "\x81\x84"
-     "\x36\x00\x31\xE2\x81\x84\x37\x00\x31\xE2\x81\x84\x38\x00\x31\xE2\x81\x84"
-     "\x39\x00"
-     "\x31\xE6\x97\xA5\x00\x31\xE6\x9C\x88\x00\x31\xE7\x82\xB9\x00\x32\x00\x32"
-     "\x2C\x00"
-     "\x32\x2E\x00\x32\x30\x00\x32\x30\x2E\x00\x32\x30\xE6\x97\xA5\x00\x32\x30"
-     "\xE7\x82"
-     "\xB9\x00\x32\x31\x00\x32\x31\xE6\x97\xA5\x00\x32\x31\xE7\x82\xB9\x00\x32"
-     "\x32\x00"
-     "\x32\x32\xE6\x97\xA5\x00\x32\x32\xE7\x82\xB9\x00\x32\x33\x00\x32\x33\xE6"
-     "\x97\xA5"
-     "\x00\x32\x33\xE7\x82\xB9\x00\x32\x34\x00\x32\x34\xE6\x97\xA5\x00\x32\x34"
-     "\xE7\x82"
-     "\xB9\x00\x32\x35\x00\x32\x35\xE6\x97\xA5\x00\x32\x36\x00\x32\x36\xE6\x97"
-     "\xA5\x00"
-     "\x32\x37\x00\x32\x37\xE6\x97\xA5\x00\x32\x38\x00\x32\x38\xE6\x97\xA5\x00"
-     "\x32\x39"
-     "\x00\x32\x39\xE6\x97\xA5\x00\x32\xE2\x81\x84\x33\x00\x32\xE2\x81\x84\x35"
-     "\x00\x32"
-     "\xE6\x97\xA5\x00\x32\xE6\x9C\x88\x00\x32\xE7\x82\xB9\x00\x33\x00\x33\x2C"
-     "\x00\x33"
-     "\x2E\x00\x33\x30\x00\x33\x30\xE6\x97\xA5\x00\x33\x31\x00\x33\x31\xE6\x97"
-     "\xA5\x00"
-     "\x33\x32\x00\x33\x33\x00\x33\x34\x00\x33\x35\x00\x33\x36\x00\x33\x37\x00"
-     "\x33\x38"
-     "\x00\x33\x39\x00\x33\xE2\x81\x84\x34\x00\x33\xE2\x81\x84\x35\x00\x33\xE2"
-     "\x81\x84"
-     "\x38\x00\x33\xE6\x97\xA5\x00\x33\xE6\x9C\x88\x00\x33\xE7\x82\xB9\x00\x34"
-     "\x00\x34"
-     "\x2C\x00\x34\x2E\x00\x34\x30\x00\x34\x31\x00\x34\x32\x00\x34\x33\x00\x34"
-     "\x34\x00"
-     "\x34\x35\x00\x34\x36\x00\x34\x37\x00\x34\x38\x00\x34\x39\x00\x34\xE2\x81"
-     "\x84\x35"
-     "\x00\x34\xE6\x97\xA5\x00\x34\xE6\x9C\x88\x00\x34\xE7\x82\xB9\x00\x35\x00"
-     "\x35\x2C"
-     "\x00\x35\x2E\x00\x35\x30\x00\x35\xE2\x81\x84\x36\x00\x35\xE2\x81\x84\x38"
-     "\x00\x35"
-     "\xE6\x97\xA5\x00\x35\xE6\x9C\x88\x00\x35\xE7\x82\xB9\x00\x36\x00\x36\x2C"
-     "\x00\x36"
-     "\x2E\x00\x36\xE6\x97\xA5\x00\x36\xE6\x9C\x88\x00\x36\xE7\x82\xB9\x00\x37"
-     "\x00\x37"
-     "\x2C\x00\x37\x2E\x00\x37\xE2\x81\x84\x38\x00\x37\xE6\x97\xA5\x00\x37\xE6"
-     "\x9C\x88"
-     "\x00\x37\xE7\x82\xB9\x00\x38\x00\x38\x2C\x00\x38\x2E\x00\x38\xE6\x97\xA5"
-     "\x00\x38"
-     "\xE6\x9C\x88\x00\x38\xE7\x82\xB9\x00\x39\x00\x39\x2C\x00\x39\x2E\x00\x39"
-     "\xE6\x97"
-     "\xA5\x00\x39\xE6\x9C\x88\x00\x39\xE7\x82\xB9\x00\x3A\x00\x3A\x3A\x3D\x00"
-     "\x3B\x00"
-     "\x3C\x00\x3D\x00\x3D\x3D\x00\x3D\x3D\x3D\x00\x3E\x00\x3F\x00\x3F\x21\x00"
-     "\x3F\x3F"
-     "\x00\x40\x00\x5B\x00\x5C\x00\x5D\x00\x5E\x00\x5F\x00\x60\x00\x61\x00\x61"
-     "\x2E\x6D"
-     "\x2E\x00\x61\x2F\x63\x00\x61\x2F\x73\x00\x61\x75\x00\x61\xCA\xBE\x00\x61"
-     "\xE2\x88"
-     "\x95\x6D\x00\x62\x00\x62\x61\x72\x00\x62\x71\x00\x63\x00\x63\x2F\x6F\x00"
-     "\x63\x2F"
-     "\x75\x00\x63\x61\x6C\x00\x63\x63\x00\x63\x64\x00\x63\x6D\x00\x63\x6D\x32"
-     "\x00\x63"
-     "\x6D\x33\x00\x63\x6F\x2E\x00\x63\xE2\x88\x95\x6B\x67\x00\x64\x00\x64\x61"
-     "\x00\x64"
-     "\x62\x00\x64\x6A\x00\x64\x6C\x00\x64\x6D\x00\x64\x6D\x32\x00\x64\x6D\x33"
-     "\x00\x64"
-     "\x7A\x00\x64\xC5\xBE\x00\x65\x00\x65\x72\x67\x00\x65\x76\x00\x66\x00\x66"
-     "\x61\x78"
-     "\x00\x66\x66\x00\x66\x66\x69\x00\x66\x66\x6C\x00\x66\x69\x00\x66\x6C\x00"
-     "\x66\x6D"
-     "\x00\x67\x00\x67\x61\x6C\x00\x67\x62\x00\x67\x68\x7A\x00\x67\x70\x61\x00"
-     "\x67\x79"
-     "\x00\x68\x00\x68\x61\x00\x68\x67\x00\x68\x70\x00\x68\x70\x61\x00\x68\x76"
-     "\x00\x68"
-     "\x7A\x00\x69\x00\x69\x69\x00\x69\x69\x69\x00\x69\x6A\x00\x69\x6E\x00\x69"
-     "\x75\x00"
-     "\x69\x76\x00\x69\x78\x00\x6A\x00\x6B\x00\x6B\x61\x00\x6B\x62\x00\x6B\x63"
-     "\x61\x6C"
-     "\x00\x6B\x67\x00\x6B\x68\x7A\x00\x6B\x6B\x00\x6B\x6C\x00\x6B\x6D\x00\x6B"
-     "\x6D\x32"
-     "\x00\x6B\x6D\x33\x00\x6B\x70\x61\x00\x6B\x74\x00\x6B\x76\x00\x6B\x77\x00"
-     "\x6B\xCF"
-     "\x89\x00\x6C\x00\x6C\x6A\x00\x6C\x6D\x00\x6C\x6E\x00\x6C\x6F\x67\x00\x6C"
-     "\x74\x64"
-     "\x00\x6C\x78\x00\x6C\xC2\xB7\x00\x6D\x00\x6D\x32\x00\x6D\x33\x00\x6D\x61"
-     "\x00\x6D"
-     "\x62\x00\x6D\x63\x00\x6D\x64\x00\x6D\x67\x00\x6D\x68\x7A\x00\x6D\x69\x6C"
-     "\x00\x6D"
-     "\x6C\x00\x6D\x6D\x00\x6D\x6D\x32\x00\x6D\x6D\x33\x00\x6D\x6F\x6C\x00\x6D"
-     "\x70\x61"
-     "\x00\x6D\x72\x00\x6D\x73\x00\x6D\x76\x00\x6D\x77\x00\x6D\xCF\x89\x00\x6D"
-     "\xE2\x88"
-     "\x95\x73\x00\x6D\xE2\x88\x95\x73\x32\x00\x6E\x00\x6E\x61\x00\x6E\x66\x00"
-     "\x6E\x6A"
-     "\x00\x6E\x6D\x00\x6E\x6F\x00\x6E\x73\x00\x6E\x76\x00\x6E\x77\x00\x6F\x00"
-     "\x6F\x76"
-     "\x00\x70\x00\x70\x2E\x6D\x2E\x00\x70\x61\x00\x70\x63\x00\x70\x66\x00\x70"
-     "\x68\x00"
-     "\x70\x70\x6D\x00\x70\x70\x76\x00\x70\x72\x00\x70\x73\x00\x70\x74\x65\x00"
-     "\x70\x76"
-     "\x00\x70\x77\x00\x71\x00\x72\x00\x72\x61\x64\x00\x72\x61\x64\xE2\x88\x95"
-     "\x73\x00"
-     "\x72\x61\x64\xE2\x88\x95\x73\x32\x00\x72\x73\x00\x73\x00\x73\x64\x00\x73"
-     "\x6D\x00"
-     "\x73\x72\x00\x73\x73\x00\x73\x74\x00\x73\x76\x00\x74\x00\x74\x65\x6C\x00"
-     "\x74\x68"
-     "\x7A\x00\x74\x6D\x00\x75\x00\x76\x00\x76\x69\x00\x76\x69\x69\x00\x76\x69"
-     "\x69\x69"
-     "\x00\x76\xE2\x88\x95\x6D\x00\x77\x00\x77\x62\x00\x77\x63\x00\x77\x7A\x00"
-     "\x78\x00"
-     "\x78\x69\x00\x78\x69\x69\x00\x79\x00\x7A\x00\x7B\x00\x7C\x00\x7D\x00\xC2"
-     "\xA2\x00"
-     "\xC2\xA3\x00\xC2\xA5\x00\xC2\xA6\x00\xC2\xAC\x00\xC2\xB0\x63\x00\xC2\xB0"
-     "\x66\x00"
-     "\xC2\xB7\x00\xC3\x9F\x00\xC3\xA0\x00\xC3\xA1\x00\xC3\xA2\x00\xC3\xA3\x00"
-     "\xC3\xA4"
-     "\x00\xC3\xA5\x00\xC3\xA6\x00\xC3\xA7\x00\xC3\xA8\x00\xC3\xA9\x00\xC3\xAA"
-     "\x00\xC3"
-     "\xAB\x00\xC3\xAC\x00\xC3\xAD\x00\xC3\xAE\x00\xC3\xAF\x00\xC3\xB0\x00\xC3"
-     "\xB1\x00"
-     "\xC3\xB2\x00\xC3\xB3\x00\xC3\xB4\x00\xC3\xB5\x00\xC3\xB6\x00\xC3\xB8\x00"
-     "\xC3\xB9"
-     "\x00\xC3\xBA\x00\xC3\xBB\x00\xC3\xBC\x00\xC3\xBD\x00\xC3\xBE\x00\xC3\xBF"
-     "\x00\xC4"
-     "\x81\x00\xC4\x83\x00\xC4\x85\x00\xC4\x87\x00\xC4\x89\x00\xC4\x8B\x00\xC4"
-     "\x8D\x00"
-     "\xC4\x8F\x00\xC4\x91\x00\xC4\x93\x00\xC4\x95\x00\xC4\x97\x00\xC4\x99\x00"
-     "\xC4\x9B"
-     "\x00\xC4\x9D\x00\xC4\x9F\x00\xC4\xA1\x00\xC4\xA3\x00\xC4\xA5\x00\xC4\xA7"
-     "\x00\xC4"
-     "\xA9\x00\xC4\xAB\x00\xC4\xAD\x00\xC4\xAF\x00\xC4\xB0\x00\xC4\xB1\x00\xC4"
-     "\xB5\x00"
-     "\xC4\xB7\x00\xC4\xBA\x00\xC4\xBC\x00\xC4\xBE\x00\xC5\x82\x00\xC5\x84\x00"
-     "\xC5\x86"
-     "\x00\xC5\x88\x00\xC5\x8B\x00\xC5\x8D\x00\xC5\x8F\x00\xC5\x91\x00\xC5\x93"
-     "\x00\xC5"
-     "\x95\x00\xC5\x97\x00\xC5\x99\x00\xC5\x9B\x00\xC5\x9D\x00\xC5\x9F\x00\xC5"
-     "\xA1\x00"
-     "\xC5\xA3\x00\xC5\xA5\x00\xC5\xA7\x00\xC5\xA9\x00\xC5\xAB\x00\xC5\xAD\x00"
-     "\xC5\xAF"
-     "\x00\xC5\xB1\x00\xC5\xB3\x00\xC5\xB5\x00\xC5\xB7\x00\xC5\xBA\x00\xC5\xBC"
-     "\x00\xC5"
-     "\xBE\x00\xC6\x80\x00\xC6\x83\x00\xC6\x85\x00\xC6\x88\x00\xC6\x8C\x00\xC6"
-     "\x92\x00"
-     "\xC6\x95\x00\xC6\x99\x00\xC6\x9A\x00\xC6\x9E\x00\xC6\xA1\x00\xC6\xA3\x00"
-     "\xC6\xA5"
-     "\x00\xC6\xA8\x00\xC6\xAB\x00\xC6\xAD\x00\xC6\xB0\x00\xC6\xB4\x00\xC6\xB6"
-     "\x00\xC6"
-     "\xB9\x00\xC6\xBD\x00\xC6\xBF\x00\xC7\x80\x00\xC7\x81\x00\xC7\x82\x00\xC7"
-     "\x8E\x00"
-     "\xC7\x90\x00\xC7\x92\x00\xC7\x94\x00\xC7\x96\x00\xC7\x98\x00\xC7\x9A\x00"
-     "\xC7\x9C"
-     "\x00\xC7\x9D\x00\xC7\x9F\x00\xC7\xA1\x00\xC7\xA3\x00\xC7\xA5\x00\xC7\xA7"
-     "\x00\xC7"
-     "\xA9\x00\xC7\xAB\x00\xC7\xAD\x00\xC7\xAF\x00\xC7\xB0\x00\xC7\xB5\x00\xC7"
-     "\xB9\x00"
-     "\xC7\xBB\x00\xC7\xBD\x00\xC7\xBF\x00\xC8\x81\x00\xC8\x83\x00\xC8\x85\x00"
-     "\xC8\x87"
-     "\x00\xC8\x89\x00\xC8\x8B\x00\xC8\x8D\x00\xC8\x8F\x00\xC8\x91\x00\xC8\x93"
-     "\x00\xC8"
-     "\x95\x00\xC8\x97\x00\xC8\x99\x00\xC8\x9B\x00\xC8\x9D\x00\xC8\x9F\x00\xC8"
-     "\xA3\x00"
-     "\xC8\xA5\x00\xC8\xA7\x00\xC8\xA9\x00\xC8\xAB\x00\xC8\xAD\x00\xC8\xAF\x00"
-     "\xC8\xB1"
-     "\x00\xC8\xB3\x00\xC8\xB7\x00\xC8\xBC\x00\xC8\xBF\x00\xC9\x80\x00\xC9\x82"
-     "\x00\xC9"
-     "\x87\x00\xC9\x89\x00\xC9\x8B\x00\xC9\x8D\x00\xC9\x8F\x00\xC9\x90\x00\xC9"
-     "\x91\x00"
-     "\xC9\x92\x00\xC9\x93\x00\xC9\x94\x00\xC9\x95\x00\xC9\x96\x00\xC9\x97\x00"
-     "\xC9\x98"
-     "\x00\xC9\x99\x00\xC9\x9B\x00\xC9\x9C\x00\xC9\x9E\x00\xC9\x9F\x00\xC9\xA0"
-     "\x00\xC9"
-     "\xA1\x00\xC9\xA2\x00\xC9\xA3\x00\xC9\xA4\x00\xC9\xA5\x00\xC9\xA6\x00\xC9"
-     "\xA7\x00"
-     "\xC9\xA8\x00\xC9\xA9\x00\xC9\xAA\x00\xC9\xAB\x00\xC9\xAC\x00\xC9\xAD\x00"
-     "\xC9\xAE"
-     "\x00\xC9\xAF\x00\xC9\xB0\x00\xC9\xB1\x00\xC9\xB2\x00\xC9\xB3\x00\xC9\xB4"
-     "\x00\xC9"
-     "\xB5\x00\xC9\xB6\x00\xC9\xB7\x00\xC9\xB8\x00\xC9\xB9\x00\xC9\xBA\x00\xC9"
-     "\xBB\x00"
-     "\xC9\xBD\x00\xC9\xBE\x00\xCA\x80\x00\xCA\x81\x00\xCA\x82\x00\xCA\x83\x00"
-     "\xCA\x84"
-     "\x00\xCA\x87\x00\xCA\x88\x00\xCA\x89\x00\xCA\x8A\x00\xCA\x8B\x00\xCA\x8C"
-     "\x00\xCA"
-     "\x8D\x00\xCA\x8E\x00\xCA\x8F\x00\xCA\x90\x00\xCA\x91\x00\xCA\x92\x00\xCA"
-     "\x95\x00"
-     "\xCA\x98\x00\xCA\x99\x00\xCA\x9B\x00\xCA\x9C\x00\xCA\x9D\x00\xCA\x9E\x00"
-     "\xCA\x9F"
-     "\x00\xCA\xA1\x00\xCA\xA2\x00\xCA\xA3\x00\xCA\xA4\x00\xCA\xA5\x00\xCA\xA6"
-     "\x00\xCA"
-     "\xA7\x00\xCA\xA8\x00\xCA\xA9\x00\xCA\xAA\x00\xCA\xAB\x00\xCA\xB9\x00\xCA"
-     "\xBC\x6E"
-     "\x00\xCB\x90\x00\xCB\x91\x00\xCC\x80\x00\xCC\x81\x00\xCC\x88\xCC\x81\x00"
-     "\xCC\x93"
-     "\x00\xCD\xB1\x00\xCD\xB3\x00\xCD\xB7\x00\xCD\xBB\x00\xCD\xBC\x00\xCD\xBD"
-     "\x00\xCE"
-     "\x90\x00\xCE\xAC\x00\xCE\xAD\x00\xCE\xAE\x00\xCE\xAF\x00\xCE\xB0\x00\xCE"
-     "\xB1\x00"
-     "\xCE\xB2\x00\xCE\xB3\x00\xCE\xB4\x00\xCE\xB5\x00\xCE\xB6\x00\xCE\xB7\x00"
-     "\xCE\xB8"
-     "\x00\xCE\xB9\x00\xCE\xBA\x00\xCE\xBB\x00\xCE\xBC\x00\xCE\xBC\x61\x00\xCE"
-     "\xBC\x66"
-     "\x00\xCE\xBC\x67\x00\xCE\xBC\x6C\x00\xCE\xBC\x6D\x00\xCE\xBC\x73\x00\xCE"
-     "\xBC\x76"
-     "\x00\xCE\xBC\x77\x00\xCE\xBD\x00\xCE\xBE\x00\xCE\xBF\x00\xCF\x80\x00\xCF"
-     "\x81\x00"
-     "\xCF\x83\x00\xCF\x84\x00\xCF\x85\x00\xCF\x86\x00\xCF\x87\x00\xCF\x88\x00"
-     "\xCF\x89"
-     "\x00\xCF\x8A\x00\xCF\x8B\x00\xCF\x8C\x00\xCF\x8D\x00\xCF\x8E\x00\xCF\x97"
-     "\x00\xCF"
-     "\x99\x00\xCF\x9B\x00\xCF\x9D\x00\xCF\x9F\x00\xCF\xA1\x00\xCF\xA3\x00\xCF"
-     "\xA5\x00"
-     "\xCF\xA7\x00\xCF\xA9\x00\xCF\xAB\x00\xCF\xAD\x00\xCF\xAF\x00\xCF\xB3\x00"
-     "\xCF\xB8"
-     "\x00\xCF\xBB\x00\xD0\xB0\x00\xD0\xB1\x00\xD0\xB2\x00\xD0\xB3\x00\xD0\xB4"
-     "\x00\xD0"
-     "\xB5\x00\xD0\xB6\x00\xD0\xB7\x00\xD0\xB8\x00\xD0\xB9\x00\xD0\xBA\x00\xD0"
-     "\xBB\x00"
-     "\xD0\xBC\x00\xD0\xBD\x00\xD0\xBE\x00\xD0\xBF\x00\xD1\x80\x00\xD1\x81\x00"
-     "\xD1\x82"
-     "\x00\xD1\x83\x00\xD1\x84\x00\xD1\x85\x00\xD1\x86\x00\xD1\x87\x00\xD1\x88"
-     "\x00\xD1"
-     "\x89\x00\xD1\x8A\x00\xD1\x8B\x00\xD1\x8C\x00\xD1\x8D\x00\xD1\x8E\x00\xD1"
-     "\x8F\x00"
-     "\xD1\x90\x00\xD1\x91\x00\xD1\x92\x00\xD1\x93\x00\xD1\x94\x00\xD1\x95\x00"
-     "\xD1\x96"
-     "\x00\xD1\x97\x00\xD1\x98\x00\xD1\x99\x00\xD1\x9A\x00\xD1\x9B\x00\xD1\x9C"
-     "\x00\xD1"
-     "\x9D\x00\xD1\x9E\x00\xD1\x9F\x00\xD1\xA1\x00\xD1\xA3\x00\xD1\xA5\x00\xD1"
-     "\xA7\x00"
-     "\xD1\xA9\x00\xD1\xAB\x00\xD1\xAD\x00\xD1\xAF\x00\xD1\xB1\x00\xD1\xB3\x00"
-     "\xD1\xB5"
-     "\x00\xD1\xB7\x00\xD1\xB9\x00\xD1\xBB\x00\xD1\xBD\x00\xD1\xBF\x00\xD2\x81"
-     "\x00\xD2"
-     "\x8B\x00\xD2\x8D\x00\xD2\x8F\x00\xD2\x91\x00\xD2\x93\x00\xD2\x95\x00\xD2"
-     "\x97\x00"
-     "\xD2\x99\x00\xD2\x9B\x00\xD2\x9D\x00\xD2\x9F\x00\xD2\xA1\x00\xD2\xA3\x00"
-     "\xD2\xA5"
-     "\x00\xD2\xA7\x00\xD2\xA9\x00\xD2\xAB\x00\xD2\xAD\x00\xD2\xAF\x00\xD2\xB1"
-     "\x00\xD2"
-     "\xB3\x00\xD2\xB5\x00\xD2\xB7\x00\xD2\xB9\x00\xD2\xBB\x00\xD2\xBD\x00\xD2"
-     "\xBF\x00"
-     "\xD3\x82\x00\xD3\x84\x00\xD3\x86\x00\xD3\x88\x00\xD3\x8A\x00\xD3\x8C\x00"
-     "\xD3\x8E"
-     "\x00\xD3\x8F\x00\xD3\x91\x00\xD3\x93\x00\xD3\x95\x00\xD3\x97\x00\xD3\x99"
-     "\x00\xD3"
-     "\x9B\x00\xD3\x9D\x00\xD3\x9F\x00\xD3\xA1\x00\xD3\xA3\x00\xD3\xA5\x00\xD3"
-     "\xA7\x00"
-     "\xD3\xA9\x00\xD3\xAB\x00\xD3\xAD\x00\xD3\xAF\x00\xD3\xB1\x00\xD3\xB3\x00"
-     "\xD3\xB5"
-     "\x00\xD3\xB7\x00\xD3\xB9\x00\xD3\xBB\x00\xD3\xBD\x00\xD3\xBF\x00\xD4\x81"
-     "\x00\xD4"
-     "\x83\x00\xD4\x85\x00\xD4\x87\x00\xD4\x89\x00\xD4\x8B\x00\xD4\x8D\x00\xD4"
-     "\x8F\x00"
-     "\xD4\x91\x00\xD4\x93\x00\xD4\x95\x00\xD4\x97\x00\xD4\x99\x00\xD4\x9B\x00"
-     "\xD4\x9D"
-     "\x00\xD4\x9F\x00\xD4\xA1\x00\xD4\xA3\x00\xD4\xA5\x00\xD4\xA7\x00\xD4\xA9"
-     "\x00\xD4"
-     "\xAB\x00\xD4\xAD\x00\xD4\xAF\x00\xD5\xA1\x00\xD5\xA2\x00\xD5\xA3\x00\xD5"
-     "\xA4\x00"
-     "\xD5\xA5\x00\xD5\xA5\xD6\x82\x00\xD5\xA6\x00\xD5\xA7\x00\xD5\xA8\x00\xD5"
-     "\xA9\x00"
-     "\xD5\xAA\x00\xD5\xAB\x00\xD5\xAC\x00\xD5\xAD\x00\xD5\xAE\x00\xD5\xAF\x00"
-     "\xD5\xB0"
-     "\x00\xD5\xB1\x00\xD5\xB2\x00\xD5\xB3\x00\xD5\xB4\x00\xD5\xB4\xD5\xA5\x00"
-     "\xD5\xB4"
-     "\xD5\xAB\x00\xD5\xB4\xD5\xAD\x00\xD5\xB4\xD5\xB6\x00\xD5\xB5\x00\xD5\xB6"
-     "\x00\xD5"
-     "\xB7\x00\xD5\xB8\x00\xD5\xB9\x00\xD5\xBA\x00\xD5\xBB\x00\xD5\xBC\x00\xD5"
-     "\xBD\x00"
-     "\xD5\xBE\x00\xD5\xBE\xD5\xB6\x00\xD5\xBF\x00\xD6\x80\x00\xD6\x81\x00\xD6"
-     "\x82\x00"
-     "\xD6\x83\x00\xD6\x84\x00\xD6\x85\x00\xD6\x86\x00\xD7\x90\x00\xD7\x90\xD6"
-     "\xB7\x00"
-     "\xD7\x90\xD6\xB8\x00\xD7\x90\xD6\xBC\x00\xD7\x90\xD7\x9C\x00\xD7\x91\x00"
-     "\xD7\x91"
-     "\xD6\xBC\x00\xD7\x91\xD6\xBF\x00\xD7\x92\x00\xD7\x92\xD6\xBC\x00\xD7\x93"
-     "\x00\xD7"
-     "\x93\xD6\xBC\x00\xD7\x94\x00\xD7\x94\xD6\xBC\x00\xD7\x95\xD6\xB9\x00\xD7"
-     "\x95\xD6"
-     "\xBC\x00\xD7\x96\xD6\xBC\x00\xD7\x98\xD6\xBC\x00\xD7\x99\xD6\xB4\x00\xD7"
-     "\x99\xD6"
-     "\xBC\x00\xD7\x9A\xD6\xBC\x00\xD7\x9B\x00\xD7\x9B\xD6\xBC\x00\xD7\x9B\xD6"
-     "\xBF\x00"
-     "\xD7\x9C\x00\xD7\x9C\xD6\xBC\x00\xD7\x9D\x00\xD7\x9E\xD6\xBC\x00\xD7\xA0"
-     "\xD6\xBC"
-     "\x00\xD7\xA1\xD6\xBC\x00\xD7\xA2\x00\xD7\xA3\xD6\xBC\x00\xD7\xA4\xD6\xBC"
-     "\x00\xD7"
-     "\xA4\xD6\xBF\x00\xD7\xA6\xD6\xBC\x00\xD7\xA7\xD6\xBC\x00\xD7\xA8\x00\xD7"
-     "\xA8\xD6"
-     "\xBC\x00\xD7\xA9\xD6\xBC\x00\xD7\xA9\xD6\xBC\xD7\x81\x00\xD7\xA9\xD6\xBC"
-     "\xD7\x82"
-     "\x00\xD7\xA9\xD7\x81\x00\xD7\xA9\xD7\x82\x00\xD7\xAA\x00\xD7\xAA\xD6\xBC"
-     "\x00\xD7"
-     "\xB2\xD6\xB7\x00\xD8\xA1\x00\xD8\xA2\x00\xD8\xA3\x00\xD8\xA4\x00\xD8\xA5"
-     "\x00\xD8"
-     "\xA6\x00\xD8\xA6\xD8\xA7\x00\xD8\xA6\xD8\xAC\x00\xD8\xA6\xD8\xAD\x00\xD8"
-     "\xA6\xD8"
-     "\xAE\x00\xD8\xA6\xD8\xB1\x00\xD8\xA6\xD8\xB2\x00\xD8\xA6\xD9\x85\x00\xD8"
-     "\xA6\xD9"
-     "\x86\x00\xD8\xA6\xD9\x87\x00\xD8\xA6\xD9\x88\x00\xD8\xA6\xD9\x89\x00\xD8"
-     "\xA6\xD9"
-     "\x8A\x00\xD8\xA6\xDB\x86\x00\xD8\xA6\xDB\x87\x00\xD8\xA6\xDB\x88\x00\xD8"
-     "\xA6\xDB"
-     "\x90\x00\xD8\xA6\xDB\x95\x00\xD8\xA7\x00\xD8\xA7\xD9\x83\xD8\xA8\xD8\xB1"
-     "\x00\xD8"
-     "\xA7\xD9\x84\xD9\x84\xD9\x87\x00\xD8\xA7\xD9\x8B\x00\xD8\xA7\xD9\xB4\x00"
-     "\xD8\xA8"
-     "\x00\xD8\xA8\xD8\xAC\x00\xD8\xA8\xD8\xAD\x00\xD8\xA8\xD8\xAD\xD9\x8A\x00"
-     "\xD8\xA8"
-     "\xD8\xAE\x00\xD8\xA8\xD8\xAE\xD9\x8A\x00\xD8\xA8\xD8\xB1\x00\xD8\xA8\xD8"
-     "\xB2\x00"
-     "\xD8\xA8\xD9\x85\x00\xD8\xA8\xD9\x86\x00\xD8\xA8\xD9\x87\x00\xD8\xA8\xD9"
-     "\x89\x00"
-     "\xD8\xA8\xD9\x8A\x00\xD8\xA9\x00\xD8\xAA\x00\xD8\xAA\xD8\xAC\x00\xD8\xAA"
-     "\xD8\xAC"
-     "\xD9\x85\x00\xD8\xAA\xD8\xAC\xD9\x89\x00\xD8\xAA\xD8\xAC\xD9\x8A\x00\xD8"
-     "\xAA\xD8"
-     "\xAD\x00\xD8\xAA\xD8\xAD\xD8\xAC\x00\xD8\xAA\xD8\xAD\xD9\x85\x00\xD8\xAA"
-     "\xD8\xAE"
-     "\x00\xD8\xAA\xD8\xAE\xD9\x85\x00\xD8\xAA\xD8\xAE\xD9\x89\x00\xD8\xAA\xD8"
-     "\xAE\xD9"
-     "\x8A\x00\xD8\xAA\xD8\xB1\x00\xD8\xAA\xD8\xB2\x00\xD8\xAA\xD9\x85\x00\xD8"
-     "\xAA\xD9"
-     "\x85\xD8\xAC\x00\xD8\xAA\xD9\x85\xD8\xAD\x00\xD8\xAA\xD9\x85\xD8\xAE\x00"
-     "\xD8\xAA"
-     "\xD9\x85\xD9\x89\x00\xD8\xAA\xD9\x85\xD9\x8A\x00\xD8\xAA\xD9\x86\x00\xD8"
-     "\xAA\xD9"
-     "\x87\x00\xD8\xAA\xD9\x89\x00\xD8\xAA\xD9\x8A\x00\xD8\xAB\x00\xD8\xAB\xD8"
-     "\xAC\x00"
-     "\xD8\xAB\xD8\xB1\x00\xD8\xAB\xD8\xB2\x00\xD8\xAB\xD9\x85\x00\xD8\xAB\xD9"
-     "\x86\x00"
-     "\xD8\xAB\xD9\x87\x00\xD8\xAB\xD9\x89\x00\xD8\xAB\xD9\x8A\x00\xD8\xAC\x00"
-     "\xD8\xAC"
-     "\xD8\xAD\x00\xD8\xAC\xD8\xAD\xD9\x89\x00\xD8\xAC\xD8\xAD\xD9\x8A\x00\xD8"
-     "\xAC\xD9"
-     "\x84\x20\xD8\xAC\xD9\x84\xD8\xA7\xD9\x84\xD9\x87\x00\xD8\xAC\xD9\x85\x00"
-     "\xD8\xAC"
-     "\xD9\x85\xD8\xAD\x00\xD8\xAC\xD9\x85\xD9\x89\x00\xD8\xAC\xD9\x85\xD9\x8A"
-     "\x00\xD8"
-     "\xAC\xD9\x89\x00\xD8\xAC\xD9\x8A\x00\xD8\xAD\x00\xD8\xAD\xD8\xAC\x00\xD8"
-     "\xAD\xD8"
-     "\xAC\xD9\x8A\x00\xD8\xAD\xD9\x85\x00\xD8\xAD\xD9\x85\xD9\x89\x00\xD8\xAD"
-     "\xD9\x85"
-     "\xD9\x8A\x00\xD8\xAD\xD9\x89\x00\xD8\xAD\xD9\x8A\x00\xD8\xAE\x00\xD8\xAE"
-     "\xD8\xAC"
-     "\x00\xD8\xAE\xD8\xAD\x00\xD8\xAE\xD9\x85\x00\xD8\xAE\xD9\x89\x00\xD8\xAE"
-     "\xD9\x8A"
-     "\x00\xD8\xAF\x00\xD8\xB0\x00\xD8\xB0\xD9\xB0\x00\xD8\xB1\x00\xD8\xB1\xD8"
-     "\xB3\xD9"
-     "\x88\xD9\x84\x00\xD8\xB1\xD9\xB0\x00\xD8\xB1\xDB\x8C\xD8\xA7\xD9\x84\x00"
-     "\xD8\xB2"
-     "\x00\xD8\xB3\x00\xD8\xB3\xD8\xAC\x00\xD8\xB3\xD8\xAC\xD8\xAD\x00\xD8\xB3"
-     "\xD8\xAC"
-     "\xD9\x89\x00\xD8\xB3\xD8\xAD\x00\xD8\xB3\xD8\xAD\xD8\xAC\x00\xD8\xB3\xD8"
-     "\xAE\x00"
-     "\xD8\xB3\xD8\xAE\xD9\x89\x00\xD8\xB3\xD8\xAE\xD9\x8A\x00\xD8\xB3\xD8\xB1"
-     "\x00\xD8"
-     "\xB3\xD9\x85\x00\xD8\xB3\xD9\x85\xD8\xAC\x00\xD8\xB3\xD9\x85\xD8\xAD\x00"
-     "\xD8\xB3"
-     "\xD9\x85\xD9\x85\x00\xD8\xB3\xD9\x87\x00\xD8\xB3\xD9\x89\x00\xD8\xB3\xD9"
-     "\x8A\x00"
-     "\xD8\xB4\x00\xD8\xB4\xD8\xAC\x00\xD8\xB4\xD8\xAC\xD9\x8A\x00\xD8\xB4\xD8"
-     "\xAD\x00"
-     "\xD8\xB4\xD8\xAD\xD9\x85\x00\xD8\xB4\xD8\xAD\xD9\x8A\x00\xD8\xB4\xD8\xAE"
-     "\x00\xD8"
-     "\xB4\xD8\xB1\x00\xD8\xB4\xD9\x85\x00\xD8\xB4\xD9\x85\xD8\xAE\x00\xD8\xB4"
-     "\xD9\x85"
-     "\xD9\x85\x00\xD8\xB4\xD9\x87\x00\xD8\xB4\xD9\x89\x00\xD8\xB4\xD9\x8A\x00"
-     "\xD8\xB5"
-     "\x00\xD8\xB5\xD8\xAD\x00\xD8\xB5\xD8\xAD\xD8\xAD\x00\xD8\xB5\xD8\xAD\xD9"
-     "\x8A\x00"
-     "\xD8\xB5\xD8\xAE\x00\xD8\xB5\xD8\xB1\x00\xD8\xB5\xD9\x84\xD8\xB9\xD9\x85"
-     "\x00\xD8"
-     "\xB5\xD9\x84\xD9\x89\x00\xD8\xB5\xD9\x84\xD9\x89\x20\xD8\xA7\xD9\x84\xD9"
-     "\x84\xD9"
-     "\x87\x20\xD8\xB9\xD9\x84\xD9\x8A\xD9\x87\x20\xD9\x88\xD8\xB3\xD9\x84\xD9"
-     "\x85\x00"
-     "\xD8\xB5\xD9\x84\xDB\x92\x00\xD8\xB5\xD9\x85\x00\xD8\xB5\xD9\x85\xD9\x85"
-     "\x00\xD8"
-     "\xB5\xD9\x89\x00\xD8\xB5\xD9\x8A\x00\xD8\xB6\x00\xD8\xB6\xD8\xAC\x00\xD8"
-     "\xB6\xD8"
-     "\xAD\x00\xD8\xB6\xD8\xAD\xD9\x89\x00\xD8\xB6\xD8\xAD\xD9\x8A\x00\xD8\xB6"
-     "\xD8\xAE"
-     "\x00\xD8\xB6\xD8\xAE\xD9\x85\x00\xD8\xB6\xD8\xB1\x00\xD8\xB6\xD9\x85\x00"
-     "\xD8\xB6"
-     "\xD9\x89\x00\xD8\xB6\xD9\x8A\x00\xD8\xB7\x00\xD8\xB7\xD8\xAD\x00\xD8\xB7"
-     "\xD9\x85"
-     "\x00\xD8\xB7\xD9\x85\xD8\xAD\x00\xD8\xB7\xD9\x85\xD9\x85\x00\xD8\xB7\xD9"
-     "\x85\xD9"
-     "\x8A\x00\xD8\xB7\xD9\x89\x00\xD8\xB7\xD9\x8A\x00\xD8\xB8\x00\xD8\xB8\xD9"
-     "\x85\x00"
-     "\xD8\xB9\x00\xD8\xB9\xD8\xAC\x00\xD8\xB9\xD8\xAC\xD9\x85\x00\xD8\xB9\xD9"
-     "\x84\xD9"
-     "\x8A\xD9\x87\x00\xD8\xB9\xD9\x85\x00\xD8\xB9\xD9\x85\xD9\x85\x00\xD8\xB9"
-     "\xD9\x85"
-     "\xD9\x89\x00\xD8\xB9\xD9\x85\xD9\x8A\x00\xD8\xB9\xD9\x89\x00\xD8\xB9\xD9"
-     "\x8A\x00"
-     "\xD8\xBA\x00\xD8\xBA\xD8\xAC\x00\xD8\xBA\xD9\x85\x00\xD8\xBA\xD9\x85\xD9"
-     "\x85\x00"
-     "\xD8\xBA\xD9\x85\xD9\x89\x00\xD8\xBA\xD9\x85\xD9\x8A\x00\xD8\xBA\xD9\x89"
-     "\x00\xD8"
-     "\xBA\xD9\x8A\x00\xD9\x80\xD9\x8B\x00\xD9\x80\xD9\x8E\x00\xD9\x80\xD9\x8E"
-     "\xD9\x91"
-     "\x00\xD9\x80\xD9\x8F\x00\xD9\x80\xD9\x8F\xD9\x91\x00\xD9\x80\xD9\x90\x00"
-     "\xD9\x80"
-     "\xD9\x90\xD9\x91\x00\xD9\x80\xD9\x91\x00\xD9\x80\xD9\x92\x00\xD9\x81\x00"
-     "\xD9\x81"
-     "\xD8\xAC\x00\xD9\x81\xD8\xAD\x00\xD9\x81\xD8\xAE\x00\xD9\x81\xD8\xAE\xD9"
-     "\x85\x00"
-     "\xD9\x81\xD9\x85\x00\xD9\x81\xD9\x85\xD9\x8A\x00\xD9\x81\xD9\x89\x00\xD9"
-     "\x81\xD9"
-     "\x8A\x00\xD9\x82\x00\xD9\x82\xD8\xAD\x00\xD9\x82\xD9\x84\xDB\x92\x00\xD9"
-     "\x82\xD9"
-     "\x85\x00\xD9\x82\xD9\x85\xD8\xAD\x00\xD9\x82\xD9\x85\xD9\x85\x00\xD9\x82"
-     "\xD9\x85"
-     "\xD9\x8A\x00\xD9\x82\xD9\x89\x00\xD9\x82\xD9\x8A\x00\xD9\x83\x00\xD9\x83"
-     "\xD8\xA7"
-     "\x00\xD9\x83\xD8\xAC\x00\xD9\x83\xD8\xAD\x00\xD9\x83\xD8\xAE\x00\xD9\x83"
-     "\xD9\x84"
-     "\x00\xD9\x83\xD9\x85\x00\xD9\x83\xD9\x85\xD9\x85\x00\xD9\x83\xD9\x85\xD9"
-     "\x8A\x00"
-     "\xD9\x83\xD9\x89\x00\xD9\x83\xD9\x8A\x00\xD9\x84\x00\xD9\x84\xD8\xA2\x00"
-     "\xD9\x84"
-     "\xD8\xA3\x00\xD9\x84\xD8\xA5\x00\xD9\x84\xD8\xA7\x00\xD9\x84\xD8\xAC\x00"
-     "\xD9\x84"
-     "\xD8\xAC\xD8\xAC\x00\xD9\x84\xD8\xAC\xD9\x85\x00\xD9\x84\xD8\xAC\xD9\x8A"
-     "\x00\xD9"
-     "\x84\xD8\xAD\x00\xD9\x84\xD8\xAD\xD9\x85\x00\xD9\x84\xD8\xAD\xD9\x89\x00"
-     "\xD9\x84"
-     "\xD8\xAD\xD9\x8A\x00\xD9\x84\xD8\xAE\x00\xD9\x84\xD8\xAE\xD9\x85\x00\xD9"
-     "\x84\xD9"
-     "\x85\x00\xD9\x84\xD9\x85\xD8\xAD\x00\xD9\x84\xD9\x85\xD9\x8A\x00\xD9\x84"
-     "\xD9\x87"
-     "\x00\xD9\x84\xD9\x89\x00\xD9\x84\xD9\x8A\x00\xD9\x85\x00\xD9\x85\xD8\xA7"
-     "\x00\xD9"
-     "\x85\xD8\xAC\x00\xD9\x85\xD8\xAC\xD8\xAD\x00\xD9\x85\xD8\xAC\xD8\xAE\x00"
-     "\xD9\x85"
-     "\xD8\xAC\xD9\x85\x00\xD9\x85\xD8\xAC\xD9\x8A\x00\xD9\x85\xD8\xAD\x00\xD9"
-     "\x85\xD8"
-     "\xAD\xD8\xAC\x00\xD9\x85\xD8\xAD\xD9\x85\x00\xD9\x85\xD8\xAD\xD9\x85\xD8"
-     "\xAF\x00"
-     "\xD9\x85\xD8\xAD\xD9\x8A\x00\xD9\x85\xD8\xAE\x00\xD9\x85\xD8\xAE\xD8\xAC"
-     "\x00\xD9"
-     "\x85\xD8\xAE\xD9\x85\x00\xD9\x85\xD8\xAE\xD9\x8A\x00\xD9\x85\xD9\x85\x00"
-     "\xD9\x85"
-     "\xD9\x85\xD9\x8A\x00\xD9\x85\xD9\x89\x00\xD9\x85\xD9\x8A\x00\xD9\x86\x00"
-     "\xD9\x86"
-     "\xD8\xAC\x00\xD9\x86\xD8\xAC\xD8\xAD\x00\xD9\x86\xD8\xAC\xD9\x85\x00\xD9"
-     "\x86\xD8"
-     "\xAC\xD9\x89\x00\xD9\x86\xD8\xAC\xD9\x8A\x00\xD9\x86\xD8\xAD\x00\xD9\x86"
-     "\xD8\xAD"
-     "\xD9\x85\x00\xD9\x86\xD8\xAD\xD9\x89\x00\xD9\x86\xD8\xAD\xD9\x8A\x00\xD9"
-     "\x86\xD8"
-     "\xAE\x00\xD9\x86\xD8\xB1\x00\xD9\x86\xD8\xB2\x00\xD9\x86\xD9\x85\x00\xD9"
-     "\x86\xD9"
-     "\x85\xD9\x89\x00\xD9\x86\xD9\x85\xD9\x8A\x00\xD9\x86\xD9\x86\x00\xD9\x86"
-     "\xD9\x87"
-     "\x00\xD9\x86\xD9\x89\x00\xD9\x86\xD9\x8A\x00\xD9\x87\x00\xD9\x87\xD8\xAC"
-     "\x00\xD9"
-     "\x87\xD9\x85\x00\xD9\x87\xD9\x85\xD8\xAC\x00\xD9\x87\xD9\x85\xD9\x85\x00"
-     "\xD9\x87"
-     "\xD9\x89\x00\xD9\x87\xD9\x8A\x00\xD9\x87\xD9\xB0\x00\xD9\x88\x00\xD9\x88"
-     "\xD8\xB3"
-     "\xD9\x84\xD9\x85\x00\xD9\x88\xD9\xB4\x00\xD9\x89\x00\xD9\x89\xD9\xB0\x00"
-     "\xD9\x8A"
-     "\x00\xD9\x8A\xD8\xAC\x00\xD9\x8A\xD8\xAC\xD9\x8A\x00\xD9\x8A\xD8\xAD\x00"
-     "\xD9\x8A"
-     "\xD8\xAD\xD9\x8A\x00\xD9\x8A\xD8\xAE\x00\xD9\x8A\xD8\xB1\x00\xD9\x8A\xD8"
-     "\xB2\x00"
-     "\xD9\x8A\xD9\x85\x00\xD9\x8A\xD9\x85\xD9\x85\x00\xD9\x8A\xD9\x85\xD9\x8A"
-     "\x00\xD9"
-     "\x8A\xD9\x86\x00\xD9\x8A\xD9\x87\x00\xD9\x8A\xD9\x89\x00\xD9\x8A\xD9\x8A"
-     "\x00\xD9"
-     "\x8A\xD9\xB4\x00\xD9\xAE\x00\xD9\xAF\x00\xD9\xB1\x00\xD9\xB9\x00\xD9\xBA"
-     "\x00\xD9"
-     "\xBB\x00\xD9\xBE\x00\xD9\xBF\x00\xDA\x80\x00\xDA\x83\x00\xDA\x84\x00\xDA"
-     "\x86\x00"
-     "\xDA\x87\x00\xDA\x88\x00\xDA\x8C\x00\xDA\x8D\x00\xDA\x8E\x00\xDA\x91\x00"
-     "\xDA\x98"
-     "\x00\xDA\xA1\x00\xDA\xA4\x00\xDA\xA6\x00\xDA\xA9\x00\xDA\xAD\x00\xDA\xAF"
-     "\x00\xDA"
-     "\xB1\x00\xDA\xB3\x00\xDA\xBA\x00\xDA\xBB\x00\xDA\xBE\x00\xDB\x80\x00\xDB"
-     "\x81\x00"
-     "\xDB\x82\x00\xDB\x85\x00\xDB\x86\x00\xDB\x87\x00\xDB\x87\xD9\xB4\x00\xDB"
-     "\x88\x00"
-     "\xDB\x89\x00\xDB\x8B\x00\xDB\x8C\x00\xDB\x90\x00\xDB\x92\x00\xDB\x93\x00"
-     "\xE0\xA4"
-     "\x95\xE0\xA4\xBC\x00\xE0\xA4\x96\xE0\xA4\xBC\x00\xE0\xA4\x97\xE0\xA4\xBC"
-     "\x00\xE0"
-     "\xA4\x9C\xE0\xA4\xBC\x00\xE0\xA4\xA1\xE0\xA4\xBC\x00\xE0\xA4\xA2\xE0\xA4"
-     "\xBC\x00"
-     "\xE0\xA4\xA9\x00\xE0\xA4\xAB\xE0\xA4\xBC\x00\xE0\xA4\xAF\xE0\xA4\xBC\x00"
-     "\xE0\xA4"
-     "\xB1\x00\xE0\xA4\xB4\x00\xE0\xA6\xA1\xE0\xA6\xBC\x00\xE0\xA6\xA2\xE0\xA6"
-     "\xBC\x00"
-     "\xE0\xA6\xAF\xE0\xA6\xBC\x00\xE0\xA7\x8B\x00\xE0\xA7\x8C\x00\xE0\xA8\x96"
-     "\xE0\xA8"
-     "\xBC\x00\xE0\xA8\x97\xE0\xA8\xBC\x00\xE0\xA8\x9C\xE0\xA8\xBC\x00\xE0\xA8"
-     "\xAB\xE0"
-     "\xA8\xBC\x00\xE0\xA8\xB2\xE0\xA8\xBC\x00\xE0\xA8\xB8\xE0\xA8\xBC\x00\xE0"
-     "\xAC\xA1"
-     "\xE0\xAC\xBC\x00\xE0\xAC\xA2\xE0\xAC\xBC\x00\xE0\xAD\x88\x00\xE0\xAD\x8B"
-     "\x00\xE0"
-     "\xAD\x8C\x00\xE0\xAE\x94\x00\xE0\xAF\x8A\x00\xE0\xAF\x8B\x00\xE0\xAF\x8C"
-     "\x00\xE0"
-     "\xB1\x88\x00\xE0\xB3\x80\x00\xE0\xB3\x87\x00\xE0\xB3\x88\x00\xE0\xB3\x8A"
-     "\x00\xE0"
-     "\xB3\x8B\x00\xE0\xB5\x8A\x00\xE0\xB5\x8B\x00\xE0\xB5\x8C\x00\xE0\xB7\x9A"
-     "\x00\xE0"
-     "\xB7\x9C\x00\xE0\xB7\x9D\x00\xE0\xB7\x9E\x00\xE0\xB9\x8D\xE0\xB8\xB2\x00"
-     "\xE0\xBA"
-     "\xAB\xE0\xBA\x99\x00\xE0\xBA\xAB\xE0\xBA\xA1\x00\xE0\xBB\x8D\xE0\xBA\xB2"
-     "\x00\xE0"
-     "\xBC\x8B\x00\xE0\xBD\x80\xE0\xBE\xB5\x00\xE0\xBD\x82\xE0\xBE\xB7\x00\xE0"
-     "\xBD\x8C"
-     "\xE0\xBE\xB7\x00\xE0\xBD\x91\xE0\xBE\xB7\x00\xE0\xBD\x96\xE0\xBE\xB7\x00"
-     "\xE0\xBD"
-     "\x9B\xE0\xBE\xB7\x00\xE0\xBD\xB1\xE0\xBD\xB2\x00\xE0\xBD\xB1\xE0\xBD\xB4"
-     "\x00\xE0"
-     "\xBD\xB1\xE0\xBE\x80\x00\xE0\xBE\x90\xE0\xBE\xB5\x00\xE0\xBE\x92\xE0\xBE"
-     "\xB7\x00"
-     "\xE0\xBE\x9C\xE0\xBE\xB7\x00\xE0\xBE\xA1\xE0\xBE\xB7\x00\xE0\xBE\xA6\xE0"
-     "\xBE\xB7"
-     "\x00\xE0\xBE\xAB\xE0\xBE\xB7\x00\xE0\xBE\xB2\xE0\xBD\xB1\xE0\xBE\x80\x00"
-     "\xE0\xBE"
-     "\xB2\xE0\xBE\x80\x00\xE0\xBE\xB3\xE0\xBD\xB1\xE0\xBE\x80\x00\xE0\xBE\xB3"
-     "\xE0\xBE"
-     "\x80\x00\xE1\x80\xA6\x00\xE1\x83\x90\x00\xE1\x83\x91\x00\xE1\x83\x92\x00"
-     "\xE1\x83"
-     "\x93\x00\xE1\x83\x94\x00\xE1\x83\x95\x00\xE1\x83\x96\x00\xE1\x83\x97\x00"
-     "\xE1\x83"
-     "\x98\x00\xE1\x83\x99\x00\xE1\x83\x9A\x00\xE1\x83\x9B\x00\xE1\x83\x9C\x00"
-     "\xE1\x83"
-     "\x9D\x00\xE1\x83\x9E\x00\xE1\x83\x9F\x00\xE1\x83\xA0\x00\xE1\x83\xA1\x00"
-     "\xE1\x83"
-     "\xA2\x00\xE1\x83\xA3\x00\xE1\x83\xA4\x00\xE1\x83\xA5\x00\xE1\x83\xA6\x00"
-     "\xE1\x83"
-     "\xA7\x00\xE1\x83\xA8\x00\xE1\x83\xA9\x00\xE1\x83\xAA\x00\xE1\x83\xAB\x00"
-     "\xE1\x83"
-     "\xAC\x00\xE1\x83\xAD\x00\xE1\x83\xAE\x00\xE1\x83\xAF\x00\xE1\x83\xB0\x00"
-     "\xE1\x83"
-     "\xB1\x00\xE1\x83\xB2\x00\xE1\x83\xB3\x00\xE1\x83\xB4\x00\xE1\x83\xB5\x00"
-     "\xE1\x83"
-     "\xB6\x00\xE1\x83\xB7\x00\xE1\x83\xB8\x00\xE1\x83\xB9\x00\xE1\x83\xBA\x00"
-     "\xE1\x83"
-     "\xBD\x00\xE1\x83\xBE\x00\xE1\x83\xBF\x00\xE1\x84\x80\x00\xE1\x84\x81\x00"
-     "\xE1\x84"
-     "\x82\x00\xE1\x84\x83\x00\xE1\x84\x84\x00\xE1\x84\x85\x00\xE1\x84\x86\x00"
-     "\xE1\x84"
-     "\x87\x00\xE1\x84\x88\x00\xE1\x84\x89\x00\xE1\x84\x8A\x00\xE1\x84\x8B\x00"
-     "\xE1\x84"
-     "\x8C\x00\xE1\x84\x8D\x00\xE1\x84\x8E\x00\xE1\x84\x8F\x00\xE1\x84\x90\x00"
-     "\xE1\x84"
-     "\x91\x00\xE1\x84\x92\x00\xE1\x84\x94\x00\xE1\x84\x95\x00\xE1\x84\x9A\x00"
-     "\xE1\x84"
-     "\x9C\x00\xE1\x84\x9D\x00\xE1\x84\x9E\x00\xE1\x84\xA0\x00\xE1\x84\xA1\x00"
-     "\xE1\x84"
-     "\xA2\x00\xE1\x84\xA3\x00\xE1\x84\xA7\x00\xE1\x84\xA9\x00\xE1\x84\xAB\x00"
-     "\xE1\x84"
-     "\xAC\x00\xE1\x84\xAD\x00\xE1\x84\xAE\x00\xE1\x84\xAF\x00\xE1\x84\xB2\x00"
-     "\xE1\x84"
-     "\xB6\x00\xE1\x85\x80\x00\xE1\x85\x87\x00\xE1\x85\x8C\x00\xE1\x85\x97\x00"
-     "\xE1\x85"
-     "\x98\x00\xE1\x85\x99\x00\xE1\x85\xA0\x00\xE1\x85\xA1\x00\xE1\x85\xA2\x00"
-     "\xE1\x85"
-     "\xA3\x00\xE1\x85\xA4\x00\xE1\x85\xA5\x00\xE1\x85\xA6\x00\xE1\x85\xA7\x00"
-     "\xE1\x85"
-     "\xA8\x00\xE1\x85\xA9\x00\xE1\x85\xAA\x00\xE1\x85\xAB\x00\xE1\x85\xAC\x00"
-     "\xE1\x85"
-     "\xAD\x00\xE1\x85\xAE\x00\xE1\x85\xAF\x00\xE1\x85\xB0\x00\xE1\x85\xB1\x00"
-     "\xE1\x85"
-     "\xB2\x00\xE1\x85\xB3\x00\xE1\x85\xB4\x00\xE1\x85\xB5\x00\xE1\x86\x84\x00"
-     "\xE1\x86"
-     "\x85\x00\xE1\x86\x88\x00\xE1\x86\x91\x00\xE1\x86\x92\x00\xE1\x86\x94\x00"
-     "\xE1\x86"
-     "\x9E\x00\xE1\x86\xA1\x00\xE1\x86\xAA\x00\xE1\x86\xAC\x00\xE1\x86\xAD\x00"
-     "\xE1\x86"
-     "\xB0\x00\xE1\x86\xB1\x00\xE1\x86\xB2\x00\xE1\x86\xB3\x00\xE1\x86\xB4\x00"
-     "\xE1\x86"
-     "\xB5\x00\xE1\x87\x87\x00\xE1\x87\x88\x00\xE1\x87\x8C\x00\xE1\x87\x8E\x00"
-     "\xE1\x87"
-     "\x93\x00\xE1\x87\x97\x00\xE1\x87\x99\x00\xE1\x87\x9D\x00\xE1\x87\x9F\x00"
-     "\xE1\x87"
-     "\xB1\x00\xE1\x87\xB2\x00\xE1\x8E\xA0\x00\xE1\x8E\xA1\x00\xE1\x8E\xA2\x00"
-     "\xE1\x8E"
-     "\xA3\x00\xE1\x8E\xA4\x00\xE1\x8E\xA5\x00\xE1\x8E\xA6\x00\xE1\x8E\xA7\x00"
-     "\xE1\x8E"
-     "\xA8\x00\xE1\x8E\xA9\x00\xE1\x8E\xAA\x00\xE1\x8E\xAB\x00\xE1\x8E\xAC\x00"
-     "\xE1\x8E"
-     "\xAD\x00\xE1\x8E\xAE\x00\xE1\x8E\xAF\x00\xE1\x8E\xB0\x00\xE1\x8E\xB1\x00"
-     "\xE1\x8E"
-     "\xB2\x00\xE1\x8E\xB3\x00\xE1\x8E\xB4\x00\xE1\x8E\xB5\x00\xE1\x8E\xB6\x00"
-     "\xE1\x8E"
-     "\xB7\x00\xE1\x8E\xB8\x00\xE1\x8E\xB9\x00\xE1\x8E\xBA\x00\xE1\x8E\xBB\x00"
-     "\xE1\x8E"
-     "\xBC\x00\xE1\x8E\xBD\x00\xE1\x8E\xBE\x00\xE1\x8E\xBF\x00\xE1\x8F\x80\x00"
-     "\xE1\x8F"
-     "\x81\x00\xE1\x8F\x82\x00\xE1\x8F\x83\x00\xE1\x8F\x84\x00\xE1\x8F\x85\x00"
-     "\xE1\x8F"
-     "\x86\x00\xE1\x8F\x87\x00\xE1\x8F\x88\x00\xE1\x8F\x89\x00\xE1\x8F\x8A\x00"
-     "\xE1\x8F"
-     "\x8B\x00\xE1\x8F\x8C\x00\xE1\x8F\x8D\x00\xE1\x8F\x8E\x00\xE1\x8F\x8F\x00"
-     "\xE1\x8F"
-     "\x90\x00\xE1\x8F\x91\x00\xE1\x8F\x92\x00\xE1\x8F\x93\x00\xE1\x8F\x94\x00"
-     "\xE1\x8F"
-     "\x95\x00\xE1\x8F\x96\x00\xE1\x8F\x97\x00\xE1\x8F\x98\x00\xE1\x8F\x99\x00"
-     "\xE1\x8F"
-     "\x9A\x00\xE1\x8F\x9B\x00\xE1\x8F\x9C\x00\xE1\x8F\x9D\x00\xE1\x8F\x9E\x00"
-     "\xE1\x8F"
-     "\x9F\x00\xE1\x8F\xA0\x00\xE1\x8F\xA1\x00\xE1\x8F\xA2\x00\xE1\x8F\xA3\x00"
-     "\xE1\x8F"
-     "\xA4\x00\xE1\x8F\xA5\x00\xE1\x8F\xA6\x00\xE1\x8F\xA7\x00\xE1\x8F\xA8\x00"
-     "\xE1\x8F"
-     "\xA9\x00\xE1\x8F\xAA\x00\xE1\x8F\xAB\x00\xE1\x8F\xAC\x00\xE1\x8F\xAD\x00"
-     "\xE1\x8F"
-     "\xAE\x00\xE1\x8F\xAF\x00\xE1\x8F\xB0\x00\xE1\x8F\xB1\x00\xE1\x8F\xB2\x00"
-     "\xE1\x8F"
-     "\xB3\x00\xE1\x8F\xB4\x00\xE1\x8F\xB5\x00\xE1\xAC\x86\x00\xE1\xAC\x88\x00"
-     "\xE1\xAC"
-     "\x8A\x00\xE1\xAC\x8C\x00\xE1\xAC\x8E\x00\xE1\xAC\x92\x00\xE1\xAC\xBB\x00"
-     "\xE1\xAC"
-     "\xBD\x00\xE1\xAD\x80\x00\xE1\xAD\x81\x00\xE1\xAD\x83\x00\xE1\xB4\x82\x00"
-     "\xE1\xB4"
-     "\x96\x00\xE1\xB4\x97\x00\xE1\xB4\x9C\x00\xE1\xB4\x9D\x00\xE1\xB4\xA5\x00"
-     "\xE1\xB5"
-     "\xB9\x00\xE1\xB5\xBB\x00\xE1\xB5\xBD\x00\xE1\xB6\x85\x00\xE1\xB6\x8E\x00"
-     "\xE1\xB6"
-     "\x91\x00\xE1\xB8\x81\x00\xE1\xB8\x83\x00\xE1\xB8\x85\x00\xE1\xB8\x87\x00"
-     "\xE1\xB8"
-     "\x89\x00\xE1\xB8\x8B\x00\xE1\xB8\x8D\x00\xE1\xB8\x8F\x00\xE1\xB8\x91\x00"
-     "\xE1\xB8"
-     "\x93\x00\xE1\xB8\x95\x00\xE1\xB8\x97\x00\xE1\xB8\x99\x00\xE1\xB8\x9B\x00"
-     "\xE1\xB8"
-     "\x9D\x00\xE1\xB8\x9F\x00\xE1\xB8\xA1\x00\xE1\xB8\xA3\x00\xE1\xB8\xA5\x00"
-     "\xE1\xB8"
-     "\xA7\x00\xE1\xB8\xA9\x00\xE1\xB8\xAB\x00\xE1\xB8\xAD\x00\xE1\xB8\xAF\x00"
-     "\xE1\xB8"
-     "\xB1\x00\xE1\xB8\xB3\x00\xE1\xB8\xB5\x00\xE1\xB8\xB7\x00\xE1\xB8\xB9\x00"
-     "\xE1\xB8"
-     "\xBB\x00\xE1\xB8\xBD\x00\xE1\xB8\xBF\x00\xE1\xB9\x81\x00\xE1\xB9\x83\x00"
-     "\xE1\xB9"
-     "\x85\x00\xE1\xB9\x87\x00\xE1\xB9\x89\x00\xE1\xB9\x8B\x00\xE1\xB9\x8D\x00"
-     "\xE1\xB9"
-     "\x8F\x00\xE1\xB9\x91\x00\xE1\xB9\x93\x00\xE1\xB9\x95\x00\xE1\xB9\x97\x00"
-     "\xE1\xB9"
-     "\x99\x00\xE1\xB9\x9B\x00\xE1\xB9\x9D\x00\xE1\xB9\x9F\x00\xE1\xB9\xA1\x00"
-     "\xE1\xB9"
-     "\xA3\x00\xE1\xB9\xA5\x00\xE1\xB9\xA7\x00\xE1\xB9\xA9\x00\xE1\xB9\xAB\x00"
-     "\xE1\xB9"
-     "\xAD\x00\xE1\xB9\xAF\x00\xE1\xB9\xB1\x00\xE1\xB9\xB3\x00\xE1\xB9\xB5\x00"
-     "\xE1\xB9"
-     "\xB7\x00\xE1\xB9\xB9\x00\xE1\xB9\xBB\x00\xE1\xB9\xBD\x00\xE1\xB9\xBF\x00"
-     "\xE1\xBA"
-     "\x81\x00\xE1\xBA\x83\x00\xE1\xBA\x85\x00\xE1\xBA\x87\x00\xE1\xBA\x89\x00"
-     "\xE1\xBA"
-     "\x8B\x00\xE1\xBA\x8D\x00\xE1\xBA\x8F\x00\xE1\xBA\x91\x00\xE1\xBA\x93\x00"
-     "\xE1\xBA"
-     "\x95\x00\xE1\xBA\x96\x00\xE1\xBA\x97\x00\xE1\xBA\x98\x00\xE1\xBA\x99\x00"
-     "\xE1\xBA"
-     "\xA1\x00\xE1\xBA\xA3\x00\xE1\xBA\xA5\x00\xE1\xBA\xA7\x00\xE1\xBA\xA9\x00"
-     "\xE1\xBA"
-     "\xAB\x00\xE1\xBA\xAD\x00\xE1\xBA\xAF\x00\xE1\xBA\xB1\x00\xE1\xBA\xB3\x00"
-     "\xE1\xBA"
-     "\xB5\x00\xE1\xBA\xB7\x00\xE1\xBA\xB9\x00\xE1\xBA\xBB\x00\xE1\xBA\xBD\x00"
-     "\xE1\xBA"
-     "\xBF\x00\xE1\xBB\x81\x00\xE1\xBB\x83\x00\xE1\xBB\x85\x00\xE1\xBB\x87\x00"
-     "\xE1\xBB"
-     "\x89\x00\xE1\xBB\x8B\x00\xE1\xBB\x8D\x00\xE1\xBB\x8F\x00\xE1\xBB\x91\x00"
-     "\xE1\xBB"
-     "\x93\x00\xE1\xBB\x95\x00\xE1\xBB\x97\x00\xE1\xBB\x99\x00\xE1\xBB\x9B\x00"
-     "\xE1\xBB"
-     "\x9D\x00\xE1\xBB\x9F\x00\xE1\xBB\xA1\x00\xE1\xBB\xA3\x00\xE1\xBB\xA5\x00"
-     "\xE1\xBB"
-     "\xA7\x00\xE1\xBB\xA9\x00\xE1\xBB\xAB\x00\xE1\xBB\xAD\x00\xE1\xBB\xAF\x00"
-     "\xE1\xBB"
-     "\xB1\x00\xE1\xBB\xB3\x00\xE1\xBB\xB5\x00\xE1\xBB\xB7\x00\xE1\xBB\xB9\x00"
-     "\xE1\xBB"
-     "\xBB\x00\xE1\xBB\xBD\x00\xE1\xBB\xBF\x00\xE1\xBC\x80\x00\xE1\xBC\x81\x00"
-     "\xE1\xBC"
-     "\x82\x00\xE1\xBC\x83\x00\xE1\xBC\x84\x00\xE1\xBC\x85\x00\xE1\xBC\x86\x00"
-     "\xE1\xBC"
-     "\x87\x00\xE1\xBC\x90\x00\xE1\xBC\x91\x00\xE1\xBC\x92\x00\xE1\xBC\x93\x00"
-     "\xE1\xBC"
-     "\x94\x00\xE1\xBC\x95\x00\xE1\xBC\xA0\x00\xE1\xBC\xA1\x00\xE1\xBC\xA2\x00"
-     "\xE1\xBC"
-     "\xA3\x00\xE1\xBC\xA4\x00\xE1\xBC\xA5\x00\xE1\xBC\xA6\x00\xE1\xBC\xA7\x00"
-     "\xE1\xBC"
-     "\xB0\x00\xE1\xBC\xB1\x00\xE1\xBC\xB2\x00\xE1\xBC\xB3\x00\xE1\xBC\xB4\x00"
-     "\xE1\xBC"
-     "\xB5\x00\xE1\xBC\xB6\x00\xE1\xBC\xB7\x00\xE1\xBD\x80\x00\xE1\xBD\x81\x00"
-     "\xE1\xBD"
-     "\x82\x00\xE1\xBD\x83\x00\xE1\xBD\x84\x00\xE1\xBD\x85\x00\xE1\xBD\x90\x00"
-     "\xE1\xBD"
-     "\x91\x00\xE1\xBD\x92\x00\xE1\xBD\x93\x00\xE1\xBD\x94\x00\xE1\xBD\x95\x00"
-     "\xE1\xBD"
-     "\x96\x00\xE1\xBD\x97\x00\xE1\xBD\xA0\x00\xE1\xBD\xA1\x00\xE1\xBD\xA2\x00"
-     "\xE1\xBD"
-     "\xA3\x00\xE1\xBD\xA4\x00\xE1\xBD\xA5\x00\xE1\xBD\xA6\x00\xE1\xBD\xA7\x00"
-     "\xE1\xBD"
-     "\xB0\x00\xE1\xBD\xB2\x00\xE1\xBD\xB4\x00\xE1\xBD\xB6\x00\xE1\xBD\xB8\x00"
-     "\xE1\xBD"
-     "\xBA\x00\xE1\xBD\xBC\x00\xE1\xBE\x80\x00\xE1\xBE\x81\x00\xE1\xBE\x82\x00"
-     "\xE1\xBE"
-     "\x83\x00\xE1\xBE\x84\x00\xE1\xBE\x85\x00\xE1\xBE\x86\x00\xE1\xBE\x87\x00"
-     "\xE1\xBE"
-     "\x90\x00\xE1\xBE\x91\x00\xE1\xBE\x92\x00\xE1\xBE\x93\x00\xE1\xBE\x94\x00"
-     "\xE1\xBE"
-     "\x95\x00\xE1\xBE\x96\x00\xE1\xBE\x97\x00\xE1\xBE\xA0\x00\xE1\xBE\xA1\x00"
-     "\xE1\xBE"
-     "\xA2\x00\xE1\xBE\xA3\x00\xE1\xBE\xA4\x00\xE1\xBE\xA5\x00\xE1\xBE\xA6\x00"
-     "\xE1\xBE"
-     "\xA7\x00\xE1\xBE\xB0\x00\xE1\xBE\xB1\x00\xE1\xBE\xB2\x00\xE1\xBE\xB3\x00"
-     "\xE1\xBE"
-     "\xB4\x00\xE1\xBE\xB6\x00\xE1\xBE\xB7\x00\xE1\xBF\x82\x00\xE1\xBF\x83\x00"
-     "\xE1\xBF"
-     "\x84\x00\xE1\xBF\x86\x00\xE1\xBF\x87\x00\xE1\xBF\x90\x00\xE1\xBF\x91\x00"
-     "\xE1\xBF"
-     "\x92\x00\xE1\xBF\x96\x00\xE1\xBF\x97\x00\xE1\xBF\xA0\x00\xE1\xBF\xA1\x00"
-     "\xE1\xBF"
-     "\xA2\x00\xE1\xBF\xA4\x00\xE1\xBF\xA5\x00\xE1\xBF\xA6\x00\xE1\xBF\xA7\x00"
-     "\xE1\xBF"
-     "\xB2\x00\xE1\xBF\xB3\x00\xE1\xBF\xB4\x00\xE1\xBF\xB6\x00\xE1\xBF\xB7\x00"
-     "\xE2\x80"
-     "\x90\x00\xE2\x80\x93\x00\xE2\x80\x94\x00\xE2\x80\xB2\xE2\x80\xB2\x00\xE2"
-     "\x80\xB2"
-     "\xE2\x80\xB2\xE2\x80\xB2\x00\xE2\x80\xB2\xE2\x80\xB2\xE2\x80\xB2\xE2\x80"
-     "\xB2\x00"
-     "\xE2\x80\xB5\xE2\x80\xB5\x00\xE2\x80\xB5\xE2\x80\xB5\xE2\x80\xB5\x00\xE2"
-     "\x82\xA9"
-     "\x00\xE2\x85\x8E\x00\xE2\x86\x84\x00\xE2\x86\x90\x00\xE2\x86\x91\x00\xE2"
-     "\x86\x92"
-     "\x00\xE2\x86\x93\x00\xE2\x86\x9A\x00\xE2\x86\x9B\x00\xE2\x86\xAE\x00\xE2"
-     "\x87\x8D"
-     "\x00\xE2\x87\x8E\x00\xE2\x87\x8F\x00\xE2\x88\x82\x00\xE2\x88\x84\x00\xE2"
-     "\x88\x87"
-     "\x00\xE2\x88\x89\x00\xE2\x88\x8C\x00\xE2\x88\x91\x00\xE2\x88\x92\x00\xE2"
-     "\x88\xA4"
-     "\x00\xE2\x88\xA6\x00\xE2\x88\xAB\xE2\x88\xAB\x00\xE2\x88\xAB\xE2\x88\xAB"
-     "\xE2\x88"
-     "\xAB\x00\xE2\x88\xAB\xE2\x88\xAB\xE2\x88\xAB\xE2\x88\xAB\x00\xE2\x88\xAE"
-     "\xE2\x88"
-     "\xAE\x00\xE2\x88\xAE\xE2\x88\xAE\xE2\x88\xAE\x00\xE2\x89\x81\x00\xE2\x89"
-     "\x84\x00"
-     "\xE2\x89\x87\x00\xE2\x89\x89\x00\xE2\x89\xA0\x00\xE2\x89\xA2\x00\xE2\x89"
-     "\xAD\x00"
-     "\xE2\x89\xAE\x00\xE2\x89\xAF\x00\xE2\x89\xB0\x00\xE2\x89\xB1\x00\xE2\x89"
-     "\xB4\x00"
-     "\xE2\x89\xB5\x00\xE2\x89\xB8\x00\xE2\x89\xB9\x00\xE2\x8A\x80\x00\xE2\x8A"
-     "\x81\x00"
-     "\xE2\x8A\x84\x00\xE2\x8A\x85\x00\xE2\x8A\x88\x00\xE2\x8A\x89\x00\xE2\x8A"
-     "\xAC\x00"
-     "\xE2\x8A\xAD\x00\xE2\x8A\xAE\x00\xE2\x8A\xAF\x00\xE2\x8B\xA0\x00\xE2\x8B"
-     "\xA1\x00"
-     "\xE2\x8B\xA2\x00\xE2\x8B\xA3\x00\xE2\x8B\xAA\x00\xE2\x8B\xAB\x00\xE2\x8B"
-     "\xAC\x00"
-     "\xE2\x8B\xAD\x00\xE2\x94\x82\x00\xE2\x96\xA0\x00\xE2\x97\x8B\x00\xE2\xA6"
-     "\x85\x00"
-     "\xE2\xA6\x86\x00\xE2\xAB\x9D\xCC\xB8\x00\xE2\xB0\xB0\x00\xE2\xB0\xB1\x00"
-     "\xE2\xB0"
-     "\xB2\x00\xE2\xB0\xB3\x00\xE2\xB0\xB4\x00\xE2\xB0\xB5\x00\xE2\xB0\xB6\x00"
-     "\xE2\xB0"
-     "\xB7\x00\xE2\xB0\xB8\x00\xE2\xB0\xB9\x00\xE2\xB0\xBA\x00\xE2\xB0\xBB\x00"
-     "\xE2\xB0"
-     "\xBC\x00\xE2\xB0\xBD\x00\xE2\xB0\xBE\x00\xE2\xB0\xBF\x00\xE2\xB1\x80\x00"
-     "\xE2\xB1"
-     "\x81\x00\xE2\xB1\x82\x00\xE2\xB1\x83\x00\xE2\xB1\x84\x00\xE2\xB1\x85\x00"
-     "\xE2\xB1"
-     "\x86\x00\xE2\xB1\x87\x00\xE2\xB1\x88\x00\xE2\xB1\x89\x00\xE2\xB1\x8A\x00"
-     "\xE2\xB1"
-     "\x8B\x00\xE2\xB1\x8C\x00\xE2\xB1\x8D\x00\xE2\xB1\x8E\x00\xE2\xB1\x8F\x00"
-     "\xE2\xB1"
-     "\x90\x00\xE2\xB1\x91\x00\xE2\xB1\x92\x00\xE2\xB1\x93\x00\xE2\xB1\x94\x00"
-     "\xE2\xB1"
-     "\x95\x00\xE2\xB1\x96\x00\xE2\xB1\x97\x00\xE2\xB1\x98\x00\xE2\xB1\x99\x00"
-     "\xE2\xB1"
-     "\x9A\x00\xE2\xB1\x9B\x00\xE2\xB1\x9C\x00\xE2\xB1\x9D\x00\xE2\xB1\x9E\x00"
-     "\xE2\xB1"
-     "\x9F\x00\xE2\xB1\xA1\x00\xE2\xB1\xA5\x00\xE2\xB1\xA6\x00\xE2\xB1\xA8\x00"
-     "\xE2\xB1"
-     "\xAA\x00\xE2\xB1\xAC\x00\xE2\xB1\xB1\x00\xE2\xB1\xB3\x00\xE2\xB1\xB6\x00"
-     "\xE2\xB2"
-     "\x81\x00\xE2\xB2\x83\x00\xE2\xB2\x85\x00\xE2\xB2\x87\x00\xE2\xB2\x89\x00"
-     "\xE2\xB2"
-     "\x8B\x00\xE2\xB2\x8D\x00\xE2\xB2\x8F\x00\xE2\xB2\x91\x00\xE2\xB2\x93\x00"
-     "\xE2\xB2"
-     "\x95\x00\xE2\xB2\x97\x00\xE2\xB2\x99\x00\xE2\xB2\x9B\x00\xE2\xB2\x9D\x00"
-     "\xE2\xB2"
-     "\x9F\x00\xE2\xB2\xA1\x00\xE2\xB2\xA3\x00\xE2\xB2\xA5\x00\xE2\xB2\xA7\x00"
-     "\xE2\xB2"
-     "\xA9\x00\xE2\xB2\xAB\x00\xE2\xB2\xAD\x00\xE2\xB2\xAF\x00\xE2\xB2\xB1\x00"
-     "\xE2\xB2"
-     "\xB3\x00\xE2\xB2\xB5\x00\xE2\xB2\xB7\x00\xE2\xB2\xB9\x00\xE2\xB2\xBB\x00"
-     "\xE2\xB2"
-     "\xBD\x00\xE2\xB2\xBF\x00\xE2\xB3\x81\x00\xE2\xB3\x83\x00\xE2\xB3\x85\x00"
-     "\xE2\xB3"
-     "\x87\x00\xE2\xB3\x89\x00\xE2\xB3\x8B\x00\xE2\xB3\x8D\x00\xE2\xB3\x8F\x00"
-     "\xE2\xB3"
-     "\x91\x00\xE2\xB3\x93\x00\xE2\xB3\x95\x00\xE2\xB3\x97\x00\xE2\xB3\x99\x00"
-     "\xE2\xB3"
-     "\x9B\x00\xE2\xB3\x9D\x00\xE2\xB3\x9F\x00\xE2\xB3\xA1\x00\xE2\xB3\xA3\x00"
-     "\xE2\xB3"
-     "\xAC\x00\xE2\xB3\xAE\x00\xE2\xB3\xB3\x00\xE2\xB4\x80\x00\xE2\xB4\x81\x00"
-     "\xE2\xB4"
-     "\x82\x00\xE2\xB4\x83\x00\xE2\xB4\x84\x00\xE2\xB4\x85\x00\xE2\xB4\x86\x00"
-     "\xE2\xB4"
-     "\x87\x00\xE2\xB4\x88\x00\xE2\xB4\x89\x00\xE2\xB4\x8A\x00\xE2\xB4\x8B\x00"
-     "\xE2\xB4"
-     "\x8C\x00\xE2\xB4\x8D\x00\xE2\xB4\x8E\x00\xE2\xB4\x8F\x00\xE2\xB4\x90\x00"
-     "\xE2\xB4"
-     "\x91\x00\xE2\xB4\x92\x00\xE2\xB4\x93\x00\xE2\xB4\x94\x00\xE2\xB4\x95\x00"
-     "\xE2\xB4"
-     "\x96\x00\xE2\xB4\x97\x00\xE2\xB4\x98\x00\xE2\xB4\x99\x00\xE2\xB4\x9A\x00"
-     "\xE2\xB4"
-     "\x9B\x00\xE2\xB4\x9C\x00\xE2\xB4\x9D\x00\xE2\xB4\x9E\x00\xE2\xB4\x9F\x00"
-     "\xE2\xB4"
-     "\xA0\x00\xE2\xB4\xA1\x00\xE2\xB4\xA2\x00\xE2\xB4\xA3\x00\xE2\xB4\xA4\x00"
-     "\xE2\xB4"
-     "\xA5\x00\xE2\xB4\xA7\x00\xE2\xB4\xAD\x00\xE2\xB5\xA1\x00\xE3\x80\x81\x00"
-     "\xE3\x80"
-     "\x82\x00\xE3\x80\x88\x00\xE3\x80\x89\x00\xE3\x80\x8A\x00\xE3\x80\x8B\x00"
-     "\xE3\x80"
-     "\x8C\x00\xE3\x80\x8D\x00\xE3\x80\x8E\x00\xE3\x80\x8F\x00\xE3\x80\x90\x00"
-     "\xE3\x80"
-     "\x91\x00\xE3\x80\x92\x00\xE3\x80\x94\x00\xE3\x80\x94\x73\xE3\x80\x95\x00"
-     "\xE3\x80"
-     "\x94\xE4\xB8\x89\xE3\x80\x95\x00\xE3\x80\x94\xE4\xBA\x8C\xE3\x80\x95\x00"
-     "\xE3\x80"
-     "\x94\xE5\x8B\x9D\xE3\x80\x95\x00\xE3\x80\x94\xE5\xAE\x89\xE3\x80\x95\x00"
-     "\xE3\x80"
-     "\x94\xE6\x89\x93\xE3\x80\x95\x00\xE3\x80\x94\xE6\x95\x97\xE3\x80\x95\x00"
-     "\xE3\x80"
-     "\x94\xE6\x9C\xAC\xE3\x80\x95\x00\xE3\x80\x94\xE7\x82\xB9\xE3\x80\x95\x00"
-     "\xE3\x80"
-     "\x94\xE7\x9B\x97\xE3\x80\x95\x00\xE3\x80\x95\x00\xE3\x80\x96\x00\xE3\x80"
-     "\x97\x00"
-     "\xE3\x81\x8C\x00\xE3\x81\x8E\x00\xE3\x81\x90\x00\xE3\x81\x92\x00\xE3\x81"
-     "\x94\x00"
-     "\xE3\x81\x96\x00\xE3\x81\x98\x00\xE3\x81\x9A\x00\xE3\x81\x9C\x00\xE3\x81"
-     "\x9E\x00"
-     "\xE3\x81\xA0\x00\xE3\x81\xA2\x00\xE3\x81\xA5\x00\xE3\x81\xA7\x00\xE3\x81"
-     "\xA9\x00"
-     "\xE3\x81\xB0\x00\xE3\x81\xB1\x00\xE3\x81\xB3\x00\xE3\x81\xB4\x00\xE3\x81"
-     "\xB6\x00"
-     "\xE3\x81\xB7\x00\xE3\x81\xB9\x00\xE3\x81\xBA\x00\xE3\x81\xBB\xE3\x81\x8B"
-     "\x00\xE3"
-     "\x81\xBC\x00\xE3\x81\xBD\x00\xE3\x82\x88\xE3\x82\x8A\x00\xE3\x82\x94\x00"
-     "\xE3\x82"
-     "\x99\x00\xE3\x82\x9A\x00\xE3\x82\x9E\x00\xE3\x82\xA1\x00\xE3\x82\xA2\x00"
-     "\xE3\x82"
-     "\xA2\xE3\x83\x91\xE3\x83\xBC\xE3\x83\x88\x00\xE3\x82\xA2\xE3\x83\xAB\xE3"
-     "\x83\x95"
-     "\xE3\x82\xA1\x00\xE3\x82\xA2\xE3\x83\xB3\xE3\x83\x9A\xE3\x82\xA2\x00\xE3"
-     "\x82\xA2"
-     "\xE3\x83\xBC\xE3\x83\xAB\x00\xE3\x82\xA3\x00\xE3\x82\xA4\x00\xE3\x82\xA4"
-     "\xE3\x83"
-     "\x8B\xE3\x83\xB3\xE3\x82\xB0\x00\xE3\x82\xA4\xE3\x83\xB3\xE3\x83\x81\x00"
-     "\xE3\x82"
-     "\xA5\x00\xE3\x82\xA6\x00\xE3\x82\xA6\xE3\x82\xA9\xE3\x83\xB3\x00\xE3\x82"
-     "\xA7\x00"
-     "\xE3\x82\xA8\x00\xE3\x82\xA8\xE3\x82\xB9\xE3\x82\xAF\xE3\x83\xBC\xE3\x83"
-     "\x89\x00"
-     "\xE3\x82\xA8\xE3\x83\xBC\xE3\x82\xAB\xE3\x83\xBC\x00\xE3\x82\xA9\x00\xE3"
-     "\x82\xAA"
-     "\x00\xE3\x82\xAA\xE3\x83\xB3\xE3\x82\xB9\x00\xE3\x82\xAA\xE3\x83\xBC\xE3"
-     "\x83\xA0"
-     "\x00\xE3\x82\xAB\x00\xE3\x82\xAB\xE3\x82\xA4\xE3\x83\xAA\x00\xE3\x82\xAB"
-     "\xE3\x83"
-     "\xA9\xE3\x83\x83\xE3\x83\x88\x00\xE3\x82\xAB\xE3\x83\xAD\xE3\x83\xAA\xE3"
-     "\x83\xBC"
-     "\x00\xE3\x82\xAC\x00\xE3\x82\xAC\xE3\x83\xAD\xE3\x83\xB3\x00\xE3\x82\xAC"
-     "\xE3\x83"
-     "\xB3\xE3\x83\x9E\x00\xE3\x82\xAD\x00\xE3\x82\xAD\xE3\x83\xA5\xE3\x83\xAA"
-     "\xE3\x83"
-     "\xBC\x00\xE3\x82\xAD\xE3\x83\xAD\x00\xE3\x82\xAD\xE3\x83\xAD\xE3\x82\xB0"
-     "\xE3\x83"
-     "\xA9\xE3\x83\xA0\x00\xE3\x82\xAD\xE3\x83\xAD\xE3\x83\xA1\xE3\x83\xBC\xE3"
-     "\x83\x88"
-     "\xE3\x83\xAB\x00\xE3\x82\xAD\xE3\x83\xAD\xE3\x83\xAF\xE3\x83\x83\xE3\x83"
-     "\x88\x00"
-     "\xE3\x82\xAE\x00\xE3\x82\xAE\xE3\x82\xAC\x00\xE3\x82\xAE\xE3\x83\x8B\xE3"
-     "\x83\xBC"
-     "\x00\xE3\x82\xAE\xE3\x83\xAB\xE3\x83\x80\xE3\x83\xBC\x00\xE3\x82\xAF\x00"
-     "\xE3\x82"
-     "\xAF\xE3\x83\xAB\xE3\x82\xBC\xE3\x82\xA4\xE3\x83\xAD\x00\xE3\x82\xAF\xE3"
-     "\x83\xAD"
-     "\xE3\x83\xBC\xE3\x83\x8D\x00\xE3\x82\xB0\x00\xE3\x82\xB0\xE3\x83\xA9\xE3"
-     "\x83\xA0"
-     "\x00\xE3\x82\xB0\xE3\x83\xA9\xE3\x83\xA0\xE3\x83\x88\xE3\x83\xB3\x00\xE3"
-     "\x82\xB1"
-     "\x00\xE3\x82\xB1\xE3\x83\xBC\xE3\x82\xB9\x00\xE3\x82\xB2\x00\xE3\x82\xB3"
-     "\x00\xE3"
-     "\x82\xB3\xE3\x82\xB3\x00\xE3\x82\xB3\xE3\x83\x88\x00\xE3\x82\xB3\xE3\x83"
-     "\xAB\xE3"
-     "\x83\x8A\x00\xE3\x82\xB3\xE3\x83\xBC\xE3\x83\x9D\x00\xE3\x82\xB4\x00\xE3"
-     "\x82\xB5"
-     "\x00\xE3\x82\xB5\xE3\x82\xA4\xE3\x82\xAF\xE3\x83\xAB\x00\xE3\x82\xB5\xE3"
-     "\x83\xB3"
-     "\xE3\x83\x81\xE3\x83\xBC\xE3\x83\xA0\x00\xE3\x82\xB6\x00\xE3\x82\xB7\x00"
-     "\xE3\x82"
-     "\xB7\xE3\x83\xAA\xE3\x83\xB3\xE3\x82\xB0\x00\xE3\x82\xB8\x00\xE3\x82\xB9"
-     "\x00\xE3"
-     "\x82\xBA\x00\xE3\x82\xBB\x00\xE3\x82\xBB\xE3\x83\xB3\xE3\x83\x81\x00\xE3"
-     "\x82\xBB"
-     "\xE3\x83\xB3\xE3\x83\x88\x00\xE3\x82\xBC\x00\xE3\x82\xBD\x00\xE3\x82\xBE"
-     "\x00\xE3"
-     "\x82\xBF\x00\xE3\x83\x80\x00\xE3\x83\x80\xE3\x83\xBC\xE3\x82\xB9\x00\xE3"
-     "\x83\x81"
-     "\x00\xE3\x83\x82\x00\xE3\x83\x83\x00\xE3\x83\x84\x00\xE3\x83\x85\x00\xE3"
-     "\x83\x86"
-     "\x00\xE3\x83\x87\x00\xE3\x83\x87\xE3\x82\xB7\x00\xE3\x83\x88\x00\xE3\x83"
-     "\x88\xE3"
-     "\x83\xB3\x00\xE3\x83\x89\x00\xE3\x83\x89\xE3\x83\xAB\x00\xE3\x83\x8A\x00"
-     "\xE3\x83"
-     "\x8A\xE3\x83\x8E\x00\xE3\x83\x8B\x00\xE3\x83\x8C\x00\xE3\x83\x8D\x00\xE3"
-     "\x83\x8E"
-     "\x00\xE3\x83\x8E\xE3\x83\x83\xE3\x83\x88\x00\xE3\x83\x8F\x00\xE3\x83\x8F"
-     "\xE3\x82"
-     "\xA4\xE3\x83\x84\x00\xE3\x83\x90\x00\xE3\x83\x90\xE3\x83\xBC\xE3\x83\xAC"
-     "\xE3\x83"
-     "\xAB\x00\xE3\x83\x91\x00\xE3\x83\x91\xE3\x83\xBC\xE3\x82\xBB\xE3\x83\xB3"
-     "\xE3\x83"
-     "\x88\x00\xE3\x83\x91\xE3\x83\xBC\xE3\x83\x84\x00\xE3\x83\x92\x00\xE3\x83"
-     "\x93\x00"
-     "\xE3\x83\x93\xE3\x83\xAB\x00\xE3\x83\x94\x00\xE3\x83\x94\xE3\x82\xA2\xE3"
-     "\x82\xB9"
-     "\xE3\x83\x88\xE3\x83\xAB\x00\xE3\x83\x94\xE3\x82\xAF\xE3\x83\xAB\x00\xE3"
-     "\x83\x94"
-     "\xE3\x82\xB3\x00\xE3\x83\x95\x00\xE3\x83\x95\xE3\x82\xA1\xE3\x83\xA9\xE3"
-     "\x83\x83"
-     "\xE3\x83\x89\x00\xE3\x83\x95\xE3\x82\xA3\xE3\x83\xBC\xE3\x83\x88\x00\xE3"
-     "\x83\x95"
-     "\xE3\x83\xA9\xE3\x83\xB3\x00\xE3\x83\x96\x00\xE3\x83\x96\xE3\x83\x83\xE3"
-     "\x82\xB7"
-     "\xE3\x82\xA7\xE3\x83\xAB\x00\xE3\x83\x97\x00\xE3\x83\x98\x00\xE3\x83\x98"
-     "\xE3\x82"
-     "\xAF\xE3\x82\xBF\xE3\x83\xBC\xE3\x83\xAB\x00\xE3\x83\x98\xE3\x83\xAB\xE3"
-     "\x83\x84"
-     "\x00\xE3\x83\x99\x00\xE3\x83\x99\xE3\x83\xBC\xE3\x82\xBF\x00\xE3\x83\x9A"
-     "\x00\xE3"
-     "\x83\x9A\xE3\x82\xBD\x00\xE3\x83\x9A\xE3\x83\x8B\xE3\x83\x92\x00\xE3\x83"
-     "\x9A\xE3"
-     "\x83\xB3\xE3\x82\xB9\x00\xE3\x83\x9A\xE3\x83\xBC\xE3\x82\xB8\x00\xE3\x83"
-     "\x9B\x00"
-     "\xE3\x83\x9B\xE3\x83\xB3\x00\xE3\x83\x9B\xE3\x83\xBC\xE3\x83\xAB\x00\xE3"
-     "\x83\x9B"
-     "\xE3\x83\xBC\xE3\x83\xB3\x00\xE3\x83\x9C\x00\xE3\x83\x9C\xE3\x83\xAB\xE3"
-     "\x83\x88"
-     "\x00\xE3\x83\x9D\x00\xE3\x83\x9D\xE3\x82\xA4\xE3\x83\xB3\xE3\x83\x88\x00"
-     "\xE3\x83"
-     "\x9D\xE3\x83\xB3\xE3\x83\x89\x00\xE3\x83\x9E\x00\xE3\x83\x9E\xE3\x82\xA4"
-     "\xE3\x82"
-     "\xAF\xE3\x83\xAD\x00\xE3\x83\x9E\xE3\x82\xA4\xE3\x83\xAB\x00\xE3\x83\x9E"
-     "\xE3\x83"
-     "\x83\xE3\x83\x8F\x00\xE3\x83\x9E\xE3\x83\xAB\xE3\x82\xAF\x00\xE3\x83\x9E"
-     "\xE3\x83"
-     "\xB3\xE3\x82\xB7\xE3\x83\xA7\xE3\x83\xB3\x00\xE3\x83\x9F\x00\xE3\x83\x9F"
-     "\xE3\x82"
-     "\xAF\xE3\x83\xAD\xE3\x83\xB3\x00\xE3\x83\x9F\xE3\x83\xAA\x00\xE3\x83\x9F"
-     "\xE3\x83"
-     "\xAA\xE3\x83\x90\xE3\x83\xBC\xE3\x83\xAB\x00\xE3\x83\xA0\x00\xE3\x83\xA1"
-     "\x00\xE3"
-     "\x83\xA1\xE3\x82\xAC\x00\xE3\x83\xA1\xE3\x82\xAC\xE3\x83\x88\xE3\x83\xB3"
-     "\x00\xE3"
-     "\x83\xA1\xE3\x83\xBC\xE3\x83\x88\xE3\x83\xAB\x00\xE3\x83\xA2\x00\xE3\x83"
-     "\xA3\x00"
-     "\xE3\x83\xA4\x00\xE3\x83\xA4\xE3\x83\xBC\xE3\x83\x89\x00\xE3\x83\xA4\xE3"
-     "\x83\xBC"
-     "\xE3\x83\xAB\x00\xE3\x83\xA5\x00\xE3\x83\xA6\x00\xE3\x83\xA6\xE3\x82\xA2"
-     "\xE3\x83"
-     "\xB3\x00\xE3\x83\xA7\x00\xE3\x83\xA8\x00\xE3\x83\xA9\x00\xE3\x83\xAA\x00"
-     "\xE3\x83"
-     "\xAA\xE3\x83\x83\xE3\x83\x88\xE3\x83\xAB\x00\xE3\x83\xAA\xE3\x83\xA9\x00"
-     "\xE3\x83"
-     "\xAB\x00\xE3\x83\xAB\xE3\x83\x94\xE3\x83\xBC\x00\xE3\x83\xAB\xE3\x83\xBC"
-     "\xE3\x83"
-     "\x96\xE3\x83\xAB\x00\xE3\x83\xAC\x00\xE3\x83\xAC\xE3\x83\xA0\x00\xE3\x83"
-     "\xAC\xE3"
-     "\x83\xB3\xE3\x83\x88\xE3\x82\xB2\xE3\x83\xB3\x00\xE3\x83\xAD\x00\xE3\x83"
-     "\xAF\x00"
-     "\xE3\x83\xAF\xE3\x83\x83\xE3\x83\x88\x00\xE3\x83\xB0\x00\xE3\x83\xB1\x00"
-     "\xE3\x83"
-     "\xB2\x00\xE3\x83\xB3\x00\xE3\x83\xB4\x00\xE3\x83\xB7\x00\xE3\x83\xB8\x00"
-     "\xE3\x83"
-     "\xB9\x00\xE3\x83\xBA\x00\xE3\x83\xBB\x00\xE3\x83\xBC\x00\xE3\x83\xBE\x00"
-     "\xE3\x92"
-     "\x9E\x00\xE3\x92\xB9\x00\xE3\x92\xBB\x00\xE3\x93\x9F\x00\xE3\x94\x95\x00"
-     "\xE3\x9B"
-     "\xAE\x00\xE3\x9B\xBC\x00\xE3\x9E\x81\x00\xE3\xA0\xAF\x00\xE3\xA1\xA2\x00"
-     "\xE3\xA1"
-     "\xBC\x00\xE3\xA3\x87\x00\xE3\xA3\xA3\x00\xE3\xA4\x9C\x00\xE3\xA4\xBA\x00"
-     "\xE3\xA8"
-     "\xAE\x00\xE3\xA9\xAC\x00\xE3\xAB\xA4\x00\xE3\xAC\x88\x00\xE3\xAC\x99\x00"
-     "\xE3\xAD"
-     "\x89\x00\xE3\xAE\x9D\x00\xE3\xB0\x98\x00\xE3\xB1\x8E\x00\xE3\xB4\xB3\x00"
-     "\xE3\xB6"
-     "\x96\x00\xE3\xBA\xAC\x00\xE3\xBA\xB8\x00\xE3\xBC\x9B\x00\xE3\xBF\xBC\x00"
-     "\xE4\x80"
-     "\x88\x00\xE4\x80\x98\x00\xE4\x80\xB9\x00\xE4\x81\x86\x00\xE4\x82\x96\x00"
-     "\xE4\x83"
-     "\xA3\x00\xE4\x84\xAF\x00\xE4\x88\x82\x00\xE4\x88\xA7\x00\xE4\x8A\xA0\x00"
-     "\xE4\x8C"
-     "\x81\x00\xE4\x8C\xB4\x00\xE4\x8D\x99\x00\xE4\x8F\x95\x00\xE4\x8F\x99\x00"
-     "\xE4\x90"
-     "\x8B\x00\xE4\x91\xAB\x00\xE4\x94\xAB\x00\xE4\x95\x9D\x00\xE4\x95\xA1\x00"
-     "\xE4\x95"
-     "\xAB\x00\xE4\x97\x97\x00\xE4\x97\xB9\x00\xE4\x98\xB5\x00\xE4\x9A\xBE\x00"
-     "\xE4\x9B"
-     "\x87\x00\xE4\xA6\x95\x00\xE4\xA7\xA6\x00\xE4\xA9\xAE\x00\xE4\xA9\xB6\x00"
-     "\xE4\xAA"
-     "\xB2\x00\xE4\xAC\xB3\x00\xE4\xAF\x8E\x00\xE4\xB3\x8E\x00\xE4\xB3\xAD\x00"
-     "\xE4\xB3"
-     "\xB8\x00\xE4\xB5\x96\x00\xE4\xB8\x80\x00\xE4\xB8\x81\x00\xE4\xB8\x83\x00"
-     "\xE4\xB8"
-     "\x89\x00\xE4\xB8\x8A\x00\xE4\xB8\x8B\x00\xE4\xB8\x8D\x00\xE4\xB8\x99\x00"
-     "\xE4\xB8"
-     "\xA6\x00\xE4\xB8\xA8\x00\xE4\xB8\xAD\x00\xE4\xB8\xB2\x00\xE4\xB8\xB6\x00"
-     "\xE4\xB8"
-     "\xB8\x00\xE4\xB8\xB9\x00\xE4\xB8\xBD\x00\xE4\xB8\xBF\x00\xE4\xB9\x81\x00"
-     "\xE4\xB9"
-     "\x99\x00\xE4\xB9\x9D\x00\xE4\xBA\x82\x00\xE4\xBA\x85\x00\xE4\xBA\x86\x00"
-     "\xE4\xBA"
-     "\x8C\x00\xE4\xBA\x94\x00\xE4\xBA\xA0\x00\xE4\xBA\xA4\x00\xE4\xBA\xAE\x00"
-     "\xE4\xBA"
-     "\xBA\x00\xE4\xBB\x80\x00\xE4\xBB\x8C\x00\xE4\xBB\xA4\x00\xE4\xBB\xA4\xE5"
-     "\x92\x8C"
-     "\x00\xE4\xBC\x81\x00\xE4\xBC\x91\x00\xE4\xBD\xA0\x00\xE4\xBE\x80\x00\xE4"
-     "\xBE\x86"
-     "\x00\xE4\xBE\x8B\x00\xE4\xBE\xAE\x00\xE4\xBE\xBB\x00\xE4\xBE\xBF\x00\xE5"
-     "\x80\x82"
-     "\x00\xE5\x80\xAB\x00\xE5\x81\xBA\x00\xE5\x82\x99\x00\xE5\x83\x8F\x00\xE5"
-     "\x83\x9A"
-     "\x00\xE5\x83\xA7\x00\xE5\x84\xAA\x00\xE5\x84\xBF\x00\xE5\x85\x80\x00\xE5"
-     "\x85\x85"
-     "\x00\xE5\x85\x8D\x00\xE5\x85\x94\x00\xE5\x85\xA4\x00\xE5\x85\xA5\x00\xE5"
-     "\x85\xA7"
-     "\x00\xE5\x85\xA8\x00\xE5\x85\xA9\x00\xE5\x85\xAB\x00\xE5\x85\xAD\x00\xE5"
-     "\x85\xB7"
-     "\x00\xE5\x86\x80\x00\xE5\x86\x82\x00\xE5\x86\x8D\x00\xE5\x86\x92\x00\xE5"
-     "\x86\x95"
-     "\x00\xE5\x86\x96\x00\xE5\x86\x97\x00\xE5\x86\x99\x00\xE5\x86\xA4\x00\xE5"
-     "\x86\xAB"
-     "\x00\xE5\x86\xAC\x00\xE5\x86\xB5\x00\xE5\x86\xB7\x00\xE5\x87\x89\x00\xE5"
-     "\x87\x8C"
-     "\x00\xE5\x87\x9C\x00\xE5\x87\x9E\x00\xE5\x87\xA0\x00\xE5\x87\xB5\x00\xE5"
-     "\x88\x80"
-     "\x00\xE5\x88\x83\x00\xE5\x88\x87\x00\xE5\x88\x97\x00\xE5\x88\x9D\x00\xE5"
-     "\x88\xA9"
-     "\x00\xE5\x88\xBA\x00\xE5\x88\xBB\x00\xE5\x89\x86\x00\xE5\x89\x8D\x00\xE5"
-     "\x89\xB2"
-     "\x00\xE5\x89\xB7\x00\xE5\x8A\x89\x00\xE5\x8A\x9B\x00\xE5\x8A\xA3\x00\xE5"
-     "\x8A\xB3"
-     "\x00\xE5\x8A\xB4\x00\xE5\x8B\x87\x00\xE5\x8B\x89\x00\xE5\x8B\x92\x00\xE5"
-     "\x8B\x9E"
-     "\x00\xE5\x8B\xA4\x00\xE5\x8B\xB5\x00\xE5\x8B\xB9\x00\xE5\x8B\xBA\x00\xE5"
-     "\x8C\x85"
-     "\x00\xE5\x8C\x86\x00\xE5\x8C\x95\x00\xE5\x8C\x97\x00\xE5\x8C\x9A\x00\xE5"
-     "\x8C\xB8"
-     "\x00\xE5\x8C\xBB\x00\xE5\x8C\xBF\x00\xE5\x8D\x81\x00\xE5\x8D\x84\x00\xE5"
-     "\x8D\x85"
-     "\x00\xE5\x8D\x89\x00\xE5\x8D\x91\x00\xE5\x8D\x94\x00\xE5\x8D\x9A\x00\xE5"
-     "\x8D\x9C"
-     "\x00\xE5\x8D\xA9\x00\xE5\x8D\xB0\x00\xE5\x8D\xB3\x00\xE5\x8D\xB5\x00\xE5"
-     "\x8D\xBD"
-     "\x00\xE5\x8D\xBF\x00\xE5\x8E\x82\x00\xE5\x8E\xB6\x00\xE5\x8F\x83\x00\xE5"
-     "\x8F\x88"
-     "\x00\xE5\x8F\x8A\x00\xE5\x8F\x8C\x00\xE5\x8F\x9F\x00\xE5\x8F\xA3\x00\xE5"
-     "\x8F\xA5"
-     "\x00\xE5\x8F\xAB\x00\xE5\x8F\xAF\x00\xE5\x8F\xB1\x00\xE5\x8F\xB3\x00\xE5"
-     "\x90\x86"
-     "\x00\xE5\x90\x88\x00\xE5\x90\x8D\x00\xE5\x90\x8F\x00\xE5\x90\x9D\x00\xE5"
-     "\x90\xB8"
-     "\x00\xE5\x90\xB9\x00\xE5\x91\x82\x00\xE5\x91\x88\x00\xE5\x91\xA8\x00\xE5"
-     "\x92\x9E"
-     "\x00\xE5\x92\xA2\x00\xE5\x92\xBD\x00\xE5\x93\xB6\x00\xE5\x94\x90\x00\xE5"
-     "\x95\x8F"
-     "\x00\xE5\x95\x93\x00\xE5\x95\x95\x00\xE5\x95\xA3\x00\xE5\x96\x84\x00\xE5"
-     "\x96\x87"
-     "\x00\xE5\x96\x99\x00\xE5\x96\x9D\x00\xE5\x96\xAB\x00\xE5\x96\xB3\x00\xE5"
-     "\x96\xB6"
-     "\x00\xE5\x97\x80\x00\xE5\x97\x82\x00\xE5\x97\xA2\x00\xE5\x98\x86\x00\xE5"
-     "\x99\x91"
-     "\x00\xE5\x99\xA8\x00\xE5\x99\xB4\x00\xE5\x9B\x97\x00\xE5\x9B\x9B\x00\xE5"
-     "\x9B\xB9"
-     "\x00\xE5\x9C\x96\x00\xE5\x9C\x97\x00\xE5\x9C\x9F\x00\xE5\x9C\xB0\x00\xE5"
-     "\x9E\x8B"
-     "\x00\xE5\x9F\x8E\x00\xE5\x9F\xB4\x00\xE5\xA0\x8D\x00\xE5\xA0\xB1\x00\xE5"
-     "\xA0\xB2"
-     "\x00\xE5\xA1\x80\x00\xE5\xA1\x9A\x00\xE5\xA1\x9E\x00\xE5\xA2\xA8\x00\xE5"
-     "\xA2\xAC"
-     "\x00\xE5\xA2\xB3\x00\xE5\xA3\x98\x00\xE5\xA3\x9F\x00\xE5\xA3\xAB\x00\xE5"
-     "\xA3\xAE"
-     "\x00\xE5\xA3\xB0\x00\xE5\xA3\xB2\x00\xE5\xA3\xB7\x00\xE5\xA4\x82\x00\xE5"
-     "\xA4\x86"
-     "\x00\xE5\xA4\x8A\x00\xE5\xA4\x95\x00\xE5\xA4\x9A\x00\xE5\xA4\x9C\x00\xE5"
-     "\xA4\xA2"
-     "\x00\xE5\xA4\xA7\x00\xE5\xA4\xA7\xE6\xAD\xA3\x00\xE5\xA4\xA9\x00\xE5\xA5"
-     "\x84\x00"
-     "\xE5\xA5\x88\x00\xE5\xA5\x91\x00\xE5\xA5\x94\x00\xE5\xA5\xA2\x00\xE5\xA5"
-     "\xB3\x00"
-     "\xE5\xA7\x98\x00\xE5\xA7\xAC\x00\xE5\xA8\x9B\x00\xE5\xA8\xA7\x00\xE5\xA9"
-     "\xA2\x00"
-     "\xE5\xA9\xA6\x00\xE5\xAA\xB5\x00\xE5\xAC\x88\x00\xE5\xAC\xA8\x00\xE5\xAC"
-     "\xBE\x00"
-     "\xE5\xAD\x90\x00\xE5\xAD\x97\x00\xE5\xAD\xA6\x00\xE5\xAE\x80\x00\xE5\xAE"
-     "\x85\x00"
-     "\xE5\xAE\x97\x00\xE5\xAF\x83\x00\xE5\xAF\x98\x00\xE5\xAF\xA7\x00\xE5\xAF"
-     "\xAE\x00"
-     "\xE5\xAF\xB3\x00\xE5\xAF\xB8\x00\xE5\xAF\xBF\x00\xE5\xB0\x86\x00\xE5\xB0"
-     "\x8F\x00"
-     "\xE5\xB0\xA2\x00\xE5\xB0\xB8\x00\xE5\xB0\xBF\x00\xE5\xB1\xA0\x00\xE5\xB1"
-     "\xA2\x00"
-     "\xE5\xB1\xA4\x00\xE5\xB1\xA5\x00\xE5\xB1\xAE\x00\xE5\xB1\xB1\x00\xE5\xB2"
-     "\x8D\x00"
-     "\xE5\xB3\x80\x00\xE5\xB4\x99\x00\xE5\xB5\x83\x00\xE5\xB5\x90\x00\xE5\xB5"
-     "\xAB\x00"
-     "\xE5\xB5\xAE\x00\xE5\xB5\xBC\x00\xE5\xB6\xB2\x00\xE5\xB6\xBA\x00\xE5\xB7"
-     "\x9B\x00"
-     "\xE5\xB7\xA1\x00\xE5\xB7\xA2\x00\xE5\xB7\xA5\x00\xE5\xB7\xA6\x00\xE5\xB7"
-     "\xB1\x00"
-     "\xE5\xB7\xBD\x00\xE5\xB7\xBE\x00\xE5\xB8\xA8\x00\xE5\xB8\xBD\x00\xE5\xB9"
-     "\xA9\x00"
-     "\xE5\xB9\xB2\x00\xE5\xB9\xB3\xE6\x88\x90\x00\xE5\xB9\xB4\x00\xE5\xB9\xBA"
-     "\x00\xE5"
-     "\xB9\xBC\x00\xE5\xB9\xBF\x00\xE5\xBA\xA6\x00\xE5\xBA\xB0\x00\xE5\xBA\xB3"
-     "\x00\xE5"
-     "\xBA\xB6\x00\xE5\xBB\x89\x00\xE5\xBB\x8A\x00\xE5\xBB\x92\x00\xE5\xBB\x93"
-     "\x00\xE5"
-     "\xBB\x99\x00\xE5\xBB\xAC\x00\xE5\xBB\xB4\x00\xE5\xBB\xBE\x00\xE5\xBC\x84"
-     "\x00\xE5"
-     "\xBC\x8B\x00\xE5\xBC\x93\x00\xE5\xBC\xA2\x00\xE5\xBD\x90\x00\xE5\xBD\x93"
-     "\x00\xE5"
-     "\xBD\xA1\x00\xE5\xBD\xA2\x00\xE5\xBD\xA9\x00\xE5\xBD\xAB\x00\xE5\xBD\xB3"
-     "\x00\xE5"
-     "\xBE\x8B\x00\xE5\xBE\x8C\x00\xE5\xBE\x97\x00\xE5\xBE\x9A\x00\xE5\xBE\xA9"
-     "\x00\xE5"
-     "\xBE\xAD\x00\xE5\xBF\x83\x00\xE5\xBF\x8D\x00\xE5\xBF\x97\x00\xE5\xBF\xB5"
-     "\x00\xE5"
-     "\xBF\xB9\x00\xE6\x80\x92\x00\xE6\x80\x9C\x00\xE6\x81\xB5\x00\xE6\x82\x81"
-     "\x00\xE6"
-     "\x82\x94\x00\xE6\x83\x87\x00\xE6\x83\x98\x00\xE6\x83\xA1\x00\xE6\x84\x88"
-     "\x00\xE6"
-     "\x85\x84\x00\xE6\x85\x88\x00\xE6\x85\x8C\x00\xE6\x85\x8E\x00\xE6\x85\xA0"
-     "\x00\xE6"
-     "\x85\xA8\x00\xE6\x85\xBA\x00\xE6\x86\x8E\x00\xE6\x86\x90\x00\xE6\x86\xA4"
-     "\x00\xE6"
-     "\x86\xAF\x00\xE6\x86\xB2\x00\xE6\x87\x9E\x00\xE6\x87\xB2\x00\xE6\x87\xB6"
-     "\x00\xE6"
-     "\x88\x80\x00\xE6\x88\x88\x00\xE6\x88\x90\x00\xE6\x88\x9B\x00\xE6\x88\xAE"
-     "\x00\xE6"
-     "\x88\xB4\x00\xE6\x88\xB6\x00\xE6\x89\x8B\x00\xE6\x89\x93\x00\xE6\x89\x9D"
-     "\x00\xE6"
-     "\x8A\x95\x00\xE6\x8A\xB1\x00\xE6\x8B\x89\x00\xE6\x8B\x8F\x00\xE6\x8B\x93"
-     "\x00\xE6"
-     "\x8B\x94\x00\xE6\x8B\xBC\x00\xE6\x8B\xBE\x00\xE6\x8C\x87\x00\xE6\x8C\xBD"
-     "\x00\xE6"
-     "\x8D\x90\x00\xE6\x8D\x95\x00\xE6\x8D\xA8\x00\xE6\x8D\xBB\x00\xE6\x8E\x83"
-     "\x00\xE6"
-     "\x8E\xA0\x00\xE6\x8E\xA9\x00\xE6\x8F\x84\x00\xE6\x8F\x85\x00\xE6\x8F\xA4"
-     "\x00\xE6"
-     "\x90\x9C\x00\xE6\x90\xA2\x00\xE6\x91\x92\x00\xE6\x91\xA9\x00\xE6\x91\xB7"
-     "\x00\xE6"
-     "\x91\xBE\x00\xE6\x92\x9A\x00\xE6\x92\x9D\x00\xE6\x93\x84\x00\xE6\x94\xAF"
-     "\x00\xE6"
-     "\x94\xB4\x00\xE6\x95\x8F\x00\xE6\x95\x96\x00\xE6\x95\xAC\x00\xE6\x95\xB8"
-     "\x00\xE6"
-     "\x96\x87\x00\xE6\x96\x97\x00\xE6\x96\x99\x00\xE6\x96\xA4\x00\xE6\x96\xB0"
-     "\x00\xE6"
-     "\x96\xB9\x00\xE6\x97\x85\x00\xE6\x97\xA0\x00\xE6\x97\xA2\x00\xE6\x97\xA3"
-     "\x00\xE6"
-     "\x97\xA5\x00\xE6\x98\x8E\xE6\xB2\xBB\x00\xE6\x98\x93\x00\xE6\x98\xA0\x00"
-     "\xE6\x98"
-     "\xAD\xE5\x92\x8C\x00\xE6\x99\x89\x00\xE6\x99\xB4\x00\xE6\x9A\x88\x00\xE6"
-     "\x9A\x91"
-     "\x00\xE6\x9A\x9C\x00\xE6\x9A\xB4\x00\xE6\x9B\x86\x00\xE6\x9B\xB0\x00\xE6"
-     "\x9B\xB4"
-     "\x00\xE6\x9B\xB8\x00\xE6\x9C\x80\x00\xE6\x9C\x88\x00\xE6\x9C\x89\x00\xE6"
-     "\x9C\x97"
-     "\x00\xE6\x9C\x9B\x00\xE6\x9C\xA1\x00\xE6\x9C\xA8\x00\xE6\x9D\x8E\x00\xE6"
-     "\x9D\x93"
-     "\x00\xE6\x9D\x96\x00\xE6\x9D\x9E\x00\xE6\x9D\xBB\x00\xE6\x9E\x85\x00\xE6"
-     "\x9E\x97"
-     "\x00\xE6\x9F\xB3\x00\xE6\x9F\xBA\x00\xE6\xA0\x97\x00\xE6\xA0\x9F\x00\xE6"
-     "\xA0\xAA"
-     "\x00\xE6\xA0\xAA\xE5\xBC\x8F\xE4\xBC\x9A\xE7\xA4\xBE\x00\xE6\xA1\x92\x00"
-     "\xE6\xA2"
-     "\x81\x00\xE6\xA2\x85\x00\xE6\xA2\x8E\x00\xE6\xA2\xA8\x00\xE6\xA4\x94\x00"
-     "\xE6\xA5"
-     "\x82\x00\xE6\xA6\xA3\x00\xE6\xA7\xAA\x00\xE6\xA8\x82\x00\xE6\xA8\x93\x00"
-     "\xE6\xAA"
-     "\xA8\x00\xE6\xAB\x93\x00\xE6\xAB\x9B\x00\xE6\xAC\x84\x00\xE6\xAC\xA0\x00"
-     "\xE6\xAC"
-     "\xA1\x00\xE6\xAD\x94\x00\xE6\xAD\xA2\x00\xE6\xAD\xA3\x00\xE6\xAD\xB2\x00"
-     "\xE6\xAD"
-     "\xB7\x00\xE6\xAD\xB9\x00\xE6\xAE\x9F\x00\xE6\xAE\xAE\x00\xE6\xAE\xB3\x00"
-     "\xE6\xAE"
-     "\xBA\x00\xE6\xAE\xBB\x00\xE6\xAF\x8B\x00\xE6\xAF\x8D\x00\xE6\xAF\x94\x00"
-     "\xE6\xAF"
-     "\x9B\x00\xE6\xB0\x8F\x00\xE6\xB0\x94\x00\xE6\xB0\xB4\x00\xE6\xB1\x8E\x00"
-     "\xE6\xB1"
-     "\xA7\x00\xE6\xB2\x88\x00\xE6\xB2\xBF\x00\xE6\xB3\x8C\x00\xE6\xB3\x8D\x00"
-     "\xE6\xB3"
-     "\xA5\x00\xE6\xB3\xA8\x00\xE6\xB4\x96\x00\xE6\xB4\x9B\x00\xE6\xB4\x9E\x00"
-     "\xE6\xB4"
-     "\xB4\x00\xE6\xB4\xBE\x00\xE6\xB5\x81\x00\xE6\xB5\xA9\x00\xE6\xB5\xAA\x00"
-     "\xE6\xB5"
-     "\xB7\x00\xE6\xB5\xB8\x00\xE6\xB6\x85\x00\xE6\xB7\x8B\x00\xE6\xB7\x9A\x00"
-     "\xE6\xB7"
-     "\xAA\x00\xE6\xB7\xB9\x00\xE6\xB8\x9A\x00\xE6\xB8\xAF\x00\xE6\xB9\xAE\x00"
-     "\xE6\xBA"
-     "\x80\x00\xE6\xBA\x9C\x00\xE6\xBA\xBA\x00\xE6\xBB\x87\x00\xE6\xBB\x8B\x00"
-     "\xE6\xBB"
-     "\x91\x00\xE6\xBB\x9B\x00\xE6\xBC\x8F\x00\xE6\xBC\x94\x00\xE6\xBC\xA2\x00"
-     "\xE6\xBC"
-     "\xA3\x00\xE6\xBD\xAE\x00\xE6\xBF\x86\x00\xE6\xBF\xAB\x00\xE6\xBF\xBE\x00"
-     "\xE7\x80"
-     "\x9B\x00\xE7\x80\x9E\x00\xE7\x80\xB9\x00\xE7\x81\x8A\x00\xE7\x81\xAB\x00"
-     "\xE7\x81"
-     "\xB0\x00\xE7\x81\xB7\x00\xE7\x81\xBD\x00\xE7\x82\x99\x00\xE7\x82\xAD\x00"
-     "\xE7\x83"
-     "\x88\x00\xE7\x83\x99\x00\xE7\x84\xA1\x00\xE7\x85\x85\x00\xE7\x85\x89\x00"
-     "\xE7\x85"
-     "\xAE\x00\xE7\x86\x9C\x00\xE7\x87\x8E\x00\xE7\x87\x90\x00\xE7\x88\x90\x00"
-     "\xE7\x88"
-     "\x9B\x00\xE7\x88\xA8\x00\xE7\x88\xAA\x00\xE7\x88\xAB\x00\xE7\x88\xB5\x00"
-     "\xE7\x88"
-     "\xB6\x00\xE7\x88\xBB\x00\xE7\x88\xBF\x00\xE7\x89\x87\x00\xE7\x89\x90\x00"
-     "\xE7\x89"
-     "\x99\x00\xE7\x89\x9B\x00\xE7\x89\xA2\x00\xE7\x89\xB9\x00\xE7\x8A\x80\x00"
-     "\xE7\x8A"
-     "\x95\x00\xE7\x8A\xAC\x00\xE7\x8A\xAF\x00\xE7\x8B\x80\x00\xE7\x8B\xBC\x00"
-     "\xE7\x8C"
-     "\xAA\x00\xE7\x8D\xB5\x00\xE7\x8D\xBA\x00\xE7\x8E\x84\x00\xE7\x8E\x87\x00"
-     "\xE7\x8E"
-     "\x89\x00\xE7\x8E\x8B\x00\xE7\x8E\xA5\x00\xE7\x8E\xB2\x00\xE7\x8F\x9E\x00"
-     "\xE7\x90"
-     "\x86\x00\xE7\x90\x89\x00\xE7\x90\xA2\x00\xE7\x91\x87\x00\xE7\x91\x9C\x00"
-     "\xE7\x91"
-     "\xA9\x00\xE7\x91\xB1\x00\xE7\x92\x85\x00\xE7\x92\x89\x00\xE7\x92\x98\x00"
-     "\xE7\x93"
-     "\x8A\x00\xE7\x93\x9C\x00\xE7\x93\xA6\x00\xE7\x94\x86\x00\xE7\x94\x98\x00"
-     "\xE7\x94"
-     "\x9F\x00\xE7\x94\xA4\x00\xE7\x94\xA8\x00\xE7\x94\xB0\x00\xE7\x94\xB2\x00"
-     "\xE7\x94"
-     "\xB3\x00\xE7\x94\xB7\x00\xE7\x94\xBB\x00\xE7\x94\xBE\x00\xE7\x95\x99\x00"
-     "\xE7\x95"
-     "\xA5\x00\xE7\x95\xB0\x00\xE7\x96\x8B\x00\xE7\x96\x92\x00\xE7\x97\xA2\x00"
-     "\xE7\x98"
-     "\x90\x00\xE7\x98\x9D\x00\xE7\x98\x9F\x00\xE7\x99\x82\x00\xE7\x99\xA9\x00"
-     "\xE7\x99"
-     "\xB6\x00\xE7\x99\xBD\x00\xE7\x9A\xAE\x00\xE7\x9A\xBF\x00\xE7\x9B\x8A\x00"
-     "\xE7\x9B"
-     "\x9B\x00\xE7\x9B\xA3\x00\xE7\x9B\xA7\x00\xE7\x9B\xAE\x00\xE7\x9B\xB4\x00"
-     "\xE7\x9C"
-     "\x81\x00\xE7\x9C\x9E\x00\xE7\x9C\x9F\x00\xE7\x9D\x80\x00\xE7\x9D\x8A\x00"
-     "\xE7\x9E"
-     "\x8B\x00\xE7\x9E\xA7\x00\xE7\x9F\x9B\x00\xE7\x9F\xA2\x00\xE7\x9F\xB3\x00"
-     "\xE7\xA1"
-     "\x8E\x00\xE7\xA1\xAB\x00\xE7\xA2\x8C\x00\xE7\xA2\x91\x00\xE7\xA3\x8A\x00"
-     "\xE7\xA3"
-     "\x8C\x00\xE7\xA3\xBB\x00\xE7\xA4\xAA\x00\xE7\xA4\xBA\x00\xE7\xA4\xBC\x00"
-     "\xE7\xA4"
-     "\xBE\x00\xE7\xA5\x88\x00\xE7\xA5\x89\x00\xE7\xA5\x90\x00\xE7\xA5\x96\x00"
-     "\xE7\xA5"
-     "\x9D\x00\xE7\xA5\x9E\x00\xE7\xA5\xA5\x00\xE7\xA5\xBF\x00\xE7\xA6\x81\x00"
-     "\xE7\xA6"
-     "\x8D\x00\xE7\xA6\x8E\x00\xE7\xA6\x8F\x00\xE7\xA6\xAE\x00\xE7\xA6\xB8\x00"
-     "\xE7\xA6"
-     "\xBE\x00\xE7\xA7\x8A\x00\xE7\xA7\x98\x00\xE7\xA7\xAB\x00\xE7\xA8\x9C\x00"
-     "\xE7\xA9"
-     "\x80\x00\xE7\xA9\x8A\x00\xE7\xA9\x8F\x00\xE7\xA9\xB4\x00\xE7\xA9\xBA\x00"
-     "\xE7\xAA"
-     "\x81\x00\xE7\xAA\xB1\x00\xE7\xAB\x8B\x00\xE7\xAB\xAE\x00\xE7\xAB\xB9\x00"
-     "\xE7\xAC"
-     "\xA0\x00\xE7\xAE\x8F\x00\xE7\xAF\x80\x00\xE7\xAF\x86\x00\xE7\xAF\x89\x00"
-     "\xE7\xB0"
-     "\xBE\x00\xE7\xB1\xA0\x00\xE7\xB1\xB3\x00\xE7\xB1\xBB\x00\xE7\xB2\x92\x00"
-     "\xE7\xB2"
-     "\xBE\x00\xE7\xB3\x92\x00\xE7\xB3\x96\x00\xE7\xB3\xA3\x00\xE7\xB3\xA7\x00"
-     "\xE7\xB3"
-     "\xA8\x00\xE7\xB3\xB8\x00\xE7\xB4\x80\x00\xE7\xB4\x90\x00\xE7\xB4\xA2\x00"
-     "\xE7\xB4"
-     "\xAF\x00\xE7\xB5\x82\x00\xE7\xB5\x9B\x00\xE7\xB5\xA3\x00\xE7\xB6\xA0\x00"
-     "\xE7\xB6"
-     "\xBE\x00\xE7\xB7\x87\x00\xE7\xB7\xB4\x00\xE7\xB8\x82\x00\xE7\xB8\x89\x00"
-     "\xE7\xB8"
-     "\xB7\x00\xE7\xB9\x81\x00\xE7\xB9\x85\x00\xE7\xBC\xB6\x00\xE7\xBC\xBE\x00"
-     "\xE7\xBD"
-     "\x91\x00\xE7\xBD\xB2\x00\xE7\xBD\xB9\x00\xE7\xBD\xBA\x00\xE7\xBE\x85\x00"
-     "\xE7\xBE"
-     "\x8A\x00\xE7\xBE\x95\x00\xE7\xBE\x9A\x00\xE7\xBE\xBD\x00\xE7\xBF\xBA\x00"
-     "\xE8\x80"
-     "\x81\x00\xE8\x80\x85\x00\xE8\x80\x8C\x00\xE8\x80\x92\x00\xE8\x80\xB3\x00"
-     "\xE8\x81"
-     "\x86\x00\xE8\x81\xA0\x00\xE8\x81\xAF\x00\xE8\x81\xB0\x00\xE8\x81\xBE\x00"
-     "\xE8\x81"
-     "\xBF\x00\xE8\x82\x89\x00\xE8\x82\x8B\x00\xE8\x82\xAD\x00\xE8\x82\xB2\x00"
-     "\xE8\x84"
-     "\x83\x00\xE8\x84\xBE\x00\xE8\x87\x98\x00\xE8\x87\xA3\x00\xE8\x87\xA8\x00"
-     "\xE8\x87"
-     "\xAA\x00\xE8\x87\xAD\x00\xE8\x87\xB3\x00\xE8\x87\xBC\x00\xE8\x88\x81\x00"
-     "\xE8\x88"
-     "\x84\x00\xE8\x88\x8C\x00\xE8\x88\x98\x00\xE8\x88\x9B\x00\xE8\x88\x9F\x00"
-     "\xE8\x89"
-     "\xAE\x00\xE8\x89\xAF\x00\xE8\x89\xB2\x00\xE8\x89\xB8\x00\xE8\x89\xB9\x00"
-     "\xE8\x8A"
-     "\x8B\x00\xE8\x8A\x91\x00\xE8\x8A\x9D\x00\xE8\x8A\xB1\x00\xE8\x8A\xB3\x00"
-     "\xE8\x8A"
-     "\xBD\x00\xE8\x8B\xA5\x00\xE8\x8B\xA6\x00\xE8\x8C\x9D\x00\xE8\x8C\xA3\x00"
-     "\xE8\x8C"
-     "\xB6\x00\xE8\x8D\x92\x00\xE8\x8D\x93\x00\xE8\x8D\xA3\x00\xE8\x8E\xAD\x00"
-     "\xE8\x8E"
-     "\xBD\x00\xE8\x8F\x89\x00\xE8\x8F\x8A\x00\xE8\x8F\x8C\x00\xE8\x8F\x9C\x00"
-     "\xE8\x8F"
-     "\xA7\x00\xE8\x8F\xAF\x00\xE8\x8F\xB1\x00\xE8\x90\xBD\x00\xE8\x91\x89\x00"
-     "\xE8\x91"
-     "\x97\x00\xE8\x93\xAE\x00\xE8\x93\xB1\x00\xE8\x93\xB3\x00\xE8\x93\xBC\x00"
-     "\xE8\x94"
-     "\x96\x00\xE8\x95\xA4\x00\xE8\x97\x8D\x00\xE8\x97\xBA\x00\xE8\x98\x86\x00"
-     "\xE8\x98"
-     "\x92\x00\xE8\x98\xAD\x00\xE8\x98\xBF\x00\xE8\x99\x8D\x00\xE8\x99\x90\x00"
-     "\xE8\x99"
-     "\x9C\x00\xE8\x99\xA7\x00\xE8\x99\xA9\x00\xE8\x99\xAB\x00\xE8\x9A\x88\x00"
-     "\xE8\x9A"
-     "\xA9\x00\xE8\x9B\xA2\x00\xE8\x9C\x8E\x00\xE8\x9C\xA8\x00\xE8\x9D\xAB\x00"
-     "\xE8\x9D"
-     "\xB9\x00\xE8\x9E\x86\x00\xE8\x9E\xBA\x00\xE8\x9F\xA1\x00\xE8\xA0\x81\x00"
-     "\xE8\xA0"
-     "\x9F\x00\xE8\xA1\x80\x00\xE8\xA1\x8C\x00\xE8\xA1\xA0\x00\xE8\xA1\xA3\x00"
-     "\xE8\xA3"
-     "\x82\x00\xE8\xA3\x8F\x00\xE8\xA3\x97\x00\xE8\xA3\x9E\x00\xE8\xA3\xA1\x00"
-     "\xE8\xA3"
-     "\xB8\x00\xE8\xA3\xBA\x00\xE8\xA4\x90\x00\xE8\xA5\x81\x00\xE8\xA5\xA4\x00"
-     "\xE8\xA5"
-     "\xBE\x00\xE8\xA6\x86\x00\xE8\xA6\x8B\x00\xE8\xA6\x96\x00\xE8\xA7\x92\x00"
-     "\xE8\xA7"
-     "\xA3\x00\xE8\xA8\x80\x00\xE8\xAA\xA0\x00\xE8\xAA\xAA\x00\xE8\xAA\xBF\x00"
-     "\xE8\xAB"
-     "\x8B\x00\xE8\xAB\x92\x00\xE8\xAB\x96\x00\xE8\xAB\xAD\x00\xE8\xAB\xB8\x00"
-     "\xE8\xAB"
-     "\xBE\x00\xE8\xAC\x81\x00\xE8\xAC\xB9\x00\xE8\xAD\x98\x00\xE8\xAE\x80\x00"
-     "\xE8\xAE"
-     "\x8A\x00\xE8\xB0\xB7\x00\xE8\xB1\x86\x00\xE8\xB1\x88\x00\xE8\xB1\x95\x00"
-     "\xE8\xB1"
-     "\xB8\x00\xE8\xB2\x9D\x00\xE8\xB2\xA1\x00\xE8\xB2\xA9\x00\xE8\xB2\xAB\x00"
-     "\xE8\xB3"
-     "\x81\x00\xE8\xB3\x82\x00\xE8\xB3\x87\x00\xE8\xB3\x88\x00\xE8\xB3\x93\x00"
-     "\xE8\xB4"
-     "\x88\x00\xE8\xB4\x9B\x00\xE8\xB5\xA4\x00\xE8\xB5\xB0\x00\xE8\xB5\xB7\x00"
-     "\xE8\xB6"
-     "\xB3\x00\xE8\xB6\xBC\x00\xE8\xB7\x8B\x00\xE8\xB7\xAF\x00\xE8\xB7\xB0\x00"
-     "\xE8\xBA"
-     "\xAB\x00\xE8\xBB\x8A\x00\xE8\xBB\x94\x00\xE8\xBC\xA6\x00\xE8\xBC\xAA\x00"
-     "\xE8\xBC"
-     "\xB8\x00\xE8\xBC\xBB\x00\xE8\xBD\xA2\x00\xE8\xBE\x9B\x00\xE8\xBE\x9E\x00"
-     "\xE8\xBE"
-     "\xB0\x00\xE8\xBE\xB5\x00\xE8\xBE\xB6\x00\xE9\x80\xA3\x00\xE9\x80\xB8\x00"
-     "\xE9\x81"
-     "\x8A\x00\xE9\x81\xA9\x00\xE9\x81\xB2\x00\xE9\x81\xBC\x00\xE9\x82\x8F\x00"
-     "\xE9\x82"
-     "\x91\x00\xE9\x82\x94\x00\xE9\x83\x8E\x00\xE9\x83\x9E\x00\xE9\x83\xB1\x00"
-     "\xE9\x83"
-     "\xBD\x00\xE9\x84\x91\x00\xE9\x84\x9B\x00\xE9\x85\x89\x00\xE9\x85\x8D\x00"
-     "\xE9\x85"
-     "\xAA\x00\xE9\x86\x99\x00\xE9\x86\xB4\x00\xE9\x87\x86\x00\xE9\x87\x8C\x00"
-     "\xE9\x87"
-     "\x8F\x00\xE9\x87\x91\x00\xE9\x88\xB4\x00\xE9\x88\xB8\x00\xE9\x89\xB6\x00"
-     "\xE9\x89"
-     "\xBC\x00\xE9\x8B\x97\x00\xE9\x8B\x98\x00\xE9\x8C\x84\x00\xE9\x8D\x8A\x00"
-     "\xE9\x8F"
-     "\xB9\x00\xE9\x90\x95\x00\xE9\x95\xB7\x00\xE9\x96\x80\x00\xE9\x96\x8B\x00"
-     "\xE9\x96"
-     "\xAD\x00\xE9\x96\xB7\x00\xE9\x98\x9C\x00\xE9\x98\xAE\x00\xE9\x99\x8B\x00"
-     "\xE9\x99"
-     "\x8D\x00\xE9\x99\xB5\x00\xE9\x99\xB8\x00\xE9\x99\xBC\x00\xE9\x9A\x86\x00"
-     "\xE9\x9A"
-     "\xA3\x00\xE9\x9A\xB6\x00\xE9\x9A\xB7\x00\xE9\x9A\xB8\x00\xE9\x9A\xB9\x00"
-     "\xE9\x9B"
-     "\x83\x00\xE9\x9B\xA2\x00\xE9\x9B\xA3\x00\xE9\x9B\xA8\x00\xE9\x9B\xB6\x00"
-     "\xE9\x9B"
-     "\xB7\x00\xE9\x9C\xA3\x00\xE9\x9C\xB2\x00\xE9\x9D\x88\x00\xE9\x9D\x91\x00"
-     "\xE9\x9D"
-     "\x96\x00\xE9\x9D\x9E\x00\xE9\x9D\xA2\x00\xE9\x9D\xA9\x00\xE9\x9F\x8B\x00"
-     "\xE9\x9F"
-     "\x9B\x00\xE9\x9F\xA0\x00\xE9\x9F\xAD\x00\xE9\x9F\xB3\x00\xE9\x9F\xBF\x00"
-     "\xE9\xA0"
-     "\x81\x00\xE9\xA0\x85\x00\xE9\xA0\x8B\x00\xE9\xA0\x98\x00\xE9\xA0\xA9\x00"
-     "\xE9\xA0"
-     "\xBB\x00\xE9\xA1\x9E\x00\xE9\xA2\xA8\x00\xE9\xA3\x9B\x00\xE9\xA3\x9F\x00"
-     "\xE9\xA3"
-     "\xA2\x00\xE9\xA3\xAF\x00\xE9\xA3\xBC\x00\xE9\xA4\xA8\x00\xE9\xA4\xA9\x00"
-     "\xE9\xA6"
-     "\x96\x00\xE9\xA6\x99\x00\xE9\xA6\xA7\x00\xE9\xA6\xAC\x00\xE9\xA7\x82\x00"
-     "\xE9\xA7"
-     "\xB1\x00\xE9\xA7\xBE\x00\xE9\xA9\xAA\x00\xE9\xAA\xA8\x00\xE9\xAB\x98\x00"
-     "\xE9\xAB"
-     "\x9F\x00\xE9\xAC\x92\x00\xE9\xAC\xA5\x00\xE9\xAC\xAF\x00\xE9\xAC\xB2\x00"
-     "\xE9\xAC"
-     "\xBC\x00\xE9\xAD\x9A\x00\xE9\xAD\xAF\x00\xE9\xB1\x80\x00\xE9\xB1\x97\x00"
-     "\xE9\xB3"
-     "\xA5\x00\xE9\xB3\xBD\x00\xE9\xB5\xA7\x00\xE9\xB6\xB4\x00\xE9\xB7\xBA\x00"
-     "\xE9\xB8"
-     "\x9E\x00\xE9\xB9\xB5\x00\xE9\xB9\xBF\x00\xE9\xBA\x97\x00\xE9\xBA\x9F\x00"
-     "\xE9\xBA"
-     "\xA5\x00\xE9\xBA\xBB\x00\xE9\xBB\x83\x00\xE9\xBB\x8D\x00\xE9\xBB\x8E\x00"
-     "\xE9\xBB"
-     "\x91\x00\xE9\xBB\xB9\x00\xE9\xBB\xBD\x00\xE9\xBB\xBE\x00\xE9\xBC\x85\x00"
-     "\xE9\xBC"
-     "\x8E\x00\xE9\xBC\x8F\x00\xE9\xBC\x93\x00\xE9\xBC\x96\x00\xE9\xBC\xA0\x00"
-     "\xE9\xBC"
-     "\xBB\x00\xE9\xBD\x83\x00\xE9\xBD\x8A\x00\xE9\xBD\x92\x00\xE9\xBE\x8D\x00"
-     "\xE9\xBE"
-     "\x8E\x00\xE9\xBE\x9C\x00\xE9\xBE\x9F\x00\xE9\xBE\xA0\x00\xEA\x99\x81\x00"
-     "\xEA\x99"
-     "\x83\x00\xEA\x99\x85\x00\xEA\x99\x87\x00\xEA\x99\x89\x00\xEA\x99\x8B\x00"
-     "\xEA\x99"
-     "\x8D\x00\xEA\x99\x8F\x00\xEA\x99\x91\x00\xEA\x99\x93\x00\xEA\x99\x95\x00"
-     "\xEA\x99"
-     "\x97\x00\xEA\x99\x99\x00\xEA\x99\x9B\x00\xEA\x99\x9D\x00\xEA\x99\x9F\x00"
-     "\xEA\x99"
-     "\xA1\x00\xEA\x99\xA3\x00\xEA\x99\xA5\x00\xEA\x99\xA7\x00\xEA\x99\xA9\x00"
-     "\xEA\x99"
-     "\xAB\x00\xEA\x99\xAD\x00\xEA\x9A\x81\x00\xEA\x9A\x83\x00\xEA\x9A\x85\x00"
-     "\xEA\x9A"
-     "\x87\x00\xEA\x9A\x89\x00\xEA\x9A\x8B\x00\xEA\x9A\x8D\x00\xEA\x9A\x8F\x00"
-     "\xEA\x9A"
-     "\x91\x00\xEA\x9A\x93\x00\xEA\x9A\x95\x00\xEA\x9A\x97\x00\xEA\x9A\x99\x00"
-     "\xEA\x9A"
-     "\x9B\x00\xEA\x9C\xA3\x00\xEA\x9C\xA5\x00\xEA\x9C\xA7\x00\xEA\x9C\xA9\x00"
-     "\xEA\x9C"
-     "\xAB\x00\xEA\x9C\xAD\x00\xEA\x9C\xAF\x00\xEA\x9C\xB3\x00\xEA\x9C\xB5\x00"
-     "\xEA\x9C"
-     "\xB7\x00\xEA\x9C\xB9\x00\xEA\x9C\xBB\x00\xEA\x9C\xBD\x00\xEA\x9C\xBF\x00"
-     "\xEA\x9D"
-     "\x81\x00\xEA\x9D\x83\x00\xEA\x9D\x85\x00\xEA\x9D\x87\x00\xEA\x9D\x89\x00"
-     "\xEA\x9D"
-     "\x8B\x00\xEA\x9D\x8D\x00\xEA\x9D\x8F\x00\xEA\x9D\x91\x00\xEA\x9D\x93\x00"
-     "\xEA\x9D"
-     "\x95\x00\xEA\x9D\x97\x00\xEA\x9D\x99\x00\xEA\x9D\x9B\x00\xEA\x9D\x9D\x00"
-     "\xEA\x9D"
-     "\x9F\x00\xEA\x9D\xA1\x00\xEA\x9D\xA3\x00\xEA\x9D\xA5\x00\xEA\x9D\xA7\x00"
-     "\xEA\x9D"
-     "\xA9\x00\xEA\x9D\xAB\x00\xEA\x9D\xAD\x00\xEA\x9D\xAF\x00\xEA\x9D\xBA\x00"
-     "\xEA\x9D"
-     "\xBC\x00\xEA\x9D\xBF\x00\xEA\x9E\x81\x00\xEA\x9E\x83\x00\xEA\x9E\x85\x00"
-     "\xEA\x9E"
-     "\x87\x00\xEA\x9E\x8C\x00\xEA\x9E\x8E\x00\xEA\x9E\x91\x00\xEA\x9E\x93\x00"
-     "\xEA\x9E"
-     "\x94\x00\xEA\x9E\x97\x00\xEA\x9E\x99\x00\xEA\x9E\x9B\x00\xEA\x9E\x9D\x00"
-     "\xEA\x9E"
-     "\x9F\x00\xEA\x9E\xA1\x00\xEA\x9E\xA3\x00\xEA\x9E\xA5\x00\xEA\x9E\xA7\x00"
-     "\xEA\x9E"
-     "\xA9\x00\xEA\x9E\xB5\x00\xEA\x9E\xB7\x00\xEA\x9E\xB9\x00\xEA\x9E\xBB\x00"
-     "\xEA\x9E"
-     "\xBD\x00\xEA\x9E\xBF\x00\xEA\x9F\x81\x00\xEA\x9F\x83\x00\xEA\x9F\x88\x00"
-     "\xEA\x9F"
-     "\x8A\x00\xEA\x9F\x91\x00\xEA\x9F\x97\x00\xEA\x9F\x99\x00\xEA\x9F\xB6\x00"
-     "\xEA\xAC"
-     "\xB7\x00\xEA\xAD\x92\x00\xEA\xAD\x93\x00\xEA\xAD\xA6\x00\xEA\xAD\xA7\x00"
-     "\xEA\xB0"
-     "\x80\x00\xEA\xB0\x81\x00\xEA\xB0\x82\x00\xEA\xB0\x83\x00\xEA\xB0\x84\x00"
-     "\xEA\xB0"
-     "\x85\x00\xEA\xB0\x86\x00\xEA\xB0\x87\x00\xEA\xB0\x88\x00\xEA\xB0\x89\x00"
-     "\xEA\xB0"
-     "\x8A\x00\xEA\xB0\x8B\x00\xEA\xB0\x8C\x00\xEA\xB0\x8D\x00\xEA\xB0\x8E\x00"
-     "\xEA\xB0"
-     "\x8F\x00\xEA\xB0\x90\x00\xEA\xB0\x91\x00\xEA\xB0\x92\x00\xEA\xB0\x93\x00"
-     "\xEA\xB0"
-     "\x94\x00\xEA\xB0\x95\x00\xEA\xB0\x96\x00\xEA\xB0\x97\x00\xEA\xB0\x98\x00"
-     "\xEA\xB0"
-     "\x99\x00\xEA\xB0\x9A\x00\xEA\xB0\x9B\x00\xEA\xB0\x9C\x00\xEA\xB0\x9D\x00"
-     "\xEA\xB0"
-     "\x9E\x00\xEA\xB0\x9F\x00\xEA\xB0\xA0\x00\xEA\xB0\xA1\x00\xEA\xB0\xA2\x00"
-     "\xEA\xB0"
-     "\xA3\x00\xEA\xB0\xA4\x00\xEA\xB0\xA5\x00\xEA\xB0\xA6\x00\xEA\xB0\xA7\x00"
-     "\xEA\xB0"
-     "\xA8\x00\xEA\xB0\xA9\x00\xEA\xB0\xAA\x00\xEA\xB0\xAB\x00\xEA\xB0\xAC\x00"
-     "\xEA\xB0"
-     "\xAD\x00\xEA\xB0\xAE\x00\xEA\xB0\xAF\x00\xEA\xB0\xB0\x00\xEA\xB0\xB1\x00"
-     "\xEA\xB0"
-     "\xB2\x00\xEA\xB0\xB3\x00\xEA\xB0\xB4\x00\xEA\xB0\xB5\x00\xEA\xB0\xB6\x00"
-     "\xEA\xB0"
-     "\xB7\x00\xEA\xB0\xB8\x00\xEA\xB0\xB9\x00\xEA\xB0\xBA\x00\xEA\xB0\xBB\x00"
-     "\xEA\xB0"
-     "\xBC\x00\xEA\xB0\xBD\x00\xEA\xB0\xBE\x00\xEA\xB0\xBF\x00\xEA\xB1\x80\x00"
-     "\xEA\xB1"
-     "\x81\x00\xEA\xB1\x82\x00\xEA\xB1\x83\x00\xEA\xB1\x84\x00\xEA\xB1\x85\x00"
-     "\xEA\xB1"
-     "\x86\x00\xEA\xB1\x87\x00\xEA\xB1\x88\x00\xEA\xB1\x89\x00\xEA\xB1\x8A\x00"
-     "\xEA\xB1"
-     "\x8B\x00\xEA\xB1\x8C\x00\xEA\xB1\x8D\x00\xEA\xB1\x8E\x00\xEA\xB1\x8F\x00"
-     "\xEA\xB1"
-     "\x90\x00\xEA\xB1\x91\x00\xEA\xB1\x92\x00\xEA\xB1\x93\x00\xEA\xB1\x94\x00"
-     "\xEA\xB1"
-     "\x95\x00\xEA\xB1\x96\x00\xEA\xB1\x97\x00\xEA\xB1\x98\x00\xEA\xB1\x99\x00"
-     "\xEA\xB1"
-     "\x9A\x00\xEA\xB1\x9B\x00\xEA\xB1\x9C\x00\xEA\xB1\x9D\x00\xEA\xB1\x9E\x00"
-     "\xEA\xB1"
-     "\x9F\x00\xEA\xB1\xA0\x00\xEA\xB1\xA1\x00\xEA\xB1\xA2\x00\xEA\xB1\xA3\x00"
-     "\xEA\xB1"
-     "\xA4\x00\xEA\xB1\xA5\x00\xEA\xB1\xA6\x00\xEA\xB1\xA7\x00\xEA\xB1\xA8\x00"
-     "\xEA\xB1"
-     "\xA9\x00\xEA\xB1\xAA\x00\xEA\xB1\xAB\x00\xEA\xB1\xAC\x00\xEA\xB1\xAD\x00"
-     "\xEA\xB1"
-     "\xAE\x00\xEA\xB1\xAF\x00\xEA\xB1\xB0\x00\xEA\xB1\xB1\x00\xEA\xB1\xB2\x00"
-     "\xEA\xB1"
-     "\xB3\x00\xEA\xB1\xB4\x00\xEA\xB1\xB5\x00\xEA\xB1\xB6\x00\xEA\xB1\xB7\x00"
-     "\xEA\xB1"
-     "\xB8\x00\xEA\xB1\xB9\x00\xEA\xB1\xBA\x00\xEA\xB1\xBB\x00\xEA\xB1\xBC\x00"
-     "\xEA\xB1"
-     "\xBD\x00\xEA\xB1\xBE\x00\xEA\xB1\xBF\x00\xEA\xB2\x80\x00\xEA\xB2\x81\x00"
-     "\xEA\xB2"
-     "\x82\x00\xEA\xB2\x83\x00\xEA\xB2\x84\x00\xEA\xB2\x85\x00\xEA\xB2\x86\x00"
-     "\xEA\xB2"
-     "\x87\x00\xEA\xB2\x88\x00\xEA\xB2\x89\x00\xEA\xB2\x8A\x00\xEA\xB2\x8B\x00"
-     "\xEA\xB2"
-     "\x8C\x00\xEA\xB2\x8D\x00\xEA\xB2\x8E\x00\xEA\xB2\x8F\x00\xEA\xB2\x90\x00"
-     "\xEA\xB2"
-     "\x91\x00\xEA\xB2\x92\x00\xEA\xB2\x93\x00\xEA\xB2\x94\x00\xEA\xB2\x95\x00"
-     "\xEA\xB2"
-     "\x96\x00\xEA\xB2\x97\x00\xEA\xB2\x98\x00\xEA\xB2\x99\x00\xEA\xB2\x9A\x00"
-     "\xEA\xB2"
-     "\x9B\x00\xEA\xB2\x9C\x00\xEA\xB2\x9D\x00\xEA\xB2\x9E\x00\xEA\xB2\x9F\x00"
-     "\xEA\xB2"
-     "\xA0\x00\xEA\xB2\xA1\x00\xEA\xB2\xA2\x00\xEA\xB2\xA3\x00\xEA\xB2\xA4\x00"
-     "\xEA\xB2"
-     "\xA5\x00\xEA\xB2\xA6\x00\xEA\xB2\xA7\x00\xEA\xB2\xA8\x00\xEA\xB2\xA9\x00"
-     "\xEA\xB2"
-     "\xAA\x00\xEA\xB2\xAB\x00\xEA\xB2\xAC\x00\xEA\xB2\xAD\x00\xEA\xB2\xAE\x00"
-     "\xEA\xB2"
-     "\xAF\x00\xEA\xB2\xB0\x00\xEA\xB2\xB1\x00\xEA\xB2\xB2\x00\xEA\xB2\xB3\x00"
-     "\xEA\xB2"
-     "\xB4\x00\xEA\xB2\xB5\x00\xEA\xB2\xB6\x00\xEA\xB2\xB7\x00\xEA\xB2\xB8\x00"
-     "\xEA\xB2"
-     "\xB9\x00\xEA\xB2\xBA\x00\xEA\xB2\xBB\x00\xEA\xB2\xBC\x00\xEA\xB2\xBD\x00"
-     "\xEA\xB2"
-     "\xBE\x00\xEA\xB2\xBF\x00\xEA\xB3\x80\x00\xEA\xB3\x81\x00\xEA\xB3\x82\x00"
-     "\xEA\xB3"
-     "\x83\x00\xEA\xB3\x84\x00\xEA\xB3\x85\x00\xEA\xB3\x86\x00\xEA\xB3\x87\x00"
-     "\xEA\xB3"
-     "\x88\x00\xEA\xB3\x89\x00\xEA\xB3\x8A\x00\xEA\xB3\x8B\x00\xEA\xB3\x8C\x00"
-     "\xEA\xB3"
-     "\x8D\x00\xEA\xB3\x8E\x00\xEA\xB3\x8F\x00\xEA\xB3\x90\x00\xEA\xB3\x91\x00"
-     "\xEA\xB3"
-     "\x92\x00\xEA\xB3\x93\x00\xEA\xB3\x94\x00\xEA\xB3\x95\x00\xEA\xB3\x96\x00"
-     "\xEA\xB3"
-     "\x97\x00\xEA\xB3\x98\x00\xEA\xB3\x99\x00\xEA\xB3\x9A\x00\xEA\xB3\x9B\x00"
-     "\xEA\xB3"
-     "\x9C\x00\xEA\xB3\x9D\x00\xEA\xB3\x9E\x00\xEA\xB3\x9F\x00\xEA\xB3\xA0\x00"
-     "\xEA\xB3"
-     "\xA1\x00\xEA\xB3\xA2\x00\xEA\xB3\xA3\x00\xEA\xB3\xA4\x00\xEA\xB3\xA5\x00"
-     "\xEA\xB3"
-     "\xA6\x00\xEA\xB3\xA7\x00\xEA\xB3\xA8\x00\xEA\xB3\xA9\x00\xEA\xB3\xAA\x00"
-     "\xEA\xB3"
-     "\xAB\x00\xEA\xB3\xAC\x00\xEA\xB3\xAD\x00\xEA\xB3\xAE\x00\xEA\xB3\xAF\x00"
-     "\xEA\xB3"
-     "\xB0\x00\xEA\xB3\xB1\x00\xEA\xB3\xB2\x00\xEA\xB3\xB3\x00\xEA\xB3\xB4\x00"
-     "\xEA\xB3"
-     "\xB5\x00\xEA\xB3\xB6\x00\xEA\xB3\xB7\x00\xEA\xB3\xB8\x00\xEA\xB3\xB9\x00"
-     "\xEA\xB3"
-     "\xBA\x00\xEA\xB3\xBB\x00\xEA\xB3\xBC\x00\xEA\xB3\xBD\x00\xEA\xB3\xBE\x00"
-     "\xEA\xB3"
-     "\xBF\x00\xEA\xB4\x80\x00\xEA\xB4\x81\x00\xEA\xB4\x82\x00\xEA\xB4\x83\x00"
-     "\xEA\xB4"
-     "\x84\x00\xEA\xB4\x85\x00\xEA\xB4\x86\x00\xEA\xB4\x87\x00\xEA\xB4\x88\x00"
-     "\xEA\xB4"
-     "\x89\x00\xEA\xB4\x8A\x00\xEA\xB4\x8B\x00\xEA\xB4\x8C\x00\xEA\xB4\x8D\x00"
-     "\xEA\xB4"
-     "\x8E\x00\xEA\xB4\x8F\x00\xEA\xB4\x90\x00\xEA\xB4\x91\x00\xEA\xB4\x92\x00"
-     "\xEA\xB4"
-     "\x93\x00\xEA\xB4\x94\x00\xEA\xB4\x95\x00\xEA\xB4\x96\x00\xEA\xB4\x97\x00"
-     "\xEA\xB4"
-     "\x98\x00\xEA\xB4\x99\x00\xEA\xB4\x9A\x00\xEA\xB4\x9B\x00\xEA\xB4\x9C\x00"
-     "\xEA\xB4"
-     "\x9D\x00\xEA\xB4\x9E\x00\xEA\xB4\x9F\x00\xEA\xB4\xA0\x00\xEA\xB4\xA1\x00"
-     "\xEA\xB4"
-     "\xA2\x00\xEA\xB4\xA3\x00\xEA\xB4\xA4\x00\xEA\xB4\xA5\x00\xEA\xB4\xA6\x00"
-     "\xEA\xB4"
-     "\xA7\x00\xEA\xB4\xA8\x00\xEA\xB4\xA9\x00\xEA\xB4\xAA\x00\xEA\xB4\xAB\x00"
-     "\xEA\xB4"
-     "\xAC\x00\xEA\xB4\xAD\x00\xEA\xB4\xAE\x00\xEA\xB4\xAF\x00\xEA\xB4\xB0\x00"
-     "\xEA\xB4"
-     "\xB1\x00\xEA\xB4\xB2\x00\xEA\xB4\xB3\x00\xEA\xB4\xB4\x00\xEA\xB4\xB5\x00"
-     "\xEA\xB4"
-     "\xB6\x00\xEA\xB4\xB7\x00\xEA\xB4\xB8\x00\xEA\xB4\xB9\x00\xEA\xB4\xBA\x00"
-     "\xEA\xB4"
-     "\xBB\x00\xEA\xB4\xBC\x00\xEA\xB4\xBD\x00\xEA\xB4\xBE\x00\xEA\xB4\xBF\x00"
-     "\xEA\xB5"
-     "\x80\x00\xEA\xB5\x81\x00\xEA\xB5\x82\x00\xEA\xB5\x83\x00\xEA\xB5\x84\x00"
-     "\xEA\xB5"
-     "\x85\x00\xEA\xB5\x86\x00\xEA\xB5\x87\x00\xEA\xB5\x88\x00\xEA\xB5\x89\x00"
-     "\xEA\xB5"
-     "\x8A\x00\xEA\xB5\x8B\x00\xEA\xB5\x8C\x00\xEA\xB5\x8D\x00\xEA\xB5\x8E\x00"
-     "\xEA\xB5"
-     "\x8F\x00\xEA\xB5\x90\x00\xEA\xB5\x91\x00\xEA\xB5\x92\x00\xEA\xB5\x93\x00"
-     "\xEA\xB5"
-     "\x94\x00\xEA\xB5\x95\x00\xEA\xB5\x96\x00\xEA\xB5\x97\x00\xEA\xB5\x98\x00"
-     "\xEA\xB5"
-     "\x99\x00\xEA\xB5\x9A\x00\xEA\xB5\x9B\x00\xEA\xB5\x9C\x00\xEA\xB5\x9D\x00"
-     "\xEA\xB5"
-     "\x9E\x00\xEA\xB5\x9F\x00\xEA\xB5\xA0\x00\xEA\xB5\xA1\x00\xEA\xB5\xA2\x00"
-     "\xEA\xB5"
-     "\xA3\x00\xEA\xB5\xA4\x00\xEA\xB5\xA5\x00\xEA\xB5\xA6\x00\xEA\xB5\xA7\x00"
-     "\xEA\xB5"
-     "\xA8\x00\xEA\xB5\xA9\x00\xEA\xB5\xAA\x00\xEA\xB5\xAB\x00\xEA\xB5\xAC\x00"
-     "\xEA\xB5"
-     "\xAD\x00\xEA\xB5\xAE\x00\xEA\xB5\xAF\x00\xEA\xB5\xB0\x00\xEA\xB5\xB1\x00"
-     "\xEA\xB5"
-     "\xB2\x00\xEA\xB5\xB3\x00\xEA\xB5\xB4\x00\xEA\xB5\xB5\x00\xEA\xB5\xB6\x00"
-     "\xEA\xB5"
-     "\xB7\x00\xEA\xB5\xB8\x00\xEA\xB5\xB9\x00\xEA\xB5\xBA\x00\xEA\xB5\xBB\x00"
-     "\xEA\xB5"
-     "\xBC\x00\xEA\xB5\xBD\x00\xEA\xB5\xBE\x00\xEA\xB5\xBF\x00\xEA\xB6\x80\x00"
-     "\xEA\xB6"
-     "\x81\x00\xEA\xB6\x82\x00\xEA\xB6\x83\x00\xEA\xB6\x84\x00\xEA\xB6\x85\x00"
-     "\xEA\xB6"
-     "\x86\x00\xEA\xB6\x87\x00\xEA\xB6\x88\x00\xEA\xB6\x89\x00\xEA\xB6\x8A\x00"
-     "\xEA\xB6"
-     "\x8B\x00\xEA\xB6\x8C\x00\xEA\xB6\x8D\x00\xEA\xB6\x8E\x00\xEA\xB6\x8F\x00"
-     "\xEA\xB6"
-     "\x90\x00\xEA\xB6\x91\x00\xEA\xB6\x92\x00\xEA\xB6\x93\x00\xEA\xB6\x94\x00"
-     "\xEA\xB6"
-     "\x95\x00\xEA\xB6\x96\x00\xEA\xB6\x97\x00\xEA\xB6\x98\x00\xEA\xB6\x99\x00"
-     "\xEA\xB6"
-     "\x9A\x00\xEA\xB6\x9B\x00\xEA\xB6\x9C\x00\xEA\xB6\x9D\x00\xEA\xB6\x9E\x00"
-     "\xEA\xB6"
-     "\x9F\x00\xEA\xB6\xA0\x00\xEA\xB6\xA1\x00\xEA\xB6\xA2\x00\xEA\xB6\xA3\x00"
-     "\xEA\xB6"
-     "\xA4\x00\xEA\xB6\xA5\x00\xEA\xB6\xA6\x00\xEA\xB6\xA7\x00\xEA\xB6\xA8\x00"
-     "\xEA\xB6"
-     "\xA9\x00\xEA\xB6\xAA\x00\xEA\xB6\xAB\x00\xEA\xB6\xAC\x00\xEA\xB6\xAD\x00"
-     "\xEA\xB6"
-     "\xAE\x00\xEA\xB6\xAF\x00\xEA\xB6\xB0\x00\xEA\xB6\xB1\x00\xEA\xB6\xB2\x00"
-     "\xEA\xB6"
-     "\xB3\x00\xEA\xB6\xB4\x00\xEA\xB6\xB5\x00\xEA\xB6\xB6\x00\xEA\xB6\xB7\x00"
-     "\xEA\xB6"
-     "\xB8\x00\xEA\xB6\xB9\x00\xEA\xB6\xBA\x00\xEA\xB6\xBB\x00\xEA\xB6\xBC\x00"
-     "\xEA\xB6"
-     "\xBD\x00\xEA\xB6\xBE\x00\xEA\xB6\xBF\x00\xEA\xB7\x80\x00\xEA\xB7\x81\x00"
-     "\xEA\xB7"
-     "\x82\x00\xEA\xB7\x83\x00\xEA\xB7\x84\x00\xEA\xB7\x85\x00\xEA\xB7\x86\x00"
-     "\xEA\xB7"
-     "\x87\x00\xEA\xB7\x88\x00\xEA\xB7\x89\x00\xEA\xB7\x8A\x00\xEA\xB7\x8B\x00"
-     "\xEA\xB7"
-     "\x8C\x00\xEA\xB7\x8D\x00\xEA\xB7\x8E\x00\xEA\xB7\x8F\x00\xEA\xB7\x90\x00"
-     "\xEA\xB7"
-     "\x91\x00\xEA\xB7\x92\x00\xEA\xB7\x93\x00\xEA\xB7\x94\x00\xEA\xB7\x95\x00"
-     "\xEA\xB7"
-     "\x96\x00\xEA\xB7\x97\x00\xEA\xB7\x98\x00\xEA\xB7\x99\x00\xEA\xB7\x9A\x00"
-     "\xEA\xB7"
-     "\x9B\x00\xEA\xB7\x9C\x00\xEA\xB7\x9D\x00\xEA\xB7\x9E\x00\xEA\xB7\x9F\x00"
-     "\xEA\xB7"
-     "\xA0\x00\xEA\xB7\xA1\x00\xEA\xB7\xA2\x00\xEA\xB7\xA3\x00\xEA\xB7\xA4\x00"
-     "\xEA\xB7"
-     "\xA5\x00\xEA\xB7\xA6\x00\xEA\xB7\xA7\x00\xEA\xB7\xA8\x00\xEA\xB7\xA9\x00"
-     "\xEA\xB7"
-     "\xAA\x00\xEA\xB7\xAB\x00\xEA\xB7\xAC\x00\xEA\xB7\xAD\x00\xEA\xB7\xAE\x00"
-     "\xEA\xB7"
-     "\xAF\x00\xEA\xB7\xB0\x00\xEA\xB7\xB1\x00\xEA\xB7\xB2\x00\xEA\xB7\xB3\x00"
-     "\xEA\xB7"
-     "\xB4\x00\xEA\xB7\xB5\x00\xEA\xB7\xB6\x00\xEA\xB7\xB7\x00\xEA\xB7\xB8\x00"
-     "\xEA\xB7"
-     "\xB9\x00\xEA\xB7\xBA\x00\xEA\xB7\xBB\x00\xEA\xB7\xBC\x00\xEA\xB7\xBD\x00"
-     "\xEA\xB7"
-     "\xBE\x00\xEA\xB7\xBF\x00\xEA\xB8\x80\x00\xEA\xB8\x81\x00\xEA\xB8\x82\x00"
-     "\xEA\xB8"
-     "\x83\x00\xEA\xB8\x84\x00\xEA\xB8\x85\x00\xEA\xB8\x86\x00\xEA\xB8\x87\x00"
-     "\xEA\xB8"
-     "\x88\x00\xEA\xB8\x89\x00\xEA\xB8\x8A\x00\xEA\xB8\x8B\x00\xEA\xB8\x8C\x00"
-     "\xEA\xB8"
-     "\x8D\x00\xEA\xB8\x8E\x00\xEA\xB8\x8F\x00\xEA\xB8\x90\x00\xEA\xB8\x91\x00"
-     "\xEA\xB8"
-     "\x92\x00\xEA\xB8\x93\x00\xEA\xB8\x94\x00\xEA\xB8\x95\x00\xEA\xB8\x96\x00"
-     "\xEA\xB8"
-     "\x97\x00\xEA\xB8\x98\x00\xEA\xB8\x99\x00\xEA\xB8\x9A\x00\xEA\xB8\x9B\x00"
-     "\xEA\xB8"
-     "\x9C\x00\xEA\xB8\x9D\x00\xEA\xB8\x9E\x00\xEA\xB8\x9F\x00\xEA\xB8\xA0\x00"
-     "\xEA\xB8"
-     "\xA1\x00\xEA\xB8\xA2\x00\xEA\xB8\xA3\x00\xEA\xB8\xA4\x00\xEA\xB8\xA5\x00"
-     "\xEA\xB8"
-     "\xA6\x00\xEA\xB8\xA7\x00\xEA\xB8\xA8\x00\xEA\xB8\xA9\x00\xEA\xB8\xAA\x00"
-     "\xEA\xB8"
-     "\xAB\x00\xEA\xB8\xAC\x00\xEA\xB8\xAD\x00\xEA\xB8\xAE\x00\xEA\xB8\xAF\x00"
-     "\xEA\xB8"
-     "\xB0\x00\xEA\xB8\xB1\x00\xEA\xB8\xB2\x00\xEA\xB8\xB3\x00\xEA\xB8\xB4\x00"
-     "\xEA\xB8"
-     "\xB5\x00\xEA\xB8\xB6\x00\xEA\xB8\xB7\x00\xEA\xB8\xB8\x00\xEA\xB8\xB9\x00"
-     "\xEA\xB8"
-     "\xBA\x00\xEA\xB8\xBB\x00\xEA\xB8\xBC\x00\xEA\xB8\xBD\x00\xEA\xB8\xBE\x00"
-     "\xEA\xB8"
-     "\xBF\x00\xEA\xB9\x80\x00\xEA\xB9\x81\x00\xEA\xB9\x82\x00\xEA\xB9\x83\x00"
-     "\xEA\xB9"
-     "\x84\x00\xEA\xB9\x85\x00\xEA\xB9\x86\x00\xEA\xB9\x87\x00\xEA\xB9\x88\x00"
-     "\xEA\xB9"
-     "\x89\x00\xEA\xB9\x8A\x00\xEA\xB9\x8B\x00\xEA\xB9\x8C\x00\xEA\xB9\x8D\x00"
-     "\xEA\xB9"
-     "\x8E\x00\xEA\xB9\x8F\x00\xEA\xB9\x90\x00\xEA\xB9\x91\x00\xEA\xB9\x92\x00"
-     "\xEA\xB9"
-     "\x93\x00\xEA\xB9\x94\x00\xEA\xB9\x95\x00\xEA\xB9\x96\x00\xEA\xB9\x97\x00"
-     "\xEA\xB9"
-     "\x98\x00\xEA\xB9\x99\x00\xEA\xB9\x9A\x00\xEA\xB9\x9B\x00\xEA\xB9\x9C\x00"
-     "\xEA\xB9"
-     "\x9D\x00\xEA\xB9\x9E\x00\xEA\xB9\x9F\x00\xEA\xB9\xA0\x00\xEA\xB9\xA1\x00"
-     "\xEA\xB9"
-     "\xA2\x00\xEA\xB9\xA3\x00\xEA\xB9\xA4\x00\xEA\xB9\xA5\x00\xEA\xB9\xA6\x00"
-     "\xEA\xB9"
-     "\xA7\x00\xEA\xB9\xA8\x00\xEA\xB9\xA9\x00\xEA\xB9\xAA\x00\xEA\xB9\xAB\x00"
-     "\xEA\xB9"
-     "\xAC\x00\xEA\xB9\xAD\x00\xEA\xB9\xAE\x00\xEA\xB9\xAF\x00\xEA\xB9\xB0\x00"
-     "\xEA\xB9"
-     "\xB1\x00\xEA\xB9\xB2\x00\xEA\xB9\xB3\x00\xEA\xB9\xB4\x00\xEA\xB9\xB5\x00"
-     "\xEA\xB9"
-     "\xB6\x00\xEA\xB9\xB7\x00\xEA\xB9\xB8\x00\xEA\xB9\xB9\x00\xEA\xB9\xBA\x00"
-     "\xEA\xB9"
-     "\xBB\x00\xEA\xB9\xBC\x00\xEA\xB9\xBD\x00\xEA\xB9\xBE\x00\xEA\xB9\xBF\x00"
-     "\xEA\xBA"
-     "\x80\x00\xEA\xBA\x81\x00\xEA\xBA\x82\x00\xEA\xBA\x83\x00\xEA\xBA\x84\x00"
-     "\xEA\xBA"
-     "\x85\x00\xEA\xBA\x86\x00\xEA\xBA\x87\x00\xEA\xBA\x88\x00\xEA\xBA\x89\x00"
-     "\xEA\xBA"
-     "\x8A\x00\xEA\xBA\x8B\x00\xEA\xBA\x8C\x00\xEA\xBA\x8D\x00\xEA\xBA\x8E\x00"
-     "\xEA\xBA"
-     "\x8F\x00\xEA\xBA\x90\x00\xEA\xBA\x91\x00\xEA\xBA\x92\x00\xEA\xBA\x93\x00"
-     "\xEA\xBA"
-     "\x94\x00\xEA\xBA\x95\x00\xEA\xBA\x96\x00\xEA\xBA\x97\x00\xEA\xBA\x98\x00"
-     "\xEA\xBA"
-     "\x99\x00\xEA\xBA\x9A\x00\xEA\xBA\x9B\x00\xEA\xBA\x9C\x00\xEA\xBA\x9D\x00"
-     "\xEA\xBA"
-     "\x9E\x00\xEA\xBA\x9F\x00\xEA\xBA\xA0\x00\xEA\xBA\xA1\x00\xEA\xBA\xA2\x00"
-     "\xEA\xBA"
-     "\xA3\x00\xEA\xBA\xA4\x00\xEA\xBA\xA5\x00\xEA\xBA\xA6\x00\xEA\xBA\xA7\x00"
-     "\xEA\xBA"
-     "\xA8\x00\xEA\xBA\xA9\x00\xEA\xBA\xAA\x00\xEA\xBA\xAB\x00\xEA\xBA\xAC\x00"
-     "\xEA\xBA"
-     "\xAD\x00\xEA\xBA\xAE\x00\xEA\xBA\xAF\x00\xEA\xBA\xB0\x00\xEA\xBA\xB1\x00"
-     "\xEA\xBA"
-     "\xB2\x00\xEA\xBA\xB3\x00\xEA\xBA\xB4\x00\xEA\xBA\xB5\x00\xEA\xBA\xB6\x00"
-     "\xEA\xBA"
-     "\xB7\x00\xEA\xBA\xB8\x00\xEA\xBA\xB9\x00\xEA\xBA\xBA\x00\xEA\xBA\xBB\x00"
-     "\xEA\xBA"
-     "\xBC\x00\xEA\xBA\xBD\x00\xEA\xBA\xBE\x00\xEA\xBA\xBF\x00\xEA\xBB\x80\x00"
-     "\xEA\xBB"
-     "\x81\x00\xEA\xBB\x82\x00\xEA\xBB\x83\x00\xEA\xBB\x84\x00\xEA\xBB\x85\x00"
-     "\xEA\xBB"
-     "\x86\x00\xEA\xBB\x87\x00\xEA\xBB\x88\x00\xEA\xBB\x89\x00\xEA\xBB\x8A\x00"
-     "\xEA\xBB"
-     "\x8B\x00\xEA\xBB\x8C\x00\xEA\xBB\x8D\x00\xEA\xBB\x8E\x00\xEA\xBB\x8F\x00"
-     "\xEA\xBB"
-     "\x90\x00\xEA\xBB\x91\x00\xEA\xBB\x92\x00\xEA\xBB\x93\x00\xEA\xBB\x94\x00"
-     "\xEA\xBB"
-     "\x95\x00\xEA\xBB\x96\x00\xEA\xBB\x97\x00\xEA\xBB\x98\x00\xEA\xBB\x99\x00"
-     "\xEA\xBB"
-     "\x9A\x00\xEA\xBB\x9B\x00\xEA\xBB\x9C\x00\xEA\xBB\x9D\x00\xEA\xBB\x9E\x00"
-     "\xEA\xBB"
-     "\x9F\x00\xEA\xBB\xA0\x00\xEA\xBB\xA1\x00\xEA\xBB\xA2\x00\xEA\xBB\xA3\x00"
-     "\xEA\xBB"
-     "\xA4\x00\xEA\xBB\xA5\x00\xEA\xBB\xA6\x00\xEA\xBB\xA7\x00\xEA\xBB\xA8\x00"
-     "\xEA\xBB"
-     "\xA9\x00\xEA\xBB\xAA\x00\xEA\xBB\xAB\x00\xEA\xBB\xAC\x00\xEA\xBB\xAD\x00"
-     "\xEA\xBB"
-     "\xAE\x00\xEA\xBB\xAF\x00\xEA\xBB\xB0\x00\xEA\xBB\xB1\x00\xEA\xBB\xB2\x00"
-     "\xEA\xBB"
-     "\xB3\x00\xEA\xBB\xB4\x00\xEA\xBB\xB5\x00\xEA\xBB\xB6\x00\xEA\xBB\xB7\x00"
-     "\xEA\xBB"
-     "\xB8\x00\xEA\xBB\xB9\x00\xEA\xBB\xBA\x00\xEA\xBB\xBB\x00\xEA\xBB\xBC\x00"
-     "\xEA\xBB"
-     "\xBD\x00\xEA\xBB\xBE\x00\xEA\xBB\xBF\x00\xEA\xBC\x80\x00\xEA\xBC\x81\x00"
-     "\xEA\xBC"
-     "\x82\x00\xEA\xBC\x83\x00\xEA\xBC\x84\x00\xEA\xBC\x85\x00\xEA\xBC\x86\x00"
-     "\xEA\xBC"
-     "\x87\x00\xEA\xBC\x88\x00\xEA\xBC\x89\x00\xEA\xBC\x8A\x00\xEA\xBC\x8B\x00"
-     "\xEA\xBC"
-     "\x8C\x00\xEA\xBC\x8D\x00\xEA\xBC\x8E\x00\xEA\xBC\x8F\x00\xEA\xBC\x90\x00"
-     "\xEA\xBC"
-     "\x91\x00\xEA\xBC\x92\x00\xEA\xBC\x93\x00\xEA\xBC\x94\x00\xEA\xBC\x95\x00"
-     "\xEA\xBC"
-     "\x96\x00\xEA\xBC\x97\x00\xEA\xBC\x98\x00\xEA\xBC\x99\x00\xEA\xBC\x9A\x00"
-     "\xEA\xBC"
-     "\x9B\x00\xEA\xBC\x9C\x00\xEA\xBC\x9D\x00\xEA\xBC\x9E\x00\xEA\xBC\x9F\x00"
-     "\xEA\xBC"
-     "\xA0\x00\xEA\xBC\xA1\x00\xEA\xBC\xA2\x00\xEA\xBC\xA3\x00\xEA\xBC\xA4\x00"
-     "\xEA\xBC"
-     "\xA5\x00\xEA\xBC\xA6\x00\xEA\xBC\xA7\x00\xEA\xBC\xA8\x00\xEA\xBC\xA9\x00"
-     "\xEA\xBC"
-     "\xAA\x00\xEA\xBC\xAB\x00\xEA\xBC\xAC\x00\xEA\xBC\xAD\x00\xEA\xBC\xAE\x00"
-     "\xEA\xBC"
-     "\xAF\x00\xEA\xBC\xB0\x00\xEA\xBC\xB1\x00\xEA\xBC\xB2\x00\xEA\xBC\xB3\x00"
-     "\xEA\xBC"
-     "\xB4\x00\xEA\xBC\xB5\x00\xEA\xBC\xB6\x00\xEA\xBC\xB7\x00\xEA\xBC\xB8\x00"
-     "\xEA\xBC"
-     "\xB9\x00\xEA\xBC\xBA\x00\xEA\xBC\xBB\x00\xEA\xBC\xBC\x00\xEA\xBC\xBD\x00"
-     "\xEA\xBC"
-     "\xBE\x00\xEA\xBC\xBF\x00\xEA\xBD\x80\x00\xEA\xBD\x81\x00\xEA\xBD\x82\x00"
-     "\xEA\xBD"
-     "\x83\x00\xEA\xBD\x84\x00\xEA\xBD\x85\x00\xEA\xBD\x86\x00\xEA\xBD\x87\x00"
-     "\xEA\xBD"
-     "\x88\x00\xEA\xBD\x89\x00\xEA\xBD\x8A\x00\xEA\xBD\x8B\x00\xEA\xBD\x8C\x00"
-     "\xEA\xBD"
-     "\x8D\x00\xEA\xBD\x8E\x00\xEA\xBD\x8F\x00\xEA\xBD\x90\x00\xEA\xBD\x91\x00"
-     "\xEA\xBD"
-     "\x92\x00\xEA\xBD\x93\x00\xEA\xBD\x94\x00\xEA\xBD\x95\x00\xEA\xBD\x96\x00"
-     "\xEA\xBD"
-     "\x97\x00\xEA\xBD\x98\x00\xEA\xBD\x99\x00\xEA\xBD\x9A\x00\xEA\xBD\x9B\x00"
-     "\xEA\xBD"
-     "\x9C\x00\xEA\xBD\x9D\x00\xEA\xBD\x9E\x00\xEA\xBD\x9F\x00\xEA\xBD\xA0\x00"
-     "\xEA\xBD"
-     "\xA1\x00\xEA\xBD\xA2\x00\xEA\xBD\xA3\x00\xEA\xBD\xA4\x00\xEA\xBD\xA5\x00"
-     "\xEA\xBD"
-     "\xA6\x00\xEA\xBD\xA7\x00\xEA\xBD\xA8\x00\xEA\xBD\xA9\x00\xEA\xBD\xAA\x00"
-     "\xEA\xBD"
-     "\xAB\x00\xEA\xBD\xAC\x00\xEA\xBD\xAD\x00\xEA\xBD\xAE\x00\xEA\xBD\xAF\x00"
-     "\xEA\xBD"
-     "\xB0\x00\xEA\xBD\xB1\x00\xEA\xBD\xB2\x00\xEA\xBD\xB3\x00\xEA\xBD\xB4\x00"
-     "\xEA\xBD"
-     "\xB5\x00\xEA\xBD\xB6\x00\xEA\xBD\xB7\x00\xEA\xBD\xB8\x00\xEA\xBD\xB9\x00"
-     "\xEA\xBD"
-     "\xBA\x00\xEA\xBD\xBB\x00\xEA\xBD\xBC\x00\xEA\xBD\xBD\x00\xEA\xBD\xBE\x00"
-     "\xEA\xBD"
-     "\xBF\x00\xEA\xBE\x80\x00\xEA\xBE\x81\x00\xEA\xBE\x82\x00\xEA\xBE\x83\x00"
-     "\xEA\xBE"
-     "\x84\x00\xEA\xBE\x85\x00\xEA\xBE\x86\x00\xEA\xBE\x87\x00\xEA\xBE\x88\x00"
-     "\xEA\xBE"
-     "\x89\x00\xEA\xBE\x8A\x00\xEA\xBE\x8B\x00\xEA\xBE\x8C\x00\xEA\xBE\x8D\x00"
-     "\xEA\xBE"
-     "\x8E\x00\xEA\xBE\x8F\x00\xEA\xBE\x90\x00\xEA\xBE\x91\x00\xEA\xBE\x92\x00"
-     "\xEA\xBE"
-     "\x93\x00\xEA\xBE\x94\x00\xEA\xBE\x95\x00\xEA\xBE\x96\x00\xEA\xBE\x97\x00"
-     "\xEA\xBE"
-     "\x98\x00\xEA\xBE\x99\x00\xEA\xBE\x9A\x00\xEA\xBE\x9B\x00\xEA\xBE\x9C\x00"
-     "\xEA\xBE"
-     "\x9D\x00\xEA\xBE\x9E\x00\xEA\xBE\x9F\x00\xEA\xBE\xA0\x00\xEA\xBE\xA1\x00"
-     "\xEA\xBE"
-     "\xA2\x00\xEA\xBE\xA3\x00\xEA\xBE\xA4\x00\xEA\xBE\xA5\x00\xEA\xBE\xA6\x00"
-     "\xEA\xBE"
-     "\xA7\x00\xEA\xBE\xA8\x00\xEA\xBE\xA9\x00\xEA\xBE\xAA\x00\xEA\xBE\xAB\x00"
-     "\xEA\xBE"
-     "\xAC\x00\xEA\xBE\xAD\x00\xEA\xBE\xAE\x00\xEA\xBE\xAF\x00\xEA\xBE\xB0\x00"
-     "\xEA\xBE"
-     "\xB1\x00\xEA\xBE\xB2\x00\xEA\xBE\xB3\x00\xEA\xBE\xB4\x00\xEA\xBE\xB5\x00"
-     "\xEA\xBE"
-     "\xB6\x00\xEA\xBE\xB7\x00\xEA\xBE\xB8\x00\xEA\xBE\xB9\x00\xEA\xBE\xBA\x00"
-     "\xEA\xBE"
-     "\xBB\x00\xEA\xBE\xBC\x00\xEA\xBE\xBD\x00\xEA\xBE\xBE\x00\xEA\xBE\xBF\x00"
-     "\xEA\xBF"
-     "\x80\x00\xEA\xBF\x81\x00\xEA\xBF\x82\x00\xEA\xBF\x83\x00\xEA\xBF\x84\x00"
-     "\xEA\xBF"
-     "\x85\x00\xEA\xBF\x86\x00\xEA\xBF\x87\x00\xEA\xBF\x88\x00\xEA\xBF\x89\x00"
-     "\xEA\xBF"
-     "\x8A\x00\xEA\xBF\x8B\x00\xEA\xBF\x8C\x00\xEA\xBF\x8D\x00\xEA\xBF\x8E\x00"
-     "\xEA\xBF"
-     "\x8F\x00\xEA\xBF\x90\x00\xEA\xBF\x91\x00\xEA\xBF\x92\x00\xEA\xBF\x93\x00"
-     "\xEA\xBF"
-     "\x94\x00\xEA\xBF\x95\x00\xEA\xBF\x96\x00\xEA\xBF\x97\x00\xEA\xBF\x98\x00"
-     "\xEA\xBF"
-     "\x99\x00\xEA\xBF\x9A\x00\xEA\xBF\x9B\x00\xEA\xBF\x9C\x00\xEA\xBF\x9D\x00"
-     "\xEA\xBF"
-     "\x9E\x00\xEA\xBF\x9F\x00\xEA\xBF\xA0\x00\xEA\xBF\xA1\x00\xEA\xBF\xA2\x00"
-     "\xEA\xBF"
-     "\xA3\x00\xEA\xBF\xA4\x00\xEA\xBF\xA5\x00\xEA\xBF\xA6\x00\xEA\xBF\xA7\x00"
-     "\xEA\xBF"
-     "\xA8\x00\xEA\xBF\xA9\x00\xEA\xBF\xAA\x00\xEA\xBF\xAB\x00\xEA\xBF\xAC\x00"
-     "\xEA\xBF"
-     "\xAD\x00\xEA\xBF\xAE\x00\xEA\xBF\xAF\x00\xEA\xBF\xB0\x00\xEA\xBF\xB1\x00"
-     "\xEA\xBF"
-     "\xB2\x00\xEA\xBF\xB3\x00\xEA\xBF\xB4\x00\xEA\xBF\xB5\x00\xEA\xBF\xB6\x00"
-     "\xEA\xBF"
-     "\xB7\x00\xEA\xBF\xB8\x00\xEA\xBF\xB9\x00\xEA\xBF\xBA\x00\xEA\xBF\xBB\x00"
-     "\xEA\xBF"
-     "\xBC\x00\xEA\xBF\xBD\x00\xEA\xBF\xBE\x00\xEA\xBF\xBF\x00\xEB\x80\x80\x00"
-     "\xEB\x80"
-     "\x81\x00\xEB\x80\x82\x00\xEB\x80\x83\x00\xEB\x80\x84\x00\xEB\x80\x85\x00"
-     "\xEB\x80"
-     "\x86\x00\xEB\x80\x87\x00\xEB\x80\x88\x00\xEB\x80\x89\x00\xEB\x80\x8A\x00"
-     "\xEB\x80"
-     "\x8B\x00\xEB\x80\x8C\x00\xEB\x80\x8D\x00\xEB\x80\x8E\x00\xEB\x80\x8F\x00"
-     "\xEB\x80"
-     "\x90\x00\xEB\x80\x91\x00\xEB\x80\x92\x00\xEB\x80\x93\x00\xEB\x80\x94\x00"
-     "\xEB\x80"
-     "\x95\x00\xEB\x80\x96\x00\xEB\x80\x97\x00\xEB\x80\x98\x00\xEB\x80\x99\x00"
-     "\xEB\x80"
-     "\x9A\x00\xEB\x80\x9B\x00\xEB\x80\x9C\x00\xEB\x80\x9D\x00\xEB\x80\x9E\x00"
-     "\xEB\x80"
-     "\x9F\x00\xEB\x80\xA0\x00\xEB\x80\xA1\x00\xEB\x80\xA2\x00\xEB\x80\xA3\x00"
-     "\xEB\x80"
-     "\xA4\x00\xEB\x80\xA5\x00\xEB\x80\xA6\x00\xEB\x80\xA7\x00\xEB\x80\xA8\x00"
-     "\xEB\x80"
-     "\xA9\x00\xEB\x80\xAA\x00\xEB\x80\xAB\x00\xEB\x80\xAC\x00\xEB\x80\xAD\x00"
-     "\xEB\x80"
-     "\xAE\x00\xEB\x80\xAF\x00\xEB\x80\xB0\x00\xEB\x80\xB1\x00\xEB\x80\xB2\x00"
-     "\xEB\x80"
-     "\xB3\x00\xEB\x80\xB4\x00\xEB\x80\xB5\x00\xEB\x80\xB6\x00\xEB\x80\xB7\x00"
-     "\xEB\x80"
-     "\xB8\x00\xEB\x80\xB9\x00\xEB\x80\xBA\x00\xEB\x80\xBB\x00\xEB\x80\xBC\x00"
-     "\xEB\x80"
-     "\xBD\x00\xEB\x80\xBE\x00\xEB\x80\xBF\x00\xEB\x81\x80\x00\xEB\x81\x81\x00"
-     "\xEB\x81"
-     "\x82\x00\xEB\x81\x83\x00\xEB\x81\x84\x00\xEB\x81\x85\x00\xEB\x81\x86\x00"
-     "\xEB\x81"
-     "\x87\x00\xEB\x81\x88\x00\xEB\x81\x89\x00\xEB\x81\x8A\x00\xEB\x81\x8B\x00"
-     "\xEB\x81"
-     "\x8C\x00\xEB\x81\x8D\x00\xEB\x81\x8E\x00\xEB\x81\x8F\x00\xEB\x81\x90\x00"
-     "\xEB\x81"
-     "\x91\x00\xEB\x81\x92\x00\xEB\x81\x93\x00\xEB\x81\x94\x00\xEB\x81\x95\x00"
-     "\xEB\x81"
-     "\x96\x00\xEB\x81\x97\x00\xEB\x81\x98\x00\xEB\x81\x99\x00\xEB\x81\x9A\x00"
-     "\xEB\x81"
-     "\x9B\x00\xEB\x81\x9C\x00\xEB\x81\x9D\x00\xEB\x81\x9E\x00\xEB\x81\x9F\x00"
-     "\xEB\x81"
-     "\xA0\x00\xEB\x81\xA1\x00\xEB\x81\xA2\x00\xEB\x81\xA3\x00\xEB\x81\xA4\x00"
-     "\xEB\x81"
-     "\xA5\x00\xEB\x81\xA6\x00\xEB\x81\xA7\x00\xEB\x81\xA8\x00\xEB\x81\xA9\x00"
-     "\xEB\x81"
-     "\xAA\x00\xEB\x81\xAB\x00\xEB\x81\xAC\x00\xEB\x81\xAD\x00\xEB\x81\xAE\x00"
-     "\xEB\x81"
-     "\xAF\x00\xEB\x81\xB0\x00\xEB\x81\xB1\x00\xEB\x81\xB2\x00\xEB\x81\xB3\x00"
-     "\xEB\x81"
-     "\xB4\x00\xEB\x81\xB5\x00\xEB\x81\xB6\x00\xEB\x81\xB7\x00\xEB\x81\xB8\x00"
-     "\xEB\x81"
-     "\xB9\x00\xEB\x81\xBA\x00\xEB\x81\xBB\x00\xEB\x81\xBC\x00\xEB\x81\xBD\x00"
-     "\xEB\x81"
-     "\xBE\x00\xEB\x81\xBF\x00\xEB\x82\x80\x00\xEB\x82\x81\x00\xEB\x82\x82\x00"
-     "\xEB\x82"
-     "\x83\x00\xEB\x82\x84\x00\xEB\x82\x85\x00\xEB\x82\x86\x00\xEB\x82\x87\x00"
-     "\xEB\x82"
-     "\x88\x00\xEB\x82\x89\x00\xEB\x82\x8A\x00\xEB\x82\x8B\x00\xEB\x82\x8C\x00"
-     "\xEB\x82"
-     "\x8D\x00\xEB\x82\x8E\x00\xEB\x82\x8F\x00\xEB\x82\x90\x00\xEB\x82\x91\x00"
-     "\xEB\x82"
-     "\x92\x00\xEB\x82\x93\x00\xEB\x82\x94\x00\xEB\x82\x95\x00\xEB\x82\x96\x00"
-     "\xEB\x82"
-     "\x97\x00\xEB\x82\x98\x00\xEB\x82\x99\x00\xEB\x82\x9A\x00\xEB\x82\x9B\x00"
-     "\xEB\x82"
-     "\x9C\x00\xEB\x82\x9D\x00\xEB\x82\x9E\x00\xEB\x82\x9F\x00\xEB\x82\xA0\x00"
-     "\xEB\x82"
-     "\xA1\x00\xEB\x82\xA2\x00\xEB\x82\xA3\x00\xEB\x82\xA4\x00\xEB\x82\xA5\x00"
-     "\xEB\x82"
-     "\xA6\x00\xEB\x82\xA7\x00\xEB\x82\xA8\x00\xEB\x82\xA9\x00\xEB\x82\xAA\x00"
-     "\xEB\x82"
-     "\xAB\x00\xEB\x82\xAC\x00\xEB\x82\xAD\x00\xEB\x82\xAE\x00\xEB\x82\xAF\x00"
-     "\xEB\x82"
-     "\xB0\x00\xEB\x82\xB1\x00\xEB\x82\xB2\x00\xEB\x82\xB3\x00\xEB\x82\xB4\x00"
-     "\xEB\x82"
-     "\xB5\x00\xEB\x82\xB6\x00\xEB\x82\xB7\x00\xEB\x82\xB8\x00\xEB\x82\xB9\x00"
-     "\xEB\x82"
-     "\xBA\x00\xEB\x82\xBB\x00\xEB\x82\xBC\x00\xEB\x82\xBD\x00\xEB\x82\xBE\x00"
-     "\xEB\x82"
-     "\xBF\x00\xEB\x83\x80\x00\xEB\x83\x81\x00\xEB\x83\x82\x00\xEB\x83\x83\x00"
-     "\xEB\x83"
-     "\x84\x00\xEB\x83\x85\x00\xEB\x83\x86\x00\xEB\x83\x87\x00\xEB\x83\x88\x00"
-     "\xEB\x83"
-     "\x89\x00\xEB\x83\x8A\x00\xEB\x83\x8B\x00\xEB\x83\x8C\x00\xEB\x83\x8D\x00"
-     "\xEB\x83"
-     "\x8E\x00\xEB\x83\x8F\x00\xEB\x83\x90\x00\xEB\x83\x91\x00\xEB\x83\x92\x00"
-     "\xEB\x83"
-     "\x93\x00\xEB\x83\x94\x00\xEB\x83\x95\x00\xEB\x83\x96\x00\xEB\x83\x97\x00"
-     "\xEB\x83"
-     "\x98\x00\xEB\x83\x99\x00\xEB\x83\x9A\x00\xEB\x83\x9B\x00\xEB\x83\x9C\x00"
-     "\xEB\x83"
-     "\x9D\x00\xEB\x83\x9E\x00\xEB\x83\x9F\x00\xEB\x83\xA0\x00\xEB\x83\xA1\x00"
-     "\xEB\x83"
-     "\xA2\x00\xEB\x83\xA3\x00\xEB\x83\xA4\x00\xEB\x83\xA5\x00\xEB\x83\xA6\x00"
-     "\xEB\x83"
-     "\xA7\x00\xEB\x83\xA8\x00\xEB\x83\xA9\x00\xEB\x83\xAA\x00\xEB\x83\xAB\x00"
-     "\xEB\x83"
-     "\xAC\x00\xEB\x83\xAD\x00\xEB\x83\xAE\x00\xEB\x83\xAF\x00\xEB\x83\xB0\x00"
-     "\xEB\x83"
-     "\xB1\x00\xEB\x83\xB2\x00\xEB\x83\xB3\x00\xEB\x83\xB4\x00\xEB\x83\xB5\x00"
-     "\xEB\x83"
-     "\xB6\x00\xEB\x83\xB7\x00\xEB\x83\xB8\x00\xEB\x83\xB9\x00\xEB\x83\xBA\x00"
-     "\xEB\x83"
-     "\xBB\x00\xEB\x83\xBC\x00\xEB\x83\xBD\x00\xEB\x83\xBE\x00\xEB\x83\xBF\x00"
-     "\xEB\x84"
-     "\x80\x00\xEB\x84\x81\x00\xEB\x84\x82\x00\xEB\x84\x83\x00\xEB\x84\x84\x00"
-     "\xEB\x84"
-     "\x85\x00\xEB\x84\x86\x00\xEB\x84\x87\x00\xEB\x84\x88\x00\xEB\x84\x89\x00"
-     "\xEB\x84"
-     "\x8A\x00\xEB\x84\x8B\x00\xEB\x84\x8C\x00\xEB\x84\x8D\x00\xEB\x84\x8E\x00"
-     "\xEB\x84"
-     "\x8F\x00\xEB\x84\x90\x00\xEB\x84\x91\x00\xEB\x84\x92\x00\xEB\x84\x93\x00"
-     "\xEB\x84"
-     "\x94\x00\xEB\x84\x95\x00\xEB\x84\x96\x00\xEB\x84\x97\x00\xEB\x84\x98\x00"
-     "\xEB\x84"
-     "\x99\x00\xEB\x84\x9A\x00\xEB\x84\x9B\x00\xEB\x84\x9C\x00\xEB\x84\x9D\x00"
-     "\xEB\x84"
-     "\x9E\x00\xEB\x84\x9F\x00\xEB\x84\xA0\x00\xEB\x84\xA1\x00\xEB\x84\xA2\x00"
-     "\xEB\x84"
-     "\xA3\x00\xEB\x84\xA4\x00\xEB\x84\xA5\x00\xEB\x84\xA6\x00\xEB\x84\xA7\x00"
-     "\xEB\x84"
-     "\xA8\x00\xEB\x84\xA9\x00\xEB\x84\xAA\x00\xEB\x84\xAB\x00\xEB\x84\xAC\x00"
-     "\xEB\x84"
-     "\xAD\x00\xEB\x84\xAE\x00\xEB\x84\xAF\x00\xEB\x84\xB0\x00\xEB\x84\xB1\x00"
-     "\xEB\x84"
-     "\xB2\x00\xEB\x84\xB3\x00\xEB\x84\xB4\x00\xEB\x84\xB5\x00\xEB\x84\xB6\x00"
-     "\xEB\x84"
-     "\xB7\x00\xEB\x84\xB8\x00\xEB\x84\xB9\x00\xEB\x84\xBA\x00\xEB\x84\xBB\x00"
-     "\xEB\x84"
-     "\xBC\x00\xEB\x84\xBD\x00\xEB\x84\xBE\x00\xEB\x84\xBF\x00\xEB\x85\x80\x00"
-     "\xEB\x85"
-     "\x81\x00\xEB\x85\x82\x00\xEB\x85\x83\x00\xEB\x85\x84\x00\xEB\x85\x85\x00"
-     "\xEB\x85"
-     "\x86\x00\xEB\x85\x87\x00\xEB\x85\x88\x00\xEB\x85\x89\x00\xEB\x85\x8A\x00"
-     "\xEB\x85"
-     "\x8B\x00\xEB\x85\x8C\x00\xEB\x85\x8D\x00\xEB\x85\x8E\x00\xEB\x85\x8F\x00"
-     "\xEB\x85"
-     "\x90\x00\xEB\x85\x91\x00\xEB\x85\x92\x00\xEB\x85\x93\x00\xEB\x85\x94\x00"
-     "\xEB\x85"
-     "\x95\x00\xEB\x85\x96\x00\xEB\x85\x97\x00\xEB\x85\x98\x00\xEB\x85\x99\x00"
-     "\xEB\x85"
-     "\x9A\x00\xEB\x85\x9B\x00\xEB\x85\x9C\x00\xEB\x85\x9D\x00\xEB\x85\x9E\x00"
-     "\xEB\x85"
-     "\x9F\x00\xEB\x85\xA0\x00\xEB\x85\xA1\x00\xEB\x85\xA2\x00\xEB\x85\xA3\x00"
-     "\xEB\x85"
-     "\xA4\x00\xEB\x85\xA5\x00\xEB\x85\xA6\x00\xEB\x85\xA7\x00\xEB\x85\xA8\x00"
-     "\xEB\x85"
-     "\xA9\x00\xEB\x85\xAA\x00\xEB\x85\xAB\x00\xEB\x85\xAC\x00\xEB\x85\xAD\x00"
-     "\xEB\x85"
-     "\xAE\x00\xEB\x85\xAF\x00\xEB\x85\xB0\x00\xEB\x85\xB1\x00\xEB\x85\xB2\x00"
-     "\xEB\x85"
-     "\xB3\x00\xEB\x85\xB4\x00\xEB\x85\xB5\x00\xEB\x85\xB6\x00\xEB\x85\xB7\x00"
-     "\xEB\x85"
-     "\xB8\x00\xEB\x85\xB9\x00\xEB\x85\xBA\x00\xEB\x85\xBB\x00\xEB\x85\xBC\x00"
-     "\xEB\x85"
-     "\xBD\x00\xEB\x85\xBE\x00\xEB\x85\xBF\x00\xEB\x86\x80\x00\xEB\x86\x81\x00"
-     "\xEB\x86"
-     "\x82\x00\xEB\x86\x83\x00\xEB\x86\x84\x00\xEB\x86\x85\x00\xEB\x86\x86\x00"
-     "\xEB\x86"
-     "\x87\x00\xEB\x86\x88\x00\xEB\x86\x89\x00\xEB\x86\x8A\x00\xEB\x86\x8B\x00"
-     "\xEB\x86"
-     "\x8C\x00\xEB\x86\x8D\x00\xEB\x86\x8E\x00\xEB\x86\x8F\x00\xEB\x86\x90\x00"
-     "\xEB\x86"
-     "\x91\x00\xEB\x86\x92\x00\xEB\x86\x93\x00\xEB\x86\x94\x00\xEB\x86\x95\x00"
-     "\xEB\x86"
-     "\x96\x00\xEB\x86\x97\x00\xEB\x86\x98\x00\xEB\x86\x99\x00\xEB\x86\x9A\x00"
-     "\xEB\x86"
-     "\x9B\x00\xEB\x86\x9C\x00\xEB\x86\x9D\x00\xEB\x86\x9E\x00\xEB\x86\x9F\x00"
-     "\xEB\x86"
-     "\xA0\x00\xEB\x86\xA1\x00\xEB\x86\xA2\x00\xEB\x86\xA3\x00\xEB\x86\xA4\x00"
-     "\xEB\x86"
-     "\xA5\x00\xEB\x86\xA6\x00\xEB\x86\xA7\x00\xEB\x86\xA8\x00\xEB\x86\xA9\x00"
-     "\xEB\x86"
-     "\xAA\x00\xEB\x86\xAB\x00\xEB\x86\xAC\x00\xEB\x86\xAD\x00\xEB\x86\xAE\x00"
-     "\xEB\x86"
-     "\xAF\x00\xEB\x86\xB0\x00\xEB\x86\xB1\x00\xEB\x86\xB2\x00\xEB\x86\xB3\x00"
-     "\xEB\x86"
-     "\xB4\x00\xEB\x86\xB5\x00\xEB\x86\xB6\x00\xEB\x86\xB7\x00\xEB\x86\xB8\x00"
-     "\xEB\x86"
-     "\xB9\x00\xEB\x86\xBA\x00\xEB\x86\xBB\x00\xEB\x86\xBC\x00\xEB\x86\xBD\x00"
-     "\xEB\x86"
-     "\xBE\x00\xEB\x86\xBF\x00\xEB\x87\x80\x00\xEB\x87\x81\x00\xEB\x87\x82\x00"
-     "\xEB\x87"
-     "\x83\x00\xEB\x87\x84\x00\xEB\x87\x85\x00\xEB\x87\x86\x00\xEB\x87\x87\x00"
-     "\xEB\x87"
-     "\x88\x00\xEB\x87\x89\x00\xEB\x87\x8A\x00\xEB\x87\x8B\x00\xEB\x87\x8C\x00"
-     "\xEB\x87"
-     "\x8D\x00\xEB\x87\x8E\x00\xEB\x87\x8F\x00\xEB\x87\x90\x00\xEB\x87\x91\x00"
-     "\xEB\x87"
-     "\x92\x00\xEB\x87\x93\x00\xEB\x87\x94\x00\xEB\x87\x95\x00\xEB\x87\x96\x00"
-     "\xEB\x87"
-     "\x97\x00\xEB\x87\x98\x00\xEB\x87\x99\x00\xEB\x87\x9A\x00\xEB\x87\x9B\x00"
-     "\xEB\x87"
-     "\x9C\x00\xEB\x87\x9D\x00\xEB\x87\x9E\x00\xEB\x87\x9F\x00\xEB\x87\xA0\x00"
-     "\xEB\x87"
-     "\xA1\x00\xEB\x87\xA2\x00\xEB\x87\xA3\x00\xEB\x87\xA4\x00\xEB\x87\xA5\x00"
-     "\xEB\x87"
-     "\xA6\x00\xEB\x87\xA7\x00\xEB\x87\xA8\x00\xEB\x87\xA9\x00\xEB\x87\xAA\x00"
-     "\xEB\x87"
-     "\xAB\x00\xEB\x87\xAC\x00\xEB\x87\xAD\x00\xEB\x87\xAE\x00\xEB\x87\xAF\x00"
-     "\xEB\x87"
-     "\xB0\x00\xEB\x87\xB1\x00\xEB\x87\xB2\x00\xEB\x87\xB3\x00\xEB\x87\xB4\x00"
-     "\xEB\x87"
-     "\xB5\x00\xEB\x87\xB6\x00\xEB\x87\xB7\x00\xEB\x87\xB8\x00\xEB\x87\xB9\x00"
-     "\xEB\x87"
-     "\xBA\x00\xEB\x87\xBB\x00\xEB\x87\xBC\x00\xEB\x87\xBD\x00\xEB\x87\xBE\x00"
-     "\xEB\x87"
-     "\xBF\x00\xEB\x88\x80\x00\xEB\x88\x81\x00\xEB\x88\x82\x00\xEB\x88\x83\x00"
-     "\xEB\x88"
-     "\x84\x00\xEB\x88\x85\x00\xEB\x88\x86\x00\xEB\x88\x87\x00\xEB\x88\x88\x00"
-     "\xEB\x88"
-     "\x89\x00\xEB\x88\x8A\x00\xEB\x88\x8B\x00\xEB\x88\x8C\x00\xEB\x88\x8D\x00"
-     "\xEB\x88"
-     "\x8E\x00\xEB\x88\x8F\x00\xEB\x88\x90\x00\xEB\x88\x91\x00\xEB\x88\x92\x00"
-     "\xEB\x88"
-     "\x93\x00\xEB\x88\x94\x00\xEB\x88\x95\x00\xEB\x88\x96\x00\xEB\x88\x97\x00"
-     "\xEB\x88"
-     "\x98\x00\xEB\x88\x99\x00\xEB\x88\x9A\x00\xEB\x88\x9B\x00\xEB\x88\x9C\x00"
-     "\xEB\x88"
-     "\x9D\x00\xEB\x88\x9E\x00\xEB\x88\x9F\x00\xEB\x88\xA0\x00\xEB\x88\xA1\x00"
-     "\xEB\x88"
-     "\xA2\x00\xEB\x88\xA3\x00\xEB\x88\xA4\x00\xEB\x88\xA5\x00\xEB\x88\xA6\x00"
-     "\xEB\x88"
-     "\xA7\x00\xEB\x88\xA8\x00\xEB\x88\xA9\x00\xEB\x88\xAA\x00\xEB\x88\xAB\x00"
-     "\xEB\x88"
-     "\xAC\x00\xEB\x88\xAD\x00\xEB\x88\xAE\x00\xEB\x88\xAF\x00\xEB\x88\xB0\x00"
-     "\xEB\x88"
-     "\xB1\x00\xEB\x88\xB2\x00\xEB\x88\xB3\x00\xEB\x88\xB4\x00\xEB\x88\xB5\x00"
-     "\xEB\x88"
-     "\xB6\x00\xEB\x88\xB7\x00\xEB\x88\xB8\x00\xEB\x88\xB9\x00\xEB\x88\xBA\x00"
-     "\xEB\x88"
-     "\xBB\x00\xEB\x88\xBC\x00\xEB\x88\xBD\x00\xEB\x88\xBE\x00\xEB\x88\xBF\x00"
-     "\xEB\x89"
-     "\x80\x00\xEB\x89\x81\x00\xEB\x89\x82\x00\xEB\x89\x83\x00\xEB\x89\x84\x00"
-     "\xEB\x89"
-     "\x85\x00\xEB\x89\x86\x00\xEB\x89\x87\x00\xEB\x89\x88\x00\xEB\x89\x89\x00"
-     "\xEB\x89"
-     "\x8A\x00\xEB\x89\x8B\x00\xEB\x89\x8C\x00\xEB\x89\x8D\x00\xEB\x89\x8E\x00"
-     "\xEB\x89"
-     "\x8F\x00\xEB\x89\x90\x00\xEB\x89\x91\x00\xEB\x89\x92\x00\xEB\x89\x93\x00"
-     "\xEB\x89"
-     "\x94\x00\xEB\x89\x95\x00\xEB\x89\x96\x00\xEB\x89\x97\x00\xEB\x89\x98\x00"
-     "\xEB\x89"
-     "\x99\x00\xEB\x89\x9A\x00\xEB\x89\x9B\x00\xEB\x89\x9C\x00\xEB\x89\x9D\x00"
-     "\xEB\x89"
-     "\x9E\x00\xEB\x89\x9F\x00\xEB\x89\xA0\x00\xEB\x89\xA1\x00\xEB\x89\xA2\x00"
-     "\xEB\x89"
-     "\xA3\x00\xEB\x89\xA4\x00\xEB\x89\xA5\x00\xEB\x89\xA6\x00\xEB\x89\xA7\x00"
-     "\xEB\x89"
-     "\xA8\x00\xEB\x89\xA9\x00\xEB\x89\xAA\x00\xEB\x89\xAB\x00\xEB\x89\xAC\x00"
-     "\xEB\x89"
-     "\xAD\x00\xEB\x89\xAE\x00\xEB\x89\xAF\x00\xEB\x89\xB0\x00\xEB\x89\xB1\x00"
-     "\xEB\x89"
-     "\xB2\x00\xEB\x89\xB3\x00\xEB\x89\xB4\x00\xEB\x89\xB5\x00\xEB\x89\xB6\x00"
-     "\xEB\x89"
-     "\xB7\x00\xEB\x89\xB8\x00\xEB\x89\xB9\x00\xEB\x89\xBA\x00\xEB\x89\xBB\x00"
-     "\xEB\x89"
-     "\xBC\x00\xEB\x89\xBD\x00\xEB\x89\xBE\x00\xEB\x89\xBF\x00\xEB\x8A\x80\x00"
-     "\xEB\x8A"
-     "\x81\x00\xEB\x8A\x82\x00\xEB\x8A\x83\x00\xEB\x8A\x84\x00\xEB\x8A\x85\x00"
-     "\xEB\x8A"
-     "\x86\x00\xEB\x8A\x87\x00\xEB\x8A\x88\x00\xEB\x8A\x89\x00\xEB\x8A\x8A\x00"
-     "\xEB\x8A"
-     "\x8B\x00\xEB\x8A\x8C\x00\xEB\x8A\x8D\x00\xEB\x8A\x8E\x00\xEB\x8A\x8F\x00"
-     "\xEB\x8A"
-     "\x90\x00\xEB\x8A\x91\x00\xEB\x8A\x92\x00\xEB\x8A\x93\x00\xEB\x8A\x94\x00"
-     "\xEB\x8A"
-     "\x95\x00\xEB\x8A\x96\x00\xEB\x8A\x97\x00\xEB\x8A\x98\x00\xEB\x8A\x99\x00"
-     "\xEB\x8A"
-     "\x9A\x00\xEB\x8A\x9B\x00\xEB\x8A\x9C\x00\xEB\x8A\x9D\x00\xEB\x8A\x9E\x00"
-     "\xEB\x8A"
-     "\x9F\x00\xEB\x8A\xA0\x00\xEB\x8A\xA1\x00\xEB\x8A\xA2\x00\xEB\x8A\xA3\x00"
-     "\xEB\x8A"
-     "\xA4\x00\xEB\x8A\xA5\x00\xEB\x8A\xA6\x00\xEB\x8A\xA7\x00\xEB\x8A\xA8\x00"
-     "\xEB\x8A"
-     "\xA9\x00\xEB\x8A\xAA\x00\xEB\x8A\xAB\x00\xEB\x8A\xAC\x00\xEB\x8A\xAD\x00"
-     "\xEB\x8A"
-     "\xAE\x00\xEB\x8A\xAF\x00\xEB\x8A\xB0\x00\xEB\x8A\xB1\x00\xEB\x8A\xB2\x00"
-     "\xEB\x8A"
-     "\xB3\x00\xEB\x8A\xB4\x00\xEB\x8A\xB5\x00\xEB\x8A\xB6\x00\xEB\x8A\xB7\x00"
-     "\xEB\x8A"
-     "\xB8\x00\xEB\x8A\xB9\x00\xEB\x8A\xBA\x00\xEB\x8A\xBB\x00\xEB\x8A\xBC\x00"
-     "\xEB\x8A"
-     "\xBD\x00\xEB\x8A\xBE\x00\xEB\x8A\xBF\x00\xEB\x8B\x80\x00\xEB\x8B\x81\x00"
-     "\xEB\x8B"
-     "\x82\x00\xEB\x8B\x83\x00\xEB\x8B\x84\x00\xEB\x8B\x85\x00\xEB\x8B\x86\x00"
-     "\xEB\x8B"
-     "\x87\x00\xEB\x8B\x88\x00\xEB\x8B\x89\x00\xEB\x8B\x8A\x00\xEB\x8B\x8B\x00"
-     "\xEB\x8B"
-     "\x8C\x00\xEB\x8B\x8D\x00\xEB\x8B\x8E\x00\xEB\x8B\x8F\x00\xEB\x8B\x90\x00"
-     "\xEB\x8B"
-     "\x91\x00\xEB\x8B\x92\x00\xEB\x8B\x93\x00\xEB\x8B\x94\x00\xEB\x8B\x95\x00"
-     "\xEB\x8B"
-     "\x96\x00\xEB\x8B\x97\x00\xEB\x8B\x98\x00\xEB\x8B\x99\x00\xEB\x8B\x9A\x00"
-     "\xEB\x8B"
-     "\x9B\x00\xEB\x8B\x9C\x00\xEB\x8B\x9D\x00\xEB\x8B\x9E\x00\xEB\x8B\x9F\x00"
-     "\xEB\x8B"
-     "\xA0\x00\xEB\x8B\xA1\x00\xEB\x8B\xA2\x00\xEB\x8B\xA3\x00\xEB\x8B\xA4\x00"
-     "\xEB\x8B"
-     "\xA5\x00\xEB\x8B\xA6\x00\xEB\x8B\xA7\x00\xEB\x8B\xA8\x00\xEB\x8B\xA9\x00"
-     "\xEB\x8B"
-     "\xAA\x00\xEB\x8B\xAB\x00\xEB\x8B\xAC\x00\xEB\x8B\xAD\x00\xEB\x8B\xAE\x00"
-     "\xEB\x8B"
-     "\xAF\x00\xEB\x8B\xB0\x00\xEB\x8B\xB1\x00\xEB\x8B\xB2\x00\xEB\x8B\xB3\x00"
-     "\xEB\x8B"
-     "\xB4\x00\xEB\x8B\xB5\x00\xEB\x8B\xB6\x00\xEB\x8B\xB7\x00\xEB\x8B\xB8\x00"
-     "\xEB\x8B"
-     "\xB9\x00\xEB\x8B\xBA\x00\xEB\x8B\xBB\x00\xEB\x8B\xBC\x00\xEB\x8B\xBD\x00"
-     "\xEB\x8B"
-     "\xBE\x00\xEB\x8B\xBF\x00\xEB\x8C\x80\x00\xEB\x8C\x81\x00\xEB\x8C\x82\x00"
-     "\xEB\x8C"
-     "\x83\x00\xEB\x8C\x84\x00\xEB\x8C\x85\x00\xEB\x8C\x86\x00\xEB\x8C\x87\x00"
-     "\xEB\x8C"
-     "\x88\x00\xEB\x8C\x89\x00\xEB\x8C\x8A\x00\xEB\x8C\x8B\x00\xEB\x8C\x8C\x00"
-     "\xEB\x8C"
-     "\x8D\x00\xEB\x8C\x8E\x00\xEB\x8C\x8F\x00\xEB\x8C\x90\x00\xEB\x8C\x91\x00"
-     "\xEB\x8C"
-     "\x92\x00\xEB\x8C\x93\x00\xEB\x8C\x94\x00\xEB\x8C\x95\x00\xEB\x8C\x96\x00"
-     "\xEB\x8C"
-     "\x97\x00\xEB\x8C\x98\x00\xEB\x8C\x99\x00\xEB\x8C\x9A\x00\xEB\x8C\x9B\x00"
-     "\xEB\x8C"
-     "\x9C\x00\xEB\x8C\x9D\x00\xEB\x8C\x9E\x00\xEB\x8C\x9F\x00\xEB\x8C\xA0\x00"
-     "\xEB\x8C"
-     "\xA1\x00\xEB\x8C\xA2\x00\xEB\x8C\xA3\x00\xEB\x8C\xA4\x00\xEB\x8C\xA5\x00"
-     "\xEB\x8C"
-     "\xA6\x00\xEB\x8C\xA7\x00\xEB\x8C\xA8\x00\xEB\x8C\xA9\x00\xEB\x8C\xAA\x00"
-     "\xEB\x8C"
-     "\xAB\x00\xEB\x8C\xAC\x00\xEB\x8C\xAD\x00\xEB\x8C\xAE\x00\xEB\x8C\xAF\x00"
-     "\xEB\x8C"
-     "\xB0\x00\xEB\x8C\xB1\x00\xEB\x8C\xB2\x00\xEB\x8C\xB3\x00\xEB\x8C\xB4\x00"
-     "\xEB\x8C"
-     "\xB5\x00\xEB\x8C\xB6\x00\xEB\x8C\xB7\x00\xEB\x8C\xB8\x00\xEB\x8C\xB9\x00"
-     "\xEB\x8C"
-     "\xBA\x00\xEB\x8C\xBB\x00\xEB\x8C\xBC\x00\xEB\x8C\xBD\x00\xEB\x8C\xBE\x00"
-     "\xEB\x8C"
-     "\xBF\x00\xEB\x8D\x80\x00\xEB\x8D\x81\x00\xEB\x8D\x82\x00\xEB\x8D\x83\x00"
-     "\xEB\x8D"
-     "\x84\x00\xEB\x8D\x85\x00\xEB\x8D\x86\x00\xEB\x8D\x87\x00\xEB\x8D\x88\x00"
-     "\xEB\x8D"
-     "\x89\x00\xEB\x8D\x8A\x00\xEB\x8D\x8B\x00\xEB\x8D\x8C\x00\xEB\x8D\x8D\x00"
-     "\xEB\x8D"
-     "\x8E\x00\xEB\x8D\x8F\x00\xEB\x8D\x90\x00\xEB\x8D\x91\x00\xEB\x8D\x92\x00"
-     "\xEB\x8D"
-     "\x93\x00\xEB\x8D\x94\x00\xEB\x8D\x95\x00\xEB\x8D\x96\x00\xEB\x8D\x97\x00"
-     "\xEB\x8D"
-     "\x98\x00\xEB\x8D\x99\x00\xEB\x8D\x9A\x00\xEB\x8D\x9B\x00\xEB\x8D\x9C\x00"
-     "\xEB\x8D"
-     "\x9D\x00\xEB\x8D\x9E\x00\xEB\x8D\x9F\x00\xEB\x8D\xA0\x00\xEB\x8D\xA1\x00"
-     "\xEB\x8D"
-     "\xA2\x00\xEB\x8D\xA3\x00\xEB\x8D\xA4\x00\xEB\x8D\xA5\x00\xEB\x8D\xA6\x00"
-     "\xEB\x8D"
-     "\xA7\x00\xEB\x8D\xA8\x00\xEB\x8D\xA9\x00\xEB\x8D\xAA\x00\xEB\x8D\xAB\x00"
-     "\xEB\x8D"
-     "\xAC\x00\xEB\x8D\xAD\x00\xEB\x8D\xAE\x00\xEB\x8D\xAF\x00\xEB\x8D\xB0\x00"
-     "\xEB\x8D"
-     "\xB1\x00\xEB\x8D\xB2\x00\xEB\x8D\xB3\x00\xEB\x8D\xB4\x00\xEB\x8D\xB5\x00"
-     "\xEB\x8D"
-     "\xB6\x00\xEB\x8D\xB7\x00\xEB\x8D\xB8\x00\xEB\x8D\xB9\x00\xEB\x8D\xBA\x00"
-     "\xEB\x8D"
-     "\xBB\x00\xEB\x8D\xBC\x00\xEB\x8D\xBD\x00\xEB\x8D\xBE\x00\xEB\x8D\xBF\x00"
-     "\xEB\x8E"
-     "\x80\x00\xEB\x8E\x81\x00\xEB\x8E\x82\x00\xEB\x8E\x83\x00\xEB\x8E\x84\x00"
-     "\xEB\x8E"
-     "\x85\x00\xEB\x8E\x86\x00\xEB\x8E\x87\x00\xEB\x8E\x88\x00\xEB\x8E\x89\x00"
-     "\xEB\x8E"
-     "\x8A\x00\xEB\x8E\x8B\x00\xEB\x8E\x8C\x00\xEB\x8E\x8D\x00\xEB\x8E\x8E\x00"
-     "\xEB\x8E"
-     "\x8F\x00\xEB\x8E\x90\x00\xEB\x8E\x91\x00\xEB\x8E\x92\x00\xEB\x8E\x93\x00"
-     "\xEB\x8E"
-     "\x94\x00\xEB\x8E\x95\x00\xEB\x8E\x96\x00\xEB\x8E\x97\x00\xEB\x8E\x98\x00"
-     "\xEB\x8E"
-     "\x99\x00\xEB\x8E\x9A\x00\xEB\x8E\x9B\x00\xEB\x8E\x9C\x00\xEB\x8E\x9D\x00"
-     "\xEB\x8E"
-     "\x9E\x00\xEB\x8E\x9F\x00\xEB\x8E\xA0\x00\xEB\x8E\xA1\x00\xEB\x8E\xA2\x00"
-     "\xEB\x8E"
-     "\xA3\x00\xEB\x8E\xA4\x00\xEB\x8E\xA5\x00\xEB\x8E\xA6\x00\xEB\x8E\xA7\x00"
-     "\xEB\x8E"
-     "\xA8\x00\xEB\x8E\xA9\x00\xEB\x8E\xAA\x00\xEB\x8E\xAB\x00\xEB\x8E\xAC\x00"
-     "\xEB\x8E"
-     "\xAD\x00\xEB\x8E\xAE\x00\xEB\x8E\xAF\x00\xEB\x8E\xB0\x00\xEB\x8E\xB1\x00"
-     "\xEB\x8E"
-     "\xB2\x00\xEB\x8E\xB3\x00\xEB\x8E\xB4\x00\xEB\x8E\xB5\x00\xEB\x8E\xB6\x00"
-     "\xEB\x8E"
-     "\xB7\x00\xEB\x8E\xB8\x00\xEB\x8E\xB9\x00\xEB\x8E\xBA\x00\xEB\x8E\xBB\x00"
-     "\xEB\x8E"
-     "\xBC\x00\xEB\x8E\xBD\x00\xEB\x8E\xBE\x00\xEB\x8E\xBF\x00\xEB\x8F\x80\x00"
-     "\xEB\x8F"
-     "\x81\x00\xEB\x8F\x82\x00\xEB\x8F\x83\x00\xEB\x8F\x84\x00\xEB\x8F\x85\x00"
-     "\xEB\x8F"
-     "\x86\x00\xEB\x8F\x87\x00\xEB\x8F\x88\x00\xEB\x8F\x89\x00\xEB\x8F\x8A\x00"
-     "\xEB\x8F"
-     "\x8B\x00\xEB\x8F\x8C\x00\xEB\x8F\x8D\x00\xEB\x8F\x8E\x00\xEB\x8F\x8F\x00"
-     "\xEB\x8F"
-     "\x90\x00\xEB\x8F\x91\x00\xEB\x8F\x92\x00\xEB\x8F\x93\x00\xEB\x8F\x94\x00"
-     "\xEB\x8F"
-     "\x95\x00\xEB\x8F\x96\x00\xEB\x8F\x97\x00\xEB\x8F\x98\x00\xEB\x8F\x99\x00"
-     "\xEB\x8F"
-     "\x9A\x00\xEB\x8F\x9B\x00\xEB\x8F\x9C\x00\xEB\x8F\x9D\x00\xEB\x8F\x9E\x00"
-     "\xEB\x8F"
-     "\x9F\x00\xEB\x8F\xA0\x00\xEB\x8F\xA1\x00\xEB\x8F\xA2\x00\xEB\x8F\xA3\x00"
-     "\xEB\x8F"
-     "\xA4\x00\xEB\x8F\xA5\x00\xEB\x8F\xA6\x00\xEB\x8F\xA7\x00\xEB\x8F\xA8\x00"
-     "\xEB\x8F"
-     "\xA9\x00\xEB\x8F\xAA\x00\xEB\x8F\xAB\x00\xEB\x8F\xAC\x00\xEB\x8F\xAD\x00"
-     "\xEB\x8F"
-     "\xAE\x00\xEB\x8F\xAF\x00\xEB\x8F\xB0\x00\xEB\x8F\xB1\x00\xEB\x8F\xB2\x00"
-     "\xEB\x8F"
-     "\xB3\x00\xEB\x8F\xB4\x00\xEB\x8F\xB5\x00\xEB\x8F\xB6\x00\xEB\x8F\xB7\x00"
-     "\xEB\x8F"
-     "\xB8\x00\xEB\x8F\xB9\x00\xEB\x8F\xBA\x00\xEB\x8F\xBB\x00\xEB\x8F\xBC\x00"
-     "\xEB\x8F"
-     "\xBD\x00\xEB\x8F\xBE\x00\xEB\x8F\xBF\x00\xEB\x90\x80\x00\xEB\x90\x81\x00"
-     "\xEB\x90"
-     "\x82\x00\xEB\x90\x83\x00\xEB\x90\x84\x00\xEB\x90\x85\x00\xEB\x90\x86\x00"
-     "\xEB\x90"
-     "\x87\x00\xEB\x90\x88\x00\xEB\x90\x89\x00\xEB\x90\x8A\x00\xEB\x90\x8B\x00"
-     "\xEB\x90"
-     "\x8C\x00\xEB\x90\x8D\x00\xEB\x90\x8E\x00\xEB\x90\x8F\x00\xEB\x90\x90\x00"
-     "\xEB\x90"
-     "\x91\x00\xEB\x90\x92\x00\xEB\x90\x93\x00\xEB\x90\x94\x00\xEB\x90\x95\x00"
-     "\xEB\x90"
-     "\x96\x00\xEB\x90\x97\x00\xEB\x90\x98\x00\xEB\x90\x99\x00\xEB\x90\x9A\x00"
-     "\xEB\x90"
-     "\x9B\x00\xEB\x90\x9C\x00\xEB\x90\x9D\x00\xEB\x90\x9E\x00\xEB\x90\x9F\x00"
-     "\xEB\x90"
-     "\xA0\x00\xEB\x90\xA1\x00\xEB\x90\xA2\x00\xEB\x90\xA3\x00\xEB\x90\xA4\x00"
-     "\xEB\x90"
-     "\xA5\x00\xEB\x90\xA6\x00\xEB\x90\xA7\x00\xEB\x90\xA8\x00\xEB\x90\xA9\x00"
-     "\xEB\x90"
-     "\xAA\x00\xEB\x90\xAB\x00\xEB\x90\xAC\x00\xEB\x90\xAD\x00\xEB\x90\xAE\x00"
-     "\xEB\x90"
-     "\xAF\x00\xEB\x90\xB0\x00\xEB\x90\xB1\x00\xEB\x90\xB2\x00\xEB\x90\xB3\x00"
-     "\xEB\x90"
-     "\xB4\x00\xEB\x90\xB5\x00\xEB\x90\xB6\x00\xEB\x90\xB7\x00\xEB\x90\xB8\x00"
-     "\xEB\x90"
-     "\xB9\x00\xEB\x90\xBA\x00\xEB\x90\xBB\x00\xEB\x90\xBC\x00\xEB\x90\xBD\x00"
-     "\xEB\x90"
-     "\xBE\x00\xEB\x90\xBF\x00\xEB\x91\x80\x00\xEB\x91\x81\x00\xEB\x91\x82\x00"
-     "\xEB\x91"
-     "\x83\x00\xEB\x91\x84\x00\xEB\x91\x85\x00\xEB\x91\x86\x00\xEB\x91\x87\x00"
-     "\xEB\x91"
-     "\x88\x00\xEB\x91\x89\x00\xEB\x91\x8A\x00\xEB\x91\x8B\x00\xEB\x91\x8C\x00"
-     "\xEB\x91"
-     "\x8D\x00\xEB\x91\x8E\x00\xEB\x91\x8F\x00\xEB\x91\x90\x00\xEB\x91\x91\x00"
-     "\xEB\x91"
-     "\x92\x00\xEB\x91\x93\x00\xEB\x91\x94\x00\xEB\x91\x95\x00\xEB\x91\x96\x00"
-     "\xEB\x91"
-     "\x97\x00\xEB\x91\x98\x00\xEB\x91\x99\x00\xEB\x91\x9A\x00\xEB\x91\x9B\x00"
-     "\xEB\x91"
-     "\x9C\x00\xEB\x91\x9D\x00\xEB\x91\x9E\x00\xEB\x91\x9F\x00\xEB\x91\xA0\x00"
-     "\xEB\x91"
-     "\xA1\x00\xEB\x91\xA2\x00\xEB\x91\xA3\x00\xEB\x91\xA4\x00\xEB\x91\xA5\x00"
-     "\xEB\x91"
-     "\xA6\x00\xEB\x91\xA7\x00\xEB\x91\xA8\x00\xEB\x91\xA9\x00\xEB\x91\xAA\x00"
-     "\xEB\x91"
-     "\xAB\x00\xEB\x91\xAC\x00\xEB\x91\xAD\x00\xEB\x91\xAE\x00\xEB\x91\xAF\x00"
-     "\xEB\x91"
-     "\xB0\x00\xEB\x91\xB1\x00\xEB\x91\xB2\x00\xEB\x91\xB3\x00\xEB\x91\xB4\x00"
-     "\xEB\x91"
-     "\xB5\x00\xEB\x91\xB6\x00\xEB\x91\xB7\x00\xEB\x91\xB8\x00\xEB\x91\xB9\x00"
-     "\xEB\x91"
-     "\xBA\x00\xEB\x91\xBB\x00\xEB\x91\xBC\x00\xEB\x91\xBD\x00\xEB\x91\xBE\x00"
-     "\xEB\x91"
-     "\xBF\x00\xEB\x92\x80\x00\xEB\x92\x81\x00\xEB\x92\x82\x00\xEB\x92\x83\x00"
-     "\xEB\x92"
-     "\x84\x00\xEB\x92\x85\x00\xEB\x92\x86\x00\xEB\x92\x87\x00\xEB\x92\x88\x00"
-     "\xEB\x92"
-     "\x89\x00\xEB\x92\x8A\x00\xEB\x92\x8B\x00\xEB\x92\x8C\x00\xEB\x92\x8D\x00"
-     "\xEB\x92"
-     "\x8E\x00\xEB\x92\x8F\x00\xEB\x92\x90\x00\xEB\x92\x91\x00\xEB\x92\x92\x00"
-     "\xEB\x92"
-     "\x93\x00\xEB\x92\x94\x00\xEB\x92\x95\x00\xEB\x92\x96\x00\xEB\x92\x97\x00"
-     "\xEB\x92"
-     "\x98\x00\xEB\x92\x99\x00\xEB\x92\x9A\x00\xEB\x92\x9B\x00\xEB\x92\x9C\x00"
-     "\xEB\x92"
-     "\x9D\x00\xEB\x92\x9E\x00\xEB\x92\x9F\x00\xEB\x92\xA0\x00\xEB\x92\xA1\x00"
-     "\xEB\x92"
-     "\xA2\x00\xEB\x92\xA3\x00\xEB\x92\xA4\x00\xEB\x92\xA5\x00\xEB\x92\xA6\x00"
-     "\xEB\x92"
-     "\xA7\x00\xEB\x92\xA8\x00\xEB\x92\xA9\x00\xEB\x92\xAA\x00\xEB\x92\xAB\x00"
-     "\xEB\x92"
-     "\xAC\x00\xEB\x92\xAD\x00\xEB\x92\xAE\x00\xEB\x92\xAF\x00\xEB\x92\xB0\x00"
-     "\xEB\x92"
-     "\xB1\x00\xEB\x92\xB2\x00\xEB\x92\xB3\x00\xEB\x92\xB4\x00\xEB\x92\xB5\x00"
-     "\xEB\x92"
-     "\xB6\x00\xEB\x92\xB7\x00\xEB\x92\xB8\x00\xEB\x92\xB9\x00\xEB\x92\xBA\x00"
-     "\xEB\x92"
-     "\xBB\x00\xEB\x92\xBC\x00\xEB\x92\xBD\x00\xEB\x92\xBE\x00\xEB\x92\xBF\x00"
-     "\xEB\x93"
-     "\x80\x00\xEB\x93\x81\x00\xEB\x93\x82\x00\xEB\x93\x83\x00\xEB\x93\x84\x00"
-     "\xEB\x93"
-     "\x85\x00\xEB\x93\x86\x00\xEB\x93\x87\x00\xEB\x93\x88\x00\xEB\x93\x89\x00"
-     "\xEB\x93"
-     "\x8A\x00\xEB\x93\x8B\x00\xEB\x93\x8C\x00\xEB\x93\x8D\x00\xEB\x93\x8E\x00"
-     "\xEB\x93"
-     "\x8F\x00\xEB\x93\x90\x00\xEB\x93\x91\x00\xEB\x93\x92\x00\xEB\x93\x93\x00"
-     "\xEB\x93"
-     "\x94\x00\xEB\x93\x95\x00\xEB\x93\x96\x00\xEB\x93\x97\x00\xEB\x93\x98\x00"
-     "\xEB\x93"
-     "\x99\x00\xEB\x93\x9A\x00\xEB\x93\x9B\x00\xEB\x93\x9C\x00\xEB\x93\x9D\x00"
-     "\xEB\x93"
-     "\x9E\x00\xEB\x93\x9F\x00\xEB\x93\xA0\x00\xEB\x93\xA1\x00\xEB\x93\xA2\x00"
-     "\xEB\x93"
-     "\xA3\x00\xEB\x93\xA4\x00\xEB\x93\xA5\x00\xEB\x93\xA6\x00\xEB\x93\xA7\x00"
-     "\xEB\x93"
-     "\xA8\x00\xEB\x93\xA9\x00\xEB\x93\xAA\x00\xEB\x93\xAB\x00\xEB\x93\xAC\x00"
-     "\xEB\x93"
-     "\xAD\x00\xEB\x93\xAE\x00\xEB\x93\xAF\x00\xEB\x93\xB0\x00\xEB\x93\xB1\x00"
-     "\xEB\x93"
-     "\xB2\x00\xEB\x93\xB3\x00\xEB\x93\xB4\x00\xEB\x93\xB5\x00\xEB\x93\xB6\x00"
-     "\xEB\x93"
-     "\xB7\x00\xEB\x93\xB8\x00\xEB\x93\xB9\x00\xEB\x93\xBA\x00\xEB\x93\xBB\x00"
-     "\xEB\x93"
-     "\xBC\x00\xEB\x93\xBD\x00\xEB\x93\xBE\x00\xEB\x93\xBF\x00\xEB\x94\x80\x00"
-     "\xEB\x94"
-     "\x81\x00\xEB\x94\x82\x00\xEB\x94\x83\x00\xEB\x94\x84\x00\xEB\x94\x85\x00"
-     "\xEB\x94"
-     "\x86\x00\xEB\x94\x87\x00\xEB\x94\x88\x00\xEB\x94\x89\x00\xEB\x94\x8A\x00"
-     "\xEB\x94"
-     "\x8B\x00\xEB\x94\x8C\x00\xEB\x94\x8D\x00\xEB\x94\x8E\x00\xEB\x94\x8F\x00"
-     "\xEB\x94"
-     "\x90\x00\xEB\x94\x91\x00\xEB\x94\x92\x00\xEB\x94\x93\x00\xEB\x94\x94\x00"
-     "\xEB\x94"
-     "\x95\x00\xEB\x94\x96\x00\xEB\x94\x97\x00\xEB\x94\x98\x00\xEB\x94\x99\x00"
-     "\xEB\x94"
-     "\x9A\x00\xEB\x94\x9B\x00\xEB\x94\x9C\x00\xEB\x94\x9D\x00\xEB\x94\x9E\x00"
-     "\xEB\x94"
-     "\x9F\x00\xEB\x94\xA0\x00\xEB\x94\xA1\x00\xEB\x94\xA2\x00\xEB\x94\xA3\x00"
-     "\xEB\x94"
-     "\xA4\x00\xEB\x94\xA5\x00\xEB\x94\xA6\x00\xEB\x94\xA7\x00\xEB\x94\xA8\x00"
-     "\xEB\x94"
-     "\xA9\x00\xEB\x94\xAA\x00\xEB\x94\xAB\x00\xEB\x94\xAC\x00\xEB\x94\xAD\x00"
-     "\xEB\x94"
-     "\xAE\x00\xEB\x94\xAF\x00\xEB\x94\xB0\x00\xEB\x94\xB1\x00\xEB\x94\xB2\x00"
-     "\xEB\x94"
-     "\xB3\x00\xEB\x94\xB4\x00\xEB\x94\xB5\x00\xEB\x94\xB6\x00\xEB\x94\xB7\x00"
-     "\xEB\x94"
-     "\xB8\x00\xEB\x94\xB9\x00\xEB\x94\xBA\x00\xEB\x94\xBB\x00\xEB\x94\xBC\x00"
-     "\xEB\x94"
-     "\xBD\x00\xEB\x94\xBE\x00\xEB\x94\xBF\x00\xEB\x95\x80\x00\xEB\x95\x81\x00"
-     "\xEB\x95"
-     "\x82\x00\xEB\x95\x83\x00\xEB\x95\x84\x00\xEB\x95\x85\x00\xEB\x95\x86\x00"
-     "\xEB\x95"
-     "\x87\x00\xEB\x95\x88\x00\xEB\x95\x89\x00\xEB\x95\x8A\x00\xEB\x95\x8B\x00"
-     "\xEB\x95"
-     "\x8C\x00\xEB\x95\x8D\x00\xEB\x95\x8E\x00\xEB\x95\x8F\x00\xEB\x95\x90\x00"
-     "\xEB\x95"
-     "\x91\x00\xEB\x95\x92\x00\xEB\x95\x93\x00\xEB\x95\x94\x00\xEB\x95\x95\x00"
-     "\xEB\x95"
-     "\x96\x00\xEB\x95\x97\x00\xEB\x95\x98\x00\xEB\x95\x99\x00\xEB\x95\x9A\x00"
-     "\xEB\x95"
-     "\x9B\x00\xEB\x95\x9C\x00\xEB\x95\x9D\x00\xEB\x95\x9E\x00\xEB\x95\x9F\x00"
-     "\xEB\x95"
-     "\xA0\x00\xEB\x95\xA1\x00\xEB\x95\xA2\x00\xEB\x95\xA3\x00\xEB\x95\xA4\x00"
-     "\xEB\x95"
-     "\xA5\x00\xEB\x95\xA6\x00\xEB\x95\xA7\x00\xEB\x95\xA8\x00\xEB\x95\xA9\x00"
-     "\xEB\x95"
-     "\xAA\x00\xEB\x95\xAB\x00\xEB\x95\xAC\x00\xEB\x95\xAD\x00\xEB\x95\xAE\x00"
-     "\xEB\x95"
-     "\xAF\x00\xEB\x95\xB0\x00\xEB\x95\xB1\x00\xEB\x95\xB2\x00\xEB\x95\xB3\x00"
-     "\xEB\x95"
-     "\xB4\x00\xEB\x95\xB5\x00\xEB\x95\xB6\x00\xEB\x95\xB7\x00\xEB\x95\xB8\x00"
-     "\xEB\x95"
-     "\xB9\x00\xEB\x95\xBA\x00\xEB\x95\xBB\x00\xEB\x95\xBC\x00\xEB\x95\xBD\x00"
-     "\xEB\x95"
-     "\xBE\x00\xEB\x95\xBF\x00\xEB\x96\x80\x00\xEB\x96\x81\x00\xEB\x96\x82\x00"
-     "\xEB\x96"
-     "\x83\x00\xEB\x96\x84\x00\xEB\x96\x85\x00\xEB\x96\x86\x00\xEB\x96\x87\x00"
-     "\xEB\x96"
-     "\x88\x00\xEB\x96\x89\x00\xEB\x96\x8A\x00\xEB\x96\x8B\x00\xEB\x96\x8C\x00"
-     "\xEB\x96"
-     "\x8D\x00\xEB\x96\x8E\x00\xEB\x96\x8F\x00\xEB\x96\x90\x00\xEB\x96\x91\x00"
-     "\xEB\x96"
-     "\x92\x00\xEB\x96\x93\x00\xEB\x96\x94\x00\xEB\x96\x95\x00\xEB\x96\x96\x00"
-     "\xEB\x96"
-     "\x97\x00\xEB\x96\x98\x00\xEB\x96\x99\x00\xEB\x96\x9A\x00\xEB\x96\x9B\x00"
-     "\xEB\x96"
-     "\x9C\x00\xEB\x96\x9D\x00\xEB\x96\x9E\x00\xEB\x96\x9F\x00\xEB\x96\xA0\x00"
-     "\xEB\x96"
-     "\xA1\x00\xEB\x96\xA2\x00\xEB\x96\xA3\x00\xEB\x96\xA4\x00\xEB\x96\xA5\x00"
-     "\xEB\x96"
-     "\xA6\x00\xEB\x96\xA7\x00\xEB\x96\xA8\x00\xEB\x96\xA9\x00\xEB\x96\xAA\x00"
-     "\xEB\x96"
-     "\xAB\x00\xEB\x96\xAC\x00\xEB\x96\xAD\x00\xEB\x96\xAE\x00\xEB\x96\xAF\x00"
-     "\xEB\x96"
-     "\xB0\x00\xEB\x96\xB1\x00\xEB\x96\xB2\x00\xEB\x96\xB3\x00\xEB\x96\xB4\x00"
-     "\xEB\x96"
-     "\xB5\x00\xEB\x96\xB6\x00\xEB\x96\xB7\x00\xEB\x96\xB8\x00\xEB\x96\xB9\x00"
-     "\xEB\x96"
-     "\xBA\x00\xEB\x96\xBB\x00\xEB\x96\xBC\x00\xEB\x96\xBD\x00\xEB\x96\xBE\x00"
-     "\xEB\x96"
-     "\xBF\x00\xEB\x97\x80\x00\xEB\x97\x81\x00\xEB\x97\x82\x00\xEB\x97\x83\x00"
-     "\xEB\x97"
-     "\x84\x00\xEB\x97\x85\x00\xEB\x97\x86\x00\xEB\x97\x87\x00\xEB\x97\x88\x00"
-     "\xEB\x97"
-     "\x89\x00\xEB\x97\x8A\x00\xEB\x97\x8B\x00\xEB\x97\x8C\x00\xEB\x97\x8D\x00"
-     "\xEB\x97"
-     "\x8E\x00\xEB\x97\x8F\x00\xEB\x97\x90\x00\xEB\x97\x91\x00\xEB\x97\x92\x00"
-     "\xEB\x97"
-     "\x93\x00\xEB\x97\x94\x00\xEB\x97\x95\x00\xEB\x97\x96\x00\xEB\x97\x97\x00"
-     "\xEB\x97"
-     "\x98\x00\xEB\x97\x99\x00\xEB\x97\x9A\x00\xEB\x97\x9B\x00\xEB\x97\x9C\x00"
-     "\xEB\x97"
-     "\x9D\x00\xEB\x97\x9E\x00\xEB\x97\x9F\x00\xEB\x97\xA0\x00\xEB\x97\xA1\x00"
-     "\xEB\x97"
-     "\xA2\x00\xEB\x97\xA3\x00\xEB\x97\xA4\x00\xEB\x97\xA5\x00\xEB\x97\xA6\x00"
-     "\xEB\x97"
-     "\xA7\x00\xEB\x97\xA8\x00\xEB\x97\xA9\x00\xEB\x97\xAA\x00\xEB\x97\xAB\x00"
-     "\xEB\x97"
-     "\xAC\x00\xEB\x97\xAD\x00\xEB\x97\xAE\x00\xEB\x97\xAF\x00\xEB\x97\xB0\x00"
-     "\xEB\x97"
-     "\xB1\x00\xEB\x97\xB2\x00\xEB\x97\xB3\x00\xEB\x97\xB4\x00\xEB\x97\xB5\x00"
-     "\xEB\x97"
-     "\xB6\x00\xEB\x97\xB7\x00\xEB\x97\xB8\x00\xEB\x97\xB9\x00\xEB\x97\xBA\x00"
-     "\xEB\x97"
-     "\xBB\x00\xEB\x97\xBC\x00\xEB\x97\xBD\x00\xEB\x97\xBE\x00\xEB\x97\xBF\x00"
-     "\xEB\x98"
-     "\x80\x00\xEB\x98\x81\x00\xEB\x98\x82\x00\xEB\x98\x83\x00\xEB\x98\x84\x00"
-     "\xEB\x98"
-     "\x85\x00\xEB\x98\x86\x00\xEB\x98\x87\x00\xEB\x98\x88\x00\xEB\x98\x89\x00"
-     "\xEB\x98"
-     "\x8A\x00\xEB\x98\x8B\x00\xEB\x98\x8C\x00\xEB\x98\x8D\x00\xEB\x98\x8E\x00"
-     "\xEB\x98"
-     "\x8F\x00\xEB\x98\x90\x00\xEB\x98\x91\x00\xEB\x98\x92\x00\xEB\x98\x93\x00"
-     "\xEB\x98"
-     "\x94\x00\xEB\x98\x95\x00\xEB\x98\x96\x00\xEB\x98\x97\x00\xEB\x98\x98\x00"
-     "\xEB\x98"
-     "\x99\x00\xEB\x98\x9A\x00\xEB\x98\x9B\x00\xEB\x98\x9C\x00\xEB\x98\x9D\x00"
-     "\xEB\x98"
-     "\x9E\x00\xEB\x98\x9F\x00\xEB\x98\xA0\x00\xEB\x98\xA1\x00\xEB\x98\xA2\x00"
-     "\xEB\x98"
-     "\xA3\x00\xEB\x98\xA4\x00\xEB\x98\xA5\x00\xEB\x98\xA6\x00\xEB\x98\xA7\x00"
-     "\xEB\x98"
-     "\xA8\x00\xEB\x98\xA9\x00\xEB\x98\xAA\x00\xEB\x98\xAB\x00\xEB\x98\xAC\x00"
-     "\xEB\x98"
-     "\xAD\x00\xEB\x98\xAE\x00\xEB\x98\xAF\x00\xEB\x98\xB0\x00\xEB\x98\xB1\x00"
-     "\xEB\x98"
-     "\xB2\x00\xEB\x98\xB3\x00\xEB\x98\xB4\x00\xEB\x98\xB5\x00\xEB\x98\xB6\x00"
-     "\xEB\x98"
-     "\xB7\x00\xEB\x98\xB8\x00\xEB\x98\xB9\x00\xEB\x98\xBA\x00\xEB\x98\xBB\x00"
-     "\xEB\x98"
-     "\xBC\x00\xEB\x98\xBD\x00\xEB\x98\xBE\x00\xEB\x98\xBF\x00\xEB\x99\x80\x00"
-     "\xEB\x99"
-     "\x81\x00\xEB\x99\x82\x00\xEB\x99\x83\x00\xEB\x99\x84\x00\xEB\x99\x85\x00"
-     "\xEB\x99"
-     "\x86\x00\xEB\x99\x87\x00\xEB\x99\x88\x00\xEB\x99\x89\x00\xEB\x99\x8A\x00"
-     "\xEB\x99"
-     "\x8B\x00\xEB\x99\x8C\x00\xEB\x99\x8D\x00\xEB\x99\x8E\x00\xEB\x99\x8F\x00"
-     "\xEB\x99"
-     "\x90\x00\xEB\x99\x91\x00\xEB\x99\x92\x00\xEB\x99\x93\x00\xEB\x99\x94\x00"
-     "\xEB\x99"
-     "\x95\x00\xEB\x99\x96\x00\xEB\x99\x97\x00\xEB\x99\x98\x00\xEB\x99\x99\x00"
-     "\xEB\x99"
-     "\x9A\x00\xEB\x99\x9B\x00\xEB\x99\x9C\x00\xEB\x99\x9D\x00\xEB\x99\x9E\x00"
-     "\xEB\x99"
-     "\x9F\x00\xEB\x99\xA0\x00\xEB\x99\xA1\x00\xEB\x99\xA2\x00\xEB\x99\xA3\x00"
-     "\xEB\x99"
-     "\xA4\x00\xEB\x99\xA5\x00\xEB\x99\xA6\x00\xEB\x99\xA7\x00\xEB\x99\xA8\x00"
-     "\xEB\x99"
-     "\xA9\x00\xEB\x99\xAA\x00\xEB\x99\xAB\x00\xEB\x99\xAC\x00\xEB\x99\xAD\x00"
-     "\xEB\x99"
-     "\xAE\x00\xEB\x99\xAF\x00\xEB\x99\xB0\x00\xEB\x99\xB1\x00\xEB\x99\xB2\x00"
-     "\xEB\x99"
-     "\xB3\x00\xEB\x99\xB4\x00\xEB\x99\xB5\x00\xEB\x99\xB6\x00\xEB\x99\xB7\x00"
-     "\xEB\x99"
-     "\xB8\x00\xEB\x99\xB9\x00\xEB\x99\xBA\x00\xEB\x99\xBB\x00\xEB\x99\xBC\x00"
-     "\xEB\x99"
-     "\xBD\x00\xEB\x99\xBE\x00\xEB\x99\xBF\x00\xEB\x9A\x80\x00\xEB\x9A\x81\x00"
-     "\xEB\x9A"
-     "\x82\x00\xEB\x9A\x83\x00\xEB\x9A\x84\x00\xEB\x9A\x85\x00\xEB\x9A\x86\x00"
-     "\xEB\x9A"
-     "\x87\x00\xEB\x9A\x88\x00\xEB\x9A\x89\x00\xEB\x9A\x8A\x00\xEB\x9A\x8B\x00"
-     "\xEB\x9A"
-     "\x8C\x00\xEB\x9A\x8D\x00\xEB\x9A\x8E\x00\xEB\x9A\x8F\x00\xEB\x9A\x90\x00"
-     "\xEB\x9A"
-     "\x91\x00\xEB\x9A\x92\x00\xEB\x9A\x93\x00\xEB\x9A\x94\x00\xEB\x9A\x95\x00"
-     "\xEB\x9A"
-     "\x96\x00\xEB\x9A\x97\x00\xEB\x9A\x98\x00\xEB\x9A\x99\x00\xEB\x9A\x9A\x00"
-     "\xEB\x9A"
-     "\x9B\x00\xEB\x9A\x9C\x00\xEB\x9A\x9D\x00\xEB\x9A\x9E\x00\xEB\x9A\x9F\x00"
-     "\xEB\x9A"
-     "\xA0\x00\xEB\x9A\xA1\x00\xEB\x9A\xA2\x00\xEB\x9A\xA3\x00\xEB\x9A\xA4\x00"
-     "\xEB\x9A"
-     "\xA5\x00\xEB\x9A\xA6\x00\xEB\x9A\xA7\x00\xEB\x9A\xA8\x00\xEB\x9A\xA9\x00"
-     "\xEB\x9A"
-     "\xAA\x00\xEB\x9A\xAB\x00\xEB\x9A\xAC\x00\xEB\x9A\xAD\x00\xEB\x9A\xAE\x00"
-     "\xEB\x9A"
-     "\xAF\x00\xEB\x9A\xB0\x00\xEB\x9A\xB1\x00\xEB\x9A\xB2\x00\xEB\x9A\xB3\x00"
-     "\xEB\x9A"
-     "\xB4\x00\xEB\x9A\xB5\x00\xEB\x9A\xB6\x00\xEB\x9A\xB7\x00\xEB\x9A\xB8\x00"
-     "\xEB\x9A"
-     "\xB9\x00\xEB\x9A\xBA\x00\xEB\x9A\xBB\x00\xEB\x9A\xBC\x00\xEB\x9A\xBD\x00"
-     "\xEB\x9A"
-     "\xBE\x00\xEB\x9A\xBF\x00\xEB\x9B\x80\x00\xEB\x9B\x81\x00\xEB\x9B\x82\x00"
-     "\xEB\x9B"
-     "\x83\x00\xEB\x9B\x84\x00\xEB\x9B\x85\x00\xEB\x9B\x86\x00\xEB\x9B\x87\x00"
-     "\xEB\x9B"
-     "\x88\x00\xEB\x9B\x89\x00\xEB\x9B\x8A\x00\xEB\x9B\x8B\x00\xEB\x9B\x8C\x00"
-     "\xEB\x9B"
-     "\x8D\x00\xEB\x9B\x8E\x00\xEB\x9B\x8F\x00\xEB\x9B\x90\x00\xEB\x9B\x91\x00"
-     "\xEB\x9B"
-     "\x92\x00\xEB\x9B\x93\x00\xEB\x9B\x94\x00\xEB\x9B\x95\x00\xEB\x9B\x96\x00"
-     "\xEB\x9B"
-     "\x97\x00\xEB\x9B\x98\x00\xEB\x9B\x99\x00\xEB\x9B\x9A\x00\xEB\x9B\x9B\x00"
-     "\xEB\x9B"
-     "\x9C\x00\xEB\x9B\x9D\x00\xEB\x9B\x9E\x00\xEB\x9B\x9F\x00\xEB\x9B\xA0\x00"
-     "\xEB\x9B"
-     "\xA1\x00\xEB\x9B\xA2\x00\xEB\x9B\xA3\x00\xEB\x9B\xA4\x00\xEB\x9B\xA5\x00"
-     "\xEB\x9B"
-     "\xA6\x00\xEB\x9B\xA7\x00\xEB\x9B\xA8\x00\xEB\x9B\xA9\x00\xEB\x9B\xAA\x00"
-     "\xEB\x9B"
-     "\xAB\x00\xEB\x9B\xAC\x00\xEB\x9B\xAD\x00\xEB\x9B\xAE\x00\xEB\x9B\xAF\x00"
-     "\xEB\x9B"
-     "\xB0\x00\xEB\x9B\xB1\x00\xEB\x9B\xB2\x00\xEB\x9B\xB3\x00\xEB\x9B\xB4\x00"
-     "\xEB\x9B"
-     "\xB5\x00\xEB\x9B\xB6\x00\xEB\x9B\xB7\x00\xEB\x9B\xB8\x00\xEB\x9B\xB9\x00"
-     "\xEB\x9B"
-     "\xBA\x00\xEB\x9B\xBB\x00\xEB\x9B\xBC\x00\xEB\x9B\xBD\x00\xEB\x9B\xBE\x00"
-     "\xEB\x9B"
-     "\xBF\x00\xEB\x9C\x80\x00\xEB\x9C\x81\x00\xEB\x9C\x82\x00\xEB\x9C\x83\x00"
-     "\xEB\x9C"
-     "\x84\x00\xEB\x9C\x85\x00\xEB\x9C\x86\x00\xEB\x9C\x87\x00\xEB\x9C\x88\x00"
-     "\xEB\x9C"
-     "\x89\x00\xEB\x9C\x8A\x00\xEB\x9C\x8B\x00\xEB\x9C\x8C\x00\xEB\x9C\x8D\x00"
-     "\xEB\x9C"
-     "\x8E\x00\xEB\x9C\x8F\x00\xEB\x9C\x90\x00\xEB\x9C\x91\x00\xEB\x9C\x92\x00"
-     "\xEB\x9C"
-     "\x93\x00\xEB\x9C\x94\x00\xEB\x9C\x95\x00\xEB\x9C\x96\x00\xEB\x9C\x97\x00"
-     "\xEB\x9C"
-     "\x98\x00\xEB\x9C\x99\x00\xEB\x9C\x9A\x00\xEB\x9C\x9B\x00\xEB\x9C\x9C\x00"
-     "\xEB\x9C"
-     "\x9D\x00\xEB\x9C\x9E\x00\xEB\x9C\x9F\x00\xEB\x9C\xA0\x00\xEB\x9C\xA1\x00"
-     "\xEB\x9C"
-     "\xA2\x00\xEB\x9C\xA3\x00\xEB\x9C\xA4\x00\xEB\x9C\xA5\x00\xEB\x9C\xA6\x00"
-     "\xEB\x9C"
-     "\xA7\x00\xEB\x9C\xA8\x00\xEB\x9C\xA9\x00\xEB\x9C\xAA\x00\xEB\x9C\xAB\x00"
-     "\xEB\x9C"
-     "\xAC\x00\xEB\x9C\xAD\x00\xEB\x9C\xAE\x00\xEB\x9C\xAF\x00\xEB\x9C\xB0\x00"
-     "\xEB\x9C"
-     "\xB1\x00\xEB\x9C\xB2\x00\xEB\x9C\xB3\x00\xEB\x9C\xB4\x00\xEB\x9C\xB5\x00"
-     "\xEB\x9C"
-     "\xB6\x00\xEB\x9C\xB7\x00\xEB\x9C\xB8\x00\xEB\x9C\xB9\x00\xEB\x9C\xBA\x00"
-     "\xEB\x9C"
-     "\xBB\x00\xEB\x9C\xBC\x00\xEB\x9C\xBD\x00\xEB\x9C\xBE\x00\xEB\x9C\xBF\x00"
-     "\xEB\x9D"
-     "\x80\x00\xEB\x9D\x81\x00\xEB\x9D\x82\x00\xEB\x9D\x83\x00\xEB\x9D\x84\x00"
-     "\xEB\x9D"
-     "\x85\x00\xEB\x9D\x86\x00\xEB\x9D\x87\x00\xEB\x9D\x88\x00\xEB\x9D\x89\x00"
-     "\xEB\x9D"
-     "\x8A\x00\xEB\x9D\x8B\x00\xEB\x9D\x8C\x00\xEB\x9D\x8D\x00\xEB\x9D\x8E\x00"
-     "\xEB\x9D"
-     "\x8F\x00\xEB\x9D\x90\x00\xEB\x9D\x91\x00\xEB\x9D\x92\x00\xEB\x9D\x93\x00"
-     "\xEB\x9D"
-     "\x94\x00\xEB\x9D\x95\x00\xEB\x9D\x96\x00\xEB\x9D\x97\x00\xEB\x9D\x98\x00"
-     "\xEB\x9D"
-     "\x99\x00\xEB\x9D\x9A\x00\xEB\x9D\x9B\x00\xEB\x9D\x9C\x00\xEB\x9D\x9D\x00"
-     "\xEB\x9D"
-     "\x9E\x00\xEB\x9D\x9F\x00\xEB\x9D\xA0\x00\xEB\x9D\xA1\x00\xEB\x9D\xA2\x00"
-     "\xEB\x9D"
-     "\xA3\x00\xEB\x9D\xA4\x00\xEB\x9D\xA5\x00\xEB\x9D\xA6\x00\xEB\x9D\xA7\x00"
-     "\xEB\x9D"
-     "\xA8\x00\xEB\x9D\xA9\x00\xEB\x9D\xAA\x00\xEB\x9D\xAB\x00\xEB\x9D\xAC\x00"
-     "\xEB\x9D"
-     "\xAD\x00\xEB\x9D\xAE\x00\xEB\x9D\xAF\x00\xEB\x9D\xB0\x00\xEB\x9D\xB1\x00"
-     "\xEB\x9D"
-     "\xB2\x00\xEB\x9D\xB3\x00\xEB\x9D\xB4\x00\xEB\x9D\xB5\x00\xEB\x9D\xB6\x00"
-     "\xEB\x9D"
-     "\xB7\x00\xEB\x9D\xB8\x00\xEB\x9D\xB9\x00\xEB\x9D\xBA\x00\xEB\x9D\xBB\x00"
-     "\xEB\x9D"
-     "\xBC\x00\xEB\x9D\xBD\x00\xEB\x9D\xBE\x00\xEB\x9D\xBF\x00\xEB\x9E\x80\x00"
-     "\xEB\x9E"
-     "\x81\x00\xEB\x9E\x82\x00\xEB\x9E\x83\x00\xEB\x9E\x84\x00\xEB\x9E\x85\x00"
-     "\xEB\x9E"
-     "\x86\x00\xEB\x9E\x87\x00\xEB\x9E\x88\x00\xEB\x9E\x89\x00\xEB\x9E\x8A\x00"
-     "\xEB\x9E"
-     "\x8B\x00\xEB\x9E\x8C\x00\xEB\x9E\x8D\x00\xEB\x9E\x8E\x00\xEB\x9E\x8F\x00"
-     "\xEB\x9E"
-     "\x90\x00\xEB\x9E\x91\x00\xEB\x9E\x92\x00\xEB\x9E\x93\x00\xEB\x9E\x94\x00"
-     "\xEB\x9E"
-     "\x95\x00\xEB\x9E\x96\x00\xEB\x9E\x97\x00\xEB\x9E\x98\x00\xEB\x9E\x99\x00"
-     "\xEB\x9E"
-     "\x9A\x00\xEB\x9E\x9B\x00\xEB\x9E\x9C\x00\xEB\x9E\x9D\x00\xEB\x9E\x9E\x00"
-     "\xEB\x9E"
-     "\x9F\x00\xEB\x9E\xA0\x00\xEB\x9E\xA1\x00\xEB\x9E\xA2\x00\xEB\x9E\xA3\x00"
-     "\xEB\x9E"
-     "\xA4\x00\xEB\x9E\xA5\x00\xEB\x9E\xA6\x00\xEB\x9E\xA7\x00\xEB\x9E\xA8\x00"
-     "\xEB\x9E"
-     "\xA9\x00\xEB\x9E\xAA\x00\xEB\x9E\xAB\x00\xEB\x9E\xAC\x00\xEB\x9E\xAD\x00"
-     "\xEB\x9E"
-     "\xAE\x00\xEB\x9E\xAF\x00\xEB\x9E\xB0\x00\xEB\x9E\xB1\x00\xEB\x9E\xB2\x00"
-     "\xEB\x9E"
-     "\xB3\x00\xEB\x9E\xB4\x00\xEB\x9E\xB5\x00\xEB\x9E\xB6\x00\xEB\x9E\xB7\x00"
-     "\xEB\x9E"
-     "\xB8\x00\xEB\x9E\xB9\x00\xEB\x9E\xBA\x00\xEB\x9E\xBB\x00\xEB\x9E\xBC\x00"
-     "\xEB\x9E"
-     "\xBD\x00\xEB\x9E\xBE\x00\xEB\x9E\xBF\x00\xEB\x9F\x80\x00\xEB\x9F\x81\x00"
-     "\xEB\x9F"
-     "\x82\x00\xEB\x9F\x83\x00\xEB\x9F\x84\x00\xEB\x9F\x85\x00\xEB\x9F\x86\x00"
-     "\xEB\x9F"
-     "\x87\x00\xEB\x9F\x88\x00\xEB\x9F\x89\x00\xEB\x9F\x8A\x00\xEB\x9F\x8B\x00"
-     "\xEB\x9F"
-     "\x8C\x00\xEB\x9F\x8D\x00\xEB\x9F\x8E\x00\xEB\x9F\x8F\x00\xEB\x9F\x90\x00"
-     "\xEB\x9F"
-     "\x91\x00\xEB\x9F\x92\x00\xEB\x9F\x93\x00\xEB\x9F\x94\x00\xEB\x9F\x95\x00"
-     "\xEB\x9F"
-     "\x96\x00\xEB\x9F\x97\x00\xEB\x9F\x98\x00\xEB\x9F\x99\x00\xEB\x9F\x9A\x00"
-     "\xEB\x9F"
-     "\x9B\x00\xEB\x9F\x9C\x00\xEB\x9F\x9D\x00\xEB\x9F\x9E\x00\xEB\x9F\x9F\x00"
-     "\xEB\x9F"
-     "\xA0\x00\xEB\x9F\xA1\x00\xEB\x9F\xA2\x00\xEB\x9F\xA3\x00\xEB\x9F\xA4\x00"
-     "\xEB\x9F"
-     "\xA5\x00\xEB\x9F\xA6\x00\xEB\x9F\xA7\x00\xEB\x9F\xA8\x00\xEB\x9F\xA9\x00"
-     "\xEB\x9F"
-     "\xAA\x00\xEB\x9F\xAB\x00\xEB\x9F\xAC\x00\xEB\x9F\xAD\x00\xEB\x9F\xAE\x00"
-     "\xEB\x9F"
-     "\xAF\x00\xEB\x9F\xB0\x00\xEB\x9F\xB1\x00\xEB\x9F\xB2\x00\xEB\x9F\xB3\x00"
-     "\xEB\x9F"
-     "\xB4\x00\xEB\x9F\xB5\x00\xEB\x9F\xB6\x00\xEB\x9F\xB7\x00\xEB\x9F\xB8\x00"
-     "\xEB\x9F"
-     "\xB9\x00\xEB\x9F\xBA\x00\xEB\x9F\xBB\x00\xEB\x9F\xBC\x00\xEB\x9F\xBD\x00"
-     "\xEB\x9F"
-     "\xBE\x00\xEB\x9F\xBF\x00\xEB\xA0\x80\x00\xEB\xA0\x81\x00\xEB\xA0\x82\x00"
-     "\xEB\xA0"
-     "\x83\x00\xEB\xA0\x84\x00\xEB\xA0\x85\x00\xEB\xA0\x86\x00\xEB\xA0\x87\x00"
-     "\xEB\xA0"
-     "\x88\x00\xEB\xA0\x89\x00\xEB\xA0\x8A\x00\xEB\xA0\x8B\x00\xEB\xA0\x8C\x00"
-     "\xEB\xA0"
-     "\x8D\x00\xEB\xA0\x8E\x00\xEB\xA0\x8F\x00\xEB\xA0\x90\x00\xEB\xA0\x91\x00"
-     "\xEB\xA0"
-     "\x92\x00\xEB\xA0\x93\x00\xEB\xA0\x94\x00\xEB\xA0\x95\x00\xEB\xA0\x96\x00"
-     "\xEB\xA0"
-     "\x97\x00\xEB\xA0\x98\x00\xEB\xA0\x99\x00\xEB\xA0\x9A\x00\xEB\xA0\x9B\x00"
-     "\xEB\xA0"
-     "\x9C\x00\xEB\xA0\x9D\x00\xEB\xA0\x9E\x00\xEB\xA0\x9F\x00\xEB\xA0\xA0\x00"
-     "\xEB\xA0"
-     "\xA1\x00\xEB\xA0\xA2\x00\xEB\xA0\xA3\x00\xEB\xA0\xA4\x00\xEB\xA0\xA5\x00"
-     "\xEB\xA0"
-     "\xA6\x00\xEB\xA0\xA7\x00\xEB\xA0\xA8\x00\xEB\xA0\xA9\x00\xEB\xA0\xAA\x00"
-     "\xEB\xA0"
-     "\xAB\x00\xEB\xA0\xAC\x00\xEB\xA0\xAD\x00\xEB\xA0\xAE\x00\xEB\xA0\xAF\x00"
-     "\xEB\xA0"
-     "\xB0\x00\xEB\xA0\xB1\x00\xEB\xA0\xB2\x00\xEB\xA0\xB3\x00\xEB\xA0\xB4\x00"
-     "\xEB\xA0"
-     "\xB5\x00\xEB\xA0\xB6\x00\xEB\xA0\xB7\x00\xEB\xA0\xB8\x00\xEB\xA0\xB9\x00"
-     "\xEB\xA0"
-     "\xBA\x00\xEB\xA0\xBB\x00\xEB\xA0\xBC\x00\xEB\xA0\xBD\x00\xEB\xA0\xBE\x00"
-     "\xEB\xA0"
-     "\xBF\x00\xEB\xA1\x80\x00\xEB\xA1\x81\x00\xEB\xA1\x82\x00\xEB\xA1\x83\x00"
-     "\xEB\xA1"
-     "\x84\x00\xEB\xA1\x85\x00\xEB\xA1\x86\x00\xEB\xA1\x87\x00\xEB\xA1\x88\x00"
-     "\xEB\xA1"
-     "\x89\x00\xEB\xA1\x8A\x00\xEB\xA1\x8B\x00\xEB\xA1\x8C\x00\xEB\xA1\x8D\x00"
-     "\xEB\xA1"
-     "\x8E\x00\xEB\xA1\x8F\x00\xEB\xA1\x90\x00\xEB\xA1\x91\x00\xEB\xA1\x92\x00"
-     "\xEB\xA1"
-     "\x93\x00\xEB\xA1\x94\x00\xEB\xA1\x95\x00\xEB\xA1\x96\x00\xEB\xA1\x97\x00"
-     "\xEB\xA1"
-     "\x98\x00\xEB\xA1\x99\x00\xEB\xA1\x9A\x00\xEB\xA1\x9B\x00\xEB\xA1\x9C\x00"
-     "\xEB\xA1"
-     "\x9D\x00\xEB\xA1\x9E\x00\xEB\xA1\x9F\x00\xEB\xA1\xA0\x00\xEB\xA1\xA1\x00"
-     "\xEB\xA1"
-     "\xA2\x00\xEB\xA1\xA3\x00\xEB\xA1\xA4\x00\xEB\xA1\xA5\x00\xEB\xA1\xA6\x00"
-     "\xEB\xA1"
-     "\xA7\x00\xEB\xA1\xA8\x00\xEB\xA1\xA9\x00\xEB\xA1\xAA\x00\xEB\xA1\xAB\x00"
-     "\xEB\xA1"
-     "\xAC\x00\xEB\xA1\xAD\x00\xEB\xA1\xAE\x00\xEB\xA1\xAF\x00\xEB\xA1\xB0\x00"
-     "\xEB\xA1"
-     "\xB1\x00\xEB\xA1\xB2\x00\xEB\xA1\xB3\x00\xEB\xA1\xB4\x00\xEB\xA1\xB5\x00"
-     "\xEB\xA1"
-     "\xB6\x00\xEB\xA1\xB7\x00\xEB\xA1\xB8\x00\xEB\xA1\xB9\x00\xEB\xA1\xBA\x00"
-     "\xEB\xA1"
-     "\xBB\x00\xEB\xA1\xBC\x00\xEB\xA1\xBD\x00\xEB\xA1\xBE\x00\xEB\xA1\xBF\x00"
-     "\xEB\xA2"
-     "\x80\x00\xEB\xA2\x81\x00\xEB\xA2\x82\x00\xEB\xA2\x83\x00\xEB\xA2\x84\x00"
-     "\xEB\xA2"
-     "\x85\x00\xEB\xA2\x86\x00\xEB\xA2\x87\x00\xEB\xA2\x88\x00\xEB\xA2\x89\x00"
-     "\xEB\xA2"
-     "\x8A\x00\xEB\xA2\x8B\x00\xEB\xA2\x8C\x00\xEB\xA2\x8D\x00\xEB\xA2\x8E\x00"
-     "\xEB\xA2"
-     "\x8F\x00\xEB\xA2\x90\x00\xEB\xA2\x91\x00\xEB\xA2\x92\x00\xEB\xA2\x93\x00"
-     "\xEB\xA2"
-     "\x94\x00\xEB\xA2\x95\x00\xEB\xA2\x96\x00\xEB\xA2\x97\x00\xEB\xA2\x98\x00"
-     "\xEB\xA2"
-     "\x99\x00\xEB\xA2\x9A\x00\xEB\xA2\x9B\x00\xEB\xA2\x9C\x00\xEB\xA2\x9D\x00"
-     "\xEB\xA2"
-     "\x9E\x00\xEB\xA2\x9F\x00\xEB\xA2\xA0\x00\xEB\xA2\xA1\x00\xEB\xA2\xA2\x00"
-     "\xEB\xA2"
-     "\xA3\x00\xEB\xA2\xA4\x00\xEB\xA2\xA5\x00\xEB\xA2\xA6\x00\xEB\xA2\xA7\x00"
-     "\xEB\xA2"
-     "\xA8\x00\xEB\xA2\xA9\x00\xEB\xA2\xAA\x00\xEB\xA2\xAB\x00\xEB\xA2\xAC\x00"
-     "\xEB\xA2"
-     "\xAD\x00\xEB\xA2\xAE\x00\xEB\xA2\xAF\x00\xEB\xA2\xB0\x00\xEB\xA2\xB1\x00"
-     "\xEB\xA2"
-     "\xB2\x00\xEB\xA2\xB3\x00\xEB\xA2\xB4\x00\xEB\xA2\xB5\x00\xEB\xA2\xB6\x00"
-     "\xEB\xA2"
-     "\xB7\x00\xEB\xA2\xB8\x00\xEB\xA2\xB9\x00\xEB\xA2\xBA\x00\xEB\xA2\xBB\x00"
-     "\xEB\xA2"
-     "\xBC\x00\xEB\xA2\xBD\x00\xEB\xA2\xBE\x00\xEB\xA2\xBF\x00\xEB\xA3\x80\x00"
-     "\xEB\xA3"
-     "\x81\x00\xEB\xA3\x82\x00\xEB\xA3\x83\x00\xEB\xA3\x84\x00\xEB\xA3\x85\x00"
-     "\xEB\xA3"
-     "\x86\x00\xEB\xA3\x87\x00\xEB\xA3\x88\x00\xEB\xA3\x89\x00\xEB\xA3\x8A\x00"
-     "\xEB\xA3"
-     "\x8B\x00\xEB\xA3\x8C\x00\xEB\xA3\x8D\x00\xEB\xA3\x8E\x00\xEB\xA3\x8F\x00"
-     "\xEB\xA3"
-     "\x90\x00\xEB\xA3\x91\x00\xEB\xA3\x92\x00\xEB\xA3\x93\x00\xEB\xA3\x94\x00"
-     "\xEB\xA3"
-     "\x95\x00\xEB\xA3\x96\x00\xEB\xA3\x97\x00\xEB\xA3\x98\x00\xEB\xA3\x99\x00"
-     "\xEB\xA3"
-     "\x9A\x00\xEB\xA3\x9B\x00\xEB\xA3\x9C\x00\xEB\xA3\x9D\x00\xEB\xA3\x9E\x00"
-     "\xEB\xA3"
-     "\x9F\x00\xEB\xA3\xA0\x00\xEB\xA3\xA1\x00\xEB\xA3\xA2\x00\xEB\xA3\xA3\x00"
-     "\xEB\xA3"
-     "\xA4\x00\xEB\xA3\xA5\x00\xEB\xA3\xA6\x00\xEB\xA3\xA7\x00\xEB\xA3\xA8\x00"
-     "\xEB\xA3"
-     "\xA9\x00\xEB\xA3\xAA\x00\xEB\xA3\xAB\x00\xEB\xA3\xAC\x00\xEB\xA3\xAD\x00"
-     "\xEB\xA3"
-     "\xAE\x00\xEB\xA3\xAF\x00\xEB\xA3\xB0\x00\xEB\xA3\xB1\x00\xEB\xA3\xB2\x00"
-     "\xEB\xA3"
-     "\xB3\x00\xEB\xA3\xB4\x00\xEB\xA3\xB5\x00\xEB\xA3\xB6\x00\xEB\xA3\xB7\x00"
-     "\xEB\xA3"
-     "\xB8\x00\xEB\xA3\xB9\x00\xEB\xA3\xBA\x00\xEB\xA3\xBB\x00\xEB\xA3\xBC\x00"
-     "\xEB\xA3"
-     "\xBD\x00\xEB\xA3\xBE\x00\xEB\xA3\xBF\x00\xEB\xA4\x80\x00\xEB\xA4\x81\x00"
-     "\xEB\xA4"
-     "\x82\x00\xEB\xA4\x83\x00\xEB\xA4\x84\x00\xEB\xA4\x85\x00\xEB\xA4\x86\x00"
-     "\xEB\xA4"
-     "\x87\x00\xEB\xA4\x88\x00\xEB\xA4\x89\x00\xEB\xA4\x8A\x00\xEB\xA4\x8B\x00"
-     "\xEB\xA4"
-     "\x8C\x00\xEB\xA4\x8D\x00\xEB\xA4\x8E\x00\xEB\xA4\x8F\x00\xEB\xA4\x90\x00"
-     "\xEB\xA4"
-     "\x91\x00\xEB\xA4\x92\x00\xEB\xA4\x93\x00\xEB\xA4\x94\x00\xEB\xA4\x95\x00"
-     "\xEB\xA4"
-     "\x96\x00\xEB\xA4\x97\x00\xEB\xA4\x98\x00\xEB\xA4\x99\x00\xEB\xA4\x9A\x00"
-     "\xEB\xA4"
-     "\x9B\x00\xEB\xA4\x9C\x00\xEB\xA4\x9D\x00\xEB\xA4\x9E\x00\xEB\xA4\x9F\x00"
-     "\xEB\xA4"
-     "\xA0\x00\xEB\xA4\xA1\x00\xEB\xA4\xA2\x00\xEB\xA4\xA3\x00\xEB\xA4\xA4\x00"
-     "\xEB\xA4"
-     "\xA5\x00\xEB\xA4\xA6\x00\xEB\xA4\xA7\x00\xEB\xA4\xA8\x00\xEB\xA4\xA9\x00"
-     "\xEB\xA4"
-     "\xAA\x00\xEB\xA4\xAB\x00\xEB\xA4\xAC\x00\xEB\xA4\xAD\x00\xEB\xA4\xAE\x00"
-     "\xEB\xA4"
-     "\xAF\x00\xEB\xA4\xB0\x00\xEB\xA4\xB1\x00\xEB\xA4\xB2\x00\xEB\xA4\xB3\x00"
-     "\xEB\xA4"
-     "\xB4\x00\xEB\xA4\xB5\x00\xEB\xA4\xB6\x00\xEB\xA4\xB7\x00\xEB\xA4\xB8\x00"
-     "\xEB\xA4"
-     "\xB9\x00\xEB\xA4\xBA\x00\xEB\xA4\xBB\x00\xEB\xA4\xBC\x00\xEB\xA4\xBD\x00"
-     "\xEB\xA4"
-     "\xBE\x00\xEB\xA4\xBF\x00\xEB\xA5\x80\x00\xEB\xA5\x81\x00\xEB\xA5\x82\x00"
-     "\xEB\xA5"
-     "\x83\x00\xEB\xA5\x84\x00\xEB\xA5\x85\x00\xEB\xA5\x86\x00\xEB\xA5\x87\x00"
-     "\xEB\xA5"
-     "\x88\x00\xEB\xA5\x89\x00\xEB\xA5\x8A\x00\xEB\xA5\x8B\x00\xEB\xA5\x8C\x00"
-     "\xEB\xA5"
-     "\x8D\x00\xEB\xA5\x8E\x00\xEB\xA5\x8F\x00\xEB\xA5\x90\x00\xEB\xA5\x91\x00"
-     "\xEB\xA5"
-     "\x92\x00\xEB\xA5\x93\x00\xEB\xA5\x94\x00\xEB\xA5\x95\x00\xEB\xA5\x96\x00"
-     "\xEB\xA5"
-     "\x97\x00\xEB\xA5\x98\x00\xEB\xA5\x99\x00\xEB\xA5\x9A\x00\xEB\xA5\x9B\x00"
-     "\xEB\xA5"
-     "\x9C\x00\xEB\xA5\x9D\x00\xEB\xA5\x9E\x00\xEB\xA5\x9F\x00\xEB\xA5\xA0\x00"
-     "\xEB\xA5"
-     "\xA1\x00\xEB\xA5\xA2\x00\xEB\xA5\xA3\x00\xEB\xA5\xA4\x00\xEB\xA5\xA5\x00"
-     "\xEB\xA5"
-     "\xA6\x00\xEB\xA5\xA7\x00\xEB\xA5\xA8\x00\xEB\xA5\xA9\x00\xEB\xA5\xAA\x00"
-     "\xEB\xA5"
-     "\xAB\x00\xEB\xA5\xAC\x00\xEB\xA5\xAD\x00\xEB\xA5\xAE\x00\xEB\xA5\xAF\x00"
-     "\xEB\xA5"
-     "\xB0\x00\xEB\xA5\xB1\x00\xEB\xA5\xB2\x00\xEB\xA5\xB3\x00\xEB\xA5\xB4\x00"
-     "\xEB\xA5"
-     "\xB5\x00\xEB\xA5\xB6\x00\xEB\xA5\xB7\x00\xEB\xA5\xB8\x00\xEB\xA5\xB9\x00"
-     "\xEB\xA5"
-     "\xBA\x00\xEB\xA5\xBB\x00\xEB\xA5\xBC\x00\xEB\xA5\xBD\x00\xEB\xA5\xBE\x00"
-     "\xEB\xA5"
-     "\xBF\x00\xEB\xA6\x80\x00\xEB\xA6\x81\x00\xEB\xA6\x82\x00\xEB\xA6\x83\x00"
-     "\xEB\xA6"
-     "\x84\x00\xEB\xA6\x85\x00\xEB\xA6\x86\x00\xEB\xA6\x87\x00\xEB\xA6\x88\x00"
-     "\xEB\xA6"
-     "\x89\x00\xEB\xA6\x8A\x00\xEB\xA6\x8B\x00\xEB\xA6\x8C\x00\xEB\xA6\x8D\x00"
-     "\xEB\xA6"
-     "\x8E\x00\xEB\xA6\x8F\x00\xEB\xA6\x90\x00\xEB\xA6\x91\x00\xEB\xA6\x92\x00"
-     "\xEB\xA6"
-     "\x93\x00\xEB\xA6\x94\x00\xEB\xA6\x95\x00\xEB\xA6\x96\x00\xEB\xA6\x97\x00"
-     "\xEB\xA6"
-     "\x98\x00\xEB\xA6\x99\x00\xEB\xA6\x9A\x00\xEB\xA6\x9B\x00\xEB\xA6\x9C\x00"
-     "\xEB\xA6"
-     "\x9D\x00\xEB\xA6\x9E\x00\xEB\xA6\x9F\x00\xEB\xA6\xA0\x00\xEB\xA6\xA1\x00"
-     "\xEB\xA6"
-     "\xA2\x00\xEB\xA6\xA3\x00\xEB\xA6\xA4\x00\xEB\xA6\xA5\x00\xEB\xA6\xA6\x00"
-     "\xEB\xA6"
-     "\xA7\x00\xEB\xA6\xA8\x00\xEB\xA6\xA9\x00\xEB\xA6\xAA\x00\xEB\xA6\xAB\x00"
-     "\xEB\xA6"
-     "\xAC\x00\xEB\xA6\xAD\x00\xEB\xA6\xAE\x00\xEB\xA6\xAF\x00\xEB\xA6\xB0\x00"
-     "\xEB\xA6"
-     "\xB1\x00\xEB\xA6\xB2\x00\xEB\xA6\xB3\x00\xEB\xA6\xB4\x00\xEB\xA6\xB5\x00"
-     "\xEB\xA6"
-     "\xB6\x00\xEB\xA6\xB7\x00\xEB\xA6\xB8\x00\xEB\xA6\xB9\x00\xEB\xA6\xBA\x00"
-     "\xEB\xA6"
-     "\xBB\x00\xEB\xA6\xBC\x00\xEB\xA6\xBD\x00\xEB\xA6\xBE\x00\xEB\xA6\xBF\x00"
-     "\xEB\xA7"
-     "\x80\x00\xEB\xA7\x81\x00\xEB\xA7\x82\x00\xEB\xA7\x83\x00\xEB\xA7\x84\x00"
-     "\xEB\xA7"
-     "\x85\x00\xEB\xA7\x86\x00\xEB\xA7\x87\x00\xEB\xA7\x88\x00\xEB\xA7\x89\x00"
-     "\xEB\xA7"
-     "\x8A\x00\xEB\xA7\x8B\x00\xEB\xA7\x8C\x00\xEB\xA7\x8D\x00\xEB\xA7\x8E\x00"
-     "\xEB\xA7"
-     "\x8F\x00\xEB\xA7\x90\x00\xEB\xA7\x91\x00\xEB\xA7\x92\x00\xEB\xA7\x93\x00"
-     "\xEB\xA7"
-     "\x94\x00\xEB\xA7\x95\x00\xEB\xA7\x96\x00\xEB\xA7\x97\x00\xEB\xA7\x98\x00"
-     "\xEB\xA7"
-     "\x99\x00\xEB\xA7\x9A\x00\xEB\xA7\x9B\x00\xEB\xA7\x9C\x00\xEB\xA7\x9D\x00"
-     "\xEB\xA7"
-     "\x9E\x00\xEB\xA7\x9F\x00\xEB\xA7\xA0\x00\xEB\xA7\xA1\x00\xEB\xA7\xA2\x00"
-     "\xEB\xA7"
-     "\xA3\x00\xEB\xA7\xA4\x00\xEB\xA7\xA5\x00\xEB\xA7\xA6\x00\xEB\xA7\xA7\x00"
-     "\xEB\xA7"
-     "\xA8\x00\xEB\xA7\xA9\x00\xEB\xA7\xAA\x00\xEB\xA7\xAB\x00\xEB\xA7\xAC\x00"
-     "\xEB\xA7"
-     "\xAD\x00\xEB\xA7\xAE\x00\xEB\xA7\xAF\x00\xEB\xA7\xB0\x00\xEB\xA7\xB1\x00"
-     "\xEB\xA7"
-     "\xB2\x00\xEB\xA7\xB3\x00\xEB\xA7\xB4\x00\xEB\xA7\xB5\x00\xEB\xA7\xB6\x00"
-     "\xEB\xA7"
-     "\xB7\x00\xEB\xA7\xB8\x00\xEB\xA7\xB9\x00\xEB\xA7\xBA\x00\xEB\xA7\xBB\x00"
-     "\xEB\xA7"
-     "\xBC\x00\xEB\xA7\xBD\x00\xEB\xA7\xBE\x00\xEB\xA7\xBF\x00\xEB\xA8\x80\x00"
-     "\xEB\xA8"
-     "\x81\x00\xEB\xA8\x82\x00\xEB\xA8\x83\x00\xEB\xA8\x84\x00\xEB\xA8\x85\x00"
-     "\xEB\xA8"
-     "\x86\x00\xEB\xA8\x87\x00\xEB\xA8\x88\x00\xEB\xA8\x89\x00\xEB\xA8\x8A\x00"
-     "\xEB\xA8"
-     "\x8B\x00\xEB\xA8\x8C\x00\xEB\xA8\x8D\x00\xEB\xA8\x8E\x00\xEB\xA8\x8F\x00"
-     "\xEB\xA8"
-     "\x90\x00\xEB\xA8\x91\x00\xEB\xA8\x92\x00\xEB\xA8\x93\x00\xEB\xA8\x94\x00"
-     "\xEB\xA8"
-     "\x95\x00\xEB\xA8\x96\x00\xEB\xA8\x97\x00\xEB\xA8\x98\x00\xEB\xA8\x99\x00"
-     "\xEB\xA8"
-     "\x9A\x00\xEB\xA8\x9B\x00\xEB\xA8\x9C\x00\xEB\xA8\x9D\x00\xEB\xA8\x9E\x00"
-     "\xEB\xA8"
-     "\x9F\x00\xEB\xA8\xA0\x00\xEB\xA8\xA1\x00\xEB\xA8\xA2\x00\xEB\xA8\xA3\x00"
-     "\xEB\xA8"
-     "\xA4\x00\xEB\xA8\xA5\x00\xEB\xA8\xA6\x00\xEB\xA8\xA7\x00\xEB\xA8\xA8\x00"
-     "\xEB\xA8"
-     "\xA9\x00\xEB\xA8\xAA\x00\xEB\xA8\xAB\x00\xEB\xA8\xAC\x00\xEB\xA8\xAD\x00"
-     "\xEB\xA8"
-     "\xAE\x00\xEB\xA8\xAF\x00\xEB\xA8\xB0\x00\xEB\xA8\xB1\x00\xEB\xA8\xB2\x00"
-     "\xEB\xA8"
-     "\xB3\x00\xEB\xA8\xB4\x00\xEB\xA8\xB5\x00\xEB\xA8\xB6\x00\xEB\xA8\xB7\x00"
-     "\xEB\xA8"
-     "\xB8\x00\xEB\xA8\xB9\x00\xEB\xA8\xBA\x00\xEB\xA8\xBB\x00\xEB\xA8\xBC\x00"
-     "\xEB\xA8"
-     "\xBD\x00\xEB\xA8\xBE\x00\xEB\xA8\xBF\x00\xEB\xA9\x80\x00\xEB\xA9\x81\x00"
-     "\xEB\xA9"
-     "\x82\x00\xEB\xA9\x83\x00\xEB\xA9\x84\x00\xEB\xA9\x85\x00\xEB\xA9\x86\x00"
-     "\xEB\xA9"
-     "\x87\x00\xEB\xA9\x88\x00\xEB\xA9\x89\x00\xEB\xA9\x8A\x00\xEB\xA9\x8B\x00"
-     "\xEB\xA9"
-     "\x8C\x00\xEB\xA9\x8D\x00\xEB\xA9\x8E\x00\xEB\xA9\x8F\x00\xEB\xA9\x90\x00"
-     "\xEB\xA9"
-     "\x91\x00\xEB\xA9\x92\x00\xEB\xA9\x93\x00\xEB\xA9\x94\x00\xEB\xA9\x95\x00"
-     "\xEB\xA9"
-     "\x96\x00\xEB\xA9\x97\x00\xEB\xA9\x98\x00\xEB\xA9\x99\x00\xEB\xA9\x9A\x00"
-     "\xEB\xA9"
-     "\x9B\x00\xEB\xA9\x9C\x00\xEB\xA9\x9D\x00\xEB\xA9\x9E\x00\xEB\xA9\x9F\x00"
-     "\xEB\xA9"
-     "\xA0\x00\xEB\xA9\xA1\x00\xEB\xA9\xA2\x00\xEB\xA9\xA3\x00\xEB\xA9\xA4\x00"
-     "\xEB\xA9"
-     "\xA5\x00\xEB\xA9\xA6\x00\xEB\xA9\xA7\x00\xEB\xA9\xA8\x00\xEB\xA9\xA9\x00"
-     "\xEB\xA9"
-     "\xAA\x00\xEB\xA9\xAB\x00\xEB\xA9\xAC\x00\xEB\xA9\xAD\x00\xEB\xA9\xAE\x00"
-     "\xEB\xA9"
-     "\xAF\x00\xEB\xA9\xB0\x00\xEB\xA9\xB1\x00\xEB\xA9\xB2\x00\xEB\xA9\xB3\x00"
-     "\xEB\xA9"
-     "\xB4\x00\xEB\xA9\xB5\x00\xEB\xA9\xB6\x00\xEB\xA9\xB7\x00\xEB\xA9\xB8\x00"
-     "\xEB\xA9"
-     "\xB9\x00\xEB\xA9\xBA\x00\xEB\xA9\xBB\x00\xEB\xA9\xBC\x00\xEB\xA9\xBD\x00"
-     "\xEB\xA9"
-     "\xBE\x00\xEB\xA9\xBF\x00\xEB\xAA\x80\x00\xEB\xAA\x81\x00\xEB\xAA\x82\x00"
-     "\xEB\xAA"
-     "\x83\x00\xEB\xAA\x84\x00\xEB\xAA\x85\x00\xEB\xAA\x86\x00\xEB\xAA\x87\x00"
-     "\xEB\xAA"
-     "\x88\x00\xEB\xAA\x89\x00\xEB\xAA\x8A\x00\xEB\xAA\x8B\x00\xEB\xAA\x8C\x00"
-     "\xEB\xAA"
-     "\x8D\x00\xEB\xAA\x8E\x00\xEB\xAA\x8F\x00\xEB\xAA\x90\x00\xEB\xAA\x91\x00"
-     "\xEB\xAA"
-     "\x92\x00\xEB\xAA\x93\x00\xEB\xAA\x94\x00\xEB\xAA\x95\x00\xEB\xAA\x96\x00"
-     "\xEB\xAA"
-     "\x97\x00\xEB\xAA\x98\x00\xEB\xAA\x99\x00\xEB\xAA\x9A\x00\xEB\xAA\x9B\x00"
-     "\xEB\xAA"
-     "\x9C\x00\xEB\xAA\x9D\x00\xEB\xAA\x9E\x00\xEB\xAA\x9F\x00\xEB\xAA\xA0\x00"
-     "\xEB\xAA"
-     "\xA1\x00\xEB\xAA\xA2\x00\xEB\xAA\xA3\x00\xEB\xAA\xA4\x00\xEB\xAA\xA5\x00"
-     "\xEB\xAA"
-     "\xA6\x00\xEB\xAA\xA7\x00\xEB\xAA\xA8\x00\xEB\xAA\xA9\x00\xEB\xAA\xAA\x00"
-     "\xEB\xAA"
-     "\xAB\x00\xEB\xAA\xAC\x00\xEB\xAA\xAD\x00\xEB\xAA\xAE\x00\xEB\xAA\xAF\x00"
-     "\xEB\xAA"
-     "\xB0\x00\xEB\xAA\xB1\x00\xEB\xAA\xB2\x00\xEB\xAA\xB3\x00\xEB\xAA\xB4\x00"
-     "\xEB\xAA"
-     "\xB5\x00\xEB\xAA\xB6\x00\xEB\xAA\xB7\x00\xEB\xAA\xB8\x00\xEB\xAA\xB9\x00"
-     "\xEB\xAA"
-     "\xBA\x00\xEB\xAA\xBB\x00\xEB\xAA\xBC\x00\xEB\xAA\xBD\x00\xEB\xAA\xBE\x00"
-     "\xEB\xAA"
-     "\xBF\x00\xEB\xAB\x80\x00\xEB\xAB\x81\x00\xEB\xAB\x82\x00\xEB\xAB\x83\x00"
-     "\xEB\xAB"
-     "\x84\x00\xEB\xAB\x85\x00\xEB\xAB\x86\x00\xEB\xAB\x87\x00\xEB\xAB\x88\x00"
-     "\xEB\xAB"
-     "\x89\x00\xEB\xAB\x8A\x00\xEB\xAB\x8B\x00\xEB\xAB\x8C\x00\xEB\xAB\x8D\x00"
-     "\xEB\xAB"
-     "\x8E\x00\xEB\xAB\x8F\x00\xEB\xAB\x90\x00\xEB\xAB\x91\x00\xEB\xAB\x92\x00"
-     "\xEB\xAB"
-     "\x93\x00\xEB\xAB\x94\x00\xEB\xAB\x95\x00\xEB\xAB\x96\x00\xEB\xAB\x97\x00"
-     "\xEB\xAB"
-     "\x98\x00\xEB\xAB\x99\x00\xEB\xAB\x9A\x00\xEB\xAB\x9B\x00\xEB\xAB\x9C\x00"
-     "\xEB\xAB"
-     "\x9D\x00\xEB\xAB\x9E\x00\xEB\xAB\x9F\x00\xEB\xAB\xA0\x00\xEB\xAB\xA1\x00"
-     "\xEB\xAB"
-     "\xA2\x00\xEB\xAB\xA3\x00\xEB\xAB\xA4\x00\xEB\xAB\xA5\x00\xEB\xAB\xA6\x00"
-     "\xEB\xAB"
-     "\xA7\x00\xEB\xAB\xA8\x00\xEB\xAB\xA9\x00\xEB\xAB\xAA\x00\xEB\xAB\xAB\x00"
-     "\xEB\xAB"
-     "\xAC\x00\xEB\xAB\xAD\x00\xEB\xAB\xAE\x00\xEB\xAB\xAF\x00\xEB\xAB\xB0\x00"
-     "\xEB\xAB"
-     "\xB1\x00\xEB\xAB\xB2\x00\xEB\xAB\xB3\x00\xEB\xAB\xB4\x00\xEB\xAB\xB5\x00"
-     "\xEB\xAB"
-     "\xB6\x00\xEB\xAB\xB7\x00\xEB\xAB\xB8\x00\xEB\xAB\xB9\x00\xEB\xAB\xBA\x00"
-     "\xEB\xAB"
-     "\xBB\x00\xEB\xAB\xBC\x00\xEB\xAB\xBD\x00\xEB\xAB\xBE\x00\xEB\xAB\xBF\x00"
-     "\xEB\xAC"
-     "\x80\x00\xEB\xAC\x81\x00\xEB\xAC\x82\x00\xEB\xAC\x83\x00\xEB\xAC\x84\x00"
-     "\xEB\xAC"
-     "\x85\x00\xEB\xAC\x86\x00\xEB\xAC\x87\x00\xEB\xAC\x88\x00\xEB\xAC\x89\x00"
-     "\xEB\xAC"
-     "\x8A\x00\xEB\xAC\x8B\x00\xEB\xAC\x8C\x00\xEB\xAC\x8D\x00\xEB\xAC\x8E\x00"
-     "\xEB\xAC"
-     "\x8F\x00\xEB\xAC\x90\x00\xEB\xAC\x91\x00\xEB\xAC\x92\x00\xEB\xAC\x93\x00"
-     "\xEB\xAC"
-     "\x94\x00\xEB\xAC\x95\x00\xEB\xAC\x96\x00\xEB\xAC\x97\x00\xEB\xAC\x98\x00"
-     "\xEB\xAC"
-     "\x99\x00\xEB\xAC\x9A\x00\xEB\xAC\x9B\x00\xEB\xAC\x9C\x00\xEB\xAC\x9D\x00"
-     "\xEB\xAC"
-     "\x9E\x00\xEB\xAC\x9F\x00\xEB\xAC\xA0\x00\xEB\xAC\xA1\x00\xEB\xAC\xA2\x00"
-     "\xEB\xAC"
-     "\xA3\x00\xEB\xAC\xA4\x00\xEB\xAC\xA5\x00\xEB\xAC\xA6\x00\xEB\xAC\xA7\x00"
-     "\xEB\xAC"
-     "\xA8\x00\xEB\xAC\xA9\x00\xEB\xAC\xAA\x00\xEB\xAC\xAB\x00\xEB\xAC\xAC\x00"
-     "\xEB\xAC"
-     "\xAD\x00\xEB\xAC\xAE\x00\xEB\xAC\xAF\x00\xEB\xAC\xB0\x00\xEB\xAC\xB1\x00"
-     "\xEB\xAC"
-     "\xB2\x00\xEB\xAC\xB3\x00\xEB\xAC\xB4\x00\xEB\xAC\xB5\x00\xEB\xAC\xB6\x00"
-     "\xEB\xAC"
-     "\xB7\x00\xEB\xAC\xB8\x00\xEB\xAC\xB9\x00\xEB\xAC\xBA\x00\xEB\xAC\xBB\x00"
-     "\xEB\xAC"
-     "\xBC\x00\xEB\xAC\xBD\x00\xEB\xAC\xBE\x00\xEB\xAC\xBF\x00\xEB\xAD\x80\x00"
-     "\xEB\xAD"
-     "\x81\x00\xEB\xAD\x82\x00\xEB\xAD\x83\x00\xEB\xAD\x84\x00\xEB\xAD\x85\x00"
-     "\xEB\xAD"
-     "\x86\x00\xEB\xAD\x87\x00\xEB\xAD\x88\x00\xEB\xAD\x89\x00\xEB\xAD\x8A\x00"
-     "\xEB\xAD"
-     "\x8B\x00\xEB\xAD\x8C\x00\xEB\xAD\x8D\x00\xEB\xAD\x8E\x00\xEB\xAD\x8F\x00"
-     "\xEB\xAD"
-     "\x90\x00\xEB\xAD\x91\x00\xEB\xAD\x92\x00\xEB\xAD\x93\x00\xEB\xAD\x94\x00"
-     "\xEB\xAD"
-     "\x95\x00\xEB\xAD\x96\x00\xEB\xAD\x97\x00\xEB\xAD\x98\x00\xEB\xAD\x99\x00"
-     "\xEB\xAD"
-     "\x9A\x00\xEB\xAD\x9B\x00\xEB\xAD\x9C\x00\xEB\xAD\x9D\x00\xEB\xAD\x9E\x00"
-     "\xEB\xAD"
-     "\x9F\x00\xEB\xAD\xA0\x00\xEB\xAD\xA1\x00\xEB\xAD\xA2\x00\xEB\xAD\xA3\x00"
-     "\xEB\xAD"
-     "\xA4\x00\xEB\xAD\xA5\x00\xEB\xAD\xA6\x00\xEB\xAD\xA7\x00\xEB\xAD\xA8\x00"
-     "\xEB\xAD"
-     "\xA9\x00\xEB\xAD\xAA\x00\xEB\xAD\xAB\x00\xEB\xAD\xAC\x00\xEB\xAD\xAD\x00"
-     "\xEB\xAD"
-     "\xAE\x00\xEB\xAD\xAF\x00\xEB\xAD\xB0\x00\xEB\xAD\xB1\x00\xEB\xAD\xB2\x00"
-     "\xEB\xAD"
-     "\xB3\x00\xEB\xAD\xB4\x00\xEB\xAD\xB5\x00\xEB\xAD\xB6\x00\xEB\xAD\xB7\x00"
-     "\xEB\xAD"
-     "\xB8\x00\xEB\xAD\xB9\x00\xEB\xAD\xBA\x00\xEB\xAD\xBB\x00\xEB\xAD\xBC\x00"
-     "\xEB\xAD"
-     "\xBD\x00\xEB\xAD\xBE\x00\xEB\xAD\xBF\x00\xEB\xAE\x80\x00\xEB\xAE\x81\x00"
-     "\xEB\xAE"
-     "\x82\x00\xEB\xAE\x83\x00\xEB\xAE\x84\x00\xEB\xAE\x85\x00\xEB\xAE\x86\x00"
-     "\xEB\xAE"
-     "\x87\x00\xEB\xAE\x88\x00\xEB\xAE\x89\x00\xEB\xAE\x8A\x00\xEB\xAE\x8B\x00"
-     "\xEB\xAE"
-     "\x8C\x00\xEB\xAE\x8D\x00\xEB\xAE\x8E\x00\xEB\xAE\x8F\x00\xEB\xAE\x90\x00"
-     "\xEB\xAE"
-     "\x91\x00\xEB\xAE\x92\x00\xEB\xAE\x93\x00\xEB\xAE\x94\x00\xEB\xAE\x95\x00"
-     "\xEB\xAE"
-     "\x96\x00\xEB\xAE\x97\x00\xEB\xAE\x98\x00\xEB\xAE\x99\x00\xEB\xAE\x9A\x00"
-     "\xEB\xAE"
-     "\x9B\x00\xEB\xAE\x9C\x00\xEB\xAE\x9D\x00\xEB\xAE\x9E\x00\xEB\xAE\x9F\x00"
-     "\xEB\xAE"
-     "\xA0\x00\xEB\xAE\xA1\x00\xEB\xAE\xA2\x00\xEB\xAE\xA3\x00\xEB\xAE\xA4\x00"
-     "\xEB\xAE"
-     "\xA5\x00\xEB\xAE\xA6\x00\xEB\xAE\xA7\x00\xEB\xAE\xA8\x00\xEB\xAE\xA9\x00"
-     "\xEB\xAE"
-     "\xAA\x00\xEB\xAE\xAB\x00\xEB\xAE\xAC\x00\xEB\xAE\xAD\x00\xEB\xAE\xAE\x00"
-     "\xEB\xAE"
-     "\xAF\x00\xEB\xAE\xB0\x00\xEB\xAE\xB1\x00\xEB\xAE\xB2\x00\xEB\xAE\xB3\x00"
-     "\xEB\xAE"
-     "\xB4\x00\xEB\xAE\xB5\x00\xEB\xAE\xB6\x00\xEB\xAE\xB7\x00\xEB\xAE\xB8\x00"
-     "\xEB\xAE"
-     "\xB9\x00\xEB\xAE\xBA\x00\xEB\xAE\xBB\x00\xEB\xAE\xBC\x00\xEB\xAE\xBD\x00"
-     "\xEB\xAE"
-     "\xBE\x00\xEB\xAE\xBF\x00\xEB\xAF\x80\x00\xEB\xAF\x81\x00\xEB\xAF\x82\x00"
-     "\xEB\xAF"
-     "\x83\x00\xEB\xAF\x84\x00\xEB\xAF\x85\x00\xEB\xAF\x86\x00\xEB\xAF\x87\x00"
-     "\xEB\xAF"
-     "\x88\x00\xEB\xAF\x89\x00\xEB\xAF\x8A\x00\xEB\xAF\x8B\x00\xEB\xAF\x8C\x00"
-     "\xEB\xAF"
-     "\x8D\x00\xEB\xAF\x8E\x00\xEB\xAF\x8F\x00\xEB\xAF\x90\x00\xEB\xAF\x91\x00"
-     "\xEB\xAF"
-     "\x92\x00\xEB\xAF\x93\x00\xEB\xAF\x94\x00\xEB\xAF\x95\x00\xEB\xAF\x96\x00"
-     "\xEB\xAF"
-     "\x97\x00\xEB\xAF\x98\x00\xEB\xAF\x99\x00\xEB\xAF\x9A\x00\xEB\xAF\x9B\x00"
-     "\xEB\xAF"
-     "\x9C\x00\xEB\xAF\x9D\x00\xEB\xAF\x9E\x00\xEB\xAF\x9F\x00\xEB\xAF\xA0\x00"
-     "\xEB\xAF"
-     "\xA1\x00\xEB\xAF\xA2\x00\xEB\xAF\xA3\x00\xEB\xAF\xA4\x00\xEB\xAF\xA5\x00"
-     "\xEB\xAF"
-     "\xA6\x00\xEB\xAF\xA7\x00\xEB\xAF\xA8\x00\xEB\xAF\xA9\x00\xEB\xAF\xAA\x00"
-     "\xEB\xAF"
-     "\xAB\x00\xEB\xAF\xAC\x00\xEB\xAF\xAD\x00\xEB\xAF\xAE\x00\xEB\xAF\xAF\x00"
-     "\xEB\xAF"
-     "\xB0\x00\xEB\xAF\xB1\x00\xEB\xAF\xB2\x00\xEB\xAF\xB3\x00\xEB\xAF\xB4\x00"
-     "\xEB\xAF"
-     "\xB5\x00\xEB\xAF\xB6\x00\xEB\xAF\xB7\x00\xEB\xAF\xB8\x00\xEB\xAF\xB9\x00"
-     "\xEB\xAF"
-     "\xBA\x00\xEB\xAF\xBB\x00\xEB\xAF\xBC\x00\xEB\xAF\xBD\x00\xEB\xAF\xBE\x00"
-     "\xEB\xAF"
-     "\xBF\x00\xEB\xB0\x80\x00\xEB\xB0\x81\x00\xEB\xB0\x82\x00\xEB\xB0\x83\x00"
-     "\xEB\xB0"
-     "\x84\x00\xEB\xB0\x85\x00\xEB\xB0\x86\x00\xEB\xB0\x87\x00\xEB\xB0\x88\x00"
-     "\xEB\xB0"
-     "\x89\x00\xEB\xB0\x8A\x00\xEB\xB0\x8B\x00\xEB\xB0\x8C\x00\xEB\xB0\x8D\x00"
-     "\xEB\xB0"
-     "\x8E\x00\xEB\xB0\x8F\x00\xEB\xB0\x90\x00\xEB\xB0\x91\x00\xEB\xB0\x92\x00"
-     "\xEB\xB0"
-     "\x93\x00\xEB\xB0\x94\x00\xEB\xB0\x95\x00\xEB\xB0\x96\x00\xEB\xB0\x97\x00"
-     "\xEB\xB0"
-     "\x98\x00\xEB\xB0\x99\x00\xEB\xB0\x9A\x00\xEB\xB0\x9B\x00\xEB\xB0\x9C\x00"
-     "\xEB\xB0"
-     "\x9D\x00\xEB\xB0\x9E\x00\xEB\xB0\x9F\x00\xEB\xB0\xA0\x00\xEB\xB0\xA1\x00"
-     "\xEB\xB0"
-     "\xA2\x00\xEB\xB0\xA3\x00\xEB\xB0\xA4\x00\xEB\xB0\xA5\x00\xEB\xB0\xA6\x00"
-     "\xEB\xB0"
-     "\xA7\x00\xEB\xB0\xA8\x00\xEB\xB0\xA9\x00\xEB\xB0\xAA\x00\xEB\xB0\xAB\x00"
-     "\xEB\xB0"
-     "\xAC\x00\xEB\xB0\xAD\x00\xEB\xB0\xAE\x00\xEB\xB0\xAF\x00\xEB\xB0\xB0\x00"
-     "\xEB\xB0"
-     "\xB1\x00\xEB\xB0\xB2\x00\xEB\xB0\xB3\x00\xEB\xB0\xB4\x00\xEB\xB0\xB5\x00"
-     "\xEB\xB0"
-     "\xB6\x00\xEB\xB0\xB7\x00\xEB\xB0\xB8\x00\xEB\xB0\xB9\x00\xEB\xB0\xBA\x00"
-     "\xEB\xB0"
-     "\xBB\x00\xEB\xB0\xBC\x00\xEB\xB0\xBD\x00\xEB\xB0\xBE\x00\xEB\xB0\xBF\x00"
-     "\xEB\xB1"
-     "\x80\x00\xEB\xB1\x81\x00\xEB\xB1\x82\x00\xEB\xB1\x83\x00\xEB\xB1\x84\x00"
-     "\xEB\xB1"
-     "\x85\x00\xEB\xB1\x86\x00\xEB\xB1\x87\x00\xEB\xB1\x88\x00\xEB\xB1\x89\x00"
-     "\xEB\xB1"
-     "\x8A\x00\xEB\xB1\x8B\x00\xEB\xB1\x8C\x00\xEB\xB1\x8D\x00\xEB\xB1\x8E\x00"
-     "\xEB\xB1"
-     "\x8F\x00\xEB\xB1\x90\x00\xEB\xB1\x91\x00\xEB\xB1\x92\x00\xEB\xB1\x93\x00"
-     "\xEB\xB1"
-     "\x94\x00\xEB\xB1\x95\x00\xEB\xB1\x96\x00\xEB\xB1\x97\x00\xEB\xB1\x98\x00"
-     "\xEB\xB1"
-     "\x99\x00\xEB\xB1\x9A\x00\xEB\xB1\x9B\x00\xEB\xB1\x9C\x00\xEB\xB1\x9D\x00"
-     "\xEB\xB1"
-     "\x9E\x00\xEB\xB1\x9F\x00\xEB\xB1\xA0\x00\xEB\xB1\xA1\x00\xEB\xB1\xA2\x00"
-     "\xEB\xB1"
-     "\xA3\x00\xEB\xB1\xA4\x00\xEB\xB1\xA5\x00\xEB\xB1\xA6\x00\xEB\xB1\xA7\x00"
-     "\xEB\xB1"
-     "\xA8\x00\xEB\xB1\xA9\x00\xEB\xB1\xAA\x00\xEB\xB1\xAB\x00\xEB\xB1\xAC\x00"
-     "\xEB\xB1"
-     "\xAD\x00\xEB\xB1\xAE\x00\xEB\xB1\xAF\x00\xEB\xB1\xB0\x00\xEB\xB1\xB1\x00"
-     "\xEB\xB1"
-     "\xB2\x00\xEB\xB1\xB3\x00\xEB\xB1\xB4\x00\xEB\xB1\xB5\x00\xEB\xB1\xB6\x00"
-     "\xEB\xB1"
-     "\xB7\x00\xEB\xB1\xB8\x00\xEB\xB1\xB9\x00\xEB\xB1\xBA\x00\xEB\xB1\xBB\x00"
-     "\xEB\xB1"
-     "\xBC\x00\xEB\xB1\xBD\x00\xEB\xB1\xBE\x00\xEB\xB1\xBF\x00\xEB\xB2\x80\x00"
-     "\xEB\xB2"
-     "\x81\x00\xEB\xB2\x82\x00\xEB\xB2\x83\x00\xEB\xB2\x84\x00\xEB\xB2\x85\x00"
-     "\xEB\xB2"
-     "\x86\x00\xEB\xB2\x87\x00\xEB\xB2\x88\x00\xEB\xB2\x89\x00\xEB\xB2\x8A\x00"
-     "\xEB\xB2"
-     "\x8B\x00\xEB\xB2\x8C\x00\xEB\xB2\x8D\x00\xEB\xB2\x8E\x00\xEB\xB2\x8F\x00"
-     "\xEB\xB2"
-     "\x90\x00\xEB\xB2\x91\x00\xEB\xB2\x92\x00\xEB\xB2\x93\x00\xEB\xB2\x94\x00"
-     "\xEB\xB2"
-     "\x95\x00\xEB\xB2\x96\x00\xEB\xB2\x97\x00\xEB\xB2\x98\x00\xEB\xB2\x99\x00"
-     "\xEB\xB2"
-     "\x9A\x00\xEB\xB2\x9B\x00\xEB\xB2\x9C\x00\xEB\xB2\x9D\x00\xEB\xB2\x9E\x00"
-     "\xEB\xB2"
-     "\x9F\x00\xEB\xB2\xA0\x00\xEB\xB2\xA1\x00\xEB\xB2\xA2\x00\xEB\xB2\xA3\x00"
-     "\xEB\xB2"
-     "\xA4\x00\xEB\xB2\xA5\x00\xEB\xB2\xA6\x00\xEB\xB2\xA7\x00\xEB\xB2\xA8\x00"
-     "\xEB\xB2"
-     "\xA9\x00\xEB\xB2\xAA\x00\xEB\xB2\xAB\x00\xEB\xB2\xAC\x00\xEB\xB2\xAD\x00"
-     "\xEB\xB2"
-     "\xAE\x00\xEB\xB2\xAF\x00\xEB\xB2\xB0\x00\xEB\xB2\xB1\x00\xEB\xB2\xB2\x00"
-     "\xEB\xB2"
-     "\xB3\x00\xEB\xB2\xB4\x00\xEB\xB2\xB5\x00\xEB\xB2\xB6\x00\xEB\xB2\xB7\x00"
-     "\xEB\xB2"
-     "\xB8\x00\xEB\xB2\xB9\x00\xEB\xB2\xBA\x00\xEB\xB2\xBB\x00\xEB\xB2\xBC\x00"
-     "\xEB\xB2"
-     "\xBD\x00\xEB\xB2\xBE\x00\xEB\xB2\xBF\x00\xEB\xB3\x80\x00\xEB\xB3\x81\x00"
-     "\xEB\xB3"
-     "\x82\x00\xEB\xB3\x83\x00\xEB\xB3\x84\x00\xEB\xB3\x85\x00\xEB\xB3\x86\x00"
-     "\xEB\xB3"
-     "\x87\x00\xEB\xB3\x88\x00\xEB\xB3\x89\x00\xEB\xB3\x8A\x00\xEB\xB3\x8B\x00"
-     "\xEB\xB3"
-     "\x8C\x00\xEB\xB3\x8D\x00\xEB\xB3\x8E\x00\xEB\xB3\x8F\x00\xEB\xB3\x90\x00"
-     "\xEB\xB3"
-     "\x91\x00\xEB\xB3\x92\x00\xEB\xB3\x93\x00\xEB\xB3\x94\x00\xEB\xB3\x95\x00"
-     "\xEB\xB3"
-     "\x96\x00\xEB\xB3\x97\x00\xEB\xB3\x98\x00\xEB\xB3\x99\x00\xEB\xB3\x9A\x00"
-     "\xEB\xB3"
-     "\x9B\x00\xEB\xB3\x9C\x00\xEB\xB3\x9D\x00\xEB\xB3\x9E\x00\xEB\xB3\x9F\x00"
-     "\xEB\xB3"
-     "\xA0\x00\xEB\xB3\xA1\x00\xEB\xB3\xA2\x00\xEB\xB3\xA3\x00\xEB\xB3\xA4\x00"
-     "\xEB\xB3"
-     "\xA5\x00\xEB\xB3\xA6\x00\xEB\xB3\xA7\x00\xEB\xB3\xA8\x00\xEB\xB3\xA9\x00"
-     "\xEB\xB3"
-     "\xAA\x00\xEB\xB3\xAB\x00\xEB\xB3\xAC\x00\xEB\xB3\xAD\x00\xEB\xB3\xAE\x00"
-     "\xEB\xB3"
-     "\xAF\x00\xEB\xB3\xB0\x00\xEB\xB3\xB1\x00\xEB\xB3\xB2\x00\xEB\xB3\xB3\x00"
-     "\xEB\xB3"
-     "\xB4\x00\xEB\xB3\xB5\x00\xEB\xB3\xB6\x00\xEB\xB3\xB7\x00\xEB\xB3\xB8\x00"
-     "\xEB\xB3"
-     "\xB9\x00\xEB\xB3\xBA\x00\xEB\xB3\xBB\x00\xEB\xB3\xBC\x00\xEB\xB3\xBD\x00"
-     "\xEB\xB3"
-     "\xBE\x00\xEB\xB3\xBF\x00\xEB\xB4\x80\x00\xEB\xB4\x81\x00\xEB\xB4\x82\x00"
-     "\xEB\xB4"
-     "\x83\x00\xEB\xB4\x84\x00\xEB\xB4\x85\x00\xEB\xB4\x86\x00\xEB\xB4\x87\x00"
-     "\xEB\xB4"
-     "\x88\x00\xEB\xB4\x89\x00\xEB\xB4\x8A\x00\xEB\xB4\x8B\x00\xEB\xB4\x8C\x00"
-     "\xEB\xB4"
-     "\x8D\x00\xEB\xB4\x8E\x00\xEB\xB4\x8F\x00\xEB\xB4\x90\x00\xEB\xB4\x91\x00"
-     "\xEB\xB4"
-     "\x92\x00\xEB\xB4\x93\x00\xEB\xB4\x94\x00\xEB\xB4\x95\x00\xEB\xB4\x96\x00"
-     "\xEB\xB4"
-     "\x97\x00\xEB\xB4\x98\x00\xEB\xB4\x99\x00\xEB\xB4\x9A\x00\xEB\xB4\x9B\x00"
-     "\xEB\xB4"
-     "\x9C\x00\xEB\xB4\x9D\x00\xEB\xB4\x9E\x00\xEB\xB4\x9F\x00\xEB\xB4\xA0\x00"
-     "\xEB\xB4"
-     "\xA1\x00\xEB\xB4\xA2\x00\xEB\xB4\xA3\x00\xEB\xB4\xA4\x00\xEB\xB4\xA5\x00"
-     "\xEB\xB4"
-     "\xA6\x00\xEB\xB4\xA7\x00\xEB\xB4\xA8\x00\xEB\xB4\xA9\x00\xEB\xB4\xAA\x00"
-     "\xEB\xB4"
-     "\xAB\x00\xEB\xB4\xAC\x00\xEB\xB4\xAD\x00\xEB\xB4\xAE\x00\xEB\xB4\xAF\x00"
-     "\xEB\xB4"
-     "\xB0\x00\xEB\xB4\xB1\x00\xEB\xB4\xB2\x00\xEB\xB4\xB3\x00\xEB\xB4\xB4\x00"
-     "\xEB\xB4"
-     "\xB5\x00\xEB\xB4\xB6\x00\xEB\xB4\xB7\x00\xEB\xB4\xB8\x00\xEB\xB4\xB9\x00"
-     "\xEB\xB4"
-     "\xBA\x00\xEB\xB4\xBB\x00\xEB\xB4\xBC\x00\xEB\xB4\xBD\x00\xEB\xB4\xBE\x00"
-     "\xEB\xB4"
-     "\xBF\x00\xEB\xB5\x80\x00\xEB\xB5\x81\x00\xEB\xB5\x82\x00\xEB\xB5\x83\x00"
-     "\xEB\xB5"
-     "\x84\x00\xEB\xB5\x85\x00\xEB\xB5\x86\x00\xEB\xB5\x87\x00\xEB\xB5\x88\x00"
-     "\xEB\xB5"
-     "\x89\x00\xEB\xB5\x8A\x00\xEB\xB5\x8B\x00\xEB\xB5\x8C\x00\xEB\xB5\x8D\x00"
-     "\xEB\xB5"
-     "\x8E\x00\xEB\xB5\x8F\x00\xEB\xB5\x90\x00\xEB\xB5\x91\x00\xEB\xB5\x92\x00"
-     "\xEB\xB5"
-     "\x93\x00\xEB\xB5\x94\x00\xEB\xB5\x95\x00\xEB\xB5\x96\x00\xEB\xB5\x97\x00"
-     "\xEB\xB5"
-     "\x98\x00\xEB\xB5\x99\x00\xEB\xB5\x9A\x00\xEB\xB5\x9B\x00\xEB\xB5\x9C\x00"
-     "\xEB\xB5"
-     "\x9D\x00\xEB\xB5\x9E\x00\xEB\xB5\x9F\x00\xEB\xB5\xA0\x00\xEB\xB5\xA1\x00"
-     "\xEB\xB5"
-     "\xA2\x00\xEB\xB5\xA3\x00\xEB\xB5\xA4\x00\xEB\xB5\xA5\x00\xEB\xB5\xA6\x00"
-     "\xEB\xB5"
-     "\xA7\x00\xEB\xB5\xA8\x00\xEB\xB5\xA9\x00\xEB\xB5\xAA\x00\xEB\xB5\xAB\x00"
-     "\xEB\xB5"
-     "\xAC\x00\xEB\xB5\xAD\x00\xEB\xB5\xAE\x00\xEB\xB5\xAF\x00\xEB\xB5\xB0\x00"
-     "\xEB\xB5"
-     "\xB1\x00\xEB\xB5\xB2\x00\xEB\xB5\xB3\x00\xEB\xB5\xB4\x00\xEB\xB5\xB5\x00"
-     "\xEB\xB5"
-     "\xB6\x00\xEB\xB5\xB7\x00\xEB\xB5\xB8\x00\xEB\xB5\xB9\x00\xEB\xB5\xBA\x00"
-     "\xEB\xB5"
-     "\xBB\x00\xEB\xB5\xBC\x00\xEB\xB5\xBD\x00\xEB\xB5\xBE\x00\xEB\xB5\xBF\x00"
-     "\xEB\xB6"
-     "\x80\x00\xEB\xB6\x81\x00\xEB\xB6\x82\x00\xEB\xB6\x83\x00\xEB\xB6\x84\x00"
-     "\xEB\xB6"
-     "\x85\x00\xEB\xB6\x86\x00\xEB\xB6\x87\x00\xEB\xB6\x88\x00\xEB\xB6\x89\x00"
-     "\xEB\xB6"
-     "\x8A\x00\xEB\xB6\x8B\x00\xEB\xB6\x8C\x00\xEB\xB6\x8D\x00\xEB\xB6\x8E\x00"
-     "\xEB\xB6"
-     "\x8F\x00\xEB\xB6\x90\x00\xEB\xB6\x91\x00\xEB\xB6\x92\x00\xEB\xB6\x93\x00"
-     "\xEB\xB6"
-     "\x94\x00\xEB\xB6\x95\x00\xEB\xB6\x96\x00\xEB\xB6\x97\x00\xEB\xB6\x98\x00"
-     "\xEB\xB6"
-     "\x99\x00\xEB\xB6\x9A\x00\xEB\xB6\x9B\x00\xEB\xB6\x9C\x00\xEB\xB6\x9D\x00"
-     "\xEB\xB6"
-     "\x9E\x00\xEB\xB6\x9F\x00\xEB\xB6\xA0\x00\xEB\xB6\xA1\x00\xEB\xB6\xA2\x00"
-     "\xEB\xB6"
-     "\xA3\x00\xEB\xB6\xA4\x00\xEB\xB6\xA5\x00\xEB\xB6\xA6\x00\xEB\xB6\xA7\x00"
-     "\xEB\xB6"
-     "\xA8\x00\xEB\xB6\xA9\x00\xEB\xB6\xAA\x00\xEB\xB6\xAB\x00\xEB\xB6\xAC\x00"
-     "\xEB\xB6"
-     "\xAD\x00\xEB\xB6\xAE\x00\xEB\xB6\xAF\x00\xEB\xB6\xB0\x00\xEB\xB6\xB1\x00"
-     "\xEB\xB6"
-     "\xB2\x00\xEB\xB6\xB3\x00\xEB\xB6\xB4\x00\xEB\xB6\xB5\x00\xEB\xB6\xB6\x00"
-     "\xEB\xB6"
-     "\xB7\x00\xEB\xB6\xB8\x00\xEB\xB6\xB9\x00\xEB\xB6\xBA\x00\xEB\xB6\xBB\x00"
-     "\xEB\xB6"
-     "\xBC\x00\xEB\xB6\xBD\x00\xEB\xB6\xBE\x00\xEB\xB6\xBF\x00\xEB\xB7\x80\x00"
-     "\xEB\xB7"
-     "\x81\x00\xEB\xB7\x82\x00\xEB\xB7\x83\x00\xEB\xB7\x84\x00\xEB\xB7\x85\x00"
-     "\xEB\xB7"
-     "\x86\x00\xEB\xB7\x87\x00\xEB\xB7\x88\x00\xEB\xB7\x89\x00\xEB\xB7\x8A\x00"
-     "\xEB\xB7"
-     "\x8B\x00\xEB\xB7\x8C\x00\xEB\xB7\x8D\x00\xEB\xB7\x8E\x00\xEB\xB7\x8F\x00"
-     "\xEB\xB7"
-     "\x90\x00\xEB\xB7\x91\x00\xEB\xB7\x92\x00\xEB\xB7\x93\x00\xEB\xB7\x94\x00"
-     "\xEB\xB7"
-     "\x95\x00\xEB\xB7\x96\x00\xEB\xB7\x97\x00\xEB\xB7\x98\x00\xEB\xB7\x99\x00"
-     "\xEB\xB7"
-     "\x9A\x00\xEB\xB7\x9B\x00\xEB\xB7\x9C\x00\xEB\xB7\x9D\x00\xEB\xB7\x9E\x00"
-     "\xEB\xB7"
-     "\x9F\x00\xEB\xB7\xA0\x00\xEB\xB7\xA1\x00\xEB\xB7\xA2\x00\xEB\xB7\xA3\x00"
-     "\xEB\xB7"
-     "\xA4\x00\xEB\xB7\xA5\x00\xEB\xB7\xA6\x00\xEB\xB7\xA7\x00\xEB\xB7\xA8\x00"
-     "\xEB\xB7"
-     "\xA9\x00\xEB\xB7\xAA\x00\xEB\xB7\xAB\x00\xEB\xB7\xAC\x00\xEB\xB7\xAD\x00"
-     "\xEB\xB7"
-     "\xAE\x00\xEB\xB7\xAF\x00\xEB\xB7\xB0\x00\xEB\xB7\xB1\x00\xEB\xB7\xB2\x00"
-     "\xEB\xB7"
-     "\xB3\x00\xEB\xB7\xB4\x00\xEB\xB7\xB5\x00\xEB\xB7\xB6\x00\xEB\xB7\xB7\x00"
-     "\xEB\xB7"
-     "\xB8\x00\xEB\xB7\xB9\x00\xEB\xB7\xBA\x00\xEB\xB7\xBB\x00\xEB\xB7\xBC\x00"
-     "\xEB\xB7"
-     "\xBD\x00\xEB\xB7\xBE\x00\xEB\xB7\xBF\x00\xEB\xB8\x80\x00\xEB\xB8\x81\x00"
-     "\xEB\xB8"
-     "\x82\x00\xEB\xB8\x83\x00\xEB\xB8\x84\x00\xEB\xB8\x85\x00\xEB\xB8\x86\x00"
-     "\xEB\xB8"
-     "\x87\x00\xEB\xB8\x88\x00\xEB\xB8\x89\x00\xEB\xB8\x8A\x00\xEB\xB8\x8B\x00"
-     "\xEB\xB8"
-     "\x8C\x00\xEB\xB8\x8D\x00\xEB\xB8\x8E\x00\xEB\xB8\x8F\x00\xEB\xB8\x90\x00"
-     "\xEB\xB8"
-     "\x91\x00\xEB\xB8\x92\x00\xEB\xB8\x93\x00\xEB\xB8\x94\x00\xEB\xB8\x95\x00"
-     "\xEB\xB8"
-     "\x96\x00\xEB\xB8\x97\x00\xEB\xB8\x98\x00\xEB\xB8\x99\x00\xEB\xB8\x9A\x00"
-     "\xEB\xB8"
-     "\x9B\x00\xEB\xB8\x9C\x00\xEB\xB8\x9D\x00\xEB\xB8\x9E\x00\xEB\xB8\x9F\x00"
-     "\xEB\xB8"
-     "\xA0\x00\xEB\xB8\xA1\x00\xEB\xB8\xA2\x00\xEB\xB8\xA3\x00\xEB\xB8\xA4\x00"
-     "\xEB\xB8"
-     "\xA5\x00\xEB\xB8\xA6\x00\xEB\xB8\xA7\x00\xEB\xB8\xA8\x00\xEB\xB8\xA9\x00"
-     "\xEB\xB8"
-     "\xAA\x00\xEB\xB8\xAB\x00\xEB\xB8\xAC\x00\xEB\xB8\xAD\x00\xEB\xB8\xAE\x00"
-     "\xEB\xB8"
-     "\xAF\x00\xEB\xB8\xB0\x00\xEB\xB8\xB1\x00\xEB\xB8\xB2\x00\xEB\xB8\xB3\x00"
-     "\xEB\xB8"
-     "\xB4\x00\xEB\xB8\xB5\x00\xEB\xB8\xB6\x00\xEB\xB8\xB7\x00\xEB\xB8\xB8\x00"
-     "\xEB\xB8"
-     "\xB9\x00\xEB\xB8\xBA\x00\xEB\xB8\xBB\x00\xEB\xB8\xBC\x00\xEB\xB8\xBD\x00"
-     "\xEB\xB8"
-     "\xBE\x00\xEB\xB8\xBF\x00\xEB\xB9\x80\x00\xEB\xB9\x81\x00\xEB\xB9\x82\x00"
-     "\xEB\xB9"
-     "\x83\x00\xEB\xB9\x84\x00\xEB\xB9\x85\x00\xEB\xB9\x86\x00\xEB\xB9\x87\x00"
-     "\xEB\xB9"
-     "\x88\x00\xEB\xB9\x89\x00\xEB\xB9\x8A\x00\xEB\xB9\x8B\x00\xEB\xB9\x8C\x00"
-     "\xEB\xB9"
-     "\x8D\x00\xEB\xB9\x8E\x00\xEB\xB9\x8F\x00\xEB\xB9\x90\x00\xEB\xB9\x91\x00"
-     "\xEB\xB9"
-     "\x92\x00\xEB\xB9\x93\x00\xEB\xB9\x94\x00\xEB\xB9\x95\x00\xEB\xB9\x96\x00"
-     "\xEB\xB9"
-     "\x97\x00\xEB\xB9\x98\x00\xEB\xB9\x99\x00\xEB\xB9\x9A\x00\xEB\xB9\x9B\x00"
-     "\xEB\xB9"
-     "\x9C\x00\xEB\xB9\x9D\x00\xEB\xB9\x9E\x00\xEB\xB9\x9F\x00\xEB\xB9\xA0\x00"
-     "\xEB\xB9"
-     "\xA1\x00\xEB\xB9\xA2\x00\xEB\xB9\xA3\x00\xEB\xB9\xA4\x00\xEB\xB9\xA5\x00"
-     "\xEB\xB9"
-     "\xA6\x00\xEB\xB9\xA7\x00\xEB\xB9\xA8\x00\xEB\xB9\xA9\x00\xEB\xB9\xAA\x00"
-     "\xEB\xB9"
-     "\xAB\x00\xEB\xB9\xAC\x00\xEB\xB9\xAD\x00\xEB\xB9\xAE\x00\xEB\xB9\xAF\x00"
-     "\xEB\xB9"
-     "\xB0\x00\xEB\xB9\xB1\x00\xEB\xB9\xB2\x00\xEB\xB9\xB3\x00\xEB\xB9\xB4\x00"
-     "\xEB\xB9"
-     "\xB5\x00\xEB\xB9\xB6\x00\xEB\xB9\xB7\x00\xEB\xB9\xB8\x00\xEB\xB9\xB9\x00"
-     "\xEB\xB9"
-     "\xBA\x00\xEB\xB9\xBB\x00\xEB\xB9\xBC\x00\xEB\xB9\xBD\x00\xEB\xB9\xBE\x00"
-     "\xEB\xB9"
-     "\xBF\x00\xEB\xBA\x80\x00\xEB\xBA\x81\x00\xEB\xBA\x82\x00\xEB\xBA\x83\x00"
-     "\xEB\xBA"
-     "\x84\x00\xEB\xBA\x85\x00\xEB\xBA\x86\x00\xEB\xBA\x87\x00\xEB\xBA\x88\x00"
-     "\xEB\xBA"
-     "\x89\x00\xEB\xBA\x8A\x00\xEB\xBA\x8B\x00\xEB\xBA\x8C\x00\xEB\xBA\x8D\x00"
-     "\xEB\xBA"
-     "\x8E\x00\xEB\xBA\x8F\x00\xEB\xBA\x90\x00\xEB\xBA\x91\x00\xEB\xBA\x92\x00"
-     "\xEB\xBA"
-     "\x93\x00\xEB\xBA\x94\x00\xEB\xBA\x95\x00\xEB\xBA\x96\x00\xEB\xBA\x97\x00"
-     "\xEB\xBA"
-     "\x98\x00\xEB\xBA\x99\x00\xEB\xBA\x9A\x00\xEB\xBA\x9B\x00\xEB\xBA\x9C\x00"
-     "\xEB\xBA"
-     "\x9D\x00\xEB\xBA\x9E\x00\xEB\xBA\x9F\x00\xEB\xBA\xA0\x00\xEB\xBA\xA1\x00"
-     "\xEB\xBA"
-     "\xA2\x00\xEB\xBA\xA3\x00\xEB\xBA\xA4\x00\xEB\xBA\xA5\x00\xEB\xBA\xA6\x00"
-     "\xEB\xBA"
-     "\xA7\x00\xEB\xBA\xA8\x00\xEB\xBA\xA9\x00\xEB\xBA\xAA\x00\xEB\xBA\xAB\x00"
-     "\xEB\xBA"
-     "\xAC\x00\xEB\xBA\xAD\x00\xEB\xBA\xAE\x00\xEB\xBA\xAF\x00\xEB\xBA\xB0\x00"
-     "\xEB\xBA"
-     "\xB1\x00\xEB\xBA\xB2\x00\xEB\xBA\xB3\x00\xEB\xBA\xB4\x00\xEB\xBA\xB5\x00"
-     "\xEB\xBA"
-     "\xB6\x00\xEB\xBA\xB7\x00\xEB\xBA\xB8\x00\xEB\xBA\xB9\x00\xEB\xBA\xBA\x00"
-     "\xEB\xBA"
-     "\xBB\x00\xEB\xBA\xBC\x00\xEB\xBA\xBD\x00\xEB\xBA\xBE\x00\xEB\xBA\xBF\x00"
-     "\xEB\xBB"
-     "\x80\x00\xEB\xBB\x81\x00\xEB\xBB\x82\x00\xEB\xBB\x83\x00\xEB\xBB\x84\x00"
-     "\xEB\xBB"
-     "\x85\x00\xEB\xBB\x86\x00\xEB\xBB\x87\x00\xEB\xBB\x88\x00\xEB\xBB\x89\x00"
-     "\xEB\xBB"
-     "\x8A\x00\xEB\xBB\x8B\x00\xEB\xBB\x8C\x00\xEB\xBB\x8D\x00\xEB\xBB\x8E\x00"
-     "\xEB\xBB"
-     "\x8F\x00\xEB\xBB\x90\x00\xEB\xBB\x91\x00\xEB\xBB\x92\x00\xEB\xBB\x93\x00"
-     "\xEB\xBB"
-     "\x94\x00\xEB\xBB\x95\x00\xEB\xBB\x96\x00\xEB\xBB\x97\x00\xEB\xBB\x98\x00"
-     "\xEB\xBB"
-     "\x99\x00\xEB\xBB\x9A\x00\xEB\xBB\x9B\x00\xEB\xBB\x9C\x00\xEB\xBB\x9D\x00"
-     "\xEB\xBB"
-     "\x9E\x00\xEB\xBB\x9F\x00\xEB\xBB\xA0\x00\xEB\xBB\xA1\x00\xEB\xBB\xA2\x00"
-     "\xEB\xBB"
-     "\xA3\x00\xEB\xBB\xA4\x00\xEB\xBB\xA5\x00\xEB\xBB\xA6\x00\xEB\xBB\xA7\x00"
-     "\xEB\xBB"
-     "\xA8\x00\xEB\xBB\xA9\x00\xEB\xBB\xAA\x00\xEB\xBB\xAB\x00\xEB\xBB\xAC\x00"
-     "\xEB\xBB"
-     "\xAD\x00\xEB\xBB\xAE\x00\xEB\xBB\xAF\x00\xEB\xBB\xB0\x00\xEB\xBB\xB1\x00"
-     "\xEB\xBB"
-     "\xB2\x00\xEB\xBB\xB3\x00\xEB\xBB\xB4\x00\xEB\xBB\xB5\x00\xEB\xBB\xB6\x00"
-     "\xEB\xBB"
-     "\xB7\x00\xEB\xBB\xB8\x00\xEB\xBB\xB9\x00\xEB\xBB\xBA\x00\xEB\xBB\xBB\x00"
-     "\xEB\xBB"
-     "\xBC\x00\xEB\xBB\xBD\x00\xEB\xBB\xBE\x00\xEB\xBB\xBF\x00\xEB\xBC\x80\x00"
-     "\xEB\xBC"
-     "\x81\x00\xEB\xBC\x82\x00\xEB\xBC\x83\x00\xEB\xBC\x84\x00\xEB\xBC\x85\x00"
-     "\xEB\xBC"
-     "\x86\x00\xEB\xBC\x87\x00\xEB\xBC\x88\x00\xEB\xBC\x89\x00\xEB\xBC\x8A\x00"
-     "\xEB\xBC"
-     "\x8B\x00\xEB\xBC\x8C\x00\xEB\xBC\x8D\x00\xEB\xBC\x8E\x00\xEB\xBC\x8F\x00"
-     "\xEB\xBC"
-     "\x90\x00\xEB\xBC\x91\x00\xEB\xBC\x92\x00\xEB\xBC\x93\x00\xEB\xBC\x94\x00"
-     "\xEB\xBC"
-     "\x95\x00\xEB\xBC\x96\x00\xEB\xBC\x97\x00\xEB\xBC\x98\x00\xEB\xBC\x99\x00"
-     "\xEB\xBC"
-     "\x9A\x00\xEB\xBC\x9B\x00\xEB\xBC\x9C\x00\xEB\xBC\x9D\x00\xEB\xBC\x9E\x00"
-     "\xEB\xBC"
-     "\x9F\x00\xEB\xBC\xA0\x00\xEB\xBC\xA1\x00\xEB\xBC\xA2\x00\xEB\xBC\xA3\x00"
-     "\xEB\xBC"
-     "\xA4\x00\xEB\xBC\xA5\x00\xEB\xBC\xA6\x00\xEB\xBC\xA7\x00\xEB\xBC\xA8\x00"
-     "\xEB\xBC"
-     "\xA9\x00\xEB\xBC\xAA\x00\xEB\xBC\xAB\x00\xEB\xBC\xAC\x00\xEB\xBC\xAD\x00"
-     "\xEB\xBC"
-     "\xAE\x00\xEB\xBC\xAF\x00\xEB\xBC\xB0\x00\xEB\xBC\xB1\x00\xEB\xBC\xB2\x00"
-     "\xEB\xBC"
-     "\xB3\x00\xEB\xBC\xB4\x00\xEB\xBC\xB5\x00\xEB\xBC\xB6\x00\xEB\xBC\xB7\x00"
-     "\xEB\xBC"
-     "\xB8\x00\xEB\xBC\xB9\x00\xEB\xBC\xBA\x00\xEB\xBC\xBB\x00\xEB\xBC\xBC\x00"
-     "\xEB\xBC"
-     "\xBD\x00\xEB\xBC\xBE\x00\xEB\xBC\xBF\x00\xEB\xBD\x80\x00\xEB\xBD\x81\x00"
-     "\xEB\xBD"
-     "\x82\x00\xEB\xBD\x83\x00\xEB\xBD\x84\x00\xEB\xBD\x85\x00\xEB\xBD\x86\x00"
-     "\xEB\xBD"
-     "\x87\x00\xEB\xBD\x88\x00\xEB\xBD\x89\x00\xEB\xBD\x8A\x00\xEB\xBD\x8B\x00"
-     "\xEB\xBD"
-     "\x8C\x00\xEB\xBD\x8D\x00\xEB\xBD\x8E\x00\xEB\xBD\x8F\x00\xEB\xBD\x90\x00"
-     "\xEB\xBD"
-     "\x91\x00\xEB\xBD\x92\x00\xEB\xBD\x93\x00\xEB\xBD\x94\x00\xEB\xBD\x95\x00"
-     "\xEB\xBD"
-     "\x96\x00\xEB\xBD\x97\x00\xEB\xBD\x98\x00\xEB\xBD\x99\x00\xEB\xBD\x9A\x00"
-     "\xEB\xBD"
-     "\x9B\x00\xEB\xBD\x9C\x00\xEB\xBD\x9D\x00\xEB\xBD\x9E\x00\xEB\xBD\x9F\x00"
-     "\xEB\xBD"
-     "\xA0\x00\xEB\xBD\xA1\x00\xEB\xBD\xA2\x00\xEB\xBD\xA3\x00\xEB\xBD\xA4\x00"
-     "\xEB\xBD"
-     "\xA5\x00\xEB\xBD\xA6\x00\xEB\xBD\xA7\x00\xEB\xBD\xA8\x00\xEB\xBD\xA9\x00"
-     "\xEB\xBD"
-     "\xAA\x00\xEB\xBD\xAB\x00\xEB\xBD\xAC\x00\xEB\xBD\xAD\x00\xEB\xBD\xAE\x00"
-     "\xEB\xBD"
-     "\xAF\x00\xEB\xBD\xB0\x00\xEB\xBD\xB1\x00\xEB\xBD\xB2\x00\xEB\xBD\xB3\x00"
-     "\xEB\xBD"
-     "\xB4\x00\xEB\xBD\xB5\x00\xEB\xBD\xB6\x00\xEB\xBD\xB7\x00\xEB\xBD\xB8\x00"
-     "\xEB\xBD"
-     "\xB9\x00\xEB\xBD\xBA\x00\xEB\xBD\xBB\x00\xEB\xBD\xBC\x00\xEB\xBD\xBD\x00"
-     "\xEB\xBD"
-     "\xBE\x00\xEB\xBD\xBF\x00\xEB\xBE\x80\x00\xEB\xBE\x81\x00\xEB\xBE\x82\x00"
-     "\xEB\xBE"
-     "\x83\x00\xEB\xBE\x84\x00\xEB\xBE\x85\x00\xEB\xBE\x86\x00\xEB\xBE\x87\x00"
-     "\xEB\xBE"
-     "\x88\x00\xEB\xBE\x89\x00\xEB\xBE\x8A\x00\xEB\xBE\x8B\x00\xEB\xBE\x8C\x00"
-     "\xEB\xBE"
-     "\x8D\x00\xEB\xBE\x8E\x00\xEB\xBE\x8F\x00\xEB\xBE\x90\x00\xEB\xBE\x91\x00"
-     "\xEB\xBE"
-     "\x92\x00\xEB\xBE\x93\x00\xEB\xBE\x94\x00\xEB\xBE\x95\x00\xEB\xBE\x96\x00"
-     "\xEB\xBE"
-     "\x97\x00\xEB\xBE\x98\x00\xEB\xBE\x99\x00\xEB\xBE\x9A\x00\xEB\xBE\x9B\x00"
-     "\xEB\xBE"
-     "\x9C\x00\xEB\xBE\x9D\x00\xEB\xBE\x9E\x00\xEB\xBE\x9F\x00\xEB\xBE\xA0\x00"
-     "\xEB\xBE"
-     "\xA1\x00\xEB\xBE\xA2\x00\xEB\xBE\xA3\x00\xEB\xBE\xA4\x00\xEB\xBE\xA5\x00"
-     "\xEB\xBE"
-     "\xA6\x00\xEB\xBE\xA7\x00\xEB\xBE\xA8\x00\xEB\xBE\xA9\x00\xEB\xBE\xAA\x00"
-     "\xEB\xBE"
-     "\xAB\x00\xEB\xBE\xAC\x00\xEB\xBE\xAD\x00\xEB\xBE\xAE\x00\xEB\xBE\xAF\x00"
-     "\xEB\xBE"
-     "\xB0\x00\xEB\xBE\xB1\x00\xEB\xBE\xB2\x00\xEB\xBE\xB3\x00\xEB\xBE\xB4\x00"
-     "\xEB\xBE"
-     "\xB5\x00\xEB\xBE\xB6\x00\xEB\xBE\xB7\x00\xEB\xBE\xB8\x00\xEB\xBE\xB9\x00"
-     "\xEB\xBE"
-     "\xBA\x00\xEB\xBE\xBB\x00\xEB\xBE\xBC\x00\xEB\xBE\xBD\x00\xEB\xBE\xBE\x00"
-     "\xEB\xBE"
-     "\xBF\x00\xEB\xBF\x80\x00\xEB\xBF\x81\x00\xEB\xBF\x82\x00\xEB\xBF\x83\x00"
-     "\xEB\xBF"
-     "\x84\x00\xEB\xBF\x85\x00\xEB\xBF\x86\x00\xEB\xBF\x87\x00\xEB\xBF\x88\x00"
-     "\xEB\xBF"
-     "\x89\x00\xEB\xBF\x8A\x00\xEB\xBF\x8B\x00\xEB\xBF\x8C\x00\xEB\xBF\x8D\x00"
-     "\xEB\xBF"
-     "\x8E\x00\xEB\xBF\x8F\x00\xEB\xBF\x90\x00\xEB\xBF\x91\x00\xEB\xBF\x92\x00"
-     "\xEB\xBF"
-     "\x93\x00\xEB\xBF\x94\x00\xEB\xBF\x95\x00\xEB\xBF\x96\x00\xEB\xBF\x97\x00"
-     "\xEB\xBF"
-     "\x98\x00\xEB\xBF\x99\x00\xEB\xBF\x9A\x00\xEB\xBF\x9B\x00\xEB\xBF\x9C\x00"
-     "\xEB\xBF"
-     "\x9D\x00\xEB\xBF\x9E\x00\xEB\xBF\x9F\x00\xEB\xBF\xA0\x00\xEB\xBF\xA1\x00"
-     "\xEB\xBF"
-     "\xA2\x00\xEB\xBF\xA3\x00\xEB\xBF\xA4\x00\xEB\xBF\xA5\x00\xEB\xBF\xA6\x00"
-     "\xEB\xBF"
-     "\xA7\x00\xEB\xBF\xA8\x00\xEB\xBF\xA9\x00\xEB\xBF\xAA\x00\xEB\xBF\xAB\x00"
-     "\xEB\xBF"
-     "\xAC\x00\xEB\xBF\xAD\x00\xEB\xBF\xAE\x00\xEB\xBF\xAF\x00\xEB\xBF\xB0\x00"
-     "\xEB\xBF"
-     "\xB1\x00\xEB\xBF\xB2\x00\xEB\xBF\xB3\x00\xEB\xBF\xB4\x00\xEB\xBF\xB5\x00"
-     "\xEB\xBF"
-     "\xB6\x00\xEB\xBF\xB7\x00\xEB\xBF\xB8\x00\xEB\xBF\xB9\x00\xEB\xBF\xBA\x00"
-     "\xEB\xBF"
-     "\xBB\x00\xEB\xBF\xBC\x00\xEB\xBF\xBD\x00\xEB\xBF\xBE\x00\xEB\xBF\xBF\x00"
-     "\xEC\x80"
-     "\x80\x00\xEC\x80\x81\x00\xEC\x80\x82\x00\xEC\x80\x83\x00\xEC\x80\x84\x00"
-     "\xEC\x80"
-     "\x85\x00\xEC\x80\x86\x00\xEC\x80\x87\x00\xEC\x80\x88\x00\xEC\x80\x89\x00"
-     "\xEC\x80"
-     "\x8A\x00\xEC\x80\x8B\x00\xEC\x80\x8C\x00\xEC\x80\x8D\x00\xEC\x80\x8E\x00"
-     "\xEC\x80"
-     "\x8F\x00\xEC\x80\x90\x00\xEC\x80\x91\x00\xEC\x80\x92\x00\xEC\x80\x93\x00"
-     "\xEC\x80"
-     "\x94\x00\xEC\x80\x95\x00\xEC\x80\x96\x00\xEC\x80\x97\x00\xEC\x80\x98\x00"
-     "\xEC\x80"
-     "\x99\x00\xEC\x80\x9A\x00\xEC\x80\x9B\x00\xEC\x80\x9C\x00\xEC\x80\x9D\x00"
-     "\xEC\x80"
-     "\x9E\x00\xEC\x80\x9F\x00\xEC\x80\xA0\x00\xEC\x80\xA1\x00\xEC\x80\xA2\x00"
-     "\xEC\x80"
-     "\xA3\x00\xEC\x80\xA4\x00\xEC\x80\xA5\x00\xEC\x80\xA6\x00\xEC\x80\xA7\x00"
-     "\xEC\x80"
-     "\xA8\x00\xEC\x80\xA9\x00\xEC\x80\xAA\x00\xEC\x80\xAB\x00\xEC\x80\xAC\x00"
-     "\xEC\x80"
-     "\xAD\x00\xEC\x80\xAE\x00\xEC\x80\xAF\x00\xEC\x80\xB0\x00\xEC\x80\xB1\x00"
-     "\xEC\x80"
-     "\xB2\x00\xEC\x80\xB3\x00\xEC\x80\xB4\x00\xEC\x80\xB5\x00\xEC\x80\xB6\x00"
-     "\xEC\x80"
-     "\xB7\x00\xEC\x80\xB8\x00\xEC\x80\xB9\x00\xEC\x80\xBA\x00\xEC\x80\xBB\x00"
-     "\xEC\x80"
-     "\xBC\x00\xEC\x80\xBD\x00\xEC\x80\xBE\x00\xEC\x80\xBF\x00\xEC\x81\x80\x00"
-     "\xEC\x81"
-     "\x81\x00\xEC\x81\x82\x00\xEC\x81\x83\x00\xEC\x81\x84\x00\xEC\x81\x85\x00"
-     "\xEC\x81"
-     "\x86\x00\xEC\x81\x87\x00\xEC\x81\x88\x00\xEC\x81\x89\x00\xEC\x81\x8A\x00"
-     "\xEC\x81"
-     "\x8B\x00\xEC\x81\x8C\x00\xEC\x81\x8D\x00\xEC\x81\x8E\x00\xEC\x81\x8F\x00"
-     "\xEC\x81"
-     "\x90\x00\xEC\x81\x91\x00\xEC\x81\x92\x00\xEC\x81\x93\x00\xEC\x81\x94\x00"
-     "\xEC\x81"
-     "\x95\x00\xEC\x81\x96\x00\xEC\x81\x97\x00\xEC\x81\x98\x00\xEC\x81\x99\x00"
-     "\xEC\x81"
-     "\x9A\x00\xEC\x81\x9B\x00\xEC\x81\x9C\x00\xEC\x81\x9D\x00\xEC\x81\x9E\x00"
-     "\xEC\x81"
-     "\x9F\x00\xEC\x81\xA0\x00\xEC\x81\xA1\x00\xEC\x81\xA2\x00\xEC\x81\xA3\x00"
-     "\xEC\x81"
-     "\xA4\x00\xEC\x81\xA5\x00\xEC\x81\xA6\x00\xEC\x81\xA7\x00\xEC\x81\xA8\x00"
-     "\xEC\x81"
-     "\xA9\x00\xEC\x81\xAA\x00\xEC\x81\xAB\x00\xEC\x81\xAC\x00\xEC\x81\xAD\x00"
-     "\xEC\x81"
-     "\xAE\x00\xEC\x81\xAF\x00\xEC\x81\xB0\x00\xEC\x81\xB1\x00\xEC\x81\xB2\x00"
-     "\xEC\x81"
-     "\xB3\x00\xEC\x81\xB4\x00\xEC\x81\xB5\x00\xEC\x81\xB6\x00\xEC\x81\xB7\x00"
-     "\xEC\x81"
-     "\xB8\x00\xEC\x81\xB9\x00\xEC\x81\xBA\x00\xEC\x81\xBB\x00\xEC\x81\xBC\x00"
-     "\xEC\x81"
-     "\xBD\x00\xEC\x81\xBE\x00\xEC\x81\xBF\x00\xEC\x82\x80\x00\xEC\x82\x81\x00"
-     "\xEC\x82"
-     "\x82\x00\xEC\x82\x83\x00\xEC\x82\x84\x00\xEC\x82\x85\x00\xEC\x82\x86\x00"
-     "\xEC\x82"
-     "\x87\x00\xEC\x82\x88\x00\xEC\x82\x89\x00\xEC\x82\x8A\x00\xEC\x82\x8B\x00"
-     "\xEC\x82"
-     "\x8C\x00\xEC\x82\x8D\x00\xEC\x82\x8E\x00\xEC\x82\x8F\x00\xEC\x82\x90\x00"
-     "\xEC\x82"
-     "\x91\x00\xEC\x82\x92\x00\xEC\x82\x93\x00\xEC\x82\x94\x00\xEC\x82\x95\x00"
-     "\xEC\x82"
-     "\x96\x00\xEC\x82\x97\x00\xEC\x82\x98\x00\xEC\x82\x99\x00\xEC\x82\x9A\x00"
-     "\xEC\x82"
-     "\x9B\x00\xEC\x82\x9C\x00\xEC\x82\x9D\x00\xEC\x82\x9E\x00\xEC\x82\x9F\x00"
-     "\xEC\x82"
-     "\xA0\x00\xEC\x82\xA1\x00\xEC\x82\xA2\x00\xEC\x82\xA3\x00\xEC\x82\xA4\x00"
-     "\xEC\x82"
-     "\xA5\x00\xEC\x82\xA6\x00\xEC\x82\xA7\x00\xEC\x82\xA8\x00\xEC\x82\xA9\x00"
-     "\xEC\x82"
-     "\xAA\x00\xEC\x82\xAB\x00\xEC\x82\xAC\x00\xEC\x82\xAD\x00\xEC\x82\xAE\x00"
-     "\xEC\x82"
-     "\xAF\x00\xEC\x82\xB0\x00\xEC\x82\xB1\x00\xEC\x82\xB2\x00\xEC\x82\xB3\x00"
-     "\xEC\x82"
-     "\xB4\x00\xEC\x82\xB5\x00\xEC\x82\xB6\x00\xEC\x82\xB7\x00\xEC\x82\xB8\x00"
-     "\xEC\x82"
-     "\xB9\x00\xEC\x82\xBA\x00\xEC\x82\xBB\x00\xEC\x82\xBC\x00\xEC\x82\xBD\x00"
-     "\xEC\x82"
-     "\xBE\x00\xEC\x82\xBF\x00\xEC\x83\x80\x00\xEC\x83\x81\x00\xEC\x83\x82\x00"
-     "\xEC\x83"
-     "\x83\x00\xEC\x83\x84\x00\xEC\x83\x85\x00\xEC\x83\x86\x00\xEC\x83\x87\x00"
-     "\xEC\x83"
-     "\x88\x00\xEC\x83\x89\x00\xEC\x83\x8A\x00\xEC\x83\x8B\x00\xEC\x83\x8C\x00"
-     "\xEC\x83"
-     "\x8D\x00\xEC\x83\x8E\x00\xEC\x83\x8F\x00\xEC\x83\x90\x00\xEC\x83\x91\x00"
-     "\xEC\x83"
-     "\x92\x00\xEC\x83\x93\x00\xEC\x83\x94\x00\xEC\x83\x95\x00\xEC\x83\x96\x00"
-     "\xEC\x83"
-     "\x97\x00\xEC\x83\x98\x00\xEC\x83\x99\x00\xEC\x83\x9A\x00\xEC\x83\x9B\x00"
-     "\xEC\x83"
-     "\x9C\x00\xEC\x83\x9D\x00\xEC\x83\x9E\x00\xEC\x83\x9F\x00\xEC\x83\xA0\x00"
-     "\xEC\x83"
-     "\xA1\x00\xEC\x83\xA2\x00\xEC\x83\xA3\x00\xEC\x83\xA4\x00\xEC\x83\xA5\x00"
-     "\xEC\x83"
-     "\xA6\x00\xEC\x83\xA7\x00\xEC\x83\xA8\x00\xEC\x83\xA9\x00\xEC\x83\xAA\x00"
-     "\xEC\x83"
-     "\xAB\x00\xEC\x83\xAC\x00\xEC\x83\xAD\x00\xEC\x83\xAE\x00\xEC\x83\xAF\x00"
-     "\xEC\x83"
-     "\xB0\x00\xEC\x83\xB1\x00\xEC\x83\xB2\x00\xEC\x83\xB3\x00\xEC\x83\xB4\x00"
-     "\xEC\x83"
-     "\xB5\x00\xEC\x83\xB6\x00\xEC\x83\xB7\x00\xEC\x83\xB8\x00\xEC\x83\xB9\x00"
-     "\xEC\x83"
-     "\xBA\x00\xEC\x83\xBB\x00\xEC\x83\xBC\x00\xEC\x83\xBD\x00\xEC\x83\xBE\x00"
-     "\xEC\x83"
-     "\xBF\x00\xEC\x84\x80\x00\xEC\x84\x81\x00\xEC\x84\x82\x00\xEC\x84\x83\x00"
-     "\xEC\x84"
-     "\x84\x00\xEC\x84\x85\x00\xEC\x84\x86\x00\xEC\x84\x87\x00\xEC\x84\x88\x00"
-     "\xEC\x84"
-     "\x89\x00\xEC\x84\x8A\x00\xEC\x84\x8B\x00\xEC\x84\x8C\x00\xEC\x84\x8D\x00"
-     "\xEC\x84"
-     "\x8E\x00\xEC\x84\x8F\x00\xEC\x84\x90\x00\xEC\x84\x91\x00\xEC\x84\x92\x00"
-     "\xEC\x84"
-     "\x93\x00\xEC\x84\x94\x00\xEC\x84\x95\x00\xEC\x84\x96\x00\xEC\x84\x97\x00"
-     "\xEC\x84"
-     "\x98\x00\xEC\x84\x99\x00\xEC\x84\x9A\x00\xEC\x84\x9B\x00\xEC\x84\x9C\x00"
-     "\xEC\x84"
-     "\x9D\x00\xEC\x84\x9E\x00\xEC\x84\x9F\x00\xEC\x84\xA0\x00\xEC\x84\xA1\x00"
-     "\xEC\x84"
-     "\xA2\x00\xEC\x84\xA3\x00\xEC\x84\xA4\x00\xEC\x84\xA5\x00\xEC\x84\xA6\x00"
-     "\xEC\x84"
-     "\xA7\x00\xEC\x84\xA8\x00\xEC\x84\xA9\x00\xEC\x84\xAA\x00\xEC\x84\xAB\x00"
-     "\xEC\x84"
-     "\xAC\x00\xEC\x84\xAD\x00\xEC\x84\xAE\x00\xEC\x84\xAF\x00\xEC\x84\xB0\x00"
-     "\xEC\x84"
-     "\xB1\x00\xEC\x84\xB2\x00\xEC\x84\xB3\x00\xEC\x84\xB4\x00\xEC\x84\xB5\x00"
-     "\xEC\x84"
-     "\xB6\x00\xEC\x84\xB7\x00\xEC\x84\xB8\x00\xEC\x84\xB9\x00\xEC\x84\xBA\x00"
-     "\xEC\x84"
-     "\xBB\x00\xEC\x84\xBC\x00\xEC\x84\xBD\x00\xEC\x84\xBE\x00\xEC\x84\xBF\x00"
-     "\xEC\x85"
-     "\x80\x00\xEC\x85\x81\x00\xEC\x85\x82\x00\xEC\x85\x83\x00\xEC\x85\x84\x00"
-     "\xEC\x85"
-     "\x85\x00\xEC\x85\x86\x00\xEC\x85\x87\x00\xEC\x85\x88\x00\xEC\x85\x89\x00"
-     "\xEC\x85"
-     "\x8A\x00\xEC\x85\x8B\x00\xEC\x85\x8C\x00\xEC\x85\x8D\x00\xEC\x85\x8E\x00"
-     "\xEC\x85"
-     "\x8F\x00\xEC\x85\x90\x00\xEC\x85\x91\x00\xEC\x85\x92\x00\xEC\x85\x93\x00"
-     "\xEC\x85"
-     "\x94\x00\xEC\x85\x95\x00\xEC\x85\x96\x00\xEC\x85\x97\x00\xEC\x85\x98\x00"
-     "\xEC\x85"
-     "\x99\x00\xEC\x85\x9A\x00\xEC\x85\x9B\x00\xEC\x85\x9C\x00\xEC\x85\x9D\x00"
-     "\xEC\x85"
-     "\x9E\x00\xEC\x85\x9F\x00\xEC\x85\xA0\x00\xEC\x85\xA1\x00\xEC\x85\xA2\x00"
-     "\xEC\x85"
-     "\xA3\x00\xEC\x85\xA4\x00\xEC\x85\xA5\x00\xEC\x85\xA6\x00\xEC\x85\xA7\x00"
-     "\xEC\x85"
-     "\xA8\x00\xEC\x85\xA9\x00\xEC\x85\xAA\x00\xEC\x85\xAB\x00\xEC\x85\xAC\x00"
-     "\xEC\x85"
-     "\xAD\x00\xEC\x85\xAE\x00\xEC\x85\xAF\x00\xEC\x85\xB0\x00\xEC\x85\xB1\x00"
-     "\xEC\x85"
-     "\xB2\x00\xEC\x85\xB3\x00\xEC\x85\xB4\x00\xEC\x85\xB5\x00\xEC\x85\xB6\x00"
-     "\xEC\x85"
-     "\xB7\x00\xEC\x85\xB8\x00\xEC\x85\xB9\x00\xEC\x85\xBA\x00\xEC\x85\xBB\x00"
-     "\xEC\x85"
-     "\xBC\x00\xEC\x85\xBD\x00\xEC\x85\xBE\x00\xEC\x85\xBF\x00\xEC\x86\x80\x00"
-     "\xEC\x86"
-     "\x81\x00\xEC\x86\x82\x00\xEC\x86\x83\x00\xEC\x86\x84\x00\xEC\x86\x85\x00"
-     "\xEC\x86"
-     "\x86\x00\xEC\x86\x87\x00\xEC\x86\x88\x00\xEC\x86\x89\x00\xEC\x86\x8A\x00"
-     "\xEC\x86"
-     "\x8B\x00\xEC\x86\x8C\x00\xEC\x86\x8D\x00\xEC\x86\x8E\x00\xEC\x86\x8F\x00"
-     "\xEC\x86"
-     "\x90\x00\xEC\x86\x91\x00\xEC\x86\x92\x00\xEC\x86\x93\x00\xEC\x86\x94\x00"
-     "\xEC\x86"
-     "\x95\x00\xEC\x86\x96\x00\xEC\x86\x97\x00\xEC\x86\x98\x00\xEC\x86\x99\x00"
-     "\xEC\x86"
-     "\x9A\x00\xEC\x86\x9B\x00\xEC\x86\x9C\x00\xEC\x86\x9D\x00\xEC\x86\x9E\x00"
-     "\xEC\x86"
-     "\x9F\x00\xEC\x86\xA0\x00\xEC\x86\xA1\x00\xEC\x86\xA2\x00\xEC\x86\xA3\x00"
-     "\xEC\x86"
-     "\xA4\x00\xEC\x86\xA5\x00\xEC\x86\xA6\x00\xEC\x86\xA7\x00\xEC\x86\xA8\x00"
-     "\xEC\x86"
-     "\xA9\x00\xEC\x86\xAA\x00\xEC\x86\xAB\x00\xEC\x86\xAC\x00\xEC\x86\xAD\x00"
-     "\xEC\x86"
-     "\xAE\x00\xEC\x86\xAF\x00\xEC\x86\xB0\x00\xEC\x86\xB1\x00\xEC\x86\xB2\x00"
-     "\xEC\x86"
-     "\xB3\x00\xEC\x86\xB4\x00\xEC\x86\xB5\x00\xEC\x86\xB6\x00\xEC\x86\xB7\x00"
-     "\xEC\x86"
-     "\xB8\x00\xEC\x86\xB9\x00\xEC\x86\xBA\x00\xEC\x86\xBB\x00\xEC\x86\xBC\x00"
-     "\xEC\x86"
-     "\xBD\x00\xEC\x86\xBE\x00\xEC\x86\xBF\x00\xEC\x87\x80\x00\xEC\x87\x81\x00"
-     "\xEC\x87"
-     "\x82\x00\xEC\x87\x83\x00\xEC\x87\x84\x00\xEC\x87\x85\x00\xEC\x87\x86\x00"
-     "\xEC\x87"
-     "\x87\x00\xEC\x87\x88\x00\xEC\x87\x89\x00\xEC\x87\x8A\x00\xEC\x87\x8B\x00"
-     "\xEC\x87"
-     "\x8C\x00\xEC\x87\x8D\x00\xEC\x87\x8E\x00\xEC\x87\x8F\x00\xEC\x87\x90\x00"
-     "\xEC\x87"
-     "\x91\x00\xEC\x87\x92\x00\xEC\x87\x93\x00\xEC\x87\x94\x00\xEC\x87\x95\x00"
-     "\xEC\x87"
-     "\x96\x00\xEC\x87\x97\x00\xEC\x87\x98\x00\xEC\x87\x99\x00\xEC\x87\x9A\x00"
-     "\xEC\x87"
-     "\x9B\x00\xEC\x87\x9C\x00\xEC\x87\x9D\x00\xEC\x87\x9E\x00\xEC\x87\x9F\x00"
-     "\xEC\x87"
-     "\xA0\x00\xEC\x87\xA1\x00\xEC\x87\xA2\x00\xEC\x87\xA3\x00\xEC\x87\xA4\x00"
-     "\xEC\x87"
-     "\xA5\x00\xEC\x87\xA6\x00\xEC\x87\xA7\x00\xEC\x87\xA8\x00\xEC\x87\xA9\x00"
-     "\xEC\x87"
-     "\xAA\x00\xEC\x87\xAB\x00\xEC\x87\xAC\x00\xEC\x87\xAD\x00\xEC\x87\xAE\x00"
-     "\xEC\x87"
-     "\xAF\x00\xEC\x87\xB0\x00\xEC\x87\xB1\x00\xEC\x87\xB2\x00\xEC\x87\xB3\x00"
-     "\xEC\x87"
-     "\xB4\x00\xEC\x87\xB5\x00\xEC\x87\xB6\x00\xEC\x87\xB7\x00\xEC\x87\xB8\x00"
-     "\xEC\x87"
-     "\xB9\x00\xEC\x87\xBA\x00\xEC\x87\xBB\x00\xEC\x87\xBC\x00\xEC\x87\xBD\x00"
-     "\xEC\x87"
-     "\xBE\x00\xEC\x87\xBF\x00\xEC\x88\x80\x00\xEC\x88\x81\x00\xEC\x88\x82\x00"
-     "\xEC\x88"
-     "\x83\x00\xEC\x88\x84\x00\xEC\x88\x85\x00\xEC\x88\x86\x00\xEC\x88\x87\x00"
-     "\xEC\x88"
-     "\x88\x00\xEC\x88\x89\x00\xEC\x88\x8A\x00\xEC\x88\x8B\x00\xEC\x88\x8C\x00"
-     "\xEC\x88"
-     "\x8D\x00\xEC\x88\x8E\x00\xEC\x88\x8F\x00\xEC\x88\x90\x00\xEC\x88\x91\x00"
-     "\xEC\x88"
-     "\x92\x00\xEC\x88\x93\x00\xEC\x88\x94\x00\xEC\x88\x95\x00\xEC\x88\x96\x00"
-     "\xEC\x88"
-     "\x97\x00\xEC\x88\x98\x00\xEC\x88\x99\x00\xEC\x88\x9A\x00\xEC\x88\x9B\x00"
-     "\xEC\x88"
-     "\x9C\x00\xEC\x88\x9D\x00\xEC\x88\x9E\x00\xEC\x88\x9F\x00\xEC\x88\xA0\x00"
-     "\xEC\x88"
-     "\xA1\x00\xEC\x88\xA2\x00\xEC\x88\xA3\x00\xEC\x88\xA4\x00\xEC\x88\xA5\x00"
-     "\xEC\x88"
-     "\xA6\x00\xEC\x88\xA7\x00\xEC\x88\xA8\x00\xEC\x88\xA9\x00\xEC\x88\xAA\x00"
-     "\xEC\x88"
-     "\xAB\x00\xEC\x88\xAC\x00\xEC\x88\xAD\x00\xEC\x88\xAE\x00\xEC\x88\xAF\x00"
-     "\xEC\x88"
-     "\xB0\x00\xEC\x88\xB1\x00\xEC\x88\xB2\x00\xEC\x88\xB3\x00\xEC\x88\xB4\x00"
-     "\xEC\x88"
-     "\xB5\x00\xEC\x88\xB6\x00\xEC\x88\xB7\x00\xEC\x88\xB8\x00\xEC\x88\xB9\x00"
-     "\xEC\x88"
-     "\xBA\x00\xEC\x88\xBB\x00\xEC\x88\xBC\x00\xEC\x88\xBD\x00\xEC\x88\xBE\x00"
-     "\xEC\x88"
-     "\xBF\x00\xEC\x89\x80\x00\xEC\x89\x81\x00\xEC\x89\x82\x00\xEC\x89\x83\x00"
-     "\xEC\x89"
-     "\x84\x00\xEC\x89\x85\x00\xEC\x89\x86\x00\xEC\x89\x87\x00\xEC\x89\x88\x00"
-     "\xEC\x89"
-     "\x89\x00\xEC\x89\x8A\x00\xEC\x89\x8B\x00\xEC\x89\x8C\x00\xEC\x89\x8D\x00"
-     "\xEC\x89"
-     "\x8E\x00\xEC\x89\x8F\x00\xEC\x89\x90\x00\xEC\x89\x91\x00\xEC\x89\x92\x00"
-     "\xEC\x89"
-     "\x93\x00\xEC\x89\x94\x00\xEC\x89\x95\x00\xEC\x89\x96\x00\xEC\x89\x97\x00"
-     "\xEC\x89"
-     "\x98\x00\xEC\x89\x99\x00\xEC\x89\x9A\x00\xEC\x89\x9B\x00\xEC\x89\x9C\x00"
-     "\xEC\x89"
-     "\x9D\x00\xEC\x89\x9E\x00\xEC\x89\x9F\x00\xEC\x89\xA0\x00\xEC\x89\xA1\x00"
-     "\xEC\x89"
-     "\xA2\x00\xEC\x89\xA3\x00\xEC\x89\xA4\x00\xEC\x89\xA5\x00\xEC\x89\xA6\x00"
-     "\xEC\x89"
-     "\xA7\x00\xEC\x89\xA8\x00\xEC\x89\xA9\x00\xEC\x89\xAA\x00\xEC\x89\xAB\x00"
-     "\xEC\x89"
-     "\xAC\x00\xEC\x89\xAD\x00\xEC\x89\xAE\x00\xEC\x89\xAF\x00\xEC\x89\xB0\x00"
-     "\xEC\x89"
-     "\xB1\x00\xEC\x89\xB2\x00\xEC\x89\xB3\x00\xEC\x89\xB4\x00\xEC\x89\xB5\x00"
-     "\xEC\x89"
-     "\xB6\x00\xEC\x89\xB7\x00\xEC\x89\xB8\x00\xEC\x89\xB9\x00\xEC\x89\xBA\x00"
-     "\xEC\x89"
-     "\xBB\x00\xEC\x89\xBC\x00\xEC\x89\xBD\x00\xEC\x89\xBE\x00\xEC\x89\xBF\x00"
-     "\xEC\x8A"
-     "\x80\x00\xEC\x8A\x81\x00\xEC\x8A\x82\x00\xEC\x8A\x83\x00\xEC\x8A\x84\x00"
-     "\xEC\x8A"
-     "\x85\x00\xEC\x8A\x86\x00\xEC\x8A\x87\x00\xEC\x8A\x88\x00\xEC\x8A\x89\x00"
-     "\xEC\x8A"
-     "\x8A\x00\xEC\x8A\x8B\x00\xEC\x8A\x8C\x00\xEC\x8A\x8D\x00\xEC\x8A\x8E\x00"
-     "\xEC\x8A"
-     "\x8F\x00\xEC\x8A\x90\x00\xEC\x8A\x91\x00\xEC\x8A\x92\x00\xEC\x8A\x93\x00"
-     "\xEC\x8A"
-     "\x94\x00\xEC\x8A\x95\x00\xEC\x8A\x96\x00\xEC\x8A\x97\x00\xEC\x8A\x98\x00"
-     "\xEC\x8A"
-     "\x99\x00\xEC\x8A\x9A\x00\xEC\x8A\x9B\x00\xEC\x8A\x9C\x00\xEC\x8A\x9D\x00"
-     "\xEC\x8A"
-     "\x9E\x00\xEC\x8A\x9F\x00\xEC\x8A\xA0\x00\xEC\x8A\xA1\x00\xEC\x8A\xA2\x00"
-     "\xEC\x8A"
-     "\xA3\x00\xEC\x8A\xA4\x00\xEC\x8A\xA5\x00\xEC\x8A\xA6\x00\xEC\x8A\xA7\x00"
-     "\xEC\x8A"
-     "\xA8\x00\xEC\x8A\xA9\x00\xEC\x8A\xAA\x00\xEC\x8A\xAB\x00\xEC\x8A\xAC\x00"
-     "\xEC\x8A"
-     "\xAD\x00\xEC\x8A\xAE\x00\xEC\x8A\xAF\x00\xEC\x8A\xB0\x00\xEC\x8A\xB1\x00"
-     "\xEC\x8A"
-     "\xB2\x00\xEC\x8A\xB3\x00\xEC\x8A\xB4\x00\xEC\x8A\xB5\x00\xEC\x8A\xB6\x00"
-     "\xEC\x8A"
-     "\xB7\x00\xEC\x8A\xB8\x00\xEC\x8A\xB9\x00\xEC\x8A\xBA\x00\xEC\x8A\xBB\x00"
-     "\xEC\x8A"
-     "\xBC\x00\xEC\x8A\xBD\x00\xEC\x8A\xBE\x00\xEC\x8A\xBF\x00\xEC\x8B\x80\x00"
-     "\xEC\x8B"
-     "\x81\x00\xEC\x8B\x82\x00\xEC\x8B\x83\x00\xEC\x8B\x84\x00\xEC\x8B\x85\x00"
-     "\xEC\x8B"
-     "\x86\x00\xEC\x8B\x87\x00\xEC\x8B\x88\x00\xEC\x8B\x89\x00\xEC\x8B\x8A\x00"
-     "\xEC\x8B"
-     "\x8B\x00\xEC\x8B\x8C\x00\xEC\x8B\x8D\x00\xEC\x8B\x8E\x00\xEC\x8B\x8F\x00"
-     "\xEC\x8B"
-     "\x90\x00\xEC\x8B\x91\x00\xEC\x8B\x92\x00\xEC\x8B\x93\x00\xEC\x8B\x94\x00"
-     "\xEC\x8B"
-     "\x95\x00\xEC\x8B\x96\x00\xEC\x8B\x97\x00\xEC\x8B\x98\x00\xEC\x8B\x99\x00"
-     "\xEC\x8B"
-     "\x9A\x00\xEC\x8B\x9B\x00\xEC\x8B\x9C\x00\xEC\x8B\x9D\x00\xEC\x8B\x9E\x00"
-     "\xEC\x8B"
-     "\x9F\x00\xEC\x8B\xA0\x00\xEC\x8B\xA1\x00\xEC\x8B\xA2\x00\xEC\x8B\xA3\x00"
-     "\xEC\x8B"
-     "\xA4\x00\xEC\x8B\xA5\x00\xEC\x8B\xA6\x00\xEC\x8B\xA7\x00\xEC\x8B\xA8\x00"
-     "\xEC\x8B"
-     "\xA9\x00\xEC\x8B\xAA\x00\xEC\x8B\xAB\x00\xEC\x8B\xAC\x00\xEC\x8B\xAD\x00"
-     "\xEC\x8B"
-     "\xAE\x00\xEC\x8B\xAF\x00\xEC\x8B\xB0\x00\xEC\x8B\xB1\x00\xEC\x8B\xB2\x00"
-     "\xEC\x8B"
-     "\xB3\x00\xEC\x8B\xB4\x00\xEC\x8B\xB5\x00\xEC\x8B\xB6\x00\xEC\x8B\xB7\x00"
-     "\xEC\x8B"
-     "\xB8\x00\xEC\x8B\xB9\x00\xEC\x8B\xBA\x00\xEC\x8B\xBB\x00\xEC\x8B\xBC\x00"
-     "\xEC\x8B"
-     "\xBD\x00\xEC\x8B\xBE\x00\xEC\x8B\xBF\x00\xEC\x8C\x80\x00\xEC\x8C\x81\x00"
-     "\xEC\x8C"
-     "\x82\x00\xEC\x8C\x83\x00\xEC\x8C\x84\x00\xEC\x8C\x85\x00\xEC\x8C\x86\x00"
-     "\xEC\x8C"
-     "\x87\x00\xEC\x8C\x88\x00\xEC\x8C\x89\x00\xEC\x8C\x8A\x00\xEC\x8C\x8B\x00"
-     "\xEC\x8C"
-     "\x8C\x00\xEC\x8C\x8D\x00\xEC\x8C\x8E\x00\xEC\x8C\x8F\x00\xEC\x8C\x90\x00"
-     "\xEC\x8C"
-     "\x91\x00\xEC\x8C\x92\x00\xEC\x8C\x93\x00\xEC\x8C\x94\x00\xEC\x8C\x95\x00"
-     "\xEC\x8C"
-     "\x96\x00\xEC\x8C\x97\x00\xEC\x8C\x98\x00\xEC\x8C\x99\x00\xEC\x8C\x9A\x00"
-     "\xEC\x8C"
-     "\x9B\x00\xEC\x8C\x9C\x00\xEC\x8C\x9D\x00\xEC\x8C\x9E\x00\xEC\x8C\x9F\x00"
-     "\xEC\x8C"
-     "\xA0\x00\xEC\x8C\xA1\x00\xEC\x8C\xA2\x00\xEC\x8C\xA3\x00\xEC\x8C\xA4\x00"
-     "\xEC\x8C"
-     "\xA5\x00\xEC\x8C\xA6\x00\xEC\x8C\xA7\x00\xEC\x8C\xA8\x00\xEC\x8C\xA9\x00"
-     "\xEC\x8C"
-     "\xAA\x00\xEC\x8C\xAB\x00\xEC\x8C\xAC\x00\xEC\x8C\xAD\x00\xEC\x8C\xAE\x00"
-     "\xEC\x8C"
-     "\xAF\x00\xEC\x8C\xB0\x00\xEC\x8C\xB1\x00\xEC\x8C\xB2\x00\xEC\x8C\xB3\x00"
-     "\xEC\x8C"
-     "\xB4\x00\xEC\x8C\xB5\x00\xEC\x8C\xB6\x00\xEC\x8C\xB7\x00\xEC\x8C\xB8\x00"
-     "\xEC\x8C"
-     "\xB9\x00\xEC\x8C\xBA\x00\xEC\x8C\xBB\x00\xEC\x8C\xBC\x00\xEC\x8C\xBD\x00"
-     "\xEC\x8C"
-     "\xBE\x00\xEC\x8C\xBF\x00\xEC\x8D\x80\x00\xEC\x8D\x81\x00\xEC\x8D\x82\x00"
-     "\xEC\x8D"
-     "\x83\x00\xEC\x8D\x84\x00\xEC\x8D\x85\x00\xEC\x8D\x86\x00\xEC\x8D\x87\x00"
-     "\xEC\x8D"
-     "\x88\x00\xEC\x8D\x89\x00\xEC\x8D\x8A\x00\xEC\x8D\x8B\x00\xEC\x8D\x8C\x00"
-     "\xEC\x8D"
-     "\x8D\x00\xEC\x8D\x8E\x00\xEC\x8D\x8F\x00\xEC\x8D\x90\x00\xEC\x8D\x91\x00"
-     "\xEC\x8D"
-     "\x92\x00\xEC\x8D\x93\x00\xEC\x8D\x94\x00\xEC\x8D\x95\x00\xEC\x8D\x96\x00"
-     "\xEC\x8D"
-     "\x97\x00\xEC\x8D\x98\x00\xEC\x8D\x99\x00\xEC\x8D\x9A\x00\xEC\x8D\x9B\x00"
-     "\xEC\x8D"
-     "\x9C\x00\xEC\x8D\x9D\x00\xEC\x8D\x9E\x00\xEC\x8D\x9F\x00\xEC\x8D\xA0\x00"
-     "\xEC\x8D"
-     "\xA1\x00\xEC\x8D\xA2\x00\xEC\x8D\xA3\x00\xEC\x8D\xA4\x00\xEC\x8D\xA5\x00"
-     "\xEC\x8D"
-     "\xA6\x00\xEC\x8D\xA7\x00\xEC\x8D\xA8\x00\xEC\x8D\xA9\x00\xEC\x8D\xAA\x00"
-     "\xEC\x8D"
-     "\xAB\x00\xEC\x8D\xAC\x00\xEC\x8D\xAD\x00\xEC\x8D\xAE\x00\xEC\x8D\xAF\x00"
-     "\xEC\x8D"
-     "\xB0\x00\xEC\x8D\xB1\x00\xEC\x8D\xB2\x00\xEC\x8D\xB3\x00\xEC\x8D\xB4\x00"
-     "\xEC\x8D"
-     "\xB5\x00\xEC\x8D\xB6\x00\xEC\x8D\xB7\x00\xEC\x8D\xB8\x00\xEC\x8D\xB9\x00"
-     "\xEC\x8D"
-     "\xBA\x00\xEC\x8D\xBB\x00\xEC\x8D\xBC\x00\xEC\x8D\xBD\x00\xEC\x8D\xBE\x00"
-     "\xEC\x8D"
-     "\xBF\x00\xEC\x8E\x80\x00\xEC\x8E\x81\x00\xEC\x8E\x82\x00\xEC\x8E\x83\x00"
-     "\xEC\x8E"
-     "\x84\x00\xEC\x8E\x85\x00\xEC\x8E\x86\x00\xEC\x8E\x87\x00\xEC\x8E\x88\x00"
-     "\xEC\x8E"
-     "\x89\x00\xEC\x8E\x8A\x00\xEC\x8E\x8B\x00\xEC\x8E\x8C\x00\xEC\x8E\x8D\x00"
-     "\xEC\x8E"
-     "\x8E\x00\xEC\x8E\x8F\x00\xEC\x8E\x90\x00\xEC\x8E\x91\x00\xEC\x8E\x92\x00"
-     "\xEC\x8E"
-     "\x93\x00\xEC\x8E\x94\x00\xEC\x8E\x95\x00\xEC\x8E\x96\x00\xEC\x8E\x97\x00"
-     "\xEC\x8E"
-     "\x98\x00\xEC\x8E\x99\x00\xEC\x8E\x9A\x00\xEC\x8E\x9B\x00\xEC\x8E\x9C\x00"
-     "\xEC\x8E"
-     "\x9D\x00\xEC\x8E\x9E\x00\xEC\x8E\x9F\x00\xEC\x8E\xA0\x00\xEC\x8E\xA1\x00"
-     "\xEC\x8E"
-     "\xA2\x00\xEC\x8E\xA3\x00\xEC\x8E\xA4\x00\xEC\x8E\xA5\x00\xEC\x8E\xA6\x00"
-     "\xEC\x8E"
-     "\xA7\x00\xEC\x8E\xA8\x00\xEC\x8E\xA9\x00\xEC\x8E\xAA\x00\xEC\x8E\xAB\x00"
-     "\xEC\x8E"
-     "\xAC\x00\xEC\x8E\xAD\x00\xEC\x8E\xAE\x00\xEC\x8E\xAF\x00\xEC\x8E\xB0\x00"
-     "\xEC\x8E"
-     "\xB1\x00\xEC\x8E\xB2\x00\xEC\x8E\xB3\x00\xEC\x8E\xB4\x00\xEC\x8E\xB5\x00"
-     "\xEC\x8E"
-     "\xB6\x00\xEC\x8E\xB7\x00\xEC\x8E\xB8\x00\xEC\x8E\xB9\x00\xEC\x8E\xBA\x00"
-     "\xEC\x8E"
-     "\xBB\x00\xEC\x8E\xBC\x00\xEC\x8E\xBD\x00\xEC\x8E\xBE\x00\xEC\x8E\xBF\x00"
-     "\xEC\x8F"
-     "\x80\x00\xEC\x8F\x81\x00\xEC\x8F\x82\x00\xEC\x8F\x83\x00\xEC\x8F\x84\x00"
-     "\xEC\x8F"
-     "\x85\x00\xEC\x8F\x86\x00\xEC\x8F\x87\x00\xEC\x8F\x88\x00\xEC\x8F\x89\x00"
-     "\xEC\x8F"
-     "\x8A\x00\xEC\x8F\x8B\x00\xEC\x8F\x8C\x00\xEC\x8F\x8D\x00\xEC\x8F\x8E\x00"
-     "\xEC\x8F"
-     "\x8F\x00\xEC\x8F\x90\x00\xEC\x8F\x91\x00\xEC\x8F\x92\x00\xEC\x8F\x93\x00"
-     "\xEC\x8F"
-     "\x94\x00\xEC\x8F\x95\x00\xEC\x8F\x96\x00\xEC\x8F\x97\x00\xEC\x8F\x98\x00"
-     "\xEC\x8F"
-     "\x99\x00\xEC\x8F\x9A\x00\xEC\x8F\x9B\x00\xEC\x8F\x9C\x00\xEC\x8F\x9D\x00"
-     "\xEC\x8F"
-     "\x9E\x00\xEC\x8F\x9F\x00\xEC\x8F\xA0\x00\xEC\x8F\xA1\x00\xEC\x8F\xA2\x00"
-     "\xEC\x8F"
-     "\xA3\x00\xEC\x8F\xA4\x00\xEC\x8F\xA5\x00\xEC\x8F\xA6\x00\xEC\x8F\xA7\x00"
-     "\xEC\x8F"
-     "\xA8\x00\xEC\x8F\xA9\x00\xEC\x8F\xAA\x00\xEC\x8F\xAB\x00\xEC\x8F\xAC\x00"
-     "\xEC\x8F"
-     "\xAD\x00\xEC\x8F\xAE\x00\xEC\x8F\xAF\x00\xEC\x8F\xB0\x00\xEC\x8F\xB1\x00"
-     "\xEC\x8F"
-     "\xB2\x00\xEC\x8F\xB3\x00\xEC\x8F\xB4\x00\xEC\x8F\xB5\x00\xEC\x8F\xB6\x00"
-     "\xEC\x8F"
-     "\xB7\x00\xEC\x8F\xB8\x00\xEC\x8F\xB9\x00\xEC\x8F\xBA\x00\xEC\x8F\xBB\x00"
-     "\xEC\x8F"
-     "\xBC\x00\xEC\x8F\xBD\x00\xEC\x8F\xBE\x00\xEC\x8F\xBF\x00\xEC\x90\x80\x00"
-     "\xEC\x90"
-     "\x81\x00\xEC\x90\x82\x00\xEC\x90\x83\x00\xEC\x90\x84\x00\xEC\x90\x85\x00"
-     "\xEC\x90"
-     "\x86\x00\xEC\x90\x87\x00\xEC\x90\x88\x00\xEC\x90\x89\x00\xEC\x90\x8A\x00"
-     "\xEC\x90"
-     "\x8B\x00\xEC\x90\x8C\x00\xEC\x90\x8D\x00\xEC\x90\x8E\x00\xEC\x90\x8F\x00"
-     "\xEC\x90"
-     "\x90\x00\xEC\x90\x91\x00\xEC\x90\x92\x00\xEC\x90\x93\x00\xEC\x90\x94\x00"
-     "\xEC\x90"
-     "\x95\x00\xEC\x90\x96\x00\xEC\x90\x97\x00\xEC\x90\x98\x00\xEC\x90\x99\x00"
-     "\xEC\x90"
-     "\x9A\x00\xEC\x90\x9B\x00\xEC\x90\x9C\x00\xEC\x90\x9D\x00\xEC\x90\x9E\x00"
-     "\xEC\x90"
-     "\x9F\x00\xEC\x90\xA0\x00\xEC\x90\xA1\x00\xEC\x90\xA2\x00\xEC\x90\xA3\x00"
-     "\xEC\x90"
-     "\xA4\x00\xEC\x90\xA5\x00\xEC\x90\xA6\x00\xEC\x90\xA7\x00\xEC\x90\xA8\x00"
-     "\xEC\x90"
-     "\xA9\x00\xEC\x90\xAA\x00\xEC\x90\xAB\x00\xEC\x90\xAC\x00\xEC\x90\xAD\x00"
-     "\xEC\x90"
-     "\xAE\x00\xEC\x90\xAF\x00\xEC\x90\xB0\x00\xEC\x90\xB1\x00\xEC\x90\xB2\x00"
-     "\xEC\x90"
-     "\xB3\x00\xEC\x90\xB4\x00\xEC\x90\xB5\x00\xEC\x90\xB6\x00\xEC\x90\xB7\x00"
-     "\xEC\x90"
-     "\xB8\x00\xEC\x90\xB9\x00\xEC\x90\xBA\x00\xEC\x90\xBB\x00\xEC\x90\xBC\x00"
-     "\xEC\x90"
-     "\xBD\x00\xEC\x90\xBE\x00\xEC\x90\xBF\x00\xEC\x91\x80\x00\xEC\x91\x81\x00"
-     "\xEC\x91"
-     "\x82\x00\xEC\x91\x83\x00\xEC\x91\x84\x00\xEC\x91\x85\x00\xEC\x91\x86\x00"
-     "\xEC\x91"
-     "\x87\x00\xEC\x91\x88\x00\xEC\x91\x89\x00\xEC\x91\x8A\x00\xEC\x91\x8B\x00"
-     "\xEC\x91"
-     "\x8C\x00\xEC\x91\x8D\x00\xEC\x91\x8E\x00\xEC\x91\x8F\x00\xEC\x91\x90\x00"
-     "\xEC\x91"
-     "\x91\x00\xEC\x91\x92\x00\xEC\x91\x93\x00\xEC\x91\x94\x00\xEC\x91\x95\x00"
-     "\xEC\x91"
-     "\x96\x00\xEC\x91\x97\x00\xEC\x91\x98\x00\xEC\x91\x99\x00\xEC\x91\x9A\x00"
-     "\xEC\x91"
-     "\x9B\x00\xEC\x91\x9C\x00\xEC\x91\x9D\x00\xEC\x91\x9E\x00\xEC\x91\x9F\x00"
-     "\xEC\x91"
-     "\xA0\x00\xEC\x91\xA1\x00\xEC\x91\xA2\x00\xEC\x91\xA3\x00\xEC\x91\xA4\x00"
-     "\xEC\x91"
-     "\xA5\x00\xEC\x91\xA6\x00\xEC\x91\xA7\x00\xEC\x91\xA8\x00\xEC\x91\xA9\x00"
-     "\xEC\x91"
-     "\xAA\x00\xEC\x91\xAB\x00\xEC\x91\xAC\x00\xEC\x91\xAD\x00\xEC\x91\xAE\x00"
-     "\xEC\x91"
-     "\xAF\x00\xEC\x91\xB0\x00\xEC\x91\xB1\x00\xEC\x91\xB2\x00\xEC\x91\xB3\x00"
-     "\xEC\x91"
-     "\xB4\x00\xEC\x91\xB5\x00\xEC\x91\xB6\x00\xEC\x91\xB7\x00\xEC\x91\xB8\x00"
-     "\xEC\x91"
-     "\xB9\x00\xEC\x91\xBA\x00\xEC\x91\xBB\x00\xEC\x91\xBC\x00\xEC\x91\xBD\x00"
-     "\xEC\x91"
-     "\xBE\x00\xEC\x91\xBF\x00\xEC\x92\x80\x00\xEC\x92\x81\x00\xEC\x92\x82\x00"
-     "\xEC\x92"
-     "\x83\x00\xEC\x92\x84\x00\xEC\x92\x85\x00\xEC\x92\x86\x00\xEC\x92\x87\x00"
-     "\xEC\x92"
-     "\x88\x00\xEC\x92\x89\x00\xEC\x92\x8A\x00\xEC\x92\x8B\x00\xEC\x92\x8C\x00"
-     "\xEC\x92"
-     "\x8D\x00\xEC\x92\x8E\x00\xEC\x92\x8F\x00\xEC\x92\x90\x00\xEC\x92\x91\x00"
-     "\xEC\x92"
-     "\x92\x00\xEC\x92\x93\x00\xEC\x92\x94\x00\xEC\x92\x95\x00\xEC\x92\x96\x00"
-     "\xEC\x92"
-     "\x97\x00\xEC\x92\x98\x00\xEC\x92\x99\x00\xEC\x92\x9A\x00\xEC\x92\x9B\x00"
-     "\xEC\x92"
-     "\x9C\x00\xEC\x92\x9D\x00\xEC\x92\x9E\x00\xEC\x92\x9F\x00\xEC\x92\xA0\x00"
-     "\xEC\x92"
-     "\xA1\x00\xEC\x92\xA2\x00\xEC\x92\xA3\x00\xEC\x92\xA4\x00\xEC\x92\xA5\x00"
-     "\xEC\x92"
-     "\xA6\x00\xEC\x92\xA7\x00\xEC\x92\xA8\x00\xEC\x92\xA9\x00\xEC\x92\xAA\x00"
-     "\xEC\x92"
-     "\xAB\x00\xEC\x92\xAC\x00\xEC\x92\xAD\x00\xEC\x92\xAE\x00\xEC\x92\xAF\x00"
-     "\xEC\x92"
-     "\xB0\x00\xEC\x92\xB1\x00\xEC\x92\xB2\x00\xEC\x92\xB3\x00\xEC\x92\xB4\x00"
-     "\xEC\x92"
-     "\xB5\x00\xEC\x92\xB6\x00\xEC\x92\xB7\x00\xEC\x92\xB8\x00\xEC\x92\xB9\x00"
-     "\xEC\x92"
-     "\xBA\x00\xEC\x92\xBB\x00\xEC\x92\xBC\x00\xEC\x92\xBD\x00\xEC\x92\xBE\x00"
-     "\xEC\x92"
-     "\xBF\x00\xEC\x93\x80\x00\xEC\x93\x81\x00\xEC\x93\x82\x00\xEC\x93\x83\x00"
-     "\xEC\x93"
-     "\x84\x00\xEC\x93\x85\x00\xEC\x93\x86\x00\xEC\x93\x87\x00\xEC\x93\x88\x00"
-     "\xEC\x93"
-     "\x89\x00\xEC\x93\x8A\x00\xEC\x93\x8B\x00\xEC\x93\x8C\x00\xEC\x93\x8D\x00"
-     "\xEC\x93"
-     "\x8E\x00\xEC\x93\x8F\x00\xEC\x93\x90\x00\xEC\x93\x91\x00\xEC\x93\x92\x00"
-     "\xEC\x93"
-     "\x93\x00\xEC\x93\x94\x00\xEC\x93\x95\x00\xEC\x93\x96\x00\xEC\x93\x97\x00"
-     "\xEC\x93"
-     "\x98\x00\xEC\x93\x99\x00\xEC\x93\x9A\x00\xEC\x93\x9B\x00\xEC\x93\x9C\x00"
-     "\xEC\x93"
-     "\x9D\x00\xEC\x93\x9E\x00\xEC\x93\x9F\x00\xEC\x93\xA0\x00\xEC\x93\xA1\x00"
-     "\xEC\x93"
-     "\xA2\x00\xEC\x93\xA3\x00\xEC\x93\xA4\x00\xEC\x93\xA5\x00\xEC\x93\xA6\x00"
-     "\xEC\x93"
-     "\xA7\x00\xEC\x93\xA8\x00\xEC\x93\xA9\x00\xEC\x93\xAA\x00\xEC\x93\xAB\x00"
-     "\xEC\x93"
-     "\xAC\x00\xEC\x93\xAD\x00\xEC\x93\xAE\x00\xEC\x93\xAF\x00\xEC\x93\xB0\x00"
-     "\xEC\x93"
-     "\xB1\x00\xEC\x93\xB2\x00\xEC\x93\xB3\x00\xEC\x93\xB4\x00\xEC\x93\xB5\x00"
-     "\xEC\x93"
-     "\xB6\x00\xEC\x93\xB7\x00\xEC\x93\xB8\x00\xEC\x93\xB9\x00\xEC\x93\xBA\x00"
-     "\xEC\x93"
-     "\xBB\x00\xEC\x93\xBC\x00\xEC\x93\xBD\x00\xEC\x93\xBE\x00\xEC\x93\xBF\x00"
-     "\xEC\x94"
-     "\x80\x00\xEC\x94\x81\x00\xEC\x94\x82\x00\xEC\x94\x83\x00\xEC\x94\x84\x00"
-     "\xEC\x94"
-     "\x85\x00\xEC\x94\x86\x00\xEC\x94\x87\x00\xEC\x94\x88\x00\xEC\x94\x89\x00"
-     "\xEC\x94"
-     "\x8A\x00\xEC\x94\x8B\x00\xEC\x94\x8C\x00\xEC\x94\x8D\x00\xEC\x94\x8E\x00"
-     "\xEC\x94"
-     "\x8F\x00\xEC\x94\x90\x00\xEC\x94\x91\x00\xEC\x94\x92\x00\xEC\x94\x93\x00"
-     "\xEC\x94"
-     "\x94\x00\xEC\x94\x95\x00\xEC\x94\x96\x00\xEC\x94\x97\x00\xEC\x94\x98\x00"
-     "\xEC\x94"
-     "\x99\x00\xEC\x94\x9A\x00\xEC\x94\x9B\x00\xEC\x94\x9C\x00\xEC\x94\x9D\x00"
-     "\xEC\x94"
-     "\x9E\x00\xEC\x94\x9F\x00\xEC\x94\xA0\x00\xEC\x94\xA1\x00\xEC\x94\xA2\x00"
-     "\xEC\x94"
-     "\xA3\x00\xEC\x94\xA4\x00\xEC\x94\xA5\x00\xEC\x94\xA6\x00\xEC\x94\xA7\x00"
-     "\xEC\x94"
-     "\xA8\x00\xEC\x94\xA9\x00\xEC\x94\xAA\x00\xEC\x94\xAB\x00\xEC\x94\xAC\x00"
-     "\xEC\x94"
-     "\xAD\x00\xEC\x94\xAE\x00\xEC\x94\xAF\x00\xEC\x94\xB0\x00\xEC\x94\xB1\x00"
-     "\xEC\x94"
-     "\xB2\x00\xEC\x94\xB3\x00\xEC\x94\xB4\x00\xEC\x94\xB5\x00\xEC\x94\xB6\x00"
-     "\xEC\x94"
-     "\xB7\x00\xEC\x94\xB8\x00\xEC\x94\xB9\x00\xEC\x94\xBA\x00\xEC\x94\xBB\x00"
-     "\xEC\x94"
-     "\xBC\x00\xEC\x94\xBD\x00\xEC\x94\xBE\x00\xEC\x94\xBF\x00\xEC\x95\x80\x00"
-     "\xEC\x95"
-     "\x81\x00\xEC\x95\x82\x00\xEC\x95\x83\x00\xEC\x95\x84\x00\xEC\x95\x85\x00"
-     "\xEC\x95"
-     "\x86\x00\xEC\x95\x87\x00\xEC\x95\x88\x00\xEC\x95\x89\x00\xEC\x95\x8A\x00"
-     "\xEC\x95"
-     "\x8B\x00\xEC\x95\x8C\x00\xEC\x95\x8D\x00\xEC\x95\x8E\x00\xEC\x95\x8F\x00"
-     "\xEC\x95"
-     "\x90\x00\xEC\x95\x91\x00\xEC\x95\x92\x00\xEC\x95\x93\x00\xEC\x95\x94\x00"
-     "\xEC\x95"
-     "\x95\x00\xEC\x95\x96\x00\xEC\x95\x97\x00\xEC\x95\x98\x00\xEC\x95\x99\x00"
-     "\xEC\x95"
-     "\x9A\x00\xEC\x95\x9B\x00\xEC\x95\x9C\x00\xEC\x95\x9D\x00\xEC\x95\x9E\x00"
-     "\xEC\x95"
-     "\x9F\x00\xEC\x95\xA0\x00\xEC\x95\xA1\x00\xEC\x95\xA2\x00\xEC\x95\xA3\x00"
-     "\xEC\x95"
-     "\xA4\x00\xEC\x95\xA5\x00\xEC\x95\xA6\x00\xEC\x95\xA7\x00\xEC\x95\xA8\x00"
-     "\xEC\x95"
-     "\xA9\x00\xEC\x95\xAA\x00\xEC\x95\xAB\x00\xEC\x95\xAC\x00\xEC\x95\xAD\x00"
-     "\xEC\x95"
-     "\xAE\x00\xEC\x95\xAF\x00\xEC\x95\xB0\x00\xEC\x95\xB1\x00\xEC\x95\xB2\x00"
-     "\xEC\x95"
-     "\xB3\x00\xEC\x95\xB4\x00\xEC\x95\xB5\x00\xEC\x95\xB6\x00\xEC\x95\xB7\x00"
-     "\xEC\x95"
-     "\xB8\x00\xEC\x95\xB9\x00\xEC\x95\xBA\x00\xEC\x95\xBB\x00\xEC\x95\xBC\x00"
-     "\xEC\x95"
-     "\xBD\x00\xEC\x95\xBE\x00\xEC\x95\xBF\x00\xEC\x96\x80\x00\xEC\x96\x81\x00"
-     "\xEC\x96"
-     "\x82\x00\xEC\x96\x83\x00\xEC\x96\x84\x00\xEC\x96\x85\x00\xEC\x96\x86\x00"
-     "\xEC\x96"
-     "\x87\x00\xEC\x96\x88\x00\xEC\x96\x89\x00\xEC\x96\x8A\x00\xEC\x96\x8B\x00"
-     "\xEC\x96"
-     "\x8C\x00\xEC\x96\x8D\x00\xEC\x96\x8E\x00\xEC\x96\x8F\x00\xEC\x96\x90\x00"
-     "\xEC\x96"
-     "\x91\x00\xEC\x96\x92\x00\xEC\x96\x93\x00\xEC\x96\x94\x00\xEC\x96\x95\x00"
-     "\xEC\x96"
-     "\x96\x00\xEC\x96\x97\x00\xEC\x96\x98\x00\xEC\x96\x99\x00\xEC\x96\x9A\x00"
-     "\xEC\x96"
-     "\x9B\x00\xEC\x96\x9C\x00\xEC\x96\x9D\x00\xEC\x96\x9E\x00\xEC\x96\x9F\x00"
-     "\xEC\x96"
-     "\xA0\x00\xEC\x96\xA1\x00\xEC\x96\xA2\x00\xEC\x96\xA3\x00\xEC\x96\xA4\x00"
-     "\xEC\x96"
-     "\xA5\x00\xEC\x96\xA6\x00\xEC\x96\xA7\x00\xEC\x96\xA8\x00\xEC\x96\xA9\x00"
-     "\xEC\x96"
-     "\xAA\x00\xEC\x96\xAB\x00\xEC\x96\xAC\x00\xEC\x96\xAD\x00\xEC\x96\xAE\x00"
-     "\xEC\x96"
-     "\xAF\x00\xEC\x96\xB0\x00\xEC\x96\xB1\x00\xEC\x96\xB2\x00\xEC\x96\xB3\x00"
-     "\xEC\x96"
-     "\xB4\x00\xEC\x96\xB5\x00\xEC\x96\xB6\x00\xEC\x96\xB7\x00\xEC\x96\xB8\x00"
-     "\xEC\x96"
-     "\xB9\x00\xEC\x96\xBA\x00\xEC\x96\xBB\x00\xEC\x96\xBC\x00\xEC\x96\xBD\x00"
-     "\xEC\x96"
-     "\xBE\x00\xEC\x96\xBF\x00\xEC\x97\x80\x00\xEC\x97\x81\x00\xEC\x97\x82\x00"
-     "\xEC\x97"
-     "\x83\x00\xEC\x97\x84\x00\xEC\x97\x85\x00\xEC\x97\x86\x00\xEC\x97\x87\x00"
-     "\xEC\x97"
-     "\x88\x00\xEC\x97\x89\x00\xEC\x97\x8A\x00\xEC\x97\x8B\x00\xEC\x97\x8C\x00"
-     "\xEC\x97"
-     "\x8D\x00\xEC\x97\x8E\x00\xEC\x97\x8F\x00\xEC\x97\x90\x00\xEC\x97\x91\x00"
-     "\xEC\x97"
-     "\x92\x00\xEC\x97\x93\x00\xEC\x97\x94\x00\xEC\x97\x95\x00\xEC\x97\x96\x00"
-     "\xEC\x97"
-     "\x97\x00\xEC\x97\x98\x00\xEC\x97\x99\x00\xEC\x97\x9A\x00\xEC\x97\x9B\x00"
-     "\xEC\x97"
-     "\x9C\x00\xEC\x97\x9D\x00\xEC\x97\x9E\x00\xEC\x97\x9F\x00\xEC\x97\xA0\x00"
-     "\xEC\x97"
-     "\xA1\x00\xEC\x97\xA2\x00\xEC\x97\xA3\x00\xEC\x97\xA4\x00\xEC\x97\xA5\x00"
-     "\xEC\x97"
-     "\xA6\x00\xEC\x97\xA7\x00\xEC\x97\xA8\x00\xEC\x97\xA9\x00\xEC\x97\xAA\x00"
-     "\xEC\x97"
-     "\xAB\x00\xEC\x97\xAC\x00\xEC\x97\xAD\x00\xEC\x97\xAE\x00\xEC\x97\xAF\x00"
-     "\xEC\x97"
-     "\xB0\x00\xEC\x97\xB1\x00\xEC\x97\xB2\x00\xEC\x97\xB3\x00\xEC\x97\xB4\x00"
-     "\xEC\x97"
-     "\xB5\x00\xEC\x97\xB6\x00\xEC\x97\xB7\x00\xEC\x97\xB8\x00\xEC\x97\xB9\x00"
-     "\xEC\x97"
-     "\xBA\x00\xEC\x97\xBB\x00\xEC\x97\xBC\x00\xEC\x97\xBD\x00\xEC\x97\xBE\x00"
-     "\xEC\x97"
-     "\xBF\x00\xEC\x98\x80\x00\xEC\x98\x81\x00\xEC\x98\x82\x00\xEC\x98\x83\x00"
-     "\xEC\x98"
-     "\x84\x00\xEC\x98\x85\x00\xEC\x98\x86\x00\xEC\x98\x87\x00\xEC\x98\x88\x00"
-     "\xEC\x98"
-     "\x89\x00\xEC\x98\x8A\x00\xEC\x98\x8B\x00\xEC\x98\x8C\x00\xEC\x98\x8D\x00"
-     "\xEC\x98"
-     "\x8E\x00\xEC\x98\x8F\x00\xEC\x98\x90\x00\xEC\x98\x91\x00\xEC\x98\x92\x00"
-     "\xEC\x98"
-     "\x93\x00\xEC\x98\x94\x00\xEC\x98\x95\x00\xEC\x98\x96\x00\xEC\x98\x97\x00"
-     "\xEC\x98"
-     "\x98\x00\xEC\x98\x99\x00\xEC\x98\x9A\x00\xEC\x98\x9B\x00\xEC\x98\x9C\x00"
-     "\xEC\x98"
-     "\x9D\x00\xEC\x98\x9E\x00\xEC\x98\x9F\x00\xEC\x98\xA0\x00\xEC\x98\xA1\x00"
-     "\xEC\x98"
-     "\xA2\x00\xEC\x98\xA3\x00\xEC\x98\xA4\x00\xEC\x98\xA5\x00\xEC\x98\xA6\x00"
-     "\xEC\x98"
-     "\xA7\x00\xEC\x98\xA8\x00\xEC\x98\xA9\x00\xEC\x98\xAA\x00\xEC\x98\xAB\x00"
-     "\xEC\x98"
-     "\xAC\x00\xEC\x98\xAD\x00\xEC\x98\xAE\x00\xEC\x98\xAF\x00\xEC\x98\xB0\x00"
-     "\xEC\x98"
-     "\xB1\x00\xEC\x98\xB2\x00\xEC\x98\xB3\x00\xEC\x98\xB4\x00\xEC\x98\xB5\x00"
-     "\xEC\x98"
-     "\xB6\x00\xEC\x98\xB7\x00\xEC\x98\xB8\x00\xEC\x98\xB9\x00\xEC\x98\xBA\x00"
-     "\xEC\x98"
-     "\xBB\x00\xEC\x98\xBC\x00\xEC\x98\xBD\x00\xEC\x98\xBE\x00\xEC\x98\xBF\x00"
-     "\xEC\x99"
-     "\x80\x00\xEC\x99\x81\x00\xEC\x99\x82\x00\xEC\x99\x83\x00\xEC\x99\x84\x00"
-     "\xEC\x99"
-     "\x85\x00\xEC\x99\x86\x00\xEC\x99\x87\x00\xEC\x99\x88\x00\xEC\x99\x89\x00"
-     "\xEC\x99"
-     "\x8A\x00\xEC\x99\x8B\x00\xEC\x99\x8C\x00\xEC\x99\x8D\x00\xEC\x99\x8E\x00"
-     "\xEC\x99"
-     "\x8F\x00\xEC\x99\x90\x00\xEC\x99\x91\x00\xEC\x99\x92\x00\xEC\x99\x93\x00"
-     "\xEC\x99"
-     "\x94\x00\xEC\x99\x95\x00\xEC\x99\x96\x00\xEC\x99\x97\x00\xEC\x99\x98\x00"
-     "\xEC\x99"
-     "\x99\x00\xEC\x99\x9A\x00\xEC\x99\x9B\x00\xEC\x99\x9C\x00\xEC\x99\x9D\x00"
-     "\xEC\x99"
-     "\x9E\x00\xEC\x99\x9F\x00\xEC\x99\xA0\x00\xEC\x99\xA1\x00\xEC\x99\xA2\x00"
-     "\xEC\x99"
-     "\xA3\x00\xEC\x99\xA4\x00\xEC\x99\xA5\x00\xEC\x99\xA6\x00\xEC\x99\xA7\x00"
-     "\xEC\x99"
-     "\xA8\x00\xEC\x99\xA9\x00\xEC\x99\xAA\x00\xEC\x99\xAB\x00\xEC\x99\xAC\x00"
-     "\xEC\x99"
-     "\xAD\x00\xEC\x99\xAE\x00\xEC\x99\xAF\x00\xEC\x99\xB0\x00\xEC\x99\xB1\x00"
-     "\xEC\x99"
-     "\xB2\x00\xEC\x99\xB3\x00\xEC\x99\xB4\x00\xEC\x99\xB5\x00\xEC\x99\xB6\x00"
-     "\xEC\x99"
-     "\xB7\x00\xEC\x99\xB8\x00\xEC\x99\xB9\x00\xEC\x99\xBA\x00\xEC\x99\xBB\x00"
-     "\xEC\x99"
-     "\xBC\x00\xEC\x99\xBD\x00\xEC\x99\xBE\x00\xEC\x99\xBF\x00\xEC\x9A\x80\x00"
-     "\xEC\x9A"
-     "\x81\x00\xEC\x9A\x82\x00\xEC\x9A\x83\x00\xEC\x9A\x84\x00\xEC\x9A\x85\x00"
-     "\xEC\x9A"
-     "\x86\x00\xEC\x9A\x87\x00\xEC\x9A\x88\x00\xEC\x9A\x89\x00\xEC\x9A\x8A\x00"
-     "\xEC\x9A"
-     "\x8B\x00\xEC\x9A\x8C\x00\xEC\x9A\x8D\x00\xEC\x9A\x8E\x00\xEC\x9A\x8F\x00"
-     "\xEC\x9A"
-     "\x90\x00\xEC\x9A\x91\x00\xEC\x9A\x92\x00\xEC\x9A\x93\x00\xEC\x9A\x94\x00"
-     "\xEC\x9A"
-     "\x95\x00\xEC\x9A\x96\x00\xEC\x9A\x97\x00\xEC\x9A\x98\x00\xEC\x9A\x99\x00"
-     "\xEC\x9A"
-     "\x9A\x00\xEC\x9A\x9B\x00\xEC\x9A\x9C\x00\xEC\x9A\x9D\x00\xEC\x9A\x9E\x00"
-     "\xEC\x9A"
-     "\x9F\x00\xEC\x9A\xA0\x00\xEC\x9A\xA1\x00\xEC\x9A\xA2\x00\xEC\x9A\xA3\x00"
-     "\xEC\x9A"
-     "\xA4\x00\xEC\x9A\xA5\x00\xEC\x9A\xA6\x00\xEC\x9A\xA7\x00\xEC\x9A\xA8\x00"
-     "\xEC\x9A"
-     "\xA9\x00\xEC\x9A\xAA\x00\xEC\x9A\xAB\x00\xEC\x9A\xAC\x00\xEC\x9A\xAD\x00"
-     "\xEC\x9A"
-     "\xAE\x00\xEC\x9A\xAF\x00\xEC\x9A\xB0\x00\xEC\x9A\xB1\x00\xEC\x9A\xB2\x00"
-     "\xEC\x9A"
-     "\xB3\x00\xEC\x9A\xB4\x00\xEC\x9A\xB5\x00\xEC\x9A\xB6\x00\xEC\x9A\xB7\x00"
-     "\xEC\x9A"
-     "\xB8\x00\xEC\x9A\xB9\x00\xEC\x9A\xBA\x00\xEC\x9A\xBB\x00\xEC\x9A\xBC\x00"
-     "\xEC\x9A"
-     "\xBD\x00\xEC\x9A\xBE\x00\xEC\x9A\xBF\x00\xEC\x9B\x80\x00\xEC\x9B\x81\x00"
-     "\xEC\x9B"
-     "\x82\x00\xEC\x9B\x83\x00\xEC\x9B\x84\x00\xEC\x9B\x85\x00\xEC\x9B\x86\x00"
-     "\xEC\x9B"
-     "\x87\x00\xEC\x9B\x88\x00\xEC\x9B\x89\x00\xEC\x9B\x8A\x00\xEC\x9B\x8B\x00"
-     "\xEC\x9B"
-     "\x8C\x00\xEC\x9B\x8D\x00\xEC\x9B\x8E\x00\xEC\x9B\x8F\x00\xEC\x9B\x90\x00"
-     "\xEC\x9B"
-     "\x91\x00\xEC\x9B\x92\x00\xEC\x9B\x93\x00\xEC\x9B\x94\x00\xEC\x9B\x95\x00"
-     "\xEC\x9B"
-     "\x96\x00\xEC\x9B\x97\x00\xEC\x9B\x98\x00\xEC\x9B\x99\x00\xEC\x9B\x9A\x00"
-     "\xEC\x9B"
-     "\x9B\x00\xEC\x9B\x9C\x00\xEC\x9B\x9D\x00\xEC\x9B\x9E\x00\xEC\x9B\x9F\x00"
-     "\xEC\x9B"
-     "\xA0\x00\xEC\x9B\xA1\x00\xEC\x9B\xA2\x00\xEC\x9B\xA3\x00\xEC\x9B\xA4\x00"
-     "\xEC\x9B"
-     "\xA5\x00\xEC\x9B\xA6\x00\xEC\x9B\xA7\x00\xEC\x9B\xA8\x00\xEC\x9B\xA9\x00"
-     "\xEC\x9B"
-     "\xAA\x00\xEC\x9B\xAB\x00\xEC\x9B\xAC\x00\xEC\x9B\xAD\x00\xEC\x9B\xAE\x00"
-     "\xEC\x9B"
-     "\xAF\x00\xEC\x9B\xB0\x00\xEC\x9B\xB1\x00\xEC\x9B\xB2\x00\xEC\x9B\xB3\x00"
-     "\xEC\x9B"
-     "\xB4\x00\xEC\x9B\xB5\x00\xEC\x9B\xB6\x00\xEC\x9B\xB7\x00\xEC\x9B\xB8\x00"
-     "\xEC\x9B"
-     "\xB9\x00\xEC\x9B\xBA\x00\xEC\x9B\xBB\x00\xEC\x9B\xBC\x00\xEC\x9B\xBD\x00"
-     "\xEC\x9B"
-     "\xBE\x00\xEC\x9B\xBF\x00\xEC\x9C\x80\x00\xEC\x9C\x81\x00\xEC\x9C\x82\x00"
-     "\xEC\x9C"
-     "\x83\x00\xEC\x9C\x84\x00\xEC\x9C\x85\x00\xEC\x9C\x86\x00\xEC\x9C\x87\x00"
-     "\xEC\x9C"
-     "\x88\x00\xEC\x9C\x89\x00\xEC\x9C\x8A\x00\xEC\x9C\x8B\x00\xEC\x9C\x8C\x00"
-     "\xEC\x9C"
-     "\x8D\x00\xEC\x9C\x8E\x00\xEC\x9C\x8F\x00\xEC\x9C\x90\x00\xEC\x9C\x91\x00"
-     "\xEC\x9C"
-     "\x92\x00\xEC\x9C\x93\x00\xEC\x9C\x94\x00\xEC\x9C\x95\x00\xEC\x9C\x96\x00"
-     "\xEC\x9C"
-     "\x97\x00\xEC\x9C\x98\x00\xEC\x9C\x99\x00\xEC\x9C\x9A\x00\xEC\x9C\x9B\x00"
-     "\xEC\x9C"
-     "\x9C\x00\xEC\x9C\x9D\x00\xEC\x9C\x9E\x00\xEC\x9C\x9F\x00\xEC\x9C\xA0\x00"
-     "\xEC\x9C"
-     "\xA1\x00\xEC\x9C\xA2\x00\xEC\x9C\xA3\x00\xEC\x9C\xA4\x00\xEC\x9C\xA5\x00"
-     "\xEC\x9C"
-     "\xA6\x00\xEC\x9C\xA7\x00\xEC\x9C\xA8\x00\xEC\x9C\xA9\x00\xEC\x9C\xAA\x00"
-     "\xEC\x9C"
-     "\xAB\x00\xEC\x9C\xAC\x00\xEC\x9C\xAD\x00\xEC\x9C\xAE\x00\xEC\x9C\xAF\x00"
-     "\xEC\x9C"
-     "\xB0\x00\xEC\x9C\xB1\x00\xEC\x9C\xB2\x00\xEC\x9C\xB3\x00\xEC\x9C\xB4\x00"
-     "\xEC\x9C"
-     "\xB5\x00\xEC\x9C\xB6\x00\xEC\x9C\xB7\x00\xEC\x9C\xB8\x00\xEC\x9C\xB9\x00"
-     "\xEC\x9C"
-     "\xBA\x00\xEC\x9C\xBB\x00\xEC\x9C\xBC\x00\xEC\x9C\xBD\x00\xEC\x9C\xBE\x00"
-     "\xEC\x9C"
-     "\xBF\x00\xEC\x9D\x80\x00\xEC\x9D\x81\x00\xEC\x9D\x82\x00\xEC\x9D\x83\x00"
-     "\xEC\x9D"
-     "\x84\x00\xEC\x9D\x85\x00\xEC\x9D\x86\x00\xEC\x9D\x87\x00\xEC\x9D\x88\x00"
-     "\xEC\x9D"
-     "\x89\x00\xEC\x9D\x8A\x00\xEC\x9D\x8B\x00\xEC\x9D\x8C\x00\xEC\x9D\x8D\x00"
-     "\xEC\x9D"
-     "\x8E\x00\xEC\x9D\x8F\x00\xEC\x9D\x90\x00\xEC\x9D\x91\x00\xEC\x9D\x92\x00"
-     "\xEC\x9D"
-     "\x93\x00\xEC\x9D\x94\x00\xEC\x9D\x95\x00\xEC\x9D\x96\x00\xEC\x9D\x97\x00"
-     "\xEC\x9D"
-     "\x98\x00\xEC\x9D\x99\x00\xEC\x9D\x9A\x00\xEC\x9D\x9B\x00\xEC\x9D\x9C\x00"
-     "\xEC\x9D"
-     "\x9D\x00\xEC\x9D\x9E\x00\xEC\x9D\x9F\x00\xEC\x9D\xA0\x00\xEC\x9D\xA1\x00"
-     "\xEC\x9D"
-     "\xA2\x00\xEC\x9D\xA3\x00\xEC\x9D\xA4\x00\xEC\x9D\xA5\x00\xEC\x9D\xA6\x00"
-     "\xEC\x9D"
-     "\xA7\x00\xEC\x9D\xA8\x00\xEC\x9D\xA9\x00\xEC\x9D\xAA\x00\xEC\x9D\xAB\x00"
-     "\xEC\x9D"
-     "\xAC\x00\xEC\x9D\xAD\x00\xEC\x9D\xAE\x00\xEC\x9D\xAF\x00\xEC\x9D\xB0\x00"
-     "\xEC\x9D"
-     "\xB1\x00\xEC\x9D\xB2\x00\xEC\x9D\xB3\x00\xEC\x9D\xB4\x00\xEC\x9D\xB5\x00"
-     "\xEC\x9D"
-     "\xB6\x00\xEC\x9D\xB7\x00\xEC\x9D\xB8\x00\xEC\x9D\xB9\x00\xEC\x9D\xBA\x00"
-     "\xEC\x9D"
-     "\xBB\x00\xEC\x9D\xBC\x00\xEC\x9D\xBD\x00\xEC\x9D\xBE\x00\xEC\x9D\xBF\x00"
-     "\xEC\x9E"
-     "\x80\x00\xEC\x9E\x81\x00\xEC\x9E\x82\x00\xEC\x9E\x83\x00\xEC\x9E\x84\x00"
-     "\xEC\x9E"
-     "\x85\x00\xEC\x9E\x86\x00\xEC\x9E\x87\x00\xEC\x9E\x88\x00\xEC\x9E\x89\x00"
-     "\xEC\x9E"
-     "\x8A\x00\xEC\x9E\x8B\x00\xEC\x9E\x8C\x00\xEC\x9E\x8D\x00\xEC\x9E\x8E\x00"
-     "\xEC\x9E"
-     "\x8F\x00\xEC\x9E\x90\x00\xEC\x9E\x91\x00\xEC\x9E\x92\x00\xEC\x9E\x93\x00"
-     "\xEC\x9E"
-     "\x94\x00\xEC\x9E\x95\x00\xEC\x9E\x96\x00\xEC\x9E\x97\x00\xEC\x9E\x98\x00"
-     "\xEC\x9E"
-     "\x99\x00\xEC\x9E\x9A\x00\xEC\x9E\x9B\x00\xEC\x9E\x9C\x00\xEC\x9E\x9D\x00"
-     "\xEC\x9E"
-     "\x9E\x00\xEC\x9E\x9F\x00\xEC\x9E\xA0\x00\xEC\x9E\xA1\x00\xEC\x9E\xA2\x00"
-     "\xEC\x9E"
-     "\xA3\x00\xEC\x9E\xA4\x00\xEC\x9E\xA5\x00\xEC\x9E\xA6\x00\xEC\x9E\xA7\x00"
-     "\xEC\x9E"
-     "\xA8\x00\xEC\x9E\xA9\x00\xEC\x9E\xAA\x00\xEC\x9E\xAB\x00\xEC\x9E\xAC\x00"
-     "\xEC\x9E"
-     "\xAD\x00\xEC\x9E\xAE\x00\xEC\x9E\xAF\x00\xEC\x9E\xB0\x00\xEC\x9E\xB1\x00"
-     "\xEC\x9E"
-     "\xB2\x00\xEC\x9E\xB3\x00\xEC\x9E\xB4\x00\xEC\x9E\xB5\x00\xEC\x9E\xB6\x00"
-     "\xEC\x9E"
-     "\xB7\x00\xEC\x9E\xB8\x00\xEC\x9E\xB9\x00\xEC\x9E\xBA\x00\xEC\x9E\xBB\x00"
-     "\xEC\x9E"
-     "\xBC\x00\xEC\x9E\xBD\x00\xEC\x9E\xBE\x00\xEC\x9E\xBF\x00\xEC\x9F\x80\x00"
-     "\xEC\x9F"
-     "\x81\x00\xEC\x9F\x82\x00\xEC\x9F\x83\x00\xEC\x9F\x84\x00\xEC\x9F\x85\x00"
-     "\xEC\x9F"
-     "\x86\x00\xEC\x9F\x87\x00\xEC\x9F\x88\x00\xEC\x9F\x89\x00\xEC\x9F\x8A\x00"
-     "\xEC\x9F"
-     "\x8B\x00\xEC\x9F\x8C\x00\xEC\x9F\x8D\x00\xEC\x9F\x8E\x00\xEC\x9F\x8F\x00"
-     "\xEC\x9F"
-     "\x90\x00\xEC\x9F\x91\x00\xEC\x9F\x92\x00\xEC\x9F\x93\x00\xEC\x9F\x94\x00"
-     "\xEC\x9F"
-     "\x95\x00\xEC\x9F\x96\x00\xEC\x9F\x97\x00\xEC\x9F\x98\x00\xEC\x9F\x99\x00"
-     "\xEC\x9F"
-     "\x9A\x00\xEC\x9F\x9B\x00\xEC\x9F\x9C\x00\xEC\x9F\x9D\x00\xEC\x9F\x9E\x00"
-     "\xEC\x9F"
-     "\x9F\x00\xEC\x9F\xA0\x00\xEC\x9F\xA1\x00\xEC\x9F\xA2\x00\xEC\x9F\xA3\x00"
-     "\xEC\x9F"
-     "\xA4\x00\xEC\x9F\xA5\x00\xEC\x9F\xA6\x00\xEC\x9F\xA7\x00\xEC\x9F\xA8\x00"
-     "\xEC\x9F"
-     "\xA9\x00\xEC\x9F\xAA\x00\xEC\x9F\xAB\x00\xEC\x9F\xAC\x00\xEC\x9F\xAD\x00"
-     "\xEC\x9F"
-     "\xAE\x00\xEC\x9F\xAF\x00\xEC\x9F\xB0\x00\xEC\x9F\xB1\x00\xEC\x9F\xB2\x00"
-     "\xEC\x9F"
-     "\xB3\x00\xEC\x9F\xB4\x00\xEC\x9F\xB5\x00\xEC\x9F\xB6\x00\xEC\x9F\xB7\x00"
-     "\xEC\x9F"
-     "\xB8\x00\xEC\x9F\xB9\x00\xEC\x9F\xBA\x00\xEC\x9F\xBB\x00\xEC\x9F\xBC\x00"
-     "\xEC\x9F"
-     "\xBD\x00\xEC\x9F\xBE\x00\xEC\x9F\xBF\x00\xEC\xA0\x80\x00\xEC\xA0\x81\x00"
-     "\xEC\xA0"
-     "\x82\x00\xEC\xA0\x83\x00\xEC\xA0\x84\x00\xEC\xA0\x85\x00\xEC\xA0\x86\x00"
-     "\xEC\xA0"
-     "\x87\x00\xEC\xA0\x88\x00\xEC\xA0\x89\x00\xEC\xA0\x8A\x00\xEC\xA0\x8B\x00"
-     "\xEC\xA0"
-     "\x8C\x00\xEC\xA0\x8D\x00\xEC\xA0\x8E\x00\xEC\xA0\x8F\x00\xEC\xA0\x90\x00"
-     "\xEC\xA0"
-     "\x91\x00\xEC\xA0\x92\x00\xEC\xA0\x93\x00\xEC\xA0\x94\x00\xEC\xA0\x95\x00"
-     "\xEC\xA0"
-     "\x96\x00\xEC\xA0\x97\x00\xEC\xA0\x98\x00\xEC\xA0\x99\x00\xEC\xA0\x9A\x00"
-     "\xEC\xA0"
-     "\x9B\x00\xEC\xA0\x9C\x00\xEC\xA0\x9D\x00\xEC\xA0\x9E\x00\xEC\xA0\x9F\x00"
-     "\xEC\xA0"
-     "\xA0\x00\xEC\xA0\xA1\x00\xEC\xA0\xA2\x00\xEC\xA0\xA3\x00\xEC\xA0\xA4\x00"
-     "\xEC\xA0"
-     "\xA5\x00\xEC\xA0\xA6\x00\xEC\xA0\xA7\x00\xEC\xA0\xA8\x00\xEC\xA0\xA9\x00"
-     "\xEC\xA0"
-     "\xAA\x00\xEC\xA0\xAB\x00\xEC\xA0\xAC\x00\xEC\xA0\xAD\x00\xEC\xA0\xAE\x00"
-     "\xEC\xA0"
-     "\xAF\x00\xEC\xA0\xB0\x00\xEC\xA0\xB1\x00\xEC\xA0\xB2\x00\xEC\xA0\xB3\x00"
-     "\xEC\xA0"
-     "\xB4\x00\xEC\xA0\xB5\x00\xEC\xA0\xB6\x00\xEC\xA0\xB7\x00\xEC\xA0\xB8\x00"
-     "\xEC\xA0"
-     "\xB9\x00\xEC\xA0\xBA\x00\xEC\xA0\xBB\x00\xEC\xA0\xBC\x00\xEC\xA0\xBD\x00"
-     "\xEC\xA0"
-     "\xBE\x00\xEC\xA0\xBF\x00\xEC\xA1\x80\x00\xEC\xA1\x81\x00\xEC\xA1\x82\x00"
-     "\xEC\xA1"
-     "\x83\x00\xEC\xA1\x84\x00\xEC\xA1\x85\x00\xEC\xA1\x86\x00\xEC\xA1\x87\x00"
-     "\xEC\xA1"
-     "\x88\x00\xEC\xA1\x89\x00\xEC\xA1\x8A\x00\xEC\xA1\x8B\x00\xEC\xA1\x8C\x00"
-     "\xEC\xA1"
-     "\x8D\x00\xEC\xA1\x8E\x00\xEC\xA1\x8F\x00\xEC\xA1\x90\x00\xEC\xA1\x91\x00"
-     "\xEC\xA1"
-     "\x92\x00\xEC\xA1\x93\x00\xEC\xA1\x94\x00\xEC\xA1\x95\x00\xEC\xA1\x96\x00"
-     "\xEC\xA1"
-     "\x97\x00\xEC\xA1\x98\x00\xEC\xA1\x99\x00\xEC\xA1\x9A\x00\xEC\xA1\x9B\x00"
-     "\xEC\xA1"
-     "\x9C\x00\xEC\xA1\x9D\x00\xEC\xA1\x9E\x00\xEC\xA1\x9F\x00\xEC\xA1\xA0\x00"
-     "\xEC\xA1"
-     "\xA1\x00\xEC\xA1\xA2\x00\xEC\xA1\xA3\x00\xEC\xA1\xA4\x00\xEC\xA1\xA5\x00"
-     "\xEC\xA1"
-     "\xA6\x00\xEC\xA1\xA7\x00\xEC\xA1\xA8\x00\xEC\xA1\xA9\x00\xEC\xA1\xAA\x00"
-     "\xEC\xA1"
-     "\xAB\x00\xEC\xA1\xAC\x00\xEC\xA1\xAD\x00\xEC\xA1\xAE\x00\xEC\xA1\xAF\x00"
-     "\xEC\xA1"
-     "\xB0\x00\xEC\xA1\xB1\x00\xEC\xA1\xB2\x00\xEC\xA1\xB3\x00\xEC\xA1\xB4\x00"
-     "\xEC\xA1"
-     "\xB5\x00\xEC\xA1\xB6\x00\xEC\xA1\xB7\x00\xEC\xA1\xB8\x00\xEC\xA1\xB9\x00"
-     "\xEC\xA1"
-     "\xBA\x00\xEC\xA1\xBB\x00\xEC\xA1\xBC\x00\xEC\xA1\xBD\x00\xEC\xA1\xBE\x00"
-     "\xEC\xA1"
-     "\xBF\x00\xEC\xA2\x80\x00\xEC\xA2\x81\x00\xEC\xA2\x82\x00\xEC\xA2\x83\x00"
-     "\xEC\xA2"
-     "\x84\x00\xEC\xA2\x85\x00\xEC\xA2\x86\x00\xEC\xA2\x87\x00\xEC\xA2\x88\x00"
-     "\xEC\xA2"
-     "\x89\x00\xEC\xA2\x8A\x00\xEC\xA2\x8B\x00\xEC\xA2\x8C\x00\xEC\xA2\x8D\x00"
-     "\xEC\xA2"
-     "\x8E\x00\xEC\xA2\x8F\x00\xEC\xA2\x90\x00\xEC\xA2\x91\x00\xEC\xA2\x92\x00"
-     "\xEC\xA2"
-     "\x93\x00\xEC\xA2\x94\x00\xEC\xA2\x95\x00\xEC\xA2\x96\x00\xEC\xA2\x97\x00"
-     "\xEC\xA2"
-     "\x98\x00\xEC\xA2\x99\x00\xEC\xA2\x9A\x00\xEC\xA2\x9B\x00\xEC\xA2\x9C\x00"
-     "\xEC\xA2"
-     "\x9D\x00\xEC\xA2\x9E\x00\xEC\xA2\x9F\x00\xEC\xA2\xA0\x00\xEC\xA2\xA1\x00"
-     "\xEC\xA2"
-     "\xA2\x00\xEC\xA2\xA3\x00\xEC\xA2\xA4\x00\xEC\xA2\xA5\x00\xEC\xA2\xA6\x00"
-     "\xEC\xA2"
-     "\xA7\x00\xEC\xA2\xA8\x00\xEC\xA2\xA9\x00\xEC\xA2\xAA\x00\xEC\xA2\xAB\x00"
-     "\xEC\xA2"
-     "\xAC\x00\xEC\xA2\xAD\x00\xEC\xA2\xAE\x00\xEC\xA2\xAF\x00\xEC\xA2\xB0\x00"
-     "\xEC\xA2"
-     "\xB1\x00\xEC\xA2\xB2\x00\xEC\xA2\xB3\x00\xEC\xA2\xB4\x00\xEC\xA2\xB5\x00"
-     "\xEC\xA2"
-     "\xB6\x00\xEC\xA2\xB7\x00\xEC\xA2\xB8\x00\xEC\xA2\xB9\x00\xEC\xA2\xBA\x00"
-     "\xEC\xA2"
-     "\xBB\x00\xEC\xA2\xBC\x00\xEC\xA2\xBD\x00\xEC\xA2\xBE\x00\xEC\xA2\xBF\x00"
-     "\xEC\xA3"
-     "\x80\x00\xEC\xA3\x81\x00\xEC\xA3\x82\x00\xEC\xA3\x83\x00\xEC\xA3\x84\x00"
-     "\xEC\xA3"
-     "\x85\x00\xEC\xA3\x86\x00\xEC\xA3\x87\x00\xEC\xA3\x88\x00\xEC\xA3\x89\x00"
-     "\xEC\xA3"
-     "\x8A\x00\xEC\xA3\x8B\x00\xEC\xA3\x8C\x00\xEC\xA3\x8D\x00\xEC\xA3\x8E\x00"
-     "\xEC\xA3"
-     "\x8F\x00\xEC\xA3\x90\x00\xEC\xA3\x91\x00\xEC\xA3\x92\x00\xEC\xA3\x93\x00"
-     "\xEC\xA3"
-     "\x94\x00\xEC\xA3\x95\x00\xEC\xA3\x96\x00\xEC\xA3\x97\x00\xEC\xA3\x98\x00"
-     "\xEC\xA3"
-     "\x99\x00\xEC\xA3\x9A\x00\xEC\xA3\x9B\x00\xEC\xA3\x9C\x00\xEC\xA3\x9D\x00"
-     "\xEC\xA3"
-     "\x9E\x00\xEC\xA3\x9F\x00\xEC\xA3\xA0\x00\xEC\xA3\xA1\x00\xEC\xA3\xA2\x00"
-     "\xEC\xA3"
-     "\xA3\x00\xEC\xA3\xA4\x00\xEC\xA3\xA5\x00\xEC\xA3\xA6\x00\xEC\xA3\xA7\x00"
-     "\xEC\xA3"
-     "\xA8\x00\xEC\xA3\xA9\x00\xEC\xA3\xAA\x00\xEC\xA3\xAB\x00\xEC\xA3\xAC\x00"
-     "\xEC\xA3"
-     "\xAD\x00\xEC\xA3\xAE\x00\xEC\xA3\xAF\x00\xEC\xA3\xB0\x00\xEC\xA3\xB1\x00"
-     "\xEC\xA3"
-     "\xB2\x00\xEC\xA3\xB3\x00\xEC\xA3\xB4\x00\xEC\xA3\xB5\x00\xEC\xA3\xB6\x00"
-     "\xEC\xA3"
-     "\xB7\x00\xEC\xA3\xB8\x00\xEC\xA3\xB9\x00\xEC\xA3\xBA\x00\xEC\xA3\xBB\x00"
-     "\xEC\xA3"
-     "\xBC\x00\xEC\xA3\xBC\xEC\x9D\x98\x00\xEC\xA3\xBD\x00\xEC\xA3\xBE\x00\xEC"
-     "\xA3\xBF"
-     "\x00\xEC\xA4\x80\x00\xEC\xA4\x81\x00\xEC\xA4\x82\x00\xEC\xA4\x83\x00\xEC"
-     "\xA4\x84"
-     "\x00\xEC\xA4\x85\x00\xEC\xA4\x86\x00\xEC\xA4\x87\x00\xEC\xA4\x88\x00\xEC"
-     "\xA4\x89"
-     "\x00\xEC\xA4\x8A\x00\xEC\xA4\x8B\x00\xEC\xA4\x8C\x00\xEC\xA4\x8D\x00\xEC"
-     "\xA4\x8E"
-     "\x00\xEC\xA4\x8F\x00\xEC\xA4\x90\x00\xEC\xA4\x91\x00\xEC\xA4\x92\x00\xEC"
-     "\xA4\x93"
-     "\x00\xEC\xA4\x94\x00\xEC\xA4\x95\x00\xEC\xA4\x96\x00\xEC\xA4\x97\x00\xEC"
-     "\xA4\x98"
-     "\x00\xEC\xA4\x99\x00\xEC\xA4\x9A\x00\xEC\xA4\x9B\x00\xEC\xA4\x9C\x00\xEC"
-     "\xA4\x9D"
-     "\x00\xEC\xA4\x9E\x00\xEC\xA4\x9F\x00\xEC\xA4\xA0\x00\xEC\xA4\xA1\x00\xEC"
-     "\xA4\xA2"
-     "\x00\xEC\xA4\xA3\x00\xEC\xA4\xA4\x00\xEC\xA4\xA5\x00\xEC\xA4\xA6\x00\xEC"
-     "\xA4\xA7"
-     "\x00\xEC\xA4\xA8\x00\xEC\xA4\xA9\x00\xEC\xA4\xAA\x00\xEC\xA4\xAB\x00\xEC"
-     "\xA4\xAC"
-     "\x00\xEC\xA4\xAD\x00\xEC\xA4\xAE\x00\xEC\xA4\xAF\x00\xEC\xA4\xB0\x00\xEC"
-     "\xA4\xB1"
-     "\x00\xEC\xA4\xB2\x00\xEC\xA4\xB3\x00\xEC\xA4\xB4\x00\xEC\xA4\xB5\x00\xEC"
-     "\xA4\xB6"
-     "\x00\xEC\xA4\xB7\x00\xEC\xA4\xB8\x00\xEC\xA4\xB9\x00\xEC\xA4\xBA\x00\xEC"
-     "\xA4\xBB"
-     "\x00\xEC\xA4\xBC\x00\xEC\xA4\xBD\x00\xEC\xA4\xBE\x00\xEC\xA4\xBF\x00\xEC"
-     "\xA5\x80"
-     "\x00\xEC\xA5\x81\x00\xEC\xA5\x82\x00\xEC\xA5\x83\x00\xEC\xA5\x84\x00\xEC"
-     "\xA5\x85"
-     "\x00\xEC\xA5\x86\x00\xEC\xA5\x87\x00\xEC\xA5\x88\x00\xEC\xA5\x89\x00\xEC"
-     "\xA5\x8A"
-     "\x00\xEC\xA5\x8B\x00\xEC\xA5\x8C\x00\xEC\xA5\x8D\x00\xEC\xA5\x8E\x00\xEC"
-     "\xA5\x8F"
-     "\x00\xEC\xA5\x90\x00\xEC\xA5\x91\x00\xEC\xA5\x92\x00\xEC\xA5\x93\x00\xEC"
-     "\xA5\x94"
-     "\x00\xEC\xA5\x95\x00\xEC\xA5\x96\x00\xEC\xA5\x97\x00\xEC\xA5\x98\x00\xEC"
-     "\xA5\x99"
-     "\x00\xEC\xA5\x9A\x00\xEC\xA5\x9B\x00\xEC\xA5\x9C\x00\xEC\xA5\x9D\x00\xEC"
-     "\xA5\x9E"
-     "\x00\xEC\xA5\x9F\x00\xEC\xA5\xA0\x00\xEC\xA5\xA1\x00\xEC\xA5\xA2\x00\xEC"
-     "\xA5\xA3"
-     "\x00\xEC\xA5\xA4\x00\xEC\xA5\xA5\x00\xEC\xA5\xA6\x00\xEC\xA5\xA7\x00\xEC"
-     "\xA5\xA8"
-     "\x00\xEC\xA5\xA9\x00\xEC\xA5\xAA\x00\xEC\xA5\xAB\x00\xEC\xA5\xAC\x00\xEC"
-     "\xA5\xAD"
-     "\x00\xEC\xA5\xAE\x00\xEC\xA5\xAF\x00\xEC\xA5\xB0\x00\xEC\xA5\xB1\x00\xEC"
-     "\xA5\xB2"
-     "\x00\xEC\xA5\xB3\x00\xEC\xA5\xB4\x00\xEC\xA5\xB5\x00\xEC\xA5\xB6\x00\xEC"
-     "\xA5\xB7"
-     "\x00\xEC\xA5\xB8\x00\xEC\xA5\xB9\x00\xEC\xA5\xBA\x00\xEC\xA5\xBB\x00\xEC"
-     "\xA5\xBC"
-     "\x00\xEC\xA5\xBD\x00\xEC\xA5\xBE\x00\xEC\xA5\xBF\x00\xEC\xA6\x80\x00\xEC"
-     "\xA6\x81"
-     "\x00\xEC\xA6\x82\x00\xEC\xA6\x83\x00\xEC\xA6\x84\x00\xEC\xA6\x85\x00\xEC"
-     "\xA6\x86"
-     "\x00\xEC\xA6\x87\x00\xEC\xA6\x88\x00\xEC\xA6\x89\x00\xEC\xA6\x8A\x00\xEC"
-     "\xA6\x8B"
-     "\x00\xEC\xA6\x8C\x00\xEC\xA6\x8D\x00\xEC\xA6\x8E\x00\xEC\xA6\x8F\x00\xEC"
-     "\xA6\x90"
-     "\x00\xEC\xA6\x91\x00\xEC\xA6\x92\x00\xEC\xA6\x93\x00\xEC\xA6\x94\x00\xEC"
-     "\xA6\x95"
-     "\x00\xEC\xA6\x96\x00\xEC\xA6\x97\x00\xEC\xA6\x98\x00\xEC\xA6\x99\x00\xEC"
-     "\xA6\x9A"
-     "\x00\xEC\xA6\x9B\x00\xEC\xA6\x9C\x00\xEC\xA6\x9D\x00\xEC\xA6\x9E\x00\xEC"
-     "\xA6\x9F"
-     "\x00\xEC\xA6\xA0\x00\xEC\xA6\xA1\x00\xEC\xA6\xA2\x00\xEC\xA6\xA3\x00\xEC"
-     "\xA6\xA4"
-     "\x00\xEC\xA6\xA5\x00\xEC\xA6\xA6\x00\xEC\xA6\xA7\x00\xEC\xA6\xA8\x00\xEC"
-     "\xA6\xA9"
-     "\x00\xEC\xA6\xAA\x00\xEC\xA6\xAB\x00\xEC\xA6\xAC\x00\xEC\xA6\xAD\x00\xEC"
-     "\xA6\xAE"
-     "\x00\xEC\xA6\xAF\x00\xEC\xA6\xB0\x00\xEC\xA6\xB1\x00\xEC\xA6\xB2\x00\xEC"
-     "\xA6\xB3"
-     "\x00\xEC\xA6\xB4\x00\xEC\xA6\xB5\x00\xEC\xA6\xB6\x00\xEC\xA6\xB7\x00\xEC"
-     "\xA6\xB8"
-     "\x00\xEC\xA6\xB9\x00\xEC\xA6\xBA\x00\xEC\xA6\xBB\x00\xEC\xA6\xBC\x00\xEC"
-     "\xA6\xBD"
-     "\x00\xEC\xA6\xBE\x00\xEC\xA6\xBF\x00\xEC\xA7\x80\x00\xEC\xA7\x81\x00\xEC"
-     "\xA7\x82"
-     "\x00\xEC\xA7\x83\x00\xEC\xA7\x84\x00\xEC\xA7\x85\x00\xEC\xA7\x86\x00\xEC"
-     "\xA7\x87"
-     "\x00\xEC\xA7\x88\x00\xEC\xA7\x89\x00\xEC\xA7\x8A\x00\xEC\xA7\x8B\x00\xEC"
-     "\xA7\x8C"
-     "\x00\xEC\xA7\x8D\x00\xEC\xA7\x8E\x00\xEC\xA7\x8F\x00\xEC\xA7\x90\x00\xEC"
-     "\xA7\x91"
-     "\x00\xEC\xA7\x92\x00\xEC\xA7\x93\x00\xEC\xA7\x94\x00\xEC\xA7\x95\x00\xEC"
-     "\xA7\x96"
-     "\x00\xEC\xA7\x97\x00\xEC\xA7\x98\x00\xEC\xA7\x99\x00\xEC\xA7\x9A\x00\xEC"
-     "\xA7\x9B"
-     "\x00\xEC\xA7\x9C\x00\xEC\xA7\x9D\x00\xEC\xA7\x9E\x00\xEC\xA7\x9F\x00\xEC"
-     "\xA7\xA0"
-     "\x00\xEC\xA7\xA1\x00\xEC\xA7\xA2\x00\xEC\xA7\xA3\x00\xEC\xA7\xA4\x00\xEC"
-     "\xA7\xA5"
-     "\x00\xEC\xA7\xA6\x00\xEC\xA7\xA7\x00\xEC\xA7\xA8\x00\xEC\xA7\xA9\x00\xEC"
-     "\xA7\xAA"
-     "\x00\xEC\xA7\xAB\x00\xEC\xA7\xAC\x00\xEC\xA7\xAD\x00\xEC\xA7\xAE\x00\xEC"
-     "\xA7\xAF"
-     "\x00\xEC\xA7\xB0\x00\xEC\xA7\xB1\x00\xEC\xA7\xB2\x00\xEC\xA7\xB3\x00\xEC"
-     "\xA7\xB4"
-     "\x00\xEC\xA7\xB5\x00\xEC\xA7\xB6\x00\xEC\xA7\xB7\x00\xEC\xA7\xB8\x00\xEC"
-     "\xA7\xB9"
-     "\x00\xEC\xA7\xBA\x00\xEC\xA7\xBB\x00\xEC\xA7\xBC\x00\xEC\xA7\xBD\x00\xEC"
-     "\xA7\xBE"
-     "\x00\xEC\xA7\xBF\x00\xEC\xA8\x80\x00\xEC\xA8\x81\x00\xEC\xA8\x82\x00\xEC"
-     "\xA8\x83"
-     "\x00\xEC\xA8\x84\x00\xEC\xA8\x85\x00\xEC\xA8\x86\x00\xEC\xA8\x87\x00\xEC"
-     "\xA8\x88"
-     "\x00\xEC\xA8\x89\x00\xEC\xA8\x8A\x00\xEC\xA8\x8B\x00\xEC\xA8\x8C\x00\xEC"
-     "\xA8\x8D"
-     "\x00\xEC\xA8\x8E\x00\xEC\xA8\x8F\x00\xEC\xA8\x90\x00\xEC\xA8\x91\x00\xEC"
-     "\xA8\x92"
-     "\x00\xEC\xA8\x93\x00\xEC\xA8\x94\x00\xEC\xA8\x95\x00\xEC\xA8\x96\x00\xEC"
-     "\xA8\x97"
-     "\x00\xEC\xA8\x98\x00\xEC\xA8\x99\x00\xEC\xA8\x9A\x00\xEC\xA8\x9B\x00\xEC"
-     "\xA8\x9C"
-     "\x00\xEC\xA8\x9D\x00\xEC\xA8\x9E\x00\xEC\xA8\x9F\x00\xEC\xA8\xA0\x00\xEC"
-     "\xA8\xA1"
-     "\x00\xEC\xA8\xA2\x00\xEC\xA8\xA3\x00\xEC\xA8\xA4\x00\xEC\xA8\xA5\x00\xEC"
-     "\xA8\xA6"
-     "\x00\xEC\xA8\xA7\x00\xEC\xA8\xA8\x00\xEC\xA8\xA9\x00\xEC\xA8\xAA\x00\xEC"
-     "\xA8\xAB"
-     "\x00\xEC\xA8\xAC\x00\xEC\xA8\xAD\x00\xEC\xA8\xAE\x00\xEC\xA8\xAF\x00\xEC"
-     "\xA8\xB0"
-     "\x00\xEC\xA8\xB1\x00\xEC\xA8\xB2\x00\xEC\xA8\xB3\x00\xEC\xA8\xB4\x00\xEC"
-     "\xA8\xB5"
-     "\x00\xEC\xA8\xB6\x00\xEC\xA8\xB7\x00\xEC\xA8\xB8\x00\xEC\xA8\xB9\x00\xEC"
-     "\xA8\xBA"
-     "\x00\xEC\xA8\xBB\x00\xEC\xA8\xBC\x00\xEC\xA8\xBD\x00\xEC\xA8\xBE\x00\xEC"
-     "\xA8\xBF"
-     "\x00\xEC\xA9\x80\x00\xEC\xA9\x81\x00\xEC\xA9\x82\x00\xEC\xA9\x83\x00\xEC"
-     "\xA9\x84"
-     "\x00\xEC\xA9\x85\x00\xEC\xA9\x86\x00\xEC\xA9\x87\x00\xEC\xA9\x88\x00\xEC"
-     "\xA9\x89"
-     "\x00\xEC\xA9\x8A\x00\xEC\xA9\x8B\x00\xEC\xA9\x8C\x00\xEC\xA9\x8D\x00\xEC"
-     "\xA9\x8E"
-     "\x00\xEC\xA9\x8F\x00\xEC\xA9\x90\x00\xEC\xA9\x91\x00\xEC\xA9\x92\x00\xEC"
-     "\xA9\x93"
-     "\x00\xEC\xA9\x94\x00\xEC\xA9\x95\x00\xEC\xA9\x96\x00\xEC\xA9\x97\x00\xEC"
-     "\xA9\x98"
-     "\x00\xEC\xA9\x99\x00\xEC\xA9\x9A\x00\xEC\xA9\x9B\x00\xEC\xA9\x9C\x00\xEC"
-     "\xA9\x9D"
-     "\x00\xEC\xA9\x9E\x00\xEC\xA9\x9F\x00\xEC\xA9\xA0\x00\xEC\xA9\xA1\x00\xEC"
-     "\xA9\xA2"
-     "\x00\xEC\xA9\xA3\x00\xEC\xA9\xA4\x00\xEC\xA9\xA5\x00\xEC\xA9\xA6\x00\xEC"
-     "\xA9\xA7"
-     "\x00\xEC\xA9\xA8\x00\xEC\xA9\xA9\x00\xEC\xA9\xAA\x00\xEC\xA9\xAB\x00\xEC"
-     "\xA9\xAC"
-     "\x00\xEC\xA9\xAD\x00\xEC\xA9\xAE\x00\xEC\xA9\xAF\x00\xEC\xA9\xB0\x00\xEC"
-     "\xA9\xB1"
-     "\x00\xEC\xA9\xB2\x00\xEC\xA9\xB3\x00\xEC\xA9\xB4\x00\xEC\xA9\xB5\x00\xEC"
-     "\xA9\xB6"
-     "\x00\xEC\xA9\xB7\x00\xEC\xA9\xB8\x00\xEC\xA9\xB9\x00\xEC\xA9\xBA\x00\xEC"
-     "\xA9\xBB"
-     "\x00\xEC\xA9\xBC\x00\xEC\xA9\xBD\x00\xEC\xA9\xBE\x00\xEC\xA9\xBF\x00\xEC"
-     "\xAA\x80"
-     "\x00\xEC\xAA\x81\x00\xEC\xAA\x82\x00\xEC\xAA\x83\x00\xEC\xAA\x84\x00\xEC"
-     "\xAA\x85"
-     "\x00\xEC\xAA\x86\x00\xEC\xAA\x87\x00\xEC\xAA\x88\x00\xEC\xAA\x89\x00\xEC"
-     "\xAA\x8A"
-     "\x00\xEC\xAA\x8B\x00\xEC\xAA\x8C\x00\xEC\xAA\x8D\x00\xEC\xAA\x8E\x00\xEC"
-     "\xAA\x8F"
-     "\x00\xEC\xAA\x90\x00\xEC\xAA\x91\x00\xEC\xAA\x92\x00\xEC\xAA\x93\x00\xEC"
-     "\xAA\x94"
-     "\x00\xEC\xAA\x95\x00\xEC\xAA\x96\x00\xEC\xAA\x97\x00\xEC\xAA\x98\x00\xEC"
-     "\xAA\x99"
-     "\x00\xEC\xAA\x9A\x00\xEC\xAA\x9B\x00\xEC\xAA\x9C\x00\xEC\xAA\x9D\x00\xEC"
-     "\xAA\x9E"
-     "\x00\xEC\xAA\x9F\x00\xEC\xAA\xA0\x00\xEC\xAA\xA1\x00\xEC\xAA\xA2\x00\xEC"
-     "\xAA\xA3"
-     "\x00\xEC\xAA\xA4\x00\xEC\xAA\xA5\x00\xEC\xAA\xA6\x00\xEC\xAA\xA7\x00\xEC"
-     "\xAA\xA8"
-     "\x00\xEC\xAA\xA9\x00\xEC\xAA\xAA\x00\xEC\xAA\xAB\x00\xEC\xAA\xAC\x00\xEC"
-     "\xAA\xAD"
-     "\x00\xEC\xAA\xAE\x00\xEC\xAA\xAF\x00\xEC\xAA\xB0\x00\xEC\xAA\xB1\x00\xEC"
-     "\xAA\xB2"
-     "\x00\xEC\xAA\xB3\x00\xEC\xAA\xB4\x00\xEC\xAA\xB5\x00\xEC\xAA\xB6\x00\xEC"
-     "\xAA\xB7"
-     "\x00\xEC\xAA\xB8\x00\xEC\xAA\xB9\x00\xEC\xAA\xBA\x00\xEC\xAA\xBB\x00\xEC"
-     "\xAA\xBC"
-     "\x00\xEC\xAA\xBD\x00\xEC\xAA\xBE\x00\xEC\xAA\xBF\x00\xEC\xAB\x80\x00\xEC"
-     "\xAB\x81"
-     "\x00\xEC\xAB\x82\x00\xEC\xAB\x83\x00\xEC\xAB\x84\x00\xEC\xAB\x85\x00\xEC"
-     "\xAB\x86"
-     "\x00\xEC\xAB\x87\x00\xEC\xAB\x88\x00\xEC\xAB\x89\x00\xEC\xAB\x8A\x00\xEC"
-     "\xAB\x8B"
-     "\x00\xEC\xAB\x8C\x00\xEC\xAB\x8D\x00\xEC\xAB\x8E\x00\xEC\xAB\x8F\x00\xEC"
-     "\xAB\x90"
-     "\x00\xEC\xAB\x91\x00\xEC\xAB\x92\x00\xEC\xAB\x93\x00\xEC\xAB\x94\x00\xEC"
-     "\xAB\x95"
-     "\x00\xEC\xAB\x96\x00\xEC\xAB\x97\x00\xEC\xAB\x98\x00\xEC\xAB\x99\x00\xEC"
-     "\xAB\x9A"
-     "\x00\xEC\xAB\x9B\x00\xEC\xAB\x9C\x00\xEC\xAB\x9D\x00\xEC\xAB\x9E\x00\xEC"
-     "\xAB\x9F"
-     "\x00\xEC\xAB\xA0\x00\xEC\xAB\xA1\x00\xEC\xAB\xA2\x00\xEC\xAB\xA3\x00\xEC"
-     "\xAB\xA4"
-     "\x00\xEC\xAB\xA5\x00\xEC\xAB\xA6\x00\xEC\xAB\xA7\x00\xEC\xAB\xA8\x00\xEC"
-     "\xAB\xA9"
-     "\x00\xEC\xAB\xAA\x00\xEC\xAB\xAB\x00\xEC\xAB\xAC\x00\xEC\xAB\xAD\x00\xEC"
-     "\xAB\xAE"
-     "\x00\xEC\xAB\xAF\x00\xEC\xAB\xB0\x00\xEC\xAB\xB1\x00\xEC\xAB\xB2\x00\xEC"
-     "\xAB\xB3"
-     "\x00\xEC\xAB\xB4\x00\xEC\xAB\xB5\x00\xEC\xAB\xB6\x00\xEC\xAB\xB7\x00\xEC"
-     "\xAB\xB8"
-     "\x00\xEC\xAB\xB9\x00\xEC\xAB\xBA\x00\xEC\xAB\xBB\x00\xEC\xAB\xBC\x00\xEC"
-     "\xAB\xBD"
-     "\x00\xEC\xAB\xBE\x00\xEC\xAB\xBF\x00\xEC\xAC\x80\x00\xEC\xAC\x81\x00\xEC"
-     "\xAC\x82"
-     "\x00\xEC\xAC\x83\x00\xEC\xAC\x84\x00\xEC\xAC\x85\x00\xEC\xAC\x86\x00\xEC"
-     "\xAC\x87"
-     "\x00\xEC\xAC\x88\x00\xEC\xAC\x89\x00\xEC\xAC\x8A\x00\xEC\xAC\x8B\x00\xEC"
-     "\xAC\x8C"
-     "\x00\xEC\xAC\x8D\x00\xEC\xAC\x8E\x00\xEC\xAC\x8F\x00\xEC\xAC\x90\x00\xEC"
-     "\xAC\x91"
-     "\x00\xEC\xAC\x92\x00\xEC\xAC\x93\x00\xEC\xAC\x94\x00\xEC\xAC\x95\x00\xEC"
-     "\xAC\x96"
-     "\x00\xEC\xAC\x97\x00\xEC\xAC\x98\x00\xEC\xAC\x99\x00\xEC\xAC\x9A\x00\xEC"
-     "\xAC\x9B"
-     "\x00\xEC\xAC\x9C\x00\xEC\xAC\x9D\x00\xEC\xAC\x9E\x00\xEC\xAC\x9F\x00\xEC"
-     "\xAC\xA0"
-     "\x00\xEC\xAC\xA1\x00\xEC\xAC\xA2\x00\xEC\xAC\xA3\x00\xEC\xAC\xA4\x00\xEC"
-     "\xAC\xA5"
-     "\x00\xEC\xAC\xA6\x00\xEC\xAC\xA7\x00\xEC\xAC\xA8\x00\xEC\xAC\xA9\x00\xEC"
-     "\xAC\xAA"
-     "\x00\xEC\xAC\xAB\x00\xEC\xAC\xAC\x00\xEC\xAC\xAD\x00\xEC\xAC\xAE\x00\xEC"
-     "\xAC\xAF"
-     "\x00\xEC\xAC\xB0\x00\xEC\xAC\xB1\x00\xEC\xAC\xB2\x00\xEC\xAC\xB3\x00\xEC"
-     "\xAC\xB4"
-     "\x00\xEC\xAC\xB5\x00\xEC\xAC\xB6\x00\xEC\xAC\xB7\x00\xEC\xAC\xB8\x00\xEC"
-     "\xAC\xB9"
-     "\x00\xEC\xAC\xBA\x00\xEC\xAC\xBB\x00\xEC\xAC\xBC\x00\xEC\xAC\xBD\x00\xEC"
-     "\xAC\xBE"
-     "\x00\xEC\xAC\xBF\x00\xEC\xAD\x80\x00\xEC\xAD\x81\x00\xEC\xAD\x82\x00\xEC"
-     "\xAD\x83"
-     "\x00\xEC\xAD\x84\x00\xEC\xAD\x85\x00\xEC\xAD\x86\x00\xEC\xAD\x87\x00\xEC"
-     "\xAD\x88"
-     "\x00\xEC\xAD\x89\x00\xEC\xAD\x8A\x00\xEC\xAD\x8B\x00\xEC\xAD\x8C\x00\xEC"
-     "\xAD\x8D"
-     "\x00\xEC\xAD\x8E\x00\xEC\xAD\x8F\x00\xEC\xAD\x90\x00\xEC\xAD\x91\x00\xEC"
-     "\xAD\x92"
-     "\x00\xEC\xAD\x93\x00\xEC\xAD\x94\x00\xEC\xAD\x95\x00\xEC\xAD\x96\x00\xEC"
-     "\xAD\x97"
-     "\x00\xEC\xAD\x98\x00\xEC\xAD\x99\x00\xEC\xAD\x9A\x00\xEC\xAD\x9B\x00\xEC"
-     "\xAD\x9C"
-     "\x00\xEC\xAD\x9D\x00\xEC\xAD\x9E\x00\xEC\xAD\x9F\x00\xEC\xAD\xA0\x00\xEC"
-     "\xAD\xA1"
-     "\x00\xEC\xAD\xA2\x00\xEC\xAD\xA3\x00\xEC\xAD\xA4\x00\xEC\xAD\xA5\x00\xEC"
-     "\xAD\xA6"
-     "\x00\xEC\xAD\xA7\x00\xEC\xAD\xA8\x00\xEC\xAD\xA9\x00\xEC\xAD\xAA\x00\xEC"
-     "\xAD\xAB"
-     "\x00\xEC\xAD\xAC\x00\xEC\xAD\xAD\x00\xEC\xAD\xAE\x00\xEC\xAD\xAF\x00\xEC"
-     "\xAD\xB0"
-     "\x00\xEC\xAD\xB1\x00\xEC\xAD\xB2\x00\xEC\xAD\xB3\x00\xEC\xAD\xB4\x00\xEC"
-     "\xAD\xB5"
-     "\x00\xEC\xAD\xB6\x00\xEC\xAD\xB7\x00\xEC\xAD\xB8\x00\xEC\xAD\xB9\x00\xEC"
-     "\xAD\xBA"
-     "\x00\xEC\xAD\xBB\x00\xEC\xAD\xBC\x00\xEC\xAD\xBD\x00\xEC\xAD\xBE\x00\xEC"
-     "\xAD\xBF"
-     "\x00\xEC\xAE\x80\x00\xEC\xAE\x81\x00\xEC\xAE\x82\x00\xEC\xAE\x83\x00\xEC"
-     "\xAE\x84"
-     "\x00\xEC\xAE\x85\x00\xEC\xAE\x86\x00\xEC\xAE\x87\x00\xEC\xAE\x88\x00\xEC"
-     "\xAE\x89"
-     "\x00\xEC\xAE\x8A\x00\xEC\xAE\x8B\x00\xEC\xAE\x8C\x00\xEC\xAE\x8D\x00\xEC"
-     "\xAE\x8E"
-     "\x00\xEC\xAE\x8F\x00\xEC\xAE\x90\x00\xEC\xAE\x91\x00\xEC\xAE\x92\x00\xEC"
-     "\xAE\x93"
-     "\x00\xEC\xAE\x94\x00\xEC\xAE\x95\x00\xEC\xAE\x96\x00\xEC\xAE\x97\x00\xEC"
-     "\xAE\x98"
-     "\x00\xEC\xAE\x99\x00\xEC\xAE\x9A\x00\xEC\xAE\x9B\x00\xEC\xAE\x9C\x00\xEC"
-     "\xAE\x9D"
-     "\x00\xEC\xAE\x9E\x00\xEC\xAE\x9F\x00\xEC\xAE\xA0\x00\xEC\xAE\xA1\x00\xEC"
-     "\xAE\xA2"
-     "\x00\xEC\xAE\xA3\x00\xEC\xAE\xA4\x00\xEC\xAE\xA5\x00\xEC\xAE\xA6\x00\xEC"
-     "\xAE\xA7"
-     "\x00\xEC\xAE\xA8\x00\xEC\xAE\xA9\x00\xEC\xAE\xAA\x00\xEC\xAE\xAB\x00\xEC"
-     "\xAE\xAC"
-     "\x00\xEC\xAE\xAD\x00\xEC\xAE\xAE\x00\xEC\xAE\xAF\x00\xEC\xAE\xB0\x00\xEC"
-     "\xAE\xB1"
-     "\x00\xEC\xAE\xB2\x00\xEC\xAE\xB3\x00\xEC\xAE\xB4\x00\xEC\xAE\xB5\x00\xEC"
-     "\xAE\xB6"
-     "\x00\xEC\xAE\xB7\x00\xEC\xAE\xB8\x00\xEC\xAE\xB9\x00\xEC\xAE\xBA\x00\xEC"
-     "\xAE\xBB"
-     "\x00\xEC\xAE\xBC\x00\xEC\xAE\xBD\x00\xEC\xAE\xBE\x00\xEC\xAE\xBF\x00\xEC"
-     "\xAF\x80"
-     "\x00\xEC\xAF\x81\x00\xEC\xAF\x82\x00\xEC\xAF\x83\x00\xEC\xAF\x84\x00\xEC"
-     "\xAF\x85"
-     "\x00\xEC\xAF\x86\x00\xEC\xAF\x87\x00\xEC\xAF\x88\x00\xEC\xAF\x89\x00\xEC"
-     "\xAF\x8A"
-     "\x00\xEC\xAF\x8B\x00\xEC\xAF\x8C\x00\xEC\xAF\x8D\x00\xEC\xAF\x8E\x00\xEC"
-     "\xAF\x8F"
-     "\x00\xEC\xAF\x90\x00\xEC\xAF\x91\x00\xEC\xAF\x92\x00\xEC\xAF\x93\x00\xEC"
-     "\xAF\x94"
-     "\x00\xEC\xAF\x95\x00\xEC\xAF\x96\x00\xEC\xAF\x97\x00\xEC\xAF\x98\x00\xEC"
-     "\xAF\x99"
-     "\x00\xEC\xAF\x9A\x00\xEC\xAF\x9B\x00\xEC\xAF\x9C\x00\xEC\xAF\x9D\x00\xEC"
-     "\xAF\x9E"
-     "\x00\xEC\xAF\x9F\x00\xEC\xAF\xA0\x00\xEC\xAF\xA1\x00\xEC\xAF\xA2\x00\xEC"
-     "\xAF\xA3"
-     "\x00\xEC\xAF\xA4\x00\xEC\xAF\xA5\x00\xEC\xAF\xA6\x00\xEC\xAF\xA7\x00\xEC"
-     "\xAF\xA8"
-     "\x00\xEC\xAF\xA9\x00\xEC\xAF\xAA\x00\xEC\xAF\xAB\x00\xEC\xAF\xAC\x00\xEC"
-     "\xAF\xAD"
-     "\x00\xEC\xAF\xAE\x00\xEC\xAF\xAF\x00\xEC\xAF\xB0\x00\xEC\xAF\xB1\x00\xEC"
-     "\xAF\xB2"
-     "\x00\xEC\xAF\xB3\x00\xEC\xAF\xB4\x00\xEC\xAF\xB5\x00\xEC\xAF\xB6\x00\xEC"
-     "\xAF\xB7"
-     "\x00\xEC\xAF\xB8\x00\xEC\xAF\xB9\x00\xEC\xAF\xBA\x00\xEC\xAF\xBB\x00\xEC"
-     "\xAF\xBC"
-     "\x00\xEC\xAF\xBD\x00\xEC\xAF\xBE\x00\xEC\xAF\xBF\x00\xEC\xB0\x80\x00\xEC"
-     "\xB0\x81"
-     "\x00\xEC\xB0\x82\x00\xEC\xB0\x83\x00\xEC\xB0\x84\x00\xEC\xB0\x85\x00\xEC"
-     "\xB0\x86"
-     "\x00\xEC\xB0\x87\x00\xEC\xB0\x88\x00\xEC\xB0\x89\x00\xEC\xB0\x8A\x00\xEC"
-     "\xB0\x8B"
-     "\x00\xEC\xB0\x8C\x00\xEC\xB0\x8D\x00\xEC\xB0\x8E\x00\xEC\xB0\x8F\x00\xEC"
-     "\xB0\x90"
-     "\x00\xEC\xB0\x91\x00\xEC\xB0\x92\x00\xEC\xB0\x93\x00\xEC\xB0\x94\x00\xEC"
-     "\xB0\x95"
-     "\x00\xEC\xB0\x96\x00\xEC\xB0\x97\x00\xEC\xB0\x98\x00\xEC\xB0\x99\x00\xEC"
-     "\xB0\x9A"
-     "\x00\xEC\xB0\x9B\x00\xEC\xB0\x9C\x00\xEC\xB0\x9D\x00\xEC\xB0\x9E\x00\xEC"
-     "\xB0\x9F"
-     "\x00\xEC\xB0\xA0\x00\xEC\xB0\xA1\x00\xEC\xB0\xA2\x00\xEC\xB0\xA3\x00\xEC"
-     "\xB0\xA4"
-     "\x00\xEC\xB0\xA5\x00\xEC\xB0\xA6\x00\xEC\xB0\xA7\x00\xEC\xB0\xA8\x00\xEC"
-     "\xB0\xA9"
-     "\x00\xEC\xB0\xAA\x00\xEC\xB0\xAB\x00\xEC\xB0\xAC\x00\xEC\xB0\xAD\x00\xEC"
-     "\xB0\xAE"
-     "\x00\xEC\xB0\xAF\x00\xEC\xB0\xB0\x00\xEC\xB0\xB1\x00\xEC\xB0\xB2\x00\xEC"
-     "\xB0\xB3"
-     "\x00\xEC\xB0\xB4\x00\xEC\xB0\xB5\x00\xEC\xB0\xB6\x00\xEC\xB0\xB7\x00\xEC"
-     "\xB0\xB8"
-     "\x00\xEC\xB0\xB8\xEA\xB3\xA0\x00\xEC\xB0\xB9\x00\xEC\xB0\xBA\x00\xEC\xB0"
-     "\xBB\x00"
-     "\xEC\xB0\xBC\x00\xEC\xB0\xBD\x00\xEC\xB0\xBE\x00\xEC\xB0\xBF\x00\xEC\xB1"
-     "\x80\x00"
-     "\xEC\xB1\x81\x00\xEC\xB1\x82\x00\xEC\xB1\x83\x00\xEC\xB1\x84\x00\xEC\xB1"
-     "\x85\x00"
-     "\xEC\xB1\x86\x00\xEC\xB1\x87\x00\xEC\xB1\x88\x00\xEC\xB1\x89\x00\xEC\xB1"
-     "\x8A\x00"
-     "\xEC\xB1\x8B\x00\xEC\xB1\x8C\x00\xEC\xB1\x8D\x00\xEC\xB1\x8E\x00\xEC\xB1"
-     "\x8F\x00"
-     "\xEC\xB1\x90\x00\xEC\xB1\x91\x00\xEC\xB1\x92\x00\xEC\xB1\x93\x00\xEC\xB1"
-     "\x94\x00"
-     "\xEC\xB1\x95\x00\xEC\xB1\x96\x00\xEC\xB1\x97\x00\xEC\xB1\x98\x00\xEC\xB1"
-     "\x99\x00"
-     "\xEC\xB1\x9A\x00\xEC\xB1\x9B\x00\xEC\xB1\x9C\x00\xEC\xB1\x9D\x00\xEC\xB1"
-     "\x9E\x00"
-     "\xEC\xB1\x9F\x00\xEC\xB1\xA0\x00\xEC\xB1\xA1\x00\xEC\xB1\xA2\x00\xEC\xB1"
-     "\xA3\x00"
-     "\xEC\xB1\xA4\x00\xEC\xB1\xA5\x00\xEC\xB1\xA6\x00\xEC\xB1\xA7\x00\xEC\xB1"
-     "\xA8\x00"
-     "\xEC\xB1\xA9\x00\xEC\xB1\xAA\x00\xEC\xB1\xAB\x00\xEC\xB1\xAC\x00\xEC\xB1"
-     "\xAD\x00"
-     "\xEC\xB1\xAE\x00\xEC\xB1\xAF\x00\xEC\xB1\xB0\x00\xEC\xB1\xB1\x00\xEC\xB1"
-     "\xB2\x00"
-     "\xEC\xB1\xB3\x00\xEC\xB1\xB4\x00\xEC\xB1\xB5\x00\xEC\xB1\xB6\x00\xEC\xB1"
-     "\xB7\x00"
-     "\xEC\xB1\xB8\x00\xEC\xB1\xB9\x00\xEC\xB1\xBA\x00\xEC\xB1\xBB\x00\xEC\xB1"
-     "\xBC\x00"
-     "\xEC\xB1\xBD\x00\xEC\xB1\xBE\x00\xEC\xB1\xBF\x00\xEC\xB2\x80\x00\xEC\xB2"
-     "\x81\x00"
-     "\xEC\xB2\x82\x00\xEC\xB2\x83\x00\xEC\xB2\x84\x00\xEC\xB2\x85\x00\xEC\xB2"
-     "\x86\x00"
-     "\xEC\xB2\x87\x00\xEC\xB2\x88\x00\xEC\xB2\x89\x00\xEC\xB2\x8A\x00\xEC\xB2"
-     "\x8B\x00"
-     "\xEC\xB2\x8C\x00\xEC\xB2\x8D\x00\xEC\xB2\x8E\x00\xEC\xB2\x8F\x00\xEC\xB2"
-     "\x90\x00"
-     "\xEC\xB2\x91\x00\xEC\xB2\x92\x00\xEC\xB2\x93\x00\xEC\xB2\x94\x00\xEC\xB2"
-     "\x95\x00"
-     "\xEC\xB2\x96\x00\xEC\xB2\x97\x00\xEC\xB2\x98\x00\xEC\xB2\x99\x00\xEC\xB2"
-     "\x9A\x00"
-     "\xEC\xB2\x9B\x00\xEC\xB2\x9C\x00\xEC\xB2\x9D\x00\xEC\xB2\x9E\x00\xEC\xB2"
-     "\x9F\x00"
-     "\xEC\xB2\xA0\x00\xEC\xB2\xA1\x00\xEC\xB2\xA2\x00\xEC\xB2\xA3\x00\xEC\xB2"
-     "\xA4\x00"
-     "\xEC\xB2\xA5\x00\xEC\xB2\xA6\x00\xEC\xB2\xA7\x00\xEC\xB2\xA8\x00\xEC\xB2"
-     "\xA9\x00"
-     "\xEC\xB2\xAA\x00\xEC\xB2\xAB\x00\xEC\xB2\xAC\x00\xEC\xB2\xAD\x00\xEC\xB2"
-     "\xAE\x00"
-     "\xEC\xB2\xAF\x00\xEC\xB2\xB0\x00\xEC\xB2\xB1\x00\xEC\xB2\xB2\x00\xEC\xB2"
-     "\xB3\x00"
-     "\xEC\xB2\xB4\x00\xEC\xB2\xB5\x00\xEC\xB2\xB6\x00\xEC\xB2\xB7\x00\xEC\xB2"
-     "\xB8\x00"
-     "\xEC\xB2\xB9\x00\xEC\xB2\xBA\x00\xEC\xB2\xBB\x00\xEC\xB2\xBC\x00\xEC\xB2"
-     "\xBD\x00"
-     "\xEC\xB2\xBE\x00\xEC\xB2\xBF\x00\xEC\xB3\x80\x00\xEC\xB3\x81\x00\xEC\xB3"
-     "\x82\x00"
-     "\xEC\xB3\x83\x00\xEC\xB3\x84\x00\xEC\xB3\x85\x00\xEC\xB3\x86\x00\xEC\xB3"
-     "\x87\x00"
-     "\xEC\xB3\x88\x00\xEC\xB3\x89\x00\xEC\xB3\x8A\x00\xEC\xB3\x8B\x00\xEC\xB3"
-     "\x8C\x00"
-     "\xEC\xB3\x8D\x00\xEC\xB3\x8E\x00\xEC\xB3\x8F\x00\xEC\xB3\x90\x00\xEC\xB3"
-     "\x91\x00"
-     "\xEC\xB3\x92\x00\xEC\xB3\x93\x00\xEC\xB3\x94\x00\xEC\xB3\x95\x00\xEC\xB3"
-     "\x96\x00"
-     "\xEC\xB3\x97\x00\xEC\xB3\x98\x00\xEC\xB3\x99\x00\xEC\xB3\x9A\x00\xEC\xB3"
-     "\x9B\x00"
-     "\xEC\xB3\x9C\x00\xEC\xB3\x9D\x00\xEC\xB3\x9E\x00\xEC\xB3\x9F\x00\xEC\xB3"
-     "\xA0\x00"
-     "\xEC\xB3\xA1\x00\xEC\xB3\xA2\x00\xEC\xB3\xA3\x00\xEC\xB3\xA4\x00\xEC\xB3"
-     "\xA5\x00"
-     "\xEC\xB3\xA6\x00\xEC\xB3\xA7\x00\xEC\xB3\xA8\x00\xEC\xB3\xA9\x00\xEC\xB3"
-     "\xAA\x00"
-     "\xEC\xB3\xAB\x00\xEC\xB3\xAC\x00\xEC\xB3\xAD\x00\xEC\xB3\xAE\x00\xEC\xB3"
-     "\xAF\x00"
-     "\xEC\xB3\xB0\x00\xEC\xB3\xB1\x00\xEC\xB3\xB2\x00\xEC\xB3\xB3\x00\xEC\xB3"
-     "\xB4\x00"
-     "\xEC\xB3\xB5\x00\xEC\xB3\xB6\x00\xEC\xB3\xB7\x00\xEC\xB3\xB8\x00\xEC\xB3"
-     "\xB9\x00"
-     "\xEC\xB3\xBA\x00\xEC\xB3\xBB\x00\xEC\xB3\xBC\x00\xEC\xB3\xBD\x00\xEC\xB3"
-     "\xBE\x00"
-     "\xEC\xB3\xBF\x00\xEC\xB4\x80\x00\xEC\xB4\x81\x00\xEC\xB4\x82\x00\xEC\xB4"
-     "\x83\x00"
-     "\xEC\xB4\x84\x00\xEC\xB4\x85\x00\xEC\xB4\x86\x00\xEC\xB4\x87\x00\xEC\xB4"
-     "\x88\x00"
-     "\xEC\xB4\x89\x00\xEC\xB4\x8A\x00\xEC\xB4\x8B\x00\xEC\xB4\x8C\x00\xEC\xB4"
-     "\x8D\x00"
-     "\xEC\xB4\x8E\x00\xEC\xB4\x8F\x00\xEC\xB4\x90\x00\xEC\xB4\x91\x00\xEC\xB4"
-     "\x92\x00"
-     "\xEC\xB4\x93\x00\xEC\xB4\x94\x00\xEC\xB4\x95\x00\xEC\xB4\x96\x00\xEC\xB4"
-     "\x97\x00"
-     "\xEC\xB4\x98\x00\xEC\xB4\x99\x00\xEC\xB4\x9A\x00\xEC\xB4\x9B\x00\xEC\xB4"
-     "\x9C\x00"
-     "\xEC\xB4\x9D\x00\xEC\xB4\x9E\x00\xEC\xB4\x9F\x00\xEC\xB4\xA0\x00\xEC\xB4"
-     "\xA1\x00"
-     "\xEC\xB4\xA2\x00\xEC\xB4\xA3\x00\xEC\xB4\xA4\x00\xEC\xB4\xA5\x00\xEC\xB4"
-     "\xA6\x00"
-     "\xEC\xB4\xA7\x00\xEC\xB4\xA8\x00\xEC\xB4\xA9\x00\xEC\xB4\xAA\x00\xEC\xB4"
-     "\xAB\x00"
-     "\xEC\xB4\xAC\x00\xEC\xB4\xAD\x00\xEC\xB4\xAE\x00\xEC\xB4\xAF\x00\xEC\xB4"
-     "\xB0\x00"
-     "\xEC\xB4\xB1\x00\xEC\xB4\xB2\x00\xEC\xB4\xB3\x00\xEC\xB4\xB4\x00\xEC\xB4"
-     "\xB5\x00"
-     "\xEC\xB4\xB6\x00\xEC\xB4\xB7\x00\xEC\xB4\xB8\x00\xEC\xB4\xB9\x00\xEC\xB4"
-     "\xBA\x00"
-     "\xEC\xB4\xBB\x00\xEC\xB4\xBC\x00\xEC\xB4\xBD\x00\xEC\xB4\xBE\x00\xEC\xB4"
-     "\xBF\x00"
-     "\xEC\xB5\x80\x00\xEC\xB5\x81\x00\xEC\xB5\x82\x00\xEC\xB5\x83\x00\xEC\xB5"
-     "\x84\x00"
-     "\xEC\xB5\x85\x00\xEC\xB5\x86\x00\xEC\xB5\x87\x00\xEC\xB5\x88\x00\xEC\xB5"
-     "\x89\x00"
-     "\xEC\xB5\x8A\x00\xEC\xB5\x8B\x00\xEC\xB5\x8C\x00\xEC\xB5\x8D\x00\xEC\xB5"
-     "\x8E\x00"
-     "\xEC\xB5\x8F\x00\xEC\xB5\x90\x00\xEC\xB5\x91\x00\xEC\xB5\x92\x00\xEC\xB5"
-     "\x93\x00"
-     "\xEC\xB5\x94\x00\xEC\xB5\x95\x00\xEC\xB5\x96\x00\xEC\xB5\x97\x00\xEC\xB5"
-     "\x98\x00"
-     "\xEC\xB5\x99\x00\xEC\xB5\x9A\x00\xEC\xB5\x9B\x00\xEC\xB5\x9C\x00\xEC\xB5"
-     "\x9D\x00"
-     "\xEC\xB5\x9E\x00\xEC\xB5\x9F\x00\xEC\xB5\xA0\x00\xEC\xB5\xA1\x00\xEC\xB5"
-     "\xA2\x00"
-     "\xEC\xB5\xA3\x00\xEC\xB5\xA4\x00\xEC\xB5\xA5\x00\xEC\xB5\xA6\x00\xEC\xB5"
-     "\xA7\x00"
-     "\xEC\xB5\xA8\x00\xEC\xB5\xA9\x00\xEC\xB5\xAA\x00\xEC\xB5\xAB\x00\xEC\xB5"
-     "\xAC\x00"
-     "\xEC\xB5\xAD\x00\xEC\xB5\xAE\x00\xEC\xB5\xAF\x00\xEC\xB5\xB0\x00\xEC\xB5"
-     "\xB1\x00"
-     "\xEC\xB5\xB2\x00\xEC\xB5\xB3\x00\xEC\xB5\xB4\x00\xEC\xB5\xB5\x00\xEC\xB5"
-     "\xB6\x00"
-     "\xEC\xB5\xB7\x00\xEC\xB5\xB8\x00\xEC\xB5\xB9\x00\xEC\xB5\xBA\x00\xEC\xB5"
-     "\xBB\x00"
-     "\xEC\xB5\xBC\x00\xEC\xB5\xBD\x00\xEC\xB5\xBE\x00\xEC\xB5\xBF\x00\xEC\xB6"
-     "\x80\x00"
-     "\xEC\xB6\x81\x00\xEC\xB6\x82\x00\xEC\xB6\x83\x00\xEC\xB6\x84\x00\xEC\xB6"
-     "\x85\x00"
-     "\xEC\xB6\x86\x00\xEC\xB6\x87\x00\xEC\xB6\x88\x00\xEC\xB6\x89\x00\xEC\xB6"
-     "\x8A\x00"
-     "\xEC\xB6\x8B\x00\xEC\xB6\x8C\x00\xEC\xB6\x8D\x00\xEC\xB6\x8E\x00\xEC\xB6"
-     "\x8F\x00"
-     "\xEC\xB6\x90\x00\xEC\xB6\x91\x00\xEC\xB6\x92\x00\xEC\xB6\x93\x00\xEC\xB6"
-     "\x94\x00"
-     "\xEC\xB6\x95\x00\xEC\xB6\x96\x00\xEC\xB6\x97\x00\xEC\xB6\x98\x00\xEC\xB6"
-     "\x99\x00"
-     "\xEC\xB6\x9A\x00\xEC\xB6\x9B\x00\xEC\xB6\x9C\x00\xEC\xB6\x9D\x00\xEC\xB6"
-     "\x9E\x00"
-     "\xEC\xB6\x9F\x00\xEC\xB6\xA0\x00\xEC\xB6\xA1\x00\xEC\xB6\xA2\x00\xEC\xB6"
-     "\xA3\x00"
-     "\xEC\xB6\xA4\x00\xEC\xB6\xA5\x00\xEC\xB6\xA6\x00\xEC\xB6\xA7\x00\xEC\xB6"
-     "\xA8\x00"
-     "\xEC\xB6\xA9\x00\xEC\xB6\xAA\x00\xEC\xB6\xAB\x00\xEC\xB6\xAC\x00\xEC\xB6"
-     "\xAD\x00"
-     "\xEC\xB6\xAE\x00\xEC\xB6\xAF\x00\xEC\xB6\xB0\x00\xEC\xB6\xB1\x00\xEC\xB6"
-     "\xB2\x00"
-     "\xEC\xB6\xB3\x00\xEC\xB6\xB4\x00\xEC\xB6\xB5\x00\xEC\xB6\xB6\x00\xEC\xB6"
-     "\xB7\x00"
-     "\xEC\xB6\xB8\x00\xEC\xB6\xB9\x00\xEC\xB6\xBA\x00\xEC\xB6\xBB\x00\xEC\xB6"
-     "\xBC\x00"
-     "\xEC\xB6\xBD\x00\xEC\xB6\xBE\x00\xEC\xB6\xBF\x00\xEC\xB7\x80\x00\xEC\xB7"
-     "\x81\x00"
-     "\xEC\xB7\x82\x00\xEC\xB7\x83\x00\xEC\xB7\x84\x00\xEC\xB7\x85\x00\xEC\xB7"
-     "\x86\x00"
-     "\xEC\xB7\x87\x00\xEC\xB7\x88\x00\xEC\xB7\x89\x00\xEC\xB7\x8A\x00\xEC\xB7"
-     "\x8B\x00"
-     "\xEC\xB7\x8C\x00\xEC\xB7\x8D\x00\xEC\xB7\x8E\x00\xEC\xB7\x8F\x00\xEC\xB7"
-     "\x90\x00"
-     "\xEC\xB7\x91\x00\xEC\xB7\x92\x00\xEC\xB7\x93\x00\xEC\xB7\x94\x00\xEC\xB7"
-     "\x95\x00"
-     "\xEC\xB7\x96\x00\xEC\xB7\x97\x00\xEC\xB7\x98\x00\xEC\xB7\x99\x00\xEC\xB7"
-     "\x9A\x00"
-     "\xEC\xB7\x9B\x00\xEC\xB7\x9C\x00\xEC\xB7\x9D\x00\xEC\xB7\x9E\x00\xEC\xB7"
-     "\x9F\x00"
-     "\xEC\xB7\xA0\x00\xEC\xB7\xA1\x00\xEC\xB7\xA2\x00\xEC\xB7\xA3\x00\xEC\xB7"
-     "\xA4\x00"
-     "\xEC\xB7\xA5\x00\xEC\xB7\xA6\x00\xEC\xB7\xA7\x00\xEC\xB7\xA8\x00\xEC\xB7"
-     "\xA9\x00"
-     "\xEC\xB7\xAA\x00\xEC\xB7\xAB\x00\xEC\xB7\xAC\x00\xEC\xB7\xAD\x00\xEC\xB7"
-     "\xAE\x00"
-     "\xEC\xB7\xAF\x00\xEC\xB7\xB0\x00\xEC\xB7\xB1\x00\xEC\xB7\xB2\x00\xEC\xB7"
-     "\xB3\x00"
-     "\xEC\xB7\xB4\x00\xEC\xB7\xB5\x00\xEC\xB7\xB6\x00\xEC\xB7\xB7\x00\xEC\xB7"
-     "\xB8\x00"
-     "\xEC\xB7\xB9\x00\xEC\xB7\xBA\x00\xEC\xB7\xBB\x00\xEC\xB7\xBC\x00\xEC\xB7"
-     "\xBD\x00"
-     "\xEC\xB7\xBE\x00\xEC\xB7\xBF\x00\xEC\xB8\x80\x00\xEC\xB8\x81\x00\xEC\xB8"
-     "\x82\x00"
-     "\xEC\xB8\x83\x00\xEC\xB8\x84\x00\xEC\xB8\x85\x00\xEC\xB8\x86\x00\xEC\xB8"
-     "\x87\x00"
-     "\xEC\xB8\x88\x00\xEC\xB8\x89\x00\xEC\xB8\x8A\x00\xEC\xB8\x8B\x00\xEC\xB8"
-     "\x8C\x00"
-     "\xEC\xB8\x8D\x00\xEC\xB8\x8E\x00\xEC\xB8\x8F\x00\xEC\xB8\x90\x00\xEC\xB8"
-     "\x91\x00"
-     "\xEC\xB8\x92\x00\xEC\xB8\x93\x00\xEC\xB8\x94\x00\xEC\xB8\x95\x00\xEC\xB8"
-     "\x96\x00"
-     "\xEC\xB8\x97\x00\xEC\xB8\x98\x00\xEC\xB8\x99\x00\xEC\xB8\x9A\x00\xEC\xB8"
-     "\x9B\x00"
-     "\xEC\xB8\x9C\x00\xEC\xB8\x9D\x00\xEC\xB8\x9E\x00\xEC\xB8\x9F\x00\xEC\xB8"
-     "\xA0\x00"
-     "\xEC\xB8\xA1\x00\xEC\xB8\xA2\x00\xEC\xB8\xA3\x00\xEC\xB8\xA4\x00\xEC\xB8"
-     "\xA5\x00"
-     "\xEC\xB8\xA6\x00\xEC\xB8\xA7\x00\xEC\xB8\xA8\x00\xEC\xB8\xA9\x00\xEC\xB8"
-     "\xAA\x00"
-     "\xEC\xB8\xAB\x00\xEC\xB8\xAC\x00\xEC\xB8\xAD\x00\xEC\xB8\xAE\x00\xEC\xB8"
-     "\xAF\x00"
-     "\xEC\xB8\xB0\x00\xEC\xB8\xB1\x00\xEC\xB8\xB2\x00\xEC\xB8\xB3\x00\xEC\xB8"
-     "\xB4\x00"
-     "\xEC\xB8\xB5\x00\xEC\xB8\xB6\x00\xEC\xB8\xB7\x00\xEC\xB8\xB8\x00\xEC\xB8"
-     "\xB9\x00"
-     "\xEC\xB8\xBA\x00\xEC\xB8\xBB\x00\xEC\xB8\xBC\x00\xEC\xB8\xBD\x00\xEC\xB8"
-     "\xBE\x00"
-     "\xEC\xB8\xBF\x00\xEC\xB9\x80\x00\xEC\xB9\x81\x00\xEC\xB9\x82\x00\xEC\xB9"
-     "\x83\x00"
-     "\xEC\xB9\x84\x00\xEC\xB9\x85\x00\xEC\xB9\x86\x00\xEC\xB9\x87\x00\xEC\xB9"
-     "\x88\x00"
-     "\xEC\xB9\x89\x00\xEC\xB9\x8A\x00\xEC\xB9\x8B\x00\xEC\xB9\x8C\x00\xEC\xB9"
-     "\x8D\x00"
-     "\xEC\xB9\x8E\x00\xEC\xB9\x8F\x00\xEC\xB9\x90\x00\xEC\xB9\x91\x00\xEC\xB9"
-     "\x92\x00"
-     "\xEC\xB9\x93\x00\xEC\xB9\x94\x00\xEC\xB9\x95\x00\xEC\xB9\x96\x00\xEC\xB9"
-     "\x97\x00"
-     "\xEC\xB9\x98\x00\xEC\xB9\x99\x00\xEC\xB9\x9A\x00\xEC\xB9\x9B\x00\xEC\xB9"
-     "\x9C\x00"
-     "\xEC\xB9\x9D\x00\xEC\xB9\x9E\x00\xEC\xB9\x9F\x00\xEC\xB9\xA0\x00\xEC\xB9"
-     "\xA1\x00"
-     "\xEC\xB9\xA2\x00\xEC\xB9\xA3\x00\xEC\xB9\xA4\x00\xEC\xB9\xA5\x00\xEC\xB9"
-     "\xA6\x00"
-     "\xEC\xB9\xA7\x00\xEC\xB9\xA8\x00\xEC\xB9\xA9\x00\xEC\xB9\xAA\x00\xEC\xB9"
-     "\xAB\x00"
-     "\xEC\xB9\xAC\x00\xEC\xB9\xAD\x00\xEC\xB9\xAE\x00\xEC\xB9\xAF\x00\xEC\xB9"
-     "\xB0\x00"
-     "\xEC\xB9\xB1\x00\xEC\xB9\xB2\x00\xEC\xB9\xB3\x00\xEC\xB9\xB4\x00\xEC\xB9"
-     "\xB5\x00"
-     "\xEC\xB9\xB6\x00\xEC\xB9\xB7\x00\xEC\xB9\xB8\x00\xEC\xB9\xB9\x00\xEC\xB9"
-     "\xBA\x00"
-     "\xEC\xB9\xBB\x00\xEC\xB9\xBC\x00\xEC\xB9\xBD\x00\xEC\xB9\xBE\x00\xEC\xB9"
-     "\xBF\x00"
-     "\xEC\xBA\x80\x00\xEC\xBA\x81\x00\xEC\xBA\x82\x00\xEC\xBA\x83\x00\xEC\xBA"
-     "\x84\x00"
-     "\xEC\xBA\x85\x00\xEC\xBA\x86\x00\xEC\xBA\x87\x00\xEC\xBA\x88\x00\xEC\xBA"
-     "\x89\x00"
-     "\xEC\xBA\x8A\x00\xEC\xBA\x8B\x00\xEC\xBA\x8C\x00\xEC\xBA\x8D\x00\xEC\xBA"
-     "\x8E\x00"
-     "\xEC\xBA\x8F\x00\xEC\xBA\x90\x00\xEC\xBA\x91\x00\xEC\xBA\x92\x00\xEC\xBA"
-     "\x93\x00"
-     "\xEC\xBA\x94\x00\xEC\xBA\x95\x00\xEC\xBA\x96\x00\xEC\xBA\x97\x00\xEC\xBA"
-     "\x98\x00"
-     "\xEC\xBA\x99\x00\xEC\xBA\x9A\x00\xEC\xBA\x9B\x00\xEC\xBA\x9C\x00\xEC\xBA"
-     "\x9D\x00"
-     "\xEC\xBA\x9E\x00\xEC\xBA\x9F\x00\xEC\xBA\xA0\x00\xEC\xBA\xA1\x00\xEC\xBA"
-     "\xA2\x00"
-     "\xEC\xBA\xA3\x00\xEC\xBA\xA4\x00\xEC\xBA\xA5\x00\xEC\xBA\xA6\x00\xEC\xBA"
-     "\xA7\x00"
-     "\xEC\xBA\xA8\x00\xEC\xBA\xA9\x00\xEC\xBA\xAA\x00\xEC\xBA\xAB\x00\xEC\xBA"
-     "\xAC\x00"
-     "\xEC\xBA\xAD\x00\xEC\xBA\xAE\x00\xEC\xBA\xAF\x00\xEC\xBA\xB0\x00\xEC\xBA"
-     "\xB1\x00"
-     "\xEC\xBA\xB2\x00\xEC\xBA\xB3\x00\xEC\xBA\xB4\x00\xEC\xBA\xB5\x00\xEC\xBA"
-     "\xB6\x00"
-     "\xEC\xBA\xB7\x00\xEC\xBA\xB8\x00\xEC\xBA\xB9\x00\xEC\xBA\xBA\x00\xEC\xBA"
-     "\xBB\x00"
-     "\xEC\xBA\xBC\x00\xEC\xBA\xBD\x00\xEC\xBA\xBE\x00\xEC\xBA\xBF\x00\xEC\xBB"
-     "\x80\x00"
-     "\xEC\xBB\x81\x00\xEC\xBB\x82\x00\xEC\xBB\x83\x00\xEC\xBB\x84\x00\xEC\xBB"
-     "\x85\x00"
-     "\xEC\xBB\x86\x00\xEC\xBB\x87\x00\xEC\xBB\x88\x00\xEC\xBB\x89\x00\xEC\xBB"
-     "\x8A\x00"
-     "\xEC\xBB\x8B\x00\xEC\xBB\x8C\x00\xEC\xBB\x8D\x00\xEC\xBB\x8E\x00\xEC\xBB"
-     "\x8F\x00"
-     "\xEC\xBB\x90\x00\xEC\xBB\x91\x00\xEC\xBB\x92\x00\xEC\xBB\x93\x00\xEC\xBB"
-     "\x94\x00"
-     "\xEC\xBB\x95\x00\xEC\xBB\x96\x00\xEC\xBB\x97\x00\xEC\xBB\x98\x00\xEC\xBB"
-     "\x99\x00"
-     "\xEC\xBB\x9A\x00\xEC\xBB\x9B\x00\xEC\xBB\x9C\x00\xEC\xBB\x9D\x00\xEC\xBB"
-     "\x9E\x00"
-     "\xEC\xBB\x9F\x00\xEC\xBB\xA0\x00\xEC\xBB\xA1\x00\xEC\xBB\xA2\x00\xEC\xBB"
-     "\xA3\x00"
-     "\xEC\xBB\xA4\x00\xEC\xBB\xA5\x00\xEC\xBB\xA6\x00\xEC\xBB\xA7\x00\xEC\xBB"
-     "\xA8\x00"
-     "\xEC\xBB\xA9\x00\xEC\xBB\xAA\x00\xEC\xBB\xAB\x00\xEC\xBB\xAC\x00\xEC\xBB"
-     "\xAD\x00"
-     "\xEC\xBB\xAE\x00\xEC\xBB\xAF\x00\xEC\xBB\xB0\x00\xEC\xBB\xB1\x00\xEC\xBB"
-     "\xB2\x00"
-     "\xEC\xBB\xB3\x00\xEC\xBB\xB4\x00\xEC\xBB\xB5\x00\xEC\xBB\xB6\x00\xEC\xBB"
-     "\xB7\x00"
-     "\xEC\xBB\xB8\x00\xEC\xBB\xB9\x00\xEC\xBB\xBA\x00\xEC\xBB\xBB\x00\xEC\xBB"
-     "\xBC\x00"
-     "\xEC\xBB\xBD\x00\xEC\xBB\xBE\x00\xEC\xBB\xBF\x00\xEC\xBC\x80\x00\xEC\xBC"
-     "\x81\x00"
-     "\xEC\xBC\x82\x00\xEC\xBC\x83\x00\xEC\xBC\x84\x00\xEC\xBC\x85\x00\xEC\xBC"
-     "\x86\x00"
-     "\xEC\xBC\x87\x00\xEC\xBC\x88\x00\xEC\xBC\x89\x00\xEC\xBC\x8A\x00\xEC\xBC"
-     "\x8B\x00"
-     "\xEC\xBC\x8C\x00\xEC\xBC\x8D\x00\xEC\xBC\x8E\x00\xEC\xBC\x8F\x00\xEC\xBC"
-     "\x90\x00"
-     "\xEC\xBC\x91\x00\xEC\xBC\x92\x00\xEC\xBC\x93\x00\xEC\xBC\x94\x00\xEC\xBC"
-     "\x95\x00"
-     "\xEC\xBC\x96\x00\xEC\xBC\x97\x00\xEC\xBC\x98\x00\xEC\xBC\x99\x00\xEC\xBC"
-     "\x9A\x00"
-     "\xEC\xBC\x9B\x00\xEC\xBC\x9C\x00\xEC\xBC\x9D\x00\xEC\xBC\x9E\x00\xEC\xBC"
-     "\x9F\x00"
-     "\xEC\xBC\xA0\x00\xEC\xBC\xA1\x00\xEC\xBC\xA2\x00\xEC\xBC\xA3\x00\xEC\xBC"
-     "\xA4\x00"
-     "\xEC\xBC\xA5\x00\xEC\xBC\xA6\x00\xEC\xBC\xA7\x00\xEC\xBC\xA8\x00\xEC\xBC"
-     "\xA9\x00"
-     "\xEC\xBC\xAA\x00\xEC\xBC\xAB\x00\xEC\xBC\xAC\x00\xEC\xBC\xAD\x00\xEC\xBC"
-     "\xAE\x00"
-     "\xEC\xBC\xAF\x00\xEC\xBC\xB0\x00\xEC\xBC\xB1\x00\xEC\xBC\xB2\x00\xEC\xBC"
-     "\xB3\x00"
-     "\xEC\xBC\xB4\x00\xEC\xBC\xB5\x00\xEC\xBC\xB6\x00\xEC\xBC\xB7\x00\xEC\xBC"
-     "\xB8\x00"
-     "\xEC\xBC\xB9\x00\xEC\xBC\xBA\x00\xEC\xBC\xBB\x00\xEC\xBC\xBC\x00\xEC\xBC"
-     "\xBD\x00"
-     "\xEC\xBC\xBE\x00\xEC\xBC\xBF\x00\xEC\xBD\x80\x00\xEC\xBD\x81\x00\xEC\xBD"
-     "\x82\x00"
-     "\xEC\xBD\x83\x00\xEC\xBD\x84\x00\xEC\xBD\x85\x00\xEC\xBD\x86\x00\xEC\xBD"
-     "\x87\x00"
-     "\xEC\xBD\x88\x00\xEC\xBD\x89\x00\xEC\xBD\x8A\x00\xEC\xBD\x8B\x00\xEC\xBD"
-     "\x8C\x00"
-     "\xEC\xBD\x8D\x00\xEC\xBD\x8E\x00\xEC\xBD\x8F\x00\xEC\xBD\x90\x00\xEC\xBD"
-     "\x91\x00"
-     "\xEC\xBD\x92\x00\xEC\xBD\x93\x00\xEC\xBD\x94\x00\xEC\xBD\x95\x00\xEC\xBD"
-     "\x96\x00"
-     "\xEC\xBD\x97\x00\xEC\xBD\x98\x00\xEC\xBD\x99\x00\xEC\xBD\x9A\x00\xEC\xBD"
-     "\x9B\x00"
-     "\xEC\xBD\x9C\x00\xEC\xBD\x9D\x00\xEC\xBD\x9E\x00\xEC\xBD\x9F\x00\xEC\xBD"
-     "\xA0\x00"
-     "\xEC\xBD\xA1\x00\xEC\xBD\xA2\x00\xEC\xBD\xA3\x00\xEC\xBD\xA4\x00\xEC\xBD"
-     "\xA5\x00"
-     "\xEC\xBD\xA6\x00\xEC\xBD\xA7\x00\xEC\xBD\xA8\x00\xEC\xBD\xA9\x00\xEC\xBD"
-     "\xAA\x00"
-     "\xEC\xBD\xAB\x00\xEC\xBD\xAC\x00\xEC\xBD\xAD\x00\xEC\xBD\xAE\x00\xEC\xBD"
-     "\xAF\x00"
-     "\xEC\xBD\xB0\x00\xEC\xBD\xB1\x00\xEC\xBD\xB2\x00\xEC\xBD\xB3\x00\xEC\xBD"
-     "\xB4\x00"
-     "\xEC\xBD\xB5\x00\xEC\xBD\xB6\x00\xEC\xBD\xB7\x00\xEC\xBD\xB8\x00\xEC\xBD"
-     "\xB9\x00"
-     "\xEC\xBD\xBA\x00\xEC\xBD\xBB\x00\xEC\xBD\xBC\x00\xEC\xBD\xBD\x00\xEC\xBD"
-     "\xBE\x00"
-     "\xEC\xBD\xBF\x00\xEC\xBE\x80\x00\xEC\xBE\x81\x00\xEC\xBE\x82\x00\xEC\xBE"
-     "\x83\x00"
-     "\xEC\xBE\x84\x00\xEC\xBE\x85\x00\xEC\xBE\x86\x00\xEC\xBE\x87\x00\xEC\xBE"
-     "\x88\x00"
-     "\xEC\xBE\x89\x00\xEC\xBE\x8A\x00\xEC\xBE\x8B\x00\xEC\xBE\x8C\x00\xEC\xBE"
-     "\x8D\x00"
-     "\xEC\xBE\x8E\x00\xEC\xBE\x8F\x00\xEC\xBE\x90\x00\xEC\xBE\x91\x00\xEC\xBE"
-     "\x92\x00"
-     "\xEC\xBE\x93\x00\xEC\xBE\x94\x00\xEC\xBE\x95\x00\xEC\xBE\x96\x00\xEC\xBE"
-     "\x97\x00"
-     "\xEC\xBE\x98\x00\xEC\xBE\x99\x00\xEC\xBE\x9A\x00\xEC\xBE\x9B\x00\xEC\xBE"
-     "\x9C\x00"
-     "\xEC\xBE\x9D\x00\xEC\xBE\x9E\x00\xEC\xBE\x9F\x00\xEC\xBE\xA0\x00\xEC\xBE"
-     "\xA1\x00"
-     "\xEC\xBE\xA2\x00\xEC\xBE\xA3\x00\xEC\xBE\xA4\x00\xEC\xBE\xA5\x00\xEC\xBE"
-     "\xA6\x00"
-     "\xEC\xBE\xA7\x00\xEC\xBE\xA8\x00\xEC\xBE\xA9\x00\xEC\xBE\xAA\x00\xEC\xBE"
-     "\xAB\x00"
-     "\xEC\xBE\xAC\x00\xEC\xBE\xAD\x00\xEC\xBE\xAE\x00\xEC\xBE\xAF\x00\xEC\xBE"
-     "\xB0\x00"
-     "\xEC\xBE\xB1\x00\xEC\xBE\xB2\x00\xEC\xBE\xB3\x00\xEC\xBE\xB4\x00\xEC\xBE"
-     "\xB5\x00"
-     "\xEC\xBE\xB6\x00\xEC\xBE\xB7\x00\xEC\xBE\xB8\x00\xEC\xBE\xB9\x00\xEC\xBE"
-     "\xBA\x00"
-     "\xEC\xBE\xBB\x00\xEC\xBE\xBC\x00\xEC\xBE\xBD\x00\xEC\xBE\xBE\x00\xEC\xBE"
-     "\xBF\x00"
-     "\xEC\xBF\x80\x00\xEC\xBF\x81\x00\xEC\xBF\x82\x00\xEC\xBF\x83\x00\xEC\xBF"
-     "\x84\x00"
-     "\xEC\xBF\x85\x00\xEC\xBF\x86\x00\xEC\xBF\x87\x00\xEC\xBF\x88\x00\xEC\xBF"
-     "\x89\x00"
-     "\xEC\xBF\x8A\x00\xEC\xBF\x8B\x00\xEC\xBF\x8C\x00\xEC\xBF\x8D\x00\xEC\xBF"
-     "\x8E\x00"
-     "\xEC\xBF\x8F\x00\xEC\xBF\x90\x00\xEC\xBF\x91\x00\xEC\xBF\x92\x00\xEC\xBF"
-     "\x93\x00"
-     "\xEC\xBF\x94\x00\xEC\xBF\x95\x00\xEC\xBF\x96\x00\xEC\xBF\x97\x00\xEC\xBF"
-     "\x98\x00"
-     "\xEC\xBF\x99\x00\xEC\xBF\x9A\x00\xEC\xBF\x9B\x00\xEC\xBF\x9C\x00\xEC\xBF"
-     "\x9D\x00"
-     "\xEC\xBF\x9E\x00\xEC\xBF\x9F\x00\xEC\xBF\xA0\x00\xEC\xBF\xA1\x00\xEC\xBF"
-     "\xA2\x00"
-     "\xEC\xBF\xA3\x00\xEC\xBF\xA4\x00\xEC\xBF\xA5\x00\xEC\xBF\xA6\x00\xEC\xBF"
-     "\xA7\x00"
-     "\xEC\xBF\xA8\x00\xEC\xBF\xA9\x00\xEC\xBF\xAA\x00\xEC\xBF\xAB\x00\xEC\xBF"
-     "\xAC\x00"
-     "\xEC\xBF\xAD\x00\xEC\xBF\xAE\x00\xEC\xBF\xAF\x00\xEC\xBF\xB0\x00\xEC\xBF"
-     "\xB1\x00"
-     "\xEC\xBF\xB2\x00\xEC\xBF\xB3\x00\xEC\xBF\xB4\x00\xEC\xBF\xB5\x00\xEC\xBF"
-     "\xB6\x00"
-     "\xEC\xBF\xB7\x00\xEC\xBF\xB8\x00\xEC\xBF\xB9\x00\xEC\xBF\xBA\x00\xEC\xBF"
-     "\xBB\x00"
-     "\xEC\xBF\xBC\x00\xEC\xBF\xBD\x00\xEC\xBF\xBE\x00\xEC\xBF\xBF\x00\xED\x80"
-     "\x80\x00"
-     "\xED\x80\x81\x00\xED\x80\x82\x00\xED\x80\x83\x00\xED\x80\x84\x00\xED\x80"
-     "\x85\x00"
-     "\xED\x80\x86\x00\xED\x80\x87\x00\xED\x80\x88\x00\xED\x80\x89\x00\xED\x80"
-     "\x8A\x00"
-     "\xED\x80\x8B\x00\xED\x80\x8C\x00\xED\x80\x8D\x00\xED\x80\x8E\x00\xED\x80"
-     "\x8F\x00"
-     "\xED\x80\x90\x00\xED\x80\x91\x00\xED\x80\x92\x00\xED\x80\x93\x00\xED\x80"
-     "\x94\x00"
-     "\xED\x80\x95\x00\xED\x80\x96\x00\xED\x80\x97\x00\xED\x80\x98\x00\xED\x80"
-     "\x99\x00"
-     "\xED\x80\x9A\x00\xED\x80\x9B\x00\xED\x80\x9C\x00\xED\x80\x9D\x00\xED\x80"
-     "\x9E\x00"
-     "\xED\x80\x9F\x00\xED\x80\xA0\x00\xED\x80\xA1\x00\xED\x80\xA2\x00\xED\x80"
-     "\xA3\x00"
-     "\xED\x80\xA4\x00\xED\x80\xA5\x00\xED\x80\xA6\x00\xED\x80\xA7\x00\xED\x80"
-     "\xA8\x00"
-     "\xED\x80\xA9\x00\xED\x80\xAA\x00\xED\x80\xAB\x00\xED\x80\xAC\x00\xED\x80"
-     "\xAD\x00"
-     "\xED\x80\xAE\x00\xED\x80\xAF\x00\xED\x80\xB0\x00\xED\x80\xB1\x00\xED\x80"
-     "\xB2\x00"
-     "\xED\x80\xB3\x00\xED\x80\xB4\x00\xED\x80\xB5\x00\xED\x80\xB6\x00\xED\x80"
-     "\xB7\x00"
-     "\xED\x80\xB8\x00\xED\x80\xB9\x00\xED\x80\xBA\x00\xED\x80\xBB\x00\xED\x80"
-     "\xBC\x00"
-     "\xED\x80\xBD\x00\xED\x80\xBE\x00\xED\x80\xBF\x00\xED\x81\x80\x00\xED\x81"
-     "\x81\x00"
-     "\xED\x81\x82\x00\xED\x81\x83\x00\xED\x81\x84\x00\xED\x81\x85\x00\xED\x81"
-     "\x86\x00"
-     "\xED\x81\x87\x00\xED\x81\x88\x00\xED\x81\x89\x00\xED\x81\x8A\x00\xED\x81"
-     "\x8B\x00"
-     "\xED\x81\x8C\x00\xED\x81\x8D\x00\xED\x81\x8E\x00\xED\x81\x8F\x00\xED\x81"
-     "\x90\x00"
-     "\xED\x81\x91\x00\xED\x81\x92\x00\xED\x81\x93\x00\xED\x81\x94\x00\xED\x81"
-     "\x95\x00"
-     "\xED\x81\x96\x00\xED\x81\x97\x00\xED\x81\x98\x00\xED\x81\x99\x00\xED\x81"
-     "\x9A\x00"
-     "\xED\x81\x9B\x00\xED\x81\x9C\x00\xED\x81\x9D\x00\xED\x81\x9E\x00\xED\x81"
-     "\x9F\x00"
-     "\xED\x81\xA0\x00\xED\x81\xA1\x00\xED\x81\xA2\x00\xED\x81\xA3\x00\xED\x81"
-     "\xA4\x00"
-     "\xED\x81\xA5\x00\xED\x81\xA6\x00\xED\x81\xA7\x00\xED\x81\xA8\x00\xED\x81"
-     "\xA9\x00"
-     "\xED\x81\xAA\x00\xED\x81\xAB\x00\xED\x81\xAC\x00\xED\x81\xAD\x00\xED\x81"
-     "\xAE\x00"
-     "\xED\x81\xAF\x00\xED\x81\xB0\x00\xED\x81\xB1\x00\xED\x81\xB2\x00\xED\x81"
-     "\xB3\x00"
-     "\xED\x81\xB4\x00\xED\x81\xB5\x00\xED\x81\xB6\x00\xED\x81\xB7\x00\xED\x81"
-     "\xB8\x00"
-     "\xED\x81\xB9\x00\xED\x81\xBA\x00\xED\x81\xBB\x00\xED\x81\xBC\x00\xED\x81"
-     "\xBD\x00"
-     "\xED\x81\xBE\x00\xED\x81\xBF\x00\xED\x82\x80\x00\xED\x82\x81\x00\xED\x82"
-     "\x82\x00"
-     "\xED\x82\x83\x00\xED\x82\x84\x00\xED\x82\x85\x00\xED\x82\x86\x00\xED\x82"
-     "\x87\x00"
-     "\xED\x82\x88\x00\xED\x82\x89\x00\xED\x82\x8A\x00\xED\x82\x8B\x00\xED\x82"
-     "\x8C\x00"
-     "\xED\x82\x8D\x00\xED\x82\x8E\x00\xED\x82\x8F\x00\xED\x82\x90\x00\xED\x82"
-     "\x91\x00"
-     "\xED\x82\x92\x00\xED\x82\x93\x00\xED\x82\x94\x00\xED\x82\x95\x00\xED\x82"
-     "\x96\x00"
-     "\xED\x82\x97\x00\xED\x82\x98\x00\xED\x82\x99\x00\xED\x82\x9A\x00\xED\x82"
-     "\x9B\x00"
-     "\xED\x82\x9C\x00\xED\x82\x9D\x00\xED\x82\x9E\x00\xED\x82\x9F\x00\xED\x82"
-     "\xA0\x00"
-     "\xED\x82\xA1\x00\xED\x82\xA2\x00\xED\x82\xA3\x00\xED\x82\xA4\x00\xED\x82"
-     "\xA5\x00"
-     "\xED\x82\xA6\x00\xED\x82\xA7\x00\xED\x82\xA8\x00\xED\x82\xA9\x00\xED\x82"
-     "\xAA\x00"
-     "\xED\x82\xAB\x00\xED\x82\xAC\x00\xED\x82\xAD\x00\xED\x82\xAE\x00\xED\x82"
-     "\xAF\x00"
-     "\xED\x82\xB0\x00\xED\x82\xB1\x00\xED\x82\xB2\x00\xED\x82\xB3\x00\xED\x82"
-     "\xB4\x00"
-     "\xED\x82\xB5\x00\xED\x82\xB6\x00\xED\x82\xB7\x00\xED\x82\xB8\x00\xED\x82"
-     "\xB9\x00"
-     "\xED\x82\xBA\x00\xED\x82\xBB\x00\xED\x82\xBC\x00\xED\x82\xBD\x00\xED\x82"
-     "\xBE\x00"
-     "\xED\x82\xBF\x00\xED\x83\x80\x00\xED\x83\x81\x00\xED\x83\x82\x00\xED\x83"
-     "\x83\x00"
-     "\xED\x83\x84\x00\xED\x83\x85\x00\xED\x83\x86\x00\xED\x83\x87\x00\xED\x83"
-     "\x88\x00"
-     "\xED\x83\x89\x00\xED\x83\x8A\x00\xED\x83\x8B\x00\xED\x83\x8C\x00\xED\x83"
-     "\x8D\x00"
-     "\xED\x83\x8E\x00\xED\x83\x8F\x00\xED\x83\x90\x00\xED\x83\x91\x00\xED\x83"
-     "\x92\x00"
-     "\xED\x83\x93\x00\xED\x83\x94\x00\xED\x83\x95\x00\xED\x83\x96\x00\xED\x83"
-     "\x97\x00"
-     "\xED\x83\x98\x00\xED\x83\x99\x00\xED\x83\x9A\x00\xED\x83\x9B\x00\xED\x83"
-     "\x9C\x00"
-     "\xED\x83\x9D\x00\xED\x83\x9E\x00\xED\x83\x9F\x00\xED\x83\xA0\x00\xED\x83"
-     "\xA1\x00"
-     "\xED\x83\xA2\x00\xED\x83\xA3\x00\xED\x83\xA4\x00\xED\x83\xA5\x00\xED\x83"
-     "\xA6\x00"
-     "\xED\x83\xA7\x00\xED\x83\xA8\x00\xED\x83\xA9\x00\xED\x83\xAA\x00\xED\x83"
-     "\xAB\x00"
-     "\xED\x83\xAC\x00\xED\x83\xAD\x00\xED\x83\xAE\x00\xED\x83\xAF\x00\xED\x83"
-     "\xB0\x00"
-     "\xED\x83\xB1\x00\xED\x83\xB2\x00\xED\x83\xB3\x00\xED\x83\xB4\x00\xED\x83"
-     "\xB5\x00"
-     "\xED\x83\xB6\x00\xED\x83\xB7\x00\xED\x83\xB8\x00\xED\x83\xB9\x00\xED\x83"
-     "\xBA\x00"
-     "\xED\x83\xBB\x00\xED\x83\xBC\x00\xED\x83\xBD\x00\xED\x83\xBE\x00\xED\x83"
-     "\xBF\x00"
-     "\xED\x84\x80\x00\xED\x84\x81\x00\xED\x84\x82\x00\xED\x84\x83\x00\xED\x84"
-     "\x84\x00"
-     "\xED\x84\x85\x00\xED\x84\x86\x00\xED\x84\x87\x00\xED\x84\x88\x00\xED\x84"
-     "\x89\x00"
-     "\xED\x84\x8A\x00\xED\x84\x8B\x00\xED\x84\x8C\x00\xED\x84\x8D\x00\xED\x84"
-     "\x8E\x00"
-     "\xED\x84\x8F\x00\xED\x84\x90\x00\xED\x84\x91\x00\xED\x84\x92\x00\xED\x84"
-     "\x93\x00"
-     "\xED\x84\x94\x00\xED\x84\x95\x00\xED\x84\x96\x00\xED\x84\x97\x00\xED\x84"
-     "\x98\x00"
-     "\xED\x84\x99\x00\xED\x84\x9A\x00\xED\x84\x9B\x00\xED\x84\x9C\x00\xED\x84"
-     "\x9D\x00"
-     "\xED\x84\x9E\x00\xED\x84\x9F\x00\xED\x84\xA0\x00\xED\x84\xA1\x00\xED\x84"
-     "\xA2\x00"
-     "\xED\x84\xA3\x00\xED\x84\xA4\x00\xED\x84\xA5\x00\xED\x84\xA6\x00\xED\x84"
-     "\xA7\x00"
-     "\xED\x84\xA8\x00\xED\x84\xA9\x00\xED\x84\xAA\x00\xED\x84\xAB\x00\xED\x84"
-     "\xAC\x00"
-     "\xED\x84\xAD\x00\xED\x84\xAE\x00\xED\x84\xAF\x00\xED\x84\xB0\x00\xED\x84"
-     "\xB1\x00"
-     "\xED\x84\xB2\x00\xED\x84\xB3\x00\xED\x84\xB4\x00\xED\x84\xB5\x00\xED\x84"
-     "\xB6\x00"
-     "\xED\x84\xB7\x00\xED\x84\xB8\x00\xED\x84\xB9\x00\xED\x84\xBA\x00\xED\x84"
-     "\xBB\x00"
-     "\xED\x84\xBC\x00\xED\x84\xBD\x00\xED\x84\xBE\x00\xED\x84\xBF\x00\xED\x85"
-     "\x80\x00"
-     "\xED\x85\x81\x00\xED\x85\x82\x00\xED\x85\x83\x00\xED\x85\x84\x00\xED\x85"
-     "\x85\x00"
-     "\xED\x85\x86\x00\xED\x85\x87\x00\xED\x85\x88\x00\xED\x85\x89\x00\xED\x85"
-     "\x8A\x00"
-     "\xED\x85\x8B\x00\xED\x85\x8C\x00\xED\x85\x8D\x00\xED\x85\x8E\x00\xED\x85"
-     "\x8F\x00"
-     "\xED\x85\x90\x00\xED\x85\x91\x00\xED\x85\x92\x00\xED\x85\x93\x00\xED\x85"
-     "\x94\x00"
-     "\xED\x85\x95\x00\xED\x85\x96\x00\xED\x85\x97\x00\xED\x85\x98\x00\xED\x85"
-     "\x99\x00"
-     "\xED\x85\x9A\x00\xED\x85\x9B\x00\xED\x85\x9C\x00\xED\x85\x9D\x00\xED\x85"
-     "\x9E\x00"
-     "\xED\x85\x9F\x00\xED\x85\xA0\x00\xED\x85\xA1\x00\xED\x85\xA2\x00\xED\x85"
-     "\xA3\x00"
-     "\xED\x85\xA4\x00\xED\x85\xA5\x00\xED\x85\xA6\x00\xED\x85\xA7\x00\xED\x85"
-     "\xA8\x00"
-     "\xED\x85\xA9\x00\xED\x85\xAA\x00\xED\x85\xAB\x00\xED\x85\xAC\x00\xED\x85"
-     "\xAD\x00"
-     "\xED\x85\xAE\x00\xED\x85\xAF\x00\xED\x85\xB0\x00\xED\x85\xB1\x00\xED\x85"
-     "\xB2\x00"
-     "\xED\x85\xB3\x00\xED\x85\xB4\x00\xED\x85\xB5\x00\xED\x85\xB6\x00\xED\x85"
-     "\xB7\x00"
-     "\xED\x85\xB8\x00\xED\x85\xB9\x00\xED\x85\xBA\x00\xED\x85\xBB\x00\xED\x85"
-     "\xBC\x00"
-     "\xED\x85\xBD\x00\xED\x85\xBE\x00\xED\x85\xBF\x00\xED\x86\x80\x00\xED\x86"
-     "\x81\x00"
-     "\xED\x86\x82\x00\xED\x86\x83\x00\xED\x86\x84\x00\xED\x86\x85\x00\xED\x86"
-     "\x86\x00"
-     "\xED\x86\x87\x00\xED\x86\x88\x00\xED\x86\x89\x00\xED\x86\x8A\x00\xED\x86"
-     "\x8B\x00"
-     "\xED\x86\x8C\x00\xED\x86\x8D\x00\xED\x86\x8E\x00\xED\x86\x8F\x00\xED\x86"
-     "\x90\x00"
-     "\xED\x86\x91\x00\xED\x86\x92\x00\xED\x86\x93\x00\xED\x86\x94\x00\xED\x86"
-     "\x95\x00"
-     "\xED\x86\x96\x00\xED\x86\x97\x00\xED\x86\x98\x00\xED\x86\x99\x00\xED\x86"
-     "\x9A\x00"
-     "\xED\x86\x9B\x00\xED\x86\x9C\x00\xED\x86\x9D\x00\xED\x86\x9E\x00\xED\x86"
-     "\x9F\x00"
-     "\xED\x86\xA0\x00\xED\x86\xA1\x00\xED\x86\xA2\x00\xED\x86\xA3\x00\xED\x86"
-     "\xA4\x00"
-     "\xED\x86\xA5\x00\xED\x86\xA6\x00\xED\x86\xA7\x00\xED\x86\xA8\x00\xED\x86"
-     "\xA9\x00"
-     "\xED\x86\xAA\x00\xED\x86\xAB\x00\xED\x86\xAC\x00\xED\x86\xAD\x00\xED\x86"
-     "\xAE\x00"
-     "\xED\x86\xAF\x00\xED\x86\xB0\x00\xED\x86\xB1\x00\xED\x86\xB2\x00\xED\x86"
-     "\xB3\x00"
-     "\xED\x86\xB4\x00\xED\x86\xB5\x00\xED\x86\xB6\x00\xED\x86\xB7\x00\xED\x86"
-     "\xB8\x00"
-     "\xED\x86\xB9\x00\xED\x86\xBA\x00\xED\x86\xBB\x00\xED\x86\xBC\x00\xED\x86"
-     "\xBD\x00"
-     "\xED\x86\xBE\x00\xED\x86\xBF\x00\xED\x87\x80\x00\xED\x87\x81\x00\xED\x87"
-     "\x82\x00"
-     "\xED\x87\x83\x00\xED\x87\x84\x00\xED\x87\x85\x00\xED\x87\x86\x00\xED\x87"
-     "\x87\x00"
-     "\xED\x87\x88\x00\xED\x87\x89\x00\xED\x87\x8A\x00\xED\x87\x8B\x00\xED\x87"
-     "\x8C\x00"
-     "\xED\x87\x8D\x00\xED\x87\x8E\x00\xED\x87\x8F\x00\xED\x87\x90\x00\xED\x87"
-     "\x91\x00"
-     "\xED\x87\x92\x00\xED\x87\x93\x00\xED\x87\x94\x00\xED\x87\x95\x00\xED\x87"
-     "\x96\x00"
-     "\xED\x87\x97\x00\xED\x87\x98\x00\xED\x87\x99\x00\xED\x87\x9A\x00\xED\x87"
-     "\x9B\x00"
-     "\xED\x87\x9C\x00\xED\x87\x9D\x00\xED\x87\x9E\x00\xED\x87\x9F\x00\xED\x87"
-     "\xA0\x00"
-     "\xED\x87\xA1\x00\xED\x87\xA2\x00\xED\x87\xA3\x00\xED\x87\xA4\x00\xED\x87"
-     "\xA5\x00"
-     "\xED\x87\xA6\x00\xED\x87\xA7\x00\xED\x87\xA8\x00\xED\x87\xA9\x00\xED\x87"
-     "\xAA\x00"
-     "\xED\x87\xAB\x00\xED\x87\xAC\x00\xED\x87\xAD\x00\xED\x87\xAE\x00\xED\x87"
-     "\xAF\x00"
-     "\xED\x87\xB0\x00\xED\x87\xB1\x00\xED\x87\xB2\x00\xED\x87\xB3\x00\xED\x87"
-     "\xB4\x00"
-     "\xED\x87\xB5\x00\xED\x87\xB6\x00\xED\x87\xB7\x00\xED\x87\xB8\x00\xED\x87"
-     "\xB9\x00"
-     "\xED\x87\xBA\x00\xED\x87\xBB\x00\xED\x87\xBC\x00\xED\x87\xBD\x00\xED\x87"
-     "\xBE\x00"
-     "\xED\x87\xBF\x00\xED\x88\x80\x00\xED\x88\x81\x00\xED\x88\x82\x00\xED\x88"
-     "\x83\x00"
-     "\xED\x88\x84\x00\xED\x88\x85\x00\xED\x88\x86\x00\xED\x88\x87\x00\xED\x88"
-     "\x88\x00"
-     "\xED\x88\x89\x00\xED\x88\x8A\x00\xED\x88\x8B\x00\xED\x88\x8C\x00\xED\x88"
-     "\x8D\x00"
-     "\xED\x88\x8E\x00\xED\x88\x8F\x00\xED\x88\x90\x00\xED\x88\x91\x00\xED\x88"
-     "\x92\x00"
-     "\xED\x88\x93\x00\xED\x88\x94\x00\xED\x88\x95\x00\xED\x88\x96\x00\xED\x88"
-     "\x97\x00"
-     "\xED\x88\x98\x00\xED\x88\x99\x00\xED\x88\x9A\x00\xED\x88\x9B\x00\xED\x88"
-     "\x9C\x00"
-     "\xED\x88\x9D\x00\xED\x88\x9E\x00\xED\x88\x9F\x00\xED\x88\xA0\x00\xED\x88"
-     "\xA1\x00"
-     "\xED\x88\xA2\x00\xED\x88\xA3\x00\xED\x88\xA4\x00\xED\x88\xA5\x00\xED\x88"
-     "\xA6\x00"
-     "\xED\x88\xA7\x00\xED\x88\xA8\x00\xED\x88\xA9\x00\xED\x88\xAA\x00\xED\x88"
-     "\xAB\x00"
-     "\xED\x88\xAC\x00\xED\x88\xAD\x00\xED\x88\xAE\x00\xED\x88\xAF\x00\xED\x88"
-     "\xB0\x00"
-     "\xED\x88\xB1\x00\xED\x88\xB2\x00\xED\x88\xB3\x00\xED\x88\xB4\x00\xED\x88"
-     "\xB5\x00"
-     "\xED\x88\xB6\x00\xED\x88\xB7\x00\xED\x88\xB8\x00\xED\x88\xB9\x00\xED\x88"
-     "\xBA\x00"
-     "\xED\x88\xBB\x00\xED\x88\xBC\x00\xED\x88\xBD\x00\xED\x88\xBE\x00\xED\x88"
-     "\xBF\x00"
-     "\xED\x89\x80\x00\xED\x89\x81\x00\xED\x89\x82\x00\xED\x89\x83\x00\xED\x89"
-     "\x84\x00"
-     "\xED\x89\x85\x00\xED\x89\x86\x00\xED\x89\x87\x00\xED\x89\x88\x00\xED\x89"
-     "\x89\x00"
-     "\xED\x89\x8A\x00\xED\x89\x8B\x00\xED\x89\x8C\x00\xED\x89\x8D\x00\xED\x89"
-     "\x8E\x00"
-     "\xED\x89\x8F\x00\xED\x89\x90\x00\xED\x89\x91\x00\xED\x89\x92\x00\xED\x89"
-     "\x93\x00"
-     "\xED\x89\x94\x00\xED\x89\x95\x00\xED\x89\x96\x00\xED\x89\x97\x00\xED\x89"
-     "\x98\x00"
-     "\xED\x89\x99\x00\xED\x89\x9A\x00\xED\x89\x9B\x00\xED\x89\x9C\x00\xED\x89"
-     "\x9D\x00"
-     "\xED\x89\x9E\x00\xED\x89\x9F\x00\xED\x89\xA0\x00\xED\x89\xA1\x00\xED\x89"
-     "\xA2\x00"
-     "\xED\x89\xA3\x00\xED\x89\xA4\x00\xED\x89\xA5\x00\xED\x89\xA6\x00\xED\x89"
-     "\xA7\x00"
-     "\xED\x89\xA8\x00\xED\x89\xA9\x00\xED\x89\xAA\x00\xED\x89\xAB\x00\xED\x89"
-     "\xAC\x00"
-     "\xED\x89\xAD\x00\xED\x89\xAE\x00\xED\x89\xAF\x00\xED\x89\xB0\x00\xED\x89"
-     "\xB1\x00"
-     "\xED\x89\xB2\x00\xED\x89\xB3\x00\xED\x89\xB4\x00\xED\x89\xB5\x00\xED\x89"
-     "\xB6\x00"
-     "\xED\x89\xB7\x00\xED\x89\xB8\x00\xED\x89\xB9\x00\xED\x89\xBA\x00\xED\x89"
-     "\xBB\x00"
-     "\xED\x89\xBC\x00\xED\x89\xBD\x00\xED\x89\xBE\x00\xED\x89\xBF\x00\xED\x8A"
-     "\x80\x00"
-     "\xED\x8A\x81\x00\xED\x8A\x82\x00\xED\x8A\x83\x00\xED\x8A\x84\x00\xED\x8A"
-     "\x85\x00"
-     "\xED\x8A\x86\x00\xED\x8A\x87\x00\xED\x8A\x88\x00\xED\x8A\x89\x00\xED\x8A"
-     "\x8A\x00"
-     "\xED\x8A\x8B\x00\xED\x8A\x8C\x00\xED\x8A\x8D\x00\xED\x8A\x8E\x00\xED\x8A"
-     "\x8F\x00"
-     "\xED\x8A\x90\x00\xED\x8A\x91\x00\xED\x8A\x92\x00\xED\x8A\x93\x00\xED\x8A"
-     "\x94\x00"
-     "\xED\x8A\x95\x00\xED\x8A\x96\x00\xED\x8A\x97\x00\xED\x8A\x98\x00\xED\x8A"
-     "\x99\x00"
-     "\xED\x8A\x9A\x00\xED\x8A\x9B\x00\xED\x8A\x9C\x00\xED\x8A\x9D\x00\xED\x8A"
-     "\x9E\x00"
-     "\xED\x8A\x9F\x00\xED\x8A\xA0\x00\xED\x8A\xA1\x00\xED\x8A\xA2\x00\xED\x8A"
-     "\xA3\x00"
-     "\xED\x8A\xA4\x00\xED\x8A\xA5\x00\xED\x8A\xA6\x00\xED\x8A\xA7\x00\xED\x8A"
-     "\xA8\x00"
-     "\xED\x8A\xA9\x00\xED\x8A\xAA\x00\xED\x8A\xAB\x00\xED\x8A\xAC\x00\xED\x8A"
-     "\xAD\x00"
-     "\xED\x8A\xAE\x00\xED\x8A\xAF\x00\xED\x8A\xB0\x00\xED\x8A\xB1\x00\xED\x8A"
-     "\xB2\x00"
-     "\xED\x8A\xB3\x00\xED\x8A\xB4\x00\xED\x8A\xB5\x00\xED\x8A\xB6\x00\xED\x8A"
-     "\xB7\x00"
-     "\xED\x8A\xB8\x00\xED\x8A\xB9\x00\xED\x8A\xBA\x00\xED\x8A\xBB\x00\xED\x8A"
-     "\xBC\x00"
-     "\xED\x8A\xBD\x00\xED\x8A\xBE\x00\xED\x8A\xBF\x00\xED\x8B\x80\x00\xED\x8B"
-     "\x81\x00"
-     "\xED\x8B\x82\x00\xED\x8B\x83\x00\xED\x8B\x84\x00\xED\x8B\x85\x00\xED\x8B"
-     "\x86\x00"
-     "\xED\x8B\x87\x00\xED\x8B\x88\x00\xED\x8B\x89\x00\xED\x8B\x8A\x00\xED\x8B"
-     "\x8B\x00"
-     "\xED\x8B\x8C\x00\xED\x8B\x8D\x00\xED\x8B\x8E\x00\xED\x8B\x8F\x00\xED\x8B"
-     "\x90\x00"
-     "\xED\x8B\x91\x00\xED\x8B\x92\x00\xED\x8B\x93\x00\xED\x8B\x94\x00\xED\x8B"
-     "\x95\x00"
-     "\xED\x8B\x96\x00\xED\x8B\x97\x00\xED\x8B\x98\x00\xED\x8B\x99\x00\xED\x8B"
-     "\x9A\x00"
-     "\xED\x8B\x9B\x00\xED\x8B\x9C\x00\xED\x8B\x9D\x00\xED\x8B\x9E\x00\xED\x8B"
-     "\x9F\x00"
-     "\xED\x8B\xA0\x00\xED\x8B\xA1\x00\xED\x8B\xA2\x00\xED\x8B\xA3\x00\xED\x8B"
-     "\xA4\x00"
-     "\xED\x8B\xA5\x00\xED\x8B\xA6\x00\xED\x8B\xA7\x00\xED\x8B\xA8\x00\xED\x8B"
-     "\xA9\x00"
-     "\xED\x8B\xAA\x00\xED\x8B\xAB\x00\xED\x8B\xAC\x00\xED\x8B\xAD\x00\xED\x8B"
-     "\xAE\x00"
-     "\xED\x8B\xAF\x00\xED\x8B\xB0\x00\xED\x8B\xB1\x00\xED\x8B\xB2\x00\xED\x8B"
-     "\xB3\x00"
-     "\xED\x8B\xB4\x00\xED\x8B\xB5\x00\xED\x8B\xB6\x00\xED\x8B\xB7\x00\xED\x8B"
-     "\xB8\x00"
-     "\xED\x8B\xB9\x00\xED\x8B\xBA\x00\xED\x8B\xBB\x00\xED\x8B\xBC\x00\xED\x8B"
-     "\xBD\x00"
-     "\xED\x8B\xBE\x00\xED\x8B\xBF\x00\xED\x8C\x80\x00\xED\x8C\x81\x00\xED\x8C"
-     "\x82\x00"
-     "\xED\x8C\x83\x00\xED\x8C\x84\x00\xED\x8C\x85\x00\xED\x8C\x86\x00\xED\x8C"
-     "\x87\x00"
-     "\xED\x8C\x88\x00\xED\x8C\x89\x00\xED\x8C\x8A\x00\xED\x8C\x8B\x00\xED\x8C"
-     "\x8C\x00"
-     "\xED\x8C\x8D\x00\xED\x8C\x8E\x00\xED\x8C\x8F\x00\xED\x8C\x90\x00\xED\x8C"
-     "\x91\x00"
-     "\xED\x8C\x92\x00\xED\x8C\x93\x00\xED\x8C\x94\x00\xED\x8C\x95\x00\xED\x8C"
-     "\x96\x00"
-     "\xED\x8C\x97\x00\xED\x8C\x98\x00\xED\x8C\x99\x00\xED\x8C\x9A\x00\xED\x8C"
-     "\x9B\x00"
-     "\xED\x8C\x9C\x00\xED\x8C\x9D\x00\xED\x8C\x9E\x00\xED\x8C\x9F\x00\xED\x8C"
-     "\xA0\x00"
-     "\xED\x8C\xA1\x00\xED\x8C\xA2\x00\xED\x8C\xA3\x00\xED\x8C\xA4\x00\xED\x8C"
-     "\xA5\x00"
-     "\xED\x8C\xA6\x00\xED\x8C\xA7\x00\xED\x8C\xA8\x00\xED\x8C\xA9\x00\xED\x8C"
-     "\xAA\x00"
-     "\xED\x8C\xAB\x00\xED\x8C\xAC\x00\xED\x8C\xAD\x00\xED\x8C\xAE\x00\xED\x8C"
-     "\xAF\x00"
-     "\xED\x8C\xB0\x00\xED\x8C\xB1\x00\xED\x8C\xB2\x00\xED\x8C\xB3\x00\xED\x8C"
-     "\xB4\x00"
-     "\xED\x8C\xB5\x00\xED\x8C\xB6\x00\xED\x8C\xB7\x00\xED\x8C\xB8\x00\xED\x8C"
-     "\xB9\x00"
-     "\xED\x8C\xBA\x00\xED\x8C\xBB\x00\xED\x8C\xBC\x00\xED\x8C\xBD\x00\xED\x8C"
-     "\xBE\x00"
-     "\xED\x8C\xBF\x00\xED\x8D\x80\x00\xED\x8D\x81\x00\xED\x8D\x82\x00\xED\x8D"
-     "\x83\x00"
-     "\xED\x8D\x84\x00\xED\x8D\x85\x00\xED\x8D\x86\x00\xED\x8D\x87\x00\xED\x8D"
-     "\x88\x00"
-     "\xED\x8D\x89\x00\xED\x8D\x8A\x00\xED\x8D\x8B\x00\xED\x8D\x8C\x00\xED\x8D"
-     "\x8D\x00"
-     "\xED\x8D\x8E\x00\xED\x8D\x8F\x00\xED\x8D\x90\x00\xED\x8D\x91\x00\xED\x8D"
-     "\x92\x00"
-     "\xED\x8D\x93\x00\xED\x8D\x94\x00\xED\x8D\x95\x00\xED\x8D\x96\x00\xED\x8D"
-     "\x97\x00"
-     "\xED\x8D\x98\x00\xED\x8D\x99\x00\xED\x8D\x9A\x00\xED\x8D\x9B\x00\xED\x8D"
-     "\x9C\x00"
-     "\xED\x8D\x9D\x00\xED\x8D\x9E\x00\xED\x8D\x9F\x00\xED\x8D\xA0\x00\xED\x8D"
-     "\xA1\x00"
-     "\xED\x8D\xA2\x00\xED\x8D\xA3\x00\xED\x8D\xA4\x00\xED\x8D\xA5\x00\xED\x8D"
-     "\xA6\x00"
-     "\xED\x8D\xA7\x00\xED\x8D\xA8\x00\xED\x8D\xA9\x00\xED\x8D\xAA\x00\xED\x8D"
-     "\xAB\x00"
-     "\xED\x8D\xAC\x00\xED\x8D\xAD\x00\xED\x8D\xAE\x00\xED\x8D\xAF\x00\xED\x8D"
-     "\xB0\x00"
-     "\xED\x8D\xB1\x00\xED\x8D\xB2\x00\xED\x8D\xB3\x00\xED\x8D\xB4\x00\xED\x8D"
-     "\xB5\x00"
-     "\xED\x8D\xB6\x00\xED\x8D\xB7\x00\xED\x8D\xB8\x00\xED\x8D\xB9\x00\xED\x8D"
-     "\xBA\x00"
-     "\xED\x8D\xBB\x00\xED\x8D\xBC\x00\xED\x8D\xBD\x00\xED\x8D\xBE\x00\xED\x8D"
-     "\xBF\x00"
-     "\xED\x8E\x80\x00\xED\x8E\x81\x00\xED\x8E\x82\x00\xED\x8E\x83\x00\xED\x8E"
-     "\x84\x00"
-     "\xED\x8E\x85\x00\xED\x8E\x86\x00\xED\x8E\x87\x00\xED\x8E\x88\x00\xED\x8E"
-     "\x89\x00"
-     "\xED\x8E\x8A\x00\xED\x8E\x8B\x00\xED\x8E\x8C\x00\xED\x8E\x8D\x00\xED\x8E"
-     "\x8E\x00"
-     "\xED\x8E\x8F\x00\xED\x8E\x90\x00\xED\x8E\x91\x00\xED\x8E\x92\x00\xED\x8E"
-     "\x93\x00"
-     "\xED\x8E\x94\x00\xED\x8E\x95\x00\xED\x8E\x96\x00\xED\x8E\x97\x00\xED\x8E"
-     "\x98\x00"
-     "\xED\x8E\x99\x00\xED\x8E\x9A\x00\xED\x8E\x9B\x00\xED\x8E\x9C\x00\xED\x8E"
-     "\x9D\x00"
-     "\xED\x8E\x9E\x00\xED\x8E\x9F\x00\xED\x8E\xA0\x00\xED\x8E\xA1\x00\xED\x8E"
-     "\xA2\x00"
-     "\xED\x8E\xA3\x00\xED\x8E\xA4\x00\xED\x8E\xA5\x00\xED\x8E\xA6\x00\xED\x8E"
-     "\xA7\x00"
-     "\xED\x8E\xA8\x00\xED\x8E\xA9\x00\xED\x8E\xAA\x00\xED\x8E\xAB\x00\xED\x8E"
-     "\xAC\x00"
-     "\xED\x8E\xAD\x00\xED\x8E\xAE\x00\xED\x8E\xAF\x00\xED\x8E\xB0\x00\xED\x8E"
-     "\xB1\x00"
-     "\xED\x8E\xB2\x00\xED\x8E\xB3\x00\xED\x8E\xB4\x00\xED\x8E\xB5\x00\xED\x8E"
-     "\xB6\x00"
-     "\xED\x8E\xB7\x00\xED\x8E\xB8\x00\xED\x8E\xB9\x00\xED\x8E\xBA\x00\xED\x8E"
-     "\xBB\x00"
-     "\xED\x8E\xBC\x00\xED\x8E\xBD\x00\xED\x8E\xBE\x00\xED\x8E\xBF\x00\xED\x8F"
-     "\x80\x00"
-     "\xED\x8F\x81\x00\xED\x8F\x82\x00\xED\x8F\x83\x00\xED\x8F\x84\x00\xED\x8F"
-     "\x85\x00"
-     "\xED\x8F\x86\x00\xED\x8F\x87\x00\xED\x8F\x88\x00\xED\x8F\x89\x00\xED\x8F"
-     "\x8A\x00"
-     "\xED\x8F\x8B\x00\xED\x8F\x8C\x00\xED\x8F\x8D\x00\xED\x8F\x8E\x00\xED\x8F"
-     "\x8F\x00"
-     "\xED\x8F\x90\x00\xED\x8F\x91\x00\xED\x8F\x92\x00\xED\x8F\x93\x00\xED\x8F"
-     "\x94\x00"
-     "\xED\x8F\x95\x00\xED\x8F\x96\x00\xED\x8F\x97\x00\xED\x8F\x98\x00\xED\x8F"
-     "\x99\x00"
-     "\xED\x8F\x9A\x00\xED\x8F\x9B\x00\xED\x8F\x9C\x00\xED\x8F\x9D\x00\xED\x8F"
-     "\x9E\x00"
-     "\xED\x8F\x9F\x00\xED\x8F\xA0\x00\xED\x8F\xA1\x00\xED\x8F\xA2\x00\xED\x8F"
-     "\xA3\x00"
-     "\xED\x8F\xA4\x00\xED\x8F\xA5\x00\xED\x8F\xA6\x00\xED\x8F\xA7\x00\xED\x8F"
-     "\xA8\x00"
-     "\xED\x8F\xA9\x00\xED\x8F\xAA\x00\xED\x8F\xAB\x00\xED\x8F\xAC\x00\xED\x8F"
-     "\xAD\x00"
-     "\xED\x8F\xAE\x00\xED\x8F\xAF\x00\xED\x8F\xB0\x00\xED\x8F\xB1\x00\xED\x8F"
-     "\xB2\x00"
-     "\xED\x8F\xB3\x00\xED\x8F\xB4\x00\xED\x8F\xB5\x00\xED\x8F\xB6\x00\xED\x8F"
-     "\xB7\x00"
-     "\xED\x8F\xB8\x00\xED\x8F\xB9\x00\xED\x8F\xBA\x00\xED\x8F\xBB\x00\xED\x8F"
-     "\xBC\x00"
-     "\xED\x8F\xBD\x00\xED\x8F\xBE\x00\xED\x8F\xBF\x00\xED\x90\x80\x00\xED\x90"
-     "\x81\x00"
-     "\xED\x90\x82\x00\xED\x90\x83\x00\xED\x90\x84\x00\xED\x90\x85\x00\xED\x90"
-     "\x86\x00"
-     "\xED\x90\x87\x00\xED\x90\x88\x00\xED\x90\x89\x00\xED\x90\x8A\x00\xED\x90"
-     "\x8B\x00"
-     "\xED\x90\x8C\x00\xED\x90\x8D\x00\xED\x90\x8E\x00\xED\x90\x8F\x00\xED\x90"
-     "\x90\x00"
-     "\xED\x90\x91\x00\xED\x90\x92\x00\xED\x90\x93\x00\xED\x90\x94\x00\xED\x90"
-     "\x95\x00"
-     "\xED\x90\x96\x00\xED\x90\x97\x00\xED\x90\x98\x00\xED\x90\x99\x00\xED\x90"
-     "\x9A\x00"
-     "\xED\x90\x9B\x00\xED\x90\x9C\x00\xED\x90\x9D\x00\xED\x90\x9E\x00\xED\x90"
-     "\x9F\x00"
-     "\xED\x90\xA0\x00\xED\x90\xA1\x00\xED\x90\xA2\x00\xED\x90\xA3\x00\xED\x90"
-     "\xA4\x00"
-     "\xED\x90\xA5\x00\xED\x90\xA6\x00\xED\x90\xA7\x00\xED\x90\xA8\x00\xED\x90"
-     "\xA9\x00"
-     "\xED\x90\xAA\x00\xED\x90\xAB\x00\xED\x90\xAC\x00\xED\x90\xAD\x00\xED\x90"
-     "\xAE\x00"
-     "\xED\x90\xAF\x00\xED\x90\xB0\x00\xED\x90\xB1\x00\xED\x90\xB2\x00\xED\x90"
-     "\xB3\x00"
-     "\xED\x90\xB4\x00\xED\x90\xB5\x00\xED\x90\xB6\x00\xED\x90\xB7\x00\xED\x90"
-     "\xB8\x00"
-     "\xED\x90\xB9\x00\xED\x90\xBA\x00\xED\x90\xBB\x00\xED\x90\xBC\x00\xED\x90"
-     "\xBD\x00"
-     "\xED\x90\xBE\x00\xED\x90\xBF\x00\xED\x91\x80\x00\xED\x91\x81\x00\xED\x91"
-     "\x82\x00"
-     "\xED\x91\x83\x00\xED\x91\x84\x00\xED\x91\x85\x00\xED\x91\x86\x00\xED\x91"
-     "\x87\x00"
-     "\xED\x91\x88\x00\xED\x91\x89\x00\xED\x91\x8A\x00\xED\x91\x8B\x00\xED\x91"
-     "\x8C\x00"
-     "\xED\x91\x8D\x00\xED\x91\x8E\x00\xED\x91\x8F\x00\xED\x91\x90\x00\xED\x91"
-     "\x91\x00"
-     "\xED\x91\x92\x00\xED\x91\x93\x00\xED\x91\x94\x00\xED\x91\x95\x00\xED\x91"
-     "\x96\x00"
-     "\xED\x91\x97\x00\xED\x91\x98\x00\xED\x91\x99\x00\xED\x91\x9A\x00\xED\x91"
-     "\x9B\x00"
-     "\xED\x91\x9C\x00\xED\x91\x9D\x00\xED\x91\x9E\x00\xED\x91\x9F\x00\xED\x91"
-     "\xA0\x00"
-     "\xED\x91\xA1\x00\xED\x91\xA2\x00\xED\x91\xA3\x00\xED\x91\xA4\x00\xED\x91"
-     "\xA5\x00"
-     "\xED\x91\xA6\x00\xED\x91\xA7\x00\xED\x91\xA8\x00\xED\x91\xA9\x00\xED\x91"
-     "\xAA\x00"
-     "\xED\x91\xAB\x00\xED\x91\xAC\x00\xED\x91\xAD\x00\xED\x91\xAE\x00\xED\x91"
-     "\xAF\x00"
-     "\xED\x91\xB0\x00\xED\x91\xB1\x00\xED\x91\xB2\x00\xED\x91\xB3\x00\xED\x91"
-     "\xB4\x00"
-     "\xED\x91\xB5\x00\xED\x91\xB6\x00\xED\x91\xB7\x00\xED\x91\xB8\x00\xED\x91"
-     "\xB9\x00"
-     "\xED\x91\xBA\x00\xED\x91\xBB\x00\xED\x91\xBC\x00\xED\x91\xBD\x00\xED\x91"
-     "\xBE\x00"
-     "\xED\x91\xBF\x00\xED\x92\x80\x00\xED\x92\x81\x00\xED\x92\x82\x00\xED\x92"
-     "\x83\x00"
-     "\xED\x92\x84\x00\xED\x92\x85\x00\xED\x92\x86\x00\xED\x92\x87\x00\xED\x92"
-     "\x88\x00"
-     "\xED\x92\x89\x00\xED\x92\x8A\x00\xED\x92\x8B\x00\xED\x92\x8C\x00\xED\x92"
-     "\x8D\x00"
-     "\xED\x92\x8E\x00\xED\x92\x8F\x00\xED\x92\x90\x00\xED\x92\x91\x00\xED\x92"
-     "\x92\x00"
-     "\xED\x92\x93\x00\xED\x92\x94\x00\xED\x92\x95\x00\xED\x92\x96\x00\xED\x92"
-     "\x97\x00"
-     "\xED\x92\x98\x00\xED\x92\x99\x00\xED\x92\x9A\x00\xED\x92\x9B\x00\xED\x92"
-     "\x9C\x00"
-     "\xED\x92\x9D\x00\xED\x92\x9E\x00\xED\x92\x9F\x00\xED\x92\xA0\x00\xED\x92"
-     "\xA1\x00"
-     "\xED\x92\xA2\x00\xED\x92\xA3\x00\xED\x92\xA4\x00\xED\x92\xA5\x00\xED\x92"
-     "\xA6\x00"
-     "\xED\x92\xA7\x00\xED\x92\xA8\x00\xED\x92\xA9\x00\xED\x92\xAA\x00\xED\x92"
-     "\xAB\x00"
-     "\xED\x92\xAC\x00\xED\x92\xAD\x00\xED\x92\xAE\x00\xED\x92\xAF\x00\xED\x92"
-     "\xB0\x00"
-     "\xED\x92\xB1\x00\xED\x92\xB2\x00\xED\x92\xB3\x00\xED\x92\xB4\x00\xED\x92"
-     "\xB5\x00"
-     "\xED\x92\xB6\x00\xED\x92\xB7\x00\xED\x92\xB8\x00\xED\x92\xB9\x00\xED\x92"
-     "\xBA\x00"
-     "\xED\x92\xBB\x00\xED\x92\xBC\x00\xED\x92\xBD\x00\xED\x92\xBE\x00\xED\x92"
-     "\xBF\x00"
-     "\xED\x93\x80\x00\xED\x93\x81\x00\xED\x93\x82\x00\xED\x93\x83\x00\xED\x93"
-     "\x84\x00"
-     "\xED\x93\x85\x00\xED\x93\x86\x00\xED\x93\x87\x00\xED\x93\x88\x00\xED\x93"
-     "\x89\x00"
-     "\xED\x93\x8A\x00\xED\x93\x8B\x00\xED\x93\x8C\x00\xED\x93\x8D\x00\xED\x93"
-     "\x8E\x00"
-     "\xED\x93\x8F\x00\xED\x93\x90\x00\xED\x93\x91\x00\xED\x93\x92\x00\xED\x93"
-     "\x93\x00"
-     "\xED\x93\x94\x00\xED\x93\x95\x00\xED\x93\x96\x00\xED\x93\x97\x00\xED\x93"
-     "\x98\x00"
-     "\xED\x93\x99\x00\xED\x93\x9A\x00\xED\x93\x9B\x00\xED\x93\x9C\x00\xED\x93"
-     "\x9D\x00"
-     "\xED\x93\x9E\x00\xED\x93\x9F\x00\xED\x93\xA0\x00\xED\x93\xA1\x00\xED\x93"
-     "\xA2\x00"
-     "\xED\x93\xA3\x00\xED\x93\xA4\x00\xED\x93\xA5\x00\xED\x93\xA6\x00\xED\x93"
-     "\xA7\x00"
-     "\xED\x93\xA8\x00\xED\x93\xA9\x00\xED\x93\xAA\x00\xED\x93\xAB\x00\xED\x93"
-     "\xAC\x00"
-     "\xED\x93\xAD\x00\xED\x93\xAE\x00\xED\x93\xAF\x00\xED\x93\xB0\x00\xED\x93"
-     "\xB1\x00"
-     "\xED\x93\xB2\x00\xED\x93\xB3\x00\xED\x93\xB4\x00\xED\x93\xB5\x00\xED\x93"
-     "\xB6\x00"
-     "\xED\x93\xB7\x00\xED\x93\xB8\x00\xED\x93\xB9\x00\xED\x93\xBA\x00\xED\x93"
-     "\xBB\x00"
-     "\xED\x93\xBC\x00\xED\x93\xBD\x00\xED\x93\xBE\x00\xED\x93\xBF\x00\xED\x94"
-     "\x80\x00"
-     "\xED\x94\x81\x00\xED\x94\x82\x00\xED\x94\x83\x00\xED\x94\x84\x00\xED\x94"
-     "\x85\x00"
-     "\xED\x94\x86\x00\xED\x94\x87\x00\xED\x94\x88\x00\xED\x94\x89\x00\xED\x94"
-     "\x8A\x00"
-     "\xED\x94\x8B\x00\xED\x94\x8C\x00\xED\x94\x8D\x00\xED\x94\x8E\x00\xED\x94"
-     "\x8F\x00"
-     "\xED\x94\x90\x00\xED\x94\x91\x00\xED\x94\x92\x00\xED\x94\x93\x00\xED\x94"
-     "\x94\x00"
-     "\xED\x94\x95\x00\xED\x94\x96\x00\xED\x94\x97\x00\xED\x94\x98\x00\xED\x94"
-     "\x99\x00"
-     "\xED\x94\x9A\x00\xED\x94\x9B\x00\xED\x94\x9C\x00\xED\x94\x9D\x00\xED\x94"
-     "\x9E\x00"
-     "\xED\x94\x9F\x00\xED\x94\xA0\x00\xED\x94\xA1\x00\xED\x94\xA2\x00\xED\x94"
-     "\xA3\x00"
-     "\xED\x94\xA4\x00\xED\x94\xA5\x00\xED\x94\xA6\x00\xED\x94\xA7\x00\xED\x94"
-     "\xA8\x00"
-     "\xED\x94\xA9\x00\xED\x94\xAA\x00\xED\x94\xAB\x00\xED\x94\xAC\x00\xED\x94"
-     "\xAD\x00"
-     "\xED\x94\xAE\x00\xED\x94\xAF\x00\xED\x94\xB0\x00\xED\x94\xB1\x00\xED\x94"
-     "\xB2\x00"
-     "\xED\x94\xB3\x00\xED\x94\xB4\x00\xED\x94\xB5\x00\xED\x94\xB6\x00\xED\x94"
-     "\xB7\x00"
-     "\xED\x94\xB8\x00\xED\x94\xB9\x00\xED\x94\xBA\x00\xED\x94\xBB\x00\xED\x94"
-     "\xBC\x00"
-     "\xED\x94\xBD\x00\xED\x94\xBE\x00\xED\x94\xBF\x00\xED\x95\x80\x00\xED\x95"
-     "\x81\x00"
-     "\xED\x95\x82\x00\xED\x95\x83\x00\xED\x95\x84\x00\xED\x95\x85\x00\xED\x95"
-     "\x86\x00"
-     "\xED\x95\x87\x00\xED\x95\x88\x00\xED\x95\x89\x00\xED\x95\x8A\x00\xED\x95"
-     "\x8B\x00"
-     "\xED\x95\x8C\x00\xED\x95\x8D\x00\xED\x95\x8E\x00\xED\x95\x8F\x00\xED\x95"
-     "\x90\x00"
-     "\xED\x95\x91\x00\xED\x95\x92\x00\xED\x95\x93\x00\xED\x95\x94\x00\xED\x95"
-     "\x95\x00"
-     "\xED\x95\x96\x00\xED\x95\x97\x00\xED\x95\x98\x00\xED\x95\x99\x00\xED\x95"
-     "\x9A\x00"
-     "\xED\x95\x9B\x00\xED\x95\x9C\x00\xED\x95\x9D\x00\xED\x95\x9E\x00\xED\x95"
-     "\x9F\x00"
-     "\xED\x95\xA0\x00\xED\x95\xA1\x00\xED\x95\xA2\x00\xED\x95\xA3\x00\xED\x95"
-     "\xA4\x00"
-     "\xED\x95\xA5\x00\xED\x95\xA6\x00\xED\x95\xA7\x00\xED\x95\xA8\x00\xED\x95"
-     "\xA9\x00"
-     "\xED\x95\xAA\x00\xED\x95\xAB\x00\xED\x95\xAC\x00\xED\x95\xAD\x00\xED\x95"
-     "\xAE\x00"
-     "\xED\x95\xAF\x00\xED\x95\xB0\x00\xED\x95\xB1\x00\xED\x95\xB2\x00\xED\x95"
-     "\xB3\x00"
-     "\xED\x95\xB4\x00\xED\x95\xB5\x00\xED\x95\xB6\x00\xED\x95\xB7\x00\xED\x95"
-     "\xB8\x00"
-     "\xED\x95\xB9\x00\xED\x95\xBA\x00\xED\x95\xBB\x00\xED\x95\xBC\x00\xED\x95"
-     "\xBD\x00"
-     "\xED\x95\xBE\x00\xED\x95\xBF\x00\xED\x96\x80\x00\xED\x96\x81\x00\xED\x96"
-     "\x82\x00"
-     "\xED\x96\x83\x00\xED\x96\x84\x00\xED\x96\x85\x00\xED\x96\x86\x00\xED\x96"
-     "\x87\x00"
-     "\xED\x96\x88\x00\xED\x96\x89\x00\xED\x96\x8A\x00\xED\x96\x8B\x00\xED\x96"
-     "\x8C\x00"
-     "\xED\x96\x8D\x00\xED\x96\x8E\x00\xED\x96\x8F\x00\xED\x96\x90\x00\xED\x96"
-     "\x91\x00"
-     "\xED\x96\x92\x00\xED\x96\x93\x00\xED\x96\x94\x00\xED\x96\x95\x00\xED\x96"
-     "\x96\x00"
-     "\xED\x96\x97\x00\xED\x96\x98\x00\xED\x96\x99\x00\xED\x96\x9A\x00\xED\x96"
-     "\x9B\x00"
-     "\xED\x96\x9C\x00\xED\x96\x9D\x00\xED\x96\x9E\x00\xED\x96\x9F\x00\xED\x96"
-     "\xA0\x00"
-     "\xED\x96\xA1\x00\xED\x96\xA2\x00\xED\x96\xA3\x00\xED\x96\xA4\x00\xED\x96"
-     "\xA5\x00"
-     "\xED\x96\xA6\x00\xED\x96\xA7\x00\xED\x96\xA8\x00\xED\x96\xA9\x00\xED\x96"
-     "\xAA\x00"
-     "\xED\x96\xAB\x00\xED\x96\xAC\x00\xED\x96\xAD\x00\xED\x96\xAE\x00\xED\x96"
-     "\xAF\x00"
-     "\xED\x96\xB0\x00\xED\x96\xB1\x00\xED\x96\xB2\x00\xED\x96\xB3\x00\xED\x96"
-     "\xB4\x00"
-     "\xED\x96\xB5\x00\xED\x96\xB6\x00\xED\x96\xB7\x00\xED\x96\xB8\x00\xED\x96"
-     "\xB9\x00"
-     "\xED\x96\xBA\x00\xED\x96\xBB\x00\xED\x96\xBC\x00\xED\x96\xBD\x00\xED\x96"
-     "\xBE\x00"
-     "\xED\x96\xBF\x00\xED\x97\x80\x00\xED\x97\x81\x00\xED\x97\x82\x00\xED\x97"
-     "\x83\x00"
-     "\xED\x97\x84\x00\xED\x97\x85\x00\xED\x97\x86\x00\xED\x97\x87\x00\xED\x97"
-     "\x88\x00"
-     "\xED\x97\x89\x00\xED\x97\x8A\x00\xED\x97\x8B\x00\xED\x97\x8C\x00\xED\x97"
-     "\x8D\x00"
-     "\xED\x97\x8E\x00\xED\x97\x8F\x00\xED\x97\x90\x00\xED\x97\x91\x00\xED\x97"
-     "\x92\x00"
-     "\xED\x97\x93\x00\xED\x97\x94\x00\xED\x97\x95\x00\xED\x97\x96\x00\xED\x97"
-     "\x97\x00"
-     "\xED\x97\x98\x00\xED\x97\x99\x00\xED\x97\x9A\x00\xED\x97\x9B\x00\xED\x97"
-     "\x9C\x00"
-     "\xED\x97\x9D\x00\xED\x97\x9E\x00\xED\x97\x9F\x00\xED\x97\xA0\x00\xED\x97"
-     "\xA1\x00"
-     "\xED\x97\xA2\x00\xED\x97\xA3\x00\xED\x97\xA4\x00\xED\x97\xA5\x00\xED\x97"
-     "\xA6\x00"
-     "\xED\x97\xA7\x00\xED\x97\xA8\x00\xED\x97\xA9\x00\xED\x97\xAA\x00\xED\x97"
-     "\xAB\x00"
-     "\xED\x97\xAC\x00\xED\x97\xAD\x00\xED\x97\xAE\x00\xED\x97\xAF\x00\xED\x97"
-     "\xB0\x00"
-     "\xED\x97\xB1\x00\xED\x97\xB2\x00\xED\x97\xB3\x00\xED\x97\xB4\x00\xED\x97"
-     "\xB5\x00"
-     "\xED\x97\xB6\x00\xED\x97\xB7\x00\xED\x97\xB8\x00\xED\x97\xB9\x00\xED\x97"
-     "\xBA\x00"
-     "\xED\x97\xBB\x00\xED\x97\xBC\x00\xED\x97\xBD\x00\xED\x97\xBE\x00\xED\x97"
-     "\xBF\x00"
-     "\xED\x98\x80\x00\xED\x98\x81\x00\xED\x98\x82\x00\xED\x98\x83\x00\xED\x98"
-     "\x84\x00"
-     "\xED\x98\x85\x00\xED\x98\x86\x00\xED\x98\x87\x00\xED\x98\x88\x00\xED\x98"
-     "\x89\x00"
-     "\xED\x98\x8A\x00\xED\x98\x8B\x00\xED\x98\x8C\x00\xED\x98\x8D\x00\xED\x98"
-     "\x8E\x00"
-     "\xED\x98\x8F\x00\xED\x98\x90\x00\xED\x98\x91\x00\xED\x98\x92\x00\xED\x98"
-     "\x93\x00"
-     "\xED\x98\x94\x00\xED\x98\x95\x00\xED\x98\x96\x00\xED\x98\x97\x00\xED\x98"
-     "\x98\x00"
-     "\xED\x98\x99\x00\xED\x98\x9A\x00\xED\x98\x9B\x00\xED\x98\x9C\x00\xED\x98"
-     "\x9D\x00"
-     "\xED\x98\x9E\x00\xED\x98\x9F\x00\xED\x98\xA0\x00\xED\x98\xA1\x00\xED\x98"
-     "\xA2\x00"
-     "\xED\x98\xA3\x00\xED\x98\xA4\x00\xED\x98\xA5\x00\xED\x98\xA6\x00\xED\x98"
-     "\xA7\x00"
-     "\xED\x98\xA8\x00\xED\x98\xA9\x00\xED\x98\xAA\x00\xED\x98\xAB\x00\xED\x98"
-     "\xAC\x00"
-     "\xED\x98\xAD\x00\xED\x98\xAE\x00\xED\x98\xAF\x00\xED\x98\xB0\x00\xED\x98"
-     "\xB1\x00"
-     "\xED\x98\xB2\x00\xED\x98\xB3\x00\xED\x98\xB4\x00\xED\x98\xB5\x00\xED\x98"
-     "\xB6\x00"
-     "\xED\x98\xB7\x00\xED\x98\xB8\x00\xED\x98\xB9\x00\xED\x98\xBA\x00\xED\x98"
-     "\xBB\x00"
-     "\xED\x98\xBC\x00\xED\x98\xBD\x00\xED\x98\xBE\x00\xED\x98\xBF\x00\xED\x99"
-     "\x80\x00"
-     "\xED\x99\x81\x00\xED\x99\x82\x00\xED\x99\x83\x00\xED\x99\x84\x00\xED\x99"
-     "\x85\x00"
-     "\xED\x99\x86\x00\xED\x99\x87\x00\xED\x99\x88\x00\xED\x99\x89\x00\xED\x99"
-     "\x8A\x00"
-     "\xED\x99\x8B\x00\xED\x99\x8C\x00\xED\x99\x8D\x00\xED\x99\x8E\x00\xED\x99"
-     "\x8F\x00"
-     "\xED\x99\x90\x00\xED\x99\x91\x00\xED\x99\x92\x00\xED\x99\x93\x00\xED\x99"
-     "\x94\x00"
-     "\xED\x99\x95\x00\xED\x99\x96\x00\xED\x99\x97\x00\xED\x99\x98\x00\xED\x99"
-     "\x99\x00"
-     "\xED\x99\x9A\x00\xED\x99\x9B\x00\xED\x99\x9C\x00\xED\x99\x9D\x00\xED\x99"
-     "\x9E\x00"
-     "\xED\x99\x9F\x00\xED\x99\xA0\x00\xED\x99\xA1\x00\xED\x99\xA2\x00\xED\x99"
-     "\xA3\x00"
-     "\xED\x99\xA4\x00\xED\x99\xA5\x00\xED\x99\xA6\x00\xED\x99\xA7\x00\xED\x99"
-     "\xA8\x00"
-     "\xED\x99\xA9\x00\xED\x99\xAA\x00\xED\x99\xAB\x00\xED\x99\xAC\x00\xED\x99"
-     "\xAD\x00"
-     "\xED\x99\xAE\x00\xED\x99\xAF\x00\xED\x99\xB0\x00\xED\x99\xB1\x00\xED\x99"
-     "\xB2\x00"
-     "\xED\x99\xB3\x00\xED\x99\xB4\x00\xED\x99\xB5\x00\xED\x99\xB6\x00\xED\x99"
-     "\xB7\x00"
-     "\xED\x99\xB8\x00\xED\x99\xB9\x00\xED\x99\xBA\x00\xED\x99\xBB\x00\xED\x99"
-     "\xBC\x00"
-     "\xED\x99\xBD\x00\xED\x99\xBE\x00\xED\x99\xBF\x00\xED\x9A\x80\x00\xED\x9A"
-     "\x81\x00"
-     "\xED\x9A\x82\x00\xED\x9A\x83\x00\xED\x9A\x84\x00\xED\x9A\x85\x00\xED\x9A"
-     "\x86\x00"
-     "\xED\x9A\x87\x00\xED\x9A\x88\x00\xED\x9A\x89\x00\xED\x9A\x8A\x00\xED\x9A"
-     "\x8B\x00"
-     "\xED\x9A\x8C\x00\xED\x9A\x8D\x00\xED\x9A\x8E\x00\xED\x9A\x8F\x00\xED\x9A"
-     "\x90\x00"
-     "\xED\x9A\x91\x00\xED\x9A\x92\x00\xED\x9A\x93\x00\xED\x9A\x94\x00\xED\x9A"
-     "\x95\x00"
-     "\xED\x9A\x96\x00\xED\x9A\x97\x00\xED\x9A\x98\x00\xED\x9A\x99\x00\xED\x9A"
-     "\x9A\x00"
-     "\xED\x9A\x9B\x00\xED\x9A\x9C\x00\xED\x9A\x9D\x00\xED\x9A\x9E\x00\xED\x9A"
-     "\x9F\x00"
-     "\xED\x9A\xA0\x00\xED\x9A\xA1\x00\xED\x9A\xA2\x00\xED\x9A\xA3\x00\xED\x9A"
-     "\xA4\x00"
-     "\xED\x9A\xA5\x00\xED\x9A\xA6\x00\xED\x9A\xA7\x00\xED\x9A\xA8\x00\xED\x9A"
-     "\xA9\x00"
-     "\xED\x9A\xAA\x00\xED\x9A\xAB\x00\xED\x9A\xAC\x00\xED\x9A\xAD\x00\xED\x9A"
-     "\xAE\x00"
-     "\xED\x9A\xAF\x00\xED\x9A\xB0\x00\xED\x9A\xB1\x00\xED\x9A\xB2\x00\xED\x9A"
-     "\xB3\x00"
-     "\xED\x9A\xB4\x00\xED\x9A\xB5\x00\xED\x9A\xB6\x00\xED\x9A\xB7\x00\xED\x9A"
-     "\xB8\x00"
-     "\xED\x9A\xB9\x00\xED\x9A\xBA\x00\xED\x9A\xBB\x00\xED\x9A\xBC\x00\xED\x9A"
-     "\xBD\x00"
-     "\xED\x9A\xBE\x00\xED\x9A\xBF\x00\xED\x9B\x80\x00\xED\x9B\x81\x00\xED\x9B"
-     "\x82\x00"
-     "\xED\x9B\x83\x00\xED\x9B\x84\x00\xED\x9B\x85\x00\xED\x9B\x86\x00\xED\x9B"
-     "\x87\x00"
-     "\xED\x9B\x88\x00\xED\x9B\x89\x00\xED\x9B\x8A\x00\xED\x9B\x8B\x00\xED\x9B"
-     "\x8C\x00"
-     "\xED\x9B\x8D\x00\xED\x9B\x8E\x00\xED\x9B\x8F\x00\xED\x9B\x90\x00\xED\x9B"
-     "\x91\x00"
-     "\xED\x9B\x92\x00\xED\x9B\x93\x00\xED\x9B\x94\x00\xED\x9B\x95\x00\xED\x9B"
-     "\x96\x00"
-     "\xED\x9B\x97\x00\xED\x9B\x98\x00\xED\x9B\x99\x00\xED\x9B\x9A\x00\xED\x9B"
-     "\x9B\x00"
-     "\xED\x9B\x9C\x00\xED\x9B\x9D\x00\xED\x9B\x9E\x00\xED\x9B\x9F\x00\xED\x9B"
-     "\xA0\x00"
-     "\xED\x9B\xA1\x00\xED\x9B\xA2\x00\xED\x9B\xA3\x00\xED\x9B\xA4\x00\xED\x9B"
-     "\xA5\x00"
-     "\xED\x9B\xA6\x00\xED\x9B\xA7\x00\xED\x9B\xA8\x00\xED\x9B\xA9\x00\xED\x9B"
-     "\xAA\x00"
-     "\xED\x9B\xAB\x00\xED\x9B\xAC\x00\xED\x9B\xAD\x00\xED\x9B\xAE\x00\xED\x9B"
-     "\xAF\x00"
-     "\xED\x9B\xB0\x00\xED\x9B\xB1\x00\xED\x9B\xB2\x00\xED\x9B\xB3\x00\xED\x9B"
-     "\xB4\x00"
-     "\xED\x9B\xB5\x00\xED\x9B\xB6\x00\xED\x9B\xB7\x00\xED\x9B\xB8\x00\xED\x9B"
-     "\xB9\x00"
-     "\xED\x9B\xBA\x00\xED\x9B\xBB\x00\xED\x9B\xBC\x00\xED\x9B\xBD\x00\xED\x9B"
-     "\xBE\x00"
-     "\xED\x9B\xBF\x00\xED\x9C\x80\x00\xED\x9C\x81\x00\xED\x9C\x82\x00\xED\x9C"
-     "\x83\x00"
-     "\xED\x9C\x84\x00\xED\x9C\x85\x00\xED\x9C\x86\x00\xED\x9C\x87\x00\xED\x9C"
-     "\x88\x00"
-     "\xED\x9C\x89\x00\xED\x9C\x8A\x00\xED\x9C\x8B\x00\xED\x9C\x8C\x00\xED\x9C"
-     "\x8D\x00"
-     "\xED\x9C\x8E\x00\xED\x9C\x8F\x00\xED\x9C\x90\x00\xED\x9C\x91\x00\xED\x9C"
-     "\x92\x00"
-     "\xED\x9C\x93\x00\xED\x9C\x94\x00\xED\x9C\x95\x00\xED\x9C\x96\x00\xED\x9C"
-     "\x97\x00"
-     "\xED\x9C\x98\x00\xED\x9C\x99\x00\xED\x9C\x9A\x00\xED\x9C\x9B\x00\xED\x9C"
-     "\x9C\x00"
-     "\xED\x9C\x9D\x00\xED\x9C\x9E\x00\xED\x9C\x9F\x00\xED\x9C\xA0\x00\xED\x9C"
-     "\xA1\x00"
-     "\xED\x9C\xA2\x00\xED\x9C\xA3\x00\xED\x9C\xA4\x00\xED\x9C\xA5\x00\xED\x9C"
-     "\xA6\x00"
-     "\xED\x9C\xA7\x00\xED\x9C\xA8\x00\xED\x9C\xA9\x00\xED\x9C\xAA\x00\xED\x9C"
-     "\xAB\x00"
-     "\xED\x9C\xAC\x00\xED\x9C\xAD\x00\xED\x9C\xAE\x00\xED\x9C\xAF\x00\xED\x9C"
-     "\xB0\x00"
-     "\xED\x9C\xB1\x00\xED\x9C\xB2\x00\xED\x9C\xB3\x00\xED\x9C\xB4\x00\xED\x9C"
-     "\xB5\x00"
-     "\xED\x9C\xB6\x00\xED\x9C\xB7\x00\xED\x9C\xB8\x00\xED\x9C\xB9\x00\xED\x9C"
-     "\xBA\x00"
-     "\xED\x9C\xBB\x00\xED\x9C\xBC\x00\xED\x9C\xBD\x00\xED\x9C\xBE\x00\xED\x9C"
-     "\xBF\x00"
-     "\xED\x9D\x80\x00\xED\x9D\x81\x00\xED\x9D\x82\x00\xED\x9D\x83\x00\xED\x9D"
-     "\x84\x00"
-     "\xED\x9D\x85\x00\xED\x9D\x86\x00\xED\x9D\x87\x00\xED\x9D\x88\x00\xED\x9D"
-     "\x89\x00"
-     "\xED\x9D\x8A\x00\xED\x9D\x8B\x00\xED\x9D\x8C\x00\xED\x9D\x8D\x00\xED\x9D"
-     "\x8E\x00"
-     "\xED\x9D\x8F\x00\xED\x9D\x90\x00\xED\x9D\x91\x00\xED\x9D\x92\x00\xED\x9D"
-     "\x93\x00"
-     "\xED\x9D\x94\x00\xED\x9D\x95\x00\xED\x9D\x96\x00\xED\x9D\x97\x00\xED\x9D"
-     "\x98\x00"
-     "\xED\x9D\x99\x00\xED\x9D\x9A\x00\xED\x9D\x9B\x00\xED\x9D\x9C\x00\xED\x9D"
-     "\x9D\x00"
-     "\xED\x9D\x9E\x00\xED\x9D\x9F\x00\xED\x9D\xA0\x00\xED\x9D\xA1\x00\xED\x9D"
-     "\xA2\x00"
-     "\xED\x9D\xA3\x00\xED\x9D\xA4\x00\xED\x9D\xA5\x00\xED\x9D\xA6\x00\xED\x9D"
-     "\xA7\x00"
-     "\xED\x9D\xA8\x00\xED\x9D\xA9\x00\xED\x9D\xAA\x00\xED\x9D\xAB\x00\xED\x9D"
-     "\xAC\x00"
-     "\xED\x9D\xAD\x00\xED\x9D\xAE\x00\xED\x9D\xAF\x00\xED\x9D\xB0\x00\xED\x9D"
-     "\xB1\x00"
-     "\xED\x9D\xB2\x00\xED\x9D\xB3\x00\xED\x9D\xB4\x00\xED\x9D\xB5\x00\xED\x9D"
-     "\xB6\x00"
-     "\xED\x9D\xB7\x00\xED\x9D\xB8\x00\xED\x9D\xB9\x00\xED\x9D\xBA\x00\xED\x9D"
-     "\xBB\x00"
-     "\xED\x9D\xBC\x00\xED\x9D\xBD\x00\xED\x9D\xBE\x00\xED\x9D\xBF\x00\xED\x9E"
-     "\x80\x00"
-     "\xED\x9E\x81\x00\xED\x9E\x82\x00\xED\x9E\x83\x00\xED\x9E\x84\x00\xED\x9E"
-     "\x85\x00"
-     "\xED\x9E\x86\x00\xED\x9E\x87\x00\xED\x9E\x88\x00\xED\x9E\x89\x00\xED\x9E"
-     "\x8A\x00"
-     "\xED\x9E\x8B\x00\xED\x9E\x8C\x00\xED\x9E\x8D\x00\xED\x9E\x8E\x00\xED\x9E"
-     "\x8F\x00"
-     "\xED\x9E\x90\x00\xED\x9E\x91\x00\xED\x9E\x92\x00\xED\x9E\x93\x00\xED\x9E"
-     "\x94\x00"
-     "\xED\x9E\x95\x00\xED\x9E\x96\x00\xED\x9E\x97\x00\xED\x9E\x98\x00\xED\x9E"
-     "\x99\x00"
-     "\xED\x9E\x9A\x00\xED\x9E\x9B\x00\xED\x9E\x9C\x00\xED\x9E\x9D\x00\xED\x9E"
-     "\x9E\x00"
-     "\xED\x9E\x9F\x00\xED\x9E\xA0\x00\xED\x9E\xA1\x00\xED\x9E\xA2\x00\xED\x9E"
-     "\xA3\x00"
-     "\xF0\x90\x90\xA8\x00\xF0\x90\x90\xA9\x00\xF0\x90\x90\xAA\x00\xF0\x90\x90"
-     "\xAB\x00"
-     "\xF0\x90\x90\xAC\x00\xF0\x90\x90\xAD\x00\xF0\x90\x90\xAE\x00\xF0\x90\x90"
-     "\xAF\x00"
-     "\xF0\x90\x90\xB0\x00\xF0\x90\x90\xB1\x00\xF0\x90\x90\xB2\x00\xF0\x90\x90"
-     "\xB3\x00"
-     "\xF0\x90\x90\xB4\x00\xF0\x90\x90\xB5\x00\xF0\x90\x90\xB6\x00\xF0\x90\x90"
-     "\xB7\x00"
-     "\xF0\x90\x90\xB8\x00\xF0\x90\x90\xB9\x00\xF0\x90\x90\xBA\x00\xF0\x90\x90"
-     "\xBB\x00"
-     "\xF0\x90\x90\xBC\x00\xF0\x90\x90\xBD\x00\xF0\x90\x90\xBE\x00\xF0\x90\x90"
-     "\xBF\x00"
-     "\xF0\x90\x91\x80\x00\xF0\x90\x91\x81\x00\xF0\x90\x91\x82\x00\xF0\x90\x91"
-     "\x83\x00"
-     "\xF0\x90\x91\x84\x00\xF0\x90\x91\x85\x00\xF0\x90\x91\x86\x00\xF0\x90\x91"
-     "\x87\x00"
-     "\xF0\x90\x91\x88\x00\xF0\x90\x91\x89\x00\xF0\x90\x91\x8A\x00\xF0\x90\x91"
-     "\x8B\x00"
-     "\xF0\x90\x91\x8C\x00\xF0\x90\x91\x8D\x00\xF0\x90\x91\x8E\x00\xF0\x90\x91"
-     "\x8F\x00"
-     "\xF0\x90\x93\x98\x00\xF0\x90\x93\x99\x00\xF0\x90\x93\x9A\x00\xF0\x90\x93"
-     "\x9B\x00"
-     "\xF0\x90\x93\x9C\x00\xF0\x90\x93\x9D\x00\xF0\x90\x93\x9E\x00\xF0\x90\x93"
-     "\x9F\x00"
-     "\xF0\x90\x93\xA0\x00\xF0\x90\x93\xA1\x00\xF0\x90\x93\xA2\x00\xF0\x90\x93"
-     "\xA3\x00"
-     "\xF0\x90\x93\xA4\x00\xF0\x90\x93\xA5\x00\xF0\x90\x93\xA6\x00\xF0\x90\x93"
-     "\xA7\x00"
-     "\xF0\x90\x93\xA8\x00\xF0\x90\x93\xA9\x00\xF0\x90\x93\xAA\x00\xF0\x90\x93"
-     "\xAB\x00"
-     "\xF0\x90\x93\xAC\x00\xF0\x90\x93\xAD\x00\xF0\x90\x93\xAE\x00\xF0\x90\x93"
-     "\xAF\x00"
-     "\xF0\x90\x93\xB0\x00\xF0\x90\x93\xB1\x00\xF0\x90\x93\xB2\x00\xF0\x90\x93"
-     "\xB3\x00"
-     "\xF0\x90\x93\xB4\x00\xF0\x90\x93\xB5\x00\xF0\x90\x93\xB6\x00\xF0\x90\x93"
-     "\xB7\x00"
-     "\xF0\x90\x93\xB8\x00\xF0\x90\x93\xB9\x00\xF0\x90\x93\xBA\x00\xF0\x90\x93"
-     "\xBB\x00"
-     "\xF0\x90\x96\x97\x00\xF0\x90\x96\x98\x00\xF0\x90\x96\x99\x00\xF0\x90\x96"
-     "\x9A\x00"
-     "\xF0\x90\x96\x9B\x00\xF0\x90\x96\x9C\x00\xF0\x90\x96\x9D\x00\xF0\x90\x96"
-     "\x9E\x00"
-     "\xF0\x90\x96\x9F\x00\xF0\x90\x96\xA0\x00\xF0\x90\x96\xA1\x00\xF0\x90\x96"
-     "\xA3\x00"
-     "\xF0\x90\x96\xA4\x00\xF0\x90\x96\xA5\x00\xF0\x90\x96\xA6\x00\xF0\x90\x96"
-     "\xA7\x00"
-     "\xF0\x90\x96\xA8\x00\xF0\x90\x96\xA9\x00\xF0\x90\x96\xAA\x00\xF0\x90\x96"
-     "\xAB\x00"
-     "\xF0\x90\x96\xAC\x00\xF0\x90\x96\xAD\x00\xF0\x90\x96\xAE\x00\xF0\x90\x96"
-     "\xAF\x00"
-     "\xF0\x90\x96\xB0\x00\xF0\x90\x96\xB1\x00\xF0\x90\x96\xB3\x00\xF0\x90\x96"
-     "\xB4\x00"
-     "\xF0\x90\x96\xB5\x00\xF0\x90\x96\xB6\x00\xF0\x90\x96\xB7\x00\xF0\x90\x96"
-     "\xB8\x00"
-     "\xF0\x90\x96\xB9\x00\xF0\x90\x96\xBB\x00\xF0\x90\x96\xBC\x00\xF0\x90\xB3"
-     "\x80\x00"
-     "\xF0\x90\xB3\x81\x00\xF0\x90\xB3\x82\x00\xF0\x90\xB3\x83\x00\xF0\x90\xB3"
-     "\x84\x00"
-     "\xF0\x90\xB3\x85\x00\xF0\x90\xB3\x86\x00\xF0\x90\xB3\x87\x00\xF0\x90\xB3"
-     "\x88\x00"
-     "\xF0\x90\xB3\x89\x00\xF0\x90\xB3\x8A\x00\xF0\x90\xB3\x8B\x00\xF0\x90\xB3"
-     "\x8C\x00"
-     "\xF0\x90\xB3\x8D\x00\xF0\x90\xB3\x8E\x00\xF0\x90\xB3\x8F\x00\xF0\x90\xB3"
-     "\x90\x00"
-     "\xF0\x90\xB3\x91\x00\xF0\x90\xB3\x92\x00\xF0\x90\xB3\x93\x00\xF0\x90\xB3"
-     "\x94\x00"
-     "\xF0\x90\xB3\x95\x00\xF0\x90\xB3\x96\x00\xF0\x90\xB3\x97\x00\xF0\x90\xB3"
-     "\x98\x00"
-     "\xF0\x90\xB3\x99\x00\xF0\x90\xB3\x9A\x00\xF0\x90\xB3\x9B\x00\xF0\x90\xB3"
-     "\x9C\x00"
-     "\xF0\x90\xB3\x9D\x00\xF0\x90\xB3\x9E\x00\xF0\x90\xB3\x9F\x00\xF0\x90\xB3"
-     "\xA0\x00"
-     "\xF0\x90\xB3\xA1\x00\xF0\x90\xB3\xA2\x00\xF0\x90\xB3\xA3\x00\xF0\x90\xB3"
-     "\xA4\x00"
-     "\xF0\x90\xB3\xA5\x00\xF0\x90\xB3\xA6\x00\xF0\x90\xB3\xA7\x00\xF0\x90\xB3"
-     "\xA8\x00"
-     "\xF0\x90\xB3\xA9\x00\xF0\x90\xB3\xAA\x00\xF0\x90\xB3\xAB\x00\xF0\x90\xB3"
-     "\xAC\x00"
-     "\xF0\x90\xB3\xAD\x00\xF0\x90\xB3\xAE\x00\xF0\x90\xB3\xAF\x00\xF0\x90\xB3"
-     "\xB0\x00"
-     "\xF0\x90\xB3\xB1\x00\xF0\x90\xB3\xB2\x00\xF0\x91\x82\x9A\x00\xF0\x91\x82"
-     "\x9C\x00"
-     "\xF0\x91\x82\xAB\x00\xF0\x91\x84\xAE\x00\xF0\x91\x84\xAF\x00\xF0\x91\x8D"
-     "\x8B\x00"
-     "\xF0\x91\x8D\x8C\x00\xF0\x91\x92\xBB\x00\xF0\x91\x92\xBC\x00\xF0\x91\x92"
-     "\xBE\x00"
-     "\xF0\x91\x96\xBA\x00\xF0\x91\x96\xBB\x00\xF0\x91\xA3\x80\x00\xF0\x91\xA3"
-     "\x81\x00"
-     "\xF0\x91\xA3\x82\x00\xF0\x91\xA3\x83\x00\xF0\x91\xA3\x84\x00\xF0\x91\xA3"
-     "\x85\x00"
-     "\xF0\x91\xA3\x86\x00\xF0\x91\xA3\x87\x00\xF0\x91\xA3\x88\x00\xF0\x91\xA3"
-     "\x89\x00"
-     "\xF0\x91\xA3\x8A\x00\xF0\x91\xA3\x8B\x00\xF0\x91\xA3\x8C\x00\xF0\x91\xA3"
-     "\x8D\x00"
-     "\xF0\x91\xA3\x8E\x00\xF0\x91\xA3\x8F\x00\xF0\x91\xA3\x90\x00\xF0\x91\xA3"
-     "\x91\x00"
-     "\xF0\x91\xA3\x92\x00\xF0\x91\xA3\x93\x00\xF0\x91\xA3\x94\x00\xF0\x91\xA3"
-     "\x95\x00"
-     "\xF0\x91\xA3\x96\x00\xF0\x91\xA3\x97\x00\xF0\x91\xA3\x98\x00\xF0\x91\xA3"
-     "\x99\x00"
-     "\xF0\x91\xA3\x9A\x00\xF0\x91\xA3\x9B\x00\xF0\x91\xA3\x9C\x00\xF0\x91\xA3"
-     "\x9D\x00"
-     "\xF0\x91\xA3\x9E\x00\xF0\x91\xA3\x9F\x00\xF0\x91\xA4\xB8\x00\xF0\x96\xB9"
-     "\xA0\x00"
-     "\xF0\x96\xB9\xA1\x00\xF0\x96\xB9\xA2\x00\xF0\x96\xB9\xA3\x00\xF0\x96\xB9"
-     "\xA4\x00"
-     "\xF0\x96\xB9\xA5\x00\xF0\x96\xB9\xA6\x00\xF0\x96\xB9\xA7\x00\xF0\x96\xB9"
-     "\xA8\x00"
-     "\xF0\x96\xB9\xA9\x00\xF0\x96\xB9\xAA\x00\xF0\x96\xB9\xAB\x00\xF0\x96\xB9"
-     "\xAC\x00"
-     "\xF0\x96\xB9\xAD\x00\xF0\x96\xB9\xAE\x00\xF0\x96\xB9\xAF\x00\xF0\x96\xB9"
-     "\xB0\x00"
-     "\xF0\x96\xB9\xB1\x00\xF0\x96\xB9\xB2\x00\xF0\x96\xB9\xB3\x00\xF0\x96\xB9"
-     "\xB4\x00"
-     "\xF0\x96\xB9\xB5\x00\xF0\x96\xB9\xB6\x00\xF0\x96\xB9\xB7\x00\xF0\x96\xB9"
-     "\xB8\x00"
-     "\xF0\x96\xB9\xB9\x00\xF0\x96\xB9\xBA\x00\xF0\x96\xB9\xBB\x00\xF0\x96\xB9"
-     "\xBC\x00"
-     "\xF0\x96\xB9\xBD\x00\xF0\x96\xB9\xBE\x00\xF0\x96\xB9\xBF\x00\xF0\x9D\x85"
-     "\x97\xF0"
-     "\x9D\x85\xA5\x00\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\x00\xF0\x9D\x85\x98\xF0"
-     "\x9D\x85"
-     "\xA5\xF0\x9D\x85\xAE\x00\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\xF0\x9D\x85\xAF"
-     "\x00\xF0"
-     "\x9D\x85\x98\xF0\x9D\x85\xA5\xF0\x9D\x85\xB0\x00\xF0\x9D\x85\x98\xF0\x9D"
-     "\x85\xA5"
-     "\xF0\x9D\x85\xB1\x00\xF0\x9D\x85\x98\xF0\x9D\x85\xA5\xF0\x9D\x85\xB2\x00"
-     "\xF0\x9D"
-     "\x86\xB9\xF0\x9D\x85\xA5\x00\xF0\x9D\x86\xB9\xF0\x9D\x85\xA5\xF0\x9D\x85"
-     "\xAE\x00"
-     "\xF0\x9D\x86\xB9\xF0\x9D\x85\xA5\xF0\x9D\x85\xAF\x00\xF0\x9D\x86\xBA\xF0"
-     "\x9D\x85"
-     "\xA5\x00\xF0\x9D\x86\xBA\xF0\x9D\x85\xA5\xF0\x9D\x85\xAE\x00\xF0\x9D\x86"
-     "\xBA\xF0"
-     "\x9D\x85\xA5\xF0\x9D\x85\xAF\x00\xF0\x9D\xBC\x84\x00\xF0\x9D\xBC\x85\x00"
-     "\xF0\x9D"
-     "\xBC\x86\x00\xF0\x9D\xBC\x88\x00\xF0\x9D\xBC\x8A\x00\xF0\x9D\xBC\x9E\x00"
-     "\xF0\x9E"
-     "\xA4\xA2\x00\xF0\x9E\xA4\xA3\x00\xF0\x9E\xA4\xA4\x00\xF0\x9E\xA4\xA5\x00"
-     "\xF0\x9E"
-     "\xA4\xA6\x00\xF0\x9E\xA4\xA7\x00\xF0\x9E\xA4\xA8\x00\xF0\x9E\xA4\xA9\x00"
-     "\xF0\x9E"
-     "\xA4\xAA\x00\xF0\x9E\xA4\xAB\x00\xF0\x9E\xA4\xAC\x00\xF0\x9E\xA4\xAD\x00"
-     "\xF0\x9E"
-     "\xA4\xAE\x00\xF0\x9E\xA4\xAF\x00\xF0\x9E\xA4\xB0\x00\xF0\x9E\xA4\xB1\x00"
-     "\xF0\x9E"
-     "\xA4\xB2\x00\xF0\x9E\xA4\xB3\x00\xF0\x9E\xA4\xB4\x00\xF0\x9E\xA4\xB5\x00"
-     "\xF0\x9E"
-     "\xA4\xB6\x00\xF0\x9E\xA4\xB7\x00\xF0\x9E\xA4\xB8\x00\xF0\x9E\xA4\xB9\x00"
-     "\xF0\x9E"
-     "\xA4\xBA\x00\xF0\x9E\xA4\xBB\x00\xF0\x9E\xA4\xBC\x00\xF0\x9E\xA4\xBD\x00"
-     "\xF0\x9E"
-     "\xA4\xBE\x00\xF0\x9E\xA4\xBF\x00\xF0\x9E\xA5\x80\x00\xF0\x9E\xA5\x81\x00"
-     "\xF0\x9E"
-     "\xA5\x82\x00\xF0\x9E\xA5\x83\x00\xF0\xA0\x84\xA2\x00\xF0\xA0\x94\x9C\x00"
-     "\xF0\xA0"
-     "\x94\xA5\x00\xF0\xA0\x95\x8B\x00\xF0\xA0\x98\xBA\x00\xF0\xA0\xA0\x84\x00"
-     "\xF0\xA0"
-     "\xA3\x9E\x00\xF0\xA0\xA8\xAC\x00\xF0\xA0\xAD\xA3\x00\xF0\xA1\x93\xA4\x00"
-     "\xF0\xA1"
-     "\x9A\xA8\x00\xF0\xA1\x9B\xAA\x00\xF0\xA1\xA7\x88\x00\xF0\xA1\xAC\x98\x00"
-     "\xF0\xA1"
-     "\xB4\x8B\x00\xF0\xA1\xB7\xA4\x00\xF0\xA1\xB7\xA6\x00\xF0\xA2\x86\x83\x00"
-     "\xF0\xA2"
-     "\x86\x9F\x00\xF0\xA2\x8C\xB1\x00\xF0\xA2\x9B\x94\x00\xF0\xA2\xA1\x84\x00"
-     "\xF0\xA2"
-     "\xA1\x8A\x00\xF0\xA2\xAC\x8C\x00\xF0\xA2\xAF\xB1\x00\xF0\xA3\x80\x8A\x00"
-     "\xF0\xA3"
-     "\x8A\xB8\x00\xF0\xA3\x8D\x9F\x00\xF0\xA3\x8E\x93\x00\xF0\xA3\x8E\x9C\x00"
-     "\xF0\xA3"
-     "\x8F\x83\x00\xF0\xA3\x8F\x95\x00\xF0\xA3\x91\xAD\x00\xF0\xA3\x9A\xA3\x00"
-     "\xF0\xA3"
-     "\xA2\xA7\x00\xF0\xA3\xAA\x8D\x00\xF0\xA3\xAB\xBA\x00\xF0\xA3\xB2\xBC\x00"
-     "\xF0\xA3"
-     "\xB4\x9E\x00\xF0\xA3\xBB\x91\x00\xF0\xA3\xBD\x9E\x00\xF0\xA3\xBE\x8E\x00"
-     "\xF0\xA4"
-     "\x89\xA3\x00\xF0\xA4\x8B\xAE\x00\xF0\xA4\x8E\xAB\x00\xF0\xA4\x98\x88\x00"
-     "\xF0\xA4"
-     "\x9C\xB5\x00\xF0\xA4\xA0\x94\x00\xF0\xA4\xB0\xB6\x00\xF0\xA4\xB2\x92\x00"
-     "\xF0\xA4"
-     "\xBE\xA1\x00\xF0\xA4\xBE\xB8\x00\xF0\xA5\x81\x84\x00\xF0\xA5\x83\xB2\x00"
-     "\xF0\xA5"
-     "\x83\xB3\x00\xF0\xA5\x84\x99\x00\xF0\xA5\x84\xB3\x00\xF0\xA5\x89\x89\x00"
-     "\xF0\xA5"
-     "\x90\x9D\x00\xF0\xA5\x98\xA6\x00\xF0\xA5\x9A\x9A\x00\xF0\xA5\x9B\x85\x00"
-     "\xF0\xA5"
-     "\xA5\xBC\x00\xF0\xA5\xAA\xA7\x00\xF0\xA5\xAE\xAB\x00\xF0\xA5\xB2\x80\x00"
-     "\xF0\xA5"
-     "\xB3\x90\x00\xF0\xA5\xBE\x86\x00\xF0\xA6\x87\x9A\x00\xF0\xA6\x88\xA8\x00"
-     "\xF0\xA6"
-     "\x89\x87\x00\xF0\xA6\x8B\x99\x00\xF0\xA6\x8C\xBE\x00\xF0\xA6\x93\x9A\x00"
-     "\xF0\xA6"
-     "\x94\xA3\x00\xF0\xA6\x96\xA8\x00\xF0\xA6\x9E\xA7\x00\xF0\xA6\x9E\xB5\x00"
-     "\xF0\xA6"
-     "\xAC\xBC\x00\xF0\xA6\xB0\xB6\x00\xF0\xA6\xB3\x95\x00\xF0\xA6\xB5\xAB\x00"
-     "\xF0\xA6"
-     "\xBC\xAC\x00\xF0\xA6\xBE\xB1\x00\xF0\xA7\x83\x92\x00\xF0\xA7\x8F\x8A\x00"
-     "\xF0\xA7"
-     "\x99\xA7\x00\xF0\xA7\xA2\xAE\x00\xF0\xA7\xA5\xA6\x00\xF0\xA7\xB2\xA8\x00"
-     "\xF0\xA7"
-     "\xBB\x93\x00\xF0\xA7\xBC\xAF\x00\xF0\xA8\x97\x92\x00\xF0\xA8\x97\xAD\x00"
-     "\xF0\xA8"
-     "\x9C\xAE\x00\xF0\xA8\xAF\xBA\x00\xF0\xA8\xB5\xB7\x00\xF0\xA9\x85\x85\x00"
-     "\xF0\xA9"
-     "\x87\x9F\x00\xF0\xA9\x88\x9A\x00\xF0\xA9\x90\x8A\x00\xF0\xA9\x92\x96\x00"
-     "\xF0\xA9"
-     "\x96\xB6\x00\xF0\xA9\xAC\xB0\x00\xF0\xAA\x83\x8E\x00\xF0\xAA\x84\x85\x00"
-     "\xF0\xAA"
-     "\x88\x8E\x00\xF0\xAA\x8A\x91\x00\xF0\xAA\x8E\x92\x00\xF0\xAA\x98\x80"
-     "\x00"},
-};
-#endif
-constexpr size_t kNormalizationRules_size = 4;
-
-}  // namespace
-}  // namespace sentencepiece
-#endif  // NORMALIZATION_RULE_H_
diff --git a/projects/llm_framework/include/sentencepiece/normalizer.h b/projects/llm_framework/include/sentencepiece/normalizer.h
deleted file mode 100644
index c79813c7..00000000
--- a/projects/llm_framework/include/sentencepiece/normalizer.h
+++ /dev/null
@@ -1,159 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef NORMALIZER_NORMALIZER_H_
-#define NORMALIZER_NORMALIZER_H_
-
-#include <memory>
-#include <set>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "common.h"
-#include "sentencepiece_model.pb.h"
-#include "sentencepiece_processor.h"
-#include "third_party/absl/strings/string_view.h"
-#include "third_party/darts_clone/darts.h"
-
-namespace sentencepiece {
-namespace normalizer {
-
-// Given a list of strings, finds the longest string which is a
-// prefix of a query.
-class PrefixMatcher {
- public:
-  // Initializes the PrefixMatcher with `dic`.
-  explicit PrefixMatcher(const std::set<absl::string_view> &dic);
-
-  // Finds the longest string in dic, which is a prefix of `w`.
-  // Returns the UTF8 byte length of matched string.
-  // `found` is set if a prefix match exists.
-  // If no entry is found, consumes one Unicode character.
-  int PrefixMatch(absl::string_view w, bool *found = nullptr) const;
-
-  // Replaces entries in `w` with `out`.
-  std::string GlobalReplace(absl::string_view w, absl::string_view out) const;
-
- private:
-  std::unique_ptr<Darts::DoubleArray> trie_;
-};
-
-// Normalizer implements a simple text normalizer with
-// user-defined string-to-string rules and leftmost longest
-// matching. The rules of Normalizer are built with
-// Builder::CompileCharsMap() method. Pre-compiled rules are
-// also available via Builder::GetPrecompiledCharsMap(<name>) method.
-//
-// The motivation of Normalizer is to make flexible, user-customizable
-// and self-contained normalizer.  All the logic of normalization is
-// encoded in the model proto which allows us to define language/task
-// dependent normalization rules without breaking the default rule.
-class Normalizer {
- public:
-  // Instantiates Normalizer with |spec|.
-  // |spec| should not be deleted until Normalizer is destroyed.
-  explicit Normalizer(const NormalizerSpec &spec);
-  Normalizer(const NormalizerSpec &spec, const TrainerSpec &trainer_Spec);
-  virtual ~Normalizer();
-
-  virtual void SetPrefixMatcher(const PrefixMatcher *matcher) {
-    matcher_ = matcher;
-  }
-
-  // Returns Status.
-  // Normalizes function is valid only when status is OK.
-  virtual util::Status status() const { return status_; }
-
-  // Normalizes a plain utf8 string into an internal representation for
-  // Sentencepiece model. |norm_to_orig| stores the byte-alignment from
-  // normalized string to the original input.
-  // This function can do the following normalizations:
-  // - Character normalization.
-  //   (NFKC / full-width to half-width conversion etc).
-  // - Adds a prefix space.
-  // - Replaces a space with a meta symbol.
-  // - Removing heading, tailing and other redundant spaces.
-  virtual util::Status Normalize(absl::string_view input,
-                                 std::string *normalized,
-                                 std::vector<size_t> *norm_to_orig) const;
-
-  // Returns a normalized string without alignments.
-  // This function is used in sentencepiece training.
-  virtual std::string Normalize(absl::string_view input) const;
-
-  friend class Builder;
-
- private:
-  FRIEND_TEST(NormalizerTest, EncodeDecodePrecompiledCharsMapTest);
-
-  void Init();
-
-  // Normalizes the prefix of |input| and returns the pair of
-  // normalized prefix and length we must consume after
-  // normalization.
-  // Here's the sample code for the full text normalization.
-  //
-  // string output;
-  // absl::string_view input = "...";
-  // while (!input.empty()) {
-  //   const auto p = normalizer.NormalizePrefix(input);
-  //   output.append(p.first.data(), p.first.size());
-  //   input.remove_prefix(p.second);
-  // }
-  std::pair<absl::string_view, int> NormalizePrefix(
-      absl::string_view input) const;
-
-  // Encodes trie_blob and normalized string and return compiled blob.
-  static std::string EncodePrecompiledCharsMap(absl::string_view trie_blob,
-                                               absl::string_view normalized);
-
-  // Decodes blob into trie_blob and normalized string.
-  static util::Status DecodePrecompiledCharsMap(absl::string_view blob,
-                                                absl::string_view *trie_blob,
-                                                absl::string_view *normalized,
-                                                std::string *buffer = nullptr);
-
-  // Maximum size of the return value of Trie, which corresponds
-  // to the maximum size of shared common prefix in the chars map.
-  static constexpr int kMaxTrieResultsSize = 32;
-
-  // Internal trie for efficient longest matching.
-  std::unique_ptr<Darts::DoubleArray> trie_;
-
-  // "\0" delimitered output string.
-  // the value of |trie_| stores pointers to this string.
-  const char *normalized_ = nullptr;
-
-  // Spec for normalization.
-  const NormalizerSpec *spec_;
-
-  // Prefix matcher;
-  const PrefixMatcher *matcher_ = nullptr;
-
-  // Split hello world into "hello_" and "world_" instead of
-  // "_hello" and "_world".
-  const bool treat_whitespace_as_suffix_ = false;
-
-#ifdef IS_BIG_ENDIAN
-  // Stores the blob for TRIE encoded in big-endian.
-  std::string precompiled_charsmap_buffer_;
-#endif
-
-  // Normalizer's status.
-  util::Status status_;
-};
-}  // namespace normalizer
-}  // namespace sentencepiece
-#endif  // NORMALIZER_NORMALIZER_H_
diff --git a/projects/llm_framework/include/sentencepiece/pretokenizer_for_training.h b/projects/llm_framework/include/sentencepiece/pretokenizer_for_training.h
deleted file mode 100644
index fa54f95c..00000000
--- a/projects/llm_framework/include/sentencepiece/pretokenizer_for_training.h
+++ /dev/null
@@ -1,63 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef PRETOKENIZER_FOR_TRAINING_H_
-#define PRETOKENIZER_FOR_TRAINING_H_
-
-#include <memory>
-#include <string>
-
-#include "common.h"
-#include "sentencepiece.pb.h"
-#include "sentencepiece_processor.h"
-#include "third_party/absl/strings/string_view.h"
-
-namespace sentencepiece {
-namespace pretokenizer {
-
-class PretokenizerForTrainingInterface {
- public:
-  PretokenizerForTrainingInterface() {}
-  virtual ~PretokenizerForTrainingInterface() {}
-  virtual util::Status status() const = 0;
-
-  // Puts kUPPBoundaryStr before and after the pre-tokenizer's segmentation
-  // when there are no spaces between these tokens.
-  // Example1:
-  // input: 東京です
-  // segmentation: piece[0] = {0, 6}, piece[1] = {6, 12}
-  // output: 東京<tab>です (here kUPPBoundaryStr is <tab>)
-  //
-  // Example2:
-  // input: I love sentencepiece
-  // segmentation: piece[0] = {0, 1}, piece[1] = {2, 6},
-  //               piece[2] = {7, 15}, piece[3] = {15, 20}
-  // output: I love sentence<tab>piece.
-  std::vector<std::string> PreTokenize(absl::string_view text) const;
-
-  // Returns pre-tokenized result.
-  // Note that the pre-tokenized constraint is specified with the
-  // byte offsets (SentencePiece::begin, SentencePiece::end) over
-  // the input text.
-  virtual SentencePieceText Tokenize(absl::string_view text) const = 0;
-
- private:
-  static std::string Preprocess(absl::string_view text);
-  static std::vector<std::string> Postprocess(const SentencePieceText &spt);
-};
-
-}  // namespace pretokenizer
-}  // namespace sentencepiece
-
-#endif  // PRETOKENIZER_FOR_TRAINING_H_
diff --git a/projects/llm_framework/include/sentencepiece/sentencepiece_processor.h b/projects/llm_framework/include/sentencepiece/sentencepiece_processor.h
deleted file mode 100644
index dd3f092e..00000000
--- a/projects/llm_framework/include/sentencepiece/sentencepiece_processor.h
+++ /dev/null
@@ -1,752 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef SENTENCEPIECE_PROCESSOR_H_
-#define SENTENCEPIECE_PROCESSOR_H_
-
-#include <cstring>
-#include <memory>
-#include <string>
-#include <string_view>
-#include <utility>
-#include <vector>
-
-#ifndef SWIG
-namespace absl {
-using std::string_view;
-}  // namespace absl
-#endif  // SWIG
-
-namespace sentencepiece {
-namespace util {
-
-enum class StatusCode : int {
-  kOk = 0,
-  kCancelled = 1,
-  kUnknown = 2,
-  kInvalidArgument = 3,
-  kDeadlineExceeded = 4,
-  kNotFound = 5,
-  kAlreadyExists = 6,
-  kPermissionDenied = 7,
-  kResourceExhausted = 8,
-  kFailedPrecondition = 9,
-  kAborted = 10,
-  kOutOfRange = 11,
-  kUnimplemented = 12,
-  kInternal = 13,
-  kUnavailable = 14,
-  kDataLoss = 15,
-  kUnauthenticated = 16,
-};
-
-class Status {
- public:
-  Status();
-  ~Status();
-  Status(StatusCode code, absl::string_view error_message);
-  Status(const Status &s);
-  void operator=(const Status &s);
-  bool operator==(const Status &s) const;
-  bool operator!=(const Status &s) const;
-  inline bool ok() const { return rep_ == nullptr; }
-
-  void set_error_message(const char *str);
-  const char *error_message() const;
-  const char *message() const { return error_message(); }
-  StatusCode code() const;
-  std::string ToString() const;
-
-  void IgnoreError();
-
- private:
-  struct Rep;
-  std::unique_ptr<Rep> rep_;
-};
-}  // namespace util
-
-// SentencePieceProcessor:
-// Simple and language independent tokenizer and de-tokenizer for
-// Neural Network Machine Translation.
-//
-// SentencePieceProcessor provides Encode() and Decode() methods,
-// which correspond to tokenization and de-tokenization respectively.
-//
-// - Encode:
-//   Given a raw source sentence, encode it into a sequence
-//   of pieces or vocabulary ids.
-//
-// - Decode:
-//   Given a sequence of pieces or vocabulary ids, decode it
-//   into a de-tokenized raw sentence.
-//
-// SentencePieceProcessor provides a lossless data conversion
-// that allows the original raw sentence to be perfectly reconstructed
-// from the encoded data, i.e., Decode(Encode(input)) == input.
-// This characteristics is useful, as we can make the de-tokenization
-// completely language independent.
-//
-// Usage:
-//   SentencePieceProcessor sp;
-//   sp.Load("//path/to/model");
-//
-//   vector<string> sps;
-//   sp.Encode("hello world.", &sps).IgnoreError();
-//
-//   vector<int> ids;
-//   sp.Encode("hello world.", &ids).IgnoreError();
-//
-//   string detok;
-//   sp.Decode(sps, &detok);
-//   CHECK_EQ("hello world.", detok).IgnoreError();
-//
-//   sp.Decode(ids, &detok);
-//   CHECK_EQ("hello world.", detok).IgnoreError();
-//
-//  We can also use SentencePieceText which manages the byte-offsets
-//  between user input (output) and internal sentence pieces.
-//
-//   SentencePieceText spt;
-//   sp.Encode("hello world.", &spt);
-//   // Emits the byte range of each piece.
-//   for (const auto &piece : spt.pieces()) {
-//      LOG(INFO) << piece.begin() << " " << piece.end();
-//   }
-//
-//   sp.Decode({0, 1, 2, 3..}, &spt);
-//   for (const auto &piece : spt.pieces()) {
-//      LOG(INFO) << piece.begin() << " " << piece.end();
-//   }
-//
-
-class NBestSentencePieceText;
-class ModelInterface;
-class SentencePieceText;
-class ModelProto;
-class NormalizerSpec;
-
-namespace normalizer {
-class Normalizer;
-}  // namespace normalizer
-
-#ifndef SWIGGO
-namespace util {
-// Redefine std::string for serialized_proto interface as Python's string is
-// a Unicode string. We can enforce the return value to be raw byte sequence
-// with SWIG's typemap.
-using bytes = std::string;
-}  // namespace util
-#endif  // SWIGGO
-
-class NBestSentencePieceText;
-class ModelInterface;
-class SentencePieceText;
-class SentencePieceText_SentencePiece;
-
-// Wrapper class of SentencePieceText
-// This wrapper only allows an immutable access to the proto and
-// hides the actual implementation of protobuf.
-// See sentencepiece.proto for the details of this class.
-class ImmutableSentencePieceText_ImmutableSentencePiece {
- public:
-  ImmutableSentencePieceText_ImmutableSentencePiece();
-  ~ImmutableSentencePieceText_ImmutableSentencePiece() = default;
-
-  const std::string &piece() const;
-  const std::string &surface() const;
-  uint32_t id() const;
-  uint32_t begin() const;
-  uint32_t end() const;
-
-  friend class ImmutableSentencePieceText;
-
- private:
-  explicit ImmutableSentencePieceText_ImmutableSentencePiece(
-      const SentencePieceText_SentencePiece &sp);
-  const SentencePieceText_SentencePiece *sp_ = nullptr;
-};
-
-class ImmutableSentencePieceText {
- public:
-  ImmutableSentencePieceText();
-  virtual ~ImmutableSentencePieceText();
-
-  std::vector<ImmutableSentencePieceText_ImmutableSentencePiece> pieces() const;
-
-  size_t pieces_size() const;
-  ImmutableSentencePieceText_ImmutableSentencePiece pieces(int index) const;
-
-  const std::string &text() const;
-  float score() const;
-
-  util::bytes SerializeAsString() const;
-
-  // Returns the actual mutable proto.
-  // Do not use this outside of SentencePieceProcessor, as
-  // it returns the raw pointer managed by the shared_ptr.
-  SentencePieceText *mutable_proto();
-
-  // Converts the utf8 byte spans into Unicode char span.
-  void ConvertToUnicodeSpans();
-
-  friend class ImmutableNBestSentencePieceText;
-
- private:
-  explicit ImmutableSentencePieceText(const SentencePieceText &spt);
-  const SentencePieceText *spt_ = nullptr;
-  std::shared_ptr<SentencePieceText> rep_;
-};
-
-// Wrapper class of SentencePieceText
-// This wrapper only allows an immutable access to the proto and
-// hides the actual implementation of protobuf.
-// See sentencepiece.proto for the details of this class.
-class ImmutableNBestSentencePieceText {
- public:
-  ImmutableNBestSentencePieceText();
-  virtual ~ImmutableNBestSentencePieceText();
-
-  std::vector<ImmutableSentencePieceText> nbests() const;
-
-  size_t nbests_size() const;
-  ImmutableSentencePieceText nbests(int index) const;
-
-  util::bytes SerializeAsString() const;
-
-  // Returns the actual mutable proto.
-  // Do not use this outside of SentencePieceProcessor, as
-  // it returns the raw pointer managed by the shared_ptr.
-  NBestSentencePieceText *mutable_proto();
-
-  void ConvertToUnicodeSpans();
-
- private:
-  std::shared_ptr<NBestSentencePieceText> rep_;
-};
-
-class SentencePieceProcessor {
- public:
-  SentencePieceProcessor();
-  virtual ~SentencePieceProcessor();
-
-  // Loads model from `filename`.
-  // Returns false if `filename` cannot be loaded.
-  virtual util::Status Load(absl::string_view filename);
-
-  // Loads model from `filename`.
-  // Crash if `filename` cannot be loaded.
-  virtual void LoadOrDie(absl::string_view filename);
-
-  // Loads model from `model_proto`.
-  // `model_proto` is copied.
-  virtual util::Status Load(const ModelProto &model_proto);
-
-  // Loads model from `model_proto`.
-  // `model_proto` is moved.
-  virtual util::Status Load(std::unique_ptr<ModelProto> model_proto);
-
-  // Loads model from `serialized`, which is a string-serialized model proto.
-  // Useful to load the model from a platform independent blob object.
-  virtual util::Status LoadFromSerializedProto(absl::string_view serialized);
-
-  // Returns the status. Encode/Decode methods are valid when status is OK.
-  virtual util::Status status() const;
-
-  // Sets encode extra_option sequence.
-  virtual util::Status SetEncodeExtraOptions(absl::string_view extra_option);
-
-  // Sets decode extra_option sequence.
-  virtual util::Status SetDecodeExtraOptions(absl::string_view extra_option);
-
-  //////////////////////////////////////////////////////////////
-  // Vocabulary restriction.
-  // Background:
-  // https://github.com/rsennrich/subword-nmt#best-practice-advice-for-byte-pair-encoding-in-nmt
-
-  // Restricts the vocabulary set.
-  // The input sentences are encoded into the tokens in `valid_vocab`.
-  virtual util::Status SetVocabulary(
-      const std::vector<absl::string_view> &valid_vocab);
-
-  // Reverts the vocabulary restriction.
-  virtual util::Status ResetVocabulary();
-
-  // Loads the valid vocabulary set from `filename` in TSV format.
-  // Format:  <token> <tab> <freq>.
-  // Any token with frequency < threshold will be treated as OOV.
-  virtual util::Status LoadVocabulary(absl::string_view filename,
-                                      int threshold);
-
-  //////////////////////////////////////////////////////////////
-  // Simple Encode and Decode API.
-  //
-  // Given a UTF8 input, encodes it into a sequence of sentence pieces.
-  virtual util::Status Encode(absl::string_view input,
-                              std::vector<std::string> *pieces) const;
-
-  // Given a UTF8 input, encodes it into a sequence of ids.
-  virtual util::Status Encode(absl::string_view input,
-                              std::vector<int> *ids) const;
-
-  // Given a sequence of pieces, decodes it into a detokenized output.
-  virtual util::Status Decode(const std::vector<std::string> &pieces,
-                              std::string *detokenized) const;
-
-  // Given a sequence of pieces, decodes it into a detokenized output.
-  virtual util::Status Decode(const std::vector<absl::string_view> &pieces,
-                              std::string *detokenized) const;
-
-  // Given a sequence of ids, decodes it into a detokenized output.
-  virtual util::Status Decode(const std::vector<int> &ids,
-                              std::string *detokenized) const;
-
-  //////////////////////////////////////////////////////////////
-  // NBest API.
-  //
-  // Same as Encode, but returns nbest results.
-  virtual util::Status NBestEncode(
-      absl::string_view input, int nbest_size,
-      std::vector<std::vector<std::string>> *pieces) const;
-
-  // Same as Encode, but returns nbest results.
-  virtual util::Status NBestEncode(absl::string_view input, int nbest_size,
-                                   std::vector<std::vector<int>> *ids) const;
-
-  //////////////////////////////////////////////////////////////
-  // Sampling API.
-  //
-  // Unigram and BPE support sampling mode.
-  // - Unigram (--model_type=unigram):
-  // `nbest_size`: When `nbest_size` is positive value, approximately samples
-  // one segmentation from nbest candidates. When `nbest_size` is negative
-  // value, samples one segmentation from the hypotheses (Lattice) according to
-  // the generation probabilities using forward-filtering and backward-sampling
-  // algorithm.
-  // `alpha`: Smoothing parameter (inverse temperature). The best segmentation
-  // (Viterbi segmentation) is more likely sampled when setting larger alpha.
-  // When alpha is 0.0, one segmentation is uniformly sampled from the nbest or
-  // lattice. `nbest_size` and `alpha` correspond to parameters `l` and `alpha`
-  // in https://arxiv.org/abs/1804.10959  (nbest_size < 0 means l = infinity)
-  //
-  // - BPE (--model_type=bpe):
-  // `alpha`: The dropout probability `p` of bpe merge operations in
-  // https://arxiv.org/abs/1910.13267 Nbest-based sampling is not supported so
-  // nbest_size parameter is ignored in BPE.
-  virtual util::Status SampleEncode(absl::string_view input, int nbest_size,
-                                    float alpha,
-                                    std::vector<std::string> *pieces) const;
-
-  // Same as above, but returns a sequence of ids.
-  virtual util::Status SampleEncode(absl::string_view input, int nbest_size,
-                                    float alpha, std::vector<int> *ids) const;
-
-  //////////////////////////////////////////////////////////////
-  // SampleEncodeAndScore API.
-  //
-  // Sample `samples` many tokenisations from the segmentation lattice.
-  // These methods are only available in model_type=unigram.
-  //
-  // `alpha`: smoothing parameter (inverse temperature). The same as `alpha` in
-  // `Sample` method.
-  // 'wor`: If `wor` is true, the samples are taken without replacement, and the
-  // scores are the inclusion probabilities of the elements in the sample;
-  // otherwise the samples are taken with replacement and the scores are the
-  // log-probs of sample elements
-  // `include_best`: If `include_best` is true, the best tokenisation is always
-  // included in the sample, and the remaining elements are sampled excluding
-  // the best.
-  virtual util::Status SampleEncodeAndScore(
-      absl::string_view input, int num_samples, float alpha, bool wor,
-      bool include_best,
-      std::vector<std::pair<std::vector<std::string>, float>> *pieces) const;
-
-  // Same as above, but returns a sequence of ids.
-  virtual util::Status SampleEncodeAndScore(
-      absl::string_view input, int num_samples, float alpha, bool wor,
-      bool include_best,
-      std::vector<std::pair<std::vector<int>, float>> *ids) const;
-
-  //////////////////////////////////////////////////////////////
-  // Entropy API.
-  //
-  // This only available in model_type=unigram.
-  // Calculate entropy of possible tokenisations
-  virtual util::Status CalculateEntropy(absl::string_view input, float alpha,
-                                        float *entropy) const;
-
-  //////////////////////////////////////////////////////////////
-  // Advanced API returning SentencePieceText, which manages
-  // utf8-byte alignments between user-input/detokenized text
-  // and internal sentencepiece sequence.
-  //
-  // Given a UTF8 input, encodes it into SentencePieceText.
-  //
-  // When using these APIs, sentencepiece.pb.h header files must be included.
-  // We can also use ImutableSentencePieceText as follows.
-  //
-  // ImmutableSentencePieceText spt;
-  // Encode("hello", spt.mutable_proto()).IgnoreError();
-  // std::cout << spt.pieces_size() << std::endl;
-  virtual util::Status Encode(absl::string_view input,
-                              SentencePieceText *spt) const;
-
-  virtual util::Status NBestEncode(absl::string_view input, int nbest_size,
-                                   NBestSentencePieceText *nbest_spt) const;
-
-  virtual util::Status SampleEncode(absl::string_view input, int nbest_size,
-                                    float alpha, SentencePieceText *spt) const;
-
-  virtual util::Status SampleEncodeAndScore(
-      absl::string_view input, int num_samples, float alpha, bool wor,
-      bool include_best, NBestSentencePieceText *samples_spt) const;
-
-  // DEPRECATED: Remove this API and use std::vector<std::string_view>
-  virtual util::Status Decode(const std::vector<std::string> &pieces,
-                              SentencePieceText *spt) const;
-
-  virtual util::Status Decode(const std::vector<absl::string_view> &pieces,
-                              SentencePieceText *spt) const;
-
-  virtual util::Status Decode(const std::vector<int> &ids,
-                              SentencePieceText *spt) const;
-#ifdef SWIG
-#define SPP_SWIG_CHECK_AND_THROW \
-  if (!status.ok()) throw status;
-#else
-#define SPP_SWIG_CHECK_AND_THROW \
-  if (!status.ok()) {            \
-  }
-#endif  // SWIG
-
-#define DEFINE_SPP_DIRECT_FUNC_IMPL(FuncName, OutType, ...) \
-  OutType output;                                           \
-  const auto status = FuncName(__VA_ARGS__, &output);       \
-  SPP_SWIG_CHECK_AND_THROW;                                 \
-  return output;
-
-#define DEFINE_SPP_SERIALIZED_PROTO_IMPL(FuncName, OutType, ...)     \
-  OutType output;                                                    \
-  const auto status = FuncName(__VA_ARGS__, output.mutable_proto()); \
-  SPP_SWIG_CHECK_AND_THROW;                                          \
-  return output.SerializeAsString();
-
-#define DEFINE_SPP_IMMUTABLE_PROTO_IMPL(FuncName, OutType, ...)      \
-  OutType output;                                                    \
-  const auto status = FuncName(__VA_ARGS__, output.mutable_proto()); \
-  SPP_SWIG_CHECK_AND_THROW;                                          \
-  return output;
-
-  //////////////////////////////////////////////////////////////
-  // Handy methods that return the result directly.
-  // These functions ignore internal errors.
-  virtual std::vector<std::string> EncodeAsPieces(
-      absl::string_view input) const {
-    DEFINE_SPP_DIRECT_FUNC_IMPL(Encode, std::vector<std::string>, input);
-  }
-
-  virtual std::vector<int> EncodeAsIds(absl::string_view input) const {
-    DEFINE_SPP_DIRECT_FUNC_IMPL(Encode, std::vector<int>, input);
-  }
-
-  virtual std::vector<std::vector<std::string>> NBestEncodeAsPieces(
-      absl::string_view input, int nbest_size) const {
-    DEFINE_SPP_DIRECT_FUNC_IMPL(
-        NBestEncode, std::vector<std::vector<std::string>>, input, nbest_size);
-  }
-
-  virtual std::vector<std::vector<int>> NBestEncodeAsIds(
-      absl::string_view input, int nbest_size) const {
-    DEFINE_SPP_DIRECT_FUNC_IMPL(NBestEncode, std::vector<std::vector<int>>,
-                                input, nbest_size);
-  }
-
-  virtual std::vector<std::string> SampleEncodeAsPieces(absl::string_view input,
-                                                        int nbest_size,
-                                                        float alpha) const {
-    DEFINE_SPP_DIRECT_FUNC_IMPL(SampleEncode, std::vector<std::string>, input,
-                                nbest_size, alpha);
-  }
-
-  virtual std::vector<int> SampleEncodeAsIds(absl::string_view input,
-                                             int nbest_size,
-                                             float alpha) const {
-    DEFINE_SPP_DIRECT_FUNC_IMPL(SampleEncode, std::vector<int>, input,
-                                nbest_size, alpha);
-  }
-
-  virtual std::vector<std::pair<std::vector<std::string>, float>>
-  SampleEncodeAndScoreAsPieces(absl::string_view input, int num_samples,
-                               float alpha, bool wor, bool include_best) const {
-    using _T = std::vector<std::pair<std::vector<std::string>, float>>;
-    DEFINE_SPP_DIRECT_FUNC_IMPL(SampleEncodeAndScore, _T, input, num_samples,
-                                alpha, wor, include_best);
-  }
-
-  virtual std::vector<std::pair<std::vector<int>, float>>
-  SampleEncodeAndScoreAsIds(absl::string_view input, int num_samples,
-                            float alpha, bool wor, bool include_best) const {
-    using _T = std::vector<std::pair<std::vector<int>, float>>;
-    DEFINE_SPP_DIRECT_FUNC_IMPL(SampleEncodeAndScore, _T, input, num_samples,
-                                alpha, wor, include_best);
-  }
-
-  // DEPRECATED: Remove this API and use std::vector<std::string_view>
-  virtual std::string DecodePieces(
-      const std::vector<std::string> &pieces) const {
-    DEFINE_SPP_DIRECT_FUNC_IMPL(Decode, std::string, pieces);
-  }
-
-  virtual std::string DecodePieces(
-      const std::vector<absl::string_view> &pieces) const {
-    DEFINE_SPP_DIRECT_FUNC_IMPL(Decode, std::string, pieces);
-  }
-
-  virtual std::string DecodeIds(const std::vector<int> &ids) const {
-    DEFINE_SPP_DIRECT_FUNC_IMPL(Decode, std::string, ids);
-  }
-
-  virtual float CalculateEntropy(absl::string_view text, float alpha) const {
-    DEFINE_SPP_DIRECT_FUNC_IMPL(CalculateEntropy, float, text, alpha);
-  }
-
-  //////////////////////////////////////////////////////////////
-  // SerializedProto API. (DEPRECATED). Use ImmutableProto API.
-  // They are used in Python interface. Returns serialized proto.
-  // In python module, we can get access to the full Proto after
-  // deserialzing the returned byte sequence.
-  virtual util::bytes EncodeAsSerializedProto(absl::string_view input) const {
-    DEFINE_SPP_SERIALIZED_PROTO_IMPL(Encode, ImmutableSentencePieceText, input);
-  }
-
-  virtual util::bytes SampleEncodeAsSerializedProto(absl::string_view input,
-                                                    int nbest_size,
-                                                    float alpha) const {
-    DEFINE_SPP_SERIALIZED_PROTO_IMPL(SampleEncode, ImmutableSentencePieceText,
-                                     input, nbest_size, alpha);
-  }
-
-  virtual util::bytes NBestEncodeAsSerializedProto(absl::string_view input,
-                                                   int nbest_size) const {
-    DEFINE_SPP_SERIALIZED_PROTO_IMPL(
-        NBestEncode, ImmutableNBestSentencePieceText, input, nbest_size);
-  }
-
-  virtual util::bytes SampleEncodeAndScoreAsSerializedProto(
-      absl::string_view input, int num_samples, float alpha, bool wor,
-      bool include_best) const {
-    DEFINE_SPP_SERIALIZED_PROTO_IMPL(SampleEncodeAndScore,
-                                     ImmutableNBestSentencePieceText, input,
-                                     num_samples, alpha, wor, include_best);
-  }
-
-  // TODO(taku): Remove this API and use std::vector<std::string_view>
-  virtual util::bytes DecodePiecesAsSerializedProto(
-      const std::vector<std::string> &pieces) const {
-    DEFINE_SPP_SERIALIZED_PROTO_IMPL(Decode, ImmutableSentencePieceText,
-                                     pieces);
-  }
-
-  virtual util::bytes DecodePiecesAsSerializedProto(
-      const std::vector<absl::string_view> &pieces) const {
-    DEFINE_SPP_SERIALIZED_PROTO_IMPL(Decode, ImmutableSentencePieceText,
-                                     pieces);
-  }
-
-  virtual util::bytes DecodeIdsAsSerializedProto(
-      const std::vector<int> &ids) const {
-    DEFINE_SPP_SERIALIZED_PROTO_IMPL(Decode, ImmutableSentencePieceText, ids);
-  }
-
-  //////////////////////////////////////////////////////////////
-  // ImmutableProto API.
-  virtual ImmutableSentencePieceText EncodeAsImmutableProto(
-      absl::string_view input) const {
-    DEFINE_SPP_IMMUTABLE_PROTO_IMPL(Encode, ImmutableSentencePieceText, input);
-  }
-
-  virtual ImmutableSentencePieceText SampleEncodeAsImmutableProto(
-      absl::string_view input, int nbest_size, float alpha) const {
-    DEFINE_SPP_IMMUTABLE_PROTO_IMPL(SampleEncode, ImmutableSentencePieceText,
-                                    input, nbest_size, alpha);
-  }
-
-  virtual ImmutableNBestSentencePieceText NBestEncodeAsImmutableProto(
-      absl::string_view input, int nbest_size) const {
-    DEFINE_SPP_IMMUTABLE_PROTO_IMPL(
-        NBestEncode, ImmutableNBestSentencePieceText, input, nbest_size);
-  }
-
-  virtual ImmutableNBestSentencePieceText SampleEncodeAndScoreAsImmutableProto(
-      absl::string_view input, int num_samples, float alpha, bool wor,
-      bool include_best) const {
-    DEFINE_SPP_IMMUTABLE_PROTO_IMPL(SampleEncodeAndScore,
-                                    ImmutableNBestSentencePieceText, input,
-                                    num_samples, alpha, wor, include_best);
-  }
-
-  // TODO(taku): Remove this API and use std::vector<std::string_view>
-  virtual ImmutableSentencePieceText DecodePiecesAsImmutableProto(
-      const std::vector<std::string> &pieces) const {
-    DEFINE_SPP_IMMUTABLE_PROTO_IMPL(Decode, ImmutableSentencePieceText, pieces);
-  }
-
-  virtual ImmutableSentencePieceText DecodePiecesAsImmutableProto(
-      const std::vector<absl::string_view> &pieces) const {
-    DEFINE_SPP_IMMUTABLE_PROTO_IMPL(Decode, ImmutableSentencePieceText, pieces);
-  }
-
-  virtual ImmutableSentencePieceText DecodeIdsAsImmutableProto(
-      const std::vector<int> &ids) const {
-    DEFINE_SPP_IMMUTABLE_PROTO_IMPL(Decode, ImmutableSentencePieceText, ids);
-  }
-
-#undef DEFINE_SPP_DIRECT_FUNC_IMPL
-#undef DEFINE_SPP_SERIALIZED_PROTO_IMPL
-#undef DEFINE_SPP_IMMUTABLE_PROTO_IMPL
-
-  //////////////////////////////////////////////////////////////
-  // Normalization methods.
-
-  // Normalize `input`.
-  virtual util::Status Normalize(absl::string_view input,
-                                 std::string *normalized) const;
-
-  // Normalize `input`. Stores the utf8-byte offset from
-  // the normalized string to the original input.
-  virtual util::Status Normalize(absl::string_view input,
-                                 std::string *normalized,
-                                 std::vector<size_t> *norm_to_orig) const;
-
-  virtual std::string Normalize(absl::string_view input) const;
-
-  //////////////////////////////////////////////////////////////
-  // Vocabulary management methods.
-  //
-  // Returns the size of sentence pieces, which is the same as
-  // the size of vocabulary for NMT.
-  virtual int GetPieceSize() const;
-
-  // Returns the vocab id of `piece`.
-  // Returns UNK(0) if `piece` is unknown.
-  virtual int PieceToId(absl::string_view piece) const;
-
-  // Returns the string representation of vocab with `id`.
-  virtual const std::string &IdToPiece(int id) const;
-
-  // Returns the score of `id`.
-  // Usually score is an emission log probability of unigram language
-  // model.
-  virtual float GetScore(int id) const;
-
-  // Returns true if `id` is unknown symbol.
-  virtual bool IsUnknown(int id) const;
-
-  // Returns true if `id` is control symbol.
-  virtual bool IsControl(int id) const;
-
-  // Returns true if `id` is unused symbol.
-  virtual bool IsUnused(int id) const;
-
-  // Returns true if `id` is byte symbol.
-  virtual bool IsByte(int id) const;
-
-  // Returns the reserved id.
-  // Returns -1 if not defined.
-
-  // Returns unknown (<unk>) id.
-  virtual int unk_id() const;
-
-  // Returns BOS (<s>) id.
-  virtual int bos_id() const;
-
-  // Returns EOS (</s>) id.
-  virtual int eos_id() const;
-
-  // Returns PAD (<pad>) id.
-  virtual int pad_id() const;
-
-  //////////////////////////////////////////////////////////////
-  // Model management.
-  //
-  // Allows injection of a mock model instance. `model` is moved.
-  void SetModel(std::unique_ptr<ModelInterface> &&model);
-
-  // Allows injection of a normalizer instance. `normalizer` is moved.
-  void SetNormalizer(std::unique_ptr<normalizer::Normalizer> &&normalizer);
-
-  // Returns immutable model proto. Useful to obtain extended
-  // or experimental parameters encoded in model_proto.
-  const ModelProto &model_proto() const;
-
-  // returns immutable model proto as std::string.
-  // Useful to save the state of this instance via Python's pickle object.
-  util::bytes serialized_model_proto() const;
-
-  // Returns mutable normalizer_spec.
-  // Updating the intenral normalization during the encoding/decoding are not
-  // recommended and may result in unexpected behavior. Use at your own risk.
-  NormalizerSpec *mutable_normalizer_spec() const;
-
- private:
-  enum ExtraOption { REVERSE, BOS, EOS, UNK_PIECE };
-
-  util::Status ParseExtraOptions(absl::string_view extra_option,
-                                 std::vector<ExtraOption> *extra_options) const;
-
-  util::Status ApplyExtraOptions(const std::vector<ExtraOption> &extra_options,
-                                 SentencePieceText *spt) const;
-
-  util::Status PopulateSentencePieceText(
-      absl::string_view input, absl::string_view normalized,
-      const std::vector<size_t> &norm_to_orig,
-      const std::vector<std::pair<absl::string_view, int>> &result,
-      SentencePieceText *spt) const;
-
-  std::unique_ptr<ModelInterface> model_;
-  std::unique_ptr<normalizer::Normalizer> normalizer_;
-  std::unique_ptr<normalizer::Normalizer> denormalizer_;
-
-  // Underlying model protocol buffer. The same lifetime as model_.
-  std::unique_ptr<ModelProto> model_proto_;
-
-  std::vector<ExtraOption> encode_extra_options_;
-  std::vector<ExtraOption> decode_extra_options_;
-};
-
-// Set seed value of random generator.
-// Do not set static_cast<unique_int>(-1),
-// as this seed is reserved for initializing from
-// std::random_device.
-void SetRandomGeneratorSeed(unsigned int seed);
-
-// Set the global log level. The default loglevel is 0.
-// The log is emitted only when min_log_level >= output_log_level.
-void SetMinLogLevel(int v);
-
-// IO related functions to absorb model formats.
-namespace io {
-// Loads `model_proto` from `filename`.
-// We can instantiate SentencePieceProcessor as follows:
-//
-//  auto model_proto = absl::make_unique<ModelProto>();
-//  io::LoadModelProto("//path/spm.model", model_proto.get());
-//  SentencePieceProcessor sp;
-//  CHECK_OK(sp.Load(std::move(model_proto)));
-util::Status LoadModelProto(absl::string_view, ModelProto *model_proto);
-
-// Saves `model_proto` as `filename`.
-util::Status SaveModelProto(absl::string_view, const ModelProto &model_proto);
-}  // namespace io
-}  // namespace sentencepiece
-#endif  // SENTENCEPIECE_PROCESSOR_H_
diff --git a/projects/llm_framework/include/sentencepiece/sentencepiece_trainer.h b/projects/llm_framework/include/sentencepiece/sentencepiece_trainer.h
deleted file mode 100644
index e9455180..00000000
--- a/projects/llm_framework/include/sentencepiece/sentencepiece_trainer.h
+++ /dev/null
@@ -1,206 +0,0 @@
-// Copyright 2018 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef SENTENCEPIECE_TRAINER_H_
-#define SENTENCEPIECE_TRAINER_H_
-
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-#include "sentencepiece_processor.h"
-
-namespace sentencepiece {
-
-class TrainerSpec;
-class NormalizerSpec;
-class ModelProto;
-
-namespace pretokenizer {
-class PretokenizerForTrainingInterface;
-}  // namespace pretokenizer
-
-namespace normalizer {
-class Normalizer;
-}  // namespace normalizer
-
-// Iterator over the training sentences.
-// Training sentences are loaded sequentially as follows:
-//
-// for (; !it.done(); it.Next()) {
-//    const std::string &s = it.value();
-// }
-// RETURN_IF_ERROR(it.status());
-//
-class SentenceIterator {
- public:
-  virtual ~SentenceIterator() {}
-  // Returns true if iteration finishes (including error case).
-  // Uses SentenceIterator::status() method to know whether
-  // all sentences are loaded successfully.
-  virtual bool done() const = 0;
-  virtual void Next() = 0;
-  virtual const std::string &value() const = 0;
-  virtual util::Status status() const = 0;
-};
-
-class SentencePieceTrainer {
- public:
-  // Trains SentencePiece model with `trainer_spec`.
-  // Default `normalizer_spec` is used.
-  // When `sentence_iterator` is passed, load sentences from the iterator.
-  static util::Status Train(const TrainerSpec &trainer_spec,
-                            SentenceIterator *sentence_iterator = nullptr,
-                            std::string *serialized_model_proto = nullptr);
-
-  // Trains SentencePiece model with `trainer_spec` and
-  // `normalizer_spec`.
-  // When `sentence_iterator` is passed, load sentences from the iterator.
-  static util::Status Train(const TrainerSpec &trainer_spec,
-                            const NormalizerSpec &normalizer_spec,
-                            SentenceIterator *sentence_iterator = nullptr,
-                            std::string *serialized_model_proto = nullptr);
-
-  // Trains SentencePiece model with `trainer_spec`, `normalizer_spec`
-  // and `denormalizer_spec`.
-  // When `sentence_iterator` is passed, load sentences from the iterator.
-  static util::Status Train(const TrainerSpec &trainer_spec,
-                            const NormalizerSpec &normalizer_spec,
-                            const NormalizerSpec &denormalizer_spec,
-                            SentenceIterator *sentence_iterator = nullptr,
-                            std::string *serialized_model_proto = nullptr);
-  // Trains SentencePiece model with command-line string in `args`,
-  // e.g.,
-  // '--input=data --model_prefix=m --vocab_size=8192 model_type=unigram'
-  // When `sentence_iterator` is passed, load sentences from the iterator.
-  static util::Status Train(absl::string_view args,
-                            SentenceIterator *sentence_iterator = nullptr,
-                            std::string *serialized_model_proto = nullptr);
-
-  // Trains SentencePiece model with mapin `kwargs`.
-  // e.g., {{"input", "data"}, {"model_prefix, "m"}, {"vocab_size", "8192"}...}
-  static util::Status Train(
-      const std::unordered_map<std::string, std::string> &kwargs,
-      SentenceIterator *sentence_iterator = nullptr,
-      std::string *serialized_model_proto = nullptr);
-
-  // The same as above, but passes the list of sentences.
-  static util::Status Train(absl::string_view args,
-                            const std::vector<std::string> &sentences,
-                            std::string *serialized_model_proto = nullptr);
-
-  // The same as above, but passes the list of sentences.
-  static util::Status Train(
-      const std::unordered_map<std::string, std::string> &kwargs,
-      const std::vector<std::string> &sentences,
-      std::string *serialized_model_proto = nullptr);
-
-  // Handy function to make a normalizer spec from the pre-compiled
-  // normalization name. Do not use this method in production as it crashes
-  // When `name` is invalid. Useful for unittesting.
-  static NormalizerSpec GetNormalizerSpec(absl::string_view name);
-
-  // Populates necessary fields (precompiled_charmap) from
-  // `NormalizerSpec::name` or `NormalizerSpec::normalization_rule_tsv`.
-  static util::Status PopulateNormalizerSpec(NormalizerSpec *normalizer_spec,
-                                             bool is_denormalizer = false);
-
-  // Overrides `trainer_spec`, `normalizer_spec`, `denormalizer_spec` with the
-  // std::unordered_map in `kargs`.
-  static util::Status MergeSpecsFromArgs(
-      const std::unordered_map<std::string, std::string> &kwargs,
-      TrainerSpec *trainer_spec, NormalizerSpec *normalizer_spec,
-      NormalizerSpec *denormalizer_spec);
-
-  // Overrides `trainer_spec`, `normalizer_spec`, `denormalizer_spec` with the
-  // command line flags in `args`.
-  static util::Status MergeSpecsFromArgs(absl::string_view args,
-                                         TrainerSpec *trainer_spec,
-                                         NormalizerSpec *normalizer_spec,
-                                         NormalizerSpec *denormalizer_spec);
-
-  // Injects global pre-tokenizer that are applied in training time.
-  // Pretokenizer is only used for extracting pieces.
-  // TODO(taku): It would be better to inject per `trainer_spec`.
-  static util::Status SetPretokenizerForTraining(
-      const pretokenizer::PretokenizerForTrainingInterface *pretokenizer);
-
-  // Returns the current pretokenizer. if no pretokenizer is defined, returns
-  // nullptr.
-  static const pretokenizer::PretokenizerForTrainingInterface *
-  GetPretokenizerForTraining();
-
-  // Helper function to set `field_name=value` in `message`.
-  // When `field_name` is repeated, multiple values can be passed
-  // with comma-separated values. `field_name` must not be a nested message.
-  // The body of these functions are automatically generated with
-  // data/gen_spec_parser.pl
-  static util::Status SetProtoField(absl::string_view name,
-                                    absl::string_view value,
-                                    TrainerSpec *message);
-
-  static util::Status SetProtoField(absl::string_view name,
-                                    absl::string_view value,
-                                    NormalizerSpec *message);
-
-  // Populates model type from string representation, e.g., "bpe".
-  // Supported model: "unigram", "bpe", "word", "char".
-  static util::Status PopulateModelTypeFromString(absl::string_view type,
-                                                  TrainerSpec *trainer_spec);
-
- private:
-  SentencePieceTrainer() {}
-  ~SentencePieceTrainer() {}
-};
-
-class SentencePieceNormalizer {
- public:
-  SentencePieceNormalizer();
-  virtual ~SentencePieceNormalizer();
-
-  virtual util::Status Load(std::unique_ptr<ModelProto> model_proto);
-
-  virtual util::Status Load(absl::string_view filename);
-
-  virtual util::Status LoadFromSerializedProto(absl::string_view serialized);
-
-  virtual util::Status LoadFromRuleTSV(absl::string_view filename);
-
-  virtual util::Status LoadFromRuleName(absl::string_view name);
-
-  virtual util::Status Normalize(absl::string_view input,
-                                 std::string *normalized) const;
-
-  virtual util::Status Normalize(absl::string_view input,
-                                 std::string *normalized,
-                                 std::vector<size_t> *norm_to_orig) const;
-
-  virtual std::string Normalize(absl::string_view input) const;
-
-  virtual NormalizerSpec *mutable_normalizer_spec() const;
-
-  virtual std::string serialized_model_proto() const;
-
- private:
-  std::unique_ptr<normalizer::Normalizer> normalizer_;
-  std::unique_ptr<ModelProto> model_proto_;
-};
-
-// Converts the utf8 byte spans into Unicode char span.
-void ConvertToUnicodeAlignment(absl::string_view orig, absl::string_view norm,
-                               std::vector<size_t> *norm_to_orig);
-
-}  // namespace sentencepiece
-
-#endif  // SENTENCEPIECE_TRAINER_H_
diff --git a/projects/llm_framework/include/sentencepiece/spec_parser.h b/projects/llm_framework/include/sentencepiece/spec_parser.h
deleted file mode 100644
index 724d11bc..00000000
--- a/projects/llm_framework/include/sentencepiece/spec_parser.h
+++ /dev/null
@@ -1,284 +0,0 @@
-// Copyright 2016 Google LLC.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef SPEC_PARSER_H_
-#define SPEC_PARSER_H_
-
-#include <string>
-#include <vector>
-
-#include "sentencepiece_processor.h"
-#include "third_party/absl/strings/ascii.h"
-#include "third_party/absl/strings/str_split.h"
-#include "util.h"
-
-namespace sentencepiece {
-
-#define PARSE_STRING(param_name)                   \
-  if (name == #param_name) {                       \
-    message->set_##param_name(std::string(value)); \
-    return util::OkStatus();                       \
-  }
-
-#define PARSE_REPEATED_STRING(param_name)                       \
-  if (name == #param_name) {                                    \
-    for (const std::string &val : util::StrSplitAsCSV(value)) { \
-      message->add_##param_name(val);                           \
-    }                                                           \
-    return util::OkStatus();                                    \
-  }
-
-#define PARSE_BYTE(param_name)                             \
-  if (name == #param_name) {                               \
-    message->set_##param_name(value.data(), value.size()); \
-    return util::OkStatus();                               \
-  }
-
-#define PARSE_INT32(param_name)                                               \
-  if (name == #param_name) {                                                  \
-    int32 v;                                                                  \
-    if (!string_util::lexical_cast(value, &v))                                \
-      return util::StatusBuilder(util::StatusCode::kInvalidArgument, GTL_LOC) \
-             << "cannot parse \"" << value << "\" as int.";                   \
-    message->set_##param_name(v);                                             \
-    return util::OkStatus();                                                  \
-  }
-
-#define PARSE_UINT64(param_name)                                              \
-  if (name == #param_name) {                                                  \
-    uint64 v;                                                                 \
-    if (!string_util::lexical_cast(value, &v))                                \
-      return util::StatusBuilder(util::StatusCode::kInvalidArgument, GTL_LOC) \
-             << "cannot parse \"" << value << "\" as int.";                   \
-    message->set_##param_name(v);                                             \
-    return util::OkStatus();                                                  \
-  }
-
-#define PARSE_DOUBLE(param_name)                                              \
-  if (name == #param_name) {                                                  \
-    double v;                                                                 \
-    if (!string_util::lexical_cast(value, &v))                                \
-      return util::StatusBuilder(util::StatusCode::kInvalidArgument, GTL_LOC) \
-             << "cannot parse \"" << value << "\" as int.";                   \
-    message->set_##param_name(v);                                             \
-    return util::OkStatus();                                                  \
-  }
-
-#define PARSE_BOOL(param_name)                                                \
-  if (name == #param_name) {                                                  \
-    bool v;                                                                   \
-    if (!string_util::lexical_cast(value.empty() ? "true" : value, &v))       \
-      return util::StatusBuilder(util::StatusCode::kInvalidArgument, GTL_LOC) \
-             << "cannot parse \"" << value << "\" as bool.";                  \
-    message->set_##param_name(v);                                             \
-    return util::OkStatus();                                                  \
-  }
-
-#define PARSE_ENUM(param_name, map_name)                                      \
-  if (name == #param_name) {                                                  \
-    const auto it = map_name.find(absl::AsciiStrToUpper(value));              \
-    if (it == map_name.end())                                                 \
-      return util::StatusBuilder(util::StatusCode::kInvalidArgument, GTL_LOC) \
-             << "unknown enumeration value of \"" << value << "\" as "        \
-             << #map_name;                                                    \
-    message->set_##param_name(it->second);                                    \
-    return util::OkStatus();                                                  \
-  }
-
-#define PRINT_PARAM(param_name) \
-  os << "  " << #param_name << ": " << message.param_name() << "\n";
-
-#define PRINT_REPEATED_STRING(param_name)    \
-  for (const auto &v : message.param_name()) \
-    os << "  " << #param_name << ": " << v << "\n";
-
-#define PRINT_ENUM(param_name, map_name)               \
-  const auto it = map_name.find(message.param_name()); \
-  if (it == map_name.end())                            \
-    os << "  " << #param_name << ": unknown\n";        \
-  else                                                 \
-    os << "  " << #param_name << ": " << it->second << "\n";
-
-inline std::string PrintProto(const TrainerSpec &message,
-                              absl::string_view name) {
-  std::ostringstream os;
-
-  os << name << " {\n";
-
-  PRINT_REPEATED_STRING(input);
-  PRINT_PARAM(input_format);
-  PRINT_PARAM(model_prefix);
-
-  static const std::map<TrainerSpec::ModelType, std::string> kModelType_Map = {
-      {TrainerSpec::UNIGRAM, "UNIGRAM"},
-      {TrainerSpec::BPE, "BPE"},
-      {TrainerSpec::WORD, "WORD"},
-      {TrainerSpec::CHAR, "CHAR"},
-  };
-
-  PRINT_ENUM(model_type, kModelType_Map);
-  PRINT_PARAM(vocab_size);
-  PRINT_REPEATED_STRING(accept_language);
-  PRINT_PARAM(self_test_sample_size);
-  PRINT_PARAM(character_coverage);
-  PRINT_PARAM(input_sentence_size);
-  PRINT_PARAM(shuffle_input_sentence);
-  PRINT_PARAM(seed_sentencepiece_size);
-  PRINT_PARAM(shrinking_factor);
-  PRINT_PARAM(max_sentence_length);
-  PRINT_PARAM(num_threads);
-  PRINT_PARAM(num_sub_iterations);
-  PRINT_PARAM(max_sentencepiece_length);
-  PRINT_PARAM(split_by_unicode_script);
-  PRINT_PARAM(split_by_number);
-  PRINT_PARAM(split_by_whitespace);
-  PRINT_PARAM(split_digits);
-  PRINT_PARAM(pretokenization_delimiter);
-  PRINT_PARAM(treat_whitespace_as_suffix);
-  PRINT_PARAM(allow_whitespace_only_pieces);
-  PRINT_REPEATED_STRING(control_symbols);
-  PRINT_REPEATED_STRING(user_defined_symbols);
-  PRINT_PARAM(required_chars);
-  PRINT_PARAM(byte_fallback);
-  PRINT_PARAM(vocabulary_output_piece_score);
-  PRINT_PARAM(train_extremely_large_corpus);
-  PRINT_PARAM(seed_sentencepieces_file);
-  PRINT_PARAM(hard_vocab_limit);
-  PRINT_PARAM(use_all_vocab);
-  PRINT_PARAM(unk_id);
-  PRINT_PARAM(bos_id);
-  PRINT_PARAM(eos_id);
-  PRINT_PARAM(pad_id);
-  PRINT_PARAM(unk_piece);
-  PRINT_PARAM(bos_piece);
-  PRINT_PARAM(eos_piece);
-  PRINT_PARAM(pad_piece);
-  PRINT_PARAM(unk_surface);
-  PRINT_PARAM(enable_differential_privacy);
-  PRINT_PARAM(differential_privacy_noise_level);
-  PRINT_PARAM(differential_privacy_clipping_threshold);
-
-  os << "}\n";
-
-  return os.str();
-}
-
-inline std::string PrintProto(const NormalizerSpec &message,
-                              absl::string_view name) {
-  std::ostringstream os;
-
-  os << name << " {\n";
-
-  PRINT_PARAM(name);
-  PRINT_PARAM(add_dummy_prefix);
-  PRINT_PARAM(remove_extra_whitespaces);
-  PRINT_PARAM(escape_whitespaces);
-  PRINT_PARAM(normalization_rule_tsv);
-
-  os << "}\n";
-
-  return os.str();
-}
-
-util::Status SentencePieceTrainer::SetProtoField(absl::string_view name,
-                                                 absl::string_view value,
-                                                 TrainerSpec *message) {
-  CHECK_OR_RETURN(message);
-
-  PARSE_REPEATED_STRING(input);
-  PARSE_STRING(input_format);
-  PARSE_STRING(model_prefix);
-
-  static const std::map<std::string, TrainerSpec::ModelType> kModelType_Map = {
-      {"UNIGRAM", TrainerSpec::UNIGRAM},
-      {"BPE", TrainerSpec::BPE},
-      {"WORD", TrainerSpec::WORD},
-      {"CHAR", TrainerSpec::CHAR},
-  };
-
-  PARSE_ENUM(model_type, kModelType_Map);
-  PARSE_INT32(vocab_size);
-  PARSE_REPEATED_STRING(accept_language);
-  PARSE_INT32(self_test_sample_size);
-  PARSE_DOUBLE(character_coverage);
-  PARSE_UINT64(input_sentence_size);
-  PARSE_BOOL(shuffle_input_sentence);
-  PARSE_INT32(seed_sentencepiece_size);
-  PARSE_DOUBLE(shrinking_factor);
-  PARSE_INT32(max_sentence_length);
-  PARSE_INT32(num_threads);
-  PARSE_INT32(num_sub_iterations);
-  PARSE_INT32(max_sentencepiece_length);
-  PARSE_BOOL(split_by_unicode_script);
-  PARSE_BOOL(split_by_number);
-  PARSE_BOOL(split_by_whitespace);
-  PARSE_BOOL(split_digits);
-  PARSE_STRING(pretokenization_delimiter);
-  PARSE_BOOL(treat_whitespace_as_suffix);
-  PARSE_BOOL(allow_whitespace_only_pieces);
-  PARSE_REPEATED_STRING(control_symbols);
-  PARSE_REPEATED_STRING(user_defined_symbols);
-  PARSE_STRING(required_chars);
-  PARSE_BOOL(byte_fallback);
-  PARSE_BOOL(hard_vocab_limit);
-  PARSE_BOOL(vocabulary_output_piece_score);
-  PARSE_BOOL(train_extremely_large_corpus);
-  PARSE_STRING(seed_sentencepieces_file);
-  PARSE_BOOL(use_all_vocab);
-  PARSE_INT32(unk_id);
-  PARSE_INT32(bos_id);
-  PARSE_INT32(eos_id);
-  PARSE_INT32(pad_id);
-  PARSE_STRING(unk_piece);
-  PARSE_STRING(bos_piece);
-  PARSE_STRING(eos_piece);
-  PARSE_STRING(pad_piece);
-  PARSE_STRING(unk_surface);
-  PARSE_BOOL(enable_differential_privacy);
-  PARSE_DOUBLE(differential_privacy_noise_level);
-  PARSE_UINT64(differential_privacy_clipping_threshold);
-
-  return util::StatusBuilder(util::StatusCode::kNotFound, GTL_LOC)
-         << "unknown field name \"" << name << "\" in TrainerSpec.";
-}
-
-util::Status SentencePieceTrainer::SetProtoField(absl::string_view name,
-                                                 absl::string_view value,
-                                                 NormalizerSpec *message) {
-  CHECK_OR_RETURN(message);
-
-  PARSE_STRING(name);
-  PARSE_BYTE(precompiled_charsmap);
-  PARSE_BOOL(add_dummy_prefix);
-  PARSE_BOOL(remove_extra_whitespaces);
-  PARSE_BOOL(escape_whitespaces);
-  PARSE_STRING(normalization_rule_tsv);
-
-  return util::StatusBuilder(util::StatusCode::kNotFound, GTL_LOC)
-         << "unknown field name \"" << name << "\" in NormalizerSpec.";
-}
-
-#undef PARSE_STRING
-#undef PARSE_REPEATED_STRING
-#undef PARSE_BOOL
-#undef PARSE_BYTE
-#undef PARSE_INT32
-#undef PARSE_DUOBLE
-#undef PARSE_ENUM
-#undef PRINT_MAP
-#undef PRINT_REPEATED_STRING
-#undef PRINT_ENUM
-}  // namespace sentencepiece
-
-#endif  // SPEC_PARSER_H_
diff --git a/projects/llm_framework/include/sentencepiece/testharness.h b/projects/llm_framework/include/sentencepiece/testharness.h
deleted file mode 100644
index 3984c351..00000000
--- a/projects/llm_framework/include/sentencepiece/testharness.h
+++ /dev/null
@@ -1,227 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef TESTHARNESS_H_
-#define TESTHARNESS_H_
-
-#include <cmath>
-#include <iostream>
-#include <sstream>
-#include <string>
-
-#include "common.h"
-#include "third_party/absl/flags/flag.h"
-#include "third_party/absl/flags/parse.h"
-#include "third_party/absl/strings/string_view.h"
-
-ABSL_DECLARE_FLAG(std::string, test_tmpdir);
-ABSL_DECLARE_FLAG(std::string, test_srcdir);
-
-namespace testing {
-inline std::string TempDir() { return absl::GetFlag(FLAGS_test_tmpdir); }
-inline std::string SrcDir() { return absl::GetFlag(FLAGS_test_srcdir); }
-}  // namespace testing
-
-namespace sentencepiece {
-namespace test {
-// Run some of the tests registered by the TEST() macro.
-// TEST(Foo, Hello) { ... }
-// TEST(Foo, World) { ... }
-//
-// Returns 0 if all tests pass.
-// Dies or returns a non-zero value if some test fails.
-int RunAllTests();
-
-// An instance of Tester is allocated to hold temporary state during
-// the execution of an assertion.
-class Tester {
- public:
-  Tester(const char *fname, int line) : ok_(true), fname_(fname), line_(line) {}
-
-  ~Tester() {
-    if (!ok_) {
-      std::cerr << "[       NG ] " << fname_ << ":" << line_ << ":" << ss_.str()
-                << std::endl;
-      exit(-1);
-    }
-  }
-
-  Tester &Is(bool b, const char *msg) {
-    if (!b) {
-      ss_ << " failed: " << msg;
-      ok_ = false;
-    }
-    return *this;
-  }
-
-  Tester &IsNear(double val1, double val2, double abs_error, const char *msg1,
-                 const char *msg2) {
-    const double diff = std::fabs(val1 - val2);
-    if (diff > abs_error) {
-      ss_ << "The difference between (" << msg1 << ") and (" << msg2 << ") is "
-          << diff << ", which exceeds " << abs_error << ", where\n"
-          << msg1 << " evaluates to " << val1 << ",\n"
-          << msg2 << " evaluates to " << val2;
-      ok_ = false;
-    }
-    return *this;
-  }
-
-#define BINARY_OP(name, op)                                                  \
-  template <class X, class Y>                                                \
-  Tester &name(const X &x, const Y &y, const char *msg1, const char *msg2) { \
-    if (!(x op y)) {                                                         \
-      ss_ << " failed: " << msg1 << (" " #op " ") << msg2;                   \
-      ok_ = false;                                                           \
-    }                                                                        \
-    return *this;                                                            \
-  }
-
-  BINARY_OP(IsEq, ==)
-  BINARY_OP(IsNe, !=)
-  BINARY_OP(IsGe, >=)
-  BINARY_OP(IsGt, >)
-  BINARY_OP(IsLe, <=)
-  BINARY_OP(IsLt, <)
-#undef BINARY_OP
-
-  // Attach the specified value to the error message if an error has occurred
-  template <class V>
-  Tester &operator<<(const V &value) {
-    if (!ok_) {
-      ss_ << " " << value;
-    }
-    return *this;
-  }
-
- private:
-  bool ok_;
-  const char *fname_;
-  int line_;
-  std::stringstream ss_;
-};
-
-#define EXPECT_TRUE(c) \
-  sentencepiece::test::Tester(__FILE__, __LINE__).Is((c), #c)
-#define EXPECT_FALSE(c) \
-  sentencepiece::test::Tester(__FILE__, __LINE__).Is((!(c)), #c)
-#define EXPECT_STREQ(a, b)                        \
-  sentencepiece::test::Tester(__FILE__, __LINE__) \
-      .IsEq(std::string(a), std::string(b), #a, #b)
-#define EXPECT_EQ(a, b) \
-  sentencepiece::test::Tester(__FILE__, __LINE__).IsEq((a), (b), #a, #b)
-#define EXPECT_NE(a, b) \
-  sentencepiece::test::Tester(__FILE__, __LINE__).IsNe((a), (b), #a, #b)
-#define EXPECT_GE(a, b) \
-  sentencepiece::test::Tester(__FILE__, __LINE__).IsGe((a), (b), #a, #b)
-#define EXPECT_GT(a, b) \
-  sentencepiece::test::Tester(__FILE__, __LINE__).IsGt((a), (b), #a, #b)
-#define EXPECT_LE(a, b) \
-  sentencepiece::test::Tester(__FILE__, __LINE__).IsLe((a), (b), #a, #b)
-#define EXPECT_LT(a, b) \
-  sentencepiece::test::Tester(__FILE__, __LINE__).IsLt((a), (b), #a, #b)
-#define EXPECT_NEAR(a, b, c) \
-  sentencepiece::test::Tester(__FILE__, __LINE__).IsNear((a), (b), (c), #a, #b)
-#define EXPECT_OK(c) EXPECT_EQ(c, ::sentencepiece::util::OkStatus())
-#define EXPECT_NOT_OK(c) EXPECT_NE(c, ::sentencepiece::util::OkStatus())
-
-#define EXPECT_DEATH(statement, condition)   \
-  {                                          \
-    sentencepiece::error::SetTestCounter(1); \
-    statement;                               \
-    sentencepiece::error::SetTestCounter(0); \
-  };
-
-#define ASSERT_TRUE EXPECT_TRUE
-#define ASSERT_FALSE EXPECT_FALSE
-#define ASSERT_STREQ EXPECT_STREQ
-#define ASSERT_EQ EXPECT_EQ
-#define ASSERT_NE EXPECT_NE
-#define ASSERT_GE EXPECT_GE
-#define ASSERT_GT EXPECT_GT
-#define ASSERT_LE EXPECT_LE
-#define ASSERT_LT EXPECT_LT
-#define ASSERT_NEAR EXPECT_NEAR
-#define ASSERT_NOT_OK EXPECT_NOT_OK
-#define ASSERT_DEATH ASSERT_DEATH
-
-template <typename T>
-class TestWithParam {
- public:
-  using ParamType = T;
-  virtual void SetUp() {}
-  virtual void TearDown() {}
-  virtual ~TestWithParam() {}
-  virtual ParamType GetParam() const { return ParamType(); }
-};
-
-template <typename T>
-std::vector<T> ValuesIn(const std::vector<T> &v) {
-  return v;
-}
-
-#define TCONCAT(a, b, c) TCONCAT1(a, b, c)
-#define TCONCAT1(a, b, c) a##b##c
-
-#define INSTANTIATE_TEST_SUITE_P(suite_base, base, params)           \
-  std::vector<base::ParamType> TCONCAT(base, _get_params_, base)() { \
-    return params;                                                   \
-  }
-
-#define TEST(base, name)                                                       \
-  class TCONCAT(base, _Test_, name) {                                          \
-   public:                                                                     \
-    void _Run();                                                               \
-    static void _RunIt() {                                                     \
-      TCONCAT(base, _Test_, name) t;                                           \
-      t._Run();                                                                \
-    }                                                                          \
-  };                                                                           \
-  bool TCONCAT(base, _Test_ignored_, name) =                                   \
-      sentencepiece::test::RegisterTest(#base, #name,                          \
-                                        &TCONCAT(base, _Test_, name)::_RunIt); \
-  void TCONCAT(base, _Test_, name)::_Run()
-
-#define TEST_P(base, name)                                          \
-  std::vector<base::ParamType> TCONCAT(base, _get_params_, base)(); \
-  class TCONCAT(base, _Test_p_, name) : public base {               \
-   public:                                                          \
-    std::vector<ParamType> GetParams() const {                      \
-      return TCONCAT(base, _get_params_, base)();                   \
-    }                                                               \
-    ParamType param_;                                               \
-    void SetParam(const ParamType &param) { param_ = param; }       \
-    ParamType GetParam() const { return param_; }                   \
-    void _Run();                                                    \
-    static void _RunIt() {                                          \
-      TCONCAT(base, _Test_p_, name) t;                              \
-      for (const auto &param : t.GetParams()) {                     \
-        t.SetParam(param);                                          \
-        t.SetUp();                                                  \
-        t._Run();                                                   \
-        t.TearDown();                                               \
-      }                                                             \
-    }                                                               \
-  };                                                                \
-  bool TCONCAT(base, _Test_p_ignored_, name) =                      \
-      sentencepiece::test::RegisterTest(                            \
-          #base, #name, &TCONCAT(base, _Test_p_, name)::_RunIt);    \
-  void TCONCAT(base, _Test_p_, name)::_Run()
-
-// Register the specified test.  Typically not used directly, but
-// invoked via the macro expansion of TEST.
-extern bool RegisterTest(const char *base, const char *name, void (*func)());
-}  // namespace test
-}  // namespace sentencepiece
-#endif  // TESTHARNESS_H_
diff --git a/projects/llm_framework/include/sentencepiece/trainer_factory.h b/projects/llm_framework/include/sentencepiece/trainer_factory.h
deleted file mode 100644
index a11cbc0f..00000000
--- a/projects/llm_framework/include/sentencepiece/trainer_factory.h
+++ /dev/null
@@ -1,33 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef TRAINER_FACTORY_H_
-#define TRAINER_FACTORY_H_
-
-#include <memory>
-
-#include "sentencepiece_model.pb.h"
-#include "trainer_interface.h"
-
-namespace sentencepiece {
-
-class TrainerFactory {
- public:
-  // Creates Trainer instance from |trainer_spec| and |normalizer_spec|.
-  static std::unique_ptr<TrainerInterface> Create(
-      const TrainerSpec &trainer_spec, const NormalizerSpec &normalizer_spec,
-      const NormalizerSpec &denormalizer_spec);
-};
-}  // namespace sentencepiece
-#endif  // TRAINER_FACTORY_H_
diff --git a/projects/llm_framework/include/sentencepiece/trainer_interface.h b/projects/llm_framework/include/sentencepiece/trainer_interface.h
deleted file mode 100644
index 8d625a95..00000000
--- a/projects/llm_framework/include/sentencepiece/trainer_interface.h
+++ /dev/null
@@ -1,183 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef TRAINER_INTERFACE_H_
-#define TRAINER_INTERFACE_H_
-
-#include <algorithm>
-#include <map>
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "common.h"
-#include "filesystem.h"
-#include "sentencepiece_model.pb.h"
-#include "sentencepiece_processor.h"
-#include "sentencepiece_trainer.h"
-#include "third_party/absl/container/flat_hash_map.h"
-#include "util.h"
-
-namespace sentencepiece {
-
-template <typename K, typename V>
-std::vector<std::pair<K, V>> Sorted(const std::vector<std::pair<K, V>> &m) {
-  std::vector<std::pair<K, V>> v = m;
-  std::sort(v.begin(), v.end(),
-            [](const std::pair<K, V> &p1, const std::pair<K, V> &p2) {
-              return (p1.second > p2.second ||
-                      (p1.second == p2.second && p1.first < p2.first));
-            });
-  return v;
-}
-
-template <typename K, typename V>
-std::vector<std::pair<K, V>> Sorted(const absl::flat_hash_map<K, V> &m) {
-  std::vector<std::pair<K, V>> v(m.begin(), m.end());
-  return Sorted(v);
-}
-
-class MultiFileSentenceIterator : public SentenceIterator {
- public:
-  explicit MultiFileSentenceIterator(const std::vector<std::string> &files);
-  ~MultiFileSentenceIterator() {}
-
-  bool done() const override;
-  void Next() override;
-  const std::string &value() const override { return value_; }
-  util::Status status() const override;
-
- private:
-  void TryRead();
-
-  bool read_done_ = false;
-  size_t file_index_ = 0;
-  std::vector<std::string> files_;
-  std::string value_;
-  std::unique_ptr<filesystem::ReadableFile> fp_;
-};
-
-// Base trainer class
-class TrainerInterface {
- public:
-  using Sentence = std::pair<std::string, int64>;
-  using Sentences = std::vector<Sentence>;
-
-  static const char32 kWSChar;
-  static const char32 kUNKChar;
-  static const char32 kUPPBoundaryChar;
-  static const char kWSStr[];
-  static const char kUNKStr[];
-  static const char kUPPBoundaryStr[];
-
-  TrainerInterface(const TrainerSpec &trainer_spec,
-                   const NormalizerSpec &normalizer_spec,
-                   const NormalizerSpec &denormalizer_spec);
-
-  virtual ~TrainerInterface();
-
-  // Loads sentence from `sentence_iterator` and stores the model
-  // to `output_model_proto`.
-  virtual util::Status Train(SentenceIterator *sentence_iterator,
-                             ModelProto *output_model_proto) {
-    sentence_iterator_ = sentence_iterator;
-    output_model_proto_ = output_model_proto;
-    return Train();
-  }
-
-  virtual util::Status Train() { return status(); }
-
-  virtual util::Status status() const { return status_; }
-
-  FRIEND_TEST(TrainerInterfaceTest, IsValidSentencePieceTest);
-  FRIEND_TEST(TrainerInterfaceTest, OverrideSpecialPiecesTest);
-  FRIEND_TEST(TrainerInterfaceTest, BytePiecesTest);
-  FRIEND_TEST(TrainerInterfaceTest, SerializeTest);
-  FRIEND_TEST(TrainerInterfaceTest, CharactersTest);
-
-  // Loads all sentences from spec.input() or SentenceIterator.
-  // It loads at most input_sentence_size sentences.
-  util::Status LoadSentences();
-
- protected:
-  // Returns true if |piece| is valid sentence piece.
-  // The result is affected by
-  // max_sentencepiece_length, split_by_whiespace, split_by_unicode_script.
-  bool IsValidSentencePiece(const string_util::UnicodeText &piece) const;
-
-  // Splits all sentencecs by whitespaces and
-  // replace the |sentences_| with tokenized string.
-  // e.g.,
-  //  [ ["hello world ", 1], ["hi world]" ] =>
-  //  [ ["hello", 1], ["hi", 1], ["world", 2] ]
-  void SplitSentencesByWhitespace();
-
-  // Save model files into spec.model_prefix().
-  util::Status Save() const;
-
-  // Set of characters which must be included in the final vocab.
-  // The value of this map stores the frequency.
-  absl::flat_hash_map<char32, int64> required_chars_;
-
-  // Final output pieces
-  std::vector<std::pair<std::string, float>> final_pieces_;
-
-  // All sentences.
-  Sentences sentences_;
-
-  // Trainer spec.
-  TrainerSpec trainer_spec_;
-
-  // Normalizer spec
-  NormalizerSpec normalizer_spec_;
-
-  // Denormalizer spec
-  NormalizerSpec denormalizer_spec_;
-
-  // Reserved control pieces. e.g., <unk>, <s>, </s>.
-  // key is vocab id.
-  std::map<int, std::pair<std::string, ModelProto::SentencePiece::Type>>
-      meta_pieces_;
-
-  // Detect errors on initialization.
-  util::Status status_;
-
-  // Loads sentences from SentenceIterator if not null.
-  SentenceIterator *sentence_iterator_ = nullptr;
-
-  // Emits model to this proto instead of file.
-  ModelProto *output_model_proto_ = nullptr;
-
- private:
-  // Serialize final_pieces_ to |model_proto|.
-  util::Status Serialize(ModelProto *model_proto) const;
-
-  // Saves the best sentence split with the current model for debugging.
-  util::Status SaveSplits(absl::string_view filename) const;
-
-  // Saves model file.
-  util::Status SaveModel(absl::string_view filename) const;
-
-  // Saves vocabulary file for NMT.
-  util::Status SaveVocab(absl::string_view filename) const;
-
-  // Initializes `meta_pieces_` from TrainerSpec.
-  util::Status InitMetaPieces();
-
-  // Randomly sampled raw sentences for self-testing.
-  std::vector<std::string> self_test_samples_;
-};
-}  // namespace sentencepiece
-#endif  // TRAINER_INTERFACE_H_
diff --git a/projects/llm_framework/include/sentencepiece/unicode_script.h b/projects/llm_framework/include/sentencepiece/unicode_script.h
deleted file mode 100644
index 67042c00..00000000
--- a/projects/llm_framework/include/sentencepiece/unicode_script.h
+++ /dev/null
@@ -1,165 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef UNICODE_SCRIPT_H_
-#define UNICODE_SCRIPT_H_
-
-#include "common.h"
-
-namespace sentencepiece {
-namespace unicode_script {
-enum ScriptType : int32_t {
-  U_Adlam,
-  U_Ahom,
-  U_Anatolian_Hieroglyphs,
-  U_Arabic,
-  U_Armenian,
-  U_Avestan,
-  U_Balinese,
-  U_Bamum,
-  U_Bassa_Vah,
-  U_Batak,
-  U_Bengali,
-  U_Bhaiksuki,
-  U_Bopomofo,
-  U_Brahmi,
-  U_Braille,
-  U_Buginese,
-  U_Buhid,
-  U_Canadian_Aboriginal,
-  U_Carian,
-  U_Caucasian_Albanian,
-  U_Chakma,
-  U_Cham,
-  U_Cherokee,
-  U_Common,
-  U_Coptic,
-  U_Cuneiform,
-  U_Cypriot,
-  U_Cyrillic,
-  U_Deseret,
-  U_Devanagari,
-  U_Duployan,
-  U_Egyptian_Hieroglyphs,
-  U_Elbasan,
-  U_Ethiopic,
-  U_Georgian,
-  U_Glagolitic,
-  U_Gothic,
-  U_Grantha,
-  U_Greek,
-  U_Gujarati,
-  U_Gurmukhi,
-  U_Han,
-  U_Hangul,
-  U_Hanunoo,
-  U_Hatran,
-  U_Hebrew,
-  U_Hiragana,
-  U_Imperial_Aramaic,
-  U_Inherited,
-  U_Inscriptional_Pahlavi,
-  U_Inscriptional_Parthian,
-  U_Javanese,
-  U_Kaithi,
-  U_Kannada,
-  U_Katakana,
-  U_Kayah_Li,
-  U_Kharoshthi,
-  U_Khmer,
-  U_Khojki,
-  U_Khudawadi,
-  U_Lao,
-  U_Latin,
-  U_Lepcha,
-  U_Limbu,
-  U_Linear_A,
-  U_Linear_B,
-  U_Lisu,
-  U_Lycian,
-  U_Lydian,
-  U_Mahajani,
-  U_Malayalam,
-  U_Mandaic,
-  U_Manichaean,
-  U_Marchen,
-  U_Meetei_Mayek,
-  U_Mende_Kikakui,
-  U_Meroitic_Cursive,
-  U_Meroitic_Hieroglyphs,
-  U_Miao,
-  U_Modi,
-  U_Mongolian,
-  U_Mro,
-  U_Multani,
-  U_Myanmar,
-  U_Nabataean,
-  U_New_Tai_Lue,
-  U_Newa,
-  U_Nko,
-  U_Ogham,
-  U_Ol_Chiki,
-  U_Old_Hungarian,
-  U_Old_Italic,
-  U_Old_North_Arabian,
-  U_Old_Permic,
-  U_Old_Persian,
-  U_Old_South_Arabian,
-  U_Old_Turkic,
-  U_Oriya,
-  U_Osage,
-  U_Osmanya,
-  U_Pahawh_Hmong,
-  U_Palmyrene,
-  U_Pau_Cin_Hau,
-  U_Phags_Pa,
-  U_Phoenician,
-  U_Psalter_Pahlavi,
-  U_Rejang,
-  U_Runic,
-  U_Samaritan,
-  U_Saurashtra,
-  U_Sharada,
-  U_Shavian,
-  U_Siddham,
-  U_SignWriting,
-  U_Sinhala,
-  U_Sora_Sompeng,
-  U_Sundanese,
-  U_Syloti_Nagri,
-  U_Syriac,
-  U_Tagalog,
-  U_Tagbanwa,
-  U_Tai_Le,
-  U_Tai_Tham,
-  U_Tai_Viet,
-  U_Takri,
-  U_Tamil,
-  U_Tangut,
-  U_Telugu,
-  U_Thaana,
-  U_Thai,
-  U_Tibetan,
-  U_Tifinagh,
-  U_Tirhuta,
-  U_Ugaritic,
-  U_Vai,
-  U_Warang_Citi,
-  U_Yi
-};
-
-ScriptType GetScript(char32 c);
-}  // namespace unicode_script
-}  // namespace sentencepiece
-#endif  // UNICODE_SCRIPT
diff --git a/projects/llm_framework/include/sentencepiece/unicode_script_map.h b/projects/llm_framework/include/sentencepiece/unicode_script_map.h
deleted file mode 100644
index f2e67e9e..00000000
--- a/projects/llm_framework/include/sentencepiece/unicode_script_map.h
+++ /dev/null
@@ -1,1955 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef UNICODE_SCRIPT_DATA_H_
-#define UNICODE_SCRIPT_DATA_H_
-#include "third_party/absl/container/flat_hash_map.h"
-namespace sentencepiece {
-namespace unicode_script {
-namespace {
-void InitTable(absl::flat_hash_map<char32, ScriptType> *smap) {
-  for (char32 c = 0x0000; c <= 0x001F; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x0020] = U_Common;
-  for (char32 c = 0x0021; c <= 0x0023; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x0024] = U_Common;
-  for (char32 c = 0x0025; c <= 0x0027; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x0028] = U_Common;
-  (*smap)[0x0029] = U_Common;
-  (*smap)[0x002A] = U_Common;
-  (*smap)[0x002B] = U_Common;
-  (*smap)[0x002C] = U_Common;
-  (*smap)[0x002D] = U_Common;
-  for (char32 c = 0x002E; c <= 0x002F; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x0030; c <= 0x0039; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x003A; c <= 0x003B; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x003C; c <= 0x003E; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x003F; c <= 0x0040; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x005B] = U_Common;
-  (*smap)[0x005C] = U_Common;
-  (*smap)[0x005D] = U_Common;
-  (*smap)[0x005E] = U_Common;
-  (*smap)[0x005F] = U_Common;
-  (*smap)[0x0060] = U_Common;
-  (*smap)[0x007B] = U_Common;
-  (*smap)[0x007C] = U_Common;
-  (*smap)[0x007D] = U_Common;
-  (*smap)[0x007E] = U_Common;
-  for (char32 c = 0x007F; c <= 0x009F; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x00A0] = U_Common;
-  (*smap)[0x00A1] = U_Common;
-  for (char32 c = 0x00A2; c <= 0x00A5; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x00A6] = U_Common;
-  (*smap)[0x00A7] = U_Common;
-  (*smap)[0x00A8] = U_Common;
-  (*smap)[0x00A9] = U_Common;
-  (*smap)[0x00AB] = U_Common;
-  (*smap)[0x00AC] = U_Common;
-  (*smap)[0x00AD] = U_Common;
-  (*smap)[0x00AE] = U_Common;
-  (*smap)[0x00AF] = U_Common;
-  (*smap)[0x00B0] = U_Common;
-  (*smap)[0x00B1] = U_Common;
-  for (char32 c = 0x00B2; c <= 0x00B3; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x00B4] = U_Common;
-  (*smap)[0x00B5] = U_Common;
-  for (char32 c = 0x00B6; c <= 0x00B7; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x00B8] = U_Common;
-  (*smap)[0x00B9] = U_Common;
-  (*smap)[0x00BB] = U_Common;
-  for (char32 c = 0x00BC; c <= 0x00BE; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x00BF] = U_Common;
-  (*smap)[0x00D7] = U_Common;
-  (*smap)[0x00F7] = U_Common;
-  for (char32 c = 0x02B9; c <= 0x02C1; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x02C2; c <= 0x02C5; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x02C6; c <= 0x02D1; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x02D2; c <= 0x02DF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x02E5; c <= 0x02E9; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x02EC] = U_Common;
-  (*smap)[0x02ED] = U_Common;
-  (*smap)[0x02EE] = U_Common;
-  for (char32 c = 0x02EF; c <= 0x02FF; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x0374] = U_Common;
-  (*smap)[0x037E] = U_Common;
-  (*smap)[0x0385] = U_Common;
-  (*smap)[0x0387] = U_Common;
-  (*smap)[0x0589] = U_Common;
-  (*smap)[0x0605] = U_Common;
-  (*smap)[0x060C] = U_Common;
-  (*smap)[0x061B] = U_Common;
-  (*smap)[0x061C] = U_Common;
-  (*smap)[0x061F] = U_Common;
-  (*smap)[0x0640] = U_Common;
-  (*smap)[0x06DD] = U_Common;
-  (*smap)[0x08E2] = U_Common;
-  for (char32 c = 0x0964; c <= 0x0965; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x0E3F] = U_Common;
-  for (char32 c = 0x0FD5; c <= 0x0FD8; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x10FB] = U_Common;
-  for (char32 c = 0x16EB; c <= 0x16ED; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1735; c <= 0x1736; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1802; c <= 0x1803; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1805] = U_Common;
-  (*smap)[0x1CD3] = U_Common;
-  (*smap)[0x1CE1] = U_Common;
-  for (char32 c = 0x1CE9; c <= 0x1CEC; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1CEE; c <= 0x1CF1; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1CF2; c <= 0x1CF3; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1CF5; c <= 0x1CF6; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2000; c <= 0x200A; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x200B] = U_Common;
-  for (char32 c = 0x200E; c <= 0x200F; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2010; c <= 0x2015; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2016; c <= 0x2017; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2018] = U_Common;
-  (*smap)[0x2019] = U_Common;
-  (*smap)[0x201A] = U_Common;
-  for (char32 c = 0x201B; c <= 0x201C; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x201D] = U_Common;
-  (*smap)[0x201E] = U_Common;
-  (*smap)[0x201F] = U_Common;
-  for (char32 c = 0x2020; c <= 0x2027; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2028] = U_Common;
-  (*smap)[0x2029] = U_Common;
-  for (char32 c = 0x202A; c <= 0x202E; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x202F] = U_Common;
-  for (char32 c = 0x2030; c <= 0x2038; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2039] = U_Common;
-  (*smap)[0x203A] = U_Common;
-  for (char32 c = 0x203B; c <= 0x203E; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x203F; c <= 0x2040; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2041; c <= 0x2043; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2044] = U_Common;
-  (*smap)[0x2045] = U_Common;
-  (*smap)[0x2046] = U_Common;
-  for (char32 c = 0x2047; c <= 0x2051; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2052] = U_Common;
-  (*smap)[0x2053] = U_Common;
-  (*smap)[0x2054] = U_Common;
-  for (char32 c = 0x2055; c <= 0x205E; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x205F] = U_Common;
-  for (char32 c = 0x2060; c <= 0x2064; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2066; c <= 0x206F; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2070] = U_Common;
-  for (char32 c = 0x2074; c <= 0x2079; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x207A; c <= 0x207C; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x207D] = U_Common;
-  (*smap)[0x207E] = U_Common;
-  for (char32 c = 0x2080; c <= 0x2089; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x208A; c <= 0x208C; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x208D] = U_Common;
-  (*smap)[0x208E] = U_Common;
-  for (char32 c = 0x20A0; c <= 0x20BE; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2100; c <= 0x2101; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2102] = U_Common;
-  for (char32 c = 0x2103; c <= 0x2106; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2107] = U_Common;
-  for (char32 c = 0x2108; c <= 0x2109; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x210A; c <= 0x2113; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2114] = U_Common;
-  (*smap)[0x2115] = U_Common;
-  for (char32 c = 0x2116; c <= 0x2117; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2118] = U_Common;
-  for (char32 c = 0x2119; c <= 0x211D; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x211E; c <= 0x2123; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2124] = U_Common;
-  (*smap)[0x2125] = U_Common;
-  (*smap)[0x2127] = U_Common;
-  (*smap)[0x2128] = U_Common;
-  (*smap)[0x2129] = U_Common;
-  for (char32 c = 0x212C; c <= 0x212D; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x212E] = U_Common;
-  for (char32 c = 0x212F; c <= 0x2131; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2133; c <= 0x2134; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2135; c <= 0x2138; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2139] = U_Common;
-  for (char32 c = 0x213A; c <= 0x213B; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x213C; c <= 0x213F; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2140; c <= 0x2144; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2145; c <= 0x2149; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x214A] = U_Common;
-  (*smap)[0x214B] = U_Common;
-  for (char32 c = 0x214C; c <= 0x214D; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x214F] = U_Common;
-  for (char32 c = 0x2150; c <= 0x215F; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2189] = U_Common;
-  for (char32 c = 0x218A; c <= 0x218B; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2190; c <= 0x2194; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2195; c <= 0x2199; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x219A; c <= 0x219B; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x219C; c <= 0x219F; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x21A0] = U_Common;
-  for (char32 c = 0x21A1; c <= 0x21A2; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x21A3] = U_Common;
-  for (char32 c = 0x21A4; c <= 0x21A5; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x21A6] = U_Common;
-  for (char32 c = 0x21A7; c <= 0x21AD; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x21AE] = U_Common;
-  for (char32 c = 0x21AF; c <= 0x21CD; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x21CE; c <= 0x21CF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x21D0; c <= 0x21D1; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x21D2] = U_Common;
-  (*smap)[0x21D3] = U_Common;
-  (*smap)[0x21D4] = U_Common;
-  for (char32 c = 0x21D5; c <= 0x21F3; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x21F4; c <= 0x22FF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2300; c <= 0x2307; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2308] = U_Common;
-  (*smap)[0x2309] = U_Common;
-  (*smap)[0x230A] = U_Common;
-  (*smap)[0x230B] = U_Common;
-  for (char32 c = 0x230C; c <= 0x231F; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2320; c <= 0x2321; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2322; c <= 0x2328; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2329] = U_Common;
-  (*smap)[0x232A] = U_Common;
-  for (char32 c = 0x232B; c <= 0x237B; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x237C] = U_Common;
-  for (char32 c = 0x237D; c <= 0x239A; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x239B; c <= 0x23B3; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x23B4; c <= 0x23DB; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x23DC; c <= 0x23E1; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x23E2; c <= 0x23FE; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2400; c <= 0x2426; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2440; c <= 0x244A; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2460; c <= 0x249B; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x249C; c <= 0x24E9; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x24EA; c <= 0x24FF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2500; c <= 0x25B6; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x25B7] = U_Common;
-  for (char32 c = 0x25B8; c <= 0x25C0; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x25C1] = U_Common;
-  for (char32 c = 0x25C2; c <= 0x25F7; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x25F8; c <= 0x25FF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2600; c <= 0x266E; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x266F] = U_Common;
-  for (char32 c = 0x2670; c <= 0x2767; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2768] = U_Common;
-  (*smap)[0x2769] = U_Common;
-  (*smap)[0x276A] = U_Common;
-  (*smap)[0x276B] = U_Common;
-  (*smap)[0x276C] = U_Common;
-  (*smap)[0x276D] = U_Common;
-  (*smap)[0x276E] = U_Common;
-  (*smap)[0x276F] = U_Common;
-  (*smap)[0x2770] = U_Common;
-  (*smap)[0x2771] = U_Common;
-  (*smap)[0x2772] = U_Common;
-  (*smap)[0x2773] = U_Common;
-  (*smap)[0x2774] = U_Common;
-  (*smap)[0x2775] = U_Common;
-  for (char32 c = 0x2776; c <= 0x2793; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2794; c <= 0x27BF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x27C0; c <= 0x27C4; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x27C5] = U_Common;
-  (*smap)[0x27C6] = U_Common;
-  for (char32 c = 0x27C7; c <= 0x27E5; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x27E6] = U_Common;
-  (*smap)[0x27E7] = U_Common;
-  (*smap)[0x27E8] = U_Common;
-  (*smap)[0x27E9] = U_Common;
-  (*smap)[0x27EA] = U_Common;
-  (*smap)[0x27EB] = U_Common;
-  (*smap)[0x27EC] = U_Common;
-  (*smap)[0x27ED] = U_Common;
-  (*smap)[0x27EE] = U_Common;
-  (*smap)[0x27EF] = U_Common;
-  for (char32 c = 0x27F0; c <= 0x27FF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2900; c <= 0x2982; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2983] = U_Common;
-  (*smap)[0x2984] = U_Common;
-  (*smap)[0x2985] = U_Common;
-  (*smap)[0x2986] = U_Common;
-  (*smap)[0x2987] = U_Common;
-  (*smap)[0x2988] = U_Common;
-  (*smap)[0x2989] = U_Common;
-  (*smap)[0x298A] = U_Common;
-  (*smap)[0x298B] = U_Common;
-  (*smap)[0x298C] = U_Common;
-  (*smap)[0x298D] = U_Common;
-  (*smap)[0x298E] = U_Common;
-  (*smap)[0x298F] = U_Common;
-  (*smap)[0x2990] = U_Common;
-  (*smap)[0x2991] = U_Common;
-  (*smap)[0x2992] = U_Common;
-  (*smap)[0x2993] = U_Common;
-  (*smap)[0x2994] = U_Common;
-  (*smap)[0x2995] = U_Common;
-  (*smap)[0x2996] = U_Common;
-  (*smap)[0x2997] = U_Common;
-  (*smap)[0x2998] = U_Common;
-  for (char32 c = 0x2999; c <= 0x29D7; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x29D8] = U_Common;
-  (*smap)[0x29D9] = U_Common;
-  (*smap)[0x29DA] = U_Common;
-  (*smap)[0x29DB] = U_Common;
-  for (char32 c = 0x29DC; c <= 0x29FB; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x29FC] = U_Common;
-  (*smap)[0x29FD] = U_Common;
-  for (char32 c = 0x29FE; c <= 0x2AFF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2B00; c <= 0x2B2F; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2B30; c <= 0x2B44; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2B45; c <= 0x2B46; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2B47; c <= 0x2B4C; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2B4D; c <= 0x2B73; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2B76; c <= 0x2B95; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2B98; c <= 0x2BB9; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2BBD; c <= 0x2BC8; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2BCA; c <= 0x2BD1; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2BEC; c <= 0x2BEF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2E00; c <= 0x2E01; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2E02] = U_Common;
-  (*smap)[0x2E03] = U_Common;
-  (*smap)[0x2E04] = U_Common;
-  (*smap)[0x2E05] = U_Common;
-  for (char32 c = 0x2E06; c <= 0x2E08; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2E09] = U_Common;
-  (*smap)[0x2E0A] = U_Common;
-  (*smap)[0x2E0B] = U_Common;
-  (*smap)[0x2E0C] = U_Common;
-  (*smap)[0x2E0D] = U_Common;
-  for (char32 c = 0x2E0E; c <= 0x2E16; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2E17] = U_Common;
-  for (char32 c = 0x2E18; c <= 0x2E19; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2E1A] = U_Common;
-  (*smap)[0x2E1B] = U_Common;
-  (*smap)[0x2E1C] = U_Common;
-  (*smap)[0x2E1D] = U_Common;
-  for (char32 c = 0x2E1E; c <= 0x2E1F; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2E20] = U_Common;
-  (*smap)[0x2E21] = U_Common;
-  (*smap)[0x2E22] = U_Common;
-  (*smap)[0x2E23] = U_Common;
-  (*smap)[0x2E24] = U_Common;
-  (*smap)[0x2E25] = U_Common;
-  (*smap)[0x2E26] = U_Common;
-  (*smap)[0x2E27] = U_Common;
-  (*smap)[0x2E28] = U_Common;
-  (*smap)[0x2E29] = U_Common;
-  for (char32 c = 0x2E2A; c <= 0x2E2E; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2E2F] = U_Common;
-  for (char32 c = 0x2E30; c <= 0x2E39; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2E3A; c <= 0x2E3B; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2E3C; c <= 0x2E3F; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x2E40] = U_Common;
-  (*smap)[0x2E41] = U_Common;
-  (*smap)[0x2E42] = U_Common;
-  for (char32 c = 0x2E43; c <= 0x2E44; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x2FF0; c <= 0x2FFB; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x3000] = U_Common;
-  for (char32 c = 0x3001; c <= 0x3003; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x3004] = U_Common;
-  (*smap)[0x3006] = U_Common;
-  (*smap)[0x3008] = U_Common;
-  (*smap)[0x3009] = U_Common;
-  (*smap)[0x300A] = U_Common;
-  (*smap)[0x300B] = U_Common;
-  (*smap)[0x300C] = U_Common;
-  (*smap)[0x300D] = U_Common;
-  (*smap)[0x300E] = U_Common;
-  (*smap)[0x300F] = U_Common;
-  (*smap)[0x3010] = U_Common;
-  (*smap)[0x3011] = U_Common;
-  for (char32 c = 0x3012; c <= 0x3013; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x3014] = U_Common;
-  (*smap)[0x3015] = U_Common;
-  (*smap)[0x3016] = U_Common;
-  (*smap)[0x3017] = U_Common;
-  (*smap)[0x3018] = U_Common;
-  (*smap)[0x3019] = U_Common;
-  (*smap)[0x301A] = U_Common;
-  (*smap)[0x301B] = U_Common;
-  (*smap)[0x301C] = U_Common;
-  (*smap)[0x301D] = U_Common;
-  for (char32 c = 0x301E; c <= 0x301F; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x3020] = U_Common;
-  (*smap)[0x3030] = U_Common;
-  for (char32 c = 0x3031; c <= 0x3035; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x3036; c <= 0x3037; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x303C] = U_Common;
-  (*smap)[0x303D] = U_Common;
-  for (char32 c = 0x303E; c <= 0x303F; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x309B; c <= 0x309C; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x30A0] = U_Common;
-  (*smap)[0x30FB] = U_Common;
-  (*smap)[0x30FC] = U_Common;
-  for (char32 c = 0x3190; c <= 0x3191; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x3192; c <= 0x3195; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x3196; c <= 0x319F; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x31C0; c <= 0x31E3; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x3220; c <= 0x3229; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x322A; c <= 0x3247; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x3248; c <= 0x324F; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x3250] = U_Common;
-  for (char32 c = 0x3251; c <= 0x325F; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x327F] = U_Common;
-  for (char32 c = 0x3280; c <= 0x3289; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x328A; c <= 0x32B0; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x32B1; c <= 0x32BF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x32C0; c <= 0x32CF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x3358; c <= 0x33FF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x4DC0; c <= 0x4DFF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xA700; c <= 0xA716; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xA717; c <= 0xA71F; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xA720; c <= 0xA721; ++c) (*smap)[c] = U_Common;
-  (*smap)[0xA788] = U_Common;
-  for (char32 c = 0xA789; c <= 0xA78A; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xA830; c <= 0xA835; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xA836; c <= 0xA837; ++c) (*smap)[c] = U_Common;
-  (*smap)[0xA838] = U_Common;
-  (*smap)[0xA839] = U_Common;
-  (*smap)[0xA92E] = U_Common;
-  (*smap)[0xA9CF] = U_Common;
-  (*smap)[0xAB5B] = U_Common;
-  (*smap)[0xFD3E] = U_Common;
-  (*smap)[0xFD3F] = U_Common;
-  for (char32 c = 0xFE10; c <= 0xFE16; ++c) (*smap)[c] = U_Common;
-  (*smap)[0xFE17] = U_Common;
-  (*smap)[0xFE18] = U_Common;
-  (*smap)[0xFE19] = U_Common;
-  (*smap)[0xFE30] = U_Common;
-  for (char32 c = 0xFE31; c <= 0xFE32; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xFE33; c <= 0xFE34; ++c) (*smap)[c] = U_Common;
-  (*smap)[0xFE35] = U_Common;
-  (*smap)[0xFE36] = U_Common;
-  (*smap)[0xFE37] = U_Common;
-  (*smap)[0xFE38] = U_Common;
-  (*smap)[0xFE39] = U_Common;
-  (*smap)[0xFE3A] = U_Common;
-  (*smap)[0xFE3B] = U_Common;
-  (*smap)[0xFE3C] = U_Common;
-  (*smap)[0xFE3D] = U_Common;
-  (*smap)[0xFE3E] = U_Common;
-  (*smap)[0xFE3F] = U_Common;
-  (*smap)[0xFE40] = U_Common;
-  (*smap)[0xFE41] = U_Common;
-  (*smap)[0xFE42] = U_Common;
-  (*smap)[0xFE43] = U_Common;
-  (*smap)[0xFE44] = U_Common;
-  for (char32 c = 0xFE45; c <= 0xFE46; ++c) (*smap)[c] = U_Common;
-  (*smap)[0xFE47] = U_Common;
-  (*smap)[0xFE48] = U_Common;
-  for (char32 c = 0xFE49; c <= 0xFE4C; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xFE4D; c <= 0xFE4F; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xFE50; c <= 0xFE52; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xFE54; c <= 0xFE57; ++c) (*smap)[c] = U_Common;
-  (*smap)[0xFE58] = U_Common;
-  (*smap)[0xFE59] = U_Common;
-  (*smap)[0xFE5A] = U_Common;
-  (*smap)[0xFE5B] = U_Common;
-  (*smap)[0xFE5C] = U_Common;
-  (*smap)[0xFE5D] = U_Common;
-  (*smap)[0xFE5E] = U_Common;
-  for (char32 c = 0xFE5F; c <= 0xFE61; ++c) (*smap)[c] = U_Common;
-  (*smap)[0xFE62] = U_Common;
-  (*smap)[0xFE63] = U_Common;
-  for (char32 c = 0xFE64; c <= 0xFE66; ++c) (*smap)[c] = U_Common;
-  (*smap)[0xFE68] = U_Common;
-  (*smap)[0xFE69] = U_Common;
-  for (char32 c = 0xFE6A; c <= 0xFE6B; ++c) (*smap)[c] = U_Common;
-  (*smap)[0xFEFF] = U_Common;
-  for (char32 c = 0xFF01; c <= 0xFF03; ++c) (*smap)[c] = U_Common;
-  (*smap)[0xFF04] = U_Common;
-  for (char32 c = 0xFF05; c <= 0xFF07; ++c) (*smap)[c] = U_Common;
-  (*smap)[0xFF08] = U_Common;
-  (*smap)[0xFF09] = U_Common;
-  (*smap)[0xFF0A] = U_Common;
-  (*smap)[0xFF0B] = U_Common;
-  (*smap)[0xFF0C] = U_Common;
-  (*smap)[0xFF0D] = U_Common;
-  for (char32 c = 0xFF0E; c <= 0xFF0F; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xFF10; c <= 0xFF19; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xFF1A; c <= 0xFF1B; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xFF1C; c <= 0xFF1E; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xFF1F; c <= 0xFF20; ++c) (*smap)[c] = U_Common;
-  (*smap)[0xFF3B] = U_Common;
-  (*smap)[0xFF3C] = U_Common;
-  (*smap)[0xFF3D] = U_Common;
-  (*smap)[0xFF3E] = U_Common;
-  (*smap)[0xFF3F] = U_Common;
-  (*smap)[0xFF40] = U_Common;
-  (*smap)[0xFF5B] = U_Common;
-  (*smap)[0xFF5C] = U_Common;
-  (*smap)[0xFF5D] = U_Common;
-  (*smap)[0xFF5E] = U_Common;
-  (*smap)[0xFF5F] = U_Common;
-  (*smap)[0xFF60] = U_Common;
-  (*smap)[0xFF61] = U_Common;
-  (*smap)[0xFF62] = U_Common;
-  (*smap)[0xFF63] = U_Common;
-  for (char32 c = 0xFF64; c <= 0xFF65; ++c) (*smap)[c] = U_Common;
-  (*smap)[0xFF70] = U_Common;
-  for (char32 c = 0xFF9E; c <= 0xFF9F; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xFFE0; c <= 0xFFE1; ++c) (*smap)[c] = U_Common;
-  (*smap)[0xFFE2] = U_Common;
-  (*smap)[0xFFE3] = U_Common;
-  (*smap)[0xFFE4] = U_Common;
-  for (char32 c = 0xFFE5; c <= 0xFFE6; ++c) (*smap)[c] = U_Common;
-  (*smap)[0xFFE8] = U_Common;
-  for (char32 c = 0xFFE9; c <= 0xFFEC; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xFFED; c <= 0xFFEE; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xFFF9; c <= 0xFFFB; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0xFFFC; c <= 0xFFFD; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x10100; c <= 0x10102; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x10107; c <= 0x10133; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x10137; c <= 0x1013F; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x10190; c <= 0x1019B; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x101D0; c <= 0x101FC; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x102E1; c <= 0x102FB; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1BCA0; c <= 0x1BCA3; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D000; c <= 0x1D0F5; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D100; c <= 0x1D126; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D129; c <= 0x1D164; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D165; c <= 0x1D166; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D16A; c <= 0x1D16C; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D16D; c <= 0x1D172; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D173; c <= 0x1D17A; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D183; c <= 0x1D184; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D18C; c <= 0x1D1A9; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D1AE; c <= 0x1D1E8; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D300; c <= 0x1D356; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D360; c <= 0x1D371; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D400; c <= 0x1D454; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D456; c <= 0x1D49C; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D49E; c <= 0x1D49F; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1D4A2] = U_Common;
-  for (char32 c = 0x1D4A5; c <= 0x1D4A6; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D4A9; c <= 0x1D4AC; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D4AE; c <= 0x1D4B9; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1D4BB] = U_Common;
-  for (char32 c = 0x1D4BD; c <= 0x1D4C3; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D4C5; c <= 0x1D505; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D507; c <= 0x1D50A; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D50D; c <= 0x1D514; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D516; c <= 0x1D51C; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D51E; c <= 0x1D539; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D53B; c <= 0x1D53E; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D540; c <= 0x1D544; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1D546] = U_Common;
-  for (char32 c = 0x1D54A; c <= 0x1D550; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D552; c <= 0x1D6A5; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D6A8; c <= 0x1D6C0; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1D6C1] = U_Common;
-  for (char32 c = 0x1D6C2; c <= 0x1D6DA; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1D6DB] = U_Common;
-  for (char32 c = 0x1D6DC; c <= 0x1D6FA; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1D6FB] = U_Common;
-  for (char32 c = 0x1D6FC; c <= 0x1D714; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1D715] = U_Common;
-  for (char32 c = 0x1D716; c <= 0x1D734; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1D735] = U_Common;
-  for (char32 c = 0x1D736; c <= 0x1D74E; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1D74F] = U_Common;
-  for (char32 c = 0x1D750; c <= 0x1D76E; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1D76F] = U_Common;
-  for (char32 c = 0x1D770; c <= 0x1D788; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1D789] = U_Common;
-  for (char32 c = 0x1D78A; c <= 0x1D7A8; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1D7A9] = U_Common;
-  for (char32 c = 0x1D7AA; c <= 0x1D7C2; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1D7C3] = U_Common;
-  for (char32 c = 0x1D7C4; c <= 0x1D7CB; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1D7CE; c <= 0x1D7FF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F000; c <= 0x1F02B; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F030; c <= 0x1F093; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F0A0; c <= 0x1F0AE; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F0B1; c <= 0x1F0BF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F0C1; c <= 0x1F0CF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F0D1; c <= 0x1F0F5; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F100; c <= 0x1F10C; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F110; c <= 0x1F12E; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F130; c <= 0x1F16B; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F170; c <= 0x1F1AC; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F1E6; c <= 0x1F1FF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F201; c <= 0x1F202; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F210; c <= 0x1F23B; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F240; c <= 0x1F248; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F250; c <= 0x1F251; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F300; c <= 0x1F3FA; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F3FB; c <= 0x1F3FF; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F400; c <= 0x1F6D2; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F6E0; c <= 0x1F6EC; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F6F0; c <= 0x1F6F6; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F700; c <= 0x1F773; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F780; c <= 0x1F7D4; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F800; c <= 0x1F80B; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F810; c <= 0x1F847; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F850; c <= 0x1F859; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F860; c <= 0x1F887; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F890; c <= 0x1F8AD; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F910; c <= 0x1F91E; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F920; c <= 0x1F927; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1F930] = U_Common;
-  for (char32 c = 0x1F933; c <= 0x1F93E; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F940; c <= 0x1F94B; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F950; c <= 0x1F95E; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x1F980; c <= 0x1F991; ++c) (*smap)[c] = U_Common;
-  (*smap)[0x1F9C0] = U_Common;
-  (*smap)[0xE0001] = U_Common;
-  for (char32 c = 0xE0020; c <= 0xE007F; ++c) (*smap)[c] = U_Common;
-  for (char32 c = 0x0041; c <= 0x005A; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x0061; c <= 0x007A; ++c) (*smap)[c] = U_Latin;
-  (*smap)[0x00AA] = U_Latin;
-  (*smap)[0x00BA] = U_Latin;
-  for (char32 c = 0x00C0; c <= 0x00D6; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x00D8; c <= 0x00F6; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x00F8; c <= 0x01BA; ++c) (*smap)[c] = U_Latin;
-  (*smap)[0x01BB] = U_Latin;
-  for (char32 c = 0x01BC; c <= 0x01BF; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x01C0; c <= 0x01C3; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x01C4; c <= 0x0293; ++c) (*smap)[c] = U_Latin;
-  (*smap)[0x0294] = U_Latin;
-  for (char32 c = 0x0295; c <= 0x02AF; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x02B0; c <= 0x02B8; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x02E0; c <= 0x02E4; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x1D00; c <= 0x1D25; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x1D2C; c <= 0x1D5C; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x1D62; c <= 0x1D65; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x1D6B; c <= 0x1D77; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x1D79; c <= 0x1D9A; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x1D9B; c <= 0x1DBE; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x1E00; c <= 0x1EFF; ++c) (*smap)[c] = U_Latin;
-  (*smap)[0x2071] = U_Latin;
-  (*smap)[0x207F] = U_Latin;
-  for (char32 c = 0x2090; c <= 0x209C; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x212A; c <= 0x212B; ++c) (*smap)[c] = U_Latin;
-  (*smap)[0x2132] = U_Latin;
-  (*smap)[0x214E] = U_Latin;
-  for (char32 c = 0x2160; c <= 0x2182; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x2183; c <= 0x2184; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x2185; c <= 0x2188; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x2C60; c <= 0x2C7B; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x2C7C; c <= 0x2C7D; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x2C7E; c <= 0x2C7F; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0xA722; c <= 0xA76F; ++c) (*smap)[c] = U_Latin;
-  (*smap)[0xA770] = U_Latin;
-  for (char32 c = 0xA771; c <= 0xA787; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0xA78B; c <= 0xA78E; ++c) (*smap)[c] = U_Latin;
-  (*smap)[0xA78F] = U_Latin;
-  for (char32 c = 0xA790; c <= 0xA7AE; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0xA7B0; c <= 0xA7B7; ++c) (*smap)[c] = U_Latin;
-  (*smap)[0xA7F7] = U_Latin;
-  for (char32 c = 0xA7F8; c <= 0xA7F9; ++c) (*smap)[c] = U_Latin;
-  (*smap)[0xA7FA] = U_Latin;
-  for (char32 c = 0xA7FB; c <= 0xA7FF; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0xAB30; c <= 0xAB5A; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0xAB5C; c <= 0xAB5F; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0xAB60; c <= 0xAB64; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0xFB00; c <= 0xFB06; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0xFF21; c <= 0xFF3A; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0xFF41; c <= 0xFF5A; ++c) (*smap)[c] = U_Latin;
-  for (char32 c = 0x0370; c <= 0x0373; ++c) (*smap)[c] = U_Greek;
-  (*smap)[0x0375] = U_Greek;
-  for (char32 c = 0x0376; c <= 0x0377; ++c) (*smap)[c] = U_Greek;
-  (*smap)[0x037A] = U_Greek;
-  for (char32 c = 0x037B; c <= 0x037D; ++c) (*smap)[c] = U_Greek;
-  (*smap)[0x037F] = U_Greek;
-  (*smap)[0x0384] = U_Greek;
-  (*smap)[0x0386] = U_Greek;
-  for (char32 c = 0x0388; c <= 0x038A; ++c) (*smap)[c] = U_Greek;
-  (*smap)[0x038C] = U_Greek;
-  for (char32 c = 0x038E; c <= 0x03A1; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x03A3; c <= 0x03E1; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x03F0; c <= 0x03F5; ++c) (*smap)[c] = U_Greek;
-  (*smap)[0x03F6] = U_Greek;
-  for (char32 c = 0x03F7; c <= 0x03FF; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1D26; c <= 0x1D2A; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1D5D; c <= 0x1D61; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1D66; c <= 0x1D6A; ++c) (*smap)[c] = U_Greek;
-  (*smap)[0x1DBF] = U_Greek;
-  for (char32 c = 0x1F00; c <= 0x1F15; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1F18; c <= 0x1F1D; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1F20; c <= 0x1F45; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1F48; c <= 0x1F4D; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1F50; c <= 0x1F57; ++c) (*smap)[c] = U_Greek;
-  (*smap)[0x1F59] = U_Greek;
-  (*smap)[0x1F5B] = U_Greek;
-  (*smap)[0x1F5D] = U_Greek;
-  for (char32 c = 0x1F5F; c <= 0x1F7D; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1F80; c <= 0x1FB4; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1FB6; c <= 0x1FBC; ++c) (*smap)[c] = U_Greek;
-  (*smap)[0x1FBD] = U_Greek;
-  (*smap)[0x1FBE] = U_Greek;
-  for (char32 c = 0x1FBF; c <= 0x1FC1; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1FC2; c <= 0x1FC4; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1FC6; c <= 0x1FCC; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1FCD; c <= 0x1FCF; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1FD0; c <= 0x1FD3; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1FD6; c <= 0x1FDB; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1FDD; c <= 0x1FDF; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1FE0; c <= 0x1FEC; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1FED; c <= 0x1FEF; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1FF2; c <= 0x1FF4; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1FF6; c <= 0x1FFC; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1FFD; c <= 0x1FFE; ++c) (*smap)[c] = U_Greek;
-  (*smap)[0x2126] = U_Greek;
-  (*smap)[0xAB65] = U_Greek;
-  for (char32 c = 0x10140; c <= 0x10174; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x10175; c <= 0x10178; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x10179; c <= 0x10189; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1018A; c <= 0x1018B; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1018C; c <= 0x1018E; ++c) (*smap)[c] = U_Greek;
-  (*smap)[0x101A0] = U_Greek;
-  for (char32 c = 0x1D200; c <= 0x1D241; ++c) (*smap)[c] = U_Greek;
-  for (char32 c = 0x1D242; c <= 0x1D244; ++c) (*smap)[c] = U_Greek;
-  (*smap)[0x1D245] = U_Greek;
-  for (char32 c = 0x0400; c <= 0x0481; ++c) (*smap)[c] = U_Cyrillic;
-  (*smap)[0x0482] = U_Cyrillic;
-  for (char32 c = 0x0483; c <= 0x0484; ++c) (*smap)[c] = U_Cyrillic;
-  (*smap)[0x0487] = U_Cyrillic;
-  for (char32 c = 0x0488; c <= 0x0489; ++c) (*smap)[c] = U_Cyrillic;
-  for (char32 c = 0x048A; c <= 0x052F; ++c) (*smap)[c] = U_Cyrillic;
-  for (char32 c = 0x1C80; c <= 0x1C88; ++c) (*smap)[c] = U_Cyrillic;
-  (*smap)[0x1D2B] = U_Cyrillic;
-  (*smap)[0x1D78] = U_Cyrillic;
-  for (char32 c = 0x2DE0; c <= 0x2DFF; ++c) (*smap)[c] = U_Cyrillic;
-  for (char32 c = 0xA640; c <= 0xA66D; ++c) (*smap)[c] = U_Cyrillic;
-  (*smap)[0xA66E] = U_Cyrillic;
-  (*smap)[0xA66F] = U_Cyrillic;
-  for (char32 c = 0xA670; c <= 0xA672; ++c) (*smap)[c] = U_Cyrillic;
-  (*smap)[0xA673] = U_Cyrillic;
-  for (char32 c = 0xA674; c <= 0xA67D; ++c) (*smap)[c] = U_Cyrillic;
-  (*smap)[0xA67E] = U_Cyrillic;
-  (*smap)[0xA67F] = U_Cyrillic;
-  for (char32 c = 0xA680; c <= 0xA69B; ++c) (*smap)[c] = U_Cyrillic;
-  for (char32 c = 0xA69C; c <= 0xA69D; ++c) (*smap)[c] = U_Cyrillic;
-  for (char32 c = 0xA69E; c <= 0xA69F; ++c) (*smap)[c] = U_Cyrillic;
-  for (char32 c = 0xFE2E; c <= 0xFE2F; ++c) (*smap)[c] = U_Cyrillic;
-  for (char32 c = 0x0531; c <= 0x0556; ++c) (*smap)[c] = U_Armenian;
-  (*smap)[0x0559] = U_Armenian;
-  for (char32 c = 0x055A; c <= 0x055F; ++c) (*smap)[c] = U_Armenian;
-  for (char32 c = 0x0561; c <= 0x0587; ++c) (*smap)[c] = U_Armenian;
-  (*smap)[0x058A] = U_Armenian;
-  for (char32 c = 0x058D; c <= 0x058E; ++c) (*smap)[c] = U_Armenian;
-  (*smap)[0x058F] = U_Armenian;
-  for (char32 c = 0xFB13; c <= 0xFB17; ++c) (*smap)[c] = U_Armenian;
-  for (char32 c = 0x0591; c <= 0x05BD; ++c) (*smap)[c] = U_Hebrew;
-  (*smap)[0x05BE] = U_Hebrew;
-  (*smap)[0x05BF] = U_Hebrew;
-  (*smap)[0x05C0] = U_Hebrew;
-  for (char32 c = 0x05C1; c <= 0x05C2; ++c) (*smap)[c] = U_Hebrew;
-  (*smap)[0x05C3] = U_Hebrew;
-  for (char32 c = 0x05C4; c <= 0x05C5; ++c) (*smap)[c] = U_Hebrew;
-  (*smap)[0x05C6] = U_Hebrew;
-  (*smap)[0x05C7] = U_Hebrew;
-  for (char32 c = 0x05D0; c <= 0x05EA; ++c) (*smap)[c] = U_Hebrew;
-  for (char32 c = 0x05F0; c <= 0x05F2; ++c) (*smap)[c] = U_Hebrew;
-  for (char32 c = 0x05F3; c <= 0x05F4; ++c) (*smap)[c] = U_Hebrew;
-  (*smap)[0xFB1D] = U_Hebrew;
-  (*smap)[0xFB1E] = U_Hebrew;
-  for (char32 c = 0xFB1F; c <= 0xFB28; ++c) (*smap)[c] = U_Hebrew;
-  (*smap)[0xFB29] = U_Hebrew;
-  for (char32 c = 0xFB2A; c <= 0xFB36; ++c) (*smap)[c] = U_Hebrew;
-  for (char32 c = 0xFB38; c <= 0xFB3C; ++c) (*smap)[c] = U_Hebrew;
-  (*smap)[0xFB3E] = U_Hebrew;
-  for (char32 c = 0xFB40; c <= 0xFB41; ++c) (*smap)[c] = U_Hebrew;
-  for (char32 c = 0xFB43; c <= 0xFB44; ++c) (*smap)[c] = U_Hebrew;
-  for (char32 c = 0xFB46; c <= 0xFB4F; ++c) (*smap)[c] = U_Hebrew;
-  for (char32 c = 0x0600; c <= 0x0604; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x0606; c <= 0x0608; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x0609; c <= 0x060A; ++c) (*smap)[c] = U_Arabic;
-  (*smap)[0x060B] = U_Arabic;
-  (*smap)[0x060D] = U_Arabic;
-  for (char32 c = 0x060E; c <= 0x060F; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x0610; c <= 0x061A; ++c) (*smap)[c] = U_Arabic;
-  (*smap)[0x061E] = U_Arabic;
-  for (char32 c = 0x0620; c <= 0x063F; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x0641; c <= 0x064A; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x0656; c <= 0x065F; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x0660; c <= 0x0669; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x066A; c <= 0x066D; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x066E; c <= 0x066F; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x0671; c <= 0x06D3; ++c) (*smap)[c] = U_Arabic;
-  (*smap)[0x06D4] = U_Arabic;
-  (*smap)[0x06D5] = U_Arabic;
-  for (char32 c = 0x06D6; c <= 0x06DC; ++c) (*smap)[c] = U_Arabic;
-  (*smap)[0x06DE] = U_Arabic;
-  for (char32 c = 0x06DF; c <= 0x06E4; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x06E5; c <= 0x06E6; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x06E7; c <= 0x06E8; ++c) (*smap)[c] = U_Arabic;
-  (*smap)[0x06E9] = U_Arabic;
-  for (char32 c = 0x06EA; c <= 0x06ED; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x06EE; c <= 0x06EF; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x06F0; c <= 0x06F9; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x06FA; c <= 0x06FC; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x06FD; c <= 0x06FE; ++c) (*smap)[c] = U_Arabic;
-  (*smap)[0x06FF] = U_Arabic;
-  for (char32 c = 0x0750; c <= 0x077F; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x08A0; c <= 0x08B4; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x08B6; c <= 0x08BD; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x08D4; c <= 0x08E1; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x08E3; c <= 0x08FF; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0xFB50; c <= 0xFBB1; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0xFBB2; c <= 0xFBC1; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0xFBD3; c <= 0xFD3D; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0xFD50; c <= 0xFD8F; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0xFD92; c <= 0xFDC7; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0xFDF0; c <= 0xFDFB; ++c) (*smap)[c] = U_Arabic;
-  (*smap)[0xFDFC] = U_Arabic;
-  (*smap)[0xFDFD] = U_Arabic;
-  for (char32 c = 0xFE70; c <= 0xFE74; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0xFE76; c <= 0xFEFC; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x10E60; c <= 0x10E7E; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x1EE00; c <= 0x1EE03; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x1EE05; c <= 0x1EE1F; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x1EE21; c <= 0x1EE22; ++c) (*smap)[c] = U_Arabic;
-  (*smap)[0x1EE24] = U_Arabic;
-  (*smap)[0x1EE27] = U_Arabic;
-  for (char32 c = 0x1EE29; c <= 0x1EE32; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x1EE34; c <= 0x1EE37; ++c) (*smap)[c] = U_Arabic;
-  (*smap)[0x1EE39] = U_Arabic;
-  (*smap)[0x1EE3B] = U_Arabic;
-  (*smap)[0x1EE42] = U_Arabic;
-  (*smap)[0x1EE47] = U_Arabic;
-  (*smap)[0x1EE49] = U_Arabic;
-  (*smap)[0x1EE4B] = U_Arabic;
-  for (char32 c = 0x1EE4D; c <= 0x1EE4F; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x1EE51; c <= 0x1EE52; ++c) (*smap)[c] = U_Arabic;
-  (*smap)[0x1EE54] = U_Arabic;
-  (*smap)[0x1EE57] = U_Arabic;
-  (*smap)[0x1EE59] = U_Arabic;
-  (*smap)[0x1EE5B] = U_Arabic;
-  (*smap)[0x1EE5D] = U_Arabic;
-  (*smap)[0x1EE5F] = U_Arabic;
-  for (char32 c = 0x1EE61; c <= 0x1EE62; ++c) (*smap)[c] = U_Arabic;
-  (*smap)[0x1EE64] = U_Arabic;
-  for (char32 c = 0x1EE67; c <= 0x1EE6A; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x1EE6C; c <= 0x1EE72; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x1EE74; c <= 0x1EE77; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x1EE79; c <= 0x1EE7C; ++c) (*smap)[c] = U_Arabic;
-  (*smap)[0x1EE7E] = U_Arabic;
-  for (char32 c = 0x1EE80; c <= 0x1EE89; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x1EE8B; c <= 0x1EE9B; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x1EEA1; c <= 0x1EEA3; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x1EEA5; c <= 0x1EEA9; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x1EEAB; c <= 0x1EEBB; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x1EEF0; c <= 0x1EEF1; ++c) (*smap)[c] = U_Arabic;
-  for (char32 c = 0x0700; c <= 0x070D; ++c) (*smap)[c] = U_Syriac;
-  (*smap)[0x070F] = U_Syriac;
-  (*smap)[0x0710] = U_Syriac;
-  (*smap)[0x0711] = U_Syriac;
-  for (char32 c = 0x0712; c <= 0x072F; ++c) (*smap)[c] = U_Syriac;
-  for (char32 c = 0x0730; c <= 0x074A; ++c) (*smap)[c] = U_Syriac;
-  for (char32 c = 0x074D; c <= 0x074F; ++c) (*smap)[c] = U_Syriac;
-  for (char32 c = 0x0780; c <= 0x07A5; ++c) (*smap)[c] = U_Thaana;
-  for (char32 c = 0x07A6; c <= 0x07B0; ++c) (*smap)[c] = U_Thaana;
-  (*smap)[0x07B1] = U_Thaana;
-  for (char32 c = 0x0900; c <= 0x0902; ++c) (*smap)[c] = U_Devanagari;
-  (*smap)[0x0903] = U_Devanagari;
-  for (char32 c = 0x0904; c <= 0x0939; ++c) (*smap)[c] = U_Devanagari;
-  (*smap)[0x093A] = U_Devanagari;
-  (*smap)[0x093B] = U_Devanagari;
-  (*smap)[0x093C] = U_Devanagari;
-  (*smap)[0x093D] = U_Devanagari;
-  for (char32 c = 0x093E; c <= 0x0940; ++c) (*smap)[c] = U_Devanagari;
-  for (char32 c = 0x0941; c <= 0x0948; ++c) (*smap)[c] = U_Devanagari;
-  for (char32 c = 0x0949; c <= 0x094C; ++c) (*smap)[c] = U_Devanagari;
-  (*smap)[0x094D] = U_Devanagari;
-  for (char32 c = 0x094E; c <= 0x094F; ++c) (*smap)[c] = U_Devanagari;
-  (*smap)[0x0950] = U_Devanagari;
-  for (char32 c = 0x0953; c <= 0x0957; ++c) (*smap)[c] = U_Devanagari;
-  for (char32 c = 0x0958; c <= 0x0961; ++c) (*smap)[c] = U_Devanagari;
-  for (char32 c = 0x0962; c <= 0x0963; ++c) (*smap)[c] = U_Devanagari;
-  for (char32 c = 0x0966; c <= 0x096F; ++c) (*smap)[c] = U_Devanagari;
-  (*smap)[0x0970] = U_Devanagari;
-  (*smap)[0x0971] = U_Devanagari;
-  for (char32 c = 0x0972; c <= 0x097F; ++c) (*smap)[c] = U_Devanagari;
-  for (char32 c = 0xA8E0; c <= 0xA8F1; ++c) (*smap)[c] = U_Devanagari;
-  for (char32 c = 0xA8F2; c <= 0xA8F7; ++c) (*smap)[c] = U_Devanagari;
-  for (char32 c = 0xA8F8; c <= 0xA8FA; ++c) (*smap)[c] = U_Devanagari;
-  (*smap)[0xA8FB] = U_Devanagari;
-  (*smap)[0xA8FC] = U_Devanagari;
-  (*smap)[0xA8FD] = U_Devanagari;
-  (*smap)[0x0980] = U_Bengali;
-  (*smap)[0x0981] = U_Bengali;
-  for (char32 c = 0x0982; c <= 0x0983; ++c) (*smap)[c] = U_Bengali;
-  for (char32 c = 0x0985; c <= 0x098C; ++c) (*smap)[c] = U_Bengali;
-  for (char32 c = 0x098F; c <= 0x0990; ++c) (*smap)[c] = U_Bengali;
-  for (char32 c = 0x0993; c <= 0x09A8; ++c) (*smap)[c] = U_Bengali;
-  for (char32 c = 0x09AA; c <= 0x09B0; ++c) (*smap)[c] = U_Bengali;
-  (*smap)[0x09B2] = U_Bengali;
-  for (char32 c = 0x09B6; c <= 0x09B9; ++c) (*smap)[c] = U_Bengali;
-  (*smap)[0x09BC] = U_Bengali;
-  (*smap)[0x09BD] = U_Bengali;
-  for (char32 c = 0x09BE; c <= 0x09C0; ++c) (*smap)[c] = U_Bengali;
-  for (char32 c = 0x09C1; c <= 0x09C4; ++c) (*smap)[c] = U_Bengali;
-  for (char32 c = 0x09C7; c <= 0x09C8; ++c) (*smap)[c] = U_Bengali;
-  for (char32 c = 0x09CB; c <= 0x09CC; ++c) (*smap)[c] = U_Bengali;
-  (*smap)[0x09CD] = U_Bengali;
-  (*smap)[0x09CE] = U_Bengali;
-  (*smap)[0x09D7] = U_Bengali;
-  for (char32 c = 0x09DC; c <= 0x09DD; ++c) (*smap)[c] = U_Bengali;
-  for (char32 c = 0x09DF; c <= 0x09E1; ++c) (*smap)[c] = U_Bengali;
-  for (char32 c = 0x09E2; c <= 0x09E3; ++c) (*smap)[c] = U_Bengali;
-  for (char32 c = 0x09E6; c <= 0x09EF; ++c) (*smap)[c] = U_Bengali;
-  for (char32 c = 0x09F0; c <= 0x09F1; ++c) (*smap)[c] = U_Bengali;
-  for (char32 c = 0x09F2; c <= 0x09F3; ++c) (*smap)[c] = U_Bengali;
-  for (char32 c = 0x09F4; c <= 0x09F9; ++c) (*smap)[c] = U_Bengali;
-  (*smap)[0x09FA] = U_Bengali;
-  (*smap)[0x09FB] = U_Bengali;
-  for (char32 c = 0x0A01; c <= 0x0A02; ++c) (*smap)[c] = U_Gurmukhi;
-  (*smap)[0x0A03] = U_Gurmukhi;
-  for (char32 c = 0x0A05; c <= 0x0A0A; ++c) (*smap)[c] = U_Gurmukhi;
-  for (char32 c = 0x0A0F; c <= 0x0A10; ++c) (*smap)[c] = U_Gurmukhi;
-  for (char32 c = 0x0A13; c <= 0x0A28; ++c) (*smap)[c] = U_Gurmukhi;
-  for (char32 c = 0x0A2A; c <= 0x0A30; ++c) (*smap)[c] = U_Gurmukhi;
-  for (char32 c = 0x0A32; c <= 0x0A33; ++c) (*smap)[c] = U_Gurmukhi;
-  for (char32 c = 0x0A35; c <= 0x0A36; ++c) (*smap)[c] = U_Gurmukhi;
-  for (char32 c = 0x0A38; c <= 0x0A39; ++c) (*smap)[c] = U_Gurmukhi;
-  (*smap)[0x0A3C] = U_Gurmukhi;
-  for (char32 c = 0x0A3E; c <= 0x0A40; ++c) (*smap)[c] = U_Gurmukhi;
-  for (char32 c = 0x0A41; c <= 0x0A42; ++c) (*smap)[c] = U_Gurmukhi;
-  for (char32 c = 0x0A47; c <= 0x0A48; ++c) (*smap)[c] = U_Gurmukhi;
-  for (char32 c = 0x0A4B; c <= 0x0A4D; ++c) (*smap)[c] = U_Gurmukhi;
-  (*smap)[0x0A51] = U_Gurmukhi;
-  for (char32 c = 0x0A59; c <= 0x0A5C; ++c) (*smap)[c] = U_Gurmukhi;
-  (*smap)[0x0A5E] = U_Gurmukhi;
-  for (char32 c = 0x0A66; c <= 0x0A6F; ++c) (*smap)[c] = U_Gurmukhi;
-  for (char32 c = 0x0A70; c <= 0x0A71; ++c) (*smap)[c] = U_Gurmukhi;
-  for (char32 c = 0x0A72; c <= 0x0A74; ++c) (*smap)[c] = U_Gurmukhi;
-  (*smap)[0x0A75] = U_Gurmukhi;
-  for (char32 c = 0x0A81; c <= 0x0A82; ++c) (*smap)[c] = U_Gujarati;
-  (*smap)[0x0A83] = U_Gujarati;
-  for (char32 c = 0x0A85; c <= 0x0A8D; ++c) (*smap)[c] = U_Gujarati;
-  for (char32 c = 0x0A8F; c <= 0x0A91; ++c) (*smap)[c] = U_Gujarati;
-  for (char32 c = 0x0A93; c <= 0x0AA8; ++c) (*smap)[c] = U_Gujarati;
-  for (char32 c = 0x0AAA; c <= 0x0AB0; ++c) (*smap)[c] = U_Gujarati;
-  for (char32 c = 0x0AB2; c <= 0x0AB3; ++c) (*smap)[c] = U_Gujarati;
-  for (char32 c = 0x0AB5; c <= 0x0AB9; ++c) (*smap)[c] = U_Gujarati;
-  (*smap)[0x0ABC] = U_Gujarati;
-  (*smap)[0x0ABD] = U_Gujarati;
-  for (char32 c = 0x0ABE; c <= 0x0AC0; ++c) (*smap)[c] = U_Gujarati;
-  for (char32 c = 0x0AC1; c <= 0x0AC5; ++c) (*smap)[c] = U_Gujarati;
-  for (char32 c = 0x0AC7; c <= 0x0AC8; ++c) (*smap)[c] = U_Gujarati;
-  (*smap)[0x0AC9] = U_Gujarati;
-  for (char32 c = 0x0ACB; c <= 0x0ACC; ++c) (*smap)[c] = U_Gujarati;
-  (*smap)[0x0ACD] = U_Gujarati;
-  (*smap)[0x0AD0] = U_Gujarati;
-  for (char32 c = 0x0AE0; c <= 0x0AE1; ++c) (*smap)[c] = U_Gujarati;
-  for (char32 c = 0x0AE2; c <= 0x0AE3; ++c) (*smap)[c] = U_Gujarati;
-  for (char32 c = 0x0AE6; c <= 0x0AEF; ++c) (*smap)[c] = U_Gujarati;
-  (*smap)[0x0AF0] = U_Gujarati;
-  (*smap)[0x0AF1] = U_Gujarati;
-  (*smap)[0x0AF9] = U_Gujarati;
-  (*smap)[0x0B01] = U_Oriya;
-  for (char32 c = 0x0B02; c <= 0x0B03; ++c) (*smap)[c] = U_Oriya;
-  for (char32 c = 0x0B05; c <= 0x0B0C; ++c) (*smap)[c] = U_Oriya;
-  for (char32 c = 0x0B0F; c <= 0x0B10; ++c) (*smap)[c] = U_Oriya;
-  for (char32 c = 0x0B13; c <= 0x0B28; ++c) (*smap)[c] = U_Oriya;
-  for (char32 c = 0x0B2A; c <= 0x0B30; ++c) (*smap)[c] = U_Oriya;
-  for (char32 c = 0x0B32; c <= 0x0B33; ++c) (*smap)[c] = U_Oriya;
-  for (char32 c = 0x0B35; c <= 0x0B39; ++c) (*smap)[c] = U_Oriya;
-  (*smap)[0x0B3C] = U_Oriya;
-  (*smap)[0x0B3D] = U_Oriya;
-  (*smap)[0x0B3E] = U_Oriya;
-  (*smap)[0x0B3F] = U_Oriya;
-  (*smap)[0x0B40] = U_Oriya;
-  for (char32 c = 0x0B41; c <= 0x0B44; ++c) (*smap)[c] = U_Oriya;
-  for (char32 c = 0x0B47; c <= 0x0B48; ++c) (*smap)[c] = U_Oriya;
-  for (char32 c = 0x0B4B; c <= 0x0B4C; ++c) (*smap)[c] = U_Oriya;
-  (*smap)[0x0B4D] = U_Oriya;
-  (*smap)[0x0B56] = U_Oriya;
-  (*smap)[0x0B57] = U_Oriya;
-  for (char32 c = 0x0B5C; c <= 0x0B5D; ++c) (*smap)[c] = U_Oriya;
-  for (char32 c = 0x0B5F; c <= 0x0B61; ++c) (*smap)[c] = U_Oriya;
-  for (char32 c = 0x0B62; c <= 0x0B63; ++c) (*smap)[c] = U_Oriya;
-  for (char32 c = 0x0B66; c <= 0x0B6F; ++c) (*smap)[c] = U_Oriya;
-  (*smap)[0x0B70] = U_Oriya;
-  (*smap)[0x0B71] = U_Oriya;
-  for (char32 c = 0x0B72; c <= 0x0B77; ++c) (*smap)[c] = U_Oriya;
-  (*smap)[0x0B82] = U_Tamil;
-  (*smap)[0x0B83] = U_Tamil;
-  for (char32 c = 0x0B85; c <= 0x0B8A; ++c) (*smap)[c] = U_Tamil;
-  for (char32 c = 0x0B8E; c <= 0x0B90; ++c) (*smap)[c] = U_Tamil;
-  for (char32 c = 0x0B92; c <= 0x0B95; ++c) (*smap)[c] = U_Tamil;
-  for (char32 c = 0x0B99; c <= 0x0B9A; ++c) (*smap)[c] = U_Tamil;
-  (*smap)[0x0B9C] = U_Tamil;
-  for (char32 c = 0x0B9E; c <= 0x0B9F; ++c) (*smap)[c] = U_Tamil;
-  for (char32 c = 0x0BA3; c <= 0x0BA4; ++c) (*smap)[c] = U_Tamil;
-  for (char32 c = 0x0BA8; c <= 0x0BAA; ++c) (*smap)[c] = U_Tamil;
-  for (char32 c = 0x0BAE; c <= 0x0BB9; ++c) (*smap)[c] = U_Tamil;
-  for (char32 c = 0x0BBE; c <= 0x0BBF; ++c) (*smap)[c] = U_Tamil;
-  (*smap)[0x0BC0] = U_Tamil;
-  for (char32 c = 0x0BC1; c <= 0x0BC2; ++c) (*smap)[c] = U_Tamil;
-  for (char32 c = 0x0BC6; c <= 0x0BC8; ++c) (*smap)[c] = U_Tamil;
-  for (char32 c = 0x0BCA; c <= 0x0BCC; ++c) (*smap)[c] = U_Tamil;
-  (*smap)[0x0BCD] = U_Tamil;
-  (*smap)[0x0BD0] = U_Tamil;
-  (*smap)[0x0BD7] = U_Tamil;
-  for (char32 c = 0x0BE6; c <= 0x0BEF; ++c) (*smap)[c] = U_Tamil;
-  for (char32 c = 0x0BF0; c <= 0x0BF2; ++c) (*smap)[c] = U_Tamil;
-  for (char32 c = 0x0BF3; c <= 0x0BF8; ++c) (*smap)[c] = U_Tamil;
-  (*smap)[0x0BF9] = U_Tamil;
-  (*smap)[0x0BFA] = U_Tamil;
-  (*smap)[0x0C00] = U_Telugu;
-  for (char32 c = 0x0C01; c <= 0x0C03; ++c) (*smap)[c] = U_Telugu;
-  for (char32 c = 0x0C05; c <= 0x0C0C; ++c) (*smap)[c] = U_Telugu;
-  for (char32 c = 0x0C0E; c <= 0x0C10; ++c) (*smap)[c] = U_Telugu;
-  for (char32 c = 0x0C12; c <= 0x0C28; ++c) (*smap)[c] = U_Telugu;
-  for (char32 c = 0x0C2A; c <= 0x0C39; ++c) (*smap)[c] = U_Telugu;
-  (*smap)[0x0C3D] = U_Telugu;
-  for (char32 c = 0x0C3E; c <= 0x0C40; ++c) (*smap)[c] = U_Telugu;
-  for (char32 c = 0x0C41; c <= 0x0C44; ++c) (*smap)[c] = U_Telugu;
-  for (char32 c = 0x0C46; c <= 0x0C48; ++c) (*smap)[c] = U_Telugu;
-  for (char32 c = 0x0C4A; c <= 0x0C4D; ++c) (*smap)[c] = U_Telugu;
-  for (char32 c = 0x0C55; c <= 0x0C56; ++c) (*smap)[c] = U_Telugu;
-  for (char32 c = 0x0C58; c <= 0x0C5A; ++c) (*smap)[c] = U_Telugu;
-  for (char32 c = 0x0C60; c <= 0x0C61; ++c) (*smap)[c] = U_Telugu;
-  for (char32 c = 0x0C62; c <= 0x0C63; ++c) (*smap)[c] = U_Telugu;
-  for (char32 c = 0x0C66; c <= 0x0C6F; ++c) (*smap)[c] = U_Telugu;
-  for (char32 c = 0x0C78; c <= 0x0C7E; ++c) (*smap)[c] = U_Telugu;
-  (*smap)[0x0C7F] = U_Telugu;
-  (*smap)[0x0C80] = U_Kannada;
-  (*smap)[0x0C81] = U_Kannada;
-  for (char32 c = 0x0C82; c <= 0x0C83; ++c) (*smap)[c] = U_Kannada;
-  for (char32 c = 0x0C85; c <= 0x0C8C; ++c) (*smap)[c] = U_Kannada;
-  for (char32 c = 0x0C8E; c <= 0x0C90; ++c) (*smap)[c] = U_Kannada;
-  for (char32 c = 0x0C92; c <= 0x0CA8; ++c) (*smap)[c] = U_Kannada;
-  for (char32 c = 0x0CAA; c <= 0x0CB3; ++c) (*smap)[c] = U_Kannada;
-  for (char32 c = 0x0CB5; c <= 0x0CB9; ++c) (*smap)[c] = U_Kannada;
-  (*smap)[0x0CBC] = U_Kannada;
-  (*smap)[0x0CBD] = U_Kannada;
-  (*smap)[0x0CBE] = U_Kannada;
-  (*smap)[0x0CBF] = U_Kannada;
-  for (char32 c = 0x0CC0; c <= 0x0CC4; ++c) (*smap)[c] = U_Kannada;
-  (*smap)[0x0CC6] = U_Kannada;
-  for (char32 c = 0x0CC7; c <= 0x0CC8; ++c) (*smap)[c] = U_Kannada;
-  for (char32 c = 0x0CCA; c <= 0x0CCB; ++c) (*smap)[c] = U_Kannada;
-  for (char32 c = 0x0CCC; c <= 0x0CCD; ++c) (*smap)[c] = U_Kannada;
-  for (char32 c = 0x0CD5; c <= 0x0CD6; ++c) (*smap)[c] = U_Kannada;
-  (*smap)[0x0CDE] = U_Kannada;
-  for (char32 c = 0x0CE0; c <= 0x0CE1; ++c) (*smap)[c] = U_Kannada;
-  for (char32 c = 0x0CE2; c <= 0x0CE3; ++c) (*smap)[c] = U_Kannada;
-  for (char32 c = 0x0CE6; c <= 0x0CEF; ++c) (*smap)[c] = U_Kannada;
-  for (char32 c = 0x0CF1; c <= 0x0CF2; ++c) (*smap)[c] = U_Kannada;
-  (*smap)[0x0D01] = U_Malayalam;
-  for (char32 c = 0x0D02; c <= 0x0D03; ++c) (*smap)[c] = U_Malayalam;
-  for (char32 c = 0x0D05; c <= 0x0D0C; ++c) (*smap)[c] = U_Malayalam;
-  for (char32 c = 0x0D0E; c <= 0x0D10; ++c) (*smap)[c] = U_Malayalam;
-  for (char32 c = 0x0D12; c <= 0x0D3A; ++c) (*smap)[c] = U_Malayalam;
-  (*smap)[0x0D3D] = U_Malayalam;
-  for (char32 c = 0x0D3E; c <= 0x0D40; ++c) (*smap)[c] = U_Malayalam;
-  for (char32 c = 0x0D41; c <= 0x0D44; ++c) (*smap)[c] = U_Malayalam;
-  for (char32 c = 0x0D46; c <= 0x0D48; ++c) (*smap)[c] = U_Malayalam;
-  for (char32 c = 0x0D4A; c <= 0x0D4C; ++c) (*smap)[c] = U_Malayalam;
-  (*smap)[0x0D4D] = U_Malayalam;
-  (*smap)[0x0D4E] = U_Malayalam;
-  (*smap)[0x0D4F] = U_Malayalam;
-  for (char32 c = 0x0D54; c <= 0x0D56; ++c) (*smap)[c] = U_Malayalam;
-  (*smap)[0x0D57] = U_Malayalam;
-  for (char32 c = 0x0D58; c <= 0x0D5E; ++c) (*smap)[c] = U_Malayalam;
-  for (char32 c = 0x0D5F; c <= 0x0D61; ++c) (*smap)[c] = U_Malayalam;
-  for (char32 c = 0x0D62; c <= 0x0D63; ++c) (*smap)[c] = U_Malayalam;
-  for (char32 c = 0x0D66; c <= 0x0D6F; ++c) (*smap)[c] = U_Malayalam;
-  for (char32 c = 0x0D70; c <= 0x0D78; ++c) (*smap)[c] = U_Malayalam;
-  (*smap)[0x0D79] = U_Malayalam;
-  for (char32 c = 0x0D7A; c <= 0x0D7F; ++c) (*smap)[c] = U_Malayalam;
-  for (char32 c = 0x0D82; c <= 0x0D83; ++c) (*smap)[c] = U_Sinhala;
-  for (char32 c = 0x0D85; c <= 0x0D96; ++c) (*smap)[c] = U_Sinhala;
-  for (char32 c = 0x0D9A; c <= 0x0DB1; ++c) (*smap)[c] = U_Sinhala;
-  for (char32 c = 0x0DB3; c <= 0x0DBB; ++c) (*smap)[c] = U_Sinhala;
-  (*smap)[0x0DBD] = U_Sinhala;
-  for (char32 c = 0x0DC0; c <= 0x0DC6; ++c) (*smap)[c] = U_Sinhala;
-  (*smap)[0x0DCA] = U_Sinhala;
-  for (char32 c = 0x0DCF; c <= 0x0DD1; ++c) (*smap)[c] = U_Sinhala;
-  for (char32 c = 0x0DD2; c <= 0x0DD4; ++c) (*smap)[c] = U_Sinhala;
-  (*smap)[0x0DD6] = U_Sinhala;
-  for (char32 c = 0x0DD8; c <= 0x0DDF; ++c) (*smap)[c] = U_Sinhala;
-  for (char32 c = 0x0DE6; c <= 0x0DEF; ++c) (*smap)[c] = U_Sinhala;
-  for (char32 c = 0x0DF2; c <= 0x0DF3; ++c) (*smap)[c] = U_Sinhala;
-  (*smap)[0x0DF4] = U_Sinhala;
-  for (char32 c = 0x111E1; c <= 0x111F4; ++c) (*smap)[c] = U_Sinhala;
-  for (char32 c = 0x0E01; c <= 0x0E30; ++c) (*smap)[c] = U_Thai;
-  (*smap)[0x0E31] = U_Thai;
-  for (char32 c = 0x0E32; c <= 0x0E33; ++c) (*smap)[c] = U_Thai;
-  for (char32 c = 0x0E34; c <= 0x0E3A; ++c) (*smap)[c] = U_Thai;
-  for (char32 c = 0x0E40; c <= 0x0E45; ++c) (*smap)[c] = U_Thai;
-  (*smap)[0x0E46] = U_Thai;
-  for (char32 c = 0x0E47; c <= 0x0E4E; ++c) (*smap)[c] = U_Thai;
-  (*smap)[0x0E4F] = U_Thai;
-  for (char32 c = 0x0E50; c <= 0x0E59; ++c) (*smap)[c] = U_Thai;
-  for (char32 c = 0x0E5A; c <= 0x0E5B; ++c) (*smap)[c] = U_Thai;
-  for (char32 c = 0x0E81; c <= 0x0E82; ++c) (*smap)[c] = U_Lao;
-  (*smap)[0x0E84] = U_Lao;
-  for (char32 c = 0x0E87; c <= 0x0E88; ++c) (*smap)[c] = U_Lao;
-  (*smap)[0x0E8A] = U_Lao;
-  (*smap)[0x0E8D] = U_Lao;
-  for (char32 c = 0x0E94; c <= 0x0E97; ++c) (*smap)[c] = U_Lao;
-  for (char32 c = 0x0E99; c <= 0x0E9F; ++c) (*smap)[c] = U_Lao;
-  for (char32 c = 0x0EA1; c <= 0x0EA3; ++c) (*smap)[c] = U_Lao;
-  (*smap)[0x0EA5] = U_Lao;
-  (*smap)[0x0EA7] = U_Lao;
-  for (char32 c = 0x0EAA; c <= 0x0EAB; ++c) (*smap)[c] = U_Lao;
-  for (char32 c = 0x0EAD; c <= 0x0EB0; ++c) (*smap)[c] = U_Lao;
-  (*smap)[0x0EB1] = U_Lao;
-  for (char32 c = 0x0EB2; c <= 0x0EB3; ++c) (*smap)[c] = U_Lao;
-  for (char32 c = 0x0EB4; c <= 0x0EB9; ++c) (*smap)[c] = U_Lao;
-  for (char32 c = 0x0EBB; c <= 0x0EBC; ++c) (*smap)[c] = U_Lao;
-  (*smap)[0x0EBD] = U_Lao;
-  for (char32 c = 0x0EC0; c <= 0x0EC4; ++c) (*smap)[c] = U_Lao;
-  (*smap)[0x0EC6] = U_Lao;
-  for (char32 c = 0x0EC8; c <= 0x0ECD; ++c) (*smap)[c] = U_Lao;
-  for (char32 c = 0x0ED0; c <= 0x0ED9; ++c) (*smap)[c] = U_Lao;
-  for (char32 c = 0x0EDC; c <= 0x0EDF; ++c) (*smap)[c] = U_Lao;
-  (*smap)[0x0F00] = U_Tibetan;
-  for (char32 c = 0x0F01; c <= 0x0F03; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x0F04; c <= 0x0F12; ++c) (*smap)[c] = U_Tibetan;
-  (*smap)[0x0F13] = U_Tibetan;
-  (*smap)[0x0F14] = U_Tibetan;
-  for (char32 c = 0x0F15; c <= 0x0F17; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x0F18; c <= 0x0F19; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x0F1A; c <= 0x0F1F; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x0F20; c <= 0x0F29; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x0F2A; c <= 0x0F33; ++c) (*smap)[c] = U_Tibetan;
-  (*smap)[0x0F34] = U_Tibetan;
-  (*smap)[0x0F35] = U_Tibetan;
-  (*smap)[0x0F36] = U_Tibetan;
-  (*smap)[0x0F37] = U_Tibetan;
-  (*smap)[0x0F38] = U_Tibetan;
-  (*smap)[0x0F39] = U_Tibetan;
-  (*smap)[0x0F3A] = U_Tibetan;
-  (*smap)[0x0F3B] = U_Tibetan;
-  (*smap)[0x0F3C] = U_Tibetan;
-  (*smap)[0x0F3D] = U_Tibetan;
-  for (char32 c = 0x0F3E; c <= 0x0F3F; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x0F40; c <= 0x0F47; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x0F49; c <= 0x0F6C; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x0F71; c <= 0x0F7E; ++c) (*smap)[c] = U_Tibetan;
-  (*smap)[0x0F7F] = U_Tibetan;
-  for (char32 c = 0x0F80; c <= 0x0F84; ++c) (*smap)[c] = U_Tibetan;
-  (*smap)[0x0F85] = U_Tibetan;
-  for (char32 c = 0x0F86; c <= 0x0F87; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x0F88; c <= 0x0F8C; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x0F8D; c <= 0x0F97; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x0F99; c <= 0x0FBC; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x0FBE; c <= 0x0FC5; ++c) (*smap)[c] = U_Tibetan;
-  (*smap)[0x0FC6] = U_Tibetan;
-  for (char32 c = 0x0FC7; c <= 0x0FCC; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x0FCE; c <= 0x0FCF; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x0FD0; c <= 0x0FD4; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x0FD9; c <= 0x0FDA; ++c) (*smap)[c] = U_Tibetan;
-  for (char32 c = 0x1000; c <= 0x102A; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x102B; c <= 0x102C; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x102D; c <= 0x1030; ++c) (*smap)[c] = U_Myanmar;
-  (*smap)[0x1031] = U_Myanmar;
-  for (char32 c = 0x1032; c <= 0x1037; ++c) (*smap)[c] = U_Myanmar;
-  (*smap)[0x1038] = U_Myanmar;
-  for (char32 c = 0x1039; c <= 0x103A; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x103B; c <= 0x103C; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x103D; c <= 0x103E; ++c) (*smap)[c] = U_Myanmar;
-  (*smap)[0x103F] = U_Myanmar;
-  for (char32 c = 0x1040; c <= 0x1049; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x104A; c <= 0x104F; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x1050; c <= 0x1055; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x1056; c <= 0x1057; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x1058; c <= 0x1059; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x105A; c <= 0x105D; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x105E; c <= 0x1060; ++c) (*smap)[c] = U_Myanmar;
-  (*smap)[0x1061] = U_Myanmar;
-  for (char32 c = 0x1062; c <= 0x1064; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x1065; c <= 0x1066; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x1067; c <= 0x106D; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x106E; c <= 0x1070; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x1071; c <= 0x1074; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x1075; c <= 0x1081; ++c) (*smap)[c] = U_Myanmar;
-  (*smap)[0x1082] = U_Myanmar;
-  for (char32 c = 0x1083; c <= 0x1084; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x1085; c <= 0x1086; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x1087; c <= 0x108C; ++c) (*smap)[c] = U_Myanmar;
-  (*smap)[0x108D] = U_Myanmar;
-  (*smap)[0x108E] = U_Myanmar;
-  (*smap)[0x108F] = U_Myanmar;
-  for (char32 c = 0x1090; c <= 0x1099; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x109A; c <= 0x109C; ++c) (*smap)[c] = U_Myanmar;
-  (*smap)[0x109D] = U_Myanmar;
-  for (char32 c = 0x109E; c <= 0x109F; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0xA9E0; c <= 0xA9E4; ++c) (*smap)[c] = U_Myanmar;
-  (*smap)[0xA9E5] = U_Myanmar;
-  (*smap)[0xA9E6] = U_Myanmar;
-  for (char32 c = 0xA9E7; c <= 0xA9EF; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0xA9F0; c <= 0xA9F9; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0xA9FA; c <= 0xA9FE; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0xAA60; c <= 0xAA6F; ++c) (*smap)[c] = U_Myanmar;
-  (*smap)[0xAA70] = U_Myanmar;
-  for (char32 c = 0xAA71; c <= 0xAA76; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0xAA77; c <= 0xAA79; ++c) (*smap)[c] = U_Myanmar;
-  (*smap)[0xAA7A] = U_Myanmar;
-  (*smap)[0xAA7B] = U_Myanmar;
-  (*smap)[0xAA7C] = U_Myanmar;
-  (*smap)[0xAA7D] = U_Myanmar;
-  for (char32 c = 0xAA7E; c <= 0xAA7F; ++c) (*smap)[c] = U_Myanmar;
-  for (char32 c = 0x10A0; c <= 0x10C5; ++c) (*smap)[c] = U_Georgian;
-  (*smap)[0x10C7] = U_Georgian;
-  (*smap)[0x10CD] = U_Georgian;
-  for (char32 c = 0x10D0; c <= 0x10FA; ++c) (*smap)[c] = U_Georgian;
-  (*smap)[0x10FC] = U_Georgian;
-  for (char32 c = 0x10FD; c <= 0x10FF; ++c) (*smap)[c] = U_Georgian;
-  for (char32 c = 0x2D00; c <= 0x2D25; ++c) (*smap)[c] = U_Georgian;
-  (*smap)[0x2D27] = U_Georgian;
-  (*smap)[0x2D2D] = U_Georgian;
-  for (char32 c = 0x1100; c <= 0x11FF; ++c) (*smap)[c] = U_Hangul;
-  for (char32 c = 0x302E; c <= 0x302F; ++c) (*smap)[c] = U_Hangul;
-  for (char32 c = 0x3131; c <= 0x318E; ++c) (*smap)[c] = U_Hangul;
-  for (char32 c = 0x3200; c <= 0x321E; ++c) (*smap)[c] = U_Hangul;
-  for (char32 c = 0x3260; c <= 0x327E; ++c) (*smap)[c] = U_Hangul;
-  for (char32 c = 0xA960; c <= 0xA97C; ++c) (*smap)[c] = U_Hangul;
-  for (char32 c = 0xAC00; c <= 0xD7A3; ++c) (*smap)[c] = U_Hangul;
-  for (char32 c = 0xD7B0; c <= 0xD7C6; ++c) (*smap)[c] = U_Hangul;
-  for (char32 c = 0xD7CB; c <= 0xD7FB; ++c) (*smap)[c] = U_Hangul;
-  for (char32 c = 0xFFA0; c <= 0xFFBE; ++c) (*smap)[c] = U_Hangul;
-  for (char32 c = 0xFFC2; c <= 0xFFC7; ++c) (*smap)[c] = U_Hangul;
-  for (char32 c = 0xFFCA; c <= 0xFFCF; ++c) (*smap)[c] = U_Hangul;
-  for (char32 c = 0xFFD2; c <= 0xFFD7; ++c) (*smap)[c] = U_Hangul;
-  for (char32 c = 0xFFDA; c <= 0xFFDC; ++c) (*smap)[c] = U_Hangul;
-  for (char32 c = 0x1200; c <= 0x1248; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x124A; c <= 0x124D; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x1250; c <= 0x1256; ++c) (*smap)[c] = U_Ethiopic;
-  (*smap)[0x1258] = U_Ethiopic;
-  for (char32 c = 0x125A; c <= 0x125D; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x1260; c <= 0x1288; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x128A; c <= 0x128D; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x1290; c <= 0x12B0; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x12B2; c <= 0x12B5; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x12B8; c <= 0x12BE; ++c) (*smap)[c] = U_Ethiopic;
-  (*smap)[0x12C0] = U_Ethiopic;
-  for (char32 c = 0x12C2; c <= 0x12C5; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x12C8; c <= 0x12D6; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x12D8; c <= 0x1310; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x1312; c <= 0x1315; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x1318; c <= 0x135A; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x135D; c <= 0x135F; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x1360; c <= 0x1368; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x1369; c <= 0x137C; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x1380; c <= 0x138F; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x1390; c <= 0x1399; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x2D80; c <= 0x2D96; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x2DA0; c <= 0x2DA6; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x2DA8; c <= 0x2DAE; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x2DB0; c <= 0x2DB6; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x2DB8; c <= 0x2DBE; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x2DC0; c <= 0x2DC6; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x2DC8; c <= 0x2DCE; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x2DD0; c <= 0x2DD6; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x2DD8; c <= 0x2DDE; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0xAB01; c <= 0xAB06; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0xAB09; c <= 0xAB0E; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0xAB11; c <= 0xAB16; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0xAB20; c <= 0xAB26; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0xAB28; c <= 0xAB2E; ++c) (*smap)[c] = U_Ethiopic;
-  for (char32 c = 0x13A0; c <= 0x13F5; ++c) (*smap)[c] = U_Cherokee;
-  for (char32 c = 0x13F8; c <= 0x13FD; ++c) (*smap)[c] = U_Cherokee;
-  for (char32 c = 0xAB70; c <= 0xABBF; ++c) (*smap)[c] = U_Cherokee;
-  (*smap)[0x1400] = U_Canadian_Aboriginal;
-  for (char32 c = 0x1401; c <= 0x166C; ++c) (*smap)[c] = U_Canadian_Aboriginal;
-  for (char32 c = 0x166D; c <= 0x166E; ++c) (*smap)[c] = U_Canadian_Aboriginal;
-  for (char32 c = 0x166F; c <= 0x167F; ++c) (*smap)[c] = U_Canadian_Aboriginal;
-  for (char32 c = 0x18B0; c <= 0x18F5; ++c) (*smap)[c] = U_Canadian_Aboriginal;
-  (*smap)[0x1680] = U_Ogham;
-  for (char32 c = 0x1681; c <= 0x169A; ++c) (*smap)[c] = U_Ogham;
-  (*smap)[0x169B] = U_Ogham;
-  (*smap)[0x169C] = U_Ogham;
-  for (char32 c = 0x16A0; c <= 0x16EA; ++c) (*smap)[c] = U_Runic;
-  for (char32 c = 0x16EE; c <= 0x16F0; ++c) (*smap)[c] = U_Runic;
-  for (char32 c = 0x16F1; c <= 0x16F8; ++c) (*smap)[c] = U_Runic;
-  for (char32 c = 0x1780; c <= 0x17B3; ++c) (*smap)[c] = U_Khmer;
-  for (char32 c = 0x17B4; c <= 0x17B5; ++c) (*smap)[c] = U_Khmer;
-  (*smap)[0x17B6] = U_Khmer;
-  for (char32 c = 0x17B7; c <= 0x17BD; ++c) (*smap)[c] = U_Khmer;
-  for (char32 c = 0x17BE; c <= 0x17C5; ++c) (*smap)[c] = U_Khmer;
-  (*smap)[0x17C6] = U_Khmer;
-  for (char32 c = 0x17C7; c <= 0x17C8; ++c) (*smap)[c] = U_Khmer;
-  for (char32 c = 0x17C9; c <= 0x17D3; ++c) (*smap)[c] = U_Khmer;
-  for (char32 c = 0x17D4; c <= 0x17D6; ++c) (*smap)[c] = U_Khmer;
-  (*smap)[0x17D7] = U_Khmer;
-  for (char32 c = 0x17D8; c <= 0x17DA; ++c) (*smap)[c] = U_Khmer;
-  (*smap)[0x17DB] = U_Khmer;
-  (*smap)[0x17DC] = U_Khmer;
-  (*smap)[0x17DD] = U_Khmer;
-  for (char32 c = 0x17E0; c <= 0x17E9; ++c) (*smap)[c] = U_Khmer;
-  for (char32 c = 0x17F0; c <= 0x17F9; ++c) (*smap)[c] = U_Khmer;
-  for (char32 c = 0x19E0; c <= 0x19FF; ++c) (*smap)[c] = U_Khmer;
-  for (char32 c = 0x1800; c <= 0x1801; ++c) (*smap)[c] = U_Mongolian;
-  (*smap)[0x1804] = U_Mongolian;
-  (*smap)[0x1806] = U_Mongolian;
-  for (char32 c = 0x1807; c <= 0x180A; ++c) (*smap)[c] = U_Mongolian;
-  for (char32 c = 0x180B; c <= 0x180D; ++c) (*smap)[c] = U_Mongolian;
-  (*smap)[0x180E] = U_Mongolian;
-  for (char32 c = 0x1810; c <= 0x1819; ++c) (*smap)[c] = U_Mongolian;
-  for (char32 c = 0x1820; c <= 0x1842; ++c) (*smap)[c] = U_Mongolian;
-  (*smap)[0x1843] = U_Mongolian;
-  for (char32 c = 0x1844; c <= 0x1877; ++c) (*smap)[c] = U_Mongolian;
-  for (char32 c = 0x1880; c <= 0x1884; ++c) (*smap)[c] = U_Mongolian;
-  for (char32 c = 0x1885; c <= 0x1886; ++c) (*smap)[c] = U_Mongolian;
-  for (char32 c = 0x1887; c <= 0x18A8; ++c) (*smap)[c] = U_Mongolian;
-  (*smap)[0x18A9] = U_Mongolian;
-  (*smap)[0x18AA] = U_Mongolian;
-  for (char32 c = 0x11660; c <= 0x1166C; ++c) (*smap)[c] = U_Mongolian;
-  for (char32 c = 0x3041; c <= 0x3096; ++c) (*smap)[c] = U_Hiragana;
-  for (char32 c = 0x309D; c <= 0x309E; ++c) (*smap)[c] = U_Hiragana;
-  (*smap)[0x309F] = U_Hiragana;
-  (*smap)[0x1B001] = U_Hiragana;
-  (*smap)[0x1F200] = U_Hiragana;
-  for (char32 c = 0x30A1; c <= 0x30FA; ++c) (*smap)[c] = U_Katakana;
-  for (char32 c = 0x30FD; c <= 0x30FE; ++c) (*smap)[c] = U_Katakana;
-  (*smap)[0x30FF] = U_Katakana;
-  for (char32 c = 0x31F0; c <= 0x31FF; ++c) (*smap)[c] = U_Katakana;
-  for (char32 c = 0x32D0; c <= 0x32FE; ++c) (*smap)[c] = U_Katakana;
-  for (char32 c = 0x3300; c <= 0x3357; ++c) (*smap)[c] = U_Katakana;
-  for (char32 c = 0xFF66; c <= 0xFF6F; ++c) (*smap)[c] = U_Katakana;
-  for (char32 c = 0xFF71; c <= 0xFF9D; ++c) (*smap)[c] = U_Katakana;
-  (*smap)[0x1B000] = U_Katakana;
-  for (char32 c = 0x02EA; c <= 0x02EB; ++c) (*smap)[c] = U_Bopomofo;
-  for (char32 c = 0x3105; c <= 0x312D; ++c) (*smap)[c] = U_Bopomofo;
-  for (char32 c = 0x31A0; c <= 0x31BA; ++c) (*smap)[c] = U_Bopomofo;
-  for (char32 c = 0x2E80; c <= 0x2E99; ++c) (*smap)[c] = U_Han;
-  for (char32 c = 0x2E9B; c <= 0x2EF3; ++c) (*smap)[c] = U_Han;
-  for (char32 c = 0x2F00; c <= 0x2FD5; ++c) (*smap)[c] = U_Han;
-  (*smap)[0x3005] = U_Han;
-  (*smap)[0x3007] = U_Han;
-  for (char32 c = 0x3021; c <= 0x3029; ++c) (*smap)[c] = U_Han;
-  for (char32 c = 0x3038; c <= 0x303A; ++c) (*smap)[c] = U_Han;
-  (*smap)[0x303B] = U_Han;
-  for (char32 c = 0x3400; c <= 0x4DB5; ++c) (*smap)[c] = U_Han;
-  for (char32 c = 0x4E00; c <= 0x9FD5; ++c) (*smap)[c] = U_Han;
-  for (char32 c = 0xF900; c <= 0xFA6D; ++c) (*smap)[c] = U_Han;
-  for (char32 c = 0xFA70; c <= 0xFAD9; ++c) (*smap)[c] = U_Han;
-  for (char32 c = 0x20000; c <= 0x2A6D6; ++c) (*smap)[c] = U_Han;
-  for (char32 c = 0x2A700; c <= 0x2B734; ++c) (*smap)[c] = U_Han;
-  for (char32 c = 0x2B740; c <= 0x2B81D; ++c) (*smap)[c] = U_Han;
-  for (char32 c = 0x2B820; c <= 0x2CEA1; ++c) (*smap)[c] = U_Han;
-  for (char32 c = 0x2F800; c <= 0x2FA1D; ++c) (*smap)[c] = U_Han;
-  for (char32 c = 0xA000; c <= 0xA014; ++c) (*smap)[c] = U_Yi;
-  (*smap)[0xA015] = U_Yi;
-  for (char32 c = 0xA016; c <= 0xA48C; ++c) (*smap)[c] = U_Yi;
-  for (char32 c = 0xA490; c <= 0xA4C6; ++c) (*smap)[c] = U_Yi;
-  for (char32 c = 0x10300; c <= 0x1031F; ++c) (*smap)[c] = U_Old_Italic;
-  for (char32 c = 0x10320; c <= 0x10323; ++c) (*smap)[c] = U_Old_Italic;
-  for (char32 c = 0x10330; c <= 0x10340; ++c) (*smap)[c] = U_Gothic;
-  (*smap)[0x10341] = U_Gothic;
-  for (char32 c = 0x10342; c <= 0x10349; ++c) (*smap)[c] = U_Gothic;
-  (*smap)[0x1034A] = U_Gothic;
-  for (char32 c = 0x10400; c <= 0x1044F; ++c) (*smap)[c] = U_Deseret;
-  for (char32 c = 0x0300; c <= 0x036F; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x0485; c <= 0x0486; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x064B; c <= 0x0655; ++c) (*smap)[c] = U_Inherited;
-  (*smap)[0x0670] = U_Inherited;
-  for (char32 c = 0x0951; c <= 0x0952; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x1AB0; c <= 0x1ABD; ++c) (*smap)[c] = U_Inherited;
-  (*smap)[0x1ABE] = U_Inherited;
-  for (char32 c = 0x1CD0; c <= 0x1CD2; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x1CD4; c <= 0x1CE0; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x1CE2; c <= 0x1CE8; ++c) (*smap)[c] = U_Inherited;
-  (*smap)[0x1CED] = U_Inherited;
-  (*smap)[0x1CF4] = U_Inherited;
-  for (char32 c = 0x1CF8; c <= 0x1CF9; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x1DC0; c <= 0x1DF5; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x1DFB; c <= 0x1DFF; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x200C; c <= 0x200D; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x20D0; c <= 0x20DC; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x20DD; c <= 0x20E0; ++c) (*smap)[c] = U_Inherited;
-  (*smap)[0x20E1] = U_Inherited;
-  for (char32 c = 0x20E2; c <= 0x20E4; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x20E5; c <= 0x20F0; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x302A; c <= 0x302D; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x3099; c <= 0x309A; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0xFE00; c <= 0xFE0F; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0xFE20; c <= 0xFE2D; ++c) (*smap)[c] = U_Inherited;
-  (*smap)[0x101FD] = U_Inherited;
-  (*smap)[0x102E0] = U_Inherited;
-  for (char32 c = 0x1D167; c <= 0x1D169; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x1D17B; c <= 0x1D182; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x1D185; c <= 0x1D18B; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x1D1AA; c <= 0x1D1AD; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0xE0100; c <= 0xE01EF; ++c) (*smap)[c] = U_Inherited;
-  for (char32 c = 0x1700; c <= 0x170C; ++c) (*smap)[c] = U_Tagalog;
-  for (char32 c = 0x170E; c <= 0x1711; ++c) (*smap)[c] = U_Tagalog;
-  for (char32 c = 0x1712; c <= 0x1714; ++c) (*smap)[c] = U_Tagalog;
-  for (char32 c = 0x1720; c <= 0x1731; ++c) (*smap)[c] = U_Hanunoo;
-  for (char32 c = 0x1732; c <= 0x1734; ++c) (*smap)[c] = U_Hanunoo;
-  for (char32 c = 0x1740; c <= 0x1751; ++c) (*smap)[c] = U_Buhid;
-  for (char32 c = 0x1752; c <= 0x1753; ++c) (*smap)[c] = U_Buhid;
-  for (char32 c = 0x1760; c <= 0x176C; ++c) (*smap)[c] = U_Tagbanwa;
-  for (char32 c = 0x176E; c <= 0x1770; ++c) (*smap)[c] = U_Tagbanwa;
-  for (char32 c = 0x1772; c <= 0x1773; ++c) (*smap)[c] = U_Tagbanwa;
-  for (char32 c = 0x1900; c <= 0x191E; ++c) (*smap)[c] = U_Limbu;
-  for (char32 c = 0x1920; c <= 0x1922; ++c) (*smap)[c] = U_Limbu;
-  for (char32 c = 0x1923; c <= 0x1926; ++c) (*smap)[c] = U_Limbu;
-  for (char32 c = 0x1927; c <= 0x1928; ++c) (*smap)[c] = U_Limbu;
-  for (char32 c = 0x1929; c <= 0x192B; ++c) (*smap)[c] = U_Limbu;
-  for (char32 c = 0x1930; c <= 0x1931; ++c) (*smap)[c] = U_Limbu;
-  (*smap)[0x1932] = U_Limbu;
-  for (char32 c = 0x1933; c <= 0x1938; ++c) (*smap)[c] = U_Limbu;
-  for (char32 c = 0x1939; c <= 0x193B; ++c) (*smap)[c] = U_Limbu;
-  (*smap)[0x1940] = U_Limbu;
-  for (char32 c = 0x1944; c <= 0x1945; ++c) (*smap)[c] = U_Limbu;
-  for (char32 c = 0x1946; c <= 0x194F; ++c) (*smap)[c] = U_Limbu;
-  for (char32 c = 0x1950; c <= 0x196D; ++c) (*smap)[c] = U_Tai_Le;
-  for (char32 c = 0x1970; c <= 0x1974; ++c) (*smap)[c] = U_Tai_Le;
-  for (char32 c = 0x10000; c <= 0x1000B; ++c) (*smap)[c] = U_Linear_B;
-  for (char32 c = 0x1000D; c <= 0x10026; ++c) (*smap)[c] = U_Linear_B;
-  for (char32 c = 0x10028; c <= 0x1003A; ++c) (*smap)[c] = U_Linear_B;
-  for (char32 c = 0x1003C; c <= 0x1003D; ++c) (*smap)[c] = U_Linear_B;
-  for (char32 c = 0x1003F; c <= 0x1004D; ++c) (*smap)[c] = U_Linear_B;
-  for (char32 c = 0x10050; c <= 0x1005D; ++c) (*smap)[c] = U_Linear_B;
-  for (char32 c = 0x10080; c <= 0x100FA; ++c) (*smap)[c] = U_Linear_B;
-  for (char32 c = 0x10380; c <= 0x1039D; ++c) (*smap)[c] = U_Ugaritic;
-  (*smap)[0x1039F] = U_Ugaritic;
-  for (char32 c = 0x10450; c <= 0x1047F; ++c) (*smap)[c] = U_Shavian;
-  for (char32 c = 0x10480; c <= 0x1049D; ++c) (*smap)[c] = U_Osmanya;
-  for (char32 c = 0x104A0; c <= 0x104A9; ++c) (*smap)[c] = U_Osmanya;
-  for (char32 c = 0x10800; c <= 0x10805; ++c) (*smap)[c] = U_Cypriot;
-  (*smap)[0x10808] = U_Cypriot;
-  for (char32 c = 0x1080A; c <= 0x10835; ++c) (*smap)[c] = U_Cypriot;
-  for (char32 c = 0x10837; c <= 0x10838; ++c) (*smap)[c] = U_Cypriot;
-  (*smap)[0x1083C] = U_Cypriot;
-  (*smap)[0x1083F] = U_Cypriot;
-  for (char32 c = 0x2800; c <= 0x28FF; ++c) (*smap)[c] = U_Braille;
-  for (char32 c = 0x1A00; c <= 0x1A16; ++c) (*smap)[c] = U_Buginese;
-  for (char32 c = 0x1A17; c <= 0x1A18; ++c) (*smap)[c] = U_Buginese;
-  for (char32 c = 0x1A19; c <= 0x1A1A; ++c) (*smap)[c] = U_Buginese;
-  (*smap)[0x1A1B] = U_Buginese;
-  for (char32 c = 0x1A1E; c <= 0x1A1F; ++c) (*smap)[c] = U_Buginese;
-  for (char32 c = 0x03E2; c <= 0x03EF; ++c) (*smap)[c] = U_Coptic;
-  for (char32 c = 0x2C80; c <= 0x2CE4; ++c) (*smap)[c] = U_Coptic;
-  for (char32 c = 0x2CE5; c <= 0x2CEA; ++c) (*smap)[c] = U_Coptic;
-  for (char32 c = 0x2CEB; c <= 0x2CEE; ++c) (*smap)[c] = U_Coptic;
-  for (char32 c = 0x2CEF; c <= 0x2CF1; ++c) (*smap)[c] = U_Coptic;
-  for (char32 c = 0x2CF2; c <= 0x2CF3; ++c) (*smap)[c] = U_Coptic;
-  for (char32 c = 0x2CF9; c <= 0x2CFC; ++c) (*smap)[c] = U_Coptic;
-  (*smap)[0x2CFD] = U_Coptic;
-  for (char32 c = 0x2CFE; c <= 0x2CFF; ++c) (*smap)[c] = U_Coptic;
-  for (char32 c = 0x1980; c <= 0x19AB; ++c) (*smap)[c] = U_New_Tai_Lue;
-  for (char32 c = 0x19B0; c <= 0x19C9; ++c) (*smap)[c] = U_New_Tai_Lue;
-  for (char32 c = 0x19D0; c <= 0x19D9; ++c) (*smap)[c] = U_New_Tai_Lue;
-  (*smap)[0x19DA] = U_New_Tai_Lue;
-  for (char32 c = 0x19DE; c <= 0x19DF; ++c) (*smap)[c] = U_New_Tai_Lue;
-  for (char32 c = 0x2C00; c <= 0x2C2E; ++c) (*smap)[c] = U_Glagolitic;
-  for (char32 c = 0x2C30; c <= 0x2C5E; ++c) (*smap)[c] = U_Glagolitic;
-  for (char32 c = 0x1E000; c <= 0x1E006; ++c) (*smap)[c] = U_Glagolitic;
-  for (char32 c = 0x1E008; c <= 0x1E018; ++c) (*smap)[c] = U_Glagolitic;
-  for (char32 c = 0x1E01B; c <= 0x1E021; ++c) (*smap)[c] = U_Glagolitic;
-  for (char32 c = 0x1E023; c <= 0x1E024; ++c) (*smap)[c] = U_Glagolitic;
-  for (char32 c = 0x1E026; c <= 0x1E02A; ++c) (*smap)[c] = U_Glagolitic;
-  for (char32 c = 0x2D30; c <= 0x2D67; ++c) (*smap)[c] = U_Tifinagh;
-  (*smap)[0x2D6F] = U_Tifinagh;
-  (*smap)[0x2D70] = U_Tifinagh;
-  (*smap)[0x2D7F] = U_Tifinagh;
-  for (char32 c = 0xA800; c <= 0xA801; ++c) (*smap)[c] = U_Syloti_Nagri;
-  (*smap)[0xA802] = U_Syloti_Nagri;
-  for (char32 c = 0xA803; c <= 0xA805; ++c) (*smap)[c] = U_Syloti_Nagri;
-  (*smap)[0xA806] = U_Syloti_Nagri;
-  for (char32 c = 0xA807; c <= 0xA80A; ++c) (*smap)[c] = U_Syloti_Nagri;
-  (*smap)[0xA80B] = U_Syloti_Nagri;
-  for (char32 c = 0xA80C; c <= 0xA822; ++c) (*smap)[c] = U_Syloti_Nagri;
-  for (char32 c = 0xA823; c <= 0xA824; ++c) (*smap)[c] = U_Syloti_Nagri;
-  for (char32 c = 0xA825; c <= 0xA826; ++c) (*smap)[c] = U_Syloti_Nagri;
-  (*smap)[0xA827] = U_Syloti_Nagri;
-  for (char32 c = 0xA828; c <= 0xA82B; ++c) (*smap)[c] = U_Syloti_Nagri;
-  for (char32 c = 0x103A0; c <= 0x103C3; ++c) (*smap)[c] = U_Old_Persian;
-  for (char32 c = 0x103C8; c <= 0x103CF; ++c) (*smap)[c] = U_Old_Persian;
-  (*smap)[0x103D0] = U_Old_Persian;
-  for (char32 c = 0x103D1; c <= 0x103D5; ++c) (*smap)[c] = U_Old_Persian;
-  (*smap)[0x10A00] = U_Kharoshthi;
-  for (char32 c = 0x10A01; c <= 0x10A03; ++c) (*smap)[c] = U_Kharoshthi;
-  for (char32 c = 0x10A05; c <= 0x10A06; ++c) (*smap)[c] = U_Kharoshthi;
-  for (char32 c = 0x10A0C; c <= 0x10A0F; ++c) (*smap)[c] = U_Kharoshthi;
-  for (char32 c = 0x10A10; c <= 0x10A13; ++c) (*smap)[c] = U_Kharoshthi;
-  for (char32 c = 0x10A15; c <= 0x10A17; ++c) (*smap)[c] = U_Kharoshthi;
-  for (char32 c = 0x10A19; c <= 0x10A33; ++c) (*smap)[c] = U_Kharoshthi;
-  for (char32 c = 0x10A38; c <= 0x10A3A; ++c) (*smap)[c] = U_Kharoshthi;
-  (*smap)[0x10A3F] = U_Kharoshthi;
-  for (char32 c = 0x10A40; c <= 0x10A47; ++c) (*smap)[c] = U_Kharoshthi;
-  for (char32 c = 0x10A50; c <= 0x10A58; ++c) (*smap)[c] = U_Kharoshthi;
-  for (char32 c = 0x1B00; c <= 0x1B03; ++c) (*smap)[c] = U_Balinese;
-  (*smap)[0x1B04] = U_Balinese;
-  for (char32 c = 0x1B05; c <= 0x1B33; ++c) (*smap)[c] = U_Balinese;
-  (*smap)[0x1B34] = U_Balinese;
-  (*smap)[0x1B35] = U_Balinese;
-  for (char32 c = 0x1B36; c <= 0x1B3A; ++c) (*smap)[c] = U_Balinese;
-  (*smap)[0x1B3B] = U_Balinese;
-  (*smap)[0x1B3C] = U_Balinese;
-  for (char32 c = 0x1B3D; c <= 0x1B41; ++c) (*smap)[c] = U_Balinese;
-  (*smap)[0x1B42] = U_Balinese;
-  for (char32 c = 0x1B43; c <= 0x1B44; ++c) (*smap)[c] = U_Balinese;
-  for (char32 c = 0x1B45; c <= 0x1B4B; ++c) (*smap)[c] = U_Balinese;
-  for (char32 c = 0x1B50; c <= 0x1B59; ++c) (*smap)[c] = U_Balinese;
-  for (char32 c = 0x1B5A; c <= 0x1B60; ++c) (*smap)[c] = U_Balinese;
-  for (char32 c = 0x1B61; c <= 0x1B6A; ++c) (*smap)[c] = U_Balinese;
-  for (char32 c = 0x1B6B; c <= 0x1B73; ++c) (*smap)[c] = U_Balinese;
-  for (char32 c = 0x1B74; c <= 0x1B7C; ++c) (*smap)[c] = U_Balinese;
-  for (char32 c = 0x12000; c <= 0x12399; ++c) (*smap)[c] = U_Cuneiform;
-  for (char32 c = 0x12400; c <= 0x1246E; ++c) (*smap)[c] = U_Cuneiform;
-  for (char32 c = 0x12470; c <= 0x12474; ++c) (*smap)[c] = U_Cuneiform;
-  for (char32 c = 0x12480; c <= 0x12543; ++c) (*smap)[c] = U_Cuneiform;
-  for (char32 c = 0x10900; c <= 0x10915; ++c) (*smap)[c] = U_Phoenician;
-  for (char32 c = 0x10916; c <= 0x1091B; ++c) (*smap)[c] = U_Phoenician;
-  (*smap)[0x1091F] = U_Phoenician;
-  for (char32 c = 0xA840; c <= 0xA873; ++c) (*smap)[c] = U_Phags_Pa;
-  for (char32 c = 0xA874; c <= 0xA877; ++c) (*smap)[c] = U_Phags_Pa;
-  for (char32 c = 0x07C0; c <= 0x07C9; ++c) (*smap)[c] = U_Nko;
-  for (char32 c = 0x07CA; c <= 0x07EA; ++c) (*smap)[c] = U_Nko;
-  for (char32 c = 0x07EB; c <= 0x07F3; ++c) (*smap)[c] = U_Nko;
-  for (char32 c = 0x07F4; c <= 0x07F5; ++c) (*smap)[c] = U_Nko;
-  (*smap)[0x07F6] = U_Nko;
-  for (char32 c = 0x07F7; c <= 0x07F9; ++c) (*smap)[c] = U_Nko;
-  (*smap)[0x07FA] = U_Nko;
-  for (char32 c = 0x1B80; c <= 0x1B81; ++c) (*smap)[c] = U_Sundanese;
-  (*smap)[0x1B82] = U_Sundanese;
-  for (char32 c = 0x1B83; c <= 0x1BA0; ++c) (*smap)[c] = U_Sundanese;
-  (*smap)[0x1BA1] = U_Sundanese;
-  for (char32 c = 0x1BA2; c <= 0x1BA5; ++c) (*smap)[c] = U_Sundanese;
-  for (char32 c = 0x1BA6; c <= 0x1BA7; ++c) (*smap)[c] = U_Sundanese;
-  for (char32 c = 0x1BA8; c <= 0x1BA9; ++c) (*smap)[c] = U_Sundanese;
-  (*smap)[0x1BAA] = U_Sundanese;
-  for (char32 c = 0x1BAB; c <= 0x1BAD; ++c) (*smap)[c] = U_Sundanese;
-  for (char32 c = 0x1BAE; c <= 0x1BAF; ++c) (*smap)[c] = U_Sundanese;
-  for (char32 c = 0x1BB0; c <= 0x1BB9; ++c) (*smap)[c] = U_Sundanese;
-  for (char32 c = 0x1BBA; c <= 0x1BBF; ++c) (*smap)[c] = U_Sundanese;
-  for (char32 c = 0x1CC0; c <= 0x1CC7; ++c) (*smap)[c] = U_Sundanese;
-  for (char32 c = 0x1C00; c <= 0x1C23; ++c) (*smap)[c] = U_Lepcha;
-  for (char32 c = 0x1C24; c <= 0x1C2B; ++c) (*smap)[c] = U_Lepcha;
-  for (char32 c = 0x1C2C; c <= 0x1C33; ++c) (*smap)[c] = U_Lepcha;
-  for (char32 c = 0x1C34; c <= 0x1C35; ++c) (*smap)[c] = U_Lepcha;
-  for (char32 c = 0x1C36; c <= 0x1C37; ++c) (*smap)[c] = U_Lepcha;
-  for (char32 c = 0x1C3B; c <= 0x1C3F; ++c) (*smap)[c] = U_Lepcha;
-  for (char32 c = 0x1C40; c <= 0x1C49; ++c) (*smap)[c] = U_Lepcha;
-  for (char32 c = 0x1C4D; c <= 0x1C4F; ++c) (*smap)[c] = U_Lepcha;
-  for (char32 c = 0x1C50; c <= 0x1C59; ++c) (*smap)[c] = U_Ol_Chiki;
-  for (char32 c = 0x1C5A; c <= 0x1C77; ++c) (*smap)[c] = U_Ol_Chiki;
-  for (char32 c = 0x1C78; c <= 0x1C7D; ++c) (*smap)[c] = U_Ol_Chiki;
-  for (char32 c = 0x1C7E; c <= 0x1C7F; ++c) (*smap)[c] = U_Ol_Chiki;
-  for (char32 c = 0xA500; c <= 0xA60B; ++c) (*smap)[c] = U_Vai;
-  (*smap)[0xA60C] = U_Vai;
-  for (char32 c = 0xA60D; c <= 0xA60F; ++c) (*smap)[c] = U_Vai;
-  for (char32 c = 0xA610; c <= 0xA61F; ++c) (*smap)[c] = U_Vai;
-  for (char32 c = 0xA620; c <= 0xA629; ++c) (*smap)[c] = U_Vai;
-  for (char32 c = 0xA62A; c <= 0xA62B; ++c) (*smap)[c] = U_Vai;
-  for (char32 c = 0xA880; c <= 0xA881; ++c) (*smap)[c] = U_Saurashtra;
-  for (char32 c = 0xA882; c <= 0xA8B3; ++c) (*smap)[c] = U_Saurashtra;
-  for (char32 c = 0xA8B4; c <= 0xA8C3; ++c) (*smap)[c] = U_Saurashtra;
-  for (char32 c = 0xA8C4; c <= 0xA8C5; ++c) (*smap)[c] = U_Saurashtra;
-  for (char32 c = 0xA8CE; c <= 0xA8CF; ++c) (*smap)[c] = U_Saurashtra;
-  for (char32 c = 0xA8D0; c <= 0xA8D9; ++c) (*smap)[c] = U_Saurashtra;
-  for (char32 c = 0xA900; c <= 0xA909; ++c) (*smap)[c] = U_Kayah_Li;
-  for (char32 c = 0xA90A; c <= 0xA925; ++c) (*smap)[c] = U_Kayah_Li;
-  for (char32 c = 0xA926; c <= 0xA92D; ++c) (*smap)[c] = U_Kayah_Li;
-  (*smap)[0xA92F] = U_Kayah_Li;
-  for (char32 c = 0xA930; c <= 0xA946; ++c) (*smap)[c] = U_Rejang;
-  for (char32 c = 0xA947; c <= 0xA951; ++c) (*smap)[c] = U_Rejang;
-  for (char32 c = 0xA952; c <= 0xA953; ++c) (*smap)[c] = U_Rejang;
-  (*smap)[0xA95F] = U_Rejang;
-  for (char32 c = 0x10280; c <= 0x1029C; ++c) (*smap)[c] = U_Lycian;
-  for (char32 c = 0x102A0; c <= 0x102D0; ++c) (*smap)[c] = U_Carian;
-  for (char32 c = 0x10920; c <= 0x10939; ++c) (*smap)[c] = U_Lydian;
-  (*smap)[0x1093F] = U_Lydian;
-  for (char32 c = 0xAA00; c <= 0xAA28; ++c) (*smap)[c] = U_Cham;
-  for (char32 c = 0xAA29; c <= 0xAA2E; ++c) (*smap)[c] = U_Cham;
-  for (char32 c = 0xAA2F; c <= 0xAA30; ++c) (*smap)[c] = U_Cham;
-  for (char32 c = 0xAA31; c <= 0xAA32; ++c) (*smap)[c] = U_Cham;
-  for (char32 c = 0xAA33; c <= 0xAA34; ++c) (*smap)[c] = U_Cham;
-  for (char32 c = 0xAA35; c <= 0xAA36; ++c) (*smap)[c] = U_Cham;
-  for (char32 c = 0xAA40; c <= 0xAA42; ++c) (*smap)[c] = U_Cham;
-  (*smap)[0xAA43] = U_Cham;
-  for (char32 c = 0xAA44; c <= 0xAA4B; ++c) (*smap)[c] = U_Cham;
-  (*smap)[0xAA4C] = U_Cham;
-  (*smap)[0xAA4D] = U_Cham;
-  for (char32 c = 0xAA50; c <= 0xAA59; ++c) (*smap)[c] = U_Cham;
-  for (char32 c = 0xAA5C; c <= 0xAA5F; ++c) (*smap)[c] = U_Cham;
-  for (char32 c = 0x1A20; c <= 0x1A54; ++c) (*smap)[c] = U_Tai_Tham;
-  (*smap)[0x1A55] = U_Tai_Tham;
-  (*smap)[0x1A56] = U_Tai_Tham;
-  (*smap)[0x1A57] = U_Tai_Tham;
-  for (char32 c = 0x1A58; c <= 0x1A5E; ++c) (*smap)[c] = U_Tai_Tham;
-  (*smap)[0x1A60] = U_Tai_Tham;
-  (*smap)[0x1A61] = U_Tai_Tham;
-  (*smap)[0x1A62] = U_Tai_Tham;
-  for (char32 c = 0x1A63; c <= 0x1A64; ++c) (*smap)[c] = U_Tai_Tham;
-  for (char32 c = 0x1A65; c <= 0x1A6C; ++c) (*smap)[c] = U_Tai_Tham;
-  for (char32 c = 0x1A6D; c <= 0x1A72; ++c) (*smap)[c] = U_Tai_Tham;
-  for (char32 c = 0x1A73; c <= 0x1A7C; ++c) (*smap)[c] = U_Tai_Tham;
-  (*smap)[0x1A7F] = U_Tai_Tham;
-  for (char32 c = 0x1A80; c <= 0x1A89; ++c) (*smap)[c] = U_Tai_Tham;
-  for (char32 c = 0x1A90; c <= 0x1A99; ++c) (*smap)[c] = U_Tai_Tham;
-  for (char32 c = 0x1AA0; c <= 0x1AA6; ++c) (*smap)[c] = U_Tai_Tham;
-  (*smap)[0x1AA7] = U_Tai_Tham;
-  for (char32 c = 0x1AA8; c <= 0x1AAD; ++c) (*smap)[c] = U_Tai_Tham;
-  for (char32 c = 0xAA80; c <= 0xAAAF; ++c) (*smap)[c] = U_Tai_Viet;
-  (*smap)[0xAAB0] = U_Tai_Viet;
-  (*smap)[0xAAB1] = U_Tai_Viet;
-  for (char32 c = 0xAAB2; c <= 0xAAB4; ++c) (*smap)[c] = U_Tai_Viet;
-  for (char32 c = 0xAAB5; c <= 0xAAB6; ++c) (*smap)[c] = U_Tai_Viet;
-  for (char32 c = 0xAAB7; c <= 0xAAB8; ++c) (*smap)[c] = U_Tai_Viet;
-  for (char32 c = 0xAAB9; c <= 0xAABD; ++c) (*smap)[c] = U_Tai_Viet;
-  for (char32 c = 0xAABE; c <= 0xAABF; ++c) (*smap)[c] = U_Tai_Viet;
-  (*smap)[0xAAC0] = U_Tai_Viet;
-  (*smap)[0xAAC1] = U_Tai_Viet;
-  (*smap)[0xAAC2] = U_Tai_Viet;
-  for (char32 c = 0xAADB; c <= 0xAADC; ++c) (*smap)[c] = U_Tai_Viet;
-  (*smap)[0xAADD] = U_Tai_Viet;
-  for (char32 c = 0xAADE; c <= 0xAADF; ++c) (*smap)[c] = U_Tai_Viet;
-  for (char32 c = 0x10B00; c <= 0x10B35; ++c) (*smap)[c] = U_Avestan;
-  for (char32 c = 0x10B39; c <= 0x10B3F; ++c) (*smap)[c] = U_Avestan;
-  for (char32 c = 0x13000; c <= 0x1342E; ++c)
-    (*smap)[c] = U_Egyptian_Hieroglyphs;
-  for (char32 c = 0x0800; c <= 0x0815; ++c) (*smap)[c] = U_Samaritan;
-  for (char32 c = 0x0816; c <= 0x0819; ++c) (*smap)[c] = U_Samaritan;
-  (*smap)[0x081A] = U_Samaritan;
-  for (char32 c = 0x081B; c <= 0x0823; ++c) (*smap)[c] = U_Samaritan;
-  (*smap)[0x0824] = U_Samaritan;
-  for (char32 c = 0x0825; c <= 0x0827; ++c) (*smap)[c] = U_Samaritan;
-  (*smap)[0x0828] = U_Samaritan;
-  for (char32 c = 0x0829; c <= 0x082D; ++c) (*smap)[c] = U_Samaritan;
-  for (char32 c = 0x0830; c <= 0x083E; ++c) (*smap)[c] = U_Samaritan;
-  for (char32 c = 0xA4D0; c <= 0xA4F7; ++c) (*smap)[c] = U_Lisu;
-  for (char32 c = 0xA4F8; c <= 0xA4FD; ++c) (*smap)[c] = U_Lisu;
-  for (char32 c = 0xA4FE; c <= 0xA4FF; ++c) (*smap)[c] = U_Lisu;
-  for (char32 c = 0xA6A0; c <= 0xA6E5; ++c) (*smap)[c] = U_Bamum;
-  for (char32 c = 0xA6E6; c <= 0xA6EF; ++c) (*smap)[c] = U_Bamum;
-  for (char32 c = 0xA6F0; c <= 0xA6F1; ++c) (*smap)[c] = U_Bamum;
-  for (char32 c = 0xA6F2; c <= 0xA6F7; ++c) (*smap)[c] = U_Bamum;
-  for (char32 c = 0x16800; c <= 0x16A38; ++c) (*smap)[c] = U_Bamum;
-  for (char32 c = 0xA980; c <= 0xA982; ++c) (*smap)[c] = U_Javanese;
-  (*smap)[0xA983] = U_Javanese;
-  for (char32 c = 0xA984; c <= 0xA9B2; ++c) (*smap)[c] = U_Javanese;
-  (*smap)[0xA9B3] = U_Javanese;
-  for (char32 c = 0xA9B4; c <= 0xA9B5; ++c) (*smap)[c] = U_Javanese;
-  for (char32 c = 0xA9B6; c <= 0xA9B9; ++c) (*smap)[c] = U_Javanese;
-  for (char32 c = 0xA9BA; c <= 0xA9BB; ++c) (*smap)[c] = U_Javanese;
-  (*smap)[0xA9BC] = U_Javanese;
-  for (char32 c = 0xA9BD; c <= 0xA9C0; ++c) (*smap)[c] = U_Javanese;
-  for (char32 c = 0xA9C1; c <= 0xA9CD; ++c) (*smap)[c] = U_Javanese;
-  for (char32 c = 0xA9D0; c <= 0xA9D9; ++c) (*smap)[c] = U_Javanese;
-  for (char32 c = 0xA9DE; c <= 0xA9DF; ++c) (*smap)[c] = U_Javanese;
-  for (char32 c = 0xAAE0; c <= 0xAAEA; ++c) (*smap)[c] = U_Meetei_Mayek;
-  (*smap)[0xAAEB] = U_Meetei_Mayek;
-  for (char32 c = 0xAAEC; c <= 0xAAED; ++c) (*smap)[c] = U_Meetei_Mayek;
-  for (char32 c = 0xAAEE; c <= 0xAAEF; ++c) (*smap)[c] = U_Meetei_Mayek;
-  for (char32 c = 0xAAF0; c <= 0xAAF1; ++c) (*smap)[c] = U_Meetei_Mayek;
-  (*smap)[0xAAF2] = U_Meetei_Mayek;
-  for (char32 c = 0xAAF3; c <= 0xAAF4; ++c) (*smap)[c] = U_Meetei_Mayek;
-  (*smap)[0xAAF5] = U_Meetei_Mayek;
-  (*smap)[0xAAF6] = U_Meetei_Mayek;
-  for (char32 c = 0xABC0; c <= 0xABE2; ++c) (*smap)[c] = U_Meetei_Mayek;
-  for (char32 c = 0xABE3; c <= 0xABE4; ++c) (*smap)[c] = U_Meetei_Mayek;
-  (*smap)[0xABE5] = U_Meetei_Mayek;
-  for (char32 c = 0xABE6; c <= 0xABE7; ++c) (*smap)[c] = U_Meetei_Mayek;
-  (*smap)[0xABE8] = U_Meetei_Mayek;
-  for (char32 c = 0xABE9; c <= 0xABEA; ++c) (*smap)[c] = U_Meetei_Mayek;
-  (*smap)[0xABEB] = U_Meetei_Mayek;
-  (*smap)[0xABEC] = U_Meetei_Mayek;
-  (*smap)[0xABED] = U_Meetei_Mayek;
-  for (char32 c = 0xABF0; c <= 0xABF9; ++c) (*smap)[c] = U_Meetei_Mayek;
-  for (char32 c = 0x10840; c <= 0x10855; ++c) (*smap)[c] = U_Imperial_Aramaic;
-  (*smap)[0x10857] = U_Imperial_Aramaic;
-  for (char32 c = 0x10858; c <= 0x1085F; ++c) (*smap)[c] = U_Imperial_Aramaic;
-  for (char32 c = 0x10A60; c <= 0x10A7C; ++c) (*smap)[c] = U_Old_South_Arabian;
-  for (char32 c = 0x10A7D; c <= 0x10A7E; ++c) (*smap)[c] = U_Old_South_Arabian;
-  (*smap)[0x10A7F] = U_Old_South_Arabian;
-  for (char32 c = 0x10B40; c <= 0x10B55; ++c)
-    (*smap)[c] = U_Inscriptional_Parthian;
-  for (char32 c = 0x10B58; c <= 0x10B5F; ++c)
-    (*smap)[c] = U_Inscriptional_Parthian;
-  for (char32 c = 0x10B60; c <= 0x10B72; ++c)
-    (*smap)[c] = U_Inscriptional_Pahlavi;
-  for (char32 c = 0x10B78; c <= 0x10B7F; ++c)
-    (*smap)[c] = U_Inscriptional_Pahlavi;
-  for (char32 c = 0x10C00; c <= 0x10C48; ++c) (*smap)[c] = U_Old_Turkic;
-  for (char32 c = 0x11080; c <= 0x11081; ++c) (*smap)[c] = U_Kaithi;
-  (*smap)[0x11082] = U_Kaithi;
-  for (char32 c = 0x11083; c <= 0x110AF; ++c) (*smap)[c] = U_Kaithi;
-  for (char32 c = 0x110B0; c <= 0x110B2; ++c) (*smap)[c] = U_Kaithi;
-  for (char32 c = 0x110B3; c <= 0x110B6; ++c) (*smap)[c] = U_Kaithi;
-  for (char32 c = 0x110B7; c <= 0x110B8; ++c) (*smap)[c] = U_Kaithi;
-  for (char32 c = 0x110B9; c <= 0x110BA; ++c) (*smap)[c] = U_Kaithi;
-  for (char32 c = 0x110BB; c <= 0x110BC; ++c) (*smap)[c] = U_Kaithi;
-  (*smap)[0x110BD] = U_Kaithi;
-  for (char32 c = 0x110BE; c <= 0x110C1; ++c) (*smap)[c] = U_Kaithi;
-  for (char32 c = 0x1BC0; c <= 0x1BE5; ++c) (*smap)[c] = U_Batak;
-  (*smap)[0x1BE6] = U_Batak;
-  (*smap)[0x1BE7] = U_Batak;
-  for (char32 c = 0x1BE8; c <= 0x1BE9; ++c) (*smap)[c] = U_Batak;
-  for (char32 c = 0x1BEA; c <= 0x1BEC; ++c) (*smap)[c] = U_Batak;
-  (*smap)[0x1BED] = U_Batak;
-  (*smap)[0x1BEE] = U_Batak;
-  for (char32 c = 0x1BEF; c <= 0x1BF1; ++c) (*smap)[c] = U_Batak;
-  for (char32 c = 0x1BF2; c <= 0x1BF3; ++c) (*smap)[c] = U_Batak;
-  for (char32 c = 0x1BFC; c <= 0x1BFF; ++c) (*smap)[c] = U_Batak;
-  (*smap)[0x11000] = U_Brahmi;
-  (*smap)[0x11001] = U_Brahmi;
-  (*smap)[0x11002] = U_Brahmi;
-  for (char32 c = 0x11003; c <= 0x11037; ++c) (*smap)[c] = U_Brahmi;
-  for (char32 c = 0x11038; c <= 0x11046; ++c) (*smap)[c] = U_Brahmi;
-  for (char32 c = 0x11047; c <= 0x1104D; ++c) (*smap)[c] = U_Brahmi;
-  for (char32 c = 0x11052; c <= 0x11065; ++c) (*smap)[c] = U_Brahmi;
-  for (char32 c = 0x11066; c <= 0x1106F; ++c) (*smap)[c] = U_Brahmi;
-  (*smap)[0x1107F] = U_Brahmi;
-  for (char32 c = 0x0840; c <= 0x0858; ++c) (*smap)[c] = U_Mandaic;
-  for (char32 c = 0x0859; c <= 0x085B; ++c) (*smap)[c] = U_Mandaic;
-  (*smap)[0x085E] = U_Mandaic;
-  for (char32 c = 0x11100; c <= 0x11102; ++c) (*smap)[c] = U_Chakma;
-  for (char32 c = 0x11103; c <= 0x11126; ++c) (*smap)[c] = U_Chakma;
-  for (char32 c = 0x11127; c <= 0x1112B; ++c) (*smap)[c] = U_Chakma;
-  (*smap)[0x1112C] = U_Chakma;
-  for (char32 c = 0x1112D; c <= 0x11134; ++c) (*smap)[c] = U_Chakma;
-  for (char32 c = 0x11136; c <= 0x1113F; ++c) (*smap)[c] = U_Chakma;
-  for (char32 c = 0x11140; c <= 0x11143; ++c) (*smap)[c] = U_Chakma;
-  for (char32 c = 0x109A0; c <= 0x109B7; ++c) (*smap)[c] = U_Meroitic_Cursive;
-  for (char32 c = 0x109BC; c <= 0x109BD; ++c) (*smap)[c] = U_Meroitic_Cursive;
-  for (char32 c = 0x109BE; c <= 0x109BF; ++c) (*smap)[c] = U_Meroitic_Cursive;
-  for (char32 c = 0x109C0; c <= 0x109CF; ++c) (*smap)[c] = U_Meroitic_Cursive;
-  for (char32 c = 0x109D2; c <= 0x109FF; ++c) (*smap)[c] = U_Meroitic_Cursive;
-  for (char32 c = 0x10980; c <= 0x1099F; ++c)
-    (*smap)[c] = U_Meroitic_Hieroglyphs;
-  for (char32 c = 0x16F00; c <= 0x16F44; ++c) (*smap)[c] = U_Miao;
-  (*smap)[0x16F50] = U_Miao;
-  for (char32 c = 0x16F51; c <= 0x16F7E; ++c) (*smap)[c] = U_Miao;
-  for (char32 c = 0x16F8F; c <= 0x16F92; ++c) (*smap)[c] = U_Miao;
-  for (char32 c = 0x16F93; c <= 0x16F9F; ++c) (*smap)[c] = U_Miao;
-  for (char32 c = 0x11180; c <= 0x11181; ++c) (*smap)[c] = U_Sharada;
-  (*smap)[0x11182] = U_Sharada;
-  for (char32 c = 0x11183; c <= 0x111B2; ++c) (*smap)[c] = U_Sharada;
-  for (char32 c = 0x111B3; c <= 0x111B5; ++c) (*smap)[c] = U_Sharada;
-  for (char32 c = 0x111B6; c <= 0x111BE; ++c) (*smap)[c] = U_Sharada;
-  for (char32 c = 0x111BF; c <= 0x111C0; ++c) (*smap)[c] = U_Sharada;
-  for (char32 c = 0x111C1; c <= 0x111C4; ++c) (*smap)[c] = U_Sharada;
-  for (char32 c = 0x111C5; c <= 0x111C9; ++c) (*smap)[c] = U_Sharada;
-  for (char32 c = 0x111CA; c <= 0x111CC; ++c) (*smap)[c] = U_Sharada;
-  (*smap)[0x111CD] = U_Sharada;
-  for (char32 c = 0x111D0; c <= 0x111D9; ++c) (*smap)[c] = U_Sharada;
-  (*smap)[0x111DA] = U_Sharada;
-  (*smap)[0x111DB] = U_Sharada;
-  (*smap)[0x111DC] = U_Sharada;
-  for (char32 c = 0x111DD; c <= 0x111DF; ++c) (*smap)[c] = U_Sharada;
-  for (char32 c = 0x110D0; c <= 0x110E8; ++c) (*smap)[c] = U_Sora_Sompeng;
-  for (char32 c = 0x110F0; c <= 0x110F9; ++c) (*smap)[c] = U_Sora_Sompeng;
-  for (char32 c = 0x11680; c <= 0x116AA; ++c) (*smap)[c] = U_Takri;
-  (*smap)[0x116AB] = U_Takri;
-  (*smap)[0x116AC] = U_Takri;
-  (*smap)[0x116AD] = U_Takri;
-  for (char32 c = 0x116AE; c <= 0x116AF; ++c) (*smap)[c] = U_Takri;
-  for (char32 c = 0x116B0; c <= 0x116B5; ++c) (*smap)[c] = U_Takri;
-  (*smap)[0x116B6] = U_Takri;
-  (*smap)[0x116B7] = U_Takri;
-  for (char32 c = 0x116C0; c <= 0x116C9; ++c) (*smap)[c] = U_Takri;
-  for (char32 c = 0x10530; c <= 0x10563; ++c) (*smap)[c] = U_Caucasian_Albanian;
-  (*smap)[0x1056F] = U_Caucasian_Albanian;
-  for (char32 c = 0x16AD0; c <= 0x16AED; ++c) (*smap)[c] = U_Bassa_Vah;
-  for (char32 c = 0x16AF0; c <= 0x16AF4; ++c) (*smap)[c] = U_Bassa_Vah;
-  (*smap)[0x16AF5] = U_Bassa_Vah;
-  for (char32 c = 0x1BC00; c <= 0x1BC6A; ++c) (*smap)[c] = U_Duployan;
-  for (char32 c = 0x1BC70; c <= 0x1BC7C; ++c) (*smap)[c] = U_Duployan;
-  for (char32 c = 0x1BC80; c <= 0x1BC88; ++c) (*smap)[c] = U_Duployan;
-  for (char32 c = 0x1BC90; c <= 0x1BC99; ++c) (*smap)[c] = U_Duployan;
-  (*smap)[0x1BC9C] = U_Duployan;
-  for (char32 c = 0x1BC9D; c <= 0x1BC9E; ++c) (*smap)[c] = U_Duployan;
-  (*smap)[0x1BC9F] = U_Duployan;
-  for (char32 c = 0x10500; c <= 0x10527; ++c) (*smap)[c] = U_Elbasan;
-  for (char32 c = 0x11300; c <= 0x11301; ++c) (*smap)[c] = U_Grantha;
-  for (char32 c = 0x11302; c <= 0x11303; ++c) (*smap)[c] = U_Grantha;
-  for (char32 c = 0x11305; c <= 0x1130C; ++c) (*smap)[c] = U_Grantha;
-  for (char32 c = 0x1130F; c <= 0x11310; ++c) (*smap)[c] = U_Grantha;
-  for (char32 c = 0x11313; c <= 0x11328; ++c) (*smap)[c] = U_Grantha;
-  for (char32 c = 0x1132A; c <= 0x11330; ++c) (*smap)[c] = U_Grantha;
-  for (char32 c = 0x11332; c <= 0x11333; ++c) (*smap)[c] = U_Grantha;
-  for (char32 c = 0x11335; c <= 0x11339; ++c) (*smap)[c] = U_Grantha;
-  (*smap)[0x1133C] = U_Grantha;
-  (*smap)[0x1133D] = U_Grantha;
-  for (char32 c = 0x1133E; c <= 0x1133F; ++c) (*smap)[c] = U_Grantha;
-  (*smap)[0x11340] = U_Grantha;
-  for (char32 c = 0x11341; c <= 0x11344; ++c) (*smap)[c] = U_Grantha;
-  for (char32 c = 0x11347; c <= 0x11348; ++c) (*smap)[c] = U_Grantha;
-  for (char32 c = 0x1134B; c <= 0x1134D; ++c) (*smap)[c] = U_Grantha;
-  (*smap)[0x11350] = U_Grantha;
-  (*smap)[0x11357] = U_Grantha;
-  for (char32 c = 0x1135D; c <= 0x11361; ++c) (*smap)[c] = U_Grantha;
-  for (char32 c = 0x11362; c <= 0x11363; ++c) (*smap)[c] = U_Grantha;
-  for (char32 c = 0x11366; c <= 0x1136C; ++c) (*smap)[c] = U_Grantha;
-  for (char32 c = 0x11370; c <= 0x11374; ++c) (*smap)[c] = U_Grantha;
-  for (char32 c = 0x16B00; c <= 0x16B2F; ++c) (*smap)[c] = U_Pahawh_Hmong;
-  for (char32 c = 0x16B30; c <= 0x16B36; ++c) (*smap)[c] = U_Pahawh_Hmong;
-  for (char32 c = 0x16B37; c <= 0x16B3B; ++c) (*smap)[c] = U_Pahawh_Hmong;
-  for (char32 c = 0x16B3C; c <= 0x16B3F; ++c) (*smap)[c] = U_Pahawh_Hmong;
-  for (char32 c = 0x16B40; c <= 0x16B43; ++c) (*smap)[c] = U_Pahawh_Hmong;
-  (*smap)[0x16B44] = U_Pahawh_Hmong;
-  (*smap)[0x16B45] = U_Pahawh_Hmong;
-  for (char32 c = 0x16B50; c <= 0x16B59; ++c) (*smap)[c] = U_Pahawh_Hmong;
-  for (char32 c = 0x16B5B; c <= 0x16B61; ++c) (*smap)[c] = U_Pahawh_Hmong;
-  for (char32 c = 0x16B63; c <= 0x16B77; ++c) (*smap)[c] = U_Pahawh_Hmong;
-  for (char32 c = 0x16B7D; c <= 0x16B8F; ++c) (*smap)[c] = U_Pahawh_Hmong;
-  for (char32 c = 0x11200; c <= 0x11211; ++c) (*smap)[c] = U_Khojki;
-  for (char32 c = 0x11213; c <= 0x1122B; ++c) (*smap)[c] = U_Khojki;
-  for (char32 c = 0x1122C; c <= 0x1122E; ++c) (*smap)[c] = U_Khojki;
-  for (char32 c = 0x1122F; c <= 0x11231; ++c) (*smap)[c] = U_Khojki;
-  for (char32 c = 0x11232; c <= 0x11233; ++c) (*smap)[c] = U_Khojki;
-  (*smap)[0x11234] = U_Khojki;
-  (*smap)[0x11235] = U_Khojki;
-  for (char32 c = 0x11236; c <= 0x11237; ++c) (*smap)[c] = U_Khojki;
-  for (char32 c = 0x11238; c <= 0x1123D; ++c) (*smap)[c] = U_Khojki;
-  (*smap)[0x1123E] = U_Khojki;
-  for (char32 c = 0x10600; c <= 0x10736; ++c) (*smap)[c] = U_Linear_A;
-  for (char32 c = 0x10740; c <= 0x10755; ++c) (*smap)[c] = U_Linear_A;
-  for (char32 c = 0x10760; c <= 0x10767; ++c) (*smap)[c] = U_Linear_A;
-  for (char32 c = 0x11150; c <= 0x11172; ++c) (*smap)[c] = U_Mahajani;
-  (*smap)[0x11173] = U_Mahajani;
-  for (char32 c = 0x11174; c <= 0x11175; ++c) (*smap)[c] = U_Mahajani;
-  (*smap)[0x11176] = U_Mahajani;
-  for (char32 c = 0x10AC0; c <= 0x10AC7; ++c) (*smap)[c] = U_Manichaean;
-  (*smap)[0x10AC8] = U_Manichaean;
-  for (char32 c = 0x10AC9; c <= 0x10AE4; ++c) (*smap)[c] = U_Manichaean;
-  for (char32 c = 0x10AE5; c <= 0x10AE6; ++c) (*smap)[c] = U_Manichaean;
-  for (char32 c = 0x10AEB; c <= 0x10AEF; ++c) (*smap)[c] = U_Manichaean;
-  for (char32 c = 0x10AF0; c <= 0x10AF6; ++c) (*smap)[c] = U_Manichaean;
-  for (char32 c = 0x1E800; c <= 0x1E8C4; ++c) (*smap)[c] = U_Mende_Kikakui;
-  for (char32 c = 0x1E8C7; c <= 0x1E8CF; ++c) (*smap)[c] = U_Mende_Kikakui;
-  for (char32 c = 0x1E8D0; c <= 0x1E8D6; ++c) (*smap)[c] = U_Mende_Kikakui;
-  for (char32 c = 0x11600; c <= 0x1162F; ++c) (*smap)[c] = U_Modi;
-  for (char32 c = 0x11630; c <= 0x11632; ++c) (*smap)[c] = U_Modi;
-  for (char32 c = 0x11633; c <= 0x1163A; ++c) (*smap)[c] = U_Modi;
-  for (char32 c = 0x1163B; c <= 0x1163C; ++c) (*smap)[c] = U_Modi;
-  (*smap)[0x1163D] = U_Modi;
-  (*smap)[0x1163E] = U_Modi;
-  for (char32 c = 0x1163F; c <= 0x11640; ++c) (*smap)[c] = U_Modi;
-  for (char32 c = 0x11641; c <= 0x11643; ++c) (*smap)[c] = U_Modi;
-  (*smap)[0x11644] = U_Modi;
-  for (char32 c = 0x11650; c <= 0x11659; ++c) (*smap)[c] = U_Modi;
-  for (char32 c = 0x16A40; c <= 0x16A5E; ++c) (*smap)[c] = U_Mro;
-  for (char32 c = 0x16A60; c <= 0x16A69; ++c) (*smap)[c] = U_Mro;
-  for (char32 c = 0x16A6E; c <= 0x16A6F; ++c) (*smap)[c] = U_Mro;
-  for (char32 c = 0x10A80; c <= 0x10A9C; ++c) (*smap)[c] = U_Old_North_Arabian;
-  for (char32 c = 0x10A9D; c <= 0x10A9F; ++c) (*smap)[c] = U_Old_North_Arabian;
-  for (char32 c = 0x10880; c <= 0x1089E; ++c) (*smap)[c] = U_Nabataean;
-  for (char32 c = 0x108A7; c <= 0x108AF; ++c) (*smap)[c] = U_Nabataean;
-  for (char32 c = 0x10860; c <= 0x10876; ++c) (*smap)[c] = U_Palmyrene;
-  for (char32 c = 0x10877; c <= 0x10878; ++c) (*smap)[c] = U_Palmyrene;
-  for (char32 c = 0x10879; c <= 0x1087F; ++c) (*smap)[c] = U_Palmyrene;
-  for (char32 c = 0x11AC0; c <= 0x11AF8; ++c) (*smap)[c] = U_Pau_Cin_Hau;
-  for (char32 c = 0x10350; c <= 0x10375; ++c) (*smap)[c] = U_Old_Permic;
-  for (char32 c = 0x10376; c <= 0x1037A; ++c) (*smap)[c] = U_Old_Permic;
-  for (char32 c = 0x10B80; c <= 0x10B91; ++c) (*smap)[c] = U_Psalter_Pahlavi;
-  for (char32 c = 0x10B99; c <= 0x10B9C; ++c) (*smap)[c] = U_Psalter_Pahlavi;
-  for (char32 c = 0x10BA9; c <= 0x10BAF; ++c) (*smap)[c] = U_Psalter_Pahlavi;
-  for (char32 c = 0x11580; c <= 0x115AE; ++c) (*smap)[c] = U_Siddham;
-  for (char32 c = 0x115AF; c <= 0x115B1; ++c) (*smap)[c] = U_Siddham;
-  for (char32 c = 0x115B2; c <= 0x115B5; ++c) (*smap)[c] = U_Siddham;
-  for (char32 c = 0x115B8; c <= 0x115BB; ++c) (*smap)[c] = U_Siddham;
-  for (char32 c = 0x115BC; c <= 0x115BD; ++c) (*smap)[c] = U_Siddham;
-  (*smap)[0x115BE] = U_Siddham;
-  for (char32 c = 0x115BF; c <= 0x115C0; ++c) (*smap)[c] = U_Siddham;
-  for (char32 c = 0x115C1; c <= 0x115D7; ++c) (*smap)[c] = U_Siddham;
-  for (char32 c = 0x115D8; c <= 0x115DB; ++c) (*smap)[c] = U_Siddham;
-  for (char32 c = 0x115DC; c <= 0x115DD; ++c) (*smap)[c] = U_Siddham;
-  for (char32 c = 0x112B0; c <= 0x112DE; ++c) (*smap)[c] = U_Khudawadi;
-  (*smap)[0x112DF] = U_Khudawadi;
-  for (char32 c = 0x112E0; c <= 0x112E2; ++c) (*smap)[c] = U_Khudawadi;
-  for (char32 c = 0x112E3; c <= 0x112EA; ++c) (*smap)[c] = U_Khudawadi;
-  for (char32 c = 0x112F0; c <= 0x112F9; ++c) (*smap)[c] = U_Khudawadi;
-  for (char32 c = 0x11480; c <= 0x114AF; ++c) (*smap)[c] = U_Tirhuta;
-  for (char32 c = 0x114B0; c <= 0x114B2; ++c) (*smap)[c] = U_Tirhuta;
-  for (char32 c = 0x114B3; c <= 0x114B8; ++c) (*smap)[c] = U_Tirhuta;
-  (*smap)[0x114B9] = U_Tirhuta;
-  (*smap)[0x114BA] = U_Tirhuta;
-  for (char32 c = 0x114BB; c <= 0x114BE; ++c) (*smap)[c] = U_Tirhuta;
-  for (char32 c = 0x114BF; c <= 0x114C0; ++c) (*smap)[c] = U_Tirhuta;
-  (*smap)[0x114C1] = U_Tirhuta;
-  for (char32 c = 0x114C2; c <= 0x114C3; ++c) (*smap)[c] = U_Tirhuta;
-  for (char32 c = 0x114C4; c <= 0x114C5; ++c) (*smap)[c] = U_Tirhuta;
-  (*smap)[0x114C6] = U_Tirhuta;
-  (*smap)[0x114C7] = U_Tirhuta;
-  for (char32 c = 0x114D0; c <= 0x114D9; ++c) (*smap)[c] = U_Tirhuta;
-  for (char32 c = 0x118A0; c <= 0x118DF; ++c) (*smap)[c] = U_Warang_Citi;
-  for (char32 c = 0x118E0; c <= 0x118E9; ++c) (*smap)[c] = U_Warang_Citi;
-  for (char32 c = 0x118EA; c <= 0x118F2; ++c) (*smap)[c] = U_Warang_Citi;
-  (*smap)[0x118FF] = U_Warang_Citi;
-  for (char32 c = 0x11700; c <= 0x11719; ++c) (*smap)[c] = U_Ahom;
-  for (char32 c = 0x1171D; c <= 0x1171F; ++c) (*smap)[c] = U_Ahom;
-  for (char32 c = 0x11720; c <= 0x11721; ++c) (*smap)[c] = U_Ahom;
-  for (char32 c = 0x11722; c <= 0x11725; ++c) (*smap)[c] = U_Ahom;
-  (*smap)[0x11726] = U_Ahom;
-  for (char32 c = 0x11727; c <= 0x1172B; ++c) (*smap)[c] = U_Ahom;
-  for (char32 c = 0x11730; c <= 0x11739; ++c) (*smap)[c] = U_Ahom;
-  for (char32 c = 0x1173A; c <= 0x1173B; ++c) (*smap)[c] = U_Ahom;
-  for (char32 c = 0x1173C; c <= 0x1173E; ++c) (*smap)[c] = U_Ahom;
-  (*smap)[0x1173F] = U_Ahom;
-  for (char32 c = 0x14400; c <= 0x14646; ++c)
-    (*smap)[c] = U_Anatolian_Hieroglyphs;
-  for (char32 c = 0x108E0; c <= 0x108F2; ++c) (*smap)[c] = U_Hatran;
-  for (char32 c = 0x108F4; c <= 0x108F5; ++c) (*smap)[c] = U_Hatran;
-  for (char32 c = 0x108FB; c <= 0x108FF; ++c) (*smap)[c] = U_Hatran;
-  for (char32 c = 0x11280; c <= 0x11286; ++c) (*smap)[c] = U_Multani;
-  (*smap)[0x11288] = U_Multani;
-  for (char32 c = 0x1128A; c <= 0x1128D; ++c) (*smap)[c] = U_Multani;
-  for (char32 c = 0x1128F; c <= 0x1129D; ++c) (*smap)[c] = U_Multani;
-  for (char32 c = 0x1129F; c <= 0x112A8; ++c) (*smap)[c] = U_Multani;
-  (*smap)[0x112A9] = U_Multani;
-  for (char32 c = 0x10C80; c <= 0x10CB2; ++c) (*smap)[c] = U_Old_Hungarian;
-  for (char32 c = 0x10CC0; c <= 0x10CF2; ++c) (*smap)[c] = U_Old_Hungarian;
-  for (char32 c = 0x10CFA; c <= 0x10CFF; ++c) (*smap)[c] = U_Old_Hungarian;
-  for (char32 c = 0x1D800; c <= 0x1D9FF; ++c) (*smap)[c] = U_SignWriting;
-  for (char32 c = 0x1DA00; c <= 0x1DA36; ++c) (*smap)[c] = U_SignWriting;
-  for (char32 c = 0x1DA37; c <= 0x1DA3A; ++c) (*smap)[c] = U_SignWriting;
-  for (char32 c = 0x1DA3B; c <= 0x1DA6C; ++c) (*smap)[c] = U_SignWriting;
-  for (char32 c = 0x1DA6D; c <= 0x1DA74; ++c) (*smap)[c] = U_SignWriting;
-  (*smap)[0x1DA75] = U_SignWriting;
-  for (char32 c = 0x1DA76; c <= 0x1DA83; ++c) (*smap)[c] = U_SignWriting;
-  (*smap)[0x1DA84] = U_SignWriting;
-  for (char32 c = 0x1DA85; c <= 0x1DA86; ++c) (*smap)[c] = U_SignWriting;
-  for (char32 c = 0x1DA87; c <= 0x1DA8B; ++c) (*smap)[c] = U_SignWriting;
-  for (char32 c = 0x1DA9B; c <= 0x1DA9F; ++c) (*smap)[c] = U_SignWriting;
-  for (char32 c = 0x1DAA1; c <= 0x1DAAF; ++c) (*smap)[c] = U_SignWriting;
-  for (char32 c = 0x1E900; c <= 0x1E943; ++c) (*smap)[c] = U_Adlam;
-  for (char32 c = 0x1E944; c <= 0x1E94A; ++c) (*smap)[c] = U_Adlam;
-  for (char32 c = 0x1E950; c <= 0x1E959; ++c) (*smap)[c] = U_Adlam;
-  for (char32 c = 0x1E95E; c <= 0x1E95F; ++c) (*smap)[c] = U_Adlam;
-  for (char32 c = 0x11C00; c <= 0x11C08; ++c) (*smap)[c] = U_Bhaiksuki;
-  for (char32 c = 0x11C0A; c <= 0x11C2E; ++c) (*smap)[c] = U_Bhaiksuki;
-  (*smap)[0x11C2F] = U_Bhaiksuki;
-  for (char32 c = 0x11C30; c <= 0x11C36; ++c) (*smap)[c] = U_Bhaiksuki;
-  for (char32 c = 0x11C38; c <= 0x11C3D; ++c) (*smap)[c] = U_Bhaiksuki;
-  (*smap)[0x11C3E] = U_Bhaiksuki;
-  (*smap)[0x11C3F] = U_Bhaiksuki;
-  (*smap)[0x11C40] = U_Bhaiksuki;
-  for (char32 c = 0x11C41; c <= 0x11C45; ++c) (*smap)[c] = U_Bhaiksuki;
-  for (char32 c = 0x11C50; c <= 0x11C59; ++c) (*smap)[c] = U_Bhaiksuki;
-  for (char32 c = 0x11C5A; c <= 0x11C6C; ++c) (*smap)[c] = U_Bhaiksuki;
-  for (char32 c = 0x11C70; c <= 0x11C71; ++c) (*smap)[c] = U_Marchen;
-  for (char32 c = 0x11C72; c <= 0x11C8F; ++c) (*smap)[c] = U_Marchen;
-  for (char32 c = 0x11C92; c <= 0x11CA7; ++c) (*smap)[c] = U_Marchen;
-  (*smap)[0x11CA9] = U_Marchen;
-  for (char32 c = 0x11CAA; c <= 0x11CB0; ++c) (*smap)[c] = U_Marchen;
-  (*smap)[0x11CB1] = U_Marchen;
-  for (char32 c = 0x11CB2; c <= 0x11CB3; ++c) (*smap)[c] = U_Marchen;
-  (*smap)[0x11CB4] = U_Marchen;
-  for (char32 c = 0x11CB5; c <= 0x11CB6; ++c) (*smap)[c] = U_Marchen;
-  for (char32 c = 0x11400; c <= 0x11434; ++c) (*smap)[c] = U_Newa;
-  for (char32 c = 0x11435; c <= 0x11437; ++c) (*smap)[c] = U_Newa;
-  for (char32 c = 0x11438; c <= 0x1143F; ++c) (*smap)[c] = U_Newa;
-  for (char32 c = 0x11440; c <= 0x11441; ++c) (*smap)[c] = U_Newa;
-  for (char32 c = 0x11442; c <= 0x11444; ++c) (*smap)[c] = U_Newa;
-  (*smap)[0x11445] = U_Newa;
-  (*smap)[0x11446] = U_Newa;
-  for (char32 c = 0x11447; c <= 0x1144A; ++c) (*smap)[c] = U_Newa;
-  for (char32 c = 0x1144B; c <= 0x1144F; ++c) (*smap)[c] = U_Newa;
-  for (char32 c = 0x11450; c <= 0x11459; ++c) (*smap)[c] = U_Newa;
-  (*smap)[0x1145B] = U_Newa;
-  (*smap)[0x1145D] = U_Newa;
-  for (char32 c = 0x104B0; c <= 0x104D3; ++c) (*smap)[c] = U_Osage;
-  for (char32 c = 0x104D8; c <= 0x104FB; ++c) (*smap)[c] = U_Osage;
-  (*smap)[0x16FE0] = U_Tangut;
-  for (char32 c = 0x17000; c <= 0x187EC; ++c) (*smap)[c] = U_Tangut;
-  for (char32 c = 0x18800; c <= 0x18AF2; ++c) (*smap)[c] = U_Tangut;
-}
-}  // namespace
-}  // namespace unicode_script
-}  // namespace sentencepiece
-#endif  // UNICODE_SCRIPT_DATA_H_
diff --git a/projects/llm_framework/include/sentencepiece/unigram_model.h b/projects/llm_framework/include/sentencepiece/unigram_model.h
deleted file mode 100644
index aa4f28f3..00000000
--- a/projects/llm_framework/include/sentencepiece/unigram_model.h
+++ /dev/null
@@ -1,217 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef UNIGRAM_MODEL_H_
-#define UNIGRAM_MODEL_H_
-
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "common.h"
-#include "freelist.h"
-#include "model_interface.h"
-#include "sentencepiece_model.pb.h"
-#include "third_party/darts_clone/darts.h"
-
-namespace sentencepiece {
-namespace unigram {
-
-// Lattice represents a search space of sentence piece segmentation.
-class Lattice {
- public:
-  Lattice();
-  virtual ~Lattice();
-
-  struct Node {
-    absl::string_view piece;  // Sentence piece representation.
-    uint32 pos;               // Unicode position in the sentence.
-    uint32 length;            // Unicode length, not UT8 byte.
-    uint32 node_id;           // unique id in the current lattice.
-    int id;                   // vocab id. (maybe -1 for UNK)
-    float score;              // logprob of this sentencepiece.
-    float backtrace_score;    // backtrace info used in Viterbi.
-    Node *prev;               // best previous node on Viterbi path.
-
-    std::string DebugString() const;
-  };
-
-  // Returns bos node.
-  Node *bos_node() const;
-
-  // Returns eos node.
-  Node *eos_node() const;
-
-  // Returns nodes starting at |pos|.
-  const std::vector<Node *> &begin_nodes(int pos) const;
-
-  // Returns nodes ending at |pos|.
-  const std::vector<Node *> &end_nodes(int pos) const;
-
-  // Returns Unicode character length.
-  int size() const;
-
-  // Returns multi-byte (utf8) length.
-  int utf8_size() const;
-
-  // Returns the substring of sentence. sentence[pos:]
-  const char *surface(int pos) const;
-
-  // Returns immutable sentence. The same as surface(0)
-  const char *sentence() const;
-
-  // Clears the lattice.
-  void Clear();
-
-  // Sets new sentence.
-  void SetSentence(absl::string_view sentence);
-
-  // Inserts a new node at [pos, pos + length - 1].
-  // After calling this method, The caller must set Node::score and Node::id.
-  Node *Insert(int pos, int length);
-
-  using LatticePathWithScore = std::pair<std::vector<Node *>, float>;
-
-  // Returns Viterbi path. All nodes must be populated in advance.
-  LatticePathWithScore Viterbi();
-
-  // Runs forwards/backwards algorithm, returns vector with normalised
-  // transition probs.
-  std::vector<float> ForwardAlgorithm(float theta) const;
-  std::vector<float> BackwardAlgorithm(float theta) const;
-
-  // Returns n-best results.
-  std::vector<LatticePathWithScore> NBest(size_t nbest_size, bool sample,
-                                          float theta);
-
-  // Samples one path from the lattice according to the
-  // generation probability (Product of piece probabilities).
-  // `theta` is a smoothing parameter.
-  std::vector<Node *> Sample(float theta);
-
-  // Calculates the entropy of the lattice.
-  float CalculateEntropy(float theta) const;
-
-  // Populates marginal probability of every node in this lattice.
-  // |freq| is the frequency of the sentence.
-  //  for (auto *node : all_nodes_) {
-  //    (*expected)[node->id] += marginal_prob_of_node * freq;
-  //  }
-  // Returns the log-likelihood of this sentence.
-  float PopulateMarginal(float freq, std::vector<float> *expected) const;
-
- private:
-  // Returns new node.
-  // Lattice class has the ownership of the returned value.
-  Node *NewNode();
-
-  absl::string_view sentence_;
-  std::vector<const char *> surface_;
-  std::vector<std::vector<Node *>> begin_nodes_;
-  std::vector<std::vector<Node *>> end_nodes_;
-  model::FreeList<Node> node_allocator_;
-};
-
-class Model : public ModelInterface {
- public:
-  explicit Model(const ModelProto &model_proto);
-  Model() {}
-  ~Model() override;
-
-  EncodeResult Encode(absl::string_view normalized) const override;
-
-  NBestEncodeResult NBestEncode(absl::string_view normalized,
-                                int nbest_size) const override;
-
-  EncodeResult SampleEncode(absl::string_view normalized,
-                            float theta) const override;
-
-  NBestEncodeResult SampleEncodeAndScore(absl::string_view normalized,
-                                         float theta, int samples, bool wor,
-                                         bool include_best) const override;
-
-  float CalculateEntropy(absl::string_view normalized,
-                         float theta) const override;
-
-  bool IsSampleEncodeAvailable() const override { return true; }
-
-  bool IsSampleEncodeAndScoreAvailable() const override { return true; }
-
-  bool IsCalculateEntropyAvailable() const override { return true; }
-
-  bool IsNBestEncodeAvailable() const override { return true; }
-
-  // Returns the minimum score in sentence pieces.
-  // min_score() - 10 is used for the cost of unknown sentence.
-  float min_score() const { return min_score_; }
-
-  // Returns the maximum score in sentence pieces.
-  // max_score() is used for the cost of user defined symbols.
-  float max_score() const { return max_score_; }
-
-  // Populates all sentence pieces to the |lattice|.
-  // After calling this function, lattice.Viterbi() returns the
-  // best segmentation.
-  void PopulateNodes(Lattice *lattice) const;
-
-  // Returns a vocab id of |piece|.
-  int PieceToId(absl::string_view piece) const override;
-
-  // Verifies if two outputs are equivalent by comparing their scores.
-  bool VerifyOutputsEquivalent(absl::string_view expected,
-                               absl::string_view actual) const override;
-
-  enum EncoderVersion {
-    kOptimized,  // The optimized encoder.
-    kOriginal    // The original encoder.
-  };
-
-  void SetEncoderVersion(EncoderVersion encoder_version) {
-    encoder_version_ = encoder_version;
-  }
-
-  // Returns the current encoder version in use.
-  EncoderVersion GetEncoderVersion() const { return encoder_version_; }
-
- protected:
-  // Builds a Trie index.
-  void BuildTrie(std::vector<std::pair<absl::string_view, int>> *pieces);
-
-  // The optimized Viterbi encode.
-  // Main differences from the original function:
-  // 1. Memorizes the best path at each postion so far,
-  // 2. No need to store the Lattice nodes,
-  // 3. Works in utf-8 directly,
-  // 4. Defines a new struct with fewer fields than Lattice,
-  // 5. Does not depend on `class Lattice` nor call `SetSentence()`,
-  // `PopulateNodes()`, or `Viterbi()`. It does everything in one function.
-  // For detailed explanations please see the comments inside the function body.
-  EncodeResult EncodeOptimized(absl::string_view normalized) const;
-
-  float min_score_ = 0.0;
-  float max_score_ = 0.0;
-  std::unique_ptr<Darts::DoubleArray> trie_;
-
-  // Maximum size of the return value of Trie, which corresponds
-  // to the maximum size of shared common prefix in the sentence pieces.
-  int trie_results_size_;
-
-  // encoder version.
-  EncoderVersion encoder_version_ = kOptimized;
-};
-
-}  // namespace unigram
-}  // namespace sentencepiece
-#endif  // UNIGRAM_MODEL_H_
diff --git a/projects/llm_framework/include/sentencepiece/unigram_model_trainer.h b/projects/llm_framework/include/sentencepiece/unigram_model_trainer.h
deleted file mode 100644
index c6562e65..00000000
--- a/projects/llm_framework/include/sentencepiece/unigram_model_trainer.h
+++ /dev/null
@@ -1,115 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef UNIGRAM_MODEL_TRAINER_H_
-#define UNIGRAM_MODEL_TRAINER_H_
-
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "sentencepiece_model.pb.h"
-#include "third_party/absl/strings/string_view.h"
-#include "trainer_interface.h"
-#include "unigram_model.h"
-#include "util.h"
-
-namespace sentencepiece {
-namespace unigram {
-
-using string_util::UnicodeText;
-
-class TrainerModel : public Model {
- public:
-  using SentencePieces = std::vector<std::pair<std::string, float>>;
-
-  TrainerModel() {}
-  TrainerModel(const ModelProto &model_proto) = delete;
-  TrainerModel(const TrainerSpec &trainer_spec,
-               const NormalizerSpec &normalizaiton_spec);
-  ~TrainerModel() override;
-
-  // Returns the sentencepieces.
-  // The meta symbols, e.g., </s> are NOT included.
-  const SentencePieces &GetSentencePieces() const;
-
-  // Sets sentencepieces. The sentencepieces are moved.
-  // The meta symbols, e.g., </s> are NOT included.
-  void SetSentencePieces(SentencePieces &&sentencepieces);
-
-  EncodeResult Encode(absl::string_view normalized) const override {
-    return {};
-  }
-
- private:
-  SentencePieces sentencepieces_;
-  TrainerSpec trainer_spec_;
-  NormalizerSpec normalizer_spec_;
-  ModelProto model_proto_data_;
-};
-
-class Trainer : public TrainerInterface {
- public:
-  Trainer(const TrainerSpec &trainer_spec,
-          const NormalizerSpec &normalizer_spec,
-          const NormalizerSpec &denormalizer_spec)
-      : TrainerInterface::TrainerInterface(trainer_spec, normalizer_spec,
-                                           denormalizer_spec) {}
-
-  TrainerModel::SentencePieces MakeSeedSentencePieces();
-
-  util::Status Train() override;
-
- private:
-  FRIEND_TEST(TrainerTest, IsValidSentencePieceTest);
-
-  // Makes seed pieces from the training corpus.
-  // The size of seed pieces is determined by seed_sentencepiece_size.
-  // node_int_type should be of integer type (int32 or int64),
-  // determined by train_extremely_large_corpus.
-  template <typename node_int_type>
-  TrainerModel::SentencePieces MakeSeedSentencePiecesInternal();
-
-  // Executes the E step of EM and returns expected count.
-  // The index of return array is the vocab id.
-  // |objective| is a negative likelihood of the current model.
-  // |num_token| is the number of total tokens to tokenize
-  // training corpus.
-  std::vector<float> RunEStep(const TrainerModel &model, float *objective,
-                              int64 *num_tokens) const;
-
-  // Executes the M step of EM with the expected frequency and
-  // returns new pieces.
-  TrainerModel::SentencePieces RunMStep(
-      const TrainerModel &model, const std::vector<float> &expected) const;
-
-  // Heuristically prunes the current pieces.
-  // This is called after each EM sub-iteration.
-  TrainerModel::SentencePieces PruneSentencePieces(
-      const TrainerModel &model) const;
-
-  // Makes the final sentence pieces by incorporating the required characters
-  // and control/user defined symbols.
-  TrainerModel::SentencePieces FinalizeSentencePieces(
-      const TrainerModel &model) const;
-
-  // When the size of SentencePieces becomes less than desired_vocab_size_,
-  // break the main training loop. desired_vocab_size_ = 1.1 * vocab_size_
-  // for now.
-  int desired_vocab_size_;
-};
-}  // namespace unigram
-}  // namespace sentencepiece
-#endif  // UNIGRAM_MODEL_TRAINER_H_
diff --git a/projects/llm_framework/include/sentencepiece/util.h b/projects/llm_framework/include/sentencepiece/util.h
deleted file mode 100644
index cd843275..00000000
--- a/projects/llm_framework/include/sentencepiece/util.h
+++ /dev/null
@@ -1,442 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef UTIL_H_
-#define UTIL_H_
-
-#include <stdio.h>
-#include <string.h>
-
-#include <algorithm>
-#include <functional>
-#include <memory>
-#include <random>
-#include <sstream>
-#include <string>
-#include <thread>
-#include <utility>
-#include <vector>
-
-#include "common.h"
-#include "sentencepiece_processor.h"
-#include "third_party/absl/strings/string_view.h"
-
-#ifdef SPM_NO_THREADLOCAL
-#include <pthread.h>
-#endif
-
-namespace sentencepiece {
-template <typename T>
-std::ostream &operator<<(std::ostream &out, const std::vector<T> &v) {
-  for (const auto n : v) {
-    out << " " << n;
-  }
-  return out;
-}
-
-uint32 GetRandomGeneratorSeed();
-
-// String utilities
-namespace string_util {
-
-template <typename Target>
-inline bool lexical_cast(absl::string_view arg, Target *result) {
-  std::stringstream ss;
-  return (ss << arg.data() && ss >> *result);
-}
-
-template <>
-inline bool lexical_cast(absl::string_view arg, bool *result) {
-  const char *kTrue[] = {"1", "t", "true", "y", "yes"};
-  const char *kFalse[] = {"0", "f", "false", "n", "no"};
-  std::string lower_value = std::string(arg);
-  std::transform(lower_value.begin(), lower_value.end(), lower_value.begin(),
-                 ::tolower);
-  for (size_t i = 0; i < 5; ++i) {
-    if (lower_value == kTrue[i]) {
-      *result = true;
-      return true;
-    } else if (lower_value == kFalse[i]) {
-      *result = false;
-      return true;
-    }
-  }
-
-  return false;
-}
-
-template <>
-inline bool lexical_cast(absl::string_view arg, std::string *result) {
-  *result = std::string(arg);
-  return true;
-}
-
-template <typename T>
-inline bool DecodePOD(absl::string_view str, T *result) {
-  if (sizeof(*result) != str.size()) {
-    return false;
-  }
-  memcpy(result, str.data(), sizeof(T));
-  return true;
-}
-
-template <typename T>
-inline std::string EncodePOD(const T &value) {
-  std::string s;
-  s.resize(sizeof(T));
-  memcpy(const_cast<char *>(s.data()), &value, sizeof(T));
-  return s;
-}
-
-template <typename T>
-inline std::string IntToHex(T value) {
-  std::ostringstream os;
-  os << std::hex << std::uppercase << value;
-  return os.str();
-}
-
-template <typename T>
-inline T HexToInt(absl::string_view value) {
-  T n;
-  std::istringstream is(value.data());
-  is >> std::hex >> n;
-  return n;
-}
-
-template <typename T>
-inline size_t Itoa(T val, char *s) {
-  char *org = s;
-
-  if (val < 0) {
-    *s++ = '-';
-    val = -val;
-  }
-  char *t = s;
-
-  T mod = 0;
-  while (val) {
-    mod = val % 10;
-    *t++ = static_cast<char>(mod) + '0';
-    val /= 10;
-  }
-
-  if (s == t) {
-    *t++ = '0';
-  }
-
-  *t = '\0';
-  std::reverse(s, t);
-  return static_cast<size_t>(t - org);
-}
-
-template <typename T>
-std::string SimpleItoa(T val) {
-  char buf[32];
-  Itoa<T>(val, buf);
-  return std::string(buf);
-}
-
-// Return length of a single UTF-8 source character
-inline size_t OneCharLen(const char *src) {
-  return "\1\1\1\1\1\1\1\1\1\1\1\1\2\2\3\4"[(*src & 0xFF) >> 4];
-}
-
-// Return (x & 0xC0) == 0x80;
-// Since trail bytes are always in [0x80, 0xBF], we can optimize:
-inline bool IsTrailByte(char x) { return static_cast<signed char>(x) < -0x40; }
-
-inline bool IsValidCodepoint(char32 c) {
-  return (static_cast<uint32>(c) < 0xD800) || (c >= 0xE000 && c <= 0x10FFFF);
-}
-
-bool IsStructurallyValid(absl::string_view str);
-
-using UnicodeText = std::vector<char32>;
-
-char32 DecodeUTF8(const char *begin, const char *end, size_t *mblen);
-
-inline char32 DecodeUTF8(absl::string_view input, size_t *mblen) {
-  return DecodeUTF8(input.data(), input.data() + input.size(), mblen);
-}
-
-inline bool IsValidDecodeUTF8(absl::string_view input, size_t *mblen) {
-  const char32 c = DecodeUTF8(input, mblen);
-  return c != kUnicodeError || *mblen == 3;
-}
-
-size_t EncodeUTF8(char32 c, char *output);
-
-std::string UnicodeCharToUTF8(const char32 c);
-
-UnicodeText UTF8ToUnicodeText(absl::string_view utf8);
-
-std::string UnicodeTextToUTF8(const UnicodeText &utext);
-
-}  // namespace string_util
-
-// other map/ptr utilties
-namespace port {
-
-template <class Collection, class Key>
-bool ContainsKey(const Collection &collection, const Key &key) {
-  return collection.find(key) != collection.end();
-}
-
-template <class Collection>
-const typename Collection::value_type::second_type &FindOrDie(
-    const Collection &collection,
-    const typename Collection::value_type::first_type &key) {
-  typename Collection::const_iterator it = collection.find(key);
-  CHECK(it != collection.end()) << "Map key not found: " << key;
-  return it->second;
-}
-
-template <class Collection>
-const typename Collection::value_type::second_type &FindWithDefault(
-    const Collection &collection,
-    const typename Collection::value_type::first_type &key,
-    const typename Collection::value_type::second_type &value) {
-  typename Collection::const_iterator it = collection.find(key);
-  if (it == collection.end()) {
-    return value;
-  }
-  return it->second;
-}
-
-template <class Collection>
-bool InsertIfNotPresent(Collection *const collection,
-                        const typename Collection::value_type &vt) {
-  return collection->insert(vt).second;
-}
-
-template <class Collection>
-bool InsertIfNotPresent(
-    Collection *const collection,
-    const typename Collection::value_type::first_type &key,
-    const typename Collection::value_type::second_type &value) {
-  return InsertIfNotPresent(collection,
-                            typename Collection::value_type(key, value));
-}
-
-template <class Collection>
-void InsertOrDie(Collection *const collection,
-                 const typename Collection::value_type::first_type &key,
-                 const typename Collection::value_type::second_type &data) {
-  CHECK(InsertIfNotPresent(collection, key, data)) << "duplicate key";
-}
-
-// hash
-inline void mix(uint64 &a, uint64 &b, uint64 &c) {  // 64bit version
-  a -= b;
-  a -= c;
-  a ^= (c >> 43);
-  b -= c;
-  b -= a;
-  b ^= (a << 9);
-  c -= a;
-  c -= b;
-  c ^= (b >> 8);
-  a -= b;
-  a -= c;
-  a ^= (c >> 38);
-  b -= c;
-  b -= a;
-  b ^= (a << 23);
-  c -= a;
-  c -= b;
-  c ^= (b >> 5);
-  a -= b;
-  a -= c;
-  a ^= (c >> 35);
-  b -= c;
-  b -= a;
-  b ^= (a << 49);
-  c -= a;
-  c -= b;
-  c ^= (b >> 11);
-  a -= b;
-  a -= c;
-  a ^= (c >> 12);
-  b -= c;
-  b -= a;
-  b ^= (a << 18);
-  c -= a;
-  c -= b;
-  c ^= (b >> 22);
-}
-
-inline uint64 FingerprintCat(uint64 x, uint64 y) {
-  uint64 b = 0xe08c1d668b756f82;  // more of the golden ratio
-  mix(x, b, y);
-  return y;
-}
-
-}  // namespace port
-
-namespace random {
-
-std::mt19937 *GetRandomGenerator();
-
-template <typename T>
-class ReservoirSampler {
- public:
-  explicit ReservoirSampler(std::vector<T> *sampled, uint64 size)
-      : sampled_(sampled), size_(size), engine_(GetRandomGeneratorSeed()) {}
-  explicit ReservoirSampler(std::vector<T> *sampled, uint64 size, uint64 seed)
-      : sampled_(sampled), size_(size), engine_(seed) {}
-  virtual ~ReservoirSampler() {}
-
-  void Add(const T &item) {
-    if (size_ == 0) return;
-
-    ++total_;
-    if (sampled_->size() < size_) {
-      sampled_->push_back(item);
-    } else {
-      std::uniform_int_distribution<uint64> dist(0, total_ - 1);
-      const uint64 n = dist(engine_);
-      if (n < sampled_->size()) (*sampled_)[n] = item;
-    }
-  }
-
-  uint64 total_size() const { return total_; }
-
- private:
-  std::vector<T> *sampled_ = nullptr;
-  uint64 size_ = 0;
-  uint64 total_ = 0;
-  std::mt19937 engine_;
-};
-
-}  // namespace random
-
-namespace util {
-
-inline std::string JoinPath(absl::string_view path) {
-  return std::string(path.data(), path.size());
-}
-
-template <typename... T>
-inline std::string JoinPath(absl::string_view first, const T &...rest) {
-#ifdef OS_WIN
-  return JoinPath(first) + "\\" + JoinPath(rest...);
-#else
-  return JoinPath(first) + "/" + JoinPath(rest...);
-#endif
-}
-
-std::string StrError(int errnum);
-
-std::vector<std::string> StrSplitAsCSV(absl::string_view text);
-
-#ifdef OS_WIN
-std::wstring Utf8ToWide(const absl::string_view input);
-#endif
-
-inline Status OkStatus() { return Status(); }
-
-#define DECLARE_ERROR(FUNC)                                \
-  inline util::Status FUNC##Error(absl::string_view str) { \
-    return util::Status(StatusCode::k##FUNC, str.data());  \
-  }                                                        \
-  inline bool Is##FUNC(const util::Status &status) {       \
-    return status.code() == StatusCode::k##FUNC;           \
-  }
-
-DECLARE_ERROR(Cancelled)
-DECLARE_ERROR(InvalidArgument)
-DECLARE_ERROR(NotFound)
-DECLARE_ERROR(AlreadyExists)
-DECLARE_ERROR(ResourceExhausted)
-DECLARE_ERROR(Unavailable)
-DECLARE_ERROR(FailedPrecondition)
-DECLARE_ERROR(OutOfRange)
-DECLARE_ERROR(Unimplemented)
-DECLARE_ERROR(Internal)
-DECLARE_ERROR(Aborted)
-DECLARE_ERROR(DeadlineExceeded)
-DECLARE_ERROR(DataLoss)
-DECLARE_ERROR(Unknown)
-DECLARE_ERROR(PermissionDenied)
-DECLARE_ERROR(Unauthenticated)
-
-#define GTL_LOC (0)
-
-class StatusBuilder {
- public:
-  explicit StatusBuilder(StatusCode code) : code_(code) {}
-  explicit StatusBuilder(StatusCode code, int loc) : code_(code) {}
-
-  template <typename T>
-  StatusBuilder &operator<<(const T &value) {
-    os_ << value;
-    return *this;
-  }
-
-  operator Status() const { return Status(code_, os_.str()); }
-
- private:
-  StatusCode code_;
-  std::ostringstream os_;
-};
-
-#define CHECK_OR_RETURN(condition)                           \
-  if (condition) {                                           \
-  } else /* NOLINT */                                        \
-    return ::sentencepiece::util::StatusBuilder(             \
-               ::sentencepiece::util::StatusCode::kInternal) \
-           << __FILE__ << "(" << __LINE__ << ") [" << #condition << "] "
-
-#define CHECK_EQ_OR_RETURN(a, b) CHECK_OR_RETURN((a) == (b))
-#define CHECK_NE_OR_RETURN(a, b) CHECK_OR_RETURN((a) != (b))
-#define CHECK_GE_OR_RETURN(a, b) CHECK_OR_RETURN((a) >= (b))
-#define CHECK_LE_OR_RETURN(a, b) CHECK_OR_RETURN((a) <= (b))
-#define CHECK_GT_OR_RETURN(a, b) CHECK_OR_RETURN((a) > (b))
-#define CHECK_LT_OR_RETURN(a, b) CHECK_OR_RETURN((a) < (b))
-
-}  // namespace util
-
-namespace port {
-template <typename T>
-void STLDeleteElements(std::vector<T *> *vec) {
-  for (auto item : *vec) {
-    delete item;
-  }
-  vec->clear();
-}
-}  // namespace port
-
-class ThreadPool {
- public:
-  ThreadPool(int32 n) {}
-  virtual ~ThreadPool() {
-    for (auto &task : tasks_) {
-      task.join();
-    }
-  }
-
-  void Schedule(std::function<void()> closure) { tasks_.emplace_back(closure); }
-  void StartWorkers() {}
-
- private:
-  std::vector<std::thread> tasks_;
-};
-
-namespace log_domain {
-
-double LogSum(const std::vector<double> &xs);
-
-}  // namespace log_domain
-}  // namespace sentencepiece
-#endif  // UTIL_H_
diff --git a/projects/llm_framework/include/sentencepiece/word_model.h b/projects/llm_framework/include/sentencepiece/word_model.h
deleted file mode 100644
index 34470f96..00000000
--- a/projects/llm_framework/include/sentencepiece/word_model.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef WORD_MODEL_H_
-#define WORD_MODEL_H_
-
-#include "model_interface.h"
-#include "sentencepiece_model.pb.h"
-
-namespace sentencepiece {
-namespace word {
-
-// Tokenize text with whitespaces.
-class Model : public ModelInterface {
- public:
-  explicit Model(const ModelProto &model_proto);
-  ~Model() override;
-
-  EncodeResult Encode(absl::string_view normalized) const override;
-};
-}  // namespace word
-}  // namespace sentencepiece
-#endif  // WORD_MODEL_H_
diff --git a/projects/llm_framework/include/sentencepiece/word_model_trainer.h b/projects/llm_framework/include/sentencepiece/word_model_trainer.h
deleted file mode 100644
index 76f8f32a..00000000
--- a/projects/llm_framework/include/sentencepiece/word_model_trainer.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Copyright 2016 Google Inc.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.!
-
-#ifndef WORD_MODEL_TRAINER_H_
-#define WORD_MODEL_TRAINER_H_
-
-#include "sentencepiece_model.pb.h"
-#include "trainer_interface.h"
-
-namespace sentencepiece {
-namespace word {
-
-// Trainer class for word model.
-//
-// Word model simply counts the frequency of
-// space-delimited tokens, then keep top
-// |vocab_size| frequent tokens.
-class Trainer : public TrainerInterface {
- public:
-  Trainer(const TrainerSpec &trainer_spec,
-          const NormalizerSpec &normalizer_spec,
-          const NormalizerSpec &denormalizer_spec)
-      : TrainerInterface::TrainerInterface(trainer_spec, normalizer_spec,
-                                           denormalizer_spec) {}
-
-  util::Status Train() override;
-};
-}  // namespace word
-}  // namespace sentencepiece
-#endif  // WORD_MODEL_TRAINER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/allocator.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/allocator.h
deleted file mode 100644
index e3e987ed..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/allocator.h
+++ /dev/null
@@ -1,444 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2018 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_ALLOCATOR_H
-#define NCNN_ALLOCATOR_H
-
-#ifdef _WIN32
-#define WIN32_LEAN_AND_MEAN
-#include <windows.h>
-#endif
-
-#include "platform.h"
-
-#include <stdlib.h>
-
-#if NCNN_PLATFORM_API
-#if __ANDROID_API__ >= 26
-#include <android/hardware_buffer.h>
-#endif // __ANDROID_API__ >= 26
-#endif // NCNN_PLATFORM_API
-
-namespace ncnn {
-
-// the alignment of all the allocated buffers
-#if NCNN_AVX512
-#define NCNN_MALLOC_ALIGN 64
-#elif NCNN_AVX
-#define NCNN_MALLOC_ALIGN 32
-#else
-#define NCNN_MALLOC_ALIGN 16
-#endif
-
-// we have some optimized kernels that may overread buffer a bit in loop
-// it is common to interleave next-loop data load with arithmetic instructions
-// allocating more bytes keeps us safe from SEGV_ACCERR failure
-#define NCNN_MALLOC_OVERREAD 64
-
-// Aligns a pointer to the specified number of bytes
-// ptr Aligned pointer
-// n Alignment size that must be a power of two
-template<typename _Tp>
-static NCNN_FORCEINLINE _Tp* alignPtr(_Tp* ptr, int n = (int)sizeof(_Tp))
-{
-    return (_Tp*)(((size_t)ptr + n - 1) & -n);
-}
-
-// Aligns a buffer size to the specified number of bytes
-// The function returns the minimum number that is greater or equal to sz and is divisible by n
-// sz Buffer size to align
-// n Alignment size that must be a power of two
-static NCNN_FORCEINLINE size_t alignSize(size_t sz, int n)
-{
-    return (sz + n - 1) & -n;
-}
-
-static NCNN_FORCEINLINE void* fastMalloc(size_t size)
-{
-#if _MSC_VER
-    return _aligned_malloc(size, NCNN_MALLOC_ALIGN);
-#elif (defined(__unix__) || defined(__APPLE__)) && _POSIX_C_SOURCE >= 200112L || (__ANDROID__ && __ANDROID_API__ >= 17)
-    void* ptr = 0;
-    if (posix_memalign(&ptr, NCNN_MALLOC_ALIGN, size + NCNN_MALLOC_OVERREAD))
-        ptr = 0;
-    return ptr;
-#elif __ANDROID__ && __ANDROID_API__ < 17
-    return memalign(NCNN_MALLOC_ALIGN, size + NCNN_MALLOC_OVERREAD);
-#else
-    unsigned char* udata = (unsigned char*)malloc(size + sizeof(void*) + NCNN_MALLOC_ALIGN + NCNN_MALLOC_OVERREAD);
-    if (!udata)
-        return 0;
-    unsigned char** adata = alignPtr((unsigned char**)udata + 1, NCNN_MALLOC_ALIGN);
-    adata[-1] = udata;
-    return adata;
-#endif
-}
-
-static NCNN_FORCEINLINE void fastFree(void* ptr)
-{
-    if (ptr)
-    {
-#if _MSC_VER
-        _aligned_free(ptr);
-#elif (defined(__unix__) || defined(__APPLE__)) && _POSIX_C_SOURCE >= 200112L || (__ANDROID__ && __ANDROID_API__ >= 17)
-        free(ptr);
-#elif __ANDROID__ && __ANDROID_API__ < 17
-        free(ptr);
-#else
-        unsigned char* udata = ((unsigned char**)ptr)[-1];
-        free(udata);
-#endif
-    }
-}
-
-#if NCNN_THREADS
-// exchange-add operation for atomic operations on reference counters
-#if defined __riscv && !defined __riscv_atomic
-// riscv target without A extension
-static NCNN_FORCEINLINE int NCNN_XADD(int* addr, int delta)
-{
-    int tmp = *addr;
-    *addr += delta;
-    return tmp;
-}
-#elif defined __INTEL_COMPILER && !(defined WIN32 || defined _WIN32)
-// atomic increment on the linux version of the Intel(tm) compiler
-#define NCNN_XADD(addr, delta) (int)_InterlockedExchangeAdd(const_cast<void*>(reinterpret_cast<volatile void*>(addr)), delta)
-#elif defined __GNUC__
-#if defined __clang__ && __clang_major__ >= 3 && !defined __ANDROID__ && !defined __EMSCRIPTEN__ && !defined(__CUDACC__)
-#ifdef __ATOMIC_ACQ_REL
-#define NCNN_XADD(addr, delta) __c11_atomic_fetch_add((_Atomic(int)*)(addr), delta, __ATOMIC_ACQ_REL)
-#else
-#define NCNN_XADD(addr, delta) __atomic_fetch_add((_Atomic(int)*)(addr), delta, 4)
-#endif
-#else
-#if defined __ATOMIC_ACQ_REL && !defined __clang__
-// version for gcc >= 4.7
-#define NCNN_XADD(addr, delta) (int)__atomic_fetch_add((unsigned*)(addr), (unsigned)(delta), __ATOMIC_ACQ_REL)
-#else
-#define NCNN_XADD(addr, delta) (int)__sync_fetch_and_add((unsigned*)(addr), (unsigned)(delta))
-#endif
-#endif
-#elif defined _MSC_VER && !defined RC_INVOKED
-#define NCNN_XADD(addr, delta) (int)_InterlockedExchangeAdd((long volatile*)addr, delta)
-#else
-// thread-unsafe branch
-static NCNN_FORCEINLINE int NCNN_XADD(int* addr, int delta)
-{
-    int tmp = *addr;
-    *addr += delta;
-    return tmp;
-}
-#endif
-#else  // NCNN_THREADS
-static NCNN_FORCEINLINE int NCNN_XADD(int* addr, int delta)
-{
-    int tmp = *addr;
-    *addr += delta;
-    return tmp;
-}
-#endif // NCNN_THREADS
-
-class NCNN_EXPORT Allocator
-{
-public:
-    virtual ~Allocator();
-    virtual void* fastMalloc(size_t size) = 0;
-    virtual void fastFree(void* ptr) = 0;
-};
-
-class PoolAllocatorPrivate;
-class NCNN_EXPORT PoolAllocator : public Allocator
-{
-public:
-    PoolAllocator();
-    ~PoolAllocator();
-
-    // ratio range 0 ~ 1
-    // default cr = 0
-    void set_size_compare_ratio(float scr);
-
-    // budget drop threshold
-    // default threshold = 10
-    void set_size_drop_threshold(size_t);
-
-    // release all budgets immediately
-    void clear();
-
-    virtual void* fastMalloc(size_t size);
-    virtual void fastFree(void* ptr);
-
-private:
-    PoolAllocator(const PoolAllocator&);
-    PoolAllocator& operator=(const PoolAllocator&);
-
-private:
-    PoolAllocatorPrivate* const d;
-};
-
-class UnlockedPoolAllocatorPrivate;
-class NCNN_EXPORT UnlockedPoolAllocator : public Allocator
-{
-public:
-    UnlockedPoolAllocator();
-    ~UnlockedPoolAllocator();
-
-    // ratio range 0 ~ 1
-    // default cr = 0
-    void set_size_compare_ratio(float scr);
-
-    // budget drop threshold
-    // default threshold = 10
-    void set_size_drop_threshold(size_t);
-
-    // release all budgets immediately
-    void clear();
-
-    virtual void* fastMalloc(size_t size);
-    virtual void fastFree(void* ptr);
-
-private:
-    UnlockedPoolAllocator(const UnlockedPoolAllocator&);
-    UnlockedPoolAllocator& operator=(const UnlockedPoolAllocator&);
-
-private:
-    UnlockedPoolAllocatorPrivate* const d;
-};
-
-#if NCNN_VULKAN
-
-class VulkanDevice;
-
-class NCNN_EXPORT VkBufferMemory
-{
-public:
-    VkBuffer buffer;
-
-    // the base offset assigned by allocator
-    size_t offset;
-    size_t capacity;
-
-    VkDeviceMemory memory;
-    void* mapped_ptr;
-
-    // buffer state, modified by command functions internally
-    mutable VkAccessFlags access_flags;
-    mutable VkPipelineStageFlags stage_flags;
-
-    // initialize and modified by mat
-    int refcount;
-};
-
-class NCNN_EXPORT VkImageMemory
-{
-public:
-    VkImage image;
-    VkImageView imageview;
-
-    // underlying info assigned by allocator
-    int width;
-    int height;
-    int depth;
-    VkFormat format;
-
-    VkDeviceMemory memory;
-    void* mapped_ptr;
-
-    // the base offset assigned by allocator
-    size_t bind_offset;
-    size_t bind_capacity;
-
-    // image state, modified by command functions internally
-    mutable VkAccessFlags access_flags;
-    mutable VkImageLayout image_layout;
-    mutable VkPipelineStageFlags stage_flags;
-
-    // in-execution state, modified by command functions internally
-    mutable int command_refcount;
-
-    // initialize and modified by mat
-    int refcount;
-};
-
-class NCNN_EXPORT VkAllocator
-{
-public:
-    explicit VkAllocator(const VulkanDevice* _vkdev);
-    virtual ~VkAllocator();
-
-    virtual void clear();
-
-    virtual VkBufferMemory* fastMalloc(size_t size) = 0;
-    virtual void fastFree(VkBufferMemory* ptr) = 0;
-    virtual int flush(VkBufferMemory* ptr);
-    virtual int invalidate(VkBufferMemory* ptr);
-
-    virtual VkImageMemory* fastMalloc(int w, int h, int c, size_t elemsize, int elempack) = 0;
-    virtual void fastFree(VkImageMemory* ptr) = 0;
-
-public:
-    const VulkanDevice* vkdev;
-    uint32_t buffer_memory_type_index;
-    uint32_t image_memory_type_index;
-    uint32_t reserved_type_index;
-    bool mappable;
-    bool coherent;
-
-protected:
-    VkBuffer create_buffer(size_t size, VkBufferUsageFlags usage);
-    VkDeviceMemory allocate_memory(size_t size, uint32_t memory_type_index);
-    VkDeviceMemory allocate_dedicated_memory(size_t size, uint32_t memory_type_index, VkImage image, VkBuffer buffer);
-
-    VkImage create_image(int width, int height, int depth, VkFormat format, VkImageTiling tiling, VkImageUsageFlags usage);
-    VkImageView create_imageview(VkImage image, VkFormat format);
-};
-
-class VkBlobAllocatorPrivate;
-class NCNN_EXPORT VkBlobAllocator : public VkAllocator
-{
-public:
-    explicit VkBlobAllocator(const VulkanDevice* vkdev, size_t preferred_block_size = 16 * 1024 * 1024); // 16M
-    virtual ~VkBlobAllocator();
-
-public:
-    // release all budgets immediately
-    virtual void clear();
-
-    virtual VkBufferMemory* fastMalloc(size_t size);
-    virtual void fastFree(VkBufferMemory* ptr);
-    virtual VkImageMemory* fastMalloc(int w, int h, int c, size_t elemsize, int elempack);
-    virtual void fastFree(VkImageMemory* ptr);
-
-private:
-    VkBlobAllocator(const VkBlobAllocator&);
-    VkBlobAllocator& operator=(const VkBlobAllocator&);
-
-private:
-    VkBlobAllocatorPrivate* const d;
-};
-
-class VkWeightAllocatorPrivate;
-class NCNN_EXPORT VkWeightAllocator : public VkAllocator
-{
-public:
-    explicit VkWeightAllocator(const VulkanDevice* vkdev, size_t preferred_block_size = 8 * 1024 * 1024); // 8M
-    virtual ~VkWeightAllocator();
-
-public:
-    // release all blocks immediately
-    virtual void clear();
-
-public:
-    virtual VkBufferMemory* fastMalloc(size_t size);
-    virtual void fastFree(VkBufferMemory* ptr);
-    virtual VkImageMemory* fastMalloc(int w, int h, int c, size_t elemsize, int elempack);
-    virtual void fastFree(VkImageMemory* ptr);
-
-private:
-    VkWeightAllocator(const VkWeightAllocator&);
-    VkWeightAllocator& operator=(const VkWeightAllocator&);
-
-private:
-    VkWeightAllocatorPrivate* const d;
-};
-
-class VkStagingAllocatorPrivate;
-class NCNN_EXPORT VkStagingAllocator : public VkAllocator
-{
-public:
-    explicit VkStagingAllocator(const VulkanDevice* vkdev);
-    virtual ~VkStagingAllocator();
-
-public:
-    // ratio range 0 ~ 1
-    // default cr = 0.75
-    void set_size_compare_ratio(float scr);
-
-    // release all budgets immediately
-    virtual void clear();
-
-    virtual VkBufferMemory* fastMalloc(size_t size);
-    virtual void fastFree(VkBufferMemory* ptr);
-    virtual VkImageMemory* fastMalloc(int w, int h, int c, size_t elemsize, int elempack);
-    virtual void fastFree(VkImageMemory* ptr);
-
-private:
-    VkStagingAllocator(const VkStagingAllocator&);
-    VkStagingAllocator& operator=(const VkStagingAllocator&);
-
-private:
-    VkStagingAllocatorPrivate* const d;
-};
-
-class VkWeightStagingAllocatorPrivate;
-class NCNN_EXPORT VkWeightStagingAllocator : public VkAllocator
-{
-public:
-    explicit VkWeightStagingAllocator(const VulkanDevice* vkdev);
-    virtual ~VkWeightStagingAllocator();
-
-public:
-    virtual VkBufferMemory* fastMalloc(size_t size);
-    virtual void fastFree(VkBufferMemory* ptr);
-    virtual VkImageMemory* fastMalloc(int w, int h, int c, size_t elemsize, int elempack);
-    virtual void fastFree(VkImageMemory* ptr);
-
-private:
-    VkWeightStagingAllocator(const VkWeightStagingAllocator&);
-    VkWeightStagingAllocator& operator=(const VkWeightStagingAllocator&);
-
-private:
-    VkWeightStagingAllocatorPrivate* const d;
-};
-
-#if NCNN_PLATFORM_API
-#if __ANDROID_API__ >= 26
-class NCNN_EXPORT VkAndroidHardwareBufferImageAllocator : public VkAllocator
-{
-public:
-    VkAndroidHardwareBufferImageAllocator(const VulkanDevice* _vkdev, AHardwareBuffer* _hb);
-    virtual ~VkAndroidHardwareBufferImageAllocator();
-
-public:
-    virtual VkBufferMemory* fastMalloc(size_t size);
-    virtual void fastFree(VkBufferMemory* ptr);
-    virtual VkImageMemory* fastMalloc(int w, int h, int c, size_t elemsize, int elempack);
-    virtual void fastFree(VkImageMemory* ptr);
-
-private:
-    VkAndroidHardwareBufferImageAllocator(const VkAndroidHardwareBufferImageAllocator&);
-    VkAndroidHardwareBufferImageAllocator& operator=(const VkAndroidHardwareBufferImageAllocator&);
-
-public:
-    int init();
-
-    int width() const;
-    int height() const;
-    uint64_t external_format() const;
-
-public:
-    AHardwareBuffer* hb;
-    AHardwareBuffer_Desc bufferDesc;
-    VkAndroidHardwareBufferFormatPropertiesANDROID bufferFormatProperties;
-    VkAndroidHardwareBufferPropertiesANDROID bufferProperties;
-    VkSamplerYcbcrConversionKHR samplerYcbcrConversion;
-};
-#endif // __ANDROID_API__ >= 26
-#endif // NCNN_PLATFORM_API
-
-#endif // NCNN_VULKAN
-
-} // namespace ncnn
-
-#endif // NCNN_ALLOCATOR_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/benchmark.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/benchmark.h
deleted file mode 100644
index ed42c1ac..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/benchmark.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_BENCHMARK_H
-#define NCNN_BENCHMARK_H
-
-#include "layer.h"
-#include "mat.h"
-#include "platform.h"
-
-namespace ncnn {
-
-// get now timestamp in ms
-NCNN_EXPORT double get_current_time();
-
-// sleep milliseconds
-NCNN_EXPORT void sleep(unsigned long long int milliseconds = 1000);
-
-#if NCNN_BENCHMARK
-
-NCNN_EXPORT void benchmark(const Layer* layer, double start, double end);
-NCNN_EXPORT void benchmark(const Layer* layer, const Mat& bottom_blob, Mat& top_blob, double start, double end);
-
-#endif // NCNN_BENCHMARK
-
-} // namespace ncnn
-
-#endif // NCNN_BENCHMARK_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/blob.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/blob.h
deleted file mode 100644
index c9f144fb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/blob.h
+++ /dev/null
@@ -1,44 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_BLOB_H
-#define NCNN_BLOB_H
-
-#include "mat.h"
-#include "platform.h"
-
-namespace ncnn {
-
-class NCNN_EXPORT Blob
-{
-public:
-    // empty
-    Blob();
-
-public:
-#if NCNN_STRING
-    // blob name
-    std::string name;
-#endif // NCNN_STRING
-    // layer index which produce this blob as output
-    int producer;
-    // layer index which need this blob as input
-    int consumer;
-    // shape hint
-    Mat shape;
-};
-
-} // namespace ncnn
-
-#endif // NCNN_BLOB_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/c_api.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/c_api.h
deleted file mode 100644
index ddd8bfff..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/c_api.h
+++ /dev/null
@@ -1,370 +0,0 @@
-/* Tencent is pleased to support the open source community by making ncnn available.
- *
- * Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
- *
- * Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
- * in compliance with the License. You may obtain a copy of the License at
- *
- * https://opensource.org/licenses/BSD-3-Clause
- *
- * Unless required by applicable law or agreed to in writing, software distributed
- * under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
- * CONDITIONS OF ANY KIND, either express or implied. See the License for the
- * specific language governing permissions and limitations under the License.
- */
-
-#ifndef NCNN_C_API_H
-#define NCNN_C_API_H
-
-#include "platform.h"
-
-#if NCNN_C_API
-
-#include <stddef.h>
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-NCNN_EXPORT const char* ncnn_version();
-
-/* allocator api */
-typedef struct __ncnn_allocator_t* ncnn_allocator_t;
-struct NCNN_EXPORT __ncnn_allocator_t
-{
-    void* pthis;
-
-    void* (*fast_malloc)(ncnn_allocator_t allocator, size_t size);
-    void (*fast_free)(ncnn_allocator_t allocator, void* ptr);
-};
-
-NCNN_EXPORT ncnn_allocator_t ncnn_allocator_create_pool_allocator();
-NCNN_EXPORT ncnn_allocator_t ncnn_allocator_create_unlocked_pool_allocator();
-NCNN_EXPORT void ncnn_allocator_destroy(ncnn_allocator_t allocator);
-
-/* option api */
-typedef struct __ncnn_option_t* ncnn_option_t;
-
-NCNN_EXPORT ncnn_option_t ncnn_option_create();
-NCNN_EXPORT void ncnn_option_destroy(ncnn_option_t opt);
-
-NCNN_EXPORT int ncnn_option_get_num_threads(const ncnn_option_t opt);
-NCNN_EXPORT void ncnn_option_set_num_threads(ncnn_option_t opt, int num_threads);
-
-NCNN_EXPORT int ncnn_option_get_use_local_pool_allocator(const ncnn_option_t opt);
-NCNN_EXPORT void ncnn_option_set_use_local_pool_allocator(ncnn_option_t opt, int use_local_pool_allocator);
-
-NCNN_EXPORT void ncnn_option_set_blob_allocator(ncnn_option_t opt, ncnn_allocator_t allocator);
-NCNN_EXPORT void ncnn_option_set_workspace_allocator(ncnn_option_t opt, ncnn_allocator_t allocator);
-
-NCNN_EXPORT int ncnn_option_get_use_vulkan_compute(const ncnn_option_t opt);
-NCNN_EXPORT void ncnn_option_set_use_vulkan_compute(ncnn_option_t opt, int use_vulkan_compute);
-
-/* mat api */
-typedef struct __ncnn_mat_t* ncnn_mat_t;
-
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create();
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_1d(int w, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_2d(int w, int h, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_3d(int w, int h, int c, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_4d(int w, int h, int d, int c, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_external_1d(int w, void* data, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_external_2d(int w, int h, void* data, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_external_3d(int w, int h, int c, void* data, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_external_4d(int w, int h, int d, int c, void* data, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_1d_elem(int w, size_t elemsize, int elempack, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_2d_elem(int w, int h, size_t elemsize, int elempack, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_3d_elem(int w, int h, int c, size_t elemsize, int elempack, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_4d_elem(int w, int h, int d, int c, size_t elemsize, int elempack, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_external_1d_elem(int w, void* data, size_t elemsize, int elempack, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_external_2d_elem(int w, int h, void* data, size_t elemsize, int elempack, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_external_3d_elem(int w, int h, int c, void* data, size_t elemsize, int elempack, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_create_external_4d_elem(int w, int h, int d, int c, void* data, size_t elemsize, int elempack, ncnn_allocator_t allocator);
-NCNN_EXPORT void ncnn_mat_destroy(ncnn_mat_t mat);
-
-NCNN_EXPORT void ncnn_mat_fill_float(ncnn_mat_t mat, float v);
-
-NCNN_EXPORT ncnn_mat_t ncnn_mat_clone(const ncnn_mat_t mat, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_reshape_1d(const ncnn_mat_t mat, int w, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_reshape_2d(const ncnn_mat_t mat, int w, int h, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_reshape_3d(const ncnn_mat_t mat, int w, int h, int c, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_reshape_4d(const ncnn_mat_t mat, int w, int h, int d, int c, ncnn_allocator_t allocator);
-
-NCNN_EXPORT int ncnn_mat_get_dims(const ncnn_mat_t mat);
-NCNN_EXPORT int ncnn_mat_get_w(const ncnn_mat_t mat);
-NCNN_EXPORT int ncnn_mat_get_h(const ncnn_mat_t mat);
-NCNN_EXPORT int ncnn_mat_get_d(const ncnn_mat_t mat);
-NCNN_EXPORT int ncnn_mat_get_c(const ncnn_mat_t mat);
-NCNN_EXPORT size_t ncnn_mat_get_elemsize(const ncnn_mat_t mat);
-NCNN_EXPORT int ncnn_mat_get_elempack(const ncnn_mat_t mat);
-NCNN_EXPORT size_t ncnn_mat_get_cstep(const ncnn_mat_t mat);
-NCNN_EXPORT void* ncnn_mat_get_data(const ncnn_mat_t mat);
-
-NCNN_EXPORT void* ncnn_mat_get_channel_data(const ncnn_mat_t mat, int c);
-
-#if NCNN_PIXEL
-
-/* mat pixel api */
-#define NCNN_MAT_PIXEL_RGB       1
-#define NCNN_MAT_PIXEL_BGR       2
-#define NCNN_MAT_PIXEL_GRAY      3
-#define NCNN_MAT_PIXEL_RGBA      4
-#define NCNN_MAT_PIXEL_BGRA      5
-#define NCNN_MAT_PIXEL_X2Y(X, Y) (X | (Y << 16))
-NCNN_EXPORT ncnn_mat_t ncnn_mat_from_pixels(const unsigned char* pixels, int type, int w, int h, int stride, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_from_pixels_resize(const unsigned char* pixels, int type, int w, int h, int stride, int target_width, int target_height, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_from_pixels_roi(const unsigned char* pixels, int type, int w, int h, int stride, int roix, int roiy, int roiw, int roih, ncnn_allocator_t allocator);
-NCNN_EXPORT ncnn_mat_t ncnn_mat_from_pixels_roi_resize(const unsigned char* pixels, int type, int w, int h, int stride, int roix, int roiy, int roiw, int roih, int target_width, int target_height, ncnn_allocator_t allocator);
-NCNN_EXPORT void ncnn_mat_to_pixels(const ncnn_mat_t mat, unsigned char* pixels, int type, int stride);
-NCNN_EXPORT void ncnn_mat_to_pixels_resize(const ncnn_mat_t mat, unsigned char* pixels, int type, int target_width, int target_height, int target_stride);
-
-#endif /* NCNN_PIXEL */
-
-NCNN_EXPORT void ncnn_mat_substract_mean_normalize(ncnn_mat_t mat, const float* mean_vals, const float* norm_vals);
-
-NCNN_EXPORT void ncnn_convert_packing(const ncnn_mat_t src, ncnn_mat_t* dst, int elempack, const ncnn_option_t opt);
-NCNN_EXPORT void ncnn_flatten(const ncnn_mat_t src, ncnn_mat_t* dst, const ncnn_option_t opt);
-
-/* blob api */
-typedef struct __ncnn_blob_t* ncnn_blob_t;
-
-#if NCNN_STRING
-NCNN_EXPORT const char* ncnn_blob_get_name(const ncnn_blob_t blob);
-#endif /* NCNN_STRING */
-
-NCNN_EXPORT int ncnn_blob_get_producer(const ncnn_blob_t blob);
-NCNN_EXPORT int ncnn_blob_get_consumer(const ncnn_blob_t blob);
-
-NCNN_EXPORT void ncnn_blob_get_shape(const ncnn_blob_t blob, int* dims, int* w, int* h, int* c);
-
-/* paramdict api */
-typedef struct __ncnn_paramdict_t* ncnn_paramdict_t;
-
-NCNN_EXPORT ncnn_paramdict_t ncnn_paramdict_create();
-NCNN_EXPORT void ncnn_paramdict_destroy(ncnn_paramdict_t pd);
-
-NCNN_EXPORT int ncnn_paramdict_get_type(const ncnn_paramdict_t pd, int id);
-
-NCNN_EXPORT int ncnn_paramdict_get_int(const ncnn_paramdict_t pd, int id, int def);
-NCNN_EXPORT float ncnn_paramdict_get_float(const ncnn_paramdict_t pd, int id, float def);
-NCNN_EXPORT ncnn_mat_t ncnn_paramdict_get_array(const ncnn_paramdict_t pd, int id, const ncnn_mat_t def);
-
-NCNN_EXPORT void ncnn_paramdict_set_int(ncnn_paramdict_t pd, int id, int i);
-NCNN_EXPORT void ncnn_paramdict_set_float(ncnn_paramdict_t pd, int id, float f);
-NCNN_EXPORT void ncnn_paramdict_set_array(ncnn_paramdict_t pd, int id, const ncnn_mat_t v);
-
-/* datareader api */
-typedef struct __ncnn_datareader_t* ncnn_datareader_t;
-struct NCNN_EXPORT __ncnn_datareader_t
-{
-    void* pthis;
-
-#if NCNN_STRING
-    int (*scan)(ncnn_datareader_t dr, const char* format, void* p);
-#endif /* NCNN_STRING */
-    size_t (*read)(ncnn_datareader_t dr, void* buf, size_t size);
-};
-
-NCNN_EXPORT ncnn_datareader_t ncnn_datareader_create();
-#if NCNN_STDIO
-NCNN_EXPORT ncnn_datareader_t ncnn_datareader_create_from_stdio(FILE* fp);
-#endif /* NCNN_STDIO */
-NCNN_EXPORT ncnn_datareader_t ncnn_datareader_create_from_memory(const unsigned char** mem);
-NCNN_EXPORT void ncnn_datareader_destroy(ncnn_datareader_t dr);
-
-/* modelbin api */
-typedef struct __ncnn_modelbin_t* ncnn_modelbin_t;
-struct NCNN_EXPORT __ncnn_modelbin_t
-{
-    void* pthis;
-
-    ncnn_mat_t (*load_1d)(const ncnn_modelbin_t mb, int w, int type);
-    ncnn_mat_t (*load_2d)(const ncnn_modelbin_t mb, int w, int h, int type);
-    ncnn_mat_t (*load_3d)(const ncnn_modelbin_t mb, int w, int h, int c, int type);
-};
-
-NCNN_EXPORT ncnn_modelbin_t ncnn_modelbin_create_from_datareader(const ncnn_datareader_t dr);
-NCNN_EXPORT ncnn_modelbin_t ncnn_modelbin_create_from_mat_array(const ncnn_mat_t* weights, int n);
-NCNN_EXPORT void ncnn_modelbin_destroy(ncnn_modelbin_t mb);
-
-/* layer api */
-typedef struct __ncnn_layer_t* ncnn_layer_t;
-struct NCNN_EXPORT __ncnn_layer_t
-{
-    void* pthis;
-
-    int (*load_param)(ncnn_layer_t layer, const ncnn_paramdict_t pd);
-    int (*load_model)(ncnn_layer_t layer, const ncnn_modelbin_t mb);
-
-    int (*create_pipeline)(ncnn_layer_t layer, const ncnn_option_t opt);
-    int (*destroy_pipeline)(ncnn_layer_t layer, const ncnn_option_t opt);
-
-    int (*forward_1)(const ncnn_layer_t layer, const ncnn_mat_t bottom_blob, ncnn_mat_t* top_blob, const ncnn_option_t opt);
-    int (*forward_n)(const ncnn_layer_t layer, const ncnn_mat_t* bottom_blobs, int n, ncnn_mat_t* top_blobs, int n2, const ncnn_option_t opt);
-
-    int (*forward_inplace_1)(const ncnn_layer_t layer, ncnn_mat_t bottom_top_blob, const ncnn_option_t opt);
-    int (*forward_inplace_n)(const ncnn_layer_t layer, ncnn_mat_t* bottom_top_blobs, int n, const ncnn_option_t opt);
-};
-
-NCNN_EXPORT ncnn_layer_t ncnn_layer_create();
-NCNN_EXPORT ncnn_layer_t ncnn_layer_create_by_typeindex(int typeindex);
-#if NCNN_STRING
-NCNN_EXPORT ncnn_layer_t ncnn_layer_create_by_type(const char* type);
-NCNN_EXPORT int ncnn_layer_type_to_index(const char* type);
-#endif /* NCNN_STRING */
-NCNN_EXPORT void ncnn_layer_destroy(ncnn_layer_t layer);
-
-#if NCNN_STRING
-NCNN_EXPORT const char* ncnn_layer_get_name(const ncnn_layer_t layer);
-#endif /* NCNN_STRING */
-
-NCNN_EXPORT int ncnn_layer_get_typeindex(const ncnn_layer_t layer);
-#if NCNN_STRING
-NCNN_EXPORT const char* ncnn_layer_get_type(const ncnn_layer_t layer);
-#endif /* NCNN_STRING */
-
-NCNN_EXPORT int ncnn_layer_get_one_blob_only(const ncnn_layer_t layer);
-NCNN_EXPORT int ncnn_layer_get_support_inplace(const ncnn_layer_t layer);
-NCNN_EXPORT int ncnn_layer_get_support_vulkan(const ncnn_layer_t layer);
-NCNN_EXPORT int ncnn_layer_get_support_packing(const ncnn_layer_t layer);
-NCNN_EXPORT int ncnn_layer_get_support_bf16_storage(const ncnn_layer_t layer);
-NCNN_EXPORT int ncnn_layer_get_support_fp16_storage(const ncnn_layer_t layer);
-NCNN_EXPORT int ncnn_layer_get_support_image_storage(const ncnn_layer_t layer);
-
-NCNN_EXPORT void ncnn_layer_set_one_blob_only(ncnn_layer_t layer, int enable);
-NCNN_EXPORT void ncnn_layer_set_support_inplace(ncnn_layer_t layer, int enable);
-NCNN_EXPORT void ncnn_layer_set_support_vulkan(ncnn_layer_t layer, int enable);
-NCNN_EXPORT void ncnn_layer_set_support_packing(ncnn_layer_t layer, int enable);
-NCNN_EXPORT void ncnn_layer_set_support_bf16_storage(ncnn_layer_t layer, int enable);
-NCNN_EXPORT void ncnn_layer_set_support_fp16_storage(ncnn_layer_t layer, int enable);
-NCNN_EXPORT void ncnn_layer_set_support_image_storage(ncnn_layer_t layer, int enable);
-
-NCNN_EXPORT int ncnn_layer_get_bottom_count(const ncnn_layer_t layer);
-NCNN_EXPORT int ncnn_layer_get_bottom(const ncnn_layer_t layer, int i);
-NCNN_EXPORT int ncnn_layer_get_top_count(const ncnn_layer_t layer);
-NCNN_EXPORT int ncnn_layer_get_top(const ncnn_layer_t layer, int i);
-
-NCNN_EXPORT void ncnn_blob_get_bottom_shape(const ncnn_layer_t layer, int i, int* dims, int* w, int* h, int* c);
-NCNN_EXPORT void ncnn_blob_get_top_shape(const ncnn_layer_t layer, int i, int* dims, int* w, int* h, int* c);
-
-/* layer factory function */
-typedef ncnn_layer_t (*ncnn_layer_creator_t)(void* userdata);
-typedef void (*ncnn_layer_destroyer_t)(ncnn_layer_t layer, void* userdata);
-
-typedef struct __ncnn_net_custom_layer_factory_t* ncnn_net_custom_layer_factory_t;
-struct __ncnn_net_custom_layer_factory_t
-{
-    ncnn_layer_creator_t creator;
-    ncnn_layer_destroyer_t destroyer;
-    void* userdata;
-    ncnn_net_custom_layer_factory_t next;
-};
-
-/* net api */
-typedef struct __ncnn_net_t* ncnn_net_t;
-struct __ncnn_net_t
-{
-    void* pthis;
-
-    ncnn_net_custom_layer_factory_t custom_layer_factory;
-};
-
-NCNN_EXPORT ncnn_net_t ncnn_net_create();
-NCNN_EXPORT void ncnn_net_destroy(ncnn_net_t net);
-
-NCNN_EXPORT ncnn_option_t ncnn_net_get_option(ncnn_net_t net);
-NCNN_EXPORT void ncnn_net_set_option(ncnn_net_t net, ncnn_option_t opt);
-
-#if NCNN_STRING
-NCNN_EXPORT void ncnn_net_register_custom_layer_by_type(ncnn_net_t net, const char* type, ncnn_layer_creator_t creator, ncnn_layer_destroyer_t destroyer, void* userdata);
-#endif /* NCNN_STRING */
-NCNN_EXPORT void ncnn_net_register_custom_layer_by_typeindex(ncnn_net_t net, int typeindex, ncnn_layer_creator_t creator, ncnn_layer_destroyer_t destroyer, void* userdata);
-
-#if NCNN_STDIO
-#if NCNN_STRING
-NCNN_EXPORT int ncnn_net_load_param(ncnn_net_t net, const char* path);
-#endif /* NCNN_STRING */
-NCNN_EXPORT int ncnn_net_load_param_bin(ncnn_net_t net, const char* path);
-NCNN_EXPORT int ncnn_net_load_model(ncnn_net_t net, const char* path);
-#endif /* NCNN_STDIO */
-
-#if NCNN_STDIO
-#if NCNN_STRING
-NCNN_EXPORT int ncnn_net_load_param_memory(ncnn_net_t net, const char* mem);
-#endif /* NCNN_STRING */
-#endif /* NCNN_STDIO */
-NCNN_EXPORT int ncnn_net_load_param_bin_memory(ncnn_net_t net, const unsigned char* mem);
-NCNN_EXPORT int ncnn_net_load_model_memory(ncnn_net_t net, const unsigned char* mem);
-
-#if NCNN_STRING
-NCNN_EXPORT int ncnn_net_load_param_datareader(ncnn_net_t net, const ncnn_datareader_t dr);
-#endif /* NCNN_STRING */
-NCNN_EXPORT int ncnn_net_load_param_bin_datareader(ncnn_net_t net, const ncnn_datareader_t dr);
-NCNN_EXPORT int ncnn_net_load_model_datareader(ncnn_net_t net, const ncnn_datareader_t dr);
-
-NCNN_EXPORT void ncnn_net_clear(ncnn_net_t net);
-
-NCNN_EXPORT int ncnn_net_get_input_count(const ncnn_net_t net);
-NCNN_EXPORT int ncnn_net_get_output_count(const ncnn_net_t net);
-#if NCNN_STRING
-NCNN_EXPORT const char* ncnn_net_get_input_name(const ncnn_net_t net, int i);
-NCNN_EXPORT const char* ncnn_net_get_output_name(const ncnn_net_t net, int i);
-#endif /* NCNN_STRING */
-NCNN_EXPORT int ncnn_net_get_input_index(const ncnn_net_t net, int i);
-NCNN_EXPORT int ncnn_net_get_output_index(const ncnn_net_t net, int i);
-
-/* extractor api */
-typedef struct __ncnn_extractor_t* ncnn_extractor_t;
-
-NCNN_EXPORT ncnn_extractor_t ncnn_extractor_create(ncnn_net_t net);
-NCNN_EXPORT void ncnn_extractor_destroy(ncnn_extractor_t ex);
-
-NCNN_EXPORT void ncnn_extractor_set_option(ncnn_extractor_t ex, const ncnn_option_t opt);
-
-#if NCNN_STRING
-NCNN_EXPORT int ncnn_extractor_input(ncnn_extractor_t ex, const char* name, const ncnn_mat_t mat);
-NCNN_EXPORT int ncnn_extractor_extract(ncnn_extractor_t ex, const char* name, ncnn_mat_t* mat);
-#endif /* NCNN_STRING */
-NCNN_EXPORT int ncnn_extractor_input_index(ncnn_extractor_t ex, int index, const ncnn_mat_t mat);
-NCNN_EXPORT int ncnn_extractor_extract_index(ncnn_extractor_t ex, int index, ncnn_mat_t* mat);
-
-/* mat process api */
-#define NCNN_BORDER_CONSTANT    0
-#define NCNN_BORDER_REPLICATE   1
-#define NCNN_BORDER_REFLECT     2
-#define NCNN_BORDER_TRANSPARENT -233
-NCNN_EXPORT void ncnn_copy_make_border(const ncnn_mat_t src, ncnn_mat_t dst, int top, int bottom, int left, int right, int type, float v, const ncnn_option_t opt);
-NCNN_EXPORT void ncnn_copy_make_border_3d(const ncnn_mat_t src, ncnn_mat_t dst, int top, int bottom, int left, int right, int front, int behind, int type, float v, const ncnn_option_t opt);
-NCNN_EXPORT void ncnn_copy_cut_border(const ncnn_mat_t src, ncnn_mat_t dst, int top, int bottom, int left, int right, const ncnn_option_t opt);
-NCNN_EXPORT void ncnn_copy_cut_border_3d(const ncnn_mat_t src, ncnn_mat_t dst, int top, int bottom, int left, int right, int front, int behind, const ncnn_option_t opt);
-
-#if NCNN_PIXEL_DRAWING
-/* mat pixel drawing api*/
-NCNN_EXPORT void ncnn_draw_rectangle_c1(unsigned char* pixels, int w, int h, int rx, int ry, int rw, int rh, unsigned int color, int thickness);
-NCNN_EXPORT void ncnn_draw_rectangle_c2(unsigned char* pixels, int w, int h, int rx, int ry, int rw, int rh, unsigned int color, int thickness);
-NCNN_EXPORT void ncnn_draw_rectangle_c3(unsigned char* pixels, int w, int h, int rx, int ry, int rw, int rh, unsigned int color, int thickness);
-NCNN_EXPORT void ncnn_draw_rectangle_c4(unsigned char* pixels, int w, int h, int rx, int ry, int rw, int rh, unsigned int color, int thickness);
-
-NCNN_EXPORT void ncnn_draw_text_c1(unsigned char* pixels, int w, int h, const char* text, int x, int y, int fontpixelsize, unsigned int color);
-NCNN_EXPORT void ncnn_draw_text_c2(unsigned char* pixels, int w, int h, const char* text, int x, int y, int fontpixelsize, unsigned int color);
-NCNN_EXPORT void ncnn_draw_text_c3(unsigned char* pixels, int w, int h, const char* text, int x, int y, int fontpixelsize, unsigned int color);
-NCNN_EXPORT void ncnn_draw_text_c4(unsigned char* pixels, int w, int h, const char* text, int x, int y, int fontpixelsize, unsigned int color);
-
-NCNN_EXPORT void ncnn_draw_circle_c1(unsigned char* pixels, int w, int h, int cx, int cy, int radius, unsigned int color, int thickness);
-NCNN_EXPORT void ncnn_draw_circle_c2(unsigned char* pixels, int w, int h, int cx, int cy, int radius, unsigned int color, int thickness);
-NCNN_EXPORT void ncnn_draw_circle_c3(unsigned char* pixels, int w, int h, int cx, int cy, int radius, unsigned int color, int thickness);
-NCNN_EXPORT void ncnn_draw_circle_c4(unsigned char* pixels, int w, int h, int cx, int cy, int radius, unsigned int color, int thickness);
-
-NCNN_EXPORT void ncnn_draw_line_c1(unsigned char* pixels, int w, int h, int x0, int y0, int x1, int y1, unsigned int color, int thickness);
-NCNN_EXPORT void ncnn_draw_line_c2(unsigned char* pixels, int w, int h, int x0, int y0, int x1, int y1, unsigned int color, int thickness);
-NCNN_EXPORT void ncnn_draw_line_c3(unsigned char* pixels, int w, int h, int x0, int y0, int x1, int y1, unsigned int color, int thickness);
-NCNN_EXPORT void ncnn_draw_line_c4(unsigned char* pixels, int w, int h, int x0, int y0, int x1, int y1, unsigned int color, int thickness);
-#endif /* NCNN_PIXEL_DRAWING */
-
-#ifdef __cplusplus
-} /* extern "C" */
-#endif
-
-#endif /* NCNN_C_API */
-
-#endif /* NCNN_C_API_H */
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/command.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/command.h
deleted file mode 100644
index 88912739..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/command.h
+++ /dev/null
@@ -1,134 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_COMMAND_H
-#define NCNN_COMMAND_H
-
-#include "platform.h"
-
-#if NCNN_VULKAN
-
-#include "mat.h"
-
-namespace ncnn {
-
-class Pipeline;
-#if NCNN_PLATFORM_API
-#if __ANDROID_API__ >= 26
-class ImportAndroidHardwareBufferPipeline;
-#endif // __ANDROID_API__ >= 26
-#endif // NCNN_PLATFORM_API
-class VkComputePrivate;
-class NCNN_EXPORT VkCompute
-{
-public:
-    explicit VkCompute(const VulkanDevice* vkdev);
-    virtual ~VkCompute();
-
-public:
-    void record_upload(const Mat& src, VkMat& dst, const Option& opt);
-
-    void record_upload(const Mat& src, VkImageMat& dst, const Option& opt);
-
-    void record_download(const VkMat& src, Mat& dst, const Option& opt);
-
-    void record_download(const VkImageMat& src, Mat& dst, const Option& opt);
-
-    void record_buffer_to_image(const VkMat& src, VkImageMat& dst, const Option& opt);
-
-    void record_image_to_buffer(const VkImageMat& src, VkMat& dst, const Option& opt);
-
-    void record_clone(const Mat& src, VkMat& dst, const Option& opt);
-
-    void record_clone(const Mat& src, VkImageMat& dst, const Option& opt);
-
-    void record_clone(const VkMat& src, Mat& dst, const Option& opt);
-
-    void record_clone(const VkImageMat& src, Mat& dst, const Option& opt);
-
-    void record_clone(const VkMat& src, VkMat& dst, const Option& opt);
-
-    void record_clone(const VkImageMat& src, VkImageMat& dst, const Option& opt);
-
-    void record_clone(const VkMat& src, VkImageMat& dst, const Option& opt);
-
-    void record_clone(const VkImageMat& src, VkMat& dst, const Option& opt);
-
-    void record_pipeline(const Pipeline* pipeline, const std::vector<VkMat>& bindings, const std::vector<vk_constant_type>& constants, const VkMat& dispatcher);
-
-    void record_pipeline(const Pipeline* pipeline, const std::vector<VkImageMat>& bindings, const std::vector<vk_constant_type>& constants, const VkImageMat& dispatcher);
-
-    void record_pipeline(const Pipeline* pipeline, const std::vector<VkMat>& buffer_bindings, const std::vector<VkImageMat>& image_bindings, const std::vector<vk_constant_type>& constants, const VkMat& dispatcher);
-    void record_pipeline(const Pipeline* pipeline, const std::vector<VkMat>& buffer_bindings, const std::vector<VkImageMat>& image_bindings, const std::vector<vk_constant_type>& constants, const VkImageMat& dispatcher);
-    void record_pipeline(const Pipeline* pipeline, const std::vector<VkMat>& buffer_bindings, const std::vector<VkImageMat>& image_bindings, const std::vector<vk_constant_type>& constants, const Mat& dispatcher);
-
-#if NCNN_BENCHMARK
-    void record_write_timestamp(uint32_t query);
-#endif // NCNN_BENCHMARK
-
-#if NCNN_PLATFORM_API
-#if __ANDROID_API__ >= 26
-    void record_import_android_hardware_buffer(const ImportAndroidHardwareBufferPipeline* pipeline, const VkImageMat& src, const VkMat& dst);
-
-    void record_import_android_hardware_buffer(const ImportAndroidHardwareBufferPipeline* pipeline, const VkImageMat& src, const VkImageMat& dst);
-#endif // __ANDROID_API__ >= 26
-#endif // NCNN_PLATFORM_API
-
-    int submit_and_wait();
-
-    int reset();
-
-#if NCNN_BENCHMARK
-    int create_query_pool(uint32_t query_count);
-
-    int get_query_pool_results(uint32_t first_query, uint32_t query_count, std::vector<uint64_t>& results);
-#endif // NCNN_BENCHMARK
-
-protected:
-    const VulkanDevice* vkdev;
-
-    void barrier_readwrite(const VkMat& binding);
-    void barrier_readwrite(const VkImageMat& binding);
-    void barrier_readonly(const VkImageMat& binding);
-
-private:
-    VkComputePrivate* const d;
-};
-
-class VkTransferPrivate;
-class NCNN_EXPORT VkTransfer
-{
-public:
-    explicit VkTransfer(const VulkanDevice* vkdev);
-    virtual ~VkTransfer();
-
-public:
-    void record_upload(const Mat& src, VkMat& dst, const Option& opt, bool flatten = true);
-
-    void record_upload(const Mat& src, VkImageMat& dst, const Option& opt);
-
-    int submit_and_wait();
-
-protected:
-    const VulkanDevice* vkdev;
-
-private:
-    VkTransferPrivate* const d;
-};
-
-} // namespace ncnn
-
-#endif // NCNN_VULKAN
-
-#endif // NCNN_COMMAND_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/cpu.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/cpu.h
deleted file mode 100644
index 7d6bfce1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/cpu.h
+++ /dev/null
@@ -1,178 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_CPU_H
-#define NCNN_CPU_H
-
-#include <stddef.h>
-
-#if (defined _WIN32 && !(defined __MINGW32__))
-#define WIN32_LEAN_AND_MEAN
-#include <windows.h>
-#endif
-#if defined __ANDROID__ || defined __linux__
-#include <sched.h> // cpu_set_t
-#endif
-
-#include "platform.h"
-
-namespace ncnn {
-
-class NCNN_EXPORT CpuSet
-{
-public:
-    CpuSet();
-    void enable(int cpu);
-    void disable(int cpu);
-    void disable_all();
-    bool is_enabled(int cpu) const;
-    int num_enabled() const;
-
-public:
-#if (defined _WIN32 && !(defined __MINGW32__))
-    ULONG_PTR mask;
-#endif
-#if defined __ANDROID__ || defined __linux__
-    cpu_set_t cpu_set;
-#endif
-#if __APPLE__
-    unsigned int policy;
-#endif
-};
-
-// test optional cpu features
-// edsp = armv7 edsp
-NCNN_EXPORT int cpu_support_arm_edsp();
-// neon = armv7 neon or aarch64 asimd
-NCNN_EXPORT int cpu_support_arm_neon();
-// vfpv4 = armv7 fp16 + fma
-NCNN_EXPORT int cpu_support_arm_vfpv4();
-// asimdhp = aarch64 asimd half precision
-NCNN_EXPORT int cpu_support_arm_asimdhp();
-// cpuid = aarch64 cpuid info
-NCNN_EXPORT int cpu_support_arm_cpuid();
-// asimddp = aarch64 asimd dot product
-NCNN_EXPORT int cpu_support_arm_asimddp();
-// asimdfhm = aarch64 asimd fhm
-NCNN_EXPORT int cpu_support_arm_asimdfhm();
-// bf16 = aarch64 bf16
-NCNN_EXPORT int cpu_support_arm_bf16();
-// i8mm = aarch64 i8mm
-NCNN_EXPORT int cpu_support_arm_i8mm();
-// sve = aarch64 sve
-NCNN_EXPORT int cpu_support_arm_sve();
-// sve2 = aarch64 sve2
-NCNN_EXPORT int cpu_support_arm_sve2();
-// svebf16 = aarch64 svebf16
-NCNN_EXPORT int cpu_support_arm_svebf16();
-// svei8mm = aarch64 svei8mm
-NCNN_EXPORT int cpu_support_arm_svei8mm();
-// svef32mm = aarch64 svef32mm
-NCNN_EXPORT int cpu_support_arm_svef32mm();
-
-// avx = x86 avx
-NCNN_EXPORT int cpu_support_x86_avx();
-// fma = x86 fma
-NCNN_EXPORT int cpu_support_x86_fma();
-// xop = x86 xop
-NCNN_EXPORT int cpu_support_x86_xop();
-// f16c = x86 f16c
-NCNN_EXPORT int cpu_support_x86_f16c();
-// avx2 = x86 avx2 + fma + f16c
-NCNN_EXPORT int cpu_support_x86_avx2();
-// avx_vnni = x86 avx vnni
-NCNN_EXPORT int cpu_support_x86_avx_vnni();
-// avx512 = x86 avx512f + avx512cd + avx512bw + avx512dq + avx512vl
-NCNN_EXPORT int cpu_support_x86_avx512();
-// avx512_vnni = x86 avx512 vnni
-NCNN_EXPORT int cpu_support_x86_avx512_vnni();
-// avx512_bf16 = x86 avx512 bf16
-NCNN_EXPORT int cpu_support_x86_avx512_bf16();
-// avx512_fp16 = x86 avx512 fp16
-NCNN_EXPORT int cpu_support_x86_avx512_fp16();
-
-// lsx = loongarch lsx
-NCNN_EXPORT int cpu_support_loongarch_lsx();
-// lasx = loongarch lasx
-NCNN_EXPORT int cpu_support_loongarch_lasx();
-
-// msa = mips mas
-NCNN_EXPORT int cpu_support_mips_msa();
-// mmi = loongson mmi
-NCNN_EXPORT int cpu_support_loongson_mmi();
-
-// v = riscv vector
-NCNN_EXPORT int cpu_support_riscv_v();
-// zfh = riscv half-precision float
-NCNN_EXPORT int cpu_support_riscv_zfh();
-// vlenb = riscv vector length in bytes
-NCNN_EXPORT int cpu_riscv_vlenb();
-
-// cpu info
-NCNN_EXPORT int get_cpu_count();
-NCNN_EXPORT int get_little_cpu_count();
-NCNN_EXPORT int get_big_cpu_count();
-
-NCNN_EXPORT int get_physical_cpu_count();
-NCNN_EXPORT int get_physical_little_cpu_count();
-NCNN_EXPORT int get_physical_big_cpu_count();
-
-// cpu l2 varies from 64k to 1M, but l3 can be zero
-NCNN_EXPORT int get_cpu_level2_cache_size();
-NCNN_EXPORT int get_cpu_level3_cache_size();
-
-// bind all threads on little clusters if powersave enabled
-// affects HMP arch cpu like ARM big.LITTLE
-// only implemented on android at the moment
-// switching powersave is expensive and not thread-safe
-// 0 = all cores enabled(default)
-// 1 = only little clusters enabled
-// 2 = only big clusters enabled
-// return 0 if success for setter function
-NCNN_EXPORT int get_cpu_powersave();
-NCNN_EXPORT int set_cpu_powersave(int powersave);
-
-// convenient wrapper
-NCNN_EXPORT const CpuSet& get_cpu_thread_affinity_mask(int powersave);
-
-// set explicit thread affinity
-NCNN_EXPORT int set_cpu_thread_affinity(const CpuSet& thread_affinity_mask);
-
-// runtime thread affinity info
-NCNN_EXPORT int is_current_thread_running_on_a53_a55();
-
-// misc function wrapper for openmp routines
-NCNN_EXPORT int get_omp_num_threads();
-NCNN_EXPORT void set_omp_num_threads(int num_threads);
-
-NCNN_EXPORT int get_omp_dynamic();
-NCNN_EXPORT void set_omp_dynamic(int dynamic);
-
-NCNN_EXPORT int get_omp_thread_num();
-
-NCNN_EXPORT int get_kmp_blocktime();
-NCNN_EXPORT void set_kmp_blocktime(int time_ms);
-
-// need to flush denormals on Intel Chipset.
-// Other architectures such as ARM can be added as needed.
-// 0 = DAZ OFF, FTZ OFF
-// 1 = DAZ ON , FTZ OFF
-// 2 = DAZ OFF, FTZ ON
-// 3 = DAZ ON,  FTZ ON
-NCNN_EXPORT int get_flush_denormals();
-NCNN_EXPORT int set_flush_denormals(int flush_denormals);
-
-} // namespace ncnn
-
-#endif // NCNN_CPU_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/alsa.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/alsa.h
deleted file mode 100644
index 0200e906..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/alsa.h
+++ /dev/null
@@ -1,76 +0,0 @@
-/**
- * Copyright (c)  2023  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_ALSA_H_
-#define SHERPA_NCNN_CSRC_ALSA_H_
-
-#include <memory>
-#include <vector>
-
-#include "alsa/asoundlib.h"
-#include "sherpa-ncnn/csrc/resample.h"
-
-namespace sherpa_ncnn {
-
-class Alsa {
- public:
-  explicit Alsa(const char *device_name);
-  ~Alsa();
-
-  // This is a blocking read.
-  //
-  // @param num_samples  Number of samples to read.
-  //
-  // The returned value is valid until the next call to Read().
-  const std::vector<float> &Read(int32_t num_samples);
-
-  int32_t GetExpectedSampleRate() const { return expected_sample_rate_; }
-  int32_t GetActualSampleRate() const { return actual_sample_rate_; }
-
- private:
-  const std::vector<float> &Read16(int32_t num_samples);
-  const std::vector<float> &Read32(int32_t num_samples);
-
- private:
-  snd_pcm_t *capture_handle_;
-  int32_t expected_sample_rate_ = 16000;
-  int32_t actual_sample_rate_;
-
-  int32_t actual_channel_count_ = 1;
-
-  // If there are multipel channels, we use this channel for recognition
-  int32_t channel_to_use_ = 0;
-
-  std::unique_ptr<LinearResample> resampler_;
-
-  // If it is 16, we use samples16_
-  // If it is 32, we use samples32_
-  //
-  // It can only be 16 or 32.
-  int32_t pcm_format_ = 16;
-
-  std::vector<int16_t> samples16_;  // directly from the microphone
-  std::vector<int32_t> samples32_;  // directly from the microphone
-
-  std::vector<float> samples1_;  // normalized version of samples_
-  std::vector<float> samples2_;  // possibly resampled from samples1_
-};
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_ALSA_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/context-graph.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/context-graph.h
deleted file mode 100644
index 326c4fa0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/context-graph.h
+++ /dev/null
@@ -1,91 +0,0 @@
-// sherpa-ncnn/csrc/context-graph.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_NCNN_CSRC_CONTEXT_GRAPH_H_
-#define SHERPA_NCNN_CSRC_CONTEXT_GRAPH_H_
-
-#include <memory>
-#include <string>
-#include <tuple>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-namespace sherpa_ncnn {
-
-class ContextGraph;
-using ContextGraphPtr = std::shared_ptr<ContextGraph>;
-
-struct ContextState {
-  int32_t token;
-  float token_score;
-  float node_score;
-  float output_score;
-  int32_t level;
-  float ac_threshold;
-  bool is_end;
-  std::string phrase;
-  std::unordered_map<int32_t, std::unique_ptr<ContextState>> next;
-  const ContextState *fail = nullptr;
-  const ContextState *output = nullptr;
-
-  ContextState() = default;
-  ContextState(int32_t token, float token_score, float node_score,
-               float output_score, int32_t level = 0, float ac_threshold = 0.0f,
-               bool is_end = false, const std::string &phrase = {})
-      : token(token),
-        token_score(token_score),
-        node_score(node_score),
-        output_score(output_score),
-        level(level),
-        ac_threshold(ac_threshold),
-        is_end(is_end),
-        phrase(phrase) {}
-};
-
-class ContextGraph {
- public:
-  ContextGraph() = default;
-  ContextGraph(const std::vector<std::vector<int32_t>> &token_ids,
-               float context_score, float ac_threshold,
-               const std::vector<float> &scores = {},
-               const std::vector<std::string> &phrases = {},
-               const std::vector<float> &ac_thresholds = {})
-      : context_score_(context_score), ac_threshold_(ac_threshold) {
-    root_ = std::make_unique<ContextState>(-1, 0, 0, 0);
-    root_->fail = root_.get();
-    Build(token_ids, scores, phrases, ac_thresholds);
-  }
-
-  ContextGraph(const std::vector<std::vector<int32_t>> &token_ids,
-               float context_score, const std::vector<float> &scores = {},
-               const std::vector<std::string> &phrases = {})
-      : ContextGraph(token_ids, context_score, 0.0f, scores, phrases,
-                     std::vector<float>()) {}
-
-  std::tuple<float, const ContextState *, const ContextState *> ForwardOneStep(
-      const ContextState *state, int32_t token_id,
-      bool strict_mode = true) const;
-
-  std::pair<bool, const ContextState *> IsMatched(
-      const ContextState *state) const;
-
-  std::pair<float, const ContextState *> Finalize(
-      const ContextState *state) const;
-
-  const ContextState *Root() const { return root_.get(); }
-
- private:
-  float context_score_;
-  float ac_threshold_;
-  std::unique_ptr<ContextState> root_;
-  void Build(const std::vector<std::vector<int32_t>> &token_ids,
-             const std::vector<float> &scores,
-             const std::vector<std::string> &phrases,
-             const std::vector<float> &ac_thresholds) const;
-  void FillFailOutput() const;
-};
-
-}  // namespace sherpa_ncnn
-#endif  // SHERPA_NCNN_CSRC_CONTEXT_GRAPH_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/conv-emformer-model.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/conv-emformer-model.h
deleted file mode 100644
index 71241ee0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/conv-emformer-model.h
+++ /dev/null
@@ -1,107 +0,0 @@
-// sherpa-ncnn/csrc/conv-emformer-model.h
-//
-// Copyright (c)  2022  Xiaomi Corporation
-
-#ifndef SHERPA_NCNN_CSRC_CONV_EMFORMER_MODEL_H_
-#define SHERPA_NCNN_CSRC_CONV_EMFORMER_MODEL_H_
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "net.h"  // NOLINT
-#include "sherpa-ncnn/csrc/model.h"
-
-namespace sherpa_ncnn {
-// Please refer to https://github.com/k2-fsa/icefall/pull/717
-// for how the model is converted from icefall to ncnn
-class ConvEmformerModel : public Model {
- public:
-  explicit ConvEmformerModel(const ModelConfig &config);
-#if __ANDROID_API__ >= 9
-  ConvEmformerModel(AAssetManager *mgr, const ModelConfig &config);
-#endif
-
-  ncnn::Net &GetEncoder() override { return encoder_; }
-  ncnn::Net &GetDecoder() override { return decoder_; }
-  ncnn::Net &GetJoiner() override { return joiner_; }
-
-  std::vector<ncnn::Mat> GetEncoderInitStates() const override;
-
-  std::pair<ncnn::Mat, std::vector<ncnn::Mat>> RunEncoder(
-      ncnn::Mat &features, const std::vector<ncnn::Mat> &states) override;
-
-  std::pair<ncnn::Mat, std::vector<ncnn::Mat>> RunEncoder(
-      ncnn::Mat &features, const std::vector<ncnn::Mat> &states,
-      ncnn::Extractor *extractor) override;
-
-  ncnn::Mat RunDecoder(ncnn::Mat &decoder_input) override;
-
-  ncnn::Mat RunDecoder(ncnn::Mat &decoder_input,
-                       ncnn::Extractor *extractor) override;
-
-  ncnn::Mat RunJoiner(ncnn::Mat &encoder_out, ncnn::Mat &decoder_out) override;
-
-  ncnn::Mat RunJoiner(ncnn::Mat &encoder_out, ncnn::Mat &decoder_out,
-                      ncnn::Extractor *extractor) override;
-
-  int32_t Segment() const override {
-    // chunk_length 32
-    // right_context 8
-    // subsampling factor 4
-    //
-    // segment = 32 + (8 + 2 * 4 + 3) = 32 + 19 = 51
-    return chunk_length_ + (right_context_length_ + 2 * 4 + 3);
-  }
-
-  // Advance the feature extract by this number of frames after
-  // running the encoder network
-  int32_t Offset() const override { return chunk_length_; }
-
- private:
-  void InitEncoder(const std::string &encoder_param,
-                   const std::string &encoder_bin);
-  void InitDecoder(const std::string &decoder_param,
-                   const std::string &decoder_bin);
-  void InitJoiner(const std::string &joiner_param,
-                  const std::string &joiner_bin);
-
-  void InitEncoderPostProcessing();
-
-#if __ANDROID_API__ >= 9
-  void InitEncoder(AAssetManager *mgr, const std::string &encoder_param,
-                   const std::string &encoder_bin);
-  void InitDecoder(AAssetManager *mgr, const std::string &decoder_param,
-                   const std::string &decoder_bin);
-  void InitJoiner(AAssetManager *mgr, const std::string &joiner_param,
-                  const std::string &joiner_bin);
-#endif
-
-  void InitEncoderInputOutputIndexes();
-  void InitDecoderInputOutputIndexes();
-  void InitJoinerInputOutputIndexes();
-
- private:
-  ncnn::Net encoder_;
-  ncnn::Net decoder_;
-  ncnn::Net joiner_;
-
-  int32_t num_layers_ = 12;               // arg1
-  int32_t memory_size_ = 32;              // arg2
-  int32_t cnn_module_kernel_ = 31;        // arg3
-  int32_t left_context_length_ = 32 / 4;  // arg4
-  int32_t chunk_length_ = 32;             // arg5
-  int32_t right_context_length_ = 8;      // arg6
-  int32_t d_model_ = 512;                 // arg7
-
-  std::vector<int32_t> encoder_input_indexes_;
-  std::vector<int32_t> encoder_output_indexes_;
-
-  std::vector<int32_t> decoder_input_indexes_;
-  std::vector<int32_t> decoder_output_indexes_;
-
-  std::vector<int32_t> joiner_input_indexes_;
-  std::vector<int32_t> joiner_output_indexes_;
-};
-
-}  // namespace sherpa_ncnn
-#endif  // SHERPA_NCNN_CSRC_CONV_EMFORMER_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/decoder.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/decoder.h
deleted file mode 100644
index 22a17a79..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/decoder.h
+++ /dev/null
@@ -1,102 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- * Copyright (c)  2022                     (Pingfeng Luo)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_DECODER_H_
-#define SHERPA_NCNN_CSRC_DECODER_H_
-#include <string>
-#include <vector>
-
-#include "mat.h"  // NOLINT
-#include "sherpa-ncnn/csrc/hypothesis.h"
-
-namespace sherpa_ncnn {
-
-struct DecoderConfig {
-  // supported values are: modified_beam_search, greedy_search
-  std::string method = "greedy_search";
-
-  int32_t num_active_paths = 4;  // only used by modified beam search
-
-  DecoderConfig() = default;
-
-  DecoderConfig(const std::string &method, int32_t num_active_paths)
-      : method(method), num_active_paths(num_active_paths) {}
-
-  std::string ToString() const;
-};
-
-struct DecoderResult {
-  /// Number of frames we have decoded so far, counted after subsampling
-  int32_t frame_offset = 0;
-
-  /// The decoded token IDs so far
-  std::vector<int32_t> tokens;
-
-  /// number of trailing blank frames decoded so far
-  int32_t num_trailing_blanks = 0;
-
-  std::vector<int32_t> timestamps;
-
-  // Cache the decoder_out just before endpointing
-  ncnn::Mat decoder_out;
-
-  // used only for modified_beam_search
-  Hypotheses hyps;
-};
-
-class Stream;
-
-class Decoder {
- public:
-  virtual ~Decoder() = default;
-
-  /* Return an empty result.
-   *
-   * To simplify the decoding code, we add `context_size` blanks
-   * to the beginning of the decoding result, which will be
-   * stripped by calling `StripPrecedingBlanks()`.
-   */
-  virtual DecoderResult GetEmptyResult() const = 0;
-
-  /** Strip blanks added by `GetEmptyResult()`.
-   *
-   * @param r It is changed in-place.
-   */
-  virtual void StripLeadingBlanks(DecoderResult * /*r*/) const {}
-
-  /** Run transducer beam search given the output from the encoder model.
-   *
-   * @param encoder_out A 3-D tensor of shape (N, T, joiner_dim)
-   * @param result  It is modified in-place.
-   *
-   * @note There is no need to pass encoder_out_length here since for the
-   * online decoding case, each utterance has the same number of frames
-   * and there are no paddings.
-   */
-  virtual void Decode(ncnn::Mat encoder_out, DecoderResult *result) = 0;
-
-  virtual void Decode(ncnn::Mat encoder_out, Stream *s, DecoderResult *result) {
-    NCNN_LOGE("Please override it!");
-    exit(-1);
-  }
-};
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/display.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/display.h
deleted file mode 100644
index a8972e75..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/display.h
+++ /dev/null
@@ -1,108 +0,0 @@
-/**
- * Copyright (c)  2023  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_DISPLAY_H_
-#define SHERPA_NCNN_CSRC_DISPLAY_H_
-#include <stdio.h>
-
-#include <string>
-
-namespace sherpa_ncnn {
-
-// this class works only on Linux/macOS
-class Display {
- public:
-  explicit Display(int32_t max_word_per_line = 60)
-      : max_word_per_line_(max_word_per_line) {}
-
-  virtual void Print(int32_t segment_id, const std::string &s) {
-#ifdef _MSC_VER
-    if (segment_id != -1) {
-      if (last_segment_ != segment_id) {
-          fprintf(stderr, "\n%d:%s", segment_id, s.c_str());
-          last_segment_ = segment_id;
-      } else {
-          fprintf(stderr, "\r%d:%s", segment_id, s.c_str());
-      }
-    } else {
-      fprintf(stderr, "%s\n", s.c_str());
-    }
-    return;
-#endif
-    if (last_segment_ == segment_id) {
-      Clear();
-    } else {
-      if (last_segment_ != -1) {
-        fprintf(stderr, "\n\r");
-      }
-      last_segment_ = segment_id;
-      num_previous_lines_ = 0;
-    }
-
-    if (segment_id != -1) {
-      fprintf(stderr, "\r%d:", segment_id);
-    }
-
-    int32_t i = 0;
-    for (size_t n = 0; n < s.size();) {
-      if (s[n] > 0 && s[n] < 0x7f) {
-        fprintf(stderr, "%c", s[n]);
-        ++n;
-      } else {
-        // Each Chinese character occupies 3 bytes for UTF-8 encoding.
-        std::string tmp(s.begin() + n, s.begin() + n + 3);
-        fprintf(stderr, "%s", tmp.data());
-        n += 3;
-      }
-
-      ++i;
-      if (i >= max_word_per_line_ && n + 1 < s.size() &&
-          (s[n] == ' ' || s[n] < 0)) {
-        fprintf(stderr, "\n\r ");
-        ++num_previous_lines_;
-        i = 0;
-      }
-    }
-  }
-
- private:
-  // Clear the output for the current segment
-  void Clear() {
-    ClearCurrentLine();
-    while (num_previous_lines_ > 0) {
-      GoUpOneLine();
-      ClearCurrentLine();
-      --num_previous_lines_;
-    }
-  }
-
-  // Clear the current line
-  void ClearCurrentLine() const { fprintf(stderr, "\33[2K\r"); }
-
-  // Move the cursor to the previous line
-  void GoUpOneLine() const { fprintf(stderr, "\033[1A\r"); }
-
- private:
-  int32_t max_word_per_line_;
-  int32_t num_previous_lines_ = 0;
-  int32_t last_segment_ = -1;
-};
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_DISPLAY_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/endpoint.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/endpoint.h
deleted file mode 100644
index 7198a4de..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/endpoint.h
+++ /dev/null
@@ -1,84 +0,0 @@
-/**
- * Copyright      2022  (authors: Pingfeng Luo)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-#ifndef SHERPA_NCNN_CSRC_ENDPOINT_H_
-#define SHERPA_NCNN_CSRC_ENDPOINT_H_
-
-#include <string>
-#include <vector>
-
-namespace sherpa_ncnn {
-
-struct EndpointRule {
-  // If True, for this endpointing rule to apply there must
-  // be nonsilence in the best-path traceback.
-  // For decoding, a non-blank token is considered as non-silence
-  bool must_contain_nonsilence;
-  // This endpointing rule requires duration of trailing silence
-  // (in seconds) to be >= this value.
-  float min_trailing_silence;
-  // This endpointing rule requires utterance-length (in seconds)
-  // to be >= this value.
-  float min_utterance_length;
-
-  explicit EndpointRule(const bool must_contain_nonsilence = true,
-                        const float min_trailing_silence = 2.0,
-                        const float min_utterance_length = 0)
-      : must_contain_nonsilence(must_contain_nonsilence),
-        min_trailing_silence(min_trailing_silence),
-        min_utterance_length(min_utterance_length) {}
-
-  std::string ToString() const;
-};
-
-struct EndpointConfig {
-  // For default setting,
-  // rule1 times out after 2.4 seconds of silence, even if we decoded nothing.
-  // rule2 times out after 1.2 seconds of silence after decoding something.
-  // rule3 times out after the utterance is 20 seconds long, regardless of
-  // anything else.
-  EndpointRule rule1;
-  EndpointRule rule2;
-  EndpointRule rule3;
-
-  EndpointConfig(const EndpointRule &rule1, const EndpointRule &rule2,
-                 const EndpointRule &rule3)
-      : rule1(rule1), rule2(rule2), rule3(rule3) {}
-
-  EndpointConfig()
-      : rule1(false, 2.4, 0), rule2(true, 1.4, 0), rule3(false, 0, 20) {}
-
-  std::string ToString() const;
-};
-
-class Endpoint {
- public:
-  explicit Endpoint(const EndpointConfig &config) : config_(config) {}
-
-  /// This function returns true if this set of endpointing rules thinks we
-  /// should terminate decoding.
-  bool IsEndpoint(const int num_frames_decoded,
-                  const int trailing_silence_frames,
-                  const float frame_shift_in_seconds) const;
-
- private:
-  EndpointConfig config_;
-};
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_ENDPOINT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/features.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/features.h
deleted file mode 100644
index 51af2bb0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/features.h
+++ /dev/null
@@ -1,82 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_FEATURES_H_
-#define SHERPA_NCNN_CSRC_FEATURES_H_
-
-#include <memory>
-#include <string>
-
-namespace ncnn {
-class Mat;
-}
-
-namespace sherpa_ncnn {
-
-struct FeatureExtractorConfig {
-  int32_t sampling_rate = 16000;
-  int32_t feature_dim = 80;
-
-  std::string ToString() const;
-};
-
-class FeatureExtractor {
- public:
-  explicit FeatureExtractor(const FeatureExtractorConfig &config);
-  ~FeatureExtractor();
-
-  /**
-     @param sampling_rate The sampling_rate of the input waveform. We will
-                          do resample if it is different from
-                          config.sampling_rate.
-                          Caution: You MUST not use a different sampling rate
-                          across different calls for AcceptWaveform().
-     @param waveform Pointer to a 1-D array of size n
-     @param n Number of entries in waveform
-   */
-  void AcceptWaveform(int32_t sampling_rate, const float *waveform, int32_t n);
-
-  // InputFinished() tells the class you won't be providing any
-  // more waveform.  This will help flush out the last frame or two
-  // of features, in the case where snip-edges == false; it also
-  // affects the return value of IsLastFrame().
-  void InputFinished();
-
-  int32_t NumFramesReady() const;
-
-  // Note: IsLastFrame() will only ever return true if you have called
-  // InputFinished() (and this frame is the last frame).
-  bool IsLastFrame(int32_t frame) const;
-
-  /** Get n frames starting from the given frame index.
-   *
-   * @param frame_index  The starting frame index
-   * @param n  Number of frames to get.
-   * @return Return a 2-D tensor of shape (n, feature_dim).
-   *         ans.w == feature_dim; ans.h == n
-   */
-  ncnn::Mat GetFrames(int32_t frame_index, int32_t n) const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_FEATURES_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/greedy-search-decoder.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/greedy-search-decoder.h
deleted file mode 100644
index 440a22de..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/greedy-search-decoder.h
+++ /dev/null
@@ -1,47 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- * Copyright (c)  2022                     (Pingfeng Luo)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_GREEDY_SEARCH_DECODER_H_
-#define SHERPA_NCNN_CSRC_GREEDY_SEARCH_DECODER_H_
-
-#include "sherpa-ncnn/csrc/decoder.h"
-#include "sherpa-ncnn/csrc/model.h"
-
-namespace sherpa_ncnn {
-
-class GreedySearchDecoder : public Decoder {
- public:
-  explicit GreedySearchDecoder(Model *model) : model_(model) {}
-
-  DecoderResult GetEmptyResult() const override;
-
-  void StripLeadingBlanks(DecoderResult * /*r*/) const override;
-
-  void Decode(ncnn::Mat encoder_out, DecoderResult *result) override;
-
- private:
-  ncnn::Mat BuildDecoderInput(const DecoderResult &result) const;
-
- private:
-  Model *model_;  // not owned
-};
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_GREEDY_SEARCH_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/hypothesis.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/hypothesis.h
deleted file mode 100644
index c4349d01..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/hypothesis.h
+++ /dev/null
@@ -1,133 +0,0 @@
-/**
- * Copyright      2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_HYPOTHESIS_H_
-#define SHERPA_NCNN_CSRC_HYPOTHESIS_H_
-
-#include <sstream>
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include "sherpa-ncnn/csrc/context-graph.h"
-
-namespace sherpa_ncnn {
-
-struct Hypothesis {
-  // The predicted tokens so far. Newly predicated tokens are appended.
-  std::vector<int32_t> ys;
-
-  // timestamps[i] contains the frame number after subsampling
-  // on which ys[i] is decoded.
-  std::vector<int32_t> timestamps;
-
-  // The total score of ys in log space.
-  double log_prob = 0;
-  const ContextState *context_state;
-  int32_t num_trailing_blanks = 0;
-
-  Hypothesis() = default;
-  Hypothesis(const std::vector<int32_t> &ys, double log_prob,
-             const ContextState *context_state = nullptr)
-      : ys(ys), log_prob(log_prob), context_state(context_state) {}
-
-  // If two Hypotheses have the same `Key`, then they contain
-  // the same token sequence.
-  std::string Key() const {
-    // TODO(fangjun): Use a hash function?
-    std::ostringstream os;
-    std::string sep = "-";
-    for (auto i : ys) {
-      os << i << sep;
-      sep = "-";
-    }
-    return os.str();
-  }
-
-  // For debugging
-  std::string ToString() const {
-    std::ostringstream os;
-    os << "(" << Key() << ", " << log_prob << ")";
-    return os.str();
-  }
-};
-
-class Hypotheses {
- public:
-  Hypotheses() = default;
-
-  explicit Hypotheses(std::vector<Hypothesis> hyps) {
-    for (auto &h : hyps) {
-      hyps_dict_[h.Key()] = std::move(h);
-    }
-  }
-
-  explicit Hypotheses(std::unordered_map<std::string, Hypothesis> hyps_dict)
-      : hyps_dict_(std::move(hyps_dict)) {}
-
-  // Add hyp to this object. If it already exists, its log_prob
-  // is updated with the given hyp using log-sum-exp.
-  void Add(Hypothesis hyp);
-
-  // Get the hyp that has the largest log_prob.
-  // If length_norm is true, hyp's log_prob is divided by
-  // len(hyp.ys) before comparison.
-  Hypothesis GetMostProbable(bool length_norm) const;
-
-  // Get the k hyps that have the largest log_prob.
-  // If length_norm is true, hyp's log_prob is divided by
-  // len(hyp.ys) before comparison.
-  std::vector<Hypothesis> GetTopK(int32_t k, bool length_norm) const;
-
-  int32_t Size() const { return hyps_dict_.size(); }
-
-  std::string ToString() const {
-    std::ostringstream os;
-    for (const auto &p : hyps_dict_) {
-      os << p.second.ToString() << "\n";
-    }
-    return os.str();
-  }
-
-  const auto begin() const { return hyps_dict_.begin(); }
-  const auto end() const { return hyps_dict_.end(); }
-  auto begin() { return hyps_dict_.begin(); }
-  auto end() { return hyps_dict_.end(); }
-
-  void Clear() { hyps_dict_.clear(); }
-
- private:
-  // Return a list of hyps contained in this object.
-  std::vector<Hypothesis> Vec() const {
-    std::vector<Hypothesis> ans;
-    ans.reserve(hyps_dict_.size());
-    for (const auto &p : hyps_dict_) {
-      ans.push_back(p.second);
-    }
-    return ans;
-  }
-
- private:
-  using Map = std ::unordered_map<std::string, Hypothesis>;
-  Map hyps_dict_;
-};
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_HYPOTHESIS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/lstm-model.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/lstm-model.h
deleted file mode 100644
index f944da66..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/lstm-model.h
+++ /dev/null
@@ -1,132 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_LSTM_MODEL_H_
-#define SHERPA_NCNN_CSRC_LSTM_MODEL_H_
-
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "net.h"  // NOLINT
-#include "sherpa-ncnn/csrc/model.h"
-
-namespace sherpa_ncnn {
-
-class LstmModel : public Model {
- public:
-  explicit LstmModel(const ModelConfig &config);
-#if __ANDROID_API__ >= 9
-  LstmModel(AAssetManager *mgr, const ModelConfig &config);
-#endif
-
-  ncnn::Net &GetEncoder() override { return encoder_; }
-  ncnn::Net &GetDecoder() override { return decoder_; }
-  ncnn::Net &GetJoiner() override { return joiner_; }
-
-  std::vector<ncnn::Mat> GetEncoderInitStates() const override;
-
-  /** Run the encoder network.
-   *
-   * @param features  A 2-d mat of shape (num_frames, feature_dim).
-   *                  Note: features.w = feature_dim.
-   *                        features.h = num_frames.
-   * @param states Contains two tensors:
-   *          - hx  Hidden state of the LSTM model. You can leave it to empty
-   *                on the first invocation. It is changed in-place.
-   *
-   *          - cx  Hidden cell state of the LSTM model. You can leave it to
-   *                empty on the first invocation. It is changed in-place.
-   *
-   *          - Note: on the first invocation, you can pass an empty vector.
-   *
-   * @return Return a pair containing:
-   *   - the output of the encoder. Its shape is (num_out_frames, encoder_dim).
-   *     Note: ans.w == encoder_dim; ans.h == num_out_frames
-   *
-   *   - next_states, a vector containing hx and cx for the next invocation
-   */
-  std::pair<ncnn::Mat, std::vector<ncnn::Mat>> RunEncoder(
-      ncnn::Mat &features, const std::vector<ncnn::Mat> &states) override;
-
-  std::pair<ncnn::Mat, std::vector<ncnn::Mat>> RunEncoder(
-      ncnn::Mat &features, const std::vector<ncnn::Mat> &states,
-      ncnn::Extractor *extractor) override;
-
-  ncnn::Mat RunDecoder(ncnn::Mat &decoder_input) override;
-
-  ncnn::Mat RunDecoder(ncnn::Mat &decoder_input,
-                       ncnn::Extractor *extractor) override;
-
-  ncnn::Mat RunJoiner(ncnn::Mat &encoder_out, ncnn::Mat &decoder_out) override;
-
-  ncnn::Mat RunJoiner(ncnn::Mat &encoder_out, ncnn::Mat &decoder_out,
-                      ncnn::Extractor *extractor) override;
-
-  int32_t Segment() const override { return 9; }
-
-  // Advance the feature extract by this number of frames after
-  // running the encoder network
-  int32_t Offset() const override { return 4; }
-
- private:
-  void InitEncoder(const std::string &encoder_param,
-                   const std::string &encoder_bin);
-  void InitDecoder(const std::string &decoder_param,
-                   const std::string &decoder_bin);
-  void InitJoiner(const std::string &joiner_param,
-                  const std::string &joiner_bin);
-
-  void InitEncoderPostProcessing();
-
-#if __ANDROID_API__ >= 9
-  void InitEncoder(AAssetManager *mgr, const std::string &encoder_param,
-                   const std::string &encoder_bin);
-  void InitDecoder(AAssetManager *mgr, const std::string &decoder_param,
-                   const std::string &decoder_bin);
-  void InitJoiner(AAssetManager *mgr, const std::string &joiner_param,
-                  const std::string &joiner_bin);
-#endif
-
-  void InitEncoderInputOutputIndexes();
-  void InitDecoderInputOutputIndexes();
-  void InitJoinerInputOutputIndexes();
-
- private:
-  // arg0: 3
-  int32_t num_encoder_layers_ = 12;  // arg1
-  int32_t encoder_dim_ = 512;        // arg2, i.e., d_model
-  int32_t rnn_hidden_size_ = 1024;   // arg3
-
-  ncnn::Net encoder_;
-  ncnn::Net decoder_;
-  ncnn::Net joiner_;
-
-  std::vector<int32_t> encoder_input_indexes_;
-  std::vector<int32_t> encoder_output_indexes_;
-
-  std::vector<int32_t> decoder_input_indexes_;
-  std::vector<int32_t> decoder_output_indexes_;
-
-  std::vector<int32_t> joiner_input_indexes_;
-  std::vector<int32_t> joiner_output_indexes_;
-};
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_LSTM_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/math.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/math.h
deleted file mode 100644
index 8017e6c1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/math.h
+++ /dev/null
@@ -1,120 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Daniel Povey)
- * Copyright (c)  2022                     (Pingfeng Luo)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-// This file is copied from k2/csrc/utils.h
-#ifndef SHERPA_NCNN_CSRC_MATH_H_
-#define SHERPA_NCNN_CSRC_MATH_H_
-
-#include <algorithm>
-#include <cassert>
-#include <cmath>
-#include <numeric>
-#include <vector>
-
-namespace sherpa_ncnn {
-
-// logf(FLT_EPSILON)
-#define SHERPA_NCNN_MIN_LOG_DIFF_FLOAT -15.9423847198486328125f
-
-// log(DBL_EPSILON)
-#define SHERPA_NCNN_MIN_LOG_DIFF_DOUBLE \
-  -36.0436533891171535515240975655615329742431640625
-
-template <typename T>
-struct LogAdd;
-
-template <>
-struct LogAdd<double> {
-  double operator()(double x, double y) const {
-    double diff;
-
-    if (x < y) {
-      diff = x - y;
-      x = y;
-    } else {
-      diff = y - x;
-    }
-    // diff is negative.  x is now the larger one.
-
-    if (diff >= SHERPA_NCNN_MIN_LOG_DIFF_DOUBLE) {
-      double res;
-      res = x + log1p(exp(diff));
-      return res;
-    }
-
-    return x;  // return the larger one.
-  }
-};
-
-template <>
-struct LogAdd<float> {
-  float operator()(float x, float y) const {
-    float diff;
-
-    if (x < y) {
-      diff = x - y;
-      x = y;
-    } else {
-      diff = y - x;
-    }
-    // diff is negative.  x is now the larger one.
-
-    if (diff >= SHERPA_NCNN_MIN_LOG_DIFF_DOUBLE) {
-      float res;
-      res = x + log1pf(expf(diff));
-      return res;
-    }
-
-    return x;  // return the larger one.
-  }
-};
-
-template <class T>
-void LogSoftmax(T *input, int32_t input_len) {
-  assert(input);
-
-  T m = *std::max_element(input, input + input_len);
-
-  T sum = 0.0;
-  for (int32_t i = 0; i < input_len; i++) {
-    sum += exp(input[i] - m);
-  }
-
-  T offset = m + log(sum);
-  for (int32_t i = 0; i < input_len; i++) {
-    input[i] -= offset;
-  }
-}
-
-template <class T>
-std::vector<int32_t> TopkIndex(const T *vec, int32_t size, int32_t topk) {
-  std::vector<int32_t> vec_index(size);
-  std::iota(vec_index.begin(), vec_index.end(), 0);
-
-  std::partial_sort(vec_index.begin(), vec_index.begin() + topk,
-                    vec_index.end(), [vec](int32_t index_1, int32_t index_2) {
-                      return vec[index_1] > vec[index_2];
-                    });
-
-  int32_t k_num = std::min<int32_t>(size, topk);
-  std::vector<int32_t> index(vec_index.begin(), vec_index.begin() + k_num);
-  return index;
-}
-
-}  // namespace sherpa_ncnn
-#endif  // SHERPA_NCNN_CSRC_MATH_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/meta-data.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/meta-data.h
deleted file mode 100644
index a4e13cfa..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/meta-data.h
+++ /dev/null
@@ -1,72 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_META_DATA_H_
-#define SHERPA_NCNN_CSRC_META_DATA_H_
-
-#include "layer.h"  // NOLINT
-#include "net.h"    // NOLINT
-
-namespace sherpa_ncnn {
-
-class MetaData : public ncnn::Layer {
- public:
-  int load_param(const ncnn::ParamDict &pd) override;
-
-  // arg0 is the model type:
-  //  1 - ConvEmformer
-  //  2 - Zipformer
-  //  3 - LSTM
-  //
-  //  arg15 is the model version, defaults to 0
-  int32_t arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7;
-  int32_t arg8, arg9, arg10, arg11, arg12, arg13, arg14, arg15;
-
-  ncnn::Mat arg16, arg17, arg18, arg19, arg20, arg21, arg22, arg23;
-
-  float arg24, arg25, arg26, arg27, arg28, arg29, arg30, arg31;
-};
-
-/*
-In encoder.ncnn.param, you can use
-
-SherpaMetaData sherpa_meta_data1 0 0 0=1 1=12 24=1.5
-
-For instace, suppose you have a encoder.ncnn.param looks like below:
-
-
-7767517
-1060 1342
-Input                    in0                      0 1 in0
-
-You can change it to
-
-7767517
-1061 1342
-SherpaMetaData            sherpa_meta_data1       0 0 0=1
-Input                    in0                      0 1 in0
-
-Note: You first need to change 1060 to 1061 since we add one layer
-
- */
-
-void RegisterMetaDataLayer(ncnn::Net &net);
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_META_DATA_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/microphone.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/microphone.h
deleted file mode 100644
index f5faede6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/microphone.h
+++ /dev/null
@@ -1,34 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_MICROPHONE_H_
-#define SHERPA_NCNN_CSRC_MICROPHONE_H_
-
-namespace sherpa_ncnn {
-
-class Microphone {
- public:
-  Microphone();
-  ~Microphone();
-
- private:
-};
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_MICROPHONE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/model.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/model.h
deleted file mode 100644
index b518fbad..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/model.h
+++ /dev/null
@@ -1,147 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_MODEL_H_
-#define SHERPA_NCNN_CSRC_MODEL_H_
-
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "net.h"  // NOLINT
-
-namespace sherpa_ncnn {
-
-struct ModelConfig {
-  std::string encoder_param;  // path to encoder.ncnn.param
-  std::string encoder_bin;    // path to encoder.ncnn.bin
-  std::string decoder_param;  // path to decoder.ncnn.param
-  std::string decoder_bin;    // path to decoder.ncnn.bin
-  std::string joiner_param;   // path to joiner.ncnn.param
-  std::string joiner_bin;     // path to joiner.ncnn.bin
-  std::string tokens;         // path to tokens.txt
-  bool use_vulkan_compute = true;
-
-  ncnn::Option encoder_opt;
-  ncnn::Option decoder_opt;
-  ncnn::Option joiner_opt;
-
-  std::string ToString() const;
-};
-
-class Model {
- public:
-  virtual ~Model() = default;
-
-  static void RegisterCustomLayers(ncnn::Net &net);
-
-  /** Create a model from a config. */
-  static std::unique_ptr<Model> Create(const ModelConfig &config);
-
-#if __ANDROID_API__ >= 9
-  static std::unique_ptr<Model> Create(AAssetManager *mgr,
-                                       const ModelConfig &config);
-#endif
-
-  // Return the encoder network.
-  virtual ncnn::Net &GetEncoder() = 0;
-
-  // Return the decoder network.
-  virtual ncnn::Net &GetDecoder() = 0;
-
-  // Return the joiner network.
-  virtual ncnn::Net &GetJoiner() = 0;
-
-  virtual std::vector<ncnn::Mat> GetEncoderInitStates() const = 0;
-
-  /** Run the encoder network.
-   *
-   * @param features  A 2-d mat of shape (num_frames, feature_dim).
-   *                  Note: features.w = feature_dim.
-   *                        features.h = num_frames.
-   * @param states It contains the states for the encoder network. Its exact
-   *               content is determined by the underlying network.
-   *
-   * @return Return a pair containing:
-   *   - encoder_out
-   *   - next_states
-   */
-  virtual std::pair<ncnn::Mat, std::vector<ncnn::Mat>> RunEncoder(
-      ncnn::Mat &features, const std::vector<ncnn::Mat> &states) = 0;
-
-  /** Run the encoder network with a user provided extractor.
-   */
-  virtual std::pair<ncnn::Mat, std::vector<ncnn::Mat>> RunEncoder(
-      ncnn::Mat &features, const std::vector<ncnn::Mat> &states,
-      ncnn::Extractor *extractor) = 0;
-
-  /** Run the decoder network.
-   *
-   * @param  decoder_input A mat of shape (context_size,). Note: Its underlying
-   *                       content consists of integers, though its type is
-   *                       float.
-   *
-   * @return Return a mat of shape (decoder_dim,)
-   */
-  virtual ncnn::Mat RunDecoder(ncnn::Mat &decoder_input) = 0;
-
-  /** Run the decoder network with a user provided extractor.
-   */
-  virtual ncnn::Mat RunDecoder(ncnn::Mat &decoder_input,
-                               ncnn::Extractor *extractor) = 0;
-
-  /** Run the joiner network.
-   *
-   * @param encoder_out  A mat of shape (encoder_dim,)
-   * @param decoder_out  A mat of shape (decoder_dim,)
-   *
-   * @return Return the joiner output which is of shape (vocab_size,)
-   */
-  virtual ncnn::Mat RunJoiner(ncnn::Mat &encoder_out,
-                              ncnn::Mat &decoder_out) = 0;
-
-  /** Run the joiner network with a user provided extractor.
-   */
-  virtual ncnn::Mat RunJoiner(ncnn::Mat &encoder_out, ncnn::Mat &decoder_out,
-                              ncnn::Extractor *extractor) = 0;
-
-  virtual int32_t ContextSize() const { return 2; }
-
-  virtual int32_t BlankId() const { return 0; }
-
-  // The encoder takes this number of frames as input
-  virtual int32_t Segment() const = 0;
-
-  // Advance the feature extractor by this number of frames after
-  // running the encoder network
-  virtual int32_t Offset() const = 0;
-
- protected:
-  static void InitNet(ncnn::Net &net, const std::string &param,
-                      const std::string &bin);
-
-#if __ANDROID_API__ >= 9
-  static void InitNet(AAssetManager *mgr, ncnn::Net &net,
-                      const std::string &param, const std::string &bin);
-#endif
-};
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/modified-beam-search-decoder.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/modified-beam-search-decoder.h
deleted file mode 100644
index e84a0c49..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/modified-beam-search-decoder.h
+++ /dev/null
@@ -1,55 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- * Copyright (c)  2022                     (Pingfeng Luo)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_MODIFIED_BEAM_SEARCH_DECODER_H_
-#define SHERPA_NCNN_CSRC_MODIFIED_BEAM_SEARCH_DECODER_H_
-
-#include <vector>
-
-#include "mat.h"  // NOLINT
-#include "sherpa-ncnn/csrc/decoder.h"
-#include "sherpa-ncnn/csrc/model.h"
-#include "sherpa-ncnn/csrc/stream.h"
-#include "sherpa-ncnn/csrc/context-graph.h"
-
-namespace sherpa_ncnn {
-
-class ModifiedBeamSearchDecoder : public Decoder {
- public:
-  ModifiedBeamSearchDecoder(Model *model, int32_t num_active_paths)
-      : model_(model), num_active_paths_(num_active_paths) {}
-
-  DecoderResult GetEmptyResult() const override;
-
-  void StripLeadingBlanks(DecoderResult *r) const override;
-
-  void Decode(ncnn::Mat encoder_out, DecoderResult *result) override;
-  void Decode(ncnn::Mat encoder_out, Stream *s, DecoderResult *result) override;
-
- private:
-  ncnn::Mat BuildDecoderInput(const std::vector<Hypothesis> &hyps) const;
-
- private:
-  Model *model_;  // not owned
-  int32_t num_active_paths_;
-};
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_MODIFIED_BEAM_SEARCH_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/poolingmodulenoproj.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/poolingmodulenoproj.h
deleted file mode 100644
index 96104902..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/poolingmodulenoproj.h
+++ /dev/null
@@ -1,43 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-#ifndef SHERPA_NCNN_CSRC_POOLINGMODULENOPROJ_H_
-#define SHERPA_NCNN_CSRC_POOLINGMODULENOPROJ_H_
-
-#include <utility>
-#include <vector>
-
-#include "layer.h"  // NOLINT
-#include "net.h"    // NOLINT
-
-namespace sherpa_ncnn {
-
-// used only by zipformer
-class PoolingModuleNoProj : public ncnn::Layer {
- public:
-  PoolingModuleNoProj();
-
-  int32_t forward(const std::vector<ncnn::Mat> &bottom_blobs,
-                  std::vector<ncnn::Mat> &top_blobs,
-                  const ncnn::Option &opt) const override;
-};
-
-void RegisterPoolingModuleNoProjLayer(ncnn::Net &net);
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_POOLINGMODULENOPROJ_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/recognizer.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/recognizer.h
deleted file mode 100644
index 45a63b6b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/recognizer.h
+++ /dev/null
@@ -1,126 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- * Copyright (c)  2022                     (Pingfeng Luo)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_RECOGNIZER_H_
-#define SHERPA_NCNN_CSRC_RECOGNIZER_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "sherpa-ncnn/csrc/endpoint.h"
-#include "sherpa-ncnn/csrc/features.h"
-#include "sherpa-ncnn/csrc/hypothesis.h"
-#include "sherpa-ncnn/csrc/model.h"
-#include "sherpa-ncnn/csrc/stream.h"
-#include "sherpa-ncnn/csrc/symbol-table.h"
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-namespace sherpa_ncnn {
-
-struct RecognitionResult {
-  std::string text;
-  std::vector<float> timestamps;
-  std::vector<int32_t> tokens;
-
-  // String based tokens
-  std::vector<std::string> stokens;
-
-  std::string ToString() const;
-};
-
-struct RecognizerConfig {
-  FeatureExtractorConfig feat_config;
-  ModelConfig model_config;
-  DecoderConfig decoder_config;
-  EndpointConfig endpoint_config;
-  bool enable_endpoint = false;
-
-  std::string hotwords_file;
-
-  /// used only for modified_beam_search
-  float hotwords_score = 1.5;
-
-  RecognizerConfig() = default;
-
-  RecognizerConfig(const FeatureExtractorConfig &feat_config,
-                   const ModelConfig &model_config,
-                   const DecoderConfig decoder_config,
-                   const EndpointConfig &endpoint_config, bool enable_endpoint,
-                   const std::string &hotwords_file, float hotwords_score)
-      : feat_config(feat_config),
-        model_config(model_config),
-        decoder_config(decoder_config),
-        endpoint_config(endpoint_config),
-        enable_endpoint(enable_endpoint),
-        hotwords_file(hotwords_file),
-        hotwords_score(hotwords_score) {}
-
-  std::string ToString() const;
-};
-
-class Recognizer {
- public:
-  explicit Recognizer(const RecognizerConfig &config);
-
-#if __ANDROID_API__ >= 9
-  Recognizer(AAssetManager *mgr, const RecognizerConfig &config);
-#endif
-
-  ~Recognizer();
-
-  /// Create a stream for decoding.
-  std::unique_ptr<Stream> CreateStream() const;
-
-  /**
-   * Return true if the given stream has enough frames for decoding.
-   * Return false otherwise
-   */
-  bool IsReady(Stream *s) const;
-
-  void DecodeStream(Stream *s) const;
-
-  // Return true if we detect an endpoint for this stream.
-  // Note: If this function returns true, you usually want to
-  // invoke Reset(s).
-  bool IsEndpoint(Stream *s) const;
-
-  // Clear the state of this stream. If IsEndpoint(s) returns true,
-  // after calling this function, IsEndpoint(s) will return false
-  void Reset(Stream *s) const;
-
-  RecognitionResult GetResult(Stream *s) const;
-
-  // Return the contained model
-  //
-  // The user should not free it.
-  const Model *GetModel() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_RECOGNIZER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/resample.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/resample.h
deleted file mode 100644
index 66a99422..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/resample.h
+++ /dev/null
@@ -1,144 +0,0 @@
-/**
- * Copyright     2013  Pegah Ghahremani
- *               2014  IMSL, PKU-HKUST (author: Wei Shi)
- *               2014  Yanqing Sun, Junjie Wang
- *               2014  Johns Hopkins University (author: Daniel Povey)
- * Copyright     2023  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-// this file is copied and modified from
-// kaldi/src/feat/resample.h
-#ifndef SHERPA_NCNN_CSRC_RESAMPLE_H_
-#define SHERPA_NCNN_CSRC_RESAMPLE_H_
-
-#include <cstdint>
-#include <vector>
-
-namespace sherpa_ncnn {
-
-/*
-   We require that the input and output sampling rate be specified as
-   integers, as this is an easy way to specify that their ratio be rational.
-*/
-
-class LinearResample {
- public:
-  /// Constructor.  We make the input and output sample rates integers, because
-  /// we are going to need to find a common divisor.  This should just remind
-  /// you that they need to be integers.  The filter cutoff needs to be less
-  /// than samp_rate_in_hz/2 and less than samp_rate_out_hz/2.  num_zeros
-  /// controls the sharpness of the filter, more == sharper but less efficient.
-  /// We suggest around 4 to 10 for normal use.
-  LinearResample(int32_t samp_rate_in_hz, int32_t samp_rate_out_hz,
-                 float filter_cutoff_hz, int32_t num_zeros);
-
-  /// Calling the function Reset() resets the state of the object prior to
-  /// processing a new signal; it is only necessary if you have called
-  /// Resample(x, x_size, false, y) for some signal, leading to a remainder of
-  /// the signal being called, but then abandon processing the signal before
-  /// calling Resample(x, x_size, true, y) for the last piece.  Call it
-  /// unnecessarily between signals will not do any harm.
-  void Reset();
-
-  /// This function does the resampling.  If you call it with flush == true and
-  /// you have never called it with flush == false, it just resamples the input
-  /// signal (it resizes the output to a suitable number of samples).
-  ///
-  /// You can also use this function to process a signal a piece at a time.
-  /// suppose you break it into piece1, piece2, ... pieceN.  You can call
-  /// \code{.cc}
-  /// Resample(piece1, piece1_size, false, &output1);
-  /// Resample(piece2, piece2_size, false, &output2);
-  /// Resample(piece3, piece3_size, true, &output3);
-  /// \endcode
-  /// If you call it with flush == false, it won't output the last few samples
-  /// but will remember them, so that if you later give it a second piece of
-  /// the input signal it can process it correctly.
-  /// If your most recent call to the object was with flush == false, it will
-  /// have internal state; you can remove this by calling Reset().
-  /// Empty input is acceptable.
-  void Resample(const float *input, int32_t input_dim, bool flush,
-                std::vector<float> *output);
-
-  //// Return the input and output sampling rates (for checks, for example)
-  int32_t GetInputSamplingRate() const { return samp_rate_in_; }
-  int32_t GetOutputSamplingRate() const { return samp_rate_out_; }
-
- private:
-  void SetIndexesAndWeights();
-
-  float FilterFunc(float) const;
-
-  /// This function outputs the number of output samples we will output
-  /// for a signal with "input_num_samp" input samples.  If flush == true,
-  /// we return the largest n such that
-  /// (n/samp_rate_out_) is in the interval [ 0, input_num_samp/samp_rate_in_ ),
-  /// and note that the interval is half-open.  If flush == false,
-  /// define window_width as num_zeros / (2.0 * filter_cutoff_);
-  /// we return the largest n such that (n/samp_rate_out_) is in the interval
-  /// [ 0, input_num_samp/samp_rate_in_ - window_width ).
-  int64_t GetNumOutputSamples(int64_t input_num_samp, bool flush) const;
-
-  /// Given an output-sample index, this function outputs to *first_samp_in the
-  /// first input-sample index that we have a weight on (may be negative),
-  /// and to *samp_out_wrapped the index into weights_ where we can get the
-  /// corresponding weights on the input.
-  inline void GetIndexes(int64_t samp_out, int64_t *first_samp_in,
-                         int32_t *samp_out_wrapped) const;
-
-  void SetRemainder(const float *input, int32_t input_dim);
-
- private:
-  // The following variables are provided by the user.
-  int32_t samp_rate_in_;
-  int32_t samp_rate_out_;
-  float filter_cutoff_;
-  int32_t num_zeros_;
-
-  int32_t input_samples_in_unit_;  ///< The number of input samples in the
-                                   ///< smallest repeating unit: num_samp_in_ =
-                                   ///< samp_rate_in_hz / Gcd(samp_rate_in_hz,
-                                   ///< samp_rate_out_hz)
-
-  int32_t output_samples_in_unit_;  ///< The number of output samples in the
-                                    ///< smallest repeating unit: num_samp_out_
-                                    ///< = samp_rate_out_hz /
-                                    ///< Gcd(samp_rate_in_hz, samp_rate_out_hz)
-
-  /// The first input-sample index that we sum over, for this output-sample
-  /// index.  May be negative; any truncation at the beginning is handled
-  /// separately.  This is just for the first few output samples, but we can
-  /// extrapolate the correct input-sample index for arbitrary output samples.
-  std::vector<int32_t> first_index_;
-
-  /// Weights on the input samples, for this output-sample index.
-  std::vector<std::vector<float>> weights_;
-
-  // the following variables keep track of where we are in a particular signal,
-  // if it is being provided over multiple calls to Resample().
-
-  int64_t input_sample_offset_;   ///< The number of input samples we have
-                                  ///< already received for this signal
-                                  ///< (including anything in remainder_)
-  int64_t output_sample_offset_;  ///< The number of samples we have already
-                                  ///< output for this signal.
-  std::vector<float> input_remainder_;  ///< A small trailing part of the
-                                        ///< previously seen input signal.
-};
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_RESAMPLE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/simpleupsample.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/simpleupsample.h
deleted file mode 100644
index 679dceff..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/simpleupsample.h
+++ /dev/null
@@ -1,52 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_SIMPLEUPSAMPLE_H_
-#define SHERPA_NCNN_CSRC_SIMPLEUPSAMPLE_H_
-
-#include <utility>
-
-#include "layer.h"  // NOLINT
-#include "net.h"    // NOLINT
-
-namespace sherpa_ncnn {
-
-class SimpleUpsample : public ncnn::Layer {
- public:
-  SimpleUpsample();
-
-  int32_t load_param(const ncnn::ParamDict &pd) override;
-
-  int32_t load_model(const ncnn::ModelBin &mb) override;
-
-  int32_t forward(const ncnn::Mat &bottom_blob, ncnn::Mat &top_blob,
-                  const ncnn::Option &opt) const override;
-
- public:
-  int32_t upsample;
-  int32_t num_channels;
-  int32_t bias_data_size;
-
-  ncnn::Mat bias;
-};
-
-void RegisterTensorSimpleUpsampleLayer(ncnn::Net &net);
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_SIMPLEUPSAMPLE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/stack.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/stack.h
deleted file mode 100644
index ee3e054c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/stack.h
+++ /dev/null
@@ -1,48 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_STACK_H_
-#define SHERPA_NCNN_CSRC_STACK_H_
-
-#include <utility>
-#include <vector>
-
-#include "layer.h"  // NOLINT
-#include "net.h"    // NOLINT
-
-namespace sherpa_ncnn {
-
-class Stack : public ncnn::Layer {
- public:
-  Stack();
-
-  int32_t load_param(const ncnn::ParamDict &pd) override;
-
-  int32_t forward(const std::vector<ncnn::Mat> &bottom_blobs,
-                  std::vector<ncnn::Mat> &top_blobs,
-                  const ncnn::Option &opt) const override;
-
- public:
-  int32_t axis;
-};
-
-void RegisterStackLayer(ncnn::Net &net);
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_STACK_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/stream.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/stream.h
deleted file mode 100644
index 45f7a71b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/stream.h
+++ /dev/null
@@ -1,108 +0,0 @@
-/**
- * Copyright (c)  2023  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_STREAM_H_
-#define SHERPA_NCNN_CSRC_STREAM_H_
-
-#include <memory>
-#include <vector>
-
-#include "sherpa-ncnn/csrc/context-graph.h"
-#include "sherpa-ncnn/csrc/decoder.h"
-#include "sherpa-ncnn/csrc/features.h"
-
-namespace sherpa_ncnn {
-class Stream {
- public:
-  explicit Stream(const FeatureExtractorConfig &config = {},
-                  ContextGraphPtr context_graph = nullptr);
-  ~Stream();
-
-  /**
-     @param sampling_rate The sampling_rate of the input waveform. We will
-                          do resample if it is different from
-                          config.sampling_rate.
-                          Caution: You MUST not use a different sampling rate
-                          across different calls for AcceptWaveform().
-     @param waveform Pointer to a 1-D array of size n
-     @param n Number of entries in waveform
-   */
-  void AcceptWaveform(int32_t sampling_rate, const float *waveform, int32_t n);
-
-  /**
-   * InputFinished() tells the class you won't be providing any
-   * more waveform.  This will help flush out the last frame or two
-   * of features, in the case where snip-edges == false; it also
-   * affects the return value of IsLastFrame().
-   */
-  void InputFinished();
-
-  int32_t NumFramesReady() const;
-
-  /** Note: IsLastFrame() will only ever return true if you have called
-   * InputFinished() (and this frame is the last frame).
-   */
-  bool IsLastFrame(int32_t frame) const;
-
-  /** Get n frames starting from the given frame index.
-   *
-   * @param frame_index  The starting frame index
-   * @param n  Number of frames to get.
-   * @return Return a 2-D tensor of shape (n, feature_dim).
-   *         which is flattened into a 1-D vector (flattened in in row major)
-   */
-  ncnn::Mat GetFrames(int32_t frame_index, int32_t n) const;
-
-  void Reset();
-
-  /**
-   * Finalize the decoding result. This is mainly for decoding with hotwords
-   * (i.e. providing context_graph). It will cancel the boosting score of the
-   * partial matching paths. For example, the hotword is "BCD", the path "ABC"
-   * gets boosting score of "BC" but it fails to match the whole hotword "BCD",
-   * so we have to cancel the scores of "BC" at the end.
-   */
-  void Finalize();
-
-  // Return a reference to the number of processed frames so far
-  // before subsampling..
-  // Initially, it is 0. It is always less than NumFramesReady().
-  //
-  // The returned reference is valid as long as this object is alive.
-  int32_t &GetNumProcessedFrames();
-
-  void SetResult(const DecoderResult &r);
-  DecoderResult &GetResult();
-
-  void SetStates(const std::vector<ncnn::Mat> &states);
-  std::vector<ncnn::Mat> &GetStates();
-  /**
-   * Get the context graph corresponding to this stream.
-   *
-   * @return Return the context graph for this stream.
-   */
-  const ContextGraphPtr &GetContextGraph() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_STREAM_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/symbol-table.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/symbol-table.h
deleted file mode 100644
index 768433bd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/symbol-table.h
+++ /dev/null
@@ -1,74 +0,0 @@
-/**
- * Copyright      2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_SYMBOL_TABLE_H_
-#define SHERPA_NCNN_CSRC_SYMBOL_TABLE_H_
-
-#include <string>
-#include <unordered_map>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-namespace sherpa_ncnn {
-
-/// It manages mapping between symbols and integer IDs.
-class SymbolTable {
- public:
-  SymbolTable() = default;
-  /// Construct a symbol table from a file.
-  /// Each line in the file contains two fields:
-  ///
-  ///    sym ID
-  ///
-  /// Fields are separated by space(s).
-  explicit SymbolTable(const std::string &filename);
-
-#if __ANDROID_API__ >= 9
-  SymbolTable(AAssetManager *mgr, const std::string &filename);
-#endif
-
-  /// Return a string representation of this symbol table
-  std::string ToString() const;
-
-  /// Return the symbol corresponding to the given ID.
-  const std::string &operator[](int32_t id) const;
-  /// Return the ID corresponding to the given symbol.
-  int32_t operator[](const std::string &sym) const;
-
-  /// Return true if there is a symbol with the given ID.
-  bool contains(int32_t id) const;
-
-  /// Return true if there is a given symbol in the symbol table.
-  bool contains(const std::string &sym) const;
-
- private:
-  void Init(std::istream &is);
-
- private:
-  std::unordered_map<std::string, int32_t> sym2id_;
-  std::unordered_map<int32_t, std::string> id2sym_;
-};
-
-std::ostream &operator<<(std::ostream &os, const SymbolTable &symbol_table);
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_SYMBOL_TABLE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/tensorasstrided.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/tensorasstrided.h
deleted file mode 100644
index 00d91eef..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/tensorasstrided.h
+++ /dev/null
@@ -1,48 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_TENSORASSTRIDED_H_
-#define SHERPA_NCNN_CSRC_TENSORASSTRIDED_H_
-
-#include <utility>
-
-#include "layer.h"  // NOLINT
-#include "net.h"    // NOLINT
-
-namespace sherpa_ncnn {
-
-class TensorAsStrided : public ncnn::Layer {
- public:
-  TensorAsStrided();
-
-  int32_t load_param(const ncnn::ParamDict &pd) override;
-
-  int32_t forward(const ncnn::Mat &bottom_blob, ncnn::Mat &top_blob,
-                  const ncnn::Option &opt) const override;
-
- public:
-  ncnn::Mat sizes;
-  ncnn::Mat strides;
-  int32_t storage_offset;
-};
-
-void RegisterTensorAsStridedLayer(ncnn::Net &net);
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_TENSORASSTRIDED_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/wave-reader.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/wave-reader.h
deleted file mode 100644
index 78e5448e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/wave-reader.h
+++ /dev/null
@@ -1,45 +0,0 @@
-/**
- * Copyright      2021  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef SHERPA_NCNN_CSRC_WAVE_READER_H_
-#define SHERPA_NCNN_CSRC_WAVE_READER_H_
-
-#include <istream>
-#include <string>
-#include <vector>
-
-namespace sherpa_ncnn {
-
-/** Read a wave file with expected sample rate.
-
-    @param filename Path to a wave file. It MUST be single channel, PCM encoded.
-    @param expected_sample_rate  Expected sample rate of the wave file. If the
-                               sample rate don't match, it throws an exception.
-    @param is_ok On return it is true if the reading succeeded; false otherwise.
-
-    @return Return wave samples normalized to the range [-1, 1).
- */
-std::vector<float> ReadWave(const std::string &filename,
-                            float expected_sample_rate, bool *is_ok);
-
-std::vector<float> ReadWave(std::istream &is, float expected_sample_rate,
-                            bool *is_ok);
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_WAVE_READER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/zipformer-model.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/zipformer-model.h
deleted file mode 100644
index 74c25815..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/csrc/zipformer-model.h
+++ /dev/null
@@ -1,111 +0,0 @@
-// sherpa-ncnn/csrc/zipformer-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_NCNN_CSRC_ZIPFORMER_MODEL_H_
-#define SHERPA_NCNN_CSRC_ZIPFORMER_MODEL_H_
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "net.h"  // NOLINT
-#include "sherpa-ncnn/csrc/model.h"
-
-namespace sherpa_ncnn {
-// Please refer to https://github.com/k2-fsa/icefall/pull/906
-// for how the model is converted from icefall to ncnn
-class ZipformerModel : public Model {
- public:
-  explicit ZipformerModel(const ModelConfig &config);
-#if __ANDROID_API__ >= 9
-  ZipformerModel(AAssetManager *mgr, const ModelConfig &config);
-#endif
-
-  ncnn::Net &GetEncoder() override { return encoder_; }
-  ncnn::Net &GetDecoder() override { return decoder_; }
-  ncnn::Net &GetJoiner() override { return joiner_; }
-
-  std::vector<ncnn::Mat> GetEncoderInitStates() const override;
-
-  std::pair<ncnn::Mat, std::vector<ncnn::Mat>> RunEncoder(
-      ncnn::Mat &features, const std::vector<ncnn::Mat> &states) override;
-
-  std::pair<ncnn::Mat, std::vector<ncnn::Mat>> RunEncoder(
-      ncnn::Mat &features, const std::vector<ncnn::Mat> &states,
-      ncnn::Extractor *extractor) override;
-
-  ncnn::Mat RunDecoder(ncnn::Mat &decoder_input) override;
-
-  ncnn::Mat RunDecoder(ncnn::Mat &decoder_input,
-                       ncnn::Extractor *extractor) override;
-
-  ncnn::Mat RunJoiner(ncnn::Mat &encoder_out, ncnn::Mat &decoder_out) override;
-
-  ncnn::Mat RunJoiner(ncnn::Mat &encoder_out, ncnn::Mat &decoder_out,
-                      ncnn::Extractor *extractor) override;
-
-  int32_t Segment() const override {
-    // pad_length 7, because the subsampling expression is
-    // ((x_len - 7) // 2 + 1)//2, we need to pad 7 frames
-    //
-    // decode chunk length before subsample is 32 frames
-    //
-    // So each segment is pad_length + decode_chunk_length = 7 + 32 = 39
-    return decode_chunk_length_ + pad_length_;
-  }
-
-  // Advance the feature extract by this number of frames after
-  // running the encoder network
-  int32_t Offset() const override { return decode_chunk_length_; }
-
- private:
-  void InitEncoder(const std::string &encoder_param,
-                   const std::string &encoder_bin);
-  void InitDecoder(const std::string &decoder_param,
-                   const std::string &decoder_bin);
-  void InitJoiner(const std::string &joiner_param,
-                  const std::string &joiner_bin);
-
-  void InitEncoderPostProcessing();
-
-#if __ANDROID_API__ >= 9
-  void InitEncoder(AAssetManager *mgr, const std::string &encoder_param,
-                   const std::string &encoder_bin);
-  void InitDecoder(AAssetManager *mgr, const std::string &decoder_param,
-                   const std::string &decoder_bin);
-  void InitJoiner(AAssetManager *mgr, const std::string &joiner_param,
-                  const std::string &joiner_bin);
-#endif
-
-  void InitEncoderInputOutputIndexes();
-  void InitDecoderInputOutputIndexes();
-  void InitJoinerInputOutputIndexes();
-
- private:
-  ncnn::Net encoder_;
-  ncnn::Net decoder_;
-  ncnn::Net joiner_;
-
-  int32_t decode_chunk_length_ = 32;  // arg1, before subsampling
-  int32_t num_left_chunks_ = 4;       // arg2
-  int32_t pad_length_ = 7;            // arg3
-
-  std::vector<int32_t> num_encoder_layers_;              // arg16
-  std::vector<int32_t> encoder_dims_;                    // arg17
-  std::vector<int32_t> attention_dims_;                  // arg18
-  std::vector<int32_t> zipformer_downsampling_factors_;  // arg19
-  std::vector<int32_t> cnn_module_kernels_;              // arg20
-
-  std::vector<int32_t> encoder_input_indexes_;
-  std::vector<int32_t> encoder_output_indexes_;
-
-  std::vector<int32_t> decoder_input_indexes_;
-  std::vector<int32_t> decoder_output_indexes_;
-
-  std::vector<int32_t> joiner_input_indexes_;
-  std::vector<int32_t> joiner_output_indexes_;
-};
-
-}  // namespace sherpa_ncnn
-
-#endif  // SHERPA_NCNN_CSRC_ZIPFORMER_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/datareader.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/datareader.h
deleted file mode 100644
index ed2aba3c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/datareader.h
+++ /dev/null
@@ -1,122 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_DATAREADER_H
-#define NCNN_DATAREADER_H
-
-#include "platform.h"
-#if NCNN_STDIO
-#include <stdio.h>
-#endif
-
-#if NCNN_PLATFORM_API
-#if __ANDROID_API__ >= 9
-#include <android/asset_manager.h>
-#endif
-#endif // NCNN_PLATFORM_API
-
-namespace ncnn {
-
-// data read wrapper
-class NCNN_EXPORT DataReader
-{
-public:
-    DataReader();
-    virtual ~DataReader();
-
-#if NCNN_STRING
-    // parse plain param text
-    // return 1 if scan success
-    virtual int scan(const char* format, void* p) const;
-#endif // NCNN_STRING
-
-    // read binary param and model data
-    // return bytes read
-    virtual size_t read(void* buf, size_t size) const;
-
-    // get model data reference
-    // return bytes referenced
-    virtual size_t reference(size_t size, const void** buf) const;
-};
-
-#if NCNN_STDIO
-class DataReaderFromStdioPrivate;
-class NCNN_EXPORT DataReaderFromStdio : public DataReader
-{
-public:
-    explicit DataReaderFromStdio(FILE* fp);
-    virtual ~DataReaderFromStdio();
-
-#if NCNN_STRING
-    virtual int scan(const char* format, void* p) const;
-#endif // NCNN_STRING
-    virtual size_t read(void* buf, size_t size) const;
-
-private:
-    DataReaderFromStdio(const DataReaderFromStdio&);
-    DataReaderFromStdio& operator=(const DataReaderFromStdio&);
-
-private:
-    DataReaderFromStdioPrivate* const d;
-};
-#endif // NCNN_STDIO
-
-class DataReaderFromMemoryPrivate;
-class NCNN_EXPORT DataReaderFromMemory : public DataReader
-{
-public:
-    explicit DataReaderFromMemory(const unsigned char*& mem);
-    virtual ~DataReaderFromMemory();
-
-#if NCNN_STRING
-    virtual int scan(const char* format, void* p) const;
-#endif // NCNN_STRING
-    virtual size_t read(void* buf, size_t size) const;
-    virtual size_t reference(size_t size, const void** buf) const;
-
-private:
-    DataReaderFromMemory(const DataReaderFromMemory&);
-    DataReaderFromMemory& operator=(const DataReaderFromMemory&);
-
-private:
-    DataReaderFromMemoryPrivate* const d;
-};
-
-#if NCNN_PLATFORM_API
-#if __ANDROID_API__ >= 9
-class DataReaderFromAndroidAssetPrivate;
-class NCNN_EXPORT DataReaderFromAndroidAsset : public DataReader
-{
-public:
-    explicit DataReaderFromAndroidAsset(AAsset* asset);
-    virtual ~DataReaderFromAndroidAsset();
-
-#if NCNN_STRING
-    virtual int scan(const char* format, void* p) const;
-#endif // NCNN_STRING
-    virtual size_t read(void* buf, size_t size) const;
-
-private:
-    DataReaderFromAndroidAsset(const DataReaderFromAndroidAsset&);
-    DataReaderFromAndroidAsset& operator=(const DataReaderFromAndroidAsset&);
-
-private:
-    DataReaderFromAndroidAssetPrivate* const d;
-};
-#endif // __ANDROID_API__ >= 9
-#endif // NCNN_PLATFORM_API
-
-} // namespace ncnn
-
-#endif // NCNN_DATAREADER_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/gpu.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/gpu.h
deleted file mode 100644
index a8d7ff91..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/gpu.h
+++ /dev/null
@@ -1,480 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2018 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_GPU_H
-#define NCNN_GPU_H
-
-#include "platform.h"
-
-#if NCNN_VULKAN
-
-#include "mat.h"
-
-namespace ncnn {
-
-// instance
-
-// Create VkInstance and initialize some objects that need to be calculated by GPU
-// Creates a VkInstance object, Checks the extended attributes supported by the Vulkan instance concerned,
-// Initializes, and creates Vulkan validation layers (if ENABLE_VALIDATION_LAYER is enabled),
-// Iterates over all supported physical devices, etc.
-NCNN_EXPORT int create_gpu_instance(const char* driver_path = 0);
-
-// Get global VkInstance variable
-// Must be called after create_gpu_instance() and before destroy_gpu_instance()
-NCNN_EXPORT VkInstance get_gpu_instance();
-
-// Destroy VkInstance object and free the memory of the associated object
-// Usually called in the destructor of the main program exit
-NCNN_EXPORT void destroy_gpu_instance();
-
-// vulkan core
-extern PFN_vkAllocateCommandBuffers vkAllocateCommandBuffers;
-extern PFN_vkAllocateDescriptorSets vkAllocateDescriptorSets;
-extern PFN_vkAllocateMemory vkAllocateMemory;
-extern PFN_vkBeginCommandBuffer vkBeginCommandBuffer;
-extern PFN_vkBindBufferMemory vkBindBufferMemory;
-extern PFN_vkBindImageMemory vkBindImageMemory;
-extern PFN_vkCmdBeginQuery vkCmdBeginQuery;
-extern PFN_vkCmdBindDescriptorSets vkCmdBindDescriptorSets;
-extern PFN_vkCmdBindIndexBuffer vkCmdBindIndexBuffer;
-extern PFN_vkCmdBindPipeline vkCmdBindPipeline;
-extern PFN_vkCmdCopyBuffer vkCmdCopyBuffer;
-extern PFN_vkCmdCopyBufferToImage vkCmdCopyBufferToImage;
-extern PFN_vkCmdCopyImage vkCmdCopyImage;
-extern PFN_vkCmdCopyImageToBuffer vkCmdCopyImageToBuffer;
-extern PFN_vkCmdCopyQueryPoolResults vkCmdCopyQueryPoolResults;
-extern PFN_vkCmdDispatch vkCmdDispatch;
-extern PFN_vkCmdDispatchIndirect vkCmdDispatchIndirect;
-extern PFN_vkCmdEndQuery vkCmdEndQuery;
-extern PFN_vkCmdExecuteCommands vkCmdExecuteCommands;
-extern PFN_vkCmdFillBuffer vkCmdFillBuffer;
-extern PFN_vkCmdPipelineBarrier vkCmdPipelineBarrier;
-extern PFN_vkCmdPushConstants vkCmdPushConstants;
-extern PFN_vkCmdResetQueryPool vkCmdResetQueryPool;
-extern PFN_vkCmdResolveImage vkCmdResolveImage;
-extern PFN_vkCmdUpdateBuffer vkCmdUpdateBuffer;
-extern PFN_vkCmdWriteTimestamp vkCmdWriteTimestamp;
-extern PFN_vkCreateBuffer vkCreateBuffer;
-extern PFN_vkCreateBufferView vkCreateBufferView;
-extern PFN_vkCreateCommandPool vkCreateCommandPool;
-extern PFN_vkCreateComputePipelines vkCreateComputePipelines;
-extern PFN_vkCreateDescriptorPool vkCreateDescriptorPool;
-extern PFN_vkCreateDescriptorSetLayout vkCreateDescriptorSetLayout;
-extern PFN_vkCreateDevice vkCreateDevice;
-extern PFN_vkCreateFence vkCreateFence;
-extern PFN_vkCreateImage vkCreateImage;
-extern PFN_vkCreateImageView vkCreateImageView;
-extern PFN_vkCreatePipelineCache vkCreatePipelineCache;
-extern PFN_vkCreatePipelineLayout vkCreatePipelineLayout;
-extern PFN_vkCreateQueryPool vkCreateQueryPool;
-extern PFN_vkCreateSampler vkCreateSampler;
-extern PFN_vkCreateSemaphore vkCreateSemaphore;
-extern PFN_vkCreateShaderModule vkCreateShaderModule;
-extern PFN_vkDestroyBuffer vkDestroyBuffer;
-extern PFN_vkDestroyBufferView vkDestroyBufferView;
-extern PFN_vkDestroyCommandPool vkDestroyCommandPool;
-extern PFN_vkDestroyDescriptorPool vkDestroyDescriptorPool;
-extern PFN_vkDestroyDescriptorSetLayout vkDestroyDescriptorSetLayout;
-extern PFN_vkDestroyDevice vkDestroyDevice;
-extern PFN_vkDestroyFence vkDestroyFence;
-extern PFN_vkDestroyImage vkDestroyImage;
-extern PFN_vkDestroyImageView vkDestroyImageView;
-extern PFN_vkDestroyInstance vkDestroyInstance;
-extern PFN_vkDestroyPipeline vkDestroyPipeline;
-extern PFN_vkDestroyPipelineCache vkDestroyPipelineCache;
-extern PFN_vkDestroyPipelineLayout vkDestroyPipelineLayout;
-extern PFN_vkDestroyQueryPool vkDestroyQueryPool;
-extern PFN_vkDestroySampler vkDestroySampler;
-extern PFN_vkDestroySemaphore vkDestroySemaphore;
-extern PFN_vkDestroyShaderModule vkDestroyShaderModule;
-extern PFN_vkDeviceWaitIdle vkDeviceWaitIdle;
-extern PFN_vkEndCommandBuffer vkEndCommandBuffer;
-extern PFN_vkEnumerateDeviceExtensionProperties vkEnumerateDeviceExtensionProperties;
-extern PFN_vkEnumerateDeviceLayerProperties vkEnumerateDeviceLayerProperties;
-extern PFN_vkEnumeratePhysicalDevices vkEnumeratePhysicalDevices;
-extern PFN_vkFlushMappedMemoryRanges vkFlushMappedMemoryRanges;
-extern PFN_vkFreeCommandBuffers vkFreeCommandBuffers;
-extern PFN_vkFreeDescriptorSets vkFreeDescriptorSets;
-extern PFN_vkFreeMemory vkFreeMemory;
-extern PFN_vkGetBufferMemoryRequirements vkGetBufferMemoryRequirements;
-extern PFN_vkGetDeviceMemoryCommitment vkGetDeviceMemoryCommitment;
-extern PFN_vkGetDeviceProcAddr vkGetDeviceProcAddr;
-extern PFN_vkGetDeviceQueue vkGetDeviceQueue;
-extern PFN_vkGetFenceStatus vkGetFenceStatus;
-extern PFN_vkGetImageMemoryRequirements vkGetImageMemoryRequirements;
-extern PFN_vkGetImageSubresourceLayout vkGetImageSubresourceLayout;
-extern PFN_vkGetPhysicalDeviceFeatures vkGetPhysicalDeviceFeatures;
-extern PFN_vkGetPhysicalDeviceFormatProperties vkGetPhysicalDeviceFormatProperties;
-extern PFN_vkGetPhysicalDeviceImageFormatProperties vkGetPhysicalDeviceImageFormatProperties;
-extern PFN_vkGetPhysicalDeviceMemoryProperties vkGetPhysicalDeviceMemoryProperties;
-extern PFN_vkGetPhysicalDeviceProperties vkGetPhysicalDeviceProperties;
-extern PFN_vkGetPhysicalDeviceQueueFamilyProperties vkGetPhysicalDeviceQueueFamilyProperties;
-extern PFN_vkGetPipelineCacheData vkGetPipelineCacheData;
-extern PFN_vkGetQueryPoolResults vkGetQueryPoolResults;
-extern PFN_vkInvalidateMappedMemoryRanges vkInvalidateMappedMemoryRanges;
-extern PFN_vkMapMemory vkMapMemory;
-extern PFN_vkMergePipelineCaches vkMergePipelineCaches;
-extern PFN_vkQueueSubmit vkQueueSubmit;
-extern PFN_vkQueueWaitIdle vkQueueWaitIdle;
-extern PFN_vkResetCommandBuffer vkResetCommandBuffer;
-extern PFN_vkResetCommandPool vkResetCommandPool;
-extern PFN_vkResetDescriptorPool vkResetDescriptorPool;
-extern PFN_vkResetFences vkResetFences;
-extern PFN_vkUnmapMemory vkUnmapMemory;
-extern PFN_vkUpdateDescriptorSets vkUpdateDescriptorSets;
-extern PFN_vkWaitForFences vkWaitForFences;
-
-// instance extension capability
-extern int support_VK_KHR_external_memory_capabilities;
-extern int support_VK_KHR_get_physical_device_properties2;
-extern int support_VK_KHR_get_surface_capabilities2;
-extern int support_VK_KHR_surface;
-extern int support_VK_EXT_debug_utils;
-extern int support_VK_EXT_validation_features;
-extern int support_VK_EXT_validation_flags;
-#if __ANDROID_API__ >= 26
-extern int support_VK_KHR_android_surface;
-#endif // __ANDROID_API__ >= 26
-
-// VK_KHR_cooperative_matrix
-extern PFN_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR;
-
-// VK_KHR_external_memory_capabilities
-extern PFN_vkGetPhysicalDeviceExternalBufferPropertiesKHR vkGetPhysicalDeviceExternalBufferPropertiesKHR;
-
-// VK_KHR_get_physical_device_properties2
-extern PFN_vkGetPhysicalDeviceFeatures2KHR vkGetPhysicalDeviceFeatures2KHR;
-extern PFN_vkGetPhysicalDeviceProperties2KHR vkGetPhysicalDeviceProperties2KHR;
-extern PFN_vkGetPhysicalDeviceFormatProperties2KHR vkGetPhysicalDeviceFormatProperties2KHR;
-extern PFN_vkGetPhysicalDeviceImageFormatProperties2KHR vkGetPhysicalDeviceImageFormatProperties2KHR;
-extern PFN_vkGetPhysicalDeviceQueueFamilyProperties2KHR vkGetPhysicalDeviceQueueFamilyProperties2KHR;
-extern PFN_vkGetPhysicalDeviceMemoryProperties2KHR vkGetPhysicalDeviceMemoryProperties2KHR;
-
-// VK_KHR_get_surface_capabilities2
-extern PFN_vkGetPhysicalDeviceSurfaceCapabilities2KHR vkGetPhysicalDeviceSurfaceCapabilities2KHR;
-extern PFN_vkGetPhysicalDeviceSurfaceFormats2KHR vkGetPhysicalDeviceSurfaceFormats2KHR;
-
-// VK_KHR_surface
-extern PFN_vkDestroySurfaceKHR vkDestroySurfaceKHR;
-extern PFN_vkGetPhysicalDeviceSurfaceSupportKHR vkGetPhysicalDeviceSurfaceSupportKHR;
-extern PFN_vkGetPhysicalDeviceSurfaceCapabilitiesKHR vkGetPhysicalDeviceSurfaceCapabilitiesKHR;
-extern PFN_vkGetPhysicalDeviceSurfaceFormatsKHR vkGetPhysicalDeviceSurfaceFormatsKHR;
-extern PFN_vkGetPhysicalDeviceSurfacePresentModesKHR vkGetPhysicalDeviceSurfacePresentModesKHR;
-
-#if __ANDROID_API__ >= 26
-// VK_KHR_android_surface
-extern PFN_vkCreateAndroidSurfaceKHR vkCreateAndroidSurfaceKHR;
-#endif // __ANDROID_API__ >= 26
-
-// VK_NV_cooperative_matrix
-extern PFN_vkGetPhysicalDeviceCooperativeMatrixPropertiesNV vkGetPhysicalDeviceCooperativeMatrixPropertiesNV;
-
-// get info
-NCNN_EXPORT int get_gpu_count();
-NCNN_EXPORT int get_default_gpu_index();
-
-class GpuInfoPrivate;
-class NCNN_EXPORT GpuInfo
-{
-public:
-    explicit GpuInfo();
-    virtual ~GpuInfo();
-
-    // vulkan physical device
-    VkPhysicalDevice physical_device() const;
-
-    // memory properties
-    const VkPhysicalDeviceMemoryProperties& physical_device_memory_properties() const;
-
-    // info
-    uint32_t api_version() const;
-    uint32_t driver_version() const;
-    uint32_t vendor_id() const;
-    uint32_t device_id() const;
-    const char* device_name() const;
-    uint8_t* pipeline_cache_uuid() const;
-
-    // 0 = discrete gpu
-    // 1 = integrated gpu
-    // 2 = virtual gpu
-    // 3 = cpu
-    int type() const;
-
-    // hardware limit
-    uint32_t max_shared_memory_size() const;
-    uint32_t max_workgroup_count_x() const;
-    uint32_t max_workgroup_count_y() const;
-    uint32_t max_workgroup_count_z() const;
-    uint32_t max_workgroup_invocations() const;
-    uint32_t max_workgroup_size_x() const;
-    uint32_t max_workgroup_size_y() const;
-    uint32_t max_workgroup_size_z() const;
-    size_t memory_map_alignment() const;
-    size_t buffer_offset_alignment() const;
-    size_t non_coherent_atom_size() const;
-    size_t buffer_image_granularity() const;
-    uint32_t max_image_dimension_1d() const;
-    uint32_t max_image_dimension_2d() const;
-    uint32_t max_image_dimension_3d() const;
-    float timestamp_period() const;
-
-    // runtime
-    uint32_t compute_queue_family_index() const;
-    uint32_t graphics_queue_family_index() const;
-    uint32_t transfer_queue_family_index() const;
-
-    uint32_t compute_queue_count() const;
-    uint32_t graphics_queue_count() const;
-    uint32_t transfer_queue_count() const;
-
-    // property
-    bool unified_compute_transfer_queue() const;
-
-    // subgroup
-    uint32_t subgroup_size() const;
-    bool support_subgroup_basic() const;
-    bool support_subgroup_vote() const;
-    bool support_subgroup_ballot() const;
-    bool support_subgroup_shuffle() const;
-
-    // bug is not feature
-    bool bug_storage_buffer_no_l1() const;
-    bool bug_corrupted_online_pipeline_cache() const;
-    bool bug_buffer_image_load_zero() const;
-
-    // but sometimes bug is a feature
-    bool bug_implicit_fp16_arithmetic() const;
-
-    // fp16 and int8 feature
-    bool support_fp16_packed() const;
-    bool support_fp16_storage() const;
-    bool support_fp16_uniform() const;
-    bool support_fp16_arithmetic() const;
-    bool support_int8_packed() const;
-    bool support_int8_storage() const;
-    bool support_int8_uniform() const;
-    bool support_int8_arithmetic() const;
-
-    // ycbcr conversion feature
-    bool support_ycbcr_conversion() const;
-
-    // cooperative matrix feature
-    bool support_cooperative_matrix() const;
-    bool support_cooperative_matrix_8_8_16() const;
-    bool support_cooperative_matrix_16_8_8() const;
-    bool support_cooperative_matrix_16_8_16() const;
-    bool support_cooperative_matrix_16_16_16() const;
-
-    // extension capability
-    int support_VK_KHR_8bit_storage() const;
-    int support_VK_KHR_16bit_storage() const;
-    int support_VK_KHR_bind_memory2() const;
-    int support_VK_KHR_buffer_device_address() const;
-    int support_VK_KHR_create_renderpass2() const;
-    int support_VK_KHR_cooperative_matrix() const;
-    int support_VK_KHR_dedicated_allocation() const;
-    int support_VK_KHR_descriptor_update_template() const;
-    int support_VK_KHR_external_memory() const;
-    int support_VK_KHR_get_memory_requirements2() const;
-    int support_VK_KHR_maintenance1() const;
-    int support_VK_KHR_maintenance2() const;
-    int support_VK_KHR_maintenance3() const;
-    int support_VK_KHR_multiview() const;
-    int support_VK_KHR_portability_subset() const;
-    int support_VK_KHR_push_descriptor() const;
-    int support_VK_KHR_sampler_ycbcr_conversion() const;
-    int support_VK_KHR_shader_float16_int8() const;
-    int support_VK_KHR_shader_float_controls() const;
-    int support_VK_KHR_storage_buffer_storage_class() const;
-    int support_VK_KHR_swapchain() const;
-    int support_VK_EXT_buffer_device_address() const;
-    int support_VK_EXT_descriptor_indexing() const;
-    int support_VK_EXT_memory_budget() const;
-    int support_VK_EXT_memory_priority() const;
-    int support_VK_EXT_queue_family_foreign() const;
-    int support_VK_AMD_device_coherent_memory() const;
-#if __ANDROID_API__ >= 26
-    int support_VK_ANDROID_external_memory_android_hardware_buffer() const;
-#endif // __ANDROID_API__ >= 26
-    int support_VK_NV_cooperative_matrix() const;
-
-private:
-    GpuInfo(const GpuInfo&);
-    GpuInfo& operator=(const GpuInfo&);
-
-private:
-    friend int create_gpu_instance(const char* driver_path);
-    GpuInfoPrivate* const d;
-};
-
-NCNN_EXPORT const GpuInfo& get_gpu_info(int device_index = get_default_gpu_index());
-
-class VkAllocator;
-class VkCompute;
-class Option;
-class PipelineCache;
-class VulkanDevicePrivate;
-class NCNN_EXPORT VulkanDevice
-{
-public:
-    VulkanDevice(int device_index = get_default_gpu_index());
-    ~VulkanDevice();
-
-    const GpuInfo& info;
-
-    VkDevice vkdevice() const;
-
-    VkShaderModule compile_shader_module(const uint32_t* spv_data, size_t spv_data_size) const;
-
-    // with fixed workgroup size
-    VkShaderModule compile_shader_module(const uint32_t* spv_data, size_t spv_data_size, uint32_t local_size_x, uint32_t local_size_y, uint32_t local_size_z) const;
-
-    // helper for creating pipeline
-    int create_descriptorset_layout(int binding_count, const int* binding_types, VkDescriptorSetLayout* descriptorset_layout) const;
-    int create_pipeline_layout(int push_constant_count, VkDescriptorSetLayout descriptorset_layout, VkPipelineLayout* pipeline_layout) const;
-    int create_pipeline(VkShaderModule shader_module, VkPipelineLayout pipeline_layout, const std::vector<vk_specialization_type>& specializations, VkPipeline* pipeline) const;
-    int create_descriptor_update_template(int binding_count, const int* binding_types, VkDescriptorSetLayout descriptorset_layout, VkPipelineLayout pipeline_layout, VkDescriptorUpdateTemplateKHR* descriptor_update_template) const;
-
-    uint32_t find_memory_index(uint32_t memory_type_bits, VkFlags required, VkFlags preferred, VkFlags preferred_not) const;
-    bool is_mappable(uint32_t memory_type_index) const;
-    bool is_coherent(uint32_t memory_type_index) const;
-
-    VkQueue acquire_queue(uint32_t queue_family_index) const;
-    void reclaim_queue(uint32_t queue_family_index, VkQueue queue) const;
-
-    // allocator on this device
-    VkAllocator* acquire_blob_allocator() const;
-    void reclaim_blob_allocator(VkAllocator* allocator) const;
-
-    VkAllocator* acquire_staging_allocator() const;
-    void reclaim_staging_allocator(VkAllocator* allocator) const;
-
-    // immutable sampler for texelfetch
-    const VkSampler* immutable_texelfetch_sampler() const;
-
-    // dummy buffer image
-    VkMat get_dummy_buffer() const;
-    VkImageMat get_dummy_image() const;
-    VkImageMat get_dummy_image_readonly() const;
-
-    // pipeline cache on this device
-    const PipelineCache* get_pipeline_cache() const;
-
-    // test image allocation
-    bool shape_support_image_storage(const Mat& shape) const;
-
-    // current gpu heap memory budget in MB
-    uint32_t get_heap_budget() const;
-
-    // utility operator
-    void convert_packing(const VkMat& src, VkMat& dst, int dst_elempack, VkCompute& cmd, const Option& opt) const;
-    void convert_packing(const VkImageMat& src, VkImageMat& dst, int dst_elempack, VkCompute& cmd, const Option& opt) const;
-    void convert_packing(const VkMat& src, VkImageMat& dst, int dst_elempack, VkCompute& cmd, const Option& opt) const;
-    void convert_packing(const VkImageMat& src, VkMat& dst, int dst_elempack, VkCompute& cmd, const Option& opt) const;
-
-    // VK_KHR_bind_memory2
-    PFN_vkBindBufferMemory2KHR vkBindBufferMemory2KHR;
-    PFN_vkBindImageMemory2KHR vkBindImageMemory2KHR;
-
-    // VK_KHR_buffer_device_address
-    PFN_vkGetBufferDeviceAddressKHR vkGetBufferDeviceAddressKHR;
-    PFN_vkGetBufferOpaqueCaptureAddressKHR vkGetBufferOpaqueCaptureAddressKHR;
-    PFN_vkGetDeviceMemoryOpaqueCaptureAddressKHR vkGetDeviceMemoryOpaqueCaptureAddressKHR;
-
-    // VK_KHR_descriptor_update_template
-    PFN_vkCreateDescriptorUpdateTemplateKHR vkCreateDescriptorUpdateTemplateKHR;
-    PFN_vkDestroyDescriptorUpdateTemplateKHR vkDestroyDescriptorUpdateTemplateKHR;
-    PFN_vkUpdateDescriptorSetWithTemplateKHR vkUpdateDescriptorSetWithTemplateKHR;
-
-    // VK_KHR_get_memory_requirements2
-    PFN_vkGetImageMemoryRequirements2KHR vkGetImageMemoryRequirements2KHR;
-    PFN_vkGetBufferMemoryRequirements2KHR vkGetBufferMemoryRequirements2KHR;
-
-    // VK_KHR_maintenance1
-    PFN_vkTrimCommandPoolKHR vkTrimCommandPoolKHR;
-
-    // VK_KHR_maintenance3
-    PFN_vkGetDescriptorSetLayoutSupportKHR vkGetDescriptorSetLayoutSupportKHR;
-
-    // VK_KHR_push_descriptor
-    PFN_vkCmdPushDescriptorSetWithTemplateKHR vkCmdPushDescriptorSetWithTemplateKHR;
-    PFN_vkCmdPushDescriptorSetKHR vkCmdPushDescriptorSetKHR;
-
-    // VK_KHR_sampler_ycbcr_conversion
-    PFN_vkCreateSamplerYcbcrConversionKHR vkCreateSamplerYcbcrConversionKHR;
-    PFN_vkDestroySamplerYcbcrConversionKHR vkDestroySamplerYcbcrConversionKHR;
-
-    // VK_KHR_swapchain
-    PFN_vkCreateSwapchainKHR vkCreateSwapchainKHR;
-    PFN_vkDestroySwapchainKHR vkDestroySwapchainKHR;
-    PFN_vkGetSwapchainImagesKHR vkGetSwapchainImagesKHR;
-    PFN_vkAcquireNextImageKHR vkAcquireNextImageKHR;
-    PFN_vkQueuePresentKHR vkQueuePresentKHR;
-
-    // VK_EXT_buffer_device_address
-    PFN_vkGetBufferDeviceAddressEXT vkGetBufferDeviceAddressEXT;
-
-#if __ANDROID_API__ >= 26
-    // VK_ANDROID_external_memory_android_hardware_buffer
-    PFN_vkGetAndroidHardwareBufferPropertiesANDROID vkGetAndroidHardwareBufferPropertiesANDROID;
-    PFN_vkGetMemoryAndroidHardwareBufferANDROID vkGetMemoryAndroidHardwareBufferANDROID;
-#endif // __ANDROID_API__ >= 26
-
-protected:
-    // device extension
-    int init_device_extension();
-
-private:
-    VulkanDevice(const VulkanDevice&);
-    VulkanDevice& operator=(const VulkanDevice&);
-
-private:
-    VulkanDevicePrivate* const d;
-};
-
-NCNN_EXPORT VulkanDevice* get_gpu_device(int device_index = get_default_gpu_index());
-
-// online spirv compilation
-NCNN_EXPORT int compile_spirv_module(const char* comp_string, const Option& opt, std::vector<uint32_t>& spirv);
-NCNN_EXPORT int compile_spirv_module(const char* comp_data, int comp_data_size, const Option& opt, std::vector<uint32_t>& spirv);
-NCNN_EXPORT int compile_spirv_module(int shader_type_index, const Option& opt, std::vector<uint32_t>& spirv);
-
-// info from spirv
-class NCNN_EXPORT ShaderInfo
-{
-public:
-    int specialization_count;
-    int binding_count;
-    int push_constant_count;
-
-    // 0 = null
-    // 1 = storage buffer
-    // 2 = storage image
-    // 3 = combined image sampler
-    int binding_types[16]; // 16 is large enough I think ...
-
-    int reserved_0;
-    int reserved_1;
-    int reserved_2;
-    int reserved_3;
-};
-
-NCNN_EXPORT int resolve_shader_info(const uint32_t* spv_data, size_t spv_data_size, ShaderInfo& shader_info);
-
-} // namespace ncnn
-
-#endif // NCNN_VULKAN
-
-#endif // NCNN_GPU_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer.h
deleted file mode 100644
index d44713de..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer.h
+++ /dev/null
@@ -1,230 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_LAYER_H
-#define NCNN_LAYER_H
-
-#include "mat.h"
-#include "modelbin.h"
-#include "option.h"
-#include "paramdict.h"
-#include "platform.h"
-
-#if NCNN_VULKAN
-#include "command.h"
-#include "pipeline.h"
-#endif // NCNN_VULKAN
-
-namespace ncnn {
-
-class NCNN_EXPORT Layer
-{
-public:
-    // empty
-    Layer();
-    // virtual destructor
-    virtual ~Layer();
-
-    // load layer specific parameter from parsed dict
-    // return 0 if success
-    virtual int load_param(const ParamDict& pd);
-
-    // load layer specific weight data from model binary
-    // return 0 if success
-    virtual int load_model(const ModelBin& mb);
-
-    // layer implementation specific setup
-    // return 0 if success
-    virtual int create_pipeline(const Option& opt);
-
-    // layer implementation specific clean
-    // return 0 if success
-    virtual int destroy_pipeline(const Option& opt);
-
-public:
-    // one input and one output blob
-    bool one_blob_only;
-
-    // support inplace inference
-    bool support_inplace;
-
-    // support vulkan compute
-    bool support_vulkan;
-
-    // accept input blob with packed storage
-    bool support_packing;
-
-    // accept bf16
-    bool support_bf16_storage;
-
-    // accept fp16
-    bool support_fp16_storage;
-
-    // accept int8
-    bool support_int8_storage;
-
-    // shader image storage
-    bool support_image_storage;
-
-    // shader tensor storage
-    bool support_tensor_storage;
-
-    bool support_reserved_00;
-
-    bool support_reserved_0;
-    bool support_reserved_1;
-    bool support_reserved_2;
-    bool support_reserved_3;
-    bool support_reserved_4;
-    bool support_reserved_5;
-    bool support_reserved_6;
-    bool support_reserved_7;
-    bool support_reserved_8;
-    bool support_reserved_9;
-
-    // feature disabled set
-    int featmask;
-
-public:
-    // implement inference
-    // return 0 if success
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    // implement inplace inference
-    // return 0 if success
-    virtual int forward_inplace(std::vector<Mat>& bottom_top_blobs, const Option& opt) const;
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-#if NCNN_VULKAN
-public:
-    // upload weight blob from host to device
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-public:
-    // implement inference
-    // return 0 if success
-    virtual int forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-    // implement inference
-    // return 0 if success
-    virtual int forward(const std::vector<VkImageMat>& bottom_blobs, std::vector<VkImageMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-    // implement inplace inference
-    // return 0 if success
-    virtual int forward_inplace(std::vector<VkMat>& bottom_top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-    // implement inplace inference
-    // return 0 if success
-    virtual int forward_inplace(std::vector<VkImageMat>& bottom_top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    // assigned immediately after creating this layer
-    const VulkanDevice* vkdev;
-#endif // NCNN_VULKAN
-
-public:
-    // custom user data
-    void* userdata;
-    // layer type index
-    int typeindex;
-#if NCNN_STRING
-    // layer type name
-    std::string type;
-    // layer name
-    std::string name;
-#endif // NCNN_STRING
-    // blob index which this layer needs as input
-    std::vector<int> bottoms;
-    // blob index which this layer produces as output
-    std::vector<int> tops;
-    // shape hint
-    std::vector<Mat> bottom_shapes;
-    std::vector<Mat> top_shapes;
-};
-
-// layer factory function
-typedef Layer* (*layer_creator_func)(void*);
-typedef void (*layer_destroyer_func)(Layer*, void*);
-
-struct layer_registry_entry
-{
-#if NCNN_STRING
-    // layer type name
-    const char* name;
-#endif // NCNN_STRING
-    // layer factory entry
-    layer_creator_func creator;
-};
-
-struct custom_layer_registry_entry
-{
-#if NCNN_STRING
-    // layer type name
-    const char* name;
-#endif // NCNN_STRING
-    // layer factory entry
-    layer_creator_func creator;
-    layer_destroyer_func destroyer;
-    void* userdata;
-};
-
-struct overwrite_builtin_layer_registry_entry
-{
-    // layer type index
-    int typeindex;
-    // layer factory entry
-    layer_creator_func creator;
-    layer_destroyer_func destroyer;
-    void* userdata;
-};
-
-#if NCNN_STRING
-// get layer type from type name
-NCNN_EXPORT int layer_to_index(const char* type);
-// create layer from type name
-NCNN_EXPORT Layer* create_layer(const char* type);
-NCNN_EXPORT Layer* create_layer_naive(const char* type);
-NCNN_EXPORT Layer* create_layer_cpu(const char* type);
-#if NCNN_VULKAN
-NCNN_EXPORT Layer* create_layer_vulkan(const char* type);
-#endif // NCNN_VULKAN
-#endif // NCNN_STRING
-// create layer from layer type
-NCNN_EXPORT Layer* create_layer(int index);
-NCNN_EXPORT Layer* create_layer_naive(int index);
-NCNN_EXPORT Layer* create_layer_cpu(int index);
-#if NCNN_VULKAN
-NCNN_EXPORT Layer* create_layer_vulkan(int index);
-#endif // NCNN_VULKAN
-
-#define DEFINE_LAYER_CREATOR(name)                          \
-    ::ncnn::Layer* name##_layer_creator(void* /*userdata*/) \
-    {                                                       \
-        return new name;                                    \
-    }
-
-#define DEFINE_LAYER_DESTROYER(name)                                      \
-    void name##_layer_destroyer(::ncnn::Layer* layer, void* /*userdata*/) \
-    {                                                                     \
-        delete layer;                                                     \
-    }
-
-} // namespace ncnn
-
-#endif // NCNN_LAYER_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/absval.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/absval.h
deleted file mode 100644
index 39783e4d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/absval.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ABSVAL_H
-#define LAYER_ABSVAL_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class AbsVal : public Layer
-{
-public:
-    AbsVal();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ABSVAL_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/argmax.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/argmax.h
deleted file mode 100644
index 05d5ca40..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/argmax.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ARGMAX_H
-#define LAYER_ARGMAX_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class ArgMax : public Layer
-{
-public:
-    ArgMax();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int out_max_val;
-    int topk;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ARGMAX_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/absval_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/absval_arm.h
deleted file mode 100644
index c1cea1df..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/absval_arm.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ABSVAL_ARM_H
-#define LAYER_ABSVAL_ARM_H
-
-#include "absval.h"
-
-namespace ncnn {
-
-class AbsVal_arm : public AbsVal
-{
-public:
-    AbsVal_arm();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ABSVAL_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/arm_activation.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/arm_activation.h
deleted file mode 100644
index aca3e574..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/arm_activation.h
+++ /dev/null
@@ -1,221 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef ARM_ACTIVATION_H
-#define ARM_ACTIVATION_H
-
-#include "fused_activation.h"
-
-#if __ARM_NEON
-#include <arm_neon.h>
-#include "neon_mathfun.h"
-
-static inline float32x4_t activation_ps(float32x4_t _v, int activation_type, const ncnn::Mat& activation_params)
-{
-    if (activation_type == 1)
-    {
-        const float32x4_t _zero = vdupq_n_f32(0.f);
-        _v = vmaxq_f32(_v, _zero);
-    }
-    else if (activation_type == 2)
-    {
-        const float32x4_t _zero = vdupq_n_f32(0.f);
-        const float32x4_t _slope = vdupq_n_f32(activation_params[0]);
-        const uint32x4_t _lemask = vcleq_f32(_v, _zero);
-        float32x4_t _ps = vmulq_f32(_v, _slope);
-        _v = vbslq_f32(_lemask, _ps, _v);
-    }
-    else if (activation_type == 3)
-    {
-        const float32x4_t _min = vdupq_n_f32(activation_params[0]);
-        const float32x4_t _max = vdupq_n_f32(activation_params[1]);
-        _v = vmaxq_f32(_v, _min);
-        _v = vminq_f32(_v, _max);
-    }
-    else if (activation_type == 4)
-    {
-        _v = sigmoid_ps(_v);
-    }
-    else if (activation_type == 5)
-    {
-        _v = vmulq_f32(_v, tanh_ps(log_ps(vaddq_f32(exp_ps(_v), vdupq_n_f32(1.f)))));
-    }
-    else if (activation_type == 6)
-    {
-        const float alpha = activation_params[0];
-        const float beta = activation_params[1];
-        const float32x4_t _zero = vdupq_n_f32(0.f);
-        const float32x4_t _one = vdupq_n_f32(1.f);
-        float32x4_t _ans = vdupq_n_f32(beta);
-        _ans = vmlaq_n_f32(_ans, _v, alpha);
-        _ans = vmaxq_f32(_ans, _zero);
-        _ans = vminq_f32(_ans, _one);
-        _v = vmulq_f32(_ans, _v);
-    }
-
-    return _v;
-}
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-#include "arm_usability.h"
-#include "neon_mathfun_fp16s.h"
-
-static inline __fp16 activation_ss_f16(__fp16 v, int activation_type, const ncnn::Mat& activation_params)
-{
-    if (activation_type == 1)
-    {
-        v = std::max(v, (__fp16)0.f);
-    }
-    else if (activation_type == 2)
-    {
-        const __fp16 slope = (__fp16)(activation_params[0]);
-        v = v > 0.f ? v : v * slope;
-    }
-    else if (activation_type == 3)
-    {
-        const __fp16 min = (__fp16)(activation_params[0]);
-        const __fp16 max = (__fp16)(activation_params[1]);
-        if (v < min)
-            v = min;
-        if (v > max)
-            v = max;
-    }
-    else if (activation_type == 4)
-    {
-        v = (__fp16)1.f / ((__fp16)1.f + (__fp16)expf(-v));
-    }
-    else if (activation_type == 5)
-    {
-        v = v * (__fp16)tanhf(logf(expf((float)v) + 1.f));
-    }
-    else if (activation_type == 6)
-    {
-        const __fp16 alpha = (__fp16)(activation_params[0]);
-        const __fp16 beta = (__fp16)(activation_params[1]);
-        const __fp16 lower = -beta / alpha;
-        const __fp16 upper = ((__fp16)1.f / alpha) + lower;
-        if (v < lower)
-            v = (__fp16)0.f;
-        else if (v > upper)
-            ;
-        else
-            v = v * (v * alpha + beta);
-    }
-
-    return v;
-}
-
-static inline float16x4_t activation_ps_f16(float16x4_t _v, int activation_type, const ncnn::Mat& activation_params)
-{
-    if (activation_type == 1)
-    {
-        const float16x4_t _zero = vdup_n_f16(0.f);
-        _v = vmax_f16(_v, _zero);
-    }
-    else if (activation_type == 2)
-    {
-        const float16x4_t _zero = vdup_n_f16(0.f);
-#if _MSC_VER
-        const float16x4_t _slope = vcvt_f16_f32(vdupq_n_f32(activation_params[0]));
-#else
-        const float16x4_t _slope = vdup_n_f16((__fp16)activation_params[0]);
-#endif
-        const uint16x4_t _lemask = vcle_f16(_v, _zero);
-        float16x4_t _ps = vmul_f16(_v, _slope);
-        _v = vbsl_f16(_lemask, _ps, _v);
-    }
-    else if (activation_type == 3)
-    {
-        const float16x4_t _min = vdup_n_f16((__fp16)activation_params[0]);
-        const float16x4_t _max = vdup_n_f16((__fp16)activation_params[1]);
-        _v = vmax_f16(_v, _min);
-        _v = vmin_f16(_v, _max);
-    }
-    else if (activation_type == 4)
-    {
-        _v = sigmoid_ps_f16(_v);
-    }
-    else if (activation_type == 5)
-    {
-        _v = vmul_f16(_v, tanh_ps_f16(log_ps_f16(vadd_f16(exp_ps_f16(_v), vdup_n_f16(1.f)))));
-    }
-    else if (activation_type == 6)
-    {
-        const __fp16 alpha = (__fp16)activation_params[0];
-        const __fp16 beta = (__fp16)activation_params[1];
-        const float16x4_t _zero = vdup_n_f16(0.f);
-        const float16x4_t _one = vdup_n_f16(1.f);
-        float16x4_t _ans = vdup_n_f16(beta);
-        _ans = vfma_n_f16(_ans, _v, alpha);
-        _ans = vmax_f16(_ans, _zero);
-        _ans = vmin_f16(_ans, _one);
-        _v = vmul_f16(_ans, _v);
-    }
-
-    return _v;
-}
-
-static inline float16x8_t activation_ps_f16(float16x8_t _v, int activation_type, const ncnn::Mat& activation_params)
-{
-    if (activation_type == 1)
-    {
-        const float16x8_t _zero = vdupq_n_f16(0.f);
-        _v = vmaxq_f16(_v, _zero);
-    }
-    else if (activation_type == 2)
-    {
-        const float16x8_t _zero = vdupq_n_f16(0.f);
-#if _MSC_VER
-        const float16x4_t _slope0 = vcvt_f16_f32(vdupq_n_f32(activation_params[0]));
-        const float16x8_t _slope = vcombine_f16(_slope0, _slope0);
-#else
-        const float16x8_t _slope = vdupq_n_f16((__fp16)activation_params[0]);
-#endif
-        const uint16x8_t _lemask = vcleq_f16(_v, _zero);
-        float16x8_t _ps = vmulq_f16(_v, _slope);
-        _v = vbslq_f16(_lemask, _ps, _v);
-    }
-    else if (activation_type == 3)
-    {
-        const float16x8_t _min = vdupq_n_f16((__fp16)activation_params[0]);
-        const float16x8_t _max = vdupq_n_f16((__fp16)activation_params[1]);
-        _v = vmaxq_f16(_v, _min);
-        _v = vminq_f16(_v, _max);
-    }
-    else if (activation_type == 4)
-    {
-        _v = sigmoid_ps_f16(_v);
-    }
-    else if (activation_type == 5)
-    {
-        _v = vmulq_f16(_v, tanh_ps_f16(log_ps_f16(vaddq_f16(exp_ps_f16(_v), vdupq_n_f16(1.f)))));
-    }
-    else if (activation_type == 6)
-    {
-        const __fp16 alpha_fp16 = (__fp16)activation_params[0];
-        const __fp16 beta_fp16 = (__fp16)activation_params[1];
-        const float16x8_t _zero = vdupq_n_f16(0.f);
-        const float16x8_t _one = vdupq_n_f16(1.f);
-        float16x8_t _ans = vdupq_n_f16(beta_fp16);
-        _ans = vfmaq_n_f16(_ans, _v, alpha_fp16);
-        _ans = vmaxq_f16(_ans, _zero);
-        _ans = vminq_f16(_ans, _one);
-        _v = vmulq_f16(_ans, _v);
-    }
-    return _v;
-}
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-#endif // __ARM_NEON
-
-#endif // ARM_ACTIVATION_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/arm_usability.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/arm_usability.h
deleted file mode 100644
index 5d6afb7b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/arm_usability.h
+++ /dev/null
@@ -1,770 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef ARM_USABILITY_H
-#define ARM_USABILITY_H
-
-static inline signed char float2int8(float v)
-{
-    int int32 = (int)roundf(v);
-    if (int32 > 127) return 127;
-    if (int32 < -127) return -127;
-    return (signed char)int32;
-}
-
-#if __ARM_NEON
-#include <arm_neon.h>
-
-static inline uint16x4_t float2bfloat(float32x4_t _v)
-{
-    return vshrn_n_u32(vreinterpretq_u32_f32(_v), 16);
-}
-static inline float32x4_t bfloat2float(uint16x4_t _v)
-{
-    return vreinterpretq_f32_u32(vshll_n_u16(_v, 16));
-}
-
-static inline int8x8_t float2int8(float32x4_t _vlow, float32x4_t _vhigh)
-{
-#if __aarch64__
-    int32x4_t _vlow32 = vcvtaq_s32_f32(_vlow);
-    int32x4_t _vhigh32 = vcvtaq_s32_f32(_vhigh);
-#else
-    // vcvtq_s32_f32 is round to zero
-    // simulate round to nearest via +/-0.5
-    float32x4_t _p5 = vdupq_n_f32(0.5f);
-    int32x4_t _signmask = vdupq_n_s32(1 << 31);
-    int32x4_t _signlow = vandq_s32(vreinterpretq_s32_f32(_vlow), _signmask);
-    int32x4_t _signhigh = vandq_s32(vreinterpretq_s32_f32(_vhigh), _signmask);
-    float32x4_t _p5low = vreinterpretq_f32_s32(vorrq_s32(vreinterpretq_s32_f32(_p5), _signlow));
-    float32x4_t _p5high = vreinterpretq_f32_s32(vorrq_s32(vreinterpretq_s32_f32(_p5), _signhigh));
-    float32x4_t _vlow5 = vaddq_f32(_vlow, _p5low);
-    float32x4_t _vhigh5 = vaddq_f32(_vhigh, _p5high);
-    int32x4_t _vlow32 = vcvtq_s32_f32(_vlow5);
-    int32x4_t _vhigh32 = vcvtq_s32_f32(_vhigh5);
-#endif
-    int16x8_t _v16 = vcombine_s16(vqmovn_s32(_vlow32), vqmovn_s32(_vhigh32));
-    int8x8_t _v8 = vqmovn_s16(_v16);
-    return vmax_s8(_v8, vdup_n_s8(-127));
-}
-
-static inline int8x8_t float2int8relu(float32x4_t _vlow, float32x4_t _vhigh)
-{
-#if __aarch64__
-    int32x4_t _vlow32 = vcvtaq_s32_f32(_vlow);
-    int32x4_t _vhigh32 = vcvtaq_s32_f32(_vhigh);
-#else
-    // vcvtq_s32_f32 is round to zero
-    // simulate round to nearest via +/-0.5
-    float32x4_t _p5 = vdupq_n_f32(0.5f);
-    int32x4_t _signmask = vdupq_n_s32(1 << 31);
-    int32x4_t _signlow = vandq_s32(vreinterpretq_s32_f32(_vlow), _signmask);
-    int32x4_t _signhigh = vandq_s32(vreinterpretq_s32_f32(_vhigh), _signmask);
-    float32x4_t _p5low = vreinterpretq_f32_s32(vorrq_s32(vreinterpretq_s32_f32(_p5), _signlow));
-    float32x4_t _p5high = vreinterpretq_f32_s32(vorrq_s32(vreinterpretq_s32_f32(_p5), _signhigh));
-    float32x4_t _vlow5 = vaddq_f32(_vlow, _p5low);
-    float32x4_t _vhigh5 = vaddq_f32(_vhigh, _p5high);
-    int32x4_t _vlow32 = vcvtq_s32_f32(_vlow5);
-    int32x4_t _vhigh32 = vcvtq_s32_f32(_vhigh5);
-#endif
-    int16x8_t _v16 = vcombine_s16(vqmovn_s32(_vlow32), vqmovn_s32(_vhigh32));
-    int8x8_t _v8 = vqmovn_s16(_v16);
-    return vmax_s8(_v8, vdup_n_s8(0));
-}
-
-static inline int8x8_t float2int8leakyrelu(float32x4_t _vlow, float32x4_t _vhigh, float32x4_t _slope)
-{
-    float32x4_t _vlow_leaky = vmulq_f32(_vlow, _slope);
-    float32x4_t _vhigh_leaky = vmulq_f32(_vhigh, _slope);
-#if __aarch64__
-    int32x4_t _vlow32 = vcvtaq_s32_f32(_vlow);
-    int32x4_t _vhigh32 = vcvtaq_s32_f32(_vhigh);
-    int32x4_t _vlow32_leaky = vcvtaq_s32_f32(_vlow_leaky);
-    int32x4_t _vhigh32_leaky = vcvtaq_s32_f32(_vhigh_leaky);
-#else
-    // vcvtq_s32_f32 is round to zero
-    // simulate round to nearest via +/-0.5
-    float32x4_t _p5 = vdupq_n_f32(0.5f);
-    int32x4_t _signmask = vdupq_n_s32(1 << 31);
-    int32x4_t _signlow = vandq_s32(vreinterpretq_s32_f32(_vlow), _signmask);
-    int32x4_t _signhigh = vandq_s32(vreinterpretq_s32_f32(_vhigh), _signmask);
-    float32x4_t _p5low = vreinterpretq_f32_s32(vorrq_s32(vreinterpretq_s32_f32(_p5), _signlow));
-    float32x4_t _p5high = vreinterpretq_f32_s32(vorrq_s32(vreinterpretq_s32_f32(_p5), _signhigh));
-    float32x4_t _vlow5 = vaddq_f32(_vlow, _p5low);
-    float32x4_t _vhigh5 = vaddq_f32(_vhigh, _p5high);
-    int32x4_t _vlow32 = vcvtq_s32_f32(_vlow5);
-    int32x4_t _vhigh32 = vcvtq_s32_f32(_vhigh5);
-
-    int32x4_t _signlow_leaky = vandq_s32(vreinterpretq_s32_f32(_vlow_leaky), _signmask);
-    int32x4_t _signhigh_leaky = vandq_s32(vreinterpretq_s32_f32(_vhigh_leaky), _signmask);
-    float32x4_t _p5low_leaky = vreinterpretq_f32_s32(vorrq_s32(vreinterpretq_s32_f32(_p5), _signlow_leaky));
-    float32x4_t _p5high_leaky = vreinterpretq_f32_s32(vorrq_s32(vreinterpretq_s32_f32(_p5), _signhigh_leaky));
-    float32x4_t _vlow5_leaky = vaddq_f32(_vlow_leaky, _p5low_leaky);
-    float32x4_t _vhigh5_leaky = vaddq_f32(_vhigh_leaky, _p5high_leaky);
-    int32x4_t _vlow32_leaky = vcvtq_s32_f32(_vlow5_leaky);
-    int32x4_t _vhigh32_leaky = vcvtq_s32_f32(_vhigh5_leaky);
-#endif
-    int16x8_t _v16 = vcombine_s16(vqmovn_s32(_vlow32), vqmovn_s32(_vhigh32));
-    int16x8_t _v16_leaky = vcombine_s16(vqmovn_s32(_vlow32_leaky), vqmovn_s32(_vhigh32_leaky));
-    int8x8_t _v8 = vqmovn_s16(_v16);
-    int8x8_t _v8_leaky = vqmovn_s16(_v16_leaky);
-    return vmax_s8(_v8, _v8_leaky);
-}
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-#ifdef _MSC_VER
-struct __fp16
-{
-    __fp16()
-    {
-        _u16 = 0;
-    }
-
-    __fp16(float f32)
-    {
-        _u16 = vget_lane_u16(vreinterpretq_u16_f16(vcvt_f16_f32(vdupq_n_f32(f32))), 0);
-    }
-
-    __fp16(__n16 n16)
-    {
-        _u16 = n16.n16_u16[0];
-    }
-
-    operator const float() const
-    {
-        return vgetq_lane_f32(vcvt_f32_f16(vreinterpretq_f16_u16(vdup_n_u16(_u16))), 0);
-    }
-
-    __fp16& operator+=(const __fp16& b)
-    {
-        float a = (float)*this;
-        float f32 = (a + (float)b);
-        _u16 = vget_lane_u16(vreinterpretq_u16_f16(vcvt_f16_f32(vdupq_n_f32(f32))), 0);
-        return *this;
-    }
-
-    __fp16& operator-=(const __fp16& b)
-    {
-        float a = (float)*this;
-        float f32 = (a - (float)b);
-        _u16 = vget_lane_u16(vreinterpretq_u16_f16(vcvt_f16_f32(vdupq_n_f32(f32))), 0);
-        return *this;
-    }
-
-    __fp16& operator*=(const __fp16& b)
-    {
-        float a = (float)*this;
-        float f32 = (a * (float)b);
-        _u16 = vget_lane_u16(vreinterpretq_u16_f16(vcvt_f16_f32(vdupq_n_f32(f32))), 0);
-        return *this;
-    }
-
-    __fp16& operator/=(const __fp16& b)
-    {
-        float a = (float)*this;
-        float f32 = (a / (float)b);
-        _u16 = vget_lane_u16(vreinterpretq_u16_f16(vcvt_f16_f32(vdupq_n_f32(f32))), 0);
-        return *this;
-    }
-
-    unsigned short _u16;
-};
-
-static inline __fp16 operator-(const __fp16& a)
-{
-    return __fp16(-(float)a);
-}
-static inline __fp16 operator+(const __fp16& a, const __fp16& b)
-{
-    return __fp16((float)a + (float)b);
-}
-static inline __fp16 operator-(const __fp16& a, const __fp16& b)
-{
-    return __fp16((float)a - (float)b);
-}
-static inline __fp16 operator*(const __fp16& a, const __fp16& b)
-{
-    return __fp16((float)a * (float)b);
-}
-static inline __fp16 operator/(const __fp16& a, const __fp16& b)
-{
-    return __fp16((float)a / (float)b);
-}
-
-static inline float16x4_t vdup_n_f16(const __fp16& f16)
-{
-    return vreinterpret_f16_u16(vdup_n_u16(f16._u16));
-}
-
-static inline float16x8_t vdupq_n_f16(const __fp16& f16)
-{
-    return vreinterpretq_f16_u16(vdupq_n_u16(f16._u16));
-}
-
-static inline __fp16 vmaxv_f16(float16x4_t a)
-{
-    return __fp16(vmaxvq_f32(vcvt_f32_f16(a)));
-}
-
-static inline __fp16 vmaxvq_f16(float16x8_t a)
-{
-    float x = vmaxvq_f32(vcvt_f32_f16(vget_low_f16(a)));
-    float y = vmaxvq_f32(vcvt_f32_f16(vget_high_f16(a)));
-    return __fp16(x > y ? x : y);
-}
-
-#define vld1q_f16 vld1q_u16
-#define vst1q_f16 vst1q_u16
-
-#define vld2_f16 vld2_u16
-#define vst2_f16 vst2_u16
-
-#define vld2q_f16 vld2q_u16
-#define vst2q_f16 vst2q_u16
-
-#define vld4_f16 vld4_u16
-#define vst4_f16 vst4_u16
-
-#define vld4q_f16 vld4q_u16
-#define vst4q_f16 vst4q_u16
-
-#define vld1q_dup_f16 vld1q_dup_u16
-
-#define vset_lane_f16(x, v, i)  vset_lane_u16(x._u16, (uint16x4_t)v, i)
-#define vsetq_lane_f16(x, v, i) vsetq_lane_u16(x._u16, (uint16x8_t)v, i)
-
-#define vfma_n_f16(va, vb, x)  vfma_f16(va, vb, vdup_n_f16(x))
-#define vfmaq_n_f16(va, vb, x) vfmaq_f16(va, vb, vdupq_n_f16(x))
-
-#endif
-
-static inline signed char float2int8(__fp16 v)
-{
-    int int32 = (int)roundf(v);
-    if (int32 > 127) return 127;
-    if (int32 < -127) return -127;
-    return (signed char)int32;
-}
-
-static inline int8x8_t float2int8(float16x8_t _v)
-{
-    int16x8_t _v16 = vcvtaq_s16_f16(_v);
-    int8x8_t _v8 = vqmovn_s16(_v16);
-    return vmax_s8(_v8, vdup_n_s8(-127));
-}
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-
-static inline void transpose4x4_u16(uint16x4_t& _r0, uint16x4_t& _r1, uint16x4_t& _r2, uint16x4_t& _r3)
-{
-    uint16x4x2_t _r01z = vzip_u16(_r0, _r1);
-    uint16x4x2_t _r23z = vzip_u16(_r2, _r3);
-    uint32x2x2_t _r01 = vzip_u32(vreinterpret_u32_u16(_r01z.val[0]), vreinterpret_u32_u16(_r23z.val[0]));
-    uint32x2x2_t _r23 = vzip_u32(vreinterpret_u32_u16(_r01z.val[1]), vreinterpret_u32_u16(_r23z.val[1]));
-    _r0 = vreinterpret_u16_u32(_r01.val[0]);
-    _r1 = vreinterpret_u16_u32(_r01.val[1]);
-    _r2 = vreinterpret_u16_u32(_r23.val[0]);
-    _r3 = vreinterpret_u16_u32(_r23.val[1]);
-}
-
-static inline void transpose4x8_u16(uint16x4_t& _r0, uint16x4_t& _r1, uint16x4_t& _r2, uint16x4_t& _r3, uint16x4_t& _r4, uint16x4_t& _r5, uint16x4_t& _r6, uint16x4_t& _r7)
-{
-    uint16x4x2_t _r01z = vzip_u16(_r0, _r1);
-    uint16x4x2_t _r23z = vzip_u16(_r2, _r3);
-    uint16x4x2_t _r45z = vzip_u16(_r4, _r5);
-    uint16x4x2_t _r67z = vzip_u16(_r6, _r7);
-    uint32x2x2_t _r01_0 = vzip_u32(vreinterpret_u32_u16(_r01z.val[0]), vreinterpret_u32_u16(_r23z.val[0]));
-    uint32x2x2_t _r23_0 = vzip_u32(vreinterpret_u32_u16(_r01z.val[1]), vreinterpret_u32_u16(_r23z.val[1]));
-    uint32x2x2_t _r01_1 = vzip_u32(vreinterpret_u32_u16(_r45z.val[0]), vreinterpret_u32_u16(_r67z.val[0]));
-    uint32x2x2_t _r23_1 = vzip_u32(vreinterpret_u32_u16(_r45z.val[1]), vreinterpret_u32_u16(_r67z.val[1]));
-    _r0 = vreinterpret_u16_u32(_r01_0.val[0]);
-    _r1 = vreinterpret_u16_u32(_r01_1.val[0]);
-    _r2 = vreinterpret_u16_u32(_r01_0.val[1]);
-    _r3 = vreinterpret_u16_u32(_r01_1.val[1]);
-    _r4 = vreinterpret_u16_u32(_r23_0.val[0]);
-    _r5 = vreinterpret_u16_u32(_r23_1.val[0]);
-    _r6 = vreinterpret_u16_u32(_r23_0.val[1]);
-    _r7 = vreinterpret_u16_u32(_r23_1.val[1]);
-}
-
-static inline void transpose4x12_u16(uint16x4_t& _r0, uint16x4_t& _r1, uint16x4_t& _r2, uint16x4_t& _r3, uint16x4_t& _r4, uint16x4_t& _r5, uint16x4_t& _r6, uint16x4_t& _r7, uint16x4_t& _r8, uint16x4_t& _r9, uint16x4_t& _ra, uint16x4_t& _rb)
-{
-    uint16x4x2_t _r01z = vzip_u16(_r0, _r1);
-    uint16x4x2_t _r23z = vzip_u16(_r2, _r3);
-    uint16x4x2_t _r45z = vzip_u16(_r4, _r5);
-    uint16x4x2_t _r67z = vzip_u16(_r6, _r7);
-    uint16x4x2_t _r89z = vzip_u16(_r8, _r9);
-    uint16x4x2_t _rabz = vzip_u16(_ra, _rb);
-    uint32x2x2_t _r01_0 = vzip_u32(vreinterpret_u32_u16(_r01z.val[0]), vreinterpret_u32_u16(_r23z.val[0]));
-    uint32x2x2_t _r23_0 = vzip_u32(vreinterpret_u32_u16(_r01z.val[1]), vreinterpret_u32_u16(_r23z.val[1]));
-    uint32x2x2_t _r01_1 = vzip_u32(vreinterpret_u32_u16(_r45z.val[0]), vreinterpret_u32_u16(_r67z.val[0]));
-    uint32x2x2_t _r23_1 = vzip_u32(vreinterpret_u32_u16(_r45z.val[1]), vreinterpret_u32_u16(_r67z.val[1]));
-    uint32x2x2_t _r01_2 = vzip_u32(vreinterpret_u32_u16(_r89z.val[0]), vreinterpret_u32_u16(_rabz.val[0]));
-    uint32x2x2_t _r23_2 = vzip_u32(vreinterpret_u32_u16(_r89z.val[1]), vreinterpret_u32_u16(_rabz.val[1]));
-    _r0 = vreinterpret_u16_u32(_r01_0.val[0]);
-    _r1 = vreinterpret_u16_u32(_r01_1.val[0]);
-    _r2 = vreinterpret_u16_u32(_r01_2.val[0]);
-    _r3 = vreinterpret_u16_u32(_r01_0.val[1]);
-    _r4 = vreinterpret_u16_u32(_r01_1.val[1]);
-    _r5 = vreinterpret_u16_u32(_r01_2.val[1]);
-    _r6 = vreinterpret_u16_u32(_r23_0.val[0]);
-    _r7 = vreinterpret_u16_u32(_r23_1.val[0]);
-    _r8 = vreinterpret_u16_u32(_r23_2.val[0]);
-    _r9 = vreinterpret_u16_u32(_r23_0.val[1]);
-    _ra = vreinterpret_u16_u32(_r23_1.val[1]);
-    _rb = vreinterpret_u16_u32(_r23_2.val[1]);
-}
-
-static inline void transpose8x4_u16(uint16x8_t& _r0, uint16x8_t& _r1, uint16x8_t& _r2, uint16x8_t& _r3)
-{
-    uint16x8x2_t _r01t = vzipq_u16(_r0, _r1);
-    uint16x8x2_t _r23t = vzipq_u16(_r2, _r3);
-    uint32x4x2_t _r01_0 = vzipq_u32(vreinterpretq_u32_u16(_r01t.val[0]), vreinterpretq_u32_u16(_r23t.val[0]));
-    uint32x4x2_t _r23_0 = vzipq_u32(vreinterpretq_u32_u16(_r01t.val[1]), vreinterpretq_u32_u16(_r23t.val[1]));
-    _r0 = vreinterpretq_u16_u32(_r01_0.val[0]);
-    _r1 = vreinterpretq_u16_u32(_r01_0.val[1]);
-    _r2 = vreinterpretq_u16_u32(_r23_0.val[0]);
-    _r3 = vreinterpretq_u16_u32(_r23_0.val[1]);
-}
-
-static inline void transpose8x8_u16(uint16x8_t& _r0, uint16x8_t& _r1, uint16x8_t& _r2, uint16x8_t& _r3, uint16x8_t& _r4, uint16x8_t& _r5, uint16x8_t& _r6, uint16x8_t& _r7)
-{
-    uint16x8x2_t _r01t = vzipq_u16(_r0, _r1);
-    uint16x8x2_t _r23t = vzipq_u16(_r2, _r3);
-    uint16x8x2_t _r45t = vzipq_u16(_r4, _r5);
-    uint16x8x2_t _r67t = vzipq_u16(_r6, _r7);
-    uint32x4x2_t _r01_0 = vzipq_u32(vreinterpretq_u32_u16(_r01t.val[0]), vreinterpretq_u32_u16(_r23t.val[0]));
-    uint32x4x2_t _r23_0 = vzipq_u32(vreinterpretq_u32_u16(_r01t.val[1]), vreinterpretq_u32_u16(_r23t.val[1]));
-    uint32x4x2_t _r01_1 = vzipq_u32(vreinterpretq_u32_u16(_r45t.val[0]), vreinterpretq_u32_u16(_r67t.val[0]));
-    uint32x4x2_t _r23_1 = vzipq_u32(vreinterpretq_u32_u16(_r45t.val[1]), vreinterpretq_u32_u16(_r67t.val[1]));
-    _r0 = vreinterpretq_u16_u32(vcombine_u32(vget_low_u32(_r01_0.val[0]), vget_low_u32(_r01_1.val[0])));
-    _r1 = vreinterpretq_u16_u32(vcombine_u32(vget_high_u32(_r01_0.val[0]), vget_high_u32(_r01_1.val[0])));
-    _r2 = vreinterpretq_u16_u32(vcombine_u32(vget_low_u32(_r01_0.val[1]), vget_low_u32(_r01_1.val[1])));
-    _r3 = vreinterpretq_u16_u32(vcombine_u32(vget_high_u32(_r01_0.val[1]), vget_high_u32(_r01_1.val[1])));
-    _r4 = vreinterpretq_u16_u32(vcombine_u32(vget_low_u32(_r23_0.val[0]), vget_low_u32(_r23_1.val[0])));
-    _r5 = vreinterpretq_u16_u32(vcombine_u32(vget_high_u32(_r23_0.val[0]), vget_high_u32(_r23_1.val[0])));
-    _r6 = vreinterpretq_u16_u32(vcombine_u32(vget_low_u32(_r23_0.val[1]), vget_low_u32(_r23_1.val[1])));
-    _r7 = vreinterpretq_u16_u32(vcombine_u32(vget_high_u32(_r23_0.val[1]), vget_high_u32(_r23_1.val[1])));
-}
-
-static inline void transpose8x12_u16(uint16x8_t& _r0, uint16x8_t& _r1, uint16x8_t& _r2, uint16x8_t& _r3, uint16x8_t& _r4, uint16x8_t& _r5, uint16x8_t& _r6, uint16x8_t& _r7, uint16x8_t& _r8, uint16x8_t& _r9, uint16x8_t& _ra, uint16x8_t& _rb)
-{
-    uint16x8x2_t _r01t = vzipq_u16(_r0, _r1);
-    uint16x8x2_t _r23t = vzipq_u16(_r2, _r3);
-    uint16x8x2_t _r45t = vzipq_u16(_r4, _r5);
-    uint16x8x2_t _r67t = vzipq_u16(_r6, _r7);
-    uint16x8x2_t _r89t = vzipq_u16(_r8, _r9);
-    uint16x8x2_t _rabt = vzipq_u16(_ra, _rb);
-    uint32x4x2_t _r01_0 = vzipq_u32(vreinterpretq_u32_u16(_r01t.val[0]), vreinterpretq_u32_u16(_r23t.val[0]));
-    uint32x4x2_t _r23_0 = vzipq_u32(vreinterpretq_u32_u16(_r01t.val[1]), vreinterpretq_u32_u16(_r23t.val[1]));
-    uint32x4x2_t _r01_1 = vzipq_u32(vreinterpretq_u32_u16(_r45t.val[0]), vreinterpretq_u32_u16(_r67t.val[0]));
-    uint32x4x2_t _r23_1 = vzipq_u32(vreinterpretq_u32_u16(_r45t.val[1]), vreinterpretq_u32_u16(_r67t.val[1]));
-    uint32x4x2_t _r01_2 = vzipq_u32(vreinterpretq_u32_u16(_r89t.val[0]), vreinterpretq_u32_u16(_rabt.val[0]));
-    uint32x4x2_t _r23_2 = vzipq_u32(vreinterpretq_u32_u16(_r89t.val[1]), vreinterpretq_u32_u16(_rabt.val[1]));
-    _r0 = vreinterpretq_u16_u32(vcombine_u32(vget_low_u32(_r01_0.val[0]), vget_low_u32(_r01_1.val[0])));
-    _r1 = vreinterpretq_u16_u32(vcombine_u32(vget_low_u32(_r01_2.val[0]), vget_high_u32(_r01_0.val[0])));
-    _r2 = vreinterpretq_u16_u32(vcombine_u32(vget_high_u32(_r01_1.val[0]), vget_high_u32(_r01_2.val[0])));
-    _r3 = vreinterpretq_u16_u32(vcombine_u32(vget_low_u32(_r01_0.val[1]), vget_low_u32(_r01_1.val[1])));
-    _r4 = vreinterpretq_u16_u32(vcombine_u32(vget_low_u32(_r01_2.val[1]), vget_high_u32(_r01_0.val[1])));
-    _r5 = vreinterpretq_u16_u32(vcombine_u32(vget_high_u32(_r01_1.val[1]), vget_high_u32(_r01_2.val[1])));
-    _r6 = vreinterpretq_u16_u32(vcombine_u32(vget_low_u32(_r23_0.val[0]), vget_low_u32(_r23_1.val[0])));
-    _r7 = vreinterpretq_u16_u32(vcombine_u32(vget_low_u32(_r23_2.val[0]), vget_high_u32(_r23_0.val[0])));
-    _r8 = vreinterpretq_u16_u32(vcombine_u32(vget_high_u32(_r23_1.val[0]), vget_high_u32(_r23_2.val[0])));
-    _r9 = vreinterpretq_u16_u32(vcombine_u32(vget_low_u32(_r23_0.val[1]), vget_low_u32(_r23_1.val[1])));
-    _ra = vreinterpretq_u16_u32(vcombine_u32(vget_low_u32(_r23_2.val[1]), vget_high_u32(_r23_0.val[1])));
-    _rb = vreinterpretq_u16_u32(vcombine_u32(vget_high_u32(_r23_1.val[1]), vget_high_u32(_r23_2.val[1])));
-}
-
-static inline void transpose4x4_ps(float32x4_t& _r0, float32x4_t& _r1, float32x4_t& _r2, float32x4_t& _r3)
-{
-    float32x4x2_t _r01z = vzipq_f32(_r0, _r1);
-    float32x4x2_t _r23z = vzipq_f32(_r2, _r3);
-    _r0 = vcombine_f32(vget_low_f32(_r01z.val[0]), vget_low_f32(_r23z.val[0]));
-    _r1 = vcombine_f32(vget_high_f32(_r01z.val[0]), vget_high_f32(_r23z.val[0]));
-    _r2 = vcombine_f32(vget_low_f32(_r01z.val[1]), vget_low_f32(_r23z.val[1]));
-    _r3 = vcombine_f32(vget_high_f32(_r01z.val[1]), vget_high_f32(_r23z.val[1]));
-}
-
-static inline void transpose4x8_ps(float32x4_t& _r0, float32x4_t& _r1, float32x4_t& _r2, float32x4_t& _r3, float32x4_t& _r4, float32x4_t& _r5, float32x4_t& _r6, float32x4_t& _r7)
-{
-    float32x4x2_t _r01z = vzipq_f32(_r0, _r1);
-    float32x4x2_t _r23z = vzipq_f32(_r2, _r3);
-    float32x4x2_t _r45z = vzipq_f32(_r4, _r5);
-    float32x4x2_t _r67z = vzipq_f32(_r6, _r7);
-    _r0 = vcombine_f32(vget_low_f32(_r01z.val[0]), vget_low_f32(_r23z.val[0]));
-    _r1 = vcombine_f32(vget_low_f32(_r45z.val[0]), vget_low_f32(_r67z.val[0]));
-    _r2 = vcombine_f32(vget_high_f32(_r01z.val[0]), vget_high_f32(_r23z.val[0]));
-    _r3 = vcombine_f32(vget_high_f32(_r45z.val[0]), vget_high_f32(_r67z.val[0]));
-    _r4 = vcombine_f32(vget_low_f32(_r01z.val[1]), vget_low_f32(_r23z.val[1]));
-    _r5 = vcombine_f32(vget_low_f32(_r45z.val[1]), vget_low_f32(_r67z.val[1]));
-    _r6 = vcombine_f32(vget_high_f32(_r01z.val[1]), vget_high_f32(_r23z.val[1]));
-    _r7 = vcombine_f32(vget_high_f32(_r45z.val[1]), vget_high_f32(_r67z.val[1]));
-}
-
-static inline void transpose4x12_ps(float32x4_t& _r0, float32x4_t& _r1, float32x4_t& _r2, float32x4_t& _r3, float32x4_t& _r4, float32x4_t& _r5, float32x4_t& _r6, float32x4_t& _r7, float32x4_t& _r8, float32x4_t& _r9, float32x4_t& _ra, float32x4_t& _rb)
-{
-    float32x4x2_t _r01z = vzipq_f32(_r0, _r1);
-    float32x4x2_t _r23z = vzipq_f32(_r2, _r3);
-    float32x4x2_t _r45z = vzipq_f32(_r4, _r5);
-    float32x4x2_t _r67z = vzipq_f32(_r6, _r7);
-    float32x4x2_t _r89z = vzipq_f32(_r8, _r9);
-    float32x4x2_t _rabz = vzipq_f32(_ra, _rb);
-    _r0 = vcombine_f32(vget_low_f32(_r01z.val[0]), vget_low_f32(_r23z.val[0]));
-    _r1 = vcombine_f32(vget_low_f32(_r45z.val[0]), vget_low_f32(_r67z.val[0]));
-    _r2 = vcombine_f32(vget_low_f32(_r89z.val[0]), vget_low_f32(_rabz.val[0]));
-    _r3 = vcombine_f32(vget_high_f32(_r01z.val[0]), vget_high_f32(_r23z.val[0]));
-    _r4 = vcombine_f32(vget_high_f32(_r45z.val[0]), vget_high_f32(_r67z.val[0]));
-    _r5 = vcombine_f32(vget_high_f32(_r89z.val[0]), vget_high_f32(_rabz.val[0]));
-    _r6 = vcombine_f32(vget_low_f32(_r01z.val[1]), vget_low_f32(_r23z.val[1]));
-    _r7 = vcombine_f32(vget_low_f32(_r45z.val[1]), vget_low_f32(_r67z.val[1]));
-    _r8 = vcombine_f32(vget_low_f32(_r89z.val[1]), vget_low_f32(_rabz.val[1]));
-    _r9 = vcombine_f32(vget_high_f32(_r01z.val[1]), vget_high_f32(_r23z.val[1]));
-    _ra = vcombine_f32(vget_high_f32(_r45z.val[1]), vget_high_f32(_r67z.val[1]));
-    _rb = vcombine_f32(vget_high_f32(_r89z.val[1]), vget_high_f32(_rabz.val[1]));
-}
-
-static inline void transpose8x4_ps(float32x4_t& _r0l, float32x4_t& _r0h,
-                                   float32x4_t& _r1l, float32x4_t& _r1h,
-                                   float32x4_t& _r2l, float32x4_t& _r2h,
-                                   float32x4_t& _r3l, float32x4_t& _r3h)
-{
-    float32x4x2_t _r01lz = vzipq_f32(_r0l, _r1l);
-    float32x4x2_t _r23lz = vzipq_f32(_r2l, _r3l);
-    float32x4x2_t _r01hz = vzipq_f32(_r0h, _r1h);
-    float32x4x2_t _r23hz = vzipq_f32(_r2h, _r3h);
-    _r0l = vcombine_f32(vget_low_f32(_r01lz.val[0]), vget_low_f32(_r23lz.val[0]));
-    _r0h = vcombine_f32(vget_high_f32(_r01lz.val[0]), vget_high_f32(_r23lz.val[0]));
-    _r1l = vcombine_f32(vget_low_f32(_r01lz.val[1]), vget_low_f32(_r23lz.val[1]));
-    _r1h = vcombine_f32(vget_high_f32(_r01lz.val[1]), vget_high_f32(_r23lz.val[1]));
-    _r2l = vcombine_f32(vget_low_f32(_r01hz.val[0]), vget_low_f32(_r23hz.val[0]));
-    _r2h = vcombine_f32(vget_high_f32(_r01hz.val[0]), vget_high_f32(_r23hz.val[0]));
-    _r3l = vcombine_f32(vget_low_f32(_r01hz.val[1]), vget_low_f32(_r23hz.val[1]));
-    _r3h = vcombine_f32(vget_high_f32(_r01hz.val[1]), vget_high_f32(_r23hz.val[1]));
-}
-
-static inline void transpose12x4_ps(float32x4_t& _r0l, float32x4_t& _r0m, float32x4_t& _r0h,
-                                    float32x4_t& _r1l, float32x4_t& _r1m, float32x4_t& _r1h,
-                                    float32x4_t& _r2l, float32x4_t& _r2m, float32x4_t& _r2h,
-                                    float32x4_t& _r3l, float32x4_t& _r3m, float32x4_t& _r3h)
-{
-    float32x4x2_t _r01lz = vzipq_f32(_r0l, _r1l);
-    float32x4x2_t _r23lz = vzipq_f32(_r2l, _r3l);
-    float32x4x2_t _r01mz = vzipq_f32(_r0m, _r1m);
-    float32x4x2_t _r23mz = vzipq_f32(_r2m, _r3m);
-    float32x4x2_t _r01hz = vzipq_f32(_r0h, _r1h);
-    float32x4x2_t _r23hz = vzipq_f32(_r2h, _r3h);
-    _r0l = vcombine_f32(vget_low_f32(_r01lz.val[0]), vget_low_f32(_r23lz.val[0]));
-    _r0m = vcombine_f32(vget_high_f32(_r01lz.val[0]), vget_high_f32(_r23lz.val[0]));
-    _r0h = vcombine_f32(vget_low_f32(_r01lz.val[1]), vget_low_f32(_r23lz.val[1]));
-    _r1l = vcombine_f32(vget_high_f32(_r01lz.val[1]), vget_high_f32(_r23lz.val[1]));
-    _r1m = vcombine_f32(vget_low_f32(_r01mz.val[0]), vget_low_f32(_r23mz.val[0]));
-    _r1h = vcombine_f32(vget_high_f32(_r01mz.val[0]), vget_high_f32(_r23mz.val[0]));
-    _r2l = vcombine_f32(vget_low_f32(_r01mz.val[1]), vget_low_f32(_r23mz.val[1]));
-    _r2m = vcombine_f32(vget_high_f32(_r01mz.val[1]), vget_high_f32(_r23mz.val[1]));
-    _r2h = vcombine_f32(vget_low_f32(_r01hz.val[0]), vget_low_f32(_r23hz.val[0]));
-    _r3l = vcombine_f32(vget_high_f32(_r01hz.val[0]), vget_high_f32(_r23hz.val[0]));
-    _r3m = vcombine_f32(vget_low_f32(_r01hz.val[1]), vget_low_f32(_r23hz.val[1]));
-    _r3h = vcombine_f32(vget_high_f32(_r01hz.val[1]), vget_high_f32(_r23hz.val[1]));
-}
-
-#if __aarch64__
-static inline void transpose8x8_ps(float32x4_t& _r0l, float32x4_t& _r0h,
-                                   float32x4_t& _r1l, float32x4_t& _r1h,
-                                   float32x4_t& _r2l, float32x4_t& _r2h,
-                                   float32x4_t& _r3l, float32x4_t& _r3h,
-                                   float32x4_t& _r4l, float32x4_t& _r4h,
-                                   float32x4_t& _r5l, float32x4_t& _r5h,
-                                   float32x4_t& _r6l, float32x4_t& _r6h,
-                                   float32x4_t& _r7l, float32x4_t& _r7h)
-{
-    float32x4x2_t _r01lz = vzipq_f32(_r0l, _r1l);
-    float32x4x2_t _r23lz = vzipq_f32(_r2l, _r3l);
-    float32x4x2_t _r01hz = vzipq_f32(_r0h, _r1h);
-    float32x4x2_t _r23hz = vzipq_f32(_r2h, _r3h);
-    float32x4x2_t _r45lz = vzipq_f32(_r4l, _r5l);
-    float32x4x2_t _r67lz = vzipq_f32(_r6l, _r7l);
-    float32x4x2_t _r45hz = vzipq_f32(_r4h, _r5h);
-    float32x4x2_t _r67hz = vzipq_f32(_r6h, _r7h);
-    _r0l = vcombine_f32(vget_low_f32(_r01lz.val[0]), vget_low_f32(_r23lz.val[0]));
-    _r0h = vcombine_f32(vget_low_f32(_r45lz.val[0]), vget_low_f32(_r67lz.val[0]));
-    _r1l = vcombine_f32(vget_high_f32(_r01lz.val[0]), vget_high_f32(_r23lz.val[0]));
-    _r1h = vcombine_f32(vget_high_f32(_r45lz.val[0]), vget_high_f32(_r67lz.val[0]));
-    _r2l = vcombine_f32(vget_low_f32(_r01lz.val[1]), vget_low_f32(_r23lz.val[1]));
-    _r2h = vcombine_f32(vget_low_f32(_r45lz.val[1]), vget_low_f32(_r67lz.val[1]));
-    _r3l = vcombine_f32(vget_high_f32(_r01lz.val[1]), vget_high_f32(_r23lz.val[1]));
-    _r3h = vcombine_f32(vget_high_f32(_r45lz.val[1]), vget_high_f32(_r67lz.val[1]));
-    _r4l = vcombine_f32(vget_low_f32(_r01hz.val[0]), vget_low_f32(_r23hz.val[0]));
-    _r4h = vcombine_f32(vget_low_f32(_r45hz.val[0]), vget_low_f32(_r67hz.val[0]));
-    _r5l = vcombine_f32(vget_high_f32(_r01hz.val[0]), vget_high_f32(_r23hz.val[0]));
-    _r5h = vcombine_f32(vget_high_f32(_r45hz.val[0]), vget_high_f32(_r67hz.val[0]));
-    _r6l = vcombine_f32(vget_low_f32(_r01hz.val[1]), vget_low_f32(_r23hz.val[1]));
-    _r6h = vcombine_f32(vget_low_f32(_r45hz.val[1]), vget_low_f32(_r67hz.val[1]));
-    _r7l = vcombine_f32(vget_high_f32(_r01hz.val[1]), vget_high_f32(_r23hz.val[1]));
-    _r7h = vcombine_f32(vget_high_f32(_r45hz.val[1]), vget_high_f32(_r67hz.val[1]));
-}
-
-static inline void transpose8x12_ps(float32x4_t& _r0l, float32x4_t& _r0h,
-                                    float32x4_t& _r1l, float32x4_t& _r1h,
-                                    float32x4_t& _r2l, float32x4_t& _r2h,
-                                    float32x4_t& _r3l, float32x4_t& _r3h,
-                                    float32x4_t& _r4l, float32x4_t& _r4h,
-                                    float32x4_t& _r5l, float32x4_t& _r5h,
-                                    float32x4_t& _r6l, float32x4_t& _r6h,
-                                    float32x4_t& _r7l, float32x4_t& _r7h,
-                                    float32x4_t& _r8l, float32x4_t& _r8h,
-                                    float32x4_t& _r9l, float32x4_t& _r9h,
-                                    float32x4_t& _ral, float32x4_t& _rah,
-                                    float32x4_t& _rbl, float32x4_t& _rbh)
-{
-    float32x4x2_t _r01lz = vzipq_f32(_r0l, _r1l);
-    float32x4x2_t _r23lz = vzipq_f32(_r2l, _r3l);
-    float32x4x2_t _r01hz = vzipq_f32(_r0h, _r1h);
-    float32x4x2_t _r23hz = vzipq_f32(_r2h, _r3h);
-    float32x4x2_t _r45lz = vzipq_f32(_r4l, _r5l);
-    float32x4x2_t _r67lz = vzipq_f32(_r6l, _r7l);
-    float32x4x2_t _r45hz = vzipq_f32(_r4h, _r5h);
-    float32x4x2_t _r67hz = vzipq_f32(_r6h, _r7h);
-    float32x4x2_t _r89lz = vzipq_f32(_r8l, _r9l);
-    float32x4x2_t _rablz = vzipq_f32(_ral, _rbl);
-    float32x4x2_t _r89hz = vzipq_f32(_r8h, _r9h);
-    float32x4x2_t _rabhz = vzipq_f32(_rah, _rbh);
-    _r0l = vcombine_f32(vget_low_f32(_r01lz.val[0]), vget_low_f32(_r23lz.val[0]));
-    _r0h = vcombine_f32(vget_low_f32(_r45lz.val[0]), vget_low_f32(_r67lz.val[0]));
-    _r1l = vcombine_f32(vget_low_f32(_r89lz.val[0]), vget_low_f32(_rablz.val[0]));
-    _r1h = vcombine_f32(vget_high_f32(_r01lz.val[0]), vget_high_f32(_r23lz.val[0]));
-    _r2l = vcombine_f32(vget_high_f32(_r45lz.val[0]), vget_high_f32(_r67lz.val[0]));
-    _r2h = vcombine_f32(vget_high_f32(_r89lz.val[0]), vget_high_f32(_rablz.val[0]));
-    _r3l = vcombine_f32(vget_low_f32(_r01lz.val[1]), vget_low_f32(_r23lz.val[1]));
-    _r3h = vcombine_f32(vget_low_f32(_r45lz.val[1]), vget_low_f32(_r67lz.val[1]));
-    _r4l = vcombine_f32(vget_low_f32(_r89lz.val[1]), vget_low_f32(_rablz.val[1]));
-    _r4h = vcombine_f32(vget_high_f32(_r01lz.val[1]), vget_high_f32(_r23lz.val[1]));
-    _r5l = vcombine_f32(vget_high_f32(_r45lz.val[1]), vget_high_f32(_r67lz.val[1]));
-    _r5h = vcombine_f32(vget_high_f32(_r89lz.val[1]), vget_high_f32(_rablz.val[1]));
-    _r6l = vcombine_f32(vget_low_f32(_r01hz.val[0]), vget_low_f32(_r23hz.val[0]));
-    _r6h = vcombine_f32(vget_low_f32(_r45hz.val[0]), vget_low_f32(_r67hz.val[0]));
-    _r7l = vcombine_f32(vget_low_f32(_r89hz.val[0]), vget_low_f32(_rabhz.val[0]));
-    _r7h = vcombine_f32(vget_high_f32(_r01hz.val[0]), vget_high_f32(_r23hz.val[0]));
-    _r8l = vcombine_f32(vget_high_f32(_r45hz.val[0]), vget_high_f32(_r67hz.val[0]));
-    _r8h = vcombine_f32(vget_high_f32(_r89hz.val[0]), vget_high_f32(_rabhz.val[0]));
-    _r9l = vcombine_f32(vget_low_f32(_r01hz.val[1]), vget_low_f32(_r23hz.val[1]));
-    _r9h = vcombine_f32(vget_low_f32(_r45hz.val[1]), vget_low_f32(_r67hz.val[1]));
-    _ral = vcombine_f32(vget_low_f32(_r89hz.val[1]), vget_low_f32(_rabhz.val[1]));
-    _rah = vcombine_f32(vget_high_f32(_r01hz.val[1]), vget_high_f32(_r23hz.val[1]));
-    _rbl = vcombine_f32(vget_high_f32(_r45hz.val[1]), vget_high_f32(_r67hz.val[1]));
-    _rbh = vcombine_f32(vget_high_f32(_r89hz.val[1]), vget_high_f32(_rabhz.val[1]));
-}
-
-static inline void transpose12x8_ps(float32x4_t& _r0l, float32x4_t& _r0m, float32x4_t& _r0h,
-                                    float32x4_t& _r1l, float32x4_t& _r1m, float32x4_t& _r1h,
-                                    float32x4_t& _r2l, float32x4_t& _r2m, float32x4_t& _r2h,
-                                    float32x4_t& _r3l, float32x4_t& _r3m, float32x4_t& _r3h,
-                                    float32x4_t& _r4l, float32x4_t& _r4m, float32x4_t& _r4h,
-                                    float32x4_t& _r5l, float32x4_t& _r5m, float32x4_t& _r5h,
-                                    float32x4_t& _r6l, float32x4_t& _r6m, float32x4_t& _r6h,
-                                    float32x4_t& _r7l, float32x4_t& _r7m, float32x4_t& _r7h)
-{
-    float32x4x2_t _r01lz = vzipq_f32(_r0l, _r1l);
-    float32x4x2_t _r23lz = vzipq_f32(_r2l, _r3l);
-    float32x4x2_t _r01mz = vzipq_f32(_r0m, _r1m);
-    float32x4x2_t _r23mz = vzipq_f32(_r2m, _r3m);
-    float32x4x2_t _r01hz = vzipq_f32(_r0h, _r1h);
-    float32x4x2_t _r23hz = vzipq_f32(_r2h, _r3h);
-    float32x4x2_t _r45lz = vzipq_f32(_r4l, _r5l);
-    float32x4x2_t _r67lz = vzipq_f32(_r6l, _r7l);
-    float32x4x2_t _r45mz = vzipq_f32(_r4m, _r5m);
-    float32x4x2_t _r67mz = vzipq_f32(_r6m, _r7m);
-    float32x4x2_t _r45hz = vzipq_f32(_r4h, _r5h);
-    float32x4x2_t _r67hz = vzipq_f32(_r6h, _r7h);
-    _r0l = vcombine_f32(vget_low_f32(_r01lz.val[0]), vget_low_f32(_r23lz.val[0]));
-    _r0m = vcombine_f32(vget_low_f32(_r45lz.val[0]), vget_low_f32(_r67lz.val[0]));
-    _r0h = vcombine_f32(vget_high_f32(_r01lz.val[0]), vget_high_f32(_r23lz.val[0]));
-    _r1l = vcombine_f32(vget_high_f32(_r45lz.val[0]), vget_high_f32(_r67lz.val[0]));
-    _r1m = vcombine_f32(vget_low_f32(_r01lz.val[1]), vget_low_f32(_r23lz.val[1]));
-    _r1h = vcombine_f32(vget_low_f32(_r45lz.val[1]), vget_low_f32(_r67lz.val[1]));
-    _r2l = vcombine_f32(vget_high_f32(_r01lz.val[1]), vget_high_f32(_r23lz.val[1]));
-    _r2m = vcombine_f32(vget_high_f32(_r45lz.val[1]), vget_high_f32(_r67lz.val[1]));
-    _r2h = vcombine_f32(vget_low_f32(_r01mz.val[0]), vget_low_f32(_r23mz.val[0]));
-    _r3l = vcombine_f32(vget_low_f32(_r45mz.val[0]), vget_low_f32(_r67mz.val[0]));
-    _r3m = vcombine_f32(vget_high_f32(_r01mz.val[0]), vget_high_f32(_r23mz.val[0]));
-    _r3h = vcombine_f32(vget_high_f32(_r45mz.val[0]), vget_high_f32(_r67mz.val[0]));
-    _r4l = vcombine_f32(vget_low_f32(_r01mz.val[1]), vget_low_f32(_r23mz.val[1]));
-    _r4m = vcombine_f32(vget_low_f32(_r45mz.val[1]), vget_low_f32(_r67mz.val[1]));
-    _r4h = vcombine_f32(vget_high_f32(_r01mz.val[1]), vget_high_f32(_r23mz.val[1]));
-    _r5l = vcombine_f32(vget_high_f32(_r45mz.val[1]), vget_high_f32(_r67mz.val[1]));
-    _r5m = vcombine_f32(vget_low_f32(_r01hz.val[0]), vget_low_f32(_r23hz.val[0]));
-    _r5h = vcombine_f32(vget_low_f32(_r45hz.val[0]), vget_low_f32(_r67hz.val[0]));
-    _r6l = vcombine_f32(vget_high_f32(_r01hz.val[0]), vget_high_f32(_r23hz.val[0]));
-    _r6m = vcombine_f32(vget_high_f32(_r45hz.val[0]), vget_high_f32(_r67hz.val[0]));
-    _r6h = vcombine_f32(vget_low_f32(_r01hz.val[1]), vget_low_f32(_r23hz.val[1]));
-    _r7l = vcombine_f32(vget_low_f32(_r45hz.val[1]), vget_low_f32(_r67hz.val[1]));
-    _r7m = vcombine_f32(vget_high_f32(_r01hz.val[1]), vget_high_f32(_r23hz.val[1]));
-    _r7h = vcombine_f32(vget_high_f32(_r45hz.val[1]), vget_high_f32(_r67hz.val[1]));
-}
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-static inline void transpose4x4_ph(float16x4_t& _r0, float16x4_t& _r1, float16x4_t& _r2, float16x4_t& _r3)
-{
-    uint16x4x2_t _r01z = vzip_u16(vreinterpret_u16_f16(_r0), vreinterpret_u16_f16(_r1));
-    uint16x4x2_t _r23z = vzip_u16(vreinterpret_u16_f16(_r2), vreinterpret_u16_f16(_r3));
-    uint32x2x2_t _r01 = vzip_u32(vreinterpret_u32_u16(_r01z.val[0]), vreinterpret_u32_u16(_r23z.val[0]));
-    uint32x2x2_t _r23 = vzip_u32(vreinterpret_u32_u16(_r01z.val[1]), vreinterpret_u32_u16(_r23z.val[1]));
-    _r0 = vreinterpret_f16_u32(_r01.val[0]);
-    _r1 = vreinterpret_f16_u32(_r01.val[1]);
-    _r2 = vreinterpret_f16_u32(_r23.val[0]);
-    _r3 = vreinterpret_f16_u32(_r23.val[1]);
-}
-
-static inline void transpose4x8_ph(float16x4_t& _r0, float16x4_t& _r1, float16x4_t& _r2, float16x4_t& _r3, float16x4_t& _r4, float16x4_t& _r5, float16x4_t& _r6, float16x4_t& _r7)
-{
-    uint16x4x2_t _r01z = vzip_u16(vreinterpret_u16_f16(_r0), vreinterpret_u16_f16(_r1));
-    uint16x4x2_t _r23z = vzip_u16(vreinterpret_u16_f16(_r2), vreinterpret_u16_f16(_r3));
-    uint16x4x2_t _r45z = vzip_u16(vreinterpret_u16_f16(_r4), vreinterpret_u16_f16(_r5));
-    uint16x4x2_t _r67z = vzip_u16(vreinterpret_u16_f16(_r6), vreinterpret_u16_f16(_r7));
-    uint32x2x2_t _r01_0 = vzip_u32(vreinterpret_u32_u16(_r01z.val[0]), vreinterpret_u32_u16(_r23z.val[0]));
-    uint32x2x2_t _r23_0 = vzip_u32(vreinterpret_u32_u16(_r01z.val[1]), vreinterpret_u32_u16(_r23z.val[1]));
-    uint32x2x2_t _r01_1 = vzip_u32(vreinterpret_u32_u16(_r45z.val[0]), vreinterpret_u32_u16(_r67z.val[0]));
-    uint32x2x2_t _r23_1 = vzip_u32(vreinterpret_u32_u16(_r45z.val[1]), vreinterpret_u32_u16(_r67z.val[1]));
-    _r0 = vreinterpret_f16_u32(_r01_0.val[0]);
-    _r1 = vreinterpret_f16_u32(_r01_1.val[0]);
-    _r2 = vreinterpret_f16_u32(_r01_0.val[1]);
-    _r3 = vreinterpret_f16_u32(_r01_1.val[1]);
-    _r4 = vreinterpret_f16_u32(_r23_0.val[0]);
-    _r5 = vreinterpret_f16_u32(_r23_1.val[0]);
-    _r6 = vreinterpret_f16_u32(_r23_0.val[1]);
-    _r7 = vreinterpret_f16_u32(_r23_1.val[1]);
-}
-
-static inline void transpose4x12_ph(float16x4_t& _r0, float16x4_t& _r1, float16x4_t& _r2, float16x4_t& _r3, float16x4_t& _r4, float16x4_t& _r5, float16x4_t& _r6, float16x4_t& _r7, float16x4_t& _r8, float16x4_t& _r9, float16x4_t& _ra, float16x4_t& _rb)
-{
-    uint16x4x2_t _r01z = vzip_u16(vreinterpret_u16_f16(_r0), vreinterpret_u16_f16(_r1));
-    uint16x4x2_t _r23z = vzip_u16(vreinterpret_u16_f16(_r2), vreinterpret_u16_f16(_r3));
-    uint16x4x2_t _r45z = vzip_u16(vreinterpret_u16_f16(_r4), vreinterpret_u16_f16(_r5));
-    uint16x4x2_t _r67z = vzip_u16(vreinterpret_u16_f16(_r6), vreinterpret_u16_f16(_r7));
-    uint16x4x2_t _r89z = vzip_u16(vreinterpret_u16_f16(_r8), vreinterpret_u16_f16(_r9));
-    uint16x4x2_t _rabz = vzip_u16(vreinterpret_u16_f16(_ra), vreinterpret_u16_f16(_rb));
-    uint32x2x2_t _r01_0 = vzip_u32(vreinterpret_u32_u16(_r01z.val[0]), vreinterpret_u32_u16(_r23z.val[0]));
-    uint32x2x2_t _r23_0 = vzip_u32(vreinterpret_u32_u16(_r01z.val[1]), vreinterpret_u32_u16(_r23z.val[1]));
-    uint32x2x2_t _r01_1 = vzip_u32(vreinterpret_u32_u16(_r45z.val[0]), vreinterpret_u32_u16(_r67z.val[0]));
-    uint32x2x2_t _r23_1 = vzip_u32(vreinterpret_u32_u16(_r45z.val[1]), vreinterpret_u32_u16(_r67z.val[1]));
-    uint32x2x2_t _r01_2 = vzip_u32(vreinterpret_u32_u16(_r89z.val[0]), vreinterpret_u32_u16(_rabz.val[0]));
-    uint32x2x2_t _r23_2 = vzip_u32(vreinterpret_u32_u16(_r89z.val[1]), vreinterpret_u32_u16(_rabz.val[1]));
-    _r0 = vreinterpret_f16_u32(_r01_0.val[0]);
-    _r1 = vreinterpret_f16_u32(_r01_1.val[0]);
-    _r2 = vreinterpret_f16_u32(_r01_2.val[0]);
-    _r3 = vreinterpret_f16_u32(_r01_0.val[1]);
-    _r4 = vreinterpret_f16_u32(_r01_1.val[1]);
-    _r5 = vreinterpret_f16_u32(_r01_2.val[1]);
-    _r6 = vreinterpret_f16_u32(_r23_0.val[0]);
-    _r7 = vreinterpret_f16_u32(_r23_1.val[0]);
-    _r8 = vreinterpret_f16_u32(_r23_2.val[0]);
-    _r9 = vreinterpret_f16_u32(_r23_0.val[1]);
-    _ra = vreinterpret_f16_u32(_r23_1.val[1]);
-    _rb = vreinterpret_f16_u32(_r23_2.val[1]);
-}
-
-static inline void transpose8x4_ph(float16x8_t& _r0, float16x8_t& _r1, float16x8_t& _r2, float16x8_t& _r3)
-{
-    uint16x8x2_t _r01t = vzipq_u16(vreinterpretq_u16_f16(_r0), vreinterpretq_u16_f16(_r1));
-    uint16x8x2_t _r23t = vzipq_u16(vreinterpretq_u16_f16(_r2), vreinterpretq_u16_f16(_r3));
-    uint32x4x2_t _r01 = vzipq_u32(vreinterpretq_u32_u16(_r01t.val[0]), vreinterpretq_u32_u16(_r23t.val[0]));
-    uint32x4x2_t _r23 = vzipq_u32(vreinterpretq_u32_u16(_r01t.val[1]), vreinterpretq_u32_u16(_r23t.val[1]));
-    _r0 = vreinterpretq_f16_u32(_r01.val[0]);
-    _r1 = vreinterpretq_f16_u32(_r01.val[1]);
-    _r2 = vreinterpretq_f16_u32(_r23.val[0]);
-    _r3 = vreinterpretq_f16_u32(_r23.val[1]);
-}
-
-static inline void transpose8x8_ph(float16x8_t& _r0, float16x8_t& _r1, float16x8_t& _r2, float16x8_t& _r3, float16x8_t& _r4, float16x8_t& _r5, float16x8_t& _r6, float16x8_t& _r7)
-{
-    uint16x8x2_t _r01t = vzipq_u16(vreinterpretq_u16_f16(_r0), vreinterpretq_u16_f16(_r1));
-    uint16x8x2_t _r23t = vzipq_u16(vreinterpretq_u16_f16(_r2), vreinterpretq_u16_f16(_r3));
-    uint16x8x2_t _r45t = vzipq_u16(vreinterpretq_u16_f16(_r4), vreinterpretq_u16_f16(_r5));
-    uint16x8x2_t _r67t = vzipq_u16(vreinterpretq_u16_f16(_r6), vreinterpretq_u16_f16(_r7));
-    uint32x4x2_t _r01_0 = vzipq_u32(vreinterpretq_u32_u16(_r01t.val[0]), vreinterpretq_u32_u16(_r23t.val[0]));
-    uint32x4x2_t _r23_0 = vzipq_u32(vreinterpretq_u32_u16(_r01t.val[1]), vreinterpretq_u32_u16(_r23t.val[1]));
-    uint32x4x2_t _r01_1 = vzipq_u32(vreinterpretq_u32_u16(_r45t.val[0]), vreinterpretq_u32_u16(_r67t.val[0]));
-    uint32x4x2_t _r23_1 = vzipq_u32(vreinterpretq_u32_u16(_r45t.val[1]), vreinterpretq_u32_u16(_r67t.val[1]));
-    _r0 = vreinterpretq_f16_u32(vcombine_u32(vget_low_u32(_r01_0.val[0]), vget_low_u32(_r01_1.val[0])));
-    _r1 = vreinterpretq_f16_u32(vcombine_u32(vget_high_u32(_r01_0.val[0]), vget_high_u32(_r01_1.val[0])));
-    _r2 = vreinterpretq_f16_u32(vcombine_u32(vget_low_u32(_r01_0.val[1]), vget_low_u32(_r01_1.val[1])));
-    _r3 = vreinterpretq_f16_u32(vcombine_u32(vget_high_u32(_r01_0.val[1]), vget_high_u32(_r01_1.val[1])));
-    _r4 = vreinterpretq_f16_u32(vcombine_u32(vget_low_u32(_r23_0.val[0]), vget_low_u32(_r23_1.val[0])));
-    _r5 = vreinterpretq_f16_u32(vcombine_u32(vget_high_u32(_r23_0.val[0]), vget_high_u32(_r23_1.val[0])));
-    _r6 = vreinterpretq_f16_u32(vcombine_u32(vget_low_u32(_r23_0.val[1]), vget_low_u32(_r23_1.val[1])));
-    _r7 = vreinterpretq_f16_u32(vcombine_u32(vget_high_u32(_r23_0.val[1]), vget_high_u32(_r23_1.val[1])));
-}
-
-static inline void transpose8x12_ph(float16x8_t& _r0, float16x8_t& _r1, float16x8_t& _r2, float16x8_t& _r3, float16x8_t& _r4, float16x8_t& _r5, float16x8_t& _r6, float16x8_t& _r7, float16x8_t& _r8, float16x8_t& _r9, float16x8_t& _ra, float16x8_t& _rb)
-{
-    uint16x8x2_t _r01t = vzipq_u16(vreinterpretq_u16_f16(_r0), vreinterpretq_u16_f16(_r1));
-    uint16x8x2_t _r23t = vzipq_u16(vreinterpretq_u16_f16(_r2), vreinterpretq_u16_f16(_r3));
-    uint16x8x2_t _r45t = vzipq_u16(vreinterpretq_u16_f16(_r4), vreinterpretq_u16_f16(_r5));
-    uint16x8x2_t _r67t = vzipq_u16(vreinterpretq_u16_f16(_r6), vreinterpretq_u16_f16(_r7));
-    uint16x8x2_t _r89t = vzipq_u16(vreinterpretq_u16_f16(_r8), vreinterpretq_u16_f16(_r9));
-    uint16x8x2_t _rabt = vzipq_u16(vreinterpretq_u16_f16(_ra), vreinterpretq_u16_f16(_rb));
-    uint32x4x2_t _r01_0 = vzipq_u32(vreinterpretq_u32_u16(_r01t.val[0]), vreinterpretq_u32_u16(_r23t.val[0]));
-    uint32x4x2_t _r23_0 = vzipq_u32(vreinterpretq_u32_u16(_r01t.val[1]), vreinterpretq_u32_u16(_r23t.val[1]));
-    uint32x4x2_t _r01_1 = vzipq_u32(vreinterpretq_u32_u16(_r45t.val[0]), vreinterpretq_u32_u16(_r67t.val[0]));
-    uint32x4x2_t _r23_1 = vzipq_u32(vreinterpretq_u32_u16(_r45t.val[1]), vreinterpretq_u32_u16(_r67t.val[1]));
-    uint32x4x2_t _r01_2 = vzipq_u32(vreinterpretq_u32_u16(_r89t.val[0]), vreinterpretq_u32_u16(_rabt.val[0]));
-    uint32x4x2_t _r23_2 = vzipq_u32(vreinterpretq_u32_u16(_r89t.val[1]), vreinterpretq_u32_u16(_rabt.val[1]));
-    _r0 = vreinterpretq_f16_u32(vcombine_u32(vget_low_u32(_r01_0.val[0]), vget_low_u32(_r01_1.val[0])));
-    _r1 = vreinterpretq_f16_u32(vcombine_u32(vget_low_u32(_r01_2.val[0]), vget_high_u32(_r01_0.val[0])));
-    _r2 = vreinterpretq_f16_u32(vcombine_u32(vget_high_u32(_r01_1.val[0]), vget_high_u32(_r01_2.val[0])));
-    _r3 = vreinterpretq_f16_u32(vcombine_u32(vget_low_u32(_r01_0.val[1]), vget_low_u32(_r01_1.val[1])));
-    _r4 = vreinterpretq_f16_u32(vcombine_u32(vget_low_u32(_r01_2.val[1]), vget_high_u32(_r01_0.val[1])));
-    _r5 = vreinterpretq_f16_u32(vcombine_u32(vget_high_u32(_r01_1.val[1]), vget_high_u32(_r01_2.val[1])));
-    _r6 = vreinterpretq_f16_u32(vcombine_u32(vget_low_u32(_r23_0.val[0]), vget_low_u32(_r23_1.val[0])));
-    _r7 = vreinterpretq_f16_u32(vcombine_u32(vget_low_u32(_r23_2.val[0]), vget_high_u32(_r23_0.val[0])));
-    _r8 = vreinterpretq_f16_u32(vcombine_u32(vget_high_u32(_r23_1.val[0]), vget_high_u32(_r23_2.val[0])));
-    _r9 = vreinterpretq_f16_u32(vcombine_u32(vget_low_u32(_r23_0.val[1]), vget_low_u32(_r23_1.val[1])));
-    _ra = vreinterpretq_f16_u32(vcombine_u32(vget_low_u32(_r23_2.val[1]), vget_high_u32(_r23_0.val[1])));
-    _rb = vreinterpretq_f16_u32(vcombine_u32(vget_high_u32(_r23_1.val[1]), vget_high_u32(_r23_2.val[1])));
-}
-
-static inline void transpose12x4_ph(float16x4_t& _r0l, float16x4_t& _r0m, float16x4_t& _r0h,
-                                    float16x4_t& _r1l, float16x4_t& _r1m, float16x4_t& _r1h,
-                                    float16x4_t& _r2l, float16x4_t& _r2m, float16x4_t& _r2h,
-                                    float16x4_t& _r3l, float16x4_t& _r3m, float16x4_t& _r3h)
-{
-    uint16x4x2_t _r01lz = vzip_u16(vreinterpret_u16_f16(_r0l), vreinterpret_u16_f16(_r1l));
-    uint16x4x2_t _r23lz = vzip_u16(vreinterpret_u16_f16(_r2l), vreinterpret_u16_f16(_r3l));
-    uint16x4x2_t _r01mz = vzip_u16(vreinterpret_u16_f16(_r0m), vreinterpret_u16_f16(_r1m));
-    uint16x4x2_t _r23mz = vzip_u16(vreinterpret_u16_f16(_r2m), vreinterpret_u16_f16(_r3m));
-    uint16x4x2_t _r01hz = vzip_u16(vreinterpret_u16_f16(_r0h), vreinterpret_u16_f16(_r1h));
-    uint16x4x2_t _r23hz = vzip_u16(vreinterpret_u16_f16(_r2h), vreinterpret_u16_f16(_r3h));
-    uint32x2x2_t _r01 = vzip_u32(vreinterpret_u32_u16(_r01lz.val[0]), vreinterpret_u32_u16(_r23lz.val[0]));
-    uint32x2x2_t _r23 = vzip_u32(vreinterpret_u32_u16(_r01lz.val[1]), vreinterpret_u32_u16(_r23lz.val[1]));
-    uint32x2x2_t _r45 = vzip_u32(vreinterpret_u32_u16(_r01mz.val[0]), vreinterpret_u32_u16(_r23mz.val[0]));
-    uint32x2x2_t _r67 = vzip_u32(vreinterpret_u32_u16(_r01mz.val[1]), vreinterpret_u32_u16(_r23mz.val[1]));
-    uint32x2x2_t _r89 = vzip_u32(vreinterpret_u32_u16(_r01hz.val[0]), vreinterpret_u32_u16(_r23hz.val[0]));
-    uint32x2x2_t _rab = vzip_u32(vreinterpret_u32_u16(_r01hz.val[1]), vreinterpret_u32_u16(_r23hz.val[1]));
-    _r0l = vreinterpret_f16_u32(_r01.val[0]);
-    _r0m = vreinterpret_f16_u32(_r01.val[1]);
-    _r0h = vreinterpret_f16_u32(_r23.val[0]);
-    _r1l = vreinterpret_f16_u32(_r23.val[1]);
-    _r1m = vreinterpret_f16_u32(_r45.val[0]);
-    _r1h = vreinterpret_f16_u32(_r45.val[1]);
-    _r2l = vreinterpret_f16_u32(_r67.val[0]);
-    _r2m = vreinterpret_f16_u32(_r67.val[1]);
-    _r2h = vreinterpret_f16_u32(_r89.val[0]);
-    _r3l = vreinterpret_f16_u32(_r89.val[1]);
-    _r3m = vreinterpret_f16_u32(_rab.val[0]);
-    _r3h = vreinterpret_f16_u32(_rab.val[1]);
-}
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-
-#endif // __aarch64__
-#endif // __ARM_NEON
-
-#endif // ARM_USABILITY_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/batchnorm_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/batchnorm_arm.h
deleted file mode 100644
index 1393bb30..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/batchnorm_arm.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BATCHNORM_ARM_H
-#define LAYER_BATCHNORM_ARM_H
-
-#include "batchnorm.h"
-
-namespace ncnn {
-
-class BatchNorm_arm : public BatchNorm
-{
-public:
-    BatchNorm_arm();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_inplace_bf16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BATCHNORM_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/bias_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/bias_arm.h
deleted file mode 100644
index 5f08facf..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/bias_arm.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BIAS_ARM_H
-#define LAYER_BIAS_ARM_H
-
-#include "bias.h"
-
-namespace ncnn {
-
-class Bias_arm : public Bias
-{
-public:
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BIAS_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/binaryop_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/binaryop_arm.h
deleted file mode 100644
index 1337065e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/binaryop_arm.h
+++ /dev/null
@@ -1,44 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BINARYOP_ARM_H
-#define LAYER_BINARYOP_ARM_H
-
-#include "binaryop.h"
-
-namespace ncnn {
-
-class BinaryOp_arm : public BinaryOp
-{
-public:
-    BinaryOp_arm();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_fp16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_bf16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-    int forward_inplace_bf16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BINARYOP_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/cast_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/cast_arm.h
deleted file mode 100644
index fc32c70d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/cast_arm.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CAST_ARM_H
-#define LAYER_CAST_ARM_H
-
-#include "cast.h"
-
-namespace ncnn {
-
-class Cast_arm : public Cast
-{
-public:
-    Cast_arm();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CAST_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/cast_bf16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/cast_bf16.h
deleted file mode 100644
index ab196687..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/cast_bf16.h
+++ /dev/null
@@ -1,309 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if NCNN_RUNTIME_CPU && NCNN_ARM84BF16 && __aarch64__ && !__ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-void cast_fp32_to_bf16_neon_bf16(const Mat& bottom_blob, Mat& top_blob, const Option& opt);
-void cast_bf16_to_fp32_neon_bf16(const Mat& bottom_blob, Mat& top_blob, const Option& opt);
-#endif
-
-static void cast_fp32_to_bf16_neon(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-#if NCNN_RUNTIME_CPU && NCNN_ARM84BF16 && __aarch64__ && !__ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-    if (ncnn::cpu_support_arm_bf16())
-    {
-        cast_fp32_to_bf16_neon_bf16(bottom_blob, top_blob, opt);
-        return;
-    }
-#endif
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int d = bottom_blob.d;
-    const int channels = bottom_blob.c;
-    const int elempack = bottom_blob.elempack;
-
-    const int size = w * h * d * elempack;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* ptr = bottom_blob.channel(q);
-#if __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-        __bf16* outptr = top_blob.channel(q);
-#else
-        unsigned short* outptr = top_blob.channel(q);
-#endif
-
-        int i = 0;
-#if __ARM_NEON
-        for (; i + 15 < size; i += 16)
-        {
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-#if __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-            asm volatile(
-                "prfm   pldl1keep, [%0, #512]   \n"
-                "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-                "bfcvtn v0.4h, v0.4s            \n"
-                "bfcvtn2 v0.8h, v1.4s           \n"
-                "bfcvtn v1.4h, v2.4s            \n"
-                "bfcvtn2 v1.8h, v3.4s           \n"
-                "st1    {v0.8h, v1.8h}, [%1], #32 \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "v0", "v1");
-#else  // __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-            asm volatile(
-                "prfm   pldl1keep, [%0, #512]   \n"
-                "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-                "shrn   v0.4h, v0.4s, #16       \n"
-                "shrn   v1.4h, v1.4s, #16       \n"
-                "shrn   v2.4h, v2.4s, #16       \n"
-                "shrn   v3.4h, v3.4s, #16       \n"
-                "st1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%1], #32 \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "v0", "v1", "v2", "v3");
-#endif // __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-#else  // __aarch64__
-            asm volatile(
-                "pld        [%0, #512]      \n"
-                "vldm       %0!, {d0-d7}    \n"
-                "vshrn.u32  d0, q0, #16     \n"
-                "vshrn.u32  d1, q1, #16     \n"
-                "vshrn.u32  d2, q2, #16     \n"
-                "vshrn.u32  d3, q3, #16     \n"
-                "vst1.u16   {d0-d3}, [%1]!  \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "q0", "q1", "q2", "q3");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _p0_fp32 = vld1q_f32(ptr);
-            float32x4_t _p1_fp32 = vld1q_f32(ptr + 4);
-            float32x4_t _p2_fp32 = vld1q_f32(ptr + 8);
-            float32x4_t _p3_fp32 = vld1q_f32(ptr + 12);
-#if __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-            bfloat16x4_t _p0_bf16 = vcvt_bf16_f32(_p0_fp32);
-            bfloat16x4_t _p1_bf16 = vcvt_bf16_f32(_p1_fp32);
-            bfloat16x4_t _p2_bf16 = vcvt_bf16_f32(_p2_fp32);
-            bfloat16x4_t _p3_bf16 = vcvt_bf16_f32(_p3_fp32);
-            bfloat16x8_t _p_bf16 = vcombine_bf16(_p0_bf16, _p1_bf16);
-            bfloat16x8_t _q_bf16 = vcombine_bf16(_p2_bf16, _p3_bf16);
-            vst1q_bf16(outptr, _p_bf16);
-            vst1q_bf16(outptr + 8, _q_bf16);
-#else
-            uint16x4_t _p0_bf16 = float2bfloat(_p0_fp32);
-            uint16x4_t _p1_bf16 = float2bfloat(_p1_fp32);
-            uint16x4_t _p2_bf16 = float2bfloat(_p2_fp32);
-            uint16x4_t _p3_bf16 = float2bfloat(_p3_fp32);
-            uint16x8_t _p_bf16 = vcombine_u16(_p0_bf16, _p1_bf16);
-            uint16x8_t _q_bf16 = vcombine_u16(_p2_bf16, _p3_bf16);
-            vst1q_u16(outptr, _p_bf16);
-            vst1q_u16(outptr + 8, _q_bf16);
-#endif
-            ptr += 16;
-            outptr += 16;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; i + 7 < size; i += 8)
-        {
-            float32x4_t _p0_fp32 = vld1q_f32(ptr);
-            float32x4_t _p1_fp32 = vld1q_f32(ptr + 4);
-#if __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-            bfloat16x4_t _p0_bf16 = vcvt_bf16_f32(_p0_fp32);
-            bfloat16x4_t _p1_bf16 = vcvt_bf16_f32(_p1_fp32);
-            bfloat16x8_t _p_bf16 = vcombine_bf16(_p0_bf16, _p1_bf16);
-            vst1q_bf16(outptr, _p_bf16);
-#else
-            uint16x4_t _p0_bf16 = float2bfloat(_p0_fp32);
-            uint16x4_t _p1_bf16 = float2bfloat(_p1_fp32);
-            uint16x8_t _p_bf16 = vcombine_u16(_p0_bf16, _p1_bf16);
-            vst1q_u16(outptr, _p_bf16);
-#endif
-            ptr += 8;
-            outptr += 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-            float32x4_t _p_fp32 = vld1q_f32(ptr);
-#if __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-            bfloat16x4_t _p_bf16 = vcvt_bf16_f32(_p_fp32);
-            vst1_bf16(outptr, _p_bf16);
-#else
-            uint16x4_t _p_bf16 = float2bfloat(_p_fp32);
-            vst1_u16(outptr, _p_bf16);
-#endif
-            ptr += 4;
-            outptr += 4;
-        }
-#endif
-        for (; i < size; i++)
-        {
-#if NCNN_GNU_INLINE_ASM && __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-            asm volatile(
-                "ldr    s0, [%0], #4    \n"
-                "bfcvt  h0, s0          \n"
-                "str    h0, [%1], #2    \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "s0");
-            // because intrinsic cause ndk clang crash
-            // *outptr++ = vcvth_bf16_f32(*ptr++);
-#else
-#if __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-            *(unsigned short*)outptr = float32_to_bfloat16(*ptr++);
-            outptr++;
-#else
-            *outptr++ = float32_to_bfloat16(*ptr++);
-#endif
-#endif
-        }
-    }
-}
-
-static void cast_bf16_to_fp32_neon(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-#if NCNN_RUNTIME_CPU && NCNN_ARM84BF16 && __aarch64__ && !__ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-    if (ncnn::cpu_support_arm_bf16())
-    {
-        cast_bf16_to_fp32_neon_bf16(bottom_blob, top_blob, opt);
-        return;
-    }
-#endif
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int d = bottom_blob.d;
-    const int channels = bottom_blob.c;
-    const int elempack = bottom_blob.elempack;
-
-    const int size = w * h * d * elempack;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-#if __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-        const __bf16* ptr = bottom_blob.channel(q);
-#else
-        const unsigned short* ptr = bottom_blob.channel(q);
-#endif
-        float* outptr = top_blob.channel(q);
-
-        int i = 0;
-#if __ARM_NEON
-        for (; i + 15 < size; i += 16)
-        {
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "prfm   pldl1keep, [%0, #256]   \n"
-                "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%0], #32 \n"
-                "shll   v0.4s, v0.4h, #16       \n"
-                "shll   v1.4s, v1.4h, #16       \n"
-                "shll   v2.4s, v2.4h, #16       \n"
-                "shll   v3.4s, v3.4h, #16       \n"
-                "st1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "v0", "v1", "v2", "v3");
-#else  // __aarch64__
-            asm volatile(
-                "pld        [%0, #256]      \n"
-                "vld1.u16   {d4-d7}, [%0]!  \n"
-                "vshll.u16  q0, d4, #16     \n"
-                "vshll.u16  q1, d5, #16     \n"
-                "vshll.u16  q2, d6, #16     \n"
-                "vshll.u16  q3, d7, #16     \n"
-                "vstm       %1!, {d0-d7}    \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "q0", "q1", "q2", "q3");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-#if __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-            bfloat16x8_t _p_bf16 = vld1q_bf16(ptr);
-            bfloat16x8_t _q_bf16 = vld1q_bf16(ptr + 8);
-            float32x4_t _p0_fp32 = vcvt_f32_bf16(vget_low_bf16(_p_bf16));
-            float32x4_t _p1_fp32 = vcvt_f32_bf16(vget_high_bf16(_p_bf16));
-            float32x4_t _p2_fp32 = vcvt_f32_bf16(vget_low_bf16(_q_bf16));
-            float32x4_t _p3_fp32 = vcvt_f32_bf16(vget_high_bf16(_q_bf16));
-#else
-            uint16x8_t _p_bf16 = vld1q_u16(ptr);
-            uint16x8_t _q_bf16 = vld1q_u16(ptr + 8);
-            float32x4_t _p0_fp32 = bfloat2float(vget_low_u16(_p_bf16));
-            float32x4_t _p1_fp32 = bfloat2float(vget_high_u16(_p_bf16));
-            float32x4_t _p2_fp32 = bfloat2float(vget_low_u16(_q_bf16));
-            float32x4_t _p3_fp32 = bfloat2float(vget_high_u16(_q_bf16));
-#endif
-            vst1q_f32(outptr, _p0_fp32);
-            vst1q_f32(outptr + 4, _p1_fp32);
-            vst1q_f32(outptr + 8, _p2_fp32);
-            vst1q_f32(outptr + 12, _p3_fp32);
-            ptr += 16;
-            outptr += 16;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; i + 7 < size; i += 8)
-        {
-#if __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-            bfloat16x8_t _p_bf16 = vld1q_bf16(ptr);
-            float32x4_t _p0_fp32 = vcvt_f32_bf16(vget_low_bf16(_p_bf16));
-            float32x4_t _p1_fp32 = vcvt_f32_bf16(vget_high_bf16(_p_bf16));
-#else
-            uint16x8_t _p_bf16 = vld1q_u16(ptr);
-            float32x4_t _p0_fp32 = bfloat2float(vget_low_u16(_p_bf16));
-            float32x4_t _p1_fp32 = bfloat2float(vget_high_u16(_p_bf16));
-#endif
-            vst1q_f32(outptr, _p0_fp32);
-            vst1q_f32(outptr + 4, _p1_fp32);
-            ptr += 8;
-            outptr += 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-#if __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-            bfloat16x4_t _p_bf16 = vld1_bf16(ptr);
-            float32x4_t _p_fp32 = vcvt_f32_bf16(_p_bf16);
-#else
-            uint16x4_t _p_bf16 = vld1_u16(ptr);
-            float32x4_t _p_fp32 = bfloat2float(_p_bf16);
-#endif
-            vst1q_f32(outptr, _p_fp32);
-            ptr += 4;
-            outptr += 4;
-        }
-#endif
-        for (; i < size; i++)
-        {
-#if __ARM_FEATURE_BF16_VECTOR_ARITHMETIC
-            *outptr++ = vcvtah_f32_bf16(*ptr++);
-#else
-            *outptr++ = bfloat16_to_float32(*ptr++);
-#endif
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/cast_fp16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/cast_fp16.h
deleted file mode 100644
index 7d5866d0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/cast_fp16.h
+++ /dev/null
@@ -1,325 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if NCNN_RUNTIME_CPU && NCNN_VFPV4 && __ARM_NEON && !(__ARM_FP & 2)
-void cast_fp32_to_fp16_neon_vfpv4(const Mat& bottom_blob, Mat& top_blob, const Option& opt);
-void cast_fp16_to_fp32_neon_vfpv4(const Mat& bottom_blob, Mat& top_blob, const Option& opt);
-#endif
-
-static void cast_fp32_to_fp16_neon(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-#if NCNN_RUNTIME_CPU && NCNN_VFPV4 && __ARM_NEON && !(__ARM_FP & 2)
-    if (ncnn::cpu_support_arm_vfpv4())
-    {
-        cast_fp32_to_fp16_neon_vfpv4(bottom_blob, top_blob, opt);
-        return;
-    }
-#endif
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int d = bottom_blob.d;
-    const int channels = bottom_blob.c;
-    const int elempack = bottom_blob.elempack;
-
-    const int size = w * h * d * elempack;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* ptr = bottom_blob.channel(q);
-        unsigned short* outptr = top_blob.channel(q);
-
-        int i = 0;
-#if (__ARM_FP & 2)
-        for (; i + 15 < size; i += 16)
-        {
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "prfm   pldl1keep, [%0, #512]       \n"
-                "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-                "fcvtn  v0.4h, v0.4s                \n"
-                "fcvtn  v1.4h, v1.4s                \n"
-                "fcvtn  v2.4h, v2.4s                \n"
-                "fcvtn  v3.4h, v3.4s                \n"
-                "st1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%1], #32 \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "v0", "v1", "v2", "v3");
-#else  // __aarch64__
-            asm volatile(
-                "pld        [%0, #512]      \n"
-                "vldm       %0!, {d0-d7}    \n"
-                "vcvt.f16.f32 d0, q0        \n"
-                "vcvt.f16.f32 d1, q1        \n"
-                "vcvt.f16.f32 d2, q2        \n"
-                "vcvt.f16.f32 d3, q3        \n"
-                "vst1.u16   {d0-d3}, [%1]!  \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "q0", "q1", "q2", "q3");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _p0_fp32 = vld1q_f32(ptr);
-            float32x4_t _p1_fp32 = vld1q_f32(ptr + 4);
-            float32x4_t _p2_fp32 = vld1q_f32(ptr + 8);
-            float32x4_t _p3_fp32 = vld1q_f32(ptr + 12);
-            uint16x4_t _p0_fp16 = (uint16x4_t)vcvt_f16_f32(_p0_fp32);
-            uint16x4_t _p1_fp16 = (uint16x4_t)vcvt_f16_f32(_p1_fp32);
-            uint16x4_t _p2_fp16 = (uint16x4_t)vcvt_f16_f32(_p2_fp32);
-            uint16x4_t _p3_fp16 = (uint16x4_t)vcvt_f16_f32(_p3_fp32);
-            uint16x8_t _p_fp16 = vcombine_u16(_p0_fp16, _p1_fp16);
-            uint16x8_t _q_fp16 = vcombine_u16(_p2_fp16, _p3_fp16);
-            vst1q_u16(outptr, _p_fp16);
-            vst1q_u16(outptr + 8, _q_fp16);
-            ptr += 16;
-            outptr += 16;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; i + 7 < size; i += 8)
-        {
-#if NCNN_GNU_INLINE_ASM
-            // This is originally implemented with neon fp16 intrinsics.
-            // In the new version of gcc, __ARM_FP16_FORMAT_IEEE or __ARM_FP16_FORMAT_ALTERNATIVE needs to be defined to use the float16x4_t type.
-            // That leads to compiler error when compiled with -mfpu=neon-vfpv4 but without -mfp16-format=ieee flag.
-            // We could add more macro conditions to differentiate between old and new versions, but that's pretty ugly!
-            // Just use all inline assembly here ~
-            //          --- nihui
-#if __aarch64__
-            asm volatile(
-                "ld1    {v0.4s, v1.4s}, [%0], #32   \n"
-                "fcvtn  v0.4h, v0.4s                \n"
-                "fcvtn  v1.4h, v1.4s                \n"
-                "st1    {v0.4h, v1.4h}, [%1], #16   \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "v0", "v1");
-#else  // __aarch64__
-            asm volatile(
-                "vld1.f32   {d0-d3}, [%0]!  \n"
-                "vcvt.f16.f32 d0, q0        \n"
-                "vcvt.f16.f32 d1, q1        \n"
-                "vst1.u16   {d0-d1}, [%1]!  \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "q0", "q1");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _p0_fp32 = vld1q_f32(ptr);
-            float32x4_t _p1_fp32 = vld1q_f32(ptr + 4);
-            uint16x4_t _p0_fp16 = (uint16x4_t)vcvt_f16_f32(_p0_fp32);
-            uint16x4_t _p1_fp16 = (uint16x4_t)vcvt_f16_f32(_p1_fp32);
-            uint16x8_t _p_fp16 = vcombine_u16(_p0_fp16, _p1_fp16);
-            vst1q_u16(outptr, _p_fp16);
-            ptr += 8;
-            outptr += 8;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; i + 3 < size; i += 4)
-        {
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "ld1    {v0.4s}, [%0], #16  \n"
-                "fcvtn  v0.4h, v0.4s        \n"
-                "st1    {v0.4h}, [%1], #8   \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "v0");
-#else  // __aarch64__
-            asm volatile(
-                "vld1.f32   {d0-d1}, [%0]!  \n"
-                "vcvt.f16.f32 d0, q0        \n"
-                "vst1.u16   {d0}, [%1]!     \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "q0");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _p_fp32 = vld1q_f32(ptr);
-            uint16x4_t _p_fp16 = (uint16x4_t)vcvt_f16_f32(_p_fp32);
-            vst1_u16(outptr, _p_fp16);
-            ptr += 4;
-            outptr += 4;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-#endif // (__ARM_FP & 2)
-        for (; i < size; i++)
-        {
-            *outptr++ = float32_to_float16(*ptr++);
-        }
-    }
-}
-
-static void cast_fp16_to_fp32_neon(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-#if NCNN_RUNTIME_CPU && NCNN_VFPV4 && __ARM_NEON && !(__ARM_FP & 2)
-    if (ncnn::cpu_support_arm_vfpv4())
-    {
-        cast_fp16_to_fp32_neon_vfpv4(bottom_blob, top_blob, opt);
-        return;
-    }
-#endif
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int d = bottom_blob.d;
-    const int channels = bottom_blob.c;
-    const int elempack = bottom_blob.elempack;
-
-    const int size = w * h * d * elempack;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const unsigned short* ptr = bottom_blob.channel(q);
-        float* outptr = top_blob.channel(q);
-
-        int i = 0;
-#if (__ARM_FP & 2)
-        for (; i + 15 < size; i += 16)
-        {
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "prfm   pldl1keep, [%0, #256]       \n"
-                "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%0], #32 \n"
-                "fcvtl  v0.4s, v0.4h                \n"
-                "fcvtl  v1.4s, v1.4h                \n"
-                "fcvtl  v2.4s, v2.4h                \n"
-                "fcvtl  v3.4s, v3.4h                \n"
-                "st1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "v0", "v1", "v2", "v3");
-#else  // __aarch64__
-            asm volatile(
-                "pld        [%0, #256]      \n"
-                "vld1.u16   {d4-d7}, [%0]!  \n"
-                "vcvt.f32.f16 q0, d4        \n"
-                "vcvt.f32.f16 q1, d5        \n"
-                "vcvt.f32.f16 q2, d6        \n"
-                "vcvt.f32.f16 q3, d7        \n"
-                "vstm       %1!, {d0-d7}    \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "q0", "q1", "q2", "q3");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            uint16x8_t _p_fp16 = vld1q_u16(ptr);
-            uint16x8_t _q_fp16 = vld1q_u16(ptr + 8);
-            float32x4_t _p0_fp32 = vcvt_f32_f16((float16x4_t)vget_low_u16(_p_fp16));
-            float32x4_t _p1_fp32 = vcvt_f32_f16((float16x4_t)vget_high_u16(_p_fp16));
-            float32x4_t _p2_fp32 = vcvt_f32_f16((float16x4_t)vget_low_u16(_q_fp16));
-            float32x4_t _p3_fp32 = vcvt_f32_f16((float16x4_t)vget_high_u16(_q_fp16));
-            vst1q_f32(outptr, _p0_fp32);
-            vst1q_f32(outptr + 4, _p1_fp32);
-            vst1q_f32(outptr + 8, _p2_fp32);
-            vst1q_f32(outptr + 12, _p3_fp32);
-            ptr += 16;
-            outptr += 16;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; i + 7 < size; i += 8)
-        {
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "ld1    {v0.4h, v1.4h}, [%0], #16   \n"
-                "fcvtl  v0.4s, v0.4h                \n"
-                "fcvtl  v1.4s, v1.4h                \n"
-                "st1    {v0.4s, v1.4s}, [%1], #32   \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "v0", "v1");
-#else  // __aarch64__
-            asm volatile(
-                "vld1.u16   {d4-d5}, [%0]!  \n"
-                "vcvt.f32.f16 q0, d4        \n"
-                "vcvt.f32.f16 q1, d5        \n"
-                "vst1.f32   {d0-d3}, [%1]!  \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "q0", "q1", "q2");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            uint16x8_t _p_fp16 = vld1q_u16(ptr);
-            float32x4_t _p0_fp32 = vcvt_f32_f16((float16x4_t)vget_low_u16(_p_fp16));
-            float32x4_t _p1_fp32 = vcvt_f32_f16((float16x4_t)vget_high_u16(_p_fp16));
-            vst1q_f32(outptr, _p0_fp32);
-            vst1q_f32(outptr + 4, _p1_fp32);
-            ptr += 8;
-            outptr += 8;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; i + 3 < size; i += 4)
-        {
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "ld1    {v0.4h}, [%0], #8   \n"
-                "fcvtl  v0.4s, v0.4h        \n"
-                "st1    {v0.4s}, [%1], #16  \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "v0");
-#else  // __aarch64__
-            asm volatile(
-                "vld1.u16   {d2}, [%0]!     \n"
-                "vcvt.f32.f16 q0, d2        \n"
-                "vst1.f32   {d0-d1}, [%1]!  \n"
-                : "=r"(ptr),   // %0
-                "=r"(outptr) // %1
-                : "0"(ptr),
-                "1"(outptr)
-                : "memory", "q0", "q1");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            uint16x4_t _p_fp16 = vld1_u16(ptr);
-            float32x4_t _p_fp32 = vcvt_f32_f16((float16x4_t)_p_fp16);
-            vst1q_f32(outptr, _p_fp32);
-            ptr += 4;
-            outptr += 4;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-#endif // (__ARM_FP & 2)
-        for (; i < size; i++)
-        {
-            *outptr++ = float16_to_float32(*ptr++);
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/clip_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/clip_arm.h
deleted file mode 100644
index ef281d24..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/clip_arm.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CLIP_ARM_H
-#define LAYER_CLIP_ARM_H
-
-#include "clip.h"
-
-namespace ncnn {
-
-class Clip_arm : public Clip
-{
-public:
-    Clip_arm();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_inplace_bf16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CLIP_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/concat_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/concat_arm.h
deleted file mode 100644
index 9491a280..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/concat_arm.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONCAT_ARM_H
-#define LAYER_CONCAT_ARM_H
-
-#include "concat.h"
-
-namespace ncnn {
-
-class Concat_arm : public Concat
-{
-public:
-    Concat_arm();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    int forward_bf16s_fp16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONCAT_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_arm.h
deleted file mode 100644
index 48babb91..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_arm.h
+++ /dev/null
@@ -1,54 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTION1D_ARM_H
-#define LAYER_CONVOLUTION1D_ARM_H
-
-#include "convolution1d.h"
-
-namespace ncnn {
-
-class Convolution1D_arm : public Convolution1D
-{
-public:
-    Convolution1D_arm();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int create_pipeline_bf16s(const Option& opt);
-    int forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Mat weight_data_tm;
-
-    // fp16
-    Mat bias_data_fp16;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTION1D_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_packed.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_packed.h
deleted file mode 100644
index 7b487413..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_packed.h
+++ /dev/null
@@ -1,1293 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution1d_transform_kernel_packed(const Mat& kernel, Mat& kernel_tm, int inh, int outh, int kernel_w)
-{
-    // src = kw-inh-outh
-    // dst = pb-pa-kw-inh/pa-outh/pb
-
-    // clang-format off
-    // *INDENT-OFF*
-#if __ARM_NEON
-#if __aarch64__
-    if (outh >= 8)
-    {
-        if (inh >= 8)
-            kernel_tm.create(8 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2);
-        else if (inh >= 4)
-            kernel_tm.create(8 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2);
-        else if (inh >= 2)
-            kernel_tm.create(8 * 2 * kernel_w, inh / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2);
-        else
-            kernel_tm.create(8 * kernel_w, inh, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2);
-    }
-    else
-#endif // __aarch64__
-    if (outh >= 4)
-    {
-#if __aarch64__
-        if (inh >= 8)
-            kernel_tm.create(4 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2);
-        else
-#endif // __aarch64__
-        if (inh >= 4)
-            kernel_tm.create(4 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2);
-        else if (inh >= 2)
-            kernel_tm.create(4 * 2 * kernel_w, inh / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2);
-        else
-            kernel_tm.create(4 * kernel_w, inh, outh / 4 + (outh % 4) / 2 + outh % 2);
-    }
-    else
-#endif // __ARM_NEON
-    if (outh >= 2)
-    {
-#if __ARM_NEON
-#if __aarch64__
-        if (inh >= 8)
-            kernel_tm.create(2 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 2 + outh % 2);
-        else
-#endif // __aarch64__
-        if (inh >= 4)
-            kernel_tm.create(2 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 2 + outh % 2);
-        else
-#endif // __ARM_NEON
-        if (inh >= 2)
-            kernel_tm.create(2 * 2 * kernel_w, inh / 2 + inh % 2, outh / 2 + outh % 2);
-        else
-            kernel_tm.create(2 * kernel_w, inh, outh / 2 + outh % 2);
-    }
-    else
-    {
-#if __ARM_NEON
-#if __aarch64__
-        if (inh >= 8)
-            kernel_tm.create(8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh);
-        else
-#endif // __aarch64__
-        if (inh >= 4)
-            kernel_tm.create(4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh);
-        else
-#endif // __ARM_NEON
-        if (inh >= 2)
-            kernel_tm.create(2 * kernel_w, inh / 2 + inh % 2, outh);
-        else
-            kernel_tm.create(kernel_w, inh, outh);
-    }
-    // *INDENT-ON*
-    // clang-format on
-
-    int q = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; q + 7 < outh; q += 8)
-    {
-        const float* kptr0 = (const float*)kernel + q * inh * kernel_w;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inh * kernel_w;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inh * kernel_w;
-        const float* kptr4 = (const float*)kernel + (q + 4) * inh * kernel_w;
-        const float* kptr5 = (const float*)kernel + (q + 5) * inh * kernel_w;
-        const float* kptr6 = (const float*)kernel + (q + 6) * inh * kernel_w;
-        const float* kptr7 = (const float*)kernel + (q + 7) * inh * kernel_w;
-
-        float* g00 = kernel_tm.channel(q / 8);
-
-        int p = 0;
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-                const float* k4 = kptr4 + p * kernel_w;
-                const float* k5 = kptr5 + p * kernel_w;
-                const float* k6 = kptr6 + p * kernel_w;
-                const float* k7 = kptr7 + p * kernel_w;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[k];
-                    g00[1] = k1[k];
-                    g00[2] = k2[k];
-                    g00[3] = k3[k];
-                    g00[4] = k4[k];
-                    g00[5] = k5[k];
-                    g00[6] = k6[k];
-                    g00[7] = k7[k];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    k4 += kernel_w;
-                    k5 += kernel_w;
-                    k6 += kernel_w;
-                    k7 += kernel_w;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-                const float* k4 = kptr4 + p * kernel_w;
-                const float* k5 = kptr5 + p * kernel_w;
-                const float* k6 = kptr6 + p * kernel_w;
-                const float* k7 = kptr7 + p * kernel_w;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k0[k];
-                    g00[1] = k1[k];
-                    g00[2] = k2[k];
-                    g00[3] = k3[k];
-                    g00[4] = k4[k];
-                    g00[5] = k5[k];
-                    g00[6] = k6[k];
-                    g00[7] = k7[k];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    k4 += kernel_w;
-                    k5 += kernel_w;
-                    k6 += kernel_w;
-                    k7 += kernel_w;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-                const float* k4 = kptr4 + p * kernel_w;
-                const float* k5 = kptr5 + p * kernel_w;
-                const float* k6 = kptr6 + p * kernel_w;
-                const float* k7 = kptr7 + p * kernel_w;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = k0[k];
-                    g00[1] = k1[k];
-                    g00[2] = k2[k];
-                    g00[3] = k3[k];
-                    g00[4] = k4[k];
-                    g00[5] = k5[k];
-                    g00[6] = k6[k];
-                    g00[7] = k7[k];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    k4 += kernel_w;
-                    k5 += kernel_w;
-                    k6 += kernel_w;
-                    k7 += kernel_w;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p < inh; p++)
-        {
-            const float* k0 = kptr0 + p * kernel_w;
-            const float* k1 = kptr1 + p * kernel_w;
-            const float* k2 = kptr2 + p * kernel_w;
-            const float* k3 = kptr3 + p * kernel_w;
-            const float* k4 = kptr4 + p * kernel_w;
-            const float* k5 = kptr5 + p * kernel_w;
-            const float* k6 = kptr6 + p * kernel_w;
-            const float* k7 = kptr7 + p * kernel_w;
-
-            for (int k = 0; k < kernel_w; k++)
-            {
-                g00[0] = k0[k];
-                g00[1] = k1[k];
-                g00[2] = k2[k];
-                g00[3] = k3[k];
-                g00[4] = k4[k];
-                g00[5] = k5[k];
-                g00[6] = k6[k];
-                g00[7] = k7[k];
-                g00 += 8;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; q + 3 < outh; q += 4)
-    {
-        const float* kptr0 = (const float*)kernel + q * inh * kernel_w;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inh * kernel_w;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inh * kernel_w;
-
-#if __aarch64__
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4);
-#else
-        float* g00 = kernel_tm.channel(q / 4);
-#endif
-
-        int p = 0;
-#if __aarch64__
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[k];
-                    g00[1] = k1[k];
-                    g00[2] = k2[k];
-                    g00[3] = k3[k];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    g00 += 4;
-                }
-            }
-        }
-#endif // __aarch64__
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k0[k];
-                    g00[1] = k1[k];
-                    g00[2] = k2[k];
-                    g00[3] = k3[k];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    g00 += 4;
-                }
-            }
-        }
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = k0[k];
-                    g00[1] = k1[k];
-                    g00[2] = k2[k];
-                    g00[3] = k3[k];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    g00 += 4;
-                }
-            }
-        }
-        for (; p < inh; p++)
-        {
-            const float* k0 = kptr0 + p * kernel_w;
-            const float* k1 = kptr1 + p * kernel_w;
-            const float* k2 = kptr2 + p * kernel_w;
-            const float* k3 = kptr3 + p * kernel_w;
-
-            for (int k = 0; k < kernel_w; k++)
-            {
-                g00[0] = k0[k];
-                g00[1] = k1[k];
-                g00[2] = k2[k];
-                g00[3] = k3[k];
-                g00 += 4;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; q + 1 < outh; q += 2)
-    {
-        const float* kptr0 = (const float*)kernel + q * inh * kernel_w;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w;
-
-#if __aarch64__
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2);
-#elif __ARM_NEON
-        float* g00 = kernel_tm.channel(q / 4 + (q % 4) / 2);
-#else
-        float* g00 = kernel_tm.channel(q / 2);
-#endif
-
-        int p = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w + k;
-                const float* k1 = kptr1 + p * kernel_w + k;
-
-                g00[0] = k0[0];
-                g00[1] = k0[kernel_w];
-                g00[2] = k0[kernel_w * 2];
-                g00[3] = k0[kernel_w * 3];
-                g00[4] = k0[kernel_w * 4];
-                g00[5] = k0[kernel_w * 5];
-                g00[6] = k0[kernel_w * 6];
-                g00[7] = k0[kernel_w * 7];
-                g00[8] = k1[0];
-                g00[9] = k1[kernel_w];
-                g00[10] = k1[kernel_w * 2];
-                g00[11] = k1[kernel_w * 3];
-                g00[12] = k1[kernel_w * 4];
-                g00[13] = k1[kernel_w * 5];
-                g00[14] = k1[kernel_w * 6];
-                g00[15] = k1[kernel_w * 7];
-                g00 += 16;
-            }
-        }
-#endif // __aarch64__
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w + k;
-                const float* k1 = kptr1 + p * kernel_w + k;
-
-                g00[0] = k0[0];
-                g00[1] = k0[kernel_w];
-                g00[2] = k0[kernel_w * 2];
-                g00[3] = k0[kernel_w * 3];
-                g00[4] = k1[0];
-                g00[5] = k1[kernel_w];
-                g00[6] = k1[kernel_w * 2];
-                g00[7] = k1[kernel_w * 3];
-                g00 += 8;
-            }
-        }
-#endif // __ARM_NEON
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = k0[k];
-                    g00[1] = k1[k];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    g00 += 2;
-                }
-            }
-        }
-        for (; p < inh; p++)
-        {
-            const float* k0 = kptr0 + p * kernel_w;
-            const float* k1 = kptr1 + p * kernel_w;
-
-            for (int k = 0; k < kernel_w; k++)
-            {
-                g00[0] = k0[k];
-                g00[1] = k1[k];
-                g00 += 2;
-            }
-        }
-    }
-    for (; q < outh; q++)
-    {
-        const float* kptr = (const float*)kernel + q * inh * kernel_w;
-
-#if __aarch64__
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2 + q % 2);
-#elif __ARM_NEON
-        float* g00 = kernel_tm.channel(q / 4 + (q % 4) / 2 + q % 2);
-#else
-        float* g00 = kernel_tm.channel(q / 2 + q % 2);
-#endif
-
-        int p = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr + p * kernel_w;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[k];
-                    k0 += kernel_w;
-                    g00 += 1;
-                }
-            }
-        }
-#endif // __aarch64__
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr + p * kernel_w;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k0[k];
-                    k0 += kernel_w;
-                    g00 += 1;
-                }
-            }
-        }
-#endif // __ARM_NEON
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr + p * kernel_w;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = k0[k];
-                    k0 += kernel_w;
-                    g00 += 1;
-                }
-            }
-        }
-        for (; p < inh; p++)
-        {
-            const float* k0 = kptr + p * kernel_w;
-
-            for (int k = 0; k < kernel_w; k++)
-            {
-                g00[0] = k0[k];
-                g00++;
-            }
-        }
-    }
-}
-
-static void convolution1d_packed(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int kernel_w, int dilation_w, int stride_w, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int elempack = bottom_blob.elempack;
-    const int inh = bottom_blob.h * elempack;
-
-    const int N = bottom_blob.w * elempack;
-
-    const int outw = top_blob.w;
-    const int out_elempack = top_blob.elempack;
-    const int outh = top_blob.h * out_elempack;
-
-    const int M = top_blob.w * out_elempack;
-
-    const float* bias_data_ptr = bias_data;
-
-    int nn_outh = 0;
-    int remain_outh_start = 0;
-#if __ARM_NEON
-#if __aarch64__
-    nn_outh = (outh - remain_outh_start) / 8;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = remain_outh_start + pp * 8;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-        const int out_elempack = top_blob.elempack;
-
-        float* outptr = top_blob.row(p / out_elempack);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
-            float32x4_t _sum1 = vdupq_n_f32(0.f);
-            float32x4_t _sum2 = vdupq_n_f32(0.f);
-            float32x4_t _sum3 = vdupq_n_f32(0.f);
-            float32x4_t _sum4 = vdupq_n_f32(0.f);
-            float32x4_t _sum5 = vdupq_n_f32(0.f);
-            float32x4_t _sum6 = vdupq_n_f32(0.f);
-            float32x4_t _sum7 = vdupq_n_f32(0.f);
-
-            if (bias_data_ptr)
-            {
-                _sum0 = vld1q_f32(bias_data_ptr + p);
-                _sum1 = vld1q_f32(bias_data_ptr + p + 4);
-            }
-
-            const float* kptr = weight_data_tm.channel(p / 8);
-
-            int q = 0;
-            for (; q + 7 < inh; q += 8)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    float32x4_t _r1;
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1q_f32(r0);
-                        _r1 = vld1q_f32(r0 + N);
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float32x4_t();
-                        _r1 = float32x4_t();
-                        _r0 = vsetq_lane_f32(r0[0], _r0, 0);
-                        _r0 = vsetq_lane_f32(r0[N], _r0, 1);
-                        _r0 = vsetq_lane_f32(r0[N * 2], _r0, 2);
-                        _r0 = vsetq_lane_f32(r0[N * 3], _r0, 3);
-                        _r1 = vsetq_lane_f32(r0[N * 4], _r1, 0);
-                        _r1 = vsetq_lane_f32(r0[N * 5], _r1, 1);
-                        _r1 = vsetq_lane_f32(r0[N * 6], _r1, 2);
-                        _r1 = vsetq_lane_f32(r0[N * 7], _r1, 3);
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w0 = vld1q_f32(kptr);
-                    float32x4_t _w1 = vld1q_f32(kptr + 4);
-                    float32x4_t _w2 = vld1q_f32(kptr + 4 * 2);
-                    float32x4_t _w3 = vld1q_f32(kptr + 4 * 3);
-                    float32x4_t _w4 = vld1q_f32(kptr + 4 * 4);
-                    float32x4_t _w5 = vld1q_f32(kptr + 4 * 5);
-                    float32x4_t _w6 = vld1q_f32(kptr + 4 * 6);
-                    float32x4_t _w7 = vld1q_f32(kptr + 4 * 7);
-                    float32x4_t _w8 = vld1q_f32(kptr + 4 * 8);
-                    float32x4_t _w9 = vld1q_f32(kptr + 4 * 9);
-                    float32x4_t _wa = vld1q_f32(kptr + 4 * 10);
-                    float32x4_t _wb = vld1q_f32(kptr + 4 * 11);
-                    float32x4_t _wc = vld1q_f32(kptr + 4 * 12);
-                    float32x4_t _wd = vld1q_f32(kptr + 4 * 13);
-                    float32x4_t _we = vld1q_f32(kptr + 4 * 14);
-                    float32x4_t _wf = vld1q_f32(kptr + 4 * 15);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 0);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 1);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 1);
-                    _sum4 = vfmaq_laneq_f32(_sum4, _w4, _r0, 2);
-                    _sum5 = vfmaq_laneq_f32(_sum5, _w5, _r0, 2);
-                    _sum6 = vfmaq_laneq_f32(_sum6, _w6, _r0, 3);
-                    _sum7 = vfmaq_laneq_f32(_sum7, _w7, _r0, 3);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w8, _r1, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w9, _r1, 0);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _wa, _r1, 1);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _wb, _r1, 1);
-                    _sum4 = vfmaq_laneq_f32(_sum4, _wc, _r1, 2);
-                    _sum5 = vfmaq_laneq_f32(_sum5, _wd, _r1, 2);
-                    _sum6 = vfmaq_laneq_f32(_sum6, _we, _r1, 3);
-                    _sum7 = vfmaq_laneq_f32(_sum7, _wf, _r1, 3);
-
-                    kptr += 64;
-                }
-            }
-            for (; q + 3 < inh; q += 4)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1q_f32(r0);
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float32x4_t();
-                        _r0 = vsetq_lane_f32(r0[0], _r0, 0);
-                        _r0 = vsetq_lane_f32(r0[N], _r0, 1);
-                        _r0 = vsetq_lane_f32(r0[N * 2], _r0, 2);
-                        _r0 = vsetq_lane_f32(r0[N * 3], _r0, 3);
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w0 = vld1q_f32(kptr);
-                    float32x4_t _w1 = vld1q_f32(kptr + 4);
-                    float32x4_t _w2 = vld1q_f32(kptr + 4 * 2);
-                    float32x4_t _w3 = vld1q_f32(kptr + 4 * 3);
-                    float32x4_t _w4 = vld1q_f32(kptr + 4 * 4);
-                    float32x4_t _w5 = vld1q_f32(kptr + 4 * 5);
-                    float32x4_t _w6 = vld1q_f32(kptr + 4 * 6);
-                    float32x4_t _w7 = vld1q_f32(kptr + 4 * 7);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 0);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 1);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 1);
-                    _sum4 = vfmaq_laneq_f32(_sum4, _w4, _r0, 2);
-                    _sum5 = vfmaq_laneq_f32(_sum5, _w5, _r0, 2);
-                    _sum6 = vfmaq_laneq_f32(_sum6, _w6, _r0, 3);
-                    _sum7 = vfmaq_laneq_f32(_sum7, _w7, _r0, 3);
-
-                    kptr += 32;
-                }
-            }
-            for (; q + 1 < inh; q += 2)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val0;
-                    float val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = r0[0];
-                        val1 = r0[N];
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w0 = vld1q_f32(kptr);
-                    float32x4_t _w1 = vld1q_f32(kptr + 4);
-                    float32x4_t _w2 = vld1q_f32(kptr + 8);
-                    float32x4_t _w3 = vld1q_f32(kptr + 12);
-                    _sum0 = vfmaq_n_f32(_sum0, _w0, val0);
-                    _sum1 = vfmaq_n_f32(_sum1, _w1, val0);
-                    _sum2 = vfmaq_n_f32(_sum2, _w2, val1);
-                    _sum3 = vfmaq_n_f32(_sum3, _w3, val1);
-
-                    kptr += 16;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _val;
-                    // if (elempack == 1)
-                    {
-                        _val = vdupq_n_f32(r0[0]);
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w0 = vld1q_f32(kptr);
-                    float32x4_t _w1 = vld1q_f32(kptr + 4);
-                    _sum0 = vfmaq_f32(_sum0, _w0, _val);
-                    _sum1 = vfmaq_f32(_sum1, _w1, _val);
-
-                    kptr += 8;
-                }
-            }
-
-            _sum0 = vaddq_f32(_sum0, _sum2);
-            _sum1 = vaddq_f32(_sum1, _sum3);
-            _sum4 = vaddq_f32(_sum4, _sum6);
-            _sum5 = vaddq_f32(_sum5, _sum7);
-            _sum0 = vaddq_f32(_sum0, _sum4);
-            _sum1 = vaddq_f32(_sum1, _sum5);
-
-            _sum0 = activation_ps(_sum0, activation_type, activation_params);
-            _sum1 = activation_ps(_sum1, activation_type, activation_params);
-
-            if (out_elempack == 4)
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + M, _sum1);
-                outptr += 4;
-            }
-            else // if (out_elempack == 1)
-            {
-                outptr[0] = vgetq_lane_f32(_sum0, 0);
-                outptr[M] = vgetq_lane_f32(_sum0, 1);
-                outptr[M * 2] = vgetq_lane_f32(_sum0, 2);
-                outptr[M * 3] = vgetq_lane_f32(_sum0, 3);
-                outptr[M * 4] = vgetq_lane_f32(_sum1, 0);
-                outptr[M * 5] = vgetq_lane_f32(_sum1, 1);
-                outptr[M * 6] = vgetq_lane_f32(_sum1, 2);
-                outptr[M * 7] = vgetq_lane_f32(_sum1, 3);
-                outptr += 1;
-            }
-        }
-    }
-    remain_outh_start += nn_outh * 8;
-    nn_outh = (outh - remain_outh_start) / 4;
-#else // __aarch64__
-    nn_outh = (outh - remain_outh_start) / 4;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __aarch64__
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = remain_outh_start + pp * 4;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-        const int out_elempack = top_blob.elempack;
-
-        float* outptr = top_blob.row(p / out_elempack);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
-            float32x4_t _sum1 = vdupq_n_f32(0.f);
-            float32x4_t _sum2 = vdupq_n_f32(0.f);
-            float32x4_t _sum3 = vdupq_n_f32(0.f);
-
-            if (bias_data_ptr)
-            {
-                _sum0 = vld1q_f32(bias_data_ptr + p);
-            }
-
-#if __aarch64__
-            const float* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4);
-#else
-            const float* kptr = weight_data_tm.channel(p / 4);
-#endif
-
-            int q = 0;
-#if __aarch64__
-            for (; q + 7 < inh; q += 8)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    float32x4_t _r1;
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1q_f32(r0);
-                        _r1 = vld1q_f32(r0 + N);
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float32x4_t();
-                        _r1 = float32x4_t();
-                        _r0 = vsetq_lane_f32(r0[0], _r0, 0);
-                        _r0 = vsetq_lane_f32(r0[N], _r0, 1);
-                        _r0 = vsetq_lane_f32(r0[N * 2], _r0, 2);
-                        _r0 = vsetq_lane_f32(r0[N * 3], _r0, 3);
-                        _r1 = vsetq_lane_f32(r0[N * 4], _r1, 0);
-                        _r1 = vsetq_lane_f32(r0[N * 5], _r1, 1);
-                        _r1 = vsetq_lane_f32(r0[N * 6], _r1, 2);
-                        _r1 = vsetq_lane_f32(r0[N * 7], _r1, 3);
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w0 = vld1q_f32(kptr);
-                    float32x4_t _w1 = vld1q_f32(kptr + 4);
-                    float32x4_t _w2 = vld1q_f32(kptr + 8);
-                    float32x4_t _w3 = vld1q_f32(kptr + 12);
-                    float32x4_t _w4 = vld1q_f32(kptr + 16);
-                    float32x4_t _w5 = vld1q_f32(kptr + 20);
-                    float32x4_t _w6 = vld1q_f32(kptr + 24);
-                    float32x4_t _w7 = vld1q_f32(kptr + 28);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 1);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 2);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 3);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w4, _r1, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w5, _r1, 1);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w6, _r1, 2);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w7, _r1, 3);
-
-                    kptr += 32;
-                }
-            }
-#endif // __aarch64__
-            for (; q + 3 < inh; q += 4)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1q_f32(r0);
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float32x4_t();
-                        _r0 = vsetq_lane_f32(r0[0], _r0, 0);
-                        _r0 = vsetq_lane_f32(r0[N], _r0, 1);
-                        _r0 = vsetq_lane_f32(r0[N * 2], _r0, 2);
-                        _r0 = vsetq_lane_f32(r0[N * 3], _r0, 3);
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w0 = vld1q_f32(kptr);
-                    float32x4_t _w1 = vld1q_f32(kptr + 4);
-                    float32x4_t _w2 = vld1q_f32(kptr + 8);
-                    float32x4_t _w3 = vld1q_f32(kptr + 12);
-#if __aarch64__
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 1);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 2);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 3);
-#else
-                    _sum0 = vmlaq_lane_f32(_sum0, _w0, vget_low_f32(_r0), 0);
-                    _sum1 = vmlaq_lane_f32(_sum1, _w1, vget_low_f32(_r0), 1);
-                    _sum2 = vmlaq_lane_f32(_sum2, _w2, vget_high_f32(_r0), 0);
-                    _sum3 = vmlaq_lane_f32(_sum3, _w3, vget_high_f32(_r0), 1);
-#endif
-
-                    kptr += 16;
-                }
-            }
-            for (; q + 1 < inh; q += 2)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val0;
-                    float val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = r0[0];
-                        val1 = r0[N];
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w0 = vld1q_f32(kptr);
-                    float32x4_t _w1 = vld1q_f32(kptr + 4);
-#if __aarch64__
-                    _sum0 = vfmaq_n_f32(_sum0, _w0, val0);
-                    _sum1 = vfmaq_n_f32(_sum1, _w1, val1);
-#else
-                    _sum0 = vmlaq_n_f32(_sum0, _w0, val0);
-                    _sum1 = vmlaq_n_f32(_sum1, _w1, val1);
-#endif
-
-                    kptr += 8;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _val;
-                    // if (elempack == 1)
-                    {
-                        _val = vdupq_n_f32(r0[0]);
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w = vld1q_f32(kptr);
-#if __aarch64__
-                    _sum0 = vfmaq_f32(_sum0, _val, _w);
-#else
-                    _sum0 = vmlaq_f32(_sum0, _val, _w);
-#endif
-
-                    kptr += 4;
-                }
-            }
-
-            _sum0 = vaddq_f32(_sum0, _sum1);
-            _sum2 = vaddq_f32(_sum2, _sum3);
-            _sum0 = vaddq_f32(_sum0, _sum2);
-
-            _sum0 = activation_ps(_sum0, activation_type, activation_params);
-
-            if (out_elempack == 4)
-            {
-                vst1q_f32(outptr, _sum0);
-                outptr += 4;
-            }
-            else // if (out_elempack == 1)
-            {
-                outptr[0] = vgetq_lane_f32(_sum0, 0);
-                outptr[M] = vgetq_lane_f32(_sum0, 1);
-                outptr[M * 2] = vgetq_lane_f32(_sum0, 2);
-                outptr[M * 3] = vgetq_lane_f32(_sum0, 3);
-                outptr += 1;
-            }
-        }
-    }
-    remain_outh_start += nn_outh * 4;
-    nn_outh = (outh - remain_outh_start) / 2;
-#else // __ARM_NEON
-    nn_outh = (outh - remain_outh_start) / 2;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __ARM_NEON
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = remain_outh_start + pp * 2;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-
-        float* outptr0 = top_blob.row(p);
-        float* outptr1 = top_blob.row(p + 1);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float sum0 = 0.f;
-            float sum1 = 0.f;
-
-            if (bias_data_ptr)
-            {
-                sum0 = bias_data_ptr[p];
-                sum1 = bias_data_ptr[p + 1];
-            }
-
-#if __aarch64__
-            const float* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2);
-#elif __ARM_NEON
-            const float* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2);
-#else
-            const float* kptr = weight_data_tm.channel(p / 2);
-#endif
-
-            int q = 0;
-#if __ARM_NEON
-#if __aarch64__
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
-            float32x4_t _sum1 = vdupq_n_f32(0.f);
-            float32x4_t _sum2 = vdupq_n_f32(0.f);
-            float32x4_t _sum3 = vdupq_n_f32(0.f);
-            for (; q + 7 < inh; q += 8)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    float32x4_t _r1;
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1q_f32(r0);
-                        _r1 = vld1q_f32(r0 + N);
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float32x4_t();
-                        _r1 = float32x4_t();
-                        _r0 = vsetq_lane_f32(r0[0], _r0, 0);
-                        _r0 = vsetq_lane_f32(r0[N], _r0, 1);
-                        _r0 = vsetq_lane_f32(r0[N * 2], _r0, 2);
-                        _r0 = vsetq_lane_f32(r0[N * 3], _r0, 3);
-                        _r1 = vsetq_lane_f32(r0[N * 4], _r1, 0);
-                        _r1 = vsetq_lane_f32(r0[N * 5], _r1, 1);
-                        _r1 = vsetq_lane_f32(r0[N * 6], _r1, 2);
-                        _r1 = vsetq_lane_f32(r0[N * 7], _r1, 3);
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w0 = vld1q_f32(kptr);
-                    float32x4_t _w1 = vld1q_f32(kptr + 4);
-                    float32x4_t _w2 = vld1q_f32(kptr + 8);
-                    float32x4_t _w3 = vld1q_f32(kptr + 12);
-                    _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                    _sum1 = vfmaq_f32(_sum1, _r1, _w1);
-                    _sum2 = vfmaq_f32(_sum2, _r0, _w2);
-                    _sum3 = vfmaq_f32(_sum3, _r1, _w3);
-
-                    kptr += 16;
-                }
-            }
-            _sum0 = vaddq_f32(_sum0, _sum1);
-            _sum2 = vaddq_f32(_sum2, _sum3);
-            sum0 += vaddvq_f32(_sum0);
-            sum1 += vaddvq_f32(_sum2);
-            _sum0 = vdupq_n_f32(0.f);
-            _sum1 = vdupq_n_f32(0.f);
-#else  // __aarch64__
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
-            float32x4_t _sum1 = vdupq_n_f32(0.f);
-#endif // __aarch64__
-            for (; q + 3 < inh; q += 4)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1q_f32(r0);
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float32x4_t();
-                        _r0 = vsetq_lane_f32(r0[0], _r0, 0);
-                        _r0 = vsetq_lane_f32(r0[N], _r0, 1);
-                        _r0 = vsetq_lane_f32(r0[N * 2], _r0, 2);
-                        _r0 = vsetq_lane_f32(r0[N * 3], _r0, 3);
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w0 = vld1q_f32(kptr);
-                    float32x4_t _w1 = vld1q_f32(kptr + 4);
-#if __aarch64__
-                    _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                    _sum1 = vfmaq_f32(_sum1, _r0, _w1);
-#else
-                    _sum0 = vmlaq_f32(_sum0, _r0, _w0);
-                    _sum1 = vmlaq_f32(_sum1, _r0, _w1);
-#endif
-
-                    kptr += 8;
-                }
-            }
-#if __aarch64__
-            sum0 += vaddvq_f32(_sum0);
-            sum1 += vaddvq_f32(_sum1);
-#else
-            float32x2_t _ss0 = vadd_f32(vget_low_f32(_sum0), vget_high_f32(_sum0));
-            float32x2_t _ss1 = vadd_f32(vget_low_f32(_sum1), vget_high_f32(_sum1));
-            float32x2_t _ss = vpadd_f32(_ss0, _ss1);
-            sum0 += vget_lane_f32(_ss, 0);
-            sum1 += vget_lane_f32(_ss, 1);
-#endif
-#endif // __ARM_NEON
-            for (; q + 1 < inh; q += 2)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val0;
-                    float val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = r0[0];
-                        val1 = r0[N];
-                        r0 += dilation_w;
-                    }
-
-                    sum0 += val0 * kptr[0];
-                    sum1 += val0 * kptr[1];
-                    sum0 += val1 * kptr[2];
-                    sum1 += val1 * kptr[3];
-
-                    kptr += 4;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val;
-                    // if (elempack == 1)
-                    {
-                        val = r0[0];
-                        r0 += dilation_w;
-                    }
-
-                    sum0 += val * kptr[0];
-                    sum1 += val * kptr[1];
-
-                    kptr += 2;
-                }
-            }
-
-            sum0 = activation_ss(sum0, activation_type, activation_params);
-            sum1 = activation_ss(sum1, activation_type, activation_params);
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-            outptr0 += 1;
-            outptr1 += 1;
-        }
-    }
-    remain_outh_start += nn_outh * 2;
-    for (int p = remain_outh_start; p < outh; p++)
-    {
-        float* outptr = top_blob.row(p);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float sum = 0.f;
-
-            if (bias_data_ptr)
-            {
-                sum = bias_data_ptr[p];
-            }
-
-#if __aarch64__
-            const float* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#elif __ARM_NEON
-            const float* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2 + p % 2);
-#else
-            const float* kptr = weight_data_tm.channel(p / 2 + p % 2);
-#endif
-
-            int q = 0;
-#if __ARM_NEON
-#if __aarch64__
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
-            float32x4_t _sum1 = vdupq_n_f32(0.f);
-            for (; q + 7 < inh; q += 8)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    float32x4_t _r1;
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1q_f32(r0);
-                        _r1 = vld1q_f32(r0 + N);
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float32x4_t();
-                        _r1 = float32x4_t();
-                        _r0 = vsetq_lane_f32(r0[0], _r0, 0);
-                        _r0 = vsetq_lane_f32(r0[N], _r0, 1);
-                        _r0 = vsetq_lane_f32(r0[N * 2], _r0, 2);
-                        _r0 = vsetq_lane_f32(r0[N * 3], _r0, 3);
-                        _r1 = vsetq_lane_f32(r0[N * 4], _r1, 0);
-                        _r1 = vsetq_lane_f32(r0[N * 5], _r1, 1);
-                        _r1 = vsetq_lane_f32(r0[N * 6], _r1, 2);
-                        _r1 = vsetq_lane_f32(r0[N * 7], _r1, 3);
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w0 = vld1q_f32(kptr);
-                    float32x4_t _w1 = vld1q_f32(kptr + 4);
-                    _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                    _sum1 = vfmaq_f32(_sum1, _r1, _w1);
-
-                    kptr += 8;
-                }
-            }
-            _sum0 = vaddq_f32(_sum0, _sum1);
-            sum += vaddvq_f32(_sum0);
-#endif // __aarch64__
-            float32x4_t _sum = vdupq_n_f32(0.f);
-            for (; q + 3 < inh; q += 4)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1q_f32(r0);
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float32x4_t();
-                        _r0 = vsetq_lane_f32(r0[0], _r0, 0);
-                        _r0 = vsetq_lane_f32(r0[N], _r0, 1);
-                        _r0 = vsetq_lane_f32(r0[N * 2], _r0, 2);
-                        _r0 = vsetq_lane_f32(r0[N * 3], _r0, 3);
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w = vld1q_f32(kptr);
-#if __aarch64__
-                    _sum = vfmaq_f32(_sum, _r0, _w);
-#else
-                    _sum = vmlaq_f32(_sum, _r0, _w);
-#endif
-
-                    kptr += 4;
-                }
-            }
-#if __aarch64__
-            sum += vaddvq_f32(_sum);
-#else
-            float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-            _ss = vpadd_f32(_ss, _ss);
-            sum += vget_lane_f32(_ss, 0);
-#endif
-#endif // __ARM_NEON
-            for (; q + 1 < inh; q += 2)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val0;
-                    float val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = r0[0];
-                        val1 = r0[N];
-                        r0 += dilation_w;
-                    }
-
-                    sum += val0 * kptr[0];
-                    sum += val1 * kptr[1];
-
-                    kptr += 2;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val;
-                    // if (elempack == 1)
-                    {
-                        val = r0[0];
-                        r0 += dilation_w;
-                    }
-
-                    sum += val * kptr[0];
-
-                    kptr += 1;
-                }
-            }
-
-            sum = activation_ss(sum, activation_type, activation_params);
-
-            outptr[0] = sum;
-            outptr += 1;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_packed_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_packed_bf16s.h
deleted file mode 100644
index 8f626b57..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_packed_bf16s.h
+++ /dev/null
@@ -1,1330 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution1d_transform_kernel_packed_bf16s(const Mat& kernel, Mat& kernel_tm, int inh, int outh, int kernel_w)
-{
-    // src = kw-inh-outh
-    // dst = pb-pa-kw-inh/pa-outh/pb
-
-    // clang-format off
-    // *INDENT-OFF*
-#if __ARM_NEON
-#if __aarch64__
-    if (outh >= 8)
-    {
-        if (inh >= 8)
-            kernel_tm.create(8 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-        else if (inh >= 4)
-            kernel_tm.create(8 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-        else if (inh >= 2)
-            kernel_tm.create(8 * 2 * kernel_w, inh / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-        else
-            kernel_tm.create(8 * kernel_w, inh, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-    }
-    else
-#endif // __aarch64__
-    if (outh >= 4)
-    {
-#if __aarch64__
-        if (inh >= 8)
-            kernel_tm.create(4 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-        else
-#endif // __aarch64__
-        if (inh >= 4)
-            kernel_tm.create(4 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-        else if (inh >= 2)
-            kernel_tm.create(4 * 2 * kernel_w, inh / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-        else
-            kernel_tm.create(4 * kernel_w, inh, outh / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-    }
-    else
-#endif // __ARM_NEON
-    if (outh >= 2)
-    {
-#if __ARM_NEON
-#if __aarch64__
-        if (inh >= 8)
-            kernel_tm.create(2 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 2 + outh % 2, (size_t)2u);
-        else
-#endif // __aarch64__
-        if (inh >= 4)
-            kernel_tm.create(2 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 2 + outh % 2, (size_t)2u);
-        else
-#endif // __ARM_NEON
-        if (inh >= 2)
-            kernel_tm.create(2 * 2 * kernel_w, inh / 2 + inh % 2, outh / 2 + outh % 2, (size_t)2u);
-        else
-            kernel_tm.create(2 * kernel_w, inh, outh / 2 + outh % 2, (size_t)2u);
-    }
-    else
-    {
-#if __ARM_NEON
-#if __aarch64__
-        if (inh >= 8)
-            kernel_tm.create(8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh, (size_t)2u);
-        else
-#endif // __aarch64__
-        if (inh >= 4)
-            kernel_tm.create(4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh, (size_t)2u);
-        else
-#endif // __ARM_NEON
-        if (inh >= 2)
-            kernel_tm.create(2 * kernel_w, inh / 2 + inh % 2, outh, (size_t)2u);
-        else
-            kernel_tm.create(kernel_w, inh, outh, (size_t)2u);
-    }
-    // *INDENT-ON*
-    // clang-format on
-
-    int q = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; q + 7 < outh; q += 8)
-    {
-        const float* kptr0 = (const float*)kernel + q * inh * kernel_w;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inh * kernel_w;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inh * kernel_w;
-        const float* kptr4 = (const float*)kernel + (q + 4) * inh * kernel_w;
-        const float* kptr5 = (const float*)kernel + (q + 5) * inh * kernel_w;
-        const float* kptr6 = (const float*)kernel + (q + 6) * inh * kernel_w;
-        const float* kptr7 = (const float*)kernel + (q + 7) * inh * kernel_w;
-
-        unsigned short* g00 = kernel_tm.channel(q / 8);
-
-        int p = 0;
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-                const float* k4 = kptr4 + p * kernel_w;
-                const float* k5 = kptr5 + p * kernel_w;
-                const float* k6 = kptr6 + p * kernel_w;
-                const float* k7 = kptr7 + p * kernel_w;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    g00[1] = float32_to_bfloat16(k1[k]);
-                    g00[2] = float32_to_bfloat16(k2[k]);
-                    g00[3] = float32_to_bfloat16(k3[k]);
-                    g00[4] = float32_to_bfloat16(k4[k]);
-                    g00[5] = float32_to_bfloat16(k5[k]);
-                    g00[6] = float32_to_bfloat16(k6[k]);
-                    g00[7] = float32_to_bfloat16(k7[k]);
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    k4 += kernel_w;
-                    k5 += kernel_w;
-                    k6 += kernel_w;
-                    k7 += kernel_w;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-                const float* k4 = kptr4 + p * kernel_w;
-                const float* k5 = kptr5 + p * kernel_w;
-                const float* k6 = kptr6 + p * kernel_w;
-                const float* k7 = kptr7 + p * kernel_w;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    g00[1] = float32_to_bfloat16(k1[k]);
-                    g00[2] = float32_to_bfloat16(k2[k]);
-                    g00[3] = float32_to_bfloat16(k3[k]);
-                    g00[4] = float32_to_bfloat16(k4[k]);
-                    g00[5] = float32_to_bfloat16(k5[k]);
-                    g00[6] = float32_to_bfloat16(k6[k]);
-                    g00[7] = float32_to_bfloat16(k7[k]);
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    k4 += kernel_w;
-                    k5 += kernel_w;
-                    k6 += kernel_w;
-                    k7 += kernel_w;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-                const float* k4 = kptr4 + p * kernel_w;
-                const float* k5 = kptr5 + p * kernel_w;
-                const float* k6 = kptr6 + p * kernel_w;
-                const float* k7 = kptr7 + p * kernel_w;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    g00[1] = float32_to_bfloat16(k1[k]);
-                    g00[2] = float32_to_bfloat16(k2[k]);
-                    g00[3] = float32_to_bfloat16(k3[k]);
-                    g00[4] = float32_to_bfloat16(k4[k]);
-                    g00[5] = float32_to_bfloat16(k5[k]);
-                    g00[6] = float32_to_bfloat16(k6[k]);
-                    g00[7] = float32_to_bfloat16(k7[k]);
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    k4 += kernel_w;
-                    k5 += kernel_w;
-                    k6 += kernel_w;
-                    k7 += kernel_w;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p < inh; p++)
-        {
-            const float* k0 = kptr0 + p * kernel_w;
-            const float* k1 = kptr1 + p * kernel_w;
-            const float* k2 = kptr2 + p * kernel_w;
-            const float* k3 = kptr3 + p * kernel_w;
-            const float* k4 = kptr4 + p * kernel_w;
-            const float* k5 = kptr5 + p * kernel_w;
-            const float* k6 = kptr6 + p * kernel_w;
-            const float* k7 = kptr7 + p * kernel_w;
-
-            for (int k = 0; k < kernel_w; k++)
-            {
-                g00[0] = float32_to_bfloat16(k0[k]);
-                g00[1] = float32_to_bfloat16(k1[k]);
-                g00[2] = float32_to_bfloat16(k2[k]);
-                g00[3] = float32_to_bfloat16(k3[k]);
-                g00[4] = float32_to_bfloat16(k4[k]);
-                g00[5] = float32_to_bfloat16(k5[k]);
-                g00[6] = float32_to_bfloat16(k6[k]);
-                g00[7] = float32_to_bfloat16(k7[k]);
-                g00 += 8;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; q + 3 < outh; q += 4)
-    {
-        const float* kptr0 = (const float*)kernel + q * inh * kernel_w;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inh * kernel_w;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inh * kernel_w;
-
-#if __aarch64__
-        unsigned short* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4);
-#else
-        unsigned short* g00 = kernel_tm.channel(q / 4);
-#endif
-
-        int p = 0;
-#if __aarch64__
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    g00[1] = float32_to_bfloat16(k1[k]);
-                    g00[2] = float32_to_bfloat16(k2[k]);
-                    g00[3] = float32_to_bfloat16(k3[k]);
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    g00 += 4;
-                }
-            }
-        }
-#endif // __aarch64__
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    g00[1] = float32_to_bfloat16(k1[k]);
-                    g00[2] = float32_to_bfloat16(k2[k]);
-                    g00[3] = float32_to_bfloat16(k3[k]);
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    g00 += 4;
-                }
-            }
-        }
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    g00[1] = float32_to_bfloat16(k1[k]);
-                    g00[2] = float32_to_bfloat16(k2[k]);
-                    g00[3] = float32_to_bfloat16(k3[k]);
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    g00 += 4;
-                }
-            }
-        }
-        for (; p < inh; p++)
-        {
-            const float* k0 = kptr0 + p * kernel_w;
-            const float* k1 = kptr1 + p * kernel_w;
-            const float* k2 = kptr2 + p * kernel_w;
-            const float* k3 = kptr3 + p * kernel_w;
-
-            for (int k = 0; k < kernel_w; k++)
-            {
-                g00[0] = float32_to_bfloat16(k0[k]);
-                g00[1] = float32_to_bfloat16(k1[k]);
-                g00[2] = float32_to_bfloat16(k2[k]);
-                g00[3] = float32_to_bfloat16(k3[k]);
-                g00 += 4;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; q + 1 < outh; q += 2)
-    {
-        const float* kptr0 = (const float*)kernel + q * inh * kernel_w;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w;
-
-#if __aarch64__
-        unsigned short* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2);
-#elif __ARM_NEON
-        unsigned short* g00 = kernel_tm.channel(q / 4 + (q % 4) / 2);
-#else
-        unsigned short* g00 = kernel_tm.channel(q / 2);
-#endif
-
-        int p = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w + k;
-                const float* k1 = kptr1 + p * kernel_w + k;
-
-                g00[0] = float32_to_bfloat16(k0[0]);
-                g00[1] = float32_to_bfloat16(k0[kernel_w]);
-                g00[2] = float32_to_bfloat16(k0[kernel_w * 2]);
-                g00[3] = float32_to_bfloat16(k0[kernel_w * 3]);
-                g00[4] = float32_to_bfloat16(k0[kernel_w * 4]);
-                g00[5] = float32_to_bfloat16(k0[kernel_w * 5]);
-                g00[6] = float32_to_bfloat16(k0[kernel_w * 6]);
-                g00[7] = float32_to_bfloat16(k0[kernel_w * 7]);
-                g00[8] = float32_to_bfloat16(k1[0]);
-                g00[9] = float32_to_bfloat16(k1[kernel_w]);
-                g00[10] = float32_to_bfloat16(k1[kernel_w * 2]);
-                g00[11] = float32_to_bfloat16(k1[kernel_w * 3]);
-                g00[12] = float32_to_bfloat16(k1[kernel_w * 4]);
-                g00[13] = float32_to_bfloat16(k1[kernel_w * 5]);
-                g00[14] = float32_to_bfloat16(k1[kernel_w * 6]);
-                g00[15] = float32_to_bfloat16(k1[kernel_w * 7]);
-                g00 += 16;
-            }
-        }
-#endif // __aarch64__
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w + k;
-                const float* k1 = kptr1 + p * kernel_w + k;
-
-                g00[0] = float32_to_bfloat16(k0[0]);
-                g00[1] = float32_to_bfloat16(k0[kernel_w]);
-                g00[2] = float32_to_bfloat16(k0[kernel_w * 2]);
-                g00[3] = float32_to_bfloat16(k0[kernel_w * 3]);
-                g00[4] = float32_to_bfloat16(k1[0]);
-                g00[5] = float32_to_bfloat16(k1[kernel_w]);
-                g00[6] = float32_to_bfloat16(k1[kernel_w * 2]);
-                g00[7] = float32_to_bfloat16(k1[kernel_w * 3]);
-                g00 += 8;
-            }
-        }
-#endif // __ARM_NEON
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    g00[1] = float32_to_bfloat16(k1[k]);
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    g00 += 2;
-                }
-            }
-        }
-        for (; p < inh; p++)
-        {
-            const float* k0 = kptr0 + p * kernel_w;
-            const float* k1 = kptr1 + p * kernel_w;
-
-            for (int k = 0; k < kernel_w; k++)
-            {
-                g00[0] = float32_to_bfloat16(k0[k]);
-                g00[1] = float32_to_bfloat16(k1[k]);
-                g00 += 2;
-            }
-        }
-    }
-    for (; q < outh; q++)
-    {
-        const float* kptr = (const float*)kernel + q * inh * kernel_w;
-
-#if __aarch64__
-        unsigned short* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2 + q % 2);
-#elif __ARM_NEON
-        unsigned short* g00 = kernel_tm.channel(q / 4 + (q % 4) / 2 + q % 2);
-#else
-        unsigned short* g00 = kernel_tm.channel(q / 2 + q % 2);
-#endif
-
-        int p = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr + p * kernel_w;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    k0 += kernel_w;
-                    g00 += 1;
-                }
-            }
-        }
-#endif // __aarch64__
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr + p * kernel_w;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    k0 += kernel_w;
-                    g00 += 1;
-                }
-            }
-        }
-#endif // __ARM_NEON
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr + p * kernel_w;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    k0 += kernel_w;
-                    g00 += 1;
-                }
-            }
-        }
-        for (; p < inh; p++)
-        {
-            const float* k0 = kptr + p * kernel_w;
-
-            for (int k = 0; k < kernel_w; k++)
-            {
-                g00[0] = float32_to_bfloat16(k0[k]);
-                g00++;
-            }
-        }
-    }
-}
-
-static void convolution1d_packed_bf16s(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int kernel_w, int dilation_w, int stride_w, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int elempack = bottom_blob.elempack;
-    const int inh = bottom_blob.h * elempack;
-
-    const int N = bottom_blob.w * elempack;
-
-    const int outw = top_blob.w;
-    const int out_elempack = top_blob.elempack;
-    const int outh = top_blob.h * out_elempack;
-
-    const int M = top_blob.w * out_elempack;
-
-    const float* bias_data_ptr = bias_data;
-
-    int nn_outh = 0;
-    int remain_outh_start = 0;
-#if __ARM_NEON
-#if __aarch64__
-    nn_outh = (outh - remain_outh_start) / 8;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = remain_outh_start + pp * 8;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-        const int out_elempack = top_blob.elempack;
-
-        unsigned short* outptr = top_blob.row<unsigned short>(p / out_elempack);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
-            float32x4_t _sum1 = vdupq_n_f32(0.f);
-            float32x4_t _sum2 = vdupq_n_f32(0.f);
-            float32x4_t _sum3 = vdupq_n_f32(0.f);
-            float32x4_t _sum4 = vdupq_n_f32(0.f);
-            float32x4_t _sum5 = vdupq_n_f32(0.f);
-            float32x4_t _sum6 = vdupq_n_f32(0.f);
-            float32x4_t _sum7 = vdupq_n_f32(0.f);
-
-            if (bias_data_ptr)
-            {
-                _sum0 = vld1q_f32(bias_data_ptr + p);
-                _sum1 = vld1q_f32(bias_data_ptr + p + 4);
-            }
-
-            const unsigned short* kptr = weight_data_tm.channel(p / 8);
-
-            int q = 0;
-            for (; q + 7 < inh; q += 8)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    float32x4_t _r1;
-                    if (elempack == 4)
-                    {
-                        _r0 = bfloat2float(vld1_u16(r0));
-                        _r1 = bfloat2float(vld1_u16(r0 + N));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        uint16x8_t _r_u16 = uint16x8_t();
-                        _r_u16 = vsetq_lane_u16(r0[0], _r_u16, 0);
-                        _r_u16 = vsetq_lane_u16(r0[N], _r_u16, 1);
-                        _r_u16 = vsetq_lane_u16(r0[N * 2], _r_u16, 2);
-                        _r_u16 = vsetq_lane_u16(r0[N * 3], _r_u16, 3);
-                        _r_u16 = vsetq_lane_u16(r0[N * 4], _r_u16, 4);
-                        _r_u16 = vsetq_lane_u16(r0[N * 5], _r_u16, 5);
-                        _r_u16 = vsetq_lane_u16(r0[N * 6], _r_u16, 6);
-                        _r_u16 = vsetq_lane_u16(r0[N * 7], _r_u16, 7);
-                        _r0 = bfloat2float(vget_low_u16(_r_u16));
-                        _r1 = bfloat2float(vget_high_u16(_r_u16));
-                        r0 += dilation_w;
-                    }
-
-                    uint16x8_t _w01 = vld1q_u16(kptr);
-                    uint16x8_t _w23 = vld1q_u16(kptr + 8);
-                    uint16x8_t _w45 = vld1q_u16(kptr + 16);
-                    uint16x8_t _w67 = vld1q_u16(kptr + 24);
-                    uint16x8_t _w89 = vld1q_u16(kptr + 32);
-                    uint16x8_t _wab = vld1q_u16(kptr + 40);
-                    uint16x8_t _wcd = vld1q_u16(kptr + 48);
-                    uint16x8_t _wef = vld1q_u16(kptr + 56);
-                    float32x4_t _w0 = bfloat2float(vget_low_u16(_w01));
-                    float32x4_t _w1 = bfloat2float(vget_high_u16(_w01));
-                    float32x4_t _w2 = bfloat2float(vget_low_u16(_w23));
-                    float32x4_t _w3 = bfloat2float(vget_high_u16(_w23));
-                    float32x4_t _w4 = bfloat2float(vget_low_u16(_w45));
-                    float32x4_t _w5 = bfloat2float(vget_high_u16(_w45));
-                    float32x4_t _w6 = bfloat2float(vget_low_u16(_w67));
-                    float32x4_t _w7 = bfloat2float(vget_high_u16(_w67));
-                    float32x4_t _w8 = bfloat2float(vget_low_u16(_w89));
-                    float32x4_t _w9 = bfloat2float(vget_high_u16(_w89));
-                    float32x4_t _wa = bfloat2float(vget_low_u16(_wab));
-                    float32x4_t _wb = bfloat2float(vget_high_u16(_wab));
-                    float32x4_t _wc = bfloat2float(vget_low_u16(_wcd));
-                    float32x4_t _wd = bfloat2float(vget_high_u16(_wcd));
-                    float32x4_t _we = bfloat2float(vget_low_u16(_wef));
-                    float32x4_t _wf = bfloat2float(vget_high_u16(_wef));
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 0);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 1);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 1);
-                    _sum4 = vfmaq_laneq_f32(_sum4, _w4, _r0, 2);
-                    _sum5 = vfmaq_laneq_f32(_sum5, _w5, _r0, 2);
-                    _sum6 = vfmaq_laneq_f32(_sum6, _w6, _r0, 3);
-                    _sum7 = vfmaq_laneq_f32(_sum7, _w7, _r0, 3);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w8, _r1, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w9, _r1, 0);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _wa, _r1, 1);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _wb, _r1, 1);
-                    _sum4 = vfmaq_laneq_f32(_sum4, _wc, _r1, 2);
-                    _sum5 = vfmaq_laneq_f32(_sum5, _wd, _r1, 2);
-                    _sum6 = vfmaq_laneq_f32(_sum6, _we, _r1, 3);
-                    _sum7 = vfmaq_laneq_f32(_sum7, _wf, _r1, 3);
-
-                    kptr += 64;
-                }
-            }
-            for (; q + 3 < inh; q += 4)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = bfloat2float(vld1_u16(r0));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        uint16x4_t _r_u16 = uint16x4_t();
-                        _r_u16 = vset_lane_u16(r0[0], _r_u16, 0);
-                        _r_u16 = vset_lane_u16(r0[N], _r_u16, 1);
-                        _r_u16 = vset_lane_u16(r0[N * 2], _r_u16, 2);
-                        _r_u16 = vset_lane_u16(r0[N * 3], _r_u16, 3);
-                        _r0 = bfloat2float(_r_u16);
-                        r0 += dilation_w;
-                    }
-
-                    uint16x8_t _w01 = vld1q_u16(kptr);
-                    uint16x8_t _w23 = vld1q_u16(kptr + 8);
-                    uint16x8_t _w45 = vld1q_u16(kptr + 16);
-                    uint16x8_t _w67 = vld1q_u16(kptr + 24);
-                    float32x4_t _w0 = bfloat2float(vget_low_u16(_w01));
-                    float32x4_t _w1 = bfloat2float(vget_high_u16(_w01));
-                    float32x4_t _w2 = bfloat2float(vget_low_u16(_w23));
-                    float32x4_t _w3 = bfloat2float(vget_high_u16(_w23));
-                    float32x4_t _w4 = bfloat2float(vget_low_u16(_w45));
-                    float32x4_t _w5 = bfloat2float(vget_high_u16(_w45));
-                    float32x4_t _w6 = bfloat2float(vget_low_u16(_w67));
-                    float32x4_t _w7 = bfloat2float(vget_high_u16(_w67));
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 0);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 1);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 1);
-                    _sum4 = vfmaq_laneq_f32(_sum4, _w4, _r0, 2);
-                    _sum5 = vfmaq_laneq_f32(_sum5, _w5, _r0, 2);
-                    _sum6 = vfmaq_laneq_f32(_sum6, _w6, _r0, 3);
-                    _sum7 = vfmaq_laneq_f32(_sum7, _w7, _r0, 3);
-
-                    kptr += 32;
-                }
-            }
-            for (; q + 1 < inh; q += 2)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val0;
-                    float val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = bfloat16_to_float32(r0[0]);
-                        val1 = bfloat16_to_float32(r0[N]);
-                        r0 += dilation_w;
-                    }
-
-                    uint16x8_t _w01 = vld1q_u16(kptr);
-                    uint16x8_t _w23 = vld1q_u16(kptr + 8);
-                    float32x4_t _w0 = bfloat2float(vget_low_u16(_w01));
-                    float32x4_t _w1 = bfloat2float(vget_high_u16(_w01));
-                    float32x4_t _w2 = bfloat2float(vget_low_u16(_w23));
-                    float32x4_t _w3 = bfloat2float(vget_high_u16(_w23));
-                    _sum0 = vfmaq_n_f32(_sum0, _w0, val0);
-                    _sum1 = vfmaq_n_f32(_sum1, _w1, val0);
-                    _sum2 = vfmaq_n_f32(_sum2, _w2, val1);
-                    _sum3 = vfmaq_n_f32(_sum3, _w3, val1);
-
-                    kptr += 16;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _val;
-                    // if (elempack == 1)
-                    {
-                        _val = bfloat2float(vdup_n_u16(r0[0]));
-                        r0 += dilation_w;
-                    }
-
-                    uint16x8_t _w = vld1q_u16(kptr);
-                    float32x4_t _w0 = bfloat2float(vget_low_u16(_w));
-                    float32x4_t _w1 = bfloat2float(vget_high_u16(_w));
-                    _sum0 = vfmaq_f32(_sum0, _w0, _val);
-                    _sum1 = vfmaq_f32(_sum1, _w1, _val);
-
-                    kptr += 8;
-                }
-            }
-
-            _sum0 = vaddq_f32(_sum0, _sum2);
-            _sum1 = vaddq_f32(_sum1, _sum3);
-            _sum4 = vaddq_f32(_sum4, _sum6);
-            _sum5 = vaddq_f32(_sum5, _sum7);
-            _sum0 = vaddq_f32(_sum0, _sum4);
-            _sum1 = vaddq_f32(_sum1, _sum5);
-
-            _sum0 = activation_ps(_sum0, activation_type, activation_params);
-            _sum1 = activation_ps(_sum1, activation_type, activation_params);
-
-            if (out_elempack == 4)
-            {
-                vst1_u16(outptr, float2bfloat(_sum0));
-                vst1_u16(outptr + M, float2bfloat(_sum1));
-                outptr += 4;
-            }
-            else // if (out_elempack == 1)
-            {
-                uint16x4_t _sum0_u16 = float2bfloat(_sum0);
-                uint16x4_t _sum1_u16 = float2bfloat(_sum1);
-                outptr[0] = vget_lane_u16(_sum0_u16, 0);
-                outptr[M] = vget_lane_u16(_sum0_u16, 1);
-                outptr[M * 2] = vget_lane_u16(_sum0_u16, 2);
-                outptr[M * 3] = vget_lane_u16(_sum0_u16, 3);
-                outptr[M * 4] = vget_lane_u16(_sum1_u16, 0);
-                outptr[M * 5] = vget_lane_u16(_sum1_u16, 1);
-                outptr[M * 6] = vget_lane_u16(_sum1_u16, 2);
-                outptr[M * 7] = vget_lane_u16(_sum1_u16, 3);
-                outptr += 1;
-            }
-        }
-    }
-    remain_outh_start += nn_outh * 8;
-    nn_outh = (outh - remain_outh_start) / 4;
-#else // __aarch64__
-    nn_outh = (outh - remain_outh_start) / 4;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __aarch64__
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = remain_outh_start + pp * 4;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-        const int out_elempack = top_blob.elempack;
-
-        unsigned short* outptr = top_blob.row<unsigned short>(p / out_elempack);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
-            float32x4_t _sum1 = vdupq_n_f32(0.f);
-            float32x4_t _sum2 = vdupq_n_f32(0.f);
-            float32x4_t _sum3 = vdupq_n_f32(0.f);
-
-            if (bias_data_ptr)
-            {
-                _sum0 = vld1q_f32(bias_data_ptr + p);
-            }
-
-#if __aarch64__
-            const unsigned short* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4);
-#else
-            const unsigned short* kptr = weight_data_tm.channel(p / 4);
-#endif
-
-            int q = 0;
-#if __aarch64__
-            for (; q + 7 < inh; q += 8)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    float32x4_t _r1;
-                    if (elempack == 4)
-                    {
-                        _r0 = bfloat2float(vld1_u16(r0));
-                        _r1 = bfloat2float(vld1_u16(r0 + N));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        uint16x8_t _r_u16 = uint16x8_t();
-                        _r_u16 = vsetq_lane_u16(r0[0], _r_u16, 0);
-                        _r_u16 = vsetq_lane_u16(r0[N], _r_u16, 1);
-                        _r_u16 = vsetq_lane_u16(r0[N * 2], _r_u16, 2);
-                        _r_u16 = vsetq_lane_u16(r0[N * 3], _r_u16, 3);
-                        _r_u16 = vsetq_lane_u16(r0[N * 4], _r_u16, 4);
-                        _r_u16 = vsetq_lane_u16(r0[N * 5], _r_u16, 5);
-                        _r_u16 = vsetq_lane_u16(r0[N * 6], _r_u16, 6);
-                        _r_u16 = vsetq_lane_u16(r0[N * 7], _r_u16, 7);
-                        _r0 = bfloat2float(vget_low_u16(_r_u16));
-                        _r1 = bfloat2float(vget_high_u16(_r_u16));
-                        r0 += dilation_w;
-                    }
-
-                    uint16x8_t _w01 = vld1q_u16(kptr);
-                    uint16x8_t _w23 = vld1q_u16(kptr + 8);
-                    uint16x8_t _w45 = vld1q_u16(kptr + 16);
-                    uint16x8_t _w67 = vld1q_u16(kptr + 24);
-                    float32x4_t _w0 = bfloat2float(vget_low_u16(_w01));
-                    float32x4_t _w1 = bfloat2float(vget_high_u16(_w01));
-                    float32x4_t _w2 = bfloat2float(vget_low_u16(_w23));
-                    float32x4_t _w3 = bfloat2float(vget_high_u16(_w23));
-                    float32x4_t _w4 = bfloat2float(vget_low_u16(_w45));
-                    float32x4_t _w5 = bfloat2float(vget_high_u16(_w45));
-                    float32x4_t _w6 = bfloat2float(vget_low_u16(_w67));
-                    float32x4_t _w7 = bfloat2float(vget_high_u16(_w67));
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 1);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 2);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 3);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w4, _r1, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w5, _r1, 1);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w6, _r1, 2);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w7, _r1, 3);
-
-                    kptr += 32;
-                }
-            }
-#endif // __aarch64__
-            for (; q + 3 < inh; q += 4)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = bfloat2float(vld1_u16(r0));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        uint16x4_t _r_u16 = uint16x4_t();
-                        _r_u16 = vset_lane_u16(r0[0], _r_u16, 0);
-                        _r_u16 = vset_lane_u16(r0[N], _r_u16, 1);
-                        _r_u16 = vset_lane_u16(r0[N * 2], _r_u16, 2);
-                        _r_u16 = vset_lane_u16(r0[N * 3], _r_u16, 3);
-                        _r0 = bfloat2float(_r_u16);
-                        r0 += dilation_w;
-                    }
-
-                    uint16x8_t _w01 = vld1q_u16(kptr);
-                    uint16x8_t _w23 = vld1q_u16(kptr + 8);
-                    float32x4_t _w0 = bfloat2float(vget_low_u16(_w01));
-                    float32x4_t _w1 = bfloat2float(vget_high_u16(_w01));
-                    float32x4_t _w2 = bfloat2float(vget_low_u16(_w23));
-                    float32x4_t _w3 = bfloat2float(vget_high_u16(_w23));
-#if __aarch64__
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 1);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 2);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 3);
-#else
-                    _sum0 = vmlaq_lane_f32(_sum0, _w0, vget_low_f32(_r0), 0);
-                    _sum1 = vmlaq_lane_f32(_sum1, _w1, vget_low_f32(_r0), 1);
-                    _sum2 = vmlaq_lane_f32(_sum2, _w2, vget_high_f32(_r0), 0);
-                    _sum3 = vmlaq_lane_f32(_sum3, _w3, vget_high_f32(_r0), 1);
-#endif
-
-                    kptr += 16;
-                }
-            }
-            for (; q + 1 < inh; q += 2)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val0;
-                    float val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = bfloat16_to_float32(r0[0]);
-                        val1 = bfloat16_to_float32(r0[N]);
-                        r0 += dilation_w;
-                    }
-
-                    uint16x8_t _w = vld1q_u16(kptr);
-                    float32x4_t _w0 = bfloat2float(vget_low_u16(_w));
-                    float32x4_t _w1 = bfloat2float(vget_high_u16(_w));
-#if __aarch64__
-                    _sum0 = vfmaq_n_f32(_sum0, _w0, val0);
-                    _sum1 = vfmaq_n_f32(_sum1, _w1, val1);
-#else
-                    _sum0 = vmlaq_n_f32(_sum0, _w0, val0);
-                    _sum1 = vmlaq_n_f32(_sum1, _w1, val1);
-#endif
-
-                    kptr += 8;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _val;
-                    // if (elempack == 1)
-                    {
-                        _val = bfloat2float(vdup_n_u16(r0[0]));
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w = bfloat2float(vld1_u16(kptr));
-#if __aarch64__
-                    _sum0 = vfmaq_f32(_sum0, _val, _w);
-#else
-                    _sum0 = vmlaq_f32(_sum0, _val, _w);
-#endif
-
-                    kptr += 4;
-                }
-            }
-
-            _sum0 = vaddq_f32(_sum0, _sum1);
-            _sum2 = vaddq_f32(_sum2, _sum3);
-            _sum0 = vaddq_f32(_sum0, _sum2);
-
-            _sum0 = activation_ps(_sum0, activation_type, activation_params);
-
-            if (out_elempack == 4)
-            {
-                vst1_u16(outptr, float2bfloat(_sum0));
-                outptr += 4;
-            }
-            else // if (out_elempack == 1)
-            {
-                uint16x4_t _sum0_u16 = float2bfloat(_sum0);
-                outptr[0] = vget_lane_u16(_sum0_u16, 0);
-                outptr[M] = vget_lane_u16(_sum0_u16, 1);
-                outptr[M * 2] = vget_lane_u16(_sum0_u16, 2);
-                outptr[M * 3] = vget_lane_u16(_sum0_u16, 3);
-                outptr += 1;
-            }
-        }
-    }
-    remain_outh_start += nn_outh * 4;
-    nn_outh = (outh - remain_outh_start) / 2;
-#else // __ARM_NEON
-    nn_outh = (outh - remain_outh_start) / 2;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __ARM_NEON
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = remain_outh_start + pp * 2;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-
-        unsigned short* outptr0 = top_blob.row<unsigned short>(p);
-        unsigned short* outptr1 = top_blob.row<unsigned short>(p + 1);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float sum0 = 0.f;
-            float sum1 = 0.f;
-
-            if (bias_data_ptr)
-            {
-                sum0 = bias_data_ptr[p];
-                sum1 = bias_data_ptr[p + 1];
-            }
-
-#if __aarch64__
-            const unsigned short* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2);
-#elif __ARM_NEON
-            const unsigned short* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2);
-#else
-            const unsigned short* kptr = weight_data_tm.channel(p / 2);
-#endif
-
-            int q = 0;
-#if __ARM_NEON
-#if __aarch64__
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
-            float32x4_t _sum1 = vdupq_n_f32(0.f);
-            float32x4_t _sum2 = vdupq_n_f32(0.f);
-            float32x4_t _sum3 = vdupq_n_f32(0.f);
-            for (; q + 7 < inh; q += 8)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    float32x4_t _r1;
-                    if (elempack == 4)
-                    {
-                        _r0 = bfloat2float(vld1_u16(r0));
-                        _r1 = bfloat2float(vld1_u16(r0 + N));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        uint16x8_t _r01_u16 = uint16x8_t();
-                        _r01_u16 = vsetq_lane_u16(r0[0], _r01_u16, 0);
-                        _r01_u16 = vsetq_lane_u16(r0[N], _r01_u16, 1);
-                        _r01_u16 = vsetq_lane_u16(r0[N * 2], _r01_u16, 2);
-                        _r01_u16 = vsetq_lane_u16(r0[N * 3], _r01_u16, 3);
-                        _r01_u16 = vsetq_lane_u16(r0[N * 4], _r01_u16, 4);
-                        _r01_u16 = vsetq_lane_u16(r0[N * 5], _r01_u16, 5);
-                        _r01_u16 = vsetq_lane_u16(r0[N * 6], _r01_u16, 6);
-                        _r01_u16 = vsetq_lane_u16(r0[N * 7], _r01_u16, 7);
-                        _r0 = bfloat2float(vget_low_u16(_r01_u16));
-                        _r1 = bfloat2float(vget_high_u16(_r01_u16));
-                        r0 += dilation_w;
-                    }
-
-                    uint16x8_t _w01 = vld1q_u16(kptr);
-                    uint16x8_t _w23 = vld1q_u16(kptr + 8);
-                    float32x4_t _w0 = bfloat2float(vget_low_u16(_w01));
-                    float32x4_t _w1 = bfloat2float(vget_high_u16(_w01));
-                    float32x4_t _w2 = bfloat2float(vget_low_u16(_w23));
-                    float32x4_t _w3 = bfloat2float(vget_high_u16(_w23));
-                    _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                    _sum1 = vfmaq_f32(_sum1, _r1, _w1);
-                    _sum2 = vfmaq_f32(_sum2, _r0, _w2);
-                    _sum3 = vfmaq_f32(_sum3, _r1, _w3);
-
-                    kptr += 16;
-                }
-            }
-            _sum0 = vaddq_f32(_sum0, _sum1);
-            _sum2 = vaddq_f32(_sum2, _sum3);
-            sum0 += vaddvq_f32(_sum0);
-            sum1 += vaddvq_f32(_sum2);
-            _sum0 = vdupq_n_f32(0.f);
-            _sum1 = vdupq_n_f32(0.f);
-#else  // __aarch64__
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
-            float32x4_t _sum1 = vdupq_n_f32(0.f);
-#endif // __aarch64__
-            for (; q + 3 < inh; q += 4)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = bfloat2float(vld1_u16(r0));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        uint16x4_t _r0_u16 = uint16x4_t();
-                        _r0_u16 = vset_lane_u16(r0[0], _r0_u16, 0);
-                        _r0_u16 = vset_lane_u16(r0[N], _r0_u16, 1);
-                        _r0_u16 = vset_lane_u16(r0[N * 2], _r0_u16, 2);
-                        _r0_u16 = vset_lane_u16(r0[N * 3], _r0_u16, 3);
-                        _r0 = bfloat2float(_r0_u16);
-                        r0 += dilation_w;
-                    }
-
-                    uint16x8_t _w = vld1q_u16(kptr);
-                    float32x4_t _w0 = bfloat2float(vget_low_u16(_w));
-                    float32x4_t _w1 = bfloat2float(vget_high_u16(_w));
-#if __aarch64__
-                    _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                    _sum1 = vfmaq_f32(_sum1, _r0, _w1);
-#else
-                    _sum0 = vmlaq_f32(_sum0, _r0, _w0);
-                    _sum1 = vmlaq_f32(_sum1, _r0, _w1);
-#endif
-
-                    kptr += 8;
-                }
-            }
-#if __aarch64__
-            sum0 += vaddvq_f32(_sum0);
-            sum1 += vaddvq_f32(_sum1);
-#else
-            float32x2_t _ss0 = vadd_f32(vget_low_f32(_sum0), vget_high_f32(_sum0));
-            float32x2_t _ss1 = vadd_f32(vget_low_f32(_sum1), vget_high_f32(_sum1));
-            float32x2_t _ss = vpadd_f32(_ss0, _ss1);
-            sum0 += vget_lane_f32(_ss, 0);
-            sum1 += vget_lane_f32(_ss, 1);
-#endif
-#endif // __ARM_NEON
-            for (; q + 1 < inh; q += 2)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val0;
-                    float val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = bfloat16_to_float32(r0[0]);
-                        val1 = bfloat16_to_float32(r0[N]);
-                        r0 += dilation_w;
-                    }
-
-                    sum0 += val0 * bfloat16_to_float32(kptr[0]);
-                    sum1 += val0 * bfloat16_to_float32(kptr[1]);
-                    sum0 += val1 * bfloat16_to_float32(kptr[2]);
-                    sum1 += val1 * bfloat16_to_float32(kptr[3]);
-
-                    kptr += 4;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val;
-                    // if (elempack == 1)
-                    {
-                        val = bfloat16_to_float32(r0[0]);
-                        r0 += dilation_w;
-                    }
-
-                    sum0 += val * bfloat16_to_float32(kptr[0]);
-                    sum1 += val * bfloat16_to_float32(kptr[1]);
-
-                    kptr += 2;
-                }
-            }
-
-            sum0 = activation_ss(sum0, activation_type, activation_params);
-            sum1 = activation_ss(sum1, activation_type, activation_params);
-
-            outptr0[0] = float32_to_bfloat16(sum0);
-            outptr1[0] = float32_to_bfloat16(sum1);
-            outptr0 += 1;
-            outptr1 += 1;
-        }
-    }
-    remain_outh_start += nn_outh * 2;
-    for (int p = remain_outh_start; p < outh; p++)
-    {
-        unsigned short* outptr = top_blob.row<unsigned short>(p);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float sum = 0.f;
-
-            if (bias_data_ptr)
-            {
-                sum = bias_data_ptr[p];
-            }
-
-#if __aarch64__
-            const unsigned short* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#elif __ARM_NEON
-            const unsigned short* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2 + p % 2);
-#else
-            const unsigned short* kptr = weight_data_tm.channel(p / 2 + p % 2);
-#endif
-
-            int q = 0;
-#if __ARM_NEON
-#if __aarch64__
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
-            float32x4_t _sum1 = vdupq_n_f32(0.f);
-            for (; q + 7 < inh; q += 8)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    float32x4_t _r1;
-                    if (elempack == 4)
-                    {
-                        _r0 = bfloat2float(vld1_u16(r0));
-                        _r1 = bfloat2float(vld1_u16(r0 + N));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        uint16x8_t _r01_u16 = uint16x8_t();
-                        _r01_u16 = vsetq_lane_u16(r0[0], _r01_u16, 0);
-                        _r01_u16 = vsetq_lane_u16(r0[N], _r01_u16, 1);
-                        _r01_u16 = vsetq_lane_u16(r0[N * 2], _r01_u16, 2);
-                        _r01_u16 = vsetq_lane_u16(r0[N * 3], _r01_u16, 3);
-                        _r01_u16 = vsetq_lane_u16(r0[N * 4], _r01_u16, 4);
-                        _r01_u16 = vsetq_lane_u16(r0[N * 5], _r01_u16, 5);
-                        _r01_u16 = vsetq_lane_u16(r0[N * 6], _r01_u16, 6);
-                        _r01_u16 = vsetq_lane_u16(r0[N * 7], _r01_u16, 7);
-                        _r0 = bfloat2float(vget_low_u16(_r01_u16));
-                        _r1 = bfloat2float(vget_high_u16(_r01_u16));
-                        r0 += dilation_w;
-                    }
-
-                    uint16x8_t _w = vld1q_u16(kptr);
-                    float32x4_t _w0 = bfloat2float(vget_low_u16(_w));
-                    float32x4_t _w1 = bfloat2float(vget_high_u16(_w));
-                    _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                    _sum1 = vfmaq_f32(_sum1, _r1, _w1);
-
-                    kptr += 8;
-                }
-            }
-            _sum0 = vaddq_f32(_sum0, _sum1);
-            sum += vaddvq_f32(_sum0);
-#endif // __aarch64__
-            float32x4_t _sum = vdupq_n_f32(0.f);
-            for (; q + 3 < inh; q += 4)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = bfloat2float(vld1_u16(r0));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        uint16x4_t _r0_u16 = uint16x4_t();
-                        _r0_u16 = vset_lane_u16(r0[0], _r0_u16, 0);
-                        _r0_u16 = vset_lane_u16(r0[N], _r0_u16, 1);
-                        _r0_u16 = vset_lane_u16(r0[N * 2], _r0_u16, 2);
-                        _r0_u16 = vset_lane_u16(r0[N * 3], _r0_u16, 3);
-                        _r0 = bfloat2float(_r0_u16);
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w = bfloat2float(vld1_u16(kptr));
-#if __aarch64__
-                    _sum = vfmaq_f32(_sum, _r0, _w);
-#else
-                    _sum = vmlaq_f32(_sum, _r0, _w);
-#endif
-
-                    kptr += 4;
-                }
-            }
-#if __aarch64__
-            sum += vaddvq_f32(_sum);
-#else
-            float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-            _ss = vpadd_f32(_ss, _ss);
-            sum += vget_lane_f32(_ss, 0);
-#endif
-#endif // __ARM_NEON
-            for (; q + 1 < inh; q += 2)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val0;
-                    float val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = bfloat16_to_float32(r0[0]);
-                        val1 = bfloat16_to_float32(r0[N]);
-                        r0 += dilation_w;
-                    }
-
-                    sum += val0 * bfloat16_to_float32(kptr[0]);
-                    sum += val1 * bfloat16_to_float32(kptr[1]);
-
-                    kptr += 2;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const unsigned short* r0 = bottom_blob.row<const unsigned short>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val;
-                    // if (elempack == 1)
-                    {
-                        val = bfloat16_to_float32(r0[0]);
-                        r0 += dilation_w;
-                    }
-
-                    sum += val * bfloat16_to_float32(kptr[0]);
-
-                    kptr += 1;
-                }
-            }
-
-            sum = activation_ss(sum, activation_type, activation_params);
-
-            outptr[0] = float32_to_bfloat16(sum);
-            outptr += 1;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_packed_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_packed_fp16s.h
deleted file mode 100644
index 7b0c3e49..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution1d_packed_fp16s.h
+++ /dev/null
@@ -1,1883 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution1d_transform_kernel_packed_fp16s(const Mat& kernel, Mat& kernel_tm, int inh, int outh, int kernel_w)
-{
-    // src = kw-inh-outh
-    // dst = pb-pa-kw-inh/pa-outh/pb
-
-    // clang-format off
-    // *INDENT-OFF*
-    if (outh >= 8)
-    {
-        if (inh >= 8)
-            kernel_tm.create(8 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-        else if (inh >= 4)
-            kernel_tm.create(8 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-        else if (inh >= 2)
-            kernel_tm.create(8 * 2 * kernel_w, inh / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-        else
-            kernel_tm.create(8 * kernel_w, inh, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-    }
-    else if (outh >= 4)
-    {
-        if (inh >= 8)
-            kernel_tm.create(4 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-        else if (inh >= 4)
-            kernel_tm.create(4 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-        else if (inh >= 2)
-            kernel_tm.create(4 * 2 * kernel_w, inh / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-        else
-            kernel_tm.create(4 * kernel_w, inh, outh / 4 + (outh % 4) / 2 + outh % 2, (size_t)2u);
-    }
-    else if (outh >= 2)
-    {
-        if (inh >= 8)
-            kernel_tm.create(2 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 2 + outh % 2, (size_t)2u);
-        else if (inh >= 4)
-            kernel_tm.create(2 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 2 + outh % 2, (size_t)2u);
-        else if (inh >= 2)
-            kernel_tm.create(2 * 2 * kernel_w, inh / 2 + inh % 2, outh / 2 + outh % 2, (size_t)2u);
-        else
-            kernel_tm.create(2 * kernel_w, inh, outh / 2 + outh % 2, (size_t)2u);
-    }
-    else
-    {
-        if (inh >= 8)
-            kernel_tm.create(8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh, (size_t)2u);
-        else if (inh >= 4)
-            kernel_tm.create(4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh, (size_t)2u);
-        else if (inh >= 2)
-            kernel_tm.create(2 * kernel_w, inh / 2 + inh % 2, outh, (size_t)2u);
-        else
-            kernel_tm.create(kernel_w, inh, outh, (size_t)2u);
-    }
-    // *INDENT-ON*
-    // clang-format on
-
-    int q = 0;
-    for (; q + 7 < outh; q += 8)
-    {
-        const float* kptr0 = (const float*)kernel + q * inh * kernel_w;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inh * kernel_w;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inh * kernel_w;
-        const float* kptr4 = (const float*)kernel + (q + 4) * inh * kernel_w;
-        const float* kptr5 = (const float*)kernel + (q + 5) * inh * kernel_w;
-        const float* kptr6 = (const float*)kernel + (q + 6) * inh * kernel_w;
-        const float* kptr7 = (const float*)kernel + (q + 7) * inh * kernel_w;
-
-        __fp16* g00 = kernel_tm.channel(q / 8);
-
-        int p = 0;
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-                const float* k4 = kptr4 + p * kernel_w;
-                const float* k5 = kptr5 + p * kernel_w;
-                const float* k6 = kptr6 + p * kernel_w;
-                const float* k7 = kptr7 + p * kernel_w;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    g00[1] = (__fp16)k1[k];
-                    g00[2] = (__fp16)k2[k];
-                    g00[3] = (__fp16)k3[k];
-                    g00[4] = (__fp16)k4[k];
-                    g00[5] = (__fp16)k5[k];
-                    g00[6] = (__fp16)k6[k];
-                    g00[7] = (__fp16)k7[k];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    k4 += kernel_w;
-                    k5 += kernel_w;
-                    k6 += kernel_w;
-                    k7 += kernel_w;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-                const float* k4 = kptr4 + p * kernel_w;
-                const float* k5 = kptr5 + p * kernel_w;
-                const float* k6 = kptr6 + p * kernel_w;
-                const float* k7 = kptr7 + p * kernel_w;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    g00[1] = (__fp16)k1[k];
-                    g00[2] = (__fp16)k2[k];
-                    g00[3] = (__fp16)k3[k];
-                    g00[4] = (__fp16)k4[k];
-                    g00[5] = (__fp16)k5[k];
-                    g00[6] = (__fp16)k6[k];
-                    g00[7] = (__fp16)k7[k];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    k4 += kernel_w;
-                    k5 += kernel_w;
-                    k6 += kernel_w;
-                    k7 += kernel_w;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-                const float* k4 = kptr4 + p * kernel_w;
-                const float* k5 = kptr5 + p * kernel_w;
-                const float* k6 = kptr6 + p * kernel_w;
-                const float* k7 = kptr7 + p * kernel_w;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    g00[1] = (__fp16)k1[k];
-                    g00[2] = (__fp16)k2[k];
-                    g00[3] = (__fp16)k3[k];
-                    g00[4] = (__fp16)k4[k];
-                    g00[5] = (__fp16)k5[k];
-                    g00[6] = (__fp16)k6[k];
-                    g00[7] = (__fp16)k7[k];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    k4 += kernel_w;
-                    k5 += kernel_w;
-                    k6 += kernel_w;
-                    k7 += kernel_w;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p < inh; p++)
-        {
-            const float* k0 = kptr0 + p * kernel_w;
-            const float* k1 = kptr1 + p * kernel_w;
-            const float* k2 = kptr2 + p * kernel_w;
-            const float* k3 = kptr3 + p * kernel_w;
-            const float* k4 = kptr4 + p * kernel_w;
-            const float* k5 = kptr5 + p * kernel_w;
-            const float* k6 = kptr6 + p * kernel_w;
-            const float* k7 = kptr7 + p * kernel_w;
-
-            for (int k = 0; k < kernel_w; k++)
-            {
-                g00[0] = (__fp16)k0[k];
-                g00[1] = (__fp16)k1[k];
-                g00[2] = (__fp16)k2[k];
-                g00[3] = (__fp16)k3[k];
-                g00[4] = (__fp16)k4[k];
-                g00[5] = (__fp16)k5[k];
-                g00[6] = (__fp16)k6[k];
-                g00[7] = (__fp16)k7[k];
-                g00 += 8;
-            }
-        }
-    }
-    for (; q + 3 < outh; q += 4)
-    {
-        const float* kptr0 = (const float*)kernel + q * inh * kernel_w;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inh * kernel_w;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inh * kernel_w;
-
-        __fp16* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4);
-
-        int p = 0;
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    g00[1] = (__fp16)k1[k];
-                    g00[2] = (__fp16)k2[k];
-                    g00[3] = (__fp16)k3[k];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    g00 += 4;
-                }
-            }
-        }
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    g00[1] = (__fp16)k1[k];
-                    g00[2] = (__fp16)k2[k];
-                    g00[3] = (__fp16)k3[k];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    g00 += 4;
-                }
-            }
-        }
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-                const float* k2 = kptr2 + p * kernel_w;
-                const float* k3 = kptr3 + p * kernel_w;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    g00[1] = (__fp16)k1[k];
-                    g00[2] = (__fp16)k2[k];
-                    g00[3] = (__fp16)k3[k];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    g00 += 4;
-                }
-            }
-        }
-        for (; p < inh; p++)
-        {
-            const float* k0 = kptr0 + p * kernel_w;
-            const float* k1 = kptr1 + p * kernel_w;
-            const float* k2 = kptr2 + p * kernel_w;
-            const float* k3 = kptr3 + p * kernel_w;
-
-            for (int k = 0; k < kernel_w; k++)
-            {
-                g00[0] = (__fp16)k0[k];
-                g00[1] = (__fp16)k1[k];
-                g00[2] = (__fp16)k2[k];
-                g00[3] = (__fp16)k3[k];
-                g00 += 4;
-            }
-        }
-    }
-    for (; q + 1 < outh; q += 2)
-    {
-        const float* kptr0 = (const float*)kernel + q * inh * kernel_w;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w;
-
-        __fp16* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2);
-
-        int p = 0;
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w + k;
-                const float* k1 = kptr1 + p * kernel_w + k;
-
-                g00[0] = (__fp16)k0[0];
-                g00[1] = (__fp16)k0[kernel_w];
-                g00[2] = (__fp16)k0[kernel_w * 2];
-                g00[3] = (__fp16)k0[kernel_w * 3];
-                g00[4] = (__fp16)k0[kernel_w * 4];
-                g00[5] = (__fp16)k0[kernel_w * 5];
-                g00[6] = (__fp16)k0[kernel_w * 6];
-                g00[7] = (__fp16)k0[kernel_w * 7];
-                g00[8] = (__fp16)k1[0];
-                g00[9] = (__fp16)k1[kernel_w];
-                g00[10] = (__fp16)k1[kernel_w * 2];
-                g00[11] = (__fp16)k1[kernel_w * 3];
-                g00[12] = (__fp16)k1[kernel_w * 4];
-                g00[13] = (__fp16)k1[kernel_w * 5];
-                g00[14] = (__fp16)k1[kernel_w * 6];
-                g00[15] = (__fp16)k1[kernel_w * 7];
-                g00 += 16;
-            }
-        }
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w + k;
-                const float* k1 = kptr1 + p * kernel_w + k;
-
-                g00[0] = (__fp16)k0[0];
-                g00[1] = (__fp16)k0[kernel_w];
-                g00[2] = (__fp16)k0[kernel_w * 2];
-                g00[3] = (__fp16)k0[kernel_w * 3];
-                g00[4] = (__fp16)k1[0];
-                g00[5] = (__fp16)k1[kernel_w];
-                g00[6] = (__fp16)k1[kernel_w * 2];
-                g00[7] = (__fp16)k1[kernel_w * 3];
-                g00 += 8;
-            }
-        }
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + p * kernel_w;
-                const float* k1 = kptr1 + p * kernel_w;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    g00[1] = (__fp16)k1[k];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    g00 += 2;
-                }
-            }
-        }
-        for (; p < inh; p++)
-        {
-            const float* k0 = kptr0 + p * kernel_w;
-            const float* k1 = kptr1 + p * kernel_w;
-
-            for (int k = 0; k < kernel_w; k++)
-            {
-                g00[0] = (__fp16)k0[k];
-                g00[1] = (__fp16)k1[k];
-                g00 += 2;
-            }
-        }
-    }
-    for (; q < outh; q++)
-    {
-        const float* kptr = (const float*)kernel + q * inh * kernel_w;
-
-        __fp16* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2 + q % 2);
-
-        int p = 0;
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr + p * kernel_w;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    k0 += kernel_w;
-                    g00 += 1;
-                }
-            }
-        }
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr + p * kernel_w;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    k0 += kernel_w;
-                    g00 += 1;
-                }
-            }
-        }
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr + p * kernel_w;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    k0 += kernel_w;
-                    g00 += 1;
-                }
-            }
-        }
-        for (; p < inh; p++)
-        {
-            const float* k0 = kptr + p * kernel_w;
-
-            for (int k = 0; k < kernel_w; k++)
-            {
-                g00[0] = (__fp16)k0[k];
-                g00++;
-            }
-        }
-    }
-}
-
-static void convolution1d_packed_fp16s(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int kernel_w, int dilation_w, int stride_w, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int elempack = bottom_blob.elempack;
-    const int inh = bottom_blob.h * elempack;
-
-    const int N = bottom_blob.w * elempack;
-
-    const int outw = top_blob.w;
-    const int out_elempack = top_blob.elempack;
-    const int outh = top_blob.h * out_elempack;
-
-    const int M = top_blob.w * out_elempack;
-
-    const float* bias_data_ptr = bias_data;
-
-    int nn_outh = 0;
-    int remain_outh_start = 0;
-    nn_outh = (outh - remain_outh_start) / 8;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = remain_outh_start + pp * 8;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-        const int out_elempack = top_blob.elempack;
-
-        __fp16* outptr = top_blob.row<__fp16>(p / out_elempack);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
-            float32x4_t _sum1 = vdupq_n_f32(0.f);
-            float32x4_t _sum2 = vdupq_n_f32(0.f);
-            float32x4_t _sum3 = vdupq_n_f32(0.f);
-            float32x4_t _sum4 = vdupq_n_f32(0.f);
-            float32x4_t _sum5 = vdupq_n_f32(0.f);
-            float32x4_t _sum6 = vdupq_n_f32(0.f);
-            float32x4_t _sum7 = vdupq_n_f32(0.f);
-
-            if (bias_data_ptr)
-            {
-                _sum0 = vld1q_f32(bias_data_ptr + p);
-                _sum1 = vld1q_f32(bias_data_ptr + p + 4);
-            }
-
-            const __fp16* kptr = weight_data_tm.channel(p / 8);
-
-            int q = 0;
-            for (; q + 7 < inh; q += 8)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    float32x4_t _r1;
-                    if (elempack == 4)
-                    {
-                        _r0 = vcvt_f32_f16(vld1_f16(r0));
-                        _r1 = vcvt_f32_f16(vld1_f16(r0 + N));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        float16x8_t _r_f16 = float16x8_t();
-                        _r_f16 = vsetq_lane_f16(r0[0], _r_f16, 0);
-                        _r_f16 = vsetq_lane_f16(r0[N], _r_f16, 1);
-                        _r_f16 = vsetq_lane_f16(r0[N * 2], _r_f16, 2);
-                        _r_f16 = vsetq_lane_f16(r0[N * 3], _r_f16, 3);
-                        _r_f16 = vsetq_lane_f16(r0[N * 4], _r_f16, 4);
-                        _r_f16 = vsetq_lane_f16(r0[N * 5], _r_f16, 5);
-                        _r_f16 = vsetq_lane_f16(r0[N * 6], _r_f16, 6);
-                        _r_f16 = vsetq_lane_f16(r0[N * 7], _r_f16, 7);
-                        _r0 = vcvt_f32_f16(vget_low_f16(_r_f16));
-                        _r1 = vcvt_f32_f16(vget_high_f16(_r_f16));
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w01 = vld1q_f16(kptr);
-                    float16x8_t _w23 = vld1q_f16(kptr + 8);
-                    float16x8_t _w45 = vld1q_f16(kptr + 16);
-                    float16x8_t _w67 = vld1q_f16(kptr + 24);
-                    float16x8_t _w89 = vld1q_f16(kptr + 32);
-                    float16x8_t _wab = vld1q_f16(kptr + 40);
-                    float16x8_t _wcd = vld1q_f16(kptr + 48);
-                    float16x8_t _wef = vld1q_f16(kptr + 56);
-                    float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01));
-                    float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01));
-                    float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23));
-                    float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23));
-                    float32x4_t _w4 = vcvt_f32_f16(vget_low_f16(_w45));
-                    float32x4_t _w5 = vcvt_f32_f16(vget_high_f16(_w45));
-                    float32x4_t _w6 = vcvt_f32_f16(vget_low_f16(_w67));
-                    float32x4_t _w7 = vcvt_f32_f16(vget_high_f16(_w67));
-                    float32x4_t _w8 = vcvt_f32_f16(vget_low_f16(_w89));
-                    float32x4_t _w9 = vcvt_f32_f16(vget_high_f16(_w89));
-                    float32x4_t _wa = vcvt_f32_f16(vget_low_f16(_wab));
-                    float32x4_t _wb = vcvt_f32_f16(vget_high_f16(_wab));
-                    float32x4_t _wc = vcvt_f32_f16(vget_low_f16(_wcd));
-                    float32x4_t _wd = vcvt_f32_f16(vget_high_f16(_wcd));
-                    float32x4_t _we = vcvt_f32_f16(vget_low_f16(_wef));
-                    float32x4_t _wf = vcvt_f32_f16(vget_high_f16(_wef));
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 0);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 1);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 1);
-                    _sum4 = vfmaq_laneq_f32(_sum4, _w4, _r0, 2);
-                    _sum5 = vfmaq_laneq_f32(_sum5, _w5, _r0, 2);
-                    _sum6 = vfmaq_laneq_f32(_sum6, _w6, _r0, 3);
-                    _sum7 = vfmaq_laneq_f32(_sum7, _w7, _r0, 3);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w8, _r1, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w9, _r1, 0);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _wa, _r1, 1);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _wb, _r1, 1);
-                    _sum4 = vfmaq_laneq_f32(_sum4, _wc, _r1, 2);
-                    _sum5 = vfmaq_laneq_f32(_sum5, _wd, _r1, 2);
-                    _sum6 = vfmaq_laneq_f32(_sum6, _we, _r1, 3);
-                    _sum7 = vfmaq_laneq_f32(_sum7, _wf, _r1, 3);
-
-                    kptr += 64;
-                }
-            }
-            for (; q + 3 < inh; q += 4)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = vcvt_f32_f16(vld1_f16(r0));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        float16x4_t _r_f16 = float16x4_t();
-                        _r_f16 = vset_lane_f16(r0[0], _r_f16, 0);
-                        _r_f16 = vset_lane_f16(r0[N], _r_f16, 1);
-                        _r_f16 = vset_lane_f16(r0[N * 2], _r_f16, 2);
-                        _r_f16 = vset_lane_f16(r0[N * 3], _r_f16, 3);
-                        _r0 = vcvt_f32_f16(_r_f16);
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w01 = vld1q_f16(kptr);
-                    float16x8_t _w23 = vld1q_f16(kptr + 8);
-                    float16x8_t _w45 = vld1q_f16(kptr + 16);
-                    float16x8_t _w67 = vld1q_f16(kptr + 24);
-                    float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01));
-                    float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01));
-                    float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23));
-                    float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23));
-                    float32x4_t _w4 = vcvt_f32_f16(vget_low_f16(_w45));
-                    float32x4_t _w5 = vcvt_f32_f16(vget_high_f16(_w45));
-                    float32x4_t _w6 = vcvt_f32_f16(vget_low_f16(_w67));
-                    float32x4_t _w7 = vcvt_f32_f16(vget_high_f16(_w67));
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 0);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 1);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 1);
-                    _sum4 = vfmaq_laneq_f32(_sum4, _w4, _r0, 2);
-                    _sum5 = vfmaq_laneq_f32(_sum5, _w5, _r0, 2);
-                    _sum6 = vfmaq_laneq_f32(_sum6, _w6, _r0, 3);
-                    _sum7 = vfmaq_laneq_f32(_sum7, _w7, _r0, 3);
-
-                    kptr += 32;
-                }
-            }
-            for (; q + 1 < inh; q += 2)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val0;
-                    float val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = (float)(r0[0]);
-                        val1 = (float)(r0[N]);
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w01 = vld1q_f16(kptr);
-                    float16x8_t _w23 = vld1q_f16(kptr + 8);
-                    float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01));
-                    float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01));
-                    float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23));
-                    float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23));
-                    _sum0 = vfmaq_n_f32(_sum0, _w0, val0);
-                    _sum1 = vfmaq_n_f32(_sum1, _w1, val0);
-                    _sum2 = vfmaq_n_f32(_sum2, _w2, val1);
-                    _sum3 = vfmaq_n_f32(_sum3, _w3, val1);
-
-                    kptr += 16;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _val;
-                    // if (elempack == 1)
-                    {
-                        _val = vcvt_f32_f16(vdup_n_f16(r0[0]));
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w = vld1q_f16(kptr);
-                    float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w));
-                    float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w));
-                    _sum0 = vfmaq_f32(_sum0, _w0, _val);
-                    _sum1 = vfmaq_f32(_sum1, _w1, _val);
-
-                    kptr += 8;
-                }
-            }
-
-            _sum0 = vaddq_f32(_sum0, _sum2);
-            _sum1 = vaddq_f32(_sum1, _sum3);
-            _sum4 = vaddq_f32(_sum4, _sum6);
-            _sum5 = vaddq_f32(_sum5, _sum7);
-            _sum0 = vaddq_f32(_sum0, _sum4);
-            _sum1 = vaddq_f32(_sum1, _sum5);
-
-            _sum0 = activation_ps(_sum0, activation_type, activation_params);
-            _sum1 = activation_ps(_sum1, activation_type, activation_params);
-
-            if (out_elempack == 4)
-            {
-                vst1_f16(outptr, vcvt_f16_f32(_sum0));
-                vst1_f16(outptr + M, vcvt_f16_f32(_sum1));
-                outptr += 4;
-            }
-            else // if (out_elempack == 1)
-            {
-                float16x4_t _sum0_f16 = vcvt_f16_f32(_sum0);
-                float16x4_t _sum1_f16 = vcvt_f16_f32(_sum1);
-                outptr[0] = vget_lane_f16(_sum0_f16, 0);
-                outptr[M] = vget_lane_f16(_sum0_f16, 1);
-                outptr[M * 2] = vget_lane_f16(_sum0_f16, 2);
-                outptr[M * 3] = vget_lane_f16(_sum0_f16, 3);
-                outptr[M * 4] = vget_lane_f16(_sum1_f16, 0);
-                outptr[M * 5] = vget_lane_f16(_sum1_f16, 1);
-                outptr[M * 6] = vget_lane_f16(_sum1_f16, 2);
-                outptr[M * 7] = vget_lane_f16(_sum1_f16, 3);
-                outptr += 1;
-            }
-        }
-    }
-    remain_outh_start += nn_outh * 8;
-    nn_outh = (outh - remain_outh_start) / 4;
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = remain_outh_start + pp * 4;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-        const int out_elempack = top_blob.elempack;
-
-        __fp16* outptr = top_blob.row<__fp16>(p / out_elempack);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
-            float32x4_t _sum1 = vdupq_n_f32(0.f);
-            float32x4_t _sum2 = vdupq_n_f32(0.f);
-            float32x4_t _sum3 = vdupq_n_f32(0.f);
-
-            if (bias_data_ptr)
-            {
-                _sum0 = vld1q_f32(bias_data_ptr + p);
-            }
-
-            const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4);
-
-            int q = 0;
-            for (; q + 7 < inh; q += 8)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    float32x4_t _r1;
-                    if (elempack == 4)
-                    {
-                        _r0 = vcvt_f32_f16(vld1_f16(r0));
-                        _r1 = vcvt_f32_f16(vld1_f16(r0 + N));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        float16x8_t _r_f16 = float16x8_t();
-                        _r_f16 = vsetq_lane_f16(r0[0], _r_f16, 0);
-                        _r_f16 = vsetq_lane_f16(r0[N], _r_f16, 1);
-                        _r_f16 = vsetq_lane_f16(r0[N * 2], _r_f16, 2);
-                        _r_f16 = vsetq_lane_f16(r0[N * 3], _r_f16, 3);
-                        _r_f16 = vsetq_lane_f16(r0[N * 4], _r_f16, 4);
-                        _r_f16 = vsetq_lane_f16(r0[N * 5], _r_f16, 5);
-                        _r_f16 = vsetq_lane_f16(r0[N * 6], _r_f16, 6);
-                        _r_f16 = vsetq_lane_f16(r0[N * 7], _r_f16, 7);
-                        _r0 = vcvt_f32_f16(vget_low_f16(_r_f16));
-                        _r1 = vcvt_f32_f16(vget_high_f16(_r_f16));
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w01 = vld1q_f16(kptr);
-                    float16x8_t _w23 = vld1q_f16(kptr + 8);
-                    float16x8_t _w45 = vld1q_f16(kptr + 16);
-                    float16x8_t _w67 = vld1q_f16(kptr + 24);
-                    float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01));
-                    float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01));
-                    float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23));
-                    float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23));
-                    float32x4_t _w4 = vcvt_f32_f16(vget_low_f16(_w45));
-                    float32x4_t _w5 = vcvt_f32_f16(vget_high_f16(_w45));
-                    float32x4_t _w6 = vcvt_f32_f16(vget_low_f16(_w67));
-                    float32x4_t _w7 = vcvt_f32_f16(vget_high_f16(_w67));
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 1);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 2);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 3);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w4, _r1, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w5, _r1, 1);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w6, _r1, 2);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w7, _r1, 3);
-
-                    kptr += 32;
-                }
-            }
-            for (; q + 3 < inh; q += 4)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = vcvt_f32_f16(vld1_f16(r0));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        float16x4_t _r_f16 = float16x4_t();
-                        _r_f16 = vset_lane_f16(r0[0], _r_f16, 0);
-                        _r_f16 = vset_lane_f16(r0[N], _r_f16, 1);
-                        _r_f16 = vset_lane_f16(r0[N * 2], _r_f16, 2);
-                        _r_f16 = vset_lane_f16(r0[N * 3], _r_f16, 3);
-                        _r0 = vcvt_f32_f16(_r_f16);
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w01 = vld1q_f16(kptr);
-                    float16x8_t _w23 = vld1q_f16(kptr + 8);
-                    float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01));
-                    float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01));
-                    float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23));
-                    float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23));
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 1);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 2);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 3);
-
-                    kptr += 16;
-                }
-            }
-            for (; q + 1 < inh; q += 2)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val0;
-                    float val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = (float)(r0[0]);
-                        val1 = (float)(r0[N]);
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w = vld1q_f16(kptr);
-                    float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w));
-                    float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w));
-                    _sum0 = vfmaq_n_f32(_sum0, _w0, val0);
-                    _sum1 = vfmaq_n_f32(_sum1, _w1, val1);
-
-                    kptr += 8;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _val;
-                    // if (elempack == 1)
-                    {
-                        _val = vcvt_f32_f16(vdup_n_f16(r0[0]));
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w = vcvt_f32_f16(vld1_f16(kptr));
-                    _sum0 = vfmaq_f32(_sum0, _val, _w);
-
-                    kptr += 4;
-                }
-            }
-
-            _sum0 = vaddq_f32(_sum0, _sum1);
-            _sum2 = vaddq_f32(_sum2, _sum3);
-            _sum0 = vaddq_f32(_sum0, _sum2);
-
-            _sum0 = activation_ps(_sum0, activation_type, activation_params);
-
-            if (out_elempack == 4)
-            {
-                vst1_f16(outptr, vcvt_f16_f32(_sum0));
-                outptr += 4;
-            }
-            else // if (out_elempack == 1)
-            {
-                float16x4_t _sum0_f16 = vcvt_f16_f32(_sum0);
-                outptr[0] = vget_lane_f16(_sum0_f16, 0);
-                outptr[M] = vget_lane_f16(_sum0_f16, 1);
-                outptr[M * 2] = vget_lane_f16(_sum0_f16, 2);
-                outptr[M * 3] = vget_lane_f16(_sum0_f16, 3);
-                outptr += 1;
-            }
-        }
-    }
-    remain_outh_start += nn_outh * 4;
-    nn_outh = (outh - remain_outh_start) / 2;
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = remain_outh_start + pp * 2;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-
-        __fp16* outptr0 = top_blob.row<__fp16>(p);
-        __fp16* outptr1 = top_blob.row<__fp16>(p + 1);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float sum0 = 0.f;
-            float sum1 = 0.f;
-
-            if (bias_data_ptr)
-            {
-                sum0 = bias_data_ptr[p];
-                sum1 = bias_data_ptr[p + 1];
-            }
-
-            const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2);
-
-            int q = 0;
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
-            float32x4_t _sum1 = vdupq_n_f32(0.f);
-            float32x4_t _sum2 = vdupq_n_f32(0.f);
-            float32x4_t _sum3 = vdupq_n_f32(0.f);
-            for (; q + 7 < inh; q += 8)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    float32x4_t _r1;
-                    if (elempack == 4)
-                    {
-                        _r0 = vcvt_f32_f16(vld1_f16(r0));
-                        _r1 = vcvt_f32_f16(vld1_f16(r0 + N));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        float16x8_t _r01_f16 = float16x8_t();
-                        _r01_f16 = vsetq_lane_f16(r0[0], _r01_f16, 0);
-                        _r01_f16 = vsetq_lane_f16(r0[N], _r01_f16, 1);
-                        _r01_f16 = vsetq_lane_f16(r0[N * 2], _r01_f16, 2);
-                        _r01_f16 = vsetq_lane_f16(r0[N * 3], _r01_f16, 3);
-                        _r01_f16 = vsetq_lane_f16(r0[N * 4], _r01_f16, 4);
-                        _r01_f16 = vsetq_lane_f16(r0[N * 5], _r01_f16, 5);
-                        _r01_f16 = vsetq_lane_f16(r0[N * 6], _r01_f16, 6);
-                        _r01_f16 = vsetq_lane_f16(r0[N * 7], _r01_f16, 7);
-                        _r0 = vcvt_f32_f16(vget_low_f16(_r01_f16));
-                        _r1 = vcvt_f32_f16(vget_high_f16(_r01_f16));
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w01 = vld1q_f16(kptr);
-                    float16x8_t _w23 = vld1q_f16(kptr + 8);
-                    float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01));
-                    float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01));
-                    float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23));
-                    float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23));
-                    _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                    _sum1 = vfmaq_f32(_sum1, _r1, _w1);
-                    _sum2 = vfmaq_f32(_sum2, _r0, _w2);
-                    _sum3 = vfmaq_f32(_sum3, _r1, _w3);
-
-                    kptr += 16;
-                }
-            }
-            _sum0 = vaddq_f32(_sum0, _sum1);
-            _sum2 = vaddq_f32(_sum2, _sum3);
-            sum0 += vaddvq_f32(_sum0);
-            sum1 += vaddvq_f32(_sum2);
-            _sum0 = vdupq_n_f32(0.f);
-            _sum1 = vdupq_n_f32(0.f);
-            for (; q + 3 < inh; q += 4)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = vcvt_f32_f16(vld1_f16(r0));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        float16x4_t _r0_f16 = float16x4_t();
-                        _r0_f16 = vset_lane_f16(r0[0], _r0_f16, 0);
-                        _r0_f16 = vset_lane_f16(r0[N], _r0_f16, 1);
-                        _r0_f16 = vset_lane_f16(r0[N * 2], _r0_f16, 2);
-                        _r0_f16 = vset_lane_f16(r0[N * 3], _r0_f16, 3);
-                        _r0 = vcvt_f32_f16(_r0_f16);
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w = vld1q_f16(kptr);
-                    float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w));
-                    float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w));
-                    _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                    _sum1 = vfmaq_f32(_sum1, _r0, _w1);
-
-                    kptr += 8;
-                }
-            }
-            sum0 += vaddvq_f32(_sum0);
-            sum1 += vaddvq_f32(_sum1);
-            for (; q + 1 < inh; q += 2)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val0;
-                    float val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = (float)(r0[0]);
-                        val1 = (float)(r0[N]);
-                        r0 += dilation_w;
-                    }
-
-                    sum0 += val0 * (float)(kptr[0]);
-                    sum1 += val0 * (float)(kptr[1]);
-                    sum0 += val1 * (float)(kptr[2]);
-                    sum1 += val1 * (float)(kptr[3]);
-
-                    kptr += 4;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val;
-                    // if (elempack == 1)
-                    {
-                        val = (float)(r0[0]);
-                        r0 += dilation_w;
-                    }
-
-                    sum0 += val * (float)(kptr[0]);
-                    sum1 += val * (float)(kptr[1]);
-
-                    kptr += 2;
-                }
-            }
-
-            sum0 = activation_ss(sum0, activation_type, activation_params);
-            sum1 = activation_ss(sum1, activation_type, activation_params);
-
-            outptr0[0] = (__fp16)(sum0);
-            outptr1[0] = (__fp16)(sum1);
-            outptr0 += 1;
-            outptr1 += 1;
-        }
-    }
-    remain_outh_start += nn_outh * 2;
-    for (int p = remain_outh_start; p < outh; p++)
-    {
-        __fp16* outptr = top_blob.row<__fp16>(p);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float sum = 0.f;
-
-            if (bias_data_ptr)
-            {
-                sum = bias_data_ptr[p];
-            }
-
-            const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-
-            int q = 0;
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
-            float32x4_t _sum1 = vdupq_n_f32(0.f);
-            for (; q + 7 < inh; q += 8)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    float32x4_t _r1;
-                    if (elempack == 4)
-                    {
-                        _r0 = vcvt_f32_f16(vld1_f16(r0));
-                        _r1 = vcvt_f32_f16(vld1_f16(r0 + N));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        float16x8_t _r01_f16 = float16x8_t();
-                        _r01_f16 = vsetq_lane_f16(r0[0], _r01_f16, 0);
-                        _r01_f16 = vsetq_lane_f16(r0[N], _r01_f16, 1);
-                        _r01_f16 = vsetq_lane_f16(r0[N * 2], _r01_f16, 2);
-                        _r01_f16 = vsetq_lane_f16(r0[N * 3], _r01_f16, 3);
-                        _r01_f16 = vsetq_lane_f16(r0[N * 4], _r01_f16, 4);
-                        _r01_f16 = vsetq_lane_f16(r0[N * 5], _r01_f16, 5);
-                        _r01_f16 = vsetq_lane_f16(r0[N * 6], _r01_f16, 6);
-                        _r01_f16 = vsetq_lane_f16(r0[N * 7], _r01_f16, 7);
-                        _r0 = vcvt_f32_f16(vget_low_f16(_r01_f16));
-                        _r1 = vcvt_f32_f16(vget_high_f16(_r01_f16));
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w = vld1q_f16(kptr);
-                    float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w));
-                    float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w));
-                    _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                    _sum1 = vfmaq_f32(_sum1, _r1, _w1);
-
-                    kptr += 8;
-                }
-            }
-            _sum0 = vaddq_f32(_sum0, _sum1);
-            sum += vaddvq_f32(_sum0);
-            float32x4_t _sum = vdupq_n_f32(0.f);
-            for (; q + 3 < inh; q += 4)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float32x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = vcvt_f32_f16(vld1_f16(r0));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        float16x4_t _r0_f16 = float16x4_t();
-                        _r0_f16 = vset_lane_f16(r0[0], _r0_f16, 0);
-                        _r0_f16 = vset_lane_f16(r0[N], _r0_f16, 1);
-                        _r0_f16 = vset_lane_f16(r0[N * 2], _r0_f16, 2);
-                        _r0_f16 = vset_lane_f16(r0[N * 3], _r0_f16, 3);
-                        _r0 = vcvt_f32_f16(_r0_f16);
-                        r0 += dilation_w;
-                    }
-
-                    float32x4_t _w = vcvt_f32_f16(vld1_f16(kptr));
-                    _sum = vfmaq_f32(_sum, _r0, _w);
-
-                    kptr += 4;
-                }
-            }
-            sum += vaddvq_f32(_sum);
-            for (; q + 1 < inh; q += 2)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val0;
-                    float val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = (float)(r0[0]);
-                        val1 = (float)(r0[N]);
-                        r0 += dilation_w;
-                    }
-
-                    sum += val0 * (float)(kptr[0]);
-                    sum += val1 * (float)(kptr[1]);
-
-                    kptr += 2;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float val;
-                    // if (elempack == 1)
-                    {
-                        val = (float)(r0[0]);
-                        r0 += dilation_w;
-                    }
-
-                    sum += val * (float)(kptr[0]);
-
-                    kptr += 1;
-                }
-            }
-
-            sum = activation_ss(sum, activation_type, activation_params);
-
-            outptr[0] = (__fp16)(sum);
-            outptr += 1;
-        }
-    }
-}
-
-static void convolution1d_packed_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int kernel_w, int dilation_w, int stride_w, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int elempack = bottom_blob.elempack;
-    const int inh = bottom_blob.h * elempack;
-
-    const int N = bottom_blob.w * elempack;
-
-    const int outw = top_blob.w;
-    const int out_elempack = top_blob.elempack;
-    const int outh = top_blob.h * out_elempack;
-
-    const int M = top_blob.w * out_elempack;
-
-    const __fp16* bias_data_ptr = bias_data;
-
-    int nn_outh = 0;
-    int remain_outh_start = 0;
-    nn_outh = (outh - remain_outh_start) / 8;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = remain_outh_start + pp * 8;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-        const int out_elempack = top_blob.elempack;
-
-        __fp16* outptr = top_blob.row<__fp16>(p / out_elempack);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float16x8_t _sum0 = vdupq_n_f16(0.f);
-            float16x8_t _sum1 = vdupq_n_f16(0.f);
-            float16x8_t _sum2 = vdupq_n_f16(0.f);
-            float16x8_t _sum3 = vdupq_n_f16(0.f);
-
-            if (bias_data_ptr)
-            {
-                _sum0 = vld1q_f16(bias_data_ptr + p);
-            }
-
-            const __fp16* kptr = weight_data_tm.channel(p / 8);
-
-            int q = 0;
-            for (; q + 7 < inh; q += 8)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float16x8_t _r0;
-                    if (elempack == 8)
-                    {
-                        _r0 = vld1q_f16(r0);
-                        r0 += dilation_w * 8;
-                    }
-                    else if (elempack == 4)
-                    {
-                        _r0 = vcombine_f16(vld1_f16(r0), vld1_f16(r0 + N));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float16x8_t();
-                        _r0 = vsetq_lane_f16(r0[0], _r0, 0);
-                        _r0 = vsetq_lane_f16(r0[N], _r0, 1);
-                        _r0 = vsetq_lane_f16(r0[N * 2], _r0, 2);
-                        _r0 = vsetq_lane_f16(r0[N * 3], _r0, 3);
-                        _r0 = vsetq_lane_f16(r0[N * 4], _r0, 4);
-                        _r0 = vsetq_lane_f16(r0[N * 5], _r0, 5);
-                        _r0 = vsetq_lane_f16(r0[N * 6], _r0, 6);
-                        _r0 = vsetq_lane_f16(r0[N * 7], _r0, 7);
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w0 = vld1q_f16(kptr);
-                    float16x8_t _w1 = vld1q_f16(kptr + 8);
-                    float16x8_t _w2 = vld1q_f16(kptr + 8 * 2);
-                    float16x8_t _w3 = vld1q_f16(kptr + 8 * 3);
-                    float16x8_t _w4 = vld1q_f16(kptr + 8 * 4);
-                    float16x8_t _w5 = vld1q_f16(kptr + 8 * 5);
-                    float16x8_t _w6 = vld1q_f16(kptr + 8 * 6);
-                    float16x8_t _w7 = vld1q_f16(kptr + 8 * 7);
-                    _sum0 = vfmaq_laneq_f16(_sum0, _w0, _r0, 0);
-                    _sum1 = vfmaq_laneq_f16(_sum1, _w1, _r0, 1);
-                    _sum2 = vfmaq_laneq_f16(_sum2, _w2, _r0, 2);
-                    _sum3 = vfmaq_laneq_f16(_sum3, _w3, _r0, 3);
-                    _sum0 = vfmaq_laneq_f16(_sum0, _w4, _r0, 4);
-                    _sum1 = vfmaq_laneq_f16(_sum1, _w5, _r0, 5);
-                    _sum2 = vfmaq_laneq_f16(_sum2, _w6, _r0, 6);
-                    _sum3 = vfmaq_laneq_f16(_sum3, _w7, _r0, 7);
-
-                    kptr += 64;
-                }
-            }
-            for (; q + 3 < inh; q += 4)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float16x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1_f16(r0);
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float16x4_t();
-                        _r0 = vset_lane_f16(r0[0], _r0, 0);
-                        _r0 = vset_lane_f16(r0[N], _r0, 1);
-                        _r0 = vset_lane_f16(r0[N * 2], _r0, 2);
-                        _r0 = vset_lane_f16(r0[N * 3], _r0, 3);
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w0 = vld1q_f16(kptr);
-                    float16x8_t _w1 = vld1q_f16(kptr + 8);
-                    float16x8_t _w2 = vld1q_f16(kptr + 8 * 2);
-                    float16x8_t _w3 = vld1q_f16(kptr + 8 * 3);
-                    _sum0 = vfmaq_lane_f16(_sum0, _w0, _r0, 0);
-                    _sum1 = vfmaq_lane_f16(_sum1, _w1, _r0, 1);
-                    _sum2 = vfmaq_lane_f16(_sum2, _w2, _r0, 2);
-                    _sum3 = vfmaq_lane_f16(_sum3, _w3, _r0, 3);
-
-                    kptr += 32;
-                }
-            }
-            for (; q + 1 < inh; q += 2)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    __fp16 val0;
-                    __fp16 val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = r0[0];
-                        val1 = r0[N];
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w0 = vld1q_f16(kptr);
-                    float16x8_t _w1 = vld1q_f16(kptr + 8);
-                    _sum0 = vfmaq_n_f16(_sum0, _w0, val0);
-                    _sum1 = vfmaq_n_f16(_sum1, _w1, val1);
-
-                    kptr += 16;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float16x8_t _val;
-                    // if (elempack == 1)
-                    {
-                        _val = vdupq_n_f16(r0[0]);
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w0 = vld1q_f16(kptr);
-                    _sum0 = vfmaq_f16(_sum0, _w0, _val);
-
-                    kptr += 8;
-                }
-            }
-
-            _sum0 = vaddq_f16(_sum0, _sum1);
-            _sum2 = vaddq_f16(_sum2, _sum3);
-            _sum0 = vaddq_f16(_sum0, _sum2);
-
-            _sum0 = activation_ps_f16(_sum0, activation_type, activation_params);
-
-            if (out_elempack == 8)
-            {
-                vst1q_f16(outptr, _sum0);
-                outptr += 8;
-            }
-            else if (out_elempack == 4)
-            {
-                vst1_f16(outptr, vget_low_f16(_sum0));
-                vst1_f16(outptr + M, vget_high_f16(_sum0));
-                outptr += 4;
-            }
-            else // if (out_elempack == 1)
-            {
-                outptr[0] = vgetq_lane_f16(_sum0, 0);
-                outptr[M] = vgetq_lane_f16(_sum0, 1);
-                outptr[M * 2] = vgetq_lane_f16(_sum0, 2);
-                outptr[M * 3] = vgetq_lane_f16(_sum0, 3);
-                outptr[M * 4] = vgetq_lane_f16(_sum0, 4);
-                outptr[M * 5] = vgetq_lane_f16(_sum0, 5);
-                outptr[M * 6] = vgetq_lane_f16(_sum0, 6);
-                outptr[M * 7] = vgetq_lane_f16(_sum0, 7);
-                outptr += 1;
-            }
-        }
-    }
-    remain_outh_start += nn_outh * 8;
-    nn_outh = (outh - remain_outh_start) / 4;
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = remain_outh_start + pp * 4;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-        const int out_elempack = top_blob.elempack;
-
-        __fp16* outptr = top_blob.row<__fp16>(p / out_elempack);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float16x4_t _sum0 = vdup_n_f16(0.f);
-            float16x4_t _sum1 = vdup_n_f16(0.f);
-            float16x4_t _sum2 = vdup_n_f16(0.f);
-            float16x4_t _sum3 = vdup_n_f16(0.f);
-
-            if (bias_data_ptr)
-            {
-                _sum0 = vld1_f16(bias_data_ptr + p);
-            }
-
-            const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4);
-
-            int q = 0;
-            for (; q + 7 < inh; q += 8)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float16x4_t _r0;
-                    float16x4_t _r1;
-                    if (elempack == 8)
-                    {
-                        float16x8_t _r01 = vld1q_f16(r0);
-                        _r0 = vget_low_f16(_r01);
-                        _r1 = vget_high_f16(_r01);
-                        r0 += dilation_w * 8;
-                    }
-                    else if (elempack == 4)
-                    {
-                        _r0 = vld1_f16(r0);
-                        _r1 = vld1_f16(r0 + N);
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float16x4_t();
-                        _r1 = float16x4_t();
-                        _r0 = vset_lane_f16(r0[0], _r0, 0);
-                        _r0 = vset_lane_f16(r0[N], _r0, 1);
-                        _r0 = vset_lane_f16(r0[N * 2], _r0, 2);
-                        _r0 = vset_lane_f16(r0[N * 3], _r0, 3);
-                        _r1 = vset_lane_f16(r0[N * 4], _r1, 0);
-                        _r1 = vset_lane_f16(r0[N * 5], _r1, 1);
-                        _r1 = vset_lane_f16(r0[N * 6], _r1, 2);
-                        _r1 = vset_lane_f16(r0[N * 7], _r1, 3);
-                        r0 += dilation_w;
-                    }
-
-                    float16x4_t _w0 = vld1_f16(kptr);
-                    float16x4_t _w1 = vld1_f16(kptr + 4);
-                    float16x4_t _w2 = vld1_f16(kptr + 8);
-                    float16x4_t _w3 = vld1_f16(kptr + 12);
-                    float16x4_t _w4 = vld1_f16(kptr + 16);
-                    float16x4_t _w5 = vld1_f16(kptr + 20);
-                    float16x4_t _w6 = vld1_f16(kptr + 24);
-                    float16x4_t _w7 = vld1_f16(kptr + 28);
-                    _sum0 = vfma_lane_f16(_sum0, _w0, _r0, 0);
-                    _sum1 = vfma_lane_f16(_sum1, _w1, _r0, 1);
-                    _sum2 = vfma_lane_f16(_sum2, _w2, _r0, 2);
-                    _sum3 = vfma_lane_f16(_sum3, _w3, _r0, 3);
-                    _sum0 = vfma_lane_f16(_sum0, _w4, _r1, 0);
-                    _sum1 = vfma_lane_f16(_sum1, _w5, _r1, 1);
-                    _sum2 = vfma_lane_f16(_sum2, _w6, _r1, 2);
-                    _sum3 = vfma_lane_f16(_sum3, _w7, _r1, 3);
-
-                    kptr += 32;
-                }
-            }
-            for (; q + 3 < inh; q += 4)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float16x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1_f16(r0);
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float16x4_t();
-                        _r0 = vset_lane_f16(r0[0], _r0, 0);
-                        _r0 = vset_lane_f16(r0[N], _r0, 1);
-                        _r0 = vset_lane_f16(r0[N * 2], _r0, 2);
-                        _r0 = vset_lane_f16(r0[N * 3], _r0, 3);
-                        r0 += dilation_w;
-                    }
-
-                    float16x4_t _w0 = vld1_f16(kptr);
-                    float16x4_t _w1 = vld1_f16(kptr + 4);
-                    float16x4_t _w2 = vld1_f16(kptr + 8);
-                    float16x4_t _w3 = vld1_f16(kptr + 12);
-                    _sum0 = vfma_lane_f16(_sum0, _w0, _r0, 0);
-                    _sum1 = vfma_lane_f16(_sum1, _w1, _r0, 1);
-                    _sum2 = vfma_lane_f16(_sum2, _w2, _r0, 2);
-                    _sum3 = vfma_lane_f16(_sum3, _w3, _r0, 3);
-
-                    kptr += 16;
-                }
-            }
-            for (; q + 1 < inh; q += 2)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    __fp16 val0;
-                    __fp16 val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = r0[0];
-                        val1 = r0[N];
-                        r0 += dilation_w;
-                    }
-
-                    float16x4_t _w0 = vld1_f16(kptr);
-                    float16x4_t _w1 = vld1_f16(kptr + 4);
-                    _sum0 = vfma_n_f16(_sum0, _w0, val0);
-                    _sum1 = vfma_n_f16(_sum1, _w1, val1);
-
-                    kptr += 8;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float16x4_t _val;
-                    // if (elempack == 1)
-                    {
-                        _val = vdup_n_f16(r0[0]);
-                        r0 += dilation_w;
-                    }
-
-                    float16x4_t _w = vld1_f16(kptr);
-                    _sum0 = vfma_f16(_sum0, _val, _w);
-
-                    kptr += 4;
-                }
-            }
-
-            _sum0 = vadd_f16(_sum0, _sum1);
-            _sum2 = vadd_f16(_sum2, _sum3);
-            _sum0 = vadd_f16(_sum0, _sum2);
-
-            _sum0 = activation_ps_f16(_sum0, activation_type, activation_params);
-
-            if (out_elempack == 4)
-            {
-                vst1_f16(outptr, _sum0);
-                outptr += 4;
-            }
-            else // if (out_elempack == 1)
-            {
-                outptr[0] = vget_lane_f16(_sum0, 0);
-                outptr[M] = vget_lane_f16(_sum0, 1);
-                outptr[M * 2] = vget_lane_f16(_sum0, 2);
-                outptr[M * 3] = vget_lane_f16(_sum0, 3);
-                outptr += 1;
-            }
-        }
-    }
-    remain_outh_start += nn_outh * 4;
-    nn_outh = (outh - remain_outh_start) / 2;
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = remain_outh_start + pp * 2;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-
-        __fp16* outptr0 = top_blob.row<__fp16>(p);
-        __fp16* outptr1 = top_blob.row<__fp16>(p + 1);
-
-        for (int j = 0; j < outw; j++)
-        {
-            __fp16 sum0 = 0.f;
-            __fp16 sum1 = 0.f;
-
-            if (bias_data_ptr)
-            {
-                sum0 = bias_data_ptr[p];
-                sum1 = bias_data_ptr[p + 1];
-            }
-
-            const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2);
-
-            int q = 0;
-            float16x8_t _sum0 = vdupq_n_f16(0.f);
-            float16x8_t _sum1 = vdupq_n_f16(0.f);
-            for (; q + 7 < inh; q += 8)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float16x8_t _r0;
-                    if (elempack == 8)
-                    {
-                        _r0 = vld1q_f16(r0);
-                        r0 += dilation_w * 8;
-                    }
-                    else if (elempack == 4)
-                    {
-                        _r0 = vcombine_f16(vld1_f16(r0), vld1_f16(r0 + N));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float16x8_t();
-                        _r0 = vsetq_lane_f16(r0[0], _r0, 0);
-                        _r0 = vsetq_lane_f16(r0[N], _r0, 1);
-                        _r0 = vsetq_lane_f16(r0[N * 2], _r0, 2);
-                        _r0 = vsetq_lane_f16(r0[N * 3], _r0, 3);
-                        _r0 = vsetq_lane_f16(r0[N * 4], _r0, 4);
-                        _r0 = vsetq_lane_f16(r0[N * 5], _r0, 5);
-                        _r0 = vsetq_lane_f16(r0[N * 6], _r0, 6);
-                        _r0 = vsetq_lane_f16(r0[N * 7], _r0, 7);
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w0 = vld1q_f16(kptr);
-                    float16x8_t _w1 = vld1q_f16(kptr + 8);
-                    _sum0 = vfmaq_f16(_sum0, _r0, _w0);
-                    _sum1 = vfmaq_f16(_sum1, _r0, _w1);
-
-                    kptr += 16;
-                }
-            }
-            for (; q + 3 < inh; q += 4)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float16x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1_f16(r0);
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float16x4_t();
-                        _r0 = vset_lane_f16(r0[0], _r0, 0);
-                        _r0 = vset_lane_f16(r0[N], _r0, 1);
-                        _r0 = vset_lane_f16(r0[N * 2], _r0, 2);
-                        _r0 = vset_lane_f16(r0[N * 3], _r0, 3);
-                        r0 += dilation_w;
-                    }
-
-                    float16x4_t _w0 = vld1_f16(kptr);
-                    float16x4_t _w1 = vld1_f16(kptr + 4);
-                    _sum0 = vcombine_f16(vfma_f16(vget_low_f16(_sum0), _r0, _w0), vget_high_f16(_sum0));
-                    _sum1 = vcombine_f16(vfma_f16(vget_low_f16(_sum1), _r0, _w1), vget_high_f16(_sum1));
-
-                    kptr += 8;
-                }
-            }
-            for (; q + 1 < inh; q += 2)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    __fp16 val0;
-                    __fp16 val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = r0[0];
-                        val1 = r0[N];
-                        r0 += dilation_w;
-                    }
-
-                    sum0 += val0 * kptr[0];
-                    sum1 += val0 * kptr[1];
-                    sum0 += val1 * kptr[2];
-                    sum1 += val1 * kptr[3];
-
-                    kptr += 4;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    __fp16 val;
-                    // if (elempack == 1)
-                    {
-                        val = r0[0];
-                        r0 += dilation_w;
-                    }
-
-                    sum0 += val * kptr[0];
-                    sum1 += val * kptr[1];
-
-                    kptr += 2;
-                }
-            }
-
-            float16x4_t _ss0 = vadd_f16(vget_low_f16(_sum0), vget_high_f16(_sum0));
-            float16x4_t _ss1 = vadd_f16(vget_low_f16(_sum1), vget_high_f16(_sum1));
-            float16x4_t _ss = vpadd_f16(_ss0, _ss1);
-            _ss = vpadd_f16(_ss, _ss);
-            sum0 += vget_lane_f16(_ss, 0);
-            sum1 += vget_lane_f16(_ss, 1);
-
-            sum0 = activation_ss_f16(sum0, activation_type, activation_params);
-            sum1 = activation_ss_f16(sum1, activation_type, activation_params);
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-            outptr0 += 1;
-            outptr1 += 1;
-        }
-    }
-    remain_outh_start += nn_outh * 2;
-    for (int p = remain_outh_start; p < outh; p++)
-    {
-        __fp16* outptr = top_blob.row<__fp16>(p);
-
-        for (int j = 0; j < outw; j++)
-        {
-            __fp16 sum = 0.f;
-
-            if (bias_data_ptr)
-            {
-                sum = bias_data_ptr[p];
-            }
-
-            const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-
-            int q = 0;
-            float16x8_t _sum = vdupq_n_f16(0.f);
-            for (; q + 7 < inh; q += 8)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float16x8_t _r0;
-                    if (elempack == 8)
-                    {
-                        _r0 = vld1q_f16(r0);
-                        r0 += dilation_w * 8;
-                    }
-                    else if (elempack == 4)
-                    {
-                        _r0 = vcombine_f16(vld1_f16(r0), vld1_f16(r0 + N));
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float16x8_t();
-                        _r0 = vsetq_lane_f16(r0[0], _r0, 0);
-                        _r0 = vsetq_lane_f16(r0[N], _r0, 1);
-                        _r0 = vsetq_lane_f16(r0[N * 2], _r0, 2);
-                        _r0 = vsetq_lane_f16(r0[N * 3], _r0, 3);
-                        _r0 = vsetq_lane_f16(r0[N * 4], _r0, 4);
-                        _r0 = vsetq_lane_f16(r0[N * 5], _r0, 5);
-                        _r0 = vsetq_lane_f16(r0[N * 6], _r0, 6);
-                        _r0 = vsetq_lane_f16(r0[N * 7], _r0, 7);
-                        r0 += dilation_w;
-                    }
-
-                    float16x8_t _w0 = vld1q_f16(kptr);
-                    _sum = vfmaq_f16(_sum, _r0, _w0);
-
-                    kptr += 8;
-                }
-            }
-            for (; q + 3 < inh; q += 4)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q / elempack) + j * stride_w * elempack;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    float16x4_t _r0;
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1_f16(r0);
-                        r0 += dilation_w * 4;
-                    }
-                    else // if (elempack == 1)
-                    {
-                        _r0 = float16x4_t();
-                        _r0 = vset_lane_f16(r0[0], _r0, 0);
-                        _r0 = vset_lane_f16(r0[N], _r0, 1);
-                        _r0 = vset_lane_f16(r0[N * 2], _r0, 2);
-                        _r0 = vset_lane_f16(r0[N * 3], _r0, 3);
-                        r0 += dilation_w;
-                    }
-
-                    float16x4_t _w = vld1_f16(kptr);
-                    _sum = vcombine_f16(vfma_f16(vget_low_f16(_sum), _r0, _w), vget_high_f16(_sum));
-
-                    kptr += 4;
-                }
-            }
-            for (; q + 1 < inh; q += 2)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    __fp16 val0;
-                    __fp16 val1;
-                    // if (elempack == 1)
-                    {
-                        val0 = r0[0];
-                        val1 = r0[N];
-                        r0 += dilation_w;
-                    }
-
-                    sum += val0 * kptr[0];
-                    sum += val1 * kptr[1];
-
-                    kptr += 2;
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const __fp16* r0 = bottom_blob.row<const __fp16>(q) + j * stride_w;
-
-                for (int k = 0; k < kernel_w; k++)
-                {
-                    __fp16 val;
-                    // if (elempack == 1)
-                    {
-                        val = r0[0];
-                        r0 += dilation_w;
-                    }
-
-                    sum += val * kptr[0];
-
-                    kptr += 1;
-                }
-            }
-
-            float16x4_t _ss = vadd_f16(vget_low_f16(_sum), vget_high_f16(_sum));
-            _ss = vpadd_f16(_ss, _ss);
-            _ss = vpadd_f16(_ss, _ss);
-            sum += vget_lane_f16(_ss, 0);
-
-            sum = activation_ss_f16(sum, activation_type, activation_params);
-
-            outptr[0] = sum;
-            outptr += 1;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_1x1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_1x1.h
deleted file mode 100644
index 727cf17f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_1x1.h
+++ /dev/null
@@ -1,2598 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    int nn_outch = 0;
-    int remain_outch_start = 0;
-
-#if __ARM_NEON && __aarch64__
-
-    nn_outch = outch >> 3;
-    remain_outch_start = nn_outch << 3;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 8;
-
-        Mat out0 = top_blob.channel(p);
-        Mat out1 = top_blob.channel(p + 1);
-        Mat out2 = top_blob.channel(p + 2);
-        Mat out3 = top_blob.channel(p + 3);
-        Mat out4 = top_blob.channel(p + 4);
-        Mat out5 = top_blob.channel(p + 5);
-        Mat out6 = top_blob.channel(p + 6);
-        Mat out7 = top_blob.channel(p + 7);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        const float bias1 = bias ? bias[p + 1] : 0.f;
-        const float bias2 = bias ? bias[p + 2] : 0.f;
-        const float bias3 = bias ? bias[p + 3] : 0.f;
-        const float bias4 = bias ? bias[p + 4] : 0.f;
-        const float bias5 = bias ? bias[p + 5] : 0.f;
-        const float bias6 = bias ? bias[p + 6] : 0.f;
-        const float bias7 = bias ? bias[p + 7] : 0.f;
-
-        out0.fill(bias0);
-        out1.fill(bias1);
-        out2.fill(bias2);
-        out3.fill(bias3);
-        out4.fill(bias4);
-        out5.fill(bias5);
-        out6.fill(bias6);
-        out7.fill(bias7);
-
-        int q = 0;
-
-        for (; q + 7 < inch; q += 8)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-            float* outptr2 = out2;
-            float* outptr3 = out3;
-            float* outptr4 = out4;
-            float* outptr5 = out5;
-            float* outptr6 = out6;
-            float* outptr7 = out7;
-
-            const float* img0 = bottom_blob.channel(q);
-            const float* img1 = bottom_blob.channel(q + 1);
-            const float* img2 = bottom_blob.channel(q + 2);
-            const float* img3 = bottom_blob.channel(q + 3);
-            const float* img4 = bottom_blob.channel(q + 4);
-            const float* img5 = bottom_blob.channel(q + 5);
-            const float* img6 = bottom_blob.channel(q + 6);
-            const float* img7 = bottom_blob.channel(q + 7);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float* kernel1 = kernel + (p + 1) * inch + q;
-            const float* kernel2 = kernel + (p + 2) * inch + q;
-            const float* kernel3 = kernel + (p + 3) * inch + q;
-            const float* kernel4 = kernel + (p + 4) * inch + q;
-            const float* kernel5 = kernel + (p + 5) * inch + q;
-            const float* kernel6 = kernel + (p + 6) * inch + q;
-            const float* kernel7 = kernel + (p + 7) * inch + q;
-
-            const float* r0 = img0;
-            const float* r1 = img1;
-            const float* r2 = img2;
-            const float* r3 = img3;
-            const float* r4 = img4;
-            const float* r5 = img5;
-            const float* r6 = img6;
-            const float* r7 = img7;
-
-            int size = outw * outh;
-
-            int nn = size >> 2;
-            int remain = size & 3;
-
-            float32x4_t _k0 = vld1q_f32(kernel0);
-            float32x4_t _k1 = vld1q_f32(kernel1);
-            float32x4_t _k2 = vld1q_f32(kernel2);
-            float32x4_t _k3 = vld1q_f32(kernel3);
-            float32x4_t _k4 = vld1q_f32(kernel4);
-            float32x4_t _k5 = vld1q_f32(kernel5);
-            float32x4_t _k6 = vld1q_f32(kernel6);
-            float32x4_t _k7 = vld1q_f32(kernel7);
-
-            float32x4_t _k0n = vld1q_f32(kernel0 + 4);
-            float32x4_t _k1n = vld1q_f32(kernel1 + 4);
-            float32x4_t _k2n = vld1q_f32(kernel2 + 4);
-            float32x4_t _k3n = vld1q_f32(kernel3 + 4);
-            float32x4_t _k4n = vld1q_f32(kernel4 + 4);
-            float32x4_t _k5n = vld1q_f32(kernel5 + 4);
-            float32x4_t _k6n = vld1q_f32(kernel6 + 4);
-            float32x4_t _k7n = vld1q_f32(kernel7 + 4);
-
-#ifdef __clang__
-            // gcc reject over 30 oprands :(
-            if (nn > 0)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%9, #128]       \n"
-                    "ld1    {v17.4s}, [%9], #16         \n"
-
-                    "prfm   pldl1keep, [%1, #128]       \n"
-                    "ld1    {v18.4s}, [%1]              \n"
-
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "ld1    {v19.4s}, [%2]              \n"
-
-                    "0:                                 \n"
-
-                    "fmla   v18.4s, v17.4s, %34.s[0]    \n"
-
-                    "prfm   pldl1keep, [%3, #128]       \n"
-                    "ld1    {v20.4s}, [%3]              \n"
-
-                    "fmla   v19.4s, v17.4s, %35.s[0]    \n"
-
-                    "prfm   pldl1keep, [%4, #128]       \n"
-                    "ld1    {v21.4s}, [%4]              \n"
-
-                    "fmla   v20.4s, v17.4s, %36.s[0]    \n"
-
-                    "prfm   pldl1keep, [%5, #128]       \n"
-                    "ld1    {v22.4s}, [%5]              \n"
-
-                    "fmla   v21.4s, v17.4s, %37.s[0]    \n"
-
-                    "prfm   pldl1keep, [%6, #128]       \n"
-                    "ld1    {v23.4s}, [%6]              \n"
-
-                    "fmla   v22.4s, v17.4s, %38.s[0]    \n"
-
-                    "prfm   pldl1keep, [%10, #128]      \n"
-                    "ld1    {v16.4s}, [%10], #16        \n"
-
-                    "fmla   v23.4s, v17.4s, %39.s[0]    \n"
-
-                    "prfm   pldl1keep, [%7, #128]       \n"
-                    "ld1    {v24.4s}, [%7]              \n"
-
-                    "fmla   v18.4s, v16.4s, %34.s[1]    \n"
-                    "fmla   v19.4s, v16.4s, %35.s[1]    \n"
-
-                    "prfm   pldl1keep, [%8, #128]       \n"
-                    "ld1    {v25.4s}, [%8]              \n"
-
-                    "fmla   v24.4s, v17.4s, %40.s[0]    \n"
-                    "fmla   v25.4s, v17.4s, %41.s[0]    \n"
-
-                    "fmla   v20.4s, v16.4s, %36.s[1]    \n"
-                    "fmla   v21.4s, v16.4s, %37.s[1]    \n"
-
-                    "prfm   pldl1keep, [%11, #128]      \n"
-                    "ld1    {v17.4s}, [%11], #16        \n"
-
-                    "fmla   v22.4s, v16.4s, %38.s[1]    \n"
-                    "fmla   v23.4s, v16.4s, %39.s[1]    \n"
-
-                    "fmla   v18.4s, v17.4s, %34.s[2]    \n"
-                    "fmla   v19.4s, v17.4s, %35.s[2]    \n"
-
-                    "fmla   v24.4s, v16.4s, %40.s[1]    \n"
-                    "fmla   v25.4s, v16.4s, %41.s[1]    \n"
-
-                    "fmla   v20.4s, v17.4s, %36.s[2]    \n"
-                    "fmla   v21.4s, v17.4s, %37.s[2]    \n"
-
-                    "prfm   pldl1keep, [%12, #128]      \n"
-                    "ld1    {v16.4s}, [%12], #16        \n"
-
-                    "fmla   v22.4s, v17.4s, %38.s[2]    \n"
-                    "fmla   v23.4s, v17.4s, %39.s[2]    \n"
-
-                    "fmla   v18.4s, v16.4s, %34.s[3]    \n"
-                    "fmla   v19.4s, v16.4s, %35.s[3]    \n"
-
-                    "fmla   v24.4s, v17.4s, %40.s[2]    \n"
-                    "fmla   v25.4s, v17.4s, %41.s[2]    \n"
-
-                    "fmla   v20.4s, v16.4s, %36.s[3]    \n"
-                    "fmla   v21.4s, v16.4s, %37.s[3]    \n"
-
-                    "prfm   pldl1keep, [%13, #128]      \n"
-                    "ld1    {v17.4s}, [%13], #16        \n"
-
-                    "fmla   v22.4s, v16.4s, %38.s[3]    \n"
-                    "fmla   v23.4s, v16.4s, %39.s[3]    \n"
-
-                    "fmla   v18.4s, v17.4s, %42.s[0]    \n"
-                    "fmla   v19.4s, v17.4s, %43.s[0]    \n"
-
-                    "fmla   v24.4s, v16.4s, %40.s[3]    \n"
-                    "fmla   v25.4s, v16.4s, %41.s[3]    \n"
-
-                    "fmla   v20.4s, v17.4s, %44.s[0]    \n"
-                    "fmla   v21.4s, v17.4s, %45.s[0]    \n"
-
-                    "prfm   pldl1keep, [%14, #128]      \n"
-                    "ld1    {v16.4s}, [%14], #16        \n"
-
-                    "fmla   v22.4s, v17.4s, %46.s[0]    \n"
-                    "fmla   v23.4s, v17.4s, %47.s[0]    \n"
-
-                    "fmla   v18.4s, v16.4s, %42.s[1]    \n"
-                    "fmla   v19.4s, v16.4s, %43.s[1]    \n"
-
-                    "fmla   v24.4s, v17.4s, %48.s[0]    \n"
-                    "fmla   v25.4s, v17.4s, %49.s[0]    \n"
-
-                    "fmla   v20.4s, v16.4s, %44.s[1]    \n"
-                    "fmla   v21.4s, v16.4s, %45.s[1]    \n"
-
-                    "prfm   pldl1keep, [%15, #128]      \n"
-                    "ld1    {v17.4s}, [%15], #16        \n"
-
-                    "fmla   v22.4s, v16.4s, %46.s[1]    \n"
-                    "fmla   v23.4s, v16.4s, %47.s[1]    \n"
-
-                    "fmla   v18.4s, v17.4s, %42.s[2]    \n"
-                    "fmla   v19.4s, v17.4s, %43.s[2]    \n"
-
-                    "fmla   v24.4s, v16.4s, %48.s[1]    \n"
-                    "fmla   v25.4s, v16.4s, %49.s[1]    \n"
-
-                    "fmla   v20.4s, v17.4s, %44.s[2]    \n"
-                    "fmla   v21.4s, v17.4s, %45.s[2]    \n"
-
-                    "prfm   pldl1keep, [%16, #128]      \n"
-                    "ld1    {v16.4s}, [%16], #16        \n"
-
-                    "fmla   v22.4s, v17.4s, %46.s[2]    \n"
-                    "fmla   v23.4s, v17.4s, %47.s[2]    \n"
-
-                    "fmla   v18.4s, v16.4s, %42.s[3]    \n"
-                    "fmla   v19.4s, v16.4s, %43.s[3]    \n"
-
-                    "fmla   v24.4s, v17.4s, %48.s[2]    \n"
-                    "fmla   v25.4s, v17.4s, %49.s[2]    \n"
-
-                    "fmla   v20.4s, v16.4s, %44.s[3]    \n"
-                    "fmla   v21.4s, v16.4s, %45.s[3]    \n"
-
-                    "st1    {v18.4s}, [%1], #16         \n"
-
-                    "fmla   v22.4s, v16.4s, %46.s[3]    \n"
-
-                    "st1    {v19.4s}, [%2], #16         \n"
-
-                    "fmla   v23.4s, v16.4s, %47.s[3]    \n"
-
-                    "st1    {v20.4s}, [%3], #16         \n"
-
-                    "prfm   pldl1keep, [%9, #128]       \n"
-                    "ld1    {v17.4s}, [%9], #16         \n"
-
-                    "fmla   v24.4s, v16.4s, %48.s[3]    \n"
-
-                    "st1    {v21.4s}, [%4], #16         \n"
-
-                    "fmla   v25.4s, v16.4s, %49.s[3]    \n"
-
-                    "st1    {v22.4s}, [%5], #16         \n"
-
-                    "prfm   pldl1keep, [%1, #128]       \n"
-                    "ld1    {v18.4s}, [%1]              \n"
-
-                    "st1    {v23.4s}, [%6], #16         \n"
-
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "ld1    {v19.4s}, [%2]              \n"
-
-                    "st1    {v24.4s}, [%7], #16         \n"
-
-                    "subs   %w0, %w0, #1                \n"
-
-                    "st1    {v25.4s}, [%8], #16         \n"
-
-                    "bne    0b                          \n"
-                    "sub    %9, %9, #16                 \n"
-                    : "=r"(nn),      // %0
-                    "=r"(outptr0), // %1
-                    "=r"(outptr1), // %2
-                    "=r"(outptr2), // %3
-                    "=r"(outptr3), // %4
-                    "=r"(outptr4), // %5
-                    "=r"(outptr5), // %6
-                    "=r"(outptr6), // %7
-                    "=r"(outptr7), // %8
-                    "=r"(r0),      // %9
-                    "=r"(r1),      // %10
-                    "=r"(r2),      // %11
-                    "=r"(r3),      // %12
-                    "=r"(r4),      // %13
-                    "=r"(r5),      // %14
-                    "=r"(r6),      // %15
-                    "=r"(r7)       // %16
-                    : "0"(nn),
-                    "1"(outptr0),
-                    "2"(outptr1),
-                    "3"(outptr2),
-                    "4"(outptr3),
-                    "5"(outptr4),
-                    "6"(outptr5),
-                    "7"(outptr6),
-                    "8"(outptr7),
-                    "9"(r0),
-                    "10"(r1),
-                    "11"(r2),
-                    "12"(r3),
-                    "13"(r4),
-                    "14"(r5),
-                    "15"(r6),
-                    "16"(r7),
-                    "w"(_k0),                                                                            // %34
-                    "w"(_k1),                                                                            // %35
-                    "w"(_k2),                                                                            // %36
-                    "w"(_k3),                                                                            // %37
-                    "w"(_k4),                                                                            // %38
-                    "w"(_k5),                                                                            // %39
-                    "w"(_k6),                                                                            // %40
-                    "w"(_k7),                                                                            // %41
-                    "w"(_k0n),                                                                           // %42
-                    "w"(_k1n),                                                                           // %43
-                    "w"(_k2n),                                                                           // %44
-                    "w"(_k3n),                                                                           // %45
-                    "w"(_k4n),                                                                           // %46
-                    "w"(_k5n),                                                                           // %47
-                    "w"(_k6n),                                                                           // %48
-                    "w"(_k7n)                                                                            // %49
-                    : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25" //, "v26", "v27", "v28", "v29", "v30", "v31"
-                );
-            }
-#else
-            for (; nn > 0; nn--)
-            {
-                float32x4_t _p = vld1q_f32(r0);
-
-                float32x4_t _out0p = vld1q_f32(outptr0);
-                float32x4_t _out1p = vld1q_f32(outptr1);
-                float32x4_t _out2p = vld1q_f32(outptr2);
-                float32x4_t _out3p = vld1q_f32(outptr3);
-                float32x4_t _out4p = vld1q_f32(outptr4);
-                float32x4_t _out5p = vld1q_f32(outptr5);
-                float32x4_t _out6p = vld1q_f32(outptr6);
-                float32x4_t _out7p = vld1q_f32(outptr7);
-
-                _out0p = vfmaq_laneq_f32(_out0p, _p, _k0, 0);
-                _out1p = vfmaq_laneq_f32(_out1p, _p, _k1, 0);
-                _out2p = vfmaq_laneq_f32(_out2p, _p, _k2, 0);
-                _out3p = vfmaq_laneq_f32(_out3p, _p, _k3, 0);
-                _out4p = vfmaq_laneq_f32(_out4p, _p, _k4, 0);
-                _out5p = vfmaq_laneq_f32(_out5p, _p, _k5, 0);
-                _out6p = vfmaq_laneq_f32(_out6p, _p, _k6, 0);
-                _out7p = vfmaq_laneq_f32(_out7p, _p, _k7, 0);
-
-                float32x4_t _p1 = vld1q_f32(r1);
-
-                _out0p = vfmaq_laneq_f32(_out0p, _p1, _k0, 1);
-                _out1p = vfmaq_laneq_f32(_out1p, _p1, _k1, 1);
-                _out2p = vfmaq_laneq_f32(_out2p, _p1, _k2, 1);
-                _out3p = vfmaq_laneq_f32(_out3p, _p1, _k3, 1);
-                _out4p = vfmaq_laneq_f32(_out4p, _p1, _k4, 1);
-                _out5p = vfmaq_laneq_f32(_out5p, _p1, _k5, 1);
-                _out6p = vfmaq_laneq_f32(_out6p, _p1, _k6, 1);
-                _out7p = vfmaq_laneq_f32(_out7p, _p1, _k7, 1);
-
-                float32x4_t _p2 = vld1q_f32(r2);
-
-                _out0p = vfmaq_laneq_f32(_out0p, _p2, _k0, 2);
-                _out1p = vfmaq_laneq_f32(_out1p, _p2, _k1, 2);
-                _out2p = vfmaq_laneq_f32(_out2p, _p2, _k2, 2);
-                _out3p = vfmaq_laneq_f32(_out3p, _p2, _k3, 2);
-                _out4p = vfmaq_laneq_f32(_out4p, _p2, _k4, 2);
-                _out5p = vfmaq_laneq_f32(_out5p, _p2, _k5, 2);
-                _out6p = vfmaq_laneq_f32(_out6p, _p2, _k6, 2);
-                _out7p = vfmaq_laneq_f32(_out7p, _p2, _k7, 2);
-
-                float32x4_t _p3 = vld1q_f32(r3);
-
-                _out0p = vfmaq_laneq_f32(_out0p, _p3, _k0, 3);
-                _out1p = vfmaq_laneq_f32(_out1p, _p3, _k1, 3);
-                _out2p = vfmaq_laneq_f32(_out2p, _p3, _k2, 3);
-                _out3p = vfmaq_laneq_f32(_out3p, _p3, _k3, 3);
-                _out4p = vfmaq_laneq_f32(_out4p, _p3, _k4, 3);
-                _out5p = vfmaq_laneq_f32(_out5p, _p3, _k5, 3);
-                _out6p = vfmaq_laneq_f32(_out6p, _p3, _k6, 3);
-                _out7p = vfmaq_laneq_f32(_out7p, _p3, _k7, 3);
-
-                float32x4_t _p4 = vld1q_f32(r4);
-
-                _out0p = vfmaq_laneq_f32(_out0p, _p4, _k0n, 0);
-                _out1p = vfmaq_laneq_f32(_out1p, _p4, _k1n, 0);
-                _out2p = vfmaq_laneq_f32(_out2p, _p4, _k2n, 0);
-                _out3p = vfmaq_laneq_f32(_out3p, _p4, _k3n, 0);
-                _out4p = vfmaq_laneq_f32(_out4p, _p4, _k4n, 0);
-                _out5p = vfmaq_laneq_f32(_out5p, _p4, _k5n, 0);
-                _out6p = vfmaq_laneq_f32(_out6p, _p4, _k6n, 0);
-                _out7p = vfmaq_laneq_f32(_out7p, _p4, _k7n, 0);
-
-                float32x4_t _p5 = vld1q_f32(r5);
-
-                _out0p = vfmaq_laneq_f32(_out0p, _p5, _k0n, 1);
-                _out1p = vfmaq_laneq_f32(_out1p, _p5, _k1n, 1);
-                _out2p = vfmaq_laneq_f32(_out2p, _p5, _k2n, 1);
-                _out3p = vfmaq_laneq_f32(_out3p, _p5, _k3n, 1);
-                _out4p = vfmaq_laneq_f32(_out4p, _p5, _k4n, 1);
-                _out5p = vfmaq_laneq_f32(_out5p, _p5, _k5n, 1);
-                _out6p = vfmaq_laneq_f32(_out6p, _p5, _k6n, 1);
-                _out7p = vfmaq_laneq_f32(_out7p, _p5, _k7n, 1);
-
-                float32x4_t _p6 = vld1q_f32(r6);
-
-                _out0p = vfmaq_laneq_f32(_out0p, _p6, _k0n, 2);
-                _out1p = vfmaq_laneq_f32(_out1p, _p6, _k1n, 2);
-                _out2p = vfmaq_laneq_f32(_out2p, _p6, _k2n, 2);
-                _out3p = vfmaq_laneq_f32(_out3p, _p6, _k3n, 2);
-                _out4p = vfmaq_laneq_f32(_out4p, _p6, _k4n, 2);
-                _out5p = vfmaq_laneq_f32(_out5p, _p6, _k5n, 2);
-                _out6p = vfmaq_laneq_f32(_out6p, _p6, _k6n, 2);
-                _out7p = vfmaq_laneq_f32(_out7p, _p6, _k7n, 2);
-
-                float32x4_t _p7 = vld1q_f32(r7);
-
-                _out0p = vfmaq_laneq_f32(_out0p, _p7, _k0n, 3);
-                _out1p = vfmaq_laneq_f32(_out1p, _p7, _k1n, 3);
-                _out2p = vfmaq_laneq_f32(_out2p, _p7, _k2n, 3);
-                _out3p = vfmaq_laneq_f32(_out3p, _p7, _k3n, 3);
-                _out4p = vfmaq_laneq_f32(_out4p, _p7, _k4n, 3);
-                _out5p = vfmaq_laneq_f32(_out5p, _p7, _k5n, 3);
-                _out6p = vfmaq_laneq_f32(_out6p, _p7, _k6n, 3);
-                _out7p = vfmaq_laneq_f32(_out7p, _p7, _k7n, 3);
-
-                vst1q_f32(outptr0, _out0p);
-                vst1q_f32(outptr1, _out1p);
-                vst1q_f32(outptr2, _out2p);
-                vst1q_f32(outptr3, _out3p);
-                vst1q_f32(outptr4, _out4p);
-                vst1q_f32(outptr5, _out5p);
-                vst1q_f32(outptr6, _out6p);
-                vst1q_f32(outptr7, _out7p);
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                r3 += 4;
-                r4 += 4;
-                r5 += 4;
-                r6 += 4;
-                r7 += 4;
-                outptr0 += 4;
-                outptr1 += 4;
-                outptr2 += 4;
-                outptr3 += 4;
-                outptr4 += 4;
-                outptr5 += 4;
-                outptr6 += 4;
-                outptr7 += 4;
-            }
-#endif
-            for (; remain > 0; remain--)
-            {
-                // TODO neon optimize
-                float sum0 = *r0 * kernel0[0] + *r1 * kernel0[1] + *r2 * kernel0[2] + *r3 * kernel0[3] + *r4 * kernel0[4] + *r5 * kernel0[5] + *r6 * kernel0[6] + *r7 * kernel0[7];
-                float sum1 = *r0 * kernel1[0] + *r1 * kernel1[1] + *r2 * kernel1[2] + *r3 * kernel1[3] + *r4 * kernel1[4] + *r5 * kernel1[5] + *r6 * kernel1[6] + *r7 * kernel1[7];
-                float sum2 = *r0 * kernel2[0] + *r1 * kernel2[1] + *r2 * kernel2[2] + *r3 * kernel2[3] + *r4 * kernel2[4] + *r5 * kernel2[5] + *r6 * kernel2[6] + *r7 * kernel2[7];
-                float sum3 = *r0 * kernel3[0] + *r1 * kernel3[1] + *r2 * kernel3[2] + *r3 * kernel3[3] + *r4 * kernel3[4] + *r5 * kernel3[5] + *r6 * kernel3[6] + *r7 * kernel3[7];
-                float sum4 = *r0 * kernel4[0] + *r1 * kernel4[1] + *r2 * kernel4[2] + *r3 * kernel4[3] + *r4 * kernel4[4] + *r5 * kernel4[5] + *r6 * kernel4[6] + *r7 * kernel4[7];
-                float sum5 = *r0 * kernel5[0] + *r1 * kernel5[1] + *r2 * kernel5[2] + *r3 * kernel5[3] + *r4 * kernel5[4] + *r5 * kernel5[5] + *r6 * kernel5[6] + *r7 * kernel5[7];
-                float sum6 = *r0 * kernel6[0] + *r1 * kernel6[1] + *r2 * kernel6[2] + *r3 * kernel6[3] + *r4 * kernel6[4] + *r5 * kernel6[5] + *r6 * kernel6[6] + *r7 * kernel6[7];
-                float sum7 = *r0 * kernel7[0] + *r1 * kernel7[1] + *r2 * kernel7[2] + *r3 * kernel7[3] + *r4 * kernel7[4] + *r5 * kernel7[5] + *r6 * kernel7[6] + *r7 * kernel7[7];
-
-                *outptr0 += sum0;
-                *outptr1 += sum1;
-                *outptr2 += sum2;
-                *outptr3 += sum3;
-                *outptr4 += sum4;
-                *outptr5 += sum5;
-                *outptr6 += sum6;
-                *outptr7 += sum7;
-
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                r4++;
-                r5++;
-                r6++;
-                r7++;
-                outptr0++;
-                outptr1++;
-                outptr2++;
-                outptr3++;
-                outptr4++;
-                outptr5++;
-                outptr6++;
-                outptr7++;
-            }
-        }
-
-        for (; q < inch; q++)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-            float* outptr2 = out2;
-            float* outptr3 = out3;
-            float* outptr4 = out4;
-            float* outptr5 = out5;
-            float* outptr6 = out6;
-            float* outptr7 = out7;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float* kernel1 = kernel + (p + 1) * inch + q;
-            const float* kernel2 = kernel + (p + 2) * inch + q;
-            const float* kernel3 = kernel + (p + 3) * inch + q;
-            const float* kernel4 = kernel + (p + 4) * inch + q;
-            const float* kernel5 = kernel + (p + 5) * inch + q;
-            const float* kernel6 = kernel + (p + 6) * inch + q;
-            const float* kernel7 = kernel + (p + 7) * inch + q;
-
-            const float k0 = kernel0[0];
-            const float k1 = kernel1[0];
-            const float k2 = kernel2[0];
-            const float k3 = kernel3[0];
-            const float k4 = kernel4[0];
-            const float k5 = kernel5[0];
-            const float k6 = kernel6[0];
-            const float k7 = kernel7[0];
-
-            const float* r0 = img0;
-
-            int size = outw * outh;
-
-            int nn = size >> 2;
-            int remain = size & 3;
-
-            float32x4_t _k0 = vdupq_n_f32(k0);
-            float32x4_t _k1 = vdupq_n_f32(k1);
-            float32x4_t _k2 = vdupq_n_f32(k2);
-            float32x4_t _k3 = vdupq_n_f32(k3);
-            float32x4_t _k4 = vdupq_n_f32(k4);
-            float32x4_t _k5 = vdupq_n_f32(k5);
-            float32x4_t _k6 = vdupq_n_f32(k6);
-            float32x4_t _k7 = vdupq_n_f32(k7);
-
-            for (; nn > 0; nn--)
-            {
-                float32x4_t _p = vld1q_f32(r0);
-
-                float32x4_t _out0p = vld1q_f32(outptr0);
-                float32x4_t _out1p = vld1q_f32(outptr1);
-                float32x4_t _out2p = vld1q_f32(outptr2);
-                float32x4_t _out3p = vld1q_f32(outptr3);
-                float32x4_t _out4p = vld1q_f32(outptr4);
-                float32x4_t _out5p = vld1q_f32(outptr5);
-                float32x4_t _out6p = vld1q_f32(outptr6);
-                float32x4_t _out7p = vld1q_f32(outptr7);
-
-                _out0p = vfmaq_f32(_out0p, _p, _k0);
-                _out1p = vfmaq_f32(_out1p, _p, _k1);
-                _out2p = vfmaq_f32(_out2p, _p, _k2);
-                _out3p = vfmaq_f32(_out3p, _p, _k3);
-                _out4p = vfmaq_f32(_out4p, _p, _k4);
-                _out5p = vfmaq_f32(_out5p, _p, _k5);
-                _out6p = vfmaq_f32(_out6p, _p, _k6);
-                _out7p = vfmaq_f32(_out7p, _p, _k7);
-
-                vst1q_f32(outptr0, _out0p);
-                vst1q_f32(outptr1, _out1p);
-                vst1q_f32(outptr2, _out2p);
-                vst1q_f32(outptr3, _out3p);
-                vst1q_f32(outptr4, _out4p);
-                vst1q_f32(outptr5, _out5p);
-                vst1q_f32(outptr6, _out6p);
-                vst1q_f32(outptr7, _out7p);
-
-                r0 += 4;
-                outptr0 += 4;
-                outptr1 += 4;
-                outptr2 += 4;
-                outptr3 += 4;
-                outptr4 += 4;
-                outptr5 += 4;
-                outptr6 += 4;
-                outptr7 += 4;
-            }
-            for (; remain > 0; remain--)
-            {
-                // TODO neon optimize
-                float sum0 = *r0 * k0;
-                float sum1 = *r0 * k1;
-                float sum2 = *r0 * k2;
-                float sum3 = *r0 * k3;
-                float sum4 = *r0 * k4;
-                float sum5 = *r0 * k5;
-                float sum6 = *r0 * k6;
-                float sum7 = *r0 * k7;
-
-                *outptr0 += sum0;
-                *outptr1 += sum1;
-                *outptr2 += sum2;
-                *outptr3 += sum3;
-                *outptr4 += sum4;
-                *outptr5 += sum5;
-                *outptr6 += sum6;
-                *outptr7 += sum7;
-
-                r0++;
-                outptr0++;
-                outptr1++;
-                outptr2++;
-                outptr3++;
-                outptr4++;
-                outptr5++;
-                outptr6++;
-                outptr7++;
-            }
-        }
-    }
-
-#else
-
-    nn_outch = outch / 6;
-    remain_outch_start = nn_outch * 6;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 6;
-
-        Mat out0 = top_blob.channel(p);
-        Mat out1 = top_blob.channel(p + 1);
-        Mat out2 = top_blob.channel(p + 2);
-        Mat out3 = top_blob.channel(p + 3);
-        Mat out4 = top_blob.channel(p + 4);
-        Mat out5 = top_blob.channel(p + 5);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        const float bias1 = bias ? bias[p + 1] : 0.f;
-        const float bias2 = bias ? bias[p + 2] : 0.f;
-        const float bias3 = bias ? bias[p + 3] : 0.f;
-        const float bias4 = bias ? bias[p + 4] : 0.f;
-        const float bias5 = bias ? bias[p + 5] : 0.f;
-
-        out0.fill(bias0);
-        out1.fill(bias1);
-        out2.fill(bias2);
-        out3.fill(bias3);
-        out4.fill(bias4);
-        out5.fill(bias5);
-
-        int q = 0;
-
-        for (; q + 3 < inch; q += 4)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-            float* outptr2 = out2;
-            float* outptr3 = out3;
-            float* outptr4 = out4;
-            float* outptr5 = out5;
-
-            const float* img0 = bottom_blob.channel(q);
-            const float* img1 = bottom_blob.channel(q + 1);
-            const float* img2 = bottom_blob.channel(q + 2);
-            const float* img3 = bottom_blob.channel(q + 3);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float* kernel1 = kernel + (p + 1) * inch + q;
-            const float* kernel2 = kernel + (p + 2) * inch + q;
-            const float* kernel3 = kernel + (p + 3) * inch + q;
-            const float* kernel4 = kernel + (p + 4) * inch + q;
-            const float* kernel5 = kernel + (p + 5) * inch + q;
-
-            const float* r0 = img0;
-            const float* r1 = img1;
-            const float* r2 = img2;
-            const float* r3 = img3;
-
-            int size = outw * outh;
-
-#if __ARM_NEON
-            int nn = size >> 2;
-            int remain = size & 3;
-#else
-            int remain = size;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-            float32x4_t _k0 = vld1q_f32(kernel0);
-            float32x4_t _k1 = vld1q_f32(kernel1);
-            float32x4_t _k2 = vld1q_f32(kernel2);
-            float32x4_t _k3 = vld1q_f32(kernel3);
-            float32x4_t _k4 = vld1q_f32(kernel4);
-            float32x4_t _k5 = vld1q_f32(kernel5);
-
-            for (; nn > 0; nn--)
-            {
-                asm volatile(
-                    "pld        [%6, #128]              \n"
-                    "vld1.f32   {d24-d25}, [%6 :128]!   \n" // q12 = r0
-
-                    "pld        [%0, #128]              \n"
-                    "vld1.f32   {d12-d13}, [%0 :128]    \n" // q6 = outptr0
-
-                    "pld        [%1, #128]              \n"
-                    "vld1.f32   {d14-d15}, [%1 :128]    \n" // q7 = outptr1
-
-                    "vmla.f32   q6, q12, %e20[0]        \n"
-
-                    "pld        [%2, #128]              \n"
-                    "vld1.f32   {d16-d17}, [%2 :128]    \n" // q8 = outptr2
-
-                    "vmla.f32   q7, q12, %e21[0]        \n"
-
-                    "pld        [%3, #128]              \n"
-                    "vld1.f32   {d18-d19}, [%3 :128]    \n" // q9 = outptr3
-
-                    "vmla.f32   q8, q12, %e22[0]        \n"
-
-                    "pld        [%7, #128]              \n"
-                    "vld1.f32   {d26-d27}, [%7 :128]!   \n" // q13 = r1
-
-                    "vmla.f32   q9, q12, %e23[0]        \n"
-
-                    "pld        [%4, #128]              \n"
-                    "vld1.f32   {d20-d21}, [%4 :128]    \n" // q10 = outptr4
-
-                    "vmla.f32   q6, q13, %e20[1]        \n"
-                    "vmla.f32   q7, q13, %e21[1]        \n"
-
-                    "pld        [%5, #128]              \n"
-                    "vld1.f32   {d22-d23}, [%5 :128]    \n" // q11 = outptr5
-
-                    "vmla.f32   q10, q12, %e24[0]       \n"
-                    "vmla.f32   q11, q12, %e25[0]       \n"
-
-                    "vmla.f32   q8, q13, %e22[1]        \n"
-                    "vmla.f32   q9, q13, %e23[1]        \n"
-
-                    "pld        [%8, #128]              \n"
-                    "vld1.f32   {d28-d29}, [%8 :128]!   \n" // q14 = r2
-
-                    "vmla.f32   q10, q13, %e24[1]       \n"
-                    "vmla.f32   q11, q13, %e25[1]       \n"
-
-                    "vmla.f32   q6, q14, %f20[0]        \n"
-                    "vmla.f32   q7, q14, %f21[0]        \n"
-                    "vmla.f32   q8, q14, %f22[0]        \n"
-                    "vmla.f32   q9, q14, %f23[0]        \n"
-
-                    "pld        [%9, #128]             \n"
-                    "vld1.f32   {d30-d31}, [%9 :128]!  \n" // q15 = r3
-
-                    "vmla.f32   q10, q14, %f24[0]       \n"
-                    "vmla.f32   q11, q14, %f25[0]       \n"
-
-                    "vmla.f32   q6, q15, %f20[1]        \n"
-                    "vmla.f32   q7, q15, %f21[1]        \n"
-                    "vmla.f32   q8, q15, %f22[1]        \n"
-                    "vmla.f32   q9, q15, %f23[1]        \n"
-
-                    "vmla.f32   q10, q15, %f24[1]       \n"
-                    "vmla.f32   q11, q15, %f25[1]       \n"
-
-                    "vst1.f32   {d12-d13}, [%0 :128]!   \n"
-                    "vst1.f32   {d14-d15}, [%1 :128]!   \n"
-                    "vst1.f32   {d16-d17}, [%2 :128]!   \n"
-                    "vst1.f32   {d18-d19}, [%3 :128]!   \n"
-                    "vst1.f32   {d20-d21}, [%4 :128]!   \n"
-                    "vst1.f32   {d22-d23}, [%5 :128]!   \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(outptr2), // %2
-                    "=r"(outptr3), // %3
-                    "=r"(outptr4), // %4
-                    "=r"(outptr5), // %5
-                    "=r"(r0),      // %6
-                    "=r"(r1),      // %7
-                    "=r"(r2),      // %8
-                    "=r"(r3)       // %9
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(outptr2),
-                    "3"(outptr3),
-                    "4"(outptr4),
-                    "5"(outptr5),
-                    "6"(r0),
-                    "7"(r1),
-                    "8"(r2),
-                    "9"(r3),
-                    "w"(_k0), // %20
-                    "w"(_k1), // %21
-                    "w"(_k2), // %22
-                    "w"(_k3), // %23
-                    "w"(_k4), // %24
-                    "w"(_k5)  // %25
-                    : "memory", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-            }
-#endif // __ARM_NEON
-
-            for (; remain > 0; remain--)
-            {
-                // TODO neon optimize
-                float sum0 = *r0 * kernel0[0] + *r1 * kernel0[1] + *r2 * kernel0[2] + *r3 * kernel0[3];
-                float sum1 = *r0 * kernel1[0] + *r1 * kernel1[1] + *r2 * kernel1[2] + *r3 * kernel1[3];
-                float sum2 = *r0 * kernel2[0] + *r1 * kernel2[1] + *r2 * kernel2[2] + *r3 * kernel2[3];
-                float sum3 = *r0 * kernel3[0] + *r1 * kernel3[1] + *r2 * kernel3[2] + *r3 * kernel3[3];
-                float sum4 = *r0 * kernel4[0] + *r1 * kernel4[1] + *r2 * kernel4[2] + *r3 * kernel4[3];
-                float sum5 = *r0 * kernel5[0] + *r1 * kernel5[1] + *r2 * kernel5[2] + *r3 * kernel5[3];
-
-                *outptr0 += sum0;
-                *outptr1 += sum1;
-                *outptr2 += sum2;
-                *outptr3 += sum3;
-                *outptr4 += sum4;
-                *outptr5 += sum5;
-
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                outptr0++;
-                outptr1++;
-                outptr2++;
-                outptr3++;
-                outptr4++;
-                outptr5++;
-            }
-        }
-
-        for (; q < inch; q++)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-            float* outptr2 = out2;
-            float* outptr3 = out3;
-            float* outptr4 = out4;
-            float* outptr5 = out5;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float* kernel1 = kernel + (p + 1) * inch + q;
-            const float* kernel2 = kernel + (p + 2) * inch + q;
-            const float* kernel3 = kernel + (p + 3) * inch + q;
-            const float* kernel4 = kernel + (p + 4) * inch + q;
-            const float* kernel5 = kernel + (p + 5) * inch + q;
-
-            const float k0 = kernel0[0];
-            const float k1 = kernel1[0];
-            const float k2 = kernel2[0];
-            const float k3 = kernel3[0];
-            const float k4 = kernel4[0];
-            const float k5 = kernel5[0];
-
-            const float* r0 = img0;
-
-            int size = outw * outh;
-
-#if __ARM_NEON
-            int nn = size >> 2;
-            int remain = size & 3;
-#else
-            int remain = size;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-            float32x4_t _k0 = vdupq_n_f32(k0);
-            float32x4_t _k1 = vdupq_n_f32(k1);
-            float32x4_t _k2 = vdupq_n_f32(k2);
-            float32x4_t _k3 = vdupq_n_f32(k3);
-            float32x4_t _k4 = vdupq_n_f32(k4);
-            float32x4_t _k5 = vdupq_n_f32(k5);
-
-            if (nn > 0)
-            {
-                asm volatile(
-                    "pld        [%7, #128]              \n"
-                    "vld1.f32   {d24-d25}, [%7 :128]!   \n" // q12 = r0
-
-                    "pld        [%1, #128]              \n"
-                    "vld1.f32   {d12-d13}, [%1 :128]    \n" // q6 = outptr0
-
-                    "0:                                 \n"
-
-                    "pld        [%2, #128]              \n"
-                    "vld1.f32   {d14-d15}, [%2 :128]    \n" // q7 = outptr1
-
-                    "vmla.f32   q6, q12, %q16           \n"
-
-                    "pld        [%3, #128]              \n"
-                    "vld1.f32   {d16-d17}, [%3 :128]    \n" // q8 = outptr2
-
-                    "vmla.f32   q7, q12, %q17           \n"
-
-                    "pld        [%4, #128]              \n"
-                    "vld1.f32   {d18-d19}, [%4 :128]    \n" // q9 = outptr3
-
-                    "vmla.f32   q8, q12, %q18           \n"
-
-                    "pld        [%5, #128]              \n"
-                    "vld1.f32   {d20-d21}, [%5 :128]    \n" // q10 = outptr4
-
-                    "vmla.f32   q9, q12, %q19           \n"
-
-                    "pld        [%6, #128]              \n"
-                    "vld1.f32   {d22-d23}, [%6 :128]    \n" // q11 = outptr5
-
-                    "vmla.f32   q10, q12, %q20          \n"
-                    "vmla.f32   q11, q12, %q21          \n"
-
-                    "pld        [%7, #128]              \n"
-                    "vld1.f32   {d24-d25}, [%7 :128]!   \n" // q12 = r0
-
-                    "vst1.f32   {d12-d13}, [%1 :128]!   \n"
-                    "vst1.f32   {d14-d15}, [%2 :128]!   \n"
-
-                    "pld        [%1, #128]              \n"
-                    "vld1.f32   {d12-d13}, [%1 :128]    \n" // q6 = outptr0
-
-                    "vst1.f32   {d16-d17}, [%3 :128]!   \n"
-                    "vst1.f32   {d18-d19}, [%4 :128]!   \n"
-
-                    "subs       %0, #1                  \n"
-
-                    "vst1.f32   {d20-d21}, [%5 :128]!   \n"
-                    "vst1.f32   {d22-d23}, [%6 :128]!   \n"
-
-                    "bne        0b                      \n"
-
-                    "sub        %7, #16                 \n"
-
-                    : "=r"(nn),      // %0
-                    "=r"(outptr0), // %1
-                    "=r"(outptr1), // %2
-                    "=r"(outptr2), // %3
-                    "=r"(outptr3), // %4
-                    "=r"(outptr4), // %5
-                    "=r"(outptr5), // %6
-                    "=r"(r0)       // %7
-                    : "0"(nn),
-                    "1"(outptr0),
-                    "2"(outptr1),
-                    "3"(outptr2),
-                    "4"(outptr3),
-                    "5"(outptr4),
-                    "6"(outptr5),
-                    "7"(r0),
-                    "w"(_k0), // %16
-                    "w"(_k1), // %17
-                    "w"(_k2), // %18
-                    "w"(_k3), // %19
-                    "w"(_k4), // %20
-                    "w"(_k5)  // %21
-                    : "cc", "memory", "q6", "q7", "q8", "q9", "q10", "q11", "q12");
-            }
-#endif // __ARM_NEON
-
-            for (; remain > 0; remain--)
-            {
-                // TODO neon optimize
-                float sum0 = *r0 * k0;
-                float sum1 = *r0 * k1;
-                float sum2 = *r0 * k2;
-                float sum3 = *r0 * k3;
-                float sum4 = *r0 * k4;
-                float sum5 = *r0 * k5;
-
-                *outptr0 += sum0;
-                *outptr1 += sum1;
-                *outptr2 += sum2;
-                *outptr3 += sum3;
-                *outptr4 += sum4;
-                *outptr5 += sum5;
-
-                r0++;
-                outptr0++;
-                outptr1++;
-                outptr2++;
-                outptr3++;
-                outptr4++;
-                outptr5++;
-            }
-        }
-    }
-#endif // __ARM_NEON && __aarch64__
-
-    nn_outch = (outch - remain_outch_start) >> 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = remain_outch_start + pp * 4;
-
-        Mat out0 = top_blob.channel(p);
-        Mat out1 = top_blob.channel(p + 1);
-        Mat out2 = top_blob.channel(p + 2);
-        Mat out3 = top_blob.channel(p + 3);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        const float bias1 = bias ? bias[p + 1] : 0.f;
-        const float bias2 = bias ? bias[p + 2] : 0.f;
-        const float bias3 = bias ? bias[p + 3] : 0.f;
-
-        out0.fill(bias0);
-        out1.fill(bias1);
-        out2.fill(bias2);
-        out3.fill(bias3);
-
-        int q = 0;
-
-        for (; q + 3 < inch; q += 4)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-            float* outptr2 = out2;
-            float* outptr3 = out3;
-
-            const float* img0 = bottom_blob.channel(q);
-            const float* img1 = bottom_blob.channel(q + 1);
-            const float* img2 = bottom_blob.channel(q + 2);
-            const float* img3 = bottom_blob.channel(q + 3);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float* kernel1 = kernel + (p + 1) * inch + q;
-            const float* kernel2 = kernel + (p + 2) * inch + q;
-            const float* kernel3 = kernel + (p + 3) * inch + q;
-
-            const float* r0 = img0;
-            const float* r1 = img1;
-            const float* r2 = img2;
-            const float* r3 = img3;
-
-            int size = outw * outh;
-
-#if __ARM_NEON
-            int nn = size >> 3;
-            int remain = size & 7;
-#else
-            int remain = size;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-            float32x4_t _k0 = vld1q_f32(kernel0);
-            float32x4_t _k1 = vld1q_f32(kernel1);
-            float32x4_t _k2 = vld1q_f32(kernel2);
-            float32x4_t _k3 = vld1q_f32(kernel3);
-
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v6.4s, v7.4s}, [%5], #32   \n"
-
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v8.4s, v9.4s}, [%1]        \n"
-
-                    "0:                                 \n"
-
-                    "fmla   v8.4s, v6.4s, %18.s[0]      \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v10.4s, v11.4s}, [%2]      \n"
-
-                    "fmla   v9.4s, v7.4s, %18.s[0]      \n"
-
-                    "fmla   v10.4s, v6.4s, %19.s[0]     \n"
-
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v12.4s, v13.4s}, [%3]      \n"
-
-                    "fmla   v11.4s, v7.4s, %19.s[0]     \n"
-
-                    "fmla   v12.4s, v6.4s, %20.s[0]     \n"
-
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v14.4s, v15.4s}, [%4]      \n"
-
-                    "fmla   v13.4s, v7.4s, %20.s[0]     \n"
-
-                    "prfm   pldl1keep, [%6, #256]       \n"
-                    "ld1    {v4.4s, v5.4s}, [%6], #32   \n"
-
-                    "fmla   v14.4s, v6.4s, %21.s[0]     \n"
-                    "fmla   v15.4s, v7.4s, %21.s[0]     \n"
-
-                    "fmla   v8.4s, v4.4s, %18.s[1]      \n"
-                    "fmla   v9.4s, v5.4s, %18.s[1]      \n"
-
-                    "fmla   v10.4s, v4.4s, %19.s[1]     \n"
-                    "fmla   v11.4s, v5.4s, %19.s[1]     \n"
-
-                    "fmla   v12.4s, v4.4s, %20.s[1]     \n"
-                    "fmla   v13.4s, v5.4s, %20.s[1]     \n"
-
-                    "prfm   pldl1keep, [%7, #256]       \n"
-                    "ld1    {v6.4s, v7.4s}, [%7], #32   \n"
-
-                    "fmla   v14.4s, v4.4s, %21.s[1]     \n"
-                    "fmla   v15.4s, v5.4s, %21.s[1]     \n"
-
-                    "fmla   v8.4s, v6.4s, %18.s[2]      \n"
-                    "fmla   v9.4s, v7.4s, %18.s[2]      \n"
-
-                    "fmla   v10.4s, v6.4s, %19.s[2]     \n"
-                    "fmla   v11.4s, v7.4s, %19.s[2]     \n"
-
-                    "fmla   v12.4s, v6.4s, %20.s[2]     \n"
-                    "fmla   v13.4s, v7.4s, %20.s[2]     \n"
-
-                    "prfm   pldl1keep, [%8, #256]       \n"
-                    "ld1    {v4.4s, v5.4s}, [%8], #32   \n"
-
-                    "fmla   v14.4s, v6.4s, %21.s[2]     \n"
-                    "fmla   v15.4s, v7.4s, %21.s[2]     \n"
-
-                    "fmla   v8.4s, v4.4s, %18.s[3]      \n"
-                    "fmla   v9.4s, v5.4s, %18.s[3]      \n"
-
-                    "fmla   v10.4s, v4.4s, %19.s[3]     \n"
-                    "fmla   v11.4s, v5.4s, %19.s[3]     \n"
-
-                    "st1    {v8.4s, v9.4s}, [%1], #32   \n"
-
-                    "fmla   v12.4s, v4.4s, %20.s[3]     \n"
-                    "fmla   v13.4s, v5.4s, %20.s[3]     \n"
-
-                    "st1    {v10.4s, v11.4s}, [%2], #32 \n"
-
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v6.4s, v7.4s}, [%5], #32   \n"
-
-                    "fmla   v14.4s, v4.4s, %21.s[3]     \n"
-                    "fmla   v15.4s, v5.4s, %21.s[3]     \n"
-
-                    "st1    {v12.4s, v13.4s}, [%3], #32 \n"
-
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v8.4s, v9.4s}, [%1]        \n"
-
-                    "subs   %w0, %w0, #1                \n"
-
-                    "st1    {v14.4s, v15.4s}, [%4], #32 \n"
-
-                    "bne    0b                          \n"
-                    "sub    %5, %5, #32                 \n"
-                    : "=r"(nn),      // %0
-                    "=r"(outptr0), // %1
-                    "=r"(outptr1), // %2
-                    "=r"(outptr2), // %3
-                    "=r"(outptr3), // %4
-                    "=r"(r0),      // %5
-                    "=r"(r1),      // %6
-                    "=r"(r2),      // %7
-                    "=r"(r3)       // %8
-                    : "0"(nn),
-                    "1"(outptr0),
-                    "2"(outptr1),
-                    "3"(outptr2),
-                    "4"(outptr3),
-                    "5"(r0),
-                    "6"(r1),
-                    "7"(r2),
-                    "8"(r3),
-                    "w"(_k0), // %18
-                    "w"(_k1), // %19
-                    "w"(_k2), // %20
-                    "w"(_k3)  // %21
-                    : "cc", "memory", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    "pld        [%5, #256]              \n"
-                    "vld1.f32   {d12-d15}, [%5 :128]!   \n"
-                    "pld        [%1, #256]              \n"
-                    "vld1.f32   {d16-d19}, [%1 :128]    \n"
-                    "0:                                 \n"
-
-                    "vmla.f32   q8, q6, %e18[0]         \n"
-
-                    "pld        [%2, #256]              \n"
-                    "vld1.f32   {d20-d23}, [%2 :128]    \n"
-                    "vmla.f32   q9, q7, %e18[0]         \n"
-
-                    "vmla.f32   q10, q6, %e19[0]        \n"
-
-                    "pld        [%3, #256]              \n"
-                    "vld1.f32   {d24-d27}, [%3 :128]    \n"
-                    "vmla.f32   q11, q7, %e19[0]        \n"
-
-                    "vmla.f32   q12, q6, %e20[0]        \n"
-
-                    "pld        [%4, #256]              \n"
-                    "vld1.f32   {d28-d31}, [%4 :128]    \n"
-                    "vmla.f32   q13, q7, %e20[0]        \n"
-
-                    "pld        [%6, #256]              \n"
-                    "vld1.f32   {d8-d11}, [%6 :128]!    \n"
-
-                    "vmla.f32   q14, q6, %e21[0]        \n"
-                    "vmla.f32   q15, q7, %e21[0]        \n"
-
-                    "vmla.f32   q8, q4, %e18[1]         \n"
-                    "vmla.f32   q9, q5, %e18[1]         \n"
-
-                    "vmla.f32   q10, q4, %e19[1]        \n"
-                    "vmla.f32   q11, q5, %e19[1]        \n"
-
-                    "vmla.f32   q12, q4, %e20[1]        \n"
-                    "vmla.f32   q13, q5, %e20[1]        \n"
-
-                    "pld        [%7, #256]              \n"
-                    "vld1.f32   {d12-d15}, [%7 :128]!   \n"
-
-                    "vmla.f32   q14, q4, %e21[1]        \n"
-                    "vmla.f32   q15, q5, %e21[1]        \n"
-
-                    "vmla.f32   q8, q6, %f18[0]         \n"
-                    "vmla.f32   q9, q7, %f18[0]         \n"
-
-                    "vmla.f32   q10, q6, %f19[0]        \n"
-                    "vmla.f32   q11, q7, %f19[0]        \n"
-
-                    "vmla.f32   q12, q6, %f20[0]        \n"
-                    "vmla.f32   q13, q7, %f20[0]        \n"
-
-                    "pld        [%8, #256]              \n"
-                    "vld1.f32   {d8-d11}, [%8 :128]!    \n"
-
-                    "vmla.f32   q14, q6, %f21[0]        \n"
-                    "vmla.f32   q15, q7, %f21[0]        \n"
-
-                    "vmla.f32   q8, q4, %f18[1]         \n"
-                    "vmla.f32   q9, q5, %f18[1]         \n"
-
-                    "vmla.f32   q10, q4, %f19[1]        \n"
-                    "vmla.f32   q11, q5, %f19[1]        \n"
-
-                    "vmla.f32   q12, q4, %f20[1]        \n"
-                    "vst1.f32   {d16-d19}, [%1 :128]!   \n"
-
-                    "vmla.f32   q13, q5, %f20[1]        \n"
-
-                    "vst1.f32   {d20-d23}, [%2 :128]!   \n"
-
-                    "vmla.f32   q14, q4, %f21[1]        \n"
-                    "pld        [%5, #256]              \n"
-                    "vld1.f32   {d12-d15}, [%5 :128]!   \n"
-
-                    "vmla.f32   q15, q5, %f21[1]        \n"
-
-                    "vst1.f32   {d24-d27}, [%3 :128]!   \n"
-
-                    "pld        [%1, #256]              \n"
-                    "vld1.f32   {d16-d19}, [%1 :128]    \n"
-
-                    "subs       %0, #1                  \n"
-                    "vst1.f32   {d28-d31}, [%4 :128]!   \n"
-
-                    "bne        0b                      \n"
-                    "sub        %5, #32                 \n"
-                    : "=r"(nn),      // %0
-                    "=r"(outptr0), // %1
-                    "=r"(outptr1), // %2
-                    "=r"(outptr2), // %3
-                    "=r"(outptr3), // %4
-                    "=r"(r0),      // %5
-                    "=r"(r1),      // %6
-                    "=r"(r2),      // %7
-                    "=r"(r3)       // %8
-                    : "0"(nn),
-                    "1"(outptr0),
-                    "2"(outptr1),
-                    "3"(outptr2),
-                    "4"(outptr3),
-                    "5"(r0),
-                    "6"(r1),
-                    "7"(r2),
-                    "8"(r3),
-                    "w"(_k0), // %18
-                    "w"(_k1), // %19
-                    "w"(_k2), // %20
-                    "w"(_k3)  // %21
-                    : "cc", "memory", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-                // TODO neon optimize
-                float sum0 = *r0 * kernel0[0] + *r1 * kernel0[1] + *r2 * kernel0[2] + *r3 * kernel0[3];
-                float sum1 = *r0 * kernel1[0] + *r1 * kernel1[1] + *r2 * kernel1[2] + *r3 * kernel1[3];
-                float sum2 = *r0 * kernel2[0] + *r1 * kernel2[1] + *r2 * kernel2[2] + *r3 * kernel2[3];
-                float sum3 = *r0 * kernel3[0] + *r1 * kernel3[1] + *r2 * kernel3[2] + *r3 * kernel3[3];
-
-                *outptr0 += sum0;
-                *outptr1 += sum1;
-                *outptr2 += sum2;
-                *outptr3 += sum3;
-
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                outptr0++;
-                outptr1++;
-                outptr2++;
-                outptr3++;
-            }
-        }
-
-        for (; q < inch; q++)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-            float* outptr2 = out2;
-            float* outptr3 = out3;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float* kernel1 = kernel + (p + 1) * inch + q;
-            const float* kernel2 = kernel + (p + 2) * inch + q;
-            const float* kernel3 = kernel + (p + 3) * inch + q;
-
-            const float k0 = kernel0[0];
-            const float k1 = kernel1[0];
-            const float k2 = kernel2[0];
-            const float k3 = kernel3[0];
-
-            const float* r0 = img0;
-
-            int size = outw * outh;
-
-#if __ARM_NEON
-            int nn = size >> 3;
-            int remain = size & 7;
-#else
-            int remain = size;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-            float32x4_t _k0 = vdupq_n_f32(k0);
-            float32x4_t _k1 = vdupq_n_f32(k1);
-            float32x4_t _k2 = vdupq_n_f32(k2);
-            float32x4_t _k3 = vdupq_n_f32(k3);
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    "prfm       pldl1keep, [%5, #256]          \n"
-                    "ld1        {v6.4s, v7.4s}, [%5], #32      \n"
-                    "0:                                        \n"
-                    "prfm       pldl1keep, [%1, #256]          \n"
-                    "ld1        {v8.4s, v9.4s}, [%1]           \n"
-                    "fmla       v8.4s, v6.4s, %12.4s           \n"
-                    "fmla       v9.4s, v7.4s, %12.4s           \n"
-
-                    "prfm       pldl1keep, [%2, #256]          \n"
-                    "ld1        {v10.4s, v11.4s}, [%2]         \n"
-                    "fmla       v10.4s, v6.4s, %13.4s          \n"
-                    "fmla       v11.4s, v7.4s, %13.4s          \n"
-
-                    "st1        {v8.4s, v9.4s}, [%1], #32      \n"
-
-                    "prfm       pldl1keep, [%3, #256]          \n"
-                    "ld1        {v12.4s, v13.4s}, [%3]         \n"
-                    "fmla       v12.4s, v6.4s, %14.4s          \n"
-                    "fmla       v13.4s, v7.4s, %14.4s          \n"
-
-                    "st1        {v10.4s, v11.4s}, [%2], #32    \n"
-
-                    "prfm       pldl1keep, [%4, #256]          \n"
-                    "ld1        {v14.4s, v15.4s}, [%4]         \n"
-                    "fmla       v14.4s, v6.4s, %15.4s          \n"
-                    "fmla       v15.4s, v7.4s, %15.4s          \n"
-
-                    "st1        {v12.4s, v13.4s}, [%3], #32    \n"
-
-                    "prfm       pldl1keep, [%5, #256]          \n"
-                    "ld1        {v6.4s, v7.4s}, [%5], #32      \n"
-                    "subs       %w0, %w0, #1                   \n"
-                    "st1        {v14.4s, v15.4s}, [%4], #32    \n"
-                    "bne        0b                             \n"
-                    "sub        %5, %5, #32                    \n"
-                    : "=r"(nn),      // %0
-                    "=r"(outptr0), // %1
-                    "=r"(outptr1), // %2
-                    "=r"(outptr2), // %3
-                    "=r"(outptr3), // %4
-                    "=r"(r0)       // %5
-                    : "0"(nn),
-                    "1"(outptr0),
-                    "2"(outptr1),
-                    "3"(outptr2),
-                    "4"(outptr3),
-                    "5"(r0),
-                    "w"(_k0), // %12
-                    "w"(_k1), // %13
-                    "w"(_k2), // %14
-                    "w"(_k3)  // %15
-                    : "cc", "memory", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    "pld        [%5, #256]              \n"
-                    "vld1.f32   {d12-d15}, [%5 :128]!   \n"
-                    "0:                                 \n"
-                    "pld        [%1, #256]              \n"
-                    "vld1.f32   {d16-d19}, [%1 :128]    \n"
-                    "vmla.f32   q8, q6, %q12            \n"
-                    "vmla.f32   q9, q7, %q12            \n"
-
-                    "pld        [%2, #256]              \n"
-                    "vld1.f32   {d20-d23}, [%2 :128]    \n"
-                    "vmla.f32   q10, q6, %q13           \n"
-                    "vmla.f32   q11, q7, %q13           \n"
-
-                    "vst1.f32   {d16-d19}, [%1 :128]!   \n"
-
-                    "pld        [%3, #256]              \n"
-                    "vld1.f32   {d24-d27}, [%3 :128]    \n"
-                    "vmla.f32   q12, q6, %q14           \n"
-                    "vmla.f32   q13, q7, %q14           \n"
-
-                    "vst1.f32   {d20-d23}, [%2 :128]!   \n"
-
-                    "pld        [%4, #256]              \n"
-                    "vld1.f32   {d28-d31}, [%4 :128]    \n"
-                    "vmla.f32   q14, q6, %q15           \n"
-                    "vmla.f32   q15, q7, %q15           \n"
-
-                    "vst1.f32   {d24-d27}, [%3 :128]!   \n"
-
-                    "pld        [%5, #256]              \n"
-                    "vld1.f32   {d12-d15}, [%5 :128]!   \n"
-                    "subs       %0, #1                  \n"
-                    "vst1.f32   {d28-d31}, [%4 :128]!   \n"
-                    "bne        0b                      \n"
-                    "sub        %5, #32                 \n"
-                    : "=r"(nn),      // %0
-                    "=r"(outptr0), // %1
-                    "=r"(outptr1), // %2
-                    "=r"(outptr2), // %3
-                    "=r"(outptr3), // %4
-                    "=r"(r0)       // %5
-                    : "0"(nn),
-                    "1"(outptr0),
-                    "2"(outptr1),
-                    "3"(outptr2),
-                    "4"(outptr3),
-                    "5"(r0),
-                    "w"(_k0), // %12
-                    "w"(_k1), // %13
-                    "w"(_k2), // %14
-                    "w"(_k3)  // %15
-                    : "cc", "memory", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-                // TODO neon optimize
-                float sum0 = *r0 * k0;
-                float sum1 = *r0 * k1;
-                float sum2 = *r0 * k2;
-                float sum3 = *r0 * k3;
-
-                *outptr0 += sum0;
-                *outptr1 += sum1;
-                *outptr2 += sum2;
-                *outptr3 += sum3;
-
-                r0++;
-                outptr0++;
-                outptr1++;
-                outptr2++;
-                outptr3++;
-            }
-        }
-    }
-
-    remain_outch_start += nn_outch << 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        int q = 0;
-
-        for (; q + 3 < inch; q += 4)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-            const float* img1 = bottom_blob.channel(q + 1);
-            const float* img2 = bottom_blob.channel(q + 2);
-            const float* img3 = bottom_blob.channel(q + 3);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float k0 = kernel0[0];
-            const float k1 = kernel0[1];
-            const float k2 = kernel0[2];
-            const float k3 = kernel0[3];
-
-            const float* r0 = img0;
-            const float* r1 = img1;
-            const float* r2 = img2;
-            const float* r3 = img3;
-
-            int size = outw * outh;
-
-#if __ARM_NEON
-            int nn = size >> 3;
-            int remain = size & 7;
-#else
-            int remain = size;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-            float32x4_t _k0 = vdupq_n_f32(k0);
-            float32x4_t _k1 = vdupq_n_f32(k1);
-            float32x4_t _k2 = vdupq_n_f32(k2);
-            float32x4_t _k3 = vdupq_n_f32(k3);
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    "prfm       pldl1keep, [%2, #256]          \n"
-                    "ld1        {v2.4s, v3.4s}, [%2], #32      \n"
-                    "0:                                        \n"
-                    "prfm       pldl1keep, [%1, #256]          \n"
-                    "ld1        {v0.4s, v1.4s}, [%1]           \n"
-                    "fmla       v0.4s, v2.4s, %12.4s           \n"
-                    "fmla       v1.4s, v3.4s, %12.4s           \n"
-
-                    "prfm       pldl1keep, [%3, #256]          \n"
-                    "ld1        {v2.4s, v3.4s}, [%3], #32      \n"
-                    "fmla       v0.4s, v2.4s, %13.4s           \n"
-                    "fmla       v1.4s, v3.4s, %13.4s           \n"
-
-                    "prfm       pldl1keep, [%4, #256]          \n"
-                    "ld1        {v2.4s, v3.4s}, [%4], #32      \n"
-                    "fmla       v0.4s, v2.4s, %14.4s           \n"
-                    "fmla       v1.4s, v3.4s, %14.4s           \n"
-
-                    "prfm       pldl1keep, [%5, #256]          \n"
-                    "ld1        {v2.4s, v3.4s}, [%5], #32      \n"
-                    "fmla       v0.4s, v2.4s, %15.4s           \n"
-                    "fmla       v1.4s, v3.4s, %15.4s           \n"
-
-                    "prfm       pldl1keep, [%2, #256]          \n"
-                    "ld1        {v2.4s, v3.4s}, [%2], #32      \n"
-                    "subs       %w0, %w0, #1                   \n"
-                    "st1        {v0.4s, v1.4s}, [%1], #32      \n"
-                    "bne        0b                             \n"
-                    "sub        %2, %2, #32                    \n"
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0),     // %2
-                    "=r"(r1),     // %3
-                    "=r"(r2),     // %4
-                    "=r"(r3)      // %5
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "w"(_k0), // %12
-                    "w"(_k1), // %13
-                    "w"(_k2), // %14
-                    "w"(_k3)  // %15
-                    : "cc", "memory", "v0", "v1", "v2", "v3");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    "pld        [%2, #256]          \n"
-                    "vld1.f32   {d4-d7}, [%2 :128]! \n"
-                    "0:                             \n"
-                    "pld        [%1, #256]          \n"
-                    "vld1.f32   {d0-d3}, [%1 :128]  \n"
-                    "vmla.f32   q0, q2, %q12        \n"
-                    "vmla.f32   q1, q3, %q12        \n"
-                    "pld        [%3, #256]          \n"
-                    "vld1.f32   {d4-d7}, [%3 :128]! \n"
-                    "vmla.f32   q0, q2, %q13        \n"
-                    "vmla.f32   q1, q3, %q13        \n"
-                    "pld        [%4, #256]          \n"
-                    "vld1.f32   {d4-d7}, [%4 :128]! \n"
-                    "vmla.f32   q0, q2, %q14        \n"
-                    "vmla.f32   q1, q3, %q14        \n"
-                    "pld        [%5, #256]          \n"
-                    "vld1.f32   {d4-d7}, [%5 :128]! \n"
-                    "vmla.f32   q0, q2, %q15        \n"
-                    "vmla.f32   q1, q3, %q15        \n"
-                    "pld        [%2, #256]          \n"
-                    "vld1.f32   {d4-d7}, [%2 :128]! \n"
-                    "subs       %0, #1              \n"
-                    "vst1.f32   {d0-d3}, [%1 :128]! \n"
-                    "bne        0b                  \n"
-                    "sub        %2, #32             \n"
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0),     // %2
-                    "=r"(r1),     // %3
-                    "=r"(r2),     // %4
-                    "=r"(r3)      // %5
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "w"(_k0), // %12
-                    "w"(_k1), // %13
-                    "w"(_k2), // %14
-                    "w"(_k3)  // %15
-                    : "cc", "memory", "q0", "q1", "q2", "q3");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-                float sum = *r0 * k0;
-                float sum1 = *r1 * k1;
-                float sum2 = *r2 * k2;
-                float sum3 = *r3 * k3;
-
-                *outptr += sum + sum1 + sum2 + sum3;
-
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                outptr++;
-            }
-        }
-
-        for (; q < inch; q++)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float k0 = kernel0[0];
-
-            const float* r0 = img0;
-
-            int size = outw * outh;
-
-#if __ARM_NEON
-            int nn = size >> 3;
-            int remain = size & 7;
-#else
-            int remain = size;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-            float32x4_t _k0 = vdupq_n_f32(k0);
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    "prfm       pldl1keep, [%2, #256]          \n"
-                    "ld1        {v2.4s, v3.4s}, [%2], #32      \n"
-                    "0:                                        \n"
-                    "prfm       pldl1keep, [%1, #256]          \n"
-                    "ld1        {v0.4s, v1.4s}, [%1]           \n"
-                    "fmla       v0.4s, v2.4s, %6.4s            \n"
-                    "fmla       v1.4s, v3.4s, %6.4s            \n"
-                    "prfm       pldl1keep, [%2, #256]          \n"
-                    "ld1        {v2.4s, v3.4s}, [%2], #32      \n"
-                    "subs       %w0, %w0, #1                   \n"
-                    "st1        {v0.4s, v1.4s}, [%1], #32      \n"
-                    "bne        0b                             \n"
-                    "sub        %2, %2, #32                    \n"
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0)      // %2
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "w"(_k0) // %6
-                    : "cc", "memory", "v0", "v1", "v2", "v3");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    "pld        [%2, #256]          \n"
-                    "vld1.f32   {d4-d7}, [%2 :128]! \n"
-                    "0:                             \n"
-                    "pld        [%1, #256]          \n"
-                    "vld1.f32   {d0-d3}, [%1 :128]  \n"
-                    "vmla.f32   q0, q2, %q6         \n"
-                    "vmla.f32   q1, q3, %q6         \n"
-                    "pld        [%2, #256]          \n"
-                    "vld1.f32   {d4-d7}, [%2 :128]! \n"
-                    "subs       %0, #1              \n"
-                    "vst1.f32   {d0-d3}, [%1 :128]! \n"
-                    "bne        0b                  \n"
-                    "sub        %2, #32             \n"
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0)      // %2
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "w"(_k0) // %6
-                    : "cc", "memory", "q0", "q1", "q2", "q3");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-                float sum = *r0 * k0;
-
-                *outptr += sum;
-
-                r0++;
-                outptr++;
-            }
-        }
-    }
-}
-
-static void conv1x1s2_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    int nn_outch = outch >> 2;
-    int remain_outch_start = nn_outch << 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 4;
-
-        Mat out0 = top_blob.channel(p);
-        Mat out1 = top_blob.channel(p + 1);
-        Mat out2 = top_blob.channel(p + 2);
-        Mat out3 = top_blob.channel(p + 3);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        const float bias1 = bias ? bias[p + 1] : 0.f;
-        const float bias2 = bias ? bias[p + 2] : 0.f;
-        const float bias3 = bias ? bias[p + 3] : 0.f;
-
-        out0.fill(bias0);
-        out1.fill(bias1);
-        out2.fill(bias2);
-        out3.fill(bias3);
-
-        int q = 0;
-
-        for (; q + 3 < inch; q += 4)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-            float* outptr2 = out2;
-            float* outptr3 = out3;
-
-            const float* img0 = bottom_blob.channel(q);
-            const float* img1 = bottom_blob.channel(q + 1);
-            const float* img2 = bottom_blob.channel(q + 2);
-            const float* img3 = bottom_blob.channel(q + 3);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float* kernel1 = kernel + (p + 1) * inch + q;
-            const float* kernel2 = kernel + (p + 2) * inch + q;
-            const float* kernel3 = kernel + (p + 3) * inch + q;
-
-            const float* r0 = img0;
-            const float* r1 = img1;
-            const float* r2 = img2;
-            const float* r3 = img3;
-
-            for (int i = 0; i < outh; i++)
-            {
-                int size = outw;
-
-#if __ARM_NEON
-                int nn = size >> 3;
-                int remain = size & 7;
-#else
-                int remain = size;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-                float32x4_t _k0 = vld1q_f32(kernel0);
-                float32x4_t _k1 = vld1q_f32(kernel1);
-                float32x4_t _k2 = vld1q_f32(kernel2);
-                float32x4_t _k3 = vld1q_f32(kernel3);
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "0:                                        \n"
-
-                        "prfm       pldl1keep, [%5, #512]          \n"
-                        "ld2        {v4.4s, v5.4s}, [%5], #32      \n"
-                        "ld2        {v6.4s, v7.4s}, [%5], #32      \n"
-                        "and        v5.16b, v6.16b, v6.16b         \n" // v4 v5
-
-                        "prfm       pldl1keep, [%1, #256]          \n"
-                        "ld1        {v8.4s, v9.4s}, [%1]           \n"
-
-                        "fmla       v8.4s, v4.4s, %18.s[0]         \n"
-                        "fmla       v9.4s, v5.4s, %18.s[0]         \n"
-
-                        "prfm       pldl1keep, [%2, #256]          \n"
-                        "ld1        {v10.4s, v11.4s}, [%2]         \n"
-
-                        "fmla       v10.4s, v4.4s, %19.s[0]        \n"
-                        "fmla       v11.4s, v5.4s, %19.s[0]        \n"
-
-                        "prfm       pldl1keep, [%3, #256]          \n"
-                        "ld1        {v12.4s, v13.4s}, [%3]         \n"
-
-                        "fmla       v12.4s, v4.4s, %20.s[0]        \n"
-                        "fmla       v13.4s, v5.4s, %20.s[0]        \n"
-
-                        "prfm       pldl1keep, [%4, #256]          \n"
-                        "ld1        {v14.4s, v15.4s}, [%4]         \n"
-
-                        "prfm       pldl1keep, [%6, #512]          \n"
-                        "ld2        {v6.4s, v7.4s}, [%6], #32      \n"
-
-                        "fmla       v14.4s, v4.4s, %21.s[0]        \n"
-                        "fmla       v15.4s, v5.4s, %21.s[0]        \n"
-
-                        "ld2        {v4.4s, v5.4s}, [%6], #32      \n"
-                        "and        v7.16b, v4.16b, v4.16b         \n" // v6 v7
-
-                        "fmla       v8.4s, v6.4s, %18.s[1]         \n"
-                        "fmla       v9.4s, v7.4s, %18.s[1]         \n"
-
-                        "fmla       v10.4s, v6.4s, %19.s[1]        \n"
-                        "fmla       v11.4s, v7.4s, %19.s[1]        \n"
-
-                        "fmla       v12.4s, v6.4s, %20.s[1]        \n"
-                        "fmla       v13.4s, v7.4s, %20.s[1]        \n"
-
-                        "prfm       pldl1keep, [%7, #512]          \n"
-                        "ld2        {v4.4s, v5.4s}, [%7], #32      \n"
-
-                        "fmla       v14.4s, v6.4s, %21.s[1]        \n"
-                        "fmla       v15.4s, v7.4s, %21.s[1]        \n"
-
-                        "ld2        {v6.4s, v7.4s}, [%7], #32      \n"
-                        "and        v5.16b, v6.16b, v6.16b         \n" // v4 v5
-
-                        "fmla       v8.4s, v4.4s, %18.s[2]         \n"
-                        "fmla       v9.4s, v5.4s, %18.s[2]         \n"
-
-                        "fmla       v10.4s, v4.4s, %19.s[2]        \n"
-                        "fmla       v11.4s, v5.4s, %19.s[2]        \n"
-
-                        "fmla       v12.4s, v4.4s, %20.s[2]        \n"
-                        "fmla       v13.4s, v5.4s, %20.s[2]        \n"
-
-                        "prfm       pldl1keep, [%8, #512]          \n"
-                        "ld2        {v6.4s, v7.4s}, [%8], #32      \n"
-
-                        "fmla       v14.4s, v4.4s, %21.s[2]        \n"
-                        "fmla       v15.4s, v5.4s, %21.s[2]        \n"
-
-                        "ld2        {v4.4s, v5.4s}, [%8], #32      \n"
-                        "and        v7.16b, v4.16b, v4.16b         \n" // v6 v7
-
-                        "fmla       v8.4s, v6.4s, %18.s[3]         \n"
-                        "fmla       v9.4s, v7.4s, %18.s[3]         \n"
-
-                        "fmla       v10.4s, v6.4s, %19.s[3]        \n"
-                        "fmla       v11.4s, v7.4s, %19.s[3]        \n"
-
-                        "st1        {v8.4s, v9.4s}, [%1], #32      \n"
-
-                        "fmla       v12.4s, v6.4s, %20.s[3]        \n"
-                        "fmla       v13.4s, v7.4s, %20.s[3]        \n"
-
-                        "st1        {v10.4s, v11.4s}, [%2], #32    \n"
-
-                        "fmla       v14.4s, v6.4s, %21.s[3]        \n"
-                        "fmla       v15.4s, v7.4s, %21.s[3]        \n"
-
-                        "st1        {v12.4s, v13.4s}, [%3], #32    \n"
-
-                        "subs       %w0, %w0, #1                   \n"
-                        "st1        {v14.4s, v15.4s}, [%4], #32    \n"
-
-                        "bne        0b                             \n"
-                        : "=r"(nn),      // %0
-                        "=r"(outptr0), // %1
-                        "=r"(outptr1), // %2
-                        "=r"(outptr2), // %3
-                        "=r"(outptr3), // %4
-                        "=r"(r0),      // %5
-                        "=r"(r1),      // %6
-                        "=r"(r2),      // %7
-                        "=r"(r3)       // %8
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(outptr1),
-                        "3"(outptr2),
-                        "4"(outptr3),
-                        "5"(r0),
-                        "6"(r1),
-                        "7"(r2),
-                        "8"(r3),
-                        "w"(_k0), // %18
-                        "w"(_k1), // %19
-                        "w"(_k2), // %20
-                        "w"(_k3)  // %21
-                        : "cc", "memory", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "0:                             \n"
-
-                        "pld        [%5, #512]          \n"
-                        "vld2.f32   {d8-d11}, [%5]!     \n"
-                        "vld2.f32   {d12-d15}, [%5]!    \n"
-                        "vand       q5, q6, q6          \n" // q4 q5
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.f32   {d16-d19}, [%1]     \n"
-
-                        "vmla.f32   q8, q4, %e18[0]     \n"
-                        "vmla.f32   q9, q5, %e18[0]     \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.f32   {d20-d23}, [%2]     \n"
-
-                        "vmla.f32   q10, q4, %e19[0]    \n"
-                        "vmla.f32   q11, q5, %e19[0]    \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.f32   {d24-d27}, [%3]     \n"
-
-                        "vmla.f32   q12, q4, %e20[0]    \n"
-                        "vmla.f32   q13, q5, %e20[0]    \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.f32   {d28-d31}, [%4]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vld2.f32   {d12-d15}, [%6]!    \n"
-
-                        "vmla.f32   q14, q4, %e21[0]    \n"
-                        "vmla.f32   q15, q5, %e21[0]    \n"
-
-                        "vld2.f32   {d8-d11}, [%6]!     \n"
-                        "vand       q7, q4, q4          \n" // q6 q7
-
-                        "vmla.f32   q8, q6, %e18[1]     \n"
-                        "vmla.f32   q9, q7, %e18[1]     \n"
-
-                        "vmla.f32   q10, q6, %e19[1]    \n"
-                        "vmla.f32   q11, q7, %e19[1]    \n"
-
-                        "vmla.f32   q12, q6, %e20[1]    \n"
-                        "vmla.f32   q13, q7, %e20[1]    \n"
-
-                        "pld        [%7, #512]          \n"
-                        "vld2.f32   {d8-d11}, [%7]!     \n"
-
-                        "vmla.f32   q14, q6, %e21[1]    \n"
-                        "vmla.f32   q15, q7, %e21[1]    \n"
-
-                        "vld2.f32   {d12-d15}, [%7]!    \n"
-                        "vand       q5, q6, q6          \n" // q4 q5
-
-                        "vmla.f32   q8, q4, %f18[0]     \n"
-                        "vmla.f32   q9, q5, %f18[0]     \n"
-
-                        "vmla.f32   q10, q4, %f19[0]    \n"
-                        "vmla.f32   q11, q5, %f19[0]    \n"
-
-                        "vmla.f32   q12, q4, %f20[0]    \n"
-                        "vmla.f32   q13, q5, %f20[0]    \n"
-
-                        "pld        [%8, #512]          \n"
-                        "vld2.f32   {d12-d15}, [%8]!    \n"
-
-                        "vmla.f32   q14, q4, %f21[0]    \n"
-                        "vmla.f32   q15, q5, %f21[0]    \n"
-
-                        "vld2.f32   {d8-d11}, [%8]!     \n"
-                        "vand       q7, q4, q4          \n" // q6 q7
-
-                        "vmla.f32   q8, q6, %f18[1]     \n"
-                        "vmla.f32   q9, q7, %f18[1]     \n"
-
-                        "vmla.f32   q10, q6, %f19[1]    \n"
-                        "vmla.f32   q11, q7, %f19[1]    \n"
-
-                        "vst1.f32   {d16-d19}, [%1]!    \n"
-
-                        "vmla.f32   q12, q6, %f20[1]    \n"
-                        "vmla.f32   q13, q7, %f20[1]    \n"
-
-                        "vst1.f32   {d20-d23}, [%2]!    \n"
-
-                        "vmla.f32   q14, q6, %f21[1]    \n"
-                        "vmla.f32   q15, q7, %f21[1]    \n"
-
-                        "vst1.f32   {d24-d27}, [%3]!    \n"
-
-                        "subs       %0, #1              \n"
-                        "vst1.f32   {d28-d31}, [%4]!    \n"
-
-                        "bne        0b                  \n"
-                        : "=r"(nn),      // %0
-                        "=r"(outptr0), // %1
-                        "=r"(outptr1), // %2
-                        "=r"(outptr2), // %3
-                        "=r"(outptr3), // %4
-                        "=r"(r0),      // %5
-                        "=r"(r1),      // %6
-                        "=r"(r2),      // %7
-                        "=r"(r3)       // %8
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(outptr1),
-                        "3"(outptr2),
-                        "4"(outptr3),
-                        "5"(r0),
-                        "6"(r1),
-                        "7"(r2),
-                        "8"(r3),
-                        "w"(_k0), // %18
-                        "w"(_k1), // %19
-                        "w"(_k2), // %20
-                        "w"(_k3)  // %21
-                        : "cc", "memory", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-                    // TODO neon optimize
-                    float sum0 = *r0 * kernel0[0] + *r1 * kernel0[1] + *r2 * kernel0[2] + *r3 * kernel0[3];
-                    float sum1 = *r0 * kernel1[0] + *r1 * kernel1[1] + *r2 * kernel1[2] + *r3 * kernel1[3];
-                    float sum2 = *r0 * kernel2[0] + *r1 * kernel2[1] + *r2 * kernel2[2] + *r3 * kernel2[3];
-                    float sum3 = *r0 * kernel3[0] + *r1 * kernel3[1] + *r2 * kernel3[2] + *r3 * kernel3[3];
-
-                    *outptr0 += sum0;
-                    *outptr1 += sum1;
-                    *outptr2 += sum2;
-                    *outptr3 += sum3;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    r3 += 2;
-                    outptr0++;
-                    outptr1++;
-                    outptr2++;
-                    outptr3++;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-            }
-        }
-
-        for (; q < inch; q++)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-            float* outptr2 = out2;
-            float* outptr3 = out3;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float* kernel1 = kernel + (p + 1) * inch + q;
-            const float* kernel2 = kernel + (p + 2) * inch + q;
-            const float* kernel3 = kernel + (p + 3) * inch + q;
-
-            const float k0 = kernel0[0];
-            const float k1 = kernel1[0];
-            const float k2 = kernel2[0];
-            const float k3 = kernel3[0];
-
-            const float* r0 = img0;
-
-            for (int i = 0; i < outh; i++)
-            {
-                int size = outw;
-
-#if __ARM_NEON
-                int nn = size >> 3;
-                int remain = size & 7;
-#else
-                int remain = size;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-                float32x4_t _k0 = vdupq_n_f32(k0);
-                float32x4_t _k1 = vdupq_n_f32(k1);
-                float32x4_t _k2 = vdupq_n_f32(k2);
-                float32x4_t _k3 = vdupq_n_f32(k3);
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "0:                                        \n"
-
-                        "prfm       pldl1keep, [%5, #512]          \n"
-                        "ld2        {v4.4s, v5.4s}, [%5], #32      \n"
-                        "ld2        {v6.4s, v7.4s}, [%5], #32      \n"
-                        "and        v5.16b, v6.16b, v6.16b         \n"
-
-                        "prfm       pldl1keep, [%1, #256]          \n"
-                        "ld1        {v8.4s, v9.4s}, [%1]           \n"
-
-                        "fmla       v8.4s, v4.4s, %12.4s           \n"
-                        "fmla       v9.4s, v5.4s, %12.4s           \n"
-
-                        "prfm       pldl1keep, [%2, #256]          \n"
-                        "ld1        {v10.4s, v11.4s}, [%2]         \n"
-
-                        "fmla       v10.4s, v4.4s, %13.4s          \n"
-                        "fmla       v11.4s, v5.4s, %13.4s          \n"
-
-                        "prfm       pldl1keep, [%3, #256]          \n"
-                        "ld1        {v12.4s, v13.4s}, [%3]         \n"
-
-                        "st1        {v8.4s, v9.4s}, [%1], #32      \n"
-
-                        "fmla       v12.4s, v4.4s, %14.4s          \n"
-                        "fmla       v13.4s, v5.4s, %14.4s          \n"
-
-                        "prfm       pldl1keep, [%4, #256]          \n"
-                        "ld1        {v14.4s, v15.4s}, [%4]         \n"
-
-                        "st1        {v10.4s, v11.4s}, [%2], #32    \n"
-
-                        "fmla       v14.4s, v4.4s, %15.4s          \n"
-                        "fmla       v15.4s, v5.4s, %15.4s          \n"
-
-                        "st1        {v12.4s, v13.4s}, [%3], #32    \n"
-                        "subs       %w0, %w0, #1                   \n"
-
-                        "st1        {v14.4s, v15.4s}, [%4], #32    \n"
-                        "bne        0b                             \n"
-                        : "=r"(nn),      // %0
-                        "=r"(outptr0), // %1
-                        "=r"(outptr1), // %2
-                        "=r"(outptr2), // %3
-                        "=r"(outptr3), // %4
-                        "=r"(r0)       // %5
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(outptr1),
-                        "3"(outptr2),
-                        "4"(outptr3),
-                        "5"(r0),
-                        "w"(_k0), // %12
-                        "w"(_k1), // %13
-                        "w"(_k2), // %14
-                        "w"(_k3)  // %15
-                        : "cc", "memory", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "0:                             \n"
-
-                        "pld        [%5, #512]          \n"
-                        "vld2.f32   {d8-d11}, [%5]!     \n"
-                        "vld2.f32   {d12-d15}, [%5]!    \n"
-                        "vand       q5, q6, q6          \n" // q4 q5
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.f32   {d16-d19}, [%1]     \n"
-
-                        "vmla.f32   q8, q4, %q12        \n"
-                        "vmla.f32   q9, q5, %q12        \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.f32   {d20-d23}, [%2]     \n"
-
-                        "vmla.f32   q10, q4, %q13       \n"
-                        "vmla.f32   q11, q5, %q13       \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.f32   {d24-d27}, [%3]     \n"
-
-                        "vst1.f32   {d16-d19}, [%1]!    \n"
-
-                        "vmla.f32   q12, q4, %q14       \n"
-                        "vmla.f32   q13, q5, %q14       \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.f32   {d28-d31}, [%4]     \n"
-
-                        "vst1.f32   {d20-d23}, [%2]!    \n"
-
-                        "vmla.f32   q14, q4, %q15       \n"
-                        "vmla.f32   q15, q5, %q15       \n"
-
-                        "vst1.f32   {d24-d27}, [%3]!    \n"
-                        "subs       %0, #1              \n"
-                        "vst1.f32   {d28-d31}, [%4]!    \n"
-                        "bne        0b                  \n"
-                        : "=r"(nn),      // %0
-                        "=r"(outptr0), // %1
-                        "=r"(outptr1), // %2
-                        "=r"(outptr2), // %3
-                        "=r"(outptr3), // %4
-                        "=r"(r0)       // %5
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(outptr1),
-                        "3"(outptr2),
-                        "4"(outptr3),
-                        "5"(r0),
-                        "w"(_k0), // %12
-                        "w"(_k1), // %13
-                        "w"(_k2), // %14
-                        "w"(_k3)  // %15
-                        : "cc", "memory", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-                    // TODO neon optimize
-                    float sum0 = *r0 * k0;
-                    float sum1 = *r0 * k1;
-                    float sum2 = *r0 * k2;
-                    float sum3 = *r0 * k3;
-
-                    *outptr0 += sum0;
-                    *outptr1 += sum1;
-                    *outptr2 += sum2;
-                    *outptr3 += sum3;
-
-                    r0 += 2;
-                    outptr0++;
-                    outptr1++;
-                    outptr2++;
-                    outptr3++;
-                }
-
-                r0 += tailstep;
-            }
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        int q = 0;
-
-        for (; q + 3 < inch; q += 4)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-            const float* img1 = bottom_blob.channel(q + 1);
-            const float* img2 = bottom_blob.channel(q + 2);
-            const float* img3 = bottom_blob.channel(q + 3);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float k0 = kernel0[0];
-            const float k1 = kernel0[1];
-            const float k2 = kernel0[2];
-            const float k3 = kernel0[3];
-
-            const float* r0 = img0;
-            const float* r1 = img1;
-            const float* r2 = img2;
-            const float* r3 = img3;
-
-            for (int i = 0; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 3;
-                int remain = outw & 7;
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-                float32x4_t _k0 = vdupq_n_f32(k0);
-                float32x4_t _k1 = vdupq_n_f32(k1);
-                float32x4_t _k2 = vdupq_n_f32(k2);
-                float32x4_t _k3 = vdupq_n_f32(k3);
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "prfm       pldl1keep, [%2, #512]          \n"
-                        "ld2        {v2.4s, v3.4s}, [%2], #32      \n"
-                        "ld2        {v8.4s, v9.4s}, [%2], #32      \n"
-                        "0:                                        \n"
-
-                        "prfm       pldl1keep, [%1, #256]          \n"
-                        "ld1        {v0.4s, v1.4s}, [%1]           \n"
-                        "fmla       v0.4s, v2.4s, %12.4s           \n"
-                        "fmla       v1.4s, v8.4s, %12.4s           \n"
-
-                        "prfm       pldl1keep, [%3, #512]          \n"
-                        "ld2        {v2.4s, v3.4s}, [%3], #32      \n"
-                        "ld2        {v8.4s, v9.4s}, [%3], #32      \n"
-                        "fmla       v0.4s, v2.4s, %13.4s           \n"
-                        "fmla       v1.4s, v8.4s, %13.4s           \n"
-
-                        "prfm       pldl1keep, [%4, #512]          \n"
-                        "ld2        {v2.4s, v3.4s}, [%4], #32      \n"
-                        "ld2        {v8.4s, v9.4s}, [%4], #32      \n"
-                        "fmla       v0.4s, v2.4s, %14.4s           \n"
-                        "fmla       v1.4s, v8.4s, %14.4s           \n"
-
-                        "prfm       pldl1keep, [%5, #512]          \n"
-                        "ld2        {v2.4s, v3.4s}, [%5], #32      \n"
-                        "ld2        {v8.4s, v9.4s}, [%5], #32      \n"
-                        "fmla       v0.4s, v2.4s, %15.4s           \n"
-                        "fmla       v1.4s, v8.4s, %15.4s           \n"
-
-                        "prfm       pldl1keep, [%2, #512]          \n"
-                        "ld2        {v2.4s, v3.4s}, [%2], #32      \n"
-                        "ld2        {v8.4s, v9.4s}, [%2], #32      \n"
-
-                        "subs       %w0, %w0, #1                   \n"
-                        "st1        {v0.4s, v1.4s}, [%1], #32      \n"
-                        "bne        0b                             \n"
-                        "sub        %2, %2, #64                    \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2),     // %4
-                        "=r"(r3)      // %5
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "w"(_k0), // %12
-                        "w"(_k1), // %13
-                        "w"(_k2), // %14
-                        "w"(_k3)  // %15
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v8", "v9");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "pld        [%2, #512]          \n"
-                        "vld2.f32   {d4-d7}, [%2]!      \n"
-                        "vld2.f32   {d16-d19}, [%2]!    \n"
-                        "0:                             \n"
-                        "pld        [%1, #256]          \n"
-                        "vld1.f32   {d0-d3}, [%1]       \n"
-                        "vmla.f32   q0, q2, %q12        \n"
-                        "vmla.f32   q1, q8, %q12        \n"
-                        "pld        [%3, #512]          \n"
-                        "vld2.f32   {d4-d7}, [%3]!      \n"
-                        "vld2.f32   {d16-d19}, [%3]!    \n"
-                        "vmla.f32   q0, q2, %q13        \n"
-                        "vmla.f32   q1, q8, %q13        \n"
-                        "pld        [%4, #512]          \n"
-                        "vld2.f32   {d4-d7}, [%4]!      \n"
-                        "vld2.f32   {d16-d19}, [%4]!    \n"
-                        "vmla.f32   q0, q2, %q14        \n"
-                        "vmla.f32   q1, q8, %q14        \n"
-                        "pld        [%5, #512]          \n"
-                        "vld2.f32   {d4-d7}, [%5]!      \n"
-                        "vld2.f32   {d16-d19}, [%5]!    \n"
-                        "vmla.f32   q0, q2, %q15        \n"
-                        "vmla.f32   q1, q8, %q15        \n"
-                        "pld        [%2, #512]          \n"
-                        "vld2.f32   {d4-d7}, [%2]!      \n"
-                        "vld2.f32   {d16-d19}, [%2]!    \n"
-                        "subs       %0, #1              \n"
-                        "vst1.f32   {d0-d3}, [%1]!      \n"
-                        "bne        0b                  \n"
-                        "sub        %2, #64             \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2),     // %4
-                        "=r"(r3)      // %5
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "w"(_k0), // %12
-                        "w"(_k1), // %13
-                        "w"(_k2), // %14
-                        "w"(_k3)  // %15
-                        : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-                    float sum = *r0 * k0;
-                    float sum1 = *r1 * k1;
-                    float sum2 = *r2 * k2;
-                    float sum3 = *r3 * k3;
-
-                    *outptr += sum + sum1 + sum2 + sum3;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    r3 += 2;
-                    outptr++;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-            }
-        }
-
-        for (; q < inch; q++)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float k0 = kernel0[0];
-
-            const float* r0 = img0;
-
-            for (int i = 0; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 3;
-                int remain = outw & 7;
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-                float32x4_t _k0 = vdupq_n_f32(k0);
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "prfm       pldl1keep, [%2, #512]          \n"
-                        "ld2        {v2.4s, v3.4s}, [%2], #32      \n"
-                        "ld2        {v8.4s, v9.4s}, [%2], #32      \n"
-
-                        "0:                                        \n"
-
-                        "prfm       pldl1keep, [%1, #256]          \n"
-                        "ld1        {v0.4s, v1.4s}, [%1]           \n"
-                        "fmla       v0.4s, v2.4s, %6.4s            \n"
-                        "fmla       v1.4s, v8.4s, %6.4s            \n"
-
-                        "prfm       pldl1keep, [%2, #512]          \n"
-                        "ld2        {v2.4s, v3.4s}, [%2], #32      \n"
-                        "ld2        {v8.4s, v9.4s}, [%2], #32      \n"
-
-                        "subs       %w0, %w0, #1                   \n"
-                        "st1        {v0.4s, v1.4s}, [%1], #32      \n"
-                        "bne        0b                             \n"
-                        "sub        %2, %2, #64                    \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0)      // %2
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "w"(_k0) // %6
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v8", "v9");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "pld        [%2, #512]          \n"
-                        "vld2.f32   {d4-d7}, [%2]!      \n"
-                        "vld2.f32   {d16-d19}, [%2]!    \n"
-                        "0:                             \n"
-                        "pld        [%1, #256]          \n"
-                        "vld1.f32   {d0-d3}, [%1]       \n"
-                        "vmla.f32   q0, q2, %q6         \n"
-                        "vmla.f32   q1, q8, %q6         \n"
-                        "pld        [%2, #512]          \n"
-                        "vld2.f32   {d4-d7}, [%2]!      \n"
-                        "vld2.f32   {d16-d19}, [%2]!    \n"
-                        "subs       %0, #1              \n"
-                        "vst1.f32   {d0-d3}, [%1]!      \n"
-                        "bne        0b                  \n"
-                        "sub        %2, #64             \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0)      // %2
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "w"(_k0) // %6
-                        : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-                    float sum = *r0 * k0;
-
-                    *outptr += sum;
-
-                    r0 += 2;
-                    outptr++;
-                }
-
-                r0 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_2x2.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_2x2.h
deleted file mode 100644
index 2bedf4bf..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_2x2.h
+++ /dev/null
@@ -1,443 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv2x2s1_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        int q = 0;
-
-        for (; q + 1 < inch; q += 2)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-            const float* img1 = bottom_blob.channel(q + 1);
-
-            const float* kernel0 = kernel + p * inch * 4 + q * 4;
-            const float* kernel1 = kernel0 + 4;
-
-            const float* r00 = img0;
-            const float* r01 = img0 + w;
-
-            const float* r10 = img1;
-            const float* r11 = img1 + w;
-
-#if __ARM_NEON
-            float32x4_t _k0 = vld1q_f32(kernel0);
-            float32x4_t _k1 = vld1q_f32(kernel1);
-#endif // __ARM_NEON
-
-            for (int i = 0; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw & 3;
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "prfm       pldl1keep, [%1, #128]          \n"
-                        "ld1        {v0.4s}, [%1], #16             \n"
-                        "prfm       pldl1keep, [%2, #128]          \n"
-                        "ld1        {v2.4s}, [%2], #16             \n"
-                        "prfm       pldl1keep, [%3, #128]          \n"
-                        "ld1        {v12.4s}, [%3], #16            \n"
-                        "prfm       pldl1keep, [%4, #128]          \n"
-                        "ld1        {v14.4s}, [%4], #16            \n"
-
-                        "0:                                        \n"
-                        "prfm       pldl1keep, [%5, #128]          \n"
-                        "ld1        {v9.4s}, [%5]                  \n"
-
-                        "fmul       v8.4s, v0.4s, %12.s[0]         \n"
-                        "fmla       v9.4s, v2.4s, %12.s[2]         \n"
-
-                        "prfm       pldl1keep, [%1, #128]          \n"
-                        "ld1        {v1.4s}, [%1], #16             \n"
-
-                        "prfm       pldl1keep, [%2, #128]          \n"
-                        "ld1        {v3.4s}, [%2], #16             \n"
-
-                        "ext        v10.16b, v0.16b, v1.16b, #4    \n"
-                        "ext        v11.16b, v2.16b, v3.16b, #4    \n"
-
-                        "fmla       v8.4s, v12.4s, %13.s[0]        \n"
-                        "fmla       v9.4s, v14.4s, %13.s[2]        \n"
-
-                        "prfm       pldl1keep, [%3, #128]          \n"
-                        "ld1        {v13.4s}, [%3], #16            \n"
-
-                        "prfm       pldl1keep, [%4, #128]          \n"
-                        "ld1        {v15.4s}, [%4], #16            \n"
-
-                        "fmla       v8.4s, v10.4s, %12.s[1]        \n"
-                        "fmla       v9.4s, v11.4s, %12.s[3]        \n"
-
-                        "ext        v10.16b, v12.16b, v13.16b, #4  \n"
-                        "ext        v11.16b, v14.16b, v15.16b, #4  \n"
-
-                        "fmla       v8.4s, v10.4s, %13.s[1]        \n"
-                        "fmla       v9.4s, v11.4s, %13.s[3]        \n"
-
-                        "orr        v0.16b, v1.16b, v1.16b         \n"
-                        "orr        v2.16b, v3.16b, v3.16b         \n"
-
-                        "fadd       v8.4s, v8.4s, v9.4s            \n"
-
-                        "orr        v12.16b, v13.16b, v13.16b      \n"
-                        "orr        v14.16b, v15.16b, v15.16b      \n"
-
-                        "subs       %w0, %w0, #1                   \n"
-                        "st1        {v8.4s}, [%5], #16             \n"
-                        "bne        0b                             \n"
-                        "sub        %1, %1, #16                    \n"
-                        "sub        %2, %2, #16                    \n"
-                        "sub        %3, %3, #16                    \n"
-                        "sub        %4, %4, #16                    \n"
-                        : "=r"(nn),    // %0
-                        "=r"(r00),   // %1
-                        "=r"(r01),   // %2
-                        "=r"(r10),   // %3
-                        "=r"(r11),   // %4
-                        "=r"(outptr) // %5
-                        : "0"(nn),
-                        "1"(r00),
-                        "2"(r01),
-                        "3"(r10),
-                        "4"(r11),
-                        "5"(outptr),
-                        "w"(_k0), // %12
-                        "w"(_k1)  // %13
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%1]!      \n"
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d4-d5}, [%2]!      \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.f32   {d24-d25}, [%3]!    \n"
-                        "pld        [%4, #128]          \n"
-                        "vld1.f32   {d28-d29}, [%4]!    \n"
-
-                        "0:                             \n"
-                        "pld        [%5, #128]          \n"
-                        "vld1.f32   {d18-d19}, [%5]     \n" // q9 = sum
-
-                        "vmul.f32   q8, q0, %e12[0]     \n"
-                        "vmla.f32   q9, q2, %f12[0]     \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d2-d3}, [%1]!      \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d6-d7}, [%2]!      \n"
-
-                        "vext.f32   q10, q0, q1, #1     \n"
-                        "vext.f32   q11, q2, q3, #1     \n"
-
-                        "vmla.f32   q8, q12, %e13[0]    \n"
-                        "vmla.f32   q9, q14, %f13[0]    \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.f32   {d26-d27}, [%3]!    \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.f32   {d30-d31}, [%4]!    \n"
-
-                        "vmla.f32   q8, q10, %e12[1]    \n"
-                        "vmla.f32   q9, q11, %f12[1]    \n"
-
-                        "vext.f32   q10, q12, q13, #1   \n"
-                        "vext.f32   q11, q14, q15, #1   \n"
-
-                        "vmla.f32   q8, q10, %e13[1]    \n"
-                        "vmla.f32   q9, q11, %f13[1]    \n"
-
-                        "vorr       q0, q1, q1          \n"
-                        "vorr       q2, q3, q3          \n"
-
-                        "vadd.f32   q8, q8, q9          \n"
-
-                        "vorr       q12, q13, q13       \n"
-                        "vorr       q14, q15, q15       \n"
-
-                        "subs       %0, #1              \n"
-
-                        "vst1.f32   {d16-d17}, [%5]!    \n"
-
-                        "bne        0b                  \n"
-                        "sub        %1, #16             \n"
-                        "sub        %2, #16             \n"
-                        "sub        %3, #16             \n"
-                        "sub        %4, #16             \n"
-                        : "=r"(nn),    // %0
-                        "=r"(r00),   // %1
-                        "=r"(r01),   // %2
-                        "=r"(r10),   // %3
-                        "=r"(r11),   // %4
-                        "=r"(outptr) // %5
-                        : "0"(nn),
-                        "1"(r00),
-                        "2"(r01),
-                        "3"(r10),
-                        "4"(r11),
-                        "5"(outptr),
-                        "w"(_k0), // %12
-                        "w"(_k1)  // %13
-                        : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-
-                for (; remain > 0; remain--)
-                {
-#if __ARM_NEON
-                    float32x2_t _r00 = vld1_f32(r00);
-                    float32x2_t _r01 = vld1_f32(r01);
-                    float32x4_t _r00r1 = vcombine_f32(_r00, _r01);
-                    float32x4_t _s0s1 = vmulq_f32(_r00r1, _k0);
-
-                    float32x2_t _r10 = vld1_f32(r10);
-                    float32x2_t _r11 = vld1_f32(r11);
-                    float32x4_t _r10r1 = vcombine_f32(_r10, _r11);
-                    _s0s1 = vmlaq_f32(_s0s1, _r10r1, _k1);
-
-                    float32x2_t _s = vadd_f32(vget_low_f32(_s0s1), vget_high_f32(_s0s1));
-                    _s = vpadd_f32(_s, _s);
-                    *outptr += vget_lane_f32(_s, 0);
-#else
-                    float sum = 0.f;
-
-                    sum += r00[0] * kernel0[0];
-                    sum += r00[1] * kernel0[1];
-                    sum += r01[0] * kernel0[2];
-                    sum += r01[1] * kernel0[3];
-
-                    sum += r10[0] * kernel1[0];
-                    sum += r10[1] * kernel1[1];
-                    sum += r11[0] * kernel1[2];
-                    sum += r11[1] * kernel1[3];
-
-                    *outptr += sum;
-#endif // __ARM_NEON
-
-                    r00 += 1;
-                    r01 += 1;
-                    r10 += 1;
-                    r11 += 1;
-                    outptr++;
-                }
-
-                r00 += 1;
-                r01 += 1;
-                r10 += 1;
-                r11 += 1;
-            }
-        }
-
-        for (; q < inch; q++)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch * 4 + q * 4;
-
-            const float* r0 = img0;
-            const float* r1 = img0 + w;
-
-#if __ARM_NEON
-            float32x4_t _k0 = vdupq_n_f32(kernel0[0]);
-            float32x4_t _k1 = vdupq_n_f32(kernel0[1]);
-            float32x4_t _k2 = vdupq_n_f32(kernel0[2]);
-            float32x4_t _k3 = vdupq_n_f32(kernel0[3]);
-#endif // __ARM_NEON
-
-            for (int i = 0; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw & 3;
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "prfm       pldl1keep, [%1, #128]          \n"
-                        "ld1        {v0.4s}, [%1], #16             \n"
-                        "prfm       pldl1keep, [%2, #128]          \n"
-                        "ld1        {v2.4s}, [%2], #16             \n"
-
-                        "0:                                        \n"
-                        "prfm       pldl1keep, [%3, #128]          \n"
-                        "ld1        {v9.4s}, [%3]                  \n"
-
-                        "fmul       v8.4s, v0.4s, %8.4s            \n"
-                        "fmla       v9.4s, v2.4s, %10.4s           \n"
-
-                        "prfm       pldl1keep, [%1, #128]          \n"
-                        "ld1        {v1.4s}, [%1], #16             \n"
-                        "ext        v10.16b, v0.16b, v1.16b, #4    \n"
-
-                        "fmla       v8.4s, v10.4s, %9.4s           \n"
-
-                        "prfm       pldl1keep, [%2, #128]          \n"
-                        "ld1        {v3.4s}, [%2], #16             \n"
-                        "ext        v11.16b, v2.16b, v3.16b, #4    \n"
-
-                        "fmla       v9.4s, v11.4s, %11.4s          \n"
-
-                        "orr        v0.16b, v1.16b, v1.16b         \n"
-                        "fadd       v8.4s, v8.4s, v9.4s            \n"
-                        "orr        v2.16b, v3.16b, v3.16b         \n"
-
-                        "subs       %w0, %w0, #1                   \n"
-                        "st1        {v8.4s}, [%3], #16             \n"
-                        "bne        0b                             \n"
-                        "sub        %1, %1, #16                    \n"
-                        "sub        %2, %2, #16                    \n"
-                        : "=r"(nn),    // %0
-                        "=r"(r0),    // %1
-                        "=r"(r1),    // %2
-                        "=r"(outptr) // %3
-                        : "0"(nn),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(outptr),
-                        "w"(_k0), // %8
-                        "w"(_k1), // %9
-                        "w"(_k2), // %10
-                        "w"(_k3)  // %11
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v8", "v9", "v10", "v11");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%1]!      \n"
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d4-d5}, [%2]!      \n"
-
-                        "0:                             \n"
-                        "pld        [%3, #128]          \n"
-                        "vld1.f32   {d18-d19}, [%3]     \n" // q9 = sum
-
-                        "vmul.f32   q8, q0, %q8         \n"
-                        "vmla.f32   q9, q2, %q10        \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d2-d3}, [%1]!      \n"
-                        "vext.f32   q10, q0, q1, #1     \n"
-
-                        "vmla.f32   q8, q10, %q9        \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d6-d7}, [%2]!      \n"
-                        "vext.f32   q11, q2, q3, #1     \n"
-
-                        "vmla.f32   q9, q11, %q11       \n"
-
-                        "vorr       q0, q1, q1          \n"
-                        "vadd.f32   q8, q8, q9          \n"
-                        "vorr       q2, q3, q3          \n"
-
-                        "subs       %0, #1              \n"
-                        "vst1.f32   {d16-d17}, [%3]!    \n"
-                        "bne        0b                  \n"
-                        "sub        %1, #16             \n"
-                        "sub        %2, #16             \n"
-                        : "=r"(nn),    // %0
-                        "=r"(r0),    // %1
-                        "=r"(r1),    // %2
-                        "=r"(outptr) // %3
-                        : "0"(nn),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(outptr),
-                        "w"(_k0), // %8
-                        "w"(_k1), // %9
-                        "w"(_k2), // %10
-                        "w"(_k3)  // %11
-                        : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-                float32x4_t _k0123 = vld1q_f32(kernel0);
-#endif
-
-                for (; remain > 0; remain--)
-                {
-#if __ARM_NEON
-                    float32x2_t _r0 = vld1_f32(r0);
-                    float32x2_t _r1 = vld1_f32(r1);
-                    float32x4_t _r0r1 = vcombine_f32(_r0, _r1);
-                    float32x4_t _s0s1 = vmulq_f32(_r0r1, _k0123);
-                    float32x2_t _s = vadd_f32(vget_low_f32(_s0s1), vget_high_f32(_s0s1));
-                    _s = vpadd_f32(_s, _s);
-                    *outptr += vget_lane_f32(_s, 0);
-#else
-                    float sum = 0.f;
-                    sum += r0[0] * kernel0[0];
-                    sum += r0[1] * kernel0[1];
-                    sum += r1[0] * kernel0[2];
-                    sum += r1[1] * kernel0[3];
-                    *outptr += sum;
-#endif
-
-                    r0 += 1;
-                    r1 += 1;
-                    outptr++;
-                }
-
-                r0 += 1;
-                r1 += 1;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3.h
deleted file mode 100644
index aa235054..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3.h
+++ /dev/null
@@ -1,3013 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        Mat out0 = top_blob.channel(p);
-        Mat out1 = top_blob.channel(p + 1);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        const float bias1 = bias ? bias[p + 1] : 0.f;
-
-        out0.fill(bias0);
-        out1.fill(bias1);
-
-        const float* k0 = kernel + p * inch * 9;
-        const float* k1 = kernel + (p + 1) * inch * 9;
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-            float* outptr0n = outptr0 + outw;
-            float* outptr1n = outptr1 + outw;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0;
-            const float* r1 = img0 + w;
-            const float* r2 = img0 + w * 2;
-            const float* r3 = img0 + w * 3;
-
-#if __ARM_NEON
-            float32x4_t _k00 = vld1q_f32(k0);
-            float32x4_t _k03 = vld1q_f32(k0 + 3);
-            float32x4_t _k06 = vld1q_f32(k0 + 6);
-
-            float32x4_t _k10 = vld1q_f32(k1);
-            float32x4_t _k13 = vld1q_f32(k1 + 3);
-            float32x4_t _k16 = vld1q_f32(k1 + 6);
-#endif // __ARM_NEON
-
-            int i = 0;
-
-            for (; i + 1 < outh; i += 2)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw & 3;
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%5]        \n" // r0
-                        "add    %5, %5, #16                 \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v14.4s, v15.4s}, [%8]      \n" // r3
-                        "add    %8, %8, #16                 \n"
-
-                        "ext    v10.16b, v8.16b, v9.16b, #4 \n"
-                        "ext    v11.16b, v14.16b, v15.16b, #8 \n"
-
-                        "0:                                 \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v6.4s}, [%1]               \n" // _sum0
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v7.4s}, [%2]               \n" // _sum1
-
-                        "fmla   v6.4s, v8.4s, %18.s[0]      \n"
-                        "fmla   v7.4s, v8.4s, %21.s[0]      \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v12.4s}, [%3]              \n" // _sum0n
-
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v13.4s}, [%4]              \n" // _sum1n
-
-                        "fmla   v12.4s, v14.4s, %20.s[0]    \n"
-                        "fmla   v13.4s, v14.4s, %23.s[0]    \n"
-
-                        "ext    v8.16b, v8.16b, v9.16b, #8  \n"
-                        "ext    v9.16b, v14.16b, v15.16b, #4 \n"
-
-                        "fmla   v6.4s, v10.4s, %18.s[1]     \n"
-                        "fmla   v7.4s, v10.4s, %21.s[1]     \n"
-                        "fmla   v12.4s, v11.4s, %20.s[2]    \n"
-                        "fmla   v13.4s, v11.4s, %23.s[2]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v14.4s, v15.4s}, [%6]      \n" // r1
-                        "add    %6, %6, #16                 \n"
-
-                        "fmla   v6.4s, v8.4s, %18.s[2]      \n"
-                        "fmla   v7.4s, v8.4s, %21.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, %20.s[1]     \n"
-                        "fmla   v13.4s, v9.4s, %23.s[1]     \n"
-
-                        "ext    v10.16b, v14.16b, v15.16b, #4 \n"
-
-                        "fmla   v6.4s, v14.4s, %19.s[0]     \n"
-                        "fmla   v7.4s, v14.4s, %22.s[0]     \n"
-                        "fmla   v12.4s, v14.4s, %18.s[0]    \n"
-                        "fmla   v13.4s, v14.4s, %21.s[0]    \n"
-
-                        "ext    v11.16b, v14.16b, v15.16b, #8 \n"
-
-                        "fmla   v6.4s, v10.4s, %19.s[1]     \n"
-                        "fmla   v7.4s, v10.4s, %22.s[1]     \n"
-                        "fmla   v12.4s, v10.4s, %18.s[1]    \n"
-                        "fmla   v13.4s, v10.4s, %21.s[1]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%7]        \n" // r2
-                        "add    %7, %7, #16                 \n"
-
-                        "fmla   v6.4s, v11.4s, %19.s[2]     \n"
-                        "fmla   v7.4s, v11.4s, %22.s[2]     \n"
-                        "fmla   v12.4s, v11.4s, %18.s[2]    \n"
-                        "fmla   v13.4s, v11.4s, %21.s[2]    \n"
-
-                        "ext    v10.16b, v8.16b, v9.16b, #4 \n"
-
-                        "fmla   v6.4s, v8.4s, %20.s[0]      \n"
-                        "fmla   v7.4s, v8.4s, %23.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, %19.s[0]     \n"
-                        "fmla   v13.4s, v8.4s, %22.s[0]     \n"
-
-                        "ext    v11.16b, v8.16b, v9.16b, #8 \n"
-
-                        "fmla   v6.4s, v10.4s, %20.s[1]     \n"
-                        "fmla   v7.4s, v10.4s, %23.s[1]     \n"
-                        "fmla   v12.4s, v10.4s, %19.s[1]    \n"
-                        "fmla   v13.4s, v10.4s, %22.s[1]    \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%5]        \n" // r0
-                        "add    %5, %5, #16                 \n"
-
-                        "fmla   v6.4s, v11.4s, %20.s[2]     \n"
-                        "fmla   v7.4s, v11.4s, %23.s[2]     \n"
-                        "fmla   v12.4s, v11.4s, %19.s[2]    \n"
-                        "fmla   v13.4s, v11.4s, %22.s[2]    \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v14.4s, v15.4s}, [%8]      \n" // r3
-                        "add    %8, %8, #16                 \n"
-
-                        "ext    v10.16b, v8.16b, v9.16b, #4 \n"
-
-                        "st1    {v6.4s}, [%1], #16          \n"
-                        "st1    {v7.4s}, [%2], #16          \n"
-
-                        "ext    v11.16b, v14.16b, v15.16b, #8 \n"
-
-                        "st1    {v12.4s}, [%3], #16         \n"
-                        "st1    {v13.4s}, [%4], #16         \n"
-
-                        "subs   %w0, %w0, #1                \n"
-                        "bne    0b                          \n"
-
-                        "sub    %5, %5, #16                 \n"
-                        "sub    %8, %8, #16                 \n"
-                        : "=r"(nn),       // %0
-                        "=r"(outptr0),  // %1
-                        "=r"(outptr1),  // %2
-                        "=r"(outptr0n), // %3
-                        "=r"(outptr1n), // %4
-                        "=r"(r0),       // %5
-                        "=r"(r1),       // %6
-                        "=r"(r2),       // %7
-                        "=r"(r3)        // %8
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(outptr1),
-                        "3"(outptr0n),
-                        "4"(outptr1n),
-                        "5"(r0),
-                        "6"(r1),
-                        "7"(r2),
-                        "8"(r3),
-                        "w"(_k00), // %18
-                        "w"(_k03), // %19
-                        "w"(_k06), // %20
-                        "w"(_k10), // %21
-                        "w"(_k13), // %22
-                        "w"(_k16)  // %23
-                        : "cc", "memory", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-
-                        "pld        [%5, #192]          \n"
-                        "vld1.f32   {d16-d18}, [%5 :64] \n" // r0
-                        "add        %5, #16             \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.f32   {d28-d30}, [%8]     \n" // r3
-                        "add        %8, #16             \n"
-
-                        "vext.32    q10, q8, q9, #1     \n"
-                        "vext.32    q11, q14, q15, #2   \n"
-
-                        "0:                             \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d12-d13}, [%1 :64] \n" // _sum0
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d14-d15}, [%2 :64] \n" // _sum1
-
-                        "vmla.f32   q6, q8, %e18[0]     \n"
-                        "vmla.f32   q7, q8, %e21[0]     \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.f32   {d24-d25}, [%3]     \n" // _sum0n
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.f32   {d26-d27}, [%4]     \n" // _sum1n
-
-                        "vmla.f32   q12, q14, %e20[0]   \n"
-                        "vmla.f32   q13, q14, %e23[0]   \n"
-
-                        "vext.32    q8, q8, q9, #2      \n"
-                        "vext.32    q9, q14, q15, #1    \n"
-
-                        "vmla.f32   q6, q10, %e18[1]    \n"
-                        "vmla.f32   q7, q10, %e21[1]    \n"
-                        "vmla.f32   q12, q11, %f20[0]   \n"
-                        "vmla.f32   q13, q11, %f23[0]   \n"
-
-                        "pld        [%6, #192]          \n"
-                        "vld1.f32   {d28-d30}, [%6]     \n" // r1
-                        "add        %6, #16             \n"
-
-                        "vmla.f32   q6, q8, %f18[0]     \n"
-                        "vmla.f32   q7, q8, %f21[0]     \n"
-                        "vmla.f32   q12, q9, %e20[1]    \n"
-                        "vmla.f32   q13, q9, %e23[1]    \n"
-
-                        "vext.32    q10, q14, q15, #1   \n"
-
-                        "vmla.f32   q6, q14, %e19[0]    \n"
-                        "vmla.f32   q7, q14, %e22[0]    \n"
-                        "vmla.f32   q12, q14, %e18[0]   \n"
-                        "vmla.f32   q13, q14, %e21[0]   \n"
-
-                        "vext.32    q11, q14, q15, #2   \n"
-
-                        "vmla.f32   q6, q10, %e19[1]    \n"
-                        "vmla.f32   q7, q10, %e22[1]    \n"
-                        "vmla.f32   q12, q10, %e18[1]   \n"
-                        "vmla.f32   q13, q10, %e21[1]   \n"
-
-                        "pld        [%7, #192]          \n"
-                        "vld1.f32   {d16-d18}, [%7 :64] \n" // r2
-                        "add        %7, #16             \n"
-
-                        "vmla.f32   q6, q11, %f19[0]    \n"
-                        "vmla.f32   q7, q11, %f22[0]    \n"
-                        "vmla.f32   q12, q11, %f18[0]   \n"
-                        "vmla.f32   q13, q11, %f21[0]   \n"
-
-                        "vext.32    q10, q8, q9, #1     \n"
-
-                        "vmla.f32   q6, q8, %e20[0]     \n"
-                        "vmla.f32   q7, q8, %e23[0]     \n"
-                        "vmla.f32   q12, q8, %e19[0]    \n"
-                        "vmla.f32   q13, q8, %e22[0]    \n"
-
-                        "vext.32    q11, q8, q9, #2     \n"
-
-                        "vmla.f32   q6, q10, %e20[1]    \n"
-                        "vmla.f32   q7, q10, %e23[1]    \n"
-                        "vmla.f32   q12, q10, %e19[1]   \n"
-                        "vmla.f32   q13, q10, %e22[1]   \n"
-
-                        "pld        [%5, #192]          \n"
-                        "vld1.f32   {d16-d18}, [%5 :64] \n" // r0
-                        "add        %5, #16             \n"
-
-                        "vmla.f32   q6, q11, %f20[0]    \n"
-                        "vmla.f32   q7, q11, %f23[0]    \n"
-                        "vmla.f32   q12, q11, %f19[0]   \n"
-                        "vmla.f32   q13, q11, %f22[0]   \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.f32   {d28-d30}, [%8]     \n" // r3
-                        "add        %8, #16             \n"
-
-                        "vext.32    q10, q8, q9, #1     \n"
-
-                        "vst1.f32   {d12-d13}, [%1 : 64]!\n"
-                        "vst1.f32   {d14-d15}, [%2 : 64]!\n"
-
-                        "vext.32    q11, q14, q15, #2   \n"
-
-                        "vst1.f32   {d24-d25}, [%3]!    \n"
-                        "vst1.f32   {d26-d27}, [%4]!    \n"
-
-                        "subs       %0, #1              \n"
-                        "bne        0b                  \n"
-
-                        "sub        %5, #16             \n"
-                        "sub        %8, #16             \n"
-                        : "=r"(nn),       // %0
-                        "=r"(outptr0),  // %1
-                        "=r"(outptr1),  // %2
-                        "=r"(outptr0n), // %3
-                        "=r"(outptr1n), // %4
-                        "=r"(r0),       // %5
-                        "=r"(r1),       // %6
-                        "=r"(r2),       // %7
-                        "=r"(r3)        // %8
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(outptr1),
-                        "3"(outptr0n),
-                        "4"(outptr1n),
-                        "5"(r0),
-                        "6"(r1),
-                        "7"(r2),
-                        "8"(r3),
-                        "w"(_k00), // %18
-                        "w"(_k03), // %19
-                        "w"(_k06), // %20
-                        "w"(_k10), // %21
-                        "w"(_k13), // %22
-                        "w"(_k16)  // %23
-                        : "cc", "memory", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-#if __ARM_NEON
-                    float32x4_t _r00 = vld1q_f32(r0);
-                    float32x4_t _r10 = vld1q_f32(r1);
-                    float32x4_t _r20 = vld1q_f32(r2);
-                    float32x4_t _r30 = vld1q_f32(r3);
-
-                    float32x4_t _sum0 = vmulq_f32(_r00, _k00);
-                    float32x4_t _sum1 = vmulq_f32(_r00, _k10);
-                    _sum0 = vmlaq_f32(_sum0, _r10, _k03);
-                    _sum1 = vmlaq_f32(_sum1, _r10, _k13);
-                    _sum0 = vmlaq_f32(_sum0, _r20, _k06);
-                    _sum1 = vmlaq_f32(_sum1, _r20, _k16);
-
-                    float32x4_t _sum0n = vmulq_f32(_r10, _k00);
-                    float32x4_t _sum1n = vmulq_f32(_r10, _k10);
-                    _sum0n = vmlaq_f32(_sum0n, _r20, _k03);
-                    _sum1n = vmlaq_f32(_sum1n, _r20, _k13);
-                    _sum0n = vmlaq_f32(_sum0n, _r30, _k06);
-                    _sum1n = vmlaq_f32(_sum1n, _r30, _k16);
-
-                    _sum0 = vsetq_lane_f32(*outptr0, _sum0, 3);
-                    _sum1 = vsetq_lane_f32(*outptr1, _sum1, 3);
-                    _sum0n = vsetq_lane_f32(*outptr0n, _sum0n, 3);
-                    _sum1n = vsetq_lane_f32(*outptr1n, _sum1n, 3);
-#if __aarch64__
-                    *outptr0 = vaddvq_f32(_sum0);
-                    *outptr1 = vaddvq_f32(_sum1);
-                    *outptr0n = vaddvq_f32(_sum0n);
-                    *outptr1n = vaddvq_f32(_sum1n);
-#else
-                    float32x2_t _ss0 = vadd_f32(vget_low_f32(_sum0), vget_high_f32(_sum0));
-                    float32x2_t _ss1 = vadd_f32(vget_low_f32(_sum1), vget_high_f32(_sum1));
-                    float32x2_t _ss0n = vadd_f32(vget_low_f32(_sum0n), vget_high_f32(_sum0n));
-                    float32x2_t _ss1n = vadd_f32(vget_low_f32(_sum1n), vget_high_f32(_sum1n));
-
-                    float32x2_t _ss01 = vpadd_f32(_ss0, _ss1);
-                    float32x2_t _ss01n = vpadd_f32(_ss0n, _ss1n);
-
-                    *outptr0 = vget_lane_f32(_ss01, 0);
-                    *outptr1 = vget_lane_f32(_ss01, 1);
-                    *outptr0n = vget_lane_f32(_ss01n, 0);
-                    *outptr1n = vget_lane_f32(_ss01n, 1);
-#endif // __aarch64__
-#else
-                    float sum0 = 0.f;
-                    float sum0n = 0.f;
-                    float sum1 = 0.f;
-                    float sum1n = 0.f;
-
-                    sum0 += r0[0] * k0[0];
-                    sum0 += r0[1] * k0[1];
-                    sum0 += r0[2] * k0[2];
-                    sum0 += r1[0] * k0[3];
-                    sum0 += r1[1] * k0[4];
-                    sum0 += r1[2] * k0[5];
-                    sum0 += r2[0] * k0[6];
-                    sum0 += r2[1] * k0[7];
-                    sum0 += r2[2] * k0[8];
-
-                    sum1 += r0[0] * k1[0];
-                    sum1 += r0[1] * k1[1];
-                    sum1 += r0[2] * k1[2];
-                    sum1 += r1[0] * k1[3];
-                    sum1 += r1[1] * k1[4];
-                    sum1 += r1[2] * k1[5];
-                    sum1 += r2[0] * k1[6];
-                    sum1 += r2[1] * k1[7];
-                    sum1 += r2[2] * k1[8];
-
-                    sum0n += r1[0] * k0[0];
-                    sum0n += r1[1] * k0[1];
-                    sum0n += r1[2] * k0[2];
-                    sum0n += r2[0] * k0[3];
-                    sum0n += r2[1] * k0[4];
-                    sum0n += r2[2] * k0[5];
-                    sum0n += r3[0] * k0[6];
-                    sum0n += r3[1] * k0[7];
-                    sum0n += r3[2] * k0[8];
-
-                    sum1n += r1[0] * k1[0];
-                    sum1n += r1[1] * k1[1];
-                    sum1n += r1[2] * k1[2];
-                    sum1n += r2[0] * k1[3];
-                    sum1n += r2[1] * k1[4];
-                    sum1n += r2[2] * k1[5];
-                    sum1n += r3[0] * k1[6];
-                    sum1n += r3[1] * k1[7];
-                    sum1n += r3[2] * k1[8];
-
-                    *outptr0 += sum0;
-                    *outptr1 += sum1;
-                    *outptr0n += sum0n;
-                    *outptr1n += sum1n;
-#endif // __ARM_NEON
-                    r0++;
-                    r1++;
-                    r2++;
-                    r3++;
-                    outptr0++;
-                    outptr1++;
-                    outptr0n++;
-                    outptr1n++;
-                }
-
-                r0 += 2 + w;
-                r1 += 2 + w;
-                r2 += 2 + w;
-                r3 += 2 + w;
-
-                outptr0 += outw;
-                outptr1 += outw;
-                outptr0n += outw;
-                outptr1n += outw;
-            }
-
-            for (; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw & 3;
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "0:                                 \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%3]        \n" // r0
-                        "add    %3, %3, #16                 \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v6.4s}, [%1]               \n" // _sum0
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v7.4s}, [%2]               \n" // _sum1
-
-                        "fmul   v14.4s, v8.4s, %12.s[0]     \n"
-                        "fmul   v15.4s, v8.4s, %15.s[0]     \n"
-
-                        "ext    v10.16b, v8.16b, v9.16b, #4 \n"
-                        "ext    v11.16b, v8.16b, v9.16b, #8 \n"
-
-                        "fmla   v6.4s, v10.4s, %12.s[1]     \n"
-                        "fmla   v7.4s, v10.4s, %15.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%4]        \n" // r1
-                        "add    %4, %4, #16                 \n"
-
-                        "fmla   v14.4s, v11.4s, %12.s[2]    \n"
-                        "fmla   v15.4s, v11.4s, %15.s[2]    \n"
-
-                        "fmla   v6.4s, v8.4s, %13.s[0]      \n"
-                        "fmla   v7.4s, v8.4s, %16.s[0]      \n"
-
-                        "ext    v10.16b, v8.16b, v9.16b, #4 \n"
-                        "ext    v11.16b, v8.16b, v9.16b, #8 \n"
-
-                        "fmla   v14.4s, v10.4s, %13.s[1]    \n"
-                        "fmla   v15.4s, v10.4s, %16.s[1]    \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%5]        \n" // r2
-                        "add    %5, %5, #16                 \n"
-
-                        "fmla   v6.4s, v11.4s, %13.s[2]     \n"
-                        "fmla   v7.4s, v11.4s, %16.s[2]     \n"
-
-                        "fmla   v14.4s, v8.4s, %14.s[0]     \n"
-                        "fmla   v15.4s, v8.4s, %17.s[0]     \n"
-
-                        "ext    v10.16b, v8.16b, v9.16b, #4 \n"
-                        "ext    v11.16b, v8.16b, v9.16b, #8 \n"
-
-                        "fmla   v6.4s, v10.4s, %14.s[1]     \n"
-                        "fmla   v7.4s, v10.4s, %17.s[1]     \n"
-
-                        "fmla   v14.4s, v11.4s, %14.s[2]    \n"
-                        "fmla   v15.4s, v11.4s, %17.s[2]    \n"
-
-                        "fadd   v6.4s, v6.4s, v14.4s        \n"
-                        "fadd   v7.4s, v7.4s, v15.4s        \n"
-
-                        "st1    {v6.4s}, [%1], #16          \n"
-                        "st1    {v7.4s}, [%2], #16          \n"
-
-                        "subs   %w0, %w0, #1                \n"
-                        "bne    0b                          \n"
-
-                        : "=r"(nn),      // %0
-                        "=r"(outptr0), // %1
-                        "=r"(outptr1), // %2
-                        "=r"(r0),      // %3
-                        "=r"(r1),      // %4
-                        "=r"(r2)       // %5
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(outptr1),
-                        "3"(r0),
-                        "4"(r1),
-                        "5"(r2),
-                        "w"(_k00), // %12
-                        "w"(_k03), // %13
-                        "w"(_k06), // %14
-                        "w"(_k10), // %15
-                        "w"(_k13), // %16
-                        "w"(_k16)  // %17
-                        : "cc", "memory", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "0:                             \n"
-
-                        "pld        [%3, #192]          \n"
-                        "vld1.f32   {d16-d18}, [%3]     \n" // r0
-                        "add        %3, #16             \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d12-d13}, [%1]     \n" // _sum0
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d14-d15}, [%2]     \n" // _sum1
-
-                        "vmul.f32   q14, q8, %e12[0]    \n"
-                        "vmul.f32   q15, q8, %e15[0]    \n"
-
-                        "vext.32    q10, q8, q9, #1     \n"
-                        "vext.32    q11, q8, q9, #2     \n"
-
-                        "vmla.f32   q6, q10, %e12[1]    \n"
-                        "vmla.f32   q7, q10, %e15[1]    \n"
-
-                        "pld        [%4, #192]          \n"
-                        "vld1.f32   {d16-d18}, [%4]     \n" // r1
-                        "add        %4, #16             \n"
-
-                        "vmla.f32   q14, q11, %f12[0]   \n"
-                        "vmla.f32   q15, q11, %f15[0]   \n"
-
-                        "vmla.f32   q6, q8, %e13[0]     \n"
-                        "vmla.f32   q7, q8, %e16[0]     \n"
-
-                        "vext.32    q10, q8, q9, #1     \n"
-                        "vext.32    q11, q8, q9, #2     \n"
-
-                        "vmla.f32   q14, q10, %e13[1]   \n"
-                        "vmla.f32   q15, q10, %e16[1]   \n"
-
-                        "pld        [%5, #192]          \n"
-                        "vld1.f32   {d16-d18}, [%5]     \n" // r2
-                        "add        %5, #16             \n"
-
-                        "vmla.f32   q6, q11, %f13[0]    \n"
-                        "vmla.f32   q7, q11, %f16[0]    \n"
-
-                        "vmla.f32   q14, q8, %e14[0]    \n"
-                        "vmla.f32   q15, q8, %e17[0]    \n"
-
-                        "vext.32    q10, q8, q9, #1     \n"
-                        "vext.32    q11, q8, q9, #2     \n"
-
-                        "vmla.f32   q6, q10, %e14[1]    \n"
-                        "vmla.f32   q7, q10, %e17[1]    \n"
-
-                        "vmla.f32   q14, q11, %f14[0]   \n"
-                        "vmla.f32   q15, q11, %f17[0]   \n"
-
-                        "vadd.f32   q6, q6, q14         \n"
-                        "vadd.f32   q7, q7, q15         \n"
-
-                        "vst1.f32   {d12-d13}, [%1]!    \n"
-
-                        "vst1.f32   {d14-d15}, [%2]!    \n"
-
-                        "subs       %0, #1              \n"
-                        "bne        0b                  \n"
-
-                        : "=r"(nn),      // %0
-                        "=r"(outptr0), // %1
-                        "=r"(outptr1), // %2
-                        "=r"(r0),      // %3
-                        "=r"(r1),      // %4
-                        "=r"(r2)       // %5
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(outptr1),
-                        "3"(r0),
-                        "4"(r1),
-                        "5"(r2),
-                        "w"(_k00), // %12
-                        "w"(_k03), // %13
-                        "w"(_k06), // %14
-                        "w"(_k10), // %15
-                        "w"(_k13), // %16
-                        "w"(_k16)  // %17
-                        : "cc", "memory", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-#if __ARM_NEON
-                    float32x4_t _r00 = vld1q_f32(r0);
-                    float32x4_t _r10 = vld1q_f32(r1);
-                    float32x4_t _r20 = vld1q_f32(r2);
-
-                    float32x4_t _sum0 = vmulq_f32(_r00, _k00);
-                    float32x4_t _sum1 = vmulq_f32(_r00, _k10);
-                    _sum0 = vmlaq_f32(_sum0, _r10, _k03);
-                    _sum1 = vmlaq_f32(_sum1, _r10, _k13);
-                    _sum0 = vmlaq_f32(_sum0, _r20, _k06);
-                    _sum1 = vmlaq_f32(_sum1, _r20, _k16);
-
-                    _sum0 = vsetq_lane_f32(*outptr0, _sum0, 3);
-                    _sum1 = vsetq_lane_f32(*outptr1, _sum1, 3);
-#if __aarch64__
-                    *outptr0 = vaddvq_f32(_sum0);
-                    *outptr1 = vaddvq_f32(_sum1);
-#else
-                    float32x2_t _ss0 = vadd_f32(vget_low_f32(_sum0), vget_high_f32(_sum0));
-                    float32x2_t _ss1 = vadd_f32(vget_low_f32(_sum1), vget_high_f32(_sum1));
-                    float32x2_t _ss01 = vpadd_f32(_ss0, _ss1);
-
-                    *outptr0 = vget_lane_f32(_ss01, 0);
-                    *outptr1 = vget_lane_f32(_ss01, 1);
-#endif // __aarch64__
-#else
-                    float sum0 = 0.f;
-                    float sum1 = 0.f;
-
-                    sum0 += r0[0] * k0[0];
-                    sum0 += r0[1] * k0[1];
-                    sum0 += r0[2] * k0[2];
-                    sum0 += r1[0] * k0[3];
-                    sum0 += r1[1] * k0[4];
-                    sum0 += r1[2] * k0[5];
-                    sum0 += r2[0] * k0[6];
-                    sum0 += r2[1] * k0[7];
-                    sum0 += r2[2] * k0[8];
-
-                    sum1 += r0[0] * k1[0];
-                    sum1 += r0[1] * k1[1];
-                    sum1 += r0[2] * k1[2];
-                    sum1 += r1[0] * k1[3];
-                    sum1 += r1[1] * k1[4];
-                    sum1 += r1[2] * k1[5];
-                    sum1 += r2[0] * k1[6];
-                    sum1 += r2[1] * k1[7];
-                    sum1 += r2[2] * k1[8];
-
-                    *outptr0 += sum0;
-                    *outptr1 += sum1;
-#endif // __ARM_NEON
-                    r0++;
-                    r1++;
-                    r2++;
-                    outptr0++;
-                    outptr1++;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9;
-            k1 += 9;
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        const float* kernel0 = kernel + p * inch * 9;
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr = out;
-            float* outptr2 = outptr + outw;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0;
-            const float* r1 = img0 + w;
-            const float* r2 = img0 + w * 2;
-            const float* r3 = img0 + w * 3;
-
-#if __ARM_NEON
-            float32x4_t _k0123 = vld1q_f32(kernel0);
-            float32x4_t _k3456 = vld1q_f32(kernel0 + 3);
-            float32x4_t _k6789 = vld1q_f32(kernel0 + 6);
-#else
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-#endif // __ARM_NEON
-
-            int i = 0;
-
-            for (; i + 1 < outh; i += 2)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw & 3;
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v9.4s, v10.4s}, [%3]       \n" // r0
-                        "add    %3, %3, #16                 \n"
-
-                        "ext    v11.16b, v9.16b, v10.16b, #4 \n"
-                        "ext    v12.16b, v9.16b, v10.16b, #8 \n"
-
-                        "0:                                 \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v7.4s}, [%1]               \n" // _sum
-
-                        "fmla   v7.4s, v9.4s, %14.s[0]      \n"
-                        "fmul   v6.4s, v11.4s, %14.s[1]     \n"
-                        "fmul   v13.4s, v12.4s, %14.s[2]    \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v9.4s, v10.4s}, [%4]       \n" // r1
-                        "add    %4, %4, #16                 \n"
-
-                        "fmla   v7.4s, v9.4s, %15.s[0]      \n"
-
-                        "ext    v11.16b, v9.16b, v10.16b, #4 \n"
-                        "ext    v12.16b, v9.16b, v10.16b, #8 \n"
-
-                        "fmla   v6.4s, v11.4s, %15.s[1]     \n"
-                        "fmla   v13.4s, v12.4s, %15.s[2]    \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v8.4s}, [%2]               \n" // _sum2
-
-                        "fmla   v8.4s, v9.4s, %14.s[0]      \n"
-                        "fmul   v14.4s, v11.4s, %14.s[1]    \n"
-                        "fmul   v15.4s, v12.4s, %14.s[2]    \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v9.4s, v10.4s}, [%5]       \n" // r2
-                        "add    %5, %5, #16                 \n"
-
-                        "fmla   v7.4s, v9.4s, %16.s[0]      \n"
-
-                        "ext    v11.16b, v9.16b, v10.16b, #4 \n"
-                        "ext    v12.16b, v9.16b, v10.16b, #8 \n"
-
-                        "fmla   v6.4s, v11.4s, %16.s[1]     \n"
-                        "fmla   v13.4s, v12.4s, %16.s[2]    \n"
-
-                        "fmla   v8.4s, v9.4s, %15.s[0]      \n"
-                        "fmla   v14.4s, v11.4s, %15.s[1]    \n"
-                        "fmla   v15.4s, v12.4s, %15.s[2]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v9.4s, v10.4s}, [%6]       \n" // r3
-                        "add    %6, %6, #16                 \n"
-
-                        "fmla   v8.4s, v9.4s, %16.s[0]      \n"
-
-                        "ext    v11.16b, v9.16b, v10.16b, #4 \n"
-                        "ext    v12.16b, v9.16b, v10.16b, #8 \n"
-
-                        "fmla   v14.4s, v11.4s, %16.s[1]    \n"
-                        "fmla   v15.4s, v12.4s, %16.s[2]    \n"
-
-                        "fadd   v7.4s, v7.4s, v6.4s         \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v9.4s, v10.4s}, [%3]       \n" // r0
-
-                        "fadd   v8.4s, v8.4s, v14.4s        \n"
-                        "fadd   v7.4s, v7.4s, v13.4s        \n"
-                        "fadd   v8.4s, v8.4s, v15.4s        \n"
-
-                        "ext    v11.16b, v9.16b, v10.16b, #4 \n"
-                        "ext    v12.16b, v9.16b, v10.16b, #8 \n"
-
-                        "add    %3, %3, #16                 \n"
-
-                        "st1    {v7.4s}, [%1], #16          \n"
-                        "st1    {v8.4s}, [%2], #16          \n"
-
-                        "subs   %w0, %w0, #1                \n"
-                        "bne    0b                          \n"
-
-                        "sub    %3, %3, #16                 \n"
-                        : "=r"(nn),      // %0
-                        "=r"(outptr),  // %1
-                        "=r"(outptr2), // %2
-                        "=r"(r0),      // %3
-                        "=r"(r1),      // %4
-                        "=r"(r2),      // %5
-                        "=r"(r3)       // %6
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(outptr2),
-                        "3"(r0),
-                        "4"(r1),
-                        "5"(r2),
-                        "6"(r3),
-                        "w"(_k0123), // %14
-                        "w"(_k3456), // %15
-                        "w"(_k6789)  // %16
-                        : "cc", "memory", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "pld        [%3, #192]          \n"
-                        "vld1.f32   {d18-d20}, [%3 :64] \n" // r0
-                        "add        %3, #16             \n"
-
-                        "vext.32    q11, q9, q10, #1    \n"
-                        "vext.32    q12, q9, q10, #2    \n"
-
-                        "0:                             \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d14-d15}, [%1 :64] \n" // _sum
-
-                        "vmla.f32   q7, q9, %e14[0]     \n"
-                        "vmul.f32   q6, q11, %e14[1]    \n"
-                        "vmul.f32   q13, q12, %f14[0]   \n"
-
-                        "pld        [%4, #192]          \n"
-                        "vld1.f32   {d18-d20}, [%4]     \n" // r1
-                        "add        %4, #16             \n"
-
-                        "vmla.f32   q7, q9, %e15[0]     \n"
-
-                        "vext.32    q11, q9, q10, #1    \n"
-                        "vext.32    q12, q9, q10, #2    \n"
-
-                        "vmla.f32   q6, q11, %e15[1]    \n"
-                        "vmla.f32   q13, q12, %f15[0]   \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d16-d17}, [%2]     \n" // _sum2
-
-                        "vmla.f32   q8, q9, %e14[0]     \n"
-                        "vmul.f32   q14, q11, %e14[1]   \n"
-                        "vmul.f32   q15, q12, %f14[0]   \n"
-
-                        "pld        [%5, #192]          \n"
-                        "vld1.f32   {d18-d20}, [%5 :64] \n" // r2
-                        "add        %5, #16             \n"
-
-                        "vmla.f32   q7, q9, %e16[0]     \n"
-
-                        "vext.32    q11, q9, q10, #1    \n"
-                        "vext.32    q12, q9, q10, #2    \n"
-
-                        "vmla.f32   q6, q11, %e16[1]    \n"
-                        "vmla.f32   q13, q12, %f16[0]   \n"
-
-                        "vmla.f32   q8, q9, %e15[0]     \n"
-                        "vmla.f32   q14, q11, %e15[1]   \n"
-                        "vmla.f32   q15, q12, %f15[0]   \n"
-
-                        "pld        [%6, #192]          \n"
-                        "vld1.f32   {d18-d20}, [%6]     \n" // r3
-                        "add        %6, #16             \n"
-
-                        "vmla.f32   q8, q9, %e16[0]     \n"
-
-                        "vext.32    q11, q9, q10, #1    \n"
-                        "vext.32    q12, q9, q10, #2    \n"
-
-                        "vmla.f32   q14, q11, %e16[1]   \n"
-                        "vmla.f32   q15, q12, %f16[0]   \n"
-
-                        "vadd.f32   q7, q7, q6          \n"
-
-                        "pld        [%3, #192]          \n"
-                        "vld1.f32   {d18-d20}, [%3 :64] \n" // r0
-
-                        "vadd.f32   q8, q8, q14         \n"
-                        "vadd.f32   q7, q7, q13         \n"
-                        "vadd.f32   q8, q8, q15         \n"
-
-                        "vext.32    q11, q9, q10, #1    \n"
-                        "vext.32    q12, q9, q10, #2    \n"
-
-                        "add        %3, #16             \n"
-
-                        "vst1.f32   {d14-d15}, [%1]!    \n"
-                        "vst1.f32   {d16-d17}, [%2]!    \n"
-
-                        "subs       %0, #1              \n"
-                        "bne        0b                  \n"
-
-                        "sub        %3, #16             \n"
-                        : "=r"(nn),      // %0
-                        "=r"(outptr),  // %1
-                        "=r"(outptr2), // %2
-                        "=r"(r0),      // %3
-                        "=r"(r1),      // %4
-                        "=r"(r2),      // %5
-                        "=r"(r3)       // %6
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(outptr2),
-                        "3"(r0),
-                        "4"(r1),
-                        "5"(r2),
-                        "6"(r3),
-                        "w"(_k0123), // %14
-                        "w"(_k3456), // %15
-                        "w"(_k6789)  // %16
-                        : "cc", "memory", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-#if __ARM_NEON
-                    float32x4_t _r00 = vld1q_f32(r0);
-                    float32x4_t _r10 = vld1q_f32(r1);
-                    float32x4_t _r20 = vld1q_f32(r2);
-                    float32x4_t _r30 = vld1q_f32(r3);
-
-                    float32x4_t _sum = vmulq_f32(_r00, _k0123);
-                    _sum = vmlaq_f32(_sum, _r10, _k3456);
-                    _sum = vmlaq_f32(_sum, _r20, _k6789);
-
-                    float32x4_t _sum2 = vmulq_f32(_r10, _k0123);
-                    _sum2 = vmlaq_f32(_sum2, _r20, _k3456);
-                    _sum2 = vmlaq_f32(_sum2, _r30, _k6789);
-
-                    _sum = vsetq_lane_f32(*outptr, _sum, 3);
-                    _sum2 = vsetq_lane_f32(*outptr2, _sum2, 3);
-
-#if __aarch64__
-                    *outptr = vaddvq_f32(_sum);
-                    *outptr2 = vaddvq_f32(_sum2);
-#else
-                    float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-                    float32x2_t _ss2 = vadd_f32(vget_low_f32(_sum2), vget_high_f32(_sum2));
-
-                    float32x2_t _sss2 = vpadd_f32(_ss, _ss2);
-
-                    *outptr = vget_lane_f32(_sss2, 0);
-                    *outptr2 = vget_lane_f32(_sss2, 1);
-#endif // __aarch64__
-#else
-                    float sum = 0;
-                    float sum2 = 0;
-
-                    sum += r0[0] * k0[0];
-                    sum += r0[1] * k0[1];
-                    sum += r0[2] * k0[2];
-                    sum += r1[0] * k1[0];
-                    sum += r1[1] * k1[1];
-                    sum += r1[2] * k1[2];
-                    sum += r2[0] * k2[0];
-                    sum += r2[1] * k2[1];
-                    sum += r2[2] * k2[2];
-
-                    sum2 += r1[0] * k0[0];
-                    sum2 += r1[1] * k0[1];
-                    sum2 += r1[2] * k0[2];
-                    sum2 += r2[0] * k1[0];
-                    sum2 += r2[1] * k1[1];
-                    sum2 += r2[2] * k1[2];
-                    sum2 += r3[0] * k2[0];
-                    sum2 += r3[1] * k2[1];
-                    sum2 += r3[2] * k2[2];
-
-                    *outptr += sum;
-                    *outptr2 += sum2;
-#endif
-                    r0++;
-                    r1++;
-                    r2++;
-                    r3++;
-                    outptr++;
-                    outptr2++;
-                }
-
-                r0 += 2 + w;
-                r1 += 2 + w;
-                r2 += 2 + w;
-                r3 += 2 + w;
-
-                outptr += outw;
-                outptr2 += outw;
-            }
-
-            for (; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw & 3;
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%2]        \n" // r0
-                        "add    %2, %2, #16                 \n"
-
-                        "ext    v10.16b, v8.16b, v9.16b, #4 \n"
-                        "ext    v11.16b, v8.16b, v9.16b, #8 \n"
-
-                        "0:                                 \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v7.4s}, [%1]               \n" // _sum
-
-                        "fmla   v7.4s, v8.4s, %10.s[0]      \n"
-                        "fmul   v13.4s, v10.4s, %10.s[1]    \n"
-                        "fmul   v14.4s, v11.4s, %10.s[2]    \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%3]        \n" // r1
-                        "add    %3, %3, #16                 \n"
-
-                        "fmla   v7.4s, v8.4s, %11.s[0]      \n"
-
-                        "ext    v10.16b, v8.16b, v9.16b, #4 \n"
-                        "ext    v11.16b, v8.16b, v9.16b, #8 \n"
-
-                        "fmla   v13.4s, v10.4s, %11.s[1]    \n"
-                        "fmla   v14.4s, v11.4s, %11.s[2]    \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%4]        \n" // r2
-                        "add    %4, %4, #16                 \n"
-
-                        "fmla   v7.4s, v8.4s, %12.s[0]      \n"
-
-                        "ext    v10.16b, v8.16b, v9.16b, #4 \n"
-                        "ext    v11.16b, v8.16b, v9.16b, #8 \n"
-
-                        "fmla   v13.4s, v10.4s, %12.s[1]    \n"
-                        "fmla   v14.4s, v11.4s, %12.s[2]    \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%2]        \n" // r0
-                        "add    %2, %2, #16                 \n"
-
-                        "fadd   v7.4s, v7.4s, v13.4s        \n"
-                        "fadd   v7.4s, v7.4s, v14.4s        \n"
-
-                        "ext    v10.16b, v8.16b, v9.16b, #4 \n"
-                        "ext    v11.16b, v8.16b, v9.16b, #8 \n"
-
-                        "st1    {v7.4s}, [%1], #16          \n"
-
-                        "subs   %w0, %w0, #1                \n"
-                        "bne    0b                          \n"
-
-                        "sub    %2, %2, #16                 \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2)      // %4
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k0123), // %10
-                        "w"(_k3456), // %11
-                        "w"(_k6789)  // %12
-                        : "cc", "memory", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "pld        [%2, #192]          \n"
-                        "vld1.f32   {d16-d18}, [%2]     \n" // r0
-                        "add        %2, #16             \n"
-
-                        "vext.32    q10, q8, q9, #1     \n"
-                        "vext.32    q11, q8, q9, #2     \n"
-
-                        "0:                             \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d14-d15}, [%1]     \n" // _sum
-
-                        "vmla.f32   q7, q8, %e10[0]     \n"
-                        "vmul.f32   q13, q10, %e10[1]   \n"
-                        "vmul.f32   q14, q11, %f10[0]   \n"
-
-                        "pld        [%3, #192]          \n"
-                        "vld1.f32   {d16-d18}, [%3]     \n" // r1
-                        "add        %3, #16             \n"
-
-                        "vmla.f32   q7, q8, %e11[0]     \n"
-
-                        "vext.32    q10, q8, q9, #1     \n"
-                        "vext.32    q11, q8, q9, #2     \n"
-
-                        "vmla.f32   q13, q10, %e11[1]   \n"
-                        "vmla.f32   q14, q11, %f11[0]   \n"
-
-                        "pld        [%4, #192]          \n"
-                        "vld1.f32   {d16-d18}, [%4]     \n" // r2
-                        "add        %4, #16             \n"
-
-                        "vmla.f32   q7, q8, %e12[0]     \n"
-
-                        "vext.32    q10, q8, q9, #1     \n"
-                        "vext.32    q11, q8, q9, #2     \n"
-
-                        "vmla.f32   q13, q10, %e12[1]   \n"
-                        "vmla.f32   q14, q11, %f12[0]   \n"
-
-                        "pld        [%2, #192]          \n"
-                        "vld1.f32   {d16-d18}, [%2]     \n" // r0
-                        "add        %2, #16             \n"
-
-                        "vadd.f32   q7, q7, q13         \n"
-                        "vadd.f32   q7, q7, q14         \n"
-
-                        "vext.32    q10, q8, q9, #1     \n"
-                        "vext.32    q11, q8, q9, #2     \n"
-
-                        "vst1.f32   {d14-d15}, [%1]!    \n"
-
-                        "subs       %0, #1              \n"
-                        "bne        0b                  \n"
-
-                        "sub        %2, #16             \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2)      // %4
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k0123), // %10
-                        "w"(_k3456), // %11
-                        "w"(_k6789)  // %12
-                        : "cc", "memory", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-#if __ARM_NEON
-                    float32x4_t _r00 = vld1q_f32(r0);
-                    float32x4_t _r10 = vld1q_f32(r1);
-                    float32x4_t _r20 = vld1q_f32(r2);
-
-                    float32x4_t _sum = vmulq_f32(_r00, _k0123);
-                    _sum = vmlaq_f32(_sum, _r10, _k3456);
-                    _sum = vmlaq_f32(_sum, _r20, _k6789);
-
-                    _sum = vsetq_lane_f32(*outptr, _sum, 3);
-
-#if __aarch64__
-                    *outptr = vaddvq_f32(_sum);
-#else
-                    float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-                    _ss = vpadd_f32(_ss, _ss);
-
-                    *outptr = vget_lane_f32(_ss, 0);
-#endif // __aarch64__
-#else
-                    float sum = 0;
-
-                    sum += r0[0] * k0[0];
-                    sum += r0[1] * k0[1];
-                    sum += r0[2] * k0[2];
-                    sum += r1[0] * k1[0];
-                    sum += r1[1] * k1[1];
-                    sum += r1[2] * k1[2];
-                    sum += r2[0] * k2[0];
-                    sum += r2[1] * k2[1];
-                    sum += r2[2] * k2[2];
-
-                    *outptr += sum;
-#endif
-                    r0++;
-                    r1++;
-                    r2++;
-                    outptr++;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            kernel0 += 9;
-        }
-    }
-}
-
-static void conv3x3s2_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        Mat out0 = top_blob.channel(p);
-        Mat out1 = top_blob.channel(p + 1);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        const float bias1 = bias ? bias[p + 1] : 0.f;
-
-        out0.fill(bias0);
-        out1.fill(bias1);
-
-        const float* k0 = kernel + p * inch * 9;
-        const float* k1 = kernel + (p + 1) * inch * 9;
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0;
-            const float* r1 = img0 + w;
-            const float* r2 = img0 + w * 2;
-
-#if __ARM_NEON
-            float32x4_t _k00 = vld1q_f32(k0);
-            float32x4_t _k03 = vld1q_f32(k0 + 3);
-            float32x4_t _k06 = vld1q_f32(k0 + 6);
-
-            float32x4_t _k10 = vld1q_f32(k1);
-            float32x4_t _k13 = vld1q_f32(k1 + 3);
-            float32x4_t _k16 = vld1q_f32(k1 + 6);
-#endif // __ARM_NEON
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw & 3;
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld2    {v8.4s, v9.4s}, [%3], #32   \n" // v8 v9 = r0
-
-                        "0:                                 \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v6.4s}, [%1]               \n" // v6 = _sum0
-
-                        "fmul   v12.4s, v8.4s, %12.s[0]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v7.4s}, [%2]               \n" // v7 = _sum1
-
-                        "fmul   v13.4s, v8.4s, %15.s[0]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld2    {v10.4s, v11.4s}, [%3]      \n" // v10
-
-                        "fmla   v6.4s, v9.4s, %12.s[1]      \n"
-
-                        "ext    v14.16b, v8.16b, v10.16b, #4\n"
-
-                        "fmla   v7.4s, v9.4s, %15.s[1]      \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld2    {v8.4s, v9.4s}, [%4], #32   \n" // r1
-
-                        "fmla   v12.4s, v14.4s, %12.s[2]    \n"
-                        "fmla   v13.4s, v14.4s, %15.s[2]    \n"
-
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld2    {v10.4s, v11.4s}, [%4]      \n"
-
-                        "fmla   v6.4s, v8.4s, %13.s[0]      \n"
-                        "fmla   v7.4s, v8.4s, %16.s[0]      \n"
-
-                        "ext    v14.16b, v8.16b, v10.16b, #4\n"
-
-                        "fmla   v12.4s, v9.4s, %13.s[1]     \n"
-                        "fmla   v13.4s, v9.4s, %16.s[1]     \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld2    {v8.4s, v9.4s}, [%5], #32   \n" // r2
-
-                        "fmla   v6.4s, v14.4s, %13.s[2]     \n"
-                        "fmla   v7.4s, v14.4s, %16.s[2]     \n"
-
-                        "prfm   pldl1keep, [%5, #128]       \n"
-                        "ld2    {v10.4s, v11.4s}, [%5]      \n"
-
-                        "fmla   v12.4s, v8.4s, %14.s[0]     \n"
-                        "fmla   v13.4s, v8.4s, %17.s[0]     \n"
-
-                        "ext    v14.16b, v8.16b, v10.16b, #4\n"
-
-                        "fmla   v6.4s, v9.4s, %14.s[1]      \n"
-                        "fmla   v7.4s, v9.4s, %17.s[1]      \n"
-
-                        "fmla   v12.4s, v14.4s, %14.s[2]    \n"
-                        "fmla   v13.4s, v14.4s, %17.s[2]    \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld2    {v8.4s, v9.4s}, [%3], #32   \n" // v8 v9 = r0
-
-                        "fadd   v6.4s, v6.4s, v12.4s        \n"
-                        "fadd   v7.4s, v7.4s, v13.4s        \n"
-
-                        "subs   %w0, %w0, #1                \n"
-
-                        "st1    {v6.4s}, [%1], #16          \n"
-                        "st1    {v7.4s}, [%2], #16          \n"
-
-                        "bne    0b                          \n"
-                        "sub    %3, %3, #32                 \n"
-
-                        : "=r"(nn),      // %0
-                        "=r"(outptr0), // %1
-                        "=r"(outptr1), // %2
-                        "=r"(r0),      // %3
-                        "=r"(r1),      // %4
-                        "=r"(r2)       // %5
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(outptr1),
-                        "3"(r0),
-                        "4"(r1),
-                        "5"(r2),
-                        "w"(_k00), // %12
-                        "w"(_k03), // %13
-                        "w"(_k06), // %14
-                        "w"(_k10), // %15
-                        "w"(_k13), // %16
-                        "w"(_k16)  // %17
-                        : "cc", "memory", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "pld        [%3, #256]          \n"
-                        "vld2.f32   {d16-d19}, [%3]!    \n" // q8 q9 = r0
-
-                        "0:                             \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d12-d13}, [%1]     \n" // q6 = _sum0
-
-                        "vmul.f32   q12, q8, %e12[0]    \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d14-d15}, [%2]     \n" // q7 = _sum1
-
-                        "vmul.f32   q13, q8, %e15[0]    \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld2.f32   {d20-d21}, [%3]     \n" // q10
-
-                        "vmla.f32   q6, q9, %e12[1]     \n"
-
-                        "vext.32    q11, q8, q10, #1    \n"
-
-                        "vmla.f32   q7, q9, %e15[1]     \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld2.f32   {d16-d19}, [%4]!    \n" // r1
-
-                        "vmla.f32   q12, q11, %f12[0]   \n"
-                        "vmla.f32   q13, q11, %f15[0]   \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld2.f32   {d20-d21}, [%4]     \n"
-
-                        "vmla.f32   q6, q8, %e13[0]     \n"
-                        "vmla.f32   q7, q8, %e16[0]     \n"
-
-                        "vext.32    q11, q8, q10, #1    \n"
-
-                        "vmla.f32   q12, q9, %e13[1]    \n"
-                        "vmla.f32   q13, q9, %e16[1]    \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld2.f32   {d16-d19}, [%5]!    \n" // r2
-
-                        "vmla.f32   q6, q11, %f13[0]    \n"
-                        "vmla.f32   q7, q11, %f16[0]    \n"
-
-                        "pld        [%5, #128]          \n"
-                        "vld2.f32   {d20-d21}, [%5]     \n"
-
-                        "vmla.f32   q12, q8, %e14[0]    \n"
-                        "vmla.f32   q13, q8, %e17[0]    \n"
-
-                        "vext.32    q11, q8, q10, #1    \n"
-
-                        "vmla.f32   q6, q9, %e14[1]     \n"
-                        "vmla.f32   q7, q9, %e17[1]     \n"
-
-                        "vmla.f32   q12, q11, %f14[0]   \n"
-                        "vmla.f32   q13, q11, %f17[0]   \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld2.f32   {d16-d19}, [%3]!    \n" // q8 q9 = r0
-
-                        "vadd.f32   q6, q6, q12         \n"
-                        "vadd.f32   q7, q7, q13         \n"
-
-                        "subs       %0, #1              \n"
-
-                        "vst1.f32   {d12-d13}, [%1]!    \n"
-                        "vst1.f32   {d14-d15}, [%2]!    \n"
-
-                        "bne        0b                  \n"
-                        "sub        %3, #32             \n"
-
-                        : "=r"(nn),      // %0
-                        "=r"(outptr0), // %1
-                        "=r"(outptr1), // %2
-                        "=r"(r0),      // %3
-                        "=r"(r1),      // %4
-                        "=r"(r2)       // %5
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(outptr1),
-                        "3"(r0),
-                        "4"(r1),
-                        "5"(r2),
-                        "w"(_k00), // %12
-                        "w"(_k03), // %13
-                        "w"(_k06), // %14
-                        "w"(_k10), // %15
-                        "w"(_k13), // %16
-                        "w"(_k16)  // %17
-                        : "cc", "memory", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-#if __ARM_NEON
-                    float32x4_t _r00 = vld1q_f32(r0);
-                    float32x4_t _r10 = vld1q_f32(r1);
-                    float32x4_t _r20 = vld1q_f32(r2);
-
-                    float32x4_t _sum0 = vmulq_f32(_r00, _k00);
-                    float32x4_t _sum1 = vmulq_f32(_r00, _k10);
-                    _sum0 = vmlaq_f32(_sum0, _r10, _k03);
-                    _sum1 = vmlaq_f32(_sum1, _r10, _k13);
-                    _sum0 = vmlaq_f32(_sum0, _r20, _k06);
-                    _sum1 = vmlaq_f32(_sum1, _r20, _k16);
-
-                    _sum0 = vsetq_lane_f32(*outptr0, _sum0, 3);
-                    _sum1 = vsetq_lane_f32(*outptr1, _sum1, 3);
-#if __aarch64__
-                    *outptr0 = vaddvq_f32(_sum0);
-                    *outptr1 = vaddvq_f32(_sum1);
-#else
-                    float32x2_t _ss0 = vadd_f32(vget_low_f32(_sum0), vget_high_f32(_sum0));
-                    float32x2_t _ss1 = vadd_f32(vget_low_f32(_sum1), vget_high_f32(_sum1));
-                    float32x2_t _ss01 = vpadd_f32(_ss0, _ss1);
-
-                    *outptr0 = vget_lane_f32(_ss01, 0);
-                    *outptr1 = vget_lane_f32(_ss01, 1);
-#endif // __aarch64__
-#else
-                    float sum0 = 0.f;
-                    float sum1 = 0.f;
-
-                    sum0 += r0[0] * k0[0];
-                    sum0 += r0[1] * k0[1];
-                    sum0 += r0[2] * k0[2];
-                    sum0 += r1[0] * k0[3];
-                    sum0 += r1[1] * k0[4];
-                    sum0 += r1[2] * k0[5];
-                    sum0 += r2[0] * k0[6];
-                    sum0 += r2[1] * k0[7];
-                    sum0 += r2[2] * k0[8];
-
-                    sum1 += r0[0] * k1[0];
-                    sum1 += r0[1] * k1[1];
-                    sum1 += r0[2] * k1[2];
-                    sum1 += r1[0] * k1[3];
-                    sum1 += r1[1] * k1[4];
-                    sum1 += r1[2] * k1[5];
-                    sum1 += r2[0] * k1[6];
-                    sum1 += r2[1] * k1[7];
-                    sum1 += r2[2] * k1[8];
-
-                    *outptr0 += sum0;
-                    *outptr1 += sum1;
-#endif // __ARM_NEON
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0++;
-                    outptr1++;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9;
-            k1 += 9;
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        const float* kernel0 = kernel + p * inch * 9;
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0;
-            const float* r1 = img0 + w;
-            const float* r2 = img0 + w * 2;
-
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-#if __ARM_NEON
-            float32x4_t _k0123 = vld1q_f32(k0);
-            float32x4_t _k3456 = vld1q_f32(k1);
-            float32x4_t _k6789 = vld1q_f32(k2);
-#endif // __ARM_NEON
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw & 3;
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "prfm       pldl1keep, [%2, #256]          \n"
-                        "ld2        {v2.4s, v3.4s}, [%2], #32      \n"
-                        "0:                                        \n"
-
-                        "prfm       pldl1keep, [%1, #128]          \n"
-                        "ld1        {v0.4s}, [%1]                  \n"
-
-                        "fmla       v0.4s,  v2.4s, %10.s[0]        \n"
-                        "fmul       v10.4s, v3.4s, %10.s[1]        \n"
-
-                        "prfm       pldl1keep, [%2, #256]          \n"
-                        "ld2        {v8.4s, v9.4s}, [%2]           \n"
-                        "ext        v1.16b, v2.16b, v8.16b, #4     \n"
-
-                        "fmul       v11.4s, v1.4s, %10.s[2]        \n"
-
-                        "prfm       pldl1keep, [%3, #256]          \n"
-                        "ld2        {v2.4s, v3.4s}, [%3], #32      \n"
-
-                        "fmla       v0.4s,  v2.4s, %11.s[0]        \n"
-                        "fmla       v10.4s, v3.4s, %11.s[1]        \n"
-
-                        "prfm       pldl1keep, [%3, #256]          \n"
-                        "ld2        {v8.4s, v9.4s}, [%3]           \n"
-                        "ext        v1.16b, v2.16b, v8.16b, #4     \n"
-
-                        "fmla       v11.4s, v1.4s, %11.s[2]        \n"
-
-                        "prfm       pldl1keep, [%4, #256]          \n"
-                        "ld2        {v2.4s, v3.4s}, [%4], #32      \n"
-
-                        "fmla       v0.4s,  v2.4s, %12.s[0]        \n"
-                        "fmla       v10.4s, v3.4s, %12.s[1]        \n"
-
-                        "prfm       pldl1keep, [%4, #256]          \n"
-                        "ld2        {v8.4s, v9.4s}, [%4]           \n"
-                        "ext        v1.16b, v2.16b, v8.16b, #4     \n"
-
-                        "fmla       v11.4s, v1.4s, %12.s[2]        \n"
-
-                        "prfm       pldl1keep, [%2, #256]          \n"
-                        "ld2        {v2.4s, v3.4s}, [%2], #32      \n"
-
-                        "fadd       v0.4s, v0.4s, v10.4s           \n"
-                        "fadd       v0.4s, v0.4s, v11.4s           \n"
-
-                        "subs       %w0, %w0, #1                   \n"
-                        "st1        {v0.4s}, [%1], #16             \n"
-                        "bne        0b                             \n"
-                        "sub        %2, %2, #32                    \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2)      // %4
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k0123), // %10
-                        "w"(_k3456), // %11
-                        "w"(_k6789)  // %12
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "pld        [%2, #256]          \n"
-                        "vld2.f32   {d4-d7}, [%2]!      \n"
-
-                        "0:                             \n"
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%1]       \n"
-
-                        "vmla.f32   q0, q2, %e10[0]     \n"
-                        "vmul.f32   q10, q3, %e10[1]    \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld2.f32   {d16-d17}, [%2]     \n"
-                        "vext.32    q1, q2, q8, #1      \n"
-
-                        "vmul.f32   q11, q1, %f10[0]    \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld2.f32   {d4-d7}, [%3]!      \n"
-
-                        "vmla.f32   q0, q2, %e11[0]     \n"
-                        "vmla.f32   q10, q3, %e11[1]    \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld2.f32   {d16-d17}, [%3]     \n"
-                        "vext.32    q1, q2, q8, #1      \n"
-
-                        "vmla.f32   q11, q1, %f11[0]    \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld2.f32   {d4-d7}, [%4]!      \n"
-
-                        "vmla.f32   q0, q2, %e12[0]     \n"
-                        "vmla.f32   q10, q3, %e12[1]    \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld2.f32   {d16-d17}, [%4]     \n"
-                        "vext.32    q1, q2, q8, #1      \n"
-
-                        "vmla.f32   q11, q1, %f12[0]    \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld2.f32   {d4-d7}, [%2]!      \n"
-
-                        "vadd.f32   q0, q0, q10         \n"
-                        "vadd.f32   q0, q0, q11         \n"
-
-                        "subs       %0, #1              \n"
-                        "vst1.f32   {d0-d1}, [%1]!      \n"
-                        "bne        0b                  \n"
-                        "sub        %2, #32             \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2)      // %4
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k0123), // %10
-                        "w"(_k3456), // %11
-                        "w"(_k6789)  // %12
-                        : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-#if __ARM_NEON
-                    float32x4_t _r00 = vld1q_f32(r0);
-                    float32x4_t _r10 = vld1q_f32(r1);
-                    float32x4_t _r20 = vld1q_f32(r2);
-
-                    float32x4_t _sum = vmulq_f32(_r00, _k0123);
-                    _sum = vmlaq_f32(_sum, _r10, _k3456);
-                    _sum = vmlaq_f32(_sum, _r20, _k6789);
-
-                    _sum = vsetq_lane_f32(*outptr, _sum, 3);
-
-#if __aarch64__
-                    *outptr = vaddvq_f32(_sum);
-#else
-                    float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-                    _ss = vpadd_f32(_ss, _ss);
-
-                    *outptr = vget_lane_f32(_ss, 0);
-#endif // __aarch64__
-#else
-                    float sum = 0;
-
-                    sum += r0[0] * k0[0];
-                    sum += r0[1] * k0[1];
-                    sum += r0[2] * k0[2];
-                    sum += r1[0] * k1[0];
-                    sum += r1[1] * k1[1];
-                    sum += r1[2] * k1[2];
-                    sum += r2[0] * k2[0];
-                    sum += r2[1] * k2[1];
-                    sum += r2[2] * k2[2];
-
-                    *outptr += sum;
-#endif // __ARM_NEON
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr++;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            kernel0 += 9;
-        }
-    }
-}
-
-static void conv3x3s2_transform_kernel_neon(const Mat& _kernel, Mat& kernel_tm, int inch, int outch)
-{
-    kernel_tm.create(8 * 9, inch, outch / 8 + outch % 8);
-
-    const float* kernel = _kernel;
-
-    int p = 0;
-    for (; p + 7 < outch; p += 8)
-    {
-        const float* k0 = kernel + (p + 0) * inch * 9;
-        const float* k1 = kernel + (p + 1) * inch * 9;
-        const float* k2 = kernel + (p + 2) * inch * 9;
-        const float* k3 = kernel + (p + 3) * inch * 9;
-        const float* k4 = kernel + (p + 4) * inch * 9;
-        const float* k5 = kernel + (p + 5) * inch * 9;
-        const float* k6 = kernel + (p + 6) * inch * 9;
-        const float* k7 = kernel + (p + 7) * inch * 9;
-
-        float* ktmp = kernel_tm.channel(p / 8);
-
-        for (int q = 0; q < inch; q++)
-        {
-            for (int k = 0; k < 9; k++)
-            {
-                ktmp[0] = k0[k];
-                ktmp[1] = k1[k];
-                ktmp[2] = k2[k];
-                ktmp[3] = k3[k];
-                ktmp[4] = k4[k];
-                ktmp[5] = k5[k];
-                ktmp[6] = k6[k];
-                ktmp[7] = k7[k];
-                ktmp += 8;
-            }
-
-            k0 += 9;
-            k1 += 9;
-            k2 += 9;
-            k3 += 9;
-            k4 += 9;
-            k5 += 9;
-            k6 += 9;
-            k7 += 9;
-        }
-    }
-    for (; p < outch; p++)
-    {
-        const float* k0 = kernel + (p + 0) * inch * 9;
-
-        float* ktmp = kernel_tm.channel(p / 8 + p % 8);
-
-        for (int q = 0; q < inch; q++)
-        {
-            for (int k = 0; k < 9; k++)
-            {
-                ktmp[k] = k0[k];
-            }
-            ktmp += 9;
-
-            k0 += 9;
-        }
-    }
-}
-
-static void conv3x3s2_packed_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    //     const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    int nn_outch = outch >> 3;
-    int remain_outch_start = nn_outch << 3;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 8;
-
-        Mat out0 = top_blob.channel(p + 0);
-        Mat out1 = top_blob.channel(p + 1);
-        Mat out2 = top_blob.channel(p + 2);
-        Mat out3 = top_blob.channel(p + 3);
-        Mat out4 = top_blob.channel(p + 4);
-        Mat out5 = top_blob.channel(p + 5);
-        Mat out6 = top_blob.channel(p + 6);
-        Mat out7 = top_blob.channel(p + 7);
-
-        const float bias0 = bias ? bias[p + 0] : 0.f;
-        const float bias1 = bias ? bias[p + 1] : 0.f;
-        const float bias2 = bias ? bias[p + 2] : 0.f;
-        const float bias3 = bias ? bias[p + 3] : 0.f;
-        const float bias4 = bias ? bias[p + 4] : 0.f;
-        const float bias5 = bias ? bias[p + 5] : 0.f;
-        const float bias6 = bias ? bias[p + 6] : 0.f;
-        const float bias7 = bias ? bias[p + 7] : 0.f;
-
-        out0.fill(bias0);
-        out1.fill(bias1);
-        out2.fill(bias2);
-        out3.fill(bias3);
-        out4.fill(bias4);
-        out5.fill(bias5);
-        out6.fill(bias6);
-        out7.fill(bias7);
-
-        const float* ktmp = _kernel.channel(p / 8);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-            float* outptr2 = out2;
-            float* outptr3 = out3;
-            float* outptr4 = out4;
-            float* outptr5 = out5;
-            float* outptr6 = out6;
-            float* outptr7 = out7;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0;
-            const float* r1 = img0 + w;
-            const float* r2 = img0 + w * 2;
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw & 3;
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "0:                                 \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v8.4s}, [%1]               \n"
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v9.4s}, [%2]               \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v10.4s}, [%3]              \n"
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v11.4s}, [%4]              \n"
-
-                        ///
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld2    {v4.4s, v5.4s}, [%9], #32   \n" // v4=00 v5=01
-
-                        "ld1    {v0.4s, v1.4s}, [%12], #32  \n"
-
-                        "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                        "fmla   v9.4s, v4.4s, v0.s[1]       \n"
-
-                        "prfm   pldl1keep, [%5, #128]       \n"
-                        "ld1    {v12.4s}, [%5]              \n"
-                        "prfm   pldl1keep, [%6, #128]       \n"
-                        "ld1    {v13.4s}, [%6]              \n"
-
-                        "fmla   v10.4s, v4.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v4.4s, v0.s[3]      \n"
-
-                        "prfm   pldl1keep, [%7, #128]       \n"
-                        "ld1    {v14.4s}, [%7]              \n"
-                        "prfm   pldl1keep, [%8, #128]       \n"
-                        "ld1    {v15.4s}, [%8]              \n"
-
-                        "ld1    {v2.4s, v3.4s}, [%12], #32  \n"
-
-                        "fmla   v12.4s, v4.4s, v1.s[0]      \n"
-                        "fmla   v13.4s, v4.4s, v1.s[1]      \n"
-                        "fmla   v14.4s, v4.4s, v1.s[2]      \n"
-                        "fmla   v15.4s, v4.4s, v1.s[3]      \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld2    {v6.4s, v7.4s}, [%9]        \n" // v6
-
-                        "fmla   v8.4s, v5.4s, v2.s[0]       \n"
-                        "fmla   v9.4s, v5.4s, v2.s[1]       \n"
-                        "fmla   v10.4s, v5.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v5.4s, v2.s[3]      \n"
-
-                        "ext    v6.16b, v4.16b, v6.16b, #4  \n" // v6=02
-
-                        "ld1    {v0.4s, v1.4s}, [%12], #32  \n"
-
-                        "fmla   v12.4s, v5.4s, v3.s[0]      \n"
-                        "fmla   v13.4s, v5.4s, v3.s[1]      \n"
-                        "fmla   v14.4s, v5.4s, v3.s[2]      \n"
-                        "fmla   v15.4s, v5.4s, v3.s[3]      \n"
-
-                        ///
-                        "prfm   pldl1keep, [%10, #256]      \n"
-                        "ld2    {v4.4s, v5.4s}, [%10], #32  \n" // v4=10 v5=11
-
-                        "fmla   v8.4s, v6.4s, v0.s[0]       \n"
-                        "fmla   v9.4s, v6.4s, v0.s[1]       \n"
-                        "fmla   v10.4s, v6.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v6.4s, v0.s[3]      \n"
-
-                        "ld1    {v2.4s, v3.4s}, [%12], #32  \n"
-
-                        "fmla   v12.4s, v6.4s, v1.s[0]      \n"
-                        "fmla   v13.4s, v6.4s, v1.s[1]      \n"
-                        "fmla   v14.4s, v6.4s, v1.s[2]      \n"
-                        "fmla   v15.4s, v6.4s, v1.s[3]      \n"
-
-                        "fmla   v8.4s, v4.4s, v2.s[0]       \n"
-                        "fmla   v9.4s, v4.4s, v2.s[1]       \n"
-                        "fmla   v10.4s, v4.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v4.4s, v2.s[3]      \n"
-
-                        "ld1    {v0.4s, v1.4s}, [%12], #32  \n"
-
-                        "fmla   v12.4s, v4.4s, v3.s[0]      \n"
-                        "fmla   v13.4s, v4.4s, v3.s[1]      \n"
-                        "fmla   v14.4s, v4.4s, v3.s[2]      \n"
-                        "fmla   v15.4s, v4.4s, v3.s[3]      \n"
-
-                        "prfm   pldl1keep, [%10, #256]      \n"
-                        "ld2    {v6.4s, v7.4s}, [%10]       \n" // v6
-
-                        "fmla   v8.4s, v5.4s, v0.s[0]       \n"
-                        "fmla   v9.4s, v5.4s, v0.s[1]       \n"
-                        "fmla   v10.4s, v5.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v5.4s, v0.s[3]      \n"
-
-                        "ld1    {v2.4s, v3.4s}, [%12], #32  \n"
-
-                        "ext    v6.16b, v4.16b, v6.16b, #4  \n" // v6=12
-
-                        "fmla   v12.4s, v5.4s, v1.s[0]      \n"
-                        "fmla   v13.4s, v5.4s, v1.s[1]      \n"
-                        "fmla   v14.4s, v5.4s, v1.s[2]      \n"
-                        "fmla   v15.4s, v5.4s, v1.s[3]      \n"
-
-                        ///
-                        "prfm   pldl1keep, [%11, #256]      \n"
-                        "ld2    {v4.4s, v5.4s}, [%11], #32  \n" // v4=20 v5=21
-
-                        "fmla   v8.4s, v6.4s, v2.s[0]       \n"
-                        "fmla   v9.4s, v6.4s, v2.s[1]       \n"
-                        "fmla   v10.4s, v6.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v6.4s, v2.s[3]      \n"
-
-                        "ld1    {v0.4s, v1.4s}, [%12], #32  \n"
-
-                        "fmla   v12.4s, v6.4s, v3.s[0]      \n"
-                        "fmla   v13.4s, v6.4s, v3.s[1]      \n"
-                        "fmla   v14.4s, v6.4s, v3.s[2]      \n"
-                        "fmla   v15.4s, v6.4s, v3.s[3]      \n"
-
-                        "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                        "fmla   v9.4s, v4.4s, v0.s[1]       \n"
-                        "fmla   v10.4s, v4.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v4.4s, v0.s[3]      \n"
-
-                        "ld1    {v2.4s, v3.4s}, [%12], #32  \n"
-
-                        "fmla   v12.4s, v4.4s, v1.s[0]      \n"
-                        "fmla   v13.4s, v4.4s, v1.s[1]      \n"
-                        "fmla   v14.4s, v4.4s, v1.s[2]      \n"
-                        "fmla   v15.4s, v4.4s, v1.s[3]      \n"
-
-                        "prfm   pldl1keep, [%11, #256]      \n"
-                        "ld2    {v6.4s, v7.4s}, [%11]       \n" // v6
-
-                        "fmla   v8.4s, v5.4s, v2.s[0]       \n"
-                        "fmla   v9.4s, v5.4s, v2.s[1]       \n"
-                        "fmla   v10.4s, v5.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v5.4s, v2.s[3]      \n"
-
-                        "ext    v6.16b, v4.16b, v6.16b, #4  \n" // v6=22
-
-                        "ld1    {v0.4s, v1.4s}, [%12], #32  \n"
-
-                        "fmla   v12.4s, v5.4s, v3.s[0]      \n"
-                        "fmla   v13.4s, v5.4s, v3.s[1]      \n"
-                        "fmla   v14.4s, v5.4s, v3.s[2]      \n"
-                        "fmla   v15.4s, v5.4s, v3.s[3]      \n"
-
-                        "fmla   v8.4s, v6.4s, v0.s[0]       \n"
-                        "fmla   v9.4s, v6.4s, v0.s[1]       \n"
-                        "fmla   v10.4s, v6.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v6.4s, v0.s[3]      \n"
-
-                        "fmla   v12.4s, v6.4s, v1.s[0]      \n"
-                        "fmla   v13.4s, v6.4s, v1.s[1]      \n"
-
-                        "st1    {v8.4s}, [%1], #16          \n"
-                        "st1    {v9.4s}, [%2], #16          \n"
-
-                        "fmla   v14.4s, v6.4s, v1.s[2]      \n"
-                        "fmla   v15.4s, v6.4s, v1.s[3]      \n"
-
-                        "st1    {v10.4s}, [%3], #16         \n"
-                        "st1    {v11.4s}, [%4], #16         \n"
-
-                        "sub    %12, %12, #288              \n"
-
-                        "st1    {v12.4s}, [%5], #16         \n"
-                        "st1    {v13.4s}, [%6], #16         \n"
-
-                        "subs   %w0, %w0, #1                \n"
-
-                        "st1    {v14.4s}, [%7], #16         \n"
-                        "st1    {v15.4s}, [%8], #16         \n"
-
-                        "bne    0b                          \n"
-                        : "=r"(nn),      // %0
-                        "=r"(outptr0), // %1
-                        "=r"(outptr1), // %2
-                        "=r"(outptr2), // %3
-                        "=r"(outptr3), // %4
-                        "=r"(outptr4), // %5
-                        "=r"(outptr5), // %6
-                        "=r"(outptr6), // %7
-                        "=r"(outptr7), // %8
-                        "=r"(r0),      // %9
-                        "=r"(r1),      // %10
-                        "=r"(r2),      // %11
-                        "=r"(ktmp)     // %12
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(outptr1),
-                        "3"(outptr2),
-                        "4"(outptr3),
-                        "5"(outptr4),
-                        "6"(outptr5),
-                        "7"(outptr6),
-                        "8"(outptr7),
-                        "9"(r0),
-                        "10"(r1),
-                        "11"(r2),
-                        "12"(ktmp)
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                }
-#else  // __aarch64__
-                for (; nn > 0; nn--)
-                {
-                    asm volatile(
-                        "pld        [%0, #128]          \n"
-                        "vld1.f32   {d16-d17}, [%0]     \n"
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d18-d19}, [%1]     \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d20-d21}, [%2]     \n"
-                        "pld        [%3, #128]          \n"
-                        "vld1.f32   {d22-d23}, [%3]     \n"
-
-                        ///
-                        "pld        [%8, #256]          \n"
-                        "vld2.f32   {d8-d11}, [%8]!     \n" // q4=00 q5=01
-
-                        "vld1.f32   {d0-d3}, [%11 :128]! \n"
-
-                        "vmla.f32   q8, q4, d0[0]       \n"
-                        "vmla.f32   q9, q4, d0[1]       \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.f32   {d24-d25}, [%4]     \n"
-                        "pld        [%5, #128]          \n"
-                        "vld1.f32   {d26-d27}, [%5]     \n"
-
-                        "vmla.f32   q10, q4, d1[0]      \n"
-                        "vmla.f32   q11, q4, d1[1]      \n"
-
-                        "pld        [%6, #128]          \n"
-                        "vld1.f32   {d28-d29}, [%6]     \n"
-                        "pld        [%7, #128]          \n"
-                        "vld1.f32   {d30-d31}, [%7]     \n"
-
-                        "vld1.f32   {d4-d7}, [%11 :128]! \n"
-
-                        "vmla.f32   q12, q4, d2[0]      \n"
-                        "vmla.f32   q13, q4, d2[1]      \n"
-                        "vmla.f32   q14, q4, d3[0]      \n"
-                        "vmla.f32   q15, q4, d3[1]      \n"
-
-                        "pld        [%8, #128]          \n"
-                        "vld2.f32   {d12-d13}, [%8]     \n" // q6
-
-                        "vmla.f32   q8, q5, d4[0]       \n"
-                        "vmla.f32   q9, q5, d4[1]       \n"
-                        "vmla.f32   q10, q5, d5[0]      \n"
-                        "vmla.f32   q11, q5, d5[1]      \n"
-
-                        "vext.f32   q6, q4, q6, #1      \n" // q6=02
-
-                        "vld1.f32   {d0-d3}, [%11 :128]! \n"
-
-                        "vmla.f32   q12, q5, d6[0]      \n"
-                        "vmla.f32   q13, q5, d6[1]      \n"
-                        "vmla.f32   q14, q5, d7[0]      \n"
-                        "vmla.f32   q15, q5, d7[1]      \n"
-
-                        ///
-                        "pld        [%9, #256]          \n"
-                        "vld2.f32   {d8-d11}, [%9]!     \n" // q4=10 q5=11
-
-                        "vmla.f32   q8, q6, d0[0]       \n"
-                        "vmla.f32   q9, q6, d0[1]       \n"
-                        "vmla.f32   q10, q6, d1[0]      \n"
-                        "vmla.f32   q11, q6, d1[1]      \n"
-
-                        "vld1.f32   {d4-d7}, [%11 :128]! \n"
-
-                        "vmla.f32   q12, q6, d2[0]      \n"
-                        "vmla.f32   q13, q6, d2[1]      \n"
-                        "vmla.f32   q14, q6, d3[0]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "vmla.f32   q8, q4, d4[0]       \n"
-                        "vmla.f32   q9, q4, d4[1]       \n"
-                        "vmla.f32   q10, q4, d5[0]      \n"
-                        "vmla.f32   q11, q4, d5[1]      \n"
-
-                        "vld1.f32   {d0-d3}, [%11 :128]! \n"
-
-                        "vmla.f32   q12, q4, d6[0]      \n"
-                        "vmla.f32   q13, q4, d6[1]      \n"
-                        "vmla.f32   q14, q4, d7[0]      \n"
-                        "vmla.f32   q15, q4, d7[1]      \n"
-
-                        "pld        [%9, #128]          \n"
-                        "vld2.f32   {d12-d13}, [%9]     \n" // q6
-
-                        "vmla.f32   q8, q5, d0[0]       \n"
-                        "vmla.f32   q9, q5, d0[1]       \n"
-                        "vmla.f32   q10, q5, d1[0]      \n"
-                        "vmla.f32   q11, q5, d1[1]      \n"
-
-                        "vld1.f32   {d4-d7}, [%11 :128]! \n"
-
-                        "vext.f32   q6, q4, q6, #1      \n" // q6=12
-
-                        "vmla.f32   q12, q5, d2[0]      \n"
-                        "vmla.f32   q13, q5, d2[1]      \n"
-                        "vmla.f32   q14, q5, d3[0]      \n"
-                        "vmla.f32   q15, q5, d3[1]      \n"
-
-                        ///
-                        "pld        [%10, #256]         \n"
-                        "vld2.f32   {d8-d11}, [%10]!    \n" // q4=20 q5=21
-
-                        "vmla.f32   q8, q6, d4[0]       \n"
-                        "vmla.f32   q9, q6, d4[1]       \n"
-                        "vmla.f32   q10, q6, d5[0]      \n"
-                        "vmla.f32   q11, q6, d5[1]      \n"
-
-                        "vld1.f32   {d0-d3}, [%11 :128]! \n"
-
-                        "vmla.f32   q12, q6, d6[0]      \n"
-                        "vmla.f32   q13, q6, d6[1]      \n"
-                        "vmla.f32   q14, q6, d7[0]      \n"
-                        "vmla.f32   q15, q6, d7[1]      \n"
-
-                        "vmla.f32   q8, q4, d0[0]       \n"
-                        "vmla.f32   q9, q4, d0[1]       \n"
-                        "vmla.f32   q10, q4, d1[0]      \n"
-                        "vmla.f32   q11, q4, d1[1]      \n"
-
-                        "vld1.f32   {d4-d7}, [%11 :128]! \n"
-
-                        "vmla.f32   q12, q4, d2[0]      \n"
-                        "vmla.f32   q13, q4, d2[1]      \n"
-                        "vmla.f32   q14, q4, d3[0]      \n"
-                        "vmla.f32   q15, q4, d3[1]      \n"
-
-                        "pld        [%10, #128]         \n"
-                        "vld2.f32   {d12-d13}, [%10]    \n" // q6
-
-                        "vmla.f32   q8, q5, d4[0]       \n"
-                        "vmla.f32   q9, q5, d4[1]       \n"
-                        "vmla.f32   q10, q5, d5[0]      \n"
-                        "vmla.f32   q11, q5, d5[1]      \n"
-
-                        "vext.f32   q6, q4, q6, #1      \n" // q6=22
-
-                        "vld1.f32   {d0-d3}, [%11 :128]! \n"
-
-                        "vmla.f32   q12, q5, d6[0]      \n"
-                        "vmla.f32   q13, q5, d6[1]      \n"
-                        "vmla.f32   q14, q5, d7[0]      \n"
-                        "vmla.f32   q15, q5, d7[1]      \n"
-
-                        "vmla.f32   q8, q6, d0[0]       \n"
-                        "vmla.f32   q9, q6, d0[1]       \n"
-                        "vmla.f32   q10, q6, d1[0]      \n"
-                        "vmla.f32   q11, q6, d1[1]      \n"
-
-                        "vmla.f32   q12, q6, d2[0]      \n"
-                        "vmla.f32   q13, q6, d2[1]      \n"
-
-                        "vst1.f32   {d16-d17}, [%0]!    \n"
-                        "vst1.f32   {d18-d19}, [%1]!    \n"
-
-                        "vmla.f32   q14, q6, d3[0]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "vst1.f32   {d20-d21}, [%2]!    \n"
-                        "vst1.f32   {d22-d23}, [%3]!    \n"
-
-                        "sub        %11, %11, #288      \n"
-
-                        "vst1.f32   {d24-d25}, [%4]!    \n"
-                        "vst1.f32   {d26-d27}, [%5]!    \n"
-                        "vst1.f32   {d28-d29}, [%6]!    \n"
-                        "vst1.f32   {d30-d31}, [%7]!    \n"
-                        : "=r"(outptr0), // %0
-                        "=r"(outptr1), // %1
-                        "=r"(outptr2), // %2
-                        "=r"(outptr3), // %3
-                        "=r"(outptr4), // %4
-                        "=r"(outptr5), // %5
-                        "=r"(outptr6), // %6
-                        "=r"(outptr7), // %7
-                        "=r"(r0),      // %8
-                        "=r"(r1),      // %9
-                        "=r"(r2),      // %10
-                        "=r"(ktmp)     // %11
-                        : "0"(outptr0),
-                        "1"(outptr1),
-                        "2"(outptr2),
-                        "3"(outptr3),
-                        "4"(outptr4),
-                        "5"(outptr5),
-                        "6"(outptr6),
-                        "7"(outptr7),
-                        "8"(r0),
-                        "9"(r1),
-                        "10"(r2),
-                        "11"(ktmp)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-#if __ARM_NEON
-#if __aarch64__
-                    asm volatile(
-                        "ld1    {v10.4s, v11.4s}, [%11], #32    \n"
-
-                        "prfm   pldl1keep, [%8, #128]   \n"
-                        "ld1    {v0.4s}, [%8]           \n"
-
-                        "ld1    {v12.4s, v13.4s}, [%11], #32    \n"
-
-                        "ld1    {v8.s}[0], [%0]         \n"
-                        "ld1    {v8.s}[1], [%1]         \n"
-                        "ld1    {v8.s}[2], [%2]         \n"
-                        "ld1    {v8.s}[3], [%3]         \n"
-
-                        "fmul   v14.4s, v10.4s, v0.s[0] \n"
-                        "fmul   v15.4s, v11.4s, v0.s[0] \n"
-
-                        "ld1    {v9.s}[0], [%4]         \n"
-                        "ld1    {v9.s}[1], [%5]         \n"
-                        "ld1    {v9.s}[2], [%6]         \n"
-                        "ld1    {v9.s}[3], [%7]         \n"
-
-                        "ld1    {v10.4s, v11.4s}, [%11], #32    \n"
-
-                        "fmla   v8.4s, v12.4s, v0.s[1]  \n"
-                        "fmla   v9.4s, v13.4s, v0.s[1]  \n"
-
-                        "ld1    {v12.4s, v13.4s}, [%11], #32    \n"
-
-                        "fmla   v14.4s, v10.4s, v0.s[2] \n"
-                        "fmla   v15.4s, v11.4s, v0.s[2] \n"
-
-                        "prfm   pldl1keep, [%9, #128]   \n"
-                        "ld1    {v1.4s}, [%9]           \n"
-
-                        "ld1    {v10.4s, v11.4s}, [%11], #32    \n"
-
-                        "fmla   v8.4s, v12.4s, v1.s[0]  \n"
-                        "fmla   v9.4s, v13.4s, v1.s[0]  \n"
-
-                        "ld1    {v12.4s, v13.4s}, [%11], #32    \n"
-
-                        "fmla   v14.4s, v10.4s, v1.s[1] \n"
-                        "fmla   v15.4s, v11.4s, v1.s[1] \n"
-
-                        "ld1    {v10.4s, v11.4s}, [%11], #32    \n"
-
-                        "fmla   v8.4s, v12.4s, v1.s[2]  \n"
-                        "fmla   v9.4s, v13.4s, v1.s[2]  \n"
-
-                        "prfm   pldl1keep, [%10, #128]  \n"
-                        "ld1    {v0.4s}, [%10]          \n"
-
-                        "ld1    {v12.4s, v13.4s}, [%11], #32    \n"
-
-                        "fmla   v14.4s, v10.4s, v0.s[0] \n"
-                        "fmla   v15.4s, v11.4s, v0.s[0] \n"
-
-                        "ld1    {v10.4s, v11.4s}, [%11], #32    \n"
-
-                        "fmla   v8.4s, v12.4s, v0.s[1]  \n"
-                        "fmla   v9.4s, v13.4s, v0.s[1]  \n"
-
-                        "fmla   v14.4s, v10.4s, v0.s[2] \n"
-                        "fmla   v15.4s, v11.4s, v0.s[2] \n"
-
-                        "fadd   v8.4s, v8.4s, v14.4s    \n"
-                        "fadd   v9.4s, v9.4s, v15.4s    \n"
-
-                        "sub    %11, %11, #288          \n"
-
-                        "st1    {v8.s}[0], [%0], #4     \n"
-                        "st1    {v8.s}[1], [%1], #4     \n"
-                        "st1    {v8.s}[2], [%2], #4     \n"
-                        "st1    {v8.s}[3], [%3], #4     \n"
-
-                        "st1    {v9.s}[0], [%4], #4     \n"
-                        "st1    {v9.s}[1], [%5], #4     \n"
-                        "st1    {v9.s}[2], [%6], #4     \n"
-                        "st1    {v9.s}[3], [%7], #4     \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(outptr1), // %1
-                        "=r"(outptr2), // %2
-                        "=r"(outptr3), // %3
-                        "=r"(outptr4), // %4
-                        "=r"(outptr5), // %5
-                        "=r"(outptr6), // %6
-                        "=r"(outptr7), // %7
-                        "=r"(r0),      // %8
-                        "=r"(r1),      // %9
-                        "=r"(r2),      // %10
-                        "=r"(ktmp)     // %11
-                        : "0"(outptr0),
-                        "1"(outptr1),
-                        "2"(outptr2),
-                        "3"(outptr3),
-                        "4"(outptr4),
-                        "5"(outptr5),
-                        "6"(outptr6),
-                        "7"(outptr7),
-                        "8"(r0),
-                        "9"(r1),
-                        "10"(r2),
-                        "11"(ktmp)
-                        : "memory", "v0", "v1", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-#else  // __aarch64__
-                    asm volatile(
-                        "vld1.f32   {d20-d23}, [%11 :128]! \n"
-
-                        "pld        [%8, #128]      \n"
-                        "vld1.f32   {d0-d1}, [%8]   \n"
-
-                        "vld1.f32   {d24-d27}, [%11 :128]! \n"
-
-                        "vld1.f32   {d16[0]}, [%0]  \n"
-                        "vld1.f32   {d16[1]}, [%1]  \n"
-                        "vld1.f32   {d17[0]}, [%2]  \n"
-                        "vld1.f32   {d17[1]}, [%3]  \n"
-
-                        "vmul.f32   q14, q10, d0[0] \n"
-                        "vmul.f32   q15, q11, d0[0] \n"
-
-                        "vld1.f32   {d18[0]}, [%4]  \n"
-                        "vld1.f32   {d18[1]}, [%5]  \n"
-                        "vld1.f32   {d19[0]}, [%6]  \n"
-                        "vld1.f32   {d19[1]}, [%7]  \n"
-
-                        "vld1.f32   {d20-d23}, [%11 :128]! \n"
-
-                        "vmla.f32   q8, q12, d0[1]  \n"
-                        "vmla.f32   q9, q13, d0[1]  \n"
-
-                        "vld1.f32   {d24-d27}, [%11 :128]! \n"
-
-                        "vmla.f32   q14, q10, d1[0] \n"
-                        "vmla.f32   q15, q11, d1[0] \n"
-
-                        "pld        [%9, #128]      \n"
-                        "vld1.f32   {d2-d3}, [%9]   \n"
-
-                        "vld1.f32   {d20-d23}, [%11 :128]! \n"
-
-                        "vmla.f32   q8, q12, d2[0]  \n"
-                        "vmla.f32   q9, q13, d2[0]  \n"
-
-                        "vld1.f32   {d24-d27}, [%11 :128]! \n"
-
-                        "vmla.f32   q14, q10, d2[1] \n"
-                        "vmla.f32   q15, q11, d2[1] \n"
-
-                        "vld1.f32   {d20-d23}, [%11 :128]! \n"
-
-                        "vmla.f32   q8, q12, d3[0]  \n"
-                        "vmla.f32   q9, q13, d3[0]  \n"
-
-                        "pld        [%10, #128]     \n"
-                        "vld1.f32   {d0-d1}, [%10]  \n"
-
-                        "vld1.f32   {d24-d27}, [%11 :128]! \n"
-
-                        "vmla.f32   q14, q10, d0[0] \n"
-                        "vmla.f32   q15, q11, d0[0] \n"
-
-                        "vld1.f32   {d20-d23}, [%11 :128]! \n"
-
-                        "vmla.f32   q8, q12, d0[1]  \n"
-                        "vmla.f32   q9, q13, d0[1]  \n"
-
-                        "vmla.f32   q14, q10, d1[0] \n"
-                        "vmla.f32   q15, q11, d1[0] \n"
-
-                        "vadd.f32   q8, q8, q14     \n"
-                        "vadd.f32   q9, q9, q15     \n"
-
-                        "sub        %11, %11, #288  \n"
-
-                        "vst1.f32   {d16[0]}, [%0]! \n"
-                        "vst1.f32   {d16[1]}, [%1]! \n"
-                        "vst1.f32   {d17[0]}, [%2]! \n"
-                        "vst1.f32   {d17[1]}, [%3]! \n"
-
-                        "vst1.f32   {d18[0]}, [%4]! \n"
-                        "vst1.f32   {d18[1]}, [%5]! \n"
-                        "vst1.f32   {d19[0]}, [%6]! \n"
-                        "vst1.f32   {d19[1]}, [%7]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(outptr1), // %1
-                        "=r"(outptr2), // %2
-                        "=r"(outptr3), // %3
-                        "=r"(outptr4), // %4
-                        "=r"(outptr5), // %5
-                        "=r"(outptr6), // %6
-                        "=r"(outptr7), // %7
-                        "=r"(r0),      // %8
-                        "=r"(r1),      // %9
-                        "=r"(r2),      // %10
-                        "=r"(ktmp)     // %11
-                        : "0"(outptr0),
-                        "1"(outptr1),
-                        "2"(outptr2),
-                        "3"(outptr3),
-                        "4"(outptr4),
-                        "5"(outptr5),
-                        "6"(outptr6),
-                        "7"(outptr7),
-                        "8"(r0),
-                        "9"(r1),
-                        "10"(r2),
-                        "11"(ktmp)
-                        : "memory", "q0", "q1", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // __ARM_NEON
-                    float sum0 = 0.f;
-                    float sum1 = 0.f;
-                    float sum2 = 0.f;
-                    float sum3 = 0.f;
-                    float sum4 = 0.f;
-                    float sum5 = 0.f;
-                    float sum6 = 0.f;
-                    float sum7 = 0.f;
-
-                    sum0 += r0[0] * ktmp[0];
-                    sum1 += r0[0] * ktmp[1];
-                    sum2 += r0[0] * ktmp[2];
-                    sum3 += r0[0] * ktmp[3];
-                    sum4 += r0[0] * ktmp[4];
-                    sum5 += r0[0] * ktmp[5];
-                    sum6 += r0[0] * ktmp[6];
-                    sum7 += r0[0] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += r0[1] * ktmp[0];
-                    sum1 += r0[1] * ktmp[1];
-                    sum2 += r0[1] * ktmp[2];
-                    sum3 += r0[1] * ktmp[3];
-                    sum4 += r0[1] * ktmp[4];
-                    sum5 += r0[1] * ktmp[5];
-                    sum6 += r0[1] * ktmp[6];
-                    sum7 += r0[1] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += r0[2] * ktmp[0];
-                    sum1 += r0[2] * ktmp[1];
-                    sum2 += r0[2] * ktmp[2];
-                    sum3 += r0[2] * ktmp[3];
-                    sum4 += r0[2] * ktmp[4];
-                    sum5 += r0[2] * ktmp[5];
-                    sum6 += r0[2] * ktmp[6];
-                    sum7 += r0[2] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += r1[0] * ktmp[0];
-                    sum1 += r1[0] * ktmp[1];
-                    sum2 += r1[0] * ktmp[2];
-                    sum3 += r1[0] * ktmp[3];
-                    sum4 += r1[0] * ktmp[4];
-                    sum5 += r1[0] * ktmp[5];
-                    sum6 += r1[0] * ktmp[6];
-                    sum7 += r1[0] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += r1[1] * ktmp[0];
-                    sum1 += r1[1] * ktmp[1];
-                    sum2 += r1[1] * ktmp[2];
-                    sum3 += r1[1] * ktmp[3];
-                    sum4 += r1[1] * ktmp[4];
-                    sum5 += r1[1] * ktmp[5];
-                    sum6 += r1[1] * ktmp[6];
-                    sum7 += r1[1] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += r1[2] * ktmp[0];
-                    sum1 += r1[2] * ktmp[1];
-                    sum2 += r1[2] * ktmp[2];
-                    sum3 += r1[2] * ktmp[3];
-                    sum4 += r1[2] * ktmp[4];
-                    sum5 += r1[2] * ktmp[5];
-                    sum6 += r1[2] * ktmp[6];
-                    sum7 += r1[2] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += r2[0] * ktmp[0];
-                    sum1 += r2[0] * ktmp[1];
-                    sum2 += r2[0] * ktmp[2];
-                    sum3 += r2[0] * ktmp[3];
-                    sum4 += r2[0] * ktmp[4];
-                    sum5 += r2[0] * ktmp[5];
-                    sum6 += r2[0] * ktmp[6];
-                    sum7 += r2[0] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += r2[1] * ktmp[0];
-                    sum1 += r2[1] * ktmp[1];
-                    sum2 += r2[1] * ktmp[2];
-                    sum3 += r2[1] * ktmp[3];
-                    sum4 += r2[1] * ktmp[4];
-                    sum5 += r2[1] * ktmp[5];
-                    sum6 += r2[1] * ktmp[6];
-                    sum7 += r2[1] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += r2[2] * ktmp[0];
-                    sum1 += r2[2] * ktmp[1];
-                    sum2 += r2[2] * ktmp[2];
-                    sum3 += r2[2] * ktmp[3];
-                    sum4 += r2[2] * ktmp[4];
-                    sum5 += r2[2] * ktmp[5];
-                    sum6 += r2[2] * ktmp[6];
-                    sum7 += r2[2] * ktmp[7];
-                    ktmp += 8;
-
-                    *outptr0 += sum0;
-                    *outptr1 += sum1;
-                    *outptr2 += sum2;
-                    *outptr3 += sum3;
-                    *outptr4 += sum4;
-                    *outptr5 += sum5;
-                    *outptr6 += sum6;
-                    *outptr7 += sum7;
-
-                    ktmp -= 8 * 9;
-
-                    outptr0++;
-                    outptr1++;
-                    outptr2++;
-                    outptr3++;
-                    outptr4++;
-                    outptr5++;
-                    outptr6++;
-                    outptr7++;
-#endif // __ARM_NEON
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            ktmp += 8 * 9;
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        const float* ktmp = _kernel.channel(p / 8 + p % 8);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0;
-            const float* r1 = img0 + w;
-            const float* r2 = img0 + w * 2;
-
-            const float* k0 = ktmp;
-            const float* k1 = ktmp + 3;
-            const float* k2 = ktmp + 6;
-
-#if __ARM_NEON
-            float32x4_t _k0123 = vld1q_f32(k0);
-            float32x4_t _k3456 = vld1q_f32(k1);
-            float32x4_t _k6789 = vld1q_f32(k2);
-#endif // __ARM_NEON
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw & 3;
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "prfm       pldl1keep, [%2, #256]          \n"
-                        "ld2        {v2.4s, v3.4s}, [%2], #32      \n"
-                        "0:                                        \n"
-
-                        "prfm       pldl1keep, [%1, #128]          \n"
-                        "ld1        {v0.4s}, [%1]                  \n"
-
-                        "fmla       v0.4s,  v2.4s, %10.s[0]        \n"
-                        "fmul       v10.4s, v3.4s, %10.s[1]        \n"
-
-                        "prfm       pldl1keep, [%2, #256]          \n"
-                        "ld2        {v8.4s, v9.4s}, [%2]           \n"
-                        "ext        v1.16b, v2.16b, v8.16b, #4     \n"
-
-                        "fmul       v11.4s, v1.4s, %10.s[2]        \n"
-
-                        "prfm       pldl1keep, [%3, #256]          \n"
-                        "ld2        {v2.4s, v3.4s}, [%3], #32      \n"
-
-                        "fmla       v0.4s,  v2.4s, %11.s[0]        \n"
-                        "fmla       v10.4s, v3.4s, %11.s[1]        \n"
-
-                        "prfm       pldl1keep, [%3, #256]          \n"
-                        "ld2        {v8.4s, v9.4s}, [%3]           \n"
-                        "ext        v1.16b, v2.16b, v8.16b, #4     \n"
-
-                        "fmla       v11.4s, v1.4s, %11.s[2]        \n"
-
-                        "prfm       pldl1keep, [%4, #256]          \n"
-                        "ld2        {v2.4s, v3.4s}, [%4], #32      \n"
-
-                        "fmla       v0.4s,  v2.4s, %12.s[0]        \n"
-                        "fmla       v10.4s, v3.4s, %12.s[1]        \n"
-
-                        "prfm       pldl1keep, [%4, #256]          \n"
-                        "ld2        {v8.4s, v9.4s}, [%4]           \n"
-                        "ext        v1.16b, v2.16b, v8.16b, #4     \n"
-
-                        "fmla       v11.4s, v1.4s, %12.s[2]        \n"
-
-                        "prfm       pldl1keep, [%2, #256]          \n"
-                        "ld2        {v2.4s, v3.4s}, [%2], #32      \n"
-
-                        "fadd       v0.4s, v0.4s, v10.4s           \n"
-                        "fadd       v0.4s, v0.4s, v11.4s           \n"
-
-                        "subs       %w0, %w0, #1                   \n"
-                        "st1        {v0.4s}, [%1], #16             \n"
-                        "bne        0b                             \n"
-                        "sub        %2, %2, #32                    \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2)      // %4
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k0123), // %10
-                        "w"(_k3456), // %11
-                        "w"(_k6789)  // %12
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "pld        [%2, #256]          \n"
-                        "vld2.f32   {d4-d7}, [%2]!      \n"
-
-                        "0:                             \n"
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%1]       \n"
-
-                        "vmla.f32   q0, q2, %e10[0]     \n"
-                        "vmul.f32   q10, q3, %e10[1]    \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld2.f32   {d16-d17}, [%2]     \n"
-                        "vext.32    q1, q2, q8, #1      \n"
-
-                        "vmul.f32   q11, q1, %f10[0]    \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld2.f32   {d4-d7}, [%3]!      \n"
-
-                        "vmla.f32   q0, q2, %e11[0]     \n"
-                        "vmla.f32   q10, q3, %e11[1]    \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld2.f32   {d16-d17}, [%3]     \n"
-                        "vext.32    q1, q2, q8, #1      \n"
-
-                        "vmla.f32   q11, q1, %f11[0]    \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld2.f32   {d4-d7}, [%4]!      \n"
-
-                        "vmla.f32   q0, q2, %e12[0]     \n"
-                        "vmla.f32   q10, q3, %e12[1]    \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld2.f32   {d16-d17}, [%4]     \n"
-                        "vext.32    q1, q2, q8, #1      \n"
-
-                        "vmla.f32   q11, q1, %f12[0]    \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld2.f32   {d4-d7}, [%2]!      \n"
-
-                        "vadd.f32   q0, q0, q10         \n"
-                        "vadd.f32   q0, q0, q11         \n"
-
-                        "subs       %0, #1              \n"
-                        "vst1.f32   {d0-d1}, [%1]!      \n"
-                        "bne        0b                  \n"
-                        "sub        %2, #32             \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2)      // %4
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k0123), // %10
-                        "w"(_k3456), // %11
-                        "w"(_k6789)  // %12
-                        : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-#if __ARM_NEON
-                    float32x4_t _r00 = vld1q_f32(r0);
-                    float32x4_t _r10 = vld1q_f32(r1);
-                    float32x4_t _r20 = vld1q_f32(r2);
-
-                    float32x4_t _sum = vmulq_f32(_r00, _k0123);
-                    _sum = vmlaq_f32(_sum, _r10, _k3456);
-                    _sum = vmlaq_f32(_sum, _r20, _k6789);
-
-                    _sum = vsetq_lane_f32(*outptr, _sum, 3);
-
-#if __aarch64__
-                    *outptr = vaddvq_f32(_sum);
-#else
-                    float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-                    _ss = vpadd_f32(_ss, _ss);
-
-                    *outptr = vget_lane_f32(_ss, 0);
-#endif // __aarch64__
-#else
-                    float sum = 0;
-
-                    sum += r0[0] * ktmp[0];
-                    sum += r0[1] * ktmp[1];
-                    sum += r0[2] * ktmp[2];
-                    sum += r1[0] * ktmp[3];
-                    sum += r1[1] * ktmp[4];
-                    sum += r1[2] * ktmp[5];
-                    sum += r2[0] * ktmp[6];
-                    sum += r2[1] * ktmp[7];
-                    sum += r2[2] * ktmp[8];
-
-                    *outptr += sum;
-#endif // __ARM_NEON
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr++;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            ktmp += 9;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_int8.h
deleted file mode 100644
index 1868b5d6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_int8.h
+++ /dev/null
@@ -1,1190 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s2_transform_kernel_int8_neon(const Mat& _kernel, Mat& kernel_tm, int inch, int outch)
-{
-    kernel_tm.create(8 * 9, inch, outch / 8 + outch % 8, (size_t)1u);
-
-    const signed char* kernel = _kernel;
-
-    int p = 0;
-    for (; p + 7 < outch; p += 8)
-    {
-        const signed char* k0 = kernel + (p + 0) * inch * 9;
-        const signed char* k1 = kernel + (p + 1) * inch * 9;
-        const signed char* k2 = kernel + (p + 2) * inch * 9;
-        const signed char* k3 = kernel + (p + 3) * inch * 9;
-        const signed char* k4 = kernel + (p + 4) * inch * 9;
-        const signed char* k5 = kernel + (p + 5) * inch * 9;
-        const signed char* k6 = kernel + (p + 6) * inch * 9;
-        const signed char* k7 = kernel + (p + 7) * inch * 9;
-
-        signed char* ktmp = kernel_tm.channel(p / 8);
-
-        for (int q = 0; q < inch; q++)
-        {
-            for (int k = 0; k < 9; k++)
-            {
-                ktmp[0] = k0[k];
-                ktmp[1] = k1[k];
-                ktmp[2] = k2[k];
-                ktmp[3] = k3[k];
-                ktmp[4] = k4[k];
-                ktmp[5] = k5[k];
-                ktmp[6] = k6[k];
-                ktmp[7] = k7[k];
-                ktmp += 8;
-            }
-
-            k0 += 9;
-            k1 += 9;
-            k2 += 9;
-            k3 += 9;
-            k4 += 9;
-            k5 += 9;
-            k6 += 9;
-            k7 += 9;
-        }
-    }
-    for (; p < outch; p++)
-    {
-        const signed char* k0 = kernel + (p + 0) * inch * 9;
-
-        signed char* ktmp = kernel_tm.channel(p / 8 + p % 8);
-
-        for (int q = 0; q < inch; q++)
-        {
-            for (int k = 0; k < 9; k++)
-            {
-                ktmp[k] = k0[k];
-            }
-            ktmp += 9;
-
-            k0 += 9;
-        }
-    }
-}
-
-static void conv3x3s2_packed_int8_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    int nn_outch = outch >> 3;
-    int remain_outch_start = nn_outch << 3;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 8;
-
-        Mat out0 = top_blob.channel(p + 0);
-        Mat out1 = top_blob.channel(p + 1);
-        Mat out2 = top_blob.channel(p + 2);
-        Mat out3 = top_blob.channel(p + 3);
-        Mat out4 = top_blob.channel(p + 4);
-        Mat out5 = top_blob.channel(p + 5);
-        Mat out6 = top_blob.channel(p + 6);
-        Mat out7 = top_blob.channel(p + 7);
-
-        out0.fill(0);
-        out1.fill(0);
-        out2.fill(0);
-        out3.fill(0);
-        out4.fill(0);
-        out5.fill(0);
-        out6.fill(0);
-        out7.fill(0);
-
-        const signed char* ktmp = _kernel.channel(p / 8);
-
-        for (int q = 0; q < inch; q++)
-        {
-            int* outptr0 = out0;
-            int* outptr1 = out1;
-            int* outptr2 = out2;
-            int* outptr3 = out3;
-            int* outptr4 = out4;
-            int* outptr5 = out5;
-            int* outptr6 = out6;
-            int* outptr7 = out7;
-
-            const signed char* img0 = bottom_blob.channel(q);
-
-            const signed char* r0 = img0;
-            const signed char* r1 = img0 + w;
-            const signed char* r2 = img0 + w * 2;
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-#if __ARM_NEON
-#if __aarch64__
-                int nn = outw >> 3;
-                int remain = outw & 7;
-#else
-                int nn = outw >> 2;
-                int remain = outw & 3;
-#endif // __aarch64__
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "0:                                   \n"
-
-                        "ld1    {v0.8b, v1.8b, v2.8b}, [%12], #24  \n" //ktmp
-                        "ld2    {v3.8b, v4.8b}, [%9], #16     \n"      //r0-r2
-                        "ld2    {v5.8b, v6.8b}, [%9]          \n"
-
-                        "ld1    {v8.4s, v9.4s}, [%1]          \n" //out0
-                        "ld1    {v10.4s, v11.4s}, [%2]        \n" //out1
-                        "ld1    {v12.4s, v13.4s}, [%3]        \n" //out2
-                        "ld1    {v14.4s, v15.4s}, [%4]        \n" //out3
-                        "ld1    {v16.4s, v17.4s}, [%5]        \n" //out4
-                        "ld1    {v18.4s, v19.4s}, [%6]        \n" //out5
-                        "ld1    {v20.4s, v21.4s}, [%7]        \n" //out6
-                        "ld1    {v22.4s, v23.4s}, [%8]        \n" //out7
-
-                        "ext    v7.8b, v3.8b, v5.8b, #1       \n"
-
-                        "sshll  v0.8h, v0.8b, #0              \n" //(k00-k70)
-                        "sshll  v1.8h, v1.8b, #0              \n" //(k01-k71)
-                        "sshll  v2.8h, v2.8b, #0              \n" //(k02-k72)
-                        "sshll  v3.8h, v3.8b, #0              \n" // r0
-                        "sshll  v4.8h, v4.8b, #0              \n" // r1
-                        "sshll  v7.8h, v7.8b, #0              \n" // r2
-
-                        // r0
-                        "smlal  v8.4s, v3.4h, v0.h[0]         \n" // out0 += (r00-r07)*k00
-                        "smlal2  v9.4s, v3.8h, v0.h[0]        \n"
-                        "smlal  v10.4s, v3.4h, v0.h[1]        \n" // out1 += (r00-r07)*k10
-                        "smlal2  v11.4s, v3.8h, v0.h[1]       \n"
-                        "smlal  v12.4s, v3.4h, v0.h[2]        \n" // out2 += (r00-r07)*k20
-                        "smlal2  v13.4s, v3.8h, v0.h[2]       \n"
-                        "smlal  v14.4s, v3.4h, v0.h[3]        \n" // out3 += (r00-r07)*k30
-                        "smlal2  v15.4s, v3.8h, v0.h[3]       \n"
-                        "smlal  v16.4s, v3.4h, v0.h[4]        \n" // out4 += (r00-r07)*k40
-                        "smlal2  v17.4s, v3.8h, v0.h[4]       \n"
-                        "smlal  v18.4s, v3.4h, v0.h[5]        \n" // out5 += (r00-r07)*k50
-                        "smlal2  v19.4s, v3.8h, v0.h[5]       \n"
-                        "smlal  v20.4s, v3.4h, v0.h[6]        \n" // out6 += (r00-r07)*k60
-                        "smlal2  v21.4s, v3.8h, v0.h[6]       \n"
-                        "smlal  v22.4s, v3.4h, v0.h[7]        \n" // out7 += (r00-r07)*k70
-                        "smlal2  v23.4s, v3.8h, v0.h[7]       \n"
-                        // r1
-                        "smlal  v8.4s, v4.4h, v1.h[0]         \n" // out0 += (r10-r17)*k01
-                        "smlal2  v9.4s, v4.8h, v1.h[0]        \n"
-                        "smlal  v10.4s, v4.4h, v1.h[1]        \n" // out1 += (r10-r17)*k11
-                        "smlal2  v11.4s, v4.8h, v1.h[1]       \n"
-                        "smlal  v12.4s, v4.4h, v1.h[2]        \n" // out2 += (r10-r17)*k21
-                        "smlal2  v13.4s, v4.8h, v1.h[2]       \n"
-                        "smlal  v14.4s, v4.4h, v1.h[3]        \n" // out3 += (r10-r17)*k31
-                        "smlal2  v15.4s, v4.8h, v1.h[3]       \n"
-                        "smlal  v16.4s, v4.4h, v1.h[4]        \n" // out4 += (r10-r17)*k41
-                        "smlal2  v17.4s, v4.8h, v1.h[4]       \n"
-                        "smlal  v18.4s, v4.4h, v1.h[5]        \n" // out5 += (r10-r17)*k51
-                        "smlal2  v19.4s, v4.8h, v1.h[5]       \n"
-                        "smlal  v20.4s, v4.4h, v1.h[6]        \n" // out6 += (r10-r17)*k61
-                        "smlal2  v21.4s, v4.8h, v1.h[6]       \n"
-                        "smlal  v22.4s, v4.4h, v1.h[7]        \n" // out7 += (r10-r17)*k71
-                        "smlal2  v23.4s, v4.8h, v1.h[7]       \n"
-                        // r2
-                        "smlal  v8.4s, v7.4h, v2.h[0]         \n" // out0 += (r20-r27)*k02
-                        "smlal2  v9.4s, v7.8h, v2.h[0]        \n"
-                        "smlal  v10.4s, v7.4h, v2.h[1]        \n" // out1 += (r20-r27)*k12
-                        "smlal2  v11.4s, v7.8h, v2.h[1]       \n"
-                        "smlal  v12.4s, v7.4h, v2.h[2]        \n" // out2 += (r20-r27)*k22
-                        "smlal2  v13.4s, v7.8h, v2.h[2]       \n"
-                        "smlal  v14.4s, v7.4h, v2.h[3]        \n" // out3 += (r20-r27)*k32
-                        "smlal2  v15.4s, v7.8h, v2.h[3]       \n"
-                        "smlal  v16.4s, v7.4h, v2.h[4]        \n" // out4 += (r20-r27)*k42
-                        "smlal2  v17.4s, v7.8h, v2.h[4]       \n"
-                        "smlal  v18.4s, v7.4h, v2.h[5]        \n" // out5 += (r20-r27)*k52
-                        "smlal2  v19.4s, v7.8h, v2.h[5]       \n"
-                        "smlal  v20.4s, v7.4h, v2.h[6]        \n" // out6 += (r20-r27)*k62
-                        "smlal2  v21.4s, v7.8h, v2.h[6]       \n"
-                        "smlal  v22.4s, v7.4h, v2.h[7]        \n" // out7 += (r20-r27)*k72
-                        "smlal2  v23.4s, v7.8h, v2.h[7]       \n"
-
-                        "ld1    {v0.8b, v1.8b, v2.8b}, [%12], #24  \n" //ktmp
-                        "ld2    {v3.8b, v4.8b}, [%10], #16    \n"      //r3-r5
-                        "ld2    {v5.8b, v6.8b}, [%10]         \n"
-
-                        "ext    v7.8b, v3.8b, v5.8b, #1       \n"
-
-                        "sshll  v0.8h, v0.8b, #0              \n" //(k03-k73)
-                        "sshll  v1.8h, v1.8b, #0              \n" //(k04-k74)
-                        "sshll  v2.8h, v2.8b, #0              \n" //(k05-k75)
-                        "sshll  v3.8h, v3.8b, #0              \n" // r3
-                        "sshll  v4.8h, v4.8b, #0              \n" // r4
-                        "sshll  v7.8h, v7.8b, #0              \n" // r5
-
-                        // r3
-                        "smlal  v8.4s, v3.4h, v0.h[0]         \n" // out0 += (r30-r37)*k03
-                        "smlal2  v9.4s, v3.8h, v0.h[0]        \n"
-                        "smlal  v10.4s, v3.4h, v0.h[1]        \n" // out1 += (r30-r37)*k13
-                        "smlal2  v11.4s, v3.8h, v0.h[1]       \n"
-                        "smlal  v12.4s, v3.4h, v0.h[2]        \n" // out2 += (r30-r37)*k23
-                        "smlal2  v13.4s, v3.8h, v0.h[2]       \n"
-                        "smlal  v14.4s, v3.4h, v0.h[3]        \n" // out3 += (r30-r37)*k33
-                        "smlal2  v15.4s, v3.8h, v0.h[3]       \n"
-                        "smlal  v16.4s, v3.4h, v0.h[4]        \n" // out4 += (r30-r37)*k43
-                        "smlal2  v17.4s, v3.8h, v0.h[4]       \n"
-                        "smlal  v18.4s, v3.4h, v0.h[5]        \n" // out5 += (r30-r37)*k53
-                        "smlal2  v19.4s, v3.8h, v0.h[5]       \n"
-                        "smlal  v20.4s, v3.4h, v0.h[6]        \n" // out6 += (r30-r37)*k63
-                        "smlal2  v21.4s, v3.8h, v0.h[6]       \n"
-                        "smlal  v22.4s, v3.4h, v0.h[7]        \n" // out7 += (r30-r37)*k73
-                        "smlal2  v23.4s, v3.8h, v0.h[7]       \n"
-                        // r4
-                        "smlal  v8.4s, v4.4h, v1.h[0]         \n" // out0 += (r40-r47)*k04
-                        "smlal2  v9.4s, v4.8h, v1.h[0]        \n"
-                        "smlal  v10.4s, v4.4h, v1.h[1]        \n" // out1 += (r40-r47)*k14
-                        "smlal2  v11.4s, v4.8h, v1.h[1]       \n"
-                        "smlal  v12.4s, v4.4h, v1.h[2]        \n" // out2 += (r40-r47)*k24
-                        "smlal2  v13.4s, v4.8h, v1.h[2]       \n"
-                        "smlal  v14.4s, v4.4h, v1.h[3]        \n" // out3 += (r40-r47)*k34
-                        "smlal2  v15.4s, v4.8h, v1.h[3]       \n"
-                        "smlal  v16.4s, v4.4h, v1.h[4]        \n" // out4 += (r40-r47)*k44
-                        "smlal2  v17.4s, v4.8h, v1.h[4]       \n"
-                        "smlal  v18.4s, v4.4h, v1.h[5]        \n" // out5 += (r40-r47)*k54
-                        "smlal2  v19.4s, v4.8h, v1.h[5]       \n"
-                        "smlal  v20.4s, v4.4h, v1.h[6]        \n" // out6 += (r40-r47)*k64
-                        "smlal2  v21.4s, v4.8h, v1.h[6]       \n"
-                        "smlal  v22.4s, v4.4h, v1.h[7]        \n" // out7 += (r40-r47)*k74
-                        "smlal2  v23.4s, v4.8h, v1.h[7]       \n"
-                        // r5
-                        "smlal  v8.4s, v7.4h, v2.h[0]         \n" // out0 += (r50-r57)*k05
-                        "smlal2  v9.4s, v7.8h, v2.h[0]        \n"
-                        "smlal  v10.4s, v7.4h, v2.h[1]        \n" // out1 += (r50-r57)*k15
-                        "smlal2  v11.4s, v7.8h, v2.h[1]       \n"
-                        "smlal  v12.4s, v7.4h, v2.h[2]        \n" // out2 += (r50-r57)*k25
-                        "smlal2  v13.4s, v7.8h, v2.h[2]       \n"
-                        "smlal  v14.4s, v7.4h, v2.h[3]        \n" // out3 += (r50-r57)*k35
-                        "smlal2  v15.4s, v7.8h, v2.h[3]       \n"
-                        "smlal  v16.4s, v7.4h, v2.h[4]        \n" // out4 += (r50-r57)*k45
-                        "smlal2  v17.4s, v7.8h, v2.h[4]       \n"
-                        "smlal  v18.4s, v7.4h, v2.h[5]        \n" // out5 += (r50-r57)*k55
-                        "smlal2  v19.4s, v7.8h, v2.h[5]       \n"
-                        "smlal  v20.4s, v7.4h, v2.h[6]        \n" // out6 += (r50-r57)*k65
-                        "smlal2  v21.4s, v7.8h, v2.h[6]       \n"
-                        "smlal  v22.4s, v7.4h, v2.h[7]        \n" // out7 += (r50-r57)*k75
-                        "smlal2  v23.4s, v7.8h, v2.h[7]       \n"
-
-                        "ld1    {v0.8b, v1.8b, v2.8b}, [%12], #24  \n" //ktmp
-                        "ld2    {v3.8b, v4.8b}, [%11], #16    \n"      //r6-r8
-                        "ld2    {v5.8b, v6.8b}, [%11]         \n"
-
-                        "ext    v7.8b, v3.8b, v5.8b, #1       \n"
-
-                        "sshll  v0.8h, v0.8b, #0              \n" //(k06-k76)
-                        "sshll  v1.8h, v1.8b, #0              \n" //(k07-k77)
-                        "sshll  v2.8h, v2.8b, #0              \n" //(k08-k78)
-                        "sshll  v3.8h, v3.8b, #0              \n" // r6
-                        "sshll  v4.8h, v4.8b, #0              \n" // r7
-                        "sshll  v7.8h, v7.8b, #0              \n" // r8
-
-                        // r6
-                        "smlal  v8.4s, v3.4h, v0.h[0]         \n" // out0 += (r60-r67)*k06
-                        "smlal2  v9.4s, v3.8h, v0.h[0]        \n"
-                        "smlal  v10.4s, v3.4h, v0.h[1]        \n" // out1 += (r60-r67)*k16
-                        "smlal2  v11.4s, v3.8h, v0.h[1]       \n"
-                        "smlal  v12.4s, v3.4h, v0.h[2]        \n" // out2 += (r60-r67)*k26
-                        "smlal2  v13.4s, v3.8h, v0.h[2]       \n"
-                        "smlal  v14.4s, v3.4h, v0.h[3]        \n" // out3 += (r60-r67)*k36
-                        "smlal2  v15.4s, v3.8h, v0.h[3]       \n"
-                        "smlal  v16.4s, v3.4h, v0.h[4]        \n" // out4 += (r60-r67)*k46
-                        "smlal2  v17.4s, v3.8h, v0.h[4]       \n"
-                        "smlal  v18.4s, v3.4h, v0.h[5]        \n" // out5 += (r60-r67)*k56
-                        "smlal2  v19.4s, v3.8h, v0.h[5]       \n"
-                        "smlal  v20.4s, v3.4h, v0.h[6]        \n" // out6 += (r60-r67)*k66
-                        "smlal2  v21.4s, v3.8h, v0.h[6]       \n"
-                        "smlal  v22.4s, v3.4h, v0.h[7]        \n" // out7 += (r60-r67)*k76
-                        "smlal2  v23.4s, v3.8h, v0.h[7]       \n"
-                        // r7
-                        "smlal  v8.4s, v4.4h, v1.h[0]         \n" // out0 += (r70-r77)*k07
-                        "smlal2  v9.4s, v4.8h, v1.h[0]        \n"
-                        "smlal  v10.4s, v4.4h, v1.h[1]        \n" // out1 += (r70-r77)*k17
-                        "smlal2  v11.4s, v4.8h, v1.h[1]       \n"
-                        "smlal  v12.4s, v4.4h, v1.h[2]        \n" // out2 += (r70-r77)*k27
-                        "smlal2  v13.4s, v4.8h, v1.h[2]       \n"
-                        "smlal  v14.4s, v4.4h, v1.h[3]        \n" // out3 += (r70-r77)*k37
-                        "smlal2  v15.4s, v4.8h, v1.h[3]       \n"
-                        "smlal  v16.4s, v4.4h, v1.h[4]        \n" // out4 += (r70-r77)*k47
-                        "smlal2  v17.4s, v4.8h, v1.h[4]       \n"
-                        "smlal  v18.4s, v4.4h, v1.h[5]        \n" // out5 += (r70-r77)*k57
-                        "smlal2  v19.4s, v4.8h, v1.h[5]       \n"
-                        "smlal  v20.4s, v4.4h, v1.h[6]        \n" // out6 += (r70-r77)*k67
-                        "smlal2  v21.4s, v4.8h, v1.h[6]       \n"
-                        "smlal  v22.4s, v4.4h, v1.h[7]        \n" // out7 += (r70-r77)*k77
-                        "smlal2  v23.4s, v4.8h, v1.h[7]       \n"
-                        // r8
-                        "smlal  v8.4s, v7.4h, v2.h[0]         \n" // out0 += (r80-r87)*k08
-                        "smlal2  v9.4s, v7.8h, v2.h[0]        \n"
-                        "smlal  v10.4s, v7.4h, v2.h[1]        \n" // out1 += (r80-r87)*k18
-                        "smlal2  v11.4s, v7.8h, v2.h[1]       \n"
-                        "smlal  v12.4s, v7.4h, v2.h[2]        \n" // out2 += (r80-r87)*k28
-                        "smlal2  v13.4s, v7.8h, v2.h[2]       \n"
-                        "smlal  v14.4s, v7.4h, v2.h[3]        \n" // out3 += (r80-r87)*k38
-                        "smlal2  v15.4s, v7.8h, v2.h[3]       \n"
-                        "smlal  v16.4s, v7.4h, v2.h[4]        \n" // out4 += (r80-r87)*k48
-                        "smlal2  v17.4s, v7.8h, v2.h[4]       \n"
-                        "smlal  v18.4s, v7.4h, v2.h[5]        \n" // out5 += (r80-r87)*k58
-                        "smlal2  v19.4s, v7.8h, v2.h[5]       \n"
-                        "smlal  v20.4s, v7.4h, v2.h[6]        \n" // out6 += (r80-r87)*k68
-                        "smlal2  v21.4s, v7.8h, v2.h[6]       \n"
-                        "smlal  v22.4s, v7.4h, v2.h[7]        \n" // out7 += (r80-r87)*k78
-                        "smlal2  v23.4s, v7.8h, v2.h[7]       \n"
-
-                        "st1    {v8.4s, v9.4s}, [%1], #32     \n"
-                        "st1    {v10.4s, v11.4s}, [%2], #32   \n"
-                        "st1    {v12.4s, v13.4s}, [%3], #32   \n"
-                        "st1    {v14.4s, v15.4s}, [%4], #32   \n"
-                        "st1    {v16.4s, v17.4s}, [%5], #32   \n"
-                        "st1    {v18.4s, v19.4s}, [%6], #32   \n"
-                        "st1    {v20.4s, v21.4s}, [%7], #32   \n"
-                        "st1    {v22.4s, v23.4s}, [%8], #32   \n"
-
-                        "subs   %w0, %w0, #1                  \n"
-                        "sub    %12, %12, #72                 \n" // reset ktmp
-
-                        "bne    0b                            \n"
-
-                        : "=r"(nn),      // %0
-                        "=r"(outptr0), // %1
-                        "=r"(outptr1), // %2
-                        "=r"(outptr2), // %3
-                        "=r"(outptr3), // %4
-                        "=r"(outptr4), // %5
-                        "=r"(outptr5), // %6
-                        "=r"(outptr6), // %7
-                        "=r"(outptr7), // %8
-                        "=r"(r0),      // %9
-                        "=r"(r1),      // %10
-                        "=r"(r2),      // %11
-                        "=r"(ktmp)     // %12
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(outptr1),
-                        "3"(outptr2),
-                        "4"(outptr3),
-                        "5"(outptr4),
-                        "6"(outptr5),
-                        "7"(outptr6),
-                        "8"(outptr7),
-                        "9"(r0),
-                        "10"(r1),
-                        "11"(r2),
-                        "12"(ktmp)
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-                }
-#else  // __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "0:                             \n"
-                        "pld        [%1, #128]          \n"
-                        "vld1.s32   {d16-d17}, [%1]     \n" // out0
-                        "pld        [%2, #128]          \n"
-                        "vld1.s32   {d18-d19}, [%2]     \n" // out1
-                        "pld        [%3, #128]          \n"
-                        "vld1.s32   {d20-d21}, [%3]     \n" // out2
-                        "pld        [%4, #128]          \n"
-                        "vld1.s32   {d22-d23}, [%4]     \n" // out3
-
-                        // r0
-                        "pld        [%9, #64]          \n"
-                        "vld2.s8    {d8-d9}, [%9]       \n" // d8(a00 a02 a04 a06 a08 a010 a012 a014), d9(a01 a03 a05 a07 a09 a011 a013 a015)
-                        "add        %9, #8              \n"
-                        "pld        [%12, #64]         \n"
-                        "vld1.s8    {d0-d2}, [%12]!     \n" // d0(k00-k70) d1(k01-k71) d2(k02-k72)
-
-                        "pld        [%5, #128]          \n"
-                        "vld1.s32   {d24-d25}, [%5]     \n" // out4
-                        "pld        [%6, #128]          \n"
-                        "vld1.s32   {d26-d27}, [%6]     \n" // out5
-
-                        "vmovl.s8   q2, d2              \n" // q2(k02-k72)
-                        "vmovl.s8   q1, d1              \n" // q1(k01-k71)
-                        "vmovl.s8   q0, d0              \n" // q0(k00-k70)
-                        "vext.s8    d12, d8, d8, #1     \n" // d12(a02 a04 a06 a08 x x x x)
-
-                        "pld        [%7, #128]          \n"
-                        "vld1.s32   {d28-d29}, [%7]     \n" // out6
-
-                        "vmovl.s8   q5, d9              \n" // q5(a01 a03 a05 a07 a09 a011 a013 a015) d11
-                        "vmovl.s8   q4, d8              \n" // q4(a00 a02 a04 a06 a08 a010 a012 a014) d9
-                        "vmovl.s8   q6, d12             \n" // q6(a02 a04 a06 a08 a010 a012 a014 a016) d13
-
-                        "pld        [%8, #128]          \n"
-                        "vld1.s32   {d30-d31}, [%8]     \n" // out7
-
-                        "vmlal.s16  q8, d8, d0[0]       \n" // sum0 += (a00 a02 a04 a06) * k00
-                        "vmlal.s16  q9, d8, d0[1]       \n" // sum1 += (a00 a02 a04 a06) * k10
-                        "vmlal.s16  q10, d8, d0[2]      \n" // sum2 += (a00 a02 a04 a06) * k20
-                        "vmlal.s16  q11, d8, d0[3]      \n" // sum3 += (a00 a02 a04 a06) * k30
-                        "vmlal.s16  q12, d8, d1[0]      \n" // sum4 += (a00 a02 a04 a06) * k40
-                        "vmlal.s16  q13, d8, d1[1]      \n" // sum5 += (a00 a02 a04 a06) * k50
-                        "vmlal.s16  q14, d8, d1[2]      \n" // sum6 += (a00 a02 a04 a06) * k60
-                        "vmlal.s16  q15, d8, d1[3]      \n" // sum7 += (a00 a02 a04 a06) * k70
-
-                        "vmlal.s16  q8, d10, d2[0]      \n" // sum0 += (a01-a07) * k01
-                        "vmlal.s16  q9, d10, d2[1]      \n" // sum1 += (a01-a07) * k11
-                        "vmlal.s16  q10, d10, d2[2]     \n" // sum2 += (a01-a07) * k21
-                        "vmlal.s16  q11, d10, d2[3]     \n" // sum3 += (a01-a07) * k31
-                        "vmlal.s16  q12, d10, d3[0]     \n" // sum4 += (a01-a07) * k41
-                        "vmlal.s16  q13, d10, d3[1]     \n" // sum5 += (a01-a07) * k51
-                        "vmlal.s16  q14, d10, d3[2]     \n" // sum6 += (a01-a07) * k61
-                        "vmlal.s16  q15, d10, d3[3]     \n" // sum7 += (a01-a07) * k71
-
-                        "pld        [%10, #64]         \n"
-                        "vld2.s8    {d8-d9}, [%10]      \n" // d8(a10 a12 a14 a16 a18 a110 a112 a114), d9(a11 a13 a15 a17 a19 a111 a113 a115)
-                        "add        %10, #8             \n"
-
-                        "vmlal.s16  q8, d12, d4[0]      \n" // sum0 += (a02-a08) * k02
-                        "vmlal.s16  q9, d12, d4[1]      \n" // sum1 += (a02-a08) * k12
-                        "vmlal.s16  q10, d12, d4[2]     \n" // sum2 += (a02-a08) * k22
-                        "vmlal.s16  q11, d12, d4[3]     \n" // sum3 += (a02-a08) * k32
-
-                        "pld        [%12, #64]         \n"
-                        "vld1.s8    {d0-d2}, [%12]!     \n" // d0(k03-k73) d1(k04-k74) d2(k05-k75)
-
-                        "vmlal.s16  q12, d12, d5[0]     \n" // sum4 += (a02-a08) * k42
-                        "vmlal.s16  q13, d12, d5[1]     \n" // sum5 += (a02-a08) * k52
-                        "vmlal.s16  q14, d12, d5[2]     \n" // sum6 += (a02-a08) * k62
-                        "vmlal.s16  q15, d12, d5[3]     \n" // sum7 += (a02-a08) * k72
-
-                        // r1
-                        "vext.s8    d12, d8, d8, #1     \n" // d12(a12 a14 a16 a18 x x x x)
-
-                        "vmovl.s8   q2, d2              \n" // q2(k05-k75)
-                        "vmovl.s8   q1, d1              \n" // q1(k04-k74)
-                        "vmovl.s8   q0, d0              \n" // q0(k03-k73)
-                        "vmovl.s8   q5, d9              \n" // q5(a11-a115)
-                        "vmovl.s8   q4, d8              \n" // q4(a10-a114)
-                        "vmovl.s8   q6, d12             \n" // q6(a12-a116)
-
-                        "vmlal.s16  q8, d8, d0[0]       \n" // sum0 += (a10-a16) * k03
-                        "vmlal.s16  q9, d8, d0[1]       \n" // sum1 += (a10-a16) * k13
-                        "vmlal.s16  q10, d8, d0[2]      \n" // sum2 += (a10-a16) * k23
-                        "vmlal.s16  q11, d8, d0[3]      \n" // sum3 += (a10-a16) * k33
-                        "vmlal.s16  q12, d8, d1[0]      \n" // sum4 += (a10-a16) * k43
-                        "vmlal.s16  q13, d8, d1[1]      \n" // sum5 += (a10-a16) * k53
-                        "vmlal.s16  q14, d8, d1[2]      \n" // sum6 += (a10-a16) * k63
-                        "vmlal.s16  q15, d8, d1[3]      \n" // sum7 += (a10-a16) * k73
-
-                        "vmlal.s16  q8, d10, d2[0]      \n" // sum0 += (a11-a17) * k04
-                        "vmlal.s16  q9, d10, d2[1]      \n" // sum1 += (a11-a17) * k14
-                        "vmlal.s16  q10, d10, d2[2]     \n" // sum2 += (a11-a17) * k24
-                        "vmlal.s16  q11, d10, d2[3]     \n" // sum3 += (a11-a17) * k34
-                        "vmlal.s16  q12, d10, d3[0]     \n" // sum4 += (a11-a17) * k44
-                        "vmlal.s16  q13, d10, d3[1]     \n" // sum5 += (a11-a17) * k54
-                        "vmlal.s16  q14, d10, d3[2]     \n" // sum6 += (a11-a17) * k64
-                        "vmlal.s16  q15, d10, d3[3]     \n" // sum7 += (a11-a17) * k74
-
-                        "pld        [%11, #64]         \n"
-                        "vld2.s8    {d8-d9}, [%11]      \n" // d8(a20 a22 a24 a26 a28 a210 a212 a214), d9(a21 a23 a25 a27 a29 a211 a213 a215)
-                        "add        %11, #8             \n"
-
-                        "vmlal.s16  q8, d12, d4[0]      \n" // sum0 += (a12-a18) * k05
-                        "vmlal.s16  q9, d12, d4[1]      \n" // sum1 += (a12-a18) * k15
-                        "vmlal.s16  q10, d12, d4[2]     \n" // sum2 += (a12-a18) * k25
-                        "vmlal.s16  q11, d12, d4[3]     \n" // sum3 += (a12-a18) * k35
-
-                        "pld        [%12, #64]         \n"
-                        "vld1.s8    {d0-d2}, [%12]!     \n" // d0(k06-k76) d1(k07-k77) d2(k08-k78)
-
-                        "vmlal.s16  q12, d12, d5[0]     \n" // sum4 += (a12-a18) * k45
-                        "vmlal.s16  q13, d12, d5[1]     \n" // sum5 += (a12-a18) * k55
-                        "vmlal.s16  q14, d12, d5[2]     \n" // sum6 += (a12-a18) * k65
-                        "vmlal.s16  q15, d12, d5[3]     \n" // sum7 += (a12-a18) * k75
-
-                        // r2
-                        "vext.s8    d12, d8, d8, #1     \n" // d12(a22 a24 a26 a28 x x x x)
-
-                        "vmovl.s8   q2, d2              \n" // q2(k08-k78)
-                        "vmovl.s8   q1, d1              \n" // q1(k07-k77)
-                        "vmovl.s8   q0, d0              \n" // q0(k06-k76)
-                        "vmovl.s8   q5, d9              \n" // q5(a21-a215)
-                        "vmovl.s8   q4, d8              \n" // q4(a20-a214)
-                        "vmovl.s8   q6, d12             \n" // q6(a22-a216)
-
-                        "vmlal.s16  q8, d8, d0[0]       \n" // sum0 += (a20-a26) * k06
-                        "vmlal.s16  q9, d8, d0[1]       \n" // sum1 += (a20-a26) * k16
-                        "vmlal.s16  q10, d8, d0[2]      \n" // sum2 += (a20-a26) * k26
-                        "vmlal.s16  q11, d8, d0[3]      \n" // sum3 += (a20-a26) * k36
-                        "vmlal.s16  q12, d8, d1[0]      \n" // sum4 += (a20-a26) * k46
-                        "vmlal.s16  q13, d8, d1[1]      \n" // sum5 += (a20-a26) * k56
-                        "vmlal.s16  q14, d8, d1[2]      \n" // sum6 += (a20-a26) * k66
-                        "vmlal.s16  q15, d8, d1[3]      \n" // sum7 += (a20-a26) * k76
-
-                        "vmlal.s16  q8, d10, d2[0]      \n" // sum0 += (a21-a27) * k07
-                        "vmlal.s16  q9, d10, d2[1]      \n" // sum1 += (a21-a27) * k17
-                        "vmlal.s16  q10, d10, d2[2]     \n" // sum2 += (a21-a27) * k27
-                        "vmlal.s16  q11, d10, d2[3]     \n" // sum3 += (a21-a27) * k37
-                        "vmlal.s16  q12, d10, d3[0]     \n" // sum4 += (a21-a27) * k47
-                        "vmlal.s16  q13, d10, d3[1]     \n" // sum5 += (a21-a27) * k57
-                        "vmlal.s16  q14, d10, d3[2]     \n" // sum6 += (a21-a27) * k67
-                        "vmlal.s16  q15, d10, d3[3]     \n" // sum7 += (a21-a27) * k77
-
-                        "vmlal.s16  q8, d12, d4[0]      \n" // sum0 += (a22-a28) * k08
-                        "vmlal.s16  q9, d12, d4[1]      \n" // sum1 += (a22-a28) * k18
-                        "vmlal.s16  q10, d12, d4[2]     \n" // sum2 += (a22-a28) * k28
-                        "vmlal.s16  q11, d12, d4[3]     \n" // sum3 += (a22-a28) * k38
-                        "vmlal.s16  q12, d12, d5[0]     \n" // sum4 += (a22-a28) * k48
-                        "vmlal.s16  q13, d12, d5[1]     \n" // sum5 += (a22-a28) * k58
-                        "vmlal.s16  q14, d12, d5[2]     \n" // sum6 += (a22-a28) * k68
-                        "vmlal.s16  q15, d12, d5[3]     \n" // sum7 += (a22-a28) * k78
-
-                        // save s32 to memory
-                        "sub        %12, %12, #72       \n"
-                        "vst1.s32   {d16-d17}, [%1]!    \n" // out0
-                        "vst1.s32   {d18-d19}, [%2]!    \n" // out1
-                        "vst1.s32   {d20-d21}, [%3]!    \n" // out2
-                        "vst1.s32   {d22-d23}, [%4]!    \n" // out3
-                        "subs       %0, #1              \n"
-                        "vst1.s32   {d24-d25}, [%5]!    \n" // out4
-                        "vst1.s32   {d26-d27}, [%6]!    \n" // out5
-                        "vst1.s32   {d28-d29}, [%7]!    \n" // out6
-                        "vst1.s32   {d30-d31}, [%8]!    \n" // out7
-
-                        "bne        0b                  \n"
-                        : "=r"(nn),      // %0
-                        "=r"(outptr0), // %1
-                        "=r"(outptr1), // %2
-                        "=r"(outptr2), // %3
-                        "=r"(outptr3), // %4
-                        "=r"(outptr4), // %5
-                        "=r"(outptr5), // %6
-                        "=r"(outptr6), // %7
-                        "=r"(outptr7), // %8
-                        "=r"(r0),      // %9
-                        "=r"(r1),      // %10
-                        "=r"(r2),      // %11
-                        "=r"(ktmp)     // %12
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(outptr1),
-                        "3"(outptr2),
-                        "4"(outptr3),
-                        "5"(outptr4),
-                        "6"(outptr5),
-                        "7"(outptr6),
-                        "8"(outptr7),
-                        "9"(r0),
-                        "10"(r1),
-                        "11"(r2),
-                        "12"(ktmp)
-                        : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-#if __ARM_NEON
-#if __aarch64__
-                    int8x8_t _r0_s8 = vld1_s8(r0); // (a00 a01 a02 ....)
-                    int8x8_t _r1_s8 = vld1_s8(r1); // (a10 a11 a12 ....)
-                    int8x8_t _r2_s8 = vld1_s8(r2); // (a20 a21 a22 ....)
-
-                    int16x8_t _r0 = vmovl_s8(_r0_s8);
-                    int16x8_t _r1 = vmovl_s8(_r1_s8);
-                    int16x8_t _r2 = vmovl_s8(_r2_s8);
-
-                    int32x4_t _sum03 = {};
-                    int32x4_t _sum47 = {};
-
-                    _sum03 = vld1q_lane_s32(outptr0, _sum03, 0); // out0
-                    _sum03 = vld1q_lane_s32(outptr1, _sum03, 1); // out1
-                    _sum03 = vld1q_lane_s32(outptr2, _sum03, 2); // out2
-                    _sum03 = vld1q_lane_s32(outptr3, _sum03, 3); // out3
-                    _sum47 = vld1q_lane_s32(outptr4, _sum47, 0); // out4
-                    _sum47 = vld1q_lane_s32(outptr5, _sum47, 1); // out5
-                    _sum47 = vld1q_lane_s32(outptr6, _sum47, 2); // out6
-                    _sum47 = vld1q_lane_s32(outptr7, _sum47, 3); // out7
-
-                    // k0 - k2
-                    int8x8_t _k0_8 = vld1_s8(ktmp);      //(k00-k70)
-                    int8x8_t _k1_8 = vld1_s8(ktmp + 8);  //(k01-k71)
-                    int8x8_t _k2_8 = vld1_s8(ktmp + 16); //(k02-k72)
-
-                    int16x8_t _k0 = vmovl_s8(_k0_8);
-                    int16x8_t _k1 = vmovl_s8(_k1_8);
-                    int16x8_t _k2 = vmovl_s8(_k2_8);
-
-                    int32x4_t _sum0 = vmull_laneq_s16(vget_low_s16(_k0), _r0, 0);
-                    int32x4_t _sum0n = vmull_laneq_s16(vget_high_s16(_k0), _r0, 0);
-                    int32x4_t _sum1 = vmull_laneq_s16(vget_low_s16(_k1), _r0, 1);
-                    int32x4_t _sum1n = vmull_laneq_s16(vget_high_s16(_k1), _r0, 1);
-                    _sum03 = vmlal_laneq_s16(_sum03, vget_low_s16(_k2), _r0, 2);
-                    _sum47 = vmlal_laneq_s16(_sum47, vget_high_s16(_k2), _r0, 2);
-
-                    // k3 - k5
-                    _k0_8 = vld1_s8(ktmp + 24); //(k03-k73)
-                    _k1_8 = vld1_s8(ktmp + 32); //(k04-k74)
-                    _k2_8 = vld1_s8(ktmp + 40); //(k05-k75)
-
-                    _k0 = vmovl_s8(_k0_8);
-                    _k1 = vmovl_s8(_k1_8);
-                    _k2 = vmovl_s8(_k2_8);
-
-                    _sum0 = vmlal_laneq_s16(_sum0, vget_low_s16(_k0), _r1, 0);
-                    _sum0n = vmlal_laneq_s16(_sum0n, vget_high_s16(_k0), _r1, 0);
-                    _sum1 = vmlal_laneq_s16(_sum1, vget_low_s16(_k1), _r1, 1);
-                    _sum1n = vmlal_laneq_s16(_sum1n, vget_high_s16(_k1), _r1, 1);
-                    _sum03 = vmlal_laneq_s16(_sum03, vget_low_s16(_k2), _r1, 2);
-                    _sum47 = vmlal_laneq_s16(_sum47, vget_high_s16(_k2), _r1, 2);
-
-                    // k6 - k8
-                    _k0_8 = vld1_s8(ktmp + 48); //(k06-k76)
-                    _k1_8 = vld1_s8(ktmp + 56); //(k07-k77)
-                    _k2_8 = vld1_s8(ktmp + 64); //(k08-k78)
-
-                    _k0 = vmovl_s8(_k0_8);
-                    _k1 = vmovl_s8(_k1_8);
-                    _k2 = vmovl_s8(_k2_8);
-
-                    _sum0 = vmlal_laneq_s16(_sum0, vget_low_s16(_k0), _r2, 0);
-                    _sum0n = vmlal_laneq_s16(_sum0n, vget_high_s16(_k0), _r2, 0);
-                    _sum1 = vmlal_laneq_s16(_sum1, vget_low_s16(_k1), _r2, 1);
-                    _sum1n = vmlal_laneq_s16(_sum1n, vget_high_s16(_k1), _r2, 1);
-                    _sum03 = vmlal_laneq_s16(_sum03, vget_low_s16(_k2), _r2, 2);
-                    _sum47 = vmlal_laneq_s16(_sum47, vget_high_s16(_k2), _r2, 2);
-
-                    _sum0 = vaddq_s32(_sum0, _sum1);
-                    _sum0n = vaddq_s32(_sum0n, _sum1n);
-                    _sum03 = vaddq_s32(_sum03, _sum0);
-                    _sum47 = vaddq_s32(_sum47, _sum0n);
-
-                    vst1q_lane_s32(outptr0, _sum03, 0);
-                    vst1q_lane_s32(outptr1, _sum03, 1);
-                    vst1q_lane_s32(outptr2, _sum03, 2);
-                    vst1q_lane_s32(outptr3, _sum03, 3);
-                    vst1q_lane_s32(outptr4, _sum47, 0);
-                    vst1q_lane_s32(outptr5, _sum47, 1);
-                    vst1q_lane_s32(outptr6, _sum47, 2);
-                    vst1q_lane_s32(outptr7, _sum47, 3);
-
-                    outptr0++;
-                    outptr1++;
-                    outptr2++;
-                    outptr3++;
-                    outptr4++;
-                    outptr5++;
-                    outptr6++;
-                    outptr7++;
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%8, #64]          \n"
-                        "vld1.s8    {d0}, [%8]         \n" // d0(a00 a01 a02 ....)
-                        "pld        [%9, #64]          \n"
-                        "vld1.s8    {d2}, [%9]         \n" // d2(a10 a11 a12 ....)
-                        "pld        [%10, #64]         \n"
-                        "vld1.s8    {d4}, [%10]        \n" // d4(a20 a21 a22 ....)
-
-                        "pld        [%11, #64]         \n"
-                        "vld1.s8    {d6-d8}, [%11]!    \n" // d6(k00-k70) d7(k01-k71) d8(k02-k72)
-
-                        "vmovl.s8   q0, d0             \n" // d0(a00 a01 a02 x)
-                        "vmovl.s8   q1, d2             \n" // d2(a10 a11 a12 x)
-                        "vmovl.s8   q2, d4             \n" // d4(a20 a21 a22 x)
-
-                        "vmovl.s8   q5, d8             \n" // d10(k02-k32) d11(k42-k72)
-                        "vmovl.s8   q4, d7             \n" // d8(k01-k31) d9(k41-k71)
-                        "vmovl.s8   q3, d6             \n" // d6(k00-k30) d7(k40-k70)
-
-                        "vld1.s32   {d20[0]}, [%0]     \n" // out0 q10
-                        "vld1.s32   {d20[1]}, [%1]     \n" // out1
-                        "vld1.s32   {d21[0]}, [%2]     \n" // out2
-                        "vld1.s32   {d21[1]}, [%3]     \n" // out3
-
-                        "pld        [%11, #64]         \n"
-                        "vld1.s8    {d24-d26}, [%11]!  \n"
-                        "vmovl.s8   q14, d26           \n" // d28(k05-k35) d29(k45-k75)
-                        "vmovl.s8   q13, d25           \n" // d26(k04-k34) d27(k44-k74)
-                        "vmovl.s8   q12, d24           \n" // d24(k03-k33) d25(k43-k73)
-
-                        "vld1.s32   {d22[0]}, [%4]     \n" // out4 q11
-                        "vld1.s32   {d22[1]}, [%5]     \n" // out5
-                        "vld1.s32   {d23[0]}, [%6]     \n" // out6
-                        "vld1.s32   {d23[1]}, [%7]     \n" // out7
-
-                        "vmull.s16  q6, d6, d0[0]      \n" // a00 x (k00-k30)
-                        "vmull.s16  q7, d7, d0[0]      \n" // a00 x (k40-k70)
-                        "vmull.s16  q8, d8, d0[1]      \n" // a01 x (k01-k31)
-                        "vmull.s16  q9, d9, d0[1]      \n" // a01 x (k41-k71)
-                        "vmlal.s16  q10, d10, d0[2]    \n" // a02 x (k02-k32)
-                        "vmlal.s16  q11, d11, d0[2]    \n" // a02 x (k42-k72)
-
-                        "pld        [%11, #64]         \n"
-                        "vld1.s8    {d6-d8}, [%11]!    \n"
-                        "vmovl.s8   q5, d8             \n" // d10(k08-k38) d11(k48-k78)
-                        "vmovl.s8   q4, d7             \n" // d8(k07-k37) d9(k47-k77)
-                        "vmovl.s8   q3, d6             \n" // d6(k06-k36) d7(k46-k76)
-
-                        "vmlal.s16  q6, d24, d2[0]     \n" // a10 x (k03-k33)
-                        "vmlal.s16  q7, d25, d2[0]     \n" // a10 x (k43-k73)
-                        "vmlal.s16  q8, d26, d2[1]     \n" // a11 x (k04-k34)
-                        "vmlal.s16  q9, d27, d2[1]     \n" // a11 x (k44-k74)
-                        "vmlal.s16  q10, d28, d2[2]    \n" // a12 x (k05-k35)
-                        "vmlal.s16  q11, d29, d2[2]    \n" // a12 x (k45-k75)
-
-                        "vmlal.s16  q6, d6, d4[0]      \n" // a20 x (k06-k36)
-                        "vmlal.s16  q7, d7, d4[0]      \n" // a20 x (k46-k76)
-                        "vmlal.s16  q8, d8, d4[1]      \n" // a21 x (k07-k37)
-                        "vmlal.s16  q9, d9, d4[1]      \n" // a21 x (k47-k77)
-                        "vmlal.s16  q10, d10, d4[2]    \n" // a22 x (k08-k38)
-                        "vmlal.s16  q11, d11, d4[2]    \n" // a22 x (k48-k78)
-
-                        "vadd.s32   q8, q8, q6         \n"
-                        "vadd.s32   q9, q9, q7         \n"
-
-                        "sub        %11, %11, #72      \n"
-
-                        "vadd.s32   q10, q10, q8       \n"
-                        "vadd.s32   q11, q11, q9       \n"
-
-                        "vst1.s32   {d20[0]}, [%0]!    \n" // out0
-                        "vst1.s32   {d20[1]}, [%1]!    \n" // out1
-                        "vst1.s32   {d21[0]}, [%2]!    \n" // out2
-                        "vst1.s32   {d21[1]}, [%3]!    \n" // out3
-                        "vst1.s32   {d22[0]}, [%4]!    \n" // out4
-                        "vst1.s32   {d22[1]}, [%5]!    \n" // out5
-                        "vst1.s32   {d23[0]}, [%6]!    \n" // out6
-                        "vst1.s32   {d23[1]}, [%7]!    \n" // out7
-
-                        : "=r"(outptr0), // %0
-                        "=r"(outptr1), // %1
-                        "=r"(outptr2), // %2
-                        "=r"(outptr3), // %3
-                        "=r"(outptr4), // %4
-                        "=r"(outptr5), // %5
-                        "=r"(outptr6), // %6
-                        "=r"(outptr7), // %7
-                        "=r"(r0),      // %8
-                        "=r"(r1),      // %9
-                        "=r"(r2),      // %10
-                        "=r"(ktmp)     // %11
-                        : "0"(outptr0),
-                        "1"(outptr1),
-                        "2"(outptr2),
-                        "3"(outptr3),
-                        "4"(outptr4),
-                        "5"(outptr5),
-                        "6"(outptr6),
-                        "7"(outptr7),
-                        "8"(r0),
-                        "9"(r1),
-                        "10"(r2),
-                        "11"(ktmp)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // __ARM_NEON
-                    int sum0 = 0;
-                    int sum1 = 0;
-                    int sum2 = 0;
-                    int sum3 = 0;
-                    int sum4 = 0;
-                    int sum5 = 0;
-                    int sum6 = 0;
-                    int sum7 = 0;
-
-                    sum0 += (int)r0[0] * ktmp[0];
-                    sum1 += (int)r0[0] * ktmp[1];
-                    sum2 += (int)r0[0] * ktmp[2];
-                    sum3 += (int)r0[0] * ktmp[3];
-                    sum4 += (int)r0[0] * ktmp[4];
-                    sum5 += (int)r0[0] * ktmp[5];
-                    sum6 += (int)r0[0] * ktmp[6];
-                    sum7 += (int)r0[0] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += (int)r0[1] * ktmp[0];
-                    sum1 += (int)r0[1] * ktmp[1];
-                    sum2 += (int)r0[1] * ktmp[2];
-                    sum3 += (int)r0[1] * ktmp[3];
-                    sum4 += (int)r0[1] * ktmp[4];
-                    sum5 += (int)r0[1] * ktmp[5];
-                    sum6 += (int)r0[1] * ktmp[6];
-                    sum7 += (int)r0[1] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += (int)r0[2] * ktmp[0];
-                    sum1 += (int)r0[2] * ktmp[1];
-                    sum2 += (int)r0[2] * ktmp[2];
-                    sum3 += (int)r0[2] * ktmp[3];
-                    sum4 += (int)r0[2] * ktmp[4];
-                    sum5 += (int)r0[2] * ktmp[5];
-                    sum6 += (int)r0[2] * ktmp[6];
-                    sum7 += (int)r0[2] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += (int)r1[0] * ktmp[0];
-                    sum1 += (int)r1[0] * ktmp[1];
-                    sum2 += (int)r1[0] * ktmp[2];
-                    sum3 += (int)r1[0] * ktmp[3];
-                    sum4 += (int)r1[0] * ktmp[4];
-                    sum5 += (int)r1[0] * ktmp[5];
-                    sum6 += (int)r1[0] * ktmp[6];
-                    sum7 += (int)r1[0] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += (int)r1[1] * ktmp[0];
-                    sum1 += (int)r1[1] * ktmp[1];
-                    sum2 += (int)r1[1] * ktmp[2];
-                    sum3 += (int)r1[1] * ktmp[3];
-                    sum4 += (int)r1[1] * ktmp[4];
-                    sum5 += (int)r1[1] * ktmp[5];
-                    sum6 += (int)r1[1] * ktmp[6];
-                    sum7 += (int)r1[1] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += (int)r1[2] * ktmp[0];
-                    sum1 += (int)r1[2] * ktmp[1];
-                    sum2 += (int)r1[2] * ktmp[2];
-                    sum3 += (int)r1[2] * ktmp[3];
-                    sum4 += (int)r1[2] * ktmp[4];
-                    sum5 += (int)r1[2] * ktmp[5];
-                    sum6 += (int)r1[2] * ktmp[6];
-                    sum7 += (int)r1[2] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += (int)r2[0] * ktmp[0];
-                    sum1 += (int)r2[0] * ktmp[1];
-                    sum2 += (int)r2[0] * ktmp[2];
-                    sum3 += (int)r2[0] * ktmp[3];
-                    sum4 += (int)r2[0] * ktmp[4];
-                    sum5 += (int)r2[0] * ktmp[5];
-                    sum6 += (int)r2[0] * ktmp[6];
-                    sum7 += (int)r2[0] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += (int)r2[1] * ktmp[0];
-                    sum1 += (int)r2[1] * ktmp[1];
-                    sum2 += (int)r2[1] * ktmp[2];
-                    sum3 += (int)r2[1] * ktmp[3];
-                    sum4 += (int)r2[1] * ktmp[4];
-                    sum5 += (int)r2[1] * ktmp[5];
-                    sum6 += (int)r2[1] * ktmp[6];
-                    sum7 += (int)r2[1] * ktmp[7];
-                    ktmp += 8;
-
-                    sum0 += (int)r2[2] * ktmp[0];
-                    sum1 += (int)r2[2] * ktmp[1];
-                    sum2 += (int)r2[2] * ktmp[2];
-                    sum3 += (int)r2[2] * ktmp[3];
-                    sum4 += (int)r2[2] * ktmp[4];
-                    sum5 += (int)r2[2] * ktmp[5];
-                    sum6 += (int)r2[2] * ktmp[6];
-                    sum7 += (int)r2[2] * ktmp[7];
-                    ktmp += 8;
-
-                    *outptr0 += sum0;
-                    *outptr1 += sum1;
-                    *outptr2 += sum2;
-                    *outptr3 += sum3;
-                    *outptr4 += sum4;
-                    *outptr5 += sum5;
-                    *outptr6 += sum6;
-                    *outptr7 += sum7;
-
-                    ktmp -= 8 * 9;
-
-                    outptr0++;
-                    outptr1++;
-                    outptr2++;
-                    outptr3++;
-                    outptr4++;
-                    outptr5++;
-                    outptr6++;
-                    outptr7++;
-#endif // __ARM_NEON
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            ktmp += 8 * 9;
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        out.fill(0);
-
-        const signed char* ktmp = _kernel.channel(p / 8 + p % 8);
-
-        for (int q = 0; q < inch; q++)
-        {
-            int* outptr = out;
-
-            const signed char* img0 = bottom_blob.channel(q);
-
-            const signed char* r0 = img0;
-            const signed char* r1 = img0 + w;
-            const signed char* r2 = img0 + w * 2;
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 3;
-                int remain = outw & 7;
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "0:                                   \n"
-
-                        "ld1    {v0.8b, v1.8b}, [%5]          \n" //ktmp
-                        "ld2    {v2.8b, v3.8b}, [%2], #16     \n" //r0-r2
-                        "ld2    {v4.8b, v5.8b}, [%2]          \n"
-
-                        "ld2    {v6.8b, v7.8b}, [%3], #16     \n" //r3-r5
-                        "ld2    {v8.8b, v9.8b}, [%3]          \n"
-
-                        "ld2    {v10.8b, v11.8b}, [%4], #16   \n" //r6-r8
-                        "ld2    {v12.8b, v13.8b}, [%4]        \n"
-
-                        "ld1    {v14.4s, v15.4s}, [%1]        \n" //out0
-
-                        "ext    v4.8b, v2.8b, v4.8b, #1       \n"
-                        "ext    v8.8b, v6.8b, v8.8b, #1       \n"
-                        "ext    v12.8b, v10.8b, v12.8b, #1    \n"
-
-                        "sshll  v0.8h, v0.8b, #0              \n" //(k0-k7)
-                        "sshll  v1.8h, v1.8b, #0              \n" //(k8)
-                        "sshll  v2.8h, v2.8b, #0              \n" // r0
-                        "sshll  v3.8h, v3.8b, #0              \n" // r1
-                        "sshll  v4.8h, v4.8b, #0              \n" // r2
-                        "sshll  v6.8h, v6.8b, #0              \n" // r3
-                        "sshll  v7.8h, v7.8b, #0              \n" // r4
-                        "sshll  v8.8h, v8.8b, #0              \n" // r5
-                        "sshll  v10.8h, v10.8b, #0            \n" // r6
-                        "sshll  v11.8h, v11.8b, #0            \n" // r7
-                        "sshll  v12.8h, v12.8b, #0            \n" // r8
-
-                        // r0
-                        "smull  v16.4s, v2.4h, v0.h[0]        \n" // out = r0*k0
-                        "smull2  v17.4s, v2.8h, v0.h[0]       \n"
-                        "smull  v18.4s, v3.4h, v0.h[1]        \n" // outn = r1*k1
-                        "smull2  v19.4s, v3.8h, v0.h[1]       \n"
-                        "smlal  v16.4s, v4.4h, v0.h[2]        \n" // out = r2*k2
-                        "smlal2  v17.4s, v4.8h, v0.h[2]       \n"
-                        "smlal  v18.4s, v6.4h, v0.h[3]        \n" // outn = r3*k3
-                        "smlal2  v19.4s, v6.8h, v0.h[3]       \n"
-                        "smlal  v16.4s, v7.4h, v0.h[4]        \n" // out = r4*k4
-                        "smlal2  v17.4s, v7.8h, v0.h[4]       \n"
-                        "smlal  v18.4s, v8.4h, v0.h[5]        \n" // outn = r5*k5
-                        "smlal2  v19.4s, v8.8h, v0.h[5]       \n"
-                        "smlal  v16.4s, v10.4h, v0.h[6]       \n" // out = r6*k6
-                        "smlal2  v17.4s, v10.8h, v0.h[6]      \n"
-                        "smlal  v18.4s, v11.4h, v0.h[7]       \n" // outn = r7*k7
-                        "smlal2  v19.4s, v11.8h, v0.h[7]      \n"
-                        "smlal  v16.4s, v12.4h, v1.h[0]       \n" // out = r8*k8
-                        "smlal2  v17.4s, v12.8h, v1.h[0]      \n"
-
-                        "add    v8.4s, v16.4s, v18.4s         \n"
-                        "add    v9.4s, v17.4s, v19.4s         \n"
-
-                        "st1    {v8.4s, v9.4s}, [%1], #32     \n"
-
-                        "subs   %w0, %w0, #1                  \n"
-
-                        "bne    0b                            \n"
-
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2),     // %4
-                        "=r"(ktmp)    // %5
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(ktmp)
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "vld1.s8    {d0-d1}, [%5]       \n" // d0(k0 - k7) d1(k8 ...)
-                        "vmovl.s8   q1, d1              \n" // d2(k8 ...)
-                        "vmovl.s8   q0, d0              \n" // d0(k0 - k3) d1(k4 - k7)
-                        "0:                             \n"
-                        "pld        [%2, #192]          \n"
-                        "vld2.s8    {d4-d5}, [%2]!      \n" // r0 d4(a00 a02 ... a014) d5(a01 a03 ... a015)
-                        "vld2.s8    {d8-d9}, [%2]       \n" //    d8(a016 ....)
-                        "vld2.s8    {d10-d11}, [%3]!    \n" // r1 d10(a10 a12 ... a114) d11(a11 a13 ... a115)
-                        "vld2.s8    {d14-d15}, [%3]     \n" //    d14(a116 ....)
-                        "vld2.s8    {d16-d17}, [%4]!    \n" // r2 d16(a20 a22 ... a214) d17(a21 a23 ... a215)
-                        "vld2.s8    {d20-d21}, [%4]     \n" //    d20(a216 ....)
-                        "vld1.s32   {d22-d25}, [%1]     \n" // q11(out0 - out3) q12(out4 - out7)
-
-                        "vext.s8    d8, d4, d8, #1      \n" //  d8(a02 a04 ... a016)
-                        "vext.s8    d14, d10, d14, #1   \n" // d14(a12 a14 ... a116)
-                        "vext.s8    d20, d16, d20, #1   \n" // d20(a22 a24 ... a216)
-
-                        "vmovl.s8   q3, d5              \n" // q3(a01 a03 ... a015)
-                        "vmovl.s8   q2, d4              \n" // q2(a00 a02 ... a014)
-                        "vmovl.s8   q4, d8              \n" // q4(a02 a04 ... a016)
-
-                        "vmovl.s8   q6, d11             \n" // q6(a11 a13 ... a115)
-                        "vmovl.s8   q5, d10             \n" // q5(a10 a12 ... a114)
-                        "vmovl.s8   q7, d14             \n" // q7(a12 a14 ... a116)
-
-                        "vmovl.s8   q9, d17             \n" // q9(a21 a23 ... a215)
-                        "vmovl.s8   q8, d16             \n" // q8(a20 a22 ... a214)
-                        "vmovl.s8   q10, d20            \n" // q10(a22 a24 ... a216)
-
-                        "vmlal.s16  q11, d4, d0[0]      \n" // k0
-                        "vmlal.s16  q12, d5, d0[0]      \n"
-                        "vmull.s16  q13, d6, d0[1]      \n" // k1
-                        "vmull.s16  q14, d7, d0[1]      \n"
-                        "vmlal.s16  q11, d8, d0[2]      \n" // k2
-                        "vmlal.s16  q12, d9, d0[2]      \n"
-
-                        "vmlal.s16  q13, d12, d1[0]     \n" // k4
-                        "vmlal.s16  q14, d13, d1[0]     \n"
-                        "vmlal.s16  q11, d10, d0[3]     \n" // k3
-                        "vmlal.s16  q12, d11, d0[3]     \n"
-                        "vmlal.s16  q13, d14, d1[1]     \n" // k5
-                        "vmlal.s16  q14, d15, d1[1]     \n"
-
-                        "vmlal.s16  q11, d16, d1[2]     \n" // k6
-                        "vmlal.s16  q12, d17, d1[2]     \n"
-                        "vmlal.s16  q13, d18, d1[3]     \n" // k7
-                        "vmlal.s16  q14, d19, d1[3]     \n"
-                        "vmlal.s16  q11, d20, d2[0]     \n" // k8
-                        "vmlal.s16  q12, d21, d2[0]     \n"
-
-                        "vadd.s32   q11, q11, q13       \n"
-                        "vadd.s32   q12, q12, q14       \n"
-
-                        "vst1.32    {d22-d25}, [%1]!    \n"
-
-                        "subs       %0, #1              \n"
-                        "bne        0b                  \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2),     // %4
-                        "=r"(ktmp)    // %5
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(ktmp)
-                        : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                if (remain > 0)
-                {
-#if __ARM_NEON
-                    int8x8_t _k01234567s8 = vld1_s8(ktmp);
-                    int8x8_t _k8xxxxxxxs8 = vld1_s8(ktmp + 8);
-                    int8x8_t _k34567xxxs8 = vext_s8(_k01234567s8, _k01234567s8, 3);
-                    int8x8_t _k678xxxxxs8 = vext_s8(_k01234567s8, _k8xxxxxxxs8, 6);
-                    int16x8_t _k0123_s16 = vmovl_s8(_k01234567s8);
-                    int16x8_t _k3456_s16 = vmovl_s8(_k34567xxxs8);
-                    int16x8_t _k678x_s16 = vmovl_s8(_k678xxxxxs8);
-#endif
-                    for (; remain > 0; remain--)
-                    {
-#if __ARM_NEON
-                        int8x8_t _r00s8 = vld1_s8(r0);
-                        int8x8_t _r10s8 = vld1_s8(r1);
-                        int8x8_t _r20s8 = vld1_s8(r2);
-
-                        int16x8_t _r00s16 = vmovl_s8(_r00s8);
-                        int16x8_t _r10s16 = vmovl_s8(_r10s8);
-                        int16x8_t _r20s16 = vmovl_s8(_r20s8);
-
-                        int32x4_t _sum = vmull_s16(vget_low_s16(_r00s16), vget_low_s16(_k0123_s16));
-                        _sum = vmlal_s16(_sum, vget_low_s16(_r10s16), vget_low_s16(_k3456_s16));
-                        _sum = vmlal_s16(_sum, vget_low_s16(_r20s16), vget_low_s16(_k678x_s16));
-
-                        _sum = vsetq_lane_s32(*outptr, _sum, 3);
-
-#if __aarch64__
-                        *outptr = vaddvq_s32(_sum);
-#else
-                        int32x2_t _ss = vadd_s32(vget_low_s32(_sum), vget_high_s32(_sum));
-                        _ss = vpadd_s32(_ss, _ss);
-
-                        *outptr = vget_lane_s32(_ss, 0);
-#endif // __aarch64__
-#else
-                        int sum = 0;
-
-                        sum += (int)r0[0] * ktmp[0];
-                        sum += (int)r0[1] * ktmp[1];
-                        sum += (int)r0[2] * ktmp[2];
-                        sum += (int)r1[0] * ktmp[3];
-                        sum += (int)r1[1] * ktmp[4];
-                        sum += (int)r1[2] * ktmp[5];
-                        sum += (int)r2[0] * ktmp[6];
-                        sum += (int)r2[1] * ktmp[7];
-                        sum += (int)r2[2] * ktmp[8];
-
-                        *outptr += sum;
-#endif // __ARM_NEON
-                        r0 += 2;
-                        r1 += 2;
-                        r2 += 2;
-                        outptr++;
-                    }
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            ktmp += 9;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to4.h
deleted file mode 100644
index e1e452a9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to4.h
+++ /dev/null
@@ -1,1449 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack1to4_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    int remain_outch_start = 0;
-
-#if __ARM_NEON && __aarch64__
-    int nn_outch = 0;
-    nn_outch = outch >> 1;
-    remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        Mat out0 = top_blob.channel(p);
-        Mat out1 = top_blob.channel(p + 1);
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        float32x4_t _bias1 = bias ? vld1q_f32((const float*)bias + (p + 1) * 4) : vdupq_n_f32(0.f);
-        out0.fill(_bias0);
-        out1.fill(_bias1);
-
-        const float* k0 = kernel.channel(p);
-        const float* k1 = kernel.channel(p + 1);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            float32x4_t _k00_0 = vld1q_f32(k0);
-            float32x4_t _k01_0 = vld1q_f32(k0 + 4);
-            float32x4_t _k02_0 = vld1q_f32(k0 + 8);
-            float32x4_t _k10_0 = vld1q_f32(k0 + 12);
-            float32x4_t _k11_0 = vld1q_f32(k0 + 16);
-            float32x4_t _k12_0 = vld1q_f32(k0 + 20);
-            float32x4_t _k20_0 = vld1q_f32(k0 + 24);
-            float32x4_t _k21_0 = vld1q_f32(k0 + 28);
-            float32x4_t _k22_0 = vld1q_f32(k0 + 32);
-
-            float32x4_t _k00_1 = vld1q_f32(k1);
-            float32x4_t _k01_1 = vld1q_f32(k1 + 4);
-            float32x4_t _k02_1 = vld1q_f32(k1 + 8);
-            float32x4_t _k10_1 = vld1q_f32(k1 + 12);
-            float32x4_t _k11_1 = vld1q_f32(k1 + 16);
-            float32x4_t _k12_1 = vld1q_f32(k1 + 20);
-            float32x4_t _k20_1 = vld1q_f32(k1 + 24);
-            float32x4_t _k21_1 = vld1q_f32(k1 + 28);
-            float32x4_t _k22_1 = vld1q_f32(k1 + 32);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-
-                for (; j + 3 < outw; j += 4)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0] \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%1] \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v0.4s}, [%2], #16          \n"
-
-                        "ld1    {v1.2s}, [%2]               \n"
-
-                        "fmla   v24.4s, %10.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %10.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %10.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %10.4s, v0.s[3]     \n"
-                        "fmla   v28.4s, %19.4s, v0.s[0]     \n"
-                        "fmla   v29.4s, %19.4s, v0.s[1]     \n"
-                        "fmla   v30.4s, %19.4s, v0.s[2]     \n"
-                        "fmla   v31.4s, %19.4s, v0.s[3]     \n"
-
-                        "fmla   v24.4s, %11.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %11.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %11.4s, v0.s[3]     \n"
-                        "fmla   v27.4s, %11.4s, v1.s[0]     \n"
-                        "fmla   v28.4s, %20.4s, v0.s[1]     \n"
-                        "fmla   v29.4s, %20.4s, v0.s[2]     \n"
-                        "fmla   v30.4s, %20.4s, v0.s[3]     \n"
-                        "fmla   v31.4s, %20.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v2.4s}, [%3], #16          \n"
-
-                        "ld1    {v3.2s}, [%3]               \n"
-
-                        "fmla   v24.4s, %12.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %12.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %12.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %12.4s, v1.s[1]     \n"
-                        "fmla   v28.4s, %21.4s, v0.s[2]     \n"
-                        "fmla   v29.4s, %21.4s, v0.s[3]     \n"
-                        "fmla   v30.4s, %21.4s, v1.s[0]     \n"
-                        "fmla   v31.4s, %21.4s, v1.s[1]     \n"
-
-                        "fmla   v24.4s, %13.4s, v2.s[0]     \n"
-                        "fmla   v25.4s, %13.4s, v2.s[1]     \n"
-                        "fmla   v26.4s, %13.4s, v2.s[2]     \n"
-                        "fmla   v27.4s, %13.4s, v2.s[3]     \n"
-                        "fmla   v28.4s, %22.4s, v2.s[0]     \n"
-                        "fmla   v29.4s, %22.4s, v2.s[1]     \n"
-                        "fmla   v30.4s, %22.4s, v2.s[2]     \n"
-                        "fmla   v31.4s, %22.4s, v2.s[3]     \n"
-
-                        "fmla   v24.4s, %14.4s, v2.s[1]     \n"
-                        "fmla   v25.4s, %14.4s, v2.s[2]     \n"
-                        "fmla   v26.4s, %14.4s, v2.s[3]     \n"
-                        "fmla   v27.4s, %14.4s, v3.s[0]     \n"
-                        "fmla   v28.4s, %23.4s, v2.s[1]     \n"
-                        "fmla   v29.4s, %23.4s, v2.s[2]     \n"
-                        "fmla   v30.4s, %23.4s, v2.s[3]     \n"
-                        "fmla   v31.4s, %23.4s, v3.s[0]     \n"
-
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v0.4s}, [%4], #16          \n"
-
-                        "ld1    {v1.2s}, [%4]               \n"
-
-                        "fmla   v24.4s, %15.4s, v2.s[2]     \n"
-                        "fmla   v25.4s, %15.4s, v2.s[3]     \n"
-                        "fmla   v26.4s, %15.4s, v3.s[0]     \n"
-                        "fmla   v27.4s, %15.4s, v3.s[1]     \n"
-                        "fmla   v28.4s, %24.4s, v2.s[2]     \n"
-                        "fmla   v29.4s, %24.4s, v2.s[3]     \n"
-                        "fmla   v30.4s, %24.4s, v3.s[0]     \n"
-                        "fmla   v31.4s, %24.4s, v3.s[1]     \n"
-
-                        "fmla   v24.4s, %16.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %16.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %16.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %16.4s, v0.s[3]     \n"
-                        "fmla   v28.4s, %25.4s, v0.s[0]     \n"
-                        "fmla   v29.4s, %25.4s, v0.s[1]     \n"
-                        "fmla   v30.4s, %25.4s, v0.s[2]     \n"
-                        "fmla   v31.4s, %25.4s, v0.s[3]     \n"
-
-                        "fmla   v24.4s, %17.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %17.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %17.4s, v0.s[3]     \n"
-                        "fmla   v27.4s, %17.4s, v1.s[0]     \n"
-                        "fmla   v28.4s, %26.4s, v0.s[1]     \n"
-                        "fmla   v29.4s, %26.4s, v0.s[2]     \n"
-                        "fmla   v30.4s, %26.4s, v0.s[3]     \n"
-                        "fmla   v31.4s, %26.4s, v1.s[0]     \n"
-
-                        "fmla   v24.4s, %18.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %18.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %18.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %18.4s, v1.s[1]     \n"
-                        "fmla   v28.4s, %27.4s, v0.s[2]     \n"
-                        "fmla   v29.4s, %27.4s, v0.s[3]     \n"
-                        "fmla   v30.4s, %27.4s, v1.s[0]     \n"
-                        "fmla   v31.4s, %27.4s, v1.s[1]     \n"
-
-                        "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%1], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(outptr1), // %1
-                        "=r"(r0),      // %2
-                        "=r"(r1),      // %3
-                        "=r"(r2)       // %4
-                        : "0"(outptr0),
-                        "1"(outptr1),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00_0), // %10
-                        "w"(_k01_0), // %11
-                        "w"(_k02_0), // %12
-                        "w"(_k10_0), // %13
-                        "w"(_k11_0), // %14
-                        "w"(_k12_0), // %15
-                        "w"(_k20_0), // %16
-                        "w"(_k21_0), // %17
-                        "w"(_k22_0), // %18
-                        "w"(_k00_1), // %19
-                        "w"(_k01_1), // %20
-                        "w"(_k02_1), // %21
-                        "w"(_k10_1), // %22
-                        "w"(_k11_1), // %23
-                        "w"(_k12_1), // %24
-                        "w"(_k20_1), // %25
-                        "w"(_k21_1), // %26
-                        "w"(_k22_1)  // %27
-                        : "memory", "v0", "v1", "v2", "v3", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v24.4s, v25.4s}, [%0]      \n"
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v26.4s, v27.4s}, [%1]      \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v0.4s}, [%2]               \n"
-                        "add    %2, %2, #8                  \n"
-
-                        "fmla   v24.4s, %10.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %10.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %19.4s, v0.s[0]     \n"
-                        "fmla   v27.4s, %19.4s, v0.s[1]     \n"
-
-                        "fmla   v24.4s, %11.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %11.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %20.4s, v0.s[1]     \n"
-                        "fmla   v27.4s, %20.4s, v0.s[2]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v1.4s}, [%3]               \n"
-
-                        "fmla   v24.4s, %12.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %12.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %21.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %21.4s, v0.s[3]     \n"
-
-                        "add    %3, %3, #8                  \n"
-
-                        "fmla   v24.4s, %13.4s, v1.s[0]     \n"
-                        "fmla   v25.4s, %13.4s, v1.s[1]     \n"
-                        "fmla   v26.4s, %22.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %22.4s, v1.s[1]     \n"
-
-                        "fmla   v24.4s, %14.4s, v1.s[1]     \n"
-                        "fmla   v25.4s, %14.4s, v1.s[2]     \n"
-                        "fmla   v26.4s, %23.4s, v1.s[1]     \n"
-                        "fmla   v27.4s, %23.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v0.4s}, [%4]               \n"
-
-                        "fmla   v24.4s, %15.4s, v1.s[2]     \n"
-                        "fmla   v25.4s, %15.4s, v1.s[3]     \n"
-                        "fmla   v26.4s, %24.4s, v1.s[2]     \n"
-                        "fmla   v27.4s, %24.4s, v1.s[3]     \n"
-
-                        "add    %4, %4, #8                  \n"
-
-                        "fmla   v24.4s, %16.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %16.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %25.4s, v0.s[0]     \n"
-                        "fmla   v27.4s, %25.4s, v0.s[1]     \n"
-
-                        "fmla   v24.4s, %17.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %17.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %26.4s, v0.s[1]     \n"
-                        "fmla   v27.4s, %26.4s, v0.s[2]     \n"
-
-                        "fmla   v24.4s, %18.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %18.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %27.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %27.4s, v0.s[3]     \n"
-
-                        "st1    {v24.4s, v25.4s}, [%0], #32 \n"
-                        "st1    {v26.4s, v27.4s}, [%1], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(outptr1), // %1
-                        "=r"(r0),      // %2
-                        "=r"(r1),      // %3
-                        "=r"(r2)       // %4
-                        : "0"(outptr0),
-                        "1"(outptr1),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00_0), // %10
-                        "w"(_k01_0), // %11
-                        "w"(_k02_0), // %12
-                        "w"(_k10_0), // %13
-                        "w"(_k11_0), // %14
-                        "w"(_k12_0), // %15
-                        "w"(_k20_0), // %16
-                        "w"(_k21_0), // %17
-                        "w"(_k22_0), // %18
-                        "w"(_k00_1), // %19
-                        "w"(_k01_1), // %20
-                        "w"(_k02_1), // %21
-                        "w"(_k10_1), // %22
-                        "w"(_k11_1), // %23
-                        "w"(_k12_1), // %24
-                        "w"(_k20_1), // %25
-                        "w"(_k21_1), // %26
-                        "w"(_k22_1)  // %27
-                        : "memory", "v0", "v1", "v24", "v25", "v26", "v27");
-                }
-                for (; j < outw; j++)
-                {
-                    float32x4_t _sum00 = vld1q_f32(outptr0);
-                    float32x4_t _sum10 = vld1q_f32(outptr1);
-
-                    float32x4_t _r0 = vld1q_f32(r0);
-                    float32x4_t _r1 = vld1q_f32(r1);
-                    float32x4_t _r2 = vld1q_f32(r2);
-
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k00_0, _r0, 0);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k01_0, _r0, 1);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k02_0, _r0, 2);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k10_0, _r1, 0);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k11_0, _r1, 1);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k12_0, _r1, 2);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k20_0, _r2, 0);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k21_0, _r2, 1);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k22_0, _r2, 2);
-
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k00_1, _r0, 0);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k01_1, _r0, 1);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k02_1, _r0, 2);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k10_1, _r1, 0);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k11_1, _r1, 1);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k12_1, _r1, 2);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k20_1, _r2, 0);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k21_1, _r2, 1);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k22_1, _r2, 2);
-
-                    vst1q_f32(outptr0, _sum00);
-                    vst1q_f32(outptr1, _sum10);
-
-                    r0 += 1;
-                    r1 += 1;
-                    r2 += 1;
-                    outptr0 += 4;
-                    outptr1 += 4;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * 4;
-            k1 += 9 * 4;
-        }
-    }
-#endif // __ARM_NEON && __aarch64__
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        out0.fill(_bias0);
-
-        const float* k0 = kernel.channel(p);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            float32x4_t _k00 = vld1q_f32(k0);
-            float32x4_t _k01 = vld1q_f32(k0 + 4);
-            float32x4_t _k02 = vld1q_f32(k0 + 8);
-            float32x4_t _k10 = vld1q_f32(k0 + 12);
-            float32x4_t _k11 = vld1q_f32(k0 + 16);
-            float32x4_t _k12 = vld1q_f32(k0 + 20);
-            float32x4_t _k20 = vld1q_f32(k0 + 24);
-            float32x4_t _k21 = vld1q_f32(k0 + 28);
-            float32x4_t _k22 = vld1q_f32(k0 + 32);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-
-#if __aarch64__
-                for (; j + 7 < outw; j += 8)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%1], #32   \n"
-
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0] \n"
-
-                        "fmla   v24.4s, %8.4s, v0.s[0]      \n"
-                        "fmla   v25.4s, %8.4s, v0.s[1]      \n"
-                        "fmla   v26.4s, %8.4s, v0.s[2]      \n"
-                        "fmla   v27.4s, %8.4s, v0.s[3]      \n"
-                        "fmla   v28.4s, %8.4s, v1.s[0]      \n"
-                        "fmla   v29.4s, %8.4s, v1.s[1]      \n"
-                        "fmla   v30.4s, %8.4s, v1.s[2]      \n"
-                        "fmla   v31.4s, %8.4s, v1.s[3]      \n"
-
-                        "ld1    {v2.2s}, [%1]               \n"
-
-                        "fmla   v24.4s, %9.4s, v0.s[1]      \n"
-                        "fmla   v25.4s, %9.4s, v0.s[2]      \n"
-                        "fmla   v26.4s, %9.4s, v0.s[3]      \n"
-                        "fmla   v27.4s, %9.4s, v1.s[0]      \n"
-                        "fmla   v28.4s, %9.4s, v1.s[1]      \n"
-                        "fmla   v29.4s, %9.4s, v1.s[2]      \n"
-                        "fmla   v30.4s, %9.4s, v1.s[3]      \n"
-                        "fmla   v31.4s, %9.4s, v2.s[0]      \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v4.4s, v5.4s}, [%2], #32   \n"
-
-                        "fmla   v24.4s, %10.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %10.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %10.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %10.4s, v1.s[1]     \n"
-                        "fmla   v28.4s, %10.4s, v1.s[2]     \n"
-                        "fmla   v29.4s, %10.4s, v1.s[3]     \n"
-                        "fmla   v30.4s, %10.4s, v2.s[0]     \n"
-                        "fmla   v31.4s, %10.4s, v2.s[1]     \n"
-
-                        "ld1    {v2.2s}, [%2]               \n"
-
-                        "fmla   v24.4s, %11.4s, v4.s[0]     \n"
-                        "fmla   v25.4s, %11.4s, v4.s[1]     \n"
-                        "fmla   v26.4s, %11.4s, v4.s[2]     \n"
-                        "fmla   v27.4s, %11.4s, v4.s[3]     \n"
-                        "fmla   v28.4s, %11.4s, v5.s[0]     \n"
-                        "fmla   v29.4s, %11.4s, v5.s[1]     \n"
-                        "fmla   v30.4s, %11.4s, v5.s[2]     \n"
-                        "fmla   v31.4s, %11.4s, v5.s[3]     \n"
-
-                        "fmla   v24.4s, %12.4s, v4.s[1]     \n"
-                        "fmla   v25.4s, %12.4s, v4.s[2]     \n"
-                        "fmla   v26.4s, %12.4s, v4.s[3]     \n"
-                        "fmla   v27.4s, %12.4s, v5.s[0]     \n"
-                        "fmla   v28.4s, %12.4s, v5.s[1]     \n"
-                        "fmla   v29.4s, %12.4s, v5.s[2]     \n"
-                        "fmla   v30.4s, %12.4s, v5.s[3]     \n"
-                        "fmla   v31.4s, %12.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%3], #32   \n"
-
-                        "fmla   v24.4s, %13.4s, v4.s[2]     \n"
-                        "fmla   v25.4s, %13.4s, v4.s[3]     \n"
-                        "fmla   v26.4s, %13.4s, v5.s[0]     \n"
-                        "fmla   v27.4s, %13.4s, v5.s[1]     \n"
-                        "fmla   v28.4s, %13.4s, v5.s[2]     \n"
-                        "fmla   v29.4s, %13.4s, v5.s[3]     \n"
-                        "fmla   v30.4s, %13.4s, v2.s[0]     \n"
-                        "fmla   v31.4s, %13.4s, v2.s[1]     \n"
-
-                        "ld1    {v2.2s}, [%3]               \n"
-
-                        "fmla   v24.4s, %14.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %14.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %14.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %14.4s, v0.s[3]     \n"
-                        "fmla   v28.4s, %14.4s, v1.s[0]     \n"
-                        "fmla   v29.4s, %14.4s, v1.s[1]     \n"
-                        "fmla   v30.4s, %14.4s, v1.s[2]     \n"
-                        "fmla   v31.4s, %14.4s, v1.s[3]     \n"
-
-                        "fmla   v24.4s, %15.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %15.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %15.4s, v0.s[3]     \n"
-                        "fmla   v27.4s, %15.4s, v1.s[0]     \n"
-                        "fmla   v28.4s, %15.4s, v1.s[1]     \n"
-                        "fmla   v29.4s, %15.4s, v1.s[2]     \n"
-                        "fmla   v30.4s, %15.4s, v1.s[3]     \n"
-                        "fmla   v31.4s, %15.4s, v2.s[0]     \n"
-
-                        "sub    %0, %0, #64                 \n"
-
-                        "fmla   v24.4s, %16.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %16.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %16.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %16.4s, v1.s[1]     \n"
-                        "fmla   v28.4s, %16.4s, v1.s[2]     \n"
-                        "fmla   v29.4s, %16.4s, v1.s[3]     \n"
-                        "fmla   v30.4s, %16.4s, v2.s[0]     \n"
-                        "fmla   v31.4s, %16.4s, v2.s[1]     \n"
-
-                        "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "v0", "v1", "v2", "v4", "v5", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-#endif // __aarch64__
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0] \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.4s}, [%1], #16          \n"
-
-                        "fmla   v24.4s, %8.4s, v0.s[0]      \n"
-                        "fmla   v25.4s, %8.4s, v0.s[1]      \n"
-                        "fmla   v26.4s, %8.4s, v0.s[2]      \n"
-                        "fmla   v27.4s, %8.4s, v0.s[3]      \n"
-
-                        "ld1    {v1.2s}, [%1]               \n"
-
-                        "fmla   v24.4s, %9.4s, v0.s[1]      \n"
-                        "fmla   v25.4s, %9.4s, v0.s[2]      \n"
-                        "fmla   v26.4s, %9.4s, v0.s[3]      \n"
-                        "fmla   v27.4s, %9.4s, v1.s[0]      \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v2.4s}, [%2], #16          \n"
-
-                        "fmla   v24.4s, %10.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %10.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %10.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %10.4s, v1.s[1]     \n"
-
-                        "ld1    {v3.2s}, [%2]               \n"
-
-                        "fmla   v24.4s, %11.4s, v2.s[0]     \n"
-                        "fmla   v25.4s, %11.4s, v2.s[1]     \n"
-                        "fmla   v26.4s, %11.4s, v2.s[2]     \n"
-                        "fmla   v27.4s, %11.4s, v2.s[3]     \n"
-
-                        "fmla   v24.4s, %12.4s, v2.s[1]     \n"
-                        "fmla   v25.4s, %12.4s, v2.s[2]     \n"
-                        "fmla   v26.4s, %12.4s, v2.s[3]     \n"
-                        "fmla   v27.4s, %12.4s, v3.s[0]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.4s}, [%3], #16          \n"
-
-                        "fmla   v24.4s, %13.4s, v2.s[2]     \n"
-                        "fmla   v25.4s, %13.4s, v2.s[3]     \n"
-                        "fmla   v26.4s, %13.4s, v3.s[0]     \n"
-                        "fmla   v27.4s, %13.4s, v3.s[1]     \n"
-
-                        "ld1    {v1.2s}, [%3]               \n"
-
-                        "fmla   v24.4s, %14.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %14.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %14.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %14.4s, v0.s[3]     \n"
-
-                        "fmla   v24.4s, %15.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %15.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %15.4s, v0.s[3]     \n"
-                        "fmla   v27.4s, %15.4s, v1.s[0]     \n"
-
-                        "fmla   v24.4s, %16.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %16.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %16.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %16.4s, v1.s[1]     \n"
-
-                        "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "v0", "v1", "v2", "v3", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #512]          \n"
-                        "vldm       %0, {d24-d31}       \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%1]!      \n"
-
-                        "vmla.f32   q12, %q8, d0[0]     \n"
-                        "vmla.f32   q13, %q8, d0[1]     \n"
-                        "vmla.f32   q14, %q8, d1[0]     \n"
-                        "vmla.f32   q15, %q8, d1[1]     \n"
-
-                        "vld1.f32   {d2}, [%1]          \n"
-
-                        "vmla.f32   q12, %q9, d0[1]     \n"
-                        "vmla.f32   q13, %q9, d1[0]     \n"
-                        "vmla.f32   q14, %q9, d1[1]     \n"
-                        "vmla.f32   q15, %q9, d2[0]     \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d4-d5}, [%2]!      \n"
-
-                        "vmla.f32   q12, %q10, d1[0]    \n"
-                        "vmla.f32   q13, %q10, d1[1]    \n"
-                        "vmla.f32   q14, %q10, d2[0]    \n"
-                        "vmla.f32   q15, %q10, d2[1]    \n"
-
-                        "vmla.f32   q12, %q11, d4[0]    \n"
-                        "vmla.f32   q13, %q11, d4[1]    \n"
-                        "vmla.f32   q14, %q11, d5[0]    \n"
-                        "vmla.f32   q15, %q11, d5[1]    \n"
-
-                        "vld1.f32   {d3}, [%2]          \n"
-
-                        "vmla.f32   q12, %q12, d4[1]    \n"
-                        "vmla.f32   q13, %q12, d5[0]    \n"
-                        "vmla.f32   q14, %q12, d5[1]    \n"
-                        "vmla.f32   q15, %q12, d3[0]    \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%3]!      \n"
-
-                        "vmla.f32   q12, %q13, d5[0]    \n"
-                        "vmla.f32   q13, %q13, d5[1]    \n"
-                        "vmla.f32   q14, %q13, d3[0]    \n"
-                        "vmla.f32   q15, %q13, d3[1]    \n"
-
-                        "vmla.f32   q12, %q14, d0[0]    \n"
-                        "vmla.f32   q13, %q14, d0[1]    \n"
-                        "vmla.f32   q14, %q14, d1[0]    \n"
-                        "vmla.f32   q15, %q14, d1[1]    \n"
-
-                        "vld1.f32   {d2}, [%3]          \n"
-
-                        "vmla.f32   q12, %q15, d0[1]    \n"
-                        "vmla.f32   q13, %q15, d1[0]    \n"
-                        "vmla.f32   q14, %q15, d1[1]    \n"
-                        "vmla.f32   q15, %q15, d2[0]    \n"
-
-                        "vmla.f32   q12, %q16, d1[0]    \n"
-                        "vmla.f32   q13, %q16, d1[1]    \n"
-                        "vmla.f32   q14, %q16, d2[0]    \n"
-                        "vmla.f32   q15, %q16, d2[1]    \n"
-
-                        "vstm       %0!, {d24-d31}      \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "q0", "q1", "q2", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v24.4s, v25.4s}, [%0]      \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.4s}, [%1]               \n"
-
-                        "fmul   v26.4s, %8.4s, v0.s[0]      \n"
-                        "fmul   v27.4s, %8.4s, v0.s[1]      \n"
-                        "fmla   v24.4s, %9.4s, v0.s[1]      \n"
-                        "fmla   v25.4s, %9.4s, v0.s[2]      \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v1.4s}, [%2]               \n"
-
-                        "fmla   v26.4s, %10.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %10.4s, v0.s[3]     \n"
-                        "fmla   v24.4s, %11.4s, v1.s[0]     \n"
-                        "fmla   v25.4s, %11.4s, v1.s[1]     \n"
-
-                        "add    %1, %1, #8                  \n"
-
-                        "fmla   v26.4s, %12.4s, v1.s[1]     \n"
-                        "fmla   v27.4s, %12.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.4s}, [%3]               \n"
-
-                        "fmla   v24.4s, %13.4s, v1.s[2]     \n"
-                        "fmla   v25.4s, %13.4s, v1.s[3]     \n"
-                        "fmla   v26.4s, %14.4s, v0.s[0]     \n"
-                        "fmla   v27.4s, %14.4s, v0.s[1]     \n"
-
-                        "add    %2, %2, #8                  \n"
-
-                        "fmla   v24.4s, %15.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %15.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %16.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %16.4s, v0.s[3]     \n"
-
-                        "add    %3, %3, #8                  \n"
-
-                        "fadd   v24.4s, v24.4s, v26.4s      \n"
-                        "fadd   v25.4s, v25.4s, v27.4s      \n"
-
-                        "st1    {v24.4s, v25.4s}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "v0", "v1", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #256]          \n"
-                        "vld1.f32   {d24-d27}, [%0 :128] \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%1]       \n"
-
-                        "vmul.f32   q14, %q8, d0[0]     \n"
-                        "vmul.f32   q15, %q8, d0[1]     \n"
-                        "vmla.f32   q12, %q9, d0[1]     \n"
-                        "vmla.f32   q13, %q9, d1[0]     \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d2-d3}, [%2]       \n"
-
-                        "vmla.f32   q14, %q10, d1[0]    \n"
-                        "vmla.f32   q15, %q10, d1[1]    \n"
-                        "vmla.f32   q12, %q11, d2[0]    \n"
-                        "vmla.f32   q13, %q11, d2[1]    \n"
-
-                        "add        %1, %1, #8          \n"
-
-                        "vmla.f32   q14, %q12, d2[1]    \n"
-                        "vmla.f32   q15, %q12, d3[0]    \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%3]       \n"
-
-                        "vmla.f32   q12, %q13, d3[0]    \n"
-                        "vmla.f32   q13, %q13, d3[1]    \n"
-                        "vmla.f32   q14, %q14, d0[0]    \n"
-                        "vmla.f32   q15, %q14, d0[1]    \n"
-
-                        "add        %2, %2, #8          \n"
-
-                        "vmla.f32   q12, %q15, d0[1]    \n"
-                        "vmla.f32   q13, %q15, d1[0]    \n"
-                        "vmla.f32   q14, %q16, d1[0]    \n"
-                        "vmla.f32   q15, %q16, d1[1]    \n"
-
-                        "add        %3, %3, #8          \n"
-
-                        "vadd.f32   q12, q12, q14       \n"
-                        "vadd.f32   q13, q13, q15       \n"
-
-                        "vst1.f32   {d24-d27}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "q0", "q1", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-                    float32x4_t _sum0 = vld1q_f32(outptr0);
-
-                    float32x4_t _r0 = vld1q_f32(r0);
-                    float32x4_t _r1 = vld1q_f32(r1);
-                    float32x4_t _r2 = vld1q_f32(r2);
-
-#if __aarch64__
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k00, _r0, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k01, _r0, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k02, _r0, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k10, _r1, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k11, _r1, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k12, _r1, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k20, _r2, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k21, _r2, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k22, _r2, 2);
-#else
-                    _sum0 = vmlaq_lane_f32(_sum0, _k00, vget_low_f32(_r0), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k01, vget_low_f32(_r0), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k02, vget_high_f32(_r0), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k10, vget_low_f32(_r1), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k11, vget_low_f32(_r1), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k12, vget_high_f32(_r1), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k20, vget_low_f32(_r2), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k21, vget_low_f32(_r2), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k22, vget_high_f32(_r2), 0);
-#endif
-
-                    vst1q_f32(outptr0, _sum0);
-
-                    r0 += 1;
-                    r1 += 1;
-                    r2 += 1;
-                    outptr0 += 4;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * 4;
-        }
-    }
-}
-
-static void conv3x3s2_pack1to4_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* bias = _bias;
-
-    int remain_outch_start = 0;
-
-#if __ARM_NEON && __aarch64__
-    int nn_outch = 0;
-    nn_outch = outch >> 1;
-    remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        Mat out0 = top_blob.channel(p);
-        Mat out1 = top_blob.channel(p + 1);
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        float32x4_t _bias1 = bias ? vld1q_f32((const float*)bias + (p + 1) * 4) : vdupq_n_f32(0.f);
-        out0.fill(_bias0);
-        out1.fill(_bias1);
-
-        const float* k0 = kernel.channel(p);
-        const float* k1 = kernel.channel(p + 1);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            float32x4_t _k00_0 = vld1q_f32(k0);
-            float32x4_t _k01_0 = vld1q_f32(k0 + 4);
-            float32x4_t _k02_0 = vld1q_f32(k0 + 8);
-            float32x4_t _k10_0 = vld1q_f32(k0 + 12);
-            float32x4_t _k11_0 = vld1q_f32(k0 + 16);
-            float32x4_t _k12_0 = vld1q_f32(k0 + 20);
-            float32x4_t _k20_0 = vld1q_f32(k0 + 24);
-            float32x4_t _k21_0 = vld1q_f32(k0 + 28);
-            float32x4_t _k22_0 = vld1q_f32(k0 + 32);
-
-            float32x4_t _k00_1 = vld1q_f32(k1);
-            float32x4_t _k01_1 = vld1q_f32(k1 + 4);
-            float32x4_t _k02_1 = vld1q_f32(k1 + 8);
-            float32x4_t _k10_1 = vld1q_f32(k1 + 12);
-            float32x4_t _k11_1 = vld1q_f32(k1 + 16);
-            float32x4_t _k12_1 = vld1q_f32(k1 + 20);
-            float32x4_t _k20_1 = vld1q_f32(k1 + 24);
-            float32x4_t _k21_1 = vld1q_f32(k1 + 28);
-            float32x4_t _k22_1 = vld1q_f32(k1 + 32);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int nn = outw >> 2;
-                int remain = outw & 3;
-
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "0:                                 \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v6.4s, v7.4s, v8.4s, v9.4s}, [%1] \n" // sum0
-
-                        // r0
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%3], #32   \n"
-                        "ld1r   {v4.4s}, [%3]               \n"
-
-                        "fmla   v6.4s, %12.4s, v0.s[0]      \n"
-                        "fmla   v7.4s, %12.4s, v0.s[2]      \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%2] \n" // sum1
-
-                        "fmla   v8.4s, %12.4s, v1.s[0]      \n"
-                        "fmla   v9.4s, %12.4s, v1.s[2]      \n"
-
-                        "fmla   v10.4s, %21.4s, v0.s[0]     \n"
-                        "fmla   v11.4s, %21.4s, v0.s[2]     \n"
-                        "fmla   v12.4s, %21.4s, v1.s[0]     \n"
-                        "fmla   v13.4s, %21.4s, v1.s[2]     \n"
-
-                        "fmla   v6.4s, %13.4s, v0.s[1]      \n"
-                        "fmla   v7.4s, %13.4s, v0.s[3]      \n"
-                        "fmla   v8.4s, %13.4s, v1.s[1]      \n"
-                        "fmla   v9.4s, %13.4s, v1.s[3]      \n"
-                        "fmla   v10.4s, %22.4s, v0.s[1]     \n"
-                        "fmla   v11.4s, %22.4s, v0.s[3]     \n"
-                        "fmla   v12.4s, %22.4s, v1.s[1]     \n"
-                        "fmla   v13.4s, %22.4s, v1.s[3]     \n"
-
-                        // r1
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v2.4s, v3.4s}, [%4], #32   \n"
-                        "ld1r   {v5.4s}, [%4]               \n"
-
-                        "fmla   v6.4s, %14.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %14.4s, v1.s[0]      \n"
-                        "fmla   v8.4s, %14.4s, v1.s[2]      \n"
-                        "fmla   v9.4s, %14.4s, v4.s[0]      \n"
-                        "fmla   v10.4s, %23.4s, v0.s[2]     \n"
-                        "fmla   v11.4s, %23.4s, v1.s[0]     \n"
-                        "fmla   v12.4s, %23.4s, v1.s[2]     \n"
-                        "fmla   v13.4s, %23.4s, v4.s[0]     \n"
-
-                        "fmla   v6.4s, %15.4s, v2.s[0]      \n"
-                        "fmla   v7.4s, %15.4s, v2.s[2]      \n"
-                        "fmla   v8.4s, %15.4s, v3.s[0]      \n"
-                        "fmla   v9.4s, %15.4s, v3.s[2]      \n"
-                        "fmla   v10.4s, %24.4s, v2.s[0]     \n"
-                        "fmla   v11.4s, %24.4s, v2.s[2]     \n"
-                        "fmla   v12.4s, %24.4s, v3.s[0]     \n"
-                        "fmla   v13.4s, %24.4s, v3.s[2]     \n"
-
-                        "fmla   v6.4s, %16.4s, v2.s[1]      \n"
-                        "fmla   v7.4s, %16.4s, v2.s[3]      \n"
-                        "fmla   v8.4s, %16.4s, v3.s[1]      \n"
-                        "fmla   v9.4s, %16.4s, v3.s[3]      \n"
-                        "fmla   v10.4s, %25.4s, v2.s[1]     \n"
-                        "fmla   v11.4s, %25.4s, v2.s[3]     \n"
-                        "fmla   v12.4s, %25.4s, v3.s[1]     \n"
-                        "fmla   v13.4s, %25.4s, v3.s[3]     \n"
-
-                        // r2
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%5], #32   \n"
-                        "ld1r   {v4.4s}, [%5]               \n"
-
-                        "fmla   v6.4s, %17.4s, v2.s[2]      \n"
-                        "fmla   v7.4s, %17.4s, v3.s[0]      \n"
-                        "fmla   v8.4s, %17.4s, v3.s[2]      \n"
-                        "fmla   v9.4s, %17.4s, v5.s[0]      \n"
-                        "fmla   v10.4s, %26.4s, v2.s[2]     \n"
-                        "fmla   v11.4s, %26.4s, v3.s[0]     \n"
-                        "fmla   v12.4s, %26.4s, v3.s[2]     \n"
-                        "fmla   v13.4s, %26.4s, v5.s[0]     \n"
-
-                        "fmla   v6.4s, %18.4s, v0.s[0]      \n"
-                        "fmla   v7.4s, %18.4s, v0.s[2]      \n"
-                        "fmla   v8.4s, %18.4s, v1.s[0]      \n"
-                        "fmla   v9.4s, %18.4s, v1.s[2]      \n"
-                        "fmla   v10.4s, %27.4s, v0.s[0]     \n"
-                        "fmla   v11.4s, %27.4s, v0.s[2]     \n"
-                        "fmla   v12.4s, %27.4s, v1.s[0]     \n"
-                        "fmla   v13.4s, %27.4s, v1.s[2]     \n"
-
-                        "fmla   v6.4s, %19.4s, v0.s[1]      \n"
-                        "fmla   v7.4s, %19.4s, v0.s[3]      \n"
-                        "fmla   v8.4s, %19.4s, v1.s[1]      \n"
-                        "fmla   v9.4s, %19.4s, v1.s[3]      \n"
-                        "fmla   v10.4s, %28.4s, v0.s[1]     \n"
-                        "fmla   v11.4s, %28.4s, v0.s[3]     \n"
-                        "fmla   v12.4s, %28.4s, v1.s[1]     \n"
-                        "fmla   v13.4s, %28.4s, v1.s[3]     \n"
-
-                        "fmla   v6.4s, %20.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %20.4s, v1.s[0]      \n"
-                        "fmla   v8.4s, %20.4s, v1.s[2]      \n"
-                        "fmla   v9.4s, %20.4s, v4.s[0]      \n"
-                        "fmla   v10.4s, %29.4s, v0.s[2]     \n"
-                        "fmla   v11.4s, %29.4s, v1.s[0]     \n"
-                        "fmla   v12.4s, %29.4s, v1.s[2]     \n"
-                        "fmla   v13.4s, %29.4s, v4.s[0]     \n"
-
-                        "subs   %w0, %w0, #1                \n"
-
-                        "st1    {v6.4s, v7.4s, v8.4s, v9.4s}, [%1], #64 \n"
-                        "st1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%2], #64 \n"
-
-                        "bne    0b                          \n"
-
-                        : "=r"(nn),      // %0
-                        "=r"(outptr0), // %1
-                        "=r"(outptr1), // %2
-                        "=r"(r0),      // %3
-                        "=r"(r1),      // %4
-                        "=r"(r2)       // %5
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(outptr1),
-                        "3"(r0),
-                        "4"(r1),
-                        "5"(r2),
-                        "w"(_k00_0), // %12
-                        "w"(_k01_0), // %13
-                        "w"(_k02_0), // %14
-                        "w"(_k10_0), // %15
-                        "w"(_k11_0), // %16
-                        "w"(_k12_0), // %17
-                        "w"(_k20_0), // %18
-                        "w"(_k21_0), // %19
-                        "w"(_k22_0), // %20
-                        "w"(_k00_1), // %21
-                        "w"(_k01_1), // %22
-                        "w"(_k02_1), // %23
-                        "w"(_k10_1), // %24
-                        "w"(_k11_1), // %25
-                        "w"(_k12_1), // %26
-                        "w"(_k20_1), // %27
-                        "w"(_k21_1), // %28
-                        "w"(_k22_1)  // %29
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13");
-                }
-
-                for (; remain > 0; remain--)
-                {
-                    float32x4_t _sum0 = vld1q_f32(outptr0);
-                    float32x4_t _sum1 = vld1q_f32(outptr1);
-
-                    float32x4_t _r0 = vld1q_f32(r0);
-                    float32x4_t _r1 = vld1q_f32(r1);
-                    float32x4_t _r2 = vld1q_f32(r2);
-
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k00_0, _r0, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k01_0, _r0, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k02_0, _r0, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k10_0, _r1, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k11_0, _r1, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k12_0, _r1, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k20_0, _r2, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k21_0, _r2, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k22_0, _r2, 2);
-
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k00_1, _r0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k01_1, _r0, 1);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k02_1, _r0, 2);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k10_1, _r1, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k11_1, _r1, 1);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k12_1, _r1, 2);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k20_1, _r2, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k21_1, _r2, 1);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k22_1, _r2, 2);
-
-                    vst1q_f32(outptr0, _sum0);
-                    vst1q_f32(outptr1, _sum1);
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0 += 4;
-                    outptr1 += 4;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * 4;
-            k1 += 9 * 4;
-        }
-    }
-#endif // __ARM_NEON && __aarch64__
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        out0.fill(_bias0);
-
-        const float* k0 = kernel.channel(p);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            float32x4_t _k00 = vld1q_f32(k0);
-            float32x4_t _k01 = vld1q_f32(k0 + 4);
-            float32x4_t _k02 = vld1q_f32(k0 + 8);
-            float32x4_t _k10 = vld1q_f32(k0 + 12);
-            float32x4_t _k11 = vld1q_f32(k0 + 16);
-            float32x4_t _k12 = vld1q_f32(k0 + 20);
-            float32x4_t _k20 = vld1q_f32(k0 + 24);
-            float32x4_t _k21 = vld1q_f32(k0 + 28);
-            float32x4_t _k22 = vld1q_f32(k0 + 32);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int nn = outw >> 2;
-                int remain = outw & 3;
-
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "0:                                 \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v6.4s, v7.4s, v8.4s, v9.4s}, [%1] \n" // sum0
-
-                        // r0
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%2], #32   \n"
-                        "ld1r   {v4.4s}, [%2]               \n"
-
-                        "fmla   v6.4s, %10.4s, v0.s[0]      \n"
-                        "fmla   v7.4s, %10.4s, v0.s[2]      \n"
-                        "fmla   v8.4s, %10.4s, v1.s[0]      \n"
-                        "fmla   v9.4s, %10.4s, v1.s[2]      \n"
-
-                        "fmla   v6.4s, %11.4s, v0.s[1]      \n"
-                        "fmla   v7.4s, %11.4s, v0.s[3]      \n"
-                        "fmla   v8.4s, %11.4s, v1.s[1]      \n"
-                        "fmla   v9.4s, %11.4s, v1.s[3]      \n"
-
-                        // r1
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v2.4s, v3.4s}, [%3], #32   \n"
-                        "ld1r   {v5.4s}, [%3]               \n"
-
-                        "fmla   v6.4s, %12.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %12.4s, v1.s[0]      \n"
-                        "fmla   v8.4s, %12.4s, v1.s[2]      \n"
-                        "fmla   v9.4s, %12.4s, v4.s[0]      \n"
-
-                        "fmla   v6.4s, %13.4s, v2.s[0]      \n"
-                        "fmla   v7.4s, %13.4s, v2.s[2]      \n"
-                        "fmla   v8.4s, %13.4s, v3.s[0]      \n"
-                        "fmla   v9.4s, %13.4s, v3.s[2]      \n"
-
-                        "fmla   v6.4s, %14.4s, v2.s[1]      \n"
-                        "fmla   v7.4s, %14.4s, v2.s[3]      \n"
-                        "fmla   v8.4s, %14.4s, v3.s[1]      \n"
-                        "fmla   v9.4s, %14.4s, v3.s[3]      \n"
-
-                        // r2
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%4], #32   \n"
-                        "ld1r   {v4.4s}, [%4]               \n"
-
-                        "fmla   v6.4s, %15.4s, v2.s[2]      \n"
-                        "fmla   v7.4s, %15.4s, v3.s[0]      \n"
-                        "fmla   v8.4s, %15.4s, v3.s[2]      \n"
-                        "fmla   v9.4s, %15.4s, v5.s[0]      \n"
-
-                        "fmla   v6.4s, %16.4s, v0.s[0]      \n"
-                        "fmla   v7.4s, %16.4s, v0.s[2]      \n"
-                        "fmla   v8.4s, %16.4s, v1.s[0]      \n"
-                        "fmla   v9.4s, %16.4s, v1.s[2]      \n"
-
-                        "fmla   v6.4s, %17.4s, v0.s[1]      \n"
-                        "fmla   v7.4s, %17.4s, v0.s[3]      \n"
-                        "fmla   v8.4s, %17.4s, v1.s[1]      \n"
-                        "fmla   v9.4s, %17.4s, v1.s[3]      \n"
-
-                        "fmla   v6.4s, %18.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %18.4s, v1.s[0]      \n"
-                        "fmla   v8.4s, %18.4s, v1.s[2]      \n"
-                        "fmla   v9.4s, %18.4s, v4.s[0]      \n"
-
-                        "subs   %w0, %w0, #1                \n"
-
-                        "st1    {v6.4s, v7.4s, v8.4s, v9.4s}, [%1], #64 \n"
-
-                        "bne    0b                          \n"
-
-                        : "=r"(nn),      // %0
-                        "=r"(outptr0), // %1
-                        "=r"(r0),      // %2
-                        "=r"(r1),      // %3
-                        "=r"(r2)       // %4
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00), // %10
-                        "w"(_k01), // %11
-                        "w"(_k02), // %12
-                        "w"(_k10), // %13
-                        "w"(_k11), // %14
-                        "w"(_k12), // %15
-                        "w"(_k20), // %16
-                        "w"(_k21), // %17
-                        "w"(_k22)  // %18
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9");
-                }
-#else  // __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "0:                             \n"
-
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1, {d0-d7}         \n" // sum0
-
-                        // r0
-                        "pld        [%2, #256]          \n"
-                        "vld1.f32   {d8-d11}, [%2]!     \n"
-                        "vld1.f32   {d12[]}, [%2]       \n"
-
-                        "vmla.f32   q0, %q10, d8[0]     \n"
-                        "vmla.f32   q1, %q10, d9[0]     \n"
-                        "vmla.f32   q2, %q10, d10[0]    \n"
-                        "vmla.f32   q3, %q10, d11[0]    \n"
-
-                        "vmla.f32   q0, %q11, d8[1]     \n"
-                        "vmla.f32   q1, %q11, d9[1]     \n"
-                        "vmla.f32   q2, %q11, d10[1]    \n"
-                        "vmla.f32   q3, %q11, d11[1]    \n"
-
-                        "vmla.f32   q0, %q12, d9[0]     \n"
-                        "vmla.f32   q1, %q12, d10[0]    \n"
-                        "vmla.f32   q2, %q12, d11[0]    \n"
-
-                        // r1
-                        "pld        [%3, #256]          \n"
-                        "vld1.f32   {d8-d11}, [%3]!     \n"
-                        "vld1.f32   {d13[]}, [%3]       \n"
-
-                        "vmla.f32   q3, %q12, d12[0]    \n"
-
-                        "vmla.f32   q0, %q13, d8[0]     \n"
-                        "vmla.f32   q1, %q13, d9[0]     \n"
-                        "vmla.f32   q2, %q13, d10[0]    \n"
-                        "vmla.f32   q3, %q13, d11[0]    \n"
-
-                        "vmla.f32   q0, %q14, d8[1]     \n"
-                        "vmla.f32   q1, %q14, d9[1]     \n"
-                        "vmla.f32   q2, %q14, d10[1]    \n"
-                        "vmla.f32   q3, %q14, d11[1]    \n"
-
-                        "vmla.f32   q0, %q15, d9[0]     \n"
-                        "vmla.f32   q1, %q15, d10[0]    \n"
-                        "vmla.f32   q2, %q15, d11[0]    \n"
-
-                        // r2
-                        "pld        [%4, #256]          \n"
-                        "vld1.f32   {d8-d11}, [%4]!     \n"
-                        "vld1.f32   {d12[]}, [%4]       \n"
-
-                        "vmla.f32   q3, %q15, d13[0]    \n"
-
-                        "vmla.f32   q0, %q16, d8[0]     \n"
-                        "vmla.f32   q1, %q16, d9[0]     \n"
-                        "vmla.f32   q2, %q16, d10[0]    \n"
-                        "vmla.f32   q3, %q16, d11[0]    \n"
-
-                        "vmla.f32   q0, %q17, d8[1]     \n"
-                        "vmla.f32   q1, %q17, d9[1]     \n"
-                        "vmla.f32   q2, %q17, d10[1]    \n"
-                        "vmla.f32   q3, %q17, d11[1]    \n"
-
-                        "vmla.f32   q0, %q18, d9[0]     \n"
-                        "vmla.f32   q1, %q18, d10[0]    \n"
-                        "vmla.f32   q2, %q18, d11[0]    \n"
-                        "vmla.f32   q3, %q18, d12[0]    \n"
-
-                        "subs       %0, %0, #1          \n"
-
-                        "vstm       %1!, {d0-d7}        \n"
-
-                        "bne        0b                  \n"
-
-                        : "=r"(nn),      // %0
-                        "=r"(outptr0), // %1
-                        "=r"(r0),      // %2
-                        "=r"(r1),      // %3
-                        "=r"(r2)       // %4
-                        : "0"(nn),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00), // %10
-                        "w"(_k01), // %11
-                        "w"(_k02), // %12
-                        "w"(_k10), // %13
-                        "w"(_k11), // %14
-                        "w"(_k12), // %15
-                        "w"(_k20), // %16
-                        "w"(_k21), // %17
-                        "w"(_k22)  // %18
-                        : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6");
-                }
-#endif // __aarch64__
-
-                for (; remain > 0; remain--)
-                {
-                    float32x4_t _sum0 = vld1q_f32(outptr0);
-
-                    float32x4_t _r0 = vld1q_f32(r0);
-                    float32x4_t _r1 = vld1q_f32(r1);
-                    float32x4_t _r2 = vld1q_f32(r2);
-
-#if __aarch64__
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k00, _r0, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k01, _r0, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k02, _r0, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k10, _r1, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k11, _r1, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k12, _r1, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k20, _r2, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k21, _r2, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k22, _r2, 2);
-#else
-                    _sum0 = vmlaq_lane_f32(_sum0, _k00, vget_low_f32(_r0), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k01, vget_low_f32(_r0), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k02, vget_high_f32(_r0), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k10, vget_low_f32(_r1), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k11, vget_low_f32(_r1), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k12, vget_high_f32(_r1), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k20, vget_low_f32(_r2), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k21, vget_low_f32(_r2), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k22, vget_high_f32(_r2), 0);
-#endif
-
-                    vst1q_f32(outptr0, _sum0);
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0 += 4;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to4_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to4_bf16s.h
deleted file mode 100644
index cf599775..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to4_bf16s.h
+++ /dev/null
@@ -1,3732 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack1to4_bf16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-#if __ARM_NEON && __aarch64__
-    Mat top_blob_fp32(outw, outh, opt.num_threads, (size_t)4u * 4 * 2, 4 * 2, opt.workspace_allocator);
-#else
-    Mat top_blob_fp32(outw, outh, opt.num_threads, (size_t)4u * 4, 4, opt.workspace_allocator);
-#endif
-
-    const float* bias = _bias;
-
-    int remain_outch_start = 0;
-
-#if __ARM_NEON && __aarch64__
-    int nn_outch = 0;
-    nn_outch = outch >> 1;
-    remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        Mat out0 = top_blob_fp32.channel(get_omp_thread_num());
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        float32x4_t _bias1 = bias ? vld1q_f32((const float*)bias + (p + 1) * 4) : vdupq_n_f32(0.f);
-        {
-            float* ptr = (float*)out0;
-
-            for (int i = 0; i < outh; i++)
-            {
-                int j = 0;
-
-                for (; j + 3 < outw; j += 4)
-                {
-                    vst1q_f32(ptr, _bias0);
-                    vst1q_f32(ptr + 4, _bias0);
-                    vst1q_f32(ptr + 8, _bias0);
-                    vst1q_f32(ptr + 12, _bias0);
-                    vst1q_f32(ptr + 16, _bias1);
-                    vst1q_f32(ptr + 20, _bias1);
-                    vst1q_f32(ptr + 24, _bias1);
-                    vst1q_f32(ptr + 28, _bias1);
-                    ptr += 32;
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    vst1q_f32(ptr, _bias0);
-                    vst1q_f32(ptr + 4, _bias0);
-                    vst1q_f32(ptr + 8, _bias1);
-                    vst1q_f32(ptr + 12, _bias1);
-                    ptr += 16;
-                }
-                for (; j < outw; j++)
-                {
-                    vst1q_f32(ptr, _bias0);
-                    vst1q_f32(ptr + 4, _bias1);
-                    ptr += 8;
-                }
-            }
-        }
-
-        const unsigned short* k0 = kernel.channel(p);
-        const unsigned short* k1 = kernel.channel(p + 1);
-
-        int q = 0;
-        for (; q < inch - 1; q++)
-        {
-            float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<const unsigned short>(0);
-            const unsigned short* r1 = img0.row<const unsigned short>(1);
-            const unsigned short* r2 = img0.row<const unsigned short>(2);
-
-            float32x4_t _k00_0 = bfloat2float(vld1_u16(k0));
-            float32x4_t _k01_0 = bfloat2float(vld1_u16(k0 + 4));
-            float32x4_t _k02_0 = bfloat2float(vld1_u16(k0 + 8));
-            float32x4_t _k10_0 = bfloat2float(vld1_u16(k0 + 12));
-            float32x4_t _k11_0 = bfloat2float(vld1_u16(k0 + 16));
-            float32x4_t _k12_0 = bfloat2float(vld1_u16(k0 + 20));
-            float32x4_t _k20_0 = bfloat2float(vld1_u16(k0 + 24));
-            float32x4_t _k21_0 = bfloat2float(vld1_u16(k0 + 28));
-            float32x4_t _k22_0 = bfloat2float(vld1_u16(k0 + 32));
-
-            float32x4_t _k00_1 = bfloat2float(vld1_u16(k1));
-            float32x4_t _k01_1 = bfloat2float(vld1_u16(k1 + 4));
-            float32x4_t _k02_1 = bfloat2float(vld1_u16(k1 + 8));
-            float32x4_t _k10_1 = bfloat2float(vld1_u16(k1 + 12));
-            float32x4_t _k11_1 = bfloat2float(vld1_u16(k1 + 16));
-            float32x4_t _k12_1 = bfloat2float(vld1_u16(k1 + 20));
-            float32x4_t _k20_1 = bfloat2float(vld1_u16(k1 + 24));
-            float32x4_t _k21_1 = bfloat2float(vld1_u16(k1 + 28));
-            float32x4_t _k22_1 = bfloat2float(vld1_u16(k1 + 32));
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-
-                for (; j + 3 < outw; j += 4)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1], #8           \n"
-                        "ld1    {v1.s}[0], [%1]             \n"
-
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0] \n"
-
-                        "fmla   v24.4s, %8.4s, v0.s[0]      \n"
-                        "fmla   v25.4s, %8.4s, v0.s[1]      \n"
-                        "fmla   v26.4s, %8.4s, v0.s[2]      \n"
-                        "fmla   v27.4s, %8.4s, v0.s[3]      \n"
-                        "fmla   v28.4s, %17.4s, v0.s[0]     \n"
-                        "fmla   v29.4s, %17.4s, v0.s[1]     \n"
-                        "fmla   v30.4s, %17.4s, v0.s[2]     \n"
-                        "fmla   v31.4s, %17.4s, v0.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v2.4h}, [%2], #8           \n"
-                        "ld1    {v3.s}[0], [%2]             \n"
-
-                        "fmla   v24.4s, %9.4s, v0.s[1]      \n"
-                        "fmla   v25.4s, %9.4s, v0.s[2]      \n"
-                        "fmla   v26.4s, %9.4s, v0.s[3]      \n"
-                        "fmla   v27.4s, %9.4s, v1.s[0]      \n"
-                        "fmla   v28.4s, %18.4s, v0.s[1]     \n"
-                        "fmla   v29.4s, %18.4s, v0.s[2]     \n"
-                        "fmla   v30.4s, %18.4s, v0.s[3]     \n"
-                        "fmla   v31.4s, %18.4s, v1.s[0]     \n"
-
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v24.4s, %10.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %10.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %10.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %10.4s, v1.s[1]     \n"
-                        "fmla   v28.4s, %19.4s, v0.s[2]     \n"
-                        "fmla   v29.4s, %19.4s, v0.s[3]     \n"
-                        "fmla   v30.4s, %19.4s, v1.s[0]     \n"
-                        "fmla   v31.4s, %19.4s, v1.s[1]     \n"
-
-                        "fmla   v24.4s, %11.4s, v2.s[0]     \n"
-                        "fmla   v25.4s, %11.4s, v2.s[1]     \n"
-                        "fmla   v26.4s, %11.4s, v2.s[2]     \n"
-                        "fmla   v27.4s, %11.4s, v2.s[3]     \n"
-                        "fmla   v28.4s, %20.4s, v2.s[0]     \n"
-                        "fmla   v29.4s, %20.4s, v2.s[1]     \n"
-                        "fmla   v30.4s, %20.4s, v2.s[2]     \n"
-                        "fmla   v31.4s, %20.4s, v2.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v0.4h}, [%3], #8           \n"
-                        "ld1    {v1.s}[0], [%3]             \n"
-
-                        "fmla   v24.4s, %12.4s, v2.s[1]     \n"
-                        "fmla   v25.4s, %12.4s, v2.s[2]     \n"
-                        "fmla   v26.4s, %12.4s, v2.s[3]     \n"
-                        "fmla   v27.4s, %12.4s, v3.s[0]     \n"
-                        "fmla   v28.4s, %21.4s, v2.s[1]     \n"
-                        "fmla   v29.4s, %21.4s, v2.s[2]     \n"
-                        "fmla   v30.4s, %21.4s, v2.s[3]     \n"
-                        "fmla   v31.4s, %21.4s, v3.s[0]     \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v24.4s, %13.4s, v2.s[2]     \n"
-                        "fmla   v25.4s, %13.4s, v2.s[3]     \n"
-                        "fmla   v26.4s, %13.4s, v3.s[0]     \n"
-                        "fmla   v27.4s, %13.4s, v3.s[1]     \n"
-                        "fmla   v28.4s, %22.4s, v2.s[2]     \n"
-                        "fmla   v29.4s, %22.4s, v2.s[3]     \n"
-                        "fmla   v30.4s, %22.4s, v3.s[0]     \n"
-                        "fmla   v31.4s, %22.4s, v3.s[1]     \n"
-
-                        "fmla   v24.4s, %14.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %14.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %14.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %14.4s, v0.s[3]     \n"
-                        "fmla   v28.4s, %23.4s, v0.s[0]     \n"
-                        "fmla   v29.4s, %23.4s, v0.s[1]     \n"
-                        "fmla   v30.4s, %23.4s, v0.s[2]     \n"
-                        "fmla   v31.4s, %23.4s, v0.s[3]     \n"
-
-                        "fmla   v24.4s, %15.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %15.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %15.4s, v0.s[3]     \n"
-                        "fmla   v27.4s, %15.4s, v1.s[0]     \n"
-                        "fmla   v28.4s, %24.4s, v0.s[1]     \n"
-                        "fmla   v29.4s, %24.4s, v0.s[2]     \n"
-                        "fmla   v30.4s, %24.4s, v0.s[3]     \n"
-                        "fmla   v31.4s, %24.4s, v1.s[0]     \n"
-
-                        "sub    %0, %0, #64                 \n"
-
-                        "fmla   v24.4s, %16.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %16.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %16.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %16.4s, v1.s[1]     \n"
-                        "fmla   v28.4s, %25.4s, v0.s[2]     \n"
-                        "fmla   v29.4s, %25.4s, v0.s[3]     \n"
-                        "fmla   v30.4s, %25.4s, v1.s[0]     \n"
-                        "fmla   v31.4s, %25.4s, v1.s[1]     \n"
-
-                        "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00_0), // %8
-                        "w"(_k01_0), // %9
-                        "w"(_k02_0), // %10
-                        "w"(_k10_0), // %11
-                        "w"(_k11_0), // %12
-                        "w"(_k12_0), // %13
-                        "w"(_k20_0), // %14
-                        "w"(_k21_0), // %15
-                        "w"(_k22_0), // %16
-                        "w"(_k00_1), // %17
-                        "w"(_k01_1), // %18
-                        "w"(_k02_1), // %19
-                        "w"(_k10_1), // %20
-                        "w"(_k11_1), // %21
-                        "w"(_k12_1), // %22
-                        "w"(_k20_1), // %23
-                        "w"(_k21_1), // %24
-                        "w"(_k22_1)  // %25
-                        : "memory", "v0", "v1", "v2", "v3", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1]               \n"
-
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0] \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v24.4s, %8.4s, v0.s[0]      \n"
-                        "fmla   v25.4s, %8.4s, v0.s[1]      \n"
-                        "fmla   v26.4s, %17.4s, v0.s[0]     \n"
-                        "fmla   v27.4s, %17.4s, v0.s[1]     \n"
-
-                        "prfm   pldl1keep, [%2, #64]       \n"
-                        "ld1    {v1.4h}, [%2]               \n"
-
-                        "fmla   v24.4s, %9.4s, v0.s[1]      \n"
-                        "fmla   v25.4s, %9.4s, v0.s[2]      \n"
-                        "fmla   v26.4s, %18.4s, v0.s[1]     \n"
-                        "fmla   v27.4s, %18.4s, v0.s[2]     \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v24.4s, %10.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %10.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %19.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %19.4s, v0.s[3]     \n"
-
-                        "fmla   v24.4s, %11.4s, v1.s[0]     \n"
-                        "fmla   v25.4s, %11.4s, v1.s[1]     \n"
-                        "fmla   v26.4s, %20.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %20.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%3, #64]       \n"
-                        "ld1    {v0.4h}, [%3]               \n"
-
-                        "fmla   v24.4s, %12.4s, v1.s[1]     \n"
-                        "fmla   v25.4s, %12.4s, v1.s[2]     \n"
-                        "fmla   v26.4s, %21.4s, v1.s[1]     \n"
-                        "fmla   v27.4s, %21.4s, v1.s[2]     \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v24.4s, %13.4s, v1.s[2]     \n"
-                        "fmla   v25.4s, %13.4s, v1.s[3]     \n"
-                        "fmla   v26.4s, %22.4s, v1.s[2]     \n"
-                        "fmla   v27.4s, %22.4s, v1.s[3]     \n"
-
-                        "fmla   v24.4s, %14.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %14.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %23.4s, v0.s[0]     \n"
-                        "fmla   v27.4s, %23.4s, v0.s[1]     \n"
-
-                        "add    %1, %1, #4                  \n"
-
-                        "fmla   v24.4s, %15.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %15.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %24.4s, v0.s[1]     \n"
-                        "fmla   v27.4s, %24.4s, v0.s[2]     \n"
-
-                        "add    %2, %2, #4                  \n"
-
-                        "fmla   v24.4s, %16.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %16.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %25.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %25.4s, v0.s[3]     \n"
-
-                        "add    %3, %3, #4                  \n"
-
-                        "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00_0), // %8
-                        "w"(_k01_0), // %9
-                        "w"(_k02_0), // %10
-                        "w"(_k10_0), // %11
-                        "w"(_k11_0), // %12
-                        "w"(_k12_0), // %13
-                        "w"(_k20_0), // %14
-                        "w"(_k21_0), // %15
-                        "w"(_k22_0), // %16
-                        "w"(_k00_1), // %17
-                        "w"(_k01_1), // %18
-                        "w"(_k02_1), // %19
-                        "w"(_k10_1), // %20
-                        "w"(_k11_1), // %21
-                        "w"(_k12_1), // %22
-                        "w"(_k20_1), // %23
-                        "w"(_k21_1), // %24
-                        "w"(_k22_1)  // %25
-                        : "memory", "v0", "v1", "v24", "v25", "v26", "v27");
-                }
-                for (; j < outw; j++)
-                {
-                    float32x4_t _sum00 = vld1q_f32(outptr0);
-                    float32x4_t _sum10 = vld1q_f32(outptr0 + 4);
-
-                    float32x4_t _r0 = bfloat2float(vld1_u16(r0));
-                    float32x4_t _r1 = bfloat2float(vld1_u16(r1));
-                    float32x4_t _r2 = bfloat2float(vld1_u16(r2));
-
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k00_0, _r0, 0);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k01_0, _r0, 1);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k02_0, _r0, 2);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k10_0, _r1, 0);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k11_0, _r1, 1);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k12_0, _r1, 2);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k20_0, _r2, 0);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k21_0, _r2, 1);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k22_0, _r2, 2);
-
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k00_1, _r0, 0);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k01_1, _r0, 1);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k02_1, _r0, 2);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k10_1, _r1, 0);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k11_1, _r1, 1);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k12_1, _r1, 2);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k20_1, _r2, 0);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k21_1, _r2, 1);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k22_1, _r2, 2);
-
-                    vst1q_f32(outptr0, _sum00);
-                    vst1q_f32(outptr0 + 4, _sum10);
-
-                    r0 += 1;
-                    r1 += 1;
-                    r2 += 1;
-                    outptr0 += 8;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * 4;
-            k1 += 9 * 4;
-        }
-        for (; q < inch; q++)
-        {
-            unsigned short* outptr0_bf16 = top_blob.channel(p);
-            unsigned short* outptr1_bf16 = top_blob.channel(p + 1);
-
-            const float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<const unsigned short>(0);
-            const unsigned short* r1 = img0.row<const unsigned short>(1);
-            const unsigned short* r2 = img0.row<const unsigned short>(2);
-
-            float32x4_t _k00_0 = bfloat2float(vld1_u16(k0));
-            float32x4_t _k01_0 = bfloat2float(vld1_u16(k0 + 4));
-            float32x4_t _k02_0 = bfloat2float(vld1_u16(k0 + 8));
-            float32x4_t _k10_0 = bfloat2float(vld1_u16(k0 + 12));
-            float32x4_t _k11_0 = bfloat2float(vld1_u16(k0 + 16));
-            float32x4_t _k12_0 = bfloat2float(vld1_u16(k0 + 20));
-            float32x4_t _k20_0 = bfloat2float(vld1_u16(k0 + 24));
-            float32x4_t _k21_0 = bfloat2float(vld1_u16(k0 + 28));
-            float32x4_t _k22_0 = bfloat2float(vld1_u16(k0 + 32));
-
-            float32x4_t _k00_1 = bfloat2float(vld1_u16(k1));
-            float32x4_t _k01_1 = bfloat2float(vld1_u16(k1 + 4));
-            float32x4_t _k02_1 = bfloat2float(vld1_u16(k1 + 8));
-            float32x4_t _k10_1 = bfloat2float(vld1_u16(k1 + 12));
-            float32x4_t _k11_1 = bfloat2float(vld1_u16(k1 + 16));
-            float32x4_t _k12_1 = bfloat2float(vld1_u16(k1 + 20));
-            float32x4_t _k20_1 = bfloat2float(vld1_u16(k1 + 24));
-            float32x4_t _k21_1 = bfloat2float(vld1_u16(k1 + 28));
-            float32x4_t _k22_1 = bfloat2float(vld1_u16(k1 + 32));
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-
-                for (; j + 3 < outw; j += 4)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v0.4h}, [%3], #8           \n"
-                        "ld1    {v1.s}[0], [%3]             \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%2], #64 \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%2], #64 \n"
-
-                        "fmla   v24.4s, %12.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %12.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %12.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %12.4s, v0.s[3]     \n"
-                        "fmla   v28.4s, %21.4s, v0.s[0]     \n"
-                        "fmla   v29.4s, %21.4s, v0.s[1]     \n"
-                        "fmla   v30.4s, %21.4s, v0.s[2]     \n"
-                        "fmla   v31.4s, %21.4s, v0.s[3]     \n"
-
-                        "fmla   v24.4s, %13.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %13.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %13.4s, v0.s[3]     \n"
-                        "fmla   v27.4s, %13.4s, v1.s[0]     \n"
-                        "fmla   v28.4s, %22.4s, v0.s[1]     \n"
-                        "fmla   v29.4s, %22.4s, v0.s[2]     \n"
-                        "fmla   v30.4s, %22.4s, v0.s[3]     \n"
-                        "fmla   v31.4s, %22.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%4, #64]        \n"
-                        "ld1    {v2.4h}, [%4], #8           \n"
-                        "ld1    {v3.s}[0], [%4]             \n"
-
-                        "fmla   v24.4s, %14.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %14.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %14.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %14.4s, v1.s[1]     \n"
-                        "fmla   v28.4s, %23.4s, v0.s[2]     \n"
-                        "fmla   v29.4s, %23.4s, v0.s[3]     \n"
-                        "fmla   v30.4s, %23.4s, v1.s[0]     \n"
-                        "fmla   v31.4s, %23.4s, v1.s[1]     \n"
-
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v24.4s, %15.4s, v2.s[0]     \n"
-                        "fmla   v25.4s, %15.4s, v2.s[1]     \n"
-                        "fmla   v26.4s, %15.4s, v2.s[2]     \n"
-                        "fmla   v27.4s, %15.4s, v2.s[3]     \n"
-                        "fmla   v28.4s, %24.4s, v2.s[0]     \n"
-                        "fmla   v29.4s, %24.4s, v2.s[1]     \n"
-                        "fmla   v30.4s, %24.4s, v2.s[2]     \n"
-                        "fmla   v31.4s, %24.4s, v2.s[3]     \n"
-
-                        "fmla   v24.4s, %16.4s, v2.s[1]     \n"
-                        "fmla   v25.4s, %16.4s, v2.s[2]     \n"
-                        "fmla   v26.4s, %16.4s, v2.s[3]     \n"
-                        "fmla   v27.4s, %16.4s, v3.s[0]     \n"
-                        "fmla   v28.4s, %25.4s, v2.s[1]     \n"
-                        "fmla   v29.4s, %25.4s, v2.s[2]     \n"
-                        "fmla   v30.4s, %25.4s, v2.s[3]     \n"
-                        "fmla   v31.4s, %25.4s, v3.s[0]     \n"
-
-                        "prfm   pldl1keep, [%5, #64]        \n"
-                        "ld1    {v0.4h}, [%5], #8           \n"
-                        "ld1    {v1.s}[0], [%5]             \n"
-
-                        "fmla   v24.4s, %17.4s, v2.s[2]     \n"
-                        "fmla   v25.4s, %17.4s, v2.s[3]     \n"
-                        "fmla   v26.4s, %17.4s, v3.s[0]     \n"
-                        "fmla   v27.4s, %17.4s, v3.s[1]     \n"
-                        "fmla   v28.4s, %26.4s, v2.s[2]     \n"
-                        "fmla   v29.4s, %26.4s, v2.s[3]     \n"
-                        "fmla   v30.4s, %26.4s, v3.s[0]     \n"
-                        "fmla   v31.4s, %26.4s, v3.s[1]     \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v24.4s, %18.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %18.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %18.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %18.4s, v0.s[3]     \n"
-                        "fmla   v28.4s, %27.4s, v0.s[0]     \n"
-                        "fmla   v29.4s, %27.4s, v0.s[1]     \n"
-                        "fmla   v30.4s, %27.4s, v0.s[2]     \n"
-                        "fmla   v31.4s, %27.4s, v0.s[3]     \n"
-
-                        "fmla   v24.4s, %19.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %19.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %19.4s, v0.s[3]     \n"
-                        "fmla   v27.4s, %19.4s, v1.s[0]     \n"
-                        "fmla   v28.4s, %28.4s, v0.s[1]     \n"
-                        "fmla   v29.4s, %28.4s, v0.s[2]     \n"
-                        "fmla   v30.4s, %28.4s, v0.s[3]     \n"
-                        "fmla   v31.4s, %28.4s, v1.s[0]     \n"
-
-                        "fmla   v24.4s, %20.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %20.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %20.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %20.4s, v1.s[1]     \n"
-                        "fmla   v28.4s, %29.4s, v0.s[2]     \n"
-                        "fmla   v29.4s, %29.4s, v0.s[3]     \n"
-                        "fmla   v30.4s, %29.4s, v1.s[0]     \n"
-                        "fmla   v31.4s, %29.4s, v1.s[1]     \n"
-
-                        "shrn   v24.4h, v24.4s, #16         \n"
-                        "shrn   v25.4h, v25.4s, #16         \n"
-                        "shrn   v26.4h, v26.4s, #16         \n"
-                        "shrn   v27.4h, v27.4s, #16         \n"
-                        "shrn   v28.4h, v28.4s, #16         \n"
-                        "shrn   v29.4h, v29.4s, #16         \n"
-                        "shrn   v30.4h, v30.4s, #16         \n"
-                        "shrn   v31.4h, v31.4s, #16         \n"
-
-                        "st1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%0], #32 \n"
-                        "st1    {v28.4h, v29.4h, v30.4h, v31.4h}, [%1], #32 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr1_bf16), // %1
-                        "=r"(outptr0),      // %2
-                        "=r"(r0),           // %3
-                        "=r"(r1),           // %4
-                        "=r"(r2)            // %5
-                        : "0"(outptr0_bf16),
-                        "1"(outptr1_bf16),
-                        "2"(outptr0),
-                        "3"(r0),
-                        "4"(r1),
-                        "5"(r2),
-                        "w"(_k00_0), // %12
-                        "w"(_k01_0), // %13
-                        "w"(_k02_0), // %14
-                        "w"(_k10_0), // %15
-                        "w"(_k11_0), // %16
-                        "w"(_k12_0), // %17
-                        "w"(_k20_0), // %18
-                        "w"(_k21_0), // %19
-                        "w"(_k22_0), // %20
-                        "w"(_k00_1), // %21
-                        "w"(_k01_1), // %22
-                        "w"(_k02_1), // %23
-                        "w"(_k10_1), // %24
-                        "w"(_k11_1), // %25
-                        "w"(_k12_1), // %26
-                        "w"(_k20_1), // %27
-                        "w"(_k21_1), // %28
-                        "w"(_k22_1)  // %29
-                        : "memory", "v0", "v1", "v2", "v3", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v0.4h}, [%3]               \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%2], #64 \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v24.4s, %12.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %12.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %21.4s, v0.s[0]     \n"
-                        "fmla   v27.4s, %21.4s, v0.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #64]        \n"
-                        "ld1    {v1.4h}, [%4]               \n"
-
-                        "fmla   v24.4s, %13.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %13.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %22.4s, v0.s[1]     \n"
-                        "fmla   v27.4s, %22.4s, v0.s[2]     \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v24.4s, %14.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %14.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %23.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %23.4s, v0.s[3]     \n"
-
-                        "fmla   v24.4s, %15.4s, v1.s[0]     \n"
-                        "fmla   v25.4s, %15.4s, v1.s[1]     \n"
-                        "fmla   v26.4s, %24.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %24.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%5, #64]        \n"
-                        "ld1    {v0.4h}, [%5]               \n"
-
-                        "fmla   v24.4s, %16.4s, v1.s[1]     \n"
-                        "fmla   v25.4s, %16.4s, v1.s[2]     \n"
-                        "fmla   v26.4s, %25.4s, v1.s[1]     \n"
-                        "fmla   v27.4s, %25.4s, v1.s[2]     \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v24.4s, %17.4s, v1.s[2]     \n"
-                        "fmla   v25.4s, %17.4s, v1.s[3]     \n"
-                        "fmla   v26.4s, %26.4s, v1.s[2]     \n"
-                        "fmla   v27.4s, %26.4s, v1.s[3]     \n"
-
-                        "fmla   v24.4s, %18.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %18.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %27.4s, v0.s[0]     \n"
-                        "fmla   v27.4s, %27.4s, v0.s[1]     \n"
-
-                        "fmla   v24.4s, %19.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %19.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %28.4s, v0.s[1]     \n"
-                        "fmla   v27.4s, %28.4s, v0.s[2]     \n"
-
-                        "add    %3, %3, #4                  \n"
-
-                        "fmla   v24.4s, %20.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %20.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %29.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %29.4s, v0.s[3]     \n"
-
-                        "add    %4, %4, #4                  \n"
-
-                        "shrn   v24.4h, v24.4s, #16         \n"
-                        "shrn   v25.4h, v25.4s, #16         \n"
-                        "shrn   v26.4h, v26.4s, #16         \n"
-                        "shrn   v27.4h, v27.4s, #16         \n"
-
-                        "add    %5, %5, #4                  \n"
-
-                        "st1    {v24.4h, v25.4h}, [%0], #16 \n"
-                        "st1    {v26.4h, v27.4h}, [%1], #16 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr1_bf16), // %1
-                        "=r"(outptr0),      // %2
-                        "=r"(r0),           // %3
-                        "=r"(r1),           // %4
-                        "=r"(r2)            // %5
-                        : "0"(outptr0_bf16),
-                        "1"(outptr1_bf16),
-                        "2"(outptr0),
-                        "3"(r0),
-                        "4"(r1),
-                        "5"(r2),
-                        "w"(_k00_0), // %12
-                        "w"(_k01_0), // %13
-                        "w"(_k02_0), // %14
-                        "w"(_k10_0), // %15
-                        "w"(_k11_0), // %16
-                        "w"(_k12_0), // %17
-                        "w"(_k20_0), // %18
-                        "w"(_k21_0), // %19
-                        "w"(_k22_0), // %20
-                        "w"(_k00_1), // %21
-                        "w"(_k01_1), // %22
-                        "w"(_k02_1), // %23
-                        "w"(_k10_1), // %24
-                        "w"(_k11_1), // %25
-                        "w"(_k12_1), // %26
-                        "w"(_k20_1), // %27
-                        "w"(_k21_1), // %28
-                        "w"(_k22_1)  // %29
-                        : "memory", "v0", "v1", "v24", "v25", "v26", "v27");
-                }
-                for (; j < outw; j++)
-                {
-                    float32x4_t _sum00 = vld1q_f32(outptr0);
-                    float32x4_t _sum10 = vld1q_f32(outptr0 + 4);
-
-                    float32x4_t _r0 = bfloat2float(vld1_u16(r0));
-                    float32x4_t _r1 = bfloat2float(vld1_u16(r1));
-                    float32x4_t _r2 = bfloat2float(vld1_u16(r2));
-
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k00_0, _r0, 0);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k01_0, _r0, 1);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k02_0, _r0, 2);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k10_0, _r1, 0);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k11_0, _r1, 1);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k12_0, _r1, 2);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k20_0, _r2, 0);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k21_0, _r2, 1);
-                    _sum00 = vfmaq_laneq_f32(_sum00, _k22_0, _r2, 2);
-
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k00_1, _r0, 0);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k01_1, _r0, 1);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k02_1, _r0, 2);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k10_1, _r1, 0);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k11_1, _r1, 1);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k12_1, _r1, 2);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k20_1, _r2, 0);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k21_1, _r2, 1);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _k22_1, _r2, 2);
-
-                    vst1_u16(outptr0_bf16, float2bfloat(_sum00));
-                    vst1_u16(outptr1_bf16, float2bfloat(_sum10));
-
-                    r0 += 1;
-                    r1 += 1;
-                    r2 += 1;
-                    outptr0 += 8;
-                    outptr0_bf16 += 4;
-                    outptr1_bf16 += 4;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * 4;
-            k1 += 9 * 4;
-        }
-    }
-#endif // __ARM_NEON && __aarch64__
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out0 = top_blob_fp32.channel(get_omp_thread_num());
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        out0.fill(_bias0);
-
-        const unsigned short* k0 = kernel.channel(p);
-
-        int q = 0;
-        for (; q < inch - 1; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<unsigned short>(0);
-            const unsigned short* r1 = img0.row<unsigned short>(1);
-            const unsigned short* r2 = img0.row<unsigned short>(2);
-
-            float32x4_t _k00 = bfloat2float(vld1_u16(k0));
-            float32x4_t _k01 = bfloat2float(vld1_u16(k0 + 4));
-            float32x4_t _k02 = bfloat2float(vld1_u16(k0 + 8));
-            float32x4_t _k10 = bfloat2float(vld1_u16(k0 + 12));
-            float32x4_t _k11 = bfloat2float(vld1_u16(k0 + 16));
-            float32x4_t _k12 = bfloat2float(vld1_u16(k0 + 20));
-            float32x4_t _k20 = bfloat2float(vld1_u16(k0 + 24));
-            float32x4_t _k21 = bfloat2float(vld1_u16(k0 + 28));
-            float32x4_t _k22 = bfloat2float(vld1_u16(k0 + 32));
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-
-#if __aarch64__
-                for (; j + 7 < outw; j += 8)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-
-                        //                         "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0] \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%1], #16   \n"
-                        "ld1    {v2.s}[0], [%1]             \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v24.4s, %8.4s, v0.s[0]      \n"
-                        "fmla   v25.4s, %8.4s, v0.s[1]      \n"
-                        "fmla   v26.4s, %8.4s, v0.s[2]      \n"
-                        "fmla   v27.4s, %8.4s, v0.s[3]      \n"
-                        "fmla   v28.4s, %8.4s, v1.s[0]      \n"
-                        "fmla   v29.4s, %8.4s, v1.s[1]      \n"
-                        "fmla   v30.4s, %8.4s, v1.s[2]      \n"
-                        "fmla   v31.4s, %8.4s, v1.s[3]      \n"
-
-                        "fmla   v24.4s, %9.4s, v0.s[1]      \n"
-                        "fmla   v25.4s, %9.4s, v0.s[2]      \n"
-                        "fmla   v26.4s, %9.4s, v0.s[3]      \n"
-                        "fmla   v27.4s, %9.4s, v1.s[0]      \n"
-                        "fmla   v28.4s, %9.4s, v1.s[1]      \n"
-                        "fmla   v29.4s, %9.4s, v1.s[2]      \n"
-                        "fmla   v30.4s, %9.4s, v1.s[3]      \n"
-                        "fmla   v31.4s, %9.4s, v2.s[0]      \n"
-
-                        "fmla   v24.4s, %10.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %10.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %10.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %10.4s, v1.s[1]     \n"
-                        "fmla   v28.4s, %10.4s, v1.s[2]     \n"
-                        "fmla   v29.4s, %10.4s, v1.s[3]     \n"
-                        "fmla   v30.4s, %10.4s, v2.s[0]     \n"
-                        "fmla   v31.4s, %10.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%2], #16   \n"
-                        "ld1    {v2.s}[0], [%2]             \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v24.4s, %11.4s, v4.s[0]     \n"
-                        "fmla   v25.4s, %11.4s, v4.s[1]     \n"
-                        "fmla   v26.4s, %11.4s, v4.s[2]     \n"
-                        "fmla   v27.4s, %11.4s, v4.s[3]     \n"
-                        "fmla   v28.4s, %11.4s, v5.s[0]     \n"
-                        "fmla   v29.4s, %11.4s, v5.s[1]     \n"
-                        "fmla   v30.4s, %11.4s, v5.s[2]     \n"
-                        "fmla   v31.4s, %11.4s, v5.s[3]     \n"
-
-                        "fmla   v24.4s, %12.4s, v4.s[1]     \n"
-                        "fmla   v25.4s, %12.4s, v4.s[2]     \n"
-                        "fmla   v26.4s, %12.4s, v4.s[3]     \n"
-                        "fmla   v27.4s, %12.4s, v5.s[0]     \n"
-                        "fmla   v28.4s, %12.4s, v5.s[1]     \n"
-                        "fmla   v29.4s, %12.4s, v5.s[2]     \n"
-                        "fmla   v30.4s, %12.4s, v5.s[3]     \n"
-                        "fmla   v31.4s, %12.4s, v2.s[0]     \n"
-
-                        "fmla   v24.4s, %13.4s, v4.s[2]     \n"
-                        "fmla   v25.4s, %13.4s, v4.s[3]     \n"
-                        "fmla   v26.4s, %13.4s, v5.s[0]     \n"
-                        "fmla   v27.4s, %13.4s, v5.s[1]     \n"
-                        "fmla   v28.4s, %13.4s, v5.s[2]     \n"
-                        "fmla   v29.4s, %13.4s, v5.s[3]     \n"
-                        "fmla   v30.4s, %13.4s, v2.s[0]     \n"
-                        "fmla   v31.4s, %13.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%3], #16   \n"
-                        "ld1    {v2.s}[0], [%3]             \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v24.4s, %14.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %14.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %14.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %14.4s, v0.s[3]     \n"
-                        "fmla   v28.4s, %14.4s, v1.s[0]     \n"
-                        "fmla   v29.4s, %14.4s, v1.s[1]     \n"
-                        "fmla   v30.4s, %14.4s, v1.s[2]     \n"
-                        "fmla   v31.4s, %14.4s, v1.s[3]     \n"
-
-                        "fmla   v24.4s, %15.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %15.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %15.4s, v0.s[3]     \n"
-                        "fmla   v27.4s, %15.4s, v1.s[0]     \n"
-                        "fmla   v28.4s, %15.4s, v1.s[1]     \n"
-                        "fmla   v29.4s, %15.4s, v1.s[2]     \n"
-                        "fmla   v30.4s, %15.4s, v1.s[3]     \n"
-                        "fmla   v31.4s, %15.4s, v2.s[0]     \n"
-
-                        "sub    %0, %0, #64                 \n"
-
-                        "fmla   v24.4s, %16.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %16.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %16.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %16.4s, v1.s[1]     \n"
-                        "fmla   v28.4s, %16.4s, v1.s[2]     \n"
-                        "fmla   v29.4s, %16.4s, v1.s[3]     \n"
-                        "fmla   v30.4s, %16.4s, v2.s[0]     \n"
-                        "fmla   v31.4s, %16.4s, v2.s[1]     \n"
-
-                        "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "v0", "v1", "v2", "v4", "v5", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-#endif // __aarch64__
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1], #8           \n"
-
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0] \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "ld1    {v1.s}[0], [%1]             \n"
-
-                        "fmla   v24.4s, %8.4s, v0.s[0]      \n"
-                        "fmla   v25.4s, %8.4s, v0.s[1]      \n"
-
-                        "fmla   v26.4s, %8.4s, v0.s[2]      \n"
-                        "fmla   v27.4s, %8.4s, v0.s[3]      \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v24.4s, %9.4s, v0.s[1]      \n"
-                        "fmla   v25.4s, %9.4s, v0.s[2]      \n"
-
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v2.4h}, [%2], #8           \n"
-
-                        "fmla   v26.4s, %9.4s, v0.s[3]      \n"
-                        "fmla   v27.4s, %9.4s, v1.s[0]      \n"
-
-                        "ld1    {v3.s}[0], [%2]             \n"
-
-                        "fmla   v24.4s, %10.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %10.4s, v0.s[3]     \n"
-
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v26.4s, %10.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %10.4s, v1.s[1]     \n"
-
-                        "fmla   v24.4s, %11.4s, v2.s[0]     \n"
-                        "fmla   v25.4s, %11.4s, v2.s[1]     \n"
-
-                        "fmla   v26.4s, %11.4s, v2.s[2]     \n"
-                        "fmla   v27.4s, %11.4s, v2.s[3]     \n"
-
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v24.4s, %12.4s, v2.s[1]     \n"
-                        "fmla   v25.4s, %12.4s, v2.s[2]     \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v0.4h}, [%3], #8           \n"
-
-                        "fmla   v26.4s, %12.4s, v2.s[3]     \n"
-                        "fmla   v27.4s, %12.4s, v3.s[0]     \n"
-
-                        "ld1    {v1.s}[0], [%3]             \n"
-
-                        "fmla   v24.4s, %13.4s, v2.s[2]     \n"
-                        "fmla   v25.4s, %13.4s, v2.s[3]     \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v26.4s, %13.4s, v3.s[0]     \n"
-                        "fmla   v27.4s, %13.4s, v3.s[1]     \n"
-
-                        "fmla   v24.4s, %14.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %14.4s, v0.s[1]     \n"
-
-                        "fmla   v26.4s, %14.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %14.4s, v0.s[3]     \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v24.4s, %15.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %15.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %15.4s, v0.s[3]     \n"
-                        "fmla   v27.4s, %15.4s, v1.s[0]     \n"
-
-                        "fmla   v24.4s, %16.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %16.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %16.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %16.4s, v1.s[1]     \n"
-
-                        "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "v0", "v1", "v2", "v3", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #512]          \n"
-                        "vldm       %0, {d24-d31}       \n"
-
-                        "pld        [%1, #64]           \n"
-                        "vld1.u16   {d1}, [%1]!         \n"
-                        "vld1.u32   {d2[0]}, [%1]       \n"
-
-                        "vshll.u16  q0, d1, #16         \n"
-                        "vshll.u16  q1, d2, #16         \n"
-
-                        "vmla.f32   q12, %q8, d0[0]     \n"
-                        "vmla.f32   q13, %q8, d0[1]     \n"
-                        "vmla.f32   q14, %q8, d1[0]     \n"
-                        "vmla.f32   q15, %q8, d1[1]     \n"
-
-                        "vmla.f32   q12, %q9, d0[1]     \n"
-                        "vmla.f32   q13, %q9, d1[0]     \n"
-                        "vmla.f32   q14, %q9, d1[1]     \n"
-                        "vmla.f32   q15, %q9, d2[0]     \n"
-
-                        "vmla.f32   q12, %q10, d1[0]    \n"
-                        "vmla.f32   q13, %q10, d1[1]    \n"
-                        "vmla.f32   q14, %q10, d2[0]    \n"
-                        "vmla.f32   q15, %q10, d2[1]    \n"
-
-                        "pld        [%2, #64]           \n"
-                        "vld1.u16   {d5}, [%2]!         \n"
-                        "vld1.u32   {d3[0]}, [%2]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q12, %q11, d4[0]    \n"
-                        "vmla.f32   q13, %q11, d4[1]    \n"
-                        "vmla.f32   q14, %q11, d5[0]    \n"
-                        "vmla.f32   q15, %q11, d5[1]    \n"
-
-                        "vmla.f32   q12, %q12, d4[1]    \n"
-                        "vmla.f32   q13, %q12, d5[0]    \n"
-                        "vmla.f32   q14, %q12, d5[1]    \n"
-                        "vmla.f32   q15, %q12, d2[0]    \n"
-
-                        "vmla.f32   q12, %q13, d5[0]    \n"
-                        "vmla.f32   q13, %q13, d5[1]    \n"
-                        "vmla.f32   q14, %q13, d2[0]    \n"
-                        "vmla.f32   q15, %q13, d2[1]    \n"
-
-                        "pld        [%3, #64]           \n"
-                        "vld1.u16   {d1}, [%3]!         \n"
-                        "vld1.u32   {d2[0]}, [%3]       \n"
-
-                        "vshll.u16  q0, d1, #16         \n"
-                        "vshll.u16  q1, d2, #16         \n"
-
-                        "vmla.f32   q12, %q14, d0[0]    \n"
-                        "vmla.f32   q13, %q14, d0[1]    \n"
-                        "vmla.f32   q14, %q14, d1[0]    \n"
-                        "vmla.f32   q15, %q14, d1[1]    \n"
-
-                        "vmla.f32   q12, %q15, d0[1]    \n"
-                        "vmla.f32   q13, %q15, d1[0]    \n"
-                        "vmla.f32   q14, %q15, d1[1]    \n"
-                        "vmla.f32   q15, %q15, d2[0]    \n"
-
-                        "vmla.f32   q12, %q16, d1[0]    \n"
-                        "vmla.f32   q13, %q16, d1[1]    \n"
-                        "vmla.f32   q14, %q16, d2[0]    \n"
-                        "vmla.f32   q15, %q16, d2[1]    \n"
-
-                        "vstm       %0!, {d24-d31}      \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "q0", "q1", "q2", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1]               \n"
-
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v28.4s, v29.4s}, [%0]      \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmul   v24.4s, %8.4s, v0.s[0]      \n"
-                        "fmul   v25.4s, %8.4s, v0.s[1]      \n"
-
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v1.4h}, [%2]               \n"
-
-                        "fmul   v26.4s, %9.4s, v0.s[1]      \n"
-                        "fmul   v27.4s, %9.4s, v0.s[2]      \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v28.4s, %10.4s, v0.s[2]     \n"
-                        "fmla   v29.4s, %10.4s, v0.s[3]     \n"
-
-                        "fmla   v24.4s, %11.4s, v1.s[0]     \n"
-                        "fmla   v25.4s, %11.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v0.4h}, [%3]               \n"
-
-                        "fmla   v26.4s, %12.4s, v1.s[1]     \n"
-                        "fmla   v27.4s, %12.4s, v1.s[2]     \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v28.4s, %13.4s, v1.s[2]     \n"
-                        "fmla   v29.4s, %13.4s, v1.s[3]     \n"
-
-                        "fmla   v24.4s, %14.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %14.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %15.4s, v0.s[1]     \n"
-                        "fmla   v27.4s, %15.4s, v0.s[2]     \n"
-
-                        "fmla   v28.4s, %16.4s, v0.s[2]     \n"
-                        "fmla   v29.4s, %16.4s, v0.s[3]     \n"
-
-                        "add    %1, %1, #4                  \n"
-
-                        "fadd   v24.4s, v24.4s, v26.4s      \n"
-                        "fadd   v25.4s, v25.4s, v27.4s      \n"
-
-                        "add    %2, %2, #4                  \n"
-
-                        "fadd   v28.4s, v28.4s, v24.4s      \n"
-                        "fadd   v29.4s, v29.4s, v25.4s      \n"
-
-                        "add    %3, %3, #4                  \n"
-
-                        "st1    {v28.4s, v29.4s}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "v0", "v1", "v24", "v25", "v26", "v27", "v28", "v29");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #64]           \n"
-                        "vld1.u16   {d1}, [%1]          \n"
-
-                        "pld        [%0, #256]          \n"
-                        "vld1.f32   {d24-d27}, [%0 :128] \n"
-
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmul.f32   q14, %q8, d0[0]     \n"
-                        "vmul.f32   q15, %q8, d0[1]     \n"
-                        "vmla.f32   q12, %q9, d0[1]     \n"
-                        "vmla.f32   q13, %q9, d1[0]     \n"
-
-                        "pld        [%2, #64]           \n"
-                        "vld1.u16   {d3}, [%2]          \n"
-
-                        "vmla.f32   q14, %q10, d1[0]    \n"
-                        "vmla.f32   q15, %q10, d1[1]    \n"
-
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q12, %q11, d2[0]    \n"
-                        "vmla.f32   q13, %q11, d2[1]    \n"
-
-                        "vmla.f32   q14, %q12, d2[1]    \n"
-                        "vmla.f32   q15, %q12, d3[0]    \n"
-
-                        "pld        [%3, #64]           \n"
-                        "vld1.u16   {d1}, [%3]          \n"
-
-                        "vmla.f32   q12, %q13, d3[0]    \n"
-                        "vmla.f32   q13, %q13, d3[1]    \n"
-
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q14, %q14, d0[0]    \n"
-                        "vmla.f32   q15, %q14, d0[1]    \n"
-
-                        "vmla.f32   q12, %q15, d0[1]    \n"
-                        "vmla.f32   q13, %q15, d1[0]    \n"
-
-                        "add        %1, %1, #4          \n"
-
-                        "vmla.f32   q14, %q16, d1[0]    \n"
-                        "vmla.f32   q15, %q16, d1[1]    \n"
-
-                        "add        %2, %2, #4          \n"
-
-                        "vadd.f32   q12, q12, q14       \n"
-                        "vadd.f32   q13, q13, q15       \n"
-
-                        "add        %3, %3, #4          \n"
-
-                        "vst1.f32   {d24-d27}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "q0", "q1", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-                    float32x4_t _sum0 = vld1q_f32(outptr0);
-
-                    float32x4_t _r0 = bfloat2float(vld1_u16(r0));
-                    float32x4_t _r1 = bfloat2float(vld1_u16(r1));
-                    float32x4_t _r2 = bfloat2float(vld1_u16(r2));
-
-#if __aarch64__
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k00, _r0, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k01, _r0, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k02, _r0, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k10, _r1, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k11, _r1, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k12, _r1, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k20, _r2, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k21, _r2, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k22, _r2, 2);
-#else
-                    _sum0 = vmlaq_lane_f32(_sum0, _k00, vget_low_f32(_r0), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k01, vget_low_f32(_r0), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k02, vget_high_f32(_r0), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k10, vget_low_f32(_r1), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k11, vget_low_f32(_r1), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k12, vget_high_f32(_r1), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k20, vget_low_f32(_r2), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k21, vget_low_f32(_r2), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k22, vget_high_f32(_r2), 0);
-#endif
-
-                    vst1q_f32(outptr0, _sum0);
-
-                    r0 += 1;
-                    r1 += 1;
-                    r2 += 1;
-                    outptr0 += 4;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * 4;
-        }
-        for (; q < inch; q++)
-        {
-            unsigned short* outptr0_bf16 = top_blob.channel(p);
-
-            const float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<unsigned short>(0);
-            const unsigned short* r1 = img0.row<unsigned short>(1);
-            const unsigned short* r2 = img0.row<unsigned short>(2);
-
-            float32x4_t _k00 = bfloat2float(vld1_u16(k0));
-            float32x4_t _k01 = bfloat2float(vld1_u16(k0 + 4));
-            float32x4_t _k02 = bfloat2float(vld1_u16(k0 + 8));
-            float32x4_t _k10 = bfloat2float(vld1_u16(k0 + 12));
-            float32x4_t _k11 = bfloat2float(vld1_u16(k0 + 16));
-            float32x4_t _k12 = bfloat2float(vld1_u16(k0 + 20));
-            float32x4_t _k20 = bfloat2float(vld1_u16(k0 + 24));
-            float32x4_t _k21 = bfloat2float(vld1_u16(k0 + 28));
-            float32x4_t _k22 = bfloat2float(vld1_u16(k0 + 32));
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-
-#if __aarch64__
-                for (; j + 7 < outw; j += 8)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%1], #64 \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%1], #64 \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%2], #16   \n"
-                        "ld1    {v2.s}[0], [%2]             \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v24.4s, %10.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %10.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %10.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %10.4s, v0.s[3]     \n"
-                        "fmla   v28.4s, %10.4s, v1.s[0]     \n"
-                        "fmla   v29.4s, %10.4s, v1.s[1]     \n"
-                        "fmla   v30.4s, %10.4s, v1.s[2]     \n"
-                        "fmla   v31.4s, %10.4s, v1.s[3]     \n"
-
-                        "fmla   v24.4s, %11.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %11.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %11.4s, v0.s[3]     \n"
-                        "fmla   v27.4s, %11.4s, v1.s[0]     \n"
-                        "fmla   v28.4s, %11.4s, v1.s[1]     \n"
-                        "fmla   v29.4s, %11.4s, v1.s[2]     \n"
-                        "fmla   v30.4s, %11.4s, v1.s[3]     \n"
-                        "fmla   v31.4s, %11.4s, v2.s[0]     \n"
-
-                        "fmla   v24.4s, %12.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %12.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %12.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %12.4s, v1.s[1]     \n"
-                        "fmla   v28.4s, %12.4s, v1.s[2]     \n"
-                        "fmla   v29.4s, %12.4s, v1.s[3]     \n"
-                        "fmla   v30.4s, %12.4s, v2.s[0]     \n"
-                        "fmla   v31.4s, %12.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%3], #16   \n"
-                        "ld1    {v2.s}[0], [%3]             \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v24.4s, %13.4s, v4.s[0]     \n"
-                        "fmla   v25.4s, %13.4s, v4.s[1]     \n"
-                        "fmla   v26.4s, %13.4s, v4.s[2]     \n"
-                        "fmla   v27.4s, %13.4s, v4.s[3]     \n"
-                        "fmla   v28.4s, %13.4s, v5.s[0]     \n"
-                        "fmla   v29.4s, %13.4s, v5.s[1]     \n"
-                        "fmla   v30.4s, %13.4s, v5.s[2]     \n"
-                        "fmla   v31.4s, %13.4s, v5.s[3]     \n"
-
-                        "fmla   v24.4s, %14.4s, v4.s[1]     \n"
-                        "fmla   v25.4s, %14.4s, v4.s[2]     \n"
-                        "fmla   v26.4s, %14.4s, v4.s[3]     \n"
-                        "fmla   v27.4s, %14.4s, v5.s[0]     \n"
-                        "fmla   v28.4s, %14.4s, v5.s[1]     \n"
-                        "fmla   v29.4s, %14.4s, v5.s[2]     \n"
-                        "fmla   v30.4s, %14.4s, v5.s[3]     \n"
-                        "fmla   v31.4s, %14.4s, v2.s[0]     \n"
-
-                        "fmla   v24.4s, %15.4s, v4.s[2]     \n"
-                        "fmla   v25.4s, %15.4s, v4.s[3]     \n"
-                        "fmla   v26.4s, %15.4s, v5.s[0]     \n"
-                        "fmla   v27.4s, %15.4s, v5.s[1]     \n"
-                        "fmla   v28.4s, %15.4s, v5.s[2]     \n"
-                        "fmla   v29.4s, %15.4s, v5.s[3]     \n"
-                        "fmla   v30.4s, %15.4s, v2.s[0]     \n"
-                        "fmla   v31.4s, %15.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%4], #16   \n"
-                        "ld1    {v2.s}[0], [%4]             \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v24.4s, %16.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %16.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %16.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %16.4s, v0.s[3]     \n"
-                        "fmla   v28.4s, %16.4s, v1.s[0]     \n"
-                        "fmla   v29.4s, %16.4s, v1.s[1]     \n"
-                        "fmla   v30.4s, %16.4s, v1.s[2]     \n"
-                        "fmla   v31.4s, %16.4s, v1.s[3]     \n"
-
-                        "fmla   v24.4s, %17.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %17.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %17.4s, v0.s[3]     \n"
-                        "fmla   v27.4s, %17.4s, v1.s[0]     \n"
-                        "fmla   v28.4s, %17.4s, v1.s[1]     \n"
-                        "fmla   v29.4s, %17.4s, v1.s[2]     \n"
-                        "fmla   v30.4s, %17.4s, v1.s[3]     \n"
-                        "fmla   v31.4s, %17.4s, v2.s[0]     \n"
-
-                        "fmla   v24.4s, %18.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %18.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %18.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %18.4s, v1.s[1]     \n"
-                        "fmla   v28.4s, %18.4s, v1.s[2]     \n"
-                        "fmla   v29.4s, %18.4s, v1.s[3]     \n"
-                        "fmla   v30.4s, %18.4s, v2.s[0]     \n"
-                        "fmla   v31.4s, %18.4s, v2.s[1]     \n"
-
-                        "shrn   v24.4h, v24.4s, #16         \n"
-                        "shrn   v25.4h, v25.4s, #16         \n"
-                        "shrn   v26.4h, v26.4s, #16         \n"
-                        "shrn   v27.4h, v27.4s, #16         \n"
-                        "shrn   v28.4h, v28.4s, #16         \n"
-                        "shrn   v29.4h, v29.4s, #16         \n"
-                        "shrn   v30.4h, v30.4s, #16         \n"
-                        "shrn   v31.4h, v31.4s, #16         \n"
-
-                        "st1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%0], #32 \n"
-                        "st1    {v28.4h, v29.4h, v30.4h, v31.4h}, [%0], #32 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2)            // %4
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00), // %10
-                        "w"(_k01), // %11
-                        "w"(_k02), // %12
-                        "w"(_k10), // %13
-                        "w"(_k11), // %14
-                        "w"(_k12), // %15
-                        "w"(_k20), // %16
-                        "w"(_k21), // %17
-                        "w"(_k22)  // %18
-                        : "memory", "v0", "v1", "v2", "v4", "v5", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-#endif // __aarch64__
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v0.4h}, [%2], #8           \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%1], #64 \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "ld1    {v1.s}[0], [%2]             \n"
-
-                        "fmla   v24.4s, %10.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %10.4s, v0.s[1]     \n"
-
-                        "fmla   v26.4s, %10.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %10.4s, v0.s[3]     \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v24.4s, %11.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %11.4s, v0.s[2]     \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v2.4h}, [%3], #8           \n"
-
-                        "fmla   v26.4s, %11.4s, v0.s[3]     \n"
-                        "fmla   v27.4s, %11.4s, v1.s[0]     \n"
-
-                        "ld1    {v3.s}[0], [%3]             \n"
-
-                        "fmla   v24.4s, %12.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %12.4s, v0.s[3]     \n"
-
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v26.4s, %12.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %12.4s, v1.s[1]     \n"
-
-                        "fmla   v24.4s, %13.4s, v2.s[0]     \n"
-                        "fmla   v25.4s, %13.4s, v2.s[1]     \n"
-
-                        "fmla   v26.4s, %13.4s, v2.s[2]     \n"
-                        "fmla   v27.4s, %13.4s, v2.s[3]     \n"
-
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v24.4s, %14.4s, v2.s[1]     \n"
-                        "fmla   v25.4s, %14.4s, v2.s[2]     \n"
-
-                        "prfm   pldl1keep, [%4, #64]        \n"
-                        "ld1    {v0.4h}, [%4], #8           \n"
-
-                        "fmla   v26.4s, %14.4s, v2.s[3]     \n"
-                        "fmla   v27.4s, %14.4s, v3.s[0]     \n"
-
-                        "ld1    {v1.s}[0], [%4]             \n"
-
-                        "fmla   v24.4s, %15.4s, v2.s[2]     \n"
-                        "fmla   v25.4s, %15.4s, v2.s[3]     \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v26.4s, %15.4s, v3.s[0]     \n"
-                        "fmla   v27.4s, %15.4s, v3.s[1]     \n"
-
-                        "fmla   v24.4s, %16.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %16.4s, v0.s[1]     \n"
-
-                        "fmla   v26.4s, %16.4s, v0.s[2]     \n"
-                        "fmla   v27.4s, %16.4s, v0.s[3]     \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v24.4s, %17.4s, v0.s[1]     \n"
-                        "fmla   v25.4s, %17.4s, v0.s[2]     \n"
-                        "fmla   v26.4s, %17.4s, v0.s[3]     \n"
-                        "fmla   v27.4s, %17.4s, v1.s[0]     \n"
-
-                        "fmla   v24.4s, %18.4s, v0.s[2]     \n"
-                        "fmla   v25.4s, %18.4s, v0.s[3]     \n"
-                        "fmla   v26.4s, %18.4s, v1.s[0]     \n"
-                        "fmla   v27.4s, %18.4s, v1.s[1]     \n"
-
-                        "shrn   v24.4h, v24.4s, #16         \n"
-                        "shrn   v25.4h, v25.4s, #16         \n"
-                        "shrn   v26.4h, v26.4s, #16         \n"
-                        "shrn   v27.4h, v27.4s, #16         \n"
-
-                        "st1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%0], #32 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2)            // %4
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00), // %10
-                        "w"(_k01), // %11
-                        "w"(_k02), // %12
-                        "w"(_k10), // %13
-                        "w"(_k11), // %14
-                        "w"(_k12), // %15
-                        "w"(_k20), // %16
-                        "w"(_k21), // %17
-                        "w"(_k22)  // %18
-                        : "memory", "v0", "v1", "v2", "v3", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1!, {d24-d31}      \n"
-
-                        "pld        [%2, #64]           \n"
-                        "vld1.u16   {d1}, [%2]!         \n"
-                        "vld1.u32   {d2[0]}, [%2]       \n"
-
-                        "vshll.u16  q0, d1, #16         \n"
-                        "vshll.u16  q1, d2, #16         \n"
-
-                        "vmla.f32   q12, %q10, d0[0]    \n"
-                        "vmla.f32   q13, %q10, d0[1]    \n"
-                        "vmla.f32   q14, %q10, d1[0]    \n"
-                        "vmla.f32   q15, %q10, d1[1]    \n"
-
-                        "vmla.f32   q12, %q11, d0[1]    \n"
-                        "vmla.f32   q13, %q11, d1[0]    \n"
-                        "vmla.f32   q14, %q11, d1[1]    \n"
-                        "vmla.f32   q15, %q11, d2[0]    \n"
-
-                        "vmla.f32   q12, %q12, d1[0]    \n"
-                        "vmla.f32   q13, %q12, d1[1]    \n"
-                        "vmla.f32   q14, %q12, d2[0]    \n"
-                        "vmla.f32   q15, %q12, d2[1]    \n"
-
-                        "pld        [%3, #64]           \n"
-                        "vld1.u16   {d5}, [%3]!         \n"
-                        "vld1.u32   {d3[0]}, [%3]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q12, %q13, d4[0]    \n"
-                        "vmla.f32   q13, %q13, d4[1]    \n"
-                        "vmla.f32   q14, %q13, d5[0]    \n"
-                        "vmla.f32   q15, %q13, d5[1]    \n"
-
-                        "vmla.f32   q12, %q14, d4[1]    \n"
-                        "vmla.f32   q13, %q14, d5[0]    \n"
-                        "vmla.f32   q14, %q14, d5[1]    \n"
-                        "vmla.f32   q15, %q14, d2[0]    \n"
-
-                        "vmla.f32   q12, %q15, d5[0]    \n"
-                        "vmla.f32   q13, %q15, d5[1]    \n"
-                        "vmla.f32   q14, %q15, d2[0]    \n"
-                        "vmla.f32   q15, %q15, d2[1]    \n"
-
-                        "pld        [%4, #64]           \n"
-                        "vld1.u16   {d1}, [%4]!         \n"
-                        "vld1.u32   {d2[0]}, [%4]       \n"
-
-                        "vshll.u16  q0, d1, #16         \n"
-                        "vshll.u16  q1, d2, #16         \n"
-
-                        "vmla.f32   q12, %q16, d0[0]    \n"
-                        "vmla.f32   q13, %q16, d0[1]    \n"
-                        "vmla.f32   q14, %q16, d1[0]    \n"
-                        "vmla.f32   q15, %q16, d1[1]    \n"
-
-                        "vmla.f32   q12, %q17, d0[1]    \n"
-                        "vmla.f32   q13, %q17, d1[0]    \n"
-                        "vmla.f32   q14, %q17, d1[1]    \n"
-                        "vmla.f32   q15, %q17, d2[0]    \n"
-
-                        "vmla.f32   q12, %q18, d1[0]    \n"
-                        "vmla.f32   q13, %q18, d1[1]    \n"
-                        "vmla.f32   q14, %q18, d2[0]    \n"
-                        "vmla.f32   q15, %q18, d2[1]    \n"
-
-                        "vshrn.s32  d24, q12, #16       \n"
-                        "vshrn.s32  d25, q13, #16       \n"
-                        "vshrn.s32  d26, q14, #16       \n"
-                        "vshrn.s32  d27, q15, #16       \n"
-
-                        "vst1.u16   {d24-d27}, [%0 :64]! \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2)            // %4
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00), // %10
-                        "w"(_k01), // %11
-                        "w"(_k02), // %12
-                        "w"(_k10), // %13
-                        "w"(_k11), // %14
-                        "w"(_k12), // %15
-                        "w"(_k20), // %16
-                        "w"(_k21), // %17
-                        "w"(_k22)  // %18
-                        : "memory", "q0", "q1", "q2", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v0.4h}, [%2]               \n"
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v28.4s, v29.4s}, [%1], #32 \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmul   v24.4s, %10.4s, v0.s[0]     \n"
-                        "fmul   v25.4s, %10.4s, v0.s[1]     \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v1.4h}, [%3]               \n"
-
-                        "fmul   v26.4s, %11.4s, v0.s[1]     \n"
-                        "fmul   v27.4s, %11.4s, v0.s[2]     \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v28.4s, %12.4s, v0.s[2]     \n"
-                        "fmla   v29.4s, %12.4s, v0.s[3]     \n"
-
-                        "fmla   v24.4s, %13.4s, v1.s[0]     \n"
-                        "fmla   v25.4s, %13.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #64]        \n"
-                        "ld1    {v0.4h}, [%4]               \n"
-
-                        "fmla   v26.4s, %14.4s, v1.s[1]     \n"
-                        "fmla   v27.4s, %14.4s, v1.s[2]     \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v28.4s, %15.4s, v1.s[2]     \n"
-                        "fmla   v29.4s, %15.4s, v1.s[3]     \n"
-
-                        "fmla   v24.4s, %16.4s, v0.s[0]     \n"
-                        "fmla   v25.4s, %16.4s, v0.s[1]     \n"
-                        "fmla   v26.4s, %17.4s, v0.s[1]     \n"
-                        "fmla   v27.4s, %17.4s, v0.s[2]     \n"
-
-                        "fmla   v28.4s, %18.4s, v0.s[2]     \n"
-                        "fmla   v29.4s, %18.4s, v0.s[3]     \n"
-
-                        "add    %2, %2, #4                  \n"
-
-                        "fadd   v24.4s, v24.4s, v26.4s      \n"
-                        "fadd   v25.4s, v25.4s, v27.4s      \n"
-
-                        "add    %3, %3, #4                  \n"
-
-                        "fadd   v28.4s, v28.4s, v24.4s      \n"
-                        "fadd   v29.4s, v29.4s, v25.4s      \n"
-
-                        "add    %4, %4, #4                  \n"
-
-                        "shrn   v28.4h, v28.4s, #16         \n"
-                        "shrn   v29.4h, v29.4s, #16         \n"
-
-                        "st1    {v28.4h, v29.4h}, [%0], #16 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2)            // %4
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00), // %10
-                        "w"(_k01), // %11
-                        "w"(_k02), // %12
-                        "w"(_k10), // %13
-                        "w"(_k11), // %14
-                        "w"(_k12), // %15
-                        "w"(_k20), // %16
-                        "w"(_k21), // %17
-                        "w"(_k22)  // %18
-                        : "memory", "v0", "v1", "v24", "v25", "v26", "v27", "v28", "v29");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%2, #64]           \n"
-                        "vld1.u16   {d1}, [%2]          \n"
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.f32   {d24-d27}, [%1 :128]! \n"
-
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmul.f32   q14, %q10, d0[0]    \n"
-                        "vmul.f32   q15, %q10, d0[1]    \n"
-                        "vmla.f32   q12, %q11, d0[1]    \n"
-                        "vmla.f32   q13, %q11, d1[0]    \n"
-
-                        "pld        [%3, #64]           \n"
-                        "vld1.u16   {d3}, [%3]          \n"
-
-                        "vmla.f32   q14, %q12, d1[0]    \n"
-                        "vmla.f32   q15, %q12, d1[1]    \n"
-
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q12, %q13, d2[0]    \n"
-                        "vmla.f32   q13, %q13, d2[1]    \n"
-
-                        "vmla.f32   q14, %q14, d2[1]    \n"
-                        "vmla.f32   q15, %q14, d3[0]    \n"
-
-                        "pld        [%4, #64]           \n"
-                        "vld1.u16   {d1}, [%4]          \n"
-
-                        "vmla.f32   q12, %q15, d3[0]    \n"
-                        "vmla.f32   q13, %q15, d3[1]    \n"
-
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q14, %q16, d0[0]    \n"
-                        "vmla.f32   q15, %q16, d0[1]    \n"
-
-                        "vmla.f32   q12, %q17, d0[1]    \n"
-                        "vmla.f32   q13, %q17, d1[0]    \n"
-
-                        "add        %2, %2, #4          \n"
-
-                        "vmla.f32   q14, %q18, d1[0]    \n"
-                        "vmla.f32   q15, %q18, d1[1]    \n"
-
-                        "add        %3, %3, #4          \n"
-
-                        "vadd.f32   q12, q12, q14       \n"
-                        "vadd.f32   q13, q13, q15       \n"
-
-                        "add        %4, %4, #4          \n"
-
-                        "vshrn.s32  d24, q12, #16       \n"
-                        "vshrn.s32  d25, q13, #16       \n"
-
-                        "vst1.f32   {d24-d25}, [%0 :64]! \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2)            // %4
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00), // %10
-                        "w"(_k01), // %11
-                        "w"(_k02), // %12
-                        "w"(_k10), // %13
-                        "w"(_k11), // %14
-                        "w"(_k12), // %15
-                        "w"(_k20), // %16
-                        "w"(_k21), // %17
-                        "w"(_k22)  // %18
-                        : "memory", "q0", "q1", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-                    float32x4_t _sum0 = vld1q_f32(outptr0);
-
-                    float32x4_t _r0 = bfloat2float(vld1_u16(r0));
-                    float32x4_t _r1 = bfloat2float(vld1_u16(r1));
-                    float32x4_t _r2 = bfloat2float(vld1_u16(r2));
-
-#if __aarch64__
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k00, _r0, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k01, _r0, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k02, _r0, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k10, _r1, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k11, _r1, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k12, _r1, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k20, _r2, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k21, _r2, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k22, _r2, 2);
-#else
-                    _sum0 = vmlaq_lane_f32(_sum0, _k00, vget_low_f32(_r0), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k01, vget_low_f32(_r0), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k02, vget_high_f32(_r0), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k10, vget_low_f32(_r1), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k11, vget_low_f32(_r1), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k12, vget_high_f32(_r1), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k20, vget_low_f32(_r2), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k21, vget_low_f32(_r2), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k22, vget_high_f32(_r2), 0);
-#endif
-
-                    vst1_u16(outptr0_bf16, float2bfloat(_sum0));
-
-                    r0 += 1;
-                    r1 += 1;
-                    r2 += 1;
-                    outptr0 += 4;
-                    outptr0_bf16 += 4;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * 4;
-        }
-    }
-}
-
-static void conv3x3s2_pack1to4_bf16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-#if __ARM_NEON && __aarch64__
-    Mat top_blob_fp32(outw, outh, opt.num_threads, (size_t)4u * 4 * 2, 4 * 2, opt.workspace_allocator);
-#else
-    Mat top_blob_fp32(outw, outh, opt.num_threads, (size_t)4u * 4, 4, opt.workspace_allocator);
-#endif
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* bias = _bias;
-
-    int remain_outch_start = 0;
-
-#if __ARM_NEON && __aarch64__
-    int nn_outch = 0;
-    nn_outch = outch >> 1;
-    remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        Mat out0 = top_blob_fp32.channel(get_omp_thread_num());
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        float32x4_t _bias1 = bias ? vld1q_f32((const float*)bias + (p + 1) * 4) : vdupq_n_f32(0.f);
-        {
-            float* ptr = (float*)out0;
-
-            for (int i = 0; i < outh; i++)
-            {
-                int j = 0;
-
-                for (; j + 3 < outw; j += 4)
-                {
-                    vst1q_f32(ptr, _bias0);
-                    vst1q_f32(ptr + 4, _bias0);
-                    vst1q_f32(ptr + 8, _bias0);
-                    vst1q_f32(ptr + 12, _bias0);
-                    vst1q_f32(ptr + 16, _bias1);
-                    vst1q_f32(ptr + 20, _bias1);
-                    vst1q_f32(ptr + 24, _bias1);
-                    vst1q_f32(ptr + 28, _bias1);
-                    ptr += 32;
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    vst1q_f32(ptr, _bias0);
-                    vst1q_f32(ptr + 4, _bias0);
-                    vst1q_f32(ptr + 8, _bias1);
-                    vst1q_f32(ptr + 12, _bias1);
-                    ptr += 16;
-                }
-                for (; j < outw; j++)
-                {
-                    vst1q_f32(ptr, _bias0);
-                    vst1q_f32(ptr + 4, _bias1);
-                    ptr += 8;
-                }
-            }
-        }
-
-        const unsigned short* k0 = kernel.channel(p);
-        const unsigned short* k1 = kernel.channel(p + 1);
-
-        int q = 0;
-        for (; q < inch - 1; q++)
-        {
-            float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<const unsigned short>(0);
-            const unsigned short* r1 = img0.row<const unsigned short>(1);
-            const unsigned short* r2 = img0.row<const unsigned short>(2);
-
-            float32x4_t _k00_0 = bfloat2float(vld1_u16(k0));
-            float32x4_t _k01_0 = bfloat2float(vld1_u16(k0 + 4));
-            float32x4_t _k02_0 = bfloat2float(vld1_u16(k0 + 8));
-            float32x4_t _k10_0 = bfloat2float(vld1_u16(k0 + 12));
-            float32x4_t _k11_0 = bfloat2float(vld1_u16(k0 + 16));
-            float32x4_t _k12_0 = bfloat2float(vld1_u16(k0 + 20));
-            float32x4_t _k20_0 = bfloat2float(vld1_u16(k0 + 24));
-            float32x4_t _k21_0 = bfloat2float(vld1_u16(k0 + 28));
-            float32x4_t _k22_0 = bfloat2float(vld1_u16(k0 + 32));
-
-            float32x4_t _k00_1 = bfloat2float(vld1_u16(k1));
-            float32x4_t _k01_1 = bfloat2float(vld1_u16(k1 + 4));
-            float32x4_t _k02_1 = bfloat2float(vld1_u16(k1 + 8));
-            float32x4_t _k10_1 = bfloat2float(vld1_u16(k1 + 12));
-            float32x4_t _k11_1 = bfloat2float(vld1_u16(k1 + 16));
-            float32x4_t _k12_1 = bfloat2float(vld1_u16(k1 + 20));
-            float32x4_t _k20_1 = bfloat2float(vld1_u16(k1 + 24));
-            float32x4_t _k21_1 = bfloat2float(vld1_u16(k1 + 28));
-            float32x4_t _k22_1 = bfloat2float(vld1_u16(k1 + 32));
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-
-                for (; j + 3 < outw; j += 4)
-                {
-                    asm volatile(
-                        // r0
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%1], #16   \n"
-
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v6.4s, v7.4s, v8.4s, v9.4s}, [%0], #64 \n" // sum0
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        //                         "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%0] \n" // sum1
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v6.4s, %8.4s, v0.s[0]       \n"
-                        "fmla   v7.4s, %8.4s, v0.s[2]       \n"
-                        "fmla   v8.4s, %8.4s, v1.s[0]       \n"
-                        "fmla   v9.4s, %8.4s, v1.s[2]       \n"
-                        "fmla   v10.4s, %17.4s, v0.s[0]     \n"
-                        "fmla   v11.4s, %17.4s, v0.s[2]     \n"
-                        "fmla   v12.4s, %17.4s, v1.s[0]     \n"
-                        "fmla   v13.4s, %17.4s, v1.s[2]     \n"
-
-                        "ld1    {v4.h}[0], [%1]             \n"
-
-                        "fmla   v6.4s, %9.4s, v0.s[1]       \n"
-                        "fmla   v7.4s, %9.4s, v0.s[3]       \n"
-                        "fmla   v8.4s, %9.4s, v1.s[1]       \n"
-                        "fmla   v9.4s, %9.4s, v1.s[3]       \n"
-                        "fmla   v10.4s, %18.4s, v0.s[1]     \n"
-                        "fmla   v11.4s, %18.4s, v0.s[3]     \n"
-                        "fmla   v12.4s, %18.4s, v1.s[1]     \n"
-                        "fmla   v13.4s, %18.4s, v1.s[3]     \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        // r1
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v2.4h, v3.4h}, [%2], #16   \n"
-
-                        "fmla   v6.4s, %10.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %10.4s, v1.s[0]      \n"
-                        "fmla   v8.4s, %10.4s, v1.s[2]      \n"
-                        "fmla   v9.4s, %10.4s, v4.s[0]      \n"
-
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v10.4s, %19.4s, v0.s[2]     \n"
-                        "fmla   v11.4s, %19.4s, v1.s[0]     \n"
-                        "fmla   v12.4s, %19.4s, v1.s[2]     \n"
-                        "fmla   v13.4s, %19.4s, v4.s[0]     \n"
-
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v6.4s, %11.4s, v2.s[0]      \n"
-                        "fmla   v7.4s, %11.4s, v2.s[2]      \n"
-                        "fmla   v8.4s, %11.4s, v3.s[0]      \n"
-                        "fmla   v9.4s, %11.4s, v3.s[2]      \n"
-                        "fmla   v10.4s, %20.4s, v2.s[0]     \n"
-                        "fmla   v11.4s, %20.4s, v2.s[2]     \n"
-                        "fmla   v12.4s, %20.4s, v3.s[0]     \n"
-                        "fmla   v13.4s, %20.4s, v3.s[2]     \n"
-
-                        "ld1    {v5.h}[0], [%2]             \n"
-
-                        "fmla   v6.4s, %12.4s, v2.s[1]      \n"
-                        "fmla   v7.4s, %12.4s, v2.s[3]      \n"
-                        "fmla   v8.4s, %12.4s, v3.s[1]      \n"
-                        "fmla   v9.4s, %12.4s, v3.s[3]      \n"
-
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v10.4s, %21.4s, v2.s[1]     \n"
-                        "fmla   v11.4s, %21.4s, v2.s[3]     \n"
-                        "fmla   v12.4s, %21.4s, v3.s[1]     \n"
-                        "fmla   v13.4s, %21.4s, v3.s[3]     \n"
-
-                        // r2
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%3], #16   \n"
-
-                        "fmla   v6.4s, %13.4s, v2.s[2]      \n"
-                        "fmla   v7.4s, %13.4s, v3.s[0]      \n"
-                        "fmla   v8.4s, %13.4s, v3.s[2]      \n"
-                        "fmla   v9.4s, %13.4s, v5.s[0]      \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v10.4s, %22.4s, v2.s[2]     \n"
-                        "fmla   v11.4s, %22.4s, v3.s[0]     \n"
-                        "fmla   v12.4s, %22.4s, v3.s[2]     \n"
-                        "fmla   v13.4s, %22.4s, v5.s[0]     \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v6.4s, %14.4s, v0.s[0]      \n"
-                        "fmla   v7.4s, %14.4s, v0.s[2]      \n"
-                        "fmla   v8.4s, %14.4s, v1.s[0]      \n"
-                        "fmla   v9.4s, %14.4s, v1.s[2]      \n"
-                        "fmla   v10.4s, %23.4s, v0.s[0]     \n"
-                        "fmla   v11.4s, %23.4s, v0.s[2]     \n"
-                        "fmla   v12.4s, %23.4s, v1.s[0]     \n"
-                        "fmla   v13.4s, %23.4s, v1.s[2]     \n"
-
-                        "ld1    {v4.h}[0], [%3]             \n"
-
-                        "fmla   v6.4s, %15.4s, v0.s[1]      \n"
-                        "fmla   v7.4s, %15.4s, v0.s[3]      \n"
-                        "fmla   v8.4s, %15.4s, v1.s[1]      \n"
-                        "fmla   v9.4s, %15.4s, v1.s[3]      \n"
-                        "fmla   v10.4s, %24.4s, v0.s[1]     \n"
-                        "fmla   v11.4s, %24.4s, v0.s[3]     \n"
-                        "fmla   v12.4s, %24.4s, v1.s[1]     \n"
-                        "fmla   v13.4s, %24.4s, v1.s[3]     \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "fmla   v6.4s, %16.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %16.4s, v1.s[0]      \n"
-                        "fmla   v8.4s, %16.4s, v1.s[2]      \n"
-                        "fmla   v9.4s, %16.4s, v4.s[0]      \n"
-
-                        "sub    %0, %0, #64                 \n"
-
-                        "fmla   v10.4s, %25.4s, v0.s[2]     \n"
-                        "fmla   v11.4s, %25.4s, v1.s[0]     \n"
-                        "fmla   v12.4s, %25.4s, v1.s[2]     \n"
-                        "fmla   v13.4s, %25.4s, v4.s[0]     \n"
-
-                        "st1    {v6.4s, v7.4s, v8.4s, v9.4s}, [%0], #64 \n"
-                        "st1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00_0), // %8
-                        "w"(_k01_0), // %9
-                        "w"(_k02_0), // %10
-                        "w"(_k10_0), // %11
-                        "w"(_k11_0), // %12
-                        "w"(_k12_0), // %13
-                        "w"(_k20_0), // %14
-                        "w"(_k21_0), // %15
-                        "w"(_k22_0), // %16
-                        "w"(_k00_1), // %17
-                        "w"(_k01_1), // %18
-                        "w"(_k02_1), // %19
-                        "w"(_k10_1), // %20
-                        "w"(_k11_1), // %21
-                        "w"(_k12_1), // %22
-                        "w"(_k20_1), // %23
-                        "w"(_k21_1), // %24
-                        "w"(_k22_1)  // %25
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13");
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    asm volatile(
-                        // r0
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1], #8           \n"
-
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%0] \n" // sum0 sum1
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v10.4s, %8.4s, v0.s[0]      \n"
-                        "fmla   v11.4s, %8.4s, v0.s[2]      \n"
-                        "fmla   v12.4s, %17.4s, v0.s[0]     \n"
-                        "fmla   v13.4s, %17.4s, v0.s[2]     \n"
-
-                        "ld1    {v1.h}[0], [%1]             \n"
-
-                        "fmla   v10.4s, %9.4s, v0.s[1]      \n"
-                        "fmla   v11.4s, %9.4s, v0.s[3]      \n"
-                        "fmla   v12.4s, %18.4s, v0.s[1]     \n"
-                        "fmla   v13.4s, %18.4s, v0.s[3]     \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        // r1
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v2.4h}, [%2], #8           \n"
-
-                        "fmla   v10.4s, %10.4s, v0.s[2]     \n"
-                        "fmla   v11.4s, %10.4s, v1.s[0]     \n"
-                        "fmla   v12.4s, %19.4s, v0.s[2]     \n"
-                        "fmla   v13.4s, %19.4s, v1.s[0]     \n"
-
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v10.4s, %11.4s, v2.s[0]     \n"
-                        "fmla   v11.4s, %11.4s, v2.s[2]     \n"
-                        "fmla   v12.4s, %20.4s, v2.s[0]     \n"
-                        "fmla   v13.4s, %20.4s, v2.s[2]     \n"
-
-                        "ld1    {v3.h}[0], [%2]             \n"
-
-                        "fmla   v10.4s, %12.4s, v2.s[1]     \n"
-                        "fmla   v11.4s, %12.4s, v2.s[3]     \n"
-                        "fmla   v12.4s, %21.4s, v2.s[1]     \n"
-                        "fmla   v13.4s, %21.4s, v2.s[3]     \n"
-
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        // r2
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v0.4h}, [%3], #8           \n"
-
-                        "fmla   v10.4s, %13.4s, v2.s[2]     \n"
-                        "fmla   v11.4s, %13.4s, v3.s[0]     \n"
-                        "fmla   v12.4s, %22.4s, v2.s[2]     \n"
-                        "fmla   v13.4s, %22.4s, v3.s[0]     \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v10.4s, %14.4s, v0.s[0]     \n"
-                        "fmla   v11.4s, %14.4s, v0.s[2]     \n"
-                        "fmla   v12.4s, %23.4s, v0.s[0]     \n"
-                        "fmla   v13.4s, %23.4s, v0.s[2]     \n"
-
-                        "ld1    {v1.h}[0], [%3]             \n"
-
-                        "fmla   v10.4s, %15.4s, v0.s[1]     \n"
-                        "fmla   v11.4s, %15.4s, v0.s[3]     \n"
-                        "fmla   v12.4s, %24.4s, v0.s[1]     \n"
-                        "fmla   v13.4s, %24.4s, v0.s[3]     \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v10.4s, %16.4s, v0.s[2]     \n"
-                        "fmla   v11.4s, %16.4s, v1.s[0]     \n"
-                        "fmla   v12.4s, %25.4s, v0.s[2]     \n"
-                        "fmla   v13.4s, %25.4s, v1.s[0]     \n"
-
-                        "st1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00_0), // %8
-                        "w"(_k01_0), // %9
-                        "w"(_k02_0), // %10
-                        "w"(_k10_0), // %11
-                        "w"(_k11_0), // %12
-                        "w"(_k12_0), // %13
-                        "w"(_k20_0), // %14
-                        "w"(_k21_0), // %15
-                        "w"(_k22_0), // %16
-                        "w"(_k00_1), // %17
-                        "w"(_k01_1), // %18
-                        "w"(_k02_1), // %19
-                        "w"(_k10_1), // %20
-                        "w"(_k11_1), // %21
-                        "w"(_k12_1), // %22
-                        "w"(_k20_1), // %23
-                        "w"(_k21_1), // %24
-                        "w"(_k22_1)  // %25
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13");
-                }
-                for (; j < outw; j++)
-                {
-                    float32x4_t _sum0 = vld1q_f32(outptr0);
-                    float32x4_t _sum1 = vld1q_f32(outptr0 + 4);
-
-                    float32x4_t _r0 = bfloat2float(vld1_u16(r0));
-                    float32x4_t _r1 = bfloat2float(vld1_u16(r1));
-                    float32x4_t _r2 = bfloat2float(vld1_u16(r2));
-
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k00_0, _r0, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k01_0, _r0, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k02_0, _r0, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k10_0, _r1, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k11_0, _r1, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k12_0, _r1, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k20_0, _r2, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k21_0, _r2, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k22_0, _r2, 2);
-
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k00_1, _r0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k01_1, _r0, 1);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k02_1, _r0, 2);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k10_1, _r1, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k11_1, _r1, 1);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k12_1, _r1, 2);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k20_1, _r2, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k21_1, _r2, 1);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k22_1, _r2, 2);
-
-                    vst1q_f32(outptr0, _sum0);
-                    vst1q_f32(outptr0 + 4, _sum1);
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0 += 8;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * 4;
-            k1 += 9 * 4;
-        }
-        for (; q < inch; q++)
-        {
-            unsigned short* outptr0_bf16 = top_blob.channel(p);
-            unsigned short* outptr1_bf16 = top_blob.channel(p + 1);
-
-            const float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<const unsigned short>(0);
-            const unsigned short* r1 = img0.row<const unsigned short>(1);
-            const unsigned short* r2 = img0.row<const unsigned short>(2);
-
-            float32x4_t _k00_0 = bfloat2float(vld1_u16(k0));
-            float32x4_t _k01_0 = bfloat2float(vld1_u16(k0 + 4));
-            float32x4_t _k02_0 = bfloat2float(vld1_u16(k0 + 8));
-            float32x4_t _k10_0 = bfloat2float(vld1_u16(k0 + 12));
-            float32x4_t _k11_0 = bfloat2float(vld1_u16(k0 + 16));
-            float32x4_t _k12_0 = bfloat2float(vld1_u16(k0 + 20));
-            float32x4_t _k20_0 = bfloat2float(vld1_u16(k0 + 24));
-            float32x4_t _k21_0 = bfloat2float(vld1_u16(k0 + 28));
-            float32x4_t _k22_0 = bfloat2float(vld1_u16(k0 + 32));
-
-            float32x4_t _k00_1 = bfloat2float(vld1_u16(k1));
-            float32x4_t _k01_1 = bfloat2float(vld1_u16(k1 + 4));
-            float32x4_t _k02_1 = bfloat2float(vld1_u16(k1 + 8));
-            float32x4_t _k10_1 = bfloat2float(vld1_u16(k1 + 12));
-            float32x4_t _k11_1 = bfloat2float(vld1_u16(k1 + 16));
-            float32x4_t _k12_1 = bfloat2float(vld1_u16(k1 + 20));
-            float32x4_t _k20_1 = bfloat2float(vld1_u16(k1 + 24));
-            float32x4_t _k21_1 = bfloat2float(vld1_u16(k1 + 28));
-            float32x4_t _k22_1 = bfloat2float(vld1_u16(k1 + 32));
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-
-                for (; j + 3 < outw; j += 4)
-                {
-                    asm volatile(
-                        // r0
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%3], #16   \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v6.4s, v7.4s, v8.4s, v9.4s}, [%2], #64 \n" // sum0
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%2], #64 \n" // sum1
-
-                        "fmla   v6.4s, %12.4s, v0.s[0]      \n"
-                        "fmla   v7.4s, %12.4s, v0.s[2]      \n"
-                        "fmla   v8.4s, %12.4s, v1.s[0]      \n"
-                        "fmla   v9.4s, %12.4s, v1.s[2]      \n"
-                        "fmla   v10.4s, %21.4s, v0.s[0]     \n"
-                        "fmla   v11.4s, %21.4s, v0.s[2]     \n"
-                        "fmla   v12.4s, %21.4s, v1.s[0]     \n"
-                        "fmla   v13.4s, %21.4s, v1.s[2]     \n"
-
-                        "ld1    {v4.h}[0], [%3]             \n"
-
-                        "fmla   v6.4s, %13.4s, v0.s[1]      \n"
-                        "fmla   v7.4s, %13.4s, v0.s[3]      \n"
-                        "fmla   v8.4s, %13.4s, v1.s[1]      \n"
-                        "fmla   v9.4s, %13.4s, v1.s[3]      \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "fmla   v10.4s, %22.4s, v0.s[1]     \n"
-                        "fmla   v11.4s, %22.4s, v0.s[3]     \n"
-                        "fmla   v12.4s, %22.4s, v1.s[1]     \n"
-                        "fmla   v13.4s, %22.4s, v1.s[3]     \n"
-
-                        // r1
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v2.4h, v3.4h}, [%4], #16   \n"
-
-                        "fmla   v6.4s, %14.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %14.4s, v1.s[0]      \n"
-                        "fmla   v8.4s, %14.4s, v1.s[2]      \n"
-                        "fmla   v9.4s, %14.4s, v4.s[0]      \n"
-
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v10.4s, %23.4s, v0.s[2]     \n"
-                        "fmla   v11.4s, %23.4s, v1.s[0]     \n"
-                        "fmla   v12.4s, %23.4s, v1.s[2]     \n"
-                        "fmla   v13.4s, %23.4s, v4.s[0]     \n"
-
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v6.4s, %15.4s, v2.s[0]      \n"
-                        "fmla   v7.4s, %15.4s, v2.s[2]      \n"
-                        "fmla   v8.4s, %15.4s, v3.s[0]      \n"
-                        "fmla   v9.4s, %15.4s, v3.s[2]      \n"
-                        "fmla   v10.4s, %24.4s, v2.s[0]     \n"
-                        "fmla   v11.4s, %24.4s, v2.s[2]     \n"
-                        "fmla   v12.4s, %24.4s, v3.s[0]     \n"
-                        "fmla   v13.4s, %24.4s, v3.s[2]     \n"
-
-                        "ld1    {v5.h}[0], [%4]             \n"
-
-                        "fmla   v6.4s, %16.4s, v2.s[1]      \n"
-                        "fmla   v7.4s, %16.4s, v2.s[3]      \n"
-                        "fmla   v8.4s, %16.4s, v3.s[1]      \n"
-                        "fmla   v9.4s, %16.4s, v3.s[3]      \n"
-
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v10.4s, %25.4s, v2.s[1]     \n"
-                        "fmla   v11.4s, %25.4s, v2.s[3]     \n"
-                        "fmla   v12.4s, %25.4s, v3.s[1]     \n"
-                        "fmla   v13.4s, %25.4s, v3.s[3]     \n"
-
-                        // r2
-                        "prfm   pldl1keep, [%5, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%5], #16   \n"
-
-                        "fmla   v6.4s, %17.4s, v2.s[2]      \n"
-                        "fmla   v7.4s, %17.4s, v3.s[0]      \n"
-                        "fmla   v8.4s, %17.4s, v3.s[2]      \n"
-                        "fmla   v9.4s, %17.4s, v5.s[0]      \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v10.4s, %26.4s, v2.s[2]     \n"
-                        "fmla   v11.4s, %26.4s, v3.s[0]     \n"
-                        "fmla   v12.4s, %26.4s, v3.s[2]     \n"
-                        "fmla   v13.4s, %26.4s, v5.s[0]     \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v6.4s, %18.4s, v0.s[0]      \n"
-                        "fmla   v7.4s, %18.4s, v0.s[2]      \n"
-                        "fmla   v8.4s, %18.4s, v1.s[0]      \n"
-                        "fmla   v9.4s, %18.4s, v1.s[2]      \n"
-                        "fmla   v10.4s, %27.4s, v0.s[0]     \n"
-                        "fmla   v11.4s, %27.4s, v0.s[2]     \n"
-                        "fmla   v12.4s, %27.4s, v1.s[0]     \n"
-                        "fmla   v13.4s, %27.4s, v1.s[2]     \n"
-
-                        "ld1    {v4.h}[0], [%5]             \n"
-
-                        "fmla   v6.4s, %19.4s, v0.s[1]      \n"
-                        "fmla   v7.4s, %19.4s, v0.s[3]      \n"
-                        "fmla   v8.4s, %19.4s, v1.s[1]      \n"
-                        "fmla   v9.4s, %19.4s, v1.s[3]      \n"
-                        "fmla   v10.4s, %28.4s, v0.s[1]     \n"
-                        "fmla   v11.4s, %28.4s, v0.s[3]     \n"
-                        "fmla   v12.4s, %28.4s, v1.s[1]     \n"
-                        "fmla   v13.4s, %28.4s, v1.s[3]     \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "fmla   v6.4s, %20.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %20.4s, v1.s[0]      \n"
-                        "fmla   v8.4s, %20.4s, v1.s[2]      \n"
-                        "fmla   v9.4s, %20.4s, v4.s[0]      \n"
-                        "fmla   v10.4s, %29.4s, v0.s[2]     \n"
-                        "fmla   v11.4s, %29.4s, v1.s[0]     \n"
-                        "fmla   v12.4s, %29.4s, v1.s[2]     \n"
-                        "fmla   v13.4s, %29.4s, v4.s[0]     \n"
-
-                        "shrn   v6.4h, v6.4s, #16           \n"
-                        "shrn   v7.4h, v7.4s, #16           \n"
-                        "shrn   v8.4h, v8.4s, #16           \n"
-                        "shrn   v9.4h, v9.4s, #16           \n"
-                        "shrn   v10.4h, v10.4s, #16         \n"
-                        "shrn   v11.4h, v11.4s, #16         \n"
-
-                        "st1    {v6.4h, v7.4h, v8.4h, v9.4h}, [%0], #32 \n"
-
-                        "shrn   v12.4h, v12.4s, #16         \n"
-                        "shrn   v13.4h, v13.4s, #16         \n"
-
-                        "st1    {v10.4h, v11.4h, v12.4h, v13.4h}, [%1], #32 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr1_bf16), // %1
-                        "=r"(outptr0),      // %2
-                        "=r"(r0),           // %3
-                        "=r"(r1),           // %4
-                        "=r"(r2)            // %5
-                        : "0"(outptr0_bf16),
-                        "1"(outptr1_bf16),
-                        "2"(outptr0),
-                        "3"(r0),
-                        "4"(r1),
-                        "5"(r2),
-                        "w"(_k00_0), // %12
-                        "w"(_k01_0), // %13
-                        "w"(_k02_0), // %14
-                        "w"(_k10_0), // %15
-                        "w"(_k11_0), // %16
-                        "w"(_k12_0), // %17
-                        "w"(_k20_0), // %18
-                        "w"(_k21_0), // %19
-                        "w"(_k22_0), // %20
-                        "w"(_k00_1), // %21
-                        "w"(_k01_1), // %22
-                        "w"(_k02_1), // %23
-                        "w"(_k10_1), // %24
-                        "w"(_k11_1), // %25
-                        "w"(_k12_1), // %26
-                        "w"(_k20_1), // %27
-                        "w"(_k21_1), // %28
-                        "w"(_k22_1)  // %29
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13");
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    asm volatile(
-                        // r0
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v0.4h}, [%3], #8           \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%2], #64 \n" // sum0 sum1
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v10.4s, %12.4s, v0.s[0]     \n"
-                        "fmla   v11.4s, %12.4s, v0.s[2]     \n"
-                        "fmla   v12.4s, %21.4s, v0.s[0]     \n"
-                        "fmla   v13.4s, %21.4s, v0.s[2]     \n"
-
-                        "ld1    {v1.h}[0], [%3]             \n"
-
-                        "fmla   v10.4s, %13.4s, v0.s[1]     \n"
-                        "fmla   v11.4s, %13.4s, v0.s[3]     \n"
-                        "fmla   v12.4s, %22.4s, v0.s[1]     \n"
-                        "fmla   v13.4s, %22.4s, v0.s[3]     \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        // r1
-                        "prfm   pldl1keep, [%4, #64]        \n"
-                        "ld1    {v2.4h}, [%4], #8           \n"
-
-                        "fmla   v10.4s, %14.4s, v0.s[2]     \n"
-                        "fmla   v11.4s, %14.4s, v1.s[0]     \n"
-                        "fmla   v12.4s, %23.4s, v0.s[2]     \n"
-                        "fmla   v13.4s, %23.4s, v1.s[0]     \n"
-
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v10.4s, %15.4s, v2.s[0]     \n"
-                        "fmla   v11.4s, %15.4s, v2.s[2]     \n"
-                        "fmla   v12.4s, %24.4s, v2.s[0]     \n"
-                        "fmla   v13.4s, %24.4s, v2.s[2]     \n"
-
-                        "ld1    {v3.h}[0], [%4]             \n"
-
-                        "fmla   v10.4s, %16.4s, v2.s[1]     \n"
-                        "fmla   v11.4s, %16.4s, v2.s[3]     \n"
-                        "fmla   v12.4s, %25.4s, v2.s[1]     \n"
-                        "fmla   v13.4s, %25.4s, v2.s[3]     \n"
-
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        // r2
-                        "prfm   pldl1keep, [%5, #64]        \n"
-                        "ld1    {v0.4h}, [%5], #8           \n"
-
-                        "fmla   v10.4s, %17.4s, v2.s[2]     \n"
-                        "fmla   v11.4s, %17.4s, v3.s[0]     \n"
-                        "fmla   v12.4s, %26.4s, v2.s[2]     \n"
-                        "fmla   v13.4s, %26.4s, v3.s[0]     \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v10.4s, %18.4s, v0.s[0]     \n"
-                        "fmla   v11.4s, %18.4s, v0.s[2]     \n"
-                        "fmla   v12.4s, %27.4s, v0.s[0]     \n"
-                        "fmla   v13.4s, %27.4s, v0.s[2]     \n"
-
-                        "ld1    {v1.h}[0], [%5]             \n"
-
-                        "fmla   v10.4s, %19.4s, v0.s[1]     \n"
-                        "fmla   v11.4s, %19.4s, v0.s[3]     \n"
-                        "fmla   v12.4s, %28.4s, v0.s[1]     \n"
-                        "fmla   v13.4s, %28.4s, v0.s[3]     \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v10.4s, %20.4s, v0.s[2]     \n"
-                        "fmla   v11.4s, %20.4s, v1.s[0]     \n"
-                        "fmla   v12.4s, %29.4s, v0.s[2]     \n"
-                        "fmla   v13.4s, %29.4s, v1.s[0]     \n"
-
-                        "shrn   v10.4h, v10.4s, #16         \n"
-                        "shrn   v11.4h, v11.4s, #16         \n"
-                        "shrn   v12.4h, v12.4s, #16         \n"
-                        "shrn   v13.4h, v13.4s, #16         \n"
-
-                        "st1    {v10.4h, v11.4h}, [%0], #16 \n"
-                        "st1    {v12.4h, v13.4h}, [%1], #16 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr1_bf16), // %1
-                        "=r"(outptr0),      // %2
-                        "=r"(r0),           // %3
-                        "=r"(r1),           // %4
-                        "=r"(r2)            // %5
-                        : "0"(outptr0_bf16),
-                        "1"(outptr1_bf16),
-                        "2"(outptr0),
-                        "3"(r0),
-                        "4"(r1),
-                        "5"(r2),
-                        "w"(_k00_0), // %12
-                        "w"(_k01_0), // %13
-                        "w"(_k02_0), // %14
-                        "w"(_k10_0), // %15
-                        "w"(_k11_0), // %16
-                        "w"(_k12_0), // %17
-                        "w"(_k20_0), // %18
-                        "w"(_k21_0), // %19
-                        "w"(_k22_0), // %20
-                        "w"(_k00_1), // %21
-                        "w"(_k01_1), // %22
-                        "w"(_k02_1), // %23
-                        "w"(_k10_1), // %24
-                        "w"(_k11_1), // %25
-                        "w"(_k12_1), // %26
-                        "w"(_k20_1), // %27
-                        "w"(_k21_1), // %28
-                        "w"(_k22_1)  // %29
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13");
-                }
-                for (; j < outw; j++)
-                {
-                    float32x4_t _sum0 = vld1q_f32(outptr0);
-                    float32x4_t _sum1 = vld1q_f32(outptr0 + 4);
-
-                    float32x4_t _r0 = bfloat2float(vld1_u16(r0));
-                    float32x4_t _r1 = bfloat2float(vld1_u16(r1));
-                    float32x4_t _r2 = bfloat2float(vld1_u16(r2));
-
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k00_0, _r0, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k01_0, _r0, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k02_0, _r0, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k10_0, _r1, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k11_0, _r1, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k12_0, _r1, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k20_0, _r2, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k21_0, _r2, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k22_0, _r2, 2);
-
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k00_1, _r0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k01_1, _r0, 1);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k02_1, _r0, 2);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k10_1, _r1, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k11_1, _r1, 1);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k12_1, _r1, 2);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k20_1, _r2, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k21_1, _r2, 1);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _k22_1, _r2, 2);
-
-                    vst1_u16(outptr0_bf16, float2bfloat(_sum0));
-                    vst1_u16(outptr1_bf16, float2bfloat(_sum1));
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0 += 8;
-                    outptr0_bf16 += 4;
-                    outptr1_bf16 += 4;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * 4;
-            k1 += 9 * 4;
-        }
-    }
-#endif // __ARM_NEON && __aarch64__
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out0 = top_blob_fp32.channel(get_omp_thread_num());
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        out0.fill(_bias0);
-
-        const unsigned short* k0 = kernel.channel(p);
-
-        int q = 0;
-        for (; q < inch - 1; q++)
-        {
-            float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<const unsigned short>(0);
-            const unsigned short* r1 = img0.row<const unsigned short>(1);
-            const unsigned short* r2 = img0.row<const unsigned short>(2);
-
-            float32x4_t _k00 = bfloat2float(vld1_u16(k0));
-            float32x4_t _k01 = bfloat2float(vld1_u16(k0 + 4));
-            float32x4_t _k02 = bfloat2float(vld1_u16(k0 + 8));
-            float32x4_t _k10 = bfloat2float(vld1_u16(k0 + 12));
-            float32x4_t _k11 = bfloat2float(vld1_u16(k0 + 16));
-            float32x4_t _k12 = bfloat2float(vld1_u16(k0 + 20));
-            float32x4_t _k20 = bfloat2float(vld1_u16(k0 + 24));
-            float32x4_t _k21 = bfloat2float(vld1_u16(k0 + 28));
-            float32x4_t _k22 = bfloat2float(vld1_u16(k0 + 32));
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        // r0
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%1], #16   \n"
-
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v6.4s, v7.4s, v8.4s, v9.4s}, [%0] \n" // sum0
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v6.4s, %8.4s, v0.s[0]       \n"
-                        "fmla   v7.4s, %8.4s, v0.s[2]       \n"
-                        "fmla   v8.4s, %8.4s, v1.s[0]       \n"
-                        "fmla   v9.4s, %8.4s, v1.s[2]       \n"
-
-                        "ld1    {v4.h}[0], [%1]             \n"
-
-                        "fmla   v6.4s, %9.4s, v0.s[1]       \n"
-                        "fmla   v7.4s, %9.4s, v0.s[3]       \n"
-                        "fmla   v8.4s, %9.4s, v1.s[1]       \n"
-                        "fmla   v9.4s, %9.4s, v1.s[3]       \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        // r1
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v2.4h, v3.4h}, [%2], #16   \n"
-
-                        "fmla   v6.4s, %10.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %10.4s, v1.s[0]      \n"
-                        "fmla   v8.4s, %10.4s, v1.s[2]      \n"
-                        "fmla   v9.4s, %10.4s, v4.s[0]      \n"
-
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v6.4s, %11.4s, v2.s[0]      \n"
-                        "fmla   v7.4s, %11.4s, v2.s[2]      \n"
-                        "fmla   v8.4s, %11.4s, v3.s[0]      \n"
-                        "fmla   v9.4s, %11.4s, v3.s[2]      \n"
-
-                        "ld1    {v5.h}[0], [%2]             \n"
-
-                        "fmla   v6.4s, %12.4s, v2.s[1]      \n"
-                        "fmla   v7.4s, %12.4s, v2.s[3]      \n"
-                        "fmla   v8.4s, %12.4s, v3.s[1]      \n"
-                        "fmla   v9.4s, %12.4s, v3.s[3]      \n"
-
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        // r2
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%3], #16   \n"
-
-                        "fmla   v6.4s, %13.4s, v2.s[2]      \n"
-                        "fmla   v7.4s, %13.4s, v3.s[0]      \n"
-                        "fmla   v8.4s, %13.4s, v3.s[2]      \n"
-                        "fmla   v9.4s, %13.4s, v5.s[0]      \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v6.4s, %14.4s, v0.s[0]      \n"
-                        "fmla   v7.4s, %14.4s, v0.s[2]      \n"
-                        "fmla   v8.4s, %14.4s, v1.s[0]      \n"
-                        "fmla   v9.4s, %14.4s, v1.s[2]      \n"
-
-                        "ld1    {v4.h}[0], [%3]             \n"
-
-                        "fmla   v6.4s, %15.4s, v0.s[1]      \n"
-                        "fmla   v7.4s, %15.4s, v0.s[3]      \n"
-                        "fmla   v8.4s, %15.4s, v1.s[1]      \n"
-                        "fmla   v9.4s, %15.4s, v1.s[3]      \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "fmla   v6.4s, %16.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %16.4s, v1.s[0]      \n"
-                        "fmla   v8.4s, %16.4s, v1.s[2]      \n"
-                        "fmla   v9.4s, %16.4s, v4.s[0]      \n"
-
-                        "st1    {v6.4s, v7.4s, v8.4s, v9.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9");
-#else  // __aarch64__
-                    asm volatile(
-                        // r0
-                        "pld        [%1, #128]          \n"
-                        "vld1.u16   {d12-d13}, [%1]!    \n"
-
-                        "pld        [%0, #512]          \n"
-                        "vldm       %0, {d0-d7}         \n" // sum0
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-
-                        "vld1.u16   {d12[0]}, [%1]      \n"
-
-                        "vmla.f32   q0, %q8, d8[0]      \n"
-                        "vmla.f32   q1, %q8, d9[0]      \n"
-
-                        "vmla.f32   q2, %q8, d10[0]     \n"
-                        "vmla.f32   q3, %q8, d11[0]     \n"
-
-                        "vmla.f32   q0, %q9, d8[1]      \n"
-                        "vmla.f32   q1, %q9, d9[1]      \n"
-
-                        "vshl.u32   d8, d12, #16        \n"
-
-                        "vmla.f32   q2, %q9, d10[1]     \n"
-                        "vmla.f32   q3, %q9, d11[1]     \n"
-
-                        // r1
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d12-d13}, [%2]!    \n"
-
-                        "vmla.f32   q0, %q10, d9[0]     \n"
-                        "vmla.f32   q1, %q10, d10[0]    \n"
-                        "vmla.f32   q2, %q10, d11[0]    \n"
-                        "vmla.f32   q3, %q10, d8[0]     \n"
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-
-                        "vld1.u16   {d12[0]}, [%2]      \n"
-
-                        "vmla.f32   q0, %q11, d8[0]     \n"
-                        "vmla.f32   q1, %q11, d9[0]     \n"
-
-                        "vmla.f32   q2, %q11, d10[0]    \n"
-                        "vmla.f32   q3, %q11, d11[0]    \n"
-
-                        "vmla.f32   q0, %q12, d8[1]     \n"
-                        "vmla.f32   q1, %q12, d9[1]     \n"
-
-                        "vshl.u32   d8, d12, #16        \n"
-
-                        "vmla.f32   q2, %q12, d10[1]    \n"
-                        "vmla.f32   q3, %q12, d11[1]    \n"
-
-                        // r2
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d12-d13}, [%3]!    \n"
-
-                        "vmla.f32   q0, %q13, d9[0]     \n"
-                        "vmla.f32   q1, %q13, d10[0]    \n"
-                        "vmla.f32   q2, %q13, d11[0]    \n"
-                        "vmla.f32   q3, %q13, d8[0]     \n"
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-
-                        "vld1.u16   {d12[0]}, [%3]      \n"
-
-                        "vmla.f32   q0, %q14, d8[0]     \n"
-                        "vmla.f32   q1, %q14, d9[0]     \n"
-
-                        "vmla.f32   q2, %q14, d10[0]    \n"
-                        "vmla.f32   q3, %q14, d11[0]    \n"
-
-                        "vmla.f32   q0, %q15, d8[1]     \n"
-                        "vmla.f32   q1, %q15, d9[1]     \n"
-
-                        "vshl.u32   d8, d12, #16        \n"
-
-                        "vmla.f32   q2, %q15, d10[1]    \n"
-                        "vmla.f32   q3, %q15, d11[1]    \n"
-
-                        "vmla.f32   q0, %q16, d9[0]     \n"
-                        "vmla.f32   q1, %q16, d10[0]    \n"
-                        "vmla.f32   q2, %q16, d11[0]    \n"
-                        "vmla.f32   q3, %q16, d8[0]     \n"
-
-                        "vstm       %0!, {d0-d7}        \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        // r0
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1], #8           \n"
-
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%0]        \n" // sum0
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmul   v6.4s, %8.4s, v0.s[0]       \n"
-                        "fmul   v7.4s, %8.4s, v0.s[2]       \n"
-
-                        "ld1    {v1.h}[0], [%1]             \n"
-
-                        "fmla   v8.4s, %9.4s, v0.s[1]       \n"
-                        "fmla   v9.4s, %9.4s, v0.s[3]       \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        // r1
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v2.4h}, [%2], #8           \n"
-
-                        "fmla   v6.4s, %10.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %10.4s, v1.s[0]      \n"
-
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v8.4s, %11.4s, v2.s[0]      \n"
-                        "fmla   v9.4s, %11.4s, v2.s[2]      \n"
-
-                        "ld1    {v3.h}[0], [%2]             \n"
-
-                        "fmla   v6.4s, %12.4s, v2.s[1]      \n"
-                        "fmla   v7.4s, %12.4s, v2.s[3]      \n"
-
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        // r2
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v0.4h}, [%3], #8           \n"
-
-                        "fmla   v8.4s, %13.4s, v2.s[2]      \n"
-                        "fmla   v9.4s, %13.4s, v3.s[0]      \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v6.4s, %14.4s, v0.s[0]      \n"
-                        "fmla   v7.4s, %14.4s, v0.s[2]      \n"
-
-                        "ld1    {v1.h}[0], [%3]             \n"
-
-                        "fmla   v8.4s, %15.4s, v0.s[1]      \n"
-                        "fmla   v9.4s, %15.4s, v0.s[3]      \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v6.4s, %16.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %16.4s, v1.s[0]      \n"
-
-                        "fadd   v8.4s, v8.4s, v6.4s         \n"
-                        "fadd   v9.4s, v9.4s, v7.4s         \n"
-
-                        "st1    {v8.4s, v9.4s}, [%0], #32   \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9");
-#else  // __aarch64__
-                    asm volatile(
-                        // r0
-                        "pld        [%1, #64]           \n"
-                        "vld1.u16   {d9}, [%1]!         \n"
-
-                        "pld        [%0, #256]          \n"
-                        "vld1.f32   {d4-d7}, [%0]       \n" // sum0
-
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmul.f32   q0, %q8, d8[0]      \n"
-                        "vmul.f32   q1, %q8, d9[0]      \n"
-
-                        "vld1.u16   {d11[]}, [%1]       \n"
-
-                        "vmla.f32   q2, %q9, d8[1]      \n"
-                        "vmla.f32   q3, %q9, d9[1]      \n"
-
-                        "vshll.u16  q5, d11, #16        \n"
-
-                        // r1
-                        "pld        [%2, #64]           \n"
-                        "vld1.u16   {d13}, [%2]!        \n"
-
-                        "vmla.f32   q0, %q10, d9[0]     \n"
-                        "vmla.f32   q1, %q10, d10[0]    \n"
-
-                        "vshll.u16  q6, d13, #16        \n"
-
-                        "vmla.f32   q2, %q11, d12[0]    \n"
-                        "vmla.f32   q3, %q11, d13[0]    \n"
-
-                        "vld1.u16   {d9[]}, [%2]        \n"
-
-                        "vmla.f32   q0, %q12, d12[1]    \n"
-                        "vmla.f32   q1, %q12, d13[1]    \n"
-
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        // r2
-                        "pld        [%3, #64]           \n"
-                        "vld1.u16   {d11}, [%3]!        \n"
-
-                        "vmla.f32   q2, %q13, d13[0]    \n"
-                        "vmla.f32   q3, %q13, d8[0]     \n"
-
-                        "vshll.u16  q5, d11, #16        \n"
-
-                        "vmla.f32   q0, %q14, d10[0]    \n"
-                        "vmla.f32   q1, %q14, d11[0]    \n"
-
-                        "vld1.u16   {d13[]}, [%3]       \n"
-
-                        "vmla.f32   q2, %q15, d10[1]    \n"
-                        "vmla.f32   q3, %q15, d11[1]    \n"
-
-                        "vshll.u16  q6, d13, #16        \n"
-
-                        "vmla.f32   q0, %q16, d11[0]    \n"
-                        "vmla.f32   q1, %q16, d12[0]    \n"
-
-                        "vadd.f32   q2, q2, q0          \n"
-                        "vadd.f32   q3, q3, q1          \n"
-
-                        "vst1.f32   {d4-d7}, [%0]!      \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-                    float32x4_t _sum0 = vld1q_f32(outptr0);
-
-                    float32x4_t _r0 = bfloat2float(vld1_u16(r0));
-                    float32x4_t _r1 = bfloat2float(vld1_u16(r1));
-                    float32x4_t _r2 = bfloat2float(vld1_u16(r2));
-
-#if __aarch64__
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k00, _r0, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k01, _r0, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k02, _r0, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k10, _r1, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k11, _r1, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k12, _r1, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k20, _r2, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k21, _r2, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k22, _r2, 2);
-#else
-                    _sum0 = vmlaq_lane_f32(_sum0, _k00, vget_low_f32(_r0), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k01, vget_low_f32(_r0), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k02, vget_high_f32(_r0), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k10, vget_low_f32(_r1), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k11, vget_low_f32(_r1), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k12, vget_high_f32(_r1), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k20, vget_low_f32(_r2), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k21, vget_low_f32(_r2), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k22, vget_high_f32(_r2), 0);
-#endif
-
-                    vst1q_f32(outptr0, _sum0);
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0 += 4;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * 4;
-        }
-        for (; q < inch; q++)
-        {
-            unsigned short* outptr0_bf16 = top_blob.channel(p);
-
-            const float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<const unsigned short>(0);
-            const unsigned short* r1 = img0.row<const unsigned short>(1);
-            const unsigned short* r2 = img0.row<const unsigned short>(2);
-
-            float32x4_t _k00 = bfloat2float(vld1_u16(k0));
-            float32x4_t _k01 = bfloat2float(vld1_u16(k0 + 4));
-            float32x4_t _k02 = bfloat2float(vld1_u16(k0 + 8));
-            float32x4_t _k10 = bfloat2float(vld1_u16(k0 + 12));
-            float32x4_t _k11 = bfloat2float(vld1_u16(k0 + 16));
-            float32x4_t _k12 = bfloat2float(vld1_u16(k0 + 20));
-            float32x4_t _k20 = bfloat2float(vld1_u16(k0 + 24));
-            float32x4_t _k21 = bfloat2float(vld1_u16(k0 + 28));
-            float32x4_t _k22 = bfloat2float(vld1_u16(k0 + 32));
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        // r0
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%2], #16   \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v6.4s, v7.4s, v8.4s, v9.4s}, [%1], #64 \n" // sum0
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v6.4s, %10.4s, v0.s[0]      \n"
-                        "fmla   v7.4s, %10.4s, v0.s[2]      \n"
-                        "fmla   v8.4s, %10.4s, v1.s[0]      \n"
-                        "fmla   v9.4s, %10.4s, v1.s[2]      \n"
-
-                        "ld1    {v4.h}[0], [%2]             \n"
-
-                        "fmla   v6.4s, %11.4s, v0.s[1]      \n"
-                        "fmla   v7.4s, %11.4s, v0.s[3]      \n"
-                        "fmla   v8.4s, %11.4s, v1.s[1]      \n"
-                        "fmla   v9.4s, %11.4s, v1.s[3]      \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        // r1
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v2.4h, v3.4h}, [%3], #16   \n"
-
-                        "fmla   v6.4s, %12.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %12.4s, v1.s[0]      \n"
-                        "fmla   v8.4s, %12.4s, v1.s[2]      \n"
-                        "fmla   v9.4s, %12.4s, v4.s[0]      \n"
-
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v6.4s, %13.4s, v2.s[0]      \n"
-                        "fmla   v7.4s, %13.4s, v2.s[2]      \n"
-                        "fmla   v8.4s, %13.4s, v3.s[0]      \n"
-                        "fmla   v9.4s, %13.4s, v3.s[2]      \n"
-
-                        "ld1    {v5.h}[0], [%3]             \n"
-
-                        "fmla   v6.4s, %14.4s, v2.s[1]      \n"
-                        "fmla   v7.4s, %14.4s, v2.s[3]      \n"
-                        "fmla   v8.4s, %14.4s, v3.s[1]      \n"
-                        "fmla   v9.4s, %14.4s, v3.s[3]      \n"
-
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        // r2
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%4], #16   \n"
-
-                        "fmla   v6.4s, %15.4s, v2.s[2]      \n"
-                        "fmla   v7.4s, %15.4s, v3.s[0]      \n"
-                        "fmla   v8.4s, %15.4s, v3.s[2]      \n"
-                        "fmla   v9.4s, %15.4s, v5.s[0]      \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v6.4s, %16.4s, v0.s[0]      \n"
-                        "fmla   v7.4s, %16.4s, v0.s[2]      \n"
-                        "fmla   v8.4s, %16.4s, v1.s[0]      \n"
-                        "fmla   v9.4s, %16.4s, v1.s[2]      \n"
-
-                        "ld1    {v4.h}[0], [%4]             \n"
-
-                        "fmla   v6.4s, %17.4s, v0.s[1]      \n"
-                        "fmla   v7.4s, %17.4s, v0.s[3]      \n"
-                        "fmla   v8.4s, %17.4s, v1.s[1]      \n"
-                        "fmla   v9.4s, %17.4s, v1.s[3]      \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "fmla   v6.4s, %18.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %18.4s, v1.s[0]      \n"
-                        "fmla   v8.4s, %18.4s, v1.s[2]      \n"
-                        "fmla   v9.4s, %18.4s, v4.s[0]      \n"
-
-                        "shrn   v6.4h, v6.4s, #16           \n"
-                        "shrn   v7.4h, v7.4s, #16           \n"
-                        "shrn   v8.4h, v8.4s, #16           \n"
-                        "shrn   v9.4h, v9.4s, #16           \n"
-
-                        "st1    {v6.4h, v7.4h, v8.4h, v9.4h}, [%0], #32 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2)            // %4
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00), // %10
-                        "w"(_k01), // %11
-                        "w"(_k02), // %12
-                        "w"(_k10), // %13
-                        "w"(_k11), // %14
-                        "w"(_k12), // %15
-                        "w"(_k20), // %16
-                        "w"(_k21), // %17
-                        "w"(_k22)  // %18
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9");
-#else  // __aarch64__
-                    asm volatile(
-                        // r0
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d12-d13}, [%2]!    \n"
-
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1!, {d0-d7}        \n" // sum0
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-
-                        "vld1.u16   {d12[0]}, [%2]      \n"
-
-                        "vmla.f32   q0, %q10, d8[0]     \n"
-                        "vmla.f32   q1, %q10, d9[0]     \n"
-
-                        "vmla.f32   q2, %q10, d10[0]    \n"
-                        "vmla.f32   q3, %q10, d11[0]    \n"
-
-                        "vmla.f32   q0, %q11, d8[1]     \n"
-                        "vmla.f32   q1, %q11, d9[1]     \n"
-
-                        "vshl.u32   d8, d12, #16        \n"
-
-                        "vmla.f32   q2, %q11, d10[1]    \n"
-                        "vmla.f32   q3, %q11, d11[1]    \n"
-
-                        // r1
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d12-d13}, [%3]!    \n"
-
-                        "vmla.f32   q0, %q12, d9[0]     \n"
-                        "vmla.f32   q1, %q12, d10[0]    \n"
-                        "vmla.f32   q2, %q12, d11[0]    \n"
-                        "vmla.f32   q3, %q12, d8[0]     \n"
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-
-                        "vld1.u16   {d12[0]}, [%3]      \n"
-
-                        "vmla.f32   q0, %q13, d8[0]     \n"
-                        "vmla.f32   q1, %q13, d9[0]     \n"
-
-                        "vmla.f32   q2, %q13, d10[0]    \n"
-                        "vmla.f32   q3, %q13, d11[0]    \n"
-
-                        "vmla.f32   q0, %q14, d8[1]     \n"
-                        "vmla.f32   q1, %q14, d9[1]     \n"
-
-                        "vshl.u32   d8, d12, #16        \n"
-
-                        "vmla.f32   q2, %q14, d10[1]    \n"
-                        "vmla.f32   q3, %q14, d11[1]    \n"
-
-                        // r2
-                        "pld        [%4, #128]          \n"
-                        "vld1.u16   {d12-d13}, [%4]!    \n"
-
-                        "vmla.f32   q0, %q15, d9[0]     \n"
-                        "vmla.f32   q1, %q15, d10[0]    \n"
-                        "vmla.f32   q2, %q15, d11[0]    \n"
-                        "vmla.f32   q3, %q15, d8[0]     \n"
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-
-                        "vld1.u16   {d12[0]}, [%4]      \n"
-
-                        "vmla.f32   q0, %q16, d8[0]     \n"
-                        "vmla.f32   q1, %q16, d9[0]     \n"
-
-                        "vmla.f32   q2, %q16, d10[0]    \n"
-                        "vmla.f32   q3, %q16, d11[0]    \n"
-
-                        "vmla.f32   q0, %q17, d8[1]     \n"
-                        "vmla.f32   q1, %q17, d9[1]     \n"
-
-                        "vshl.u32   d8, d12, #16        \n"
-
-                        "vmla.f32   q2, %q17, d10[1]    \n"
-                        "vmla.f32   q3, %q17, d11[1]    \n"
-
-                        "vmla.f32   q0, %q18, d9[0]     \n"
-                        "vmla.f32   q1, %q18, d10[0]    \n"
-                        "vmla.f32   q2, %q18, d11[0]    \n"
-                        "vmla.f32   q3, %q18, d8[0]     \n"
-
-                        "vshrn.u32  d0, q0, #16         \n"
-                        "vshrn.u32  d1, q1, #16         \n"
-                        "vshrn.u32  d2, q2, #16         \n"
-                        "vshrn.u32  d3, q3, #16         \n"
-
-                        "vst1.u16   {d0-d3}, [%0 :64]!  \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2)            // %4
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00), // %10
-                        "w"(_k01), // %11
-                        "w"(_k02), // %12
-                        "w"(_k10), // %13
-                        "w"(_k11), // %14
-                        "w"(_k12), // %15
-                        "w"(_k20), // %16
-                        "w"(_k21), // %17
-                        "w"(_k22)  // %18
-                        : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        // r0
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v0.4h}, [%2], #8           \n"
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%1], #32   \n" // sum0
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmul   v6.4s, %10.4s, v0.s[0]      \n"
-                        "fmul   v7.4s, %10.4s, v0.s[2]      \n"
-
-                        "ld1    {v1.h}[0], [%2]             \n"
-
-                        "fmla   v8.4s, %11.4s, v0.s[1]      \n"
-                        "fmla   v9.4s, %11.4s, v0.s[3]      \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        // r1
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v2.4h}, [%3], #8           \n"
-
-                        "fmla   v6.4s, %12.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %12.4s, v1.s[0]      \n"
-
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v8.4s, %13.4s, v2.s[0]      \n"
-                        "fmla   v9.4s, %13.4s, v2.s[2]      \n"
-
-                        "ld1    {v3.h}[0], [%3]             \n"
-
-                        "fmla   v6.4s, %14.4s, v2.s[1]      \n"
-                        "fmla   v7.4s, %14.4s, v2.s[3]      \n"
-
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        // r2
-                        "prfm   pldl1keep, [%4, #64]        \n"
-                        "ld1    {v0.4h}, [%4], #8           \n"
-
-                        "fmla   v8.4s, %15.4s, v2.s[2]      \n"
-                        "fmla   v9.4s, %15.4s, v3.s[0]      \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "fmla   v6.4s, %16.4s, v0.s[0]      \n"
-                        "fmla   v7.4s, %16.4s, v0.s[2]      \n"
-
-                        "ld1    {v1.h}[0], [%4]             \n"
-
-                        "fmla   v8.4s, %17.4s, v0.s[1]      \n"
-                        "fmla   v9.4s, %17.4s, v0.s[3]      \n"
-
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v6.4s, %18.4s, v0.s[2]      \n"
-                        "fmla   v7.4s, %18.4s, v1.s[0]      \n"
-
-                        "fadd   v8.4s, v8.4s, v6.4s         \n"
-                        "fadd   v9.4s, v9.4s, v7.4s         \n"
-
-                        "shrn   v8.4h, v8.4s, #16           \n"
-                        "shrn   v9.4h, v9.4s, #16           \n"
-
-                        "st1    {v8.4h, v9.4h}, [%0], #16   \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2)            // %4
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00), // %10
-                        "w"(_k01), // %11
-                        "w"(_k02), // %12
-                        "w"(_k10), // %13
-                        "w"(_k11), // %14
-                        "w"(_k12), // %15
-                        "w"(_k20), // %16
-                        "w"(_k21), // %17
-                        "w"(_k22)  // %18
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9");
-#else  // __aarch64__
-                    asm volatile(
-                        // r0
-                        "pld        [%2, #64]           \n"
-                        "vld1.u16   {d9}, [%2]!         \n"
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.f32   {d4-d7}, [%1]!      \n" // sum0
-
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmul.f32   q0, %q10, d8[0]     \n"
-                        "vmul.f32   q1, %q10, d9[0]     \n"
-
-                        "vld1.u16   {d11[]}, [%2]       \n"
-
-                        "vmla.f32   q2, %q11, d8[1]     \n"
-                        "vmla.f32   q3, %q11, d9[1]     \n"
-
-                        "vshll.u16  q5, d11, #16        \n"
-
-                        // r1
-                        "pld        [%3, #64]           \n"
-                        "vld1.u16   {d13}, [%3]!        \n"
-
-                        "vmla.f32   q0, %q12, d9[0]     \n"
-                        "vmla.f32   q1, %q12, d10[0]    \n"
-
-                        "vshll.u16  q6, d13, #16        \n"
-
-                        "vmla.f32   q2, %q13, d12[0]    \n"
-                        "vmla.f32   q3, %q13, d13[0]    \n"
-
-                        "vld1.u16   {d9[]}, [%3]        \n"
-
-                        "vmla.f32   q0, %q14, d12[1]    \n"
-                        "vmla.f32   q1, %q14, d13[1]    \n"
-
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        // r2
-                        "pld        [%4, #64]           \n"
-                        "vld1.u16   {d11}, [%4]!        \n"
-
-                        "vmla.f32   q2, %q15, d13[0]    \n"
-                        "vmla.f32   q3, %q15, d8[0]     \n"
-
-                        "vshll.u16  q5, d11, #16        \n"
-
-                        "vmla.f32   q0, %q16, d10[0]    \n"
-                        "vmla.f32   q1, %q16, d11[0]    \n"
-
-                        "vld1.u16   {d13[]}, [%4]       \n"
-
-                        "vmla.f32   q2, %q17, d10[1]    \n"
-                        "vmla.f32   q3, %q17, d11[1]    \n"
-
-                        "vshll.u16  q6, d13, #16        \n"
-
-                        "vmla.f32   q0, %q18, d11[0]    \n"
-                        "vmla.f32   q1, %q18, d12[0]    \n"
-
-                        "vadd.f32   q2, q2, q0          \n"
-                        "vadd.f32   q3, q3, q1          \n"
-
-                        "vshrn.u32  d2, q2, #16         \n"
-                        "vshrn.u32  d3, q3, #16         \n"
-
-                        "vst1.u16   {d2-d3}, [%0 :64]!  \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2)            // %4
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00), // %10
-                        "w"(_k01), // %11
-                        "w"(_k02), // %12
-                        "w"(_k10), // %13
-                        "w"(_k11), // %14
-                        "w"(_k12), // %15
-                        "w"(_k20), // %16
-                        "w"(_k21), // %17
-                        "w"(_k22)  // %18
-                        : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-                    float32x4_t _sum0 = vld1q_f32(outptr0);
-
-                    float32x4_t _r0 = bfloat2float(vld1_u16(r0));
-                    float32x4_t _r1 = bfloat2float(vld1_u16(r1));
-                    float32x4_t _r2 = bfloat2float(vld1_u16(r2));
-
-#if __aarch64__
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k00, _r0, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k01, _r0, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k02, _r0, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k10, _r1, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k11, _r1, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k12, _r1, 2);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k20, _r2, 0);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k21, _r2, 1);
-                    _sum0 = vfmaq_laneq_f32(_sum0, _k22, _r2, 2);
-#else
-                    _sum0 = vmlaq_lane_f32(_sum0, _k00, vget_low_f32(_r0), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k01, vget_low_f32(_r0), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k02, vget_high_f32(_r0), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k10, vget_low_f32(_r1), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k11, vget_low_f32(_r1), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k12, vget_high_f32(_r1), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k20, vget_low_f32(_r2), 0);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k21, vget_low_f32(_r2), 1);
-                    _sum0 = vmlaq_lane_f32(_sum0, _k22, vget_high_f32(_r2), 0);
-#endif
-
-                    vst1_u16(outptr0_bf16, float2bfloat(_sum0));
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0 += 4;
-                    outptr0_bf16 += 4;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to4_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to4_fp16s.h
deleted file mode 100644
index f8a055cd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to4_fp16s.h
+++ /dev/null
@@ -1,646 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack1to4_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float16x4_t _bias0 = bias ? vld1_f16(bias + p * 4) : vdup_n_f16((__fp16)0.f);
-        out0.fill(_bias0);
-
-        const __fp16* k0 = kernel.channel(p);
-
-        int q = 0;
-        for (; q < inch; q++)
-        {
-            __fp16* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const __fp16* r0 = img0.row<const __fp16>(0);
-            const __fp16* r1 = img0.row<const __fp16>(1);
-            const __fp16* r2 = img0.row<const __fp16>(2);
-
-            float16x4_t _k00 = vld1_f16(k0);
-            float16x4_t _k01 = vld1_f16(k0 + 4);
-            float16x4_t _k02 = vld1_f16(k0 + 8);
-            float16x4_t _k10 = vld1_f16(k0 + 12);
-            float16x4_t _k11 = vld1_f16(k0 + 16);
-            float16x4_t _k12 = vld1_f16(k0 + 20);
-            float16x4_t _k20 = vld1_f16(k0 + 24);
-            float16x4_t _k21 = vld1_f16(k0 + 28);
-            float16x4_t _k22 = vld1_f16(k0 + 32);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%0], #32 \n" // sum0 sum1 sum2 sum3
-
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h, v31.4h}, [%0] \n" // sum4 sum5 sum6 sum7
-
-                        "sub    %0, %0, #32                 \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.8h}, [%1], #16          \n" // r0
-                        "ld1    {v1.4h}, [%1]               \n"
-
-                        "fmla   v24.4h, %8.4h, v0.h[0]      \n"
-                        "fmla   v25.4h, %8.4h, v0.h[1]      \n"
-                        "fmla   v26.4h, %8.4h, v0.h[2]      \n"
-                        "fmla   v27.4h, %8.4h, v0.h[3]      \n"
-                        "fmla   v28.4h, %8.4h, v0.h[4]      \n"
-                        "fmla   v29.4h, %8.4h, v0.h[5]      \n"
-                        "fmla   v30.4h, %8.4h, v0.h[6]      \n"
-                        "fmla   v31.4h, %8.4h, v0.h[7]      \n"
-
-                        "fmla   v24.4h, %9.4h, v0.h[1]      \n"
-                        "fmla   v25.4h, %9.4h, v0.h[2]      \n"
-                        "fmla   v26.4h, %9.4h, v0.h[3]      \n"
-                        "fmla   v27.4h, %9.4h, v0.h[4]      \n"
-                        "fmla   v28.4h, %9.4h, v0.h[5]      \n"
-                        "fmla   v29.4h, %9.4h, v0.h[6]      \n"
-                        "fmla   v30.4h, %9.4h, v0.h[7]      \n"
-                        "fmla   v31.4h, %9.4h, v1.h[0]      \n"
-
-                        "fmla   v24.4h, %10.4h, v0.h[2]     \n"
-                        "fmla   v25.4h, %10.4h, v0.h[3]     \n"
-                        "fmla   v26.4h, %10.4h, v0.h[4]     \n"
-                        "fmla   v27.4h, %10.4h, v0.h[5]     \n"
-                        "fmla   v28.4h, %10.4h, v0.h[6]     \n"
-                        "fmla   v29.4h, %10.4h, v0.h[7]     \n"
-                        "fmla   v30.4h, %10.4h, v1.h[0]     \n"
-                        "fmla   v31.4h, %10.4h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v2.8h}, [%2], #16          \n" // r1
-                        "ld1    {v3.4h}, [%2]               \n"
-
-                        "fmla   v24.4h, %11.4h, v2.h[0]     \n"
-                        "fmla   v25.4h, %11.4h, v2.h[1]     \n"
-                        "fmla   v26.4h, %11.4h, v2.h[2]     \n"
-                        "fmla   v27.4h, %11.4h, v2.h[3]     \n"
-                        "fmla   v28.4h, %11.4h, v2.h[4]     \n"
-                        "fmla   v29.4h, %11.4h, v2.h[5]     \n"
-                        "fmla   v30.4h, %11.4h, v2.h[6]     \n"
-                        "fmla   v31.4h, %11.4h, v2.h[7]     \n"
-
-                        "fmla   v24.4h, %12.4h, v2.h[1]     \n"
-                        "fmla   v25.4h, %12.4h, v2.h[2]     \n"
-                        "fmla   v26.4h, %12.4h, v2.h[3]     \n"
-                        "fmla   v27.4h, %12.4h, v2.h[4]     \n"
-                        "fmla   v28.4h, %12.4h, v2.h[5]     \n"
-                        "fmla   v29.4h, %12.4h, v2.h[6]     \n"
-                        "fmla   v30.4h, %12.4h, v2.h[7]     \n"
-                        "fmla   v31.4h, %12.4h, v3.h[0]     \n"
-
-                        "fmla   v24.4h, %13.4h, v2.h[2]     \n"
-                        "fmla   v25.4h, %13.4h, v2.h[3]     \n"
-                        "fmla   v26.4h, %13.4h, v2.h[4]     \n"
-                        "fmla   v27.4h, %13.4h, v2.h[5]     \n"
-                        "fmla   v28.4h, %13.4h, v2.h[6]     \n"
-                        "fmla   v29.4h, %13.4h, v2.h[7]     \n"
-                        "fmla   v30.4h, %13.4h, v3.h[0]     \n"
-                        "fmla   v31.4h, %13.4h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v4.8h}, [%3], #16          \n" // r2
-                        "ld1    {v5.4h}, [%3]               \n"
-
-                        "fmla   v24.4h, %14.4h, v4.h[0]     \n"
-                        "fmla   v25.4h, %14.4h, v4.h[1]     \n"
-                        "fmla   v26.4h, %14.4h, v4.h[2]     \n"
-                        "fmla   v27.4h, %14.4h, v4.h[3]     \n"
-                        "fmla   v28.4h, %14.4h, v4.h[4]     \n"
-                        "fmla   v29.4h, %14.4h, v4.h[5]     \n"
-                        "fmla   v30.4h, %14.4h, v4.h[6]     \n"
-                        "fmla   v31.4h, %14.4h, v4.h[7]     \n"
-
-                        "fmla   v24.4h, %15.4h, v4.h[1]     \n"
-                        "fmla   v25.4h, %15.4h, v4.h[2]     \n"
-                        "fmla   v26.4h, %15.4h, v4.h[3]     \n"
-                        "fmla   v27.4h, %15.4h, v4.h[4]     \n"
-                        "fmla   v28.4h, %15.4h, v4.h[5]     \n"
-                        "fmla   v29.4h, %15.4h, v4.h[6]     \n"
-                        "fmla   v30.4h, %15.4h, v4.h[7]     \n"
-                        "fmla   v31.4h, %15.4h, v5.h[0]     \n"
-
-                        "fmla   v24.4h, %16.4h, v4.h[2]     \n"
-                        "fmla   v25.4h, %16.4h, v4.h[3]     \n"
-                        "fmla   v26.4h, %16.4h, v4.h[4]     \n"
-                        "fmla   v27.4h, %16.4h, v4.h[5]     \n"
-                        "fmla   v28.4h, %16.4h, v4.h[6]     \n"
-                        "fmla   v29.4h, %16.4h, v4.h[7]     \n"
-                        "fmla   v30.4h, %16.4h, v5.h[0]     \n"
-                        "fmla   v31.4h, %16.4h, v5.h[1]     \n"
-
-                        "st1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%0], #32 \n"
-                        "st1    {v28.4h, v29.4h, v30.4h, v31.4h}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-                for (; j + 3 < outw; j += 4)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h, v31.4h}, [%0] \n" // sum0 sum1 sum2 sum3
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.8h}, [%1]               \n" // r0
-
-                        "fmla   v28.4h, %8.4h, v0.h[0]      \n"
-                        "fmla   v29.4h, %8.4h, v0.h[1]      \n"
-                        "fmla   v30.4h, %8.4h, v0.h[2]      \n"
-                        "fmla   v31.4h, %8.4h, v0.h[3]      \n"
-
-                        "fmla   v28.4h, %9.4h, v0.h[1]      \n"
-                        "fmla   v29.4h, %9.4h, v0.h[2]      \n"
-                        "fmla   v30.4h, %9.4h, v0.h[3]      \n"
-                        "fmla   v31.4h, %9.4h, v0.h[4]      \n"
-
-                        "fmla   v28.4h, %10.4h, v0.h[2]     \n"
-                        "fmla   v29.4h, %10.4h, v0.h[3]     \n"
-                        "fmla   v30.4h, %10.4h, v0.h[4]     \n"
-                        "fmla   v31.4h, %10.4h, v0.h[5]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v1.8h}, [%2]               \n" // r1
-
-                        "fmla   v28.4h, %11.4h, v1.h[0]     \n"
-                        "fmla   v29.4h, %11.4h, v1.h[1]     \n"
-                        "fmla   v30.4h, %11.4h, v1.h[2]     \n"
-                        "fmla   v31.4h, %11.4h, v1.h[3]     \n"
-
-                        "fmla   v28.4h, %12.4h, v1.h[1]     \n"
-                        "fmla   v29.4h, %12.4h, v1.h[2]     \n"
-                        "fmla   v30.4h, %12.4h, v1.h[3]     \n"
-                        "fmla   v31.4h, %12.4h, v1.h[4]     \n"
-
-                        "fmla   v28.4h, %13.4h, v1.h[2]     \n"
-                        "fmla   v29.4h, %13.4h, v1.h[3]     \n"
-                        "fmla   v30.4h, %13.4h, v1.h[4]     \n"
-                        "fmla   v31.4h, %13.4h, v1.h[5]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v2.8h}, [%3]               \n" // r2
-
-                        "fmla   v28.4h, %14.4h, v2.h[0]     \n"
-                        "fmla   v29.4h, %14.4h, v2.h[1]     \n"
-                        "fmla   v30.4h, %14.4h, v2.h[2]     \n"
-                        "fmla   v31.4h, %14.4h, v2.h[3]     \n"
-
-                        "fmla   v28.4h, %15.4h, v2.h[1]     \n"
-                        "fmla   v29.4h, %15.4h, v2.h[2]     \n"
-                        "fmla   v30.4h, %15.4h, v2.h[3]     \n"
-                        "fmla   v31.4h, %15.4h, v2.h[4]     \n"
-
-                        "fmla   v28.4h, %16.4h, v2.h[2]     \n"
-                        "fmla   v29.4h, %16.4h, v2.h[3]     \n"
-                        "fmla   v30.4h, %16.4h, v2.h[4]     \n"
-                        "fmla   v31.4h, %16.4h, v2.h[5]     \n"
-
-                        "add    %1, %1, #8                  \n"
-                        "add    %2, %2, #8                  \n"
-                        "add    %3, %3, #8                  \n"
-
-                        "st1    {v28.4h, v29.4h, v30.4h, v31.4h}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v28", "v29", "v30", "v31");
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v30.4h, v31.4h}, [%0]      \n" // sum0 sum1
-
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1]               \n" // r0
-
-                        "fmla   v30.4h, %8.4h, v0.h[0]      \n"
-                        "fmla   v31.4h, %8.4h, v0.h[1]      \n"
-                        "fmla   v30.4h, %9.4h, v0.h[1]      \n"
-                        "fmla   v31.4h, %9.4h, v0.h[2]      \n"
-                        "fmla   v30.4h, %10.4h, v0.h[2]     \n"
-                        "fmla   v31.4h, %10.4h, v0.h[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v1.4h}, [%2]               \n" // r1
-
-                        "fmla   v30.4h, %11.4h, v1.h[0]     \n"
-                        "fmla   v31.4h, %11.4h, v1.h[1]     \n"
-                        "fmla   v30.4h, %12.4h, v1.h[1]     \n"
-                        "fmla   v31.4h, %12.4h, v1.h[2]     \n"
-                        "fmla   v30.4h, %13.4h, v1.h[2]     \n"
-                        "fmla   v31.4h, %13.4h, v1.h[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v2.4h}, [%3]               \n" // r2
-
-                        "fmla   v30.4h, %14.4h, v2.h[0]     \n"
-                        "fmla   v31.4h, %14.4h, v2.h[1]     \n"
-                        "fmla   v30.4h, %15.4h, v2.h[1]     \n"
-                        "fmla   v31.4h, %15.4h, v2.h[2]     \n"
-                        "fmla   v30.4h, %16.4h, v2.h[2]     \n"
-                        "fmla   v31.4h, %16.4h, v2.h[3]     \n"
-
-                        "add    %1, %1, #4                  \n"
-                        "add    %2, %2, #4                  \n"
-                        "add    %3, %3, #4                  \n"
-
-                        "st1    {v30.4h, v31.4h}, [%0], #16 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v30", "v31");
-                }
-                for (; j < outw; j++)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #64]        \n"
-                        "ld1    {v30.4h}, [%0]              \n" // sum0
-
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1]               \n" // r0
-
-                        "fmla   v30.4h, %8.4h, v0.h[0]      \n"
-                        "fmla   v30.4h, %9.4h, v0.h[1]      \n"
-                        "fmla   v30.4h, %10.4h, v0.h[2]     \n"
-
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v1.4h}, [%2]               \n" // r1
-
-                        "fmla   v30.4h, %11.4h, v1.h[0]     \n"
-                        "fmla   v30.4h, %12.4h, v1.h[1]     \n"
-                        "fmla   v30.4h, %13.4h, v1.h[2]     \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v2.4h}, [%3]               \n" // r2
-
-                        "fmla   v30.4h, %14.4h, v2.h[0]     \n"
-                        "fmla   v30.4h, %15.4h, v2.h[1]     \n"
-                        "fmla   v30.4h, %16.4h, v2.h[2]     \n"
-
-                        "add    %1, %1, #2                  \n"
-                        "add    %2, %2, #2                  \n"
-                        "add    %3, %3, #2                  \n"
-
-                        "st1    {v30.4h}, [%0], #8          \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v30");
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * 4;
-        }
-    }
-}
-
-static void conv3x3s2_pack1to4_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float16x4_t _bias0 = bias ? vld1_f16(bias + p * 4) : vdup_n_f16((__fp16)0.f);
-        out0.fill(_bias0);
-
-        const __fp16* k0 = kernel.channel(p);
-
-        int q = 0;
-        for (; q < inch; q++)
-        {
-            __fp16* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const __fp16* r0 = img0.row<const __fp16>(0);
-            const __fp16* r1 = img0.row<const __fp16>(1);
-            const __fp16* r2 = img0.row<const __fp16>(2);
-
-            float16x4_t _k00 = vld1_f16(k0);
-            float16x4_t _k01 = vld1_f16(k0 + 4);
-            float16x4_t _k02 = vld1_f16(k0 + 8);
-            float16x4_t _k10 = vld1_f16(k0 + 12);
-            float16x4_t _k11 = vld1_f16(k0 + 16);
-            float16x4_t _k12 = vld1_f16(k0 + 20);
-            float16x4_t _k20 = vld1_f16(k0 + 24);
-            float16x4_t _k21 = vld1_f16(k0 + 28);
-            float16x4_t _k22 = vld1_f16(k0 + 32);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h, v31.4h}, [%0] \n" // sum0 sum1 sum2 sum3
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.8h}, [%1], #16          \n" // r0
-                        "ld1    {v1.h}[0], [%1]             \n"
-
-                        "fmla   v28.4h, %8.4h, v0.h[0]      \n"
-                        "fmla   v29.4h, %8.4h, v0.h[2]      \n"
-                        "fmla   v30.4h, %8.4h, v0.h[4]      \n"
-                        "fmla   v31.4h, %8.4h, v0.h[6]      \n"
-
-                        "fmla   v28.4h, %9.4h, v0.h[1]      \n"
-                        "fmla   v29.4h, %9.4h, v0.h[3]      \n"
-                        "fmla   v30.4h, %9.4h, v0.h[5]      \n"
-                        "fmla   v31.4h, %9.4h, v0.h[7]      \n"
-
-                        "fmla   v28.4h, %10.4h, v0.h[2]     \n"
-                        "fmla   v29.4h, %10.4h, v0.h[4]     \n"
-                        "fmla   v30.4h, %10.4h, v0.h[6]     \n"
-                        "fmla   v31.4h, %10.4h, v1.h[0]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v2.8h}, [%2], #16          \n" // r1
-                        "ld1    {v3.h}[0], [%2]             \n"
-
-                        "fmla   v28.4h, %11.4h, v2.h[0]     \n"
-                        "fmla   v29.4h, %11.4h, v2.h[2]     \n"
-                        "fmla   v30.4h, %11.4h, v2.h[4]     \n"
-                        "fmla   v31.4h, %11.4h, v2.h[6]     \n"
-
-                        "fmla   v28.4h, %12.4h, v2.h[1]     \n"
-                        "fmla   v29.4h, %12.4h, v2.h[3]     \n"
-                        "fmla   v30.4h, %12.4h, v2.h[5]     \n"
-                        "fmla   v31.4h, %12.4h, v2.h[7]     \n"
-
-                        "fmla   v28.4h, %13.4h, v2.h[2]     \n"
-                        "fmla   v29.4h, %13.4h, v2.h[4]     \n"
-                        "fmla   v30.4h, %13.4h, v2.h[6]     \n"
-                        "fmla   v31.4h, %13.4h, v3.h[0]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v4.8h}, [%3], #16          \n" // r2
-                        "ld1    {v5.h}[0], [%3]             \n"
-
-                        "fmla   v28.4h, %14.4h, v4.h[0]     \n"
-                        "fmla   v29.4h, %14.4h, v4.h[2]     \n"
-                        "fmla   v30.4h, %14.4h, v4.h[4]     \n"
-                        "fmla   v31.4h, %14.4h, v4.h[6]     \n"
-
-                        "fmla   v28.4h, %15.4h, v4.h[1]     \n"
-                        "fmla   v29.4h, %15.4h, v4.h[3]     \n"
-                        "fmla   v30.4h, %15.4h, v4.h[5]     \n"
-                        "fmla   v31.4h, %15.4h, v4.h[7]     \n"
-
-                        "fmla   v28.4h, %16.4h, v4.h[2]     \n"
-                        "fmla   v29.4h, %16.4h, v4.h[4]     \n"
-                        "fmla   v30.4h, %16.4h, v4.h[6]     \n"
-                        "fmla   v31.4h, %16.4h, v5.h[0]     \n"
-
-                        "st1    {v28.4h, v29.4h, v30.4h, v31.4h}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v28", "v29", "v30", "v31");
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v30.4h, v31.4h}, [%0]      \n" // sum0 sum1
-
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1], #8           \n" // r0
-                        "ld1    {v1.h}[0], [%1]             \n"
-
-                        "fmla   v30.4h, %8.4h, v0.h[0]      \n"
-                        "fmla   v31.4h, %8.4h, v0.h[2]      \n"
-                        "fmla   v30.4h, %9.4h, v0.h[1]      \n"
-                        "fmla   v31.4h, %9.4h, v0.h[3]      \n"
-                        "fmla   v30.4h, %10.4h, v0.h[2]     \n"
-                        "fmla   v31.4h, %10.4h, v1.h[0]     \n"
-
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v2.4h}, [%2], #8           \n" // r1
-                        "ld1    {v3.h}[0], [%2]             \n"
-
-                        "fmla   v30.4h, %11.4h, v2.h[0]     \n"
-                        "fmla   v31.4h, %11.4h, v2.h[2]     \n"
-                        "fmla   v30.4h, %12.4h, v2.h[1]     \n"
-                        "fmla   v31.4h, %12.4h, v2.h[3]     \n"
-                        "fmla   v30.4h, %13.4h, v2.h[2]     \n"
-                        "fmla   v31.4h, %13.4h, v3.h[0]     \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v4.4h}, [%3], #8           \n" // r2
-                        "ld1    {v5.h}[0], [%3]             \n"
-
-                        "fmla   v30.4h, %14.4h, v4.h[0]     \n"
-                        "fmla   v31.4h, %14.4h, v4.h[2]     \n"
-                        "fmla   v30.4h, %15.4h, v4.h[1]     \n"
-                        "fmla   v31.4h, %15.4h, v4.h[3]     \n"
-                        "fmla   v30.4h, %16.4h, v4.h[2]     \n"
-                        "fmla   v31.4h, %16.4h, v5.h[0]     \n"
-
-                        "st1    {v30.4h, v31.4h}, [%0], #16 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v30", "v31");
-                }
-                for (; j < outw; j++)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #64]        \n"
-                        "ld1    {v30.4h}, [%0]              \n" // sum0
-
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1]               \n" // r0
-
-                        "fmla   v30.4h, %8.4h, v0.h[0]      \n"
-                        "fmla   v30.4h, %9.4h, v0.h[1]      \n"
-                        "fmla   v30.4h, %10.4h, v0.h[2]     \n"
-
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v1.4h}, [%2]               \n" // r1
-
-                        "fmla   v30.4h, %11.4h, v1.h[0]     \n"
-                        "fmla   v30.4h, %12.4h, v1.h[1]     \n"
-                        "fmla   v30.4h, %13.4h, v1.h[2]     \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v2.4h}, [%3]               \n" // r2
-
-                        "fmla   v30.4h, %14.4h, v2.h[0]     \n"
-                        "fmla   v30.4h, %15.4h, v2.h[1]     \n"
-                        "fmla   v30.4h, %16.4h, v2.h[2]     \n"
-
-                        "add    %1, %1, #4                  \n"
-                        "add    %2, %2, #4                  \n"
-                        "add    %3, %3, #4                  \n"
-
-                        "st1    {v30.4h}, [%0], #8          \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v30");
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to8_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to8_fp16s.h
deleted file mode 100644
index bd03d450..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack1to8_fp16s.h
+++ /dev/null
@@ -1,646 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack1to8_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float16x8_t _bias0 = bias ? vld1q_f16(bias + p * 8) : vdupq_n_f16((__fp16)0.f);
-        out0.fill(_bias0);
-
-        const __fp16* k0 = kernel.channel(p);
-
-        int q = 0;
-        for (; q < inch; q++)
-        {
-            __fp16* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const __fp16* r0 = img0.row<const __fp16>(0);
-            const __fp16* r1 = img0.row<const __fp16>(1);
-            const __fp16* r2 = img0.row<const __fp16>(2);
-
-            float16x8_t _k00 = vld1q_f16(k0);
-            float16x8_t _k01 = vld1q_f16(k0 + 8);
-            float16x8_t _k02 = vld1q_f16(k0 + 16);
-            float16x8_t _k10 = vld1q_f16(k0 + 24);
-            float16x8_t _k11 = vld1q_f16(k0 + 32);
-            float16x8_t _k12 = vld1q_f16(k0 + 40);
-            float16x8_t _k20 = vld1q_f16(k0 + 48);
-            float16x8_t _k21 = vld1q_f16(k0 + 56);
-            float16x8_t _k22 = vld1q_f16(k0 + 64);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n" // sum0 sum1 sum2 sum3
-
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0] \n" // sum4 sum5 sum6 sum7
-
-                        "sub    %0, %0, #64                 \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.8h}, [%1], #16          \n" // r0
-                        "ld1    {v1.4h}, [%1]               \n"
-
-                        "fmla   v24.8h, %8.8h, v0.h[0]      \n"
-                        "fmla   v25.8h, %8.8h, v0.h[1]      \n"
-                        "fmla   v26.8h, %8.8h, v0.h[2]      \n"
-                        "fmla   v27.8h, %8.8h, v0.h[3]      \n"
-                        "fmla   v28.8h, %8.8h, v0.h[4]      \n"
-                        "fmla   v29.8h, %8.8h, v0.h[5]      \n"
-                        "fmla   v30.8h, %8.8h, v0.h[6]      \n"
-                        "fmla   v31.8h, %8.8h, v0.h[7]      \n"
-
-                        "fmla   v24.8h, %9.8h, v0.h[1]      \n"
-                        "fmla   v25.8h, %9.8h, v0.h[2]      \n"
-                        "fmla   v26.8h, %9.8h, v0.h[3]      \n"
-                        "fmla   v27.8h, %9.8h, v0.h[4]      \n"
-                        "fmla   v28.8h, %9.8h, v0.h[5]      \n"
-                        "fmla   v29.8h, %9.8h, v0.h[6]      \n"
-                        "fmla   v30.8h, %9.8h, v0.h[7]      \n"
-                        "fmla   v31.8h, %9.8h, v1.h[0]      \n"
-
-                        "fmla   v24.8h, %10.8h, v0.h[2]     \n"
-                        "fmla   v25.8h, %10.8h, v0.h[3]     \n"
-                        "fmla   v26.8h, %10.8h, v0.h[4]     \n"
-                        "fmla   v27.8h, %10.8h, v0.h[5]     \n"
-                        "fmla   v28.8h, %10.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, %10.8h, v0.h[7]     \n"
-                        "fmla   v30.8h, %10.8h, v1.h[0]     \n"
-                        "fmla   v31.8h, %10.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v2.8h}, [%2], #16          \n" // r1
-                        "ld1    {v3.4h}, [%2]               \n"
-
-                        "fmla   v24.8h, %11.8h, v2.h[0]     \n"
-                        "fmla   v25.8h, %11.8h, v2.h[1]     \n"
-                        "fmla   v26.8h, %11.8h, v2.h[2]     \n"
-                        "fmla   v27.8h, %11.8h, v2.h[3]     \n"
-                        "fmla   v28.8h, %11.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, %11.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, %11.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, %11.8h, v2.h[7]     \n"
-
-                        "fmla   v24.8h, %12.8h, v2.h[1]     \n"
-                        "fmla   v25.8h, %12.8h, v2.h[2]     \n"
-                        "fmla   v26.8h, %12.8h, v2.h[3]     \n"
-                        "fmla   v27.8h, %12.8h, v2.h[4]     \n"
-                        "fmla   v28.8h, %12.8h, v2.h[5]     \n"
-                        "fmla   v29.8h, %12.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, %12.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, %12.8h, v3.h[0]     \n"
-
-                        "fmla   v24.8h, %13.8h, v2.h[2]     \n"
-                        "fmla   v25.8h, %13.8h, v2.h[3]     \n"
-                        "fmla   v26.8h, %13.8h, v2.h[4]     \n"
-                        "fmla   v27.8h, %13.8h, v2.h[5]     \n"
-                        "fmla   v28.8h, %13.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, %13.8h, v2.h[7]     \n"
-                        "fmla   v30.8h, %13.8h, v3.h[0]     \n"
-                        "fmla   v31.8h, %13.8h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v4.8h}, [%3], #16          \n" // r2
-                        "ld1    {v5.4h}, [%3]               \n"
-
-                        "fmla   v24.8h, %14.8h, v4.h[0]     \n"
-                        "fmla   v25.8h, %14.8h, v4.h[1]     \n"
-                        "fmla   v26.8h, %14.8h, v4.h[2]     \n"
-                        "fmla   v27.8h, %14.8h, v4.h[3]     \n"
-                        "fmla   v28.8h, %14.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, %14.8h, v4.h[5]     \n"
-                        "fmla   v30.8h, %14.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, %14.8h, v4.h[7]     \n"
-
-                        "fmla   v24.8h, %15.8h, v4.h[1]     \n"
-                        "fmla   v25.8h, %15.8h, v4.h[2]     \n"
-                        "fmla   v26.8h, %15.8h, v4.h[3]     \n"
-                        "fmla   v27.8h, %15.8h, v4.h[4]     \n"
-                        "fmla   v28.8h, %15.8h, v4.h[5]     \n"
-                        "fmla   v29.8h, %15.8h, v4.h[6]     \n"
-                        "fmla   v30.8h, %15.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, %15.8h, v5.h[0]     \n"
-
-                        "fmla   v24.8h, %16.8h, v4.h[2]     \n"
-                        "fmla   v25.8h, %16.8h, v4.h[3]     \n"
-                        "fmla   v26.8h, %16.8h, v4.h[4]     \n"
-                        "fmla   v27.8h, %16.8h, v4.h[5]     \n"
-                        "fmla   v28.8h, %16.8h, v4.h[6]     \n"
-                        "fmla   v29.8h, %16.8h, v4.h[7]     \n"
-                        "fmla   v30.8h, %16.8h, v5.h[0]     \n"
-                        "fmla   v31.8h, %16.8h, v5.h[1]     \n"
-
-                        "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                        "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-                for (; j + 3 < outw; j += 4)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0] \n" // sum0 sum1 sum2 sum3
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.8h}, [%1]               \n" // r0
-
-                        "fmla   v28.8h, %8.8h, v0.h[0]      \n"
-                        "fmla   v29.8h, %8.8h, v0.h[1]      \n"
-                        "fmla   v30.8h, %8.8h, v0.h[2]      \n"
-                        "fmla   v31.8h, %8.8h, v0.h[3]      \n"
-
-                        "fmla   v28.8h, %9.8h, v0.h[1]      \n"
-                        "fmla   v29.8h, %9.8h, v0.h[2]      \n"
-                        "fmla   v30.8h, %9.8h, v0.h[3]      \n"
-                        "fmla   v31.8h, %9.8h, v0.h[4]      \n"
-
-                        "fmla   v28.8h, %10.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, %10.8h, v0.h[3]     \n"
-                        "fmla   v30.8h, %10.8h, v0.h[4]     \n"
-                        "fmla   v31.8h, %10.8h, v0.h[5]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v1.8h}, [%2]               \n" // r1
-
-                        "fmla   v28.8h, %11.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, %11.8h, v1.h[1]     \n"
-                        "fmla   v30.8h, %11.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, %11.8h, v1.h[3]     \n"
-
-                        "fmla   v28.8h, %12.8h, v1.h[1]     \n"
-                        "fmla   v29.8h, %12.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, %12.8h, v1.h[3]     \n"
-                        "fmla   v31.8h, %12.8h, v1.h[4]     \n"
-
-                        "fmla   v28.8h, %13.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, %13.8h, v1.h[3]     \n"
-                        "fmla   v30.8h, %13.8h, v1.h[4]     \n"
-                        "fmla   v31.8h, %13.8h, v1.h[5]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v2.8h}, [%3]               \n" // r2
-
-                        "fmla   v28.8h, %14.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, %14.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, %14.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, %14.8h, v2.h[3]     \n"
-
-                        "fmla   v28.8h, %15.8h, v2.h[1]     \n"
-                        "fmla   v29.8h, %15.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, %15.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, %15.8h, v2.h[4]     \n"
-
-                        "fmla   v28.8h, %16.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, %16.8h, v2.h[3]     \n"
-                        "fmla   v30.8h, %16.8h, v2.h[4]     \n"
-                        "fmla   v31.8h, %16.8h, v2.h[5]     \n"
-
-                        "add    %1, %1, #8                  \n"
-                        "add    %2, %2, #8                  \n"
-                        "add    %3, %3, #8                  \n"
-
-                        "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v28", "v29", "v30", "v31");
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v30.8h, v31.8h}, [%0]      \n" // sum0 sum1
-
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1]               \n" // r0
-
-                        "fmla   v30.8h, %8.8h, v0.h[0]      \n"
-                        "fmla   v31.8h, %8.8h, v0.h[1]      \n"
-                        "fmla   v30.8h, %9.8h, v0.h[1]      \n"
-                        "fmla   v31.8h, %9.8h, v0.h[2]      \n"
-                        "fmla   v30.8h, %10.8h, v0.h[2]     \n"
-                        "fmla   v31.8h, %10.8h, v0.h[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v1.4h}, [%2]               \n" // r1
-
-                        "fmla   v30.8h, %11.8h, v1.h[0]     \n"
-                        "fmla   v31.8h, %11.8h, v1.h[1]     \n"
-                        "fmla   v30.8h, %12.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, %12.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, %13.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, %13.8h, v1.h[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v2.4h}, [%3]               \n" // r2
-
-                        "fmla   v30.8h, %14.8h, v2.h[0]     \n"
-                        "fmla   v31.8h, %14.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, %15.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, %15.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, %16.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, %16.8h, v2.h[3]     \n"
-
-                        "add    %1, %1, #4                  \n"
-                        "add    %2, %2, #4                  \n"
-                        "add    %3, %3, #4                  \n"
-
-                        "st1    {v30.8h, v31.8h}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v30", "v31");
-                }
-                for (; j < outw; j++)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v30.8h}, [%0]              \n" // sum0
-
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1]               \n" // r0
-
-                        "fmla   v30.8h, %8.8h, v0.h[0]      \n"
-                        "fmla   v30.8h, %9.8h, v0.h[1]      \n"
-                        "fmla   v30.8h, %10.8h, v0.h[2]     \n"
-
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v1.4h}, [%2]               \n" // r1
-
-                        "fmla   v30.8h, %11.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, %12.8h, v1.h[1]     \n"
-                        "fmla   v30.8h, %13.8h, v1.h[2]     \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v2.4h}, [%3]               \n" // r2
-
-                        "fmla   v30.8h, %14.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, %15.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, %16.8h, v2.h[2]     \n"
-
-                        "add    %1, %1, #2                  \n"
-                        "add    %2, %2, #2                  \n"
-                        "add    %3, %3, #2                  \n"
-
-                        "st1    {v30.8h}, [%0], #16         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v30");
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * 8;
-        }
-    }
-}
-
-static void conv3x3s2_pack1to8_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float16x8_t _bias0 = bias ? vld1q_f16(bias + p * 8) : vdupq_n_f16((__fp16)0.f);
-        out0.fill(_bias0);
-
-        const __fp16* k0 = kernel.channel(p);
-
-        int q = 0;
-        for (; q < inch; q++)
-        {
-            __fp16* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const __fp16* r0 = img0.row<const __fp16>(0);
-            const __fp16* r1 = img0.row<const __fp16>(1);
-            const __fp16* r2 = img0.row<const __fp16>(2);
-
-            float16x8_t _k00 = vld1q_f16(k0);
-            float16x8_t _k01 = vld1q_f16(k0 + 8);
-            float16x8_t _k02 = vld1q_f16(k0 + 16);
-            float16x8_t _k10 = vld1q_f16(k0 + 24);
-            float16x8_t _k11 = vld1q_f16(k0 + 32);
-            float16x8_t _k12 = vld1q_f16(k0 + 40);
-            float16x8_t _k20 = vld1q_f16(k0 + 48);
-            float16x8_t _k21 = vld1q_f16(k0 + 56);
-            float16x8_t _k22 = vld1q_f16(k0 + 64);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0] \n" // sum0 sum1 sum2 sum3
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.8h}, [%1], #16          \n" // r0
-                        "ld1    {v1.h}[0], [%1]             \n"
-
-                        "fmla   v28.8h, %8.8h, v0.h[0]      \n"
-                        "fmla   v29.8h, %8.8h, v0.h[2]      \n"
-                        "fmla   v30.8h, %8.8h, v0.h[4]      \n"
-                        "fmla   v31.8h, %8.8h, v0.h[6]      \n"
-
-                        "fmla   v28.8h, %9.8h, v0.h[1]      \n"
-                        "fmla   v29.8h, %9.8h, v0.h[3]      \n"
-                        "fmla   v30.8h, %9.8h, v0.h[5]      \n"
-                        "fmla   v31.8h, %9.8h, v0.h[7]      \n"
-
-                        "fmla   v28.8h, %10.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, %10.8h, v0.h[4]     \n"
-                        "fmla   v30.8h, %10.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, %10.8h, v1.h[0]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v2.8h}, [%2], #16          \n" // r1
-                        "ld1    {v3.h}[0], [%2]             \n"
-
-                        "fmla   v28.8h, %11.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, %11.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, %11.8h, v2.h[4]     \n"
-                        "fmla   v31.8h, %11.8h, v2.h[6]     \n"
-
-                        "fmla   v28.8h, %12.8h, v2.h[1]     \n"
-                        "fmla   v29.8h, %12.8h, v2.h[3]     \n"
-                        "fmla   v30.8h, %12.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, %12.8h, v2.h[7]     \n"
-
-                        "fmla   v28.8h, %13.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, %13.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, %13.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, %13.8h, v3.h[0]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v4.8h}, [%3], #16          \n" // r2
-                        "ld1    {v5.h}[0], [%3]             \n"
-
-                        "fmla   v28.8h, %14.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, %14.8h, v4.h[2]     \n"
-                        "fmla   v30.8h, %14.8h, v4.h[4]     \n"
-                        "fmla   v31.8h, %14.8h, v4.h[6]     \n"
-
-                        "fmla   v28.8h, %15.8h, v4.h[1]     \n"
-                        "fmla   v29.8h, %15.8h, v4.h[3]     \n"
-                        "fmla   v30.8h, %15.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, %15.8h, v4.h[7]     \n"
-
-                        "fmla   v28.8h, %16.8h, v4.h[2]     \n"
-                        "fmla   v29.8h, %16.8h, v4.h[4]     \n"
-                        "fmla   v30.8h, %16.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, %16.8h, v5.h[0]     \n"
-
-                        "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v28", "v29", "v30", "v31");
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v30.8h, v31.8h}, [%0]      \n" // sum0 sum1
-
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1], #8           \n" // r0
-                        "ld1    {v1.h}[0], [%1]             \n"
-
-                        "fmla   v30.8h, %8.8h, v0.h[0]      \n"
-                        "fmla   v31.8h, %8.8h, v0.h[2]      \n"
-                        "fmla   v30.8h, %9.8h, v0.h[1]      \n"
-                        "fmla   v31.8h, %9.8h, v0.h[3]      \n"
-                        "fmla   v30.8h, %10.8h, v0.h[2]     \n"
-                        "fmla   v31.8h, %10.8h, v1.h[0]     \n"
-
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v2.4h}, [%2], #8           \n" // r1
-                        "ld1    {v3.h}[0], [%2]             \n"
-
-                        "fmla   v30.8h, %11.8h, v2.h[0]     \n"
-                        "fmla   v31.8h, %11.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, %12.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, %12.8h, v2.h[3]     \n"
-                        "fmla   v30.8h, %13.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, %13.8h, v3.h[0]     \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v4.4h}, [%3], #8           \n" // r2
-                        "ld1    {v5.h}[0], [%3]             \n"
-
-                        "fmla   v30.8h, %14.8h, v4.h[0]     \n"
-                        "fmla   v31.8h, %14.8h, v4.h[2]     \n"
-                        "fmla   v30.8h, %15.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, %15.8h, v4.h[3]     \n"
-                        "fmla   v30.8h, %16.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, %16.8h, v5.h[0]     \n"
-
-                        "st1    {v30.8h, v31.8h}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v30", "v31");
-                }
-                for (; j < outw; j++)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v30.8h}, [%0]              \n" // sum0
-
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1]               \n" // r0
-
-                        "fmla   v30.8h, %8.8h, v0.h[0]      \n"
-                        "fmla   v30.8h, %9.8h, v0.h[1]      \n"
-                        "fmla   v30.8h, %10.8h, v0.h[2]     \n"
-
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v1.4h}, [%2]               \n" // r1
-
-                        "fmla   v30.8h, %11.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, %12.8h, v1.h[1]     \n"
-                        "fmla   v30.8h, %13.8h, v1.h[2]     \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v2.4h}, [%3]               \n" // r2
-
-                        "fmla   v30.8h, %14.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, %15.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, %16.8h, v2.h[2]     \n"
-
-                        "add    %1, %1, #4                  \n"
-                        "add    %2, %2, #4                  \n"
-                        "add    %3, %3, #4                  \n"
-
-                        "st1    {v30.8h}, [%0], #16         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "cc", "memory", "v0", "v1", "v2", "v30");
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * 8;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4.h
deleted file mode 100644
index 5255d162..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4.h
+++ /dev/null
@@ -1,1072 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s2_pack4_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        out0.fill(_bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            const float* kptr = (const float*)kernel.channel(p).row(q);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0] \n" // sum0 sum1 sum2 sum3
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n" // r04 r05 r06 r07
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4], #64 \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%4], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v28.4s}, [%1]              \n" // r08
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%2], #64 \n" // r10 r11 r12 r13
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%4], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v28.s[3]    \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%2], #64 \n" // r14 r15 r16 r17
-
-                        "fmla   v20.4s, v24.4s, v8.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v10.s[0]    \n"
-                        "fmla   v22.4s, v24.4s, v12.s[0]    \n"
-                        "fmla   v23.4s, v24.4s, v14.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v8.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v10.s[1]    \n"
-                        "fmla   v22.4s, v25.4s, v12.s[1]    \n"
-                        "fmla   v23.4s, v25.4s, v14.s[1]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v8.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v10.s[2]    \n"
-                        "fmla   v22.4s, v26.4s, v12.s[2]    \n"
-                        "fmla   v23.4s, v26.4s, v14.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v8.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v10.s[3]    \n"
-                        "fmla   v22.4s, v27.4s, v12.s[3]    \n"
-                        "fmla   v23.4s, v27.4s, v14.s[3]    \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v28.4s}, [%2]              \n" // r18
-
-                        "fmla   v20.4s, v16.4s, v9.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v11.s[0]    \n"
-                        "fmla   v22.4s, v16.4s, v13.s[0]    \n"
-                        "fmla   v23.4s, v16.4s, v15.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v9.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v11.s[1]    \n"
-                        "fmla   v22.4s, v17.4s, v13.s[1]    \n"
-                        "fmla   v23.4s, v17.4s, v15.s[1]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%4], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v9.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v11.s[2]    \n"
-                        "fmla   v22.4s, v18.4s, v13.s[2]    \n"
-                        "fmla   v23.4s, v18.4s, v15.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v9.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v11.s[3]    \n"
-                        "fmla   v22.4s, v19.4s, v13.s[3]    \n"
-                        "fmla   v23.4s, v19.4s, v15.s[3]    \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%3], #64 \n" // r20 r21 r22 r23
-
-                        "fmla   v20.4s, v24.4s, v10.s[0]    \n"
-                        "fmla   v21.4s, v24.4s, v12.s[0]    \n"
-                        "fmla   v22.4s, v24.4s, v14.s[0]    \n"
-                        "fmla   v23.4s, v24.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v10.s[1]    \n"
-                        "fmla   v21.4s, v25.4s, v12.s[1]    \n"
-                        "fmla   v22.4s, v25.4s, v14.s[1]    \n"
-                        "fmla   v23.4s, v25.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v10.s[2]    \n"
-                        "fmla   v21.4s, v26.4s, v12.s[2]    \n"
-                        "fmla   v22.4s, v26.4s, v14.s[2]    \n"
-                        "fmla   v23.4s, v26.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v10.s[3]    \n"
-                        "fmla   v21.4s, v27.4s, v12.s[3]    \n"
-                        "fmla   v22.4s, v27.4s, v14.s[3]    \n"
-                        "fmla   v23.4s, v27.4s, v28.s[3]    \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%3], #64 \n" // r24 r25 r26 r27
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%4], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v28.4s}, [%3]              \n" // r28
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        //                         "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4] \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v28.s[1]    \n"
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v28.s[3]    \n"
-
-                        "sub    %4, %4, #512                \n" // kptr -= 8 * 16;
-
-                        "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(kptr)     // %4
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #512]          \n"
-                        "vldm       %0, {d24-d31}       \n" // sum0 sum1 sum2 sum3
-
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1!, {d0-d7}        \n" // r00 r01 r02 r03
-
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1!, {d8-d15}       \n" // r04 r05 r06 r07
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%1 :128]  \n" // r08
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%2, #512]          \n"
-                        "vldm       %2!, {d8-d15}       \n" // r10 r11 r12 r13
-
-                        "pld        [%2, #512]          \n"
-                        "vldm       %2!, {d0-d7}        \n" // r14 r15 r16 r17
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d12[0]     \n"
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d13[0]    \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d8-d9}, [%2 :128]  \n" // r18
-
-                        "vmla.f32   q12, q8, d10[0]     \n"
-                        "vmla.f32   q13, q8, d14[0]     \n"
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d10[1]     \n"
-                        "vmla.f32   q13, q9, d14[1]     \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d11[0]    \n"
-                        "vmla.f32   q13, q10, d15[0]    \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d11[1]    \n"
-                        "vmla.f32   q13, q11, d15[1]    \n"
-                        "vmla.f32   q14, q11, d3[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d12[0]     \n"
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d12[1]     \n"
-                        "vmla.f32   q13, q9, d0[1]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-                        "vmla.f32   q12, q10, d13[0]    \n"
-                        "vmla.f32   q13, q10, d1[0]     \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d13[1]    \n"
-                        "vmla.f32   q13, q11, d1[1]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%3, #512]          \n"
-                        "vldm       %3!, {d0-d7}        \n" // r20 r21 r22 r23
-
-                        "pld        [%3, #512]          \n"
-                        "vldm       %3!, {d8-d15}       \n" // r24 r25 r26 r27
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%3 :128]  \n" // r28
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        //                         "pld        [%4, #512]          \n"
-                        "vldm       %4, {d16-d23}       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "sub        %4, %4, #512        \n" // kptr -= 8 * 16;
-
-                        "vstm       %0!, {d24-d31}      \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(kptr)     // %4
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v20.4s, v21.4s}, [%0]      \n" // sum0 sum1
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4], #64 \n"
-
-                        "fmul   v22.4s, v16.4s, v0.s[0]     \n"
-                        "fmul   v23.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%4], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v4.4s}, [%1]               \n" // r04
-
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%4], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n" // r10 r11 r12 r13
-
-                        "fmla   v22.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v4.4s}, [%2]               \n" // r14
-
-                        "fmla   v22.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%4], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%3], #64 \n" // r20 r21 r22 r23
-
-                        "fmla   v22.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%4], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v4.4s}, [%3]               \n" // r24
-
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-
-                        //                         "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4] \n"
-
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-
-                        "fadd   v20.4s, v20.4s, v22.4s      \n"
-                        "fadd   v21.4s, v21.4s, v23.4s      \n"
-
-                        "sub    %4, %4, #512                \n" // kptr -= 8 * 16;
-
-                        "st1    {v20.4s, v21.4s}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(kptr)     // %4
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #256]          \n"
-                        "vld1.f32   {d24-d27}, [%0 :128] \n" // sum0 sum1
-
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1!, {d0-d7}        \n" // r00 r01 r02 r03
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmul.f32   q14, q8, d0[0]      \n"
-                        "vmul.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d8-d9}, [%1 :128]  \n" // r04
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-
-                        "pld        [%2, #512]          \n"
-                        "vldm       %2!, {d0-d7}        \n" // r10 r11 r12 r13
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d8-d9}, [%2 :128]  \n" // r14
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-
-                        "pld        [%3, #512]          \n"
-                        "vldm       %3!, {d0-d7}        \n" // r20 r21 r22 r23
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.f32   {d8-d9}, [%3 :128]  \n" // r24
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-
-                        //                         "pld        [%4, #512]          \n"
-                        "vldm       %4, {d16-d23}       \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-
-                        "vadd.f32   q12, q12, q14       \n"
-                        "vadd.f32   q13, q13, q15       \n"
-
-                        "sub        %4, %4, #512        \n" // kptr -= 8 * 16;
-
-                        "vst1.f32   {d24-d27}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(kptr)     // %4
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v20.4s}, [%0]              \n" // sum0
-
-                        "prfm   pldl1keep, [%1, #384]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s}, [%1] \n" // r00 r01 r02
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4], #64 \n"
-
-                        "fmul   v21.4s, v16.4s, v0.s[0]     \n"
-                        "fmul   v22.4s, v17.4s, v0.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%4], #64 \n"
-
-                        "fmul   v23.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #384]       \n"
-                        "ld1    {v3.4s, v4.4s, v5.4s}, [%2] \n" // r10 r11 r12
-
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%4], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%4], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #384]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s}, [%3] \n" // r20 r21 r22
-
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v5.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v0.s[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%4], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-
-                        //                         "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%4] \n"
-
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-
-                        "add    %1, %1, #32                 \n"
-
-                        "fadd   v22.4s, v21.4s, v22.4s      \n"
-
-                        "add    %2, %2, #32                 \n"
-
-                        "fadd   v23.4s, v23.4s, v22.4s      \n"
-
-                        "add    %3, %3, #32                 \n"
-
-                        "fadd   v20.4s, v20.4s, v23.4s      \n"
-
-                        "sub    %4, %4, #512                \n" // kptr -= 8 * 16;
-
-                        "st1    {v20.4s}, [%0], #16         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(kptr)     // %4
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #128]          \n"
-                        "vld1.f32   {d24-d25}, [%0 :128] \n" // sum0
-
-                        "pld        [%1, #384]          \n"
-                        "vldm       %1, {d0-d5}         \n" // r00 r01 r02
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmul.f32   q13, q8, d0[0]      \n"
-                        "vmul.f32   q14, q9, d0[1]      \n"
-                        "vmul.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-
-                        "pld        [%2, #384]          \n"
-                        "vldm       %2, {d0-d5}         \n" // r10 r11 r12
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-
-                        "pld        [%3, #384]          \n"
-                        "vldm       %3, {d0-d5}         \n" // r20 r21 r22
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-
-                        //                         "pld        [%4, #512]          \n"
-                        "vldm       %4, {d16-d23}       \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-
-                        "vadd.f32   q14, q14, q13       \n"
-
-                        "add        %1, %1, #32         \n"
-
-                        "vadd.f32   q15, q15, q14       \n"
-
-                        "add        %2, %2, #32         \n"
-
-                        "vadd.f32   q12, q12, q15       \n"
-
-                        "add        %3, %3, #32         \n"
-
-                        "sub        %4, %4, #512        \n" // kptr -= 8 * 16;
-
-                        "vst1.f32   {d24-d25}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(kptr)     // %4
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4_bf16s.h
deleted file mode 100644
index 9f47aa4c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4_bf16s.h
+++ /dev/null
@@ -1,3003 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s2_pack4_bf16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    Mat top_blob_fp32(outw, outh, opt.num_threads, (size_t)4u * 4, 4, opt.workspace_allocator);
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob_fp32.channel(get_omp_thread_num());
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        out0.fill(_bias0);
-
-        int q = 0;
-        for (; q < inch - 1; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<const unsigned short>(0);
-            const unsigned short* r1 = img0.row<const unsigned short>(1);
-            const unsigned short* r2 = img0.row<const unsigned short>(2);
-
-            const unsigned short* kptr = (const unsigned short*)kernel.channel(p).row<const unsigned short>(q);
-
-#if __aarch64__
-            // 16 * 9
-            uint16x8_t _k00_01 = vld1q_u16(kptr);
-            uint16x8_t _k00_23 = vld1q_u16(kptr + 8);
-            uint16x8_t _k01_01 = vld1q_u16(kptr + 16);
-            uint16x8_t _k01_23 = vld1q_u16(kptr + 24);
-            uint16x8_t _k02_01 = vld1q_u16(kptr + 32);
-            uint16x8_t _k02_23 = vld1q_u16(kptr + 40);
-            uint16x8_t _k10_01 = vld1q_u16(kptr + 48);
-            uint16x8_t _k10_23 = vld1q_u16(kptr + 56);
-            uint16x8_t _k11_01 = vld1q_u16(kptr + 64);
-            uint16x8_t _k11_23 = vld1q_u16(kptr + 72);
-            uint16x8_t _k12_01 = vld1q_u16(kptr + 80);
-            uint16x8_t _k12_23 = vld1q_u16(kptr + 88);
-            uint16x8_t _k20_01 = vld1q_u16(kptr + 96);
-            uint16x8_t _k20_23 = vld1q_u16(kptr + 104);
-            uint16x8_t _k21_01 = vld1q_u16(kptr + 112);
-            uint16x8_t _k21_23 = vld1q_u16(kptr + 120);
-            uint16x8_t _k22_01 = vld1q_u16(kptr + 128);
-            uint16x8_t _k22_23 = vld1q_u16(kptr + 136);
-#endif // __aarch64__
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%0] \n" // sum0 sum1 sum2 sum3
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%1], #64 \n" // r00 r01 r02 r03
-
-                        "shll   v0.4s, v4.4h, #16           \n"
-                        "shll2  v1.4s, v4.8h, #16           \n"
-                        "shll   v2.4s, v5.4h, #16           \n"
-                        "shll2  v3.4s, v5.8h, #16           \n"
-
-                        "shll   v4.4s, v6.4h, #16           \n"
-                        "shll2  v5.4s, v6.8h, #16           \n"
-                        "shll   v6.4s, v7.4h, #16           \n"
-                        "shll2  v7.4s, v7.8h, #16           \n"
-
-                        "shll   v8.4s, %8.4h, #16           \n"
-                        "shll2  v9.4s, %8.8h, #16           \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v4.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v6.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v0.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v2.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v4.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v6.s[1]      \n"
-
-                        "shll   v8.4s, %9.4h, #16           \n"
-                        "shll2  v9.4s, %9.8h, #16           \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v6.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v0.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v4.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v6.s[3]      \n"
-
-                        "shll   v8.4s, %10.4h, #16          \n"
-                        "shll2  v9.4s, %10.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v5.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v7.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v1.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v3.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v5.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v7.s[1]      \n"
-
-                        "shll   v8.4s, %11.4h, #16          \n"
-                        "shll2  v9.4s, %11.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v5.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v7.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v1.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v3.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v5.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v7.s[3]      \n"
-
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1]               \n" // r08
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "shll   v8.4s, %12.4h, #16          \n"
-                        "shll2  v9.4s, %12.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v6.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v0.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v2.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v4.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v6.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[1]      \n"
-
-                        "shll   v8.4s, %13.4h, #16          \n"
-                        "shll2  v9.4s, %13.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v6.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v4.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v6.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[3]      \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%2], #64 \n" // r10 r11 r12 r13
-
-                        "shll   v0.4s, v4.4h, #16           \n"
-                        "shll2  v1.4s, v4.8h, #16           \n"
-                        "shll   v2.4s, v5.4h, #16           \n"
-                        "shll2  v3.4s, v5.8h, #16           \n"
-
-                        "shll   v4.4s, v6.4h, #16           \n"
-                        "shll2  v5.4s, v6.8h, #16           \n"
-                        "shll   v6.4s, v7.4h, #16           \n"
-                        "shll2  v7.4s, v7.8h, #16           \n"
-
-                        "shll   v8.4s, %14.4h, #16          \n"
-                        "shll2  v9.4s, %14.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v4.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v6.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v0.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v2.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v4.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v6.s[1]      \n"
-
-                        "shll   v8.4s, %15.4h, #16          \n"
-                        "shll2  v9.4s, %15.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v6.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v0.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v4.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v6.s[3]      \n"
-
-                        "shll   v8.4s, %16.4h, #16          \n"
-                        "shll2  v9.4s, %16.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v5.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v7.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v1.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v3.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v5.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v7.s[1]      \n"
-
-                        "shll   v8.4s, %17.4h, #16          \n"
-                        "shll2  v9.4s, %17.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v5.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v7.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v1.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v3.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v5.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v7.s[3]      \n"
-
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v0.4h}, [%2]               \n" // r18
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "shll   v8.4s, %18.4h, #16          \n"
-                        "shll2  v9.4s, %18.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v6.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v0.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v2.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v4.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v6.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[1]      \n"
-
-                        "shll   v8.4s, %19.4h, #16          \n"
-                        "shll2  v9.4s, %19.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v6.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v4.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v6.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[3]      \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%3], #64 \n" // r20 r21 r22 r23
-
-                        "shll   v0.4s, v4.4h, #16           \n"
-                        "shll2  v1.4s, v4.8h, #16           \n"
-                        "shll   v2.4s, v5.4h, #16           \n"
-                        "shll2  v3.4s, v5.8h, #16           \n"
-
-                        "shll   v4.4s, v6.4h, #16           \n"
-                        "shll2  v5.4s, v6.8h, #16           \n"
-                        "shll   v6.4s, v7.4h, #16           \n"
-                        "shll2  v7.4s, v7.8h, #16           \n"
-
-                        "shll   v8.4s, %20.4h, #16          \n"
-                        "shll2  v9.4s, %20.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v4.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v6.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v0.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v2.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v4.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v6.s[1]      \n"
-
-                        "shll   v8.4s, %21.4h, #16          \n"
-                        "shll2  v9.4s, %21.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v6.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v0.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v4.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v6.s[3]      \n"
-
-                        "shll   v8.4s, %22.4h, #16          \n"
-                        "shll2  v9.4s, %22.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v5.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v7.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v1.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v3.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v5.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v7.s[1]      \n"
-
-                        "shll   v8.4s, %23.4h, #16          \n"
-                        "shll2  v9.4s, %23.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v5.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v7.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v1.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v3.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v5.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v7.s[3]      \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v0.4h}, [%3]               \n" // r28
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "shll   v8.4s, %24.4h, #16          \n"
-                        "shll2  v9.4s, %24.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v6.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v0.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v2.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v4.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v6.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[1]      \n"
-
-                        "shll   v8.4s, %25.4h, #16          \n"
-                        "shll2  v9.4s, %25.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v6.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v4.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v6.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[3]      \n"
-
-                        "st1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00_01), // %8
-                        "w"(_k00_23), // %9
-                        "w"(_k01_01), // %10
-                        "w"(_k01_23), // %11
-                        "w"(_k02_01), // %12
-                        "w"(_k02_23), // %13
-                        "w"(_k10_01), // %14
-                        "w"(_k10_23), // %15
-                        "w"(_k11_01), // %16
-                        "w"(_k11_23), // %17
-                        "w"(_k12_01), // %18
-                        "w"(_k12_23), // %19
-                        "w"(_k20_01), // %20
-                        "w"(_k20_23), // %21
-                        "w"(_k21_01), // %22
-                        "w"(_k21_23), // %23
-                        "w"(_k22_01), // %24
-                        "w"(_k22_23)  // %25
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #512]          \n"
-                        "vldm       %0, {d24-d31}       \n" // sum0 sum1 sum2 sum3
-
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1!, {d8-d15}       \n" // r00 r01 r02 r03 r04 r05 r06 r07
-
-                        "vshll.u16  q0, d8, #16         \n"
-                        "vshll.u16  q1, d9, #16         \n"
-                        "vshll.u16  q2, d10, #16        \n"
-                        "vshll.u16  q3, d11, #16        \n"
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%1, #64]           \n"
-                        "vld1.f32   {d1}, [%1 :64]      \n" // r08
-
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%2, #512]          \n"
-                        "vldm       %2!, {d8-d15}       \n" // r10 r11 r12 r13 r14 r15 r16 r17
-
-                        "vshll.u16  q0, d8, #16         \n"
-                        "vshll.u16  q1, d9, #16         \n"
-                        "vshll.u16  q2, d10, #16        \n"
-                        "vshll.u16  q3, d11, #16        \n"
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%2, #64]           \n"
-                        "vld1.f32   {d1}, [%2 :64]      \n" // r18
-
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vldm       %3!, {d8-d15}       \n" // r20 r21 r22 r23 r24 r25 r26 r27
-
-                        "vshll.u16  q0, d8, #16         \n"
-                        "vshll.u16  q1, d9, #16         \n"
-                        "vshll.u16  q2, d10, #16        \n"
-                        "vshll.u16  q3, d11, #16        \n"
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%3, #64]           \n"
-                        "vld1.f32   {d1}, [%3 :64]      \n" // r28
-
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        //                         "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128] \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "sub        %4, %4, #256        \n" // kptr -= 8 * 16;
-
-                        "vstm       %0!, {d24-d31}      \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(kptr)     // %4
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v12.4s, v13.4s}, [%0]      \n" // sum0 sum1
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%1], #32 \n" // r00 r01 r02 r03
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "shll   v6.4s, %8.4h, #16           \n"
-                        "shll2  v7.4s, %8.8h, #16           \n"
-                        "shll   v8.4s, %9.4h, #16           \n"
-                        "shll2  v9.4s, %9.8h, #16           \n"
-
-                        "fmul   v10.4s, v6.4s, v0.s[0]      \n"
-                        "fmul   v11.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v0.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v2.s[1]      \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v0.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v2.s[3]      \n"
-
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v4.4h}, [%1]               \n" // r04
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "shll   v6.4s, %10.4h, #16          \n"
-                        "shll2  v7.4s, %10.8h, #16          \n"
-                        "shll   v8.4s, %11.4h, #16          \n"
-                        "shll2  v9.4s, %11.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v3.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v1.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v3.s[1]      \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v1.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v3.s[3]      \n"
-
-                        "shll   v6.4s, %12.4h, #16          \n"
-                        "shll2  v7.4s, %12.8h, #16          \n"
-                        "shll   v8.4s, %13.4h, #16          \n"
-                        "shll2  v9.4s, %13.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v4.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v2.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v4.s[1]      \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v4.s[3]      \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n" // r10 r11 r12 r13
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "shll   v6.4s, %14.4h, #16          \n"
-                        "shll2  v7.4s, %14.8h, #16          \n"
-                        "shll   v8.4s, %15.4h, #16          \n"
-                        "shll2  v9.4s, %15.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v0.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v0.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v2.s[1]      \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v0.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v2.s[3]      \n"
-
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v4.4h}, [%2]               \n" // r14
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "shll   v6.4s, %16.4h, #16          \n"
-                        "shll2  v7.4s, %16.8h, #16          \n"
-                        "shll   v8.4s, %17.4h, #16          \n"
-                        "shll2  v9.4s, %17.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v3.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v1.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v3.s[1]      \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v1.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v3.s[3]      \n"
-
-                        "shll   v6.4s, %18.4h, #16          \n"
-                        "shll2  v7.4s, %18.8h, #16          \n"
-                        "shll   v8.4s, %19.4h, #16          \n"
-                        "shll2  v9.4s, %19.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v4.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v2.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v4.s[1]      \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v4.s[3]      \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%3], #32 \n" // r20 r21 r22 r23
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "shll   v6.4s, %20.4h, #16          \n"
-                        "shll2  v7.4s, %20.8h, #16          \n"
-                        "shll   v8.4s, %21.4h, #16          \n"
-                        "shll2  v9.4s, %21.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v0.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v0.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v2.s[1]      \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v0.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v2.s[3]      \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v4.4h}, [%3]               \n" // r24
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "shll   v6.4s, %22.4h, #16          \n"
-                        "shll2  v7.4s, %22.8h, #16          \n"
-                        "shll   v8.4s, %23.4h, #16          \n"
-                        "shll2  v9.4s, %23.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v3.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v1.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v3.s[1]      \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v1.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v3.s[3]      \n"
-
-                        "shll   v6.4s, %24.4h, #16          \n"
-                        "shll2  v7.4s, %24.8h, #16          \n"
-                        "shll   v8.4s, %25.4h, #16          \n"
-                        "shll2  v9.4s, %25.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v4.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v2.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v4.s[1]      \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v4.s[3]      \n"
-
-                        "fadd   v12.4s, v10.4s, v12.4s      \n"
-                        "fadd   v13.4s, v11.4s, v13.4s      \n"
-
-                        "st1    {v12.4s, v13.4s}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00_01), // %8
-                        "w"(_k00_23), // %9
-                        "w"(_k01_01), // %10
-                        "w"(_k01_23), // %11
-                        "w"(_k02_01), // %12
-                        "w"(_k02_23), // %13
-                        "w"(_k10_01), // %14
-                        "w"(_k10_23), // %15
-                        "w"(_k11_01), // %16
-                        "w"(_k11_23), // %17
-                        "w"(_k12_01), // %18
-                        "w"(_k12_23), // %19
-                        "w"(_k20_01), // %20
-                        "w"(_k20_23), // %21
-                        "w"(_k21_01), // %22
-                        "w"(_k21_23), // %23
-                        "w"(_k22_01), // %24
-                        "w"(_k22_23)  // %25
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #256]          \n"
-                        "vld1.f32   {d28-d31}, [%0 :128] \n" // sum0 sum1
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%1 :64]!  \n" // r00 r01 r02 r03
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmul.f32   q12, q8, d0[0]      \n"
-                        "vmul.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%1, #64]           \n"
-                        "vld1.f32   {d9}, [%1 :64]      \n" // r04
-
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q11, d3[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%2 :64]!  \n" // r10 r11 r12 r13
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%2, #64]           \n"
-                        "vld1.f32   {d9}, [%2 :64]      \n" // r14
-
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q11, d3[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%3 :64]!  \n" // r20 r21 r22 r23
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%3, #64]           \n"
-                        "vld1.f32   {d9}, [%3 :64]      \n" // r24
-
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q11, d3[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        //                         "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128] \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "vadd.f32   q14, q12, q14       \n"
-                        "vadd.f32   q15, q13, q15       \n"
-
-                        "sub        %4, %4, #256        \n" // kptr -= 8 * 16;
-
-                        "vst1.f32   {d28-d31}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(kptr)     // %4
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v13.4s}, [%0]              \n" // sum0
-
-                        "prfm   pldl1keep, [%1, #192]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%1] \n" // r00 r01 r02
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "shll   v6.4s, %8.4h, #16           \n"
-                        "shll2  v7.4s, %8.8h, #16           \n"
-
-                        "fmul   v10.4s, v6.4s, v0.s[0]      \n"
-                        "fmul   v11.4s, v7.4s, v0.s[1]      \n"
-
-                        "shll   v8.4s, %9.4h, #16           \n"
-                        "shll2  v9.4s, %9.8h, #16           \n"
-
-                        "fmul   v12.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[3]      \n"
-
-                        "shll   v6.4s, %10.4h, #16          \n"
-                        "shll2  v7.4s, %10.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v1.s[1]      \n"
-
-                        "shll   v8.4s, %11.4h, #16          \n"
-                        "shll2  v9.4s, %11.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v1.s[3]      \n"
-
-                        "shll   v6.4s, %12.4h, #16          \n"
-                        "shll2  v7.4s, %12.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v2.s[1]      \n"
-
-                        "shll   v8.4s, %13.4h, #16          \n"
-                        "shll2  v9.4s, %13.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v2.s[3]      \n"
-
-                        "prfm   pldl1keep, [%2, #192]       \n"
-                        "ld1    {v3.4h, v4.4h, v5.4h}, [%2] \n" // r10 r11 r12
-
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "shll   v6.4s, %14.4h, #16          \n"
-                        "shll2  v7.4s, %14.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v3.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v3.s[1]      \n"
-
-                        "shll   v8.4s, %15.4h, #16          \n"
-                        "shll2  v9.4s, %15.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v3.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v3.s[3]      \n"
-
-                        "shll   v6.4s, %16.4h, #16          \n"
-                        "shll2  v7.4s, %16.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v4.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v4.s[1]      \n"
-
-                        "shll   v8.4s, %17.4h, #16          \n"
-                        "shll2  v9.4s, %17.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v4.s[3]      \n"
-
-                        "shll   v6.4s, %18.4h, #16          \n"
-                        "shll2  v7.4s, %18.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v5.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v5.s[1]      \n"
-
-                        "shll   v8.4s, %19.4h, #16          \n"
-                        "shll2  v9.4s, %19.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v5.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v5.s[3]      \n"
-
-                        "prfm   pldl1keep, [%3, #192]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%3] \n" // r20 r21 r22
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "shll   v6.4s, %20.4h, #16          \n"
-                        "shll2  v7.4s, %20.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v0.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v0.s[1]      \n"
-
-                        "shll   v8.4s, %21.4h, #16          \n"
-                        "shll2  v9.4s, %21.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[3]      \n"
-
-                        "shll   v6.4s, %22.4h, #16          \n"
-                        "shll2  v7.4s, %22.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v1.s[1]      \n"
-
-                        "shll   v8.4s, %23.4h, #16          \n"
-                        "shll2  v9.4s, %23.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v1.s[3]      \n"
-
-                        "shll   v6.4s, %24.4h, #16          \n"
-                        "shll2  v7.4s, %24.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v2.s[1]      \n"
-
-                        "shll   v8.4s, %25.4h, #16          \n"
-                        "shll2  v9.4s, %25.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v2.s[3]      \n"
-
-                        "fadd   v11.4s, v10.4s, v11.4s      \n"
-
-                        "add    %1, %1, #16                 \n"
-                        "fadd   v13.4s, v12.4s, v13.4s      \n"
-
-                        "add    %2, %2, #16                 \n"
-                        "fadd   v13.4s, v11.4s, v13.4s      \n"
-
-                        "add    %3, %3, #16                 \n"
-
-                        "st1    {v13.4s}, [%0], #16         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00_01), // %8
-                        "w"(_k00_23), // %9
-                        "w"(_k01_01), // %10
-                        "w"(_k01_23), // %11
-                        "w"(_k02_01), // %12
-                        "w"(_k02_23), // %13
-                        "w"(_k10_01), // %14
-                        "w"(_k10_23), // %15
-                        "w"(_k11_01), // %16
-                        "w"(_k11_23), // %17
-                        "w"(_k12_01), // %18
-                        "w"(_k12_23), // %19
-                        "w"(_k20_01), // %20
-                        "w"(_k20_23), // %21
-                        "w"(_k21_01), // %22
-                        "w"(_k21_23), // %23
-                        "w"(_k22_01), // %24
-                        "w"(_k22_23)  // %25
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #128]          \n"
-                        "vld1.f32   {d30-d31}, [%0 :128] \n" // sum0
-
-                        "pld        [%1, #192]          \n"
-                        "vld1.u16   {d2-d4}, [%1 :64]   \n" // r00 r01 r02
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vshll.u16  q2, d4, #16         \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmul.f32   q12, q8, d0[0]      \n"
-                        "vmul.f32   q13, q9, d0[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmul.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q11, d3[1]     \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%2, #192]          \n"
-                        "vld1.u16   {d2-d4}, [%2 :64]   \n" // r10 r11 r12
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vshll.u16  q2, d4, #16         \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q9, d0[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q11, d3[1]     \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%3, #192]          \n"
-                        "vld1.u16   {d2-d4}, [%3 :64]   \n" // r20 r21 r22
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vshll.u16  q2, d4, #16         \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q9, d0[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q11, d3[1]     \n"
-
-                        //                         "pld        [%4, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%4 :128] \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "add        %1, %1, #16         \n"
-                        "vadd.f32   q13, q12, q13       \n"
-
-                        "add        %2, %2, #16         \n"
-                        "vadd.f32   q15, q14, q15       \n"
-
-                        "add        %3, %3, #16         \n"
-                        "vadd.f32   q15, q13, q15       \n"
-
-                        "sub        %4, %4, #256        \n" // kptr -= 8 * 16 * 2;
-
-                        "vst1.f32   {d30-d31}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(kptr)     // %4
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-        }
-        for (; q < inch; q++)
-        {
-            unsigned short* outptr0_bf16 = top_blob.channel(p);
-
-            const float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<const unsigned short>(0);
-            const unsigned short* r1 = img0.row<const unsigned short>(1);
-            const unsigned short* r2 = img0.row<const unsigned short>(2);
-
-            const unsigned short* kptr = (const unsigned short*)kernel.channel(p).row<const unsigned short>(q);
-
-#if __aarch64__
-            // 16 * 9
-            uint16x8_t _k00_01 = vld1q_u16(kptr);
-            uint16x8_t _k00_23 = vld1q_u16(kptr + 8);
-            uint16x8_t _k01_01 = vld1q_u16(kptr + 16);
-            uint16x8_t _k01_23 = vld1q_u16(kptr + 24);
-            uint16x8_t _k02_01 = vld1q_u16(kptr + 32);
-            uint16x8_t _k02_23 = vld1q_u16(kptr + 40);
-            uint16x8_t _k10_01 = vld1q_u16(kptr + 48);
-            uint16x8_t _k10_23 = vld1q_u16(kptr + 56);
-            uint16x8_t _k11_01 = vld1q_u16(kptr + 64);
-            uint16x8_t _k11_23 = vld1q_u16(kptr + 72);
-            uint16x8_t _k12_01 = vld1q_u16(kptr + 80);
-            uint16x8_t _k12_23 = vld1q_u16(kptr + 88);
-            uint16x8_t _k20_01 = vld1q_u16(kptr + 96);
-            uint16x8_t _k20_23 = vld1q_u16(kptr + 104);
-            uint16x8_t _k21_01 = vld1q_u16(kptr + 112);
-            uint16x8_t _k21_23 = vld1q_u16(kptr + 120);
-            uint16x8_t _k22_01 = vld1q_u16(kptr + 128);
-            uint16x8_t _k22_23 = vld1q_u16(kptr + 136);
-#endif // __aarch64__
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%1], #64 \n" // sum0 sum1 sum2 sum3
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%2], #64 \n" // r00 r01 r02 r03
-
-                        "shll   v0.4s, v4.4h, #16           \n"
-                        "shll2  v1.4s, v4.8h, #16           \n"
-                        "shll   v2.4s, v5.4h, #16           \n"
-                        "shll2  v3.4s, v5.8h, #16           \n"
-
-                        "shll   v4.4s, v6.4h, #16           \n"
-                        "shll2  v5.4s, v6.8h, #16           \n"
-                        "shll   v6.4s, v7.4h, #16           \n"
-                        "shll2  v7.4s, v7.8h, #16           \n"
-
-                        "shll   v8.4s, %10.4h, #16          \n"
-                        "shll2  v9.4s, %10.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v4.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v6.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v0.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v2.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v4.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v6.s[1]      \n"
-
-                        "shll   v8.4s, %11.4h, #16          \n"
-                        "shll2  v9.4s, %11.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v6.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v0.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v4.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v6.s[3]      \n"
-
-                        "shll   v8.4s, %12.4h, #16          \n"
-                        "shll2  v9.4s, %12.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v5.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v7.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v1.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v3.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v5.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v7.s[1]      \n"
-
-                        "shll   v8.4s, %13.4h, #16          \n"
-                        "shll2  v9.4s, %13.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v5.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v7.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v1.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v3.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v5.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v7.s[3]      \n"
-
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v0.4h}, [%2]               \n" // r08
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "shll   v8.4s, %14.4h, #16          \n"
-                        "shll2  v9.4s, %14.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v6.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v0.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v2.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v4.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v6.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[1]      \n"
-
-                        "shll   v8.4s, %15.4h, #16          \n"
-                        "shll2  v9.4s, %15.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v6.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v4.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v6.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[3]      \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%3], #64 \n" // r10 r11 r12 r13
-
-                        "shll   v0.4s, v4.4h, #16           \n"
-                        "shll2  v1.4s, v4.8h, #16           \n"
-                        "shll   v2.4s, v5.4h, #16           \n"
-                        "shll2  v3.4s, v5.8h, #16           \n"
-
-                        "shll   v4.4s, v6.4h, #16           \n"
-                        "shll2  v5.4s, v6.8h, #16           \n"
-                        "shll   v6.4s, v7.4h, #16           \n"
-                        "shll2  v7.4s, v7.8h, #16           \n"
-
-                        "shll   v8.4s, %16.4h, #16          \n"
-                        "shll2  v9.4s, %16.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v4.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v6.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v0.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v2.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v4.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v6.s[1]      \n"
-
-                        "shll   v8.4s, %17.4h, #16          \n"
-                        "shll2  v9.4s, %17.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v6.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v0.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v4.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v6.s[3]      \n"
-
-                        "shll   v8.4s, %18.4h, #16          \n"
-                        "shll2  v9.4s, %18.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v5.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v7.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v1.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v3.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v5.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v7.s[1]      \n"
-
-                        "shll   v8.4s, %19.4h, #16          \n"
-                        "shll2  v9.4s, %19.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v5.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v7.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v1.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v3.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v5.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v7.s[3]      \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v0.4h}, [%3]               \n" // r18
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "shll   v8.4s, %20.4h, #16          \n"
-                        "shll2  v9.4s, %20.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v6.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v0.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v2.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v4.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v6.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[1]      \n"
-
-                        "shll   v8.4s, %21.4h, #16          \n"
-                        "shll2  v9.4s, %21.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v6.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v4.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v6.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[3]      \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%4], #64 \n" // r20 r21 r22 r23
-
-                        "shll   v0.4s, v4.4h, #16           \n"
-                        "shll2  v1.4s, v4.8h, #16           \n"
-                        "shll   v2.4s, v5.4h, #16           \n"
-                        "shll2  v3.4s, v5.8h, #16           \n"
-
-                        "shll   v4.4s, v6.4h, #16           \n"
-                        "shll2  v5.4s, v6.8h, #16           \n"
-                        "shll   v6.4s, v7.4h, #16           \n"
-                        "shll2  v7.4s, v7.8h, #16           \n"
-
-                        "shll   v8.4s, %22.4h, #16          \n"
-                        "shll2  v9.4s, %22.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v4.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v6.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v0.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v2.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v4.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v6.s[1]      \n"
-
-                        "shll   v8.4s, %23.4h, #16          \n"
-                        "shll2  v9.4s, %23.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v6.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v0.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v4.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v6.s[3]      \n"
-
-                        "shll   v8.4s, %24.4h, #16          \n"
-                        "shll2  v9.4s, %24.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v5.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v7.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v1.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v3.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v5.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v7.s[1]      \n"
-
-                        "shll   v8.4s, %25.4h, #16          \n"
-                        "shll2  v9.4s, %25.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v5.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v7.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v1.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v3.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v5.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v7.s[3]      \n"
-
-                        "prfm   pldl1keep, [%4, #64]        \n"
-                        "ld1    {v0.4h}, [%4]               \n" // r28
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "shll   v8.4s, %26.4h, #16          \n"
-                        "shll2  v9.4s, %26.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[0]      \n"
-                        "fmla   v12.4s, v8.4s, v6.s[0]      \n"
-                        "fmla   v13.4s, v8.4s, v0.s[0]      \n"
-                        "fmla   v10.4s, v9.4s, v2.s[1]      \n"
-                        "fmla   v11.4s, v9.4s, v4.s[1]      \n"
-                        "fmla   v12.4s, v9.4s, v6.s[1]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[1]      \n"
-
-                        "shll   v8.4s, %27.4h, #16          \n"
-                        "shll2  v9.4s, %27.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v12.4s, v8.4s, v6.s[2]      \n"
-                        "fmla   v13.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v10.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v11.4s, v9.4s, v4.s[3]      \n"
-                        "fmla   v12.4s, v9.4s, v6.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[3]      \n"
-
-                        "shrn   v10.4h, v10.4s, #16         \n"
-                        "shrn   v11.4h, v11.4s, #16         \n"
-                        "shrn   v12.4h, v12.4s, #16         \n"
-                        "shrn   v13.4h, v13.4s, #16         \n"
-
-                        "st1    {v10.4h, v11.4h, v12.4h, v13.4h}, [%0], #32 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2)            // %4
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00_01), // %10
-                        "w"(_k00_23), // %11
-                        "w"(_k01_01), // %12
-                        "w"(_k01_23), // %13
-                        "w"(_k02_01), // %14
-                        "w"(_k02_23), // %15
-                        "w"(_k10_01), // %16
-                        "w"(_k10_23), // %17
-                        "w"(_k11_01), // %18
-                        "w"(_k11_23), // %19
-                        "w"(_k12_01), // %20
-                        "w"(_k12_23), // %21
-                        "w"(_k20_01), // %22
-                        "w"(_k20_23), // %23
-                        "w"(_k21_01), // %24
-                        "w"(_k21_23), // %25
-                        "w"(_k22_01), // %26
-                        "w"(_k22_23)  // %27
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1!, {d24-d31}      \n" // sum0 sum1 sum2 sum3
-
-                        "pld        [%2, #512]          \n"
-                        "vldm       %2!, {d8-d15}       \n" // r00 r01 r02 r03 r04 r05 r06 r07
-
-                        "vshll.u16  q0, d8, #16         \n"
-                        "vshll.u16  q1, d9, #16         \n"
-                        "vshll.u16  q2, d10, #16        \n"
-                        "vshll.u16  q3, d11, #16        \n"
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%2, #64]           \n"
-                        "vld1.f32   {d1}, [%2 :64]      \n" // r08
-
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%3, #512]          \n"
-                        "vldm       %3!, {d8-d15}       \n" // r10 r11 r12 r13 r14 r15 r16 r17
-
-                        "vshll.u16  q0, d8, #16         \n"
-                        "vshll.u16  q1, d9, #16         \n"
-                        "vshll.u16  q2, d10, #16        \n"
-                        "vshll.u16  q3, d11, #16        \n"
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%3, #64]           \n"
-                        "vld1.f32   {d1}, [%3 :64]      \n" // r18
-
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vldm       %4!, {d8-d15}       \n" // r20 r21 r22 r23 r24 r25 r26 r27
-
-                        "vshll.u16  q0, d8, #16         \n"
-                        "vshll.u16  q1, d9, #16         \n"
-                        "vshll.u16  q2, d10, #16        \n"
-                        "vshll.u16  q3, d11, #16        \n"
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%4, #64]           \n"
-                        "vld1.f32   {d1}, [%4 :64]      \n" // r28
-
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        //                         "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128] \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "sub        %5, %5, #256        \n" // kptr -= 8 * 16;
-
-                        "vshrn.u32  d24, q12, #16       \n"
-                        "vshrn.u32  d25, q13, #16       \n"
-                        "vshrn.u32  d26, q14, #16       \n"
-                        "vshrn.u32  d27, q15, #16       \n"
-
-                        "vst1.f32   {d24-d27}, [%0 :64]! \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(kptr)          // %5
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v12.4s, v13.4s}, [%1], #32 \n" // sum0 sum1
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n" // r00 r01 r02 r03
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "shll   v6.4s, %10.4h, #16          \n"
-                        "shll2  v7.4s, %10.8h, #16          \n"
-
-                        "fmul   v10.4s, v6.4s, v0.s[0]      \n"
-                        "fmul   v11.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v0.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v2.s[1]      \n"
-
-                        "shll   v8.4s, %11.4h, #16          \n"
-                        "shll2  v9.4s, %11.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v0.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v2.s[3]      \n"
-
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v4.4h}, [%2]               \n" // r04
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "shll   v6.4s, %12.4h, #16          \n"
-                        "shll2  v7.4s, %12.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v3.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v1.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v3.s[1]      \n"
-
-                        "shll   v8.4s, %13.4h, #16          \n"
-                        "shll2  v9.4s, %13.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v1.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v3.s[3]      \n"
-
-                        "shll   v6.4s, %14.4h, #16          \n"
-                        "shll2  v7.4s, %14.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v4.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v2.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v4.s[1]      \n"
-
-                        "shll   v8.4s, %15.4h, #16          \n"
-                        "shll2  v9.4s, %15.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v4.s[3]      \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%3], #32 \n" // r10 r11 r12 r13
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "shll   v6.4s, %16.4h, #16          \n"
-                        "shll2  v7.4s, %16.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v0.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v0.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v2.s[1]      \n"
-
-                        "shll   v8.4s, %17.4h, #16          \n"
-                        "shll2  v9.4s, %17.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v0.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v2.s[3]      \n"
-
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v4.4h}, [%3]               \n" // r14
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "shll   v6.4s, %18.4h, #16          \n"
-                        "shll2  v7.4s, %18.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v3.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v1.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v3.s[1]      \n"
-
-                        "shll   v8.4s, %19.4h, #16          \n"
-                        "shll2  v9.4s, %19.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v1.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v3.s[3]      \n"
-
-                        "shll   v6.4s, %20.4h, #16          \n"
-                        "shll2  v7.4s, %20.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v4.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v2.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v4.s[1]      \n"
-
-                        "shll   v8.4s, %21.4h, #16          \n"
-                        "shll2  v9.4s, %21.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v4.s[3]      \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%4], #32 \n" // r20 r21 r22 r23
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "shll   v6.4s, %22.4h, #16          \n"
-                        "shll2  v7.4s, %22.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v0.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v0.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v2.s[1]      \n"
-
-                        "shll   v8.4s, %23.4h, #16          \n"
-                        "shll2  v9.4s, %23.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v0.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v2.s[3]      \n"
-
-                        "prfm   pldl1keep, [%4, #64]        \n"
-                        "ld1    {v4.4h}, [%4]               \n" // r24
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "shll   v6.4s, %24.4h, #16          \n"
-                        "shll2  v7.4s, %24.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v3.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v1.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v3.s[1]      \n"
-
-                        "shll   v8.4s, %25.4h, #16          \n"
-                        "shll2  v9.4s, %25.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v3.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v1.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v3.s[3]      \n"
-
-                        "shll   v6.4s, %26.4h, #16          \n"
-                        "shll2  v7.4s, %26.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v6.4s, v4.s[0]      \n"
-                        "fmla   v12.4s, v7.4s, v2.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v4.s[1]      \n"
-
-                        "shll   v8.4s, %27.4h, #16          \n"
-                        "shll2  v9.4s, %27.8h, #16          \n"
-
-                        "fmla   v10.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v11.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v12.4s, v9.4s, v2.s[3]      \n"
-                        "fmla   v13.4s, v9.4s, v4.s[3]      \n"
-
-                        "fadd   v12.4s, v10.4s, v12.4s      \n"
-                        "fadd   v13.4s, v11.4s, v13.4s      \n"
-
-                        "shrn   v12.4h, v12.4s, #16         \n"
-                        "shrn   v13.4h, v13.4s, #16         \n"
-
-                        "st1    {v12.4h, v13.4h}, [%0], #16 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2)            // %4
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00_01), // %10
-                        "w"(_k00_23), // %11
-                        "w"(_k01_01), // %12
-                        "w"(_k01_23), // %13
-                        "w"(_k02_01), // %14
-                        "w"(_k02_23), // %15
-                        "w"(_k10_01), // %16
-                        "w"(_k10_23), // %17
-                        "w"(_k11_01), // %18
-                        "w"(_k11_23), // %19
-                        "w"(_k12_01), // %20
-                        "w"(_k12_23), // %21
-                        "w"(_k20_01), // %22
-                        "w"(_k20_23), // %23
-                        "w"(_k21_01), // %24
-                        "w"(_k21_23), // %25
-                        "w"(_k22_01), // %26
-                        "w"(_k22_23)  // %27
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #256]          \n"
-                        "vld1.f32   {d28-d31}, [%1 :128]! \n" // sum0 sum1
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%2 :64]!  \n" // r00 r01 r02 r03
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmul.f32   q12, q8, d0[0]      \n"
-                        "vmul.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%2, #64]           \n"
-                        "vld1.f32   {d9}, [%2 :64]      \n" // r04
-
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q11, d3[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%3 :64]!  \n" // r10 r11 r12 r13
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%3, #64]           \n"
-                        "vld1.f32   {d9}, [%3 :64]      \n" // r14
-
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q11, d3[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%4 :64]!  \n" // r20 r21 r22 r23
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%4, #64]           \n"
-                        "vld1.f32   {d9}, [%4 :64]      \n" // r24
-
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q11, d3[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        //                         "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128] \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "vadd.f32   q14, q12, q14       \n"
-                        "vadd.f32   q15, q13, q15       \n"
-
-                        "sub        %5, %5, #256        \n" // kptr -= 8 * 16;
-
-                        "vshrn.u32  d28, q14, #16       \n"
-                        "vshrn.u32  d29, q15, #16       \n"
-
-                        "vst1.f32   {d28-d29}, [%0 :64]! \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(kptr)          // %5
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v13.4s}, [%1], #16         \n" // sum0
-
-                        "prfm   pldl1keep, [%2, #192]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%2] \n" // r00 r01 r02
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "shll   v6.4s, %10.4h, #16          \n"
-                        "shll2  v7.4s, %10.8h, #16          \n"
-
-                        "fmul   v10.4s, v6.4s, v0.s[0]      \n"
-                        "fmul   v11.4s, v7.4s, v0.s[1]      \n"
-
-                        "shll   v8.4s, %11.4h, #16          \n"
-                        "shll2  v9.4s, %11.8h, #16          \n"
-
-                        "fmul   v12.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[3]      \n"
-
-                        "shll   v6.4s, %12.4h, #16          \n"
-                        "shll2  v7.4s, %12.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v1.s[1]      \n"
-
-                        "shll   v8.4s, %13.4h, #16          \n"
-                        "shll2  v9.4s, %13.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v1.s[3]      \n"
-
-                        "shll   v6.4s, %14.4h, #16          \n"
-                        "shll2  v7.4s, %14.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v2.s[1]      \n"
-
-                        "shll   v8.4s, %15.4h, #16          \n"
-                        "shll2  v9.4s, %15.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v2.s[3]      \n"
-
-                        "prfm   pldl1keep, [%3, #192]       \n"
-                        "ld1    {v3.4h, v4.4h, v5.4h}, [%3] \n" // r10 r11 r12
-
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "shll   v6.4s, %16.4h, #16          \n"
-                        "shll2  v7.4s, %16.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v3.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v3.s[1]      \n"
-
-                        "shll   v8.4s, %17.4h, #16          \n"
-                        "shll2  v9.4s, %17.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v3.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v3.s[3]      \n"
-
-                        "shll   v6.4s, %18.4h, #16          \n"
-                        "shll2  v7.4s, %18.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v4.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v4.s[1]      \n"
-
-                        "shll   v8.4s, %19.4h, #16          \n"
-                        "shll2  v9.4s, %19.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v4.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v4.s[3]      \n"
-
-                        "shll   v6.4s, %20.4h, #16          \n"
-                        "shll2  v7.4s, %20.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v5.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v5.s[1]      \n"
-
-                        "shll   v8.4s, %21.4h, #16          \n"
-                        "shll2  v9.4s, %21.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v5.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v5.s[3]      \n"
-
-                        "prfm   pldl1keep, [%4, #192]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%4] \n" // r20 r21 r22
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "shll   v6.4s, %22.4h, #16          \n"
-                        "shll2  v7.4s, %22.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v0.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v0.s[1]      \n"
-
-                        "shll   v8.4s, %23.4h, #16          \n"
-                        "shll2  v9.4s, %23.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v0.s[3]      \n"
-
-                        "shll   v6.4s, %24.4h, #16          \n"
-                        "shll2  v7.4s, %24.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v1.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v1.s[1]      \n"
-
-                        "shll   v8.4s, %25.4h, #16          \n"
-                        "shll2  v9.4s, %25.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v1.s[3]      \n"
-
-                        "shll   v6.4s, %26.4h, #16          \n"
-                        "shll2  v7.4s, %26.8h, #16          \n"
-
-                        "fmla   v10.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v11.4s, v7.4s, v2.s[1]      \n"
-
-                        "shll   v8.4s, %27.4h, #16          \n"
-                        "shll2  v9.4s, %27.8h, #16          \n"
-
-                        "fmla   v12.4s, v8.4s, v2.s[2]      \n"
-                        "fmla   v13.4s, v9.4s, v2.s[3]      \n"
-
-                        "fadd   v11.4s, v10.4s, v11.4s      \n"
-
-                        "add    %2, %2, #16                 \n"
-                        "fadd   v13.4s, v12.4s, v13.4s      \n"
-
-                        "add    %3, %3, #16                 \n"
-                        "fadd   v13.4s, v11.4s, v13.4s      \n"
-
-                        "add    %4, %4, #16                 \n"
-                        "shrn   v13.4h, v13.4s, #16         \n"
-
-                        "st1    {v13.4h}, [%0], #8          \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2)            // %4
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00_01), // %10
-                        "w"(_k00_23), // %11
-                        "w"(_k01_01), // %12
-                        "w"(_k01_23), // %13
-                        "w"(_k02_01), // %14
-                        "w"(_k02_23), // %15
-                        "w"(_k10_01), // %16
-                        "w"(_k10_23), // %17
-                        "w"(_k11_01), // %18
-                        "w"(_k11_23), // %19
-                        "w"(_k12_01), // %20
-                        "w"(_k12_23), // %21
-                        "w"(_k20_01), // %22
-                        "w"(_k20_23), // %23
-                        "w"(_k21_01), // %24
-                        "w"(_k21_23), // %25
-                        "w"(_k22_01), // %26
-                        "w"(_k22_23)  // %27
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d30-d31}, [%1 :128]! \n" // sum0
-
-                        "pld        [%2, #192]          \n"
-                        "vld1.u16   {d2-d4}, [%2 :64]   \n" // r00 r01 r02
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vshll.u16  q2, d4, #16         \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmul.f32   q12, q8, d0[0]      \n"
-                        "vmul.f32   q13, q9, d0[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmul.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q11, d3[1]     \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%3, #192]          \n"
-                        "vld1.u16   {d2-d4}, [%3 :64]   \n" // r10 r11 r12
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vshll.u16  q2, d4, #16         \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q9, d0[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q11, d3[1]     \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%4, #192]          \n"
-                        "vld1.u16   {d2-d4}, [%4 :64]   \n" // r20 r21 r22
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vshll.u16  q2, d4, #16         \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q9, d0[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q11, d3[1]     \n"
-
-                        //                         "pld        [%5, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%5 :128] \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "add        %2, %2, #16         \n"
-                        "vadd.f32   q13, q12, q13       \n"
-
-                        "add        %3, %3, #16         \n"
-                        "vadd.f32   q15, q14, q15       \n"
-
-                        "add        %4, %4, #16         \n"
-                        "vadd.f32   q15, q13, q15       \n"
-
-                        "sub        %5, %5, #256        \n" // kptr -= 8 * 16 * 2;
-
-                        "vshrn.u32  d31, q15, #16       \n"
-
-                        "vst1.u16   {d31}, [%0 :64]!    \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(kptr)          // %5
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4_fp16s.h
deleted file mode 100644
index 52ef0271..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4_fp16s.h
+++ /dev/null
@@ -1,574 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack4_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float16x4_t _bias0 = bias ? vld1_f16(bias + p * 4) : vdup_n_f16((__fp16)0.f);
-        out0.fill(_bias0);
-
-        int q = 0;
-        for (; q < inch; q++)
-        {
-            __fp16* outptr0 = out0.row<__fp16>(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const __fp16* r0 = img0.row<const __fp16>(0);
-            const __fp16* r1 = img0.row<const __fp16>(1);
-            const __fp16* r2 = img0.row<const __fp16>(2);
-
-            const __fp16* kptr = kernel.channel(p).row<const __fp16>(q);
-
-            // 16 * 9
-            float16x8_t _k00_01 = vld1q_f16(kptr);
-            float16x8_t _k00_23 = vld1q_f16(kptr + 8);
-            float16x8_t _k01_01 = vld1q_f16(kptr + 16);
-            float16x8_t _k01_23 = vld1q_f16(kptr + 24);
-            float16x8_t _k02_01 = vld1q_f16(kptr + 32);
-            float16x8_t _k02_23 = vld1q_f16(kptr + 40);
-            float16x8_t _k10_01 = vld1q_f16(kptr + 48);
-            float16x8_t _k10_23 = vld1q_f16(kptr + 56);
-            float16x8_t _k11_01 = vld1q_f16(kptr + 64);
-            float16x8_t _k11_23 = vld1q_f16(kptr + 72);
-            float16x8_t _k12_01 = vld1q_f16(kptr + 80);
-            float16x8_t _k12_23 = vld1q_f16(kptr + 88);
-            float16x8_t _k20_01 = vld1q_f16(kptr + 96);
-            float16x8_t _k20_23 = vld1q_f16(kptr + 104);
-            float16x8_t _k21_01 = vld1q_f16(kptr + 112);
-            float16x8_t _k21_23 = vld1q_f16(kptr + 120);
-            float16x8_t _k22_01 = vld1q_f16(kptr + 128);
-            float16x8_t _k22_23 = vld1q_f16(kptr + 136);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v10.4h, v11.4h, v12.4h, v13.4h}, [%0] \n" // sum0 sum1 sum2 sum3
-
-                        "prfm   pldl1keep, [%1, #384]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h}, [%1] \n" // r00 r01 r02 r03 r04 r05
-
-                        "ext    v6.16b, %8.16b, %8.16b, #8  \n"
-                        "fmla   v10.4h, %8.4h, v0.h[0]      \n"
-                        "fmla   v11.4h, %8.4h, v0.h[4]      \n"
-                        "fmla   v12.4h, %8.4h, v1.h[0]      \n"
-                        "fmla   v13.4h, %8.4h, v1.h[4]      \n"
-                        "fmla   v10.4h, v6.4h, v0.h[1]      \n"
-                        "fmla   v11.4h, v6.4h, v0.h[5]      \n"
-                        "fmla   v12.4h, v6.4h, v1.h[1]      \n"
-                        "fmla   v13.4h, v6.4h, v1.h[5]      \n"
-                        "ext    v7.16b, %9.16b, %9.16b, #8  \n"
-                        "fmla   v10.4h, %9.4h, v0.h[2]      \n"
-                        "fmla   v11.4h, %9.4h, v0.h[6]      \n"
-                        "fmla   v12.4h, %9.4h, v1.h[2]      \n"
-                        "fmla   v13.4h, %9.4h, v1.h[6]      \n"
-                        "fmla   v10.4h, v7.4h, v0.h[3]      \n"
-                        "fmla   v11.4h, v7.4h, v0.h[7]      \n"
-                        "fmla   v12.4h, v7.4h, v1.h[3]      \n"
-                        "fmla   v13.4h, v7.4h, v1.h[7]      \n"
-
-                        "ext    v8.16b, %10.16b, %10.16b, #8 \n"
-                        "fmla   v10.4h, %10.4h, v0.h[4]     \n"
-                        "fmla   v11.4h, %10.4h, v1.h[0]     \n"
-                        "fmla   v12.4h, %10.4h, v1.h[4]     \n"
-                        "fmla   v13.4h, %10.4h, v2.h[0]     \n"
-                        "fmla   v10.4h, v8.4h, v0.h[5]      \n"
-                        "fmla   v11.4h, v8.4h, v1.h[1]      \n"
-                        "fmla   v12.4h, v8.4h, v1.h[5]      \n"
-                        "fmla   v13.4h, v8.4h, v2.h[1]      \n"
-                        "ext    v9.16b, %11.16b, %11.16b, #8 \n"
-                        "fmla   v10.4h, %11.4h, v0.h[6]     \n"
-                        "fmla   v11.4h, %11.4h, v1.h[2]     \n"
-                        "fmla   v12.4h, %11.4h, v1.h[6]     \n"
-                        "fmla   v13.4h, %11.4h, v2.h[2]     \n"
-                        "fmla   v10.4h, v9.4h, v0.h[7]      \n"
-                        "fmla   v11.4h, v9.4h, v1.h[3]      \n"
-                        "fmla   v12.4h, v9.4h, v1.h[7]      \n"
-                        "fmla   v13.4h, v9.4h, v2.h[3]      \n"
-
-                        "prfm   pldl1keep, [%2, #384]       \n"
-                        "ld1    {v3.8h, v4.8h, v5.8h}, [%2] \n" // r10 r11 r12 r13 r14 r15
-
-                        "ext    v6.16b, %12.16b, %12.16b, #8 \n"
-                        "fmla   v10.4h, %12.4h, v1.h[0]     \n"
-                        "fmla   v11.4h, %12.4h, v1.h[4]     \n"
-                        "fmla   v12.4h, %12.4h, v2.h[0]     \n"
-                        "fmla   v13.4h, %12.4h, v2.h[4]     \n"
-                        "fmla   v10.4h, v6.4h, v1.h[1]      \n"
-                        "fmla   v11.4h, v6.4h, v1.h[5]      \n"
-                        "fmla   v12.4h, v6.4h, v2.h[1]      \n"
-                        "fmla   v13.4h, v6.4h, v2.h[5]      \n"
-                        "ext    v7.16b, %13.16b, %13.16b, #8 \n"
-                        "fmla   v10.4h, %13.4h, v1.h[2]     \n"
-                        "fmla   v11.4h, %13.4h, v1.h[6]     \n"
-                        "fmla   v12.4h, %13.4h, v2.h[2]     \n"
-                        "fmla   v13.4h, %13.4h, v2.h[6]     \n"
-                        "fmla   v10.4h, v7.4h, v1.h[3]      \n"
-                        "fmla   v11.4h, v7.4h, v1.h[7]      \n"
-                        "fmla   v12.4h, v7.4h, v2.h[3]      \n"
-                        "fmla   v13.4h, v7.4h, v2.h[7]      \n"
-
-                        "ext    v8.16b, %14.16b, %14.16b, #8 \n"
-                        "fmla   v10.4h, %14.4h, v3.h[0]     \n"
-                        "fmla   v11.4h, %14.4h, v3.h[4]     \n"
-                        "fmla   v12.4h, %14.4h, v4.h[0]     \n"
-                        "fmla   v13.4h, %14.4h, v4.h[4]     \n"
-                        "fmla   v10.4h, v8.4h, v3.h[1]      \n"
-                        "fmla   v11.4h, v8.4h, v3.h[5]      \n"
-                        "fmla   v12.4h, v8.4h, v4.h[1]      \n"
-                        "fmla   v13.4h, v8.4h, v4.h[5]      \n"
-                        "ext    v9.16b, %15.16b, %15.16b, #8 \n"
-                        "fmla   v10.4h, %15.4h, v3.h[2]     \n"
-                        "fmla   v11.4h, %15.4h, v3.h[6]     \n"
-                        "fmla   v12.4h, %15.4h, v4.h[2]     \n"
-                        "fmla   v13.4h, %15.4h, v4.h[6]     \n"
-                        "fmla   v10.4h, v9.4h, v3.h[3]      \n"
-                        "fmla   v11.4h, v9.4h, v3.h[7]      \n"
-                        "fmla   v12.4h, v9.4h, v4.h[3]      \n"
-                        "fmla   v13.4h, v9.4h, v4.h[7]      \n"
-
-                        "ext    v6.16b, %16.16b, %16.16b, #8 \n"
-                        "fmla   v10.4h, %16.4h, v3.h[4]     \n"
-                        "fmla   v11.4h, %16.4h, v4.h[0]     \n"
-                        "fmla   v12.4h, %16.4h, v4.h[4]     \n"
-                        "fmla   v13.4h, %16.4h, v5.h[0]     \n"
-                        "fmla   v10.4h, v6.4h, v3.h[5]      \n"
-                        "fmla   v11.4h, v6.4h, v4.h[1]      \n"
-                        "fmla   v12.4h, v6.4h, v4.h[5]      \n"
-                        "fmla   v13.4h, v6.4h, v5.h[1]      \n"
-                        "ext    v7.16b, %17.16b, %17.16b, #8 \n"
-                        "fmla   v10.4h, %17.4h, v3.h[6]     \n"
-                        "fmla   v11.4h, %17.4h, v4.h[2]     \n"
-                        "fmla   v12.4h, %17.4h, v4.h[6]     \n"
-                        "fmla   v13.4h, %17.4h, v5.h[2]     \n"
-                        "fmla   v10.4h, v7.4h, v3.h[7]      \n"
-                        "fmla   v11.4h, v7.4h, v4.h[3]      \n"
-                        "fmla   v12.4h, v7.4h, v4.h[7]      \n"
-                        "fmla   v13.4h, v7.4h, v5.h[3]      \n"
-
-                        "prfm   pldl1keep, [%3, #384]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h}, [%3] \n" // r20 r21 r22 r23 r24 r25
-
-                        "ext    v8.16b, %18.16b, %18.16b, #8 \n"
-                        "fmla   v10.4h, %18.4h, v4.h[0]     \n"
-                        "fmla   v11.4h, %18.4h, v4.h[4]     \n"
-                        "fmla   v12.4h, %18.4h, v5.h[0]     \n"
-                        "fmla   v13.4h, %18.4h, v5.h[4]     \n"
-                        "fmla   v10.4h, v8.4h, v4.h[1]      \n"
-                        "fmla   v11.4h, v8.4h, v4.h[5]      \n"
-                        "fmla   v12.4h, v8.4h, v5.h[1]      \n"
-                        "fmla   v13.4h, v8.4h, v5.h[5]      \n"
-                        "ext    v9.16b, %19.16b, %19.16b, #8 \n"
-                        "fmla   v10.4h, %19.4h, v4.h[2]     \n"
-                        "fmla   v11.4h, %19.4h, v4.h[6]     \n"
-                        "fmla   v12.4h, %19.4h, v5.h[2]     \n"
-                        "fmla   v13.4h, %19.4h, v5.h[6]     \n"
-                        "fmla   v10.4h, v9.4h, v4.h[3]      \n"
-                        "fmla   v11.4h, v9.4h, v4.h[7]      \n"
-                        "fmla   v12.4h, v9.4h, v5.h[3]      \n"
-                        "fmla   v13.4h, v9.4h, v5.h[7]      \n"
-
-                        "ext    v6.16b, %20.16b, %20.16b, #8 \n"
-                        "fmla   v10.4h, %20.4h, v0.h[0]     \n"
-                        "fmla   v11.4h, %20.4h, v0.h[4]     \n"
-                        "fmla   v12.4h, %20.4h, v1.h[0]     \n"
-                        "fmla   v13.4h, %20.4h, v1.h[4]     \n"
-                        "fmla   v10.4h, v6.4h, v0.h[1]      \n"
-                        "fmla   v11.4h, v6.4h, v0.h[5]      \n"
-                        "fmla   v12.4h, v6.4h, v1.h[1]      \n"
-                        "fmla   v13.4h, v6.4h, v1.h[5]      \n"
-                        "ext    v7.16b, %21.16b, %21.16b, #8 \n"
-                        "fmla   v10.4h, %21.4h, v0.h[2]     \n"
-                        "fmla   v11.4h, %21.4h, v0.h[6]     \n"
-                        "fmla   v12.4h, %21.4h, v1.h[2]     \n"
-                        "fmla   v13.4h, %21.4h, v1.h[6]     \n"
-                        "fmla   v10.4h, v7.4h, v0.h[3]      \n"
-                        "fmla   v11.4h, v7.4h, v0.h[7]      \n"
-                        "fmla   v12.4h, v7.4h, v1.h[3]      \n"
-                        "fmla   v13.4h, v7.4h, v1.h[7]      \n"
-
-                        "ext    v8.16b, %22.16b, %22.16b, #8 \n"
-                        "fmla   v10.4h, %22.4h, v0.h[4]     \n"
-                        "fmla   v11.4h, %22.4h, v1.h[0]     \n"
-                        "fmla   v12.4h, %22.4h, v1.h[4]     \n"
-                        "fmla   v13.4h, %22.4h, v2.h[0]     \n"
-                        "fmla   v10.4h, v8.4h, v0.h[5]      \n"
-                        "fmla   v11.4h, v8.4h, v1.h[1]      \n"
-                        "fmla   v12.4h, v8.4h, v1.h[5]      \n"
-                        "fmla   v13.4h, v8.4h, v2.h[1]      \n"
-                        "ext    v9.16b, %23.16b, %23.16b, #8 \n"
-                        "fmla   v10.4h, %23.4h, v0.h[6]     \n"
-                        "fmla   v11.4h, %23.4h, v1.h[2]     \n"
-                        "fmla   v12.4h, %23.4h, v1.h[6]     \n"
-                        "fmla   v13.4h, %23.4h, v2.h[2]     \n"
-                        "fmla   v10.4h, v9.4h, v0.h[7]      \n"
-                        "fmla   v11.4h, v9.4h, v1.h[3]      \n"
-                        "fmla   v12.4h, v9.4h, v1.h[7]      \n"
-                        "fmla   v13.4h, v9.4h, v2.h[3]      \n"
-
-                        "ext    v6.16b, %24.16b, %24.16b, #8 \n"
-                        "fmla   v10.4h, %24.4h, v1.h[0]     \n"
-                        "fmla   v11.4h, %24.4h, v1.h[4]     \n"
-                        "fmla   v12.4h, %24.4h, v2.h[0]     \n"
-                        "fmla   v13.4h, %24.4h, v2.h[4]     \n"
-
-                        "add    %1, %1, #32                 \n"
-
-                        "fmla   v10.4h, v6.4h, v1.h[1]      \n"
-                        "fmla   v11.4h, v6.4h, v1.h[5]      \n"
-                        "fmla   v12.4h, v6.4h, v2.h[1]      \n"
-                        "fmla   v13.4h, v6.4h, v2.h[5]      \n"
-                        "ext    v7.16b, %25.16b, %25.16b, #8 \n"
-                        "fmla   v10.4h, %25.4h, v1.h[2]     \n"
-                        "fmla   v11.4h, %25.4h, v1.h[6]     \n"
-                        "fmla   v12.4h, %25.4h, v2.h[2]     \n"
-                        "fmla   v13.4h, %25.4h, v2.h[6]     \n"
-
-                        "add    %2, %2, #32                 \n"
-
-                        "fmla   v10.4h, v7.4h, v1.h[3]      \n"
-                        "fmla   v11.4h, v7.4h, v1.h[7]      \n"
-                        "fmla   v12.4h, v7.4h, v2.h[3]      \n"
-                        "fmla   v13.4h, v7.4h, v2.h[7]      \n"
-
-                        "add    %3, %3, #32                 \n"
-
-                        "st1    {v10.4h, v11.4h, v12.4h, v13.4h}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00_01), // %8
-                        "w"(_k00_23), // %9
-                        "w"(_k01_01), // %10
-                        "w"(_k01_23), // %11
-                        "w"(_k02_01), // %12
-                        "w"(_k02_23), // %13
-                        "w"(_k10_01), // %14
-                        "w"(_k10_23), // %15
-                        "w"(_k11_01), // %16
-                        "w"(_k11_23), // %17
-                        "w"(_k12_01), // %18
-                        "w"(_k12_23), // %19
-                        "w"(_k20_01), // %20
-                        "w"(_k20_23), // %21
-                        "w"(_k21_01), // %22
-                        "w"(_k21_23), // %23
-                        "w"(_k22_01), // %24
-                        "w"(_k22_23)  // %25
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13");
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v0.8h, v1.8h}, [%1]        \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v12.4h, v13.4h}, [%0]      \n" // sum0 sum1
-
-                        "ext    v4.16b, %8.16b, %8.16b, #8  \n"
-                        "fmul   v10.4h, %8.4h, v0.h[0]      \n"
-                        "fmul   v11.4h, %8.4h, v0.h[4]      \n"
-                        "fmla   v12.4h, v4.4h, v0.h[1]      \n"
-                        "fmla   v13.4h, v4.4h, v0.h[5]      \n"
-                        "ext    v5.16b, %9.16b, %9.16b, #8  \n"
-                        "fmla   v10.4h, %9.4h, v0.h[2]      \n"
-                        "fmla   v11.4h, %9.4h, v0.h[6]      \n"
-                        "fmla   v12.4h, v5.4h, v0.h[3]      \n"
-                        "fmla   v13.4h, v5.4h, v0.h[7]      \n"
-
-                        "ext    v6.16b, %10.16b, %10.16b, #8 \n"
-                        "fmla   v10.4h, %10.4h, v0.h[4]     \n"
-                        "fmla   v11.4h, %10.4h, v1.h[0]     \n"
-                        "fmla   v12.4h, v6.4h, v0.h[5]      \n"
-                        "fmla   v13.4h, v6.4h, v1.h[1]      \n"
-                        "ext    v7.16b, %11.16b, %11.16b, #8 \n"
-                        "fmla   v10.4h, %11.4h, v0.h[6]     \n"
-                        "fmla   v11.4h, %11.4h, v1.h[2]     \n"
-                        "fmla   v12.4h, v7.4h, v0.h[7]      \n"
-                        "fmla   v13.4h, v7.4h, v1.h[3]      \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v2.8h, v3.8h}, [%2]        \n" // r10 r11 r12 r13
-
-                        "ext    v8.16b, %12.16b, %12.16b, #8 \n"
-                        "fmla   v10.4h, %12.4h, v1.h[0]     \n"
-                        "fmla   v11.4h, %12.4h, v1.h[4]     \n"
-                        "fmla   v12.4h, v8.4h, v1.h[1]      \n"
-                        "fmla   v13.4h, v8.4h, v1.h[5]      \n"
-                        "ext    v9.16b, %13.16b, %13.16b, #8 \n"
-                        "fmla   v10.4h, %13.4h, v1.h[2]     \n"
-                        "fmla   v11.4h, %13.4h, v1.h[6]     \n"
-                        "fmla   v12.4h, v9.4h, v1.h[3]      \n"
-                        "fmla   v13.4h, v9.4h, v1.h[7]      \n"
-
-                        "ext    v4.16b, %14.16b, %14.16b, #8 \n"
-                        "fmla   v10.4h, %14.4h, v2.h[0]     \n"
-                        "fmla   v11.4h, %14.4h, v2.h[4]     \n"
-                        "fmla   v12.4h, v4.4h, v2.h[1]      \n"
-                        "fmla   v13.4h, v4.4h, v2.h[5]      \n"
-                        "ext    v5.16b, %15.16b, %15.16b, #8 \n"
-                        "fmla   v10.4h, %15.4h, v2.h[2]     \n"
-                        "fmla   v11.4h, %15.4h, v2.h[6]     \n"
-                        "fmla   v12.4h, v5.4h, v2.h[3]      \n"
-                        "fmla   v13.4h, v5.4h, v2.h[7]      \n"
-
-                        "ext    v6.16b, %16.16b, %16.16b, #8 \n"
-                        "fmla   v10.4h, %16.4h, v2.h[4]     \n"
-                        "fmla   v11.4h, %16.4h, v3.h[0]     \n"
-                        "fmla   v12.4h, v6.4h, v2.h[5]      \n"
-                        "fmla   v13.4h, v6.4h, v3.h[1]      \n"
-                        "ext    v7.16b, %17.16b, %17.16b, #8 \n"
-                        "fmla   v10.4h, %17.4h, v2.h[6]     \n"
-                        "fmla   v11.4h, %17.4h, v3.h[2]     \n"
-                        "fmla   v12.4h, v7.4h, v2.h[7]      \n"
-                        "fmla   v13.4h, v7.4h, v3.h[3]      \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.8h, v1.8h}, [%3]        \n" // r20 r21 r22 r23
-
-                        "ext    v8.16b, %18.16b, %18.16b, #8 \n"
-                        "fmla   v10.4h, %18.4h, v3.h[0]     \n"
-                        "fmla   v11.4h, %18.4h, v3.h[4]     \n"
-                        "fmla   v12.4h, v8.4h, v3.h[1]      \n"
-                        "fmla   v13.4h, v8.4h, v3.h[5]      \n"
-                        "ext    v9.16b, %19.16b, %19.16b, #8 \n"
-                        "fmla   v10.4h, %19.4h, v3.h[2]     \n"
-                        "fmla   v11.4h, %19.4h, v3.h[6]     \n"
-                        "fmla   v12.4h, v9.4h, v3.h[3]      \n"
-                        "fmla   v13.4h, v9.4h, v3.h[7]      \n"
-
-                        "ext    v4.16b, %20.16b, %20.16b, #8 \n"
-                        "fmla   v10.4h, %20.4h, v0.h[0]     \n"
-                        "fmla   v11.4h, %20.4h, v0.h[4]     \n"
-                        "fmla   v12.4h, v4.4h, v0.h[1]      \n"
-                        "fmla   v13.4h, v4.4h, v0.h[5]      \n"
-                        "ext    v5.16b, %21.16b, %21.16b, #8 \n"
-                        "fmla   v10.4h, %21.4h, v0.h[2]     \n"
-                        "fmla   v11.4h, %21.4h, v0.h[6]     \n"
-                        "fmla   v12.4h, v5.4h, v0.h[3]      \n"
-                        "fmla   v13.4h, v5.4h, v0.h[7]      \n"
-
-                        "ext    v6.16b, %22.16b, %22.16b, #8 \n"
-                        "fmla   v10.4h, %22.4h, v0.h[4]     \n"
-                        "fmla   v11.4h, %22.4h, v1.h[0]     \n"
-                        "fmla   v12.4h, v6.4h, v0.h[5]      \n"
-                        "fmla   v13.4h, v6.4h, v1.h[1]      \n"
-                        "ext    v7.16b, %23.16b, %23.16b, #8 \n"
-                        "fmla   v10.4h, %23.4h, v0.h[6]     \n"
-                        "fmla   v11.4h, %23.4h, v1.h[2]     \n"
-                        "fmla   v12.4h, v7.4h, v0.h[7]      \n"
-                        "fmla   v13.4h, v7.4h, v1.h[3]      \n"
-
-                        "ext    v8.16b, %24.16b, %24.16b, #8 \n"
-                        "fmla   v10.4h, %24.4h, v1.h[0]     \n"
-                        "fmla   v11.4h, %24.4h, v1.h[4]     \n"
-                        "fmla   v12.4h, v8.4h, v1.h[1]      \n"
-                        "fmla   v13.4h, v8.4h, v1.h[5]      \n"
-                        "ext    v9.16b, %25.16b, %25.16b, #8 \n"
-                        "fmla   v10.4h, %25.4h, v1.h[2]     \n"
-                        "fmla   v11.4h, %25.4h, v1.h[6]     \n"
-                        "fmla   v12.4h, v9.4h, v1.h[3]      \n"
-                        "fmla   v13.4h, v9.4h, v1.h[7]      \n"
-
-                        "add    %1, %1, #16                 \n"
-
-                        "fadd   v10.4h, v10.4h, v12.4h      \n"
-
-                        "add    %2, %2, #16                 \n"
-
-                        "fadd   v11.4h, v11.4h, v13.4h      \n"
-
-                        "add    %3, %3, #16                 \n"
-
-                        "st1    {v10.4h, v11.4h}, [%0], #16 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00_01), // %8
-                        "w"(_k00_23), // %9
-                        "w"(_k01_01), // %10
-                        "w"(_k01_23), // %11
-                        "w"(_k02_01), // %12
-                        "w"(_k02_23), // %13
-                        "w"(_k10_01), // %14
-                        "w"(_k10_23), // %15
-                        "w"(_k11_01), // %16
-                        "w"(_k11_23), // %17
-                        "w"(_k12_01), // %18
-                        "w"(_k12_23), // %19
-                        "w"(_k20_01), // %20
-                        "w"(_k20_23), // %21
-                        "w"(_k21_01), // %22
-                        "w"(_k21_23), // %23
-                        "w"(_k22_01), // %24
-                        "w"(_k22_23)  // %25
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13");
-                }
-                for (; j < outw; j++)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #192]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%1] \n" // r00 r01 r02
-
-                        "prfm   pldl1keep, [%0, #64]        \n"
-                        "ld1    {v13.4h}, [%0]              \n" // sum0
-
-                        "ext    v6.16b, %8.16b, %8.16b, #8  \n"
-                        "fmul   v10.4h, %8.4h, v0.h[0]      \n"
-                        "fmul   v11.4h, v6.4h, v0.h[1]      \n"
-                        "ext    v7.16b, %9.16b, %9.16b, #8  \n"
-                        "fmul   v12.4h, %9.4h, v0.h[2]      \n"
-                        "fmla   v13.4h, v7.4h, v0.h[3]      \n"
-
-                        "ext    v8.16b, %10.16b, %10.16b, #8 \n"
-                        "fmla   v10.4h, %10.4h, v1.h[0]     \n"
-                        "fmla   v11.4h, v8.4h, v1.h[1]      \n"
-                        "ext    v9.16b, %11.16b, %11.16b, #8 \n"
-                        "fmla   v12.4h, %11.4h, v1.h[2]     \n"
-                        "fmla   v13.4h, v9.4h, v1.h[3]      \n"
-
-                        "prfm   pldl1keep, [%2, #192]       \n"
-                        "ld1    {v3.4h, v4.4h, v5.4h}, [%2] \n" // r10 r11 r12
-
-                        "ext    v6.16b, %12.16b, %12.16b, #8 \n"
-                        "fmla   v10.4h, %12.4h, v2.h[0]     \n"
-                        "fmla   v11.4h, v6.4h, v2.h[1]      \n"
-                        "ext    v7.16b, %13.16b, %13.16b, #8 \n"
-                        "fmla   v12.4h, %13.4h, v2.h[2]     \n"
-                        "fmla   v13.4h, v7.4h, v2.h[3]      \n"
-
-                        "ext    v8.16b, %14.16b, %14.16b, #8 \n"
-                        "fmla   v10.4h, %14.4h, v3.h[0]     \n"
-                        "fmla   v11.4h, v8.4h, v3.h[1]      \n"
-                        "ext    v9.16b, %15.16b, %15.16b, #8 \n"
-                        "fmla   v12.4h, %15.4h, v3.h[2]     \n"
-                        "fmla   v13.4h, v9.4h, v3.h[3]      \n"
-
-                        "ext    v6.16b, %16.16b, %16.16b, #8 \n"
-                        "fmla   v10.4h, %16.4h, v4.h[0]     \n"
-                        "fmla   v11.4h, v6.4h, v4.h[1]      \n"
-                        "ext    v7.16b, %17.16b, %17.16b, #8 \n"
-                        "fmla   v12.4h, %17.4h, v4.h[2]     \n"
-                        "fmla   v13.4h, v7.4h, v4.h[3]      \n"
-
-                        "prfm   pldl1keep, [%3, #192]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%3] \n" // r20 r21 r22
-
-                        "ext    v8.16b, %18.16b, %18.16b, #8 \n"
-                        "fmla   v10.4h, %18.4h, v5.h[0]     \n"
-                        "fmla   v11.4h, v8.4h, v5.h[1]      \n"
-                        "ext    v9.16b, %19.16b, %19.16b, #8 \n"
-                        "fmla   v12.4h, %19.4h, v5.h[2]     \n"
-                        "fmla   v13.4h, v9.4h, v5.h[3]      \n"
-
-                        "ext    v6.16b, %20.16b, %20.16b, #8 \n"
-                        "fmla   v10.4h, %20.4h, v0.h[0]     \n"
-                        "fmla   v11.4h, v6.4h, v0.h[1]      \n"
-                        "ext    v7.16b, %21.16b, %21.16b, #8 \n"
-                        "fmla   v12.4h, %21.4h, v0.h[2]     \n"
-                        "fmla   v13.4h, v7.4h, v0.h[3]      \n"
-
-                        "ext    v8.16b, %22.16b, %22.16b, #8 \n"
-                        "fmla   v10.4h, %22.4h, v1.h[0]     \n"
-                        "fmla   v11.4h, v8.4h, v1.h[1]      \n"
-                        "ext    v9.16b, %23.16b, %23.16b, #8 \n"
-                        "fmla   v12.4h, %23.4h, v1.h[2]     \n"
-                        "fmla   v13.4h, v9.4h, v1.h[3]      \n"
-
-                        "ext    v6.16b, %24.16b, %24.16b, #8 \n"
-                        "fmla   v10.4h, %24.4h, v2.h[0]     \n"
-                        "fmla   v11.4h, v6.4h, v2.h[1]      \n"
-                        "ext    v7.16b, %25.16b, %25.16b, #8 \n"
-                        "fmla   v12.4h, %25.4h, v2.h[2]     \n"
-                        "fmla   v13.4h, v7.4h, v2.h[3]      \n"
-
-                        "fadd   v10.4h, v10.4h, v11.4h      \n"
-
-                        "add    %1, %1, #8                  \n"
-
-                        "fadd   v12.4h, v12.4h, v13.4h      \n"
-
-                        "add    %2, %2, #8                  \n"
-
-                        "fadd   v10.4h, v10.4h, v12.4h      \n"
-
-                        "add    %3, %3, #8                  \n"
-
-                        "st1    {v10.4h}, [%0], #8          \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00_01), // %8
-                        "w"(_k00_23), // %9
-                        "w"(_k01_01), // %10
-                        "w"(_k01_23), // %11
-                        "w"(_k02_01), // %12
-                        "w"(_k02_23), // %13
-                        "w"(_k10_01), // %14
-                        "w"(_k10_23), // %15
-                        "w"(_k11_01), // %16
-                        "w"(_k11_23), // %17
-                        "w"(_k12_01), // %18
-                        "w"(_k12_23), // %19
-                        "w"(_k20_01), // %20
-                        "w"(_k20_23), // %21
-                        "w"(_k21_01), // %22
-                        "w"(_k21_23), // %23
-                        "w"(_k22_01), // %24
-                        "w"(_k22_23)  // %25
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13");
-                }
-
-                r0 += 8;
-                r1 += 8;
-                r2 += 8;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4to1.h
deleted file mode 100644
index 206a93bd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack4to1.h
+++ /dev/null
@@ -1,1153 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack4to1_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    int remain_outch_start = 0;
-
-#if __ARM_NEON && __aarch64__
-    int nn_outch = 0;
-    nn_outch = outch >> 1;
-    remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        Mat out0 = top_blob.channel(p);
-        Mat out1 = top_blob.channel(p + 1);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        const float bias1 = bias ? bias[p + 1] : 0.f;
-        out0.fill(bias0);
-        out1.fill(bias1);
-
-        const float* k0 = kernel.channel(p);
-        const float* k1 = kernel.channel(p + 1);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            float32x4_t _k00_0 = vld1q_f32(k0);
-            float32x4_t _k01_0 = vld1q_f32(k0 + 4);
-            float32x4_t _k02_0 = vld1q_f32(k0 + 8);
-            float32x4_t _k10_0 = vld1q_f32(k0 + 12);
-            float32x4_t _k11_0 = vld1q_f32(k0 + 16);
-            float32x4_t _k12_0 = vld1q_f32(k0 + 20);
-            float32x4_t _k20_0 = vld1q_f32(k0 + 24);
-            float32x4_t _k21_0 = vld1q_f32(k0 + 28);
-            float32x4_t _k22_0 = vld1q_f32(k0 + 32);
-
-            float32x4_t _k00_1 = vld1q_f32(k1);
-            float32x4_t _k01_1 = vld1q_f32(k1 + 4);
-            float32x4_t _k02_1 = vld1q_f32(k1 + 8);
-            float32x4_t _k10_1 = vld1q_f32(k1 + 12);
-            float32x4_t _k11_1 = vld1q_f32(k1 + 16);
-            float32x4_t _k12_1 = vld1q_f32(k1 + 20);
-            float32x4_t _k20_1 = vld1q_f32(k1 + 24);
-            float32x4_t _k21_1 = vld1q_f32(k1 + 28);
-            float32x4_t _k22_1 = vld1q_f32(k1 + 32);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-
-                for (; j + 3 < outw; j += 4)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n" // r00 r01 r02 r03
-
-                        "fmul   v16.4s, %10.4s, v0.4s       \n"
-                        "fmul   v17.4s, %19.4s, v0.4s       \n"
-                        "fmul   v18.4s, %10.4s, v1.4s       \n"
-                        "fmul   v19.4s, %19.4s, v1.4s       \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v4.4s, v5.4s}, [%2]        \n" // r04 r05
-
-                        "fmul   v6.4s, %10.4s, v2.4s        \n"
-                        "fmul   v7.4s, %19.4s, v2.4s        \n"
-                        "fmul   v8.4s, %10.4s, v3.4s        \n"
-                        "fmul   v9.4s, %19.4s, v3.4s        \n"
-
-                        "fmla   v16.4s, %11.4s, v1.4s       \n"
-                        "fmla   v17.4s, %20.4s, v1.4s       \n"
-                        "fmla   v18.4s, %11.4s, v2.4s       \n"
-                        "fmla   v19.4s, %20.4s, v2.4s       \n"
-                        "fmla   v6.4s, %11.4s, v3.4s        \n"
-                        "fmla   v7.4s, %20.4s, v3.4s        \n"
-                        "fmla   v8.4s, %11.4s, v4.4s        \n"
-                        "fmla   v9.4s, %20.4s, v4.4s        \n"
-
-                        "fmla   v16.4s, %12.4s, v2.4s       \n"
-                        "fmla   v17.4s, %21.4s, v2.4s       \n"
-                        "fmla   v18.4s, %12.4s, v3.4s       \n"
-                        "fmla   v19.4s, %21.4s, v3.4s       \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%3], #64 \n" // r10 r11 r12 r12
-
-                        "fmla   v6.4s, %12.4s, v4.4s        \n"
-                        "fmla   v7.4s, %21.4s, v4.4s        \n"
-                        "fmla   v8.4s, %12.4s, v5.4s        \n"
-                        "fmla   v9.4s, %21.4s, v5.4s        \n"
-
-                        "fmla   v16.4s, %13.4s, v0.4s       \n"
-                        "fmla   v17.4s, %22.4s, v0.4s       \n"
-                        "fmla   v18.4s, %13.4s, v1.4s       \n"
-                        "fmla   v19.4s, %22.4s, v1.4s       \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v4.4s, v5.4s}, [%3]        \n" // r14 r15
-
-                        "fmla   v6.4s, %13.4s, v2.4s        \n"
-                        "fmla   v7.4s, %22.4s, v2.4s        \n"
-                        "fmla   v8.4s, %13.4s, v3.4s        \n"
-                        "fmla   v9.4s, %22.4s, v3.4s        \n"
-
-                        "fmla   v16.4s, %14.4s, v1.4s       \n"
-                        "fmla   v17.4s, %23.4s, v1.4s       \n"
-                        "fmla   v18.4s, %14.4s, v2.4s       \n"
-                        "fmla   v19.4s, %23.4s, v2.4s       \n"
-                        "fmla   v6.4s, %14.4s, v3.4s        \n"
-                        "fmla   v7.4s, %23.4s, v3.4s        \n"
-                        "fmla   v8.4s, %14.4s, v4.4s        \n"
-                        "fmla   v9.4s, %23.4s, v4.4s        \n"
-
-                        "fmla   v16.4s, %15.4s, v2.4s       \n"
-                        "fmla   v17.4s, %24.4s, v2.4s       \n"
-                        "fmla   v18.4s, %15.4s, v3.4s       \n"
-                        "fmla   v19.4s, %24.4s, v3.4s       \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%4], #64 \n" // r20 r21 r22 r22
-
-                        "fmla   v6.4s, %15.4s, v4.4s        \n"
-                        "fmla   v7.4s, %24.4s, v4.4s        \n"
-                        "fmla   v8.4s, %15.4s, v5.4s        \n"
-                        "fmla   v9.4s, %24.4s, v5.4s        \n"
-
-                        "fmla   v16.4s, %16.4s, v0.4s       \n"
-                        "fmla   v17.4s, %25.4s, v0.4s       \n"
-                        "fmla   v18.4s, %16.4s, v1.4s       \n"
-                        "fmla   v19.4s, %25.4s, v1.4s       \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v4.4s, v5.4s}, [%4]        \n" // r24 r25
-
-                        "fmla   v6.4s, %16.4s, v2.4s        \n"
-                        "fmla   v7.4s, %25.4s, v2.4s        \n"
-                        "fmla   v8.4s, %16.4s, v3.4s        \n"
-                        "fmla   v9.4s, %25.4s, v3.4s        \n"
-
-                        "fmla   v16.4s, %17.4s, v1.4s       \n"
-                        "fmla   v17.4s, %26.4s, v1.4s       \n"
-                        "fmla   v18.4s, %17.4s, v2.4s       \n"
-                        "fmla   v19.4s, %26.4s, v2.4s       \n"
-                        "fmla   v6.4s, %17.4s, v3.4s        \n"
-                        "fmla   v7.4s, %26.4s, v3.4s        \n"
-                        "fmla   v8.4s, %17.4s, v4.4s        \n"
-                        "fmla   v9.4s, %26.4s, v4.4s        \n"
-
-                        "fmla   v16.4s, %18.4s, v2.4s       \n"
-                        "fmla   v17.4s, %27.4s, v2.4s       \n"
-                        "fmla   v18.4s, %18.4s, v3.4s       \n"
-                        "fmla   v19.4s, %27.4s, v3.4s       \n"
-                        "fmla   v6.4s, %18.4s, v4.4s        \n"
-                        "fmla   v7.4s, %27.4s, v4.4s        \n"
-                        "fmla   v8.4s, %18.4s, v5.4s        \n"
-                        "fmla   v9.4s, %27.4s, v5.4s        \n"
-
-                        "ld1    {v0.4s}, [%0]               \n" // sum00 sum01 sum02 sum03
-                        "ld1    {v1.4s}, [%1]               \n" // sum10 sum11 sum12 sum13
-
-                        "faddp  v16.4s, v16.4s, v16.4s      \n"
-                        "faddp  v17.4s, v17.4s, v17.4s      \n"
-                        "faddp  v18.4s, v18.4s, v18.4s      \n"
-                        "faddp  v19.4s, v19.4s, v19.4s      \n"
-                        "faddp  v6.4s, v6.4s, v6.4s         \n"
-                        "faddp  v7.4s, v7.4s, v7.4s         \n"
-                        "faddp  v8.4s, v8.4s, v8.4s         \n"
-                        "faddp  v9.4s, v9.4s, v9.4s         \n"
-
-                        "faddp  v16.2s, v16.2s, v18.2s      \n"
-                        "faddp  v17.2s, v17.2s, v19.2s      \n"
-                        "faddp  v6.2s, v6.2s, v8.2s         \n"
-                        "faddp  v7.2s, v7.2s, v9.2s         \n"
-
-                        "trn1   v16.2d, v16.2d, v6.2d       \n"
-                        "trn1   v17.2d, v17.2d, v7.2d       \n"
-
-                        "fadd   v0.4s, v0.4s, v16.4s        \n"
-                        "fadd   v1.4s, v1.4s, v17.4s        \n"
-
-                        "st1    {v0.4s}, [%0], #16          \n"
-                        "st1    {v1.4s}, [%1], #16          \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(outptr1), // %1
-                        "=r"(r0),      // %2
-                        "=r"(r1),      // %3
-                        "=r"(r2)       // %4
-                        : "0"(outptr0),
-                        "1"(outptr1),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00_0), // %10
-                        "w"(_k01_0), // %11
-                        "w"(_k02_0), // %12
-                        "w"(_k10_0), // %13
-                        "w"(_k11_0), // %14
-                        "w"(_k12_0), // %15
-                        "w"(_k20_0), // %16
-                        "w"(_k21_0), // %17
-                        "w"(_k22_0), // %18
-                        "w"(_k00_1), // %19
-                        "w"(_k01_1), // %20
-                        "w"(_k02_1), // %21
-                        "w"(_k10_1), // %22
-                        "w"(_k11_1), // %23
-                        "w"(_k12_1), // %24
-                        "w"(_k20_1), // %25
-                        "w"(_k21_1), // %26
-                        "w"(_k22_1)  // %27
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v16", "v17", "v18", "v19");
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2] \n" // r00 r01 r02 r03
-
-                        "fmul   v16.4s, %10.4s, v0.4s       \n"
-                        "fmul   v17.4s, %19.4s, v0.4s       \n"
-                        "fmul   v18.4s, %10.4s, v1.4s       \n"
-                        "fmul   v19.4s, %19.4s, v1.4s       \n"
-                        "fmla   v16.4s, %11.4s, v1.4s       \n"
-                        "fmla   v17.4s, %20.4s, v1.4s       \n"
-                        "fmla   v18.4s, %11.4s, v2.4s       \n"
-                        "fmla   v19.4s, %20.4s, v2.4s       \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%3] \n" // r10 r11 r12 r12
-
-                        "fmla   v16.4s, %12.4s, v2.4s       \n"
-                        "fmla   v17.4s, %21.4s, v2.4s       \n"
-                        "fmla   v18.4s, %12.4s, v3.4s       \n"
-                        "fmla   v19.4s, %21.4s, v3.4s       \n"
-
-                        "fmla   v16.4s, %13.4s, v4.4s       \n"
-                        "fmla   v17.4s, %22.4s, v4.4s       \n"
-                        "fmla   v18.4s, %13.4s, v5.4s       \n"
-                        "fmla   v19.4s, %22.4s, v5.4s       \n"
-                        "fmla   v16.4s, %14.4s, v5.4s       \n"
-                        "fmla   v17.4s, %23.4s, v5.4s       \n"
-                        "fmla   v18.4s, %14.4s, v6.4s       \n"
-                        "fmla   v19.4s, %23.4s, v6.4s       \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%4] \n" // r20 r21 r22 r22
-
-                        "fmla   v16.4s, %15.4s, v6.4s       \n"
-                        "fmla   v17.4s, %24.4s, v6.4s       \n"
-                        "fmla   v18.4s, %15.4s, v7.4s       \n"
-                        "fmla   v19.4s, %24.4s, v7.4s       \n"
-
-                        "fmla   v16.4s, %16.4s, v0.4s       \n"
-                        "fmla   v17.4s, %25.4s, v0.4s       \n"
-                        "fmla   v18.4s, %16.4s, v1.4s       \n"
-                        "fmla   v19.4s, %25.4s, v1.4s       \n"
-                        "fmla   v16.4s, %17.4s, v1.4s       \n"
-                        "fmla   v17.4s, %26.4s, v1.4s       \n"
-                        "fmla   v18.4s, %17.4s, v2.4s       \n"
-                        "fmla   v19.4s, %26.4s, v2.4s       \n"
-                        "fmla   v16.4s, %18.4s, v2.4s       \n"
-                        "fmla   v17.4s, %27.4s, v2.4s       \n"
-                        "fmla   v18.4s, %18.4s, v3.4s       \n"
-                        "fmla   v19.4s, %27.4s, v3.4s       \n"
-
-                        "ld1    {v4.2s}, [%0]               \n" // sum00 sum01
-                        "ld1    {v5.2s}, [%1]               \n" // sum10 sum11
-
-                        "faddp  v16.4s, v16.4s, v16.4s      \n"
-                        "faddp  v17.4s, v17.4s, v17.4s      \n"
-                        "faddp  v18.4s, v18.4s, v18.4s      \n"
-                        "faddp  v19.4s, v19.4s, v19.4s      \n"
-
-                        "add    %2, %2, #32                 \n"
-
-                        "faddp  v16.2s, v16.2s, v18.2s      \n"
-                        "faddp  v17.2s, v17.2s, v19.2s      \n"
-
-                        "add    %3, %3, #32                 \n"
-
-                        "fadd   v4.2s, v4.2s, v16.2s        \n"
-                        "fadd   v5.2s, v5.2s, v17.2s        \n"
-
-                        "add    %4, %4, #32                 \n"
-
-                        "st1    {v4.2s}, [%0], #8           \n"
-                        "st1    {v5.2s}, [%1], #8           \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(outptr1), // %1
-                        "=r"(r0),      // %2
-                        "=r"(r1),      // %3
-                        "=r"(r2)       // %4
-                        : "0"(outptr0),
-                        "1"(outptr1),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00_0), // %10
-                        "w"(_k01_0), // %11
-                        "w"(_k02_0), // %12
-                        "w"(_k10_0), // %13
-                        "w"(_k11_0), // %14
-                        "w"(_k12_0), // %15
-                        "w"(_k20_0), // %16
-                        "w"(_k21_0), // %17
-                        "w"(_k22_0), // %18
-                        "w"(_k00_1), // %19
-                        "w"(_k01_1), // %20
-                        "w"(_k02_1), // %21
-                        "w"(_k10_1), // %22
-                        "w"(_k11_1), // %23
-                        "w"(_k12_1), // %24
-                        "w"(_k20_1), // %25
-                        "w"(_k21_1), // %26
-                        "w"(_k22_1)  // %27
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19");
-                }
-                for (; j < outw; j++)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%2, #384]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s}, [%2] \n" // r00 r01 r02
-
-                        "fmul   v16.4s, %10.4s, v0.4s       \n"
-                        "fmul   v17.4s, %19.4s, v0.4s       \n"
-                        "fmul   v18.4s, %11.4s, v1.4s       \n"
-                        "fmul   v19.4s, %20.4s, v1.4s       \n"
-
-                        "prfm   pldl1keep, [%3, #384]       \n"
-                        "ld1    {v3.4s, v4.4s, v5.4s}, [%3] \n" // r10 r11 r12
-
-                        "fmla   v16.4s, %12.4s, v2.4s       \n"
-                        "fmla   v17.4s, %21.4s, v2.4s       \n"
-
-                        "fmla   v18.4s, %13.4s, v3.4s       \n"
-                        "fmla   v19.4s, %22.4s, v3.4s       \n"
-                        "fmla   v16.4s, %14.4s, v4.4s       \n"
-                        "fmla   v17.4s, %23.4s, v4.4s       \n"
-
-                        "prfm   pldl1keep, [%4, #384]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s}, [%4] \n" // r20 r21 r22
-
-                        "fmla   v18.4s, %15.4s, v5.4s       \n"
-                        "fmla   v19.4s, %24.4s, v5.4s       \n"
-
-                        "fmla   v16.4s, %16.4s, v0.4s       \n"
-                        "fmla   v17.4s, %25.4s, v0.4s       \n"
-                        "fmla   v18.4s, %17.4s, v1.4s       \n"
-                        "fmla   v19.4s, %26.4s, v1.4s       \n"
-                        "fmla   v16.4s, %18.4s, v2.4s       \n"
-                        "fmla   v17.4s, %27.4s, v2.4s       \n"
-
-                        "ld1    {v3.s}[0], [%0]             \n" // sum00
-                        "ld1    {v4.s}[0], [%1]             \n" // sum10
-
-                        "fadd   v16.4s, v16.4s, v18.4s      \n"
-                        "fadd   v17.4s, v17.4s, v19.4s      \n"
-
-                        "add    %2, %2, #16                 \n"
-
-                        "faddp  v16.4s, v16.4s, v16.4s      \n"
-                        "faddp  v17.4s, v17.4s, v17.4s      \n"
-
-                        "add    %3, %3, #16                 \n"
-
-                        "faddp  v16.2s, v16.2s, v16.2s      \n"
-                        "faddp  v17.2s, v17.2s, v17.2s      \n"
-
-                        "add    %4, %4, #16                 \n"
-
-                        "fadd   v3.2s, v3.2s, v16.2s        \n"
-                        "fadd   v4.2s, v4.2s, v17.2s        \n"
-
-                        "st1    {v3.s}[0], [%0], #4         \n"
-                        "st1    {v4.s}[0], [%1], #4         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(outptr1), // %1
-                        "=r"(r0),      // %2
-                        "=r"(r1),      // %3
-                        "=r"(r2)       // %4
-                        : "0"(outptr0),
-                        "1"(outptr1),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "w"(_k00_0), // %10
-                        "w"(_k01_0), // %11
-                        "w"(_k02_0), // %12
-                        "w"(_k10_0), // %13
-                        "w"(_k11_0), // %14
-                        "w"(_k12_0), // %15
-                        "w"(_k20_0), // %16
-                        "w"(_k21_0), // %17
-                        "w"(_k22_0), // %18
-                        "w"(_k00_1), // %19
-                        "w"(_k01_1), // %20
-                        "w"(_k02_1), // %21
-                        "w"(_k10_1), // %22
-                        "w"(_k11_1), // %23
-                        "w"(_k12_1), // %24
-                        "w"(_k20_1), // %25
-                        "w"(_k21_1), // %26
-                        "w"(_k22_1)  // %27
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v19");
-                }
-
-                r0 += 2 * 4;
-                r1 += 2 * 4;
-                r2 += 2 * 4;
-            }
-
-            k0 += 9 * 4;
-            k1 += 9 * 4;
-        }
-    }
-#endif // __ARM_NEON && __aarch64__
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        out0.fill(bias0);
-
-        const float* k0 = kernel.channel(p);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            float32x4_t _k00 = vld1q_f32(k0);
-            float32x4_t _k01 = vld1q_f32(k0 + 4);
-            float32x4_t _k02 = vld1q_f32(k0 + 8);
-            float32x4_t _k10 = vld1q_f32(k0 + 12);
-            float32x4_t _k11 = vld1q_f32(k0 + 16);
-            float32x4_t _k12 = vld1q_f32(k0 + 20);
-            float32x4_t _k20 = vld1q_f32(k0 + 24);
-            float32x4_t _k21 = vld1q_f32(k0 + 28);
-            float32x4_t _k22 = vld1q_f32(k0 + 32);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-
-#if __aarch64__
-                for (; j + 7 < outw; j += 8)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n" // r04 r05 r06 r07
-
-                        "fmul   v16.4s, %8.4s, v0.4s        \n"
-                        "fmul   v17.4s, %8.4s, v1.4s        \n"
-                        "fmul   v18.4s, %8.4s, v2.4s        \n"
-                        "fmul   v19.4s, %8.4s, v3.4s        \n"
-                        "fmul   v20.4s, %8.4s, v4.4s        \n"
-                        "fmul   v21.4s, %8.4s, v5.4s        \n"
-                        "fmul   v22.4s, %8.4s, v6.4s        \n"
-                        "fmul   v23.4s, %8.4s, v7.4s        \n"
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%1]        \n" // r08 r09
-
-                        "fmla   v16.4s, %9.4s, v1.4s        \n"
-                        "fmla   v17.4s, %9.4s, v2.4s        \n"
-                        "fmla   v18.4s, %9.4s, v3.4s        \n"
-                        "fmla   v19.4s, %9.4s, v4.4s        \n"
-                        "fmla   v20.4s, %9.4s, v5.4s        \n"
-                        "fmla   v21.4s, %9.4s, v6.4s        \n"
-                        "fmla   v22.4s, %9.4s, v7.4s        \n"
-                        "fmla   v23.4s, %9.4s, v8.4s        \n"
-
-                        "fmla   v16.4s, %10.4s, v2.4s       \n"
-                        "fmla   v17.4s, %10.4s, v3.4s       \n"
-                        "fmla   v18.4s, %10.4s, v4.4s       \n"
-                        "fmla   v19.4s, %10.4s, v5.4s       \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n" // r10 r11 r12 r13
-
-                        "fmla   v20.4s, %10.4s, v6.4s       \n"
-                        "fmla   v21.4s, %10.4s, v7.4s       \n"
-                        "fmla   v22.4s, %10.4s, v8.4s       \n"
-                        "fmla   v23.4s, %10.4s, v9.4s       \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%2], #64 \n" // r14 r15 r16 r17
-
-                        "fmla   v16.4s, %11.4s, v0.4s       \n"
-                        "fmla   v17.4s, %11.4s, v1.4s       \n"
-                        "fmla   v18.4s, %11.4s, v2.4s       \n"
-                        "fmla   v19.4s, %11.4s, v3.4s       \n"
-                        "fmla   v20.4s, %11.4s, v4.4s       \n"
-                        "fmla   v21.4s, %11.4s, v5.4s       \n"
-                        "fmla   v22.4s, %11.4s, v6.4s       \n"
-                        "fmla   v23.4s, %11.4s, v7.4s       \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%2]        \n" // r18 r19
-
-                        "fmla   v16.4s, %12.4s, v1.4s       \n"
-                        "fmla   v17.4s, %12.4s, v2.4s       \n"
-                        "fmla   v18.4s, %12.4s, v3.4s       \n"
-                        "fmla   v19.4s, %12.4s, v4.4s       \n"
-                        "fmla   v20.4s, %12.4s, v5.4s       \n"
-                        "fmla   v21.4s, %12.4s, v6.4s       \n"
-                        "fmla   v22.4s, %12.4s, v7.4s       \n"
-                        "fmla   v23.4s, %12.4s, v8.4s       \n"
-
-                        "fmla   v16.4s, %13.4s, v2.4s       \n"
-                        "fmla   v17.4s, %13.4s, v3.4s       \n"
-                        "fmla   v18.4s, %13.4s, v4.4s       \n"
-                        "fmla   v19.4s, %13.4s, v5.4s       \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%3], #64 \n" // r20 r21 r22 r23
-
-                        "fmla   v20.4s, %13.4s, v6.4s       \n"
-                        "fmla   v21.4s, %13.4s, v7.4s       \n"
-                        "fmla   v22.4s, %13.4s, v8.4s       \n"
-                        "fmla   v23.4s, %13.4s, v9.4s       \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%3], #64 \n" // r24 r25 r26 r27
-
-                        "fmla   v16.4s, %14.4s, v0.4s       \n"
-                        "fmla   v17.4s, %14.4s, v1.4s       \n"
-                        "fmla   v18.4s, %14.4s, v2.4s       \n"
-                        "fmla   v19.4s, %14.4s, v3.4s       \n"
-                        "fmla   v20.4s, %14.4s, v4.4s       \n"
-                        "fmla   v21.4s, %14.4s, v5.4s       \n"
-                        "fmla   v22.4s, %14.4s, v6.4s       \n"
-                        "fmla   v23.4s, %14.4s, v7.4s       \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%3]        \n" // r28 r29
-
-                        "fmla   v16.4s, %15.4s, v1.4s       \n"
-                        "fmla   v17.4s, %15.4s, v2.4s       \n"
-                        "fmla   v18.4s, %15.4s, v3.4s       \n"
-                        "fmla   v19.4s, %15.4s, v4.4s       \n"
-                        "fmla   v20.4s, %15.4s, v5.4s       \n"
-                        "fmla   v21.4s, %15.4s, v6.4s       \n"
-                        "fmla   v22.4s, %15.4s, v7.4s       \n"
-                        "fmla   v23.4s, %15.4s, v8.4s       \n"
-
-                        "fmla   v16.4s, %16.4s, v2.4s       \n"
-                        "fmla   v17.4s, %16.4s, v3.4s       \n"
-                        "fmla   v18.4s, %16.4s, v4.4s       \n"
-                        "fmla   v19.4s, %16.4s, v5.4s       \n"
-                        "fmla   v20.4s, %16.4s, v6.4s       \n"
-                        "fmla   v21.4s, %16.4s, v7.4s       \n"
-                        "fmla   v22.4s, %16.4s, v8.4s       \n"
-                        "fmla   v23.4s, %16.4s, v9.4s       \n"
-
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%0]        \n" // sum0 sum1 sum2 sum3 sum4 sum5 sum6 sum7
-
-                        "faddp  v16.4s, v16.4s, v17.4s      \n"
-                        "faddp  v18.4s, v18.4s, v19.4s      \n"
-                        "faddp  v20.4s, v20.4s, v21.4s      \n"
-                        "faddp  v22.4s, v22.4s, v23.4s      \n"
-
-                        "faddp  v16.4s, v16.4s, v18.4s      \n"
-                        "faddp  v20.4s, v20.4s, v22.4s      \n"
-
-                        "fadd   v0.4s, v0.4s, v16.4s        \n"
-                        "fadd   v1.4s, v1.4s, v20.4s        \n"
-
-                        "st1    {v0.4s, v1.4s}, [%0], #32   \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-                }
-#endif // __aarch64__
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%1]        \n" // r04 r05
-
-                        "fmul   v16.4s, %8.4s, v0.4s        \n"
-                        "fmul   v17.4s, %8.4s, v1.4s        \n"
-                        "fmul   v18.4s, %8.4s, v2.4s        \n"
-                        "fmul   v19.4s, %8.4s, v3.4s        \n"
-
-                        "fmla   v16.4s, %9.4s, v1.4s        \n"
-                        "fmla   v17.4s, %9.4s, v2.4s        \n"
-                        "fmla   v18.4s, %9.4s, v3.4s        \n"
-                        "fmla   v19.4s, %9.4s, v8.4s        \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%2], #64 \n" // r10 r11 r12 r13
-
-                        "fmla   v16.4s, %10.4s, v2.4s       \n"
-                        "fmla   v17.4s, %10.4s, v3.4s       \n"
-                        "fmla   v18.4s, %10.4s, v8.4s       \n"
-                        "fmla   v19.4s, %10.4s, v9.4s       \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%2]        \n" // r14 r15
-
-                        "fmla   v16.4s, %11.4s, v4.4s       \n"
-                        "fmla   v17.4s, %11.4s, v5.4s       \n"
-                        "fmla   v18.4s, %11.4s, v6.4s       \n"
-                        "fmla   v19.4s, %11.4s, v7.4s       \n"
-
-                        "fmla   v16.4s, %12.4s, v5.4s       \n"
-                        "fmla   v17.4s, %12.4s, v6.4s       \n"
-                        "fmla   v18.4s, %12.4s, v7.4s       \n"
-                        "fmla   v19.4s, %12.4s, v8.4s       \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%3], #64 \n" // r20 r21 r22 r23
-
-                        "fmla   v16.4s, %13.4s, v6.4s       \n"
-                        "fmla   v17.4s, %13.4s, v7.4s       \n"
-                        "fmla   v18.4s, %13.4s, v8.4s       \n"
-                        "fmla   v19.4s, %13.4s, v9.4s       \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v8.4s, v9.4s}, [%3]        \n" // r24 r25
-
-                        "fmla   v16.4s, %14.4s, v0.4s       \n"
-                        "fmla   v17.4s, %14.4s, v1.4s       \n"
-                        "fmla   v18.4s, %14.4s, v2.4s       \n"
-                        "fmla   v19.4s, %14.4s, v3.4s       \n"
-
-                        "fmla   v16.4s, %15.4s, v1.4s       \n"
-                        "fmla   v17.4s, %15.4s, v2.4s       \n"
-                        "fmla   v18.4s, %15.4s, v3.4s       \n"
-                        "fmla   v19.4s, %15.4s, v8.4s       \n"
-
-                        "fmla   v16.4s, %16.4s, v2.4s       \n"
-                        "fmla   v17.4s, %16.4s, v3.4s       \n"
-                        "fmla   v18.4s, %16.4s, v8.4s       \n"
-                        "fmla   v19.4s, %16.4s, v9.4s       \n"
-
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v0.4s}, [%0]               \n" // sum0 sum1 sum2 sum3
-
-                        "faddp  v16.4s, v16.4s, v17.4s      \n"
-                        "faddp  v18.4s, v18.4s, v19.4s      \n"
-
-                        "faddp  v16.4s, v16.4s, v18.4s      \n"
-
-                        "fadd   v0.4s, v0.4s, v16.4s        \n"
-
-                        "st1    {v0.4s}, [%0], #16          \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v16", "v17", "v18", "v19");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%1 :128]! \n" // r00 r01
-
-                        "vmul.f32   q3, %q8, q0     \n"
-
-                        "pld        [%1, #128]      \n"
-                        "vld1.f32   {d4-d5}, [%1 :128]! \n" // r02
-
-                        "vmul.f32   q4, %q8, q1     \n"
-                        "vmla.f32   q3, %q9, q1     \n"
-
-                        "pld        [%1, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%1 :128]! \n" // r03 r04
-
-                        "vmul.f32   q5, %q8, q2     \n"
-                        "vmla.f32   q4, %q9, q2     \n"
-                        "vmla.f32   q3, %q10, q2    \n"
-
-                        "vmul.f32   q6, %q8, q0     \n"
-                        "vmla.f32   q5, %q9, q0     \n"
-                        "vmla.f32   q4, %q10, q0    \n"
-
-                        "pld        [%1, #128]      \n"
-                        "vld1.f32   {d4-d5}, [%1 :128] \n" // r05
-
-                        "vmla.f32   q6, %q9, q1     \n"
-                        "vmla.f32   q5, %q10, q1    \n"
-
-                        "pld        [%2, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%2 :128]! \n" // r10 r11
-
-                        "vmla.f32   q6, %q10, q2    \n"
-
-                        "vmla.f32   q3, %q11, q0    \n"
-
-                        "pld        [%2, #128]      \n"
-                        "vld1.f32   {d4-d5}, [%2 :128]! \n" // r12
-
-                        "vmla.f32   q4, %q11, q1    \n"
-                        "vmla.f32   q3, %q12, q1    \n"
-
-                        "pld        [%2, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%2 :128]! \n" // r13 r14
-
-                        "vmla.f32   q5, %q11, q2    \n"
-                        "vmla.f32   q4, %q12, q2    \n"
-                        "vmla.f32   q3, %q13, q2    \n"
-
-                        "vmla.f32   q6, %q11, q0    \n"
-                        "vmla.f32   q5, %q12, q0    \n"
-                        "vmla.f32   q4, %q13, q0    \n"
-
-                        "pld        [%2, #128]      \n"
-                        "vld1.f32   {d4-d5}, [%2 :128] \n" // r15
-
-                        "vmla.f32   q6, %q12, q1    \n"
-                        "vmla.f32   q5, %q13, q1    \n"
-
-                        "pld        [%3, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%3 :128]! \n" // r20 r21
-
-                        "vmla.f32   q6, %q13, q2    \n"
-
-                        "vmla.f32   q3, %q14, q0    \n"
-
-                        "pld        [%3, #128]      \n"
-                        "vld1.f32   {d4-d5}, [%3 :128]! \n" // r22
-
-                        "vmla.f32   q4, %q14, q1    \n"
-                        "vmla.f32   q3, %q15, q1    \n"
-
-                        "pld        [%3, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%3 :128]! \n" // r23 r24
-
-                        "vmla.f32   q5, %q14, q2    \n"
-                        "vmla.f32   q4, %q15, q2    \n"
-                        "vmla.f32   q3, %q16, q2    \n"
-
-                        "vmla.f32   q6, %q14, q0    \n"
-                        "vmla.f32   q5, %q15, q0    \n"
-                        "vmla.f32   q4, %q16, q0    \n"
-
-                        "pld        [%3, #128]      \n"
-                        "vld1.f32   {d4-d5}, [%3 :128] \n" // r25
-
-                        "vmla.f32   q6, %q15, q1    \n"
-                        "vmla.f32   q5, %q16, q1    \n"
-
-                        "vld1.f32   {d0-d1}, [%0]   \n" // sum0 sum1 sum2 sum3
-
-                        "vmla.f32   q6, %q16, q2    \n"
-
-                        "vadd.f32   d6, d6, d7      \n"
-                        "vadd.f32   d8, d8, d9      \n"
-                        "vadd.f32   d10, d10, d11   \n"
-                        "vadd.f32   d12, d12, d13   \n"
-
-                        "sub        %1, %1, #16     \n"
-
-                        "vpadd.f32  d6, d6, d8      \n"
-                        "vpadd.f32  d7, d10, d12    \n"
-
-                        "sub        %2, %2, #16     \n"
-
-                        "vadd.f32   q0, q0, q3      \n"
-
-                        "sub        %3, %3, #16     \n"
-
-                        "vst1.f32   {d0-d1}, [%0]!  \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1] \n" // r00 r01 r02 r03
-
-                        "fmul   v16.4s, %8.4s, v0.4s        \n"
-                        "fmul   v17.4s, %8.4s, v1.4s        \n"
-                        "fmul   v18.4s, %9.4s, v1.4s        \n"
-                        "fmul   v19.4s, %9.4s, v2.4s        \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%2] \n" // r10 r11 r12 r13
-
-                        "fmla   v16.4s, %10.4s, v2.4s       \n"
-                        "fmla   v17.4s, %10.4s, v3.4s       \n"
-
-                        "fmla   v18.4s, %11.4s, v4.4s       \n"
-                        "fmla   v19.4s, %11.4s, v5.4s       \n"
-                        "fmla   v16.4s, %12.4s, v5.4s       \n"
-                        "fmla   v17.4s, %12.4s, v6.4s       \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%3] \n" // r20 r21 r22 r23
-
-                        "fmla   v18.4s, %13.4s, v6.4s       \n"
-                        "fmla   v19.4s, %13.4s, v7.4s       \n"
-
-                        "fmla   v16.4s, %14.4s, v0.4s       \n"
-                        "fmla   v17.4s, %14.4s, v1.4s       \n"
-                        "fmla   v18.4s, %15.4s, v1.4s       \n"
-                        "fmla   v19.4s, %15.4s, v2.4s       \n"
-                        "fmla   v16.4s, %16.4s, v2.4s       \n"
-                        "fmla   v17.4s, %16.4s, v3.4s       \n"
-
-                        "ld1    {v0.2s}, [%0]               \n" // sum0 sum1
-
-                        "fadd   v16.4s, v16.4s, v18.4s      \n"
-                        "fadd   v17.4s, v17.4s, v19.4s      \n"
-
-                        "add    %1, %1, #32                 \n"
-
-                        "faddp  v16.4s, v16.4s, v17.4s      \n"
-
-                        "add    %2, %2, #32                 \n"
-
-                        "faddp  v16.4s, v16.4s, v16.4s      \n"
-
-                        "add    %3, %3, #32                 \n"
-
-                        "fadd   v0.2s, v0.2s, v16.2s        \n"
-
-                        "st1    {v0.2s}, [%0], #8           \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%1 :128]! \n" // r00 r01
-
-                        "vmul.f32   q5, %q8, q0     \n"
-                        "vmul.f32   q6, %q8, q1     \n"
-                        "vmul.f32   q2, %q9, q1     \n"
-
-                        "pld        [%1, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%1 :128] \n" // r02 r03
-
-                        "vmul.f32   q3, %q9, q0     \n"
-                        "vmla.f32   q5, %q10, q0    \n"
-                        "vmla.f32   q6, %q10, q1    \n"
-
-                        "pld        [%2, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%2 :128]! \n" // r10 r11
-
-                        "vmla.f32   q2, %q11, q0    \n"
-                        "vmla.f32   q3, %q11, q1    \n"
-                        "vmla.f32   q5, %q12, q1    \n"
-
-                        "pld        [%2, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%2 :128] \n" // r12 r13
-
-                        "vmla.f32   q6, %q12, q0    \n"
-                        "vmla.f32   q2, %q13, q0    \n"
-                        "vmla.f32   q3, %q13, q1    \n"
-
-                        "pld        [%3, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%3 :128]! \n" // r20 r21
-
-                        "vmla.f32   q5, %q14, q0    \n"
-                        "vmla.f32   q6, %q14, q1    \n"
-                        "vmla.f32   q2, %q15, q1    \n"
-
-                        "pld        [%3, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%3 :128] \n" // r22 r23
-
-                        "vmla.f32   q3, %q15, q0    \n"
-                        "vmla.f32   q5, %q16, q0    \n"
-                        "vmla.f32   q6, %q16, q1    \n"
-
-                        "vld1.f32   {d8}, [%0]      \n" // sum0 sum1
-
-                        "vadd.f32   q5, q5, q2      \n"
-                        "vadd.f32   q6, q6, q3      \n"
-
-                        "vadd.f32   d10, d10, d11   \n"
-                        "vadd.f32   d12, d12, d13   \n"
-
-                        "vpadd.f32  d10, d10, d12   \n"
-
-                        "vadd.f32   d8, d8, d10     \n"
-
-                        "vst1.f32   {d8}, [%0]!     \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #384]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s}, [%1] \n" // r00 r01 r02
-
-                        "eor    v16.16b, v16.16b, v16.16b   \n"
-                        "ld1    {v16.s}[0], [%0]            \n" // sum0
-
-                        "fmul   v17.4s, %8.4s, v0.4s        \n"
-                        "fmul   v18.4s, %9.4s, v1.4s        \n"
-
-                        "prfm   pldl1keep, [%2, #384]       \n"
-                        "ld1    {v3.4s, v4.4s, v5.4s}, [%2] \n" // r10 r11 r12
-
-                        "fmla   v16.4s, %10.4s, v2.4s       \n"
-
-                        "fmla   v17.4s, %11.4s, v3.4s       \n"
-                        "fmla   v18.4s, %12.4s, v4.4s       \n"
-
-                        "prfm   pldl1keep, [%3, #384]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s}, [%3] \n" // r20 r21 r22
-
-                        "fmla   v16.4s, %13.4s, v5.4s       \n"
-
-                        "fmla   v17.4s, %14.4s, v0.4s       \n"
-                        "fmla   v18.4s, %15.4s, v1.4s       \n"
-                        "fmla   v16.4s, %16.4s, v2.4s       \n"
-
-                        "fadd   v17.4s, v17.4s, v18.4s      \n"
-                        "fadd   v16.4s, v16.4s, v17.4s      \n"
-
-                        "add    %1, %1, #16                 \n"
-
-                        "faddp  v16.4s, v16.4s, v16.4s      \n"
-
-                        "add    %2, %2, #16                 \n"
-
-                        "faddp  v16.2s, v16.2s, v16.2s      \n"
-
-                        "add    %3, %3, #16                 \n"
-
-                        "st1    {v16.s}[0], [%0], #4        \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #384]      \n"
-                        "vldm       %1, {d0-d5}     \n" // r00 r01 r02
-
-                        "veor       q3, q3          \n"
-                        "vld1.f32   {d6[0]}, [%0]   \n" // sum0
-
-                        "vmul.f32   q4, %q8, q0     \n"
-                        "vmul.f32   q5, %q9, q1     \n"
-                        "vmla.f32   q3, %q10, q2    \n"
-
-                        "pld        [%2, #384]      \n"
-                        "vldm       %2, {d0-d5}     \n" // r10 r11 r12
-
-                        "vmla.f32   q4, %q11, q0    \n"
-                        "vmla.f32   q5, %q12, q1    \n"
-                        "vmla.f32   q3, %q13, q2    \n"
-
-                        "pld        [%3, #384]      \n"
-                        "vldm       %3, {d0-d5}     \n" // r20 r21 r22
-
-                        "vmla.f32   q4, %q14, q0    \n"
-                        "vmla.f32   q5, %q15, q1    \n"
-                        "vmla.f32   q3, %q16, q2    \n"
-
-                        "vadd.f32   q4, q4, q5      \n"
-                        "vadd.f32   q3, q3, q4      \n"
-
-                        "add        %1, %1, #16     \n"
-
-                        "vadd.f32   d6, d6, d7      \n"
-
-                        "add        %2, %2, #16     \n"
-
-                        "vpadd.f32  d6, d6, d6      \n"
-
-                        "add        %3, %3, #16     \n"
-
-                        "vst1.f32   {d6[0]}, [%0]!  \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2)       // %3
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "w"(_k00), // %8
-                        "w"(_k01), // %9
-                        "w"(_k02), // %10
-                        "w"(_k10), // %11
-                        "w"(_k11), // %12
-                        "w"(_k12), // %13
-                        "w"(_k20), // %14
-                        "w"(_k21), // %15
-                        "w"(_k22)  // %16
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5");
-#endif // __aarch64__
-                }
-
-                r0 += 2 * 4;
-                r1 += 2 * 4;
-                r2 += 2 * 4;
-            }
-
-            k0 += 9 * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack8_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack8_fp16s.h
deleted file mode 100644
index fb4480bc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_pack8_fp16s.h
+++ /dev/null
@@ -1,1952 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack8_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float16x8_t _bias0 = bias ? vld1q_f16(bias + p * 8) : vdupq_n_f16(0.f);
-        out0.fill(_bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            __fp16* outptr0 = out0.row<__fp16>(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const __fp16* r0 = img0.row<const __fp16>(0);
-            const __fp16* r1 = img0.row<const __fp16>(1);
-            const __fp16* r2 = img0.row<const __fp16>(2);
-
-            const __fp16* kptr = kernel.channel(p).row<const __fp16>(q);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%1], #64 \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0] \n" // sum0
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v4.8h, v5.8h}, [%1]        \n" // r04 r05
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v3.h[0]     \n"
-
-                        "fmla   v28.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v3.h[2]     \n"
-
-                        "fmla   v28.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v3.h[4]     \n"
-
-                        "fmla   v28.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v3.h[6]     \n"
-
-                        "fmla   v28.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v4.h[0]     \n"
-
-                        "fmla   v28.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v4.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v4.h[2]     \n"
-
-                        "fmla   v28.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v4.h[4]     \n"
-
-                        "fmla   v28.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v4.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v4.h[6]     \n"
-
-                        "fmla   v28.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v5.h[0]     \n"
-
-                        "fmla   v28.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v5.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v5.h[2]     \n"
-
-                        "fmla   v28.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v5.h[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v8.8h, v9.8h, v10.8h, v11.8h}, [%2], #64 \n" // r10 r11 r12 r13
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v5.h[4]     \n"
-
-                        "fmla   v28.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v5.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v5.h[6]     \n"
-
-                        "fmla   v28.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v5.h[7]     \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v12.8h, v13.8h}, [%2]      \n" // r14 r15
-
-                        "fmla   v28.8h, v16.8h, v8.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v11.h[0]    \n"
-
-                        "fmla   v28.8h, v17.8h, v8.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v9.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v11.h[1]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v8.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v11.h[2]    \n"
-
-                        "fmla   v28.8h, v19.8h, v8.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v9.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v10.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v11.h[3]    \n"
-
-                        "fmla   v28.8h, v20.8h, v8.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v11.h[4]    \n"
-
-                        "fmla   v28.8h, v21.8h, v8.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v11.h[5]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v8.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v11.h[6]    \n"
-
-                        "fmla   v28.8h, v23.8h, v8.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v9.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v10.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v11.h[7]    \n"
-
-                        "fmla   v28.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v30.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v12.h[0]    \n"
-
-                        "fmla   v28.8h, v17.8h, v9.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v30.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v12.h[1]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v30.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v12.h[2]    \n"
-
-                        "fmla   v28.8h, v19.8h, v9.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v10.h[3]    \n"
-                        "fmla   v30.8h, v19.8h, v11.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v12.h[3]    \n"
-
-                        "fmla   v28.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v30.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v12.h[4]    \n"
-
-                        "fmla   v28.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v30.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v12.h[5]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v30.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v12.h[6]    \n"
-
-                        "fmla   v28.8h, v23.8h, v9.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v10.h[7]    \n"
-                        "fmla   v30.8h, v23.8h, v11.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v12.h[7]    \n"
-
-                        "fmla   v28.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v29.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v30.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v13.h[0]    \n"
-
-                        "fmla   v28.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v29.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v30.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v13.h[1]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v29.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v30.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v13.h[2]    \n"
-
-                        "fmla   v28.8h, v19.8h, v10.h[3]    \n"
-                        "fmla   v29.8h, v19.8h, v11.h[3]    \n"
-                        "fmla   v30.8h, v19.8h, v12.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v13.h[3]    \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%3], #64 \n" // r20 r21 r22 r23
-
-                        "fmla   v28.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v29.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v30.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v13.h[4]    \n"
-
-                        "fmla   v28.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v29.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v30.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v13.h[5]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v29.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v30.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v13.h[6]    \n"
-
-                        "fmla   v28.8h, v23.8h, v10.h[7]    \n"
-                        "fmla   v29.8h, v23.8h, v11.h[7]    \n"
-                        "fmla   v30.8h, v23.8h, v12.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v13.h[7]    \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v4.8h, v5.8h}, [%3]        \n" // r24 r25
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v3.h[0]     \n"
-
-                        "fmla   v28.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v3.h[2]     \n"
-
-                        "fmla   v28.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v3.h[4]     \n"
-
-                        "fmla   v28.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v3.h[6]     \n"
-
-                        "fmla   v28.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v4.h[0]     \n"
-
-                        "fmla   v28.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v4.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v4.h[2]     \n"
-
-                        "fmla   v28.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v4.h[4]     \n"
-
-                        "fmla   v28.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v4.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v4.h[6]     \n"
-
-                        "fmla   v28.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v5.h[0]     \n"
-
-                        "fmla   v28.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v5.h[1]     \n"
-
-                        // "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4] \n"
-
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v5.h[2]     \n"
-
-                        "fmla   v28.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v5.h[3]     \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v5.h[4]     \n"
-
-                        "fmla   v28.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v5.h[5]     \n"
-
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v5.h[6]     \n"
-
-                        "fmla   v28.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v5.h[7]     \n"
-
-                        "sub    %4, %4, #1088               \n" // kptr -= 8.5 * 64;
-
-                        "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(kptr)     // %4
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%1] \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v30.8h, v31.8h}, [%0]      \n" // sum0
-
-                        "fmul   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmul   v29.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v1.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v2.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v2.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%2] \n" // r10 r11 r12 r13
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v5.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v5.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v5.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v5.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v6.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v6.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v6.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v5.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v6.h[3]     \n"
-
-                        "fmla   v28.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v6.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v6.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v6.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v5.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v6.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v6.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v7.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v6.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v7.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v6.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v7.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v6.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v7.h[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%3] \n" // r20 r21 r22 r23
-
-                        "fmla   v28.8h, v20.8h, v6.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v7.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v6.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v7.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v6.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v7.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v6.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v7.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v1.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v2.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v2.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[1]     \n"
-
-                        // "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4] \n"
-
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "add    %1, %1, #32                 \n"
-
-                        "add    %2, %2, #32                 \n"
-
-                        "add    %3, %3, #32                 \n"
-
-                        "fadd   v28.8h, v28.8h, v30.8h      \n"
-                        "fadd   v29.8h, v29.8h, v31.8h      \n"
-
-                        "sub    %4, %4, #1088               \n" // kptr -= 8.5 * 64;
-
-                        "st1    {v28.8h, v29.8h}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(kptr)     // %4
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-                }
-                for (; j < outw; j++)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "prfm   pldl1keep, [%1, #384]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h}, [%1] \n" // r00 r01 r02
-
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v31.8h}, [%0]              \n" // sum0
-
-                        "fmul   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmul   v29.8h, v17.8h, v0.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmul   v30.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #384]       \n"
-                        "ld1    {v3.8h, v4.8h, v5.8h}, [%2] \n" // r10 r11 r12
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v4.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v4.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v5.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v5.h[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #384]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h}, [%3] \n" // r20 r21 r22
-
-                        "fmla   v28.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v5.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v5.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v0.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-
-                        // "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4] \n"
-
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-
-                        "add    %1, %1, #16                 \n"
-
-                        "fmla   v30.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "add    %2, %2, #16                 \n"
-
-                        "fadd   v28.8h, v28.8h, v29.8h      \n"
-                        "fadd   v30.8h, v30.8h, v31.8h      \n"
-
-                        "add    %3, %3, #16                 \n"
-
-                        "fadd   v28.8h, v28.8h, v30.8h      \n"
-
-                        "sub    %4, %4, #1088               \n" // kptr -= 8.5 * 64;
-
-                        "st1    {v28.8h}, [%0], #16         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(kptr)     // %4
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-                }
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-            }
-        }
-    }
-}
-
-static void conv3x3s2_pack8_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 8;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float16x8_t _bias0 = bias ? vld1q_f16(bias + p * 8) : vdupq_n_f16(0.f);
-        out0.fill(_bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            __fp16* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const __fp16* r0 = img0.row<const __fp16>(0);
-            const __fp16* r1 = img0.row<const __fp16>(1);
-            const __fp16* r2 = img0.row<const __fp16>(2);
-
-            const __fp16* kptr = kernel.channel(p).row<const __fp16>(q);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%1], #64 \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0] \n" // sum0
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%1], #64 \n" // r04 r05 r06 r07
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v6.h[0]     \n"
-
-                        "fmla   v28.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v6.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v6.h[2]     \n"
-
-                        "fmla   v28.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v6.h[3]     \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v6.h[4]     \n"
-
-                        "fmla   v28.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v6.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v6.h[6]     \n"
-
-                        "fmla   v28.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v6.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v7.h[0]     \n"
-
-                        "fmla   v28.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v7.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v7.h[2]     \n"
-
-                        "fmla   v28.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v5.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v7.h[3]     \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v7.h[4]     \n"
-
-                        "fmla   v28.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v7.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v7.h[6]     \n"
-
-                        "fmla   v28.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v5.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v7.h[7]     \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.8h}, [%1]               \n" // r08
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v6.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v0.h[0]     \n"
-
-                        "fmla   v28.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v6.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v0.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v6.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v0.h[2]     \n"
-
-                        "fmla   v28.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v6.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v8.8h, v9.8h, v10.8h, v11.8h}, [%2], #64 \n" // r10 r11 r12 r13
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v6.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v0.h[4]     \n"
-
-                        "fmla   v28.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v6.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v0.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v6.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v0.h[6]     \n"
-
-                        "fmla   v28.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v6.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[7]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%2], #64 \n" // r14 r15 r16 r17
-
-                        "fmla   v28.8h, v16.8h, v8.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v30.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v14.h[0]    \n"
-
-                        "fmla   v28.8h, v17.8h, v8.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v30.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v14.h[1]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v8.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v30.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v14.h[2]    \n"
-
-                        "fmla   v28.8h, v19.8h, v8.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v10.h[3]    \n"
-                        "fmla   v30.8h, v19.8h, v12.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v14.h[3]    \n"
-
-                        "fmla   v28.8h, v20.8h, v8.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v30.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v14.h[4]    \n"
-
-                        "fmla   v28.8h, v21.8h, v8.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v30.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v14.h[5]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v8.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v30.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v14.h[6]    \n"
-
-                        "fmla   v28.8h, v23.8h, v8.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v10.h[7]    \n"
-                        "fmla   v30.8h, v23.8h, v12.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v14.h[7]    \n"
-
-                        "fmla   v28.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v30.8h, v16.8h, v13.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v15.h[0]    \n"
-
-                        "fmla   v28.8h, v17.8h, v9.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v30.8h, v17.8h, v13.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v15.h[1]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v30.8h, v18.8h, v13.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v15.h[2]    \n"
-
-                        "fmla   v28.8h, v19.8h, v9.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v11.h[3]    \n"
-                        "fmla   v30.8h, v19.8h, v13.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v15.h[3]    \n"
-
-                        "fmla   v28.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v30.8h, v20.8h, v13.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v15.h[4]    \n"
-
-                        "fmla   v28.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v30.8h, v21.8h, v13.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v15.h[5]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v30.8h, v22.8h, v13.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v15.h[6]    \n"
-
-                        "fmla   v28.8h, v23.8h, v9.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v11.h[7]    \n"
-                        "fmla   v30.8h, v23.8h, v13.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v15.h[7]    \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v8.8h}, [%2]               \n" // r18
-
-                        "fmla   v28.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v29.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v30.8h, v16.8h, v14.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v8.h[0]     \n"
-
-                        "fmla   v28.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v29.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v30.8h, v17.8h, v14.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v8.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v29.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v30.8h, v18.8h, v14.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v8.h[2]     \n"
-
-                        "fmla   v28.8h, v19.8h, v10.h[3]    \n"
-                        "fmla   v29.8h, v19.8h, v12.h[3]    \n"
-                        "fmla   v30.8h, v19.8h, v14.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v8.h[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%3], #64 \n" // r20 r21 r22 r23
-
-                        "fmla   v28.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v29.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v30.8h, v20.8h, v14.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v8.h[4]     \n"
-
-                        "fmla   v28.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v29.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v30.8h, v21.8h, v14.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v8.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v29.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v30.8h, v22.8h, v14.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v8.h[6]     \n"
-
-                        "fmla   v28.8h, v23.8h, v10.h[7]    \n"
-                        "fmla   v29.8h, v23.8h, v12.h[7]    \n"
-                        "fmla   v30.8h, v23.8h, v14.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v8.h[7]     \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%3], #64 \n" // r24 r25 r26 r27
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v6.h[0]     \n"
-
-                        "fmla   v28.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v6.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v6.h[2]     \n"
-
-                        "fmla   v28.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v6.h[3]     \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v6.h[4]     \n"
-
-                        "fmla   v28.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v6.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v6.h[6]     \n"
-
-                        "fmla   v28.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v6.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v7.h[0]     \n"
-
-                        "fmla   v28.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v7.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v7.h[2]     \n"
-
-                        "fmla   v28.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v5.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v7.h[3]     \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v7.h[4]     \n"
-
-                        "fmla   v28.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v7.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v7.h[6]     \n"
-
-                        "fmla   v28.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v5.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v7.h[7]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.8h}, [%3]               \n" // r28
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v6.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v0.h[0]     \n"
-
-                        "fmla   v28.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v6.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v0.h[1]     \n"
-
-                        // "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4] \n"
-
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v6.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v0.h[2]     \n"
-
-                        "fmla   v28.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v6.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[3]     \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v6.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v0.h[4]     \n"
-
-                        "fmla   v28.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v6.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v0.h[5]     \n"
-
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v6.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v0.h[6]     \n"
-
-                        "fmla   v28.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v6.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[7]     \n"
-
-                        "sub    %4, %4, #1088               \n" // kptr -= 8.5 * 64;
-
-                        "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(kptr)     // %4
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%1], #64 \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v30.8h, v31.8h}, [%0]      \n" // sum0
-
-                        "fmul   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmul   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v2.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v2.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.8h}, [%1]               \n" // r04
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v0.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%2], #64 \n" // r10 r11 r12 r13
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v0.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v6.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v6.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v6.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v6.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v6.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v6.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v6.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v6.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v7.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v7.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v7.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v5.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v7.h[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v4.8h}, [%2]               \n" // r14
-
-                        "fmla   v28.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v7.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v7.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v7.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v5.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v7.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v6.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v6.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v4.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v6.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v6.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%3], #64 \n" // r20 r21 r22 r23
-
-                        "fmla   v28.8h, v20.8h, v6.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v6.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v4.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v6.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v6.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v2.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v2.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.8h}, [%3]               \n" // r24
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v0.h[1]     \n"
-
-                        // "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4] \n"
-
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[7]     \n"
-
-                        "fadd   v28.8h, v28.8h, v30.8h      \n"
-                        "fadd   v29.8h, v29.8h, v31.8h      \n"
-
-                        "sub    %4, %4, #1088               \n" // kptr -= 8.5 * 64;
-
-                        "st1    {v28.8h, v29.8h}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(kptr)     // %4
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-                }
-                for (; j < outw; j++)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "prfm   pldl1keep, [%1, #384]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h}, [%1] \n" // r00 r01 r02
-
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v31.8h}, [%0]              \n" // sum0
-
-                        "fmul   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmul   v29.8h, v17.8h, v0.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmul   v30.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #384]       \n"
-                        "ld1    {v3.8h, v4.8h, v5.8h}, [%2] \n" // r10 r11 r12
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v4.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v4.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v5.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v5.h[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #384]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h}, [%3] \n" // r20 r21 r22
-
-                        "fmla   v28.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v5.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v5.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v0.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-
-                        // "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4] \n"
-
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-
-                        "add    %1, %1, #32                 \n"
-
-                        "fmla   v30.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "add    %2, %2, #32                 \n"
-
-                        "fadd   v28.8h, v28.8h, v29.8h      \n"
-                        "fadd   v30.8h, v30.8h, v31.8h      \n"
-
-                        "add    %3, %3, #32                 \n"
-
-                        "fadd   v28.8h, v28.8h, v30.8h      \n"
-
-                        "sub    %4, %4, #1088               \n" // kptr -= 8.5 * 64;
-
-                        "st1    {v28.8h}, [%0], #16         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(kptr)     // %4
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd.h
deleted file mode 100644
index 291f6907..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd.h
+++ /dev/null
@@ -1,9405 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd_pack_A_tile(const Mat& A, Mat& AT, int batch, int max_ii, int max_kk)
-{
-    const int N = max_kk * batch;
-
-    for (int b = 0; b < batch; b++)
-    {
-        float* pp = AT.row(b);
-
-        int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; ii + 7 < max_ii; ii += 8)
-        {
-            const float* p0 = (const float*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                pp[2] = p0[2 * N];
-                pp[3] = p0[3 * N];
-                pp[4] = p0[4 * N];
-                pp[5] = p0[5 * N];
-                pp[6] = p0[6 * N];
-                pp[7] = p0[7 * N];
-                p0 += batch;
-                pp += 8;
-            }
-        }
-#endif // __aarch64__
-        for (; ii + 3 < max_ii; ii += 4)
-        {
-            const float* p0 = (const float*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                pp[2] = p0[2 * N];
-                pp[3] = p0[3 * N];
-                p0 += batch;
-                pp += 4;
-            }
-        }
-#endif // __ARM_NEON
-        for (; ii + 1 < max_ii; ii += 2)
-        {
-            const float* p0 = (const float*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                p0 += batch;
-                pp += 2;
-            }
-        }
-        for (; ii < max_ii; ii++)
-        {
-            const float* p0 = (const float*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                p0 += batch;
-                pp += 1;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd_transpose_pack_B_tile(const Mat& B, Mat& BT, int batch, int max_jj, int max_kk, int nT)
-{
-    #pragma omp parallel for num_threads(nT)
-    for (int b = 0; b < batch; b++)
-    {
-        float* pp = BT.row(b);
-
-        int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            const float* p0 = B;
-
-            int kk = 0;
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                // transpose 8x12
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v4.4s, v5.4s, v6.4s, v7.4s}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0]  \n"
-
-                    "uzp1   v24.4s, v0.4s, v4.4s        \n"
-                    "uzp2   v25.4s, v0.4s, v4.4s        \n"
-                    "uzp1   v26.4s, v1.4s, v5.4s        \n"
-                    "uzp2   v27.4s, v1.4s, v5.4s        \n"
-                    "uzp1   v28.4s, v2.4s, v6.4s        \n"
-                    "uzp2   v29.4s, v2.4s, v6.4s        \n"
-                    "uzp1   v30.4s, v3.4s, v7.4s        \n"
-                    "uzp2   v31.4s, v3.4s, v7.4s        \n"
-
-                    "uzp1   v0.4s, v8.4s, v12.4s        \n"
-                    "uzp2   v1.4s, v8.4s, v12.4s        \n"
-                    "uzp1   v2.4s, v9.4s, v13.4s        \n"
-                    "uzp2   v3.4s, v9.4s, v13.4s        \n"
-                    "uzp1   v4.4s, v10.4s, v14.4s       \n"
-                    "uzp2   v5.4s, v10.4s, v14.4s       \n"
-                    "uzp1   v6.4s, v11.4s, v15.4s       \n"
-                    "uzp2   v7.4s, v11.4s, v15.4s       \n"
-
-                    "sub    %0, %0, #320                \n"
-
-                    "uzp1   v8.4s, v16.4s, v20.4s       \n"
-                    "uzp2   v9.4s, v16.4s, v20.4s       \n"
-                    "uzp1   v10.4s, v17.4s, v21.4s      \n"
-                    "uzp2   v11.4s, v17.4s, v21.4s      \n"
-                    "uzp1   v12.4s, v18.4s, v22.4s      \n"
-                    "uzp2   v13.4s, v18.4s, v22.4s      \n"
-                    "uzp1   v14.4s, v19.4s, v23.4s      \n"
-                    "uzp2   v15.4s, v19.4s, v23.4s      \n"
-
-                    "st1    {v24.4s}, [%1], #16         \n"
-                    "st1    {v0.4s}, [%1], #16          \n"
-                    "st1    {v8.4s}, [%1], #16          \n"
-                    "st1    {v26.4s}, [%1], #16         \n"
-                    "st1    {v2.4s}, [%1], #16          \n"
-                    "st1    {v10.4s}, [%1], #16         \n"
-                    "st1    {v28.4s}, [%1], #16         \n"
-                    "st1    {v4.4s}, [%1], #16          \n"
-                    "st1    {v12.4s}, [%1], #16         \n"
-                    "st1    {v30.4s}, [%1], #16         \n"
-                    "st1    {v6.4s}, [%1], #16          \n"
-                    "st1    {v14.4s}, [%1], #16         \n"
-
-                    "st1    {v25.4s}, [%1], #16         \n"
-                    "st1    {v1.4s}, [%1], #16          \n"
-                    "st1    {v9.4s}, [%1], #16          \n"
-                    "st1    {v27.4s}, [%1], #16         \n"
-                    "st1    {v3.4s}, [%1], #16          \n"
-                    "st1    {v11.4s}, [%1], #16         \n"
-                    "st1    {v29.4s}, [%1], #16         \n"
-                    "st1    {v5.4s}, [%1], #16          \n"
-                    "st1    {v13.4s}, [%1], #16         \n"
-                    "st1    {v31.4s}, [%1], #16         \n"
-                    "st1    {v7.4s}, [%1], #16          \n"
-                    "st1    {v15.4s}, [%1], #16         \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                p0 += max_jj * batch * 8;
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4x4_t _r0 = vld4q_f32(p0);
-                float32x4x4_t _r1 = vld4q_f32(p0 + 16);
-                float32x4x4_t _r2 = vld4q_f32(p0 + 32);
-                float32x4x4_t _r3 = vld4q_f32(p0 + 48);
-                float32x4x4_t _r4 = vld4q_f32(p0 + 64);
-                float32x4x4_t _r5 = vld4q_f32(p0 + 80);
-                float32x4x2_t _r04l = vuzpq_f32(_r0.val[0], _r1.val[0]);
-                float32x4x2_t _r15l = vuzpq_f32(_r0.val[1], _r1.val[1]);
-                float32x4x2_t _r26l = vuzpq_f32(_r0.val[2], _r1.val[2]);
-                float32x4x2_t _r37l = vuzpq_f32(_r0.val[3], _r1.val[3]);
-                float32x4x2_t _r04m = vuzpq_f32(_r2.val[0], _r3.val[0]);
-                float32x4x2_t _r15m = vuzpq_f32(_r2.val[1], _r3.val[1]);
-                float32x4x2_t _r26m = vuzpq_f32(_r2.val[2], _r3.val[2]);
-                float32x4x2_t _r37m = vuzpq_f32(_r2.val[3], _r3.val[3]);
-                float32x4x2_t _r04h = vuzpq_f32(_r4.val[0], _r5.val[0]);
-                float32x4x2_t _r15h = vuzpq_f32(_r4.val[1], _r5.val[1]);
-                float32x4x2_t _r26h = vuzpq_f32(_r4.val[2], _r5.val[2]);
-                float32x4x2_t _r37h = vuzpq_f32(_r4.val[3], _r5.val[3]);
-                vst1q_f32(pp, _r04l.val[0]);
-                vst1q_f32(pp + 4, _r04m.val[0]);
-                vst1q_f32(pp + 4 * 2, _r04h.val[0]);
-                vst1q_f32(pp + 4 * 3, _r15l.val[0]);
-                vst1q_f32(pp + 4 * 4, _r15m.val[0]);
-                vst1q_f32(pp + 4 * 5, _r15h.val[0]);
-                vst1q_f32(pp + 4 * 6, _r26l.val[0]);
-                vst1q_f32(pp + 4 * 7, _r26m.val[0]);
-                vst1q_f32(pp + 4 * 8, _r26h.val[0]);
-                vst1q_f32(pp + 4 * 9, _r37l.val[0]);
-                vst1q_f32(pp + 4 * 10, _r37m.val[0]);
-                vst1q_f32(pp + 4 * 11, _r37h.val[0]);
-                vst1q_f32(pp + 4 * 12, _r04l.val[1]);
-                vst1q_f32(pp + 4 * 13, _r04m.val[1]);
-                vst1q_f32(pp + 4 * 14, _r04h.val[1]);
-                vst1q_f32(pp + 4 * 15, _r15l.val[1]);
-                vst1q_f32(pp + 4 * 16, _r15m.val[1]);
-                vst1q_f32(pp + 4 * 17, _r15h.val[1]);
-                vst1q_f32(pp + 4 * 18, _r26l.val[1]);
-                vst1q_f32(pp + 4 * 19, _r26m.val[1]);
-                vst1q_f32(pp + 4 * 20, _r26h.val[1]);
-                vst1q_f32(pp + 4 * 21, _r37l.val[1]);
-                vst1q_f32(pp + 4 * 22, _r37m.val[1]);
-                vst1q_f32(pp + 4 * 23, _r37h.val[1]);
-                p0 += max_jj * batch * 8;
-                pp += 96;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 8;
-            p0 += (b * max_jj + jj) * 4;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                // transpose 4x12
-                float32x4x4_t _r0 = vld4q_f32(p0);
-                float32x4x4_t _r1 = vld4q_f32(p0 + 16);
-                float32x4x4_t _r2 = vld4q_f32(p0 + 32);
-                vst1q_f32(pp, _r0.val[0]);
-                vst1q_f32(pp + 4, _r1.val[0]);
-                vst1q_f32(pp + 4 * 2, _r2.val[0]);
-                vst1q_f32(pp + 4 * 3, _r0.val[1]);
-                vst1q_f32(pp + 4 * 4, _r1.val[1]);
-                vst1q_f32(pp + 4 * 5, _r2.val[1]);
-                vst1q_f32(pp + 4 * 6, _r0.val[2]);
-                vst1q_f32(pp + 4 * 7, _r1.val[2]);
-                vst1q_f32(pp + 4 * 8, _r2.val[2]);
-                vst1q_f32(pp + 4 * 9, _r0.val[3]);
-                vst1q_f32(pp + 4 * 10, _r1.val[3]);
-                vst1q_f32(pp + 4 * 11, _r2.val[3]);
-                p0 += max_jj * batch * 4;
-                pp += 48;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                // transpose 2x12
-                float32x4x2_t _r0 = vld2q_f32(p0);
-                float32x4x2_t _r1 = vld2q_f32(p0 + 8);
-                float32x4x2_t _r2 = vld2q_f32(p0 + 16);
-                vst1q_f32(pp, _r0.val[0]);
-                vst1q_f32(pp + 4, _r1.val[0]);
-                vst1q_f32(pp + 4 * 2, _r2.val[0]);
-                vst1q_f32(pp + 4 * 3, _r0.val[1]);
-                vst1q_f32(pp + 4 * 4, _r1.val[1]);
-                vst1q_f32(pp + 4 * 5, _r2.val[1]);
-                p0 += max_jj * batch * 2;
-                pp += 24;
-            }
-            p0 -= (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                float32x4_t _r0 = vld1q_f32(p0);
-                float32x4_t _r1 = vld1q_f32(p0 + 4);
-                float32x4_t _r2 = vld1q_f32(p0 + 8);
-                vst1q_f32(pp, _r0);
-                vst1q_f32(pp + 4, _r1);
-                vst1q_f32(pp + 8, _r2);
-                p0 += max_jj * batch;
-                pp += 12;
-            }
-        }
-#endif // __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            const float* p0 = B;
-
-            int kk = 0;
-#if __aarch64__
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                // transpose 8x8
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v4.4s, v5.4s, v6.4s, v7.4s}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0] \n"
-
-                    "uzp1   v16.4s, v0.4s, v4.4s        \n"
-                    "uzp2   v24.4s, v0.4s, v4.4s        \n"
-                    "uzp1   v18.4s, v1.4s, v5.4s        \n"
-                    "uzp2   v26.4s, v1.4s, v5.4s        \n"
-                    "uzp1   v20.4s, v2.4s, v6.4s        \n"
-                    "uzp2   v28.4s, v2.4s, v6.4s        \n"
-                    "uzp1   v22.4s, v3.4s, v7.4s        \n"
-                    "uzp2   v30.4s, v3.4s, v7.4s        \n"
-
-                    "sub    %0, %0, #192                \n"
-
-                    "uzp1   v17.4s, v8.4s, v12.4s       \n"
-                    "uzp2   v25.4s, v8.4s, v12.4s       \n"
-                    "uzp1   v19.4s, v9.4s, v13.4s       \n"
-                    "uzp2   v27.4s, v9.4s, v13.4s       \n"
-                    "uzp1   v21.4s, v10.4s, v14.4s      \n"
-                    "uzp2   v29.4s, v10.4s, v14.4s      \n"
-                    "uzp1   v23.4s, v11.4s, v15.4s      \n"
-                    "uzp2   v31.4s, v11.4s, v15.4s      \n"
-
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%1], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%1], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%1], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                p0 += max_jj * batch * 8;
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4x4_t _r0 = vld4q_f32(p0);
-                float32x4x4_t _r1 = vld4q_f32(p0 + 16);
-                float32x4x4_t _r2 = vld4q_f32(p0 + 32);
-                float32x4x4_t _r3 = vld4q_f32(p0 + 48);
-                float32x4x2_t _r04l = vuzpq_f32(_r0.val[0], _r1.val[0]);
-                float32x4x2_t _r15l = vuzpq_f32(_r0.val[1], _r1.val[1]);
-                float32x4x2_t _r26l = vuzpq_f32(_r0.val[2], _r1.val[2]);
-                float32x4x2_t _r37l = vuzpq_f32(_r0.val[3], _r1.val[3]);
-                float32x4x2_t _r04h = vuzpq_f32(_r2.val[0], _r3.val[0]);
-                float32x4x2_t _r15h = vuzpq_f32(_r2.val[1], _r3.val[1]);
-                float32x4x2_t _r26h = vuzpq_f32(_r2.val[2], _r3.val[2]);
-                float32x4x2_t _r37h = vuzpq_f32(_r2.val[3], _r3.val[3]);
-                vst1q_f32(pp, _r04l.val[0]);
-                vst1q_f32(pp + 4, _r04h.val[0]);
-                vst1q_f32(pp + 4 * 2, _r15l.val[0]);
-                vst1q_f32(pp + 4 * 3, _r15h.val[0]);
-                vst1q_f32(pp + 4 * 4, _r26l.val[0]);
-                vst1q_f32(pp + 4 * 5, _r26h.val[0]);
-                vst1q_f32(pp + 4 * 6, _r37l.val[0]);
-                vst1q_f32(pp + 4 * 7, _r37h.val[0]);
-                vst1q_f32(pp + 4 * 8, _r04l.val[1]);
-                vst1q_f32(pp + 4 * 9, _r04h.val[1]);
-                vst1q_f32(pp + 4 * 10, _r15l.val[1]);
-                vst1q_f32(pp + 4 * 11, _r15h.val[1]);
-                vst1q_f32(pp + 4 * 12, _r26l.val[1]);
-                vst1q_f32(pp + 4 * 13, _r26h.val[1]);
-                vst1q_f32(pp + 4 * 14, _r37l.val[1]);
-                vst1q_f32(pp + 4 * 15, _r37h.val[1]);
-                p0 += max_jj * batch * 8;
-                pp += 64;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 8;
-#endif // __aarch64__
-            p0 += (b * max_jj + jj) * 4;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                // transpose 4x8
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v4.4s, v5.4s, v6.4s, v7.4s}, [%0] \n"
-                    "sub    %0, %0, #64                 \n"
-                    "st1    {v0.4s}, [%1], #16          \n"
-                    "st1    {v4.4s}, [%1], #16          \n"
-                    "st1    {v1.4s}, [%1], #16          \n"
-                    "st1    {v5.4s}, [%1], #16          \n"
-                    "st1    {v2.4s}, [%1], #16          \n"
-                    "st1    {v6.4s}, [%1], #16          \n"
-                    "st1    {v3.4s}, [%1], #16          \n"
-                    "st1    {v7.4s}, [%1], #16          \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%0, #512]          \n"
-                    "vldm       %0!, {d0-d7}        \n"
-                    "pld        [%0, #512]          \n"
-                    "vldm       %0, {d16-d23}       \n"
-
-                    "vtrn.32    q0, q1              \n"
-                    "vtrn.32    q2, q3              \n"
-                    "vtrn.32    q8, q9              \n"
-                    "vtrn.32    q10, q11            \n"
-                    "vswp       d1, d4              \n"
-                    "vswp       d3, d6              \n"
-                    "vswp       d17, d20            \n"
-                    "vswp       d19, d22            \n"
-                    "vswp       q1, q8              \n"
-                    "vswp       q3, q10             \n"
-
-                    "vst1.f32   {d0-d3}, [%1 :128]! \n"
-                    "vst1.f32   {d16-d19}, [%1 :128]! \n"
-                    "sub        %0, %0, #64         \n"
-                    "vst1.f32   {d4-d7}, [%1 :128]! \n"
-                    "vst1.f32   {d20-d23}, [%1 :128]! \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
-#endif // __aarch64__
-                p0 += max_jj * batch * 4;
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4x4_t _r0 = vld4q_f32(p0);
-                float32x4x4_t _r1 = vld4q_f32(p0 + 16);
-                vst1q_f32(pp, _r0.val[0]);
-                vst1q_f32(pp + 4, _r1.val[0]);
-                vst1q_f32(pp + 4 * 2, _r0.val[1]);
-                vst1q_f32(pp + 4 * 3, _r1.val[1]);
-                vst1q_f32(pp + 4 * 4, _r0.val[2]);
-                vst1q_f32(pp + 4 * 5, _r1.val[2]);
-                vst1q_f32(pp + 4 * 6, _r0.val[3]);
-                vst1q_f32(pp + 4 * 7, _r1.val[3]);
-                p0 += max_jj * batch * 4;
-                pp += 32;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 4;
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                // transpose 2x8
-                float32x4x2_t _r0 = vld2q_f32(p0);
-                float32x4x2_t _r1 = vld2q_f32(p0 + 8);
-                vst1q_f32(pp, _r0.val[0]);
-                vst1q_f32(pp + 4, _r1.val[0]);
-                vst1q_f32(pp + 4 * 2, _r0.val[1]);
-                vst1q_f32(pp + 4 * 3, _r1.val[1]);
-                p0 += max_jj * batch * 2;
-                pp += 16;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                float32x4_t _r0 = vld1q_f32(p0);
-                float32x4_t _r1 = vld1q_f32(p0 + 4);
-                vst1q_f32(pp, _r0);
-                vst1q_f32(pp + 4, _r1);
-                p0 += max_jj * batch;
-                pp += 8;
-            }
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            const float* p0 = B;
-
-            int kk = 0;
-#if __aarch64__
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                // transpose 8x4
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v4.4s, v5.4s, v6.4s, v7.4s}, [%0] \n"
-
-                    "uzp1   v8.4s, v0.4s, v4.4s         \n"
-                    "uzp2   v12.4s, v0.4s, v4.4s        \n"
-                    "uzp1   v9.4s, v1.4s, v5.4s         \n"
-                    "uzp2   v13.4s, v1.4s, v5.4s        \n"
-
-                    "sub    %0, %0, #64                 \n"
-
-                    "uzp1   v10.4s, v2.4s, v6.4s        \n"
-                    "uzp2   v14.4s, v2.4s, v6.4s        \n"
-                    "uzp1   v11.4s, v3.4s, v7.4s        \n"
-                    "uzp2   v15.4s, v3.4s, v7.4s        \n"
-
-                    "st1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%1], #64 \n"
-                    "st1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                p0 += max_jj * batch * 8;
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4x4_t _r0;
-                float32x4x4_t _r1;
-                _r0.val[0] = vld1q_f32(p0);
-                _r1.val[0] = vld1q_f32(p0 + 4);
-                _r0.val[1] = vld1q_f32(p0 + 8);
-                _r1.val[1] = vld1q_f32(p0 + 12);
-                _r0.val[2] = vld1q_f32(p0 + 16);
-                _r1.val[2] = vld1q_f32(p0 + 20);
-                _r0.val[3] = vld1q_f32(p0 + 24);
-                _r1.val[3] = vld1q_f32(p0 + 28);
-                vst4q_f32(pp, _r0);
-                vst4q_f32(pp + 16, _r1);
-                p0 += max_jj * batch * 8;
-                pp += 32;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 8;
-#endif // __aarch64__
-            p0 += (b * max_jj + jj) * 4;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                // transpose 4x4
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0] \n"
-                    "st4    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%0, #512]          \n"
-                    "vldm       %0, {d0-d7}         \n"
-                    "vtrn.32    q0, q1              \n"
-                    "vtrn.32    q2, q3              \n"
-                    "vswp       d1, d4              \n"
-                    "vswp       d3, d6              \n"
-                    "vstm       %1!, {d0-d7}        \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "q0", "q1", "q2", "q3");
-#endif // __aarch64__
-                p0 += max_jj * batch * 4;
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4x4_t _r0;
-                _r0.val[0] = vld1q_f32(p0);
-                _r0.val[1] = vld1q_f32(p0 + 4);
-                _r0.val[2] = vld1q_f32(p0 + 8);
-                _r0.val[3] = vld1q_f32(p0 + 12);
-                vst4q_f32(pp, _r0);
-                p0 += max_jj * batch * 4;
-                pp += 16;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 4;
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                // transpose 2x4
-                float32x4x2_t _r0 = vld2q_f32(p0);
-                vst1q_f32(pp, _r0.val[0]);
-                vst1q_f32(pp + 4, _r0.val[1]);
-                p0 += max_jj * batch * 2;
-                pp += 8;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                float32x4_t _r0 = vld1q_f32(p0);
-                vst1q_f32(pp, _r0);
-                p0 += max_jj * batch;
-                pp += 4;
-            }
-        }
-#endif // __ARM_NEON
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            const float* p0 = B;
-
-            int kk = 0;
-#if __ARM_NEON
-#if __aarch64__
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                // transpose 8x2
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0] \n"
-
-                    "zip1   v4.4s, v0.4s, v2.4s         \n"
-                    "zip2   v5.4s, v0.4s, v2.4s         \n"
-                    "zip1   v6.4s, v1.4s, v3.4s         \n"
-                    "zip2   v7.4s, v1.4s, v3.4s         \n"
-
-                    "st1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
-                p0 += max_jj * batch * 8;
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4x2_t _r0;
-                float32x4x2_t _r1;
-                _r0.val[0] = vld1q_f32(p0);
-                _r1.val[0] = vld1q_f32(p0 + 4);
-                _r0.val[1] = vld1q_f32(p0 + 8);
-                _r1.val[1] = vld1q_f32(p0 + 12);
-                vst2q_f32(pp, _r0);
-                vst2q_f32(pp + 8, _r1);
-                p0 += max_jj * batch * 8;
-                pp += 16;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 8;
-#endif // __aarch64__
-            p0 += (b * max_jj + jj) * 4;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                // transpose 4x2
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #256]       \n"
-                    "ld1    {v0.4s, v1.4s}, [%0]        \n"
-                    "st2    {v0.4s, v1.4s}, [%1], #32   \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%0, #256]          \n"
-                    "vld1.f32   {d0-d3}, [%0 :128]  \n"
-                    "vst2.f32   {d0-d3}, [%1 :128]! \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "q0", "q1");
-#endif // __aarch64__
-                p0 += max_jj * batch * 4;
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4x2_t _r0;
-                _r0.val[0] = vld1q_f32(p0);
-                _r0.val[1] = vld1q_f32(p0 + 4);
-                vst2q_f32(pp, _r0);
-                p0 += max_jj * batch * 4;
-                pp += 8;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 4;
-#endif // __ARM_NEON
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[2];
-                pp[2] = p0[1];
-                pp[3] = p0[3];
-                p0 += max_jj * batch * 2;
-                pp += 4;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                p0 += max_jj * batch;
-                pp += 2;
-            }
-        }
-        for (; jj < max_jj; jj++)
-        {
-            const float* p0 = B;
-
-            int kk = 0;
-#if __ARM_NEON
-#if __aarch64__
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #256]       \n"
-                    "ld1    {v0.4s, v1.4s}, [%0]        \n"
-                    "st1    {v0.4s, v1.4s}, [%1], #32   \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1");
-                p0 += max_jj * batch * 8;
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4_t _r0 = vld1q_f32(p0);
-                float32x4_t _r1 = vld1q_f32(p0 + 4);
-                vst1q_f32(pp, _r0);
-                vst1q_f32(pp + 4, _r1);
-                p0 += max_jj * batch * 8;
-                pp += 8;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 8;
-#endif // __aarch64__
-            p0 += (b * max_jj + jj) * 4;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #128]       \n"
-                    "ld1    {v0.4s}, [%0]               \n"
-                    "st1    {v0.4s}, [%1], #16          \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%0, #128]          \n"
-                    "vld1.f32   {d0-d1}, [%0]       \n"
-                    "vst1.f32   {d0-d1}, [%1]!      \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "q0");
-#endif // __aarch64__
-                p0 += max_jj * batch * 4;
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4_t _r0 = vld1q_f32(p0);
-                vst1q_f32(pp, _r0);
-                p0 += max_jj * batch * 4;
-                pp += 4;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 4;
-#endif // __ARM_NEON
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                p0 += max_jj * batch * 2;
-                pp += 2;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                p0 += max_jj * batch;
-                pp += 1;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd_gemm_transB_packed_tile(const Mat& AT_tile, const Mat& BT_tile, Mat& top_blob, int batch, int max_ii, int max_jj, int k, int max_kk, int use_a53_a55_optimized_kernel)
-{
-    // NCNN_LOGE("conv3x3s1_winograd_gemm_transB_packed_tile %d %d %d", max_ii, max_jj, max_kk);
-    float* outptr = top_blob;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const float* pAT = AT_tile.row(b) + max_kk * ii;
-            const float* pB = BT_tile.row(b);
-
-            int jj = 0;
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-                if (use_a53_a55_optimized_kernel && cpu_support_arm_asimdhp())
-                {
-                    // a55
-                    asm volatile(
-                        "cbz    %w7, 0f                     \n"
-
-                        "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                        "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                        "subs   %0, %0, #320                \n"
-                        "b      1f                          \n"
-
-                        "0:                                 \n"
-                        "eor    v8.16b, v8.16b, v8.16b      \n"
-                        "eor    v9.16b, v9.16b, v9.16b      \n"
-                        "eor    v10.16b, v10.16b, v10.16b   \n"
-                        "eor    v11.16b, v11.16b, v11.16b   \n"
-                        "eor    v12.16b, v12.16b, v12.16b   \n"
-                        "eor    v13.16b, v13.16b, v13.16b   \n"
-                        "eor    v14.16b, v14.16b, v14.16b   \n"
-                        "eor    v15.16b, v15.16b, v15.16b   \n"
-                        "eor    v16.16b, v16.16b, v16.16b   \n"
-                        "eor    v17.16b, v17.16b, v17.16b   \n"
-                        "eor    v18.16b, v18.16b, v18.16b   \n"
-                        "eor    v19.16b, v19.16b, v19.16b   \n"
-                        "eor    v20.16b, v20.16b, v20.16b   \n"
-                        "eor    v21.16b, v21.16b, v21.16b   \n"
-                        "eor    v22.16b, v22.16b, v22.16b   \n"
-                        "eor    v23.16b, v23.16b, v23.16b   \n"
-                        "eor    v24.16b, v24.16b, v24.16b   \n"
-                        "eor    v25.16b, v25.16b, v25.16b   \n"
-                        "eor    v26.16b, v26.16b, v26.16b   \n"
-                        "eor    v27.16b, v27.16b, v27.16b   \n"
-                        "eor    v28.16b, v28.16b, v28.16b   \n"
-                        "eor    v29.16b, v29.16b, v29.16b   \n"
-                        "eor    v30.16b, v30.16b, v30.16b   \n"
-                        "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                        "1:                                 \n"
-                        "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                        "cmp    w4, #0                      \n"
-                        "beq    3f                          \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v4.4s}, [%1], #16          \n"
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.4s}, [%2], #16          \n"
-
-                        "ldr    d5, [%1], #8                \n"
-                        "ldr    x25, [%1], #8               \n"
-
-                        ".align 4                           \n"
-                        "2:                                 \n"
-                        "ldr    d1, [%2], #8                \n"
-                        "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                        "ldr    x21, [%2], #8               \n"
-                        "fmla   v10.4s, v4.4s, v0.s[1]      \n"
-                        "ins    v5.d[1], x25                \n"
-                        "fmla   v12.4s, v4.4s, v0.s[2]      \n"
-                        "ldr    d2, [%2], #8                \n"
-                        "fmla   v14.4s, v4.4s, v0.s[3]      \n"
-                        "ldr    x22, [%2], #8               \n"
-                        "fmla   v9.4s, v5.4s, v0.s[0]       \n"
-                        "ldr    d6, [%1], #8                \n"
-                        "fmla   v11.4s, v5.4s, v0.s[1]      \n"
-                        "ins    v1.d[1], x21                \n"
-                        "fmla   v13.4s, v5.4s, v0.s[2]      \n"
-                        "ldr    x26, [%1], #8               \n"
-                        "fmla   v15.4s, v5.4s, v0.s[3]      \n"
-                        "ldr    d3, [%2], #8                \n"
-                        "fmla   v16.4s, v4.4s, v1.s[0]      \n"
-                        "ldr    x23, [%2], #8               \n"
-                        "fmla   v18.4s, v4.4s, v1.s[1]      \n"
-                        "ldr    d7, [%1], #8                \n"
-                        "fmla   v20.4s, v4.4s, v1.s[2]      \n"
-                        "ldr    x27, [%1], #8               \n"
-                        "fmla   v22.4s, v4.4s, v1.s[3]      \n"
-                        "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                        "fmla   v17.4s, v5.4s, v1.s[0]      \n"
-                        "ldr    d0, [%2], #8                \n"
-                        "fmla   v19.4s, v5.4s, v1.s[1]      \n"
-                        "ins    v2.d[1], x22                \n"
-                        "fmla   v21.4s, v5.4s, v1.s[2]      \n"
-                        "ldr    x20, [%2], #8               \n"
-                        "fmla   v23.4s, v5.4s, v1.s[3]      \n"
-                        "fmla   v24.4s, v4.4s, v2.s[0]      \n"
-                        "ldr    d1, [%2], #8                \n"
-                        "fmla   v26.4s, v4.4s, v2.s[1]      \n"
-                        "ins    v6.d[1], x26                \n"
-                        "fmla   v28.4s, v4.4s, v2.s[2]      \n"
-                        "ldr    x21, [%2], #8               \n"
-                        "fmla   v30.4s, v4.4s, v2.s[3]      \n"
-                        "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                        "fmla   v25.4s, v5.4s, v2.s[0]      \n"
-                        "ldr    d4, [%1], #8                \n"
-                        "fmla   v27.4s, v5.4s, v2.s[1]      \n"
-                        "ins    v3.d[1], x23                \n"
-                        "fmla   v29.4s, v5.4s, v2.s[2]      \n"
-                        "ldr    x24, [%1], #8               \n"
-                        "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-                        "fmla   v8.4s, v6.4s, v3.s[0]       \n"
-                        "ldr    d2, [%2], #8                \n"
-                        "fmla   v10.4s, v6.4s, v3.s[1]      \n"
-                        "ins    v7.d[1], x27                \n"
-                        "fmla   v12.4s, v6.4s, v3.s[2]      \n"
-                        "ldr    x22, [%2], #8               \n"
-                        "fmla   v14.4s, v6.4s, v3.s[3]      \n"
-                        "fmla   v9.4s, v7.4s, v3.s[0]       \n"
-                        "ldr    d5, [%1], #8                \n"
-                        "fmla   v11.4s, v7.4s, v3.s[1]      \n"
-                        "ins    v0.d[1], x20                \n"
-                        "fmla   v13.4s, v7.4s, v3.s[2]      \n"
-                        "ldr    x25, [%1], #8               \n"
-                        "fmla   v15.4s, v7.4s, v3.s[3]      \n"
-                        "fmla   v16.4s, v6.4s, v0.s[0]      \n"
-                        "ldr    d3, [%2], #8                \n"
-                        "fmla   v18.4s, v6.4s, v0.s[1]      \n"
-                        "ldr    x23, [%2], #8               \n"
-                        "fmla   v20.4s, v6.4s, v0.s[2]      \n"
-                        "fmla   v22.4s, v6.4s, v0.s[3]      \n"
-                        "fmla   v17.4s, v7.4s, v0.s[0]      \n"
-                        "fmla   v19.4s, v7.4s, v0.s[1]      \n"
-                        "ins    v1.d[1], x21                \n"
-                        "fmla   v21.4s, v7.4s, v0.s[2]      \n"
-                        "fmla   v23.4s, v7.4s, v0.s[3]      \n"
-                        "prfm   pldl1keep, [%2, #256]       \n" // NOTE PRELOAD
-                        "fmla   v24.4s, v6.4s, v1.s[0]      \n"
-                        "fmla   v26.4s, v6.4s, v1.s[1]      \n"
-                        "ins    v4.d[1], x24                \n"
-                        "fmla   v28.4s, v6.4s, v1.s[2]      \n"
-                        "ldr    d0, [%2], #8                \n"
-                        "fmla   v30.4s, v6.4s, v1.s[3]      \n"
-                        "ldr    x20, [%2], #8               \n"
-                        "fmla   v25.4s, v7.4s, v1.s[0]      \n"
-                        "ldr    d6, [%1], #8                \n"
-                        "fmla   v27.4s, v7.4s, v1.s[1]      \n"
-                        "ins    v2.d[1], x22                \n"
-                        "fmla   v29.4s, v7.4s, v1.s[2]      \n"
-                        "fmla   v31.4s, v7.4s, v1.s[3]      \n"
-                        "ldr    x26, [%1], #8               \n"
-                        "fmla   v8.4s, v4.4s, v2.s[0]       \n"
-                        "ldr    d1, [%2], #8                \n"
-                        "fmla   v10.4s, v4.4s, v2.s[1]      \n"
-                        "ins    v5.d[1], x25                \n"
-                        "fmla   v12.4s, v4.4s, v2.s[2]      \n"
-                        "ldr    x21, [%2], #8               \n"
-                        "fmla   v14.4s, v4.4s, v2.s[3]      \n"
-                        "ldr    d7, [%1], #8                \n"
-                        "fmla   v9.4s, v5.4s, v2.s[0]       \n"
-                        "ldr    x27, [%1], #8               \n"
-                        "fmla   v11.4s, v5.4s, v2.s[1]      \n"
-                        "ins    v3.d[1], x23                \n"
-                        "fmla   v13.4s, v5.4s, v2.s[2]      \n"
-                        "fmla   v15.4s, v5.4s, v2.s[3]      \n"
-                        "fmla   v16.4s, v4.4s, v3.s[0]      \n"
-                        "ldr    d2, [%2], #8                \n"
-                        "fmla   v18.4s, v4.4s, v3.s[1]      \n"
-                        "ldr    x22, [%2], #8               \n"
-                        "fmla   v20.4s, v4.4s, v3.s[2]      \n"
-                        "fmla   v22.4s, v4.4s, v3.s[3]      \n"
-                        "fmla   v17.4s, v5.4s, v3.s[0]      \n"
-                        "fmla   v19.4s, v5.4s, v3.s[1]      \n"
-                        "ins    v0.d[1], x20                \n"
-                        "fmla   v21.4s, v5.4s, v3.s[2]      \n"
-                        "fmla   v23.4s, v5.4s, v3.s[3]      \n"
-                        "fmla   v24.4s, v4.4s, v0.s[0]      \n"
-                        "ldr    d3, [%2], #8                \n"
-                        "fmla   v26.4s, v4.4s, v0.s[1]      \n"
-                        "ldr    x23, [%2], #8               \n"
-                        "fmla   v28.4s, v4.4s, v0.s[2]      \n"
-                        "ins    v6.d[1], x26                \n"
-                        "fmla   v30.4s, v4.4s, v0.s[3]      \n"
-                        "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                        "fmla   v25.4s, v5.4s, v0.s[0]      \n"
-                        "ldr    d4, [%1], #8                \n"
-                        "fmla   v27.4s, v5.4s, v0.s[1]      \n"
-                        "ins    v1.d[1], x21                \n"
-                        "fmla   v29.4s, v5.4s, v0.s[2]      \n"
-                        "ldr    x24, [%1], #8               \n"
-                        "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-                        "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                        "fmla   v8.4s, v6.4s, v1.s[0]       \n"
-                        "ldr    d0, [%2], #8                \n"
-                        "fmla   v10.4s, v6.4s, v1.s[1]      \n"
-                        "ins    v7.d[1], x27                \n"
-                        "fmla   v12.4s, v6.4s, v1.s[2]      \n"
-                        "ldr    x20, [%2], #8               \n"
-                        "fmla   v14.4s, v6.4s, v1.s[3]      \n"
-                        "ldr    d5, [%1], #8                \n"
-                        "fmla   v9.4s, v7.4s, v1.s[0]       \n"
-                        "ldr    x25, [%1], #8               \n"
-                        "fmla   v11.4s, v7.4s, v1.s[1]      \n"
-                        "ins    v2.d[1], x22                \n"
-                        "fmla   v13.4s, v7.4s, v1.s[2]      \n"
-                        "fmla   v15.4s, v7.4s, v1.s[3]      \n"
-                        "fmla   v16.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v18.4s, v6.4s, v2.s[1]      \n"
-                        "fmla   v20.4s, v6.4s, v2.s[2]      \n"
-                        "fmla   v22.4s, v6.4s, v2.s[3]      \n"
-                        "fmla   v17.4s, v7.4s, v2.s[0]      \n"
-                        "fmla   v19.4s, v7.4s, v2.s[1]      \n"
-                        "ins    v3.d[1], x23                \n"
-                        "fmla   v21.4s, v7.4s, v2.s[2]      \n"
-                        "fmla   v23.4s, v7.4s, v2.s[3]      \n"
-                        "fmla   v24.4s, v6.4s, v3.s[0]      \n"
-                        "fmla   v26.4s, v6.4s, v3.s[1]      \n"
-                        "fmla   v28.4s, v6.4s, v3.s[2]      \n"
-                        "ins    v4.d[1], x24                \n"
-                        "fmla   v30.4s, v6.4s, v3.s[3]      \n"
-                        "fmla   v25.4s, v7.4s, v3.s[0]      \n"
-                        "subs   w4, w4, #1                  \n"
-                        "fmla   v27.4s, v7.4s, v3.s[1]      \n"
-                        "fmla   v29.4s, v7.4s, v3.s[2]      \n"
-                        "ins    v0.d[1], x20                \n"
-                        "fmla   v31.4s, v7.4s, v3.s[3]      \n"
-                        "bne    2b                          \n"
-
-                        "sub    %1, %1, #32                 \n"
-                        "sub    %2, %2, #16                 \n"
-
-                        "3:                                 \n"
-                        "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                        "cmp    w4, #0                      \n"
-                        "beq    5f                          \n"
-
-                        "4:                                 \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s}, [%2], #48 \n"
-                        "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-
-                        "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                        "fmla   v10.4s, v4.4s, v0.s[1]      \n"
-                        "fmla   v12.4s, v4.4s, v0.s[2]      \n"
-                        "fmla   v14.4s, v4.4s, v0.s[3]      \n"
-                        "fmla   v16.4s, v4.4s, v1.s[0]      \n"
-                        "fmla   v18.4s, v4.4s, v1.s[1]      \n"
-                        "fmla   v20.4s, v4.4s, v1.s[2]      \n"
-                        "fmla   v22.4s, v4.4s, v1.s[3]      \n"
-                        "fmla   v24.4s, v4.4s, v2.s[0]      \n"
-                        "fmla   v26.4s, v4.4s, v2.s[1]      \n"
-                        "fmla   v28.4s, v4.4s, v2.s[2]      \n"
-                        "fmla   v30.4s, v4.4s, v2.s[3]      \n"
-
-                        "subs   w4, w4, #1                  \n"
-
-                        "fmla   v9.4s, v5.4s, v0.s[0]       \n"
-                        "fmla   v11.4s, v5.4s, v0.s[1]      \n"
-                        "fmla   v13.4s, v5.4s, v0.s[2]      \n"
-                        "fmla   v15.4s, v5.4s, v0.s[3]      \n"
-                        "fmla   v17.4s, v5.4s, v1.s[0]      \n"
-                        "fmla   v19.4s, v5.4s, v1.s[1]      \n"
-                        "fmla   v21.4s, v5.4s, v1.s[2]      \n"
-                        "fmla   v23.4s, v5.4s, v1.s[3]      \n"
-                        "fmla   v25.4s, v5.4s, v2.s[0]      \n"
-                        "fmla   v27.4s, v5.4s, v2.s[1]      \n"
-                        "fmla   v29.4s, v5.4s, v2.s[2]      \n"
-                        "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-
-                        "bne    4b                          \n"
-
-                        "5:                                 \n"
-                        "st1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                        "st1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                        "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                        "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                        "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr), // %0
-                        "=r"(pA),     // %1
-                        "=r"(pB)      // %2
-                        : "0"(outptr),
-                        "1"(pA),
-                        "2"(pB),
-                        "r"(max_kk), // %6
-                        "r"(k)       // %7
-                        : "cc", "memory", "x4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-                else if (use_a53_a55_optimized_kernel && !cpu_support_arm_asimdhp())
-                {
-                    // a53
-                    asm volatile(
-                        "cbz    %w7, 0f                     \n"
-
-                        "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                        "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                        "subs   %0, %0, #320                \n"
-                        "b      1f                          \n"
-
-                        "0:                                 \n"
-                        "eor    v8.16b, v8.16b, v8.16b      \n"
-                        "eor    v9.16b, v9.16b, v9.16b      \n"
-                        "eor    v10.16b, v10.16b, v10.16b   \n"
-                        "eor    v11.16b, v11.16b, v11.16b   \n"
-                        "eor    v12.16b, v12.16b, v12.16b   \n"
-                        "eor    v13.16b, v13.16b, v13.16b   \n"
-                        "eor    v14.16b, v14.16b, v14.16b   \n"
-                        "eor    v15.16b, v15.16b, v15.16b   \n"
-                        "eor    v16.16b, v16.16b, v16.16b   \n"
-                        "eor    v17.16b, v17.16b, v17.16b   \n"
-                        "eor    v18.16b, v18.16b, v18.16b   \n"
-                        "eor    v19.16b, v19.16b, v19.16b   \n"
-                        "eor    v20.16b, v20.16b, v20.16b   \n"
-                        "eor    v21.16b, v21.16b, v21.16b   \n"
-                        "eor    v22.16b, v22.16b, v22.16b   \n"
-                        "eor    v23.16b, v23.16b, v23.16b   \n"
-                        "eor    v24.16b, v24.16b, v24.16b   \n"
-                        "eor    v25.16b, v25.16b, v25.16b   \n"
-                        "eor    v26.16b, v26.16b, v26.16b   \n"
-                        "eor    v27.16b, v27.16b, v27.16b   \n"
-                        "eor    v28.16b, v28.16b, v28.16b   \n"
-                        "eor    v29.16b, v29.16b, v29.16b   \n"
-                        "eor    v30.16b, v30.16b, v30.16b   \n"
-                        "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                        "1:                                 \n"
-                        "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                        "cmp    w4, #0                      \n"
-                        "beq    3f                          \n"
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v4.4s}, [%1], #16          \n"
-
-                        "prfm   pldl1keep, [%2, #384]       \n"
-                        "ld1    {v0.4s}, [%2], #16          \n"
-
-                        "ldr    d1, [%2]                    \n"
-                        "ldr    x21, [%2, #8]               \n"
-                        "ldr    d2, [%2, #16]               \n"
-                        "ldr    x22, [%2, #24]              \n"
-                        "add    %2, %2, #32                 \n"
-
-                        ".align 4                           \n"
-                        "2:                                 \n"
-
-                        "ldr    d5, [%1]                    \n"
-                        "ins    v1.d[1], x21                \n"
-                        "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                        "ldr    x25, [%1, #8]               \n"
-                        "fmla   v10.4s, v4.4s, v0.s[1]      \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v12.4s, v4.4s, v0.s[2]      \n"
-
-                        "ldr    d6, [%1]                    \n"
-                        "ins    v2.d[1], x22                \n"
-                        "fmla   v14.4s, v4.4s, v0.s[3]      \n"
-                        "ldr    x26, [%1, #8]               \n"
-                        "fmla   v16.4s, v4.4s, v1.s[0]      \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v18.4s, v4.4s, v1.s[1]      \n"
-
-                        "nop                                \n"
-                        "prfm   pldl1keep, [%1, #256]       \n" // NOTE PRELOAD
-                        "fmla   v20.4s, v4.4s, v1.s[2]      \n"
-                        "nop                                \n"
-                        "fmla   v22.4s, v4.4s, v1.s[3]      \n"
-                        "nop                                \n"
-                        "fmla   v24.4s, v4.4s, v2.s[0]      \n"
-
-                        "ldr    d3, [%2]                    \n"
-                        "ins    v5.d[1], x25                \n"
-                        "fmla   v26.4s, v4.4s, v2.s[1]      \n"
-                        "ldr    x23, [%2, #8]               \n"
-                        "fmla   v28.4s, v4.4s, v2.s[2]      \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v30.4s, v4.4s, v2.s[3]      \n"
-
-                        "nop                                \n"
-                        "prfm   pldl1keep, [%2, #384]       \n" // NOTE PRELOAD
-                        "fmla   v9.4s, v5.4s, v0.s[0]       \n"
-                        "nop                                \n"
-                        "fmla   v11.4s, v5.4s, v0.s[1]      \n"
-                        "nop                                \n"
-                        "fmla   v13.4s, v5.4s, v0.s[2]      \n"
-
-                        "nop                                \n"
-                        "nop                                \n"
-                        "fmla   v15.4s, v5.4s, v0.s[3]      \n"
-                        "nop                                \n"
-                        "fmla   v17.4s, v5.4s, v1.s[0]      \n"
-                        "nop                                \n"
-                        "fmla   v19.4s, v5.4s, v1.s[1]      \n"
-
-                        "ldr    d0, [%2]                    \n"
-                        "ins    v6.d[1], x26                \n"
-                        "fmla   v21.4s, v5.4s, v1.s[2]      \n"
-                        "ldr    x20, [%2, #8]               \n"
-                        "fmla   v23.4s, v5.4s, v1.s[3]      \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v25.4s, v5.4s, v2.s[0]      \n"
-
-                        "ldr    d1, [%2]                    \n"
-                        "ins    v3.d[1], x23                \n"
-                        "fmla   v27.4s, v5.4s, v2.s[1]      \n"
-                        "ldr    x21, [%2, #8]               \n"
-                        "fmla   v29.4s, v5.4s, v2.s[2]      \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-
-                        "ldr    d7, [%1]                    \n"
-                        "ins    v0.d[1], x20                \n"
-                        "fmla   v8.4s, v6.4s, v3.s[0]       \n"
-                        "ldr    x27, [%1, #8]               \n"
-                        "fmla   v10.4s, v6.4s, v3.s[1]      \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v12.4s, v6.4s, v3.s[2]      \n"
-
-                        "ldr    d4, [%1]                    \n"
-                        "ins    v1.d[1], x21                \n"
-                        "fmla   v14.4s, v6.4s, v3.s[3]      \n"
-                        "ldr    x24, [%1, #8]               \n"
-                        "fmla   v16.4s, v6.4s, v0.s[0]      \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v18.4s, v6.4s, v0.s[1]      \n"
-
-                        "nop                                \n"
-                        "prfm   pldl1keep, [%1, #256]       \n" // NOTE PRELOAD
-                        "fmla   v20.4s, v6.4s, v0.s[2]      \n"
-                        "nop                                \n"
-                        "fmla   v22.4s, v6.4s, v0.s[3]      \n"
-                        "nop                                \n"
-                        "fmla   v24.4s, v6.4s, v1.s[0]      \n"
-
-                        "ldr    d2, [%2]                    \n"
-                        "ins    v7.d[1], x27                \n"
-                        "fmla   v26.4s, v6.4s, v1.s[1]      \n"
-                        "ldr    x22, [%2, #8]               \n"
-                        "fmla   v28.4s, v6.4s, v1.s[2]      \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v30.4s, v6.4s, v1.s[3]      \n"
-
-                        "nop                                \n"
-                        "prfm   pldl1keep, [%2, #384]       \n" // NOTE PRELOAD
-                        "fmla   v9.4s, v7.4s, v3.s[0]       \n"
-                        "nop                                \n"
-                        "fmla   v11.4s, v7.4s, v3.s[1]      \n"
-                        "nop                                \n"
-                        "fmla   v13.4s, v7.4s, v3.s[2]      \n"
-
-                        "nop                                \n"
-                        "nop                                \n"
-                        "fmla   v15.4s, v7.4s, v3.s[3]      \n"
-                        "nop                                \n"
-                        "fmla   v17.4s, v7.4s, v0.s[0]      \n"
-                        "nop                                \n"
-                        "fmla   v19.4s, v7.4s, v0.s[1]      \n"
-
-                        "ldr    d3, [%2]                    \n"
-                        "ins    v4.d[1], x24                \n"
-                        "fmla   v21.4s, v7.4s, v0.s[2]      \n"
-                        "ldr    x23, [%2, #8]               \n"
-                        "fmla   v23.4s, v7.4s, v0.s[3]      \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v25.4s, v7.4s, v1.s[0]      \n"
-
-                        "ldr    d0, [%2]                    \n"
-                        "ins    v2.d[1], x22                \n"
-                        "fmla   v27.4s, v7.4s, v1.s[1]      \n"
-                        "ldr    x20, [%2, #8]               \n"
-                        "fmla   v29.4s, v7.4s, v1.s[2]      \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v31.4s, v7.4s, v1.s[3]      \n"
-
-                        "ldr    d5, [%1]                    \n"
-                        "ins    v3.d[1], x23                \n"
-                        "fmla   v8.4s, v4.4s, v2.s[0]       \n"
-                        "ldr    x25, [%1, #8]               \n"
-                        "fmla   v10.4s, v4.4s, v2.s[1]      \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v12.4s, v4.4s, v2.s[2]      \n"
-
-                        "ldr    d6, [%1]                    \n"
-                        "ins    v0.d[1], x20                \n"
-                        "fmla   v14.4s, v4.4s, v2.s[3]      \n"
-                        "ldr    x26, [%1, #8]               \n"
-                        "fmla   v16.4s, v4.4s, v3.s[0]      \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v18.4s, v4.4s, v3.s[1]      \n"
-
-                        "nop                                \n"
-                        "prfm   pldl1keep, [%1, #256]       \n" // NOTE PRELOAD
-                        "fmla   v20.4s, v4.4s, v3.s[2]      \n"
-                        "nop                                \n"
-                        "fmla   v22.4s, v4.4s, v3.s[3]      \n"
-                        "nop                                \n"
-                        "fmla   v24.4s, v4.4s, v0.s[0]      \n"
-
-                        "ldr    d1, [%2]                    \n"
-                        "ins    v5.d[1], x25                \n"
-                        "fmla   v26.4s, v4.4s, v0.s[1]      \n"
-                        "ldr    x21, [%2, #8]               \n"
-                        "fmla   v28.4s, v4.4s, v0.s[2]      \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v30.4s, v4.4s, v0.s[3]      \n"
-
-                        "nop                                \n"
-                        "prfm   pldl1keep, [%2, #384]       \n" // NOTE PRELOAD
-                        "fmla   v9.4s, v5.4s, v2.s[0]       \n"
-                        "nop                                \n"
-                        "fmla   v11.4s, v5.4s, v2.s[1]      \n"
-                        "nop                                \n"
-                        "fmla   v13.4s, v5.4s, v2.s[2]      \n"
-
-                        "nop                                \n"
-                        "nop                                \n"
-                        "fmla   v15.4s, v5.4s, v2.s[3]      \n"
-                        "nop                                \n"
-                        "fmla   v17.4s, v5.4s, v3.s[0]      \n"
-                        "nop                                \n"
-                        "fmla   v19.4s, v5.4s, v3.s[1]      \n"
-
-                        "ldr    d2, [%2]                    \n"
-                        "ins    v6.d[1], x26                \n"
-                        "fmla   v21.4s, v5.4s, v3.s[2]      \n"
-                        "ldr    x22, [%2, #8]               \n"
-                        "fmla   v23.4s, v5.4s, v3.s[3]      \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v25.4s, v5.4s, v0.s[0]      \n"
-
-                        "ldr    d3, [%2]                    \n"
-                        "ins    v1.d[1], x21                \n"
-                        "fmla   v27.4s, v5.4s, v0.s[1]      \n"
-                        "ldr    x23, [%2, #8]               \n"
-                        "fmla   v29.4s, v5.4s, v0.s[2]      \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-
-                        "ldr    d7, [%1]                    \n"
-                        "ins    v2.d[1], x22                \n"
-                        "fmla   v8.4s, v6.4s, v1.s[0]       \n"
-                        "ldr    x27, [%1, #8]               \n"
-                        "fmla   v10.4s, v6.4s, v1.s[1]      \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v12.4s, v6.4s, v1.s[2]      \n"
-
-                        "ldr    d4, [%1]                    \n"
-                        "ins    v3.d[1], x23                \n"
-                        "fmla   v14.4s, v6.4s, v1.s[3]      \n"
-                        "ldr    x24, [%1, #8]               \n"
-                        "fmla   v16.4s, v6.4s, v2.s[0]      \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v18.4s, v6.4s, v2.s[1]      \n"
-
-                        "nop                                \n"
-                        "prfm   pldl1keep, [%1, #256]       \n" // NOTE PRELOAD
-                        "fmla   v20.4s, v6.4s, v2.s[2]      \n"
-                        "nop                                \n"
-                        "fmla   v22.4s, v6.4s, v2.s[3]      \n"
-                        "nop                                \n"
-                        "fmla   v24.4s, v6.4s, v3.s[0]      \n"
-
-                        "ldr    d0, [%2]                    \n"
-                        "ins    v7.d[1], x27                \n"
-                        "fmla   v26.4s, v6.4s, v3.s[1]      \n"
-                        "ldr    x20, [%2, #8]               \n"
-                        "fmla   v28.4s, v6.4s, v3.s[2]      \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v30.4s, v6.4s, v3.s[3]      \n"
-
-                        "nop                                \n"
-                        "prfm   pldl1keep, [%2, #384]       \n" // NOTE PRELOAD
-                        "fmla   v9.4s, v7.4s, v1.s[0]       \n"
-                        "nop                                \n"
-                        "fmla   v11.4s, v7.4s, v1.s[1]      \n"
-                        "nop                                \n"
-                        "fmla   v13.4s, v7.4s, v1.s[2]      \n"
-
-                        "nop                                \n"
-                        "nop                                \n"
-                        "fmla   v15.4s, v7.4s, v1.s[3]      \n"
-                        "subs   w4, w4, #1                  \n"
-                        "fmla   v17.4s, v7.4s, v2.s[0]      \n"
-                        "nop                                \n"
-                        "fmla   v19.4s, v7.4s, v2.s[1]      \n"
-
-                        "ldr    d1, [%2]                    \n"
-                        "ins    v4.d[1], x24                \n"
-                        "fmla   v21.4s, v7.4s, v2.s[2]      \n"
-                        "ldr    x21, [%2, #8]               \n"
-                        "fmla   v23.4s, v7.4s, v2.s[3]      \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v25.4s, v7.4s, v3.s[0]      \n"
-
-                        "ldr    d2, [%2]                    \n"
-                        "ins    v0.d[1], x20                \n"
-                        "fmla   v27.4s, v7.4s, v3.s[1]      \n"
-                        "ldr    x22, [%2, #8]               \n"
-                        "fmla   v29.4s, v7.4s, v3.s[2]      \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v31.4s, v7.4s, v3.s[3]      \n"
-
-                        "bne    2b                          \n"
-
-                        "sub    %1, %1, #16                 \n"
-                        "sub    %2, %2, #48                 \n"
-
-                        "3:                                 \n"
-                        "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                        "cmp    w4, #0                      \n"
-                        "beq    5f                          \n"
-
-                        "4:                                 \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s}, [%2], #48 \n"
-                        "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-
-                        "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                        "fmla   v10.4s, v4.4s, v0.s[1]      \n"
-                        "fmla   v12.4s, v4.4s, v0.s[2]      \n"
-                        "fmla   v14.4s, v4.4s, v0.s[3]      \n"
-                        "fmla   v16.4s, v4.4s, v1.s[0]      \n"
-                        "fmla   v18.4s, v4.4s, v1.s[1]      \n"
-                        "fmla   v20.4s, v4.4s, v1.s[2]      \n"
-                        "fmla   v22.4s, v4.4s, v1.s[3]      \n"
-                        "fmla   v24.4s, v4.4s, v2.s[0]      \n"
-                        "fmla   v26.4s, v4.4s, v2.s[1]      \n"
-                        "fmla   v28.4s, v4.4s, v2.s[2]      \n"
-                        "fmla   v30.4s, v4.4s, v2.s[3]      \n"
-
-                        "subs   w4, w4, #1                  \n"
-
-                        "fmla   v9.4s, v5.4s, v0.s[0]       \n"
-                        "fmla   v11.4s, v5.4s, v0.s[1]      \n"
-                        "fmla   v13.4s, v5.4s, v0.s[2]      \n"
-                        "fmla   v15.4s, v5.4s, v0.s[3]      \n"
-                        "fmla   v17.4s, v5.4s, v1.s[0]      \n"
-                        "fmla   v19.4s, v5.4s, v1.s[1]      \n"
-                        "fmla   v21.4s, v5.4s, v1.s[2]      \n"
-                        "fmla   v23.4s, v5.4s, v1.s[3]      \n"
-                        "fmla   v25.4s, v5.4s, v2.s[0]      \n"
-                        "fmla   v27.4s, v5.4s, v2.s[1]      \n"
-                        "fmla   v29.4s, v5.4s, v2.s[2]      \n"
-                        "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-
-                        "bne    4b                          \n"
-
-                        "5:                                 \n"
-                        "st1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                        "st1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                        "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                        "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                        "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr), // %0
-                        "=r"(pA),     // %1
-                        "=r"(pB)      // %2
-                        : "0"(outptr),
-                        "1"(pA),
-                        "2"(pB),
-                        "r"(max_kk), // %6
-                        "r"(k)       // %7
-                        : "cc", "memory", "x4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-                else
-                {
-                    asm volatile(
-                        "cbz    %w7, 0f                     \n"
-
-                        "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                        "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                        "subs   %0, %0, #320                \n"
-                        "b      1f                          \n"
-
-                        "0:                                 \n"
-                        "eor    v8.16b, v8.16b, v8.16b      \n"
-                        "eor    v9.16b, v9.16b, v9.16b      \n"
-                        "eor    v10.16b, v10.16b, v10.16b   \n"
-                        "eor    v11.16b, v11.16b, v11.16b   \n"
-                        "eor    v12.16b, v12.16b, v12.16b   \n"
-                        "eor    v13.16b, v13.16b, v13.16b   \n"
-                        "eor    v14.16b, v14.16b, v14.16b   \n"
-                        "eor    v15.16b, v15.16b, v15.16b   \n"
-                        "eor    v16.16b, v16.16b, v16.16b   \n"
-                        "eor    v17.16b, v17.16b, v17.16b   \n"
-                        "eor    v18.16b, v18.16b, v18.16b   \n"
-                        "eor    v19.16b, v19.16b, v19.16b   \n"
-                        "eor    v20.16b, v20.16b, v20.16b   \n"
-                        "eor    v21.16b, v21.16b, v21.16b   \n"
-                        "eor    v22.16b, v22.16b, v22.16b   \n"
-                        "eor    v23.16b, v23.16b, v23.16b   \n"
-                        "eor    v24.16b, v24.16b, v24.16b   \n"
-                        "eor    v25.16b, v25.16b, v25.16b   \n"
-                        "eor    v26.16b, v26.16b, v26.16b   \n"
-                        "eor    v27.16b, v27.16b, v27.16b   \n"
-                        "eor    v28.16b, v28.16b, v28.16b   \n"
-                        "eor    v29.16b, v29.16b, v29.16b   \n"
-                        "eor    v30.16b, v30.16b, v30.16b   \n"
-                        "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                        "1:                                 \n"
-                        "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                        "cmp    w4, #0                      \n"
-                        "beq    3f                          \n"
-
-                        "2:                                 \n"
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-
-                        "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                        "fmla   v10.4s, v4.4s, v0.s[1]      \n"
-                        "fmla   v12.4s, v4.4s, v0.s[2]      \n"
-                        "fmla   v14.4s, v4.4s, v0.s[3]      \n"
-                        "fmla   v9.4s, v5.4s, v0.s[0]       \n"
-                        "fmla   v11.4s, v5.4s, v0.s[1]      \n"
-                        "fmla   v13.4s, v5.4s, v0.s[2]      \n"
-                        "fmla   v15.4s, v5.4s, v0.s[3]      \n"
-
-                        "fmla   v16.4s, v4.4s, v1.s[0]      \n"
-                        "fmla   v18.4s, v4.4s, v1.s[1]      \n"
-                        "fmla   v20.4s, v4.4s, v1.s[2]      \n"
-                        "fmla   v22.4s, v4.4s, v1.s[3]      \n"
-                        "fmla   v17.4s, v5.4s, v1.s[0]      \n"
-                        "fmla   v19.4s, v5.4s, v1.s[1]      \n"
-                        "fmla   v21.4s, v5.4s, v1.s[2]      \n"
-                        "fmla   v23.4s, v5.4s, v1.s[3]      \n"
-
-                        "fmla   v24.4s, v4.4s, v2.s[0]      \n"
-                        "fmla   v26.4s, v4.4s, v2.s[1]      \n"
-                        "fmla   v28.4s, v4.4s, v2.s[2]      \n"
-                        "fmla   v30.4s, v4.4s, v2.s[3]      \n"
-                        "fmla   v25.4s, v5.4s, v2.s[0]      \n"
-                        "fmla   v27.4s, v5.4s, v2.s[1]      \n"
-                        "fmla   v29.4s, v5.4s, v2.s[2]      \n"
-                        "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-
-                        "fmla   v8.4s, v6.4s, v3.s[0]       \n"
-                        "fmla   v10.4s, v6.4s, v3.s[1]      \n"
-                        "fmla   v12.4s, v6.4s, v3.s[2]      \n"
-                        "fmla   v14.4s, v6.4s, v3.s[3]      \n"
-                        "fmla   v9.4s, v7.4s, v3.s[0]       \n"
-                        "fmla   v11.4s, v7.4s, v3.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v3.s[2]      \n"
-                        "fmla   v15.4s, v7.4s, v3.s[3]      \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-
-                        "fmla   v16.4s, v6.4s, v0.s[0]      \n"
-                        "fmla   v18.4s, v6.4s, v0.s[1]      \n"
-                        "fmla   v20.4s, v6.4s, v0.s[2]      \n"
-                        "fmla   v22.4s, v6.4s, v0.s[3]      \n"
-                        "fmla   v17.4s, v7.4s, v0.s[0]      \n"
-                        "fmla   v19.4s, v7.4s, v0.s[1]      \n"
-                        "fmla   v21.4s, v7.4s, v0.s[2]      \n"
-                        "fmla   v23.4s, v7.4s, v0.s[3]      \n"
-
-                        "fmla   v24.4s, v6.4s, v1.s[0]      \n"
-                        "fmla   v26.4s, v6.4s, v1.s[1]      \n"
-                        "fmla   v28.4s, v6.4s, v1.s[2]      \n"
-                        "fmla   v30.4s, v6.4s, v1.s[3]      \n"
-                        "fmla   v25.4s, v7.4s, v1.s[0]      \n"
-                        "fmla   v27.4s, v7.4s, v1.s[1]      \n"
-                        "fmla   v29.4s, v7.4s, v1.s[2]      \n"
-                        "fmla   v31.4s, v7.4s, v1.s[3]      \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n"
-
-                        "fmla   v8.4s, v4.4s, v2.s[0]       \n"
-                        "fmla   v10.4s, v4.4s, v2.s[1]      \n"
-                        "fmla   v12.4s, v4.4s, v2.s[2]      \n"
-                        "fmla   v14.4s, v4.4s, v2.s[3]      \n"
-                        "fmla   v9.4s, v5.4s, v2.s[0]       \n"
-                        "fmla   v11.4s, v5.4s, v2.s[1]      \n"
-                        "fmla   v13.4s, v5.4s, v2.s[2]      \n"
-                        "fmla   v15.4s, v5.4s, v2.s[3]      \n"
-
-                        "fmla   v16.4s, v4.4s, v3.s[0]      \n"
-                        "fmla   v18.4s, v4.4s, v3.s[1]      \n"
-                        "fmla   v20.4s, v4.4s, v3.s[2]      \n"
-                        "fmla   v22.4s, v4.4s, v3.s[3]      \n"
-                        "fmla   v17.4s, v5.4s, v3.s[0]      \n"
-                        "fmla   v19.4s, v5.4s, v3.s[1]      \n"
-                        "fmla   v21.4s, v5.4s, v3.s[2]      \n"
-                        "fmla   v23.4s, v5.4s, v3.s[3]      \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-
-                        "fmla   v24.4s, v4.4s, v0.s[0]      \n"
-                        "fmla   v26.4s, v4.4s, v0.s[1]      \n"
-                        "fmla   v28.4s, v4.4s, v0.s[2]      \n"
-                        "fmla   v30.4s, v4.4s, v0.s[3]      \n"
-                        "fmla   v25.4s, v5.4s, v0.s[0]      \n"
-                        "fmla   v27.4s, v5.4s, v0.s[1]      \n"
-                        "fmla   v29.4s, v5.4s, v0.s[2]      \n"
-                        "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-
-                        "fmla   v8.4s, v6.4s, v1.s[0]       \n"
-                        "fmla   v10.4s, v6.4s, v1.s[1]      \n"
-                        "fmla   v12.4s, v6.4s, v1.s[2]      \n"
-                        "fmla   v14.4s, v6.4s, v1.s[3]      \n"
-                        "fmla   v9.4s, v7.4s, v1.s[0]       \n"
-                        "fmla   v11.4s, v7.4s, v1.s[1]      \n"
-                        "fmla   v13.4s, v7.4s, v1.s[2]      \n"
-                        "fmla   v15.4s, v7.4s, v1.s[3]      \n"
-
-                        "fmla   v16.4s, v6.4s, v2.s[0]      \n"
-                        "fmla   v18.4s, v6.4s, v2.s[1]      \n"
-                        "fmla   v20.4s, v6.4s, v2.s[2]      \n"
-                        "fmla   v22.4s, v6.4s, v2.s[3]      \n"
-                        "fmla   v17.4s, v7.4s, v2.s[0]      \n"
-                        "fmla   v19.4s, v7.4s, v2.s[1]      \n"
-                        "fmla   v21.4s, v7.4s, v2.s[2]      \n"
-                        "fmla   v23.4s, v7.4s, v2.s[3]      \n"
-
-                        "subs   w4, w4, #1                  \n"
-
-                        "fmla   v24.4s, v6.4s, v3.s[0]      \n"
-                        "fmla   v26.4s, v6.4s, v3.s[1]      \n"
-                        "fmla   v28.4s, v6.4s, v3.s[2]      \n"
-                        "fmla   v30.4s, v6.4s, v3.s[3]      \n"
-                        "fmla   v25.4s, v7.4s, v3.s[0]      \n"
-                        "fmla   v27.4s, v7.4s, v3.s[1]      \n"
-                        "fmla   v29.4s, v7.4s, v3.s[2]      \n"
-                        "fmla   v31.4s, v7.4s, v3.s[3]      \n"
-
-                        "bne    2b                          \n"
-
-                        "3:                                 \n"
-                        "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                        "cmp    w4, #0                      \n"
-                        "beq    5f                          \n"
-
-                        "4:                                 \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s}, [%2], #48 \n"
-                        "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-
-                        "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                        "fmla   v10.4s, v4.4s, v0.s[1]      \n"
-                        "fmla   v12.4s, v4.4s, v0.s[2]      \n"
-                        "fmla   v14.4s, v4.4s, v0.s[3]      \n"
-                        "fmla   v16.4s, v4.4s, v1.s[0]      \n"
-                        "fmla   v18.4s, v4.4s, v1.s[1]      \n"
-                        "fmla   v20.4s, v4.4s, v1.s[2]      \n"
-                        "fmla   v22.4s, v4.4s, v1.s[3]      \n"
-                        "fmla   v24.4s, v4.4s, v2.s[0]      \n"
-                        "fmla   v26.4s, v4.4s, v2.s[1]      \n"
-                        "fmla   v28.4s, v4.4s, v2.s[2]      \n"
-                        "fmla   v30.4s, v4.4s, v2.s[3]      \n"
-
-                        "subs   w4, w4, #1                  \n"
-
-                        "fmla   v9.4s, v5.4s, v0.s[0]       \n"
-                        "fmla   v11.4s, v5.4s, v0.s[1]      \n"
-                        "fmla   v13.4s, v5.4s, v0.s[2]      \n"
-                        "fmla   v15.4s, v5.4s, v0.s[3]      \n"
-                        "fmla   v17.4s, v5.4s, v1.s[0]      \n"
-                        "fmla   v19.4s, v5.4s, v1.s[1]      \n"
-                        "fmla   v21.4s, v5.4s, v1.s[2]      \n"
-                        "fmla   v23.4s, v5.4s, v1.s[3]      \n"
-                        "fmla   v25.4s, v5.4s, v2.s[0]      \n"
-                        "fmla   v27.4s, v5.4s, v2.s[1]      \n"
-                        "fmla   v29.4s, v5.4s, v2.s[2]      \n"
-                        "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-
-                        "bne    4b                          \n"
-
-                        "5:                                 \n"
-                        "st1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                        "st1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                        "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                        "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                        "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr), // %0
-                        "=r"(pA),     // %1
-                        "=r"(pB)      // %2
-                        : "0"(outptr),
-                        "1"(pA),
-                        "2"(pB),
-                        "r"(max_kk), // %6
-                        "r"(k)       // %7
-                        : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4_t _sum00;
-                float32x4_t _sum01;
-                float32x4_t _sum10;
-                float32x4_t _sum11;
-                float32x4_t _sum20;
-                float32x4_t _sum21;
-                float32x4_t _sum30;
-                float32x4_t _sum31;
-                float32x4_t _sum40;
-                float32x4_t _sum41;
-                float32x4_t _sum50;
-                float32x4_t _sum51;
-                float32x4_t _sum60;
-                float32x4_t _sum61;
-                float32x4_t _sum70;
-                float32x4_t _sum71;
-                float32x4_t _sum80;
-                float32x4_t _sum81;
-                float32x4_t _sum90;
-                float32x4_t _sum91;
-                float32x4_t _suma0;
-                float32x4_t _suma1;
-                float32x4_t _sumb0;
-                float32x4_t _sumb1;
-
-                if (k == 0)
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                    _sum20 = vdupq_n_f32(0.f);
-                    _sum21 = vdupq_n_f32(0.f);
-                    _sum30 = vdupq_n_f32(0.f);
-                    _sum31 = vdupq_n_f32(0.f);
-                    _sum40 = vdupq_n_f32(0.f);
-                    _sum41 = vdupq_n_f32(0.f);
-                    _sum50 = vdupq_n_f32(0.f);
-                    _sum51 = vdupq_n_f32(0.f);
-                    _sum60 = vdupq_n_f32(0.f);
-                    _sum61 = vdupq_n_f32(0.f);
-                    _sum70 = vdupq_n_f32(0.f);
-                    _sum71 = vdupq_n_f32(0.f);
-                    _sum80 = vdupq_n_f32(0.f);
-                    _sum81 = vdupq_n_f32(0.f);
-                    _sum90 = vdupq_n_f32(0.f);
-                    _sum91 = vdupq_n_f32(0.f);
-                    _suma0 = vdupq_n_f32(0.f);
-                    _suma1 = vdupq_n_f32(0.f);
-                    _sumb0 = vdupq_n_f32(0.f);
-                    _sumb1 = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    _sum00 = vld1q_f32(outptr);
-                    _sum01 = vld1q_f32(outptr + 4 * 1);
-                    _sum10 = vld1q_f32(outptr + 4 * 2);
-                    _sum11 = vld1q_f32(outptr + 4 * 3);
-                    _sum20 = vld1q_f32(outptr + 4 * 4);
-                    _sum21 = vld1q_f32(outptr + 4 * 5);
-                    _sum30 = vld1q_f32(outptr + 4 * 6);
-                    _sum31 = vld1q_f32(outptr + 4 * 7);
-                    _sum40 = vld1q_f32(outptr + 4 * 8);
-                    _sum41 = vld1q_f32(outptr + 4 * 9);
-                    _sum50 = vld1q_f32(outptr + 4 * 10);
-                    _sum51 = vld1q_f32(outptr + 4 * 11);
-                    _sum60 = vld1q_f32(outptr + 4 * 12);
-                    _sum61 = vld1q_f32(outptr + 4 * 13);
-                    _sum70 = vld1q_f32(outptr + 4 * 14);
-                    _sum71 = vld1q_f32(outptr + 4 * 15);
-                    _sum80 = vld1q_f32(outptr + 4 * 16);
-                    _sum81 = vld1q_f32(outptr + 4 * 17);
-                    _sum90 = vld1q_f32(outptr + 4 * 18);
-                    _sum91 = vld1q_f32(outptr + 4 * 19);
-                    _suma0 = vld1q_f32(outptr + 4 * 20);
-                    _suma1 = vld1q_f32(outptr + 4 * 21);
-                    _sumb0 = vld1q_f32(outptr + 4 * 22);
-                    _sumb1 = vld1q_f32(outptr + 4 * 23);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pA0 = vld1q_f32(pA);
-                    float32x4_t _pA1 = vld1q_f32(pA + 4);
-
-                    float32x4_t _pB0 = vld1q_f32(pB);
-                    float32x4_t _pB1 = vld1q_f32(pB + 4);
-                    float32x4_t _pB2 = vld1q_f32(pB + 8);
-
-                    _sum00 = vfmaq_laneq_f32(_sum00, _pA0, _pB0, 0);
-                    _sum01 = vfmaq_laneq_f32(_sum01, _pA1, _pB0, 0);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _pA0, _pB0, 1);
-                    _sum11 = vfmaq_laneq_f32(_sum11, _pA1, _pB0, 1);
-                    _sum20 = vfmaq_laneq_f32(_sum20, _pA0, _pB0, 2);
-                    _sum21 = vfmaq_laneq_f32(_sum21, _pA1, _pB0, 2);
-                    _sum30 = vfmaq_laneq_f32(_sum30, _pA0, _pB0, 3);
-                    _sum31 = vfmaq_laneq_f32(_sum31, _pA1, _pB0, 3);
-                    _sum40 = vfmaq_laneq_f32(_sum40, _pA0, _pB1, 0);
-                    _sum41 = vfmaq_laneq_f32(_sum41, _pA1, _pB1, 0);
-                    _sum50 = vfmaq_laneq_f32(_sum50, _pA0, _pB1, 1);
-                    _sum51 = vfmaq_laneq_f32(_sum51, _pA1, _pB1, 1);
-                    _sum60 = vfmaq_laneq_f32(_sum60, _pA0, _pB1, 2);
-                    _sum61 = vfmaq_laneq_f32(_sum61, _pA1, _pB1, 2);
-                    _sum70 = vfmaq_laneq_f32(_sum70, _pA0, _pB1, 3);
-                    _sum71 = vfmaq_laneq_f32(_sum71, _pA1, _pB1, 3);
-                    _sum80 = vfmaq_laneq_f32(_sum80, _pA0, _pB2, 0);
-                    _sum81 = vfmaq_laneq_f32(_sum81, _pA1, _pB2, 0);
-                    _sum90 = vfmaq_laneq_f32(_sum90, _pA0, _pB2, 1);
-                    _sum91 = vfmaq_laneq_f32(_sum91, _pA1, _pB2, 1);
-                    _suma0 = vfmaq_laneq_f32(_suma0, _pA0, _pB2, 2);
-                    _suma1 = vfmaq_laneq_f32(_suma1, _pA1, _pB2, 2);
-                    _sumb0 = vfmaq_laneq_f32(_sumb0, _pA0, _pB2, 3);
-                    _sumb1 = vfmaq_laneq_f32(_sumb1, _pA1, _pB2, 3);
-
-                    pA += 8;
-                    pB += 12;
-                }
-
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                vst1q_f32(outptr + 4 * 4, _sum20);
-                vst1q_f32(outptr + 4 * 5, _sum21);
-                vst1q_f32(outptr + 4 * 6, _sum30);
-                vst1q_f32(outptr + 4 * 7, _sum31);
-                vst1q_f32(outptr + 4 * 8, _sum40);
-                vst1q_f32(outptr + 4 * 9, _sum41);
-                vst1q_f32(outptr + 4 * 10, _sum50);
-                vst1q_f32(outptr + 4 * 11, _sum51);
-                vst1q_f32(outptr + 4 * 12, _sum60);
-                vst1q_f32(outptr + 4 * 13, _sum61);
-                vst1q_f32(outptr + 4 * 14, _sum70);
-                vst1q_f32(outptr + 4 * 15, _sum71);
-                vst1q_f32(outptr + 4 * 16, _sum80);
-                vst1q_f32(outptr + 4 * 17, _sum81);
-                vst1q_f32(outptr + 4 * 18, _sum90);
-                vst1q_f32(outptr + 4 * 19, _sum91);
-                vst1q_f32(outptr + 4 * 20, _suma0);
-                vst1q_f32(outptr + 4 * 21, _suma1);
-                vst1q_f32(outptr + 4 * 22, _sumb0);
-                vst1q_f32(outptr + 4 * 23, _sumb1);
-                outptr += 8 * 12;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-                if (use_a53_a55_optimized_kernel && cpu_support_arm_asimdhp())
-                {
-                    // a55
-                    asm volatile(
-                        "cbz    %w7, 0f                     \n"
-
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                        "subs   %0, %0, #192                \n"
-                        "b      1f                          \n"
-
-                        "0:                                 \n"
-                        "eor    v16.16b, v16.16b, v16.16b   \n"
-                        "eor    v17.16b, v17.16b, v17.16b   \n"
-                        "eor    v18.16b, v18.16b, v18.16b   \n"
-                        "eor    v19.16b, v19.16b, v19.16b   \n"
-                        "eor    v20.16b, v20.16b, v20.16b   \n"
-                        "eor    v21.16b, v21.16b, v21.16b   \n"
-                        "eor    v22.16b, v22.16b, v22.16b   \n"
-                        "eor    v23.16b, v23.16b, v23.16b   \n"
-                        "eor    v24.16b, v24.16b, v24.16b   \n"
-                        "eor    v25.16b, v25.16b, v25.16b   \n"
-                        "eor    v26.16b, v26.16b, v26.16b   \n"
-                        "eor    v27.16b, v27.16b, v27.16b   \n"
-                        "eor    v28.16b, v28.16b, v28.16b   \n"
-                        "eor    v29.16b, v29.16b, v29.16b   \n"
-                        "eor    v30.16b, v30.16b, v30.16b   \n"
-                        "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                        "1:                                 \n"
-                        "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                        "cmp    w4, #0                      \n"
-                        "beq    3f                          \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v8.4s}, [%1], #16          \n"
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.4s}, [%2], #16          \n"
-
-                        "ldr    d1, [%2], #8                \n"
-                        "ldr    x21, [%2], #8               \n"
-
-                        ".align 4                           \n"
-                        "2:                                 \n"
-                        "ldr    d9, [%1], #8                \n"
-                        "fmla   v16.4s, v8.4s, v0.s[0]      \n"
-                        "ldr    x25, [%1], #8               \n"
-                        "fmla   v18.4s, v8.4s, v0.s[1]      \n"
-                        "ins    v1.d[1], x21                \n"
-                        "fmla   v20.4s, v8.4s, v0.s[2]      \n"
-                        "ldr    d10, [%1], #8               \n"
-                        "fmla   v22.4s, v8.4s, v0.s[3]      \n"
-                        "ldr    x26, [%1], #8               \n"
-                        "fmla   v24.4s, v8.4s, v1.s[0]      \n"
-                        "ldr    d2, [%2], #8                \n"
-                        "fmla   v26.4s, v8.4s, v1.s[1]      \n"
-                        "ins    v9.d[1], x25                \n"
-                        "fmla   v28.4s, v8.4s, v1.s[2]      \n"
-                        "ldr    x22, [%2], #8               \n"
-                        "fmla   v30.4s, v8.4s, v1.s[3]      \n"
-                        "ldr    d3, [%2], #8                \n"
-                        "fmla   v17.4s, v9.4s, v0.s[0]      \n"
-                        "ldr    x23, [%2], #8               \n"
-                        "fmla   v19.4s, v9.4s, v0.s[1]      \n"
-                        "ins    v10.d[1], x26               \n"
-                        "fmla   v21.4s, v9.4s, v0.s[2]      \n"
-                        "ldr    d11, [%1], #8               \n"
-                        "fmla   v23.4s, v9.4s, v0.s[3]      \n"
-                        "ldr    x27, [%1], #8               \n"
-                        "fmla   v25.4s, v9.4s, v1.s[0]      \n"
-                        "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                        "fmla   v27.4s, v9.4s, v1.s[1]      \n"
-                        "ins    v2.d[1], x22                \n"
-                        "fmla   v29.4s, v9.4s, v1.s[2]      \n"
-                        "ldr    d12, [%1], #8               \n"
-                        "fmla   v31.4s, v9.4s, v1.s[3]      \n"
-                        "ldr    x24, [%1], #8               \n"
-                        "fmla   v16.4s, v10.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                        "fmla   v18.4s, v10.4s, v2.s[1]     \n"
-                        "ins    v3.d[1], x23                \n"
-                        "fmla   v20.4s, v10.4s, v2.s[2]     \n"
-                        "ldr    d4, [%2], #8                \n"
-                        "fmla   v22.4s, v10.4s, v2.s[3]     \n"
-                        "ldr    x20, [%2], #8               \n"
-                        "fmla   v24.4s, v10.4s, v3.s[0]     \n"
-                        "ldr    d5, [%2], #8                \n"
-                        "fmla   v26.4s, v10.4s, v3.s[1]     \n"
-                        "ins    v11.d[1], x27               \n"
-                        "fmla   v28.4s, v10.4s, v3.s[2]     \n"
-                        "ldr    x21, [%2], #8               \n"
-                        "fmla   v30.4s, v10.4s, v3.s[3]     \n"
-                        "ldr    d13, [%1], #8               \n"
-                        "fmla   v17.4s, v11.4s, v2.s[0]     \n"
-                        "ldr    x25, [%1], #8               \n"
-                        "fmla   v19.4s, v11.4s, v2.s[1]     \n"
-                        "ins    v12.d[1], x24               \n"
-                        "fmla   v21.4s, v11.4s, v2.s[2]     \n"
-                        "ldr    d14, [%1], #8               \n"
-                        "fmla   v23.4s, v11.4s, v2.s[3]     \n"
-                        "ldr    x26, [%1], #8               \n"
-                        "fmla   v25.4s, v11.4s, v3.s[0]     \n"
-                        "ldr    d6, [%2], #8                \n"
-                        "fmla   v27.4s, v11.4s, v3.s[1]     \n"
-                        "ins    v4.d[1], x20                \n"
-                        "fmla   v29.4s, v11.4s, v3.s[2]     \n"
-                        "ldr    x22, [%2], #8               \n"
-                        "fmla   v31.4s, v11.4s, v3.s[3]     \n"
-                        "ldr    d7, [%2], #8                \n"
-                        "fmla   v16.4s, v12.4s, v4.s[0]     \n"
-                        "ldr    x23, [%2], #8               \n"
-                        "fmla   v18.4s, v12.4s, v4.s[1]     \n"
-                        "ins    v5.d[1], x21                \n"
-                        "fmla   v20.4s, v12.4s, v4.s[2]     \n"
-                        "ldr    d15, [%1], #8               \n"
-                        "fmla   v22.4s, v12.4s, v4.s[3]     \n"
-                        "ldr    x27, [%1], #8               \n"
-                        "fmla   v24.4s, v12.4s, v5.s[0]     \n"
-                        "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                        "fmla   v26.4s, v12.4s, v5.s[1]     \n"
-                        "ins    v13.d[1], x25               \n"
-                        "fmla   v28.4s, v12.4s, v5.s[2]     \n"
-                        "ldr    d8, [%1], #8                \n"
-                        "fmla   v30.4s, v12.4s, v5.s[3]     \n"
-                        "ldr    x24, [%1], #8               \n"
-                        "fmla   v17.4s, v13.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                        "fmla   v19.4s, v13.4s, v4.s[1]     \n"
-                        "ins    v14.d[1], x26               \n"
-                        "fmla   v21.4s, v13.4s, v4.s[2]     \n"
-                        "ldr    d0, [%2], #8                \n"
-                        "fmla   v23.4s, v13.4s, v4.s[3]     \n"
-                        "ldr    x20, [%2], #8               \n"
-                        "fmla   v25.4s, v13.4s, v5.s[0]     \n"
-                        "ldr    d1, [%2], #8                \n"
-                        "fmla   v27.4s, v13.4s, v5.s[1]     \n"
-                        "ins    v6.d[1], x22                \n"
-                        "fmla   v29.4s, v13.4s, v5.s[2]     \n"
-                        "ldr    x21, [%2], #8               \n"
-                        "fmla   v31.4s, v13.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v14.4s, v6.s[0]     \n"
-                        "fmla   v18.4s, v14.4s, v6.s[1]     \n"
-                        "ins    v7.d[1], x23                \n"
-                        "fmla   v20.4s, v14.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v14.4s, v6.s[3]     \n"
-                        "fmla   v24.4s, v14.4s, v7.s[0]     \n"
-                        "fmla   v26.4s, v14.4s, v7.s[1]     \n"
-                        "ins    v15.d[1], x27               \n"
-                        "fmla   v28.4s, v14.4s, v7.s[2]     \n"
-                        "fmla   v30.4s, v14.4s, v7.s[3]     \n"
-                        "fmla   v17.4s, v15.4s, v6.s[0]     \n"
-                        "fmla   v19.4s, v15.4s, v6.s[1]     \n"
-                        "ins    v8.d[1], x24                \n"
-                        "fmla   v21.4s, v15.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v15.4s, v6.s[3]     \n"
-                        "fmla   v25.4s, v15.4s, v7.s[0]     \n"
-                        "subs   w4, w4, #1                  \n"
-                        "fmla   v27.4s, v15.4s, v7.s[1]     \n"
-                        "fmla   v29.4s, v15.4s, v7.s[2]     \n"
-                        "ins    v0.d[1], x20                \n"
-                        "fmla   v31.4s, v15.4s, v7.s[3]     \n"
-                        "bne    2b                          \n"
-
-                        "sub    %1, %1, #16                 \n"
-                        "sub    %2, %2, #32                 \n"
-
-                        "3:                                 \n"
-                        "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                        "cmp    w4, #0                      \n"
-                        "beq    5f                          \n"
-
-                        "4:                                 \n"
-                        "ld1    {v0.4s, v1.4s}, [%2], #32   \n"
-                        "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-                        "fmla   v16.4s, v4.4s, v0.s[0]      \n"
-                        "fmla   v18.4s, v4.4s, v0.s[1]      \n"
-                        "fmla   v20.4s, v4.4s, v0.s[2]      \n"
-                        "fmla   v22.4s, v4.4s, v0.s[3]      \n"
-                        "fmla   v17.4s, v5.4s, v0.s[0]      \n"
-                        "fmla   v19.4s, v5.4s, v0.s[1]      \n"
-                        "fmla   v21.4s, v5.4s, v0.s[2]      \n"
-                        "fmla   v23.4s, v5.4s, v0.s[3]      \n"
-                        "subs   w4, w4, #1                  \n"
-                        "fmla   v24.4s, v4.4s, v1.s[0]      \n"
-                        "fmla   v26.4s, v4.4s, v1.s[1]      \n"
-                        "fmla   v28.4s, v4.4s, v1.s[2]      \n"
-                        "fmla   v30.4s, v4.4s, v1.s[3]      \n"
-                        "fmla   v25.4s, v5.4s, v1.s[0]      \n"
-                        "fmla   v27.4s, v5.4s, v1.s[1]      \n"
-                        "fmla   v29.4s, v5.4s, v1.s[2]      \n"
-                        "fmla   v31.4s, v5.4s, v1.s[3]      \n"
-                        "bne    4b                          \n"
-
-                        "5:                                 \n"
-                        "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                        "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                        "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr), // %0
-                        "=r"(pA),     // %1
-                        "=r"(pB)      // %2
-                        : "0"(outptr),
-                        "1"(pA),
-                        "2"(pB),
-                        "r"(max_kk), // %6
-                        "r"(k)       // %7
-                        : "cc", "memory", "x4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-                else if (use_a53_a55_optimized_kernel && !cpu_support_arm_asimdhp())
-                {
-                    // a53
-                    asm volatile(
-                        "cbz    %w7, 0f                     \n"
-
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                        "subs   %0, %0, #192                \n"
-                        "b      1f                          \n"
-
-                        "0:                                 \n"
-                        "eor    v16.16b, v16.16b, v16.16b   \n"
-                        "eor    v17.16b, v17.16b, v17.16b   \n"
-                        "eor    v18.16b, v18.16b, v18.16b   \n"
-                        "eor    v19.16b, v19.16b, v19.16b   \n"
-                        "eor    v20.16b, v20.16b, v20.16b   \n"
-                        "eor    v21.16b, v21.16b, v21.16b   \n"
-                        "eor    v22.16b, v22.16b, v22.16b   \n"
-                        "eor    v23.16b, v23.16b, v23.16b   \n"
-                        "eor    v24.16b, v24.16b, v24.16b   \n"
-                        "eor    v25.16b, v25.16b, v25.16b   \n"
-                        "eor    v26.16b, v26.16b, v26.16b   \n"
-                        "eor    v27.16b, v27.16b, v27.16b   \n"
-                        "eor    v28.16b, v28.16b, v28.16b   \n"
-                        "eor    v29.16b, v29.16b, v29.16b   \n"
-                        "eor    v30.16b, v30.16b, v30.16b   \n"
-                        "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                        "1:                                 \n"
-                        "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                        "cmp    w4, #0                      \n"
-                        "beq    3f                          \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ldr    d0, [%2]                    \n"
-                        "ldr    x20, [%2, #8]               \n"
-                        "ins    v0.d[1], x20                \n"
-                        "add    %2, %2, #16                 \n"
-
-                        "ldr    d8, [%1]                    \n"
-                        "ldr    x24, [%1, #8]               \n"
-                        "ins    v8.d[1], x24                \n"
-                        "add    %1, %1, #16                 \n"
-
-                        "ldr    d1, [%2]                    \n"
-                        "ldr    x21, [%2, #8]               \n"
-                        "add    %2, %2, #16                 \n"
-
-                        "ldr    d9, [%1]                    \n"
-                        "ldr    x25, [%1, #8]               \n"
-                        "add    %1, %1, #16                 \n"
-
-                        ".align 4                           \n"
-                        "2:                                 \n"
-
-                        "ldr    d2, [%2]                    \n"
-                        "ins    v1.d[1], x21                \n"
-                        "fmla   v16.4s, v8.4s, v0.s[0]      \n"
-                        "ldr    x22, [%2, #8]               \n"
-                        "fmla   v18.4s, v8.4s, v0.s[1]      \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v20.4s, v8.4s, v0.s[2]      \n"
-
-                        "ldr    d10, [%1]                   \n"
-                        "ins    v9.d[1], x25                \n"
-                        "fmla   v22.4s, v8.4s, v0.s[3]      \n"
-                        "ldr    x26, [%1, #8]               \n"
-                        "fmla   v24.4s, v8.4s, v1.s[0]      \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v26.4s, v8.4s, v1.s[1]      \n"
-
-                        "ldr    d3, [%2]                    \n"
-                        "ins    v2.d[1], x22                \n"
-                        "fmla   v28.4s, v8.4s, v1.s[2]      \n"
-                        "ldr    x23, [%2, #8]               \n"
-                        "fmla   v30.4s, v8.4s, v1.s[3]      \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v17.4s, v9.4s, v0.s[0]      \n"
-
-                        "nop                                \n"
-                        "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                        "fmla   v19.4s, v9.4s, v0.s[1]      \n"
-                        "nop                                \n"
-                        "fmla   v21.4s, v9.4s, v0.s[2]      \n"
-                        "nop                                \n"
-                        "fmla   v23.4s, v9.4s, v0.s[3]      \n"
-
-                        "ldr    d11, [%1]                   \n"
-                        "ins    v10.d[1], x26               \n"
-                        "fmla   v25.4s, v9.4s, v1.s[0]      \n"
-                        "ldr    x27, [%1, #8]               \n"
-                        "fmla   v27.4s, v9.4s, v1.s[1]      \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v29.4s, v9.4s, v1.s[2]      \n"
-
-                        "nop                                \n"
-                        "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                        "fmla   v31.4s, v9.4s, v1.s[3]      \n"
-                        "nop                                \n"
-                        "fmla   v16.4s, v10.4s, v2.s[0]     \n"
-                        "nop                                \n"
-                        "fmla   v18.4s, v10.4s, v2.s[1]     \n"
-
-                        "ldr    d4, [%2]                    \n"
-                        "ins    v3.d[1], x23                \n"
-                        "fmla   v20.4s, v10.4s, v2.s[2]     \n"
-                        "ldr    x20, [%2, #8]               \n"
-                        "fmla   v22.4s, v10.4s, v2.s[3]     \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v24.4s, v10.4s, v3.s[0]     \n"
-
-                        "ldr    d12, [%1]                   \n"
-                        "ins    v11.d[1], x27               \n"
-                        "fmla   v26.4s, v10.4s, v3.s[1]     \n"
-                        "ldr    x24, [%1, #8]               \n"
-                        "fmla   v28.4s, v10.4s, v3.s[2]     \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v30.4s, v10.4s, v3.s[3]     \n"
-
-                        "ldr    d5, [%2]                    \n"
-                        "ins    v4.d[1], x20                \n"
-                        "fmla   v17.4s, v11.4s, v2.s[0]     \n"
-                        "ldr    x21, [%2, #8]               \n"
-                        "fmla   v19.4s, v11.4s, v2.s[1]     \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v21.4s, v11.4s, v2.s[2]     \n"
-
-                        "ldr    d13, [%1]                   \n"
-                        "ins    v12.d[1], x24               \n"
-                        "fmla   v23.4s, v11.4s, v2.s[3]     \n"
-                        "ldr    x25, [%1, #8]               \n"
-                        "fmla   v25.4s, v11.4s, v3.s[0]     \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v27.4s, v11.4s, v3.s[1]     \n"
-
-                        "ldr    d6, [%2]                    \n"
-                        "ins    v5.d[1], x21                \n"
-                        "fmla   v29.4s, v11.4s, v3.s[2]     \n"
-                        "ldr    x22, [%2, #8]               \n"
-                        "fmla   v31.4s, v11.4s, v3.s[3]     \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v16.4s, v12.4s, v4.s[0]     \n"
-
-                        "ldr    d14, [%1]                   \n"
-                        "ins    v13.d[1], x25               \n"
-                        "fmla   v18.4s, v12.4s, v4.s[1]     \n"
-                        "ldr    x26, [%1, #8]               \n"
-                        "fmla   v20.4s, v12.4s, v4.s[2]     \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v22.4s, v12.4s, v4.s[3]     \n"
-
-                        "ldr    d7, [%2]                    \n"
-                        "ins    v6.d[1], x22                \n"
-                        "fmla   v24.4s, v12.4s, v5.s[0]     \n"
-                        "ldr    x23, [%2, #8]               \n"
-                        "fmla   v26.4s, v12.4s, v5.s[1]     \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v28.4s, v12.4s, v5.s[2]     \n"
-
-                        "nop                                \n"
-                        "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                        "fmla   v30.4s, v12.4s, v5.s[3]     \n"
-                        "nop                                \n"
-                        "fmla   v17.4s, v13.4s, v4.s[0]     \n"
-                        "nop                                \n"
-                        "fmla   v19.4s, v13.4s, v4.s[1]     \n"
-
-                        "ldr    d15, [%1]                   \n"
-                        "ins    v14.d[1], x26               \n"
-                        "fmla   v21.4s, v13.4s, v4.s[2]     \n"
-                        "ldr    x27, [%1, #8]               \n"
-                        "fmla   v23.4s, v13.4s, v4.s[3]     \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v25.4s, v13.4s, v5.s[0]     \n"
-
-                        "nop                                \n"
-                        "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                        "fmla   v27.4s, v13.4s, v5.s[1]     \n"
-                        "nop                                \n"
-                        "fmla   v29.4s, v13.4s, v5.s[2]     \n"
-                        "nop                                \n"
-                        "fmla   v31.4s, v13.4s, v5.s[3]     \n"
-
-                        "ldr    d0, [%2]                    \n"
-                        "ins    v7.d[1], x23                \n"
-                        "fmla   v16.4s, v14.4s, v6.s[0]     \n"
-                        "ldr    x20, [%2, #8]               \n"
-                        "fmla   v18.4s, v14.4s, v6.s[1]     \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v20.4s, v14.4s, v6.s[2]     \n"
-
-                        "ldr    d8, [%1]                    \n"
-                        "ins    v15.d[1], x27               \n"
-                        "fmla   v22.4s, v14.4s, v6.s[3]     \n"
-                        "ldr    x24, [%1, #8]               \n"
-                        "fmla   v24.4s, v14.4s, v7.s[0]     \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v26.4s, v14.4s, v7.s[1]     \n"
-
-                        "ldr    d1, [%2]                    \n"
-                        "ins    v0.d[1], x20                \n"
-                        "fmla   v28.4s, v14.4s, v7.s[2]     \n"
-                        "ldr    x21, [%2, #8]               \n"
-                        "fmla   v30.4s, v14.4s, v7.s[3]     \n"
-                        "add    %2, %2, #16                 \n"
-                        "fmla   v17.4s, v15.4s, v6.s[0]     \n"
-
-                        "ldr    d9, [%1]                    \n"
-                        "ins    v8.d[1], x24                \n"
-                        "fmla   v19.4s, v15.4s, v6.s[1]     \n"
-                        "ldr    x25, [%1, #8]               \n"
-                        "fmla   v21.4s, v15.4s, v6.s[2]     \n"
-                        "add    %1, %1, #16                 \n"
-                        "fmla   v23.4s, v15.4s, v6.s[3]     \n"
-
-                        "nop                                \n"
-                        "nop                                \n"
-                        "fmla   v25.4s, v15.4s, v7.s[0]     \n"
-                        "subs   w4, w4, #1                  \n"
-                        "fmla   v27.4s, v15.4s, v7.s[1]     \n"
-                        "nop                                \n"
-                        "fmla   v29.4s, v15.4s, v7.s[2]     \n"
-
-                        "nop                                \n"
-                        "nop                                \n"
-                        "fmla   v31.4s, v15.4s, v7.s[3]     \n"
-                        "nop                                \n"
-                        "nop                                \n"
-                        "nop                                \n"
-                        "nop                                \n"
-
-                        "bne    2b                          \n"
-
-                        "sub    %1, %1, #32                 \n"
-                        "sub    %2, %2, #32                 \n"
-
-                        "3:                                 \n"
-                        "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                        "cmp    w4, #0                      \n"
-                        "beq    5f                          \n"
-
-                        "4:                                 \n"
-                        "ld1    {v0.4s, v1.4s}, [%2], #32   \n"
-                        "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-                        "fmla   v16.4s, v4.4s, v0.s[0]      \n"
-                        "fmla   v18.4s, v4.4s, v0.s[1]      \n"
-                        "fmla   v20.4s, v4.4s, v0.s[2]      \n"
-                        "fmla   v22.4s, v4.4s, v0.s[3]      \n"
-                        "fmla   v17.4s, v5.4s, v0.s[0]      \n"
-                        "fmla   v19.4s, v5.4s, v0.s[1]      \n"
-                        "fmla   v21.4s, v5.4s, v0.s[2]      \n"
-                        "fmla   v23.4s, v5.4s, v0.s[3]      \n"
-                        "subs   w4, w4, #1                  \n"
-                        "fmla   v24.4s, v4.4s, v1.s[0]      \n"
-                        "fmla   v26.4s, v4.4s, v1.s[1]      \n"
-                        "fmla   v28.4s, v4.4s, v1.s[2]      \n"
-                        "fmla   v30.4s, v4.4s, v1.s[3]      \n"
-                        "fmla   v25.4s, v5.4s, v1.s[0]      \n"
-                        "fmla   v27.4s, v5.4s, v1.s[1]      \n"
-                        "fmla   v29.4s, v5.4s, v1.s[2]      \n"
-                        "fmla   v31.4s, v5.4s, v1.s[3]      \n"
-                        "bne    4b                          \n"
-
-                        "5:                                 \n"
-                        "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                        "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                        "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr), // %0
-                        "=r"(pA),     // %1
-                        "=r"(pB)      // %2
-                        : "0"(outptr),
-                        "1"(pA),
-                        "2"(pB),
-                        "r"(max_kk), // %6
-                        "r"(k)       // %7
-                        : "cc", "memory", "x4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-                else
-                {
-                    asm volatile(
-                        "cbz    %w7, 0f                     \n"
-
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                        "subs   %0, %0, #192                \n"
-                        "b      1f                          \n"
-
-                        "0:                                 \n"
-                        "eor    v16.16b, v16.16b, v16.16b   \n"
-                        "eor    v17.16b, v17.16b, v17.16b   \n"
-                        "eor    v18.16b, v18.16b, v18.16b   \n"
-                        "eor    v19.16b, v19.16b, v19.16b   \n"
-                        "eor    v20.16b, v20.16b, v20.16b   \n"
-                        "eor    v21.16b, v21.16b, v21.16b   \n"
-                        "eor    v22.16b, v22.16b, v22.16b   \n"
-                        "eor    v23.16b, v23.16b, v23.16b   \n"
-                        "eor    v24.16b, v24.16b, v24.16b   \n"
-                        "eor    v25.16b, v25.16b, v25.16b   \n"
-                        "eor    v26.16b, v26.16b, v26.16b   \n"
-                        "eor    v27.16b, v27.16b, v27.16b   \n"
-                        "eor    v28.16b, v28.16b, v28.16b   \n"
-                        "eor    v29.16b, v29.16b, v29.16b   \n"
-                        "eor    v30.16b, v30.16b, v30.16b   \n"
-                        "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                        "1:                                 \n"
-                        "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                        "cmp    w4, #0                      \n"
-                        "beq    3f                          \n"
-
-                        "2:                                 \n"
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%1], #64 \n"
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-                        "fmla   v16.4s, v8.4s, v0.s[0]      \n"
-                        "fmla   v18.4s, v8.4s, v0.s[1]      \n"
-                        "fmla   v20.4s, v8.4s, v0.s[2]      \n"
-                        "fmla   v22.4s, v8.4s, v0.s[3]      \n"
-                        "fmla   v24.4s, v8.4s, v1.s[0]      \n"
-                        "fmla   v26.4s, v8.4s, v1.s[1]      \n"
-                        "fmla   v28.4s, v8.4s, v1.s[2]      \n"
-                        "fmla   v30.4s, v8.4s, v1.s[3]      \n"
-                        "fmla   v17.4s, v9.4s, v0.s[0]      \n"
-                        "fmla   v19.4s, v9.4s, v0.s[1]      \n"
-                        "fmla   v21.4s, v9.4s, v0.s[2]      \n"
-                        "fmla   v23.4s, v9.4s, v0.s[3]      \n"
-                        "fmla   v25.4s, v9.4s, v1.s[0]      \n"
-                        "fmla   v27.4s, v9.4s, v1.s[1]      \n"
-                        "fmla   v29.4s, v9.4s, v1.s[2]      \n"
-                        "fmla   v31.4s, v9.4s, v1.s[3]      \n"
-                        "fmla   v16.4s, v10.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v10.4s, v2.s[1]     \n"
-                        "fmla   v20.4s, v10.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v10.4s, v2.s[3]     \n"
-                        "fmla   v24.4s, v10.4s, v3.s[0]     \n"
-                        "fmla   v26.4s, v10.4s, v3.s[1]     \n"
-                        "fmla   v28.4s, v10.4s, v3.s[2]     \n"
-                        "fmla   v30.4s, v10.4s, v3.s[3]     \n"
-                        "fmla   v17.4s, v11.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v11.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v11.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v11.4s, v2.s[3]     \n"
-                        "fmla   v25.4s, v11.4s, v3.s[0]     \n"
-                        "fmla   v27.4s, v11.4s, v3.s[1]     \n"
-                        "fmla   v29.4s, v11.4s, v3.s[2]     \n"
-                        "fmla   v31.4s, v11.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%1], #64 \n"
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%2], #64 \n"
-                        "fmla   v16.4s, v12.4s, v4.s[0]     \n"
-                        "fmla   v18.4s, v12.4s, v4.s[1]     \n"
-                        "fmla   v20.4s, v12.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v12.4s, v4.s[3]     \n"
-                        "fmla   v24.4s, v12.4s, v5.s[0]     \n"
-                        "fmla   v26.4s, v12.4s, v5.s[1]     \n"
-                        "fmla   v28.4s, v12.4s, v5.s[2]     \n"
-                        "fmla   v30.4s, v12.4s, v5.s[3]     \n"
-                        "fmla   v17.4s, v13.4s, v4.s[0]     \n"
-                        "fmla   v19.4s, v13.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v13.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v13.4s, v4.s[3]     \n"
-                        "fmla   v25.4s, v13.4s, v5.s[0]     \n"
-                        "fmla   v27.4s, v13.4s, v5.s[1]     \n"
-                        "fmla   v29.4s, v13.4s, v5.s[2]     \n"
-                        "fmla   v31.4s, v13.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v14.4s, v6.s[0]     \n"
-                        "fmla   v18.4s, v14.4s, v6.s[1]     \n"
-                        "fmla   v20.4s, v14.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v14.4s, v6.s[3]     \n"
-                        "fmla   v24.4s, v14.4s, v7.s[0]     \n"
-                        "fmla   v26.4s, v14.4s, v7.s[1]     \n"
-                        "fmla   v28.4s, v14.4s, v7.s[2]     \n"
-                        "fmla   v30.4s, v14.4s, v7.s[3]     \n"
-                        "subs   w4, w4, #1                  \n"
-                        "fmla   v17.4s, v15.4s, v6.s[0]     \n"
-                        "fmla   v19.4s, v15.4s, v6.s[1]     \n"
-                        "fmla   v21.4s, v15.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v15.4s, v6.s[3]     \n"
-                        "fmla   v25.4s, v15.4s, v7.s[0]     \n"
-                        "fmla   v27.4s, v15.4s, v7.s[1]     \n"
-                        "fmla   v29.4s, v15.4s, v7.s[2]     \n"
-                        "fmla   v31.4s, v15.4s, v7.s[3]     \n"
-                        "bne    2b                          \n"
-
-                        "3:                                 \n"
-                        "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                        "cmp    w4, #0                      \n"
-                        "beq    5f                          \n"
-
-                        "4:                                 \n"
-                        "ld1    {v0.4s, v1.4s}, [%2], #32   \n"
-                        "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-                        "fmla   v16.4s, v4.4s, v0.s[0]      \n"
-                        "fmla   v18.4s, v4.4s, v0.s[1]      \n"
-                        "fmla   v20.4s, v4.4s, v0.s[2]      \n"
-                        "fmla   v22.4s, v4.4s, v0.s[3]      \n"
-                        "fmla   v17.4s, v5.4s, v0.s[0]      \n"
-                        "fmla   v19.4s, v5.4s, v0.s[1]      \n"
-                        "fmla   v21.4s, v5.4s, v0.s[2]      \n"
-                        "fmla   v23.4s, v5.4s, v0.s[3]      \n"
-                        "subs   w4, w4, #1                  \n"
-                        "fmla   v24.4s, v4.4s, v1.s[0]      \n"
-                        "fmla   v26.4s, v4.4s, v1.s[1]      \n"
-                        "fmla   v28.4s, v4.4s, v1.s[2]      \n"
-                        "fmla   v30.4s, v4.4s, v1.s[3]      \n"
-                        "fmla   v25.4s, v5.4s, v1.s[0]      \n"
-                        "fmla   v27.4s, v5.4s, v1.s[1]      \n"
-                        "fmla   v29.4s, v5.4s, v1.s[2]      \n"
-                        "fmla   v31.4s, v5.4s, v1.s[3]      \n"
-                        "bne    4b                          \n"
-
-                        "5:                                 \n"
-                        "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                        "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                        "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                        "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr), // %0
-                        "=r"(pA),     // %1
-                        "=r"(pB)      // %2
-                        : "0"(outptr),
-                        "1"(pA),
-                        "2"(pB),
-                        "r"(max_kk), // %6
-                        "r"(k)       // %7
-                        : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4_t _sum00;
-                float32x4_t _sum01;
-                float32x4_t _sum10;
-                float32x4_t _sum11;
-                float32x4_t _sum20;
-                float32x4_t _sum21;
-                float32x4_t _sum30;
-                float32x4_t _sum31;
-                float32x4_t _sum40;
-                float32x4_t _sum41;
-                float32x4_t _sum50;
-                float32x4_t _sum51;
-                float32x4_t _sum60;
-                float32x4_t _sum61;
-                float32x4_t _sum70;
-                float32x4_t _sum71;
-
-                if (k == 0)
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                    _sum20 = vdupq_n_f32(0.f);
-                    _sum21 = vdupq_n_f32(0.f);
-                    _sum30 = vdupq_n_f32(0.f);
-                    _sum31 = vdupq_n_f32(0.f);
-                    _sum40 = vdupq_n_f32(0.f);
-                    _sum41 = vdupq_n_f32(0.f);
-                    _sum50 = vdupq_n_f32(0.f);
-                    _sum51 = vdupq_n_f32(0.f);
-                    _sum60 = vdupq_n_f32(0.f);
-                    _sum61 = vdupq_n_f32(0.f);
-                    _sum70 = vdupq_n_f32(0.f);
-                    _sum71 = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    _sum00 = vld1q_f32(outptr);
-                    _sum01 = vld1q_f32(outptr + 4 * 1);
-                    _sum10 = vld1q_f32(outptr + 4 * 2);
-                    _sum11 = vld1q_f32(outptr + 4 * 3);
-                    _sum20 = vld1q_f32(outptr + 4 * 4);
-                    _sum21 = vld1q_f32(outptr + 4 * 5);
-                    _sum30 = vld1q_f32(outptr + 4 * 6);
-                    _sum31 = vld1q_f32(outptr + 4 * 7);
-                    _sum40 = vld1q_f32(outptr + 4 * 8);
-                    _sum41 = vld1q_f32(outptr + 4 * 9);
-                    _sum50 = vld1q_f32(outptr + 4 * 10);
-                    _sum51 = vld1q_f32(outptr + 4 * 11);
-                    _sum60 = vld1q_f32(outptr + 4 * 12);
-                    _sum61 = vld1q_f32(outptr + 4 * 13);
-                    _sum70 = vld1q_f32(outptr + 4 * 14);
-                    _sum71 = vld1q_f32(outptr + 4 * 15);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pA0 = vld1q_f32(pA);
-                    float32x4_t _pA1 = vld1q_f32(pA + 4);
-
-                    float32x4_t _pB0 = vld1q_f32(pB);
-                    float32x4_t _pB1 = vld1q_f32(pB + 4);
-
-                    _sum00 = vfmaq_laneq_f32(_sum00, _pA0, _pB0, 0);
-                    _sum01 = vfmaq_laneq_f32(_sum01, _pA1, _pB0, 0);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _pA0, _pB0, 1);
-                    _sum11 = vfmaq_laneq_f32(_sum11, _pA1, _pB0, 1);
-                    _sum20 = vfmaq_laneq_f32(_sum20, _pA0, _pB0, 2);
-                    _sum21 = vfmaq_laneq_f32(_sum21, _pA1, _pB0, 2);
-                    _sum30 = vfmaq_laneq_f32(_sum30, _pA0, _pB0, 3);
-                    _sum31 = vfmaq_laneq_f32(_sum31, _pA1, _pB0, 3);
-                    _sum40 = vfmaq_laneq_f32(_sum40, _pA0, _pB1, 0);
-                    _sum41 = vfmaq_laneq_f32(_sum41, _pA1, _pB1, 0);
-                    _sum50 = vfmaq_laneq_f32(_sum50, _pA0, _pB1, 1);
-                    _sum51 = vfmaq_laneq_f32(_sum51, _pA1, _pB1, 1);
-                    _sum60 = vfmaq_laneq_f32(_sum60, _pA0, _pB1, 2);
-                    _sum61 = vfmaq_laneq_f32(_sum61, _pA1, _pB1, 2);
-                    _sum70 = vfmaq_laneq_f32(_sum70, _pA0, _pB1, 3);
-                    _sum71 = vfmaq_laneq_f32(_sum71, _pA1, _pB1, 3);
-
-                    pA += 8;
-                    pB += 8;
-                }
-
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                vst1q_f32(outptr + 4 * 4, _sum20);
-                vst1q_f32(outptr + 4 * 5, _sum21);
-                vst1q_f32(outptr + 4 * 6, _sum30);
-                vst1q_f32(outptr + 4 * 7, _sum31);
-                vst1q_f32(outptr + 4 * 8, _sum40);
-                vst1q_f32(outptr + 4 * 9, _sum41);
-                vst1q_f32(outptr + 4 * 10, _sum50);
-                vst1q_f32(outptr + 4 * 11, _sum51);
-                vst1q_f32(outptr + 4 * 12, _sum60);
-                vst1q_f32(outptr + 4 * 13, _sum61);
-                vst1q_f32(outptr + 4 * 14, _sum70);
-                vst1q_f32(outptr + 4 * 15, _sum71);
-                outptr += 8 * 8;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "cbz    %w7, 0f                     \n"
-
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #64                 \n"
-                    "b      1f                          \n"
-
-                    "0:                                 \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-                    "eor    v25.16b, v25.16b, v25.16b   \n"
-                    "eor    v26.16b, v26.16b, v26.16b   \n"
-                    "eor    v27.16b, v27.16b, v27.16b   \n"
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-                    "eor    v29.16b, v29.16b, v29.16b   \n"
-                    "eor    v30.16b, v30.16b, v30.16b   \n"
-                    "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                    "1:                                 \n"
-                    "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    3f                          \n"
-
-                    "2:                                 \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-                    "fmla   v24.4s, v4.4s, v0.s[0]      \n"
-                    "fmla   v26.4s, v4.4s, v0.s[1]      \n"
-                    "fmla   v28.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v30.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v25.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v27.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v29.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-                    "fmla   v24.4s, v6.4s, v1.s[0]      \n"
-                    "fmla   v26.4s, v6.4s, v1.s[1]      \n"
-                    "fmla   v28.4s, v6.4s, v1.s[2]      \n"
-                    "fmla   v30.4s, v6.4s, v1.s[3]      \n"
-                    "fmla   v25.4s, v7.4s, v1.s[0]      \n"
-                    "fmla   v27.4s, v7.4s, v1.s[1]      \n"
-                    "fmla   v29.4s, v7.4s, v1.s[2]      \n"
-                    "fmla   v31.4s, v7.4s, v1.s[3]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%1], #64 \n"
-                    "fmla   v24.4s, v8.4s, v2.s[0]      \n"
-                    "fmla   v26.4s, v8.4s, v2.s[1]      \n"
-                    "fmla   v28.4s, v8.4s, v2.s[2]      \n"
-                    "fmla   v30.4s, v8.4s, v2.s[3]      \n"
-                    "fmla   v25.4s, v9.4s, v2.s[0]      \n"
-                    "fmla   v27.4s, v9.4s, v2.s[1]      \n"
-                    "fmla   v29.4s, v9.4s, v2.s[2]      \n"
-                    "fmla   v31.4s, v9.4s, v2.s[3]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v24.4s, v10.4s, v3.s[0]     \n"
-                    "fmla   v26.4s, v10.4s, v3.s[1]     \n"
-                    "fmla   v28.4s, v10.4s, v3.s[2]     \n"
-                    "fmla   v30.4s, v10.4s, v3.s[3]     \n"
-                    "fmla   v25.4s, v11.4s, v3.s[0]     \n"
-                    "fmla   v27.4s, v11.4s, v3.s[1]     \n"
-                    "fmla   v29.4s, v11.4s, v3.s[2]     \n"
-                    "fmla   v31.4s, v11.4s, v3.s[3]     \n"
-                    "bne    2b                          \n"
-
-                    "3:                                 \n"
-                    "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "ld1    {v0.4s}, [%2], #16          \n"
-                    "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-                    "fmla   v24.4s, v4.4s, v0.s[0]      \n"
-                    "fmla   v26.4s, v4.4s, v0.s[1]      \n"
-                    "fmla   v28.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v30.4s, v4.4s, v0.s[3]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v25.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v27.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v29.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4_t _sum00;
-                float32x4_t _sum01;
-                float32x4_t _sum10;
-                float32x4_t _sum11;
-                float32x4_t _sum20;
-                float32x4_t _sum21;
-                float32x4_t _sum30;
-                float32x4_t _sum31;
-
-                if (k == 0)
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                    _sum20 = vdupq_n_f32(0.f);
-                    _sum21 = vdupq_n_f32(0.f);
-                    _sum30 = vdupq_n_f32(0.f);
-                    _sum31 = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    _sum00 = vld1q_f32(outptr);
-                    _sum01 = vld1q_f32(outptr + 4 * 1);
-                    _sum10 = vld1q_f32(outptr + 4 * 2);
-                    _sum11 = vld1q_f32(outptr + 4 * 3);
-                    _sum20 = vld1q_f32(outptr + 4 * 4);
-                    _sum21 = vld1q_f32(outptr + 4 * 5);
-                    _sum30 = vld1q_f32(outptr + 4 * 6);
-                    _sum31 = vld1q_f32(outptr + 4 * 7);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pA0 = vld1q_f32(pA);
-                    float32x4_t _pA1 = vld1q_f32(pA + 4);
-
-                    float32x4_t _pB0 = vld1q_f32(pB);
-
-                    _sum00 = vfmaq_laneq_f32(_sum00, _pA0, _pB0, 0);
-                    _sum01 = vfmaq_laneq_f32(_sum01, _pA1, _pB0, 0);
-                    _sum10 = vfmaq_laneq_f32(_sum10, _pA0, _pB0, 1);
-                    _sum11 = vfmaq_laneq_f32(_sum11, _pA1, _pB0, 1);
-                    _sum20 = vfmaq_laneq_f32(_sum20, _pA0, _pB0, 2);
-                    _sum21 = vfmaq_laneq_f32(_sum21, _pA1, _pB0, 2);
-                    _sum30 = vfmaq_laneq_f32(_sum30, _pA0, _pB0, 3);
-                    _sum31 = vfmaq_laneq_f32(_sum31, _pA1, _pB0, 3);
-
-                    pA += 8;
-                    pB += 4;
-                }
-
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                vst1q_f32(outptr + 4 * 4, _sum20);
-                vst1q_f32(outptr + 4 * 5, _sum21);
-                vst1q_f32(outptr + 4 * 6, _sum30);
-                vst1q_f32(outptr + 4 * 7, _sum31);
-                outptr += 8 * 4;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "cbz    %w7, 0f                     \n"
-
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "b      1f                          \n"
-
-                    "0:                                 \n"
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-                    "eor    v29.16b, v29.16b, v29.16b   \n"
-                    "eor    v30.16b, v30.16b, v30.16b   \n"
-                    "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                    "1:                                 \n"
-                    "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    3f                          \n"
-
-                    "2:                                 \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n"
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v0.4s, v1.4s}, [%2], #32   \n"
-                    "fmla   v28.4s, v4.4s, v0.s[0]      \n"
-                    "fmla   v30.4s, v4.4s, v0.s[1]      \n"
-                    "fmla   v29.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v31.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v28.4s, v6.4s, v0.s[2]      \n"
-                    "fmla   v30.4s, v6.4s, v0.s[3]      \n"
-                    "fmla   v29.4s, v7.4s, v0.s[2]      \n"
-                    "fmla   v31.4s, v7.4s, v0.s[3]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%1], #64 \n"
-                    "fmla   v28.4s, v8.4s, v1.s[0]      \n"
-                    "fmla   v30.4s, v8.4s, v1.s[1]      \n"
-                    "fmla   v29.4s, v9.4s, v1.s[0]      \n"
-                    "fmla   v31.4s, v9.4s, v1.s[1]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v28.4s, v10.4s, v1.s[2]     \n"
-                    "fmla   v30.4s, v10.4s, v1.s[3]     \n"
-                    "fmla   v29.4s, v11.4s, v1.s[2]     \n"
-                    "fmla   v31.4s, v11.4s, v1.s[3]     \n"
-                    "bne    2b                          \n"
-
-                    "3:                                 \n"
-                    "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "ld1    {v0.2s}, [%2], #8           \n"
-                    "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-                    "fmla   v28.4s, v4.4s, v0.s[0]      \n"
-                    "fmla   v30.4s, v4.4s, v0.s[1]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v29.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v31.4s, v5.4s, v0.s[1]      \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "x4", "v0", "v1", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4_t _sum00;
-                float32x4_t _sum01;
-                float32x4_t _sum10;
-                float32x4_t _sum11;
-
-                if (k == 0)
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    _sum00 = vld1q_f32(outptr);
-                    _sum01 = vld1q_f32(outptr + 4 * 1);
-                    _sum10 = vld1q_f32(outptr + 4 * 2);
-                    _sum11 = vld1q_f32(outptr + 4 * 3);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pA0 = vld1q_f32(pA);
-                    float32x4_t _pA1 = vld1q_f32(pA + 4);
-
-                    float32x2_t _pB0 = vld1_f32(pB);
-
-                    _sum00 = vfmaq_lane_f32(_sum00, _pA0, _pB0, 0);
-                    _sum01 = vfmaq_lane_f32(_sum01, _pA1, _pB0, 0);
-                    _sum10 = vfmaq_lane_f32(_sum10, _pA0, _pB0, 1);
-                    _sum11 = vfmaq_lane_f32(_sum11, _pA1, _pB0, 1);
-
-                    pA += 8;
-                    pB += 2;
-                }
-
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                outptr += 8 * 2;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "cbz    %w7, 0f                     \n"
-
-                    "ld1    {v30.4s, v31.4s}, [%0]      \n"
-                    "b      1f                          \n"
-
-                    "0:                                 \n"
-                    "eor    v30.16b, v30.16b, v30.16b   \n"
-                    "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                    "1:                                 \n"
-                    "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    3f                          \n"
-
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-                    "eor    v29.16b, v29.16b, v29.16b   \n"
-                    "2:                                 \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n"
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "ld1    {v0.4s}, [%2], #16          \n"
-                    "fmla   v28.4s, v4.4s, v0.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v30.4s, v6.4s, v0.s[1]      \n"
-                    "fmla   v31.4s, v7.4s, v0.s[1]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%1], #64 \n"
-                    "fmla   v28.4s, v8.4s, v0.s[2]      \n"
-                    "fmla   v29.4s, v9.4s, v0.s[2]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v30.4s, v10.4s, v0.s[3]     \n"
-                    "fmla   v31.4s, v11.4s, v0.s[3]     \n"
-                    "bne    2b                          \n"
-                    "fadd   v30.4s, v30.4s, v28.4s      \n"
-                    "fadd   v31.4s, v31.4s, v29.4s      \n"
-
-                    "3:                                 \n"
-                    "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "ld1r   {v0.4s}, [%2], #4           \n"
-                    "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-                    "fmla   v30.4s, v4.4s, v0.4s        \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v31.4s, v5.4s, v0.4s        \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "st1    {v30.4s, v31.4s}, [%0], #32 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "x4", "v0", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4_t _sum00;
-                float32x4_t _sum01;
-
-                if (k == 0)
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    _sum00 = vld1q_f32(outptr);
-                    _sum01 = vld1q_f32(outptr + 4);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pA0 = vld1q_f32(pA);
-                    float32x4_t _pA1 = vld1q_f32(pA + 4);
-
-                    float32x4_t _pB = vld1q_dup_f32(pB);
-
-                    _sum00 = vfmaq_f32(_sum00, _pA0, _pB);
-                    _sum01 = vfmaq_f32(_sum01, _pA1, _pB);
-
-                    pA += 8;
-                    pB += 1;
-                }
-
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                outptr += 8;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const float* pAT = AT_tile.row(b) + max_kk * ii;
-            const float* pB = BT_tile.row(b);
-
-            int jj = 0;
-#if __aarch64__
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "cbz    %w7, 0f                     \n"
-
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #128                \n"
-                    "b      1f                          \n"
-
-                    "0:                                 \n"
-                    "eor    v20.16b, v20.16b, v20.16b   \n"
-                    "eor    v21.16b, v21.16b, v21.16b   \n"
-                    "eor    v22.16b, v22.16b, v22.16b   \n"
-                    "eor    v23.16b, v23.16b, v23.16b   \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-                    "eor    v25.16b, v25.16b, v25.16b   \n"
-                    "eor    v26.16b, v26.16b, v26.16b   \n"
-                    "eor    v27.16b, v27.16b, v27.16b   \n"
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-                    "eor    v29.16b, v29.16b, v29.16b   \n"
-                    "eor    v30.16b, v30.16b, v30.16b   \n"
-                    "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                    "1:                                 \n"
-                    "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    3f                          \n"
-
-                    "2:                                 \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%1], #64 \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-                    "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                    "fmla   v21.4s, v16.4s, v0.s[1]     \n"
-                    "fmla   v22.4s, v16.4s, v0.s[2]     \n"
-                    "fmla   v23.4s, v16.4s, v0.s[3]     \n"
-                    "fmla   v24.4s, v16.4s, v1.s[0]     \n"
-                    "fmla   v25.4s, v16.4s, v1.s[1]     \n"
-                    "fmla   v26.4s, v16.4s, v1.s[2]     \n"
-                    "fmla   v27.4s, v16.4s, v1.s[3]     \n"
-                    "fmla   v28.4s, v16.4s, v2.s[0]     \n"
-                    "fmla   v29.4s, v16.4s, v2.s[1]     \n"
-                    "fmla   v30.4s, v16.4s, v2.s[2]     \n"
-                    "fmla   v31.4s, v16.4s, v2.s[3]     \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%2], #64 \n"
-                    "fmla   v20.4s, v17.4s, v3.s[0]     \n"
-                    "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                    "fmla   v22.4s, v17.4s, v3.s[2]     \n"
-                    "fmla   v23.4s, v17.4s, v3.s[3]     \n"
-                    "fmla   v24.4s, v17.4s, v4.s[0]     \n"
-                    "fmla   v25.4s, v17.4s, v4.s[1]     \n"
-                    "fmla   v26.4s, v17.4s, v4.s[2]     \n"
-                    "fmla   v27.4s, v17.4s, v4.s[3]     \n"
-                    "fmla   v28.4s, v17.4s, v5.s[0]     \n"
-                    "fmla   v29.4s, v17.4s, v5.s[1]     \n"
-                    "fmla   v30.4s, v17.4s, v5.s[2]     \n"
-                    "fmla   v31.4s, v17.4s, v5.s[3]     \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-                    "fmla   v20.4s, v18.4s, v6.s[0]     \n"
-                    "fmla   v21.4s, v18.4s, v6.s[1]     \n"
-                    "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                    "fmla   v23.4s, v18.4s, v6.s[3]     \n"
-                    "fmla   v24.4s, v18.4s, v7.s[0]     \n"
-                    "fmla   v25.4s, v18.4s, v7.s[1]     \n"
-                    "fmla   v26.4s, v18.4s, v7.s[2]     \n"
-                    "fmla   v27.4s, v18.4s, v7.s[3]     \n"
-                    "fmla   v28.4s, v18.4s, v0.s[0]     \n"
-                    "fmla   v29.4s, v18.4s, v0.s[1]     \n"
-                    "fmla   v30.4s, v18.4s, v0.s[2]     \n"
-                    "fmla   v31.4s, v18.4s, v0.s[3]     \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v20.4s, v19.4s, v1.s[0]     \n"
-                    "fmla   v21.4s, v19.4s, v1.s[1]     \n"
-                    "fmla   v22.4s, v19.4s, v1.s[2]     \n"
-                    "fmla   v23.4s, v19.4s, v1.s[3]     \n"
-                    "fmla   v24.4s, v19.4s, v2.s[0]     \n"
-                    "fmla   v25.4s, v19.4s, v2.s[1]     \n"
-                    "fmla   v26.4s, v19.4s, v2.s[2]     \n"
-                    "fmla   v27.4s, v19.4s, v2.s[3]     \n"
-                    "fmla   v28.4s, v19.4s, v3.s[0]     \n"
-                    "fmla   v29.4s, v19.4s, v3.s[1]     \n"
-                    "fmla   v30.4s, v19.4s, v3.s[2]     \n"
-                    "fmla   v31.4s, v19.4s, v3.s[3]     \n"
-                    "bne    2b                          \n"
-
-                    "3:                                 \n"
-                    "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s}, [%2], #48 \n"
-                    "ld1    {v16.4s}, [%1], #16         \n"
-                    "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                    "fmla   v21.4s, v16.4s, v0.s[1]     \n"
-                    "fmla   v22.4s, v16.4s, v0.s[2]     \n"
-                    "fmla   v23.4s, v16.4s, v0.s[3]     \n"
-                    "fmla   v24.4s, v16.4s, v1.s[0]     \n"
-                    "fmla   v25.4s, v16.4s, v1.s[1]     \n"
-                    "fmla   v26.4s, v16.4s, v1.s[2]     \n"
-                    "fmla   v27.4s, v16.4s, v1.s[3]     \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v28.4s, v16.4s, v2.s[0]     \n"
-                    "fmla   v29.4s, v16.4s, v2.s[1]     \n"
-                    "fmla   v30.4s, v16.4s, v2.s[2]     \n"
-                    "fmla   v31.4s, v16.4s, v2.s[3]     \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4_t _sum0;
-                float32x4_t _sum1;
-                float32x4_t _sum2;
-                float32x4_t _sum3;
-                float32x4_t _sum4;
-                float32x4_t _sum5;
-                float32x4_t _sum6;
-                float32x4_t _sum7;
-                float32x4_t _sum8;
-                float32x4_t _sum9;
-                float32x4_t _suma;
-                float32x4_t _sumb;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                    _sum2 = vdupq_n_f32(0.f);
-                    _sum3 = vdupq_n_f32(0.f);
-                    _sum4 = vdupq_n_f32(0.f);
-                    _sum5 = vdupq_n_f32(0.f);
-                    _sum6 = vdupq_n_f32(0.f);
-                    _sum7 = vdupq_n_f32(0.f);
-                    _sum8 = vdupq_n_f32(0.f);
-                    _sum9 = vdupq_n_f32(0.f);
-                    _suma = vdupq_n_f32(0.f);
-                    _sumb = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    _sum0 = vld1q_f32(outptr);
-                    _sum1 = vld1q_f32(outptr + 4);
-                    _sum2 = vld1q_f32(outptr + 8);
-                    _sum3 = vld1q_f32(outptr + 12);
-                    _sum4 = vld1q_f32(outptr + 16);
-                    _sum5 = vld1q_f32(outptr + 20);
-                    _sum6 = vld1q_f32(outptr + 24);
-                    _sum7 = vld1q_f32(outptr + 28);
-                    _sum8 = vld1q_f32(outptr + 32);
-                    _sum9 = vld1q_f32(outptr + 36);
-                    _suma = vld1q_f32(outptr + 40);
-                    _sumb = vld1q_f32(outptr + 44);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pA = vld1q_f32(pA);
-                    float32x4_t _pB0 = vld1q_f32(pB);
-                    float32x4_t _pB1 = vld1q_f32(pB + 4);
-                    float32x4_t _pB2 = vld1q_f32(pB + 8);
-
-                    _sum0 = vfmaq_laneq_f32(_sum0, _pA, _pB0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _pA, _pB0, 1);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _pA, _pB0, 2);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _pA, _pB0, 3);
-                    _sum4 = vfmaq_laneq_f32(_sum4, _pA, _pB1, 0);
-                    _sum5 = vfmaq_laneq_f32(_sum5, _pA, _pB1, 1);
-                    _sum6 = vfmaq_laneq_f32(_sum6, _pA, _pB1, 2);
-                    _sum7 = vfmaq_laneq_f32(_sum7, _pA, _pB1, 3);
-                    _sum8 = vfmaq_laneq_f32(_sum8, _pA, _pB2, 0);
-                    _sum9 = vfmaq_laneq_f32(_sum9, _pA, _pB2, 1);
-                    _suma = vfmaq_laneq_f32(_suma, _pA, _pB2, 2);
-                    _sumb = vfmaq_laneq_f32(_sumb, _pA, _pB2, 3);
-
-                    pA += 4;
-                    pB += 12;
-                }
-
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 4 * 2, _sum2);
-                vst1q_f32(outptr + 4 * 3, _sum3);
-                vst1q_f32(outptr + 4 * 4, _sum4);
-                vst1q_f32(outptr + 4 * 5, _sum5);
-                vst1q_f32(outptr + 4 * 6, _sum6);
-                vst1q_f32(outptr + 4 * 7, _sum7);
-                vst1q_f32(outptr + 4 * 8, _sum8);
-                vst1q_f32(outptr + 4 * 9, _sum9);
-                vst1q_f32(outptr + 4 * 10, _suma);
-                vst1q_f32(outptr + 4 * 11, _sumb);
-                outptr += 4 * 12;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-#endif // __aarch64__
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "cbz    %w7, 0f                     \n"
-
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #64                 \n"
-                    "b      1f                          \n"
-
-                    "0:                                 \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-                    "eor    v25.16b, v25.16b, v25.16b   \n"
-                    "eor    v26.16b, v26.16b, v26.16b   \n"
-                    "eor    v27.16b, v27.16b, v27.16b   \n"
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-                    "eor    v29.16b, v29.16b, v29.16b   \n"
-                    "eor    v30.16b, v30.16b, v30.16b   \n"
-                    "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                    "1:                                 \n"
-                    "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    3f                          \n"
-
-                    "2:                                 \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%1], #64 \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-                    "fmla   v24.4s, v16.4s, v0.s[0]     \n"
-                    "fmla   v25.4s, v16.4s, v0.s[1]     \n"
-                    "fmla   v26.4s, v16.4s, v0.s[2]     \n"
-                    "fmla   v27.4s, v16.4s, v0.s[3]     \n"
-                    "fmla   v28.4s, v16.4s, v1.s[0]     \n"
-                    "fmla   v29.4s, v16.4s, v1.s[1]     \n"
-                    "fmla   v30.4s, v16.4s, v1.s[2]     \n"
-                    "fmla   v31.4s, v16.4s, v1.s[3]     \n"
-                    "fmla   v24.4s, v17.4s, v2.s[0]     \n"
-                    "fmla   v25.4s, v17.4s, v2.s[1]     \n"
-                    "fmla   v26.4s, v17.4s, v2.s[2]     \n"
-                    "fmla   v27.4s, v17.4s, v2.s[3]     \n"
-                    "fmla   v28.4s, v17.4s, v3.s[0]     \n"
-                    "fmla   v29.4s, v17.4s, v3.s[1]     \n"
-                    "fmla   v30.4s, v17.4s, v3.s[2]     \n"
-                    "fmla   v31.4s, v17.4s, v3.s[3]     \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%2], #64 \n"
-                    "fmla   v24.4s, v18.4s, v4.s[0]     \n"
-                    "fmla   v25.4s, v18.4s, v4.s[1]     \n"
-                    "fmla   v26.4s, v18.4s, v4.s[2]     \n"
-                    "fmla   v27.4s, v18.4s, v4.s[3]     \n"
-                    "fmla   v28.4s, v18.4s, v5.s[0]     \n"
-                    "fmla   v29.4s, v18.4s, v5.s[1]     \n"
-                    "fmla   v30.4s, v18.4s, v5.s[2]     \n"
-                    "fmla   v31.4s, v18.4s, v5.s[3]     \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v24.4s, v19.4s, v6.s[0]     \n"
-                    "fmla   v25.4s, v19.4s, v6.s[1]     \n"
-                    "fmla   v26.4s, v19.4s, v6.s[2]     \n"
-                    "fmla   v27.4s, v19.4s, v6.s[3]     \n"
-                    "fmla   v28.4s, v19.4s, v7.s[0]     \n"
-                    "fmla   v29.4s, v19.4s, v7.s[1]     \n"
-                    "fmla   v30.4s, v19.4s, v7.s[2]     \n"
-                    "fmla   v31.4s, v19.4s, v7.s[3]     \n"
-                    "bne    2b                          \n"
-
-                    "3:                                 \n"
-                    "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "ld1    {v0.4s, v1.4s}, [%2], #32   \n"
-                    "ld1    {v16.4s}, [%1], #16         \n"
-                    "fmla   v24.4s, v16.4s, v0.s[0]     \n"
-                    "fmla   v25.4s, v16.4s, v0.s[1]     \n"
-                    "fmla   v26.4s, v16.4s, v0.s[2]     \n"
-                    "fmla   v27.4s, v16.4s, v0.s[3]     \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v28.4s, v16.4s, v1.s[0]     \n"
-                    "fmla   v29.4s, v16.4s, v1.s[1]     \n"
-                    "fmla   v30.4s, v16.4s, v1.s[2]     \n"
-                    "fmla   v31.4s, v16.4s, v1.s[3]     \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "cmp        %7, #0              \n"
-                    "beq        0f                  \n"
-
-                    "vldm       %0!, {d16-d23}      \n"
-                    "vldm       %0, {d24-d31}       \n"
-                    "sub        %0, %0, #64         \n"
-                    "b          1f                  \n"
-
-                    "0:                             \n"
-                    "veor       q8, q8              \n"
-                    "veor       q9, q9              \n"
-                    "veor       q10, q10            \n"
-                    "veor       q11, q11            \n"
-                    "veor       q12, q12            \n"
-                    "veor       q13, q13            \n"
-                    "veor       q14, q14            \n"
-                    "veor       q15, q15            \n"
-
-                    "1:                             \n"
-                    "lsr        r4, %6, #2          \n" // r4 = max_kk >> 2
-                    "cmp        r4, #0              \n"
-                    "beq        3f                  \n"
-
-                    "2:                             \n"
-                    "pld        [%1, #512]          \n"
-                    "vldm       %1!, {d8-d15}       \n"
-                    "pld        [%2, #512]          \n"
-                    "vldm       %2!, {d0-d7}        \n"
-                    "vmla.f32   q8, q4, d0[0]       \n"
-                    "vmla.f32   q9, q4, d0[1]       \n"
-                    "vmla.f32   q10, q4, d1[0]      \n"
-                    "vmla.f32   q11, q4, d1[1]      \n"
-                    "vmla.f32   q12, q4, d2[0]      \n"
-                    "vmla.f32   q13, q4, d2[1]      \n"
-                    "vmla.f32   q14, q4, d3[0]      \n"
-                    "vmla.f32   q15, q4, d3[1]      \n"
-                    "vmla.f32   q8, q5, d4[0]       \n"
-                    "vmla.f32   q9, q5, d4[1]       \n"
-                    "vmla.f32   q10, q5, d5[0]      \n"
-                    "vmla.f32   q11, q5, d5[1]      \n"
-                    "vmla.f32   q12, q5, d6[0]      \n"
-                    "vmla.f32   q13, q5, d6[1]      \n"
-                    "vmla.f32   q14, q5, d7[0]      \n"
-                    "vmla.f32   q15, q5, d7[1]      \n"
-                    "pld        [%2, #512]          \n"
-                    "vldm       %2!, {d0-d7}        \n"
-                    "vmla.f32   q8, q6, d0[0]       \n"
-                    "vmla.f32   q9, q6, d0[1]       \n"
-                    "vmla.f32   q10, q6, d1[0]      \n"
-                    "vmla.f32   q11, q6, d1[1]      \n"
-                    "vmla.f32   q12, q6, d2[0]      \n"
-                    "vmla.f32   q13, q6, d2[1]      \n"
-                    "vmla.f32   q14, q6, d3[0]      \n"
-                    "vmla.f32   q15, q6, d3[1]      \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmla.f32   q8, q7, d4[0]       \n"
-                    "vmla.f32   q9, q7, d4[1]       \n"
-                    "vmla.f32   q10, q7, d5[0]      \n"
-                    "vmla.f32   q11, q7, d5[1]      \n"
-                    "vmla.f32   q12, q7, d6[0]      \n"
-                    "vmla.f32   q13, q7, d6[1]      \n"
-                    "vmla.f32   q14, q7, d7[0]      \n"
-                    "vmla.f32   q15, q7, d7[1]      \n"
-                    "bne        2b                  \n"
-
-                    "3:                             \n"
-                    "and        r4, %6, #3          \n" // r4 = remain = max_kk & 3
-                    "cmp        r4, #0              \n"
-                    "beq        5f                  \n"
-
-                    "4:                             \n"
-                    "vldm       %2!, {d0-d3}        \n"
-                    "vld1.f32   {d8-d9}, [%1 :128]! \n"
-                    "vmla.f32   q8, q4, d0[0]       \n"
-                    "vmla.f32   q9, q4, d0[1]       \n"
-                    "vmla.f32   q10, q4, d1[0]      \n"
-                    "vmla.f32   q11, q4, d1[1]      \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmla.f32   q12, q4, d2[0]      \n"
-                    "vmla.f32   q13, q4, d2[1]      \n"
-                    "vmla.f32   q14, q4, d3[0]      \n"
-                    "vmla.f32   q15, q4, d3[1]      \n"
-                    "bne        4b                  \n"
-
-                    "5:                             \n"
-                    "vstm       %0!, {d16-d23}      \n"
-                    "vstm       %0!, {d24-d31}      \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4_t _sum0;
-                float32x4_t _sum1;
-                float32x4_t _sum2;
-                float32x4_t _sum3;
-                float32x4_t _sum4;
-                float32x4_t _sum5;
-                float32x4_t _sum6;
-                float32x4_t _sum7;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                    _sum2 = vdupq_n_f32(0.f);
-                    _sum3 = vdupq_n_f32(0.f);
-                    _sum4 = vdupq_n_f32(0.f);
-                    _sum5 = vdupq_n_f32(0.f);
-                    _sum6 = vdupq_n_f32(0.f);
-                    _sum7 = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    _sum0 = vld1q_f32(outptr);
-                    _sum1 = vld1q_f32(outptr + 4);
-                    _sum2 = vld1q_f32(outptr + 8);
-                    _sum3 = vld1q_f32(outptr + 12);
-                    _sum4 = vld1q_f32(outptr + 16);
-                    _sum5 = vld1q_f32(outptr + 20);
-                    _sum6 = vld1q_f32(outptr + 24);
-                    _sum7 = vld1q_f32(outptr + 28);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pA = vld1q_f32(pA);
-                    float32x4_t _pB0 = vld1q_f32(pB);
-                    float32x4_t _pB1 = vld1q_f32(pB + 4);
-
-#if __aarch64__
-                    _sum0 = vfmaq_laneq_f32(_sum0, _pA, _pB0, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _pA, _pB0, 1);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _pA, _pB0, 2);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _pA, _pB0, 3);
-                    _sum4 = vfmaq_laneq_f32(_sum4, _pA, _pB1, 0);
-                    _sum5 = vfmaq_laneq_f32(_sum5, _pA, _pB1, 1);
-                    _sum6 = vfmaq_laneq_f32(_sum6, _pA, _pB1, 2);
-                    _sum7 = vfmaq_laneq_f32(_sum7, _pA, _pB1, 3);
-#else
-                    _sum0 = vmlaq_lane_f32(_sum0, _pA, vget_low_f32(_pB0), 0);
-                    _sum1 = vmlaq_lane_f32(_sum1, _pA, vget_low_f32(_pB0), 1);
-                    _sum2 = vmlaq_lane_f32(_sum2, _pA, vget_high_f32(_pB0), 0);
-                    _sum3 = vmlaq_lane_f32(_sum3, _pA, vget_high_f32(_pB0), 1);
-                    _sum4 = vmlaq_lane_f32(_sum4, _pA, vget_low_f32(_pB1), 0);
-                    _sum5 = vmlaq_lane_f32(_sum5, _pA, vget_low_f32(_pB1), 1);
-                    _sum6 = vmlaq_lane_f32(_sum6, _pA, vget_high_f32(_pB1), 0);
-                    _sum7 = vmlaq_lane_f32(_sum7, _pA, vget_high_f32(_pB1), 1);
-#endif
-
-                    pA += 4;
-                    pB += 8;
-                }
-
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 4 * 2, _sum2);
-                vst1q_f32(outptr + 4 * 3, _sum3);
-                vst1q_f32(outptr + 4 * 4, _sum4);
-                vst1q_f32(outptr + 4 * 5, _sum5);
-                vst1q_f32(outptr + 4 * 6, _sum6);
-                vst1q_f32(outptr + 4 * 7, _sum7);
-                outptr += 4 * 8;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "cbz    %w7, 0f                     \n"
-
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0] \n"
-                    "b      1f                          \n"
-
-                    "0:                                 \n"
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-                    "eor    v29.16b, v29.16b, v29.16b   \n"
-                    "eor    v30.16b, v30.16b, v30.16b   \n"
-                    "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                    "1:                                 \n"
-                    "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    3f                          \n"
-
-                    "2:                                 \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%1], #64 \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-                    "fmla   v28.4s, v16.4s, v0.s[0]     \n"
-                    "fmla   v29.4s, v16.4s, v0.s[1]     \n"
-                    "fmla   v30.4s, v16.4s, v0.s[2]     \n"
-                    "fmla   v31.4s, v16.4s, v0.s[3]     \n"
-                    "fmla   v28.4s, v17.4s, v1.s[0]     \n"
-                    "fmla   v29.4s, v17.4s, v1.s[1]     \n"
-                    "fmla   v30.4s, v17.4s, v1.s[2]     \n"
-                    "fmla   v31.4s, v17.4s, v1.s[3]     \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v28.4s, v18.4s, v2.s[0]     \n"
-                    "fmla   v29.4s, v18.4s, v2.s[1]     \n"
-                    "fmla   v30.4s, v18.4s, v2.s[2]     \n"
-                    "fmla   v31.4s, v18.4s, v2.s[3]     \n"
-                    "fmla   v28.4s, v19.4s, v3.s[0]     \n"
-                    "fmla   v29.4s, v19.4s, v3.s[1]     \n"
-                    "fmla   v30.4s, v19.4s, v3.s[2]     \n"
-                    "fmla   v31.4s, v19.4s, v3.s[3]     \n"
-                    "bne    2b                          \n"
-
-                    "3:                                 \n"
-                    "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "ld1    {v0.4s}, [%2], #16          \n"
-                    "ld1    {v16.4s}, [%1], #16         \n"
-                    "fmla   v28.4s, v16.4s, v0.s[0]     \n"
-                    "fmla   v29.4s, v16.4s, v0.s[1]     \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v30.4s, v16.4s, v0.s[2]     \n"
-                    "fmla   v31.4s, v16.4s, v0.s[3]     \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "cmp        %7, #0              \n"
-                    "beq        0f                  \n"
-
-                    "vldm       %0, {d24-d31}       \n"
-                    "b          1f                  \n"
-
-                    "0:                             \n"
-                    "veor       q12, q12            \n"
-                    "veor       q13, q13            \n"
-                    "veor       q14, q14            \n"
-                    "veor       q15, q15            \n"
-
-                    "1:                             \n"
-                    "lsr        r4, %6, #2          \n" // r4 = max_kk >> 2
-                    "cmp        r4, #0              \n"
-                    "beq        3f                  \n"
-
-                    "2:                             \n"
-                    "pld        [%1, #512]          \n"
-                    "vldm       %1!, {d8-d15}       \n"
-                    "pld        [%2, #512]          \n"
-                    "vldm       %2!, {d0-d7}        \n"
-                    "vmla.f32   q12, q4, d0[0]      \n"
-                    "vmla.f32   q13, q4, d0[1]      \n"
-                    "vmla.f32   q14, q4, d1[0]      \n"
-                    "vmla.f32   q15, q4, d1[1]      \n"
-                    "vmla.f32   q12, q5, d2[0]      \n"
-                    "vmla.f32   q13, q5, d2[1]      \n"
-                    "vmla.f32   q14, q5, d3[0]      \n"
-                    "vmla.f32   q15, q5, d3[1]      \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmla.f32   q12, q6, d4[0]      \n"
-                    "vmla.f32   q13, q6, d4[1]      \n"
-                    "vmla.f32   q14, q6, d5[0]      \n"
-                    "vmla.f32   q15, q6, d5[1]      \n"
-                    "vmla.f32   q12, q7, d6[0]      \n"
-                    "vmla.f32   q13, q7, d6[1]      \n"
-                    "vmla.f32   q14, q7, d7[0]      \n"
-                    "vmla.f32   q15, q7, d7[1]      \n"
-                    "bne        2b                  \n"
-
-                    "3:                             \n"
-                    "and        r4, %6, #3          \n" // r4 = remain = max_kk & 3
-                    "cmp        r4, #0              \n"
-                    "beq        5f                  \n"
-
-                    "4:                             \n"
-                    "vld1.f32   {d0-d1}, [%2 :128]! \n"
-                    "vld1.f32   {d8-d9}, [%1 :128]! \n"
-                    "vmla.f32   q12, q4, d0[0]      \n"
-                    "vmla.f32   q13, q4, d0[1]      \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmla.f32   q14, q4, d1[0]      \n"
-                    "vmla.f32   q15, q4, d1[1]      \n"
-                    "bne        4b                  \n"
-
-                    "5:                             \n"
-                    "vstm       %0!, {d24-d31}      \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4_t _sum0;
-                float32x4_t _sum1;
-                float32x4_t _sum2;
-                float32x4_t _sum3;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                    _sum2 = vdupq_n_f32(0.f);
-                    _sum3 = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    _sum0 = vld1q_f32(outptr);
-                    _sum1 = vld1q_f32(outptr + 4);
-                    _sum2 = vld1q_f32(outptr + 8);
-                    _sum3 = vld1q_f32(outptr + 12);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pA = vld1q_f32(pA);
-                    float32x4_t _pB = vld1q_f32(pB);
-
-#if __aarch64__
-                    _sum0 = vfmaq_laneq_f32(_sum0, _pA, _pB, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _pA, _pB, 1);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _pA, _pB, 2);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _pA, _pB, 3);
-#else
-                    _sum0 = vmlaq_lane_f32(_sum0, _pA, vget_low_f32(_pB), 0);
-                    _sum1 = vmlaq_lane_f32(_sum1, _pA, vget_low_f32(_pB), 1);
-                    _sum2 = vmlaq_lane_f32(_sum2, _pA, vget_high_f32(_pB), 0);
-                    _sum3 = vmlaq_lane_f32(_sum3, _pA, vget_high_f32(_pB), 1);
-#endif
-
-                    pA += 4;
-                    pB += 4;
-                }
-
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 4 * 2, _sum2);
-                vst1q_f32(outptr + 4 * 3, _sum3);
-                outptr += 4 * 4;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "cbz    %w7, 0f                     \n"
-
-                    "ld1    {v30.4s, v31.4s}, [%0]      \n"
-                    "b      1f                          \n"
-
-                    "0:                                 \n"
-                    "eor    v30.16b, v30.16b, v30.16b   \n"
-                    "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                    "1:                                 \n"
-                    "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    3f                          \n"
-
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-                    "eor    v29.16b, v29.16b, v29.16b   \n"
-                    "2:                                 \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%1], #64 \n"
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v0.4s, v1.4s}, [%2], #32   \n"
-                    "fmla   v28.4s, v16.4s, v0.s[0]     \n"
-                    "fmla   v29.4s, v16.4s, v0.s[1]     \n"
-                    "fmla   v30.4s, v17.4s, v0.s[2]     \n"
-                    "fmla   v31.4s, v17.4s, v0.s[3]     \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v28.4s, v18.4s, v1.s[0]     \n"
-                    "fmla   v29.4s, v18.4s, v1.s[1]     \n"
-                    "fmla   v30.4s, v19.4s, v1.s[2]     \n"
-                    "fmla   v31.4s, v19.4s, v1.s[3]     \n"
-                    "bne    2b                          \n"
-                    "fadd   v30.4s, v30.4s, v28.4s      \n"
-                    "fadd   v31.4s, v31.4s, v29.4s      \n"
-
-                    "3:                                 \n"
-                    "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "ld1    {v0.2s}, [%2], #8           \n"
-                    "ld1    {v16.4s}, [%1], #16         \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v30.4s, v16.4s, v0.s[0]     \n"
-                    "fmla   v31.4s, v16.4s, v0.s[1]     \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "st1    {v30.4s, v31.4s}, [%0], #32 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "cmp        %7, #0              \n"
-                    "beq        0f                  \n"
-
-                    "vld1.f32   {d28-d31}, [%0 :128] \n"
-                    "b          1f                  \n"
-
-                    "0:                             \n"
-                    "veor       q14, q14            \n"
-                    "veor       q15, q15            \n"
-
-                    "1:                             \n"
-                    "lsr        r4, %6, #2          \n" // r4 = max_kk >> 2
-                    "cmp        r4, #0              \n"
-                    "beq        3f                  \n"
-
-                    "veor       q12, q12            \n"
-                    "veor       q13, q13            \n"
-                    "2:                             \n"
-                    "pld        [%1, #512]          \n"
-                    "vldm       %1!, {d8-d15}       \n"
-                    "pld        [%2, #256]          \n"
-                    "vld1.f32   {d0-d3}, [%2 :128]! \n"
-                    "vmla.f32   q12, q4, d0[0]      \n"
-                    "vmla.f32   q13, q4, d0[1]      \n"
-                    "vmla.f32   q14, q5, d1[0]      \n"
-                    "vmla.f32   q15, q5, d1[1]      \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmla.f32   q12, q6, d2[0]      \n"
-                    "vmla.f32   q13, q6, d2[1]      \n"
-                    "vmla.f32   q14, q7, d3[0]      \n"
-                    "vmla.f32   q15, q7, d3[1]      \n"
-                    "bne        2b                  \n"
-                    "vadd.f32   q14, q14, q12       \n"
-                    "vadd.f32   q15, q15, q13       \n"
-
-                    "3:                             \n"
-                    "and        r4, %6, #3          \n" // r4 = remain = max_kk & 3
-                    "cmp        r4, #0              \n"
-                    "beq        5f                  \n"
-
-                    "4:                             \n"
-                    "vld1.f32   {d0}, [%2 :64]!     \n"
-                    "vld1.f32   {d8-d9}, [%1 :128]! \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmla.f32   q14, q4, d0[0]      \n"
-                    "vmla.f32   q15, q4, d0[1]      \n"
-                    "bne        4b                  \n"
-
-                    "5:                             \n"
-                    "vst1.f32   {d28-d31}, [%0 :128]! \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4_t _sum0;
-                float32x4_t _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    _sum0 = vld1q_f32(outptr);
-                    _sum1 = vld1q_f32(outptr + 4);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pA = vld1q_f32(pA);
-                    float32x2_t _pB = vld1_f32(pB);
-
-#if __aarch64__
-                    _sum0 = vfmaq_lane_f32(_sum0, _pA, _pB, 0);
-                    _sum1 = vfmaq_lane_f32(_sum1, _pA, _pB, 1);
-#else
-                    _sum0 = vmlaq_lane_f32(_sum0, _pA, _pB, 0);
-                    _sum1 = vmlaq_lane_f32(_sum1, _pA, _pB, 1);
-#endif
-
-                    pA += 4;
-                    pB += 2;
-                }
-
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                outptr += 4 * 2;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "cbz    %w7, 0f                     \n"
-
-                    "ld1    {v31.4s}, [%0]              \n"
-                    "b      1f                          \n"
-
-                    "0:                                 \n"
-                    "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                    "1:                                 \n"
-                    "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    3f                          \n"
-
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-                    "eor    v29.16b, v29.16b, v29.16b   \n"
-                    "eor    v30.16b, v30.16b, v30.16b   \n"
-                    "2:                                 \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%1], #64 \n"
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "ld1    {v0.4s}, [%2], #16          \n"
-                    "fmla   v28.4s, v16.4s, v0.s[0]     \n"
-                    "fmla   v29.4s, v17.4s, v0.s[1]     \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v30.4s, v18.4s, v0.s[2]     \n"
-                    "fmla   v31.4s, v19.4s, v0.s[3]     \n"
-                    "bne    2b                          \n"
-                    "fadd   v30.4s, v30.4s, v28.4s      \n"
-                    "fadd   v31.4s, v31.4s, v29.4s      \n"
-                    "fadd   v31.4s, v31.4s, v30.4s      \n"
-
-                    "3:                                 \n"
-                    "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "ld1r   {v0.4s}, [%2], #4           \n"
-                    "ld1    {v16.4s}, [%1], #16         \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v31.4s, v16.4s, v0.4s       \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "st1    {v31.4s}, [%0], #16         \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "cmp        %7, #0              \n"
-                    "beq        0f                  \n"
-
-                    "vld1.f32   {d30-d31}, [%0 :128] \n"
-                    "b          1f                  \n"
-
-                    "0:                             \n"
-                    "veor       q15, q15            \n"
-
-                    "1:                             \n"
-                    "lsr        r4, %6, #2          \n" // r4 = max_kk >> 2
-                    "cmp        r4, #0              \n"
-                    "beq        3f                  \n"
-
-                    "veor       q12, q12            \n"
-                    "veor       q13, q13            \n"
-                    "veor       q14, q14            \n"
-                    "2:                             \n"
-                    "pld        [%1, #512]          \n"
-                    "vldm       %1!, {d8-d15}       \n"
-                    "pld        [%2, #128]          \n"
-                    "vld1.f32   {d0-d1}, [%2 :64]!  \n"
-                    "vmla.f32   q12, q4, d0[0]      \n"
-                    "vmla.f32   q13, q5, d0[1]      \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmla.f32   q14, q6, d1[0]      \n"
-                    "vmla.f32   q15, q7, d1[1]      \n"
-                    "bne        2b                  \n"
-                    "vadd.f32   q14, q14, q12       \n"
-                    "vadd.f32   q15, q15, q13       \n"
-                    "vadd.f32   q15, q15, q14       \n"
-
-                    "3:                             \n"
-                    "and        r4, %6, #3          \n" // r4 = remain = max_kk & 3
-                    "cmp        r4, #0              \n"
-                    "beq        5f                  \n"
-
-                    "4:                             \n"
-                    "vld1.f32   {d0[0]}, [%2]!      \n"
-                    "vld1.f32   {d8-d9}, [%1 :128]! \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmla.f32   q15, q4, d0[0]      \n"
-                    "bne        4b                  \n"
-
-                    "5:                             \n"
-                    "vst1.f32   {d30-d31}, [%0 :128]! \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4_t _sum;
-
-                if (k == 0)
-                {
-                    _sum = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    _sum = vld1q_f32(outptr);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pA = vld1q_f32(pA);
-                    float32x4_t _pB = vdupq_n_f32(pB[0]);
-
-#if __aarch64__
-                    _sum = vfmaq_f32(_sum, _pA, _pB);
-#else
-                    _sum = vmlaq_f32(_sum, _pA, _pB);
-#endif
-
-                    pA += 4;
-                    pB += 1;
-                }
-
-                vst1q_f32(outptr, _sum);
-                outptr += 4;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const float* pAT = AT_tile.row(b) + max_kk * ii;
-            const float* pB = BT_tile.row(b);
-
-            int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const float* pA = pAT;
-
-                float32x4_t _sum00;
-                float32x4_t _sum01;
-                float32x4_t _sum02;
-                float32x4_t _sum10;
-                float32x4_t _sum11;
-                float32x4_t _sum12;
-
-                if (k == 0)
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum02 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                    _sum12 = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    float32x4x2_t _tmp01 = vld2q_f32(outptr);
-                    float32x4x2_t _tmp23 = vld2q_f32(outptr + 8);
-                    float32x4x2_t _tmp45 = vld2q_f32(outptr + 16);
-                    _sum00 = _tmp01.val[0];
-                    _sum01 = _tmp23.val[0];
-                    _sum02 = _tmp45.val[0];
-                    _sum10 = _tmp01.val[1];
-                    _sum11 = _tmp23.val[1];
-                    _sum12 = _tmp45.val[1];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pB0 = vld1q_f32(pB);
-                    float32x4_t _pB1 = vld1q_f32(pB + 4);
-                    float32x4_t _pB2 = vld1q_f32(pB + 8);
-
-                    float32x2_t _pA = vld1_f32(pA);
-#if __aarch64__
-                    _sum00 = vfmaq_lane_f32(_sum00, _pB0, _pA, 0);
-                    _sum01 = vfmaq_lane_f32(_sum01, _pB1, _pA, 0);
-                    _sum02 = vfmaq_lane_f32(_sum02, _pB2, _pA, 0);
-                    _sum10 = vfmaq_lane_f32(_sum10, _pB0, _pA, 1);
-                    _sum11 = vfmaq_lane_f32(_sum11, _pB1, _pA, 1);
-                    _sum12 = vfmaq_lane_f32(_sum12, _pB2, _pA, 1);
-#else
-                    _sum00 = vmlaq_lane_f32(_sum00, _pB0, _pA, 0);
-                    _sum01 = vmlaq_lane_f32(_sum01, _pB1, _pA, 0);
-                    _sum02 = vmlaq_lane_f32(_sum02, _pB2, _pA, 0);
-                    _sum10 = vmlaq_lane_f32(_sum10, _pB0, _pA, 1);
-                    _sum11 = vmlaq_lane_f32(_sum11, _pB1, _pA, 1);
-                    _sum12 = vmlaq_lane_f32(_sum12, _pB2, _pA, 1);
-#endif
-
-                    pA += 2;
-                    pB += 12;
-                }
-
-                float32x4x2_t _tmp01;
-                _tmp01.val[0] = _sum00;
-                _tmp01.val[1] = _sum10;
-                float32x4x2_t _tmp23;
-                _tmp23.val[0] = _sum01;
-                _tmp23.val[1] = _sum11;
-                float32x4x2_t _tmp45;
-                _tmp45.val[0] = _sum02;
-                _tmp45.val[1] = _sum12;
-                vst2q_f32(outptr, _tmp01);
-                vst2q_f32(outptr + 8, _tmp23);
-                vst2q_f32(outptr + 16, _tmp45);
-                outptr += 2 * 12;
-            }
-#endif // __aarch64__
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const float* pA = pAT;
-
-                float32x4_t _sum00;
-                float32x4_t _sum01;
-                float32x4_t _sum10;
-                float32x4_t _sum11;
-
-                if (k == 0)
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    float32x4x2_t _tmp01 = vld2q_f32(outptr);
-                    float32x4x2_t _tmp23 = vld2q_f32(outptr + 8);
-                    _sum00 = _tmp01.val[0];
-                    _sum01 = _tmp23.val[0];
-                    _sum10 = _tmp01.val[1];
-                    _sum11 = _tmp23.val[1];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pB0 = vld1q_f32(pB);
-                    float32x4_t _pB1 = vld1q_f32(pB + 4);
-
-                    float32x2_t _pA = vld1_f32(pA);
-#if __aarch64__
-                    _sum00 = vfmaq_lane_f32(_sum00, _pB0, _pA, 0);
-                    _sum01 = vfmaq_lane_f32(_sum01, _pB1, _pA, 0);
-                    _sum10 = vfmaq_lane_f32(_sum10, _pB0, _pA, 1);
-                    _sum11 = vfmaq_lane_f32(_sum11, _pB1, _pA, 1);
-#else
-                    _sum00 = vmlaq_lane_f32(_sum00, _pB0, _pA, 0);
-                    _sum01 = vmlaq_lane_f32(_sum01, _pB1, _pA, 0);
-                    _sum10 = vmlaq_lane_f32(_sum10, _pB0, _pA, 1);
-                    _sum11 = vmlaq_lane_f32(_sum11, _pB1, _pA, 1);
-#endif
-
-                    pA += 2;
-                    pB += 8;
-                }
-
-                float32x4x2_t _tmp01;
-                _tmp01.val[0] = _sum00;
-                _tmp01.val[1] = _sum10;
-                float32x4x2_t _tmp23;
-                _tmp23.val[0] = _sum01;
-                _tmp23.val[1] = _sum11;
-                vst2q_f32(outptr, _tmp01);
-                vst2q_f32(outptr + 8, _tmp23);
-                outptr += 2 * 8;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const float* pA = pAT;
-
-                float32x4_t _sum0;
-                float32x4_t _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    float32x4x2_t _tmp01 = vld2q_f32(outptr);
-                    _sum0 = _tmp01.val[0];
-                    _sum1 = _tmp01.val[1];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pB = vld1q_f32(pB);
-
-                    float32x2_t _pA = vld1_f32(pA);
-#if __aarch64__
-                    _sum0 = vfmaq_lane_f32(_sum0, _pB, _pA, 0);
-                    _sum1 = vfmaq_lane_f32(_sum1, _pB, _pA, 1);
-#else
-                    _sum0 = vmlaq_lane_f32(_sum0, _pB, _pA, 0);
-                    _sum1 = vmlaq_lane_f32(_sum1, _pB, _pA, 1);
-#endif
-
-                    pA += 2;
-                    pB += 4;
-                }
-
-                float32x4x2_t _tmp01;
-                _tmp01.val[0] = _sum0;
-                _tmp01.val[1] = _sum1;
-                vst2q_f32(outptr, _tmp01);
-                outptr += 2 * 4;
-            }
-#endif // __ARM_NEON
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const float* pA = pAT;
-
-                float sum00 = 0.f;
-                float sum01 = 0.f;
-                float sum10 = 0.f;
-                float sum11 = 0.f;
-
-                if (k == 0)
-                {
-                    sum00 = 0.f;
-                    sum01 = 0.f;
-                    sum10 = 0.f;
-                    sum11 = 0.f;
-                }
-                else
-                {
-                    sum00 = outptr[0];
-                    sum01 = outptr[1];
-                    sum10 = outptr[2];
-                    sum11 = outptr[3];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    sum00 += pA[0] * pB[0];
-                    sum01 += pA[1] * pB[0];
-                    sum10 += pA[0] * pB[1];
-                    sum11 += pA[1] * pB[1];
-                    pA += 2;
-                    pB += 2;
-                }
-
-                outptr[0] = sum00;
-                outptr[1] = sum01;
-                outptr[2] = sum10;
-                outptr[3] = sum11;
-                outptr += 2 * 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const float* pA = pAT;
-
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-
-                if (k == 0)
-                {
-                    sum0 = 0.f;
-                    sum1 = 0.f;
-                }
-                else
-                {
-                    sum0 = outptr[0];
-                    sum1 = outptr[1];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    sum0 += pA[0] * pB[0];
-                    sum1 += pA[1] * pB[0];
-                    pA += 2;
-                    pB += 1;
-                }
-
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-                outptr += 2;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const float* pAT = AT_tile.row(b) + max_kk * ii;
-            const float* pB = BT_tile.row(b);
-
-            int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const float* pA = pAT;
-
-                float32x4_t _sum0;
-                float32x4_t _sum1;
-                float32x4_t _sum2;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                    _sum2 = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    _sum0 = vld1q_f32(outptr);
-                    _sum1 = vld1q_f32(outptr + 4);
-                    _sum2 = vld1q_f32(outptr + 8);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pB0 = vld1q_f32(pB);
-                    float32x4_t _pB1 = vld1q_f32(pB + 4);
-                    float32x4_t _pB2 = vld1q_f32(pB + 8);
-
-                    float32x4_t _pA0 = vdupq_n_f32(pA[0]);
-#if __aarch64__
-                    _sum0 = vfmaq_f32(_sum0, _pA0, _pB0);
-                    _sum1 = vfmaq_f32(_sum1, _pA0, _pB1);
-                    _sum2 = vfmaq_f32(_sum2, _pA0, _pB2);
-#else
-                    _sum0 = vmlaq_f32(_sum0, _pA0, _pB0);
-                    _sum1 = vmlaq_f32(_sum1, _pA0, _pB1);
-                    _sum2 = vmlaq_f32(_sum2, _pA0, _pB2);
-#endif
-
-                    pA += 1;
-                    pB += 12;
-                }
-
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 8, _sum2);
-                outptr += 12;
-            }
-#endif // __aarch64__
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const float* pA = pAT;
-
-                float32x4_t _sum0;
-                float32x4_t _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    _sum0 = vld1q_f32(outptr);
-                    _sum1 = vld1q_f32(outptr + 4);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pB0 = vld1q_f32(pB);
-                    float32x4_t _pB1 = vld1q_f32(pB + 4);
-
-                    float32x4_t _pA0 = vdupq_n_f32(pA[0]);
-#if __aarch64__
-                    _sum0 = vfmaq_f32(_sum0, _pA0, _pB0);
-                    _sum1 = vfmaq_f32(_sum1, _pA0, _pB1);
-#else
-                    _sum0 = vmlaq_f32(_sum0, _pA0, _pB0);
-                    _sum1 = vmlaq_f32(_sum1, _pA0, _pB1);
-#endif
-
-                    pA += 1;
-                    pB += 8;
-                }
-
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                outptr += 8;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const float* pA = pAT;
-
-                float32x4_t _sum;
-
-                if (k == 0)
-                {
-                    _sum = vdupq_n_f32(0.f);
-                }
-                else
-                {
-                    _sum = vld1q_f32(outptr);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float32x4_t _pB = vld1q_f32(pB);
-                    float32x4_t _pA = vdupq_n_f32(pA[0]);
-
-#if __aarch64__
-                    _sum = vfmaq_f32(_sum, _pA, _pB);
-#else
-                    _sum = vmlaq_f32(_sum, _pA, _pB);
-#endif
-
-                    pA += 1;
-                    pB += 4;
-                }
-
-                vst1q_f32(outptr, _sum);
-                outptr += 4;
-            }
-#endif // __ARM_NEON
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const float* pA = pAT;
-
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-
-                if (k == 0)
-                {
-                    sum0 = 0.f;
-                    sum1 = 0.f;
-                }
-                else
-                {
-                    sum0 = outptr[0];
-                    sum1 = outptr[1];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    sum0 += pA[0] * pB[0];
-                    sum1 += pA[0] * pB[1];
-                    pA += 1;
-                    pB += 2;
-                }
-
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-                outptr += 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const float* pA = pAT;
-
-                float sum = 0.f;
-
-                if (k == 0)
-                {
-                    sum = 0.f;
-                }
-                else
-                {
-                    sum = outptr[0];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    sum += pA[0] * pB[0];
-                    pA += 1;
-                    pB += 1;
-                }
-
-                outptr[0] = sum;
-                outptr += 1;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd_get_optimal_tile_mnk(int M, int N, int K, int B, int& TILE_M, int& TILE_N, int& TILE_K, int nT)
-{
-    // resolve optimal tile size from cache size
-    const int l2_cache_size_fp32 = (int)(get_cpu_level2_cache_size() / sizeof(float));
-
-    if (nT == 0)
-        nT = get_physical_big_cpu_count();
-
-    // we shall take B into account for batched gemm, but that will be slower on arm in practice, why ?
-    (void)B;
-
-    // solve K
-    {
-        // try not to split K
-#if __aarch64__
-        int tile_size = (l2_cache_size_fp32 - 32) / 12;
-#elif __ARM_NEON
-        int tile_size = (l2_cache_size_fp32 - 16) / 8;
-#else
-        int tile_size = (l2_cache_size_fp32 - 2) / 3;
-#endif
-
-#if __aarch64__
-        TILE_K = std::max(8, tile_size / 8 * 8);
-#elif __ARM_NEON
-        TILE_K = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_K = std::max(2, tile_size / 2 * 2);
-#endif
-
-        int nn_K = (K + TILE_K - 1) / TILE_K;
-#if __aarch64__
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 7) / 8 * 8);
-#elif __ARM_NEON
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 3) / 4 * 4);
-#else
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 1) / 2 * 2);
-#endif
-    }
-
-    // solve M
-    {
-#if __aarch64__
-        TILE_M = 8;
-#elif __ARM_NEON
-        TILE_M = 4;
-#else
-        TILE_M = 2;
-#endif
-    }
-
-    {
-        TILE_M *= std::min(nT, get_physical_cpu_count());
-
-        int nn_M = (M + TILE_M - 1) / TILE_M;
-#if __aarch64__
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-#elif __ARM_NEON
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 3) / 4 * 4);
-#else
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 1) / 2 * 2);
-#endif
-
-        if (nT > 1)
-        {
-#if __aarch64__
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 7) / 8 * 8);
-#elif __ARM_NEON
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 3) / 4 * 4);
-#else
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 1) / 2 * 2);
-#endif
-        }
-
-#if __aarch64__
-        TILE_M = std::max(8, TILE_M);
-#elif __ARM_NEON
-        TILE_M = std::max(4, TILE_M);
-#else
-        TILE_M = std::max(2, TILE_M);
-#endif
-    }
-
-    if (N > 0)
-    {
-        int tile_size;
-        if (TILE_K >= K)
-        {
-            tile_size = (l2_cache_size_fp32 - TILE_M * TILE_K) / TILE_K;
-        }
-        else
-        {
-            tile_size = (l2_cache_size_fp32 - TILE_M * TILE_K) / (TILE_M + TILE_K);
-        }
-
-#if __aarch64__
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#elif __ARM_NEON
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_N = std::max(1, tile_size);
-#endif
-
-        int nn_N = (N + TILE_N - 1) / TILE_N;
-
-#if __aarch64__
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#elif __ARM_NEON
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#else
-        TILE_N = std::min(TILE_N, (N + nn_N - 1) / nn_N);
-#endif
-
-#if __aarch64__
-        TILE_N = std::max(4, TILE_N);
-#elif __ARM_NEON
-        TILE_N = std::max(4, TILE_N);
-#else
-        TILE_N = std::max(1, TILE_N);
-#endif
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_kernel_tile(const Mat& kernel, Mat& A, int inch, int i, int max_ii, int k, int max_kk)
-{
-    // const float ktm[4][3] = {
-    //     {1.0f, 0.0f, 0.0f},
-    //     {1.0f / 2, 1.0f / 2, 1.0f / 2},
-    //     {1.0f / 2, -1.0f / 2, 1.0f / 2},
-    //     {0.0f, 0.0f, 1.0f}
-    // };
-
-    float* ptmp = A;
-
-    int ii = 0;
-    for (; ii < max_ii; ii++)
-    {
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            float tmp[4][3];
-
-            const float* k0 = (const float*)kernel + (i + ii) * inch * 9 + (k + kk) * 9;
-
-            for (int m = 0; m < 3; m++)
-            {
-                float r0 = k0[0];
-                float r1 = k0[1];
-                float r2 = k0[2];
-
-                tmp[0][m] = r0;
-                tmp[1][m] = r0 * 0.5f + r1 * 0.5f + r2 * 0.5f;
-                tmp[2][m] = r0 * 0.5f - r1 * 0.5f + r2 * 0.5f;
-                tmp[3][m] = r2;
-
-                k0 += 3;
-            }
-
-            for (int m = 0; m < 4; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-
-                float z0 = r0;
-                float z1 = r0 * 0.5f + r1 * 0.5f + r2 * 0.5f;
-                float z2 = r0 * 0.5f - r1 * 0.5f + r2 * 0.5f;
-                float z3 = r2;
-
-                ptmp[0] = z0;
-                ptmp[1] = z1;
-                ptmp[2] = z2;
-                ptmp[3] = z3;
-                ptmp += 4;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_kernel(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt)
-{
-    const int M = outch;
-    const int K = inch;
-    const int B = 16;
-
-    int TILE_M, TILE_N, TILE_K;
-    conv3x3s1_winograd_get_optimal_tile_mnk(M, 0, K, B, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    Mat A_tileX(B * TILE_M * TILE_K, 1, opt.num_threads);
-
-    AT.create(TILE_K * TILE_M, B, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat A_tile = A_tileX.channel(get_omp_thread_num());
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_ii = std::min((M - i), TILE_M);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            conv3x3s1_winograd23_transform_kernel_tile(kernel, A_tile, inch, i, max_ii, k, max_kk);
-
-            Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-            conv3x3s1_winograd_pack_A_tile(A_tile, AT_tile, B, max_ii, max_kk);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_input_tile(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    // const float itm[4][4] = {
-    //     {1.0f,  0.0f, -1.0f,  0.0f},
-    //     {0.0f,  1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  0.00f, 1.0f}
-    // };
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w - 1) / 2;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-#if __ARM_NEON
-#if __aarch64__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[4][4][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel((k + kk) / elempack).row(ti * 2) + (tj * 2) * elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float32x4_t _r00 = vdupq_n_f32(0.f);
-                float32x4_t _r01 = vdupq_n_f32(0.f);
-                float32x4_t _r10 = vdupq_n_f32(0.f);
-                float32x4_t _r11 = vdupq_n_f32(0.f);
-                float32x4_t _r20 = vdupq_n_f32(0.f);
-                float32x4_t _r21 = vdupq_n_f32(0.f);
-                float32x4_t _r30 = vdupq_n_f32(0.f);
-                float32x4_t _r31 = vdupq_n_f32(0.f);
-
-                if (ti * 2 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-
-                        _r00 = vld1q_f32(r0);
-                        _r01 = vld1q_f32(r1);
-                        if (tj * 2 + 1 < w)
-                        {
-                            _r10 = vld1q_f32(r0 + 4);
-                            _r11 = vld1q_f32(r1 + 4);
-                        }
-                        if (tj * 2 + 2 < w)
-                        {
-                            _r20 = vld1q_f32(r0 + 8);
-                            _r21 = vld1q_f32(r1 + 8);
-                        }
-                        if (tj * 2 + 3 < w)
-                        {
-                            _r30 = vld1q_f32(r0 + 12);
-                            _r31 = vld1q_f32(r1 + 12);
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-                        const float* r4 = r0 + N * 4;
-                        const float* r5 = r0 + N * 5;
-                        const float* r6 = r0 + N * 6;
-                        const float* r7 = r0 + N * 7;
-
-                        float32x4_t _t0 = vld1q_f32(r0);
-                        float32x4_t _t1 = vld1q_f32(r1);
-                        float32x4_t _t2 = vld1q_f32(r2);
-                        float32x4_t _t3 = vld1q_f32(r3);
-                        float32x4_t _t4 = vld1q_f32(r4);
-                        float32x4_t _t5 = vld1q_f32(r5);
-                        float32x4_t _t6 = vld1q_f32(r6);
-                        float32x4_t _t7 = vld1q_f32(r7);
-
-                        transpose4x4_ps(_t0, _t1, _t2, _t3);
-                        transpose4x4_ps(_t4, _t5, _t6, _t7);
-
-                        _r00 = _t0;
-                        _r01 = _t4;
-                        if (tj * 2 + 1 < w)
-                        {
-                            _r10 = _t1;
-                            _r11 = _t5;
-                        }
-                        if (tj * 2 + 2 < w)
-                        {
-                            _r20 = _t2;
-                            _r21 = _t6;
-                        }
-                        if (tj * 2 + 3 < w)
-                        {
-                            _r30 = _t3;
-                            _r31 = _t7;
-                        }
-                    }
-                }
-
-                float32x4_t _tmp00 = vsubq_f32(_r00, _r20);
-                float32x4_t _tmp01 = vsubq_f32(_r01, _r21);
-                float32x4_t _tmp10 = vaddq_f32(_r10, _r20);
-                float32x4_t _tmp11 = vaddq_f32(_r11, _r21);
-                float32x4_t _tmp20 = vsubq_f32(_r20, _r10);
-                float32x4_t _tmp21 = vsubq_f32(_r21, _r11);
-                float32x4_t _tmp30 = vsubq_f32(_r30, _r10);
-                float32x4_t _tmp31 = vsubq_f32(_r31, _r11);
-
-                vst1q_f32(tmp[0][m], _tmp00);
-                vst1q_f32(tmp[0][m] + 4, _tmp01);
-                vst1q_f32(tmp[1][m], _tmp10);
-                vst1q_f32(tmp[1][m] + 4, _tmp11);
-                vst1q_f32(tmp[2][m], _tmp20);
-                vst1q_f32(tmp[2][m] + 4, _tmp21);
-                vst1q_f32(tmp[3][m], _tmp30);
-                vst1q_f32(tmp[3][m] + 4, _tmp31);
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 16 + jj * 8;
-            float* p1 = p0 + max_jj * 8;
-            float* p2 = p0 + max_jj * 8 * 2;
-            float* p3 = p0 + max_jj * 8 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float32x4_t _r00 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r01 = vld1q_f32(tmp[m][0] + 4);
-                float32x4_t _r10 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r11 = vld1q_f32(tmp[m][1] + 4);
-                float32x4_t _r20 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r21 = vld1q_f32(tmp[m][2] + 4);
-                float32x4_t _r30 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r31 = vld1q_f32(tmp[m][3] + 4);
-
-                float32x4_t _tmp00 = vsubq_f32(_r00, _r20);
-                float32x4_t _tmp01 = vsubq_f32(_r01, _r21);
-                float32x4_t _tmp10 = vaddq_f32(_r10, _r20);
-                float32x4_t _tmp11 = vaddq_f32(_r11, _r21);
-                float32x4_t _tmp20 = vsubq_f32(_r20, _r10);
-                float32x4_t _tmp21 = vsubq_f32(_r21, _r11);
-                float32x4_t _tmp30 = vsubq_f32(_r30, _r10);
-                float32x4_t _tmp31 = vsubq_f32(_r31, _r11);
-
-                vst1q_f32(p0, _tmp00);
-                vst1q_f32(p0 + 4, _tmp01);
-                vst1q_f32(p1, _tmp10);
-                vst1q_f32(p1 + 4, _tmp11);
-                vst1q_f32(p2, _tmp20);
-                vst1q_f32(p2 + 4, _tmp21);
-                vst1q_f32(p3, _tmp30);
-                vst1q_f32(p3 + 4, _tmp31);
-
-                p0 += max_jj * 4 * 8;
-                p1 += max_jj * 4 * 8;
-                p2 += max_jj * 4 * 8;
-                p3 += max_jj * 4 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-#else // __aarch64__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-    #pragma omp parallel for num_threads(nT)
-#endif // __aarch64__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 4;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[4][4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel((k + kk) / elempack).row(ti * 2) + (tj * 2) * elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float32x4_t _r0 = vdupq_n_f32(0.f);
-                float32x4_t _r1 = vdupq_n_f32(0.f);
-                float32x4_t _r2 = vdupq_n_f32(0.f);
-                float32x4_t _r3 = vdupq_n_f32(0.f);
-
-                if (ti * 2 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1q_f32(r0);
-                        if (tj * 2 + 1 < w) _r1 = vld1q_f32(r0 + 4);
-                        if (tj * 2 + 2 < w) _r2 = vld1q_f32(r0 + 8);
-                        if (tj * 2 + 3 < w) _r3 = vld1q_f32(r0 + 12);
-                    }
-                    if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-
-                        float32x4_t _t0 = vld1q_f32(r0);
-                        float32x4_t _t1 = vld1q_f32(r1);
-                        float32x4_t _t2 = vld1q_f32(r2);
-                        float32x4_t _t3 = vld1q_f32(r3);
-
-                        transpose4x4_ps(_t0, _t1, _t2, _t3);
-
-                        _r0 = _t0;
-                        if (tj * 2 + 1 < w) _r1 = _t1;
-                        if (tj * 2 + 2 < w) _r2 = _t2;
-                        if (tj * 2 + 3 < w) _r3 = _t3;
-                    }
-                }
-
-                float32x4_t _tmp0 = vsubq_f32(_r0, _r2);
-                float32x4_t _tmp1 = vaddq_f32(_r1, _r2);
-                float32x4_t _tmp2 = vsubq_f32(_r2, _r1);
-                float32x4_t _tmp3 = vsubq_f32(_r3, _r1);
-
-                vst1q_f32(tmp[0][m], _tmp0);
-                vst1q_f32(tmp[1][m], _tmp1);
-                vst1q_f32(tmp[2][m], _tmp2);
-                vst1q_f32(tmp[3][m], _tmp3);
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 16 + jj * 4;
-            float* p1 = p0 + max_jj * 4;
-            float* p2 = p0 + max_jj * 4 * 2;
-            float* p3 = p0 + max_jj * 4 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float32x4_t _r0 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r1 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r2 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r3 = vld1q_f32(tmp[m][3]);
-
-                float32x4_t _tmp0 = vsubq_f32(_r0, _r2);
-                float32x4_t _tmp1 = vaddq_f32(_r1, _r2);
-                float32x4_t _tmp2 = vsubq_f32(_r2, _r1);
-                float32x4_t _tmp3 = vsubq_f32(_r3, _r1);
-
-                vst1q_f32(p0, _tmp0);
-                vst1q_f32(p1, _tmp1);
-                vst1q_f32(p2, _tmp2);
-                vst1q_f32(p3, _tmp3);
-
-                p0 += max_jj * 4 * 4;
-                p1 += max_jj * 4 * 4;
-                p2 += max_jj * 4 * 4;
-                p3 += max_jj * 4 * 4;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 4;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-#else // __ARM_NEON
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    #pragma omp parallel for num_threads(nT)
-#endif // __ARM_NEON
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        float tmp[4][4][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel(k + kk).row(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 4; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vdup_n_f32(0.f);
-                float32x2_t _r1 = vdup_n_f32(0.f);
-                float32x2_t _r2 = vdup_n_f32(0.f);
-                float32x2_t _r3 = vdup_n_f32(0.f);
-#else
-                float r00 = 0.f;
-                float r01 = 0.f;
-                float r10 = 0.f;
-                float r11 = 0.f;
-                float r20 = 0.f;
-                float r21 = 0.f;
-                float r30 = 0.f;
-                float r31 = 0.f;
-#endif
-
-                if (ti * 2 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-
-#if __ARM_NEON
-                        float32x4_t _t0 = vld1q_f32(r0);
-                        float32x4_t _t1 = vld1q_f32(r1);
-                        float32x4x2_t _t01 = vzipq_f32(_t0, _t1);
-
-                        _r0 = vget_low_f32(_t01.val[0]);
-                        if (tj * 2 + 1 < w) _r1 = vget_high_f32(_t01.val[0]);
-                        if (tj * 2 + 2 < w) _r2 = vget_low_f32(_t01.val[1]);
-                        if (tj * 2 + 3 < w) _r3 = vget_high_f32(_t01.val[1]);
-#else
-                        r00 = r0[0];
-                        r01 = r1[0];
-                        if (tj * 2 + 1 < w)
-                        {
-                            r10 = r0[1];
-                            r11 = r1[1];
-                        }
-                        if (tj * 2 + 2 < w)
-                        {
-                            r20 = r0[2];
-                            r21 = r1[2];
-                        }
-                        if (tj * 2 + 3 < w)
-                        {
-                            r30 = r0[3];
-                            r31 = r1[3];
-                        }
-#endif
-                    }
-                }
-
-#if __ARM_NEON
-                float32x2_t _tmp0 = vsub_f32(_r0, _r2);
-                float32x2_t _tmp1 = vadd_f32(_r1, _r2);
-                float32x2_t _tmp2 = vsub_f32(_r2, _r1);
-                float32x2_t _tmp3 = vsub_f32(_r3, _r1);
-
-                vst1_f32(tmp[0][m], _tmp0);
-                vst1_f32(tmp[1][m], _tmp1);
-                vst1_f32(tmp[2][m], _tmp2);
-                vst1_f32(tmp[3][m], _tmp3);
-#else
-                tmp[0][m][0] = r00 - r20;
-                tmp[0][m][1] = r01 - r21;
-                tmp[1][m][0] = r10 + r20;
-                tmp[1][m][1] = r11 + r21;
-                tmp[2][m][0] = r20 - r10;
-                tmp[2][m][1] = r21 - r11;
-                tmp[3][m][0] = r30 - r10;
-                tmp[3][m][1] = r31 - r11;
-#endif
-
-                r0 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 16 + jj * 2;
-            float* p1 = p0 + max_jj * 2;
-            float* p2 = p0 + max_jj * 2 * 2;
-            float* p3 = p0 + max_jj * 2 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(tmp[m][0]);
-                float32x2_t _r1 = vld1_f32(tmp[m][1]);
-                float32x2_t _r2 = vld1_f32(tmp[m][2]);
-                float32x2_t _r3 = vld1_f32(tmp[m][3]);
-
-                float32x2_t _tmp0 = vsub_f32(_r0, _r2);
-                float32x2_t _tmp1 = vadd_f32(_r1, _r2);
-                float32x2_t _tmp2 = vsub_f32(_r2, _r1);
-                float32x2_t _tmp3 = vsub_f32(_r3, _r1);
-
-                vst1_f32(p0, _tmp0);
-                vst1_f32(p1, _tmp1);
-                vst1_f32(p2, _tmp2);
-                vst1_f32(p3, _tmp3);
-#else
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-
-                p0[0] = r00 - r20;
-                p0[1] = r01 - r21;
-                p1[0] = r10 + r20;
-                p1[1] = r11 + r21;
-                p2[0] = r20 - r10;
-                p2[1] = r21 - r11;
-                p3[0] = r30 - r10;
-                p3[1] = r31 - r11;
-#endif
-
-                p0 += max_jj * 4 * 2;
-                p1 += max_jj * 4 * 2;
-                p2 += max_jj * 4 * 2;
-                p3 += max_jj * 4 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        float tmp[4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0123 = bottom_blob.channel(k + kk).row(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 4; m++)
-            {
-                float r0 = 0.f;
-                float r1 = 0.f;
-                float r2 = 0.f;
-                float r3 = 0.f;
-
-                if (ti * 2 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = r0123[0];
-                        if (tj * 2 + 1 < w) r1 = r0123[1];
-                        if (tj * 2 + 2 < w) r2 = r0123[2];
-                        if (tj * 2 + 3 < w) r3 = r0123[3];
-                    }
-                }
-
-                tmp[0][m] = r0 - r2;
-                tmp[1][m] = r1 + r2;
-                tmp[2][m] = r2 - r1;
-                tmp[3][m] = r3 - r1;
-
-                r0123 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 16 + jj;
-            float* p1 = p0 + max_jj;
-            float* p2 = p0 + max_jj * 2;
-            float* p3 = p0 + max_jj * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-
-                p0[0] = r0 - r2;
-                p1[0] = r1 + r2;
-                p2[0] = r2 - r1;
-                p3[0] = r3 - r1;
-
-                p0 += max_jj * 4;
-                p1 += max_jj * 4;
-                p2 += max_jj * 4;
-                p3 += max_jj * 4;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_output_tile(const Mat& top_tile, Mat& top_blob, const Mat& bias, int i, int max_ii, int j, int max_jj)
-{
-    // const float otm[2][4] = {
-    //     {1.0f,  1.0f,  1.0f,  0.0f},
-    //     {0.0f,  1.0f, -1.0f,  1.0f}
-    // };
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 1) / 2;
-
-    const float* biasptr = bias;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        float32x4_t _bias0 = biasptr ? vld1q_f32(biasptr + i + ii) : vdupq_n_f32(0.f);
-        float32x4_t _bias1 = biasptr ? vld1q_f32(biasptr + i + ii + 4) : vdupq_n_f32(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[2][4][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 16 + jj * 8;
-            const float* r1 = r0 + max_jj * 8;
-            const float* r2 = r0 + max_jj * 8 * 2;
-            const float* r3 = r0 + max_jj * 8 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r30 = vld1q_f32(r3);
-                float32x4_t _r31 = vld1q_f32(r3 + 4);
-
-                float32x4_t _tmp00 = vaddq_f32(vaddq_f32(_r00, _r10), _r20);
-                float32x4_t _tmp01 = vaddq_f32(vaddq_f32(_r01, _r11), _r21);
-                float32x4_t _tmp10 = vaddq_f32(vsubq_f32(_r10, _r20), _r30);
-                float32x4_t _tmp11 = vaddq_f32(vsubq_f32(_r11, _r21), _r31);
-
-                vst1q_f32(tmp[0][m], _tmp00);
-                vst1q_f32(tmp[0][m] + 4, _tmp01);
-                vst1q_f32(tmp[1][m], _tmp10);
-                vst1q_f32(tmp[1][m] + 4, _tmp11);
-
-                r0 += max_jj * 4 * 8;
-                r1 += max_jj * 4 * 8;
-                r2 += max_jj * 4 * 8;
-                r3 += max_jj * 4 * 8;
-            }
-
-            float* outptr0 = top_blob.channel((i + ii) / out_elempack).row(ti * 2) + (tj * 2) * out_elempack;
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                float32x4_t _r00 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r01 = vld1q_f32(tmp[m][0] + 4);
-                float32x4_t _r10 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r11 = vld1q_f32(tmp[m][1] + 4);
-                float32x4_t _r20 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r21 = vld1q_f32(tmp[m][2] + 4);
-                float32x4_t _r30 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r31 = vld1q_f32(tmp[m][3] + 4);
-
-                float32x4_t _tmp00 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r00, _r10), _r20));
-                float32x4_t _tmp01 = vaddq_f32(_bias1, vaddq_f32(vaddq_f32(_r01, _r11), _r21));
-                float32x4_t _tmp10 = vaddq_f32(_bias0, vaddq_f32(vsubq_f32(_r10, _r20), _r30));
-                float32x4_t _tmp11 = vaddq_f32(_bias1, vaddq_f32(vsubq_f32(_r11, _r21), _r31));
-
-                if (out_elempack == 4)
-                {
-                    float* outptr1 = outptr0 + N;
-
-                    vst1q_f32(outptr0, _tmp00);
-                    vst1q_f32(outptr1, _tmp01);
-                    if (tj * 2 + 1 < outw)
-                    {
-                        vst1q_f32(outptr0 + 4, _tmp10);
-                        vst1q_f32(outptr1 + 4, _tmp11);
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    float tmp0[8];
-                    float tmp1[8];
-                    vst1q_f32(tmp0, _tmp00);
-                    vst1q_f32(tmp0 + 4, _tmp01);
-                    vst1q_f32(tmp1, _tmp10);
-                    vst1q_f32(tmp1 + 4, _tmp11);
-
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-                    float* outptr4 = outptr0 + N * 4;
-                    float* outptr5 = outptr0 + N * 5;
-                    float* outptr6 = outptr0 + N * 6;
-                    float* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        float32x4_t _bias0 = biasptr ? vld1q_f32(biasptr + i + ii) : vdupq_n_f32(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[2][4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 16 + jj * 4;
-            const float* r1 = r0 + max_jj * 4;
-            const float* r2 = r0 + max_jj * 4 * 2;
-            const float* r3 = r0 + max_jj * 4 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float32x4_t _r0 = vld1q_f32(r0);
-                float32x4_t _r1 = vld1q_f32(r1);
-                float32x4_t _r2 = vld1q_f32(r2);
-                float32x4_t _r3 = vld1q_f32(r3);
-
-                float32x4_t _tmp0 = vaddq_f32(vaddq_f32(_r0, _r1), _r2);
-                float32x4_t _tmp1 = vaddq_f32(vsubq_f32(_r1, _r2), _r3);
-
-                vst1q_f32(tmp[0][m], _tmp0);
-                vst1q_f32(tmp[1][m], _tmp1);
-
-                r0 += max_jj * 4 * 4;
-                r1 += max_jj * 4 * 4;
-                r2 += max_jj * 4 * 4;
-                r3 += max_jj * 4 * 4;
-            }
-
-            float* outptr0 = top_blob.channel((i + ii) / out_elempack).row(ti * 2) + (tj * 2) * out_elempack;
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                float32x4_t _r0 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r1 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r2 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r3 = vld1q_f32(tmp[m][3]);
-
-                float32x4_t _tmp0 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r0, _r1), _r2));
-                float32x4_t _tmp1 = vaddq_f32(_bias0, vaddq_f32(vsubq_f32(_r1, _r2), _r3));
-
-                if (out_elempack == 4)
-                {
-                    vst1q_f32(outptr0, _tmp0);
-                    if (tj * 2 + 1 < outw) vst1q_f32(outptr0 + 4, _tmp1);
-                }
-                if (out_elempack == 1)
-                {
-                    float tmp0[4];
-                    float tmp1[4];
-                    vst1q_f32(tmp0, _tmp0);
-                    vst1q_f32(tmp1, _tmp1);
-
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-#if __ARM_NEON
-        float32x2_t _bias0 = biasptr ? vld1_f32(biasptr + i + ii) : vdup_n_f32(0.f);
-#else
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-        float bias1 = biasptr ? biasptr[i + ii + 1] : 0.f;
-#endif
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        float tmp[2][4][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 16 + jj * 2;
-            const float* r1 = r0 + max_jj * 2;
-            const float* r2 = r0 + max_jj * 2 * 2;
-            const float* r3 = r0 + max_jj * 2 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(r0);
-                float32x2_t _r1 = vld1_f32(r1);
-                float32x2_t _r2 = vld1_f32(r2);
-                float32x2_t _r3 = vld1_f32(r3);
-
-                float32x2_t _tmp0 = vadd_f32(vadd_f32(_r0, _r1), _r2);
-                float32x2_t _tmp1 = vadd_f32(vsub_f32(_r1, _r2), _r3);
-
-                vst1_f32(tmp[0][m], _tmp0);
-                vst1_f32(tmp[1][m], _tmp1);
-#else
-                tmp[0][m][0] = r0[0] + r1[0] + r2[0];
-                tmp[0][m][1] = r0[1] + r1[1] + r2[1];
-                tmp[1][m][0] = r1[0] - r2[0] + r3[0];
-                tmp[1][m][1] = r1[1] - r2[1] + r3[1];
-#endif
-
-                r0 += max_jj * 4 * 2;
-                r1 += max_jj * 4 * 2;
-                r2 += max_jj * 4 * 2;
-                r3 += max_jj * 4 * 2;
-            }
-
-            float* outptr0 = top_blob.channel(i + ii).row(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(tmp[m][0]);
-                float32x2_t _r1 = vld1_f32(tmp[m][1]);
-                float32x2_t _r2 = vld1_f32(tmp[m][2]);
-                float32x2_t _r3 = vld1_f32(tmp[m][3]);
-
-                float32x2_t _tmp0 = vadd_f32(_bias0, vadd_f32(vadd_f32(_r0, _r1), _r2));
-                float32x2_t _tmp1 = vadd_f32(_bias0, vadd_f32(vsub_f32(_r1, _r2), _r3));
-#else
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-
-                float tmp00 = bias0 + r00 + r10 + r20;
-                float tmp01 = bias1 + r01 + r11 + r21;
-                float tmp10 = bias0 + r10 - r20 + r30;
-                float tmp11 = bias1 + r11 - r21 + r31;
-#endif
-
-                // if (out_elempack == 1)
-                {
-                    float* outptr1 = outptr0 + N;
-
-#if __ARM_NEON
-                    outptr0[0] = vget_lane_f32(_tmp0, 0);
-                    outptr1[0] = vget_lane_f32(_tmp0, 1);
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = vget_lane_f32(_tmp1, 0);
-                        outptr1[1] = vget_lane_f32(_tmp1, 1);
-                    }
-#else
-                    outptr0[0] = tmp00;
-                    outptr1[0] = tmp01;
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp10;
-                        outptr1[1] = tmp11;
-                    }
-#endif
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-
-        float tmp[2][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 16 + jj;
-            const float* r1 = r0 + max_jj;
-            const float* r2 = r0 + max_jj * 2;
-            const float* r3 = r0 + max_jj * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                tmp[0][m] = r0[0] + r1[0] + r2[0];
-                tmp[1][m] = r1[0] - r2[0] + r3[0];
-
-                r0 += max_jj * 4;
-                r1 += max_jj * 4;
-                r2 += max_jj * 4;
-                r3 += max_jj * 4;
-            }
-
-            float* outptr0 = top_blob.channel(i + ii).row(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-
-                float tmp0 = bias0 + r0 + r1 + r2;
-                float tmp1 = bias0 + r1 - r2 + r3;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = tmp0;
-                    if (tj * 2 + 1 < outw) outptr0[1] = tmp1;
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, const Mat& bias, int nT, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 2n+2, winograd F(2,3)
-    int w_tiles = (outw + 1) / 2;
-    int h_tiles = (outh + 1) / 2;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 16;
-
-    // NCNN_LOGE("conv3x3s1_winograd23 %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    conv3x3s1_winograd_get_optimal_tile_mnk(M, N, K, B, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 4u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 4u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd23_transform_input_tile(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 4u, opt.workspace_allocator);
-
-        #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd23_transform_input_tile(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                conv3x3s1_winograd_gemm_transB_packed_tile(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk, opt.use_a53_a55_optimized_kernel);
-            }
-
-            // transform output
-            conv3x3s1_winograd23_transform_output_tile(top_tile, top_blob, bias, i, max_ii, j, max_jj);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_kernel_tile(const Mat& kernel, Mat& A, int inch, int i, int max_ii, int k, int max_kk)
-{
-    float* ptmp = A;
-
-    int ii = 0;
-    for (; ii < max_ii; ii++)
-    {
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            const float sq2 = 1.41421356237f;
-            // const float ktm[6][3] = {
-            //     {1.0f, 0.0f, 0.0f},
-            //     {-2.0f / 3, -sq2 / 3, -1.0f / 3},
-            //     {-2.0f / 3, sq2 / 3, -1.0f / 3},
-            //     {1.0f / 6, sq2 / 6, 1.0f / 3},
-            //     {1.0f / 6, -sq2 / 6, 1.0f / 3},
-            //     {0.0f, 0.0f, 1.0f}
-            // };
-            const float ktm0 = 2.0f / 3;
-            const float ktm1 = sq2 / 3;
-            const float ktm2 = 1.0f / 3;
-            const float ktm3 = 1.0f / 6;
-            const float ktm4 = sq2 / 6;
-
-            float tmp[6][3];
-
-            const float* k0 = (const float*)kernel + (i + ii) * inch * 9 + (k + kk) * 9;
-
-            for (int m = 0; m < 3; m++)
-            {
-                float r0 = k0[0];
-                float r1 = k0[1];
-                float r2 = k0[2];
-
-                tmp[0][m] = r0;
-                tmp[1][m] = -r0 * ktm0 - r1 * ktm1 - r2 * ktm2;
-                tmp[2][m] = -r0 * ktm0 + r1 * ktm1 - r2 * ktm2;
-                tmp[3][m] = r0 * ktm3 + r1 * ktm4 + r2 * ktm2;
-                tmp[4][m] = r0 * ktm3 - r1 * ktm4 + r2 * ktm2;
-                tmp[5][m] = r2;
-
-                k0 += 3;
-            }
-
-            for (int m = 0; m < 6; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-
-                float z0 = r0;
-                float z1 = -r0 * ktm0 - r1 * ktm1 - r2 * ktm2;
-                float z2 = -r0 * ktm0 + r1 * ktm1 - r2 * ktm2;
-                float z3 = r0 * ktm3 + r1 * ktm4 + r2 * ktm2;
-                float z4 = r0 * ktm3 - r1 * ktm4 + r2 * ktm2;
-                float z5 = r2;
-
-                ptmp[0] = z0;
-                ptmp[1] = z1;
-                ptmp[2] = z2;
-                ptmp[3] = z3;
-                ptmp[4] = z4;
-                ptmp[5] = z5;
-                ptmp += 6;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_kernel(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt)
-{
-    const int M = outch;
-    const int K = inch;
-    const int B = 36;
-
-    int TILE_M, TILE_N, TILE_K;
-    conv3x3s1_winograd_get_optimal_tile_mnk(M, 0, K, B, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    Mat A_tileX(B * TILE_M * TILE_K, 1, opt.num_threads);
-
-    AT.create(TILE_K * TILE_M, B, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat A_tile = A_tileX.channel(get_omp_thread_num());
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_ii = std::min((M - i), TILE_M);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            conv3x3s1_winograd43_transform_kernel_tile(kernel, A_tile, inch, i, max_ii, k, max_kk);
-
-            Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-            conv3x3s1_winograd_pack_A_tile(A_tile, AT_tile, B, max_ii, max_kk);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_input_tile(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    const float sq2 = 1.41421356237;
-    const float sq2_d2 = 1.41421356237 / 2;
-
-    // const float itm[6][6] = {
-    //     {1.0f,  0.0f,  -2.5f,  0.0f,  1.0f, 0.0f},
-    //     {0.0f, -sq2,   -2.0f,  sq2/2, 1.0f, 0.0f},
-    //     {0.0f,  sq2,   -2.0f, -sq2/2, 1.0f, 0.0f},
-    //     {0.0f, -sq2/2, -0.5f,  sq2,   1.0f, 0.0f},
-    //     {0.0f,  sq2/2, -0.5f, -sq2,   1.0f, 0.0f},
-    //     {0.0f,  1.0f,   0.0f,  -2.5f, 0.0f, 1.0f}
-    // };
-
-    // 0 =  r00 + r04 - 2.5f * r02
-    // 1 = -(sq2 * r01 - sq2_d2 * r03) + (r04 - 2 * r02)
-    // 2 =  (sq2 * r01 - sq2_d2 * r03) + (r04 - 2 * r02)
-    // 3 =  (sq2 * r03 - sq2_d2 * r01) + (r04 - 0.5f * r02)
-    // 4 = -(sq2 * r03 - sq2_d2 * r01) + (r04 - 0.5f * r02)
-    // 5 =  r01 + r05 - 2.5f * r03
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w + 1) / 4;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-#if __ARM_NEON
-#if __aarch64__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[6][6][8];
-
-        const float coeffs[4] = {sq2, -sq2_d2, -2.f, -0.5f};
-        float32x4_t _coeffs = vld1q_f32(coeffs);
-        float32x4_t _vm2_5 = vdupq_n_f32(-2.5f);
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel((k + kk) / elempack).row(ti * 4) + (tj * 4) * elempack;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float32x4_t _r00 = vdupq_n_f32(0.f);
-                float32x4_t _r01 = vdupq_n_f32(0.f);
-                float32x4_t _r10 = vdupq_n_f32(0.f);
-                float32x4_t _r11 = vdupq_n_f32(0.f);
-                float32x4_t _r20 = vdupq_n_f32(0.f);
-                float32x4_t _r21 = vdupq_n_f32(0.f);
-                float32x4_t _r30 = vdupq_n_f32(0.f);
-                float32x4_t _r31 = vdupq_n_f32(0.f);
-                float32x4_t _r40 = vdupq_n_f32(0.f);
-                float32x4_t _r41 = vdupq_n_f32(0.f);
-                float32x4_t _r50 = vdupq_n_f32(0.f);
-                float32x4_t _r51 = vdupq_n_f32(0.f);
-
-                if (ti * 4 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-
-                        _r00 = vld1q_f32(r0);
-                        _r01 = vld1q_f32(r1);
-                        if (tj * 4 + 1 < w)
-                        {
-                            _r10 = vld1q_f32(r0 + 4);
-                            _r11 = vld1q_f32(r1 + 4);
-                        }
-                        if (tj * 4 + 2 < w)
-                        {
-                            _r20 = vld1q_f32(r0 + 8);
-                            _r21 = vld1q_f32(r1 + 8);
-                        }
-                        if (tj * 4 + 3 < w)
-                        {
-                            _r30 = vld1q_f32(r0 + 12);
-                            _r31 = vld1q_f32(r1 + 12);
-                        }
-                        if (tj * 4 + 4 < w)
-                        {
-                            _r40 = vld1q_f32(r0 + 16);
-                            _r41 = vld1q_f32(r1 + 16);
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            _r50 = vld1q_f32(r0 + 20);
-                            _r51 = vld1q_f32(r1 + 20);
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-                        const float* r4 = r0 + N * 4;
-                        const float* r5 = r0 + N * 5;
-                        const float* r6 = r0 + N * 6;
-                        const float* r7 = r0 + N * 7;
-
-                        float32x4_t _t0 = vld1q_f32(r0);
-                        float32x4_t _t1 = vld1q_f32(r1);
-                        float32x4_t _t2 = vld1q_f32(r2);
-                        float32x4_t _t3 = vld1q_f32(r3);
-                        float32x4_t _t4 = vld1q_f32(r4);
-                        float32x4_t _t5 = vld1q_f32(r5);
-                        float32x4_t _t6 = vld1q_f32(r6);
-                        float32x4_t _t7 = vld1q_f32(r7);
-
-                        transpose4x4_ps(_t0, _t1, _t2, _t3);
-                        transpose4x4_ps(_t4, _t5, _t6, _t7);
-
-                        _r00 = _t0;
-                        _r01 = _t4;
-                        if (tj * 4 + 1 < w)
-                        {
-                            _r10 = _t1;
-                            _r11 = _t5;
-                        }
-                        if (tj * 4 + 2 < w)
-                        {
-                            _r20 = _t2;
-                            _r21 = _t6;
-                        }
-                        if (tj * 4 + 3 < w)
-                        {
-                            _r30 = _t3;
-                            _r31 = _t7;
-                        }
-                        if (tj * 4 + 4 < w)
-                        {
-                            float tmp[8] = {r0[4], r1[4], r2[4], r3[4], r4[4], r5[4], r6[4], r7[4]};
-                            _r40 = vld1q_f32(tmp);
-                            _r41 = vld1q_f32(tmp + 4);
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            float tmp[8] = {r0[5], r1[5], r2[5], r3[5], r4[5], r5[5], r6[5], r7[5]};
-                            _r50 = vld1q_f32(tmp);
-                            _r51 = vld1q_f32(tmp + 4);
-                        }
-                    }
-                }
-
-                float32x4_t _tmp12a0 = vfmaq_laneq_f32(vmulq_laneq_f32(_r10, _coeffs, 0), _r30, _coeffs, 1);
-                float32x4_t _tmp12a1 = vfmaq_laneq_f32(vmulq_laneq_f32(_r11, _coeffs, 0), _r31, _coeffs, 1);
-                float32x4_t _tmp12b0 = vfmaq_laneq_f32(_r40, _r20, _coeffs, 2);
-                float32x4_t _tmp12b1 = vfmaq_laneq_f32(_r41, _r21, _coeffs, 2);
-                float32x4_t _tmp34a0 = vfmaq_laneq_f32(vmulq_laneq_f32(_r30, _coeffs, 0), _r10, _coeffs, 1);
-                float32x4_t _tmp34a1 = vfmaq_laneq_f32(vmulq_laneq_f32(_r31, _coeffs, 0), _r11, _coeffs, 1);
-                float32x4_t _tmp34b0 = vfmaq_laneq_f32(_r40, _r20, _coeffs, 3);
-                float32x4_t _tmp34b1 = vfmaq_laneq_f32(_r41, _r21, _coeffs, 3);
-
-                float32x4_t _tmp00 = vfmaq_f32(vaddq_f32(_r00, _r40), _r20, _vm2_5);
-                float32x4_t _tmp01 = vfmaq_f32(vaddq_f32(_r01, _r41), _r21, _vm2_5);
-                float32x4_t _tmp10 = vsubq_f32(_tmp12b0, _tmp12a0);
-                float32x4_t _tmp11 = vsubq_f32(_tmp12b1, _tmp12a1);
-                float32x4_t _tmp20 = vaddq_f32(_tmp12b0, _tmp12a0);
-                float32x4_t _tmp21 = vaddq_f32(_tmp12b1, _tmp12a1);
-                float32x4_t _tmp30 = vaddq_f32(_tmp34b0, _tmp34a0);
-                float32x4_t _tmp31 = vaddq_f32(_tmp34b1, _tmp34a1);
-                float32x4_t _tmp40 = vsubq_f32(_tmp34b0, _tmp34a0);
-                float32x4_t _tmp41 = vsubq_f32(_tmp34b1, _tmp34a1);
-                float32x4_t _tmp50 = vfmaq_f32(vaddq_f32(_r10, _r50), _r30, _vm2_5);
-                float32x4_t _tmp51 = vfmaq_f32(vaddq_f32(_r11, _r51), _r31, _vm2_5);
-
-                vst1q_f32(tmp[0][m], _tmp00);
-                vst1q_f32(tmp[0][m] + 4, _tmp01);
-                vst1q_f32(tmp[1][m], _tmp10);
-                vst1q_f32(tmp[1][m] + 4, _tmp11);
-                vst1q_f32(tmp[2][m], _tmp20);
-                vst1q_f32(tmp[2][m] + 4, _tmp21);
-                vst1q_f32(tmp[3][m], _tmp30);
-                vst1q_f32(tmp[3][m] + 4, _tmp31);
-                vst1q_f32(tmp[4][m], _tmp40);
-                vst1q_f32(tmp[4][m] + 4, _tmp41);
-                vst1q_f32(tmp[5][m], _tmp50);
-                vst1q_f32(tmp[5][m] + 4, _tmp51);
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 36 + jj * 8;
-            float* p1 = p0 + max_jj * 8;
-            float* p2 = p0 + max_jj * 8 * 2;
-            float* p3 = p0 + max_jj * 8 * 3;
-            float* p4 = p0 + max_jj * 8 * 4;
-            float* p5 = p0 + max_jj * 8 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float32x4_t _r00 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r01 = vld1q_f32(tmp[m][0] + 4);
-                float32x4_t _r10 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r11 = vld1q_f32(tmp[m][1] + 4);
-                float32x4_t _r20 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r21 = vld1q_f32(tmp[m][2] + 4);
-                float32x4_t _r30 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r31 = vld1q_f32(tmp[m][3] + 4);
-                float32x4_t _r40 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r41 = vld1q_f32(tmp[m][4] + 4);
-                float32x4_t _r50 = vld1q_f32(tmp[m][5]);
-                float32x4_t _r51 = vld1q_f32(tmp[m][5] + 4);
-
-                float32x4_t _tmp12a0 = vfmaq_laneq_f32(vmulq_laneq_f32(_r10, _coeffs, 0), _r30, _coeffs, 1);
-                float32x4_t _tmp12a1 = vfmaq_laneq_f32(vmulq_laneq_f32(_r11, _coeffs, 0), _r31, _coeffs, 1);
-                float32x4_t _tmp12b0 = vfmaq_laneq_f32(_r40, _r20, _coeffs, 2);
-                float32x4_t _tmp12b1 = vfmaq_laneq_f32(_r41, _r21, _coeffs, 2);
-                float32x4_t _tmp34a0 = vfmaq_laneq_f32(vmulq_laneq_f32(_r30, _coeffs, 0), _r10, _coeffs, 1);
-                float32x4_t _tmp34a1 = vfmaq_laneq_f32(vmulq_laneq_f32(_r31, _coeffs, 0), _r11, _coeffs, 1);
-                float32x4_t _tmp34b0 = vfmaq_laneq_f32(_r40, _r20, _coeffs, 3);
-                float32x4_t _tmp34b1 = vfmaq_laneq_f32(_r41, _r21, _coeffs, 3);
-
-                float32x4_t _tmp00 = vfmaq_f32(vaddq_f32(_r00, _r40), _r20, _vm2_5);
-                float32x4_t _tmp01 = vfmaq_f32(vaddq_f32(_r01, _r41), _r21, _vm2_5);
-                float32x4_t _tmp10 = vsubq_f32(_tmp12b0, _tmp12a0);
-                float32x4_t _tmp11 = vsubq_f32(_tmp12b1, _tmp12a1);
-                float32x4_t _tmp20 = vaddq_f32(_tmp12b0, _tmp12a0);
-                float32x4_t _tmp21 = vaddq_f32(_tmp12b1, _tmp12a1);
-                float32x4_t _tmp30 = vaddq_f32(_tmp34b0, _tmp34a0);
-                float32x4_t _tmp31 = vaddq_f32(_tmp34b1, _tmp34a1);
-                float32x4_t _tmp40 = vsubq_f32(_tmp34b0, _tmp34a0);
-                float32x4_t _tmp41 = vsubq_f32(_tmp34b1, _tmp34a1);
-                float32x4_t _tmp50 = vfmaq_f32(vaddq_f32(_r10, _r50), _r30, _vm2_5);
-                float32x4_t _tmp51 = vfmaq_f32(vaddq_f32(_r11, _r51), _r31, _vm2_5);
-
-                vst1q_f32(p0, _tmp00);
-                vst1q_f32(p0 + 4, _tmp01);
-                vst1q_f32(p1, _tmp10);
-                vst1q_f32(p1 + 4, _tmp11);
-                vst1q_f32(p2, _tmp20);
-                vst1q_f32(p2 + 4, _tmp21);
-                vst1q_f32(p3, _tmp30);
-                vst1q_f32(p3 + 4, _tmp31);
-                vst1q_f32(p4, _tmp40);
-                vst1q_f32(p4 + 4, _tmp41);
-                vst1q_f32(p5, _tmp50);
-                vst1q_f32(p5 + 4, _tmp51);
-
-                p0 += max_jj * 6 * 8;
-                p1 += max_jj * 6 * 8;
-                p2 += max_jj * 6 * 8;
-                p3 += max_jj * 6 * 8;
-                p4 += max_jj * 6 * 8;
-                p5 += max_jj * 6 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-#else // __aarch64__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-    #pragma omp parallel for num_threads(nT)
-#endif // __aarch64__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 4;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[6][6][4];
-
-        const float coeffs[4] = {sq2, -sq2_d2, -2.f, -0.5f};
-        float32x4_t _coeffs = vld1q_f32(coeffs);
-        float32x4_t _vm2_5 = vdupq_n_f32(-2.5f);
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel((k + kk) / elempack).row(ti * 4) + (tj * 4) * elempack;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float32x4_t _r0 = vdupq_n_f32(0.f);
-                float32x4_t _r1 = vdupq_n_f32(0.f);
-                float32x4_t _r2 = vdupq_n_f32(0.f);
-                float32x4_t _r3 = vdupq_n_f32(0.f);
-                float32x4_t _r4 = vdupq_n_f32(0.f);
-                float32x4_t _r5 = vdupq_n_f32(0.f);
-
-                if (ti * 4 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1q_f32(r0);
-                        if (tj * 4 + 1 < w) _r1 = vld1q_f32(r0 + 4);
-                        if (tj * 4 + 2 < w) _r2 = vld1q_f32(r0 + 8);
-                        if (tj * 4 + 3 < w) _r3 = vld1q_f32(r0 + 12);
-                        if (tj * 4 + 4 < w) _r4 = vld1q_f32(r0 + 16);
-                        if (tj * 4 + 5 < w) _r5 = vld1q_f32(r0 + 20);
-                    }
-                    if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-
-                        float32x4_t _t0 = vld1q_f32(r0);
-                        float32x4_t _t1 = vld1q_f32(r1);
-                        float32x4_t _t2 = vld1q_f32(r2);
-                        float32x4_t _t3 = vld1q_f32(r3);
-
-                        transpose4x4_ps(_t0, _t1, _t2, _t3);
-
-                        _r0 = _t0;
-                        if (tj * 4 + 1 < w) _r1 = _t1;
-                        if (tj * 4 + 2 < w) _r2 = _t2;
-                        if (tj * 4 + 3 < w) _r3 = _t3;
-                        if (tj * 4 + 4 < w)
-                        {
-                            float tmp[4] = {r0[4], r1[4], r2[4], r3[4]};
-                            _r4 = vld1q_f32(tmp);
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            float tmp[4] = {r0[5], r1[5], r2[5], r3[5]};
-                            _r5 = vld1q_f32(tmp);
-                        }
-                    }
-                }
-
-#if __aarch64__
-                float32x4_t _tmp12a = vfmaq_laneq_f32(vmulq_laneq_f32(_r1, _coeffs, 0), _r3, _coeffs, 1);
-                float32x4_t _tmp12b = vfmaq_laneq_f32(_r4, _r2, _coeffs, 2);
-                float32x4_t _tmp34a = vfmaq_laneq_f32(vmulq_laneq_f32(_r3, _coeffs, 0), _r1, _coeffs, 1);
-                float32x4_t _tmp34b = vfmaq_laneq_f32(_r4, _r2, _coeffs, 3);
-#else
-                float32x4_t _tmp12a = vmlaq_lane_f32(vmulq_lane_f32(_r1, vget_low_f32(_coeffs), 0), _r3, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp12b = vmlaq_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 0);
-                float32x4_t _tmp34a = vmlaq_lane_f32(vmulq_lane_f32(_r3, vget_low_f32(_coeffs), 0), _r1, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp34b = vmlaq_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 1);
-#endif
-
-#if __aarch64__
-                float32x4_t _tmp0 = vfmaq_f32(vaddq_f32(_r0, _r4), _r2, _vm2_5);
-#else
-                float32x4_t _tmp0 = vmlaq_f32(vaddq_f32(_r0, _r4), _r2, _vm2_5);
-#endif
-                float32x4_t _tmp1 = vsubq_f32(_tmp12b, _tmp12a);
-                float32x4_t _tmp2 = vaddq_f32(_tmp12b, _tmp12a);
-                float32x4_t _tmp3 = vaddq_f32(_tmp34b, _tmp34a);
-                float32x4_t _tmp4 = vsubq_f32(_tmp34b, _tmp34a);
-#if __aarch64__
-                float32x4_t _tmp5 = vfmaq_f32(vaddq_f32(_r1, _r5), _r3, _vm2_5);
-#else
-                float32x4_t _tmp5 = vmlaq_f32(vaddq_f32(_r1, _r5), _r3, _vm2_5);
-#endif
-
-                vst1q_f32(tmp[0][m], _tmp0);
-                vst1q_f32(tmp[1][m], _tmp1);
-                vst1q_f32(tmp[2][m], _tmp2);
-                vst1q_f32(tmp[3][m], _tmp3);
-                vst1q_f32(tmp[4][m], _tmp4);
-                vst1q_f32(tmp[5][m], _tmp5);
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 36 + jj * 4;
-            float* p1 = p0 + max_jj * 4;
-            float* p2 = p0 + max_jj * 4 * 2;
-            float* p3 = p0 + max_jj * 4 * 3;
-            float* p4 = p0 + max_jj * 4 * 4;
-            float* p5 = p0 + max_jj * 4 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float32x4_t _r0 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r1 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r2 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r3 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r4 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r5 = vld1q_f32(tmp[m][5]);
-
-#if __aarch64__
-                float32x4_t _tmp12a = vfmaq_laneq_f32(vmulq_laneq_f32(_r1, _coeffs, 0), _r3, _coeffs, 1);
-                float32x4_t _tmp12b = vfmaq_laneq_f32(_r4, _r2, _coeffs, 2);
-                float32x4_t _tmp34a = vfmaq_laneq_f32(vmulq_laneq_f32(_r3, _coeffs, 0), _r1, _coeffs, 1);
-                float32x4_t _tmp34b = vfmaq_laneq_f32(_r4, _r2, _coeffs, 3);
-#else
-                float32x4_t _tmp12a = vmlaq_lane_f32(vmulq_lane_f32(_r1, vget_low_f32(_coeffs), 0), _r3, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp12b = vmlaq_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 0);
-                float32x4_t _tmp34a = vmlaq_lane_f32(vmulq_lane_f32(_r3, vget_low_f32(_coeffs), 0), _r1, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp34b = vmlaq_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 1);
-#endif
-
-#if __aarch64__
-                float32x4_t _tmp0 = vfmaq_f32(vaddq_f32(_r0, _r4), _r2, _vm2_5);
-#else
-                float32x4_t _tmp0 = vmlaq_f32(vaddq_f32(_r0, _r4), _r2, _vm2_5);
-#endif
-                float32x4_t _tmp1 = vsubq_f32(_tmp12b, _tmp12a);
-                float32x4_t _tmp2 = vaddq_f32(_tmp12b, _tmp12a);
-                float32x4_t _tmp3 = vaddq_f32(_tmp34b, _tmp34a);
-                float32x4_t _tmp4 = vsubq_f32(_tmp34b, _tmp34a);
-#if __aarch64__
-                float32x4_t _tmp5 = vfmaq_f32(vaddq_f32(_r1, _r5), _r3, _vm2_5);
-#else
-                float32x4_t _tmp5 = vmlaq_f32(vaddq_f32(_r1, _r5), _r3, _vm2_5);
-#endif
-
-                vst1q_f32(p0, _tmp0);
-                vst1q_f32(p1, _tmp1);
-                vst1q_f32(p2, _tmp2);
-                vst1q_f32(p3, _tmp3);
-                vst1q_f32(p4, _tmp4);
-                vst1q_f32(p5, _tmp5);
-
-                p0 += max_jj * 6 * 4;
-                p1 += max_jj * 6 * 4;
-                p2 += max_jj * 6 * 4;
-                p3 += max_jj * 6 * 4;
-                p4 += max_jj * 6 * 4;
-                p5 += max_jj * 6 * 4;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 4;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-#else // __ARM_NEON
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    #pragma omp parallel for num_threads(nT)
-#endif // __ARM_NEON
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        float tmp[6][6][2];
-
-#if __ARM_NEON
-        const float coeffs[4] = {sq2, -sq2_d2, -2.f, -0.5f};
-        float32x4_t _coeffs = vld1q_f32(coeffs);
-        float32x2_t _vm2_5 = vdup_n_f32(-2.5f);
-#endif
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel(k + kk).row(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 6; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vdup_n_f32(0.f);
-                float32x2_t _r1 = vdup_n_f32(0.f);
-                float32x2_t _r2 = vdup_n_f32(0.f);
-                float32x2_t _r3 = vdup_n_f32(0.f);
-                float32x2_t _r4 = vdup_n_f32(0.f);
-                float32x2_t _r5 = vdup_n_f32(0.f);
-#else
-                float r00 = 0.f;
-                float r01 = 0.f;
-                float r10 = 0.f;
-                float r11 = 0.f;
-                float r20 = 0.f;
-                float r21 = 0.f;
-                float r30 = 0.f;
-                float r31 = 0.f;
-                float r40 = 0.f;
-                float r41 = 0.f;
-                float r50 = 0.f;
-                float r51 = 0.f;
-#endif
-
-                if (ti * 4 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-
-#if __ARM_NEON
-                        float32x4_t _t0 = vld1q_f32(r0);
-                        float32x4_t _t1 = vld1q_f32(r1);
-                        float32x4x2_t _t01 = vzipq_f32(_t0, _t1);
-
-                        _r0 = vget_low_f32(_t01.val[0]);
-                        if (tj * 4 + 1 < w) _r1 = vget_high_f32(_t01.val[0]);
-                        if (tj * 4 + 2 < w) _r2 = vget_low_f32(_t01.val[1]);
-                        if (tj * 4 + 3 < w) _r3 = vget_high_f32(_t01.val[1]);
-                        if (tj * 4 + 4 < w)
-                        {
-                            float tmp[2] = {r0[4], r1[4]};
-                            _r4 = vld1_f32(tmp);
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            float tmp[2] = {r0[5], r1[5]};
-                            _r5 = vld1_f32(tmp);
-                        }
-#else
-                        r00 = r0[0];
-                        r01 = r1[0];
-                        if (tj * 4 + 1 < w)
-                        {
-                            r10 = r0[1];
-                            r11 = r1[1];
-                        }
-                        if (tj * 4 + 2 < w)
-                        {
-                            r20 = r0[2];
-                            r21 = r1[2];
-                        }
-                        if (tj * 4 + 3 < w)
-                        {
-                            r30 = r0[3];
-                            r31 = r1[3];
-                        }
-                        if (tj * 4 + 4 < w)
-                        {
-                            r40 = r0[4];
-                            r41 = r1[4];
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            r50 = r0[5];
-                            r51 = r1[5];
-                        }
-#endif
-                    }
-                }
-
-#if __ARM_NEON
-#if __aarch64__
-                float32x2_t _tmp12a = vfma_laneq_f32(vmul_laneq_f32(_r1, _coeffs, 0), _r3, _coeffs, 1);
-                float32x2_t _tmp12b = vfma_laneq_f32(_r4, _r2, _coeffs, 2);
-                float32x2_t _tmp34a = vfma_laneq_f32(vmul_laneq_f32(_r3, _coeffs, 0), _r1, _coeffs, 1);
-                float32x2_t _tmp34b = vfma_laneq_f32(_r4, _r2, _coeffs, 3);
-#else
-                float32x2_t _tmp12a = vmla_lane_f32(vmul_lane_f32(_r1, vget_low_f32(_coeffs), 0), _r3, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp12b = vmla_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 0);
-                float32x2_t _tmp34a = vmla_lane_f32(vmul_lane_f32(_r3, vget_low_f32(_coeffs), 0), _r1, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp34b = vmla_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 1);
-#endif
-
-#if __aarch64__
-                float32x2_t _tmp0 = vfma_f32(vadd_f32(_r0, _r4), _r2, _vm2_5);
-#else
-                float32x2_t _tmp0 = vmla_f32(vadd_f32(_r0, _r4), _r2, _vm2_5);
-#endif
-                float32x2_t _tmp1 = vsub_f32(_tmp12b, _tmp12a);
-                float32x2_t _tmp2 = vadd_f32(_tmp12b, _tmp12a);
-                float32x2_t _tmp3 = vadd_f32(_tmp34b, _tmp34a);
-                float32x2_t _tmp4 = vsub_f32(_tmp34b, _tmp34a);
-#if __aarch64__
-                float32x2_t _tmp5 = vfma_f32(vadd_f32(_r1, _r5), _r3, _vm2_5);
-#else
-                float32x2_t _tmp5 = vmla_f32(vadd_f32(_r1, _r5), _r3, _vm2_5);
-#endif
-
-                vst1_f32(tmp[0][m], _tmp0);
-                vst1_f32(tmp[1][m], _tmp1);
-                vst1_f32(tmp[2][m], _tmp2);
-                vst1_f32(tmp[3][m], _tmp3);
-                vst1_f32(tmp[4][m], _tmp4);
-                vst1_f32(tmp[5][m], _tmp5);
-#else
-                float tmp12a0 = sq2 * r10 - sq2_d2 * r30;
-                float tmp12a1 = sq2 * r11 - sq2_d2 * r31;
-                float tmp12b0 = r40 - 2 * r20;
-                float tmp12b1 = r41 - 2 * r21;
-                float tmp34a0 = sq2 * r30 - sq2_d2 * r10;
-                float tmp34a1 = sq2 * r31 - sq2_d2 * r11;
-                float tmp34b0 = r40 - 0.5f * r20;
-                float tmp34b1 = r41 - 0.5f * r21;
-
-                tmp[0][m][0] = r00 + r40 - 2.5f * r20;
-                tmp[0][m][1] = r01 + r41 - 2.5f * r21;
-                tmp[1][m][0] = tmp12b0 - tmp12a0;
-                tmp[1][m][1] = tmp12b1 - tmp12a1;
-                tmp[2][m][0] = tmp12b0 + tmp12a0;
-                tmp[2][m][1] = tmp12b1 + tmp12a1;
-                tmp[3][m][0] = tmp34b0 + tmp34a0;
-                tmp[3][m][1] = tmp34b1 + tmp34a1;
-                tmp[4][m][0] = tmp34b0 - tmp34a0;
-                tmp[4][m][1] = tmp34b1 - tmp34a1;
-                tmp[5][m][0] = r10 + r50 - 2.5f * r30;
-                tmp[5][m][1] = r11 + r51 - 2.5f * r31;
-#endif
-
-                r0 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 36 + jj * 2;
-            float* p1 = p0 + max_jj * 2;
-            float* p2 = p0 + max_jj * 2 * 2;
-            float* p3 = p0 + max_jj * 2 * 3;
-            float* p4 = p0 + max_jj * 2 * 4;
-            float* p5 = p0 + max_jj * 2 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(tmp[m][0]);
-                float32x2_t _r1 = vld1_f32(tmp[m][1]);
-                float32x2_t _r2 = vld1_f32(tmp[m][2]);
-                float32x2_t _r3 = vld1_f32(tmp[m][3]);
-                float32x2_t _r4 = vld1_f32(tmp[m][4]);
-                float32x2_t _r5 = vld1_f32(tmp[m][5]);
-
-#if __aarch64__
-                float32x2_t _tmp12a = vfma_laneq_f32(vmul_laneq_f32(_r1, _coeffs, 0), _r3, _coeffs, 1);
-                float32x2_t _tmp12b = vfma_laneq_f32(_r4, _r2, _coeffs, 2);
-                float32x2_t _tmp34a = vfma_laneq_f32(vmul_laneq_f32(_r3, _coeffs, 0), _r1, _coeffs, 1);
-                float32x2_t _tmp34b = vfma_laneq_f32(_r4, _r2, _coeffs, 3);
-#else
-                float32x2_t _tmp12a = vmla_lane_f32(vmul_lane_f32(_r1, vget_low_f32(_coeffs), 0), _r3, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp12b = vmla_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 0);
-                float32x2_t _tmp34a = vmla_lane_f32(vmul_lane_f32(_r3, vget_low_f32(_coeffs), 0), _r1, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp34b = vmla_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 1);
-#endif
-
-#if __aarch64__
-                float32x2_t _tmp0 = vfma_f32(vadd_f32(_r0, _r4), _r2, _vm2_5);
-#else
-                float32x2_t _tmp0 = vmla_f32(vadd_f32(_r0, _r4), _r2, _vm2_5);
-#endif
-                float32x2_t _tmp1 = vsub_f32(_tmp12b, _tmp12a);
-                float32x2_t _tmp2 = vadd_f32(_tmp12b, _tmp12a);
-                float32x2_t _tmp3 = vadd_f32(_tmp34b, _tmp34a);
-                float32x2_t _tmp4 = vsub_f32(_tmp34b, _tmp34a);
-#if __aarch64__
-                float32x2_t _tmp5 = vfma_f32(vadd_f32(_r1, _r5), _r3, _vm2_5);
-#else
-                float32x2_t _tmp5 = vmla_f32(vadd_f32(_r1, _r5), _r3, _vm2_5);
-#endif
-
-                vst1_f32(p0, _tmp0);
-                vst1_f32(p1, _tmp1);
-                vst1_f32(p2, _tmp2);
-                vst1_f32(p3, _tmp3);
-                vst1_f32(p4, _tmp4);
-                vst1_f32(p5, _tmp5);
-#else
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-                float r40 = tmp[m][4][0];
-                float r41 = tmp[m][4][1];
-                float r50 = tmp[m][5][0];
-                float r51 = tmp[m][5][1];
-
-                float tmp12a0 = sq2 * r10 - sq2_d2 * r30;
-                float tmp12a1 = sq2 * r11 - sq2_d2 * r31;
-                float tmp12b0 = r40 - 2 * r20;
-                float tmp12b1 = r41 - 2 * r21;
-                float tmp34a0 = sq2 * r30 - sq2_d2 * r10;
-                float tmp34a1 = sq2 * r31 - sq2_d2 * r11;
-                float tmp34b0 = r40 - 0.5f * r20;
-                float tmp34b1 = r41 - 0.5f * r21;
-
-                p0[0] = r00 + r40 - 2.5f * r20;
-                p0[1] = r01 + r41 - 2.5f * r21;
-                p1[0] = tmp12b0 - tmp12a0;
-                p1[1] = tmp12b1 - tmp12a1;
-                p2[0] = tmp12b0 + tmp12a0;
-                p2[1] = tmp12b1 + tmp12a1;
-                p3[0] = tmp34b0 + tmp34a0;
-                p3[1] = tmp34b1 + tmp34a1;
-                p4[0] = tmp34b0 - tmp34a0;
-                p4[1] = tmp34b1 - tmp34a1;
-                p5[0] = r10 + r50 - 2.5f * r30;
-                p5[1] = r11 + r51 - 2.5f * r31;
-#endif
-
-                p0 += max_jj * 6 * 2;
-                p1 += max_jj * 6 * 2;
-                p2 += max_jj * 6 * 2;
-                p3 += max_jj * 6 * 2;
-                p4 += max_jj * 6 * 2;
-                p5 += max_jj * 6 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        float tmp[6][6];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0123 = bottom_blob.channel(k + kk).row(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 6; m++)
-            {
-                float r0 = 0.f;
-                float r1 = 0.f;
-                float r2 = 0.f;
-                float r3 = 0.f;
-                float r4 = 0.f;
-                float r5 = 0.f;
-
-                if (ti * 4 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = r0123[0];
-                        if (tj * 4 + 1 < w) r1 = r0123[1];
-                        if (tj * 4 + 2 < w) r2 = r0123[2];
-                        if (tj * 4 + 3 < w) r3 = r0123[3];
-                        if (tj * 4 + 4 < w) r4 = r0123[4];
-                        if (tj * 4 + 5 < w) r5 = r0123[5];
-                    }
-                }
-
-                float tmp12a = sq2 * r1 - sq2_d2 * r3;
-                float tmp12b = r4 - 2 * r2;
-                float tmp34a = sq2 * r3 - sq2_d2 * r1;
-                float tmp34b = r4 - 0.5f * r2;
-
-                tmp[0][m] = r0 + r4 - 2.5f * r2;
-                tmp[1][m] = tmp12b - tmp12a;
-                tmp[2][m] = tmp12b + tmp12a;
-                tmp[3][m] = tmp34b + tmp34a;
-                tmp[4][m] = tmp34b - tmp34a;
-                tmp[5][m] = r1 + r5 - 2.5f * r3;
-
-                r0123 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 36 + jj;
-            float* p1 = p0 + max_jj;
-            float* p2 = p0 + max_jj * 2;
-            float* p3 = p0 + max_jj * 3;
-            float* p4 = p0 + max_jj * 4;
-            float* p5 = p0 + max_jj * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-                float r4 = tmp[m][4];
-                float r5 = tmp[m][5];
-
-                float tmp12a = sq2 * r1 - sq2_d2 * r3;
-                float tmp12b = r4 - 2 * r2;
-                float tmp34a = sq2 * r3 - sq2_d2 * r1;
-                float tmp34b = r4 - 0.5f * r2;
-
-                p0[0] = r0 + r4 - 2.5f * r2;
-                p1[0] = tmp12b - tmp12a;
-                p2[0] = tmp12b + tmp12a;
-                p3[0] = tmp34b + tmp34a;
-                p4[0] = tmp34b - tmp34a;
-                p5[0] = r1 + r5 - 2.5f * r3;
-
-                p0 += max_jj * 6;
-                p1 += max_jj * 6;
-                p2 += max_jj * 6;
-                p3 += max_jj * 6;
-                p4 += max_jj * 6;
-                p5 += max_jj * 6;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_output_tile(const Mat& top_tile, Mat& top_blob, const Mat& bias, int i, int max_ii, int j, int max_jj)
-{
-    const float sq2 = 1.41421356237;
-    const float sq2_m2 = 1.41421356237 * 2;
-    const float sq2_d2 = 1.41421356237 / 2;
-    const float sq2_d4 = 1.41421356237 / 4;
-
-    // const float otm[4][6] = {
-    //     {1.0f, 1.0f,   1.0f,  1.0f,  1.0f,   0.0f},
-    //     {0.0f, sq2/2, -sq2/2, sq2,   -sq2,   0.0f},
-    //     {0.0f, 0.5f,   0.5f,  2.0f,  2.0f,   0.0f},
-    //     {0.0f, sq2/4, -sq2/4, sq2*2, -sq2*2, 1.0f}
-    // };
-
-    // 0 = r00 + (r01 + r02) + (r03 + r04)
-    // 1 =       (r01 - r02) * sq2_d2 + (r03 - r04) * sq2
-    // 2 =       (r01 + r02) * 0.5f + (r03 + r04) * 2
-    // 3 = r05 + (r01 - r02) * sq2_d4 + (r03 - r04) * sq2_m2
-
-#if __ARM_NEON
-    const float coeffs[6] = {sq2, sq2_d2, sq2_d4, sq2_m2, 0.5f, 2.f};
-    float32x4_t _coeffs = vld1q_f32(coeffs);
-    float32x2_t _coeffs2 = vld1_f32(coeffs + 4);
-#endif // __ARM_NEON
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 3) / 4;
-
-    const float* biasptr = bias;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        float32x4_t _bias0 = biasptr ? vld1q_f32(biasptr + i + ii) : vdupq_n_f32(0.f);
-        float32x4_t _bias1 = biasptr ? vld1q_f32(biasptr + i + ii + 4) : vdupq_n_f32(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[4][6][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 36 + jj * 8;
-            const float* r1 = r0 + max_jj * 8;
-            const float* r2 = r0 + max_jj * 8 * 2;
-            const float* r3 = r0 + max_jj * 8 * 3;
-            const float* r4 = r0 + max_jj * 8 * 4;
-            const float* r5 = r0 + max_jj * 8 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r30 = vld1q_f32(r3);
-                float32x4_t _r31 = vld1q_f32(r3 + 4);
-                float32x4_t _r40 = vld1q_f32(r4);
-                float32x4_t _r41 = vld1q_f32(r4 + 4);
-                float32x4_t _r50 = vld1q_f32(r5);
-                float32x4_t _r51 = vld1q_f32(r5 + 4);
-
-                float32x4_t _tmp02a0 = vaddq_f32(_r10, _r20);
-                float32x4_t _tmp02a1 = vaddq_f32(_r11, _r21);
-                float32x4_t _tmp02b0 = vaddq_f32(_r30, _r40);
-                float32x4_t _tmp02b1 = vaddq_f32(_r31, _r41);
-                float32x4_t _tmp13a0 = vsubq_f32(_r10, _r20);
-                float32x4_t _tmp13a1 = vsubq_f32(_r11, _r21);
-                float32x4_t _tmp13b0 = vsubq_f32(_r30, _r40);
-                float32x4_t _tmp13b1 = vsubq_f32(_r31, _r41);
-
-                float32x4_t _tmp00 = vaddq_f32(vaddq_f32(_r00, _tmp02a0), _tmp02b0);
-                float32x4_t _tmp01 = vaddq_f32(vaddq_f32(_r01, _tmp02a1), _tmp02b1);
-                float32x4_t _tmp10 = vfmaq_laneq_f32(vmulq_laneq_f32(_tmp13a0, _coeffs, 1), _tmp13b0, _coeffs, 0);
-                float32x4_t _tmp11 = vfmaq_laneq_f32(vmulq_laneq_f32(_tmp13a1, _coeffs, 1), _tmp13b1, _coeffs, 0);
-                float32x4_t _tmp20 = vfmaq_lane_f32(vmulq_lane_f32(_tmp02a0, _coeffs2, 0), _tmp02b0, _coeffs2, 1);
-                float32x4_t _tmp21 = vfmaq_lane_f32(vmulq_lane_f32(_tmp02a1, _coeffs2, 0), _tmp02b1, _coeffs2, 1);
-                float32x4_t _tmp30 = vfmaq_laneq_f32(vfmaq_laneq_f32(_r50, _tmp13a0, _coeffs, 2), _tmp13b0, _coeffs, 3);
-                float32x4_t _tmp31 = vfmaq_laneq_f32(vfmaq_laneq_f32(_r51, _tmp13a1, _coeffs, 2), _tmp13b1, _coeffs, 3);
-
-                vst1q_f32(tmp[0][m], _tmp00);
-                vst1q_f32(tmp[0][m] + 4, _tmp01);
-                vst1q_f32(tmp[1][m], _tmp10);
-                vst1q_f32(tmp[1][m] + 4, _tmp11);
-                vst1q_f32(tmp[2][m], _tmp20);
-                vst1q_f32(tmp[2][m] + 4, _tmp21);
-                vst1q_f32(tmp[3][m], _tmp30);
-                vst1q_f32(tmp[3][m] + 4, _tmp31);
-
-                r0 += max_jj * 6 * 8;
-                r1 += max_jj * 6 * 8;
-                r2 += max_jj * 6 * 8;
-                r3 += max_jj * 6 * 8;
-                r4 += max_jj * 6 * 8;
-                r5 += max_jj * 6 * 8;
-            }
-
-            float* outptr0 = top_blob.channel((i + ii) / out_elempack).row(ti * 4) + (tj * 4) * out_elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                float32x4_t _r00 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r01 = vld1q_f32(tmp[m][0] + 4);
-                float32x4_t _r10 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r11 = vld1q_f32(tmp[m][1] + 4);
-                float32x4_t _r20 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r21 = vld1q_f32(tmp[m][2] + 4);
-                float32x4_t _r30 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r31 = vld1q_f32(tmp[m][3] + 4);
-                float32x4_t _r40 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r41 = vld1q_f32(tmp[m][4] + 4);
-                float32x4_t _r50 = vld1q_f32(tmp[m][5]);
-                float32x4_t _r51 = vld1q_f32(tmp[m][5] + 4);
-
-                float32x4_t _tmp02a0 = vaddq_f32(_r10, _r20);
-                float32x4_t _tmp02a1 = vaddq_f32(_r11, _r21);
-                float32x4_t _tmp02b0 = vaddq_f32(_r30, _r40);
-                float32x4_t _tmp02b1 = vaddq_f32(_r31, _r41);
-                float32x4_t _tmp13a0 = vsubq_f32(_r10, _r20);
-                float32x4_t _tmp13a1 = vsubq_f32(_r11, _r21);
-                float32x4_t _tmp13b0 = vsubq_f32(_r30, _r40);
-                float32x4_t _tmp13b1 = vsubq_f32(_r31, _r41);
-
-                float32x4_t _tmp00 = vaddq_f32(vaddq_f32(_r00, _tmp02a0), vaddq_f32(_tmp02b0, _bias0));
-                float32x4_t _tmp01 = vaddq_f32(vaddq_f32(_r01, _tmp02a1), vaddq_f32(_tmp02b1, _bias1));
-                float32x4_t _tmp10 = vfmaq_laneq_f32(vfmaq_laneq_f32(_bias0, _tmp13a0, _coeffs, 1), _tmp13b0, _coeffs, 0);
-                float32x4_t _tmp11 = vfmaq_laneq_f32(vfmaq_laneq_f32(_bias1, _tmp13a1, _coeffs, 1), _tmp13b1, _coeffs, 0);
-                float32x4_t _tmp20 = vfmaq_lane_f32(vfmaq_lane_f32(_bias0, _tmp02a0, _coeffs2, 0), _tmp02b0, _coeffs2, 1);
-                float32x4_t _tmp21 = vfmaq_lane_f32(vfmaq_lane_f32(_bias1, _tmp02a1, _coeffs2, 0), _tmp02b1, _coeffs2, 1);
-                float32x4_t _tmp30 = vfmaq_laneq_f32(vfmaq_laneq_f32(vaddq_f32(_r50, _bias0), _tmp13a0, _coeffs, 2), _tmp13b0, _coeffs, 3);
-                float32x4_t _tmp31 = vfmaq_laneq_f32(vfmaq_laneq_f32(vaddq_f32(_r51, _bias1), _tmp13a1, _coeffs, 2), _tmp13b1, _coeffs, 3);
-
-                if (out_elempack == 4)
-                {
-                    float* outptr1 = outptr0 + N;
-
-                    vst1q_f32(outptr0, _tmp00);
-                    vst1q_f32(outptr1, _tmp01);
-                    if (tj * 4 + 1 < outw)
-                    {
-                        vst1q_f32(outptr0 + 4, _tmp10);
-                        vst1q_f32(outptr1 + 4, _tmp11);
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        vst1q_f32(outptr0 + 8, _tmp20);
-                        vst1q_f32(outptr1 + 8, _tmp21);
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        vst1q_f32(outptr0 + 12, _tmp30);
-                        vst1q_f32(outptr1 + 12, _tmp31);
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    float tmp0[8];
-                    float tmp1[8];
-                    float tmp2[8];
-                    float tmp3[8];
-                    vst1q_f32(tmp0, _tmp00);
-                    vst1q_f32(tmp0 + 4, _tmp01);
-                    vst1q_f32(tmp1, _tmp10);
-                    vst1q_f32(tmp1 + 4, _tmp11);
-                    vst1q_f32(tmp2, _tmp20);
-                    vst1q_f32(tmp2 + 4, _tmp21);
-                    vst1q_f32(tmp3, _tmp30);
-                    vst1q_f32(tmp3 + 4, _tmp31);
-
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-                    float* outptr4 = outptr0 + N * 4;
-                    float* outptr5 = outptr0 + N * 5;
-                    float* outptr6 = outptr0 + N * 6;
-                    float* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                        outptr4[2] = tmp2[4];
-                        outptr5[2] = tmp2[5];
-                        outptr6[2] = tmp2[6];
-                        outptr7[2] = tmp2[7];
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                        outptr4[3] = tmp3[4];
-                        outptr5[3] = tmp3[5];
-                        outptr6[3] = tmp3[6];
-                        outptr7[3] = tmp3[7];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        float32x4_t _bias0 = biasptr ? vld1q_f32(biasptr + i + ii) : vdupq_n_f32(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[4][6][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 36 + jj * 4;
-            const float* r1 = r0 + max_jj * 4;
-            const float* r2 = r0 + max_jj * 4 * 2;
-            const float* r3 = r0 + max_jj * 4 * 3;
-            const float* r4 = r0 + max_jj * 4 * 4;
-            const float* r5 = r0 + max_jj * 4 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float32x4_t _r0 = vld1q_f32(r0);
-                float32x4_t _r1 = vld1q_f32(r1);
-                float32x4_t _r2 = vld1q_f32(r2);
-                float32x4_t _r3 = vld1q_f32(r3);
-                float32x4_t _r4 = vld1q_f32(r4);
-                float32x4_t _r5 = vld1q_f32(r5);
-
-                float32x4_t _tmp02a = vaddq_f32(_r1, _r2);
-                float32x4_t _tmp02b = vaddq_f32(_r3, _r4);
-                float32x4_t _tmp13a = vsubq_f32(_r1, _r2);
-                float32x4_t _tmp13b = vsubq_f32(_r3, _r4);
-
-                float32x4_t _tmp0 = vaddq_f32(vaddq_f32(_r0, _tmp02a), _tmp02b);
-#if __aarch64__
-                float32x4_t _tmp1 = vfmaq_laneq_f32(vmulq_laneq_f32(_tmp13a, _coeffs, 1), _tmp13b, _coeffs, 0);
-                float32x4_t _tmp2 = vfmaq_lane_f32(vmulq_lane_f32(_tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x4_t _tmp3 = vfmaq_laneq_f32(vfmaq_laneq_f32(_r5, _tmp13a, _coeffs, 2), _tmp13b, _coeffs, 3);
-#else
-                float32x4_t _tmp1 = vmlaq_lane_f32(vmulq_lane_f32(_tmp13a, vget_low_f32(_coeffs), 1), _tmp13b, vget_low_f32(_coeffs), 0);
-                float32x4_t _tmp2 = vmlaq_lane_f32(vmulq_lane_f32(_tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x4_t _tmp3 = vmlaq_lane_f32(vmlaq_lane_f32(_r5, _tmp13a, vget_high_f32(_coeffs), 0), _tmp13b, vget_high_f32(_coeffs), 1);
-#endif
-
-                vst1q_f32(tmp[0][m], _tmp0);
-                vst1q_f32(tmp[1][m], _tmp1);
-                vst1q_f32(tmp[2][m], _tmp2);
-                vst1q_f32(tmp[3][m], _tmp3);
-
-                r0 += max_jj * 6 * 4;
-                r1 += max_jj * 6 * 4;
-                r2 += max_jj * 6 * 4;
-                r3 += max_jj * 6 * 4;
-                r4 += max_jj * 6 * 4;
-                r5 += max_jj * 6 * 4;
-            }
-
-            float* outptr0 = top_blob.channel((i + ii) / out_elempack).row(ti * 4) + (tj * 4) * out_elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                float32x4_t _r0 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r1 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r2 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r3 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r4 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r5 = vld1q_f32(tmp[m][5]);
-
-                float32x4_t _tmp02a = vaddq_f32(_r1, _r2);
-                float32x4_t _tmp02b = vaddq_f32(_r3, _r4);
-                float32x4_t _tmp13a = vsubq_f32(_r1, _r2);
-                float32x4_t _tmp13b = vsubq_f32(_r3, _r4);
-
-                float32x4_t _tmp0 = vaddq_f32(vaddq_f32(_r0, _tmp02a), vaddq_f32(_tmp02b, _bias0));
-#if __aarch64__
-                float32x4_t _tmp1 = vfmaq_laneq_f32(vfmaq_laneq_f32(_bias0, _tmp13a, _coeffs, 1), _tmp13b, _coeffs, 0);
-                float32x4_t _tmp2 = vfmaq_lane_f32(vfmaq_lane_f32(_bias0, _tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x4_t _tmp3 = vfmaq_laneq_f32(vfmaq_laneq_f32(vaddq_f32(_r5, _bias0), _tmp13a, _coeffs, 2), _tmp13b, _coeffs, 3);
-#else
-                float32x4_t _tmp1 = vmlaq_lane_f32(vmlaq_lane_f32(_bias0, _tmp13a, vget_low_f32(_coeffs), 1), _tmp13b, vget_low_f32(_coeffs), 0);
-                float32x4_t _tmp2 = vmlaq_lane_f32(vmlaq_lane_f32(_bias0, _tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x4_t _tmp3 = vmlaq_lane_f32(vmlaq_lane_f32(vaddq_f32(_r5, _bias0), _tmp13a, vget_high_f32(_coeffs), 0), _tmp13b, vget_high_f32(_coeffs), 1);
-#endif
-
-                if (out_elempack == 4)
-                {
-                    vst1q_f32(outptr0, _tmp0);
-                    if (tj * 4 + 1 < outw) vst1q_f32(outptr0 + 4, _tmp1);
-                    if (tj * 4 + 2 < outw) vst1q_f32(outptr0 + 8, _tmp2);
-                    if (tj * 4 + 3 < outw) vst1q_f32(outptr0 + 12, _tmp3);
-                }
-                if (out_elempack == 1)
-                {
-                    float tmp0[4];
-                    float tmp1[4];
-                    float tmp2[4];
-                    float tmp3[4];
-                    vst1q_f32(tmp0, _tmp0);
-                    vst1q_f32(tmp1, _tmp1);
-                    vst1q_f32(tmp2, _tmp2);
-                    vst1q_f32(tmp3, _tmp3);
-
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-#if __ARM_NEON
-        float32x2_t _bias0 = biasptr ? vld1_f32(biasptr + i + ii) : vdup_n_f32(0.f);
-#else
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-        float bias1 = biasptr ? biasptr[i + ii + 1] : 0.f;
-#endif
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        float tmp[4][6][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 36 + jj * 2;
-            const float* r1 = r0 + max_jj * 2;
-            const float* r2 = r0 + max_jj * 2 * 2;
-            const float* r3 = r0 + max_jj * 2 * 3;
-            const float* r4 = r0 + max_jj * 2 * 4;
-            const float* r5 = r0 + max_jj * 2 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(r0);
-                float32x2_t _r1 = vld1_f32(r1);
-                float32x2_t _r2 = vld1_f32(r2);
-                float32x2_t _r3 = vld1_f32(r3);
-                float32x2_t _r4 = vld1_f32(r4);
-                float32x2_t _r5 = vld1_f32(r5);
-
-                float32x2_t _tmp02a = vadd_f32(_r1, _r2);
-                float32x2_t _tmp02b = vadd_f32(_r3, _r4);
-                float32x2_t _tmp13a = vsub_f32(_r1, _r2);
-                float32x2_t _tmp13b = vsub_f32(_r3, _r4);
-
-                float32x2_t _tmp0 = vadd_f32(vadd_f32(_r0, _tmp02a), _tmp02b);
-#if __aarch64__
-                float32x2_t _tmp1 = vfma_laneq_f32(vmul_laneq_f32(_tmp13a, _coeffs, 1), _tmp13b, _coeffs, 0);
-                float32x2_t _tmp2 = vfma_lane_f32(vmul_lane_f32(_tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x2_t _tmp3 = vfma_laneq_f32(vfma_laneq_f32(_r5, _tmp13a, _coeffs, 2), _tmp13b, _coeffs, 3);
-#else
-                float32x2_t _tmp1 = vmla_lane_f32(vmul_lane_f32(_tmp13a, vget_low_f32(_coeffs), 1), _tmp13b, vget_low_f32(_coeffs), 0);
-                float32x2_t _tmp2 = vmla_lane_f32(vmul_lane_f32(_tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x2_t _tmp3 = vmla_lane_f32(vmla_lane_f32(_r5, _tmp13a, vget_high_f32(_coeffs), 0), _tmp13b, vget_high_f32(_coeffs), 1);
-#endif
-
-                vst1_f32(tmp[0][m], _tmp0);
-                vst1_f32(tmp[1][m], _tmp1);
-                vst1_f32(tmp[2][m], _tmp2);
-                vst1_f32(tmp[3][m], _tmp3);
-#else
-                float tmp02a0 = r1[0] + r2[0];
-                float tmp02a1 = r1[1] + r2[1];
-                float tmp02b0 = r3[0] + r4[0];
-                float tmp02b1 = r3[1] + r4[1];
-                float tmp13a0 = r1[0] - r2[0];
-                float tmp13a1 = r1[1] - r2[1];
-                float tmp13b0 = r3[0] - r4[0];
-                float tmp13b1 = r3[1] - r4[1];
-
-                tmp[0][m][0] = r0[0] + tmp02a0 + tmp02b0;
-                tmp[0][m][1] = r0[1] + tmp02a1 + tmp02b1;
-                tmp[1][m][0] = tmp13a0 * sq2_d2 + tmp13b0 * sq2;
-                tmp[1][m][1] = tmp13a1 * sq2_d2 + tmp13b1 * sq2;
-                tmp[2][m][0] = tmp02a0 * 0.5f + tmp02b0 * 2;
-                tmp[2][m][1] = tmp02a1 * 0.5f + tmp02b1 * 2;
-                tmp[3][m][0] = r5[0] + tmp13a0 * sq2_d4 + tmp13b0 * sq2_m2;
-                tmp[3][m][1] = r5[1] + tmp13a1 * sq2_d4 + tmp13b1 * sq2_m2;
-#endif
-
-                r0 += max_jj * 6 * 2;
-                r1 += max_jj * 6 * 2;
-                r2 += max_jj * 6 * 2;
-                r3 += max_jj * 6 * 2;
-                r4 += max_jj * 6 * 2;
-                r5 += max_jj * 6 * 2;
-            }
-
-            float* outptr0 = top_blob.channel(i + ii).row(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(tmp[m][0]);
-                float32x2_t _r1 = vld1_f32(tmp[m][1]);
-                float32x2_t _r2 = vld1_f32(tmp[m][2]);
-                float32x2_t _r3 = vld1_f32(tmp[m][3]);
-                float32x2_t _r4 = vld1_f32(tmp[m][4]);
-                float32x2_t _r5 = vld1_f32(tmp[m][5]);
-
-                float32x2_t _tmp02a = vadd_f32(_r1, _r2);
-                float32x2_t _tmp02b = vadd_f32(_r3, _r4);
-                float32x2_t _tmp13a = vsub_f32(_r1, _r2);
-                float32x2_t _tmp13b = vsub_f32(_r3, _r4);
-
-                float32x2_t _tmp0 = vadd_f32(vadd_f32(_r0, _tmp02a), vadd_f32(_tmp02b, _bias0));
-#if __aarch64__
-                float32x2_t _tmp1 = vfma_laneq_f32(vfma_laneq_f32(_bias0, _tmp13a, _coeffs, 1), _tmp13b, _coeffs, 0);
-                float32x2_t _tmp2 = vfma_lane_f32(vfma_lane_f32(_bias0, _tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x2_t _tmp3 = vfma_laneq_f32(vfma_laneq_f32(vadd_f32(_r5, _bias0), _tmp13a, _coeffs, 2), _tmp13b, _coeffs, 3);
-#else
-                float32x2_t _tmp1 = vmla_lane_f32(vmla_lane_f32(_bias0, _tmp13a, vget_low_f32(_coeffs), 1), _tmp13b, vget_low_f32(_coeffs), 0);
-                float32x2_t _tmp2 = vmla_lane_f32(vmla_lane_f32(_bias0, _tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x2_t _tmp3 = vmla_lane_f32(vmla_lane_f32(vadd_f32(_r5, _bias0), _tmp13a, vget_high_f32(_coeffs), 0), _tmp13b, vget_high_f32(_coeffs), 1);
-#endif
-#else
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-                float r40 = tmp[m][4][0];
-                float r41 = tmp[m][4][1];
-                float r50 = tmp[m][5][0];
-                float r51 = tmp[m][5][1];
-
-                float tmp02a0 = r10 + r20;
-                float tmp02a1 = r11 + r21;
-                float tmp02b0 = r30 + r40;
-                float tmp02b1 = r31 + r41;
-                float tmp13a0 = r10 - r20;
-                float tmp13a1 = r11 - r21;
-                float tmp13b0 = r30 - r40;
-                float tmp13b1 = r31 - r41;
-
-                float tmp00 = bias0 + r00 + tmp02a0 + tmp02b0;
-                float tmp01 = bias1 + r01 + tmp02a1 + tmp02b1;
-                float tmp10 = bias0 + tmp13a0 * sq2_d2 + tmp13b0 * sq2;
-                float tmp11 = bias1 + tmp13a1 * sq2_d2 + tmp13b1 * sq2;
-                float tmp20 = bias0 + tmp02a0 * 0.5f + tmp02b0 * 2;
-                float tmp21 = bias1 + tmp02a1 * 0.5f + tmp02b1 * 2;
-                float tmp30 = bias0 + r50 + tmp13a0 * sq2_d4 + tmp13b0 * sq2_m2;
-                float tmp31 = bias1 + r51 + tmp13a1 * sq2_d4 + tmp13b1 * sq2_m2;
-#endif
-
-                // if (out_elempack == 1)
-                {
-                    float* outptr1 = outptr0 + N;
-
-#if __ARM_NEON
-                    outptr0[0] = vget_lane_f32(_tmp0, 0);
-                    outptr1[0] = vget_lane_f32(_tmp0, 1);
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = vget_lane_f32(_tmp1, 0);
-                        outptr1[1] = vget_lane_f32(_tmp1, 1);
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = vget_lane_f32(_tmp2, 0);
-                        outptr1[2] = vget_lane_f32(_tmp2, 1);
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = vget_lane_f32(_tmp3, 0);
-                        outptr1[3] = vget_lane_f32(_tmp3, 1);
-                    }
-#else
-                    outptr0[0] = tmp00;
-                    outptr1[0] = tmp01;
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp10;
-                        outptr1[1] = tmp11;
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp20;
-                        outptr1[2] = tmp21;
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp30;
-                        outptr1[3] = tmp31;
-                    }
-#endif
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-
-        float tmp[4][6];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 36 + jj;
-            const float* r1 = r0 + max_jj;
-            const float* r2 = r0 + max_jj * 2;
-            const float* r3 = r0 + max_jj * 3;
-            const float* r4 = r0 + max_jj * 4;
-            const float* r5 = r0 + max_jj * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float tmp02a = r1[0] + r2[0];
-                float tmp02b = r3[0] + r4[0];
-                float tmp13a = r1[0] - r2[0];
-                float tmp13b = r3[0] - r4[0];
-
-                tmp[0][m] = r0[0] + tmp02a + tmp02b;
-                tmp[1][m] = tmp13a * sq2_d2 + tmp13b * sq2;
-                tmp[2][m] = tmp02a * 0.5f + tmp02b * 2;
-                tmp[3][m] = r5[0] + tmp13a * sq2_d4 + tmp13b * sq2_m2;
-
-                r0 += max_jj * 6;
-                r1 += max_jj * 6;
-                r2 += max_jj * 6;
-                r3 += max_jj * 6;
-                r4 += max_jj * 6;
-                r5 += max_jj * 6;
-            }
-
-            float* outptr0 = top_blob.channel(i + ii).row(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-                float r4 = tmp[m][4];
-                float r5 = tmp[m][5];
-
-                float tmp02a = r1 + r2;
-                float tmp02b = r3 + r4;
-                float tmp13a = r1 - r2;
-                float tmp13b = r3 - r4;
-
-                float tmp0 = bias0 + r0 + tmp02a + tmp02b;
-                float tmp1 = bias0 + tmp13a * sq2_d2 + tmp13b * sq2;
-                float tmp2 = bias0 + tmp02a * 0.5f + tmp02b * 2;
-                float tmp3 = bias0 + r5 + tmp13a * sq2_d4 + tmp13b * sq2_m2;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = tmp0;
-                    if (tj * 4 + 1 < outw) outptr0[1] = tmp1;
-                    if (tj * 4 + 2 < outw) outptr0[2] = tmp2;
-                    if (tj * 4 + 3 < outw) outptr0[3] = tmp3;
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, const Mat& bias, int nT, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 4n+2, winograd F(4,3)
-    int w_tiles = (outw + 3) / 4;
-    int h_tiles = (outh + 3) / 4;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 36;
-
-    // NCNN_LOGE("conv3x3s1_winograd43 %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    conv3x3s1_winograd_get_optimal_tile_mnk(M, N, K, B, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 4u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 4u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd43_transform_input_tile(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 4u, opt.workspace_allocator);
-
-        #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd43_transform_input_tile(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                conv3x3s1_winograd_gemm_transB_packed_tile(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk, opt.use_a53_a55_optimized_kernel);
-            }
-
-            // transform output
-            conv3x3s1_winograd43_transform_output_tile(top_tile, top_blob, bias, i, max_ii, j, max_jj);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd63_transform_kernel_tile(const Mat& kernel, Mat& A, int inch, int i, int max_ii, int k, int max_kk)
-{
-    float* ptmp = A;
-
-    int ii = 0;
-    for (; ii < max_ii; ii++)
-    {
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            // const float ktm[8][3] = {
-            //     {1.0f, 0.0f, 0.0f},
-            //     {-2.0f / 9, -2.0f / 9, -2.0f / 9},
-            //     {-2.0f / 9, 2.0f / 9, -2.0f / 9},
-            //     {1.0f / 90, 1.0f / 45, 2.0f / 45},
-            //     {1.0f / 90, -1.0f / 45, 2.0f / 45},
-            //     {1.0f / 45, 1.0f / 90, 1.0f / 180},
-            //     {1.0f / 45, -1.0f / 90, 1.0f / 180},
-            //     {0.0f, 0.0f, 1.0f}
-            // };
-            const float ktm0 = 2.0f / 9;
-            const float ktm1 = 1.0f / 45;
-            const float ktm2 = 2.0f / 45;
-            const float ktm3 = 1.0f / 90;
-            const float ktm4 = 1.0f / 180;
-
-            float tmp[8][3];
-
-            const float* k0 = (const float*)kernel + (i + ii) * inch * 9 + (k + kk) * 9;
-
-            for (int m = 0; m < 3; m++)
-            {
-                float r0 = k0[0];
-                float r1 = k0[1];
-                float r2 = k0[2];
-
-                tmp[0][m] = r0;
-                tmp[1][m] = -r0 * ktm0 - r1 * ktm0 - r2 * ktm0;
-                tmp[2][m] = -r0 * ktm0 + r1 * ktm0 - r2 * ktm0;
-                tmp[3][m] = r0 * ktm3 + r1 * ktm1 + r2 * ktm2;
-                tmp[4][m] = r0 * ktm3 - r1 * ktm1 + r2 * ktm2;
-                tmp[5][m] = r0 * ktm1 + r1 * ktm3 + r2 * ktm4;
-                tmp[6][m] = r0 * ktm1 - r1 * ktm3 + r2 * ktm4;
-                tmp[7][m] = r2;
-
-                k0 += 3;
-            }
-
-            for (int m = 0; m < 8; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-
-                float z0 = r0;
-                float z1 = -r0 * ktm0 - r1 * ktm0 - r2 * ktm0;
-                float z2 = -r0 * ktm0 + r1 * ktm0 - r2 * ktm0;
-                float z3 = r0 * ktm3 + r1 * ktm1 + r2 * ktm2;
-                float z4 = r0 * ktm3 - r1 * ktm1 + r2 * ktm2;
-                float z5 = r0 * ktm1 + r1 * ktm3 + r2 * ktm4;
-                float z6 = r0 * ktm1 - r1 * ktm3 + r2 * ktm4;
-                float z7 = r2;
-
-                ptmp[0] = z0;
-                ptmp[1] = z1;
-                ptmp[2] = z2;
-                ptmp[3] = z3;
-                ptmp[4] = z4;
-                ptmp[5] = z5;
-                ptmp[6] = z6;
-                ptmp[7] = z7;
-                ptmp += 8;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd63_transform_kernel(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt)
-{
-    const int M = outch;
-    const int K = inch;
-    const int B = 64;
-
-    int TILE_M, TILE_N, TILE_K;
-    conv3x3s1_winograd_get_optimal_tile_mnk(M, 0, K, B, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    Mat A_tileX(B * TILE_M * TILE_K, 1, opt.num_threads);
-
-    AT.create(TILE_K * TILE_M, B, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat A_tile = A_tileX.channel(get_omp_thread_num());
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_ii = std::min((M - i), TILE_M);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            conv3x3s1_winograd63_transform_kernel_tile(kernel, A_tile, inch, i, max_ii, k, max_kk);
-
-            Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-            conv3x3s1_winograd_pack_A_tile(A_tile, AT_tile, B, max_ii, max_kk);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd63_transform_input_tile(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    // const float itm[8][8] = {
-    //     {1.0f, 0.0f,-5.25f, 0.00f, 5.25f, 0.00f,-1.0f, 0.0f},
-    //     {0.0f, 1.0f, 1.00f,-4.25f,-4.25f, 1.00f, 1.0f, 0.0f},
-    //     {0.0f,-1.0f, 1.00f, 4.25f,-4.25f,-1.00f, 1.0f, 0.0f},
-    //     {0.0f, 0.5f, 0.25f,-2.50f,-1.25f, 2.00f, 1.0f, 0.0f},
-    //     {0.0f,-0.5f, 0.25f, 2.50f,-1.25f,-2.00f, 1.0f, 0.0f},
-    //     {0.0f, 2.0f, 4.00f,-2.50f,-5.00f, 0.50f, 1.0f, 0.0f},
-    //     {0.0f,-2.0f, 4.00f, 2.50f,-5.00f,-0.50f, 1.0f, 0.0f},
-    //     {0.0f,-1.0f, 0.00f, 5.25f, 0.00f,-5.25f, 0.0f, 1.0f}
-    // };
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w + 3) / 6;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-#if __ARM_NEON
-#if __aarch64__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[8][8][8];
-
-        const float coeffs[8] = {5.25f, -4.25f, -1.25f, 0.25f, -2.5f, 0.5f, 2.f, 4.f};
-        float32x4_t _coeffs = vld1q_f32(coeffs);
-        float32x4_t _coeffs2 = vld1q_f32(coeffs + 4);
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel((k + kk) / elempack).row(ti * 6) + (tj * 6) * elempack;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float32x4_t _r00 = vdupq_n_f32(0.f);
-                float32x4_t _r01 = vdupq_n_f32(0.f);
-                float32x4_t _r10 = vdupq_n_f32(0.f);
-                float32x4_t _r11 = vdupq_n_f32(0.f);
-                float32x4_t _r20 = vdupq_n_f32(0.f);
-                float32x4_t _r21 = vdupq_n_f32(0.f);
-                float32x4_t _r30 = vdupq_n_f32(0.f);
-                float32x4_t _r31 = vdupq_n_f32(0.f);
-                float32x4_t _r40 = vdupq_n_f32(0.f);
-                float32x4_t _r41 = vdupq_n_f32(0.f);
-                float32x4_t _r50 = vdupq_n_f32(0.f);
-                float32x4_t _r51 = vdupq_n_f32(0.f);
-                float32x4_t _r60 = vdupq_n_f32(0.f);
-                float32x4_t _r61 = vdupq_n_f32(0.f);
-                float32x4_t _r70 = vdupq_n_f32(0.f);
-                float32x4_t _r71 = vdupq_n_f32(0.f);
-
-                if (ti * 6 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-
-                        _r00 = vld1q_f32(r0);
-                        _r01 = vld1q_f32(r1);
-                        if (tj * 6 + 1 < w)
-                        {
-                            _r10 = vld1q_f32(r0 + 4);
-                            _r11 = vld1q_f32(r1 + 4);
-                        }
-                        if (tj * 6 + 2 < w)
-                        {
-                            _r20 = vld1q_f32(r0 + 8);
-                            _r21 = vld1q_f32(r1 + 8);
-                        }
-                        if (tj * 6 + 3 < w)
-                        {
-                            _r30 = vld1q_f32(r0 + 12);
-                            _r31 = vld1q_f32(r1 + 12);
-                        }
-                        if (tj * 6 + 4 < w)
-                        {
-                            _r40 = vld1q_f32(r0 + 16);
-                            _r41 = vld1q_f32(r1 + 16);
-                        }
-                        if (tj * 6 + 5 < w)
-                        {
-                            _r50 = vld1q_f32(r0 + 20);
-                            _r51 = vld1q_f32(r1 + 20);
-                        }
-                        if (tj * 6 + 6 < w)
-                        {
-                            _r60 = vld1q_f32(r0 + 24);
-                            _r61 = vld1q_f32(r1 + 24);
-                        }
-                        if (tj * 6 + 7 < w)
-                        {
-                            _r70 = vld1q_f32(r0 + 28);
-                            _r71 = vld1q_f32(r1 + 28);
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-                        const float* r4 = r0 + N * 4;
-                        const float* r5 = r0 + N * 5;
-                        const float* r6 = r0 + N * 6;
-                        const float* r7 = r0 + N * 7;
-
-                        float32x4_t _t0 = vld1q_f32(r0);
-                        float32x4_t _t1 = vld1q_f32(r1);
-                        float32x4_t _t2 = vld1q_f32(r2);
-                        float32x4_t _t3 = vld1q_f32(r3);
-                        float32x4_t _t4 = vld1q_f32(r4);
-                        float32x4_t _t5 = vld1q_f32(r5);
-                        float32x4_t _t6 = vld1q_f32(r6);
-                        float32x4_t _t7 = vld1q_f32(r7);
-
-                        transpose4x4_ps(_t0, _t1, _t2, _t3);
-                        transpose4x4_ps(_t4, _t5, _t6, _t7);
-
-                        _r00 = _t0;
-                        _r01 = _t4;
-                        if (tj * 6 + 1 < w)
-                        {
-                            _r10 = _t1;
-                            _r11 = _t5;
-                        }
-                        if (tj * 6 + 2 < w)
-                        {
-                            _r20 = _t2;
-                            _r21 = _t6;
-                        }
-                        if (tj * 6 + 3 < w)
-                        {
-                            _r30 = _t3;
-                            _r31 = _t7;
-                        }
-                        if (tj * 6 + 4 < w)
-                        {
-                            _t0 = vld1q_f32(r0 + 4);
-                            _t1 = vld1q_f32(r1 + 4);
-                            _t2 = vld1q_f32(r2 + 4);
-                            _t3 = vld1q_f32(r3 + 4);
-                            _t4 = vld1q_f32(r4 + 4);
-                            _t5 = vld1q_f32(r5 + 4);
-                            _t6 = vld1q_f32(r6 + 4);
-                            _t7 = vld1q_f32(r7 + 4);
-
-                            transpose4x4_ps(_t0, _t1, _t2, _t3);
-                            transpose4x4_ps(_t4, _t5, _t6, _t7);
-
-                            _r40 = _t0;
-                            _r41 = _t4;
-                            if (tj * 6 + 5 < w)
-                            {
-                                _r50 = _t1;
-                                _r51 = _t5;
-                            }
-                            if (tj * 6 + 6 < w)
-                            {
-                                _r60 = _t2;
-                                _r61 = _t6;
-                            }
-                            if (tj * 6 + 7 < w)
-                            {
-                                _r70 = _t3;
-                                _r71 = _t7;
-                            }
-                        }
-                    }
-                }
-
-                float32x4_t _tmp12a0 = vfmaq_laneq_f32(vaddq_f32(_r20, _r60), _r40, _coeffs, 1);
-                float32x4_t _tmp12a1 = vfmaq_laneq_f32(vaddq_f32(_r21, _r61), _r41, _coeffs, 1);
-                float32x4_t _tmp12b0 = vfmaq_laneq_f32(vaddq_f32(_r10, _r50), _r30, _coeffs, 1);
-                float32x4_t _tmp12b1 = vfmaq_laneq_f32(vaddq_f32(_r11, _r51), _r31, _coeffs, 1);
-                float32x4_t _tmp34a0 = vfmaq_laneq_f32(vfmaq_laneq_f32(_r60, _r20, _coeffs, 3), _r40, _coeffs, 2);
-                float32x4_t _tmp34a1 = vfmaq_laneq_f32(vfmaq_laneq_f32(_r61, _r21, _coeffs, 3), _r41, _coeffs, 2);
-                float32x4_t _tmp34b0 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r10, _coeffs2, 1), _r30, _coeffs2, 0), _r50, _coeffs2, 2);
-                float32x4_t _tmp34b1 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r11, _coeffs2, 1), _r31, _coeffs2, 0), _r51, _coeffs2, 2);
-                float32x4_t _tmp56a0 = vfmaq_laneq_f32(_r60, vfmaq_laneq_f32(_r20, _r40, _coeffs, 2), _coeffs2, 3);
-                float32x4_t _tmp56a1 = vfmaq_laneq_f32(_r61, vfmaq_laneq_f32(_r21, _r41, _coeffs, 2), _coeffs2, 3);
-                float32x4_t _tmp56b0 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r10, _coeffs2, 2), _r30, _coeffs2, 0), _r50, _coeffs2, 1);
-                float32x4_t _tmp56b1 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r11, _coeffs2, 2), _r31, _coeffs2, 0), _r51, _coeffs2, 1);
-
-                float32x4_t _tmp00 = vfmaq_laneq_f32(vsubq_f32(_r00, _r60), vsubq_f32(_r40, _r20), _coeffs, 0);
-                float32x4_t _tmp01 = vfmaq_laneq_f32(vsubq_f32(_r01, _r61), vsubq_f32(_r41, _r21), _coeffs, 0);
-                float32x4_t _tmp10 = vaddq_f32(_tmp12a0, _tmp12b0);
-                float32x4_t _tmp11 = vaddq_f32(_tmp12a1, _tmp12b1);
-                float32x4_t _tmp20 = vsubq_f32(_tmp12a0, _tmp12b0);
-                float32x4_t _tmp21 = vsubq_f32(_tmp12a1, _tmp12b1);
-                float32x4_t _tmp30 = vaddq_f32(_tmp34a0, _tmp34b0);
-                float32x4_t _tmp31 = vaddq_f32(_tmp34a1, _tmp34b1);
-                float32x4_t _tmp40 = vsubq_f32(_tmp34a0, _tmp34b0);
-                float32x4_t _tmp41 = vsubq_f32(_tmp34a1, _tmp34b1);
-                float32x4_t _tmp50 = vaddq_f32(_tmp56a0, _tmp56b0);
-                float32x4_t _tmp51 = vaddq_f32(_tmp56a1, _tmp56b1);
-                float32x4_t _tmp60 = vsubq_f32(_tmp56a0, _tmp56b0);
-                float32x4_t _tmp61 = vsubq_f32(_tmp56a1, _tmp56b1);
-                float32x4_t _tmp70 = vfmaq_laneq_f32(vsubq_f32(_r70, _r10), vsubq_f32(_r30, _r50), _coeffs, 0);
-                float32x4_t _tmp71 = vfmaq_laneq_f32(vsubq_f32(_r71, _r11), vsubq_f32(_r31, _r51), _coeffs, 0);
-
-                vst1q_f32(tmp[0][m], _tmp00);
-                vst1q_f32(tmp[0][m] + 4, _tmp01);
-                vst1q_f32(tmp[1][m], _tmp10);
-                vst1q_f32(tmp[1][m] + 4, _tmp11);
-                vst1q_f32(tmp[2][m], _tmp20);
-                vst1q_f32(tmp[2][m] + 4, _tmp21);
-                vst1q_f32(tmp[3][m], _tmp30);
-                vst1q_f32(tmp[3][m] + 4, _tmp31);
-                vst1q_f32(tmp[4][m], _tmp40);
-                vst1q_f32(tmp[4][m] + 4, _tmp41);
-                vst1q_f32(tmp[5][m], _tmp50);
-                vst1q_f32(tmp[5][m] + 4, _tmp51);
-                vst1q_f32(tmp[6][m], _tmp60);
-                vst1q_f32(tmp[6][m] + 4, _tmp61);
-                vst1q_f32(tmp[7][m], _tmp70);
-                vst1q_f32(tmp[7][m] + 4, _tmp71);
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 64 + jj * 8;
-            float* p1 = p0 + max_jj * 8;
-            float* p2 = p0 + max_jj * 8 * 2;
-            float* p3 = p0 + max_jj * 8 * 3;
-            float* p4 = p0 + max_jj * 8 * 4;
-            float* p5 = p0 + max_jj * 8 * 5;
-            float* p6 = p0 + max_jj * 8 * 6;
-            float* p7 = p0 + max_jj * 8 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float32x4_t _r00 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r01 = vld1q_f32(tmp[m][0] + 4);
-                float32x4_t _r10 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r11 = vld1q_f32(tmp[m][1] + 4);
-                float32x4_t _r20 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r21 = vld1q_f32(tmp[m][2] + 4);
-                float32x4_t _r30 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r31 = vld1q_f32(tmp[m][3] + 4);
-                float32x4_t _r40 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r41 = vld1q_f32(tmp[m][4] + 4);
-                float32x4_t _r50 = vld1q_f32(tmp[m][5]);
-                float32x4_t _r51 = vld1q_f32(tmp[m][5] + 4);
-                float32x4_t _r60 = vld1q_f32(tmp[m][6]);
-                float32x4_t _r61 = vld1q_f32(tmp[m][6] + 4);
-                float32x4_t _r70 = vld1q_f32(tmp[m][7]);
-                float32x4_t _r71 = vld1q_f32(tmp[m][7] + 4);
-
-                float32x4_t _tmp12a0 = vfmaq_laneq_f32(vaddq_f32(_r20, _r60), _r40, _coeffs, 1);
-                float32x4_t _tmp12a1 = vfmaq_laneq_f32(vaddq_f32(_r21, _r61), _r41, _coeffs, 1);
-                float32x4_t _tmp12b0 = vfmaq_laneq_f32(vaddq_f32(_r10, _r50), _r30, _coeffs, 1);
-                float32x4_t _tmp12b1 = vfmaq_laneq_f32(vaddq_f32(_r11, _r51), _r31, _coeffs, 1);
-                float32x4_t _tmp34a0 = vfmaq_laneq_f32(vfmaq_laneq_f32(_r60, _r20, _coeffs, 3), _r40, _coeffs, 2);
-                float32x4_t _tmp34a1 = vfmaq_laneq_f32(vfmaq_laneq_f32(_r61, _r21, _coeffs, 3), _r41, _coeffs, 2);
-                float32x4_t _tmp34b0 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r10, _coeffs2, 1), _r30, _coeffs2, 0), _r50, _coeffs2, 2);
-                float32x4_t _tmp34b1 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r11, _coeffs2, 1), _r31, _coeffs2, 0), _r51, _coeffs2, 2);
-                float32x4_t _tmp56a0 = vfmaq_laneq_f32(_r60, vfmaq_laneq_f32(_r20, _r40, _coeffs, 2), _coeffs2, 3);
-                float32x4_t _tmp56a1 = vfmaq_laneq_f32(_r61, vfmaq_laneq_f32(_r21, _r41, _coeffs, 2), _coeffs2, 3);
-                float32x4_t _tmp56b0 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r10, _coeffs2, 2), _r30, _coeffs2, 0), _r50, _coeffs2, 1);
-                float32x4_t _tmp56b1 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r11, _coeffs2, 2), _r31, _coeffs2, 0), _r51, _coeffs2, 1);
-
-                float32x4_t _tmp00 = vfmaq_laneq_f32(vsubq_f32(_r00, _r60), vsubq_f32(_r40, _r20), _coeffs, 0);
-                float32x4_t _tmp01 = vfmaq_laneq_f32(vsubq_f32(_r01, _r61), vsubq_f32(_r41, _r21), _coeffs, 0);
-                float32x4_t _tmp10 = vaddq_f32(_tmp12a0, _tmp12b0);
-                float32x4_t _tmp11 = vaddq_f32(_tmp12a1, _tmp12b1);
-                float32x4_t _tmp20 = vsubq_f32(_tmp12a0, _tmp12b0);
-                float32x4_t _tmp21 = vsubq_f32(_tmp12a1, _tmp12b1);
-                float32x4_t _tmp30 = vaddq_f32(_tmp34a0, _tmp34b0);
-                float32x4_t _tmp31 = vaddq_f32(_tmp34a1, _tmp34b1);
-                float32x4_t _tmp40 = vsubq_f32(_tmp34a0, _tmp34b0);
-                float32x4_t _tmp41 = vsubq_f32(_tmp34a1, _tmp34b1);
-                float32x4_t _tmp50 = vaddq_f32(_tmp56a0, _tmp56b0);
-                float32x4_t _tmp51 = vaddq_f32(_tmp56a1, _tmp56b1);
-                float32x4_t _tmp60 = vsubq_f32(_tmp56a0, _tmp56b0);
-                float32x4_t _tmp61 = vsubq_f32(_tmp56a1, _tmp56b1);
-                float32x4_t _tmp70 = vfmaq_laneq_f32(vsubq_f32(_r70, _r10), vsubq_f32(_r30, _r50), _coeffs, 0);
-                float32x4_t _tmp71 = vfmaq_laneq_f32(vsubq_f32(_r71, _r11), vsubq_f32(_r31, _r51), _coeffs, 0);
-
-                vst1q_f32(p0, _tmp00);
-                vst1q_f32(p0 + 4, _tmp01);
-                vst1q_f32(p1, _tmp10);
-                vst1q_f32(p1 + 4, _tmp11);
-                vst1q_f32(p2, _tmp20);
-                vst1q_f32(p2 + 4, _tmp21);
-                vst1q_f32(p3, _tmp30);
-                vst1q_f32(p3 + 4, _tmp31);
-                vst1q_f32(p4, _tmp40);
-                vst1q_f32(p4 + 4, _tmp41);
-                vst1q_f32(p5, _tmp50);
-                vst1q_f32(p5 + 4, _tmp51);
-                vst1q_f32(p6, _tmp60);
-                vst1q_f32(p6 + 4, _tmp61);
-                vst1q_f32(p7, _tmp70);
-                vst1q_f32(p7 + 4, _tmp71);
-
-                p0 += max_jj * 8 * 8;
-                p1 += max_jj * 8 * 8;
-                p2 += max_jj * 8 * 8;
-                p3 += max_jj * 8 * 8;
-                p4 += max_jj * 8 * 8;
-                p5 += max_jj * 8 * 8;
-                p6 += max_jj * 8 * 8;
-                p7 += max_jj * 8 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-#else // __aarch64__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-    #pragma omp parallel for num_threads(nT)
-#endif // __aarch64__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 4;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[8][8][4];
-
-        const float coeffs[8] = {5.25f, -4.25f, -1.25f, 0.25f, -2.5f, 0.5f, 2.f, 4.f};
-        float32x4_t _coeffs = vld1q_f32(coeffs);
-        float32x4_t _coeffs2 = vld1q_f32(coeffs + 4);
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel((k + kk) / elempack).row(ti * 6) + (tj * 6) * elempack;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float32x4_t _r0 = vdupq_n_f32(0.f);
-                float32x4_t _r1 = vdupq_n_f32(0.f);
-                float32x4_t _r2 = vdupq_n_f32(0.f);
-                float32x4_t _r3 = vdupq_n_f32(0.f);
-                float32x4_t _r4 = vdupq_n_f32(0.f);
-                float32x4_t _r5 = vdupq_n_f32(0.f);
-                float32x4_t _r6 = vdupq_n_f32(0.f);
-                float32x4_t _r7 = vdupq_n_f32(0.f);
-
-                if (ti * 6 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1q_f32(r0);
-                        if (tj * 6 + 1 < w) _r1 = vld1q_f32(r0 + 4);
-                        if (tj * 6 + 2 < w) _r2 = vld1q_f32(r0 + 8);
-                        if (tj * 6 + 3 < w) _r3 = vld1q_f32(r0 + 12);
-                        if (tj * 6 + 4 < w) _r4 = vld1q_f32(r0 + 16);
-                        if (tj * 6 + 5 < w) _r5 = vld1q_f32(r0 + 20);
-                        if (tj * 6 + 6 < w) _r6 = vld1q_f32(r0 + 24);
-                        if (tj * 6 + 7 < w) _r7 = vld1q_f32(r0 + 28);
-                    }
-                    if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-
-                        float32x4_t _t0 = vld1q_f32(r0);
-                        float32x4_t _t1 = vld1q_f32(r1);
-                        float32x4_t _t2 = vld1q_f32(r2);
-                        float32x4_t _t3 = vld1q_f32(r3);
-
-                        transpose4x4_ps(_t0, _t1, _t2, _t3);
-
-                        _r0 = _t0;
-                        if (tj * 6 + 1 < w) _r1 = _t1;
-                        if (tj * 6 + 2 < w) _r2 = _t2;
-                        if (tj * 6 + 3 < w) _r3 = _t3;
-                        if (tj * 6 + 4 < w)
-                        {
-                            _t0 = vld1q_f32(r0 + 4);
-                            _t1 = vld1q_f32(r1 + 4);
-                            _t2 = vld1q_f32(r2 + 4);
-                            _t3 = vld1q_f32(r3 + 4);
-
-                            transpose4x4_ps(_t0, _t1, _t2, _t3);
-
-                            _r4 = _t0;
-                            if (tj * 6 + 5 < w) _r5 = _t1;
-                            if (tj * 6 + 6 < w) _r6 = _t2;
-                            if (tj * 6 + 7 < w) _r7 = _t3;
-                        }
-                    }
-                }
-
-#if __aarch64__
-                float32x4_t _tmp12a = vfmaq_laneq_f32(vaddq_f32(_r2, _r6), _r4, _coeffs, 1);
-                float32x4_t _tmp12b = vfmaq_laneq_f32(vaddq_f32(_r1, _r5), _r3, _coeffs, 1);
-                float32x4_t _tmp34a = vfmaq_laneq_f32(vfmaq_laneq_f32(_r6, _r2, _coeffs, 3), _r4, _coeffs, 2);
-                float32x4_t _tmp34b = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r1, _coeffs2, 1), _r3, _coeffs2, 0), _r5, _coeffs2, 2);
-                float32x4_t _tmp56a = vfmaq_laneq_f32(_r6, vfmaq_laneq_f32(_r2, _r4, _coeffs, 2), _coeffs2, 3);
-                float32x4_t _tmp56b = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r1, _coeffs2, 2), _r3, _coeffs2, 0), _r5, _coeffs2, 1);
-#else
-                float32x4_t _tmp12a = vmlaq_lane_f32(vaddq_f32(_r2, _r6), _r4, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp12b = vmlaq_lane_f32(vaddq_f32(_r1, _r5), _r3, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp34a = vmlaq_lane_f32(vmlaq_lane_f32(_r6, _r2, vget_high_f32(_coeffs), 1), _r4, vget_high_f32(_coeffs), 0);
-                float32x4_t _tmp34b = vmlaq_lane_f32(vmlaq_lane_f32(vmulq_lane_f32(_r1, vget_low_f32(_coeffs2), 1), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_high_f32(_coeffs2), 0);
-                float32x4_t _tmp56a = vmlaq_lane_f32(_r6, vmlaq_lane_f32(_r2, _r4, vget_high_f32(_coeffs), 0), vget_high_f32(_coeffs2), 1);
-                float32x4_t _tmp56b = vmlaq_lane_f32(vmlaq_lane_f32(vmulq_lane_f32(_r1, vget_high_f32(_coeffs2), 0), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_low_f32(_coeffs2), 1);
-#endif
-
-#if __aarch64__
-                float32x4_t _tmp0 = vfmaq_laneq_f32(vsubq_f32(_r0, _r6), vsubq_f32(_r4, _r2), _coeffs, 0);
-#else
-                float32x4_t _tmp0 = vmlaq_lane_f32(vsubq_f32(_r0, _r6), vsubq_f32(_r4, _r2), vget_low_f32(_coeffs), 0);
-#endif
-                float32x4_t _tmp1 = vaddq_f32(_tmp12a, _tmp12b);
-                float32x4_t _tmp2 = vsubq_f32(_tmp12a, _tmp12b);
-                float32x4_t _tmp3 = vaddq_f32(_tmp34a, _tmp34b);
-                float32x4_t _tmp4 = vsubq_f32(_tmp34a, _tmp34b);
-                float32x4_t _tmp5 = vaddq_f32(_tmp56a, _tmp56b);
-                float32x4_t _tmp6 = vsubq_f32(_tmp56a, _tmp56b);
-#if __aarch64__
-                float32x4_t _tmp7 = vfmaq_laneq_f32(vsubq_f32(_r7, _r1), vsubq_f32(_r3, _r5), _coeffs, 0);
-#else
-                float32x4_t _tmp7 = vmlaq_lane_f32(vsubq_f32(_r7, _r1), vsubq_f32(_r3, _r5), vget_low_f32(_coeffs), 0);
-#endif
-
-                vst1q_f32(tmp[0][m], _tmp0);
-                vst1q_f32(tmp[1][m], _tmp1);
-                vst1q_f32(tmp[2][m], _tmp2);
-                vst1q_f32(tmp[3][m], _tmp3);
-                vst1q_f32(tmp[4][m], _tmp4);
-                vst1q_f32(tmp[5][m], _tmp5);
-                vst1q_f32(tmp[6][m], _tmp6);
-                vst1q_f32(tmp[7][m], _tmp7);
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 64 + jj * 4;
-            float* p1 = p0 + max_jj * 4;
-            float* p2 = p0 + max_jj * 4 * 2;
-            float* p3 = p0 + max_jj * 4 * 3;
-            float* p4 = p0 + max_jj * 4 * 4;
-            float* p5 = p0 + max_jj * 4 * 5;
-            float* p6 = p0 + max_jj * 4 * 6;
-            float* p7 = p0 + max_jj * 4 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float32x4_t _r0 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r1 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r2 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r3 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r4 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r5 = vld1q_f32(tmp[m][5]);
-                float32x4_t _r6 = vld1q_f32(tmp[m][6]);
-                float32x4_t _r7 = vld1q_f32(tmp[m][7]);
-
-#if __aarch64__
-                float32x4_t _tmp12a = vfmaq_laneq_f32(vaddq_f32(_r2, _r6), _r4, _coeffs, 1);
-                float32x4_t _tmp12b = vfmaq_laneq_f32(vaddq_f32(_r1, _r5), _r3, _coeffs, 1);
-                float32x4_t _tmp34a = vfmaq_laneq_f32(vfmaq_laneq_f32(_r6, _r2, _coeffs, 3), _r4, _coeffs, 2);
-                float32x4_t _tmp34b = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r1, _coeffs2, 1), _r3, _coeffs2, 0), _r5, _coeffs2, 2);
-                float32x4_t _tmp56a = vfmaq_laneq_f32(_r6, vfmaq_laneq_f32(_r2, _r4, _coeffs, 2), _coeffs2, 3);
-                float32x4_t _tmp56b = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r1, _coeffs2, 2), _r3, _coeffs2, 0), _r5, _coeffs2, 1);
-#else
-                float32x4_t _tmp12a = vmlaq_lane_f32(vaddq_f32(_r2, _r6), _r4, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp12b = vmlaq_lane_f32(vaddq_f32(_r1, _r5), _r3, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp34a = vmlaq_lane_f32(vmlaq_lane_f32(_r6, _r2, vget_high_f32(_coeffs), 1), _r4, vget_high_f32(_coeffs), 0);
-                float32x4_t _tmp34b = vmlaq_lane_f32(vmlaq_lane_f32(vmulq_lane_f32(_r1, vget_low_f32(_coeffs2), 1), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_high_f32(_coeffs2), 0);
-                float32x4_t _tmp56a = vmlaq_lane_f32(_r6, vmlaq_lane_f32(_r2, _r4, vget_high_f32(_coeffs), 0), vget_high_f32(_coeffs2), 1);
-                float32x4_t _tmp56b = vmlaq_lane_f32(vmlaq_lane_f32(vmulq_lane_f32(_r1, vget_high_f32(_coeffs2), 0), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_low_f32(_coeffs2), 1);
-#endif
-
-#if __aarch64__
-                float32x4_t _tmp0 = vfmaq_laneq_f32(vsubq_f32(_r0, _r6), vsubq_f32(_r4, _r2), _coeffs, 0);
-#else
-                float32x4_t _tmp0 = vmlaq_lane_f32(vsubq_f32(_r0, _r6), vsubq_f32(_r4, _r2), vget_low_f32(_coeffs), 0);
-#endif
-                float32x4_t _tmp1 = vaddq_f32(_tmp12a, _tmp12b);
-                float32x4_t _tmp2 = vsubq_f32(_tmp12a, _tmp12b);
-                float32x4_t _tmp3 = vaddq_f32(_tmp34a, _tmp34b);
-                float32x4_t _tmp4 = vsubq_f32(_tmp34a, _tmp34b);
-                float32x4_t _tmp5 = vaddq_f32(_tmp56a, _tmp56b);
-                float32x4_t _tmp6 = vsubq_f32(_tmp56a, _tmp56b);
-#if __aarch64__
-                float32x4_t _tmp7 = vfmaq_laneq_f32(vsubq_f32(_r7, _r1), vsubq_f32(_r3, _r5), _coeffs, 0);
-#else
-                float32x4_t _tmp7 = vmlaq_lane_f32(vsubq_f32(_r7, _r1), vsubq_f32(_r3, _r5), vget_low_f32(_coeffs), 0);
-#endif
-
-                vst1q_f32(p0, _tmp0);
-                vst1q_f32(p1, _tmp1);
-                vst1q_f32(p2, _tmp2);
-                vst1q_f32(p3, _tmp3);
-                vst1q_f32(p4, _tmp4);
-                vst1q_f32(p5, _tmp5);
-                vst1q_f32(p6, _tmp6);
-                vst1q_f32(p7, _tmp7);
-
-                p0 += max_jj * 8 * 4;
-                p1 += max_jj * 8 * 4;
-                p2 += max_jj * 8 * 4;
-                p3 += max_jj * 8 * 4;
-                p4 += max_jj * 8 * 4;
-                p5 += max_jj * 8 * 4;
-                p6 += max_jj * 8 * 4;
-                p7 += max_jj * 8 * 4;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 4;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-#else // __ARM_NEON
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    #pragma omp parallel for num_threads(nT)
-#endif // __ARM_NEON
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        float tmp[8][8][2];
-
-#if __ARM_NEON
-        const float coeffs[8] = {5.25f, -4.25f, -1.25f, 0.25f, -2.5f, 0.5f, 2.f, 4.f};
-        float32x4_t _coeffs = vld1q_f32(coeffs);
-        float32x4_t _coeffs2 = vld1q_f32(coeffs + 4);
-#endif
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel(k + kk).row(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 8; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vdup_n_f32(0.f);
-                float32x2_t _r1 = vdup_n_f32(0.f);
-                float32x2_t _r2 = vdup_n_f32(0.f);
-                float32x2_t _r3 = vdup_n_f32(0.f);
-                float32x2_t _r4 = vdup_n_f32(0.f);
-                float32x2_t _r5 = vdup_n_f32(0.f);
-                float32x2_t _r6 = vdup_n_f32(0.f);
-                float32x2_t _r7 = vdup_n_f32(0.f);
-#else
-                float r00 = 0.f;
-                float r01 = 0.f;
-                float r10 = 0.f;
-                float r11 = 0.f;
-                float r20 = 0.f;
-                float r21 = 0.f;
-                float r30 = 0.f;
-                float r31 = 0.f;
-                float r40 = 0.f;
-                float r41 = 0.f;
-                float r50 = 0.f;
-                float r51 = 0.f;
-                float r60 = 0.f;
-                float r61 = 0.f;
-                float r70 = 0.f;
-                float r71 = 0.f;
-#endif
-
-                if (ti * 6 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-
-#if __ARM_NEON
-                        float32x4_t _t0 = vld1q_f32(r0);
-                        float32x4_t _t1 = vld1q_f32(r1);
-                        float32x4x2_t _t01 = vzipq_f32(_t0, _t1);
-
-                        _r0 = vget_low_f32(_t01.val[0]);
-                        if (tj * 6 + 1 < w) _r1 = vget_high_f32(_t01.val[0]);
-                        if (tj * 6 + 2 < w) _r2 = vget_low_f32(_t01.val[1]);
-                        if (tj * 6 + 3 < w) _r3 = vget_high_f32(_t01.val[1]);
-                        if (tj * 6 + 4 < w)
-                        {
-                            _t0 = vld1q_f32(r0 + 4);
-                            _t1 = vld1q_f32(r1 + 4);
-                            _t01 = vzipq_f32(_t0, _t1);
-
-                            _r4 = vget_low_f32(_t01.val[0]);
-                            if (tj * 6 + 5 < w) _r5 = vget_high_f32(_t01.val[0]);
-                            if (tj * 6 + 6 < w) _r6 = vget_low_f32(_t01.val[1]);
-                            if (tj * 6 + 7 < w) _r7 = vget_high_f32(_t01.val[1]);
-                        }
-#else
-                        r00 = r0[0];
-                        r01 = r1[0];
-                        if (tj * 6 + 1 < w)
-                        {
-                            r10 = r0[1];
-                            r11 = r1[1];
-                        }
-                        if (tj * 6 + 2 < w)
-                        {
-                            r20 = r0[2];
-                            r21 = r1[2];
-                        }
-                        if (tj * 6 + 3 < w)
-                        {
-                            r30 = r0[3];
-                            r31 = r1[3];
-                        }
-                        if (tj * 6 + 4 < w)
-                        {
-                            r40 = r0[4];
-                            r41 = r1[4];
-                        }
-                        if (tj * 6 + 5 < w)
-                        {
-                            r50 = r0[5];
-                            r51 = r1[5];
-                        }
-                        if (tj * 6 + 6 < w)
-                        {
-                            r60 = r0[6];
-                            r61 = r1[6];
-                        }
-                        if (tj * 6 + 7 < w)
-                        {
-                            r70 = r0[7];
-                            r71 = r1[7];
-                        }
-#endif
-                    }
-                }
-
-#if __ARM_NEON
-#if __aarch64__
-                float32x2_t _tmp12a = vfma_laneq_f32(vadd_f32(_r2, _r6), _r4, _coeffs, 1);
-                float32x2_t _tmp12b = vfma_laneq_f32(vadd_f32(_r1, _r5), _r3, _coeffs, 1);
-                float32x2_t _tmp34a = vfma_laneq_f32(vfma_laneq_f32(_r6, _r2, _coeffs, 3), _r4, _coeffs, 2);
-                float32x2_t _tmp34b = vfma_laneq_f32(vfma_laneq_f32(vmul_laneq_f32(_r1, _coeffs2, 1), _r3, _coeffs2, 0), _r5, _coeffs2, 2);
-                float32x2_t _tmp56a = vfma_laneq_f32(_r6, vfma_laneq_f32(_r2, _r4, _coeffs, 2), _coeffs2, 3);
-                float32x2_t _tmp56b = vfma_laneq_f32(vfma_laneq_f32(vmul_laneq_f32(_r1, _coeffs2, 2), _r3, _coeffs2, 0), _r5, _coeffs2, 1);
-#else
-                float32x2_t _tmp12a = vmla_lane_f32(vadd_f32(_r2, _r6), _r4, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp12b = vmla_lane_f32(vadd_f32(_r1, _r5), _r3, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp34a = vmla_lane_f32(vmla_lane_f32(_r6, _r2, vget_high_f32(_coeffs), 1), _r4, vget_high_f32(_coeffs), 0);
-                float32x2_t _tmp34b = vmla_lane_f32(vmla_lane_f32(vmul_lane_f32(_r1, vget_low_f32(_coeffs2), 1), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_high_f32(_coeffs2), 0);
-                float32x2_t _tmp56a = vmla_lane_f32(_r6, vmla_lane_f32(_r2, _r4, vget_high_f32(_coeffs), 0), vget_high_f32(_coeffs2), 1);
-                float32x2_t _tmp56b = vmla_lane_f32(vmla_lane_f32(vmul_lane_f32(_r1, vget_high_f32(_coeffs2), 0), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_low_f32(_coeffs2), 1);
-#endif
-
-#if __aarch64__
-                float32x2_t _tmp0 = vfma_laneq_f32(vsub_f32(_r0, _r6), vsub_f32(_r4, _r2), _coeffs, 0);
-#else
-                float32x2_t _tmp0 = vmla_lane_f32(vsub_f32(_r0, _r6), vsub_f32(_r4, _r2), vget_low_f32(_coeffs), 0);
-#endif
-                float32x2_t _tmp1 = vadd_f32(_tmp12a, _tmp12b);
-                float32x2_t _tmp2 = vsub_f32(_tmp12a, _tmp12b);
-                float32x2_t _tmp3 = vadd_f32(_tmp34a, _tmp34b);
-                float32x2_t _tmp4 = vsub_f32(_tmp34a, _tmp34b);
-                float32x2_t _tmp5 = vadd_f32(_tmp56a, _tmp56b);
-                float32x2_t _tmp6 = vsub_f32(_tmp56a, _tmp56b);
-#if __aarch64__
-                float32x2_t _tmp7 = vfma_laneq_f32(vsub_f32(_r7, _r1), vsub_f32(_r3, _r5), _coeffs, 0);
-#else
-                float32x2_t _tmp7 = vmla_lane_f32(vsub_f32(_r7, _r1), vsub_f32(_r3, _r5), vget_low_f32(_coeffs), 0);
-#endif
-
-                vst1_f32(tmp[0][m], _tmp0);
-                vst1_f32(tmp[1][m], _tmp1);
-                vst1_f32(tmp[2][m], _tmp2);
-                vst1_f32(tmp[3][m], _tmp3);
-                vst1_f32(tmp[4][m], _tmp4);
-                vst1_f32(tmp[5][m], _tmp5);
-                vst1_f32(tmp[6][m], _tmp6);
-                vst1_f32(tmp[7][m], _tmp7);
-#else
-                float tmp12a0 = r20 + r60 - r40 * 4.25f;
-                float tmp12a1 = r21 + r61 - r41 * 4.25f;
-                float tmp12b0 = r10 + r50 - r30 * 4.25f;
-                float tmp12b1 = r11 + r51 - r31 * 4.25f;
-                float tmp34a0 = r60 + r20 * 0.25f - r40 * 1.25f;
-                float tmp34a1 = r61 + r21 * 0.25f - r41 * 1.25f;
-                float tmp34b0 = r10 * 0.5f - r30 * 2.5f + r50 * 2.f;
-                float tmp34b1 = r11 * 0.5f - r31 * 2.5f + r51 * 2.f;
-                float tmp56a0 = r20 * 4.f - r40 * 5.f + r60;
-                float tmp56a1 = r21 * 4.f - r41 * 5.f + r61;
-                float tmp56b0 = r10 * 2.f - r30 * 2.5f + r50 * 0.5f;
-                float tmp56b1 = r11 * 2.f - r31 * 2.5f + r51 * 0.5f;
-
-                tmp[0][m][0] = r00 - r60 + (r40 - r20) * 5.25f;
-                tmp[0][m][1] = r01 - r61 + (r41 - r21) * 5.25f;
-                tmp[1][m][0] = tmp12a0 + tmp12b0;
-                tmp[1][m][1] = tmp12a1 + tmp12b1;
-                tmp[2][m][0] = tmp12a0 - tmp12b0;
-                tmp[2][m][1] = tmp12a1 - tmp12b1;
-                tmp[3][m][0] = tmp34a0 + tmp34b0;
-                tmp[3][m][1] = tmp34a1 + tmp34b1;
-                tmp[4][m][0] = tmp34a0 - tmp34b0;
-                tmp[4][m][1] = tmp34a1 - tmp34b1;
-                tmp[5][m][0] = tmp56a0 + tmp56b0;
-                tmp[5][m][1] = tmp56a1 + tmp56b1;
-                tmp[6][m][0] = tmp56a0 - tmp56b0;
-                tmp[6][m][1] = tmp56a1 - tmp56b1;
-                tmp[7][m][0] = r70 - r10 + (r30 - r50) * 5.25f;
-                tmp[7][m][1] = r71 - r11 + (r31 - r51) * 5.25f;
-#endif
-
-                r0 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 64 + jj * 2;
-            float* p1 = p0 + max_jj * 2;
-            float* p2 = p0 + max_jj * 2 * 2;
-            float* p3 = p0 + max_jj * 2 * 3;
-            float* p4 = p0 + max_jj * 2 * 4;
-            float* p5 = p0 + max_jj * 2 * 5;
-            float* p6 = p0 + max_jj * 2 * 6;
-            float* p7 = p0 + max_jj * 2 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(tmp[m][0]);
-                float32x2_t _r1 = vld1_f32(tmp[m][1]);
-                float32x2_t _r2 = vld1_f32(tmp[m][2]);
-                float32x2_t _r3 = vld1_f32(tmp[m][3]);
-                float32x2_t _r4 = vld1_f32(tmp[m][4]);
-                float32x2_t _r5 = vld1_f32(tmp[m][5]);
-                float32x2_t _r6 = vld1_f32(tmp[m][6]);
-                float32x2_t _r7 = vld1_f32(tmp[m][7]);
-
-#if __aarch64__
-                float32x2_t _tmp12a = vfma_laneq_f32(vadd_f32(_r2, _r6), _r4, _coeffs, 1);
-                float32x2_t _tmp12b = vfma_laneq_f32(vadd_f32(_r1, _r5), _r3, _coeffs, 1);
-                float32x2_t _tmp34a = vfma_laneq_f32(vfma_laneq_f32(_r6, _r2, _coeffs, 3), _r4, _coeffs, 2);
-                float32x2_t _tmp34b = vfma_laneq_f32(vfma_laneq_f32(vmul_laneq_f32(_r1, _coeffs2, 1), _r3, _coeffs2, 0), _r5, _coeffs2, 2);
-                float32x2_t _tmp56a = vfma_laneq_f32(_r6, vfma_laneq_f32(_r2, _r4, _coeffs, 2), _coeffs2, 3);
-                float32x2_t _tmp56b = vfma_laneq_f32(vfma_laneq_f32(vmul_laneq_f32(_r1, _coeffs2, 2), _r3, _coeffs2, 0), _r5, _coeffs2, 1);
-#else
-                float32x2_t _tmp12a = vmla_lane_f32(vadd_f32(_r2, _r6), _r4, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp12b = vmla_lane_f32(vadd_f32(_r1, _r5), _r3, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp34a = vmla_lane_f32(vmla_lane_f32(_r6, _r2, vget_high_f32(_coeffs), 1), _r4, vget_high_f32(_coeffs), 0);
-                float32x2_t _tmp34b = vmla_lane_f32(vmla_lane_f32(vmul_lane_f32(_r1, vget_low_f32(_coeffs2), 1), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_high_f32(_coeffs2), 0);
-                float32x2_t _tmp56a = vmla_lane_f32(_r6, vmla_lane_f32(_r2, _r4, vget_high_f32(_coeffs), 0), vget_high_f32(_coeffs2), 1);
-                float32x2_t _tmp56b = vmla_lane_f32(vmla_lane_f32(vmul_lane_f32(_r1, vget_high_f32(_coeffs2), 0), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_low_f32(_coeffs2), 1);
-#endif
-
-#if __aarch64__
-                float32x2_t _tmp0 = vfma_laneq_f32(vsub_f32(_r0, _r6), vsub_f32(_r4, _r2), _coeffs, 0);
-#else
-                float32x2_t _tmp0 = vmla_lane_f32(vsub_f32(_r0, _r6), vsub_f32(_r4, _r2), vget_low_f32(_coeffs), 0);
-#endif
-                float32x2_t _tmp1 = vadd_f32(_tmp12a, _tmp12b);
-                float32x2_t _tmp2 = vsub_f32(_tmp12a, _tmp12b);
-                float32x2_t _tmp3 = vadd_f32(_tmp34a, _tmp34b);
-                float32x2_t _tmp4 = vsub_f32(_tmp34a, _tmp34b);
-                float32x2_t _tmp5 = vadd_f32(_tmp56a, _tmp56b);
-                float32x2_t _tmp6 = vsub_f32(_tmp56a, _tmp56b);
-#if __aarch64__
-                float32x2_t _tmp7 = vfma_laneq_f32(vsub_f32(_r7, _r1), vsub_f32(_r3, _r5), _coeffs, 0);
-#else
-                float32x2_t _tmp7 = vmla_lane_f32(vsub_f32(_r7, _r1), vsub_f32(_r3, _r5), vget_low_f32(_coeffs), 0);
-#endif
-
-                vst1_f32(p0, _tmp0);
-                vst1_f32(p1, _tmp1);
-                vst1_f32(p2, _tmp2);
-                vst1_f32(p3, _tmp3);
-                vst1_f32(p4, _tmp4);
-                vst1_f32(p5, _tmp5);
-                vst1_f32(p6, _tmp6);
-                vst1_f32(p7, _tmp7);
-#else
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-                float r40 = tmp[m][4][0];
-                float r41 = tmp[m][4][1];
-                float r50 = tmp[m][5][0];
-                float r51 = tmp[m][5][1];
-                float r60 = tmp[m][6][0];
-                float r61 = tmp[m][6][1];
-                float r70 = tmp[m][7][0];
-                float r71 = tmp[m][7][1];
-
-                float tmp12a0 = r20 + r60 - r40 * 4.25f;
-                float tmp12a1 = r21 + r61 - r41 * 4.25f;
-                float tmp12b0 = r10 + r50 - r30 * 4.25f;
-                float tmp12b1 = r11 + r51 - r31 * 4.25f;
-                float tmp34a0 = r60 + r20 * 0.25f - r40 * 1.25f;
-                float tmp34a1 = r61 + r21 * 0.25f - r41 * 1.25f;
-                float tmp34b0 = r10 * 0.5f - r30 * 2.5f + r50 * 2.f;
-                float tmp34b1 = r11 * 0.5f - r31 * 2.5f + r51 * 2.f;
-                float tmp56a0 = r20 * 4.f - r40 * 5.f + r60;
-                float tmp56a1 = r21 * 4.f - r41 * 5.f + r61;
-                float tmp56b0 = r10 * 2.f - r30 * 2.5f + r50 * 0.5f;
-                float tmp56b1 = r11 * 2.f - r31 * 2.5f + r51 * 0.5f;
-
-                p0[0] = r00 - r60 + (r40 - r20) * 5.25f;
-                p0[1] = r01 - r61 + (r41 - r21) * 5.25f;
-                p1[0] = tmp12a0 + tmp12b0;
-                p1[1] = tmp12a1 + tmp12b1;
-                p2[0] = tmp12a0 - tmp12b0;
-                p2[1] = tmp12a1 - tmp12b1;
-                p3[0] = tmp34a0 + tmp34b0;
-                p3[1] = tmp34a1 + tmp34b1;
-                p4[0] = tmp34a0 - tmp34b0;
-                p4[1] = tmp34a1 - tmp34b1;
-                p5[0] = tmp56a0 + tmp56b0;
-                p5[1] = tmp56a1 + tmp56b1;
-                p6[0] = tmp56a0 - tmp56b0;
-                p6[1] = tmp56a1 - tmp56b1;
-                p7[0] = r70 - r10 + (r30 - r50) * 5.25f;
-                p7[1] = r71 - r11 + (r31 - r51) * 5.25f;
-#endif
-
-                p0 += max_jj * 8 * 2;
-                p1 += max_jj * 8 * 2;
-                p2 += max_jj * 8 * 2;
-                p3 += max_jj * 8 * 2;
-                p4 += max_jj * 8 * 2;
-                p5 += max_jj * 8 * 2;
-                p6 += max_jj * 8 * 2;
-                p7 += max_jj * 8 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        float tmp[8][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0123 = bottom_blob.channel(k + kk).row(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 8; m++)
-            {
-                float r0 = 0.f;
-                float r1 = 0.f;
-                float r2 = 0.f;
-                float r3 = 0.f;
-                float r4 = 0.f;
-                float r5 = 0.f;
-                float r6 = 0.f;
-                float r7 = 0.f;
-
-                if (ti * 6 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = r0123[0];
-                        if (tj * 6 + 1 < w) r1 = r0123[1];
-                        if (tj * 6 + 2 < w) r2 = r0123[2];
-                        if (tj * 6 + 3 < w) r3 = r0123[3];
-                        if (tj * 6 + 4 < w) r4 = r0123[4];
-                        if (tj * 6 + 5 < w) r5 = r0123[5];
-                        if (tj * 6 + 6 < w) r6 = r0123[6];
-                        if (tj * 6 + 7 < w) r7 = r0123[7];
-                    }
-                }
-
-                float tmp12a = r2 + r6 - r4 * 4.25f;
-                float tmp12b = r1 + r5 - r3 * 4.25f;
-                float tmp34a = r6 + r2 * 0.25f - r4 * 1.25f;
-                float tmp34b = r1 * 0.5f - r3 * 2.5f + r5 * 2.f;
-                float tmp56a = r2 * 4.f - r4 * 5.f + r6;
-                float tmp56b = r1 * 2.f - r3 * 2.5f + r5 * 0.5f;
-
-                tmp[0][m] = r0 - r6 + (r4 - r2) * 5.25f;
-                tmp[1][m] = tmp12a + tmp12b;
-                tmp[2][m] = tmp12a - tmp12b;
-                tmp[3][m] = tmp34a + tmp34b;
-                tmp[4][m] = tmp34a - tmp34b;
-                tmp[5][m] = tmp56a + tmp56b;
-                tmp[6][m] = tmp56a - tmp56b;
-                tmp[7][m] = r7 - r1 + (r3 - r5) * 5.25f;
-
-                r0123 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 64 + jj;
-            float* p1 = p0 + max_jj;
-            float* p2 = p0 + max_jj * 2;
-            float* p3 = p0 + max_jj * 3;
-            float* p4 = p0 + max_jj * 4;
-            float* p5 = p0 + max_jj * 5;
-            float* p6 = p0 + max_jj * 6;
-            float* p7 = p0 + max_jj * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-                float r4 = tmp[m][4];
-                float r5 = tmp[m][5];
-                float r6 = tmp[m][6];
-                float r7 = tmp[m][7];
-
-                float tmp12a = r2 + r6 - r4 * 4.25f;
-                float tmp12b = r1 + r5 - r3 * 4.25f;
-                float tmp34a = r6 + r2 * 0.25f - r4 * 1.25f;
-                float tmp34b = r1 * 0.5f - r3 * 2.5f + r5 * 2.f;
-                float tmp56a = r2 * 4.f - r4 * 5.f + r6;
-                float tmp56b = r1 * 2.f - r3 * 2.5f + r5 * 0.5f;
-
-                p0[0] = r0 - r6 + (r4 - r2) * 5.25f;
-                p1[0] = tmp12a + tmp12b;
-                p2[0] = tmp12a - tmp12b;
-                p3[0] = tmp34a + tmp34b;
-                p4[0] = tmp34a - tmp34b;
-                p5[0] = tmp56a + tmp56b;
-                p6[0] = tmp56a - tmp56b;
-                p7[0] = r7 - r1 + (r3 - r5) * 5.25f;
-
-                p0 += max_jj * 8;
-                p1 += max_jj * 8;
-                p2 += max_jj * 8;
-                p3 += max_jj * 8;
-                p4 += max_jj * 8;
-                p5 += max_jj * 8;
-                p6 += max_jj * 8;
-                p7 += max_jj * 8;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd63_transform_output_tile(const Mat& top_tile, Mat& top_blob, const Mat& bias, int i, int max_ii, int j, int max_jj)
-{
-    // const float otm[6][8] = {
-    //     {1.0f, 1.0f,  1.0f,  1.0f,  1.0f, 32.0f, 32.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f,  2.0f, -2.0f, 16.0f,-16.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f,  4.0f,  4.0f,  8.0f,  8.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f,  8.0f, -8.0f,  4.0f, -4.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f, 16.0f, 16.0f,  2.0f,  2.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 32.0f,-32.0f,  1.0f, -1.0f, 1.0f}
-    // };
-
-#if __ARM_NEON
-    const float coeffs[4] = {32.f, 16.f, 8.f, 4.f};
-    float32x4_t _coeffs = vld1q_f32(coeffs);
-    float32x2_t _v2 = vdup_n_f32(2.f);
-#endif
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 5) / 6;
-
-    const float* biasptr = bias;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        float32x4_t _bias0 = biasptr ? vld1q_f32(biasptr + i + ii) : vdupq_n_f32(0.f);
-        float32x4_t _bias1 = biasptr ? vld1q_f32(biasptr + i + ii + 4) : vdupq_n_f32(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[6][8][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 64 + jj * 8;
-            const float* r1 = r0 + max_jj * 8;
-            const float* r2 = r0 + max_jj * 8 * 2;
-            const float* r3 = r0 + max_jj * 8 * 3;
-            const float* r4 = r0 + max_jj * 8 * 4;
-            const float* r5 = r0 + max_jj * 8 * 5;
-            const float* r6 = r0 + max_jj * 8 * 6;
-            const float* r7 = r0 + max_jj * 8 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r30 = vld1q_f32(r3);
-                float32x4_t _r31 = vld1q_f32(r3 + 4);
-                float32x4_t _r40 = vld1q_f32(r4);
-                float32x4_t _r41 = vld1q_f32(r4 + 4);
-                float32x4_t _r50 = vld1q_f32(r5);
-                float32x4_t _r51 = vld1q_f32(r5 + 4);
-                float32x4_t _r60 = vld1q_f32(r6);
-                float32x4_t _r61 = vld1q_f32(r6 + 4);
-                float32x4_t _r70 = vld1q_f32(r7);
-                float32x4_t _r71 = vld1q_f32(r7 + 4);
-
-                float32x4_t _tmp024a0 = vaddq_f32(_r10, _r20);
-                float32x4_t _tmp024a1 = vaddq_f32(_r11, _r21);
-                float32x4_t _tmp135a0 = vsubq_f32(_r10, _r20);
-                float32x4_t _tmp135a1 = vsubq_f32(_r11, _r21);
-                float32x4_t _tmp024b0 = vaddq_f32(_r30, _r40);
-                float32x4_t _tmp024b1 = vaddq_f32(_r31, _r41);
-                float32x4_t _tmp135b0 = vsubq_f32(_r30, _r40);
-                float32x4_t _tmp135b1 = vsubq_f32(_r31, _r41);
-                float32x4_t _tmp024c0 = vaddq_f32(_r50, _r60);
-                float32x4_t _tmp024c1 = vaddq_f32(_r51, _r61);
-                float32x4_t _tmp135c0 = vsubq_f32(_r50, _r60);
-                float32x4_t _tmp135c1 = vsubq_f32(_r51, _r61);
-
-                float32x4_t _tmp00 = vaddq_f32(vaddq_f32(_r00, _tmp024a0), vfmaq_laneq_f32(_tmp024b0, _tmp024c0, _coeffs, 0));
-                float32x4_t _tmp01 = vaddq_f32(vaddq_f32(_r01, _tmp024a1), vfmaq_laneq_f32(_tmp024b1, _tmp024c1, _coeffs, 0));
-                float32x4_t _tmp10 = vfmaq_laneq_f32(vfmaq_lane_f32(_tmp135a0, _tmp135b0, _v2, 0), _tmp135c0, _coeffs, 1);
-                float32x4_t _tmp11 = vfmaq_laneq_f32(vfmaq_lane_f32(_tmp135a1, _tmp135b1, _v2, 0), _tmp135c1, _coeffs, 1);
-                float32x4_t _tmp20 = vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp024a0, _tmp024b0, _coeffs, 3), _tmp024c0, _coeffs, 2);
-                float32x4_t _tmp21 = vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp024a1, _tmp024b1, _coeffs, 3), _tmp024c1, _coeffs, 2);
-                float32x4_t _tmp30 = vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp135a0, _tmp135b0, _coeffs, 2), _tmp135c0, _coeffs, 3);
-                float32x4_t _tmp31 = vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp135a1, _tmp135b1, _coeffs, 2), _tmp135c1, _coeffs, 3);
-                float32x4_t _tmp40 = vfmaq_lane_f32(vfmaq_laneq_f32(_tmp024a0, _tmp024b0, _coeffs, 1), _tmp024c0, _v2, 0);
-                float32x4_t _tmp41 = vfmaq_lane_f32(vfmaq_laneq_f32(_tmp024a1, _tmp024b1, _coeffs, 1), _tmp024c1, _v2, 0);
-                float32x4_t _tmp50 = vaddq_f32(vaddq_f32(_r70, _tmp135a0), vfmaq_laneq_f32(_tmp135c0, _tmp135b0, _coeffs, 0));
-                float32x4_t _tmp51 = vaddq_f32(vaddq_f32(_r71, _tmp135a1), vfmaq_laneq_f32(_tmp135c1, _tmp135b1, _coeffs, 0));
-
-                vst1q_f32(tmp[0][m], _tmp00);
-                vst1q_f32(tmp[0][m] + 4, _tmp01);
-                vst1q_f32(tmp[1][m], _tmp10);
-                vst1q_f32(tmp[1][m] + 4, _tmp11);
-                vst1q_f32(tmp[2][m], _tmp20);
-                vst1q_f32(tmp[2][m] + 4, _tmp21);
-                vst1q_f32(tmp[3][m], _tmp30);
-                vst1q_f32(tmp[3][m] + 4, _tmp31);
-                vst1q_f32(tmp[4][m], _tmp40);
-                vst1q_f32(tmp[4][m] + 4, _tmp41);
-                vst1q_f32(tmp[5][m], _tmp50);
-                vst1q_f32(tmp[5][m] + 4, _tmp51);
-
-                r0 += max_jj * 8 * 8;
-                r1 += max_jj * 8 * 8;
-                r2 += max_jj * 8 * 8;
-                r3 += max_jj * 8 * 8;
-                r4 += max_jj * 8 * 8;
-                r5 += max_jj * 8 * 8;
-                r6 += max_jj * 8 * 8;
-                r7 += max_jj * 8 * 8;
-            }
-
-            float* outptr0 = top_blob.channel((i + ii) / out_elempack).row(ti * 6) + (tj * 6) * out_elempack;
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-                float32x4_t _r00 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r01 = vld1q_f32(tmp[m][0] + 4);
-                float32x4_t _r10 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r11 = vld1q_f32(tmp[m][1] + 4);
-                float32x4_t _r20 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r21 = vld1q_f32(tmp[m][2] + 4);
-                float32x4_t _r30 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r31 = vld1q_f32(tmp[m][3] + 4);
-                float32x4_t _r40 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r41 = vld1q_f32(tmp[m][4] + 4);
-                float32x4_t _r50 = vld1q_f32(tmp[m][5]);
-                float32x4_t _r51 = vld1q_f32(tmp[m][5] + 4);
-                float32x4_t _r60 = vld1q_f32(tmp[m][6]);
-                float32x4_t _r61 = vld1q_f32(tmp[m][6] + 4);
-                float32x4_t _r70 = vld1q_f32(tmp[m][7]);
-                float32x4_t _r71 = vld1q_f32(tmp[m][7] + 4);
-
-                float32x4_t _tmp024a0 = vaddq_f32(_r10, _r20);
-                float32x4_t _tmp024a1 = vaddq_f32(_r11, _r21);
-                float32x4_t _tmp135a0 = vsubq_f32(_r10, _r20);
-                float32x4_t _tmp135a1 = vsubq_f32(_r11, _r21);
-                float32x4_t _tmp024b0 = vaddq_f32(_r30, _r40);
-                float32x4_t _tmp024b1 = vaddq_f32(_r31, _r41);
-                float32x4_t _tmp135b0 = vsubq_f32(_r30, _r40);
-                float32x4_t _tmp135b1 = vsubq_f32(_r31, _r41);
-                float32x4_t _tmp024c0 = vaddq_f32(_r50, _r60);
-                float32x4_t _tmp024c1 = vaddq_f32(_r51, _r61);
-                float32x4_t _tmp135c0 = vsubq_f32(_r50, _r60);
-                float32x4_t _tmp135c1 = vsubq_f32(_r51, _r61);
-
-                float32x4_t _tmp00 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r00, _tmp024a0), vfmaq_laneq_f32(_tmp024b0, _tmp024c0, _coeffs, 0)));
-                float32x4_t _tmp01 = vaddq_f32(_bias1, vaddq_f32(vaddq_f32(_r01, _tmp024a1), vfmaq_laneq_f32(_tmp024b1, _tmp024c1, _coeffs, 0)));
-                float32x4_t _tmp10 = vaddq_f32(_bias0, vfmaq_laneq_f32(vfmaq_lane_f32(_tmp135a0, _tmp135b0, _v2, 0), _tmp135c0, _coeffs, 1));
-                float32x4_t _tmp11 = vaddq_f32(_bias1, vfmaq_laneq_f32(vfmaq_lane_f32(_tmp135a1, _tmp135b1, _v2, 0), _tmp135c1, _coeffs, 1));
-                float32x4_t _tmp20 = vaddq_f32(_bias0, vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp024a0, _tmp024b0, _coeffs, 3), _tmp024c0, _coeffs, 2));
-                float32x4_t _tmp21 = vaddq_f32(_bias1, vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp024a1, _tmp024b1, _coeffs, 3), _tmp024c1, _coeffs, 2));
-                float32x4_t _tmp30 = vaddq_f32(_bias0, vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp135a0, _tmp135b0, _coeffs, 2), _tmp135c0, _coeffs, 3));
-                float32x4_t _tmp31 = vaddq_f32(_bias1, vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp135a1, _tmp135b1, _coeffs, 2), _tmp135c1, _coeffs, 3));
-                float32x4_t _tmp40 = vaddq_f32(_bias0, vfmaq_lane_f32(vfmaq_laneq_f32(_tmp024a0, _tmp024b0, _coeffs, 1), _tmp024c0, _v2, 0));
-                float32x4_t _tmp41 = vaddq_f32(_bias1, vfmaq_lane_f32(vfmaq_laneq_f32(_tmp024a1, _tmp024b1, _coeffs, 1), _tmp024c1, _v2, 0));
-                float32x4_t _tmp50 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r70, _tmp135a0), vfmaq_laneq_f32(_tmp135c0, _tmp135b0, _coeffs, 0)));
-                float32x4_t _tmp51 = vaddq_f32(_bias1, vaddq_f32(vaddq_f32(_r71, _tmp135a1), vfmaq_laneq_f32(_tmp135c1, _tmp135b1, _coeffs, 0)));
-
-                if (out_elempack == 4)
-                {
-                    float* outptr1 = outptr0 + N;
-
-                    vst1q_f32(outptr0, _tmp00);
-                    vst1q_f32(outptr1, _tmp01);
-                    if (tj * 6 + 1 < outw)
-                    {
-                        vst1q_f32(outptr0 + 4, _tmp10);
-                        vst1q_f32(outptr1 + 4, _tmp11);
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        vst1q_f32(outptr0 + 8, _tmp20);
-                        vst1q_f32(outptr1 + 8, _tmp21);
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        vst1q_f32(outptr0 + 12, _tmp30);
-                        vst1q_f32(outptr1 + 12, _tmp31);
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        vst1q_f32(outptr0 + 16, _tmp40);
-                        vst1q_f32(outptr1 + 16, _tmp41);
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        vst1q_f32(outptr0 + 20, _tmp50);
-                        vst1q_f32(outptr1 + 20, _tmp51);
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    float tmp0[8];
-                    float tmp1[8];
-                    float tmp2[8];
-                    float tmp3[8];
-                    float tmp4[8];
-                    float tmp5[8];
-                    vst1q_f32(tmp0, _tmp00);
-                    vst1q_f32(tmp0 + 4, _tmp01);
-                    vst1q_f32(tmp1, _tmp10);
-                    vst1q_f32(tmp1 + 4, _tmp11);
-                    vst1q_f32(tmp2, _tmp20);
-                    vst1q_f32(tmp2 + 4, _tmp21);
-                    vst1q_f32(tmp3, _tmp30);
-                    vst1q_f32(tmp3 + 4, _tmp31);
-                    vst1q_f32(tmp4, _tmp40);
-                    vst1q_f32(tmp4 + 4, _tmp41);
-                    vst1q_f32(tmp5, _tmp50);
-                    vst1q_f32(tmp5 + 4, _tmp51);
-
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-                    float* outptr4 = outptr0 + N * 4;
-                    float* outptr5 = outptr0 + N * 5;
-                    float* outptr6 = outptr0 + N * 6;
-                    float* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-                    if (tj * 6 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                        outptr4[2] = tmp2[4];
-                        outptr5[2] = tmp2[5];
-                        outptr6[2] = tmp2[6];
-                        outptr7[2] = tmp2[7];
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                        outptr4[3] = tmp3[4];
-                        outptr5[3] = tmp3[5];
-                        outptr6[3] = tmp3[6];
-                        outptr7[3] = tmp3[7];
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        outptr0[4] = tmp4[0];
-                        outptr1[4] = tmp4[1];
-                        outptr2[4] = tmp4[2];
-                        outptr3[4] = tmp4[3];
-                        outptr4[4] = tmp4[4];
-                        outptr5[4] = tmp4[5];
-                        outptr6[4] = tmp4[6];
-                        outptr7[4] = tmp4[7];
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        outptr0[5] = tmp5[0];
-                        outptr1[5] = tmp5[1];
-                        outptr2[5] = tmp5[2];
-                        outptr3[5] = tmp5[3];
-                        outptr4[5] = tmp5[4];
-                        outptr5[5] = tmp5[5];
-                        outptr6[5] = tmp5[6];
-                        outptr7[5] = tmp5[7];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        float32x4_t _bias0 = biasptr ? vld1q_f32(biasptr + i + ii) : vdupq_n_f32(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[6][8][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 64 + jj * 4;
-            const float* r1 = r0 + max_jj * 4;
-            const float* r2 = r0 + max_jj * 4 * 2;
-            const float* r3 = r0 + max_jj * 4 * 3;
-            const float* r4 = r0 + max_jj * 4 * 4;
-            const float* r5 = r0 + max_jj * 4 * 5;
-            const float* r6 = r0 + max_jj * 4 * 6;
-            const float* r7 = r0 + max_jj * 4 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float32x4_t _r0 = vld1q_f32(r0);
-                float32x4_t _r1 = vld1q_f32(r1);
-                float32x4_t _r2 = vld1q_f32(r2);
-                float32x4_t _r3 = vld1q_f32(r3);
-                float32x4_t _r4 = vld1q_f32(r4);
-                float32x4_t _r5 = vld1q_f32(r5);
-                float32x4_t _r6 = vld1q_f32(r6);
-                float32x4_t _r7 = vld1q_f32(r7);
-
-                float32x4_t _tmp024a = vaddq_f32(_r1, _r2);
-                float32x4_t _tmp135a = vsubq_f32(_r1, _r2);
-                float32x4_t _tmp024b = vaddq_f32(_r3, _r4);
-                float32x4_t _tmp135b = vsubq_f32(_r3, _r4);
-                float32x4_t _tmp024c = vaddq_f32(_r5, _r6);
-                float32x4_t _tmp135c = vsubq_f32(_r5, _r6);
-
-#if __aarch64__
-                float32x4_t _tmp0 = vaddq_f32(vaddq_f32(_r0, _tmp024a), vfmaq_laneq_f32(_tmp024b, _tmp024c, _coeffs, 0));
-                float32x4_t _tmp1 = vfmaq_laneq_f32(vfmaq_lane_f32(_tmp135a, _tmp135b, _v2, 0), _tmp135c, _coeffs, 1);
-                float32x4_t _tmp2 = vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp024a, _tmp024b, _coeffs, 3), _tmp024c, _coeffs, 2);
-                float32x4_t _tmp3 = vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp135a, _tmp135b, _coeffs, 2), _tmp135c, _coeffs, 3);
-                float32x4_t _tmp4 = vfmaq_lane_f32(vfmaq_laneq_f32(_tmp024a, _tmp024b, _coeffs, 1), _tmp024c, _v2, 0);
-                float32x4_t _tmp5 = vaddq_f32(vaddq_f32(_r7, _tmp135a), vfmaq_laneq_f32(_tmp135c, _tmp135b, _coeffs, 0));
-#else
-                float32x4_t _tmp0 = vaddq_f32(vaddq_f32(_r0, _tmp024a), vmlaq_lane_f32(_tmp024b, _tmp024c, vget_low_f32(_coeffs), 0));
-                float32x4_t _tmp1 = vmlaq_lane_f32(vmlaq_lane_f32(_tmp135a, _tmp135b, _v2, 0), _tmp135c, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp2 = vmlaq_lane_f32(vmlaq_lane_f32(_tmp024a, _tmp024b, vget_high_f32(_coeffs), 1), _tmp024c, vget_high_f32(_coeffs), 0);
-                float32x4_t _tmp3 = vmlaq_lane_f32(vmlaq_lane_f32(_tmp135a, _tmp135b, vget_high_f32(_coeffs), 0), _tmp135c, vget_high_f32(_coeffs), 1);
-                float32x4_t _tmp4 = vmlaq_lane_f32(vmlaq_lane_f32(_tmp024a, _tmp024b, vget_low_f32(_coeffs), 1), _tmp024c, _v2, 0);
-                float32x4_t _tmp5 = vaddq_f32(vaddq_f32(_r7, _tmp135a), vmlaq_lane_f32(_tmp135c, _tmp135b, vget_low_f32(_coeffs), 0));
-#endif
-
-                vst1q_f32(tmp[0][m], _tmp0);
-                vst1q_f32(tmp[1][m], _tmp1);
-                vst1q_f32(tmp[2][m], _tmp2);
-                vst1q_f32(tmp[3][m], _tmp3);
-                vst1q_f32(tmp[4][m], _tmp4);
-                vst1q_f32(tmp[5][m], _tmp5);
-
-                r0 += max_jj * 8 * 4;
-                r1 += max_jj * 8 * 4;
-                r2 += max_jj * 8 * 4;
-                r3 += max_jj * 8 * 4;
-                r4 += max_jj * 8 * 4;
-                r5 += max_jj * 8 * 4;
-                r6 += max_jj * 8 * 4;
-                r7 += max_jj * 8 * 4;
-            }
-
-            float* outptr0 = top_blob.channel((i + ii) / out_elempack).row(ti * 6) + (tj * 6) * out_elempack;
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-                float32x4_t _r0 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r1 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r2 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r3 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r4 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r5 = vld1q_f32(tmp[m][5]);
-                float32x4_t _r6 = vld1q_f32(tmp[m][6]);
-                float32x4_t _r7 = vld1q_f32(tmp[m][7]);
-
-                float32x4_t _tmp024a = vaddq_f32(_r1, _r2);
-                float32x4_t _tmp135a = vsubq_f32(_r1, _r2);
-                float32x4_t _tmp024b = vaddq_f32(_r3, _r4);
-                float32x4_t _tmp135b = vsubq_f32(_r3, _r4);
-                float32x4_t _tmp024c = vaddq_f32(_r5, _r6);
-                float32x4_t _tmp135c = vsubq_f32(_r5, _r6);
-
-#if __aarch64__
-                float32x4_t _tmp0 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r0, _tmp024a), vfmaq_laneq_f32(_tmp024b, _tmp024c, _coeffs, 0)));
-                float32x4_t _tmp1 = vaddq_f32(_bias0, vfmaq_laneq_f32(vfmaq_lane_f32(_tmp135a, _tmp135b, _v2, 0), _tmp135c, _coeffs, 1));
-                float32x4_t _tmp2 = vaddq_f32(_bias0, vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp024a, _tmp024b, _coeffs, 3), _tmp024c, _coeffs, 2));
-                float32x4_t _tmp3 = vaddq_f32(_bias0, vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp135a, _tmp135b, _coeffs, 2), _tmp135c, _coeffs, 3));
-                float32x4_t _tmp4 = vaddq_f32(_bias0, vfmaq_lane_f32(vfmaq_laneq_f32(_tmp024a, _tmp024b, _coeffs, 1), _tmp024c, _v2, 0));
-                float32x4_t _tmp5 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r7, _tmp135a), vfmaq_laneq_f32(_tmp135c, _tmp135b, _coeffs, 0)));
-#else
-                float32x4_t _tmp0 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r0, _tmp024a), vmlaq_lane_f32(_tmp024b, _tmp024c, vget_low_f32(_coeffs), 0)));
-                float32x4_t _tmp1 = vaddq_f32(_bias0, vmlaq_lane_f32(vmlaq_lane_f32(_tmp135a, _tmp135b, _v2, 0), _tmp135c, vget_low_f32(_coeffs), 1));
-                float32x4_t _tmp2 = vaddq_f32(_bias0, vmlaq_lane_f32(vmlaq_lane_f32(_tmp024a, _tmp024b, vget_high_f32(_coeffs), 1), _tmp024c, vget_high_f32(_coeffs), 0));
-                float32x4_t _tmp3 = vaddq_f32(_bias0, vmlaq_lane_f32(vmlaq_lane_f32(_tmp135a, _tmp135b, vget_high_f32(_coeffs), 0), _tmp135c, vget_high_f32(_coeffs), 1));
-                float32x4_t _tmp4 = vaddq_f32(_bias0, vmlaq_lane_f32(vmlaq_lane_f32(_tmp024a, _tmp024b, vget_low_f32(_coeffs), 1), _tmp024c, _v2, 0));
-                float32x4_t _tmp5 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r7, _tmp135a), vmlaq_lane_f32(_tmp135c, _tmp135b, vget_low_f32(_coeffs), 0)));
-#endif
-
-                if (out_elempack == 4)
-                {
-                    vst1q_f32(outptr0, _tmp0);
-                    if (tj * 6 + 1 < outw) vst1q_f32(outptr0 + 4, _tmp1);
-                    if (tj * 6 + 2 < outw) vst1q_f32(outptr0 + 8, _tmp2);
-                    if (tj * 6 + 3 < outw) vst1q_f32(outptr0 + 12, _tmp3);
-                    if (tj * 6 + 4 < outw) vst1q_f32(outptr0 + 16, _tmp4);
-                    if (tj * 6 + 5 < outw) vst1q_f32(outptr0 + 20, _tmp5);
-                }
-                if (out_elempack == 1)
-                {
-                    float tmp0[4];
-                    float tmp1[4];
-                    float tmp2[4];
-                    float tmp3[4];
-                    float tmp4[4];
-                    float tmp5[4];
-                    vst1q_f32(tmp0, _tmp0);
-                    vst1q_f32(tmp1, _tmp1);
-                    vst1q_f32(tmp2, _tmp2);
-                    vst1q_f32(tmp3, _tmp3);
-                    vst1q_f32(tmp4, _tmp4);
-                    vst1q_f32(tmp5, _tmp5);
-
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    if (tj * 6 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        outptr0[4] = tmp4[0];
-                        outptr1[4] = tmp4[1];
-                        outptr2[4] = tmp4[2];
-                        outptr3[4] = tmp4[3];
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        outptr0[5] = tmp5[0];
-                        outptr1[5] = tmp5[1];
-                        outptr2[5] = tmp5[2];
-                        outptr3[5] = tmp5[3];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-#if __ARM_NEON
-        float32x2_t _bias0 = biasptr ? vld1_f32(biasptr + i + ii) : vdup_n_f32(0.f);
-#else
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-        float bias1 = biasptr ? biasptr[i + ii + 1] : 0.f;
-#endif
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        float tmp[6][8][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 64 + jj * 2;
-            const float* r1 = r0 + max_jj * 2;
-            const float* r2 = r0 + max_jj * 2 * 2;
-            const float* r3 = r0 + max_jj * 2 * 3;
-            const float* r4 = r0 + max_jj * 2 * 4;
-            const float* r5 = r0 + max_jj * 2 * 5;
-            const float* r6 = r0 + max_jj * 2 * 6;
-            const float* r7 = r0 + max_jj * 2 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(r0);
-                float32x2_t _r1 = vld1_f32(r1);
-                float32x2_t _r2 = vld1_f32(r2);
-                float32x2_t _r3 = vld1_f32(r3);
-                float32x2_t _r4 = vld1_f32(r4);
-                float32x2_t _r5 = vld1_f32(r5);
-                float32x2_t _r6 = vld1_f32(r6);
-                float32x2_t _r7 = vld1_f32(r7);
-
-                float32x2_t _tmp024a = vadd_f32(_r1, _r2);
-                float32x2_t _tmp135a = vsub_f32(_r1, _r2);
-                float32x2_t _tmp024b = vadd_f32(_r3, _r4);
-                float32x2_t _tmp135b = vsub_f32(_r3, _r4);
-                float32x2_t _tmp024c = vadd_f32(_r5, _r6);
-                float32x2_t _tmp135c = vsub_f32(_r5, _r6);
-
-#if __aarch64__
-                float32x2_t _tmp0 = vadd_f32(vadd_f32(_r0, _tmp024a), vfma_laneq_f32(_tmp024b, _tmp024c, _coeffs, 0));
-                float32x2_t _tmp1 = vfma_laneq_f32(vfma_f32(_tmp135a, _tmp135b, _v2), _tmp135c, _coeffs, 1);
-                float32x2_t _tmp2 = vfma_laneq_f32(vfma_laneq_f32(_tmp024a, _tmp024b, _coeffs, 3), _tmp024c, _coeffs, 2);
-                float32x2_t _tmp3 = vfma_laneq_f32(vfma_laneq_f32(_tmp135a, _tmp135b, _coeffs, 2), _tmp135c, _coeffs, 3);
-                float32x2_t _tmp4 = vfma_f32(vfma_laneq_f32(_tmp024a, _tmp024b, _coeffs, 1), _tmp024c, _v2);
-                float32x2_t _tmp5 = vadd_f32(vadd_f32(_r7, _tmp135a), vfma_laneq_f32(_tmp135c, _tmp135b, _coeffs, 0));
-#else
-                float32x2_t _tmp0 = vadd_f32(vadd_f32(_r0, _tmp024a), vmla_lane_f32(_tmp024b, _tmp024c, vget_low_f32(_coeffs), 0));
-                float32x2_t _tmp1 = vmla_lane_f32(vmla_f32(_tmp135a, _tmp135b, _v2), _tmp135c, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp2 = vmla_lane_f32(vmla_lane_f32(_tmp024a, _tmp024b, vget_high_f32(_coeffs), 1), _tmp024c, vget_high_f32(_coeffs), 0);
-                float32x2_t _tmp3 = vmla_lane_f32(vmla_lane_f32(_tmp135a, _tmp135b, vget_high_f32(_coeffs), 0), _tmp135c, vget_high_f32(_coeffs), 1);
-                float32x2_t _tmp4 = vmla_f32(vmla_lane_f32(_tmp024a, _tmp024b, vget_low_f32(_coeffs), 1), _tmp024c, _v2);
-                float32x2_t _tmp5 = vadd_f32(vadd_f32(_r7, _tmp135a), vmla_lane_f32(_tmp135c, _tmp135b, vget_low_f32(_coeffs), 0));
-#endif
-
-                vst1_f32(tmp[0][m], _tmp0);
-                vst1_f32(tmp[1][m], _tmp1);
-                vst1_f32(tmp[2][m], _tmp2);
-                vst1_f32(tmp[3][m], _tmp3);
-                vst1_f32(tmp[4][m], _tmp4);
-                vst1_f32(tmp[5][m], _tmp5);
-#else
-                float tmp024a0 = r1[0] + r2[0];
-                float tmp024a1 = r1[1] + r2[1];
-                float tmp135a0 = r1[0] - r2[0];
-                float tmp135a1 = r1[1] - r2[1];
-                float tmp024b0 = r3[0] + r4[0];
-                float tmp024b1 = r3[1] + r4[1];
-                float tmp135b0 = r3[0] - r4[0];
-                float tmp135b1 = r3[1] - r4[1];
-                float tmp024c0 = r5[0] + r6[0];
-                float tmp024c1 = r5[1] + r6[1];
-                float tmp135c0 = r5[0] - r6[0];
-                float tmp135c1 = r5[1] - r6[1];
-
-                tmp[0][m][0] = r0[0] + tmp024a0 + tmp024b0 + tmp024c0 * 32;
-                tmp[0][m][1] = r0[1] + tmp024a1 + tmp024b1 + tmp024c1 * 32;
-                tmp[1][m][0] = tmp135a0 + tmp135b0 + tmp135b0 + tmp135c0 * 16;
-                tmp[1][m][1] = tmp135a1 + tmp135b1 + tmp135b1 + tmp135c1 * 16;
-                tmp[2][m][0] = tmp024a0 + tmp024b0 * 4 + tmp024c0 * 8;
-                tmp[2][m][1] = tmp024a1 + tmp024b1 * 4 + tmp024c1 * 8;
-                tmp[3][m][0] = tmp135a0 + tmp135b0 * 8 + tmp135c0 * 4;
-                tmp[3][m][1] = tmp135a1 + tmp135b1 * 8 + tmp135c1 * 4;
-                tmp[4][m][0] = tmp024a0 + tmp024b0 * 16 + tmp024c0 + tmp024c0;
-                tmp[4][m][1] = tmp024a1 + tmp024b1 * 16 + tmp024c1 + tmp024c1;
-                tmp[5][m][0] = r7[0] + tmp135a0 + tmp135b0 * 32 + tmp135c0;
-                tmp[5][m][1] = r7[1] + tmp135a1 + tmp135b1 * 32 + tmp135c1;
-#endif
-
-                r0 += max_jj * 8 * 2;
-                r1 += max_jj * 8 * 2;
-                r2 += max_jj * 8 * 2;
-                r3 += max_jj * 8 * 2;
-                r4 += max_jj * 8 * 2;
-                r5 += max_jj * 8 * 2;
-                r6 += max_jj * 8 * 2;
-                r7 += max_jj * 8 * 2;
-            }
-
-            float* outptr0 = top_blob.channel(i + ii).row(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(tmp[m][0]);
-                float32x2_t _r1 = vld1_f32(tmp[m][1]);
-                float32x2_t _r2 = vld1_f32(tmp[m][2]);
-                float32x2_t _r3 = vld1_f32(tmp[m][3]);
-                float32x2_t _r4 = vld1_f32(tmp[m][4]);
-                float32x2_t _r5 = vld1_f32(tmp[m][5]);
-                float32x2_t _r6 = vld1_f32(tmp[m][6]);
-                float32x2_t _r7 = vld1_f32(tmp[m][7]);
-
-                float32x2_t _tmp024a = vadd_f32(_r1, _r2);
-                float32x2_t _tmp135a = vsub_f32(_r1, _r2);
-                float32x2_t _tmp024b = vadd_f32(_r3, _r4);
-                float32x2_t _tmp135b = vsub_f32(_r3, _r4);
-                float32x2_t _tmp024c = vadd_f32(_r5, _r6);
-                float32x2_t _tmp135c = vsub_f32(_r5, _r6);
-
-#if __aarch64__
-                float32x2_t _tmp0 = vadd_f32(_bias0, vadd_f32(vadd_f32(_r0, _tmp024a), vfma_laneq_f32(_tmp024b, _tmp024c, _coeffs, 0)));
-                float32x2_t _tmp1 = vadd_f32(_bias0, vfma_laneq_f32(vfma_f32(_tmp135a, _tmp135b, _v2), _tmp135c, _coeffs, 1));
-                float32x2_t _tmp2 = vadd_f32(_bias0, vfma_laneq_f32(vfma_laneq_f32(_tmp024a, _tmp024b, _coeffs, 3), _tmp024c, _coeffs, 2));
-                float32x2_t _tmp3 = vadd_f32(_bias0, vfma_laneq_f32(vfma_laneq_f32(_tmp135a, _tmp135b, _coeffs, 2), _tmp135c, _coeffs, 3));
-                float32x2_t _tmp4 = vadd_f32(_bias0, vfma_f32(vfma_laneq_f32(_tmp024a, _tmp024b, _coeffs, 1), _tmp024c, _v2));
-                float32x2_t _tmp5 = vadd_f32(_bias0, vadd_f32(vadd_f32(_r7, _tmp135a), vfma_laneq_f32(_tmp135c, _tmp135b, _coeffs, 0)));
-#else
-                float32x2_t _tmp0 = vadd_f32(_bias0, vadd_f32(vadd_f32(_r0, _tmp024a), vmla_lane_f32(_tmp024b, _tmp024c, vget_low_f32(_coeffs), 0)));
-                float32x2_t _tmp1 = vadd_f32(_bias0, vmla_lane_f32(vmla_f32(_tmp135a, _tmp135b, _v2), _tmp135c, vget_low_f32(_coeffs), 1));
-                float32x2_t _tmp2 = vadd_f32(_bias0, vmla_lane_f32(vmla_lane_f32(_tmp024a, _tmp024b, vget_high_f32(_coeffs), 1), _tmp024c, vget_high_f32(_coeffs), 0));
-                float32x2_t _tmp3 = vadd_f32(_bias0, vmla_lane_f32(vmla_lane_f32(_tmp135a, _tmp135b, vget_high_f32(_coeffs), 0), _tmp135c, vget_high_f32(_coeffs), 1));
-                float32x2_t _tmp4 = vadd_f32(_bias0, vmla_f32(vmla_lane_f32(_tmp024a, _tmp024b, vget_low_f32(_coeffs), 1), _tmp024c, _v2));
-                float32x2_t _tmp5 = vadd_f32(_bias0, vadd_f32(vadd_f32(_r7, _tmp135a), vmla_lane_f32(_tmp135c, _tmp135b, vget_low_f32(_coeffs), 0)));
-#endif
-#else
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-                float r40 = tmp[m][4][0];
-                float r41 = tmp[m][4][1];
-                float r50 = tmp[m][5][0];
-                float r51 = tmp[m][5][1];
-                float r60 = tmp[m][6][0];
-                float r61 = tmp[m][6][1];
-                float r70 = tmp[m][7][0];
-                float r71 = tmp[m][7][1];
-
-                float tmp024a0 = r10 + r20;
-                float tmp024a1 = r11 + r21;
-                float tmp135a0 = r10 - r20;
-                float tmp135a1 = r11 - r21;
-                float tmp024b0 = r30 + r40;
-                float tmp024b1 = r31 + r41;
-                float tmp135b0 = r30 - r40;
-                float tmp135b1 = r31 - r41;
-                float tmp024c0 = r50 + r60;
-                float tmp024c1 = r51 + r61;
-                float tmp135c0 = r50 - r60;
-                float tmp135c1 = r51 - r61;
-
-                float tmp00 = bias0 + r00 + tmp024a0 + tmp024b0 + tmp024c0 * 32;
-                float tmp01 = bias1 + r01 + tmp024a1 + tmp024b1 + tmp024c1 * 32;
-                float tmp10 = bias0 + tmp135a0 + tmp135b0 + tmp135b0 + tmp135c0 * 16;
-                float tmp11 = bias1 + tmp135a1 + tmp135b1 + tmp135b1 + tmp135c1 * 16;
-                float tmp20 = bias0 + tmp024a0 + tmp024b0 * 4 + tmp024c0 * 8;
-                float tmp21 = bias1 + tmp024a1 + tmp024b1 * 4 + tmp024c1 * 8;
-                float tmp30 = bias0 + tmp135a0 + tmp135b0 * 8 + tmp135c0 * 4;
-                float tmp31 = bias1 + tmp135a1 + tmp135b1 * 8 + tmp135c1 * 4;
-                float tmp40 = bias0 + tmp024a0 + tmp024b0 * 16 + tmp024c0 + tmp024c0;
-                float tmp41 = bias1 + tmp024a1 + tmp024b1 * 16 + tmp024c1 + tmp024c1;
-                float tmp50 = bias0 + r70 + tmp135a0 + tmp135b0 * 32 + tmp135c0;
-                float tmp51 = bias1 + r71 + tmp135a1 + tmp135b1 * 32 + tmp135c1;
-#endif
-
-                // if (out_elempack == 1)
-                {
-                    float* outptr1 = outptr0 + N;
-
-#if __ARM_NEON
-                    outptr0[0] = vget_lane_f32(_tmp0, 0);
-                    outptr1[0] = vget_lane_f32(_tmp0, 1);
-                    if (tj * 6 + 1 < outw)
-                    {
-                        outptr0[1] = vget_lane_f32(_tmp1, 0);
-                        outptr1[1] = vget_lane_f32(_tmp1, 1);
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        outptr0[2] = vget_lane_f32(_tmp2, 0);
-                        outptr1[2] = vget_lane_f32(_tmp2, 1);
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        outptr0[3] = vget_lane_f32(_tmp3, 0);
-                        outptr1[3] = vget_lane_f32(_tmp3, 1);
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        outptr0[4] = vget_lane_f32(_tmp4, 0);
-                        outptr1[4] = vget_lane_f32(_tmp4, 1);
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        outptr0[5] = vget_lane_f32(_tmp5, 0);
-                        outptr1[5] = vget_lane_f32(_tmp5, 1);
-                    }
-#else
-                    outptr0[0] = tmp00;
-                    outptr1[0] = tmp01;
-                    if (tj * 6 + 1 < outw)
-                    {
-                        outptr0[1] = tmp10;
-                        outptr1[1] = tmp11;
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        outptr0[2] = tmp20;
-                        outptr1[2] = tmp21;
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        outptr0[3] = tmp30;
-                        outptr1[3] = tmp31;
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        outptr0[4] = tmp40;
-                        outptr1[4] = tmp41;
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        outptr0[5] = tmp50;
-                        outptr1[5] = tmp51;
-                    }
-#endif
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-
-        float tmp[6][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 64 + jj;
-            const float* r1 = r0 + max_jj;
-            const float* r2 = r0 + max_jj * 2;
-            const float* r3 = r0 + max_jj * 3;
-            const float* r4 = r0 + max_jj * 4;
-            const float* r5 = r0 + max_jj * 5;
-            const float* r6 = r0 + max_jj * 6;
-            const float* r7 = r0 + max_jj * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float tmp024a = r1[0] + r2[0];
-                float tmp135a = r1[0] - r2[0];
-                float tmp024b = r3[0] + r4[0];
-                float tmp135b = r3[0] - r4[0];
-                float tmp024c = r5[0] + r6[0];
-                float tmp135c = r5[0] - r6[0];
-
-                tmp[0][m] = r0[0] + tmp024a + tmp024b + tmp024c * 32;
-                tmp[1][m] = tmp135a + tmp135b + tmp135b + tmp135c * 16;
-                tmp[2][m] = tmp024a + tmp024b * 4 + tmp024c * 8;
-                tmp[3][m] = tmp135a + tmp135b * 8 + tmp135c * 4;
-                tmp[4][m] = tmp024a + tmp024b * 16 + tmp024c + tmp024c;
-                tmp[5][m] = r7[0] + tmp135a + tmp135b * 32 + tmp135c;
-
-                r0 += max_jj * 8;
-                r1 += max_jj * 8;
-                r2 += max_jj * 8;
-                r3 += max_jj * 8;
-                r4 += max_jj * 8;
-                r5 += max_jj * 8;
-                r6 += max_jj * 8;
-                r7 += max_jj * 8;
-            }
-
-            float* outptr0 = top_blob.channel(i + ii).row(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-                float r4 = tmp[m][4];
-                float r5 = tmp[m][5];
-                float r6 = tmp[m][6];
-                float r7 = tmp[m][7];
-
-                float tmp024a = r1 + r2;
-                float tmp135a = r1 - r2;
-                float tmp024b = r3 + r4;
-                float tmp135b = r3 - r4;
-                float tmp024c = r5 + r6;
-                float tmp135c = r5 - r6;
-
-                float tmp0 = bias0 + r0 + tmp024a + tmp024b + tmp024c * 32;
-                float tmp1 = bias0 + tmp135a + tmp135b + tmp135b + tmp135c * 16;
-                float tmp2 = bias0 + tmp024a + tmp024b * 4 + tmp024c * 8;
-                float tmp3 = bias0 + tmp135a + tmp135b * 8 + tmp135c * 4;
-                float tmp4 = bias0 + tmp024a + tmp024b * 16 + tmp024c + tmp024c;
-                float tmp5 = bias0 + r7 + tmp135a + tmp135b * 32 + tmp135c;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = tmp0;
-                    if (tj * 6 + 1 < outw) outptr0[1] = tmp1;
-                    if (tj * 6 + 2 < outw) outptr0[2] = tmp2;
-                    if (tj * 6 + 3 < outw) outptr0[3] = tmp3;
-                    if (tj * 6 + 4 < outw) outptr0[4] = tmp4;
-                    if (tj * 6 + 5 < outw) outptr0[5] = tmp5;
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd63(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, const Mat& bias, int nT, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 6n+2, winograd F(6,3)
-    int w_tiles = (outw + 5) / 6;
-    int h_tiles = (outh + 5) / 6;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 64;
-
-    // NCNN_LOGE("conv3x3s1_winograd63 %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    conv3x3s1_winograd_get_optimal_tile_mnk(M, N, K, B, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 4u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 4u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd63_transform_input_tile(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 4u, opt.workspace_allocator);
-
-        #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd63_transform_input_tile(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                conv3x3s1_winograd_gemm_transB_packed_tile(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk, opt.use_a53_a55_optimized_kernel);
-            }
-
-            // transform output
-            conv3x3s1_winograd63_transform_output_tile(top_tile, top_blob, bias, i, max_ii, j, max_jj);
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd_bf16s.h
deleted file mode 100644
index 58ddceeb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd_bf16s.h
+++ /dev/null
@@ -1,4541 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static inline void conv3x3s1_winograd23_transform_input_tile_bf16s(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    // const float itm[4][4] = {
-    //     {1.0f,  0.0f, -1.0f,  0.0f},
-    //     {0.0f,  1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  0.00f, 1.0f}
-    // };
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w - 1) / 2;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-#if __ARM_NEON
-#if __aarch64__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[4][4][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const unsigned short* r0 = bottom_blob.channel((k + kk) / elempack).row<const unsigned short>(ti * 2) + (tj * 2) * elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float32x4_t _r00 = vdupq_n_f32(0.f);
-                float32x4_t _r01 = vdupq_n_f32(0.f);
-                float32x4_t _r10 = vdupq_n_f32(0.f);
-                float32x4_t _r11 = vdupq_n_f32(0.f);
-                float32x4_t _r20 = vdupq_n_f32(0.f);
-                float32x4_t _r21 = vdupq_n_f32(0.f);
-                float32x4_t _r30 = vdupq_n_f32(0.f);
-                float32x4_t _r31 = vdupq_n_f32(0.f);
-
-                if (ti * 2 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        const unsigned short* r1 = r0 + N;
-
-                        _r00 = bfloat2float(vld1_u16(r0));
-                        _r01 = bfloat2float(vld1_u16(r1));
-                        if (tj * 2 + 1 < w)
-                        {
-                            _r10 = bfloat2float(vld1_u16(r0 + 4));
-                            _r11 = bfloat2float(vld1_u16(r1 + 4));
-                        }
-                        if (tj * 2 + 2 < w)
-                        {
-                            _r20 = bfloat2float(vld1_u16(r0 + 8));
-                            _r21 = bfloat2float(vld1_u16(r1 + 8));
-                        }
-                        if (tj * 2 + 3 < w)
-                        {
-                            _r30 = bfloat2float(vld1_u16(r0 + 12));
-                            _r31 = bfloat2float(vld1_u16(r1 + 12));
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        const unsigned short* r1 = r0 + N;
-                        const unsigned short* r2 = r0 + N * 2;
-                        const unsigned short* r3 = r0 + N * 3;
-                        const unsigned short* r4 = r0 + N * 4;
-                        const unsigned short* r5 = r0 + N * 5;
-                        const unsigned short* r6 = r0 + N * 6;
-                        const unsigned short* r7 = r0 + N * 7;
-
-                        uint16x4_t _t0 = vld1_u16(r0);
-                        uint16x4_t _t1 = vld1_u16(r1);
-                        uint16x4_t _t2 = vld1_u16(r2);
-                        uint16x4_t _t3 = vld1_u16(r3);
-                        uint16x4_t _t4 = vld1_u16(r4);
-                        uint16x4_t _t5 = vld1_u16(r5);
-                        uint16x4_t _t6 = vld1_u16(r6);
-                        uint16x4_t _t7 = vld1_u16(r7);
-
-                        transpose4x4_u16(_t0, _t1, _t2, _t3);
-                        transpose4x4_u16(_t4, _t5, _t6, _t7);
-
-                        _r00 = bfloat2float(_t0);
-                        _r01 = bfloat2float(_t4);
-                        if (tj * 2 + 1 < w)
-                        {
-                            _r10 = bfloat2float(_t1);
-                            _r11 = bfloat2float(_t5);
-                        }
-                        if (tj * 2 + 2 < w)
-                        {
-                            _r20 = bfloat2float(_t2);
-                            _r21 = bfloat2float(_t6);
-                        }
-                        if (tj * 2 + 3 < w)
-                        {
-                            _r30 = bfloat2float(_t3);
-                            _r31 = bfloat2float(_t7);
-                        }
-                    }
-                }
-
-                float32x4_t _tmp00 = vsubq_f32(_r00, _r20);
-                float32x4_t _tmp01 = vsubq_f32(_r01, _r21);
-                float32x4_t _tmp10 = vaddq_f32(_r10, _r20);
-                float32x4_t _tmp11 = vaddq_f32(_r11, _r21);
-                float32x4_t _tmp20 = vsubq_f32(_r20, _r10);
-                float32x4_t _tmp21 = vsubq_f32(_r21, _r11);
-                float32x4_t _tmp30 = vsubq_f32(_r30, _r10);
-                float32x4_t _tmp31 = vsubq_f32(_r31, _r11);
-
-                vst1q_f32(tmp[0][m], _tmp00);
-                vst1q_f32(tmp[0][m] + 4, _tmp01);
-                vst1q_f32(tmp[1][m], _tmp10);
-                vst1q_f32(tmp[1][m] + 4, _tmp11);
-                vst1q_f32(tmp[2][m], _tmp20);
-                vst1q_f32(tmp[2][m] + 4, _tmp21);
-                vst1q_f32(tmp[3][m], _tmp30);
-                vst1q_f32(tmp[3][m] + 4, _tmp31);
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 16 + jj * 8;
-            float* p1 = p0 + max_jj * 8;
-            float* p2 = p0 + max_jj * 8 * 2;
-            float* p3 = p0 + max_jj * 8 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float32x4_t _r00 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r01 = vld1q_f32(tmp[m][0] + 4);
-                float32x4_t _r10 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r11 = vld1q_f32(tmp[m][1] + 4);
-                float32x4_t _r20 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r21 = vld1q_f32(tmp[m][2] + 4);
-                float32x4_t _r30 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r31 = vld1q_f32(tmp[m][3] + 4);
-
-                float32x4_t _tmp00 = vsubq_f32(_r00, _r20);
-                float32x4_t _tmp01 = vsubq_f32(_r01, _r21);
-                float32x4_t _tmp10 = vaddq_f32(_r10, _r20);
-                float32x4_t _tmp11 = vaddq_f32(_r11, _r21);
-                float32x4_t _tmp20 = vsubq_f32(_r20, _r10);
-                float32x4_t _tmp21 = vsubq_f32(_r21, _r11);
-                float32x4_t _tmp30 = vsubq_f32(_r30, _r10);
-                float32x4_t _tmp31 = vsubq_f32(_r31, _r11);
-
-                vst1q_f32(p0, _tmp00);
-                vst1q_f32(p0 + 4, _tmp01);
-                vst1q_f32(p1, _tmp10);
-                vst1q_f32(p1 + 4, _tmp11);
-                vst1q_f32(p2, _tmp20);
-                vst1q_f32(p2 + 4, _tmp21);
-                vst1q_f32(p3, _tmp30);
-                vst1q_f32(p3 + 4, _tmp31);
-
-                p0 += max_jj * 4 * 8;
-                p1 += max_jj * 4 * 8;
-                p2 += max_jj * 4 * 8;
-                p3 += max_jj * 4 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-#else // __aarch64__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-    #pragma omp parallel for num_threads(nT)
-#endif // __aarch64__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 4;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[4][4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const unsigned short* r0 = bottom_blob.channel((k + kk) / elempack).row<const unsigned short>(ti * 2) + (tj * 2) * elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float32x4_t _r0 = vdupq_n_f32(0.f);
-                float32x4_t _r1 = vdupq_n_f32(0.f);
-                float32x4_t _r2 = vdupq_n_f32(0.f);
-                float32x4_t _r3 = vdupq_n_f32(0.f);
-
-                if (ti * 2 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        _r0 = bfloat2float(vld1_u16(r0));
-                        if (tj * 2 + 1 < w) _r1 = bfloat2float(vld1_u16(r0 + 4));
-                        if (tj * 2 + 2 < w) _r2 = bfloat2float(vld1_u16(r0 + 8));
-                        if (tj * 2 + 3 < w) _r3 = bfloat2float(vld1_u16(r0 + 12));
-                    }
-                    if (elempack == 1)
-                    {
-                        const unsigned short* r1 = r0 + N;
-                        const unsigned short* r2 = r0 + N * 2;
-                        const unsigned short* r3 = r0 + N * 3;
-
-                        uint16x4_t _t0 = vld1_u16(r0);
-                        uint16x4_t _t1 = vld1_u16(r1);
-                        uint16x4_t _t2 = vld1_u16(r2);
-                        uint16x4_t _t3 = vld1_u16(r3);
-
-                        transpose4x4_u16(_t0, _t1, _t2, _t3);
-
-                        _r0 = bfloat2float(_t0);
-                        if (tj * 2 + 1 < w) _r1 = bfloat2float(_t1);
-                        if (tj * 2 + 2 < w) _r2 = bfloat2float(_t2);
-                        if (tj * 2 + 3 < w) _r3 = bfloat2float(_t3);
-                    }
-                }
-
-                float32x4_t _tmp0 = vsubq_f32(_r0, _r2);
-                float32x4_t _tmp1 = vaddq_f32(_r1, _r2);
-                float32x4_t _tmp2 = vsubq_f32(_r2, _r1);
-                float32x4_t _tmp3 = vsubq_f32(_r3, _r1);
-
-                vst1q_f32(tmp[0][m], _tmp0);
-                vst1q_f32(tmp[1][m], _tmp1);
-                vst1q_f32(tmp[2][m], _tmp2);
-                vst1q_f32(tmp[3][m], _tmp3);
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 16 + jj * 4;
-            float* p1 = p0 + max_jj * 4;
-            float* p2 = p0 + max_jj * 4 * 2;
-            float* p3 = p0 + max_jj * 4 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float32x4_t _r0 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r1 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r2 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r3 = vld1q_f32(tmp[m][3]);
-
-                float32x4_t _tmp0 = vsubq_f32(_r0, _r2);
-                float32x4_t _tmp1 = vaddq_f32(_r1, _r2);
-                float32x4_t _tmp2 = vsubq_f32(_r2, _r1);
-                float32x4_t _tmp3 = vsubq_f32(_r3, _r1);
-
-                vst1q_f32(p0, _tmp0);
-                vst1q_f32(p1, _tmp1);
-                vst1q_f32(p2, _tmp2);
-                vst1q_f32(p3, _tmp3);
-
-                p0 += max_jj * 4 * 4;
-                p1 += max_jj * 4 * 4;
-                p2 += max_jj * 4 * 4;
-                p3 += max_jj * 4 * 4;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 4;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-#else // __ARM_NEON
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    #pragma omp parallel for num_threads(nT)
-#endif // __ARM_NEON
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        float tmp[4][4][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const unsigned short* r0 = bottom_blob.channel(k + kk).row<const unsigned short>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 4; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vdup_n_f32(0.f);
-                float32x2_t _r1 = vdup_n_f32(0.f);
-                float32x2_t _r2 = vdup_n_f32(0.f);
-                float32x2_t _r3 = vdup_n_f32(0.f);
-#else
-                float r00 = 0.f;
-                float r01 = 0.f;
-                float r10 = 0.f;
-                float r11 = 0.f;
-                float r20 = 0.f;
-                float r21 = 0.f;
-                float r30 = 0.f;
-                float r31 = 0.f;
-#endif
-
-                if (ti * 2 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const unsigned short* r1 = r0 + N;
-
-#if __ARM_NEON
-                        uint16x4_t _t0 = vld1_u16(r0);
-                        uint16x4_t _t1 = vld1_u16(r1);
-                        uint16x4x2_t _t01 = vzip_u16(_t0, _t1);
-                        float32x4_t _t0_fp32 = bfloat2float(_t01.val[0]);
-                        float32x4_t _t1_fp32 = bfloat2float(_t01.val[1]);
-
-                        _r0 = vget_low_f32(_t0_fp32);
-                        if (tj * 2 + 1 < w) _r1 = vget_high_f32(_t0_fp32);
-                        if (tj * 2 + 2 < w) _r2 = vget_low_f32(_t1_fp32);
-                        if (tj * 2 + 3 < w) _r3 = vget_high_f32(_t1_fp32);
-#else
-                        r00 = bfloat16_to_float32(r0[0]);
-                        r01 = bfloat16_to_float32(r1[0]);
-                        if (tj * 2 + 1 < w)
-                        {
-                            r10 = bfloat16_to_float32(r0[1]);
-                            r11 = bfloat16_to_float32(r1[1]);
-                        }
-                        if (tj * 2 + 2 < w)
-                        {
-                            r20 = bfloat16_to_float32(r0[2]);
-                            r21 = bfloat16_to_float32(r1[2]);
-                        }
-                        if (tj * 2 + 3 < w)
-                        {
-                            r30 = bfloat16_to_float32(r0[3]);
-                            r31 = bfloat16_to_float32(r1[3]);
-                        }
-#endif
-                    }
-                }
-
-#if __ARM_NEON
-                float32x2_t _tmp0 = vsub_f32(_r0, _r2);
-                float32x2_t _tmp1 = vadd_f32(_r1, _r2);
-                float32x2_t _tmp2 = vsub_f32(_r2, _r1);
-                float32x2_t _tmp3 = vsub_f32(_r3, _r1);
-
-                vst1_f32(tmp[0][m], _tmp0);
-                vst1_f32(tmp[1][m], _tmp1);
-                vst1_f32(tmp[2][m], _tmp2);
-                vst1_f32(tmp[3][m], _tmp3);
-#else
-                tmp[0][m][0] = r00 - r20;
-                tmp[0][m][1] = r01 - r21;
-                tmp[1][m][0] = r10 + r20;
-                tmp[1][m][1] = r11 + r21;
-                tmp[2][m][0] = r20 - r10;
-                tmp[2][m][1] = r21 - r11;
-                tmp[3][m][0] = r30 - r10;
-                tmp[3][m][1] = r31 - r11;
-#endif
-
-                r0 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 16 + jj * 2;
-            float* p1 = p0 + max_jj * 2;
-            float* p2 = p0 + max_jj * 2 * 2;
-            float* p3 = p0 + max_jj * 2 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(tmp[m][0]);
-                float32x2_t _r1 = vld1_f32(tmp[m][1]);
-                float32x2_t _r2 = vld1_f32(tmp[m][2]);
-                float32x2_t _r3 = vld1_f32(tmp[m][3]);
-
-                float32x2_t _tmp0 = vsub_f32(_r0, _r2);
-                float32x2_t _tmp1 = vadd_f32(_r1, _r2);
-                float32x2_t _tmp2 = vsub_f32(_r2, _r1);
-                float32x2_t _tmp3 = vsub_f32(_r3, _r1);
-
-                vst1_f32(p0, _tmp0);
-                vst1_f32(p1, _tmp1);
-                vst1_f32(p2, _tmp2);
-                vst1_f32(p3, _tmp3);
-#else
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-
-                p0[0] = r00 - r20;
-                p0[1] = r01 - r21;
-                p1[0] = r10 + r20;
-                p1[1] = r11 + r21;
-                p2[0] = r20 - r10;
-                p2[1] = r21 - r11;
-                p3[0] = r30 - r10;
-                p3[1] = r31 - r11;
-#endif
-
-                p0 += max_jj * 4 * 2;
-                p1 += max_jj * 4 * 2;
-                p2 += max_jj * 4 * 2;
-                p3 += max_jj * 4 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        float tmp[4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const unsigned short* r0123 = bottom_blob.channel(k + kk).row<const unsigned short>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 4; m++)
-            {
-                float r0 = 0.f;
-                float r1 = 0.f;
-                float r2 = 0.f;
-                float r3 = 0.f;
-
-                if (ti * 2 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = bfloat16_to_float32(r0123[0]);
-                        if (tj * 2 + 1 < w) r1 = bfloat16_to_float32(r0123[1]);
-                        if (tj * 2 + 2 < w) r2 = bfloat16_to_float32(r0123[2]);
-                        if (tj * 2 + 3 < w) r3 = bfloat16_to_float32(r0123[3]);
-                    }
-                }
-
-                tmp[0][m] = r0 - r2;
-                tmp[1][m] = r1 + r2;
-                tmp[2][m] = r2 - r1;
-                tmp[3][m] = r3 - r1;
-
-                r0123 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 16 + jj;
-            float* p1 = p0 + max_jj;
-            float* p2 = p0 + max_jj * 2;
-            float* p3 = p0 + max_jj * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-
-                p0[0] = r0 - r2;
-                p1[0] = r1 + r2;
-                p2[0] = r2 - r1;
-                p3[0] = r3 - r1;
-
-                p0 += max_jj * 4;
-                p1 += max_jj * 4;
-                p2 += max_jj * 4;
-                p3 += max_jj * 4;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_output_tile_bf16s(const Mat& top_tile, Mat& top_blob, const Mat& bias, int i, int max_ii, int j, int max_jj)
-{
-    // const float otm[2][4] = {
-    //     {1.0f,  1.0f,  1.0f,  0.0f},
-    //     {0.0f,  1.0f, -1.0f,  1.0f}
-    // };
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 1) / 2;
-
-    const float* biasptr = bias;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        float32x4_t _bias0 = biasptr ? vld1q_f32(biasptr + i + ii) : vdupq_n_f32(0.f);
-        float32x4_t _bias1 = biasptr ? vld1q_f32(biasptr + i + ii + 4) : vdupq_n_f32(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[2][4][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 16 + jj * 8;
-            const float* r1 = r0 + max_jj * 8;
-            const float* r2 = r0 + max_jj * 8 * 2;
-            const float* r3 = r0 + max_jj * 8 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r30 = vld1q_f32(r3);
-                float32x4_t _r31 = vld1q_f32(r3 + 4);
-
-                float32x4_t _tmp00 = vaddq_f32(vaddq_f32(_r00, _r10), _r20);
-                float32x4_t _tmp01 = vaddq_f32(vaddq_f32(_r01, _r11), _r21);
-                float32x4_t _tmp10 = vaddq_f32(vsubq_f32(_r10, _r20), _r30);
-                float32x4_t _tmp11 = vaddq_f32(vsubq_f32(_r11, _r21), _r31);
-
-                vst1q_f32(tmp[0][m], _tmp00);
-                vst1q_f32(tmp[0][m] + 4, _tmp01);
-                vst1q_f32(tmp[1][m], _tmp10);
-                vst1q_f32(tmp[1][m] + 4, _tmp11);
-
-                r0 += max_jj * 4 * 8;
-                r1 += max_jj * 4 * 8;
-                r2 += max_jj * 4 * 8;
-                r3 += max_jj * 4 * 8;
-            }
-
-            unsigned short* outptr0 = top_blob.channel((i + ii) / out_elempack).row<unsigned short>(ti * 2) + (tj * 2) * out_elempack;
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                float32x4_t _r00 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r01 = vld1q_f32(tmp[m][0] + 4);
-                float32x4_t _r10 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r11 = vld1q_f32(tmp[m][1] + 4);
-                float32x4_t _r20 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r21 = vld1q_f32(tmp[m][2] + 4);
-                float32x4_t _r30 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r31 = vld1q_f32(tmp[m][3] + 4);
-
-                float32x4_t _tmp00 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r00, _r10), _r20));
-                float32x4_t _tmp01 = vaddq_f32(_bias1, vaddq_f32(vaddq_f32(_r01, _r11), _r21));
-                float32x4_t _tmp10 = vaddq_f32(_bias0, vaddq_f32(vsubq_f32(_r10, _r20), _r30));
-                float32x4_t _tmp11 = vaddq_f32(_bias1, vaddq_f32(vsubq_f32(_r11, _r21), _r31));
-
-                if (out_elempack == 4)
-                {
-                    unsigned short* outptr1 = outptr0 + N;
-
-                    vst1_u16(outptr0, float2bfloat(_tmp00));
-                    vst1_u16(outptr1, float2bfloat(_tmp01));
-                    if (tj * 2 + 1 < outw)
-                    {
-                        vst1_u16(outptr0 + 4, float2bfloat(_tmp10));
-                        vst1_u16(outptr1 + 4, float2bfloat(_tmp11));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short tmp0[8];
-                    unsigned short tmp1[8];
-                    vst1_u16(tmp0, float2bfloat(_tmp00));
-                    vst1_u16(tmp0 + 4, float2bfloat(_tmp01));
-                    vst1_u16(tmp1, float2bfloat(_tmp10));
-                    vst1_u16(tmp1 + 4, float2bfloat(_tmp11));
-
-                    unsigned short* outptr1 = outptr0 + N;
-                    unsigned short* outptr2 = outptr0 + N * 2;
-                    unsigned short* outptr3 = outptr0 + N * 3;
-                    unsigned short* outptr4 = outptr0 + N * 4;
-                    unsigned short* outptr5 = outptr0 + N * 5;
-                    unsigned short* outptr6 = outptr0 + N * 6;
-                    unsigned short* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        float32x4_t _bias0 = biasptr ? vld1q_f32(biasptr + i + ii) : vdupq_n_f32(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[2][4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 16 + jj * 4;
-            const float* r1 = r0 + max_jj * 4;
-            const float* r2 = r0 + max_jj * 4 * 2;
-            const float* r3 = r0 + max_jj * 4 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float32x4_t _r0 = vld1q_f32(r0);
-                float32x4_t _r1 = vld1q_f32(r1);
-                float32x4_t _r2 = vld1q_f32(r2);
-                float32x4_t _r3 = vld1q_f32(r3);
-
-                float32x4_t _tmp0 = vaddq_f32(vaddq_f32(_r0, _r1), _r2);
-                float32x4_t _tmp1 = vaddq_f32(vsubq_f32(_r1, _r2), _r3);
-
-                vst1q_f32(tmp[0][m], _tmp0);
-                vst1q_f32(tmp[1][m], _tmp1);
-
-                r0 += max_jj * 4 * 4;
-                r1 += max_jj * 4 * 4;
-                r2 += max_jj * 4 * 4;
-                r3 += max_jj * 4 * 4;
-            }
-
-            unsigned short* outptr0 = top_blob.channel((i + ii) / out_elempack).row<unsigned short>(ti * 2) + (tj * 2) * out_elempack;
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                float32x4_t _r0 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r1 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r2 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r3 = vld1q_f32(tmp[m][3]);
-
-                float32x4_t _tmp0 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r0, _r1), _r2));
-                float32x4_t _tmp1 = vaddq_f32(_bias0, vaddq_f32(vsubq_f32(_r1, _r2), _r3));
-
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_tmp0));
-                    if (tj * 2 + 1 < outw) vst1_u16(outptr0 + 4, float2bfloat(_tmp1));
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short tmp0[4];
-                    unsigned short tmp1[4];
-                    vst1_u16(tmp0, float2bfloat(_tmp0));
-                    vst1_u16(tmp1, float2bfloat(_tmp1));
-
-                    unsigned short* outptr1 = outptr0 + N;
-                    unsigned short* outptr2 = outptr0 + N * 2;
-                    unsigned short* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-#if __ARM_NEON
-        float32x2_t _bias0 = biasptr ? vld1_f32(biasptr + i + ii) : vdup_n_f32(0.f);
-#else
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-        float bias1 = biasptr ? biasptr[i + ii + 1] : 0.f;
-#endif
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        float tmp[2][4][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 16 + jj * 2;
-            const float* r1 = r0 + max_jj * 2;
-            const float* r2 = r0 + max_jj * 2 * 2;
-            const float* r3 = r0 + max_jj * 2 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(r0);
-                float32x2_t _r1 = vld1_f32(r1);
-                float32x2_t _r2 = vld1_f32(r2);
-                float32x2_t _r3 = vld1_f32(r3);
-
-                float32x2_t _tmp0 = vadd_f32(vadd_f32(_r0, _r1), _r2);
-                float32x2_t _tmp1 = vadd_f32(vsub_f32(_r1, _r2), _r3);
-
-                vst1_f32(tmp[0][m], _tmp0);
-                vst1_f32(tmp[1][m], _tmp1);
-#else
-                tmp[0][m][0] = r0[0] + r1[0] + r2[0];
-                tmp[0][m][1] = r0[1] + r1[1] + r2[1];
-                tmp[1][m][0] = r1[0] - r2[0] + r3[0];
-                tmp[1][m][1] = r1[1] - r2[1] + r3[1];
-#endif
-
-                r0 += max_jj * 4 * 2;
-                r1 += max_jj * 4 * 2;
-                r2 += max_jj * 4 * 2;
-                r3 += max_jj * 4 * 2;
-            }
-
-            unsigned short* outptr0 = top_blob.channel(i + ii).row<unsigned short>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(tmp[m][0]);
-                float32x2_t _r1 = vld1_f32(tmp[m][1]);
-                float32x2_t _r2 = vld1_f32(tmp[m][2]);
-                float32x2_t _r3 = vld1_f32(tmp[m][3]);
-
-                float32x2_t _tmp0 = vadd_f32(_bias0, vadd_f32(vadd_f32(_r0, _r1), _r2));
-                float32x2_t _tmp1 = vadd_f32(_bias0, vadd_f32(vsub_f32(_r1, _r2), _r3));
-#else
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-
-                float tmp00 = bias0 + r00 + r10 + r20;
-                float tmp01 = bias1 + r01 + r11 + r21;
-                float tmp10 = bias0 + r10 - r20 + r30;
-                float tmp11 = bias1 + r11 - r21 + r31;
-#endif
-
-                // if (out_elempack == 1)
-                {
-                    unsigned short* outptr1 = outptr0 + N;
-
-#if __ARM_NEON
-                    uint16x4_t _tmp01 = float2bfloat(vcombine_f32(_tmp0, _tmp1));
-
-                    outptr0[0] = vget_lane_u16(_tmp01, 0);
-                    outptr1[0] = vget_lane_u16(_tmp01, 1);
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = vget_lane_u16(_tmp01, 2);
-                        outptr1[1] = vget_lane_u16(_tmp01, 3);
-                    }
-#else
-                    outptr0[0] = float32_to_bfloat16(tmp00);
-                    outptr1[0] = float32_to_bfloat16(tmp01);
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = float32_to_bfloat16(tmp10);
-                        outptr1[1] = float32_to_bfloat16(tmp11);
-                    }
-#endif
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-
-        float tmp[2][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 16 + jj;
-            const float* r1 = r0 + max_jj;
-            const float* r2 = r0 + max_jj * 2;
-            const float* r3 = r0 + max_jj * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                tmp[0][m] = r0[0] + r1[0] + r2[0];
-                tmp[1][m] = r1[0] - r2[0] + r3[0];
-
-                r0 += max_jj * 4;
-                r1 += max_jj * 4;
-                r2 += max_jj * 4;
-                r3 += max_jj * 4;
-            }
-
-            unsigned short* outptr0 = top_blob.channel(i + ii).row<unsigned short>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-
-                float tmp0 = bias0 + r0 + r1 + r2;
-                float tmp1 = bias0 + r1 - r2 + r3;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = float32_to_bfloat16(tmp0);
-                    if (tj * 2 + 1 < outw) outptr0[1] = float32_to_bfloat16(tmp1);
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_bf16s(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, const Mat& bias, int nT, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 2n+2, winograd F(2,3)
-    int w_tiles = (outw + 1) / 2;
-    int h_tiles = (outh + 1) / 2;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 16;
-
-    // NCNN_LOGE("conv3x3s1_winograd23_bf16s %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    conv3x3s1_winograd_get_optimal_tile_mnk(M, N, K, B, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 4u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 4u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd23_transform_input_tile_bf16s(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 4u, opt.workspace_allocator);
-
-        #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd23_transform_input_tile_bf16s(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                conv3x3s1_winograd_gemm_transB_packed_tile(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk, opt.use_a53_a55_optimized_kernel);
-            }
-
-            // transform output
-            conv3x3s1_winograd23_transform_output_tile_bf16s(top_tile, top_blob, bias, i, max_ii, j, max_jj);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_input_tile_bf16s(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    const float sq2 = 1.41421356237;
-    const float sq2_d2 = 1.41421356237 / 2;
-
-    // const float itm[6][6] = {
-    //     {1.0f,  0.0f,  -2.5f,  0.0f,  1.0f, 0.0f},
-    //     {0.0f, -sq2,   -2.0f,  sq2/2, 1.0f, 0.0f},
-    //     {0.0f,  sq2,   -2.0f, -sq2/2, 1.0f, 0.0f},
-    //     {0.0f, -sq2/2, -0.5f,  sq2,   1.0f, 0.0f},
-    //     {0.0f,  sq2/2, -0.5f, -sq2,   1.0f, 0.0f},
-    //     {0.0f,  1.0f,   0.0f,  -2.5f, 0.0f, 1.0f}
-    // };
-
-    // 0 =  r00 + r04 - 2.5f * r02
-    // 1 = -(sq2 * r01 - sq2_d2 * r03) + (r04 - 2 * r02)
-    // 2 =  (sq2 * r01 - sq2_d2 * r03) + (r04 - 2 * r02)
-    // 3 =  (sq2 * r03 - sq2_d2 * r01) + (r04 - 0.5f * r02)
-    // 4 = -(sq2 * r03 - sq2_d2 * r01) + (r04 - 0.5f * r02)
-    // 5 =  r01 + r05 - 2.5f * r03
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w + 1) / 4;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-#if __ARM_NEON
-#if __aarch64__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[6][6][8];
-
-        const float coeffs[4] = {sq2, -sq2_d2, -2.f, -0.5f};
-        float32x4_t _coeffs = vld1q_f32(coeffs);
-        float32x4_t _vm2_5 = vdupq_n_f32(-2.5f);
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const unsigned short* r0 = bottom_blob.channel((k + kk) / elempack).row<const unsigned short>(ti * 4) + (tj * 4) * elempack;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float32x4_t _r00 = vdupq_n_f32(0.f);
-                float32x4_t _r01 = vdupq_n_f32(0.f);
-                float32x4_t _r10 = vdupq_n_f32(0.f);
-                float32x4_t _r11 = vdupq_n_f32(0.f);
-                float32x4_t _r20 = vdupq_n_f32(0.f);
-                float32x4_t _r21 = vdupq_n_f32(0.f);
-                float32x4_t _r30 = vdupq_n_f32(0.f);
-                float32x4_t _r31 = vdupq_n_f32(0.f);
-                float32x4_t _r40 = vdupq_n_f32(0.f);
-                float32x4_t _r41 = vdupq_n_f32(0.f);
-                float32x4_t _r50 = vdupq_n_f32(0.f);
-                float32x4_t _r51 = vdupq_n_f32(0.f);
-
-                if (ti * 4 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        const unsigned short* r1 = r0 + N;
-
-                        _r00 = bfloat2float(vld1_u16(r0));
-                        _r01 = bfloat2float(vld1_u16(r1));
-                        if (tj * 4 + 1 < w)
-                        {
-                            _r10 = bfloat2float(vld1_u16(r0 + 4));
-                            _r11 = bfloat2float(vld1_u16(r1 + 4));
-                        }
-                        if (tj * 4 + 2 < w)
-                        {
-                            _r20 = bfloat2float(vld1_u16(r0 + 8));
-                            _r21 = bfloat2float(vld1_u16(r1 + 8));
-                        }
-                        if (tj * 4 + 3 < w)
-                        {
-                            _r30 = bfloat2float(vld1_u16(r0 + 12));
-                            _r31 = bfloat2float(vld1_u16(r1 + 12));
-                        }
-                        if (tj * 4 + 4 < w)
-                        {
-                            _r40 = bfloat2float(vld1_u16(r0 + 16));
-                            _r41 = bfloat2float(vld1_u16(r1 + 16));
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            _r50 = bfloat2float(vld1_u16(r0 + 20));
-                            _r51 = bfloat2float(vld1_u16(r1 + 20));
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        const unsigned short* r1 = r0 + N;
-                        const unsigned short* r2 = r0 + N * 2;
-                        const unsigned short* r3 = r0 + N * 3;
-                        const unsigned short* r4 = r0 + N * 4;
-                        const unsigned short* r5 = r0 + N * 5;
-                        const unsigned short* r6 = r0 + N * 6;
-                        const unsigned short* r7 = r0 + N * 7;
-
-                        uint16x4_t _t0 = vld1_u16(r0);
-                        uint16x4_t _t1 = vld1_u16(r1);
-                        uint16x4_t _t2 = vld1_u16(r2);
-                        uint16x4_t _t3 = vld1_u16(r3);
-                        uint16x4_t _t4 = vld1_u16(r4);
-                        uint16x4_t _t5 = vld1_u16(r5);
-                        uint16x4_t _t6 = vld1_u16(r6);
-                        uint16x4_t _t7 = vld1_u16(r7);
-
-                        transpose4x4_u16(_t0, _t1, _t2, _t3);
-                        transpose4x4_u16(_t4, _t5, _t6, _t7);
-
-                        _r00 = bfloat2float(_t0);
-                        _r01 = bfloat2float(_t4);
-                        if (tj * 4 + 1 < w)
-                        {
-                            _r10 = bfloat2float(_t1);
-                            _r11 = bfloat2float(_t5);
-                        }
-                        if (tj * 4 + 2 < w)
-                        {
-                            _r20 = bfloat2float(_t2);
-                            _r21 = bfloat2float(_t6);
-                        }
-                        if (tj * 4 + 3 < w)
-                        {
-                            _r30 = bfloat2float(_t3);
-                            _r31 = bfloat2float(_t7);
-                        }
-                        if (tj * 4 + 4 < w)
-                        {
-                            unsigned short tmp[8] = {r0[4], r1[4], r2[4], r3[4], r4[4], r5[4], r6[4], r7[4]};
-                            _r40 = bfloat2float(vld1_u16(tmp));
-                            _r41 = bfloat2float(vld1_u16(tmp + 4));
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            unsigned short tmp[8] = {r0[5], r1[5], r2[5], r3[5], r4[5], r5[5], r6[5], r7[5]};
-                            _r50 = bfloat2float(vld1_u16(tmp));
-                            _r51 = bfloat2float(vld1_u16(tmp + 4));
-                        }
-                    }
-                }
-
-                float32x4_t _tmp12a0 = vfmaq_laneq_f32(vmulq_laneq_f32(_r10, _coeffs, 0), _r30, _coeffs, 1);
-                float32x4_t _tmp12a1 = vfmaq_laneq_f32(vmulq_laneq_f32(_r11, _coeffs, 0), _r31, _coeffs, 1);
-                float32x4_t _tmp12b0 = vfmaq_laneq_f32(_r40, _r20, _coeffs, 2);
-                float32x4_t _tmp12b1 = vfmaq_laneq_f32(_r41, _r21, _coeffs, 2);
-                float32x4_t _tmp34a0 = vfmaq_laneq_f32(vmulq_laneq_f32(_r30, _coeffs, 0), _r10, _coeffs, 1);
-                float32x4_t _tmp34a1 = vfmaq_laneq_f32(vmulq_laneq_f32(_r31, _coeffs, 0), _r11, _coeffs, 1);
-                float32x4_t _tmp34b0 = vfmaq_laneq_f32(_r40, _r20, _coeffs, 3);
-                float32x4_t _tmp34b1 = vfmaq_laneq_f32(_r41, _r21, _coeffs, 3);
-
-                float32x4_t _tmp00 = vfmaq_f32(vaddq_f32(_r00, _r40), _r20, _vm2_5);
-                float32x4_t _tmp01 = vfmaq_f32(vaddq_f32(_r01, _r41), _r21, _vm2_5);
-                float32x4_t _tmp10 = vsubq_f32(_tmp12b0, _tmp12a0);
-                float32x4_t _tmp11 = vsubq_f32(_tmp12b1, _tmp12a1);
-                float32x4_t _tmp20 = vaddq_f32(_tmp12b0, _tmp12a0);
-                float32x4_t _tmp21 = vaddq_f32(_tmp12b1, _tmp12a1);
-                float32x4_t _tmp30 = vaddq_f32(_tmp34b0, _tmp34a0);
-                float32x4_t _tmp31 = vaddq_f32(_tmp34b1, _tmp34a1);
-                float32x4_t _tmp40 = vsubq_f32(_tmp34b0, _tmp34a0);
-                float32x4_t _tmp41 = vsubq_f32(_tmp34b1, _tmp34a1);
-                float32x4_t _tmp50 = vfmaq_f32(vaddq_f32(_r10, _r50), _r30, _vm2_5);
-                float32x4_t _tmp51 = vfmaq_f32(vaddq_f32(_r11, _r51), _r31, _vm2_5);
-
-                vst1q_f32(tmp[0][m], _tmp00);
-                vst1q_f32(tmp[0][m] + 4, _tmp01);
-                vst1q_f32(tmp[1][m], _tmp10);
-                vst1q_f32(tmp[1][m] + 4, _tmp11);
-                vst1q_f32(tmp[2][m], _tmp20);
-                vst1q_f32(tmp[2][m] + 4, _tmp21);
-                vst1q_f32(tmp[3][m], _tmp30);
-                vst1q_f32(tmp[3][m] + 4, _tmp31);
-                vst1q_f32(tmp[4][m], _tmp40);
-                vst1q_f32(tmp[4][m] + 4, _tmp41);
-                vst1q_f32(tmp[5][m], _tmp50);
-                vst1q_f32(tmp[5][m] + 4, _tmp51);
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 36 + jj * 8;
-            float* p1 = p0 + max_jj * 8;
-            float* p2 = p0 + max_jj * 8 * 2;
-            float* p3 = p0 + max_jj * 8 * 3;
-            float* p4 = p0 + max_jj * 8 * 4;
-            float* p5 = p0 + max_jj * 8 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float32x4_t _r00 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r01 = vld1q_f32(tmp[m][0] + 4);
-                float32x4_t _r10 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r11 = vld1q_f32(tmp[m][1] + 4);
-                float32x4_t _r20 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r21 = vld1q_f32(tmp[m][2] + 4);
-                float32x4_t _r30 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r31 = vld1q_f32(tmp[m][3] + 4);
-                float32x4_t _r40 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r41 = vld1q_f32(tmp[m][4] + 4);
-                float32x4_t _r50 = vld1q_f32(tmp[m][5]);
-                float32x4_t _r51 = vld1q_f32(tmp[m][5] + 4);
-
-                float32x4_t _tmp12a0 = vfmaq_laneq_f32(vmulq_laneq_f32(_r10, _coeffs, 0), _r30, _coeffs, 1);
-                float32x4_t _tmp12a1 = vfmaq_laneq_f32(vmulq_laneq_f32(_r11, _coeffs, 0), _r31, _coeffs, 1);
-                float32x4_t _tmp12b0 = vfmaq_laneq_f32(_r40, _r20, _coeffs, 2);
-                float32x4_t _tmp12b1 = vfmaq_laneq_f32(_r41, _r21, _coeffs, 2);
-                float32x4_t _tmp34a0 = vfmaq_laneq_f32(vmulq_laneq_f32(_r30, _coeffs, 0), _r10, _coeffs, 1);
-                float32x4_t _tmp34a1 = vfmaq_laneq_f32(vmulq_laneq_f32(_r31, _coeffs, 0), _r11, _coeffs, 1);
-                float32x4_t _tmp34b0 = vfmaq_laneq_f32(_r40, _r20, _coeffs, 3);
-                float32x4_t _tmp34b1 = vfmaq_laneq_f32(_r41, _r21, _coeffs, 3);
-
-                float32x4_t _tmp00 = vfmaq_f32(vaddq_f32(_r00, _r40), _r20, _vm2_5);
-                float32x4_t _tmp01 = vfmaq_f32(vaddq_f32(_r01, _r41), _r21, _vm2_5);
-                float32x4_t _tmp10 = vsubq_f32(_tmp12b0, _tmp12a0);
-                float32x4_t _tmp11 = vsubq_f32(_tmp12b1, _tmp12a1);
-                float32x4_t _tmp20 = vaddq_f32(_tmp12b0, _tmp12a0);
-                float32x4_t _tmp21 = vaddq_f32(_tmp12b1, _tmp12a1);
-                float32x4_t _tmp30 = vaddq_f32(_tmp34b0, _tmp34a0);
-                float32x4_t _tmp31 = vaddq_f32(_tmp34b1, _tmp34a1);
-                float32x4_t _tmp40 = vsubq_f32(_tmp34b0, _tmp34a0);
-                float32x4_t _tmp41 = vsubq_f32(_tmp34b1, _tmp34a1);
-                float32x4_t _tmp50 = vfmaq_f32(vaddq_f32(_r10, _r50), _r30, _vm2_5);
-                float32x4_t _tmp51 = vfmaq_f32(vaddq_f32(_r11, _r51), _r31, _vm2_5);
-
-                vst1q_f32(p0, _tmp00);
-                vst1q_f32(p0 + 4, _tmp01);
-                vst1q_f32(p1, _tmp10);
-                vst1q_f32(p1 + 4, _tmp11);
-                vst1q_f32(p2, _tmp20);
-                vst1q_f32(p2 + 4, _tmp21);
-                vst1q_f32(p3, _tmp30);
-                vst1q_f32(p3 + 4, _tmp31);
-                vst1q_f32(p4, _tmp40);
-                vst1q_f32(p4 + 4, _tmp41);
-                vst1q_f32(p5, _tmp50);
-                vst1q_f32(p5 + 4, _tmp51);
-
-                p0 += max_jj * 6 * 8;
-                p1 += max_jj * 6 * 8;
-                p2 += max_jj * 6 * 8;
-                p3 += max_jj * 6 * 8;
-                p4 += max_jj * 6 * 8;
-                p5 += max_jj * 6 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-#else // __aarch64__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-    #pragma omp parallel for num_threads(nT)
-#endif // __aarch64__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 4;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[6][6][4];
-
-        const float coeffs[4] = {sq2, -sq2_d2, -2.f, -0.5f};
-        float32x4_t _coeffs = vld1q_f32(coeffs);
-        float32x4_t _vm2_5 = vdupq_n_f32(-2.5f);
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const unsigned short* r0 = bottom_blob.channel((k + kk) / elempack).row<const unsigned short>(ti * 4) + (tj * 4) * elempack;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float32x4_t _r0 = vdupq_n_f32(0.f);
-                float32x4_t _r1 = vdupq_n_f32(0.f);
-                float32x4_t _r2 = vdupq_n_f32(0.f);
-                float32x4_t _r3 = vdupq_n_f32(0.f);
-                float32x4_t _r4 = vdupq_n_f32(0.f);
-                float32x4_t _r5 = vdupq_n_f32(0.f);
-
-                if (ti * 4 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        _r0 = bfloat2float(vld1_u16(r0));
-                        if (tj * 4 + 1 < w) _r1 = bfloat2float(vld1_u16(r0 + 4));
-                        if (tj * 4 + 2 < w) _r2 = bfloat2float(vld1_u16(r0 + 8));
-                        if (tj * 4 + 3 < w) _r3 = bfloat2float(vld1_u16(r0 + 12));
-                        if (tj * 4 + 4 < w) _r4 = bfloat2float(vld1_u16(r0 + 16));
-                        if (tj * 4 + 5 < w) _r5 = bfloat2float(vld1_u16(r0 + 20));
-                    }
-                    if (elempack == 1)
-                    {
-                        const unsigned short* r1 = r0 + N;
-                        const unsigned short* r2 = r0 + N * 2;
-                        const unsigned short* r3 = r0 + N * 3;
-
-                        uint16x4_t _t0 = vld1_u16(r0);
-                        uint16x4_t _t1 = vld1_u16(r1);
-                        uint16x4_t _t2 = vld1_u16(r2);
-                        uint16x4_t _t3 = vld1_u16(r3);
-
-                        transpose4x4_u16(_t0, _t1, _t2, _t3);
-
-                        _r0 = bfloat2float(_t0);
-                        if (tj * 4 + 1 < w) _r1 = bfloat2float(_t1);
-                        if (tj * 4 + 2 < w) _r2 = bfloat2float(_t2);
-                        if (tj * 4 + 3 < w) _r3 = bfloat2float(_t3);
-                        if (tj * 4 + 4 < w)
-                        {
-                            unsigned short tmp[4] = {r0[4], r1[4], r2[4], r3[4]};
-                            _r4 = bfloat2float(vld1_u16(tmp));
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            unsigned short tmp[4] = {r0[5], r1[5], r2[5], r3[5]};
-                            _r5 = bfloat2float(vld1_u16(tmp));
-                        }
-                    }
-                }
-
-#if __aarch64__
-                float32x4_t _tmp12a = vfmaq_laneq_f32(vmulq_laneq_f32(_r1, _coeffs, 0), _r3, _coeffs, 1);
-                float32x4_t _tmp12b = vfmaq_laneq_f32(_r4, _r2, _coeffs, 2);
-                float32x4_t _tmp34a = vfmaq_laneq_f32(vmulq_laneq_f32(_r3, _coeffs, 0), _r1, _coeffs, 1);
-                float32x4_t _tmp34b = vfmaq_laneq_f32(_r4, _r2, _coeffs, 3);
-#else
-                float32x4_t _tmp12a = vmlaq_lane_f32(vmulq_lane_f32(_r1, vget_low_f32(_coeffs), 0), _r3, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp12b = vmlaq_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 0);
-                float32x4_t _tmp34a = vmlaq_lane_f32(vmulq_lane_f32(_r3, vget_low_f32(_coeffs), 0), _r1, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp34b = vmlaq_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 1);
-#endif
-
-#if __aarch64__
-                float32x4_t _tmp0 = vfmaq_f32(vaddq_f32(_r0, _r4), _r2, _vm2_5);
-#else
-                float32x4_t _tmp0 = vmlaq_f32(vaddq_f32(_r0, _r4), _r2, _vm2_5);
-#endif
-                float32x4_t _tmp1 = vsubq_f32(_tmp12b, _tmp12a);
-                float32x4_t _tmp2 = vaddq_f32(_tmp12b, _tmp12a);
-                float32x4_t _tmp3 = vaddq_f32(_tmp34b, _tmp34a);
-                float32x4_t _tmp4 = vsubq_f32(_tmp34b, _tmp34a);
-#if __aarch64__
-                float32x4_t _tmp5 = vfmaq_f32(vaddq_f32(_r1, _r5), _r3, _vm2_5);
-#else
-                float32x4_t _tmp5 = vmlaq_f32(vaddq_f32(_r1, _r5), _r3, _vm2_5);
-#endif
-
-                vst1q_f32(tmp[0][m], _tmp0);
-                vst1q_f32(tmp[1][m], _tmp1);
-                vst1q_f32(tmp[2][m], _tmp2);
-                vst1q_f32(tmp[3][m], _tmp3);
-                vst1q_f32(tmp[4][m], _tmp4);
-                vst1q_f32(tmp[5][m], _tmp5);
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 36 + jj * 4;
-            float* p1 = p0 + max_jj * 4;
-            float* p2 = p0 + max_jj * 4 * 2;
-            float* p3 = p0 + max_jj * 4 * 3;
-            float* p4 = p0 + max_jj * 4 * 4;
-            float* p5 = p0 + max_jj * 4 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float32x4_t _r0 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r1 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r2 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r3 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r4 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r5 = vld1q_f32(tmp[m][5]);
-
-#if __aarch64__
-                float32x4_t _tmp12a = vfmaq_laneq_f32(vmulq_laneq_f32(_r1, _coeffs, 0), _r3, _coeffs, 1);
-                float32x4_t _tmp12b = vfmaq_laneq_f32(_r4, _r2, _coeffs, 2);
-                float32x4_t _tmp34a = vfmaq_laneq_f32(vmulq_laneq_f32(_r3, _coeffs, 0), _r1, _coeffs, 1);
-                float32x4_t _tmp34b = vfmaq_laneq_f32(_r4, _r2, _coeffs, 3);
-#else
-                float32x4_t _tmp12a = vmlaq_lane_f32(vmulq_lane_f32(_r1, vget_low_f32(_coeffs), 0), _r3, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp12b = vmlaq_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 0);
-                float32x4_t _tmp34a = vmlaq_lane_f32(vmulq_lane_f32(_r3, vget_low_f32(_coeffs), 0), _r1, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp34b = vmlaq_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 1);
-#endif
-
-#if __aarch64__
-                float32x4_t _tmp0 = vfmaq_f32(vaddq_f32(_r0, _r4), _r2, _vm2_5);
-#else
-                float32x4_t _tmp0 = vmlaq_f32(vaddq_f32(_r0, _r4), _r2, _vm2_5);
-#endif
-                float32x4_t _tmp1 = vsubq_f32(_tmp12b, _tmp12a);
-                float32x4_t _tmp2 = vaddq_f32(_tmp12b, _tmp12a);
-                float32x4_t _tmp3 = vaddq_f32(_tmp34b, _tmp34a);
-                float32x4_t _tmp4 = vsubq_f32(_tmp34b, _tmp34a);
-#if __aarch64__
-                float32x4_t _tmp5 = vfmaq_f32(vaddq_f32(_r1, _r5), _r3, _vm2_5);
-#else
-                float32x4_t _tmp5 = vmlaq_f32(vaddq_f32(_r1, _r5), _r3, _vm2_5);
-#endif
-
-                vst1q_f32(p0, _tmp0);
-                vst1q_f32(p1, _tmp1);
-                vst1q_f32(p2, _tmp2);
-                vst1q_f32(p3, _tmp3);
-                vst1q_f32(p4, _tmp4);
-                vst1q_f32(p5, _tmp5);
-
-                p0 += max_jj * 6 * 4;
-                p1 += max_jj * 6 * 4;
-                p2 += max_jj * 6 * 4;
-                p3 += max_jj * 6 * 4;
-                p4 += max_jj * 6 * 4;
-                p5 += max_jj * 6 * 4;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 4;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-#else // __ARM_NEON
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    #pragma omp parallel for num_threads(nT)
-#endif // __ARM_NEON
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        float tmp[6][6][2];
-
-#if __ARM_NEON
-        const float coeffs[4] = {sq2, -sq2_d2, -2.f, -0.5f};
-        float32x4_t _coeffs = vld1q_f32(coeffs);
-        float32x2_t _vm2_5 = vdup_n_f32(-2.5f);
-#endif
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const unsigned short* r0 = bottom_blob.channel(k + kk).row<const unsigned short>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 6; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vdup_n_f32(0.f);
-                float32x2_t _r1 = vdup_n_f32(0.f);
-                float32x2_t _r2 = vdup_n_f32(0.f);
-                float32x2_t _r3 = vdup_n_f32(0.f);
-                float32x2_t _r4 = vdup_n_f32(0.f);
-                float32x2_t _r5 = vdup_n_f32(0.f);
-#else
-                float r00 = 0.f;
-                float r01 = 0.f;
-                float r10 = 0.f;
-                float r11 = 0.f;
-                float r20 = 0.f;
-                float r21 = 0.f;
-                float r30 = 0.f;
-                float r31 = 0.f;
-                float r40 = 0.f;
-                float r41 = 0.f;
-                float r50 = 0.f;
-                float r51 = 0.f;
-#endif
-
-                if (ti * 4 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const unsigned short* r1 = r0 + N;
-
-#if __ARM_NEON
-                        uint16x4_t _t0 = vld1_u16(r0);
-                        uint16x4_t _t1 = vld1_u16(r1);
-                        uint16x4x2_t _t01 = vzip_u16(_t0, _t1);
-                        float32x4_t _t0_fp32 = bfloat2float(_t01.val[0]);
-                        float32x4_t _t1_fp32 = bfloat2float(_t01.val[1]);
-
-                        _r0 = vget_low_f32(_t0_fp32);
-                        if (tj * 4 + 1 < w) _r1 = vget_high_f32(_t0_fp32);
-                        if (tj * 4 + 2 < w) _r2 = vget_low_f32(_t1_fp32);
-                        if (tj * 4 + 3 < w) _r3 = vget_high_f32(_t1_fp32);
-                        if (tj * 4 + 4 < w)
-                        {
-                            float tmp[2] = {bfloat16_to_float32(r0[4]), bfloat16_to_float32(r1[4])};
-                            _r4 = vld1_f32(tmp);
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            float tmp[2] = {bfloat16_to_float32(r0[5]), bfloat16_to_float32(r1[5])};
-                            _r5 = vld1_f32(tmp);
-                        }
-#else
-                        r00 = bfloat16_to_float32(r0[0]);
-                        r01 = bfloat16_to_float32(r1[0]);
-                        if (tj * 4 + 1 < w)
-                        {
-                            r10 = bfloat16_to_float32(r0[1]);
-                            r11 = bfloat16_to_float32(r1[1]);
-                        }
-                        if (tj * 4 + 2 < w)
-                        {
-                            r20 = bfloat16_to_float32(r0[2]);
-                            r21 = bfloat16_to_float32(r1[2]);
-                        }
-                        if (tj * 4 + 3 < w)
-                        {
-                            r30 = bfloat16_to_float32(r0[3]);
-                            r31 = bfloat16_to_float32(r1[3]);
-                        }
-                        if (tj * 4 + 4 < w)
-                        {
-                            r40 = bfloat16_to_float32(r0[4]);
-                            r41 = bfloat16_to_float32(r1[4]);
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            r50 = bfloat16_to_float32(r0[5]);
-                            r51 = bfloat16_to_float32(r1[5]);
-                        }
-#endif
-                    }
-                }
-
-#if __ARM_NEON
-#if __aarch64__
-                float32x2_t _tmp12a = vfma_laneq_f32(vmul_laneq_f32(_r1, _coeffs, 0), _r3, _coeffs, 1);
-                float32x2_t _tmp12b = vfma_laneq_f32(_r4, _r2, _coeffs, 2);
-                float32x2_t _tmp34a = vfma_laneq_f32(vmul_laneq_f32(_r3, _coeffs, 0), _r1, _coeffs, 1);
-                float32x2_t _tmp34b = vfma_laneq_f32(_r4, _r2, _coeffs, 3);
-#else
-                float32x2_t _tmp12a = vmla_lane_f32(vmul_lane_f32(_r1, vget_low_f32(_coeffs), 0), _r3, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp12b = vmla_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 0);
-                float32x2_t _tmp34a = vmla_lane_f32(vmul_lane_f32(_r3, vget_low_f32(_coeffs), 0), _r1, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp34b = vmla_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 1);
-#endif
-
-#if __aarch64__
-                float32x2_t _tmp0 = vfma_f32(vadd_f32(_r0, _r4), _r2, _vm2_5);
-#else
-                float32x2_t _tmp0 = vmla_f32(vadd_f32(_r0, _r4), _r2, _vm2_5);
-#endif
-                float32x2_t _tmp1 = vsub_f32(_tmp12b, _tmp12a);
-                float32x2_t _tmp2 = vadd_f32(_tmp12b, _tmp12a);
-                float32x2_t _tmp3 = vadd_f32(_tmp34b, _tmp34a);
-                float32x2_t _tmp4 = vsub_f32(_tmp34b, _tmp34a);
-#if __aarch64__
-                float32x2_t _tmp5 = vfma_f32(vadd_f32(_r1, _r5), _r3, _vm2_5);
-#else
-                float32x2_t _tmp5 = vmla_f32(vadd_f32(_r1, _r5), _r3, _vm2_5);
-#endif
-
-                vst1_f32(tmp[0][m], _tmp0);
-                vst1_f32(tmp[1][m], _tmp1);
-                vst1_f32(tmp[2][m], _tmp2);
-                vst1_f32(tmp[3][m], _tmp3);
-                vst1_f32(tmp[4][m], _tmp4);
-                vst1_f32(tmp[5][m], _tmp5);
-#else
-                float tmp12a0 = sq2 * r10 - sq2_d2 * r30;
-                float tmp12a1 = sq2 * r11 - sq2_d2 * r31;
-                float tmp12b0 = r40 - 2 * r20;
-                float tmp12b1 = r41 - 2 * r21;
-                float tmp34a0 = sq2 * r30 - sq2_d2 * r10;
-                float tmp34a1 = sq2 * r31 - sq2_d2 * r11;
-                float tmp34b0 = r40 - 0.5f * r20;
-                float tmp34b1 = r41 - 0.5f * r21;
-
-                tmp[0][m][0] = r00 + r40 - 2.5f * r20;
-                tmp[0][m][1] = r01 + r41 - 2.5f * r21;
-                tmp[1][m][0] = tmp12b0 - tmp12a0;
-                tmp[1][m][1] = tmp12b1 - tmp12a1;
-                tmp[2][m][0] = tmp12b0 + tmp12a0;
-                tmp[2][m][1] = tmp12b1 + tmp12a1;
-                tmp[3][m][0] = tmp34b0 + tmp34a0;
-                tmp[3][m][1] = tmp34b1 + tmp34a1;
-                tmp[4][m][0] = tmp34b0 - tmp34a0;
-                tmp[4][m][1] = tmp34b1 - tmp34a1;
-                tmp[5][m][0] = r10 + r50 - 2.5f * r30;
-                tmp[5][m][1] = r11 + r51 - 2.5f * r31;
-#endif
-
-                r0 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 36 + jj * 2;
-            float* p1 = p0 + max_jj * 2;
-            float* p2 = p0 + max_jj * 2 * 2;
-            float* p3 = p0 + max_jj * 2 * 3;
-            float* p4 = p0 + max_jj * 2 * 4;
-            float* p5 = p0 + max_jj * 2 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(tmp[m][0]);
-                float32x2_t _r1 = vld1_f32(tmp[m][1]);
-                float32x2_t _r2 = vld1_f32(tmp[m][2]);
-                float32x2_t _r3 = vld1_f32(tmp[m][3]);
-                float32x2_t _r4 = vld1_f32(tmp[m][4]);
-                float32x2_t _r5 = vld1_f32(tmp[m][5]);
-
-#if __aarch64__
-                float32x2_t _tmp12a = vfma_laneq_f32(vmul_laneq_f32(_r1, _coeffs, 0), _r3, _coeffs, 1);
-                float32x2_t _tmp12b = vfma_laneq_f32(_r4, _r2, _coeffs, 2);
-                float32x2_t _tmp34a = vfma_laneq_f32(vmul_laneq_f32(_r3, _coeffs, 0), _r1, _coeffs, 1);
-                float32x2_t _tmp34b = vfma_laneq_f32(_r4, _r2, _coeffs, 3);
-#else
-                float32x2_t _tmp12a = vmla_lane_f32(vmul_lane_f32(_r1, vget_low_f32(_coeffs), 0), _r3, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp12b = vmla_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 0);
-                float32x2_t _tmp34a = vmla_lane_f32(vmul_lane_f32(_r3, vget_low_f32(_coeffs), 0), _r1, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp34b = vmla_lane_f32(_r4, _r2, vget_high_f32(_coeffs), 1);
-#endif
-
-#if __aarch64__
-                float32x2_t _tmp0 = vfma_f32(vadd_f32(_r0, _r4), _r2, _vm2_5);
-#else
-                float32x2_t _tmp0 = vmla_f32(vadd_f32(_r0, _r4), _r2, _vm2_5);
-#endif
-                float32x2_t _tmp1 = vsub_f32(_tmp12b, _tmp12a);
-                float32x2_t _tmp2 = vadd_f32(_tmp12b, _tmp12a);
-                float32x2_t _tmp3 = vadd_f32(_tmp34b, _tmp34a);
-                float32x2_t _tmp4 = vsub_f32(_tmp34b, _tmp34a);
-#if __aarch64__
-                float32x2_t _tmp5 = vfma_f32(vadd_f32(_r1, _r5), _r3, _vm2_5);
-#else
-                float32x2_t _tmp5 = vmla_f32(vadd_f32(_r1, _r5), _r3, _vm2_5);
-#endif
-
-                vst1_f32(p0, _tmp0);
-                vst1_f32(p1, _tmp1);
-                vst1_f32(p2, _tmp2);
-                vst1_f32(p3, _tmp3);
-                vst1_f32(p4, _tmp4);
-                vst1_f32(p5, _tmp5);
-#else
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-                float r40 = tmp[m][4][0];
-                float r41 = tmp[m][4][1];
-                float r50 = tmp[m][5][0];
-                float r51 = tmp[m][5][1];
-
-                float tmp12a0 = sq2 * r10 - sq2_d2 * r30;
-                float tmp12a1 = sq2 * r11 - sq2_d2 * r31;
-                float tmp12b0 = r40 - 2 * r20;
-                float tmp12b1 = r41 - 2 * r21;
-                float tmp34a0 = sq2 * r30 - sq2_d2 * r10;
-                float tmp34a1 = sq2 * r31 - sq2_d2 * r11;
-                float tmp34b0 = r40 - 0.5f * r20;
-                float tmp34b1 = r41 - 0.5f * r21;
-
-                p0[0] = r00 + r40 - 2.5f * r20;
-                p0[1] = r01 + r41 - 2.5f * r21;
-                p1[0] = tmp12b0 - tmp12a0;
-                p1[1] = tmp12b1 - tmp12a1;
-                p2[0] = tmp12b0 + tmp12a0;
-                p2[1] = tmp12b1 + tmp12a1;
-                p3[0] = tmp34b0 + tmp34a0;
-                p3[1] = tmp34b1 + tmp34a1;
-                p4[0] = tmp34b0 - tmp34a0;
-                p4[1] = tmp34b1 - tmp34a1;
-                p5[0] = r10 + r50 - 2.5f * r30;
-                p5[1] = r11 + r51 - 2.5f * r31;
-#endif
-
-                p0 += max_jj * 6 * 2;
-                p1 += max_jj * 6 * 2;
-                p2 += max_jj * 6 * 2;
-                p3 += max_jj * 6 * 2;
-                p4 += max_jj * 6 * 2;
-                p5 += max_jj * 6 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        float tmp[6][6];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const unsigned short* r0123 = bottom_blob.channel(k + kk).row<const unsigned short>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 6; m++)
-            {
-                float r0 = 0.f;
-                float r1 = 0.f;
-                float r2 = 0.f;
-                float r3 = 0.f;
-                float r4 = 0.f;
-                float r5 = 0.f;
-
-                if (ti * 4 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = bfloat16_to_float32(r0123[0]);
-                        if (tj * 4 + 1 < w) r1 = bfloat16_to_float32(r0123[1]);
-                        if (tj * 4 + 2 < w) r2 = bfloat16_to_float32(r0123[2]);
-                        if (tj * 4 + 3 < w) r3 = bfloat16_to_float32(r0123[3]);
-                        if (tj * 4 + 4 < w) r4 = bfloat16_to_float32(r0123[4]);
-                        if (tj * 4 + 5 < w) r5 = bfloat16_to_float32(r0123[5]);
-                    }
-                }
-
-                float tmp12a = sq2 * r1 - sq2_d2 * r3;
-                float tmp12b = r4 - 2 * r2;
-                float tmp34a = sq2 * r3 - sq2_d2 * r1;
-                float tmp34b = r4 - 0.5f * r2;
-
-                tmp[0][m] = r0 + r4 - 2.5f * r2;
-                tmp[1][m] = tmp12b - tmp12a;
-                tmp[2][m] = tmp12b + tmp12a;
-                tmp[3][m] = tmp34b + tmp34a;
-                tmp[4][m] = tmp34b - tmp34a;
-                tmp[5][m] = r1 + r5 - 2.5f * r3;
-
-                r0123 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 36 + jj;
-            float* p1 = p0 + max_jj;
-            float* p2 = p0 + max_jj * 2;
-            float* p3 = p0 + max_jj * 3;
-            float* p4 = p0 + max_jj * 4;
-            float* p5 = p0 + max_jj * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-                float r4 = tmp[m][4];
-                float r5 = tmp[m][5];
-
-                float tmp12a = sq2 * r1 - sq2_d2 * r3;
-                float tmp12b = r4 - 2 * r2;
-                float tmp34a = sq2 * r3 - sq2_d2 * r1;
-                float tmp34b = r4 - 0.5f * r2;
-
-                p0[0] = r0 + r4 - 2.5f * r2;
-                p1[0] = tmp12b - tmp12a;
-                p2[0] = tmp12b + tmp12a;
-                p3[0] = tmp34b + tmp34a;
-                p4[0] = tmp34b - tmp34a;
-                p5[0] = r1 + r5 - 2.5f * r3;
-
-                p0 += max_jj * 6;
-                p1 += max_jj * 6;
-                p2 += max_jj * 6;
-                p3 += max_jj * 6;
-                p4 += max_jj * 6;
-                p5 += max_jj * 6;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_output_tile_bf16s(const Mat& top_tile, Mat& top_blob, const Mat& bias, int i, int max_ii, int j, int max_jj)
-{
-    const float sq2 = 1.41421356237;
-    const float sq2_m2 = 1.41421356237 * 2;
-    const float sq2_d2 = 1.41421356237 / 2;
-    const float sq2_d4 = 1.41421356237 / 4;
-
-    // const float otm[4][6] = {
-    //     {1.0f, 1.0f,   1.0f,  1.0f,  1.0f,   0.0f},
-    //     {0.0f, sq2/2, -sq2/2, sq2,   -sq2,   0.0f},
-    //     {0.0f, 0.5f,   0.5f,  2.0f,  2.0f,   0.0f},
-    //     {0.0f, sq2/4, -sq2/4, sq2*2, -sq2*2, 1.0f}
-    // };
-
-    // 0 = r00 + (r01 + r02) + (r03 + r04)
-    // 1 =       (r01 - r02) * sq2_d2 + (r03 - r04) * sq2
-    // 2 =       (r01 + r02) * 0.5f + (r03 + r04) * 2
-    // 3 = r05 + (r01 - r02) * sq2_d4 + (r03 - r04) * sq2_m2
-
-#if __ARM_NEON
-    const float coeffs[6] = {sq2, sq2_d2, sq2_d4, sq2_m2, 0.5f, 2.f};
-    float32x4_t _coeffs = vld1q_f32(coeffs);
-    float32x2_t _coeffs2 = vld1_f32(coeffs + 4);
-#endif // __ARM_NEON
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 3) / 4;
-
-    const float* biasptr = bias;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        float32x4_t _bias0 = biasptr ? vld1q_f32(biasptr + i + ii) : vdupq_n_f32(0.f);
-        float32x4_t _bias1 = biasptr ? vld1q_f32(biasptr + i + ii + 4) : vdupq_n_f32(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[4][6][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 36 + jj * 8;
-            const float* r1 = r0 + max_jj * 8;
-            const float* r2 = r0 + max_jj * 8 * 2;
-            const float* r3 = r0 + max_jj * 8 * 3;
-            const float* r4 = r0 + max_jj * 8 * 4;
-            const float* r5 = r0 + max_jj * 8 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r30 = vld1q_f32(r3);
-                float32x4_t _r31 = vld1q_f32(r3 + 4);
-                float32x4_t _r40 = vld1q_f32(r4);
-                float32x4_t _r41 = vld1q_f32(r4 + 4);
-                float32x4_t _r50 = vld1q_f32(r5);
-                float32x4_t _r51 = vld1q_f32(r5 + 4);
-
-                float32x4_t _tmp02a0 = vaddq_f32(_r10, _r20);
-                float32x4_t _tmp02a1 = vaddq_f32(_r11, _r21);
-                float32x4_t _tmp02b0 = vaddq_f32(_r30, _r40);
-                float32x4_t _tmp02b1 = vaddq_f32(_r31, _r41);
-                float32x4_t _tmp13a0 = vsubq_f32(_r10, _r20);
-                float32x4_t _tmp13a1 = vsubq_f32(_r11, _r21);
-                float32x4_t _tmp13b0 = vsubq_f32(_r30, _r40);
-                float32x4_t _tmp13b1 = vsubq_f32(_r31, _r41);
-
-                float32x4_t _tmp00 = vaddq_f32(vaddq_f32(_r00, _tmp02a0), _tmp02b0);
-                float32x4_t _tmp01 = vaddq_f32(vaddq_f32(_r01, _tmp02a1), _tmp02b1);
-                float32x4_t _tmp10 = vfmaq_laneq_f32(vmulq_laneq_f32(_tmp13a0, _coeffs, 1), _tmp13b0, _coeffs, 0);
-                float32x4_t _tmp11 = vfmaq_laneq_f32(vmulq_laneq_f32(_tmp13a1, _coeffs, 1), _tmp13b1, _coeffs, 0);
-                float32x4_t _tmp20 = vfmaq_lane_f32(vmulq_lane_f32(_tmp02a0, _coeffs2, 0), _tmp02b0, _coeffs2, 1);
-                float32x4_t _tmp21 = vfmaq_lane_f32(vmulq_lane_f32(_tmp02a1, _coeffs2, 0), _tmp02b1, _coeffs2, 1);
-                float32x4_t _tmp30 = vfmaq_laneq_f32(vfmaq_laneq_f32(_r50, _tmp13a0, _coeffs, 2), _tmp13b0, _coeffs, 3);
-                float32x4_t _tmp31 = vfmaq_laneq_f32(vfmaq_laneq_f32(_r51, _tmp13a1, _coeffs, 2), _tmp13b1, _coeffs, 3);
-
-                vst1q_f32(tmp[0][m], _tmp00);
-                vst1q_f32(tmp[0][m] + 4, _tmp01);
-                vst1q_f32(tmp[1][m], _tmp10);
-                vst1q_f32(tmp[1][m] + 4, _tmp11);
-                vst1q_f32(tmp[2][m], _tmp20);
-                vst1q_f32(tmp[2][m] + 4, _tmp21);
-                vst1q_f32(tmp[3][m], _tmp30);
-                vst1q_f32(tmp[3][m] + 4, _tmp31);
-
-                r0 += max_jj * 6 * 8;
-                r1 += max_jj * 6 * 8;
-                r2 += max_jj * 6 * 8;
-                r3 += max_jj * 6 * 8;
-                r4 += max_jj * 6 * 8;
-                r5 += max_jj * 6 * 8;
-            }
-
-            unsigned short* outptr0 = top_blob.channel((i + ii) / out_elempack).row<unsigned short>(ti * 4) + (tj * 4) * out_elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                float32x4_t _r00 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r01 = vld1q_f32(tmp[m][0] + 4);
-                float32x4_t _r10 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r11 = vld1q_f32(tmp[m][1] + 4);
-                float32x4_t _r20 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r21 = vld1q_f32(tmp[m][2] + 4);
-                float32x4_t _r30 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r31 = vld1q_f32(tmp[m][3] + 4);
-                float32x4_t _r40 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r41 = vld1q_f32(tmp[m][4] + 4);
-                float32x4_t _r50 = vld1q_f32(tmp[m][5]);
-                float32x4_t _r51 = vld1q_f32(tmp[m][5] + 4);
-
-                float32x4_t _tmp02a0 = vaddq_f32(_r10, _r20);
-                float32x4_t _tmp02a1 = vaddq_f32(_r11, _r21);
-                float32x4_t _tmp02b0 = vaddq_f32(_r30, _r40);
-                float32x4_t _tmp02b1 = vaddq_f32(_r31, _r41);
-                float32x4_t _tmp13a0 = vsubq_f32(_r10, _r20);
-                float32x4_t _tmp13a1 = vsubq_f32(_r11, _r21);
-                float32x4_t _tmp13b0 = vsubq_f32(_r30, _r40);
-                float32x4_t _tmp13b1 = vsubq_f32(_r31, _r41);
-
-                float32x4_t _tmp00 = vaddq_f32(vaddq_f32(_r00, _tmp02a0), vaddq_f32(_tmp02b0, _bias0));
-                float32x4_t _tmp01 = vaddq_f32(vaddq_f32(_r01, _tmp02a1), vaddq_f32(_tmp02b1, _bias1));
-                float32x4_t _tmp10 = vfmaq_laneq_f32(vfmaq_laneq_f32(_bias0, _tmp13a0, _coeffs, 1), _tmp13b0, _coeffs, 0);
-                float32x4_t _tmp11 = vfmaq_laneq_f32(vfmaq_laneq_f32(_bias1, _tmp13a1, _coeffs, 1), _tmp13b1, _coeffs, 0);
-                float32x4_t _tmp20 = vfmaq_lane_f32(vfmaq_lane_f32(_bias0, _tmp02a0, _coeffs2, 0), _tmp02b0, _coeffs2, 1);
-                float32x4_t _tmp21 = vfmaq_lane_f32(vfmaq_lane_f32(_bias1, _tmp02a1, _coeffs2, 0), _tmp02b1, _coeffs2, 1);
-                float32x4_t _tmp30 = vfmaq_laneq_f32(vfmaq_laneq_f32(vaddq_f32(_r50, _bias0), _tmp13a0, _coeffs, 2), _tmp13b0, _coeffs, 3);
-                float32x4_t _tmp31 = vfmaq_laneq_f32(vfmaq_laneq_f32(vaddq_f32(_r51, _bias1), _tmp13a1, _coeffs, 2), _tmp13b1, _coeffs, 3);
-
-                if (out_elempack == 4)
-                {
-                    unsigned short* outptr1 = outptr0 + N;
-
-                    vst1_u16(outptr0, float2bfloat(_tmp00));
-                    vst1_u16(outptr1, float2bfloat(_tmp01));
-                    if (tj * 4 + 1 < outw)
-                    {
-                        vst1_u16(outptr0 + 4, float2bfloat(_tmp10));
-                        vst1_u16(outptr1 + 4, float2bfloat(_tmp11));
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        vst1_u16(outptr0 + 8, float2bfloat(_tmp20));
-                        vst1_u16(outptr1 + 8, float2bfloat(_tmp21));
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        vst1_u16(outptr0 + 12, float2bfloat(_tmp30));
-                        vst1_u16(outptr1 + 12, float2bfloat(_tmp31));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short tmp0[8];
-                    unsigned short tmp1[8];
-                    unsigned short tmp2[8];
-                    unsigned short tmp3[8];
-                    vst1_u16(tmp0, float2bfloat(_tmp00));
-                    vst1_u16(tmp0 + 4, float2bfloat(_tmp01));
-                    vst1_u16(tmp1, float2bfloat(_tmp10));
-                    vst1_u16(tmp1 + 4, float2bfloat(_tmp11));
-                    vst1_u16(tmp2, float2bfloat(_tmp20));
-                    vst1_u16(tmp2 + 4, float2bfloat(_tmp21));
-                    vst1_u16(tmp3, float2bfloat(_tmp30));
-                    vst1_u16(tmp3 + 4, float2bfloat(_tmp31));
-
-                    unsigned short* outptr1 = outptr0 + N;
-                    unsigned short* outptr2 = outptr0 + N * 2;
-                    unsigned short* outptr3 = outptr0 + N * 3;
-                    unsigned short* outptr4 = outptr0 + N * 4;
-                    unsigned short* outptr5 = outptr0 + N * 5;
-                    unsigned short* outptr6 = outptr0 + N * 6;
-                    unsigned short* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                        outptr4[2] = tmp2[4];
-                        outptr5[2] = tmp2[5];
-                        outptr6[2] = tmp2[6];
-                        outptr7[2] = tmp2[7];
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                        outptr4[3] = tmp3[4];
-                        outptr5[3] = tmp3[5];
-                        outptr6[3] = tmp3[6];
-                        outptr7[3] = tmp3[7];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        float32x4_t _bias0 = biasptr ? vld1q_f32(biasptr + i + ii) : vdupq_n_f32(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[4][6][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 36 + jj * 4;
-            const float* r1 = r0 + max_jj * 4;
-            const float* r2 = r0 + max_jj * 4 * 2;
-            const float* r3 = r0 + max_jj * 4 * 3;
-            const float* r4 = r0 + max_jj * 4 * 4;
-            const float* r5 = r0 + max_jj * 4 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float32x4_t _r0 = vld1q_f32(r0);
-                float32x4_t _r1 = vld1q_f32(r1);
-                float32x4_t _r2 = vld1q_f32(r2);
-                float32x4_t _r3 = vld1q_f32(r3);
-                float32x4_t _r4 = vld1q_f32(r4);
-                float32x4_t _r5 = vld1q_f32(r5);
-
-                float32x4_t _tmp02a = vaddq_f32(_r1, _r2);
-                float32x4_t _tmp02b = vaddq_f32(_r3, _r4);
-                float32x4_t _tmp13a = vsubq_f32(_r1, _r2);
-                float32x4_t _tmp13b = vsubq_f32(_r3, _r4);
-
-                float32x4_t _tmp0 = vaddq_f32(vaddq_f32(_r0, _tmp02a), _tmp02b);
-#if __aarch64__
-                float32x4_t _tmp1 = vfmaq_laneq_f32(vmulq_laneq_f32(_tmp13a, _coeffs, 1), _tmp13b, _coeffs, 0);
-                float32x4_t _tmp2 = vfmaq_lane_f32(vmulq_lane_f32(_tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x4_t _tmp3 = vfmaq_laneq_f32(vfmaq_laneq_f32(_r5, _tmp13a, _coeffs, 2), _tmp13b, _coeffs, 3);
-#else
-                float32x4_t _tmp1 = vmlaq_lane_f32(vmulq_lane_f32(_tmp13a, vget_low_f32(_coeffs), 1), _tmp13b, vget_low_f32(_coeffs), 0);
-                float32x4_t _tmp2 = vmlaq_lane_f32(vmulq_lane_f32(_tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x4_t _tmp3 = vmlaq_lane_f32(vmlaq_lane_f32(_r5, _tmp13a, vget_high_f32(_coeffs), 0), _tmp13b, vget_high_f32(_coeffs), 1);
-#endif
-
-                vst1q_f32(tmp[0][m], _tmp0);
-                vst1q_f32(tmp[1][m], _tmp1);
-                vst1q_f32(tmp[2][m], _tmp2);
-                vst1q_f32(tmp[3][m], _tmp3);
-
-                r0 += max_jj * 6 * 4;
-                r1 += max_jj * 6 * 4;
-                r2 += max_jj * 6 * 4;
-                r3 += max_jj * 6 * 4;
-                r4 += max_jj * 6 * 4;
-                r5 += max_jj * 6 * 4;
-            }
-
-            unsigned short* outptr0 = top_blob.channel((i + ii) / out_elempack).row<unsigned short>(ti * 4) + (tj * 4) * out_elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                float32x4_t _r0 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r1 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r2 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r3 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r4 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r5 = vld1q_f32(tmp[m][5]);
-
-                float32x4_t _tmp02a = vaddq_f32(_r1, _r2);
-                float32x4_t _tmp02b = vaddq_f32(_r3, _r4);
-                float32x4_t _tmp13a = vsubq_f32(_r1, _r2);
-                float32x4_t _tmp13b = vsubq_f32(_r3, _r4);
-
-                float32x4_t _tmp0 = vaddq_f32(vaddq_f32(_r0, _tmp02a), vaddq_f32(_tmp02b, _bias0));
-#if __aarch64__
-                float32x4_t _tmp1 = vfmaq_laneq_f32(vfmaq_laneq_f32(_bias0, _tmp13a, _coeffs, 1), _tmp13b, _coeffs, 0);
-                float32x4_t _tmp2 = vfmaq_lane_f32(vfmaq_lane_f32(_bias0, _tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x4_t _tmp3 = vfmaq_laneq_f32(vfmaq_laneq_f32(vaddq_f32(_r5, _bias0), _tmp13a, _coeffs, 2), _tmp13b, _coeffs, 3);
-#else
-                float32x4_t _tmp1 = vmlaq_lane_f32(vmlaq_lane_f32(_bias0, _tmp13a, vget_low_f32(_coeffs), 1), _tmp13b, vget_low_f32(_coeffs), 0);
-                float32x4_t _tmp2 = vmlaq_lane_f32(vmlaq_lane_f32(_bias0, _tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x4_t _tmp3 = vmlaq_lane_f32(vmlaq_lane_f32(vaddq_f32(_r5, _bias0), _tmp13a, vget_high_f32(_coeffs), 0), _tmp13b, vget_high_f32(_coeffs), 1);
-#endif
-
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_tmp0));
-                    if (tj * 4 + 1 < outw) vst1_u16(outptr0 + 4, float2bfloat(_tmp1));
-                    if (tj * 4 + 2 < outw) vst1_u16(outptr0 + 8, float2bfloat(_tmp2));
-                    if (tj * 4 + 3 < outw) vst1_u16(outptr0 + 12, float2bfloat(_tmp3));
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short tmp0[4];
-                    unsigned short tmp1[4];
-                    unsigned short tmp2[4];
-                    unsigned short tmp3[4];
-                    vst1_u16(tmp0, float2bfloat(_tmp0));
-                    vst1_u16(tmp1, float2bfloat(_tmp1));
-                    vst1_u16(tmp2, float2bfloat(_tmp2));
-                    vst1_u16(tmp3, float2bfloat(_tmp3));
-
-                    unsigned short* outptr1 = outptr0 + N;
-                    unsigned short* outptr2 = outptr0 + N * 2;
-                    unsigned short* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-#if __ARM_NEON
-        float32x2_t _bias0 = biasptr ? vld1_f32(biasptr + i + ii) : vdup_n_f32(0.f);
-#else
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-        float bias1 = biasptr ? biasptr[i + ii + 1] : 0.f;
-#endif
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        float tmp[4][6][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 36 + jj * 2;
-            const float* r1 = r0 + max_jj * 2;
-            const float* r2 = r0 + max_jj * 2 * 2;
-            const float* r3 = r0 + max_jj * 2 * 3;
-            const float* r4 = r0 + max_jj * 2 * 4;
-            const float* r5 = r0 + max_jj * 2 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(r0);
-                float32x2_t _r1 = vld1_f32(r1);
-                float32x2_t _r2 = vld1_f32(r2);
-                float32x2_t _r3 = vld1_f32(r3);
-                float32x2_t _r4 = vld1_f32(r4);
-                float32x2_t _r5 = vld1_f32(r5);
-
-                float32x2_t _tmp02a = vadd_f32(_r1, _r2);
-                float32x2_t _tmp02b = vadd_f32(_r3, _r4);
-                float32x2_t _tmp13a = vsub_f32(_r1, _r2);
-                float32x2_t _tmp13b = vsub_f32(_r3, _r4);
-
-                float32x2_t _tmp0 = vadd_f32(vadd_f32(_r0, _tmp02a), _tmp02b);
-#if __aarch64__
-                float32x2_t _tmp1 = vfma_laneq_f32(vmul_laneq_f32(_tmp13a, _coeffs, 1), _tmp13b, _coeffs, 0);
-                float32x2_t _tmp2 = vfma_lane_f32(vmul_lane_f32(_tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x2_t _tmp3 = vfma_laneq_f32(vfma_laneq_f32(_r5, _tmp13a, _coeffs, 2), _tmp13b, _coeffs, 3);
-#else
-                float32x2_t _tmp1 = vmla_lane_f32(vmul_lane_f32(_tmp13a, vget_low_f32(_coeffs), 1), _tmp13b, vget_low_f32(_coeffs), 0);
-                float32x2_t _tmp2 = vmla_lane_f32(vmul_lane_f32(_tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x2_t _tmp3 = vmla_lane_f32(vmla_lane_f32(_r5, _tmp13a, vget_high_f32(_coeffs), 0), _tmp13b, vget_high_f32(_coeffs), 1);
-#endif
-
-                vst1_f32(tmp[0][m], _tmp0);
-                vst1_f32(tmp[1][m], _tmp1);
-                vst1_f32(tmp[2][m], _tmp2);
-                vst1_f32(tmp[3][m], _tmp3);
-#else
-                float tmp02a0 = r1[0] + r2[0];
-                float tmp02a1 = r1[1] + r2[1];
-                float tmp02b0 = r3[0] + r4[0];
-                float tmp02b1 = r3[1] + r4[1];
-                float tmp13a0 = r1[0] - r2[0];
-                float tmp13a1 = r1[1] - r2[1];
-                float tmp13b0 = r3[0] - r4[0];
-                float tmp13b1 = r3[1] - r4[1];
-
-                tmp[0][m][0] = r0[0] + tmp02a0 + tmp02b0;
-                tmp[0][m][1] = r0[1] + tmp02a1 + tmp02b1;
-                tmp[1][m][0] = tmp13a0 * sq2_d2 + tmp13b0 * sq2;
-                tmp[1][m][1] = tmp13a1 * sq2_d2 + tmp13b1 * sq2;
-                tmp[2][m][0] = tmp02a0 * 0.5f + tmp02b0 * 2;
-                tmp[2][m][1] = tmp02a1 * 0.5f + tmp02b1 * 2;
-                tmp[3][m][0] = r5[0] + tmp13a0 * sq2_d4 + tmp13b0 * sq2_m2;
-                tmp[3][m][1] = r5[1] + tmp13a1 * sq2_d4 + tmp13b1 * sq2_m2;
-#endif
-
-                r0 += max_jj * 6 * 2;
-                r1 += max_jj * 6 * 2;
-                r2 += max_jj * 6 * 2;
-                r3 += max_jj * 6 * 2;
-                r4 += max_jj * 6 * 2;
-                r5 += max_jj * 6 * 2;
-            }
-
-            unsigned short* outptr0 = top_blob.channel(i + ii).row<unsigned short>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(tmp[m][0]);
-                float32x2_t _r1 = vld1_f32(tmp[m][1]);
-                float32x2_t _r2 = vld1_f32(tmp[m][2]);
-                float32x2_t _r3 = vld1_f32(tmp[m][3]);
-                float32x2_t _r4 = vld1_f32(tmp[m][4]);
-                float32x2_t _r5 = vld1_f32(tmp[m][5]);
-
-                float32x2_t _tmp02a = vadd_f32(_r1, _r2);
-                float32x2_t _tmp02b = vadd_f32(_r3, _r4);
-                float32x2_t _tmp13a = vsub_f32(_r1, _r2);
-                float32x2_t _tmp13b = vsub_f32(_r3, _r4);
-
-                float32x2_t _tmp0 = vadd_f32(vadd_f32(_r0, _tmp02a), vadd_f32(_tmp02b, _bias0));
-#if __aarch64__
-                float32x2_t _tmp1 = vfma_laneq_f32(vfma_laneq_f32(_bias0, _tmp13a, _coeffs, 1), _tmp13b, _coeffs, 0);
-                float32x2_t _tmp2 = vfma_lane_f32(vfma_lane_f32(_bias0, _tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x2_t _tmp3 = vfma_laneq_f32(vfma_laneq_f32(vadd_f32(_r5, _bias0), _tmp13a, _coeffs, 2), _tmp13b, _coeffs, 3);
-#else
-                float32x2_t _tmp1 = vmla_lane_f32(vmla_lane_f32(_bias0, _tmp13a, vget_low_f32(_coeffs), 1), _tmp13b, vget_low_f32(_coeffs), 0);
-                float32x2_t _tmp2 = vmla_lane_f32(vmla_lane_f32(_bias0, _tmp02a, _coeffs2, 0), _tmp02b, _coeffs2, 1);
-                float32x2_t _tmp3 = vmla_lane_f32(vmla_lane_f32(vadd_f32(_r5, _bias0), _tmp13a, vget_high_f32(_coeffs), 0), _tmp13b, vget_high_f32(_coeffs), 1);
-#endif
-#else
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-                float r40 = tmp[m][4][0];
-                float r41 = tmp[m][4][1];
-                float r50 = tmp[m][5][0];
-                float r51 = tmp[m][5][1];
-
-                float tmp02a0 = r10 + r20;
-                float tmp02a1 = r11 + r21;
-                float tmp02b0 = r30 + r40;
-                float tmp02b1 = r31 + r41;
-                float tmp13a0 = r10 - r20;
-                float tmp13a1 = r11 - r21;
-                float tmp13b0 = r30 - r40;
-                float tmp13b1 = r31 - r41;
-
-                float tmp00 = bias0 + r00 + tmp02a0 + tmp02b0;
-                float tmp01 = bias1 + r01 + tmp02a1 + tmp02b1;
-                float tmp10 = bias0 + tmp13a0 * sq2_d2 + tmp13b0 * sq2;
-                float tmp11 = bias1 + tmp13a1 * sq2_d2 + tmp13b1 * sq2;
-                float tmp20 = bias0 + tmp02a0 * 0.5f + tmp02b0 * 2;
-                float tmp21 = bias1 + tmp02a1 * 0.5f + tmp02b1 * 2;
-                float tmp30 = bias0 + r50 + tmp13a0 * sq2_d4 + tmp13b0 * sq2_m2;
-                float tmp31 = bias1 + r51 + tmp13a1 * sq2_d4 + tmp13b1 * sq2_m2;
-#endif
-
-                // if (out_elempack == 1)
-                {
-                    unsigned short* outptr1 = outptr0 + N;
-
-#if __ARM_NEON
-                    uint16x4_t _tmp01 = float2bfloat(vcombine_f32(_tmp0, _tmp1));
-                    uint16x4_t _tmp23 = float2bfloat(vcombine_f32(_tmp2, _tmp3));
-
-                    outptr0[0] = vget_lane_u16(_tmp01, 0);
-                    outptr1[0] = vget_lane_u16(_tmp01, 1);
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = vget_lane_u16(_tmp01, 2);
-                        outptr1[1] = vget_lane_u16(_tmp01, 3);
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = vget_lane_u16(_tmp23, 0);
-                        outptr1[2] = vget_lane_u16(_tmp23, 1);
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = vget_lane_u16(_tmp23, 2);
-                        outptr1[3] = vget_lane_u16(_tmp23, 3);
-                    }
-#else
-                    outptr0[0] = float32_to_bfloat16(tmp00);
-                    outptr1[0] = float32_to_bfloat16(tmp01);
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = float32_to_bfloat16(tmp10);
-                        outptr1[1] = float32_to_bfloat16(tmp11);
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = float32_to_bfloat16(tmp20);
-                        outptr1[2] = float32_to_bfloat16(tmp21);
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = float32_to_bfloat16(tmp30);
-                        outptr1[3] = float32_to_bfloat16(tmp31);
-                    }
-#endif
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-
-        float tmp[4][6];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 36 + jj;
-            const float* r1 = r0 + max_jj;
-            const float* r2 = r0 + max_jj * 2;
-            const float* r3 = r0 + max_jj * 3;
-            const float* r4 = r0 + max_jj * 4;
-            const float* r5 = r0 + max_jj * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float tmp02a = r1[0] + r2[0];
-                float tmp02b = r3[0] + r4[0];
-                float tmp13a = r1[0] - r2[0];
-                float tmp13b = r3[0] - r4[0];
-
-                tmp[0][m] = r0[0] + tmp02a + tmp02b;
-                tmp[1][m] = tmp13a * sq2_d2 + tmp13b * sq2;
-                tmp[2][m] = tmp02a * 0.5f + tmp02b * 2;
-                tmp[3][m] = r5[0] + tmp13a * sq2_d4 + tmp13b * sq2_m2;
-
-                r0 += max_jj * 6;
-                r1 += max_jj * 6;
-                r2 += max_jj * 6;
-                r3 += max_jj * 6;
-                r4 += max_jj * 6;
-                r5 += max_jj * 6;
-            }
-
-            unsigned short* outptr0 = top_blob.channel(i + ii).row<unsigned short>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-                float r4 = tmp[m][4];
-                float r5 = tmp[m][5];
-
-                float tmp02a = r1 + r2;
-                float tmp02b = r3 + r4;
-                float tmp13a = r1 - r2;
-                float tmp13b = r3 - r4;
-
-                float tmp0 = bias0 + r0 + tmp02a + tmp02b;
-                float tmp1 = bias0 + tmp13a * sq2_d2 + tmp13b * sq2;
-                float tmp2 = bias0 + tmp02a * 0.5f + tmp02b * 2;
-                float tmp3 = bias0 + r5 + tmp13a * sq2_d4 + tmp13b * sq2_m2;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = float32_to_bfloat16(tmp0);
-                    if (tj * 4 + 1 < outw) outptr0[1] = float32_to_bfloat16(tmp1);
-                    if (tj * 4 + 2 < outw) outptr0[2] = float32_to_bfloat16(tmp2);
-                    if (tj * 4 + 3 < outw) outptr0[3] = float32_to_bfloat16(tmp3);
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_bf16s(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, const Mat& bias, int nT, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 4n+2, winograd F(4,3)
-    int w_tiles = (outw + 3) / 4;
-    int h_tiles = (outh + 3) / 4;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 36;
-
-    // NCNN_LOGE("conv3x3s1_winograd43_bf16s %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    conv3x3s1_winograd_get_optimal_tile_mnk(M, N, K, B, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 4u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 4u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd43_transform_input_tile_bf16s(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 4u, opt.workspace_allocator);
-
-        #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd43_transform_input_tile_bf16s(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                conv3x3s1_winograd_gemm_transB_packed_tile(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk, opt.use_a53_a55_optimized_kernel);
-            }
-
-            // transform output
-            conv3x3s1_winograd43_transform_output_tile_bf16s(top_tile, top_blob, bias, i, max_ii, j, max_jj);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd63_transform_input_tile_bf16s(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    // const float itm[8][8] = {
-    //     {1.0f, 0.0f,-5.25f, 0.00f, 5.25f, 0.00f,-1.0f, 0.0f},
-    //     {0.0f, 1.0f, 1.00f,-4.25f,-4.25f, 1.00f, 1.0f, 0.0f},
-    //     {0.0f,-1.0f, 1.00f, 4.25f,-4.25f,-1.00f, 1.0f, 0.0f},
-    //     {0.0f, 0.5f, 0.25f,-2.50f,-1.25f, 2.00f, 1.0f, 0.0f},
-    //     {0.0f,-0.5f, 0.25f, 2.50f,-1.25f,-2.00f, 1.0f, 0.0f},
-    //     {0.0f, 2.0f, 4.00f,-2.50f,-5.00f, 0.50f, 1.0f, 0.0f},
-    //     {0.0f,-2.0f, 4.00f, 2.50f,-5.00f,-0.50f, 1.0f, 0.0f},
-    //     {0.0f,-1.0f, 0.00f, 5.25f, 0.00f,-5.25f, 0.0f, 1.0f}
-    // };
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w + 3) / 6;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-#if __ARM_NEON
-#if __aarch64__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[8][8][8];
-
-        const float coeffs[8] = {5.25f, -4.25f, -1.25f, 0.25f, -2.5f, 0.5f, 2.f, 4.f};
-        float32x4_t _coeffs = vld1q_f32(coeffs);
-        float32x4_t _coeffs2 = vld1q_f32(coeffs + 4);
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const unsigned short* r0 = bottom_blob.channel((k + kk) / elempack).row<const unsigned short>(ti * 6) + (tj * 6) * elempack;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float32x4_t _r00 = vdupq_n_f32(0.f);
-                float32x4_t _r01 = vdupq_n_f32(0.f);
-                float32x4_t _r10 = vdupq_n_f32(0.f);
-                float32x4_t _r11 = vdupq_n_f32(0.f);
-                float32x4_t _r20 = vdupq_n_f32(0.f);
-                float32x4_t _r21 = vdupq_n_f32(0.f);
-                float32x4_t _r30 = vdupq_n_f32(0.f);
-                float32x4_t _r31 = vdupq_n_f32(0.f);
-                float32x4_t _r40 = vdupq_n_f32(0.f);
-                float32x4_t _r41 = vdupq_n_f32(0.f);
-                float32x4_t _r50 = vdupq_n_f32(0.f);
-                float32x4_t _r51 = vdupq_n_f32(0.f);
-                float32x4_t _r60 = vdupq_n_f32(0.f);
-                float32x4_t _r61 = vdupq_n_f32(0.f);
-                float32x4_t _r70 = vdupq_n_f32(0.f);
-                float32x4_t _r71 = vdupq_n_f32(0.f);
-
-                if (ti * 6 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        const unsigned short* r1 = r0 + N;
-
-                        _r00 = bfloat2float(vld1_u16(r0));
-                        _r01 = bfloat2float(vld1_u16(r1));
-                        if (tj * 6 + 1 < w)
-                        {
-                            _r10 = bfloat2float(vld1_u16(r0 + 4));
-                            _r11 = bfloat2float(vld1_u16(r1 + 4));
-                        }
-                        if (tj * 6 + 2 < w)
-                        {
-                            _r20 = bfloat2float(vld1_u16(r0 + 8));
-                            _r21 = bfloat2float(vld1_u16(r1 + 8));
-                        }
-                        if (tj * 6 + 3 < w)
-                        {
-                            _r30 = bfloat2float(vld1_u16(r0 + 12));
-                            _r31 = bfloat2float(vld1_u16(r1 + 12));
-                        }
-                        if (tj * 6 + 4 < w)
-                        {
-                            _r40 = bfloat2float(vld1_u16(r0 + 16));
-                            _r41 = bfloat2float(vld1_u16(r1 + 16));
-                        }
-                        if (tj * 6 + 5 < w)
-                        {
-                            _r50 = bfloat2float(vld1_u16(r0 + 20));
-                            _r51 = bfloat2float(vld1_u16(r1 + 20));
-                        }
-                        if (tj * 6 + 6 < w)
-                        {
-                            _r60 = bfloat2float(vld1_u16(r0 + 24));
-                            _r61 = bfloat2float(vld1_u16(r1 + 24));
-                        }
-                        if (tj * 6 + 7 < w)
-                        {
-                            _r70 = bfloat2float(vld1_u16(r0 + 28));
-                            _r71 = bfloat2float(vld1_u16(r1 + 28));
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        const unsigned short* r1 = r0 + N;
-                        const unsigned short* r2 = r0 + N * 2;
-                        const unsigned short* r3 = r0 + N * 3;
-                        const unsigned short* r4 = r0 + N * 4;
-                        const unsigned short* r5 = r0 + N * 5;
-                        const unsigned short* r6 = r0 + N * 6;
-                        const unsigned short* r7 = r0 + N * 7;
-
-                        uint16x4_t _t0 = vld1_u16(r0);
-                        uint16x4_t _t1 = vld1_u16(r1);
-                        uint16x4_t _t2 = vld1_u16(r2);
-                        uint16x4_t _t3 = vld1_u16(r3);
-                        uint16x4_t _t4 = vld1_u16(r4);
-                        uint16x4_t _t5 = vld1_u16(r5);
-                        uint16x4_t _t6 = vld1_u16(r6);
-                        uint16x4_t _t7 = vld1_u16(r7);
-
-                        transpose4x4_u16(_t0, _t1, _t2, _t3);
-                        transpose4x4_u16(_t4, _t5, _t6, _t7);
-
-                        _r00 = bfloat2float(_t0);
-                        _r01 = bfloat2float(_t4);
-                        if (tj * 6 + 1 < w)
-                        {
-                            _r10 = bfloat2float(_t1);
-                            _r11 = bfloat2float(_t5);
-                        }
-                        if (tj * 6 + 2 < w)
-                        {
-                            _r20 = bfloat2float(_t2);
-                            _r21 = bfloat2float(_t6);
-                        }
-                        if (tj * 6 + 3 < w)
-                        {
-                            _r30 = bfloat2float(_t3);
-                            _r31 = bfloat2float(_t7);
-                        }
-                        if (tj * 6 + 4 < w)
-                        {
-                            _t0 = vld1_u16(r0 + 4);
-                            _t1 = vld1_u16(r1 + 4);
-                            _t2 = vld1_u16(r2 + 4);
-                            _t3 = vld1_u16(r3 + 4);
-                            _t4 = vld1_u16(r4 + 4);
-                            _t5 = vld1_u16(r5 + 4);
-                            _t6 = vld1_u16(r6 + 4);
-                            _t7 = vld1_u16(r7 + 4);
-
-                            transpose4x4_u16(_t0, _t1, _t2, _t3);
-                            transpose4x4_u16(_t4, _t5, _t6, _t7);
-
-                            _r40 = bfloat2float(_t0);
-                            _r41 = bfloat2float(_t4);
-                            if (tj * 6 + 5 < w)
-                            {
-                                _r50 = bfloat2float(_t1);
-                                _r51 = bfloat2float(_t5);
-                            }
-                            if (tj * 6 + 6 < w)
-                            {
-                                _r60 = bfloat2float(_t2);
-                                _r61 = bfloat2float(_t6);
-                            }
-                            if (tj * 6 + 7 < w)
-                            {
-                                _r70 = bfloat2float(_t3);
-                                _r71 = bfloat2float(_t7);
-                            }
-                        }
-                    }
-                }
-
-                float32x4_t _tmp12a0 = vfmaq_laneq_f32(vaddq_f32(_r20, _r60), _r40, _coeffs, 1);
-                float32x4_t _tmp12a1 = vfmaq_laneq_f32(vaddq_f32(_r21, _r61), _r41, _coeffs, 1);
-                float32x4_t _tmp12b0 = vfmaq_laneq_f32(vaddq_f32(_r10, _r50), _r30, _coeffs, 1);
-                float32x4_t _tmp12b1 = vfmaq_laneq_f32(vaddq_f32(_r11, _r51), _r31, _coeffs, 1);
-                float32x4_t _tmp34a0 = vfmaq_laneq_f32(vfmaq_laneq_f32(_r60, _r20, _coeffs, 3), _r40, _coeffs, 2);
-                float32x4_t _tmp34a1 = vfmaq_laneq_f32(vfmaq_laneq_f32(_r61, _r21, _coeffs, 3), _r41, _coeffs, 2);
-                float32x4_t _tmp34b0 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r10, _coeffs2, 1), _r30, _coeffs2, 0), _r50, _coeffs2, 2);
-                float32x4_t _tmp34b1 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r11, _coeffs2, 1), _r31, _coeffs2, 0), _r51, _coeffs2, 2);
-                float32x4_t _tmp56a0 = vfmaq_laneq_f32(_r60, vfmaq_laneq_f32(_r20, _r40, _coeffs, 2), _coeffs2, 3);
-                float32x4_t _tmp56a1 = vfmaq_laneq_f32(_r61, vfmaq_laneq_f32(_r21, _r41, _coeffs, 2), _coeffs2, 3);
-                float32x4_t _tmp56b0 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r10, _coeffs2, 2), _r30, _coeffs2, 0), _r50, _coeffs2, 1);
-                float32x4_t _tmp56b1 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r11, _coeffs2, 2), _r31, _coeffs2, 0), _r51, _coeffs2, 1);
-
-                float32x4_t _tmp00 = vfmaq_laneq_f32(vsubq_f32(_r00, _r60), vsubq_f32(_r40, _r20), _coeffs, 0);
-                float32x4_t _tmp01 = vfmaq_laneq_f32(vsubq_f32(_r01, _r61), vsubq_f32(_r41, _r21), _coeffs, 0);
-                float32x4_t _tmp10 = vaddq_f32(_tmp12a0, _tmp12b0);
-                float32x4_t _tmp11 = vaddq_f32(_tmp12a1, _tmp12b1);
-                float32x4_t _tmp20 = vsubq_f32(_tmp12a0, _tmp12b0);
-                float32x4_t _tmp21 = vsubq_f32(_tmp12a1, _tmp12b1);
-                float32x4_t _tmp30 = vaddq_f32(_tmp34a0, _tmp34b0);
-                float32x4_t _tmp31 = vaddq_f32(_tmp34a1, _tmp34b1);
-                float32x4_t _tmp40 = vsubq_f32(_tmp34a0, _tmp34b0);
-                float32x4_t _tmp41 = vsubq_f32(_tmp34a1, _tmp34b1);
-                float32x4_t _tmp50 = vaddq_f32(_tmp56a0, _tmp56b0);
-                float32x4_t _tmp51 = vaddq_f32(_tmp56a1, _tmp56b1);
-                float32x4_t _tmp60 = vsubq_f32(_tmp56a0, _tmp56b0);
-                float32x4_t _tmp61 = vsubq_f32(_tmp56a1, _tmp56b1);
-                float32x4_t _tmp70 = vfmaq_laneq_f32(vsubq_f32(_r70, _r10), vsubq_f32(_r30, _r50), _coeffs, 0);
-                float32x4_t _tmp71 = vfmaq_laneq_f32(vsubq_f32(_r71, _r11), vsubq_f32(_r31, _r51), _coeffs, 0);
-
-                vst1q_f32(tmp[0][m], _tmp00);
-                vst1q_f32(tmp[0][m] + 4, _tmp01);
-                vst1q_f32(tmp[1][m], _tmp10);
-                vst1q_f32(tmp[1][m] + 4, _tmp11);
-                vst1q_f32(tmp[2][m], _tmp20);
-                vst1q_f32(tmp[2][m] + 4, _tmp21);
-                vst1q_f32(tmp[3][m], _tmp30);
-                vst1q_f32(tmp[3][m] + 4, _tmp31);
-                vst1q_f32(tmp[4][m], _tmp40);
-                vst1q_f32(tmp[4][m] + 4, _tmp41);
-                vst1q_f32(tmp[5][m], _tmp50);
-                vst1q_f32(tmp[5][m] + 4, _tmp51);
-                vst1q_f32(tmp[6][m], _tmp60);
-                vst1q_f32(tmp[6][m] + 4, _tmp61);
-                vst1q_f32(tmp[7][m], _tmp70);
-                vst1q_f32(tmp[7][m] + 4, _tmp71);
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 64 + jj * 8;
-            float* p1 = p0 + max_jj * 8;
-            float* p2 = p0 + max_jj * 8 * 2;
-            float* p3 = p0 + max_jj * 8 * 3;
-            float* p4 = p0 + max_jj * 8 * 4;
-            float* p5 = p0 + max_jj * 8 * 5;
-            float* p6 = p0 + max_jj * 8 * 6;
-            float* p7 = p0 + max_jj * 8 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float32x4_t _r00 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r01 = vld1q_f32(tmp[m][0] + 4);
-                float32x4_t _r10 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r11 = vld1q_f32(tmp[m][1] + 4);
-                float32x4_t _r20 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r21 = vld1q_f32(tmp[m][2] + 4);
-                float32x4_t _r30 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r31 = vld1q_f32(tmp[m][3] + 4);
-                float32x4_t _r40 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r41 = vld1q_f32(tmp[m][4] + 4);
-                float32x4_t _r50 = vld1q_f32(tmp[m][5]);
-                float32x4_t _r51 = vld1q_f32(tmp[m][5] + 4);
-                float32x4_t _r60 = vld1q_f32(tmp[m][6]);
-                float32x4_t _r61 = vld1q_f32(tmp[m][6] + 4);
-                float32x4_t _r70 = vld1q_f32(tmp[m][7]);
-                float32x4_t _r71 = vld1q_f32(tmp[m][7] + 4);
-
-                float32x4_t _tmp12a0 = vfmaq_laneq_f32(vaddq_f32(_r20, _r60), _r40, _coeffs, 1);
-                float32x4_t _tmp12a1 = vfmaq_laneq_f32(vaddq_f32(_r21, _r61), _r41, _coeffs, 1);
-                float32x4_t _tmp12b0 = vfmaq_laneq_f32(vaddq_f32(_r10, _r50), _r30, _coeffs, 1);
-                float32x4_t _tmp12b1 = vfmaq_laneq_f32(vaddq_f32(_r11, _r51), _r31, _coeffs, 1);
-                float32x4_t _tmp34a0 = vfmaq_laneq_f32(vfmaq_laneq_f32(_r60, _r20, _coeffs, 3), _r40, _coeffs, 2);
-                float32x4_t _tmp34a1 = vfmaq_laneq_f32(vfmaq_laneq_f32(_r61, _r21, _coeffs, 3), _r41, _coeffs, 2);
-                float32x4_t _tmp34b0 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r10, _coeffs2, 1), _r30, _coeffs2, 0), _r50, _coeffs2, 2);
-                float32x4_t _tmp34b1 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r11, _coeffs2, 1), _r31, _coeffs2, 0), _r51, _coeffs2, 2);
-                float32x4_t _tmp56a0 = vfmaq_laneq_f32(_r60, vfmaq_laneq_f32(_r20, _r40, _coeffs, 2), _coeffs2, 3);
-                float32x4_t _tmp56a1 = vfmaq_laneq_f32(_r61, vfmaq_laneq_f32(_r21, _r41, _coeffs, 2), _coeffs2, 3);
-                float32x4_t _tmp56b0 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r10, _coeffs2, 2), _r30, _coeffs2, 0), _r50, _coeffs2, 1);
-                float32x4_t _tmp56b1 = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r11, _coeffs2, 2), _r31, _coeffs2, 0), _r51, _coeffs2, 1);
-
-                float32x4_t _tmp00 = vfmaq_laneq_f32(vsubq_f32(_r00, _r60), vsubq_f32(_r40, _r20), _coeffs, 0);
-                float32x4_t _tmp01 = vfmaq_laneq_f32(vsubq_f32(_r01, _r61), vsubq_f32(_r41, _r21), _coeffs, 0);
-                float32x4_t _tmp10 = vaddq_f32(_tmp12a0, _tmp12b0);
-                float32x4_t _tmp11 = vaddq_f32(_tmp12a1, _tmp12b1);
-                float32x4_t _tmp20 = vsubq_f32(_tmp12a0, _tmp12b0);
-                float32x4_t _tmp21 = vsubq_f32(_tmp12a1, _tmp12b1);
-                float32x4_t _tmp30 = vaddq_f32(_tmp34a0, _tmp34b0);
-                float32x4_t _tmp31 = vaddq_f32(_tmp34a1, _tmp34b1);
-                float32x4_t _tmp40 = vsubq_f32(_tmp34a0, _tmp34b0);
-                float32x4_t _tmp41 = vsubq_f32(_tmp34a1, _tmp34b1);
-                float32x4_t _tmp50 = vaddq_f32(_tmp56a0, _tmp56b0);
-                float32x4_t _tmp51 = vaddq_f32(_tmp56a1, _tmp56b1);
-                float32x4_t _tmp60 = vsubq_f32(_tmp56a0, _tmp56b0);
-                float32x4_t _tmp61 = vsubq_f32(_tmp56a1, _tmp56b1);
-                float32x4_t _tmp70 = vfmaq_laneq_f32(vsubq_f32(_r70, _r10), vsubq_f32(_r30, _r50), _coeffs, 0);
-                float32x4_t _tmp71 = vfmaq_laneq_f32(vsubq_f32(_r71, _r11), vsubq_f32(_r31, _r51), _coeffs, 0);
-
-                vst1q_f32(p0, _tmp00);
-                vst1q_f32(p0 + 4, _tmp01);
-                vst1q_f32(p1, _tmp10);
-                vst1q_f32(p1 + 4, _tmp11);
-                vst1q_f32(p2, _tmp20);
-                vst1q_f32(p2 + 4, _tmp21);
-                vst1q_f32(p3, _tmp30);
-                vst1q_f32(p3 + 4, _tmp31);
-                vst1q_f32(p4, _tmp40);
-                vst1q_f32(p4 + 4, _tmp41);
-                vst1q_f32(p5, _tmp50);
-                vst1q_f32(p5 + 4, _tmp51);
-                vst1q_f32(p6, _tmp60);
-                vst1q_f32(p6 + 4, _tmp61);
-                vst1q_f32(p7, _tmp70);
-                vst1q_f32(p7 + 4, _tmp71);
-
-                p0 += max_jj * 8 * 8;
-                p1 += max_jj * 8 * 8;
-                p2 += max_jj * 8 * 8;
-                p3 += max_jj * 8 * 8;
-                p4 += max_jj * 8 * 8;
-                p5 += max_jj * 8 * 8;
-                p6 += max_jj * 8 * 8;
-                p7 += max_jj * 8 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-#else // __aarch64__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-    #pragma omp parallel for num_threads(nT)
-#endif // __aarch64__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 4;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[8][8][4];
-
-        const float coeffs[8] = {5.25f, -4.25f, -1.25f, 0.25f, -2.5f, 0.5f, 2.f, 4.f};
-        float32x4_t _coeffs = vld1q_f32(coeffs);
-        float32x4_t _coeffs2 = vld1q_f32(coeffs + 4);
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const unsigned short* r0 = bottom_blob.channel((k + kk) / elempack).row<const unsigned short>(ti * 6) + (tj * 6) * elempack;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float32x4_t _r0 = vdupq_n_f32(0.f);
-                float32x4_t _r1 = vdupq_n_f32(0.f);
-                float32x4_t _r2 = vdupq_n_f32(0.f);
-                float32x4_t _r3 = vdupq_n_f32(0.f);
-                float32x4_t _r4 = vdupq_n_f32(0.f);
-                float32x4_t _r5 = vdupq_n_f32(0.f);
-                float32x4_t _r6 = vdupq_n_f32(0.f);
-                float32x4_t _r7 = vdupq_n_f32(0.f);
-
-                if (ti * 6 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        _r0 = bfloat2float(vld1_u16(r0));
-                        if (tj * 6 + 1 < w) _r1 = bfloat2float(vld1_u16(r0 + 4));
-                        if (tj * 6 + 2 < w) _r2 = bfloat2float(vld1_u16(r0 + 8));
-                        if (tj * 6 + 3 < w) _r3 = bfloat2float(vld1_u16(r0 + 12));
-                        if (tj * 6 + 4 < w) _r4 = bfloat2float(vld1_u16(r0 + 16));
-                        if (tj * 6 + 5 < w) _r5 = bfloat2float(vld1_u16(r0 + 20));
-                        if (tj * 6 + 6 < w) _r6 = bfloat2float(vld1_u16(r0 + 24));
-                        if (tj * 6 + 7 < w) _r7 = bfloat2float(vld1_u16(r0 + 28));
-                    }
-                    if (elempack == 1)
-                    {
-                        const unsigned short* r1 = r0 + N;
-                        const unsigned short* r2 = r0 + N * 2;
-                        const unsigned short* r3 = r0 + N * 3;
-
-                        uint16x4_t _t0 = vld1_u16(r0);
-                        uint16x4_t _t1 = vld1_u16(r1);
-                        uint16x4_t _t2 = vld1_u16(r2);
-                        uint16x4_t _t3 = vld1_u16(r3);
-
-                        transpose4x4_u16(_t0, _t1, _t2, _t3);
-
-                        _r0 = bfloat2float(_t0);
-                        if (tj * 6 + 1 < w) _r1 = bfloat2float(_t1);
-                        if (tj * 6 + 2 < w) _r2 = bfloat2float(_t2);
-                        if (tj * 6 + 3 < w) _r3 = bfloat2float(_t3);
-                        if (tj * 6 + 4 < w)
-                        {
-                            _t0 = vld1_u16(r0 + 4);
-                            _t1 = vld1_u16(r1 + 4);
-                            _t2 = vld1_u16(r2 + 4);
-                            _t3 = vld1_u16(r3 + 4);
-
-                            transpose4x4_u16(_t0, _t1, _t2, _t3);
-
-                            _r4 = bfloat2float(_t0);
-                            if (tj * 6 + 5 < w) _r5 = bfloat2float(_t1);
-                            if (tj * 6 + 6 < w) _r6 = bfloat2float(_t2);
-                            if (tj * 6 + 7 < w) _r7 = bfloat2float(_t3);
-                        }
-                    }
-                }
-
-#if __aarch64__
-                float32x4_t _tmp12a = vfmaq_laneq_f32(vaddq_f32(_r2, _r6), _r4, _coeffs, 1);
-                float32x4_t _tmp12b = vfmaq_laneq_f32(vaddq_f32(_r1, _r5), _r3, _coeffs, 1);
-                float32x4_t _tmp34a = vfmaq_laneq_f32(vfmaq_laneq_f32(_r6, _r2, _coeffs, 3), _r4, _coeffs, 2);
-                float32x4_t _tmp34b = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r1, _coeffs2, 1), _r3, _coeffs2, 0), _r5, _coeffs2, 2);
-                float32x4_t _tmp56a = vfmaq_laneq_f32(_r6, vfmaq_laneq_f32(_r2, _r4, _coeffs, 2), _coeffs2, 3);
-                float32x4_t _tmp56b = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r1, _coeffs2, 2), _r3, _coeffs2, 0), _r5, _coeffs2, 1);
-#else
-                float32x4_t _tmp12a = vmlaq_lane_f32(vaddq_f32(_r2, _r6), _r4, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp12b = vmlaq_lane_f32(vaddq_f32(_r1, _r5), _r3, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp34a = vmlaq_lane_f32(vmlaq_lane_f32(_r6, _r2, vget_high_f32(_coeffs), 1), _r4, vget_high_f32(_coeffs), 0);
-                float32x4_t _tmp34b = vmlaq_lane_f32(vmlaq_lane_f32(vmulq_lane_f32(_r1, vget_low_f32(_coeffs2), 1), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_high_f32(_coeffs2), 0);
-                float32x4_t _tmp56a = vmlaq_lane_f32(_r6, vmlaq_lane_f32(_r2, _r4, vget_high_f32(_coeffs), 0), vget_high_f32(_coeffs2), 1);
-                float32x4_t _tmp56b = vmlaq_lane_f32(vmlaq_lane_f32(vmulq_lane_f32(_r1, vget_high_f32(_coeffs2), 0), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_low_f32(_coeffs2), 1);
-#endif
-
-#if __aarch64__
-                float32x4_t _tmp0 = vfmaq_laneq_f32(vsubq_f32(_r0, _r6), vsubq_f32(_r4, _r2), _coeffs, 0);
-#else
-                float32x4_t _tmp0 = vmlaq_lane_f32(vsubq_f32(_r0, _r6), vsubq_f32(_r4, _r2), vget_low_f32(_coeffs), 0);
-#endif
-                float32x4_t _tmp1 = vaddq_f32(_tmp12a, _tmp12b);
-                float32x4_t _tmp2 = vsubq_f32(_tmp12a, _tmp12b);
-                float32x4_t _tmp3 = vaddq_f32(_tmp34a, _tmp34b);
-                float32x4_t _tmp4 = vsubq_f32(_tmp34a, _tmp34b);
-                float32x4_t _tmp5 = vaddq_f32(_tmp56a, _tmp56b);
-                float32x4_t _tmp6 = vsubq_f32(_tmp56a, _tmp56b);
-#if __aarch64__
-                float32x4_t _tmp7 = vfmaq_laneq_f32(vsubq_f32(_r7, _r1), vsubq_f32(_r3, _r5), _coeffs, 0);
-#else
-                float32x4_t _tmp7 = vmlaq_lane_f32(vsubq_f32(_r7, _r1), vsubq_f32(_r3, _r5), vget_low_f32(_coeffs), 0);
-#endif
-
-                vst1q_f32(tmp[0][m], _tmp0);
-                vst1q_f32(tmp[1][m], _tmp1);
-                vst1q_f32(tmp[2][m], _tmp2);
-                vst1q_f32(tmp[3][m], _tmp3);
-                vst1q_f32(tmp[4][m], _tmp4);
-                vst1q_f32(tmp[5][m], _tmp5);
-                vst1q_f32(tmp[6][m], _tmp6);
-                vst1q_f32(tmp[7][m], _tmp7);
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 64 + jj * 4;
-            float* p1 = p0 + max_jj * 4;
-            float* p2 = p0 + max_jj * 4 * 2;
-            float* p3 = p0 + max_jj * 4 * 3;
-            float* p4 = p0 + max_jj * 4 * 4;
-            float* p5 = p0 + max_jj * 4 * 5;
-            float* p6 = p0 + max_jj * 4 * 6;
-            float* p7 = p0 + max_jj * 4 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float32x4_t _r0 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r1 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r2 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r3 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r4 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r5 = vld1q_f32(tmp[m][5]);
-                float32x4_t _r6 = vld1q_f32(tmp[m][6]);
-                float32x4_t _r7 = vld1q_f32(tmp[m][7]);
-
-#if __aarch64__
-                float32x4_t _tmp12a = vfmaq_laneq_f32(vaddq_f32(_r2, _r6), _r4, _coeffs, 1);
-                float32x4_t _tmp12b = vfmaq_laneq_f32(vaddq_f32(_r1, _r5), _r3, _coeffs, 1);
-                float32x4_t _tmp34a = vfmaq_laneq_f32(vfmaq_laneq_f32(_r6, _r2, _coeffs, 3), _r4, _coeffs, 2);
-                float32x4_t _tmp34b = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r1, _coeffs2, 1), _r3, _coeffs2, 0), _r5, _coeffs2, 2);
-                float32x4_t _tmp56a = vfmaq_laneq_f32(_r6, vfmaq_laneq_f32(_r2, _r4, _coeffs, 2), _coeffs2, 3);
-                float32x4_t _tmp56b = vfmaq_laneq_f32(vfmaq_laneq_f32(vmulq_laneq_f32(_r1, _coeffs2, 2), _r3, _coeffs2, 0), _r5, _coeffs2, 1);
-#else
-                float32x4_t _tmp12a = vmlaq_lane_f32(vaddq_f32(_r2, _r6), _r4, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp12b = vmlaq_lane_f32(vaddq_f32(_r1, _r5), _r3, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp34a = vmlaq_lane_f32(vmlaq_lane_f32(_r6, _r2, vget_high_f32(_coeffs), 1), _r4, vget_high_f32(_coeffs), 0);
-                float32x4_t _tmp34b = vmlaq_lane_f32(vmlaq_lane_f32(vmulq_lane_f32(_r1, vget_low_f32(_coeffs2), 1), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_high_f32(_coeffs2), 0);
-                float32x4_t _tmp56a = vmlaq_lane_f32(_r6, vmlaq_lane_f32(_r2, _r4, vget_high_f32(_coeffs), 0), vget_high_f32(_coeffs2), 1);
-                float32x4_t _tmp56b = vmlaq_lane_f32(vmlaq_lane_f32(vmulq_lane_f32(_r1, vget_high_f32(_coeffs2), 0), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_low_f32(_coeffs2), 1);
-#endif
-
-#if __aarch64__
-                float32x4_t _tmp0 = vfmaq_laneq_f32(vsubq_f32(_r0, _r6), vsubq_f32(_r4, _r2), _coeffs, 0);
-#else
-                float32x4_t _tmp0 = vmlaq_lane_f32(vsubq_f32(_r0, _r6), vsubq_f32(_r4, _r2), vget_low_f32(_coeffs), 0);
-#endif
-                float32x4_t _tmp1 = vaddq_f32(_tmp12a, _tmp12b);
-                float32x4_t _tmp2 = vsubq_f32(_tmp12a, _tmp12b);
-                float32x4_t _tmp3 = vaddq_f32(_tmp34a, _tmp34b);
-                float32x4_t _tmp4 = vsubq_f32(_tmp34a, _tmp34b);
-                float32x4_t _tmp5 = vaddq_f32(_tmp56a, _tmp56b);
-                float32x4_t _tmp6 = vsubq_f32(_tmp56a, _tmp56b);
-#if __aarch64__
-                float32x4_t _tmp7 = vfmaq_laneq_f32(vsubq_f32(_r7, _r1), vsubq_f32(_r3, _r5), _coeffs, 0);
-#else
-                float32x4_t _tmp7 = vmlaq_lane_f32(vsubq_f32(_r7, _r1), vsubq_f32(_r3, _r5), vget_low_f32(_coeffs), 0);
-#endif
-
-                vst1q_f32(p0, _tmp0);
-                vst1q_f32(p1, _tmp1);
-                vst1q_f32(p2, _tmp2);
-                vst1q_f32(p3, _tmp3);
-                vst1q_f32(p4, _tmp4);
-                vst1q_f32(p5, _tmp5);
-                vst1q_f32(p6, _tmp6);
-                vst1q_f32(p7, _tmp7);
-
-                p0 += max_jj * 8 * 4;
-                p1 += max_jj * 8 * 4;
-                p2 += max_jj * 8 * 4;
-                p3 += max_jj * 8 * 4;
-                p4 += max_jj * 8 * 4;
-                p5 += max_jj * 8 * 4;
-                p6 += max_jj * 8 * 4;
-                p7 += max_jj * 8 * 4;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 4;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-#else // __ARM_NEON
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    #pragma omp parallel for num_threads(nT)
-#endif // __ARM_NEON
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        float tmp[8][8][2];
-
-#if __ARM_NEON
-        const float coeffs[8] = {5.25f, -4.25f, -1.25f, 0.25f, -2.5f, 0.5f, 2.f, 4.f};
-        float32x4_t _coeffs = vld1q_f32(coeffs);
-        float32x4_t _coeffs2 = vld1q_f32(coeffs + 4);
-#endif
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const unsigned short* r0 = bottom_blob.channel(k + kk).row<const unsigned short>(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 8; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vdup_n_f32(0.f);
-                float32x2_t _r1 = vdup_n_f32(0.f);
-                float32x2_t _r2 = vdup_n_f32(0.f);
-                float32x2_t _r3 = vdup_n_f32(0.f);
-                float32x2_t _r4 = vdup_n_f32(0.f);
-                float32x2_t _r5 = vdup_n_f32(0.f);
-                float32x2_t _r6 = vdup_n_f32(0.f);
-                float32x2_t _r7 = vdup_n_f32(0.f);
-#else
-                float r00 = 0.f;
-                float r01 = 0.f;
-                float r10 = 0.f;
-                float r11 = 0.f;
-                float r20 = 0.f;
-                float r21 = 0.f;
-                float r30 = 0.f;
-                float r31 = 0.f;
-                float r40 = 0.f;
-                float r41 = 0.f;
-                float r50 = 0.f;
-                float r51 = 0.f;
-                float r60 = 0.f;
-                float r61 = 0.f;
-                float r70 = 0.f;
-                float r71 = 0.f;
-#endif
-
-                if (ti * 6 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const unsigned short* r1 = r0 + N;
-
-#if __ARM_NEON
-                        uint16x4_t _t0 = vld1_u16(r0);
-                        uint16x4_t _t1 = vld1_u16(r1);
-                        uint16x4x2_t _t01 = vzip_u16(_t0, _t1);
-                        float32x4_t _t0_fp32 = bfloat2float(_t01.val[0]);
-                        float32x4_t _t1_fp32 = bfloat2float(_t01.val[1]);
-
-                        _r0 = vget_low_f32(_t0_fp32);
-                        if (tj * 6 + 1 < w) _r1 = vget_high_f32(_t0_fp32);
-                        if (tj * 6 + 2 < w) _r2 = vget_low_f32(_t1_fp32);
-                        if (tj * 6 + 3 < w) _r3 = vget_high_f32(_t1_fp32);
-                        if (tj * 6 + 4 < w)
-                        {
-                            _t0 = vld1_u16(r0 + 4);
-                            _t1 = vld1_u16(r1 + 4);
-                            _t01 = vzip_u16(_t0, _t1);
-                            _t0_fp32 = bfloat2float(_t01.val[0]);
-                            _t1_fp32 = bfloat2float(_t01.val[1]);
-
-                            _r4 = vget_low_f32(_t0_fp32);
-                            if (tj * 6 + 5 < w) _r5 = vget_high_f32(_t0_fp32);
-                            if (tj * 6 + 6 < w) _r6 = vget_low_f32(_t1_fp32);
-                            if (tj * 6 + 7 < w) _r7 = vget_high_f32(_t1_fp32);
-                        }
-#else
-                        r00 = bfloat16_to_float32(r0[0]);
-                        r01 = bfloat16_to_float32(r1[0]);
-                        if (tj * 6 + 1 < w)
-                        {
-                            r10 = bfloat16_to_float32(r0[1]);
-                            r11 = bfloat16_to_float32(r1[1]);
-                        }
-                        if (tj * 6 + 2 < w)
-                        {
-                            r20 = bfloat16_to_float32(r0[2]);
-                            r21 = bfloat16_to_float32(r1[2]);
-                        }
-                        if (tj * 6 + 3 < w)
-                        {
-                            r30 = bfloat16_to_float32(r0[3]);
-                            r31 = bfloat16_to_float32(r1[3]);
-                        }
-                        if (tj * 6 + 4 < w)
-                        {
-                            r40 = bfloat16_to_float32(r0[4]);
-                            r41 = bfloat16_to_float32(r1[4]);
-                        }
-                        if (tj * 6 + 5 < w)
-                        {
-                            r50 = bfloat16_to_float32(r0[5]);
-                            r51 = bfloat16_to_float32(r1[5]);
-                        }
-                        if (tj * 6 + 6 < w)
-                        {
-                            r60 = bfloat16_to_float32(r0[6]);
-                            r61 = bfloat16_to_float32(r1[6]);
-                        }
-                        if (tj * 6 + 7 < w)
-                        {
-                            r70 = bfloat16_to_float32(r0[7]);
-                            r71 = bfloat16_to_float32(r1[7]);
-                        }
-#endif
-                    }
-                }
-
-#if __ARM_NEON
-#if __aarch64__
-                float32x2_t _tmp12a = vfma_laneq_f32(vadd_f32(_r2, _r6), _r4, _coeffs, 1);
-                float32x2_t _tmp12b = vfma_laneq_f32(vadd_f32(_r1, _r5), _r3, _coeffs, 1);
-                float32x2_t _tmp34a = vfma_laneq_f32(vfma_laneq_f32(_r6, _r2, _coeffs, 3), _r4, _coeffs, 2);
-                float32x2_t _tmp34b = vfma_laneq_f32(vfma_laneq_f32(vmul_laneq_f32(_r1, _coeffs2, 1), _r3, _coeffs2, 0), _r5, _coeffs2, 2);
-                float32x2_t _tmp56a = vfma_laneq_f32(_r6, vfma_laneq_f32(_r2, _r4, _coeffs, 2), _coeffs2, 3);
-                float32x2_t _tmp56b = vfma_laneq_f32(vfma_laneq_f32(vmul_laneq_f32(_r1, _coeffs2, 2), _r3, _coeffs2, 0), _r5, _coeffs2, 1);
-#else
-                float32x2_t _tmp12a = vmla_lane_f32(vadd_f32(_r2, _r6), _r4, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp12b = vmla_lane_f32(vadd_f32(_r1, _r5), _r3, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp34a = vmla_lane_f32(vmla_lane_f32(_r6, _r2, vget_high_f32(_coeffs), 1), _r4, vget_high_f32(_coeffs), 0);
-                float32x2_t _tmp34b = vmla_lane_f32(vmla_lane_f32(vmul_lane_f32(_r1, vget_low_f32(_coeffs2), 1), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_high_f32(_coeffs2), 0);
-                float32x2_t _tmp56a = vmla_lane_f32(_r6, vmla_lane_f32(_r2, _r4, vget_high_f32(_coeffs), 0), vget_high_f32(_coeffs2), 1);
-                float32x2_t _tmp56b = vmla_lane_f32(vmla_lane_f32(vmul_lane_f32(_r1, vget_high_f32(_coeffs2), 0), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_low_f32(_coeffs2), 1);
-#endif
-
-#if __aarch64__
-                float32x2_t _tmp0 = vfma_laneq_f32(vsub_f32(_r0, _r6), vsub_f32(_r4, _r2), _coeffs, 0);
-#else
-                float32x2_t _tmp0 = vmla_lane_f32(vsub_f32(_r0, _r6), vsub_f32(_r4, _r2), vget_low_f32(_coeffs), 0);
-#endif
-                float32x2_t _tmp1 = vadd_f32(_tmp12a, _tmp12b);
-                float32x2_t _tmp2 = vsub_f32(_tmp12a, _tmp12b);
-                float32x2_t _tmp3 = vadd_f32(_tmp34a, _tmp34b);
-                float32x2_t _tmp4 = vsub_f32(_tmp34a, _tmp34b);
-                float32x2_t _tmp5 = vadd_f32(_tmp56a, _tmp56b);
-                float32x2_t _tmp6 = vsub_f32(_tmp56a, _tmp56b);
-#if __aarch64__
-                float32x2_t _tmp7 = vfma_laneq_f32(vsub_f32(_r7, _r1), vsub_f32(_r3, _r5), _coeffs, 0);
-#else
-                float32x2_t _tmp7 = vmla_lane_f32(vsub_f32(_r7, _r1), vsub_f32(_r3, _r5), vget_low_f32(_coeffs), 0);
-#endif
-
-                vst1_f32(tmp[0][m], _tmp0);
-                vst1_f32(tmp[1][m], _tmp1);
-                vst1_f32(tmp[2][m], _tmp2);
-                vst1_f32(tmp[3][m], _tmp3);
-                vst1_f32(tmp[4][m], _tmp4);
-                vst1_f32(tmp[5][m], _tmp5);
-                vst1_f32(tmp[6][m], _tmp6);
-                vst1_f32(tmp[7][m], _tmp7);
-#else
-                float tmp12a0 = r20 + r60 - r40 * 4.25f;
-                float tmp12a1 = r21 + r61 - r41 * 4.25f;
-                float tmp12b0 = r10 + r50 - r30 * 4.25f;
-                float tmp12b1 = r11 + r51 - r31 * 4.25f;
-                float tmp34a0 = r60 + r20 * 0.25f - r40 * 1.25f;
-                float tmp34a1 = r61 + r21 * 0.25f - r41 * 1.25f;
-                float tmp34b0 = r10 * 0.5f - r30 * 2.5f + r50 * 2.f;
-                float tmp34b1 = r11 * 0.5f - r31 * 2.5f + r51 * 2.f;
-                float tmp56a0 = r20 * 4.f - r40 * 5.f + r60;
-                float tmp56a1 = r21 * 4.f - r41 * 5.f + r61;
-                float tmp56b0 = r10 * 2.f - r30 * 2.5f + r50 * 0.5f;
-                float tmp56b1 = r11 * 2.f - r31 * 2.5f + r51 * 0.5f;
-
-                tmp[0][m][0] = r00 - r60 + (r40 - r20) * 5.25f;
-                tmp[0][m][1] = r01 - r61 + (r41 - r21) * 5.25f;
-                tmp[1][m][0] = tmp12a0 + tmp12b0;
-                tmp[1][m][1] = tmp12a1 + tmp12b1;
-                tmp[2][m][0] = tmp12a0 - tmp12b0;
-                tmp[2][m][1] = tmp12a1 - tmp12b1;
-                tmp[3][m][0] = tmp34a0 + tmp34b0;
-                tmp[3][m][1] = tmp34a1 + tmp34b1;
-                tmp[4][m][0] = tmp34a0 - tmp34b0;
-                tmp[4][m][1] = tmp34a1 - tmp34b1;
-                tmp[5][m][0] = tmp56a0 + tmp56b0;
-                tmp[5][m][1] = tmp56a1 + tmp56b1;
-                tmp[6][m][0] = tmp56a0 - tmp56b0;
-                tmp[6][m][1] = tmp56a1 - tmp56b1;
-                tmp[7][m][0] = r70 - r10 + (r30 - r50) * 5.25f;
-                tmp[7][m][1] = r71 - r11 + (r31 - r51) * 5.25f;
-#endif
-
-                r0 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 64 + jj * 2;
-            float* p1 = p0 + max_jj * 2;
-            float* p2 = p0 + max_jj * 2 * 2;
-            float* p3 = p0 + max_jj * 2 * 3;
-            float* p4 = p0 + max_jj * 2 * 4;
-            float* p5 = p0 + max_jj * 2 * 5;
-            float* p6 = p0 + max_jj * 2 * 6;
-            float* p7 = p0 + max_jj * 2 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(tmp[m][0]);
-                float32x2_t _r1 = vld1_f32(tmp[m][1]);
-                float32x2_t _r2 = vld1_f32(tmp[m][2]);
-                float32x2_t _r3 = vld1_f32(tmp[m][3]);
-                float32x2_t _r4 = vld1_f32(tmp[m][4]);
-                float32x2_t _r5 = vld1_f32(tmp[m][5]);
-                float32x2_t _r6 = vld1_f32(tmp[m][6]);
-                float32x2_t _r7 = vld1_f32(tmp[m][7]);
-
-#if __aarch64__
-                float32x2_t _tmp12a = vfma_laneq_f32(vadd_f32(_r2, _r6), _r4, _coeffs, 1);
-                float32x2_t _tmp12b = vfma_laneq_f32(vadd_f32(_r1, _r5), _r3, _coeffs, 1);
-                float32x2_t _tmp34a = vfma_laneq_f32(vfma_laneq_f32(_r6, _r2, _coeffs, 3), _r4, _coeffs, 2);
-                float32x2_t _tmp34b = vfma_laneq_f32(vfma_laneq_f32(vmul_laneq_f32(_r1, _coeffs2, 1), _r3, _coeffs2, 0), _r5, _coeffs2, 2);
-                float32x2_t _tmp56a = vfma_laneq_f32(_r6, vfma_laneq_f32(_r2, _r4, _coeffs, 2), _coeffs2, 3);
-                float32x2_t _tmp56b = vfma_laneq_f32(vfma_laneq_f32(vmul_laneq_f32(_r1, _coeffs2, 2), _r3, _coeffs2, 0), _r5, _coeffs2, 1);
-#else
-                float32x2_t _tmp12a = vmla_lane_f32(vadd_f32(_r2, _r6), _r4, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp12b = vmla_lane_f32(vadd_f32(_r1, _r5), _r3, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp34a = vmla_lane_f32(vmla_lane_f32(_r6, _r2, vget_high_f32(_coeffs), 1), _r4, vget_high_f32(_coeffs), 0);
-                float32x2_t _tmp34b = vmla_lane_f32(vmla_lane_f32(vmul_lane_f32(_r1, vget_low_f32(_coeffs2), 1), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_high_f32(_coeffs2), 0);
-                float32x2_t _tmp56a = vmla_lane_f32(_r6, vmla_lane_f32(_r2, _r4, vget_high_f32(_coeffs), 0), vget_high_f32(_coeffs2), 1);
-                float32x2_t _tmp56b = vmla_lane_f32(vmla_lane_f32(vmul_lane_f32(_r1, vget_high_f32(_coeffs2), 0), _r3, vget_low_f32(_coeffs2), 0), _r5, vget_low_f32(_coeffs2), 1);
-#endif
-
-#if __aarch64__
-                float32x2_t _tmp0 = vfma_laneq_f32(vsub_f32(_r0, _r6), vsub_f32(_r4, _r2), _coeffs, 0);
-#else
-                float32x2_t _tmp0 = vmla_lane_f32(vsub_f32(_r0, _r6), vsub_f32(_r4, _r2), vget_low_f32(_coeffs), 0);
-#endif
-                float32x2_t _tmp1 = vadd_f32(_tmp12a, _tmp12b);
-                float32x2_t _tmp2 = vsub_f32(_tmp12a, _tmp12b);
-                float32x2_t _tmp3 = vadd_f32(_tmp34a, _tmp34b);
-                float32x2_t _tmp4 = vsub_f32(_tmp34a, _tmp34b);
-                float32x2_t _tmp5 = vadd_f32(_tmp56a, _tmp56b);
-                float32x2_t _tmp6 = vsub_f32(_tmp56a, _tmp56b);
-#if __aarch64__
-                float32x2_t _tmp7 = vfma_laneq_f32(vsub_f32(_r7, _r1), vsub_f32(_r3, _r5), _coeffs, 0);
-#else
-                float32x2_t _tmp7 = vmla_lane_f32(vsub_f32(_r7, _r1), vsub_f32(_r3, _r5), vget_low_f32(_coeffs), 0);
-#endif
-
-                vst1_f32(p0, _tmp0);
-                vst1_f32(p1, _tmp1);
-                vst1_f32(p2, _tmp2);
-                vst1_f32(p3, _tmp3);
-                vst1_f32(p4, _tmp4);
-                vst1_f32(p5, _tmp5);
-                vst1_f32(p6, _tmp6);
-                vst1_f32(p7, _tmp7);
-#else
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-                float r40 = tmp[m][4][0];
-                float r41 = tmp[m][4][1];
-                float r50 = tmp[m][5][0];
-                float r51 = tmp[m][5][1];
-                float r60 = tmp[m][6][0];
-                float r61 = tmp[m][6][1];
-                float r70 = tmp[m][7][0];
-                float r71 = tmp[m][7][1];
-
-                float tmp12a0 = r20 + r60 - r40 * 4.25f;
-                float tmp12a1 = r21 + r61 - r41 * 4.25f;
-                float tmp12b0 = r10 + r50 - r30 * 4.25f;
-                float tmp12b1 = r11 + r51 - r31 * 4.25f;
-                float tmp34a0 = r60 + r20 * 0.25f - r40 * 1.25f;
-                float tmp34a1 = r61 + r21 * 0.25f - r41 * 1.25f;
-                float tmp34b0 = r10 * 0.5f - r30 * 2.5f + r50 * 2.f;
-                float tmp34b1 = r11 * 0.5f - r31 * 2.5f + r51 * 2.f;
-                float tmp56a0 = r20 * 4.f - r40 * 5.f + r60;
-                float tmp56a1 = r21 * 4.f - r41 * 5.f + r61;
-                float tmp56b0 = r10 * 2.f - r30 * 2.5f + r50 * 0.5f;
-                float tmp56b1 = r11 * 2.f - r31 * 2.5f + r51 * 0.5f;
-
-                p0[0] = r00 - r60 + (r40 - r20) * 5.25f;
-                p0[1] = r01 - r61 + (r41 - r21) * 5.25f;
-                p1[0] = tmp12a0 + tmp12b0;
-                p1[1] = tmp12a1 + tmp12b1;
-                p2[0] = tmp12a0 - tmp12b0;
-                p2[1] = tmp12a1 - tmp12b1;
-                p3[0] = tmp34a0 + tmp34b0;
-                p3[1] = tmp34a1 + tmp34b1;
-                p4[0] = tmp34a0 - tmp34b0;
-                p4[1] = tmp34a1 - tmp34b1;
-                p5[0] = tmp56a0 + tmp56b0;
-                p5[1] = tmp56a1 + tmp56b1;
-                p6[0] = tmp56a0 - tmp56b0;
-                p6[1] = tmp56a1 - tmp56b1;
-                p7[0] = r70 - r10 + (r30 - r50) * 5.25f;
-                p7[1] = r71 - r11 + (r31 - r51) * 5.25f;
-#endif
-
-                p0 += max_jj * 8 * 2;
-                p1 += max_jj * 8 * 2;
-                p2 += max_jj * 8 * 2;
-                p3 += max_jj * 8 * 2;
-                p4 += max_jj * 8 * 2;
-                p5 += max_jj * 8 * 2;
-                p6 += max_jj * 8 * 2;
-                p7 += max_jj * 8 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        float tmp[8][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const unsigned short* r0123 = bottom_blob.channel(k + kk).row<const unsigned short>(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 8; m++)
-            {
-                float r0 = 0.f;
-                float r1 = 0.f;
-                float r2 = 0.f;
-                float r3 = 0.f;
-                float r4 = 0.f;
-                float r5 = 0.f;
-                float r6 = 0.f;
-                float r7 = 0.f;
-
-                if (ti * 6 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = bfloat16_to_float32(r0123[0]);
-                        if (tj * 6 + 1 < w) r1 = bfloat16_to_float32(r0123[1]);
-                        if (tj * 6 + 2 < w) r2 = bfloat16_to_float32(r0123[2]);
-                        if (tj * 6 + 3 < w) r3 = bfloat16_to_float32(r0123[3]);
-                        if (tj * 6 + 4 < w) r4 = bfloat16_to_float32(r0123[4]);
-                        if (tj * 6 + 5 < w) r5 = bfloat16_to_float32(r0123[5]);
-                        if (tj * 6 + 6 < w) r6 = bfloat16_to_float32(r0123[6]);
-                        if (tj * 6 + 7 < w) r7 = bfloat16_to_float32(r0123[7]);
-                    }
-                }
-
-                float tmp12a = r2 + r6 - r4 * 4.25f;
-                float tmp12b = r1 + r5 - r3 * 4.25f;
-                float tmp34a = r6 + r2 * 0.25f - r4 * 1.25f;
-                float tmp34b = r1 * 0.5f - r3 * 2.5f + r5 * 2.f;
-                float tmp56a = r2 * 4.f - r4 * 5.f + r6;
-                float tmp56b = r1 * 2.f - r3 * 2.5f + r5 * 0.5f;
-
-                tmp[0][m] = r0 - r6 + (r4 - r2) * 5.25f;
-                tmp[1][m] = tmp12a + tmp12b;
-                tmp[2][m] = tmp12a - tmp12b;
-                tmp[3][m] = tmp34a + tmp34b;
-                tmp[4][m] = tmp34a - tmp34b;
-                tmp[5][m] = tmp56a + tmp56b;
-                tmp[6][m] = tmp56a - tmp56b;
-                tmp[7][m] = r7 - r1 + (r3 - r5) * 5.25f;
-
-                r0123 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 64 + jj;
-            float* p1 = p0 + max_jj;
-            float* p2 = p0 + max_jj * 2;
-            float* p3 = p0 + max_jj * 3;
-            float* p4 = p0 + max_jj * 4;
-            float* p5 = p0 + max_jj * 5;
-            float* p6 = p0 + max_jj * 6;
-            float* p7 = p0 + max_jj * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-                float r4 = tmp[m][4];
-                float r5 = tmp[m][5];
-                float r6 = tmp[m][6];
-                float r7 = tmp[m][7];
-
-                float tmp12a = r2 + r6 - r4 * 4.25f;
-                float tmp12b = r1 + r5 - r3 * 4.25f;
-                float tmp34a = r6 + r2 * 0.25f - r4 * 1.25f;
-                float tmp34b = r1 * 0.5f - r3 * 2.5f + r5 * 2.f;
-                float tmp56a = r2 * 4.f - r4 * 5.f + r6;
-                float tmp56b = r1 * 2.f - r3 * 2.5f + r5 * 0.5f;
-
-                p0[0] = r0 - r6 + (r4 - r2) * 5.25f;
-                p1[0] = tmp12a + tmp12b;
-                p2[0] = tmp12a - tmp12b;
-                p3[0] = tmp34a + tmp34b;
-                p4[0] = tmp34a - tmp34b;
-                p5[0] = tmp56a + tmp56b;
-                p6[0] = tmp56a - tmp56b;
-                p7[0] = r7 - r1 + (r3 - r5) * 5.25f;
-
-                p0 += max_jj * 8;
-                p1 += max_jj * 8;
-                p2 += max_jj * 8;
-                p3 += max_jj * 8;
-                p4 += max_jj * 8;
-                p5 += max_jj * 8;
-                p6 += max_jj * 8;
-                p7 += max_jj * 8;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd63_transform_output_tile_bf16s(const Mat& top_tile, Mat& top_blob, const Mat& bias, int i, int max_ii, int j, int max_jj)
-{
-    // const float otm[6][8] = {
-    //     {1.0f, 1.0f,  1.0f,  1.0f,  1.0f, 32.0f, 32.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f,  2.0f, -2.0f, 16.0f,-16.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f,  4.0f,  4.0f,  8.0f,  8.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f,  8.0f, -8.0f,  4.0f, -4.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f, 16.0f, 16.0f,  2.0f,  2.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 32.0f,-32.0f,  1.0f, -1.0f, 1.0f}
-    // };
-
-#if __ARM_NEON
-    const float coeffs[4] = {32.f, 16.f, 8.f, 4.f};
-    float32x4_t _coeffs = vld1q_f32(coeffs);
-    float32x2_t _v2 = vdup_n_f32(2.f);
-#endif
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 5) / 6;
-
-    const float* biasptr = bias;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        float32x4_t _bias0 = biasptr ? vld1q_f32(biasptr + i + ii) : vdupq_n_f32(0.f);
-        float32x4_t _bias1 = biasptr ? vld1q_f32(biasptr + i + ii + 4) : vdupq_n_f32(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[6][8][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 64 + jj * 8;
-            const float* r1 = r0 + max_jj * 8;
-            const float* r2 = r0 + max_jj * 8 * 2;
-            const float* r3 = r0 + max_jj * 8 * 3;
-            const float* r4 = r0 + max_jj * 8 * 4;
-            const float* r5 = r0 + max_jj * 8 * 5;
-            const float* r6 = r0 + max_jj * 8 * 6;
-            const float* r7 = r0 + max_jj * 8 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r30 = vld1q_f32(r3);
-                float32x4_t _r31 = vld1q_f32(r3 + 4);
-                float32x4_t _r40 = vld1q_f32(r4);
-                float32x4_t _r41 = vld1q_f32(r4 + 4);
-                float32x4_t _r50 = vld1q_f32(r5);
-                float32x4_t _r51 = vld1q_f32(r5 + 4);
-                float32x4_t _r60 = vld1q_f32(r6);
-                float32x4_t _r61 = vld1q_f32(r6 + 4);
-                float32x4_t _r70 = vld1q_f32(r7);
-                float32x4_t _r71 = vld1q_f32(r7 + 4);
-
-                float32x4_t _tmp024a0 = vaddq_f32(_r10, _r20);
-                float32x4_t _tmp024a1 = vaddq_f32(_r11, _r21);
-                float32x4_t _tmp135a0 = vsubq_f32(_r10, _r20);
-                float32x4_t _tmp135a1 = vsubq_f32(_r11, _r21);
-                float32x4_t _tmp024b0 = vaddq_f32(_r30, _r40);
-                float32x4_t _tmp024b1 = vaddq_f32(_r31, _r41);
-                float32x4_t _tmp135b0 = vsubq_f32(_r30, _r40);
-                float32x4_t _tmp135b1 = vsubq_f32(_r31, _r41);
-                float32x4_t _tmp024c0 = vaddq_f32(_r50, _r60);
-                float32x4_t _tmp024c1 = vaddq_f32(_r51, _r61);
-                float32x4_t _tmp135c0 = vsubq_f32(_r50, _r60);
-                float32x4_t _tmp135c1 = vsubq_f32(_r51, _r61);
-
-                float32x4_t _tmp00 = vaddq_f32(vaddq_f32(_r00, _tmp024a0), vfmaq_laneq_f32(_tmp024b0, _tmp024c0, _coeffs, 0));
-                float32x4_t _tmp01 = vaddq_f32(vaddq_f32(_r01, _tmp024a1), vfmaq_laneq_f32(_tmp024b1, _tmp024c1, _coeffs, 0));
-                float32x4_t _tmp10 = vfmaq_laneq_f32(vfmaq_lane_f32(_tmp135a0, _tmp135b0, _v2, 0), _tmp135c0, _coeffs, 1);
-                float32x4_t _tmp11 = vfmaq_laneq_f32(vfmaq_lane_f32(_tmp135a1, _tmp135b1, _v2, 0), _tmp135c1, _coeffs, 1);
-                float32x4_t _tmp20 = vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp024a0, _tmp024b0, _coeffs, 3), _tmp024c0, _coeffs, 2);
-                float32x4_t _tmp21 = vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp024a1, _tmp024b1, _coeffs, 3), _tmp024c1, _coeffs, 2);
-                float32x4_t _tmp30 = vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp135a0, _tmp135b0, _coeffs, 2), _tmp135c0, _coeffs, 3);
-                float32x4_t _tmp31 = vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp135a1, _tmp135b1, _coeffs, 2), _tmp135c1, _coeffs, 3);
-                float32x4_t _tmp40 = vfmaq_lane_f32(vfmaq_laneq_f32(_tmp024a0, _tmp024b0, _coeffs, 1), _tmp024c0, _v2, 0);
-                float32x4_t _tmp41 = vfmaq_lane_f32(vfmaq_laneq_f32(_tmp024a1, _tmp024b1, _coeffs, 1), _tmp024c1, _v2, 0);
-                float32x4_t _tmp50 = vaddq_f32(vaddq_f32(_r70, _tmp135a0), vfmaq_laneq_f32(_tmp135c0, _tmp135b0, _coeffs, 0));
-                float32x4_t _tmp51 = vaddq_f32(vaddq_f32(_r71, _tmp135a1), vfmaq_laneq_f32(_tmp135c1, _tmp135b1, _coeffs, 0));
-
-                vst1q_f32(tmp[0][m], _tmp00);
-                vst1q_f32(tmp[0][m] + 4, _tmp01);
-                vst1q_f32(tmp[1][m], _tmp10);
-                vst1q_f32(tmp[1][m] + 4, _tmp11);
-                vst1q_f32(tmp[2][m], _tmp20);
-                vst1q_f32(tmp[2][m] + 4, _tmp21);
-                vst1q_f32(tmp[3][m], _tmp30);
-                vst1q_f32(tmp[3][m] + 4, _tmp31);
-                vst1q_f32(tmp[4][m], _tmp40);
-                vst1q_f32(tmp[4][m] + 4, _tmp41);
-                vst1q_f32(tmp[5][m], _tmp50);
-                vst1q_f32(tmp[5][m] + 4, _tmp51);
-
-                r0 += max_jj * 8 * 8;
-                r1 += max_jj * 8 * 8;
-                r2 += max_jj * 8 * 8;
-                r3 += max_jj * 8 * 8;
-                r4 += max_jj * 8 * 8;
-                r5 += max_jj * 8 * 8;
-                r6 += max_jj * 8 * 8;
-                r7 += max_jj * 8 * 8;
-            }
-
-            unsigned short* outptr0 = top_blob.channel((i + ii) / out_elempack).row<unsigned short>(ti * 6) + (tj * 6) * out_elempack;
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-                float32x4_t _r00 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r01 = vld1q_f32(tmp[m][0] + 4);
-                float32x4_t _r10 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r11 = vld1q_f32(tmp[m][1] + 4);
-                float32x4_t _r20 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r21 = vld1q_f32(tmp[m][2] + 4);
-                float32x4_t _r30 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r31 = vld1q_f32(tmp[m][3] + 4);
-                float32x4_t _r40 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r41 = vld1q_f32(tmp[m][4] + 4);
-                float32x4_t _r50 = vld1q_f32(tmp[m][5]);
-                float32x4_t _r51 = vld1q_f32(tmp[m][5] + 4);
-                float32x4_t _r60 = vld1q_f32(tmp[m][6]);
-                float32x4_t _r61 = vld1q_f32(tmp[m][6] + 4);
-                float32x4_t _r70 = vld1q_f32(tmp[m][7]);
-                float32x4_t _r71 = vld1q_f32(tmp[m][7] + 4);
-
-                float32x4_t _tmp024a0 = vaddq_f32(_r10, _r20);
-                float32x4_t _tmp024a1 = vaddq_f32(_r11, _r21);
-                float32x4_t _tmp135a0 = vsubq_f32(_r10, _r20);
-                float32x4_t _tmp135a1 = vsubq_f32(_r11, _r21);
-                float32x4_t _tmp024b0 = vaddq_f32(_r30, _r40);
-                float32x4_t _tmp024b1 = vaddq_f32(_r31, _r41);
-                float32x4_t _tmp135b0 = vsubq_f32(_r30, _r40);
-                float32x4_t _tmp135b1 = vsubq_f32(_r31, _r41);
-                float32x4_t _tmp024c0 = vaddq_f32(_r50, _r60);
-                float32x4_t _tmp024c1 = vaddq_f32(_r51, _r61);
-                float32x4_t _tmp135c0 = vsubq_f32(_r50, _r60);
-                float32x4_t _tmp135c1 = vsubq_f32(_r51, _r61);
-
-                float32x4_t _tmp00 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r00, _tmp024a0), vfmaq_laneq_f32(_tmp024b0, _tmp024c0, _coeffs, 0)));
-                float32x4_t _tmp01 = vaddq_f32(_bias1, vaddq_f32(vaddq_f32(_r01, _tmp024a1), vfmaq_laneq_f32(_tmp024b1, _tmp024c1, _coeffs, 0)));
-                float32x4_t _tmp10 = vaddq_f32(_bias0, vfmaq_laneq_f32(vfmaq_lane_f32(_tmp135a0, _tmp135b0, _v2, 0), _tmp135c0, _coeffs, 1));
-                float32x4_t _tmp11 = vaddq_f32(_bias1, vfmaq_laneq_f32(vfmaq_lane_f32(_tmp135a1, _tmp135b1, _v2, 0), _tmp135c1, _coeffs, 1));
-                float32x4_t _tmp20 = vaddq_f32(_bias0, vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp024a0, _tmp024b0, _coeffs, 3), _tmp024c0, _coeffs, 2));
-                float32x4_t _tmp21 = vaddq_f32(_bias1, vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp024a1, _tmp024b1, _coeffs, 3), _tmp024c1, _coeffs, 2));
-                float32x4_t _tmp30 = vaddq_f32(_bias0, vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp135a0, _tmp135b0, _coeffs, 2), _tmp135c0, _coeffs, 3));
-                float32x4_t _tmp31 = vaddq_f32(_bias1, vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp135a1, _tmp135b1, _coeffs, 2), _tmp135c1, _coeffs, 3));
-                float32x4_t _tmp40 = vaddq_f32(_bias0, vfmaq_lane_f32(vfmaq_laneq_f32(_tmp024a0, _tmp024b0, _coeffs, 1), _tmp024c0, _v2, 0));
-                float32x4_t _tmp41 = vaddq_f32(_bias1, vfmaq_lane_f32(vfmaq_laneq_f32(_tmp024a1, _tmp024b1, _coeffs, 1), _tmp024c1, _v2, 0));
-                float32x4_t _tmp50 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r70, _tmp135a0), vfmaq_laneq_f32(_tmp135c0, _tmp135b0, _coeffs, 0)));
-                float32x4_t _tmp51 = vaddq_f32(_bias1, vaddq_f32(vaddq_f32(_r71, _tmp135a1), vfmaq_laneq_f32(_tmp135c1, _tmp135b1, _coeffs, 0)));
-
-                if (out_elempack == 4)
-                {
-                    unsigned short* outptr1 = outptr0 + N;
-
-                    vst1_u16(outptr0, float2bfloat(_tmp00));
-                    vst1_u16(outptr1, float2bfloat(_tmp01));
-                    if (tj * 6 + 1 < outw)
-                    {
-                        vst1_u16(outptr0 + 4, float2bfloat(_tmp10));
-                        vst1_u16(outptr1 + 4, float2bfloat(_tmp11));
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        vst1_u16(outptr0 + 8, float2bfloat(_tmp20));
-                        vst1_u16(outptr1 + 8, float2bfloat(_tmp21));
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        vst1_u16(outptr0 + 12, float2bfloat(_tmp30));
-                        vst1_u16(outptr1 + 12, float2bfloat(_tmp31));
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        vst1_u16(outptr0 + 16, float2bfloat(_tmp40));
-                        vst1_u16(outptr1 + 16, float2bfloat(_tmp41));
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        vst1_u16(outptr0 + 20, float2bfloat(_tmp50));
-                        vst1_u16(outptr1 + 20, float2bfloat(_tmp51));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short tmp0[8];
-                    unsigned short tmp1[8];
-                    unsigned short tmp2[8];
-                    unsigned short tmp3[8];
-                    unsigned short tmp4[8];
-                    unsigned short tmp5[8];
-                    vst1_u16(tmp0, float2bfloat(_tmp00));
-                    vst1_u16(tmp0 + 4, float2bfloat(_tmp01));
-                    vst1_u16(tmp1, float2bfloat(_tmp10));
-                    vst1_u16(tmp1 + 4, float2bfloat(_tmp11));
-                    vst1_u16(tmp2, float2bfloat(_tmp20));
-                    vst1_u16(tmp2 + 4, float2bfloat(_tmp21));
-                    vst1_u16(tmp3, float2bfloat(_tmp30));
-                    vst1_u16(tmp3 + 4, float2bfloat(_tmp31));
-                    vst1_u16(tmp4, float2bfloat(_tmp40));
-                    vst1_u16(tmp4 + 4, float2bfloat(_tmp41));
-                    vst1_u16(tmp5, float2bfloat(_tmp50));
-                    vst1_u16(tmp5 + 4, float2bfloat(_tmp51));
-
-                    unsigned short* outptr1 = outptr0 + N;
-                    unsigned short* outptr2 = outptr0 + N * 2;
-                    unsigned short* outptr3 = outptr0 + N * 3;
-                    unsigned short* outptr4 = outptr0 + N * 4;
-                    unsigned short* outptr5 = outptr0 + N * 5;
-                    unsigned short* outptr6 = outptr0 + N * 6;
-                    unsigned short* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-                    if (tj * 6 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                        outptr4[2] = tmp2[4];
-                        outptr5[2] = tmp2[5];
-                        outptr6[2] = tmp2[6];
-                        outptr7[2] = tmp2[7];
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                        outptr4[3] = tmp3[4];
-                        outptr5[3] = tmp3[5];
-                        outptr6[3] = tmp3[6];
-                        outptr7[3] = tmp3[7];
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        outptr0[4] = tmp4[0];
-                        outptr1[4] = tmp4[1];
-                        outptr2[4] = tmp4[2];
-                        outptr3[4] = tmp4[3];
-                        outptr4[4] = tmp4[4];
-                        outptr5[4] = tmp4[5];
-                        outptr6[4] = tmp4[6];
-                        outptr7[4] = tmp4[7];
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        outptr0[5] = tmp5[0];
-                        outptr1[5] = tmp5[1];
-                        outptr2[5] = tmp5[2];
-                        outptr3[5] = tmp5[3];
-                        outptr4[5] = tmp5[4];
-                        outptr5[5] = tmp5[5];
-                        outptr6[5] = tmp5[6];
-                        outptr7[5] = tmp5[7];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        float32x4_t _bias0 = biasptr ? vld1q_f32(biasptr + i + ii) : vdupq_n_f32(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[6][8][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 64 + jj * 4;
-            const float* r1 = r0 + max_jj * 4;
-            const float* r2 = r0 + max_jj * 4 * 2;
-            const float* r3 = r0 + max_jj * 4 * 3;
-            const float* r4 = r0 + max_jj * 4 * 4;
-            const float* r5 = r0 + max_jj * 4 * 5;
-            const float* r6 = r0 + max_jj * 4 * 6;
-            const float* r7 = r0 + max_jj * 4 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float32x4_t _r0 = vld1q_f32(r0);
-                float32x4_t _r1 = vld1q_f32(r1);
-                float32x4_t _r2 = vld1q_f32(r2);
-                float32x4_t _r3 = vld1q_f32(r3);
-                float32x4_t _r4 = vld1q_f32(r4);
-                float32x4_t _r5 = vld1q_f32(r5);
-                float32x4_t _r6 = vld1q_f32(r6);
-                float32x4_t _r7 = vld1q_f32(r7);
-
-                float32x4_t _tmp024a = vaddq_f32(_r1, _r2);
-                float32x4_t _tmp135a = vsubq_f32(_r1, _r2);
-                float32x4_t _tmp024b = vaddq_f32(_r3, _r4);
-                float32x4_t _tmp135b = vsubq_f32(_r3, _r4);
-                float32x4_t _tmp024c = vaddq_f32(_r5, _r6);
-                float32x4_t _tmp135c = vsubq_f32(_r5, _r6);
-
-#if __aarch64__
-                float32x4_t _tmp0 = vaddq_f32(vaddq_f32(_r0, _tmp024a), vfmaq_laneq_f32(_tmp024b, _tmp024c, _coeffs, 0));
-                float32x4_t _tmp1 = vfmaq_laneq_f32(vfmaq_lane_f32(_tmp135a, _tmp135b, _v2, 0), _tmp135c, _coeffs, 1);
-                float32x4_t _tmp2 = vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp024a, _tmp024b, _coeffs, 3), _tmp024c, _coeffs, 2);
-                float32x4_t _tmp3 = vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp135a, _tmp135b, _coeffs, 2), _tmp135c, _coeffs, 3);
-                float32x4_t _tmp4 = vfmaq_lane_f32(vfmaq_laneq_f32(_tmp024a, _tmp024b, _coeffs, 1), _tmp024c, _v2, 0);
-                float32x4_t _tmp5 = vaddq_f32(vaddq_f32(_r7, _tmp135a), vfmaq_laneq_f32(_tmp135c, _tmp135b, _coeffs, 0));
-#else
-                float32x4_t _tmp0 = vaddq_f32(vaddq_f32(_r0, _tmp024a), vmlaq_lane_f32(_tmp024b, _tmp024c, vget_low_f32(_coeffs), 0));
-                float32x4_t _tmp1 = vmlaq_lane_f32(vmlaq_lane_f32(_tmp135a, _tmp135b, _v2, 0), _tmp135c, vget_low_f32(_coeffs), 1);
-                float32x4_t _tmp2 = vmlaq_lane_f32(vmlaq_lane_f32(_tmp024a, _tmp024b, vget_high_f32(_coeffs), 1), _tmp024c, vget_high_f32(_coeffs), 0);
-                float32x4_t _tmp3 = vmlaq_lane_f32(vmlaq_lane_f32(_tmp135a, _tmp135b, vget_high_f32(_coeffs), 0), _tmp135c, vget_high_f32(_coeffs), 1);
-                float32x4_t _tmp4 = vmlaq_lane_f32(vmlaq_lane_f32(_tmp024a, _tmp024b, vget_low_f32(_coeffs), 1), _tmp024c, _v2, 0);
-                float32x4_t _tmp5 = vaddq_f32(vaddq_f32(_r7, _tmp135a), vmlaq_lane_f32(_tmp135c, _tmp135b, vget_low_f32(_coeffs), 0));
-#endif
-
-                vst1q_f32(tmp[0][m], _tmp0);
-                vst1q_f32(tmp[1][m], _tmp1);
-                vst1q_f32(tmp[2][m], _tmp2);
-                vst1q_f32(tmp[3][m], _tmp3);
-                vst1q_f32(tmp[4][m], _tmp4);
-                vst1q_f32(tmp[5][m], _tmp5);
-
-                r0 += max_jj * 8 * 4;
-                r1 += max_jj * 8 * 4;
-                r2 += max_jj * 8 * 4;
-                r3 += max_jj * 8 * 4;
-                r4 += max_jj * 8 * 4;
-                r5 += max_jj * 8 * 4;
-                r6 += max_jj * 8 * 4;
-                r7 += max_jj * 8 * 4;
-            }
-
-            unsigned short* outptr0 = top_blob.channel((i + ii) / out_elempack).row<unsigned short>(ti * 6) + (tj * 6) * out_elempack;
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-                float32x4_t _r0 = vld1q_f32(tmp[m][0]);
-                float32x4_t _r1 = vld1q_f32(tmp[m][1]);
-                float32x4_t _r2 = vld1q_f32(tmp[m][2]);
-                float32x4_t _r3 = vld1q_f32(tmp[m][3]);
-                float32x4_t _r4 = vld1q_f32(tmp[m][4]);
-                float32x4_t _r5 = vld1q_f32(tmp[m][5]);
-                float32x4_t _r6 = vld1q_f32(tmp[m][6]);
-                float32x4_t _r7 = vld1q_f32(tmp[m][7]);
-
-                float32x4_t _tmp024a = vaddq_f32(_r1, _r2);
-                float32x4_t _tmp135a = vsubq_f32(_r1, _r2);
-                float32x4_t _tmp024b = vaddq_f32(_r3, _r4);
-                float32x4_t _tmp135b = vsubq_f32(_r3, _r4);
-                float32x4_t _tmp024c = vaddq_f32(_r5, _r6);
-                float32x4_t _tmp135c = vsubq_f32(_r5, _r6);
-
-#if __aarch64__
-                float32x4_t _tmp0 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r0, _tmp024a), vfmaq_laneq_f32(_tmp024b, _tmp024c, _coeffs, 0)));
-                float32x4_t _tmp1 = vaddq_f32(_bias0, vfmaq_laneq_f32(vfmaq_lane_f32(_tmp135a, _tmp135b, _v2, 0), _tmp135c, _coeffs, 1));
-                float32x4_t _tmp2 = vaddq_f32(_bias0, vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp024a, _tmp024b, _coeffs, 3), _tmp024c, _coeffs, 2));
-                float32x4_t _tmp3 = vaddq_f32(_bias0, vfmaq_laneq_f32(vfmaq_laneq_f32(_tmp135a, _tmp135b, _coeffs, 2), _tmp135c, _coeffs, 3));
-                float32x4_t _tmp4 = vaddq_f32(_bias0, vfmaq_lane_f32(vfmaq_laneq_f32(_tmp024a, _tmp024b, _coeffs, 1), _tmp024c, _v2, 0));
-                float32x4_t _tmp5 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r7, _tmp135a), vfmaq_laneq_f32(_tmp135c, _tmp135b, _coeffs, 0)));
-#else
-                float32x4_t _tmp0 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r0, _tmp024a), vmlaq_lane_f32(_tmp024b, _tmp024c, vget_low_f32(_coeffs), 0)));
-                float32x4_t _tmp1 = vaddq_f32(_bias0, vmlaq_lane_f32(vmlaq_lane_f32(_tmp135a, _tmp135b, _v2, 0), _tmp135c, vget_low_f32(_coeffs), 1));
-                float32x4_t _tmp2 = vaddq_f32(_bias0, vmlaq_lane_f32(vmlaq_lane_f32(_tmp024a, _tmp024b, vget_high_f32(_coeffs), 1), _tmp024c, vget_high_f32(_coeffs), 0));
-                float32x4_t _tmp3 = vaddq_f32(_bias0, vmlaq_lane_f32(vmlaq_lane_f32(_tmp135a, _tmp135b, vget_high_f32(_coeffs), 0), _tmp135c, vget_high_f32(_coeffs), 1));
-                float32x4_t _tmp4 = vaddq_f32(_bias0, vmlaq_lane_f32(vmlaq_lane_f32(_tmp024a, _tmp024b, vget_low_f32(_coeffs), 1), _tmp024c, _v2, 0));
-                float32x4_t _tmp5 = vaddq_f32(_bias0, vaddq_f32(vaddq_f32(_r7, _tmp135a), vmlaq_lane_f32(_tmp135c, _tmp135b, vget_low_f32(_coeffs), 0)));
-#endif
-
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_tmp0));
-                    if (tj * 6 + 1 < outw) vst1_u16(outptr0 + 4, float2bfloat(_tmp1));
-                    if (tj * 6 + 2 < outw) vst1_u16(outptr0 + 8, float2bfloat(_tmp2));
-                    if (tj * 6 + 3 < outw) vst1_u16(outptr0 + 12, float2bfloat(_tmp3));
-                    if (tj * 6 + 4 < outw) vst1_u16(outptr0 + 16, float2bfloat(_tmp4));
-                    if (tj * 6 + 5 < outw) vst1_u16(outptr0 + 20, float2bfloat(_tmp5));
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short tmp0[4];
-                    unsigned short tmp1[4];
-                    unsigned short tmp2[4];
-                    unsigned short tmp3[4];
-                    unsigned short tmp4[4];
-                    unsigned short tmp5[4];
-                    vst1_u16(tmp0, float2bfloat(_tmp0));
-                    vst1_u16(tmp1, float2bfloat(_tmp1));
-                    vst1_u16(tmp2, float2bfloat(_tmp2));
-                    vst1_u16(tmp3, float2bfloat(_tmp3));
-                    vst1_u16(tmp4, float2bfloat(_tmp4));
-                    vst1_u16(tmp5, float2bfloat(_tmp5));
-
-                    unsigned short* outptr1 = outptr0 + N;
-                    unsigned short* outptr2 = outptr0 + N * 2;
-                    unsigned short* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    if (tj * 6 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        outptr0[4] = tmp4[0];
-                        outptr1[4] = tmp4[1];
-                        outptr2[4] = tmp4[2];
-                        outptr3[4] = tmp4[3];
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        outptr0[5] = tmp5[0];
-                        outptr1[5] = tmp5[1];
-                        outptr2[5] = tmp5[2];
-                        outptr3[5] = tmp5[3];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-#if __ARM_NEON
-        float32x2_t _bias0 = biasptr ? vld1_f32(biasptr + i + ii) : vdup_n_f32(0.f);
-#else
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-        float bias1 = biasptr ? biasptr[i + ii + 1] : 0.f;
-#endif
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        float tmp[6][8][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 64 + jj * 2;
-            const float* r1 = r0 + max_jj * 2;
-            const float* r2 = r0 + max_jj * 2 * 2;
-            const float* r3 = r0 + max_jj * 2 * 3;
-            const float* r4 = r0 + max_jj * 2 * 4;
-            const float* r5 = r0 + max_jj * 2 * 5;
-            const float* r6 = r0 + max_jj * 2 * 6;
-            const float* r7 = r0 + max_jj * 2 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(r0);
-                float32x2_t _r1 = vld1_f32(r1);
-                float32x2_t _r2 = vld1_f32(r2);
-                float32x2_t _r3 = vld1_f32(r3);
-                float32x2_t _r4 = vld1_f32(r4);
-                float32x2_t _r5 = vld1_f32(r5);
-                float32x2_t _r6 = vld1_f32(r6);
-                float32x2_t _r7 = vld1_f32(r7);
-
-                float32x2_t _tmp024a = vadd_f32(_r1, _r2);
-                float32x2_t _tmp135a = vsub_f32(_r1, _r2);
-                float32x2_t _tmp024b = vadd_f32(_r3, _r4);
-                float32x2_t _tmp135b = vsub_f32(_r3, _r4);
-                float32x2_t _tmp024c = vadd_f32(_r5, _r6);
-                float32x2_t _tmp135c = vsub_f32(_r5, _r6);
-
-#if __aarch64__
-                float32x2_t _tmp0 = vadd_f32(vadd_f32(_r0, _tmp024a), vfma_laneq_f32(_tmp024b, _tmp024c, _coeffs, 0));
-                float32x2_t _tmp1 = vfma_laneq_f32(vfma_f32(_tmp135a, _tmp135b, _v2), _tmp135c, _coeffs, 1);
-                float32x2_t _tmp2 = vfma_laneq_f32(vfma_laneq_f32(_tmp024a, _tmp024b, _coeffs, 3), _tmp024c, _coeffs, 2);
-                float32x2_t _tmp3 = vfma_laneq_f32(vfma_laneq_f32(_tmp135a, _tmp135b, _coeffs, 2), _tmp135c, _coeffs, 3);
-                float32x2_t _tmp4 = vfma_f32(vfma_laneq_f32(_tmp024a, _tmp024b, _coeffs, 1), _tmp024c, _v2);
-                float32x2_t _tmp5 = vadd_f32(vadd_f32(_r7, _tmp135a), vfma_laneq_f32(_tmp135c, _tmp135b, _coeffs, 0));
-#else
-                float32x2_t _tmp0 = vadd_f32(vadd_f32(_r0, _tmp024a), vmla_lane_f32(_tmp024b, _tmp024c, vget_low_f32(_coeffs), 0));
-                float32x2_t _tmp1 = vmla_lane_f32(vmla_f32(_tmp135a, _tmp135b, _v2), _tmp135c, vget_low_f32(_coeffs), 1);
-                float32x2_t _tmp2 = vmla_lane_f32(vmla_lane_f32(_tmp024a, _tmp024b, vget_high_f32(_coeffs), 1), _tmp024c, vget_high_f32(_coeffs), 0);
-                float32x2_t _tmp3 = vmla_lane_f32(vmla_lane_f32(_tmp135a, _tmp135b, vget_high_f32(_coeffs), 0), _tmp135c, vget_high_f32(_coeffs), 1);
-                float32x2_t _tmp4 = vmla_f32(vmla_lane_f32(_tmp024a, _tmp024b, vget_low_f32(_coeffs), 1), _tmp024c, _v2);
-                float32x2_t _tmp5 = vadd_f32(vadd_f32(_r7, _tmp135a), vmla_lane_f32(_tmp135c, _tmp135b, vget_low_f32(_coeffs), 0));
-#endif
-
-                vst1_f32(tmp[0][m], _tmp0);
-                vst1_f32(tmp[1][m], _tmp1);
-                vst1_f32(tmp[2][m], _tmp2);
-                vst1_f32(tmp[3][m], _tmp3);
-                vst1_f32(tmp[4][m], _tmp4);
-                vst1_f32(tmp[5][m], _tmp5);
-#else
-                float tmp024a0 = r1[0] + r2[0];
-                float tmp024a1 = r1[1] + r2[1];
-                float tmp135a0 = r1[0] - r2[0];
-                float tmp135a1 = r1[1] - r2[1];
-                float tmp024b0 = r3[0] + r4[0];
-                float tmp024b1 = r3[1] + r4[1];
-                float tmp135b0 = r3[0] - r4[0];
-                float tmp135b1 = r3[1] - r4[1];
-                float tmp024c0 = r5[0] + r6[0];
-                float tmp024c1 = r5[1] + r6[1];
-                float tmp135c0 = r5[0] - r6[0];
-                float tmp135c1 = r5[1] - r6[1];
-
-                tmp[0][m][0] = r0[0] + tmp024a0 + tmp024b0 + tmp024c0 * 32;
-                tmp[0][m][1] = r0[1] + tmp024a1 + tmp024b1 + tmp024c1 * 32;
-                tmp[1][m][0] = tmp135a0 + tmp135b0 + tmp135b0 + tmp135c0 * 16;
-                tmp[1][m][1] = tmp135a1 + tmp135b1 + tmp135b1 + tmp135c1 * 16;
-                tmp[2][m][0] = tmp024a0 + tmp024b0 * 4 + tmp024c0 * 8;
-                tmp[2][m][1] = tmp024a1 + tmp024b1 * 4 + tmp024c1 * 8;
-                tmp[3][m][0] = tmp135a0 + tmp135b0 * 8 + tmp135c0 * 4;
-                tmp[3][m][1] = tmp135a1 + tmp135b1 * 8 + tmp135c1 * 4;
-                tmp[4][m][0] = tmp024a0 + tmp024b0 * 16 + tmp024c0 + tmp024c0;
-                tmp[4][m][1] = tmp024a1 + tmp024b1 * 16 + tmp024c1 + tmp024c1;
-                tmp[5][m][0] = r7[0] + tmp135a0 + tmp135b0 * 32 + tmp135c0;
-                tmp[5][m][1] = r7[1] + tmp135a1 + tmp135b1 * 32 + tmp135c1;
-#endif
-
-                r0 += max_jj * 8 * 2;
-                r1 += max_jj * 8 * 2;
-                r2 += max_jj * 8 * 2;
-                r3 += max_jj * 8 * 2;
-                r4 += max_jj * 8 * 2;
-                r5 += max_jj * 8 * 2;
-                r6 += max_jj * 8 * 2;
-                r7 += max_jj * 8 * 2;
-            }
-
-            unsigned short* outptr0 = top_blob.channel(i + ii).row<unsigned short>(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-#if __ARM_NEON
-                float32x2_t _r0 = vld1_f32(tmp[m][0]);
-                float32x2_t _r1 = vld1_f32(tmp[m][1]);
-                float32x2_t _r2 = vld1_f32(tmp[m][2]);
-                float32x2_t _r3 = vld1_f32(tmp[m][3]);
-                float32x2_t _r4 = vld1_f32(tmp[m][4]);
-                float32x2_t _r5 = vld1_f32(tmp[m][5]);
-                float32x2_t _r6 = vld1_f32(tmp[m][6]);
-                float32x2_t _r7 = vld1_f32(tmp[m][7]);
-
-                float32x2_t _tmp024a = vadd_f32(_r1, _r2);
-                float32x2_t _tmp135a = vsub_f32(_r1, _r2);
-                float32x2_t _tmp024b = vadd_f32(_r3, _r4);
-                float32x2_t _tmp135b = vsub_f32(_r3, _r4);
-                float32x2_t _tmp024c = vadd_f32(_r5, _r6);
-                float32x2_t _tmp135c = vsub_f32(_r5, _r6);
-
-#if __aarch64__
-                float32x2_t _tmp0 = vadd_f32(_bias0, vadd_f32(vadd_f32(_r0, _tmp024a), vfma_laneq_f32(_tmp024b, _tmp024c, _coeffs, 0)));
-                float32x2_t _tmp1 = vadd_f32(_bias0, vfma_laneq_f32(vfma_f32(_tmp135a, _tmp135b, _v2), _tmp135c, _coeffs, 1));
-                float32x2_t _tmp2 = vadd_f32(_bias0, vfma_laneq_f32(vfma_laneq_f32(_tmp024a, _tmp024b, _coeffs, 3), _tmp024c, _coeffs, 2));
-                float32x2_t _tmp3 = vadd_f32(_bias0, vfma_laneq_f32(vfma_laneq_f32(_tmp135a, _tmp135b, _coeffs, 2), _tmp135c, _coeffs, 3));
-                float32x2_t _tmp4 = vadd_f32(_bias0, vfma_f32(vfma_laneq_f32(_tmp024a, _tmp024b, _coeffs, 1), _tmp024c, _v2));
-                float32x2_t _tmp5 = vadd_f32(_bias0, vadd_f32(vadd_f32(_r7, _tmp135a), vfma_laneq_f32(_tmp135c, _tmp135b, _coeffs, 0)));
-#else
-                float32x2_t _tmp0 = vadd_f32(_bias0, vadd_f32(vadd_f32(_r0, _tmp024a), vmla_lane_f32(_tmp024b, _tmp024c, vget_low_f32(_coeffs), 0)));
-                float32x2_t _tmp1 = vadd_f32(_bias0, vmla_lane_f32(vmla_f32(_tmp135a, _tmp135b, _v2), _tmp135c, vget_low_f32(_coeffs), 1));
-                float32x2_t _tmp2 = vadd_f32(_bias0, vmla_lane_f32(vmla_lane_f32(_tmp024a, _tmp024b, vget_high_f32(_coeffs), 1), _tmp024c, vget_high_f32(_coeffs), 0));
-                float32x2_t _tmp3 = vadd_f32(_bias0, vmla_lane_f32(vmla_lane_f32(_tmp135a, _tmp135b, vget_high_f32(_coeffs), 0), _tmp135c, vget_high_f32(_coeffs), 1));
-                float32x2_t _tmp4 = vadd_f32(_bias0, vmla_f32(vmla_lane_f32(_tmp024a, _tmp024b, vget_low_f32(_coeffs), 1), _tmp024c, _v2));
-                float32x2_t _tmp5 = vadd_f32(_bias0, vadd_f32(vadd_f32(_r7, _tmp135a), vmla_lane_f32(_tmp135c, _tmp135b, vget_low_f32(_coeffs), 0)));
-#endif
-#else
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-                float r40 = tmp[m][4][0];
-                float r41 = tmp[m][4][1];
-                float r50 = tmp[m][5][0];
-                float r51 = tmp[m][5][1];
-                float r60 = tmp[m][6][0];
-                float r61 = tmp[m][6][1];
-                float r70 = tmp[m][7][0];
-                float r71 = tmp[m][7][1];
-
-                float tmp024a0 = r10 + r20;
-                float tmp024a1 = r11 + r21;
-                float tmp135a0 = r10 - r20;
-                float tmp135a1 = r11 - r21;
-                float tmp024b0 = r30 + r40;
-                float tmp024b1 = r31 + r41;
-                float tmp135b0 = r30 - r40;
-                float tmp135b1 = r31 - r41;
-                float tmp024c0 = r50 + r60;
-                float tmp024c1 = r51 + r61;
-                float tmp135c0 = r50 - r60;
-                float tmp135c1 = r51 - r61;
-
-                float tmp00 = bias0 + r00 + tmp024a0 + tmp024b0 + tmp024c0 * 32;
-                float tmp01 = bias1 + r01 + tmp024a1 + tmp024b1 + tmp024c1 * 32;
-                float tmp10 = bias0 + tmp135a0 + tmp135b0 + tmp135b0 + tmp135c0 * 16;
-                float tmp11 = bias1 + tmp135a1 + tmp135b1 + tmp135b1 + tmp135c1 * 16;
-                float tmp20 = bias0 + tmp024a0 + tmp024b0 * 4 + tmp024c0 * 8;
-                float tmp21 = bias1 + tmp024a1 + tmp024b1 * 4 + tmp024c1 * 8;
-                float tmp30 = bias0 + tmp135a0 + tmp135b0 * 8 + tmp135c0 * 4;
-                float tmp31 = bias1 + tmp135a1 + tmp135b1 * 8 + tmp135c1 * 4;
-                float tmp40 = bias0 + tmp024a0 + tmp024b0 * 16 + tmp024c0 + tmp024c0;
-                float tmp41 = bias1 + tmp024a1 + tmp024b1 * 16 + tmp024c1 + tmp024c1;
-                float tmp50 = bias0 + r70 + tmp135a0 + tmp135b0 * 32 + tmp135c0;
-                float tmp51 = bias1 + r71 + tmp135a1 + tmp135b1 * 32 + tmp135c1;
-#endif
-
-                // if (out_elempack == 1)
-                {
-                    unsigned short* outptr1 = outptr0 + N;
-
-#if __ARM_NEON
-                    uint16x4_t _tmp01 = float2bfloat(vcombine_f32(_tmp0, _tmp1));
-                    uint16x4_t _tmp23 = float2bfloat(vcombine_f32(_tmp2, _tmp3));
-                    uint16x4_t _tmp45 = float2bfloat(vcombine_f32(_tmp4, _tmp5));
-
-                    outptr0[0] = vget_lane_u16(_tmp01, 0);
-                    outptr1[0] = vget_lane_u16(_tmp01, 1);
-                    if (tj * 6 + 1 < outw)
-                    {
-                        outptr0[1] = vget_lane_u16(_tmp01, 2);
-                        outptr1[1] = vget_lane_u16(_tmp01, 3);
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        outptr0[2] = vget_lane_u16(_tmp23, 0);
-                        outptr1[2] = vget_lane_u16(_tmp23, 1);
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        outptr0[3] = vget_lane_u16(_tmp23, 2);
-                        outptr1[3] = vget_lane_u16(_tmp23, 3);
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        outptr0[4] = vget_lane_u16(_tmp45, 0);
-                        outptr1[4] = vget_lane_u16(_tmp45, 1);
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        outptr0[5] = vget_lane_u16(_tmp45, 2);
-                        outptr1[5] = vget_lane_u16(_tmp45, 3);
-                    }
-#else
-                    outptr0[0] = float32_to_bfloat16(tmp00);
-                    outptr1[0] = float32_to_bfloat16(tmp01);
-                    if (tj * 6 + 1 < outw)
-                    {
-                        outptr0[1] = float32_to_bfloat16(tmp10);
-                        outptr1[1] = float32_to_bfloat16(tmp11);
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        outptr0[2] = float32_to_bfloat16(tmp20);
-                        outptr1[2] = float32_to_bfloat16(tmp21);
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        outptr0[3] = float32_to_bfloat16(tmp30);
-                        outptr1[3] = float32_to_bfloat16(tmp31);
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        outptr0[4] = float32_to_bfloat16(tmp40);
-                        outptr1[4] = float32_to_bfloat16(tmp41);
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        outptr0[5] = float32_to_bfloat16(tmp50);
-                        outptr1[5] = float32_to_bfloat16(tmp51);
-                    }
-#endif
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-
-        float tmp[6][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 64 + jj;
-            const float* r1 = r0 + max_jj;
-            const float* r2 = r0 + max_jj * 2;
-            const float* r3 = r0 + max_jj * 3;
-            const float* r4 = r0 + max_jj * 4;
-            const float* r5 = r0 + max_jj * 5;
-            const float* r6 = r0 + max_jj * 6;
-            const float* r7 = r0 + max_jj * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float tmp024a = r1[0] + r2[0];
-                float tmp135a = r1[0] - r2[0];
-                float tmp024b = r3[0] + r4[0];
-                float tmp135b = r3[0] - r4[0];
-                float tmp024c = r5[0] + r6[0];
-                float tmp135c = r5[0] - r6[0];
-
-                tmp[0][m] = r0[0] + tmp024a + tmp024b + tmp024c * 32;
-                tmp[1][m] = tmp135a + tmp135b + tmp135b + tmp135c * 16;
-                tmp[2][m] = tmp024a + tmp024b * 4 + tmp024c * 8;
-                tmp[3][m] = tmp135a + tmp135b * 8 + tmp135c * 4;
-                tmp[4][m] = tmp024a + tmp024b * 16 + tmp024c + tmp024c;
-                tmp[5][m] = r7[0] + tmp135a + tmp135b * 32 + tmp135c;
-
-                r0 += max_jj * 8;
-                r1 += max_jj * 8;
-                r2 += max_jj * 8;
-                r3 += max_jj * 8;
-                r4 += max_jj * 8;
-                r5 += max_jj * 8;
-                r6 += max_jj * 8;
-                r7 += max_jj * 8;
-            }
-
-            unsigned short* outptr0 = top_blob.channel(i + ii).row<unsigned short>(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-                float r4 = tmp[m][4];
-                float r5 = tmp[m][5];
-                float r6 = tmp[m][6];
-                float r7 = tmp[m][7];
-
-                float tmp024a = r1 + r2;
-                float tmp135a = r1 - r2;
-                float tmp024b = r3 + r4;
-                float tmp135b = r3 - r4;
-                float tmp024c = r5 + r6;
-                float tmp135c = r5 - r6;
-
-                float tmp0 = bias0 + r0 + tmp024a + tmp024b + tmp024c * 32;
-                float tmp1 = bias0 + tmp135a + tmp135b + tmp135b + tmp135c * 16;
-                float tmp2 = bias0 + tmp024a + tmp024b * 4 + tmp024c * 8;
-                float tmp3 = bias0 + tmp135a + tmp135b * 8 + tmp135c * 4;
-                float tmp4 = bias0 + tmp024a + tmp024b * 16 + tmp024c + tmp024c;
-                float tmp5 = bias0 + r7 + tmp135a + tmp135b * 32 + tmp135c;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = float32_to_bfloat16(tmp0);
-                    if (tj * 6 + 1 < outw) outptr0[1] = float32_to_bfloat16(tmp1);
-                    if (tj * 6 + 2 < outw) outptr0[2] = float32_to_bfloat16(tmp2);
-                    if (tj * 6 + 3 < outw) outptr0[3] = float32_to_bfloat16(tmp3);
-                    if (tj * 6 + 4 < outw) outptr0[4] = float32_to_bfloat16(tmp4);
-                    if (tj * 6 + 5 < outw) outptr0[5] = float32_to_bfloat16(tmp5);
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd63_bf16s(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, const Mat& bias, int nT, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 6n+2, winograd F(6,3)
-    int w_tiles = (outw + 5) / 6;
-    int h_tiles = (outh + 5) / 6;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 64;
-
-    // NCNN_LOGE("conv3x3s1_winograd63_bf16s %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    conv3x3s1_winograd_get_optimal_tile_mnk(M, N, K, B, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 4u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 4u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd63_transform_input_tile_bf16s(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 4u, opt.workspace_allocator);
-
-        #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd63_transform_input_tile_bf16s(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                conv3x3s1_winograd_gemm_transB_packed_tile(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk, opt.use_a53_a55_optimized_kernel);
-            }
-
-            // transform output
-            conv3x3s1_winograd63_transform_output_tile_bf16s(top_tile, top_blob, bias, i, max_ii, j, max_jj);
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd_fp16s.h
deleted file mode 100644
index 7332d61a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd_fp16s.h
+++ /dev/null
@@ -1,5986 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd_pack_A_tile_fp16(const Mat& A, Mat& AT, int batch, int max_ii, int max_kk)
-{
-    const int N = max_kk * batch;
-
-    for (int b = 0; b < batch; b++)
-    {
-        unsigned short* pp = AT.row<unsigned short>(b);
-
-        int ii = 0;
-        for (; ii + 7 < max_ii; ii += 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                pp[2] = p0[2 * N];
-                pp[3] = p0[3 * N];
-                pp[4] = p0[4 * N];
-                pp[5] = p0[5 * N];
-                pp[6] = p0[6 * N];
-                pp[7] = p0[7 * N];
-                p0 += batch;
-                pp += 8;
-            }
-        }
-        for (; ii + 3 < max_ii; ii += 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                pp[2] = p0[2 * N];
-                pp[3] = p0[3 * N];
-                p0 += batch;
-                pp += 4;
-            }
-        }
-        for (; ii + 1 < max_ii; ii += 2)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                p0 += batch;
-                pp += 2;
-            }
-        }
-        for (; ii < max_ii; ii++)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                p0 += batch;
-                pp += 1;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd_transpose_pack_B_tile_fp16(const Mat& B, Mat& BT, int batch, int max_jj, int max_kk, int nT)
-{
-    #pragma omp parallel for num_threads(nT)
-    for (int b = 0; b < batch; b++)
-    {
-        unsigned short* pp = BT.row<unsigned short>(b);
-
-        int jj = 0;
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            const unsigned short* p0 = B;
-
-            int kk = 0;
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                // transpose 8x12
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v4.8h, v5.8h, v6.8h, v7.8h}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v8.8h, v9.8h, v10.8h, v11.8h}, [%0] \n"
-
-                    "uzp1   v12.8h, v0.8h, v4.8h        \n"
-                    "uzp2   v16.8h, v0.8h, v4.8h        \n"
-                    "uzp1   v13.8h, v1.8h, v5.8h        \n"
-                    "uzp2   v17.8h, v1.8h, v5.8h        \n"
-                    "uzp1   v14.8h, v2.8h, v6.8h        \n"
-                    "uzp2   v18.8h, v2.8h, v6.8h        \n"
-                    "uzp1   v15.8h, v3.8h, v7.8h        \n"
-                    "uzp2   v19.8h, v3.8h, v7.8h        \n"
-                    "uzp1   v20.8h, v8.8h, v9.8h        \n"
-                    "uzp2   v22.8h, v8.8h, v9.8h        \n"
-                    "uzp1   v21.8h, v10.8h, v11.8h      \n"
-                    "uzp2   v23.8h, v10.8h, v11.8h      \n"
-
-                    "sub    %0, %0, #128                \n"
-
-                    "ext    v24.16b, v20.16b, v20.16b, #8 \n"
-                    "ext    v26.16b, v22.16b, v22.16b, #8 \n"
-                    "ext    v25.16b, v21.16b, v21.16b, #8 \n"
-                    "ext    v27.16b, v23.16b, v23.16b, #8 \n"
-
-                    "st1    {v12.8h}, [%1], #16         \n"
-                    "st1    {v20.4h}, [%1], #8          \n"
-                    "st1    {v13.8h}, [%1], #16         \n"
-                    "st1    {v24.4h}, [%1], #8          \n"
-                    "st1    {v14.8h}, [%1], #16         \n"
-                    "st1    {v21.4h}, [%1], #8          \n"
-                    "st1    {v15.8h}, [%1], #16         \n"
-                    "st1    {v25.4h}, [%1], #8          \n"
-                    "st1    {v16.8h}, [%1], #16         \n"
-                    "st1    {v22.4h}, [%1], #8          \n"
-                    "st1    {v17.8h}, [%1], #16         \n"
-                    "st1    {v26.4h}, [%1], #8          \n"
-                    "st1    {v18.8h}, [%1], #16         \n"
-                    "st1    {v23.4h}, [%1], #8          \n"
-                    "st1    {v19.8h}, [%1], #16         \n"
-                    "st1    {v27.4h}, [%1], #8          \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-                p0 += max_jj * batch * 8;
-#else  // NCNN_GNU_INLINE_ASM
-                uint16x8x4_t _r0 = vld4q_u16(p0);
-                uint16x8x4_t _r1 = vld4q_u16(p0 + 32);
-                uint16x8x4_t _r2 = vld4q_u16(p0 + 64);
-                uint16x8x2_t _r04lm = vuzpq_u16(_r0.val[0], _r1.val[0]);
-                uint16x8x2_t _r15lm = vuzpq_u16(_r0.val[1], _r1.val[1]);
-                uint16x8x2_t _r26lm = vuzpq_u16(_r0.val[2], _r1.val[2]);
-                uint16x8x2_t _r37lm = vuzpq_u16(_r0.val[3], _r1.val[3]);
-                uint16x8x2_t _r0145h = vuzpq_u16(_r2.val[0], _r2.val[1]);
-                uint16x8x2_t _r2367h = vuzpq_u16(_r2.val[2], _r2.val[3]);
-                vst1q_u16(pp, _r04lm.val[0]);
-                vst1_u16(pp + 8, vget_low_u16(_r0145h.val[0]));
-                vst1q_u16(pp + 12, _r15lm.val[0]);
-                vst1_u16(pp + 20, vget_high_u16(_r0145h.val[0]));
-                vst1q_u16(pp + 24, _r26lm.val[0]);
-                vst1_u16(pp + 32, vget_low_u16(_r2367h.val[0]));
-                vst1q_u16(pp + 36, _r37lm.val[0]);
-                vst1_u16(pp + 44, vget_high_u16(_r2367h.val[0]));
-                vst1q_u16(pp + 48, _r04lm.val[1]);
-                vst1_u16(pp + 56, vget_low_u16(_r0145h.val[1]));
-                vst1q_u16(pp + 60, _r15lm.val[1]);
-                vst1_u16(pp + 68, vget_high_u16(_r0145h.val[1]));
-                vst1q_u16(pp + 72, _r26lm.val[1]);
-                vst1_u16(pp + 80, vget_low_u16(_r2367h.val[1]));
-                vst1q_u16(pp + 84, _r37lm.val[1]);
-                vst1_u16(pp + 92, vget_high_u16(_r2367h.val[1]));
-                p0 += max_jj * batch * 8;
-                pp += 96;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 8;
-            p0 += (b * max_jj + jj) * 4;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                // transpose 4x12
-                uint16x8x4_t _r01 = vld4q_u16(p0);
-                uint16x4x4_t _r2 = vld4_u16(p0 + 32);
-                vst1q_u16(pp, _r01.val[0]);
-                vst1_u16(pp + 8, _r2.val[0]);
-                vst1q_u16(pp + 12, _r01.val[1]);
-                vst1_u16(pp + 20, _r2.val[1]);
-                vst1q_u16(pp + 24, _r01.val[2]);
-                vst1_u16(pp + 32, _r2.val[2]);
-                vst1q_u16(pp + 36, _r01.val[3]);
-                vst1_u16(pp + 44, _r2.val[3]);
-                p0 += max_jj * batch * 4;
-                pp += 48;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                // transpose 2x12
-                uint16x8x2_t _r01 = vld2q_u16(p0);
-                uint16x4x2_t _r2 = vld2_u16(p0 + 16);
-                vst1q_u16(pp, _r01.val[0]);
-                vst1_u16(pp + 8, _r2.val[0]);
-                vst1q_u16(pp + 12, _r01.val[1]);
-                vst1_u16(pp + 20, _r2.val[1]);
-                p0 += max_jj * batch * 2;
-                pp += 24;
-            }
-            p0 -= (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                uint16x8_t _r01 = vld1q_u16(p0);
-                uint16x4_t _r2 = vld1_u16(p0 + 8);
-                vst1q_u16(pp, _r01);
-                vst1_u16(pp + 8, _r2);
-                p0 += max_jj * batch;
-                pp += 12;
-            }
-        }
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            const unsigned short* p0 = B;
-
-            int kk = 0;
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                // transpose 8x8
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v4.8h, v5.8h, v6.8h, v7.8h}, [%0] \n"
-
-                    "uzp1   v8.8h, v0.8h, v4.8h         \n"
-                    "uzp2   v12.8h, v0.8h, v4.8h        \n"
-                    "uzp1   v9.8h, v1.8h, v5.8h         \n"
-                    "uzp2   v13.8h, v1.8h, v5.8h        \n"
-
-                    "sub    %0, %0, #64                 \n"
-
-                    "uzp1   v10.8h, v2.8h, v6.8h        \n"
-                    "uzp2   v14.8h, v2.8h, v6.8h        \n"
-                    "uzp1   v11.8h, v3.8h, v7.8h        \n"
-                    "uzp2   v15.8h, v3.8h, v7.8h        \n"
-
-                    "st1    {v8.8h, v9.8h, v10.8h, v11.8h}, [%1], #64 \n"
-                    "st1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                p0 += max_jj * batch * 8;
-#else  // NCNN_GNU_INLINE_ASM
-                uint16x8x4_t _r0 = vld4q_u16(p0);
-                uint16x8x4_t _r1 = vld4q_u16(p0 + 32);
-                uint16x8x2_t _r04 = vuzpq_u16(_r0.val[0], _r1.val[0]);
-                uint16x8x2_t _r15 = vuzpq_u16(_r0.val[1], _r1.val[1]);
-                uint16x8x2_t _r26 = vuzpq_u16(_r0.val[2], _r1.val[2]);
-                uint16x8x2_t _r37 = vuzpq_u16(_r0.val[3], _r1.val[3]);
-                vst1q_u16(pp, _r04.val[0]);
-                vst1q_u16(pp + 8, _r15.val[0]);
-                vst1q_u16(pp + 8 * 2, _r26.val[0]);
-                vst1q_u16(pp + 8 * 3, _r37.val[0]);
-                vst1q_u16(pp + 8 * 4, _r04.val[1]);
-                vst1q_u16(pp + 8 * 5, _r15.val[1]);
-                vst1q_u16(pp + 8 * 6, _r26.val[1]);
-                vst1q_u16(pp + 8 * 7, _r37.val[1]);
-                p0 += max_jj * batch * 8;
-                pp += 64;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 8;
-            p0 += (b * max_jj + jj) * 4;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                // transpose 4x8
-                uint16x8x4_t _r0 = vld4q_u16(p0);
-                vst1q_u16(pp, _r0.val[0]);
-                vst1q_u16(pp + 8, _r0.val[1]);
-                vst1q_u16(pp + 16, _r0.val[2]);
-                vst1q_u16(pp + 24, _r0.val[3]);
-                p0 += max_jj * batch * 4;
-                pp += 32;
-            }
-            p0 -= (b * max_jj + jj) * 4;
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                // transpose 2x8
-                uint16x8x2_t _r0 = vld2q_u16(p0);
-                vst1q_u16(pp, _r0.val[0]);
-                vst1q_u16(pp + 8, _r0.val[1]);
-                p0 += max_jj * batch * 2;
-                pp += 16;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                uint16x8_t _r0 = vld1q_u16(p0);
-                vst1q_u16(pp, _r0);
-                p0 += max_jj * batch;
-                pp += 8;
-            }
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            const unsigned short* p0 = B;
-
-            int kk = 0;
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                // transpose 8x4
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0] \n"
-                    "st4    {v0.8h, v1.8h, v2.8h, v3.8h}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3");
-                p0 += max_jj * batch * 8;
-#else  // NCNN_GNU_INLINE_ASM
-                uint16x8x4_t _r0;
-                _r0.val[0] = vld1q_u16(p0);
-                _r0.val[1] = vld1q_u16(p0 + 8);
-                _r0.val[2] = vld1q_u16(p0 + 16);
-                _r0.val[3] = vld1q_u16(p0 + 24);
-                vst4q_u16(pp, _r0);
-                p0 += max_jj * batch * 8;
-                pp += 32;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 8;
-            p0 += (b * max_jj + jj) * 4;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                // transpose 4x4
-                uint16x4x4_t _r0;
-                _r0.val[0] = vld1_u16(p0);
-                _r0.val[1] = vld1_u16(p0 + 4);
-                _r0.val[2] = vld1_u16(p0 + 8);
-                _r0.val[3] = vld1_u16(p0 + 12);
-                vst4_u16(pp, _r0);
-                p0 += max_jj * batch * 4;
-                pp += 16;
-            }
-            p0 -= (b * max_jj + jj) * 4;
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                // transpose 2x4
-                uint16x4x2_t _r0 = vld2_u16(p0);
-                vst1_u16(pp, _r0.val[0]);
-                vst1_u16(pp + 4, _r0.val[1]);
-                p0 += max_jj * batch * 2;
-                pp += 8;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                uint16x4_t _r0 = vld1_u16(p0);
-                vst1_u16(pp, _r0);
-                p0 += max_jj * batch;
-                pp += 4;
-            }
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            const unsigned short* p0 = B;
-
-            int kk = 0;
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                // transpose 8x2
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #256]       \n"
-                    "ld1    {v0.8h, v1.8h}, [%0]        \n"
-                    "st2    {v0.8h, v1.8h}, [%1], #32   \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1");
-                p0 += max_jj * batch * 8;
-#else  // NCNN_GNU_INLINE_ASM
-                uint16x8x2_t _r0;
-                _r0.val[0] = vld1q_u16(p0);
-                _r0.val[1] = vld1q_u16(p0 + 8);
-                vst2q_u16(pp, _r0);
-                p0 += max_jj * batch * 8;
-                pp += 16;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 8;
-            p0 += (b * max_jj + jj) * 4;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                // transpose 4x2
-                uint16x4x2_t _r0;
-                _r0.val[0] = vld1_u16(p0);
-                _r0.val[1] = vld1_u16(p0 + 4);
-                vst2_u16(pp, _r0);
-                p0 += max_jj * batch * 4;
-                pp += 8;
-            }
-            p0 -= (b * max_jj + jj) * 4;
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[2];
-                pp[2] = p0[1];
-                pp[3] = p0[3];
-                p0 += max_jj * batch * 2;
-                pp += 4;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                p0 += max_jj * batch;
-                pp += 2;
-            }
-        }
-        for (; jj < max_jj; jj++)
-        {
-            const unsigned short* p0 = B;
-
-            int kk = 0;
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                uint16x8_t _r0 = vld1q_u16(p0);
-                vst1q_u16(pp, _r0);
-                p0 += max_jj * batch * 8;
-                pp += 8;
-            }
-            p0 -= (b * max_jj + jj) * 8;
-            p0 += (b * max_jj + jj) * 4;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                uint16x4_t _r0 = vld1_u16(p0);
-                vst1_u16(pp, _r0);
-                p0 += max_jj * batch * 4;
-                pp += 4;
-            }
-            p0 -= (b * max_jj + jj) * 4;
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                p0 += max_jj * batch * 2;
-                pp += 2;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                p0 += max_jj * batch;
-                pp += 1;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd_gemm_transB_packed_tile_fp16sa(const Mat& AT_tile, const Mat& BT_tile, Mat& top_blob, int batch, int max_ii, int max_jj, int k, int max_kk, int use_a53_a55_optimized_kernel)
-{
-    // NCNN_LOGE("conv3x3s1_winograd_gemm_transB_packed_tile_fp16sa %d %d %d", max_ii, max_jj, max_kk);
-    __fp16* outptr = top_blob;
-
-    int ii = 0;
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const __fp16* pAT = AT_tile.row<const __fp16>(b) + max_kk * ii;
-            const __fp16* pB = BT_tile.row<const __fp16>(b);
-
-            int jj = 0;
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const __fp16* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-                if (use_a53_a55_optimized_kernel)
-                {
-                    asm volatile(
-                        "cbz    %w7, 0f                     \n"
-
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%0], #64 \n"
-                        "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                        "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0]      \n"
-                        "subs   %0, %0, #128                \n"
-                        "b      1f                          \n"
-
-                        "0:                                 \n"
-                        "eor    v20.16b, v20.16b, v20.16b   \n"
-                        "eor    v21.16b, v21.16b, v21.16b   \n"
-                        "eor    v22.16b, v22.16b, v22.16b   \n"
-                        "eor    v23.16b, v23.16b, v23.16b   \n"
-                        "eor    v24.16b, v24.16b, v24.16b   \n"
-                        "eor    v25.16b, v25.16b, v25.16b   \n"
-                        "eor    v26.16b, v26.16b, v26.16b   \n"
-                        "eor    v27.16b, v27.16b, v27.16b   \n"
-                        "eor    v28.16b, v28.16b, v28.16b   \n"
-                        "eor    v29.16b, v29.16b, v29.16b   \n"
-                        "eor    v30.16b, v30.16b, v30.16b   \n"
-                        "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                        "1:                                 \n"
-                        "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                        "cmp    w4, #0                      \n"
-                        "beq    3f                          \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v4.8h}, [%1], #16          \n"
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.8h}, [%2], #16          \n"
-
-                        "ldr    d1, [%2], #8                \n"
-                        "ldr    x21, [%2], #8               \n"
-
-                        ".align 4                           \n"
-                        "2:                                 \n"
-                        "ldr    d5, [%1], #8                \n"
-                        "fmla   v20.8h, v4.8h, v0.h[0]      \n"
-                        "ldr    x25, [%1], #8               \n"
-                        "fmla   v21.8h, v4.8h, v0.h[1]      \n"
-                        "ldr    d2, [%2], #8                \n"
-                        "fmla   v22.8h, v4.8h, v0.h[2]      \n"
-                        "ldr    x22, [%2], #8               \n"
-                        "fmla   v23.8h, v4.8h, v0.h[3]      \n"
-                        "ldr    d6, [%1], #8                \n"
-                        "fmla   v24.8h, v4.8h, v0.h[4]      \n"
-                        "ldr    x26, [%1], #8               \n"
-                        "fmla   v25.8h, v4.8h, v0.h[5]      \n"
-                        "ins    v1.d[1], x21                \n"
-                        "fmla   v26.8h, v4.8h, v0.h[6]      \n"
-                        "ldr    d3, [%2], #8                \n"
-                        "fmla   v27.8h, v4.8h, v0.h[7]      \n"
-                        "ldr    x23, [%2], #8               \n"
-                        "fmla   v28.8h, v4.8h, v1.h[0]      \n"
-                        "prfm   pldl1keep, [%2, #256]       \n" // NOTE PRELOAD
-                        "fmla   v29.8h, v4.8h, v1.h[1]      \n"
-                        "ins    v5.d[1], x25                \n"
-                        "fmla   v30.8h, v4.8h, v1.h[2]      \n"
-                        "ldr    d8, [%2], #8                \n"
-                        "fmla   v31.8h, v4.8h, v1.h[3]      \n"
-                        "ldr    x20, [%2], #8               \n"
-                        "fmla   v20.8h, v5.8h, v1.h[4]      \n"
-                        "ldr    d7, [%1], #8                \n"
-                        "fmla   v21.8h, v5.8h, v1.h[5]      \n"
-                        "ins    v2.d[1], x22                \n"
-                        "fmla   v22.8h, v5.8h, v1.h[6]      \n"
-                        "ldr    x27, [%1], #8               \n"
-                        "fmla   v23.8h, v5.8h, v1.h[7]      \n"
-                        "ldr    d9, [%2], #8                \n"
-                        "fmla   v24.8h, v5.8h, v2.h[0]      \n"
-                        "ldr    x21, [%2], #8               \n"
-                        "fmla   v25.8h, v5.8h, v2.h[1]      \n"
-                        "ins    v6.d[1], x26                \n"
-                        "fmla   v26.8h, v5.8h, v2.h[2]      \n"
-                        "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                        "fmla   v27.8h, v5.8h, v2.h[3]      \n"
-                        "ldr    d4, [%1], #8                \n"
-                        "fmla   v28.8h, v5.8h, v2.h[4]      \n"
-                        "ldr    x24, [%1], #8               \n"
-                        "fmla   v29.8h, v5.8h, v2.h[5]      \n"
-                        "ins    v3.d[1], x23                \n"
-                        "fmla   v30.8h, v5.8h, v2.h[6]      \n"
-                        "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                        "fmla   v31.8h, v5.8h, v2.h[7]      \n"
-                        "ldr    d0, [%2], #8                \n"
-                        "fmla   v20.8h, v6.8h, v3.h[0]      \n"
-                        "fmla   v21.8h, v6.8h, v3.h[1]      \n"
-                        "fmla   v22.8h, v6.8h, v3.h[2]      \n"
-                        "fmla   v23.8h, v6.8h, v3.h[3]      \n"
-                        "fmla   v24.8h, v6.8h, v3.h[4]      \n"
-                        "fmla   v25.8h, v6.8h, v3.h[5]      \n"
-                        "ins    v8.d[1], x20                \n"
-                        "fmla   v26.8h, v6.8h, v3.h[6]      \n"
-                        "ldr    x20, [%2], #8               \n"
-                        "fmla   v27.8h, v6.8h, v3.h[7]      \n"
-                        "ldr    d1, [%2], #8                \n"
-                        "fmla   v28.8h, v6.8h, v8.h[0]      \n"
-                        "fmla   v29.8h, v6.8h, v8.h[1]      \n"
-                        "ins    v7.d[1], x27                \n"
-                        "fmla   v30.8h, v6.8h, v8.h[2]      \n"
-                        "fmla   v31.8h, v6.8h, v8.h[3]      \n"
-                        "fmla   v20.8h, v7.8h, v8.h[4]      \n"
-                        "fmla   v21.8h, v7.8h, v8.h[5]      \n"
-                        "ins    v9.d[1], x21                \n"
-                        "fmla   v22.8h, v7.8h, v8.h[6]      \n"
-                        "fmla   v23.8h, v7.8h, v8.h[7]      \n"
-                        "ldr    x21, [%2], #8               \n"
-                        "fmla   v24.8h, v7.8h, v9.h[0]      \n"
-                        "fmla   v25.8h, v7.8h, v9.h[1]      \n"
-                        "ins    v4.d[1], x24                \n"
-                        "fmla   v26.8h, v7.8h, v9.h[2]      \n"
-                        "fmla   v27.8h, v7.8h, v9.h[3]      \n"
-                        "subs   w4, w4, #1                  \n"
-                        "fmla   v28.8h, v7.8h, v9.h[4]      \n"
-                        "fmla   v29.8h, v7.8h, v9.h[5]      \n"
-                        "fmla   v30.8h, v7.8h, v9.h[6]      \n"
-                        "ins    v0.d[1], x20                \n"
-                        "fmla   v31.8h, v7.8h, v9.h[7]      \n"
-                        "bne    2b                          \n"
-
-                        "sub    %1, %1, #16                 \n"
-                        "sub    %2, %2, #32                 \n"
-
-                        "3:                                 \n"
-                        "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                        "cmp    w4, #0                      \n"
-                        "beq    5f                          \n"
-
-                        "4:                                 \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%2], #24 \n"
-                        "ld1    {v4.8h}, [%1], #16          \n"
-                        "fmla   v20.8h, v4.8h, v0.h[0]      \n"
-                        "fmla   v21.8h, v4.8h, v0.h[1]      \n"
-                        "fmla   v22.8h, v4.8h, v0.h[2]      \n"
-                        "fmla   v23.8h, v4.8h, v0.h[3]      \n"
-                        "fmla   v24.8h, v4.8h, v1.h[0]      \n"
-                        "fmla   v25.8h, v4.8h, v1.h[1]      \n"
-                        "fmla   v26.8h, v4.8h, v1.h[2]      \n"
-                        "fmla   v27.8h, v4.8h, v1.h[3]      \n"
-                        "subs   w4, w4, #1                  \n"
-                        "fmla   v28.8h, v4.8h, v2.h[0]      \n"
-                        "fmla   v29.8h, v4.8h, v2.h[1]      \n"
-                        "fmla   v30.8h, v4.8h, v2.h[2]      \n"
-                        "fmla   v31.8h, v4.8h, v2.h[3]      \n"
-                        "bne    4b                          \n"
-
-                        "5:                                 \n"
-                        "st1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%0], #64 \n"
-                        "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                        "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                        : "=r"(outptr), // %0
-                        "=r"(pA),     // %1
-                        "=r"(pB)      // %2
-                        : "0"(outptr),
-                        "1"(pA),
-                        "2"(pB),
-                        "r"(max_kk), // %6
-                        "r"(k)       // %7
-                        : "cc", "memory", "x4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-                else
-                {
-                    asm volatile(
-                        "cbz    %w7, 0f                     \n"
-
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%0], #64 \n"
-                        "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                        "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0]      \n"
-                        "subs   %0, %0, #128                \n"
-                        "b      1f                          \n"
-
-                        "0:                                 \n"
-                        "eor    v20.16b, v20.16b, v20.16b   \n"
-                        "eor    v21.16b, v21.16b, v21.16b   \n"
-                        "eor    v22.16b, v22.16b, v22.16b   \n"
-                        "eor    v23.16b, v23.16b, v23.16b   \n"
-                        "eor    v24.16b, v24.16b, v24.16b   \n"
-                        "eor    v25.16b, v25.16b, v25.16b   \n"
-                        "eor    v26.16b, v26.16b, v26.16b   \n"
-                        "eor    v27.16b, v27.16b, v27.16b   \n"
-                        "eor    v28.16b, v28.16b, v28.16b   \n"
-                        "eor    v29.16b, v29.16b, v29.16b   \n"
-                        "eor    v30.16b, v30.16b, v30.16b   \n"
-                        "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                        "1:                                 \n"
-                        "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                        "cmp    w4, #0                      \n"
-                        "beq    3f                          \n"
-
-                        "2:                                 \n"
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%1], #64 \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%2], #64 \n"
-
-                        "fmla   v20.8h, v4.8h, v0.h[0]      \n"
-                        "fmla   v21.8h, v4.8h, v0.h[1]      \n"
-                        "fmla   v22.8h, v4.8h, v0.h[2]      \n"
-                        "fmla   v23.8h, v4.8h, v0.h[3]      \n"
-                        "fmla   v24.8h, v4.8h, v0.h[4]      \n"
-                        "fmla   v25.8h, v4.8h, v0.h[5]      \n"
-                        "fmla   v26.8h, v4.8h, v0.h[6]      \n"
-                        "fmla   v27.8h, v4.8h, v0.h[7]      \n"
-                        "fmla   v28.8h, v4.8h, v1.h[0]      \n"
-                        "fmla   v29.8h, v4.8h, v1.h[1]      \n"
-                        "fmla   v30.8h, v4.8h, v1.h[2]      \n"
-                        "fmla   v31.8h, v4.8h, v1.h[3]      \n"
-
-                        "fmla   v20.8h, v5.8h, v1.h[4]      \n"
-                        "fmla   v21.8h, v5.8h, v1.h[5]      \n"
-                        "fmla   v22.8h, v5.8h, v1.h[6]      \n"
-                        "fmla   v23.8h, v5.8h, v1.h[7]      \n"
-                        "fmla   v24.8h, v5.8h, v2.h[0]      \n"
-                        "fmla   v25.8h, v5.8h, v2.h[1]      \n"
-                        "fmla   v26.8h, v5.8h, v2.h[2]      \n"
-                        "fmla   v27.8h, v5.8h, v2.h[3]      \n"
-                        "fmla   v28.8h, v5.8h, v2.h[4]      \n"
-                        "fmla   v29.8h, v5.8h, v2.h[5]      \n"
-                        "fmla   v30.8h, v5.8h, v2.h[6]      \n"
-                        "fmla   v31.8h, v5.8h, v2.h[7]      \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v8.8h, v9.8h}, [%2], #32   \n"
-
-                        "fmla   v20.8h, v6.8h, v3.h[0]      \n"
-                        "fmla   v21.8h, v6.8h, v3.h[1]      \n"
-                        "fmla   v22.8h, v6.8h, v3.h[2]      \n"
-                        "fmla   v23.8h, v6.8h, v3.h[3]      \n"
-                        "fmla   v24.8h, v6.8h, v3.h[4]      \n"
-                        "fmla   v25.8h, v6.8h, v3.h[5]      \n"
-                        "fmla   v26.8h, v6.8h, v3.h[6]      \n"
-                        "fmla   v27.8h, v6.8h, v3.h[7]      \n"
-                        "fmla   v28.8h, v6.8h, v8.h[0]      \n"
-                        "fmla   v29.8h, v6.8h, v8.h[1]      \n"
-                        "fmla   v30.8h, v6.8h, v8.h[2]      \n"
-                        "fmla   v31.8h, v6.8h, v8.h[3]      \n"
-
-                        "subs   w4, w4, #1                  \n"
-
-                        "fmla   v20.8h, v7.8h, v8.h[4]      \n"
-                        "fmla   v21.8h, v7.8h, v8.h[5]      \n"
-                        "fmla   v22.8h, v7.8h, v8.h[6]      \n"
-                        "fmla   v23.8h, v7.8h, v8.h[7]      \n"
-                        "fmla   v24.8h, v7.8h, v9.h[0]      \n"
-                        "fmla   v25.8h, v7.8h, v9.h[1]      \n"
-                        "fmla   v26.8h, v7.8h, v9.h[2]      \n"
-                        "fmla   v27.8h, v7.8h, v9.h[3]      \n"
-                        "fmla   v28.8h, v7.8h, v9.h[4]      \n"
-                        "fmla   v29.8h, v7.8h, v9.h[5]      \n"
-                        "fmla   v30.8h, v7.8h, v9.h[6]      \n"
-                        "fmla   v31.8h, v7.8h, v9.h[7]      \n"
-
-                        "bne    2b                          \n"
-
-                        "3:                                 \n"
-                        "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                        "cmp    w4, #0                      \n"
-                        "beq    5f                          \n"
-
-                        "4:                                 \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%2], #24 \n"
-                        "ld1    {v4.8h}, [%1], #16          \n"
-                        "fmla   v20.8h, v4.8h, v0.h[0]      \n"
-                        "fmla   v21.8h, v4.8h, v0.h[1]      \n"
-                        "fmla   v22.8h, v4.8h, v0.h[2]      \n"
-                        "fmla   v23.8h, v4.8h, v0.h[3]      \n"
-                        "fmla   v24.8h, v4.8h, v1.h[0]      \n"
-                        "fmla   v25.8h, v4.8h, v1.h[1]      \n"
-                        "fmla   v26.8h, v4.8h, v1.h[2]      \n"
-                        "fmla   v27.8h, v4.8h, v1.h[3]      \n"
-                        "subs   w4, w4, #1                  \n"
-                        "fmla   v28.8h, v4.8h, v2.h[0]      \n"
-                        "fmla   v29.8h, v4.8h, v2.h[1]      \n"
-                        "fmla   v30.8h, v4.8h, v2.h[2]      \n"
-                        "fmla   v31.8h, v4.8h, v2.h[3]      \n"
-                        "bne    4b                          \n"
-
-                        "5:                                 \n"
-                        "st1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%0], #64 \n"
-                        "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                        "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                        : "=r"(outptr), // %0
-                        "=r"(pA),     // %1
-                        "=r"(pB)      // %2
-                        : "0"(outptr),
-                        "1"(pA),
-                        "2"(pB),
-                        "r"(max_kk), // %6
-                        "r"(k)       // %7
-                        : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-#else  // NCNN_GNU_INLINE_ASM
-                float16x8_t _sum0;
-                float16x8_t _sum1;
-                float16x8_t _sum2;
-                float16x8_t _sum3;
-                float16x8_t _sum4;
-                float16x8_t _sum5;
-                float16x8_t _sum6;
-                float16x8_t _sum7;
-                float16x8_t _sum8;
-                float16x8_t _sum9;
-                float16x8_t _suma;
-                float16x8_t _sumb;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_f16(0.f);
-                    _sum1 = vdupq_n_f16(0.f);
-                    _sum2 = vdupq_n_f16(0.f);
-                    _sum3 = vdupq_n_f16(0.f);
-                    _sum4 = vdupq_n_f16(0.f);
-                    _sum5 = vdupq_n_f16(0.f);
-                    _sum6 = vdupq_n_f16(0.f);
-                    _sum7 = vdupq_n_f16(0.f);
-                    _sum8 = vdupq_n_f16(0.f);
-                    _sum9 = vdupq_n_f16(0.f);
-                    _suma = vdupq_n_f16(0.f);
-                    _sumb = vdupq_n_f16(0.f);
-                }
-                else
-                {
-                    _sum0 = vld1q_f16(outptr);
-                    _sum1 = vld1q_f16(outptr + 8);
-                    _sum2 = vld1q_f16(outptr + 16);
-                    _sum3 = vld1q_f16(outptr + 24);
-                    _sum4 = vld1q_f16(outptr + 32);
-                    _sum5 = vld1q_f16(outptr + 40);
-                    _sum6 = vld1q_f16(outptr + 48);
-                    _sum7 = vld1q_f16(outptr + 56);
-                    _sum8 = vld1q_f16(outptr + 64);
-                    _sum9 = vld1q_f16(outptr + 72);
-                    _suma = vld1q_f16(outptr + 80);
-                    _sumb = vld1q_f16(outptr + 88);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x8_t _pA = vld1q_f16(pA);
-                    float16x8_t _pB0 = vld1q_f16(pB);
-                    float16x4_t _pB2 = vld1_f16(pB + 8);
-                    _sum0 = vfmaq_laneq_f16(_sum0, _pA, _pB0, 0);
-                    _sum1 = vfmaq_laneq_f16(_sum1, _pA, _pB0, 1);
-                    _sum2 = vfmaq_laneq_f16(_sum2, _pA, _pB0, 2);
-                    _sum3 = vfmaq_laneq_f16(_sum3, _pA, _pB0, 3);
-                    _sum4 = vfmaq_laneq_f16(_sum4, _pA, _pB0, 4);
-                    _sum5 = vfmaq_laneq_f16(_sum5, _pA, _pB0, 5);
-                    _sum6 = vfmaq_laneq_f16(_sum6, _pA, _pB0, 6);
-                    _sum7 = vfmaq_laneq_f16(_sum7, _pA, _pB0, 7);
-                    _sum8 = vfmaq_lane_f16(_sum8, _pA, _pB2, 0);
-                    _sum9 = vfmaq_lane_f16(_sum9, _pA, _pB2, 1);
-                    _suma = vfmaq_lane_f16(_suma, _pA, _pB2, 2);
-                    _sumb = vfmaq_lane_f16(_sumb, _pA, _pB2, 3);
-
-                    pA += 8;
-                    pB += 12;
-                }
-
-                vst1q_f16(outptr, _sum0);
-                vst1q_f16(outptr + 8, _sum1);
-                vst1q_f16(outptr + 8 * 2, _sum2);
-                vst1q_f16(outptr + 8 * 3, _sum3);
-                vst1q_f16(outptr + 8 * 4, _sum4);
-                vst1q_f16(outptr + 8 * 5, _sum5);
-                vst1q_f16(outptr + 8 * 6, _sum6);
-                vst1q_f16(outptr + 8 * 7, _sum7);
-                vst1q_f16(outptr + 8 * 8, _sum8);
-                vst1q_f16(outptr + 8 * 9, _sum9);
-                vst1q_f16(outptr + 8 * 10, _suma);
-                vst1q_f16(outptr + 8 * 11, _sumb);
-                outptr += 8 * 12;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const __fp16* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-                if (use_a53_a55_optimized_kernel)
-                {
-                    asm volatile(
-                        "cbz    %w7, 0f                     \n"
-
-                        "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                        "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0]      \n"
-                        "subs   %0, %0, #64                 \n"
-                        "b      1f                          \n"
-
-                        "0:                                 \n"
-                        "eor    v24.16b, v24.16b, v24.16b   \n"
-                        "eor    v25.16b, v25.16b, v25.16b   \n"
-                        "eor    v26.16b, v26.16b, v26.16b   \n"
-                        "eor    v27.16b, v27.16b, v27.16b   \n"
-                        "eor    v28.16b, v28.16b, v28.16b   \n"
-                        "eor    v29.16b, v29.16b, v29.16b   \n"
-                        "eor    v30.16b, v30.16b, v30.16b   \n"
-                        "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                        "1:                                 \n"
-                        "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                        "cmp    w4, #0                      \n"
-                        "beq    3f                          \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v4.8h}, [%1], #16          \n"
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.8h}, [%2], #16          \n"
-
-                        "ldr    d5, [%1], #8                \n"
-                        "ldr    x25, [%1], #8               \n"
-
-                        ".align 4                           \n"
-                        "2:                                 \n"
-                        "ldr    d1, [%2], #8                \n"
-                        "fmla   v24.8h, v4.8h, v0.h[0]      \n"
-                        "ldr    x21, [%2], #8               \n"
-                        "fmla   v25.8h, v4.8h, v0.h[1]      \n"
-                        "ins    v5.d[1], x25                \n"
-                        "fmla   v26.8h, v4.8h, v0.h[2]      \n"
-                        "ldr    d6, [%1], #8                \n"
-                        "fmla   v27.8h, v4.8h, v0.h[3]      \n"
-                        "ldr    x26, [%1], #8               \n"
-                        "fmla   v28.8h, v4.8h, v0.h[4]      \n"
-                        "ldr    d2, [%2], #8                \n"
-                        "fmla   v29.8h, v4.8h, v0.h[5]      \n"
-                        "ins    v1.d[1], x21                \n"
-                        "fmla   v30.8h, v4.8h, v0.h[6]      \n"
-                        "ldr    x22, [%2], #8               \n"
-                        "fmla   v31.8h, v4.8h, v0.h[7]      \n"
-                        "ldr    d7, [%1], #8                \n"
-                        "fmla   v24.8h, v5.8h, v1.h[0]      \n"
-                        "ldr    x27, [%1], #8               \n"
-                        "fmla   v25.8h, v5.8h, v1.h[1]      \n"
-                        "ins    v6.d[1], x26                \n"
-                        "fmla   v26.8h, v5.8h, v1.h[2]      \n"
-                        "ldr    d3, [%2], #8                \n"
-                        "fmla   v27.8h, v5.8h, v1.h[3]      \n"
-                        "ldr    x23, [%2], #8               \n"
-                        "fmla   v28.8h, v5.8h, v1.h[4]      \n"
-                        "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                        "fmla   v29.8h, v5.8h, v1.h[5]      \n"
-                        "ins    v2.d[1], x22                \n"
-                        "fmla   v30.8h, v5.8h, v1.h[6]      \n"
-                        "ldr    d4, [%1], #8                \n"
-                        "fmla   v31.8h, v5.8h, v1.h[7]      \n"
-                        "ldr    x24, [%1], #8               \n"
-                        "fmla   v24.8h, v6.8h, v2.h[0]      \n"
-                        "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                        "fmla   v25.8h, v6.8h, v2.h[1]      \n"
-                        "ins    v7.d[1], x27                \n"
-                        "fmla   v26.8h, v6.8h, v2.h[2]      \n"
-                        "ldr    d0, [%2], #8                \n"
-                        "fmla   v27.8h, v6.8h, v2.h[3]      \n"
-                        "ldr    x20, [%2], #8               \n"
-                        "fmla   v28.8h, v6.8h, v2.h[4]      \n"
-                        "ldr    d5, [%1], #8                \n"
-                        "fmla   v29.8h, v6.8h, v2.h[5]      \n"
-                        "ins    v3.d[1], x23                \n"
-                        "fmla   v30.8h, v6.8h, v2.h[6]      \n"
-                        "ldr    x25, [%1], #8               \n"
-                        "fmla   v31.8h, v6.8h, v2.h[7]      \n"
-                        "fmla   v24.8h, v7.8h, v3.h[0]      \n"
-                        "fmla   v25.8h, v7.8h, v3.h[1]      \n"
-                        "fmla   v26.8h, v7.8h, v3.h[2]      \n"
-                        "ins    v4.d[1], x24                \n"
-                        "fmla   v27.8h, v7.8h, v3.h[3]      \n"
-                        "fmla   v28.8h, v7.8h, v3.h[4]      \n"
-                        "subs   w4, w4, #1                  \n"
-                        "fmla   v29.8h, v7.8h, v3.h[5]      \n"
-                        "fmla   v30.8h, v7.8h, v3.h[6]      \n"
-                        "ins    v0.d[1], x20                \n"
-                        "fmla   v31.8h, v7.8h, v3.h[7]      \n"
-                        "bne    2b                          \n"
-
-                        "sub    %1, %1, #32                 \n"
-                        "sub    %2, %2, #16                 \n"
-
-                        "3:                                 \n"
-                        "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                        "cmp    w4, #0                      \n"
-                        "beq    5f                          \n"
-
-                        "4:                                 \n"
-                        "ld1    {v0.8h}, [%2], #16          \n"
-                        "ld1    {v4.8h}, [%1], #16          \n"
-                        "fmla   v24.8h, v4.8h, v0.h[0]      \n"
-                        "fmla   v25.8h, v4.8h, v0.h[1]      \n"
-                        "fmla   v26.8h, v4.8h, v0.h[2]      \n"
-                        "fmla   v27.8h, v4.8h, v0.h[3]      \n"
-                        "subs   w4, w4, #1                  \n"
-                        "fmla   v28.8h, v4.8h, v0.h[4]      \n"
-                        "fmla   v29.8h, v4.8h, v0.h[5]      \n"
-                        "fmla   v30.8h, v4.8h, v0.h[6]      \n"
-                        "fmla   v31.8h, v4.8h, v0.h[7]      \n"
-                        "bne    4b                          \n"
-
-                        "5:                                 \n"
-                        "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                        "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                        : "=r"(outptr), // %0
-                        "=r"(pA),     // %1
-                        "=r"(pB)      // %2
-                        : "0"(outptr),
-                        "1"(pA),
-                        "2"(pB),
-                        "r"(max_kk), // %6
-                        "r"(k)       // %7
-                        : "cc", "memory", "x4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-                else
-                {
-                    asm volatile(
-                        "cbz    %w7, 0f                     \n"
-
-                        "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                        "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0]      \n"
-                        "subs   %0, %0, #64                 \n"
-                        "b      1f                          \n"
-
-                        "0:                                 \n"
-                        "eor    v24.16b, v24.16b, v24.16b   \n"
-                        "eor    v25.16b, v25.16b, v25.16b   \n"
-                        "eor    v26.16b, v26.16b, v26.16b   \n"
-                        "eor    v27.16b, v27.16b, v27.16b   \n"
-                        "eor    v28.16b, v28.16b, v28.16b   \n"
-                        "eor    v29.16b, v29.16b, v29.16b   \n"
-                        "eor    v30.16b, v30.16b, v30.16b   \n"
-                        "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                        "1:                                 \n"
-                        "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                        "cmp    w4, #0                      \n"
-                        "beq    3f                          \n"
-
-                        "2:                                 \n"
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%1], #64 \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%2], #64 \n"
-
-                        "fmla   v24.8h, v4.8h, v0.h[0]      \n"
-                        "fmla   v25.8h, v4.8h, v0.h[1]      \n"
-                        "fmla   v26.8h, v4.8h, v0.h[2]      \n"
-                        "fmla   v27.8h, v4.8h, v0.h[3]      \n"
-                        "fmla   v28.8h, v4.8h, v0.h[4]      \n"
-                        "fmla   v29.8h, v4.8h, v0.h[5]      \n"
-                        "fmla   v30.8h, v4.8h, v0.h[6]      \n"
-                        "fmla   v31.8h, v4.8h, v0.h[7]      \n"
-
-                        "fmla   v24.8h, v5.8h, v1.h[0]      \n"
-                        "fmla   v25.8h, v5.8h, v1.h[1]      \n"
-                        "fmla   v26.8h, v5.8h, v1.h[2]      \n"
-                        "fmla   v27.8h, v5.8h, v1.h[3]      \n"
-                        "fmla   v28.8h, v5.8h, v1.h[4]      \n"
-                        "fmla   v29.8h, v5.8h, v1.h[5]      \n"
-                        "fmla   v30.8h, v5.8h, v1.h[6]      \n"
-                        "fmla   v31.8h, v5.8h, v1.h[7]      \n"
-
-                        "fmla   v24.8h, v6.8h, v2.h[0]      \n"
-                        "fmla   v25.8h, v6.8h, v2.h[1]      \n"
-                        "fmla   v26.8h, v6.8h, v2.h[2]      \n"
-                        "fmla   v27.8h, v6.8h, v2.h[3]      \n"
-                        "fmla   v28.8h, v6.8h, v2.h[4]      \n"
-                        "fmla   v29.8h, v6.8h, v2.h[5]      \n"
-                        "fmla   v30.8h, v6.8h, v2.h[6]      \n"
-                        "fmla   v31.8h, v6.8h, v2.h[7]      \n"
-
-                        "subs   w4, w4, #1                  \n"
-
-                        "fmla   v24.8h, v7.8h, v3.h[0]      \n"
-                        "fmla   v25.8h, v7.8h, v3.h[1]      \n"
-                        "fmla   v26.8h, v7.8h, v3.h[2]      \n"
-                        "fmla   v27.8h, v7.8h, v3.h[3]      \n"
-                        "fmla   v28.8h, v7.8h, v3.h[4]      \n"
-                        "fmla   v29.8h, v7.8h, v3.h[5]      \n"
-                        "fmla   v30.8h, v7.8h, v3.h[6]      \n"
-                        "fmla   v31.8h, v7.8h, v3.h[7]      \n"
-
-                        "bne    2b                          \n"
-
-                        "3:                                 \n"
-                        "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                        "cmp    w4, #0                      \n"
-                        "beq    5f                          \n"
-
-                        "4:                                 \n"
-                        "ld1    {v0.8h}, [%2], #16          \n"
-                        "ld1    {v4.8h}, [%1], #16          \n"
-                        "fmla   v24.8h, v4.8h, v0.h[0]      \n"
-                        "fmla   v25.8h, v4.8h, v0.h[1]      \n"
-                        "fmla   v26.8h, v4.8h, v0.h[2]      \n"
-                        "fmla   v27.8h, v4.8h, v0.h[3]      \n"
-                        "subs   w4, w4, #1                  \n"
-                        "fmla   v28.8h, v4.8h, v0.h[4]      \n"
-                        "fmla   v29.8h, v4.8h, v0.h[5]      \n"
-                        "fmla   v30.8h, v4.8h, v0.h[6]      \n"
-                        "fmla   v31.8h, v4.8h, v0.h[7]      \n"
-                        "bne    4b                          \n"
-
-                        "5:                                 \n"
-                        "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                        "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                        : "=r"(outptr), // %0
-                        "=r"(pA),     // %1
-                        "=r"(pB)      // %2
-                        : "0"(outptr),
-                        "1"(pA),
-                        "2"(pB),
-                        "r"(max_kk), // %6
-                        "r"(k)       // %7
-                        : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-#else  // NCNN_GNU_INLINE_ASM
-                float16x8_t _sum0;
-                float16x8_t _sum1;
-                float16x8_t _sum2;
-                float16x8_t _sum3;
-                float16x8_t _sum4;
-                float16x8_t _sum5;
-                float16x8_t _sum6;
-                float16x8_t _sum7;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_f16(0.f);
-                    _sum1 = vdupq_n_f16(0.f);
-                    _sum2 = vdupq_n_f16(0.f);
-                    _sum3 = vdupq_n_f16(0.f);
-                    _sum4 = vdupq_n_f16(0.f);
-                    _sum5 = vdupq_n_f16(0.f);
-                    _sum6 = vdupq_n_f16(0.f);
-                    _sum7 = vdupq_n_f16(0.f);
-                }
-                else
-                {
-                    _sum0 = vld1q_f16(outptr);
-                    _sum1 = vld1q_f16(outptr + 8);
-                    _sum2 = vld1q_f16(outptr + 16);
-                    _sum3 = vld1q_f16(outptr + 24);
-                    _sum4 = vld1q_f16(outptr + 32);
-                    _sum5 = vld1q_f16(outptr + 40);
-                    _sum6 = vld1q_f16(outptr + 48);
-                    _sum7 = vld1q_f16(outptr + 56);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x8_t _pA = vld1q_f16(pA);
-                    float16x8_t _pB = vld1q_f16(pB);
-                    _sum0 = vfmaq_laneq_f16(_sum0, _pA, _pB, 0);
-                    _sum1 = vfmaq_laneq_f16(_sum1, _pA, _pB, 1);
-                    _sum2 = vfmaq_laneq_f16(_sum2, _pA, _pB, 2);
-                    _sum3 = vfmaq_laneq_f16(_sum3, _pA, _pB, 3);
-                    _sum4 = vfmaq_laneq_f16(_sum4, _pA, _pB, 4);
-                    _sum5 = vfmaq_laneq_f16(_sum5, _pA, _pB, 5);
-                    _sum6 = vfmaq_laneq_f16(_sum6, _pA, _pB, 6);
-                    _sum7 = vfmaq_laneq_f16(_sum7, _pA, _pB, 7);
-
-                    pA += 8;
-                    pB += 8;
-                }
-
-                vst1q_f16(outptr, _sum0);
-                vst1q_f16(outptr + 8, _sum1);
-                vst1q_f16(outptr + 8 * 2, _sum2);
-                vst1q_f16(outptr + 8 * 3, _sum3);
-                vst1q_f16(outptr + 8 * 4, _sum4);
-                vst1q_f16(outptr + 8 * 5, _sum5);
-                vst1q_f16(outptr + 8 * 6, _sum6);
-                vst1q_f16(outptr + 8 * 7, _sum7);
-                outptr += 8 * 8;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const __fp16* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "cbz    %w7, 0f                     \n"
-
-                    "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0] \n"
-                    "b      1f                          \n"
-
-                    "0:                                 \n"
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-                    "eor    v29.16b, v29.16b, v29.16b   \n"
-                    "eor    v30.16b, v30.16b, v30.16b   \n"
-                    "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                    "1:                                 \n"
-                    "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    3f                          \n"
-
-                    "2:                                 \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%1], #64 \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v0.8h, v1.8h}, [%2], #32   \n"
-
-                    "fmla   v28.8h, v4.8h, v0.h[0]      \n"
-                    "fmla   v29.8h, v4.8h, v0.h[1]      \n"
-                    "fmla   v30.8h, v4.8h, v0.h[2]      \n"
-                    "fmla   v31.8h, v4.8h, v0.h[3]      \n"
-
-                    "fmla   v28.8h, v5.8h, v0.h[4]      \n"
-                    "fmla   v29.8h, v5.8h, v0.h[5]      \n"
-                    "fmla   v30.8h, v5.8h, v0.h[6]      \n"
-                    "fmla   v31.8h, v5.8h, v0.h[7]      \n"
-
-                    "fmla   v28.8h, v6.8h, v1.h[0]      \n"
-                    "fmla   v29.8h, v6.8h, v1.h[1]      \n"
-                    "fmla   v30.8h, v6.8h, v1.h[2]      \n"
-                    "fmla   v31.8h, v6.8h, v1.h[3]      \n"
-
-                    "subs   w4, w4, #1                  \n"
-
-                    "fmla   v28.8h, v7.8h, v1.h[4]      \n"
-                    "fmla   v29.8h, v7.8h, v1.h[5]      \n"
-                    "fmla   v30.8h, v7.8h, v1.h[6]      \n"
-                    "fmla   v31.8h, v7.8h, v1.h[7]      \n"
-
-                    "bne    2b                          \n"
-
-                    "3:                                 \n"
-                    "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "ld1    {v0.4h}, [%2], #8           \n"
-                    "ld1    {v4.8h}, [%1], #16          \n"
-                    "fmla   v28.8h, v4.8h, v0.h[0]      \n"
-                    "fmla   v29.8h, v4.8h, v0.h[1]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v30.8h, v4.8h, v0.h[2]      \n"
-                    "fmla   v31.8h, v4.8h, v0.h[3]      \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "x4", "v0", "v1", "v4", "v5", "v6", "v7", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-                float16x8_t _sum0;
-                float16x8_t _sum1;
-                float16x8_t _sum2;
-                float16x8_t _sum3;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_f16(0.f);
-                    _sum1 = vdupq_n_f16(0.f);
-                    _sum2 = vdupq_n_f16(0.f);
-                    _sum3 = vdupq_n_f16(0.f);
-                }
-                else
-                {
-                    _sum0 = vld1q_f16(outptr);
-                    _sum1 = vld1q_f16(outptr + 8);
-                    _sum2 = vld1q_f16(outptr + 16);
-                    _sum3 = vld1q_f16(outptr + 24);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x8_t _pA = vld1q_f16(pA);
-                    float16x4_t _pB = vld1_f16(pB);
-                    _sum0 = vfmaq_lane_f16(_sum0, _pA, _pB, 0);
-                    _sum1 = vfmaq_lane_f16(_sum1, _pA, _pB, 1);
-                    _sum2 = vfmaq_lane_f16(_sum2, _pA, _pB, 2);
-                    _sum3 = vfmaq_lane_f16(_sum3, _pA, _pB, 3);
-
-                    pA += 8;
-                    pB += 4;
-                }
-
-                vst1q_f16(outptr, _sum0);
-                vst1q_f16(outptr + 8, _sum1);
-                vst1q_f16(outptr + 8 * 2, _sum2);
-                vst1q_f16(outptr + 8 * 3, _sum3);
-                outptr += 8 * 4;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const __fp16* pA = pAT;
-
-                float16x8_t _sum0;
-                float16x8_t _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_f16(0.f);
-                    _sum1 = vdupq_n_f16(0.f);
-                }
-                else
-                {
-                    _sum0 = vld1q_f16(outptr);
-                    _sum1 = vld1q_f16(outptr + 8);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x8_t _pA = vld1q_f16(pA);
-                    _sum0 = vfmaq_n_f16(_sum0, _pA, pB[0]);
-                    _sum1 = vfmaq_n_f16(_sum1, _pA, pB[1]);
-
-                    pA += 8;
-                    pB += 2;
-                }
-
-                vst1q_f16(outptr, _sum0);
-                vst1q_f16(outptr + 8, _sum1);
-                outptr += 8 * 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const __fp16* pA = pAT;
-
-                float16x8_t _sum;
-
-                if (k == 0)
-                {
-                    _sum = vdupq_n_f16(0.f);
-                }
-                else
-                {
-                    _sum = vld1q_f16(outptr);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x8_t _pA = vld1q_f16(pA);
-                    _sum = vfmaq_n_f16(_sum, _pA, pB[0]);
-
-                    pA += 8;
-                    pB += 1;
-                }
-
-                vst1q_f16(outptr, _sum);
-                outptr += 8;
-            }
-        }
-    }
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const __fp16* pAT = AT_tile.row<const __fp16>(b) + max_kk * ii;
-            const __fp16* pB = BT_tile.row<const __fp16>(b);
-
-            int jj = 0;
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const __fp16* pA = pAT;
-
-                float16x4_t _sum0;
-                float16x4_t _sum1;
-                float16x4_t _sum2;
-                float16x4_t _sum3;
-                float16x4_t _sum4;
-                float16x4_t _sum5;
-                float16x4_t _sum6;
-                float16x4_t _sum7;
-                float16x4_t _sum8;
-                float16x4_t _sum9;
-                float16x4_t _suma;
-                float16x4_t _sumb;
-
-                if (k == 0)
-                {
-                    _sum0 = vdup_n_f16(0.f);
-                    _sum1 = vdup_n_f16(0.f);
-                    _sum2 = vdup_n_f16(0.f);
-                    _sum3 = vdup_n_f16(0.f);
-                    _sum4 = vdup_n_f16(0.f);
-                    _sum5 = vdup_n_f16(0.f);
-                    _sum6 = vdup_n_f16(0.f);
-                    _sum7 = vdup_n_f16(0.f);
-                    _sum8 = vdup_n_f16(0.f);
-                    _sum9 = vdup_n_f16(0.f);
-                    _suma = vdup_n_f16(0.f);
-                    _sumb = vdup_n_f16(0.f);
-                }
-                else
-                {
-                    _sum0 = vld1_f16(outptr);
-                    _sum1 = vld1_f16(outptr + 4);
-                    _sum2 = vld1_f16(outptr + 8);
-                    _sum3 = vld1_f16(outptr + 12);
-                    _sum4 = vld1_f16(outptr + 16);
-                    _sum5 = vld1_f16(outptr + 20);
-                    _sum6 = vld1_f16(outptr + 24);
-                    _sum7 = vld1_f16(outptr + 28);
-                    _sum8 = vld1_f16(outptr + 32);
-                    _sum9 = vld1_f16(outptr + 36);
-                    _suma = vld1_f16(outptr + 40);
-                    _sumb = vld1_f16(outptr + 44);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x4_t _pA = vld1_f16(pA);
-                    float16x8_t _pB0 = vld1q_f16(pB);
-                    float16x4_t _pB2 = vld1_f16(pB + 8);
-                    _sum0 = vfma_laneq_f16(_sum0, _pA, _pB0, 0);
-                    _sum1 = vfma_laneq_f16(_sum1, _pA, _pB0, 1);
-                    _sum2 = vfma_laneq_f16(_sum2, _pA, _pB0, 2);
-                    _sum3 = vfma_laneq_f16(_sum3, _pA, _pB0, 3);
-                    _sum4 = vfma_laneq_f16(_sum4, _pA, _pB0, 4);
-                    _sum5 = vfma_laneq_f16(_sum5, _pA, _pB0, 5);
-                    _sum6 = vfma_laneq_f16(_sum6, _pA, _pB0, 6);
-                    _sum7 = vfma_laneq_f16(_sum7, _pA, _pB0, 7);
-                    _sum8 = vfma_lane_f16(_sum8, _pA, _pB2, 0);
-                    _sum9 = vfma_lane_f16(_sum9, _pA, _pB2, 1);
-                    _suma = vfma_lane_f16(_suma, _pA, _pB2, 2);
-                    _sumb = vfma_lane_f16(_sumb, _pA, _pB2, 3);
-
-                    pA += 4;
-                    pB += 12;
-                }
-
-                vst1_f16(outptr, _sum0);
-                vst1_f16(outptr + 4, _sum1);
-                vst1_f16(outptr + 4 * 2, _sum2);
-                vst1_f16(outptr + 4 * 3, _sum3);
-                vst1_f16(outptr + 4 * 4, _sum4);
-                vst1_f16(outptr + 4 * 5, _sum5);
-                vst1_f16(outptr + 4 * 6, _sum6);
-                vst1_f16(outptr + 4 * 7, _sum7);
-                vst1_f16(outptr + 4 * 8, _sum8);
-                vst1_f16(outptr + 4 * 9, _sum9);
-                vst1_f16(outptr + 4 * 10, _suma);
-                vst1_f16(outptr + 4 * 11, _sumb);
-                outptr += 4 * 12;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const __fp16* pA = pAT;
-
-                float16x4_t _sum0;
-                float16x4_t _sum1;
-                float16x4_t _sum2;
-                float16x4_t _sum3;
-                float16x4_t _sum4;
-                float16x4_t _sum5;
-                float16x4_t _sum6;
-                float16x4_t _sum7;
-
-                if (k == 0)
-                {
-                    _sum0 = vdup_n_f16(0.f);
-                    _sum1 = vdup_n_f16(0.f);
-                    _sum2 = vdup_n_f16(0.f);
-                    _sum3 = vdup_n_f16(0.f);
-                    _sum4 = vdup_n_f16(0.f);
-                    _sum5 = vdup_n_f16(0.f);
-                    _sum6 = vdup_n_f16(0.f);
-                    _sum7 = vdup_n_f16(0.f);
-                }
-                else
-                {
-                    _sum0 = vld1_f16(outptr);
-                    _sum1 = vld1_f16(outptr + 4);
-                    _sum2 = vld1_f16(outptr + 8);
-                    _sum3 = vld1_f16(outptr + 12);
-                    _sum4 = vld1_f16(outptr + 16);
-                    _sum5 = vld1_f16(outptr + 20);
-                    _sum6 = vld1_f16(outptr + 24);
-                    _sum7 = vld1_f16(outptr + 28);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x4_t _pA = vld1_f16(pA);
-                    float16x8_t _pB = vld1q_f16(pB);
-                    _sum0 = vfma_laneq_f16(_sum0, _pA, _pB, 0);
-                    _sum1 = vfma_laneq_f16(_sum1, _pA, _pB, 1);
-                    _sum2 = vfma_laneq_f16(_sum2, _pA, _pB, 2);
-                    _sum3 = vfma_laneq_f16(_sum3, _pA, _pB, 3);
-                    _sum4 = vfma_laneq_f16(_sum4, _pA, _pB, 4);
-                    _sum5 = vfma_laneq_f16(_sum5, _pA, _pB, 5);
-                    _sum6 = vfma_laneq_f16(_sum6, _pA, _pB, 6);
-                    _sum7 = vfma_laneq_f16(_sum7, _pA, _pB, 7);
-
-                    pA += 4;
-                    pB += 8;
-                }
-
-                vst1_f16(outptr, _sum0);
-                vst1_f16(outptr + 4, _sum1);
-                vst1_f16(outptr + 4 * 2, _sum2);
-                vst1_f16(outptr + 4 * 3, _sum3);
-                vst1_f16(outptr + 4 * 4, _sum4);
-                vst1_f16(outptr + 4 * 5, _sum5);
-                vst1_f16(outptr + 4 * 6, _sum6);
-                vst1_f16(outptr + 4 * 7, _sum7);
-                outptr += 4 * 8;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const __fp16* pA = pAT;
-
-                float16x4_t _sum0;
-                float16x4_t _sum1;
-                float16x4_t _sum2;
-                float16x4_t _sum3;
-
-                if (k == 0)
-                {
-                    _sum0 = vdup_n_f16(0.f);
-                    _sum1 = vdup_n_f16(0.f);
-                    _sum2 = vdup_n_f16(0.f);
-                    _sum3 = vdup_n_f16(0.f);
-                }
-                else
-                {
-                    _sum0 = vld1_f16(outptr);
-                    _sum1 = vld1_f16(outptr + 4);
-                    _sum2 = vld1_f16(outptr + 8);
-                    _sum3 = vld1_f16(outptr + 12);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x4_t _pA = vld1_f16(pA);
-                    float16x4_t _pB = vld1_f16(pB);
-                    _sum0 = vfma_lane_f16(_sum0, _pA, _pB, 0);
-                    _sum1 = vfma_lane_f16(_sum1, _pA, _pB, 1);
-                    _sum2 = vfma_lane_f16(_sum2, _pA, _pB, 2);
-                    _sum3 = vfma_lane_f16(_sum3, _pA, _pB, 3);
-
-                    pA += 4;
-                    pB += 4;
-                }
-
-                vst1_f16(outptr, _sum0);
-                vst1_f16(outptr + 4, _sum1);
-                vst1_f16(outptr + 4 * 2, _sum2);
-                vst1_f16(outptr + 4 * 3, _sum3);
-                outptr += 4 * 4;
-            }
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const __fp16* pA = pAT;
-
-                float16x4_t _sum0;
-                float16x4_t _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = vdup_n_f16(0.f);
-                    _sum1 = vdup_n_f16(0.f);
-                }
-                else
-                {
-                    _sum0 = vld1_f16(outptr);
-                    _sum1 = vld1_f16(outptr + 4);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x4_t _pA = vld1_f16(pA);
-                    _sum0 = vfma_n_f16(_sum0, _pA, pB[0]);
-                    _sum1 = vfma_n_f16(_sum1, _pA, pB[1]);
-
-                    pA += 4;
-                    pB += 2;
-                }
-
-                vst1_f16(outptr, _sum0);
-                vst1_f16(outptr + 4, _sum1);
-                outptr += 4 * 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const __fp16* pA = pAT;
-
-                float16x4_t _sum;
-
-                if (k == 0)
-                {
-                    _sum = vdup_n_f16(0.f);
-                }
-                else
-                {
-                    _sum = vld1_f16(outptr);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x4_t _pA = vld1_f16(pA);
-                    _sum = vfma_n_f16(_sum, _pA, pB[0]);
-
-                    pA += 4;
-                    pB += 1;
-                }
-
-                vst1_f16(outptr, _sum);
-                outptr += 4;
-            }
-        }
-    }
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const __fp16* pAT = AT_tile.row<const __fp16>(b) + max_kk * ii;
-            const __fp16* pB = BT_tile.row<const __fp16>(b);
-
-            int jj = 0;
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const __fp16* pA = pAT;
-
-                float16x8_t _sum01;
-                float16x4_t _sum2;
-                float16x8_t _sum34;
-                float16x4_t _sum5;
-
-                if (k == 0)
-                {
-                    _sum01 = vdupq_n_f16(0.f);
-                    _sum2 = vdup_n_f16(0.f);
-                    _sum34 = vdupq_n_f16(0.f);
-                    _sum5 = vdup_n_f16(0.f);
-                }
-                else
-                {
-                    float16x8x2_t _tmp0123 = vld2q_f16(outptr);
-                    float16x4x2_t _tmp45 = vld2_f16(outptr + 16);
-                    _sum01 = _tmp0123.val[0];
-                    _sum2 = _tmp45.val[0];
-                    _sum34 = _tmp0123.val[1];
-                    _sum5 = _tmp45.val[1];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x8_t _pB0 = vld1q_f16(pB);
-                    float16x4_t _pB2 = vld1_f16(pB + 8);
-                    _sum01 = vfmaq_n_f16(_sum01, _pB0, pA[0]);
-                    _sum2 = vfma_n_f16(_sum2, _pB2, pA[0]);
-                    _sum34 = vfmaq_n_f16(_sum34, _pB0, pA[1]);
-                    _sum5 = vfma_n_f16(_sum5, _pB2, pA[1]);
-                    pA += 2;
-                    pB += 12;
-                }
-
-                float16x8x2_t _tmp0123;
-                _tmp0123.val[0] = _sum01;
-                _tmp0123.val[1] = _sum34;
-                float16x4x2_t _tmp45;
-                _tmp45.val[0] = _sum2;
-                _tmp45.val[1] = _sum5;
-                vst2q_f16(outptr, _tmp0123);
-                vst2_f16(outptr + 16, _tmp45);
-                outptr += 2 * 12;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const __fp16* pA = pAT;
-
-                float16x8_t _sum01;
-                float16x8_t _sum23;
-
-                if (k == 0)
-                {
-                    _sum01 = vdupq_n_f16(0.f);
-                    _sum23 = vdupq_n_f16(0.f);
-                }
-                else
-                {
-                    float16x8x2_t _tmp0123 = vld2q_f16(outptr);
-                    _sum01 = _tmp0123.val[0];
-                    _sum23 = _tmp0123.val[1];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x8_t _pB = vld1q_f16(pB);
-                    _sum01 = vfmaq_n_f16(_sum01, _pB, pA[0]);
-                    _sum23 = vfmaq_n_f16(_sum23, _pB, pA[1]);
-                    pA += 2;
-                    pB += 8;
-                }
-
-                float16x8x2_t _tmp0123;
-                _tmp0123.val[0] = _sum01;
-                _tmp0123.val[1] = _sum23;
-                vst2q_f16(outptr, _tmp0123);
-                outptr += 2 * 8;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const __fp16* pA = pAT;
-
-                float16x4_t _sum0;
-                float16x4_t _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = vdup_n_f16(0.f);
-                    _sum1 = vdup_n_f16(0.f);
-                }
-                else
-                {
-                    float16x4x2_t _tmp01 = vld2_f16(outptr);
-                    _sum0 = _tmp01.val[0];
-                    _sum1 = _tmp01.val[1];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x4_t _pB = vld1_f16(pB);
-                    _sum0 = vfma_n_f16(_sum0, _pB, pA[0]);
-                    _sum1 = vfma_n_f16(_sum1, _pB, pA[1]);
-                    pA += 2;
-                    pB += 4;
-                }
-
-                float16x4x2_t _tmp01;
-                _tmp01.val[0] = _sum0;
-                _tmp01.val[1] = _sum1;
-                vst2_f16(outptr, _tmp01);
-                outptr += 2 * 4;
-            }
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const __fp16* pA = pAT;
-
-                __fp16 sum00 = 0.f;
-                __fp16 sum01 = 0.f;
-                __fp16 sum10 = 0.f;
-                __fp16 sum11 = 0.f;
-
-                if (k == 0)
-                {
-                    sum00 = 0.f;
-                    sum01 = 0.f;
-                    sum10 = 0.f;
-                    sum11 = 0.f;
-                }
-                else
-                {
-                    sum00 = outptr[0];
-                    sum01 = outptr[1];
-                    sum10 = outptr[2];
-                    sum11 = outptr[3];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    sum00 += pA[0] * pB[0];
-                    sum01 += pA[1] * pB[0];
-                    sum10 += pA[0] * pB[1];
-                    sum11 += pA[1] * pB[1];
-                    pA += 2;
-                    pB += 2;
-                }
-
-                outptr[0] = sum00;
-                outptr[1] = sum01;
-                outptr[2] = sum10;
-                outptr[3] = sum11;
-                outptr += 2 * 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const __fp16* pA = pAT;
-
-                __fp16 sum0 = 0.f;
-                __fp16 sum1 = 0.f;
-
-                if (k == 0)
-                {
-                    sum0 = 0.f;
-                    sum1 = 0.f;
-                }
-                else
-                {
-                    sum0 = outptr[0];
-                    sum1 = outptr[1];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    sum0 += pA[0] * pB[0];
-                    sum1 += pA[1] * pB[0];
-                    pA += 2;
-                    pB += 1;
-                }
-
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-                outptr += 2;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const __fp16* pAT = AT_tile.row<const __fp16>(b) + max_kk * ii;
-            const __fp16* pB = BT_tile.row<const __fp16>(b);
-
-            int jj = 0;
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const __fp16* pA = pAT;
-
-                float16x8_t _sum01;
-                float16x4_t _sum2;
-
-                if (k == 0)
-                {
-                    _sum01 = vdupq_n_f16(0.f);
-                    _sum2 = vdup_n_f16(0.f);
-                }
-                else
-                {
-                    _sum01 = vld1q_f16(outptr);
-                    _sum2 = vld1_f16(outptr + 8);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x8_t _pB0 = vld1q_f16(pB);
-                    float16x4_t _pB2 = vld1_f16(pB + 8);
-                    _sum01 = vfmaq_n_f16(_sum01, _pB0, pA[0]);
-                    _sum2 = vfma_n_f16(_sum2, _pB2, pA[0]);
-                    pA += 1;
-                    pB += 12;
-                }
-
-                vst1q_f16(outptr, _sum01);
-                vst1_f16(outptr + 8, _sum2);
-                outptr += 12;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const __fp16* pA = pAT;
-
-                float16x8_t _sum01;
-
-                if (k == 0)
-                {
-                    _sum01 = vdupq_n_f16(0.f);
-                }
-                else
-                {
-                    _sum01 = vld1q_f16(outptr);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x8_t _pB = vld1q_f16(pB);
-                    _sum01 = vfmaq_n_f16(_sum01, _pB, pA[0]);
-                    pA += 1;
-                    pB += 8;
-                }
-
-                vst1q_f16(outptr, _sum01);
-                outptr += 8;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const __fp16* pA = pAT;
-
-                float16x4_t _sum;
-
-                if (k == 0)
-                {
-                    _sum = vdup_n_f16(0.f);
-                }
-                else
-                {
-                    _sum = vld1_f16(outptr);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    float16x4_t _pB = vld1_f16(pB);
-                    _sum = vfma_n_f16(_sum, _pB, pA[0]);
-                    pA += 1;
-                    pB += 4;
-                }
-
-                vst1_f16(outptr, _sum);
-                outptr += 4;
-            }
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const __fp16* pA = pAT;
-
-                __fp16 sum0 = 0.f;
-                __fp16 sum1 = 0.f;
-
-                if (k == 0)
-                {
-                    sum0 = 0.f;
-                    sum1 = 0.f;
-                }
-                else
-                {
-                    sum0 = outptr[0];
-                    sum1 = outptr[1];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    sum0 += pA[0] * pB[0];
-                    sum1 += pA[0] * pB[1];
-                    pA += 1;
-                    pB += 2;
-                }
-
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-                outptr += 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const __fp16* pA = pAT;
-
-                __fp16 sum = 0.f;
-
-                if (k == 0)
-                {
-                    sum = 0.f;
-                }
-                else
-                {
-                    sum = outptr[0];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    sum += pA[0] * pB[0];
-                    pA += 1;
-                    pB += 1;
-                }
-
-                outptr[0] = sum;
-                outptr += 1;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd_get_optimal_tile_mnk_fp16(int M, int N, int K, int B, int& TILE_M, int& TILE_N, int& TILE_K, int nT)
-{
-    // resolve optimal tile size from cache size
-    const int l2_cache_size_fp16 = (int)(get_cpu_level2_cache_size() / sizeof(unsigned short));
-
-    if (nT == 0)
-        nT = get_physical_big_cpu_count();
-
-    // we shall take B into account for batched gemm, but that will be slower on arm in practice, why ?
-    (void)B;
-
-    // solve K
-    {
-        // try not to split K
-        int tile_size = (l2_cache_size_fp16 - 32) / 12;
-
-        TILE_K = std::max(8, tile_size / 8 * 8);
-
-        int nn_K = (K + TILE_K - 1) / TILE_K;
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 7) / 8 * 8);
-    }
-
-    // solve M
-    {
-        TILE_M = 8;
-    }
-
-    {
-        TILE_M *= std::min(nT, get_physical_cpu_count());
-
-        int nn_M = (M + TILE_M - 1) / TILE_M;
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-
-        if (nT > 1)
-        {
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 7) / 8 * 8);
-        }
-    }
-
-    if (N > 0)
-    {
-        int tile_size;
-        if (TILE_K >= K)
-        {
-            tile_size = (l2_cache_size_fp16 - TILE_M * TILE_K) / TILE_K;
-        }
-        else
-        {
-            tile_size = (l2_cache_size_fp16 - TILE_M * TILE_K) / (TILE_M + TILE_K);
-        }
-
-        TILE_N = std::max(4, tile_size / 4 * 4);
-
-        int nn_N = (N + TILE_N - 1) / TILE_N;
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_kernel_tile_fp16sa(const Mat& kernel, Mat& A, int inch, int i, int max_ii, int k, int max_kk)
-{
-    // const float ktm[4][3] = {
-    //     {1.0f, 0.0f, 0.0f},
-    //     {1.0f / 2, 1.0f / 2, 1.0f / 2},
-    //     {1.0f / 2, -1.0f / 2, 1.0f / 2},
-    //     {0.0f, 0.0f, 1.0f}
-    // };
-
-    __fp16* ptmp = A;
-
-    int ii = 0;
-    for (; ii < max_ii; ii++)
-    {
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            float tmp[4][3];
-
-            const float* k0 = (const float*)kernel + (i + ii) * inch * 9 + (k + kk) * 9;
-
-            for (int m = 0; m < 3; m++)
-            {
-                float r0 = k0[0];
-                float r1 = k0[1];
-                float r2 = k0[2];
-
-                tmp[0][m] = r0;
-                tmp[1][m] = r0 * 0.5f + r1 * 0.5f + r2 * 0.5f;
-                tmp[2][m] = r0 * 0.5f - r1 * 0.5f + r2 * 0.5f;
-                tmp[3][m] = r2;
-
-                k0 += 3;
-            }
-
-            for (int m = 0; m < 4; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-
-                float z0 = r0;
-                float z1 = r0 * 0.5f + r1 * 0.5f + r2 * 0.5f;
-                float z2 = r0 * 0.5f - r1 * 0.5f + r2 * 0.5f;
-                float z3 = r2;
-
-                ptmp[0] = (__fp16)z0;
-                ptmp[1] = (__fp16)z1;
-                ptmp[2] = (__fp16)z2;
-                ptmp[3] = (__fp16)z3;
-                ptmp += 4;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_kernel_fp16sa(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt)
-{
-    const int M = outch;
-    const int K = inch;
-    const int B = 16;
-
-    int TILE_M, TILE_N, TILE_K;
-    conv3x3s1_winograd_get_optimal_tile_mnk_fp16(M, 0, K, B, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    Mat A_tileX(B * TILE_M * TILE_K, 1, opt.num_threads, (size_t)2u);
-
-    AT.create(TILE_K * TILE_M, B, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M, (size_t)2u);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat A_tile = A_tileX.channel(get_omp_thread_num());
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_ii = std::min((M - i), TILE_M);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            conv3x3s1_winograd23_transform_kernel_tile_fp16sa(kernel, A_tile, inch, i, max_ii, k, max_kk);
-
-            Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-            conv3x3s1_winograd_pack_A_tile_fp16(A_tile, AT_tile, B, max_ii, max_kk);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_input_tile_fp16sa(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    // const float itm[4][4] = {
-    //     {1.0f,  0.0f, -1.0f,  0.0f},
-    //     {0.0f,  1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  0.00f, 1.0f}
-    // };
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w - 1) / 2;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        __fp16 tmp[4][4][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = bottom_blob.channel((k + kk) / elempack).row<const __fp16>(ti * 2) + (tj * 2) * elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float16x8_t _r0 = vdupq_n_f16(0.f);
-                float16x8_t _r1 = vdupq_n_f16(0.f);
-                float16x8_t _r2 = vdupq_n_f16(0.f);
-                float16x8_t _r3 = vdupq_n_f16(0.f);
-
-                if (ti * 2 + m < h)
-                {
-                    if (elempack == 8)
-                    {
-                        _r0 = vld1q_f16(r0);
-                        if (tj * 2 + 1 < w) _r1 = vld1q_f16(r0 + 8);
-                        if (tj * 2 + 2 < w) _r2 = vld1q_f16(r0 + 16);
-                        if (tj * 2 + 3 < w) _r3 = vld1q_f16(r0 + 24);
-                    }
-                    if (elempack == 4)
-                    {
-                        const __fp16* r1 = r0 + N;
-
-                        _r0 = vcombine_f16(vld1_f16(r0), vld1_f16(r1));
-                        if (tj * 2 + 1 < w)
-                        {
-                            _r1 = vcombine_f16(vld1_f16(r0 + 4), vld1_f16(r1 + 4));
-                        }
-                        if (tj * 2 + 2 < w)
-                        {
-                            _r2 = vcombine_f16(vld1_f16(r0 + 8), vld1_f16(r1 + 8));
-                        }
-                        if (tj * 2 + 3 < w)
-                        {
-                            _r3 = vcombine_f16(vld1_f16(r0 + 12), vld1_f16(r1 + 12));
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        const __fp16* r1 = r0 + N;
-                        const __fp16* r2 = r0 + N * 2;
-                        const __fp16* r3 = r0 + N * 3;
-                        const __fp16* r4 = r0 + N * 4;
-                        const __fp16* r5 = r0 + N * 5;
-                        const __fp16* r6 = r0 + N * 6;
-                        const __fp16* r7 = r0 + N * 7;
-
-                        float16x4_t _t0 = vld1_f16(r0);
-                        float16x4_t _t1 = vld1_f16(r1);
-                        float16x4_t _t2 = vld1_f16(r2);
-                        float16x4_t _t3 = vld1_f16(r3);
-                        float16x4_t _t4 = vld1_f16(r4);
-                        float16x4_t _t5 = vld1_f16(r5);
-                        float16x4_t _t6 = vld1_f16(r6);
-                        float16x4_t _t7 = vld1_f16(r7);
-
-                        transpose4x4_ph(_t0, _t1, _t2, _t3);
-                        transpose4x4_ph(_t4, _t5, _t6, _t7);
-
-                        _r0 = vcombine_f16(_t0, _t4);
-                        if (tj * 2 + 1 < w)
-                        {
-                            _r1 = vcombine_f16(_t1, _t5);
-                        }
-                        if (tj * 2 + 2 < w)
-                        {
-                            _r2 = vcombine_f16(_t2, _t6);
-                        }
-                        if (tj * 2 + 3 < w)
-                        {
-                            _r3 = vcombine_f16(_t3, _t7);
-                        }
-                    }
-                }
-
-                float16x8_t _tmp0 = vsubq_f16(_r0, _r2);
-                float16x8_t _tmp1 = vaddq_f16(_r1, _r2);
-                float16x8_t _tmp2 = vsubq_f16(_r2, _r1);
-                float16x8_t _tmp3 = vsubq_f16(_r3, _r1);
-
-                vst1q_f16(tmp[0][m], _tmp0);
-                vst1q_f16(tmp[1][m], _tmp1);
-                vst1q_f16(tmp[2][m], _tmp2);
-                vst1q_f16(tmp[3][m], _tmp3);
-
-                r0 += w * elempack;
-            }
-
-            __fp16* p0 = (__fp16*)B + kk * max_jj * 16 + jj * 8;
-            __fp16* p1 = p0 + max_jj * 8;
-            __fp16* p2 = p0 + max_jj * 8 * 2;
-            __fp16* p3 = p0 + max_jj * 8 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float16x8_t _r0 = vld1q_f16(tmp[m][0]);
-                float16x8_t _r1 = vld1q_f16(tmp[m][1]);
-                float16x8_t _r2 = vld1q_f16(tmp[m][2]);
-                float16x8_t _r3 = vld1q_f16(tmp[m][3]);
-
-                float16x8_t _tmp0 = vsubq_f16(_r0, _r2);
-                float16x8_t _tmp1 = vaddq_f16(_r1, _r2);
-                float16x8_t _tmp2 = vsubq_f16(_r2, _r1);
-                float16x8_t _tmp3 = vsubq_f16(_r3, _r1);
-
-                vst1q_f16(p0, _tmp0);
-                vst1q_f16(p1, _tmp1);
-                vst1q_f16(p2, _tmp2);
-                vst1q_f16(p3, _tmp3);
-
-                p0 += max_jj * 4 * 8;
-                p1 += max_jj * 4 * 8;
-                p2 += max_jj * 4 * 8;
-                p3 += max_jj * 4 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 4;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        __fp16 tmp[4][4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = bottom_blob.channel((k + kk) / elempack).row<const __fp16>(ti * 2) + (tj * 2) * elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float16x4_t _r0 = vdup_n_f16(0.f);
-                float16x4_t _r1 = vdup_n_f16(0.f);
-                float16x4_t _r2 = vdup_n_f16(0.f);
-                float16x4_t _r3 = vdup_n_f16(0.f);
-
-                if (ti * 2 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1_f16(r0);
-                        if (tj * 2 + 1 < w) _r1 = vld1_f16(r0 + 4);
-                        if (tj * 2 + 2 < w) _r2 = vld1_f16(r0 + 8);
-                        if (tj * 2 + 3 < w) _r3 = vld1_f16(r0 + 12);
-                    }
-                    if (elempack == 1)
-                    {
-                        const __fp16* r1 = r0 + N;
-                        const __fp16* r2 = r0 + N * 2;
-                        const __fp16* r3 = r0 + N * 3;
-
-                        float16x4_t _t0 = vld1_f16(r0);
-                        float16x4_t _t1 = vld1_f16(r1);
-                        float16x4_t _t2 = vld1_f16(r2);
-                        float16x4_t _t3 = vld1_f16(r3);
-
-                        transpose4x4_ph(_t0, _t1, _t2, _t3);
-
-                        _r0 = _t0;
-                        if (tj * 2 + 1 < w) _r1 = _t1;
-                        if (tj * 2 + 2 < w) _r2 = _t2;
-                        if (tj * 2 + 3 < w) _r3 = _t3;
-                    }
-                }
-
-                float16x4_t _tmp0 = vsub_f16(_r0, _r2);
-                float16x4_t _tmp1 = vadd_f16(_r1, _r2);
-                float16x4_t _tmp2 = vsub_f16(_r2, _r1);
-                float16x4_t _tmp3 = vsub_f16(_r3, _r1);
-
-                vst1_f16(tmp[0][m], _tmp0);
-                vst1_f16(tmp[1][m], _tmp1);
-                vst1_f16(tmp[2][m], _tmp2);
-                vst1_f16(tmp[3][m], _tmp3);
-
-                r0 += w * elempack;
-            }
-
-            __fp16* p0 = (__fp16*)B + kk * max_jj * 16 + jj * 4;
-            __fp16* p1 = p0 + max_jj * 4;
-            __fp16* p2 = p0 + max_jj * 4 * 2;
-            __fp16* p3 = p0 + max_jj * 4 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float16x4_t _r0 = vld1_f16(tmp[m][0]);
-                float16x4_t _r1 = vld1_f16(tmp[m][1]);
-                float16x4_t _r2 = vld1_f16(tmp[m][2]);
-                float16x4_t _r3 = vld1_f16(tmp[m][3]);
-
-                float16x4_t _tmp0 = vsub_f16(_r0, _r2);
-                float16x4_t _tmp1 = vadd_f16(_r1, _r2);
-                float16x4_t _tmp2 = vsub_f16(_r2, _r1);
-                float16x4_t _tmp3 = vsub_f16(_r3, _r1);
-
-                vst1_f16(p0, _tmp0);
-                vst1_f16(p1, _tmp1);
-                vst1_f16(p2, _tmp2);
-                vst1_f16(p3, _tmp3);
-
-                p0 += max_jj * 4 * 4;
-                p1 += max_jj * 4 * 4;
-                p2 += max_jj * 4 * 4;
-                p3 += max_jj * 4 * 4;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 4;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        __fp16 tmp[4][4][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = bottom_blob.channel(k + kk).row<const __fp16>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 4; m++)
-            {
-                __fp16 r00 = 0.f;
-                __fp16 r01 = 0.f;
-                __fp16 r10 = 0.f;
-                __fp16 r11 = 0.f;
-                __fp16 r20 = 0.f;
-                __fp16 r21 = 0.f;
-                __fp16 r30 = 0.f;
-                __fp16 r31 = 0.f;
-
-                if (ti * 2 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const __fp16* r1 = r0 + N;
-
-                        r00 = r0[0];
-                        r01 = r1[0];
-                        if (tj * 2 + 1 < w)
-                        {
-                            r10 = r0[1];
-                            r11 = r1[1];
-                        }
-                        if (tj * 2 + 2 < w)
-                        {
-                            r20 = r0[2];
-                            r21 = r1[2];
-                        }
-                        if (tj * 2 + 3 < w)
-                        {
-                            r30 = r0[3];
-                            r31 = r1[3];
-                        }
-                    }
-                }
-
-                tmp[0][m][0] = r00 - r20;
-                tmp[0][m][1] = r01 - r21;
-                tmp[1][m][0] = r10 + r20;
-                tmp[1][m][1] = r11 + r21;
-                tmp[2][m][0] = r20 - r10;
-                tmp[2][m][1] = r21 - r11;
-                tmp[3][m][0] = r30 - r10;
-                tmp[3][m][1] = r31 - r11;
-
-                r0 += w;
-            }
-
-            __fp16* p0 = (__fp16*)B + kk * max_jj * 16 + jj * 2;
-            __fp16* p1 = p0 + max_jj * 2;
-            __fp16* p2 = p0 + max_jj * 2 * 2;
-            __fp16* p3 = p0 + max_jj * 2 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                __fp16 r00 = tmp[m][0][0];
-                __fp16 r01 = tmp[m][0][1];
-                __fp16 r10 = tmp[m][1][0];
-                __fp16 r11 = tmp[m][1][1];
-                __fp16 r20 = tmp[m][2][0];
-                __fp16 r21 = tmp[m][2][1];
-                __fp16 r30 = tmp[m][3][0];
-                __fp16 r31 = tmp[m][3][1];
-
-                p0[0] = r00 - r20;
-                p0[1] = r01 - r21;
-                p1[0] = r10 + r20;
-                p1[1] = r11 + r21;
-                p2[0] = r20 - r10;
-                p2[1] = r21 - r11;
-                p3[0] = r30 - r10;
-                p3[1] = r31 - r11;
-
-                p0 += max_jj * 4 * 2;
-                p1 += max_jj * 4 * 2;
-                p2 += max_jj * 4 * 2;
-                p3 += max_jj * 4 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        __fp16 tmp[4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0123 = bottom_blob.channel(k + kk).row<const __fp16>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 4; m++)
-            {
-                __fp16 r0 = 0.f;
-                __fp16 r1 = 0.f;
-                __fp16 r2 = 0.f;
-                __fp16 r3 = 0.f;
-
-                if (ti * 2 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = r0123[0];
-                        if (tj * 2 + 1 < w) r1 = r0123[1];
-                        if (tj * 2 + 2 < w) r2 = r0123[2];
-                        if (tj * 2 + 3 < w) r3 = r0123[3];
-                    }
-                }
-
-                tmp[0][m] = r0 - r2;
-                tmp[1][m] = r1 + r2;
-                tmp[2][m] = r2 - r1;
-                tmp[3][m] = r3 - r1;
-
-                r0123 += w;
-            }
-
-            __fp16* p0 = (__fp16*)B + kk * max_jj * 16 + jj;
-            __fp16* p1 = p0 + max_jj;
-            __fp16* p2 = p0 + max_jj * 2;
-            __fp16* p3 = p0 + max_jj * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                __fp16 r0 = tmp[m][0];
-                __fp16 r1 = tmp[m][1];
-                __fp16 r2 = tmp[m][2];
-                __fp16 r3 = tmp[m][3];
-
-                p0[0] = r0 - r2;
-                p1[0] = r1 + r2;
-                p2[0] = r2 - r1;
-                p3[0] = r3 - r1;
-
-                p0 += max_jj * 4;
-                p1 += max_jj * 4;
-                p2 += max_jj * 4;
-                p3 += max_jj * 4;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_output_tile_fp16sa(const Mat& top_tile, Mat& top_blob, const Mat& bias, int i, int max_ii, int j, int max_jj)
-{
-    // const float otm[2][4] = {
-    //     {1.0f,  1.0f,  1.0f,  0.0f},
-    //     {0.0f,  1.0f, -1.0f,  1.0f}
-    // };
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 1) / 2;
-
-    const __fp16* biasptr = bias;
-
-    int ii = 0;
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        float16x8_t _bias0 = biasptr ? vld1q_f16(biasptr + i + ii) : vdupq_n_f16(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        __fp16 tmp[2][4][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = (const __fp16*)top_tile + ii * max_jj * 16 + jj * 8;
-            const __fp16* r1 = r0 + max_jj * 8;
-            const __fp16* r2 = r0 + max_jj * 8 * 2;
-            const __fp16* r3 = r0 + max_jj * 8 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float16x8_t _r0 = vld1q_f16(r0);
-                float16x8_t _r1 = vld1q_f16(r1);
-                float16x8_t _r2 = vld1q_f16(r2);
-                float16x8_t _r3 = vld1q_f16(r3);
-
-                float16x8_t _tmp0 = vaddq_f16(vaddq_f16(_r0, _r1), _r2);
-                float16x8_t _tmp1 = vaddq_f16(vsubq_f16(_r1, _r2), _r3);
-
-                vst1q_f16(tmp[0][m], _tmp0);
-                vst1q_f16(tmp[1][m], _tmp1);
-
-                r0 += max_jj * 4 * 8;
-                r1 += max_jj * 4 * 8;
-                r2 += max_jj * 4 * 8;
-                r3 += max_jj * 4 * 8;
-            }
-
-            __fp16* outptr0 = top_blob.channel((i + ii) / out_elempack).row<__fp16>(ti * 2) + (tj * 2) * out_elempack;
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                float16x8_t _r0 = vld1q_f16(tmp[m][0]);
-                float16x8_t _r1 = vld1q_f16(tmp[m][1]);
-                float16x8_t _r2 = vld1q_f16(tmp[m][2]);
-                float16x8_t _r3 = vld1q_f16(tmp[m][3]);
-
-                float16x8_t _tmp0 = vaddq_f16(_bias0, vaddq_f16(vaddq_f16(_r0, _r1), _r2));
-                float16x8_t _tmp1 = vaddq_f16(_bias0, vaddq_f16(vsubq_f16(_r1, _r2), _r3));
-
-                if (out_elempack == 8)
-                {
-                    vst1q_f16(outptr0, _tmp0);
-                    if (tj * 2 + 1 < outw)
-                    {
-                        vst1q_f16(outptr0 + 8, _tmp1);
-                    }
-                }
-                if (out_elempack == 4)
-                {
-                    __fp16* outptr1 = outptr0 + N;
-
-                    vst1_f16(outptr0, vget_low_f16(_tmp0));
-                    vst1_f16(outptr1, vget_high_f16(_tmp0));
-                    if (tj * 2 + 1 < outw)
-                    {
-                        vst1_f16(outptr0 + 4, vget_low_f16(_tmp1));
-                        vst1_f16(outptr1 + 4, vget_high_f16(_tmp1));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    __fp16 tmp0[8];
-                    __fp16 tmp1[8];
-                    vst1q_f16(tmp0, _tmp0);
-                    vst1q_f16(tmp1, _tmp1);
-
-                    __fp16* outptr1 = outptr0 + N;
-                    __fp16* outptr2 = outptr0 + N * 2;
-                    __fp16* outptr3 = outptr0 + N * 3;
-                    __fp16* outptr4 = outptr0 + N * 4;
-                    __fp16* outptr5 = outptr0 + N * 5;
-                    __fp16* outptr6 = outptr0 + N * 6;
-                    __fp16* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        float16x4_t _bias0 = biasptr ? vld1_f16(biasptr + i + ii) : vdup_n_f16(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        __fp16 tmp[2][4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = (const __fp16*)top_tile + ii * max_jj * 16 + jj * 4;
-            const __fp16* r1 = r0 + max_jj * 4;
-            const __fp16* r2 = r0 + max_jj * 4 * 2;
-            const __fp16* r3 = r0 + max_jj * 4 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float16x4_t _r0 = vld1_f16(r0);
-                float16x4_t _r1 = vld1_f16(r1);
-                float16x4_t _r2 = vld1_f16(r2);
-                float16x4_t _r3 = vld1_f16(r3);
-
-                float16x4_t _tmp0 = vadd_f16(vadd_f16(_r0, _r1), _r2);
-                float16x4_t _tmp1 = vadd_f16(vsub_f16(_r1, _r2), _r3);
-
-                vst1_f16(tmp[0][m], _tmp0);
-                vst1_f16(tmp[1][m], _tmp1);
-
-                r0 += max_jj * 4 * 4;
-                r1 += max_jj * 4 * 4;
-                r2 += max_jj * 4 * 4;
-                r3 += max_jj * 4 * 4;
-            }
-
-            __fp16* outptr0 = top_blob.channel((i + ii) / out_elempack).row<__fp16>(ti * 2) + (tj * 2) * out_elempack;
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                float16x4_t _r0 = vld1_f16(tmp[m][0]);
-                float16x4_t _r1 = vld1_f16(tmp[m][1]);
-                float16x4_t _r2 = vld1_f16(tmp[m][2]);
-                float16x4_t _r3 = vld1_f16(tmp[m][3]);
-
-                float16x4_t _tmp0 = vadd_f16(_bias0, vadd_f16(vadd_f16(_r0, _r1), _r2));
-                float16x4_t _tmp1 = vadd_f16(_bias0, vadd_f16(vsub_f16(_r1, _r2), _r3));
-
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr0, _tmp0);
-                    if (tj * 2 + 1 < outw) vst1_f16(outptr0 + 4, _tmp1);
-                }
-                if (out_elempack == 1)
-                {
-                    __fp16 tmp0[4];
-                    __fp16 tmp1[4];
-                    vst1_f16(tmp0, _tmp0);
-                    vst1_f16(tmp1, _tmp1);
-
-                    __fp16* outptr1 = outptr0 + N;
-                    __fp16* outptr2 = outptr0 + N * 2;
-                    __fp16* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        __fp16 bias0 = biasptr ? biasptr[i + ii] : 0.f;
-        __fp16 bias1 = biasptr ? biasptr[i + ii + 1] : 0.f;
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        __fp16 tmp[2][4][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = (const __fp16*)top_tile + ii * max_jj * 16 + jj * 2;
-            const __fp16* r1 = r0 + max_jj * 2;
-            const __fp16* r2 = r0 + max_jj * 2 * 2;
-            const __fp16* r3 = r0 + max_jj * 2 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                tmp[0][m][0] = r0[0] + r1[0] + r2[0];
-                tmp[0][m][1] = r0[1] + r1[1] + r2[1];
-                tmp[1][m][0] = r1[0] - r2[0] + r3[0];
-                tmp[1][m][1] = r1[1] - r2[1] + r3[1];
-
-                r0 += max_jj * 4 * 2;
-                r1 += max_jj * 4 * 2;
-                r2 += max_jj * 4 * 2;
-                r3 += max_jj * 4 * 2;
-            }
-
-            __fp16* outptr0 = top_blob.channel(i + ii).row<__fp16>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                __fp16 r00 = tmp[m][0][0];
-                __fp16 r01 = tmp[m][0][1];
-                __fp16 r10 = tmp[m][1][0];
-                __fp16 r11 = tmp[m][1][1];
-                __fp16 r20 = tmp[m][2][0];
-                __fp16 r21 = tmp[m][2][1];
-                __fp16 r30 = tmp[m][3][0];
-                __fp16 r31 = tmp[m][3][1];
-
-                __fp16 tmp00 = bias0 + r00 + r10 + r20;
-                __fp16 tmp01 = bias1 + r01 + r11 + r21;
-                __fp16 tmp10 = bias0 + r10 - r20 + r30;
-                __fp16 tmp11 = bias1 + r11 - r21 + r31;
-
-                // if (out_elempack == 1)
-                {
-                    __fp16* outptr1 = outptr0 + N;
-
-                    outptr0[0] = tmp00;
-                    outptr1[0] = tmp01;
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp10;
-                        outptr1[1] = tmp11;
-                    }
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        __fp16 bias0 = biasptr ? biasptr[i + ii] : 0.f;
-
-        __fp16 tmp[2][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = (const __fp16*)top_tile + ii * max_jj * 16 + jj;
-            const __fp16* r1 = r0 + max_jj;
-            const __fp16* r2 = r0 + max_jj * 2;
-            const __fp16* r3 = r0 + max_jj * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                tmp[0][m] = r0[0] + r1[0] + r2[0];
-                tmp[1][m] = r1[0] - r2[0] + r3[0];
-
-                r0 += max_jj * 4;
-                r1 += max_jj * 4;
-                r2 += max_jj * 4;
-                r3 += max_jj * 4;
-            }
-
-            __fp16* outptr0 = top_blob.channel(i + ii).row<__fp16>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                __fp16 r0 = tmp[m][0];
-                __fp16 r1 = tmp[m][1];
-                __fp16 r2 = tmp[m][2];
-                __fp16 r3 = tmp[m][3];
-
-                __fp16 tmp0 = bias0 + r0 + r1 + r2;
-                __fp16 tmp1 = bias0 + r1 - r2 + r3;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = tmp0;
-                    if (tj * 2 + 1 < outw) outptr0[1] = tmp1;
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, const Mat& bias, int nT, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 2n+2, winograd F(2,3)
-    int w_tiles = (outw + 1) / 2;
-    int h_tiles = (outh + 1) / 2;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 16;
-
-    // NCNN_LOGE("conv3x3s1_winograd23_fp16sa %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    conv3x3s1_winograd_get_optimal_tile_mnk_fp16(M, N, K, B, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 2u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 2u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd23_transform_input_tile_fp16sa(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile_fp16(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 2u, opt.workspace_allocator);
-
-        #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd23_transform_input_tile_fp16sa(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile_fp16(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 2u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                conv3x3s1_winograd_gemm_transB_packed_tile_fp16sa(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk, opt.use_a53_a55_optimized_kernel);
-            }
-
-            // transform output
-            conv3x3s1_winograd23_transform_output_tile_fp16sa(top_tile, top_blob, bias, i, max_ii, j, max_jj);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_kernel_tile_fp16sa(const Mat& kernel, Mat& A, int inch, int i, int max_ii, int k, int max_kk)
-{
-    __fp16* ptmp = A;
-
-    int ii = 0;
-    for (; ii < max_ii; ii++)
-    {
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            const float sq2 = 1.41421356237f;
-            // const float ktm[6][3] = {
-            //     {1.0f, 0.0f, 0.0f},
-            //     {-2.0f / 3, -sq2 / 3, -1.0f / 3},
-            //     {-2.0f / 3, sq2 / 3, -1.0f / 3},
-            //     {1.0f / 6, sq2 / 6, 1.0f / 3},
-            //     {1.0f / 6, -sq2 / 6, 1.0f / 3},
-            //     {0.0f, 0.0f, 1.0f}
-            // };
-            const float ktm0 = 2.0f / 3;
-            const float ktm1 = sq2 / 3;
-            const float ktm2 = 1.0f / 3;
-            const float ktm3 = 1.0f / 6;
-            const float ktm4 = sq2 / 6;
-
-            float tmp[6][3];
-
-            const float* k0 = (const float*)kernel + (i + ii) * inch * 9 + (k + kk) * 9;
-
-            for (int m = 0; m < 3; m++)
-            {
-                float r0 = k0[0];
-                float r1 = k0[1];
-                float r2 = k0[2];
-
-                tmp[0][m] = r0;
-                tmp[1][m] = -r0 * ktm0 - r1 * ktm1 - r2 * ktm2;
-                tmp[2][m] = -r0 * ktm0 + r1 * ktm1 - r2 * ktm2;
-                tmp[3][m] = r0 * ktm3 + r1 * ktm4 + r2 * ktm2;
-                tmp[4][m] = r0 * ktm3 - r1 * ktm4 + r2 * ktm2;
-                tmp[5][m] = r2;
-
-                k0 += 3;
-            }
-
-            for (int m = 0; m < 6; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-
-                float z0 = r0;
-                float z1 = -r0 * ktm0 - r1 * ktm1 - r2 * ktm2;
-                float z2 = -r0 * ktm0 + r1 * ktm1 - r2 * ktm2;
-                float z3 = r0 * ktm3 + r1 * ktm4 + r2 * ktm2;
-                float z4 = r0 * ktm3 - r1 * ktm4 + r2 * ktm2;
-                float z5 = r2;
-
-                ptmp[0] = (__fp16)z0;
-                ptmp[1] = (__fp16)z1;
-                ptmp[2] = (__fp16)z2;
-                ptmp[3] = (__fp16)z3;
-                ptmp[4] = (__fp16)z4;
-                ptmp[5] = (__fp16)z5;
-                ptmp += 6;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_kernel_fp16sa(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt)
-{
-    const int M = outch;
-    const int K = inch;
-    const int B = 36;
-
-    int TILE_M, TILE_N, TILE_K;
-    conv3x3s1_winograd_get_optimal_tile_mnk_fp16(M, 0, K, B, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    Mat A_tileX(B * TILE_M * TILE_K, 1, opt.num_threads, (size_t)2u);
-
-    AT.create(TILE_K * TILE_M, B, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M, (size_t)2u);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat A_tile = A_tileX.channel(get_omp_thread_num());
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_ii = std::min((M - i), TILE_M);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            conv3x3s1_winograd43_transform_kernel_tile_fp16sa(kernel, A_tile, inch, i, max_ii, k, max_kk);
-
-            Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-            conv3x3s1_winograd_pack_A_tile_fp16(A_tile, AT_tile, B, max_ii, max_kk);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_input_tile_fp16sa(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    const __fp16 sq2 = 1.41421356237;
-    const __fp16 msq2_d2 = -1.41421356237 / 2;
-
-    // const float itm[6][6] = {
-    //     {1.0f,  0.0f,  -2.5f,  0.0f,  1.0f, 0.0f},
-    //     {0.0f, -sq2,   -2.0f,  sq2/2, 1.0f, 0.0f},
-    //     {0.0f,  sq2,   -2.0f, -sq2/2, 1.0f, 0.0f},
-    //     {0.0f, -sq2/2, -0.5f,  sq2,   1.0f, 0.0f},
-    //     {0.0f,  sq2/2, -0.5f, -sq2,   1.0f, 0.0f},
-    //     {0.0f,  1.0f,   0.0f,  -2.5f, 0.0f, 1.0f}
-    // };
-
-    // 0 =  r00 + r04 - 2.5f * r02
-    // 1 = -(sq2 * r01 - sq2_d2 * r03) + (r04 - 2 * r02)
-    // 2 =  (sq2 * r01 - sq2_d2 * r03) + (r04 - 2 * r02)
-    // 3 =  (sq2 * r03 - sq2_d2 * r01) + (r04 - 0.5f * r02)
-    // 4 = -(sq2 * r03 - sq2_d2 * r01) + (r04 - 0.5f * r02)
-    // 5 =  r01 + r05 - 2.5f * r03
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w + 1) / 4;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        __fp16 tmp[6][6][8];
-
-        const __fp16 coeffs[8] = {sq2, msq2_d2, -2.f, -0.5f, -2.5f, 0.f, 0.f, 0.f};
-        float16x8_t _coeffs = vld1q_f16(coeffs);
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = bottom_blob.channel((k + kk) / elempack).row<const __fp16>(ti * 4) + (tj * 4) * elempack;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float16x8_t _r0 = vdupq_n_f16(0.f);
-                float16x8_t _r1 = vdupq_n_f16(0.f);
-                float16x8_t _r2 = vdupq_n_f16(0.f);
-                float16x8_t _r3 = vdupq_n_f16(0.f);
-                float16x8_t _r4 = vdupq_n_f16(0.f);
-                float16x8_t _r5 = vdupq_n_f16(0.f);
-
-                if (ti * 4 + m < h)
-                {
-                    if (elempack == 8)
-                    {
-                        _r0 = vld1q_f16(r0);
-                        if (tj * 4 + 1 < w) _r1 = vld1q_f16(r0 + 8);
-                        if (tj * 4 + 2 < w) _r2 = vld1q_f16(r0 + 16);
-                        if (tj * 4 + 3 < w) _r3 = vld1q_f16(r0 + 24);
-                        if (tj * 4 + 4 < w) _r4 = vld1q_f16(r0 + 32);
-                        if (tj * 4 + 5 < w) _r5 = vld1q_f16(r0 + 40);
-                    }
-                    if (elempack == 4)
-                    {
-                        const __fp16* r1 = r0 + N;
-
-                        _r0 = vcombine_f16(vld1_f16(r0), vld1_f16(r1));
-                        if (tj * 4 + 1 < w)
-                        {
-                            _r1 = vcombine_f16(vld1_f16(r0 + 4), vld1_f16(r1 + 4));
-                        }
-                        if (tj * 4 + 2 < w)
-                        {
-                            _r2 = vcombine_f16(vld1_f16(r0 + 8), vld1_f16(r1 + 8));
-                        }
-                        if (tj * 4 + 3 < w)
-                        {
-                            _r3 = vcombine_f16(vld1_f16(r0 + 12), vld1_f16(r1 + 12));
-                        }
-                        if (tj * 4 + 4 < w)
-                        {
-                            _r4 = vcombine_f16(vld1_f16(r0 + 16), vld1_f16(r1 + 16));
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            _r5 = vcombine_f16(vld1_f16(r0 + 20), vld1_f16(r1 + 20));
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        const __fp16* r1 = r0 + N;
-                        const __fp16* r2 = r0 + N * 2;
-                        const __fp16* r3 = r0 + N * 3;
-                        const __fp16* r4 = r0 + N * 4;
-                        const __fp16* r5 = r0 + N * 5;
-                        const __fp16* r6 = r0 + N * 6;
-                        const __fp16* r7 = r0 + N * 7;
-
-                        float16x4_t _t0 = vld1_f16(r0);
-                        float16x4_t _t1 = vld1_f16(r1);
-                        float16x4_t _t2 = vld1_f16(r2);
-                        float16x4_t _t3 = vld1_f16(r3);
-                        float16x4_t _t4 = vld1_f16(r4);
-                        float16x4_t _t5 = vld1_f16(r5);
-                        float16x4_t _t6 = vld1_f16(r6);
-                        float16x4_t _t7 = vld1_f16(r7);
-
-                        transpose4x4_ph(_t0, _t1, _t2, _t3);
-                        transpose4x4_ph(_t4, _t5, _t6, _t7);
-
-                        _r0 = vcombine_f16(_t0, _t4);
-                        if (tj * 4 + 1 < w)
-                        {
-                            _r1 = vcombine_f16(_t1, _t5);
-                        }
-                        if (tj * 4 + 2 < w)
-                        {
-                            _r2 = vcombine_f16(_t2, _t6);
-                        }
-                        if (tj * 4 + 3 < w)
-                        {
-                            _r3 = vcombine_f16(_t3, _t7);
-                        }
-                        if (tj * 4 + 4 < w)
-                        {
-                            __fp16 tmp[8] = {r0[4], r1[4], r2[4], r3[4], r4[4], r5[4], r6[4], r7[4]};
-                            _r4 = vld1q_f16(tmp);
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            __fp16 tmp[8] = {r0[5], r1[5], r2[5], r3[5], r4[5], r5[5], r6[5], r7[5]};
-                            _r5 = vld1q_f16(tmp);
-                        }
-                    }
-                }
-
-                float16x8_t _tmp12a = vfmaq_laneq_f16(vmulq_laneq_f16(_r1, _coeffs, 0), _r3, _coeffs, 1);
-                float16x8_t _tmp12b = vfmaq_laneq_f16(_r4, _r2, _coeffs, 2);
-                float16x8_t _tmp34a = vfmaq_laneq_f16(vmulq_laneq_f16(_r3, _coeffs, 0), _r1, _coeffs, 1);
-                float16x8_t _tmp34b = vfmaq_laneq_f16(_r4, _r2, _coeffs, 3);
-
-                float16x8_t _tmp0 = vfmaq_laneq_f16(vaddq_f16(_r0, _r4), _r2, _coeffs, 4);
-                float16x8_t _tmp1 = vsubq_f16(_tmp12b, _tmp12a);
-                float16x8_t _tmp2 = vaddq_f16(_tmp12b, _tmp12a);
-                float16x8_t _tmp3 = vaddq_f16(_tmp34b, _tmp34a);
-                float16x8_t _tmp4 = vsubq_f16(_tmp34b, _tmp34a);
-                float16x8_t _tmp5 = vfmaq_laneq_f16(vaddq_f16(_r1, _r5), _r3, _coeffs, 4);
-
-                vst1q_f16(tmp[0][m], _tmp0);
-                vst1q_f16(tmp[1][m], _tmp1);
-                vst1q_f16(tmp[2][m], _tmp2);
-                vst1q_f16(tmp[3][m], _tmp3);
-                vst1q_f16(tmp[4][m], _tmp4);
-                vst1q_f16(tmp[5][m], _tmp5);
-
-                r0 += w * elempack;
-            }
-
-            __fp16* p0 = (__fp16*)B + kk * max_jj * 36 + jj * 8;
-            __fp16* p1 = p0 + max_jj * 8;
-            __fp16* p2 = p0 + max_jj * 8 * 2;
-            __fp16* p3 = p0 + max_jj * 8 * 3;
-            __fp16* p4 = p0 + max_jj * 8 * 4;
-            __fp16* p5 = p0 + max_jj * 8 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float16x8_t _r0 = vld1q_f16(tmp[m][0]);
-                float16x8_t _r1 = vld1q_f16(tmp[m][1]);
-                float16x8_t _r2 = vld1q_f16(tmp[m][2]);
-                float16x8_t _r3 = vld1q_f16(tmp[m][3]);
-                float16x8_t _r4 = vld1q_f16(tmp[m][4]);
-                float16x8_t _r5 = vld1q_f16(tmp[m][5]);
-
-                float16x8_t _tmp12a = vfmaq_laneq_f16(vmulq_laneq_f16(_r1, _coeffs, 0), _r3, _coeffs, 1);
-                float16x8_t _tmp12b = vfmaq_laneq_f16(_r4, _r2, _coeffs, 2);
-                float16x8_t _tmp34a = vfmaq_laneq_f16(vmulq_laneq_f16(_r3, _coeffs, 0), _r1, _coeffs, 1);
-                float16x8_t _tmp34b = vfmaq_laneq_f16(_r4, _r2, _coeffs, 3);
-
-                float16x8_t _tmp0 = vfmaq_laneq_f16(vaddq_f16(_r0, _r4), _r2, _coeffs, 4);
-                float16x8_t _tmp1 = vsubq_f16(_tmp12b, _tmp12a);
-                float16x8_t _tmp2 = vaddq_f16(_tmp12b, _tmp12a);
-                float16x8_t _tmp3 = vaddq_f16(_tmp34b, _tmp34a);
-                float16x8_t _tmp4 = vsubq_f16(_tmp34b, _tmp34a);
-                float16x8_t _tmp5 = vfmaq_laneq_f16(vaddq_f16(_r1, _r5), _r3, _coeffs, 4);
-
-                vst1q_f16(p0, _tmp0);
-                vst1q_f16(p1, _tmp1);
-                vst1q_f16(p2, _tmp2);
-                vst1q_f16(p3, _tmp3);
-                vst1q_f16(p4, _tmp4);
-                vst1q_f16(p5, _tmp5);
-
-                p0 += max_jj * 6 * 8;
-                p1 += max_jj * 6 * 8;
-                p2 += max_jj * 6 * 8;
-                p3 += max_jj * 6 * 8;
-                p4 += max_jj * 6 * 8;
-                p5 += max_jj * 6 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 4;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        __fp16 tmp[6][6][4];
-
-        const __fp16 coeffs[8] = {sq2, msq2_d2, -2.f, -0.5f, -2.5f, 0.f, 0.f, 0.f};
-        float16x8_t _coeffs = vld1q_f16(coeffs);
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = bottom_blob.channel((k + kk) / elempack).row<const __fp16>(ti * 4) + (tj * 4) * elempack;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float16x4_t _r0 = vdup_n_f16(0.f);
-                float16x4_t _r1 = vdup_n_f16(0.f);
-                float16x4_t _r2 = vdup_n_f16(0.f);
-                float16x4_t _r3 = vdup_n_f16(0.f);
-                float16x4_t _r4 = vdup_n_f16(0.f);
-                float16x4_t _r5 = vdup_n_f16(0.f);
-
-                if (ti * 4 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1_f16(r0);
-                        if (tj * 4 + 1 < w) _r1 = vld1_f16(r0 + 4);
-                        if (tj * 4 + 2 < w) _r2 = vld1_f16(r0 + 8);
-                        if (tj * 4 + 3 < w) _r3 = vld1_f16(r0 + 12);
-                        if (tj * 4 + 4 < w) _r4 = vld1_f16(r0 + 16);
-                        if (tj * 4 + 5 < w) _r5 = vld1_f16(r0 + 20);
-                    }
-                    if (elempack == 1)
-                    {
-                        const __fp16* r1 = r0 + N;
-                        const __fp16* r2 = r0 + N * 2;
-                        const __fp16* r3 = r0 + N * 3;
-
-                        float16x4_t _t0 = vld1_f16(r0);
-                        float16x4_t _t1 = vld1_f16(r1);
-                        float16x4_t _t2 = vld1_f16(r2);
-                        float16x4_t _t3 = vld1_f16(r3);
-
-                        transpose4x4_ph(_t0, _t1, _t2, _t3);
-
-                        _r0 = _t0;
-                        if (tj * 4 + 1 < w) _r1 = _t1;
-                        if (tj * 4 + 2 < w) _r2 = _t2;
-                        if (tj * 4 + 3 < w) _r3 = _t3;
-                        if (tj * 4 + 4 < w)
-                        {
-                            __fp16 tmp[4] = {r0[4], r1[4], r2[4], r3[4]};
-                            _r4 = vld1_f16(tmp);
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            __fp16 tmp[4] = {r0[5], r1[5], r2[5], r3[5]};
-                            _r5 = vld1_f16(tmp);
-                        }
-                    }
-                }
-
-                float16x4_t _tmp12a = vfma_laneq_f16(vmul_laneq_f16(_r1, _coeffs, 0), _r3, _coeffs, 1);
-                float16x4_t _tmp12b = vfma_laneq_f16(_r4, _r2, _coeffs, 2);
-                float16x4_t _tmp34a = vfma_laneq_f16(vmul_laneq_f16(_r3, _coeffs, 0), _r1, _coeffs, 1);
-                float16x4_t _tmp34b = vfma_laneq_f16(_r4, _r2, _coeffs, 3);
-
-                float16x4_t _tmp0 = vfma_laneq_f16(vadd_f16(_r0, _r4), _r2, _coeffs, 4);
-                float16x4_t _tmp1 = vsub_f16(_tmp12b, _tmp12a);
-                float16x4_t _tmp2 = vadd_f16(_tmp12b, _tmp12a);
-                float16x4_t _tmp3 = vadd_f16(_tmp34b, _tmp34a);
-                float16x4_t _tmp4 = vsub_f16(_tmp34b, _tmp34a);
-                float16x4_t _tmp5 = vfma_laneq_f16(vadd_f16(_r1, _r5), _r3, _coeffs, 4);
-
-                vst1_f16(tmp[0][m], _tmp0);
-                vst1_f16(tmp[1][m], _tmp1);
-                vst1_f16(tmp[2][m], _tmp2);
-                vst1_f16(tmp[3][m], _tmp3);
-                vst1_f16(tmp[4][m], _tmp4);
-                vst1_f16(tmp[5][m], _tmp5);
-
-                r0 += w * elempack;
-            }
-
-            __fp16* p0 = (__fp16*)B + kk * max_jj * 36 + jj * 4;
-            __fp16* p1 = p0 + max_jj * 4;
-            __fp16* p2 = p0 + max_jj * 4 * 2;
-            __fp16* p3 = p0 + max_jj * 4 * 3;
-            __fp16* p4 = p0 + max_jj * 4 * 4;
-            __fp16* p5 = p0 + max_jj * 4 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float16x4_t _r0 = vld1_f16(tmp[m][0]);
-                float16x4_t _r1 = vld1_f16(tmp[m][1]);
-                float16x4_t _r2 = vld1_f16(tmp[m][2]);
-                float16x4_t _r3 = vld1_f16(tmp[m][3]);
-                float16x4_t _r4 = vld1_f16(tmp[m][4]);
-                float16x4_t _r5 = vld1_f16(tmp[m][5]);
-
-                float16x4_t _tmp12a = vfma_laneq_f16(vmul_laneq_f16(_r1, _coeffs, 0), _r3, _coeffs, 1);
-                float16x4_t _tmp12b = vfma_laneq_f16(_r4, _r2, _coeffs, 2);
-                float16x4_t _tmp34a = vfma_laneq_f16(vmul_laneq_f16(_r3, _coeffs, 0), _r1, _coeffs, 1);
-                float16x4_t _tmp34b = vfma_laneq_f16(_r4, _r2, _coeffs, 3);
-
-                float16x4_t _tmp0 = vfma_laneq_f16(vadd_f16(_r0, _r4), _r2, _coeffs, 4);
-                float16x4_t _tmp1 = vsub_f16(_tmp12b, _tmp12a);
-                float16x4_t _tmp2 = vadd_f16(_tmp12b, _tmp12a);
-                float16x4_t _tmp3 = vadd_f16(_tmp34b, _tmp34a);
-                float16x4_t _tmp4 = vsub_f16(_tmp34b, _tmp34a);
-                float16x4_t _tmp5 = vfma_laneq_f16(vadd_f16(_r1, _r5), _r3, _coeffs, 4);
-
-                vst1_f16(p0, _tmp0);
-                vst1_f16(p1, _tmp1);
-                vst1_f16(p2, _tmp2);
-                vst1_f16(p3, _tmp3);
-                vst1_f16(p4, _tmp4);
-                vst1_f16(p5, _tmp5);
-
-                p0 += max_jj * 6 * 4;
-                p1 += max_jj * 6 * 4;
-                p2 += max_jj * 6 * 4;
-                p3 += max_jj * 6 * 4;
-                p4 += max_jj * 6 * 4;
-                p5 += max_jj * 6 * 4;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 4;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        __fp16 tmp[6][6][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = bottom_blob.channel(k + kk).row<const __fp16>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 6; m++)
-            {
-                __fp16 r00 = 0.f;
-                __fp16 r01 = 0.f;
-                __fp16 r10 = 0.f;
-                __fp16 r11 = 0.f;
-                __fp16 r20 = 0.f;
-                __fp16 r21 = 0.f;
-                __fp16 r30 = 0.f;
-                __fp16 r31 = 0.f;
-                __fp16 r40 = 0.f;
-                __fp16 r41 = 0.f;
-                __fp16 r50 = 0.f;
-                __fp16 r51 = 0.f;
-
-                if (ti * 4 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const __fp16* r1 = r0 + N;
-
-                        r00 = r0[0];
-                        r01 = r1[0];
-                        if (tj * 4 + 1 < w)
-                        {
-                            r10 = r0[1];
-                            r11 = r1[1];
-                        }
-                        if (tj * 4 + 2 < w)
-                        {
-                            r20 = r0[2];
-                            r21 = r1[2];
-                        }
-                        if (tj * 4 + 3 < w)
-                        {
-                            r30 = r0[3];
-                            r31 = r1[3];
-                        }
-                        if (tj * 4 + 4 < w)
-                        {
-                            r40 = r0[4];
-                            r41 = r1[4];
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            r50 = r0[5];
-                            r51 = r1[5];
-                        }
-                    }
-                }
-
-                __fp16 tmp12a0 = sq2 * r10 + msq2_d2 * r30;
-                __fp16 tmp12a1 = sq2 * r11 + msq2_d2 * r31;
-                __fp16 tmp12b0 = r40 - (__fp16)2.f * r20;
-                __fp16 tmp12b1 = r41 - (__fp16)2.f * r21;
-                __fp16 tmp34a0 = sq2 * r30 + msq2_d2 * r10;
-                __fp16 tmp34a1 = sq2 * r31 + msq2_d2 * r11;
-                __fp16 tmp34b0 = r40 - (__fp16)0.5f * r20;
-                __fp16 tmp34b1 = r41 - (__fp16)0.5f * r21;
-
-                tmp[0][m][0] = r00 + r40 - (__fp16)2.5f * r20;
-                tmp[0][m][1] = r01 + r41 - (__fp16)2.5f * r21;
-                tmp[1][m][0] = tmp12b0 - tmp12a0;
-                tmp[1][m][1] = tmp12b1 - tmp12a1;
-                tmp[2][m][0] = tmp12b0 + tmp12a0;
-                tmp[2][m][1] = tmp12b1 + tmp12a1;
-                tmp[3][m][0] = tmp34b0 + tmp34a0;
-                tmp[3][m][1] = tmp34b1 + tmp34a1;
-                tmp[4][m][0] = tmp34b0 - tmp34a0;
-                tmp[4][m][1] = tmp34b1 - tmp34a1;
-                tmp[5][m][0] = r10 + r50 - (__fp16)2.5f * r30;
-                tmp[5][m][1] = r11 + r51 - (__fp16)2.5f * r31;
-
-                r0 += w;
-            }
-
-            __fp16* p0 = (__fp16*)B + kk * max_jj * 36 + jj * 2;
-            __fp16* p1 = p0 + max_jj * 2;
-            __fp16* p2 = p0 + max_jj * 2 * 2;
-            __fp16* p3 = p0 + max_jj * 2 * 3;
-            __fp16* p4 = p0 + max_jj * 2 * 4;
-            __fp16* p5 = p0 + max_jj * 2 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                __fp16 r00 = tmp[m][0][0];
-                __fp16 r01 = tmp[m][0][1];
-                __fp16 r10 = tmp[m][1][0];
-                __fp16 r11 = tmp[m][1][1];
-                __fp16 r20 = tmp[m][2][0];
-                __fp16 r21 = tmp[m][2][1];
-                __fp16 r30 = tmp[m][3][0];
-                __fp16 r31 = tmp[m][3][1];
-                __fp16 r40 = tmp[m][4][0];
-                __fp16 r41 = tmp[m][4][1];
-                __fp16 r50 = tmp[m][5][0];
-                __fp16 r51 = tmp[m][5][1];
-
-                __fp16 tmp12a0 = sq2 * r10 + msq2_d2 * r30;
-                __fp16 tmp12a1 = sq2 * r11 + msq2_d2 * r31;
-                __fp16 tmp12b0 = r40 - (__fp16)2.f * r20;
-                __fp16 tmp12b1 = r41 - (__fp16)2.f * r21;
-                __fp16 tmp34a0 = sq2 * r30 + msq2_d2 * r10;
-                __fp16 tmp34a1 = sq2 * r31 + msq2_d2 * r11;
-                __fp16 tmp34b0 = r40 - (__fp16)0.5f * r20;
-                __fp16 tmp34b1 = r41 - (__fp16)0.5f * r21;
-
-                p0[0] = r00 + r40 - (__fp16)2.5f * r20;
-                p0[1] = r01 + r41 - (__fp16)2.5f * r21;
-                p1[0] = tmp12b0 - tmp12a0;
-                p1[1] = tmp12b1 - tmp12a1;
-                p2[0] = tmp12b0 + tmp12a0;
-                p2[1] = tmp12b1 + tmp12a1;
-                p3[0] = tmp34b0 + tmp34a0;
-                p3[1] = tmp34b1 + tmp34a1;
-                p4[0] = tmp34b0 - tmp34a0;
-                p4[1] = tmp34b1 - tmp34a1;
-                p5[0] = r10 + r50 - (__fp16)2.5f * r30;
-                p5[1] = r11 + r51 - (__fp16)2.5f * r31;
-
-                p0 += max_jj * 6 * 2;
-                p1 += max_jj * 6 * 2;
-                p2 += max_jj * 6 * 2;
-                p3 += max_jj * 6 * 2;
-                p4 += max_jj * 6 * 2;
-                p5 += max_jj * 6 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        __fp16 tmp[6][6];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0123 = bottom_blob.channel(k + kk).row<const __fp16>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 6; m++)
-            {
-                __fp16 r0 = 0.f;
-                __fp16 r1 = 0.f;
-                __fp16 r2 = 0.f;
-                __fp16 r3 = 0.f;
-                __fp16 r4 = 0.f;
-                __fp16 r5 = 0.f;
-
-                if (ti * 4 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = r0123[0];
-                        if (tj * 4 + 1 < w) r1 = r0123[1];
-                        if (tj * 4 + 2 < w) r2 = r0123[2];
-                        if (tj * 4 + 3 < w) r3 = r0123[3];
-                        if (tj * 4 + 4 < w) r4 = r0123[4];
-                        if (tj * 4 + 5 < w) r5 = r0123[5];
-                    }
-                }
-
-                __fp16 tmp12a = sq2 * r1 + msq2_d2 * r3;
-                __fp16 tmp12b = r4 - (__fp16)2.f * r2;
-                __fp16 tmp34a = sq2 * r3 + msq2_d2 * r1;
-                __fp16 tmp34b = r4 - (__fp16)0.5f * r2;
-
-                tmp[0][m] = r0 + r4 - (__fp16)2.5f * r2;
-                tmp[1][m] = tmp12b - tmp12a;
-                tmp[2][m] = tmp12b + tmp12a;
-                tmp[3][m] = tmp34b + tmp34a;
-                tmp[4][m] = tmp34b - tmp34a;
-                tmp[5][m] = r1 + r5 - (__fp16)2.5f * r3;
-
-                r0123 += w;
-            }
-
-            __fp16* p0 = (__fp16*)B + kk * max_jj * 36 + jj;
-            __fp16* p1 = p0 + max_jj;
-            __fp16* p2 = p0 + max_jj * 2;
-            __fp16* p3 = p0 + max_jj * 3;
-            __fp16* p4 = p0 + max_jj * 4;
-            __fp16* p5 = p0 + max_jj * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                __fp16 r0 = tmp[m][0];
-                __fp16 r1 = tmp[m][1];
-                __fp16 r2 = tmp[m][2];
-                __fp16 r3 = tmp[m][3];
-                __fp16 r4 = tmp[m][4];
-                __fp16 r5 = tmp[m][5];
-
-                __fp16 tmp12a = sq2 * r1 + msq2_d2 * r3;
-                __fp16 tmp12b = r4 - (__fp16)2.f * r2;
-                __fp16 tmp34a = sq2 * r3 + msq2_d2 * r1;
-                __fp16 tmp34b = r4 - (__fp16)0.5f * r2;
-
-                p0[0] = r0 + r4 - (__fp16)2.5f * r2;
-                p1[0] = tmp12b - tmp12a;
-                p2[0] = tmp12b + tmp12a;
-                p3[0] = tmp34b + tmp34a;
-                p4[0] = tmp34b - tmp34a;
-                p5[0] = r1 + r5 - (__fp16)2.5f * r3;
-
-                p0 += max_jj * 6;
-                p1 += max_jj * 6;
-                p2 += max_jj * 6;
-                p3 += max_jj * 6;
-                p4 += max_jj * 6;
-                p5 += max_jj * 6;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_output_tile_fp16sa(const Mat& top_tile, Mat& top_blob, const Mat& bias, int i, int max_ii, int j, int max_jj)
-{
-    const __fp16 sq2 = 1.41421356237;
-    const __fp16 sq2_m2 = 1.41421356237 * 2;
-    const __fp16 sq2_d2 = 1.41421356237 / 2;
-    const __fp16 sq2_d4 = 1.41421356237 / 4;
-
-    // const float otm[4][6] = {
-    //     {1.0f, 1.0f,   1.0f,  1.0f,  1.0f,   0.0f},
-    //     {0.0f, sq2/2, -sq2/2, sq2,   -sq2,   0.0f},
-    //     {0.0f, 0.5f,   0.5f,  2.0f,  2.0f,   0.0f},
-    //     {0.0f, sq2/4, -sq2/4, sq2*2, -sq2*2, 1.0f}
-    // };
-
-    // 0 = r00 + (r01 + r02) + (r03 + r04)
-    // 1 =       (r01 - r02) * sq2_d2 + (r03 - r04) * sq2
-    // 2 =       (r01 + r02) * 0.5f + (r03 + r04) * 2
-    // 3 = r05 + (r01 - r02) * sq2_d4 + (r03 - r04) * sq2_m2
-
-    const __fp16 coeffs[8] = {sq2, sq2_d2, sq2_d4, sq2_m2, 0.5f, 2.f, 0.f, 0.f};
-    float16x8_t _coeffs = vld1q_f16(coeffs);
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 3) / 4;
-
-    const __fp16* biasptr = bias;
-
-    int ii = 0;
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        float16x8_t _bias0 = biasptr ? vld1q_f16(biasptr + i + ii) : vdupq_n_f16(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        __fp16 tmp[4][6][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = (const __fp16*)top_tile + ii * max_jj * 36 + jj * 8;
-            const __fp16* r1 = r0 + max_jj * 8;
-            const __fp16* r2 = r0 + max_jj * 8 * 2;
-            const __fp16* r3 = r0 + max_jj * 8 * 3;
-            const __fp16* r4 = r0 + max_jj * 8 * 4;
-            const __fp16* r5 = r0 + max_jj * 8 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float16x8_t _r0 = vld1q_f16(r0);
-                float16x8_t _r1 = vld1q_f16(r1);
-                float16x8_t _r2 = vld1q_f16(r2);
-                float16x8_t _r3 = vld1q_f16(r3);
-                float16x8_t _r4 = vld1q_f16(r4);
-                float16x8_t _r5 = vld1q_f16(r5);
-
-                float16x8_t _tmp02a = vaddq_f16(_r1, _r2);
-                float16x8_t _tmp02b = vaddq_f16(_r3, _r4);
-                float16x8_t _tmp13a = vsubq_f16(_r1, _r2);
-                float16x8_t _tmp13b = vsubq_f16(_r3, _r4);
-
-                float16x8_t _tmp0 = vaddq_f16(vaddq_f16(_r0, _tmp02a), _tmp02b);
-                float16x8_t _tmp1 = vfmaq_laneq_f16(vmulq_laneq_f16(_tmp13a, _coeffs, 1), _tmp13b, _coeffs, 0);
-                float16x8_t _tmp2 = vfmaq_laneq_f16(vmulq_laneq_f16(_tmp02a, _coeffs, 4), _tmp02b, _coeffs, 5);
-                float16x8_t _tmp3 = vfmaq_laneq_f16(vfmaq_laneq_f16(_r5, _tmp13a, _coeffs, 2), _tmp13b, _coeffs, 3);
-
-                vst1q_f16(tmp[0][m], _tmp0);
-                vst1q_f16(tmp[1][m], _tmp1);
-                vst1q_f16(tmp[2][m], _tmp2);
-                vst1q_f16(tmp[3][m], _tmp3);
-
-                r0 += max_jj * 6 * 8;
-                r1 += max_jj * 6 * 8;
-                r2 += max_jj * 6 * 8;
-                r3 += max_jj * 6 * 8;
-                r4 += max_jj * 6 * 8;
-                r5 += max_jj * 6 * 8;
-            }
-
-            __fp16* outptr0 = top_blob.channel((i + ii) / out_elempack).row<__fp16>(ti * 4) + (tj * 4) * out_elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                float16x8_t _r0 = vld1q_f16(tmp[m][0]);
-                float16x8_t _r1 = vld1q_f16(tmp[m][1]);
-                float16x8_t _r2 = vld1q_f16(tmp[m][2]);
-                float16x8_t _r3 = vld1q_f16(tmp[m][3]);
-                float16x8_t _r4 = vld1q_f16(tmp[m][4]);
-                float16x8_t _r5 = vld1q_f16(tmp[m][5]);
-
-                float16x8_t _tmp02a = vaddq_f16(_r1, _r2);
-                float16x8_t _tmp02b = vaddq_f16(_r3, _r4);
-                float16x8_t _tmp13a = vsubq_f16(_r1, _r2);
-                float16x8_t _tmp13b = vsubq_f16(_r3, _r4);
-
-                float16x8_t _tmp0 = vaddq_f16(vaddq_f16(_r0, _tmp02a), vaddq_f16(_tmp02b, _bias0));
-                float16x8_t _tmp1 = vfmaq_laneq_f16(vfmaq_laneq_f16(_bias0, _tmp13a, _coeffs, 1), _tmp13b, _coeffs, 0);
-                float16x8_t _tmp2 = vfmaq_laneq_f16(vfmaq_laneq_f16(_bias0, _tmp02a, _coeffs, 4), _tmp02b, _coeffs, 5);
-                float16x8_t _tmp3 = vfmaq_laneq_f16(vfmaq_laneq_f16(vaddq_f16(_r5, _bias0), _tmp13a, _coeffs, 2), _tmp13b, _coeffs, 3);
-
-                if (out_elempack == 8)
-                {
-                    vst1q_f16(outptr0, _tmp0);
-                    if (tj * 4 + 1 < outw) vst1q_f16(outptr0 + 8, _tmp1);
-                    if (tj * 4 + 2 < outw) vst1q_f16(outptr0 + 16, _tmp2);
-                    if (tj * 4 + 3 < outw) vst1q_f16(outptr0 + 24, _tmp3);
-                }
-                if (out_elempack == 4)
-                {
-                    __fp16* outptr1 = outptr0 + N;
-
-                    vst1_f16(outptr0, vget_low_f16(_tmp0));
-                    vst1_f16(outptr1, vget_high_f16(_tmp0));
-                    if (tj * 4 + 1 < outw)
-                    {
-                        vst1_f16(outptr0 + 4, vget_low_f16(_tmp1));
-                        vst1_f16(outptr1 + 4, vget_high_f16(_tmp1));
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        vst1_f16(outptr0 + 8, vget_low_f16(_tmp2));
-                        vst1_f16(outptr1 + 8, vget_high_f16(_tmp2));
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        vst1_f16(outptr0 + 12, vget_low_f16(_tmp3));
-                        vst1_f16(outptr1 + 12, vget_high_f16(_tmp3));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    __fp16 tmp0[8];
-                    __fp16 tmp1[8];
-                    __fp16 tmp2[8];
-                    __fp16 tmp3[8];
-                    vst1q_f16(tmp0, _tmp0);
-                    vst1q_f16(tmp1, _tmp1);
-                    vst1q_f16(tmp2, _tmp2);
-                    vst1q_f16(tmp3, _tmp3);
-
-                    __fp16* outptr1 = outptr0 + N;
-                    __fp16* outptr2 = outptr0 + N * 2;
-                    __fp16* outptr3 = outptr0 + N * 3;
-                    __fp16* outptr4 = outptr0 + N * 4;
-                    __fp16* outptr5 = outptr0 + N * 5;
-                    __fp16* outptr6 = outptr0 + N * 6;
-                    __fp16* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                        outptr4[2] = tmp2[4];
-                        outptr5[2] = tmp2[5];
-                        outptr6[2] = tmp2[6];
-                        outptr7[2] = tmp2[7];
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                        outptr4[3] = tmp3[4];
-                        outptr5[3] = tmp3[5];
-                        outptr6[3] = tmp3[6];
-                        outptr7[3] = tmp3[7];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        float16x4_t _bias0 = biasptr ? vld1_f16(biasptr + i + ii) : vdup_n_f16(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        __fp16 tmp[4][6][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = (const __fp16*)top_tile + ii * max_jj * 36 + jj * 4;
-            const __fp16* r1 = r0 + max_jj * 4;
-            const __fp16* r2 = r0 + max_jj * 4 * 2;
-            const __fp16* r3 = r0 + max_jj * 4 * 3;
-            const __fp16* r4 = r0 + max_jj * 4 * 4;
-            const __fp16* r5 = r0 + max_jj * 4 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float16x4_t _r0 = vld1_f16(r0);
-                float16x4_t _r1 = vld1_f16(r1);
-                float16x4_t _r2 = vld1_f16(r2);
-                float16x4_t _r3 = vld1_f16(r3);
-                float16x4_t _r4 = vld1_f16(r4);
-                float16x4_t _r5 = vld1_f16(r5);
-
-                float16x4_t _tmp02a = vadd_f16(_r1, _r2);
-                float16x4_t _tmp02b = vadd_f16(_r3, _r4);
-                float16x4_t _tmp13a = vsub_f16(_r1, _r2);
-                float16x4_t _tmp13b = vsub_f16(_r3, _r4);
-
-                float16x4_t _tmp0 = vadd_f16(vadd_f16(_r0, _tmp02a), _tmp02b);
-                float16x4_t _tmp1 = vfma_laneq_f16(vmul_laneq_f16(_tmp13a, _coeffs, 1), _tmp13b, _coeffs, 0);
-                float16x4_t _tmp2 = vfma_laneq_f16(vmul_laneq_f16(_tmp02a, _coeffs, 4), _tmp02b, _coeffs, 5);
-                float16x4_t _tmp3 = vfma_laneq_f16(vfma_laneq_f16(_r5, _tmp13a, _coeffs, 2), _tmp13b, _coeffs, 3);
-
-                vst1_f16(tmp[0][m], _tmp0);
-                vst1_f16(tmp[1][m], _tmp1);
-                vst1_f16(tmp[2][m], _tmp2);
-                vst1_f16(tmp[3][m], _tmp3);
-
-                r0 += max_jj * 6 * 4;
-                r1 += max_jj * 6 * 4;
-                r2 += max_jj * 6 * 4;
-                r3 += max_jj * 6 * 4;
-                r4 += max_jj * 6 * 4;
-                r5 += max_jj * 6 * 4;
-            }
-
-            __fp16* outptr0 = top_blob.channel((i + ii) / out_elempack).row<__fp16>(ti * 4) + (tj * 4) * out_elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                float16x4_t _r0 = vld1_f16(tmp[m][0]);
-                float16x4_t _r1 = vld1_f16(tmp[m][1]);
-                float16x4_t _r2 = vld1_f16(tmp[m][2]);
-                float16x4_t _r3 = vld1_f16(tmp[m][3]);
-                float16x4_t _r4 = vld1_f16(tmp[m][4]);
-                float16x4_t _r5 = vld1_f16(tmp[m][5]);
-
-                float16x4_t _tmp02a = vadd_f16(_r1, _r2);
-                float16x4_t _tmp02b = vadd_f16(_r3, _r4);
-                float16x4_t _tmp13a = vsub_f16(_r1, _r2);
-                float16x4_t _tmp13b = vsub_f16(_r3, _r4);
-
-                float16x4_t _tmp0 = vadd_f16(vadd_f16(_r0, _tmp02a), vadd_f16(_tmp02b, _bias0));
-                float16x4_t _tmp1 = vfma_laneq_f16(vfma_laneq_f16(_bias0, _tmp13a, _coeffs, 1), _tmp13b, _coeffs, 0);
-                float16x4_t _tmp2 = vfma_laneq_f16(vfma_laneq_f16(_bias0, _tmp02a, _coeffs, 4), _tmp02b, _coeffs, 5);
-                float16x4_t _tmp3 = vfma_laneq_f16(vfma_laneq_f16(vadd_f16(_r5, _bias0), _tmp13a, _coeffs, 2), _tmp13b, _coeffs, 3);
-
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr0, _tmp0);
-                    if (tj * 4 + 1 < outw) vst1_f16(outptr0 + 4, _tmp1);
-                    if (tj * 4 + 2 < outw) vst1_f16(outptr0 + 8, _tmp2);
-                    if (tj * 4 + 3 < outw) vst1_f16(outptr0 + 12, _tmp3);
-                }
-                if (out_elempack == 1)
-                {
-                    __fp16 tmp0[4];
-                    __fp16 tmp1[4];
-                    __fp16 tmp2[4];
-                    __fp16 tmp3[4];
-                    vst1_f16(tmp0, _tmp0);
-                    vst1_f16(tmp1, _tmp1);
-                    vst1_f16(tmp2, _tmp2);
-                    vst1_f16(tmp3, _tmp3);
-
-                    __fp16* outptr1 = outptr0 + N;
-                    __fp16* outptr2 = outptr0 + N * 2;
-                    __fp16* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        __fp16 bias0 = biasptr ? biasptr[i + ii] : 0.f;
-        __fp16 bias1 = biasptr ? biasptr[i + ii + 1] : 0.f;
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        __fp16 tmp[4][6][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = (const __fp16*)top_tile + ii * max_jj * 36 + jj * 2;
-            const __fp16* r1 = r0 + max_jj * 2;
-            const __fp16* r2 = r0 + max_jj * 2 * 2;
-            const __fp16* r3 = r0 + max_jj * 2 * 3;
-            const __fp16* r4 = r0 + max_jj * 2 * 4;
-            const __fp16* r5 = r0 + max_jj * 2 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                __fp16 tmp02a0 = r1[0] + r2[0];
-                __fp16 tmp02a1 = r1[1] + r2[1];
-                __fp16 tmp02b0 = r3[0] + r4[0];
-                __fp16 tmp02b1 = r3[1] + r4[1];
-                __fp16 tmp13a0 = r1[0] - r2[0];
-                __fp16 tmp13a1 = r1[1] - r2[1];
-                __fp16 tmp13b0 = r3[0] - r4[0];
-                __fp16 tmp13b1 = r3[1] - r4[1];
-
-                tmp[0][m][0] = r0[0] + tmp02a0 + tmp02b0;
-                tmp[0][m][1] = r0[1] + tmp02a1 + tmp02b1;
-                tmp[1][m][0] = tmp13a0 * sq2_d2 + tmp13b0 * sq2;
-                tmp[1][m][1] = tmp13a1 * sq2_d2 + tmp13b1 * sq2;
-                tmp[2][m][0] = tmp02a0 * (__fp16)0.5f + tmp02b0 * (__fp16)2;
-                tmp[2][m][1] = tmp02a1 * (__fp16)0.5f + tmp02b1 * (__fp16)2;
-                tmp[3][m][0] = r5[0] + tmp13a0 * sq2_d4 + tmp13b0 * sq2_m2;
-                tmp[3][m][1] = r5[1] + tmp13a1 * sq2_d4 + tmp13b1 * sq2_m2;
-
-                r0 += max_jj * 6 * 2;
-                r1 += max_jj * 6 * 2;
-                r2 += max_jj * 6 * 2;
-                r3 += max_jj * 6 * 2;
-                r4 += max_jj * 6 * 2;
-                r5 += max_jj * 6 * 2;
-            }
-
-            __fp16* outptr0 = top_blob.channel(i + ii).row<__fp16>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                __fp16 r00 = tmp[m][0][0];
-                __fp16 r01 = tmp[m][0][1];
-                __fp16 r10 = tmp[m][1][0];
-                __fp16 r11 = tmp[m][1][1];
-                __fp16 r20 = tmp[m][2][0];
-                __fp16 r21 = tmp[m][2][1];
-                __fp16 r30 = tmp[m][3][0];
-                __fp16 r31 = tmp[m][3][1];
-                __fp16 r40 = tmp[m][4][0];
-                __fp16 r41 = tmp[m][4][1];
-                __fp16 r50 = tmp[m][5][0];
-                __fp16 r51 = tmp[m][5][1];
-
-                __fp16 tmp02a0 = r10 + r20;
-                __fp16 tmp02a1 = r11 + r21;
-                __fp16 tmp02b0 = r30 + r40;
-                __fp16 tmp02b1 = r31 + r41;
-                __fp16 tmp13a0 = r10 - r20;
-                __fp16 tmp13a1 = r11 - r21;
-                __fp16 tmp13b0 = r30 - r40;
-                __fp16 tmp13b1 = r31 - r41;
-
-                __fp16 tmp00 = bias0 + r00 + tmp02a0 + tmp02b0;
-                __fp16 tmp01 = bias1 + r01 + tmp02a1 + tmp02b1;
-                __fp16 tmp10 = bias0 + tmp13a0 * sq2_d2 + tmp13b0 * sq2;
-                __fp16 tmp11 = bias1 + tmp13a1 * sq2_d2 + tmp13b1 * sq2;
-                __fp16 tmp20 = bias0 + tmp02a0 * (__fp16)0.5f + tmp02b0 * (__fp16)2;
-                __fp16 tmp21 = bias1 + tmp02a1 * (__fp16)0.5f + tmp02b1 * (__fp16)2;
-                __fp16 tmp30 = bias0 + r50 + tmp13a0 * sq2_d4 + tmp13b0 * sq2_m2;
-                __fp16 tmp31 = bias1 + r51 + tmp13a1 * sq2_d4 + tmp13b1 * sq2_m2;
-
-                // if (out_elempack == 1)
-                {
-                    __fp16* outptr1 = outptr0 + N;
-
-                    outptr0[0] = tmp00;
-                    outptr1[0] = tmp01;
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp10;
-                        outptr1[1] = tmp11;
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp20;
-                        outptr1[2] = tmp21;
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp30;
-                        outptr1[3] = tmp31;
-                    }
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        __fp16 bias0 = biasptr ? biasptr[i + ii] : 0.f;
-
-        __fp16 tmp[4][6];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = (const __fp16*)top_tile + ii * max_jj * 36 + jj;
-            const __fp16* r1 = r0 + max_jj;
-            const __fp16* r2 = r0 + max_jj * 2;
-            const __fp16* r3 = r0 + max_jj * 3;
-            const __fp16* r4 = r0 + max_jj * 4;
-            const __fp16* r5 = r0 + max_jj * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                __fp16 tmp02a = r1[0] + r2[0];
-                __fp16 tmp02b = r3[0] + r4[0];
-                __fp16 tmp13a = r1[0] - r2[0];
-                __fp16 tmp13b = r3[0] - r4[0];
-
-                tmp[0][m] = r0[0] + tmp02a + tmp02b;
-                tmp[1][m] = tmp13a * sq2_d2 + tmp13b * sq2;
-                tmp[2][m] = tmp02a * (__fp16)0.5f + tmp02b * (__fp16)2;
-                tmp[3][m] = r5[0] + tmp13a * sq2_d4 + tmp13b * sq2_m2;
-
-                r0 += max_jj * 6;
-                r1 += max_jj * 6;
-                r2 += max_jj * 6;
-                r3 += max_jj * 6;
-                r4 += max_jj * 6;
-                r5 += max_jj * 6;
-            }
-
-            __fp16* outptr0 = top_blob.channel(i + ii).row<__fp16>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                __fp16 r0 = tmp[m][0];
-                __fp16 r1 = tmp[m][1];
-                __fp16 r2 = tmp[m][2];
-                __fp16 r3 = tmp[m][3];
-                __fp16 r4 = tmp[m][4];
-                __fp16 r5 = tmp[m][5];
-
-                __fp16 tmp02a = r1 + r2;
-                __fp16 tmp02b = r3 + r4;
-                __fp16 tmp13a = r1 - r2;
-                __fp16 tmp13b = r3 - r4;
-
-                __fp16 tmp0 = bias0 + r0 + tmp02a + tmp02b;
-                __fp16 tmp1 = bias0 + tmp13a * sq2_d2 + tmp13b * sq2;
-                __fp16 tmp2 = bias0 + tmp02a * (__fp16)0.5f + tmp02b * (__fp16)2;
-                __fp16 tmp3 = bias0 + r5 + tmp13a * sq2_d4 + tmp13b * sq2_m2;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = tmp0;
-                    if (tj * 4 + 1 < outw) outptr0[1] = tmp1;
-                    if (tj * 4 + 2 < outw) outptr0[2] = tmp2;
-                    if (tj * 4 + 3 < outw) outptr0[3] = tmp3;
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, const Mat& bias, int nT, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 4n+2, winograd F(4,3)
-    int w_tiles = (outw + 3) / 4;
-    int h_tiles = (outh + 3) / 4;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 36;
-
-    // NCNN_LOGE("conv3x3s1_winograd43_fp16sa %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    conv3x3s1_winograd_get_optimal_tile_mnk_fp16(M, N, K, B, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 2u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 2u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd43_transform_input_tile_fp16sa(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile_fp16(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 2u, opt.workspace_allocator);
-
-        #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd43_transform_input_tile_fp16sa(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile_fp16(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 2u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                conv3x3s1_winograd_gemm_transB_packed_tile_fp16sa(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk, opt.use_a53_a55_optimized_kernel);
-            }
-
-            // transform output
-            conv3x3s1_winograd43_transform_output_tile_fp16sa(top_tile, top_blob, bias, i, max_ii, j, max_jj);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd63_transform_kernel_tile_fp16sa(const Mat& kernel, Mat& A, int inch, int i, int max_ii, int k, int max_kk)
-{
-    __fp16* ptmp = A;
-
-    int ii = 0;
-    for (; ii < max_ii; ii++)
-    {
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            // const float ktm[8][3] = {
-            //     {1.0f, 0.0f, 0.0f},
-            //     {-2.0f / 9, -2.0f / 9, -2.0f / 9},
-            //     {-2.0f / 9, 2.0f / 9, -2.0f / 9},
-            //     {1.0f / 90, 1.0f / 45, 2.0f / 45},
-            //     {1.0f / 90, -1.0f / 45, 2.0f / 45},
-            //     {1.0f / 45, 1.0f / 90, 1.0f / 180},
-            //     {1.0f / 45, -1.0f / 90, 1.0f / 180},
-            //     {0.0f, 0.0f, 1.0f}
-            // };
-            const float ktm0 = 2.0f / 9;
-            const float ktm1 = 1.0f / 45;
-            const float ktm2 = 2.0f / 45;
-            const float ktm3 = 1.0f / 90;
-            const float ktm4 = 1.0f / 180;
-
-            float tmp[8][3];
-
-            const float* k0 = (const float*)kernel + (i + ii) * inch * 9 + (k + kk) * 9;
-
-            for (int m = 0; m < 3; m++)
-            {
-                float r0 = k0[0];
-                float r1 = k0[1];
-                float r2 = k0[2];
-
-                tmp[0][m] = r0;
-                tmp[1][m] = -r0 * ktm0 - r1 * ktm0 - r2 * ktm0;
-                tmp[2][m] = -r0 * ktm0 + r1 * ktm0 - r2 * ktm0;
-                tmp[3][m] = r0 * ktm3 + r1 * ktm1 + r2 * ktm2;
-                tmp[4][m] = r0 * ktm3 - r1 * ktm1 + r2 * ktm2;
-                tmp[5][m] = r0 * ktm1 + r1 * ktm3 + r2 * ktm4;
-                tmp[6][m] = r0 * ktm1 - r1 * ktm3 + r2 * ktm4;
-                tmp[7][m] = r2;
-
-                k0 += 3;
-            }
-
-            for (int m = 0; m < 8; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-
-                float z0 = r0;
-                float z1 = -r0 * ktm0 - r1 * ktm0 - r2 * ktm0;
-                float z2 = -r0 * ktm0 + r1 * ktm0 - r2 * ktm0;
-                float z3 = r0 * ktm3 + r1 * ktm1 + r2 * ktm2;
-                float z4 = r0 * ktm3 - r1 * ktm1 + r2 * ktm2;
-                float z5 = r0 * ktm1 + r1 * ktm3 + r2 * ktm4;
-                float z6 = r0 * ktm1 - r1 * ktm3 + r2 * ktm4;
-                float z7 = r2;
-
-                ptmp[0] = (__fp16)z0;
-                ptmp[1] = (__fp16)z1;
-                ptmp[2] = (__fp16)z2;
-                ptmp[3] = (__fp16)z3;
-                ptmp[4] = (__fp16)z4;
-                ptmp[5] = (__fp16)z5;
-                ptmp[6] = (__fp16)z6;
-                ptmp[7] = (__fp16)z7;
-                ptmp += 8;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd63_transform_kernel_fp16sa(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt)
-{
-    const int M = outch;
-    const int K = inch;
-    const int B = 64;
-
-    int TILE_M, TILE_N, TILE_K;
-    conv3x3s1_winograd_get_optimal_tile_mnk_fp16(M, 0, K, B, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    Mat A_tileX(B * TILE_M * TILE_K, 1, opt.num_threads, (size_t)2u);
-
-    AT.create(TILE_K * TILE_M, B, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M, (size_t)2u);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat A_tile = A_tileX.channel(get_omp_thread_num());
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_ii = std::min((M - i), TILE_M);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            conv3x3s1_winograd63_transform_kernel_tile_fp16sa(kernel, A_tile, inch, i, max_ii, k, max_kk);
-
-            Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-            conv3x3s1_winograd_pack_A_tile_fp16(A_tile, AT_tile, B, max_ii, max_kk);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd63_transform_input_tile_fp16sa(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    // const float itm[8][8] = {
-    //     {1.0f, 0.0f,-5.25f, 0.00f, 5.25f, 0.00f,-1.0f, 0.0f},
-    //     {0.0f, 1.0f, 1.00f,-4.25f,-4.25f, 1.00f, 1.0f, 0.0f},
-    //     {0.0f,-1.0f, 1.00f, 4.25f,-4.25f,-1.00f, 1.0f, 0.0f},
-    //     {0.0f, 0.5f, 0.25f,-2.50f,-1.25f, 2.00f, 1.0f, 0.0f},
-    //     {0.0f,-0.5f, 0.25f, 2.50f,-1.25f,-2.00f, 1.0f, 0.0f},
-    //     {0.0f, 2.0f, 4.00f,-2.50f,-5.00f, 0.50f, 1.0f, 0.0f},
-    //     {0.0f,-2.0f, 4.00f, 2.50f,-5.00f,-0.50f, 1.0f, 0.0f},
-    //     {0.0f,-1.0f, 0.00f, 5.25f, 0.00f,-5.25f, 0.0f, 1.0f}
-    // };
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w + 3) / 6;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        __fp16 tmp[8][8][8];
-
-        const __fp16 coeffs[8] = {5.25f, -4.25f, -1.25f, 0.25f, -2.5f, 0.5f, 2.f, 4.f};
-        float16x8_t _coeffs = vld1q_f16(coeffs);
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = bottom_blob.channel((k + kk) / elempack).row<const __fp16>(ti * 6) + (tj * 6) * elempack;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float16x8_t _r0 = vdupq_n_f16(0.f);
-                float16x8_t _r1 = vdupq_n_f16(0.f);
-                float16x8_t _r2 = vdupq_n_f16(0.f);
-                float16x8_t _r3 = vdupq_n_f16(0.f);
-                float16x8_t _r4 = vdupq_n_f16(0.f);
-                float16x8_t _r5 = vdupq_n_f16(0.f);
-                float16x8_t _r6 = vdupq_n_f16(0.f);
-                float16x8_t _r7 = vdupq_n_f16(0.f);
-
-                if (ti * 6 + m < h)
-                {
-                    if (elempack == 8)
-                    {
-                        _r0 = vld1q_f16(r0);
-                        if (tj * 6 + 1 < w) _r1 = vld1q_f16(r0 + 8);
-                        if (tj * 6 + 2 < w) _r2 = vld1q_f16(r0 + 16);
-                        if (tj * 6 + 3 < w) _r3 = vld1q_f16(r0 + 24);
-                        if (tj * 6 + 4 < w) _r4 = vld1q_f16(r0 + 32);
-                        if (tj * 6 + 5 < w) _r5 = vld1q_f16(r0 + 40);
-                        if (tj * 6 + 6 < w) _r6 = vld1q_f16(r0 + 48);
-                        if (tj * 6 + 7 < w) _r7 = vld1q_f16(r0 + 56);
-                    }
-                    if (elempack == 4)
-                    {
-                        const __fp16* r1 = r0 + N;
-
-                        _r0 = vcombine_f16(vld1_f16(r0), vld1_f16(r1));
-                        if (tj * 6 + 1 < w)
-                        {
-                            _r1 = vcombine_f16(vld1_f16(r0 + 4), vld1_f16(r1 + 4));
-                        }
-                        if (tj * 6 + 2 < w)
-                        {
-                            _r2 = vcombine_f16(vld1_f16(r0 + 8), vld1_f16(r1 + 8));
-                        }
-                        if (tj * 6 + 3 < w)
-                        {
-                            _r3 = vcombine_f16(vld1_f16(r0 + 12), vld1_f16(r1 + 12));
-                        }
-                        if (tj * 6 + 4 < w)
-                        {
-                            _r4 = vcombine_f16(vld1_f16(r0 + 16), vld1_f16(r1 + 16));
-                        }
-                        if (tj * 6 + 5 < w)
-                        {
-                            _r5 = vcombine_f16(vld1_f16(r0 + 20), vld1_f16(r1 + 20));
-                        }
-                        if (tj * 6 + 6 < w)
-                        {
-                            _r6 = vcombine_f16(vld1_f16(r0 + 24), vld1_f16(r1 + 24));
-                        }
-                        if (tj * 6 + 7 < w)
-                        {
-                            _r7 = vcombine_f16(vld1_f16(r0 + 28), vld1_f16(r1 + 28));
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        const __fp16* r1 = r0 + N;
-                        const __fp16* r2 = r0 + N * 2;
-                        const __fp16* r3 = r0 + N * 3;
-                        const __fp16* r4 = r0 + N * 4;
-                        const __fp16* r5 = r0 + N * 5;
-                        const __fp16* r6 = r0 + N * 6;
-                        const __fp16* r7 = r0 + N * 7;
-
-                        float16x4_t _t0 = vld1_f16(r0);
-                        float16x4_t _t1 = vld1_f16(r1);
-                        float16x4_t _t2 = vld1_f16(r2);
-                        float16x4_t _t3 = vld1_f16(r3);
-                        float16x4_t _t4 = vld1_f16(r4);
-                        float16x4_t _t5 = vld1_f16(r5);
-                        float16x4_t _t6 = vld1_f16(r6);
-                        float16x4_t _t7 = vld1_f16(r7);
-
-                        transpose4x4_ph(_t0, _t1, _t2, _t3);
-                        transpose4x4_ph(_t4, _t5, _t6, _t7);
-
-                        _r0 = vcombine_f16(_t0, _t4);
-                        if (tj * 6 + 1 < w)
-                        {
-                            _r1 = vcombine_f16(_t1, _t5);
-                        }
-                        if (tj * 6 + 2 < w)
-                        {
-                            _r2 = vcombine_f16(_t2, _t6);
-                        }
-                        if (tj * 6 + 3 < w)
-                        {
-                            _r3 = vcombine_f16(_t3, _t7);
-                        }
-                        if (tj * 6 + 4 < w)
-                        {
-                            _t0 = vld1_f16(r0 + 4);
-                            _t1 = vld1_f16(r1 + 4);
-                            _t2 = vld1_f16(r2 + 4);
-                            _t3 = vld1_f16(r3 + 4);
-                            _t4 = vld1_f16(r4 + 4);
-                            _t5 = vld1_f16(r5 + 4);
-                            _t6 = vld1_f16(r6 + 4);
-                            _t7 = vld1_f16(r7 + 4);
-
-                            transpose4x4_ph(_t0, _t1, _t2, _t3);
-                            transpose4x4_ph(_t4, _t5, _t6, _t7);
-
-                            _r4 = vcombine_f16(_t0, _t4);
-                            if (tj * 6 + 5 < w)
-                            {
-                                _r5 = vcombine_f16(_t1, _t5);
-                            }
-                            if (tj * 6 + 6 < w)
-                            {
-                                _r6 = vcombine_f16(_t2, _t6);
-                            }
-                            if (tj * 6 + 7 < w)
-                            {
-                                _r7 = vcombine_f16(_t3, _t7);
-                            }
-                        }
-                    }
-                }
-
-                float16x8_t _tmp12a = vfmaq_laneq_f16(vaddq_f16(_r2, _r6), _r4, _coeffs, 1);
-                float16x8_t _tmp12b = vfmaq_laneq_f16(vaddq_f16(_r1, _r5), _r3, _coeffs, 1);
-                float16x8_t _tmp34a = vfmaq_laneq_f16(vfmaq_laneq_f16(_r6, _r2, _coeffs, 3), _r4, _coeffs, 2);
-                float16x8_t _tmp34b = vfmaq_laneq_f16(vfmaq_laneq_f16(vmulq_laneq_f16(_r1, _coeffs, 5), _r3, _coeffs, 4), _r5, _coeffs, 6);
-                float16x8_t _tmp56a = vfmaq_laneq_f16(_r6, vfmaq_laneq_f16(_r2, _r4, _coeffs, 2), _coeffs, 7);
-                float16x8_t _tmp56b = vfmaq_laneq_f16(vfmaq_laneq_f16(vmulq_laneq_f16(_r1, _coeffs, 6), _r3, _coeffs, 4), _r5, _coeffs, 5);
-
-                float16x8_t _tmp0 = vfmaq_laneq_f16(vsubq_f16(_r0, _r6), vsubq_f16(_r4, _r2), _coeffs, 0);
-                float16x8_t _tmp1 = vaddq_f16(_tmp12a, _tmp12b);
-                float16x8_t _tmp2 = vsubq_f16(_tmp12a, _tmp12b);
-                float16x8_t _tmp3 = vaddq_f16(_tmp34a, _tmp34b);
-                float16x8_t _tmp4 = vsubq_f16(_tmp34a, _tmp34b);
-                float16x8_t _tmp5 = vaddq_f16(_tmp56a, _tmp56b);
-                float16x8_t _tmp6 = vsubq_f16(_tmp56a, _tmp56b);
-                float16x8_t _tmp7 = vfmaq_laneq_f16(vsubq_f16(_r7, _r1), vsubq_f16(_r3, _r5), _coeffs, 0);
-
-                vst1q_f16(tmp[0][m], _tmp0);
-                vst1q_f16(tmp[1][m], _tmp1);
-                vst1q_f16(tmp[2][m], _tmp2);
-                vst1q_f16(tmp[3][m], _tmp3);
-                vst1q_f16(tmp[4][m], _tmp4);
-                vst1q_f16(tmp[5][m], _tmp5);
-                vst1q_f16(tmp[6][m], _tmp6);
-                vst1q_f16(tmp[7][m], _tmp7);
-
-                r0 += w * elempack;
-            }
-
-            __fp16* p0 = (__fp16*)B + kk * max_jj * 64 + jj * 8;
-            __fp16* p1 = p0 + max_jj * 8;
-            __fp16* p2 = p0 + max_jj * 8 * 2;
-            __fp16* p3 = p0 + max_jj * 8 * 3;
-            __fp16* p4 = p0 + max_jj * 8 * 4;
-            __fp16* p5 = p0 + max_jj * 8 * 5;
-            __fp16* p6 = p0 + max_jj * 8 * 6;
-            __fp16* p7 = p0 + max_jj * 8 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float16x8_t _r0 = vld1q_f16(tmp[m][0]);
-                float16x8_t _r1 = vld1q_f16(tmp[m][1]);
-                float16x8_t _r2 = vld1q_f16(tmp[m][2]);
-                float16x8_t _r3 = vld1q_f16(tmp[m][3]);
-                float16x8_t _r4 = vld1q_f16(tmp[m][4]);
-                float16x8_t _r5 = vld1q_f16(tmp[m][5]);
-                float16x8_t _r6 = vld1q_f16(tmp[m][6]);
-                float16x8_t _r7 = vld1q_f16(tmp[m][7]);
-
-                float16x8_t _tmp12a = vfmaq_laneq_f16(vaddq_f16(_r2, _r6), _r4, _coeffs, 1);
-                float16x8_t _tmp12b = vfmaq_laneq_f16(vaddq_f16(_r1, _r5), _r3, _coeffs, 1);
-                float16x8_t _tmp34a = vfmaq_laneq_f16(vfmaq_laneq_f16(_r6, _r2, _coeffs, 3), _r4, _coeffs, 2);
-                float16x8_t _tmp34b = vfmaq_laneq_f16(vfmaq_laneq_f16(vmulq_laneq_f16(_r1, _coeffs, 5), _r3, _coeffs, 4), _r5, _coeffs, 6);
-                float16x8_t _tmp56a = vfmaq_laneq_f16(_r6, vfmaq_laneq_f16(_r2, _r4, _coeffs, 2), _coeffs, 7);
-                float16x8_t _tmp56b = vfmaq_laneq_f16(vfmaq_laneq_f16(vmulq_laneq_f16(_r1, _coeffs, 6), _r3, _coeffs, 4), _r5, _coeffs, 5);
-
-                float16x8_t _tmp0 = vfmaq_laneq_f16(vsubq_f16(_r0, _r6), vsubq_f16(_r4, _r2), _coeffs, 0);
-                float16x8_t _tmp1 = vaddq_f16(_tmp12a, _tmp12b);
-                float16x8_t _tmp2 = vsubq_f16(_tmp12a, _tmp12b);
-                float16x8_t _tmp3 = vaddq_f16(_tmp34a, _tmp34b);
-                float16x8_t _tmp4 = vsubq_f16(_tmp34a, _tmp34b);
-                float16x8_t _tmp5 = vaddq_f16(_tmp56a, _tmp56b);
-                float16x8_t _tmp6 = vsubq_f16(_tmp56a, _tmp56b);
-                float16x8_t _tmp7 = vfmaq_laneq_f16(vsubq_f16(_r7, _r1), vsubq_f16(_r3, _r5), _coeffs, 0);
-
-                vst1q_f16(p0, _tmp0);
-                vst1q_f16(p1, _tmp1);
-                vst1q_f16(p2, _tmp2);
-                vst1q_f16(p3, _tmp3);
-                vst1q_f16(p4, _tmp4);
-                vst1q_f16(p5, _tmp5);
-                vst1q_f16(p6, _tmp6);
-                vst1q_f16(p7, _tmp7);
-
-                p0 += max_jj * 8 * 8;
-                p1 += max_jj * 8 * 8;
-                p2 += max_jj * 8 * 8;
-                p3 += max_jj * 8 * 8;
-                p4 += max_jj * 8 * 8;
-                p5 += max_jj * 8 * 8;
-                p6 += max_jj * 8 * 8;
-                p7 += max_jj * 8 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 4;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        __fp16 tmp[8][8][4];
-
-        const __fp16 coeffs[8] = {5.25f, -4.25f, -1.25f, 0.25f, -2.5f, 0.5f, 2.f, 4.f};
-        float16x8_t _coeffs = vld1q_f16(coeffs);
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = bottom_blob.channel((k + kk) / elempack).row<const __fp16>(ti * 6) + (tj * 6) * elempack;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float16x4_t _r0 = vdup_n_f16(0.f);
-                float16x4_t _r1 = vdup_n_f16(0.f);
-                float16x4_t _r2 = vdup_n_f16(0.f);
-                float16x4_t _r3 = vdup_n_f16(0.f);
-                float16x4_t _r4 = vdup_n_f16(0.f);
-                float16x4_t _r5 = vdup_n_f16(0.f);
-                float16x4_t _r6 = vdup_n_f16(0.f);
-                float16x4_t _r7 = vdup_n_f16(0.f);
-
-                if (ti * 6 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        _r0 = vld1_f16(r0);
-                        if (tj * 6 + 1 < w) _r1 = vld1_f16(r0 + 4);
-                        if (tj * 6 + 2 < w) _r2 = vld1_f16(r0 + 8);
-                        if (tj * 6 + 3 < w) _r3 = vld1_f16(r0 + 12);
-                        if (tj * 6 + 4 < w) _r4 = vld1_f16(r0 + 16);
-                        if (tj * 6 + 5 < w) _r5 = vld1_f16(r0 + 20);
-                        if (tj * 6 + 6 < w) _r6 = vld1_f16(r0 + 24);
-                        if (tj * 6 + 7 < w) _r7 = vld1_f16(r0 + 28);
-                    }
-                    if (elempack == 1)
-                    {
-                        const __fp16* r1 = r0 + N;
-                        const __fp16* r2 = r0 + N * 2;
-                        const __fp16* r3 = r0 + N * 3;
-
-                        float16x4_t _t0 = vld1_f16(r0);
-                        float16x4_t _t1 = vld1_f16(r1);
-                        float16x4_t _t2 = vld1_f16(r2);
-                        float16x4_t _t3 = vld1_f16(r3);
-
-                        transpose4x4_ph(_t0, _t1, _t2, _t3);
-
-                        _r0 = _t0;
-                        if (tj * 6 + 1 < w) _r1 = _t1;
-                        if (tj * 6 + 2 < w) _r2 = _t2;
-                        if (tj * 6 + 3 < w) _r3 = _t3;
-                        if (tj * 6 + 4 < w)
-                        {
-                            _t0 = vld1_f16(r0 + 4);
-                            _t1 = vld1_f16(r1 + 4);
-                            _t2 = vld1_f16(r2 + 4);
-                            _t3 = vld1_f16(r3 + 4);
-
-                            transpose4x4_ph(_t0, _t1, _t2, _t3);
-
-                            _r4 = _t0;
-                            if (tj * 6 + 5 < w) _r5 = _t1;
-                            if (tj * 6 + 6 < w) _r6 = _t2;
-                            if (tj * 6 + 7 < w) _r7 = _t3;
-                        }
-                    }
-                }
-
-                float16x4_t _tmp12a = vfma_laneq_f16(vadd_f16(_r2, _r6), _r4, _coeffs, 1);
-                float16x4_t _tmp12b = vfma_laneq_f16(vadd_f16(_r1, _r5), _r3, _coeffs, 1);
-                float16x4_t _tmp34a = vfma_laneq_f16(vfma_laneq_f16(_r6, _r2, _coeffs, 3), _r4, _coeffs, 2);
-                float16x4_t _tmp34b = vfma_laneq_f16(vfma_laneq_f16(vmul_laneq_f16(_r1, _coeffs, 5), _r3, _coeffs, 4), _r5, _coeffs, 6);
-                float16x4_t _tmp56a = vfma_laneq_f16(_r6, vfma_laneq_f16(_r2, _r4, _coeffs, 2), _coeffs, 7);
-                float16x4_t _tmp56b = vfma_laneq_f16(vfma_laneq_f16(vmul_laneq_f16(_r1, _coeffs, 6), _r3, _coeffs, 4), _r5, _coeffs, 5);
-
-                float16x4_t _tmp0 = vfma_laneq_f16(vsub_f16(_r0, _r6), vsub_f16(_r4, _r2), _coeffs, 0);
-                float16x4_t _tmp1 = vadd_f16(_tmp12a, _tmp12b);
-                float16x4_t _tmp2 = vsub_f16(_tmp12a, _tmp12b);
-                float16x4_t _tmp3 = vadd_f16(_tmp34a, _tmp34b);
-                float16x4_t _tmp4 = vsub_f16(_tmp34a, _tmp34b);
-                float16x4_t _tmp5 = vadd_f16(_tmp56a, _tmp56b);
-                float16x4_t _tmp6 = vsub_f16(_tmp56a, _tmp56b);
-                float16x4_t _tmp7 = vfma_laneq_f16(vsub_f16(_r7, _r1), vsub_f16(_r3, _r5), _coeffs, 0);
-
-                vst1_f16(tmp[0][m], _tmp0);
-                vst1_f16(tmp[1][m], _tmp1);
-                vst1_f16(tmp[2][m], _tmp2);
-                vst1_f16(tmp[3][m], _tmp3);
-                vst1_f16(tmp[4][m], _tmp4);
-                vst1_f16(tmp[5][m], _tmp5);
-                vst1_f16(tmp[6][m], _tmp6);
-                vst1_f16(tmp[7][m], _tmp7);
-
-                r0 += w * elempack;
-            }
-
-            __fp16* p0 = (__fp16*)B + kk * max_jj * 64 + jj * 4;
-            __fp16* p1 = p0 + max_jj * 4;
-            __fp16* p2 = p0 + max_jj * 4 * 2;
-            __fp16* p3 = p0 + max_jj * 4 * 3;
-            __fp16* p4 = p0 + max_jj * 4 * 4;
-            __fp16* p5 = p0 + max_jj * 4 * 5;
-            __fp16* p6 = p0 + max_jj * 4 * 6;
-            __fp16* p7 = p0 + max_jj * 4 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float16x4_t _r0 = vld1_f16(tmp[m][0]);
-                float16x4_t _r1 = vld1_f16(tmp[m][1]);
-                float16x4_t _r2 = vld1_f16(tmp[m][2]);
-                float16x4_t _r3 = vld1_f16(tmp[m][3]);
-                float16x4_t _r4 = vld1_f16(tmp[m][4]);
-                float16x4_t _r5 = vld1_f16(tmp[m][5]);
-                float16x4_t _r6 = vld1_f16(tmp[m][6]);
-                float16x4_t _r7 = vld1_f16(tmp[m][7]);
-
-                float16x4_t _tmp12a = vfma_laneq_f16(vadd_f16(_r2, _r6), _r4, _coeffs, 1);
-                float16x4_t _tmp12b = vfma_laneq_f16(vadd_f16(_r1, _r5), _r3, _coeffs, 1);
-                float16x4_t _tmp34a = vfma_laneq_f16(vfma_laneq_f16(_r6, _r2, _coeffs, 3), _r4, _coeffs, 2);
-                float16x4_t _tmp34b = vfma_laneq_f16(vfma_laneq_f16(vmul_laneq_f16(_r1, _coeffs, 5), _r3, _coeffs, 4), _r5, _coeffs, 6);
-                float16x4_t _tmp56a = vfma_laneq_f16(_r6, vfma_laneq_f16(_r2, _r4, _coeffs, 2), _coeffs, 7);
-                float16x4_t _tmp56b = vfma_laneq_f16(vfma_laneq_f16(vmul_laneq_f16(_r1, _coeffs, 6), _r3, _coeffs, 4), _r5, _coeffs, 5);
-
-                float16x4_t _tmp0 = vfma_laneq_f16(vsub_f16(_r0, _r6), vsub_f16(_r4, _r2), _coeffs, 0);
-                float16x4_t _tmp1 = vadd_f16(_tmp12a, _tmp12b);
-                float16x4_t _tmp2 = vsub_f16(_tmp12a, _tmp12b);
-                float16x4_t _tmp3 = vadd_f16(_tmp34a, _tmp34b);
-                float16x4_t _tmp4 = vsub_f16(_tmp34a, _tmp34b);
-                float16x4_t _tmp5 = vadd_f16(_tmp56a, _tmp56b);
-                float16x4_t _tmp6 = vsub_f16(_tmp56a, _tmp56b);
-                float16x4_t _tmp7 = vfma_laneq_f16(vsub_f16(_r7, _r1), vsub_f16(_r3, _r5), _coeffs, 0);
-
-                vst1_f16(p0, _tmp0);
-                vst1_f16(p1, _tmp1);
-                vst1_f16(p2, _tmp2);
-                vst1_f16(p3, _tmp3);
-                vst1_f16(p4, _tmp4);
-                vst1_f16(p5, _tmp5);
-                vst1_f16(p6, _tmp6);
-                vst1_f16(p7, _tmp7);
-
-                p0 += max_jj * 8 * 4;
-                p1 += max_jj * 8 * 4;
-                p2 += max_jj * 8 * 4;
-                p3 += max_jj * 8 * 4;
-                p4 += max_jj * 8 * 4;
-                p5 += max_jj * 8 * 4;
-                p6 += max_jj * 8 * 4;
-                p7 += max_jj * 8 * 4;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 4;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        __fp16 tmp[8][8][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = bottom_blob.channel(k + kk).row<const __fp16>(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 8; m++)
-            {
-                __fp16 r00 = 0.f;
-                __fp16 r01 = 0.f;
-                __fp16 r10 = 0.f;
-                __fp16 r11 = 0.f;
-                __fp16 r20 = 0.f;
-                __fp16 r21 = 0.f;
-                __fp16 r30 = 0.f;
-                __fp16 r31 = 0.f;
-                __fp16 r40 = 0.f;
-                __fp16 r41 = 0.f;
-                __fp16 r50 = 0.f;
-                __fp16 r51 = 0.f;
-                __fp16 r60 = 0.f;
-                __fp16 r61 = 0.f;
-                __fp16 r70 = 0.f;
-                __fp16 r71 = 0.f;
-
-                if (ti * 6 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const __fp16* r1 = r0 + N;
-
-                        r00 = r0[0];
-                        r01 = r1[0];
-                        if (tj * 6 + 1 < w)
-                        {
-                            r10 = r0[1];
-                            r11 = r1[1];
-                        }
-                        if (tj * 6 + 2 < w)
-                        {
-                            r20 = r0[2];
-                            r21 = r1[2];
-                        }
-                        if (tj * 6 + 3 < w)
-                        {
-                            r30 = r0[3];
-                            r31 = r1[3];
-                        }
-                        if (tj * 6 + 4 < w)
-                        {
-                            r40 = r0[4];
-                            r41 = r1[4];
-                        }
-                        if (tj * 6 + 5 < w)
-                        {
-                            r50 = r0[5];
-                            r51 = r1[5];
-                        }
-                        if (tj * 6 + 6 < w)
-                        {
-                            r60 = r0[6];
-                            r61 = r1[6];
-                        }
-                        if (tj * 6 + 7 < w)
-                        {
-                            r70 = r0[7];
-                            r71 = r1[7];
-                        }
-                    }
-                }
-
-                __fp16 tmp12a0 = r20 + r60 - r40 * (__fp16)4.25f;
-                __fp16 tmp12a1 = r21 + r61 - r41 * (__fp16)4.25f;
-                __fp16 tmp12b0 = r10 + r50 - r30 * (__fp16)4.25f;
-                __fp16 tmp12b1 = r11 + r51 - r31 * (__fp16)4.25f;
-                __fp16 tmp34a0 = r60 + r20 * (__fp16)0.25f - r40 * (__fp16)1.25f;
-                __fp16 tmp34a1 = r61 + r21 * (__fp16)0.25f - r41 * (__fp16)1.25f;
-                __fp16 tmp34b0 = r10 * (__fp16)0.5f - r30 * (__fp16)2.5f + r50 * (__fp16)2.f;
-                __fp16 tmp34b1 = r11 * (__fp16)0.5f - r31 * (__fp16)2.5f + r51 * (__fp16)2.f;
-                __fp16 tmp56a0 = r20 * (__fp16)4.f - r40 * (__fp16)5.f + r60;
-                __fp16 tmp56a1 = r21 * (__fp16)4.f - r41 * (__fp16)5.f + r61;
-                __fp16 tmp56b0 = r10 * (__fp16)2.f - r30 * (__fp16)2.5f + r50 * (__fp16)0.5f;
-                __fp16 tmp56b1 = r11 * (__fp16)2.f - r31 * (__fp16)2.5f + r51 * (__fp16)0.5f;
-
-                tmp[0][m][0] = r00 - r60 + (r40 - r20) * (__fp16)5.25f;
-                tmp[0][m][1] = r01 - r61 + (r41 - r21) * (__fp16)5.25f;
-                tmp[1][m][0] = tmp12a0 + tmp12b0;
-                tmp[1][m][1] = tmp12a1 + tmp12b1;
-                tmp[2][m][0] = tmp12a0 - tmp12b0;
-                tmp[2][m][1] = tmp12a1 - tmp12b1;
-                tmp[3][m][0] = tmp34a0 + tmp34b0;
-                tmp[3][m][1] = tmp34a1 + tmp34b1;
-                tmp[4][m][0] = tmp34a0 - tmp34b0;
-                tmp[4][m][1] = tmp34a1 - tmp34b1;
-                tmp[5][m][0] = tmp56a0 + tmp56b0;
-                tmp[5][m][1] = tmp56a1 + tmp56b1;
-                tmp[6][m][0] = tmp56a0 - tmp56b0;
-                tmp[6][m][1] = tmp56a1 - tmp56b1;
-                tmp[7][m][0] = r70 - r10 + (r30 - r50) * (__fp16)5.25f;
-                tmp[7][m][1] = r71 - r11 + (r31 - r51) * (__fp16)5.25f;
-
-                r0 += w;
-            }
-
-            __fp16* p0 = (__fp16*)B + kk * max_jj * 64 + jj * 2;
-            __fp16* p1 = p0 + max_jj * 2;
-            __fp16* p2 = p0 + max_jj * 2 * 2;
-            __fp16* p3 = p0 + max_jj * 2 * 3;
-            __fp16* p4 = p0 + max_jj * 2 * 4;
-            __fp16* p5 = p0 + max_jj * 2 * 5;
-            __fp16* p6 = p0 + max_jj * 2 * 6;
-            __fp16* p7 = p0 + max_jj * 2 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                __fp16 r00 = tmp[m][0][0];
-                __fp16 r01 = tmp[m][0][1];
-                __fp16 r10 = tmp[m][1][0];
-                __fp16 r11 = tmp[m][1][1];
-                __fp16 r20 = tmp[m][2][0];
-                __fp16 r21 = tmp[m][2][1];
-                __fp16 r30 = tmp[m][3][0];
-                __fp16 r31 = tmp[m][3][1];
-                __fp16 r40 = tmp[m][4][0];
-                __fp16 r41 = tmp[m][4][1];
-                __fp16 r50 = tmp[m][5][0];
-                __fp16 r51 = tmp[m][5][1];
-                __fp16 r60 = tmp[m][6][0];
-                __fp16 r61 = tmp[m][6][1];
-                __fp16 r70 = tmp[m][7][0];
-                __fp16 r71 = tmp[m][7][1];
-
-                __fp16 tmp12a0 = r20 + r60 - r40 * (__fp16)4.25f;
-                __fp16 tmp12a1 = r21 + r61 - r41 * (__fp16)4.25f;
-                __fp16 tmp12b0 = r10 + r50 - r30 * (__fp16)4.25f;
-                __fp16 tmp12b1 = r11 + r51 - r31 * (__fp16)4.25f;
-                __fp16 tmp34a0 = r60 + r20 * (__fp16)0.25f - r40 * (__fp16)1.25f;
-                __fp16 tmp34a1 = r61 + r21 * (__fp16)0.25f - r41 * (__fp16)1.25f;
-                __fp16 tmp34b0 = r10 * (__fp16)0.5f - r30 * (__fp16)2.5f + r50 * (__fp16)2.f;
-                __fp16 tmp34b1 = r11 * (__fp16)0.5f - r31 * (__fp16)2.5f + r51 * (__fp16)2.f;
-                __fp16 tmp56a0 = r20 * (__fp16)4.f - r40 * (__fp16)5.f + r60;
-                __fp16 tmp56a1 = r21 * (__fp16)4.f - r41 * (__fp16)5.f + r61;
-                __fp16 tmp56b0 = r10 * (__fp16)2.f - r30 * (__fp16)2.5f + r50 * (__fp16)0.5f;
-                __fp16 tmp56b1 = r11 * (__fp16)2.f - r31 * (__fp16)2.5f + r51 * (__fp16)0.5f;
-
-                p0[0] = r00 - r60 + (r40 - r20) * (__fp16)5.25f;
-                p0[1] = r01 - r61 + (r41 - r21) * (__fp16)5.25f;
-                p1[0] = tmp12a0 + tmp12b0;
-                p1[1] = tmp12a1 + tmp12b1;
-                p2[0] = tmp12a0 - tmp12b0;
-                p2[1] = tmp12a1 - tmp12b1;
-                p3[0] = tmp34a0 + tmp34b0;
-                p3[1] = tmp34a1 + tmp34b1;
-                p4[0] = tmp34a0 - tmp34b0;
-                p4[1] = tmp34a1 - tmp34b1;
-                p5[0] = tmp56a0 + tmp56b0;
-                p5[1] = tmp56a1 + tmp56b1;
-                p6[0] = tmp56a0 - tmp56b0;
-                p6[1] = tmp56a1 - tmp56b1;
-                p7[0] = r70 - r10 + (r30 - r50) * (__fp16)5.25f;
-                p7[1] = r71 - r11 + (r31 - r51) * (__fp16)5.25f;
-
-                p0 += max_jj * 8 * 2;
-                p1 += max_jj * 8 * 2;
-                p2 += max_jj * 8 * 2;
-                p3 += max_jj * 8 * 2;
-                p4 += max_jj * 8 * 2;
-                p5 += max_jj * 8 * 2;
-                p6 += max_jj * 8 * 2;
-                p7 += max_jj * 8 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        __fp16 tmp[8][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0123 = bottom_blob.channel(k + kk).row<const __fp16>(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 8; m++)
-            {
-                __fp16 r0 = 0.f;
-                __fp16 r1 = 0.f;
-                __fp16 r2 = 0.f;
-                __fp16 r3 = 0.f;
-                __fp16 r4 = 0.f;
-                __fp16 r5 = 0.f;
-                __fp16 r6 = 0.f;
-                __fp16 r7 = 0.f;
-
-                if (ti * 6 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = r0123[0];
-                        if (tj * 6 + 1 < w) r1 = r0123[1];
-                        if (tj * 6 + 2 < w) r2 = r0123[2];
-                        if (tj * 6 + 3 < w) r3 = r0123[3];
-                        if (tj * 6 + 4 < w) r4 = r0123[4];
-                        if (tj * 6 + 5 < w) r5 = r0123[5];
-                        if (tj * 6 + 6 < w) r6 = r0123[6];
-                        if (tj * 6 + 7 < w) r7 = r0123[7];
-                    }
-                }
-
-                __fp16 tmp12a = r2 + r6 - r4 * (__fp16)4.25f;
-                __fp16 tmp12b = r1 + r5 - r3 * (__fp16)4.25f;
-                __fp16 tmp34a = r6 + r2 * (__fp16)0.25f - r4 * (__fp16)1.25f;
-                __fp16 tmp34b = r1 * (__fp16)0.5f - r3 * (__fp16)2.5f + r5 * (__fp16)2.f;
-                __fp16 tmp56a = r2 * (__fp16)4.f - r4 * (__fp16)5.f + r6;
-                __fp16 tmp56b = r1 * (__fp16)2.f - r3 * (__fp16)2.5f + r5 * (__fp16)0.5f;
-
-                tmp[0][m] = r0 - r6 + (r4 - r2) * (__fp16)5.25f;
-                tmp[1][m] = tmp12a + tmp12b;
-                tmp[2][m] = tmp12a - tmp12b;
-                tmp[3][m] = tmp34a + tmp34b;
-                tmp[4][m] = tmp34a - tmp34b;
-                tmp[5][m] = tmp56a + tmp56b;
-                tmp[6][m] = tmp56a - tmp56b;
-                tmp[7][m] = r7 - r1 + (r3 - r5) * (__fp16)5.25f;
-
-                r0123 += w;
-            }
-
-            __fp16* p0 = (__fp16*)B + kk * max_jj * 64 + jj;
-            __fp16* p1 = p0 + max_jj;
-            __fp16* p2 = p0 + max_jj * 2;
-            __fp16* p3 = p0 + max_jj * 3;
-            __fp16* p4 = p0 + max_jj * 4;
-            __fp16* p5 = p0 + max_jj * 5;
-            __fp16* p6 = p0 + max_jj * 6;
-            __fp16* p7 = p0 + max_jj * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                __fp16 r0 = tmp[m][0];
-                __fp16 r1 = tmp[m][1];
-                __fp16 r2 = tmp[m][2];
-                __fp16 r3 = tmp[m][3];
-                __fp16 r4 = tmp[m][4];
-                __fp16 r5 = tmp[m][5];
-                __fp16 r6 = tmp[m][6];
-                __fp16 r7 = tmp[m][7];
-
-                __fp16 tmp12a = r2 + r6 - r4 * (__fp16)4.25f;
-                __fp16 tmp12b = r1 + r5 - r3 * (__fp16)4.25f;
-                __fp16 tmp34a = r6 + r2 * (__fp16)0.25f - r4 * (__fp16)1.25f;
-                __fp16 tmp34b = r1 * (__fp16)0.5f - r3 * (__fp16)2.5f + r5 * (__fp16)2.f;
-                __fp16 tmp56a = r2 * (__fp16)4.f - r4 * (__fp16)5.f + r6;
-                __fp16 tmp56b = r1 * (__fp16)2.f - r3 * (__fp16)2.5f + r5 * (__fp16)0.5f;
-
-                p0[0] = r0 - r6 + (r4 - r2) * (__fp16)5.25f;
-                p1[0] = tmp12a + tmp12b;
-                p2[0] = tmp12a - tmp12b;
-                p3[0] = tmp34a + tmp34b;
-                p4[0] = tmp34a - tmp34b;
-                p5[0] = tmp56a + tmp56b;
-                p6[0] = tmp56a - tmp56b;
-                p7[0] = r7 - r1 + (r3 - r5) * (__fp16)5.25f;
-
-                p0 += max_jj * 8;
-                p1 += max_jj * 8;
-                p2 += max_jj * 8;
-                p3 += max_jj * 8;
-                p4 += max_jj * 8;
-                p5 += max_jj * 8;
-                p6 += max_jj * 8;
-                p7 += max_jj * 8;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd63_transform_output_tile_fp16sa(const Mat& top_tile, Mat& top_blob, const Mat& bias, int i, int max_ii, int j, int max_jj)
-{
-    // const float otm[6][8] = {
-    //     {1.0f, 1.0f,  1.0f,  1.0f,  1.0f, 32.0f, 32.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f,  2.0f, -2.0f, 16.0f,-16.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f,  4.0f,  4.0f,  8.0f,  8.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f,  8.0f, -8.0f,  4.0f, -4.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f, 16.0f, 16.0f,  2.0f,  2.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 32.0f,-32.0f,  1.0f, -1.0f, 1.0f}
-    // };
-
-    const __fp16 coeffs[8] = {32.f, 16.f, 8.f, 4.f, 2.f, 0.f, 0.f, 0.f};
-    float16x8_t _coeffs = vld1q_f16(coeffs);
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 5) / 6;
-
-    const __fp16* biasptr = bias;
-
-    int ii = 0;
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        float16x8_t _bias0 = biasptr ? vld1q_f16(biasptr + i + ii) : vdupq_n_f16(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        __fp16 tmp[6][8][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = (const __fp16*)top_tile + ii * max_jj * 64 + jj * 8;
-            const __fp16* r1 = r0 + max_jj * 8;
-            const __fp16* r2 = r0 + max_jj * 8 * 2;
-            const __fp16* r3 = r0 + max_jj * 8 * 3;
-            const __fp16* r4 = r0 + max_jj * 8 * 4;
-            const __fp16* r5 = r0 + max_jj * 8 * 5;
-            const __fp16* r6 = r0 + max_jj * 8 * 6;
-            const __fp16* r7 = r0 + max_jj * 8 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float16x8_t _r0 = vld1q_f16(r0);
-                float16x8_t _r1 = vld1q_f16(r1);
-                float16x8_t _r2 = vld1q_f16(r2);
-                float16x8_t _r3 = vld1q_f16(r3);
-                float16x8_t _r4 = vld1q_f16(r4);
-                float16x8_t _r5 = vld1q_f16(r5);
-                float16x8_t _r6 = vld1q_f16(r6);
-                float16x8_t _r7 = vld1q_f16(r7);
-
-                float16x8_t _tmp024a = vaddq_f16(_r1, _r2);
-                float16x8_t _tmp135a = vsubq_f16(_r1, _r2);
-                float16x8_t _tmp024b = vaddq_f16(_r3, _r4);
-                float16x8_t _tmp135b = vsubq_f16(_r3, _r4);
-                float16x8_t _tmp024c = vaddq_f16(_r5, _r6);
-                float16x8_t _tmp135c = vsubq_f16(_r5, _r6);
-
-                float16x8_t _tmp0 = vaddq_f16(vaddq_f16(_r0, _tmp024a), vfmaq_laneq_f16(_tmp024b, _tmp024c, _coeffs, 0));
-                float16x8_t _tmp1 = vfmaq_laneq_f16(vfmaq_laneq_f16(_tmp135a, _tmp135b, _coeffs, 4), _tmp135c, _coeffs, 1);
-                float16x8_t _tmp2 = vfmaq_laneq_f16(vfmaq_laneq_f16(_tmp024a, _tmp024b, _coeffs, 3), _tmp024c, _coeffs, 2);
-                float16x8_t _tmp3 = vfmaq_laneq_f16(vfmaq_laneq_f16(_tmp135a, _tmp135b, _coeffs, 2), _tmp135c, _coeffs, 3);
-                float16x8_t _tmp4 = vfmaq_laneq_f16(vfmaq_laneq_f16(_tmp024a, _tmp024b, _coeffs, 1), _tmp024c, _coeffs, 4);
-                float16x8_t _tmp5 = vaddq_f16(vaddq_f16(_r7, _tmp135a), vfmaq_laneq_f16(_tmp135c, _tmp135b, _coeffs, 0));
-
-                vst1q_f16(tmp[0][m], _tmp0);
-                vst1q_f16(tmp[1][m], _tmp1);
-                vst1q_f16(tmp[2][m], _tmp2);
-                vst1q_f16(tmp[3][m], _tmp3);
-                vst1q_f16(tmp[4][m], _tmp4);
-                vst1q_f16(tmp[5][m], _tmp5);
-
-                r0 += max_jj * 8 * 8;
-                r1 += max_jj * 8 * 8;
-                r2 += max_jj * 8 * 8;
-                r3 += max_jj * 8 * 8;
-                r4 += max_jj * 8 * 8;
-                r5 += max_jj * 8 * 8;
-                r6 += max_jj * 8 * 8;
-                r7 += max_jj * 8 * 8;
-            }
-
-            __fp16* outptr0 = top_blob.channel((i + ii) / out_elempack).row<__fp16>(ti * 6) + (tj * 6) * out_elempack;
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-                float16x8_t _r0 = vld1q_f16(tmp[m][0]);
-                float16x8_t _r1 = vld1q_f16(tmp[m][1]);
-                float16x8_t _r2 = vld1q_f16(tmp[m][2]);
-                float16x8_t _r3 = vld1q_f16(tmp[m][3]);
-                float16x8_t _r4 = vld1q_f16(tmp[m][4]);
-                float16x8_t _r5 = vld1q_f16(tmp[m][5]);
-                float16x8_t _r6 = vld1q_f16(tmp[m][6]);
-                float16x8_t _r7 = vld1q_f16(tmp[m][7]);
-
-                float16x8_t _tmp024a = vaddq_f16(_r1, _r2);
-                float16x8_t _tmp135a = vsubq_f16(_r1, _r2);
-                float16x8_t _tmp024b = vaddq_f16(_r3, _r4);
-                float16x8_t _tmp135b = vsubq_f16(_r3, _r4);
-                float16x8_t _tmp024c = vaddq_f16(_r5, _r6);
-                float16x8_t _tmp135c = vsubq_f16(_r5, _r6);
-
-                float16x8_t _tmp0 = vaddq_f16(_bias0, vaddq_f16(vaddq_f16(_r0, _tmp024a), vfmaq_laneq_f16(_tmp024b, _tmp024c, _coeffs, 0)));
-                float16x8_t _tmp1 = vaddq_f16(_bias0, vfmaq_laneq_f16(vfmaq_laneq_f16(_tmp135a, _tmp135b, _coeffs, 4), _tmp135c, _coeffs, 1));
-                float16x8_t _tmp2 = vaddq_f16(_bias0, vfmaq_laneq_f16(vfmaq_laneq_f16(_tmp024a, _tmp024b, _coeffs, 3), _tmp024c, _coeffs, 2));
-                float16x8_t _tmp3 = vaddq_f16(_bias0, vfmaq_laneq_f16(vfmaq_laneq_f16(_tmp135a, _tmp135b, _coeffs, 2), _tmp135c, _coeffs, 3));
-                float16x8_t _tmp4 = vaddq_f16(_bias0, vfmaq_laneq_f16(vfmaq_laneq_f16(_tmp024a, _tmp024b, _coeffs, 1), _tmp024c, _coeffs, 4));
-                float16x8_t _tmp5 = vaddq_f16(_bias0, vaddq_f16(vaddq_f16(_r7, _tmp135a), vfmaq_laneq_f16(_tmp135c, _tmp135b, _coeffs, 0)));
-
-                if (out_elempack == 8)
-                {
-                    vst1q_f16(outptr0, _tmp0);
-                    if (tj * 6 + 1 < outw) vst1q_f16(outptr0 + 8, _tmp1);
-                    if (tj * 6 + 2 < outw) vst1q_f16(outptr0 + 16, _tmp2);
-                    if (tj * 6 + 3 < outw) vst1q_f16(outptr0 + 24, _tmp3);
-                    if (tj * 6 + 4 < outw) vst1q_f16(outptr0 + 32, _tmp4);
-                    if (tj * 6 + 5 < outw) vst1q_f16(outptr0 + 40, _tmp5);
-                }
-                if (out_elempack == 4)
-                {
-                    __fp16* outptr1 = outptr0 + N;
-
-                    vst1_f16(outptr0, vget_low_f16(_tmp0));
-                    vst1_f16(outptr1, vget_high_f16(_tmp0));
-                    if (tj * 6 + 1 < outw)
-                    {
-                        vst1_f16(outptr0 + 4, vget_low_f16(_tmp1));
-                        vst1_f16(outptr1 + 4, vget_high_f16(_tmp1));
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        vst1_f16(outptr0 + 8, vget_low_f16(_tmp2));
-                        vst1_f16(outptr1 + 8, vget_high_f16(_tmp2));
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        vst1_f16(outptr0 + 12, vget_low_f16(_tmp3));
-                        vst1_f16(outptr1 + 12, vget_high_f16(_tmp3));
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        vst1_f16(outptr0 + 16, vget_low_f16(_tmp4));
-                        vst1_f16(outptr1 + 16, vget_high_f16(_tmp4));
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        vst1_f16(outptr0 + 20, vget_low_f16(_tmp5));
-                        vst1_f16(outptr1 + 20, vget_high_f16(_tmp5));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    __fp16 tmp0[8];
-                    __fp16 tmp1[8];
-                    __fp16 tmp2[8];
-                    __fp16 tmp3[8];
-                    __fp16 tmp4[8];
-                    __fp16 tmp5[8];
-                    vst1q_f16(tmp0, _tmp0);
-                    vst1q_f16(tmp1, _tmp1);
-                    vst1q_f16(tmp2, _tmp2);
-                    vst1q_f16(tmp3, _tmp3);
-                    vst1q_f16(tmp4, _tmp4);
-                    vst1q_f16(tmp5, _tmp5);
-
-                    __fp16* outptr1 = outptr0 + N;
-                    __fp16* outptr2 = outptr0 + N * 2;
-                    __fp16* outptr3 = outptr0 + N * 3;
-                    __fp16* outptr4 = outptr0 + N * 4;
-                    __fp16* outptr5 = outptr0 + N * 5;
-                    __fp16* outptr6 = outptr0 + N * 6;
-                    __fp16* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-                    if (tj * 6 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                        outptr4[2] = tmp2[4];
-                        outptr5[2] = tmp2[5];
-                        outptr6[2] = tmp2[6];
-                        outptr7[2] = tmp2[7];
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                        outptr4[3] = tmp3[4];
-                        outptr5[3] = tmp3[5];
-                        outptr6[3] = tmp3[6];
-                        outptr7[3] = tmp3[7];
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        outptr0[4] = tmp4[0];
-                        outptr1[4] = tmp4[1];
-                        outptr2[4] = tmp4[2];
-                        outptr3[4] = tmp4[3];
-                        outptr4[4] = tmp4[4];
-                        outptr5[4] = tmp4[5];
-                        outptr6[4] = tmp4[6];
-                        outptr7[4] = tmp4[7];
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        outptr0[5] = tmp5[0];
-                        outptr1[5] = tmp5[1];
-                        outptr2[5] = tmp5[2];
-                        outptr3[5] = tmp5[3];
-                        outptr4[5] = tmp5[4];
-                        outptr5[5] = tmp5[5];
-                        outptr6[5] = tmp5[6];
-                        outptr7[5] = tmp5[7];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        float16x4_t _bias0 = biasptr ? vld1_f16(biasptr + i + ii) : vdup_n_f16(0.f);
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        __fp16 tmp[6][8][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = (const __fp16*)top_tile + ii * max_jj * 64 + jj * 4;
-            const __fp16* r1 = r0 + max_jj * 4;
-            const __fp16* r2 = r0 + max_jj * 4 * 2;
-            const __fp16* r3 = r0 + max_jj * 4 * 3;
-            const __fp16* r4 = r0 + max_jj * 4 * 4;
-            const __fp16* r5 = r0 + max_jj * 4 * 5;
-            const __fp16* r6 = r0 + max_jj * 4 * 6;
-            const __fp16* r7 = r0 + max_jj * 4 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float16x4_t _r0 = vld1_f16(r0);
-                float16x4_t _r1 = vld1_f16(r1);
-                float16x4_t _r2 = vld1_f16(r2);
-                float16x4_t _r3 = vld1_f16(r3);
-                float16x4_t _r4 = vld1_f16(r4);
-                float16x4_t _r5 = vld1_f16(r5);
-                float16x4_t _r6 = vld1_f16(r6);
-                float16x4_t _r7 = vld1_f16(r7);
-
-                float16x4_t _tmp024a = vadd_f16(_r1, _r2);
-                float16x4_t _tmp135a = vsub_f16(_r1, _r2);
-                float16x4_t _tmp024b = vadd_f16(_r3, _r4);
-                float16x4_t _tmp135b = vsub_f16(_r3, _r4);
-                float16x4_t _tmp024c = vadd_f16(_r5, _r6);
-                float16x4_t _tmp135c = vsub_f16(_r5, _r6);
-
-                float16x4_t _tmp0 = vadd_f16(vadd_f16(_r0, _tmp024a), vfma_laneq_f16(_tmp024b, _tmp024c, _coeffs, 0));
-                float16x4_t _tmp1 = vfma_laneq_f16(vfma_laneq_f16(_tmp135a, _tmp135b, _coeffs, 4), _tmp135c, _coeffs, 1);
-                float16x4_t _tmp2 = vfma_laneq_f16(vfma_laneq_f16(_tmp024a, _tmp024b, _coeffs, 3), _tmp024c, _coeffs, 2);
-                float16x4_t _tmp3 = vfma_laneq_f16(vfma_laneq_f16(_tmp135a, _tmp135b, _coeffs, 2), _tmp135c, _coeffs, 3);
-                float16x4_t _tmp4 = vfma_laneq_f16(vfma_laneq_f16(_tmp024a, _tmp024b, _coeffs, 1), _tmp024c, _coeffs, 4);
-                float16x4_t _tmp5 = vadd_f16(vadd_f16(_r7, _tmp135a), vfma_laneq_f16(_tmp135c, _tmp135b, _coeffs, 0));
-
-                vst1_f16(tmp[0][m], _tmp0);
-                vst1_f16(tmp[1][m], _tmp1);
-                vst1_f16(tmp[2][m], _tmp2);
-                vst1_f16(tmp[3][m], _tmp3);
-                vst1_f16(tmp[4][m], _tmp4);
-                vst1_f16(tmp[5][m], _tmp5);
-
-                r0 += max_jj * 8 * 4;
-                r1 += max_jj * 8 * 4;
-                r2 += max_jj * 8 * 4;
-                r3 += max_jj * 8 * 4;
-                r4 += max_jj * 8 * 4;
-                r5 += max_jj * 8 * 4;
-                r6 += max_jj * 8 * 4;
-                r7 += max_jj * 8 * 4;
-            }
-
-            __fp16* outptr0 = top_blob.channel((i + ii) / out_elempack).row<__fp16>(ti * 6) + (tj * 6) * out_elempack;
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-                float16x4_t _r0 = vld1_f16(tmp[m][0]);
-                float16x4_t _r1 = vld1_f16(tmp[m][1]);
-                float16x4_t _r2 = vld1_f16(tmp[m][2]);
-                float16x4_t _r3 = vld1_f16(tmp[m][3]);
-                float16x4_t _r4 = vld1_f16(tmp[m][4]);
-                float16x4_t _r5 = vld1_f16(tmp[m][5]);
-                float16x4_t _r6 = vld1_f16(tmp[m][6]);
-                float16x4_t _r7 = vld1_f16(tmp[m][7]);
-
-                float16x4_t _tmp024a = vadd_f16(_r1, _r2);
-                float16x4_t _tmp135a = vsub_f16(_r1, _r2);
-                float16x4_t _tmp024b = vadd_f16(_r3, _r4);
-                float16x4_t _tmp135b = vsub_f16(_r3, _r4);
-                float16x4_t _tmp024c = vadd_f16(_r5, _r6);
-                float16x4_t _tmp135c = vsub_f16(_r5, _r6);
-
-                float16x4_t _tmp0 = vadd_f16(_bias0, vadd_f16(vadd_f16(_r0, _tmp024a), vfma_laneq_f16(_tmp024b, _tmp024c, _coeffs, 0)));
-                float16x4_t _tmp1 = vadd_f16(_bias0, vfma_laneq_f16(vfma_laneq_f16(_tmp135a, _tmp135b, _coeffs, 4), _tmp135c, _coeffs, 1));
-                float16x4_t _tmp2 = vadd_f16(_bias0, vfma_laneq_f16(vfma_laneq_f16(_tmp024a, _tmp024b, _coeffs, 3), _tmp024c, _coeffs, 2));
-                float16x4_t _tmp3 = vadd_f16(_bias0, vfma_laneq_f16(vfma_laneq_f16(_tmp135a, _tmp135b, _coeffs, 2), _tmp135c, _coeffs, 3));
-                float16x4_t _tmp4 = vadd_f16(_bias0, vfma_laneq_f16(vfma_laneq_f16(_tmp024a, _tmp024b, _coeffs, 1), _tmp024c, _coeffs, 4));
-                float16x4_t _tmp5 = vadd_f16(_bias0, vadd_f16(vadd_f16(_r7, _tmp135a), vfma_laneq_f16(_tmp135c, _tmp135b, _coeffs, 0)));
-
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr0, _tmp0);
-                    if (tj * 6 + 1 < outw) vst1_f16(outptr0 + 4, _tmp1);
-                    if (tj * 6 + 2 < outw) vst1_f16(outptr0 + 8, _tmp2);
-                    if (tj * 6 + 3 < outw) vst1_f16(outptr0 + 12, _tmp3);
-                    if (tj * 6 + 4 < outw) vst1_f16(outptr0 + 16, _tmp4);
-                    if (tj * 6 + 5 < outw) vst1_f16(outptr0 + 20, _tmp5);
-                }
-                if (out_elempack == 1)
-                {
-                    __fp16 tmp0[4];
-                    __fp16 tmp1[4];
-                    __fp16 tmp2[4];
-                    __fp16 tmp3[4];
-                    __fp16 tmp4[4];
-                    __fp16 tmp5[4];
-                    vst1_f16(tmp0, _tmp0);
-                    vst1_f16(tmp1, _tmp1);
-                    vst1_f16(tmp2, _tmp2);
-                    vst1_f16(tmp3, _tmp3);
-                    vst1_f16(tmp4, _tmp4);
-                    vst1_f16(tmp5, _tmp5);
-
-                    __fp16* outptr1 = outptr0 + N;
-                    __fp16* outptr2 = outptr0 + N * 2;
-                    __fp16* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    if (tj * 6 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        outptr0[4] = tmp4[0];
-                        outptr1[4] = tmp4[1];
-                        outptr2[4] = tmp4[2];
-                        outptr3[4] = tmp4[3];
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        outptr0[5] = tmp5[0];
-                        outptr1[5] = tmp5[1];
-                        outptr2[5] = tmp5[2];
-                        outptr3[5] = tmp5[3];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        __fp16 bias0 = biasptr ? biasptr[i + ii] : 0.f;
-        __fp16 bias1 = biasptr ? biasptr[i + ii + 1] : 0.f;
-
-#ifdef _MSC_VER
-        __declspec(align(8))
-#else
-        __attribute__((aligned(8)))
-#endif
-        __fp16 tmp[6][8][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = (const __fp16*)top_tile + ii * max_jj * 64 + jj * 2;
-            const __fp16* r1 = r0 + max_jj * 2;
-            const __fp16* r2 = r0 + max_jj * 2 * 2;
-            const __fp16* r3 = r0 + max_jj * 2 * 3;
-            const __fp16* r4 = r0 + max_jj * 2 * 4;
-            const __fp16* r5 = r0 + max_jj * 2 * 5;
-            const __fp16* r6 = r0 + max_jj * 2 * 6;
-            const __fp16* r7 = r0 + max_jj * 2 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                __fp16 tmp024a0 = r1[0] + r2[0];
-                __fp16 tmp024a1 = r1[1] + r2[1];
-                __fp16 tmp135a0 = r1[0] - r2[0];
-                __fp16 tmp135a1 = r1[1] - r2[1];
-                __fp16 tmp024b0 = r3[0] + r4[0];
-                __fp16 tmp024b1 = r3[1] + r4[1];
-                __fp16 tmp135b0 = r3[0] - r4[0];
-                __fp16 tmp135b1 = r3[1] - r4[1];
-                __fp16 tmp024c0 = r5[0] + r6[0];
-                __fp16 tmp024c1 = r5[1] + r6[1];
-                __fp16 tmp135c0 = r5[0] - r6[0];
-                __fp16 tmp135c1 = r5[1] - r6[1];
-
-                tmp[0][m][0] = r0[0] + tmp024a0 + tmp024b0 + tmp024c0 * (__fp16)32;
-                tmp[0][m][1] = r0[1] + tmp024a1 + tmp024b1 + tmp024c1 * (__fp16)32;
-                tmp[1][m][0] = tmp135a0 + tmp135b0 + tmp135b0 + tmp135c0 * (__fp16)16;
-                tmp[1][m][1] = tmp135a1 + tmp135b1 + tmp135b1 + tmp135c1 * (__fp16)16;
-                tmp[2][m][0] = tmp024a0 + tmp024b0 * (__fp16)4 + tmp024c0 * (__fp16)8;
-                tmp[2][m][1] = tmp024a1 + tmp024b1 * (__fp16)4 + tmp024c1 * (__fp16)8;
-                tmp[3][m][0] = tmp135a0 + tmp135b0 * (__fp16)8 + tmp135c0 * (__fp16)4;
-                tmp[3][m][1] = tmp135a1 + tmp135b1 * (__fp16)8 + tmp135c1 * (__fp16)4;
-                tmp[4][m][0] = tmp024a0 + tmp024b0 * (__fp16)16 + tmp024c0 + tmp024c0;
-                tmp[4][m][1] = tmp024a1 + tmp024b1 * (__fp16)16 + tmp024c1 + tmp024c1;
-                tmp[5][m][0] = r7[0] + tmp135a0 + tmp135b0 * (__fp16)32 + tmp135c0;
-                tmp[5][m][1] = r7[1] + tmp135a1 + tmp135b1 * (__fp16)32 + tmp135c1;
-
-                r0 += max_jj * 8 * 2;
-                r1 += max_jj * 8 * 2;
-                r2 += max_jj * 8 * 2;
-                r3 += max_jj * 8 * 2;
-                r4 += max_jj * 8 * 2;
-                r5 += max_jj * 8 * 2;
-                r6 += max_jj * 8 * 2;
-                r7 += max_jj * 8 * 2;
-            }
-
-            __fp16* outptr0 = top_blob.channel(i + ii).row<__fp16>(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-                __fp16 r00 = tmp[m][0][0];
-                __fp16 r01 = tmp[m][0][1];
-                __fp16 r10 = tmp[m][1][0];
-                __fp16 r11 = tmp[m][1][1];
-                __fp16 r20 = tmp[m][2][0];
-                __fp16 r21 = tmp[m][2][1];
-                __fp16 r30 = tmp[m][3][0];
-                __fp16 r31 = tmp[m][3][1];
-                __fp16 r40 = tmp[m][4][0];
-                __fp16 r41 = tmp[m][4][1];
-                __fp16 r50 = tmp[m][5][0];
-                __fp16 r51 = tmp[m][5][1];
-                __fp16 r60 = tmp[m][6][0];
-                __fp16 r61 = tmp[m][6][1];
-                __fp16 r70 = tmp[m][7][0];
-                __fp16 r71 = tmp[m][7][1];
-
-                __fp16 tmp024a0 = r10 + r20;
-                __fp16 tmp024a1 = r11 + r21;
-                __fp16 tmp135a0 = r10 - r20;
-                __fp16 tmp135a1 = r11 - r21;
-                __fp16 tmp024b0 = r30 + r40;
-                __fp16 tmp024b1 = r31 + r41;
-                __fp16 tmp135b0 = r30 - r40;
-                __fp16 tmp135b1 = r31 - r41;
-                __fp16 tmp024c0 = r50 + r60;
-                __fp16 tmp024c1 = r51 + r61;
-                __fp16 tmp135c0 = r50 - r60;
-                __fp16 tmp135c1 = r51 - r61;
-
-                __fp16 tmp00 = bias0 + r00 + tmp024a0 + tmp024b0 + tmp024c0 * (__fp16)32;
-                __fp16 tmp01 = bias1 + r01 + tmp024a1 + tmp024b1 + tmp024c1 * (__fp16)32;
-                __fp16 tmp10 = bias0 + tmp135a0 + tmp135b0 + tmp135b0 + tmp135c0 * (__fp16)16;
-                __fp16 tmp11 = bias1 + tmp135a1 + tmp135b1 + tmp135b1 + tmp135c1 * (__fp16)16;
-                __fp16 tmp20 = bias0 + tmp024a0 + tmp024b0 * (__fp16)4 + tmp024c0 * (__fp16)8;
-                __fp16 tmp21 = bias1 + tmp024a1 + tmp024b1 * (__fp16)4 + tmp024c1 * (__fp16)8;
-                __fp16 tmp30 = bias0 + tmp135a0 + tmp135b0 * (__fp16)8 + tmp135c0 * (__fp16)4;
-                __fp16 tmp31 = bias1 + tmp135a1 + tmp135b1 * (__fp16)8 + tmp135c1 * (__fp16)4;
-                __fp16 tmp40 = bias0 + tmp024a0 + tmp024b0 * (__fp16)16 + tmp024c0 + tmp024c0;
-                __fp16 tmp41 = bias1 + tmp024a1 + tmp024b1 * (__fp16)16 + tmp024c1 + tmp024c1;
-                __fp16 tmp50 = bias0 + r70 + tmp135a0 + tmp135b0 * (__fp16)32 + tmp135c0;
-                __fp16 tmp51 = bias1 + r71 + tmp135a1 + tmp135b1 * (__fp16)32 + tmp135c1;
-
-                // if (out_elempack == 1)
-                {
-                    __fp16* outptr1 = outptr0 + N;
-
-                    outptr0[0] = tmp00;
-                    outptr1[0] = tmp01;
-                    if (tj * 6 + 1 < outw)
-                    {
-                        outptr0[1] = tmp10;
-                        outptr1[1] = tmp11;
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        outptr0[2] = tmp20;
-                        outptr1[2] = tmp21;
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        outptr0[3] = tmp30;
-                        outptr1[3] = tmp31;
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        outptr0[4] = tmp40;
-                        outptr1[4] = tmp41;
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        outptr0[5] = tmp50;
-                        outptr1[5] = tmp51;
-                    }
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        __fp16 bias0 = biasptr ? biasptr[i + ii] : 0.f;
-
-        __fp16 tmp[6][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const __fp16* r0 = (const __fp16*)top_tile + ii * max_jj * 64 + jj;
-            const __fp16* r1 = r0 + max_jj;
-            const __fp16* r2 = r0 + max_jj * 2;
-            const __fp16* r3 = r0 + max_jj * 3;
-            const __fp16* r4 = r0 + max_jj * 4;
-            const __fp16* r5 = r0 + max_jj * 5;
-            const __fp16* r6 = r0 + max_jj * 6;
-            const __fp16* r7 = r0 + max_jj * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                __fp16 tmp024a = r1[0] + r2[0];
-                __fp16 tmp135a = r1[0] - r2[0];
-                __fp16 tmp024b = r3[0] + r4[0];
-                __fp16 tmp135b = r3[0] - r4[0];
-                __fp16 tmp024c = r5[0] + r6[0];
-                __fp16 tmp135c = r5[0] - r6[0];
-
-                tmp[0][m] = r0[0] + tmp024a + tmp024b + tmp024c * (__fp16)32;
-                tmp[1][m] = tmp135a + tmp135b + tmp135b + tmp135c * (__fp16)16;
-                tmp[2][m] = tmp024a + tmp024b * (__fp16)4 + tmp024c * (__fp16)8;
-                tmp[3][m] = tmp135a + tmp135b * (__fp16)8 + tmp135c * (__fp16)4;
-                tmp[4][m] = tmp024a + tmp024b * (__fp16)16 + tmp024c + tmp024c;
-                tmp[5][m] = r7[0] + tmp135a + tmp135b * (__fp16)32 + tmp135c;
-
-                r0 += max_jj * 8;
-                r1 += max_jj * 8;
-                r2 += max_jj * 8;
-                r3 += max_jj * 8;
-                r4 += max_jj * 8;
-                r5 += max_jj * 8;
-                r6 += max_jj * 8;
-                r7 += max_jj * 8;
-            }
-
-            __fp16* outptr0 = top_blob.channel(i + ii).row<__fp16>(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-                __fp16 r0 = tmp[m][0];
-                __fp16 r1 = tmp[m][1];
-                __fp16 r2 = tmp[m][2];
-                __fp16 r3 = tmp[m][3];
-                __fp16 r4 = tmp[m][4];
-                __fp16 r5 = tmp[m][5];
-                __fp16 r6 = tmp[m][6];
-                __fp16 r7 = tmp[m][7];
-
-                __fp16 tmp024a = r1 + r2;
-                __fp16 tmp135a = r1 - r2;
-                __fp16 tmp024b = r3 + r4;
-                __fp16 tmp135b = r3 - r4;
-                __fp16 tmp024c = r5 + r6;
-                __fp16 tmp135c = r5 - r6;
-
-                __fp16 tmp0 = bias0 + r0 + tmp024a + tmp024b + tmp024c * (__fp16)32;
-                __fp16 tmp1 = bias0 + tmp135a + tmp135b + tmp135b + tmp135c * (__fp16)16;
-                __fp16 tmp2 = bias0 + tmp024a + tmp024b * (__fp16)4 + tmp024c * (__fp16)8;
-                __fp16 tmp3 = bias0 + tmp135a + tmp135b * (__fp16)8 + tmp135c * (__fp16)4;
-                __fp16 tmp4 = bias0 + tmp024a + tmp024b * (__fp16)16 + tmp024c + tmp024c;
-                __fp16 tmp5 = bias0 + r7 + tmp135a + tmp135b * (__fp16)32 + tmp135c;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = tmp0;
-                    if (tj * 6 + 1 < outw) outptr0[1] = tmp1;
-                    if (tj * 6 + 2 < outw) outptr0[2] = tmp2;
-                    if (tj * 6 + 3 < outw) outptr0[3] = tmp3;
-                    if (tj * 6 + 4 < outw) outptr0[4] = tmp4;
-                    if (tj * 6 + 5 < outw) outptr0[5] = tmp5;
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd63_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, const Mat& bias, int nT, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 6n+2, winograd F(6,3)
-    int w_tiles = (outw + 5) / 6;
-    int h_tiles = (outh + 5) / 6;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 64;
-
-    // NCNN_LOGE("conv3x3s1_winograd63_fp16sa %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    conv3x3s1_winograd_get_optimal_tile_mnk_fp16(M, N, K, B, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 2u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 2u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd63_transform_input_tile_fp16sa(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile_fp16(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 2u, opt.workspace_allocator);
-
-        #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd63_transform_input_tile_fp16sa(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            conv3x3s1_winograd_transpose_pack_B_tile_fp16(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 2u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                conv3x3s1_winograd_gemm_transB_packed_tile_fp16sa(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk, opt.use_a53_a55_optimized_kernel);
-            }
-
-            // transform output
-            conv3x3s1_winograd63_transform_output_tile_fp16sa(top_tile, top_blob, bias, i, max_ii, j, max_jj);
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd_int8.h
deleted file mode 100644
index ab108b3f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_3x3_winograd_int8.h
+++ /dev/null
@@ -1,5719 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pack_A_tile_int8(const Mat& A, Mat& AT, int batch, int max_ii, int max_kk)
-{
-    const int N = max_kk * batch;
-
-    for (int b = 0; b < batch; b++)
-    {
-        short* pp = AT.row<short>(b);
-
-        int ii = 0;
-#if __ARM_NEON
-        for (; ii + 7 < max_ii; ii += 8)
-        {
-            const short* p0 = (const short*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                pp[2] = p0[N * 2];
-                pp[3] = p0[N * 3];
-                pp[4] = p0[N * 4];
-                pp[5] = p0[N * 5];
-                pp[6] = p0[N * 6];
-                pp[7] = p0[N * 7];
-                p0 += batch;
-                pp += 8;
-            }
-        }
-        for (; ii + 3 < max_ii; ii += 4)
-        {
-            const short* p0 = (const short*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                pp[2] = p0[N * 2];
-                pp[3] = p0[N * 3];
-                p0 += batch;
-                pp += 4;
-            }
-        }
-#endif // __ARM_NEON
-        for (; ii + 1 < max_ii; ii += 2)
-        {
-            const short* p0 = (const short*)A + ii * N + b;
-
-            int kk = 0;
-#if !__ARM_NEON && __ARM_FEATURE_SIMD32 && NCNN_GNU_INLINE_ASM
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[batch];
-                pp[2] = p0[N];
-                pp[3] = p0[batch + N];
-                p0 += batch * 2;
-                pp += 4;
-            }
-#endif
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                p0 += batch;
-                pp += 2;
-            }
-        }
-        for (; ii < max_ii; ii++)
-        {
-            const short* p0 = (const short*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                p0 += batch;
-                pp += 1;
-            }
-        }
-    }
-}
-
-static void transpose_pack_B_tile_int8(const Mat& B, Mat& BT, int batch, int max_jj, int max_kk, int nT)
-{
-    // NCNN_LOGE("transpose_pack_B_tile_int8 %d %d", max_jj, max_kk);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int b = 0; b < batch; b++)
-    {
-        short* pp = BT.row<short>(b);
-
-        int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            const short* p0 = B;
-
-            int kk = 0;
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                // transpose 8x12
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]   \n"
-                    "prfm   pldl1keep, [%0, #1024]  \n"
-                    "ld4    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0], #64 \n"
-                    "ld4    {v4.8h, v5.8h, v6.8h, v7.8h}, [%0], #64 \n"
-                    "ld4    {v16.8h, v17.8h, v18.8h, v19.8h}, [%0] \n"
-                    "sub    %0, %0, #128            \n"
-                    "uzp1   v20.8h, v0.8h, v4.8h    \n"
-                    "uzp2   v26.8h, v0.8h, v4.8h    \n"
-                    "uzp1   v23.8h, v2.8h, v6.8h    \n"
-                    "uzp2   v29.8h, v2.8h, v6.8h    \n"
-                    "uzp1   v21.8h, v16.8h, v1.8h   \n"
-                    "uzp2   v27.8h, v16.8h, v1.8h   \n"
-                    "uzp1   v22.8h, v5.8h, v17.8h   \n"
-                    "uzp2   v28.8h, v5.8h, v17.8h   \n"
-                    "uzp1   v24.8h, v18.8h, v3.8h   \n"
-                    "uzp2   v30.8h, v18.8h, v3.8h   \n"
-                    "uzp1   v25.8h, v7.8h, v19.8h   \n"
-                    "uzp2   v31.8h, v7.8h, v19.8h   \n"
-                    "st1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%1], #64 \n"
-                    "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%1], #64 \n"
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                p0 += max_jj * batch * 8;
-#else  // NCNN_GNU_INLINE_ASM
-                int16x8x4_t _r0 = vld4q_s16(p0);
-                int16x8x4_t _r1 = vld4q_s16(p0 + 32);
-                int16x8x4_t _r2 = vld4q_s16(p0 + 64);
-                int16x8x2_t _t0 = vuzpq_s16(_r0.val[0], _r1.val[0]);
-                int16x8x2_t _t1 = vuzpq_s16(_r2.val[0], _r0.val[1]);
-                int16x8x2_t _t2 = vuzpq_s16(_r1.val[1], _r2.val[1]);
-                int16x8x2_t _t3 = vuzpq_s16(_r0.val[2], _r1.val[2]);
-                int16x8x2_t _t4 = vuzpq_s16(_r2.val[2], _r0.val[3]);
-                int16x8x2_t _t5 = vuzpq_s16(_r1.val[3], _r2.val[3]);
-                vst1q_s16(pp, _t0.val[0]);
-                vst1q_s16(pp + 8, _t1.val[0]);
-                vst1q_s16(pp + 16, _t2.val[0]);
-                vst1q_s16(pp + 24, _t3.val[0]);
-                vst1q_s16(pp + 32, _t4.val[0]);
-                vst1q_s16(pp + 40, _t5.val[0]);
-                vst1q_s16(pp + 48, _t0.val[1]);
-                vst1q_s16(pp + 56, _t1.val[1]);
-                vst1q_s16(pp + 64, _t2.val[1]);
-                vst1q_s16(pp + 72, _t3.val[1]);
-                vst1q_s16(pp + 80, _t4.val[1]);
-                vst1q_s16(pp + 88, _t5.val[1]);
-                p0 += max_jj * batch * 8;
-                pp += 96;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 8;
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                int16x8x2_t _r01 = vld2q_s16(p0);
-                int16x4x2_t _r2 = vld2_s16(p0 + 16);
-                vst1q_s16(pp, _r01.val[0]);
-                vst1_s16(pp + 8, _r2.val[0]);
-                vst1q_s16(pp + 12, _r01.val[1]);
-                vst1_s16(pp + 20, _r2.val[1]);
-                p0 += max_jj * batch * 2;
-                pp += 24;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                int16x8_t _r0 = vld1q_s16(p0);
-                int16x4_t _r1 = vld1_s16(p0 + 8);
-                vst1q_s16(pp, _r0);
-                vst1_s16(pp + 8, _r1);
-                p0 += max_jj * batch;
-                pp += 12;
-            }
-        }
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            const short* p0 = B;
-
-            int kk = 0;
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                // transpose 8x8
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]   \n"
-                    "prfm   pldl1keep, [%0, #1024]  \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0], #64 \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%0] \n"
-                    "sub    %0, %0, #64             \n"
-                    "zip1   v16.8h, v0.8h, v4.8h    \n"
-                    "zip2   v20.8h, v0.8h, v4.8h    \n"
-                    "zip1   v17.8h, v1.8h, v5.8h    \n"
-                    "zip2   v21.8h, v1.8h, v5.8h    \n"
-                    "zip1   v18.8h, v2.8h, v6.8h    \n"
-                    "zip2   v22.8h, v2.8h, v6.8h    \n"
-                    "zip1   v19.8h, v3.8h, v7.8h    \n"
-                    "zip2   v23.8h, v3.8h, v7.8h    \n"
-                    "st4    {v16.8h, v17.8h, v18.8h, v19.8h}, [%1], #64 \n"
-                    "st4    {v20.8h, v21.8h, v22.8h, v23.8h}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-                p0 += max_jj * batch * 8;
-#else  // NCNN_GNU_INLINE_ASM
-                int16x8_t _r0 = vld1q_s16(p0);
-                int16x8_t _r1 = vld1q_s16(p0 + 8);
-                int16x8_t _r2 = vld1q_s16(p0 + 16);
-                int16x8_t _r3 = vld1q_s16(p0 + 24);
-                int16x8_t _r4 = vld1q_s16(p0 + 32);
-                int16x8_t _r5 = vld1q_s16(p0 + 40);
-                int16x8_t _r6 = vld1q_s16(p0 + 48);
-                int16x8_t _r7 = vld1q_s16(p0 + 56);
-                int16x8x2_t _r04 = vzipq_s16(_r0, _r4);
-                int16x8x2_t _r15 = vzipq_s16(_r1, _r5);
-                int16x8x2_t _r26 = vzipq_s16(_r2, _r6);
-                int16x8x2_t _r37 = vzipq_s16(_r3, _r7);
-                int16x8x4_t _r0123;
-                _r0123.val[0] = _r04.val[0];
-                _r0123.val[1] = _r15.val[0];
-                _r0123.val[2] = _r26.val[0];
-                _r0123.val[3] = _r37.val[0];
-                int16x8x4_t _r4567;
-                _r4567.val[0] = _r04.val[1];
-                _r4567.val[1] = _r15.val[1];
-                _r4567.val[2] = _r26.val[1];
-                _r4567.val[3] = _r37.val[1];
-                vst4q_s16(pp, _r0123);
-                vst4q_s16(pp + 32, _r4567);
-                p0 += max_jj * batch * 8;
-                pp += 64;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 8;
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                int16x8x2_t _r01 = vld2q_s16(p0);
-                vst1q_s16(pp, _r01.val[0]);
-                vst1q_s16(pp + 8, _r01.val[1]);
-                p0 += max_jj * batch * 2;
-                pp += 16;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                int16x8_t _r0 = vld1q_s16(p0);
-                vst1q_s16(pp, _r0);
-                p0 += max_jj * batch;
-                pp += 8;
-            }
-        }
-#endif // __aarch64__
-        for (; jj + 5 < max_jj; jj += 6)
-        {
-            const short* p0 = B;
-
-            int kk = 0;
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #768]   \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h}, [%0], #48 \n"
-                    "ld1    {v3.8h, v4.8h, v5.8h}, [%0] \n"
-                    "sub    %0, %0, #48             \n"
-                    "zip1   v16.8h, v0.8h, v3.8h    \n"
-                    "zip2   v20.8h, v0.8h, v3.8h    \n"
-                    "zip1   v17.8h, v1.8h, v4.8h    \n"
-                    "zip2   v21.8h, v1.8h, v4.8h    \n"
-                    "zip1   v18.8h, v2.8h, v5.8h    \n"
-                    "zip2   v22.8h, v2.8h, v5.8h    \n"
-                    "st3    {v16.8h, v17.8h, v18.8h}, [%1], #48 \n"
-                    "st3    {v20.8h, v21.8h, v22.8h}, [%1], #48 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v20", "v21", "v22");
-                p0 += max_jj * batch * 8;
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%0, #768]          \n"
-                    "vldm       %0, {d0-d11}        \n"
-                    "vzip.16    q0, q3              \n"
-                    "vzip.16    q1, q4              \n"
-                    "vzip.16    q2, q5              \n"
-                    "vst3.s16   {d0,d2,d4}, [%1]!   \n"
-                    "vst3.s16   {d1,d3,d5}, [%1]!   \n"
-                    "vst3.s16   {d6,d8,d10}, [%1]!  \n"
-                    "vst3.s16   {d7,d9,d11}, [%1]!  \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "q0", "q1", "q2", "q3", "q4", "q5");
-                p0 += max_jj * batch * 8;
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                int16x8_t _r0 = vld1q_s16(p0);
-                int16x8_t _r1 = vld1q_s16(p0 + 8);
-                int16x8_t _r2 = vld1q_s16(p0 + 16);
-                int16x8_t _r3 = vld1q_s16(p0 + 24);
-                int16x8_t _r4 = vld1q_s16(p0 + 32);
-                int16x8_t _r5 = vld1q_s16(p0 + 40);
-                int16x8x2_t _r03 = vzipq_s16(_r0, _r3);
-                int16x8x2_t _r14 = vzipq_s16(_r1, _r4);
-                int16x8x2_t _r25 = vzipq_s16(_r2, _r5);
-                int16x8x3_t _r012;
-                _r012.val[0] = _r03.val[0];
-                _r012.val[1] = _r14.val[0];
-                _r012.val[2] = _r25.val[0];
-                int16x8x3_t _r345;
-                _r345.val[0] = _r03.val[1];
-                _r345.val[1] = _r14.val[1];
-                _r345.val[2] = _r25.val[1];
-                vst3q_s16(pp, _r012);
-                vst3q_s16(pp + 24, _r345);
-                p0 += max_jj * batch * 8;
-                pp += 48;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 8;
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                int16x8x2_t _r01 = vld2q_s16(p0);
-                int32x4x2_t _r01x = vtrnq_s32(vreinterpretq_s32_s16(_r01.val[0]), vreinterpretq_s32_s16(_r01.val[1]));
-                int32x2x3_t _r012;
-                _r012.val[0] = vget_low_s32(_r01x.val[0]);
-                _r012.val[1] = vget_low_s32(_r01x.val[1]);
-                _r012.val[2] = vget_high_s32(_r01x.val[0]);
-                vst3_s32((int*)pp, _r012);
-                p0 += max_jj * batch * 2;
-                pp += 12;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                int16x4_t _r0 = vld1_s16(p0);
-                vst1_s16(pp, _r0);
-                pp[4] = p0[4];
-                pp[5] = p0[5];
-                p0 += max_jj * batch;
-                pp += 6;
-            }
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            const short* p0 = B;
-
-            int kk = 0;
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]   \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0] \n"
-                    "st4    {v0.8h, v1.8h, v2.8h, v3.8h}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3");
-                p0 += max_jj * batch * 8;
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%0, #512]          \n"
-                    "vldm       %0, {d0-d7}         \n"
-                    "vst4.s16   {d0,d2,d4,d6}, [%1]! \n"
-                    "vst4.s16   {d1,d3,d5,d7}, [%1]! \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "q0", "q1", "q2", "q3");
-                p0 += max_jj * batch * 8;
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                int16x8x4_t _r0123;
-                _r0123.val[0] = vld1q_s16(p0);
-                _r0123.val[1] = vld1q_s16(p0 + 8);
-                _r0123.val[2] = vld1q_s16(p0 + 16);
-                _r0123.val[3] = vld1q_s16(p0 + 24);
-                vst4q_s16(pp, _r0123);
-                p0 += max_jj * batch * 8;
-                pp += 32;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 8;
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                int16x4x2_t _r01 = vld2_s16(p0);
-                vst1_s16(pp, _r01.val[0]);
-                vst1_s16(pp + 4, _r01.val[1]);
-                p0 += max_jj * batch * 2;
-                pp += 8;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                int16x4_t _r0 = vld1_s16(p0);
-                vst1_s16(pp, _r0);
-                p0 += max_jj * batch;
-                pp += 4;
-            }
-        }
-#endif // __ARM_NEON
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            const short* p0 = B;
-
-            int kk = 0;
-#if __ARM_NEON
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #256]   \n"
-                    "ld1    {v0.8h, v1.8h}, [%0]    \n"
-                    "st2    {v0.8h, v1.8h}, [%1], #32 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1");
-                p0 += max_jj * batch * 8;
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%0, #256]          \n"
-                    "vld1.s16   {d0-d3}, [%0]       \n"
-                    "vst2.s16   {d0-d3}, [%1]!      \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "q0", "q1");
-                p0 += max_jj * batch * 8;
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                int16x8x2_t _r01;
-                _r01.val[0] = vld1q_s16(p0);
-                _r01.val[1] = vld1q_s16(p0 + 8);
-                vst2q_s16(pp, _r01);
-                p0 += max_jj * batch * 8;
-                pp += 16;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 8;
-#endif // __ARM_NEON
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-#if !__ARM_NEON && __ARM_FEATURE_SIMD32 && NCNN_GNU_INLINE_ASM
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                pp[2] = p0[2];
-                pp[3] = p0[3];
-#else
-                pp[0] = p0[0];
-                pp[1] = p0[2];
-                pp[2] = p0[1];
-                pp[3] = p0[3];
-#endif
-                p0 += max_jj * batch * 2;
-                pp += 4;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                p0 += max_jj * batch;
-                pp += 2;
-            }
-        }
-        for (; jj < max_jj; jj++)
-        {
-            const short* p0 = B;
-
-            int kk = 0;
-#if __ARM_NEON
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #128]   \n"
-                    "ld1    {v0.8h}, [%0]           \n"
-                    "st1    {v0.8h}, [%1], #16      \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0");
-                p0 += max_jj * batch * 8;
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%0, #128]          \n"
-                    "vld1.s16   {d0-d1}, [%0]       \n"
-                    "vst1.s16   {d0-d1}, [%1]!      \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "q0");
-                p0 += max_jj * batch * 8;
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                int16x8_t _r0 = vld1q_s16(p0);
-                vst1q_s16(pp, _r0);
-                p0 += max_jj * batch * 8;
-                pp += 8;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            p0 -= (b * max_jj + jj) * 8;
-#endif // __ARM_NEON
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                p0 += max_jj * batch * 2;
-                pp += 2;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                p0 += max_jj * batch;
-                pp += 1;
-            }
-        }
-    }
-}
-
-static void gemm_transB_packed_tile_int8(const Mat& AT_tile, const Mat& BT_tile, Mat& top_blob, int batch, int max_ii, int max_jj, int k, int max_kk)
-{
-    // return;
-    // NCNN_LOGE("gemm_transB_packed_tile_int8 %d %d %d", max_ii, max_jj, max_kk);
-
-    int* outptr = top_blob;
-
-    int ii = 0;
-#if __ARM_NEON
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const short* pAT = AT_tile.row<const short>(b) + max_kk * ii;
-            const short* pB = BT_tile.row<const short>(b);
-
-            int jj = 0;
-#if __aarch64__
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "cmp    %w7, #0                     \n"
-                    "beq    0f                          \n"
-
-                    "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "sub    %0, %0, #320                \n"
-                    "b      1f                          \n"
-
-                    "0:                                 \n"
-                    "eor    v8.16b, v8.16b, v8.16b      \n"
-                    "eor    v9.16b, v9.16b, v9.16b      \n"
-                    "eor    v10.16b, v10.16b, v10.16b   \n"
-                    "eor    v11.16b, v11.16b, v11.16b   \n"
-                    "eor    v12.16b, v12.16b, v12.16b   \n"
-                    "eor    v13.16b, v13.16b, v13.16b   \n"
-                    "eor    v14.16b, v14.16b, v14.16b   \n"
-                    "eor    v15.16b, v15.16b, v15.16b   \n"
-                    "eor    v16.16b, v16.16b, v16.16b   \n"
-                    "eor    v17.16b, v17.16b, v17.16b   \n"
-                    "eor    v18.16b, v18.16b, v18.16b   \n"
-                    "eor    v19.16b, v19.16b, v19.16b   \n"
-                    "eor    v20.16b, v20.16b, v20.16b   \n"
-                    "eor    v21.16b, v21.16b, v21.16b   \n"
-                    "eor    v22.16b, v22.16b, v22.16b   \n"
-                    "eor    v23.16b, v23.16b, v23.16b   \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-                    "eor    v25.16b, v25.16b, v25.16b   \n"
-                    "eor    v26.16b, v26.16b, v26.16b   \n"
-                    "eor    v27.16b, v27.16b, v27.16b   \n"
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-                    "eor    v29.16b, v29.16b, v29.16b   \n"
-                    "eor    v30.16b, v30.16b, v30.16b   \n"
-                    "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                    "1:                                 \n"
-                    "lsr    w4, %w6, #3                 \n" // w4 = max_kk >> 3
-                    "cmp    w4, #0                      \n"
-                    "beq    3f                          \n"
-
-                    "ld1    {v4.8h, v5.8h}, [%1], #32   \n"
-                    "ld1    {v0.8h, v1.8h}, [%2], #32   \n"
-                    ".align 4                           \n"
-                    "2:                                 \n"
-                    "smlal  v8.4s, v4.4h, v0.h[0]       \n"
-                    "smlal  v10.4s, v4.4h, v0.h[1]      \n"
-                    "ld1    {v2.8h, v3.8h}, [%2], #32   \n"
-                    "smlal2 v9.4s, v4.8h, v0.h[0]       \n"
-                    "smlal2 v11.4s, v4.8h, v0.h[1]      \n"
-                    "ld1    {v6.8h, v7.8h}, [%1], #32   \n"
-                    "smlal  v12.4s, v4.4h, v0.h[2]      \n"
-                    "smlal  v14.4s, v4.4h, v0.h[3]      \n"
-                    "smlal2 v13.4s, v4.8h, v0.h[2]      \n"
-                    "smlal2 v15.4s, v4.8h, v0.h[3]      \n"
-                    "smlal  v16.4s, v4.4h, v0.h[4]      \n"
-                    "smlal  v18.4s, v4.4h, v0.h[5]      \n"
-                    "smlal2 v17.4s, v4.8h, v0.h[4]      \n"
-                    "smlal2 v19.4s, v4.8h, v0.h[5]      \n"
-                    "smlal  v20.4s, v4.4h, v0.h[6]      \n"
-                    "smlal  v22.4s, v4.4h, v0.h[7]      \n"
-                    "smlal2 v21.4s, v4.8h, v0.h[6]      \n"
-                    "smlal2 v23.4s, v4.8h, v0.h[7]      \n"
-                    "smlal  v24.4s, v4.4h, v1.h[0]      \n"
-                    "smlal  v26.4s, v4.4h, v1.h[1]      \n"
-                    "smlal2 v25.4s, v4.8h, v1.h[0]      \n"
-                    "smlal2 v27.4s, v4.8h, v1.h[1]      \n"
-                    "smlal  v28.4s, v4.4h, v1.h[2]      \n"
-                    "smlal  v30.4s, v4.4h, v1.h[3]      \n"
-                    "smlal2 v29.4s, v4.8h, v1.h[2]      \n"
-                    "smlal2 v31.4s, v4.8h, v1.h[3]      \n"
-                    "smlal  v8.4s, v5.4h, v1.h[4]       \n"
-                    "smlal  v10.4s, v5.4h, v1.h[5]      \n"
-                    "smlal2 v9.4s, v5.8h, v1.h[4]       \n"
-                    "smlal2 v11.4s, v5.8h, v1.h[5]      \n"
-                    "smlal  v12.4s, v5.4h, v1.h[6]      \n"
-                    "smlal  v14.4s, v5.4h, v1.h[7]      \n"
-                    "smlal2 v13.4s, v5.8h, v1.h[6]      \n"
-                    "smlal2 v15.4s, v5.8h, v1.h[7]      \n"
-                    "smlal  v16.4s, v5.4h, v2.h[0]      \n"
-                    "smlal  v18.4s, v5.4h, v2.h[1]      \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.8h, v1.8h}, [%2], #32   \n"
-                    "smlal2 v17.4s, v5.8h, v2.h[0]      \n"
-                    "smlal2 v19.4s, v5.8h, v2.h[1]      \n"
-                    "smlal  v20.4s, v5.4h, v2.h[2]      \n"
-                    "smlal  v22.4s, v5.4h, v2.h[3]      \n"
-                    "smlal2 v21.4s, v5.8h, v2.h[2]      \n"
-                    "smlal2 v23.4s, v5.8h, v2.h[3]      \n"
-                    "smlal  v24.4s, v5.4h, v2.h[4]      \n"
-                    "smlal  v26.4s, v5.4h, v2.h[5]      \n"
-                    "smlal2 v25.4s, v5.8h, v2.h[4]      \n"
-                    "smlal2 v27.4s, v5.8h, v2.h[5]      \n"
-                    "smlal  v28.4s, v5.4h, v2.h[6]      \n"
-                    "smlal  v30.4s, v5.4h, v2.h[7]      \n"
-                    "smlal2 v29.4s, v5.8h, v2.h[6]      \n"
-                    "smlal2 v31.4s, v5.8h, v2.h[7]      \n"
-                    "smlal  v8.4s, v6.4h, v3.h[0]       \n"
-                    "smlal  v10.4s, v6.4h, v3.h[1]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.8h, v5.8h}, [%1], #32   \n"
-                    "smlal2 v9.4s, v6.8h, v3.h[0]       \n"
-                    "smlal2 v11.4s, v6.8h, v3.h[1]      \n"
-                    "smlal  v12.4s, v6.4h, v3.h[2]      \n"
-                    "smlal  v14.4s, v6.4h, v3.h[3]      \n"
-                    "smlal2 v13.4s, v6.8h, v3.h[2]      \n"
-                    "smlal2 v15.4s, v6.8h, v3.h[3]      \n"
-                    "smlal  v16.4s, v6.4h, v3.h[4]      \n"
-                    "smlal  v18.4s, v6.4h, v3.h[5]      \n"
-                    "smlal2 v17.4s, v6.8h, v3.h[4]      \n"
-                    "smlal2 v19.4s, v6.8h, v3.h[5]      \n"
-                    "smlal  v20.4s, v6.4h, v3.h[6]      \n"
-                    "smlal  v22.4s, v6.4h, v3.h[7]      \n"
-                    "smlal2 v21.4s, v6.8h, v3.h[6]      \n"
-                    "smlal2 v23.4s, v6.8h, v3.h[7]      \n"
-                    "smlal  v24.4s, v6.4h, v0.h[0]      \n"
-                    "smlal  v26.4s, v6.4h, v0.h[1]      \n"
-                    "ld1    {v2.8h, v3.8h}, [%2], #32   \n"
-                    "smlal2 v25.4s, v6.8h, v0.h[0]      \n"
-                    "smlal2 v27.4s, v6.8h, v0.h[1]      \n"
-                    "smlal  v28.4s, v6.4h, v0.h[2]      \n"
-                    "smlal  v30.4s, v6.4h, v0.h[3]      \n"
-                    "smlal2 v29.4s, v6.8h, v0.h[2]      \n"
-                    "smlal2 v31.4s, v6.8h, v0.h[3]      \n"
-                    "smlal  v8.4s, v7.4h, v0.h[4]       \n"
-                    "smlal  v10.4s, v7.4h, v0.h[5]      \n"
-                    "smlal2 v9.4s, v7.8h, v0.h[4]       \n"
-                    "smlal2 v11.4s, v7.8h, v0.h[5]      \n"
-                    "smlal  v12.4s, v7.4h, v0.h[6]      \n"
-                    "smlal  v14.4s, v7.4h, v0.h[7]      \n"
-                    "smlal2 v13.4s, v7.8h, v0.h[6]      \n"
-                    "smlal2 v15.4s, v7.8h, v0.h[7]      \n"
-                    "smlal  v16.4s, v7.4h, v1.h[0]      \n"
-                    "smlal  v18.4s, v7.4h, v1.h[1]      \n"
-                    "smlal2 v17.4s, v7.8h, v1.h[0]      \n"
-                    "smlal2 v19.4s, v7.8h, v1.h[1]      \n"
-                    "smlal  v20.4s, v7.4h, v1.h[2]      \n"
-                    "smlal  v22.4s, v7.4h, v1.h[3]      \n"
-                    "smlal2 v21.4s, v7.8h, v1.h[2]      \n"
-                    "smlal2 v23.4s, v7.8h, v1.h[3]      \n"
-                    "smlal  v24.4s, v7.4h, v1.h[4]      \n"
-                    "smlal  v26.4s, v7.4h, v1.h[5]      \n"
-                    "smlal2 v25.4s, v7.8h, v1.h[4]      \n"
-                    "smlal2 v27.4s, v7.8h, v1.h[5]      \n"
-                    "smlal  v28.4s, v7.4h, v1.h[6]      \n"
-                    "smlal  v30.4s, v7.4h, v1.h[7]      \n"
-                    "smlal2 v29.4s, v7.8h, v1.h[6]      \n"
-                    "smlal2 v31.4s, v7.8h, v1.h[7]      \n"
-                    "smlal  v8.4s, v4.4h, v2.h[0]       \n"
-                    "smlal  v10.4s, v4.4h, v2.h[1]      \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.8h, v1.8h}, [%2], #32   \n"
-                    "smlal2 v9.4s, v4.8h, v2.h[0]       \n"
-                    "smlal2 v11.4s, v4.8h, v2.h[1]      \n"
-                    "ld1    {v6.8h, v7.8h}, [%1], #32   \n"
-                    "smlal  v12.4s, v4.4h, v2.h[2]      \n"
-                    "smlal  v14.4s, v4.4h, v2.h[3]      \n"
-                    "smlal2 v13.4s, v4.8h, v2.h[2]      \n"
-                    "smlal2 v15.4s, v4.8h, v2.h[3]      \n"
-                    "smlal  v16.4s, v4.4h, v2.h[4]      \n"
-                    "smlal  v18.4s, v4.4h, v2.h[5]      \n"
-                    "smlal2 v17.4s, v4.8h, v2.h[4]      \n"
-                    "smlal2 v19.4s, v4.8h, v2.h[5]      \n"
-                    "smlal  v20.4s, v4.4h, v2.h[6]      \n"
-                    "smlal  v22.4s, v4.4h, v2.h[7]      \n"
-                    "smlal2 v21.4s, v4.8h, v2.h[6]      \n"
-                    "smlal2 v23.4s, v4.8h, v2.h[7]      \n"
-                    "smlal  v24.4s, v4.4h, v3.h[0]      \n"
-                    "smlal  v26.4s, v4.4h, v3.h[1]      \n"
-                    "smlal2 v25.4s, v4.8h, v3.h[0]      \n"
-                    "smlal2 v27.4s, v4.8h, v3.h[1]      \n"
-                    "smlal  v28.4s, v4.4h, v3.h[2]      \n"
-                    "smlal  v30.4s, v4.4h, v3.h[3]      \n"
-                    "smlal2 v29.4s, v4.8h, v3.h[2]      \n"
-                    "smlal2 v31.4s, v4.8h, v3.h[3]      \n"
-                    "smlal  v8.4s, v5.4h, v3.h[4]       \n"
-                    "smlal  v10.4s, v5.4h, v3.h[5]      \n"
-                    "smlal2 v9.4s, v5.8h, v3.h[4]       \n"
-                    "smlal2 v11.4s, v5.8h, v3.h[5]      \n"
-                    "smlal  v12.4s, v5.4h, v3.h[6]      \n"
-                    "smlal  v14.4s, v5.4h, v3.h[7]      \n"
-                    "smlal2 v13.4s, v5.8h, v3.h[6]      \n"
-                    "smlal2 v15.4s, v5.8h, v3.h[7]      \n"
-                    "smlal  v16.4s, v5.4h, v0.h[0]      \n"
-                    "smlal  v18.4s, v5.4h, v0.h[1]      \n"
-                    "ld1    {v2.8h, v3.8h}, [%2], #32   \n"
-                    "smlal2 v17.4s, v5.8h, v0.h[0]      \n"
-                    "smlal2 v19.4s, v5.8h, v0.h[1]      \n"
-                    "smlal  v20.4s, v5.4h, v0.h[2]      \n"
-                    "smlal  v22.4s, v5.4h, v0.h[3]      \n"
-                    "smlal2 v21.4s, v5.8h, v0.h[2]      \n"
-                    "smlal2 v23.4s, v5.8h, v0.h[3]      \n"
-                    "smlal  v24.4s, v5.4h, v0.h[4]      \n"
-                    "smlal  v26.4s, v5.4h, v0.h[5]      \n"
-                    "smlal2 v25.4s, v5.8h, v0.h[4]      \n"
-                    "smlal2 v27.4s, v5.8h, v0.h[5]      \n"
-                    "smlal  v28.4s, v5.4h, v0.h[6]      \n"
-                    "smlal  v30.4s, v5.4h, v0.h[7]      \n"
-                    "smlal2 v29.4s, v5.8h, v0.h[6]      \n"
-                    "smlal2 v31.4s, v5.8h, v0.h[7]      \n"
-                    "smlal  v8.4s, v6.4h, v1.h[0]       \n"
-                    "smlal  v10.4s, v6.4h, v1.h[1]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.8h, v5.8h}, [%1], #32   \n"
-                    "smlal2 v9.4s, v6.8h, v1.h[0]       \n"
-                    "smlal2 v11.4s, v6.8h, v1.h[1]      \n"
-                    "smlal  v12.4s, v6.4h, v1.h[2]      \n"
-                    "smlal  v14.4s, v6.4h, v1.h[3]      \n"
-                    "smlal2 v13.4s, v6.8h, v1.h[2]      \n"
-                    "smlal2 v15.4s, v6.8h, v1.h[3]      \n"
-                    "smlal  v16.4s, v6.4h, v1.h[4]      \n"
-                    "smlal  v18.4s, v6.4h, v1.h[5]      \n"
-                    "smlal2 v17.4s, v6.8h, v1.h[4]      \n"
-                    "smlal2 v19.4s, v6.8h, v1.h[5]      \n"
-                    "smlal  v20.4s, v6.4h, v1.h[6]      \n"
-                    "smlal  v22.4s, v6.4h, v1.h[7]      \n"
-                    "smlal2 v21.4s, v6.8h, v1.h[6]      \n"
-                    "smlal2 v23.4s, v6.8h, v1.h[7]      \n"
-                    "smlal  v24.4s, v6.4h, v2.h[0]      \n"
-                    "smlal  v26.4s, v6.4h, v2.h[1]      \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.8h, v1.8h}, [%2], #32   \n"
-                    "smlal2 v25.4s, v6.8h, v2.h[0]      \n"
-                    "smlal2 v27.4s, v6.8h, v2.h[1]      \n"
-                    "smlal  v28.4s, v6.4h, v2.h[2]      \n"
-                    "smlal  v30.4s, v6.4h, v2.h[3]      \n"
-                    "smlal2 v29.4s, v6.8h, v2.h[2]      \n"
-                    "smlal2 v31.4s, v6.8h, v2.h[3]      \n"
-                    "smlal  v8.4s, v7.4h, v2.h[4]       \n"
-                    "smlal  v10.4s, v7.4h, v2.h[5]      \n"
-                    "smlal2 v9.4s, v7.8h, v2.h[4]       \n"
-                    "smlal2 v11.4s, v7.8h, v2.h[5]      \n"
-                    "smlal  v12.4s, v7.4h, v2.h[6]      \n"
-                    "smlal  v14.4s, v7.4h, v2.h[7]      \n"
-                    "smlal2 v13.4s, v7.8h, v2.h[6]      \n"
-                    "smlal2 v15.4s, v7.8h, v2.h[7]      \n"
-                    "smlal  v16.4s, v7.4h, v3.h[0]      \n"
-                    "smlal  v18.4s, v7.4h, v3.h[1]      \n"
-                    "smlal2 v17.4s, v7.8h, v3.h[0]      \n"
-                    "smlal2 v19.4s, v7.8h, v3.h[1]      \n"
-                    "smlal  v20.4s, v7.4h, v3.h[2]      \n"
-                    "smlal  v22.4s, v7.4h, v3.h[3]      \n"
-                    "smlal2 v21.4s, v7.8h, v3.h[2]      \n"
-                    "smlal2 v23.4s, v7.8h, v3.h[3]      \n"
-                    "smlal  v24.4s, v7.4h, v3.h[4]      \n"
-                    "smlal  v26.4s, v7.4h, v3.h[5]      \n"
-                    "smlal2 v25.4s, v7.8h, v3.h[4]      \n"
-                    "smlal2 v27.4s, v7.8h, v3.h[5]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "smlal  v28.4s, v7.4h, v3.h[6]      \n"
-                    "smlal  v30.4s, v7.4h, v3.h[7]      \n"
-                    "smlal2 v29.4s, v7.8h, v3.h[6]      \n"
-                    "smlal2 v31.4s, v7.8h, v3.h[7]      \n"
-                    "bne    2b                          \n"
-                    "sub    %1, %1, #32                 \n"
-                    "sub    %2, %2, #32                 \n"
-
-                    "3:                                 \n"
-                    "and    w4, %w6, #7                 \n" // w4 = remain = max_kk & 7
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "ld1    {v4.8h}, [%1], #16          \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h}, [%2], #24 \n"
-                    "smlal  v8.4s, v4.4h, v0.h[0]       \n"
-                    "smlal  v10.4s, v4.4h, v0.h[1]      \n"
-                    "smlal2 v9.4s, v4.8h, v0.h[0]       \n"
-                    "smlal2 v11.4s, v4.8h, v0.h[1]      \n"
-                    "smlal  v12.4s, v4.4h, v0.h[2]      \n"
-                    "smlal  v14.4s, v4.4h, v0.h[3]      \n"
-                    "smlal2 v13.4s, v4.8h, v0.h[2]      \n"
-                    "smlal2 v15.4s, v4.8h, v0.h[3]      \n"
-                    "smlal  v16.4s, v4.4h, v1.h[0]      \n"
-                    "smlal  v18.4s, v4.4h, v1.h[1]      \n"
-                    "smlal2 v17.4s, v4.8h, v1.h[0]      \n"
-                    "smlal2 v19.4s, v4.8h, v1.h[1]      \n"
-                    "smlal  v20.4s, v4.4h, v1.h[2]      \n"
-                    "smlal  v22.4s, v4.4h, v1.h[3]      \n"
-                    "smlal2 v21.4s, v4.8h, v1.h[2]      \n"
-                    "smlal2 v23.4s, v4.8h, v1.h[3]      \n"
-                    "smlal  v24.4s, v4.4h, v2.h[0]      \n"
-                    "smlal  v26.4s, v4.4h, v2.h[1]      \n"
-                    "smlal2 v25.4s, v4.8h, v2.h[0]      \n"
-                    "smlal2 v27.4s, v4.8h, v2.h[1]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "smlal  v28.4s, v4.4h, v2.h[2]      \n"
-                    "smlal  v30.4s, v4.4h, v2.h[3]      \n"
-                    "smlal2 v29.4s, v4.8h, v2.h[2]      \n"
-                    "smlal2 v31.4s, v4.8h, v2.h[3]      \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "st1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                    "st1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-                int32x4_t _sum2;
-                int32x4_t _sum3;
-                int32x4_t _sum4;
-                int32x4_t _sum5;
-                int32x4_t _sum6;
-                int32x4_t _sum7;
-                int32x4_t _sum8;
-                int32x4_t _sum9;
-                int32x4_t _suma;
-                int32x4_t _sumb;
-                int32x4_t _sumc;
-                int32x4_t _sumd;
-                int32x4_t _sume;
-                int32x4_t _sumf;
-                int32x4_t _sumg;
-                int32x4_t _sumh;
-                int32x4_t _sumi;
-                int32x4_t _sumj;
-                int32x4_t _sumk;
-                int32x4_t _suml;
-                int32x4_t _summ;
-                int32x4_t _sumn;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                    _sum2 = vdupq_n_s32(0);
-                    _sum3 = vdupq_n_s32(0);
-                    _sum4 = vdupq_n_s32(0);
-                    _sum5 = vdupq_n_s32(0);
-                    _sum6 = vdupq_n_s32(0);
-                    _sum7 = vdupq_n_s32(0);
-                    _sum8 = vdupq_n_s32(0);
-                    _sum9 = vdupq_n_s32(0);
-                    _suma = vdupq_n_s32(0);
-                    _sumb = vdupq_n_s32(0);
-                    _sumc = vdupq_n_s32(0);
-                    _sumd = vdupq_n_s32(0);
-                    _sume = vdupq_n_s32(0);
-                    _sumf = vdupq_n_s32(0);
-                    _sumg = vdupq_n_s32(0);
-                    _sumh = vdupq_n_s32(0);
-                    _sumi = vdupq_n_s32(0);
-                    _sumj = vdupq_n_s32(0);
-                    _sumk = vdupq_n_s32(0);
-                    _suml = vdupq_n_s32(0);
-                    _summ = vdupq_n_s32(0);
-                    _sumn = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                    _sum1 = vld1q_s32(outptr + 4);
-                    _sum2 = vld1q_s32(outptr + 8);
-                    _sum3 = vld1q_s32(outptr + 12);
-                    _sum4 = vld1q_s32(outptr + 16);
-                    _sum5 = vld1q_s32(outptr + 20);
-                    _sum6 = vld1q_s32(outptr + 24);
-                    _sum7 = vld1q_s32(outptr + 28);
-                    _sum8 = vld1q_s32(outptr + 32);
-                    _sum9 = vld1q_s32(outptr + 36);
-                    _suma = vld1q_s32(outptr + 40);
-                    _sumb = vld1q_s32(outptr + 44);
-                    _sumc = vld1q_s32(outptr + 48);
-                    _sumd = vld1q_s32(outptr + 52);
-                    _sume = vld1q_s32(outptr + 56);
-                    _sumf = vld1q_s32(outptr + 60);
-                    _sumg = vld1q_s32(outptr + 64);
-                    _sumh = vld1q_s32(outptr + 68);
-                    _sumi = vld1q_s32(outptr + 72);
-                    _sumj = vld1q_s32(outptr + 76);
-                    _sumk = vld1q_s32(outptr + 80);
-                    _suml = vld1q_s32(outptr + 84);
-                    _summ = vld1q_s32(outptr + 88);
-                    _sumn = vld1q_s32(outptr + 92);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x8_t _pA = vld1q_s16(pA);
-                    int16x8_t _pB = vld1q_s16(pB);
-                    int16x4_t _pB2 = vld1_s16(pB + 8);
-                    _sum0 = vmlal_laneq_s16(_sum0, vget_low_s16(_pA), _pB, 0);
-                    _sum1 = vmlal_laneq_s16(_sum1, vget_high_s16(_pA), _pB, 0);
-                    _sum2 = vmlal_laneq_s16(_sum2, vget_low_s16(_pA), _pB, 1);
-                    _sum3 = vmlal_laneq_s16(_sum3, vget_high_s16(_pA), _pB, 1);
-                    _sum4 = vmlal_laneq_s16(_sum4, vget_low_s16(_pA), _pB, 2);
-                    _sum5 = vmlal_laneq_s16(_sum5, vget_high_s16(_pA), _pB, 2);
-                    _sum6 = vmlal_laneq_s16(_sum6, vget_low_s16(_pA), _pB, 3);
-                    _sum7 = vmlal_laneq_s16(_sum7, vget_high_s16(_pA), _pB, 3);
-                    _sum8 = vmlal_laneq_s16(_sum8, vget_low_s16(_pA), _pB, 4);
-                    _sum9 = vmlal_laneq_s16(_sum9, vget_high_s16(_pA), _pB, 4);
-                    _suma = vmlal_laneq_s16(_suma, vget_low_s16(_pA), _pB, 5);
-                    _sumb = vmlal_laneq_s16(_sumb, vget_high_s16(_pA), _pB, 5);
-                    _sumc = vmlal_laneq_s16(_sumc, vget_low_s16(_pA), _pB, 6);
-                    _sumd = vmlal_laneq_s16(_sumd, vget_high_s16(_pA), _pB, 6);
-                    _sume = vmlal_laneq_s16(_sume, vget_low_s16(_pA), _pB, 7);
-                    _sumf = vmlal_laneq_s16(_sumf, vget_high_s16(_pA), _pB, 7);
-                    _sumg = vmlal_lane_s16(_sumg, vget_low_s16(_pA), _pB2, 0);
-                    _sumh = vmlal_lane_s16(_sumh, vget_high_s16(_pA), _pB2, 0);
-                    _sumi = vmlal_lane_s16(_sumi, vget_low_s16(_pA), _pB2, 1);
-                    _sumj = vmlal_lane_s16(_sumj, vget_high_s16(_pA), _pB2, 1);
-                    _sumk = vmlal_lane_s16(_sumk, vget_low_s16(_pA), _pB2, 2);
-                    _suml = vmlal_lane_s16(_suml, vget_high_s16(_pA), _pB2, 2);
-                    _summ = vmlal_lane_s16(_summ, vget_low_s16(_pA), _pB2, 3);
-                    _sumn = vmlal_lane_s16(_sumn, vget_high_s16(_pA), _pB2, 3);
-                    pA += 8;
-                    pB += 12;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-                vst1q_s32(outptr + 16, _sum4);
-                vst1q_s32(outptr + 20, _sum5);
-                vst1q_s32(outptr + 24, _sum6);
-                vst1q_s32(outptr + 28, _sum7);
-                vst1q_s32(outptr + 32, _sum8);
-                vst1q_s32(outptr + 36, _sum9);
-                vst1q_s32(outptr + 40, _suma);
-                vst1q_s32(outptr + 44, _sumb);
-                vst1q_s32(outptr + 48, _sumc);
-                vst1q_s32(outptr + 52, _sumd);
-                vst1q_s32(outptr + 56, _sume);
-                vst1q_s32(outptr + 60, _sumf);
-                vst1q_s32(outptr + 64, _sumg);
-                vst1q_s32(outptr + 68, _sumh);
-                vst1q_s32(outptr + 72, _sumi);
-                vst1q_s32(outptr + 76, _sumj);
-                vst1q_s32(outptr + 80, _sumk);
-                vst1q_s32(outptr + 84, _suml);
-                vst1q_s32(outptr + 88, _summ);
-                vst1q_s32(outptr + 92, _sumn);
-                outptr += 96;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "cmp    %w7, #0                     \n"
-                    "beq    0f                          \n"
-
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "sub    %0, %0, #192                \n"
-                    "b      1f                          \n"
-
-                    "0:                                 \n"
-                    "eor    v16.16b, v16.16b, v16.16b   \n"
-                    "eor    v17.16b, v17.16b, v17.16b   \n"
-                    "eor    v18.16b, v18.16b, v18.16b   \n"
-                    "eor    v19.16b, v19.16b, v19.16b   \n"
-                    "eor    v20.16b, v20.16b, v20.16b   \n"
-                    "eor    v21.16b, v21.16b, v21.16b   \n"
-                    "eor    v22.16b, v22.16b, v22.16b   \n"
-                    "eor    v23.16b, v23.16b, v23.16b   \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-                    "eor    v25.16b, v25.16b, v25.16b   \n"
-                    "eor    v26.16b, v26.16b, v26.16b   \n"
-                    "eor    v27.16b, v27.16b, v27.16b   \n"
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-                    "eor    v29.16b, v29.16b, v29.16b   \n"
-                    "eor    v30.16b, v30.16b, v30.16b   \n"
-                    "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                    "1:                                 \n"
-                    "lsr    w4, %w6, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    3f                          \n"
-
-                    "ld1    {v4.8h, v5.8h}, [%1], #32   \n"
-                    "ld1    {v0.8h, v1.8h}, [%2], #32   \n"
-                    ".align 4                           \n"
-                    "2:                                 \n"
-                    "smlal  v16.4s, v4.4h, v0.h[0]      \n"
-                    "smlal  v18.4s, v4.4h, v0.h[1]      \n"
-                    "ld1    {v6.8h, v7.8h}, [%1], #32   \n"
-                    "smlal2 v17.4s, v4.8h, v0.h[0]      \n"
-                    "smlal2 v19.4s, v4.8h, v0.h[1]      \n"
-                    "ld1    {v2.8h, v3.8h}, [%2], #32   \n"
-                    "smlal  v20.4s, v4.4h, v0.h[2]      \n"
-                    "smlal  v22.4s, v4.4h, v0.h[3]      \n"
-                    "smlal2 v21.4s, v4.8h, v0.h[2]      \n"
-                    "smlal2 v23.4s, v4.8h, v0.h[3]      \n"
-                    "smlal  v24.4s, v4.4h, v0.h[4]      \n"
-                    "smlal  v26.4s, v4.4h, v0.h[5]      \n"
-                    "smlal2 v25.4s, v4.8h, v0.h[4]      \n"
-                    "smlal2 v27.4s, v4.8h, v0.h[5]      \n"
-                    "smlal  v28.4s, v4.4h, v0.h[6]      \n"
-                    "smlal  v30.4s, v4.4h, v0.h[7]      \n"
-                    "smlal2 v29.4s, v4.8h, v0.h[6]      \n"
-                    "smlal2 v31.4s, v4.8h, v0.h[7]      \n"
-                    "smlal  v16.4s, v5.4h, v1.h[0]      \n"
-                    "smlal  v18.4s, v5.4h, v1.h[1]      \n"
-                    "smlal2 v17.4s, v5.8h, v1.h[0]      \n"
-                    "smlal2 v19.4s, v5.8h, v1.h[1]      \n"
-                    "smlal  v20.4s, v5.4h, v1.h[2]      \n"
-                    "smlal  v22.4s, v5.4h, v1.h[3]      \n"
-                    "smlal2 v21.4s, v5.8h, v1.h[2]      \n"
-                    "smlal2 v23.4s, v5.8h, v1.h[3]      \n"
-                    "smlal  v24.4s, v5.4h, v1.h[4]      \n"
-                    "smlal  v26.4s, v5.4h, v1.h[5]      \n"
-                    "smlal2 v25.4s, v5.8h, v1.h[4]      \n"
-                    "smlal2 v27.4s, v5.8h, v1.h[5]      \n"
-                    "smlal  v28.4s, v5.4h, v1.h[6]      \n"
-                    "smlal  v30.4s, v5.4h, v1.h[7]      \n"
-                    "smlal2 v29.4s, v5.8h, v1.h[6]      \n"
-                    "smlal2 v31.4s, v5.8h, v1.h[7]      \n"
-                    "smlal  v16.4s, v6.4h, v2.h[0]      \n"
-                    "smlal  v18.4s, v6.4h, v2.h[1]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.8h, v5.8h}, [%1], #32   \n"
-                    "smlal2 v17.4s, v6.8h, v2.h[0]      \n"
-                    "smlal2 v19.4s, v6.8h, v2.h[1]      \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.8h, v1.8h}, [%2], #32   \n"
-                    "smlal  v20.4s, v6.4h, v2.h[2]      \n"
-                    "smlal  v22.4s, v6.4h, v2.h[3]      \n"
-                    "smlal2 v21.4s, v6.8h, v2.h[2]      \n"
-                    "smlal2 v23.4s, v6.8h, v2.h[3]      \n"
-                    "smlal  v24.4s, v6.4h, v2.h[4]      \n"
-                    "smlal  v26.4s, v6.4h, v2.h[5]      \n"
-                    "smlal2 v25.4s, v6.8h, v2.h[4]      \n"
-                    "smlal2 v27.4s, v6.8h, v2.h[5]      \n"
-                    "smlal  v28.4s, v6.4h, v2.h[6]      \n"
-                    "smlal  v30.4s, v6.4h, v2.h[7]      \n"
-                    "smlal2 v29.4s, v6.8h, v2.h[6]      \n"
-                    "smlal2 v31.4s, v6.8h, v2.h[7]      \n"
-                    "smlal  v16.4s, v7.4h, v3.h[0]      \n"
-                    "smlal  v18.4s, v7.4h, v3.h[1]      \n"
-                    "smlal2 v17.4s, v7.8h, v3.h[0]      \n"
-                    "smlal2 v19.4s, v7.8h, v3.h[1]      \n"
-                    "smlal  v20.4s, v7.4h, v3.h[2]      \n"
-                    "smlal  v22.4s, v7.4h, v3.h[3]      \n"
-                    "smlal2 v21.4s, v7.8h, v3.h[2]      \n"
-                    "smlal2 v23.4s, v7.8h, v3.h[3]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "smlal  v24.4s, v7.4h, v3.h[4]      \n"
-                    "smlal  v26.4s, v7.4h, v3.h[5]      \n"
-                    "smlal2 v25.4s, v7.8h, v3.h[4]      \n"
-                    "smlal2 v27.4s, v7.8h, v3.h[5]      \n"
-                    "smlal  v28.4s, v7.4h, v3.h[6]      \n"
-                    "smlal  v30.4s, v7.4h, v3.h[7]      \n"
-                    "smlal2 v29.4s, v7.8h, v3.h[6]      \n"
-                    "smlal2 v31.4s, v7.8h, v3.h[7]      \n"
-                    "bne    2b                          \n"
-                    "sub    %1, %1, #32                 \n"
-                    "sub    %2, %2, #32                 \n"
-
-                    "3:                                 \n"
-                    "and    w4, %w6, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "ld1    {v4.8h}, [%1], #16          \n"
-                    "ld1    {v0.8h}, [%2], #16          \n"
-                    "smlal  v16.4s, v4.4h, v0.h[0]      \n"
-                    "smlal  v18.4s, v4.4h, v0.h[1]      \n"
-                    "smlal2 v17.4s, v4.8h, v0.h[0]      \n"
-                    "smlal2 v19.4s, v4.8h, v0.h[1]      \n"
-                    "smlal  v20.4s, v4.4h, v0.h[2]      \n"
-                    "smlal  v22.4s, v4.4h, v0.h[3]      \n"
-                    "smlal2 v21.4s, v4.8h, v0.h[2]      \n"
-                    "smlal2 v23.4s, v4.8h, v0.h[3]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "smlal  v24.4s, v4.4h, v0.h[4]      \n"
-                    "smlal  v26.4s, v4.4h, v0.h[5]      \n"
-                    "smlal2 v25.4s, v4.8h, v0.h[4]      \n"
-                    "smlal2 v27.4s, v4.8h, v0.h[5]      \n"
-                    "smlal  v28.4s, v4.4h, v0.h[6]      \n"
-                    "smlal  v30.4s, v4.4h, v0.h[7]      \n"
-                    "smlal2 v29.4s, v4.8h, v0.h[6]      \n"
-                    "smlal2 v31.4s, v4.8h, v0.h[7]      \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-                int32x4_t _sum2;
-                int32x4_t _sum3;
-                int32x4_t _sum4;
-                int32x4_t _sum5;
-                int32x4_t _sum6;
-                int32x4_t _sum7;
-                int32x4_t _sum8;
-                int32x4_t _sum9;
-                int32x4_t _suma;
-                int32x4_t _sumb;
-                int32x4_t _sumc;
-                int32x4_t _sumd;
-                int32x4_t _sume;
-                int32x4_t _sumf;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                    _sum2 = vdupq_n_s32(0);
-                    _sum3 = vdupq_n_s32(0);
-                    _sum4 = vdupq_n_s32(0);
-                    _sum5 = vdupq_n_s32(0);
-                    _sum6 = vdupq_n_s32(0);
-                    _sum7 = vdupq_n_s32(0);
-                    _sum8 = vdupq_n_s32(0);
-                    _sum9 = vdupq_n_s32(0);
-                    _suma = vdupq_n_s32(0);
-                    _sumb = vdupq_n_s32(0);
-                    _sumc = vdupq_n_s32(0);
-                    _sumd = vdupq_n_s32(0);
-                    _sume = vdupq_n_s32(0);
-                    _sumf = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                    _sum1 = vld1q_s32(outptr + 4);
-                    _sum2 = vld1q_s32(outptr + 8);
-                    _sum3 = vld1q_s32(outptr + 12);
-                    _sum4 = vld1q_s32(outptr + 16);
-                    _sum5 = vld1q_s32(outptr + 20);
-                    _sum6 = vld1q_s32(outptr + 24);
-                    _sum7 = vld1q_s32(outptr + 28);
-                    _sum8 = vld1q_s32(outptr + 32);
-                    _sum9 = vld1q_s32(outptr + 36);
-                    _suma = vld1q_s32(outptr + 40);
-                    _sumb = vld1q_s32(outptr + 44);
-                    _sumc = vld1q_s32(outptr + 48);
-                    _sumd = vld1q_s32(outptr + 52);
-                    _sume = vld1q_s32(outptr + 56);
-                    _sumf = vld1q_s32(outptr + 60);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x8_t _pA = vld1q_s16(pA);
-                    int16x8_t _pB = vld1q_s16(pB);
-                    _sum0 = vmlal_laneq_s16(_sum0, vget_low_s16(_pA), _pB, 0);
-                    _sum1 = vmlal_laneq_s16(_sum1, vget_high_s16(_pA), _pB, 0);
-                    _sum2 = vmlal_laneq_s16(_sum2, vget_low_s16(_pA), _pB, 1);
-                    _sum3 = vmlal_laneq_s16(_sum3, vget_high_s16(_pA), _pB, 1);
-                    _sum4 = vmlal_laneq_s16(_sum4, vget_low_s16(_pA), _pB, 2);
-                    _sum5 = vmlal_laneq_s16(_sum5, vget_high_s16(_pA), _pB, 2);
-                    _sum6 = vmlal_laneq_s16(_sum6, vget_low_s16(_pA), _pB, 3);
-                    _sum7 = vmlal_laneq_s16(_sum7, vget_high_s16(_pA), _pB, 3);
-                    _sum8 = vmlal_laneq_s16(_sum8, vget_low_s16(_pA), _pB, 4);
-                    _sum9 = vmlal_laneq_s16(_sum9, vget_high_s16(_pA), _pB, 4);
-                    _suma = vmlal_laneq_s16(_suma, vget_low_s16(_pA), _pB, 5);
-                    _sumb = vmlal_laneq_s16(_sumb, vget_high_s16(_pA), _pB, 5);
-                    _sumc = vmlal_laneq_s16(_sumc, vget_low_s16(_pA), _pB, 6);
-                    _sumd = vmlal_laneq_s16(_sumd, vget_high_s16(_pA), _pB, 6);
-                    _sume = vmlal_laneq_s16(_sume, vget_low_s16(_pA), _pB, 7);
-                    _sumf = vmlal_laneq_s16(_sumf, vget_high_s16(_pA), _pB, 7);
-                    pA += 8;
-                    pB += 8;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-                vst1q_s32(outptr + 16, _sum4);
-                vst1q_s32(outptr + 20, _sum5);
-                vst1q_s32(outptr + 24, _sum6);
-                vst1q_s32(outptr + 28, _sum7);
-                vst1q_s32(outptr + 32, _sum8);
-                vst1q_s32(outptr + 36, _sum9);
-                vst1q_s32(outptr + 40, _suma);
-                vst1q_s32(outptr + 44, _sumb);
-                vst1q_s32(outptr + 48, _sumc);
-                vst1q_s32(outptr + 52, _sumd);
-                vst1q_s32(outptr + 56, _sume);
-                vst1q_s32(outptr + 60, _sumf);
-                outptr += 64;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-#endif // __aarch64__
-            for (; jj + 5 < max_jj; jj += 6)
-            {
-                const short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "prfm   pldl1keep, [%2, #384]       \n"
-                    "cmp    %w7, #0                     \n"
-                    "beq    0f                          \n"
-
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "sub    %0, %0, #128                \n"
-                    "b      1f                          \n"
-
-                    "0:                                 \n"
-                    "eor    v20.16b, v20.16b, v20.16b   \n"
-                    "eor    v21.16b, v21.16b, v21.16b   \n"
-                    "eor    v22.16b, v22.16b, v22.16b   \n"
-                    "eor    v23.16b, v23.16b, v23.16b   \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-                    "eor    v25.16b, v25.16b, v25.16b   \n"
-                    "eor    v26.16b, v26.16b, v26.16b   \n"
-                    "eor    v27.16b, v27.16b, v27.16b   \n"
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-                    "eor    v29.16b, v29.16b, v29.16b   \n"
-                    "eor    v30.16b, v30.16b, v30.16b   \n"
-                    "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                    "1:                                 \n"
-                    "lsr    w4, %w6, #3                 \n" // w4 = max_kk >> 3
-                    "cmp    w4, #0                      \n"
-                    "beq    3f                          \n"
-
-                    "ld1    {v6.8h, v7.8h}, [%1], #32   \n"
-                    "ld1    {v0.8h, v1.8h}, [%2], #32   \n"
-                    ".align 4                           \n"
-                    "2:                                 \n"
-                    "smlal  v20.4s, v6.4h, v0.h[0]      \n"
-                    "smlal  v22.4s, v6.4h, v0.h[1]      \n"
-                    "ld1    {v8.8h, v9.8h}, [%1], #32   \n"
-                    "smlal2 v21.4s, v6.8h, v0.h[0]      \n"
-                    "smlal2 v23.4s, v6.8h, v0.h[1]      \n"
-                    "ld1    {v2.8h, v3.8h}, [%2], #32   \n"
-                    "smlal  v24.4s, v6.4h, v0.h[2]      \n"
-                    "smlal  v26.4s, v6.4h, v0.h[3]      \n"
-                    "smlal2 v25.4s, v6.8h, v0.h[2]      \n"
-                    "smlal2 v27.4s, v6.8h, v0.h[3]      \n"
-                    "smlal  v28.4s, v6.4h, v0.h[4]      \n"
-                    "smlal  v30.4s, v6.4h, v0.h[5]      \n"
-                    "smlal2 v29.4s, v6.8h, v0.h[4]      \n"
-                    "smlal2 v31.4s, v6.8h, v0.h[5]      \n"
-                    "smlal  v20.4s, v7.4h, v0.h[6]      \n"
-                    "smlal  v22.4s, v7.4h, v0.h[7]      \n"
-                    "smlal2 v21.4s, v7.8h, v0.h[6]      \n"
-                    "smlal2 v23.4s, v7.8h, v0.h[7]      \n"
-                    "smlal  v24.4s, v7.4h, v1.h[0]      \n"
-                    "smlal  v26.4s, v7.4h, v1.h[1]      \n"
-                    "smlal2 v25.4s, v7.8h, v1.h[0]      \n"
-                    "smlal2 v27.4s, v7.8h, v1.h[1]      \n"
-                    "smlal  v28.4s, v7.4h, v1.h[2]      \n"
-                    "smlal  v30.4s, v7.4h, v1.h[3]      \n"
-                    "smlal2 v29.4s, v7.8h, v1.h[2]      \n"
-                    "smlal2 v31.4s, v7.8h, v1.h[3]      \n"
-                    "smlal  v20.4s, v8.4h, v1.h[4]      \n"
-                    "smlal  v22.4s, v8.4h, v1.h[5]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v6.8h, v7.8h}, [%1], #32   \n"
-                    "smlal2 v21.4s, v8.8h, v1.h[4]      \n"
-                    "smlal2 v23.4s, v8.8h, v1.h[5]      \n"
-                    "smlal  v24.4s, v8.4h, v1.h[6]      \n"
-                    "smlal  v26.4s, v8.4h, v1.h[7]      \n"
-                    "smlal2 v25.4s, v8.8h, v1.h[6]      \n"
-                    "smlal2 v27.4s, v8.8h, v1.h[7]      \n"
-                    "smlal  v28.4s, v8.4h, v2.h[0]      \n"
-                    "smlal  v30.4s, v8.4h, v2.h[1]      \n"
-                    "ld1    {v4.8h, v5.8h}, [%2], #32   \n"
-                    "smlal2 v29.4s, v8.8h, v2.h[0]      \n"
-                    "smlal2 v31.4s, v8.8h, v2.h[1]      \n"
-                    "smlal  v20.4s, v9.4h, v2.h[2]      \n"
-                    "smlal  v22.4s, v9.4h, v2.h[3]      \n"
-                    "smlal2 v21.4s, v9.8h, v2.h[2]      \n"
-                    "smlal2 v23.4s, v9.8h, v2.h[3]      \n"
-                    "smlal  v24.4s, v9.4h, v2.h[4]      \n"
-                    "smlal  v26.4s, v9.4h, v2.h[5]      \n"
-                    "smlal2 v25.4s, v9.8h, v2.h[4]      \n"
-                    "smlal2 v27.4s, v9.8h, v2.h[5]      \n"
-                    "smlal  v28.4s, v9.4h, v2.h[6]      \n"
-                    "smlal  v30.4s, v9.4h, v2.h[7]      \n"
-                    "smlal2 v29.4s, v9.8h, v2.h[6]      \n"
-                    "smlal2 v31.4s, v9.8h, v2.h[7]      \n"
-                    "smlal  v20.4s, v6.4h, v3.h[0]      \n"
-                    "smlal  v22.4s, v6.4h, v3.h[1]      \n"
-                    "ld1    {v8.8h, v9.8h}, [%1], #32   \n"
-                    "smlal2 v21.4s, v6.8h, v3.h[0]      \n"
-                    "smlal2 v23.4s, v6.8h, v3.h[1]      \n"
-                    "smlal  v24.4s, v6.4h, v3.h[2]      \n"
-                    "smlal  v26.4s, v6.4h, v3.h[3]      \n"
-                    "smlal2 v25.4s, v6.8h, v3.h[2]      \n"
-                    "smlal2 v27.4s, v6.8h, v3.h[3]      \n"
-                    "smlal  v28.4s, v6.4h, v3.h[4]      \n"
-                    "smlal  v30.4s, v6.4h, v3.h[5]      \n"
-                    "smlal2 v29.4s, v6.8h, v3.h[4]      \n"
-                    "smlal2 v31.4s, v6.8h, v3.h[5]      \n"
-                    "smlal  v20.4s, v7.4h, v3.h[6]      \n"
-                    "smlal  v22.4s, v7.4h, v3.h[7]      \n"
-                    "smlal2 v21.4s, v7.8h, v3.h[6]      \n"
-                    "smlal2 v23.4s, v7.8h, v3.h[7]      \n"
-                    "smlal  v24.4s, v7.4h, v4.h[0]      \n"
-                    "smlal  v26.4s, v7.4h, v4.h[1]      \n"
-                    "prfm   pldl1keep, [%2, #384]       \n"
-                    "ld1    {v0.8h, v1.8h}, [%2], #32   \n"
-                    "smlal2 v25.4s, v7.8h, v4.h[0]      \n"
-                    "smlal2 v27.4s, v7.8h, v4.h[1]      \n"
-                    "smlal  v28.4s, v7.4h, v4.h[2]      \n"
-                    "smlal  v30.4s, v7.4h, v4.h[3]      \n"
-                    "smlal2 v29.4s, v7.8h, v4.h[2]      \n"
-                    "smlal2 v31.4s, v7.8h, v4.h[3]      \n"
-                    "smlal  v20.4s, v8.4h, v4.h[4]      \n"
-                    "smlal  v22.4s, v8.4h, v4.h[5]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v6.8h, v7.8h}, [%1], #32   \n"
-                    "smlal2 v21.4s, v8.8h, v4.h[4]      \n"
-                    "smlal2 v23.4s, v8.8h, v4.h[5]      \n"
-                    "smlal  v24.4s, v8.4h, v4.h[6]      \n"
-                    "smlal  v26.4s, v8.4h, v4.h[7]      \n"
-                    "smlal2 v25.4s, v8.8h, v4.h[6]      \n"
-                    "smlal2 v27.4s, v8.8h, v4.h[7]      \n"
-                    "smlal  v28.4s, v8.4h, v5.h[0]      \n"
-                    "smlal  v30.4s, v8.4h, v5.h[1]      \n"
-                    "smlal2 v29.4s, v8.8h, v5.h[0]      \n"
-                    "smlal2 v31.4s, v8.8h, v5.h[1]      \n"
-                    "smlal  v20.4s, v9.4h, v5.h[2]      \n"
-                    "smlal  v22.4s, v9.4h, v5.h[3]      \n"
-                    "smlal2 v21.4s, v9.8h, v5.h[2]      \n"
-                    "smlal2 v23.4s, v9.8h, v5.h[3]      \n"
-                    "smlal  v24.4s, v9.4h, v5.h[4]      \n"
-                    "smlal  v26.4s, v9.4h, v5.h[5]      \n"
-                    "smlal2 v25.4s, v9.8h, v5.h[4]      \n"
-                    "smlal2 v27.4s, v9.8h, v5.h[5]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "smlal  v28.4s, v9.4h, v5.h[6]      \n"
-                    "smlal  v30.4s, v9.4h, v5.h[7]      \n"
-                    "smlal2 v29.4s, v9.8h, v5.h[6]      \n"
-                    "smlal2 v31.4s, v9.8h, v5.h[7]      \n"
-                    "bne    2b                          \n"
-                    "sub    %1, %1, #32                 \n"
-                    "sub    %2, %2, #32                 \n"
-
-                    "3:                                 \n"
-                    "and    w4, %w6, #7                 \n" // w4 = remain = max_kk & 7
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "ld1    {v4.8h}, [%1], #16          \n"
-                    "ld1    {v0.8h}, [%2]               \n"
-                    "add    %2, %2, #12                 \n"
-                    "smlal  v20.4s, v4.4h, v0.h[0]      \n"
-                    "smlal  v22.4s, v4.4h, v0.h[1]      \n"
-                    "smlal2 v21.4s, v4.8h, v0.h[0]      \n"
-                    "smlal2 v23.4s, v4.8h, v0.h[1]      \n"
-                    "smlal  v24.4s, v4.4h, v0.h[2]      \n"
-                    "smlal  v26.4s, v4.4h, v0.h[3]      \n"
-                    "smlal2 v25.4s, v4.8h, v0.h[2]      \n"
-                    "smlal2 v27.4s, v4.8h, v0.h[3]      \n"
-                    "smlal  v28.4s, v4.4h, v0.h[4]      \n"
-                    "smlal  v30.4s, v4.4h, v0.h[5]      \n"
-                    "smlal2 v29.4s, v4.8h, v0.h[4]      \n"
-                    "smlal2 v31.4s, v4.8h, v0.h[5]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%1, #512]          \n"
-                    "pld        [%2, #384]          \n"
-                    "cmp        %7, #0              \n"
-                    "beq        0f                  \n"
-
-                    "vldm       %0!, {d8-d15}       \n"
-                    "vldm       %0, {d16-d31}       \n"
-                    "sub        %0, %0, #64         \n"
-                    "b          1f                  \n"
-
-                    "0:                             \n"
-                    "veor       q4, q4              \n"
-                    "veor       q5, q5              \n"
-                    "veor       q6, q6              \n"
-                    "veor       q7, q7              \n"
-                    "veor       q8, q8              \n"
-                    "veor       q9, q9              \n"
-                    "veor       q10, q10            \n"
-                    "veor       q11, q11            \n"
-                    "veor       q12, q12            \n"
-                    "veor       q13, q13            \n"
-                    "veor       q14, q14            \n"
-                    "veor       q15, q15            \n"
-
-                    "1:                             \n"
-                    "lsr        r4, %6, #3          \n" // r4 = max_kk >> 3
-                    "cmp        r4, #0              \n"
-                    "beq        3f                  \n"
-
-                    "vld1.s16   {d4-d5}, [%1]!      \n"
-                    "vld1.s16   {d0-d1}, [%2]!      \n"
-                    ".align 4                       \n"
-                    "2:                             \n"
-                    "vmlal.s16  q4, d4, d0[0]       \n"
-                    "vld1.s16   {d6-d7}, [%1]!      \n"
-                    "vmlal.s16  q6, d4, d0[1]       \n"
-                    "vld1.s16   {d2-d3}, [%2]!      \n"
-                    "vmlal.s16  q8, d4, d0[2]       \n"
-                    "vmlal.s16  q10, d4, d0[3]      \n"
-                    "vmlal.s16  q5, d5, d0[0]       \n"
-                    "vmlal.s16  q7, d5, d0[1]       \n"
-                    "vmlal.s16  q9, d5, d0[2]       \n"
-                    "vmlal.s16  q11, d5, d0[3]      \n"
-                    "vmlal.s16  q12, d4, d1[0]      \n"
-                    "vmlal.s16  q14, d4, d1[1]      \n"
-                    "vmlal.s16  q13, d5, d1[0]      \n"
-                    "vmlal.s16  q15, d5, d1[1]      \n"
-                    "vmlal.s16  q4, d6, d1[2]       \n"
-                    "vld1.s16   {d4-d5}, [%1]!      \n"
-                    "vmlal.s16  q6, d6, d1[3]       \n"
-                    "vmlal.s16  q5, d7, d1[2]       \n"
-                    "vmlal.s16  q7, d7, d1[3]       \n"
-                    "vmlal.s16  q8, d6, d2[0]       \n"
-                    "pld        [%2, #384]          \n"
-                    "vld1.s16   {d0-d1}, [%2]!      \n"
-                    "vmlal.s16  q10, d6, d2[1]      \n"
-                    "vmlal.s16  q12, d6, d2[2]      \n"
-                    "vmlal.s16  q14, d6, d2[3]      \n"
-                    "vmlal.s16  q9, d7, d2[0]       \n"
-                    "vmlal.s16  q11, d7, d2[1]      \n"
-                    "vmlal.s16  q13, d7, d2[2]      \n"
-                    "vmlal.s16  q15, d7, d2[3]      \n"
-                    "vmlal.s16  q4, d4, d3[0]       \n"
-                    "vld1.s16   {d6-d7}, [%1]!      \n"
-                    "vmlal.s16  q6, d4, d3[1]       \n"
-                    "vmlal.s16  q8, d4, d3[2]       \n"
-                    "vmlal.s16  q10, d4, d3[3]      \n"
-                    "vmlal.s16  q5, d5, d3[0]       \n"
-                    "vmlal.s16  q7, d5, d3[1]       \n"
-                    "vmlal.s16  q9, d5, d3[2]       \n"
-                    "vmlal.s16  q11, d5, d3[3]      \n"
-                    "vmlal.s16  q12, d4, d0[0]      \n"
-                    "vld1.s16   {d2-d3}, [%2]!      \n"
-                    "vmlal.s16  q14, d4, d0[1]      \n"
-                    "vmlal.s16  q13, d5, d0[0]      \n"
-                    "vmlal.s16  q15, d5, d0[1]      \n"
-                    "vmlal.s16  q4, d6, d0[2]       \n"
-                    "pld        [%1, #512]          \n"
-                    "vld1.s16   {d4-d5}, [%1]!      \n"
-                    "vmlal.s16  q6, d6, d0[3]       \n"
-                    "vmlal.s16  q5, d7, d0[2]       \n"
-                    "vmlal.s16  q7, d7, d0[3]       \n"
-                    "vmlal.s16  q8, d6, d1[0]       \n"
-                    "vmlal.s16  q10, d6, d1[1]      \n"
-                    "vmlal.s16  q12, d6, d1[2]      \n"
-                    "vmlal.s16  q14, d6, d1[3]      \n"
-                    "vmlal.s16  q9, d7, d1[0]       \n"
-                    "vmlal.s16  q11, d7, d1[1]      \n"
-                    "vmlal.s16  q13, d7, d1[2]      \n"
-                    "vmlal.s16  q15, d7, d1[3]      \n"
-                    "vmlal.s16  q4, d4, d2[0]       \n"
-                    "vld1.s16   {d6-d7}, [%1]!      \n"
-                    "vmlal.s16  q6, d4, d2[1]       \n"
-                    "vld1.s16   {d0-d1}, [%2]!      \n"
-                    "vmlal.s16  q8, d4, d2[2]       \n"
-                    "vmlal.s16  q10, d4, d2[3]      \n"
-                    "vmlal.s16  q5, d5, d2[0]       \n"
-                    "vmlal.s16  q7, d5, d2[1]       \n"
-                    "vmlal.s16  q9, d5, d2[2]       \n"
-                    "vmlal.s16  q11, d5, d2[3]      \n"
-                    "vmlal.s16  q12, d4, d3[0]      \n"
-                    "vmlal.s16  q14, d4, d3[1]      \n"
-                    "vmlal.s16  q13, d5, d3[0]      \n"
-                    "vmlal.s16  q15, d5, d3[1]      \n"
-                    "vmlal.s16  q4, d6, d3[2]       \n"
-                    "vld1.s16   {d4-d5}, [%1]!      \n"
-                    "vmlal.s16  q6, d6, d3[3]       \n"
-                    "vmlal.s16  q5, d7, d3[2]       \n"
-                    "vmlal.s16  q7, d7, d3[3]       \n"
-                    "vmlal.s16  q8, d6, d0[0]       \n"
-                    "pld        [%2, #384]          \n"
-                    "vld1.s16   {d2-d3}, [%2]!      \n"
-                    "vmlal.s16  q10, d6, d0[1]      \n"
-                    "vmlal.s16  q12, d6, d0[2]      \n"
-                    "vmlal.s16  q14, d6, d0[3]      \n"
-                    "vmlal.s16  q9, d7, d0[0]       \n"
-                    "vmlal.s16  q11, d7, d0[1]      \n"
-                    "vmlal.s16  q13, d7, d0[2]      \n"
-                    "vmlal.s16  q15, d7, d0[3]      \n"
-                    "vmlal.s16  q4, d4, d1[0]       \n"
-                    "vld1.s16   {d6-d7}, [%1]!      \n"
-                    "vmlal.s16  q6, d4, d1[1]       \n"
-                    "vmlal.s16  q8, d4, d1[2]       \n"
-                    "vmlal.s16  q10, d4, d1[3]      \n"
-                    "vmlal.s16  q5, d5, d1[0]       \n"
-                    "vmlal.s16  q7, d5, d1[1]       \n"
-                    "vmlal.s16  q9, d5, d1[2]       \n"
-                    "vmlal.s16  q11, d5, d1[3]      \n"
-                    "vmlal.s16  q12, d4, d2[0]      \n"
-                    "vld1.s16   {d0-d1}, [%2]!      \n"
-                    "vmlal.s16  q14, d4, d2[1]      \n"
-                    "vmlal.s16  q13, d5, d2[0]      \n"
-                    "vmlal.s16  q15, d5, d2[1]      \n"
-                    "vmlal.s16  q4, d6, d2[2]       \n"
-                    "pld        [%1, #512]          \n"
-                    "vld1.s16   {d4-d5}, [%1]!      \n"
-                    "vmlal.s16  q6, d6, d2[3]       \n"
-                    "vmlal.s16  q5, d7, d2[2]       \n"
-                    "vmlal.s16  q7, d7, d2[3]       \n"
-                    "vmlal.s16  q8, d6, d3[0]       \n"
-                    "vmlal.s16  q10, d6, d3[1]      \n"
-                    "vmlal.s16  q12, d6, d3[2]      \n"
-                    "vmlal.s16  q14, d6, d3[3]      \n"
-                    "vmlal.s16  q9, d7, d3[0]       \n"
-                    "vmlal.s16  q11, d7, d3[1]      \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmlal.s16  q13, d7, d3[2]      \n"
-                    "vmlal.s16  q15, d7, d3[3]      \n"
-                    "bne        2b                  \n"
-                    "sub        %1, %1, #16         \n"
-                    "sub        %2, %2, #16         \n"
-
-                    "3:                             \n"
-                    "and        r4, %6, #7          \n" // w4 = remain = max_kk & 7
-                    "cmp        r4, #0              \n"
-                    "beq        5f                  \n"
-
-                    "4:                             \n"
-                    "vld1.s16   {d0-d1}, [%1]!      \n"
-                    "vld1.s16   {d2-d3}, [%2]       \n"
-                    "add        %2, %2, #12         \n"
-                    "vmlal.s16  q4, d0, d2[0]       \n"
-                    "vmlal.s16  q6, d0, d2[1]       \n"
-                    "vmlal.s16  q8, d0, d2[2]       \n"
-                    "vmlal.s16  q10, d0, d2[3]      \n"
-                    "vmlal.s16  q5, d1, d2[0]       \n"
-                    "vmlal.s16  q7, d1, d2[1]       \n"
-                    "vmlal.s16  q9, d1, d2[2]       \n"
-                    "vmlal.s16  q11, d1, d2[3]      \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmlal.s16  q12, d0, d3[0]      \n"
-                    "vmlal.s16  q14, d0, d3[1]      \n"
-                    "vmlal.s16  q13, d1, d3[0]      \n"
-                    "vmlal.s16  q15, d1, d3[1]      \n"
-                    "bne        4b                  \n"
-
-                    "5:                             \n"
-                    "vstm       %0!, {d8-d15}       \n"
-                    "vstm       %0!, {d16-d31}      \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-                int32x4_t _sum2;
-                int32x4_t _sum3;
-                int32x4_t _sum4;
-                int32x4_t _sum5;
-                int32x4_t _sum6;
-                int32x4_t _sum7;
-                int32x4_t _sum8;
-                int32x4_t _sum9;
-                int32x4_t _suma;
-                int32x4_t _sumb;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                    _sum2 = vdupq_n_s32(0);
-                    _sum3 = vdupq_n_s32(0);
-                    _sum4 = vdupq_n_s32(0);
-                    _sum5 = vdupq_n_s32(0);
-                    _sum6 = vdupq_n_s32(0);
-                    _sum7 = vdupq_n_s32(0);
-                    _sum8 = vdupq_n_s32(0);
-                    _sum9 = vdupq_n_s32(0);
-                    _suma = vdupq_n_s32(0);
-                    _sumb = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                    _sum1 = vld1q_s32(outptr + 4);
-                    _sum2 = vld1q_s32(outptr + 8);
-                    _sum3 = vld1q_s32(outptr + 12);
-                    _sum4 = vld1q_s32(outptr + 16);
-                    _sum5 = vld1q_s32(outptr + 20);
-                    _sum6 = vld1q_s32(outptr + 24);
-                    _sum7 = vld1q_s32(outptr + 28);
-                    _sum8 = vld1q_s32(outptr + 32);
-                    _sum9 = vld1q_s32(outptr + 36);
-                    _suma = vld1q_s32(outptr + 40);
-                    _sumb = vld1q_s32(outptr + 44);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x8_t _pA = vld1q_s16(pA);
-                    int16x8_t _pB = vld1q_s16(pB);
-                    _sum0 = vmlal_lane_s16(_sum0, vget_low_s16(_pA), vget_low_s16(_pB), 0);
-                    _sum1 = vmlal_lane_s16(_sum1, vget_high_s16(_pA), vget_low_s16(_pB), 0);
-                    _sum2 = vmlal_lane_s16(_sum2, vget_low_s16(_pA), vget_low_s16(_pB), 1);
-                    _sum3 = vmlal_lane_s16(_sum3, vget_high_s16(_pA), vget_low_s16(_pB), 1);
-                    _sum4 = vmlal_lane_s16(_sum4, vget_low_s16(_pA), vget_low_s16(_pB), 2);
-                    _sum5 = vmlal_lane_s16(_sum5, vget_high_s16(_pA), vget_low_s16(_pB), 2);
-                    _sum6 = vmlal_lane_s16(_sum6, vget_low_s16(_pA), vget_low_s16(_pB), 3);
-                    _sum7 = vmlal_lane_s16(_sum7, vget_high_s16(_pA), vget_low_s16(_pB), 3);
-                    _sum8 = vmlal_lane_s16(_sum8, vget_low_s16(_pA), vget_high_s16(_pB), 0);
-                    _sum9 = vmlal_lane_s16(_sum9, vget_high_s16(_pA), vget_high_s16(_pB), 0);
-                    _suma = vmlal_lane_s16(_suma, vget_low_s16(_pA), vget_high_s16(_pB), 1);
-                    _sumb = vmlal_lane_s16(_sumb, vget_high_s16(_pA), vget_high_s16(_pB), 1);
-                    pA += 8;
-                    pB += 6;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-                vst1q_s32(outptr + 16, _sum4);
-                vst1q_s32(outptr + 20, _sum5);
-                vst1q_s32(outptr + 24, _sum6);
-                vst1q_s32(outptr + 28, _sum7);
-                vst1q_s32(outptr + 32, _sum8);
-                vst1q_s32(outptr + 36, _sum9);
-                vst1q_s32(outptr + 40, _suma);
-                vst1q_s32(outptr + 44, _sumb);
-                outptr += 48;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "cmp    %w7, #0                     \n"
-                    "beq    0f                          \n"
-
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "sub    %0, %0, #64                 \n"
-                    "b      1f                          \n"
-
-                    "0:                                 \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-                    "eor    v25.16b, v25.16b, v25.16b   \n"
-                    "eor    v26.16b, v26.16b, v26.16b   \n"
-                    "eor    v27.16b, v27.16b, v27.16b   \n"
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-                    "eor    v29.16b, v29.16b, v29.16b   \n"
-                    "eor    v30.16b, v30.16b, v30.16b   \n"
-                    "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                    "1:                                 \n"
-                    "lsr    w4, %w6, #3                 \n" // w4 = max_kk >> 3
-                    "cmp    w4, #0                      \n"
-                    "beq    3f                          \n"
-
-                    "ld1    {v4.8h, v5.8h}, [%1], #32   \n"
-                    "ld1    {v0.8h, v1.8h}, [%2], #32   \n"
-                    ".align 4                           \n"
-                    "2:                                 \n"
-                    "smlal  v24.4s, v4.4h, v0.h[0]      \n"
-                    "smlal  v26.4s, v4.4h, v0.h[1]      \n"
-                    "ld1    {v6.8h, v7.8h}, [%1], #32   \n"
-                    "smlal2 v25.4s, v4.8h, v0.h[0]      \n"
-                    "smlal2 v27.4s, v4.8h, v0.h[1]      \n"
-                    "ld1    {v2.8h, v3.8h}, [%2], #32   \n"
-                    "smlal  v28.4s, v4.4h, v0.h[2]      \n"
-                    "smlal  v30.4s, v4.4h, v0.h[3]      \n"
-                    "smlal2 v29.4s, v4.8h, v0.h[2]      \n"
-                    "smlal2 v31.4s, v4.8h, v0.h[3]      \n"
-                    "smlal  v24.4s, v5.4h, v0.h[4]      \n"
-                    "smlal  v26.4s, v5.4h, v0.h[5]      \n"
-                    "smlal2 v25.4s, v5.8h, v0.h[4]      \n"
-                    "smlal2 v27.4s, v5.8h, v0.h[5]      \n"
-                    "smlal  v28.4s, v5.4h, v0.h[6]      \n"
-                    "smlal  v30.4s, v5.4h, v0.h[7]      \n"
-                    "smlal2 v29.4s, v5.8h, v0.h[6]      \n"
-                    "smlal2 v31.4s, v5.8h, v0.h[7]      \n"
-                    "smlal  v24.4s, v6.4h, v1.h[0]      \n"
-                    "smlal  v26.4s, v6.4h, v1.h[1]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.8h, v5.8h}, [%1], #32   \n"
-                    "smlal2 v25.4s, v6.8h, v1.h[0]      \n"
-                    "smlal2 v27.4s, v6.8h, v1.h[1]      \n"
-                    "smlal  v28.4s, v6.4h, v1.h[2]      \n"
-                    "smlal  v30.4s, v6.4h, v1.h[3]      \n"
-                    "smlal2 v29.4s, v6.8h, v1.h[2]      \n"
-                    "smlal2 v31.4s, v6.8h, v1.h[3]      \n"
-                    "smlal  v24.4s, v7.4h, v1.h[4]      \n"
-                    "smlal  v26.4s, v7.4h, v1.h[5]      \n"
-                    "smlal2 v25.4s, v7.8h, v1.h[4]      \n"
-                    "smlal2 v27.4s, v7.8h, v1.h[5]      \n"
-                    "smlal  v28.4s, v7.4h, v1.h[6]      \n"
-                    "smlal  v30.4s, v7.4h, v1.h[7]      \n"
-                    "smlal2 v29.4s, v7.8h, v1.h[6]      \n"
-                    "smlal2 v31.4s, v7.8h, v1.h[7]      \n"
-                    "smlal  v24.4s, v4.4h, v2.h[0]      \n"
-                    "smlal  v26.4s, v4.4h, v2.h[1]      \n"
-                    "ld1    {v6.8h, v7.8h}, [%1], #32   \n"
-                    "smlal2 v25.4s, v4.8h, v2.h[0]      \n"
-                    "smlal2 v27.4s, v4.8h, v2.h[1]      \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.8h, v1.8h}, [%2], #32   \n"
-                    "smlal  v28.4s, v4.4h, v2.h[2]      \n"
-                    "smlal  v30.4s, v4.4h, v2.h[3]      \n"
-                    "smlal2 v29.4s, v4.8h, v2.h[2]      \n"
-                    "smlal2 v31.4s, v4.8h, v2.h[3]      \n"
-                    "smlal  v24.4s, v5.4h, v2.h[4]      \n"
-                    "smlal  v26.4s, v5.4h, v2.h[5]      \n"
-                    "smlal2 v25.4s, v5.8h, v2.h[4]      \n"
-                    "smlal2 v27.4s, v5.8h, v2.h[5]      \n"
-                    "smlal  v28.4s, v5.4h, v2.h[6]      \n"
-                    "smlal  v30.4s, v5.4h, v2.h[7]      \n"
-                    "smlal2 v29.4s, v5.8h, v2.h[6]      \n"
-                    "smlal2 v31.4s, v5.8h, v2.h[7]      \n"
-                    "smlal  v24.4s, v6.4h, v3.h[0]      \n"
-                    "smlal  v26.4s, v6.4h, v3.h[1]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.8h, v5.8h}, [%1], #32   \n"
-                    "smlal2 v25.4s, v6.8h, v3.h[0]      \n"
-                    "smlal2 v27.4s, v6.8h, v3.h[1]      \n"
-                    "smlal  v28.4s, v6.4h, v3.h[2]      \n"
-                    "smlal  v30.4s, v6.4h, v3.h[3]      \n"
-                    "smlal2 v29.4s, v6.8h, v3.h[2]      \n"
-                    "smlal2 v31.4s, v6.8h, v3.h[3]      \n"
-                    "smlal  v24.4s, v7.4h, v3.h[4]      \n"
-                    "smlal  v26.4s, v7.4h, v3.h[5]      \n"
-                    "smlal2 v25.4s, v7.8h, v3.h[4]      \n"
-                    "smlal2 v27.4s, v7.8h, v3.h[5]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "smlal  v28.4s, v7.4h, v3.h[6]      \n"
-                    "smlal  v30.4s, v7.4h, v3.h[7]      \n"
-                    "smlal2 v29.4s, v7.8h, v3.h[6]      \n"
-                    "smlal2 v31.4s, v7.8h, v3.h[7]      \n"
-                    "bne    2b                          \n"
-                    "sub    %1, %1, #32                 \n"
-                    "sub    %2, %2, #32                 \n"
-
-                    "3:                                 \n"
-                    "and    w4, %w6, #7                 \n" // w4 = remain = max_kk & 7
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "ld1    {v4.8h}, [%1], #16          \n"
-                    "ld1    {v0.4h}, [%2], #8           \n"
-                    "smlal  v24.4s, v4.4h, v0.h[0]      \n"
-                    "smlal  v26.4s, v4.4h, v0.h[1]      \n"
-                    "smlal2 v25.4s, v4.8h, v0.h[0]      \n"
-                    "smlal2 v27.4s, v4.8h, v0.h[1]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "smlal  v28.4s, v4.4h, v0.h[2]      \n"
-                    "smlal  v30.4s, v4.4h, v0.h[3]      \n"
-                    "smlal2 v29.4s, v4.8h, v0.h[2]      \n"
-                    "smlal2 v31.4s, v4.8h, v0.h[3]      \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%1, #512]          \n"
-                    "pld        [%2, #256]          \n"
-                    "cmp        %7, #0              \n"
-                    "beq        0f                  \n"
-
-                    "vldm       %0, {d16-d31}       \n"
-                    "b          1f                  \n"
-
-                    "0:                             \n"
-                    "veor       q8, q8              \n"
-                    "veor       q9, q9              \n"
-                    "veor       q10, q10            \n"
-                    "veor       q11, q11            \n"
-                    "veor       q12, q12            \n"
-                    "veor       q13, q13            \n"
-                    "veor       q14, q14            \n"
-                    "veor       q15, q15            \n"
-
-                    "1:                             \n"
-                    "lsr        r4, %6, #2          \n" // r4 = max_kk >> 2
-                    "cmp        r4, #0              \n"
-                    "beq        3f                  \n"
-
-                    "vld1.s16   {d4-d5}, [%1]!      \n"
-                    "vld1.s16   {d0-d1}, [%2]!      \n"
-                    ".align 4                       \n"
-                    "2:                             \n"
-                    "vmlal.s16  q8, d4, d0[0]       \n"
-                    "vld1.s16   {d6-d7}, [%1]!      \n"
-                    "vmlal.s16  q10, d4, d0[1]      \n"
-                    "vmlal.s16  q12, d4, d0[2]      \n"
-                    "vmlal.s16  q14, d4, d0[3]      \n"
-                    "vmlal.s16  q9, d5, d0[0]       \n"
-                    "vld1.s16   {d8-d9}, [%1]!      \n"
-                    "vmlal.s16  q11, d5, d0[1]      \n"
-                    "vld1.s16   {d2-d3}, [%2]!      \n"
-                    "vmlal.s16  q13, d5, d0[2]      \n"
-                    "vmlal.s16  q15, d5, d0[3]      \n"
-                    "vmlal.s16  q8, d6, d1[0]       \n"
-                    "vmlal.s16  q10, d6, d1[1]      \n"
-                    "vmlal.s16  q12, d6, d1[2]      \n"
-                    "vmlal.s16  q14, d6, d1[3]      \n"
-                    "vmlal.s16  q9, d7, d1[0]       \n"
-                    "vld1.s16   {d10-d11}, [%1]!    \n"
-                    "vmlal.s16  q11, d7, d1[1]      \n"
-                    "vmlal.s16  q13, d7, d1[2]      \n"
-                    "vmlal.s16  q15, d7, d1[3]      \n"
-                    "vmlal.s16  q8, d8, d2[0]       \n"
-                    "vmlal.s16  q10, d8, d2[1]      \n"
-                    "vmlal.s16  q12, d8, d2[2]      \n"
-                    "vmlal.s16  q14, d8, d2[3]      \n"
-                    "vmlal.s16  q9, d9, d2[0]       \n"
-                    "pld        [%1, #512]          \n"
-                    "vld1.s16   {d4-d5}, [%1]!      \n"
-                    "vmlal.s16  q11, d9, d2[1]      \n"
-                    "pld        [%2, #256]          \n"
-                    "vld1.s16   {d0-d1}, [%2]!      \n"
-                    "vmlal.s16  q13, d9, d2[2]      \n"
-                    "vmlal.s16  q15, d9, d2[3]      \n"
-                    "vmlal.s16  q8, d10, d3[0]      \n"
-                    "vmlal.s16  q10, d10, d3[1]     \n"
-                    "vmlal.s16  q12, d10, d3[2]     \n"
-                    "vmlal.s16  q14, d10, d3[3]     \n"
-                    "vmlal.s16  q9, d11, d3[0]      \n"
-                    "vmlal.s16  q11, d11, d3[1]     \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmlal.s16  q13, d11, d3[2]     \n"
-                    "vmlal.s16  q15, d11, d3[3]     \n"
-                    "bne        2b                  \n"
-                    "sub        %1, %1, #16         \n"
-                    "sub        %2, %2, #16         \n"
-
-                    "3:                             \n"
-                    "and        r4, %6, #3          \n" // w4 = remain = max_kk & 3
-                    "cmp        r4, #0              \n"
-                    "beq        5f                  \n"
-
-                    "4:                             \n"
-                    "vld1.s16   {d0-d1}, [%1]!      \n"
-                    "vld1.s16   {d2}, [%2]!         \n"
-                    "vmlal.s16  q8, d0, d2[0]       \n"
-                    "vmlal.s16  q10, d0, d2[1]      \n"
-                    "vmlal.s16  q12, d0, d2[2]      \n"
-                    "vmlal.s16  q14, d0, d2[3]      \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmlal.s16  q9, d1, d2[0]       \n"
-                    "vmlal.s16  q11, d1, d2[1]      \n"
-                    "vmlal.s16  q13, d1, d2[2]      \n"
-                    "vmlal.s16  q15, d1, d2[3]      \n"
-                    "bne        4b                  \n"
-
-                    "5:                             \n"
-                    "vstm       %0!, {d16-d31}      \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-                int32x4_t _sum2;
-                int32x4_t _sum3;
-                int32x4_t _sum4;
-                int32x4_t _sum5;
-                int32x4_t _sum6;
-                int32x4_t _sum7;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                    _sum2 = vdupq_n_s32(0);
-                    _sum3 = vdupq_n_s32(0);
-                    _sum4 = vdupq_n_s32(0);
-                    _sum5 = vdupq_n_s32(0);
-                    _sum6 = vdupq_n_s32(0);
-                    _sum7 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                    _sum1 = vld1q_s32(outptr + 4);
-                    _sum2 = vld1q_s32(outptr + 8);
-                    _sum3 = vld1q_s32(outptr + 12);
-                    _sum4 = vld1q_s32(outptr + 16);
-                    _sum5 = vld1q_s32(outptr + 20);
-                    _sum6 = vld1q_s32(outptr + 24);
-                    _sum7 = vld1q_s32(outptr + 28);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x8_t _pA = vld1q_s16(pA);
-                    int16x4_t _pB = vld1_s16(pB);
-                    _sum0 = vmlal_lane_s16(_sum0, vget_low_s16(_pA), _pB, 0);
-                    _sum1 = vmlal_lane_s16(_sum1, vget_high_s16(_pA), _pB, 0);
-                    _sum2 = vmlal_lane_s16(_sum2, vget_low_s16(_pA), _pB, 1);
-                    _sum3 = vmlal_lane_s16(_sum3, vget_high_s16(_pA), _pB, 1);
-                    _sum4 = vmlal_lane_s16(_sum4, vget_low_s16(_pA), _pB, 2);
-                    _sum5 = vmlal_lane_s16(_sum5, vget_high_s16(_pA), _pB, 2);
-                    _sum6 = vmlal_lane_s16(_sum6, vget_low_s16(_pA), _pB, 3);
-                    _sum7 = vmlal_lane_s16(_sum7, vget_high_s16(_pA), _pB, 3);
-                    pA += 8;
-                    pB += 4;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-                vst1q_s32(outptr + 16, _sum4);
-                vst1q_s32(outptr + 20, _sum5);
-                vst1q_s32(outptr + 24, _sum6);
-                vst1q_s32(outptr + 28, _sum7);
-                outptr += 32;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "cmp    %w7, #0                     \n"
-                    "beq    0f                          \n"
-
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0] \n"
-                    "b      1f                          \n"
-
-                    "0:                                 \n"
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-                    "eor    v29.16b, v29.16b, v29.16b   \n"
-                    "eor    v30.16b, v30.16b, v30.16b   \n"
-                    "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                    "1:                                 \n"
-                    "lsr    w4, %w6, #3                 \n" // w4 = max_kk >> 3
-                    "cmp    w4, #0                      \n"
-                    "beq    3f                          \n"
-
-                    "ld1    {v4.8h, v5.8h}, [%1], #32   \n"
-                    "ld1    {v0.8h}, [%2], #16          \n"
-                    ".align 4                           \n"
-                    "2:                                 \n"
-                    "smlal  v28.4s, v4.4h, v0.h[0]      \n"
-                    "smlal  v30.4s, v4.4h, v0.h[1]      \n"
-                    "ld1    {v6.8h, v7.8h}, [%1], #32   \n"
-                    "smlal2 v29.4s, v4.8h, v0.h[0]      \n"
-                    "smlal2 v31.4s, v4.8h, v0.h[1]      \n"
-                    "ld1    {v1.8h}, [%2], #16          \n"
-                    "smlal  v28.4s, v5.4h, v0.h[2]      \n"
-                    "smlal  v30.4s, v5.4h, v0.h[3]      \n"
-                    "smlal2 v29.4s, v5.8h, v0.h[2]      \n"
-                    "smlal2 v31.4s, v5.8h, v0.h[3]      \n"
-                    "smlal  v28.4s, v6.4h, v0.h[4]      \n"
-                    "smlal  v30.4s, v6.4h, v0.h[5]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.8h, v5.8h}, [%1], #32   \n"
-                    "smlal2 v29.4s, v6.8h, v0.h[4]      \n"
-                    "smlal2 v31.4s, v6.8h, v0.h[5]      \n"
-                    "smlal  v28.4s, v7.4h, v0.h[6]      \n"
-                    "smlal  v30.4s, v7.4h, v0.h[7]      \n"
-                    "smlal2 v29.4s, v7.8h, v0.h[6]      \n"
-                    "smlal2 v31.4s, v7.8h, v0.h[7]      \n"
-                    "smlal  v28.4s, v4.4h, v1.h[0]      \n"
-                    "smlal  v30.4s, v4.4h, v1.h[1]      \n"
-                    "ld1    {v6.8h, v7.8h}, [%1], #32   \n"
-                    "smlal2 v29.4s, v4.8h, v1.h[0]      \n"
-                    "smlal2 v31.4s, v4.8h, v1.h[1]      \n"
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v0.8h}, [%2], #16          \n"
-                    "smlal  v28.4s, v5.4h, v1.h[2]      \n"
-                    "smlal  v30.4s, v5.4h, v1.h[3]      \n"
-                    "smlal2 v29.4s, v5.8h, v1.h[2]      \n"
-                    "smlal2 v31.4s, v5.8h, v1.h[3]      \n"
-                    "smlal  v28.4s, v6.4h, v1.h[4]      \n"
-                    "smlal  v30.4s, v6.4h, v1.h[5]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.8h, v5.8h}, [%1], #32   \n"
-                    "smlal2 v29.4s, v6.8h, v1.h[4]      \n"
-                    "smlal2 v31.4s, v6.8h, v1.h[5]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "smlal  v28.4s, v7.4h, v1.h[6]      \n"
-                    "smlal  v30.4s, v7.4h, v1.h[7]      \n"
-                    "smlal2 v29.4s, v7.8h, v1.h[6]      \n"
-                    "smlal2 v31.4s, v7.8h, v1.h[7]      \n"
-                    "bne    2b                          \n"
-                    "sub    %1, %1, #32                 \n"
-                    "sub    %2, %2, #16                 \n"
-
-                    "3:                                 \n"
-                    "and    w4, %w6, #7                 \n" // w4 = remain = max_kk & 7
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "ld1    {v4.8h}, [%1], #16          \n"
-                    "ld1    {v0.4h}, [%2]               \n"
-                    "add    %2, %2, #4                  \n"
-                    "smlal  v28.4s, v4.4h, v0.h[0]      \n"
-                    "smlal  v30.4s, v4.4h, v0.h[1]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "smlal2 v29.4s, v4.8h, v0.h[0]      \n"
-                    "smlal2 v31.4s, v4.8h, v0.h[1]      \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%1, #512]          \n"
-                    "pld        [%2, #128]          \n"
-                    "cmp        %7, #0              \n"
-                    "beq        0f                  \n"
-
-                    "vldm       %0, {d24-d31}       \n"
-                    "b          1f                  \n"
-
-                    "0:                             \n"
-                    "veor       q12, q12            \n"
-                    "veor       q13, q13            \n"
-                    "veor       q14, q14            \n"
-                    "veor       q15, q15            \n"
-
-                    "1:                             \n"
-                    "lsr        r4, %6, #2          \n" // r4 = max_kk >> 2
-                    "cmp        r4, #0              \n"
-                    "beq        3f                  \n"
-
-                    "vld1.s16   {d2-d5}, [%1]!      \n"
-                    "vld1.s16   {d0}, [%2]!         \n"
-                    ".align 4                       \n"
-                    "2:                             \n"
-                    "vmlal.s16  q12, d2, d0[0]      \n"
-                    "vld1.s16   {d6-d9}, [%1]!      \n"
-                    "vmlal.s16  q14, d2, d0[1]      \n"
-                    "vld1.s16   {d1}, [%2]!         \n"
-                    "vmlal.s16  q13, d3, d0[0]      \n"
-                    "vmlal.s16  q15, d3, d0[1]      \n"
-                    "vmlal.s16  q12, d4, d0[2]      \n"
-                    "vmlal.s16  q14, d4, d0[3]      \n"
-                    "vmlal.s16  q13, d5, d0[2]      \n"
-                    "vmlal.s16  q15, d5, d0[3]      \n"
-                    "vmlal.s16  q12, d6, d1[0]      \n"
-                    "pld        [%1, #512]          \n"
-                    "vld1.s16   {d2-d5}, [%1]!      \n"
-                    "vmlal.s16  q14, d6, d1[1]      \n"
-                    "pld        [%2, #128]          \n"
-                    "vld1.s16   {d0}, [%2]!         \n"
-                    "vmlal.s16  q13, d7, d1[0]      \n"
-                    "vmlal.s16  q15, d7, d1[1]      \n"
-                    "vmlal.s16  q12, d8, d1[2]      \n"
-                    "vmlal.s16  q14, d8, d1[3]      \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmlal.s16  q13, d9, d1[2]      \n"
-                    "vmlal.s16  q15, d9, d1[3]      \n"
-                    "bne        2b                  \n"
-                    "sub        %1, %1, #32         \n"
-                    "sub        %2, %2, #8          \n"
-
-                    "3:                             \n"
-                    "and        r4, %6, #3          \n" // w4 = remain = max_kk & 3
-                    "cmp        r4, #0              \n"
-                    "beq        5f                  \n"
-
-                    "4:                             \n"
-                    "vld1.s16   {d0-d1}, [%1]!      \n"
-                    "vld1.s16   {d2}, [%2]          \n"
-                    "add        %2, %2, #4          \n"
-                    "vmlal.s16  q12, d0, d2[0]      \n"
-                    "vmlal.s16  q14, d0, d2[1]      \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmlal.s16  q13, d1, d2[0]      \n"
-                    "vmlal.s16  q15, d1, d2[1]      \n"
-                    "bne        4b                  \n"
-
-                    "5:                             \n"
-                    "vstm       %0!, {d24-d31}      \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-                int32x4_t _sum2;
-                int32x4_t _sum3;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                    _sum2 = vdupq_n_s32(0);
-                    _sum3 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                    _sum1 = vld1q_s32(outptr + 4);
-                    _sum2 = vld1q_s32(outptr + 8);
-                    _sum3 = vld1q_s32(outptr + 12);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x8_t _pA = vld1q_s16(pA);
-                    int16x4_t _pB0 = vdup_n_s16(pB[0]);
-                    int16x4_t _pB1 = vdup_n_s16(pB[1]);
-                    _sum0 = vmlal_s16(_sum0, vget_low_s16(_pA), _pB0);
-                    _sum1 = vmlal_s16(_sum1, vget_high_s16(_pA), _pB0);
-                    _sum2 = vmlal_s16(_sum2, vget_low_s16(_pA), _pB1);
-                    _sum3 = vmlal_s16(_sum3, vget_high_s16(_pA), _pB1);
-                    pA += 8;
-                    pB += 2;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-                outptr += 16;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "cmp    %w7, #0                     \n"
-                    "beq    0f                          \n"
-
-                    "ld1    {v30.4s, v31.4s}, [%0]      \n"
-                    "b      1f                          \n"
-
-                    "0:                                 \n"
-                    "eor    v30.16b, v30.16b, v30.16b   \n"
-                    "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                    "1:                                 \n"
-                    "lsr    w4, %w6, #3                 \n" // w4 = max_kk >> 3
-                    "cmp    w4, #0                      \n"
-                    "beq    3f                          \n"
-
-                    "ld1    {v4.8h, v5.8h}, [%1], #32   \n"
-                    "ld1    {v1.8h}, [%2], #16          \n"
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-                    "eor    v29.16b, v29.16b, v29.16b   \n"
-                    ".align 4                           \n"
-                    "2:                                 \n"
-                    "mov    v0.16b, v1.16b              \n"
-                    "smlal  v28.4s, v4.4h, v0.h[0]      \n"
-                    "ld1    {v6.8h, v7.8h}, [%1], #32   \n"
-                    "smlal2 v29.4s, v4.8h, v0.h[0]      \n"
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "ld1    {v1.8h}, [%2], #16          \n"
-                    "smlal  v30.4s, v5.4h, v0.h[1]      \n"
-                    "smlal2 v31.4s, v5.8h, v0.h[1]      \n"
-                    "smlal  v28.4s, v6.4h, v0.h[2]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.8h, v5.8h}, [%1], #32   \n"
-                    "smlal2 v29.4s, v6.8h, v0.h[2]      \n"
-                    "smlal  v30.4s, v7.4h, v0.h[3]      \n"
-                    "smlal2 v31.4s, v7.8h, v0.h[3]      \n"
-                    "smlal  v28.4s, v4.4h, v0.h[4]      \n"
-                    "ld1    {v6.8h, v7.8h}, [%1], #32   \n"
-                    "smlal2 v29.4s, v4.8h, v0.h[4]      \n"
-                    "smlal  v30.4s, v5.4h, v0.h[5]      \n"
-                    "smlal2 v31.4s, v5.8h, v0.h[5]      \n"
-                    "smlal  v28.4s, v6.4h, v0.h[6]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.8h, v5.8h}, [%1], #32   \n"
-                    "smlal2 v29.4s, v6.8h, v0.h[6]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "smlal  v30.4s, v7.4h, v0.h[7]      \n"
-                    "smlal2 v31.4s, v7.8h, v0.h[7]      \n"
-                    "bne    2b                          \n"
-                    "sub    %1, %1, #32                 \n"
-                    "sub    %2, %2, #16                 \n"
-                    "add    v30.4s, v30.4s, v28.4s      \n"
-                    "add    v31.4s, v31.4s, v29.4s      \n"
-
-                    "3:                                 \n"
-                    "and    w4, %w6, #7                 \n" // w4 = remain = max_kk & 7
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "ld1    {v4.8h}, [%1], #16          \n"
-                    "ld1r   {v0.4h}, [%2], #2           \n"
-                    "subs   w4, w4, #1                  \n"
-                    "smlal  v30.4s, v4.4h, v0.h[0]      \n"
-                    "smlal2 v31.4s, v4.8h, v0.h[0]      \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "st1    {v30.4s, v31.4s}, [%0], #32 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%1, #512]          \n"
-                    "pld        [%2, #64]           \n"
-                    "cmp        %7, #0              \n"
-                    "beq        0f                  \n"
-
-                    "vld1.s32   {d28-d31}, [%0]     \n"
-                    "b          1f                  \n"
-
-                    "0:                             \n"
-                    "veor       q14, q14            \n"
-                    "veor       q15, q15            \n"
-
-                    "1:                             \n"
-                    "lsr        r4, %6, #2          \n" // r4 = max_kk >> 2
-                    "cmp        r4, #0              \n"
-                    "beq        3f                  \n"
-
-                    "vld1.s16   {d2-d5}, [%1]!      \n"
-                    ".align 4                       \n"
-                    "2:                             \n"
-                    "pld        [%2, #64]           \n"
-                    "vld1.s16   {d0}, [%2]!         \n"
-                    "vmlal.s16  q14, d2, d0[0]      \n"
-                    "vld1.s16   {d6-d9}, [%1]!      \n"
-                    "vmlal.s16  q15, d3, d0[0]      \n"
-                    "vmlal.s16  q14, d4, d0[1]      \n"
-                    "vmlal.s16  q15, d5, d0[1]      \n"
-                    "vmlal.s16  q14, d6, d0[2]      \n"
-                    "pld        [%1, #512]          \n"
-                    "vld1.s16   {d2-d5}, [%1]!      \n"
-                    "vmlal.s16  q15, d7, d0[2]      \n"
-                    "vmlal.s16  q14, d8, d0[3]      \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmlal.s16  q15, d9, d0[3]      \n"
-                    "bne        2b                  \n"
-                    "sub        %1, %1, #32         \n"
-
-                    "3:                             \n"
-                    "and        r4, %6, #3          \n" // w4 = remain = max_kk & 3
-                    "cmp        r4, #0              \n"
-                    "beq        5f                  \n"
-
-                    "4:                             \n"
-                    "vld1.s16   {d0-d1}, [%1]!      \n"
-                    "vld1.s16   {d2[]}, [%2]!       \n"
-                    "subs       r4, r4, #1          \n"
-                    "vmlal.s16  q14, d0, d2[0]      \n"
-                    "vmlal.s16  q15, d1, d2[0]      \n"
-                    "bne        4b                  \n"
-
-                    "5:                             \n"
-                    "vst1.s32   {d28-d31}, [%0]!    \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB)      // %2
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "r"(max_kk), // %6
-                    "r"(k)       // %7
-                    : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                    _sum1 = vld1q_s32(outptr + 4);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x8_t _pA = vld1q_s16(pA);
-                    int16x4_t _pB = vld1_dup_s16(pB);
-                    _sum0 = vmlal_s16(_sum0, vget_low_s16(_pA), _pB);
-                    _sum1 = vmlal_s16(_sum1, vget_high_s16(_pA), _pB);
-                    pA += 8;
-                    pB += 1;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                outptr += 8;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-        }
-    }
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const short* pAT = AT_tile.row<const short>(b) + max_kk * ii;
-            const short* pB = BT_tile.row<const short>(b);
-
-            int jj = 0;
-#if __aarch64__
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const short* pA = pAT;
-
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-                int32x4_t _sum2;
-                int32x4_t _sum3;
-                int32x4_t _sum4;
-                int32x4_t _sum5;
-                int32x4_t _sum6;
-                int32x4_t _sum7;
-                int32x4_t _sum8;
-                int32x4_t _sum9;
-                int32x4_t _suma;
-                int32x4_t _sumb;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                    _sum2 = vdupq_n_s32(0);
-                    _sum3 = vdupq_n_s32(0);
-                    _sum4 = vdupq_n_s32(0);
-                    _sum5 = vdupq_n_s32(0);
-                    _sum6 = vdupq_n_s32(0);
-                    _sum7 = vdupq_n_s32(0);
-                    _sum8 = vdupq_n_s32(0);
-                    _sum9 = vdupq_n_s32(0);
-                    _suma = vdupq_n_s32(0);
-                    _sumb = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                    _sum1 = vld1q_s32(outptr + 4);
-                    _sum2 = vld1q_s32(outptr + 8);
-                    _sum3 = vld1q_s32(outptr + 12);
-                    _sum4 = vld1q_s32(outptr + 16);
-                    _sum5 = vld1q_s32(outptr + 20);
-                    _sum6 = vld1q_s32(outptr + 24);
-                    _sum7 = vld1q_s32(outptr + 28);
-                    _sum8 = vld1q_s32(outptr + 32);
-                    _sum9 = vld1q_s32(outptr + 36);
-                    _suma = vld1q_s32(outptr + 40);
-                    _sumb = vld1q_s32(outptr + 44);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x4_t _pA = vld1_s16(pA);
-                    int16x8_t _pB = vld1q_s16(pB);
-                    int16x4_t _pB2 = vld1_s16(pB + 8);
-                    _sum0 = vmlal_laneq_s16(_sum0, _pA, _pB, 0);
-                    _sum1 = vmlal_laneq_s16(_sum1, _pA, _pB, 1);
-                    _sum2 = vmlal_laneq_s16(_sum2, _pA, _pB, 2);
-                    _sum3 = vmlal_laneq_s16(_sum3, _pA, _pB, 3);
-                    _sum4 = vmlal_laneq_s16(_sum4, _pA, _pB, 4);
-                    _sum5 = vmlal_laneq_s16(_sum5, _pA, _pB, 5);
-                    _sum6 = vmlal_laneq_s16(_sum6, _pA, _pB, 6);
-                    _sum7 = vmlal_laneq_s16(_sum7, _pA, _pB, 7);
-                    _sum8 = vmlal_lane_s16(_sum8, _pA, _pB2, 0);
-                    _sum9 = vmlal_lane_s16(_sum9, _pA, _pB2, 1);
-                    _suma = vmlal_lane_s16(_suma, _pA, _pB2, 2);
-                    _sumb = vmlal_lane_s16(_sumb, _pA, _pB2, 3);
-                    pA += 4;
-                    pB += 12;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-                vst1q_s32(outptr + 16, _sum4);
-                vst1q_s32(outptr + 20, _sum5);
-                vst1q_s32(outptr + 24, _sum6);
-                vst1q_s32(outptr + 28, _sum7);
-                vst1q_s32(outptr + 32, _sum8);
-                vst1q_s32(outptr + 36, _sum9);
-                vst1q_s32(outptr + 40, _suma);
-                vst1q_s32(outptr + 44, _sumb);
-                outptr += 48;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const short* pA = pAT;
-
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-                int32x4_t _sum2;
-                int32x4_t _sum3;
-                int32x4_t _sum4;
-                int32x4_t _sum5;
-                int32x4_t _sum6;
-                int32x4_t _sum7;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                    _sum2 = vdupq_n_s32(0);
-                    _sum3 = vdupq_n_s32(0);
-                    _sum4 = vdupq_n_s32(0);
-                    _sum5 = vdupq_n_s32(0);
-                    _sum6 = vdupq_n_s32(0);
-                    _sum7 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                    _sum1 = vld1q_s32(outptr + 4);
-                    _sum2 = vld1q_s32(outptr + 8);
-                    _sum3 = vld1q_s32(outptr + 12);
-                    _sum4 = vld1q_s32(outptr + 16);
-                    _sum5 = vld1q_s32(outptr + 20);
-                    _sum6 = vld1q_s32(outptr + 24);
-                    _sum7 = vld1q_s32(outptr + 28);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x4_t _pA = vld1_s16(pA);
-                    int16x8_t _pB = vld1q_s16(pB);
-                    _sum0 = vmlal_laneq_s16(_sum0, _pA, _pB, 0);
-                    _sum1 = vmlal_laneq_s16(_sum1, _pA, _pB, 1);
-                    _sum2 = vmlal_laneq_s16(_sum2, _pA, _pB, 2);
-                    _sum3 = vmlal_laneq_s16(_sum3, _pA, _pB, 3);
-                    _sum4 = vmlal_laneq_s16(_sum4, _pA, _pB, 4);
-                    _sum5 = vmlal_laneq_s16(_sum5, _pA, _pB, 5);
-                    _sum6 = vmlal_laneq_s16(_sum6, _pA, _pB, 6);
-                    _sum7 = vmlal_laneq_s16(_sum7, _pA, _pB, 7);
-                    pA += 4;
-                    pB += 8;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-                vst1q_s32(outptr + 16, _sum4);
-                vst1q_s32(outptr + 20, _sum5);
-                vst1q_s32(outptr + 24, _sum6);
-                vst1q_s32(outptr + 28, _sum7);
-                outptr += 32;
-            }
-#endif // __aarch64__
-            for (; jj + 5 < max_jj; jj += 6)
-            {
-                const short* pA = pAT;
-
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-                int32x4_t _sum2;
-                int32x4_t _sum3;
-                int32x4_t _sum4;
-                int32x4_t _sum5;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                    _sum2 = vdupq_n_s32(0);
-                    _sum3 = vdupq_n_s32(0);
-                    _sum4 = vdupq_n_s32(0);
-                    _sum5 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                    _sum1 = vld1q_s32(outptr + 4);
-                    _sum2 = vld1q_s32(outptr + 8);
-                    _sum3 = vld1q_s32(outptr + 12);
-                    _sum4 = vld1q_s32(outptr + 16);
-                    _sum5 = vld1q_s32(outptr + 20);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x4_t _pA = vld1_s16(pA);
-                    int16x8_t _pB = vld1q_s16(pB);
-                    _sum0 = vmlal_lane_s16(_sum0, _pA, vget_low_s16(_pB), 0);
-                    _sum1 = vmlal_lane_s16(_sum1, _pA, vget_low_s16(_pB), 1);
-                    _sum2 = vmlal_lane_s16(_sum2, _pA, vget_low_s16(_pB), 2);
-                    _sum3 = vmlal_lane_s16(_sum3, _pA, vget_low_s16(_pB), 3);
-                    _sum4 = vmlal_lane_s16(_sum4, _pA, vget_high_s16(_pB), 0);
-                    _sum5 = vmlal_lane_s16(_sum5, _pA, vget_high_s16(_pB), 1);
-                    pA += 4;
-                    pB += 6;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-                vst1q_s32(outptr + 16, _sum4);
-                vst1q_s32(outptr + 20, _sum5);
-                outptr += 24;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const short* pA = pAT;
-
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-                int32x4_t _sum2;
-                int32x4_t _sum3;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                    _sum2 = vdupq_n_s32(0);
-                    _sum3 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                    _sum1 = vld1q_s32(outptr + 4);
-                    _sum2 = vld1q_s32(outptr + 8);
-                    _sum3 = vld1q_s32(outptr + 12);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x4_t _pA = vld1_s16(pA);
-                    int16x4_t _pB = vld1_s16(pB);
-                    _sum0 = vmlal_lane_s16(_sum0, _pA, _pB, 0);
-                    _sum1 = vmlal_lane_s16(_sum1, _pA, _pB, 1);
-                    _sum2 = vmlal_lane_s16(_sum2, _pA, _pB, 2);
-                    _sum3 = vmlal_lane_s16(_sum3, _pA, _pB, 3);
-                    pA += 4;
-                    pB += 4;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-                outptr += 16;
-            }
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const short* pA = pAT;
-
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                    _sum1 = vld1q_s32(outptr + 4);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x4_t _pA = vld1_s16(pA);
-                    int16x4_t _pB0 = vdup_n_s16(pB[0]);
-                    int16x4_t _pB1 = vdup_n_s16(pB[1]);
-                    _sum0 = vmlal_s16(_sum0, _pA, _pB0);
-                    _sum1 = vmlal_s16(_sum1, _pA, _pB1);
-                    pA += 4;
-                    pB += 2;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                outptr += 8;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const short* pA = pAT;
-
-                int32x4_t _sum0;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x4_t _pA = vld1_s16(pA);
-                    int16x4_t _pB = vld1_dup_s16(pB);
-                    _sum0 = vmlal_s16(_sum0, _pA, _pB);
-                    pA += 4;
-                    pB += 1;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                outptr += 4;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const short* pAT = AT_tile.row<const short>(b) + max_kk * ii;
-            const short* pB = BT_tile.row<const short>(b);
-
-            int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const short* pA = pAT;
-
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-                int32x4_t _sum2;
-                int32x4_t _sum3;
-                int32x4_t _sum4;
-                int32x4_t _sum5;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                    _sum2 = vdupq_n_s32(0);
-                    _sum3 = vdupq_n_s32(0);
-                    _sum4 = vdupq_n_s32(0);
-                    _sum5 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    int32x4x2_t _s01 = vld2q_s32(outptr);
-                    int32x4x2_t _s23 = vld2q_s32(outptr + 8);
-                    int32x4x2_t _s45 = vld2q_s32(outptr + 16);
-                    _sum0 = _s01.val[0];
-                    _sum3 = _s01.val[1];
-                    _sum1 = _s23.val[0];
-                    _sum4 = _s23.val[1];
-                    _sum2 = _s45.val[0];
-                    _sum5 = _s45.val[1];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x4_t _pA0 = vdup_n_s16(pA[0]);
-                    int16x4_t _pA1 = vdup_n_s16(pA[1]);
-                    int16x8_t _pB = vld1q_s16(pB);
-                    int16x4_t _pB2 = vld1_s16(pB + 8);
-                    _sum0 = vmlal_s16(_sum0, _pA0, vget_low_s16(_pB));
-                    _sum1 = vmlal_s16(_sum1, _pA0, vget_high_s16(_pB));
-                    _sum2 = vmlal_s16(_sum2, _pA0, _pB2);
-                    _sum3 = vmlal_s16(_sum3, _pA1, vget_low_s16(_pB));
-                    _sum4 = vmlal_s16(_sum4, _pA1, vget_high_s16(_pB));
-                    _sum5 = vmlal_s16(_sum5, _pA1, _pB2);
-                    pA += 2;
-                    pB += 12;
-                }
-
-                int32x4x2_t _s01;
-                _s01.val[0] = _sum0;
-                _s01.val[1] = _sum3;
-                int32x4x2_t _s23;
-                _s23.val[0] = _sum1;
-                _s23.val[1] = _sum4;
-                int32x4x2_t _s45;
-                _s45.val[0] = _sum2;
-                _s45.val[1] = _sum5;
-                vst2q_s32(outptr, _s01);
-                vst2q_s32(outptr + 8, _s23);
-                vst2q_s32(outptr + 16, _s45);
-                outptr += 24;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const short* pA = pAT;
-
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-                int32x4_t _sum2;
-                int32x4_t _sum3;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                    _sum2 = vdupq_n_s32(0);
-                    _sum3 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    int32x4x2_t _s01 = vld2q_s32(outptr);
-                    int32x4x2_t _s23 = vld2q_s32(outptr + 8);
-                    _sum0 = _s01.val[0];
-                    _sum2 = _s01.val[1];
-                    _sum1 = _s23.val[0];
-                    _sum3 = _s23.val[1];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x4_t _pA0 = vdup_n_s16(pA[0]);
-                    int16x4_t _pA1 = vdup_n_s16(pA[1]);
-                    int16x8_t _pB = vld1q_s16(pB);
-                    _sum0 = vmlal_s16(_sum0, _pA0, vget_low_s16(_pB));
-                    _sum1 = vmlal_s16(_sum1, _pA0, vget_high_s16(_pB));
-                    _sum2 = vmlal_s16(_sum2, _pA1, vget_low_s16(_pB));
-                    _sum3 = vmlal_s16(_sum3, _pA1, vget_high_s16(_pB));
-                    pA += 2;
-                    pB += 8;
-                }
-
-                int32x4x2_t _s01;
-                _s01.val[0] = _sum0;
-                _s01.val[1] = _sum2;
-                int32x4x2_t _s23;
-                _s23.val[0] = _sum1;
-                _s23.val[1] = _sum3;
-                vst2q_s32(outptr, _s01);
-                vst2q_s32(outptr + 8, _s23);
-                outptr += 16;
-            }
-#endif // __aarch64__
-            for (; jj + 5 < max_jj; jj += 6)
-            {
-                const short* pA = pAT;
-
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-                int32x4_t _sum2;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                    _sum2 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    int32x4x2_t _s01 = vld2q_s32(outptr);
-                    _sum0 = _s01.val[0];
-                    _sum1 = _s01.val[1];
-                    _sum2 = vld1q_s32(outptr + 8);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x4_t _pA = vreinterpret_s16_s32(vld1_dup_s32((const int*)pA));
-                    int16x8_t _pB = vld1q_s16(pB);
-                    int16x4_t _pB2 = vzip_s16(vget_high_s16(_pB), vget_high_s16(_pB)).val[0];
-                    _sum0 = vmlal_lane_s16(_sum0, vget_low_s16(_pB), _pA, 0);
-                    _sum1 = vmlal_lane_s16(_sum1, vget_low_s16(_pB), _pA, 1);
-                    _sum2 = vmlal_s16(_sum2, _pA, _pB2);
-                    pA += 2;
-                    pB += 6;
-                }
-
-                int32x4x2_t _s01;
-                _s01.val[0] = _sum0;
-                _s01.val[1] = _sum1;
-                vst2q_s32(outptr, _s01);
-                vst1q_s32(outptr + 8, _sum2);
-                outptr += 12;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const short* pA = pAT;
-
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    int32x4x2_t _s01 = vld2q_s32(outptr);
-                    _sum0 = _s01.val[0];
-                    _sum1 = _s01.val[1];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x4_t _pA0 = vdup_n_s16(pA[0]);
-                    int16x4_t _pA1 = vdup_n_s16(pA[1]);
-                    int16x4_t _pB = vld1_s16(pB);
-                    _sum0 = vmlal_s16(_sum0, _pA0, _pB);
-                    _sum1 = vmlal_s16(_sum1, _pA1, _pB);
-                    pA += 2;
-                    pB += 4;
-                }
-
-                int32x4x2_t _s01;
-                _s01.val[0] = _sum0;
-                _s01.val[1] = _sum1;
-                vst2q_s32(outptr, _s01);
-                outptr += 8;
-            }
-#endif // __ARM_NEON
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const short* pA = pAT;
-
-                int sum00 = 0;
-                int sum01 = 0;
-                int sum10 = 0;
-                int sum11 = 0;
-
-                if (k == 0)
-                {
-                    sum00 = 0;
-                    sum01 = 0;
-                    sum10 = 0;
-                    sum11 = 0;
-                }
-                else
-                {
-                    sum00 = outptr[0];
-                    sum01 = outptr[1];
-                    sum10 = outptr[2];
-                    sum11 = outptr[3];
-                }
-
-                int kk = 0;
-#if !__ARM_NEON && __ARM_FEATURE_SIMD32 && NCNN_GNU_INLINE_ASM
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    // fomit-frame-pointer implied in optimized flag spare one register
-                    // let us stay away from error: ‘asm’ operand has impossible constraints   --- nihui
-#if __OPTIMIZE__
-                    asm volatile(
-                        "ldr    r2, [%0], #4    \n" // int16x2_t _pA0 = *((int16x2_t*)pA); pA += 2;
-                        "ldr    r3, [%0], #4    \n" // int16x2_t _pA1 = *((int16x2_t*)pA); pA += 2;
-                        "ldr    r4, [%1], #4    \n" // int16x2_t _pB0 = *((int16x2_t*)pB); pB += 2;
-                        "ldr    r5, [%1], #4    \n" // int16x2_t _pB1 = *((int16x2_t*)pB); pB += 2;
-                        "smlad  %2, r2, r4, %2  \n" // sum00 = __smlad(_pA0, _pB0, sum00);
-                        "smlad  %3, r3, r4, %3  \n" // sum01 = __smlad(_pA1, _pB0, sum01);
-                        "smlad  %4, r2, r5, %4  \n" // sum10 = __smlad(_pA0, _pB1, sum10);
-                        "smlad  %5, r3, r5, %5  \n" // sum11 = __smlad(_pA1, _pB1, sum11);
-                        : "=r"(pA),
-                        "=r"(pB),
-                        "=r"(sum00),
-                        "=r"(sum01),
-                        "=r"(sum10),
-                        "=r"(sum11)
-                        : "0"(pA),
-                        "1"(pB),
-                        "2"(sum00),
-                        "3"(sum01),
-                        "4"(sum10),
-                        "5"(sum11)
-                        : "memory", "r2", "r3", "r4", "r5");
-#else
-                    int _pA0 = *((int*)pA);
-                    int _pA1 = *((int*)(pA + 2));
-                    int _pB0 = *((int*)pB);
-                    int _pB1 = *((int*)(pB + 2));
-                    asm volatile("smlad %0, %2, %3, %0"
-                                 : "=r"(sum00)
-                                 : "0"(sum00), "r"(_pA0), "r"(_pB0)
-                                 :);
-                    asm volatile("smlad %0, %2, %3, %0"
-                                 : "=r"(sum01)
-                                 : "0"(sum01), "r"(_pA1), "r"(_pB0)
-                                 :);
-                    asm volatile("smlad %0, %2, %3, %0"
-                                 : "=r"(sum10)
-                                 : "0"(sum10), "r"(_pA0), "r"(_pB1)
-                                 :);
-                    asm volatile("smlad %0, %2, %3, %0"
-                                 : "=r"(sum11)
-                                 : "0"(sum11), "r"(_pA1), "r"(_pB1)
-                                 :);
-                    pA += 4;
-                    pB += 4;
-#endif
-                }
-#endif // !__ARM_NEON && __ARM_FEATURE_SIMD32 && NCNN_GNU_INLINE_ASM
-                for (; kk < max_kk; kk++)
-                {
-                    sum00 += pA[0] * pB[0];
-                    sum01 += pA[1] * pB[0];
-                    sum10 += pA[0] * pB[1];
-                    sum11 += pA[1] * pB[1];
-                    pA += 2;
-                    pB += 2;
-                }
-
-                outptr[0] = sum00;
-                outptr[1] = sum01;
-                outptr[2] = sum10;
-                outptr[3] = sum11;
-                outptr += 2 * 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const short* pA = pAT;
-
-                int sum0 = 0;
-                int sum1 = 0;
-
-                if (k == 0)
-                {
-                    sum0 = 0;
-                    sum1 = 0;
-                }
-                else
-                {
-                    sum0 = outptr[0];
-                    sum1 = outptr[1];
-                }
-
-                int kk = 0;
-#if !__ARM_NEON && __ARM_FEATURE_SIMD32 && NCNN_GNU_INLINE_ASM
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    asm volatile(
-                        "ldr    r2, [%0], #4    \n" // int16x2_t _pA0 = *((int16x2_t*)pA); pA += 2;
-                        "ldr    r3, [%0], #4    \n" // int16x2_t _pA1 = *((int16x2_t*)pA); pA += 2;
-                        "ldr    r4, [%1], #4    \n" // int16x2_t _pB = *((int16x2_t*)pB); pB += 2;
-                        "smlad  %2, r2, r4, %2  \n" // sum0 = __smlad(_pA0, _pB, sum0);
-                        "smlad  %3, r3, r4, %3  \n" // sum1 = __smlad(_pA1, _pB, sum1);
-                        : "=r"(pA),
-                        "=r"(pB),
-                        "=r"(sum0),
-                        "=r"(sum1)
-                        : "0"(pA),
-                        "1"(pB),
-                        "2"(sum0),
-                        "3"(sum1)
-                        : "memory", "r2", "r3", "r4");
-                }
-#endif // !__ARM_NEON && __ARM_FEATURE_SIMD32 && NCNN_GNU_INLINE_ASM
-                for (; kk < max_kk; kk++)
-                {
-                    sum0 += pA[0] * pB[0];
-                    sum1 += pA[1] * pB[0];
-                    pA += 2;
-                    pB += 1;
-                }
-
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-                outptr += 2;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const short* pAT = AT_tile.row<const short>(b) + max_kk * ii;
-            const short* pB = BT_tile.row<const short>(b);
-
-            int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const short* pA = pAT;
-
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-                int32x4_t _sum2;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                    _sum2 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                    _sum1 = vld1q_s32(outptr + 4);
-                    _sum2 = vld1q_s32(outptr + 8);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x4_t _pA = vld1_dup_s16(pA);
-                    int16x8_t _pB = vld1q_s16(pB);
-                    int16x4_t _pB2 = vld1_s16(pB + 8);
-                    _sum0 = vmlal_s16(_sum0, _pA, vget_low_s16(_pB));
-                    _sum1 = vmlal_s16(_sum1, _pA, vget_high_s16(_pB));
-                    _sum2 = vmlal_s16(_sum2, _pA, _pB2);
-                    pA += 1;
-                    pB += 12;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                outptr += 12;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const short* pA = pAT;
-
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                    _sum1 = vld1q_s32(outptr + 4);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x4_t _pA = vld1_dup_s16(pA);
-                    int16x8_t _pB = vld1q_s16(pB);
-                    _sum0 = vmlal_s16(_sum0, _pA, vget_low_s16(_pB));
-                    _sum1 = vmlal_s16(_sum1, _pA, vget_high_s16(_pB));
-                    pA += 1;
-                    pB += 8;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                outptr += 8;
-            }
-#endif // __aarch64__
-            for (; jj + 5 < max_jj; jj += 6)
-            {
-                const short* pA = pAT;
-
-                int32x4_t _sum0;
-                int32x4_t _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                    _sum1 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                    _sum1 = vld1q_s32(outptr + 4);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x4_t _pA = vld1_dup_s16(pA);
-                    int16x8_t _pB = vld1q_s16(pB);
-                    _sum0 = vmlal_s16(_sum0, _pA, vget_low_s16(_pB));
-                    _sum1 = vmlal_s16(_sum1, _pA, vget_high_s16(_pB));
-                    pA += 1;
-                    pB += 6;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                vst1_s32(outptr + 4, vget_low_s32(_sum1));
-                outptr += 6;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const short* pA = pAT;
-
-                int32x4_t _sum0;
-
-                if (k == 0)
-                {
-                    _sum0 = vdupq_n_s32(0);
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    int16x4_t _pA = vld1_dup_s16(pA);
-                    int16x4_t _pB = vld1_s16(pB);
-                    _sum0 = vmlal_s16(_sum0, _pA, _pB);
-                    pA += 1;
-                    pB += 4;
-                }
-
-                vst1q_s32(outptr, _sum0);
-                outptr += 4;
-            }
-#endif // __ARM_NEON
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const short* pA = pAT;
-
-                int sum0 = 0;
-                int sum1 = 0;
-
-                if (k == 0)
-                {
-                    sum0 = 0;
-                    sum1 = 0;
-                }
-                else
-                {
-                    sum0 = outptr[0];
-                    sum1 = outptr[1];
-                }
-
-                int kk = 0;
-#if !__ARM_NEON && __ARM_FEATURE_SIMD32 && NCNN_GNU_INLINE_ASM
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    asm volatile(
-                        "ldr    r2, [%0], #4    \n" // int16x2_t _pA = *((int16x2_t*)pA); pA += 2;
-                        "ldr    r3, [%1], #4    \n" // int16x2_t _pB0 = *((int16x2_t*)pB); pB += 2;
-                        "ldr    r4, [%1], #4    \n" // int16x2_t _pB1 = *((int16x2_t*)pB); pB += 2;
-                        "smlad  %2, r2, r3, %2  \n" // sum0 = __smlad(_pA, _pB0, sum0);
-                        "smlad  %3, r2, r4, %3  \n" // sum1 = __smlad(_pA, _pB1, sum1);
-                        : "=r"(pA),
-                        "=r"(pB),
-                        "=r"(sum0),
-                        "=r"(sum1)
-                        : "0"(pA),
-                        "1"(pB),
-                        "2"(sum0),
-                        "3"(sum1)
-                        : "memory", "r2", "r3", "r4");
-                }
-#endif // !__ARM_NEON && __ARM_FEATURE_SIMD32 && NCNN_GNU_INLINE_ASM
-                for (; kk < max_kk; kk++)
-                {
-                    sum0 += pA[0] * pB[0];
-                    sum1 += pA[0] * pB[1];
-                    pA += 1;
-                    pB += 2;
-                }
-
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-                outptr += 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const short* pA = pAT;
-
-                int sum = 0;
-
-                if (k == 0)
-                {
-                    sum = 0;
-                }
-                else
-                {
-                    sum = outptr[0];
-                }
-
-                int kk = 0;
-#if !__ARM_NEON && __ARM_FEATURE_SIMD32 && NCNN_GNU_INLINE_ASM
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    asm volatile(
-                        "ldr    r2, [%0], #4    \n" // int16x2_t _pA = *((int16x2_t*)pA); pA += 2;
-                        "ldr    r3, [%1], #4    \n" // int16x2_t _pB = *((int16x2_t*)pB); pB += 2;
-                        "smlad  %2, r2, r3, %2  \n" // sum = __smlad(_pA, _pB, sum);
-                        : "=r"(pA),
-                        "=r"(pB),
-                        "=r"(sum)
-                        : "0"(pA),
-                        "1"(pB),
-                        "2"(sum)
-                        : "memory", "r2", "r3");
-                }
-#endif // !__ARM_NEON && __ARM_FEATURE_SIMD32 && NCNN_GNU_INLINE_ASM
-                for (; kk < max_kk; kk++)
-                {
-                    sum += pA[0] * pB[0];
-                    pA += 1;
-                    pB += 1;
-                }
-
-                outptr[0] = sum;
-                outptr += 1;
-            }
-        }
-    }
-}
-
-static void get_optimal_tile_mnk_int8(int M, int N, int K, int& TILE_M, int& TILE_N, int& TILE_K, int nT)
-{
-    // resolve optimal tile size from cache size
-    const int l2_cache_size_int8 = (int)(get_cpu_level2_cache_size() / sizeof(short));
-
-    if (nT == 0)
-        nT = get_physical_big_cpu_count();
-
-    // we shall take B into account for batched gemm, but that will be slower on arm in practice, why ?
-    // (void)B;
-
-    // solve K
-    {
-        // try not to split K
-#if __aarch64__
-        int tile_size = (l2_cache_size_int8 - 32) / 12;
-#elif __ARM_NEON
-        int tile_size = (l2_cache_size_int8 - 32) / 6;
-#else
-        int tile_size = (l2_cache_size_int8 - 2) / 3;
-#endif
-
-#if __aarch64__
-        TILE_K = std::max(8, tile_size / 8 * 8);
-#elif __ARM_NEON
-        TILE_K = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_K = std::max(2, tile_size / 2 * 2);
-#endif
-
-        int nn_K = (K + TILE_K - 1) / TILE_K;
-#if __aarch64__
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 7) / 8 * 8);
-#elif __ARM_NEON
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 3) / 4 * 4);
-#else
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 1) / 2 * 2);
-#endif
-    }
-
-    // solve M
-    {
-#if __ARM_NEON
-        TILE_M = 8;
-#else
-        TILE_M = 2;
-#endif
-    }
-
-    {
-        TILE_M *= std::min(nT, get_physical_cpu_count());
-
-        int nn_M = (M + TILE_M - 1) / TILE_M;
-#if __ARM_NEON
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-#else
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 1) / 2 * 2);
-#endif
-
-        if (nT > 1)
-        {
-#if __ARM_NEON
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 7) / 8 * 8);
-#else
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 1) / 2 * 2);
-#endif
-        }
-
-#if __ARM_NEON
-        TILE_M = std::max(8, TILE_M);
-#else
-        TILE_M = std::max(2, TILE_M);
-#endif
-    }
-
-    if (N > 0)
-    {
-        int tile_size;
-        if (TILE_K >= K)
-        {
-            tile_size = (l2_cache_size_int8 - TILE_M * TILE_K) / TILE_K;
-        }
-        else
-        {
-            tile_size = (l2_cache_size_int8 - TILE_M * TILE_K) / (TILE_M * 2 + TILE_K);
-        }
-
-#if __aarch64__
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#elif __ARM_NEON
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_N = std::max(1, tile_size);
-#endif
-
-        int nn_N = (N + TILE_N - 1) / TILE_N;
-
-#if __aarch64__
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#elif __ARM_NEON
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#else
-        TILE_N = std::min(TILE_N, (N + nn_N - 1) / nn_N);
-#endif
-
-#if __aarch64__
-        TILE_N = std::max(4, TILE_N);
-#elif __ARM_NEON
-        TILE_N = std::max(4, TILE_N);
-#else
-        TILE_N = std::max(1, TILE_N);
-#endif
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_kernel_tile_int8(const Mat& kernel, Mat& A, int inch, int i, int max_ii, int k, int max_kk)
-{
-    // const signed char ktm[4][3] = {
-    //     {2, 0, 0},
-    //     {1, 1, 1},
-    //     {1, -1, 1},
-    //     {0, 0, 2}
-    // };
-
-    short* ptmp = A;
-
-    int ii = 0;
-    for (; ii < max_ii; ii++)
-    {
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            short tmp[4][3];
-
-            const signed char* k0 = (const signed char*)kernel + (i + ii) * inch * 9 + (k + kk) * 9;
-
-            for (int m = 0; m < 3; m++)
-            {
-                signed char r0 = k0[0];
-                signed char r1 = k0[1];
-                signed char r2 = k0[2];
-
-                tmp[0][m] = r0 * 2;
-                tmp[1][m] = r0 + r1 + r2;
-                tmp[2][m] = r0 - r1 + r2;
-                tmp[3][m] = r2 * 2;
-
-                k0 += 3;
-            }
-
-            for (int m = 0; m < 4; m++)
-            {
-                short r0 = tmp[m][0];
-                short r1 = tmp[m][1];
-                short r2 = tmp[m][2];
-
-                short z0 = r0 * 2;
-                short z1 = r0 + r1 + r2;
-                short z2 = r0 - r1 + r2;
-                short z3 = r2 * 2;
-
-                ptmp[0] = z0;
-                ptmp[1] = z1;
-                ptmp[2] = z2;
-                ptmp[3] = z3;
-                ptmp += 4;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_kernel_int8(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt)
-{
-    const int M = outch;
-    const int K = inch;
-    const int B = 16;
-
-    int TILE_M, TILE_N, TILE_K;
-    get_optimal_tile_mnk_int8(M, 0, K, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    Mat A_tileX(B * TILE_M * TILE_K, 1, opt.num_threads, 2u, (Allocator*)0);
-
-    AT.create(TILE_K * TILE_M, B, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M, 2u, (Allocator*)0);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat A_tile = A_tileX.channel(get_omp_thread_num());
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_ii = std::min((M - i), TILE_M);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            conv3x3s1_winograd23_transform_kernel_tile_int8(kernel, A_tile, inch, i, max_ii, k, max_kk);
-
-            Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-            pack_A_tile_int8(A_tile, AT_tile, B, max_ii, max_kk);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_input_tile_int8(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    // const signed char itm[4][4] = {
-    //     {1,  0, -1,  0},
-    //     {0,  1,  1,  0},
-    //     {0, -1,  1,  0},
-    //     {0, -1,  0,  1}
-    // };
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w - 1) / 2;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-#if __ARM_NEON
-    nn_max_kk = max_kk / 8;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-        short tmp[4][4][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const signed char* r0 = bottom_blob.channel((k + kk) / elempack).row<const signed char>(ti * 2) + (tj * 2) * elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                int8x8_t _r0 = vdup_n_s8(0);
-                int8x8_t _r1 = vdup_n_s8(0);
-                int8x8_t _r2 = vdup_n_s8(0);
-                int8x8_t _r3 = vdup_n_s8(0);
-
-                if (ti * 2 + m < h)
-                {
-                    if (elempack == 8)
-                    {
-                        _r0 = vld1_s8(r0);
-                        if (tj * 2 + 1 < w) _r1 = vld1_s8(r0 + 8);
-                        if (tj * 2 + 2 < w) _r2 = vld1_s8(r0 + 16);
-                        if (tj * 2 + 3 < w) _r3 = vld1_s8(r0 + 24);
-                    }
-                    if (elempack == 1)
-                    {
-                        const signed char* r1 = r0 + N;
-                        const signed char* r2 = r0 + N * 2;
-                        const signed char* r3 = r0 + N * 3;
-                        const signed char* r4 = r0 + N * 4;
-                        const signed char* r5 = r0 + N * 5;
-                        const signed char* r6 = r0 + N * 6;
-                        const signed char* r7 = r0 + N * 7;
-
-                        int8x8_t _t0 = vld1_s8(r0);
-                        int8x8_t _t1 = vld1_s8(r1);
-                        int8x8_t _t2 = vld1_s8(r2);
-                        int8x8_t _t3 = vld1_s8(r3);
-                        int8x8_t _t4 = vld1_s8(r4);
-                        int8x8_t _t5 = vld1_s8(r5);
-                        int8x8_t _t6 = vld1_s8(r6);
-                        int8x8_t _t7 = vld1_s8(r7);
-
-                        int8x8_t _t01 = vzip_s8(_t0, _t1).val[0];
-                        int8x8_t _t23 = vzip_s8(_t2, _t3).val[0];
-                        int8x8_t _t45 = vzip_s8(_t4, _t5).val[0];
-                        int8x8_t _t67 = vzip_s8(_t6, _t7).val[0];
-                        int16x4x2_t _t0123 = vzip_s16(vreinterpret_s16_s8(_t01), vreinterpret_s16_s8(_t23));
-                        int16x4x2_t _t4567 = vzip_s16(vreinterpret_s16_s8(_t45), vreinterpret_s16_s8(_t67));
-                        int16x8_t _ta = vcombine_s16(_t0123.val[0], _t0123.val[1]);
-                        int16x8_t _tb = vcombine_s16(_t4567.val[0], _t4567.val[1]);
-                        int32x4x2_t _tab = vzipq_s32(vreinterpretq_s32_s16(_ta), vreinterpretq_s32_s16(_tb));
-
-                        _r0 = vreinterpret_s8_s32(vget_low_s32(_tab.val[0]));
-                        if (tj * 2 + 1 < w) _r1 = vreinterpret_s8_s32(vget_high_s32(_tab.val[0]));
-                        if (tj * 2 + 2 < w) _r2 = vreinterpret_s8_s32(vget_low_s32(_tab.val[1]));
-                        if (tj * 2 + 3 < w) _r3 = vreinterpret_s8_s32(vget_high_s32(_tab.val[1]));
-                    }
-                }
-
-                int16x8_t _tmp0 = vsubl_s8(_r0, _r2);
-                int16x8_t _tmp1 = vaddl_s8(_r1, _r2);
-                int16x8_t _tmp2 = vsubl_s8(_r2, _r1);
-                int16x8_t _tmp3 = vsubl_s8(_r3, _r1);
-
-                vst1q_s16(tmp[0][m], _tmp0);
-                vst1q_s16(tmp[1][m], _tmp1);
-                vst1q_s16(tmp[2][m], _tmp2);
-                vst1q_s16(tmp[3][m], _tmp3);
-
-                r0 += w * elempack;
-            }
-
-            short* p0 = (short*)B + kk * max_jj * 16 + jj * 8;
-            short* p1 = p0 + max_jj * 8;
-            short* p2 = p0 + max_jj * 8 * 2;
-            short* p3 = p0 + max_jj * 8 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                int16x8_t _r0 = vld1q_s16(tmp[m][0]);
-                int16x8_t _r1 = vld1q_s16(tmp[m][1]);
-                int16x8_t _r2 = vld1q_s16(tmp[m][2]);
-                int16x8_t _r3 = vld1q_s16(tmp[m][3]);
-
-                int16x8_t _tmp0 = vsubq_s16(_r0, _r2);
-                int16x8_t _tmp1 = vaddq_s16(_r1, _r2);
-                int16x8_t _tmp2 = vsubq_s16(_r2, _r1);
-                int16x8_t _tmp3 = vsubq_s16(_r3, _r1);
-
-                vst1q_s16(p0, _tmp0);
-                vst1q_s16(p1, _tmp1);
-                vst1q_s16(p2, _tmp2);
-                vst1q_s16(p3, _tmp3);
-
-                p0 += max_jj * 4 * 8;
-                p1 += max_jj * 4 * 8;
-                p2 += max_jj * 4 * 8;
-                p3 += max_jj * 4 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-#else // __ARM_NEON
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    #pragma omp parallel for num_threads(nT)
-#endif // __ARM_NEON
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-        short tmp[4][4][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const signed char* r0 = bottom_blob.channel(k + kk).row<const signed char>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 4; m++)
-            {
-                signed char r00 = 0;
-                signed char r01 = 0;
-                signed char r10 = 0;
-                signed char r11 = 0;
-                signed char r20 = 0;
-                signed char r21 = 0;
-                signed char r30 = 0;
-                signed char r31 = 0;
-
-                if (ti * 2 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const signed char* r1 = r0 + N;
-
-                        r00 = r0[0];
-                        r01 = r1[0];
-                        if (tj * 2 + 1 < w)
-                        {
-                            r10 = r0[1];
-                            r11 = r1[1];
-                        }
-                        if (tj * 2 + 2 < w)
-                        {
-                            r20 = r0[2];
-                            r21 = r1[2];
-                        }
-                        if (tj * 2 + 3 < w)
-                        {
-                            r30 = r0[3];
-                            r31 = r1[3];
-                        }
-                    }
-                }
-
-                tmp[0][m][0] = r00 - r20;
-                tmp[0][m][1] = r01 - r21;
-                tmp[1][m][0] = r10 + r20;
-                tmp[1][m][1] = r11 + r21;
-                tmp[2][m][0] = r20 - r10;
-                tmp[2][m][1] = r21 - r11;
-                tmp[3][m][0] = r30 - r10;
-                tmp[3][m][1] = r31 - r11;
-
-                r0 += w;
-            }
-
-            short* p0 = (short*)B + kk * max_jj * 16 + jj * 2;
-            short* p1 = p0 + max_jj * 2;
-            short* p2 = p0 + max_jj * 2 * 2;
-            short* p3 = p0 + max_jj * 2 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                short r00 = tmp[m][0][0];
-                short r01 = tmp[m][0][1];
-                short r10 = tmp[m][1][0];
-                short r11 = tmp[m][1][1];
-                short r20 = tmp[m][2][0];
-                short r21 = tmp[m][2][1];
-                short r30 = tmp[m][3][0];
-                short r31 = tmp[m][3][1];
-
-                p0[0] = r00 - r20;
-                p0[1] = r01 - r21;
-                p1[0] = r10 + r20;
-                p1[1] = r11 + r21;
-                p2[0] = r20 - r10;
-                p2[1] = r21 - r11;
-                p3[0] = r30 - r10;
-                p3[1] = r31 - r11;
-
-                p0 += max_jj * 4 * 2;
-                p1 += max_jj * 4 * 2;
-                p2 += max_jj * 4 * 2;
-                p3 += max_jj * 4 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        short tmp[4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const signed char* r0123 = bottom_blob.channel(k + kk).row<const signed char>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 4; m++)
-            {
-                signed char r0 = 0;
-                signed char r1 = 0;
-                signed char r2 = 0;
-                signed char r3 = 0;
-
-                if (ti * 2 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = r0123[0];
-                        if (tj * 2 + 1 < w) r1 = r0123[1];
-                        if (tj * 2 + 2 < w) r2 = r0123[2];
-                        if (tj * 2 + 3 < w) r3 = r0123[3];
-                    }
-                }
-
-                tmp[0][m] = r0 - r2;
-                tmp[1][m] = r1 + r2;
-                tmp[2][m] = r2 - r1;
-                tmp[3][m] = r3 - r1;
-
-                r0123 += w;
-            }
-
-            short* p0 = (short*)B + kk * max_jj * 16 + jj;
-            short* p1 = p0 + max_jj;
-            short* p2 = p0 + max_jj * 2;
-            short* p3 = p0 + max_jj * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                short r0 = tmp[m][0];
-                short r1 = tmp[m][1];
-                short r2 = tmp[m][2];
-                short r3 = tmp[m][3];
-
-                p0[0] = r0 - r2;
-                p1[0] = r1 + r2;
-                p2[0] = r2 - r1;
-                p3[0] = r3 - r1;
-
-                p0 += max_jj * 4;
-                p1 += max_jj * 4;
-                p2 += max_jj * 4;
-                p3 += max_jj * 4;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_output_tile_int8(const Mat& top_tile, Mat& top_blob, int i, int max_ii, int j, int max_jj)
-{
-    // const int otm[2][4] = {
-    //     {1,  1,  1,  0},
-    //     {0,  1, -1,  1}
-    // };
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 1) / 2;
-
-    int ii = 0;
-#if __ARM_NEON
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        int tmp[2][4][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 16 + jj * 8;
-            const int* r1 = r0 + max_jj * 8;
-            const int* r2 = r0 + max_jj * 8 * 2;
-            const int* r3 = r0 + max_jj * 8 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                int32x4_t _r00 = vld1q_s32(r0);
-                int32x4_t _r01 = vld1q_s32(r0 + 4);
-                int32x4_t _r10 = vld1q_s32(r1);
-                int32x4_t _r11 = vld1q_s32(r1 + 4);
-                int32x4_t _r20 = vld1q_s32(r2);
-                int32x4_t _r21 = vld1q_s32(r2 + 4);
-                int32x4_t _r30 = vld1q_s32(r3);
-                int32x4_t _r31 = vld1q_s32(r3 + 4);
-
-                int32x4_t _tmp00 = vaddq_s32(vaddq_s32(_r00, _r10), _r20);
-                int32x4_t _tmp01 = vaddq_s32(vaddq_s32(_r01, _r11), _r21);
-                int32x4_t _tmp10 = vaddq_s32(vsubq_s32(_r10, _r20), _r30);
-                int32x4_t _tmp11 = vaddq_s32(vsubq_s32(_r11, _r21), _r31);
-
-                vst1q_s32(tmp[0][m], _tmp00);
-                vst1q_s32(tmp[0][m] + 4, _tmp01);
-                vst1q_s32(tmp[1][m], _tmp10);
-                vst1q_s32(tmp[1][m] + 4, _tmp11);
-
-                r0 += max_jj * 4 * 8;
-                r1 += max_jj * 4 * 8;
-                r2 += max_jj * 4 * 8;
-                r3 += max_jj * 4 * 8;
-            }
-
-            int* outptr0 = top_blob.channel((i + ii) / out_elempack).row<int>(ti * 2) + (tj * 2) * out_elempack;
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                int32x4_t _r00 = vld1q_s32(tmp[m][0]);
-                int32x4_t _r01 = vld1q_s32(tmp[m][0] + 4);
-                int32x4_t _r10 = vld1q_s32(tmp[m][1]);
-                int32x4_t _r11 = vld1q_s32(tmp[m][1] + 4);
-                int32x4_t _r20 = vld1q_s32(tmp[m][2]);
-                int32x4_t _r21 = vld1q_s32(tmp[m][2] + 4);
-                int32x4_t _r30 = vld1q_s32(tmp[m][3]);
-                int32x4_t _r31 = vld1q_s32(tmp[m][3] + 4);
-
-                int32x4_t _tmp00 = vaddq_s32(vaddq_s32(_r00, _r10), _r20);
-                int32x4_t _tmp01 = vaddq_s32(vaddq_s32(_r01, _r11), _r21);
-                int32x4_t _tmp10 = vaddq_s32(vsubq_s32(_r10, _r20), _r30);
-                int32x4_t _tmp11 = vaddq_s32(vsubq_s32(_r11, _r21), _r31);
-
-                _tmp00 = vshrq_n_s32(_tmp00, 2);
-                _tmp01 = vshrq_n_s32(_tmp01, 2);
-                _tmp10 = vshrq_n_s32(_tmp10, 2);
-                _tmp11 = vshrq_n_s32(_tmp11, 2);
-
-                if (out_elempack == 8)
-                {
-                    vst1q_s32(outptr0, _tmp00);
-                    vst1q_s32(outptr0 + 4, _tmp01);
-                    if (tj * 2 + 1 < outw)
-                    {
-                        vst1q_s32(outptr0 + 8, _tmp10);
-                        vst1q_s32(outptr0 + 12, _tmp11);
-                    }
-                }
-                if (out_elempack == 4)
-                {
-                    int* outptr1 = outptr0 + N;
-
-                    vst1q_s32(outptr0, _tmp00);
-                    vst1q_s32(outptr1, _tmp01);
-                    if (tj * 2 + 1 < outw)
-                    {
-                        vst1q_s32(outptr0 + 4, _tmp10);
-                        vst1q_s32(outptr1 + 4, _tmp11);
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    int* outptr1 = outptr0 + N;
-                    int* outptr2 = outptr0 + N * 2;
-                    int* outptr3 = outptr0 + N * 3;
-                    int* outptr4 = outptr0 + N * 4;
-                    int* outptr5 = outptr0 + N * 5;
-                    int* outptr6 = outptr0 + N * 6;
-                    int* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = vgetq_lane_s32(_tmp00, 0);
-                    outptr1[0] = vgetq_lane_s32(_tmp00, 1);
-                    outptr2[0] = vgetq_lane_s32(_tmp00, 2);
-                    outptr3[0] = vgetq_lane_s32(_tmp00, 3);
-                    outptr4[0] = vgetq_lane_s32(_tmp01, 0);
-                    outptr5[0] = vgetq_lane_s32(_tmp01, 1);
-                    outptr6[0] = vgetq_lane_s32(_tmp01, 2);
-                    outptr7[0] = vgetq_lane_s32(_tmp01, 3);
-
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = vgetq_lane_s32(_tmp10, 0);
-                        outptr1[1] = vgetq_lane_s32(_tmp10, 1);
-                        outptr2[1] = vgetq_lane_s32(_tmp10, 2);
-                        outptr3[1] = vgetq_lane_s32(_tmp10, 3);
-                        outptr4[1] = vgetq_lane_s32(_tmp11, 0);
-                        outptr5[1] = vgetq_lane_s32(_tmp11, 1);
-                        outptr6[1] = vgetq_lane_s32(_tmp11, 2);
-                        outptr7[1] = vgetq_lane_s32(_tmp11, 3);
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        int tmp[2][4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 16 + jj * 4;
-            const int* r1 = r0 + max_jj * 4;
-            const int* r2 = r0 + max_jj * 4 * 2;
-            const int* r3 = r0 + max_jj * 4 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                int32x4_t _r0 = vld1q_s32(r0);
-                int32x4_t _r1 = vld1q_s32(r1);
-                int32x4_t _r2 = vld1q_s32(r2);
-                int32x4_t _r3 = vld1q_s32(r3);
-
-                int32x4_t _tmp0 = vaddq_s32(vaddq_s32(_r0, _r1), _r2);
-                int32x4_t _tmp1 = vaddq_s32(vsubq_s32(_r1, _r2), _r3);
-
-                vst1q_s32(tmp[0][m], _tmp0);
-                vst1q_s32(tmp[1][m], _tmp1);
-
-                r0 += max_jj * 4 * 4;
-                r1 += max_jj * 4 * 4;
-                r2 += max_jj * 4 * 4;
-                r3 += max_jj * 4 * 4;
-            }
-
-            int* outptr0 = top_blob.channel((i + ii) / out_elempack).row<int>(ti * 2) + (tj * 2) * out_elempack;
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                int32x4_t _r0 = vld1q_s32(tmp[m][0]);
-                int32x4_t _r1 = vld1q_s32(tmp[m][1]);
-                int32x4_t _r2 = vld1q_s32(tmp[m][2]);
-                int32x4_t _r3 = vld1q_s32(tmp[m][3]);
-
-                int32x4_t _tmp0 = vaddq_s32(vaddq_s32(_r0, _r1), _r2);
-                int32x4_t _tmp1 = vaddq_s32(vsubq_s32(_r1, _r2), _r3);
-
-                _tmp0 = vshrq_n_s32(_tmp0, 2);
-                _tmp1 = vshrq_n_s32(_tmp1, 2);
-
-                if (out_elempack == 4)
-                {
-                    vst1q_s32(outptr0, _tmp0);
-                    if (tj * 2 + 1 < outw) vst1q_s32(outptr0 + 4, _tmp1);
-                }
-                if (out_elempack == 1)
-                {
-                    int* outptr1 = outptr0 + N;
-                    int* outptr2 = outptr0 + N * 2;
-                    int* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = vgetq_lane_s32(_tmp0, 0);
-                    outptr1[0] = vgetq_lane_s32(_tmp0, 1);
-                    outptr2[0] = vgetq_lane_s32(_tmp0, 2);
-                    outptr3[0] = vgetq_lane_s32(_tmp0, 3);
-
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = vgetq_lane_s32(_tmp1, 0);
-                        outptr1[1] = vgetq_lane_s32(_tmp1, 1);
-                        outptr2[1] = vgetq_lane_s32(_tmp1, 2);
-                        outptr3[1] = vgetq_lane_s32(_tmp1, 3);
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        int tmp[2][4][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 16 + jj * 2;
-            const int* r1 = r0 + max_jj * 2;
-            const int* r2 = r0 + max_jj * 2 * 2;
-            const int* r3 = r0 + max_jj * 2 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                tmp[0][m][0] = r0[0] + r1[0] + r2[0];
-                tmp[0][m][1] = r0[1] + r1[1] + r2[1];
-                tmp[1][m][0] = r1[0] - r2[0] + r3[0];
-                tmp[1][m][1] = r1[1] - r2[1] + r3[1];
-
-                r0 += max_jj * 4 * 2;
-                r1 += max_jj * 4 * 2;
-                r2 += max_jj * 4 * 2;
-                r3 += max_jj * 4 * 2;
-            }
-
-            int* outptr0 = top_blob.channel(i + ii).row<int>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                int tmp00 = tmp[m][0][0] + tmp[m][1][0] + tmp[m][2][0];
-                int tmp01 = tmp[m][0][1] + tmp[m][1][1] + tmp[m][2][1];
-                int tmp10 = tmp[m][1][0] - tmp[m][2][0] + tmp[m][3][0];
-                int tmp11 = tmp[m][1][1] - tmp[m][2][1] + tmp[m][3][1];
-
-                tmp00 = tmp00 >> 2;
-                tmp01 = tmp01 >> 2;
-                tmp10 = tmp10 >> 2;
-                tmp11 = tmp11 >> 2;
-
-                // if (out_elempack == 1)
-                {
-                    int* outptr1 = outptr0 + N;
-
-                    outptr0[0] = tmp00;
-                    outptr1[0] = tmp01;
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp10;
-                        outptr1[1] = tmp11;
-                    }
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        int tmp[2][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 16 + jj;
-            const int* r1 = r0 + max_jj;
-            const int* r2 = r0 + max_jj * 2;
-            const int* r3 = r0 + max_jj * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                tmp[0][m] = r0[0] + r1[0] + r2[0];
-                tmp[1][m] = r1[0] - r2[0] + r3[0];
-
-                r0 += max_jj * 4;
-                r1 += max_jj * 4;
-                r2 += max_jj * 4;
-                r3 += max_jj * 4;
-            }
-
-            int* outptr0 = top_blob.channel(i + ii).row<int>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                int tmp0 = tmp[m][0] + tmp[m][1] + tmp[m][2];
-                int tmp1 = tmp[m][1] - tmp[m][2] + tmp[m][3];
-
-                tmp0 = tmp0 >> 2;
-                tmp1 = tmp1 >> 2;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = tmp0;
-                    if (tj * 2 + 1 < outw) outptr0[1] = tmp1;
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_int8(Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 2n+2, winograd F(2,3)
-    int w_tiles = (outw + 1) / 2;
-    int h_tiles = (outh + 1) / 2;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 16;
-
-    // NCNN_LOGE("conv3x3s1_winograd23_int8 %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    get_optimal_tile_mnk_int8(M, N, K, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 2u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 2u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd23_transform_input_tile_int8(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            transpose_pack_B_tile_int8(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 2u, opt.workspace_allocator);
-
-        // #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd23_transform_input_tile_int8(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            transpose_pack_B_tile_int8(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    bottom_blob.release();
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                gemm_transB_packed_tile_int8(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk);
-            }
-
-            // transform output
-            conv3x3s1_winograd23_transform_output_tile_int8(top_tile, top_blob, i, max_ii, j, max_jj);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_kernel_tile_int8(const Mat& kernel, Mat& A, int inch, int i, int max_ii, int k, int max_kk)
-{
-    // const short ktm[6][3] = {
-    //     {6, 0, 0},
-    //     {-4, -4, -4},
-    //     {-4, 4, -4},
-    //     {1, 2, 4},
-    //     {1, -2, 4},
-    //     {0, 0, 6}
-    // };
-
-    short* ptmp = A;
-
-    int ii = 0;
-    for (; ii < max_ii; ii++)
-    {
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            short tmp[6][3];
-
-            const signed char* k0 = (const signed char*)kernel + (i + ii) * inch * 9 + (k + kk) * 9;
-
-            for (int m = 0; m < 3; m++)
-            {
-                signed char r0 = k0[0];
-                signed char r1 = k0[1];
-                signed char r2 = k0[2];
-
-                tmp[0][m] = r0 * 6;
-                tmp[1][m] = -r0 * 4 - r1 * 4 - r2 * 4;
-                tmp[2][m] = -r0 * 4 + r1 * 4 - r2 * 4;
-                tmp[3][m] = r0 + r1 * 2 + r2 * 4;
-                tmp[4][m] = r0 - r1 * 2 + r2 * 4;
-                tmp[5][m] = r2 * 6;
-
-                k0 += 3;
-            }
-
-            for (int m = 0; m < 6; m++)
-            {
-                short r0 = tmp[m][0];
-                short r1 = tmp[m][1];
-                short r2 = tmp[m][2];
-
-                short z0 = r0 * 6;
-                short z1 = -r0 * 4 - r1 * 4 - r2 * 4;
-                short z2 = -r0 * 4 + r1 * 4 - r2 * 4;
-                short z3 = r0 + r1 * 2 + r2 * 4;
-                short z4 = r0 - r1 * 2 + r2 * 4;
-                short z5 = r2 * 6;
-
-                ptmp[0] = z0;
-                ptmp[1] = z1;
-                ptmp[2] = z2;
-                ptmp[3] = z3;
-                ptmp[4] = z4;
-                ptmp[5] = z5;
-                ptmp += 6;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_kernel_int8(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt)
-{
-    const int M = outch;
-    const int K = inch;
-    const int B = 36;
-
-    int TILE_M, TILE_N, TILE_K;
-    get_optimal_tile_mnk_int8(M, 0, K, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    Mat A_tileX(B * TILE_M * TILE_K, 1, opt.num_threads, 2u, (Allocator*)0);
-
-    AT.create(TILE_K * TILE_M, B, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M, 2u, (Allocator*)0);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat A_tile = A_tileX.channel(get_omp_thread_num());
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_ii = std::min((M - i), TILE_M);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            conv3x3s1_winograd43_transform_kernel_tile_int8(kernel, A_tile, inch, i, max_ii, k, max_kk);
-
-            Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-            pack_A_tile_int8(A_tile, AT_tile, B, max_ii, max_kk);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_input_tile_int8(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    // const float itm[4][4] = {
-    //     {4,  0, -5,  0, 1, 0},
-    //     {0, -4, -4,  1, 1, 0},
-    //     {0,  4, -4, -1, 1, 0},
-    //     {0, -2, -1,  2, 1, 0},
-    //     {0,  2, -1, -2, 1, 0},
-    //     {0,  4,  0, -5, 0, 1}
-    // };
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w + 1) / 4;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-#if __ARM_NEON
-    nn_max_kk = max_kk / 8;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-        short tmp[6][6][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const signed char* r0 = bottom_blob.channel((k + kk) / elempack).row<const signed char>(ti * 4) + (tj * 4) * elempack;
-
-            int8x8_t _v5 = vdup_n_s8(5);
-
-            for (int m = 0; m < 6; m++)
-            {
-                int8x8_t _r0 = vdup_n_s8(0);
-                int8x8_t _r1 = vdup_n_s8(0);
-                int8x8_t _r2 = vdup_n_s8(0);
-                int8x8_t _r3 = vdup_n_s8(0);
-                int8x8_t _r4 = vdup_n_s8(0);
-                int8x8_t _r5 = vdup_n_s8(0);
-
-                if (ti * 4 + m < h)
-                {
-                    if (elempack == 8)
-                    {
-                        _r0 = vld1_s8(r0);
-                        if (tj * 4 + 1 < w) _r1 = vld1_s8(r0 + 8);
-                        if (tj * 4 + 2 < w) _r2 = vld1_s8(r0 + 16);
-                        if (tj * 4 + 3 < w) _r3 = vld1_s8(r0 + 24);
-                        if (tj * 4 + 4 < w) _r4 = vld1_s8(r0 + 32);
-                        if (tj * 4 + 5 < w) _r5 = vld1_s8(r0 + 40);
-                    }
-                    if (elempack == 1)
-                    {
-                        const signed char* r1 = r0 + N;
-                        const signed char* r2 = r0 + N * 2;
-                        const signed char* r3 = r0 + N * 3;
-                        const signed char* r4 = r0 + N * 4;
-                        const signed char* r5 = r0 + N * 5;
-                        const signed char* r6 = r0 + N * 6;
-                        const signed char* r7 = r0 + N * 7;
-
-                        int8x8_t _t0 = vld1_s8(r0);
-                        int8x8_t _t1 = vld1_s8(r1);
-                        int8x8_t _t2 = vld1_s8(r2);
-                        int8x8_t _t3 = vld1_s8(r3);
-                        int8x8_t _t4 = vld1_s8(r4);
-                        int8x8_t _t5 = vld1_s8(r5);
-                        int8x8_t _t6 = vld1_s8(r6);
-                        int8x8_t _t7 = vld1_s8(r7);
-
-                        int8x8_t _t01 = vzip_s8(_t0, _t1).val[0];
-                        int8x8_t _t23 = vzip_s8(_t2, _t3).val[0];
-                        int8x8_t _t45 = vzip_s8(_t4, _t5).val[0];
-                        int8x8_t _t67 = vzip_s8(_t6, _t7).val[0];
-                        int16x4x2_t _t0123 = vzip_s16(vreinterpret_s16_s8(_t01), vreinterpret_s16_s8(_t23));
-                        int16x4x2_t _t4567 = vzip_s16(vreinterpret_s16_s8(_t45), vreinterpret_s16_s8(_t67));
-                        int16x8_t _ta = vcombine_s16(_t0123.val[0], _t0123.val[1]);
-                        int16x8_t _tb = vcombine_s16(_t4567.val[0], _t4567.val[1]);
-                        int32x4x2_t _tab = vzipq_s32(vreinterpretq_s32_s16(_ta), vreinterpretq_s32_s16(_tb));
-
-                        _r0 = vreinterpret_s8_s32(vget_low_s32(_tab.val[0]));
-                        if (tj * 4 + 1 < w) _r1 = vreinterpret_s8_s32(vget_high_s32(_tab.val[0]));
-                        if (tj * 4 + 2 < w) _r2 = vreinterpret_s8_s32(vget_low_s32(_tab.val[1]));
-                        if (tj * 4 + 3 < w) _r3 = vreinterpret_s8_s32(vget_high_s32(_tab.val[1]));
-                        if (tj * 4 + 4 < w)
-                        {
-                            _t01 = vzip_s8(_t0, _t1).val[1];
-                            _t23 = vzip_s8(_t2, _t3).val[1];
-                            _t45 = vzip_s8(_t4, _t5).val[1];
-                            _t67 = vzip_s8(_t6, _t7).val[1];
-                            int16x4_t _tc = vzip_s16(vreinterpret_s16_s8(_t01), vreinterpret_s16_s8(_t23)).val[0];
-                            int16x4_t _td = vzip_s16(vreinterpret_s16_s8(_t45), vreinterpret_s16_s8(_t67)).val[0];
-                            int32x2x2_t _tcd = vzip_s32(vreinterpret_s32_s16(_tc), vreinterpret_s32_s16(_td));
-
-                            _r4 = vreinterpret_s8_s32(_tcd.val[0]);
-                            if (tj * 4 + 5 < w) _r5 = vreinterpret_s8_s32(_tcd.val[1]);
-                        }
-                    }
-                }
-
-                int16x8_t _tmp12a = vsubw_s8(vshll_n_s8(_r1, 2), _r3);
-                int16x8_t _tmp12b = vsubw_s8(vshll_n_s8(_r2, 2), _r4);
-                int16x8_t _tmp34a = vshlq_n_s16(vsubl_s8(_r3, _r1), 1);
-                int16x8_t _tmp34b = vsubl_s8(_r4, _r2);
-
-                int16x8_t _tmp0 = vaddq_s16(vmovl_s8(_r4), vsubq_s16(vshll_n_s8(_r0, 2), vmull_s8(_r2, _v5)));
-                int16x8_t _tmp1 = vnegq_s16(vaddq_s16(_tmp12a, _tmp12b));
-                int16x8_t _tmp2 = vsubq_s16(_tmp12a, _tmp12b);
-                int16x8_t _tmp3 = vaddq_s16(_tmp34b, _tmp34a);
-                int16x8_t _tmp4 = vsubq_s16(_tmp34b, _tmp34a);
-                int16x8_t _tmp5 = vaddq_s16(vmovl_s8(_r5), vsubq_s16(vshll_n_s8(_r1, 2), vmull_s8(_r3, _v5)));
-
-                vst1q_s16(tmp[0][m], _tmp0);
-                vst1q_s16(tmp[1][m], _tmp1);
-                vst1q_s16(tmp[2][m], _tmp2);
-                vst1q_s16(tmp[3][m], _tmp3);
-                vst1q_s16(tmp[4][m], _tmp4);
-                vst1q_s16(tmp[5][m], _tmp5);
-
-                r0 += w * elempack;
-            }
-
-            int16x8_t _v5q = vdupq_n_s16(5);
-
-            short* p0 = (short*)B + kk * max_jj * 36 + jj * 8;
-            short* p1 = p0 + max_jj * 8;
-            short* p2 = p0 + max_jj * 8 * 2;
-            short* p3 = p0 + max_jj * 8 * 3;
-            short* p4 = p0 + max_jj * 8 * 4;
-            short* p5 = p0 + max_jj * 8 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                int16x8_t _r0 = vld1q_s16(tmp[m][0]);
-                int16x8_t _r1 = vld1q_s16(tmp[m][1]);
-                int16x8_t _r2 = vld1q_s16(tmp[m][2]);
-                int16x8_t _r3 = vld1q_s16(tmp[m][3]);
-                int16x8_t _r4 = vld1q_s16(tmp[m][4]);
-                int16x8_t _r5 = vld1q_s16(tmp[m][5]);
-
-                int16x8_t _tmp12a = vsubq_s16(_r3, vshlq_n_s16(_r1, 2));
-                int16x8_t _tmp12b = vsubq_s16(_r4, vshlq_n_s16(_r2, 2));
-                int16x8_t _tmp34a = vshlq_n_s16(vsubq_s16(_r3, _r1), 1);
-                int16x8_t _tmp34b = vsubq_s16(_r4, _r2);
-
-                int16x8_t _tmp0 = vaddq_s16(_r4, vsubq_s16(vshlq_n_s16(_r0, 2), vmulq_s16(_r2, _v5q)));
-                int16x8_t _tmp1 = vaddq_s16(_tmp12b, _tmp12a);
-                int16x8_t _tmp2 = vsubq_s16(_tmp12b, _tmp12a);
-                int16x8_t _tmp3 = vaddq_s16(_tmp34b, _tmp34a);
-                int16x8_t _tmp4 = vsubq_s16(_tmp34b, _tmp34a);
-                int16x8_t _tmp5 = vaddq_s16(_r5, vsubq_s16(vshlq_n_s16(_r1, 2), vmulq_s16(_r3, _v5q)));
-
-                vst1q_s16(p0, _tmp0);
-                vst1q_s16(p1, _tmp1);
-                vst1q_s16(p2, _tmp2);
-                vst1q_s16(p3, _tmp3);
-                vst1q_s16(p4, _tmp4);
-                vst1q_s16(p5, _tmp5);
-
-                p0 += max_jj * 6 * 8;
-                p1 += max_jj * 6 * 8;
-                p2 += max_jj * 6 * 8;
-                p3 += max_jj * 6 * 8;
-                p4 += max_jj * 6 * 8;
-                p5 += max_jj * 6 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-#else // __ARM_NEON
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    #pragma omp parallel for num_threads(nT)
-#endif // __ARM_NEON
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-        short tmp[6][6][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const signed char* r0 = bottom_blob.channel(k + kk).row<const signed char>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 6; m++)
-            {
-                signed char r00 = 0;
-                signed char r01 = 0;
-                signed char r10 = 0;
-                signed char r11 = 0;
-                signed char r20 = 0;
-                signed char r21 = 0;
-                signed char r30 = 0;
-                signed char r31 = 0;
-                signed char r40 = 0;
-                signed char r41 = 0;
-                signed char r50 = 0;
-                signed char r51 = 0;
-
-                if (ti * 4 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const signed char* r1 = r0 + N;
-
-                        r00 = r0[0];
-                        r01 = r1[0];
-                        if (tj * 4 + 1 < w)
-                        {
-                            r10 = r0[1];
-                            r11 = r1[1];
-                        }
-                        if (tj * 4 + 2 < w)
-                        {
-                            r20 = r0[2];
-                            r21 = r1[2];
-                        }
-                        if (tj * 4 + 3 < w)
-                        {
-                            r30 = r0[3];
-                            r31 = r1[3];
-                        }
-                        if (tj * 4 + 4 < w)
-                        {
-                            r40 = r0[4];
-                            r41 = r1[4];
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            r50 = r0[5];
-                            r51 = r1[5];
-                        }
-                    }
-                }
-
-                short tmp120a = r30 - r10 * 4;
-                short tmp121a = r31 - r11 * 4;
-                short tmp120b = r40 - r20 * 4;
-                short tmp121b = r41 - r21 * 4;
-                short tmp340a = (r30 - r10) * 2;
-                short tmp341a = (r31 - r11) * 2;
-                short tmp340b = r40 - r20;
-                short tmp341b = r41 - r21;
-
-                tmp[0][m][0] = r40 + r00 * 4 - r20 * 5;
-                tmp[0][m][1] = r41 + r01 * 4 - r21 * 5;
-                tmp[1][m][0] = tmp120b + tmp120a;
-                tmp[1][m][1] = tmp121b + tmp121a;
-                tmp[2][m][0] = tmp120b - tmp120a;
-                tmp[2][m][1] = tmp121b - tmp121a;
-                tmp[3][m][0] = tmp340b + tmp340a;
-                tmp[3][m][1] = tmp341b + tmp341a;
-                tmp[4][m][0] = tmp340b - tmp340a;
-                tmp[4][m][1] = tmp341b - tmp341a;
-                tmp[5][m][0] = r50 + r10 * 4 - r30 * 5;
-                tmp[5][m][1] = r51 + r11 * 4 - r31 * 5;
-
-                r0 += w;
-            }
-
-            short* p0 = (short*)B + kk * max_jj * 36 + jj * 2;
-            short* p1 = p0 + max_jj * 2;
-            short* p2 = p0 + max_jj * 2 * 2;
-            short* p3 = p0 + max_jj * 2 * 3;
-            short* p4 = p0 + max_jj * 2 * 4;
-            short* p5 = p0 + max_jj * 2 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                short r00 = tmp[m][0][0];
-                short r01 = tmp[m][0][1];
-                short r10 = tmp[m][1][0];
-                short r11 = tmp[m][1][1];
-                short r20 = tmp[m][2][0];
-                short r21 = tmp[m][2][1];
-                short r30 = tmp[m][3][0];
-                short r31 = tmp[m][3][1];
-                short r40 = tmp[m][4][0];
-                short r41 = tmp[m][4][1];
-                short r50 = tmp[m][5][0];
-                short r51 = tmp[m][5][1];
-
-                short tmp120a = r30 - r10 * 4;
-                short tmp121a = r31 - r11 * 4;
-                short tmp120b = r40 - r20 * 4;
-                short tmp121b = r41 - r21 * 4;
-                short tmp340a = (r30 - r10) * 2;
-                short tmp341a = (r31 - r11) * 2;
-                short tmp340b = r40 - r20;
-                short tmp341b = r41 - r21;
-
-                p0[0] = r40 + r00 * 4 - r20 * 5;
-                p0[1] = r41 + r01 * 4 - r21 * 5;
-                p1[0] = tmp120b + tmp120a;
-                p1[1] = tmp121b + tmp121a;
-                p2[0] = tmp120b - tmp120a;
-                p2[1] = tmp121b - tmp121a;
-                p3[0] = tmp340b + tmp340a;
-                p3[1] = tmp341b + tmp341a;
-                p4[0] = tmp340b - tmp340a;
-                p4[1] = tmp341b - tmp341a;
-                p5[0] = r50 + r10 * 4 - r30 * 5;
-                p5[1] = r51 + r11 * 4 - r31 * 5;
-
-                p0 += max_jj * 6 * 2;
-                p1 += max_jj * 6 * 2;
-                p2 += max_jj * 6 * 2;
-                p3 += max_jj * 6 * 2;
-                p4 += max_jj * 6 * 2;
-                p5 += max_jj * 6 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        short tmp[6][6];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const signed char* r0123 = bottom_blob.channel(k + kk).row<const signed char>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 6; m++)
-            {
-                signed char r0 = 0;
-                signed char r1 = 0;
-                signed char r2 = 0;
-                signed char r3 = 0;
-                signed char r4 = 0;
-                signed char r5 = 0;
-
-                if (ti * 4 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = r0123[0];
-                        if (tj * 4 + 1 < w) r1 = r0123[1];
-                        if (tj * 4 + 2 < w) r2 = r0123[2];
-                        if (tj * 4 + 3 < w) r3 = r0123[3];
-                        if (tj * 4 + 4 < w) r4 = r0123[4];
-                        if (tj * 4 + 5 < w) r5 = r0123[5];
-                    }
-                }
-
-                short tmp12a = r3 - r1 * 4;
-                short tmp12b = r4 - r2 * 4;
-                short tmp34a = (r3 - r1) * 2;
-                short tmp34b = r4 - r2;
-
-                tmp[0][m] = r4 + r0 * 4 - r2 * 5;
-                tmp[1][m] = tmp12b + tmp12a;
-                tmp[2][m] = tmp12b - tmp12a;
-                tmp[3][m] = tmp34b + tmp34a;
-                tmp[4][m] = tmp34b - tmp34a;
-                tmp[5][m] = r5 + r1 * 4 - r3 * 5;
-
-                r0123 += w;
-            }
-
-            short* p0 = (short*)B + kk * max_jj * 36 + jj;
-            short* p1 = p0 + max_jj;
-            short* p2 = p0 + max_jj * 2;
-            short* p3 = p0 + max_jj * 3;
-            short* p4 = p0 + max_jj * 4;
-            short* p5 = p0 + max_jj * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                short r0 = tmp[m][0];
-                short r1 = tmp[m][1];
-                short r2 = tmp[m][2];
-                short r3 = tmp[m][3];
-                short r4 = tmp[m][4];
-                short r5 = tmp[m][5];
-
-                short tmp12a = r3 - r1 * 4;
-                short tmp12b = r4 - r2 * 4;
-                short tmp34a = (r3 - r1) * 2;
-                short tmp34b = r4 - r2;
-
-                p0[0] = r4 + r0 * 4 - r2 * 5;
-                p1[0] = tmp12b + tmp12a;
-                p2[0] = tmp12b - tmp12a;
-                p3[0] = tmp34b + tmp34a;
-                p4[0] = tmp34b - tmp34a;
-                p5[0] = r5 + r1 * 4 - r3 * 5;
-
-                p0 += max_jj * 6;
-                p1 += max_jj * 6;
-                p2 += max_jj * 6;
-                p3 += max_jj * 6;
-                p4 += max_jj * 6;
-                p5 += max_jj * 6;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_output_tile_int8(const Mat& top_tile, Mat& top_blob, int i, int max_ii, int j, int max_jj)
-{
-    // const int otm[4][6] = {
-    //     {1, 1,  1, 1,  1, 0},
-    //     {0, 1, -1, 2, -2, 0},
-    //     {0, 1,  1, 4,  4, 0},
-    //     {0, 1, -1, 8, -8, 1}
-    // };
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 3) / 4;
-
-    int ii = 0;
-#if __ARM_NEON
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        int tmp[4][6][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 36 + jj * 8;
-            const int* r1 = r0 + max_jj * 8;
-            const int* r2 = r0 + max_jj * 8 * 2;
-            const int* r3 = r0 + max_jj * 8 * 3;
-            const int* r4 = r0 + max_jj * 8 * 4;
-            const int* r5 = r0 + max_jj * 8 * 5;
-
-            for (int m = 0; m < 5; m++)
-            {
-                int32x4_t _r00 = vld1q_s32(r0);
-                int32x4_t _r01 = vld1q_s32(r0 + 4);
-                int32x4_t _r10 = vld1q_s32(r1);
-                int32x4_t _r11 = vld1q_s32(r1 + 4);
-                int32x4_t _r20 = vld1q_s32(r2);
-                int32x4_t _r21 = vld1q_s32(r2 + 4);
-                int32x4_t _r30 = vld1q_s32(r3);
-                int32x4_t _r31 = vld1q_s32(r3 + 4);
-                int32x4_t _r40 = vld1q_s32(r4);
-                int32x4_t _r41 = vld1q_s32(r4 + 4);
-                int32x4_t _r50 = vld1q_s32(r5);
-                int32x4_t _r51 = vld1q_s32(r5 + 4);
-
-                int32x4_t _tmp02a0 = vaddq_s32(_r10, _r20);
-                int32x4_t _tmp02a1 = vaddq_s32(_r11, _r21);
-                int32x4_t _tmp02b0 = vaddq_s32(_r30, _r40);
-                int32x4_t _tmp02b1 = vaddq_s32(_r31, _r41);
-                int32x4_t _tmp13a0 = vsubq_s32(_r10, _r20);
-                int32x4_t _tmp13a1 = vsubq_s32(_r11, _r21);
-                int32x4_t _tmp13b0 = vsubq_s32(_r30, _r40);
-                int32x4_t _tmp13b1 = vsubq_s32(_r31, _r41);
-
-                int32x4_t _tmp00 = vaddq_s32(vaddq_s32(_tmp02a0, _tmp02b0), _r00);
-                int32x4_t _tmp01 = vaddq_s32(vaddq_s32(_tmp02a1, _tmp02b1), _r01);
-                int32x4_t _tmp10 = vaddq_s32(_tmp13a0, vshlq_n_s32(_tmp13b0, 1));
-                int32x4_t _tmp11 = vaddq_s32(_tmp13a1, vshlq_n_s32(_tmp13b1, 1));
-                int32x4_t _tmp20 = vaddq_s32(_tmp02a0, vshlq_n_s32(_tmp02b0, 2));
-                int32x4_t _tmp21 = vaddq_s32(_tmp02a1, vshlq_n_s32(_tmp02b1, 2));
-                int32x4_t _tmp30 = vaddq_s32(vaddq_s32(_tmp13a0, vshlq_n_s32(_tmp13b0, 3)), vshlq_n_s32(_r50, 2));
-                int32x4_t _tmp31 = vaddq_s32(vaddq_s32(_tmp13a1, vshlq_n_s32(_tmp13b1, 3)), vshlq_n_s32(_r51, 2));
-
-                vst1q_s32(tmp[0][m], _tmp00);
-                vst1q_s32(tmp[0][m] + 4, _tmp01);
-                vst1q_s32(tmp[1][m], _tmp10);
-                vst1q_s32(tmp[1][m] + 4, _tmp11);
-                vst1q_s32(tmp[2][m], _tmp20);
-                vst1q_s32(tmp[2][m] + 4, _tmp21);
-                vst1q_s32(tmp[3][m], _tmp30);
-                vst1q_s32(tmp[3][m] + 4, _tmp31);
-
-                r0 += max_jj * 6 * 8;
-                r1 += max_jj * 6 * 8;
-                r2 += max_jj * 6 * 8;
-                r3 += max_jj * 6 * 8;
-                r4 += max_jj * 6 * 8;
-                r5 += max_jj * 6 * 8;
-            }
-            for (int m = 5; m < 6; m++)
-            {
-                int32x4_t _r00 = vld1q_s32(r0);
-                int32x4_t _r01 = vld1q_s32(r0 + 4);
-                int32x4_t _r10 = vld1q_s32(r1);
-                int32x4_t _r11 = vld1q_s32(r1 + 4);
-                int32x4_t _r20 = vld1q_s32(r2);
-                int32x4_t _r21 = vld1q_s32(r2 + 4);
-                int32x4_t _r30 = vld1q_s32(r3);
-                int32x4_t _r31 = vld1q_s32(r3 + 4);
-                int32x4_t _r40 = vld1q_s32(r4);
-                int32x4_t _r41 = vld1q_s32(r4 + 4);
-                int32x4_t _r50 = vld1q_s32(r5);
-                int32x4_t _r51 = vld1q_s32(r5 + 4);
-
-                int32x4_t _tmp02a0 = vaddq_s32(_r10, _r20);
-                int32x4_t _tmp02a1 = vaddq_s32(_r11, _r21);
-                int32x4_t _tmp02b0 = vaddq_s32(_r30, _r40);
-                int32x4_t _tmp02b1 = vaddq_s32(_r31, _r41);
-                int32x4_t _tmp13a0 = vsubq_s32(_r10, _r20);
-                int32x4_t _tmp13a1 = vsubq_s32(_r11, _r21);
-                int32x4_t _tmp13b0 = vsubq_s32(_r30, _r40);
-                int32x4_t _tmp13b1 = vsubq_s32(_r31, _r41);
-
-                int32x4_t _tmp00 = vaddq_s32(vaddq_s32(_tmp02a0, _tmp02b0), _r00);
-                int32x4_t _tmp01 = vaddq_s32(vaddq_s32(_tmp02a1, _tmp02b1), _r01);
-                int32x4_t _tmp10 = vaddq_s32(_tmp13a0, vshlq_n_s32(_tmp13b0, 1));
-                int32x4_t _tmp11 = vaddq_s32(_tmp13a1, vshlq_n_s32(_tmp13b1, 1));
-                int32x4_t _tmp20 = vaddq_s32(_tmp02a0, vshlq_n_s32(_tmp02b0, 2));
-                int32x4_t _tmp21 = vaddq_s32(_tmp02a1, vshlq_n_s32(_tmp02b1, 2));
-                int32x4_t _tmp30 = vaddq_s32(vaddq_s32(_tmp13a0, vshlq_n_s32(_tmp13b0, 3)), vshlq_n_s32(_r50, 2));
-                int32x4_t _tmp31 = vaddq_s32(vaddq_s32(_tmp13a1, vshlq_n_s32(_tmp13b1, 3)), vshlq_n_s32(_r51, 2));
-
-                _tmp00 = vshlq_n_s32(_tmp00, 2);
-                _tmp01 = vshlq_n_s32(_tmp01, 2);
-                _tmp10 = vshlq_n_s32(_tmp10, 2);
-                _tmp11 = vshlq_n_s32(_tmp11, 2);
-                _tmp20 = vshlq_n_s32(_tmp20, 2);
-                _tmp21 = vshlq_n_s32(_tmp21, 2);
-                _tmp30 = vshlq_n_s32(_tmp30, 2);
-                _tmp31 = vshlq_n_s32(_tmp31, 2);
-
-                vst1q_s32(tmp[0][m], _tmp00);
-                vst1q_s32(tmp[0][m] + 4, _tmp01);
-                vst1q_s32(tmp[1][m], _tmp10);
-                vst1q_s32(tmp[1][m] + 4, _tmp11);
-                vst1q_s32(tmp[2][m], _tmp20);
-                vst1q_s32(tmp[2][m] + 4, _tmp21);
-                vst1q_s32(tmp[3][m], _tmp30);
-                vst1q_s32(tmp[3][m] + 4, _tmp31);
-
-                r0 += max_jj * 6 * 8;
-                r1 += max_jj * 6 * 8;
-                r2 += max_jj * 6 * 8;
-                r3 += max_jj * 6 * 8;
-                r4 += max_jj * 6 * 8;
-                r5 += max_jj * 6 * 8;
-            }
-
-            int* outptr0 = top_blob.channel((i + ii) / out_elempack).row<int>(ti * 4) + (tj * 4) * out_elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                int32x4_t _r00 = vld1q_s32(tmp[m][0]);
-                int32x4_t _r01 = vld1q_s32(tmp[m][0] + 4);
-                int32x4_t _r10 = vld1q_s32(tmp[m][1]);
-                int32x4_t _r11 = vld1q_s32(tmp[m][1] + 4);
-                int32x4_t _r20 = vld1q_s32(tmp[m][2]);
-                int32x4_t _r21 = vld1q_s32(tmp[m][2] + 4);
-                int32x4_t _r30 = vld1q_s32(tmp[m][3]);
-                int32x4_t _r31 = vld1q_s32(tmp[m][3] + 4);
-                int32x4_t _r40 = vld1q_s32(tmp[m][4]);
-                int32x4_t _r41 = vld1q_s32(tmp[m][4] + 4);
-                int32x4_t _r50 = vld1q_s32(tmp[m][5]);
-                int32x4_t _r51 = vld1q_s32(tmp[m][5] + 4);
-
-                int32x4_t _tmp02a0 = vaddq_s32(_r10, _r20);
-                int32x4_t _tmp02a1 = vaddq_s32(_r11, _r21);
-                int32x4_t _tmp02b0 = vaddq_s32(_r30, _r40);
-                int32x4_t _tmp02b1 = vaddq_s32(_r31, _r41);
-                int32x4_t _tmp13a0 = vsubq_s32(_r10, _r20);
-                int32x4_t _tmp13a1 = vsubq_s32(_r11, _r21);
-                int32x4_t _tmp13b0 = vsubq_s32(_r30, _r40);
-                int32x4_t _tmp13b1 = vsubq_s32(_r31, _r41);
-
-                int32x4_t _tmp00 = vaddq_s32(vaddq_s32(_tmp02a0, _tmp02b0), _r00);
-                int32x4_t _tmp01 = vaddq_s32(vaddq_s32(_tmp02a1, _tmp02b1), _r01);
-                int32x4_t _tmp10 = vaddq_s32(_tmp13a0, vshlq_n_s32(_tmp13b0, 1));
-                int32x4_t _tmp11 = vaddq_s32(_tmp13a1, vshlq_n_s32(_tmp13b1, 1));
-                int32x4_t _tmp20 = vaddq_s32(_tmp02a0, vshlq_n_s32(_tmp02b0, 2));
-                int32x4_t _tmp21 = vaddq_s32(_tmp02a1, vshlq_n_s32(_tmp02b1, 2));
-                int32x4_t _tmp30 = vaddq_s32(vaddq_s32(_tmp13a0, vshlq_n_s32(_tmp13b0, 3)), _r50);
-                int32x4_t _tmp31 = vaddq_s32(vaddq_s32(_tmp13a1, vshlq_n_s32(_tmp13b1, 3)), _r51);
-
-                // TODO use integer trick for division by 576
-                float32x4_t _v576 = vdupq_n_f32(1.0 / 576);
-                _tmp00 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(_tmp00), _v576));
-                _tmp01 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(_tmp01), _v576));
-                _tmp10 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(_tmp10), _v576));
-                _tmp11 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(_tmp11), _v576));
-                _tmp20 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(_tmp20), _v576));
-                _tmp21 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(_tmp21), _v576));
-                _tmp30 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(_tmp30), _v576));
-                _tmp31 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(_tmp31), _v576));
-
-                if (out_elempack == 8)
-                {
-                    vst1q_s32(outptr0, _tmp00);
-                    vst1q_s32(outptr0 + 4, _tmp01);
-                    if (tj * 4 + 1 < outw)
-                    {
-                        vst1q_s32(outptr0 + 8, _tmp10);
-                        vst1q_s32(outptr0 + 12, _tmp11);
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        vst1q_s32(outptr0 + 16, _tmp20);
-                        vst1q_s32(outptr0 + 20, _tmp21);
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        vst1q_s32(outptr0 + 24, _tmp30);
-                        vst1q_s32(outptr0 + 28, _tmp31);
-                    }
-                }
-                if (out_elempack == 4)
-                {
-                    int* outptr1 = outptr0 + N;
-
-                    vst1q_s32(outptr0, _tmp00);
-                    vst1q_s32(outptr1, _tmp01);
-                    if (tj * 4 + 1 < outw)
-                    {
-                        vst1q_s32(outptr0 + 4, _tmp10);
-                        vst1q_s32(outptr1 + 4, _tmp11);
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        vst1q_s32(outptr0 + 8, _tmp20);
-                        vst1q_s32(outptr1 + 8, _tmp21);
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        vst1q_s32(outptr0 + 12, _tmp30);
-                        vst1q_s32(outptr1 + 12, _tmp31);
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    int* outptr1 = outptr0 + N;
-                    int* outptr2 = outptr0 + N * 2;
-                    int* outptr3 = outptr0 + N * 3;
-                    int* outptr4 = outptr0 + N * 4;
-                    int* outptr5 = outptr0 + N * 5;
-                    int* outptr6 = outptr0 + N * 6;
-                    int* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = vgetq_lane_s32(_tmp00, 0);
-                    outptr1[0] = vgetq_lane_s32(_tmp00, 1);
-                    outptr2[0] = vgetq_lane_s32(_tmp00, 2);
-                    outptr3[0] = vgetq_lane_s32(_tmp00, 3);
-                    outptr4[0] = vgetq_lane_s32(_tmp01, 0);
-                    outptr5[0] = vgetq_lane_s32(_tmp01, 1);
-                    outptr6[0] = vgetq_lane_s32(_tmp01, 2);
-                    outptr7[0] = vgetq_lane_s32(_tmp01, 3);
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = vgetq_lane_s32(_tmp10, 0);
-                        outptr1[1] = vgetq_lane_s32(_tmp10, 1);
-                        outptr2[1] = vgetq_lane_s32(_tmp10, 2);
-                        outptr3[1] = vgetq_lane_s32(_tmp10, 3);
-                        outptr4[1] = vgetq_lane_s32(_tmp11, 0);
-                        outptr5[1] = vgetq_lane_s32(_tmp11, 1);
-                        outptr6[1] = vgetq_lane_s32(_tmp11, 2);
-                        outptr7[1] = vgetq_lane_s32(_tmp11, 3);
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = vgetq_lane_s32(_tmp20, 0);
-                        outptr1[2] = vgetq_lane_s32(_tmp20, 1);
-                        outptr2[2] = vgetq_lane_s32(_tmp20, 2);
-                        outptr3[2] = vgetq_lane_s32(_tmp20, 3);
-                        outptr4[2] = vgetq_lane_s32(_tmp21, 0);
-                        outptr5[2] = vgetq_lane_s32(_tmp21, 1);
-                        outptr6[2] = vgetq_lane_s32(_tmp21, 2);
-                        outptr7[2] = vgetq_lane_s32(_tmp21, 3);
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = vgetq_lane_s32(_tmp30, 0);
-                        outptr1[3] = vgetq_lane_s32(_tmp30, 1);
-                        outptr2[3] = vgetq_lane_s32(_tmp30, 2);
-                        outptr3[3] = vgetq_lane_s32(_tmp30, 3);
-                        outptr4[3] = vgetq_lane_s32(_tmp31, 0);
-                        outptr5[3] = vgetq_lane_s32(_tmp31, 1);
-                        outptr6[3] = vgetq_lane_s32(_tmp31, 2);
-                        outptr7[3] = vgetq_lane_s32(_tmp31, 3);
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        int tmp[4][6][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 36 + jj * 4;
-            const int* r1 = r0 + max_jj * 4;
-            const int* r2 = r0 + max_jj * 4 * 2;
-            const int* r3 = r0 + max_jj * 4 * 3;
-            const int* r4 = r0 + max_jj * 4 * 4;
-            const int* r5 = r0 + max_jj * 4 * 5;
-
-            for (int m = 0; m < 5; m++)
-            {
-                int32x4_t _r0 = vld1q_s32(r0);
-                int32x4_t _r1 = vld1q_s32(r1);
-                int32x4_t _r2 = vld1q_s32(r2);
-                int32x4_t _r3 = vld1q_s32(r3);
-                int32x4_t _r4 = vld1q_s32(r4);
-                int32x4_t _r5 = vld1q_s32(r5);
-
-                int32x4_t _tmp02a = vaddq_s32(_r1, _r2);
-                int32x4_t _tmp02b = vaddq_s32(_r3, _r4);
-                int32x4_t _tmp13a = vsubq_s32(_r1, _r2);
-                int32x4_t _tmp13b = vsubq_s32(_r3, _r4);
-
-                int32x4_t _tmp0 = vaddq_s32(vaddq_s32(_tmp02a, _tmp02b), _r0);
-                int32x4_t _tmp1 = vaddq_s32(_tmp13a, vshlq_n_s32(_tmp13b, 1));
-                int32x4_t _tmp2 = vaddq_s32(_tmp02a, vshlq_n_s32(_tmp02b, 2));
-                int32x4_t _tmp3 = vaddq_s32(vaddq_s32(_tmp13a, vshlq_n_s32(_tmp13b, 3)), vshlq_n_s32(_r5, 2));
-
-                vst1q_s32(tmp[0][m], _tmp0);
-                vst1q_s32(tmp[1][m], _tmp1);
-                vst1q_s32(tmp[2][m], _tmp2);
-                vst1q_s32(tmp[3][m], _tmp3);
-
-                r0 += max_jj * 6 * 4;
-                r1 += max_jj * 6 * 4;
-                r2 += max_jj * 6 * 4;
-                r3 += max_jj * 6 * 4;
-                r4 += max_jj * 6 * 4;
-                r5 += max_jj * 6 * 4;
-            }
-            for (int m = 5; m < 6; m++)
-            {
-                int32x4_t _r0 = vld1q_s32(r0);
-                int32x4_t _r1 = vld1q_s32(r1);
-                int32x4_t _r2 = vld1q_s32(r2);
-                int32x4_t _r3 = vld1q_s32(r3);
-                int32x4_t _r4 = vld1q_s32(r4);
-                int32x4_t _r5 = vld1q_s32(r5);
-
-                int32x4_t _tmp02a = vaddq_s32(_r1, _r2);
-                int32x4_t _tmp02b = vaddq_s32(_r3, _r4);
-                int32x4_t _tmp13a = vsubq_s32(_r1, _r2);
-                int32x4_t _tmp13b = vsubq_s32(_r3, _r4);
-
-                int32x4_t _tmp0 = vaddq_s32(vaddq_s32(_tmp02a, _tmp02b), _r0);
-                int32x4_t _tmp1 = vaddq_s32(_tmp13a, vshlq_n_s32(_tmp13b, 1));
-                int32x4_t _tmp2 = vaddq_s32(_tmp02a, vshlq_n_s32(_tmp02b, 2));
-                int32x4_t _tmp3 = vaddq_s32(vaddq_s32(_tmp13a, vshlq_n_s32(_tmp13b, 3)), vshlq_n_s32(_r5, 2));
-
-                _tmp0 = vshlq_n_s32(_tmp0, 2);
-                _tmp1 = vshlq_n_s32(_tmp1, 2);
-                _tmp2 = vshlq_n_s32(_tmp2, 2);
-                _tmp3 = vshlq_n_s32(_tmp3, 2);
-
-                vst1q_s32(tmp[0][m], _tmp0);
-                vst1q_s32(tmp[1][m], _tmp1);
-                vst1q_s32(tmp[2][m], _tmp2);
-                vst1q_s32(tmp[3][m], _tmp3);
-
-                r0 += max_jj * 6 * 4;
-                r1 += max_jj * 6 * 4;
-                r2 += max_jj * 6 * 4;
-                r3 += max_jj * 6 * 4;
-                r4 += max_jj * 6 * 4;
-                r5 += max_jj * 6 * 4;
-            }
-
-            int* outptr0 = top_blob.channel((i + ii) / out_elempack).row<int>(ti * 4) + (tj * 4) * out_elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                int32x4_t _r0 = vld1q_s32(tmp[m][0]);
-                int32x4_t _r1 = vld1q_s32(tmp[m][1]);
-                int32x4_t _r2 = vld1q_s32(tmp[m][2]);
-                int32x4_t _r3 = vld1q_s32(tmp[m][3]);
-                int32x4_t _r4 = vld1q_s32(tmp[m][4]);
-                int32x4_t _r5 = vld1q_s32(tmp[m][5]);
-
-                int32x4_t _tmp02a = vaddq_s32(_r1, _r2);
-                int32x4_t _tmp02b = vaddq_s32(_r3, _r4);
-                int32x4_t _tmp13a = vsubq_s32(_r1, _r2);
-                int32x4_t _tmp13b = vsubq_s32(_r3, _r4);
-
-                int32x4_t _tmp0 = vaddq_s32(vaddq_s32(_tmp02a, _tmp02b), _r0);
-                int32x4_t _tmp1 = vaddq_s32(_tmp13a, vshlq_n_s32(_tmp13b, 1));
-                int32x4_t _tmp2 = vaddq_s32(_tmp02a, vshlq_n_s32(_tmp02b, 2));
-                int32x4_t _tmp3 = vaddq_s32(vaddq_s32(_tmp13a, vshlq_n_s32(_tmp13b, 3)), _r5);
-
-                // TODO use integer trick for division by 576
-                float32x4_t _v576 = vdupq_n_f32(1.0 / 576);
-                _tmp0 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(_tmp0), _v576));
-                _tmp1 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(_tmp1), _v576));
-                _tmp2 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(_tmp2), _v576));
-                _tmp3 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(_tmp3), _v576));
-
-                if (out_elempack == 4)
-                {
-                    vst1q_s32(outptr0, _tmp0);
-                    if (tj * 4 + 1 < outw) vst1q_s32(outptr0 + 4, _tmp1);
-                    if (tj * 4 + 2 < outw) vst1q_s32(outptr0 + 8, _tmp2);
-                    if (tj * 4 + 3 < outw) vst1q_s32(outptr0 + 12, _tmp3);
-                }
-                if (out_elempack == 1)
-                {
-                    int* outptr1 = outptr0 + N;
-                    int* outptr2 = outptr0 + N * 2;
-                    int* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = vgetq_lane_s32(_tmp0, 0);
-                    outptr1[0] = vgetq_lane_s32(_tmp0, 1);
-                    outptr2[0] = vgetq_lane_s32(_tmp0, 2);
-                    outptr3[0] = vgetq_lane_s32(_tmp0, 3);
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = vgetq_lane_s32(_tmp1, 0);
-                        outptr1[1] = vgetq_lane_s32(_tmp1, 1);
-                        outptr2[1] = vgetq_lane_s32(_tmp1, 2);
-                        outptr3[1] = vgetq_lane_s32(_tmp1, 3);
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = vgetq_lane_s32(_tmp2, 0);
-                        outptr1[2] = vgetq_lane_s32(_tmp2, 1);
-                        outptr2[2] = vgetq_lane_s32(_tmp2, 2);
-                        outptr3[2] = vgetq_lane_s32(_tmp2, 3);
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = vgetq_lane_s32(_tmp3, 0);
-                        outptr1[3] = vgetq_lane_s32(_tmp3, 1);
-                        outptr2[3] = vgetq_lane_s32(_tmp3, 2);
-                        outptr3[3] = vgetq_lane_s32(_tmp3, 3);
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        int tmp[4][6][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 36 + jj * 2;
-            const int* r1 = r0 + max_jj * 2;
-            const int* r2 = r0 + max_jj * 2 * 2;
-            const int* r3 = r0 + max_jj * 2 * 3;
-            const int* r4 = r0 + max_jj * 2 * 4;
-            const int* r5 = r0 + max_jj * 2 * 5;
-
-            for (int m = 0; m < 5; m++)
-            {
-                int tmp02a0 = r1[0] + r2[0];
-                int tmp02a1 = r1[1] + r2[1];
-                int tmp02b0 = r3[0] + r4[0];
-                int tmp02b1 = r3[1] + r4[1];
-                int tmp13a0 = r1[0] - r2[0];
-                int tmp13a1 = r1[1] - r2[1];
-                int tmp13b0 = r3[0] - r4[0];
-                int tmp13b1 = r3[1] - r4[1];
-
-                int tmp00 = tmp02a0 + tmp02b0 + r0[0];
-                int tmp01 = tmp02a1 + tmp02b1 + r0[1];
-                int tmp10 = tmp13a0 + tmp13b0 * 2;
-                int tmp11 = tmp13a1 + tmp13b1 * 2;
-                int tmp20 = tmp02a0 + tmp02b0 * 4;
-                int tmp21 = tmp02a1 + tmp02b1 * 4;
-                int tmp30 = tmp13a0 + tmp13b0 * 8 + r5[0] * 4;
-                int tmp31 = tmp13a1 + tmp13b1 * 8 + r5[1] * 4;
-
-                tmp[0][m][0] = tmp00;
-                tmp[0][m][1] = tmp01;
-                tmp[1][m][0] = tmp10;
-                tmp[1][m][1] = tmp11;
-                tmp[2][m][0] = tmp20;
-                tmp[2][m][1] = tmp21;
-                tmp[3][m][0] = tmp30;
-                tmp[3][m][1] = tmp31;
-
-                r0 += max_jj * 6 * 2;
-                r1 += max_jj * 6 * 2;
-                r2 += max_jj * 6 * 2;
-                r3 += max_jj * 6 * 2;
-                r4 += max_jj * 6 * 2;
-                r5 += max_jj * 6 * 2;
-            }
-            for (int m = 5; m < 6; m++)
-            {
-                int tmp02a0 = r1[0] + r2[0];
-                int tmp02a1 = r1[1] + r2[1];
-                int tmp02b0 = r3[0] + r4[0];
-                int tmp02b1 = r3[1] + r4[1];
-                int tmp13a0 = r1[0] - r2[0];
-                int tmp13a1 = r1[1] - r2[1];
-                int tmp13b0 = r3[0] - r4[0];
-                int tmp13b1 = r3[1] - r4[1];
-
-                int tmp00 = tmp02a0 + tmp02b0 + r0[0];
-                int tmp01 = tmp02a1 + tmp02b1 + r0[1];
-                int tmp10 = tmp13a0 + tmp13b0 * 2;
-                int tmp11 = tmp13a1 + tmp13b1 * 2;
-                int tmp20 = tmp02a0 + tmp02b0 * 4;
-                int tmp21 = tmp02a1 + tmp02b1 * 4;
-                int tmp30 = tmp13a0 + tmp13b0 * 8 + r5[0] * 4;
-                int tmp31 = tmp13a1 + tmp13b1 * 8 + r5[1] * 4;
-
-                tmp00 = tmp00 * 4;
-                tmp01 = tmp01 * 4;
-                tmp10 = tmp10 * 4;
-                tmp11 = tmp11 * 4;
-                tmp20 = tmp20 * 4;
-                tmp21 = tmp21 * 4;
-                tmp30 = tmp30 * 4;
-                tmp31 = tmp31 * 4;
-
-                tmp[0][m][0] = tmp00;
-                tmp[0][m][1] = tmp01;
-                tmp[1][m][0] = tmp10;
-                tmp[1][m][1] = tmp11;
-                tmp[2][m][0] = tmp20;
-                tmp[2][m][1] = tmp21;
-                tmp[3][m][0] = tmp30;
-                tmp[3][m][1] = tmp31;
-
-                r0 += max_jj * 6 * 2;
-                r1 += max_jj * 6 * 2;
-                r2 += max_jj * 6 * 2;
-                r3 += max_jj * 6 * 2;
-                r4 += max_jj * 6 * 2;
-                r5 += max_jj * 6 * 2;
-            }
-
-            int* outptr0 = top_blob.channel(i + ii).row<int>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                int tmp02a0 = tmp[m][1][0] + tmp[m][2][0];
-                int tmp02a1 = tmp[m][1][1] + tmp[m][2][1];
-                int tmp02b0 = tmp[m][3][0] + tmp[m][4][0];
-                int tmp02b1 = tmp[m][3][1] + tmp[m][4][1];
-                int tmp13a0 = tmp[m][1][0] - tmp[m][2][0];
-                int tmp13a1 = tmp[m][1][1] - tmp[m][2][1];
-                int tmp13b0 = tmp[m][3][0] - tmp[m][4][0];
-                int tmp13b1 = tmp[m][3][1] - tmp[m][4][1];
-
-                int tmp00 = tmp02a0 + tmp02b0 + tmp[m][0][0];
-                int tmp01 = tmp02a1 + tmp02b1 + tmp[m][0][1];
-                int tmp10 = tmp13a0 + tmp13b0 * 2;
-                int tmp11 = tmp13a1 + tmp13b1 * 2;
-                int tmp20 = tmp02a0 + tmp02b0 * 4;
-                int tmp21 = tmp02a1 + tmp02b1 * 4;
-                int tmp30 = tmp13a0 + tmp13b0 * 8 + tmp[m][5][0];
-                int tmp31 = tmp13a1 + tmp13b1 * 8 + tmp[m][5][1];
-
-                tmp00 = tmp00 / 576;
-                tmp01 = tmp01 / 576;
-                tmp10 = tmp10 / 576;
-                tmp11 = tmp11 / 576;
-                tmp20 = tmp20 / 576;
-                tmp21 = tmp21 / 576;
-                tmp30 = tmp30 / 576;
-                tmp31 = tmp31 / 576;
-
-                // if (out_elempack == 1)
-                {
-                    int* outptr1 = outptr0 + N;
-
-                    outptr0[0] = tmp00;
-                    outptr1[0] = tmp01;
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp10;
-                        outptr1[1] = tmp11;
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp20;
-                        outptr1[2] = tmp21;
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp30;
-                        outptr1[3] = tmp31;
-                    }
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        int tmp[4][6];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 36 + jj;
-            const int* r1 = r0 + max_jj;
-            const int* r2 = r0 + max_jj * 2;
-            const int* r3 = r0 + max_jj * 3;
-            const int* r4 = r0 + max_jj * 4;
-            const int* r5 = r0 + max_jj * 5;
-
-            for (int m = 0; m < 5; m++)
-            {
-                int tmp02a = r1[0] + r2[0];
-                int tmp02b = r3[0] + r4[0];
-                int tmp13a = r1[0] - r2[0];
-                int tmp13b = r3[0] - r4[0];
-
-                int tmp0 = tmp02a + tmp02b + r0[0];
-                int tmp1 = tmp13a + tmp13b * 2;
-                int tmp2 = tmp02a + tmp02b * 4;
-                int tmp3 = tmp13a + tmp13b * 8 + r5[0] * 4;
-
-                tmp[0][m] = tmp0;
-                tmp[1][m] = tmp1;
-                tmp[2][m] = tmp2;
-                tmp[3][m] = tmp3;
-
-                r0 += max_jj * 6;
-                r1 += max_jj * 6;
-                r2 += max_jj * 6;
-                r3 += max_jj * 6;
-                r4 += max_jj * 6;
-                r5 += max_jj * 6;
-            }
-            for (int m = 5; m < 6; m++)
-            {
-                int tmp02a = r1[0] + r2[0];
-                int tmp02b = r3[0] + r4[0];
-                int tmp13a = r1[0] - r2[0];
-                int tmp13b = r3[0] - r4[0];
-
-                int tmp0 = tmp02a + tmp02b + r0[0];
-                int tmp1 = tmp13a + tmp13b * 2;
-                int tmp2 = tmp02a + tmp02b * 4;
-                int tmp3 = tmp13a + tmp13b * 8 + r5[0] * 4;
-
-                tmp0 = tmp0 * 4;
-                tmp1 = tmp1 * 4;
-                tmp2 = tmp2 * 4;
-                tmp3 = tmp3 * 4;
-
-                tmp[0][m] = tmp0;
-                tmp[1][m] = tmp1;
-                tmp[2][m] = tmp2;
-                tmp[3][m] = tmp3;
-
-                r0 += max_jj * 6;
-                r1 += max_jj * 6;
-                r2 += max_jj * 6;
-                r3 += max_jj * 6;
-                r4 += max_jj * 6;
-                r5 += max_jj * 6;
-            }
-
-            int* outptr0 = top_blob.channel(i + ii).row<int>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                int tmp02a = tmp[m][1] + tmp[m][2];
-                int tmp02b = tmp[m][3] + tmp[m][4];
-                int tmp13a = tmp[m][1] - tmp[m][2];
-                int tmp13b = tmp[m][3] - tmp[m][4];
-
-                int tmp0 = tmp02a + tmp02b + tmp[m][0];
-                int tmp1 = tmp13a + tmp13b * 2;
-                int tmp2 = tmp02a + tmp02b * 4;
-                int tmp3 = tmp13a + tmp13b * 8 + tmp[m][5];
-
-                tmp0 = tmp0 / 576;
-                tmp1 = tmp1 / 576;
-                tmp2 = tmp2 / 576;
-                tmp3 = tmp3 / 576;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = tmp0;
-                    if (tj * 4 + 1 < outw) outptr0[1] = tmp1;
-                    if (tj * 4 + 2 < outw) outptr0[2] = tmp2;
-                    if (tj * 4 + 3 < outw) outptr0[3] = tmp3;
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_int8(Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 4n+2, winograd F(4,3)
-    int w_tiles = (outw + 3) / 4;
-    int h_tiles = (outh + 3) / 4;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 36;
-
-    // NCNN_LOGE("conv3x3s1_winograd43_int8 %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    get_optimal_tile_mnk_int8(M, N, K, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 2u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 2u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd43_transform_input_tile_int8(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            transpose_pack_B_tile_int8(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 2u, opt.workspace_allocator);
-
-        #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd43_transform_input_tile_int8(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            transpose_pack_B_tile_int8(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    bottom_blob.release();
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                gemm_transB_packed_tile_int8(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk);
-            }
-
-            // transform output
-            conv3x3s1_winograd43_transform_output_tile_int8(top_tile, top_blob, i, max_ii, j, max_jj);
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_4x4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_4x4.h
deleted file mode 100644
index 93507cca..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_4x4.h
+++ /dev/null
@@ -1,400 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv4x4s4_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 4 * outw + w * 3;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch * 16 + q * 16;
-
-            const float* r0 = img0;
-            const float* r1 = img0 + w;
-            const float* r2 = img0 + w * 2;
-            const float* r3 = img0 + w * 3;
-
-#if __ARM_NEON
-            float32x4_t _k0123 = vld1q_f32(kernel0);
-            float32x4_t _k4567 = vld1q_f32(kernel0 + 4);
-            float32x4_t _k891011 = vld1q_f32(kernel0 + 8);
-            float32x4_t _k12131415 = vld1q_f32(kernel0 + 12);
-#else
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 4;
-            const float* k2 = kernel0 + 8;
-            const float* k3 = kernel0 + 12;
-#endif // __ARM_NEON
-
-            for (int i = 0; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw - (nn << 2);
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "prfm       pldl1keep, [%1, #128]          \n"
-                        "0:                                        \n"
-
-                        "prfm       pldl1keep, [%2, #512]          \n"
-                        "prfm       pldl1keep, [%3, #512]          \n"
-
-                        "ld1        {v7.4s}, [%1]                  \n" // v7 = outptr
-
-                        "ld1        {v8.4s}, [%2], #16             \n" // v8  = r0
-                        "ld1        {v9.4s}, [%3], #16             \n" // v9  = r1
-
-                        "prfm       pldl1keep, [%4, #512]          \n"
-                        "prfm       pldl1keep, [%5, #512]          \n"
-
-                        "fmul       v12.4s, v8.4s, %12.4s          \n"
-                        "fmul       v13.4s, v9.4s, %13.4s          \n"
-
-                        "ld1        {v10.4s}, [%4], #16            \n" // v10 = r2
-                        "ld1        {v11.4s}, [%5], #16            \n" // v11 = r3
-
-                        "fmla       v12.4s, v10.4s, %14.4s         \n"
-                        "fmla       v13.4s, v11.4s, %15.4s         \n"
-
-                        "fadd       v5.4s, v12.4s, v13.4s          \n"
-
-                        "ld1        {v8.4s}, [%2], #16             \n" // v8  = r0
-                        "ld1        {v9.4s}, [%3], #16             \n" // v9  = r1
-
-                        "fmul       v12.4s, v8.4s, %12.4s          \n"
-                        "fmul       v13.4s, v9.4s, %13.4s          \n"
-
-                        "ld1        {v10.4s}, [%4], #16            \n" // v10 = r2
-                        "ld1        {v11.4s}, [%5], #16            \n" // v11 = r3
-
-                        "fmla       v12.4s, v10.4s, %14.4s         \n"
-                        "fmla       v13.4s, v11.4s, %15.4s         \n"
-
-                        "fadd       v6.4s, v12.4s, v13.4s          \n"
-
-                        "ld1        {v8.4s}, [%2], #16             \n" // v8  = r0
-                        "ld1        {v9.4s}, [%3], #16             \n" // v9  = r1
-
-                        "fmul       v12.4s, v8.4s, %12.4s          \n"
-                        "fmul       v13.4s, v9.4s, %13.4s          \n"
-
-                        "ld1        {v10.4s}, [%4], #16            \n" // v10 = r2
-                        "ld1        {v11.4s}, [%5], #16            \n" // v11 = r3
-
-                        "fmla       v12.4s, v10.4s, %14.4s         \n"
-                        "fmla       v13.4s, v11.4s, %15.4s         \n"
-
-                        "fadd       v14.4s, v12.4s, v13.4s         \n"
-                        "faddp      v5.4s, v5.4s, v6.4s            \n" // Move to here to enhance ILP
-
-                        "ld1        {v8.4s}, [%2], #16             \n" // v8  = r0
-                        "ld1        {v9.4s}, [%3], #16             \n" // v9  = r1
-
-                        "fmul       v12.4s, v8.4s, %12.4s          \n"
-                        "fmul       v13.4s, v9.4s, %13.4s          \n"
-
-                        "ld1        {v10.4s}, [%4], #16            \n" // v10 = r2
-                        "ld1        {v11.4s}, [%5], #16            \n" // v11 = r3
-
-                        "fmla       v12.4s, v10.4s, %14.4s         \n"
-                        "fmla       v13.4s, v11.4s, %15.4s         \n"
-
-                        "fadd       v15.4s, v12.4s, v13.4s         \n"
-
-                        //                  "faddp      v5.4s ,  v5.4s,  v6.4s         \n"  // Move this line upward.
-                        "faddp      v14.4s, v14.4s, v15.4s         \n"
-                        "faddp      v5.4s ,  v5.4s, v14.4s         \n"
-
-                        "fadd       v7.4s, v7.4s, v5.4s            \n"
-
-                        "st1        {v7.4s}, [%1], #16             \n"
-
-                        "prfm       pldl1keep, [%1, #128]          \n"
-
-                        "subs       %w0, %w0, #1                   \n"
-                        "bne        0b                             \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2),     // %4
-                        "=r"(r3)      // %5
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "w"(_k0123),    // %12
-                        "w"(_k4567),    // %13
-                        "w"(_k891011),  // %14
-                        "w"(_k12131415) // %15
-                        : "cc", "memory", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-
-                        "pld        [%1, #128]          \n"
-
-                        "0:                             \n"
-
-                        "pld        [%2, #512]          \n"
-                        "pld        [%3, #512]          \n"
-
-                        "vld1.f32   {d14-d15}, [%1]     \n" // q7 = outptr
-
-                        "vld1.f32   {d16-d17}, [%2]!    \n" // q8  = r0
-                        "vld1.f32   {d18-d19}, [%3]!    \n" // q9  = r1
-
-                        "pld        [%4, #512]          \n"
-                        "pld        [%5, #512]          \n"
-
-                        "vmul.f32   q12, q8, %q12       \n"
-                        "vmul.f32   q13, q9, %q13       \n"
-
-                        "vld1.f32   {d20-d21}, [%4]!    \n" // q10 = r2
-                        "vld1.f32   {d22-d23}, [%5]!    \n" // q11 = r3
-
-                        "vmla.f32   q12, q10, %q14      \n"
-                        "vmla.f32   q13, q11, %q15      \n"
-
-                        "vadd.f32   q5, q12, q13        \n"
-
-                        "vld1.f32   {d16-d17}, [%2]!    \n" // q8  = r0
-                        "vld1.f32   {d18-d19}, [%3]!    \n" // q9  = r1
-
-                        "vmul.f32   q12, q8, %q12       \n"
-                        "vmul.f32   q13, q9, %q13       \n"
-
-                        "vld1.f32   {d20-d21}, [%4]!    \n" // q10 = r2
-                        "vld1.f32   {d22-d23}, [%5]!    \n" // q11 = r3
-
-                        "vmla.f32   q12, q10, %q14      \n"
-                        "vmla.f32   q13, q11, %q15      \n"
-
-                        "vadd.f32   q6, q12, q13        \n"
-
-                        "vld1.f32   {d16-d17}, [%2]!    \n" // q8  = r0
-                        "vld1.f32   {d18-d19}, [%3]!    \n" // q9  = r1
-
-                        "vmul.f32   q12, q8, %q12       \n"
-                        "vmul.f32   q13, q9, %q13       \n"
-
-                        "vld1.f32   {d20-d21}, [%4]!    \n" // q10 = r2
-                        "vld1.f32   {d22-d23}, [%5]!    \n" // q11 = r3
-
-                        "vmla.f32   q12, q10, %q14      \n"
-                        "vmla.f32   q13, q11, %q15      \n"
-
-                        "vadd.f32   q14, q12, q13       \n"
-
-                        "vld1.f32   {d16-d17}, [%2]!    \n" // q8  = r0
-                        "vld1.f32   {d18-d19}, [%3]!    \n" // q9  = r1
-
-                        "vmul.f32   q12, q8, %q12       \n"
-                        "vmul.f32   q13, q9, %q13       \n"
-
-                        "vld1.f32   {d20-d21}, [%4]!    \n" // q10 = r2
-                        "vld1.f32   {d22-d23}, [%5]!    \n" // q11 = r3
-
-                        "vmla.f32   q12, q10, %q14      \n"
-                        "vmla.f32   q13, q11, %q15      \n"
-
-                        "vadd.f32   q15, q12, q13       \n"
-
-                        "vadd.f32   d10, d10, d11       \n"
-                        "vadd.f32   d28, d28, d29       \n"
-                        "vadd.f32   d11, d12, d13       \n"
-                        "vadd.f32   d29, d30, d31       \n"
-
-                        "vpadd.f32  d10, d10, d11       \n"
-                        "vpadd.f32  d11, d28, d29       \n"
-
-                        "vadd.f32   q7, q7, q5          \n"
-
-                        "vst1.f32   {d14-d15}, [%1]!    \n"
-
-                        "pld        [%1, #128]          \n"
-
-                        "subs       %0, #1              \n"
-                        "bne        0b                  \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2),     // %4
-                        "=r"(r3)      // %5
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "w"(_k0123),    // %12
-                        "w"(_k4567),    // %13
-                        "w"(_k891011),  // %14
-                        "w"(_k12131415) // %15
-                        : "cc", "memory", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-#if __ARM_NEON
-#if __aarch64__
-                    float sum = 0.f;
-
-                    asm volatile(
-                        "ld1        {v8.4s}, [%0], #16             \n" // v8  = r0
-                        "ld1        {v9.4s}, [%1], #16             \n" // v9  = r1
-
-                        "fmul       v12.4s, v8.4s, %9.4s           \n"
-                        "fmul       v13.4s, v9.4s, %10.4s          \n"
-
-                        "ld1        {v10.4s}, [%2], #16            \n" // v10 = r2
-                        "ld1        {v11.4s}, [%3], #16            \n" // v11 = r3
-
-                        "fmla       v12.4s, v10.4s, %11.4s         \n"
-                        "fmla       v13.4s, v11.4s, %12.4s         \n"
-
-                        "fadd       v5.4s, v12.4s, v13.4s          \n"
-                        "faddp      v5.4s, v5.4s, v5.4s            \n"
-                        "faddp      s5, v5.2s                      \n"
-                        "fmov       %w4, s5                        \n"
-                        : "=r"(r0), // %0
-                        "=r"(r1), // %1
-                        "=r"(r2), // %2
-                        "=r"(r3), // %3
-                        "=r"(sum) // %4
-                        : "0"(r0),
-                        "1"(r1),
-                        "2"(r2),
-                        "3"(r3),
-                        "w"(_k0123),    // %9
-                        "w"(_k4567),    // %10
-                        "w"(_k891011),  // %11
-                        "w"(_k12131415) // %12
-                        : "cc", "memory", "v5", "v6", "v8", "v9", "v10", "v11", "v12", "v13");
-
-                    *outptr += sum;
-#else
-                    float sum = 0.f;
-
-                    asm volatile(
-                        "vld1.f32   {d16-d17}, [%0]!    \n" // q8  = r0
-                        "vld1.f32   {d18-d19}, [%1]!    \n" // q9  = r1
-
-                        "vmul.f32   q12, q8, %q9        \n"
-                        "vmul.f32   q13, q9, %q10       \n"
-
-                        "vld1.f32   {d20-d21}, [%2]!    \n" // q10 = r2
-                        "vld1.f32   {d22-d23}, [%3]!    \n" // q11 = r3
-
-                        "vmla.f32   q12, q10, %q11      \n"
-                        "vmla.f32   q13, q11, %q12      \n"
-
-                        "vadd.f32   q5, q12, q13        \n"
-                        "vadd.f32   d10, d10, d11       \n"
-                        "vpadd.f32  d10, d10, d10       \n"
-                        "vmov.f32   %4, d10[0]          \n"
-                        : "=r"(r0), // %0
-                        "=r"(r1), // %1
-                        "=r"(r2), // %2
-                        "=r"(r3), // %3
-                        "=r"(sum) // %4
-                        : "0"(r0),
-                        "1"(r1),
-                        "2"(r2),
-                        "3"(r3),
-                        "w"(_k0123),    // %9
-                        "w"(_k4567),    // %10
-                        "w"(_k891011),  // %11
-                        "w"(_k12131415) // %12
-                        : "cc", "memory", "q5", "q6", "q8", "q9", "q10", "q11", "q12", "q13");
-
-                    *outptr += sum;
-#endif // __aarch64__
-#else
-                    float sum = 0;
-
-                    sum += r0[0] * k0[0];
-                    sum += r0[1] * k0[1];
-                    sum += r0[2] * k0[2];
-                    sum += r0[3] * k0[3];
-
-                    sum += r1[0] * k1[0];
-                    sum += r1[1] * k1[1];
-                    sum += r1[2] * k1[2];
-                    sum += r1[3] * k1[3];
-
-                    sum += r2[0] * k2[0];
-                    sum += r2[1] * k2[1];
-                    sum += r2[2] * k2[2];
-                    sum += r2[3] * k2[3];
-
-                    sum += r3[0] * k3[0];
-                    sum += r3[1] * k3[1];
-                    sum += r3[2] * k3[2];
-                    sum += r3[3] * k3[3];
-
-                    *outptr += sum;
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                    r3 += 4;
-#endif // __ARM_NEON
-                    outptr++;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5.h
deleted file mode 100644
index 93529fe2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5.h
+++ /dev/null
@@ -1,1406 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv5x5s1_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr = out;
-            float* outptr2 = outptr + outw;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch * 25 + q * 25;
-
-            const float* r0 = img0;
-            const float* r1 = img0 + w;
-            const float* r2 = img0 + w * 2;
-            const float* r3 = img0 + w * 3;
-            const float* r4 = img0 + w * 4;
-            const float* r5 = img0 + w * 5;
-
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 5;
-            const float* k2 = kernel0 + 10;
-            const float* k3 = kernel0 + 15;
-            const float* k4 = kernel0 + 20;
-
-#if __ARM_NEON
-            float32x4_t _k0123 = vld1q_f32(kernel0);
-            float32x4_t _k4567 = vld1q_f32(kernel0 + 4);
-            float32x4_t _k891011 = vld1q_f32(kernel0 + 8);
-            float32x4_t _k12131415 = vld1q_f32(kernel0 + 12);
-            float32x4_t _k16171819 = vld1q_f32(kernel0 + 16);
-            float32x4_t _k20212223 = vld1q_f32(kernel0 + 20);
-            float32x4_t _k24242424 = vdupq_n_f32(kernel0[24]);
-#endif // __ARM_NEON
-
-            int i = 0;
-
-            for (; i + 1 < outh; i += 2)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw - (nn << 2);
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        // v11 = rx1 / rx3
-                        // v12 = rx2
-                        // v13 v14 = intermediate sum register
-
-                        "prfm       pldl1keep, [%1, #128]          \n"
-                        "ld1        {v7.4s}, [%1]                  \n" // v7 = out
-
-                        "0:                                        \n"
-
-                        "prfm       pldl1keep, [%2, #128]          \n"
-                        "ld1        {v8.4s}, [%2]                  \n" // v8 = out2
-
-                        // r1
-                        "prfm       pldl1keep, [%4, #256]          \n"
-                        "ld1        {v9.4s, v10.4s}, [%4]          \n" // v9 v10 = r10 r14
-                        "add        %4, %4, #16                    \n"
-
-                        "ext        v11.16b, v9.16b, v10.16b, #4   \n" //r11
-                        "fmul       v13.4s, v9.4s, %19.s[1]        \n"
-                        "fmla       v8.4s,  v9.4s, %18.s[0]        \n"
-
-                        "ext        v12.16b, v9.16b, v10.16b, #8   \n" //r12
-                        "fmla       v7.4s,  v11.4s, %19.s[2]       \n"
-                        "fmul       v14.4s, v11.4s, %18.s[1]       \n"
-
-                        "ext        v11.16b, v9.16b, v10.16b, #12  \n" //r13
-                        "fmla       v13.4s, v12.4s, %19.s[3]       \n"
-                        "fmla       v8.4s,  v12.4s, %18.s[2]       \n"
-
-                        "fmla       v7.4s,  v11.4s, %20.s[0]       \n"
-                        "fmla       v14.4s, v11.4s, %18.s[3]       \n"
-
-                        "prfm       pldl1keep, [%5, #256]          \n"
-
-                        "fmla       v13.4s, v10.4s, %20.s[1]       \n"
-                        "fmla       v8.4s,  v10.4s, %19.s[0]       \n"
-
-                        // r2
-                        "ld1        {v9.4s, v10.4s}, [%5]          \n" // v9 v10 = r20 r24
-                        "add        %5, %5, #16                    \n"
-
-                        "ext        v11.16b, v9.16b, v10.16b, #4   \n" //r21
-                        "fmla       v7.4s,  v9.4s, %20.s[2]        \n"
-                        "fmla       v14.4s, v9.4s, %19.s[1]        \n"
-
-                        "ext        v12.16b, v9.16b, v10.16b, #8   \n" //r22
-                        "fmla       v13.4s, v11.4s, %20.s[3]       \n"
-                        "fmla       v8.4s,  v11.4s, %19.s[2]       \n"
-
-                        "ext        v11.16b, v9.16b, v10.16b, #12  \n" //r23
-                        "fmla       v7.4s,  v12.4s, %21.s[0]       \n"
-                        "fmla       v14.4s, v12.4s, %19.s[3]       \n"
-
-                        "fmla       v13.4s, v11.4s, %21.s[1]       \n"
-                        "fmla       v8.4s,  v11.4s, %20.s[0]       \n"
-
-                        "prfm       pldl1keep, [%6, #256]          \n"
-
-                        "fmla       v7.4s,  v10.4s, %21.s[2]       \n"
-                        "fmla       v14.4s, v10.4s, %20.s[1]       \n"
-
-                        // r3
-                        "ld1        {v9.4s, v10.4s}, [%6]          \n" // v9 v10 = r30 r34
-                        "add        %6, %6, #16                    \n"
-
-                        "ext        v11.16b, v9.16b, v10.16b, #4   \n" //r31
-                        "fmla       v13.4s, v9.4s, %21.s[3]        \n"
-                        "fmla       v8.4s,  v9.4s, %20.s[2]        \n"
-
-                        "ext        v12.16b, v9.16b, v10.16b, #8   \n" //r32
-                        "fmla       v7.4s,  v11.4s, %22.s[0]       \n"
-                        "fmla       v14.4s, v11.4s, %20.s[3]       \n"
-
-                        "ext        v11.16b, v9.16b, v10.16b, #12  \n" //r33
-                        "fmla       v13.4s, v12.4s, %22.s[1]       \n"
-                        "fmla       v8.4s,  v12.4s, %21.s[0]       \n"
-
-                        "fmla       v7.4s,  v11.4s, %22.s[2]       \n"
-                        "fmla       v14.4s, v11.4s, %21.s[1]       \n"
-
-                        "prfm       pldl1keep, [%7, #256]          \n"
-
-                        "fmla       v13.4s, v10.4s, %22.s[3]       \n"
-                        "fmla       v8.4s,  v10.4s, %21.s[2]       \n"
-
-                        // r4
-                        "ld1        {v9.4s, v10.4s}, [%7]          \n" // v9 v10 = r40 r44
-                        "add        %7, %7, #16                    \n"
-
-                        "ext        v11.16b, v9.16b, v10.16b, #4   \n" //r41
-                        "fmla       v7.4s,  v9.4s, %23.s[0]        \n"
-                        "fmla       v14.4s, v9.4s, %21.s[3]        \n"
-
-                        "ext        v12.16b, v9.16b, v10.16b, #8   \n" //r41
-                        "fmla       v13.4s, v11.4s, %23.s[1]       \n"
-                        "fmla       v8.4s,  v11.4s, %22.s[0]       \n"
-
-                        "ext        v11.16b, v9.16b, v10.16b, #12  \n" //r41
-                        "fmla       v7.4s,  v12.4s, %23.s[2]       \n"
-                        "fmla       v14.4s, v12.4s, %22.s[1]       \n"
-
-                        "fmla       v13.4s, v11.4s, %23.s[3]       \n"
-                        "fmla       v8.4s,  v11.4s, %22.s[2]       \n"
-
-                        "prfm       pldl1keep, [%3, #256]          \n"
-
-                        "fmla       v7.4s,  v10.4s, %24.s[0]       \n"
-                        "fmla       v14.4s, v10.4s, %22.s[3]       \n"
-
-                        // r0 and r5
-                        "ld1        {v9.4s, v10.4s}, [%3]          \n" // v9 v10 = r00 r04
-                        "add        %3, %3, #16                    \n"
-
-                        "ext        v11.16b, v9.16b, v10.16b, #4   \n" //r01
-                        "fmla       v13.4s, v11.4s, %18.s[1]       \n"
-
-                        "ext        v12.16b, v9.16b, v10.16b, #8   \n" //r02
-                        "fmla       v7.4s, v12.4s, %18.s[2]        \n"
-
-                        "ext        v11.16b, v9.16b, v10.16b, #12  \n" //r03
-
-                        "prfm       pldl1keep, [%8, #256]          \n"
-
-                        "fmla       v13.4s, v11.4s, %18.s[3]       \n"
-
-                        // r5
-                        "ld1        {v11.4s, v12.4s}, [%8]         \n" // v11 v12 = r50 r54
-                        "add        %8, %8, #16                    \n"
-
-                        "fmla       v8.4s,  v11.4s, %23.s[0]       \n"
-                        "fmla       v14.4s, v12.4s, %24.s[0]       \n"
-
-                        "fmla       v7.4s,  v9.4s,  %18.s[0]       \n"
-                        "fmla       v13.4s, v10.4s, %19.s[0]       \n"
-
-                        "ext        v9.16b,  v11.16b, v12.16b, #4  \n" //r51
-                        "ext        v10.16b, v11.16b, v12.16b, #8  \n" //r52
-
-                        "fmla       v14.4s, v9.4s, %23.s[1]        \n"
-
-                        "ext        v9.16b, v11.16b, v12.16b, #12  \n" //r53
-                        "fmla       v8.4s, v10.4s, %23.s[2]        \n"
-
-                        "fmla       v14.4s, v9.4s, %23.s[3]        \n"
-
-                        "fadd       v7.4s, v7.4s, v13.4s           \n"
-
-                        "st1        {v7.4s}, [%1], #16             \n"
-
-                        "fadd       v8.4s, v8.4s, v14.4s           \n"
-
-                        "prfm       pldl1keep, [%1, #128]          \n"
-                        "ld1        {v7.4s}, [%1]                  \n" // v7 = out
-                        "st1        {v8.4s}, [%2], #16             \n"
-
-                        "subs       %w0, %w0, #1                   \n"
-                        "bne        0b                             \n"
-                        : "=r"(nn),      // %0
-                        "=r"(outptr),  // %1
-                        "=r"(outptr2), // %2
-                        "=r"(r0),      // %3
-                        "=r"(r1),      // %4
-                        "=r"(r2),      // %5
-                        "=r"(r3),      // %6
-                        "=r"(r4),      // %7
-                        "=r"(r5)       // %8
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(outptr2),
-                        "3"(r0),
-                        "4"(r1),
-                        "5"(r2),
-                        "6"(r3),
-                        "7"(r4),
-                        "8"(r5),
-                        "w"(_k0123),     // %18
-                        "w"(_k4567),     // %19
-                        "w"(_k891011),   // %20
-                        "w"(_k12131415), // %21
-                        "w"(_k16171819), // %22
-                        "w"(_k20212223), // %23
-                        "w"(_k24242424)  // %24
-                        : "cc", "memory", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        //                     "veor       q13, q13            \n"
-                        //                     "veor       q14, q14            \n"
-
-                        "pld        [%1, #128]          \n"
-
-                        "vld1.f32   {d14-d15}, [%1]     \n" // q7 = out
-
-                        "0:                             \n"
-
-                        // q11 = rx1 / rx3
-                        // q12 = rx2
-
-                        // q13 q14 = intermediate sum register
-
-                        "pld        [%2, #128]          \n"
-
-                        "vld1.f32   {d16-d17}, [%2]     \n" // q8 = out2
-
-                        "pld        [%4, #256]          \n"
-
-                        // r1
-                        "vld1.f32   {d18-d21}, [%4]     \n" // q9 q10 = r10 r14
-                        "add        %4, #16             \n"
-
-                        "vext.32    q11, q9, q10, #1    \n" // r11
-                        "vmul.f32   q13, q9, %e19[1]    \n"
-                        "vmla.f32   q8, q9, %e18[0]     \n"
-
-                        "vext.32    q12, q9, q10, #2    \n" // r12
-                        "vmla.f32   q7, q11, %f19[0]    \n"
-                        "vmul.f32   q14, q11, %e18[1]   \n"
-
-                        "vext.32    q11, q9, q10, #3    \n" // r13
-                        "vmla.f32   q13, q12, %f19[1]   \n"
-                        "vmla.f32   q8, q12, %f18[0]    \n"
-
-                        "vmla.f32   q7, q11, %e20[0]    \n"
-                        "vmla.f32   q14, q11, %f18[1]   \n"
-
-                        "pld        [%5, #256]          \n"
-
-                        "vmla.f32   q13, q10, %e20[1]   \n"
-                        "vmla.f32   q8, q10, %e19[0]    \n"
-
-                        // r2
-                        "vld1.f32   {d18-d21}, [%5]     \n" // q9 q10 = r20 r24
-                        "add        %5, #16             \n"
-
-                        "vext.32    q11, q9, q10, #1    \n" // r21
-                        "vmla.f32   q7, q9, %f20[0]     \n"
-                        "vmla.f32   q14, q9, %e19[1]    \n"
-
-                        "vext.32    q12, q9, q10, #2    \n" // r22
-                        "vmla.f32   q13, q11, %f20[1]   \n"
-                        "vmla.f32   q8, q11, %f19[0]    \n"
-
-                        "vext.32    q11, q9, q10, #3    \n" // r23
-                        "vmla.f32   q7, q12, %e21[0]    \n"
-                        "vmla.f32   q14, q12, %f19[1]   \n"
-
-                        "vmla.f32   q13, q11, %e21[1]   \n"
-                        "vmla.f32   q8, q11, %e20[0]    \n"
-
-                        "pld        [%6, #256]          \n"
-
-                        "vmla.f32   q7, q10, %f21[0]    \n"
-                        "vmla.f32   q14, q10, %e20[1]   \n"
-
-                        // r3
-                        "vld1.f32   {d18-d21}, [%6]     \n" // q9 q10 = r30 r34
-                        "add        %6, #16             \n"
-
-                        "vext.32    q11, q9, q10, #1    \n" // r31
-                        "vmla.f32   q13, q9, %f21[1]    \n"
-                        "vmla.f32   q8, q9, %f20[0]     \n"
-
-                        "vext.32    q12, q9, q10, #2    \n" // r32
-                        "vmla.f32   q7, q11, %e22[0]    \n"
-                        "vmla.f32   q14, q11, %f20[1]   \n"
-
-                        "vext.32    q11, q9, q10, #3    \n" // r33
-                        "vmla.f32   q13, q12, %e22[1]   \n"
-                        "vmla.f32   q8, q12, %e21[0]    \n"
-
-                        "vmla.f32   q7, q11, %f22[0]    \n"
-                        "vmla.f32   q14, q11, %e21[1]   \n"
-
-                        "pld        [%7, #256]          \n"
-
-                        "vmla.f32   q13, q10, %f22[1]   \n"
-                        "vmla.f32   q8, q10, %f21[0]    \n"
-
-                        // r4
-                        "vld1.f32   {d18-d21}, [%7]     \n" // q9 q10 = r40 r44
-                        "add        %7, #16             \n"
-
-                        "vext.32    q11, q9, q10, #1    \n" // r41
-                        "vmla.f32   q7, q9, %e23[0]     \n"
-                        "vmla.f32   q14, q9, %f21[1]    \n"
-
-                        "vext.32    q12, q9, q10, #2    \n" // r42
-                        "vmla.f32   q13, q11, %e23[1]   \n"
-                        "vmla.f32   q8, q11, %e22[0]    \n"
-
-                        "vext.32    q11, q9, q10, #3    \n" // r43
-                        "vmla.f32   q7, q12, %f23[0]    \n"
-                        "vmla.f32   q14, q12, %e22[1]   \n"
-
-                        "vmla.f32   q13, q11, %f23[1]   \n"
-                        "vmla.f32   q8, q11, %f22[0]    \n"
-
-                        "pld        [%3, #256]          \n"
-
-                        "vmla.f32   q7, q10, %e24[0]    \n"
-                        "vmla.f32   q14, q10, %f22[1]   \n"
-
-                        // r0 and r5
-                        "vld1.f32   {d18-d21}, [%3]     \n" // q9 q10 = r00 r04
-                        "add        %3, #16             \n"
-
-                        "vext.32    q11, q9, q10, #1    \n" // r01
-                        "vmla.f32   q13, q11, %e18[1]   \n"
-
-                        "vext.32    q12, q9, q10, #2    \n" // r02
-                        "vmla.f32   q7, q12, %f18[0]    \n"
-
-                        "vext.32    q11, q9, q10, #3    \n" // r03
-
-                        "pld        [%8, #256]          \n"
-
-                        "vmla.f32   q13, q11, %f18[1]   \n"
-
-                        // r5
-                        "vld1.f32   {d22-d25}, [%8]     \n" // q11 q12 = r50 r54
-                        "add        %8, #16             \n"
-
-                        "vmla.f32   q8, q11, %e23[0]    \n"
-                        "vmla.f32   q14, q12, %e24[0]   \n"
-
-                        "vmla.f32   q7, q9, %e18[0]     \n"
-                        "vmla.f32   q13, q10, %e19[0]   \n"
-
-                        "vext.32    q9, q11, q12, #1    \n" // r51
-                        "vext.32    q10, q11, q12, #2   \n" // r52
-
-                        "vmla.f32   q14, q9, %e23[1]    \n"
-
-                        "vext.32    q9, q11, q12, #3    \n" // r53
-                        "vmla.f32   q8, q10, %f23[0]    \n"
-
-                        "vmla.f32   q14, q9, %f23[1]    \n"
-
-                        "vadd.f32   q7, q7, q13         \n"
-
-                        //                     "veor       q13, q13            \n"
-
-                        "vst1.f32   {d14-d15}, [%1]!    \n"
-
-                        "vadd.f32   q8, q8, q14         \n"
-
-                        "pld        [%1, #128]          \n"
-
-                        "vld1.f32   {d14-d15}, [%1]     \n" // q7 = out
-
-                        //                     "veor       q14, q14            \n"
-
-                        "vst1.f32   {d16-d17}, [%2]!    \n"
-
-                        "subs       %0, #1              \n"
-                        "bne        0b                  \n"
-                        : "=r"(nn),      // %0
-                        "=r"(outptr),  // %1
-                        "=r"(outptr2), // %2
-                        "=r"(r0),      // %3
-                        "=r"(r1),      // %4
-                        "=r"(r2),      // %5
-                        "=r"(r3),      // %6
-                        "=r"(r4),      // %7
-                        "=r"(r5)       // %8
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(outptr2),
-                        "3"(r0),
-                        "4"(r1),
-                        "5"(r2),
-                        "6"(r3),
-                        "7"(r4),
-                        "8"(r5),
-                        "w"(_k0123),     // %18
-                        "w"(_k4567),     // %19
-                        "w"(_k891011),   // %20
-                        "w"(_k12131415), // %21
-                        "w"(_k16171819), // %22
-                        "w"(_k20212223), // %23
-                        "w"(_k24242424)  // %24
-                        : "cc", "memory", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-                    float sum = 0;
-                    float sum2 = 0;
-#if __ARM_NEON
-                    float32x4_t _r1 = vld1q_f32(r1);
-                    float32x4_t _k1 = vld1q_f32(k1);
-                    float32x4_t _sum = vmulq_f32(_r1, _k1);
-                    float32x4_t _sum2 = vmulq_f32(_r1, _k0123);
-
-                    float32x4_t _r2 = vld1q_f32(r2);
-                    float32x4_t _k2 = vld1q_f32(k2);
-                    _sum = vmlaq_f32(_sum, _r2, _k2);
-                    _sum2 = vmlaq_f32(_sum2, _r2, _k1);
-
-                    float32x4_t _r3 = vld1q_f32(r3);
-                    float32x4_t _k3 = vld1q_f32(k3);
-                    _sum = vmlaq_f32(_sum, _r3, _k3);
-                    _sum2 = vmlaq_f32(_sum2, _r3, _k2);
-
-                    float32x4_t _r4 = vld1q_f32(r4);
-                    _sum = vmlaq_f32(_sum, _r4, _k20212223);
-                    _sum2 = vmlaq_f32(_sum2, _r4, _k3);
-
-                    float32x4_t _r0 = vld1q_f32(r0);
-                    _sum = vmlaq_f32(_sum, _r0, _k0123);
-                    float32x4_t _r5 = vld1q_f32(r5);
-                    _sum2 = vmlaq_f32(_sum2, _r5, _k20212223);
-
-                    float32x4_t _k_t4 = {};
-
-                    _k_t4 = vsetq_lane_f32(k0[4], _k_t4, 0);
-                    _k_t4 = vsetq_lane_f32(k1[4], _k_t4, 1);
-                    _k_t4 = vsetq_lane_f32(k2[4], _k_t4, 2);
-                    _k_t4 = vsetq_lane_f32(k3[4], _k_t4, 3);
-
-                    float32x4_t _r_t4 = {};
-
-                    _r_t4 = vsetq_lane_f32(r0[4], _r_t4, 0);
-                    _r_t4 = vsetq_lane_f32(r1[4], _r_t4, 1);
-                    _r_t4 = vsetq_lane_f32(r2[4], _r_t4, 2);
-                    _r_t4 = vsetq_lane_f32(r3[4], _r_t4, 3);
-                    _sum = vmlaq_f32(_sum, _r_t4, _k_t4);
-
-                    sum = r4[4] * k4[4];
-
-                    _r_t4 = vextq_f32(_r_t4, _r_t4, 1);
-                    _r_t4 = vsetq_lane_f32(r4[4], _r_t4, 3);
-                    _sum2 = vmlaq_f32(_sum2, _r_t4, _k_t4);
-
-                    sum2 = r5[4] * k4[4];
-
-                    float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-                    float32x2_t _ss2 = vadd_f32(vget_low_f32(_sum2), vget_high_f32(_sum2));
-                    float32x2_t _ss_ss2 = vpadd_f32(_ss, _ss2);
-
-                    sum += vget_lane_f32(_ss_ss2, 0);
-                    sum2 += vget_lane_f32(_ss_ss2, 1);
-#else
-                    sum += r0[0] * k0[0];
-                    sum += r0[1] * k0[1];
-                    sum += r0[2] * k0[2];
-                    sum += r0[3] * k0[3];
-                    sum += r0[4] * k0[4];
-
-                    sum += r1[0] * k1[0];
-                    sum += r1[1] * k1[1];
-                    sum += r1[2] * k1[2];
-                    sum += r1[3] * k1[3];
-                    sum += r1[4] * k1[4];
-
-                    sum += r2[0] * k2[0];
-                    sum += r2[1] * k2[1];
-                    sum += r2[2] * k2[2];
-                    sum += r2[3] * k2[3];
-                    sum += r2[4] * k2[4];
-
-                    sum += r3[0] * k3[0];
-                    sum += r3[1] * k3[1];
-                    sum += r3[2] * k3[2];
-                    sum += r3[3] * k3[3];
-                    sum += r3[4] * k3[4];
-
-                    sum += r4[0] * k4[0];
-                    sum += r4[1] * k4[1];
-                    sum += r4[2] * k4[2];
-                    sum += r4[3] * k4[3];
-                    sum += r4[4] * k4[4];
-
-                    sum2 += r1[0] * k0[0];
-                    sum2 += r1[1] * k0[1];
-                    sum2 += r1[2] * k0[2];
-                    sum2 += r1[3] * k0[3];
-                    sum2 += r1[4] * k0[4];
-
-                    sum2 += r2[0] * k1[0];
-                    sum2 += r2[1] * k1[1];
-                    sum2 += r2[2] * k1[2];
-                    sum2 += r2[3] * k1[3];
-                    sum2 += r2[4] * k1[4];
-
-                    sum2 += r3[0] * k2[0];
-                    sum2 += r3[1] * k2[1];
-                    sum2 += r3[2] * k2[2];
-                    sum2 += r3[3] * k2[3];
-                    sum2 += r3[4] * k2[4];
-
-                    sum2 += r4[0] * k3[0];
-                    sum2 += r4[1] * k3[1];
-                    sum2 += r4[2] * k3[2];
-                    sum2 += r4[3] * k3[3];
-                    sum2 += r4[4] * k3[4];
-
-                    sum2 += r5[0] * k4[0];
-                    sum2 += r5[1] * k4[1];
-                    sum2 += r5[2] * k4[2];
-                    sum2 += r5[3] * k4[3];
-                    sum2 += r5[4] * k4[4];
-#endif // __ARM_NEON
-                    *outptr += sum;
-                    *outptr2 += sum2;
-
-                    r0++;
-                    r1++;
-                    r2++;
-                    r3++;
-                    r4++;
-                    r5++;
-                    outptr++;
-                    outptr2++;
-                }
-
-                r0 += 4 + w;
-                r1 += 4 + w;
-                r2 += 4 + w;
-                r3 += 4 + w;
-                r4 += 4 + w;
-                r5 += 4 + w;
-
-                outptr += outw;
-                outptr2 += outw;
-            }
-
-            for (; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw - (nn << 2);
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "prfm       pldl1keep, [%1, #128]          \n"
-                        "prfm       pldl1keep, [%2, #256]          \n"
-
-                        "ld1        {v8.4s, v9.4s}, [%2]           \n" // _r00 = vld1q_f32(r0+j);
-                        "add        %2, %2, #16                    \n"
-
-                        "0:                                        \n"
-
-                        "ld1        {v7.4s}, [%1]                  \n" // _sum = vld1q_f32(outptr+j);
-
-                        "ext        v10.16b, v8.16b, v9.16b, #4    \n" //_r01
-                        "ext        v11.16b, v8.16b, v9.16b, #8    \n" //_r02
-                        "ext        v12.16b, v8.16b, v9.16b, #12   \n" //_r03
-
-                        "fmla       v7.4s,   v8.4s, %14.s[0]       \n"
-                        "fmul       v13.4s, v10.4s, %14.s[1]       \n"
-
-                        "prfm       pldl1keep, [%3, #256]          \n"
-
-                        "fmul       v14.4s, v11.4s, %14.s[2]       \n"
-                        "fmul       v15.4s, v12.4s, %14.s[3]       \n"
-                        "fmla       v7.4s,   v9.4s, %15.s[0]       \n"
-
-                        "ld1        {v8.4s, v9.4s}, [%3]           \n"
-                        "add        %3, %3, #16                    \n"
-                        "ext        v10.16b, v8.16b, v9.16b, #4    \n" //_r11
-                        "ext        v11.16b, v8.16b, v9.16b, #8    \n" //_r12
-                        "ext        v12.16b, v8.16b, v9.16b, #12   \n" //_r13
-
-                        "fmla       v7.4s,   v8.4s, %15.s[1]       \n"
-                        "fmla       v13.4s, v10.4s, %15.s[2]       \n"
-
-                        "prfm       pldl1keep, [%4, #256]          \n"
-
-                        "fmla       v14.4s, v11.4s, %15.s[3]       \n"
-                        "fmla       v15.4s, v12.4s, %16.s[0]       \n"
-                        "fmla       v7.4s,   v9.4s, %16.s[1]       \n"
-
-                        "ld1        {v8.4s, v9.4s}, [%4]           \n"
-                        "add        %4, %4, #16                    \n"
-                        "ext        v10.16b, v8.16b, v9.16b, #4    \n" //_r21
-                        "ext        v11.16b, v8.16b, v9.16b, #8    \n" //_r22
-                        "ext        v12.16b, v8.16b, v9.16b, #12   \n" //_r23
-
-                        "fmla       v7.4s,   v8.4s, %16.s[2]       \n"
-                        "fmla       v13.4s, v10.4s, %16.s[3]       \n"
-
-                        "prfm       pldl1keep, [%5, #256]          \n"
-
-                        "fmla       v14.4s, v11.4s, %17.s[0]       \n"
-                        "fmla       v15.4s, v12.4s, %17.s[1]       \n"
-                        "fmla       v7.4s,   v9.4s, %17.s[2]       \n"
-
-                        "ld1        {v8.4s, v9.4s}, [%5]           \n"
-                        "add        %5, %5, #16                    \n"
-                        "ext        v10.16b, v8.16b, v9.16b, #4    \n" //_r31
-                        "ext        v11.16b, v8.16b, v9.16b, #8    \n" //_r32
-                        "ext        v12.16b, v8.16b, v9.16b, #12   \n" //_r33
-
-                        "fmla       v7.4s,   v8.4s, %17.s[3]       \n"
-                        "fmla       v13.4s, v10.4s, %18.s[0]       \n"
-
-                        "prfm       pldl1keep, [%6, #256]          \n"
-
-                        "fmla       v14.4s, v11.4s, %18.s[1]       \n"
-                        "fmla       v15.4s, v12.4s, %18.s[2]       \n"
-                        "fmla       v7.4s,   v9.4s, %18.s[3]       \n"
-
-                        "ld1        {v8.4s, v9.4s}, [%6]           \n"
-                        "add        %6, %6, #16                    \n"
-                        "ext        v10.16b, v8.16b, v9.16b, #4    \n" //_r41
-                        "ext        v11.16b, v8.16b, v9.16b, #8    \n" //_r42
-                        "ext        v12.16b, v8.16b, v9.16b, #12   \n" //_r43
-
-                        "fmla       v7.4s,   v8.4s, %19.s[0]       \n"
-                        "fmla       v13.4s, v10.4s, %19.s[1]       \n"
-                        "fmla       v14.4s, v11.4s, %19.s[2]       \n"
-                        "fmla       v15.4s, v12.4s, %19.s[3]       \n"
-                        "fmla       v7.4s,   v9.4s, %20.s[0]       \n"
-
-                        "fadd       v14.4s, v14.4s, v15.4s         \n"
-                        "fadd       v7.4s,   v7.4s, v13.4s         \n"
-
-                        "prfm       pldl1keep, [%2, #256]          \n"
-
-                        "fadd       v7.4s,   v7.4s, v14.4s         \n"
-
-                        "ld1        {v8.4s, v9.4s}, [%2]           \n"
-                        "add        %2, %2, #16                    \n"
-
-                        "st1        {v7.4s}, [%1], #16             \n"
-
-                        "prfm       pldl1keep, [%1, #128]          \n"
-
-                        "subs       %w0, %w0, #1                   \n"
-                        "bne        0b                             \n"
-
-                        "sub        %2, %2, #16                    \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2),     // %4
-                        "=r"(r3),     // %5
-                        "=r"(r4)      // %6
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "w"(_k0123),     // %14
-                        "w"(_k4567),     // %15
-                        "w"(_k891011),   // %16
-                        "w"(_k12131415), // %17
-                        "w"(_k16171819), // %18
-                        "w"(_k20212223), // %19
-                        "w"(_k24242424)  // %20
-                        : "cc", "memory", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                }
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        //                     "veor       q15, q15            \n"// _sum3 = 0;
-
-                        "pld        [%1, #128]          \n"
-
-                        "pld        [%2, #256]          \n"
-
-                        "vld1.f32   {d16-d19}, [%2]     \n" // _r00 = vld1q_f32(r0+j);
-                        "add        %2, #16             \n"
-
-                        "0:                             \n"
-
-                        "vld1.f32   {d14-d15}, [%1]     \n" // _sum = vld1q_f32(outptr+j);
-                        //                     "veor       q13, q13            \n"// _sum2 = 0;
-                        //                     "veor       q14, q14            \n"// _sum3 = 0;
-
-                        "vext.32    q10, q8, q9, #1     \n" // _r01
-                        "vext.32    q11, q8, q9, #2     \n" // _r02
-                        "vext.32    q12, q8, q9, #3     \n" // _r03
-
-                        "vmla.f32   q7, q8, %e14[0]     \n"
-                        "vmul.f32   q13, q10, %e14[1]   \n"
-
-                        "pld        [%3, #256]          \n"
-
-                        "vmul.f32   q14, q11, %f14[0]   \n"
-                        "vmul.f32   q15, q12, %f14[1]   \n"
-                        "vmla.f32   q7, q9, %e15[0]     \n"
-
-                        "vld1.f32   {d16-d19}, [%3]     \n"
-                        "add        %3, #16             \n"
-                        "vext.32    q10, q8, q9, #1     \n"
-                        "vext.32    q11, q8, q9, #2     \n"
-                        "vext.32    q12, q8, q9, #3     \n"
-
-                        "vmla.f32   q7, q8, %e15[1]     \n"
-                        "vmla.f32   q13, q10, %f15[0]   \n"
-
-                        "pld        [%4, #256]          \n"
-
-                        "vmla.f32   q14, q11, %f15[1]   \n"
-                        "vmla.f32   q15, q12, %e16[0]   \n"
-                        "vmla.f32   q7, q9, %e16[1]     \n"
-
-                        "vld1.f32   {d16-d19}, [%4]     \n"
-                        "add        %4, #16             \n"
-                        "vext.32    q10, q8, q9, #1     \n"
-                        "vext.32    q11, q8, q9, #2     \n"
-                        "vext.32    q12, q8, q9, #3     \n"
-
-                        "vmla.f32   q7, q8, %f16[0]     \n"
-                        "vmla.f32   q13, q10, %f16[1]   \n"
-
-                        "pld        [%5, #256]          \n"
-
-                        "vmla.f32   q14, q11, %e17[0]   \n"
-                        "vmla.f32   q15, q12, %e17[1]   \n"
-                        "vmla.f32   q7, q9, %f17[0]     \n"
-
-                        "vld1.f32   {d16-d19}, [%5]     \n"
-                        "add        %5, #16             \n"
-                        "vext.32    q10, q8, q9, #1     \n"
-                        "vext.32    q11, q8, q9, #2     \n"
-                        "vext.32    q12, q8, q9, #3     \n"
-
-                        "vmla.f32   q7, q8, %f17[1]     \n"
-                        "vmla.f32   q13, q10, %e18[0]   \n"
-
-                        "pld        [%6, #256]          \n"
-
-                        "vmla.f32   q14, q11, %e18[1]   \n"
-                        "vmla.f32   q15, q12, %f18[0]   \n"
-                        "vmla.f32   q7, q9, %f18[1]     \n"
-
-                        "vld1.f32   {d16-d19}, [%6]     \n"
-                        "add        %6, #16             \n"
-                        "vext.32    q10, q8, q9, #1     \n"
-                        "vext.32    q11, q8, q9, #2     \n"
-                        "vext.32    q12, q8, q9, #3     \n"
-
-                        "vmla.f32   q7, q8, %e19[0]     \n"
-                        "vmla.f32   q13, q10, %e19[1]   \n"
-                        "vmla.f32   q14, q11, %f19[0]   \n"
-                        "vmla.f32   q15, q12, %f19[1]   \n"
-                        "vmla.f32   q7, q9, %e20[0]     \n"
-
-                        "vadd.f32   q14, q14, q15       \n"
-                        "vadd.f32   q7, q7, q13         \n"
-                        //                     "veor       q15, q15            \n"// _sum3 = 0;
-
-                        "pld        [%2, #256]          \n"
-
-                        "vadd.f32   q7, q7, q14         \n"
-
-                        "vld1.f32   {d16-d19}, [%2]     \n" // _r00 = vld1q_f32(r0+j);
-                        "add        %2, #16             \n"
-
-                        "vst1.f32   {d14-d15}, [%1]!    \n"
-
-                        "pld        [%1, #128]          \n"
-
-                        "subs       %0, #1              \n"
-                        "bne        0b                  \n"
-
-                        "sub        %2, #16             \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2),     // %4
-                        "=r"(r3),     // %5
-                        "=r"(r4)      // %6
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "w"(_k0123),     // %14
-                        "w"(_k4567),     // %15
-                        "w"(_k891011),   // %16
-                        "w"(_k12131415), // %17
-                        "w"(_k16171819), // %18
-                        "w"(_k20212223), // %19
-                        "w"(_k24242424)  // %20
-                        : "cc", "memory", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-                    float sum = 0;
-#if __ARM_NEON
-                    float32x4_t _r0 = vld1q_f32(r0);
-                    float32x4_t _sum = vmulq_f32(_r0, _k0123);
-
-                    float32x4_t _r1 = vld1q_f32(r1);
-                    _sum = vmlaq_f32(_sum, _r1, vld1q_f32(k1));
-
-                    float32x4_t _r2 = vld1q_f32(r2);
-                    _sum = vmlaq_f32(_sum, _r2, vld1q_f32(k2));
-
-                    float32x4_t _r3 = vld1q_f32(r3);
-                    _sum = vmlaq_f32(_sum, _r3, vld1q_f32(k3));
-
-                    float32x4_t _r4 = vld1q_f32(r4);
-                    _sum = vmlaq_f32(_sum, _r4, _k20212223);
-
-                    float32x4_t _k_t4 = {};
-
-                    _k_t4 = vsetq_lane_f32(k0[4], _k_t4, 0);
-                    _k_t4 = vsetq_lane_f32(k1[4], _k_t4, 1);
-                    _k_t4 = vsetq_lane_f32(k2[4], _k_t4, 2);
-                    _k_t4 = vsetq_lane_f32(k3[4], _k_t4, 3);
-
-                    float32x4_t _r_t4 = {};
-
-                    _r_t4 = vsetq_lane_f32(r0[4], _r_t4, 0);
-                    _r_t4 = vsetq_lane_f32(r1[4], _r_t4, 1);
-                    _r_t4 = vsetq_lane_f32(r2[4], _r_t4, 2);
-                    _r_t4 = vsetq_lane_f32(r3[4], _r_t4, 3);
-                    _sum = vmlaq_f32(_sum, _r_t4, _k_t4);
-
-                    sum = r4[4] * k4[4];
-
-                    float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-                    _ss = vpadd_f32(_ss, _ss);
-
-                    sum += vget_lane_f32(_ss, 0);
-#else
-                    sum += r0[0] * k0[0];
-                    sum += r0[1] * k0[1];
-                    sum += r0[2] * k0[2];
-                    sum += r0[3] * k0[3];
-                    sum += r0[4] * k0[4];
-
-                    sum += r1[0] * k1[0];
-                    sum += r1[1] * k1[1];
-                    sum += r1[2] * k1[2];
-                    sum += r1[3] * k1[3];
-                    sum += r1[4] * k1[4];
-
-                    sum += r2[0] * k2[0];
-                    sum += r2[1] * k2[1];
-                    sum += r2[2] * k2[2];
-                    sum += r2[3] * k2[3];
-                    sum += r2[4] * k2[4];
-
-                    sum += r3[0] * k3[0];
-                    sum += r3[1] * k3[1];
-                    sum += r3[2] * k3[2];
-                    sum += r3[3] * k3[3];
-                    sum += r3[4] * k3[4];
-
-                    sum += r4[0] * k4[0];
-                    sum += r4[1] * k4[1];
-                    sum += r4[2] * k4[2];
-                    sum += r4[3] * k4[3];
-                    sum += r4[4] * k4[4];
-#endif
-                    *outptr += sum;
-
-                    r0++;
-                    r1++;
-                    r2++;
-                    r3++;
-                    r4++;
-                    outptr++;
-                }
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                r3 += 4;
-                r4 += 4;
-            }
-        }
-    }
-}
-
-static void conv5x5s2_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch * 25 + q * 25;
-
-            const float* r0 = img0;
-            const float* r1 = img0 + w;
-            const float* r2 = img0 + w * 2;
-            const float* r3 = img0 + w * 3;
-            const float* r4 = img0 + w * 4;
-
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 5;
-            const float* k2 = kernel0 + 10;
-            const float* k3 = kernel0 + 15;
-            const float* k4 = kernel0 + 20;
-
-#if __ARM_NEON
-            float32x4_t _k0123 = vld1q_f32(kernel0);
-            float32x4_t _k4567 = vld1q_f32(kernel0 + 4);
-            float32x4_t _k891011 = vld1q_f32(kernel0 + 8);
-            float32x4_t _k12131415 = vld1q_f32(kernel0 + 12);
-            float32x4_t _k16171819 = vld1q_f32(kernel0 + 16);
-            float32x4_t _k20212223 = vld1q_f32(kernel0 + 20);
-            float32x4_t _k24242424 = vdupq_n_f32(kernel0[24]);
-#endif // __ARM_NEON
-
-            for (int i = 0; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw - (nn << 2);
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "prfm       pldl1keep, [%2, #256]          \n"
-                        "ld2        {v8.4s, v9.4s}, [%2], #32      \n" // v8  = 0  2  4  6   q9  = 1  3  5  7
-
-                        "prfm       pldl1keep, [%2, #256]          \n"
-                        "ld2        {v10.4s, v11.4s}, [%2]         \n" // v10 = 8 10 12 14   v11 = 9 11 13 15
-
-                        "prfm       pldl1keep, [%1, #128]          \n"
-                        "0:                                        \n"
-
-                        "ld1        {v7.4s}, [%1]                  \n" // v7 = outptr
-
-                        "ext        v12.16b, v8.16b, v10.16b, #4   \n" // v12 = 2 4 6 8
-                        "ext        v11.16b, v9.16b, v11.16b, #4   \n" // v11 = 3 5 7 9
-                        "ext        v10.16b, v8.16b, v10.16b, #8   \n" // v10 = 4 6 8 10
-
-                        "fmla       v7.4s,  v8.4s, %14.s[0]        \n"
-                        "fmul       v13.4s, v9.4s, %14.s[1]        \n"
-
-                        "prfm       pldl1keep, [%3, #256]          \n"
-
-                        "fmul       v14.4s, v12.4s, %14.s[2]       \n"
-                        "fmul       v15.4s, v11.4s, %14.s[3]       \n"
-                        "fmla       v7.4s,  v10.4s, %15.s[0]       \n"
-
-                        "ld2        {v8.4s, v9.4s}, [%3], #32      \n"
-
-                        "prfm       pldl1keep, [%3, #256]          \n"
-
-                        "ld2        {v10.4s, v11.4s}, [%3]         \n"
-                        "ext        v12.16b, v8.16b, v10.16b, #4   \n"
-                        "ext        v11.16b, v9.16b, v11.16b, #4   \n"
-                        "ext        v10.16b, v8.16b, v10.16b, #8   \n"
-
-                        "fmla       v7.4s,  v8.4s, %15.s[1]        \n"
-                        "fmla       v13.4s, v9.4s, %15.s[2]        \n"
-
-                        "prfm       pldl1keep, [%4, #256]          \n"
-
-                        "fmla       v14.4s, v12.4s, %15.s[3]       \n"
-                        "fmla       v15.4s, v11.4s, %16.s[0]       \n"
-                        "fmla       v7.4s,  v10.4s, %16.s[1]       \n"
-
-                        "ld2        {v8.4s, v9.4s}, [%4], #32      \n"
-
-                        "prfm       pldl1keep, [%4, #256]          \n"
-
-                        "ld2        {v10.4s, v11.4s}, [%4]         \n"
-                        "ext        v12.16b, v8.16b, v10.16b, #4   \n"
-                        "ext        v11.16b, v9.16b, v11.16b, #4   \n"
-                        "ext        v10.16b, v8.16b, v10.16b, #8   \n"
-
-                        "fmla       v7.4s,  v8.4s, %16.s[2]        \n"
-                        "fmla       v13.4s, v9.4s, %16.s[3]        \n"
-
-                        "prfm       pldl1keep, [%5, #256]          \n"
-
-                        "fmla       v14.4s, v12.4s, %17.s[0]       \n"
-                        "fmla       v15.4s, v11.4s, %17.s[1]       \n"
-                        "fmla       v7.4s,  v10.4s, %17.s[2]       \n"
-
-                        "ld2        {v8.4s, v9.4s}, [%5], #32      \n"
-
-                        "prfm       pldl1keep, [%5, #256]          \n"
-
-                        "ld2        {v10.4s, v11.4s}, [%5]         \n"
-                        "ext        v12.16b, v8.16b, v10.16b, #4   \n"
-                        "ext        v11.16b, v9.16b, v11.16b, #4   \n"
-                        "ext        v10.16b, v8.16b, v10.16b, #8   \n"
-
-                        "fmla       v7.4s,  v8.4s, %17.s[3]        \n"
-                        "fmla       v13.4s, v9.4s, %18.s[0]        \n"
-
-                        "prfm       pldl1keep, [%6, #256]          \n"
-
-                        "fmla       v14.4s, v12.4s, %18.s[1]       \n"
-                        "fmla       v15.4s, v11.4s, %18.s[2]       \n"
-                        "fmla       v7.4s,  v10.4s, %18.s[3]       \n"
-
-                        "ld2        {v8.4s, v9.4s}, [%6], #32      \n"
-
-                        "prfm       pldl1keep, [%6, #256]          \n"
-
-                        "ld2        {v10.4s, v11.4s}, [%6]         \n"
-                        "ext        v12.16b, v8.16b, v10.16b, #4   \n"
-                        "ext        v11.16b, v9.16b, v11.16b, #4   \n"
-                        "ext        v10.16b, v8.16b, v10.16b, #8   \n"
-
-                        "fmla       v7.4s,   v8.4s, %19.s[0]       \n"
-                        "fmla       v13.4s,  v9.4s, %19.s[1]       \n"
-                        "fmla       v14.4s, v12.4s, %19.s[2]       \n"
-                        "fmla       v15.4s, v11.4s, %19.s[3]       \n"
-                        "fmla       v7.4s,  v10.4s, %20.s[0]       \n"
-
-                        "prfm       pldl1keep, [%2, #256]          \n"
-
-                        "ld2        {v8.4s, v9.4s}, [%2], #32      \n"
-
-                        "fadd       v14.4s, v14.4s, v15.4s         \n"
-                        "fadd       v7.4s,   v7.4s, v13.4s         \n"
-
-                        "prfm       pldl1keep, [%2, #256]          \n"
-
-                        "fadd       v7.4s, v7.4s, v14.4s           \n"
-
-                        "ld2        {v10.4s, v11.4s}, [%2]         \n"
-                        "st1        {v7.4s}, [%1], #16             \n"
-
-                        "prfm       pldl1keep, [%1, #128]          \n"
-
-                        "subs       %w0, %w0, #1                   \n"
-                        "bne        0b                             \n"
-
-                        "sub        %2, %2, #32                    \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2),     // %4
-                        "=r"(r3),     // %5
-                        "=r"(r4)      // %6
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "w"(_k0123),     // %14
-                        "w"(_k4567),     // %15
-                        "w"(_k891011),   // %16
-                        "w"(_k12131415), // %17
-                        "w"(_k16171819), // %18
-                        "w"(_k20212223), // %19
-                        "w"(_k24242424)  // %20
-                        : "cc", "memory", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-                }
-
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        //                     "veor       q15, q15            \n"// _sump3 = 0;
-                        //                     "veor       q13, q13            \n"// _sump2 = 0;
-                        //                     "veor       q14, q14            \n"// _sump3 = 0;
-
-                        "pld        [%2, #256]          \n"
-                        "vld2.f32   {d16-d19}, [%2]!    \n" // q8  = 0  2  4  6   q9  = 1  3  5  7
-
-                        "pld        [%2, #256]          \n"
-                        "vld2.f32   {d20-d23}, [%2]     \n" // q10 = 8 10 12 14   q11 = 9 11 13 15
-
-                        "pld        [%1, #128]          \n"
-                        "0:                             \n"
-
-                        "vld1.f32   {d14-d15}, [%1]     \n" // q7 = outptr
-
-                        "vext.32    q12, q8, q10, #1    \n" // q12 = 2 4 6 8
-                        "vext.32    q11, q9, q11, #1    \n" // q11 = 3 5 7 9
-                        "vext.32    q10, q8, q10, #2    \n" // q10 = 4 6 8 10
-
-                        "vmla.f32   q7, q8, %e14[0]     \n"
-                        "vmul.f32   q13, q9, %e14[1]    \n"
-
-                        "pld        [%3, #256]          \n"
-
-                        "vmul.f32   q14, q12, %f14[0]   \n"
-                        "vmul.f32   q15, q11, %f14[1]   \n"
-                        "vmla.f32   q7, q10, %e15[0]    \n"
-
-                        "vld2.f32   {d16-d19}, [%3]!    \n"
-
-                        "pld        [%3, #256]          \n"
-
-                        "vld2.f32   {d20-d23}, [%3]     \n"
-                        "vext.32    q12, q8, q10, #1    \n"
-                        "vext.32    q11, q9, q11, #1    \n"
-                        "vext.32    q10, q8, q10, #2    \n"
-
-                        "vmla.f32   q7, q8, %e15[1]     \n"
-                        "vmla.f32   q13, q9, %f15[0]    \n"
-
-                        "pld        [%4, #256]          \n"
-
-                        "vmla.f32   q14, q12, %f15[1]   \n"
-                        "vmla.f32   q15, q11, %e16[0]   \n"
-                        "vmla.f32   q7, q10, %e16[1]    \n"
-
-                        "vld2.f32   {d16-d19}, [%4]!    \n"
-
-                        "pld        [%4, #256]          \n"
-
-                        "vld2.f32   {d20-d23}, [%4]     \n"
-                        "vext.32    q12, q8, q10, #1    \n"
-                        "vext.32    q11, q9, q11, #1    \n"
-                        "vext.32    q10, q8, q10, #2    \n"
-
-                        "vmla.f32   q7, q8, %f16[0]     \n"
-                        "vmla.f32   q13, q9, %f16[1]    \n"
-
-                        "pld        [%5, #256]          \n"
-
-                        "vmla.f32   q14, q12, %e17[0]   \n"
-                        "vmla.f32   q15, q11, %e17[1]   \n"
-                        "vmla.f32   q7, q10, %f17[0]    \n"
-
-                        "vld2.f32   {d16-d19}, [%5]!    \n"
-
-                        "pld        [%5, #256]          \n"
-
-                        "vld2.f32   {d20-d23}, [%5]     \n"
-                        "vext.32    q12, q8, q10, #1    \n"
-                        "vext.32    q11, q9, q11, #1    \n"
-                        "vext.32    q10, q8, q10, #2    \n"
-
-                        "vmla.f32   q7, q8, %f17[1]     \n"
-                        "vmla.f32   q13, q9, %e18[0]    \n"
-
-                        "pld        [%6, #256]          \n"
-
-                        "vmla.f32   q14, q12, %e18[1]   \n"
-                        "vmla.f32   q15, q11, %f18[0]   \n"
-                        "vmla.f32   q7, q10, %f18[1]    \n"
-
-                        "vld2.f32   {d16-d19}, [%6]!    \n"
-
-                        "pld        [%6, #256]          \n"
-
-                        "vld2.f32   {d20-d23}, [%6]     \n"
-                        "vext.32    q12, q8, q10, #1    \n"
-                        "vext.32    q11, q9, q11, #1    \n"
-                        "vext.32    q10, q8, q10, #2    \n"
-
-                        "vmla.f32   q7, q8, %e19[0]     \n"
-                        "vmla.f32   q13, q9, %e19[1]    \n"
-                        "vmla.f32   q14, q12, %f19[0]   \n"
-                        "vmla.f32   q15, q11, %f19[1]   \n"
-                        "vmla.f32   q7, q10, %e20[0]    \n"
-
-                        "pld        [%2, #256]          \n"
-
-                        "vld2.f32   {d16-d19}, [%2]!    \n" // q8  = 0  2  4  6   q9  = 1  3  5  7
-
-                        "vadd.f32   q14, q14, q15       \n"
-                        "vadd.f32   q7, q7, q13         \n"
-                        //                     "veor       q15, q15            \n"// _sump3 = 0;
-                        //                     "veor       q13, q13            \n"// _sump2 = 0;
-
-                        "pld        [%2, #256]          \n"
-
-                        "vadd.f32   q7, q7, q14         \n"
-
-                        "vld2.f32   {d20-d23}, [%2]     \n" // q10 = 8 10 12 14   q11 = 9 11 13 15
-
-                        //                     "veor       q14, q14            \n"// _sump3 = 0;
-
-                        "vst1.f32   {d14-d15}, [%1]!    \n"
-
-                        "pld        [%1, #128]          \n"
-
-                        "subs       %0, #1              \n"
-                        "bne        0b                  \n"
-
-                        "sub        %2, #32             \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2),     // %4
-                        "=r"(r3),     // %5
-                        "=r"(r4)      // %6
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "w"(_k0123),     // %14
-                        "w"(_k4567),     // %15
-                        "w"(_k891011),   // %16
-                        "w"(_k12131415), // %17
-                        "w"(_k16171819), // %18
-                        "w"(_k20212223), // %19
-                        "w"(_k24242424)  // %20
-                        : "cc", "memory", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-                for (; remain > 0; remain--)
-                {
-                    float sum = 0;
-#if __ARM_NEON
-                    float32x4_t _r0 = vld1q_f32(r0);
-                    float32x4_t _sum = vmulq_f32(_r0, _k0123);
-
-                    float32x4_t _r1 = vld1q_f32(r1);
-                    _sum = vmlaq_f32(_sum, _r1, vld1q_f32(k1));
-
-                    float32x4_t _r2 = vld1q_f32(r2);
-                    _sum = vmlaq_f32(_sum, _r2, vld1q_f32(k2));
-
-                    float32x4_t _r3 = vld1q_f32(r3);
-                    _sum = vmlaq_f32(_sum, _r3, vld1q_f32(k3));
-
-                    float32x4_t _r4 = vld1q_f32(r4);
-                    _sum = vmlaq_f32(_sum, _r4, _k20212223);
-
-                    sum += r0[4] * k0[4];
-                    sum += r1[4] * k1[4];
-                    sum += r2[4] * k2[4];
-                    sum += r3[4] * k3[4];
-                    sum += r4[4] * k4[4];
-
-                    float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-                    _ss = vpadd_f32(_ss, _ss);
-
-                    sum += vget_lane_f32(_ss, 0);
-#else
-                    sum += r0[0] * k0[0];
-                    sum += r0[1] * k0[1];
-                    sum += r0[2] * k0[2];
-                    sum += r0[3] * k0[3];
-                    sum += r0[4] * k0[4];
-
-                    sum += r1[0] * k1[0];
-                    sum += r1[1] * k1[1];
-                    sum += r1[2] * k1[2];
-                    sum += r1[3] * k1[3];
-                    sum += r1[4] * k1[4];
-
-                    sum += r2[0] * k2[0];
-                    sum += r2[1] * k2[1];
-                    sum += r2[2] * k2[2];
-                    sum += r2[3] * k2[3];
-                    sum += r2[4] * k2[4];
-
-                    sum += r3[0] * k3[0];
-                    sum += r3[1] * k3[1];
-                    sum += r3[2] * k3[2];
-                    sum += r3[3] * k3[3];
-                    sum += r3[4] * k3[4];
-
-                    sum += r4[0] * k4[0];
-                    sum += r4[1] * k4[1];
-                    sum += r4[2] * k4[2];
-                    sum += r4[3] * k4[3];
-                    sum += r4[4] * k4[4];
-#endif
-                    *outptr += sum;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    r3 += 2;
-                    r4 += 2;
-                    outptr++;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-                r4 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5_pack4.h
deleted file mode 100644
index 96981b46..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5_pack4.h
+++ /dev/null
@@ -1,5007 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv5x5s1_pack4_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        out0.fill(_bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-            const float* r3 = img0.row(3);
-            const float* r4 = img0.row(4);
-
-            const float* kptr = (const float*)kernel.channel(p).row(q);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0] \n" // sum0 sum1 sum2 sum3
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1] \n" // r04 r05 r06 r07
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v5.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n" // r10 r11 r12 r13
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%2] \n" // r14 r15 r16 r17
-
-                        "fmla   v20.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v4.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v5.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%3], #64 \n" // r20 r21 r22 r23
-
-                        "fmla   v20.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%3] \n" // r24 r25 r26 r27
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v5.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%4], #64 \n" // r30 r31 r32 r33
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%4] \n" // r34 r35 r36 r37
-
-                        "fmla   v20.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v4.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v5.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%5], #64 \n" // r40 r41 r42 r43
-
-                        "fmla   v20.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%5] \n" // r44 r45 r46 r47
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v5.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v6.s[1]     \n"
-
-                        //                         "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6] \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v6.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v7.s[1]     \n"
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v7.s[3]     \n"
-
-                        "sub    %6, %6, #1536               \n" // kptr -= 24 * 16;
-
-                        "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #512]          \n"
-                        "vldm       %0, {d24-d31}       \n" // sum0 sum1 sum2 sum3
-
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1!, {d0-d7}        \n" // r00 r01 r02 r03
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1, {d8-d15}        \n" // r04 r05 r06 r07
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d7[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d6[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d7[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%2, #512]          \n"
-                        "vldm       %2!, {d0-d7}        \n" // r10 r11 r12 r13
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d10[0]     \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d11[0]    \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%2, #512]          \n"
-                        "vldm       %2, {d8-d15}        \n" // r14 r15 r16 r17
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d7[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d6[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d7[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%3, #512]          \n"
-                        "vldm       %3!, {d0-d7}        \n" // r20 r21 r22 r23
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d10[0]     \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d11[0]    \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%3, #512]          \n"
-                        "vldm       %3, {d8-d15}        \n" // r24 r25 r26 r27
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d7[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d6[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d7[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d0-d7}        \n" // r30 r31 r32 r33
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d10[0]     \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d11[0]    \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4, {d8-d15}        \n" // r34 r35 r36 r37
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d7[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d6[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d7[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%5, #512]          \n"
-                        "vldm       %5!, {d0-d7}        \n" // r40 r41 r42 r43
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d10[0]     \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d11[0]    \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%5, #512]          \n"
-                        "vldm       %5, {d8-d15}        \n" // r44 r45 r46 r47
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d7[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d6[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d7[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        //                         "pld        [%6, #512]          \n"
-                        "vldm       %6, {d16-d23}       \n"
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d10[0]     \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d11[0]    \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "sub        %6, %6, #1536       \n" // kptr -= 24 * 16;
-
-                        "vstm       %0!, {d24-d31}      \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v20.4s, v21.4s}, [%0]      \n" // sum0 sum1
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%1], #32   \n" // r00 r01
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmul   v22.4s, v16.4s, v0.s[0]     \n"
-                        "fmul   v23.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v2.4s, v3.4s, v4.4s, v5.4s}, [%1] \n" // r02 r03 r04 r05
-
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%2], #32   \n" // r10 r11
-
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-
-                        "fmla   v22.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v1.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v2.4s, v3.4s, v4.4s, v5.4s}, [%2] \n" // r12 r13 r14 r15
-
-                        "fmla   v22.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%3], #32   \n" // r20 r21
-
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v2.4s, v3.4s, v4.4s, v5.4s}, [%3] \n" // r22 r23 r24 r25
-
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%4], #32   \n" // r30 r31
-
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-
-                        "fmla   v22.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v1.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v2.4s, v3.4s, v4.4s, v5.4s}, [%4] \n" // r32 r33 r34 r35
-
-                        "fmla   v22.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%5], #32   \n" // r40 r41
-
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #512]       \n"
-                        "ld1    {v2.4s, v3.4s, v4.4s, v5.4s}, [%5] \n" // r42 r43 r44 r45
-
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-
-                        //                         "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6] \n"
-
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-
-                        "fadd   v20.4s, v20.4s, v22.4s      \n"
-                        "fadd   v21.4s, v21.4s, v23.4s      \n"
-
-                        "sub    %6, %6, #1536               \n" // kptr -= 24 * 16;
-
-                        "st1    {v20.4s, v21.4s}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #256]          \n"
-                        "vld1.f32   {d24-d27}, [%0 :128] \n" // sum0 sum1
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.f32   {d0-d3}, [%1 :128]! \n" // r00 r01
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmul.f32   q14, q8, d0[0]      \n"
-                        "vmul.f32   q15, q8, d2[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1, {d4-d11}        \n" // r02 r03 r04 r05
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.f32   {d0-d3}, [%2 :128]! \n" // r10 r11
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d2[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%2, #512]          \n"
-                        "vldm       %2, {d4-d11}        \n" // r12 r13 r14 r15
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.f32   {d0-d3}, [%3 :128]! \n" // r20 r21
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d2[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%3, #512]          \n"
-                        "vldm       %3, {d4-d11}        \n" // r22 r23 r24 r25
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.f32   {d0-d3}, [%4 :128]! \n" // r30 r31
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d2[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4, {d4-d11}        \n" // r32 r33 r34 r35
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.f32   {d0-d3}, [%5 :128]! \n" // r40 r41
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d2[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%5, #512]          \n"
-                        "vldm       %5, {d4-d11}        \n" // r42 r43 r44 r45
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-
-                        //                         "pld        [%6, #512]          \n"
-                        "vldm       %6, {d16-d23}       \n"
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-
-                        "vadd.f32   q12, q12, q14       \n"
-                        "vadd.f32   q13, q13, q15       \n"
-
-                        "sub        %6, %6, #1536       \n" // kptr -= 24 * 16;
-
-                        "vst1.f32   {d24-d27}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v20.4s}, [%0]              \n" // sum0
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.4s}, [%1], #16          \n" // r00
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v1.4s, v2.4s, v3.4s, v4.4s}, [%1] \n" // r01 r02 r03 r04
-
-                        "fmul   v21.4s, v16.4s, v0.s[0]     \n"
-                        "fmul   v22.4s, v17.4s, v0.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmul   v23.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v0.4s}, [%2], #16          \n" // r10
-
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v1.4s, v2.4s, v3.4s, v4.4s}, [%2] \n" // r11 r12 r13 r14
-
-                        "fmla   v21.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v0.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.4s}, [%3], #16          \n" // r20
-
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v1.4s, v2.4s, v3.4s, v4.4s}, [%3] \n" // r21 r22 r23 r24
-
-                        "fmla   v21.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v0.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v0.4s}, [%4], #16          \n" // r30
-
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v1.4s, v2.4s, v3.4s, v4.4s}, [%4] \n" // r31 r32 r33 r34
-
-                        "fmla   v21.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v0.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #128]       \n"
-                        "ld1    {v0.4s}, [%5], #16          \n" // r40
-
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #512]       \n"
-                        "ld1    {v1.4s, v2.4s, v3.4s, v4.4s}, [%5] \n" // r41 r42 r43 r44
-
-                        "fmla   v21.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v0.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-
-                        //                         "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6] \n"
-
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-
-                        "fadd   v22.4s, v21.4s, v22.4s      \n"
-                        "fadd   v23.4s, v22.4s, v23.4s      \n"
-                        "fadd   v20.4s, v20.4s, v23.4s      \n"
-
-                        "sub    %6, %6, #1536               \n" // kptr -= 24 * 16;
-
-                        "st1    {v20.4s}, [%0], #16         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #128]          \n"
-                        "vld1.f32   {d24-d25}, [%0 :128] \n" // sum0
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%1 :128]! \n" // r00
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmul.f32   q13, q8, d0[0]      \n"
-                        "vmul.f32   q14, q9, d0[1]      \n"
-                        "vmul.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1, {d2-d9}         \n" // r01 r02 r03 r04
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%2 :128]! \n" // r10
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%2, #512]          \n"
-                        "vldm       %2, {d2-d9}         \n" // r11 r12 r13 r14
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%3 :128]! \n" // r20
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%3, #512]          \n"
-                        "vldm       %3, {d2-d9}         \n" // r21 r22 r23 r24
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%4 :128]! \n" // r30
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4, {d2-d9}         \n" // r31 r32 r33 r34
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%5, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%5 :128]! \n" // r40
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%5, #512]          \n"
-                        "vldm       %5, {d2-d9}         \n" // r41 r42 r43 r44
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-
-                        //                         "pld        [%6, #512]          \n"
-                        "vldm       %6, {d16-d23}       \n"
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-
-                        "vadd.f32   q13, q13, q14       \n"
-                        "vadd.f32   q12, q12, q15       \n"
-                        "vadd.f32   q12, q12, q13       \n"
-
-                        "sub        %6, %6, #1536       \n" // kptr -= 24 * 16;
-
-                        "vst1.f32   {d24-d25}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-
-                r0 += 4 * 4;
-                r1 += 4 * 4;
-                r2 += 4 * 4;
-                r3 += 4 * 4;
-                r4 += 4 * 4;
-            }
-        }
-    }
-}
-
-static void conv5x5s2_pack4_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        out0.fill(_bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-            const float* r3 = img0.row(3);
-            const float* r4 = img0.row(4);
-
-            const float* kptr = (const float*)kernel.channel(p).row(q);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0] \n" // sum0 sum1 sum2 sum3
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n" // r04 r05 r06 r07
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%1, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%1] \n" // r08 r09 r010
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v28.s[3]    \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v29.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v29.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v29.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v29.s[3]    \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n" // r10 r11 r12 r13
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v23.4s, v16.4s, v30.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v28.s[1]    \n"
-                        "fmla   v23.4s, v17.4s, v30.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v23.4s, v18.4s, v30.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v28.s[3]    \n"
-                        "fmla   v23.4s, v19.4s, v30.s[3]    \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%2], #64 \n" // r14 r15 r16 r17
-
-                        "fmla   v20.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%2] \n" // r18 r19 r110
-
-                        "fmla   v20.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v28.s[3]    \n"
-
-                        "fmla   v20.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v29.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v7.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v29.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v29.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v7.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v29.s[3]    \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%3], #64 \n" // r20 r21 r22 r23
-
-                        "fmla   v20.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v28.s[0]    \n"
-                        "fmla   v23.4s, v24.4s, v30.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v28.s[1]    \n"
-                        "fmla   v23.4s, v25.4s, v30.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v28.s[2]    \n"
-                        "fmla   v23.4s, v26.4s, v30.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v28.s[3]    \n"
-                        "fmla   v23.4s, v27.4s, v30.s[3]    \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%3], #64 \n" // r24 r25 r26 r27
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%3] \n" // r28 r29 r210
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v28.s[3]    \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v29.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v29.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v29.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v29.s[3]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%4], #64 \n" // r30 r31 r32 r33
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v23.4s, v16.4s, v30.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v28.s[1]    \n"
-                        "fmla   v23.4s, v17.4s, v30.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v23.4s, v18.4s, v30.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v28.s[3]    \n"
-                        "fmla   v23.4s, v19.4s, v30.s[3]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%4], #64 \n" // r34 r35 r36 r37
-
-                        "fmla   v20.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%4] \n" // r38 r39 r310
-
-                        "fmla   v20.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v28.s[3]    \n"
-
-                        "fmla   v20.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v29.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v7.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v29.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v29.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v7.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v29.s[3]    \n"
-
-                        "prfm   pldl1keep, [%5, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%5], #64 \n" // r40 r41 r42 r43
-
-                        "fmla   v20.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v28.s[0]    \n"
-                        "fmla   v23.4s, v24.4s, v30.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v28.s[1]    \n"
-                        "fmla   v23.4s, v25.4s, v30.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v28.s[2]    \n"
-                        "fmla   v23.4s, v26.4s, v30.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v28.s[3]    \n"
-                        "fmla   v23.4s, v27.4s, v30.s[3]    \n"
-
-                        "prfm   pldl1keep, [%5, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%5], #64 \n" // r44 r45 r46 r47
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%5] \n" // r48 r49 r410
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v28.s[3]    \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v29.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v29.s[1]    \n"
-
-                        //                         "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6] \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v29.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v29.s[3]    \n"
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v23.4s, v16.4s, v30.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v28.s[1]    \n"
-                        "fmla   v23.4s, v17.4s, v30.s[1]    \n"
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v23.4s, v18.4s, v30.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v28.s[3]    \n"
-                        "fmla   v23.4s, v19.4s, v30.s[3]    \n"
-
-                        "sub    %6, %6, #1536               \n" // kptr -= 24 * 16;
-
-                        "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #512]          \n"
-                        "vldm       %0, {d24-d31}       \n" // sum0 sum1 sum2 sum3
-
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1!, {d0-d7}        \n" // r00 r01 r02 r03
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1!, {d8-d15}       \n" // r04 r05 r06 r07
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.f32   {d0-d3}, [%1 :128]! \n" // r08 r09
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d6[0]      \n"
-                        "vmla.f32   q13, q8, d10[0]     \n"
-                        "vmla.f32   q14, q8, d14[0]     \n"
-                        "vmla.f32   q15, q8, d2[0]      \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q9, d14[1]     \n"
-                        "vmla.f32   q15, q9, d2[1]      \n"
-                        "vmla.f32   q12, q10, d7[0]     \n"
-                        "vmla.f32   q13, q10, d11[0]    \n"
-                        "vmla.f32   q14, q10, d15[0]    \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "vmla.f32   q14, q11, d15[1]    \n"
-                        "vmla.f32   q15, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d4-d5}, [%1 :128]  \n" // r010
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d12[0]     \n"
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d13[0]    \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-
-                        "pld        [%2, #512]          \n"
-                        "vldm       %2!, {d8-d15}       \n" // r10 r11 r12 r13
-
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%2, #512]          \n"
-                        "vldm       %2!, {d0-d7}        \n" // r14 r15 r16 r17
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d12[0]     \n"
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d13[0]    \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d10[0]     \n"
-                        "vmla.f32   q13, q8, d14[0]     \n"
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d10[1]     \n"
-                        "vmla.f32   q13, q9, d14[1]     \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d11[0]    \n"
-                        "vmla.f32   q13, q10, d15[0]    \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d11[1]    \n"
-                        "vmla.f32   q13, q11, d15[1]    \n"
-                        "vmla.f32   q14, q11, d3[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.f32   {d8-d11}, [%2 :128]! \n" // r18 r19
-
-                        "vmla.f32   q12, q8, d12[0]     \n"
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d12[1]     \n"
-                        "vmla.f32   q13, q9, d0[1]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-                        "vmla.f32   q12, q10, d13[0]    \n"
-                        "vmla.f32   q13, q10, d1[0]     \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d13[1]    \n"
-                        "vmla.f32   q13, q11, d1[1]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d14[0]     \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d14[1]     \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d15[0]    \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d15[1]    \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d7[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d12-d13}, [%2 :128] \n" // r110
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-
-                        "pld        [%3, #512]          \n"
-                        "vldm       %3!, {d0-d7}        \n" // r20 r21 r22 r23
-
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%3, #512]          \n"
-                        "vldm       %3!, {d8-d15}       \n" // r24 r25 r26 r27
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.f32   {d0-d3}, [%3 :128]! \n" // r28 r29
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d6[0]      \n"
-                        "vmla.f32   q13, q8, d10[0]     \n"
-                        "vmla.f32   q14, q8, d14[0]     \n"
-                        "vmla.f32   q15, q8, d2[0]      \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q9, d14[1]     \n"
-                        "vmla.f32   q15, q9, d2[1]      \n"
-                        "vmla.f32   q12, q10, d7[0]     \n"
-                        "vmla.f32   q13, q10, d11[0]    \n"
-                        "vmla.f32   q14, q10, d15[0]    \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "vmla.f32   q14, q11, d15[1]    \n"
-                        "vmla.f32   q15, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.f32   {d4-d5}, [%3 :128]  \n" // r210
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d12[0]     \n"
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d13[0]    \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d8-d15}       \n" // r30 r31 r32 r33
-
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d0-d7}        \n" // r34 r35 r36 r37
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d12[0]     \n"
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d13[0]    \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d10[0]     \n"
-                        "vmla.f32   q13, q8, d14[0]     \n"
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d10[1]     \n"
-                        "vmla.f32   q13, q9, d14[1]     \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d11[0]    \n"
-                        "vmla.f32   q13, q10, d15[0]    \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d11[1]    \n"
-                        "vmla.f32   q13, q11, d15[1]    \n"
-                        "vmla.f32   q14, q11, d3[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.f32   {d8-d11}, [%4 :128]! \n" // r38 r39
-
-                        "vmla.f32   q12, q8, d12[0]     \n"
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d12[1]     \n"
-                        "vmla.f32   q13, q9, d0[1]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-                        "vmla.f32   q12, q10, d13[0]    \n"
-                        "vmla.f32   q13, q10, d1[0]     \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d13[1]    \n"
-                        "vmla.f32   q13, q11, d1[1]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d14[0]     \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d14[1]     \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d15[0]    \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d15[1]    \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d7[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.f32   {d12-d13}, [%4 :128] \n" // r310
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-
-                        "pld        [%5, #512]          \n"
-                        "vldm       %5!, {d0-d7}        \n" // r40 r41 r42 r43
-
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%5, #512]          \n"
-                        "vldm       %5!, {d8-d15}       \n" // r44 r45 r46 r47
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.f32   {d0-d3}, [%5 :128]! \n" // r48 r49
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q12, q8, d6[0]      \n"
-                        "vmla.f32   q13, q8, d10[0]     \n"
-                        "vmla.f32   q14, q8, d14[0]     \n"
-                        "vmla.f32   q15, q8, d2[0]      \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q9, d14[1]     \n"
-                        "vmla.f32   q15, q9, d2[1]      \n"
-                        "vmla.f32   q12, q10, d7[0]     \n"
-                        "vmla.f32   q13, q10, d11[0]    \n"
-                        "vmla.f32   q14, q10, d15[0]    \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "vmla.f32   q14, q11, d15[1]    \n"
-                        "vmla.f32   q15, q11, d3[1]     \n"
-
-                        //                         "pld        [%6, #512]          \n"
-                        "vldm       %6, {d16-d23}       \n"
-
-                        "pld        [%5, #128]          \n"
-                        "vld1.f32   {d4-d5}, [%5 :128]  \n" // r410
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d12[0]     \n"
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d13[0]    \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "sub        %6, %6, #1536       \n" // kptr -= 24 * 16;
-
-                        "sub        %1, %1, #32         \n"
-                        "sub        %2, %2, #32         \n"
-                        "sub        %3, %3, #32         \n"
-                        "sub        %4, %4, #32         \n"
-                        "sub        %5, %5, #32         \n"
-
-                        "vstm       %0!, {d24-d31}      \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v20.4s, v21.4s}, [%0]      \n" // sum0 sum1
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmul   v22.4s, v16.4s, v0.s[0]     \n"
-                        "fmul   v23.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-
-                        "prfm   pldl1keep, [%1, #384]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s}, [%1] \n" // r04 r05 r06
-
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n" // r10 r11 r12 r13
-
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-
-                        "fmla   v22.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #384]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s}, [%2] \n" // r14 r15 r16
-
-                        "fmla   v22.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%3], #64 \n" // r20 r21 r22 r23
-
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v6.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #384]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s}, [%3] \n" // r24 r25 r26
-
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%4], #64 \n" // r30 r31 r32 r33
-
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-
-                        "fmla   v22.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #384]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s}, [%4] \n" // r34 r35 r36
-
-                        "fmla   v22.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%5], #64 \n" // r40 r41 r42 r43
-
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v6.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #384]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s}, [%5] \n" // r44 r45 r46
-
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-
-                        //                         "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6] \n"
-
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-
-                        "fadd   v20.4s, v20.4s, v22.4s      \n"
-                        "fadd   v21.4s, v21.4s, v23.4s      \n"
-
-                        "sub    %6, %6, #1536               \n" // kptr -= 24 * 16;
-
-                        "st1    {v20.4s, v21.4s}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #256]          \n"
-                        "vld1.f32   {d24-d27}, [%0 :128] \n" // sum0 sum1
-
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1!, {d0-d7}        \n" // r00 r01 r02 r03
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmul.f32   q14, q8, d0[0]      \n"
-                        "vmul.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%1, #384]          \n"
-                        "vldm       %1, {d8-d13}        \n" // r04 r05 r06
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%2, #512]          \n"
-                        "vldm       %2!, {d0-d7}        \n" // r10 r11 r12 r13
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%2, #384]          \n"
-                        "vldm       %2, {d8-d13}        \n" // r14 r15 r16
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%3, #512]          \n"
-                        "vldm       %3!, {d0-d7}        \n" // r20 r21 r22 r23
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%3, #384]          \n"
-                        "vldm       %3, {d8-d13}        \n" // r24 r25 r26
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vldm       %4!, {d0-d7}        \n" // r30 r31 r32 r33
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%4, #384]          \n"
-                        "vldm       %4, {d8-d13}        \n" // r34 r35 r36
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%5, #512]          \n"
-                        "vldm       %5!, {d0-d7}        \n" // r40 r41 r42 r43
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%5, #384]          \n"
-                        "vldm       %5, {d8-d13}        \n" // r44 r45 r46
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-
-                        //                         "pld        [%6, #512]          \n"
-                        "vldm       %6, {d16-d23}      \n"
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-
-                        "vadd.f32   q12, q12, q14       \n"
-                        "vadd.f32   q13, q13, q15       \n"
-
-                        "sub        %6, %6, #1536       \n" // kptr -= 24 * 16;
-
-                        "vst1.f32   {d24-d27}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v20.4s}, [%0]              \n" // sum0
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%1], #32   \n" // r00 r01
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmul   v21.4s, v16.4s, v0.s[0]     \n"
-                        "fmul   v22.4s, v17.4s, v0.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmul   v23.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-
-                        "prfm   pldl1keep, [%1, #384]       \n"
-                        "ld1    {v2.4s, v3.4s, v4.4s}, [%1] \n" // r02 r03 r04
-
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%2], #32   \n" // r10 r11
-
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v0.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #384]       \n"
-                        "ld1    {v2.4s, v3.4s, v4.4s}, [%2] \n" // r12 r13 r14
-
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%3], #32   \n" // r20 r21
-
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v0.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #384]       \n"
-                        "ld1    {v2.4s, v3.4s, v4.4s}, [%3] \n" // r22 r23 r24
-
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%4], #32   \n" // r30 r31
-
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v0.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #384]       \n"
-                        "ld1    {v2.4s, v3.4s, v4.4s}, [%4] \n" // r32 r33 r34
-
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%5], #32   \n" // r40 r41
-
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v0.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #384]       \n"
-                        "ld1    {v2.4s, v3.4s, v4.4s}, [%5] \n" // r42 r43 r44
-
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%6], #64 \n"
-
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-
-                        //                         "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%6] \n"
-
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-
-                        "fadd   v22.4s, v21.4s, v22.4s      \n"
-                        "fadd   v23.4s, v22.4s, v23.4s      \n"
-                        "fadd   v20.4s, v20.4s, v23.4s      \n"
-
-                        "sub    %6, %6, #1536               \n" // kptr -= 24 * 16;
-
-                        "st1    {v20.4s}, [%0], #16         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #128]          \n"
-                        "vld1.f32   {d24-d25}, [%0 :128] \n" // sum0
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.f32   {d0-d3}, [%1 :128]! \n" // r00 r01
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmul.f32   q13, q8, d0[0]      \n"
-                        "vmul.f32   q14, q9, d0[1]      \n"
-                        "vmul.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%1, #384]          \n"
-                        "vldm       %1, {d4-d9}         \n" // r02 r03 r04
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.f32   {d0-d3}, [%2 :128]! \n" // r10 r11
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%2, #384]          \n"
-                        "vldm       %2, {d4-d9}         \n" // r12 r13 r14
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.f32   {d0-d3}, [%3 :128]! \n" // r20 r21
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%3, #384]          \n"
-                        "vldm       %3, {d4-d9}         \n" // r22 r23 r24
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.f32   {d0-d3}, [%4 :128]! \n" // r30 r31
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%4, #384]          \n"
-                        "vldm       %4, {d4-d9}         \n" // r32 r33 r34
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.f32   {d0-d3}, [%5 :128]! \n" // r40 r41
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "pld        [%5, #384]          \n"
-                        "vldm       %5, {d4-d9}         \n" // r42 r43 r44
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vldm       %6!, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-
-                        //                         "pld        [%6, #512]          \n"
-                        "vldm       %6, {d16-d23}      \n"
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-
-                        "vadd.f32   q14, q13, q14       \n"
-                        "vadd.f32   q15, q14, q15       \n"
-                        "vadd.f32   q12, q12, q15       \n"
-
-                        "sub        %6, %6, #1536       \n" // kptr -= 24 * 16;
-
-                        "vst1.f32   {d24-d25}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-                r4 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5_pack4_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5_pack4_bf16s.h
deleted file mode 100644
index 18e8eaeb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5_pack4_bf16s.h
+++ /dev/null
@@ -1,12974 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv5x5s1_pack4_bf16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    Mat top_blob_fp32(outw, outh, opt.num_threads, (size_t)4u * 4, 4, opt.workspace_allocator);
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob_fp32.channel(get_omp_thread_num());
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        out0.fill(_bias0);
-
-        int q = 0;
-        for (; q < inch - 1; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<const unsigned short>(0);
-            const unsigned short* r1 = img0.row<const unsigned short>(1);
-            const unsigned short* r2 = img0.row<const unsigned short>(2);
-            const unsigned short* r3 = img0.row<const unsigned short>(3);
-            const unsigned short* r4 = img0.row<const unsigned short>(4);
-
-            const unsigned short* kptr = kernel.channel(p).row<const unsigned short>(q);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0] \n" // sum0 sum1 sum2 sum3
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%1], #32 \n" // r00 r01 r02 r03
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%1] \n" // r04 r05 r06 r07
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v5.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n" // r10 r11 r12 r13
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%2] \n" // r14 r15 r16 r17
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v4.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v5.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%3], #32 \n" // r20 r21 r22 r23
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%3] \n" // r24 r25 r26 r27
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v5.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%4], #32 \n" // r30 r31 r32 r33
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%4] \n" // r34 r35 r36 r37
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v4.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v5.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%5], #32 \n" // r40 r41 r42 r43
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%5] \n" // r44 r45 r46 r47
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v5.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v6.s[1]     \n"
-
-                        //                         "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6] \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v6.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v7.s[1]     \n"
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v7.s[3]     \n"
-
-                        "sub    %6, %6, #768                \n" // kptr -= 24 * 16;
-
-                        "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #512]          \n"
-                        "vldm       %0, {d24-d31}       \n" // sum0 sum1 sum2 sum3
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%1 :64]!  \n" // r00 r01 r02 r03
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%1 :64] \n" // r04 r05 r06 r07
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[0]     \n"
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        "vmla.f32   q15, q11, d8[1]     \n"
-                        "vmla.f32   q12, q8, d3[0]      \n"
-                        "vmla.f32   q13, q8, d5[0]      \n"
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "vmla.f32   q14, q9, d7[1]      \n"
-                        "vmla.f32   q15, q9, d9[1]      \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[0]     \n"
-                        "vmla.f32   q13, q10, d8[0]     \n"
-                        "vmla.f32   q14, q10, d10[0]    \n"
-                        "vmla.f32   q15, q10, d12[0]    \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "vmla.f32   q14, q11, d10[1]    \n"
-                        "vmla.f32   q15, q11, d12[1]    \n"
-                        "vmla.f32   q12, q8, d7[0]      \n"
-                        "vmla.f32   q13, q8, d9[0]      \n"
-                        "vmla.f32   q14, q8, d11[0]     \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vmla.f32   q14, q9, d11[1]     \n"
-                        "vmla.f32   q15, q9, d13[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%2 :64]!  \n" // r10 r11 r12 r13
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d10[0]     \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d11[0]    \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d0[0]     \n"
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d6[0]     \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d2[1]     \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "vmla.f32   q15, q11, d6[1]     \n"
-                        "vmla.f32   q12, q8, d1[0]      \n"
-                        "vmla.f32   q13, q8, d3[0]      \n"
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vmla.f32   q13, q9, d3[1]      \n"
-                        "vmla.f32   q14, q9, d5[1]      \n"
-                        "vmla.f32   q15, q9, d7[1]      \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%2 :64] \n" // r14 r15 r16 r17
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d7[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d4[0]     \n"
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d10[0]    \n"
-                        "vmla.f32   q12, q11, d4[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "vmla.f32   q15, q11, d10[1]    \n"
-                        "vmla.f32   q12, q8, d5[0]      \n"
-                        "vmla.f32   q13, q8, d7[0]      \n"
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vmla.f32   q15, q8, d11[0]     \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vmla.f32   q14, q9, d9[1]      \n"
-                        "vmla.f32   q15, q9, d11[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d6[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d7[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%3 :64]!  \n" // r20 r21 r22 r23
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q12, q10, d8[0]     \n"
-                        "vmla.f32   q13, q10, d10[0]    \n"
-                        "vmla.f32   q14, q10, d12[0]    \n"
-                        "vmla.f32   q15, q10, d14[0]    \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "vmla.f32   q14, q11, d12[1]    \n"
-                        "vmla.f32   q15, q11, d14[1]    \n"
-                        "vmla.f32   q12, q8, d9[0]      \n"
-                        "vmla.f32   q13, q8, d11[0]     \n"
-                        "vmla.f32   q14, q8, d13[0]     \n"
-                        "vmla.f32   q15, q8, d15[0]     \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "vmla.f32   q14, q9, d13[1]     \n"
-                        "vmla.f32   q15, q9, d15[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%3 :64] \n" // r24 r25 r26 r27
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[0]     \n"
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        "vmla.f32   q15, q11, d8[1]     \n"
-                        "vmla.f32   q12, q8, d3[0]      \n"
-                        "vmla.f32   q13, q8, d5[0]      \n"
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "vmla.f32   q14, q9, d7[1]      \n"
-                        "vmla.f32   q15, q9, d9[1]      \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[0]     \n"
-                        "vmla.f32   q13, q10, d8[0]     \n"
-                        "vmla.f32   q14, q10, d10[0]    \n"
-                        "vmla.f32   q15, q10, d12[0]    \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "vmla.f32   q14, q11, d10[1]    \n"
-                        "vmla.f32   q15, q11, d12[1]    \n"
-                        "vmla.f32   q12, q8, d7[0]      \n"
-                        "vmla.f32   q13, q8, d9[0]      \n"
-                        "vmla.f32   q14, q8, d11[0]     \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vmla.f32   q14, q9, d11[1]     \n"
-                        "vmla.f32   q15, q9, d13[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%4 :64]!  \n" // r30 r31 r32 r33
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d10[0]     \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d11[0]    \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d0[0]     \n"
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d6[0]     \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d2[1]     \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "vmla.f32   q15, q11, d6[1]     \n"
-                        "vmla.f32   q12, q8, d1[0]      \n"
-                        "vmla.f32   q13, q8, d3[0]      \n"
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vmla.f32   q13, q9, d3[1]      \n"
-                        "vmla.f32   q14, q9, d5[1]      \n"
-                        "vmla.f32   q15, q9, d7[1]      \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%4 :64] \n" // r34 r35 r36 r37
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d7[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d4[0]     \n"
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d10[0]    \n"
-                        "vmla.f32   q12, q11, d4[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "vmla.f32   q15, q11, d10[1]    \n"
-                        "vmla.f32   q12, q8, d5[0]      \n"
-                        "vmla.f32   q13, q8, d7[0]      \n"
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vmla.f32   q15, q8, d11[0]     \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vmla.f32   q14, q9, d9[1]      \n"
-                        "vmla.f32   q15, q9, d11[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d6[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d7[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%5 :64]!  \n" // r40 r41 r42 r43
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q12, q10, d8[0]     \n"
-                        "vmla.f32   q13, q10, d10[0]    \n"
-                        "vmla.f32   q14, q10, d12[0]    \n"
-                        "vmla.f32   q15, q10, d14[0]    \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "vmla.f32   q14, q11, d12[1]    \n"
-                        "vmla.f32   q15, q11, d14[1]    \n"
-                        "vmla.f32   q12, q8, d9[0]      \n"
-                        "vmla.f32   q13, q8, d11[0]     \n"
-                        "vmla.f32   q14, q8, d13[0]     \n"
-                        "vmla.f32   q15, q8, d15[0]     \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "vmla.f32   q14, q9, d13[1]     \n"
-                        "vmla.f32   q15, q9, d15[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%5 :64] \n" // r44 r45 r46 r47
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[0]     \n"
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        "vmla.f32   q15, q11, d8[1]     \n"
-                        "vmla.f32   q12, q8, d3[0]      \n"
-                        "vmla.f32   q13, q8, d5[0]      \n"
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "vmla.f32   q14, q9, d7[1]      \n"
-                        "vmla.f32   q15, q9, d9[1]      \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[0]     \n"
-                        "vmla.f32   q13, q10, d8[0]     \n"
-                        "vmla.f32   q14, q10, d10[0]    \n"
-                        "vmla.f32   q15, q10, d12[0]    \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "vmla.f32   q14, q11, d10[1]    \n"
-                        "vmla.f32   q15, q11, d12[1]    \n"
-                        "vmla.f32   q12, q8, d7[0]      \n"
-                        "vmla.f32   q13, q8, d9[0]      \n"
-                        "vmla.f32   q14, q8, d11[0]     \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vmla.f32   q14, q9, d11[1]     \n"
-                        "vmla.f32   q15, q9, d13[1]     \n"
-
-                        //                         "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :64] \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d10[0]     \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d11[0]    \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "sub        %6, %6, #768        \n" // kptr -= 24 * 16;
-
-                        "vstm       %0!, {d24-d31}      \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%1], #16   \n" // r00 r01
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v20.4s, v21.4s}, [%0]      \n" // sum0 sum1
-
-                        "fmul   v22.4s, v16.4s, v0.s[0]     \n"
-                        "fmul   v23.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h, v5.4h}, [%1] \n" // r02 r03 r04 r05
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%2], #16   \n" // r10 r11
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v26.4s, v0.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h, v5.4h}, [%2] \n" // r12 r13 r14 r15
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v27.4s, v1.s[3]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v1.s[0]     \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v23.4s, v16.4s, v2.s[0]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v22.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%3], #16   \n" // r20 r21
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h, v5.4h}, [%3] \n" // r22 r23 r24 r25
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%4], #16   \n" // r30 r31
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v26.4s, v0.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h, v5.4h}, [%4] \n" // r32 r33 r34 r35
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v27.4s, v1.s[3]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v1.s[0]     \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v23.4s, v16.4s, v2.s[0]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v22.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%5], #16   \n" // r40 r41
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h, v5.4h}, [%5] \n" // r42 r43 r44 r45
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        //                         "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6] \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-
-                        "fadd   v20.4s, v20.4s, v22.4s      \n"
-                        "fadd   v21.4s, v21.4s, v23.4s      \n"
-
-                        "sub    %6, %6, #768                \n" // kptr -= 24 * 16;
-
-                        "st1    {v20.4s, v21.4s}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%1 :64]!  \n" // r00 r01
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.u16   {d8-d11}, [%1 :64]  \n" // r02 r03 r04 r05
-
-                        "vshll.u16  q8, d20, #16        \n"
-
-                        "pld        [%0, #256]          \n"
-                        "vld1.f32   {d24-d27}, [%0 :128] \n" // sum0 sum1
-
-                        "vmul.f32   q14, q8, d0[0]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vmul.f32   q15, q8, d2[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vshll.u16  q2, d8, #16         \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d2[0]     \n"
-                        "vmla.f32   q15, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d3[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q3, d9, #16         \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%2 :64]!  \n" // r10 r11
-
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d8-d11}, [%2 :64]  \n" // r12 r13 r14 r15
-
-                        "vmla.f32   q14, q10, d0[0]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vmla.f32   q15, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d2[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d1[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vshll.u16  q2, d8, #16         \n"
-                        "vmla.f32   q13, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vshll.u16  q3, d9, #16         \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d4[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%3 :64]!  \n" // r20 r21
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d10[0]    \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d11[0]     \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d8-d11}, [%3 :64]  \n" // r22 r23 r24 r25
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vmla.f32   q15, q8, d2[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vshll.u16  q2, d8, #16         \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d2[0]     \n"
-                        "vmla.f32   q15, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d3[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q3, d9, #16         \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "pld        [%4, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%4 :64]!  \n" // r30 r31
-
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d8-d11}, [%4 :64]  \n" // r32 r33 r34 r35
-
-                        "vmla.f32   q14, q10, d0[0]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vmla.f32   q15, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d2[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d1[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vshll.u16  q2, d8, #16         \n"
-                        "vmla.f32   q13, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vshll.u16  q3, d9, #16         \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d4[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "pld        [%5, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%5 :64]!  \n" // r40 r41
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d10[0]    \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d11[0]     \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d8-d11}, [%5 :64]  \n" // r42 r43 r44 r45
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vmla.f32   q15, q8, d2[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vshll.u16  q2, d8, #16         \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d2[0]     \n"
-                        "vmla.f32   q15, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d3[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q3, d9, #16         \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128] \n"
-
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-
-                        "vadd.f32   q12, q12, q14       \n"
-                        "vadd.f32   q13, q13, q15       \n"
-
-                        "sub        %6, %6, #768        \n" // kptr -= 24 * 16;
-
-                        "vst1.f32   {d24-d27}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #64]        \n"
-                        "ld1    {v0.4h}, [%1], #8           \n" // r00
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v1.4h, v2.4h, v3.4h, v4.4h}, [%1] \n" // r01 r02 r03 r04
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v20.4s}, [%0]              \n" // sum0
-
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmul   v21.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmul   v22.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmul   v23.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v0.4h}, [%2], #8           \n" // r10
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v1.4h, v2.4h, v3.4h, v4.4h}, [%2] \n" // r11 r12 r13 r14
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v22.4s, v25.4s, v0.s[1]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v23.4s, v26.4s, v0.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v1.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v0.4h}, [%3], #8           \n" // r20
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v1.4h, v2.4h, v3.4h, v4.4h}, [%3] \n" // r21 r22 r23 r24
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v22.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v23.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #64]        \n"
-                        "ld1    {v0.4h}, [%4], #8           \n" // r30
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v1.4h, v2.4h, v3.4h, v4.4h}, [%4] \n" // r31 r32 r33 r34
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v22.4s, v25.4s, v0.s[1]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v23.4s, v26.4s, v0.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v1.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #64]        \n"
-                        "ld1    {v0.4h}, [%5], #8           \n" // r40
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v1.4h, v2.4h, v3.4h, v4.4h}, [%5] \n" // r41 r42 r43 r44
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v22.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v23.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        //                         "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6] \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-
-                        "fadd   v22.4s, v21.4s, v22.4s      \n"
-                        "fadd   v23.4s, v22.4s, v23.4s      \n"
-                        "fadd   v20.4s, v20.4s, v23.4s      \n"
-
-                        "sub    %6, %6, #768                \n" // kptr -= 24 * 16;
-
-                        "st1    {v20.4s}, [%0], #16         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #64]           \n"
-                        "vld1.u16   {d1}, [%1 :64]!     \n" // r00
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "pld        [%0, #128]          \n"
-                        "vld1.f32   {d24-d25}, [%0 :128] \n" // sum0
-
-                        "vmul.f32   q13, q8, d0[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmul.f32   q14, q9, d0[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmul.f32   q15, q10, d1[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "pld        [%1, #256]          \n"
-                        "vld1.u16   {d6-d9}, [%1 :64]   \n" // r01 r02 r03 r04
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q1, d6, #16         \n"
-                        "vshll.u16  q2, d7, #16         \n"
-                        "vshll.u16  q3, d8, #16         \n"
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d2[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "pld        [%2, #64]           \n"
-                        "vld1.u16   {d1}, [%2 :64]!     \n" // r10
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d0[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d0[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d1[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d6-d9}, [%2 :64]   \n" // r11 r12 r13 r14
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q1, d6, #16         \n"
-                        "vshll.u16  q2, d7, #16         \n"
-                        "vshll.u16  q3, d8, #16         \n"
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "pld        [%3, #64]           \n"
-                        "vld1.u16   {d1}, [%3 :64]!     \n" // r20
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q13, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d6-d9}, [%3 :64]   \n" // r21 r22 r23 r24
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q1, d6, #16         \n"
-                        "vshll.u16  q2, d7, #16         \n"
-                        "vshll.u16  q3, d8, #16         \n"
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d2[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "pld        [%4, #64]           \n"
-                        "vld1.u16   {d1}, [%4 :64]!     \n" // r30
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d0[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d0[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d1[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d6-d9}, [%4 :64]   \n" // r31 r32 r33 r34
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q1, d6, #16         \n"
-                        "vshll.u16  q2, d7, #16         \n"
-                        "vshll.u16  q3, d8, #16         \n"
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "pld        [%5, #64]           \n"
-                        "vld1.u16   {d1}, [%5 :64]!     \n" // r40
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q13, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d6-d9}, [%5 :64]   \n" // r41 r42 r43 r44
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q1, d6, #16         \n"
-                        "vshll.u16  q2, d7, #16         \n"
-                        "vshll.u16  q3, d8, #16         \n"
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d2[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        //                         "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128] \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-
-                        "vadd.f32   q13, q13, q14       \n"
-                        "vadd.f32   q12, q12, q15       \n"
-                        "vadd.f32   q12, q12, q13       \n"
-
-                        "sub        %6, %6, #768        \n" // kptr -= 24 * 16;
-
-                        "vst1.f32   {d24-d25}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-
-                r0 += 4 * 4;
-                r1 += 4 * 4;
-                r2 += 4 * 4;
-                r3 += 4 * 4;
-                r4 += 4 * 4;
-            }
-        }
-        for (; q < inch; q++)
-        {
-            unsigned short* outptr0_bf16 = top_blob.channel(p);
-
-            const float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<const unsigned short>(0);
-            const unsigned short* r1 = img0.row<const unsigned short>(1);
-            const unsigned short* r2 = img0.row<const unsigned short>(2);
-            const unsigned short* r3 = img0.row<const unsigned short>(3);
-            const unsigned short* r4 = img0.row<const unsigned short>(4);
-
-            const unsigned short* kptr = kernel.channel(p).row<const unsigned short>(q);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%1], #64 \n" // sum0 sum1 sum2 sum3
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n" // r00 r01 r02 r03
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%2] \n" // r04 r05 r06 r07
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v5.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%3], #32 \n" // r10 r11 r12 r13
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%3] \n" // r14 r15 r16 r17
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v4.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v5.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%4], #32 \n" // r20 r21 r22 r23
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%4] \n" // r24 r25 r26 r27
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v5.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%5], #32 \n" // r30 r31 r32 r33
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%5] \n" // r34 r35 r36 r37
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v4.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v5.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%6], #32 \n" // r40 r41 r42 r43
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v3.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%6] \n" // r44 r45 r46 r47
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v4.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v5.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v5.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v6.s[1]     \n"
-
-                        //                         "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7] \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v6.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v7.s[1]     \n"
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v7.s[3]     \n"
-
-                        "sub    %7, %7, #768                \n" // kptr -= 24 * 16;
-
-                        "shrn   v20.4h, v20.4s, #16         \n"
-                        "shrn   v21.4h, v21.4s, #16         \n"
-                        "shrn   v22.4h, v22.4s, #16         \n"
-                        "shrn   v23.4h, v23.4s, #16         \n"
-
-                        "st1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%0], #32 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(kptr)          // %7
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1!, {d24-d31}      \n" // sum0 sum1 sum2 sum3
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%2 :64]!  \n" // r00 r01 r02 r03
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%2 :64] \n" // r04 r05 r06 r07
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[0]     \n"
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        "vmla.f32   q15, q11, d8[1]     \n"
-                        "vmla.f32   q12, q8, d3[0]      \n"
-                        "vmla.f32   q13, q8, d5[0]      \n"
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "vmla.f32   q14, q9, d7[1]      \n"
-                        "vmla.f32   q15, q9, d9[1]      \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[0]     \n"
-                        "vmla.f32   q13, q10, d8[0]     \n"
-                        "vmla.f32   q14, q10, d10[0]    \n"
-                        "vmla.f32   q15, q10, d12[0]    \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "vmla.f32   q14, q11, d10[1]    \n"
-                        "vmla.f32   q15, q11, d12[1]    \n"
-                        "vmla.f32   q12, q8, d7[0]      \n"
-                        "vmla.f32   q13, q8, d9[0]      \n"
-                        "vmla.f32   q14, q8, d11[0]     \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vmla.f32   q14, q9, d11[1]     \n"
-                        "vmla.f32   q15, q9, d13[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%3 :64]!  \n" // r10 r11 r12 r13
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d10[0]     \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d11[0]    \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d0[0]     \n"
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d6[0]     \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d2[1]     \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "vmla.f32   q15, q11, d6[1]     \n"
-                        "vmla.f32   q12, q8, d1[0]      \n"
-                        "vmla.f32   q13, q8, d3[0]      \n"
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vmla.f32   q13, q9, d3[1]      \n"
-                        "vmla.f32   q14, q9, d5[1]      \n"
-                        "vmla.f32   q15, q9, d7[1]      \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%3 :64] \n" // r14 r15 r16 r17
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d7[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d4[0]     \n"
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d10[0]    \n"
-                        "vmla.f32   q12, q11, d4[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "vmla.f32   q15, q11, d10[1]    \n"
-                        "vmla.f32   q12, q8, d5[0]      \n"
-                        "vmla.f32   q13, q8, d7[0]      \n"
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vmla.f32   q15, q8, d11[0]     \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vmla.f32   q14, q9, d9[1]      \n"
-                        "vmla.f32   q15, q9, d11[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d6[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d7[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%4 :64]!  \n" // r20 r21 r22 r23
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q12, q10, d8[0]     \n"
-                        "vmla.f32   q13, q10, d10[0]    \n"
-                        "vmla.f32   q14, q10, d12[0]    \n"
-                        "vmla.f32   q15, q10, d14[0]    \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "vmla.f32   q14, q11, d12[1]    \n"
-                        "vmla.f32   q15, q11, d14[1]    \n"
-                        "vmla.f32   q12, q8, d9[0]      \n"
-                        "vmla.f32   q13, q8, d11[0]     \n"
-                        "vmla.f32   q14, q8, d13[0]     \n"
-                        "vmla.f32   q15, q8, d15[0]     \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "vmla.f32   q14, q9, d13[1]     \n"
-                        "vmla.f32   q15, q9, d15[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%4 :64] \n" // r24 r25 r26 r27
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[0]     \n"
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        "vmla.f32   q15, q11, d8[1]     \n"
-                        "vmla.f32   q12, q8, d3[0]      \n"
-                        "vmla.f32   q13, q8, d5[0]      \n"
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "vmla.f32   q14, q9, d7[1]      \n"
-                        "vmla.f32   q15, q9, d9[1]      \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[0]     \n"
-                        "vmla.f32   q13, q10, d8[0]     \n"
-                        "vmla.f32   q14, q10, d10[0]    \n"
-                        "vmla.f32   q15, q10, d12[0]    \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "vmla.f32   q14, q11, d10[1]    \n"
-                        "vmla.f32   q15, q11, d12[1]    \n"
-                        "vmla.f32   q12, q8, d7[0]      \n"
-                        "vmla.f32   q13, q8, d9[0]      \n"
-                        "vmla.f32   q14, q8, d11[0]     \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vmla.f32   q14, q9, d11[1]     \n"
-                        "vmla.f32   q15, q9, d13[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%5 :64]!  \n" // r30 r31 r32 r33
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d10[0]     \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d11[0]    \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d0[0]     \n"
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d6[0]     \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d2[1]     \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "vmla.f32   q15, q11, d6[1]     \n"
-                        "vmla.f32   q12, q8, d1[0]      \n"
-                        "vmla.f32   q13, q8, d3[0]      \n"
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vmla.f32   q13, q9, d3[1]      \n"
-                        "vmla.f32   q14, q9, d5[1]      \n"
-                        "vmla.f32   q15, q9, d7[1]      \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%5 :64] \n" // r34 r35 r36 r37
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q8, d2[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q9, d8[1]      \n"
-                        "vmla.f32   q12, q10, d3[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d7[1]     \n"
-                        "vmla.f32   q15, q11, d9[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d4[0]     \n"
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d10[0]    \n"
-                        "vmla.f32   q12, q11, d4[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "vmla.f32   q15, q11, d10[1]    \n"
-                        "vmla.f32   q12, q8, d5[0]      \n"
-                        "vmla.f32   q13, q8, d7[0]      \n"
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vmla.f32   q15, q8, d11[0]     \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vmla.f32   q14, q9, d9[1]      \n"
-                        "vmla.f32   q15, q9, d11[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d6[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d10[0]     \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d10[1]     \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d7[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d11[0]    \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d11[1]    \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%6 :64]!  \n" // r40 r41 r42 r43
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q12, q10, d8[0]     \n"
-                        "vmla.f32   q13, q10, d10[0]    \n"
-                        "vmla.f32   q14, q10, d12[0]    \n"
-                        "vmla.f32   q15, q10, d14[0]    \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "vmla.f32   q14, q11, d12[1]    \n"
-                        "vmla.f32   q15, q11, d14[1]    \n"
-                        "vmla.f32   q12, q8, d9[0]      \n"
-                        "vmla.f32   q13, q8, d11[0]     \n"
-                        "vmla.f32   q14, q8, d13[0]     \n"
-                        "vmla.f32   q15, q8, d15[0]     \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "vmla.f32   q14, q9, d13[1]     \n"
-                        "vmla.f32   q15, q9, d15[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d5[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%6 :64] \n" // r44 r45 r46 r47
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[0]     \n"
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        "vmla.f32   q15, q11, d8[1]     \n"
-                        "vmla.f32   q12, q8, d3[0]      \n"
-                        "vmla.f32   q13, q8, d5[0]      \n"
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "vmla.f32   q14, q9, d7[1]      \n"
-                        "vmla.f32   q15, q9, d9[1]      \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d7[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[0]     \n"
-                        "vmla.f32   q13, q10, d8[0]     \n"
-                        "vmla.f32   q14, q10, d10[0]    \n"
-                        "vmla.f32   q15, q10, d12[0]    \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "vmla.f32   q14, q11, d10[1]    \n"
-                        "vmla.f32   q15, q11, d12[1]    \n"
-                        "vmla.f32   q12, q8, d7[0]      \n"
-                        "vmla.f32   q13, q8, d9[0]      \n"
-                        "vmla.f32   q14, q8, d11[0]     \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vmla.f32   q14, q9, d11[1]     \n"
-                        "vmla.f32   q15, q9, d13[1]     \n"
-
-                        //                         "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :64] \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d10[0]     \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d14[0]     \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d14[1]     \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d11[0]    \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d15[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d15[1]    \n"
-
-                        "sub        %7, %7, #768        \n" // kptr -= 24 * 16;
-
-                        "vshrn.u32  d24, q12, #16       \n"
-                        "vshrn.u32  d25, q13, #16       \n"
-                        "vshrn.u32  d26, q14, #16       \n"
-                        "vshrn.u32  d27, q15, #16       \n"
-
-                        "vst1.u16   {d24-d27}, [%0 :64]! \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(kptr)          // %7
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%2], #16   \n" // r00 r01
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v20.4s, v21.4s}, [%1], #32 \n" // sum0 sum1
-
-                        "fmul   v22.4s, v16.4s, v0.s[0]     \n"
-                        "fmul   v23.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h, v5.4h}, [%2] \n" // r02 r03 r04 r05
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%3], #16   \n" // r10 r11
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v26.4s, v0.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h, v5.4h}, [%3] \n" // r12 r13 r14 r15
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v27.4s, v1.s[3]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v1.s[0]     \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v23.4s, v16.4s, v2.s[0]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v22.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%4], #16   \n" // r20 r21
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h, v5.4h}, [%4] \n" // r22 r23 r24 r25
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%5], #16   \n" // r30 r31
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v26.4s, v0.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h, v5.4h}, [%5] \n" // r32 r33 r34 r35
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v27.4s, v1.s[3]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v1.s[0]     \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v23.4s, v16.4s, v2.s[0]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v22.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%6, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%6], #16   \n" // r40 r41
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v1.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h, v5.4h}, [%6] \n" // r42 r43 r44 r45
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v19.4s, v1.s[3]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        //                         "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7] \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-
-                        "fadd   v20.4s, v20.4s, v22.4s      \n"
-                        "fadd   v21.4s, v21.4s, v23.4s      \n"
-
-                        "sub    %7, %7, #768                \n" // kptr -= 24 * 16;
-
-                        "shrn   v20.4h, v20.4s, #16         \n"
-                        "shrn   v21.4h, v21.4s, #16         \n"
-
-                        "st1    {v20.4h, v21.4h}, [%0], #16 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(kptr)          // %7
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%2 :64]!  \n" // r00 r01
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d8-d11}, [%2 :64]  \n" // r02 r03 r04 r05
-
-                        "vshll.u16  q8, d20, #16        \n"
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.f32   {d24-d27}, [%1 :128]! \n" // sum0 sum1
-
-                        "vmul.f32   q14, q8, d0[0]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vmul.f32   q15, q8, d2[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vshll.u16  q2, d8, #16         \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d2[0]     \n"
-                        "vmla.f32   q15, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d3[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q3, d9, #16         \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%3 :64]!  \n" // r10 r11
-
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d8-d11}, [%3 :64]  \n" // r12 r13 r14 r15
-
-                        "vmla.f32   q14, q10, d0[0]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vmla.f32   q15, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d2[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d1[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vshll.u16  q2, d8, #16         \n"
-                        "vmla.f32   q13, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vshll.u16  q3, d9, #16         \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d4[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "pld        [%4, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%4 :64]!  \n" // r20 r21
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d10[0]    \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d11[0]     \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d8-d11}, [%4 :64]  \n" // r22 r23 r24 r25
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vmla.f32   q15, q8, d2[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vshll.u16  q2, d8, #16         \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d2[0]     \n"
-                        "vmla.f32   q15, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d3[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q3, d9, #16         \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "pld        [%5, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%5 :64]!  \n" // r30 r31
-
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d8-d11}, [%5 :64]  \n" // r32 r33 r34 r35
-
-                        "vmla.f32   q14, q10, d0[0]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vmla.f32   q15, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d2[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d1[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vshll.u16  q2, d8, #16         \n"
-                        "vmla.f32   q13, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vshll.u16  q3, d9, #16         \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d4[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "pld        [%6, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%6 :64]!  \n" // r40 r41
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d10[0]    \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d11[0]     \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d8-d11}, [%6 :64]  \n" // r42 r43 r44 r45
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vmla.f32   q15, q8, d2[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vshll.u16  q2, d8, #16         \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d2[0]     \n"
-                        "vmla.f32   q15, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d3[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q3, d9, #16         \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128] \n"
-
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-
-                        "vadd.f32   q12, q12, q14       \n"
-                        "vadd.f32   q13, q13, q15       \n"
-
-                        "sub        %7, %7, #768        \n" // kptr -= 24 * 16;
-
-                        "vshrn.u32  d24, q12, #16       \n"
-                        "vshrn.u32  d25, q13, #16       \n"
-
-                        "vst1.u16   {d24-d25}, [%0 :64]! \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(kptr)          // %7
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%2, #64]        \n"
-                        "ld1    {v0.4h}, [%2], #8           \n" // r00
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v1.4h, v2.4h, v3.4h, v4.4h}, [%2] \n" // r01 r02 r03 r04
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v20.4s}, [%1], #16         \n" // sum0
-
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmul   v21.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmul   v22.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmul   v23.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #64]        \n"
-                        "ld1    {v0.4h}, [%3], #8           \n" // r10
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v1.4h, v2.4h, v3.4h, v4.4h}, [%3] \n" // r11 r12 r13 r14
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v22.4s, v25.4s, v0.s[1]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v23.4s, v26.4s, v0.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v1.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #64]        \n"
-                        "ld1    {v0.4h}, [%4], #8           \n" // r20
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v1.4h, v2.4h, v3.4h, v4.4h}, [%4] \n" // r21 r22 r23 r24
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v22.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v23.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #64]        \n"
-                        "ld1    {v0.4h}, [%5], #8           \n" // r30
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v1.4h, v2.4h, v3.4h, v4.4h}, [%5] \n" // r31 r32 r33 r34
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v22.4s, v25.4s, v0.s[1]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v23.4s, v26.4s, v0.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v1.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%6, #64]        \n"
-                        "ld1    {v0.4h}, [%6], #8           \n" // r40
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v1.4h, v2.4h, v3.4h, v4.4h}, [%6] \n" // r41 r42 r43 r44
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "fmla   v22.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "fmla   v23.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        //                         "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7] \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-
-                        "fadd   v22.4s, v21.4s, v22.4s      \n"
-                        "fadd   v23.4s, v22.4s, v23.4s      \n"
-                        "fadd   v20.4s, v20.4s, v23.4s      \n"
-
-                        "sub    %7, %7, #768                \n" // kptr -= 24 * 16;
-
-                        "shrn   v20.4h, v20.4s, #16         \n"
-
-                        "st1    {v20.4h}, [%0], #8          \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(kptr)          // %7
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%2, #64]           \n"
-                        "vld1.u16   {d1}, [%2 :64]!     \n" // r00
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d24-d25}, [%1 :128]! \n" // sum0
-
-                        "vmul.f32   q13, q8, d0[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmul.f32   q14, q9, d0[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmul.f32   q15, q10, d1[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d6-d9}, [%2 :64]   \n" // r01 r02 r03 r04
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q1, d6, #16         \n"
-                        "vshll.u16  q2, d7, #16         \n"
-                        "vshll.u16  q3, d8, #16         \n"
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d2[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "pld        [%3, #64]           \n"
-                        "vld1.u16   {d1}, [%3 :64]!     \n" // r10
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d0[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d0[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d1[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d6-d9}, [%3 :64]   \n" // r11 r12 r13 r14
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q1, d6, #16         \n"
-                        "vshll.u16  q2, d7, #16         \n"
-                        "vshll.u16  q3, d8, #16         \n"
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "pld        [%4, #64]           \n"
-                        "vld1.u16   {d1}, [%4 :64]!     \n" // r20
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q13, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d6-d9}, [%4 :64]   \n" // r21 r22 r23 r24
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q1, d6, #16         \n"
-                        "vshll.u16  q2, d7, #16         \n"
-                        "vshll.u16  q3, d8, #16         \n"
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d2[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "pld        [%5, #64]           \n"
-                        "vld1.u16   {d1}, [%5 :64]!     \n" // r30
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d0[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d0[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d1[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d6-d9}, [%5 :64]   \n" // r31 r32 r33 r34
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q1, d6, #16         \n"
-                        "vshll.u16  q2, d7, #16         \n"
-                        "vshll.u16  q3, d8, #16         \n"
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "pld        [%6, #64]           \n"
-                        "vld1.u16   {d1}, [%6 :64]!     \n" // r40
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q0, d1, #16         \n"
-
-                        "vmla.f32   q13, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d6-d9}, [%6 :64]   \n" // r41 r42 r43 r44
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q1, d6, #16         \n"
-                        "vshll.u16  q2, d7, #16         \n"
-                        "vshll.u16  q3, d8, #16         \n"
-                        "vshll.u16  q4, d9, #16         \n"
-
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d2[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        //                         "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128] \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-
-                        "vadd.f32   q13, q13, q14       \n"
-                        "vadd.f32   q12, q12, q15       \n"
-                        "vadd.f32   q12, q12, q13       \n"
-
-                        "sub        %7, %7, #768        \n" // kptr -= 24 * 16;
-
-                        "vshrn.u32  d24, q12, #16       \n"
-
-                        "vst1.u16   {d24}, [%0 :64]!    \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(kptr)          // %7
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-
-                r0 += 4 * 4;
-                r1 += 4 * 4;
-                r2 += 4 * 4;
-                r3 += 4 * 4;
-                r4 += 4 * 4;
-            }
-        }
-    }
-}
-
-static void conv5x5s2_pack4_bf16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    Mat top_blob_fp32(outw, outh, opt.num_threads, (size_t)4u * 4, 4, opt.workspace_allocator);
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob_fp32.channel(get_omp_thread_num());
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        out0.fill(_bias0);
-
-        int q = 0;
-        for (; q < inch - 1; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<const unsigned short>(0);
-            const unsigned short* r1 = img0.row<const unsigned short>(1);
-            const unsigned short* r2 = img0.row<const unsigned short>(2);
-            const unsigned short* r3 = img0.row<const unsigned short>(3);
-            const unsigned short* r4 = img0.row<const unsigned short>(4);
-
-            const unsigned short* kptr = kernel.channel(p).row<const unsigned short>(q);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%1], #32 \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%1], #32 \n" // r04 r05 r06 r07
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0] \n" // sum0 sum1 sum2 sum3
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%1, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%1] \n" // r08 r09 r010
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v28.s[3]    \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v29.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v29.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v29.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v29.s[3]    \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n" // r10 r11 r12 r13
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v23.4s, v16.4s, v30.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v28.s[1]    \n"
-                        "fmla   v23.4s, v17.4s, v30.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v23.4s, v18.4s, v30.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v28.s[3]    \n"
-                        "fmla   v23.4s, v19.4s, v30.s[3]    \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%2], #32 \n" // r14 r15 r16 r17
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%2] \n" // r18 r19 r110
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v20.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v28.s[3]    \n"
-
-                        "fmla   v20.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v29.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v7.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v29.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v29.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v7.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v29.s[3]    \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%3], #32 \n" // r20 r21 r22 r23
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v28.s[0]    \n"
-                        "fmla   v23.4s, v24.4s, v30.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v28.s[1]    \n"
-                        "fmla   v23.4s, v25.4s, v30.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v28.s[2]    \n"
-                        "fmla   v23.4s, v26.4s, v30.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v28.s[3]    \n"
-                        "fmla   v23.4s, v27.4s, v30.s[3]    \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%3], #32 \n" // r24 r25 r26 r27
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%3] \n" // r28 r29 r210
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v28.s[3]    \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v29.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v29.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v29.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v29.s[3]    \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%4], #32 \n" // r30 r31 r32 r33
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v23.4s, v16.4s, v30.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v28.s[1]    \n"
-                        "fmla   v23.4s, v17.4s, v30.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v23.4s, v18.4s, v30.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v28.s[3]    \n"
-                        "fmla   v23.4s, v19.4s, v30.s[3]    \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%4], #32 \n" // r34 r35 r36 r37
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%4] \n" // r38 r39 r310
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v20.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v28.s[3]    \n"
-
-                        "fmla   v20.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v29.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v7.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v29.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v29.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v7.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v29.s[3]    \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%5], #32 \n" // r40 r41 r42 r43
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v28.s[0]    \n"
-                        "fmla   v23.4s, v24.4s, v30.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v28.s[1]    \n"
-                        "fmla   v23.4s, v25.4s, v30.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v28.s[2]    \n"
-                        "fmla   v23.4s, v26.4s, v30.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v28.s[3]    \n"
-                        "fmla   v23.4s, v27.4s, v30.s[3]    \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%5], #32 \n" // r44 r45 r46 r47
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%5] \n" // r48 r49 r410
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v28.s[3]    \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v29.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v29.s[1]    \n"
-
-                        //                         "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6] \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v29.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v29.s[3]    \n"
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v23.4s, v16.4s, v30.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v28.s[1]    \n"
-                        "fmla   v23.4s, v17.4s, v30.s[1]    \n"
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v23.4s, v18.4s, v30.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v28.s[3]    \n"
-                        "fmla   v23.4s, v19.4s, v30.s[3]    \n"
-
-                        "sub    %6, %6, #768                \n" // kptr -= 24 * 16;
-
-                        "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #512]          \n"
-                        "vldm       %0, {d24-d31}       \n" // sum0 sum1 sum2 sum3
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%1 :64]!  \n" // r00 r01 r02 r03
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%1 :64]! \n" // r04 r05 r06 r07
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[0]     \n"
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vmla.f32   q14, q10, d10[0]    \n"
-                        "vmla.f32   q15, q10, d14[0]    \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "vmla.f32   q14, q11, d10[1]    \n"
-                        "vmla.f32   q15, q11, d14[1]    \n"
-                        "vmla.f32   q12, q8, d3[0]      \n"
-                        "vmla.f32   q13, q8, d7[0]      \n"
-                        "vmla.f32   q14, q8, d11[0]     \n"
-                        "vmla.f32   q15, q8, d15[0]     \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vmla.f32   q14, q9, d11[1]     \n"
-                        "vmla.f32   q15, q9, d15[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%1 :64]!  \n" // r08 r09
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[0]     \n"
-                        "vmla.f32   q13, q10, d10[0]    \n"
-                        "vmla.f32   q14, q10, d14[0]    \n"
-                        "vmla.f32   q15, q10, d2[0]     \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "vmla.f32   q14, q11, d14[1]    \n"
-                        "vmla.f32   q15, q11, d2[1]     \n"
-                        "vmla.f32   q12, q8, d7[0]      \n"
-                        "vmla.f32   q13, q8, d11[0]     \n"
-                        "vmla.f32   q14, q8, d15[0]     \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "vmla.f32   q14, q9, d15[1]     \n"
-                        "vmla.f32   q15, q9, d3[1]      \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%1, #64]           \n"
-                        "vld1.u16   {d5}, [%1 :64]      \n" // r010
-
-                        "vshll.u16  q2, d5, #16         \n"
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d12[0]     \n"
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d13[0]    \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%2 :64]! \n" // r10 r11 r12 r13
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%2 :64]!  \n" // r14 r15 r16 r17
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q12, q10, d8[0]     \n"
-                        "vmla.f32   q13, q10, d12[0]    \n"
-                        "vmla.f32   q14, q10, d0[0]     \n"
-                        "vmla.f32   q15, q10, d4[0]     \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vmla.f32   q13, q11, d12[1]    \n"
-                        "vmla.f32   q14, q11, d0[1]     \n"
-                        "vmla.f32   q15, q11, d4[1]     \n"
-                        "vmla.f32   q12, q8, d9[0]      \n"
-                        "vmla.f32   q13, q8, d13[0]     \n"
-                        "vmla.f32   q14, q8, d1[0]      \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vmla.f32   q13, q9, d13[1]     \n"
-                        "vmla.f32   q14, q9, d1[1]      \n"
-                        "vmla.f32   q15, q9, d5[1]      \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d10[0]     \n"
-                        "vmla.f32   q13, q8, d14[0]     \n"
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d10[1]     \n"
-                        "vmla.f32   q13, q9, d14[1]     \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d11[0]    \n"
-                        "vmla.f32   q13, q10, d15[0]    \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d11[1]    \n"
-                        "vmla.f32   q13, q11, d15[1]    \n"
-                        "vmla.f32   q14, q11, d3[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d10-d11}, [%2 :64]! \n" // r18 r19
-
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vshll.u16  q5, d11, #16        \n"
-
-                        "vmla.f32   q12, q10, d12[0]    \n"
-                        "vmla.f32   q13, q10, d0[0]     \n"
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vmla.f32   q12, q11, d12[1]    \n"
-                        "vmla.f32   q13, q11, d0[1]     \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "vmla.f32   q15, q11, d8[1]     \n"
-                        "vmla.f32   q12, q8, d13[0]     \n"
-                        "vmla.f32   q13, q8, d1[0]      \n"
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vmla.f32   q12, q9, d13[1]     \n"
-                        "vmla.f32   q13, q9, d1[1]      \n"
-                        "vmla.f32   q14, q9, d5[1]      \n"
-                        "vmla.f32   q15, q9, d9[1]      \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d14[0]     \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d14[1]     \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d15[0]    \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d15[1]    \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d7[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%2, #64]           \n"
-                        "vld1.u16   {d13}, [%2 :64]     \n" // r110
-
-                        "vshll.u16  q6, d13, #16        \n"
-
-                        "vmla.f32   q12, q10, d0[0]     \n"
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d12[0]    \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "vmla.f32   q15, q11, d12[1]    \n"
-                        "vmla.f32   q12, q8, d1[0]      \n"
-                        "vmla.f32   q13, q8, d5[0]      \n"
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%3 :64]!  \n" // r20 r21 r22 r23
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q14, q9, d9[1]      \n"
-                        "vmla.f32   q15, q9, d13[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%3 :64]! \n" // r24 r25 r26 r27
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[0]     \n"
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vmla.f32   q14, q10, d10[0]    \n"
-                        "vmla.f32   q15, q10, d14[0]    \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "vmla.f32   q14, q11, d10[1]    \n"
-                        "vmla.f32   q15, q11, d14[1]    \n"
-                        "vmla.f32   q12, q8, d3[0]      \n"
-                        "vmla.f32   q13, q8, d7[0]      \n"
-                        "vmla.f32   q14, q8, d11[0]     \n"
-                        "vmla.f32   q15, q8, d15[0]     \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vmla.f32   q14, q9, d11[1]     \n"
-                        "vmla.f32   q15, q9, d15[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%3 :64]!  \n" // r28 r29
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[0]     \n"
-                        "vmla.f32   q13, q10, d10[0]    \n"
-                        "vmla.f32   q14, q10, d14[0]    \n"
-                        "vmla.f32   q15, q10, d2[0]     \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "vmla.f32   q14, q11, d14[1]    \n"
-                        "vmla.f32   q15, q11, d2[1]     \n"
-                        "vmla.f32   q12, q8, d7[0]      \n"
-                        "vmla.f32   q13, q8, d11[0]     \n"
-                        "vmla.f32   q14, q8, d15[0]     \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "vmla.f32   q14, q9, d15[1]     \n"
-                        "vmla.f32   q15, q9, d3[1]      \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%3, #64]           \n"
-                        "vld1.u16   {d5}, [%3 :64]      \n" // r210
-
-                        "vshll.u16  q2, d5, #16         \n"
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d12[0]     \n"
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d13[0]    \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%4 :64]! \n" // r30 r31 r32 r33
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%4 :64]!  \n" // r34 r35 r36 r37
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q12, q10, d8[0]     \n"
-                        "vmla.f32   q13, q10, d12[0]    \n"
-                        "vmla.f32   q14, q10, d0[0]     \n"
-                        "vmla.f32   q15, q10, d4[0]     \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vmla.f32   q13, q11, d12[1]    \n"
-                        "vmla.f32   q14, q11, d0[1]     \n"
-                        "vmla.f32   q15, q11, d4[1]     \n"
-                        "vmla.f32   q12, q8, d9[0]      \n"
-                        "vmla.f32   q13, q8, d13[0]     \n"
-                        "vmla.f32   q14, q8, d1[0]      \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vmla.f32   q13, q9, d13[1]     \n"
-                        "vmla.f32   q14, q9, d1[1]      \n"
-                        "vmla.f32   q15, q9, d5[1]      \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d10[0]     \n"
-                        "vmla.f32   q13, q8, d14[0]     \n"
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d10[1]     \n"
-                        "vmla.f32   q13, q9, d14[1]     \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d11[0]    \n"
-                        "vmla.f32   q13, q10, d15[0]    \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d11[1]    \n"
-                        "vmla.f32   q13, q11, d15[1]    \n"
-                        "vmla.f32   q14, q11, d3[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.u16   {d10-d11}, [%4 :64]! \n" // r38 r39
-
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vshll.u16  q5, d11, #16        \n"
-
-                        "vmla.f32   q12, q10, d12[0]    \n"
-                        "vmla.f32   q13, q10, d0[0]     \n"
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vmla.f32   q12, q11, d12[1]    \n"
-                        "vmla.f32   q13, q11, d0[1]     \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "vmla.f32   q15, q11, d8[1]     \n"
-                        "vmla.f32   q12, q8, d13[0]     \n"
-                        "vmla.f32   q13, q8, d1[0]      \n"
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vmla.f32   q12, q9, d13[1]     \n"
-                        "vmla.f32   q13, q9, d1[1]      \n"
-                        "vmla.f32   q14, q9, d5[1]      \n"
-                        "vmla.f32   q15, q9, d9[1]      \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d14[0]     \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d14[1]     \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d15[0]    \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d15[1]    \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d7[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%4, #64]           \n"
-                        "vld1.u16   {d13}, [%4 :64]     \n" // r310
-
-                        "vshll.u16  q6, d13, #16        \n"
-
-                        "vmla.f32   q12, q10, d0[0]     \n"
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d12[0]    \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "vmla.f32   q15, q11, d12[1]    \n"
-                        "vmla.f32   q12, q8, d1[0]      \n"
-                        "vmla.f32   q13, q8, d5[0]      \n"
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%5 :64]!  \n" // r40 r41 r42 r43
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q14, q9, d9[1]      \n"
-                        "vmla.f32   q15, q9, d13[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%5 :64]! \n" // r44 r45 r46 r47
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[0]     \n"
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vmla.f32   q14, q10, d10[0]    \n"
-                        "vmla.f32   q15, q10, d14[0]    \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "vmla.f32   q14, q11, d10[1]    \n"
-                        "vmla.f32   q15, q11, d14[1]    \n"
-                        "vmla.f32   q12, q8, d3[0]      \n"
-                        "vmla.f32   q13, q8, d7[0]      \n"
-                        "vmla.f32   q14, q8, d11[0]     \n"
-                        "vmla.f32   q15, q8, d15[0]     \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vmla.f32   q14, q9, d11[1]     \n"
-                        "vmla.f32   q15, q9, d15[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%5, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%5 :64]!  \n" // r48 r49
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[0]     \n"
-                        "vmla.f32   q13, q10, d10[0]    \n"
-                        "vmla.f32   q14, q10, d14[0]    \n"
-                        "vmla.f32   q15, q10, d2[0]     \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "vmla.f32   q14, q11, d14[1]    \n"
-                        "vmla.f32   q15, q11, d2[1]     \n"
-                        "vmla.f32   q12, q8, d7[0]      \n"
-                        "vmla.f32   q13, q8, d11[0]     \n"
-                        "vmla.f32   q14, q8, d15[0]     \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "vmla.f32   q14, q9, d15[1]     \n"
-                        "vmla.f32   q15, q9, d3[1]      \n"
-
-                        //                         "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128] \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%5, #64]           \n"
-                        "vld1.u16   {d5}, [%5 :64]      \n" // r410
-
-                        "vshll.u16  q2, d5, #16         \n"
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d12[0]     \n"
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d13[0]    \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "sub        %6, %6, #768        \n" // kptr -= 24 * 16;
-
-                        "sub        %1, %1, #16         \n"
-                        "sub        %2, %2, #16         \n"
-                        "sub        %3, %3, #16         \n"
-                        "sub        %4, %4, #16         \n"
-                        "sub        %5, %5, #16         \n"
-
-                        "vstm       %0!, {d24-d31}      \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%1], #32 \n" // r00 r01 r02 r03
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v20.4s, v21.4s}, [%0]      \n" // sum0 sum1
-
-                        "fmul   v22.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmul   v23.4s, v16.4s, v2.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "prfm   pldl1keep, [%1, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%1] \n" // r04 r05 r06
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n" // r10 r11 r12 r13
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "prfm   pldl1keep, [%2, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%2] \n" // r14 r15 r16
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v22.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%3], #32 \n" // r20 r21 r22 r23
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v6.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v6.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v16.4s, v2.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%3] \n" // r24 r25 r26
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%4], #32 \n" // r30 r31 r32 r33
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%4] \n" // r34 r35 r36
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v22.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%5], #32 \n" // r40 r41 r42 r43
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v6.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v6.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v16.4s, v2.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%5] \n" // r44 r45 r46
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        //                         "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6] \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-
-                        "fadd   v20.4s, v20.4s, v22.4s      \n"
-                        "fadd   v21.4s, v21.4s, v23.4s      \n"
-
-                        "sub    %6, %6, #768                \n" // kptr -= 24 * 16;
-
-                        "st1    {v20.4s, v21.4s}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%1 :64]!  \n" // r00 r01 r02 r03
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "pld        [%0, #256]          \n"
-                        "vld1.f32   {d24-d27}, [%0 :128] \n" // sum0 sum1
-
-                        "vmul.f32   q14, q8, d0[0]      \n"
-                        "vmul.f32   q15, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "pld        [%1, #192]          \n"
-                        "vld1.u16   {d10-d12}, [%1 :64] \n" // r04 r05 r06
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vshll.u16  q6, d12, #16        \n"
-
-                        "vmla.f32   q14, q10, d2[0]     \n"
-                        "vmla.f32   q15, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d3[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vmla.f32   q15, q10, d10[0]    \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d11[0]     \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%2 :64]!  \n" // r10 r11 r12 r13
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d0[0]     \n"
-                        "vmla.f32   q15, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d1[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "pld        [%2, #192]          \n"
-                        "vld1.u16   {d10-d12}, [%2 :64] \n" // r14 r15 r16
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vshll.u16  q6, d12, #16        \n"
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d4[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%3 :64]!  \n" // r20 r21 r22 r23
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d12[0]    \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vmla.f32   q13, q11, d12[1]    \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vmla.f32   q13, q9, d13[1]     \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "pld        [%3, #192]          \n"
-                        "vld1.u16   {d10-d12}, [%3 :64] \n" // r24 r25 r26
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vshll.u16  q6, d12, #16        \n"
-
-                        "vmla.f32   q14, q10, d2[0]     \n"
-                        "vmla.f32   q15, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d3[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vmla.f32   q15, q10, d10[0]    \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d11[0]     \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%4 :64]!  \n" // r30 r31 r32 r33
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d0[0]     \n"
-                        "vmla.f32   q15, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d1[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "pld        [%4, #192]          \n"
-                        "vld1.u16   {d10-d12}, [%4 :64] \n" // r34 r35 r36
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vshll.u16  q6, d12, #16        \n"
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d4[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%5 :64]!  \n" // r40 r41 r42 r43
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d12[0]    \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vmla.f32   q13, q11, d12[1]    \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vmla.f32   q13, q9, d13[1]     \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "pld        [%5, #192]          \n"
-                        "vld1.u16   {d10-d12}, [%5 :64] \n" // r44 r45 r46
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vshll.u16  q6, d12, #16        \n"
-
-                        "vmla.f32   q14, q10, d2[0]     \n"
-                        "vmla.f32   q15, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q8, d3[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vmla.f32   q15, q10, d10[0]    \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        //                         "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128] \n"
-
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d11[0]     \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-
-                        "vadd.f32   q12, q12, q14       \n"
-                        "vadd.f32   q13, q13, q15       \n"
-
-                        "sub        %6, %6, #768        \n" // kptr -= 24 * 16;
-
-                        "vst1.f32   {d24-d27}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v20.4s}, [%0]              \n" // sum0
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%1], #16   \n" // r00 r01
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmul   v21.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmul   v22.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmul   v23.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%1, #192]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h}, [%1] \n" // r02 r03 r04
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%2], #16   \n" // r10 r11
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v0.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v0.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%2, #192]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h}, [%2] \n" // r12 r13 r14
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v1.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%3], #16   \n" // r20 r21
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #192]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h}, [%3] \n" // r22 r23 r24
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%4], #16   \n" // r30 r31
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v0.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v0.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #192]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h}, [%4] \n" // r32 r33 r34
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v1.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%5], #16   \n" // r40 r41
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #192]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h}, [%5] \n" // r42 r43 r44
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%6], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        //                         "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6] \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-
-                        "fadd   v22.4s, v21.4s, v22.4s      \n"
-                        "fadd   v23.4s, v22.4s, v23.4s      \n"
-                        "fadd   v20.4s, v20.4s, v23.4s      \n"
-
-                        "sub    %6, %6, #768                \n" // kptr -= 24 * 16;
-
-                        "st1    {v20.4s}, [%0], #16         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%1 :64]!  \n" // r00 r01
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "pld        [%0, #128]          \n"
-                        "vld1.f32   {d24-d25}, [%0 :128] \n" // sum0
-
-                        "vmul.f32   q13, q8, d0[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmul.f32   q14, q9, d0[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmul.f32   q15, q10, d1[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "pld        [%1, #192]          \n"
-                        "vld1.u16   {d6-d8}, [%1 :64]   \n" // r02 r03 r04
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshll.u16  q4, d8, #16         \n"
-
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d2[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%2 :64]!  \n" // r10 r11
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d0[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d0[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d1[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "pld        [%2, #192]          \n"
-                        "vld1.u16   {d6-d8}, [%2 :64]   \n" // r12 r13 r14
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshll.u16  q4, d8, #16         \n"
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%3 :64]!  \n" // r20 r21
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q13, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "pld        [%3, #192]          \n"
-                        "vld1.u16   {d6-d8}, [%3 :64]   \n" // r22 r23 r24
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshll.u16  q4, d8, #16         \n"
-
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d2[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "pld        [%4, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%4 :64]!  \n" // r30 r31
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d0[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d0[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d1[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "pld        [%4, #192]          \n"
-                        "vld1.u16   {d6-d8}, [%4 :64]   \n" // r32 r33 r34
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshll.u16  q4, d8, #16         \n"
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "pld        [%5, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%5 :64]!  \n" // r40 r41
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q13, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "pld        [%5, #192]          \n"
-                        "vld1.u16   {d6-d8}, [%5 :64]   \n" // r42 r43 r44
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshll.u16  q4, d8, #16         \n"
-
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d2[1]     \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%6 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        //                         "pld        [%6, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%6 :128] \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-
-                        "vadd.f32   q14, q13, q14       \n"
-                        "vadd.f32   q15, q14, q15       \n"
-                        "vadd.f32   q12, q12, q15       \n"
-
-                        "sub        %6, %6, #768        \n" // kptr -= 24 * 16;
-
-                        "vst1.f32   {d24-d25}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-                r4 += tailstep;
-            }
-        }
-        for (; q < inch; q++)
-        {
-            unsigned short* outptr0_bf16 = top_blob.channel(p);
-
-            const float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<const unsigned short>(0);
-            const unsigned short* r1 = img0.row<const unsigned short>(1);
-            const unsigned short* r2 = img0.row<const unsigned short>(2);
-            const unsigned short* r3 = img0.row<const unsigned short>(3);
-            const unsigned short* r4 = img0.row<const unsigned short>(4);
-
-            const unsigned short* kptr = kernel.channel(p).row<const unsigned short>(q);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%2], #32 \n" // r04 r05 r06 r07
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%1], #64 \n" // sum0 sum1 sum2 sum3
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%2] \n" // r08 r09 r010
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v28.s[3]    \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v29.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v29.s[1]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v29.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v29.s[3]    \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%3], #32 \n" // r10 r11 r12 r13
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v23.4s, v16.4s, v30.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v28.s[1]    \n"
-                        "fmla   v23.4s, v17.4s, v30.s[1]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v23.4s, v18.4s, v30.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v28.s[3]    \n"
-                        "fmla   v23.4s, v19.4s, v30.s[3]    \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%3], #32 \n" // r14 r15 r16 r17
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%3] \n" // r18 r19 r110
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v20.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v28.s[3]    \n"
-
-                        "fmla   v20.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v29.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v7.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v29.s[1]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v29.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v7.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v29.s[3]    \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%4], #32 \n" // r20 r21 r22 r23
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v28.s[0]    \n"
-                        "fmla   v23.4s, v24.4s, v30.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v28.s[1]    \n"
-                        "fmla   v23.4s, v25.4s, v30.s[1]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v28.s[2]    \n"
-                        "fmla   v23.4s, v26.4s, v30.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v28.s[3]    \n"
-                        "fmla   v23.4s, v27.4s, v30.s[3]    \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%4], #32 \n" // r24 r25 r26 r27
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%4] \n" // r28 r29 r210
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v28.s[3]    \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v29.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v29.s[1]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v29.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v29.s[3]    \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%5], #32 \n" // r30 r31 r32 r33
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v23.4s, v16.4s, v30.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v28.s[1]    \n"
-                        "fmla   v23.4s, v17.4s, v30.s[1]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v23.4s, v18.4s, v30.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v28.s[3]    \n"
-                        "fmla   v23.4s, v19.4s, v30.s[3]    \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%5], #32 \n" // r34 r35 r36 r37
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%5] \n" // r38 r39 r310
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v20.4s, v16.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v28.s[3]    \n"
-
-                        "fmla   v20.4s, v16.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v7.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v29.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v7.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v29.s[1]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v7.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v29.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v7.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v29.s[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%6], #32 \n" // r40 r41 r42 r43
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v20.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v28.s[0]    \n"
-                        "fmla   v23.4s, v24.4s, v30.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v28.s[1]    \n"
-                        "fmla   v23.4s, v25.4s, v30.s[1]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v28.s[2]    \n"
-                        "fmla   v23.4s, v26.4s, v30.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v28.s[3]    \n"
-                        "fmla   v23.4s, v27.4s, v30.s[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%6], #32 \n" // r44 r45 r46 r47
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v6.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v6.s[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%6] \n" // r48 r49 r410
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v20.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v7.s[1]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v7.s[3]     \n"
-
-                        "fmla   v20.4s, v16.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v23.4s, v17.4s, v28.s[1]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v20.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v23.4s, v19.4s, v28.s[3]    \n"
-
-                        "fmla   v20.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v22.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v29.s[0]    \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v22.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v29.s[1]    \n"
-
-                        //                         "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7] \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v20.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v22.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v29.s[2]    \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v22.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v29.s[3]    \n"
-
-                        "fmla   v20.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v21.4s, v16.4s, v6.s[0]     \n"
-                        "fmla   v22.4s, v16.4s, v28.s[0]    \n"
-                        "fmla   v23.4s, v16.4s, v30.s[0]    \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "fmla   v22.4s, v17.4s, v28.s[1]    \n"
-                        "fmla   v23.4s, v17.4s, v30.s[1]    \n"
-                        "fmla   v20.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v21.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v22.4s, v18.4s, v28.s[2]    \n"
-                        "fmla   v23.4s, v18.4s, v30.s[2]    \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-                        "fmla   v22.4s, v19.4s, v28.s[3]    \n"
-                        "fmla   v23.4s, v19.4s, v30.s[3]    \n"
-
-                        "sub    %7, %7, #768                \n" // kptr -= 24 * 16;
-
-                        "shrn   v20.4h, v20.4s, #16         \n"
-                        "shrn   v21.4h, v21.4s, #16         \n"
-                        "shrn   v22.4h, v22.4s, #16         \n"
-                        "shrn   v23.4h, v23.4s, #16         \n"
-
-                        "st1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%0], #32 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(kptr)          // %7
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1!, {d24-d31}      \n" // sum0 sum1 sum2 sum3
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%2 :64]!  \n" // r00 r01 r02 r03
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%2 :64]! \n" // r04 r05 r06 r07
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[0]     \n"
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vmla.f32   q14, q10, d10[0]    \n"
-                        "vmla.f32   q15, q10, d14[0]    \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "vmla.f32   q14, q11, d10[1]    \n"
-                        "vmla.f32   q15, q11, d14[1]    \n"
-                        "vmla.f32   q12, q8, d3[0]      \n"
-                        "vmla.f32   q13, q8, d7[0]      \n"
-                        "vmla.f32   q14, q8, d11[0]     \n"
-                        "vmla.f32   q15, q8, d15[0]     \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vmla.f32   q14, q9, d11[1]     \n"
-                        "vmla.f32   q15, q9, d15[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%2 :64]!  \n" // r08 r09
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[0]     \n"
-                        "vmla.f32   q13, q10, d10[0]    \n"
-                        "vmla.f32   q14, q10, d14[0]    \n"
-                        "vmla.f32   q15, q10, d2[0]     \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "vmla.f32   q14, q11, d14[1]    \n"
-                        "vmla.f32   q15, q11, d2[1]     \n"
-                        "vmla.f32   q12, q8, d7[0]      \n"
-                        "vmla.f32   q13, q8, d11[0]     \n"
-                        "vmla.f32   q14, q8, d15[0]     \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "vmla.f32   q14, q9, d15[1]     \n"
-                        "vmla.f32   q15, q9, d3[1]      \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%2, #64]           \n"
-                        "vld1.u16   {d5}, [%2 :64]      \n" // r010
-
-                        "vshll.u16  q2, d5, #16         \n"
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d12[0]     \n"
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d13[0]    \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%3 :64]! \n" // r10 r11 r12 r13
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%3 :64]!  \n" // r14 r15 r16 r17
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q12, q10, d8[0]     \n"
-                        "vmla.f32   q13, q10, d12[0]    \n"
-                        "vmla.f32   q14, q10, d0[0]     \n"
-                        "vmla.f32   q15, q10, d4[0]     \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vmla.f32   q13, q11, d12[1]    \n"
-                        "vmla.f32   q14, q11, d0[1]     \n"
-                        "vmla.f32   q15, q11, d4[1]     \n"
-                        "vmla.f32   q12, q8, d9[0]      \n"
-                        "vmla.f32   q13, q8, d13[0]     \n"
-                        "vmla.f32   q14, q8, d1[0]      \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vmla.f32   q13, q9, d13[1]     \n"
-                        "vmla.f32   q14, q9, d1[1]      \n"
-                        "vmla.f32   q15, q9, d5[1]      \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d10[0]     \n"
-                        "vmla.f32   q13, q8, d14[0]     \n"
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d10[1]     \n"
-                        "vmla.f32   q13, q9, d14[1]     \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d11[0]    \n"
-                        "vmla.f32   q13, q10, d15[0]    \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d11[1]    \n"
-                        "vmla.f32   q13, q11, d15[1]    \n"
-                        "vmla.f32   q14, q11, d3[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d10-d11}, [%3 :64]! \n" // r18 r19
-
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vshll.u16  q5, d11, #16        \n"
-
-                        "vmla.f32   q12, q10, d12[0]    \n"
-                        "vmla.f32   q13, q10, d0[0]     \n"
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vmla.f32   q12, q11, d12[1]    \n"
-                        "vmla.f32   q13, q11, d0[1]     \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "vmla.f32   q15, q11, d8[1]     \n"
-                        "vmla.f32   q12, q8, d13[0]     \n"
-                        "vmla.f32   q13, q8, d1[0]      \n"
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vmla.f32   q12, q9, d13[1]     \n"
-                        "vmla.f32   q13, q9, d1[1]      \n"
-                        "vmla.f32   q14, q9, d5[1]      \n"
-                        "vmla.f32   q15, q9, d9[1]      \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d14[0]     \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d14[1]     \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d15[0]    \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d15[1]    \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d7[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%3, #64]           \n"
-                        "vld1.u16   {d13}, [%3 :64]     \n" // r110
-
-                        "vshll.u16  q6, d13, #16        \n"
-
-                        "vmla.f32   q12, q10, d0[0]     \n"
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d12[0]    \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "vmla.f32   q15, q11, d12[1]    \n"
-                        "vmla.f32   q12, q8, d1[0]      \n"
-                        "vmla.f32   q13, q8, d5[0]      \n"
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%4 :64]!  \n" // r20 r21 r22 r23
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q14, q9, d9[1]      \n"
-                        "vmla.f32   q15, q9, d13[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%4 :64]! \n" // r24 r25 r26 r27
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[0]     \n"
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vmla.f32   q14, q10, d10[0]    \n"
-                        "vmla.f32   q15, q10, d14[0]    \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "vmla.f32   q14, q11, d10[1]    \n"
-                        "vmla.f32   q15, q11, d14[1]    \n"
-                        "vmla.f32   q12, q8, d3[0]      \n"
-                        "vmla.f32   q13, q8, d7[0]      \n"
-                        "vmla.f32   q14, q8, d11[0]     \n"
-                        "vmla.f32   q15, q8, d15[0]     \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vmla.f32   q14, q9, d11[1]     \n"
-                        "vmla.f32   q15, q9, d15[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%4 :64]!  \n" // r28 r29
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[0]     \n"
-                        "vmla.f32   q13, q10, d10[0]    \n"
-                        "vmla.f32   q14, q10, d14[0]    \n"
-                        "vmla.f32   q15, q10, d2[0]     \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "vmla.f32   q14, q11, d14[1]    \n"
-                        "vmla.f32   q15, q11, d2[1]     \n"
-                        "vmla.f32   q12, q8, d7[0]      \n"
-                        "vmla.f32   q13, q8, d11[0]     \n"
-                        "vmla.f32   q14, q8, d15[0]     \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "vmla.f32   q14, q9, d15[1]     \n"
-                        "vmla.f32   q15, q9, d3[1]      \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%4, #64]           \n"
-                        "vld1.u16   {d5}, [%4 :64]      \n" // r210
-
-                        "vshll.u16  q2, d5, #16         \n"
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d12[0]     \n"
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d13[0]    \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%5 :64]! \n" // r30 r31 r32 r33
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%5 :64]!  \n" // r34 r35 r36 r37
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q12, q10, d8[0]     \n"
-                        "vmla.f32   q13, q10, d12[0]    \n"
-                        "vmla.f32   q14, q10, d0[0]     \n"
-                        "vmla.f32   q15, q10, d4[0]     \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vmla.f32   q13, q11, d12[1]    \n"
-                        "vmla.f32   q14, q11, d0[1]     \n"
-                        "vmla.f32   q15, q11, d4[1]     \n"
-                        "vmla.f32   q12, q8, d9[0]      \n"
-                        "vmla.f32   q13, q8, d13[0]     \n"
-                        "vmla.f32   q14, q8, d1[0]      \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vmla.f32   q13, q9, d13[1]     \n"
-                        "vmla.f32   q14, q9, d1[1]      \n"
-                        "vmla.f32   q15, q9, d5[1]      \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d10[0]     \n"
-                        "vmla.f32   q13, q8, d14[0]     \n"
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vmla.f32   q12, q9, d10[1]     \n"
-                        "vmla.f32   q13, q9, d14[1]     \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "vmla.f32   q15, q9, d6[1]      \n"
-                        "vmla.f32   q12, q10, d11[0]    \n"
-                        "vmla.f32   q13, q10, d15[0]    \n"
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vmla.f32   q12, q11, d11[1]    \n"
-                        "vmla.f32   q13, q11, d15[1]    \n"
-                        "vmla.f32   q14, q11, d3[1]     \n"
-                        "vmla.f32   q15, q11, d7[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%5, #128]          \n"
-                        "vld1.u16   {d10-d11}, [%5 :64]! \n" // r38 r39
-
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vshll.u16  q5, d11, #16        \n"
-
-                        "vmla.f32   q12, q10, d12[0]    \n"
-                        "vmla.f32   q13, q10, d0[0]     \n"
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vmla.f32   q12, q11, d12[1]    \n"
-                        "vmla.f32   q13, q11, d0[1]     \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "vmla.f32   q15, q11, d8[1]     \n"
-                        "vmla.f32   q12, q8, d13[0]     \n"
-                        "vmla.f32   q13, q8, d1[0]      \n"
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vmla.f32   q12, q9, d13[1]     \n"
-                        "vmla.f32   q13, q9, d1[1]      \n"
-                        "vmla.f32   q14, q9, d5[1]      \n"
-                        "vmla.f32   q15, q9, d9[1]      \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q12, q8, d14[0]     \n"
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vmla.f32   q12, q9, d14[1]     \n"
-                        "vmla.f32   q13, q9, d2[1]      \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "vmla.f32   q15, q9, d10[1]     \n"
-                        "vmla.f32   q12, q10, d15[0]    \n"
-                        "vmla.f32   q13, q10, d3[0]     \n"
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vmla.f32   q12, q11, d15[1]    \n"
-                        "vmla.f32   q13, q11, d3[1]     \n"
-                        "vmla.f32   q14, q11, d7[1]     \n"
-                        "vmla.f32   q15, q11, d11[1]    \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "pld        [%5, #64]           \n"
-                        "vld1.u16   {d13}, [%5 :64]     \n" // r310
-
-                        "vshll.u16  q6, d13, #16        \n"
-
-                        "vmla.f32   q12, q10, d0[0]     \n"
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d12[0]    \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "vmla.f32   q15, q11, d12[1]    \n"
-                        "vmla.f32   q12, q8, d1[0]      \n"
-                        "vmla.f32   q13, q8, d5[0]      \n"
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%6 :64]!  \n" // r40 r41 r42 r43
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q14, q9, d9[1]      \n"
-                        "vmla.f32   q15, q9, d13[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d12-d15}, [%6 :64]! \n" // r44 r45 r46 r47
-
-                        "vshll.u16  q4, d12, #16        \n"
-                        "vshll.u16  q5, d13, #16        \n"
-                        "vshll.u16  q6, d14, #16        \n"
-                        "vshll.u16  q7, d15, #16        \n"
-
-                        "vmla.f32   q12, q8, d0[0]      \n"
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vmla.f32   q15, q9, d12[1]     \n"
-                        "vmla.f32   q12, q10, d1[0]     \n"
-                        "vmla.f32   q13, q10, d5[0]     \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "vmla.f32   q14, q11, d9[1]     \n"
-                        "vmla.f32   q15, q11, d13[1]    \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[0]     \n"
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vmla.f32   q14, q10, d10[0]    \n"
-                        "vmla.f32   q15, q10, d14[0]    \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "vmla.f32   q14, q11, d10[1]    \n"
-                        "vmla.f32   q15, q11, d14[1]    \n"
-                        "vmla.f32   q12, q8, d3[0]      \n"
-                        "vmla.f32   q13, q8, d7[0]      \n"
-                        "vmla.f32   q14, q8, d11[0]     \n"
-                        "vmla.f32   q15, q8, d15[0]     \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vmla.f32   q14, q9, d11[1]     \n"
-                        "vmla.f32   q15, q9, d15[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%6, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%6 :64]!  \n" // r48 r49
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q12, q8, d4[0]      \n"
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q8, d12[0]     \n"
-                        "vmla.f32   q15, q8, d0[0]      \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "vmla.f32   q14, q9, d12[1]     \n"
-                        "vmla.f32   q15, q9, d0[1]      \n"
-                        "vmla.f32   q12, q10, d5[0]     \n"
-                        "vmla.f32   q13, q10, d9[0]     \n"
-                        "vmla.f32   q14, q10, d13[0]    \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vmla.f32   q14, q11, d13[1]    \n"
-                        "vmla.f32   q15, q11, d1[1]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vshll.u16  q9, d19, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[0]     \n"
-                        "vmla.f32   q13, q10, d10[0]    \n"
-                        "vmla.f32   q14, q10, d14[0]    \n"
-                        "vmla.f32   q15, q10, d2[0]     \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "vmla.f32   q14, q11, d14[1]    \n"
-                        "vmla.f32   q15, q11, d2[1]     \n"
-                        "vmla.f32   q12, q8, d7[0]      \n"
-                        "vmla.f32   q13, q8, d11[0]     \n"
-                        "vmla.f32   q14, q8, d15[0]     \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "vmla.f32   q14, q9, d15[1]     \n"
-                        "vmla.f32   q15, q9, d3[1]      \n"
-
-                        //                         "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128] \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "pld        [%6, #64]           \n"
-                        "vld1.u16   {d5}, [%6 :64]      \n" // r410
-
-                        "vshll.u16  q2, d5, #16         \n"
-
-                        "vmla.f32   q12, q8, d8[0]      \n"
-                        "vmla.f32   q13, q8, d12[0]     \n"
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "vmla.f32   q15, q9, d4[1]      \n"
-                        "vmla.f32   q12, q10, d9[0]     \n"
-                        "vmla.f32   q13, q10, d13[0]    \n"
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-                        "vmla.f32   q14, q11, d1[1]     \n"
-                        "vmla.f32   q15, q11, d5[1]     \n"
-
-                        "sub        %7, %7, #768        \n" // kptr -= 24 * 16;
-
-                        "sub        %2, %2, #16         \n"
-                        "sub        %3, %3, #16         \n"
-                        "sub        %4, %4, #16         \n"
-                        "sub        %5, %5, #16         \n"
-                        "sub        %6, %6, #16         \n"
-
-                        "vshrn.u32  d24, q12, #16       \n"
-                        "vshrn.u32  d25, q13, #16       \n"
-                        "vshrn.u32  d26, q14, #16       \n"
-                        "vshrn.u32  d27, q15, #16       \n"
-
-                        "vst1.u16   {d24-d27}, [%0 :64]! \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(kptr)          // %7
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n" // r00 r01 r02 r03
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v20.4s, v21.4s}, [%1], #32 \n" // sum0 sum1
-
-                        "fmul   v22.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "fmul   v23.4s, v16.4s, v2.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "prfm   pldl1keep, [%2, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%2] \n" // r04 r05 r06
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%3], #32 \n" // r10 r11 r12 r13
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%3] \n" // r14 r15 r16
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v22.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%4], #32 \n" // r20 r21 r22 r23
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v6.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v6.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v16.4s, v2.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%4] \n" // r24 r25 r26
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%5], #32 \n" // r30 r31 r32 r33
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v24.4s, v2.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v2.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%5] \n" // r34 r35 r36
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v22.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v16.4s, v3.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v3.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v24.4s, v4.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v4.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v16.4s, v5.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v5.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v5.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v5.s[3]     \n"
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%6], #32 \n" // r40 r41 r42 r43
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v22.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v24.4s, v6.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v6.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v6.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v6.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v16.4s, v2.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v0.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v0.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v2.s[3]     \n"
-                        "prfm   pldl1keep, [%6, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%6] \n" // r44 r45 r46
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v22.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v24.4s, v3.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "fmla   v23.4s, v16.4s, v4.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-
-                        "fmla   v22.4s, v18.4s, v2.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v4.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        //                         "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7] \n"
-                        "fmla   v23.4s, v24.4s, v5.s[0]     \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v20.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v5.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v5.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v5.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v22.4s, v16.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v16.4s, v6.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v20.4s, v17.4s, v4.s[1]     \n"
-                        "fmla   v21.4s, v17.4s, v6.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v22.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v18.4s, v6.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "fmla   v21.4s, v19.4s, v6.s[3]     \n"
-
-                        "fadd   v20.4s, v20.4s, v22.4s      \n"
-                        "fadd   v21.4s, v21.4s, v23.4s      \n"
-
-                        "sub    %7, %7, #768                \n" // kptr -= 24 * 16;
-
-                        "shrn   v20.4h, v20.4s, #16         \n"
-                        "shrn   v21.4h, v21.4s, #16         \n"
-
-                        "st1    {v20.4h, v21.4h}, [%0], #16 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(kptr)          // %7
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%2, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%2 :64]!  \n" // r00 r01 r02 r03
-
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.f32   {d24-d27}, [%1 :128]! \n" // sum0 sum1
-
-                        "vmul.f32   q14, q8, d0[0]      \n"
-                        "vmul.f32   q15, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "pld        [%2, #192]          \n"
-                        "vld1.u16   {d10-d12}, [%2 :64] \n" // r04 r05 r06
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vshll.u16  q6, d12, #16        \n"
-
-                        "vmla.f32   q14, q10, d2[0]     \n"
-                        "vmla.f32   q15, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d3[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vmla.f32   q15, q10, d10[0]    \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d11[0]     \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "pld        [%3, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%3 :64]!  \n" // r10 r11 r12 r13
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d0[0]     \n"
-                        "vmla.f32   q15, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d1[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "pld        [%3, #192]          \n"
-                        "vld1.u16   {d10-d12}, [%3 :64] \n" // r14 r15 r16
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vshll.u16  q6, d12, #16        \n"
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d4[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "pld        [%4, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%4 :64]!  \n" // r20 r21 r22 r23
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d12[0]    \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vmla.f32   q13, q11, d12[1]    \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vmla.f32   q13, q9, d13[1]     \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "pld        [%4, #192]          \n"
-                        "vld1.u16   {d10-d12}, [%4 :64] \n" // r24 r25 r26
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vshll.u16  q6, d12, #16        \n"
-
-                        "vmla.f32   q14, q10, d2[0]     \n"
-                        "vmla.f32   q15, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d3[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vmla.f32   q15, q10, d10[0]    \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d11[0]     \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "pld        [%5, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%5 :64]!  \n" // r30 r31 r32 r33
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d0[0]     \n"
-                        "vmla.f32   q15, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d0[1]     \n"
-                        "vmla.f32   q13, q11, d4[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d1[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "vmla.f32   q13, q9, d5[1]      \n"
-                        "pld        [%5, #192]          \n"
-                        "vld1.u16   {d10-d12}, [%5 :64] \n" // r34 r35 r36
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vshll.u16  q6, d12, #16        \n"
-
-                        "vmla.f32   q14, q8, d2[0]      \n"
-                        "vmla.f32   q15, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d2[1]      \n"
-                        "vmla.f32   q13, q9, d6[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d3[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vmla.f32   q13, q11, d7[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d4[0]     \n"
-                        "vmla.f32   q15, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d4[1]     \n"
-                        "vmla.f32   q13, q11, d8[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d5[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d6[0]      \n"
-                        "vmla.f32   q15, q8, d10[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d6[1]      \n"
-                        "vmla.f32   q13, q9, d10[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d7[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d11[0]    \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "vmla.f32   q13, q11, d11[1]    \n"
-                        "pld        [%6, #256]          \n"
-                        "vld1.u16   {d4-d7}, [%6 :64]!  \n" // r40 r41 r42 r43
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q0, d4, #16         \n"
-                        "vshll.u16  q1, d5, #16         \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q14, q10, d8[0]     \n"
-                        "vmla.f32   q15, q10, d12[0]    \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d8[1]     \n"
-                        "vmla.f32   q13, q11, d12[1]    \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vmla.f32   q13, q9, d13[1]     \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d0[0]      \n"
-                        "vmla.f32   q15, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d0[1]      \n"
-                        "vmla.f32   q13, q9, d4[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d1[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "vmla.f32   q13, q11, d5[1]     \n"
-                        "pld        [%6, #192]          \n"
-                        "vld1.u16   {d10-d12}, [%6 :64] \n" // r44 r45 r46
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q4, d10, #16        \n"
-                        "vshll.u16  q5, d11, #16        \n"
-                        "vshll.u16  q6, d12, #16        \n"
-
-                        "vmla.f32   q14, q10, d2[0]     \n"
-                        "vmla.f32   q15, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d2[1]     \n"
-                        "vmla.f32   q13, q11, d6[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q8, d3[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vmla.f32   q13, q9, d7[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d4[0]      \n"
-                        "vmla.f32   q15, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d4[1]      \n"
-                        "vmla.f32   q13, q9, d8[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-
-                        "vmla.f32   q14, q10, d5[0]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vmla.f32   q13, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q14, q10, d6[0]     \n"
-                        "vmla.f32   q15, q10, d10[0]    \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q12, q11, d6[1]     \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-                        //                         "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128] \n"
-
-                        "vmla.f32   q14, q8, d7[0]      \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d11[0]     \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vmla.f32   q13, q9, d11[1]     \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q14, q8, d8[0]      \n"
-                        "vmla.f32   q15, q8, d12[0]     \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q12, q9, d8[1]      \n"
-                        "vmla.f32   q13, q9, d12[1]     \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q14, q10, d9[0]     \n"
-                        "vmla.f32   q15, q10, d13[0]    \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vmla.f32   q13, q11, d13[1]    \n"
-
-                        "vadd.f32   q12, q12, q14       \n"
-                        "vadd.f32   q13, q13, q15       \n"
-
-                        "sub        %7, %7, #768        \n" // kptr -= 24 * 16;
-
-                        "vshrn.u32  d24, q12, #16       \n"
-                        "vshrn.u32  d25, q13, #16       \n"
-
-                        "vst1.u16   {d24-d25}, [%0 :64]! \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(kptr)          // %7
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v20.4s}, [%1], #16         \n" // sum0
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%2], #16   \n" // r00 r01
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmul   v21.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmul   v22.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmul   v23.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%2, #192]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h}, [%2] \n" // r02 r03 r04
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%3], #16   \n" // r10 r11
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v0.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v0.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%3, #192]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h}, [%3] \n" // r12 r13 r14
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v1.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%4], #16   \n" // r20 r21
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%4, #192]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h}, [%4] \n" // r22 r23 r24
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%5], #16   \n" // r30 r31
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v0.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v0.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%5, #192]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h}, [%5] \n" // r32 r33 r34
-                        "shll   v17.4s, v17.4h, #16         \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "fmla   v21.4s, v16.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v1.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v1.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v1.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v2.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v2.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v3.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v3.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v3.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v3.s[3]     \n"
-                        "prfm   pldl1keep, [%6, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%6], #16   \n" // r40 r41
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v21.4s, v24.4s, v4.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v4.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v4.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v4.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v0.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v0.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v0.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v0.s[3]     \n"
-                        "prfm   pldl1keep, [%6, #192]       \n"
-                        "ld1    {v2.4h, v3.4h, v4.4h}, [%6] \n" // r42 r43 r44
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-                        "shll   v4.4s, v4.4h, #16           \n"
-
-                        "fmla   v21.4s, v24.4s, v1.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v1.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v1.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v1.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v2.s[0]     \n"
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7], #32 \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v2.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v2.s[2]     \n"
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "fmla   v20.4s, v19.4s, v2.s[3]     \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-
-                        "fmla   v21.4s, v24.4s, v3.s[0]     \n"
-                        //                         "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7] \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-                        "fmla   v23.4s, v26.4s, v3.s[2]     \n"
-                        "shll   v16.4s, v16.4h, #16         \n"
-                        "fmla   v20.4s, v27.4s, v3.s[3]     \n"
-                        "shll   v17.4s, v17.4h, #16         \n"
-
-                        "fmla   v21.4s, v16.4s, v4.s[0]     \n"
-                        "shll   v18.4s, v18.4h, #16         \n"
-                        "fmla   v22.4s, v17.4s, v4.s[1]     \n"
-                        "shll   v19.4s, v19.4h, #16         \n"
-                        "fmla   v23.4s, v18.4s, v4.s[2]     \n"
-                        "fmla   v20.4s, v19.4s, v4.s[3]     \n"
-
-                        "fadd   v22.4s, v21.4s, v22.4s      \n"
-                        "fadd   v23.4s, v22.4s, v23.4s      \n"
-                        "fadd   v20.4s, v20.4s, v23.4s      \n"
-
-                        "sub    %7, %7, #768                \n" // kptr -= 24 * 16;
-
-                        "shrn   v20.4h, v20.4s, #16         \n"
-
-                        "st1    {v20.4h}, [%0], #8          \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(kptr)          // %7
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%2 :64]!  \n" // r00 r01
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d24-d25}, [%1 :128]! \n" // sum0
-
-                        "vmul.f32   q13, q8, d0[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmul.f32   q14, q9, d0[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmul.f32   q15, q10, d1[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "pld        [%2, #192]          \n"
-                        "vld1.u16   {d6-d8}, [%2 :64]   \n" // r02 r03 r04
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshll.u16  q4, d8, #16         \n"
-
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d2[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%3 :64]!  \n" // r10 r11
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d0[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d0[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d1[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "pld        [%3, #192]          \n"
-                        "vld1.u16   {d6-d8}, [%3 :64]   \n" // r12 r13 r14
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshll.u16  q4, d8, #16         \n"
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "pld        [%4, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%4 :64]!  \n" // r20 r21
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q13, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "pld        [%4, #192]          \n"
-                        "vld1.u16   {d6-d8}, [%4 :64]   \n" // r22 r23 r24
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshll.u16  q4, d8, #16         \n"
-
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d2[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "pld        [%5, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%5 :64]!  \n" // r30 r31
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d0[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d0[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d1[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d1[1]      \n"
-                        "pld        [%5, #192]          \n"
-                        "vld1.u16   {d6-d8}, [%5 :64]   \n" // r32 r33 r34
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshll.u16  q4, d8, #16         \n"
-
-                        "vmla.f32   q13, q8, d2[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d2[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d3[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d3[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d4[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d4[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d5[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d5[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d6[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d6[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d7[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d7[1]     \n"
-                        "pld        [%6, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%6 :64]!  \n" // r40 r41
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q13, q10, d8[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d8[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d0[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d0[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d1[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d1[1]     \n"
-                        "pld        [%6, #192]          \n"
-                        "vld1.u16   {d6-d8}, [%6 :64]   \n" // r42 r43 r44
-                        "vshll.u16  q11, d17, #16       \n"
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshll.u16  q4, d8, #16         \n"
-
-                        "vmla.f32   q13, q10, d2[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d2[1]     \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128]! \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d3[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d3[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d4[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d4[1]      \n"
-                        "pld        [%7, #256]          \n"
-                        "vld1.u16   {d16-d19}, [%7 :128]! \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d5[0]     \n"
-                        "vshll.u16  q10, d16, #16       \n"
-                        "vmla.f32   q12, q11, d5[1]     \n"
-                        "vshll.u16  q11, d17, #16       \n"
-
-                        "vmla.f32   q13, q10, d6[0]     \n"
-                        "vshll.u16  q8, d18, #16        \n"
-                        "vmla.f32   q14, q11, d6[1]     \n"
-                        //                         "pld        [%7, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%7 :128] \n"
-                        "vshll.u16  q9, d19, #16        \n"
-                        "vmla.f32   q15, q8, d7[0]      \n"
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vmla.f32   q12, q9, d7[1]      \n"
-                        "vshll.u16  q9, d21, #16        \n"
-
-                        "vmla.f32   q13, q8, d8[0]      \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vmla.f32   q14, q9, d8[1]      \n"
-                        "vshll.u16  q11, d23, #16       \n"
-                        "vmla.f32   q15, q10, d9[0]     \n"
-                        "vmla.f32   q12, q11, d9[1]     \n"
-
-                        "vadd.f32   q14, q13, q14       \n"
-                        "vadd.f32   q15, q14, q15       \n"
-                        "vadd.f32   q12, q12, q15       \n"
-
-                        "sub        %7, %7, #768        \n" // kptr -= 24 * 16;
-
-                        "vshrn.u32  d24, q12, #16       \n"
-
-                        "vst1.u16   {d24}, [%0 :64]!    \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(kptr)          // %7
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-                r4 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5_pack8_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5_pack8_fp16s.h
deleted file mode 100644
index 69e6564b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_5x5_pack8_fp16s.h
+++ /dev/null
@@ -1,3449 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv5x5s1_pack8_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float16x8_t _bias0 = bias ? vld1q_f16(bias + p * 8) : vdupq_n_f16(0.f);
-        out0.fill(_bias0);
-
-        int q = 0;
-        for (; q < inch; q++)
-        {
-            __fp16* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const __fp16* r0 = img0.row<const __fp16>(0);
-            const __fp16* r1 = img0.row<const __fp16>(1);
-            const __fp16* r2 = img0.row<const __fp16>(2);
-            const __fp16* r3 = img0.row<const __fp16>(3);
-            const __fp16* r4 = img0.row<const __fp16>(4);
-
-            const __fp16* kptr = kernel.channel(p).row<const __fp16>(q);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0] \n" // sum0 sum1 sum2 sum3
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%1], #64 \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[5]     \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%1] \n" // r04 r05 r06 r07
-
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v5.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v5.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v6.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v6.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v6.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v5.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v6.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v6.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v6.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v6.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v5.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v6.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v6.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v7.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v6.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v7.h[1]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v8.8h, v9.8h, v10.8h, v11.8h}, [%2], #64 \n" // r10 r11 r12 r13
-
-                        "fmla   v28.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v6.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v7.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v5.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v6.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v7.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v6.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v7.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v6.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v7.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v6.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v7.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v5.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v6.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v7.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v8.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v28.8h, v17.8h, v8.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v9.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v11.h[1]    \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%2] \n" // r14 r15 r16 r17
-
-                        "fmla   v28.8h, v18.8h, v8.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v28.8h, v19.8h, v8.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v9.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v10.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v11.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v8.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v28.8h, v21.8h, v8.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v8.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v28.8h, v23.8h, v8.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v9.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v10.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v11.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v30.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v28.8h, v17.8h, v9.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v30.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v28.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v30.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v28.8h, v19.8h, v9.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v10.h[3]    \n"
-                        "fmla   v30.8h, v19.8h, v11.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v12.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v30.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v28.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v30.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v30.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v28.8h, v23.8h, v9.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v10.h[7]    \n"
-                        "fmla   v30.8h, v23.8h, v11.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v12.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v29.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v30.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v13.h[0]    \n"
-                        "fmla   v28.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v29.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v30.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v13.h[1]    \n"
-                        "fmla   v28.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v29.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v30.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v13.h[2]    \n"
-                        "fmla   v28.8h, v19.8h, v10.h[3]    \n"
-                        "fmla   v29.8h, v19.8h, v11.h[3]    \n"
-                        "fmla   v30.8h, v19.8h, v12.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v13.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v29.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v30.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v13.h[4]    \n"
-                        "fmla   v28.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v29.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v30.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v13.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v29.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v30.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v13.h[6]    \n"
-                        "fmla   v28.8h, v23.8h, v10.h[7]    \n"
-                        "fmla   v29.8h, v23.8h, v11.h[7]    \n"
-                        "fmla   v30.8h, v23.8h, v12.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v13.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v29.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v30.8h, v16.8h, v13.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v14.h[0]    \n"
-                        "fmla   v28.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v29.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v30.8h, v17.8h, v13.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v14.h[1]    \n"
-                        "fmla   v28.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v29.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v30.8h, v18.8h, v13.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v14.h[2]    \n"
-                        "fmla   v28.8h, v19.8h, v11.h[3]    \n"
-                        "fmla   v29.8h, v19.8h, v12.h[3]    \n"
-                        "fmla   v30.8h, v19.8h, v13.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v14.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v29.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v30.8h, v20.8h, v13.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v14.h[4]    \n"
-                        "fmla   v28.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v29.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v30.8h, v21.8h, v13.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v14.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v29.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v30.8h, v22.8h, v13.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v14.h[6]    \n"
-                        "fmla   v28.8h, v23.8h, v11.h[7]    \n"
-                        "fmla   v29.8h, v23.8h, v12.h[7]    \n"
-                        "fmla   v30.8h, v23.8h, v13.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v14.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v29.8h, v16.8h, v13.h[0]    \n"
-                        "fmla   v30.8h, v16.8h, v14.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v15.h[0]    \n"
-                        "fmla   v28.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v29.8h, v17.8h, v13.h[1]    \n"
-                        "fmla   v30.8h, v17.8h, v14.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v15.h[1]    \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%3], #64 \n" // r20 r21 r22 r23
-
-                        "fmla   v28.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v29.8h, v18.8h, v13.h[2]    \n"
-                        "fmla   v30.8h, v18.8h, v14.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v15.h[2]    \n"
-                        "fmla   v28.8h, v19.8h, v12.h[3]    \n"
-                        "fmla   v29.8h, v19.8h, v13.h[3]    \n"
-                        "fmla   v30.8h, v19.8h, v14.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v15.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v29.8h, v20.8h, v13.h[4]    \n"
-                        "fmla   v30.8h, v20.8h, v14.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v15.h[4]    \n"
-                        "fmla   v28.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v29.8h, v21.8h, v13.h[5]    \n"
-                        "fmla   v30.8h, v21.8h, v14.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v15.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v29.8h, v22.8h, v13.h[6]    \n"
-                        "fmla   v30.8h, v22.8h, v14.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v15.h[6]    \n"
-                        "fmla   v28.8h, v23.8h, v12.h[7]    \n"
-                        "fmla   v29.8h, v23.8h, v13.h[7]    \n"
-                        "fmla   v30.8h, v23.8h, v14.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v15.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%3] \n" // r24 r25 r26 r27
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v5.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v5.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v6.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v6.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v6.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v5.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v6.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v6.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v6.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v6.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v5.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v6.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v6.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v7.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v6.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v7.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v8.8h, v9.8h, v10.8h, v11.8h}, [%4], #64 \n" // r30 r31 r32 r33
-
-                        "fmla   v28.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v6.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v7.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v5.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v6.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v7.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v6.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v7.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v6.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v7.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v6.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v7.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v5.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v6.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v7.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v8.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v28.8h, v17.8h, v8.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v9.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v11.h[1]    \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%4] \n" // r34 r35 r36 r37
-
-                        "fmla   v28.8h, v18.8h, v8.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v28.8h, v19.8h, v8.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v9.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v10.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v11.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v8.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v28.8h, v21.8h, v8.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v8.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v28.8h, v23.8h, v8.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v9.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v10.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v11.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v30.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v28.8h, v17.8h, v9.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v30.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v28.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v30.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v28.8h, v19.8h, v9.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v10.h[3]    \n"
-                        "fmla   v30.8h, v19.8h, v11.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v12.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v30.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v28.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v30.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v30.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v28.8h, v23.8h, v9.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v10.h[7]    \n"
-                        "fmla   v30.8h, v23.8h, v11.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v12.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v29.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v30.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v13.h[0]    \n"
-                        "fmla   v28.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v29.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v30.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v13.h[1]    \n"
-                        "fmla   v28.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v29.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v30.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v13.h[2]    \n"
-                        "fmla   v28.8h, v19.8h, v10.h[3]    \n"
-                        "fmla   v29.8h, v19.8h, v11.h[3]    \n"
-                        "fmla   v30.8h, v19.8h, v12.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v13.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v29.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v30.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v13.h[4]    \n"
-                        "fmla   v28.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v29.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v30.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v13.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v29.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v30.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v13.h[6]    \n"
-                        "fmla   v28.8h, v23.8h, v10.h[7]    \n"
-                        "fmla   v29.8h, v23.8h, v11.h[7]    \n"
-                        "fmla   v30.8h, v23.8h, v12.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v13.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v29.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v30.8h, v16.8h, v13.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v14.h[0]    \n"
-                        "fmla   v28.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v29.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v30.8h, v17.8h, v13.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v14.h[1]    \n"
-                        "fmla   v28.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v29.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v30.8h, v18.8h, v13.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v14.h[2]    \n"
-                        "fmla   v28.8h, v19.8h, v11.h[3]    \n"
-                        "fmla   v29.8h, v19.8h, v12.h[3]    \n"
-                        "fmla   v30.8h, v19.8h, v13.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v14.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v29.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v30.8h, v20.8h, v13.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v14.h[4]    \n"
-                        "fmla   v28.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v29.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v30.8h, v21.8h, v13.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v14.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v29.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v30.8h, v22.8h, v13.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v14.h[6]    \n"
-                        "fmla   v28.8h, v23.8h, v11.h[7]    \n"
-                        "fmla   v29.8h, v23.8h, v12.h[7]    \n"
-                        "fmla   v30.8h, v23.8h, v13.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v14.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v29.8h, v16.8h, v13.h[0]    \n"
-                        "fmla   v30.8h, v16.8h, v14.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v15.h[0]    \n"
-                        "fmla   v28.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v29.8h, v17.8h, v13.h[1]    \n"
-                        "fmla   v30.8h, v17.8h, v14.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v15.h[1]    \n"
-
-                        "prfm   pldl1keep, [%5, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%5], #64 \n" // r40 r41 r42 r43
-
-                        "fmla   v28.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v29.8h, v18.8h, v13.h[2]    \n"
-                        "fmla   v30.8h, v18.8h, v14.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v15.h[2]    \n"
-                        "fmla   v28.8h, v19.8h, v12.h[3]    \n"
-                        "fmla   v29.8h, v19.8h, v13.h[3]    \n"
-                        "fmla   v30.8h, v19.8h, v14.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v15.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v29.8h, v20.8h, v13.h[4]    \n"
-                        "fmla   v30.8h, v20.8h, v14.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v15.h[4]    \n"
-                        "fmla   v28.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v29.8h, v21.8h, v13.h[5]    \n"
-                        "fmla   v30.8h, v21.8h, v14.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v15.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v29.8h, v22.8h, v13.h[6]    \n"
-                        "fmla   v30.8h, v22.8h, v14.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v15.h[6]    \n"
-                        "fmla   v28.8h, v23.8h, v12.h[7]    \n"
-                        "fmla   v29.8h, v23.8h, v13.h[7]    \n"
-                        "fmla   v30.8h, v23.8h, v14.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v15.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%5, #512]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%5] \n" // r44 r45 r46 r47
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v5.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v5.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v6.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v6.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v6.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v5.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v6.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v6.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v6.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v6.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v5.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v6.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6] \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v6.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v7.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v30.8h, v17.8h, v6.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v7.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v6.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v7.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v5.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v6.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v7.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v6.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v7.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v6.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v7.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v6.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v7.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v29.8h, v23.8h, v5.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v6.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v7.h[7]     \n"
-
-                        "sub    %6, %6, #3136               \n" // kptr -= 24.5 * 64;
-
-                        "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v0.8h, v1.8h}, [%1], #32   \n" // r00 r01
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v30.8h, v31.8h}, [%0]      \n" // sum0 sum1
-
-                        "fmul   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmul   v29.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v1.h[5]     \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v2.8h, v3.8h, v4.8h, v5.8h}, [%1] \n" // r02 r03 r04 r05
-
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v5.h[1]     \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v8.8h, v9.8h}, [%2], #32   \n" // r10 r11
-
-                        "fmla   v28.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v5.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v5.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v16.8h, v8.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v8.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v9.h[1]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v10.8h, v11.8h, v12.8h, v13.8h}, [%2] \n" // r12 r13 r14 r15
-
-                        "fmla   v30.8h, v18.8h, v8.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v8.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v9.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v20.8h, v8.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v8.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v8.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v8.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v9.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v9.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v30.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v9.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v10.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v9.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v10.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v30.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v10.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v11.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v10.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v11.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v30.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v11.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v12.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v11.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v12.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v13.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v13.h[1]    \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.8h, v1.8h}, [%3], #32   \n" // r20 r21
-
-                        "fmla   v30.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v13.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v12.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v13.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v13.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v13.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v13.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v12.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v13.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v2.8h, v3.8h, v4.8h, v5.8h}, [%3] \n" // r22 r23 r24 r25
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v5.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v8.8h, v9.8h}, [%4], #32   \n" // r30 r31
-
-                        "fmla   v28.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v5.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v5.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v16.8h, v8.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v8.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v9.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v10.8h, v11.8h, v12.8h, v13.8h}, [%4] \n" // r32 r33 r34 r35
-
-                        "fmla   v30.8h, v18.8h, v8.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v8.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v9.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v20.8h, v8.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v8.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v8.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v8.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v9.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v9.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v30.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v9.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v10.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v9.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v10.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v30.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v10.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v11.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v10.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v11.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v30.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v11.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v12.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v11.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v12.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v31.8h, v16.8h, v13.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v13.h[1]    \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v0.8h, v1.8h}, [%5], #32   \n" // r40 r41
-
-                        "fmla   v30.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v31.8h, v18.8h, v13.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v12.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v13.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v30.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v31.8h, v20.8h, v13.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v13.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v31.8h, v22.8h, v13.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v12.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v13.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%5, #512]       \n"
-                        "ld1    {v2.8h, v3.8h, v4.8h, v5.8h}, [%5] \n" // r42 r43 r44 r45
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6] \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v5.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v5.h[7]     \n"
-
-                        "fadd   v28.8h, v28.8h, v30.8h      \n"
-                        "fadd   v29.8h, v29.8h, v31.8h      \n"
-
-                        "sub    %6, %6, #3136               \n" // kptr -= 24.5 * 64;
-
-                        "st1    {v28.8h, v29.8h}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v8", "v9", "v10", "v11", "v12", "v13", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-                }
-                for (; j < outw; j++)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.8h}, [%1], #16          \n" // r00
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v31.8h}, [%0]              \n" // sum0
-
-                        "fmul   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmul   v29.8h, v17.8h, v0.h[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmul   v30.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[5]     \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v1.8h, v2.8h, v3.8h, v4.8h}, [%1] \n" // r01 r02 r03 r04
-
-                        "fmla   v30.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v4.h[1]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v8.8h}, [%2], #16          \n" // r10
-
-                        "fmla   v30.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v8.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v8.h[1]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v9.8h, v10.8h, v11.8h, v12.8h}, [%2] \n" // r11 r12 r13 r14
-
-                        "fmla   v30.8h, v18.8h, v8.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v8.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v8.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v8.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v8.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v8.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v9.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v9.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v9.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v29.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v30.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v31.8h, v19.8h, v10.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v29.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v31.8h, v23.8h, v10.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v29.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v30.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v31.8h, v19.8h, v11.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v29.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v31.8h, v23.8h, v11.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v29.8h, v17.8h, v12.h[1]    \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.8h}, [%3], #16          \n" // r20
-
-                        "fmla   v30.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v31.8h, v19.8h, v12.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v29.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v31.8h, v23.8h, v12.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v0.h[1]     \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v1.8h, v2.8h, v3.8h, v4.8h}, [%3] \n" // r21 r22 r23 r24
-
-                        "fmla   v30.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v4.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v8.8h}, [%4], #16          \n" // r30
-
-                        "fmla   v30.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v8.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v8.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v9.8h, v10.8h, v11.8h, v12.8h}, [%4] \n" // r31 r32 r33 r34
-
-                        "fmla   v30.8h, v18.8h, v8.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v8.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v8.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v8.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v8.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v8.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v9.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v9.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v9.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v29.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v30.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v31.8h, v19.8h, v10.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v29.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v31.8h, v23.8h, v10.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v29.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v30.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v31.8h, v19.8h, v11.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v29.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v31.8h, v23.8h, v11.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v29.8h, v17.8h, v12.h[1]    \n"
-
-                        "prfm   pldl1keep, [%5, #128]       \n"
-                        "ld1    {v0.8h}, [%5], #16          \n" // r40
-
-                        "fmla   v30.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v31.8h, v19.8h, v12.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v29.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v31.8h, v23.8h, v12.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v0.h[1]     \n"
-
-                        "prfm   pldl1keep, [%5, #512]       \n"
-                        "ld1    {v1.8h, v2.8h, v3.8h, v4.8h}, [%5] \n" // r41 r42 r43 r44
-
-                        "fmla   v30.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6] \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "fadd   v28.8h, v28.8h, v29.8h      \n"
-                        "fadd   v30.8h, v30.8h, v31.8h      \n"
-                        "fadd   v28.8h, v28.8h, v30.8h      \n"
-
-                        "sub    %6, %6, #3136               \n" // kptr -= 24.5 * 64;
-
-                        "st1    {v28.8h}, [%0], #16         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v8", "v9", "v10", "v11", "v12", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-                }
-
-                r0 += 4 * 8;
-                r1 += 4 * 8;
-                r2 += 4 * 8;
-                r3 += 4 * 8;
-                r4 += 4 * 8;
-            }
-        }
-    }
-}
-
-static void conv5x5s2_pack8_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 8;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float16x8_t _bias0 = bias ? vld1q_f16(bias + p * 8) : vdupq_n_f16(0.f);
-        out0.fill(_bias0);
-
-        int q = 0;
-        for (; q < inch; q++)
-        {
-            __fp16* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const __fp16* r0 = img0.row<const __fp16>(0);
-            const __fp16* r1 = img0.row<const __fp16>(1);
-            const __fp16* r2 = img0.row<const __fp16>(2);
-            const __fp16* r3 = img0.row<const __fp16>(3);
-            const __fp16* r4 = img0.row<const __fp16>(4);
-
-            const __fp16* kptr = kernel.channel(p).row<const __fp16>(q);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 1 < outw; j += 2)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%1], #64 \n" // r00 r01 r02 r03
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v30.8h, v31.8h}, [%0]      \n" // sum0 sum1
-
-                        "fmul   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmul   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v2.h[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%1, #384]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h}, [%1] \n" // r04 r05 r06
-
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v5.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v5.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v6.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v6.h[1]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v8.8h, v9.8h, v10.8h, v11.8h}, [%2], #64 \n" // r10 r11 r12 r13
-
-                        "fmla   v28.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v6.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v6.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v6.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v6.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v6.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v6.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v8.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v8.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v28.8h, v18.8h, v8.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v8.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v10.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v8.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v8.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v8.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v8.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v10.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v9.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v11.h[1]    \n"
-
-                        "prfm   pldl1keep, [%2, #384]       \n"
-                        "ld1    {v12.8h, v13.8h, v14.8h}, [%2] \n" // r14 r15 r16
-
-                        "fmla   v28.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v9.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v11.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v9.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v11.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v29.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v28.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v29.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v10.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v12.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v29.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v29.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v10.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v12.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v29.8h, v16.8h, v13.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v13.h[1]    \n"
-                        "fmla   v28.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v29.8h, v18.8h, v13.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v11.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v13.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v29.8h, v20.8h, v13.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v13.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v29.8h, v22.8h, v13.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v11.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v13.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v29.8h, v16.8h, v14.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v14.h[1]    \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%3], #64 \n" // r20 r21 r22 r23
-
-                        "fmla   v28.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v29.8h, v18.8h, v14.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v12.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v14.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v29.8h, v20.8h, v14.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v14.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v29.8h, v22.8h, v14.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v12.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v14.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%3, #384]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h}, [%3] \n" // r24 r25 r26
-
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v5.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v5.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v6.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v6.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v8.8h, v9.8h, v10.8h, v11.8h}, [%4], #64 \n" // r30 r31 r32 r33
-
-                        "fmla   v28.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v6.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v6.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v6.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v6.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v6.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v6.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v8.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v8.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v28.8h, v18.8h, v8.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v8.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v10.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v8.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v8.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v8.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v8.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v10.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v9.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v11.h[1]    \n"
-
-                        "prfm   pldl1keep, [%4, #384]       \n"
-                        "ld1    {v12.8h, v13.8h, v14.8h}, [%4] \n" // r34 r35 r36
-
-                        "fmla   v28.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v9.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v11.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v9.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v11.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v29.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v28.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v29.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v10.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v12.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v29.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v29.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v10.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v12.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v29.8h, v16.8h, v13.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v13.h[1]    \n"
-                        "fmla   v28.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v29.8h, v18.8h, v13.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v11.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v13.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v29.8h, v20.8h, v13.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v13.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v29.8h, v22.8h, v13.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v11.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v13.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v29.8h, v16.8h, v14.h[0]    \n"
-                        "fmla   v30.8h, v17.8h, v12.h[1]    \n"
-                        "fmla   v31.8h, v17.8h, v14.h[1]    \n"
-
-                        "prfm   pldl1keep, [%5, #512]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%5], #64 \n" // r40 r41 r42 r43
-
-                        "fmla   v28.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v29.8h, v18.8h, v14.h[2]    \n"
-                        "fmla   v30.8h, v19.8h, v12.h[3]    \n"
-                        "fmla   v31.8h, v19.8h, v14.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v29.8h, v20.8h, v14.h[4]    \n"
-                        "fmla   v30.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v31.8h, v21.8h, v14.h[5]    \n"
-                        "fmla   v28.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v29.8h, v22.8h, v14.h[6]    \n"
-                        "fmla   v30.8h, v23.8h, v12.h[7]    \n"
-                        "fmla   v31.8h, v23.8h, v14.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%5, #384]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h}, [%5] \n" // r44 r45 r46
-
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v5.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v5.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v5.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v3.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v5.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v5.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v5.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v5.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v3.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v5.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6] \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v6.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v6.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v6.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v6.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v6.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v6.h[5]     \n"
-                        "fmla   v28.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v6.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v4.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v6.h[7]     \n"
-
-                        "fadd   v28.8h, v28.8h, v30.8h      \n"
-                        "fadd   v29.8h, v29.8h, v31.8h      \n"
-
-                        "sub    %6, %6, #3136               \n" // kptr -= 24.5 * 64;
-
-                        "st1    {v28.8h, v29.8h}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-                }
-                for (; j < outw; j++)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v0.8h, v1.8h}, [%1], #32   \n" // r00 r01
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v31.8h}, [%0]              \n" // sum0
-
-                        "fmul   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmul   v29.8h, v17.8h, v0.h[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmul   v30.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%1, #384]       \n"
-                        "ld1    {v2.8h, v3.8h, v4.8h}, [%1] \n" // r02 r03 r04
-
-                        "fmla   v30.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v4.h[1]     \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v8.8h, v9.8h}, [%2], #32   \n" // r10 r11
-
-                        "fmla   v30.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v8.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v8.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v8.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v8.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v8.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v8.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v8.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v8.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v9.h[1]     \n"
-
-                        "prfm   pldl1keep, [%2, #384]       \n"
-                        "ld1    {v10.8h, v11.8h, v12.8h}, [%2] \n" // r12 r13 r14
-
-                        "fmla   v30.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v9.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v9.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v29.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v30.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v31.8h, v19.8h, v10.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v29.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v31.8h, v23.8h, v10.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v29.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v30.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v31.8h, v19.8h, v11.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v29.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v31.8h, v23.8h, v11.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v29.8h, v17.8h, v12.h[1]    \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.8h, v1.8h}, [%3], #32   \n" // r20 r21
-
-                        "fmla   v30.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v31.8h, v19.8h, v12.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v29.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v31.8h, v23.8h, v12.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%3, #384]       \n"
-                        "ld1    {v2.8h, v3.8h, v4.8h}, [%3] \n" // r22 r23 r24
-
-                        "fmla   v30.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v4.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v8.8h, v9.8h}, [%4], #32   \n" // r30 r31
-
-                        "fmla   v30.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v8.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v8.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v8.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v8.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v8.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v8.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v8.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v8.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v9.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v9.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #384]       \n"
-                        "ld1    {v10.8h, v11.8h, v12.8h}, [%4] \n" // r32 r33 r34
-
-                        "fmla   v30.8h, v18.8h, v9.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v9.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v9.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v9.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v9.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v9.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v10.h[0]    \n"
-                        "fmla   v29.8h, v17.8h, v10.h[1]    \n"
-                        "fmla   v30.8h, v18.8h, v10.h[2]    \n"
-                        "fmla   v31.8h, v19.8h, v10.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v10.h[4]    \n"
-                        "fmla   v29.8h, v21.8h, v10.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v10.h[6]    \n"
-                        "fmla   v31.8h, v23.8h, v10.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v11.h[0]    \n"
-                        "fmla   v29.8h, v17.8h, v11.h[1]    \n"
-                        "fmla   v30.8h, v18.8h, v11.h[2]    \n"
-                        "fmla   v31.8h, v19.8h, v11.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v11.h[4]    \n"
-                        "fmla   v29.8h, v21.8h, v11.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v11.h[6]    \n"
-                        "fmla   v31.8h, v23.8h, v11.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v12.h[0]    \n"
-                        "fmla   v29.8h, v17.8h, v12.h[1]    \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v0.8h, v1.8h}, [%5], #32   \n" // r40 r41
-
-                        "fmla   v30.8h, v18.8h, v12.h[2]    \n"
-                        "fmla   v31.8h, v19.8h, v12.h[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v12.h[4]    \n"
-                        "fmla   v29.8h, v21.8h, v12.h[5]    \n"
-                        "fmla   v30.8h, v22.8h, v12.h[6]    \n"
-                        "fmla   v31.8h, v23.8h, v12.h[7]    \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%5, #384]       \n"
-                        "ld1    {v2.8h, v3.8h, v4.8h}, [%5] \n" // r42 r43 r44
-
-                        "fmla   v30.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v3.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n"
-
-                        "fmla   v28.8h, v20.8h, v3.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v3.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[7]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6] \n"
-
-                        "fmla   v28.8h, v16.8h, v4.h[0]     \n"
-                        "fmla   v29.8h, v17.8h, v4.h[1]     \n"
-                        "fmla   v30.8h, v18.8h, v4.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v4.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v4.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v4.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v4.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v4.h[7]     \n"
-
-                        "fadd   v28.8h, v28.8h, v29.8h      \n"
-                        "fadd   v30.8h, v30.8h, v31.8h      \n"
-                        "fadd   v28.8h, v28.8h, v30.8h      \n"
-
-                        "sub    %6, %6, #3136               \n" // kptr -= 24.5 * 64;
-
-                        "st1    {v28.8h}, [%0], #16         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(kptr)     // %6
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v8", "v9", "v10", "v11", "v12", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-                r4 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7.h
deleted file mode 100644
index d5448d92..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7.h
+++ /dev/null
@@ -1,1401 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv7x7s1_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch * 49 + q * 49;
-
-            const float* r0 = img0;
-            const float* r1 = img0 + w;
-            const float* r2 = img0 + w * 2;
-            const float* r3 = img0 + w * 3;
-            const float* r4 = img0 + w * 4;
-            const float* r5 = img0 + w * 5;
-            const float* r6 = img0 + w * 6;
-
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 7;
-            const float* k2 = kernel0 + 14;
-            const float* k3 = kernel0 + 21;
-            const float* k4 = kernel0 + 28;
-            const float* k5 = kernel0 + 35;
-            const float* k6 = kernel0 + 42;
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw - (nn << 2);
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                float32x4_t _k0123 = vld1q_f32(k0);
-                float32x4_t _k4567 = vld1q_f32(k0 + 4);
-                float32x4_t _k78910 = vld1q_f32(k1);
-                float32x4_t _k11121314 = vld1q_f32(k1 + 4);
-                float32x4_t _k14151617 = vld1q_f32(k2);
-                float32x4_t _k18192021 = vld1q_f32(k2 + 4);
-                float32x4_t _k21222324 = vld1q_f32(k3);
-                float32x4_t _k25262728 = vld1q_f32(k3 + 4);
-                float32x4_t _k28293031 = vld1q_f32(k4);
-                float32x4_t _k32333435 = vld1q_f32(k4 + 4);
-                float32x4_t _k35363738 = vld1q_f32(k5);
-                float32x4_t _k39404142 = vld1q_f32(k5 + 4);
-                float32x4_t _k42434445 = vld1q_f32(k6);
-                float32x4_t _k46474849 = vld1q_f32(k6 + 4);
-#ifdef __clang__ // __ARM_NEON && __aarch64__ && __clang__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        // v0:  input / final output
-                        // v1 v2 v3: = ri0 ri4 ri0n , i <-  1-7
-                        // v4 = ri1 / ri3 / ri6
-                        // v5 = ri2 / ri5
-                        // v9 = intermediate sum register
-                        "0:                                        \n"
-                        "prfm       pldl1keep, [%1, #128]          \n"
-                        "ld1        {v0.4s}, [%1]                  \n"
-
-                        //i = 1
-                        "prfm       pldl1keep, [%2, #384]          \n"
-                        "ld1        {v1.4s, v2.4s, v3.4s}, [%2]    \n"
-                        "add        %2, %2, #16                    \n"
-                        "ext        v4.16b, v1.16b, v2.16b, #4     \n"
-                        "fmul       v9.4s, v1.4s, %18.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v2.16b, #8     \n"
-                        "fmla       v0.4s, v4.4s, %18.s[1]         \n"
-                        "ext        v4.16b, v1.16b, v2.16b, #12    \n"
-                        "fmla       v9.4s, v5.4s, %18.s[2]         \n"
-                        "ext        v5.16b, v2.16b, v3.16b, #4     \n"
-                        "fmla       v0.4s, v4.4s, %18.s[3]         \n"
-                        "ext        v4.16b, v2.16b, v3.16b, #8     \n"
-                        "fmla       v9.4s, v2.4s, %19.s[0]         \n"
-                        "fmla       v0.4s, v5.4s, %19.s[1]         \n"
-                        "fmla       v9.4s, v4.4s, %19.s[2]         \n"
-
-                        //i = 2
-                        "prfm       pldl1keep, [%3, #384]          \n"
-                        "ld1        {v1.4s, v2.4s, v3.4s}, [%3]    \n" // v1 v2 v3: = r20 r24 r20n
-                        "add        %3, %3, #16                    \n"
-                        "ext        v4.16b, v1.16b, v2.16b, #4     \n" // v4 = r21
-                        "fmla       v9.4s, v1.4s, %20.s[0]         \n" // *+ r10
-                        "ext        v5.16b, v1.16b, v2.16b, #8     \n" // v5 = r22
-                        "fmla       v0.4s, v4.4s, %20.s[1]         \n" // *+ r11
-                        "ext        v4.16b, v1.16b, v2.16b, #12    \n" // v4 = r23
-                        "fmla       v9.4s, v5.4s, %20.s[2]         \n" // *+ r1
-                        "ext        v5.16b, v2.16b, v3.16b, #4     \n" // v5 = r25
-                        "fmla       v0.4s, v4.4s, %20.s[3]         \n" // *+ r13
-                        "ext        v4.16b, v2.16b, v3.16b, #8     \n" // v4 = r26
-                        "fmla       v9.4s, v2.4s, %21.s[0]         \n" // *+ r14
-                        "fmla       v0.4s, v5.4s, %21.s[1]         \n" // *+ r15
-                        "fmla       v9.4s, v4.4s, %21.s[2]         \n" // *+ r16
-
-                        //i = 3
-                        "prfm       pldl1keep, [%4, #384]          \n"
-                        "ld1        {v1.4s, v2.4s, v3.4s}, [%4]    \n"
-                        "add        %4, %4, #16                    \n"
-                        "ext        v4.16b, v1.16b, v2.16b, #4     \n"
-                        "fmla       v9.4s, v1.4s, %22.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v2.16b, #8     \n"
-                        "fmla       v0.4s, v4.4s, %22.s[1]         \n"
-                        "ext        v4.16b, v1.16b, v2.16b, #12    \n"
-                        "fmla       v9.4s, v5.4s, %22.s[2]         \n"
-                        "ext        v5.16b, v2.16b, v3.16b, #4     \n"
-                        "fmla       v0.4s, v4.4s, %22.s[3]         \n"
-                        "ext        v4.16b, v2.16b, v3.16b, #8     \n"
-                        "fmla       v9.4s, v2.4s, %23.s[0]         \n"
-                        "fmla       v0.4s, v5.4s, %23.s[1]         \n"
-                        "fmla       v9.4s, v4.4s, %23.s[2]         \n"
-
-                        //i = 4
-                        "prfm       pldl1keep, [%5, #384]          \n"
-                        "ld1        {v1.4s, v2.4s, v3.4s}, [%5]    \n"
-                        "add        %5, %5, #16                    \n"
-                        "ext        v4.16b, v1.16b, v2.16b, #4     \n"
-                        "fmla       v9.4s, v1.4s, %24.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v2.16b, #8     \n"
-                        "fmla       v0.4s, v4.4s, %24.s[1]         \n"
-                        "ext        v4.16b, v1.16b, v2.16b, #12    \n"
-                        "fmla       v9.4s, v5.4s, %24.s[2]         \n"
-                        "ext        v5.16b, v2.16b, v3.16b, #4     \n"
-                        "fmla       v0.4s, v4.4s, %24.s[3]         \n"
-                        "ext        v4.16b, v2.16b, v3.16b, #8     \n"
-                        "fmla       v9.4s, v2.4s, %25.s[0]         \n"
-                        "fmla       v0.4s, v5.4s, %25.s[1]         \n"
-                        "fmla       v9.4s, v4.4s, %25.s[2]         \n"
-
-                        //i = 5
-                        "prfm       pldl1keep, [%6, #384]          \n"
-                        "ld1        {v1.4s, v2.4s, v3.4s}, [%6]    \n"
-                        "add        %6, %6, #16                    \n"
-                        "ext        v4.16b, v1.16b, v2.16b, #4     \n"
-                        "fmla       v9.4s, v1.4s, %26.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v2.16b, #8     \n"
-                        "fmla       v0.4s, v4.4s, %26.s[1]         \n"
-                        "ext        v4.16b, v1.16b, v2.16b, #12    \n"
-                        "fmla       v9.4s, v5.4s, %26.s[2]         \n"
-                        "ext        v5.16b, v2.16b, v3.16b, #4     \n"
-                        "fmla       v0.4s, v4.4s, %26.s[3]         \n"
-                        "ext        v4.16b, v2.16b, v3.16b, #8     \n"
-                        "fmla       v9.4s, v2.4s, %27.s[0]         \n"
-                        "fmla       v0.4s, v5.4s, %27.s[1]         \n"
-                        "fmla       v9.4s, v4.4s, %27.s[2]         \n"
-
-                        //i = 6
-                        "prfm       pldl1keep, [%7, #384]          \n"
-                        "ld1        {v1.4s, v2.4s, v3.4s}, [%7]    \n"
-                        "add        %7, %7, #16                    \n"
-                        "ext        v4.16b, v1.16b, v2.16b, #4     \n"
-                        "fmla       v9.4s, v1.4s, %28.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v2.16b, #8     \n"
-                        "fmla       v0.4s, v4.4s, %28.s[1]         \n"
-                        "ext        v4.16b, v1.16b, v2.16b, #12    \n"
-                        "fmla       v9.4s, v5.4s, %28.s[2]         \n"
-                        "ext        v5.16b, v2.16b, v3.16b, #4     \n"
-                        "fmla       v0.4s, v4.4s, %28.s[3]         \n"
-                        "ext        v4.16b, v2.16b, v3.16b, #8     \n"
-                        "fmla       v9.4s, v2.4s, %29.s[0]         \n"
-                        "fmla       v0.4s, v5.4s, %29.s[1]         \n"
-                        "fmla       v9.4s, v4.4s, %29.s[2]         \n"
-
-                        //i = 7
-                        "prfm       pldl1keep, [%8, #384]          \n"
-                        "ld1        {v1.4s, v2.4s, v3.4s}, [%8]    \n"
-                        "add        %8, %8, #16                    \n"
-                        "ext        v4.16b, v1.16b, v2.16b, #4     \n"
-                        "fmla       v9.4s, v1.4s, %30.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v2.16b, #8     \n"
-                        "fmla       v0.4s, v4.4s, %30.s[1]         \n"
-                        "ext        v4.16b, v1.16b, v2.16b, #12    \n"
-                        "fmla       v9.4s, v5.4s, %30.s[2]         \n"
-                        "ext        v5.16b, v2.16b, v3.16b, #4     \n"
-                        "fmla       v0.4s, v4.4s, %30.s[3]         \n"
-                        "ext        v4.16b, v2.16b, v3.16b, #8     \n"
-                        "fmla       v9.4s, v2.4s, %31.s[0]         \n"
-                        "fmla       v0.4s, v5.4s, %31.s[1]         \n"
-                        "fmla       v9.4s, v4.4s, %31.s[2]         \n"
-
-                        "fadd       v0.4s, v0.4s, v9.4s            \n"
-                        "st1        {v0.4s}, [%1], #16             \n"
-                        "subs       %w0, %w0, #1                   \n"
-                        "bne        0b                             \n"
-
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2),     // %4
-                        "=r"(r3),     // %5
-                        "=r"(r4),     // %6
-                        "=r"(r5),     // %7
-                        "=r"(r6)      // %8
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(r5),
-                        "8"(r6),
-                        "w"(_k0123),     // %18
-                        "w"(_k4567),     // %19
-                        "w"(_k78910),    // %20
-                        "w"(_k11121314), // %21
-                        "w"(_k14151617), // %22
-                        "w"(_k18192021), // %23
-                        "w"(_k21222324), // %24
-                        "w"(_k25262728), // %25
-                        "w"(_k28293031), // %26
-                        "w"(_k32333435), // %27
-                        "w"(_k35363738), // %28
-                        "w"(_k39404142), // %29
-                        "w"(_k42434445), // %30
-                        "w"(_k46474849)  // %31
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v9");
-                }
-#else
-                /**
-                * __ARM_NEON && __aarch64__ defined, but __clang__ not defined
-                * When compiled with gcc, gcc does not accept over 30 operands
-                */
-                for (; nn > 0; nn--)
-                {
-                    float32x4_t _sum = vld1q_f32(outptr);
-
-                    float32x4_t _r00 = vld1q_f32(r0);             // 0 1 2 3
-                    float32x4_t _r04 = vld1q_f32(r0 + 4);         // 4 5 6 7
-                    float32x4_t _r00n = vld1q_f32(r0 + 8);        // 8 9 10 11
-                    float32x4_t _r01 = vextq_f32(_r00, _r04, 1);  // 1 2 3 4
-                    float32x4_t _r02 = vextq_f32(_r00, _r04, 2);  // 2 3 4 5
-                    float32x4_t _r03 = vextq_f32(_r00, _r04, 3);  // 3 4 5 6
-                    float32x4_t _r05 = vextq_f32(_r04, _r00n, 1); // 5 6 7 8
-                    float32x4_t _r06 = vextq_f32(_r04, _r00n, 2); // 6 7 8 9
-
-                    _sum = vfmaq_laneq_f32(_sum, _r00, _k0123, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r01, _k0123, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r02, _k0123, 2);
-                    _sum = vfmaq_laneq_f32(_sum, _r03, _k0123, 3);
-                    _sum = vfmaq_laneq_f32(_sum, _r04, _k4567, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r05, _k4567, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r06, _k4567, 2);
-
-                    float32x4_t _r10 = vld1q_f32(r1);
-                    float32x4_t _r14 = vld1q_f32(r1 + 4);
-                    float32x4_t _r10n = vld1q_f32(r1 + 8);
-                    float32x4_t _r11 = vextq_f32(_r10, _r14, 1);
-                    float32x4_t _r12 = vextq_f32(_r10, _r14, 2);
-                    float32x4_t _r13 = vextq_f32(_r10, _r14, 3);
-                    float32x4_t _r15 = vextq_f32(_r14, _r10n, 1);
-                    float32x4_t _r16 = vextq_f32(_r14, _r10n, 2);
-
-                    _sum = vfmaq_laneq_f32(_sum, _r10, _k78910, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r11, _k78910, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r12, _k78910, 2);
-                    _sum = vfmaq_laneq_f32(_sum, _r13, _k78910, 3);
-                    _sum = vfmaq_laneq_f32(_sum, _r14, _k11121314, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r15, _k11121314, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r16, _k11121314, 2);
-
-                    float32x4_t _r20 = vld1q_f32(r2);
-                    float32x4_t _r24 = vld1q_f32(r2 + 4);
-                    float32x4_t _r20n = vld1q_f32(r2 + 8);
-                    float32x4_t _r21 = vextq_f32(_r20, _r24, 1);
-                    float32x4_t _r22 = vextq_f32(_r20, _r24, 2);
-                    float32x4_t _r23 = vextq_f32(_r20, _r24, 3);
-                    float32x4_t _r25 = vextq_f32(_r24, _r20n, 1);
-                    float32x4_t _r26 = vextq_f32(_r24, _r20n, 2);
-
-                    _sum = vfmaq_laneq_f32(_sum, _r20, _k14151617, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r21, _k14151617, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r22, _k14151617, 2);
-                    _sum = vfmaq_laneq_f32(_sum, _r23, _k14151617, 3);
-                    _sum = vfmaq_laneq_f32(_sum, _r24, _k18192021, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r25, _k18192021, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r26, _k18192021, 2);
-
-                    float32x4_t _r30 = vld1q_f32(r3);
-                    float32x4_t _r34 = vld1q_f32(r3 + 4);
-                    float32x4_t _r30n = vld1q_f32(r3 + 8);
-                    float32x4_t _r31 = vextq_f32(_r30, _r34, 1);
-                    float32x4_t _r32 = vextq_f32(_r30, _r34, 2);
-                    float32x4_t _r33 = vextq_f32(_r30, _r34, 3);
-                    float32x4_t _r35 = vextq_f32(_r34, _r30n, 1);
-                    float32x4_t _r36 = vextq_f32(_r34, _r30n, 2);
-
-                    _sum = vfmaq_laneq_f32(_sum, _r30, _k21222324, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r31, _k21222324, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r32, _k21222324, 2);
-                    _sum = vfmaq_laneq_f32(_sum, _r33, _k21222324, 3);
-                    _sum = vfmaq_laneq_f32(_sum, _r34, _k25262728, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r35, _k25262728, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r36, _k25262728, 2);
-
-                    float32x4_t _r40 = vld1q_f32(r4);
-                    float32x4_t _r44 = vld1q_f32(r4 + 4);
-                    float32x4_t _r40n = vld1q_f32(r4 + 8);
-                    float32x4_t _r41 = vextq_f32(_r40, _r44, 1);
-                    float32x4_t _r42 = vextq_f32(_r40, _r44, 2);
-                    float32x4_t _r43 = vextq_f32(_r40, _r44, 3);
-                    float32x4_t _r45 = vextq_f32(_r44, _r40n, 1);
-                    float32x4_t _r46 = vextq_f32(_r44, _r40n, 2);
-
-                    _sum = vfmaq_laneq_f32(_sum, _r40, _k28293031, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r41, _k28293031, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r42, _k28293031, 2);
-                    _sum = vfmaq_laneq_f32(_sum, _r43, _k28293031, 3);
-                    _sum = vfmaq_laneq_f32(_sum, _r44, _k32333435, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r45, _k32333435, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r46, _k32333435, 2);
-
-                    float32x4_t _r50 = vld1q_f32(r5);
-                    float32x4_t _r54 = vld1q_f32(r5 + 4);
-                    float32x4_t _r50n = vld1q_f32(r5 + 8);
-                    float32x4_t _r51 = vextq_f32(_r50, _r54, 1);
-                    float32x4_t _r52 = vextq_f32(_r50, _r54, 2);
-                    float32x4_t _r53 = vextq_f32(_r50, _r54, 3);
-                    float32x4_t _r55 = vextq_f32(_r54, _r50n, 1);
-                    float32x4_t _r56 = vextq_f32(_r54, _r50n, 2);
-
-                    _sum = vfmaq_laneq_f32(_sum, _r50, _k35363738, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r51, _k35363738, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r52, _k35363738, 2);
-                    _sum = vfmaq_laneq_f32(_sum, _r53, _k35363738, 3);
-                    _sum = vfmaq_laneq_f32(_sum, _r54, _k39404142, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r55, _k39404142, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r56, _k39404142, 2);
-
-                    float32x4_t _r60 = vld1q_f32(r6);
-                    float32x4_t _r64 = vld1q_f32(r6 + 4);
-                    float32x4_t _r60n = vld1q_f32(r6 + 8);
-                    float32x4_t _r61 = vextq_f32(_r60, _r64, 1);
-                    float32x4_t _r62 = vextq_f32(_r60, _r64, 2);
-                    float32x4_t _r63 = vextq_f32(_r60, _r64, 3);
-                    float32x4_t _r65 = vextq_f32(_r64, _r60n, 1);
-                    float32x4_t _r66 = vextq_f32(_r64, _r60n, 2);
-
-                    _sum = vfmaq_laneq_f32(_sum, _r60, _k42434445, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r61, _k42434445, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r62, _k42434445, 2);
-                    _sum = vfmaq_laneq_f32(_sum, _r63, _k42434445, 3);
-                    _sum = vfmaq_laneq_f32(_sum, _r64, _k46474849, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r65, _k46474849, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r66, _k46474849, 2);
-
-                    vst1q_f32(outptr, _sum);
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                    r3 += 4;
-                    r4 += 4;
-                    r5 += 4;
-                    r6 += 4;
-                    outptr += 4;
-                }
-#endif // __clang__
-#else  //__aarch32__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "0:                             \n"
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.f32   {d24-d25}, [%1]     \n" // _sum
-                        //                     "veor       q13, q13            \n"// _sum2 = 0;
-                        //                     "veor       q14, q14            \n"// _sum3 = 0;
-                        //                     "veor       q15, q15            \n"// _sum4 = 0;
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.f32   {d8-d11}, [%9]      \n" // q4 q5 = k0123 k4567
-                        "add        %9, #28             \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%2]!      \n" // q0 = 0  1  2  3
-                        "vmla.f32   q12, q0, d8[0]      \n"
-
-                        "pld        [%2, #256]          \n"
-                        "vld1.f32   {d4-d7}, [%2]       \n" // q2 = 4  5  6  7  q3 = 8  9 10 11
-                        "vmul.f32   q13, q2, d10[0]     \n"
-
-                        "vext.32    q1, q0, q2, #1      \n" // q1 = 1  2  3  4
-                        "vext.32    q10, q2, q3, #1     \n" // q10= 5  6  7  8
-                        "vmul.f32   q14, q1, d8[1]      \n"
-                        "vmul.f32   q15, q10, d10[1]    \n"
-
-                        "vext.32    q8, q0, q2, #2      \n" // q8 = 2  3  4  5
-                        "vext.32    q11, q2, q3, #2     \n" // q11= 6  7  8  9
-                        "vmla.f32   q12, q8, d9[0]      \n"
-                        "vmla.f32   q13, q11, d11[0]    \n"
-
-                        "vext.32    q9, q0, q2, #3      \n" // q9 = 3  4  5  6
-                        "vmla.f32   q14, q9, d9[1]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.f32   {d12-d15}, [%9]     \n" // q6 q7 = k78910 k11121314
-                        "add        %9, #28             \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%3]!      \n"
-                        "vmla.f32   q15, q0, d12[0]     \n"
-
-                        "pld        [%3, #256]          \n"
-                        "vld1.f32   {d4-d7}, [%3]       \n"
-                        "vmla.f32   q12, q2, d14[0]     \n"
-
-                        "vext.32    q1, q0, q2, #1      \n"
-                        "vext.32    q10, q2, q3, #1     \n"
-                        "vmla.f32   q13, q1, d12[1]     \n"
-                        "vmla.f32   q14, q10, d14[1]    \n"
-
-                        "vext.32    q8, q0, q2, #2      \n"
-                        "vext.32    q11, q2, q3, #2     \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vmla.f32   q12, q11, d15[0]    \n"
-
-                        "vext.32    q9, q0, q2, #3      \n"
-                        "vmla.f32   q13, q9, d13[1]     \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.f32   {d8-d11}, [%9]      \n" // q4 q5 = k14151617 k18192021
-                        "add        %9, #28             \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%4]!      \n"
-                        "vmla.f32   q14, q0, d8[0]      \n"
-
-                        "pld        [%4, #256]          \n"
-                        "vld1.f32   {d4-d7}, [%4]       \n"
-                        "vmla.f32   q15, q2, d10[0]     \n"
-
-                        "vext.32    q1, q0, q2, #1      \n"
-                        "vext.32    q10, q2, q3, #1     \n"
-                        "vmla.f32   q12, q1, d8[1]      \n"
-                        "vmla.f32   q13, q10, d10[1]    \n"
-
-                        "vext.32    q8, q0, q2, #2      \n"
-                        "vext.32    q11, q2, q3, #2     \n"
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vmla.f32   q15, q11, d11[0]    \n"
-
-                        "vext.32    q9, q0, q2, #3      \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.f32   {d12-d15}, [%9]     \n" // q6 q7 = k21222324 k25262728
-                        "add        %9, #28             \n"
-
-                        "pld        [%5, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%5]!      \n"
-                        "vmla.f32   q13, q0, d12[0]     \n"
-
-                        "pld        [%5, #256]          \n"
-                        "vld1.f32   {d4-d7}, [%5]       \n"
-                        "vmla.f32   q14, q2, d14[0]     \n"
-
-                        "vext.32    q1, q0, q2, #1      \n"
-                        "vext.32    q10, q2, q3, #1     \n"
-                        "vmla.f32   q15, q1, d12[1]     \n"
-                        "vmla.f32   q12, q10, d14[1]    \n"
-
-                        "vext.32    q8, q0, q2, #2      \n"
-                        "vext.32    q11, q2, q3, #2     \n"
-                        "vmla.f32   q13, q8, d13[0]     \n"
-                        "vmla.f32   q14, q11, d15[0]    \n"
-
-                        "vext.32    q9, q0, q2, #3      \n"
-                        "vmla.f32   q15, q9, d13[1]     \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.f32   {d8-d11}, [%9]      \n" // q4 q5 = k28293031 k32333435
-                        "add        %9, #28             \n"
-
-                        "pld        [%6, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%6]!      \n"
-                        "vmla.f32   q12, q0, d8[0]      \n"
-
-                        "pld        [%6, #256]          \n"
-                        "vld1.f32   {d4-d7}, [%6]       \n"
-                        "vmla.f32   q13, q2, d10[0]     \n"
-
-                        "vext.32    q1, q0, q2, #1      \n"
-                        "vext.32    q10, q2, q3, #1     \n"
-                        "vmla.f32   q14, q1, d8[1]      \n"
-                        "vmla.f32   q15, q10, d10[1]    \n"
-
-                        "vext.32    q8, q0, q2, #2      \n"
-                        "vext.32    q11, q2, q3, #2     \n"
-                        "vmla.f32   q12, q8, d9[0]      \n"
-                        "vmla.f32   q13, q11, d11[0]    \n"
-
-                        "vext.32    q9, q0, q2, #3      \n"
-                        "vmla.f32   q14, q9, d9[1]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.f32   {d12-d15}, [%9]     \n" // q6 q7 = k35363738 k39404142
-                        "add        %9, #28             \n"
-
-                        "pld        [%7, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%7]!      \n"
-                        "vmla.f32   q15, q0, d12[0]     \n"
-
-                        "pld        [%7, #256]          \n"
-                        "vld1.f32   {d4-d7}, [%7]       \n"
-                        "vmla.f32   q12, q2, d14[0]     \n"
-
-                        "vext.32    q1, q0, q2, #1      \n"
-                        "vext.32    q10, q2, q3, #1     \n"
-                        "vmla.f32   q13, q1, d12[1]     \n"
-                        "vmla.f32   q14, q10, d14[1]    \n"
-
-                        "vext.32    q8, q0, q2, #2      \n"
-                        "vext.32    q11, q2, q3, #2     \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vmla.f32   q12, q11, d15[0]    \n"
-
-                        "vext.32    q9, q0, q2, #3      \n"
-                        "vmla.f32   q13, q9, d13[1]     \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.f32   {d8-d11}, [%9]      \n" // q4 q5 = k42434445 k46474849
-                        "sub        %9, #168            \n" // restore k0
-
-                        "pld        [%8, #128]          \n"
-                        "vld1.f32   {d0-d1}, [%8]!      \n"
-                        "vmla.f32   q14, q0, d8[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.f32   {d4-d7}, [%8]       \n"
-                        "vmla.f32   q15, q2, d10[0]     \n"
-
-                        "vext.32    q1, q0, q2, #1      \n"
-                        "vext.32    q10, q2, q3, #1     \n"
-                        "vmla.f32   q12, q1, d8[1]      \n"
-                        "vmla.f32   q13, q10, d10[1]    \n"
-
-                        "vext.32    q8, q0, q2, #2      \n"
-                        "vext.32    q11, q2, q3, #2     \n"
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vmla.f32   q15, q11, d11[0]    \n"
-
-                        "vext.32    q9, q0, q2, #3      \n"
-                        "vmla.f32   q12, q9, d9[1]      \n"
-
-                        "vadd.f32   q13, q13, q14       \n"
-                        "vadd.f32   q13, q13, q15       \n"
-                        "vadd.f32   q12, q12, q13       \n"
-
-                        "vst1.f32   {d24-d25}, [%1]!    \n"
-
-                        "subs       %0, #1              \n"
-                        "bne        0b                  \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2),     // %4
-                        "=r"(r3),     // %5
-                        "=r"(r4),     // %6
-                        "=r"(r5),     // %7
-                        "=r"(r6),     // %8
-                        "=r"(k0)      // %9
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(r5),
-                        "8"(r6),
-                        "9"(k0)
-                        : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-
-                for (; remain > 0; remain--)
-                {
-                    float sum = 0;
-
-                    sum += r0[0] * k0[0];
-                    sum += r0[1] * k0[1];
-                    sum += r0[2] * k0[2];
-                    sum += r0[3] * k0[3];
-                    sum += r0[4] * k0[4];
-                    sum += r0[5] * k0[5];
-                    sum += r0[6] * k0[6];
-
-                    sum += r1[0] * k1[0];
-                    sum += r1[1] * k1[1];
-                    sum += r1[2] * k1[2];
-                    sum += r1[3] * k1[3];
-                    sum += r1[4] * k1[4];
-                    sum += r1[5] * k1[5];
-                    sum += r1[6] * k1[6];
-
-                    sum += r2[0] * k2[0];
-                    sum += r2[1] * k2[1];
-                    sum += r2[2] * k2[2];
-                    sum += r2[3] * k2[3];
-                    sum += r2[4] * k2[4];
-                    sum += r2[5] * k2[5];
-                    sum += r2[6] * k2[6];
-
-                    sum += r3[0] * k3[0];
-                    sum += r3[1] * k3[1];
-                    sum += r3[2] * k3[2];
-                    sum += r3[3] * k3[3];
-                    sum += r3[4] * k3[4];
-                    sum += r3[5] * k3[5];
-                    sum += r3[6] * k3[6];
-
-                    sum += r4[0] * k4[0];
-                    sum += r4[1] * k4[1];
-                    sum += r4[2] * k4[2];
-                    sum += r4[3] * k4[3];
-                    sum += r4[4] * k4[4];
-                    sum += r4[5] * k4[5];
-                    sum += r4[6] * k4[6];
-
-                    sum += r5[0] * k5[0];
-                    sum += r5[1] * k5[1];
-                    sum += r5[2] * k5[2];
-                    sum += r5[3] * k5[3];
-                    sum += r5[4] * k5[4];
-                    sum += r5[5] * k5[5];
-                    sum += r5[6] * k5[6];
-
-                    sum += r6[0] * k6[0];
-                    sum += r6[1] * k6[1];
-                    sum += r6[2] * k6[2];
-                    sum += r6[3] * k6[3];
-                    sum += r6[4] * k6[4];
-                    sum += r6[5] * k6[5];
-                    sum += r6[6] * k6[6];
-
-                    *outptr += sum;
-
-                    r0++;
-                    r1++;
-                    r2++;
-                    r3++;
-                    r4++;
-                    r5++;
-                    r6++;
-                    outptr++;
-                }
-
-                r0 += 6;
-                r1 += 6;
-                r2 += 6;
-                r3 += 6;
-                r4 += 6;
-                r5 += 6;
-                r6 += 6;
-            }
-        }
-    }
-}
-
-static void conv7x7s2_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch * 49 + q * 49;
-
-            const float* r0 = img0;
-            const float* r1 = img0 + w;
-            const float* r2 = img0 + w * 2;
-            const float* r3 = img0 + w * 3;
-            const float* r4 = img0 + w * 4;
-            const float* r5 = img0 + w * 5;
-            const float* r6 = img0 + w * 6;
-
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 7;
-            const float* k2 = kernel0 + 14;
-            const float* k3 = kernel0 + 21;
-            const float* k4 = kernel0 + 28;
-            const float* k5 = kernel0 + 35;
-            const float* k6 = kernel0 + 42;
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-#if __ARM_NEON
-                int nn = outw >> 2;
-                int remain = outw - (nn << 2);
-#else
-                int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-                float32x4_t _k0123 = vld1q_f32(k0);
-                float32x4_t _k4567 = vld1q_f32(k0 + 4);
-                float32x4_t _k78910 = vld1q_f32(k1);
-                float32x4_t _k11121314 = vld1q_f32(k1 + 4);
-                float32x4_t _k14151617 = vld1q_f32(k2);
-                float32x4_t _k18192021 = vld1q_f32(k2 + 4);
-                float32x4_t _k21222324 = vld1q_f32(k3);
-                float32x4_t _k25262728 = vld1q_f32(k3 + 4);
-                float32x4_t _k28293031 = vld1q_f32(k4);
-                float32x4_t _k32333435 = vld1q_f32(k4 + 4);
-                float32x4_t _k35363738 = vld1q_f32(k5);
-                float32x4_t _k39404142 = vld1q_f32(k5 + 4);
-                float32x4_t _k42434445 = vld1q_f32(k6);
-                float32x4_t _k46474849 = vld1q_f32(k6 + 4);
-#ifdef __clang__ // __ARM_NEON && __aarch64__ && __clang__
-                if (nn > 0)
-                {
-                    asm volatile(
-                        // v0:  input / final output
-                        // v1 v2: = _ri0/_ri1  first
-                        // v3 v4: =                  then _r0_8101214/_r0_9111315
-                        // v5 = ri2 / ri4 / ri6
-                        // v6 = ri3 / ri5
-                        // v9 = intermediate sum register
-                        "0:                                        \n"
-                        "prfm       pldl1keep, [%1, #128]          \n"
-                        "ld1        {v0.4s}, [%1]                  \n"
-
-                        //i = 1
-                        "prfm       pldl1keep, [%2, #512]          \n"
-                        "ld2        {v1.4s, v2.4s}, [%2]           \n" // v1  v2 = _r00  _r01
-                        "add        %2, %2, #32                    \n"
-                        "ld2        {v3.4s, v4.4s}, [%2]           \n" // v3  v4 = _r0_8101214 / _r0_9111315
-                        "fmul       v9.4s, v1.4s, %18.s[0]         \n" // *+ _r00
-                        "ext        v5.16b, v1.16b, v3.16b, #4     \n" // v5 = _r02
-                        "fmla       v0.4s, v2.4s, %18.s[1]         \n" // *+ _r01
-                        "ext        v6.16b, v2.16b, v4.16b, #4     \n" // v6 = _r03
-                        "fmla       v9.4s, v5.4s, %18.s[2]         \n" // *+ _r02
-                        "ext        v5.16b, v1.16b, v3.16b, #8     \n" // v5 = _r04
-                        "fmla       v0.4s, v6.4s, %18.s[3]         \n" // *+ _r03
-                        "ext        v6.16b, v2.16b, v4.16b, #8     \n" // v6 = _r05
-                        "fmla       v9.4s, v5.4s, %19.s[0]         \n" // *+ _r04
-                        "ext        v5.16b, v1.16b, v3.16b, #12    \n" // v5 = _r06
-                        "fmla       v0.4s, v6.4s, %19.s[1]         \n" // *+ _r05
-                        "fmla       v9.4s, v5.4s, %19.s[2]         \n" // *+ _r06
-
-                        //i = 2
-                        "prfm       pldl1keep, [%3, #512]          \n"
-                        "ld2        {v1.4s, v2.4s}, [%3]           \n"
-                        "add        %3, %3, #32                    \n"
-                        "ld2        {v3.4s, v4.4s}, [%3]           \n"
-                        "fmla       v9.4s, v1.4s, %20.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #4     \n"
-                        "fmla       v0.4s, v2.4s, %20.s[1]         \n"
-                        "ext        v6.16b, v2.16b, v4.16b, #4     \n"
-                        "fmla       v9.4s, v5.4s, %20.s[2]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #8     \n"
-                        "fmla       v0.4s, v6.4s, %20.s[3]         \n"
-                        "ext        v6.16b, v2.16b, v4.16b, #8     \n"
-                        "fmla       v9.4s, v5.4s, %21.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #12    \n"
-                        "fmla       v0.4s, v6.4s, %21.s[1]         \n"
-                        "fmla       v9.4s, v5.4s, %21.s[2]         \n"
-
-                        //i = 3
-                        "prfm       pldl1keep, [%4, #512]          \n"
-                        "ld2        {v1.4s, v2.4s}, [%4]           \n"
-                        "add        %4, %4, #32                    \n"
-                        "ld2        {v3.4s, v4.4s}, [%4]           \n"
-                        "fmla       v9.4s, v1.4s, %22.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #4     \n"
-                        "fmla       v0.4s, v2.4s, %22.s[1]         \n"
-                        "ext        v6.16b, v2.16b, v4.16b, #4     \n"
-                        "fmla       v9.4s, v5.4s, %22.s[2]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #8     \n"
-                        "fmla       v0.4s, v6.4s, %22.s[3]         \n"
-                        "ext        v6.16b, v2.16b, v4.16b, #8     \n"
-                        "fmla       v9.4s, v5.4s, %23.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #12    \n"
-                        "fmla       v0.4s, v6.4s, %23.s[1]         \n"
-                        "fmla       v9.4s, v5.4s, %23.s[2]         \n"
-
-                        //i = 4
-                        "prfm       pldl1keep, [%5, #512]          \n"
-                        "ld2        {v1.4s, v2.4s}, [%5]           \n"
-                        "add        %5, %5, #32                    \n"
-                        "ld2        {v3.4s, v4.4s}, [%5]           \n"
-                        "fmla       v9.4s, v1.4s, %24.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #4     \n"
-                        "fmla       v0.4s, v2.4s, %24.s[1]         \n"
-                        "ext        v6.16b, v2.16b, v4.16b, #4     \n"
-                        "fmla       v9.4s, v5.4s, %24.s[2]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #8     \n"
-                        "fmla       v0.4s, v6.4s, %24.s[3]         \n"
-                        "ext        v6.16b, v2.16b, v4.16b, #8     \n"
-                        "fmla       v9.4s, v5.4s, %25.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #12    \n"
-                        "fmla       v0.4s, v6.4s, %25.s[1]         \n"
-                        "fmla       v9.4s, v5.4s, %25.s[2]         \n"
-
-                        //i = 5
-                        "prfm       pldl1keep, [%6, #512]          \n"
-                        "ld2        {v1.4s, v2.4s}, [%6]           \n"
-                        "add        %6, %6, #32                    \n"
-                        "ld2        {v3.4s, v4.4s}, [%6]           \n"
-                        "fmla       v9.4s, v1.4s, %26.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #4     \n"
-                        "fmla       v0.4s, v2.4s, %26.s[1]         \n"
-                        "ext        v6.16b, v2.16b, v4.16b, #4     \n"
-                        "fmla       v9.4s, v5.4s, %26.s[2]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #8     \n"
-                        "fmla       v0.4s, v6.4s, %26.s[3]         \n"
-                        "ext        v6.16b, v2.16b, v4.16b, #8     \n"
-                        "fmla       v9.4s, v5.4s, %27.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #12    \n"
-                        "fmla       v0.4s, v6.4s, %27.s[1]         \n"
-                        "fmla       v9.4s, v5.4s, %27.s[2]         \n"
-
-                        //i = 6
-                        "prfm       pldl1keep, [%7, #512]          \n"
-                        "ld2        {v1.4s, v2.4s}, [%7]           \n"
-                        "add        %7, %7, #32                    \n"
-                        "ld2        {v3.4s, v4.4s}, [%7]           \n"
-                        "fmla       v9.4s, v1.4s, %28.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #4     \n"
-                        "fmla       v0.4s, v2.4s, %28.s[1]         \n"
-                        "ext        v6.16b, v2.16b, v4.16b, #4     \n"
-                        "fmla       v9.4s, v5.4s, %28.s[2]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #8     \n"
-                        "fmla       v0.4s, v6.4s, %28.s[3]         \n"
-                        "ext        v6.16b, v2.16b, v4.16b, #8     \n"
-                        "fmla       v9.4s, v5.4s, %29.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #12    \n"
-                        "fmla       v0.4s, v6.4s, %29.s[1]         \n"
-                        "fmla       v9.4s, v5.4s, %29.s[2]         \n"
-
-                        //i = 7
-                        "prfm       pldl1keep, [%8, #512]          \n"
-                        "ld2        {v1.4s, v2.4s}, [%8]           \n"
-                        "add        %8, %8, #32                    \n"
-                        "ld2        {v3.4s, v4.4s}, [%8]           \n"
-                        "fmla       v9.4s, v1.4s, %30.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #4     \n"
-                        "fmla       v0.4s, v2.4s, %30.s[1]         \n"
-                        "ext        v6.16b, v2.16b, v4.16b, #4     \n"
-                        "fmla       v9.4s, v5.4s, %30.s[2]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #8     \n"
-                        "fmla       v0.4s, v6.4s, %30.s[3]         \n"
-                        "ext        v6.16b, v2.16b, v4.16b, #8     \n"
-                        "fmla       v9.4s, v5.4s, %31.s[0]         \n"
-                        "ext        v5.16b, v1.16b, v3.16b, #12    \n"
-                        "fmla       v0.4s, v6.4s, %31.s[1]         \n"
-                        "fmla       v9.4s, v5.4s, %31.s[2]         \n"
-
-                        "fadd       v0.4s, v0.4s, v9.4s            \n"
-                        "st1        {v0.4s}, [%1], #16             \n"
-                        "subs       %w0, %w0, #1                   \n"
-                        "bne        0b                             \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2),     // %4
-                        "=r"(r3),     // %5
-                        "=r"(r4),     // %6
-                        "=r"(r5),     // %7
-                        "=r"(r6)      // %8
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(r5),
-                        "8"(r6),
-                        "w"(_k0123),     // %18
-                        "w"(_k4567),     // %19
-                        "w"(_k78910),    // %20
-                        "w"(_k11121314), // %21
-                        "w"(_k14151617), // %22
-                        "w"(_k18192021), // %23
-                        "w"(_k21222324), // %24
-                        "w"(_k25262728), // %25
-                        "w"(_k28293031), // %26
-                        "w"(_k32333435), // %27
-                        "w"(_k35363738), // %28
-                        "w"(_k39404142), // %29
-                        "w"(_k42434445), // %30
-                        "w"(_k46474849)  // %31
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v9");
-                }
-#else
-                /**
-                * __ARM_NEON && __aarch64__ defined, but __clang__ not defined
-                * When compiled with gcc, gcc does not accept over 30 operands
-                */
-                for (; nn > 0; nn--)
-                {
-                    float32x4_t _sum = vld1q_f32(outptr);
-
-                    float32x4x2_t _r00_02461357 = vld2q_f32(r0);
-                    float32x4x2_t _r00nx2 = vld2q_f32(r0 + 8);
-                    float32x4_t _r0_8101214 = _r00nx2.val[0];           // 8 10 12 14
-                    float32x4_t _r0_9111315 = _r00nx2.val[1];           // 9 11 13 15
-                    float32x4_t _r00 = _r00_02461357.val[0];            // 0 2 4 6
-                    float32x4_t _r01 = _r00_02461357.val[1];            // 1 3 5 7
-                    float32x4_t _r02 = vextq_f32(_r00, _r0_8101214, 1); // 2 4 6 8
-                    float32x4_t _r03 = vextq_f32(_r01, _r0_9111315, 1); // 3 5 7 9
-                    float32x4_t _r04 = vextq_f32(_r00, _r0_8101214, 2); // 4 6 8 10
-                    float32x4_t _r05 = vextq_f32(_r01, _r0_9111315, 2); // 5 7 9 11
-                    float32x4_t _r06 = vextq_f32(_r00, _r0_8101214, 3); // 6 8 10 12
-
-                    _sum = vfmaq_laneq_f32(_sum, _r00, _k0123, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r01, _k0123, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r02, _k0123, 2);
-                    _sum = vfmaq_laneq_f32(_sum, _r03, _k0123, 3);
-                    _sum = vfmaq_laneq_f32(_sum, _r04, _k4567, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r05, _k4567, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r06, _k4567, 2);
-
-                    float32x4x2_t _r10_02461357 = vld2q_f32(r1);
-                    float32x4x2_t _r10nx2 = vld2q_f32(r1 + 8);
-                    float32x4_t _r1_8101214 = _r10nx2.val[0];
-                    float32x4_t _r1_9111315 = _r10nx2.val[1];
-                    float32x4_t _r10 = _r10_02461357.val[0];
-                    float32x4_t _r11 = _r10_02461357.val[1];
-                    float32x4_t _r12 = vextq_f32(_r10, _r1_8101214, 1);
-                    float32x4_t _r13 = vextq_f32(_r11, _r1_9111315, 1);
-                    float32x4_t _r14 = vextq_f32(_r10, _r1_8101214, 2);
-                    float32x4_t _r15 = vextq_f32(_r11, _r1_9111315, 2);
-                    float32x4_t _r16 = vextq_f32(_r10, _r1_8101214, 3);
-
-                    _sum = vfmaq_laneq_f32(_sum, _r10, _k78910, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r11, _k78910, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r12, _k78910, 2);
-                    _sum = vfmaq_laneq_f32(_sum, _r13, _k78910, 3);
-                    _sum = vfmaq_laneq_f32(_sum, _r14, _k11121314, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r15, _k11121314, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r16, _k11121314, 2);
-
-                    float32x4x2_t _r20_02461357 = vld2q_f32(r2);
-                    float32x4x2_t _r20nx2 = vld2q_f32(r2 + 8);
-                    float32x4_t _r2_8101214 = _r20nx2.val[0];
-                    float32x4_t _r2_9111315 = _r20nx2.val[1];
-                    float32x4_t _r20 = _r20_02461357.val[0];
-                    float32x4_t _r21 = _r20_02461357.val[1];
-                    float32x4_t _r22 = vextq_f32(_r20, _r2_8101214, 1);
-                    float32x4_t _r23 = vextq_f32(_r21, _r2_9111315, 1);
-                    float32x4_t _r24 = vextq_f32(_r20, _r2_8101214, 2);
-                    float32x4_t _r25 = vextq_f32(_r21, _r2_9111315, 2);
-                    float32x4_t _r26 = vextq_f32(_r20, _r2_8101214, 3);
-
-                    _sum = vfmaq_laneq_f32(_sum, _r20, _k14151617, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r21, _k14151617, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r22, _k14151617, 2);
-                    _sum = vfmaq_laneq_f32(_sum, _r23, _k14151617, 3);
-                    _sum = vfmaq_laneq_f32(_sum, _r24, _k18192021, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r25, _k18192021, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r26, _k18192021, 2);
-
-                    float32x4x2_t _r30_02461357 = vld2q_f32(r3);
-                    float32x4x2_t _r30nx2 = vld2q_f32(r3 + 8);
-                    float32x4_t _r3_8101214 = _r30nx2.val[0];
-                    float32x4_t _r3_9111315 = _r30nx2.val[1];
-                    float32x4_t _r30 = _r30_02461357.val[0];
-                    float32x4_t _r31 = _r30_02461357.val[1];
-                    float32x4_t _r32 = vextq_f32(_r30, _r3_8101214, 1);
-                    float32x4_t _r33 = vextq_f32(_r31, _r3_9111315, 1);
-                    float32x4_t _r34 = vextq_f32(_r30, _r3_8101214, 2);
-                    float32x4_t _r35 = vextq_f32(_r31, _r3_9111315, 2);
-                    float32x4_t _r36 = vextq_f32(_r30, _r3_8101214, 3);
-
-                    _sum = vfmaq_laneq_f32(_sum, _r30, _k21222324, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r31, _k21222324, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r32, _k21222324, 2);
-                    _sum = vfmaq_laneq_f32(_sum, _r33, _k21222324, 3);
-                    _sum = vfmaq_laneq_f32(_sum, _r34, _k25262728, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r35, _k25262728, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r36, _k25262728, 2);
-
-                    float32x4x2_t _r40_02461357 = vld2q_f32(r4);
-                    float32x4x2_t _r40nx2 = vld2q_f32(r4 + 8);
-                    float32x4_t _r4_8101214 = _r40nx2.val[0];
-                    float32x4_t _r4_9111315 = _r40nx2.val[1];
-                    float32x4_t _r40 = _r40_02461357.val[0];
-                    float32x4_t _r41 = _r40_02461357.val[1];
-                    float32x4_t _r42 = vextq_f32(_r40, _r4_8101214, 1);
-                    float32x4_t _r43 = vextq_f32(_r41, _r4_9111315, 1);
-                    float32x4_t _r44 = vextq_f32(_r40, _r4_8101214, 2);
-                    float32x4_t _r45 = vextq_f32(_r41, _r4_9111315, 2);
-                    float32x4_t _r46 = vextq_f32(_r40, _r4_8101214, 3);
-
-                    _sum = vfmaq_laneq_f32(_sum, _r40, _k28293031, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r41, _k28293031, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r42, _k28293031, 2);
-                    _sum = vfmaq_laneq_f32(_sum, _r43, _k28293031, 3);
-                    _sum = vfmaq_laneq_f32(_sum, _r44, _k32333435, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r45, _k32333435, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r46, _k32333435, 2);
-
-                    float32x4x2_t _r50_02461357 = vld2q_f32(r5);
-                    float32x4x2_t _r50nx2 = vld2q_f32(r5 + 8);
-                    float32x4_t _r5_8101214 = _r50nx2.val[0];
-                    float32x4_t _r5_9111315 = _r50nx2.val[1];
-                    float32x4_t _r50 = _r50_02461357.val[0];
-                    float32x4_t _r51 = _r50_02461357.val[1];
-                    float32x4_t _r52 = vextq_f32(_r50, _r5_8101214, 1);
-                    float32x4_t _r53 = vextq_f32(_r51, _r5_9111315, 1);
-                    float32x4_t _r54 = vextq_f32(_r50, _r5_8101214, 2);
-                    float32x4_t _r55 = vextq_f32(_r51, _r5_9111315, 2);
-                    float32x4_t _r56 = vextq_f32(_r50, _r5_8101214, 3);
-
-                    _sum = vfmaq_laneq_f32(_sum, _r50, _k35363738, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r51, _k35363738, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r52, _k35363738, 2);
-                    _sum = vfmaq_laneq_f32(_sum, _r53, _k35363738, 3);
-                    _sum = vfmaq_laneq_f32(_sum, _r54, _k39404142, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r55, _k39404142, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r56, _k39404142, 2);
-
-                    float32x4x2_t _r60_02461357 = vld2q_f32(r6);
-                    float32x4x2_t _r60nx2 = vld2q_f32(r6 + 8);
-                    float32x4_t _r6_8101214 = _r60nx2.val[0];
-                    float32x4_t _r6_9111315 = _r60nx2.val[1];
-                    float32x4_t _r60 = _r60_02461357.val[0];
-                    float32x4_t _r61 = _r60_02461357.val[1];
-                    float32x4_t _r62 = vextq_f32(_r60, _r6_8101214, 1);
-                    float32x4_t _r63 = vextq_f32(_r61, _r6_9111315, 1);
-                    float32x4_t _r64 = vextq_f32(_r60, _r6_8101214, 2);
-                    float32x4_t _r65 = vextq_f32(_r61, _r6_9111315, 2);
-                    float32x4_t _r66 = vextq_f32(_r60, _r6_8101214, 3);
-
-                    _sum = vfmaq_laneq_f32(_sum, _r60, _k42434445, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r61, _k42434445, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r62, _k42434445, 2);
-                    _sum = vfmaq_laneq_f32(_sum, _r63, _k42434445, 3);
-                    _sum = vfmaq_laneq_f32(_sum, _r64, _k46474849, 0);
-                    _sum = vfmaq_laneq_f32(_sum, _r65, _k46474849, 1);
-                    _sum = vfmaq_laneq_f32(_sum, _r66, _k46474849, 2);
-
-                    vst1q_f32(outptr, _sum);
-
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                    r3 += 8;
-                    r4 += 8;
-                    r5 += 8;
-                    r6 += 8;
-                    outptr += 4;
-                }
-#endif // __clang__
-#else
-                if (nn > 0)
-                {
-                    asm volatile(
-                        "0:                             \n"
-
-                        "pld        [%1, #256]          \n"
-                        "vld1.f32   {d26-d27}, [%1]     \n" // _sum
-                        //                     "veor       q14, q14            \n"// _sum2 = 0;
-                        //                     "veor       q15, q15            \n"// _sum3 = 0;
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.f32   {d8-d11}, [%9]      \n" // q4 q5 = k0123 k4567
-                        "add        %9, #28             \n"
-
-                        "pld        [%2, #512]          \n"
-                        "vld2.f32   {d0-d3}, [%2]!      \n" // q0 = 0  2  4  6  q1 = 1  3  5  7
-                        "vmla.f32   q13, q0, d8[0]      \n"
-                        "vmul.f32   q14, q1, d8[1]      \n"
-
-                        "vld2.f32   {d4-d7}, [%2]       \n" // q2 = 8 10 12 14  q3 = 9 11 13 15
-                        "vext.32    q8, q0, q2, #1      \n" // q8 = 2  4  6  8
-                        "vext.32    q9, q1, q3, #1      \n" // q9 = 3  5  7  9
-                        "vmul.f32   q15, q8, d9[0]      \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-
-                        "vext.32    q10, q0, q2, #2     \n" // q10= 4  6  8 10
-                        "vext.32    q11, q1, q3, #2     \n" // q11= 5  7  9 11
-                        "vmla.f32   q14, q10, d10[0]    \n"
-                        "vmla.f32   q15, q11, d10[1]    \n"
-
-                        "vext.32    q12, q0, q2, #3     \n" // q12= 6  8 10 12
-                        "vmla.f32   q13, q12, d11[0]    \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.f32   {d12-d15}, [%9]     \n" // q6 q7 = k78910 k11121314
-                        "add        %9, #28             \n"
-
-                        "pld        [%3, #512]          \n"
-                        "vld2.f32   {d0-d3}, [%3]!      \n"
-                        "vmla.f32   q14, q0, d12[0]     \n"
-                        "vmla.f32   q15, q1, d12[1]     \n"
-
-                        "vld2.f32   {d4-d7}, [%3]       \n"
-                        "vext.32    q8, q0, q2, #1      \n"
-                        "vext.32    q9, q1, q3, #1      \n"
-                        "vmla.f32   q13, q8, d13[0]     \n"
-                        "vmla.f32   q14, q9, d13[1]     \n"
-
-                        "vext.32    q10, q0, q2, #2     \n"
-                        "vext.32    q11, q1, q3, #2     \n"
-                        "vmla.f32   q15, q10, d14[0]    \n"
-                        "vmla.f32   q13, q11, d14[1]    \n"
-
-                        "vext.32    q12, q0, q2, #3     \n"
-                        "vmla.f32   q14, q12, d15[0]    \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.f32   {d8-d11}, [%9]      \n" // q4 q5 = k14151617 k18192021
-                        "add        %9, #28             \n"
-
-                        "pld        [%4, #512]          \n"
-                        "vld2.f32   {d0-d3}, [%4]!      \n"
-                        "vmla.f32   q15, q0, d8[0]      \n"
-                        "vmla.f32   q13, q1, d8[1]      \n"
-
-                        "vld2.f32   {d4-d7}, [%4]       \n"
-                        "vext.32    q8, q0, q2, #1      \n"
-                        "vext.32    q9, q1, q3, #1      \n"
-                        "vmla.f32   q14, q8, d9[0]      \n"
-                        "vmla.f32   q15, q9, d9[1]      \n"
-
-                        "vext.32    q10, q0, q2, #2     \n"
-                        "vext.32    q11, q1, q3, #2     \n"
-                        "vmla.f32   q13, q10, d10[0]    \n"
-                        "vmla.f32   q14, q11, d10[1]    \n"
-
-                        "vext.32    q12, q0, q2, #3     \n"
-                        "vmla.f32   q15, q12, d11[0]    \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.f32   {d12-d15}, [%9]     \n" // q6 q7 = k21222324 k25262728
-                        "add        %9, #28             \n"
-
-                        "pld        [%5, #512]          \n"
-                        "vld2.f32   {d0-d3}, [%5]!      \n"
-                        "vmla.f32   q13, q0, d12[0]     \n"
-                        "vmla.f32   q14, q1, d12[1]     \n"
-
-                        "vld2.f32   {d4-d7}, [%5]       \n"
-                        "vext.32    q8, q0, q2, #1      \n"
-                        "vext.32    q9, q1, q3, #1      \n"
-                        "vmla.f32   q15, q8, d13[0]     \n"
-                        "vmla.f32   q13, q9, d13[1]     \n"
-
-                        "vext.32    q10, q0, q2, #2     \n"
-                        "vext.32    q11, q1, q3, #2     \n"
-                        "vmla.f32   q14, q10, d14[0]    \n"
-                        "vmla.f32   q15, q11, d14[1]    \n"
-
-                        "vext.32    q12, q0, q2, #3     \n"
-                        "vmla.f32   q13, q12, d15[0]    \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.f32   {d8-d11}, [%9]      \n" // q4 q5 = k28293031 k32333435
-                        "add        %9, #28             \n"
-
-                        "pld        [%6, #512]          \n"
-                        "vld2.f32   {d0-d3}, [%6]!      \n"
-                        "vmla.f32   q14, q0, d8[0]      \n"
-                        "vmla.f32   q15, q1, d8[1]      \n"
-
-                        "vld2.f32   {d4-d7}, [%6]       \n"
-                        "vext.32    q8, q0, q2, #1      \n"
-                        "vext.32    q9, q1, q3, #1      \n"
-                        "vmla.f32   q13, q8, d9[0]      \n"
-                        "vmla.f32   q14, q9, d9[1]      \n"
-
-                        "vext.32    q10, q0, q2, #2     \n"
-                        "vext.32    q11, q1, q3, #2     \n"
-                        "vmla.f32   q15, q10, d10[0]    \n"
-                        "vmla.f32   q13, q11, d10[1]    \n"
-
-                        "vext.32    q12, q0, q2, #3     \n"
-                        "vmla.f32   q14, q12, d11[0]    \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.f32   {d12-d15}, [%9]     \n" // q6 q7 = k35363738 k39404142
-                        "add        %9, #28             \n"
-
-                        "pld        [%7, #512]          \n"
-                        "vld2.f32   {d0-d3}, [%7]!      \n"
-                        "vmla.f32   q15, q0, d12[0]     \n"
-                        "vmla.f32   q13, q1, d12[1]     \n"
-
-                        "vld2.f32   {d4-d7}, [%7]       \n"
-                        "vext.32    q8, q0, q2, #1      \n"
-                        "vext.32    q9, q1, q3, #1      \n"
-                        "vmla.f32   q14, q8, d13[0]     \n"
-                        "vmla.f32   q15, q9, d13[1]     \n"
-
-                        "vext.32    q10, q0, q2, #2     \n"
-                        "vext.32    q11, q1, q3, #2     \n"
-                        "vmla.f32   q13, q10, d14[0]    \n"
-                        "vmla.f32   q14, q11, d14[1]    \n"
-
-                        "vext.32    q12, q0, q2, #3     \n"
-                        "vmla.f32   q15, q12, d15[0]    \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.f32   {d8-d11}, [%9]      \n" // q4 q5 = k42434445 k46474849
-                        "sub        %9, #168            \n" // restore k0
-
-                        "pld        [%8, #512]          \n"
-                        "vld2.f32   {d0-d3}, [%8]!      \n"
-                        "vmla.f32   q13, q0, d8[0]      \n"
-                        "vmla.f32   q14, q1, d8[1]      \n"
-
-                        "vld2.f32   {d4-d7}, [%8]       \n"
-                        "vext.32    q8, q0, q2, #1      \n"
-                        "vext.32    q9, q1, q3, #1      \n"
-                        "vmla.f32   q15, q8, d9[0]      \n"
-                        "vmla.f32   q13, q9, d9[1]      \n"
-
-                        "vext.32    q10, q0, q2, #2     \n"
-                        "vext.32    q11, q1, q3, #2     \n"
-                        "vmla.f32   q14, q10, d10[0]    \n"
-                        "vmla.f32   q15, q11, d10[1]    \n"
-
-                        "vext.32    q12, q0, q2, #3     \n"
-                        "vmla.f32   q13, q12, d11[0]    \n"
-
-                        "vadd.f32   q14, q14, q15       \n"
-                        "vadd.f32   q13, q13, q14       \n"
-
-                        "vst1.f32   {d26-d27}, [%1]!    \n"
-
-                        "subs       %0, #1              \n"
-                        "bne        0b                  \n"
-                        : "=r"(nn),     // %0
-                        "=r"(outptr), // %1
-                        "=r"(r0),     // %2
-                        "=r"(r1),     // %3
-                        "=r"(r2),     // %4
-                        "=r"(r3),     // %5
-                        "=r"(r4),     // %6
-                        "=r"(r5),     // %7
-                        "=r"(r6),     // %8
-                        "=r"(k0)      // %9
-                        : "0"(nn),
-                        "1"(outptr),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(r5),
-                        "8"(r6),
-                        "9"(k0)
-                        : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-                }
-#endif // __aarch64__
-#endif // __ARM_NEON
-
-                for (; remain > 0; remain--)
-                {
-                    float sum = 0;
-
-                    sum += r0[0] * k0[0];
-                    sum += r0[1] * k0[1];
-                    sum += r0[2] * k0[2];
-                    sum += r0[3] * k0[3];
-                    sum += r0[4] * k0[4];
-                    sum += r0[5] * k0[5];
-                    sum += r0[6] * k0[6];
-
-                    sum += r1[0] * k1[0];
-                    sum += r1[1] * k1[1];
-                    sum += r1[2] * k1[2];
-                    sum += r1[3] * k1[3];
-                    sum += r1[4] * k1[4];
-                    sum += r1[5] * k1[5];
-                    sum += r1[6] * k1[6];
-
-                    sum += r2[0] * k2[0];
-                    sum += r2[1] * k2[1];
-                    sum += r2[2] * k2[2];
-                    sum += r2[3] * k2[3];
-                    sum += r2[4] * k2[4];
-                    sum += r2[5] * k2[5];
-                    sum += r2[6] * k2[6];
-
-                    sum += r3[0] * k3[0];
-                    sum += r3[1] * k3[1];
-                    sum += r3[2] * k3[2];
-                    sum += r3[3] * k3[3];
-                    sum += r3[4] * k3[4];
-                    sum += r3[5] * k3[5];
-                    sum += r3[6] * k3[6];
-
-                    sum += r4[0] * k4[0];
-                    sum += r4[1] * k4[1];
-                    sum += r4[2] * k4[2];
-                    sum += r4[3] * k4[3];
-                    sum += r4[4] * k4[4];
-                    sum += r4[5] * k4[5];
-                    sum += r4[6] * k4[6];
-
-                    sum += r5[0] * k5[0];
-                    sum += r5[1] * k5[1];
-                    sum += r5[2] * k5[2];
-                    sum += r5[3] * k5[3];
-                    sum += r5[4] * k5[4];
-                    sum += r5[5] * k5[5];
-                    sum += r5[6] * k5[6];
-
-                    sum += r6[0] * k6[0];
-                    sum += r6[1] * k6[1];
-                    sum += r6[2] * k6[2];
-                    sum += r6[3] * k6[3];
-                    sum += r6[4] * k6[4];
-                    sum += r6[5] * k6[5];
-                    sum += r6[6] * k6[6];
-
-                    *outptr += sum;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    r3 += 2;
-                    r4 += 2;
-                    r5 += 2;
-                    r6 += 2;
-                    outptr++;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-                r4 += tailstep;
-                r5 += tailstep;
-                r6 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7_pack1to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7_pack1to4.h
deleted file mode 100644
index 5744d499..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7_pack1to4.h
+++ /dev/null
@@ -1,2109 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv7x7s2_pack1to4_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        out0.fill(_bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-            const float* r3 = img0.row(3);
-            const float* r4 = img0.row(4);
-            const float* r5 = img0.row(5);
-            const float* r6 = img0.row(6);
-
-            const float* kptr = (const float*)kernel.channel(p).row(q);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-#if __aarch64__
-                for (; j + 7 < outw; j += 8)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n" // r0
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0] \n"
-
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[2]     \n"
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v4.4s, v5.4s}, [%1]        \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[0]     \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[1]     \n"
-
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v28.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v29.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v6.4s, v7.4s, v8.4s, v9.4s}, [%2], #64 \n" // r1
-
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v30.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v30.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v10.4s, v11.4s}, [%2]      \n"
-
-                        "fmla   v16.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v6.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v8.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v8.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v9.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v9.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v6.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v7.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v8.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v8.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v9.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v9.s[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v7.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v8.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v8.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v9.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v9.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v10.s[0]    \n"
-
-                        "fmla   v16.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v7.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v8.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v8.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v9.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v9.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v10.s[1]    \n"
-
-                        "fmla   v16.4s, v28.4s, v7.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v7.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v8.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v8.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v9.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v9.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v10.s[0]    \n"
-                        "fmla   v23.4s, v28.4s, v10.s[2]    \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v7.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v7.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v8.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v8.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v9.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v9.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v10.s[1]    \n"
-                        "fmla   v23.4s, v29.4s, v10.s[3]    \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%3], #64 \n" // r2
-
-                        "fmla   v16.4s, v30.4s, v7.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v8.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v8.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v9.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v9.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v10.s[0]    \n"
-                        "fmla   v22.4s, v30.4s, v10.s[2]    \n"
-                        "fmla   v23.4s, v30.4s, v11.s[0]    \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v4.4s, v5.4s}, [%3]        \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[0]     \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[1]     \n"
-
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v28.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v29.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v6.4s, v7.4s, v8.4s, v9.4s}, [%4], #64 \n" // r3
-
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v30.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v30.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v10.4s, v11.4s}, [%4]      \n"
-
-                        "fmla   v16.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v6.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v8.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v8.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v9.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v9.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v6.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v7.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v8.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v8.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v9.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v9.s[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v7.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v8.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v8.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v9.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v9.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v10.s[0]    \n"
-
-                        "fmla   v16.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v7.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v8.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v8.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v9.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v9.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v10.s[1]    \n"
-
-                        "fmla   v16.4s, v28.4s, v7.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v7.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v8.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v8.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v9.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v9.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v10.s[0]    \n"
-                        "fmla   v23.4s, v28.4s, v10.s[2]    \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v7.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v7.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v8.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v8.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v9.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v9.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v10.s[1]    \n"
-                        "fmla   v23.4s, v29.4s, v10.s[3]    \n"
-
-                        "prfm   pldl1keep, [%5, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%5], #64 \n" // r4
-
-                        "fmla   v16.4s, v30.4s, v7.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v8.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v8.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v9.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v9.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v10.s[0]    \n"
-                        "fmla   v22.4s, v30.4s, v10.s[2]    \n"
-                        "fmla   v23.4s, v30.4s, v11.s[0]    \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v4.4s, v5.4s}, [%5]        \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[0]     \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[1]     \n"
-
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v28.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v29.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v6.4s, v7.4s, v8.4s, v9.4s}, [%6], #64 \n" // r5
-
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v30.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v30.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v10.4s, v11.4s}, [%6]      \n"
-
-                        "fmla   v16.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v6.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v8.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v8.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v9.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v9.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v6.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v7.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v8.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v8.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v9.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v9.s[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v7.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v8.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v8.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v9.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v9.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v10.s[0]    \n"
-
-                        "fmla   v16.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v7.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v8.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v8.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v9.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v9.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v10.s[1]    \n"
-
-                        "fmla   v16.4s, v28.4s, v7.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v7.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v8.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v8.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v9.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v9.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v10.s[0]    \n"
-                        "fmla   v23.4s, v28.4s, v10.s[2]    \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v7.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v7.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v8.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v8.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v9.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v9.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v10.s[1]    \n"
-                        "fmla   v23.4s, v29.4s, v10.s[3]    \n"
-
-                        "prfm   pldl1keep, [%7, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%7], #64 \n" // r6
-
-                        "fmla   v16.4s, v30.4s, v7.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v8.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v8.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v9.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v9.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v10.s[0]    \n"
-                        "fmla   v22.4s, v30.4s, v10.s[2]    \n"
-                        "fmla   v23.4s, v30.4s, v11.s[0]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v4.4s, v5.4s}, [%7]        \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[0]     \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[1]     \n"
-
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v28.4s, v4.s[2]     \n"
-
-                        "sub    %0, %0, #64                 \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v29.4s, v4.s[3]     \n"
-
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v30.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v30.4s, v5.s[0]     \n"
-
-                        "sub    %8, %8, #784                \n"
-
-                        "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                        "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v16", "v17", "v18", "v19", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
-                }
-#endif // __aarch64__
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0] \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1] \n" // r0
-                        "add    %1, %1, #32                 \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%2, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%2] \n" // r1
-                        "add    %2, %2, #32                 \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v5.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%3, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%3] \n" // r2
-                        "add    %3, %3, #32                 \n"
-
-                        "fmla   v16.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v6.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v6.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v6.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v6.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v6.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v6.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v7.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%4, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%4] \n" // r3
-                        "add    %4, %4, #32                 \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v5.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%5, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%5] \n" // r4
-                        "add    %5, %5, #32                 \n"
-
-                        "fmla   v16.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v6.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v6.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v6.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v6.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v6.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v6.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v7.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%6, #512]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%6] \n" // r5
-                        "add    %6, %6, #32                 \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v5.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%7, #512]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%7] \n" // r6
-                        "add    %7, %7, #32                 \n"
-
-                        "fmla   v16.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v6.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v6.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v6.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v6.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v6.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v6.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v7.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-
-                        "sub    %8, %8, #784                \n"
-
-                        "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #512]      \n"
-                        "vldm       %0, {d24-d31}   \n"
-
-                        "pld        [%1, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%1]!  \n" // r0
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d10-d17}  \n"
-
-                        "vmla.f32   q12, q5, d0[0]  \n"
-                        "vmla.f32   q13, q5, d1[0]  \n"
-                        "vmla.f32   q14, q5, d2[0]  \n"
-                        "vmla.f32   q15, q5, d3[0]  \n"
-
-                        "pld        [%1, #192]      \n"
-                        "vld1.f32   {d4-d6}, [%1]   \n"
-
-                        "vmla.f32   q12, q6, d0[1]  \n"
-                        "vmla.f32   q13, q6, d1[1]  \n"
-                        "vmla.f32   q14, q6, d2[1]  \n"
-                        "vmla.f32   q15, q6, d3[1]  \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d18-d23}  \n"
-
-                        "vmla.f32   q12, q7, d1[0]  \n"
-                        "vmla.f32   q13, q7, d2[0]  \n"
-                        "vmla.f32   q14, q7, d3[0]  \n"
-                        "vmla.f32   q15, q7, d4[0]  \n"
-                        "vmla.f32   q12, q8, d1[1]  \n"
-                        "vmla.f32   q13, q8, d2[1]  \n"
-                        "vmla.f32   q14, q8, d3[1]  \n"
-                        "vmla.f32   q15, q8, d4[1]  \n"
-                        "vmla.f32   q12, q9, d2[0]  \n"
-                        "vmla.f32   q13, q9, d3[0]  \n"
-                        "vmla.f32   q14, q9, d4[0]  \n"
-                        "vmla.f32   q15, q9, d5[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d10-d17}  \n"
-
-                        "vmla.f32   q12, q10, d2[1] \n"
-                        "vmla.f32   q13, q10, d3[1] \n"
-                        "vmla.f32   q14, q10, d4[1] \n"
-                        "vmla.f32   q15, q10, d5[1] \n"
-                        "vmla.f32   q12, q11, d3[0] \n"
-                        "vmla.f32   q13, q11, d4[0] \n"
-
-                        "pld        [%2, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%2]!  \n" // r1
-
-                        "vmla.f32   q14, q11, d5[0] \n"
-                        "vmla.f32   q15, q11, d6[0] \n"
-
-                        "vmla.f32   q12, q5, d0[0]  \n"
-                        "vmla.f32   q13, q5, d1[0]  \n"
-                        "vmla.f32   q14, q5, d2[0]  \n"
-                        "vmla.f32   q15, q5, d3[0]  \n"
-
-                        "pld        [%2, #192]      \n"
-                        "vld1.f32   {d4-d6}, [%2]   \n"
-
-                        "vmla.f32   q12, q6, d0[1]  \n"
-                        "vmla.f32   q13, q6, d1[1]  \n"
-                        "vmla.f32   q14, q6, d2[1]  \n"
-                        "vmla.f32   q15, q6, d3[1]  \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d18-d23}  \n"
-
-                        "vmla.f32   q12, q7, d1[0]  \n"
-                        "vmla.f32   q13, q7, d2[0]  \n"
-                        "vmla.f32   q14, q7, d3[0]  \n"
-                        "vmla.f32   q15, q7, d4[0]  \n"
-                        "vmla.f32   q12, q8, d1[1]  \n"
-                        "vmla.f32   q13, q8, d2[1]  \n"
-                        "vmla.f32   q14, q8, d3[1]  \n"
-                        "vmla.f32   q15, q8, d4[1]  \n"
-                        "vmla.f32   q12, q9, d2[0]  \n"
-                        "vmla.f32   q13, q9, d3[0]  \n"
-                        "vmla.f32   q14, q9, d4[0]  \n"
-                        "vmla.f32   q15, q9, d5[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d10-d17}  \n"
-
-                        "vmla.f32   q12, q10, d2[1] \n"
-                        "vmla.f32   q13, q10, d3[1] \n"
-                        "vmla.f32   q14, q10, d4[1] \n"
-                        "vmla.f32   q15, q10, d5[1] \n"
-                        "vmla.f32   q12, q11, d3[0] \n"
-                        "vmla.f32   q13, q11, d4[0] \n"
-
-                        "pld        [%3, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%3]!  \n" // r2
-
-                        "vmla.f32   q14, q11, d5[0] \n"
-                        "vmla.f32   q15, q11, d6[0] \n"
-
-                        "vmla.f32   q12, q5, d0[0]  \n"
-                        "vmla.f32   q13, q5, d1[0]  \n"
-                        "vmla.f32   q14, q5, d2[0]  \n"
-                        "vmla.f32   q15, q5, d3[0]  \n"
-
-                        "pld        [%3, #192]      \n"
-                        "vld1.f32   {d4-d6}, [%3]   \n"
-
-                        "vmla.f32   q12, q6, d0[1]  \n"
-                        "vmla.f32   q13, q6, d1[1]  \n"
-                        "vmla.f32   q14, q6, d2[1]  \n"
-                        "vmla.f32   q15, q6, d3[1]  \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d18-d23}  \n"
-
-                        "vmla.f32   q12, q7, d1[0]  \n"
-                        "vmla.f32   q13, q7, d2[0]  \n"
-                        "vmla.f32   q14, q7, d3[0]  \n"
-                        "vmla.f32   q15, q7, d4[0]  \n"
-                        "vmla.f32   q12, q8, d1[1]  \n"
-                        "vmla.f32   q13, q8, d2[1]  \n"
-                        "vmla.f32   q14, q8, d3[1]  \n"
-                        "vmla.f32   q15, q8, d4[1]  \n"
-                        "vmla.f32   q12, q9, d2[0]  \n"
-                        "vmla.f32   q13, q9, d3[0]  \n"
-                        "vmla.f32   q14, q9, d4[0]  \n"
-                        "vmla.f32   q15, q9, d5[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d10-d17}  \n"
-
-                        "vmla.f32   q12, q10, d2[1] \n"
-                        "vmla.f32   q13, q10, d3[1] \n"
-                        "vmla.f32   q14, q10, d4[1] \n"
-                        "vmla.f32   q15, q10, d5[1] \n"
-                        "vmla.f32   q12, q11, d3[0] \n"
-                        "vmla.f32   q13, q11, d4[0] \n"
-
-                        "pld        [%4, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%4]!  \n" // r3
-
-                        "vmla.f32   q14, q11, d5[0] \n"
-                        "vmla.f32   q15, q11, d6[0] \n"
-
-                        "vmla.f32   q12, q5, d0[0]  \n"
-                        "vmla.f32   q13, q5, d1[0]  \n"
-                        "vmla.f32   q14, q5, d2[0]  \n"
-                        "vmla.f32   q15, q5, d3[0]  \n"
-
-                        "pld        [%4, #192]      \n"
-                        "vld1.f32   {d4-d6}, [%4]   \n"
-
-                        "vmla.f32   q12, q6, d0[1]  \n"
-                        "vmla.f32   q13, q6, d1[1]  \n"
-                        "vmla.f32   q14, q6, d2[1]  \n"
-                        "vmla.f32   q15, q6, d3[1]  \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d18-d23}  \n"
-
-                        "vmla.f32   q12, q7, d1[0]  \n"
-                        "vmla.f32   q13, q7, d2[0]  \n"
-                        "vmla.f32   q14, q7, d3[0]  \n"
-                        "vmla.f32   q15, q7, d4[0]  \n"
-                        "vmla.f32   q12, q8, d1[1]  \n"
-                        "vmla.f32   q13, q8, d2[1]  \n"
-                        "vmla.f32   q14, q8, d3[1]  \n"
-                        "vmla.f32   q15, q8, d4[1]  \n"
-                        "vmla.f32   q12, q9, d2[0]  \n"
-                        "vmla.f32   q13, q9, d3[0]  \n"
-                        "vmla.f32   q14, q9, d4[0]  \n"
-                        "vmla.f32   q15, q9, d5[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d10-d17}  \n"
-
-                        "vmla.f32   q12, q10, d2[1] \n"
-                        "vmla.f32   q13, q10, d3[1] \n"
-                        "vmla.f32   q14, q10, d4[1] \n"
-                        "vmla.f32   q15, q10, d5[1] \n"
-                        "vmla.f32   q12, q11, d3[0] \n"
-                        "vmla.f32   q13, q11, d4[0] \n"
-
-                        "pld        [%5, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%5]!  \n" // r4
-
-                        "vmla.f32   q14, q11, d5[0] \n"
-                        "vmla.f32   q15, q11, d6[0] \n"
-
-                        "vmla.f32   q12, q5, d0[0]  \n"
-                        "vmla.f32   q13, q5, d1[0]  \n"
-                        "vmla.f32   q14, q5, d2[0]  \n"
-                        "vmla.f32   q15, q5, d3[0]  \n"
-
-                        "pld        [%5, #192]      \n"
-                        "vld1.f32   {d4-d6}, [%5]   \n"
-
-                        "vmla.f32   q12, q6, d0[1]  \n"
-                        "vmla.f32   q13, q6, d1[1]  \n"
-                        "vmla.f32   q14, q6, d2[1]  \n"
-                        "vmla.f32   q15, q6, d3[1]  \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d18-d23}  \n"
-
-                        "vmla.f32   q12, q7, d1[0]  \n"
-                        "vmla.f32   q13, q7, d2[0]  \n"
-                        "vmla.f32   q14, q7, d3[0]  \n"
-                        "vmla.f32   q15, q7, d4[0]  \n"
-                        "vmla.f32   q12, q8, d1[1]  \n"
-                        "vmla.f32   q13, q8, d2[1]  \n"
-                        "vmla.f32   q14, q8, d3[1]  \n"
-                        "vmla.f32   q15, q8, d4[1]  \n"
-                        "vmla.f32   q12, q9, d2[0]  \n"
-                        "vmla.f32   q13, q9, d3[0]  \n"
-                        "vmla.f32   q14, q9, d4[0]  \n"
-                        "vmla.f32   q15, q9, d5[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d10-d17}  \n"
-
-                        "vmla.f32   q12, q10, d2[1] \n"
-                        "vmla.f32   q13, q10, d3[1] \n"
-                        "vmla.f32   q14, q10, d4[1] \n"
-                        "vmla.f32   q15, q10, d5[1] \n"
-                        "vmla.f32   q12, q11, d3[0] \n"
-                        "vmla.f32   q13, q11, d4[0] \n"
-
-                        "pld        [%6, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%6]!  \n" // r5
-
-                        "vmla.f32   q14, q11, d5[0] \n"
-                        "vmla.f32   q15, q11, d6[0] \n"
-
-                        "vmla.f32   q12, q5, d0[0]  \n"
-                        "vmla.f32   q13, q5, d1[0]  \n"
-                        "vmla.f32   q14, q5, d2[0]  \n"
-                        "vmla.f32   q15, q5, d3[0]  \n"
-
-                        "pld        [%6, #192]      \n"
-                        "vld1.f32   {d4-d6}, [%6]   \n"
-
-                        "vmla.f32   q12, q6, d0[1]  \n"
-                        "vmla.f32   q13, q6, d1[1]  \n"
-                        "vmla.f32   q14, q6, d2[1]  \n"
-                        "vmla.f32   q15, q6, d3[1]  \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d18-d23}  \n"
-
-                        "vmla.f32   q12, q7, d1[0]  \n"
-                        "vmla.f32   q13, q7, d2[0]  \n"
-                        "vmla.f32   q14, q7, d3[0]  \n"
-                        "vmla.f32   q15, q7, d4[0]  \n"
-                        "vmla.f32   q12, q8, d1[1]  \n"
-                        "vmla.f32   q13, q8, d2[1]  \n"
-                        "vmla.f32   q14, q8, d3[1]  \n"
-                        "vmla.f32   q15, q8, d4[1]  \n"
-                        "vmla.f32   q12, q9, d2[0]  \n"
-                        "vmla.f32   q13, q9, d3[0]  \n"
-                        "vmla.f32   q14, q9, d4[0]  \n"
-                        "vmla.f32   q15, q9, d5[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d10-d17}  \n"
-
-                        "vmla.f32   q12, q10, d2[1] \n"
-                        "vmla.f32   q13, q10, d3[1] \n"
-                        "vmla.f32   q14, q10, d4[1] \n"
-                        "vmla.f32   q15, q10, d5[1] \n"
-                        "vmla.f32   q12, q11, d3[0] \n"
-                        "vmla.f32   q13, q11, d4[0] \n"
-
-                        "pld        [%7, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%7]!  \n" // r6
-
-                        "vmla.f32   q14, q11, d5[0] \n"
-                        "vmla.f32   q15, q11, d6[0] \n"
-
-                        "vmla.f32   q12, q5, d0[0]  \n"
-                        "vmla.f32   q13, q5, d1[0]  \n"
-                        "vmla.f32   q14, q5, d2[0]  \n"
-                        "vmla.f32   q15, q5, d3[0]  \n"
-
-                        "pld        [%7, #192]      \n"
-                        "vld1.f32   {d4-d6}, [%7]   \n"
-
-                        "vmla.f32   q12, q6, d0[1]  \n"
-                        "vmla.f32   q13, q6, d1[1]  \n"
-                        "vmla.f32   q14, q6, d2[1]  \n"
-                        "vmla.f32   q15, q6, d3[1]  \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d18-d23}  \n"
-
-                        "vmla.f32   q12, q7, d1[0]  \n"
-                        "vmla.f32   q13, q7, d2[0]  \n"
-                        "vmla.f32   q14, q7, d3[0]  \n"
-                        "vmla.f32   q15, q7, d4[0]  \n"
-                        "vmla.f32   q12, q8, d1[1]  \n"
-                        "vmla.f32   q13, q8, d2[1]  \n"
-                        "vmla.f32   q14, q8, d3[1]  \n"
-                        "vmla.f32   q15, q8, d4[1]  \n"
-                        "vmla.f32   q12, q9, d2[0]  \n"
-                        "vmla.f32   q13, q9, d3[0]  \n"
-                        "vmla.f32   q14, q9, d4[0]  \n"
-                        "vmla.f32   q15, q9, d5[0]  \n"
-                        "vmla.f32   q12, q10, d2[1] \n"
-                        "vmla.f32   q13, q10, d3[1] \n"
-                        "vmla.f32   q14, q10, d4[1] \n"
-                        "vmla.f32   q15, q10, d5[1] \n"
-                        "vmla.f32   q12, q11, d3[0] \n"
-                        "vmla.f32   q13, q11, d4[0] \n"
-                        "vmla.f32   q14, q11, d5[0] \n"
-                        "vmla.f32   q15, q11, d6[0] \n"
-
-                        "sub        %8, %8, #784    \n"
-
-                        "vstm       %0!, {d24-d31}  \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v16.4s, v17.4s}, [%0]      \n"
-
-                        "prfm   pldl1keep, [%1, #384]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s}, [%1] \n" // r0
-                        "add    %1, %1, #16                 \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmul   v18.4s, v24.4s, v0.s[0]     \n"
-                        "fmul   v19.4s, v24.4s, v0.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%2, #384]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s}, [%2] \n" // r1
-                        "add    %2, %2, #16                 \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%3, #384]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s}, [%3] \n" // r2
-                        "add    %3, %3, #16                 \n"
-
-                        "fmla   v18.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v18.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v18.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%4, #384]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s}, [%4] \n" // r3
-                        "add    %4, %4, #16                 \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%5, #384]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s}, [%5] \n" // r4
-                        "add    %5, %5, #16                 \n"
-
-                        "fmla   v18.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v18.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v18.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%6, #384]       \n"
-                        "ld1    {v4.4s, v5.4s, v6.4s}, [%6] \n" // r5
-                        "add    %6, %6, #16                 \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%7, #384]       \n"
-                        "ld1    {v0.4s, v1.4s, v2.4s}, [%7] \n" // r6
-                        "add    %7, %7, #16                 \n"
-
-                        "fmla   v18.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v18.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v18.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v2.s[0]     \n"
-
-                        "fadd   v16.4s, v16.4s, v18.4s      \n"
-                        "fadd   v17.4s, v17.4s, v19.4s      \n"
-
-                        "sub    %8, %8, #784                \n"
-
-                        "st1    {v16.4s, v17.4s}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "v0", "v1", "v2", "v4", "v5", "v6", "v16", "v17", "v18", "v19", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #256]      \n"
-                        "vld1.f32   {d28-d31}, [%0 :128] \n"
-
-                        "pld        [%1, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%1]!  \n" // r0
-                        "vld1.f32   {d8[0]}, [%1]   \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d10-d17}  \n"
-
-                        "vmul.f32   q12, q5, d0[0]  \n"
-                        "vmul.f32   q13, q5, d1[0]  \n"
-                        "vmla.f32   q14, q6, d0[1]  \n"
-                        "vmla.f32   q15, q6, d1[1]  \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d18-d23}  \n"
-
-                        "vmla.f32   q12, q7, d1[0]  \n"
-                        "vmla.f32   q13, q7, d2[0]  \n"
-
-                        "pld        [%2, #256]      \n"
-                        "vld1.f32   {d4-d7}, [%2]!  \n" // r1
-                        "vld1.f32   {d9[0]}, [%2]   \n"
-
-                        "vmla.f32   q14, q8, d1[1]  \n"
-                        "vmla.f32   q15, q8, d2[1]  \n"
-                        "vmla.f32   q12, q9, d2[0]  \n"
-                        "vmla.f32   q13, q9, d3[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d10-d17}  \n"
-
-                        "vmla.f32   q14, q10, d2[1] \n"
-                        "vmla.f32   q15, q10, d3[1] \n"
-                        "vmla.f32   q12, q11, d3[0] \n"
-                        "vmla.f32   q13, q11, d8[0] \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d18-d23}  \n"
-
-                        "vmla.f32   q14, q5, d4[0]  \n"
-                        "vmla.f32   q15, q5, d5[0]  \n"
-                        "vmla.f32   q12, q6, d4[1]  \n"
-                        "vmla.f32   q13, q6, d5[1]  \n"
-                        "vmla.f32   q14, q7, d5[0]  \n"
-                        "vmla.f32   q15, q7, d6[0]  \n"
-
-                        "pld        [%3, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%3]!  \n" // r2
-                        "vld1.f32   {d8[0]}, [%3]   \n"
-
-                        "vmla.f32   q12, q8, d5[1]  \n"
-                        "vmla.f32   q13, q8, d6[1]  \n"
-                        "vmla.f32   q14, q9, d6[0]  \n"
-                        "vmla.f32   q15, q9, d7[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d10-d17}  \n"
-
-                        "vmla.f32   q12, q10, d6[1] \n"
-                        "vmla.f32   q13, q10, d7[1] \n"
-                        "vmla.f32   q14, q11, d7[0] \n"
-                        "vmla.f32   q15, q11, d9[0] \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d18-d23}  \n"
-
-                        "vmla.f32   q12, q5, d0[0]  \n"
-                        "vmla.f32   q13, q5, d1[0]  \n"
-                        "vmla.f32   q14, q6, d0[1]  \n"
-                        "vmla.f32   q15, q6, d1[1]  \n"
-                        "vmla.f32   q12, q7, d1[0]  \n"
-                        "vmla.f32   q13, q7, d2[0]  \n"
-
-                        "pld        [%4, #256]      \n"
-                        "vld1.f32   {d4-d7}, [%4]!  \n" // r3
-                        "vld1.f32   {d9[0]}, [%4]   \n"
-
-                        "vmla.f32   q14, q8, d1[1]  \n"
-                        "vmla.f32   q15, q8, d2[1]  \n"
-                        "vmla.f32   q12, q9, d2[0]  \n"
-                        "vmla.f32   q13, q9, d3[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d10-d17}  \n"
-
-                        "vmla.f32   q14, q10, d2[1] \n"
-                        "vmla.f32   q15, q10, d3[1] \n"
-                        "vmla.f32   q12, q11, d3[0] \n"
-                        "vmla.f32   q13, q11, d8[0] \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d18-d23}  \n"
-
-                        "vmla.f32   q14, q5, d4[0]  \n"
-                        "vmla.f32   q15, q5, d5[0]  \n"
-                        "vmla.f32   q12, q6, d4[1]  \n"
-                        "vmla.f32   q13, q6, d5[1]  \n"
-                        "vmla.f32   q14, q7, d5[0]  \n"
-                        "vmla.f32   q15, q7, d6[0]  \n"
-
-                        "pld        [%5, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%5]!  \n" // r4
-                        "vld1.f32   {d8[0]}, [%5]   \n"
-
-                        "vmla.f32   q12, q8, d5[1]  \n"
-                        "vmla.f32   q13, q8, d6[1]  \n"
-                        "vmla.f32   q14, q9, d6[0]  \n"
-                        "vmla.f32   q15, q9, d7[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d10-d17}  \n"
-
-                        "vmla.f32   q12, q10, d6[1] \n"
-                        "vmla.f32   q13, q10, d7[1] \n"
-                        "vmla.f32   q14, q11, d7[0] \n"
-                        "vmla.f32   q15, q11, d9[0] \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d18-d23}  \n"
-
-                        "vmla.f32   q12, q5, d0[0]  \n"
-                        "vmla.f32   q13, q5, d1[0]  \n"
-                        "vmla.f32   q14, q6, d0[1]  \n"
-                        "vmla.f32   q15, q6, d1[1]  \n"
-                        "vmla.f32   q12, q7, d1[0]  \n"
-                        "vmla.f32   q13, q7, d2[0]  \n"
-
-                        "pld        [%6, #256]      \n"
-                        "vld1.f32   {d4-d7}, [%6]!  \n" // r5
-                        "vld1.f32   {d9[0]}, [%6]   \n"
-
-                        "vmla.f32   q14, q8, d1[1]  \n"
-                        "vmla.f32   q15, q8, d2[1]  \n"
-                        "vmla.f32   q12, q9, d2[0]  \n"
-                        "vmla.f32   q13, q9, d3[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d10-d17}  \n"
-
-                        "vmla.f32   q14, q10, d2[1] \n"
-                        "vmla.f32   q15, q10, d3[1] \n"
-                        "vmla.f32   q12, q11, d3[0] \n"
-                        "vmla.f32   q13, q11, d8[0] \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d18-d23}  \n"
-
-                        "vmla.f32   q14, q5, d4[0]  \n"
-                        "vmla.f32   q15, q5, d5[0]  \n"
-                        "vmla.f32   q12, q6, d4[1]  \n"
-                        "vmla.f32   q13, q6, d5[1]  \n"
-                        "vmla.f32   q14, q7, d5[0]  \n"
-                        "vmla.f32   q15, q7, d6[0]  \n"
-
-                        "pld        [%7, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%7]!  \n" // r6
-                        "vld1.f32   {d8[0]}, [%7]   \n"
-
-                        "vmla.f32   q12, q8, d5[1]  \n"
-                        "vmla.f32   q13, q8, d6[1]  \n"
-                        "vmla.f32   q14, q9, d6[0]  \n"
-                        "vmla.f32   q15, q9, d7[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d10-d17}  \n"
-
-                        "vmla.f32   q12, q10, d6[1] \n"
-                        "vmla.f32   q13, q10, d7[1] \n"
-                        "vmla.f32   q14, q11, d7[0] \n"
-                        "vmla.f32   q15, q11, d9[0] \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d18-d23}  \n"
-
-                        "vmla.f32   q12, q5, d0[0]  \n"
-                        "vmla.f32   q13, q5, d1[0]  \n"
-                        "vmla.f32   q14, q6, d0[1]  \n"
-                        "vmla.f32   q15, q6, d1[1]  \n"
-
-                        "sub        %1, %1, #16     \n"
-                        "sub        %2, %2, #16     \n"
-
-                        "vmla.f32   q12, q7, d1[0]  \n"
-                        "vmla.f32   q13, q7, d2[0]  \n"
-                        "vmla.f32   q14, q8, d1[1]  \n"
-                        "vmla.f32   q15, q8, d2[1]  \n"
-
-                        "sub        %8, %8, #784    \n"
-
-                        "vmla.f32   q12, q9, d2[0]  \n"
-                        "vmla.f32   q13, q9, d3[0]  \n"
-                        "vmla.f32   q14, q10, d2[1] \n"
-                        "vmla.f32   q15, q10, d3[1] \n"
-
-                        "sub        %3, %3, #16     \n"
-                        "sub        %4, %4, #16     \n"
-
-                        "vmla.f32   q12, q11, d3[0] \n"
-                        "vmla.f32   q13, q11, d8[0] \n"
-
-                        "sub        %5, %5, #16     \n"
-                        "sub        %6, %6, #16     \n"
-
-                        "vadd.f32   q14, q14, q12   \n"
-                        "vadd.f32   q15, q15, q13   \n"
-
-                        "sub        %7, %7, #16     \n"
-
-                        "vst1.f32   {d28-d31}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v16.4s}, [%0]              \n"
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%1]        \n" // r0
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmul   v17.4s, v24.4s, v0.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmul   v18.4s, v25.4s, v0.s[1]     \n"
-                        "fmul   v19.4s, v26.4s, v0.s[2]     \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v4.4s, v5.4s}, [%2]        \n" // r1
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v18.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v30.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v18.4s, v26.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%3]        \n" // r2
-
-                        "fmla   v19.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v17.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v18.4s, v30.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v19.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v26.4s, v0.s[2]     \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v4.4s, v5.4s}, [%4]        \n" // r3
-
-                        "fmla   v18.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v30.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v18.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v19.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%5]        \n" // r4
-
-                        "fmla   v17.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v18.4s, v28.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v19.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v17.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v18.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v19.4s, v26.4s, v0.s[2]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v4.4s, v5.4s}, [%6]        \n" // r5
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v18.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v30.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v18.4s, v26.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v0.4s, v1.4s}, [%7]        \n" // r6
-
-                        "fmla   v19.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%8], #64 \n"
-
-                        "fmla   v17.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v18.4s, v30.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #384]       \n"
-                        "ld1    {v28.4s, v29.4s, v30.4s}, [%8], #48 \n"
-
-                        "fmla   v19.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v26.4s, v0.s[2]     \n"
-
-                        "add    %1, %1, #8                  \n"
-                        "add    %2, %2, #8                  \n"
-
-                        "fmla   v18.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v30.4s, v1.s[2]     \n"
-
-                        "add    %3, %3, #8                  \n"
-                        "add    %4, %4, #8                  \n"
-
-                        "fadd   v18.4s, v18.4s, v19.4s      \n"
-
-                        "add    %5, %5, #8                  \n"
-
-                        "fadd   v16.4s, v16.4s, v17.4s      \n"
-
-                        "add    %6, %6, #8                  \n"
-                        "add    %7, %7, #8                  \n"
-
-                        "fadd   v16.4s, v16.4s, v18.4s      \n"
-
-                        "sub    %8, %8, #784                \n"
-
-                        "st1    {v16.4s}, [%0], #16         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "v0", "v1", "v4", "v5", "v16", "v17", "v18", "v19", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #128]      \n"
-                        "vld1.f32   {d8-d9}, [%0 :128] \n"
-
-                        "pld        [%1, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%1]   \n" // r0
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d16-d23}  \n"
-
-                        "vmul.f32   q5, q8, d0[0]   \n"
-                        "vmul.f32   q6, q9, d0[1]   \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d24-d29}  \n"
-
-                        "vmul.f32   q7, q10, d1[0]  \n"
-                        "vmla.f32   q4, q11, d1[1]  \n"
-
-                        "pld        [%2, #256]      \n"
-                        "vld1.f32   {d4-d7}, [%2]   \n" // r1
-
-                        "vmla.f32   q5, q12, d2[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d16-d23}  \n"
-
-                        "vmla.f32   q6, q13, d2[1]  \n"
-                        "vmla.f32   q7, q14, d3[0]  \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d24-d29}  \n"
-
-                        "vmla.f32   q4, q8, d4[0]   \n"
-                        "vmla.f32   q5, q9, d4[1]   \n"
-                        "vmla.f32   q6, q10, d5[0]  \n"
-
-                        "pld        [%3, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%3]   \n" // r2
-
-                        "vmla.f32   q7, q11, d5[1]  \n"
-                        "vmla.f32   q4, q12, d6[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d16-d23}  \n"
-
-                        "vmla.f32   q5, q13, d6[1]  \n"
-                        "vmla.f32   q6, q14, d7[0]  \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d24-d29}  \n"
-
-                        "vmla.f32   q7, q8, d0[0]   \n"
-                        "vmla.f32   q4, q9, d0[1]   \n"
-                        "vmla.f32   q5, q10, d1[0]  \n"
-
-                        "pld        [%4, #256]      \n"
-                        "vld1.f32   {d4-d7}, [%4]   \n" // r3
-
-                        "vmla.f32   q6, q11, d1[1]  \n"
-                        "vmla.f32   q7, q12, d2[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d16-d23}  \n"
-
-                        "vmla.f32   q4, q13, d2[1]  \n"
-                        "vmla.f32   q5, q14, d3[0]  \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d24-d29}  \n"
-
-                        "vmla.f32   q6, q8, d4[0]   \n"
-                        "vmla.f32   q7, q9, d4[1]   \n"
-                        "vmla.f32   q4, q10, d5[0]  \n"
-
-                        "pld        [%5, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%5]   \n" // r4
-
-                        "vmla.f32   q5, q11, d5[1]  \n"
-                        "vmla.f32   q6, q12, d6[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d16-d23}  \n"
-
-                        "vmla.f32   q7, q13, d6[1]  \n"
-                        "vmla.f32   q4, q14, d7[0]  \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d24-d29}  \n"
-
-                        "vmla.f32   q5, q8, d0[0]   \n"
-                        "vmla.f32   q6, q9, d0[1]   \n"
-                        "vmla.f32   q7, q10, d1[0]  \n"
-
-                        "pld        [%6, #256]      \n"
-                        "vld1.f32   {d4-d7}, [%6]   \n" // r5
-
-                        "vmla.f32   q4, q11, d1[1]  \n"
-                        "vmla.f32   q5, q12, d2[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d16-d23}  \n"
-
-                        "vmla.f32   q6, q13, d2[1]  \n"
-                        "vmla.f32   q7, q14, d3[0]  \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d24-d29}  \n"
-
-                        "vmla.f32   q4, q8, d4[0]   \n"
-                        "vmla.f32   q5, q9, d4[1]   \n"
-                        "vmla.f32   q6, q10, d5[0]  \n"
-
-                        "pld        [%7, #256]      \n"
-                        "vld1.f32   {d0-d3}, [%7]   \n" // r6
-
-                        "vmla.f32   q7, q11, d5[1]  \n"
-                        "vmla.f32   q4, q12, d6[0]  \n"
-
-                        "pld        [%8, #512]      \n"
-                        "vldm       %8!, {d16-d23}  \n"
-
-                        "vmla.f32   q5, q13, d6[1]  \n"
-                        "vmla.f32   q6, q14, d7[0]  \n"
-
-                        "pld        [%8, #384]      \n"
-                        "vldm       %8!, {d24-d29}  \n"
-
-                        "vmla.f32   q7, q8, d0[0]   \n"
-                        "vmla.f32   q4, q9, d0[1]   \n"
-
-                        "add        %1, %1, #8      \n"
-                        "add        %2, %2, #8      \n"
-
-                        "vmla.f32   q5, q10, d1[0]  \n"
-                        "vmla.f32   q6, q11, d1[1]  \n"
-
-                        "sub        %8, %8, #784    \n"
-
-                        "vmla.f32   q7, q12, d2[0]  \n"
-                        "vmla.f32   q4, q13, d2[1]  \n"
-                        "vmla.f32   q5, q14, d3[0]  \n"
-
-                        "add        %3, %3, #8      \n"
-                        "add        %4, %4, #8      \n"
-
-                        "vadd.f32   q6, q6, q7      \n"
-
-                        "add        %5, %5, #8      \n"
-
-                        "vadd.f32   q4, q4, q5      \n"
-
-                        "add        %6, %6, #8      \n"
-
-                        "vadd.f32   q4, q4, q6      \n"
-
-                        "add        %7, %7, #8      \n"
-
-                        "vst1.f32   {d8-d9}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14");
-#endif // __aarch64__
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-                r4 += tailstep;
-                r5 += tailstep;
-                r6 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7_pack1to4_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7_pack1to4_bf16s.h
deleted file mode 100644
index 3ffd6b41..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7_pack1to4_bf16s.h
+++ /dev/null
@@ -1,5575 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv7x7s2_pack1to4_bf16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    Mat top_blob_fp32(outw, outh, opt.num_threads, (size_t)4u * 4, 4, opt.workspace_allocator);
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob_fp32.channel(get_omp_thread_num());
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + p * 4) : vdupq_n_f32(0.f);
-        out0.fill(_bias0);
-
-        int q = 0;
-        for (; q < inch - 1; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<const unsigned short>(0);
-            const unsigned short* r1 = img0.row<const unsigned short>(1);
-            const unsigned short* r2 = img0.row<const unsigned short>(2);
-            const unsigned short* r3 = img0.row<const unsigned short>(3);
-            const unsigned short* r4 = img0.row<const unsigned short>(4);
-            const unsigned short* r5 = img0.row<const unsigned short>(5);
-            const unsigned short* r6 = img0.row<const unsigned short>(6);
-
-            const unsigned short* kptr = kernel.channel(p).row<const unsigned short>(q);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-#if __aarch64__
-                for (; j + 7 < outw; j += 8)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%1], #32 \n" // r0
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0] \n"
-
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[2]     \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%1]        \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[0]     \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[1]     \n"
-
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v28.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v29.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v6.4h, v7.4h, v8.4h, v9.4h}, [%2], #32 \n" // r1
-
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-                        "shll   v8.4s, v8.4h, #16           \n"
-                        "shll   v9.4s, v9.4h, #16           \n"
-
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v30.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v30.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v10.4h, v11.4h}, [%2]      \n"
-
-                        "shll   v10.4s, v10.4h, #16         \n"
-                        "shll   v11.4s, v11.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v6.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v8.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v8.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v9.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v9.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v6.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v7.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v8.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v8.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v9.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v9.s[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v7.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v8.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v8.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v9.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v9.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v10.s[0]    \n"
-
-                        "fmla   v16.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v7.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v8.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v8.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v9.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v9.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v10.s[1]    \n"
-
-                        "fmla   v16.4s, v28.4s, v7.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v7.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v8.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v8.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v9.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v9.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v10.s[0]    \n"
-                        "fmla   v23.4s, v28.4s, v10.s[2]    \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v7.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v7.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v8.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v8.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v9.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v9.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v10.s[1]    \n"
-                        "fmla   v23.4s, v29.4s, v10.s[3]    \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%3], #32 \n" // r2
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v16.4s, v30.4s, v7.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v8.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v8.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v9.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v9.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v10.s[0]    \n"
-                        "fmla   v22.4s, v30.4s, v10.s[2]    \n"
-                        "fmla   v23.4s, v30.4s, v11.s[0]    \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%3]        \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[0]     \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[1]     \n"
-
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v28.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v29.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v6.4h, v7.4h, v8.4h, v9.4h}, [%4], #32 \n" // r3
-
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-                        "shll   v8.4s, v8.4h, #16           \n"
-                        "shll   v9.4s, v9.4h, #16           \n"
-
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v30.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v30.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v10.4h, v11.4h}, [%4]      \n"
-
-                        "shll   v10.4s, v10.4h, #16         \n"
-                        "shll   v11.4s, v11.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v6.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v8.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v8.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v9.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v9.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v6.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v7.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v8.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v8.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v9.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v9.s[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v7.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v8.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v8.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v9.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v9.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v10.s[0]    \n"
-
-                        "fmla   v16.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v7.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v8.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v8.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v9.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v9.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v10.s[1]    \n"
-
-                        "fmla   v16.4s, v28.4s, v7.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v7.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v8.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v8.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v9.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v9.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v10.s[0]    \n"
-                        "fmla   v23.4s, v28.4s, v10.s[2]    \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v7.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v7.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v8.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v8.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v9.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v9.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v10.s[1]    \n"
-                        "fmla   v23.4s, v29.4s, v10.s[3]    \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%5], #32 \n" // r4
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v16.4s, v30.4s, v7.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v8.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v8.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v9.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v9.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v10.s[0]    \n"
-                        "fmla   v22.4s, v30.4s, v10.s[2]    \n"
-                        "fmla   v23.4s, v30.4s, v11.s[0]    \n"
-
-                        "prfm   pldl1keep, [%5, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%5]        \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[0]     \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[1]     \n"
-
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v28.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v29.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v6.4h, v7.4h, v8.4h, v9.4h}, [%6], #32 \n" // r5
-
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-                        "shll   v8.4s, v8.4h, #16           \n"
-                        "shll   v9.4s, v9.4h, #16           \n"
-
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v30.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v30.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%6, #128]       \n"
-                        "ld1    {v10.4h, v11.4h}, [%6]      \n"
-
-                        "shll   v10.4s, v10.4h, #16         \n"
-                        "shll   v11.4s, v11.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v6.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v8.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v8.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v9.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v9.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v6.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v7.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v8.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v8.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v9.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v9.s[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v7.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v8.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v8.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v9.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v9.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v10.s[0]    \n"
-
-                        "fmla   v16.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v7.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v8.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v8.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v9.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v9.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v10.s[1]    \n"
-
-                        "fmla   v16.4s, v28.4s, v7.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v7.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v8.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v8.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v9.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v9.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v10.s[0]    \n"
-                        "fmla   v23.4s, v28.4s, v10.s[2]    \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v7.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v7.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v8.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v8.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v9.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v9.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v10.s[1]    \n"
-                        "fmla   v23.4s, v29.4s, v10.s[3]    \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%7], #32 \n" // r6
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v16.4s, v30.4s, v7.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v8.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v8.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v9.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v9.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v10.s[0]    \n"
-                        "fmla   v22.4s, v30.4s, v10.s[2]    \n"
-                        "fmla   v23.4s, v30.4s, v11.s[0]    \n"
-
-                        "prfm   pldl1keep, [%7, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%7]        \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[0]     \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[1]     \n"
-
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v28.4s, v4.s[2]     \n"
-
-                        "sub    %0, %0, #64                 \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v29.4s, v4.s[3]     \n"
-
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v30.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v30.4s, v5.s[0]     \n"
-
-                        "sub    %8, %8, #392                \n"
-
-                        "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                        "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v16", "v17", "v18", "v19", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
-                }
-#endif // __aarch64__
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0] \n"
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%1] \n" // r0
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%2] \n" // r1
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v5.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%3] \n" // r2
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v6.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v6.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v6.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v6.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v6.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v6.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v7.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%4] \n" // r3
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v5.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%5] \n" // r4
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v6.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v6.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v6.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v6.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v6.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v6.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v7.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%6] \n" // r5
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v5.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%7] \n" // r6
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v6.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v6.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v6.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v6.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v6.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v6.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v7.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-
-                        "add    %1, %1, #16                 \n"
-                        "add    %2, %2, #16                 \n"
-                        "add    %3, %3, #16                 \n"
-                        "add    %4, %4, #16                 \n"
-                        "add    %5, %5, #16                 \n"
-                        "add    %6, %6, #16                 \n"
-                        "add    %7, %7, #16                 \n"
-
-                        "sub    %8, %8, #392                \n"
-
-                        "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #512]          \n"
-                        "vldm       %0, {d24-d31}       \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%1]!      \n" // r0
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%8]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q5, d2[0]      \n"
-                        "vmla.f32   q15, q5, d3[0]      \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.u16   {d5-d6}, [%1]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshl.u32   d6, d6, #16         \n"
-
-                        "vmla.f32   q12, q6, d0[1]      \n"
-                        "vmla.f32   q13, q6, d1[1]      \n"
-                        "vmla.f32   q14, q6, d2[1]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%8]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q7, d3[0]      \n"
-                        "vmla.f32   q15, q7, d4[0]      \n"
-                        "vmla.f32   q12, q8, d1[1]      \n"
-                        "vmla.f32   q13, q8, d2[1]      \n"
-                        "vmla.f32   q14, q8, d3[1]      \n"
-                        "vmla.f32   q15, q8, d4[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q9, d4[0]      \n"
-                        "vmla.f32   q15, q9, d5[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%8]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[1]     \n"
-                        "vmla.f32   q13, q10, d3[1]     \n"
-                        "vmla.f32   q14, q10, d4[1]     \n"
-                        "vmla.f32   q15, q10, d5[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d4[0]     \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%2]!      \n" // r1
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q14, q11, d5[0]     \n"
-                        "vmla.f32   q15, q11, d6[0]     \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q5, d2[0]      \n"
-                        "vmla.f32   q15, q5, d3[0]      \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d5-d6}, [%2]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshl.u32   d6, d6, #16         \n"
-
-                        "vmla.f32   q12, q6, d0[1]      \n"
-                        "vmla.f32   q13, q6, d1[1]      \n"
-                        "vmla.f32   q14, q6, d2[1]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%8]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q7, d3[0]      \n"
-                        "vmla.f32   q15, q7, d4[0]      \n"
-                        "vmla.f32   q12, q8, d1[1]      \n"
-                        "vmla.f32   q13, q8, d2[1]      \n"
-                        "vmla.f32   q14, q8, d3[1]      \n"
-                        "vmla.f32   q15, q8, d4[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q9, d4[0]      \n"
-                        "vmla.f32   q15, q9, d5[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%8]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[1]     \n"
-                        "vmla.f32   q13, q10, d3[1]     \n"
-                        "vmla.f32   q14, q10, d4[1]     \n"
-                        "vmla.f32   q15, q10, d5[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d4[0]     \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%3]!      \n" // r2
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q14, q11, d5[0]     \n"
-                        "vmla.f32   q15, q11, d6[0]     \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q5, d2[0]      \n"
-                        "vmla.f32   q15, q5, d3[0]      \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d5-d6}, [%3]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshl.u32   d6, d6, #16         \n"
-
-                        "vmla.f32   q12, q6, d0[1]      \n"
-                        "vmla.f32   q13, q6, d1[1]      \n"
-                        "vmla.f32   q14, q6, d2[1]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%8]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q7, d3[0]      \n"
-                        "vmla.f32   q15, q7, d4[0]      \n"
-                        "vmla.f32   q12, q8, d1[1]      \n"
-                        "vmla.f32   q13, q8, d2[1]      \n"
-                        "vmla.f32   q14, q8, d3[1]      \n"
-                        "vmla.f32   q15, q8, d4[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q9, d4[0]      \n"
-                        "vmla.f32   q15, q9, d5[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%8]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[1]     \n"
-                        "vmla.f32   q13, q10, d3[1]     \n"
-                        "vmla.f32   q14, q10, d4[1]     \n"
-                        "vmla.f32   q15, q10, d5[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d4[0]     \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%4]!      \n" // r3
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q14, q11, d5[0]     \n"
-                        "vmla.f32   q15, q11, d6[0]     \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q5, d2[0]      \n"
-                        "vmla.f32   q15, q5, d3[0]      \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.u16   {d5-d6}, [%4]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshl.u32   d6, d6, #16         \n"
-
-                        "vmla.f32   q12, q6, d0[1]      \n"
-                        "vmla.f32   q13, q6, d1[1]      \n"
-                        "vmla.f32   q14, q6, d2[1]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%8]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q7, d3[0]      \n"
-                        "vmla.f32   q15, q7, d4[0]      \n"
-                        "vmla.f32   q12, q8, d1[1]      \n"
-                        "vmla.f32   q13, q8, d2[1]      \n"
-                        "vmla.f32   q14, q8, d3[1]      \n"
-                        "vmla.f32   q15, q8, d4[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q9, d4[0]      \n"
-                        "vmla.f32   q15, q9, d5[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%8]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[1]     \n"
-                        "vmla.f32   q13, q10, d3[1]     \n"
-                        "vmla.f32   q14, q10, d4[1]     \n"
-                        "vmla.f32   q15, q10, d5[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d4[0]     \n"
-
-                        "pld        [%5, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%5]!      \n" // r4
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q14, q11, d5[0]     \n"
-                        "vmla.f32   q15, q11, d6[0]     \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q5, d2[0]      \n"
-                        "vmla.f32   q15, q5, d3[0]      \n"
-
-                        "pld        [%5, #128]          \n"
-                        "vld1.u16   {d5-d6}, [%5]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshl.u32   d6, d6, #16         \n"
-
-                        "vmla.f32   q12, q6, d0[1]      \n"
-                        "vmla.f32   q13, q6, d1[1]      \n"
-                        "vmla.f32   q14, q6, d2[1]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%8]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q7, d3[0]      \n"
-                        "vmla.f32   q15, q7, d4[0]      \n"
-                        "vmla.f32   q12, q8, d1[1]      \n"
-                        "vmla.f32   q13, q8, d2[1]      \n"
-                        "vmla.f32   q14, q8, d3[1]      \n"
-                        "vmla.f32   q15, q8, d4[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q9, d4[0]      \n"
-                        "vmla.f32   q15, q9, d5[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%8]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[1]     \n"
-                        "vmla.f32   q13, q10, d3[1]     \n"
-                        "vmla.f32   q14, q10, d4[1]     \n"
-                        "vmla.f32   q15, q10, d5[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d4[0]     \n"
-
-                        "pld        [%6, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%6]!      \n" // r5
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q14, q11, d5[0]     \n"
-                        "vmla.f32   q15, q11, d6[0]     \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q5, d2[0]      \n"
-                        "vmla.f32   q15, q5, d3[0]      \n"
-
-                        "pld        [%6, #128]          \n"
-                        "vld1.u16   {d5-d6}, [%6]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshl.u32   d6, d6, #16         \n"
-
-                        "vmla.f32   q12, q6, d0[1]      \n"
-                        "vmla.f32   q13, q6, d1[1]      \n"
-                        "vmla.f32   q14, q6, d2[1]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%8]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q7, d3[0]      \n"
-                        "vmla.f32   q15, q7, d4[0]      \n"
-                        "vmla.f32   q12, q8, d1[1]      \n"
-                        "vmla.f32   q13, q8, d2[1]      \n"
-                        "vmla.f32   q14, q8, d3[1]      \n"
-                        "vmla.f32   q15, q8, d4[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q9, d4[0]      \n"
-                        "vmla.f32   q15, q9, d5[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%8]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[1]     \n"
-                        "vmla.f32   q13, q10, d3[1]     \n"
-                        "vmla.f32   q14, q10, d4[1]     \n"
-                        "vmla.f32   q15, q10, d5[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d4[0]     \n"
-
-                        "pld        [%7, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%7]!      \n" // r6
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q14, q11, d5[0]     \n"
-                        "vmla.f32   q15, q11, d6[0]     \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q5, d2[0]      \n"
-                        "vmla.f32   q15, q5, d3[0]      \n"
-
-                        "pld        [%7, #128]          \n"
-                        "vld1.u16   {d5-d6}, [%7]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshl.u32   d6, d6, #16         \n"
-
-                        "vmla.f32   q12, q6, d0[1]      \n"
-                        "vmla.f32   q13, q6, d1[1]      \n"
-                        "vmla.f32   q14, q6, d2[1]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%8]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q7, d3[0]      \n"
-                        "vmla.f32   q15, q7, d4[0]      \n"
-                        "vmla.f32   q12, q8, d1[1]      \n"
-                        "vmla.f32   q13, q8, d2[1]      \n"
-                        "vmla.f32   q14, q8, d3[1]      \n"
-                        "vmla.f32   q15, q8, d4[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q9, d4[0]      \n"
-                        "vmla.f32   q15, q9, d5[0]      \n"
-                        "vmla.f32   q12, q10, d2[1]     \n"
-                        "vmla.f32   q13, q10, d3[1]     \n"
-                        "vmla.f32   q14, q10, d4[1]     \n"
-                        "vmla.f32   q15, q10, d5[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d4[0]     \n"
-                        "vmla.f32   q14, q11, d5[0]     \n"
-                        "vmla.f32   q15, q11, d6[0]     \n"
-
-                        "sub        %8, %8, #392        \n"
-
-                        "vstm       %0!, {d24-d31}      \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #256]       \n"
-                        "ld1    {v16.4s, v17.4s}, [%0]      \n"
-
-                        "prfm   pldl1keep, [%1, #192]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%1] \n" // r0
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmul   v18.4s, v24.4s, v0.s[0]     \n"
-                        "fmul   v19.4s, v24.4s, v0.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%2, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%2] \n" // r1
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%3, #192]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%3] \n" // r2
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v18.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v18.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v18.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%4, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%4] \n" // r3
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%5, #192]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%5] \n" // r4
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v18.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v18.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v18.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%6, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%6] \n" // r5
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%7, #192]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%7] \n" // r6
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v18.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v18.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v18.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v2.s[0]     \n"
-
-                        "add    %1, %1, #8                  \n"
-                        "add    %2, %2, #8                  \n"
-                        "add    %3, %3, #8                  \n"
-                        "add    %4, %4, #8                  \n"
-                        "add    %5, %5, #8                  \n"
-                        "add    %6, %6, #8                  \n"
-                        "add    %7, %7, #8                  \n"
-
-                        "fadd   v16.4s, v16.4s, v18.4s      \n"
-                        "fadd   v17.4s, v17.4s, v19.4s      \n"
-
-                        "sub    %8, %8, #392                \n"
-
-                        "st1    {v16.4s, v17.4s}, [%0], #32 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "v0", "v1", "v2", "v4", "v5", "v6", "v16", "v17", "v18", "v19", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #256]          \n"
-                        "vld1.f32   {d28-d31}, [%0 :128] \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%1]!      \n" // r0
-                        "vld1.u16   {d8[0]}, [%1]       \n"
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vshl.u32   d8, d8, #16         \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%8]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmul.f32   q12, q5, d0[0]      \n"
-                        "vmul.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q6, d0[1]      \n"
-                        "vmla.f32   q15, q6, d1[1]      \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%8]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d6-d7}, [%2]!      \n" // r1
-                        "vld1.u16   {d9[0]}, [%2]       \n"
-
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshl.u32   d9, d9, #16         \n"
-
-                        "vmla.f32   q14, q8, d1[1]      \n"
-                        "vmla.f32   q15, q8, d2[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%8]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q14, q10, d2[1]     \n"
-                        "vmla.f32   q15, q10, d3[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d8[0]     \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%8]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q14, q5, d4[0]      \n"
-                        "vmla.f32   q15, q5, d5[0]      \n"
-                        "vmla.f32   q12, q6, d4[1]      \n"
-                        "vmla.f32   q13, q6, d5[1]      \n"
-                        "vmla.f32   q14, q7, d5[0]      \n"
-                        "vmla.f32   q15, q7, d6[0]      \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%3]!      \n" // r2
-                        "vld1.u16   {d8[0]}, [%3]       \n"
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vshl.u32   d8, d8, #16         \n"
-
-                        "vmla.f32   q12, q8, d5[1]      \n"
-                        "vmla.f32   q13, q8, d6[1]      \n"
-                        "vmla.f32   q14, q9, d6[0]      \n"
-                        "vmla.f32   q15, q9, d7[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%8]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[1]     \n"
-                        "vmla.f32   q13, q10, d7[1]     \n"
-                        "vmla.f32   q14, q11, d7[0]     \n"
-                        "vmla.f32   q15, q11, d9[0]     \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%8]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q6, d0[1]      \n"
-                        "vmla.f32   q15, q6, d1[1]      \n"
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.u16   {d6-d7}, [%4]!      \n" // r3
-                        "vld1.u16   {d9[0]}, [%4]       \n"
-
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshl.u32   d9, d9, #16         \n"
-
-                        "vmla.f32   q14, q8, d1[1]      \n"
-                        "vmla.f32   q15, q8, d2[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%8]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q14, q10, d2[1]     \n"
-                        "vmla.f32   q15, q10, d3[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d8[0]     \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%8]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q14, q5, d4[0]      \n"
-                        "vmla.f32   q15, q5, d5[0]      \n"
-                        "vmla.f32   q12, q6, d4[1]      \n"
-                        "vmla.f32   q13, q6, d5[1]      \n"
-                        "vmla.f32   q14, q7, d5[0]      \n"
-                        "vmla.f32   q15, q7, d6[0]      \n"
-
-                        "pld        [%5, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%5]!      \n" // r4
-                        "vld1.u16   {d8[0]}, [%5]       \n"
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vshl.u32   d8, d8, #16         \n"
-
-                        "vmla.f32   q12, q8, d5[1]      \n"
-                        "vmla.f32   q13, q8, d6[1]      \n"
-                        "vmla.f32   q14, q9, d6[0]      \n"
-                        "vmla.f32   q15, q9, d7[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%8]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[1]     \n"
-                        "vmla.f32   q13, q10, d7[1]     \n"
-                        "vmla.f32   q14, q11, d7[0]     \n"
-                        "vmla.f32   q15, q11, d9[0]     \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%8]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q6, d0[1]      \n"
-                        "vmla.f32   q15, q6, d1[1]      \n"
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-
-                        "pld        [%6, #128]          \n"
-                        "vld1.u16   {d6-d7}, [%6]!      \n" // r5
-                        "vld1.u16   {d9[0]}, [%6]       \n"
-
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshl.u32   d9, d9, #16         \n"
-
-                        "vmla.f32   q14, q8, d1[1]      \n"
-                        "vmla.f32   q15, q8, d2[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%8]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q14, q10, d2[1]     \n"
-                        "vmla.f32   q15, q10, d3[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d8[0]     \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%8]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q14, q5, d4[0]      \n"
-                        "vmla.f32   q15, q5, d5[0]      \n"
-                        "vmla.f32   q12, q6, d4[1]      \n"
-                        "vmla.f32   q13, q6, d5[1]      \n"
-                        "vmla.f32   q14, q7, d5[0]      \n"
-                        "vmla.f32   q15, q7, d6[0]      \n"
-
-                        "pld        [%7, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%7]!      \n" // r6
-                        "vld1.u16   {d8[0]}, [%7]       \n"
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vshl.u32   d8, d8, #16         \n"
-
-                        "vmla.f32   q12, q8, d5[1]      \n"
-                        "vmla.f32   q13, q8, d6[1]      \n"
-                        "vmla.f32   q14, q9, d6[0]      \n"
-                        "vmla.f32   q15, q9, d7[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%8]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[1]     \n"
-                        "vmla.f32   q13, q10, d7[1]     \n"
-                        "vmla.f32   q14, q11, d7[0]     \n"
-                        "vmla.f32   q15, q11, d9[0]     \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%8]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q6, d0[1]      \n"
-                        "vmla.f32   q15, q6, d1[1]      \n"
-
-                        "sub        %1, %1, #8          \n"
-                        "sub        %2, %2, #8          \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q8, d1[1]      \n"
-                        "vmla.f32   q15, q8, d2[1]      \n"
-
-                        "sub        %8, %8, #392        \n"
-
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q10, d2[1]     \n"
-                        "vmla.f32   q15, q10, d3[1]     \n"
-
-                        "sub        %3, %3, #8          \n"
-                        "sub        %4, %4, #8          \n"
-
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d8[0]     \n"
-
-                        "sub        %5, %5, #8          \n"
-                        "sub        %6, %6, #8          \n"
-
-                        "vadd.f32   q14, q14, q12       \n"
-                        "vadd.f32   q15, q15, q13       \n"
-
-                        "sub        %7, %7, #8          \n"
-
-                        "vst1.f32   {d28-d31}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v16.4s}, [%0]              \n"
-
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%1]        \n" // r0
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmul   v17.4s, v24.4s, v0.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmul   v18.4s, v25.4s, v0.s[1]     \n"
-                        "fmul   v19.4s, v26.4s, v0.s[2]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%2]        \n" // r1
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v18.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v30.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v18.4s, v26.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%3]        \n" // r2
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v19.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v17.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v18.4s, v30.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v19.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v26.4s, v0.s[2]     \n"
-
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%4]        \n" // r3
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v18.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v30.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v18.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v19.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%5, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%5]        \n" // r4
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v17.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v18.4s, v28.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v19.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v17.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v18.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v19.4s, v26.4s, v0.s[2]     \n"
-
-                        "prfm   pldl1keep, [%6, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%6]        \n" // r5
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v18.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v30.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v18.4s, v26.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%7, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%7]        \n" // r6
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v19.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%8], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v17.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v18.4s, v30.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%8], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v19.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v26.4s, v0.s[2]     \n"
-
-                        "add    %1, %1, #4                  \n"
-                        "add    %2, %2, #4                  \n"
-
-                        "fmla   v18.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v30.4s, v1.s[2]     \n"
-
-                        "add    %3, %3, #4                  \n"
-                        "add    %4, %4, #4                  \n"
-
-                        "fadd   v18.4s, v18.4s, v19.4s      \n"
-
-                        "add    %5, %5, #4                  \n"
-
-                        "fadd   v16.4s, v16.4s, v17.4s      \n"
-
-                        "add    %6, %6, #4                  \n"
-                        "add    %7, %7, #4                  \n"
-
-                        "fadd   v16.4s, v16.4s, v18.4s      \n"
-
-                        "sub    %8, %8, #392                \n"
-
-                        "st1    {v16.4s}, [%0], #16         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "v0", "v1", "v4", "v5", "v16", "v17", "v18", "v19", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%0, #128]          \n"
-                        "vld1.f32   {d8-d9}, [%0 :128]  \n"
-
-                        "pld        [%1, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%1]       \n" // r0
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%8]!    \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmul.f32   q5, q8, d0[0]       \n"
-                        "vmul.f32   q6, q9, d0[1]       \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d26-d28}, [%8]!    \n"
-
-                        "vshll.u16  q12, d26, #16       \n"
-                        "vshll.u16  q13, d27, #16       \n"
-                        "vshll.u16  q14, d28, #16       \n"
-
-                        "vmul.f32   q7, q10, d1[0]      \n"
-                        "vmla.f32   q4, q11, d1[1]      \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d6-d7}, [%2]       \n" // r1
-
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q5, q12, d2[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%8]!    \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q6, q13, d2[1]      \n"
-                        "vmla.f32   q7, q14, d3[0]      \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d26-d28}, [%8]!    \n"
-
-                        "vshll.u16  q12, d26, #16       \n"
-                        "vshll.u16  q13, d27, #16       \n"
-                        "vshll.u16  q14, d28, #16       \n"
-
-                        "vmla.f32   q4, q8, d4[0]       \n"
-                        "vmla.f32   q5, q9, d4[1]       \n"
-                        "vmla.f32   q6, q10, d5[0]      \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%3]       \n" // r2
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q7, q11, d5[1]      \n"
-                        "vmla.f32   q4, q12, d6[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%8]!    \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q5, q13, d6[1]      \n"
-                        "vmla.f32   q6, q14, d7[0]      \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d26-d28}, [%8]!    \n"
-
-                        "vshll.u16  q12, d26, #16       \n"
-                        "vshll.u16  q13, d27, #16       \n"
-                        "vshll.u16  q14, d28, #16       \n"
-
-                        "vmla.f32   q7, q8, d0[0]       \n"
-                        "vmla.f32   q4, q9, d0[1]       \n"
-                        "vmla.f32   q5, q10, d1[0]      \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.u16   {d6-d7}, [%4]       \n" // r3
-
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q6, q11, d1[1]      \n"
-                        "vmla.f32   q7, q12, d2[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%8]!    \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q4, q13, d2[1]      \n"
-                        "vmla.f32   q5, q14, d3[0]      \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d26-d28}, [%8]!    \n"
-
-                        "vshll.u16  q12, d26, #16       \n"
-                        "vshll.u16  q13, d27, #16       \n"
-                        "vshll.u16  q14, d28, #16       \n"
-
-                        "vmla.f32   q6, q8, d4[0]       \n"
-                        "vmla.f32   q7, q9, d4[1]       \n"
-                        "vmla.f32   q4, q10, d5[0]      \n"
-
-                        "pld        [%5, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%5]       \n" // r4
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q5, q11, d5[1]      \n"
-                        "vmla.f32   q6, q12, d6[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%8]!    \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q7, q13, d6[1]      \n"
-                        "vmla.f32   q4, q14, d7[0]      \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d26-d28}, [%8]!    \n"
-
-                        "vshll.u16  q12, d26, #16       \n"
-                        "vshll.u16  q13, d27, #16       \n"
-                        "vshll.u16  q14, d28, #16       \n"
-
-                        "vmla.f32   q5, q8, d0[0]       \n"
-                        "vmla.f32   q6, q9, d0[1]       \n"
-                        "vmla.f32   q7, q10, d1[0]      \n"
-
-                        "pld        [%6, #128]          \n"
-                        "vld1.u16   {d6-d7}, [%6]       \n" // r5
-
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q4, q11, d1[1]      \n"
-                        "vmla.f32   q5, q12, d2[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%8]!    \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q6, q13, d2[1]      \n"
-                        "vmla.f32   q7, q14, d3[0]      \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d26-d28}, [%8]!    \n"
-
-                        "vshll.u16  q12, d26, #16       \n"
-                        "vshll.u16  q13, d27, #16       \n"
-                        "vshll.u16  q14, d28, #16       \n"
-
-                        "vmla.f32   q4, q8, d4[0]       \n"
-                        "vmla.f32   q5, q9, d4[1]       \n"
-                        "vmla.f32   q6, q10, d5[0]      \n"
-
-                        "pld        [%7, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%7]       \n" // r6
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q7, q11, d5[1]      \n"
-                        "vmla.f32   q4, q12, d6[0]      \n"
-
-                        "pld        [%8, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%8]!    \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q5, q13, d6[1]      \n"
-                        "vmla.f32   q6, q14, d7[0]      \n"
-
-                        "pld        [%8, #192]          \n"
-                        "vld1.u16   {d26-d28}, [%8]!    \n"
-
-                        "vshll.u16  q12, d26, #16       \n"
-                        "vshll.u16  q13, d27, #16       \n"
-                        "vshll.u16  q14, d28, #16       \n"
-
-                        "vmla.f32   q7, q8, d0[0]       \n"
-                        "vmla.f32   q4, q9, d0[1]       \n"
-
-                        "add        %1, %1, #4          \n"
-                        "add        %2, %2, #4          \n"
-
-                        "vmla.f32   q5, q10, d1[0]      \n"
-                        "vmla.f32   q6, q11, d1[1]      \n"
-
-                        "sub        %8, %8, #392        \n"
-
-                        "vmla.f32   q7, q12, d2[0]      \n"
-                        "vmla.f32   q4, q13, d2[1]      \n"
-                        "vmla.f32   q5, q14, d3[0]      \n"
-
-                        "add        %3, %3, #4          \n"
-                        "add        %4, %4, #4          \n"
-
-                        "vadd.f32   q6, q6, q7          \n"
-
-                        "add        %5, %5, #4          \n"
-
-                        "vadd.f32   q4, q4, q5          \n"
-
-                        "add        %6, %6, #4          \n"
-
-                        "vadd.f32   q4, q4, q6          \n"
-
-                        "add        %7, %7, #4          \n"
-
-                        "vst1.f32   {d8-d9}, [%0 :128]! \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14");
-#endif // __aarch64__
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-                r4 += tailstep;
-                r5 += tailstep;
-                r6 += tailstep;
-            }
-        }
-        for (; q < inch; q++)
-        {
-            unsigned short* outptr0_bf16 = top_blob.channel(p);
-
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const unsigned short* r0 = img0.row<const unsigned short>(0);
-            const unsigned short* r1 = img0.row<const unsigned short>(1);
-            const unsigned short* r2 = img0.row<const unsigned short>(2);
-            const unsigned short* r3 = img0.row<const unsigned short>(3);
-            const unsigned short* r4 = img0.row<const unsigned short>(4);
-            const unsigned short* r5 = img0.row<const unsigned short>(5);
-            const unsigned short* r6 = img0.row<const unsigned short>(6);
-
-            const unsigned short* kptr = kernel.channel(p).row<const unsigned short>(q);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-#if __aarch64__
-                for (; j + 7 < outw; j += 8)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%1], #64 \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n" // r0
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%1], #64 \n"
-
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[2]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%2]        \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[0]     \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[1]     \n"
-
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v28.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v29.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v6.4h, v7.4h, v8.4h, v9.4h}, [%3], #32 \n" // r1
-
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-                        "shll   v8.4s, v8.4h, #16           \n"
-                        "shll   v9.4s, v9.4h, #16           \n"
-
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v30.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v30.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v10.4h, v11.4h}, [%3]      \n"
-
-                        "shll   v10.4s, v10.4h, #16         \n"
-                        "shll   v11.4s, v11.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v6.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v8.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v8.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v9.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v9.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v6.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v7.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v8.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v8.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v9.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v9.s[3]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v7.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v8.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v8.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v9.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v9.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v10.s[0]    \n"
-
-                        "fmla   v16.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v7.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v8.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v8.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v9.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v9.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v10.s[1]    \n"
-
-                        "fmla   v16.4s, v28.4s, v7.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v7.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v8.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v8.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v9.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v9.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v10.s[0]    \n"
-                        "fmla   v23.4s, v28.4s, v10.s[2]    \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v7.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v7.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v8.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v8.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v9.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v9.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v10.s[1]    \n"
-                        "fmla   v23.4s, v29.4s, v10.s[3]    \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%4], #32 \n" // r2
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v16.4s, v30.4s, v7.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v8.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v8.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v9.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v9.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v10.s[0]    \n"
-                        "fmla   v22.4s, v30.4s, v10.s[2]    \n"
-                        "fmla   v23.4s, v30.4s, v11.s[0]    \n"
-
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%4]        \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[0]     \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[1]     \n"
-
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v28.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v29.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v6.4h, v7.4h, v8.4h, v9.4h}, [%5], #32 \n" // r3
-
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-                        "shll   v8.4s, v8.4h, #16           \n"
-                        "shll   v9.4s, v9.4h, #16           \n"
-
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v30.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v30.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%5, #128]       \n"
-                        "ld1    {v10.4h, v11.4h}, [%5]      \n"
-
-                        "shll   v10.4s, v10.4h, #16         \n"
-                        "shll   v11.4s, v11.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v6.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v8.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v8.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v9.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v9.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v6.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v7.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v8.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v8.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v9.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v9.s[3]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v7.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v8.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v8.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v9.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v9.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v10.s[0]    \n"
-
-                        "fmla   v16.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v7.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v8.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v8.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v9.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v9.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v10.s[1]    \n"
-
-                        "fmla   v16.4s, v28.4s, v7.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v7.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v8.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v8.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v9.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v9.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v10.s[0]    \n"
-                        "fmla   v23.4s, v28.4s, v10.s[2]    \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v7.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v7.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v8.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v8.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v9.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v9.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v10.s[1]    \n"
-                        "fmla   v23.4s, v29.4s, v10.s[3]    \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%6], #32 \n" // r4
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v16.4s, v30.4s, v7.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v8.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v8.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v9.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v9.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v10.s[0]    \n"
-                        "fmla   v22.4s, v30.4s, v10.s[2]    \n"
-                        "fmla   v23.4s, v30.4s, v11.s[0]    \n"
-
-                        "prfm   pldl1keep, [%6, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%6]        \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[0]     \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[1]     \n"
-
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v28.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v29.4s, v4.s[3]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v6.4h, v7.4h, v8.4h, v9.4h}, [%7], #32 \n" // r5
-
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-                        "shll   v8.4s, v8.4h, #16           \n"
-                        "shll   v9.4s, v9.4h, #16           \n"
-
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v30.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v30.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%7, #128]       \n"
-                        "ld1    {v10.4h, v11.4h}, [%7]      \n"
-
-                        "shll   v10.4s, v10.4h, #16         \n"
-                        "shll   v11.4s, v11.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v6.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v6.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v7.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v7.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v8.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v8.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v9.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v9.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v6.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v6.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v7.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v7.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v8.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v8.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v9.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v9.s[3]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v26.4s, v6.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v7.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v7.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v8.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v8.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v9.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v9.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v10.s[0]    \n"
-
-                        "fmla   v16.4s, v27.4s, v6.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v7.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v7.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v8.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v8.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v9.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v9.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v10.s[1]    \n"
-
-                        "fmla   v16.4s, v28.4s, v7.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v7.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v8.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v8.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v9.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v9.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v10.s[0]    \n"
-                        "fmla   v23.4s, v28.4s, v10.s[2]    \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v7.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v7.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v8.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v8.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v9.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v9.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v10.s[1]    \n"
-                        "fmla   v23.4s, v29.4s, v10.s[3]    \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%8], #32 \n" // r6
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v16.4s, v30.4s, v7.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v8.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v8.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v9.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v9.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v10.s[0]    \n"
-                        "fmla   v22.4s, v30.4s, v10.s[2]    \n"
-                        "fmla   v23.4s, v30.4s, v11.s[0]    \n"
-
-                        "prfm   pldl1keep, [%8, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%8]        \n"
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v20.4s, v24.4s, v2.s[0]     \n"
-                        "fmla   v21.4s, v24.4s, v2.s[2]     \n"
-                        "fmla   v22.4s, v24.4s, v3.s[0]     \n"
-                        "fmla   v23.4s, v24.4s, v3.s[2]     \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v20.4s, v25.4s, v2.s[1]     \n"
-                        "fmla   v21.4s, v25.4s, v2.s[3]     \n"
-                        "fmla   v22.4s, v25.4s, v3.s[1]     \n"
-                        "fmla   v23.4s, v25.4s, v3.s[3]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-                        "fmla   v20.4s, v26.4s, v2.s[2]     \n"
-                        "fmla   v21.4s, v26.4s, v3.s[0]     \n"
-                        "fmla   v22.4s, v26.4s, v3.s[2]     \n"
-                        "fmla   v23.4s, v26.4s, v4.s[0]     \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v20.4s, v27.4s, v2.s[3]     \n"
-                        "fmla   v21.4s, v27.4s, v3.s[1]     \n"
-                        "fmla   v22.4s, v27.4s, v3.s[3]     \n"
-                        "fmla   v23.4s, v27.4s, v4.s[1]     \n"
-
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-                        "fmla   v20.4s, v28.4s, v3.s[0]     \n"
-                        "fmla   v21.4s, v28.4s, v3.s[2]     \n"
-                        "fmla   v22.4s, v28.4s, v4.s[0]     \n"
-                        "fmla   v23.4s, v28.4s, v4.s[2]     \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v20.4s, v29.4s, v3.s[1]     \n"
-                        "fmla   v21.4s, v29.4s, v3.s[3]     \n"
-                        "fmla   v22.4s, v29.4s, v4.s[1]     \n"
-                        "fmla   v23.4s, v29.4s, v4.s[3]     \n"
-
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-                        "fmla   v20.4s, v30.4s, v3.s[2]     \n"
-                        "fmla   v21.4s, v30.4s, v4.s[0]     \n"
-                        "fmla   v22.4s, v30.4s, v4.s[2]     \n"
-                        "fmla   v23.4s, v30.4s, v5.s[0]     \n"
-
-                        "sub    %9, %9, #392                \n"
-
-                        "shrn   v16.4h, v16.4s, #16         \n"
-                        "shrn   v17.4h, v17.4s, #16         \n"
-                        "shrn   v18.4h, v18.4s, #16         \n"
-                        "shrn   v19.4h, v19.4s, #16         \n"
-                        "shrn   v20.4h, v20.4s, #16         \n"
-                        "shrn   v21.4h, v21.4s, #16         \n"
-                        "shrn   v22.4h, v22.4s, #16         \n"
-                        "shrn   v23.4h, v23.4s, #16         \n"
-
-                        "st1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%0], #32 \n"
-                        "st1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%0], #32 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(r5),           // %7
-                        "=r"(r6),           // %8
-                        "=r"(kptr)          // %9
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(r5),
-                        "8"(r6),
-                        "9"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v16", "v17", "v18", "v19", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
-                }
-#endif // __aarch64__
-                for (; j + 3 < outw; j += 4)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #512]       \n"
-                        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%1], #64 \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2] \n" // r0
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%3] \n" // r1
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v5.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%4] \n" // r2
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v6.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v6.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v6.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v6.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v6.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v6.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v7.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%5] \n" // r3
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v5.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%6] \n" // r4
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v6.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v6.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v6.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v6.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v6.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v6.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v7.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%7] \n" // r5
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-                        "shll   v7.4s, v7.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v5.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v5.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v5.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #256]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%8] \n" // r6
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-                        "shll   v3.4s, v3.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v5.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v6.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v6.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v6.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v6.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v6.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v6.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v7.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v18.4s, v24.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v1.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v25.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v1.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v18.4s, v26.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v2.s[0]     \n"
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v27.4s, v1.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v2.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v18.4s, v28.4s, v2.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v2.s[2]     \n"
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v29.4s, v2.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v2.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v2.s[0]     \n"
-                        "fmla   v18.4s, v30.4s, v2.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v3.s[0]     \n"
-
-                        "add    %2, %2, #16                 \n"
-                        "add    %3, %3, #16                 \n"
-                        "add    %4, %4, #16                 \n"
-                        "add    %5, %5, #16                 \n"
-                        "add    %6, %6, #16                 \n"
-                        "add    %7, %7, #16                 \n"
-                        "add    %8, %8, #16                 \n"
-
-                        "sub    %9, %9, #392                \n"
-
-                        "shrn   v16.4h, v16.4s, #16         \n"
-                        "shrn   v17.4h, v17.4s, #16         \n"
-                        "shrn   v18.4h, v18.4s, #16         \n"
-                        "shrn   v19.4h, v19.4s, #16         \n"
-
-                        "st1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%0], #32 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(r5),           // %7
-                        "=r"(r6),           // %8
-                        "=r"(kptr)          // %9
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(r5),
-                        "8"(r6),
-                        "9"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #512]          \n"
-                        "vldm       %1!, {d24-d31}      \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%2]!      \n" // r0
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%9]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q5, d2[0]      \n"
-                        "vmla.f32   q15, q5, d3[0]      \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d5-d6}, [%2]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshl.u32   d6, d6, #16         \n"
-
-                        "vmla.f32   q12, q6, d0[1]      \n"
-                        "vmla.f32   q13, q6, d1[1]      \n"
-                        "vmla.f32   q14, q6, d2[1]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%9]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q7, d3[0]      \n"
-                        "vmla.f32   q15, q7, d4[0]      \n"
-                        "vmla.f32   q12, q8, d1[1]      \n"
-                        "vmla.f32   q13, q8, d2[1]      \n"
-                        "vmla.f32   q14, q8, d3[1]      \n"
-                        "vmla.f32   q15, q8, d4[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q9, d4[0]      \n"
-                        "vmla.f32   q15, q9, d5[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%9]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[1]     \n"
-                        "vmla.f32   q13, q10, d3[1]     \n"
-                        "vmla.f32   q14, q10, d4[1]     \n"
-                        "vmla.f32   q15, q10, d5[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d4[0]     \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%3]!      \n" // r1
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q14, q11, d5[0]     \n"
-                        "vmla.f32   q15, q11, d6[0]     \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q5, d2[0]      \n"
-                        "vmla.f32   q15, q5, d3[0]      \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d5-d6}, [%3]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshl.u32   d6, d6, #16         \n"
-
-                        "vmla.f32   q12, q6, d0[1]      \n"
-                        "vmla.f32   q13, q6, d1[1]      \n"
-                        "vmla.f32   q14, q6, d2[1]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%9]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q7, d3[0]      \n"
-                        "vmla.f32   q15, q7, d4[0]      \n"
-                        "vmla.f32   q12, q8, d1[1]      \n"
-                        "vmla.f32   q13, q8, d2[1]      \n"
-                        "vmla.f32   q14, q8, d3[1]      \n"
-                        "vmla.f32   q15, q8, d4[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q9, d4[0]      \n"
-                        "vmla.f32   q15, q9, d5[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%9]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[1]     \n"
-                        "vmla.f32   q13, q10, d3[1]     \n"
-                        "vmla.f32   q14, q10, d4[1]     \n"
-                        "vmla.f32   q15, q10, d5[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d4[0]     \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%4]!      \n" // r2
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q14, q11, d5[0]     \n"
-                        "vmla.f32   q15, q11, d6[0]     \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q5, d2[0]      \n"
-                        "vmla.f32   q15, q5, d3[0]      \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.u16   {d5-d6}, [%4]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshl.u32   d6, d6, #16         \n"
-
-                        "vmla.f32   q12, q6, d0[1]      \n"
-                        "vmla.f32   q13, q6, d1[1]      \n"
-                        "vmla.f32   q14, q6, d2[1]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%9]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q7, d3[0]      \n"
-                        "vmla.f32   q15, q7, d4[0]      \n"
-                        "vmla.f32   q12, q8, d1[1]      \n"
-                        "vmla.f32   q13, q8, d2[1]      \n"
-                        "vmla.f32   q14, q8, d3[1]      \n"
-                        "vmla.f32   q15, q8, d4[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q9, d4[0]      \n"
-                        "vmla.f32   q15, q9, d5[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%9]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[1]     \n"
-                        "vmla.f32   q13, q10, d3[1]     \n"
-                        "vmla.f32   q14, q10, d4[1]     \n"
-                        "vmla.f32   q15, q10, d5[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d4[0]     \n"
-
-                        "pld        [%5, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%5]!      \n" // r3
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q14, q11, d5[0]     \n"
-                        "vmla.f32   q15, q11, d6[0]     \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q5, d2[0]      \n"
-                        "vmla.f32   q15, q5, d3[0]      \n"
-
-                        "pld        [%5, #128]          \n"
-                        "vld1.u16   {d5-d6}, [%5]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshl.u32   d6, d6, #16         \n"
-
-                        "vmla.f32   q12, q6, d0[1]      \n"
-                        "vmla.f32   q13, q6, d1[1]      \n"
-                        "vmla.f32   q14, q6, d2[1]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%9]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q7, d3[0]      \n"
-                        "vmla.f32   q15, q7, d4[0]      \n"
-                        "vmla.f32   q12, q8, d1[1]      \n"
-                        "vmla.f32   q13, q8, d2[1]      \n"
-                        "vmla.f32   q14, q8, d3[1]      \n"
-                        "vmla.f32   q15, q8, d4[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q9, d4[0]      \n"
-                        "vmla.f32   q15, q9, d5[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%9]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[1]     \n"
-                        "vmla.f32   q13, q10, d3[1]     \n"
-                        "vmla.f32   q14, q10, d4[1]     \n"
-                        "vmla.f32   q15, q10, d5[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d4[0]     \n"
-
-                        "pld        [%6, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%6]!      \n" // r4
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q14, q11, d5[0]     \n"
-                        "vmla.f32   q15, q11, d6[0]     \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q5, d2[0]      \n"
-                        "vmla.f32   q15, q5, d3[0]      \n"
-
-                        "pld        [%6, #128]          \n"
-                        "vld1.u16   {d5-d6}, [%6]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshl.u32   d6, d6, #16         \n"
-
-                        "vmla.f32   q12, q6, d0[1]      \n"
-                        "vmla.f32   q13, q6, d1[1]      \n"
-                        "vmla.f32   q14, q6, d2[1]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%9]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q7, d3[0]      \n"
-                        "vmla.f32   q15, q7, d4[0]      \n"
-                        "vmla.f32   q12, q8, d1[1]      \n"
-                        "vmla.f32   q13, q8, d2[1]      \n"
-                        "vmla.f32   q14, q8, d3[1]      \n"
-                        "vmla.f32   q15, q8, d4[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q9, d4[0]      \n"
-                        "vmla.f32   q15, q9, d5[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%9]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[1]     \n"
-                        "vmla.f32   q13, q10, d3[1]     \n"
-                        "vmla.f32   q14, q10, d4[1]     \n"
-                        "vmla.f32   q15, q10, d5[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d4[0]     \n"
-
-                        "pld        [%7, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%7]!      \n" // r5
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q14, q11, d5[0]     \n"
-                        "vmla.f32   q15, q11, d6[0]     \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q5, d2[0]      \n"
-                        "vmla.f32   q15, q5, d3[0]      \n"
-
-                        "pld        [%7, #128]          \n"
-                        "vld1.u16   {d5-d6}, [%7]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshl.u32   d6, d6, #16         \n"
-
-                        "vmla.f32   q12, q6, d0[1]      \n"
-                        "vmla.f32   q13, q6, d1[1]      \n"
-                        "vmla.f32   q14, q6, d2[1]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%9]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q7, d3[0]      \n"
-                        "vmla.f32   q15, q7, d4[0]      \n"
-                        "vmla.f32   q12, q8, d1[1]      \n"
-                        "vmla.f32   q13, q8, d2[1]      \n"
-                        "vmla.f32   q14, q8, d3[1]      \n"
-                        "vmla.f32   q15, q8, d4[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q9, d4[0]      \n"
-                        "vmla.f32   q15, q9, d5[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%9]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d2[1]     \n"
-                        "vmla.f32   q13, q10, d3[1]     \n"
-                        "vmla.f32   q14, q10, d4[1]     \n"
-                        "vmla.f32   q15, q10, d5[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d4[0]     \n"
-
-                        "pld        [%8, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%8]!      \n" // r6
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q14, q11, d5[0]     \n"
-                        "vmla.f32   q15, q11, d6[0]     \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q5, d2[0]      \n"
-                        "vmla.f32   q15, q5, d3[0]      \n"
-
-                        "pld        [%8, #128]          \n"
-                        "vld1.u16   {d5-d6}, [%8]       \n"
-
-                        "vshll.u16  q2, d5, #16         \n"
-                        "vshl.u32   d6, d6, #16         \n"
-
-                        "vmla.f32   q12, q6, d0[1]      \n"
-                        "vmla.f32   q13, q6, d1[1]      \n"
-                        "vmla.f32   q14, q6, d2[1]      \n"
-                        "vmla.f32   q15, q6, d3[1]      \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%9]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q7, d3[0]      \n"
-                        "vmla.f32   q15, q7, d4[0]      \n"
-                        "vmla.f32   q12, q8, d1[1]      \n"
-                        "vmla.f32   q13, q8, d2[1]      \n"
-                        "vmla.f32   q14, q8, d3[1]      \n"
-                        "vmla.f32   q15, q8, d4[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q9, d4[0]      \n"
-                        "vmla.f32   q15, q9, d5[0]      \n"
-                        "vmla.f32   q12, q10, d2[1]     \n"
-                        "vmla.f32   q13, q10, d3[1]     \n"
-                        "vmla.f32   q14, q10, d4[1]     \n"
-                        "vmla.f32   q15, q10, d5[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d4[0]     \n"
-                        "vmla.f32   q14, q11, d5[0]     \n"
-                        "vmla.f32   q15, q11, d6[0]     \n"
-
-                        "sub        %9, %9, #392        \n"
-
-                        "vshrn.u32  d24, q12, #16       \n"
-                        "vshrn.u32  d25, q13, #16       \n"
-                        "vshrn.u32  d26, q14, #16       \n"
-                        "vshrn.u32  d27, q15, #16       \n"
-
-                        "vst1.u16   {d24-d27}, [%0]!    \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(r5),           // %7
-                        "=r"(r6),           // %8
-                        "=r"(kptr)          // %9
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(r5),
-                        "8"(r6),
-                        "9"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v16.4s, v17.4s}, [%1], #32 \n"
-
-                        "prfm   pldl1keep, [%2, #192]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%2] \n" // r0
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmul   v18.4s, v24.4s, v0.s[0]     \n"
-                        "fmul   v19.4s, v24.4s, v0.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%3, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%3] \n" // r1
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%4, #192]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%4] \n" // r2
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v18.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v18.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v18.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%5, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%5] \n" // r3
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%6, #192]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%6] \n" // r4
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v18.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v18.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v18.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%7, #192]       \n"
-                        "ld1    {v4.4h, v5.4h, v6.4h}, [%7] \n" // r5
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-                        "shll   v6.4s, v6.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v2.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v24.4s, v4.s[2]     \n"
-                        "fmla   v18.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v19.4s, v25.4s, v4.s[3]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-                        "fmla   v17.4s, v26.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #192]       \n"
-                        "ld1    {v0.4h, v1.4h, v2.4h}, [%8] \n" // r6
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-                        "shll   v2.4s, v2.4h, #16           \n"
-
-                        "fmla   v18.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v19.4s, v27.4s, v5.s[1]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-                        "fmla   v17.4s, v28.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v18.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v19.4s, v29.4s, v5.s[3]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-                        "fmla   v17.4s, v30.4s, v6.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v18.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v19.4s, v24.4s, v0.s[2]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v25.4s, v0.s[3]     \n"
-                        "fmla   v18.4s, v26.4s, v0.s[2]     \n"
-                        "fmla   v19.4s, v26.4s, v1.s[0]     \n"
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v27.4s, v1.s[1]     \n"
-                        "fmla   v18.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[2]     \n"
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v29.4s, v1.s[3]     \n"
-                        "fmla   v18.4s, v30.4s, v1.s[2]     \n"
-                        "fmla   v19.4s, v30.4s, v2.s[0]     \n"
-
-                        "add    %2, %2, #8                  \n"
-                        "add    %3, %3, #8                  \n"
-                        "add    %4, %4, #8                  \n"
-                        "add    %5, %5, #8                  \n"
-                        "add    %6, %6, #8                  \n"
-                        "add    %7, %7, #8                  \n"
-                        "add    %8, %8, #8                  \n"
-
-                        "fadd   v16.4s, v16.4s, v18.4s      \n"
-                        "fadd   v17.4s, v17.4s, v19.4s      \n"
-
-                        "sub    %9, %9, #392                \n"
-
-                        "shrn   v16.4h, v16.4s, #16         \n"
-                        "shrn   v17.4h, v17.4s, #16         \n"
-
-                        "st1    {v16.4h, v17.4h}, [%0], #16 \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(r5),           // %7
-                        "=r"(r6),           // %8
-                        "=r"(kptr)          // %9
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(r5),
-                        "8"(r6),
-                        "9"(kptr)
-                        : "memory", "v0", "v1", "v2", "v4", "v5", "v6", "v16", "v17", "v18", "v19", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #256]          \n"
-                        "vld1.f32   {d28-d31}, [%1 :128]! \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%2]!      \n" // r0
-                        "vld1.u16   {d8[0]}, [%2]       \n"
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vshl.u32   d8, d8, #16         \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%9]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmul.f32   q12, q5, d0[0]      \n"
-                        "vmul.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q6, d0[1]      \n"
-                        "vmla.f32   q15, q6, d1[1]      \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%9]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d6-d7}, [%3]!      \n" // r1
-                        "vld1.u16   {d9[0]}, [%3]       \n"
-
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshl.u32   d9, d9, #16         \n"
-
-                        "vmla.f32   q14, q8, d1[1]      \n"
-                        "vmla.f32   q15, q8, d2[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%9]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q14, q10, d2[1]     \n"
-                        "vmla.f32   q15, q10, d3[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d8[0]     \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%9]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q14, q5, d4[0]      \n"
-                        "vmla.f32   q15, q5, d5[0]      \n"
-                        "vmla.f32   q12, q6, d4[1]      \n"
-                        "vmla.f32   q13, q6, d5[1]      \n"
-                        "vmla.f32   q14, q7, d5[0]      \n"
-                        "vmla.f32   q15, q7, d6[0]      \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%4]!      \n" // r2
-                        "vld1.u16   {d8[0]}, [%4]       \n"
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vshl.u32   d8, d8, #16         \n"
-
-                        "vmla.f32   q12, q8, d5[1]      \n"
-                        "vmla.f32   q13, q8, d6[1]      \n"
-                        "vmla.f32   q14, q9, d6[0]      \n"
-                        "vmla.f32   q15, q9, d7[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%9]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[1]     \n"
-                        "vmla.f32   q13, q10, d7[1]     \n"
-                        "vmla.f32   q14, q11, d7[0]     \n"
-                        "vmla.f32   q15, q11, d9[0]     \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%9]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q6, d0[1]      \n"
-                        "vmla.f32   q15, q6, d1[1]      \n"
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-
-                        "pld        [%5, #128]          \n"
-                        "vld1.u16   {d6-d7}, [%5]!      \n" // r3
-                        "vld1.u16   {d9[0]}, [%5]       \n"
-
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshl.u32   d9, d9, #16         \n"
-
-                        "vmla.f32   q14, q8, d1[1]      \n"
-                        "vmla.f32   q15, q8, d2[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%9]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q14, q10, d2[1]     \n"
-                        "vmla.f32   q15, q10, d3[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d8[0]     \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%9]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q14, q5, d4[0]      \n"
-                        "vmla.f32   q15, q5, d5[0]      \n"
-                        "vmla.f32   q12, q6, d4[1]      \n"
-                        "vmla.f32   q13, q6, d5[1]      \n"
-                        "vmla.f32   q14, q7, d5[0]      \n"
-                        "vmla.f32   q15, q7, d6[0]      \n"
-
-                        "pld        [%6, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%6]!      \n" // r4
-                        "vld1.u16   {d8[0]}, [%6]       \n"
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vshl.u32   d8, d8, #16         \n"
-
-                        "vmla.f32   q12, q8, d5[1]      \n"
-                        "vmla.f32   q13, q8, d6[1]      \n"
-                        "vmla.f32   q14, q9, d6[0]      \n"
-                        "vmla.f32   q15, q9, d7[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%9]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[1]     \n"
-                        "vmla.f32   q13, q10, d7[1]     \n"
-                        "vmla.f32   q14, q11, d7[0]     \n"
-                        "vmla.f32   q15, q11, d9[0]     \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%9]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q6, d0[1]      \n"
-                        "vmla.f32   q15, q6, d1[1]      \n"
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-
-                        "pld        [%7, #128]          \n"
-                        "vld1.u16   {d6-d7}, [%7]!      \n" // r5
-                        "vld1.u16   {d9[0]}, [%7]       \n"
-
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-                        "vshl.u32   d9, d9, #16         \n"
-
-                        "vmla.f32   q14, q8, d1[1]      \n"
-                        "vmla.f32   q15, q8, d2[1]      \n"
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%9]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q14, q10, d2[1]     \n"
-                        "vmla.f32   q15, q10, d3[1]     \n"
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d8[0]     \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%9]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q14, q5, d4[0]      \n"
-                        "vmla.f32   q15, q5, d5[0]      \n"
-                        "vmla.f32   q12, q6, d4[1]      \n"
-                        "vmla.f32   q13, q6, d5[1]      \n"
-                        "vmla.f32   q14, q7, d5[0]      \n"
-                        "vmla.f32   q15, q7, d6[0]      \n"
-
-                        "pld        [%8, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%8]!      \n" // r6
-                        "vld1.u16   {d8[0]}, [%8]       \n"
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-                        "vshl.u32   d8, d8, #16         \n"
-
-                        "vmla.f32   q12, q8, d5[1]      \n"
-                        "vmla.f32   q13, q8, d6[1]      \n"
-                        "vmla.f32   q14, q9, d6[0]      \n"
-                        "vmla.f32   q15, q9, d7[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d14-d17}, [%9]!    \n"
-
-                        "vshll.u16  q5, d14, #16        \n"
-                        "vshll.u16  q6, d15, #16        \n"
-                        "vshll.u16  q7, d16, #16        \n"
-                        "vshll.u16  q8, d17, #16        \n"
-
-                        "vmla.f32   q12, q10, d6[1]     \n"
-                        "vmla.f32   q13, q10, d7[1]     \n"
-                        "vmla.f32   q14, q11, d7[0]     \n"
-                        "vmla.f32   q15, q11, d9[0]     \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d20-d22}, [%9]!    \n"
-
-                        "vshll.u16  q9, d20, #16        \n"
-                        "vshll.u16  q10, d21, #16       \n"
-                        "vshll.u16  q11, d22, #16       \n"
-
-                        "vmla.f32   q12, q5, d0[0]      \n"
-                        "vmla.f32   q13, q5, d1[0]      \n"
-                        "vmla.f32   q14, q6, d0[1]      \n"
-                        "vmla.f32   q15, q6, d1[1]      \n"
-
-                        "sub        %2, %2, #8          \n"
-                        "sub        %3, %3, #8          \n"
-
-                        "vmla.f32   q12, q7, d1[0]      \n"
-                        "vmla.f32   q13, q7, d2[0]      \n"
-                        "vmla.f32   q14, q8, d1[1]      \n"
-                        "vmla.f32   q15, q8, d2[1]      \n"
-
-                        "sub        %9, %9, #392        \n"
-
-                        "vmla.f32   q12, q9, d2[0]      \n"
-                        "vmla.f32   q13, q9, d3[0]      \n"
-                        "vmla.f32   q14, q10, d2[1]     \n"
-                        "vmla.f32   q15, q10, d3[1]     \n"
-
-                        "sub        %4, %4, #8          \n"
-                        "sub        %5, %5, #8          \n"
-
-                        "vmla.f32   q12, q11, d3[0]     \n"
-                        "vmla.f32   q13, q11, d8[0]     \n"
-
-                        "sub        %6, %6, #8          \n"
-                        "sub        %7, %7, #8          \n"
-
-                        "vadd.f32   q14, q14, q12       \n"
-                        "vadd.f32   q15, q15, q13       \n"
-
-                        "sub        %8, %8, #8          \n"
-
-                        "vshrn.u32  d28, q14, #16       \n"
-                        "vshrn.u32  d29, q15, #16       \n"
-
-                        "vst1.u16   {d28-d29}, [%0 :64]! \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(r5),           // %7
-                        "=r"(r6),           // %8
-                        "=r"(kptr)          // %9
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(r5),
-                        "8"(r6),
-                        "9"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-                }
-                for (; j < outw; j++)
-                {
-#if __aarch64__
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v16.4s}, [%1], #16         \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%2]        \n" // r0
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmul   v17.4s, v24.4s, v0.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmul   v18.4s, v25.4s, v0.s[1]     \n"
-                        "fmul   v19.4s, v26.4s, v0.s[2]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%3]        \n" // r1
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v18.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v30.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v18.4s, v26.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%4]        \n" // r2
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v19.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v17.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v18.4s, v30.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v19.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v26.4s, v0.s[2]     \n"
-
-                        "prfm   pldl1keep, [%5, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%5]        \n" // r3
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v18.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v30.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v18.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v19.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v16.4s, v26.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%6, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%6]        \n" // r4
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v17.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v18.4s, v28.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v19.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v16.4s, v30.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v17.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v18.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v19.4s, v26.4s, v0.s[2]     \n"
-
-                        "prfm   pldl1keep, [%7, #128]       \n"
-                        "ld1    {v4.4h, v5.4h}, [%7]        \n" // r5
-
-                        "shll   v4.4s, v4.4h, #16           \n"
-                        "shll   v5.4s, v5.4h, #16           \n"
-
-                        "fmla   v16.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v17.4s, v28.4s, v1.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v18.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v19.4s, v30.4s, v1.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v16.4s, v24.4s, v4.s[0]     \n"
-                        "fmla   v17.4s, v25.4s, v4.s[1]     \n"
-                        "fmla   v18.4s, v26.4s, v4.s[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #128]       \n"
-                        "ld1    {v0.4h, v1.4h}, [%8]        \n" // r6
-
-                        "shll   v0.4s, v0.4h, #16           \n"
-                        "shll   v1.4s, v1.4h, #16           \n"
-
-                        "fmla   v19.4s, v27.4s, v4.s[3]     \n"
-                        "fmla   v16.4s, v28.4s, v5.s[0]     \n"
-
-                        "prfm   pldl1keep, [%9, #256]       \n"
-                        "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%9], #32 \n"
-
-                        "shll   v24.4s, v24.4h, #16         \n"
-                        "shll   v25.4s, v25.4h, #16         \n"
-                        "shll   v26.4s, v26.4h, #16         \n"
-                        "shll   v27.4s, v27.4h, #16         \n"
-
-                        "fmla   v17.4s, v29.4s, v5.s[1]     \n"
-                        "fmla   v18.4s, v30.4s, v5.s[2]     \n"
-
-                        "prfm   pldl1keep, [%9, #192]       \n"
-                        "ld1    {v28.4h, v29.4h, v30.4h}, [%9], #24 \n"
-
-                        "shll   v28.4s, v28.4h, #16         \n"
-                        "shll   v29.4s, v29.4h, #16         \n"
-                        "shll   v30.4s, v30.4h, #16         \n"
-
-                        "fmla   v19.4s, v24.4s, v0.s[0]     \n"
-                        "fmla   v16.4s, v25.4s, v0.s[1]     \n"
-                        "fmla   v17.4s, v26.4s, v0.s[2]     \n"
-
-                        "add    %2, %2, #4                  \n"
-                        "add    %3, %3, #4                  \n"
-
-                        "fmla   v18.4s, v27.4s, v0.s[3]     \n"
-                        "fmla   v19.4s, v28.4s, v1.s[0]     \n"
-                        "fmla   v16.4s, v29.4s, v1.s[1]     \n"
-                        "fmla   v17.4s, v30.4s, v1.s[2]     \n"
-
-                        "add    %4, %4, #4                  \n"
-                        "add    %5, %5, #4                  \n"
-
-                        "fadd   v18.4s, v18.4s, v19.4s      \n"
-
-                        "add    %6, %6, #4                  \n"
-
-                        "fadd   v16.4s, v16.4s, v17.4s      \n"
-
-                        "add    %7, %7, #4                  \n"
-                        "add    %8, %8, #4                  \n"
-
-                        "fadd   v16.4s, v16.4s, v18.4s      \n"
-
-                        "sub    %9, %9, #392                \n"
-
-                        "shrn   v16.4h, v16.4s, #16         \n"
-
-                        "st1    {v16.4h}, [%0], #8          \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(r5),           // %7
-                        "=r"(r6),           // %8
-                        "=r"(kptr)          // %9
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(r5),
-                        "8"(r6),
-                        "9"(kptr)
-                        : "memory", "v0", "v1", "v4", "v5", "v16", "v17", "v18", "v19", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
-#else  // __aarch64__
-                    asm volatile(
-                        "pld        [%1, #128]          \n"
-                        "vld1.f32   {d8-d9}, [%1 :128]! \n"
-
-                        "pld        [%2, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%2]       \n" // r0
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%9]!    \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmul.f32   q5, q8, d0[0]       \n"
-                        "vmul.f32   q6, q9, d0[1]       \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d26-d28}, [%9]!    \n"
-
-                        "vshll.u16  q12, d26, #16       \n"
-                        "vshll.u16  q13, d27, #16       \n"
-                        "vshll.u16  q14, d28, #16       \n"
-
-                        "vmul.f32   q7, q10, d1[0]      \n"
-                        "vmla.f32   q4, q11, d1[1]      \n"
-
-                        "pld        [%3, #128]          \n"
-                        "vld1.u16   {d6-d7}, [%3]       \n" // r1
-
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q5, q12, d2[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%9]!    \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q6, q13, d2[1]      \n"
-                        "vmla.f32   q7, q14, d3[0]      \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d26-d28}, [%9]!    \n"
-
-                        "vshll.u16  q12, d26, #16       \n"
-                        "vshll.u16  q13, d27, #16       \n"
-                        "vshll.u16  q14, d28, #16       \n"
-
-                        "vmla.f32   q4, q8, d4[0]       \n"
-                        "vmla.f32   q5, q9, d4[1]       \n"
-                        "vmla.f32   q6, q10, d5[0]      \n"
-
-                        "pld        [%4, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%4]       \n" // r2
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q7, q11, d5[1]      \n"
-                        "vmla.f32   q4, q12, d6[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%9]!    \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q5, q13, d6[1]      \n"
-                        "vmla.f32   q6, q14, d7[0]      \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d26-d28}, [%9]!    \n"
-
-                        "vshll.u16  q12, d26, #16       \n"
-                        "vshll.u16  q13, d27, #16       \n"
-                        "vshll.u16  q14, d28, #16       \n"
-
-                        "vmla.f32   q7, q8, d0[0]       \n"
-                        "vmla.f32   q4, q9, d0[1]       \n"
-                        "vmla.f32   q5, q10, d1[0]      \n"
-
-                        "pld        [%5, #128]          \n"
-                        "vld1.u16   {d6-d7}, [%5]       \n" // r3
-
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q6, q11, d1[1]      \n"
-                        "vmla.f32   q7, q12, d2[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%9]!    \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q4, q13, d2[1]      \n"
-                        "vmla.f32   q5, q14, d3[0]      \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d26-d28}, [%9]!    \n"
-
-                        "vshll.u16  q12, d26, #16       \n"
-                        "vshll.u16  q13, d27, #16       \n"
-                        "vshll.u16  q14, d28, #16       \n"
-
-                        "vmla.f32   q6, q8, d4[0]       \n"
-                        "vmla.f32   q7, q9, d4[1]       \n"
-                        "vmla.f32   q4, q10, d5[0]      \n"
-
-                        "pld        [%6, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%6]       \n" // r4
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q5, q11, d5[1]      \n"
-                        "vmla.f32   q6, q12, d6[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%9]!    \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q7, q13, d6[1]      \n"
-                        "vmla.f32   q4, q14, d7[0]      \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d26-d28}, [%9]!    \n"
-
-                        "vshll.u16  q12, d26, #16       \n"
-                        "vshll.u16  q13, d27, #16       \n"
-                        "vshll.u16  q14, d28, #16       \n"
-
-                        "vmla.f32   q5, q8, d0[0]       \n"
-                        "vmla.f32   q6, q9, d0[1]       \n"
-                        "vmla.f32   q7, q10, d1[0]      \n"
-
-                        "pld        [%7, #128]          \n"
-                        "vld1.u16   {d6-d7}, [%7]       \n" // r5
-
-                        "vshll.u16  q2, d6, #16         \n"
-                        "vshll.u16  q3, d7, #16         \n"
-
-                        "vmla.f32   q4, q11, d1[1]      \n"
-                        "vmla.f32   q5, q12, d2[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%9]!    \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q6, q13, d2[1]      \n"
-                        "vmla.f32   q7, q14, d3[0]      \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d26-d28}, [%9]!    \n"
-
-                        "vshll.u16  q12, d26, #16       \n"
-                        "vshll.u16  q13, d27, #16       \n"
-                        "vshll.u16  q14, d28, #16       \n"
-
-                        "vmla.f32   q4, q8, d4[0]       \n"
-                        "vmla.f32   q5, q9, d4[1]       \n"
-                        "vmla.f32   q6, q10, d5[0]      \n"
-
-                        "pld        [%8, #128]          \n"
-                        "vld1.u16   {d2-d3}, [%8]       \n" // r6
-
-                        "vshll.u16  q0, d2, #16         \n"
-                        "vshll.u16  q1, d3, #16         \n"
-
-                        "vmla.f32   q7, q11, d5[1]      \n"
-                        "vmla.f32   q4, q12, d6[0]      \n"
-
-                        "pld        [%9, #256]          \n"
-                        "vld1.u16   {d20-d23}, [%9]!    \n"
-
-                        "vshll.u16  q8, d20, #16        \n"
-                        "vshll.u16  q9, d21, #16        \n"
-                        "vshll.u16  q10, d22, #16       \n"
-                        "vshll.u16  q11, d23, #16       \n"
-
-                        "vmla.f32   q5, q13, d6[1]      \n"
-                        "vmla.f32   q6, q14, d7[0]      \n"
-
-                        "pld        [%9, #192]          \n"
-                        "vld1.u16   {d26-d28}, [%9]!    \n"
-
-                        "vshll.u16  q12, d26, #16       \n"
-                        "vshll.u16  q13, d27, #16       \n"
-                        "vshll.u16  q14, d28, #16       \n"
-
-                        "vmla.f32   q7, q8, d0[0]       \n"
-                        "vmla.f32   q4, q9, d0[1]       \n"
-
-                        "add        %2, %2, #4          \n"
-                        "add        %3, %3, #4          \n"
-
-                        "vmla.f32   q5, q10, d1[0]      \n"
-                        "vmla.f32   q6, q11, d1[1]      \n"
-
-                        "sub        %9, %9, #392        \n"
-
-                        "vmla.f32   q7, q12, d2[0]      \n"
-                        "vmla.f32   q4, q13, d2[1]      \n"
-                        "vmla.f32   q5, q14, d3[0]      \n"
-
-                        "add        %4, %4, #4          \n"
-                        "add        %5, %5, #4          \n"
-
-                        "vadd.f32   q6, q6, q7          \n"
-
-                        "add        %6, %6, #4          \n"
-
-                        "vadd.f32   q4, q4, q5          \n"
-
-                        "add        %7, %7, #4          \n"
-
-                        "vadd.f32   q4, q4, q6          \n"
-
-                        "add        %8, %8, #4          \n"
-
-                        "vshrn.u32  d8, q4, #16         \n"
-
-                        "vst1.u16   {d8}, [%0 :64]!     \n"
-
-                        : "=r"(outptr0_bf16), // %0
-                        "=r"(outptr0),      // %1
-                        "=r"(r0),           // %2
-                        "=r"(r1),           // %3
-                        "=r"(r2),           // %4
-                        "=r"(r3),           // %5
-                        "=r"(r4),           // %6
-                        "=r"(r5),           // %7
-                        "=r"(r6),           // %8
-                        "=r"(kptr)          // %9
-                        : "0"(outptr0_bf16),
-                        "1"(outptr0),
-                        "2"(r0),
-                        "3"(r1),
-                        "4"(r2),
-                        "5"(r3),
-                        "6"(r4),
-                        "7"(r5),
-                        "8"(r6),
-                        "9"(kptr)
-                        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14");
-#endif // __aarch64__
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-                r4 += tailstep;
-                r5 += tailstep;
-                r6 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7_pack1to8_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7_pack1to8_fp16s.h
deleted file mode 100644
index 7af4135a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_7x7_pack1to8_fp16s.h
+++ /dev/null
@@ -1,1155 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv7x7s2_pack1to8_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        float16x8_t _bias0 = bias ? vld1q_f16(bias + p * 8) : vdupq_n_f16(0.f);
-        out0.fill(_bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            __fp16* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const __fp16* r0 = img0.row<const __fp16>(0);
-            const __fp16* r1 = img0.row<const __fp16>(1);
-            const __fp16* r2 = img0.row<const __fp16>(2);
-            const __fp16* r3 = img0.row<const __fp16>(3);
-            const __fp16* r4 = img0.row<const __fp16>(4);
-            const __fp16* r5 = img0.row<const __fp16>(5);
-            const __fp16* r6 = img0.row<const __fp16>(6);
-
-            const __fp16* kptr = kernel.channel(p).row<const __fp16>(q);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n" // sum0
-
-                        "prfm   pldl1keep, [%1, #384]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h}, [%1] \n" // r0
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0] \n" // sum0
-
-                        "fmla   v24.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v25.8h, v16.8h, v0.h[2]     \n"
-                        "fmla   v26.8h, v16.8h, v0.h[4]     \n"
-                        "fmla   v27.8h, v16.8h, v0.h[6]     \n"
-                        "fmla   v28.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, v16.8h, v1.h[4]     \n"
-                        "fmla   v31.8h, v16.8h, v1.h[6]     \n"
-
-                        "sub    %0, %0, #64                 \n"
-
-                        "fmla   v24.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v25.8h, v17.8h, v0.h[3]     \n"
-                        "fmla   v26.8h, v17.8h, v0.h[5]     \n"
-                        "fmla   v27.8h, v17.8h, v0.h[7]     \n"
-                        "fmla   v28.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[3]     \n"
-                        "fmla   v30.8h, v17.8h, v1.h[5]     \n"
-                        "fmla   v31.8h, v17.8h, v1.h[7]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v24.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v25.8h, v18.8h, v0.h[4]     \n"
-                        "fmla   v26.8h, v18.8h, v0.h[6]     \n"
-                        "fmla   v27.8h, v18.8h, v1.h[0]     \n"
-                        "fmla   v28.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v1.h[4]     \n"
-                        "fmla   v30.8h, v18.8h, v1.h[6]     \n"
-                        "fmla   v31.8h, v18.8h, v2.h[0]     \n"
-
-                        "fmla   v24.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v25.8h, v19.8h, v0.h[5]     \n"
-                        "fmla   v26.8h, v19.8h, v0.h[7]     \n"
-                        "fmla   v27.8h, v19.8h, v1.h[1]     \n"
-                        "fmla   v28.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v1.h[5]     \n"
-                        "fmla   v30.8h, v19.8h, v1.h[7]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[1]     \n"
-
-                        "prfm   pldl1keep, [%2, #384]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h}, [%2] \n" // r1
-
-                        "fmla   v24.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v25.8h, v20.8h, v0.h[6]     \n"
-                        "fmla   v26.8h, v20.8h, v1.h[0]     \n"
-                        "fmla   v27.8h, v20.8h, v1.h[2]     \n"
-                        "fmla   v28.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v1.h[6]     \n"
-                        "fmla   v30.8h, v20.8h, v2.h[0]     \n"
-                        "fmla   v31.8h, v20.8h, v2.h[2]     \n"
-
-                        "fmla   v24.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v25.8h, v21.8h, v0.h[7]     \n"
-                        "fmla   v26.8h, v21.8h, v1.h[1]     \n"
-                        "fmla   v27.8h, v21.8h, v1.h[3]     \n"
-                        "fmla   v28.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[7]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v21.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v24.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v25.8h, v22.8h, v1.h[0]     \n"
-                        "fmla   v26.8h, v22.8h, v1.h[2]     \n"
-                        "fmla   v27.8h, v22.8h, v1.h[4]     \n"
-                        "fmla   v28.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v22.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v22.8h, v2.h[4]     \n"
-
-                        "fmla   v24.8h, v23.8h, v4.h[0]     \n"
-                        "fmla   v25.8h, v23.8h, v4.h[2]     \n"
-                        "fmla   v26.8h, v23.8h, v4.h[4]     \n"
-                        "fmla   v27.8h, v23.8h, v4.h[6]     \n"
-                        "fmla   v28.8h, v23.8h, v5.h[0]     \n"
-                        "fmla   v29.8h, v23.8h, v5.h[2]     \n"
-                        "fmla   v30.8h, v23.8h, v5.h[4]     \n"
-                        "fmla   v31.8h, v23.8h, v5.h[6]     \n"
-
-                        "fmla   v24.8h, v16.8h, v4.h[1]     \n"
-                        "fmla   v25.8h, v16.8h, v4.h[3]     \n"
-                        "fmla   v26.8h, v16.8h, v4.h[5]     \n"
-                        "fmla   v27.8h, v16.8h, v4.h[7]     \n"
-                        "fmla   v28.8h, v16.8h, v5.h[1]     \n"
-                        "fmla   v29.8h, v16.8h, v5.h[3]     \n"
-                        "fmla   v30.8h, v16.8h, v5.h[5]     \n"
-                        "fmla   v31.8h, v16.8h, v5.h[7]     \n"
-
-                        "fmla   v24.8h, v17.8h, v4.h[2]     \n"
-                        "fmla   v25.8h, v17.8h, v4.h[4]     \n"
-                        "fmla   v26.8h, v17.8h, v4.h[6]     \n"
-                        "fmla   v27.8h, v17.8h, v5.h[0]     \n"
-                        "fmla   v28.8h, v17.8h, v5.h[2]     \n"
-                        "fmla   v29.8h, v17.8h, v5.h[4]     \n"
-                        "fmla   v30.8h, v17.8h, v5.h[6]     \n"
-                        "fmla   v31.8h, v17.8h, v6.h[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v24.8h, v18.8h, v4.h[3]     \n"
-                        "fmla   v25.8h, v18.8h, v4.h[5]     \n"
-                        "fmla   v26.8h, v18.8h, v4.h[7]     \n"
-                        "fmla   v27.8h, v18.8h, v5.h[1]     \n"
-                        "fmla   v28.8h, v18.8h, v5.h[3]     \n"
-                        "fmla   v29.8h, v18.8h, v5.h[5]     \n"
-                        "fmla   v30.8h, v18.8h, v5.h[7]     \n"
-                        "fmla   v31.8h, v18.8h, v6.h[1]     \n"
-
-                        "fmla   v24.8h, v19.8h, v4.h[4]     \n"
-                        "fmla   v25.8h, v19.8h, v4.h[6]     \n"
-                        "fmla   v26.8h, v19.8h, v5.h[0]     \n"
-                        "fmla   v27.8h, v19.8h, v5.h[2]     \n"
-                        "fmla   v28.8h, v19.8h, v5.h[4]     \n"
-                        "fmla   v29.8h, v19.8h, v5.h[6]     \n"
-                        "fmla   v30.8h, v19.8h, v6.h[0]     \n"
-                        "fmla   v31.8h, v19.8h, v6.h[2]     \n"
-
-                        "prfm   pldl1keep, [%3, #384]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h}, [%3] \n" // r2
-
-                        "fmla   v24.8h, v20.8h, v4.h[5]     \n"
-                        "fmla   v25.8h, v20.8h, v4.h[7]     \n"
-                        "fmla   v26.8h, v20.8h, v5.h[1]     \n"
-                        "fmla   v27.8h, v20.8h, v5.h[3]     \n"
-                        "fmla   v28.8h, v20.8h, v5.h[5]     \n"
-                        "fmla   v29.8h, v20.8h, v5.h[7]     \n"
-                        "fmla   v30.8h, v20.8h, v6.h[1]     \n"
-                        "fmla   v31.8h, v20.8h, v6.h[3]     \n"
-
-                        "fmla   v24.8h, v21.8h, v4.h[6]     \n"
-                        "fmla   v25.8h, v21.8h, v5.h[0]     \n"
-                        "fmla   v26.8h, v21.8h, v5.h[2]     \n"
-                        "fmla   v27.8h, v21.8h, v5.h[4]     \n"
-                        "fmla   v28.8h, v21.8h, v5.h[6]     \n"
-                        "fmla   v29.8h, v21.8h, v6.h[0]     \n"
-                        "fmla   v30.8h, v21.8h, v6.h[2]     \n"
-                        "fmla   v31.8h, v21.8h, v6.h[4]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v24.8h, v22.8h, v0.h[0]     \n"
-                        "fmla   v25.8h, v22.8h, v0.h[2]     \n"
-                        "fmla   v26.8h, v22.8h, v0.h[4]     \n"
-                        "fmla   v27.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v28.8h, v22.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v22.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, v22.8h, v1.h[4]     \n"
-                        "fmla   v31.8h, v22.8h, v1.h[6]     \n"
-
-                        "fmla   v24.8h, v23.8h, v0.h[1]     \n"
-                        "fmla   v25.8h, v23.8h, v0.h[3]     \n"
-                        "fmla   v26.8h, v23.8h, v0.h[5]     \n"
-                        "fmla   v27.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v28.8h, v23.8h, v1.h[1]     \n"
-                        "fmla   v29.8h, v23.8h, v1.h[3]     \n"
-                        "fmla   v30.8h, v23.8h, v1.h[5]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[7]     \n"
-
-                        "fmla   v24.8h, v16.8h, v0.h[2]     \n"
-                        "fmla   v25.8h, v16.8h, v0.h[4]     \n"
-                        "fmla   v26.8h, v16.8h, v0.h[6]     \n"
-                        "fmla   v27.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v28.8h, v16.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v16.8h, v1.h[4]     \n"
-                        "fmla   v30.8h, v16.8h, v1.h[6]     \n"
-                        "fmla   v31.8h, v16.8h, v2.h[0]     \n"
-
-                        "fmla   v24.8h, v17.8h, v0.h[3]     \n"
-                        "fmla   v25.8h, v17.8h, v0.h[5]     \n"
-                        "fmla   v26.8h, v17.8h, v0.h[7]     \n"
-                        "fmla   v27.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v28.8h, v17.8h, v1.h[3]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[5]     \n"
-                        "fmla   v30.8h, v17.8h, v1.h[7]     \n"
-                        "fmla   v31.8h, v17.8h, v2.h[1]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v24.8h, v18.8h, v0.h[4]     \n"
-                        "fmla   v25.8h, v18.8h, v0.h[6]     \n"
-                        "fmla   v26.8h, v18.8h, v1.h[0]     \n"
-                        "fmla   v27.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v28.8h, v18.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v18.8h, v1.h[6]     \n"
-                        "fmla   v30.8h, v18.8h, v2.h[0]     \n"
-                        "fmla   v31.8h, v18.8h, v2.h[2]     \n"
-
-                        "prfm   pldl1keep, [%4, #384]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h}, [%4] \n" // r3
-
-                        "fmla   v24.8h, v19.8h, v0.h[5]     \n"
-                        "fmla   v25.8h, v19.8h, v0.h[7]     \n"
-                        "fmla   v26.8h, v19.8h, v1.h[1]     \n"
-                        "fmla   v27.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v28.8h, v19.8h, v1.h[5]     \n"
-                        "fmla   v29.8h, v19.8h, v1.h[7]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[3]     \n"
-
-                        "fmla   v24.8h, v20.8h, v0.h[6]     \n"
-                        "fmla   v25.8h, v20.8h, v1.h[0]     \n"
-                        "fmla   v26.8h, v20.8h, v1.h[2]     \n"
-                        "fmla   v27.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v28.8h, v20.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v20.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v20.8h, v2.h[4]     \n"
-
-                        "fmla   v24.8h, v21.8h, v4.h[0]     \n"
-                        "fmla   v25.8h, v21.8h, v4.h[2]     \n"
-                        "fmla   v26.8h, v21.8h, v4.h[4]     \n"
-                        "fmla   v27.8h, v21.8h, v4.h[6]     \n"
-                        "fmla   v28.8h, v21.8h, v5.h[0]     \n"
-                        "fmla   v29.8h, v21.8h, v5.h[2]     \n"
-                        "fmla   v30.8h, v21.8h, v5.h[4]     \n"
-                        "fmla   v31.8h, v21.8h, v5.h[6]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v24.8h, v22.8h, v4.h[1]     \n"
-                        "fmla   v25.8h, v22.8h, v4.h[3]     \n"
-                        "fmla   v26.8h, v22.8h, v4.h[5]     \n"
-                        "fmla   v27.8h, v22.8h, v4.h[7]     \n"
-                        "fmla   v28.8h, v22.8h, v5.h[1]     \n"
-                        "fmla   v29.8h, v22.8h, v5.h[3]     \n"
-                        "fmla   v30.8h, v22.8h, v5.h[5]     \n"
-                        "fmla   v31.8h, v22.8h, v5.h[7]     \n"
-
-                        "fmla   v24.8h, v23.8h, v4.h[2]     \n"
-                        "fmla   v25.8h, v23.8h, v4.h[4]     \n"
-                        "fmla   v26.8h, v23.8h, v4.h[6]     \n"
-                        "fmla   v27.8h, v23.8h, v5.h[0]     \n"
-                        "fmla   v28.8h, v23.8h, v5.h[2]     \n"
-                        "fmla   v29.8h, v23.8h, v5.h[4]     \n"
-                        "fmla   v30.8h, v23.8h, v5.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v6.h[0]     \n"
-
-                        "fmla   v24.8h, v16.8h, v4.h[3]     \n"
-                        "fmla   v25.8h, v16.8h, v4.h[5]     \n"
-                        "fmla   v26.8h, v16.8h, v4.h[7]     \n"
-                        "fmla   v27.8h, v16.8h, v5.h[1]     \n"
-                        "fmla   v28.8h, v16.8h, v5.h[3]     \n"
-                        "fmla   v29.8h, v16.8h, v5.h[5]     \n"
-                        "fmla   v30.8h, v16.8h, v5.h[7]     \n"
-                        "fmla   v31.8h, v16.8h, v6.h[1]     \n"
-
-                        "fmla   v24.8h, v17.8h, v4.h[4]     \n"
-                        "fmla   v25.8h, v17.8h, v4.h[6]     \n"
-                        "fmla   v26.8h, v17.8h, v5.h[0]     \n"
-                        "fmla   v27.8h, v17.8h, v5.h[2]     \n"
-                        "fmla   v28.8h, v17.8h, v5.h[4]     \n"
-                        "fmla   v29.8h, v17.8h, v5.h[6]     \n"
-                        "fmla   v30.8h, v17.8h, v6.h[0]     \n"
-                        "fmla   v31.8h, v17.8h, v6.h[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v24.8h, v18.8h, v4.h[5]     \n"
-                        "fmla   v25.8h, v18.8h, v4.h[7]     \n"
-                        "fmla   v26.8h, v18.8h, v5.h[1]     \n"
-                        "fmla   v27.8h, v18.8h, v5.h[3]     \n"
-                        "fmla   v28.8h, v18.8h, v5.h[5]     \n"
-                        "fmla   v29.8h, v18.8h, v5.h[7]     \n"
-                        "fmla   v30.8h, v18.8h, v6.h[1]     \n"
-                        "fmla   v31.8h, v18.8h, v6.h[3]     \n"
-
-                        "prfm   pldl1keep, [%5, #384]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h}, [%5] \n" // r4
-
-                        "fmla   v24.8h, v19.8h, v4.h[6]     \n"
-                        "fmla   v25.8h, v19.8h, v5.h[0]     \n"
-                        "fmla   v26.8h, v19.8h, v5.h[2]     \n"
-                        "fmla   v27.8h, v19.8h, v5.h[4]     \n"
-                        "fmla   v28.8h, v19.8h, v5.h[6]     \n"
-                        "fmla   v29.8h, v19.8h, v6.h[0]     \n"
-                        "fmla   v30.8h, v19.8h, v6.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v6.h[4]     \n"
-
-                        "fmla   v24.8h, v20.8h, v0.h[0]     \n"
-                        "fmla   v25.8h, v20.8h, v0.h[2]     \n"
-                        "fmla   v26.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v27.8h, v20.8h, v0.h[6]     \n"
-                        "fmla   v28.8h, v20.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v20.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v1.h[6]     \n"
-
-                        "fmla   v24.8h, v21.8h, v0.h[1]     \n"
-                        "fmla   v25.8h, v21.8h, v0.h[3]     \n"
-                        "fmla   v26.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v27.8h, v21.8h, v0.h[7]     \n"
-                        "fmla   v28.8h, v21.8h, v1.h[1]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[3]     \n"
-                        "fmla   v30.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v1.h[7]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v24.8h, v22.8h, v0.h[2]     \n"
-                        "fmla   v25.8h, v22.8h, v0.h[4]     \n"
-                        "fmla   v26.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v27.8h, v22.8h, v1.h[0]     \n"
-                        "fmla   v28.8h, v22.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v22.8h, v1.h[4]     \n"
-                        "fmla   v30.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v2.h[0]     \n"
-
-                        "fmla   v24.8h, v23.8h, v0.h[3]     \n"
-                        "fmla   v25.8h, v23.8h, v0.h[5]     \n"
-                        "fmla   v26.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v27.8h, v23.8h, v1.h[1]     \n"
-                        "fmla   v28.8h, v23.8h, v1.h[3]     \n"
-                        "fmla   v29.8h, v23.8h, v1.h[5]     \n"
-                        "fmla   v30.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #384]       \n"
-                        "ld1    {v4.8h, v5.8h, v6.8h}, [%6] \n" // r5
-
-                        "fmla   v24.8h, v16.8h, v0.h[4]     \n"
-                        "fmla   v25.8h, v16.8h, v0.h[6]     \n"
-                        "fmla   v26.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v27.8h, v16.8h, v1.h[2]     \n"
-                        "fmla   v28.8h, v16.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v16.8h, v1.h[6]     \n"
-                        "fmla   v30.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v2.h[2]     \n"
-
-                        "fmla   v24.8h, v17.8h, v0.h[5]     \n"
-                        "fmla   v25.8h, v17.8h, v0.h[7]     \n"
-                        "fmla   v26.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v27.8h, v17.8h, v1.h[3]     \n"
-                        "fmla   v28.8h, v17.8h, v1.h[5]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[7]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v2.h[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v24.8h, v18.8h, v0.h[6]     \n"
-                        "fmla   v25.8h, v18.8h, v1.h[0]     \n"
-                        "fmla   v26.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v27.8h, v18.8h, v1.h[4]     \n"
-                        "fmla   v28.8h, v18.8h, v1.h[6]     \n"
-                        "fmla   v30.8h, v18.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[0]     \n"
-                        "fmla   v31.8h, v18.8h, v2.h[4]     \n"
-
-                        "fmla   v24.8h, v19.8h, v4.h[0]     \n"
-                        "fmla   v25.8h, v19.8h, v4.h[2]     \n"
-                        "fmla   v26.8h, v19.8h, v4.h[4]     \n"
-                        "fmla   v27.8h, v19.8h, v4.h[6]     \n"
-                        "fmla   v28.8h, v19.8h, v5.h[0]     \n"
-                        "fmla   v29.8h, v19.8h, v5.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v5.h[4]     \n"
-                        "fmla   v31.8h, v19.8h, v5.h[6]     \n"
-
-                        "fmla   v24.8h, v20.8h, v4.h[1]     \n"
-                        "fmla   v25.8h, v20.8h, v4.h[3]     \n"
-                        "fmla   v26.8h, v20.8h, v4.h[5]     \n"
-                        "fmla   v27.8h, v20.8h, v4.h[7]     \n"
-                        "fmla   v28.8h, v20.8h, v5.h[1]     \n"
-                        "fmla   v29.8h, v20.8h, v5.h[3]     \n"
-                        "fmla   v30.8h, v20.8h, v5.h[5]     \n"
-                        "fmla   v31.8h, v20.8h, v5.h[7]     \n"
-
-                        "fmla   v24.8h, v21.8h, v4.h[2]     \n"
-                        "fmla   v25.8h, v21.8h, v4.h[4]     \n"
-                        "fmla   v26.8h, v21.8h, v4.h[6]     \n"
-                        "fmla   v27.8h, v21.8h, v5.h[0]     \n"
-                        "fmla   v28.8h, v21.8h, v5.h[2]     \n"
-                        "fmla   v29.8h, v21.8h, v5.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v5.h[6]     \n"
-                        "fmla   v31.8h, v21.8h, v6.h[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v24.8h, v22.8h, v4.h[3]     \n"
-                        "fmla   v25.8h, v22.8h, v4.h[5]     \n"
-                        "fmla   v26.8h, v22.8h, v4.h[7]     \n"
-                        "fmla   v27.8h, v22.8h, v5.h[1]     \n"
-                        "fmla   v28.8h, v22.8h, v5.h[3]     \n"
-                        "fmla   v29.8h, v22.8h, v5.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v5.h[7]     \n"
-                        "fmla   v31.8h, v22.8h, v6.h[1]     \n"
-
-                        "fmla   v24.8h, v23.8h, v4.h[4]     \n"
-                        "fmla   v25.8h, v23.8h, v4.h[6]     \n"
-                        "fmla   v26.8h, v23.8h, v5.h[0]     \n"
-                        "fmla   v27.8h, v23.8h, v5.h[2]     \n"
-                        "fmla   v28.8h, v23.8h, v5.h[4]     \n"
-                        "fmla   v29.8h, v23.8h, v5.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v6.h[0]     \n"
-                        "fmla   v31.8h, v23.8h, v6.h[2]     \n"
-
-                        "prfm   pldl1keep, [%7, #384]       \n"
-                        "ld1    {v0.8h, v1.8h, v2.8h}, [%7] \n" // r6
-
-                        "fmla   v24.8h, v16.8h, v4.h[5]     \n"
-                        "fmla   v25.8h, v16.8h, v4.h[7]     \n"
-                        "fmla   v26.8h, v16.8h, v5.h[1]     \n"
-                        "fmla   v27.8h, v16.8h, v5.h[3]     \n"
-                        "fmla   v28.8h, v16.8h, v5.h[5]     \n"
-                        "fmla   v29.8h, v16.8h, v5.h[7]     \n"
-                        "fmla   v30.8h, v16.8h, v6.h[1]     \n"
-                        "fmla   v31.8h, v16.8h, v6.h[3]     \n"
-
-                        "fmla   v24.8h, v17.8h, v4.h[6]     \n"
-                        "fmla   v25.8h, v17.8h, v5.h[0]     \n"
-                        "fmla   v26.8h, v17.8h, v5.h[2]     \n"
-                        "fmla   v27.8h, v17.8h, v5.h[4]     \n"
-                        "fmla   v28.8h, v17.8h, v5.h[6]     \n"
-                        "fmla   v29.8h, v17.8h, v6.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v6.h[2]     \n"
-                        "fmla   v31.8h, v17.8h, v6.h[4]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v24.8h, v18.8h, v0.h[0]     \n"
-                        "fmla   v25.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v26.8h, v18.8h, v0.h[4]     \n"
-                        "fmla   v27.8h, v18.8h, v0.h[6]     \n"
-                        "fmla   v28.8h, v18.8h, v1.h[0]     \n"
-                        "fmla   v29.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v1.h[4]     \n"
-                        "fmla   v31.8h, v18.8h, v1.h[6]     \n"
-
-                        "fmla   v24.8h, v19.8h, v0.h[1]     \n"
-                        "fmla   v25.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v26.8h, v19.8h, v0.h[5]     \n"
-                        "fmla   v27.8h, v19.8h, v0.h[7]     \n"
-                        "fmla   v28.8h, v19.8h, v1.h[1]     \n"
-                        "fmla   v29.8h, v19.8h, v1.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v1.h[5]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[7]     \n"
-
-                        "fmla   v24.8h, v20.8h, v0.h[2]     \n"
-                        "fmla   v25.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v26.8h, v20.8h, v0.h[6]     \n"
-                        "fmla   v27.8h, v20.8h, v1.h[0]     \n"
-                        "fmla   v28.8h, v20.8h, v1.h[2]     \n"
-                        "fmla   v29.8h, v20.8h, v1.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v1.h[6]     \n"
-                        "fmla   v31.8h, v20.8h, v2.h[0]     \n"
-
-                        "add    %1, %1, #32                 \n"
-
-                        "fmla   v24.8h, v21.8h, v0.h[3]     \n"
-                        "fmla   v25.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v26.8h, v21.8h, v0.h[7]     \n"
-                        "fmla   v27.8h, v21.8h, v1.h[1]     \n"
-
-                        "add    %2, %2, #32                 \n"
-
-                        "fmla   v28.8h, v21.8h, v1.h[3]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v1.h[7]     \n"
-                        "fmla   v31.8h, v21.8h, v2.h[1]     \n"
-
-                        "prfm   pldl1keep, [%8, #128]       \n"
-                        "ld1    {v16.8h}, [%8]              \n"
-
-                        "fmla   v24.8h, v22.8h, v0.h[4]     \n"
-                        "fmla   v25.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v26.8h, v22.8h, v1.h[0]     \n"
-                        "fmla   v27.8h, v22.8h, v1.h[2]     \n"
-
-                        "add    %3, %3, #32                 \n"
-
-                        "fmla   v28.8h, v22.8h, v1.h[4]     \n"
-                        "fmla   v29.8h, v22.8h, v1.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v2.h[0]     \n"
-                        "fmla   v31.8h, v22.8h, v2.h[2]     \n"
-
-                        "add    %4, %4, #32                 \n"
-
-                        "fmla   v24.8h, v23.8h, v0.h[5]     \n"
-                        "fmla   v25.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v26.8h, v23.8h, v1.h[1]     \n"
-                        "fmla   v27.8h, v23.8h, v1.h[3]     \n"
-
-                        "add    %5, %5, #32                 \n"
-
-                        "fmla   v28.8h, v23.8h, v1.h[5]     \n"
-                        "fmla   v29.8h, v23.8h, v1.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[1]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[3]     \n"
-
-                        "add    %6, %6, #32                 \n"
-
-                        "fmla   v24.8h, v16.8h, v0.h[6]     \n"
-                        "fmla   v25.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v26.8h, v16.8h, v1.h[2]     \n"
-                        "fmla   v27.8h, v16.8h, v1.h[4]     \n"
-
-                        "add    %7, %7, #32                 \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[6]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v2.h[2]     \n"
-                        "fmla   v31.8h, v16.8h, v2.h[4]     \n"
-
-                        "sub    %8, %8, #768                \n" // kptr -= 48 * 8;
-
-                        "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                        "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "v0", "v1", "v2", "v4", "v5", "v6", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-                }
-                for (; j + 3 < outw; j += 4)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%0, #512]       \n"
-                        "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0] \n" // sum0
-
-                        "prfm   pldl1keep, [%1, #256]       \n"
-                        "ld1    {v0.8h, v1.8h}, [%1]        \n" // r0
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v16.8h, v0.h[2]     \n"
-                        "fmla   v30.8h, v16.8h, v0.h[4]     \n"
-                        "fmla   v31.8h, v16.8h, v0.h[6]     \n"
-
-                        "fmla   v28.8h, v17.8h, v0.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v0.h[3]     \n"
-                        "fmla   v30.8h, v17.8h, v0.h[5]     \n"
-                        "fmla   v31.8h, v17.8h, v0.h[7]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v18.8h, v0.h[4]     \n"
-                        "fmla   v30.8h, v18.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v18.8h, v1.h[0]     \n"
-
-                        "fmla   v28.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v29.8h, v19.8h, v0.h[5]     \n"
-                        "fmla   v30.8h, v19.8h, v0.h[7]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%2, #256]       \n"
-                        "ld1    {v2.8h, v3.8h}, [%2]        \n" // r1
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v20.8h, v0.h[6]     \n"
-                        "fmla   v30.8h, v20.8h, v1.h[0]     \n"
-                        "fmla   v31.8h, v20.8h, v1.h[2]     \n"
-
-                        "fmla   v28.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[7]     \n"
-                        "fmla   v30.8h, v21.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, v21.8h, v1.h[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v22.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, v22.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, v22.8h, v1.h[4]     \n"
-
-                        "fmla   v28.8h, v23.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v23.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[4]     \n"
-                        "fmla   v31.8h, v23.8h, v2.h[6]     \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[1]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[3]     \n"
-                        "fmla   v30.8h, v16.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v16.8h, v2.h[7]     \n"
-
-                        "fmla   v28.8h, v17.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v17.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v2.h[3]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v18.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v18.8h, v3.h[1]     \n"
-
-                        "fmla   v28.8h, v19.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v19.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v19.8h, v3.h[0]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[2]     \n"
-
-                        "prfm   pldl1keep, [%3, #256]       \n"
-                        "ld1    {v0.8h, v1.8h}, [%3]        \n" // r2
-
-                        "fmla   v28.8h, v20.8h, v2.h[5]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[7]     \n"
-                        "fmla   v30.8h, v20.8h, v3.h[1]     \n"
-                        "fmla   v31.8h, v20.8h, v3.h[3]     \n"
-
-                        "fmla   v28.8h, v21.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v21.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v21.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[4]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v22.8h, v0.h[2]     \n"
-                        "fmla   v30.8h, v22.8h, v0.h[4]     \n"
-                        "fmla   v31.8h, v22.8h, v0.h[6]     \n"
-
-                        "fmla   v28.8h, v23.8h, v0.h[1]     \n"
-                        "fmla   v29.8h, v23.8h, v0.h[3]     \n"
-                        "fmla   v30.8h, v23.8h, v0.h[5]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[7]     \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v16.8h, v0.h[4]     \n"
-                        "fmla   v30.8h, v16.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v16.8h, v1.h[0]     \n"
-
-                        "fmla   v28.8h, v17.8h, v0.h[3]     \n"
-                        "fmla   v29.8h, v17.8h, v0.h[5]     \n"
-                        "fmla   v30.8h, v17.8h, v0.h[7]     \n"
-                        "fmla   v31.8h, v17.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v18.8h, v0.h[6]     \n"
-                        "fmla   v30.8h, v18.8h, v1.h[0]     \n"
-                        "fmla   v31.8h, v18.8h, v1.h[2]     \n"
-
-                        "prfm   pldl1keep, [%4, #256]       \n"
-                        "ld1    {v2.8h, v3.8h}, [%4]        \n" // r3
-
-                        "fmla   v28.8h, v19.8h, v0.h[5]     \n"
-                        "fmla   v29.8h, v19.8h, v0.h[7]     \n"
-                        "fmla   v30.8h, v19.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[3]     \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v20.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, v20.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, v20.8h, v1.h[4]     \n"
-
-                        "fmla   v28.8h, v21.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[4]     \n"
-                        "fmla   v31.8h, v21.8h, v2.h[6]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v2.h[1]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[3]     \n"
-                        "fmla   v30.8h, v22.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v22.8h, v2.h[7]     \n"
-
-                        "fmla   v28.8h, v23.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v23.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v23.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[0]     \n"
-
-                        "fmla   v28.8h, v16.8h, v2.h[3]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v16.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v16.8h, v3.h[1]     \n"
-
-                        "prfm   pldl1keep, [%5, #256]       \n"
-                        "ld1    {v0.8h, v1.8h}, [%5]        \n" // r4
-
-                        "fmla   v28.8h, v17.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v17.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v17.8h, v3.h[0]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v2.h[5]     \n"
-                        "fmla   v29.8h, v18.8h, v2.h[7]     \n"
-                        "fmla   v30.8h, v18.8h, v3.h[1]     \n"
-                        "fmla   v31.8h, v18.8h, v3.h[3]     \n"
-
-                        "fmla   v28.8h, v19.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v19.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v19.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v3.h[4]     \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v20.8h, v0.h[2]     \n"
-                        "fmla   v30.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v31.8h, v20.8h, v0.h[6]     \n"
-
-                        "fmla   v28.8h, v21.8h, v0.h[1]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[3]     \n"
-                        "fmla   v30.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v31.8h, v21.8h, v0.h[7]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v22.8h, v0.h[4]     \n"
-                        "fmla   v30.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v22.8h, v1.h[0]     \n"
-
-                        "fmla   v28.8h, v23.8h, v0.h[3]     \n"
-                        "fmla   v29.8h, v23.8h, v0.h[5]     \n"
-                        "fmla   v30.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%6, #256]       \n"
-                        "ld1    {v2.8h, v3.8h}, [%6]        \n" // r5
-
-                        "fmla   v28.8h, v16.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v16.8h, v0.h[6]     \n"
-                        "fmla   v30.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v31.8h, v16.8h, v1.h[2]     \n"
-
-                        "fmla   v28.8h, v17.8h, v0.h[5]     \n"
-                        "fmla   v29.8h, v17.8h, v0.h[7]     \n"
-                        "fmla   v30.8h, v17.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, v17.8h, v1.h[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v18.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, v18.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, v18.8h, v1.h[4]     \n"
-
-                        "fmla   v28.8h, v19.8h, v2.h[0]     \n"
-                        "fmla   v29.8h, v19.8h, v2.h[2]     \n"
-                        "fmla   v30.8h, v19.8h, v2.h[4]     \n"
-                        "fmla   v31.8h, v19.8h, v2.h[6]     \n"
-
-                        "fmla   v28.8h, v20.8h, v2.h[1]     \n"
-                        "fmla   v29.8h, v20.8h, v2.h[3]     \n"
-                        "fmla   v30.8h, v20.8h, v2.h[5]     \n"
-                        "fmla   v31.8h, v20.8h, v2.h[7]     \n"
-
-                        "fmla   v28.8h, v21.8h, v2.h[2]     \n"
-                        "fmla   v29.8h, v21.8h, v2.h[4]     \n"
-                        "fmla   v30.8h, v21.8h, v2.h[6]     \n"
-                        "fmla   v31.8h, v21.8h, v3.h[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v28.8h, v22.8h, v2.h[3]     \n"
-                        "fmla   v29.8h, v22.8h, v2.h[5]     \n"
-                        "fmla   v30.8h, v22.8h, v2.h[7]     \n"
-                        "fmla   v31.8h, v22.8h, v3.h[1]     \n"
-
-                        "add    %1, %1, #16                 \n"
-
-                        "fmla   v28.8h, v23.8h, v2.h[4]     \n"
-                        "fmla   v29.8h, v23.8h, v2.h[6]     \n"
-                        "fmla   v30.8h, v23.8h, v3.h[0]     \n"
-                        "fmla   v31.8h, v23.8h, v3.h[2]     \n"
-
-                        "prfm   pldl1keep, [%7, #256]       \n"
-                        "ld1    {v0.8h, v1.8h}, [%7]        \n" // r6
-
-                        "fmla   v28.8h, v16.8h, v2.h[5]     \n"
-                        "fmla   v29.8h, v16.8h, v2.h[7]     \n"
-                        "fmla   v30.8h, v16.8h, v3.h[1]     \n"
-                        "fmla   v31.8h, v16.8h, v3.h[3]     \n"
-
-                        "add    %2, %2, #16                 \n"
-
-                        "fmla   v28.8h, v17.8h, v2.h[6]     \n"
-                        "fmla   v29.8h, v17.8h, v3.h[0]     \n"
-                        "fmla   v30.8h, v17.8h, v3.h[2]     \n"
-                        "fmla   v31.8h, v17.8h, v3.h[4]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v28.8h, v18.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v30.8h, v18.8h, v0.h[4]     \n"
-                        "fmla   v31.8h, v18.8h, v0.h[6]     \n"
-
-                        "add    %3, %3, #16                 \n"
-
-                        "fmla   v28.8h, v19.8h, v0.h[1]     \n"
-                        "fmla   v29.8h, v19.8h, v0.h[3]     \n"
-                        "fmla   v30.8h, v19.8h, v0.h[5]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[7]     \n"
-
-                        "add    %4, %4, #16                 \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v30.8h, v20.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v20.8h, v1.h[0]     \n"
-
-                        "add    %5, %5, #16                 \n"
-
-                        "fmla   v28.8h, v21.8h, v0.h[3]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[5]     \n"
-                        "fmla   v30.8h, v21.8h, v0.h[7]     \n"
-                        "fmla   v31.8h, v21.8h, v1.h[1]     \n"
-
-                        "prfm   pldl1keep, [%8, #128]       \n"
-                        "ld1    {v16.8h}, [%8]              \n"
-
-                        "fmla   v28.8h, v22.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v30.8h, v22.8h, v1.h[0]     \n"
-                        "fmla   v31.8h, v22.8h, v1.h[2]     \n"
-
-                        "add    %6, %6, #16                 \n"
-
-                        "fmla   v28.8h, v23.8h, v0.h[5]     \n"
-                        "fmla   v29.8h, v23.8h, v0.h[7]     \n"
-                        "fmla   v30.8h, v23.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[3]     \n"
-
-                        "add    %7, %7, #16                 \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v16.8h, v1.h[0]     \n"
-                        "fmla   v30.8h, v16.8h, v1.h[2]     \n"
-                        "fmla   v31.8h, v16.8h, v1.h[4]     \n"
-
-                        "sub    %8, %8, #768                \n" // kptr -= 48 * 8;
-
-                        "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-                }
-                for (; j < outw; j++)
-                {
-                    asm volatile(
-                        "prfm   pldl1keep, [%1, #128]       \n"
-                        "ld1    {v0.8h}, [%1]               \n" // r0
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "prfm   pldl1keep, [%0, #128]       \n"
-                        "ld1    {v31.8h}, [%0]              \n" // sum0
-
-                        "fmul   v28.8h, v16.8h, v0.h[0]     \n"
-                        "fmul   v29.8h, v17.8h, v0.h[1]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmul   v30.8h, v18.8h, v0.h[2]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[3]     \n"
-
-                        "prfm   pldl1keep, [%2, #128]       \n"
-                        "ld1    {v1.8h}, [%2]               \n" // r1
-
-                        "fmla   v28.8h, v20.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[5]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[0]     \n"
-
-                        "fmla   v28.8h, v16.8h, v1.h[1]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v1.h[3]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[4]     \n"
-
-                        "prfm   pldl1keep, [%3, #128]       \n"
-                        "ld1    {v0.8h}, [%3]               \n" // r2
-
-                        "fmla   v28.8h, v20.8h, v1.h[5]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[6]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v0.h[0]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[1]     \n"
-
-                        "prfm   pldl1keep, [%4, #128]       \n"
-                        "ld1    {v1.8h}, [%4]               \n" // r3
-
-                        "fmla   v28.8h, v16.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v17.8h, v0.h[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v0.h[4]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[5]     \n"
-
-                        "add    %1, %1, #4                  \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[6]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[0]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v1.h[1]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[2]     \n"
-
-                        "prfm   pldl1keep, [%5, #128]       \n"
-                        "ld1    {v0.8h}, [%5]               \n" // r4
-
-                        "fmla   v28.8h, v16.8h, v1.h[3]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[4]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v1.h[5]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[6]     \n"
-
-                        "add    %2, %2, #4                  \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[0]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[1]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v0.h[2]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[3]     \n"
-
-                        "prfm   pldl1keep, [%6, #128]       \n"
-                        "ld1    {v1.8h}, [%6]               \n" // r5
-
-                        "fmla   v28.8h, v16.8h, v0.h[4]     \n"
-                        "fmla   v29.8h, v17.8h, v0.h[5]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v0.h[6]     \n"
-                        "fmla   v31.8h, v19.8h, v1.h[0]     \n"
-
-                        "add    %3, %3, #4                  \n"
-
-                        "fmla   v28.8h, v20.8h, v1.h[1]     \n"
-                        "fmla   v29.8h, v21.8h, v1.h[2]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-
-                        "fmla   v30.8h, v22.8h, v1.h[3]     \n"
-                        "fmla   v31.8h, v23.8h, v1.h[4]     \n"
-
-                        "prfm   pldl1keep, [%7, #128]       \n"
-                        "ld1    {v0.8h}, [%7]               \n" // r6
-
-                        "fmla   v28.8h, v16.8h, v1.h[5]     \n"
-                        "fmla   v29.8h, v17.8h, v1.h[6]     \n"
-
-                        "prfm   pldl1keep, [%8, #512]       \n"
-                        "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-
-                        "fmla   v30.8h, v18.8h, v0.h[0]     \n"
-                        "fmla   v31.8h, v19.8h, v0.h[1]     \n"
-
-                        "add    %4, %4, #4                  \n"
-
-                        "fmla   v28.8h, v20.8h, v0.h[2]     \n"
-                        "fmla   v29.8h, v21.8h, v0.h[3]     \n"
-
-                        "prfm   pldl1keep, [%8, #128]       \n"
-                        "ld1    {v16.8h}, [%8]              \n"
-
-                        "fmla   v30.8h, v22.8h, v0.h[4]     \n"
-                        "fmla   v31.8h, v23.8h, v0.h[5]     \n"
-
-                        "add    %5, %5, #4                  \n"
-
-                        "fmla   v28.8h, v16.8h, v0.h[6]     \n"
-
-                        "add    %6, %6, #4                  \n"
-
-                        "fadd   v29.8h, v29.8h, v30.8h      \n"
-                        "fadd   v31.8h, v31.8h, v28.8h      \n"
-
-                        "add    %7, %7, #4                  \n"
-
-                        "fadd   v29.8h, v29.8h, v31.8h      \n"
-
-                        "sub    %8, %8, #768                \n" // kptr -= 48 * 8;
-
-                        "st1    {v29.8h}, [%0], #16         \n"
-
-                        : "=r"(outptr0), // %0
-                        "=r"(r0),      // %1
-                        "=r"(r1),      // %2
-                        "=r"(r2),      // %3
-                        "=r"(r3),      // %4
-                        "=r"(r4),      // %5
-                        "=r"(r5),      // %6
-                        "=r"(r6),      // %7
-                        "=r"(kptr)     // %8
-                        : "0"(outptr0),
-                        "1"(r0),
-                        "2"(r1),
-                        "3"(r2),
-                        "4"(r3),
-                        "5"(r4),
-                        "6"(r5),
-                        "7"(r6),
-                        "8"(kptr)
-                        : "memory", "v0", "v1", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-                r4 += tailstep;
-                r5 += tailstep;
-                r6 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_arm.h
deleted file mode 100644
index e25f509c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_arm.h
+++ /dev/null
@@ -1,76 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTION_ARM_H
-#define LAYER_CONVOLUTION_ARM_H
-
-#include "convolution.h"
-
-namespace ncnn {
-
-class Convolution_arm : public Convolution
-{
-public:
-    Convolution_arm();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int create_pipeline_bf16s(const Option& opt);
-    int forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_INT8
-    int create_pipeline_int8_arm(const Option& opt);
-    int forward_int8_arm(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-    int forwardDilation_arm(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    Layer* activation;
-
-    int nT;
-
-    Mat weight_data_tm;
-    Mat weight_3x3s2_data;
-
-    Mat weight_sgemm_data;
-    Mat weight_winograd23_data;
-    Mat weight_winograd43_data;
-    Mat weight_winograd63_data;
-
-    // forwardDilation
-    Layer* convolution_dilation1;
-
-    // fp16
-    Mat bias_data_fp16;
-
-#if NCNN_INT8
-    Mat scale_in_data;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTION_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm.h
deleted file mode 100644
index 45651dda..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm.h
+++ /dev/null
@@ -1,6904 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_im2col_pack_A_tile(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
-{
-    // A = (pa, maxk, inch/pa), outch
-    const int A_hstep = A.w;
-
-    float* pp = AT;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        const float* p0 = (const float*)A + (i + ii) * A_hstep + k;
-        const float* p1 = (const float*)A + (i + ii + 1) * A_hstep + k;
-        const float* p2 = (const float*)A + (i + ii + 2) * A_hstep + k;
-        const float* p3 = (const float*)A + (i + ii + 3) * A_hstep + k;
-        const float* p4 = (const float*)A + (i + ii + 4) * A_hstep + k;
-        const float* p5 = (const float*)A + (i + ii + 5) * A_hstep + k;
-        const float* p6 = (const float*)A + (i + ii + 6) * A_hstep + k;
-        const float* p7 = (const float*)A + (i + ii + 7) * A_hstep + k;
-
-        int kk = 0;
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            float32x4_t _r0l = vld1q_f32(p0);
-            float32x4_t _r0h = vld1q_f32(p0 + 4);
-            float32x4_t _r1l = vld1q_f32(p1);
-            float32x4_t _r1h = vld1q_f32(p1 + 4);
-            float32x4_t _r2l = vld1q_f32(p2);
-            float32x4_t _r2h = vld1q_f32(p2 + 4);
-            float32x4_t _r3l = vld1q_f32(p3);
-            float32x4_t _r3h = vld1q_f32(p3 + 4);
-            float32x4_t _r4l = vld1q_f32(p4);
-            float32x4_t _r4h = vld1q_f32(p4 + 4);
-            float32x4_t _r5l = vld1q_f32(p5);
-            float32x4_t _r5h = vld1q_f32(p5 + 4);
-            float32x4_t _r6l = vld1q_f32(p6);
-            float32x4_t _r6h = vld1q_f32(p6 + 4);
-            float32x4_t _r7l = vld1q_f32(p7);
-            float32x4_t _r7h = vld1q_f32(p7 + 4);
-            transpose8x8_ps(_r0l, _r0h, _r1l, _r1h, _r2l, _r2h, _r3l, _r3h, _r4l, _r4h, _r5l, _r5h, _r6l, _r6h, _r7l, _r7h);
-            vst1q_f32(pp, _r0l);
-            vst1q_f32(pp + 4, _r0h);
-            vst1q_f32(pp + 8, _r1l);
-            vst1q_f32(pp + 12, _r1h);
-            vst1q_f32(pp + 8 * 2, _r2l);
-            vst1q_f32(pp + 8 * 2 + 4, _r2h);
-            vst1q_f32(pp + 8 * 3, _r3l);
-            vst1q_f32(pp + 8 * 3 + 4, _r3h);
-            vst1q_f32(pp + 8 * 4, _r4l);
-            vst1q_f32(pp + 8 * 4 + 4, _r4h);
-            vst1q_f32(pp + 8 * 5, _r5l);
-            vst1q_f32(pp + 8 * 5 + 4, _r5h);
-            vst1q_f32(pp + 8 * 6, _r6l);
-            vst1q_f32(pp + 8 * 6 + 4, _r6h);
-            vst1q_f32(pp + 8 * 7, _r7l);
-            vst1q_f32(pp + 8 * 7 + 4, _r7h);
-            pp += 64;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-            p4 += 8;
-            p5 += 8;
-            p6 += 8;
-            p7 += 8;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp[1] = p1[0];
-            pp[2] = p2[0];
-            pp[3] = p3[0];
-            pp[4] = p4[0];
-            pp[5] = p5[0];
-            pp[6] = p6[0];
-            pp[7] = p7[0];
-            pp += 8;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-            p4++;
-            p5++;
-            p6++;
-            p7++;
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        const float* p0 = (const float*)A + (i + ii) * A_hstep + k;
-        const float* p1 = (const float*)A + (i + ii + 1) * A_hstep + k;
-        const float* p2 = (const float*)A + (i + ii + 2) * A_hstep + k;
-        const float* p3 = (const float*)A + (i + ii + 3) * A_hstep + k;
-
-        int kk = 0;
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            float32x4x4_t _r0123;
-            _r0123.val[0] = vld1q_f32(p0);
-            _r0123.val[1] = vld1q_f32(p1);
-            _r0123.val[2] = vld1q_f32(p2);
-            _r0123.val[3] = vld1q_f32(p3);
-            vst4q_f32(pp, _r0123);
-            pp += 16;
-            p0 += 4;
-            p1 += 4;
-            p2 += 4;
-            p3 += 4;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp[1] = p1[0];
-            pp[2] = p2[0];
-            pp[3] = p3[0];
-            pp += 4;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        const float* p0 = (const float*)A + (i + ii) * A_hstep + k;
-        const float* p1 = (const float*)A + (i + ii + 1) * A_hstep + k;
-
-        int kk = 0;
-#if __ARM_NEON
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            float32x4x2_t _r01;
-            _r01.val[0] = vld1q_f32(p0);
-            _r01.val[1] = vld1q_f32(p1);
-            vst2q_f32(pp, _r01);
-            pp += 8;
-            p0 += 4;
-            p1 += 4;
-        }
-#endif // __ARM_NEON
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp[1] = p1[0];
-            pp += 2;
-            p0++;
-            p1++;
-        }
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        const float* p0 = (const float*)A + (i + ii) * A_hstep + k;
-
-        int kk = 0;
-#if __ARM_NEON
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            vst1q_f32(pp, vld1q_f32(p0));
-            pp += 4;
-            p0 += 4;
-        }
-#endif // __ARM_NEON
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp += 1;
-            p0++;
-        }
-    }
-}
-
-static void convolution_gemm_transB_packed_tile(const Mat& AT_tile, const Mat& BT_tile, const Mat& CT_tile, Mat& topT_tile, Mat& top_blob, int i, int max_ii, int j, int max_jj, int k, int max_kk, bool k_end, int use_a53_a55_optimized_kernel)
-{
-    // NCNN_LOGE("convolution_gemm_transB_packed_tile %d %d %d %d %d %d", i, max_ii, j, max_jj, k, max_kk);
-
-    const int out_elempack = top_blob.elempack;
-    const int out_hstep = (int)top_blob.cstep;
-
-    const float* pAT = AT_tile;
-    const float* pBT = BT_tile;
-    const float* pC = CT_tile;
-
-    float* outptr = topT_tile;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        float* outptr0 = (float*)top_blob + (i + ii) * out_hstep + j * out_elempack;
-
-        const float* pB = pBT;
-
-        if (pC)
-        {
-            pC = (const float*)CT_tile + i + ii;
-        }
-
-        int jj = 0;
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            if (use_a53_a55_optimized_kernel && cpu_support_arm_asimdhp())
-            {
-                // a55
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #320                \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "add    x4, %8, #16                 \n"
-                    "ld1    {v8.4s}, [%8]               \n"
-                    "ld1    {v20.4s}, [x4]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v8.16b, v8.16b, v8.16b      \n"
-                    "eor    v20.16b, v20.16b, v20.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v9.16b, v8.16b              \n"
-                    "mov    v10.16b, v8.16b             \n"
-                    "mov    v11.16b, v8.16b             \n"
-                    "mov    v12.16b, v8.16b             \n"
-                    "mov    v13.16b, v8.16b             \n"
-                    "mov    v14.16b, v8.16b             \n"
-                    "mov    v15.16b, v8.16b             \n"
-                    "mov    v16.16b, v8.16b             \n"
-                    "mov    v17.16b, v8.16b             \n"
-                    "mov    v18.16b, v8.16b             \n"
-                    "mov    v19.16b, v8.16b             \n"
-
-                    "mov    v21.16b, v20.16b            \n"
-                    "mov    v22.16b, v20.16b            \n"
-                    "mov    v23.16b, v20.16b            \n"
-                    "mov    v24.16b, v20.16b            \n"
-                    "mov    v25.16b, v20.16b            \n"
-                    "mov    v26.16b, v20.16b            \n"
-                    "mov    v27.16b, v20.16b            \n"
-                    "mov    v28.16b, v20.16b            \n"
-                    "mov    v29.16b, v20.16b            \n"
-                    "mov    v30.16b, v20.16b            \n"
-                    "mov    v31.16b, v20.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.4s}, [%1], #16          \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.4s}, [%2], #16          \n"
-
-                    "ldr    d5, [%1], #8                \n"
-                    "ldr    x25, [%1], #8               \n"
-
-                    ".align 4                           \n"
-                    "4:                                 \n"
-                    "ldr    d1, [%2], #8                \n"
-                    "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                    "ldr    x21, [%2], #8               \n"
-                    "fmla   v9.4s, v4.4s, v0.s[1]       \n"
-                    "ins    v5.d[1], x25                \n"
-                    "fmla   v10.4s, v4.4s, v0.s[2]      \n"
-                    "ldr    d2, [%2], #8                \n"
-                    "fmla   v11.4s, v4.4s, v0.s[3]      \n"
-                    "ldr    x22, [%2], #8               \n"
-                    "fmla   v20.4s, v5.4s, v0.s[0]      \n"
-                    "ldr    d6, [%1], #8                \n"
-                    "fmla   v21.4s, v5.4s, v0.s[1]      \n"
-                    "ins    v1.d[1], x21                \n"
-                    "fmla   v22.4s, v5.4s, v0.s[2]      \n"
-                    "ldr    x26, [%1], #8               \n"
-                    "fmla   v23.4s, v5.4s, v0.s[3]      \n"
-                    "ldr    d3, [%2], #8                \n"
-                    "fmla   v12.4s, v4.4s, v1.s[0]      \n"
-                    "ldr    x23, [%2], #8               \n"
-                    "fmla   v13.4s, v4.4s, v1.s[1]      \n"
-                    "ldr    d7, [%1], #8                \n"
-                    "fmla   v14.4s, v4.4s, v1.s[2]      \n"
-                    "ldr    x27, [%1], #8               \n"
-                    "fmla   v15.4s, v4.4s, v1.s[3]      \n"
-                    "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                    "fmla   v24.4s, v5.4s, v1.s[0]      \n"
-                    "ldr    d0, [%2], #8                \n"
-                    "fmla   v25.4s, v5.4s, v1.s[1]      \n"
-                    "ins    v2.d[1], x22                \n"
-                    "fmla   v26.4s, v5.4s, v1.s[2]      \n"
-                    "ldr    x20, [%2], #8               \n"
-                    "fmla   v27.4s, v5.4s, v1.s[3]      \n"
-                    "fmla   v16.4s, v4.4s, v2.s[0]      \n"
-                    "ldr    d1, [%2], #8                \n"
-                    "fmla   v17.4s, v4.4s, v2.s[1]      \n"
-                    "ins    v6.d[1], x26                \n"
-                    "fmla   v18.4s, v4.4s, v2.s[2]      \n"
-                    "ldr    x21, [%2], #8               \n"
-                    "fmla   v19.4s, v4.4s, v2.s[3]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                    "fmla   v28.4s, v5.4s, v2.s[0]      \n"
-                    "ldr    d4, [%1], #8                \n"
-                    "fmla   v29.4s, v5.4s, v2.s[1]      \n"
-                    "ins    v3.d[1], x23                \n"
-                    "fmla   v30.4s, v5.4s, v2.s[2]      \n"
-                    "ldr    x24, [%1], #8               \n"
-                    "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-                    "fmla   v8.4s, v6.4s, v3.s[0]       \n"
-                    "ldr    d2, [%2], #8                \n"
-                    "fmla   v9.4s, v6.4s, v3.s[1]       \n"
-                    "ins    v7.d[1], x27                \n"
-                    "fmla   v10.4s, v6.4s, v3.s[2]      \n"
-                    "ldr    x22, [%2], #8               \n"
-                    "fmla   v11.4s, v6.4s, v3.s[3]      \n"
-                    "fmla   v20.4s, v7.4s, v3.s[0]      \n"
-                    "ldr    d5, [%1], #8                \n"
-                    "fmla   v21.4s, v7.4s, v3.s[1]      \n"
-                    "ins    v0.d[1], x20                \n"
-                    "fmla   v22.4s, v7.4s, v3.s[2]      \n"
-                    "ldr    x25, [%1], #8               \n"
-                    "fmla   v23.4s, v7.4s, v3.s[3]      \n"
-                    "fmla   v12.4s, v6.4s, v0.s[0]      \n"
-                    "ldr    d3, [%2], #8                \n"
-                    "fmla   v13.4s, v6.4s, v0.s[1]      \n"
-                    "ldr    x23, [%2], #8               \n"
-                    "fmla   v14.4s, v6.4s, v0.s[2]      \n"
-                    "fmla   v15.4s, v6.4s, v0.s[3]      \n"
-                    "fmla   v24.4s, v7.4s, v0.s[0]      \n"
-                    "fmla   v25.4s, v7.4s, v0.s[1]      \n"
-                    "ins    v1.d[1], x21                \n"
-                    "fmla   v26.4s, v7.4s, v0.s[2]      \n"
-                    "fmla   v27.4s, v7.4s, v0.s[3]      \n"
-                    "prfm   pldl1keep, [%2, #256]       \n" // NOTE PRELOAD
-                    "fmla   v16.4s, v6.4s, v1.s[0]      \n"
-                    "fmla   v17.4s, v6.4s, v1.s[1]      \n"
-                    "ins    v4.d[1], x24                \n"
-                    "fmla   v18.4s, v6.4s, v1.s[2]      \n"
-                    "ldr    d0, [%2], #8                \n"
-                    "fmla   v19.4s, v6.4s, v1.s[3]      \n"
-                    "ldr    x20, [%2], #8               \n"
-                    "fmla   v28.4s, v7.4s, v1.s[0]      \n"
-                    "ldr    d6, [%1], #8                \n"
-                    "fmla   v29.4s, v7.4s, v1.s[1]      \n"
-                    "ins    v2.d[1], x22                \n"
-                    "fmla   v30.4s, v7.4s, v1.s[2]      \n"
-                    "fmla   v31.4s, v7.4s, v1.s[3]      \n"
-                    "ldr    x26, [%1], #8               \n"
-                    "fmla   v8.4s, v4.4s, v2.s[0]       \n"
-                    "ldr    d1, [%2], #8                \n"
-                    "fmla   v9.4s, v4.4s, v2.s[1]       \n"
-                    "ins    v5.d[1], x25                \n"
-                    "fmla   v10.4s, v4.4s, v2.s[2]      \n"
-                    "ldr    x21, [%2], #8               \n"
-                    "fmla   v11.4s, v4.4s, v2.s[3]      \n"
-                    "ldr    d7, [%1], #8                \n"
-                    "fmla   v20.4s, v5.4s, v2.s[0]      \n"
-                    "ldr    x27, [%1], #8               \n"
-                    "fmla   v21.4s, v5.4s, v2.s[1]      \n"
-                    "ins    v3.d[1], x23                \n"
-                    "fmla   v22.4s, v5.4s, v2.s[2]      \n"
-                    "fmla   v23.4s, v5.4s, v2.s[3]      \n"
-                    "fmla   v12.4s, v4.4s, v3.s[0]      \n"
-                    "ldr    d2, [%2], #8                \n"
-                    "fmla   v13.4s, v4.4s, v3.s[1]      \n"
-                    "ldr    x22, [%2], #8               \n"
-                    "fmla   v14.4s, v4.4s, v3.s[2]      \n"
-                    "fmla   v15.4s, v4.4s, v3.s[3]      \n"
-                    "fmla   v24.4s, v5.4s, v3.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v3.s[1]      \n"
-                    "ins    v0.d[1], x20                \n"
-                    "fmla   v26.4s, v5.4s, v3.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v3.s[3]      \n"
-                    "fmla   v16.4s, v4.4s, v0.s[0]      \n"
-                    "ldr    d3, [%2], #8                \n"
-                    "fmla   v17.4s, v4.4s, v0.s[1]      \n"
-                    "ldr    x23, [%2], #8               \n"
-                    "fmla   v18.4s, v4.4s, v0.s[2]      \n"
-                    "ins    v6.d[1], x26                \n"
-                    "fmla   v19.4s, v4.4s, v0.s[3]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                    "fmla   v28.4s, v5.4s, v0.s[0]      \n"
-                    "ldr    d4, [%1], #8                \n"
-                    "fmla   v29.4s, v5.4s, v0.s[1]      \n"
-                    "ins    v1.d[1], x21                \n"
-                    "fmla   v30.4s, v5.4s, v0.s[2]      \n"
-                    "ldr    x24, [%1], #8               \n"
-                    "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-                    "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                    "fmla   v8.4s, v6.4s, v1.s[0]       \n"
-                    "ldr    d0, [%2], #8                \n"
-                    "fmla   v9.4s, v6.4s, v1.s[1]       \n"
-                    "ins    v7.d[1], x27                \n"
-                    "fmla   v10.4s, v6.4s, v1.s[2]      \n"
-                    "ldr    x20, [%2], #8               \n"
-                    "fmla   v11.4s, v6.4s, v1.s[3]      \n"
-                    "ldr    d5, [%1], #8                \n"
-                    "fmla   v20.4s, v7.4s, v1.s[0]      \n"
-                    "ldr    x25, [%1], #8               \n"
-                    "fmla   v21.4s, v7.4s, v1.s[1]      \n"
-                    "ins    v2.d[1], x22                \n"
-                    "fmla   v22.4s, v7.4s, v1.s[2]      \n"
-                    "fmla   v23.4s, v7.4s, v1.s[3]      \n"
-                    "fmla   v12.4s, v6.4s, v2.s[0]      \n"
-                    "fmla   v13.4s, v6.4s, v2.s[1]      \n"
-                    "fmla   v14.4s, v6.4s, v2.s[2]      \n"
-                    "fmla   v15.4s, v6.4s, v2.s[3]      \n"
-                    "fmla   v24.4s, v7.4s, v2.s[0]      \n"
-                    "fmla   v25.4s, v7.4s, v2.s[1]      \n"
-                    "ins    v3.d[1], x23                \n"
-                    "fmla   v26.4s, v7.4s, v2.s[2]      \n"
-                    "fmla   v27.4s, v7.4s, v2.s[3]      \n"
-                    "fmla   v16.4s, v6.4s, v3.s[0]      \n"
-                    "fmla   v17.4s, v6.4s, v3.s[1]      \n"
-                    "fmla   v18.4s, v6.4s, v3.s[2]      \n"
-                    "ins    v4.d[1], x24                \n"
-                    "fmla   v19.4s, v6.4s, v3.s[3]      \n"
-                    "fmla   v28.4s, v7.4s, v3.s[0]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v29.4s, v7.4s, v3.s[1]      \n"
-                    "fmla   v30.4s, v7.4s, v3.s[2]      \n"
-                    "ins    v0.d[1], x20                \n"
-                    "fmla   v31.4s, v7.4s, v3.s[3]      \n"
-                    "bne    4b                          \n"
-
-                    "sub    %1, %1, #32                 \n"
-                    "sub    %2, %2, #16                 \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s}, [%2], #48 \n"
-                    "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-
-                    "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                    "fmla   v9.4s, v4.4s, v0.s[1]       \n"
-                    "fmla   v10.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v11.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v12.4s, v4.4s, v1.s[0]      \n"
-                    "fmla   v13.4s, v4.4s, v1.s[1]      \n"
-                    "fmla   v14.4s, v4.4s, v1.s[2]      \n"
-                    "fmla   v15.4s, v4.4s, v1.s[3]      \n"
-                    "fmla   v16.4s, v4.4s, v2.s[0]      \n"
-                    "fmla   v17.4s, v4.4s, v2.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v2.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v2.s[3]      \n"
-
-                    "subs   w4, w4, #1                  \n"
-
-                    "fmla   v20.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v21.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v22.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v23.4s, v5.4s, v0.s[3]      \n"
-                    "fmla   v24.4s, v5.4s, v1.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v1.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v1.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v1.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v2.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v2.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v2.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    10f                         \n"
-
-                    // if out_elempack == 4
-                    "cmp    %w12, #4                    \n"
-                    "bne    8f                          \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 2          \n"
-                    "st1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%3], #64 \n"
-                    "st1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%3], #64 \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [x4], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [x4], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [x4] \n"
-                    "b      9f                          \n"
-
-                    // if out_elempack == 1
-                    "8:                                 \n"
-                    // transpose8x12
-                    "zip1   v6.4s, v8.4s, v9.4s         \n"
-                    "zip2   v7.4s, v8.4s, v9.4s         \n"
-                    "zip1   v8.4s, v10.4s, v11.4s       \n"
-                    "zip2   v9.4s, v10.4s, v11.4s       \n"
-                    "zip1   v10.4s, v12.4s, v13.4s      \n"
-                    "zip2   v11.4s, v12.4s, v13.4s      \n"
-                    "zip1   v12.4s, v14.4s, v15.4s      \n"
-                    "zip2   v13.4s, v14.4s, v15.4s      \n"
-                    "zip1   v14.4s, v16.4s, v17.4s      \n"
-                    "zip2   v15.4s, v16.4s, v17.4s      \n"
-                    "zip1   v16.4s, v18.4s, v19.4s      \n"
-                    "zip2   v17.4s, v18.4s, v19.4s      \n"
-
-                    "zip1   v18.4s, v20.4s, v21.4s      \n"
-                    "zip2   v19.4s, v20.4s, v21.4s      \n"
-                    "zip1   v20.4s, v22.4s, v23.4s      \n"
-                    "zip2   v21.4s, v22.4s, v23.4s      \n"
-                    "zip1   v22.4s, v24.4s, v25.4s      \n"
-                    "zip2   v23.4s, v24.4s, v25.4s      \n"
-                    "zip1   v24.4s, v26.4s, v27.4s      \n"
-                    "zip2   v25.4s, v26.4s, v27.4s      \n"
-                    "zip1   v26.4s, v28.4s, v29.4s      \n"
-                    "zip2   v27.4s, v28.4s, v29.4s      \n"
-                    "zip1   v28.4s, v30.4s, v31.4s      \n"
-                    "zip2   v29.4s, v30.4s, v31.4s      \n"
-
-                    "zip1   v0.2d, v6.2d, v8.2d         \n"
-                    "zip2   v3.2d, v6.2d, v8.2d         \n"
-                    "zip1   v1.2d, v10.2d, v12.2d       \n"
-                    "zip2   v4.2d, v10.2d, v12.2d       \n"
-                    "zip1   v2.2d, v14.2d, v16.2d       \n"
-                    "zip2   v5.2d, v14.2d, v16.2d       \n"
-
-                    "zip1   v6.2d, v7.2d, v9.2d         \n"
-                    "zip2   v9.2d, v7.2d, v9.2d         \n"
-                    "zip1   v7.2d, v11.2d, v13.2d       \n"
-                    "zip2   v10.2d, v11.2d, v13.2d      \n"
-                    "zip1   v8.2d, v15.2d, v17.2d       \n"
-                    "zip2   v11.2d, v15.2d, v17.2d      \n"
-
-                    "zip1   v12.2d, v18.2d, v20.2d      \n"
-                    "zip2   v15.2d, v18.2d, v20.2d      \n"
-                    "zip1   v13.2d, v22.2d, v24.2d      \n"
-                    "zip2   v16.2d, v22.2d, v24.2d      \n"
-                    "zip1   v14.2d, v26.2d, v28.2d      \n"
-                    "zip2   v17.2d, v26.2d, v28.2d      \n"
-
-                    "zip1   v18.2d, v19.2d, v21.2d      \n"
-                    "zip2   v21.2d, v19.2d, v21.2d      \n"
-                    "zip1   v19.2d, v23.2d, v25.2d      \n"
-                    "zip2   v22.2d, v23.2d, v25.2d      \n"
-                    "zip1   v20.2d, v27.2d, v29.2d      \n"
-                    "zip2   v23.2d, v27.2d, v29.2d      \n"
-
-                    "add    x4, %3, %w13, sxtw 2        \n"
-                    "st1    {v0.4s, v1.4s, v2.4s}, [%3], #48 \n"
-                    "st1    {v3.4s, v4.4s, v5.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v6.4s, v7.4s, v8.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v9.4s, v10.4s, v11.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v12.4s, v13.4s, v14.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v15.4s, v16.4s, v17.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v18.4s, v19.4s, v20.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v21.4s, v22.4s, v23.4s}, [x4] \n"
-
-                    "9:                                 \n"
-                    "add    %0, %0, #384                \n"
-                    "b      11f                         \n"
-
-                    "10:                                \n"
-                    "st1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                    "st1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    "11:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            else if (use_a53_a55_optimized_kernel && !cpu_support_arm_asimdhp())
-            {
-                // a53
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #320                \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "add    x4, %8, #16                 \n"
-                    "ld1    {v8.4s}, [%8]               \n"
-                    "ld1    {v20.4s}, [x4]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v8.16b, v8.16b, v8.16b      \n"
-                    "eor    v20.16b, v20.16b, v20.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v9.16b, v8.16b              \n"
-                    "mov    v10.16b, v8.16b             \n"
-                    "mov    v11.16b, v8.16b             \n"
-                    "mov    v12.16b, v8.16b             \n"
-                    "mov    v13.16b, v8.16b             \n"
-                    "mov    v14.16b, v8.16b             \n"
-                    "mov    v15.16b, v8.16b             \n"
-                    "mov    v16.16b, v8.16b             \n"
-                    "mov    v17.16b, v8.16b             \n"
-                    "mov    v18.16b, v8.16b             \n"
-                    "mov    v19.16b, v8.16b             \n"
-
-                    "mov    v21.16b, v20.16b            \n"
-                    "mov    v22.16b, v20.16b            \n"
-                    "mov    v23.16b, v20.16b            \n"
-                    "mov    v24.16b, v20.16b            \n"
-                    "mov    v25.16b, v20.16b            \n"
-                    "mov    v26.16b, v20.16b            \n"
-                    "mov    v27.16b, v20.16b            \n"
-                    "mov    v28.16b, v20.16b            \n"
-                    "mov    v29.16b, v20.16b            \n"
-                    "mov    v30.16b, v20.16b            \n"
-                    "mov    v31.16b, v20.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v4.4s}, [%1], #16          \n"
-
-                    "prfm   pldl1keep, [%2, #384]       \n"
-                    "ld1    {v0.4s}, [%2], #16          \n"
-
-                    "ldr    d1, [%2]                    \n"
-                    "ldr    x21, [%2, #8]               \n"
-                    "ldr    d2, [%2, #16]               \n"
-                    "ldr    x22, [%2, #24]              \n"
-                    "add    %2, %2, #32                 \n"
-
-                    ".align 4                           \n"
-                    "4:                                 \n"
-
-                    "ldr    d5, [%1]                    \n"
-                    "ins    v1.d[1], x21                \n"
-                    "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                    "ldr    x25, [%1, #8]               \n"
-                    "fmla   v9.4s, v4.4s, v0.s[1]       \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v10.4s, v4.4s, v0.s[2]      \n"
-
-                    "ldr    d6, [%1]                    \n"
-                    "ins    v2.d[1], x22                \n"
-                    "fmla   v11.4s, v4.4s, v0.s[3]      \n"
-                    "ldr    x26, [%1, #8]               \n"
-                    "fmla   v12.4s, v4.4s, v1.s[0]      \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v13.4s, v4.4s, v1.s[1]      \n"
-
-                    "nop                                \n"
-                    "prfm   pldl1keep, [%1, #256]       \n" // NOTE PRELOAD
-                    "fmla   v14.4s, v4.4s, v1.s[2]      \n"
-                    "nop                                \n"
-                    "fmla   v15.4s, v4.4s, v1.s[3]      \n"
-                    "nop                                \n"
-                    "fmla   v16.4s, v4.4s, v2.s[0]      \n"
-
-                    "ldr    d3, [%2]                    \n"
-                    "ins    v5.d[1], x25                \n"
-                    "fmla   v17.4s, v4.4s, v2.s[1]      \n"
-                    "ldr    x23, [%2, #8]               \n"
-                    "fmla   v18.4s, v4.4s, v2.s[2]      \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v19.4s, v4.4s, v2.s[3]      \n"
-
-                    "nop                                \n"
-                    "prfm   pldl1keep, [%2, #384]       \n" // NOTE PRELOAD
-                    "fmla   v20.4s, v5.4s, v0.s[0]      \n"
-                    "nop                                \n"
-                    "fmla   v21.4s, v5.4s, v0.s[1]      \n"
-                    "nop                                \n"
-                    "fmla   v22.4s, v5.4s, v0.s[2]      \n"
-
-                    "nop                                \n"
-                    "nop                                \n"
-                    "fmla   v23.4s, v5.4s, v0.s[3]      \n"
-                    "nop                                \n"
-                    "fmla   v24.4s, v5.4s, v1.s[0]      \n"
-                    "nop                                \n"
-                    "fmla   v25.4s, v5.4s, v1.s[1]      \n"
-
-                    "ldr    d0, [%2]                    \n"
-                    "ins    v6.d[1], x26                \n"
-                    "fmla   v26.4s, v5.4s, v1.s[2]      \n"
-                    "ldr    x20, [%2, #8]               \n"
-                    "fmla   v27.4s, v5.4s, v1.s[3]      \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v28.4s, v5.4s, v2.s[0]      \n"
-
-                    "ldr    d1, [%2]                    \n"
-                    "ins    v3.d[1], x23                \n"
-                    "fmla   v29.4s, v5.4s, v2.s[1]      \n"
-                    "ldr    x21, [%2, #8]               \n"
-                    "fmla   v30.4s, v5.4s, v2.s[2]      \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-
-                    "ldr    d7, [%1]                    \n"
-                    "ins    v0.d[1], x20                \n"
-                    "fmla   v8.4s, v6.4s, v3.s[0]       \n"
-                    "ldr    x27, [%1, #8]               \n"
-                    "fmla   v9.4s, v6.4s, v3.s[1]       \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v10.4s, v6.4s, v3.s[2]      \n"
-
-                    "ldr    d4, [%1]                    \n"
-                    "ins    v1.d[1], x21                \n"
-                    "fmla   v11.4s, v6.4s, v3.s[3]      \n"
-                    "ldr    x24, [%1, #8]               \n"
-                    "fmla   v12.4s, v6.4s, v0.s[0]      \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v13.4s, v6.4s, v0.s[1]      \n"
-
-                    "nop                                \n"
-                    "prfm   pldl1keep, [%1, #256]       \n" // NOTE PRELOAD
-                    "fmla   v14.4s, v6.4s, v0.s[2]      \n"
-                    "nop                                \n"
-                    "fmla   v15.4s, v6.4s, v0.s[3]      \n"
-                    "nop                                \n"
-                    "fmla   v16.4s, v6.4s, v1.s[0]      \n"
-
-                    "ldr    d2, [%2]                    \n"
-                    "ins    v7.d[1], x27                \n"
-                    "fmla   v17.4s, v6.4s, v1.s[1]      \n"
-                    "ldr    x22, [%2, #8]               \n"
-                    "fmla   v18.4s, v6.4s, v1.s[2]      \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v19.4s, v6.4s, v1.s[3]      \n"
-
-                    "nop                                \n"
-                    "prfm   pldl1keep, [%2, #384]       \n" // NOTE PRELOAD
-                    "fmla   v20.4s, v7.4s, v3.s[0]      \n"
-                    "nop                                \n"
-                    "fmla   v21.4s, v7.4s, v3.s[1]      \n"
-                    "nop                                \n"
-                    "fmla   v22.4s, v7.4s, v3.s[2]      \n"
-
-                    "nop                                \n"
-                    "nop                                \n"
-                    "fmla   v23.4s, v7.4s, v3.s[3]      \n"
-                    "nop                                \n"
-                    "fmla   v24.4s, v7.4s, v0.s[0]      \n"
-                    "nop                                \n"
-                    "fmla   v25.4s, v7.4s, v0.s[1]      \n"
-
-                    "ldr    d3, [%2]                    \n"
-                    "ins    v4.d[1], x24                \n"
-                    "fmla   v26.4s, v7.4s, v0.s[2]      \n"
-                    "ldr    x23, [%2, #8]               \n"
-                    "fmla   v27.4s, v7.4s, v0.s[3]      \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v28.4s, v7.4s, v1.s[0]      \n"
-
-                    "ldr    d0, [%2]                    \n"
-                    "ins    v2.d[1], x22                \n"
-                    "fmla   v29.4s, v7.4s, v1.s[1]      \n"
-                    "ldr    x20, [%2, #8]               \n"
-                    "fmla   v30.4s, v7.4s, v1.s[2]      \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v31.4s, v7.4s, v1.s[3]      \n"
-
-                    "ldr    d5, [%1]                    \n"
-                    "ins    v3.d[1], x23                \n"
-                    "fmla   v8.4s, v4.4s, v2.s[0]       \n"
-                    "ldr    x25, [%1, #8]               \n"
-                    "fmla   v9.4s, v4.4s, v2.s[1]       \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v10.4s, v4.4s, v2.s[2]      \n"
-
-                    "ldr    d6, [%1]                    \n"
-                    "ins    v0.d[1], x20                \n"
-                    "fmla   v11.4s, v4.4s, v2.s[3]      \n"
-                    "ldr    x26, [%1, #8]               \n"
-                    "fmla   v12.4s, v4.4s, v3.s[0]      \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v13.4s, v4.4s, v3.s[1]      \n"
-
-                    "nop                                \n"
-                    "prfm   pldl1keep, [%1, #256]       \n" // NOTE PRELOAD
-                    "fmla   v14.4s, v4.4s, v3.s[2]      \n"
-                    "nop                                \n"
-                    "fmla   v15.4s, v4.4s, v3.s[3]      \n"
-                    "nop                                \n"
-                    "fmla   v16.4s, v4.4s, v0.s[0]      \n"
-
-                    "ldr    d1, [%2]                    \n"
-                    "ins    v5.d[1], x25                \n"
-                    "fmla   v17.4s, v4.4s, v0.s[1]      \n"
-                    "ldr    x21, [%2, #8]               \n"
-                    "fmla   v18.4s, v4.4s, v0.s[2]      \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v19.4s, v4.4s, v0.s[3]      \n"
-
-                    "nop                                \n"
-                    "prfm   pldl1keep, [%2, #384]       \n" // NOTE PRELOAD
-                    "fmla   v20.4s, v5.4s, v2.s[0]      \n"
-                    "nop                                \n"
-                    "fmla   v21.4s, v5.4s, v2.s[1]      \n"
-                    "nop                                \n"
-                    "fmla   v22.4s, v5.4s, v2.s[2]      \n"
-
-                    "nop                                \n"
-                    "nop                                \n"
-                    "fmla   v23.4s, v5.4s, v2.s[3]      \n"
-                    "nop                                \n"
-                    "fmla   v24.4s, v5.4s, v3.s[0]      \n"
-                    "nop                                \n"
-                    "fmla   v25.4s, v5.4s, v3.s[1]      \n"
-
-                    "ldr    d2, [%2]                    \n"
-                    "ins    v6.d[1], x26                \n"
-                    "fmla   v26.4s, v5.4s, v3.s[2]      \n"
-                    "ldr    x22, [%2, #8]               \n"
-                    "fmla   v27.4s, v5.4s, v3.s[3]      \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v28.4s, v5.4s, v0.s[0]      \n"
-
-                    "ldr    d3, [%2]                    \n"
-                    "ins    v1.d[1], x21                \n"
-                    "fmla   v29.4s, v5.4s, v0.s[1]      \n"
-                    "ldr    x23, [%2, #8]               \n"
-                    "fmla   v30.4s, v5.4s, v0.s[2]      \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-
-                    "ldr    d7, [%1]                    \n"
-                    "ins    v2.d[1], x22                \n"
-                    "fmla   v8.4s, v6.4s, v1.s[0]       \n"
-                    "ldr    x27, [%1, #8]               \n"
-                    "fmla   v9.4s, v6.4s, v1.s[1]       \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v10.4s, v6.4s, v1.s[2]      \n"
-
-                    "ldr    d4, [%1]                    \n"
-                    "ins    v3.d[1], x23                \n"
-                    "fmla   v11.4s, v6.4s, v1.s[3]      \n"
-                    "ldr    x24, [%1, #8]               \n"
-                    "fmla   v12.4s, v6.4s, v2.s[0]      \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v13.4s, v6.4s, v2.s[1]      \n"
-
-                    "nop                                \n"
-                    "prfm   pldl1keep, [%1, #256]       \n" // NOTE PRELOAD
-                    "fmla   v14.4s, v6.4s, v2.s[2]      \n"
-                    "nop                                \n"
-                    "fmla   v15.4s, v6.4s, v2.s[3]      \n"
-                    "nop                                \n"
-                    "fmla   v16.4s, v6.4s, v3.s[0]      \n"
-
-                    "ldr    d0, [%2]                    \n"
-                    "ins    v7.d[1], x27                \n"
-                    "fmla   v17.4s, v6.4s, v3.s[1]      \n"
-                    "ldr    x20, [%2, #8]               \n"
-                    "fmla   v18.4s, v6.4s, v3.s[2]      \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v19.4s, v6.4s, v3.s[3]      \n"
-
-                    "nop                                \n"
-                    "prfm   pldl1keep, [%2, #384]       \n" // NOTE PRELOAD
-                    "fmla   v20.4s, v7.4s, v1.s[0]      \n"
-                    "nop                                \n"
-                    "fmla   v21.4s, v7.4s, v1.s[1]      \n"
-                    "nop                                \n"
-                    "fmla   v22.4s, v7.4s, v1.s[2]      \n"
-
-                    "nop                                \n"
-                    "nop                                \n"
-                    "fmla   v23.4s, v7.4s, v1.s[3]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v24.4s, v7.4s, v2.s[0]      \n"
-                    "nop                                \n"
-                    "fmla   v25.4s, v7.4s, v2.s[1]      \n"
-
-                    "ldr    d1, [%2]                    \n"
-                    "ins    v4.d[1], x24                \n"
-                    "fmla   v26.4s, v7.4s, v2.s[2]      \n"
-                    "ldr    x21, [%2, #8]               \n"
-                    "fmla   v27.4s, v7.4s, v2.s[3]      \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v28.4s, v7.4s, v3.s[0]      \n"
-
-                    "ldr    d2, [%2]                    \n"
-                    "ins    v0.d[1], x20                \n"
-                    "fmla   v29.4s, v7.4s, v3.s[1]      \n"
-                    "ldr    x22, [%2, #8]               \n"
-                    "fmla   v30.4s, v7.4s, v3.s[2]      \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v31.4s, v7.4s, v3.s[3]      \n"
-
-                    "bne    4b                          \n"
-
-                    "sub    %1, %1, #16                 \n"
-                    "sub    %2, %2, #48                 \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s}, [%2], #48 \n"
-                    "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-
-                    "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                    "fmla   v9.4s, v4.4s, v0.s[1]       \n"
-                    "fmla   v10.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v11.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v12.4s, v4.4s, v1.s[0]      \n"
-                    "fmla   v13.4s, v4.4s, v1.s[1]      \n"
-                    "fmla   v14.4s, v4.4s, v1.s[2]      \n"
-                    "fmla   v15.4s, v4.4s, v1.s[3]      \n"
-                    "fmla   v16.4s, v4.4s, v2.s[0]      \n"
-                    "fmla   v17.4s, v4.4s, v2.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v2.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v2.s[3]      \n"
-
-                    "subs   w4, w4, #1                  \n"
-
-                    "fmla   v20.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v21.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v22.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v23.4s, v5.4s, v0.s[3]      \n"
-                    "fmla   v24.4s, v5.4s, v1.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v1.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v1.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v1.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v2.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v2.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v2.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    10f                         \n"
-
-                    // if out_elempack == 4
-                    "cmp    %w12, #4                    \n"
-                    "bne    8f                          \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 2          \n"
-                    "st1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%3], #64 \n"
-                    "st1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%3], #64 \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [x4], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [x4], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [x4] \n"
-                    "b      9f                          \n"
-
-                    // if out_elempack == 1
-                    "8:                                 \n"
-                    // transpose8x12
-                    "zip1   v6.4s, v8.4s, v9.4s         \n"
-                    "zip2   v7.4s, v8.4s, v9.4s         \n"
-                    "zip1   v8.4s, v10.4s, v11.4s       \n"
-                    "zip2   v9.4s, v10.4s, v11.4s       \n"
-                    "zip1   v10.4s, v12.4s, v13.4s      \n"
-                    "zip2   v11.4s, v12.4s, v13.4s      \n"
-                    "zip1   v12.4s, v14.4s, v15.4s      \n"
-                    "zip2   v13.4s, v14.4s, v15.4s      \n"
-                    "zip1   v14.4s, v16.4s, v17.4s      \n"
-                    "zip2   v15.4s, v16.4s, v17.4s      \n"
-                    "zip1   v16.4s, v18.4s, v19.4s      \n"
-                    "zip2   v17.4s, v18.4s, v19.4s      \n"
-
-                    "zip1   v18.4s, v20.4s, v21.4s      \n"
-                    "zip2   v19.4s, v20.4s, v21.4s      \n"
-                    "zip1   v20.4s, v22.4s, v23.4s      \n"
-                    "zip2   v21.4s, v22.4s, v23.4s      \n"
-                    "zip1   v22.4s, v24.4s, v25.4s      \n"
-                    "zip2   v23.4s, v24.4s, v25.4s      \n"
-                    "zip1   v24.4s, v26.4s, v27.4s      \n"
-                    "zip2   v25.4s, v26.4s, v27.4s      \n"
-                    "zip1   v26.4s, v28.4s, v29.4s      \n"
-                    "zip2   v27.4s, v28.4s, v29.4s      \n"
-                    "zip1   v28.4s, v30.4s, v31.4s      \n"
-                    "zip2   v29.4s, v30.4s, v31.4s      \n"
-
-                    "zip1   v0.2d, v6.2d, v8.2d         \n"
-                    "zip2   v3.2d, v6.2d, v8.2d         \n"
-                    "zip1   v1.2d, v10.2d, v12.2d       \n"
-                    "zip2   v4.2d, v10.2d, v12.2d       \n"
-                    "zip1   v2.2d, v14.2d, v16.2d       \n"
-                    "zip2   v5.2d, v14.2d, v16.2d       \n"
-
-                    "zip1   v6.2d, v7.2d, v9.2d         \n"
-                    "zip2   v9.2d, v7.2d, v9.2d         \n"
-                    "zip1   v7.2d, v11.2d, v13.2d       \n"
-                    "zip2   v10.2d, v11.2d, v13.2d      \n"
-                    "zip1   v8.2d, v15.2d, v17.2d       \n"
-                    "zip2   v11.2d, v15.2d, v17.2d      \n"
-
-                    "zip1   v12.2d, v18.2d, v20.2d      \n"
-                    "zip2   v15.2d, v18.2d, v20.2d      \n"
-                    "zip1   v13.2d, v22.2d, v24.2d      \n"
-                    "zip2   v16.2d, v22.2d, v24.2d      \n"
-                    "zip1   v14.2d, v26.2d, v28.2d      \n"
-                    "zip2   v17.2d, v26.2d, v28.2d      \n"
-
-                    "zip1   v18.2d, v19.2d, v21.2d      \n"
-                    "zip2   v21.2d, v19.2d, v21.2d      \n"
-                    "zip1   v19.2d, v23.2d, v25.2d      \n"
-                    "zip2   v22.2d, v23.2d, v25.2d      \n"
-                    "zip1   v20.2d, v27.2d, v29.2d      \n"
-                    "zip2   v23.2d, v27.2d, v29.2d      \n"
-
-                    "add    x4, %3, %w13, sxtw 2        \n"
-                    "st1    {v0.4s, v1.4s, v2.4s}, [%3], #48 \n"
-                    "st1    {v3.4s, v4.4s, v5.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v6.4s, v7.4s, v8.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v9.4s, v10.4s, v11.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v12.4s, v13.4s, v14.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v15.4s, v16.4s, v17.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v18.4s, v19.4s, v20.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v21.4s, v22.4s, v23.4s}, [x4] \n"
-
-                    "9:                                 \n"
-                    "add    %0, %0, #384                \n"
-                    "b      11f                         \n"
-
-                    "10:                                \n"
-                    "st1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                    "st1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    "11:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            else
-            {
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #320                \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "add    x4, %8, #16                 \n"
-                    "ld1    {v8.4s}, [%8]               \n"
-                    "ld1    {v20.4s}, [x4]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v8.16b, v8.16b, v8.16b      \n"
-                    "eor    v20.16b, v20.16b, v20.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v9.16b, v8.16b              \n"
-                    "mov    v10.16b, v8.16b             \n"
-                    "mov    v11.16b, v8.16b             \n"
-                    "mov    v12.16b, v8.16b             \n"
-                    "mov    v13.16b, v8.16b             \n"
-                    "mov    v14.16b, v8.16b             \n"
-                    "mov    v15.16b, v8.16b             \n"
-                    "mov    v16.16b, v8.16b             \n"
-                    "mov    v17.16b, v8.16b             \n"
-                    "mov    v18.16b, v8.16b             \n"
-                    "mov    v19.16b, v8.16b             \n"
-
-                    "mov    v21.16b, v20.16b            \n"
-                    "mov    v22.16b, v20.16b            \n"
-                    "mov    v23.16b, v20.16b            \n"
-                    "mov    v24.16b, v20.16b            \n"
-                    "mov    v25.16b, v20.16b            \n"
-                    "mov    v26.16b, v20.16b            \n"
-                    "mov    v27.16b, v20.16b            \n"
-                    "mov    v28.16b, v20.16b            \n"
-                    "mov    v29.16b, v20.16b            \n"
-                    "mov    v30.16b, v20.16b            \n"
-                    "mov    v31.16b, v20.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-
-                    "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                    "fmla   v9.4s, v4.4s, v0.s[1]       \n"
-                    "fmla   v10.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v11.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v20.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v21.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v22.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v23.4s, v5.4s, v0.s[3]      \n"
-
-                    "fmla   v12.4s, v4.4s, v1.s[0]      \n"
-                    "fmla   v13.4s, v4.4s, v1.s[1]      \n"
-                    "fmla   v14.4s, v4.4s, v1.s[2]      \n"
-                    "fmla   v15.4s, v4.4s, v1.s[3]      \n"
-                    "fmla   v24.4s, v5.4s, v1.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v1.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v1.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v1.s[3]      \n"
-
-                    "fmla   v16.4s, v4.4s, v2.s[0]      \n"
-                    "fmla   v17.4s, v4.4s, v2.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v2.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v2.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v2.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v2.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v2.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-
-                    "fmla   v8.4s, v6.4s, v3.s[0]       \n"
-                    "fmla   v9.4s, v6.4s, v3.s[1]       \n"
-                    "fmla   v10.4s, v6.4s, v3.s[2]      \n"
-                    "fmla   v11.4s, v6.4s, v3.s[3]      \n"
-                    "fmla   v20.4s, v7.4s, v3.s[0]      \n"
-                    "fmla   v21.4s, v7.4s, v3.s[1]      \n"
-                    "fmla   v22.4s, v7.4s, v3.s[2]      \n"
-                    "fmla   v23.4s, v7.4s, v3.s[3]      \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-
-                    "fmla   v12.4s, v6.4s, v0.s[0]      \n"
-                    "fmla   v13.4s, v6.4s, v0.s[1]      \n"
-                    "fmla   v14.4s, v6.4s, v0.s[2]      \n"
-                    "fmla   v15.4s, v6.4s, v0.s[3]      \n"
-                    "fmla   v24.4s, v7.4s, v0.s[0]      \n"
-                    "fmla   v25.4s, v7.4s, v0.s[1]      \n"
-                    "fmla   v26.4s, v7.4s, v0.s[2]      \n"
-                    "fmla   v27.4s, v7.4s, v0.s[3]      \n"
-
-                    "fmla   v16.4s, v6.4s, v1.s[0]      \n"
-                    "fmla   v17.4s, v6.4s, v1.s[1]      \n"
-                    "fmla   v18.4s, v6.4s, v1.s[2]      \n"
-                    "fmla   v19.4s, v6.4s, v1.s[3]      \n"
-                    "fmla   v28.4s, v7.4s, v1.s[0]      \n"
-                    "fmla   v29.4s, v7.4s, v1.s[1]      \n"
-                    "fmla   v30.4s, v7.4s, v1.s[2]      \n"
-                    "fmla   v31.4s, v7.4s, v1.s[3]      \n"
-
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n"
-
-                    "fmla   v8.4s, v4.4s, v2.s[0]       \n"
-                    "fmla   v9.4s, v4.4s, v2.s[1]       \n"
-                    "fmla   v10.4s, v4.4s, v2.s[2]      \n"
-                    "fmla   v11.4s, v4.4s, v2.s[3]      \n"
-                    "fmla   v20.4s, v5.4s, v2.s[0]      \n"
-                    "fmla   v21.4s, v5.4s, v2.s[1]      \n"
-                    "fmla   v22.4s, v5.4s, v2.s[2]      \n"
-                    "fmla   v23.4s, v5.4s, v2.s[3]      \n"
-
-                    "fmla   v12.4s, v4.4s, v3.s[0]      \n"
-                    "fmla   v13.4s, v4.4s, v3.s[1]      \n"
-                    "fmla   v14.4s, v4.4s, v3.s[2]      \n"
-                    "fmla   v15.4s, v4.4s, v3.s[3]      \n"
-                    "fmla   v24.4s, v5.4s, v3.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v3.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v3.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v3.s[3]      \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-
-                    "fmla   v16.4s, v4.4s, v0.s[0]      \n"
-                    "fmla   v17.4s, v4.4s, v0.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-
-                    "fmla   v8.4s, v6.4s, v1.s[0]       \n"
-                    "fmla   v9.4s, v6.4s, v1.s[1]       \n"
-                    "fmla   v10.4s, v6.4s, v1.s[2]      \n"
-                    "fmla   v11.4s, v6.4s, v1.s[3]      \n"
-                    "fmla   v20.4s, v7.4s, v1.s[0]      \n"
-                    "fmla   v21.4s, v7.4s, v1.s[1]      \n"
-                    "fmla   v22.4s, v7.4s, v1.s[2]      \n"
-                    "fmla   v23.4s, v7.4s, v1.s[3]      \n"
-
-                    "fmla   v12.4s, v6.4s, v2.s[0]      \n"
-                    "fmla   v13.4s, v6.4s, v2.s[1]      \n"
-                    "fmla   v14.4s, v6.4s, v2.s[2]      \n"
-                    "fmla   v15.4s, v6.4s, v2.s[3]      \n"
-                    "fmla   v24.4s, v7.4s, v2.s[0]      \n"
-                    "fmla   v25.4s, v7.4s, v2.s[1]      \n"
-                    "fmla   v26.4s, v7.4s, v2.s[2]      \n"
-                    "fmla   v27.4s, v7.4s, v2.s[3]      \n"
-
-                    "subs   w4, w4, #1                  \n"
-
-                    "fmla   v16.4s, v6.4s, v3.s[0]      \n"
-                    "fmla   v17.4s, v6.4s, v3.s[1]      \n"
-                    "fmla   v18.4s, v6.4s, v3.s[2]      \n"
-                    "fmla   v19.4s, v6.4s, v3.s[3]      \n"
-                    "fmla   v28.4s, v7.4s, v3.s[0]      \n"
-                    "fmla   v29.4s, v7.4s, v3.s[1]      \n"
-                    "fmla   v30.4s, v7.4s, v3.s[2]      \n"
-                    "fmla   v31.4s, v7.4s, v3.s[3]      \n"
-
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s}, [%2], #48 \n"
-                    "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-
-                    "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                    "fmla   v9.4s, v4.4s, v0.s[1]       \n"
-                    "fmla   v10.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v11.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v12.4s, v4.4s, v1.s[0]      \n"
-                    "fmla   v13.4s, v4.4s, v1.s[1]      \n"
-                    "fmla   v14.4s, v4.4s, v1.s[2]      \n"
-                    "fmla   v15.4s, v4.4s, v1.s[3]      \n"
-                    "fmla   v16.4s, v4.4s, v2.s[0]      \n"
-                    "fmla   v17.4s, v4.4s, v2.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v2.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v2.s[3]      \n"
-
-                    "subs   w4, w4, #1                  \n"
-
-                    "fmla   v20.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v21.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v22.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v23.4s, v5.4s, v0.s[3]      \n"
-                    "fmla   v24.4s, v5.4s, v1.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v1.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v1.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v1.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v2.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v2.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v2.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    10f                         \n"
-
-                    // if out_elempack == 4
-                    "cmp    %w12, #4                    \n"
-                    "bne    8f                          \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 2          \n"
-                    "st1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%3], #64 \n"
-                    "st1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%3], #64 \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [x4], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [x4], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [x4] \n"
-                    "b      9f                          \n"
-
-                    // if out_elempack == 1
-                    "8:                                 \n"
-                    // transpose8x12
-                    "zip1   v6.4s, v8.4s, v9.4s         \n"
-                    "zip2   v7.4s, v8.4s, v9.4s         \n"
-                    "zip1   v8.4s, v10.4s, v11.4s       \n"
-                    "zip2   v9.4s, v10.4s, v11.4s       \n"
-                    "zip1   v10.4s, v12.4s, v13.4s      \n"
-                    "zip2   v11.4s, v12.4s, v13.4s      \n"
-                    "zip1   v12.4s, v14.4s, v15.4s      \n"
-                    "zip2   v13.4s, v14.4s, v15.4s      \n"
-                    "zip1   v14.4s, v16.4s, v17.4s      \n"
-                    "zip2   v15.4s, v16.4s, v17.4s      \n"
-                    "zip1   v16.4s, v18.4s, v19.4s      \n"
-                    "zip2   v17.4s, v18.4s, v19.4s      \n"
-
-                    "zip1   v18.4s, v20.4s, v21.4s      \n"
-                    "zip2   v19.4s, v20.4s, v21.4s      \n"
-                    "zip1   v20.4s, v22.4s, v23.4s      \n"
-                    "zip2   v21.4s, v22.4s, v23.4s      \n"
-                    "zip1   v22.4s, v24.4s, v25.4s      \n"
-                    "zip2   v23.4s, v24.4s, v25.4s      \n"
-                    "zip1   v24.4s, v26.4s, v27.4s      \n"
-                    "zip2   v25.4s, v26.4s, v27.4s      \n"
-                    "zip1   v26.4s, v28.4s, v29.4s      \n"
-                    "zip2   v27.4s, v28.4s, v29.4s      \n"
-                    "zip1   v28.4s, v30.4s, v31.4s      \n"
-                    "zip2   v29.4s, v30.4s, v31.4s      \n"
-
-                    "zip1   v0.2d, v6.2d, v8.2d         \n"
-                    "zip2   v3.2d, v6.2d, v8.2d         \n"
-                    "zip1   v1.2d, v10.2d, v12.2d       \n"
-                    "zip2   v4.2d, v10.2d, v12.2d       \n"
-                    "zip1   v2.2d, v14.2d, v16.2d       \n"
-                    "zip2   v5.2d, v14.2d, v16.2d       \n"
-
-                    "zip1   v6.2d, v7.2d, v9.2d         \n"
-                    "zip2   v9.2d, v7.2d, v9.2d         \n"
-                    "zip1   v7.2d, v11.2d, v13.2d       \n"
-                    "zip2   v10.2d, v11.2d, v13.2d      \n"
-                    "zip1   v8.2d, v15.2d, v17.2d       \n"
-                    "zip2   v11.2d, v15.2d, v17.2d      \n"
-
-                    "zip1   v12.2d, v18.2d, v20.2d      \n"
-                    "zip2   v15.2d, v18.2d, v20.2d      \n"
-                    "zip1   v13.2d, v22.2d, v24.2d      \n"
-                    "zip2   v16.2d, v22.2d, v24.2d      \n"
-                    "zip1   v14.2d, v26.2d, v28.2d      \n"
-                    "zip2   v17.2d, v26.2d, v28.2d      \n"
-
-                    "zip1   v18.2d, v19.2d, v21.2d      \n"
-                    "zip2   v21.2d, v19.2d, v21.2d      \n"
-                    "zip1   v19.2d, v23.2d, v25.2d      \n"
-                    "zip2   v22.2d, v23.2d, v25.2d      \n"
-                    "zip1   v20.2d, v27.2d, v29.2d      \n"
-                    "zip2   v23.2d, v27.2d, v29.2d      \n"
-
-                    "add    x4, %3, %w13, sxtw 2        \n"
-                    "st1    {v0.4s, v1.4s, v2.4s}, [%3], #48 \n"
-                    "st1    {v3.4s, v4.4s, v5.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v6.4s, v7.4s, v8.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v9.4s, v10.4s, v11.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v12.4s, v13.4s, v14.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v15.4s, v16.4s, v17.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v18.4s, v19.4s, v20.4s}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v21.4s, v22.4s, v23.4s}, [x4] \n"
-
-                    "9:                                 \n"
-                    "add    %0, %0, #384                \n"
-                    "b      11f                         \n"
-
-                    "10:                                \n"
-                    "st1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                    "st1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    "11:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum20;
-            float32x4_t _sum21;
-            float32x4_t _sum30;
-            float32x4_t _sum31;
-            float32x4_t _sum40;
-            float32x4_t _sum41;
-            float32x4_t _sum50;
-            float32x4_t _sum51;
-            float32x4_t _sum60;
-            float32x4_t _sum61;
-            float32x4_t _sum70;
-            float32x4_t _sum71;
-            float32x4_t _sum80;
-            float32x4_t _sum81;
-            float32x4_t _sum90;
-            float32x4_t _sum91;
-            float32x4_t _suma0;
-            float32x4_t _suma1;
-            float32x4_t _sumb0;
-            float32x4_t _sumb1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vld1q_f32(pC);
-                    _sum01 = vld1q_f32(pC + 4);
-                    _sum10 = _sum00;
-                    _sum11 = _sum01;
-                    _sum20 = _sum00;
-                    _sum21 = _sum01;
-                    _sum30 = _sum00;
-                    _sum31 = _sum01;
-                    _sum40 = _sum00;
-                    _sum41 = _sum01;
-                    _sum50 = _sum00;
-                    _sum51 = _sum01;
-                    _sum60 = _sum00;
-                    _sum61 = _sum01;
-                    _sum70 = _sum00;
-                    _sum71 = _sum01;
-                    _sum80 = _sum00;
-                    _sum81 = _sum01;
-                    _sum90 = _sum00;
-                    _sum91 = _sum01;
-                    _suma0 = _sum00;
-                    _suma1 = _sum01;
-                    _sumb0 = _sum00;
-                    _sumb1 = _sum01;
-                }
-                else
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                    _sum20 = vdupq_n_f32(0.f);
-                    _sum21 = vdupq_n_f32(0.f);
-                    _sum30 = vdupq_n_f32(0.f);
-                    _sum31 = vdupq_n_f32(0.f);
-                    _sum40 = vdupq_n_f32(0.f);
-                    _sum41 = vdupq_n_f32(0.f);
-                    _sum50 = vdupq_n_f32(0.f);
-                    _sum51 = vdupq_n_f32(0.f);
-                    _sum60 = vdupq_n_f32(0.f);
-                    _sum61 = vdupq_n_f32(0.f);
-                    _sum70 = vdupq_n_f32(0.f);
-                    _sum71 = vdupq_n_f32(0.f);
-                    _sum80 = vdupq_n_f32(0.f);
-                    _sum81 = vdupq_n_f32(0.f);
-                    _sum90 = vdupq_n_f32(0.f);
-                    _sum91 = vdupq_n_f32(0.f);
-                    _suma0 = vdupq_n_f32(0.f);
-                    _suma1 = vdupq_n_f32(0.f);
-                    _sumb0 = vdupq_n_f32(0.f);
-                    _sumb1 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-                _sum20 = vld1q_f32(outptr + 4 * 4);
-                _sum21 = vld1q_f32(outptr + 4 * 5);
-                _sum30 = vld1q_f32(outptr + 4 * 6);
-                _sum31 = vld1q_f32(outptr + 4 * 7);
-                _sum40 = vld1q_f32(outptr + 4 * 8);
-                _sum41 = vld1q_f32(outptr + 4 * 9);
-                _sum50 = vld1q_f32(outptr + 4 * 10);
-                _sum51 = vld1q_f32(outptr + 4 * 11);
-                _sum60 = vld1q_f32(outptr + 4 * 12);
-                _sum61 = vld1q_f32(outptr + 4 * 13);
-                _sum70 = vld1q_f32(outptr + 4 * 14);
-                _sum71 = vld1q_f32(outptr + 4 * 15);
-                _sum80 = vld1q_f32(outptr + 4 * 16);
-                _sum81 = vld1q_f32(outptr + 4 * 17);
-                _sum90 = vld1q_f32(outptr + 4 * 18);
-                _sum91 = vld1q_f32(outptr + 4 * 19);
-                _suma0 = vld1q_f32(outptr + 4 * 20);
-                _suma1 = vld1q_f32(outptr + 4 * 21);
-                _sumb0 = vld1q_f32(outptr + 4 * 22);
-                _sumb1 = vld1q_f32(outptr + 4 * 23);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA0 = vld1q_f32(pA);
-                float32x4_t _pA1 = vld1q_f32(pA + 4);
-
-                float32x4_t _pB0 = vld1q_f32(pB);
-                float32x4_t _pB1 = vld1q_f32(pB + 4);
-                float32x4_t _pB2 = vld1q_f32(pB + 8);
-
-                _sum00 = vfmaq_laneq_f32(_sum00, _pA0, _pB0, 0);
-                _sum01 = vfmaq_laneq_f32(_sum01, _pA1, _pB0, 0);
-                _sum10 = vfmaq_laneq_f32(_sum10, _pA0, _pB0, 1);
-                _sum11 = vfmaq_laneq_f32(_sum11, _pA1, _pB0, 1);
-                _sum20 = vfmaq_laneq_f32(_sum20, _pA0, _pB0, 2);
-                _sum21 = vfmaq_laneq_f32(_sum21, _pA1, _pB0, 2);
-                _sum30 = vfmaq_laneq_f32(_sum30, _pA0, _pB0, 3);
-                _sum31 = vfmaq_laneq_f32(_sum31, _pA1, _pB0, 3);
-                _sum40 = vfmaq_laneq_f32(_sum40, _pA0, _pB1, 0);
-                _sum41 = vfmaq_laneq_f32(_sum41, _pA1, _pB1, 0);
-                _sum50 = vfmaq_laneq_f32(_sum50, _pA0, _pB1, 1);
-                _sum51 = vfmaq_laneq_f32(_sum51, _pA1, _pB1, 1);
-                _sum60 = vfmaq_laneq_f32(_sum60, _pA0, _pB1, 2);
-                _sum61 = vfmaq_laneq_f32(_sum61, _pA1, _pB1, 2);
-                _sum70 = vfmaq_laneq_f32(_sum70, _pA0, _pB1, 3);
-                _sum71 = vfmaq_laneq_f32(_sum71, _pA1, _pB1, 3);
-                _sum80 = vfmaq_laneq_f32(_sum80, _pA0, _pB2, 0);
-                _sum81 = vfmaq_laneq_f32(_sum81, _pA1, _pB2, 0);
-                _sum90 = vfmaq_laneq_f32(_sum90, _pA0, _pB2, 1);
-                _sum91 = vfmaq_laneq_f32(_sum91, _pA1, _pB2, 1);
-                _suma0 = vfmaq_laneq_f32(_suma0, _pA0, _pB2, 2);
-                _suma1 = vfmaq_laneq_f32(_suma1, _pA1, _pB2, 2);
-                _sumb0 = vfmaq_laneq_f32(_sumb0, _pA0, _pB2, 3);
-                _sumb1 = vfmaq_laneq_f32(_sumb1, _pA1, _pB2, 3);
-
-                pA += 8;
-                pB += 12;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1q_f32(outptr0, _sum00);
-                    vst1q_f32(outptr0 + 4, _sum10);
-                    vst1q_f32(outptr0 + 4 * 2, _sum20);
-                    vst1q_f32(outptr0 + 4 * 3, _sum30);
-                    vst1q_f32(outptr0 + 4 * 4, _sum40);
-                    vst1q_f32(outptr0 + 4 * 5, _sum50);
-                    vst1q_f32(outptr0 + 4 * 6, _sum60);
-                    vst1q_f32(outptr0 + 4 * 7, _sum70);
-                    vst1q_f32(outptr0 + 4 * 8, _sum80);
-                    vst1q_f32(outptr0 + 4 * 9, _sum90);
-                    vst1q_f32(outptr0 + 4 * 10, _suma0);
-                    vst1q_f32(outptr0 + 4 * 11, _sumb0);
-
-                    vst1q_f32(outptr0 + out_hstep * 4, _sum01);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4, _sum11);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 2, _sum21);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 3, _sum31);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 4, _sum41);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 5, _sum51);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 6, _sum61);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 7, _sum71);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 8, _sum81);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 9, _sum91);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 10, _suma1);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 11, _sumb1);
-
-                    outptr0 += 48;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose8x12_ps(_sum00, _sum01, _sum10, _sum11, _sum20, _sum21, _sum30, _sum31, _sum40, _sum41, _sum50, _sum51, _sum60, _sum61, _sum70, _sum71, _sum80, _sum81, _sum90, _sum91, _suma0, _suma1, _sumb0, _sumb1);
-
-                    vst1q_f32(outptr0, _sum00);
-                    vst1q_f32(outptr0 + 4, _sum01);
-                    vst1q_f32(outptr0 + 8, _sum10);
-                    vst1q_f32(outptr0 + out_hstep, _sum11);
-                    vst1q_f32(outptr0 + out_hstep + 4, _sum20);
-                    vst1q_f32(outptr0 + out_hstep + 8, _sum21);
-                    vst1q_f32(outptr0 + out_hstep * 2, _sum30);
-                    vst1q_f32(outptr0 + out_hstep * 2 + 4, _sum31);
-                    vst1q_f32(outptr0 + out_hstep * 2 + 8, _sum40);
-                    vst1q_f32(outptr0 + out_hstep * 3, _sum41);
-                    vst1q_f32(outptr0 + out_hstep * 3 + 4, _sum50);
-                    vst1q_f32(outptr0 + out_hstep * 3 + 8, _sum51);
-                    vst1q_f32(outptr0 + out_hstep * 4, _sum60);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4, _sum61);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 8, _sum70);
-                    vst1q_f32(outptr0 + out_hstep * 5, _sum71);
-                    vst1q_f32(outptr0 + out_hstep * 5 + 4, _sum80);
-                    vst1q_f32(outptr0 + out_hstep * 5 + 8, _sum81);
-                    vst1q_f32(outptr0 + out_hstep * 6, _sum90);
-                    vst1q_f32(outptr0 + out_hstep * 6 + 4, _sum91);
-                    vst1q_f32(outptr0 + out_hstep * 6 + 8, _suma0);
-                    vst1q_f32(outptr0 + out_hstep * 7, _suma1);
-                    vst1q_f32(outptr0 + out_hstep * 7 + 4, _sumb0);
-                    vst1q_f32(outptr0 + out_hstep * 7 + 8, _sumb1);
-
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                vst1q_f32(outptr + 4 * 4, _sum20);
-                vst1q_f32(outptr + 4 * 5, _sum21);
-                vst1q_f32(outptr + 4 * 6, _sum30);
-                vst1q_f32(outptr + 4 * 7, _sum31);
-                vst1q_f32(outptr + 4 * 8, _sum40);
-                vst1q_f32(outptr + 4 * 9, _sum41);
-                vst1q_f32(outptr + 4 * 10, _sum50);
-                vst1q_f32(outptr + 4 * 11, _sum51);
-                vst1q_f32(outptr + 4 * 12, _sum60);
-                vst1q_f32(outptr + 4 * 13, _sum61);
-                vst1q_f32(outptr + 4 * 14, _sum70);
-                vst1q_f32(outptr + 4 * 15, _sum71);
-                vst1q_f32(outptr + 4 * 16, _sum80);
-                vst1q_f32(outptr + 4 * 17, _sum81);
-                vst1q_f32(outptr + 4 * 18, _sum90);
-                vst1q_f32(outptr + 4 * 19, _sum91);
-                vst1q_f32(outptr + 4 * 20, _suma0);
-                vst1q_f32(outptr + 4 * 21, _suma1);
-                vst1q_f32(outptr + 4 * 22, _sumb0);
-                vst1q_f32(outptr + 4 * 23, _sumb1);
-            }
-
-            outptr += 96;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            if (use_a53_a55_optimized_kernel && cpu_support_arm_asimdhp())
-            {
-                // a55
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #192                \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "add    x4, %8, #16                 \n"
-                    "ld1    {v16.4s}, [%8]              \n"
-                    "ld1    {v24.4s}, [x4]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v16.16b, v16.16b, v16.16b   \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v17.16b, v16.16b            \n"
-                    "mov    v18.16b, v16.16b            \n"
-                    "mov    v19.16b, v16.16b            \n"
-                    "mov    v20.16b, v16.16b            \n"
-                    "mov    v21.16b, v16.16b            \n"
-                    "mov    v22.16b, v16.16b            \n"
-                    "mov    v23.16b, v16.16b            \n"
-
-                    "mov    v25.16b, v24.16b            \n"
-                    "mov    v26.16b, v24.16b            \n"
-                    "mov    v27.16b, v24.16b            \n"
-                    "mov    v28.16b, v24.16b            \n"
-                    "mov    v29.16b, v24.16b            \n"
-                    "mov    v30.16b, v24.16b            \n"
-                    "mov    v31.16b, v24.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v8.4s}, [%1], #16          \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.4s}, [%2], #16          \n"
-
-                    "ldr    d1, [%2], #8                \n"
-                    "ldr    x21, [%2], #8               \n"
-
-                    ".align 4                           \n"
-                    "4:                                 \n"
-                    "ldr    d9, [%1], #8                \n"
-                    "fmla   v16.4s, v8.4s, v0.s[0]      \n"
-                    "ldr    x25, [%1], #8               \n"
-                    "fmla   v17.4s, v8.4s, v0.s[1]      \n"
-                    "ins    v1.d[1], x21                \n"
-                    "fmla   v18.4s, v8.4s, v0.s[2]      \n"
-                    "ldr    d10, [%1], #8               \n"
-                    "fmla   v19.4s, v8.4s, v0.s[3]      \n"
-                    "ldr    x26, [%1], #8               \n"
-                    "fmla   v20.4s, v8.4s, v1.s[0]      \n"
-                    "ldr    d2, [%2], #8                \n"
-                    "fmla   v21.4s, v8.4s, v1.s[1]      \n"
-                    "ins    v9.d[1], x25                \n"
-                    "fmla   v22.4s, v8.4s, v1.s[2]      \n"
-                    "ldr    x22, [%2], #8               \n"
-                    "fmla   v23.4s, v8.4s, v1.s[3]      \n"
-                    "ldr    d3, [%2], #8                \n"
-                    "fmla   v24.4s, v9.4s, v0.s[0]      \n"
-                    "ldr    x23, [%2], #8               \n"
-                    "fmla   v25.4s, v9.4s, v0.s[1]      \n"
-                    "ins    v10.d[1], x26               \n"
-                    "fmla   v26.4s, v9.4s, v0.s[2]      \n"
-                    "ldr    d11, [%1], #8               \n"
-                    "fmla   v27.4s, v9.4s, v0.s[3]      \n"
-                    "ldr    x27, [%1], #8               \n"
-                    "fmla   v28.4s, v9.4s, v1.s[0]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                    "fmla   v29.4s, v9.4s, v1.s[1]      \n"
-                    "ins    v2.d[1], x22                \n"
-                    "fmla   v30.4s, v9.4s, v1.s[2]      \n"
-                    "ldr    d12, [%1], #8               \n"
-                    "fmla   v31.4s, v9.4s, v1.s[3]      \n"
-                    "ldr    x24, [%1], #8               \n"
-                    "fmla   v16.4s, v10.4s, v2.s[0]     \n"
-                    "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                    "fmla   v17.4s, v10.4s, v2.s[1]     \n"
-                    "ins    v3.d[1], x23                \n"
-                    "fmla   v18.4s, v10.4s, v2.s[2]     \n"
-                    "ldr    d4, [%2], #8                \n"
-                    "fmla   v19.4s, v10.4s, v2.s[3]     \n"
-                    "ldr    x20, [%2], #8               \n"
-                    "fmla   v20.4s, v10.4s, v3.s[0]     \n"
-                    "ldr    d5, [%2], #8                \n"
-                    "fmla   v21.4s, v10.4s, v3.s[1]     \n"
-                    "ins    v11.d[1], x27               \n"
-                    "fmla   v22.4s, v10.4s, v3.s[2]     \n"
-                    "ldr    x21, [%2], #8               \n"
-                    "fmla   v23.4s, v10.4s, v3.s[3]     \n"
-                    "ldr    d13, [%1], #8               \n"
-                    "fmla   v24.4s, v11.4s, v2.s[0]     \n"
-                    "ldr    x25, [%1], #8               \n"
-                    "fmla   v25.4s, v11.4s, v2.s[1]     \n"
-                    "ins    v12.d[1], x24               \n"
-                    "fmla   v26.4s, v11.4s, v2.s[2]     \n"
-                    "ldr    d14, [%1], #8               \n"
-                    "fmla   v27.4s, v11.4s, v2.s[3]     \n"
-                    "ldr    x26, [%1], #8               \n"
-                    "fmla   v28.4s, v11.4s, v3.s[0]     \n"
-                    "ldr    d6, [%2], #8                \n"
-                    "fmla   v29.4s, v11.4s, v3.s[1]     \n"
-                    "ins    v4.d[1], x20                \n"
-                    "fmla   v30.4s, v11.4s, v3.s[2]     \n"
-                    "ldr    x22, [%2], #8               \n"
-                    "fmla   v31.4s, v11.4s, v3.s[3]     \n"
-                    "ldr    d7, [%2], #8                \n"
-                    "fmla   v16.4s, v12.4s, v4.s[0]     \n"
-                    "ldr    x23, [%2], #8               \n"
-                    "fmla   v17.4s, v12.4s, v4.s[1]     \n"
-                    "ins    v5.d[1], x21                \n"
-                    "fmla   v18.4s, v12.4s, v4.s[2]     \n"
-                    "ldr    d15, [%1], #8               \n"
-                    "fmla   v19.4s, v12.4s, v4.s[3]     \n"
-                    "ldr    x27, [%1], #8               \n"
-                    "fmla   v20.4s, v12.4s, v5.s[0]     \n"
-                    "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                    "fmla   v21.4s, v12.4s, v5.s[1]     \n"
-                    "ins    v13.d[1], x25               \n"
-                    "fmla   v22.4s, v12.4s, v5.s[2]     \n"
-                    "ldr    d8, [%1], #8                \n"
-                    "fmla   v23.4s, v12.4s, v5.s[3]     \n"
-                    "ldr    x24, [%1], #8               \n"
-                    "fmla   v24.4s, v13.4s, v4.s[0]     \n"
-                    "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                    "fmla   v25.4s, v13.4s, v4.s[1]     \n"
-                    "ins    v14.d[1], x26               \n"
-                    "fmla   v26.4s, v13.4s, v4.s[2]     \n"
-                    "ldr    d0, [%2], #8                \n"
-                    "fmla   v27.4s, v13.4s, v4.s[3]     \n"
-                    "ldr    x20, [%2], #8               \n"
-                    "fmla   v28.4s, v13.4s, v5.s[0]     \n"
-                    "ldr    d1, [%2], #8                \n"
-                    "fmla   v29.4s, v13.4s, v5.s[1]     \n"
-                    "ins    v6.d[1], x22                \n"
-                    "fmla   v30.4s, v13.4s, v5.s[2]     \n"
-                    "ldr    x21, [%2], #8               \n"
-                    "fmla   v31.4s, v13.4s, v5.s[3]     \n"
-                    "fmla   v16.4s, v14.4s, v6.s[0]     \n"
-                    "fmla   v17.4s, v14.4s, v6.s[1]     \n"
-                    "ins    v7.d[1], x23                \n"
-                    "fmla   v18.4s, v14.4s, v6.s[2]     \n"
-                    "fmla   v19.4s, v14.4s, v6.s[3]     \n"
-                    "fmla   v20.4s, v14.4s, v7.s[0]     \n"
-                    "fmla   v21.4s, v14.4s, v7.s[1]     \n"
-                    "ins    v15.d[1], x27               \n"
-                    "fmla   v22.4s, v14.4s, v7.s[2]     \n"
-                    "fmla   v23.4s, v14.4s, v7.s[3]     \n"
-                    "fmla   v24.4s, v15.4s, v6.s[0]     \n"
-                    "fmla   v25.4s, v15.4s, v6.s[1]     \n"
-                    "ins    v8.d[1], x24                \n"
-                    "fmla   v26.4s, v15.4s, v6.s[2]     \n"
-                    "fmla   v27.4s, v15.4s, v6.s[3]     \n"
-                    "fmla   v28.4s, v15.4s, v7.s[0]     \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v29.4s, v15.4s, v7.s[1]     \n"
-                    "fmla   v30.4s, v15.4s, v7.s[2]     \n"
-                    "ins    v0.d[1], x20                \n"
-                    "fmla   v31.4s, v15.4s, v7.s[3]     \n"
-                    "bne    4b                          \n"
-
-                    "sub    %1, %1, #16                 \n"
-                    "sub    %2, %2, #32                 \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v0.4s, v1.4s}, [%2], #32   \n"
-                    "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-
-                    "fmla   v16.4s, v4.4s, v0.s[0]      \n"
-                    "fmla   v17.4s, v4.4s, v0.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v20.4s, v4.4s, v1.s[0]      \n"
-                    "fmla   v21.4s, v4.4s, v1.s[1]      \n"
-                    "fmla   v22.4s, v4.4s, v1.s[2]      \n"
-                    "fmla   v23.4s, v4.4s, v1.s[3]      \n"
-
-                    "subs   w4, w4, #1                  \n"
-
-                    "fmla   v24.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v0.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v1.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v1.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v1.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v1.s[3]      \n"
-
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    10f                         \n"
-
-                    // if out_elempack == 4
-                    "cmp    %w12, #4                    \n"
-                    "bne    8f                          \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 2          \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%3], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [x4], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [x4] \n"
-                    "b      9f                          \n"
-
-                    // if out_elempack == 1
-                    "8:                                 \n"
-                    // transpose8x8
-                    "zip1   v14.4s, v16.4s, v17.4s      \n"
-                    "zip2   v15.4s, v16.4s, v17.4s      \n"
-                    "zip1   v16.4s, v18.4s, v19.4s      \n"
-                    "zip2   v17.4s, v18.4s, v19.4s      \n"
-                    "zip1   v18.4s, v20.4s, v21.4s      \n"
-                    "zip2   v19.4s, v20.4s, v21.4s      \n"
-                    "zip1   v20.4s, v22.4s, v23.4s      \n"
-                    "zip2   v21.4s, v22.4s, v23.4s      \n"
-
-                    "zip1   v22.4s, v24.4s, v25.4s      \n"
-                    "zip2   v23.4s, v24.4s, v25.4s      \n"
-                    "zip1   v24.4s, v26.4s, v27.4s      \n"
-                    "zip2   v25.4s, v26.4s, v27.4s      \n"
-                    "zip1   v26.4s, v28.4s, v29.4s      \n"
-                    "zip2   v27.4s, v28.4s, v29.4s      \n"
-                    "zip1   v28.4s, v30.4s, v31.4s      \n"
-                    "zip2   v29.4s, v30.4s, v31.4s      \n"
-
-                    "zip1   v0.2d, v14.2d, v16.2d       \n"
-                    "zip2   v2.2d, v14.2d, v16.2d       \n"
-                    "zip1   v4.2d, v15.2d, v17.2d       \n"
-                    "zip2   v6.2d, v15.2d, v17.2d       \n"
-                    "zip1   v1.2d, v18.2d, v20.2d       \n"
-                    "zip2   v3.2d, v18.2d, v20.2d       \n"
-                    "zip1   v5.2d, v19.2d, v21.2d       \n"
-                    "zip2   v7.2d, v19.2d, v21.2d       \n"
-
-                    "zip1   v8.2d, v22.2d, v24.2d       \n"
-                    "zip2   v10.2d, v22.2d, v24.2d      \n"
-                    "zip1   v12.2d, v23.2d, v25.2d      \n"
-                    "zip2   v14.2d, v23.2d, v25.2d      \n"
-                    "zip1   v9.2d, v26.2d, v28.2d       \n"
-                    "zip2   v11.2d, v26.2d, v28.2d      \n"
-                    "zip1   v13.2d, v27.2d, v29.2d      \n"
-                    "zip2   v15.2d, v27.2d, v29.2d      \n"
-
-                    "add    x4, %3, %w13, sxtw 2        \n"
-                    "st1    {v0.4s, v1.4s}, [%3], #32   \n"
-                    "st1    {v2.4s, v3.4s}, [x4]        \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v4.4s, v5.4s}, [x4]        \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v6.4s, v7.4s}, [x4]        \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v8.4s, v9.4s}, [x4]        \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v10.4s, v11.4s}, [x4]      \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v12.4s, v13.4s}, [x4]      \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v14.4s, v15.4s}, [x4]      \n"
-
-                    "9:                                 \n"
-                    "add    %0, %0, #256                \n"
-                    "b      11f                         \n"
-
-                    "10:                                \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    "11:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            else if (use_a53_a55_optimized_kernel && !cpu_support_arm_asimdhp())
-            {
-                // a53
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #192                \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "add    x4, %8, #16                 \n"
-                    "ld1    {v16.4s}, [%8]              \n"
-                    "ld1    {v24.4s}, [x4]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v16.16b, v16.16b, v16.16b   \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v17.16b, v16.16b            \n"
-                    "mov    v18.16b, v16.16b            \n"
-                    "mov    v19.16b, v16.16b            \n"
-                    "mov    v20.16b, v16.16b            \n"
-                    "mov    v21.16b, v16.16b            \n"
-                    "mov    v22.16b, v16.16b            \n"
-                    "mov    v23.16b, v16.16b            \n"
-
-                    "mov    v25.16b, v24.16b            \n"
-                    "mov    v26.16b, v24.16b            \n"
-                    "mov    v27.16b, v24.16b            \n"
-                    "mov    v28.16b, v24.16b            \n"
-                    "mov    v29.16b, v24.16b            \n"
-                    "mov    v30.16b, v24.16b            \n"
-                    "mov    v31.16b, v24.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ldr    d0, [%2]                    \n"
-                    "ldr    x20, [%2, #8]               \n"
-                    "ins    v0.d[1], x20                \n"
-                    "add    %2, %2, #16                 \n"
-
-                    "ldr    d8, [%1]                    \n"
-                    "ldr    x24, [%1, #8]               \n"
-                    "ins    v8.d[1], x24                \n"
-                    "add    %1, %1, #16                 \n"
-
-                    "ldr    d1, [%2]                    \n"
-                    "ldr    x21, [%2, #8]               \n"
-                    "add    %2, %2, #16                 \n"
-
-                    "ldr    d9, [%1]                    \n"
-                    "ldr    x25, [%1, #8]               \n"
-                    "add    %1, %1, #16                 \n"
-
-                    ".align 4                           \n"
-                    "4:                                 \n"
-
-                    "ldr    d2, [%2]                    \n"
-                    "ins    v1.d[1], x21                \n"
-                    "fmla   v16.4s, v8.4s, v0.s[0]      \n"
-                    "ldr    x22, [%2, #8]               \n"
-                    "fmla   v17.4s, v8.4s, v0.s[1]      \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v18.4s, v8.4s, v0.s[2]      \n"
-
-                    "ldr    d10, [%1]                   \n"
-                    "ins    v9.d[1], x25                \n"
-                    "fmla   v19.4s, v8.4s, v0.s[3]      \n"
-                    "ldr    x26, [%1, #8]               \n"
-                    "fmla   v20.4s, v8.4s, v1.s[0]      \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v21.4s, v8.4s, v1.s[1]      \n"
-
-                    "ldr    d3, [%2]                    \n"
-                    "ins    v2.d[1], x22                \n"
-                    "fmla   v22.4s, v8.4s, v1.s[2]      \n"
-                    "ldr    x23, [%2, #8]               \n"
-                    "fmla   v23.4s, v8.4s, v1.s[3]      \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v24.4s, v9.4s, v0.s[0]      \n"
-
-                    "nop                                \n"
-                    "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                    "fmla   v25.4s, v9.4s, v0.s[1]      \n"
-                    "nop                                \n"
-                    "fmla   v26.4s, v9.4s, v0.s[2]      \n"
-                    "nop                                \n"
-                    "fmla   v27.4s, v9.4s, v0.s[3]      \n"
-
-                    "ldr    d11, [%1]                   \n"
-                    "ins    v10.d[1], x26               \n"
-                    "fmla   v28.4s, v9.4s, v1.s[0]      \n"
-                    "ldr    x27, [%1, #8]               \n"
-                    "fmla   v29.4s, v9.4s, v1.s[1]      \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v30.4s, v9.4s, v1.s[2]      \n"
-
-                    "nop                                \n"
-                    "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                    "fmla   v31.4s, v9.4s, v1.s[3]      \n"
-                    "nop                                \n"
-                    "fmla   v16.4s, v10.4s, v2.s[0]     \n"
-                    "nop                                \n"
-                    "fmla   v17.4s, v10.4s, v2.s[1]     \n"
-
-                    "ldr    d4, [%2]                    \n"
-                    "ins    v3.d[1], x23                \n"
-                    "fmla   v18.4s, v10.4s, v2.s[2]     \n"
-                    "ldr    x20, [%2, #8]               \n"
-                    "fmla   v19.4s, v10.4s, v2.s[3]     \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v20.4s, v10.4s, v3.s[0]     \n"
-
-                    "ldr    d12, [%1]                   \n"
-                    "ins    v11.d[1], x27               \n"
-                    "fmla   v21.4s, v10.4s, v3.s[1]     \n"
-                    "ldr    x24, [%1, #8]               \n"
-                    "fmla   v22.4s, v10.4s, v3.s[2]     \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v23.4s, v10.4s, v3.s[3]     \n"
-
-                    "ldr    d5, [%2]                    \n"
-                    "ins    v4.d[1], x20                \n"
-                    "fmla   v24.4s, v11.4s, v2.s[0]     \n"
-                    "ldr    x21, [%2, #8]               \n"
-                    "fmla   v25.4s, v11.4s, v2.s[1]     \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v26.4s, v11.4s, v2.s[2]     \n"
-
-                    "ldr    d13, [%1]                   \n"
-                    "ins    v12.d[1], x24               \n"
-                    "fmla   v27.4s, v11.4s, v2.s[3]     \n"
-                    "ldr    x25, [%1, #8]               \n"
-                    "fmla   v28.4s, v11.4s, v3.s[0]     \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v29.4s, v11.4s, v3.s[1]     \n"
-
-                    "ldr    d6, [%2]                    \n"
-                    "ins    v5.d[1], x21                \n"
-                    "fmla   v30.4s, v11.4s, v3.s[2]     \n"
-                    "ldr    x22, [%2, #8]               \n"
-                    "fmla   v31.4s, v11.4s, v3.s[3]     \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v16.4s, v12.4s, v4.s[0]     \n"
-
-                    "ldr    d14, [%1]                   \n"
-                    "ins    v13.d[1], x25               \n"
-                    "fmla   v17.4s, v12.4s, v4.s[1]     \n"
-                    "ldr    x26, [%1, #8]               \n"
-                    "fmla   v18.4s, v12.4s, v4.s[2]     \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v19.4s, v12.4s, v4.s[3]     \n"
-
-                    "ldr    d7, [%2]                    \n"
-                    "ins    v6.d[1], x22                \n"
-                    "fmla   v20.4s, v12.4s, v5.s[0]     \n"
-                    "ldr    x23, [%2, #8]               \n"
-                    "fmla   v21.4s, v12.4s, v5.s[1]     \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v22.4s, v12.4s, v5.s[2]     \n"
-
-                    "nop                                \n"
-                    "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                    "fmla   v23.4s, v12.4s, v5.s[3]     \n"
-                    "nop                                \n"
-                    "fmla   v24.4s, v13.4s, v4.s[0]     \n"
-                    "nop                                \n"
-                    "fmla   v25.4s, v13.4s, v4.s[1]     \n"
-
-                    "ldr    d15, [%1]                   \n"
-                    "ins    v14.d[1], x26               \n"
-                    "fmla   v26.4s, v13.4s, v4.s[2]     \n"
-                    "ldr    x27, [%1, #8]               \n"
-                    "fmla   v27.4s, v13.4s, v4.s[3]     \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v28.4s, v13.4s, v5.s[0]     \n"
-
-                    "nop                                \n"
-                    "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                    "fmla   v29.4s, v13.4s, v5.s[1]     \n"
-                    "nop                                \n"
-                    "fmla   v30.4s, v13.4s, v5.s[2]     \n"
-                    "nop                                \n"
-                    "fmla   v31.4s, v13.4s, v5.s[3]     \n"
-
-                    "ldr    d0, [%2]                    \n"
-                    "ins    v7.d[1], x23                \n"
-                    "fmla   v16.4s, v14.4s, v6.s[0]     \n"
-                    "ldr    x20, [%2, #8]               \n"
-                    "fmla   v17.4s, v14.4s, v6.s[1]     \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v18.4s, v14.4s, v6.s[2]     \n"
-
-                    "ldr    d8, [%1]                    \n"
-                    "ins    v15.d[1], x27               \n"
-                    "fmla   v19.4s, v14.4s, v6.s[3]     \n"
-                    "ldr    x24, [%1, #8]               \n"
-                    "fmla   v20.4s, v14.4s, v7.s[0]     \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v21.4s, v14.4s, v7.s[1]     \n"
-
-                    "ldr    d1, [%2]                    \n"
-                    "ins    v0.d[1], x20                \n"
-                    "fmla   v22.4s, v14.4s, v7.s[2]     \n"
-                    "ldr    x21, [%2, #8]               \n"
-                    "fmla   v23.4s, v14.4s, v7.s[3]     \n"
-                    "add    %2, %2, #16                 \n"
-                    "fmla   v24.4s, v15.4s, v6.s[0]     \n"
-
-                    "ldr    d9, [%1]                    \n"
-                    "ins    v8.d[1], x24                \n"
-                    "fmla   v25.4s, v15.4s, v6.s[1]     \n"
-                    "ldr    x25, [%1, #8]               \n"
-                    "fmla   v26.4s, v15.4s, v6.s[2]     \n"
-                    "add    %1, %1, #16                 \n"
-                    "fmla   v27.4s, v15.4s, v6.s[3]     \n"
-
-                    "nop                                \n"
-                    "nop                                \n"
-                    "fmla   v28.4s, v15.4s, v7.s[0]     \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v29.4s, v15.4s, v7.s[1]     \n"
-                    "nop                                \n"
-                    "fmla   v30.4s, v15.4s, v7.s[2]     \n"
-
-                    "nop                                \n"
-                    "nop                                \n"
-                    "fmla   v31.4s, v15.4s, v7.s[3]     \n"
-                    "nop                                \n"
-                    "nop                                \n"
-                    "nop                                \n"
-                    "nop                                \n"
-
-                    "bne    4b                          \n"
-
-                    "sub    %1, %1, #32                 \n"
-                    "sub    %2, %2, #32                 \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v0.4s, v1.4s}, [%2], #32   \n"
-                    "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-
-                    "fmla   v16.4s, v4.4s, v0.s[0]      \n"
-                    "fmla   v17.4s, v4.4s, v0.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v20.4s, v4.4s, v1.s[0]      \n"
-                    "fmla   v21.4s, v4.4s, v1.s[1]      \n"
-                    "fmla   v22.4s, v4.4s, v1.s[2]      \n"
-                    "fmla   v23.4s, v4.4s, v1.s[3]      \n"
-
-                    "subs   w4, w4, #1                  \n"
-
-                    "fmla   v24.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v0.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v1.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v1.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v1.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v1.s[3]      \n"
-
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    10f                         \n"
-
-                    // if out_elempack == 4
-                    "cmp    %w12, #4                    \n"
-                    "bne    8f                          \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 2          \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%3], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [x4], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [x4] \n"
-                    "b      9f                          \n"
-
-                    // if out_elempack == 1
-                    "8:                                 \n"
-                    // transpose8x8
-                    "zip1   v14.4s, v16.4s, v17.4s      \n"
-                    "zip2   v15.4s, v16.4s, v17.4s      \n"
-                    "zip1   v16.4s, v18.4s, v19.4s      \n"
-                    "zip2   v17.4s, v18.4s, v19.4s      \n"
-                    "zip1   v18.4s, v20.4s, v21.4s      \n"
-                    "zip2   v19.4s, v20.4s, v21.4s      \n"
-                    "zip1   v20.4s, v22.4s, v23.4s      \n"
-                    "zip2   v21.4s, v22.4s, v23.4s      \n"
-
-                    "zip1   v22.4s, v24.4s, v25.4s      \n"
-                    "zip2   v23.4s, v24.4s, v25.4s      \n"
-                    "zip1   v24.4s, v26.4s, v27.4s      \n"
-                    "zip2   v25.4s, v26.4s, v27.4s      \n"
-                    "zip1   v26.4s, v28.4s, v29.4s      \n"
-                    "zip2   v27.4s, v28.4s, v29.4s      \n"
-                    "zip1   v28.4s, v30.4s, v31.4s      \n"
-                    "zip2   v29.4s, v30.4s, v31.4s      \n"
-
-                    "zip1   v0.2d, v14.2d, v16.2d       \n"
-                    "zip2   v2.2d, v14.2d, v16.2d       \n"
-                    "zip1   v4.2d, v15.2d, v17.2d       \n"
-                    "zip2   v6.2d, v15.2d, v17.2d       \n"
-                    "zip1   v1.2d, v18.2d, v20.2d       \n"
-                    "zip2   v3.2d, v18.2d, v20.2d       \n"
-                    "zip1   v5.2d, v19.2d, v21.2d       \n"
-                    "zip2   v7.2d, v19.2d, v21.2d       \n"
-
-                    "zip1   v8.2d, v22.2d, v24.2d       \n"
-                    "zip2   v10.2d, v22.2d, v24.2d      \n"
-                    "zip1   v12.2d, v23.2d, v25.2d      \n"
-                    "zip2   v14.2d, v23.2d, v25.2d      \n"
-                    "zip1   v9.2d, v26.2d, v28.2d       \n"
-                    "zip2   v11.2d, v26.2d, v28.2d      \n"
-                    "zip1   v13.2d, v27.2d, v29.2d      \n"
-                    "zip2   v15.2d, v27.2d, v29.2d      \n"
-
-                    "add    x4, %3, %w13, sxtw 2        \n"
-                    "st1    {v0.4s, v1.4s}, [%3], #32   \n"
-                    "st1    {v2.4s, v3.4s}, [x4]        \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v4.4s, v5.4s}, [x4]        \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v6.4s, v7.4s}, [x4]        \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v8.4s, v9.4s}, [x4]        \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v10.4s, v11.4s}, [x4]      \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v12.4s, v13.4s}, [x4]      \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v14.4s, v15.4s}, [x4]      \n"
-
-                    "9:                                 \n"
-                    "add    %0, %0, #256                \n"
-                    "b      11f                         \n"
-
-                    "10:                                \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    "11:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            else
-            {
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #192                \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "add    x4, %8, #16                 \n"
-                    "ld1    {v16.4s}, [%8]              \n"
-                    "ld1    {v24.4s}, [x4]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v16.16b, v16.16b, v16.16b   \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v17.16b, v16.16b            \n"
-                    "mov    v18.16b, v16.16b            \n"
-                    "mov    v19.16b, v16.16b            \n"
-                    "mov    v20.16b, v16.16b            \n"
-                    "mov    v21.16b, v16.16b            \n"
-                    "mov    v22.16b, v16.16b            \n"
-                    "mov    v23.16b, v16.16b            \n"
-
-                    "mov    v25.16b, v24.16b            \n"
-                    "mov    v26.16b, v24.16b            \n"
-                    "mov    v27.16b, v24.16b            \n"
-                    "mov    v28.16b, v24.16b            \n"
-                    "mov    v29.16b, v24.16b            \n"
-                    "mov    v30.16b, v24.16b            \n"
-                    "mov    v31.16b, v24.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%1], #64 \n"
-                    "fmla   v16.4s, v8.4s, v0.s[0]      \n"
-                    "fmla   v17.4s, v8.4s, v0.s[1]      \n"
-                    "fmla   v18.4s, v8.4s, v0.s[2]      \n"
-                    "fmla   v19.4s, v8.4s, v0.s[3]      \n"
-                    "fmla   v20.4s, v8.4s, v1.s[0]      \n"
-                    "fmla   v21.4s, v8.4s, v1.s[1]      \n"
-                    "fmla   v22.4s, v8.4s, v1.s[2]      \n"
-                    "fmla   v23.4s, v8.4s, v1.s[3]      \n"
-                    "fmla   v24.4s, v9.4s, v0.s[0]      \n"
-                    "fmla   v25.4s, v9.4s, v0.s[1]      \n"
-                    "fmla   v26.4s, v9.4s, v0.s[2]      \n"
-                    "fmla   v27.4s, v9.4s, v0.s[3]      \n"
-                    "fmla   v28.4s, v9.4s, v1.s[0]      \n"
-                    "fmla   v29.4s, v9.4s, v1.s[1]      \n"
-                    "fmla   v30.4s, v9.4s, v1.s[2]      \n"
-                    "fmla   v31.4s, v9.4s, v1.s[3]      \n"
-                    "fmla   v16.4s, v10.4s, v2.s[0]     \n"
-                    "fmla   v17.4s, v10.4s, v2.s[1]     \n"
-                    "fmla   v18.4s, v10.4s, v2.s[2]     \n"
-                    "fmla   v19.4s, v10.4s, v2.s[3]     \n"
-                    "fmla   v20.4s, v10.4s, v3.s[0]     \n"
-                    "fmla   v21.4s, v10.4s, v3.s[1]     \n"
-                    "fmla   v22.4s, v10.4s, v3.s[2]     \n"
-                    "fmla   v23.4s, v10.4s, v3.s[3]     \n"
-                    "fmla   v24.4s, v11.4s, v2.s[0]     \n"
-                    "fmla   v25.4s, v11.4s, v2.s[1]     \n"
-                    "fmla   v26.4s, v11.4s, v2.s[2]     \n"
-                    "fmla   v27.4s, v11.4s, v2.s[3]     \n"
-                    "fmla   v28.4s, v11.4s, v3.s[0]     \n"
-                    "fmla   v29.4s, v11.4s, v3.s[1]     \n"
-                    "fmla   v30.4s, v11.4s, v3.s[2]     \n"
-                    "fmla   v31.4s, v11.4s, v3.s[3]     \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%2], #64 \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%1], #64 \n"
-                    "fmla   v16.4s, v12.4s, v4.s[0]     \n"
-                    "fmla   v17.4s, v12.4s, v4.s[1]     \n"
-                    "fmla   v18.4s, v12.4s, v4.s[2]     \n"
-                    "fmla   v19.4s, v12.4s, v4.s[3]     \n"
-                    "fmla   v20.4s, v12.4s, v5.s[0]     \n"
-                    "fmla   v21.4s, v12.4s, v5.s[1]     \n"
-                    "fmla   v22.4s, v12.4s, v5.s[2]     \n"
-                    "fmla   v23.4s, v12.4s, v5.s[3]     \n"
-                    "fmla   v24.4s, v13.4s, v4.s[0]     \n"
-                    "fmla   v25.4s, v13.4s, v4.s[1]     \n"
-                    "fmla   v26.4s, v13.4s, v4.s[2]     \n"
-                    "fmla   v27.4s, v13.4s, v4.s[3]     \n"
-                    "fmla   v28.4s, v13.4s, v5.s[0]     \n"
-                    "fmla   v29.4s, v13.4s, v5.s[1]     \n"
-                    "fmla   v30.4s, v13.4s, v5.s[2]     \n"
-                    "fmla   v31.4s, v13.4s, v5.s[3]     \n"
-                    "fmla   v16.4s, v14.4s, v6.s[0]     \n"
-                    "fmla   v17.4s, v14.4s, v6.s[1]     \n"
-                    "fmla   v18.4s, v14.4s, v6.s[2]     \n"
-                    "fmla   v19.4s, v14.4s, v6.s[3]     \n"
-                    "fmla   v20.4s, v14.4s, v7.s[0]     \n"
-                    "fmla   v21.4s, v14.4s, v7.s[1]     \n"
-                    "fmla   v22.4s, v14.4s, v7.s[2]     \n"
-                    "fmla   v23.4s, v14.4s, v7.s[3]     \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v24.4s, v15.4s, v6.s[0]     \n"
-                    "fmla   v25.4s, v15.4s, v6.s[1]     \n"
-                    "fmla   v26.4s, v15.4s, v6.s[2]     \n"
-                    "fmla   v27.4s, v15.4s, v6.s[3]     \n"
-                    "fmla   v28.4s, v15.4s, v7.s[0]     \n"
-                    "fmla   v29.4s, v15.4s, v7.s[1]     \n"
-                    "fmla   v30.4s, v15.4s, v7.s[2]     \n"
-                    "fmla   v31.4s, v15.4s, v7.s[3]     \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v0.4s, v1.4s}, [%2], #32   \n"
-                    "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-
-                    "fmla   v16.4s, v4.4s, v0.s[0]      \n"
-                    "fmla   v17.4s, v4.4s, v0.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v20.4s, v4.4s, v1.s[0]      \n"
-                    "fmla   v21.4s, v4.4s, v1.s[1]      \n"
-                    "fmla   v22.4s, v4.4s, v1.s[2]      \n"
-                    "fmla   v23.4s, v4.4s, v1.s[3]      \n"
-
-                    "subs   w4, w4, #1                  \n"
-
-                    "fmla   v24.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v0.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v1.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v1.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v1.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v1.s[3]      \n"
-
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    10f                         \n"
-
-                    // if out_elempack == 4
-                    "cmp    %w12, #4                    \n"
-                    "bne    8f                          \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 2          \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%3], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [x4], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [x4] \n"
-                    "b      9f                          \n"
-
-                    // if out_elempack == 1
-                    "8:                                 \n"
-                    // transpose8x8
-                    "zip1   v14.4s, v16.4s, v17.4s      \n"
-                    "zip2   v15.4s, v16.4s, v17.4s      \n"
-                    "zip1   v16.4s, v18.4s, v19.4s      \n"
-                    "zip2   v17.4s, v18.4s, v19.4s      \n"
-                    "zip1   v18.4s, v20.4s, v21.4s      \n"
-                    "zip2   v19.4s, v20.4s, v21.4s      \n"
-                    "zip1   v20.4s, v22.4s, v23.4s      \n"
-                    "zip2   v21.4s, v22.4s, v23.4s      \n"
-
-                    "zip1   v22.4s, v24.4s, v25.4s      \n"
-                    "zip2   v23.4s, v24.4s, v25.4s      \n"
-                    "zip1   v24.4s, v26.4s, v27.4s      \n"
-                    "zip2   v25.4s, v26.4s, v27.4s      \n"
-                    "zip1   v26.4s, v28.4s, v29.4s      \n"
-                    "zip2   v27.4s, v28.4s, v29.4s      \n"
-                    "zip1   v28.4s, v30.4s, v31.4s      \n"
-                    "zip2   v29.4s, v30.4s, v31.4s      \n"
-
-                    "zip1   v0.2d, v14.2d, v16.2d       \n"
-                    "zip2   v2.2d, v14.2d, v16.2d       \n"
-                    "zip1   v4.2d, v15.2d, v17.2d       \n"
-                    "zip2   v6.2d, v15.2d, v17.2d       \n"
-                    "zip1   v1.2d, v18.2d, v20.2d       \n"
-                    "zip2   v3.2d, v18.2d, v20.2d       \n"
-                    "zip1   v5.2d, v19.2d, v21.2d       \n"
-                    "zip2   v7.2d, v19.2d, v21.2d       \n"
-
-                    "zip1   v8.2d, v22.2d, v24.2d       \n"
-                    "zip2   v10.2d, v22.2d, v24.2d      \n"
-                    "zip1   v12.2d, v23.2d, v25.2d      \n"
-                    "zip2   v14.2d, v23.2d, v25.2d      \n"
-                    "zip1   v9.2d, v26.2d, v28.2d       \n"
-                    "zip2   v11.2d, v26.2d, v28.2d      \n"
-                    "zip1   v13.2d, v27.2d, v29.2d      \n"
-                    "zip2   v15.2d, v27.2d, v29.2d      \n"
-
-                    "add    x4, %3, %w13, sxtw 2        \n"
-                    "st1    {v0.4s, v1.4s}, [%3], #32   \n"
-                    "st1    {v2.4s, v3.4s}, [x4]        \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v4.4s, v5.4s}, [x4]        \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v6.4s, v7.4s}, [x4]        \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v8.4s, v9.4s}, [x4]        \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v10.4s, v11.4s}, [x4]      \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v12.4s, v13.4s}, [x4]      \n"
-                    "add    x4, x4, %w13, sxtw 2        \n"
-                    "st1    {v14.4s, v15.4s}, [x4]      \n"
-
-                    "9:                                 \n"
-                    "add    %0, %0, #256                \n"
-                    "b      11f                         \n"
-
-                    "10:                                \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    "11:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum20;
-            float32x4_t _sum21;
-            float32x4_t _sum30;
-            float32x4_t _sum31;
-            float32x4_t _sum40;
-            float32x4_t _sum41;
-            float32x4_t _sum50;
-            float32x4_t _sum51;
-            float32x4_t _sum60;
-            float32x4_t _sum61;
-            float32x4_t _sum70;
-            float32x4_t _sum71;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vld1q_f32(pC);
-                    _sum01 = vld1q_f32(pC + 4);
-                    _sum10 = _sum00;
-                    _sum11 = _sum01;
-                    _sum20 = _sum00;
-                    _sum21 = _sum01;
-                    _sum30 = _sum00;
-                    _sum31 = _sum01;
-                    _sum40 = _sum00;
-                    _sum41 = _sum01;
-                    _sum50 = _sum00;
-                    _sum51 = _sum01;
-                    _sum60 = _sum00;
-                    _sum61 = _sum01;
-                    _sum70 = _sum00;
-                    _sum71 = _sum01;
-                }
-                else
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                    _sum20 = vdupq_n_f32(0.f);
-                    _sum21 = vdupq_n_f32(0.f);
-                    _sum30 = vdupq_n_f32(0.f);
-                    _sum31 = vdupq_n_f32(0.f);
-                    _sum40 = vdupq_n_f32(0.f);
-                    _sum41 = vdupq_n_f32(0.f);
-                    _sum50 = vdupq_n_f32(0.f);
-                    _sum51 = vdupq_n_f32(0.f);
-                    _sum60 = vdupq_n_f32(0.f);
-                    _sum61 = vdupq_n_f32(0.f);
-                    _sum70 = vdupq_n_f32(0.f);
-                    _sum71 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-                _sum20 = vld1q_f32(outptr + 4 * 4);
-                _sum21 = vld1q_f32(outptr + 4 * 5);
-                _sum30 = vld1q_f32(outptr + 4 * 6);
-                _sum31 = vld1q_f32(outptr + 4 * 7);
-                _sum40 = vld1q_f32(outptr + 4 * 8);
-                _sum41 = vld1q_f32(outptr + 4 * 9);
-                _sum50 = vld1q_f32(outptr + 4 * 10);
-                _sum51 = vld1q_f32(outptr + 4 * 11);
-                _sum60 = vld1q_f32(outptr + 4 * 12);
-                _sum61 = vld1q_f32(outptr + 4 * 13);
-                _sum70 = vld1q_f32(outptr + 4 * 14);
-                _sum71 = vld1q_f32(outptr + 4 * 15);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA0 = vld1q_f32(pA);
-                float32x4_t _pA1 = vld1q_f32(pA + 4);
-
-                float32x4_t _pB0 = vld1q_f32(pB);
-                float32x4_t _pB1 = vld1q_f32(pB + 4);
-
-                _sum00 = vfmaq_laneq_f32(_sum00, _pA0, _pB0, 0);
-                _sum01 = vfmaq_laneq_f32(_sum01, _pA1, _pB0, 0);
-                _sum10 = vfmaq_laneq_f32(_sum10, _pA0, _pB0, 1);
-                _sum11 = vfmaq_laneq_f32(_sum11, _pA1, _pB0, 1);
-                _sum20 = vfmaq_laneq_f32(_sum20, _pA0, _pB0, 2);
-                _sum21 = vfmaq_laneq_f32(_sum21, _pA1, _pB0, 2);
-                _sum30 = vfmaq_laneq_f32(_sum30, _pA0, _pB0, 3);
-                _sum31 = vfmaq_laneq_f32(_sum31, _pA1, _pB0, 3);
-                _sum40 = vfmaq_laneq_f32(_sum40, _pA0, _pB1, 0);
-                _sum41 = vfmaq_laneq_f32(_sum41, _pA1, _pB1, 0);
-                _sum50 = vfmaq_laneq_f32(_sum50, _pA0, _pB1, 1);
-                _sum51 = vfmaq_laneq_f32(_sum51, _pA1, _pB1, 1);
-                _sum60 = vfmaq_laneq_f32(_sum60, _pA0, _pB1, 2);
-                _sum61 = vfmaq_laneq_f32(_sum61, _pA1, _pB1, 2);
-                _sum70 = vfmaq_laneq_f32(_sum70, _pA0, _pB1, 3);
-                _sum71 = vfmaq_laneq_f32(_sum71, _pA1, _pB1, 3);
-
-                pA += 8;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1q_f32(outptr0, _sum00);
-                    vst1q_f32(outptr0 + 4, _sum10);
-                    vst1q_f32(outptr0 + 4 * 2, _sum20);
-                    vst1q_f32(outptr0 + 4 * 3, _sum30);
-                    vst1q_f32(outptr0 + 4 * 4, _sum40);
-                    vst1q_f32(outptr0 + 4 * 5, _sum50);
-                    vst1q_f32(outptr0 + 4 * 6, _sum60);
-                    vst1q_f32(outptr0 + 4 * 7, _sum70);
-
-                    vst1q_f32(outptr0 + out_hstep * 4, _sum01);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4, _sum11);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 2, _sum21);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 3, _sum31);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 4, _sum41);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 5, _sum51);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 6, _sum61);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 7, _sum71);
-
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose8x8_ps(_sum00, _sum01, _sum10, _sum11, _sum20, _sum21, _sum30, _sum31, _sum40, _sum41, _sum50, _sum51, _sum60, _sum61, _sum70, _sum71);
-
-                    vst1q_f32(outptr0, _sum00);
-                    vst1q_f32(outptr0 + 4, _sum01);
-                    vst1q_f32(outptr0 + out_hstep, _sum10);
-                    vst1q_f32(outptr0 + out_hstep + 4, _sum11);
-                    vst1q_f32(outptr0 + out_hstep * 2, _sum20);
-                    vst1q_f32(outptr0 + out_hstep * 2 + 4, _sum21);
-                    vst1q_f32(outptr0 + out_hstep * 3, _sum30);
-                    vst1q_f32(outptr0 + out_hstep * 3 + 4, _sum31);
-                    vst1q_f32(outptr0 + out_hstep * 4, _sum40);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4, _sum41);
-                    vst1q_f32(outptr0 + out_hstep * 5, _sum50);
-                    vst1q_f32(outptr0 + out_hstep * 5 + 4, _sum51);
-                    vst1q_f32(outptr0 + out_hstep * 6, _sum60);
-                    vst1q_f32(outptr0 + out_hstep * 6 + 4, _sum61);
-                    vst1q_f32(outptr0 + out_hstep * 7, _sum70);
-                    vst1q_f32(outptr0 + out_hstep * 7 + 4, _sum71);
-
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                vst1q_f32(outptr + 4 * 4, _sum20);
-                vst1q_f32(outptr + 4 * 5, _sum21);
-                vst1q_f32(outptr + 4 * 6, _sum30);
-                vst1q_f32(outptr + 4 * 7, _sum31);
-                vst1q_f32(outptr + 4 * 8, _sum40);
-                vst1q_f32(outptr + 4 * 9, _sum41);
-                vst1q_f32(outptr + 4 * 10, _sum50);
-                vst1q_f32(outptr + 4 * 11, _sum51);
-                vst1q_f32(outptr + 4 * 12, _sum60);
-                vst1q_f32(outptr + 4 * 13, _sum61);
-                vst1q_f32(outptr + 4 * 14, _sum70);
-                vst1q_f32(outptr + 4 * 15, _sum71);
-            }
-
-            outptr += 64;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                "subs   %0, %0, #64                 \n"
-                "b      3f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "add    x4, %8, #16                 \n"
-                "ld1    {v24.4s}, [%8]              \n"
-                "ld1    {v28.4s}, [x4]              \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v24.16b, v24.16b, v24.16b   \n"
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-
-                "2:                                 \n"
-                "mov    v25.16b, v24.16b            \n"
-                "mov    v26.16b, v24.16b            \n"
-                "mov    v27.16b, v24.16b            \n"
-
-                "mov    v29.16b, v28.16b            \n"
-                "mov    v30.16b, v28.16b            \n"
-                "mov    v31.16b, v28.16b            \n"
-
-                "3:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                "4:                                 \n"
-                "prfm   pldl1keep, [%2, #512]       \n"
-                "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n"
-                "fmla   v24.4s, v4.4s, v0.s[0]      \n"
-                "fmla   v25.4s, v4.4s, v0.s[1]      \n"
-                "fmla   v26.4s, v4.4s, v0.s[2]      \n"
-                "fmla   v27.4s, v4.4s, v0.s[3]      \n"
-                "fmla   v28.4s, v5.4s, v0.s[0]      \n"
-                "fmla   v29.4s, v5.4s, v0.s[1]      \n"
-                "fmla   v30.4s, v5.4s, v0.s[2]      \n"
-                "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-                "fmla   v24.4s, v6.4s, v1.s[0]      \n"
-                "fmla   v25.4s, v6.4s, v1.s[1]      \n"
-                "fmla   v26.4s, v6.4s, v1.s[2]      \n"
-                "fmla   v27.4s, v6.4s, v1.s[3]      \n"
-                "fmla   v28.4s, v7.4s, v1.s[0]      \n"
-                "fmla   v29.4s, v7.4s, v1.s[1]      \n"
-                "fmla   v30.4s, v7.4s, v1.s[2]      \n"
-                "fmla   v31.4s, v7.4s, v1.s[3]      \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%1], #64 \n"
-                "fmla   v24.4s, v8.4s, v2.s[0]      \n"
-                "fmla   v25.4s, v8.4s, v2.s[1]      \n"
-                "fmla   v26.4s, v8.4s, v2.s[2]      \n"
-                "fmla   v27.4s, v8.4s, v2.s[3]      \n"
-                "fmla   v28.4s, v9.4s, v2.s[0]      \n"
-                "fmla   v29.4s, v9.4s, v2.s[1]      \n"
-                "fmla   v30.4s, v9.4s, v2.s[2]      \n"
-                "fmla   v31.4s, v9.4s, v2.s[3]      \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v24.4s, v10.4s, v3.s[0]     \n"
-                "fmla   v25.4s, v10.4s, v3.s[1]     \n"
-                "fmla   v26.4s, v10.4s, v3.s[2]     \n"
-                "fmla   v27.4s, v10.4s, v3.s[3]     \n"
-                "fmla   v28.4s, v11.4s, v3.s[0]     \n"
-                "fmla   v29.4s, v11.4s, v3.s[1]     \n"
-                "fmla   v30.4s, v11.4s, v3.s[2]     \n"
-                "fmla   v31.4s, v11.4s, v3.s[3]     \n"
-                "bne    4b                          \n"
-
-                "5:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    7f                          \n"
-
-                "6:                                 \n"
-                "ld1    {v0.4s}, [%2], #16          \n"
-                "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-                "fmla   v24.4s, v4.4s, v0.s[0]      \n"
-                "fmla   v25.4s, v4.4s, v0.s[1]      \n"
-                "fmla   v26.4s, v4.4s, v0.s[2]      \n"
-                "fmla   v27.4s, v4.4s, v0.s[3]      \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v28.4s, v5.4s, v0.s[0]      \n"
-                "fmla   v29.4s, v5.4s, v0.s[1]      \n"
-                "fmla   v30.4s, v5.4s, v0.s[2]      \n"
-                "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-                "bne    6b                          \n"
-
-                "7:                                 \n"
-                "tst    %w11, #255                  \n"
-                "beq    10f                         \n"
-
-                // if out_elempack == 4
-                "cmp    %w12, #4                    \n"
-                "bne    8f                          \n"
-
-                "lsl    w4, %w13, #2                \n"
-                "add    x4, %3, w4, sxtw 2          \n"
-                "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%3], #64 \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [x4] \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // transpose8x4
-                "zip1   v22.4s, v24.4s, v25.4s      \n"
-                "zip2   v23.4s, v24.4s, v25.4s      \n"
-                "zip1   v24.4s, v26.4s, v27.4s      \n"
-                "zip2   v25.4s, v26.4s, v27.4s      \n"
-                "zip1   v26.4s, v28.4s, v29.4s      \n"
-                "zip2   v27.4s, v28.4s, v29.4s      \n"
-                "zip1   v28.4s, v30.4s, v31.4s      \n"
-                "zip2   v29.4s, v30.4s, v31.4s      \n"
-
-                "zip1   v0.2d, v22.2d, v24.2d       \n"
-                "zip2   v1.2d, v22.2d, v24.2d       \n"
-                "zip1   v2.2d, v23.2d, v25.2d       \n"
-                "zip2   v3.2d, v23.2d, v25.2d       \n"
-                "zip1   v4.2d, v26.2d, v28.2d       \n"
-                "zip2   v5.2d, v26.2d, v28.2d       \n"
-                "zip1   v6.2d, v27.2d, v29.2d       \n"
-                "zip2   v7.2d, v27.2d, v29.2d       \n"
-
-                "add    x4, %3, %w13, sxtw 2        \n"
-                "st1    {v0.4s}, [%3], #16          \n"
-                "st1    {v1.4s}, [x4]               \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v2.4s}, [x4]               \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v3.4s}, [x4]               \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v4.4s}, [x4]               \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v5.4s}, [x4]               \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v6.4s}, [x4]               \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v7.4s}, [x4]               \n"
-
-                "9:                                 \n"
-                "add    %0, %0, #128                \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum20;
-            float32x4_t _sum21;
-            float32x4_t _sum30;
-            float32x4_t _sum31;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vld1q_f32(pC);
-                    _sum01 = vld1q_f32(pC + 4);
-                    _sum10 = _sum00;
-                    _sum11 = _sum01;
-                    _sum20 = _sum00;
-                    _sum21 = _sum01;
-                    _sum30 = _sum00;
-                    _sum31 = _sum01;
-                }
-                else
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                    _sum20 = vdupq_n_f32(0.f);
-                    _sum21 = vdupq_n_f32(0.f);
-                    _sum30 = vdupq_n_f32(0.f);
-                    _sum31 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-                _sum20 = vld1q_f32(outptr + 4 * 4);
-                _sum21 = vld1q_f32(outptr + 4 * 5);
-                _sum30 = vld1q_f32(outptr + 4 * 6);
-                _sum31 = vld1q_f32(outptr + 4 * 7);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA0 = vld1q_f32(pA);
-                float32x4_t _pA1 = vld1q_f32(pA + 4);
-
-                float32x4_t _pB0 = vld1q_f32(pB);
-
-                _sum00 = vfmaq_laneq_f32(_sum00, _pA0, _pB0, 0);
-                _sum01 = vfmaq_laneq_f32(_sum01, _pA1, _pB0, 0);
-                _sum10 = vfmaq_laneq_f32(_sum10, _pA0, _pB0, 1);
-                _sum11 = vfmaq_laneq_f32(_sum11, _pA1, _pB0, 1);
-                _sum20 = vfmaq_laneq_f32(_sum20, _pA0, _pB0, 2);
-                _sum21 = vfmaq_laneq_f32(_sum21, _pA1, _pB0, 2);
-                _sum30 = vfmaq_laneq_f32(_sum30, _pA0, _pB0, 3);
-                _sum31 = vfmaq_laneq_f32(_sum31, _pA1, _pB0, 3);
-
-                pA += 8;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1q_f32(outptr0, _sum00);
-                    vst1q_f32(outptr0 + 4, _sum10);
-                    vst1q_f32(outptr0 + 4 * 2, _sum20);
-                    vst1q_f32(outptr0 + 4 * 3, _sum30);
-
-                    vst1q_f32(outptr0 + out_hstep * 4, _sum01);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4, _sum11);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 2, _sum21);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4 * 3, _sum31);
-
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose8x4_ps(_sum00, _sum01, _sum10, _sum11, _sum20, _sum21, _sum30, _sum31);
-
-                    vst1q_f32(outptr0, _sum00);
-                    vst1q_f32(outptr0 + out_hstep * 1, _sum01);
-                    vst1q_f32(outptr0 + out_hstep * 2, _sum10);
-                    vst1q_f32(outptr0 + out_hstep * 3, _sum11);
-                    vst1q_f32(outptr0 + out_hstep * 4, _sum20);
-                    vst1q_f32(outptr0 + out_hstep * 5, _sum21);
-                    vst1q_f32(outptr0 + out_hstep * 6, _sum30);
-                    vst1q_f32(outptr0 + out_hstep * 7, _sum31);
-
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                vst1q_f32(outptr + 4 * 4, _sum20);
-                vst1q_f32(outptr + 4 * 5, _sum21);
-                vst1q_f32(outptr + 4 * 6, _sum30);
-                vst1q_f32(outptr + 4 * 7, _sum31);
-            }
-
-            outptr += 32;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                "b      3f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "add    x4, %8, #16                 \n"
-                "ld1    {v28.4s}, [%8]              \n"
-                "ld1    {v30.4s}, [x4]              \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-                "eor    v30.16b, v30.16b, v30.16b   \n"
-
-                "2:                                 \n"
-                "mov    v29.16b, v28.16b            \n"
-                "mov    v31.16b, v30.16b            \n"
-
-                "3:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                "4:                                 \n"
-                "prfm   pldl1keep, [%2, #256]       \n"
-                "ld1    {v0.4s, v1.4s}, [%2], #32   \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n"
-                "fmla   v28.4s, v4.4s, v0.s[0]      \n"
-                "fmla   v29.4s, v4.4s, v0.s[1]      \n"
-                "fmla   v30.4s, v5.4s, v0.s[0]      \n"
-                "fmla   v31.4s, v5.4s, v0.s[1]      \n"
-                "fmla   v28.4s, v6.4s, v0.s[2]      \n"
-                "fmla   v29.4s, v6.4s, v0.s[3]      \n"
-                "fmla   v30.4s, v7.4s, v0.s[2]      \n"
-                "fmla   v31.4s, v7.4s, v0.s[3]      \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%1], #64 \n"
-                "fmla   v28.4s, v8.4s, v1.s[0]      \n"
-                "fmla   v29.4s, v8.4s, v1.s[1]      \n"
-                "fmla   v30.4s, v9.4s, v1.s[0]      \n"
-                "fmla   v31.4s, v9.4s, v1.s[1]      \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v28.4s, v10.4s, v1.s[2]     \n"
-                "fmla   v29.4s, v10.4s, v1.s[3]     \n"
-                "fmla   v30.4s, v11.4s, v1.s[2]     \n"
-                "fmla   v31.4s, v11.4s, v1.s[3]     \n"
-                "bne    4b                          \n"
-
-                "5:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    7f                          \n"
-
-                "6:                                 \n"
-                "ld1    {v0.2s}, [%2], #8           \n"
-                "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-                "fmla   v28.4s, v4.4s, v0.s[0]      \n"
-                "fmla   v29.4s, v4.4s, v0.s[1]      \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v30.4s, v5.4s, v0.s[0]      \n"
-                "fmla   v31.4s, v5.4s, v0.s[1]      \n"
-                "bne    6b                          \n"
-
-                "7:                                 \n"
-                "tst    %w11, #255                  \n"
-                "beq    10f                         \n"
-
-                // if out_elempack == 4
-                "cmp    %w12, #4                    \n"
-                "bne    8f                          \n"
-
-                "lsl    w4, %w13, #2                \n"
-                "add    x4, %3, w4, sxtw 2          \n"
-                "st1    {v28.4s, v29.4s}, [%3], #32 \n"
-                "st1    {v30.4s, v31.4s}, [x4]      \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // transpose8x2
-                "zip1   v0.4s, v28.4s, v29.4s       \n"
-                "zip2   v2.4s, v28.4s, v29.4s       \n"
-                "zip1   v4.4s, v30.4s, v31.4s       \n"
-                "zip2   v6.4s, v30.4s, v31.4s       \n"
-
-                "mov    v1.d[0], v0.d[1]            \n"
-                "mov    v3.d[0], v2.d[1]            \n"
-                "mov    v5.d[0], v4.d[1]            \n"
-                "mov    v7.d[0], v6.d[1]            \n"
-
-                "add    x4, %3, %w13, sxtw 2        \n"
-                "st1    {v0.2s}, [%3], #8           \n"
-                "st1    {v1.2s}, [x4]               \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v2.2s}, [x4]               \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v3.2s}, [x4]               \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v4.2s}, [x4]               \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v5.2s}, [x4]               \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v6.2s}, [x4]               \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v7.2s}, [x4]               \n"
-
-                "9:                                 \n"
-                "add    %0, %0, #64                 \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vld1q_f32(pC);
-                    _sum01 = vld1q_f32(pC + 4);
-                    _sum10 = _sum00;
-                    _sum11 = _sum01;
-                }
-                else
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA0 = vld1q_f32(pA);
-                float32x4_t _pA1 = vld1q_f32(pA + 4);
-
-                float32x2_t _pB0 = vld1_f32(pB);
-
-                _sum00 = vfmaq_lane_f32(_sum00, _pA0, _pB0, 0);
-                _sum01 = vfmaq_lane_f32(_sum01, _pA1, _pB0, 0);
-                _sum10 = vfmaq_lane_f32(_sum10, _pA0, _pB0, 1);
-                _sum11 = vfmaq_lane_f32(_sum11, _pA1, _pB0, 1);
-
-                pA += 8;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1q_f32(outptr0, _sum00);
-                    vst1q_f32(outptr0 + 4, _sum10);
-
-                    vst1q_f32(outptr0 + out_hstep * 4, _sum01);
-                    vst1q_f32(outptr0 + out_hstep * 4 + 4, _sum11);
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-                    float sum0[8];
-                    float sum1[8];
-                    vst1q_f32(sum0, _sum00);
-                    vst1q_f32(sum0 + 4, _sum01);
-                    vst1q_f32(sum1, _sum10);
-                    vst1q_f32(sum1 + 4, _sum11);
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[out_hstep * 4] = sum0[4];
-                    outptr0[out_hstep * 5] = sum0[5];
-                    outptr0[out_hstep * 6] = sum0[6];
-                    outptr0[out_hstep * 7] = sum0[7];
-
-                    outptr0[1] = sum1[0];
-                    outptr0[out_hstep + 1] = sum1[1];
-                    outptr0[out_hstep * 2 + 1] = sum1[2];
-                    outptr0[out_hstep * 3 + 1] = sum1[3];
-                    outptr0[out_hstep * 4 + 1] = sum1[4];
-                    outptr0[out_hstep * 5 + 1] = sum1[5];
-                    outptr0[out_hstep * 6 + 1] = sum1[6];
-                    outptr0[out_hstep * 7 + 1] = sum1[7];
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-            }
-
-            outptr += 16;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v30.4s, v31.4s}, [%0]      \n"
-                "b      3f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "ld1    {v30.4s, v31.4s}, [%8]      \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v30.16b, v30.16b, v30.16b   \n"
-                "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                "2:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    4f                          \n"
-
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-                "eor    v29.16b, v29.16b, v29.16b   \n"
-                "3:                                 \n"
-                "prfm   pldl1keep, [%2, #128]       \n"
-                "ld1    {v0.4s}, [%2], #16          \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n"
-                "fmla   v28.4s, v4.4s, v0.s[0]      \n"
-                "fmla   v29.4s, v5.4s, v0.s[0]      \n"
-                "fmla   v30.4s, v6.4s, v0.s[1]      \n"
-                "fmla   v31.4s, v7.4s, v0.s[1]      \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%1], #64 \n"
-                "fmla   v28.4s, v8.4s, v0.s[2]      \n"
-                "fmla   v29.4s, v9.4s, v0.s[2]      \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v30.4s, v10.4s, v0.s[3]     \n"
-                "fmla   v31.4s, v11.4s, v0.s[3]     \n"
-                "bne    3b                          \n"
-                "fadd   v30.4s, v30.4s, v28.4s      \n"
-                "fadd   v31.4s, v31.4s, v29.4s      \n"
-
-                "4:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    6f                          \n"
-
-                "5:                                 \n"
-                "ld1r   {v0.4s}, [%2], #4           \n"
-                "ld1    {v4.4s, v5.4s}, [%1], #32   \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v30.4s, v4.4s, v0.4s        \n"
-                "fmla   v31.4s, v5.4s, v0.4s        \n"
-                "bne    5b                          \n"
-
-                "6:                                 \n"
-                "tst    %w11, #255                  \n"
-                "beq    9f                          \n"
-
-                // if out_elempack == 4
-                "cmp    %w12, #4                    \n"
-                "bne    7f                          \n"
-
-                "lsl    w4, %w13, #2                \n"
-                "add    x4, %3, w4, sxtw 2          \n"
-                "st1    {v30.4s}, [%3], #16         \n"
-                "st1    {v31.4s}, [x4]              \n"
-                "b      8f                          \n"
-
-                // if out_elempack == 1
-                "7:                                 \n"
-                "add    x4, %3, %w13, sxtw 2        \n"
-                "st1    {v30.s}[0], [%3], #4        \n"
-                "st1    {v30.s}[1], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v30.s}[2], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v30.s}[3], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v31.s}[0], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v31.s}[1], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v31.s}[2], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v31.s}[3], [x4]            \n"
-
-                "8:                                 \n"
-                "add    %0, %0, #32                 \n"
-                "b      10f                         \n"
-
-                "9:                                 \n"
-                "st1    {v30.4s, v31.4s}, [%0], #32 \n"
-
-                "10:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vld1q_f32(pC);
-                    _sum01 = vld1q_f32(pC + 4);
-                }
-                else
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA0 = vld1q_f32(pA);
-                float32x4_t _pA1 = vld1q_f32(pA + 4);
-
-                float32x4_t _pB = vld1q_dup_f32(pB);
-
-                _sum00 = vfmaq_f32(_sum00, _pA0, _pB);
-                _sum01 = vfmaq_f32(_sum01, _pA1, _pB);
-
-                pA += 8;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1q_f32(outptr0, _sum00);
-                    vst1q_f32(outptr0 + out_hstep * 4, _sum01);
-                    outptr0 += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    float sum0[8];
-                    vst1q_f32(sum0, _sum00);
-                    vst1q_f32(sum0 + 4, _sum01);
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep * 1] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[out_hstep * 4] = sum0[4];
-                    outptr0[out_hstep * 5] = sum0[5];
-                    outptr0[out_hstep * 6] = sum0[6];
-                    outptr0[out_hstep * 7] = sum0[7];
-                    outptr0++;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-            }
-
-            outptr += 8;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-
-        pAT += max_kk * 8;
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        float* outptr0 = (float*)top_blob + (i + ii) * out_hstep + j * out_elempack;
-
-        const float* pB = pBT;
-
-        if (pC)
-        {
-            pC = (const float*)CT_tile + i + ii;
-        }
-
-        int jj = 0;
-#if __aarch64__
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                "subs   %0, %0, #128                \n"
-                "b      3f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "ld1    {v20.4s}, [%8]              \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v20.16b, v20.16b, v20.16b   \n"
-
-                "2:                                 \n"
-                "mov    v21.16b, v20.16b            \n"
-                "mov    v22.16b, v20.16b            \n"
-                "mov    v23.16b, v20.16b            \n"
-                "mov    v24.16b, v20.16b            \n"
-                "mov    v25.16b, v20.16b            \n"
-                "mov    v26.16b, v20.16b            \n"
-                "mov    v27.16b, v20.16b            \n"
-                "mov    v28.16b, v20.16b            \n"
-                "mov    v29.16b, v20.16b            \n"
-                "mov    v30.16b, v20.16b            \n"
-                "mov    v31.16b, v20.16b            \n"
-
-                "3:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                "4:                                 \n"
-                "prfm   pldl1keep, [%2, #512]       \n"
-                "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%1], #64 \n"
-                "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v21.4s, v16.4s, v0.s[1]     \n"
-                "fmla   v22.4s, v16.4s, v0.s[2]     \n"
-                "fmla   v23.4s, v16.4s, v0.s[3]     \n"
-                "fmla   v24.4s, v16.4s, v1.s[0]     \n"
-                "fmla   v25.4s, v16.4s, v1.s[1]     \n"
-                "fmla   v26.4s, v16.4s, v1.s[2]     \n"
-                "fmla   v27.4s, v16.4s, v1.s[3]     \n"
-                "fmla   v28.4s, v16.4s, v2.s[0]     \n"
-                "fmla   v29.4s, v16.4s, v2.s[1]     \n"
-                "fmla   v30.4s, v16.4s, v2.s[2]     \n"
-                "fmla   v31.4s, v16.4s, v2.s[3]     \n"
-                "prfm   pldl1keep, [%2, #512]       \n"
-                "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%2], #64 \n"
-                "fmla   v20.4s, v17.4s, v3.s[0]     \n"
-                "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                "fmla   v22.4s, v17.4s, v3.s[2]     \n"
-                "fmla   v23.4s, v17.4s, v3.s[3]     \n"
-                "fmla   v24.4s, v17.4s, v4.s[0]     \n"
-                "fmla   v25.4s, v17.4s, v4.s[1]     \n"
-                "fmla   v26.4s, v17.4s, v4.s[2]     \n"
-                "fmla   v27.4s, v17.4s, v4.s[3]     \n"
-                "fmla   v28.4s, v17.4s, v5.s[0]     \n"
-                "fmla   v29.4s, v17.4s, v5.s[1]     \n"
-                "fmla   v30.4s, v17.4s, v5.s[2]     \n"
-                "fmla   v31.4s, v17.4s, v5.s[3]     \n"
-                "prfm   pldl1keep, [%2, #512]       \n"
-                "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-                "fmla   v20.4s, v18.4s, v6.s[0]     \n"
-                "fmla   v21.4s, v18.4s, v6.s[1]     \n"
-                "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                "fmla   v23.4s, v18.4s, v6.s[3]     \n"
-                "fmla   v24.4s, v18.4s, v7.s[0]     \n"
-                "fmla   v25.4s, v18.4s, v7.s[1]     \n"
-                "fmla   v26.4s, v18.4s, v7.s[2]     \n"
-                "fmla   v27.4s, v18.4s, v7.s[3]     \n"
-                "fmla   v28.4s, v18.4s, v0.s[0]     \n"
-                "fmla   v29.4s, v18.4s, v0.s[1]     \n"
-                "fmla   v30.4s, v18.4s, v0.s[2]     \n"
-                "fmla   v31.4s, v18.4s, v0.s[3]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v20.4s, v19.4s, v1.s[0]     \n"
-                "fmla   v21.4s, v19.4s, v1.s[1]     \n"
-                "fmla   v22.4s, v19.4s, v1.s[2]     \n"
-                "fmla   v23.4s, v19.4s, v1.s[3]     \n"
-                "fmla   v24.4s, v19.4s, v2.s[0]     \n"
-                "fmla   v25.4s, v19.4s, v2.s[1]     \n"
-                "fmla   v26.4s, v19.4s, v2.s[2]     \n"
-                "fmla   v27.4s, v19.4s, v2.s[3]     \n"
-                "fmla   v28.4s, v19.4s, v3.s[0]     \n"
-                "fmla   v29.4s, v19.4s, v3.s[1]     \n"
-                "fmla   v30.4s, v19.4s, v3.s[2]     \n"
-                "fmla   v31.4s, v19.4s, v3.s[3]     \n"
-                "bne    4b                          \n"
-
-                "5:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    7f                          \n"
-
-                "6:                                 \n"
-                "ld1    {v0.4s, v1.4s, v2.4s}, [%2], #48 \n"
-                "ld1    {v16.4s}, [%1], #16         \n"
-                "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v21.4s, v16.4s, v0.s[1]     \n"
-                "fmla   v22.4s, v16.4s, v0.s[2]     \n"
-                "fmla   v23.4s, v16.4s, v0.s[3]     \n"
-                "fmla   v24.4s, v16.4s, v1.s[0]     \n"
-                "fmla   v25.4s, v16.4s, v1.s[1]     \n"
-                "fmla   v26.4s, v16.4s, v1.s[2]     \n"
-                "fmla   v27.4s, v16.4s, v1.s[3]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v28.4s, v16.4s, v2.s[0]     \n"
-                "fmla   v29.4s, v16.4s, v2.s[1]     \n"
-                "fmla   v30.4s, v16.4s, v2.s[2]     \n"
-                "fmla   v31.4s, v16.4s, v2.s[3]     \n"
-                "bne    6b                          \n"
-
-                "7:                                 \n"
-                "tst    %w11, #255                  \n"
-                "beq    10f                         \n"
-
-                // if out_elempack == 4
-                "cmp    %w12, #4                    \n"
-                "bne    8f                          \n"
-
-                "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%3], #64 \n"
-                "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%3], #64 \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%3], #64 \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // transpose4x12
-                "zip1   v18.4s, v20.4s, v21.4s      \n"
-                "zip2   v19.4s, v20.4s, v21.4s      \n"
-                "zip1   v20.4s, v22.4s, v23.4s      \n"
-                "zip2   v21.4s, v22.4s, v23.4s      \n"
-                "zip1   v22.4s, v24.4s, v25.4s      \n"
-                "zip2   v23.4s, v24.4s, v25.4s      \n"
-                "zip1   v24.4s, v26.4s, v27.4s      \n"
-                "zip2   v25.4s, v26.4s, v27.4s      \n"
-                "zip1   v26.4s, v28.4s, v29.4s      \n"
-                "zip2   v27.4s, v28.4s, v29.4s      \n"
-                "zip1   v28.4s, v30.4s, v31.4s      \n"
-                "zip2   v29.4s, v30.4s, v31.4s      \n"
-
-                "zip1   v12.2d, v18.2d, v20.2d      \n"
-                "zip2   v15.2d, v18.2d, v20.2d      \n"
-                "zip1   v13.2d, v22.2d, v24.2d      \n"
-                "zip2   v16.2d, v22.2d, v24.2d      \n"
-                "zip1   v14.2d, v26.2d, v28.2d      \n"
-                "zip2   v17.2d, v26.2d, v28.2d      \n"
-
-                "zip1   v18.2d, v19.2d, v21.2d      \n"
-                "zip2   v21.2d, v19.2d, v21.2d      \n"
-                "zip1   v19.2d, v23.2d, v25.2d      \n"
-                "zip2   v22.2d, v23.2d, v25.2d      \n"
-                "zip1   v20.2d, v27.2d, v29.2d      \n"
-                "zip2   v23.2d, v27.2d, v29.2d      \n"
-
-                "add    x4, %3, %w13, sxtw 2        \n"
-                "st1    {v12.4s, v13.4s, v14.4s}, [%3], #48 \n"
-                "st1    {v15.4s, v16.4s, v17.4s}, [x4] \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v18.4s, v19.4s, v20.4s}, [x4] \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v21.4s, v22.4s, v23.4s}, [x4] \n"
-
-                "9:                                 \n"
-                "add    %0, %0, #192                \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-            float32x4_t _sum3;
-            float32x4_t _sum4;
-            float32x4_t _sum5;
-            float32x4_t _sum6;
-            float32x4_t _sum7;
-            float32x4_t _sum8;
-            float32x4_t _sum9;
-            float32x4_t _suma;
-            float32x4_t _sumb;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1q_f32(pC);
-                    _sum1 = _sum0;
-                    _sum2 = _sum0;
-                    _sum3 = _sum0;
-                    _sum4 = _sum0;
-                    _sum5 = _sum0;
-                    _sum6 = _sum0;
-                    _sum7 = _sum0;
-                    _sum8 = _sum0;
-                    _sum9 = _sum0;
-                    _suma = _sum0;
-                    _sumb = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                    _sum2 = vdupq_n_f32(0.f);
-                    _sum3 = vdupq_n_f32(0.f);
-                    _sum4 = vdupq_n_f32(0.f);
-                    _sum5 = vdupq_n_f32(0.f);
-                    _sum6 = vdupq_n_f32(0.f);
-                    _sum7 = vdupq_n_f32(0.f);
-                    _sum8 = vdupq_n_f32(0.f);
-                    _sum9 = vdupq_n_f32(0.f);
-                    _suma = vdupq_n_f32(0.f);
-                    _sumb = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4 * 1);
-                _sum2 = vld1q_f32(outptr + 4 * 2);
-                _sum3 = vld1q_f32(outptr + 4 * 3);
-                _sum4 = vld1q_f32(outptr + 4 * 4);
-                _sum5 = vld1q_f32(outptr + 4 * 5);
-                _sum6 = vld1q_f32(outptr + 4 * 6);
-                _sum7 = vld1q_f32(outptr + 4 * 7);
-                _sum8 = vld1q_f32(outptr + 4 * 8);
-                _sum9 = vld1q_f32(outptr + 4 * 9);
-                _suma = vld1q_f32(outptr + 4 * 10);
-                _sumb = vld1q_f32(outptr + 4 * 11);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA = vld1q_f32(pA);
-                float32x4_t _pB0 = vld1q_f32(pB);
-                float32x4_t _pB1 = vld1q_f32(pB + 4);
-                float32x4_t _pB2 = vld1q_f32(pB + 8);
-
-                _sum0 = vfmaq_laneq_f32(_sum0, _pA, _pB0, 0);
-                _sum1 = vfmaq_laneq_f32(_sum1, _pA, _pB0, 1);
-                _sum2 = vfmaq_laneq_f32(_sum2, _pA, _pB0, 2);
-                _sum3 = vfmaq_laneq_f32(_sum3, _pA, _pB0, 3);
-                _sum4 = vfmaq_laneq_f32(_sum4, _pA, _pB1, 0);
-                _sum5 = vfmaq_laneq_f32(_sum5, _pA, _pB1, 1);
-                _sum6 = vfmaq_laneq_f32(_sum6, _pA, _pB1, 2);
-                _sum7 = vfmaq_laneq_f32(_sum7, _pA, _pB1, 3);
-                _sum8 = vfmaq_laneq_f32(_sum8, _pA, _pB2, 0);
-                _sum9 = vfmaq_laneq_f32(_sum9, _pA, _pB2, 1);
-                _suma = vfmaq_laneq_f32(_suma, _pA, _pB2, 2);
-                _sumb = vfmaq_laneq_f32(_sumb, _pA, _pB2, 3);
-
-                pA += 4;
-                pB += 12;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1q_f32(outptr0, _sum0);
-                    vst1q_f32(outptr0 + 4, _sum1);
-                    vst1q_f32(outptr0 + 4 * 2, _sum2);
-                    vst1q_f32(outptr0 + 4 * 3, _sum3);
-                    vst1q_f32(outptr0 + 4 * 4, _sum4);
-                    vst1q_f32(outptr0 + 4 * 5, _sum5);
-                    vst1q_f32(outptr0 + 4 * 6, _sum6);
-                    vst1q_f32(outptr0 + 4 * 7, _sum7);
-                    vst1q_f32(outptr0 + 4 * 8, _sum8);
-                    vst1q_f32(outptr0 + 4 * 9, _sum9);
-                    vst1q_f32(outptr0 + 4 * 10, _suma);
-                    vst1q_f32(outptr0 + 4 * 11, _sumb);
-                    outptr0 += 48;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose4x12_ps(_sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7, _sum8, _sum9, _suma, _sumb);
-
-                    vst1q_f32(outptr0, _sum0);
-                    vst1q_f32(outptr0 + 4, _sum1);
-                    vst1q_f32(outptr0 + 8, _sum2);
-                    vst1q_f32(outptr0 + out_hstep, _sum3);
-                    vst1q_f32(outptr0 + out_hstep + 4, _sum4);
-                    vst1q_f32(outptr0 + out_hstep + 8, _sum5);
-                    vst1q_f32(outptr0 + out_hstep * 2, _sum6);
-                    vst1q_f32(outptr0 + out_hstep * 2 + 4, _sum7);
-                    vst1q_f32(outptr0 + out_hstep * 2 + 8, _sum8);
-                    vst1q_f32(outptr0 + out_hstep * 3, _sum9);
-                    vst1q_f32(outptr0 + out_hstep * 3 + 4, _suma);
-                    vst1q_f32(outptr0 + out_hstep * 3 + 8, _sumb);
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 4 * 2, _sum2);
-                vst1q_f32(outptr + 4 * 3, _sum3);
-                vst1q_f32(outptr + 4 * 4, _sum4);
-                vst1q_f32(outptr + 4 * 5, _sum5);
-                vst1q_f32(outptr + 4 * 6, _sum6);
-                vst1q_f32(outptr + 4 * 7, _sum7);
-                vst1q_f32(outptr + 4 * 8, _sum8);
-                vst1q_f32(outptr + 4 * 9, _sum9);
-                vst1q_f32(outptr + 4 * 10, _suma);
-                vst1q_f32(outptr + 4 * 11, _sumb);
-            }
-
-            outptr += 48;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-#endif // __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                "subs   %0, %0, #64                 \n"
-                "b      3f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "ld1    {v24.4s}, [%8]              \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v24.16b, v24.16b, v24.16b   \n"
-
-                "2:                                 \n"
-                "mov    v25.16b, v24.16b            \n"
-                "mov    v26.16b, v24.16b            \n"
-                "mov    v27.16b, v24.16b            \n"
-                "mov    v28.16b, v24.16b            \n"
-                "mov    v29.16b, v24.16b            \n"
-                "mov    v30.16b, v24.16b            \n"
-                "mov    v31.16b, v24.16b            \n"
-
-                "3:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                "4:                                 \n"
-                "prfm   pldl1keep, [%2, #512]       \n"
-                "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%1], #64 \n"
-                "fmla   v24.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v25.4s, v16.4s, v0.s[1]     \n"
-                "fmla   v26.4s, v16.4s, v0.s[2]     \n"
-                "fmla   v27.4s, v16.4s, v0.s[3]     \n"
-                "fmla   v28.4s, v16.4s, v1.s[0]     \n"
-                "fmla   v29.4s, v16.4s, v1.s[1]     \n"
-                "fmla   v30.4s, v16.4s, v1.s[2]     \n"
-                "fmla   v31.4s, v16.4s, v1.s[3]     \n"
-                "fmla   v24.4s, v17.4s, v2.s[0]     \n"
-                "fmla   v25.4s, v17.4s, v2.s[1]     \n"
-                "fmla   v26.4s, v17.4s, v2.s[2]     \n"
-                "fmla   v27.4s, v17.4s, v2.s[3]     \n"
-                "fmla   v28.4s, v17.4s, v3.s[0]     \n"
-                "fmla   v29.4s, v17.4s, v3.s[1]     \n"
-                "fmla   v30.4s, v17.4s, v3.s[2]     \n"
-                "fmla   v31.4s, v17.4s, v3.s[3]     \n"
-                "prfm   pldl1keep, [%2, #512]       \n"
-                "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%2], #64 \n"
-                "fmla   v24.4s, v18.4s, v4.s[0]     \n"
-                "fmla   v25.4s, v18.4s, v4.s[1]     \n"
-                "fmla   v26.4s, v18.4s, v4.s[2]     \n"
-                "fmla   v27.4s, v18.4s, v4.s[3]     \n"
-                "fmla   v28.4s, v18.4s, v5.s[0]     \n"
-                "fmla   v29.4s, v18.4s, v5.s[1]     \n"
-                "fmla   v30.4s, v18.4s, v5.s[2]     \n"
-                "fmla   v31.4s, v18.4s, v5.s[3]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v24.4s, v19.4s, v6.s[0]     \n"
-                "fmla   v25.4s, v19.4s, v6.s[1]     \n"
-                "fmla   v26.4s, v19.4s, v6.s[2]     \n"
-                "fmla   v27.4s, v19.4s, v6.s[3]     \n"
-                "fmla   v28.4s, v19.4s, v7.s[0]     \n"
-                "fmla   v29.4s, v19.4s, v7.s[1]     \n"
-                "fmla   v30.4s, v19.4s, v7.s[2]     \n"
-                "fmla   v31.4s, v19.4s, v7.s[3]     \n"
-                "bne    4b                          \n"
-
-                "5:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    7f                          \n"
-
-                "6:                                 \n"
-                "ld1    {v0.4s, v1.4s}, [%2], #32   \n"
-                "ld1    {v16.4s}, [%1], #16         \n"
-                "fmla   v24.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v25.4s, v16.4s, v0.s[1]     \n"
-                "fmla   v26.4s, v16.4s, v0.s[2]     \n"
-                "fmla   v27.4s, v16.4s, v0.s[3]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v28.4s, v16.4s, v1.s[0]     \n"
-                "fmla   v29.4s, v16.4s, v1.s[1]     \n"
-                "fmla   v30.4s, v16.4s, v1.s[2]     \n"
-                "fmla   v31.4s, v16.4s, v1.s[3]     \n"
-                "bne    6b                          \n"
-
-                "7:                                 \n"
-                "tst    %w11, #255                  \n"
-                "beq    10f                         \n"
-
-                // if out_elempack == 4
-                "cmp    %w12, #4                    \n"
-                "bne    8f                          \n"
-
-                "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%3], #64 \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%3], #64 \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // transpose4x8
-                "zip1   v22.4s, v24.4s, v25.4s      \n"
-                "zip2   v23.4s, v24.4s, v25.4s      \n"
-                "zip1   v24.4s, v26.4s, v27.4s      \n"
-                "zip2   v25.4s, v26.4s, v27.4s      \n"
-                "zip1   v26.4s, v28.4s, v29.4s      \n"
-                "zip2   v27.4s, v28.4s, v29.4s      \n"
-                "zip1   v28.4s, v30.4s, v31.4s      \n"
-                "zip2   v29.4s, v30.4s, v31.4s      \n"
-
-                "zip1   v12.2d, v22.2d, v24.2d      \n"
-                "zip2   v14.2d, v22.2d, v24.2d      \n"
-                "zip1   v13.2d, v26.2d, v28.2d      \n"
-                "zip2   v15.2d, v26.2d, v28.2d      \n"
-
-                "zip1   v16.2d, v23.2d, v25.2d      \n"
-                "zip2   v18.2d, v23.2d, v25.2d      \n"
-                "zip1   v17.2d, v27.2d, v29.2d      \n"
-                "zip2   v19.2d, v27.2d, v29.2d      \n"
-
-                "add    x4, %3, %w13, sxtw 2        \n"
-                "st1    {v12.4s, v13.4s}, [%3], #32 \n"
-                "st1    {v14.4s, v15.4s}, [x4]      \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v16.4s, v17.4s}, [x4]      \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v18.4s, v19.4s}, [x4]      \n"
-
-                "9:                                 \n"
-                "add    %0, %0, #128                \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-            asm volatile(
-                "cmp        %10, #0             \n"
-                "beq        0f                  \n"
-
-                "vldm       %0!, {d16-d23}      \n"
-                "vldm       %0, {d24-d31}       \n"
-                "sub        %0, %0, #64         \n"
-                "b          3f                  \n"
-
-                "0:                             \n"
-                // if pC
-                "cmp        %8, #0              \n"
-                "beq        1f                  \n"
-
-                "vld1.f32   {d16-d17}, [%8]     \n"
-                "b          2f                  \n"
-
-                // else
-                "1:                             \n"
-                "veor       q8, q8              \n"
-
-                "2:                             \n"
-                "vmov       q9, q8              \n"
-                "vmov       q10, q8             \n"
-                "vmov       q11, q8             \n"
-                "vmov       q12, q8             \n"
-                "vmov       q13, q8             \n"
-                "vmov       q14, q8             \n"
-                "vmov       q15, q8             \n"
-
-                "3:                             \n"
-                "lsr        r4, %9, #2          \n" // r4 = max_kk >> 2
-                "cmp        r4, #0              \n"
-                "beq        5f                  \n"
-
-                "4:                             \n"
-                "pld        [%2, #512]          \n"
-                "vldm       %2!, {d0-d7}        \n"
-                "pld        [%1, #512]          \n"
-                "vldm       %1!, {d8-d15}       \n"
-                "vmla.f32   q8, q4, d0[0]       \n"
-                "vmla.f32   q9, q4, d0[1]       \n"
-                "vmla.f32   q10, q4, d1[0]      \n"
-                "vmla.f32   q11, q4, d1[1]      \n"
-                "vmla.f32   q12, q4, d2[0]      \n"
-                "vmla.f32   q13, q4, d2[1]      \n"
-                "vmla.f32   q14, q4, d3[0]      \n"
-                "vmla.f32   q15, q4, d3[1]      \n"
-                "vmla.f32   q8, q5, d4[0]       \n"
-                "vmla.f32   q9, q5, d4[1]       \n"
-                "vmla.f32   q10, q5, d5[0]      \n"
-                "vmla.f32   q11, q5, d5[1]      \n"
-                "vmla.f32   q12, q5, d6[0]      \n"
-                "vmla.f32   q13, q5, d6[1]      \n"
-                "vmla.f32   q14, q5, d7[0]      \n"
-                "vmla.f32   q15, q5, d7[1]      \n"
-                "pld        [%2, #512]          \n"
-                "vldm       %2!, {d0-d7}        \n"
-                "vmla.f32   q8, q6, d0[0]       \n"
-                "vmla.f32   q9, q6, d0[1]       \n"
-                "vmla.f32   q10, q6, d1[0]      \n"
-                "vmla.f32   q11, q6, d1[1]      \n"
-                "vmla.f32   q12, q6, d2[0]      \n"
-                "vmla.f32   q13, q6, d2[1]      \n"
-                "vmla.f32   q14, q6, d3[0]      \n"
-                "vmla.f32   q15, q6, d3[1]      \n"
-                "subs       r4, r4, #1          \n"
-                "vmla.f32   q8, q7, d4[0]       \n"
-                "vmla.f32   q9, q7, d4[1]       \n"
-                "vmla.f32   q10, q7, d5[0]      \n"
-                "vmla.f32   q11, q7, d5[1]      \n"
-                "vmla.f32   q12, q7, d6[0]      \n"
-                "vmla.f32   q13, q7, d6[1]      \n"
-                "vmla.f32   q14, q7, d7[0]      \n"
-                "vmla.f32   q15, q7, d7[1]      \n"
-                "bne        4b                  \n"
-
-                "5:                             \n"
-                "and        r4, %9, #3          \n" // r4 = remain = max_kk & 3
-                "cmp        r4, #0              \n"
-                "beq        7f                  \n"
-
-                "6:                             \n"
-                "vld1.f32   {d0-d3}, [%2 :128]! \n"
-                "vld1.f32   {d8-d9}, [%1 :128]! \n"
-                "vmla.f32   q8, q4, d0[0]       \n"
-                "vmla.f32   q9, q4, d0[1]       \n"
-                "vmla.f32   q10, q4, d1[0]      \n"
-                "vmla.f32   q11, q4, d1[1]      \n"
-                "subs       r4, r4, #1          \n"
-                "vmla.f32   q12, q4, d2[0]      \n"
-                "vmla.f32   q13, q4, d2[1]      \n"
-                "vmla.f32   q14, q4, d3[0]      \n"
-                "vmla.f32   q15, q4, d3[1]      \n"
-                "bne        6b                  \n"
-
-                "7:                             \n"
-                "cmp        %11, #0             \n"
-                "beq        10f                 \n"
-
-                // if out_elempack == 4
-                "cmp        %12, #4             \n"
-                "bne        8f                  \n"
-
-                "vstm       %3!, {d16-d23}      \n"
-                "vstm       %3!, {d24-d31}      \n"
-                "b          9f                  \n"
-
-                // if out_elempack == 1
-                "8:                             \n"
-                // transpose4x8
-                "vtrn.32    q8, q9              \n"
-                "vtrn.32    q10, q11            \n"
-                "vtrn.32    q12, q13            \n"
-                "vtrn.32    q14, q15            \n"
-                "vswp       d17, d20            \n"
-                "vswp       d19, d22            \n"
-                "vswp       d25, d28            \n"
-                "vswp       d27, d30            \n"
-                "vswp       q9, q12             \n"
-                "vswp       q11, q14            \n"
-
-                "add        r4, %3, %13, lsl #2 \n"
-                "vst1.f32   {d16-d19}, [%3 :128]! \n"
-                "vst1.f32   {d24-d27}, [r4 :128] \n"
-                "add        r4, r4, %13, lsl #2 \n"
-                "vst1.f32   {d20-d23}, [r4 :128] \n"
-                "add        r4, r4, %13, lsl #2 \n"
-                "vst1.f32   {d28-d31}, [r4 :128] \n"
-
-                "9:                             \n"
-                "add        %0, %0, #128        \n"
-                "b          11f                 \n"
-
-                "10:                            \n"
-                "vstm       %0!, {d16-d23}      \n"
-                "vstm       %0!, {d24-d31}      \n"
-
-                "11:                            \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-            float32x4_t _sum3;
-            float32x4_t _sum4;
-            float32x4_t _sum5;
-            float32x4_t _sum6;
-            float32x4_t _sum7;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1q_f32(pC);
-                    _sum1 = _sum0;
-                    _sum2 = _sum0;
-                    _sum3 = _sum0;
-                    _sum4 = _sum0;
-                    _sum5 = _sum0;
-                    _sum6 = _sum0;
-                    _sum7 = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                    _sum2 = vdupq_n_f32(0.f);
-                    _sum3 = vdupq_n_f32(0.f);
-                    _sum4 = vdupq_n_f32(0.f);
-                    _sum5 = vdupq_n_f32(0.f);
-                    _sum6 = vdupq_n_f32(0.f);
-                    _sum7 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4 * 1);
-                _sum2 = vld1q_f32(outptr + 4 * 2);
-                _sum3 = vld1q_f32(outptr + 4 * 3);
-                _sum4 = vld1q_f32(outptr + 4 * 4);
-                _sum5 = vld1q_f32(outptr + 4 * 5);
-                _sum6 = vld1q_f32(outptr + 4 * 6);
-                _sum7 = vld1q_f32(outptr + 4 * 7);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA = vld1q_f32(pA);
-                float32x4_t _pB0 = vld1q_f32(pB);
-                float32x4_t _pB1 = vld1q_f32(pB + 4);
-
-#if __aarch64__
-                _sum0 = vfmaq_laneq_f32(_sum0, _pA, _pB0, 0);
-                _sum1 = vfmaq_laneq_f32(_sum1, _pA, _pB0, 1);
-                _sum2 = vfmaq_laneq_f32(_sum2, _pA, _pB0, 2);
-                _sum3 = vfmaq_laneq_f32(_sum3, _pA, _pB0, 3);
-                _sum4 = vfmaq_laneq_f32(_sum4, _pA, _pB1, 0);
-                _sum5 = vfmaq_laneq_f32(_sum5, _pA, _pB1, 1);
-                _sum6 = vfmaq_laneq_f32(_sum6, _pA, _pB1, 2);
-                _sum7 = vfmaq_laneq_f32(_sum7, _pA, _pB1, 3);
-#else
-                _sum0 = vmlaq_lane_f32(_sum0, _pA, vget_low_f32(_pB0), 0);
-                _sum1 = vmlaq_lane_f32(_sum1, _pA, vget_low_f32(_pB0), 1);
-                _sum2 = vmlaq_lane_f32(_sum2, _pA, vget_high_f32(_pB0), 0);
-                _sum3 = vmlaq_lane_f32(_sum3, _pA, vget_high_f32(_pB0), 1);
-                _sum4 = vmlaq_lane_f32(_sum4, _pA, vget_low_f32(_pB1), 0);
-                _sum5 = vmlaq_lane_f32(_sum5, _pA, vget_low_f32(_pB1), 1);
-                _sum6 = vmlaq_lane_f32(_sum6, _pA, vget_high_f32(_pB1), 0);
-                _sum7 = vmlaq_lane_f32(_sum7, _pA, vget_high_f32(_pB1), 1);
-#endif
-
-                pA += 4;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1q_f32(outptr0, _sum0);
-                    vst1q_f32(outptr0 + 4, _sum1);
-                    vst1q_f32(outptr0 + 4 * 2, _sum2);
-                    vst1q_f32(outptr0 + 4 * 3, _sum3);
-                    vst1q_f32(outptr0 + 4 * 4, _sum4);
-                    vst1q_f32(outptr0 + 4 * 5, _sum5);
-                    vst1q_f32(outptr0 + 4 * 6, _sum6);
-                    vst1q_f32(outptr0 + 4 * 7, _sum7);
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose4x8_ps(_sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7);
-
-                    vst1q_f32(outptr0, _sum0);
-                    vst1q_f32(outptr0 + 4, _sum1);
-                    vst1q_f32(outptr0 + out_hstep, _sum2);
-                    vst1q_f32(outptr0 + out_hstep + 4, _sum3);
-                    vst1q_f32(outptr0 + out_hstep * 2, _sum4);
-                    vst1q_f32(outptr0 + out_hstep * 2 + 4, _sum5);
-                    vst1q_f32(outptr0 + out_hstep * 3, _sum6);
-                    vst1q_f32(outptr0 + out_hstep * 3 + 4, _sum7);
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 4 * 2, _sum2);
-                vst1q_f32(outptr + 4 * 3, _sum3);
-                vst1q_f32(outptr + 4 * 4, _sum4);
-                vst1q_f32(outptr + 4 * 5, _sum5);
-                vst1q_f32(outptr + 4 * 6, _sum6);
-                vst1q_f32(outptr + 4 * 7, _sum7);
-            }
-
-            outptr += 32;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                "b      3f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "ld1    {v28.4s}, [%8]              \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-
-                "2:                                 \n"
-                "mov    v29.16b, v28.16b            \n"
-                "mov    v30.16b, v28.16b            \n"
-                "mov    v31.16b, v28.16b            \n"
-
-                "3:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                "4:                                 \n"
-                "prfm   pldl1keep, [%2, #512]       \n"
-                "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%1], #64 \n"
-                "fmla   v28.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v29.4s, v16.4s, v0.s[1]     \n"
-                "fmla   v30.4s, v16.4s, v0.s[2]     \n"
-                "fmla   v31.4s, v16.4s, v0.s[3]     \n"
-                "fmla   v28.4s, v17.4s, v1.s[0]     \n"
-                "fmla   v29.4s, v17.4s, v1.s[1]     \n"
-                "fmla   v30.4s, v17.4s, v1.s[2]     \n"
-                "fmla   v31.4s, v17.4s, v1.s[3]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v28.4s, v18.4s, v2.s[0]     \n"
-                "fmla   v29.4s, v18.4s, v2.s[1]     \n"
-                "fmla   v30.4s, v18.4s, v2.s[2]     \n"
-                "fmla   v31.4s, v18.4s, v2.s[3]     \n"
-                "fmla   v28.4s, v19.4s, v3.s[0]     \n"
-                "fmla   v29.4s, v19.4s, v3.s[1]     \n"
-                "fmla   v30.4s, v19.4s, v3.s[2]     \n"
-                "fmla   v31.4s, v19.4s, v3.s[3]     \n"
-                "bne    4b                          \n"
-
-                "5:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    7f                          \n"
-
-                "6:                                 \n"
-                "ld1    {v0.4s}, [%2], #16          \n"
-                "ld1    {v16.4s}, [%1], #16         \n"
-                "fmla   v28.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v29.4s, v16.4s, v0.s[1]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v30.4s, v16.4s, v0.s[2]     \n"
-                "fmla   v31.4s, v16.4s, v0.s[3]     \n"
-                "bne    6b                          \n"
-
-                "7:                                 \n"
-                "tst    %w11, #255                  \n"
-                "beq    10f                         \n"
-
-                // if out_elempack == 4
-                "cmp    %w12, #4                    \n"
-                "bne    8f                          \n"
-
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%3], #64 \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // transpose4x4
-                "zip1   v26.4s, v28.4s, v29.4s      \n"
-                "zip2   v27.4s, v28.4s, v29.4s      \n"
-                "zip1   v28.4s, v30.4s, v31.4s      \n"
-                "zip2   v29.4s, v30.4s, v31.4s      \n"
-
-                "zip1   v12.2d, v26.2d, v28.2d      \n"
-                "zip2   v13.2d, v26.2d, v28.2d      \n"
-                "zip1   v14.2d, v27.2d, v29.2d      \n"
-                "zip2   v15.2d, v27.2d, v29.2d      \n"
-
-                "add    x4, %3, %w13, sxtw 2        \n"
-                "st1    {v12.4s}, [%3], #16         \n"
-                "st1    {v13.4s}, [x4]              \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v14.4s}, [x4]              \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v15.4s}, [x4]              \n"
-
-                "9:                                 \n"
-                "add    %0, %0, #64                 \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-            asm volatile(
-                "cmp        %10, #0             \n"
-                "beq        0f                  \n"
-
-                "vldm       %0, {d24-d31}       \n"
-                "b          3f                  \n"
-
-                "0:                             \n"
-                // if pC
-                "cmp        %8, #0              \n"
-                "beq        1f                  \n"
-
-                "vld1.f32   {d24-d25}, [%8]     \n"
-                "b          2f                  \n"
-
-                // else
-                "1:                             \n"
-                "veor       q12, q12            \n"
-
-                "2:                             \n"
-                "vmov       q13, q12            \n"
-                "vmov       q14, q12            \n"
-                "vmov       q15, q12            \n"
-
-                "3:                             \n"
-                "lsr        r4, %9, #2          \n" // r4 = max_kk >> 2
-                "cmp        r4, #0              \n"
-                "beq        5f                  \n"
-
-                "4:                             \n"
-                "pld        [%2, #512]          \n"
-                "vldm       %2!, {d0-d7}        \n"
-                "pld        [%1, #512]          \n"
-                "vldm       %1!, {d8-d15}       \n"
-                "vmla.f32   q12, q4, d0[0]      \n"
-                "vmla.f32   q13, q4, d0[1]      \n"
-                "vmla.f32   q14, q4, d1[0]      \n"
-                "vmla.f32   q15, q4, d1[1]      \n"
-                "vmla.f32   q12, q5, d2[0]      \n"
-                "vmla.f32   q13, q5, d2[1]      \n"
-                "vmla.f32   q14, q5, d3[0]      \n"
-                "vmla.f32   q15, q5, d3[1]      \n"
-                "subs       r4, r4, #1          \n"
-                "vmla.f32   q12, q6, d4[0]      \n"
-                "vmla.f32   q13, q6, d4[1]      \n"
-                "vmla.f32   q14, q6, d5[0]      \n"
-                "vmla.f32   q15, q6, d5[1]      \n"
-                "vmla.f32   q12, q7, d6[0]      \n"
-                "vmla.f32   q13, q7, d6[1]      \n"
-                "vmla.f32   q14, q7, d7[0]      \n"
-                "vmla.f32   q15, q7, d7[1]      \n"
-                "bne        4b                  \n"
-
-                "5:                             \n"
-                "and        r4, %9, #3          \n" // r4 = remain = max_kk & 3
-                "cmp        r4, #0              \n"
-                "beq        7f                  \n"
-
-                "6:                             \n"
-                "vld1.f32   {d0-d1}, [%2 :128]! \n"
-                "vld1.f32   {d8-d9}, [%1 :128]! \n"
-                "vmla.f32   q12, q4, d0[0]      \n"
-                "vmla.f32   q13, q4, d0[1]      \n"
-                "subs       r4, r4, #1          \n"
-                "vmla.f32   q14, q4, d1[0]      \n"
-                "vmla.f32   q15, q4, d1[1]      \n"
-                "bne        6b                  \n"
-
-                "7:                             \n"
-                "cmp        %11, #0             \n"
-                "beq        10f                 \n"
-
-                // if out_elempack == 4
-                "cmp        %12, #4             \n"
-                "bne        8f                  \n"
-
-                "vstm       %3!, {d24-d31}      \n"
-                "b          9f                  \n"
-
-                // if out_elempack == 1
-                "8:                             \n"
-                // transpose4x4
-                "vtrn.32    q12, q13            \n"
-                "vtrn.32    q14, q15            \n"
-                "vswp       d25, d28            \n"
-                "vswp       d27, d30            \n"
-
-                "add        r4, %3, %13, lsl #2 \n"
-                "vst1.f32   {d24-d25}, [%3 :128]! \n"
-                "vst1.f32   {d26-d27}, [r4 :128] \n"
-                "add        r4, r4, %13, lsl #2 \n"
-                "vst1.f32   {d28-d29}, [r4 :128] \n"
-                "add        r4, r4, %13, lsl #2 \n"
-                "vst1.f32   {d30-d31}, [r4 :128] \n"
-
-                "9:                             \n"
-                "add        %0, %0, #64         \n"
-                "b          11f                 \n"
-
-                "10:                            \n"
-                "vstm       %0!, {d24-d31}      \n"
-
-                "11:                            \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-            float32x4_t _sum3;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1q_f32(pC);
-                    _sum1 = _sum0;
-                    _sum2 = _sum0;
-                    _sum3 = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                    _sum2 = vdupq_n_f32(0.f);
-                    _sum3 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4 * 1);
-                _sum2 = vld1q_f32(outptr + 4 * 2);
-                _sum3 = vld1q_f32(outptr + 4 * 3);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA = vld1q_f32(pA);
-                float32x4_t _pB = vld1q_f32(pB);
-
-#if __aarch64__
-                _sum0 = vfmaq_laneq_f32(_sum0, _pA, _pB, 0);
-                _sum1 = vfmaq_laneq_f32(_sum1, _pA, _pB, 1);
-                _sum2 = vfmaq_laneq_f32(_sum2, _pA, _pB, 2);
-                _sum3 = vfmaq_laneq_f32(_sum3, _pA, _pB, 3);
-#else
-                _sum0 = vmlaq_lane_f32(_sum0, _pA, vget_low_f32(_pB), 0);
-                _sum1 = vmlaq_lane_f32(_sum1, _pA, vget_low_f32(_pB), 1);
-                _sum2 = vmlaq_lane_f32(_sum2, _pA, vget_high_f32(_pB), 0);
-                _sum3 = vmlaq_lane_f32(_sum3, _pA, vget_high_f32(_pB), 1);
-#endif
-
-                pA += 4;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1q_f32(outptr0, _sum0);
-                    vst1q_f32(outptr0 + 4, _sum1);
-                    vst1q_f32(outptr0 + 4 * 2, _sum2);
-                    vst1q_f32(outptr0 + 4 * 3, _sum3);
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose4x4_ps(_sum0, _sum1, _sum2, _sum3);
-
-                    vst1q_f32(outptr0, _sum0);
-                    vst1q_f32(outptr0 + out_hstep * 1, _sum1);
-                    vst1q_f32(outptr0 + out_hstep * 2, _sum2);
-                    vst1q_f32(outptr0 + out_hstep * 3, _sum3);
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 4 * 2, _sum2);
-                vst1q_f32(outptr + 4 * 3, _sum3);
-            }
-
-            outptr += 16;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v30.4s, v31.4s}, [%0]      \n"
-                "b      3f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "ld1    {v30.4s}, [%8]              \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v30.16b, v30.16b, v30.16b   \n"
-
-                "2:                                 \n"
-                "mov    v31.16b, v30.16b            \n"
-
-                "3:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-                "eor    v29.16b, v29.16b, v29.16b   \n"
-                "4:                                 \n"
-                "prfm   pldl1keep, [%2, #256]       \n"
-                "ld1    {v0.4s, v1.4s}, [%2], #32   \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%1], #64 \n"
-                "fmla   v28.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v29.4s, v16.4s, v0.s[1]     \n"
-                "fmla   v30.4s, v17.4s, v0.s[2]     \n"
-                "fmla   v31.4s, v17.4s, v0.s[3]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v28.4s, v18.4s, v1.s[0]     \n"
-                "fmla   v29.4s, v18.4s, v1.s[1]     \n"
-                "fmla   v30.4s, v19.4s, v1.s[2]     \n"
-                "fmla   v31.4s, v19.4s, v1.s[3]     \n"
-                "bne    4b                          \n"
-                "fadd   v30.4s, v30.4s, v28.4s      \n"
-                "fadd   v31.4s, v31.4s, v29.4s      \n"
-
-                "5:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    7f                          \n"
-
-                "6:                                 \n"
-                "ld1    {v0.2s}, [%2], #8           \n"
-                "ld1    {v16.4s}, [%1], #16         \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v30.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v31.4s, v16.4s, v0.s[1]     \n"
-                "bne    6b                          \n"
-
-                "7:                                 \n"
-                "tst    %w11, #255                  \n"
-                "beq    10f                         \n"
-
-                // if out_elempack == 4
-                "cmp    %w12, #4                    \n"
-                "bne    8f                          \n"
-
-                "st1    {v30.4s, v31.4s}, [%3], #32 \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // transpose4x2
-                "zip1   v28.4s, v30.4s, v31.4s      \n"
-                "zip2   v29.4s, v30.4s, v31.4s      \n"
-
-                "add    x4, %3, %w13, sxtw 2        \n"
-                "st1    {v28.d}[0], [%3], #8        \n"
-                "st1    {v28.d}[1], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v29.d}[0], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v29.d}[1], [x4]            \n"
-
-                "9:                                 \n"
-                "add    %0, %0, #32                 \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v30.4s, v31.4s}, [%0], #32 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v1", "v16", "v17", "v18", "v19", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-            asm volatile(
-                "cmp        %10, #0             \n"
-                "beq        0f                  \n"
-
-                "vld1.f32   {d28-d31}, [%0 :128] \n"
-                "b          3f                  \n"
-
-                "0:                             \n"
-                // if pC
-                "cmp        %8, #0              \n"
-                "beq        1f                  \n"
-
-                "vld1.f32   {d28-d29}, [%8]     \n"
-                "b          2f                  \n"
-
-                // else
-                "1:                             \n"
-                "veor       q14, q14            \n"
-
-                "2:                             \n"
-                "vmov       q15, q14            \n"
-
-                "3:                             \n"
-                "lsr        r4, %9, #2          \n" // r4 = max_kk >> 2
-                "cmp        r4, #0              \n"
-                "beq        5f                  \n"
-
-                "veor       q12, q12            \n"
-                "veor       q13, q13            \n"
-                "4:                             \n"
-                "pld        [%2, #256]          \n"
-                "vld1.f32   {d0-d3}, [%2 :128]! \n"
-                "pld        [%1, #512]          \n"
-                "vldm       %1!, {d8-d15}       \n"
-                "vmla.f32   q12, q4, d0[0]      \n"
-                "vmla.f32   q13, q4, d0[1]      \n"
-                "vmla.f32   q14, q5, d1[0]      \n"
-                "vmla.f32   q15, q5, d1[1]      \n"
-                "subs       r4, r4, #1          \n"
-                "vmla.f32   q12, q6, d2[0]      \n"
-                "vmla.f32   q13, q6, d2[1]      \n"
-                "vmla.f32   q14, q7, d3[0]      \n"
-                "vmla.f32   q15, q7, d3[1]      \n"
-                "bne        4b                  \n"
-                "vadd.f32   q14, q14, q12       \n"
-                "vadd.f32   q15, q15, q13       \n"
-
-                "5:                             \n"
-                "and        r4, %9, #3          \n" // r4 = remain = max_kk & 3
-                "cmp        r4, #0              \n"
-                "beq        7f                  \n"
-
-                "6:                             \n"
-                "vld1.f32   {d0}, [%2 :64]!     \n"
-                "vld1.f32   {d8-d9}, [%1 :128]! \n"
-                "subs       r4, r4, #1          \n"
-                "vmla.f32   q14, q4, d0[0]      \n"
-                "vmla.f32   q15, q4, d0[1]      \n"
-                "bne        6b                  \n"
-
-                "7:                             \n"
-                "cmp        %11, #0             \n"
-                "beq        10f                 \n"
-
-                // if out_elempack == 4
-                "cmp        %12, #4             \n"
-                "bne        8f                  \n"
-
-                "vst1.f32   {d28-d31}, [%3 :128]! \n"
-                "b          9f                  \n"
-
-                // if out_elempack == 1
-                "8:                             \n"
-                // transpose4x2
-                "vtrn.32    q14, q15            \n"
-
-                "add        r4, %3, %13, lsl #2 \n"
-                "vst1.f32   {d28}, [%3 :64]!    \n"
-                "vst1.f32   {d30}, [r4 :64]     \n"
-                "add        r4, r4, %13, lsl #2 \n"
-                "vst1.f32   {d29}, [r4 :64]     \n"
-                "add        r4, r4, %13, lsl #2 \n"
-                "vst1.f32   {d31}, [r4 :64]     \n"
-
-                "9:                             \n"
-                "add        %0, %0, #32         \n"
-                "b          11f                 \n"
-
-                "10:                            \n"
-                "vst1.f32   {d28-d31}, [%0 :128]! \n"
-
-                "11:                            \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "r4", "q0", "q1", "q4", "q5", "q6", "q7", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1q_f32(pC);
-                    _sum1 = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA = vld1q_f32(pA);
-                float32x2_t _pB = vld1_f32(pB);
-
-#if __aarch64__
-                _sum0 = vfmaq_lane_f32(_sum0, _pA, _pB, 0);
-                _sum1 = vfmaq_lane_f32(_sum1, _pA, _pB, 1);
-#else
-                _sum0 = vmlaq_lane_f32(_sum0, _pA, _pB, 0);
-                _sum1 = vmlaq_lane_f32(_sum1, _pA, _pB, 1);
-#endif
-
-                pA += 4;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1q_f32(outptr0, _sum0);
-                    vst1q_f32(outptr0 + 4, _sum1);
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-                    float sum0[4];
-                    float sum1[4];
-                    vst1q_f32(sum0, _sum0);
-                    vst1q_f32(sum1, _sum1);
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[1] = sum1[0];
-                    outptr0[out_hstep + 1] = sum1[1];
-                    outptr0[out_hstep * 2 + 1] = sum1[2];
-                    outptr0[out_hstep * 3 + 1] = sum1[3];
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-            }
-
-            outptr += 8;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            const float* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v31.4s}, [%0]              \n"
-                "b      2f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "ld1    {v31.4s}, [%8]              \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                "2:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    4f                          \n"
-
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-                "eor    v29.16b, v29.16b, v29.16b   \n"
-                "eor    v30.16b, v30.16b, v30.16b   \n"
-                "3:                                 \n"
-                "prfm   pldl1keep, [%2, #128]       \n"
-                "ld1    {v0.4s}, [%2], #16          \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%1], #64 \n"
-                "fmla   v28.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v29.4s, v17.4s, v0.s[1]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v30.4s, v18.4s, v0.s[2]     \n"
-                "fmla   v31.4s, v19.4s, v0.s[3]     \n"
-                "bne    3b                          \n"
-                "fadd   v30.4s, v30.4s, v28.4s      \n"
-                "fadd   v31.4s, v31.4s, v29.4s      \n"
-                "fadd   v31.4s, v31.4s, v30.4s      \n"
-
-                "4:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    6f                          \n"
-
-                "5:                                 \n"
-                "ld1r   {v0.4s}, [%2], #4           \n"
-                "ld1    {v16.4s}, [%1], #16         \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v31.4s, v16.4s, v0.4s       \n"
-                "bne    5b                          \n"
-
-                "6:                                 \n"
-                "tst    %w11, #255                  \n"
-                "beq    9f                          \n"
-
-                // if out_elempack == 4
-                "cmp    %w12, #4                    \n"
-                "bne    7f                          \n"
-
-                "st1    {v31.4s}, [%3], #16         \n"
-                "b      8f                          \n"
-
-                // if out_elempack == 1
-                "7:                                 \n"
-                "add    x4, %3, %w13, sxtw 2        \n"
-                "st1    {v31.s}[0], [%3], #4        \n"
-                "st1    {v31.s}[1], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v31.s}[2], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 2        \n"
-                "st1    {v31.s}[3], [x4]            \n"
-
-                "8:                                 \n"
-                "add    %0, %0, #16                 \n"
-                "b      10f                         \n"
-
-                "9:                                 \n"
-                "st1    {v31.4s}, [%0], #16         \n"
-
-                "10:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v16", "v17", "v18", "v19", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-            asm volatile(
-                "cmp        %10, #0             \n"
-                "beq        0f                  \n"
-
-                "vld1.f32   {d30-d31}, [%0 :128] \n"
-                "b          2f                  \n"
-
-                "0:                             \n"
-                // if pC
-                "cmp        %8, #0              \n"
-                "beq        1f                  \n"
-
-                "vld1.f32   {d30-d31}, [%8]     \n"
-                "b          2f                  \n"
-
-                // else
-                "1:                             \n"
-                "veor       q15, q15            \n"
-
-                "2:                             \n"
-                "lsr        r4, %9, #2          \n" // r4 = max_kk >> 2
-                "cmp        r4, #0              \n"
-                "beq        4f                  \n"
-
-                "veor       q12, q12            \n"
-                "veor       q13, q13            \n"
-                "veor       q14, q14            \n"
-                "3:                             \n"
-                "pld        [%2, #128]          \n"
-                "vld1.f32   {d0-d1}, [%2 :64]!  \n"
-                "pld        [%1, #512]          \n"
-                "vldm       %1!, {d8-d15}       \n"
-                "vmla.f32   q12, q4, d0[0]      \n"
-                "vmla.f32   q13, q5, d0[1]      \n"
-                "subs       r4, r4, #1          \n"
-                "vmla.f32   q14, q6, d1[0]      \n"
-                "vmla.f32   q15, q7, d1[1]      \n"
-                "bne        3b                  \n"
-                "vadd.f32   q14, q14, q12       \n"
-                "vadd.f32   q15, q15, q13       \n"
-                "vadd.f32   q15, q15, q14       \n"
-
-                "4:                             \n"
-                "and        r4, %9, #3          \n" // r4 = remain = max_kk & 3
-                "cmp        r4, #0              \n"
-                "beq        6f                  \n"
-
-                "5:                             \n"
-                "vld1.f32   {d0[0]}, [%2]!      \n"
-                "vld1.f32   {d8-d9}, [%1 :128]! \n"
-                "subs       r4, r4, #1          \n"
-                "vmla.f32   q15, q4, d0[0]      \n"
-                "bne        5b                  \n"
-
-                "6:                             \n"
-                "cmp        %11, #0             \n"
-                "beq        9f                  \n"
-
-                // if out_elempack == 4
-                "cmp        %12, #4             \n"
-                "bne        7f                  \n"
-
-                "vst1.f32   {d30-d31}, [%3 :128]! \n"
-                "b          8f                  \n"
-
-                // if out_elempack == 1
-                "7:                             \n"
-                "add        r4, %3, %13, lsl #2 \n"
-                "vst1.f32   {d30[0]}, [%3]!     \n"
-                "vst1.f32   {d30[1]}, [r4]      \n"
-                "add        r4, r4, %13, lsl #2 \n"
-                "vst1.f32   {d31[0]}, [r4]      \n"
-                "add        r4, r4, %13, lsl #2 \n"
-                "vst1.f32   {d31[1]}, [r4]      \n"
-
-                "8:                             \n"
-                "add        %0, %0, #16         \n"
-                "b          10f                 \n"
-
-                "9:                             \n"
-                "vst1.f32   {d30-d31}, [%0 :128]! \n"
-
-                "10:                            \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "r4", "q0", "q4", "q5", "q6", "q7", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum0;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1q_f32(pC);
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA = vld1q_f32(pA);
-                float32x4_t _pB = vdupq_n_f32(pB[0]);
-
-#if __aarch64__
-                _sum0 = vfmaq_f32(_sum0, _pA, _pB);
-#else
-                _sum0 = vmlaq_f32(_sum0, _pA, _pB);
-#endif
-
-                pA += 4;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1q_f32(outptr0, _sum0);
-                    outptr0 += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    float sum0[4];
-                    vst1q_f32(sum0, _sum0);
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0++;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-            }
-
-            outptr += 4;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-
-        pAT += max_kk * 4;
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        float* outptr0 = (float*)top_blob + (i + ii) * out_hstep + j;
-
-        const float* pB = pBT;
-
-        if (pC)
-        {
-            pC = (const float*)CT_tile + i + ii;
-        }
-
-        int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum02;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum12;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vdupq_n_f32(pC[0]);
-                    _sum01 = vdupq_n_f32(pC[0]);
-                    _sum02 = vdupq_n_f32(pC[0]);
-                    _sum10 = vdupq_n_f32(pC[1]);
-                    _sum11 = vdupq_n_f32(pC[1]);
-                    _sum12 = vdupq_n_f32(pC[1]);
-                }
-                else
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum02 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                    _sum12 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01 = vld2q_f32(outptr);
-                float32x4x2_t _tmp23 = vld2q_f32(outptr + 8);
-                float32x4x2_t _tmp45 = vld2q_f32(outptr + 16);
-                _sum00 = _tmp01.val[0];
-                _sum01 = _tmp23.val[0];
-                _sum02 = _tmp45.val[0];
-                _sum10 = _tmp01.val[1];
-                _sum11 = _tmp23.val[1];
-                _sum12 = _tmp45.val[1];
-            }
-
-            const float* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pB0 = vld1q_f32(pB);
-                float32x4_t _pB1 = vld1q_f32(pB + 4);
-                float32x4_t _pB2 = vld1q_f32(pB + 8);
-
-                float32x2_t _pA = vld1_f32(pA);
-
-                _sum00 = vfmaq_lane_f32(_sum00, _pB0, _pA, 0);
-                _sum01 = vfmaq_lane_f32(_sum01, _pB1, _pA, 0);
-                _sum02 = vfmaq_lane_f32(_sum02, _pB2, _pA, 0);
-                _sum10 = vfmaq_lane_f32(_sum10, _pB0, _pA, 1);
-                _sum11 = vfmaq_lane_f32(_sum11, _pB1, _pA, 1);
-                _sum12 = vfmaq_lane_f32(_sum12, _pB2, _pA, 1);
-
-                pA += 2;
-                pB += 12;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1q_f32(outptr0, _sum00);
-                    vst1q_f32(outptr0 + 4, _sum01);
-                    vst1q_f32(outptr0 + 8, _sum02);
-                    vst1q_f32(outptr0 + out_hstep, _sum10);
-                    vst1q_f32(outptr0 + out_hstep + 4, _sum11);
-                    vst1q_f32(outptr0 + out_hstep + 8, _sum12);
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01;
-                _tmp01.val[0] = _sum00;
-                _tmp01.val[1] = _sum10;
-                float32x4x2_t _tmp23;
-                _tmp23.val[0] = _sum01;
-                _tmp23.val[1] = _sum11;
-                float32x4x2_t _tmp45;
-                _tmp45.val[0] = _sum02;
-                _tmp45.val[1] = _sum12;
-                vst2q_f32(outptr, _tmp01);
-                vst2q_f32(outptr + 8, _tmp23);
-                vst2q_f32(outptr + 16, _tmp45);
-            }
-
-            outptr += 24;
-        }
-#endif // __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vdupq_n_f32(pC[0]);
-                    _sum01 = vdupq_n_f32(pC[0]);
-                    _sum10 = vdupq_n_f32(pC[1]);
-                    _sum11 = vdupq_n_f32(pC[1]);
-                }
-                else
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01 = vld2q_f32(outptr);
-                float32x4x2_t _tmp23 = vld2q_f32(outptr + 8);
-                _sum00 = _tmp01.val[0];
-                _sum01 = _tmp23.val[0];
-                _sum10 = _tmp01.val[1];
-                _sum11 = _tmp23.val[1];
-            }
-
-            const float* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pB0 = vld1q_f32(pB);
-                float32x4_t _pB1 = vld1q_f32(pB + 4);
-
-                float32x2_t _pA = vld1_f32(pA);
-#if __aarch64__
-                _sum00 = vfmaq_lane_f32(_sum00, _pB0, _pA, 0);
-                _sum01 = vfmaq_lane_f32(_sum01, _pB1, _pA, 0);
-                _sum10 = vfmaq_lane_f32(_sum10, _pB0, _pA, 1);
-                _sum11 = vfmaq_lane_f32(_sum11, _pB1, _pA, 1);
-#else
-                _sum00 = vmlaq_lane_f32(_sum00, _pB0, _pA, 0);
-                _sum01 = vmlaq_lane_f32(_sum01, _pB1, _pA, 0);
-                _sum10 = vmlaq_lane_f32(_sum10, _pB0, _pA, 1);
-                _sum11 = vmlaq_lane_f32(_sum11, _pB1, _pA, 1);
-#endif
-
-                pA += 2;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1q_f32(outptr0, _sum00);
-                    vst1q_f32(outptr0 + 4, _sum01);
-                    vst1q_f32(outptr0 + out_hstep, _sum10);
-                    vst1q_f32(outptr0 + out_hstep + 4, _sum11);
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01;
-                _tmp01.val[0] = _sum00;
-                _tmp01.val[1] = _sum10;
-                float32x4x2_t _tmp23;
-                _tmp23.val[0] = _sum01;
-                _tmp23.val[1] = _sum11;
-                vst2q_f32(outptr, _tmp01);
-                vst2q_f32(outptr + 8, _tmp23);
-            }
-
-            outptr += 16;
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vdupq_n_f32(pC[0]);
-                    _sum1 = vdupq_n_f32(pC[1]);
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01 = vld2q_f32(outptr);
-                _sum0 = _tmp01.val[0];
-                _sum1 = _tmp01.val[1];
-            }
-
-            const float* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pB = vld1q_f32(pB);
-
-                float32x2_t _pA = vld1_f32(pA);
-#if __aarch64__
-                _sum0 = vfmaq_lane_f32(_sum0, _pB, _pA, 0);
-                _sum1 = vfmaq_lane_f32(_sum1, _pB, _pA, 1);
-#else
-                _sum0 = vmlaq_lane_f32(_sum0, _pB, _pA, 0);
-                _sum1 = vmlaq_lane_f32(_sum1, _pB, _pA, 1);
-#endif
-
-                pA += 2;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1q_f32(outptr0, _sum0);
-                    vst1q_f32(outptr0 + out_hstep, _sum1);
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01;
-                _tmp01.val[0] = _sum0;
-                _tmp01.val[1] = _sum1;
-                vst2q_f32(outptr, _tmp01);
-            }
-
-            outptr += 8;
-        }
-#endif // __ARM_NEON
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            float sum00;
-            float sum01;
-            float sum10;
-            float sum11;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    sum00 = pC[0];
-                    sum01 = pC[1];
-                    sum10 = pC[0];
-                    sum11 = pC[1];
-                }
-                else
-                {
-                    sum00 = 0.f;
-                    sum01 = 0.f;
-                    sum10 = 0.f;
-                    sum11 = 0.f;
-                }
-            }
-            else
-            {
-                sum00 = outptr[0];
-                sum01 = outptr[1];
-                sum10 = outptr[2];
-                sum11 = outptr[3];
-            }
-
-            const float* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                sum00 += pA[0] * pB[0];
-                sum01 += pA[1] * pB[0];
-                sum10 += pA[0] * pB[1];
-                sum11 += pA[1] * pB[1];
-
-                pA += 2;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum00;
-                    outptr0[1] = sum10;
-                    outptr0[out_hstep] = sum01;
-                    outptr0[out_hstep + 1] = sum11;
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                outptr[0] = sum00;
-                outptr[1] = sum01;
-                outptr[2] = sum10;
-                outptr[3] = sum11;
-            }
-
-            outptr += 4;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            float sum0;
-            float sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    sum0 = pC[0];
-                    sum1 = pC[1];
-                }
-                else
-                {
-                    sum0 = 0.f;
-                    sum1 = 0.f;
-                }
-            }
-            else
-            {
-                sum0 = outptr[0];
-                sum1 = outptr[1];
-            }
-
-            const float* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                sum0 += pA[0] * pB[0];
-                sum1 += pA[1] * pB[0];
-                pA += 2;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum0;
-                    outptr0[out_hstep] = sum1;
-                    outptr0++;
-                }
-            }
-            else
-            {
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-            }
-
-            outptr += 2;
-        }
-
-        pAT += max_kk * 2;
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        float* outptr0 = (float*)top_blob + (i + ii) * out_hstep + j;
-
-        const float* pB = pBT;
-
-        if (pC)
-        {
-            pC = (const float*)CT_tile + i + ii;
-        }
-
-        int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vdupq_n_f32(pC[0]);
-                    _sum1 = vdupq_n_f32(pC[0]);
-                    _sum2 = vdupq_n_f32(pC[0]);
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                    _sum2 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4);
-                _sum2 = vld1q_f32(outptr + 8);
-            }
-
-            const float* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pB0 = vld1q_f32(pB);
-                float32x4_t _pB1 = vld1q_f32(pB + 4);
-                float32x4_t _pB2 = vld1q_f32(pB + 8);
-
-                float32x4_t _pA0 = vdupq_n_f32(pA[0]);
-
-                _sum0 = vfmaq_f32(_sum0, _pA0, _pB0);
-                _sum1 = vfmaq_f32(_sum1, _pA0, _pB1);
-                _sum2 = vfmaq_f32(_sum2, _pA0, _pB2);
-
-                pA += 1;
-                pB += 12;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1q_f32(outptr0, _sum0);
-                    vst1q_f32(outptr0 + 4, _sum1);
-                    vst1q_f32(outptr0 + 8, _sum2);
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 8, _sum2);
-            }
-
-            outptr += 12;
-        }
-#endif // __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vdupq_n_f32(pC[0]);
-                    _sum1 = vdupq_n_f32(pC[0]);
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4);
-            }
-
-            const float* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pB0 = vld1q_f32(pB);
-                float32x4_t _pB1 = vld1q_f32(pB + 4);
-
-                float32x4_t _pA0 = vdupq_n_f32(pA[0]);
-#if __aarch64__
-                _sum0 = vfmaq_f32(_sum0, _pA0, _pB0);
-                _sum1 = vfmaq_f32(_sum1, _pA0, _pB1);
-#else
-                _sum0 = vmlaq_f32(_sum0, _pA0, _pB0);
-                _sum1 = vmlaq_f32(_sum1, _pA0, _pB1);
-#endif
-
-                pA += 1;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1q_f32(outptr0, _sum0);
-                    vst1q_f32(outptr0 + 4, _sum1);
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-            }
-
-            outptr += 8;
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            float32x4_t _sum;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum = vdupq_n_f32(pC[0]);
-                }
-                else
-                {
-                    _sum = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum = vld1q_f32(outptr);
-            }
-
-            const float* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pB = vld1q_f32(pB);
-                float32x4_t _pA = vdupq_n_f32(pA[0]);
-
-#if __aarch64__
-                _sum = vfmaq_f32(_sum, _pA, _pB);
-#else
-                _sum = vmlaq_f32(_sum, _pA, _pB);
-#endif
-
-                pA += 1;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1q_f32(outptr0, _sum);
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum);
-            }
-
-            outptr += 4;
-        }
-#endif // __ARM_NEON
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            float sum0;
-            float sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    sum0 = pC[0];
-                    sum1 = pC[0];
-                }
-                else
-                {
-                    sum0 = 0.f;
-                    sum1 = 0.f;
-                }
-            }
-            else
-            {
-                sum0 = outptr[0];
-                sum1 = outptr[1];
-            }
-
-            const float* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                sum0 += pA[0] * pB[0];
-                sum1 += pA[0] * pB[1];
-
-                pA += 1;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum0;
-                    outptr0[1] = sum1;
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-            }
-
-            outptr += 2;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            float sum;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    sum = pC[0];
-                }
-                else
-                {
-                    sum = 0.f;
-                }
-            }
-            else
-            {
-                sum = outptr[0];
-            }
-
-            const float* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                sum += pA[0] * pB[0];
-                pA += 1;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum;
-                    outptr0++;
-                }
-            }
-            else
-            {
-                outptr[0] = sum;
-            }
-
-            outptr += 1;
-        }
-
-        pAT += max_kk;
-    }
-}
-
-static void convolution_im2col_gemm_get_optimal_tile_mnk(int M, int N, int K, int& TILE_M, int& TILE_N, int& TILE_K, int nT)
-{
-    // resolve optimal tile size from cache size
-    const int l2_cache_size_fp32 = (int)(get_cpu_level2_cache_size() / sizeof(float));
-
-    if (nT == 0)
-        nT = get_physical_big_cpu_count();
-
-    // solve K
-    {
-        // try not to split K
-#if __aarch64__
-        int tile_size = (l2_cache_size_fp32 - 32) / 12;
-#elif __ARM_NEON
-        int tile_size = (l2_cache_size_fp32 - 16) / 8;
-#else
-        int tile_size = (l2_cache_size_fp32 - 2) / 3;
-#endif
-
-#if __aarch64__
-        TILE_K = std::max(8, tile_size / 8 * 8);
-#elif __ARM_NEON
-        TILE_K = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_K = std::max(2, tile_size / 2 * 2);
-#endif
-
-        int nn_K = (K + TILE_K - 1) / TILE_K;
-#if __aarch64__
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 7) / 8 * 8);
-#elif __ARM_NEON
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 3) / 4 * 4);
-#else
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 1) / 2 * 2);
-#endif
-    }
-
-    // solve M
-    {
-#if __aarch64__
-        int nn_M = (M + 31) / 32;
-#elif __ARM_NEON
-        int nn_M = (M + 15) / 16;
-#else
-        int nn_M = (M + 7) / 8;
-#endif
-
-#if __aarch64__
-        TILE_M = std::max(8, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-#elif __ARM_NEON
-        TILE_M = std::max(4, ((M + nn_M - 1) / nn_M + 3) / 4 * 4);
-#else
-        TILE_M = std::max(2, ((M + nn_M - 1) / nn_M + 1) / 2 * 2);
-#endif
-    }
-
-    {
-        TILE_M *= std::min(nT, get_physical_cpu_count());
-
-        int nn_M = (M + TILE_M - 1) / TILE_M;
-#if __aarch64__
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-#elif __ARM_NEON
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 3) / 4 * 4);
-#else
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 1) / 2 * 2);
-#endif
-
-        if (nT > 1)
-        {
-#if __aarch64__
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 7) / 8 * 8);
-#elif __ARM_NEON
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 3) / 4 * 4);
-#else
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 1) / 2 * 2);
-#endif
-        }
-    }
-
-    if (N > 0)
-    {
-        int tile_size;
-        if (TILE_K >= K)
-        {
-            tile_size = (l2_cache_size_fp32 - TILE_M * TILE_K) / TILE_K;
-        }
-        else
-        {
-            tile_size = (l2_cache_size_fp32 - TILE_M * TILE_K) / (TILE_M + TILE_K);
-        }
-
-#if __aarch64__
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#elif __ARM_NEON
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_N = std::max(1, tile_size);
-#endif
-
-        int nn_N = (N + TILE_N - 1) / TILE_N;
-#if __aarch64__
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#elif __ARM_NEON
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#else
-        TILE_N = std::min(TILE_N, (N + nn_N - 1) / nn_N);
-#endif
-
-#if __aarch64__
-        TILE_N = std::max(4, TILE_N);
-#elif __ARM_NEON
-        TILE_N = std::max(4, TILE_N);
-#else
-        TILE_N = std::max(1, TILE_N);
-#endif
-    }
-}
-
-static void convolution_im2col_input_tile_conv1x1s1d1(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk)
-{
-    const int elempack = bottom_blob.elempack;
-
-    float* pp = B;
-
-    int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; jj + 11 < max_jj; jj += 12)
-    {
-        if (elempack == 4)
-        {
-            const float* p0 = (const float*)bottom_blob.channel(k / 4) + (j + jj) * 4;
-
-            int kk = 0;
-            for (; kk < max_kk / 4; kk++)
-            {
-                // transpose4x12
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v4.4s, v5.4s, v6.4s, v7.4s}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0] \n"
-                    "st1    {v0.4s}, [%1], #16          \n"
-                    "st1    {v4.4s}, [%1], #16          \n"
-                    "st1    {v8.4s}, [%1], #16          \n"
-                    "sub    %0, %0, #128                \n"
-                    "st1    {v1.4s}, [%1], #16          \n"
-                    "st1    {v5.4s}, [%1], #16          \n"
-                    "st1    {v9.4s}, [%1], #16          \n"
-                    "st1    {v2.4s}, [%1], #16          \n"
-                    "st1    {v6.4s}, [%1], #16          \n"
-                    "st1    {v10.4s}, [%1], #16         \n"
-                    "st1    {v3.4s}, [%1], #16          \n"
-                    "st1    {v7.4s}, [%1], #16          \n"
-                    "st1    {v11.4s}, [%1], #16         \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11");
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4x4_t _r0 = vld4q_f32(p0);
-                float32x4x4_t _r1 = vld4q_f32(p0 + 16);
-                float32x4x4_t _r2 = vld4q_f32(p0 + 32);
-                vst1q_f32(pp, _r0.val[0]);
-                vst1q_f32(pp + 4, _r1.val[0]);
-                vst1q_f32(pp + 4 * 2, _r2.val[0]);
-                vst1q_f32(pp + 4 * 3, _r0.val[1]);
-                vst1q_f32(pp + 4 * 4, _r1.val[1]);
-                vst1q_f32(pp + 4 * 5, _r2.val[1]);
-                vst1q_f32(pp + 4 * 6, _r0.val[2]);
-                vst1q_f32(pp + 4 * 7, _r1.val[2]);
-                vst1q_f32(pp + 4 * 8, _r2.val[2]);
-                vst1q_f32(pp + 4 * 9, _r0.val[3]);
-                vst1q_f32(pp + 4 * 10, _r1.val[3]);
-                vst1q_f32(pp + 4 * 11, _r2.val[3]);
-                pp += 48;
-#endif // NCNN_GNU_INLINE_ASM
-                p0 += bottom_blob.cstep * 4;
-            }
-        }
-
-        if (elempack == 1)
-        {
-            const float* p0 = (const float*)bottom_blob.channel(k) + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                float32x4_t _r0 = vld1q_f32(p0);
-                float32x4_t _r1 = vld1q_f32(p0 + 4);
-                float32x4_t _r2 = vld1q_f32(p0 + 8);
-                vst1q_f32(pp, _r0);
-                vst1q_f32(pp + 4, _r1);
-                vst1q_f32(pp + 8, _r2);
-                pp += 12;
-                p0 += bottom_blob.cstep;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-        if (elempack == 4)
-        {
-            const float* p0 = (const float*)bottom_blob.channel(k / 4) + (j + jj) * 4;
-
-            int kk = 0;
-            for (; kk < max_kk / 4; kk++)
-            {
-                // transpose4x8
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v4.4s, v5.4s, v6.4s, v7.4s}, [%0] \n"
-                    "st1    {v0.4s}, [%1], #16          \n"
-                    "st1    {v4.4s}, [%1], #16          \n"
-                    "st1    {v1.4s}, [%1], #16          \n"
-                    "st1    {v5.4s}, [%1], #16          \n"
-                    "sub    %0, %0, #64                 \n"
-                    "st1    {v2.4s}, [%1], #16          \n"
-                    "st1    {v6.4s}, [%1], #16          \n"
-                    "st1    {v3.4s}, [%1], #16          \n"
-                    "st1    {v7.4s}, [%1], #16          \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%0, #512]          \n"
-                    "vldm       %0!, {d0-d7}        \n"
-                    "pld        [%0, #512]          \n"
-                    "vldm       %0, {d16-d23}       \n"
-                    "vzip.32    q0, q1              \n"
-                    "vzip.32    q2, q3              \n"
-                    "vzip.32    q8, q9              \n"
-                    "vzip.32    q10, q11            \n"
-                    "vswp       d1, d4              \n"
-                    "vswp       d3, d6              \n"
-                    "vswp       d17, d20            \n"
-                    "vswp       d19, d22            \n"
-                    "vswp       q1, q8              \n"
-                    "vswp       q3, q10             \n"
-                    "sub        %0, %0, #64         \n"
-                    "vstm       %1!, {d0-d7}        \n"
-                    "vstm       %1!, {d16-d23}      \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4x4_t _r0 = vld4q_f32(p0);
-                float32x4x4_t _r1 = vld4q_f32(p0 + 16);
-                vst1q_f32(pp, _r0.val[0]);
-                vst1q_f32(pp + 4, _r1.val[0]);
-                vst1q_f32(pp + 4 * 2, _r0.val[1]);
-                vst1q_f32(pp + 4 * 3, _r1.val[1]);
-                vst1q_f32(pp + 4 * 4, _r0.val[2]);
-                vst1q_f32(pp + 4 * 5, _r1.val[2]);
-                vst1q_f32(pp + 4 * 6, _r0.val[3]);
-                vst1q_f32(pp + 4 * 7, _r1.val[3]);
-                pp += 32;
-#endif // NCNN_GNU_INLINE_ASM
-                p0 += bottom_blob.cstep * 4;
-            }
-        }
-
-        if (elempack == 1)
-        {
-            const float* p0 = (const float*)bottom_blob.channel(k) + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                float32x4_t _r0 = vld1q_f32(p0);
-                float32x4_t _r1 = vld1q_f32(p0 + 4);
-                vst1q_f32(pp, _r0);
-                vst1q_f32(pp + 4, _r1);
-                pp += 8;
-                p0 += bottom_blob.cstep;
-            }
-        }
-    }
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-        if (elempack == 4)
-        {
-            const float* p0 = (const float*)bottom_blob.channel(k / 4) + (j + jj) * 4;
-
-            int kk = 0;
-            for (; kk < max_kk / 4; kk++)
-            {
-                // transpose4x4
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0] \n"
-                    "st4    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%0, #512]          \n"
-                    "vldm       %0, {d0-d7}         \n"
-                    "vtrn.32    q0, q1              \n"
-                    "vtrn.32    q2, q3              \n"
-                    "vswp       d1, d4              \n"
-                    "vswp       d3, d6              \n"
-                    "vstm       %1!, {d0-d7}        \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "q0", "q1", "q2", "q3");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                float32x4x4_t _r0;
-                _r0.val[0] = vld1q_f32(p0);
-                _r0.val[1] = vld1q_f32(p0 + 4);
-                _r0.val[2] = vld1q_f32(p0 + 4 * 2);
-                _r0.val[3] = vld1q_f32(p0 + 4 * 3);
-                vst4q_f32(pp, _r0);
-                pp += 16;
-#endif // NCNN_GNU_INLINE_ASM
-                p0 += bottom_blob.cstep * 4;
-            }
-        }
-
-        if (elempack == 1)
-        {
-            const float* p0 = (const float*)bottom_blob.channel(k) + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                vst1q_f32(pp, vld1q_f32(p0));
-                pp += 4;
-                p0 += bottom_blob.cstep;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-#if __ARM_NEON
-        if (elempack == 4)
-        {
-            const float* p0 = (const float*)bottom_blob.channel(k / elempack) + (j + jj) * elempack;
-
-            int kk = 0;
-            for (; kk < max_kk / elempack; kk++)
-            {
-                // transpose4x2
-                float32x4x2_t _r0;
-                _r0.val[0] = vld1q_f32(p0);
-                _r0.val[1] = vld1q_f32(p0 + 4);
-                vst2q_f32(pp, _r0);
-                pp += 8;
-                p0 += bottom_blob.cstep * elempack;
-            }
-        }
-#endif // __ARM_NEON
-
-        if (elempack == 1)
-        {
-            const float* p0 = (const float*)bottom_blob.channel(k) + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-#if __ARM_NEON
-                vst1_f32(pp, vld1_f32(p0));
-#else
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-#endif // __ARM_NEON
-                pp += 2;
-                p0 += bottom_blob.cstep;
-            }
-        }
-    }
-    for (; jj < max_jj; jj++)
-    {
-#if __ARM_NEON
-        if (elempack == 4)
-        {
-            const float* p0 = (const float*)bottom_blob.channel(k / 4) + (j + jj) * 4;
-
-            int kk = 0;
-            for (; kk < max_kk / 4; kk++)
-            {
-                vst1q_f32(pp, vld1q_f32(p0));
-                pp += 4;
-                p0 += bottom_blob.cstep * 4;
-            }
-        }
-#endif // __ARM_NEON
-
-        if (elempack == 1)
-        {
-            const float* p0 = (const float*)bottom_blob.channel(k) + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp += 1;
-                p0 += bottom_blob.cstep;
-            }
-        }
-    }
-}
-
-static void convolution_im2col_input_tile(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h)
-{
-    if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
-    {
-        convolution_im2col_input_tile_conv1x1s1d1(bottom_blob, B, j, max_jj, k, max_kk);
-        return;
-    }
-
-    const int w = bottom_blob.w;
-    // const int channels = bottom_blob.c;
-    const int elempack = bottom_blob.elempack;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int outw = (w - kernel_extent_w) / stride_w + 1;
-
-    // j max_jj     outw*outh    split w and h
-
-    // k max_kk     pa*maxk*(inch/pa)    split inch
-
-    // k/max_kk shall be multiple of maxk
-
-    const int maxk = kernel_w * kernel_h;
-
-    float* pp = B;
-
-    int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; jj + 11 < max_jj; jj += 12)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dy2 = (j + jj + 2) / outw;
-        int dy3 = (j + jj + 3) / outw;
-        int dy4 = (j + jj + 4) / outw;
-        int dy5 = (j + jj + 5) / outw;
-        int dy6 = (j + jj + 6) / outw;
-        int dy7 = (j + jj + 7) / outw;
-        int dy8 = (j + jj + 8) / outw;
-        int dy9 = (j + jj + 9) / outw;
-        int dya = (j + jj + 10) / outw;
-        int dyb = (j + jj + 11) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-        int dx2 = (j + jj + 2) % outw;
-        int dx3 = (j + jj + 3) % outw;
-        int dx4 = (j + jj + 4) % outw;
-        int dx5 = (j + jj + 5) % outw;
-        int dx6 = (j + jj + 6) % outw;
-        int dx7 = (j + jj + 7) % outw;
-        int dx8 = (j + jj + 8) % outw;
-        int dx9 = (j + jj + 9) % outw;
-        int dxa = (j + jj + 10) % outw;
-        int dxb = (j + jj + 11) % outw;
-
-        int kk = 0;
-        for (; kk < max_kk / elempack; kk++)
-        {
-            int p = (k / elempack + kk) / maxk;
-            int uv = (k / elempack + kk) % maxk;
-            int u = uv / kernel_w;
-            int v = uv % kernel_w;
-
-            const Mat img = bottom_blob.channel(p);
-
-            int x0 = stride_w * dx0 + dilation_w * v;
-            int x1 = stride_w * dx1 + dilation_w * v;
-            int x2 = stride_w * dx2 + dilation_w * v;
-            int x3 = stride_w * dx3 + dilation_w * v;
-            int x4 = stride_w * dx4 + dilation_w * v;
-            int x5 = stride_w * dx5 + dilation_w * v;
-            int x6 = stride_w * dx6 + dilation_w * v;
-            int x7 = stride_w * dx7 + dilation_w * v;
-            int x8 = stride_w * dx8 + dilation_w * v;
-            int x9 = stride_w * dx9 + dilation_w * v;
-            int xa = stride_w * dxa + dilation_w * v;
-            int xb = stride_w * dxb + dilation_w * v;
-
-            int y0 = stride_h * dy0 + dilation_h * u;
-            int y1 = stride_h * dy1 + dilation_h * u;
-            int y2 = stride_h * dy2 + dilation_h * u;
-            int y3 = stride_h * dy3 + dilation_h * u;
-            int y4 = stride_h * dy4 + dilation_h * u;
-            int y5 = stride_h * dy5 + dilation_h * u;
-            int y6 = stride_h * dy6 + dilation_h * u;
-            int y7 = stride_h * dy7 + dilation_h * u;
-            int y8 = stride_h * dy8 + dilation_h * u;
-            int y9 = stride_h * dy9 + dilation_h * u;
-            int ya = stride_h * dya + dilation_h * u;
-            int yb = stride_h * dyb + dilation_h * u;
-
-            const float* sptr0 = img.row(y0) + x0 * elempack;
-            const float* sptr1 = img.row(y1) + x1 * elempack;
-            const float* sptr2 = img.row(y2) + x2 * elempack;
-            const float* sptr3 = img.row(y3) + x3 * elempack;
-            const float* sptr4 = img.row(y4) + x4 * elempack;
-            const float* sptr5 = img.row(y5) + x5 * elempack;
-            const float* sptr6 = img.row(y6) + x6 * elempack;
-            const float* sptr7 = img.row(y7) + x7 * elempack;
-            const float* sptr8 = img.row(y8) + x8 * elempack;
-            const float* sptr9 = img.row(y9) + x9 * elempack;
-            const float* sptra = img.row(ya) + xa * elempack;
-            const float* sptrb = img.row(yb) + xb * elempack;
-
-            if (elempack == 4)
-            {
-                float32x4_t _r0 = vld1q_f32(sptr0);
-                float32x4_t _r1 = vld1q_f32(sptr1);
-                float32x4_t _r2 = vld1q_f32(sptr2);
-                float32x4_t _r3 = vld1q_f32(sptr3);
-                float32x4_t _r4 = vld1q_f32(sptr4);
-                float32x4_t _r5 = vld1q_f32(sptr5);
-                float32x4_t _r6 = vld1q_f32(sptr6);
-                float32x4_t _r7 = vld1q_f32(sptr7);
-                float32x4_t _r8 = vld1q_f32(sptr8);
-                float32x4_t _r9 = vld1q_f32(sptr9);
-                float32x4_t _ra = vld1q_f32(sptra);
-                float32x4_t _rb = vld1q_f32(sptrb);
-                transpose4x12_ps(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7, _r8, _r9, _ra, _rb);
-                vst1q_f32(pp, _r0);
-                vst1q_f32(pp + 4, _r1);
-                vst1q_f32(pp + 4 * 2, _r2);
-                vst1q_f32(pp + 4 * 3, _r3);
-                vst1q_f32(pp + 4 * 4, _r4);
-                vst1q_f32(pp + 4 * 5, _r5);
-                vst1q_f32(pp + 4 * 6, _r6);
-                vst1q_f32(pp + 4 * 7, _r7);
-                vst1q_f32(pp + 4 * 8, _r8);
-                vst1q_f32(pp + 4 * 9, _r9);
-                vst1q_f32(pp + 4 * 10, _ra);
-                vst1q_f32(pp + 4 * 11, _rb);
-                pp += 48;
-            }
-            if (elempack == 1)
-            {
-                pp[0] = sptr0[0];
-                pp[1] = sptr1[0];
-                pp[2] = sptr2[0];
-                pp[3] = sptr3[0];
-                pp[4] = sptr4[0];
-                pp[5] = sptr5[0];
-                pp[6] = sptr6[0];
-                pp[7] = sptr7[0];
-                pp[8] = sptr8[0];
-                pp[9] = sptr9[0];
-                pp[10] = sptra[0];
-                pp[11] = sptrb[0];
-                pp += 12;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dy2 = (j + jj + 2) / outw;
-        int dy3 = (j + jj + 3) / outw;
-        int dy4 = (j + jj + 4) / outw;
-        int dy5 = (j + jj + 5) / outw;
-        int dy6 = (j + jj + 6) / outw;
-        int dy7 = (j + jj + 7) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-        int dx2 = (j + jj + 2) % outw;
-        int dx3 = (j + jj + 3) % outw;
-        int dx4 = (j + jj + 4) % outw;
-        int dx5 = (j + jj + 5) % outw;
-        int dx6 = (j + jj + 6) % outw;
-        int dx7 = (j + jj + 7) % outw;
-
-        int kk = 0;
-        for (; kk < max_kk / elempack; kk++)
-        {
-            int p = (k / elempack + kk) / maxk;
-            int uv = (k / elempack + kk) % maxk;
-            int u = uv / kernel_w;
-            int v = uv % kernel_w;
-
-            const Mat img = bottom_blob.channel(p);
-
-            int x0 = stride_w * dx0 + dilation_w * v;
-            int x1 = stride_w * dx1 + dilation_w * v;
-            int x2 = stride_w * dx2 + dilation_w * v;
-            int x3 = stride_w * dx3 + dilation_w * v;
-            int x4 = stride_w * dx4 + dilation_w * v;
-            int x5 = stride_w * dx5 + dilation_w * v;
-            int x6 = stride_w * dx6 + dilation_w * v;
-            int x7 = stride_w * dx7 + dilation_w * v;
-            int y0 = stride_h * dy0 + dilation_h * u;
-            int y1 = stride_h * dy1 + dilation_h * u;
-            int y2 = stride_h * dy2 + dilation_h * u;
-            int y3 = stride_h * dy3 + dilation_h * u;
-            int y4 = stride_h * dy4 + dilation_h * u;
-            int y5 = stride_h * dy5 + dilation_h * u;
-            int y6 = stride_h * dy6 + dilation_h * u;
-            int y7 = stride_h * dy7 + dilation_h * u;
-
-            const float* sptr0 = img.row(y0) + x0 * elempack;
-            const float* sptr1 = img.row(y1) + x1 * elempack;
-            const float* sptr2 = img.row(y2) + x2 * elempack;
-            const float* sptr3 = img.row(y3) + x3 * elempack;
-            const float* sptr4 = img.row(y4) + x4 * elempack;
-            const float* sptr5 = img.row(y5) + x5 * elempack;
-            const float* sptr6 = img.row(y6) + x6 * elempack;
-            const float* sptr7 = img.row(y7) + x7 * elempack;
-
-            if (elempack == 4)
-            {
-                float32x4_t _r0 = vld1q_f32(sptr0);
-                float32x4_t _r1 = vld1q_f32(sptr1);
-                float32x4_t _r2 = vld1q_f32(sptr2);
-                float32x4_t _r3 = vld1q_f32(sptr3);
-                float32x4_t _r4 = vld1q_f32(sptr4);
-                float32x4_t _r5 = vld1q_f32(sptr5);
-                float32x4_t _r6 = vld1q_f32(sptr6);
-                float32x4_t _r7 = vld1q_f32(sptr7);
-                transpose4x8_ps(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-                vst1q_f32(pp, _r0);
-                vst1q_f32(pp + 4, _r1);
-                vst1q_f32(pp + 4 * 2, _r2);
-                vst1q_f32(pp + 4 * 3, _r3);
-                vst1q_f32(pp + 4 * 4, _r4);
-                vst1q_f32(pp + 4 * 5, _r5);
-                vst1q_f32(pp + 4 * 6, _r6);
-                vst1q_f32(pp + 4 * 7, _r7);
-                pp += 32;
-            }
-            if (elempack == 1)
-            {
-                pp[0] = sptr0[0];
-                pp[1] = sptr1[0];
-                pp[2] = sptr2[0];
-                pp[3] = sptr3[0];
-                pp[4] = sptr4[0];
-                pp[5] = sptr5[0];
-                pp[6] = sptr6[0];
-                pp[7] = sptr7[0];
-                pp += 8;
-            }
-        }
-    }
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dy2 = (j + jj + 2) / outw;
-        int dy3 = (j + jj + 3) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-        int dx2 = (j + jj + 2) % outw;
-        int dx3 = (j + jj + 3) % outw;
-
-        int kk = 0;
-        for (; kk < max_kk / elempack; kk++)
-        {
-            int p = (k / elempack + kk) / maxk;
-            int uv = (k / elempack + kk) % maxk;
-            int u = uv / kernel_w;
-            int v = uv % kernel_w;
-
-            const Mat img = bottom_blob.channel(p);
-
-            int x0 = stride_w * dx0 + dilation_w * v;
-            int x1 = stride_w * dx1 + dilation_w * v;
-            int x2 = stride_w * dx2 + dilation_w * v;
-            int x3 = stride_w * dx3 + dilation_w * v;
-            int y0 = stride_h * dy0 + dilation_h * u;
-            int y1 = stride_h * dy1 + dilation_h * u;
-            int y2 = stride_h * dy2 + dilation_h * u;
-            int y3 = stride_h * dy3 + dilation_h * u;
-
-            const float* sptr0 = img.row(y0) + x0 * elempack;
-            const float* sptr1 = img.row(y1) + x1 * elempack;
-            const float* sptr2 = img.row(y2) + x2 * elempack;
-            const float* sptr3 = img.row(y3) + x3 * elempack;
-
-            if (elempack == 4)
-            {
-                float32x4x4_t _r0;
-                _r0.val[0] = vld1q_f32(sptr0);
-                _r0.val[1] = vld1q_f32(sptr1);
-                _r0.val[2] = vld1q_f32(sptr2);
-                _r0.val[3] = vld1q_f32(sptr3);
-                vst4q_f32(pp, _r0);
-                pp += 16;
-            }
-            if (elempack == 1)
-            {
-                pp[0] = sptr0[0];
-                pp[1] = sptr1[0];
-                pp[2] = sptr2[0];
-                pp[3] = sptr3[0];
-                pp += 4;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-
-        int kk = 0;
-        for (; kk < max_kk / elempack; kk++)
-        {
-            int p = (k / elempack + kk) / maxk;
-            int uv = (k / elempack + kk) % maxk;
-            int u = uv / kernel_w;
-            int v = uv % kernel_w;
-
-            const Mat img = bottom_blob.channel(p);
-
-            int x0 = stride_w * dx0 + dilation_w * v;
-            int x1 = stride_w * dx1 + dilation_w * v;
-            int y0 = stride_h * dy0 + dilation_h * u;
-            int y1 = stride_h * dy1 + dilation_h * u;
-
-            const float* sptr0 = img.row(y0) + x0 * elempack;
-            const float* sptr1 = img.row(y1) + x1 * elempack;
-
-#if __ARM_NEON
-            if (elempack == 4)
-            {
-                pp[0] = sptr0[0];
-                pp[1] = sptr1[0];
-                pp[2] = sptr0[1];
-                pp[3] = sptr1[1];
-                pp[4] = sptr0[2];
-                pp[5] = sptr1[2];
-                pp[6] = sptr0[3];
-                pp[7] = sptr1[3];
-                pp += 8;
-            }
-#endif // __ARM_NEON
-            if (elempack == 1)
-            {
-                pp[0] = sptr0[0];
-                pp[1] = sptr1[0];
-                pp += 2;
-            }
-        }
-    }
-    for (; jj < max_jj; jj++)
-    {
-        int dy = (j + jj) / outw;
-        int dx = (j + jj) % outw;
-
-        int kk = 0;
-        for (; kk < max_kk / elempack; kk++)
-        {
-            int p = (k / elempack + kk) / maxk;
-            int uv = (k / elempack + kk) % maxk;
-            int u = uv / kernel_w;
-            int v = uv % kernel_w;
-
-            const Mat img = bottom_blob.channel(p);
-
-            int x = stride_w * dx + dilation_w * v;
-            int y = stride_h * dy + dilation_h * u;
-
-            const float* sptr = img.row(y) + x * elempack;
-
-#if __ARM_NEON
-            if (elempack == 4)
-            {
-                pp[0] = sptr[0];
-                pp[1] = sptr[1];
-                pp[2] = sptr[2];
-                pp[3] = sptr[3];
-                pp += 4;
-            }
-#endif // __ARM_NEON
-            if (elempack == 1)
-            {
-                pp[0] = sptr[0];
-                pp += 1;
-            }
-        }
-    }
-}
-
-static void convolution_im2col_gemm_transform_kernel(const Mat& kernel, Mat& AT, int inch, int outch, int kernel_w, int kernel_h, const Option& opt)
-{
-    // NCNN_LOGE("convolution_im2col_gemm_transform_kernel");
-    const int maxk = kernel_w * kernel_h;
-
-    const int M = outch;
-    const int K = inch * maxk;
-
-    int TILE_M, TILE_N, TILE_K;
-    convolution_im2col_gemm_get_optimal_tile_mnk(M, 0, K, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    int elempack = 1;
-#if __ARM_NEON
-    if (opt.use_packing_layout)
-    {
-        elempack = inch % 4 == 0 ? 4 : 1;
-    }
-#endif // __ARM_NEON
-
-    // maxk-inch-outch to pa-maxk-inch/pa-outch
-    Mat A_data;
-    if (maxk == 1)
-    {
-        A_data = kernel.reshape(maxk * inch, outch);
-    }
-    else
-    {
-        Mat weight_data_r2 = kernel.reshape(maxk, inch, outch);
-
-        A_data.create(maxk * inch, outch);
-
-        for (int q = 0; q < outch; q += 1)
-        {
-            float* g00 = A_data.row(q);
-
-            for (int p = 0; p + (elempack - 1) < inch; p += elempack)
-            {
-                for (int k = 0; k < maxk; k++)
-                {
-                    for (int i = 0; i < elempack; i++)
-                    {
-                        const float* k00 = weight_data_r2.channel(q).row(p + i);
-                        g00[0] = k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-
-    AT.create(TILE_K * TILE_M, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat AT_tile = AT.channel(i / TILE_M).row_range(k / TILE_K, 1);
-
-            convolution_im2col_pack_A_tile(A_data, AT_tile, i, max_ii, k, max_kk);
-        }
-    }
-}
-
-static void convolution_im2col_gemm(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, const Mat& bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = top_blob.w * top_blob.h;
-    const int K = bottom_blob.c * bottom_blob.elempack * maxk;
-
-    int TILE_M, TILE_N, TILE_K;
-    convolution_im2col_gemm_get_optimal_tile_mnk(M, N, K, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 4u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-    {
-        const int ppj = ppjk / nn_K;
-        const int ppk = ppjk % nn_K;
-
-        const int j = ppj * TILE_N;
-        const int k = ppk * TILE_K;
-
-        const int max_jj = std::min((N - j), TILE_N);
-        const int max_kk = std::min((K - k), TILE_K);
-
-        Mat BT_tile = BT.channel(j / TILE_N).row_range(k / TILE_K, 1);
-
-        // im2col
-        convolution_im2col_input_tile(bottom_blob, BT_tile, j, max_jj, k, max_kk, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h);
-    }
-
-    Mat topT_tileX;
-    if (K > TILE_K)
-        topT_tileX.create(TILE_N * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat topT_tile;
-        if (K > TILE_K)
-            topT_tile = topT_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).row_range(k / TILE_K, 1);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).row_range(k / TILE_K, 1);
-
-                bool k_end = k + TILE_K >= K;
-
-                convolution_gemm_transB_packed_tile(AT_tile, BT_tile, bias, topT_tile, top_blob, i, max_ii, j, max_jj, k, max_kk, k_end, opt.use_a53_a55_optimized_kernel);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_bf16s.h
deleted file mode 100644
index f29420e9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_bf16s.h
+++ /dev/null
@@ -1,6085 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_gemm_transB_packed_tile_bf16s(const Mat& AT_tile, const Mat& BT_tile, const Mat& CT_tile, Mat& topT_tile, Mat& top_blob, int i, int max_ii, int j, int max_jj, int k, int max_kk, bool k_end, int use_a53_a55_optimized_kernel)
-{
-    // NCNN_LOGE("convolution_gemm_transB_packed_tile_bf16s %d %d %d %d %d %d", i, max_ii, j, max_jj, k, max_kk);
-
-    const int out_elempack = top_blob.elempack;
-    const int out_hstep = (int)top_blob.cstep;
-
-    const unsigned short* pAT = AT_tile;
-    const unsigned short* pBT = BT_tile;
-    const float* pC = CT_tile;
-
-    float* outptr = topT_tile;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        unsigned short* outptr0 = (unsigned short*)top_blob + (i + ii) * out_hstep + j * out_elempack;
-
-        const unsigned short* pB = pBT;
-
-        if (pC)
-        {
-            pC = (const float*)CT_tile + i + ii;
-        }
-
-        int jj = 0;
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            const unsigned short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            if (use_a53_a55_optimized_kernel && cpu_support_arm_asimdhp())
-            {
-                // a55
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #320                \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "add    x4, %8, #16                 \n"
-                    "ld1    {v8.4s}, [%8]               \n"
-                    "ld1    {v20.4s}, [x4]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v8.16b, v8.16b, v8.16b      \n"
-                    "eor    v20.16b, v20.16b, v20.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v9.16b, v8.16b              \n"
-                    "mov    v10.16b, v8.16b             \n"
-                    "mov    v11.16b, v8.16b             \n"
-                    "mov    v12.16b, v8.16b             \n"
-                    "mov    v13.16b, v8.16b             \n"
-                    "mov    v14.16b, v8.16b             \n"
-                    "mov    v15.16b, v8.16b             \n"
-                    "mov    v16.16b, v8.16b             \n"
-                    "mov    v17.16b, v8.16b             \n"
-                    "mov    v18.16b, v8.16b             \n"
-                    "mov    v19.16b, v8.16b             \n"
-
-                    "mov    v21.16b, v20.16b            \n"
-                    "mov    v22.16b, v20.16b            \n"
-                    "mov    v23.16b, v20.16b            \n"
-                    "mov    v24.16b, v20.16b            \n"
-                    "mov    v25.16b, v20.16b            \n"
-                    "mov    v26.16b, v20.16b            \n"
-                    "mov    v27.16b, v20.16b            \n"
-                    "mov    v28.16b, v20.16b            \n"
-                    "mov    v29.16b, v20.16b            \n"
-                    "mov    v30.16b, v20.16b            \n"
-                    "mov    v31.16b, v20.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.4h, v5.4h}, [%1], #16   \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h}, [%2], #24 \n"
-
-                    "shll   v4.4s, v4.4h, #16           \n"
-                    "shll   v0.4s, v0.4h, #16           \n"
-
-                    ".align 4                           \n"
-                    "4:                                 \n"
-                    "shll   v5.4s, v5.4h, #16           \n"
-                    "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                    "ldr    d6, [%1], #8                \n"
-                    "fmla   v9.4s, v4.4s, v0.s[1]       \n"
-                    "ldr    d3, [%2], #8                \n"
-                    "fmla   v10.4s, v4.4s, v0.s[2]      \n"
-                    "ldr    d7, [%1], #8                \n"
-                    "fmla   v11.4s, v4.4s, v0.s[3]      \n"
-                    "shll   v1.4s, v1.4h, #16           \n"
-                    "fmla   v20.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v21.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v22.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v23.4s, v5.4s, v0.s[3]      \n"
-                    "fmla   v12.4s, v4.4s, v1.s[0]      \n"
-                    "ldr    d0, [%2], #8                \n"
-                    "fmla   v13.4s, v4.4s, v1.s[1]      \n"
-                    "fmla   v14.4s, v4.4s, v1.s[2]      \n"
-                    "fmla   v15.4s, v4.4s, v1.s[3]      \n"
-                    "shll   v2.4s, v2.4h, #16           \n"
-                    "fmla   v24.4s, v5.4s, v1.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v1.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v1.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v1.s[3]      \n"
-                    "shll   v6.4s, v6.4h, #16           \n"
-                    "fmla   v16.4s, v4.4s, v2.s[0]      \n"
-                    "ldr    d1, [%2], #8                \n"
-                    "fmla   v17.4s, v4.4s, v2.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v2.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v2.s[3]      \n"
-                    "shll   v3.4s, v3.4h, #16           \n"
-                    "fmla   v28.4s, v5.4s, v2.s[0]      \n"
-                    "ldr    d4, [%1], #8                \n"
-                    "fmla   v29.4s, v5.4s, v2.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v2.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-                    "shll   v7.4s, v7.4h, #16           \n"
-                    "fmla   v8.4s, v6.4s, v3.s[0]       \n"
-                    "ldr    d2, [%2], #8                \n"
-                    "fmla   v9.4s, v6.4s, v3.s[1]       \n"
-                    "ldr    d5, [%1], #8                \n"
-                    "fmla   v10.4s, v6.4s, v3.s[2]      \n"
-                    "fmla   v11.4s, v6.4s, v3.s[3]      \n"
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "fmla   v20.4s, v7.4s, v3.s[0]      \n"
-                    "fmla   v21.4s, v7.4s, v3.s[1]      \n"
-                    "fmla   v22.4s, v7.4s, v3.s[2]      \n"
-                    "fmla   v23.4s, v7.4s, v3.s[3]      \n"
-                    "fmla   v12.4s, v6.4s, v0.s[0]      \n"
-                    "ldr    d3, [%2], #8                \n"
-                    "fmla   v13.4s, v6.4s, v0.s[1]      \n"
-                    "fmla   v14.4s, v6.4s, v0.s[2]      \n"
-                    "fmla   v15.4s, v6.4s, v0.s[3]      \n"
-                    "shll   v1.4s, v1.4h, #16           \n"
-                    "fmla   v24.4s, v7.4s, v0.s[0]      \n"
-                    "fmla   v25.4s, v7.4s, v0.s[1]      \n"
-                    "fmla   v26.4s, v7.4s, v0.s[2]      \n"
-                    "fmla   v27.4s, v7.4s, v0.s[3]      \n"
-                    "shll   v4.4s, v4.4h, #16           \n"
-                    "fmla   v16.4s, v6.4s, v1.s[0]      \n"
-                    "ldr    d0, [%2], #8                \n"
-                    "fmla   v17.4s, v6.4s, v1.s[1]      \n"
-                    "fmla   v18.4s, v6.4s, v1.s[2]      \n"
-                    "fmla   v19.4s, v6.4s, v1.s[3]      \n"
-                    "shll   v2.4s, v2.4h, #16           \n"
-                    "fmla   v28.4s, v7.4s, v1.s[0]      \n"
-                    "ldr    d6, [%1], #8                \n"
-                    "fmla   v29.4s, v7.4s, v1.s[1]      \n"
-                    "fmla   v30.4s, v7.4s, v1.s[2]      \n"
-                    "fmla   v31.4s, v7.4s, v1.s[3]      \n"
-                    "shll   v5.4s, v5.4h, #16           \n"
-                    "fmla   v8.4s, v4.4s, v2.s[0]       \n"
-                    "ldr    d1, [%2], #8                \n"
-                    "fmla   v9.4s, v4.4s, v2.s[1]       \n"
-                    "ldr    d7, [%1], #8                \n"
-                    "fmla   v10.4s, v4.4s, v2.s[2]      \n"
-                    "fmla   v11.4s, v4.4s, v2.s[3]      \n"
-                    "shll   v3.4s, v3.4h, #16           \n"
-                    "fmla   v20.4s, v5.4s, v2.s[0]      \n"
-                    "fmla   v21.4s, v5.4s, v2.s[1]      \n"
-                    "fmla   v22.4s, v5.4s, v2.s[2]      \n"
-                    "fmla   v23.4s, v5.4s, v2.s[3]      \n"
-                    "fmla   v12.4s, v4.4s, v3.s[0]      \n"
-                    "ldr    d2, [%2], #8                \n"
-                    "fmla   v13.4s, v4.4s, v3.s[1]      \n"
-                    "fmla   v14.4s, v4.4s, v3.s[2]      \n"
-                    "fmla   v15.4s, v4.4s, v3.s[3]      \n"
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "fmla   v24.4s, v5.4s, v3.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v3.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v3.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v3.s[3]      \n"
-                    "shll   v6.4s, v6.4h, #16           \n"
-                    "fmla   v16.4s, v4.4s, v0.s[0]      \n"
-                    "ldr    d3, [%2], #8                \n"
-                    "fmla   v17.4s, v4.4s, v0.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v0.s[2]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                    "fmla   v19.4s, v4.4s, v0.s[3]      \n"
-                    "shll   v1.4s, v1.4h, #16           \n"
-                    "fmla   v28.4s, v5.4s, v0.s[0]      \n"
-                    "ldr    d4, [%1], #8                \n"
-                    "fmla   v29.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v0.s[2]      \n"
-                    "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                    "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-                    "shll   v7.4s, v7.4h, #16           \n"
-                    "fmla   v8.4s, v6.4s, v1.s[0]       \n"
-                    "ldr    d0, [%2], #8                \n"
-                    "fmla   v9.4s, v6.4s, v1.s[1]       \n"
-                    "ldr    d5, [%1], #8                 \n"
-                    "fmla   v10.4s, v6.4s, v1.s[2]      \n"
-                    "fmla   v11.4s, v6.4s, v1.s[3]      \n"
-                    "shll   v2.4s, v2.4h, #16           \n"
-                    "fmla   v20.4s, v7.4s, v1.s[0]      \n"
-                    "fmla   v21.4s, v7.4s, v1.s[1]      \n"
-                    "fmla   v22.4s, v7.4s, v1.s[2]      \n"
-                    "fmla   v23.4s, v7.4s, v1.s[3]      \n"
-                    "fmla   v12.4s, v6.4s, v2.s[0]      \n"
-                    "ldr    d1, [%2], #8                 \n"
-                    "fmla   v13.4s, v6.4s, v2.s[1]      \n"
-                    "fmla   v14.4s, v6.4s, v2.s[2]      \n"
-                    "fmla   v15.4s, v6.4s, v2.s[3]      \n"
-                    "shll   v3.4s, v3.4h, #16           \n"
-                    "fmla   v24.4s, v7.4s, v2.s[0]      \n"
-                    "fmla   v25.4s, v7.4s, v2.s[1]      \n"
-                    "fmla   v26.4s, v7.4s, v2.s[2]      \n"
-                    "fmla   v27.4s, v7.4s, v2.s[3]      \n"
-                    "shll   v4.4s, v4.4h, #16           \n"
-                    "fmla   v16.4s, v6.4s, v3.s[0]      \n"
-                    "ldr    d2, [%2], #8                \n"
-                    "fmla   v17.4s, v6.4s, v3.s[1]      \n"
-                    "fmla   v18.4s, v6.4s, v3.s[2]      \n"
-                    "fmla   v19.4s, v6.4s, v3.s[3]      \n"
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "fmla   v28.4s, v7.4s, v3.s[0]      \n"
-                    "fmla   v29.4s, v7.4s, v3.s[1]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v30.4s, v7.4s, v3.s[2]      \n"
-                    "fmla   v31.4s, v7.4s, v3.s[3]      \n"
-                    "bne    4b                          \n"
-
-                    "sub    %1, %1, #16                 \n"
-                    "sub    %2, %2, #24                 \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h}, [%2], #24 \n"
-
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "shll   v1.4s, v1.4h, #16           \n"
-                    "shll   v2.4s, v2.4h, #16           \n"
-
-                    "ld1    {v4.4h, v5.4h}, [%1], #16   \n"
-
-                    "shll   v4.4s, v4.4h, #16           \n"
-                    "shll   v5.4s, v5.4h, #16           \n"
-
-                    "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                    "fmla   v9.4s, v4.4s, v0.s[1]       \n"
-                    "fmla   v10.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v11.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v12.4s, v4.4s, v1.s[0]      \n"
-                    "fmla   v13.4s, v4.4s, v1.s[1]      \n"
-                    "fmla   v14.4s, v4.4s, v1.s[2]      \n"
-                    "fmla   v15.4s, v4.4s, v1.s[3]      \n"
-                    "fmla   v16.4s, v4.4s, v2.s[0]      \n"
-                    "fmla   v17.4s, v4.4s, v2.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v2.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v2.s[3]      \n"
-
-                    "subs   w4, w4, #1                  \n"
-
-                    "fmla   v20.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v21.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v22.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v23.4s, v5.4s, v0.s[3]      \n"
-                    "fmla   v24.4s, v5.4s, v1.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v1.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v1.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v1.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v2.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v2.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v2.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "shrn   v0.4h, v8.4s, #16           \n"
-                    "shrn2  v0.8h, v9.4s, #16           \n"
-                    "shrn   v1.4h, v10.4s, #16          \n"
-                    "shrn2  v1.8h, v11.4s, #16          \n"
-                    "shrn   v2.4h, v12.4s, #16          \n"
-                    "shrn2  v2.8h, v13.4s, #16          \n"
-                    "shrn   v3.4h, v14.4s, #16          \n"
-                    "shrn2  v3.8h, v15.4s, #16          \n"
-                    "shrn   v4.4h, v16.4s, #16          \n"
-                    "shrn2  v4.8h, v17.4s, #16          \n"
-                    "shrn   v5.4h, v18.4s, #16          \n"
-                    "shrn2  v5.8h, v19.4s, #16          \n"
-                    "shrn   v6.4h, v20.4s, #16          \n"
-                    "shrn2  v6.8h, v21.4s, #16          \n"
-                    "shrn   v7.4h, v22.4s, #16          \n"
-                    "shrn2  v7.8h, v23.4s, #16          \n"
-                    "shrn   v8.4h, v24.4s, #16          \n"
-                    "shrn2  v8.8h, v25.4s, #16          \n"
-                    "shrn   v9.4h, v26.4s, #16          \n"
-                    "shrn2  v9.8h, v27.4s, #16          \n"
-                    "shrn   v10.4h, v28.4s, #16         \n"
-                    "shrn2  v10.8h, v29.4s, #16         \n"
-                    "shrn   v11.4h, v30.4s, #16         \n"
-                    "shrn2  v11.8h, v31.4s, #16         \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    10f                         \n"
-
-                    // if out_elempack == 4
-                    "cmp    %w12, #4                    \n"
-                    "bne    8f                          \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 1          \n"
-                    "st1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%3], #64 \n"
-                    "st1    {v4.8h, v5.8h}, [%3], #32 \n"
-                    "st1    {v6.8h, v7.8h, v8.8h, v9.8h}, [x4], #64 \n"
-                    "st1    {v10.8h, v11.8h}, [x4]      \n"
-                    "b      9f                          \n"
-
-                    // if out_elempack == 1
-                    "8:                                 \n"
-                    // transpose8x12
-                    "uzp1   v20.8h, v0.8h, v1.8h        \n"
-                    "uzp2   v21.8h, v0.8h, v1.8h        \n"
-                    "uzp1   v22.8h, v2.8h, v3.8h        \n"
-                    "uzp2   v23.8h, v2.8h, v3.8h        \n"
-                    "uzp1   v24.8h, v4.8h, v5.8h        \n"
-                    "uzp2   v25.8h, v4.8h, v5.8h        \n"
-                    "uzp1   v26.8h, v6.8h, v7.8h        \n"
-                    "uzp2   v27.8h, v6.8h, v7.8h        \n"
-                    "uzp1   v28.8h, v8.8h, v9.8h        \n"
-                    "uzp2   v29.8h, v8.8h, v9.8h        \n"
-                    "uzp1   v30.8h, v10.8h, v11.8h      \n"
-                    "uzp2   v31.8h, v10.8h, v11.8h      \n"
-
-                    "uzp1   v0.8h, v20.8h, v22.8h       \n"
-                    "uzp2   v6.8h, v20.8h, v22.8h       \n"
-                    "uzp1   v3.8h, v21.8h, v23.8h       \n"
-                    "uzp2   v9.8h, v21.8h, v23.8h       \n"
-                    "mov    v1.d[0], v0.d[1]            \n"
-                    "mov    v7.d[0], v6.d[1]            \n"
-                    "mov    v4.d[0], v3.d[1]            \n"
-                    "mov    v10.d[0], v9.d[1]           \n"
-                    "uzp1   v2.8h, v24.8h, v24.8h       \n"
-                    "uzp2   v8.8h, v24.8h, v24.8h       \n"
-                    "uzp1   v5.8h, v25.8h, v25.8h       \n"
-                    "uzp2   v11.8h, v25.8h, v25.8h      \n"
-
-                    "uzp1   v12.8h, v26.8h, v28.8h      \n"
-                    "uzp2   v18.8h, v26.8h, v28.8h      \n"
-                    "uzp1   v15.8h, v27.8h, v29.8h      \n"
-                    "uzp2   v21.8h, v27.8h, v29.8h      \n"
-                    "mov    v13.d[0], v12.d[1]          \n"
-                    "mov    v19.d[0], v18.d[1]          \n"
-                    "mov    v16.d[0], v15.d[1]          \n"
-                    "mov    v22.d[0], v21.d[1]          \n"
-                    "uzp1   v14.8h, v30.8h, v30.8h      \n"
-                    "uzp2   v20.8h, v30.8h, v30.8h      \n"
-                    "uzp1   v17.8h, v31.8h, v31.8h      \n"
-                    "uzp2   v23.8h, v31.8h, v31.8h      \n"
-
-                    "add    x4, %3, %w13, sxtw 1        \n"
-                    "st1    {v0.4h, v1.4h, v2.4h}, [%3], #24 \n"
-                    "st1    {v3.4h, v4.4h, v5.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v6.4h, v7.4h, v8.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v9.4h, v10.4h, v11.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v12.4h, v13.4h, v14.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v15.4h, v16.4h, v17.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v18.4h, v19.4h, v20.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v21.4h, v22.4h, v23.4h}, [x4] \n"
-
-                    "9:                                 \n"
-                    "add    %0, %0, #384                \n"
-                    "b      11f                         \n"
-
-                    "10:                                \n"
-                    "st1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                    "st1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    "11:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            else if (use_a53_a55_optimized_kernel && !cpu_support_arm_asimdhp())
-            {
-                // a53
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #320                \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "add    x4, %8, #16                 \n"
-                    "ld1    {v8.4s}, [%8]               \n"
-                    "ld1    {v20.4s}, [x4]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v8.16b, v8.16b, v8.16b      \n"
-                    "eor    v20.16b, v20.16b, v20.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v9.16b, v8.16b              \n"
-                    "mov    v10.16b, v8.16b             \n"
-                    "mov    v11.16b, v8.16b             \n"
-                    "mov    v12.16b, v8.16b             \n"
-                    "mov    v13.16b, v8.16b             \n"
-                    "mov    v14.16b, v8.16b             \n"
-                    "mov    v15.16b, v8.16b             \n"
-                    "mov    v16.16b, v8.16b             \n"
-                    "mov    v17.16b, v8.16b             \n"
-                    "mov    v18.16b, v8.16b             \n"
-                    "mov    v19.16b, v8.16b             \n"
-
-                    "mov    v21.16b, v20.16b            \n"
-                    "mov    v22.16b, v20.16b            \n"
-                    "mov    v23.16b, v20.16b            \n"
-                    "mov    v24.16b, v20.16b            \n"
-                    "mov    v25.16b, v20.16b            \n"
-                    "mov    v26.16b, v20.16b            \n"
-                    "mov    v27.16b, v20.16b            \n"
-                    "mov    v28.16b, v20.16b            \n"
-                    "mov    v29.16b, v20.16b            \n"
-                    "mov    v30.16b, v20.16b            \n"
-                    "mov    v31.16b, v20.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v4.4h}, [%1], #8           \n"
-
-                    "prfm   pldl1keep, [%2, #384]       \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h}, [%2], #24 \n"
-
-                    "ldr    x25, [%1]                   \n"
-                    "add    %1, %1, #8                  \n"
-
-                    "shll   v4.4s, v4.4h, #16           \n"
-                    "shll   v0.4s, v0.4h, #16           \n"
-
-                    ".align 4                           \n"
-                    "4:                                 \n"
-
-                    "shll   v1.4s, v1.4h, #16           \n"
-                    "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                    "ldr    x26, [%1]                   \n"
-                    "fmla   v9.4s, v4.4s, v0.s[1]       \n"
-                    "add    %1, %1, #8                  \n"
-                    "fmla   v10.4s, v4.4s, v0.s[2]      \n"
-
-                    "shll   v2.4s, v2.4h, #16           \n"
-                    "fmla   v11.4s, v4.4s, v0.s[3]      \n"
-                    "ldr    x23, [%2]                   \n"
-                    "fmla   v12.4s, v4.4s, v1.s[0]      \n"
-                    "add    %2, %2, #8                  \n"
-                    "fmla   v13.4s, v4.4s, v1.s[1]      \n"
-
-                    "nop                                \n"
-                    "ins    v5.d[0], x25                \n"
-                    "fmla   v14.4s, v4.4s, v1.s[2]      \n"
-                    "ldr    x20, [%2]                   \n"
-                    "fmla   v15.4s, v4.4s, v1.s[3]      \n"
-                    "add    %2, %2, #8                  \n"
-                    "fmla   v16.4s, v4.4s, v2.s[0]      \n"
-
-                    "shll   v5.4s, v5.4h, #16           \n"
-                    "fmla   v17.4s, v4.4s, v2.s[1]      \n"
-                    "ldr    x21, [%2]                   \n"
-                    "fmla   v18.4s, v4.4s, v2.s[2]      \n"
-                    "add    %2, %2, #8                  \n"
-                    "fmla   v19.4s, v4.4s, v2.s[3]      \n"
-
-                    "ins    v6.d[0], x26                \n"
-                    "ins    v3.d[0], x23                \n"
-                    "fmla   v20.4s, v5.4s, v0.s[0]      \n"
-                    "prfm   pldl1keep, [%2, #384]       \n" // NOTE PRELOAD
-                    "fmla   v21.4s, v5.4s, v0.s[1]      \n"
-                    "nop                                \n"
-                    "fmla   v22.4s, v5.4s, v0.s[2]      \n"
-
-                    "shll   v6.4s, v6.4h, #16           \n"
-                    "fmla   v23.4s, v5.4s, v0.s[3]      \n"
-                    "ldr    x27, [%1]                   \n"
-                    "fmla   v24.4s, v5.4s, v1.s[0]      \n"
-                    "add    %1, %1, #8                  \n"
-                    "fmla   v25.4s, v5.4s, v1.s[1]      \n"
-
-                    "shll   v3.4s, v3.4h, #16           \n"
-                    "fmla   v26.4s, v5.4s, v1.s[2]      \n"
-                    "prfm   pldl1keep, [%1, #256]       \n" // NOTE PRELOAD
-                    "fmla   v27.4s, v5.4s, v1.s[3]      \n"
-                    "nop                                \n"
-                    "fmla   v28.4s, v5.4s, v2.s[0]      \n"
-
-                    "ins    v0.d[0], x20                \n"
-                    "ins    v1.d[0], x21                \n"
-                    "fmla   v29.4s, v5.4s, v2.s[1]      \n"
-                    "nop                                \n"
-                    "fmla   v30.4s, v5.4s, v2.s[2]      \n"
-                    "nop                                \n"
-                    "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "fmla   v8.4s, v6.4s, v3.s[0]       \n"
-                    "ldr    x24, [%1]                   \n"
-                    "fmla   v9.4s, v6.4s, v3.s[1]       \n"
-                    "add    %1, %1, #8                  \n"
-                    "fmla   v10.4s, v6.4s, v3.s[2]      \n"
-
-                    "shll   v1.4s, v1.4h, #16           \n"
-                    "fmla   v11.4s, v6.4s, v3.s[3]      \n"
-                    "ldr    x22, [%2]                   \n"
-                    "fmla   v12.4s, v6.4s, v0.s[0]      \n"
-                    "add    %2, %2, #8                  \n"
-                    "fmla   v13.4s, v6.4s, v0.s[1]      \n"
-
-                    "nop                                \n"
-                    "ins    v7.d[0], x27                \n"
-                    "fmla   v14.4s, v6.4s, v0.s[2]      \n"
-                    "ldr    x23, [%2]                   \n"
-                    "fmla   v15.4s, v6.4s, v0.s[3]      \n"
-                    "add    %2, %2, #8                  \n"
-                    "fmla   v16.4s, v6.4s, v1.s[0]      \n"
-
-                    "shll   v7.4s, v7.4h, #16           \n"
-                    "fmla   v17.4s, v6.4s, v1.s[1]      \n"
-                    "ldr    x20, [%2]                   \n"
-                    "fmla   v18.4s, v6.4s, v1.s[2]      \n"
-                    "add    %2, %2, #8                  \n"
-                    "fmla   v19.4s, v6.4s, v1.s[3]      \n"
-
-                    "ins    v4.d[0], x24                \n"
-                    "ins    v2.d[0], x22                \n"
-                    "fmla   v20.4s, v7.4s, v3.s[0]      \n"
-                    "nop                                \n"
-                    "fmla   v21.4s, v7.4s, v3.s[1]      \n"
-                    "nop                                \n"
-                    "fmla   v22.4s, v7.4s, v3.s[2]      \n"
-
-                    "shll   v4.4s, v4.4h, #16           \n"
-                    "fmla   v23.4s, v7.4s, v3.s[3]      \n"
-                    "ldr    x25, [%1]                   \n"
-                    "fmla   v24.4s, v7.4s, v0.s[0]      \n"
-                    "add    %1, %1, #8                  \n"
-                    "fmla   v25.4s, v7.4s, v0.s[1]      \n"
-
-                    "shll   v2.4s, v2.4h, #16           \n"
-                    "fmla   v26.4s, v7.4s, v0.s[2]      \n"
-                    "nop                                \n"
-                    "fmla   v27.4s, v7.4s, v0.s[3]      \n"
-                    "nop                                \n"
-                    "fmla   v28.4s, v7.4s, v1.s[0]      \n"
-
-                    "ins    v3.d[0], x23                \n"
-                    "ins    v0.d[0], x20                \n"
-                    "fmla   v29.4s, v7.4s, v1.s[1]      \n"
-                    "nop                                \n"
-                    "fmla   v30.4s, v7.4s, v1.s[2]      \n"
-                    "nop                                \n"
-                    "fmla   v31.4s, v7.4s, v1.s[3]      \n"
-
-                    "shll   v3.4s, v3.4h, #16           \n"
-                    "fmla   v8.4s, v4.4s, v2.s[0]       \n"
-                    "ldr    x26, [%1]                   \n"
-                    "fmla   v9.4s, v4.4s, v2.s[1]       \n"
-                    "add    %1, %1, #8                  \n"
-                    "fmla   v10.4s, v4.4s, v2.s[2]      \n"
-
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "fmla   v11.4s, v4.4s, v2.s[3]      \n"
-                    "ldr    x21, [%2]                   \n"
-                    "fmla   v12.4s, v4.4s, v3.s[0]      \n"
-                    "add    %2, %2, #8                  \n"
-                    "fmla   v13.4s, v4.4s, v3.s[1]      \n"
-
-                    "nop                                \n"
-                    "ins    v5.d[0], x25                \n"
-                    "fmla   v14.4s, v4.4s, v3.s[2]      \n"
-                    "ldr    x22, [%2]                   \n"
-                    "fmla   v15.4s, v4.4s, v3.s[3]      \n"
-                    "add    %2, %2, #8                  \n"
-                    "fmla   v16.4s, v4.4s, v0.s[0]      \n"
-
-                    "shll   v5.4s, v5.4h, #16           \n"
-                    "fmla   v17.4s, v4.4s, v0.s[1]      \n"
-                    "ldr    x23, [%2]                   \n"
-                    "fmla   v18.4s, v4.4s, v0.s[2]      \n"
-                    "add    %2, %2, #8                  \n"
-                    "fmla   v19.4s, v4.4s, v0.s[3]      \n"
-
-                    "ins    v6.d[0], x26                \n"
-                    "ins    v1.d[0], x21                \n"
-                    "fmla   v20.4s, v5.4s, v2.s[0]      \n"
-                    "prfm   pldl1keep, [%2, #384]       \n" // NOTE PRELOAD
-                    "fmla   v21.4s, v5.4s, v2.s[1]      \n"
-                    "nop                                \n"
-                    "fmla   v22.4s, v5.4s, v2.s[2]      \n"
-
-                    "shll   v6.4s, v6.4h, #16           \n"
-                    "fmla   v23.4s, v5.4s, v2.s[3]      \n"
-                    "ldr    x27, [%1]                   \n"
-                    "fmla   v24.4s, v5.4s, v3.s[0]      \n"
-                    "add    %1, %1, #8                  \n"
-                    "fmla   v25.4s, v5.4s, v3.s[1]      \n"
-
-                    "shll   v1.4s, v1.4h, #16           \n"
-                    "fmla   v26.4s, v5.4s, v3.s[2]      \n"
-                    "prfm   pldl1keep, [%1, #256]       \n" // NOTE PRELOAD
-                    "fmla   v27.4s, v5.4s, v3.s[3]      \n"
-                    "nop                                \n"
-                    "fmla   v28.4s, v5.4s, v0.s[0]      \n"
-
-                    "ins    v2.d[0], x22                \n"
-                    "ins    v3.d[0], x23                \n"
-                    "fmla   v29.4s, v5.4s, v0.s[1]      \n"
-                    "nop                                \n"
-                    "fmla   v30.4s, v5.4s, v0.s[2]      \n"
-                    "nop                                \n"
-                    "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-
-                    "shll   v2.4s, v2.4h, #16           \n"
-                    "fmla   v8.4s, v6.4s, v1.s[0]       \n"
-                    "ldr    x24, [%1]                   \n"
-                    "fmla   v9.4s, v6.4s, v1.s[1]       \n"
-                    "add    %1, %1, #8                  \n"
-                    "fmla   v10.4s, v6.4s, v1.s[2]      \n"
-
-                    "shll   v3.4s, v3.4h, #16           \n"
-                    "fmla   v11.4s, v6.4s, v1.s[3]      \n"
-                    "ldr    x20, [%2]                   \n"
-                    "fmla   v12.4s, v6.4s, v2.s[0]      \n"
-                    "add    %2, %2, #8                  \n"
-                    "fmla   v13.4s, v6.4s, v2.s[1]      \n"
-
-                    "nop                                \n"
-                    "ins    v7.d[0], x27                \n"
-                    "fmla   v14.4s, v6.4s, v2.s[2]      \n"
-                    "ldr    x21, [%2]                   \n"
-                    "fmla   v15.4s, v6.4s, v2.s[3]      \n"
-                    "add    %2, %2, #8                  \n"
-                    "fmla   v16.4s, v6.4s, v3.s[0]      \n"
-
-                    "shll   v7.4s, v7.4h, #16           \n"
-                    "fmla   v17.4s, v6.4s, v3.s[1]      \n"
-                    "ldr    x22, [%2]                   \n"
-                    "fmla   v18.4s, v6.4s, v3.s[2]      \n"
-                    "add    %2, %2, #8                  \n"
-                    "fmla   v19.4s, v6.4s, v3.s[3]      \n"
-
-                    "ins    v4.d[0], x24                \n"
-                    "ins    v0.d[0], x20                \n"
-                    "fmla   v20.4s, v7.4s, v1.s[0]      \n"
-                    "nop                                \n"
-                    "fmla   v21.4s, v7.4s, v1.s[1]      \n"
-                    "nop                                \n"
-                    "fmla   v22.4s, v7.4s, v1.s[2]      \n"
-
-                    "shll   v4.4s, v4.4h, #16           \n"
-                    "fmla   v23.4s, v7.4s, v1.s[3]      \n"
-                    "ldr    x25, [%1]                   \n"
-                    "fmla   v24.4s, v7.4s, v2.s[0]      \n"
-                    "add    %1, %1, #8                  \n"
-                    "fmla   v25.4s, v7.4s, v2.s[1]      \n"
-
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "fmla   v26.4s, v7.4s, v2.s[2]      \n"
-                    "nop                                \n"
-                    "fmla   v27.4s, v7.4s, v2.s[3]      \n"
-                    "nop                                \n"
-                    "fmla   v28.4s, v7.4s, v3.s[0]      \n"
-
-                    "ins    v1.d[0], x21                \n"
-                    "ins    v2.d[0], x22                \n"
-                    "fmla   v29.4s, v7.4s, v3.s[1]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v30.4s, v7.4s, v3.s[2]      \n"
-                    "nop                                \n"
-                    "fmla   v31.4s, v7.4s, v3.s[3]      \n"
-
-                    "bne    4b                          \n"
-
-                    "sub    %1, %1, #16                 \n"
-                    "sub    %2, %2, #24                 \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h}, [%2], #24 \n"
-
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "shll   v1.4s, v1.4h, #16           \n"
-                    "shll   v2.4s, v2.4h, #16           \n"
-
-                    "ld1    {v4.4h, v5.4h}, [%1], #16   \n"
-
-                    "shll   v4.4s, v4.4h, #16           \n"
-                    "shll   v5.4s, v5.4h, #16           \n"
-
-                    "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                    "fmla   v9.4s, v4.4s, v0.s[1]       \n"
-                    "fmla   v10.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v11.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v12.4s, v4.4s, v1.s[0]      \n"
-                    "fmla   v13.4s, v4.4s, v1.s[1]      \n"
-                    "fmla   v14.4s, v4.4s, v1.s[2]      \n"
-                    "fmla   v15.4s, v4.4s, v1.s[3]      \n"
-                    "fmla   v16.4s, v4.4s, v2.s[0]      \n"
-                    "fmla   v17.4s, v4.4s, v2.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v2.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v2.s[3]      \n"
-
-                    "subs   w4, w4, #1                  \n"
-
-                    "fmla   v20.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v21.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v22.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v23.4s, v5.4s, v0.s[3]      \n"
-                    "fmla   v24.4s, v5.4s, v1.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v1.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v1.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v1.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v2.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v2.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v2.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "shrn   v0.4h, v8.4s, #16           \n"
-                    "shrn2  v0.8h, v9.4s, #16           \n"
-                    "shrn   v1.4h, v10.4s, #16          \n"
-                    "shrn2  v1.8h, v11.4s, #16          \n"
-                    "shrn   v2.4h, v12.4s, #16          \n"
-                    "shrn2  v2.8h, v13.4s, #16          \n"
-                    "shrn   v3.4h, v14.4s, #16          \n"
-                    "shrn2  v3.8h, v15.4s, #16          \n"
-                    "shrn   v4.4h, v16.4s, #16          \n"
-                    "shrn2  v4.8h, v17.4s, #16          \n"
-                    "shrn   v5.4h, v18.4s, #16          \n"
-                    "shrn2  v5.8h, v19.4s, #16          \n"
-                    "shrn   v6.4h, v20.4s, #16          \n"
-                    "shrn2  v6.8h, v21.4s, #16          \n"
-                    "shrn   v7.4h, v22.4s, #16          \n"
-                    "shrn2  v7.8h, v23.4s, #16          \n"
-                    "shrn   v8.4h, v24.4s, #16          \n"
-                    "shrn2  v8.8h, v25.4s, #16          \n"
-                    "shrn   v9.4h, v26.4s, #16          \n"
-                    "shrn2  v9.8h, v27.4s, #16          \n"
-                    "shrn   v10.4h, v28.4s, #16         \n"
-                    "shrn2  v10.8h, v29.4s, #16         \n"
-                    "shrn   v11.4h, v30.4s, #16         \n"
-                    "shrn2  v11.8h, v31.4s, #16         \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    10f                         \n"
-
-                    // if out_elempack == 4
-                    "cmp    %w12, #4                    \n"
-                    "bne    8f                          \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 1          \n"
-                    "st1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%3], #64 \n"
-                    "st1    {v4.8h, v5.8h}, [%3], #32 \n"
-                    "st1    {v6.8h, v7.8h, v8.8h, v9.8h}, [x4], #64 \n"
-                    "st1    {v10.8h, v11.8h}, [x4]      \n"
-                    "b      9f                          \n"
-
-                    // if out_elempack == 1
-                    "8:                                 \n"
-                    // transpose8x12
-                    "uzp1   v20.8h, v0.8h, v1.8h        \n"
-                    "uzp2   v21.8h, v0.8h, v1.8h        \n"
-                    "uzp1   v22.8h, v2.8h, v3.8h        \n"
-                    "uzp2   v23.8h, v2.8h, v3.8h        \n"
-                    "uzp1   v24.8h, v4.8h, v5.8h        \n"
-                    "uzp2   v25.8h, v4.8h, v5.8h        \n"
-                    "uzp1   v26.8h, v6.8h, v7.8h        \n"
-                    "uzp2   v27.8h, v6.8h, v7.8h        \n"
-                    "uzp1   v28.8h, v8.8h, v9.8h        \n"
-                    "uzp2   v29.8h, v8.8h, v9.8h        \n"
-                    "uzp1   v30.8h, v10.8h, v11.8h      \n"
-                    "uzp2   v31.8h, v10.8h, v11.8h      \n"
-
-                    "uzp1   v0.8h, v20.8h, v22.8h       \n"
-                    "uzp2   v6.8h, v20.8h, v22.8h       \n"
-                    "uzp1   v3.8h, v21.8h, v23.8h       \n"
-                    "uzp2   v9.8h, v21.8h, v23.8h       \n"
-                    "mov    v1.d[0], v0.d[1]            \n"
-                    "mov    v7.d[0], v6.d[1]            \n"
-                    "mov    v4.d[0], v3.d[1]            \n"
-                    "mov    v10.d[0], v9.d[1]           \n"
-                    "uzp1   v2.8h, v24.8h, v24.8h       \n"
-                    "uzp2   v8.8h, v24.8h, v24.8h       \n"
-                    "uzp1   v5.8h, v25.8h, v25.8h       \n"
-                    "uzp2   v11.8h, v25.8h, v25.8h      \n"
-
-                    "uzp1   v12.8h, v26.8h, v28.8h      \n"
-                    "uzp2   v18.8h, v26.8h, v28.8h      \n"
-                    "uzp1   v15.8h, v27.8h, v29.8h      \n"
-                    "uzp2   v21.8h, v27.8h, v29.8h      \n"
-                    "mov    v13.d[0], v12.d[1]          \n"
-                    "mov    v19.d[0], v18.d[1]          \n"
-                    "mov    v16.d[0], v15.d[1]          \n"
-                    "mov    v22.d[0], v21.d[1]          \n"
-                    "uzp1   v14.8h, v30.8h, v30.8h      \n"
-                    "uzp2   v20.8h, v30.8h, v30.8h      \n"
-                    "uzp1   v17.8h, v31.8h, v31.8h      \n"
-                    "uzp2   v23.8h, v31.8h, v31.8h      \n"
-
-                    "add    x4, %3, %w13, sxtw 1        \n"
-                    "st1    {v0.4h, v1.4h, v2.4h}, [%3], #24 \n"
-                    "st1    {v3.4h, v4.4h, v5.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v6.4h, v7.4h, v8.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v9.4h, v10.4h, v11.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v12.4h, v13.4h, v14.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v15.4h, v16.4h, v17.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v18.4h, v19.4h, v20.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v21.4h, v22.4h, v23.4h}, [x4] \n"
-
-                    "9:                                 \n"
-                    "add    %0, %0, #384                \n"
-                    "b      11f                         \n"
-
-                    "10:                                \n"
-                    "st1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                    "st1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    "11:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            else
-            {
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #320                \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "add    x4, %8, #16                 \n"
-                    "ld1    {v8.4s}, [%8]               \n"
-                    "ld1    {v20.4s}, [x4]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v8.16b, v8.16b, v8.16b      \n"
-                    "eor    v20.16b, v20.16b, v20.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v9.16b, v8.16b              \n"
-                    "mov    v10.16b, v8.16b             \n"
-                    "mov    v11.16b, v8.16b             \n"
-                    "mov    v12.16b, v8.16b             \n"
-                    "mov    v13.16b, v8.16b             \n"
-                    "mov    v14.16b, v8.16b             \n"
-                    "mov    v15.16b, v8.16b             \n"
-                    "mov    v16.16b, v8.16b             \n"
-                    "mov    v17.16b, v8.16b             \n"
-                    "mov    v18.16b, v8.16b             \n"
-                    "mov    v19.16b, v8.16b             \n"
-
-                    "mov    v21.16b, v20.16b            \n"
-                    "mov    v22.16b, v20.16b            \n"
-                    "mov    v23.16b, v20.16b            \n"
-                    "mov    v24.16b, v20.16b            \n"
-                    "mov    v25.16b, v20.16b            \n"
-                    "mov    v26.16b, v20.16b            \n"
-                    "mov    v27.16b, v20.16b            \n"
-                    "mov    v28.16b, v20.16b            \n"
-                    "mov    v29.16b, v20.16b            \n"
-                    "mov    v30.16b, v20.16b            \n"
-                    "mov    v31.16b, v20.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%1], #32 \n"
-
-                    "shll   v4.4s, v4.4h, #16           \n"
-                    "shll   v5.4s, v5.4h, #16           \n"
-                    "shll   v6.4s, v6.4h, #16           \n"
-                    "shll   v7.4s, v7.4h, #16           \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n"
-
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "shll   v1.4s, v1.4h, #16           \n"
-                    "shll   v2.4s, v2.4h, #16           \n"
-                    "shll   v3.4s, v3.4h, #16           \n"
-
-                    "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                    "fmla   v9.4s, v4.4s, v0.s[1]       \n"
-                    "fmla   v10.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v11.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v20.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v21.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v22.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v23.4s, v5.4s, v0.s[3]      \n"
-
-                    "fmla   v12.4s, v4.4s, v1.s[0]      \n"
-                    "fmla   v13.4s, v4.4s, v1.s[1]      \n"
-                    "fmla   v14.4s, v4.4s, v1.s[2]      \n"
-                    "fmla   v15.4s, v4.4s, v1.s[3]      \n"
-                    "fmla   v24.4s, v5.4s, v1.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v1.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v1.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v1.s[3]      \n"
-
-                    "fmla   v16.4s, v4.4s, v2.s[0]      \n"
-                    "fmla   v17.4s, v4.4s, v2.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v2.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v2.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v2.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v2.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v2.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-
-                    "fmla   v8.4s, v6.4s, v3.s[0]       \n"
-                    "fmla   v9.4s, v6.4s, v3.s[1]       \n"
-                    "fmla   v10.4s, v6.4s, v3.s[2]      \n"
-                    "fmla   v11.4s, v6.4s, v3.s[3]      \n"
-                    "fmla   v20.4s, v7.4s, v3.s[0]      \n"
-                    "fmla   v21.4s, v7.4s, v3.s[1]      \n"
-                    "fmla   v22.4s, v7.4s, v3.s[2]      \n"
-                    "fmla   v23.4s, v7.4s, v3.s[3]      \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n"
-
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "shll   v1.4s, v1.4h, #16           \n"
-                    "shll   v2.4s, v2.4h, #16           \n"
-                    "shll   v3.4s, v3.4h, #16           \n"
-
-                    "fmla   v12.4s, v6.4s, v0.s[0]      \n"
-                    "fmla   v13.4s, v6.4s, v0.s[1]      \n"
-                    "fmla   v14.4s, v6.4s, v0.s[2]      \n"
-                    "fmla   v15.4s, v6.4s, v0.s[3]      \n"
-                    "fmla   v24.4s, v7.4s, v0.s[0]      \n"
-                    "fmla   v25.4s, v7.4s, v0.s[1]      \n"
-                    "fmla   v26.4s, v7.4s, v0.s[2]      \n"
-                    "fmla   v27.4s, v7.4s, v0.s[3]      \n"
-
-                    "fmla   v16.4s, v6.4s, v1.s[0]      \n"
-                    "fmla   v17.4s, v6.4s, v1.s[1]      \n"
-                    "fmla   v18.4s, v6.4s, v1.s[2]      \n"
-                    "fmla   v19.4s, v6.4s, v1.s[3]      \n"
-                    "fmla   v28.4s, v7.4s, v1.s[0]      \n"
-                    "fmla   v29.4s, v7.4s, v1.s[1]      \n"
-                    "fmla   v30.4s, v7.4s, v1.s[2]      \n"
-                    "fmla   v31.4s, v7.4s, v1.s[3]      \n"
-
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%1], #32 \n"
-
-                    "shll   v4.4s, v4.4h, #16           \n"
-                    "shll   v5.4s, v5.4h, #16           \n"
-                    "shll   v6.4s, v6.4h, #16           \n"
-                    "shll   v7.4s, v7.4h, #16           \n"
-
-                    "fmla   v8.4s, v4.4s, v2.s[0]       \n"
-                    "fmla   v9.4s, v4.4s, v2.s[1]       \n"
-                    "fmla   v10.4s, v4.4s, v2.s[2]      \n"
-                    "fmla   v11.4s, v4.4s, v2.s[3]      \n"
-                    "fmla   v20.4s, v5.4s, v2.s[0]      \n"
-                    "fmla   v21.4s, v5.4s, v2.s[1]      \n"
-                    "fmla   v22.4s, v5.4s, v2.s[2]      \n"
-                    "fmla   v23.4s, v5.4s, v2.s[3]      \n"
-
-                    "fmla   v12.4s, v4.4s, v3.s[0]      \n"
-                    "fmla   v13.4s, v4.4s, v3.s[1]      \n"
-                    "fmla   v14.4s, v4.4s, v3.s[2]      \n"
-                    "fmla   v15.4s, v4.4s, v3.s[3]      \n"
-                    "fmla   v24.4s, v5.4s, v3.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v3.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v3.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v3.s[3]      \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n"
-
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "shll   v1.4s, v1.4h, #16           \n"
-                    "shll   v2.4s, v2.4h, #16           \n"
-                    "shll   v3.4s, v3.4h, #16           \n"
-
-                    "fmla   v16.4s, v4.4s, v0.s[0]      \n"
-                    "fmla   v17.4s, v4.4s, v0.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-
-                    "fmla   v8.4s, v6.4s, v1.s[0]       \n"
-                    "fmla   v9.4s, v6.4s, v1.s[1]       \n"
-                    "fmla   v10.4s, v6.4s, v1.s[2]      \n"
-                    "fmla   v11.4s, v6.4s, v1.s[3]      \n"
-                    "fmla   v20.4s, v7.4s, v1.s[0]      \n"
-                    "fmla   v21.4s, v7.4s, v1.s[1]      \n"
-                    "fmla   v22.4s, v7.4s, v1.s[2]      \n"
-                    "fmla   v23.4s, v7.4s, v1.s[3]      \n"
-
-                    "fmla   v12.4s, v6.4s, v2.s[0]      \n"
-                    "fmla   v13.4s, v6.4s, v2.s[1]      \n"
-                    "fmla   v14.4s, v6.4s, v2.s[2]      \n"
-                    "fmla   v15.4s, v6.4s, v2.s[3]      \n"
-                    "fmla   v24.4s, v7.4s, v2.s[0]      \n"
-                    "fmla   v25.4s, v7.4s, v2.s[1]      \n"
-                    "fmla   v26.4s, v7.4s, v2.s[2]      \n"
-                    "fmla   v27.4s, v7.4s, v2.s[3]      \n"
-
-                    "subs   w4, w4, #1                  \n"
-
-                    "fmla   v16.4s, v6.4s, v3.s[0]      \n"
-                    "fmla   v17.4s, v6.4s, v3.s[1]      \n"
-                    "fmla   v18.4s, v6.4s, v3.s[2]      \n"
-                    "fmla   v19.4s, v6.4s, v3.s[3]      \n"
-                    "fmla   v28.4s, v7.4s, v3.s[0]      \n"
-                    "fmla   v29.4s, v7.4s, v3.s[1]      \n"
-                    "fmla   v30.4s, v7.4s, v3.s[2]      \n"
-                    "fmla   v31.4s, v7.4s, v3.s[3]      \n"
-
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h}, [%2], #24 \n"
-
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "shll   v1.4s, v1.4h, #16           \n"
-                    "shll   v2.4s, v2.4h, #16           \n"
-
-                    "ld1    {v4.4h, v5.4h}, [%1], #16   \n"
-
-                    "shll   v4.4s, v4.4h, #16           \n"
-                    "shll   v5.4s, v5.4h, #16           \n"
-
-                    "fmla   v8.4s, v4.4s, v0.s[0]       \n"
-                    "fmla   v9.4s, v4.4s, v0.s[1]       \n"
-                    "fmla   v10.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v11.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v12.4s, v4.4s, v1.s[0]      \n"
-                    "fmla   v13.4s, v4.4s, v1.s[1]      \n"
-                    "fmla   v14.4s, v4.4s, v1.s[2]      \n"
-                    "fmla   v15.4s, v4.4s, v1.s[3]      \n"
-                    "fmla   v16.4s, v4.4s, v2.s[0]      \n"
-                    "fmla   v17.4s, v4.4s, v2.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v2.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v2.s[3]      \n"
-
-                    "subs   w4, w4, #1                  \n"
-
-                    "fmla   v20.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v21.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v22.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v23.4s, v5.4s, v0.s[3]      \n"
-                    "fmla   v24.4s, v5.4s, v1.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v1.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v1.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v1.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v2.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v2.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v2.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v2.s[3]      \n"
-
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "shrn   v0.4h, v8.4s, #16           \n"
-                    "shrn2  v0.8h, v9.4s, #16           \n"
-                    "shrn   v1.4h, v10.4s, #16          \n"
-                    "shrn2  v1.8h, v11.4s, #16          \n"
-                    "shrn   v2.4h, v12.4s, #16          \n"
-                    "shrn2  v2.8h, v13.4s, #16          \n"
-                    "shrn   v3.4h, v14.4s, #16          \n"
-                    "shrn2  v3.8h, v15.4s, #16          \n"
-                    "shrn   v4.4h, v16.4s, #16          \n"
-                    "shrn2  v4.8h, v17.4s, #16          \n"
-                    "shrn   v5.4h, v18.4s, #16          \n"
-                    "shrn2  v5.8h, v19.4s, #16          \n"
-                    "shrn   v6.4h, v20.4s, #16          \n"
-                    "shrn2  v6.8h, v21.4s, #16          \n"
-                    "shrn   v7.4h, v22.4s, #16          \n"
-                    "shrn2  v7.8h, v23.4s, #16          \n"
-                    "shrn   v8.4h, v24.4s, #16          \n"
-                    "shrn2  v8.8h, v25.4s, #16          \n"
-                    "shrn   v9.4h, v26.4s, #16          \n"
-                    "shrn2  v9.8h, v27.4s, #16          \n"
-                    "shrn   v10.4h, v28.4s, #16         \n"
-                    "shrn2  v10.8h, v29.4s, #16         \n"
-                    "shrn   v11.4h, v30.4s, #16         \n"
-                    "shrn2  v11.8h, v31.4s, #16         \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    10f                         \n"
-
-                    // if out_elempack == 4
-                    "cmp    %w12, #4                    \n"
-                    "bne    8f                          \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 1          \n"
-                    "st1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%3], #64 \n"
-                    "st1    {v4.8h, v5.8h}, [%3], #32 \n"
-                    "st1    {v6.8h, v7.8h, v8.8h, v9.8h}, [x4], #64 \n"
-                    "st1    {v10.8h, v11.8h}, [x4]      \n"
-                    "b      9f                          \n"
-
-                    // if out_elempack == 1
-                    "8:                                 \n"
-                    // transpose8x12
-                    "uzp1   v20.8h, v0.8h, v1.8h        \n"
-                    "uzp2   v21.8h, v0.8h, v1.8h        \n"
-                    "uzp1   v22.8h, v2.8h, v3.8h        \n"
-                    "uzp2   v23.8h, v2.8h, v3.8h        \n"
-                    "uzp1   v24.8h, v4.8h, v5.8h        \n"
-                    "uzp2   v25.8h, v4.8h, v5.8h        \n"
-                    "uzp1   v26.8h, v6.8h, v7.8h        \n"
-                    "uzp2   v27.8h, v6.8h, v7.8h        \n"
-                    "uzp1   v28.8h, v8.8h, v9.8h        \n"
-                    "uzp2   v29.8h, v8.8h, v9.8h        \n"
-                    "uzp1   v30.8h, v10.8h, v11.8h      \n"
-                    "uzp2   v31.8h, v10.8h, v11.8h      \n"
-
-                    "uzp1   v0.8h, v20.8h, v22.8h       \n"
-                    "uzp2   v6.8h, v20.8h, v22.8h       \n"
-                    "uzp1   v3.8h, v21.8h, v23.8h       \n"
-                    "uzp2   v9.8h, v21.8h, v23.8h       \n"
-                    "mov    v1.d[0], v0.d[1]            \n"
-                    "mov    v7.d[0], v6.d[1]            \n"
-                    "mov    v4.d[0], v3.d[1]            \n"
-                    "mov    v10.d[0], v9.d[1]           \n"
-                    "uzp1   v2.8h, v24.8h, v24.8h       \n"
-                    "uzp2   v8.8h, v24.8h, v24.8h       \n"
-                    "uzp1   v5.8h, v25.8h, v25.8h       \n"
-                    "uzp2   v11.8h, v25.8h, v25.8h      \n"
-
-                    "uzp1   v12.8h, v26.8h, v28.8h      \n"
-                    "uzp2   v18.8h, v26.8h, v28.8h      \n"
-                    "uzp1   v15.8h, v27.8h, v29.8h      \n"
-                    "uzp2   v21.8h, v27.8h, v29.8h      \n"
-                    "mov    v13.d[0], v12.d[1]          \n"
-                    "mov    v19.d[0], v18.d[1]          \n"
-                    "mov    v16.d[0], v15.d[1]          \n"
-                    "mov    v22.d[0], v21.d[1]          \n"
-                    "uzp1   v14.8h, v30.8h, v30.8h      \n"
-                    "uzp2   v20.8h, v30.8h, v30.8h      \n"
-                    "uzp1   v17.8h, v31.8h, v31.8h      \n"
-                    "uzp2   v23.8h, v31.8h, v31.8h      \n"
-
-                    "add    x4, %3, %w13, sxtw 1        \n"
-                    "st1    {v0.4h, v1.4h, v2.4h}, [%3], #24 \n"
-                    "st1    {v3.4h, v4.4h, v5.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v6.4h, v7.4h, v8.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v9.4h, v10.4h, v11.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v12.4h, v13.4h, v14.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v15.4h, v16.4h, v17.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v18.4h, v19.4h, v20.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v21.4h, v22.4h, v23.4h}, [x4] \n"
-
-                    "9:                                 \n"
-                    "add    %0, %0, #384                \n"
-                    "b      11f                         \n"
-
-                    "10:                                \n"
-                    "st1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64   \n"
-                    "st1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0], #64 \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    "11:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum20;
-            float32x4_t _sum21;
-            float32x4_t _sum30;
-            float32x4_t _sum31;
-            float32x4_t _sum40;
-            float32x4_t _sum41;
-            float32x4_t _sum50;
-            float32x4_t _sum51;
-            float32x4_t _sum60;
-            float32x4_t _sum61;
-            float32x4_t _sum70;
-            float32x4_t _sum71;
-            float32x4_t _sum80;
-            float32x4_t _sum81;
-            float32x4_t _sum90;
-            float32x4_t _sum91;
-            float32x4_t _suma0;
-            float32x4_t _suma1;
-            float32x4_t _sumb0;
-            float32x4_t _sumb1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vld1q_f32(pC);
-                    _sum01 = vld1q_f32(pC + 4);
-                    _sum10 = _sum00;
-                    _sum11 = _sum01;
-                    _sum20 = _sum00;
-                    _sum21 = _sum01;
-                    _sum30 = _sum00;
-                    _sum31 = _sum01;
-                    _sum40 = _sum00;
-                    _sum41 = _sum01;
-                    _sum50 = _sum00;
-                    _sum51 = _sum01;
-                    _sum60 = _sum00;
-                    _sum61 = _sum01;
-                    _sum70 = _sum00;
-                    _sum71 = _sum01;
-                    _sum80 = _sum00;
-                    _sum81 = _sum01;
-                    _sum90 = _sum00;
-                    _sum91 = _sum01;
-                    _suma0 = _sum00;
-                    _suma1 = _sum01;
-                    _sumb0 = _sum00;
-                    _sumb1 = _sum01;
-                }
-                else
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                    _sum20 = vdupq_n_f32(0.f);
-                    _sum21 = vdupq_n_f32(0.f);
-                    _sum30 = vdupq_n_f32(0.f);
-                    _sum31 = vdupq_n_f32(0.f);
-                    _sum40 = vdupq_n_f32(0.f);
-                    _sum41 = vdupq_n_f32(0.f);
-                    _sum50 = vdupq_n_f32(0.f);
-                    _sum51 = vdupq_n_f32(0.f);
-                    _sum60 = vdupq_n_f32(0.f);
-                    _sum61 = vdupq_n_f32(0.f);
-                    _sum70 = vdupq_n_f32(0.f);
-                    _sum71 = vdupq_n_f32(0.f);
-                    _sum80 = vdupq_n_f32(0.f);
-                    _sum81 = vdupq_n_f32(0.f);
-                    _sum90 = vdupq_n_f32(0.f);
-                    _sum91 = vdupq_n_f32(0.f);
-                    _suma0 = vdupq_n_f32(0.f);
-                    _suma1 = vdupq_n_f32(0.f);
-                    _sumb0 = vdupq_n_f32(0.f);
-                    _sumb1 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-                _sum20 = vld1q_f32(outptr + 4 * 4);
-                _sum21 = vld1q_f32(outptr + 4 * 5);
-                _sum30 = vld1q_f32(outptr + 4 * 6);
-                _sum31 = vld1q_f32(outptr + 4 * 7);
-                _sum40 = vld1q_f32(outptr + 4 * 8);
-                _sum41 = vld1q_f32(outptr + 4 * 9);
-                _sum50 = vld1q_f32(outptr + 4 * 10);
-                _sum51 = vld1q_f32(outptr + 4 * 11);
-                _sum60 = vld1q_f32(outptr + 4 * 12);
-                _sum61 = vld1q_f32(outptr + 4 * 13);
-                _sum70 = vld1q_f32(outptr + 4 * 14);
-                _sum71 = vld1q_f32(outptr + 4 * 15);
-                _sum80 = vld1q_f32(outptr + 4 * 16);
-                _sum81 = vld1q_f32(outptr + 4 * 17);
-                _sum90 = vld1q_f32(outptr + 4 * 18);
-                _sum91 = vld1q_f32(outptr + 4 * 19);
-                _suma0 = vld1q_f32(outptr + 4 * 20);
-                _suma1 = vld1q_f32(outptr + 4 * 21);
-                _sumb0 = vld1q_f32(outptr + 4 * 22);
-                _sumb1 = vld1q_f32(outptr + 4 * 23);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA0 = bfloat2float(vld1_u16(pA));
-                float32x4_t _pA1 = bfloat2float(vld1_u16(pA + 4));
-
-                float32x4_t _pB0 = bfloat2float(vld1_u16(pB));
-                float32x4_t _pB1 = bfloat2float(vld1_u16(pB + 4));
-                float32x4_t _pB2 = bfloat2float(vld1_u16(pB + 8));
-
-                _sum00 = vfmaq_laneq_f32(_sum00, _pA0, _pB0, 0);
-                _sum01 = vfmaq_laneq_f32(_sum01, _pA1, _pB0, 0);
-                _sum10 = vfmaq_laneq_f32(_sum10, _pA0, _pB0, 1);
-                _sum11 = vfmaq_laneq_f32(_sum11, _pA1, _pB0, 1);
-                _sum20 = vfmaq_laneq_f32(_sum20, _pA0, _pB0, 2);
-                _sum21 = vfmaq_laneq_f32(_sum21, _pA1, _pB0, 2);
-                _sum30 = vfmaq_laneq_f32(_sum30, _pA0, _pB0, 3);
-                _sum31 = vfmaq_laneq_f32(_sum31, _pA1, _pB0, 3);
-                _sum40 = vfmaq_laneq_f32(_sum40, _pA0, _pB1, 0);
-                _sum41 = vfmaq_laneq_f32(_sum41, _pA1, _pB1, 0);
-                _sum50 = vfmaq_laneq_f32(_sum50, _pA0, _pB1, 1);
-                _sum51 = vfmaq_laneq_f32(_sum51, _pA1, _pB1, 1);
-                _sum60 = vfmaq_laneq_f32(_sum60, _pA0, _pB1, 2);
-                _sum61 = vfmaq_laneq_f32(_sum61, _pA1, _pB1, 2);
-                _sum70 = vfmaq_laneq_f32(_sum70, _pA0, _pB1, 3);
-                _sum71 = vfmaq_laneq_f32(_sum71, _pA1, _pB1, 3);
-                _sum80 = vfmaq_laneq_f32(_sum80, _pA0, _pB2, 0);
-                _sum81 = vfmaq_laneq_f32(_sum81, _pA1, _pB2, 0);
-                _sum90 = vfmaq_laneq_f32(_sum90, _pA0, _pB2, 1);
-                _sum91 = vfmaq_laneq_f32(_sum91, _pA1, _pB2, 1);
-                _suma0 = vfmaq_laneq_f32(_suma0, _pA0, _pB2, 2);
-                _suma1 = vfmaq_laneq_f32(_suma1, _pA1, _pB2, 2);
-                _sumb0 = vfmaq_laneq_f32(_sumb0, _pA0, _pB2, 3);
-                _sumb1 = vfmaq_laneq_f32(_sumb1, _pA1, _pB2, 3);
-
-                pA += 8;
-                pB += 12;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum10));
-                    vst1_u16(outptr0 + 4 * 2, float2bfloat(_sum20));
-                    vst1_u16(outptr0 + 4 * 3, float2bfloat(_sum30));
-                    vst1_u16(outptr0 + 4 * 4, float2bfloat(_sum40));
-                    vst1_u16(outptr0 + 4 * 5, float2bfloat(_sum50));
-                    vst1_u16(outptr0 + 4 * 6, float2bfloat(_sum60));
-                    vst1_u16(outptr0 + 4 * 7, float2bfloat(_sum70));
-                    vst1_u16(outptr0 + 4 * 8, float2bfloat(_sum80));
-                    vst1_u16(outptr0 + 4 * 9, float2bfloat(_sum90));
-                    vst1_u16(outptr0 + 4 * 10, float2bfloat(_suma0));
-                    vst1_u16(outptr0 + 4 * 11, float2bfloat(_sumb0));
-
-                    vst1_u16(outptr0 + out_hstep * 4, float2bfloat(_sum01));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, float2bfloat(_sum11));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 2, float2bfloat(_sum21));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 3, float2bfloat(_sum31));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 4, float2bfloat(_sum41));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 5, float2bfloat(_sum51));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 6, float2bfloat(_sum61));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 7, float2bfloat(_sum71));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 8, float2bfloat(_sum81));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 9, float2bfloat(_sum91));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 10, float2bfloat(_suma1));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 11, float2bfloat(_sumb1));
-
-                    outptr0 += 48;
-                }
-                if (out_elempack == 1)
-                {
-                    uint16x8_t _t0 = vcombine_u16(float2bfloat(_sum00), float2bfloat(_sum01));
-                    uint16x8_t _t1 = vcombine_u16(float2bfloat(_sum10), float2bfloat(_sum11));
-                    uint16x8_t _t2 = vcombine_u16(float2bfloat(_sum20), float2bfloat(_sum21));
-                    uint16x8_t _t3 = vcombine_u16(float2bfloat(_sum30), float2bfloat(_sum31));
-                    uint16x8_t _t4 = vcombine_u16(float2bfloat(_sum40), float2bfloat(_sum41));
-                    uint16x8_t _t5 = vcombine_u16(float2bfloat(_sum50), float2bfloat(_sum51));
-                    uint16x8_t _t6 = vcombine_u16(float2bfloat(_sum60), float2bfloat(_sum61));
-                    uint16x8_t _t7 = vcombine_u16(float2bfloat(_sum70), float2bfloat(_sum71));
-                    uint16x8_t _t8 = vcombine_u16(float2bfloat(_sum80), float2bfloat(_sum81));
-                    uint16x8_t _t9 = vcombine_u16(float2bfloat(_sum90), float2bfloat(_sum91));
-                    uint16x8_t _ta = vcombine_u16(float2bfloat(_suma0), float2bfloat(_suma1));
-                    uint16x8_t _tb = vcombine_u16(float2bfloat(_sumb0), float2bfloat(_sumb1));
-                    transpose8x12_u16(_t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _ta, _tb);
-
-                    vst1_u16(outptr0, vget_low_u16(_t0));
-                    vst1_u16(outptr0 + 4, vget_high_u16(_t0));
-                    vst1_u16(outptr0 + 8, vget_low_u16(_t1));
-                    vst1_u16(outptr0 + out_hstep, vget_high_u16(_t1));
-                    vst1_u16(outptr0 + out_hstep + 4, vget_low_u16(_t2));
-                    vst1_u16(outptr0 + out_hstep + 8, vget_high_u16(_t2));
-                    vst1_u16(outptr0 + out_hstep * 2, vget_low_u16(_t3));
-                    vst1_u16(outptr0 + out_hstep * 2 + 4, vget_high_u16(_t3));
-                    vst1_u16(outptr0 + out_hstep * 2 + 8, vget_low_u16(_t4));
-                    vst1_u16(outptr0 + out_hstep * 3, vget_high_u16(_t4));
-                    vst1_u16(outptr0 + out_hstep * 3 + 4, vget_low_u16(_t5));
-                    vst1_u16(outptr0 + out_hstep * 3 + 8, vget_high_u16(_t5));
-                    vst1_u16(outptr0 + out_hstep * 4, vget_low_u16(_t6));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, vget_high_u16(_t6));
-                    vst1_u16(outptr0 + out_hstep * 4 + 8, vget_low_u16(_t7));
-                    vst1_u16(outptr0 + out_hstep * 5, vget_high_u16(_t7));
-                    vst1_u16(outptr0 + out_hstep * 5 + 4, vget_low_u16(_t8));
-                    vst1_u16(outptr0 + out_hstep * 5 + 8, vget_high_u16(_t8));
-                    vst1_u16(outptr0 + out_hstep * 6, vget_low_u16(_t9));
-                    vst1_u16(outptr0 + out_hstep * 6 + 4, vget_high_u16(_t9));
-                    vst1_u16(outptr0 + out_hstep * 6 + 8, vget_low_u16(_ta));
-                    vst1_u16(outptr0 + out_hstep * 7, vget_high_u16(_ta));
-                    vst1_u16(outptr0 + out_hstep * 7 + 4, vget_low_u16(_tb));
-                    vst1_u16(outptr0 + out_hstep * 7 + 8, vget_high_u16(_tb));
-
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                vst1q_f32(outptr + 4 * 4, _sum20);
-                vst1q_f32(outptr + 4 * 5, _sum21);
-                vst1q_f32(outptr + 4 * 6, _sum30);
-                vst1q_f32(outptr + 4 * 7, _sum31);
-                vst1q_f32(outptr + 4 * 8, _sum40);
-                vst1q_f32(outptr + 4 * 9, _sum41);
-                vst1q_f32(outptr + 4 * 10, _sum50);
-                vst1q_f32(outptr + 4 * 11, _sum51);
-                vst1q_f32(outptr + 4 * 12, _sum60);
-                vst1q_f32(outptr + 4 * 13, _sum61);
-                vst1q_f32(outptr + 4 * 14, _sum70);
-                vst1q_f32(outptr + 4 * 15, _sum71);
-                vst1q_f32(outptr + 4 * 16, _sum80);
-                vst1q_f32(outptr + 4 * 17, _sum81);
-                vst1q_f32(outptr + 4 * 18, _sum90);
-                vst1q_f32(outptr + 4 * 19, _sum91);
-                vst1q_f32(outptr + 4 * 20, _suma0);
-                vst1q_f32(outptr + 4 * 21, _suma1);
-                vst1q_f32(outptr + 4 * 22, _sumb0);
-                vst1q_f32(outptr + 4 * 23, _sumb1);
-            }
-
-            outptr += 96;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            const unsigned short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            if (use_a53_a55_optimized_kernel && cpu_support_arm_asimdhp())
-            {
-                // a55
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #192                \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "add    x4, %8, #16                 \n"
-                    "ld1    {v16.4s}, [%8]              \n"
-                    "ld1    {v24.4s}, [x4]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v16.16b, v16.16b, v16.16b   \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v17.16b, v16.16b            \n"
-                    "mov    v18.16b, v16.16b            \n"
-                    "mov    v19.16b, v16.16b            \n"
-                    "mov    v20.16b, v16.16b            \n"
-                    "mov    v21.16b, v16.16b            \n"
-                    "mov    v22.16b, v16.16b            \n"
-                    "mov    v23.16b, v16.16b            \n"
-
-                    "mov    v25.16b, v24.16b            \n"
-                    "mov    v26.16b, v24.16b            \n"
-                    "mov    v27.16b, v24.16b            \n"
-                    "mov    v28.16b, v24.16b            \n"
-                    "mov    v29.16b, v24.16b            \n"
-                    "mov    v30.16b, v24.16b            \n"
-                    "mov    v31.16b, v24.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v8.4h, v9.4h}, [%1], #16   \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h}, [%2], #24 \n"
-
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "shll   v8.4s, v8.4h, #16           \n"
-
-                    ".align 4                           \n"
-                    "4:                                 \n"
-                    "shll   v1.4s, v1.4h, #16           \n"
-                    "fmla   v16.4s, v8.4s, v0.s[0]      \n"
-                    "ldr    d10, [%1], #8               \n"
-                    "fmla   v17.4s, v8.4s, v0.s[1]      \n"
-                    "ldr    d3, [%2], #8                \n"
-                    "fmla   v18.4s, v8.4s, v0.s[2]      \n"
-                    "ldr    d11, [%1], #8               \n"
-                    "fmla   v19.4s, v8.4s, v0.s[3]      \n"
-                    "shll   v9.4s, v9.4h, #16           \n"
-                    "fmla   v20.4s, v8.4s, v1.s[0]      \n"
-                    "ldr    d4, [%2], #8                \n"
-                    "fmla   v21.4s, v8.4s, v1.s[1]      \n"
-                    "ldr    d12, [%1], #8               \n"
-                    "fmla   v22.4s, v8.4s, v1.s[2]      \n"
-                    "ldr    d5, [%2], #8                \n"
-                    "fmla   v23.4s, v8.4s, v1.s[3]      \n"
-                    "shll   v2.4s, v2.4h, #16           \n"
-                    "fmla   v24.4s, v9.4s, v0.s[0]      \n"
-                    "ldr    d13, [%1], #8               \n"
-                    "fmla   v25.4s, v9.4s, v0.s[1]      \n"
-                    "ldr    d6, [%2], #8                \n"
-                    "fmla   v26.4s, v9.4s, v0.s[2]      \n"
-                    "ldr    d14, [%1], #8               \n"
-                    "fmla   v27.4s, v9.4s, v0.s[3]      \n"
-                    "shll   v10.4s, v10.4h, #16         \n"
-                    "fmla   v28.4s, v9.4s, v1.s[0]      \n"
-                    "ldr    d7, [%2], #8                \n"
-                    "fmla   v29.4s, v9.4s, v1.s[1]      \n"
-                    "ldr    d15, [%1], #8               \n"
-                    "fmla   v30.4s, v9.4s, v1.s[2]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                    "fmla   v31.4s, v9.4s, v1.s[3]      \n"
-                    "shll   v3.4s, v3.4h, #16           \n"
-                    "fmla   v16.4s, v10.4s, v2.s[0]     \n"
-                    "ldr    d8, [%1], #8                \n"
-                    "fmla   v17.4s, v10.4s, v2.s[1]     \n"
-                    "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                    "fmla   v18.4s, v10.4s, v2.s[2]     \n"
-                    "ldr    d0, [%2], #8                \n"
-                    "fmla   v19.4s, v10.4s, v2.s[3]     \n"
-                    "shll   v11.4s, v11.4h, #16         \n"
-                    "fmla   v20.4s, v10.4s, v3.s[0]     \n"
-                    "ldr    d1, [%2], #8                \n"
-                    "fmla   v21.4s, v10.4s, v3.s[1]     \n"
-                    "ldr    d9, [%1], #8                \n"
-                    "fmla   v22.4s, v10.4s, v3.s[2]     \n"
-                    "fmla   v23.4s, v10.4s, v3.s[3]     \n"
-                    "shll   v4.4s, v4.4h, #16           \n"
-                    "fmla   v24.4s, v11.4s, v2.s[0]     \n"
-                    "fmla   v25.4s, v11.4s, v2.s[1]     \n"
-                    "fmla   v26.4s, v11.4s, v2.s[2]     \n"
-                    "fmla   v27.4s, v11.4s, v2.s[3]     \n"
-                    "shll   v12.4s, v12.4h, #16         \n"
-                    "fmla   v28.4s, v11.4s, v3.s[0]     \n"
-                    "ldr    d2, [%2], #8                \n"
-                    "fmla   v29.4s, v11.4s, v3.s[1]     \n"
-                    "fmla   v30.4s, v11.4s, v3.s[2]     \n"
-                    "fmla   v31.4s, v11.4s, v3.s[3]     \n"
-                    "shll   v5.4s, v5.4h, #16           \n"
-                    "fmla   v16.4s, v12.4s, v4.s[0]     \n"
-                    "fmla   v17.4s, v12.4s, v4.s[1]     \n"
-                    "fmla   v18.4s, v12.4s, v4.s[2]     \n"
-                    "fmla   v19.4s, v12.4s, v4.s[3]     \n"
-                    "shll   v13.4s, v13.4h, #16         \n"
-                    "fmla   v20.4s, v12.4s, v5.s[0]     \n"
-                    "fmla   v21.4s, v12.4s, v5.s[1]     \n"
-                    "fmla   v22.4s, v12.4s, v5.s[2]     \n"
-                    "fmla   v23.4s, v12.4s, v5.s[3]     \n"
-                    "shll   v6.4s, v6.4h, #16           \n"
-                    "fmla   v24.4s, v13.4s, v4.s[0]     \n"
-                    "fmla   v25.4s, v13.4s, v4.s[1]     \n"
-                    "fmla   v26.4s, v13.4s, v4.s[2]     \n"
-                    "fmla   v27.4s, v13.4s, v4.s[3]     \n"
-                    "shll   v14.4s, v14.4h, #16         \n"
-                    "fmla   v28.4s, v13.4s, v5.s[0]     \n"
-                    "fmla   v29.4s, v13.4s, v5.s[1]     \n"
-                    "fmla   v30.4s, v13.4s, v5.s[2]     \n"
-                    "fmla   v31.4s, v13.4s, v5.s[3]     \n"
-                    "shll   v7.4s, v7.4h, #16           \n"
-                    "fmla   v16.4s, v14.4s, v6.s[0]     \n"
-                    "fmla   v17.4s, v14.4s, v6.s[1]     \n"
-                    "fmla   v18.4s, v14.4s, v6.s[2]     \n"
-                    "fmla   v19.4s, v14.4s, v6.s[3]     \n"
-                    "shll   v15.4s, v15.4h, #16         \n"
-                    "fmla   v20.4s, v14.4s, v7.s[0]     \n"
-                    "fmla   v21.4s, v14.4s, v7.s[1]     \n"
-                    "fmla   v22.4s, v14.4s, v7.s[2]     \n"
-                    "fmla   v23.4s, v14.4s, v7.s[3]     \n"
-                    "shll   v8.4s, v8.4h, #16           \n"
-                    "fmla   v24.4s, v15.4s, v6.s[0]     \n"
-                    "fmla   v25.4s, v15.4s, v6.s[1]     \n"
-                    "fmla   v26.4s, v15.4s, v6.s[2]     \n"
-                    "fmla   v27.4s, v15.4s, v6.s[3]     \n"
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "fmla   v28.4s, v15.4s, v7.s[0]     \n"
-                    "fmla   v29.4s, v15.4s, v7.s[1]     \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v30.4s, v15.4s, v7.s[2]     \n"
-                    "fmla   v31.4s, v15.4s, v7.s[3]     \n"
-                    "bne    4b                          \n"
-
-                    "sub    %1, %1, #16                 \n"
-                    "sub    %2, %2, #24                 \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v2.8h}, [%2], #16          \n"
-                    "shll   v0.4s, v2.4h, #16           \n"
-                    "shll2  v1.4s, v2.8h, #16           \n"
-                    "ld1    {v3.8h}, [%1], #16          \n"
-                    "shll   v4.4s, v3.4h, #16           \n"
-                    "shll2  v5.4s, v3.8h, #16           \n"
-
-                    "fmla   v16.4s, v4.4s, v0.s[0]      \n"
-                    "fmla   v17.4s, v4.4s, v0.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v20.4s, v4.4s, v1.s[0]      \n"
-                    "fmla   v21.4s, v4.4s, v1.s[1]      \n"
-                    "fmla   v22.4s, v4.4s, v1.s[2]      \n"
-                    "fmla   v23.4s, v4.4s, v1.s[3]      \n"
-
-                    "subs   w4, w4, #1                  \n"
-
-                    "fmla   v24.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v0.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v1.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v1.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v1.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v1.s[3]      \n"
-
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "shrn   v0.4h, v16.4s, #16          \n"
-                    "shrn2  v0.8h, v17.4s, #16          \n"
-                    "shrn   v1.4h, v18.4s, #16          \n"
-                    "shrn2  v1.8h, v19.4s, #16          \n"
-                    "shrn   v2.4h, v20.4s, #16          \n"
-                    "shrn2  v2.8h, v21.4s, #16          \n"
-                    "shrn   v3.4h, v22.4s, #16          \n"
-                    "shrn2  v3.8h, v23.4s, #16          \n"
-                    "shrn   v4.4h, v24.4s, #16          \n"
-                    "shrn2  v4.8h, v25.4s, #16          \n"
-                    "shrn   v5.4h, v26.4s, #16          \n"
-                    "shrn2  v5.8h, v27.4s, #16          \n"
-                    "shrn   v6.4h, v28.4s, #16          \n"
-                    "shrn2  v6.8h, v29.4s, #16          \n"
-                    "shrn   v7.4h, v30.4s, #16          \n"
-                    "shrn2  v7.8h, v31.4s, #16          \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    10f                         \n"
-
-                    // if out_elempack == 4
-                    "cmp    %w12, #4                    \n"
-                    "bne    8f                          \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 1          \n"
-                    "st1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%3], #64 \n"
-                    "st1    {v4.8h, v5.8h, v6.8h, v7.8h}, [x4] \n"
-                    "b      9f                          \n"
-
-                    // if out_elempack == 1
-                    "8:                                 \n"
-                    // transpose8x8
-                    "uzp1   v24.8h, v0.8h, v1.8h        \n"
-                    "uzp2   v25.8h, v0.8h, v1.8h        \n"
-                    "uzp1   v26.8h, v2.8h, v3.8h        \n"
-                    "uzp2   v27.8h, v2.8h, v3.8h        \n"
-                    "uzp1   v28.8h, v4.8h, v5.8h        \n"
-                    "uzp2   v29.8h, v4.8h, v5.8h        \n"
-                    "uzp1   v30.8h, v6.8h, v7.8h        \n"
-                    "uzp2   v31.8h, v6.8h, v7.8h        \n"
-
-                    "uzp1   v0.8h, v24.8h, v26.8h       \n"
-                    "uzp2   v2.8h, v24.8h, v26.8h       \n"
-                    "uzp1   v1.8h, v25.8h, v27.8h       \n"
-                    "uzp2   v3.8h, v25.8h, v27.8h       \n"
-
-                    "uzp1   v4.8h, v28.8h, v30.8h       \n"
-                    "uzp2   v6.8h, v28.8h, v30.8h       \n"
-                    "uzp1   v5.8h, v29.8h, v31.8h       \n"
-                    "uzp2   v7.8h, v29.8h, v31.8h       \n"
-
-                    "add    x4, %3, %w13, sxtw 1        \n"
-                    "st1    {v0.8h}, [%3], #16          \n"
-                    "st1    {v1.8h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v2.8h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v3.8h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v4.8h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v5.8h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v6.8h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v7.8h}, [x4]               \n"
-
-                    "9:                                 \n"
-                    "add    %0, %0, #256                \n"
-                    "b      11f                         \n"
-
-                    "10:                                \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    "11:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            else
-            {
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #192                \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "add    x4, %8, #16                 \n"
-                    "ld1    {v16.4s}, [%8]              \n"
-                    "ld1    {v24.4s}, [x4]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v16.16b, v16.16b, v16.16b   \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v17.16b, v16.16b            \n"
-                    "mov    v18.16b, v16.16b            \n"
-                    "mov    v19.16b, v16.16b            \n"
-                    "mov    v20.16b, v16.16b            \n"
-                    "mov    v21.16b, v16.16b            \n"
-                    "mov    v22.16b, v16.16b            \n"
-                    "mov    v23.16b, v16.16b            \n"
-
-                    "mov    v25.16b, v24.16b            \n"
-                    "mov    v26.16b, v24.16b            \n"
-                    "mov    v27.16b, v24.16b            \n"
-                    "mov    v28.16b, v24.16b            \n"
-                    "mov    v29.16b, v24.16b            \n"
-                    "mov    v30.16b, v24.16b            \n"
-                    "mov    v31.16b, v24.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%2], #64 \n"
-                    "shll   v0.4s, v4.4h, #16           \n"
-                    "shll2  v1.4s, v4.8h, #16           \n"
-                    "shll   v2.4s, v5.4h, #16           \n"
-                    "shll2  v3.4s, v5.8h, #16           \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%1], #64 \n"
-                    "shll   v8.4s, v12.4h, #16          \n"
-                    "shll2  v9.4s, v12.8h, #16          \n"
-                    "shll   v10.4s, v13.4h, #16         \n"
-                    "shll2  v11.4s, v13.8h, #16         \n"
-                    "fmla   v16.4s, v8.4s, v0.s[0]      \n"
-                    "fmla   v17.4s, v8.4s, v0.s[1]      \n"
-                    "fmla   v18.4s, v8.4s, v0.s[2]      \n"
-                    "fmla   v19.4s, v8.4s, v0.s[3]      \n"
-                    "fmla   v20.4s, v8.4s, v1.s[0]      \n"
-                    "fmla   v21.4s, v8.4s, v1.s[1]      \n"
-                    "fmla   v22.4s, v8.4s, v1.s[2]      \n"
-                    "fmla   v23.4s, v8.4s, v1.s[3]      \n"
-                    "fmla   v24.4s, v9.4s, v0.s[0]      \n"
-                    "fmla   v25.4s, v9.4s, v0.s[1]      \n"
-                    "fmla   v26.4s, v9.4s, v0.s[2]      \n"
-                    "fmla   v27.4s, v9.4s, v0.s[3]      \n"
-                    "fmla   v28.4s, v9.4s, v1.s[0]      \n"
-                    "fmla   v29.4s, v9.4s, v1.s[1]      \n"
-                    "fmla   v30.4s, v9.4s, v1.s[2]      \n"
-                    "fmla   v31.4s, v9.4s, v1.s[3]      \n"
-                    "fmla   v16.4s, v10.4s, v2.s[0]     \n"
-                    "fmla   v17.4s, v10.4s, v2.s[1]     \n"
-                    "fmla   v18.4s, v10.4s, v2.s[2]     \n"
-                    "fmla   v19.4s, v10.4s, v2.s[3]     \n"
-                    "fmla   v20.4s, v10.4s, v3.s[0]     \n"
-                    "fmla   v21.4s, v10.4s, v3.s[1]     \n"
-                    "fmla   v22.4s, v10.4s, v3.s[2]     \n"
-                    "fmla   v23.4s, v10.4s, v3.s[3]     \n"
-                    "fmla   v24.4s, v11.4s, v2.s[0]     \n"
-                    "fmla   v25.4s, v11.4s, v2.s[1]     \n"
-                    "fmla   v26.4s, v11.4s, v2.s[2]     \n"
-                    "fmla   v27.4s, v11.4s, v2.s[3]     \n"
-                    "fmla   v28.4s, v11.4s, v3.s[0]     \n"
-                    "fmla   v29.4s, v11.4s, v3.s[1]     \n"
-                    "fmla   v30.4s, v11.4s, v3.s[2]     \n"
-                    "fmla   v31.4s, v11.4s, v3.s[3]     \n"
-                    "shll   v4.4s, v6.4h, #16           \n"
-                    "shll2  v5.4s, v6.8h, #16           \n"
-                    "shll   v6.4s, v7.4h, #16           \n"
-                    "shll2  v7.4s, v7.8h, #16           \n"
-                    "shll   v12.4s, v14.4h, #16         \n"
-                    "shll2  v13.4s, v14.8h, #16         \n"
-                    "shll   v14.4s, v15.4h, #16         \n"
-                    "shll2  v15.4s, v15.8h, #16         \n"
-                    "fmla   v16.4s, v12.4s, v4.s[0]     \n"
-                    "fmla   v17.4s, v12.4s, v4.s[1]     \n"
-                    "fmla   v18.4s, v12.4s, v4.s[2]     \n"
-                    "fmla   v19.4s, v12.4s, v4.s[3]     \n"
-                    "fmla   v20.4s, v12.4s, v5.s[0]     \n"
-                    "fmla   v21.4s, v12.4s, v5.s[1]     \n"
-                    "fmla   v22.4s, v12.4s, v5.s[2]     \n"
-                    "fmla   v23.4s, v12.4s, v5.s[3]     \n"
-                    "fmla   v24.4s, v13.4s, v4.s[0]     \n"
-                    "fmla   v25.4s, v13.4s, v4.s[1]     \n"
-                    "fmla   v26.4s, v13.4s, v4.s[2]     \n"
-                    "fmla   v27.4s, v13.4s, v4.s[3]     \n"
-                    "fmla   v28.4s, v13.4s, v5.s[0]     \n"
-                    "fmla   v29.4s, v13.4s, v5.s[1]     \n"
-                    "fmla   v30.4s, v13.4s, v5.s[2]     \n"
-                    "fmla   v31.4s, v13.4s, v5.s[3]     \n"
-                    "fmla   v16.4s, v14.4s, v6.s[0]     \n"
-                    "fmla   v17.4s, v14.4s, v6.s[1]     \n"
-                    "fmla   v18.4s, v14.4s, v6.s[2]     \n"
-                    "fmla   v19.4s, v14.4s, v6.s[3]     \n"
-                    "fmla   v20.4s, v14.4s, v7.s[0]     \n"
-                    "fmla   v21.4s, v14.4s, v7.s[1]     \n"
-                    "fmla   v22.4s, v14.4s, v7.s[2]     \n"
-                    "fmla   v23.4s, v14.4s, v7.s[3]     \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v24.4s, v15.4s, v6.s[0]     \n"
-                    "fmla   v25.4s, v15.4s, v6.s[1]     \n"
-                    "fmla   v26.4s, v15.4s, v6.s[2]     \n"
-                    "fmla   v27.4s, v15.4s, v6.s[3]     \n"
-                    "fmla   v28.4s, v15.4s, v7.s[0]     \n"
-                    "fmla   v29.4s, v15.4s, v7.s[1]     \n"
-                    "fmla   v30.4s, v15.4s, v7.s[2]     \n"
-                    "fmla   v31.4s, v15.4s, v7.s[3]     \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v2.8h}, [%2], #16          \n"
-                    "shll   v0.4s, v2.4h, #16           \n"
-                    "shll2  v1.4s, v2.8h, #16           \n"
-                    "ld1    {v3.8h}, [%1], #16          \n"
-                    "shll   v4.4s, v3.4h, #16           \n"
-                    "shll2  v5.4s, v3.8h, #16           \n"
-
-                    "fmla   v16.4s, v4.4s, v0.s[0]      \n"
-                    "fmla   v17.4s, v4.4s, v0.s[1]      \n"
-                    "fmla   v18.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v19.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v20.4s, v4.4s, v1.s[0]      \n"
-                    "fmla   v21.4s, v4.4s, v1.s[1]      \n"
-                    "fmla   v22.4s, v4.4s, v1.s[2]      \n"
-                    "fmla   v23.4s, v4.4s, v1.s[3]      \n"
-
-                    "subs   w4, w4, #1                  \n"
-
-                    "fmla   v24.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v25.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v26.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v27.4s, v5.4s, v0.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v1.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v1.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v1.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v1.s[3]      \n"
-
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "shrn   v0.4h, v16.4s, #16          \n"
-                    "shrn2  v0.8h, v17.4s, #16          \n"
-                    "shrn   v1.4h, v18.4s, #16          \n"
-                    "shrn2  v1.8h, v19.4s, #16          \n"
-                    "shrn   v2.4h, v20.4s, #16          \n"
-                    "shrn2  v2.8h, v21.4s, #16          \n"
-                    "shrn   v3.4h, v22.4s, #16          \n"
-                    "shrn2  v3.8h, v23.4s, #16          \n"
-                    "shrn   v4.4h, v24.4s, #16          \n"
-                    "shrn2  v4.8h, v25.4s, #16          \n"
-                    "shrn   v5.4h, v26.4s, #16          \n"
-                    "shrn2  v5.8h, v27.4s, #16          \n"
-                    "shrn   v6.4h, v28.4s, #16          \n"
-                    "shrn2  v6.8h, v29.4s, #16          \n"
-                    "shrn   v7.4h, v30.4s, #16          \n"
-                    "shrn2  v7.8h, v31.4s, #16          \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    10f                         \n"
-
-                    // if out_elempack == 4
-                    "cmp    %w12, #4                    \n"
-                    "bne    8f                          \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 1          \n"
-                    "st1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%3], #64 \n"
-                    "st1    {v4.8h, v5.8h, v6.8h, v7.8h}, [x4] \n"
-                    "b      9f                          \n"
-
-                    // if out_elempack == 1
-                    "8:                                 \n"
-                    // transpose8x8
-                    "uzp1   v24.8h, v0.8h, v1.8h        \n"
-                    "uzp2   v25.8h, v0.8h, v1.8h        \n"
-                    "uzp1   v26.8h, v2.8h, v3.8h        \n"
-                    "uzp2   v27.8h, v2.8h, v3.8h        \n"
-                    "uzp1   v28.8h, v4.8h, v5.8h        \n"
-                    "uzp2   v29.8h, v4.8h, v5.8h        \n"
-                    "uzp1   v30.8h, v6.8h, v7.8h        \n"
-                    "uzp2   v31.8h, v6.8h, v7.8h        \n"
-
-                    "uzp1   v0.8h, v24.8h, v26.8h       \n"
-                    "uzp2   v2.8h, v24.8h, v26.8h       \n"
-                    "uzp1   v1.8h, v25.8h, v27.8h       \n"
-                    "uzp2   v3.8h, v25.8h, v27.8h       \n"
-
-                    "uzp1   v4.8h, v28.8h, v30.8h       \n"
-                    "uzp2   v6.8h, v28.8h, v30.8h       \n"
-                    "uzp1   v5.8h, v29.8h, v31.8h       \n"
-                    "uzp2   v7.8h, v29.8h, v31.8h       \n"
-
-                    "add    x4, %3, %w13, sxtw 1        \n"
-                    "st1    {v0.8h}, [%3], #16          \n"
-                    "st1    {v1.8h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v2.8h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v3.8h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v4.8h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v5.8h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v6.8h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v7.8h}, [x4]               \n"
-
-                    "9:                                 \n"
-                    "add    %0, %0, #256                \n"
-                    "b      11f                         \n"
-
-                    "10:                                \n"
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    "11:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum20;
-            float32x4_t _sum21;
-            float32x4_t _sum30;
-            float32x4_t _sum31;
-            float32x4_t _sum40;
-            float32x4_t _sum41;
-            float32x4_t _sum50;
-            float32x4_t _sum51;
-            float32x4_t _sum60;
-            float32x4_t _sum61;
-            float32x4_t _sum70;
-            float32x4_t _sum71;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vld1q_f32(pC);
-                    _sum01 = vld1q_f32(pC + 4);
-                    _sum10 = _sum00;
-                    _sum11 = _sum01;
-                    _sum20 = _sum00;
-                    _sum21 = _sum01;
-                    _sum30 = _sum00;
-                    _sum31 = _sum01;
-                    _sum40 = _sum00;
-                    _sum41 = _sum01;
-                    _sum50 = _sum00;
-                    _sum51 = _sum01;
-                    _sum60 = _sum00;
-                    _sum61 = _sum01;
-                    _sum70 = _sum00;
-                    _sum71 = _sum01;
-                }
-                else
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                    _sum20 = vdupq_n_f32(0.f);
-                    _sum21 = vdupq_n_f32(0.f);
-                    _sum30 = vdupq_n_f32(0.f);
-                    _sum31 = vdupq_n_f32(0.f);
-                    _sum40 = vdupq_n_f32(0.f);
-                    _sum41 = vdupq_n_f32(0.f);
-                    _sum50 = vdupq_n_f32(0.f);
-                    _sum51 = vdupq_n_f32(0.f);
-                    _sum60 = vdupq_n_f32(0.f);
-                    _sum61 = vdupq_n_f32(0.f);
-                    _sum70 = vdupq_n_f32(0.f);
-                    _sum71 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-                _sum20 = vld1q_f32(outptr + 4 * 4);
-                _sum21 = vld1q_f32(outptr + 4 * 5);
-                _sum30 = vld1q_f32(outptr + 4 * 6);
-                _sum31 = vld1q_f32(outptr + 4 * 7);
-                _sum40 = vld1q_f32(outptr + 4 * 8);
-                _sum41 = vld1q_f32(outptr + 4 * 9);
-                _sum50 = vld1q_f32(outptr + 4 * 10);
-                _sum51 = vld1q_f32(outptr + 4 * 11);
-                _sum60 = vld1q_f32(outptr + 4 * 12);
-                _sum61 = vld1q_f32(outptr + 4 * 13);
-                _sum70 = vld1q_f32(outptr + 4 * 14);
-                _sum71 = vld1q_f32(outptr + 4 * 15);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA0 = bfloat2float(vld1_u16(pA));
-                float32x4_t _pA1 = bfloat2float(vld1_u16(pA + 4));
-
-                float32x4_t _pB0 = bfloat2float(vld1_u16(pB));
-                float32x4_t _pB1 = bfloat2float(vld1_u16(pB + 4));
-
-                _sum00 = vfmaq_laneq_f32(_sum00, _pA0, _pB0, 0);
-                _sum01 = vfmaq_laneq_f32(_sum01, _pA1, _pB0, 0);
-                _sum10 = vfmaq_laneq_f32(_sum10, _pA0, _pB0, 1);
-                _sum11 = vfmaq_laneq_f32(_sum11, _pA1, _pB0, 1);
-                _sum20 = vfmaq_laneq_f32(_sum20, _pA0, _pB0, 2);
-                _sum21 = vfmaq_laneq_f32(_sum21, _pA1, _pB0, 2);
-                _sum30 = vfmaq_laneq_f32(_sum30, _pA0, _pB0, 3);
-                _sum31 = vfmaq_laneq_f32(_sum31, _pA1, _pB0, 3);
-                _sum40 = vfmaq_laneq_f32(_sum40, _pA0, _pB1, 0);
-                _sum41 = vfmaq_laneq_f32(_sum41, _pA1, _pB1, 0);
-                _sum50 = vfmaq_laneq_f32(_sum50, _pA0, _pB1, 1);
-                _sum51 = vfmaq_laneq_f32(_sum51, _pA1, _pB1, 1);
-                _sum60 = vfmaq_laneq_f32(_sum60, _pA0, _pB1, 2);
-                _sum61 = vfmaq_laneq_f32(_sum61, _pA1, _pB1, 2);
-                _sum70 = vfmaq_laneq_f32(_sum70, _pA0, _pB1, 3);
-                _sum71 = vfmaq_laneq_f32(_sum71, _pA1, _pB1, 3);
-
-                pA += 8;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum10));
-                    vst1_u16(outptr0 + 4 * 2, float2bfloat(_sum20));
-                    vst1_u16(outptr0 + 4 * 3, float2bfloat(_sum30));
-                    vst1_u16(outptr0 + 4 * 4, float2bfloat(_sum40));
-                    vst1_u16(outptr0 + 4 * 5, float2bfloat(_sum50));
-                    vst1_u16(outptr0 + 4 * 6, float2bfloat(_sum60));
-                    vst1_u16(outptr0 + 4 * 7, float2bfloat(_sum70));
-
-                    vst1_u16(outptr0 + out_hstep * 4, float2bfloat(_sum01));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, float2bfloat(_sum11));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 2, float2bfloat(_sum21));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 3, float2bfloat(_sum31));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 4, float2bfloat(_sum41));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 5, float2bfloat(_sum51));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 6, float2bfloat(_sum61));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 7, float2bfloat(_sum71));
-
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-                    uint16x8_t _t0 = vcombine_u16(float2bfloat(_sum00), float2bfloat(_sum01));
-                    uint16x8_t _t1 = vcombine_u16(float2bfloat(_sum10), float2bfloat(_sum11));
-                    uint16x8_t _t2 = vcombine_u16(float2bfloat(_sum20), float2bfloat(_sum21));
-                    uint16x8_t _t3 = vcombine_u16(float2bfloat(_sum30), float2bfloat(_sum31));
-                    uint16x8_t _t4 = vcombine_u16(float2bfloat(_sum40), float2bfloat(_sum41));
-                    uint16x8_t _t5 = vcombine_u16(float2bfloat(_sum50), float2bfloat(_sum51));
-                    uint16x8_t _t6 = vcombine_u16(float2bfloat(_sum60), float2bfloat(_sum61));
-                    uint16x8_t _t7 = vcombine_u16(float2bfloat(_sum70), float2bfloat(_sum71));
-                    transpose8x8_u16(_t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7);
-
-                    vst1q_u16(outptr0, _t0);
-                    vst1q_u16(outptr0 + out_hstep, _t1);
-                    vst1q_u16(outptr0 + out_hstep * 2, _t2);
-                    vst1q_u16(outptr0 + out_hstep * 3, _t3);
-                    vst1q_u16(outptr0 + out_hstep * 4, _t4);
-                    vst1q_u16(outptr0 + out_hstep * 5, _t5);
-                    vst1q_u16(outptr0 + out_hstep * 6, _t6);
-                    vst1q_u16(outptr0 + out_hstep * 7, _t7);
-
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                vst1q_f32(outptr + 4 * 4, _sum20);
-                vst1q_f32(outptr + 4 * 5, _sum21);
-                vst1q_f32(outptr + 4 * 6, _sum30);
-                vst1q_f32(outptr + 4 * 7, _sum31);
-                vst1q_f32(outptr + 4 * 8, _sum40);
-                vst1q_f32(outptr + 4 * 9, _sum41);
-                vst1q_f32(outptr + 4 * 10, _sum50);
-                vst1q_f32(outptr + 4 * 11, _sum51);
-                vst1q_f32(outptr + 4 * 12, _sum60);
-                vst1q_f32(outptr + 4 * 13, _sum61);
-                vst1q_f32(outptr + 4 * 14, _sum70);
-                vst1q_f32(outptr + 4 * 15, _sum71);
-            }
-
-            outptr += 64;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            const unsigned short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            if (use_a53_a55_optimized_kernel && cpu_support_arm_asimdhp())
-            {
-                // a55
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #64                 \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "add    x4, %8, #16                 \n"
-                    "ld1    {v24.4s}, [%8]              \n"
-                    "ld1    {v28.4s}, [x4]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v25.16b, v24.16b            \n"
-                    "mov    v26.16b, v24.16b            \n"
-                    "mov    v27.16b, v24.16b            \n"
-
-                    "mov    v29.16b, v28.16b            \n"
-                    "mov    v30.16b, v28.16b            \n"
-                    "mov    v31.16b, v28.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.4h, v5.4h, v6.4h}, [%1], #24 \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.4h, v1.4h}, [%2], #16   \n"
-
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "shll   v4.4s, v4.4h, #16           \n"
-
-                    ".align 4                           \n"
-                    "4:                                 \n"
-                    "shll   v5.4s, v5.4h, #16           \n"
-                    "fmla   v24.4s, v4.4s, v0.s[0]      \n"
-                    "ldr    d7, [%1], #8                \n"
-                    "fmla   v25.4s, v4.4s, v0.s[1]      \n"
-                    "shll   v1.4s, v1.4h, #16           \n"
-                    "fmla   v26.4s, v4.4s, v0.s[2]      \n"
-                    "ldr    d2, [%2], #8                \n"
-                    "fmla   v27.4s, v4.4s, v0.s[3]      \n"
-                    "shll   v6.4s, v6.4h, #16           \n"
-                    "fmla   v28.4s, v5.4s, v0.s[0]      \n"
-                    "ldr    d8, [%1], #8                \n"
-                    "fmla   v29.4s, v5.4s, v0.s[1]      \n"
-                    "ldr    d9, [%1], #8                \n"
-                    "fmla   v30.4s, v5.4s, v0.s[2]      \n"
-                    "ldr    d3, [%2], #8                \n"
-                    "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-                    "shll   v7.4s, v7.4h, #16           \n"
-                    "fmla   v24.4s, v6.4s, v1.s[0]      \n"
-                    "ldr    d10, [%1], #8               \n"
-                    "fmla   v25.4s, v6.4s, v1.s[1]      \n"
-                    "shll   v2.4s, v2.4h, #16           \n"
-                    "fmla   v26.4s, v6.4s, v1.s[2]      \n"
-                    "ldr    d11, [%1], #8               \n"
-                    "fmla   v27.4s, v6.4s, v1.s[3]      \n"
-                    "shll   v8.4s, v8.4h, #16           \n"
-                    "fmla   v28.4s, v7.4s, v1.s[0]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                    "fmla   v29.4s, v7.4s, v1.s[1]      \n"
-                    "ldr    d4, [%1], #8                \n"
-                    "fmla   v30.4s, v7.4s, v1.s[2]      \n"
-                    "prfm   pldl1keep, [%2, #256]       \n" // NOTE PRELOAD
-                    "fmla   v31.4s, v7.4s, v1.s[3]      \n"
-                    "shll   v9.4s, v9.4h, #16           \n"
-                    "fmla   v24.4s, v8.4s, v2.s[0]      \n"
-                    "ldr    d0, [%2], #8                \n"
-                    "fmla   v25.4s, v8.4s, v2.s[1]      \n"
-                    "shll   v3.4s, v3.4h, #16           \n"
-                    "fmla   v26.4s, v8.4s, v2.s[2]      \n"
-                    "ldr    d5, [%1], #8                \n"
-                    "fmla   v27.4s, v8.4s, v2.s[3]      \n"
-                    "shll   v10.4s, v10.4h, #16         \n"
-                    "fmla   v28.4s, v9.4s, v2.s[0]      \n"
-                    "ldr    d1, [%2], #8                \n"
-                    "fmla   v29.4s, v9.4s, v2.s[1]      \n"
-                    "ldr    d6, [%1], #8                \n"
-                    "fmla   v30.4s, v9.4s, v2.s[2]      \n"
-                    "fmla   v31.4s, v9.4s, v2.s[3]      \n"
-                    "shll   v11.4s, v11.4h, #16         \n"
-                    "fmla   v24.4s, v10.4s, v3.s[0]     \n"
-                    "fmla   v25.4s, v10.4s, v3.s[1]     \n"
-                    "shll   v4.4s, v4.4h, #16           \n"
-                    "fmla   v26.4s, v10.4s, v3.s[2]     \n"
-                    "fmla   v27.4s, v10.4s, v3.s[3]     \n"
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "fmla   v28.4s, v11.4s, v3.s[0]     \n"
-                    "fmla   v29.4s, v11.4s, v3.s[1]     \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v30.4s, v11.4s, v3.s[2]     \n"
-                    "fmla   v31.4s, v11.4s, v3.s[3]     \n"
-                    "bne    4b                          \n"
-
-                    "sub    %1, %1, #24                 \n"
-                    "sub    %2, %2, #16                 \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v0.4h}, [%2], #8           \n"
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "ld1    {v3.8h}, [%1], #16          \n"
-                    "shll   v4.4s, v3.4h, #16           \n"
-                    "shll2  v5.4s, v3.8h, #16           \n"
-                    "fmla   v24.4s, v4.4s, v0.s[0]      \n"
-                    "fmla   v25.4s, v4.4s, v0.s[1]      \n"
-                    "fmla   v26.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v27.4s, v4.4s, v0.s[3]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v28.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "shrn   v0.4h, v24.4s, #16          \n"
-                    "shrn2  v0.8h, v25.4s, #16          \n"
-                    "shrn   v1.4h, v26.4s, #16          \n"
-                    "shrn2  v1.8h, v27.4s, #16          \n"
-                    "shrn   v2.4h, v28.4s, #16          \n"
-                    "shrn2  v2.8h, v29.4s, #16          \n"
-                    "shrn   v3.4h, v30.4s, #16          \n"
-                    "shrn2  v3.8h, v31.4s, #16          \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    10f                         \n"
-
-                    // if out_elempack == 4
-                    "cmp    %w12, #4                    \n"
-                    "bne    8f                          \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 1          \n"
-                    "st1    {v0.8h, v1.8h}, [%3], #32   \n"
-                    "st1    {v2.8h, v3.8h}, [x4]        \n"
-                    "b      9f                          \n"
-
-                    // if out_elempack == 1
-                    "8:                                 \n"
-                    // transpose8x4
-                    "uzp1   v28.8h, v0.8h, v1.8h        \n"
-                    "uzp2   v29.8h, v0.8h, v1.8h        \n"
-                    "uzp1   v30.8h, v2.8h, v3.8h        \n"
-                    "uzp2   v31.8h, v2.8h, v3.8h        \n"
-
-                    "uzp1   v0.8h, v28.8h, v29.8h       \n"
-                    "uzp2   v2.8h, v28.8h, v29.8h       \n"
-                    "uzp1   v4.8h, v30.8h, v31.8h       \n"
-                    "uzp2   v6.8h, v30.8h, v31.8h       \n"
-
-                    "mov    v1.d[0], v0.d[1]            \n"
-                    "mov    v3.d[0], v2.d[1]            \n"
-                    "mov    v5.d[0], v4.d[1]            \n"
-                    "mov    v7.d[0], v6.d[1]            \n"
-
-                    "add    x4, %3, %w13, sxtw 1        \n"
-                    "st1    {v0.4h}, [%3], #8           \n"
-                    "st1    {v1.4h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v2.4h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v3.4h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v4.4h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v5.4h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v6.4h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v7.4h}, [x4]               \n"
-
-                    "9:                                 \n"
-                    "add    %0, %0, #128                \n"
-                    "b      11f                         \n"
-
-                    "10:                                \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    "11:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            else
-            {
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                    "subs   %0, %0, #64                 \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "add    x4, %8, #16                 \n"
-                    "ld1    {v24.4s}, [%8]              \n"
-                    "ld1    {v28.4s}, [x4]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-                    "eor    v28.16b, v28.16b, v28.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v25.16b, v24.16b            \n"
-                    "mov    v26.16b, v24.16b            \n"
-                    "mov    v27.16b, v24.16b            \n"
-
-                    "mov    v29.16b, v28.16b            \n"
-                    "mov    v30.16b, v28.16b            \n"
-                    "mov    v31.16b, v28.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n"
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "shll   v1.4s, v1.4h, #16           \n"
-                    "shll   v2.4s, v2.4h, #16           \n"
-                    "shll   v3.4s, v3.4h, #16           \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%1], #64 \n"
-                    "shll   v4.4s, v12.4h, #16          \n"
-                    "shll2  v5.4s, v12.8h, #16          \n"
-                    "shll   v6.4s, v13.4h, #16          \n"
-                    "shll2  v7.4s, v13.8h, #16          \n"
-                    "fmla   v24.4s, v4.4s, v0.s[0]      \n"
-                    "fmla   v25.4s, v4.4s, v0.s[1]      \n"
-                    "fmla   v26.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v27.4s, v4.4s, v0.s[3]      \n"
-                    "fmla   v28.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-                    "fmla   v24.4s, v6.4s, v1.s[0]      \n"
-                    "fmla   v25.4s, v6.4s, v1.s[1]      \n"
-                    "fmla   v26.4s, v6.4s, v1.s[2]      \n"
-                    "fmla   v27.4s, v6.4s, v1.s[3]      \n"
-                    "fmla   v28.4s, v7.4s, v1.s[0]      \n"
-                    "fmla   v29.4s, v7.4s, v1.s[1]      \n"
-                    "fmla   v30.4s, v7.4s, v1.s[2]      \n"
-                    "fmla   v31.4s, v7.4s, v1.s[3]      \n"
-                    "shll   v8.4s, v14.4h, #16          \n"
-                    "shll2  v9.4s, v14.8h, #16          \n"
-                    "shll   v10.4s, v15.4h, #16         \n"
-                    "shll2  v11.4s, v15.8h, #16         \n"
-                    "fmla   v24.4s, v8.4s, v2.s[0]      \n"
-                    "fmla   v25.4s, v8.4s, v2.s[1]      \n"
-                    "fmla   v26.4s, v8.4s, v2.s[2]      \n"
-                    "fmla   v27.4s, v8.4s, v2.s[3]      \n"
-                    "fmla   v28.4s, v9.4s, v2.s[0]      \n"
-                    "fmla   v29.4s, v9.4s, v2.s[1]      \n"
-                    "fmla   v30.4s, v9.4s, v2.s[2]      \n"
-                    "fmla   v31.4s, v9.4s, v2.s[3]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v24.4s, v10.4s, v3.s[0]     \n"
-                    "fmla   v25.4s, v10.4s, v3.s[1]     \n"
-                    "fmla   v26.4s, v10.4s, v3.s[2]     \n"
-                    "fmla   v27.4s, v10.4s, v3.s[3]     \n"
-                    "fmla   v28.4s, v11.4s, v3.s[0]     \n"
-                    "fmla   v29.4s, v11.4s, v3.s[1]     \n"
-                    "fmla   v30.4s, v11.4s, v3.s[2]     \n"
-                    "fmla   v31.4s, v11.4s, v3.s[3]     \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v0.4h}, [%2], #8           \n"
-                    "shll   v0.4s, v0.4h, #16           \n"
-                    "ld1    {v3.8h}, [%1], #16          \n"
-                    "shll   v4.4s, v3.4h, #16           \n"
-                    "shll2  v5.4s, v3.8h, #16           \n"
-                    "fmla   v24.4s, v4.4s, v0.s[0]      \n"
-                    "fmla   v25.4s, v4.4s, v0.s[1]      \n"
-                    "fmla   v26.4s, v4.4s, v0.s[2]      \n"
-                    "fmla   v27.4s, v4.4s, v0.s[3]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v28.4s, v5.4s, v0.s[0]      \n"
-                    "fmla   v29.4s, v5.4s, v0.s[1]      \n"
-                    "fmla   v30.4s, v5.4s, v0.s[2]      \n"
-                    "fmla   v31.4s, v5.4s, v0.s[3]      \n"
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "shrn   v0.4h, v24.4s, #16          \n"
-                    "shrn2  v0.8h, v25.4s, #16          \n"
-                    "shrn   v1.4h, v26.4s, #16          \n"
-                    "shrn2  v1.8h, v27.4s, #16          \n"
-                    "shrn   v2.4h, v28.4s, #16          \n"
-                    "shrn2  v2.8h, v29.4s, #16          \n"
-                    "shrn   v3.4h, v30.4s, #16          \n"
-                    "shrn2  v3.8h, v31.4s, #16          \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    10f                         \n"
-
-                    // if out_elempack == 4
-                    "cmp    %w12, #4                    \n"
-                    "bne    8f                          \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 1          \n"
-                    "st1    {v0.8h, v1.8h}, [%3], #32   \n"
-                    "st1    {v2.8h, v3.8h}, [x4]        \n"
-                    "b      9f                          \n"
-
-                    // if out_elempack == 1
-                    "8:                                 \n"
-                    // transpose8x4
-                    "uzp1   v28.8h, v0.8h, v1.8h        \n"
-                    "uzp2   v29.8h, v0.8h, v1.8h        \n"
-                    "uzp1   v30.8h, v2.8h, v3.8h        \n"
-                    "uzp2   v31.8h, v2.8h, v3.8h        \n"
-
-                    "uzp1   v0.8h, v28.8h, v29.8h       \n"
-                    "uzp2   v2.8h, v28.8h, v29.8h       \n"
-                    "uzp1   v4.8h, v30.8h, v31.8h       \n"
-                    "uzp2   v6.8h, v30.8h, v31.8h       \n"
-
-                    "mov    v1.d[0], v0.d[1]            \n"
-                    "mov    v3.d[0], v2.d[1]            \n"
-                    "mov    v5.d[0], v4.d[1]            \n"
-                    "mov    v7.d[0], v6.d[1]            \n"
-
-                    "add    x4, %3, %w13, sxtw 1        \n"
-                    "st1    {v0.4h}, [%3], #8           \n"
-                    "st1    {v1.4h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v2.4h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v3.4h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v4.4h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v5.4h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v6.4h}, [x4]               \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v7.4h}, [x4]               \n"
-
-                    "9:                                 \n"
-                    "add    %0, %0, #128                \n"
-                    "b      11f                         \n"
-
-                    "10:                                \n"
-                    "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    "11:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum20;
-            float32x4_t _sum21;
-            float32x4_t _sum30;
-            float32x4_t _sum31;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vld1q_f32(pC);
-                    _sum01 = vld1q_f32(pC + 4);
-                    _sum10 = _sum00;
-                    _sum11 = _sum01;
-                    _sum20 = _sum00;
-                    _sum21 = _sum01;
-                    _sum30 = _sum00;
-                    _sum31 = _sum01;
-                }
-                else
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                    _sum20 = vdupq_n_f32(0.f);
-                    _sum21 = vdupq_n_f32(0.f);
-                    _sum30 = vdupq_n_f32(0.f);
-                    _sum31 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-                _sum20 = vld1q_f32(outptr + 4 * 4);
-                _sum21 = vld1q_f32(outptr + 4 * 5);
-                _sum30 = vld1q_f32(outptr + 4 * 6);
-                _sum31 = vld1q_f32(outptr + 4 * 7);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA0 = bfloat2float(vld1_u16(pA));
-                float32x4_t _pA1 = bfloat2float(vld1_u16(pA + 4));
-
-                float32x4_t _pB0 = bfloat2float(vld1_u16(pB));
-
-                _sum00 = vfmaq_laneq_f32(_sum00, _pA0, _pB0, 0);
-                _sum01 = vfmaq_laneq_f32(_sum01, _pA1, _pB0, 0);
-                _sum10 = vfmaq_laneq_f32(_sum10, _pA0, _pB0, 1);
-                _sum11 = vfmaq_laneq_f32(_sum11, _pA1, _pB0, 1);
-                _sum20 = vfmaq_laneq_f32(_sum20, _pA0, _pB0, 2);
-                _sum21 = vfmaq_laneq_f32(_sum21, _pA1, _pB0, 2);
-                _sum30 = vfmaq_laneq_f32(_sum30, _pA0, _pB0, 3);
-                _sum31 = vfmaq_laneq_f32(_sum31, _pA1, _pB0, 3);
-
-                pA += 8;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum10));
-                    vst1_u16(outptr0 + 4 * 2, float2bfloat(_sum20));
-                    vst1_u16(outptr0 + 4 * 3, float2bfloat(_sum30));
-
-                    vst1_u16(outptr0 + out_hstep * 4, float2bfloat(_sum01));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, float2bfloat(_sum11));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 2, float2bfloat(_sum21));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 3, float2bfloat(_sum31));
-
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    uint16x8_t _t0 = vcombine_u16(float2bfloat(_sum00), float2bfloat(_sum01));
-                    uint16x8_t _t1 = vcombine_u16(float2bfloat(_sum10), float2bfloat(_sum11));
-                    uint16x8_t _t2 = vcombine_u16(float2bfloat(_sum20), float2bfloat(_sum21));
-                    uint16x8_t _t3 = vcombine_u16(float2bfloat(_sum30), float2bfloat(_sum31));
-                    transpose8x4_u16(_t0, _t1, _t2, _t3);
-
-                    vst1_u16(outptr0, vget_low_u16(_t0));
-                    vst1_u16(outptr0 + out_hstep * 1, vget_high_u16(_t0));
-                    vst1_u16(outptr0 + out_hstep * 2, vget_low_u16(_t1));
-                    vst1_u16(outptr0 + out_hstep * 3, vget_high_u16(_t1));
-                    vst1_u16(outptr0 + out_hstep * 4, vget_low_u16(_t2));
-                    vst1_u16(outptr0 + out_hstep * 5, vget_high_u16(_t2));
-                    vst1_u16(outptr0 + out_hstep * 6, vget_low_u16(_t3));
-                    vst1_u16(outptr0 + out_hstep * 7, vget_high_u16(_t3));
-
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                vst1q_f32(outptr + 4 * 4, _sum20);
-                vst1q_f32(outptr + 4 * 5, _sum21);
-                vst1q_f32(outptr + 4 * 6, _sum30);
-                vst1q_f32(outptr + 4 * 7, _sum31);
-            }
-
-            outptr += 32;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            const unsigned short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                "b      3f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "add    x4, %8, #16                 \n"
-                "ld1    {v28.4s}, [%8]              \n"
-                "ld1    {v30.4s}, [x4]              \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-                "eor    v30.16b, v30.16b, v30.16b   \n"
-
-                "2:                                 \n"
-                "mov    v29.16b, v28.16b            \n"
-                "mov    v31.16b, v30.16b            \n"
-
-                "3:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                "4:                                 \n"
-                "prfm   pldl1keep, [%2, #128]       \n"
-                "ld1    {v0.4h, v1.4h}, [%2], #16   \n"
-                "shll   v0.4s, v0.4h, #16           \n"
-                "shll   v1.4s, v1.4h, #16           \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%1], #64 \n"
-                "shll   v4.4s, v12.4h, #16          \n"
-                "shll2  v5.4s, v12.8h, #16          \n"
-                "shll   v6.4s, v13.4h, #16          \n"
-                "shll2  v7.4s, v13.8h, #16          \n"
-                "fmla   v28.4s, v4.4s, v0.s[0]      \n"
-                "fmla   v29.4s, v4.4s, v0.s[1]      \n"
-                "fmla   v30.4s, v5.4s, v0.s[0]      \n"
-                "fmla   v31.4s, v5.4s, v0.s[1]      \n"
-                "fmla   v28.4s, v6.4s, v0.s[2]      \n"
-                "fmla   v29.4s, v6.4s, v0.s[3]      \n"
-                "fmla   v30.4s, v7.4s, v0.s[2]      \n"
-                "fmla   v31.4s, v7.4s, v0.s[3]      \n"
-                "shll   v8.4s, v14.4h, #16          \n"
-                "shll2  v9.4s, v14.8h, #16          \n"
-                "shll   v10.4s, v15.4h, #16         \n"
-                "shll2  v11.4s, v15.8h, #16         \n"
-                "fmla   v28.4s, v8.4s, v1.s[0]      \n"
-                "fmla   v29.4s, v8.4s, v1.s[1]      \n"
-                "fmla   v30.4s, v9.4s, v1.s[0]      \n"
-                "fmla   v31.4s, v9.4s, v1.s[1]      \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v28.4s, v10.4s, v1.s[2]     \n"
-                "fmla   v29.4s, v10.4s, v1.s[3]     \n"
-                "fmla   v30.4s, v11.4s, v1.s[2]     \n"
-                "fmla   v31.4s, v11.4s, v1.s[3]     \n"
-                "bne    4b                          \n"
-
-                "5:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    7f                          \n"
-
-                "6:                                 \n"
-                "ld1    {v0.s}[0], [%2], #4         \n"
-                "shll   v0.4s, v0.4h, #16           \n"
-                "ld1    {v3.8h}, [%1], #16          \n"
-                "shll   v4.4s, v3.4h, #16           \n"
-                "shll2  v5.4s, v3.8h, #16           \n"
-                "fmla   v28.4s, v4.4s, v0.s[0]      \n"
-                "fmla   v29.4s, v4.4s, v0.s[1]      \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v30.4s, v5.4s, v0.s[0]      \n"
-                "fmla   v31.4s, v5.4s, v0.s[1]      \n"
-                "bne    6b                          \n"
-
-                "7:                                 \n"
-                "shrn   v0.4h, v28.4s, #16          \n"
-                "shrn2  v0.8h, v29.4s, #16          \n"
-                "shrn   v1.4h, v30.4s, #16          \n"
-                "shrn2  v1.8h, v31.4s, #16          \n"
-                "tst    %w11, #255                  \n"
-                "beq    10f                         \n"
-
-                // if out_elempack == 4
-                "cmp    %w12, #4                    \n"
-                "bne    8f                          \n"
-
-                "lsl    w4, %w13, #2                \n"
-                "add    x4, %3, w4, sxtw 1          \n"
-                "st1    {v0.8h}, [%3], #16          \n"
-                "st1    {v1.8h}, [x4]               \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // transpose8x2
-                "uzp1   v2.8h, v0.8h, v1.8h         \n"
-                "uzp2   v3.8h, v0.8h, v1.8h         \n"
-                "uzp1   v0.8h, v2.8h, v3.8h         \n"
-                "uzp2   v1.8h, v2.8h, v3.8h         \n"
-
-                "add    x4, %3, %w13, sxtw 1        \n"
-                "st1    {v0.s}[0], [%3], #4         \n"
-                "st1    {v0.s}[2], [x4]             \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v1.s}[0], [x4]             \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v1.s}[2], [x4]             \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v0.s}[1], [x4]             \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v0.s}[3], [x4]             \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v1.s}[1], [x4]             \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v1.s}[3], [x4]             \n"
-
-                "9:                                 \n"
-                "add    %0, %0, #64                 \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vld1q_f32(pC);
-                    _sum01 = vld1q_f32(pC + 4);
-                    _sum10 = _sum00;
-                    _sum11 = _sum01;
-                }
-                else
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA0 = bfloat2float(vld1_u16(pA));
-                float32x4_t _pA1 = bfloat2float(vld1_u16(pA + 4));
-
-                float32x2_t _pB0 = vget_low_f32(bfloat2float(vld1_u16(pB)));
-
-                _sum00 = vfmaq_lane_f32(_sum00, _pA0, _pB0, 0);
-                _sum01 = vfmaq_lane_f32(_sum01, _pA1, _pB0, 0);
-                _sum10 = vfmaq_lane_f32(_sum10, _pA0, _pB0, 1);
-                _sum11 = vfmaq_lane_f32(_sum11, _pA1, _pB0, 1);
-
-                pA += 8;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum10));
-
-                    vst1_u16(outptr0 + out_hstep * 4, float2bfloat(_sum01));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, float2bfloat(_sum11));
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short sum0[8];
-                    unsigned short sum1[8];
-                    vst1_u16(sum0, float2bfloat(_sum00));
-                    vst1_u16(sum0 + 4, float2bfloat(_sum01));
-                    vst1_u16(sum1, float2bfloat(_sum10));
-                    vst1_u16(sum1 + 4, float2bfloat(_sum11));
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[out_hstep * 4] = sum0[4];
-                    outptr0[out_hstep * 5] = sum0[5];
-                    outptr0[out_hstep * 6] = sum0[6];
-                    outptr0[out_hstep * 7] = sum0[7];
-
-                    outptr0[1] = sum1[0];
-                    outptr0[out_hstep + 1] = sum1[1];
-                    outptr0[out_hstep * 2 + 1] = sum1[2];
-                    outptr0[out_hstep * 3 + 1] = sum1[3];
-                    outptr0[out_hstep * 4 + 1] = sum1[4];
-                    outptr0[out_hstep * 5 + 1] = sum1[5];
-                    outptr0[out_hstep * 6 + 1] = sum1[6];
-                    outptr0[out_hstep * 7 + 1] = sum1[7];
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-            }
-
-            outptr += 16;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            const unsigned short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v30.4s, v31.4s}, [%0]      \n"
-                "b      3f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "ld1    {v30.4s, v31.4s}, [%8]      \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v30.16b, v30.16b, v30.16b   \n"
-                "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                "2:                                 \n"
-
-                "3:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-                "eor    v29.16b, v29.16b, v29.16b   \n"
-                "4:                                 \n"
-                "prfm   pldl1keep, [%2, #64]        \n"
-                "ld1    {v0.4h}, [%2], #8           \n"
-                "shll   v0.4s, v0.4h, #16           \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%1], #64 \n"
-                "shll   v4.4s, v12.4h, #16          \n"
-                "shll2  v5.4s, v12.8h, #16          \n"
-                "shll   v6.4s, v13.4h, #16          \n"
-                "shll2  v7.4s, v13.8h, #16          \n"
-                "fmla   v28.4s, v4.4s, v0.s[0]      \n"
-                "fmla   v29.4s, v5.4s, v0.s[0]      \n"
-                "fmla   v30.4s, v6.4s, v0.s[1]      \n"
-                "fmla   v31.4s, v7.4s, v0.s[1]      \n"
-                "shll   v8.4s, v14.4h, #16          \n"
-                "shll2  v9.4s, v14.8h, #16          \n"
-                "shll   v10.4s, v15.4h, #16         \n"
-                "shll2  v11.4s, v15.8h, #16         \n"
-                "fmla   v28.4s, v8.4s, v0.s[2]      \n"
-                "fmla   v29.4s, v9.4s, v0.s[2]      \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v30.4s, v10.4s, v0.s[3]     \n"
-                "fmla   v31.4s, v11.4s, v0.s[3]     \n"
-                "bne    4b                          \n"
-                "fadd   v30.4s, v30.4s, v28.4s      \n"
-                "fadd   v31.4s, v31.4s, v29.4s      \n"
-
-                "5:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    7f                          \n"
-
-                "6:                                 \n"
-                "ld1r   {v0.4h}, [%2], #2           \n"
-                "shll   v0.4s, v0.4h, #16           \n"
-                "ld1    {v3.8h}, [%1], #16          \n"
-                "shll   v4.4s, v3.4h, #16           \n"
-                "shll2  v5.4s, v3.8h, #16           \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v30.4s, v4.4s, v0.4s        \n"
-                "fmla   v31.4s, v5.4s, v0.4s        \n"
-                "bne    6b                          \n"
-
-                "7:                                 \n"
-                "shrn   v30.4h, v30.4s, #16         \n"
-                "shrn   v31.4h, v31.4s, #16         \n"
-                "tst    %w11, #255                  \n"
-                "beq    10f                         \n"
-
-                // if out_elempack == 4
-                "cmp    %w12, #4                    \n"
-                "bne    8f                          \n"
-
-                "lsl    w4, %w13, #2                \n"
-                "add    x4, %3, w4, sxtw 1          \n"
-                "st1    {v30.4h}, [%3], #8          \n"
-                "st1    {v31.4h}, [x4]              \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                "add    x4, %3, %w13, sxtw 1        \n"
-                "st1    {v30.h}[0], [%3], #2        \n"
-                "st1    {v30.h}[1], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v30.h}[2], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v30.h}[3], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v31.h}[0], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v31.h}[1], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v31.h}[2], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v31.h}[3], [x4]            \n"
-
-                "9:                                 \n"
-                "add    %0, %0, #32                 \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v30.4s, v31.4s}, [%0], #32 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vld1q_f32(pC);
-                    _sum01 = vld1q_f32(pC + 4);
-                }
-                else
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA0 = bfloat2float(vld1_u16(pA));
-                float32x4_t _pA1 = bfloat2float(vld1_u16(pA + 4));
-
-                float32x4_t _pB = bfloat2float(vld1_dup_u16(pB));
-
-                _sum00 = vfmaq_f32(_sum00, _pA0, _pB);
-                _sum01 = vfmaq_f32(_sum01, _pA1, _pB);
-
-                pA += 8;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + out_hstep * 4, float2bfloat(_sum01));
-                    outptr0 += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short sum0[8];
-                    vst1_u16(sum0, float2bfloat(_sum00));
-                    vst1_u16(sum0 + 4, float2bfloat(_sum01));
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep * 1] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[out_hstep * 4] = sum0[4];
-                    outptr0[out_hstep * 5] = sum0[5];
-                    outptr0[out_hstep * 6] = sum0[6];
-                    outptr0[out_hstep * 7] = sum0[7];
-                    outptr0++;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-            }
-
-            outptr += 8;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-
-        pAT += max_kk * 8;
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        unsigned short* outptr0 = (unsigned short*)top_blob + (i + ii) * out_hstep + j * out_elempack;
-
-        const unsigned short* pB = pBT;
-
-        if (pC)
-        {
-            pC = (const float*)CT_tile + i + ii;
-        }
-
-        int jj = 0;
-#if __aarch64__
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            const unsigned short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                "subs   %0, %0, #128                \n"
-                "b      3f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "ld1    {v20.4s}, [%8]              \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v20.16b, v20.16b, v20.16b   \n"
-
-                "2:                                 \n"
-                "mov    v21.16b, v20.16b            \n"
-                "mov    v22.16b, v20.16b            \n"
-                "mov    v23.16b, v20.16b            \n"
-                "mov    v24.16b, v20.16b            \n"
-                "mov    v25.16b, v20.16b            \n"
-                "mov    v26.16b, v20.16b            \n"
-                "mov    v27.16b, v20.16b            \n"
-                "mov    v28.16b, v20.16b            \n"
-                "mov    v29.16b, v20.16b            \n"
-                "mov    v30.16b, v20.16b            \n"
-                "mov    v31.16b, v20.16b            \n"
-
-                "3:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                "4:                                 \n"
-                "prfm   pldl1keep, [%2, #512]       \n"
-                "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%2], #64 \n"
-                "shll   v0.4s, v4.4h, #16           \n"
-                "shll2  v1.4s, v4.8h, #16           \n"
-                "shll   v2.4s, v5.4h, #16           \n"
-                "shll2  v3.4s, v5.8h, #16           \n"
-                "prfm   pldl1keep, [%1, #256]       \n"
-                "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%1], #32 \n"
-                "shll   v16.4s, v16.4h, #16         \n"
-                "shll   v17.4s, v17.4h, #16         \n"
-                "shll   v18.4s, v18.4h, #16         \n"
-                "shll   v19.4s, v19.4h, #16         \n"
-                "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v21.4s, v16.4s, v0.s[1]     \n"
-                "fmla   v22.4s, v16.4s, v0.s[2]     \n"
-                "fmla   v23.4s, v16.4s, v0.s[3]     \n"
-                "fmla   v24.4s, v16.4s, v1.s[0]     \n"
-                "fmla   v25.4s, v16.4s, v1.s[1]     \n"
-                "fmla   v26.4s, v16.4s, v1.s[2]     \n"
-                "fmla   v27.4s, v16.4s, v1.s[3]     \n"
-                "fmla   v28.4s, v16.4s, v2.s[0]     \n"
-                "fmla   v29.4s, v16.4s, v2.s[1]     \n"
-                "fmla   v30.4s, v16.4s, v2.s[2]     \n"
-                "fmla   v31.4s, v16.4s, v2.s[3]     \n"
-                "shll   v4.4s, v6.4h, #16           \n"
-                "shll2  v5.4s, v6.8h, #16           \n"
-                "shll   v6.4s, v7.4h, #16           \n"
-                "shll2  v7.4s, v7.8h, #16           \n"
-                "fmla   v20.4s, v17.4s, v3.s[0]     \n"
-                "fmla   v21.4s, v17.4s, v3.s[1]     \n"
-                "fmla   v22.4s, v17.4s, v3.s[2]     \n"
-                "fmla   v23.4s, v17.4s, v3.s[3]     \n"
-                "fmla   v24.4s, v17.4s, v4.s[0]     \n"
-                "fmla   v25.4s, v17.4s, v4.s[1]     \n"
-                "fmla   v26.4s, v17.4s, v4.s[2]     \n"
-                "fmla   v27.4s, v17.4s, v4.s[3]     \n"
-                "fmla   v28.4s, v17.4s, v5.s[0]     \n"
-                "fmla   v29.4s, v17.4s, v5.s[1]     \n"
-                "fmla   v30.4s, v17.4s, v5.s[2]     \n"
-                "fmla   v31.4s, v17.4s, v5.s[3]     \n"
-                "prfm   pldl1keep, [%2, #256]       \n"
-                "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n"
-                "shll   v0.4s, v0.4h, #16           \n"
-                "shll   v1.4s, v1.4h, #16           \n"
-                "shll   v2.4s, v2.4h, #16           \n"
-                "shll   v3.4s, v3.4h, #16           \n"
-                "fmla   v20.4s, v18.4s, v6.s[0]     \n"
-                "fmla   v21.4s, v18.4s, v6.s[1]     \n"
-                "fmla   v22.4s, v18.4s, v6.s[2]     \n"
-                "fmla   v23.4s, v18.4s, v6.s[3]     \n"
-                "fmla   v24.4s, v18.4s, v7.s[0]     \n"
-                "fmla   v25.4s, v18.4s, v7.s[1]     \n"
-                "fmla   v26.4s, v18.4s, v7.s[2]     \n"
-                "fmla   v27.4s, v18.4s, v7.s[3]     \n"
-                "fmla   v28.4s, v18.4s, v0.s[0]     \n"
-                "fmla   v29.4s, v18.4s, v0.s[1]     \n"
-                "fmla   v30.4s, v18.4s, v0.s[2]     \n"
-                "fmla   v31.4s, v18.4s, v0.s[3]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v20.4s, v19.4s, v1.s[0]     \n"
-                "fmla   v21.4s, v19.4s, v1.s[1]     \n"
-                "fmla   v22.4s, v19.4s, v1.s[2]     \n"
-                "fmla   v23.4s, v19.4s, v1.s[3]     \n"
-                "fmla   v24.4s, v19.4s, v2.s[0]     \n"
-                "fmla   v25.4s, v19.4s, v2.s[1]     \n"
-                "fmla   v26.4s, v19.4s, v2.s[2]     \n"
-                "fmla   v27.4s, v19.4s, v2.s[3]     \n"
-                "fmla   v28.4s, v19.4s, v3.s[0]     \n"
-                "fmla   v29.4s, v19.4s, v3.s[1]     \n"
-                "fmla   v30.4s, v19.4s, v3.s[2]     \n"
-                "fmla   v31.4s, v19.4s, v3.s[3]     \n"
-                "bne    4b                          \n"
-
-                "5:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    7f                          \n"
-
-                "6:                                 \n"
-                "ld1    {v0.4h, v1.4h, v2.4h}, [%2], #24 \n"
-                "shll   v0.4s, v0.4h, #16           \n"
-                "shll   v1.4s, v1.4h, #16           \n"
-                "shll   v2.4s, v2.4h, #16           \n"
-                "ld1    {v16.4h}, [%1], #8          \n"
-                "shll   v16.4s, v16.4h, #16         \n"
-                "fmla   v20.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v21.4s, v16.4s, v0.s[1]     \n"
-                "fmla   v22.4s, v16.4s, v0.s[2]     \n"
-                "fmla   v23.4s, v16.4s, v0.s[3]     \n"
-                "fmla   v24.4s, v16.4s, v1.s[0]     \n"
-                "fmla   v25.4s, v16.4s, v1.s[1]     \n"
-                "fmla   v26.4s, v16.4s, v1.s[2]     \n"
-                "fmla   v27.4s, v16.4s, v1.s[3]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v28.4s, v16.4s, v2.s[0]     \n"
-                "fmla   v29.4s, v16.4s, v2.s[1]     \n"
-                "fmla   v30.4s, v16.4s, v2.s[2]     \n"
-                "fmla   v31.4s, v16.4s, v2.s[3]     \n"
-                "bne    6b                          \n"
-
-                "7:                                 \n"
-                "shrn   v0.4h, v20.4s, #16          \n"
-                "shrn2  v0.8h, v21.4s, #16          \n"
-                "shrn   v1.4h, v22.4s, #16          \n"
-                "shrn2  v1.8h, v23.4s, #16          \n"
-                "shrn   v2.4h, v24.4s, #16          \n"
-                "shrn2  v2.8h, v25.4s, #16          \n"
-                "shrn   v3.4h, v26.4s, #16          \n"
-                "shrn2  v3.8h, v27.4s, #16          \n"
-                "shrn   v4.4h, v28.4s, #16          \n"
-                "shrn2  v4.8h, v29.4s, #16          \n"
-                "shrn   v5.4h, v30.4s, #16          \n"
-                "shrn2  v5.8h, v31.4s, #16          \n"
-                "tst    %w11, #255                  \n"
-                "beq    10f                         \n"
-
-                // if out_elempack == 4
-                "cmp    %w12, #4                    \n"
-                "bne    8f                          \n"
-
-                "st1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%3], #64 \n"
-                "st1    {v4.8h, v5.8h}, [%3], #32   \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // transpose4x12
-                "uzp1   v20.8h, v0.8h, v1.8h        \n"
-                "uzp2   v21.8h, v0.8h, v1.8h        \n"
-                "uzp1   v22.8h, v2.8h, v3.8h        \n"
-                "uzp2   v23.8h, v2.8h, v3.8h        \n"
-                "uzp1   v24.8h, v4.8h, v5.8h        \n"
-                "uzp2   v25.8h, v4.8h, v5.8h        \n"
-
-                "uzp1   v0.8h, v20.8h, v21.8h       \n"
-                "uzp2   v6.8h, v20.8h, v21.8h       \n"
-                "uzp1   v1.8h, v22.8h, v23.8h       \n"
-                "uzp2   v7.8h, v22.8h, v23.8h       \n"
-                "uzp1   v2.8h, v24.8h, v25.8h       \n"
-                "uzp2   v8.8h, v24.8h, v25.8h       \n"
-
-                "mov    v3.d[0], v0.d[1]            \n"
-                "mov    v4.d[0], v1.d[1]            \n"
-                "mov    v5.d[0], v2.d[1]            \n"
-                "mov    v9.d[0], v6.d[1]            \n"
-                "mov    v10.d[0], v7.d[1]           \n"
-                "mov    v11.d[0], v8.d[1]           \n"
-
-                "add    x4, %3, %w13, sxtw 1        \n"
-                "st1    {v0.4h, v1.4h, v2.4h}, [%3], #24 \n"
-                "st1    {v3.4h, v4.4h, v5.4h}, [x4] \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v6.4h, v7.4h, v8.4h}, [x4] \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v9.4h, v10.4h, v11.4h}, [x4] \n"
-
-                "9:                                 \n"
-                "add    %0, %0, #192                \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-            float32x4_t _sum3;
-            float32x4_t _sum4;
-            float32x4_t _sum5;
-            float32x4_t _sum6;
-            float32x4_t _sum7;
-            float32x4_t _sum8;
-            float32x4_t _sum9;
-            float32x4_t _suma;
-            float32x4_t _sumb;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1q_f32(pC);
-                    _sum1 = _sum0;
-                    _sum2 = _sum0;
-                    _sum3 = _sum0;
-                    _sum4 = _sum0;
-                    _sum5 = _sum0;
-                    _sum6 = _sum0;
-                    _sum7 = _sum0;
-                    _sum8 = _sum0;
-                    _sum9 = _sum0;
-                    _suma = _sum0;
-                    _sumb = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                    _sum2 = vdupq_n_f32(0.f);
-                    _sum3 = vdupq_n_f32(0.f);
-                    _sum4 = vdupq_n_f32(0.f);
-                    _sum5 = vdupq_n_f32(0.f);
-                    _sum6 = vdupq_n_f32(0.f);
-                    _sum7 = vdupq_n_f32(0.f);
-                    _sum8 = vdupq_n_f32(0.f);
-                    _sum9 = vdupq_n_f32(0.f);
-                    _suma = vdupq_n_f32(0.f);
-                    _sumb = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4 * 1);
-                _sum2 = vld1q_f32(outptr + 4 * 2);
-                _sum3 = vld1q_f32(outptr + 4 * 3);
-                _sum4 = vld1q_f32(outptr + 4 * 4);
-                _sum5 = vld1q_f32(outptr + 4 * 5);
-                _sum6 = vld1q_f32(outptr + 4 * 6);
-                _sum7 = vld1q_f32(outptr + 4 * 7);
-                _sum8 = vld1q_f32(outptr + 4 * 8);
-                _sum9 = vld1q_f32(outptr + 4 * 9);
-                _suma = vld1q_f32(outptr + 4 * 10);
-                _sumb = vld1q_f32(outptr + 4 * 11);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA = bfloat2float(vld1_u16(pA));
-                float32x4_t _pB0 = bfloat2float(vld1_u16(pB));
-                float32x4_t _pB1 = bfloat2float(vld1_u16(pB + 4));
-                float32x4_t _pB2 = bfloat2float(vld1_u16(pB + 8));
-
-                _sum0 = vfmaq_laneq_f32(_sum0, _pA, _pB0, 0);
-                _sum1 = vfmaq_laneq_f32(_sum1, _pA, _pB0, 1);
-                _sum2 = vfmaq_laneq_f32(_sum2, _pA, _pB0, 2);
-                _sum3 = vfmaq_laneq_f32(_sum3, _pA, _pB0, 3);
-                _sum4 = vfmaq_laneq_f32(_sum4, _pA, _pB1, 0);
-                _sum5 = vfmaq_laneq_f32(_sum5, _pA, _pB1, 1);
-                _sum6 = vfmaq_laneq_f32(_sum6, _pA, _pB1, 2);
-                _sum7 = vfmaq_laneq_f32(_sum7, _pA, _pB1, 3);
-                _sum8 = vfmaq_laneq_f32(_sum8, _pA, _pB2, 0);
-                _sum9 = vfmaq_laneq_f32(_sum9, _pA, _pB2, 1);
-                _suma = vfmaq_laneq_f32(_suma, _pA, _pB2, 2);
-                _sumb = vfmaq_laneq_f32(_sumb, _pA, _pB2, 3);
-
-                pA += 4;
-                pB += 12;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum1));
-                    vst1_u16(outptr0 + 4 * 2, float2bfloat(_sum2));
-                    vst1_u16(outptr0 + 4 * 3, float2bfloat(_sum3));
-                    vst1_u16(outptr0 + 4 * 4, float2bfloat(_sum4));
-                    vst1_u16(outptr0 + 4 * 5, float2bfloat(_sum5));
-                    vst1_u16(outptr0 + 4 * 6, float2bfloat(_sum6));
-                    vst1_u16(outptr0 + 4 * 7, float2bfloat(_sum7));
-                    vst1_u16(outptr0 + 4 * 8, float2bfloat(_sum8));
-                    vst1_u16(outptr0 + 4 * 9, float2bfloat(_sum9));
-                    vst1_u16(outptr0 + 4 * 10, float2bfloat(_suma));
-                    vst1_u16(outptr0 + 4 * 11, float2bfloat(_sumb));
-                    outptr0 += 48;
-                }
-                if (out_elempack == 1)
-                {
-                    uint16x4_t _t0 = float2bfloat(_sum0);
-                    uint16x4_t _t1 = float2bfloat(_sum1);
-                    uint16x4_t _t2 = float2bfloat(_sum2);
-                    uint16x4_t _t3 = float2bfloat(_sum3);
-                    uint16x4_t _t4 = float2bfloat(_sum4);
-                    uint16x4_t _t5 = float2bfloat(_sum5);
-                    uint16x4_t _t6 = float2bfloat(_sum6);
-                    uint16x4_t _t7 = float2bfloat(_sum7);
-                    uint16x4_t _t8 = float2bfloat(_sum8);
-                    uint16x4_t _t9 = float2bfloat(_sum9);
-                    uint16x4_t _ta = float2bfloat(_suma);
-                    uint16x4_t _tb = float2bfloat(_sumb);
-                    transpose4x12_u16(_t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _ta, _tb);
-
-                    vst1_u16(outptr0, _t0);
-                    vst1_u16(outptr0 + 4, _t1);
-                    vst1_u16(outptr0 + 8, _t2);
-                    vst1_u16(outptr0 + out_hstep, _t3);
-                    vst1_u16(outptr0 + out_hstep + 4, _t4);
-                    vst1_u16(outptr0 + out_hstep + 8, _t5);
-                    vst1_u16(outptr0 + out_hstep * 2, _t6);
-                    vst1_u16(outptr0 + out_hstep * 2 + 4, _t7);
-                    vst1_u16(outptr0 + out_hstep * 2 + 8, _t8);
-                    vst1_u16(outptr0 + out_hstep * 3, _t9);
-                    vst1_u16(outptr0 + out_hstep * 3 + 4, _ta);
-                    vst1_u16(outptr0 + out_hstep * 3 + 8, _tb);
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 4 * 2, _sum2);
-                vst1q_f32(outptr + 4 * 3, _sum3);
-                vst1q_f32(outptr + 4 * 4, _sum4);
-                vst1q_f32(outptr + 4 * 5, _sum5);
-                vst1q_f32(outptr + 4 * 6, _sum6);
-                vst1q_f32(outptr + 4 * 7, _sum7);
-                vst1q_f32(outptr + 4 * 8, _sum8);
-                vst1q_f32(outptr + 4 * 9, _sum9);
-                vst1q_f32(outptr + 4 * 10, _suma);
-                vst1q_f32(outptr + 4 * 11, _sumb);
-            }
-
-            outptr += 48;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-#endif // __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            const unsigned short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                "subs   %0, %0, #64                 \n"
-                "b      3f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "ld1    {v24.4s}, [%8]              \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v24.16b, v24.16b, v24.16b   \n"
-
-                "2:                                 \n"
-                "mov    v25.16b, v24.16b            \n"
-                "mov    v26.16b, v24.16b            \n"
-                "mov    v27.16b, v24.16b            \n"
-                "mov    v28.16b, v24.16b            \n"
-                "mov    v29.16b, v24.16b            \n"
-                "mov    v30.16b, v24.16b            \n"
-                "mov    v31.16b, v24.16b            \n"
-
-                "3:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                "4:                                 \n"
-                "prfm   pldl1keep, [%2, #512]       \n"
-                "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%2], #64 \n"
-                "shll   v0.4s, v4.4h, #16           \n"
-                "shll2  v1.4s, v4.8h, #16           \n"
-                "shll   v2.4s, v5.4h, #16           \n"
-                "shll2  v3.4s, v5.8h, #16           \n"
-                "prfm   pldl1keep, [%1, #256]       \n"
-                "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%1], #32 \n"
-                "shll   v16.4s, v16.4h, #16         \n"
-                "shll   v17.4s, v17.4h, #16         \n"
-                "shll   v18.4s, v18.4h, #16         \n"
-                "shll   v19.4s, v19.4h, #16         \n"
-                "fmla   v24.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v25.4s, v16.4s, v0.s[1]     \n"
-                "fmla   v26.4s, v16.4s, v0.s[2]     \n"
-                "fmla   v27.4s, v16.4s, v0.s[3]     \n"
-                "fmla   v28.4s, v16.4s, v1.s[0]     \n"
-                "fmla   v29.4s, v16.4s, v1.s[1]     \n"
-                "fmla   v30.4s, v16.4s, v1.s[2]     \n"
-                "fmla   v31.4s, v16.4s, v1.s[3]     \n"
-                "fmla   v24.4s, v17.4s, v2.s[0]     \n"
-                "fmla   v25.4s, v17.4s, v2.s[1]     \n"
-                "fmla   v26.4s, v17.4s, v2.s[2]     \n"
-                "fmla   v27.4s, v17.4s, v2.s[3]     \n"
-                "fmla   v28.4s, v17.4s, v3.s[0]     \n"
-                "fmla   v29.4s, v17.4s, v3.s[1]     \n"
-                "fmla   v30.4s, v17.4s, v3.s[2]     \n"
-                "fmla   v31.4s, v17.4s, v3.s[3]     \n"
-                "shll   v4.4s, v6.4h, #16           \n"
-                "shll2  v5.4s, v6.8h, #16           \n"
-                "shll   v6.4s, v7.4h, #16           \n"
-                "shll2  v7.4s, v7.8h, #16           \n"
-                "fmla   v24.4s, v18.4s, v4.s[0]     \n"
-                "fmla   v25.4s, v18.4s, v4.s[1]     \n"
-                "fmla   v26.4s, v18.4s, v4.s[2]     \n"
-                "fmla   v27.4s, v18.4s, v4.s[3]     \n"
-                "fmla   v28.4s, v18.4s, v5.s[0]     \n"
-                "fmla   v29.4s, v18.4s, v5.s[1]     \n"
-                "fmla   v30.4s, v18.4s, v5.s[2]     \n"
-                "fmla   v31.4s, v18.4s, v5.s[3]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v24.4s, v19.4s, v6.s[0]     \n"
-                "fmla   v25.4s, v19.4s, v6.s[1]     \n"
-                "fmla   v26.4s, v19.4s, v6.s[2]     \n"
-                "fmla   v27.4s, v19.4s, v6.s[3]     \n"
-                "fmla   v28.4s, v19.4s, v7.s[0]     \n"
-                "fmla   v29.4s, v19.4s, v7.s[1]     \n"
-                "fmla   v30.4s, v19.4s, v7.s[2]     \n"
-                "fmla   v31.4s, v19.4s, v7.s[3]     \n"
-                "bne    4b                          \n"
-
-                "5:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    7f                          \n"
-
-                "6:                                 \n"
-                "ld1    {v0.4h, v1.4h}, [%2], #16   \n"
-                "shll   v0.4s, v0.4h, #16           \n"
-                "shll   v1.4s, v1.4h, #16           \n"
-                "ld1    {v16.4h}, [%1], #8          \n"
-                "shll   v16.4s, v16.4h, #16         \n"
-                "fmla   v24.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v25.4s, v16.4s, v0.s[1]     \n"
-                "fmla   v26.4s, v16.4s, v0.s[2]     \n"
-                "fmla   v27.4s, v16.4s, v0.s[3]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v28.4s, v16.4s, v1.s[0]     \n"
-                "fmla   v29.4s, v16.4s, v1.s[1]     \n"
-                "fmla   v30.4s, v16.4s, v1.s[2]     \n"
-                "fmla   v31.4s, v16.4s, v1.s[3]     \n"
-                "bne    6b                          \n"
-
-                "7:                                 \n"
-                "shrn   v0.4h, v24.4s, #16          \n"
-                "shrn2  v0.8h, v25.4s, #16          \n"
-                "shrn   v1.4h, v26.4s, #16          \n"
-                "shrn2  v1.8h, v27.4s, #16          \n"
-                "shrn   v2.4h, v28.4s, #16          \n"
-                "shrn2  v2.8h, v29.4s, #16          \n"
-                "shrn   v3.4h, v30.4s, #16          \n"
-                "shrn2  v3.8h, v31.4s, #16          \n"
-                "tst    %w11, #255                  \n"
-                "beq    10f                         \n"
-
-                // if out_elempack == 4
-                "cmp    %w12, #4                    \n"
-                "bne    8f                          \n"
-
-                "st1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%3], #64 \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // transpose4x8
-                "uzp1   v20.8h, v0.8h, v1.8h        \n"
-                "uzp2   v21.8h, v0.8h, v1.8h        \n"
-                "uzp1   v22.8h, v2.8h, v3.8h        \n"
-                "uzp2   v23.8h, v2.8h, v3.8h        \n"
-
-                "uzp1   v0.8h, v20.8h, v22.8h       \n"
-                "uzp2   v2.8h, v20.8h, v22.8h       \n"
-                "uzp1   v1.8h, v21.8h, v23.8h       \n"
-                "uzp2   v3.8h, v21.8h, v23.8h       \n"
-
-                "add    x4, %3, %w13, sxtw 1        \n"
-                "st1    {v0.8h}, [%3], #16          \n"
-                "st1    {v1.8h}, [x4]               \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v2.8h}, [x4]               \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v3.8h}, [x4]               \n"
-
-                "9:                                 \n"
-                "add    %0, %0, #128                \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-            asm volatile(
-                "cmp        %10, #0             \n"
-                "beq        0f                  \n"
-
-                "vldm       %0!, {d16-d23}      \n"
-                "vldm       %0, {d24-d31}       \n"
-                "sub        %0, %0, #64         \n"
-                "b          3f                  \n"
-
-                "0:                             \n"
-                // if pC
-                "cmp        %8, #0              \n"
-                "beq        1f                  \n"
-
-                "vld1.f32   {d16-d17}, [%8]     \n"
-                "b          2f                  \n"
-
-                // else
-                "1:                             \n"
-                "veor       q8, q8              \n"
-
-                "2:                             \n"
-                "vmov       q9, q8              \n"
-                "vmov       q10, q8             \n"
-                "vmov       q11, q8             \n"
-                "vmov       q12, q8             \n"
-                "vmov       q13, q8             \n"
-                "vmov       q14, q8             \n"
-                "vmov       q15, q8             \n"
-
-                "3:                             \n"
-                "lsr        r4, %9, #2          \n" // r4 = max_kk >> 2
-                "cmp        r4, #0              \n"
-                "beq        5f                  \n"
-
-                "4:                             \n"
-                "pld        [%2, #256]          \n"
-                "vld1.u16   {d4-d7}, [%2 :64]!  \n"
-                "pld        [%1, #256]          \n"
-                "vld1.u16   {d12-d15}, [%1 :64]! \n"
-                "vshll.u16  q0, d4, #16         \n"
-                "vshll.u16  q1, d5, #16         \n"
-                "vshll.u16  q2, d6, #16         \n"
-                "vshll.u16  q3, d7, #16         \n"
-                "vshll.u16  q4, d12, #16        \n"
-                "vshll.u16  q5, d13, #16        \n"
-                "vshll.u16  q6, d14, #16        \n"
-                "vshll.u16  q7, d15, #16        \n"
-                "vmla.f32   q8, q4, d0[0]       \n"
-                "vmla.f32   q9, q4, d0[1]       \n"
-                "vmla.f32   q10, q4, d1[0]      \n"
-                "vmla.f32   q11, q4, d1[1]      \n"
-                "vmla.f32   q12, q4, d2[0]      \n"
-                "vmla.f32   q13, q4, d2[1]      \n"
-                "vmla.f32   q14, q4, d3[0]      \n"
-                "vmla.f32   q15, q4, d3[1]      \n"
-                "vmla.f32   q8, q5, d4[0]       \n"
-                "vmla.f32   q9, q5, d4[1]       \n"
-                "vmla.f32   q10, q5, d5[0]      \n"
-                "vmla.f32   q11, q5, d5[1]      \n"
-                "vmla.f32   q12, q5, d6[0]      \n"
-                "vmla.f32   q13, q5, d6[1]      \n"
-                "vmla.f32   q14, q5, d7[0]      \n"
-                "vmla.f32   q15, q5, d7[1]      \n"
-                "pld        [%2, #256]          \n"
-                "vld1.u16   {d4-d7}, [%2 :64]!  \n"
-                "vshll.u16  q0, d4, #16         \n"
-                "vshll.u16  q1, d5, #16         \n"
-                "vshll.u16  q2, d6, #16         \n"
-                "vshll.u16  q3, d7, #16         \n"
-                "vmla.f32   q8, q6, d0[0]       \n"
-                "vmla.f32   q9, q6, d0[1]       \n"
-                "vmla.f32   q10, q6, d1[0]      \n"
-                "vmla.f32   q11, q6, d1[1]      \n"
-                "vmla.f32   q12, q6, d2[0]      \n"
-                "vmla.f32   q13, q6, d2[1]      \n"
-                "vmla.f32   q14, q6, d3[0]      \n"
-                "vmla.f32   q15, q6, d3[1]      \n"
-                "subs       r4, r4, #1          \n"
-                "vmla.f32   q8, q7, d4[0]       \n"
-                "vmla.f32   q9, q7, d4[1]       \n"
-                "vmla.f32   q10, q7, d5[0]      \n"
-                "vmla.f32   q11, q7, d5[1]      \n"
-                "vmla.f32   q12, q7, d6[0]      \n"
-                "vmla.f32   q13, q7, d6[1]      \n"
-                "vmla.f32   q14, q7, d7[0]      \n"
-                "vmla.f32   q15, q7, d7[1]      \n"
-                "bne        4b                  \n"
-
-                "5:                             \n"
-                "and        r4, %9, #3          \n" // r4 = remain = max_kk & 3
-                "cmp        r4, #0              \n"
-                "beq        7f                  \n"
-
-                "6:                             \n"
-                "vld1.u16   {d2-d3}, [%2 :64]!  \n"
-                "vshll.u16  q0, d2, #16         \n"
-                "vshll.u16  q1, d3, #16         \n"
-                "vld1.u16   {d9}, [%1 :64]!     \n"
-                "vshll.u16  q4, d9, #16         \n"
-                "vmla.f32   q8, q4, d0[0]       \n"
-                "vmla.f32   q9, q4, d0[1]       \n"
-                "vmla.f32   q10, q4, d1[0]      \n"
-                "vmla.f32   q11, q4, d1[1]      \n"
-                "subs       r4, r4, #1          \n"
-                "vmla.f32   q12, q4, d2[0]      \n"
-                "vmla.f32   q13, q4, d2[1]      \n"
-                "vmla.f32   q14, q4, d3[0]      \n"
-                "vmla.f32   q15, q4, d3[1]      \n"
-                "bne        6b                  \n"
-
-                "7:                             \n"
-                "vshrn.u32  d16, q8, #16        \n"
-                "vshrn.u32  d17, q9, #16        \n"
-                "vshrn.u32  d18, q10, #16       \n"
-                "vshrn.u32  d19, q11, #16       \n"
-                "vshrn.u32  d20, q12, #16       \n"
-                "vshrn.u32  d21, q13, #16       \n"
-                "vshrn.u32  d22, q14, #16       \n"
-                "vshrn.u32  d23, q15, #16       \n"
-                "cmp        %11, #0             \n"
-                "beq        10f                 \n"
-
-                // if out_elempack == 4
-                "cmp        %12, #4             \n"
-                "bne        8f                  \n"
-
-                "vstm       %3!, {d16-d23}      \n"
-                "b          9f                  \n"
-
-                // if out_elempack == 1
-                "8:                             \n"
-                // transpose4x8
-                "vuzp.16    q8, q9              \n"
-                "vuzp.16    q10, q11            \n"
-                "vuzp.16    q8, q10             \n"
-                "vuzp.16    q9, q11             \n"
-
-                "add        r4, %3, %13, lsl #1 \n"
-                "vst1.u16   {d16-d17}, [%3 :64]! \n"
-                "vst1.u16   {d18-d19}, [r4 :64] \n"
-                "add        r4, r4, %13, lsl #1 \n"
-                "vst1.u16   {d20-d21}, [r4 :64] \n"
-                "add        r4, r4, %13, lsl #1 \n"
-                "vst1.u16   {d22-d23}, [r4 :64] \n"
-
-                "9:                             \n"
-                "add        %0, %0, #128        \n"
-                "b          11f                 \n"
-
-                "10:                            \n"
-                "vstm       %0!, {d16-d23}      \n"
-                "vstm       %0!, {d24-d31}      \n"
-
-                "11:                            \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-            float32x4_t _sum3;
-            float32x4_t _sum4;
-            float32x4_t _sum5;
-            float32x4_t _sum6;
-            float32x4_t _sum7;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1q_f32(pC);
-                    _sum1 = _sum0;
-                    _sum2 = _sum0;
-                    _sum3 = _sum0;
-                    _sum4 = _sum0;
-                    _sum5 = _sum0;
-                    _sum6 = _sum0;
-                    _sum7 = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                    _sum2 = vdupq_n_f32(0.f);
-                    _sum3 = vdupq_n_f32(0.f);
-                    _sum4 = vdupq_n_f32(0.f);
-                    _sum5 = vdupq_n_f32(0.f);
-                    _sum6 = vdupq_n_f32(0.f);
-                    _sum7 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4 * 1);
-                _sum2 = vld1q_f32(outptr + 4 * 2);
-                _sum3 = vld1q_f32(outptr + 4 * 3);
-                _sum4 = vld1q_f32(outptr + 4 * 4);
-                _sum5 = vld1q_f32(outptr + 4 * 5);
-                _sum6 = vld1q_f32(outptr + 4 * 6);
-                _sum7 = vld1q_f32(outptr + 4 * 7);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA = bfloat2float(vld1_u16(pA));
-                float32x4_t _pB0 = bfloat2float(vld1_u16(pB));
-                float32x4_t _pB1 = bfloat2float(vld1_u16(pB + 4));
-
-#if __aarch64__
-                _sum0 = vfmaq_laneq_f32(_sum0, _pA, _pB0, 0);
-                _sum1 = vfmaq_laneq_f32(_sum1, _pA, _pB0, 1);
-                _sum2 = vfmaq_laneq_f32(_sum2, _pA, _pB0, 2);
-                _sum3 = vfmaq_laneq_f32(_sum3, _pA, _pB0, 3);
-                _sum4 = vfmaq_laneq_f32(_sum4, _pA, _pB1, 0);
-                _sum5 = vfmaq_laneq_f32(_sum5, _pA, _pB1, 1);
-                _sum6 = vfmaq_laneq_f32(_sum6, _pA, _pB1, 2);
-                _sum7 = vfmaq_laneq_f32(_sum7, _pA, _pB1, 3);
-#else
-                _sum0 = vmlaq_lane_f32(_sum0, _pA, vget_low_f32(_pB0), 0);
-                _sum1 = vmlaq_lane_f32(_sum1, _pA, vget_low_f32(_pB0), 1);
-                _sum2 = vmlaq_lane_f32(_sum2, _pA, vget_high_f32(_pB0), 0);
-                _sum3 = vmlaq_lane_f32(_sum3, _pA, vget_high_f32(_pB0), 1);
-                _sum4 = vmlaq_lane_f32(_sum4, _pA, vget_low_f32(_pB1), 0);
-                _sum5 = vmlaq_lane_f32(_sum5, _pA, vget_low_f32(_pB1), 1);
-                _sum6 = vmlaq_lane_f32(_sum6, _pA, vget_high_f32(_pB1), 0);
-                _sum7 = vmlaq_lane_f32(_sum7, _pA, vget_high_f32(_pB1), 1);
-#endif
-
-                pA += 4;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum1));
-                    vst1_u16(outptr0 + 4 * 2, float2bfloat(_sum2));
-                    vst1_u16(outptr0 + 4 * 3, float2bfloat(_sum3));
-                    vst1_u16(outptr0 + 4 * 4, float2bfloat(_sum4));
-                    vst1_u16(outptr0 + 4 * 5, float2bfloat(_sum5));
-                    vst1_u16(outptr0 + 4 * 6, float2bfloat(_sum6));
-                    vst1_u16(outptr0 + 4 * 7, float2bfloat(_sum7));
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-                    uint16x4_t _t0 = float2bfloat(_sum0);
-                    uint16x4_t _t1 = float2bfloat(_sum1);
-                    uint16x4_t _t2 = float2bfloat(_sum2);
-                    uint16x4_t _t3 = float2bfloat(_sum3);
-                    uint16x4_t _t4 = float2bfloat(_sum4);
-                    uint16x4_t _t5 = float2bfloat(_sum5);
-                    uint16x4_t _t6 = float2bfloat(_sum6);
-                    uint16x4_t _t7 = float2bfloat(_sum7);
-                    transpose4x8_u16(_t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7);
-
-                    vst1_u16(outptr0, _t0);
-                    vst1_u16(outptr0 + 4, _t1);
-                    vst1_u16(outptr0 + out_hstep, _t2);
-                    vst1_u16(outptr0 + out_hstep + 4, _t3);
-                    vst1_u16(outptr0 + out_hstep * 2, _t4);
-                    vst1_u16(outptr0 + out_hstep * 2 + 4, _t5);
-                    vst1_u16(outptr0 + out_hstep * 3, _t6);
-                    vst1_u16(outptr0 + out_hstep * 3 + 4, _t7);
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 4 * 2, _sum2);
-                vst1q_f32(outptr + 4 * 3, _sum3);
-                vst1q_f32(outptr + 4 * 4, _sum4);
-                vst1q_f32(outptr + 4 * 5, _sum5);
-                vst1q_f32(outptr + 4 * 6, _sum6);
-                vst1q_f32(outptr + 4 * 7, _sum7);
-            }
-
-            outptr += 32;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            const unsigned short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                "b      3f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "ld1    {v28.4s}, [%8]              \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-
-                "2:                                 \n"
-                "mov    v29.16b, v28.16b            \n"
-                "mov    v30.16b, v28.16b            \n"
-                "mov    v31.16b, v28.16b            \n"
-
-                "3:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                "4:                                 \n"
-                "prfm   pldl1keep, [%2, #256]       \n"
-                "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n"
-                "shll   v0.4s, v0.4h, #16           \n"
-                "shll   v1.4s, v1.4h, #16           \n"
-                "shll   v2.4s, v2.4h, #16           \n"
-                "shll   v3.4s, v3.4h, #16           \n"
-                "prfm   pldl1keep, [%1, #256]       \n"
-                "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%1], #32 \n"
-                "shll   v16.4s, v16.4h, #16         \n"
-                "shll   v17.4s, v17.4h, #16         \n"
-                "shll   v18.4s, v18.4h, #16         \n"
-                "shll   v19.4s, v19.4h, #16         \n"
-                "fmla   v28.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v29.4s, v16.4s, v0.s[1]     \n"
-                "fmla   v30.4s, v16.4s, v0.s[2]     \n"
-                "fmla   v31.4s, v16.4s, v0.s[3]     \n"
-                "fmla   v28.4s, v17.4s, v1.s[0]     \n"
-                "fmla   v29.4s, v17.4s, v1.s[1]     \n"
-                "fmla   v30.4s, v17.4s, v1.s[2]     \n"
-                "fmla   v31.4s, v17.4s, v1.s[3]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v28.4s, v18.4s, v2.s[0]     \n"
-                "fmla   v29.4s, v18.4s, v2.s[1]     \n"
-                "fmla   v30.4s, v18.4s, v2.s[2]     \n"
-                "fmla   v31.4s, v18.4s, v2.s[3]     \n"
-                "fmla   v28.4s, v19.4s, v3.s[0]     \n"
-                "fmla   v29.4s, v19.4s, v3.s[1]     \n"
-                "fmla   v30.4s, v19.4s, v3.s[2]     \n"
-                "fmla   v31.4s, v19.4s, v3.s[3]     \n"
-                "bne    4b                          \n"
-
-                "5:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    7f                          \n"
-
-                "6:                                 \n"
-                "ld1    {v0.4h}, [%2], #8           \n"
-                "shll   v0.4s, v0.4h, #16           \n"
-                "ld1    {v16.4h}, [%1], #8          \n"
-                "shll   v16.4s, v16.4h, #16         \n"
-                "fmla   v28.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v29.4s, v16.4s, v0.s[1]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v30.4s, v16.4s, v0.s[2]     \n"
-                "fmla   v31.4s, v16.4s, v0.s[3]     \n"
-                "bne    6b                          \n"
-
-                "7:                                 \n"
-                "shrn   v0.4h, v28.4s, #16          \n"
-                "shrn2  v0.8h, v29.4s, #16          \n"
-                "shrn   v1.4h, v30.4s, #16          \n"
-                "shrn2  v1.8h, v31.4s, #16          \n"
-                "tst    %w11, #255                  \n"
-                "beq    10f                         \n"
-
-                // if out_elempack == 4
-                "cmp    %w12, #4                    \n"
-                "bne    8f                          \n"
-
-                "st1    {v0.8h, v1.8h}, [%3], #32   \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // transpose4x4
-                "uzp1   v20.8h, v0.8h, v1.8h        \n"
-                "uzp2   v21.8h, v0.8h, v1.8h        \n"
-
-                "uzp1   v0.8h, v20.8h, v21.8h       \n"
-                "uzp2   v1.8h, v20.8h, v21.8h       \n"
-
-                "add    x4, %3, %w13, sxtw 1        \n"
-                "st1    {v0.d}[0], [%3], #8         \n"
-                "st1    {v0.d}[1], [x4]             \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v1.d}[0], [x4]             \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v1.d}[1], [x4]             \n"
-
-                "9:                                 \n"
-                "add    %0, %0, #64                 \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19", "v20", "v21", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-            asm volatile(
-                "cmp        %10, #0             \n"
-                "beq        0f                  \n"
-
-                "vldm       %0, {d24-d31}       \n"
-                "b          3f                  \n"
-
-                "0:                             \n"
-                // if pC
-                "cmp        %8, #0              \n"
-                "beq        1f                  \n"
-
-                "vld1.f32   {d24-d25}, [%8]     \n"
-                "b          2f                  \n"
-
-                // else
-                "1:                             \n"
-                "veor       q12, q12            \n"
-
-                "2:                             \n"
-                "vmov       q13, q12            \n"
-                "vmov       q14, q12            \n"
-                "vmov       q15, q12            \n"
-
-                "3:                             \n"
-                "lsr        r4, %9, #2          \n" // r4 = max_kk >> 2
-                "cmp        r4, #0              \n"
-                "beq        5f                  \n"
-
-                "4:                             \n"
-                "pld        [%2, #256]          \n"
-                "vld1.u16   {d4-d7}, [%2 :64]!  \n"
-                "pld        [%1, #256]          \n"
-                "vld1.u16   {d12-d15}, [%1 :64]! \n"
-                "vshll.u16  q0, d4, #16         \n"
-                "vshll.u16  q1, d5, #16         \n"
-                "vshll.u16  q2, d6, #16         \n"
-                "vshll.u16  q3, d7, #16         \n"
-                "vshll.u16  q4, d12, #16        \n"
-                "vshll.u16  q5, d13, #16        \n"
-                "vshll.u16  q6, d14, #16        \n"
-                "vshll.u16  q7, d15, #16        \n"
-                "vmla.f32   q12, q4, d0[0]      \n"
-                "vmla.f32   q13, q4, d0[1]      \n"
-                "vmla.f32   q14, q4, d1[0]      \n"
-                "vmla.f32   q15, q4, d1[1]      \n"
-                "vmla.f32   q12, q5, d2[0]      \n"
-                "vmla.f32   q13, q5, d2[1]      \n"
-                "vmla.f32   q14, q5, d3[0]      \n"
-                "vmla.f32   q15, q5, d3[1]      \n"
-                "subs       r4, r4, #1          \n"
-                "vmla.f32   q12, q6, d4[0]      \n"
-                "vmla.f32   q13, q6, d4[1]      \n"
-                "vmla.f32   q14, q6, d5[0]      \n"
-                "vmla.f32   q15, q6, d5[1]      \n"
-                "vmla.f32   q12, q7, d6[0]      \n"
-                "vmla.f32   q13, q7, d6[1]      \n"
-                "vmla.f32   q14, q7, d7[0]      \n"
-                "vmla.f32   q15, q7, d7[1]      \n"
-                "bne        4b                  \n"
-
-                "5:                             \n"
-                "and        r4, %9, #3          \n" // r4 = remain = max_kk & 3
-                "cmp        r4, #0              \n"
-                "beq        7f                  \n"
-
-                "6:                             \n"
-                "vld1.u16   {d0}, [%2 :64]!     \n"
-                "vshll.u16  q0, d0, #16         \n"
-                "vld1.u16   {d8}, [%1 :64]!     \n"
-                "vshll.u16  q4, d8, #16         \n"
-                "subs       r4, r4, #1          \n"
-                "vmla.f32   q12, q4, d0[0]      \n"
-                "vmla.f32   q13, q4, d0[1]      \n"
-                "vmla.f32   q14, q4, d1[0]      \n"
-                "vmla.f32   q15, q4, d1[1]      \n"
-                "bne        6b                  \n"
-
-                "7:                             \n"
-                "vshrn.u32  d24, q12, #16       \n"
-                "vshrn.u32  d25, q13, #16       \n"
-                "vshrn.u32  d26, q14, #16       \n"
-                "vshrn.u32  d27, q15, #16       \n"
-                "cmp        %11, #0             \n"
-                "beq        10f                 \n"
-
-                // if out_elempack == 4
-                "cmp        %12, #4             \n"
-                "bne        8f                  \n"
-
-                "vst1.u16   {d24-d27}, [%3]!    \n"
-                "b          9f                  \n"
-
-                // if out_elempack == 1
-                "8:                             \n"
-                // transpose4x4
-                "vuzp.16    q12, q13            \n"
-                "vuzp.16    q12, q13            \n"
-
-                "add        r4, %3, %13, lsl #1 \n"
-                "vst1.u16   {d24}, [%3]!        \n"
-                "vst1.u16   {d25}, [r4]         \n"
-                "add        r4, r4, %13, lsl #1 \n"
-                "vst1.u16   {d26}, [r4]         \n"
-                "add        r4, r4, %13, lsl #1 \n"
-                "vst1.u16   {d27}, [r4]         \n"
-
-                "9:                             \n"
-                "add        %0, %0, #64         \n"
-                "b          11f                 \n"
-
-                "10:                            \n"
-                "vstm       %0!, {d24-d31}      \n"
-
-                "11:                            \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-            float32x4_t _sum3;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1q_f32(pC);
-                    _sum1 = _sum0;
-                    _sum2 = _sum0;
-                    _sum3 = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                    _sum2 = vdupq_n_f32(0.f);
-                    _sum3 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4 * 1);
-                _sum2 = vld1q_f32(outptr + 4 * 2);
-                _sum3 = vld1q_f32(outptr + 4 * 3);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA = bfloat2float(vld1_u16(pA));
-                float32x4_t _pB = bfloat2float(vld1_u16(pB));
-
-#if __aarch64__
-                _sum0 = vfmaq_laneq_f32(_sum0, _pA, _pB, 0);
-                _sum1 = vfmaq_laneq_f32(_sum1, _pA, _pB, 1);
-                _sum2 = vfmaq_laneq_f32(_sum2, _pA, _pB, 2);
-                _sum3 = vfmaq_laneq_f32(_sum3, _pA, _pB, 3);
-#else
-                _sum0 = vmlaq_lane_f32(_sum0, _pA, vget_low_f32(_pB), 0);
-                _sum1 = vmlaq_lane_f32(_sum1, _pA, vget_low_f32(_pB), 1);
-                _sum2 = vmlaq_lane_f32(_sum2, _pA, vget_high_f32(_pB), 0);
-                _sum3 = vmlaq_lane_f32(_sum3, _pA, vget_high_f32(_pB), 1);
-#endif
-
-                pA += 4;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum1));
-                    vst1_u16(outptr0 + 4 * 2, float2bfloat(_sum2));
-                    vst1_u16(outptr0 + 4 * 3, float2bfloat(_sum3));
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    uint16x4_t _t0 = float2bfloat(_sum0);
-                    uint16x4_t _t1 = float2bfloat(_sum1);
-                    uint16x4_t _t2 = float2bfloat(_sum2);
-                    uint16x4_t _t3 = float2bfloat(_sum3);
-                    transpose4x4_u16(_t0, _t1, _t2, _t3);
-
-                    vst1_u16(outptr0, _t0);
-                    vst1_u16(outptr0 + out_hstep * 1, _t1);
-                    vst1_u16(outptr0 + out_hstep * 2, _t2);
-                    vst1_u16(outptr0 + out_hstep * 3, _t3);
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 4 * 2, _sum2);
-                vst1q_f32(outptr + 4 * 3, _sum3);
-            }
-
-            outptr += 16;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            const unsigned short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v30.4s, v31.4s}, [%0]      \n"
-                "b      3f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "ld1    {v30.4s}, [%8]              \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v30.16b, v30.16b, v30.16b   \n"
-
-                "2:                                 \n"
-                "mov    v31.16b, v30.16b            \n"
-
-                "3:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-                "eor    v29.16b, v29.16b, v29.16b   \n"
-                "4:                                 \n"
-                "prfm   pldl1keep, [%2, #128]       \n"
-                "ld1    {v0.4h, v1.4h}, [%2], #16   \n"
-                "shll   v0.4s, v0.4h, #16           \n"
-                "shll   v1.4s, v1.4h, #16           \n"
-                "prfm   pldl1keep, [%1, #256]       \n"
-                "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%1], #32 \n"
-                "shll   v16.4s, v16.4h, #16         \n"
-                "shll   v17.4s, v17.4h, #16         \n"
-                "shll   v18.4s, v18.4h, #16         \n"
-                "shll   v19.4s, v19.4h, #16         \n"
-                "fmla   v28.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v29.4s, v16.4s, v0.s[1]     \n"
-                "fmla   v30.4s, v17.4s, v0.s[2]     \n"
-                "fmla   v31.4s, v17.4s, v0.s[3]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v28.4s, v18.4s, v1.s[0]     \n"
-                "fmla   v29.4s, v18.4s, v1.s[1]     \n"
-                "fmla   v30.4s, v19.4s, v1.s[2]     \n"
-                "fmla   v31.4s, v19.4s, v1.s[3]     \n"
-                "bne    4b                          \n"
-                "fadd   v30.4s, v30.4s, v28.4s      \n"
-                "fadd   v31.4s, v31.4s, v29.4s      \n"
-
-                "5:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    7f                          \n"
-
-                "6:                                 \n"
-                "ld1    {v0.s}[0], [%2], #4         \n"
-                "shll   v0.4s, v0.4h, #16           \n"
-                "ld1    {v16.4h}, [%1], #8          \n"
-                "shll   v16.4s, v16.4h, #16         \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v30.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v31.4s, v16.4s, v0.s[1]     \n"
-                "bne    6b                          \n"
-
-                "7:                                 \n"
-                "shrn   v0.4h, v30.4s, #16          \n"
-                "shrn   v1.4h, v31.4s, #16          \n"
-                "tst    %w11, #255                  \n"
-                "beq    10f                         \n"
-
-                // if out_elempack == 4
-                "cmp    %w12, #4                    \n"
-                "bne    8f                          \n"
-
-                "st1    {v0.4h, v1.4h}, [%3], #16   \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // transpose4x2
-                "zip1   v30.4h, v0.4h, v1.4h        \n"
-                "zip2   v31.4h, v0.4h, v1.4h        \n"
-
-                "add    x4, %3, %w13, sxtw 1        \n"
-                "st1    {v30.s}[0], [%3], #4        \n"
-                "st1    {v30.s}[1], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v31.s}[0], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v31.s}[1], [x4]            \n"
-
-                "9:                                 \n"
-                "add    %0, %0, #32                 \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v30.4s, v31.4s}, [%0], #32 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v1", "v16", "v17", "v18", "v19", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-            asm volatile(
-                "cmp        %10, #0             \n"
-                "beq        0f                  \n"
-
-                "vld1.f32   {d28-d31}, [%0 :128] \n"
-                "b          3f                  \n"
-
-                "0:                             \n"
-                // if pC
-                "cmp        %8, #0              \n"
-                "beq        1f                  \n"
-
-                "vld1.f32   {d28-d29}, [%8]     \n"
-                "b          2f                  \n"
-
-                // else
-                "1:                             \n"
-                "veor       q14, q14            \n"
-
-                "2:                             \n"
-                "vmov       q15, q14            \n"
-
-                "3:                             \n"
-                "lsr        r4, %9, #2          \n" // r4 = max_kk >> 2
-                "cmp        r4, #0              \n"
-                "beq        5f                  \n"
-
-                "veor       q12, q12            \n"
-                "veor       q13, q13            \n"
-                "4:                             \n"
-                "pld        [%2, #128]          \n"
-                "vld1.u16   {d2-d3}, [%2 :64]!  \n"
-                "pld        [%1, #256]          \n"
-                "vld1.u16   {d12-d15}, [%1 :64]! \n"
-                "vshll.u16  q0, d2, #16         \n"
-                "vshll.u16  q1, d3, #16         \n"
-                "vshll.u16  q4, d12, #16        \n"
-                "vshll.u16  q5, d13, #16        \n"
-                "vshll.u16  q6, d14, #16        \n"
-                "vshll.u16  q7, d15, #16        \n"
-                "vmla.f32   q12, q4, d0[0]      \n"
-                "vmla.f32   q13, q4, d0[1]      \n"
-                "vmla.f32   q14, q5, d1[0]      \n"
-                "vmla.f32   q15, q5, d1[1]      \n"
-                "subs       r4, r4, #1          \n"
-                "vmla.f32   q12, q6, d2[0]      \n"
-                "vmla.f32   q13, q6, d2[1]      \n"
-                "vmla.f32   q14, q7, d3[0]      \n"
-                "vmla.f32   q15, q7, d3[1]      \n"
-                "bne        4b                  \n"
-                "vadd.f32   q14, q14, q12       \n"
-                "vadd.f32   q15, q15, q13       \n"
-
-                "5:                             \n"
-                "and        r4, %9, #3          \n" // r4 = remain = max_kk & 3
-                "cmp        r4, #0              \n"
-                "beq        7f                  \n"
-
-                "6:                             \n"
-                "vld1.u32   {d0[0]}, [%2]!      \n"
-                "vshll.u16  q0, d0, #16         \n"
-                "vld1.u16   {d8}, [%1 :64]!     \n"
-                "vshll.u16  q4, d8, #16         \n"
-                "subs       r4, r4, #1          \n"
-                "vmla.f32   q14, q4, d0[0]      \n"
-                "vmla.f32   q15, q4, d0[1]      \n"
-                "bne        6b                  \n"
-
-                "7:                             \n"
-                "vshrn.u32  d28, q14, #16       \n"
-                "vshrn.u32  d29, q15, #16       \n"
-                "cmp        %11, #0             \n"
-                "beq        10f                 \n"
-
-                // if out_elempack == 4
-                "cmp        %12, #4             \n"
-                "bne        8f                  \n"
-
-                "vst1.u16   {d28-d29}, [%3]!    \n"
-                "b          9f                  \n"
-
-                // if out_elempack == 1
-                "8:                             \n"
-                // transpose4x2
-                "vzip.16    d28, d29            \n"
-
-                "add        r4, %3, %13, lsl #1 \n"
-                "vst1.u32   {d28[0]}, [%3]!     \n"
-                "vst1.u32   {d28[1]}, [r4]      \n"
-                "add        r4, r4, %13, lsl #1 \n"
-                "vst1.u32   {d29[0]}, [r4]      \n"
-                "add        r4, r4, %13, lsl #1 \n"
-                "vst1.u32   {d29[1]}, [r4]      \n"
-
-                "9:                             \n"
-                "add        %0, %0, #32         \n"
-                "b          11f                 \n"
-
-                "10:                            \n"
-                "vst1.f32   {d28-d31}, [%0 :128]! \n"
-
-                "11:                            \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "r4", "q0", "q1", "q4", "q5", "q6", "q7", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1q_f32(pC);
-                    _sum1 = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA = bfloat2float(vld1_u16(pA));
-                float32x2_t _pB = vget_low_f32(bfloat2float(vld1_u16(pB)));
-
-#if __aarch64__
-                _sum0 = vfmaq_lane_f32(_sum0, _pA, _pB, 0);
-                _sum1 = vfmaq_lane_f32(_sum1, _pA, _pB, 1);
-#else
-                _sum0 = vmlaq_lane_f32(_sum0, _pA, _pB, 0);
-                _sum1 = vmlaq_lane_f32(_sum1, _pA, _pB, 1);
-#endif
-
-                pA += 4;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum1));
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short sum0[4];
-                    unsigned short sum1[4];
-                    vst1_u16(sum0, float2bfloat(_sum0));
-                    vst1_u16(sum1, float2bfloat(_sum1));
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[1] = sum1[0];
-                    outptr0[out_hstep + 1] = sum1[1];
-                    outptr0[out_hstep * 2 + 1] = sum1[2];
-                    outptr0[out_hstep * 3 + 1] = sum1[3];
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-            }
-
-            outptr += 8;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            const unsigned short* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v31.4s}, [%0]              \n"
-                "b      2f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "ld1    {v31.4s}, [%8]              \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                "2:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    4f                          \n"
-
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-                "eor    v29.16b, v29.16b, v29.16b   \n"
-                "eor    v30.16b, v30.16b, v30.16b   \n"
-                "3:                                 \n"
-                "prfm   pldl1keep, [%2, #64]        \n"
-                "ld1    {v0.4h}, [%2], #8           \n"
-                "shll   v0.4s, v0.4h, #16           \n"
-                "prfm   pldl1keep, [%1, #256]       \n"
-                "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%1], #32 \n"
-                "shll   v16.4s, v16.4h, #16         \n"
-                "shll   v17.4s, v17.4h, #16         \n"
-                "shll   v18.4s, v18.4h, #16         \n"
-                "shll   v19.4s, v19.4h, #16         \n"
-                "fmla   v28.4s, v16.4s, v0.s[0]     \n"
-                "fmla   v29.4s, v17.4s, v0.s[1]     \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v30.4s, v18.4s, v0.s[2]     \n"
-                "fmla   v31.4s, v19.4s, v0.s[3]     \n"
-                "bne    3b                          \n"
-                "fadd   v30.4s, v30.4s, v28.4s      \n"
-                "fadd   v31.4s, v31.4s, v29.4s      \n"
-                "fadd   v31.4s, v31.4s, v30.4s      \n"
-
-                "4:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    6f                          \n"
-
-                "5:                                 \n"
-                "ld1r   {v0.4h}, [%2], #2           \n"
-                "shll   v0.4s, v0.4h, #16           \n"
-                "ld1    {v16.4h}, [%1], #8          \n"
-                "shll   v16.4s, v16.4h, #16         \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v31.4s, v16.4s, v0.4s       \n"
-                "bne    5b                          \n"
-
-                "6:                                 \n"
-                "shrn   v0.4h, v31.4s, #16          \n"
-                "tst    %w11, #255                  \n"
-                "beq    9f                          \n"
-
-                // if out_elempack == 4
-                "cmp    %w12, #4                    \n"
-                "bne    7f                          \n"
-
-                "st1    {v0.4h}, [%3], #8           \n"
-                "b      8f                          \n"
-
-                // if out_elempack == 1
-                "7:                                 \n"
-                "add    x4, %3, %w13, sxtw 1        \n"
-                "st1    {v0.h}[0], [%3], #2         \n"
-                "st1    {v0.h}[1], [x4]             \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v0.h}[2], [x4]             \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v0.h}[3], [x4]             \n"
-
-                "8:                                 \n"
-                "add    %0, %0, #16                 \n"
-                "b      10f                         \n"
-
-                "9:                                 \n"
-                "st1    {v31.4s}, [%0], #16         \n"
-
-                "10:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v16", "v17", "v18", "v19", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-            asm volatile(
-                "cmp        %10, #0             \n"
-                "beq        0f                  \n"
-
-                "vld1.f32   {d30-d31}, [%0 :64] \n"
-                "b          2f                  \n"
-
-                "0:                             \n"
-                // if pC
-                "cmp        %8, #0              \n"
-                "beq        1f                  \n"
-
-                "vld1.f32   {d30-d31}, [%8]     \n"
-                "b          2f                  \n"
-
-                // else
-                "1:                             \n"
-                "veor       q15, q15            \n"
-
-                "2:                             \n"
-                "lsr        r4, %9, #2          \n" // r4 = max_kk >> 2
-                "cmp        r4, #0              \n"
-                "beq        4f                  \n"
-
-                "veor       q12, q12            \n"
-                "veor       q13, q13            \n"
-                "veor       q14, q14            \n"
-                "3:                             \n"
-                "pld        [%2, #64]           \n"
-                "vld1.u16   {d1}, [%2]!         \n"
-                "pld        [%1, #256]          \n"
-                "vld1.u16   {d12-d15}, [%1 :64]! \n"
-                "vshll.u16  q0, d1, #16         \n"
-                "vshll.u16  q4, d12, #16        \n"
-                "vshll.u16  q5, d13, #16        \n"
-                "vshll.u16  q6, d14, #16        \n"
-                "vshll.u16  q7, d15, #16        \n"
-                "vmla.f32   q12, q4, d0[0]      \n"
-                "vmla.f32   q13, q5, d0[1]      \n"
-                "vmla.f32   q14, q6, d1[0]      \n"
-                "vmla.f32   q15, q7, d1[1]      \n"
-                "subs       r4, r4, #1          \n"
-                "bne        3b                  \n"
-                "vadd.f32   q14, q14, q12       \n"
-                "vadd.f32   q15, q15, q13       \n"
-                "vadd.f32   q15, q15, q14       \n"
-
-                "4:                             \n"
-                "and        r4, %9, #3          \n" // r4 = remain = max_kk & 3
-                "cmp        r4, #0              \n"
-                "beq        6f                  \n"
-
-                "5:                             \n"
-                "vld1.u16   {d0[]}, [%2]!       \n"
-                "vshll.u16  q0, d0, #16         \n"
-                "vld1.u16   {d8}, [%1 :64]!     \n"
-                "vshll.u16  q4, d8, #16         \n"
-                "subs       r4, r4, #1          \n"
-                "vmla.f32   q15, q4, q0         \n"
-                "bne        5b                  \n"
-
-                "6:                             \n"
-                "vshrn.u32  d30, q15, #16       \n"
-                "cmp        %11, #0             \n"
-                "beq        9f                  \n"
-
-                // if out_elempack == 4
-                "cmp        %12, #4             \n"
-                "bne        7f                  \n"
-
-                "vst1.u16   {d30}, [%3]!        \n"
-                "b          8f                  \n"
-
-                // if out_elempack == 1
-                "7:                             \n"
-
-                "add        r4, %3, %13, lsl #1 \n"
-                "vst1.u16   {d30[0]}, [%3]!     \n"
-                "vst1.u16   {d30[1]}, [r4]      \n"
-                "add        r4, r4, %13, lsl #1 \n"
-                "vst1.u16   {d30[2]}, [r4]      \n"
-                "add        r4, r4, %13, lsl #1 \n"
-                "vst1.u16   {d30[3]}, [r4]      \n"
-
-                "8:                             \n"
-                "add        %0, %0, #16         \n"
-                "b          10f                 \n"
-
-                "9:                             \n"
-                "vst1.f32   {d30-d31}, [%0 :64]! \n"
-
-                "10:                            \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "r4", "q0", "q4", "q5", "q6", "q7", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            float32x4_t _sum0;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1q_f32(pC);
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA = bfloat2float(vld1_u16(pA));
-                float32x4_t _pB = bfloat2float(vdup_n_u16(pB[0]));
-
-#if __aarch64__
-                _sum0 = vfmaq_f32(_sum0, _pA, _pB);
-#else
-                _sum0 = vmlaq_f32(_sum0, _pA, _pB);
-#endif
-
-                pA += 4;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    outptr0 += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short sum0[4];
-                    vst1_u16(sum0, float2bfloat(_sum0));
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0++;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-            }
-
-            outptr += 4;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-
-        pAT += max_kk * 4;
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        unsigned short* outptr0 = (unsigned short*)top_blob + (i + ii) * out_hstep + j;
-
-        const unsigned short* pB = pBT;
-
-        if (pC)
-        {
-            pC = (const float*)CT_tile + i + ii;
-        }
-
-        int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum02;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum12;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vdupq_n_f32(pC[0]);
-                    _sum01 = vdupq_n_f32(pC[0]);
-                    _sum02 = vdupq_n_f32(pC[0]);
-                    _sum10 = vdupq_n_f32(pC[1]);
-                    _sum11 = vdupq_n_f32(pC[1]);
-                    _sum12 = vdupq_n_f32(pC[1]);
-                }
-                else
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum02 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                    _sum12 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01 = vld2q_f32(outptr);
-                float32x4x2_t _tmp23 = vld2q_f32(outptr + 8);
-                float32x4x2_t _tmp45 = vld2q_f32(outptr + 16);
-                _sum00 = _tmp01.val[0];
-                _sum01 = _tmp23.val[0];
-                _sum02 = _tmp45.val[0];
-                _sum10 = _tmp01.val[1];
-                _sum11 = _tmp23.val[1];
-                _sum12 = _tmp45.val[1];
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pB0 = bfloat2float(vld1_u16(pB));
-                float32x4_t _pB1 = bfloat2float(vld1_u16(pB + 4));
-                float32x4_t _pB2 = bfloat2float(vld1_u16(pB + 8));
-
-                float32x2_t _pA = vget_low_f32(bfloat2float(vld1_u16(pA)));
-
-                _sum00 = vfmaq_lane_f32(_sum00, _pB0, _pA, 0);
-                _sum01 = vfmaq_lane_f32(_sum01, _pB1, _pA, 0);
-                _sum02 = vfmaq_lane_f32(_sum02, _pB2, _pA, 0);
-                _sum10 = vfmaq_lane_f32(_sum10, _pB0, _pA, 1);
-                _sum11 = vfmaq_lane_f32(_sum11, _pB1, _pA, 1);
-                _sum12 = vfmaq_lane_f32(_sum12, _pB2, _pA, 1);
-
-                pA += 2;
-                pB += 12;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum01));
-                    vst1_u16(outptr0 + 8, float2bfloat(_sum02));
-                    vst1_u16(outptr0 + out_hstep, float2bfloat(_sum10));
-                    vst1_u16(outptr0 + out_hstep + 4, float2bfloat(_sum11));
-                    vst1_u16(outptr0 + out_hstep + 8, float2bfloat(_sum12));
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01;
-                _tmp01.val[0] = _sum00;
-                _tmp01.val[1] = _sum10;
-                float32x4x2_t _tmp23;
-                _tmp23.val[0] = _sum01;
-                _tmp23.val[1] = _sum11;
-                float32x4x2_t _tmp45;
-                _tmp45.val[0] = _sum02;
-                _tmp45.val[1] = _sum12;
-                vst2q_f32(outptr, _tmp01);
-                vst2q_f32(outptr + 8, _tmp23);
-                vst2q_f32(outptr + 16, _tmp45);
-            }
-
-            outptr += 24;
-        }
-#endif // __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vdupq_n_f32(pC[0]);
-                    _sum01 = vdupq_n_f32(pC[0]);
-                    _sum10 = vdupq_n_f32(pC[1]);
-                    _sum11 = vdupq_n_f32(pC[1]);
-                }
-                else
-                {
-                    _sum00 = vdupq_n_f32(0.f);
-                    _sum01 = vdupq_n_f32(0.f);
-                    _sum10 = vdupq_n_f32(0.f);
-                    _sum11 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01 = vld2q_f32(outptr);
-                float32x4x2_t _tmp23 = vld2q_f32(outptr + 8);
-                _sum00 = _tmp01.val[0];
-                _sum01 = _tmp23.val[0];
-                _sum10 = _tmp01.val[1];
-                _sum11 = _tmp23.val[1];
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pB0 = bfloat2float(vld1_u16(pB));
-                float32x4_t _pB1 = bfloat2float(vld1_u16(pB + 4));
-
-                float32x2_t _pA = vget_low_f32(bfloat2float(vld1_u16(pA)));
-#if __aarch64__
-                _sum00 = vfmaq_lane_f32(_sum00, _pB0, _pA, 0);
-                _sum01 = vfmaq_lane_f32(_sum01, _pB1, _pA, 0);
-                _sum10 = vfmaq_lane_f32(_sum10, _pB0, _pA, 1);
-                _sum11 = vfmaq_lane_f32(_sum11, _pB1, _pA, 1);
-#else
-                _sum00 = vmlaq_lane_f32(_sum00, _pB0, _pA, 0);
-                _sum01 = vmlaq_lane_f32(_sum01, _pB1, _pA, 0);
-                _sum10 = vmlaq_lane_f32(_sum10, _pB0, _pA, 1);
-                _sum11 = vmlaq_lane_f32(_sum11, _pB1, _pA, 1);
-#endif
-
-                pA += 2;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum01));
-                    vst1_u16(outptr0 + out_hstep, float2bfloat(_sum10));
-                    vst1_u16(outptr0 + out_hstep + 4, float2bfloat(_sum11));
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01;
-                _tmp01.val[0] = _sum00;
-                _tmp01.val[1] = _sum10;
-                float32x4x2_t _tmp23;
-                _tmp23.val[0] = _sum01;
-                _tmp23.val[1] = _sum11;
-                vst2q_f32(outptr, _tmp01);
-                vst2q_f32(outptr + 8, _tmp23);
-            }
-
-            outptr += 16;
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vdupq_n_f32(pC[0]);
-                    _sum1 = vdupq_n_f32(pC[1]);
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01 = vld2q_f32(outptr);
-                _sum0 = _tmp01.val[0];
-                _sum1 = _tmp01.val[1];
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pB = bfloat2float(vld1_u16(pB));
-
-                float32x2_t _pA = vget_low_f32(bfloat2float(vld1_u16(pA)));
-#if __aarch64__
-                _sum0 = vfmaq_lane_f32(_sum0, _pB, _pA, 0);
-                _sum1 = vfmaq_lane_f32(_sum1, _pB, _pA, 1);
-#else
-                _sum0 = vmlaq_lane_f32(_sum0, _pB, _pA, 0);
-                _sum1 = vmlaq_lane_f32(_sum1, _pB, _pA, 1);
-#endif
-
-                pA += 2;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + out_hstep, float2bfloat(_sum1));
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01;
-                _tmp01.val[0] = _sum0;
-                _tmp01.val[1] = _sum1;
-                vst2q_f32(outptr, _tmp01);
-            }
-
-            outptr += 8;
-        }
-#endif // __ARM_NEON
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            float sum00;
-            float sum01;
-            float sum10;
-            float sum11;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    sum00 = pC[0];
-                    sum01 = pC[1];
-                    sum10 = pC[0];
-                    sum11 = pC[1];
-                }
-                else
-                {
-                    sum00 = 0.f;
-                    sum01 = 0.f;
-                    sum10 = 0.f;
-                    sum11 = 0.f;
-                }
-            }
-            else
-            {
-                sum00 = outptr[0];
-                sum01 = outptr[1];
-                sum10 = outptr[2];
-                sum11 = outptr[3];
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                sum00 += bfloat16_to_float32(pA[0]) * bfloat16_to_float32(pB[0]);
-                sum01 += bfloat16_to_float32(pA[1]) * bfloat16_to_float32(pB[0]);
-                sum10 += bfloat16_to_float32(pA[0]) * bfloat16_to_float32(pB[1]);
-                sum11 += bfloat16_to_float32(pA[1]) * bfloat16_to_float32(pB[1]);
-
-                pA += 2;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = float32_to_bfloat16(sum00);
-                    outptr0[1] = float32_to_bfloat16(sum10);
-                    outptr0[out_hstep] = float32_to_bfloat16(sum01);
-                    outptr0[out_hstep + 1] = float32_to_bfloat16(sum11);
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                outptr[0] = sum00;
-                outptr[1] = sum01;
-                outptr[2] = sum10;
-                outptr[3] = sum11;
-            }
-
-            outptr += 4;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            float sum0;
-            float sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    sum0 = pC[0];
-                    sum1 = pC[1];
-                }
-                else
-                {
-                    sum0 = 0.f;
-                    sum1 = 0.f;
-                }
-            }
-            else
-            {
-                sum0 = outptr[0];
-                sum1 = outptr[1];
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                sum0 += bfloat16_to_float32(pA[0]) * bfloat16_to_float32(pB[0]);
-                sum1 += bfloat16_to_float32(pA[1]) * bfloat16_to_float32(pB[0]);
-                pA += 2;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = float32_to_bfloat16(sum0);
-                    outptr0[out_hstep] = float32_to_bfloat16(sum1);
-                    outptr0++;
-                }
-            }
-            else
-            {
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-            }
-
-            outptr += 2;
-        }
-
-        pAT += max_kk * 2;
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        unsigned short* outptr0 = (unsigned short*)top_blob + (i + ii) * out_hstep + j;
-
-        const unsigned short* pB = pBT;
-
-        if (pC)
-        {
-            pC = (const float*)CT_tile + i + ii;
-        }
-
-        int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vdupq_n_f32(pC[0]);
-                    _sum1 = vdupq_n_f32(pC[0]);
-                    _sum2 = vdupq_n_f32(pC[0]);
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                    _sum2 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4);
-                _sum2 = vld1q_f32(outptr + 8);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pB0 = bfloat2float(vld1_u16(pB));
-                float32x4_t _pB1 = bfloat2float(vld1_u16(pB + 4));
-                float32x4_t _pB2 = bfloat2float(vld1_u16(pB + 8));
-
-                float32x4_t _pA0 = bfloat2float(vdup_n_u16(pA[0]));
-
-                _sum0 = vfmaq_f32(_sum0, _pA0, _pB0);
-                _sum1 = vfmaq_f32(_sum1, _pA0, _pB1);
-                _sum2 = vfmaq_f32(_sum2, _pA0, _pB2);
-
-                pA += 1;
-                pB += 12;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum1));
-                    vst1_u16(outptr0 + 8, float2bfloat(_sum2));
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 8, _sum2);
-            }
-
-            outptr += 12;
-        }
-#endif // __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vdupq_n_f32(pC[0]);
-                    _sum1 = vdupq_n_f32(pC[0]);
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f32(0.f);
-                    _sum1 = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pB0 = bfloat2float(vld1_u16(pB));
-                float32x4_t _pB1 = bfloat2float(vld1_u16(pB + 4));
-
-                float32x4_t _pA0 = bfloat2float(vdup_n_u16(pA[0]));
-#if __aarch64__
-                _sum0 = vfmaq_f32(_sum0, _pA0, _pB0);
-                _sum1 = vfmaq_f32(_sum1, _pA0, _pB1);
-#else
-                _sum0 = vmlaq_f32(_sum0, _pA0, _pB0);
-                _sum1 = vmlaq_f32(_sum1, _pA0, _pB1);
-#endif
-
-                pA += 1;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum1));
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-            }
-
-            outptr += 8;
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            float32x4_t _sum;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum = vdupq_n_f32(pC[0]);
-                }
-                else
-                {
-                    _sum = vdupq_n_f32(0.f);
-                }
-            }
-            else
-            {
-                _sum = vld1q_f32(outptr);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pB = bfloat2float(vld1_u16(pB));
-                float32x4_t _pA = bfloat2float(vdup_n_u16(pA[0]));
-
-#if __aarch64__
-                _sum = vfmaq_f32(_sum, _pA, _pB);
-#else
-                _sum = vmlaq_f32(_sum, _pA, _pB);
-#endif
-
-                pA += 1;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum));
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum);
-            }
-
-            outptr += 4;
-        }
-#endif // __ARM_NEON
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            float sum0;
-            float sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    sum0 = pC[0];
-                    sum1 = pC[0];
-                }
-                else
-                {
-                    sum0 = 0.f;
-                    sum1 = 0.f;
-                }
-            }
-            else
-            {
-                sum0 = outptr[0];
-                sum1 = outptr[1];
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                sum0 += bfloat16_to_float32(pA[0]) * bfloat16_to_float32(pB[0]);
-                sum1 += bfloat16_to_float32(pA[0]) * bfloat16_to_float32(pB[1]);
-
-                pA += 1;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = float32_to_bfloat16(sum0);
-                    outptr0[1] = float32_to_bfloat16(sum1);
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-            }
-
-            outptr += 2;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            float sum;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    sum = pC[0];
-                }
-                else
-                {
-                    sum = 0.f;
-                }
-            }
-            else
-            {
-                sum = outptr[0];
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                sum += bfloat16_to_float32(pA[0]) * bfloat16_to_float32(pB[0]);
-
-                pA += 1;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = float32_to_bfloat16(sum);
-                    outptr0++;
-                }
-            }
-            else
-            {
-                outptr[0] = sum;
-            }
-
-            outptr += 1;
-        }
-
-        pAT += max_kk;
-    }
-}
-
-static void convolution_im2col_gemm_get_optimal_tile_mnk_bf16s(int M, int N, int K, int& TILE_M, int& TILE_N, int& TILE_K, int nT)
-{
-    // resolve optimal tile size from cache size
-    const int l2_cache_size_bf16 = (int)(get_cpu_level2_cache_size() / sizeof(unsigned short));
-
-    if (nT == 0)
-        nT = get_physical_big_cpu_count();
-
-    // solve K
-    {
-        // try not to split K
-#if __aarch64__
-        int tile_size = (l2_cache_size_bf16 - 32) / 12;
-#elif __ARM_NEON
-        int tile_size = (l2_cache_size_bf16 - 16) / 8;
-#else
-        int tile_size = (l2_cache_size_bf16 - 2) / 3;
-#endif
-
-#if __aarch64__
-        TILE_K = std::max(8, tile_size / 8 * 8);
-#elif __ARM_NEON
-        TILE_K = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_K = std::max(2, tile_size / 2 * 2);
-#endif
-
-        int nn_K = (K + TILE_K - 1) / TILE_K;
-#if __aarch64__
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 7) / 8 * 8);
-#elif __ARM_NEON
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 3) / 4 * 4);
-#else
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 1) / 2 * 2);
-#endif
-    }
-
-    // solve M
-    {
-#if __aarch64__
-        int nn_M = (M + 31) / 32;
-#elif __ARM_NEON
-        int nn_M = (M + 15) / 16;
-#else
-        int nn_M = (M + 7) / 8;
-#endif
-
-#if __aarch64__
-        TILE_M = std::max(8, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-#elif __ARM_NEON
-        TILE_M = std::max(4, ((M + nn_M - 1) / nn_M + 3) / 4 * 4);
-#else
-        TILE_M = std::max(2, ((M + nn_M - 1) / nn_M + 1) / 2 * 2);
-#endif
-    }
-
-    {
-        TILE_M *= std::min(nT, get_physical_cpu_count());
-
-        int nn_M = (M + TILE_M - 1) / TILE_M;
-#if __aarch64__
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-#elif __ARM_NEON
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 3) / 4 * 4);
-#else
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 1) / 2 * 2);
-#endif
-
-        if (nT > 1)
-        {
-#if __aarch64__
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 7) / 8 * 8);
-#elif __ARM_NEON
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 3) / 4 * 4);
-#else
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 1) / 2 * 2);
-#endif
-        }
-    }
-
-    if (N > 0)
-    {
-        int tile_size;
-        if (TILE_K >= K)
-        {
-            tile_size = (l2_cache_size_bf16 - TILE_M * TILE_K) / TILE_K;
-        }
-        else
-        {
-            tile_size = (l2_cache_size_bf16 - TILE_M * TILE_K) / (TILE_M * 2 + TILE_K);
-        }
-
-#if __aarch64__
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#elif __ARM_NEON
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_N = std::max(1, tile_size);
-#endif
-
-        int nn_N = (N + TILE_N - 1) / TILE_N;
-#if __aarch64__
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#elif __ARM_NEON
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#else
-        TILE_N = std::min(TILE_N, (N + nn_N - 1) / nn_N);
-#endif
-    }
-}
-
-static void convolution_im2col_gemm_transform_kernel_bf16s(const Mat& kernel, Mat& AT, int inch, int outch, int kernel_w, int kernel_h, const Option& opt)
-{
-    // NCNN_LOGE("convolution_im2col_gemm_transform_kernel");
-    const int maxk = kernel_w * kernel_h;
-
-    const int M = outch;
-    const int K = inch * maxk;
-
-    int TILE_M, TILE_N, TILE_K;
-    convolution_im2col_gemm_get_optimal_tile_mnk_bf16s(M, 0, K, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    int elempack = 1;
-#if __ARM_NEON
-    if (opt.use_packing_layout)
-    {
-        elempack = inch % 4 == 0 ? 4 : 1;
-    }
-#endif // __ARM_NEON
-
-    // maxk-inch-outch to pa-maxk-inch/pa-outch
-    Mat A_data;
-    if (maxk == 1)
-    {
-        cast_float32_to_bfloat16(kernel, A_data);
-        A_data = A_data.reshape(maxk * inch, outch);
-    }
-    else
-    {
-        Mat weight_data_r2 = kernel.reshape(maxk, inch, outch);
-
-        A_data.create(maxk * inch, outch, (size_t)2u);
-
-        for (int q = 0; q < outch; q += 1)
-        {
-            unsigned short* g00 = A_data.row<unsigned short>(q);
-
-            for (int p = 0; p + (elempack - 1) < inch; p += elempack)
-            {
-                for (int k = 0; k < maxk; k++)
-                {
-                    for (int i = 0; i < elempack; i++)
-                    {
-                        const float* k00 = weight_data_r2.channel(q).row(p + i);
-                        g00[0] = float32_to_bfloat16(k00[k]);
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-
-    AT.create(TILE_K * TILE_M, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M, (size_t)2u);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat AT_tile = AT.channel(i / TILE_M).row_range(k / TILE_K, 1);
-
-            convolution_im2col_pack_A_tile_bf16_fp16(A_data, AT_tile, i, max_ii, k, max_kk);
-        }
-    }
-}
-
-static void convolution_im2col_gemm_bf16s(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, const Mat& bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = top_blob.w * top_blob.h;
-    const int K = bottom_blob.c * bottom_blob.elempack * maxk;
-
-    int TILE_M, TILE_N, TILE_K;
-    convolution_im2col_gemm_get_optimal_tile_mnk_bf16s(M, N, K, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 2u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-    {
-        const int ppj = ppjk / nn_K;
-        const int ppk = ppjk % nn_K;
-
-        const int j = ppj * TILE_N;
-        const int k = ppk * TILE_K;
-
-        const int max_jj = std::min((N - j), TILE_N);
-        const int max_kk = std::min((K - k), TILE_K);
-
-        Mat BT_tile = BT.channel(j / TILE_N).row_range(k / TILE_K, 1);
-
-        // im2col
-        convolution_im2col_input_tile_bf16_fp16(bottom_blob, BT_tile, j, max_jj, k, max_kk, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h);
-    }
-
-    Mat topT_tileX;
-    if (K > TILE_K)
-        topT_tileX.create(TILE_N * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat topT_tile;
-        if (K > TILE_K)
-            topT_tile = topT_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).row_range(k / TILE_K, 1);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).row_range(k / TILE_K, 1);
-
-                bool k_end = k + TILE_K >= K;
-
-                convolution_gemm_transB_packed_tile_bf16s(AT_tile, BT_tile, bias, topT_tile, top_blob, i, max_ii, j, max_jj, k, max_kk, k_end, opt.use_a53_a55_optimized_kernel);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_bf16s_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_bf16s_fp16s.h
deleted file mode 100644
index aedec024..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_bf16s_fp16s.h
+++ /dev/null
@@ -1,1147 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_im2col_pack_A_tile_bf16_fp16(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
-{
-    // A = (pa, maxk, inch/pa), outch
-    const int A_hstep = A.w;
-
-    unsigned short* pp = AT;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        const unsigned short* p0 = (const unsigned short*)A + (i + ii) * A_hstep + k;
-        const unsigned short* p1 = (const unsigned short*)A + (i + ii + 1) * A_hstep + k;
-        const unsigned short* p2 = (const unsigned short*)A + (i + ii + 2) * A_hstep + k;
-        const unsigned short* p3 = (const unsigned short*)A + (i + ii + 3) * A_hstep + k;
-        const unsigned short* p4 = (const unsigned short*)A + (i + ii + 4) * A_hstep + k;
-        const unsigned short* p5 = (const unsigned short*)A + (i + ii + 5) * A_hstep + k;
-        const unsigned short* p6 = (const unsigned short*)A + (i + ii + 6) * A_hstep + k;
-        const unsigned short* p7 = (const unsigned short*)A + (i + ii + 7) * A_hstep + k;
-
-        int kk = 0;
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8_t _r0 = vld1q_u16(p0);
-            uint16x8_t _r1 = vld1q_u16(p1);
-            uint16x8_t _r2 = vld1q_u16(p2);
-            uint16x8_t _r3 = vld1q_u16(p3);
-            uint16x8_t _r4 = vld1q_u16(p4);
-            uint16x8_t _r5 = vld1q_u16(p5);
-            uint16x8_t _r6 = vld1q_u16(p6);
-            uint16x8_t _r7 = vld1q_u16(p7);
-            transpose8x8_u16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-            vst1q_u16(pp, _r0);
-            vst1q_u16(pp + 8, _r1);
-            vst1q_u16(pp + 8 * 2, _r2);
-            vst1q_u16(pp + 8 * 3, _r3);
-            vst1q_u16(pp + 8 * 4, _r4);
-            vst1q_u16(pp + 8 * 5, _r5);
-            vst1q_u16(pp + 8 * 6, _r6);
-            vst1q_u16(pp + 8 * 7, _r7);
-            pp += 64;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-            p4 += 8;
-            p5 += 8;
-            p6 += 8;
-            p7 += 8;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp[1] = p1[0];
-            pp[2] = p2[0];
-            pp[3] = p3[0];
-            pp[4] = p4[0];
-            pp[5] = p5[0];
-            pp[6] = p6[0];
-            pp[7] = p7[0];
-            pp += 8;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-            p4++;
-            p5++;
-            p6++;
-            p7++;
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        const unsigned short* p0 = (const unsigned short*)A + (i + ii) * A_hstep + k;
-        const unsigned short* p1 = (const unsigned short*)A + (i + ii + 1) * A_hstep + k;
-        const unsigned short* p2 = (const unsigned short*)A + (i + ii + 2) * A_hstep + k;
-        const unsigned short* p3 = (const unsigned short*)A + (i + ii + 3) * A_hstep + k;
-
-        int kk = 0;
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8x4_t _r0123;
-            _r0123.val[0] = vld1q_u16(p0);
-            _r0123.val[1] = vld1q_u16(p1);
-            _r0123.val[2] = vld1q_u16(p2);
-            _r0123.val[3] = vld1q_u16(p3);
-            vst4q_u16(pp, _r0123);
-            pp += 32;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4x4_t _r0123;
-            _r0123.val[0] = vld1_u16(p0);
-            _r0123.val[1] = vld1_u16(p1);
-            _r0123.val[2] = vld1_u16(p2);
-            _r0123.val[3] = vld1_u16(p3);
-            vst4_u16(pp, _r0123);
-            pp += 16;
-            p0 += 4;
-            p1 += 4;
-            p2 += 4;
-            p3 += 4;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp[1] = p1[0];
-            pp[2] = p2[0];
-            pp[3] = p3[0];
-            pp += 4;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        const unsigned short* p0 = (const unsigned short*)A + (i + ii) * A_hstep + k;
-        const unsigned short* p1 = (const unsigned short*)A + (i + ii + 1) * A_hstep + k;
-
-        int kk = 0;
-#if __ARM_NEON
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8x2_t _r01;
-            _r01.val[0] = vld1q_u16(p0);
-            _r01.val[1] = vld1q_u16(p1);
-            vst2q_u16(pp, _r01);
-            pp += 16;
-            p0 += 8;
-            p1 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4x2_t _r01;
-            _r01.val[0] = vld1_u16(p0);
-            _r01.val[1] = vld1_u16(p1);
-            vst2_u16(pp, _r01);
-            pp += 8;
-            p0 += 4;
-            p1 += 4;
-        }
-#endif // __ARM_NEON
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp[1] = p1[0];
-            pp += 2;
-            p0++;
-            p1++;
-        }
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        const unsigned short* p0 = (const unsigned short*)A + (i + ii) * A_hstep + k;
-
-        int kk = 0;
-#if __ARM_NEON
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            vst1q_u16(pp, vld1q_u16(p0));
-            pp += 8;
-            p0 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            vst1_u16(pp, vld1_u16(p0));
-            pp += 4;
-            p0 += 4;
-        }
-#endif // __ARM_NEON
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = (unsigned short)p0[0];
-            pp += 1;
-            p0++;
-        }
-    }
-}
-
-static void convolution_im2col_input_tile_conv1x1s1d1_bf16_fp16(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk)
-{
-    const int elempack = bottom_blob.elempack;
-
-    unsigned short* pp = B;
-
-    int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; jj + 11 < max_jj; jj += 12)
-    {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)bottom_blob.channel(k / 8) + (j + jj) * 8;
-
-            int kk = 0;
-            for (; kk < max_kk / 8; kk++)
-            {
-                // transpose8x12
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]   \n"
-                    "ld4    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]   \n"
-                    "ld4    {v4.8h, v5.8h, v6.8h, v7.8h}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]   \n"
-                    "ld4    {v16.8h, v17.8h, v18.8h, v19.8h}, [%0] \n"
-                    "uzp1   v20.8h, v0.8h, v4.8h    \n"
-                    "uzp2   v26.8h, v0.8h, v4.8h    \n"
-                    "uzp1   v21.8h, v16.8h, v1.8h   \n"
-                    "uzp2   v27.8h, v16.8h, v1.8h   \n"
-                    "sub    %0, %0, #128            \n"
-                    "uzp1   v22.8h, v5.8h, v17.8h   \n"
-                    "uzp2   v28.8h, v5.8h, v17.8h   \n"
-                    "uzp1   v23.8h, v2.8h, v6.8h    \n"
-                    "uzp2   v29.8h, v2.8h, v6.8h    \n"
-                    "uzp1   v24.8h, v18.8h, v3.8h   \n"
-                    "uzp2   v30.8h, v18.8h, v3.8h   \n"
-                    "uzp1   v25.8h, v7.8h, v19.8h   \n"
-                    "uzp2   v31.8h, v7.8h, v19.8h   \n"
-                    "st1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%1], #64 \n"
-                    "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%1], #64 \n"
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-                uint16x8_t _r0 = vld1q_u16(p0);
-                uint16x8_t _r1 = vld1q_u16(p0 + 8);
-                uint16x8_t _r2 = vld1q_u16(p0 + 8 * 2);
-                uint16x8_t _r3 = vld1q_u16(p0 + 8 * 3);
-                uint16x8_t _r4 = vld1q_u16(p0 + 8 * 4);
-                uint16x8_t _r5 = vld1q_u16(p0 + 8 * 5);
-                uint16x8_t _r6 = vld1q_u16(p0 + 8 * 6);
-                uint16x8_t _r7 = vld1q_u16(p0 + 8 * 7);
-                uint16x8_t _r8 = vld1q_u16(p0 + 8 * 8);
-                uint16x8_t _r9 = vld1q_u16(p0 + 8 * 9);
-                uint16x8_t _ra = vld1q_u16(p0 + 8 * 10);
-                uint16x8_t _rb = vld1q_u16(p0 + 8 * 11);
-                transpose8x12_u16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7, _r8, _r9, _ra, _rb);
-                vst1q_u16(pp, _r0);
-                vst1q_u16(pp + 8, _r1);
-                vst1q_u16(pp + 8 * 2, _r2);
-                vst1q_u16(pp + 8 * 3, _r3);
-                vst1q_u16(pp + 8 * 4, _r4);
-                vst1q_u16(pp + 8 * 5, _r5);
-                vst1q_u16(pp + 8 * 6, _r6);
-                vst1q_u16(pp + 8 * 7, _r7);
-                vst1q_u16(pp + 8 * 8, _r8);
-                vst1q_u16(pp + 8 * 9, _r9);
-                vst1q_u16(pp + 8 * 10, _ra);
-                vst1q_u16(pp + 8 * 11, _rb);
-                pp += 96;
-#endif // NCNN_GNU_INLINE_ASM
-                p0 += bottom_blob.cstep * 8;
-            }
-        }
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)bottom_blob.channel(k / 4) + (j + jj) * 4;
-
-            int kk = 0;
-            for (; kk < max_kk / 4; kk++)
-            {
-                // transpose4x12
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #256]   \n"
-                    "ld4    {v4.4h, v5.4h, v6.4h, v7.4h}, [%0]      \n"
-                    "st1    {v0.8h}, [%1], #16          \n"
-                    "st1    {v4.4h}, [%1], #8           \n"
-                    "st1    {v1.8h}, [%1], #16          \n"
-                    "st1    {v5.4h}, [%1], #8           \n"
-                    "sub    %0, %0, #64                 \n"
-                    "st1    {v2.8h}, [%1], #16          \n"
-                    "st1    {v6.4h}, [%1], #8           \n"
-                    "st1    {v3.8h}, [%1], #16          \n"
-                    "st1    {v7.4h}, [%1], #8           \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
-#else  // NCNN_GNU_INLINE_ASM
-                uint16x4x4_t _r0 = vld4_u16(p0);
-                uint16x4x4_t _r1 = vld4_u16(p0 + 16);
-                uint16x4x4_t _r2 = vld4_u16(p0 + 32);
-                vst1_u16(pp, _r0.val[0]);
-                vst1_u16(pp + 4, _r1.val[0]);
-                vst1_u16(pp + 4 * 2, _r2.val[0]);
-                vst1_u16(pp + 4 * 3, _r0.val[1]);
-                vst1_u16(pp + 4 * 4, _r1.val[1]);
-                vst1_u16(pp + 4 * 5, _r2.val[1]);
-                vst1_u16(pp + 4 * 6, _r0.val[2]);
-                vst1_u16(pp + 4 * 7, _r1.val[2]);
-                vst1_u16(pp + 4 * 8, _r2.val[2]);
-                vst1_u16(pp + 4 * 9, _r0.val[3]);
-                vst1_u16(pp + 4 * 10, _r1.val[3]);
-                vst1_u16(pp + 4 * 11, _r2.val[3]);
-                pp += 48;
-#endif // NCNN_GNU_INLINE_ASM
-                p0 += bottom_blob.cstep * 4;
-            }
-        }
-
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)bottom_blob.channel(k) + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                uint16x8_t _r01 = vld1q_u16(p0);
-                uint16x4_t _r2 = vld1_u16(p0 + 8);
-                vst1q_u16(pp, _r01);
-                vst1_u16(pp + 8, _r2);
-                pp += 12;
-                p0 += bottom_blob.cstep;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)bottom_blob.channel(k / 8) + (j + jj) * 8;
-
-            int kk = 0;
-            for (; kk < max_kk / 8; kk++)
-            {
-                // transpose8x8
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]   \n"
-                    "ld4    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0], #64 \n"
-                    "prfm   pldl1keep, [%0, #512]   \n"
-                    "ld4    {v4.8h, v5.8h, v6.8h, v7.8h}, [%0] \n"
-                    "uzp1   v16.8h, v0.8h, v4.8h    \n"
-                    "uzp2   v20.8h, v0.8h, v4.8h    \n"
-                    "uzp1   v17.8h, v1.8h, v5.8h    \n"
-                    "uzp2   v21.8h, v1.8h, v5.8h    \n"
-                    "sub    %0, %0, #64             \n"
-                    "uzp1   v18.8h, v2.8h, v6.8h    \n"
-                    "uzp2   v22.8h, v2.8h, v6.8h    \n"
-                    "uzp1   v19.8h, v3.8h, v7.8h    \n"
-                    "uzp2   v23.8h, v3.8h, v7.8h    \n"
-                    "st1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%1], #64 \n"
-                    "st1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-#else  // NCNN_GNU_INLINE_ASM
-                uint16x8_t _r0 = vld1q_u16(p0);
-                uint16x8_t _r1 = vld1q_u16(p0 + 8);
-                uint16x8_t _r2 = vld1q_u16(p0 + 8 * 2);
-                uint16x8_t _r3 = vld1q_u16(p0 + 8 * 3);
-                uint16x8_t _r4 = vld1q_u16(p0 + 8 * 4);
-                uint16x8_t _r5 = vld1q_u16(p0 + 8 * 5);
-                uint16x8_t _r6 = vld1q_u16(p0 + 8 * 6);
-                uint16x8_t _r7 = vld1q_u16(p0 + 8 * 7);
-                transpose8x8_u16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-                vst1q_u16(pp, _r0);
-                vst1q_u16(pp + 8, _r1);
-                vst1q_u16(pp + 8 * 2, _r2);
-                vst1q_u16(pp + 8 * 3, _r3);
-                vst1q_u16(pp + 8 * 4, _r4);
-                vst1q_u16(pp + 8 * 5, _r5);
-                vst1q_u16(pp + 8 * 6, _r6);
-                vst1q_u16(pp + 8 * 7, _r7);
-                pp += 64;
-#endif // NCNN_GNU_INLINE_ASM
-                p0 += bottom_blob.cstep * 8;
-            }
-        }
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)bottom_blob.channel(k / 4) + (j + jj) * 4;
-
-            int kk = 0;
-            for (; kk < max_kk / 4; kk++)
-            {
-                // transpose4x8
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0] \n"
-                    "st1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%0, #256]          \n"
-                    "vld4.u16   {d0,d2,d4,d6}, [%0 :64]! \n"
-                    "pld        [%0, #256]          \n"
-                    "vld4.u16   {d1,d3,d5,d7}, [%0 :64] \n"
-                    "sub        %0, %0, #32         \n"
-                    "vstm       %1!, {d0-d7}        \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                uint16x8x4_t _r0 = vld4q_u16(p0);
-                vst1q_u16(pp, _r0.val[0]);
-                vst1q_u16(pp + 8, _r0.val[1]);
-                vst1q_u16(pp + 16, _r0.val[2]);
-                vst1q_u16(pp + 24, _r0.val[3]);
-                pp += 32;
-#endif // NCNN_GNU_INLINE_ASM
-                p0 += bottom_blob.cstep * 4;
-            }
-        }
-
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)bottom_blob.channel(k) + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                uint16x8_t _r0 = vld1q_u16(p0);
-                vst1q_u16(pp, _r0);
-                pp += 8;
-                p0 += bottom_blob.cstep;
-            }
-        }
-    }
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)bottom_blob.channel(k / 8) + (j + jj) * 8;
-
-            int kk = 0;
-            for (; kk < max_kk / 8; kk++)
-            {
-                // transpose8x4
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0] \n"
-                    "st4    {v0.8h, v1.8h, v2.8h, v3.8h}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3");
-#else  // NCNN_GNU_INLINE_ASM
-                uint16x8x4_t _r0;
-                _r0.val[0] = vld1q_u16(p0);
-                _r0.val[1] = vld1q_u16(p0 + 8);
-                _r0.val[2] = vld1q_u16(p0 + 8 * 2);
-                _r0.val[3] = vld1q_u16(p0 + 8 * 3);
-                vst4q_u16(pp, _r0);
-                pp += 32;
-#endif // NCNN_GNU_INLINE_ASM
-                p0 += bottom_blob.cstep * 8;
-            }
-        }
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)bottom_blob.channel(k / 4) + (j + jj) * 4;
-
-            int kk = 0;
-            for (; kk < max_kk / 4; kk++)
-            {
-                // transpose4x4
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #256]       \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%0] \n"
-                    "st4    {v0.4h, v1.4h, v2.4h, v3.4h}, [%1], #32 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "v0", "v1", "v2", "v3");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%0, #256]          \n"
-                    "vld1.u16   {d0-d3}, [%0 :64]   \n"
-                    "vst4.u16   {d0-d3}, [%1 :64]!  \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp)
-                    : "memory", "q0", "q1");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                uint16x4x4_t _r0;
-                _r0.val[0] = vld1_u16(p0);
-                _r0.val[1] = vld1_u16(p0 + 4);
-                _r0.val[2] = vld1_u16(p0 + 4 * 2);
-                _r0.val[3] = vld1_u16(p0 + 4 * 3);
-                vst4_u16(pp, _r0);
-                pp += 16;
-#endif // NCNN_GNU_INLINE_ASM
-                p0 += bottom_blob.cstep * 4;
-            }
-        }
-
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)bottom_blob.channel(k) + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                uint16x4_t _r0 = vld1_u16(p0);
-                vst1_u16(pp, _r0);
-                pp += 4;
-                p0 += bottom_blob.cstep;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-#if __ARM_NEON
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)bottom_blob.channel(k / 8) + (j + jj) * 8;
-
-            int kk = 0;
-            for (; kk < max_kk / 8; kk++)
-            {
-                // transpose8x2
-                uint16x8x2_t _r0;
-                _r0.val[0] = vld1q_u16(p0);
-                _r0.val[1] = vld1q_u16(p0 + 8);
-                vst2q_u16(pp, _r0);
-                pp += 16;
-                p0 += bottom_blob.cstep * 8;
-            }
-        }
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)bottom_blob.channel(k / 4) + (j + jj) * 4;
-
-            int kk = 0;
-            for (; kk < max_kk / 4; kk++)
-            {
-                // transpose4x2
-                uint16x4x2_t _r0;
-                _r0.val[0] = vld1_u16(p0);
-                _r0.val[1] = vld1_u16(p0 + 4);
-                vst2_u16(pp, _r0);
-                pp += 8;
-                p0 += bottom_blob.cstep * 4;
-            }
-        }
-#endif // __ARM_NEON
-
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)bottom_blob.channel(k) + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                pp += 2;
-                p0 += bottom_blob.cstep;
-            }
-        }
-    }
-    for (; jj < max_jj; jj++)
-    {
-#if __ARM_NEON
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)bottom_blob.channel(k / 8) + (j + jj) * 8;
-
-            int kk = 0;
-            for (; kk < max_kk / 8; kk++)
-            {
-                vst1q_u16(pp, vld1q_u16(p0));
-                pp += 8;
-                p0 += bottom_blob.cstep * 8;
-            }
-        }
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)bottom_blob.channel(k / 4) + (j + jj) * 4;
-
-            int kk = 0;
-            for (; kk < max_kk / 4; kk++)
-            {
-                vst1_u16(pp, vld1_u16(p0));
-                pp += 4;
-                p0 += bottom_blob.cstep * 4;
-            }
-        }
-#endif // __ARM_NEON
-
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)bottom_blob.channel(k) + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp += 1;
-                p0 += bottom_blob.cstep;
-            }
-        }
-    }
-}
-
-static void convolution_im2col_input_tile_bf16_fp16(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h)
-{
-    if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
-    {
-        convolution_im2col_input_tile_conv1x1s1d1_bf16_fp16(bottom_blob, B, j, max_jj, k, max_kk);
-        return;
-    }
-
-    const int w = bottom_blob.w;
-    // const int channels = bottom_blob.c;
-    const int elempack = bottom_blob.elempack;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int outw = (w - kernel_extent_w) / stride_w + 1;
-
-    // j max_jj     outw*outh    split w and h
-
-    // k max_kk     pa*maxk*(inch/pa)    split inch
-
-    // k/max_kk shall be multiple of maxk
-
-    const int maxk = kernel_w * kernel_h;
-
-    unsigned short* pp = B;
-
-    int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; jj + 11 < max_jj; jj += 12)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dy2 = (j + jj + 2) / outw;
-        int dy3 = (j + jj + 3) / outw;
-        int dy4 = (j + jj + 4) / outw;
-        int dy5 = (j + jj + 5) / outw;
-        int dy6 = (j + jj + 6) / outw;
-        int dy7 = (j + jj + 7) / outw;
-        int dy8 = (j + jj + 8) / outw;
-        int dy9 = (j + jj + 9) / outw;
-        int dya = (j + jj + 10) / outw;
-        int dyb = (j + jj + 11) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-        int dx2 = (j + jj + 2) % outw;
-        int dx3 = (j + jj + 3) % outw;
-        int dx4 = (j + jj + 4) % outw;
-        int dx5 = (j + jj + 5) % outw;
-        int dx6 = (j + jj + 6) % outw;
-        int dx7 = (j + jj + 7) % outw;
-        int dx8 = (j + jj + 8) % outw;
-        int dx9 = (j + jj + 9) % outw;
-        int dxa = (j + jj + 10) % outw;
-        int dxb = (j + jj + 11) % outw;
-
-        int kk = 0;
-        for (; kk < max_kk / elempack; kk++)
-        {
-            int p = (k / elempack + kk) / maxk;
-            int uv = (k / elempack + kk) % maxk;
-            int u = uv / kernel_w;
-            int v = uv % kernel_w;
-
-            const Mat img = bottom_blob.channel(p);
-
-            int x0 = stride_w * dx0 + dilation_w * v;
-            int x1 = stride_w * dx1 + dilation_w * v;
-            int x2 = stride_w * dx2 + dilation_w * v;
-            int x3 = stride_w * dx3 + dilation_w * v;
-            int x4 = stride_w * dx4 + dilation_w * v;
-            int x5 = stride_w * dx5 + dilation_w * v;
-            int x6 = stride_w * dx6 + dilation_w * v;
-            int x7 = stride_w * dx7 + dilation_w * v;
-            int x8 = stride_w * dx8 + dilation_w * v;
-            int x9 = stride_w * dx9 + dilation_w * v;
-            int xa = stride_w * dxa + dilation_w * v;
-            int xb = stride_w * dxb + dilation_w * v;
-
-            int y0 = stride_h * dy0 + dilation_h * u;
-            int y1 = stride_h * dy1 + dilation_h * u;
-            int y2 = stride_h * dy2 + dilation_h * u;
-            int y3 = stride_h * dy3 + dilation_h * u;
-            int y4 = stride_h * dy4 + dilation_h * u;
-            int y5 = stride_h * dy5 + dilation_h * u;
-            int y6 = stride_h * dy6 + dilation_h * u;
-            int y7 = stride_h * dy7 + dilation_h * u;
-            int y8 = stride_h * dy8 + dilation_h * u;
-            int y9 = stride_h * dy9 + dilation_h * u;
-            int ya = stride_h * dya + dilation_h * u;
-            int yb = stride_h * dyb + dilation_h * u;
-
-            const unsigned short* sptr0 = img.row<const unsigned short>(y0) + x0 * elempack;
-            const unsigned short* sptr1 = img.row<const unsigned short>(y1) + x1 * elempack;
-            const unsigned short* sptr2 = img.row<const unsigned short>(y2) + x2 * elempack;
-            const unsigned short* sptr3 = img.row<const unsigned short>(y3) + x3 * elempack;
-            const unsigned short* sptr4 = img.row<const unsigned short>(y4) + x4 * elempack;
-            const unsigned short* sptr5 = img.row<const unsigned short>(y5) + x5 * elempack;
-            const unsigned short* sptr6 = img.row<const unsigned short>(y6) + x6 * elempack;
-            const unsigned short* sptr7 = img.row<const unsigned short>(y7) + x7 * elempack;
-            const unsigned short* sptr8 = img.row<const unsigned short>(y8) + x8 * elempack;
-            const unsigned short* sptr9 = img.row<const unsigned short>(y9) + x9 * elempack;
-            const unsigned short* sptra = img.row<const unsigned short>(ya) + xa * elempack;
-            const unsigned short* sptrb = img.row<const unsigned short>(yb) + xb * elempack;
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            if (elempack == 8)
-            {
-                uint16x8_t _r0 = vld1q_u16(sptr0);
-                uint16x8_t _r1 = vld1q_u16(sptr1);
-                uint16x8_t _r2 = vld1q_u16(sptr2);
-                uint16x8_t _r3 = vld1q_u16(sptr3);
-                uint16x8_t _r4 = vld1q_u16(sptr4);
-                uint16x8_t _r5 = vld1q_u16(sptr5);
-                uint16x8_t _r6 = vld1q_u16(sptr6);
-                uint16x8_t _r7 = vld1q_u16(sptr7);
-                uint16x8_t _r8 = vld1q_u16(sptr8);
-                uint16x8_t _r9 = vld1q_u16(sptr9);
-                uint16x8_t _ra = vld1q_u16(sptra);
-                uint16x8_t _rb = vld1q_u16(sptrb);
-                transpose8x12_u16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7, _r8, _r9, _ra, _rb);
-                vst1q_u16(pp, _r0);
-                vst1q_u16(pp + 8, _r1);
-                vst1q_u16(pp + 8 * 2, _r2);
-                vst1q_u16(pp + 8 * 3, _r3);
-                vst1q_u16(pp + 8 * 4, _r4);
-                vst1q_u16(pp + 8 * 5, _r5);
-                vst1q_u16(pp + 8 * 6, _r6);
-                vst1q_u16(pp + 8 * 7, _r7);
-                vst1q_u16(pp + 8 * 8, _r8);
-                vst1q_u16(pp + 8 * 9, _r9);
-                vst1q_u16(pp + 8 * 10, _ra);
-                vst1q_u16(pp + 8 * 11, _rb);
-                pp += 96;
-            }
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            if (elempack == 4)
-            {
-                uint16x4_t _r0 = vld1_u16(sptr0);
-                uint16x4_t _r1 = vld1_u16(sptr1);
-                uint16x4_t _r2 = vld1_u16(sptr2);
-                uint16x4_t _r3 = vld1_u16(sptr3);
-                uint16x4_t _r4 = vld1_u16(sptr4);
-                uint16x4_t _r5 = vld1_u16(sptr5);
-                uint16x4_t _r6 = vld1_u16(sptr6);
-                uint16x4_t _r7 = vld1_u16(sptr7);
-                uint16x4_t _r8 = vld1_u16(sptr8);
-                uint16x4_t _r9 = vld1_u16(sptr9);
-                uint16x4_t _ra = vld1_u16(sptra);
-                uint16x4_t _rb = vld1_u16(sptrb);
-                transpose4x12_u16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7, _r8, _r9, _ra, _rb);
-                vst1_u16(pp, _r0);
-                vst1_u16(pp + 4, _r1);
-                vst1_u16(pp + 4 * 2, _r2);
-                vst1_u16(pp + 4 * 3, _r3);
-                vst1_u16(pp + 4 * 4, _r4);
-                vst1_u16(pp + 4 * 5, _r5);
-                vst1_u16(pp + 4 * 6, _r6);
-                vst1_u16(pp + 4 * 7, _r7);
-                vst1_u16(pp + 4 * 8, _r8);
-                vst1_u16(pp + 4 * 9, _r9);
-                vst1_u16(pp + 4 * 10, _ra);
-                vst1_u16(pp + 4 * 11, _rb);
-                pp += 48;
-            }
-            if (elempack == 1)
-            {
-                pp[0] = sptr0[0];
-                pp[1] = sptr1[0];
-                pp[2] = sptr2[0];
-                pp[3] = sptr3[0];
-                pp[4] = sptr4[0];
-                pp[5] = sptr5[0];
-                pp[6] = sptr6[0];
-                pp[7] = sptr7[0];
-                pp[8] = sptr8[0];
-                pp[9] = sptr9[0];
-                pp[10] = sptra[0];
-                pp[11] = sptrb[0];
-                pp += 12;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dy2 = (j + jj + 2) / outw;
-        int dy3 = (j + jj + 3) / outw;
-        int dy4 = (j + jj + 4) / outw;
-        int dy5 = (j + jj + 5) / outw;
-        int dy6 = (j + jj + 6) / outw;
-        int dy7 = (j + jj + 7) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-        int dx2 = (j + jj + 2) % outw;
-        int dx3 = (j + jj + 3) % outw;
-        int dx4 = (j + jj + 4) % outw;
-        int dx5 = (j + jj + 5) % outw;
-        int dx6 = (j + jj + 6) % outw;
-        int dx7 = (j + jj + 7) % outw;
-
-        int kk = 0;
-        for (; kk < max_kk / elempack; kk++)
-        {
-            int p = (k / elempack + kk) / maxk;
-            int uv = (k / elempack + kk) % maxk;
-            int u = uv / kernel_w;
-            int v = uv % kernel_w;
-
-            const Mat img = bottom_blob.channel(p);
-
-            int x0 = stride_w * dx0 + dilation_w * v;
-            int x1 = stride_w * dx1 + dilation_w * v;
-            int x2 = stride_w * dx2 + dilation_w * v;
-            int x3 = stride_w * dx3 + dilation_w * v;
-            int x4 = stride_w * dx4 + dilation_w * v;
-            int x5 = stride_w * dx5 + dilation_w * v;
-            int x6 = stride_w * dx6 + dilation_w * v;
-            int x7 = stride_w * dx7 + dilation_w * v;
-            int y0 = stride_h * dy0 + dilation_h * u;
-            int y1 = stride_h * dy1 + dilation_h * u;
-            int y2 = stride_h * dy2 + dilation_h * u;
-            int y3 = stride_h * dy3 + dilation_h * u;
-            int y4 = stride_h * dy4 + dilation_h * u;
-            int y5 = stride_h * dy5 + dilation_h * u;
-            int y6 = stride_h * dy6 + dilation_h * u;
-            int y7 = stride_h * dy7 + dilation_h * u;
-
-            const unsigned short* sptr0 = img.row<const unsigned short>(y0) + x0 * elempack;
-            const unsigned short* sptr1 = img.row<const unsigned short>(y1) + x1 * elempack;
-            const unsigned short* sptr2 = img.row<const unsigned short>(y2) + x2 * elempack;
-            const unsigned short* sptr3 = img.row<const unsigned short>(y3) + x3 * elempack;
-            const unsigned short* sptr4 = img.row<const unsigned short>(y4) + x4 * elempack;
-            const unsigned short* sptr5 = img.row<const unsigned short>(y5) + x5 * elempack;
-            const unsigned short* sptr6 = img.row<const unsigned short>(y6) + x6 * elempack;
-            const unsigned short* sptr7 = img.row<const unsigned short>(y7) + x7 * elempack;
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            if (elempack == 8)
-            {
-                uint16x8_t _r0 = vld1q_u16(sptr0);
-                uint16x8_t _r1 = vld1q_u16(sptr1);
-                uint16x8_t _r2 = vld1q_u16(sptr2);
-                uint16x8_t _r3 = vld1q_u16(sptr3);
-                uint16x8_t _r4 = vld1q_u16(sptr4);
-                uint16x8_t _r5 = vld1q_u16(sptr5);
-                uint16x8_t _r6 = vld1q_u16(sptr6);
-                uint16x8_t _r7 = vld1q_u16(sptr7);
-                transpose8x8_u16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-                vst1q_u16(pp, _r0);
-                vst1q_u16(pp + 8, _r1);
-                vst1q_u16(pp + 8 * 2, _r2);
-                vst1q_u16(pp + 8 * 3, _r3);
-                vst1q_u16(pp + 8 * 4, _r4);
-                vst1q_u16(pp + 8 * 5, _r5);
-                vst1q_u16(pp + 8 * 6, _r6);
-                vst1q_u16(pp + 8 * 7, _r7);
-                pp += 64;
-            }
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            if (elempack == 4)
-            {
-                uint16x4_t _r0 = vld1_u16(sptr0);
-                uint16x4_t _r1 = vld1_u16(sptr1);
-                uint16x4_t _r2 = vld1_u16(sptr2);
-                uint16x4_t _r3 = vld1_u16(sptr3);
-                uint16x4_t _r4 = vld1_u16(sptr4);
-                uint16x4_t _r5 = vld1_u16(sptr5);
-                uint16x4_t _r6 = vld1_u16(sptr6);
-                uint16x4_t _r7 = vld1_u16(sptr7);
-                transpose4x8_u16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-                vst1_u16(pp, _r0);
-                vst1_u16(pp + 4, _r1);
-                vst1_u16(pp + 4 * 2, _r2);
-                vst1_u16(pp + 4 * 3, _r3);
-                vst1_u16(pp + 4 * 4, _r4);
-                vst1_u16(pp + 4 * 5, _r5);
-                vst1_u16(pp + 4 * 6, _r6);
-                vst1_u16(pp + 4 * 7, _r7);
-                pp += 32;
-            }
-            if (elempack == 1)
-            {
-                pp[0] = sptr0[0];
-                pp[1] = sptr1[0];
-                pp[2] = sptr2[0];
-                pp[3] = sptr3[0];
-                pp[4] = sptr4[0];
-                pp[5] = sptr5[0];
-                pp[6] = sptr6[0];
-                pp[7] = sptr7[0];
-                pp += 8;
-            }
-        }
-    }
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dy2 = (j + jj + 2) / outw;
-        int dy3 = (j + jj + 3) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-        int dx2 = (j + jj + 2) % outw;
-        int dx3 = (j + jj + 3) % outw;
-
-        int kk = 0;
-        for (; kk < max_kk / elempack; kk++)
-        {
-            int p = (k / elempack + kk) / maxk;
-            int uv = (k / elempack + kk) % maxk;
-            int u = uv / kernel_w;
-            int v = uv % kernel_w;
-
-            const Mat img = bottom_blob.channel(p);
-
-            int x0 = stride_w * dx0 + dilation_w * v;
-            int x1 = stride_w * dx1 + dilation_w * v;
-            int x2 = stride_w * dx2 + dilation_w * v;
-            int x3 = stride_w * dx3 + dilation_w * v;
-            int y0 = stride_h * dy0 + dilation_h * u;
-            int y1 = stride_h * dy1 + dilation_h * u;
-            int y2 = stride_h * dy2 + dilation_h * u;
-            int y3 = stride_h * dy3 + dilation_h * u;
-
-            const unsigned short* sptr0 = img.row<const unsigned short>(y0) + x0 * elempack;
-            const unsigned short* sptr1 = img.row<const unsigned short>(y1) + x1 * elempack;
-            const unsigned short* sptr2 = img.row<const unsigned short>(y2) + x2 * elempack;
-            const unsigned short* sptr3 = img.row<const unsigned short>(y3) + x3 * elempack;
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            if (elempack == 8)
-            {
-                uint16x8x4_t _r0;
-                _r0.val[0] = vld1q_u16(sptr0);
-                _r0.val[1] = vld1q_u16(sptr1);
-                _r0.val[2] = vld1q_u16(sptr2);
-                _r0.val[3] = vld1q_u16(sptr3);
-                vst4q_u16(pp, _r0);
-                pp += 32;
-            }
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            if (elempack == 4)
-            {
-                uint16x4x4_t _r0;
-                _r0.val[0] = vld1_u16(sptr0);
-                _r0.val[1] = vld1_u16(sptr1);
-                _r0.val[2] = vld1_u16(sptr2);
-                _r0.val[3] = vld1_u16(sptr3);
-                vst4_u16(pp, _r0);
-                pp += 16;
-            }
-            if (elempack == 1)
-            {
-                pp[0] = sptr0[0];
-                pp[1] = sptr1[0];
-                pp[2] = sptr2[0];
-                pp[3] = sptr3[0];
-                pp += 4;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-
-        int kk = 0;
-        for (; kk < max_kk / elempack; kk++)
-        {
-            int p = (k / elempack + kk) / maxk;
-            int uv = (k / elempack + kk) % maxk;
-            int u = uv / kernel_w;
-            int v = uv % kernel_w;
-
-            const Mat img = bottom_blob.channel(p);
-
-            int x0 = stride_w * dx0 + dilation_w * v;
-            int x1 = stride_w * dx1 + dilation_w * v;
-            int y0 = stride_h * dy0 + dilation_h * u;
-            int y1 = stride_h * dy1 + dilation_h * u;
-
-            const unsigned short* sptr0 = img.row<const unsigned short>(y0) + x0 * elempack;
-            const unsigned short* sptr1 = img.row<const unsigned short>(y1) + x1 * elempack;
-
-#if __ARM_NEON
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            if (elempack == 8)
-            {
-                pp[0] = sptr0[0];
-                pp[1] = sptr1[0];
-                pp[2] = sptr0[1];
-                pp[3] = sptr1[1];
-                pp[4] = sptr0[2];
-                pp[5] = sptr1[2];
-                pp[6] = sptr0[3];
-                pp[7] = sptr1[3];
-                pp[8 + 0] = sptr0[4];
-                pp[8 + 1] = sptr1[4];
-                pp[8 + 2] = sptr0[5];
-                pp[8 + 3] = sptr1[5];
-                pp[8 + 4] = sptr0[6];
-                pp[8 + 5] = sptr1[6];
-                pp[8 + 6] = sptr0[7];
-                pp[8 + 7] = sptr1[7];
-                pp += 16;
-            }
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            if (elempack == 4)
-            {
-                pp[0] = sptr0[0];
-                pp[1] = sptr1[0];
-                pp[2] = sptr0[1];
-                pp[3] = sptr1[1];
-                pp[4] = sptr0[2];
-                pp[5] = sptr1[2];
-                pp[6] = sptr0[3];
-                pp[7] = sptr1[3];
-                pp += 8;
-            }
-#endif // __ARM_NEON
-            if (elempack == 1)
-            {
-                pp[0] = sptr0[0];
-                pp[1] = sptr1[0];
-                pp += 2;
-            }
-        }
-    }
-    for (; jj < max_jj; jj++)
-    {
-        int dy = (j + jj) / outw;
-        int dx = (j + jj) % outw;
-
-        int kk = 0;
-        for (; kk < max_kk / elempack; kk++)
-        {
-            int p = (k / elempack + kk) / maxk;
-            int uv = (k / elempack + kk) % maxk;
-            int u = uv / kernel_w;
-            int v = uv % kernel_w;
-
-            const Mat img = bottom_blob.channel(p);
-
-            int x = stride_w * dx + dilation_w * v;
-            int y = stride_h * dy + dilation_h * u;
-
-            const unsigned short* sptr = img.row<const unsigned short>(y) + x * elempack;
-
-#if __ARM_NEON
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            if (elempack == 8)
-            {
-                pp[0] = sptr[0];
-                pp[1] = sptr[1];
-                pp[2] = sptr[2];
-                pp[3] = sptr[3];
-                pp[4] = sptr[4];
-                pp[5] = sptr[5];
-                pp[6] = sptr[6];
-                pp[7] = sptr[7];
-                pp += 8;
-            }
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            if (elempack == 4)
-            {
-                pp[0] = sptr[0];
-                pp[1] = sptr[1];
-                pp[2] = sptr[2];
-                pp[3] = sptr[3];
-                pp += 4;
-            }
-#endif // __ARM_NEON
-            if (elempack == 1)
-            {
-                pp[0] = sptr[0];
-                pp += 1;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_fp16s.h
deleted file mode 100644
index a4cc82d7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_fp16s.h
+++ /dev/null
@@ -1,3221 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_gemm_transB_packed_tile_fp16sa(const Mat& AT_tile, const Mat& BT_tile, const Mat& CT_tile, Mat& topT_tile, Mat& top_blob, int i, int max_ii, int j, int max_jj, int k, int max_kk, bool k_end, int use_a53_a55_optimized_kernel)
-{
-    // NCNN_LOGE("convolution_gemm_transB_packed_tile_fp16sa %d %d %d %d %d %d", i, max_ii, j, max_jj, k, max_kk);
-
-    const int out_elempack = top_blob.elempack;
-    const int out_hstep = (int)top_blob.cstep;
-
-    const __fp16* pAT = AT_tile;
-    const __fp16* pBT = BT_tile;
-    const __fp16* pC = CT_tile;
-
-    __fp16* outptr = topT_tile;
-
-    int ii = 0;
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        __fp16* outptr0 = (__fp16*)top_blob + (i + ii) * out_hstep + j * out_elempack;
-
-        const __fp16* pB = pBT;
-
-        if (pC)
-        {
-            pC = (const __fp16*)CT_tile + i + ii;
-        }
-
-        int jj = 0;
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            const __fp16* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            if (use_a53_a55_optimized_kernel)
-            {
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%0], #64 \n"
-                    "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                    "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0]      \n"
-                    "subs   %0, %0, #128                \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "ld1    {v20.8h}, [%8]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v20.16b, v20.16b, v20.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v21.16b, v20.16b            \n"
-                    "mov    v22.16b, v20.16b            \n"
-                    "mov    v23.16b, v20.16b            \n"
-                    "mov    v24.16b, v20.16b            \n"
-                    "mov    v25.16b, v20.16b            \n"
-                    "mov    v26.16b, v20.16b            \n"
-                    "mov    v27.16b, v20.16b            \n"
-                    "mov    v28.16b, v20.16b            \n"
-                    "mov    v29.16b, v20.16b            \n"
-                    "mov    v30.16b, v20.16b            \n"
-                    "mov    v31.16b, v20.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.8h}, [%1], #16          \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.8h}, [%2], #16          \n"
-
-                    "ldr    d1, [%2], #8                \n"
-                    "ldr    x21, [%2], #8               \n"
-
-                    ".align 4                           \n"
-                    "4:                                 \n"
-                    "ldr    d5, [%1], #8                \n"
-                    "fmla   v20.8h, v4.8h, v0.h[0]      \n"
-                    "ldr    x25, [%1], #8               \n"
-                    "fmla   v21.8h, v4.8h, v0.h[1]      \n"
-                    "ldr    d2, [%2], #8                \n"
-                    "fmla   v22.8h, v4.8h, v0.h[2]      \n"
-                    "ldr    x22, [%2], #8               \n"
-                    "fmla   v23.8h, v4.8h, v0.h[3]      \n"
-                    "ldr    d6, [%1], #8                \n"
-                    "fmla   v24.8h, v4.8h, v0.h[4]      \n"
-                    "ldr    x26, [%1], #8               \n"
-                    "fmla   v25.8h, v4.8h, v0.h[5]      \n"
-                    "ins    v1.d[1], x21                \n"
-                    "fmla   v26.8h, v4.8h, v0.h[6]      \n"
-                    "ldr    d3, [%2], #8                \n"
-                    "fmla   v27.8h, v4.8h, v0.h[7]      \n"
-                    "ldr    x23, [%2], #8               \n"
-                    "fmla   v28.8h, v4.8h, v1.h[0]      \n"
-                    "prfm   pldl1keep, [%2, #256]       \n" // NOTE PRELOAD
-                    "fmla   v29.8h, v4.8h, v1.h[1]      \n"
-                    "ins    v5.d[1], x25                \n"
-                    "fmla   v30.8h, v4.8h, v1.h[2]      \n"
-                    "ldr    d8, [%2], #8                \n"
-                    "fmla   v31.8h, v4.8h, v1.h[3]      \n"
-                    "ldr    x20, [%2], #8               \n"
-                    "fmla   v20.8h, v5.8h, v1.h[4]      \n"
-                    "ldr    d7, [%1], #8                \n"
-                    "fmla   v21.8h, v5.8h, v1.h[5]      \n"
-                    "ins    v2.d[1], x22                \n"
-                    "fmla   v22.8h, v5.8h, v1.h[6]      \n"
-                    "ldr    x27, [%1], #8               \n"
-                    "fmla   v23.8h, v5.8h, v1.h[7]      \n"
-                    "ldr    d9, [%2], #8                \n"
-                    "fmla   v24.8h, v5.8h, v2.h[0]      \n"
-                    "ldr    x21, [%2], #8               \n"
-                    "fmla   v25.8h, v5.8h, v2.h[1]      \n"
-                    "ins    v6.d[1], x26                \n"
-                    "fmla   v26.8h, v5.8h, v2.h[2]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                    "fmla   v27.8h, v5.8h, v2.h[3]      \n"
-                    "ldr    d4, [%1], #8                \n"
-                    "fmla   v28.8h, v5.8h, v2.h[4]      \n"
-                    "ldr    x24, [%1], #8               \n"
-                    "fmla   v29.8h, v5.8h, v2.h[5]      \n"
-                    "ins    v3.d[1], x23                \n"
-                    "fmla   v30.8h, v5.8h, v2.h[6]      \n"
-                    "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                    "fmla   v31.8h, v5.8h, v2.h[7]      \n"
-                    "ldr    d0, [%2], #8                \n"
-                    "fmla   v20.8h, v6.8h, v3.h[0]      \n"
-                    "fmla   v21.8h, v6.8h, v3.h[1]      \n"
-                    "fmla   v22.8h, v6.8h, v3.h[2]      \n"
-                    "fmla   v23.8h, v6.8h, v3.h[3]      \n"
-                    "fmla   v24.8h, v6.8h, v3.h[4]      \n"
-                    "fmla   v25.8h, v6.8h, v3.h[5]      \n"
-                    "ins    v8.d[1], x20                \n"
-                    "fmla   v26.8h, v6.8h, v3.h[6]      \n"
-                    "ldr    x20, [%2], #8               \n"
-                    "fmla   v27.8h, v6.8h, v3.h[7]      \n"
-                    "ldr    d1, [%2], #8                \n"
-                    "fmla   v28.8h, v6.8h, v8.h[0]      \n"
-                    "fmla   v29.8h, v6.8h, v8.h[1]      \n"
-                    "ins    v7.d[1], x27                \n"
-                    "fmla   v30.8h, v6.8h, v8.h[2]      \n"
-                    "fmla   v31.8h, v6.8h, v8.h[3]      \n"
-                    "fmla   v20.8h, v7.8h, v8.h[4]      \n"
-                    "fmla   v21.8h, v7.8h, v8.h[5]      \n"
-                    "ins    v9.d[1], x21                \n"
-                    "fmla   v22.8h, v7.8h, v8.h[6]      \n"
-                    "fmla   v23.8h, v7.8h, v8.h[7]      \n"
-                    "ldr    x21, [%2], #8               \n"
-                    "fmla   v24.8h, v7.8h, v9.h[0]      \n"
-                    "fmla   v25.8h, v7.8h, v9.h[1]      \n"
-                    "ins    v4.d[1], x24                \n"
-                    "fmla   v26.8h, v7.8h, v9.h[2]      \n"
-                    "fmla   v27.8h, v7.8h, v9.h[3]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v28.8h, v7.8h, v9.h[4]      \n"
-                    "fmla   v29.8h, v7.8h, v9.h[5]      \n"
-                    "fmla   v30.8h, v7.8h, v9.h[6]      \n"
-                    "ins    v0.d[1], x20                \n"
-                    "fmla   v31.8h, v7.8h, v9.h[7]      \n"
-                    "bne    4b                          \n"
-
-                    "sub    %1, %1, #16                 \n"
-                    "sub    %2, %2, #32                 \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h}, [%2], #24 \n"
-                    "ld1    {v4.8h}, [%1], #16          \n"
-                    "fmla   v20.8h, v4.8h, v0.h[0]      \n"
-                    "fmla   v21.8h, v4.8h, v0.h[1]      \n"
-                    "fmla   v22.8h, v4.8h, v0.h[2]      \n"
-                    "fmla   v23.8h, v4.8h, v0.h[3]      \n"
-                    "fmla   v24.8h, v4.8h, v1.h[0]      \n"
-                    "fmla   v25.8h, v4.8h, v1.h[1]      \n"
-                    "fmla   v26.8h, v4.8h, v1.h[2]      \n"
-                    "fmla   v27.8h, v4.8h, v1.h[3]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v28.8h, v4.8h, v2.h[0]      \n"
-                    "fmla   v29.8h, v4.8h, v2.h[1]      \n"
-                    "fmla   v30.8h, v4.8h, v2.h[2]      \n"
-                    "fmla   v31.8h, v4.8h, v2.h[3]      \n"
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    11f                         \n"
-
-                    // if out_elempack == 8
-                    "cmp    %w12, #8                    \n"
-                    "bne    8f                          \n"
-
-                    "st1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%3], #64 \n"
-                    "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%3], #64 \n"
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%3], #64 \n"
-                    "b      10f                         \n"
-
-                    // if out_elempack == 4
-                    "8:                                 \n"
-                    "cmp    %w12, #4                    \n"
-                    "bne    9f                          \n"
-
-                    "zip1   v12.2d, v20.2d, v21.2d      \n"
-                    "zip2   v18.2d, v20.2d, v21.2d      \n"
-                    "zip1   v13.2d, v22.2d, v23.2d      \n"
-                    "zip2   v19.2d, v22.2d, v23.2d      \n"
-                    "zip1   v14.2d, v24.2d, v25.2d      \n"
-                    "zip2   v20.2d, v24.2d, v25.2d      \n"
-                    "zip1   v15.2d, v26.2d, v27.2d      \n"
-                    "zip2   v21.2d, v26.2d, v27.2d      \n"
-                    "zip1   v16.2d, v28.2d, v29.2d      \n"
-                    "zip2   v22.2d, v28.2d, v29.2d      \n"
-                    "zip1   v17.2d, v30.2d, v31.2d      \n"
-                    "zip2   v23.2d, v30.2d, v31.2d      \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 1          \n"
-                    "st1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%3], #64 \n"
-                    "st1    {v16.8h, v17.8h}, [%3], #32 \n"
-                    "st1    {v18.8h, v19.8h, v20.8h, v21.8h}, [x4], #64 \n"
-                    "st1    {v22.8h, v23.8h}, [x4]      \n"
-                    "b      10f                         \n"
-
-                    // if out_elempack == 1
-                    "9:                                 \n"
-                    // transpose8x12
-                    "zip1   v18.8h, v20.8h, v21.8h      \n"
-                    "zip2   v19.8h, v20.8h, v21.8h      \n"
-                    "zip1   v20.8h, v22.8h, v23.8h      \n"
-                    "zip2   v21.8h, v22.8h, v23.8h      \n"
-                    "zip1   v22.8h, v24.8h, v25.8h      \n"
-                    "zip2   v23.8h, v24.8h, v25.8h      \n"
-                    "zip1   v24.8h, v26.8h, v27.8h      \n"
-                    "zip2   v25.8h, v26.8h, v27.8h      \n"
-                    "zip1   v26.8h, v28.8h, v29.8h      \n"
-                    "zip2   v27.8h, v28.8h, v29.8h      \n"
-                    "zip1   v28.8h, v30.8h, v31.8h      \n"
-                    "zip2   v29.8h, v30.8h, v31.8h      \n"
-
-                    "zip1   v0.4s, v18.4s, v20.4s       \n"
-                    "zip2   v3.4s, v18.4s, v20.4s       \n"
-                    "zip1   v6.4s, v19.4s, v21.4s       \n"
-                    "zip2   v9.4s, v19.4s, v21.4s       \n"
-                    "zip1   v1.4s, v22.4s, v24.4s       \n"
-                    "zip2   v4.4s, v22.4s, v24.4s       \n"
-                    "zip1   v7.4s, v23.4s, v25.4s       \n"
-                    "zip2   v10.4s, v23.4s, v25.4s      \n"
-                    "zip1   v2.4s, v26.4s, v28.4s       \n"
-                    "zip2   v5.4s, v26.4s, v28.4s       \n"
-                    "zip1   v8.4s, v27.4s, v29.4s       \n"
-                    "zip2   v11.4s, v27.4s, v29.4s      \n"
-
-                    "mov    v12.d[0], v0.d[1]           \n"
-                    "mov    v13.d[0], v1.d[1]           \n"
-                    "mov    v14.d[0], v2.d[1]           \n"
-                    "mov    v15.d[0], v3.d[1]           \n"
-                    "mov    v16.d[0], v4.d[1]           \n"
-                    "mov    v17.d[0], v5.d[1]           \n"
-                    "mov    v18.d[0], v6.d[1]           \n"
-                    "mov    v19.d[0], v7.d[1]           \n"
-                    "mov    v20.d[0], v8.d[1]           \n"
-                    "mov    v21.d[0], v9.d[1]           \n"
-                    "mov    v22.d[0], v10.d[1]          \n"
-                    "mov    v23.d[0], v11.d[1]          \n"
-
-                    "add    x4, %3, %w13, sxtw 1        \n"
-                    "st1    {v0.4h, v1.4h, v2.4h}, [%3], #24 \n"
-                    "st1    {v12.4h, v13.4h, v14.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v3.4h, v4.4h, v5.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v15.4h, v16.4h, v17.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v6.4h, v7.4h, v8.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v18.4h, v19.4h, v20.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v9.4h, v10.4h, v11.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v21.4h, v22.4h, v23.4h}, [x4] \n"
-
-                    "10:                                \n"
-                    "add    %0, %0, #192                \n"
-                    "b      12f                         \n"
-
-                    "11:                                \n"
-                    "st1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%0], #64 \n"
-                    "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                    "12:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            else
-            {
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%0], #64 \n"
-                    "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                    "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0]      \n"
-                    "subs   %0, %0, #128                \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "ld1    {v20.8h}, [%8]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v20.16b, v20.16b, v20.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v21.16b, v20.16b            \n"
-                    "mov    v22.16b, v20.16b            \n"
-                    "mov    v23.16b, v20.16b            \n"
-                    "mov    v24.16b, v20.16b            \n"
-                    "mov    v25.16b, v20.16b            \n"
-                    "mov    v26.16b, v20.16b            \n"
-                    "mov    v27.16b, v20.16b            \n"
-                    "mov    v28.16b, v20.16b            \n"
-                    "mov    v29.16b, v20.16b            \n"
-                    "mov    v30.16b, v20.16b            \n"
-                    "mov    v31.16b, v20.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%1], #64 \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%2], #64 \n"
-
-                    "fmla   v20.8h, v4.8h, v0.h[0]      \n"
-                    "fmla   v21.8h, v4.8h, v0.h[1]      \n"
-                    "fmla   v22.8h, v4.8h, v0.h[2]      \n"
-                    "fmla   v23.8h, v4.8h, v0.h[3]      \n"
-                    "fmla   v24.8h, v4.8h, v0.h[4]      \n"
-                    "fmla   v25.8h, v4.8h, v0.h[5]      \n"
-                    "fmla   v26.8h, v4.8h, v0.h[6]      \n"
-                    "fmla   v27.8h, v4.8h, v0.h[7]      \n"
-                    "fmla   v28.8h, v4.8h, v1.h[0]      \n"
-                    "fmla   v29.8h, v4.8h, v1.h[1]      \n"
-                    "fmla   v30.8h, v4.8h, v1.h[2]      \n"
-                    "fmla   v31.8h, v4.8h, v1.h[3]      \n"
-
-                    "fmla   v20.8h, v5.8h, v1.h[4]      \n"
-                    "fmla   v21.8h, v5.8h, v1.h[5]      \n"
-                    "fmla   v22.8h, v5.8h, v1.h[6]      \n"
-                    "fmla   v23.8h, v5.8h, v1.h[7]      \n"
-                    "fmla   v24.8h, v5.8h, v2.h[0]      \n"
-                    "fmla   v25.8h, v5.8h, v2.h[1]      \n"
-                    "fmla   v26.8h, v5.8h, v2.h[2]      \n"
-                    "fmla   v27.8h, v5.8h, v2.h[3]      \n"
-                    "fmla   v28.8h, v5.8h, v2.h[4]      \n"
-                    "fmla   v29.8h, v5.8h, v2.h[5]      \n"
-                    "fmla   v30.8h, v5.8h, v2.h[6]      \n"
-                    "fmla   v31.8h, v5.8h, v2.h[7]      \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v8.8h, v9.8h}, [%2], #32   \n"
-
-                    "fmla   v20.8h, v6.8h, v3.h[0]      \n"
-                    "fmla   v21.8h, v6.8h, v3.h[1]      \n"
-                    "fmla   v22.8h, v6.8h, v3.h[2]      \n"
-                    "fmla   v23.8h, v6.8h, v3.h[3]      \n"
-                    "fmla   v24.8h, v6.8h, v3.h[4]      \n"
-                    "fmla   v25.8h, v6.8h, v3.h[5]      \n"
-                    "fmla   v26.8h, v6.8h, v3.h[6]      \n"
-                    "fmla   v27.8h, v6.8h, v3.h[7]      \n"
-                    "fmla   v28.8h, v6.8h, v8.h[0]      \n"
-                    "fmla   v29.8h, v6.8h, v8.h[1]      \n"
-                    "fmla   v30.8h, v6.8h, v8.h[2]      \n"
-                    "fmla   v31.8h, v6.8h, v8.h[3]      \n"
-
-                    "subs   w4, w4, #1                  \n"
-
-                    "fmla   v20.8h, v7.8h, v8.h[4]      \n"
-                    "fmla   v21.8h, v7.8h, v8.h[5]      \n"
-                    "fmla   v22.8h, v7.8h, v8.h[6]      \n"
-                    "fmla   v23.8h, v7.8h, v8.h[7]      \n"
-                    "fmla   v24.8h, v7.8h, v9.h[0]      \n"
-                    "fmla   v25.8h, v7.8h, v9.h[1]      \n"
-                    "fmla   v26.8h, v7.8h, v9.h[2]      \n"
-                    "fmla   v27.8h, v7.8h, v9.h[3]      \n"
-                    "fmla   v28.8h, v7.8h, v9.h[4]      \n"
-                    "fmla   v29.8h, v7.8h, v9.h[5]      \n"
-                    "fmla   v30.8h, v7.8h, v9.h[6]      \n"
-                    "fmla   v31.8h, v7.8h, v9.h[7]      \n"
-
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h}, [%2], #24 \n"
-                    "ld1    {v4.8h}, [%1], #16          \n"
-                    "fmla   v20.8h, v4.8h, v0.h[0]      \n"
-                    "fmla   v21.8h, v4.8h, v0.h[1]      \n"
-                    "fmla   v22.8h, v4.8h, v0.h[2]      \n"
-                    "fmla   v23.8h, v4.8h, v0.h[3]      \n"
-                    "fmla   v24.8h, v4.8h, v1.h[0]      \n"
-                    "fmla   v25.8h, v4.8h, v1.h[1]      \n"
-                    "fmla   v26.8h, v4.8h, v1.h[2]      \n"
-                    "fmla   v27.8h, v4.8h, v1.h[3]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v28.8h, v4.8h, v2.h[0]      \n"
-                    "fmla   v29.8h, v4.8h, v2.h[1]      \n"
-                    "fmla   v30.8h, v4.8h, v2.h[2]      \n"
-                    "fmla   v31.8h, v4.8h, v2.h[3]      \n"
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    11f                         \n"
-
-                    // if out_elempack == 8
-                    "cmp    %w12, #8                    \n"
-                    "bne    8f                          \n"
-
-                    "st1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%3], #64 \n"
-                    "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%3], #64 \n"
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%3], #64 \n"
-                    "b      10f                         \n"
-
-                    // if out_elempack == 4
-                    "8:                                 \n"
-                    "cmp    %w12, #4                    \n"
-                    "bne    9f                          \n"
-
-                    "zip1   v12.2d, v20.2d, v21.2d      \n"
-                    "zip2   v18.2d, v20.2d, v21.2d      \n"
-                    "zip1   v13.2d, v22.2d, v23.2d      \n"
-                    "zip2   v19.2d, v22.2d, v23.2d      \n"
-                    "zip1   v14.2d, v24.2d, v25.2d      \n"
-                    "zip2   v20.2d, v24.2d, v25.2d      \n"
-                    "zip1   v15.2d, v26.2d, v27.2d      \n"
-                    "zip2   v21.2d, v26.2d, v27.2d      \n"
-                    "zip1   v16.2d, v28.2d, v29.2d      \n"
-                    "zip2   v22.2d, v28.2d, v29.2d      \n"
-                    "zip1   v17.2d, v30.2d, v31.2d      \n"
-                    "zip2   v23.2d, v30.2d, v31.2d      \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 1          \n"
-                    "st1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%3], #64 \n"
-                    "st1    {v16.8h, v17.8h}, [%3], #32 \n"
-                    "st1    {v18.8h, v19.8h, v20.8h, v21.8h}, [x4], #64 \n"
-                    "st1    {v22.8h, v23.8h}, [x4]      \n"
-                    "b      10f                         \n"
-
-                    // if out_elempack == 1
-                    "9:                                 \n"
-                    // transpose8x12
-                    "zip1   v18.8h, v20.8h, v21.8h      \n"
-                    "zip2   v19.8h, v20.8h, v21.8h      \n"
-                    "zip1   v20.8h, v22.8h, v23.8h      \n"
-                    "zip2   v21.8h, v22.8h, v23.8h      \n"
-                    "zip1   v22.8h, v24.8h, v25.8h      \n"
-                    "zip2   v23.8h, v24.8h, v25.8h      \n"
-                    "zip1   v24.8h, v26.8h, v27.8h      \n"
-                    "zip2   v25.8h, v26.8h, v27.8h      \n"
-                    "zip1   v26.8h, v28.8h, v29.8h      \n"
-                    "zip2   v27.8h, v28.8h, v29.8h      \n"
-                    "zip1   v28.8h, v30.8h, v31.8h      \n"
-                    "zip2   v29.8h, v30.8h, v31.8h      \n"
-
-                    "zip1   v0.4s, v18.4s, v20.4s       \n"
-                    "zip2   v3.4s, v18.4s, v20.4s       \n"
-                    "zip1   v6.4s, v19.4s, v21.4s       \n"
-                    "zip2   v9.4s, v19.4s, v21.4s       \n"
-                    "zip1   v1.4s, v22.4s, v24.4s       \n"
-                    "zip2   v4.4s, v22.4s, v24.4s       \n"
-                    "zip1   v7.4s, v23.4s, v25.4s       \n"
-                    "zip2   v10.4s, v23.4s, v25.4s      \n"
-                    "zip1   v2.4s, v26.4s, v28.4s       \n"
-                    "zip2   v5.4s, v26.4s, v28.4s       \n"
-                    "zip1   v8.4s, v27.4s, v29.4s       \n"
-                    "zip2   v11.4s, v27.4s, v29.4s      \n"
-
-                    "mov    v12.d[0], v0.d[1]           \n"
-                    "mov    v13.d[0], v1.d[1]           \n"
-                    "mov    v14.d[0], v2.d[1]           \n"
-                    "mov    v15.d[0], v3.d[1]           \n"
-                    "mov    v16.d[0], v4.d[1]           \n"
-                    "mov    v17.d[0], v5.d[1]           \n"
-                    "mov    v18.d[0], v6.d[1]           \n"
-                    "mov    v19.d[0], v7.d[1]           \n"
-                    "mov    v20.d[0], v8.d[1]           \n"
-                    "mov    v21.d[0], v9.d[1]           \n"
-                    "mov    v22.d[0], v10.d[1]          \n"
-                    "mov    v23.d[0], v11.d[1]          \n"
-
-                    "add    x4, %3, %w13, sxtw 1        \n"
-                    "st1    {v0.4h, v1.4h, v2.4h}, [%3], #24 \n"
-                    "st1    {v12.4h, v13.4h, v14.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v3.4h, v4.4h, v5.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v15.4h, v16.4h, v17.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v6.4h, v7.4h, v8.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v18.4h, v19.4h, v20.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v9.4h, v10.4h, v11.4h}, [x4] \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v21.4h, v22.4h, v23.4h}, [x4] \n"
-
-                    "10:                                \n"
-                    "add    %0, %0, #192                \n"
-                    "b      12f                         \n"
-
-                    "11:                                \n"
-                    "st1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%0], #64 \n"
-                    "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                    "12:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-#else  // NCNN_GNU_INLINE_ASM
-            float16x8_t _sum0;
-            float16x8_t _sum1;
-            float16x8_t _sum2;
-            float16x8_t _sum3;
-            float16x8_t _sum4;
-            float16x8_t _sum5;
-            float16x8_t _sum6;
-            float16x8_t _sum7;
-            float16x8_t _sum8;
-            float16x8_t _sum9;
-            float16x8_t _suma;
-            float16x8_t _sumb;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1q_f16(pC);
-                    _sum1 = _sum0;
-                    _sum2 = _sum0;
-                    _sum3 = _sum0;
-                    _sum4 = _sum0;
-                    _sum5 = _sum0;
-                    _sum6 = _sum0;
-                    _sum7 = _sum0;
-                    _sum8 = _sum0;
-                    _sum9 = _sum0;
-                    _suma = _sum0;
-                    _sumb = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f16(0.f);
-                    _sum1 = vdupq_n_f16(0.f);
-                    _sum2 = vdupq_n_f16(0.f);
-                    _sum3 = vdupq_n_f16(0.f);
-                    _sum4 = vdupq_n_f16(0.f);
-                    _sum5 = vdupq_n_f16(0.f);
-                    _sum6 = vdupq_n_f16(0.f);
-                    _sum7 = vdupq_n_f16(0.f);
-                    _sum8 = vdupq_n_f16(0.f);
-                    _sum9 = vdupq_n_f16(0.f);
-                    _suma = vdupq_n_f16(0.f);
-                    _sumb = vdupq_n_f16(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f16(outptr);
-                _sum1 = vld1q_f16(outptr + 8);
-                _sum2 = vld1q_f16(outptr + 8 * 2);
-                _sum3 = vld1q_f16(outptr + 8 * 3);
-                _sum4 = vld1q_f16(outptr + 8 * 4);
-                _sum5 = vld1q_f16(outptr + 8 * 5);
-                _sum6 = vld1q_f16(outptr + 8 * 6);
-                _sum7 = vld1q_f16(outptr + 8 * 7);
-                _sum8 = vld1q_f16(outptr + 8 * 8);
-                _sum9 = vld1q_f16(outptr + 8 * 9);
-                _suma = vld1q_f16(outptr + 8 * 10);
-                _sumb = vld1q_f16(outptr + 8 * 11);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x8_t _pA = vld1q_f16(pA);
-
-                float16x4_t _pB0 = vld1_f16(pB);
-                float16x4_t _pB1 = vld1_f16(pB + 4);
-                float16x4_t _pB2 = vld1_f16(pB + 8);
-
-                _sum0 = vfmaq_lane_f16(_sum0, _pA, _pB0, 0);
-                _sum1 = vfmaq_lane_f16(_sum1, _pA, _pB0, 1);
-                _sum2 = vfmaq_lane_f16(_sum2, _pA, _pB0, 2);
-                _sum3 = vfmaq_lane_f16(_sum3, _pA, _pB0, 3);
-                _sum4 = vfmaq_lane_f16(_sum4, _pA, _pB1, 0);
-                _sum5 = vfmaq_lane_f16(_sum5, _pA, _pB1, 1);
-                _sum6 = vfmaq_lane_f16(_sum6, _pA, _pB1, 2);
-                _sum7 = vfmaq_lane_f16(_sum7, _pA, _pB1, 3);
-                _sum8 = vfmaq_lane_f16(_sum8, _pA, _pB2, 0);
-                _sum9 = vfmaq_lane_f16(_sum9, _pA, _pB2, 1);
-                _suma = vfmaq_lane_f16(_suma, _pA, _pB2, 2);
-                _sumb = vfmaq_lane_f16(_sumb, _pA, _pB2, 3);
-
-                pA += 8;
-                pB += 12;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 8)
-                {
-                    vst1q_f16(outptr0, _sum0);
-                    vst1q_f16(outptr0 + 8, _sum1);
-                    vst1q_f16(outptr0 + 8 * 2, _sum2);
-                    vst1q_f16(outptr0 + 8 * 3, _sum3);
-                    vst1q_f16(outptr0 + 8 * 4, _sum4);
-                    vst1q_f16(outptr0 + 8 * 5, _sum5);
-                    vst1q_f16(outptr0 + 8 * 6, _sum6);
-                    vst1q_f16(outptr0 + 8 * 7, _sum7);
-                    vst1q_f16(outptr0 + 8 * 8, _sum8);
-                    vst1q_f16(outptr0 + 8 * 9, _sum9);
-                    vst1q_f16(outptr0 + 8 * 10, _suma);
-                    vst1q_f16(outptr0 + 8 * 11, _sumb);
-                    outptr0 += 96;
-                }
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr0, vget_low_f16(_sum0));
-                    vst1_f16(outptr0 + 4, vget_low_f16(_sum1));
-                    vst1_f16(outptr0 + 4 * 2, vget_low_f16(_sum2));
-                    vst1_f16(outptr0 + 4 * 3, vget_low_f16(_sum3));
-                    vst1_f16(outptr0 + 4 * 4, vget_low_f16(_sum4));
-                    vst1_f16(outptr0 + 4 * 5, vget_low_f16(_sum5));
-                    vst1_f16(outptr0 + 4 * 6, vget_low_f16(_sum6));
-                    vst1_f16(outptr0 + 4 * 7, vget_low_f16(_sum7));
-                    vst1_f16(outptr0 + 4 * 8, vget_low_f16(_sum8));
-                    vst1_f16(outptr0 + 4 * 9, vget_low_f16(_sum9));
-                    vst1_f16(outptr0 + 4 * 10, vget_low_f16(_suma));
-                    vst1_f16(outptr0 + 4 * 11, vget_low_f16(_sumb));
-
-                    vst1_f16(outptr0 + out_hstep * 4, vget_high_f16(_sum0));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4, vget_high_f16(_sum1));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 2, vget_high_f16(_sum2));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 3, vget_high_f16(_sum3));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 4, vget_high_f16(_sum4));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 5, vget_high_f16(_sum5));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 6, vget_high_f16(_sum6));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 7, vget_high_f16(_sum7));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 8, vget_high_f16(_sum8));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 9, vget_high_f16(_sum9));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 10, vget_high_f16(_suma));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 11, vget_high_f16(_sumb));
-
-                    outptr0 += 48;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose8x12_ph(_sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7, _sum8, _sum9, _suma, _sumb);
-
-                    vst1_f16(outptr0, vget_low_f16(_sum0));
-                    vst1_f16(outptr0 + 4, vget_high_f16(_sum0));
-                    vst1_f16(outptr0 + 8, vget_low_f16(_sum1));
-                    vst1_f16(outptr0 + out_hstep, vget_high_f16(_sum1));
-                    vst1_f16(outptr0 + out_hstep + 4, vget_low_f16(_sum2));
-                    vst1_f16(outptr0 + out_hstep + 8, vget_high_f16(_sum2));
-                    vst1_f16(outptr0 + out_hstep * 2, vget_low_f16(_sum3));
-                    vst1_f16(outptr0 + out_hstep * 2 + 4, vget_high_f16(_sum3));
-                    vst1_f16(outptr0 + out_hstep * 2 + 8, vget_low_f16(_sum4));
-                    vst1_f16(outptr0 + out_hstep * 3, vget_high_f16(_sum4));
-                    vst1_f16(outptr0 + out_hstep * 3 + 4, vget_low_f16(_sum5));
-                    vst1_f16(outptr0 + out_hstep * 3 + 8, vget_high_f16(_sum5));
-                    vst1_f16(outptr0 + out_hstep * 4, vget_low_f16(_sum6));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4, vget_high_f16(_sum6));
-                    vst1_f16(outptr0 + out_hstep * 4 + 8, vget_low_f16(_sum7));
-                    vst1_f16(outptr0 + out_hstep * 5, vget_high_f16(_sum7));
-                    vst1_f16(outptr0 + out_hstep * 5 + 4, vget_low_f16(_sum8));
-                    vst1_f16(outptr0 + out_hstep * 5 + 8, vget_high_f16(_sum8));
-                    vst1_f16(outptr0 + out_hstep * 6, vget_low_f16(_sum9));
-                    vst1_f16(outptr0 + out_hstep * 6 + 4, vget_high_f16(_sum9));
-                    vst1_f16(outptr0 + out_hstep * 6 + 8, vget_low_f16(_suma));
-                    vst1_f16(outptr0 + out_hstep * 7, vget_high_f16(_suma));
-                    vst1_f16(outptr0 + out_hstep * 7 + 4, vget_low_f16(_sumb));
-                    vst1_f16(outptr0 + out_hstep * 7 + 8, vget_high_f16(_sumb));
-
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                vst1q_f16(outptr, _sum0);
-                vst1q_f16(outptr + 8, _sum1);
-                vst1q_f16(outptr + 8 * 2, _sum2);
-                vst1q_f16(outptr + 8 * 3, _sum3);
-                vst1q_f16(outptr + 8 * 4, _sum4);
-                vst1q_f16(outptr + 8 * 5, _sum5);
-                vst1q_f16(outptr + 8 * 6, _sum6);
-                vst1q_f16(outptr + 8 * 7, _sum7);
-                vst1q_f16(outptr + 8 * 8, _sum8);
-                vst1q_f16(outptr + 8 * 9, _sum9);
-                vst1q_f16(outptr + 8 * 10, _suma);
-                vst1q_f16(outptr + 8 * 11, _sumb);
-            }
-
-            outptr += 96;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            const __fp16* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            if (use_a53_a55_optimized_kernel)
-            {
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                    "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0]      \n"
-                    "subs   %0, %0, #64                 \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "ld1    {v24.8h}, [%8]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v25.16b, v24.16b            \n"
-                    "mov    v26.16b, v24.16b            \n"
-                    "mov    v27.16b, v24.16b            \n"
-                    "mov    v28.16b, v24.16b            \n"
-                    "mov    v29.16b, v24.16b            \n"
-                    "mov    v30.16b, v24.16b            \n"
-                    "mov    v31.16b, v24.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.8h}, [%1], #16          \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.8h}, [%2], #16          \n"
-
-                    "ldr    d5, [%1], #8                \n"
-                    "ldr    x25, [%1], #8               \n"
-
-                    ".align 4                           \n"
-                    "4:                                 \n"
-                    "ldr    d1, [%2], #8                \n"
-                    "fmla   v24.8h, v4.8h, v0.h[0]      \n"
-                    "ldr    x21, [%2], #8               \n"
-                    "fmla   v25.8h, v4.8h, v0.h[1]      \n"
-                    "ins    v5.d[1], x25                \n"
-                    "fmla   v26.8h, v4.8h, v0.h[2]      \n"
-                    "ldr    d6, [%1], #8                \n"
-                    "fmla   v27.8h, v4.8h, v0.h[3]      \n"
-                    "ldr    x26, [%1], #8               \n"
-                    "fmla   v28.8h, v4.8h, v0.h[4]      \n"
-                    "ldr    d2, [%2], #8                \n"
-                    "fmla   v29.8h, v4.8h, v0.h[5]      \n"
-                    "ins    v1.d[1], x21                \n"
-                    "fmla   v30.8h, v4.8h, v0.h[6]      \n"
-                    "ldr    x22, [%2], #8               \n"
-                    "fmla   v31.8h, v4.8h, v0.h[7]      \n"
-                    "ldr    d7, [%1], #8                \n"
-                    "fmla   v24.8h, v5.8h, v1.h[0]      \n"
-                    "ldr    x27, [%1], #8               \n"
-                    "fmla   v25.8h, v5.8h, v1.h[1]      \n"
-                    "ins    v6.d[1], x26                \n"
-                    "fmla   v26.8h, v5.8h, v1.h[2]      \n"
-                    "ldr    d3, [%2], #8                \n"
-                    "fmla   v27.8h, v5.8h, v1.h[3]      \n"
-                    "ldr    x23, [%2], #8               \n"
-                    "fmla   v28.8h, v5.8h, v1.h[4]      \n"
-                    "prfm   pldl1keep, [%1, #512]       \n" // NOTE PRELOAD
-                    "fmla   v29.8h, v5.8h, v1.h[5]      \n"
-                    "ins    v2.d[1], x22                \n"
-                    "fmla   v30.8h, v5.8h, v1.h[6]      \n"
-                    "ldr    d4, [%1], #8                \n"
-                    "fmla   v31.8h, v5.8h, v1.h[7]      \n"
-                    "ldr    x24, [%1], #8               \n"
-                    "fmla   v24.8h, v6.8h, v2.h[0]      \n"
-                    "prfm   pldl1keep, [%2, #512]       \n" // NOTE PRELOAD
-                    "fmla   v25.8h, v6.8h, v2.h[1]      \n"
-                    "ins    v7.d[1], x27                \n"
-                    "fmla   v26.8h, v6.8h, v2.h[2]      \n"
-                    "ldr    d0, [%2], #8                \n"
-                    "fmla   v27.8h, v6.8h, v2.h[3]      \n"
-                    "ldr    x20, [%2], #8               \n"
-                    "fmla   v28.8h, v6.8h, v2.h[4]      \n"
-                    "ldr    d5, [%1], #8                \n"
-                    "fmla   v29.8h, v6.8h, v2.h[5]      \n"
-                    "ins    v3.d[1], x23                \n"
-                    "fmla   v30.8h, v6.8h, v2.h[6]      \n"
-                    "ldr    x25, [%1], #8               \n"
-                    "fmla   v31.8h, v6.8h, v2.h[7]      \n"
-                    "fmla   v24.8h, v7.8h, v3.h[0]      \n"
-                    "fmla   v25.8h, v7.8h, v3.h[1]      \n"
-                    "fmla   v26.8h, v7.8h, v3.h[2]      \n"
-                    "ins    v4.d[1], x24                \n"
-                    "fmla   v27.8h, v7.8h, v3.h[3]      \n"
-                    "fmla   v28.8h, v7.8h, v3.h[4]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v29.8h, v7.8h, v3.h[5]      \n"
-                    "fmla   v30.8h, v7.8h, v3.h[6]      \n"
-                    "ins    v0.d[1], x20                \n"
-                    "fmla   v31.8h, v7.8h, v3.h[7]      \n"
-                    "bne    4b                          \n"
-
-                    "sub    %1, %1, #32                 \n"
-                    "sub    %2, %2, #16                 \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v0.8h}, [%2], #16          \n"
-                    "ld1    {v4.8h}, [%1], #16          \n"
-                    "fmla   v24.8h, v4.8h, v0.h[0]      \n"
-                    "fmla   v25.8h, v4.8h, v0.h[1]      \n"
-                    "fmla   v26.8h, v4.8h, v0.h[2]      \n"
-                    "fmla   v27.8h, v4.8h, v0.h[3]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v28.8h, v4.8h, v0.h[4]      \n"
-                    "fmla   v29.8h, v4.8h, v0.h[5]      \n"
-                    "fmla   v30.8h, v4.8h, v0.h[6]      \n"
-                    "fmla   v31.8h, v4.8h, v0.h[7]      \n"
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    11f                         \n"
-
-                    // if out_elempack == 8
-                    "cmp    %w12, #8                    \n"
-                    "bne    8f                          \n"
-
-                    "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%3], #64 \n"
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%3], #64 \n"
-                    "b      10f                         \n"
-
-                    // if out_elempack == 4
-                    "8:                                 \n"
-                    "cmp    %w12, #4                    \n"
-                    "bne    9f                          \n"
-
-                    "zip1   v16.2d, v24.2d, v25.2d      \n"
-                    "zip2   v20.2d, v24.2d, v25.2d      \n"
-                    "zip1   v17.2d, v26.2d, v27.2d      \n"
-                    "zip2   v21.2d, v26.2d, v27.2d      \n"
-                    "zip1   v18.2d, v28.2d, v29.2d      \n"
-                    "zip2   v22.2d, v28.2d, v29.2d      \n"
-                    "zip1   v19.2d, v30.2d, v31.2d      \n"
-                    "zip2   v23.2d, v30.2d, v31.2d      \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 1          \n"
-                    "st1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%3], #64 \n"
-                    "st1    {v20.8h, v21.8h, v22.8h, v23.8h}, [x4] \n"
-                    "b      10f                         \n"
-
-                    // if out_elempack == 1
-                    "9:                                 \n"
-                    // transpose8x8
-                    "zip1   v22.8h, v24.8h, v25.8h      \n"
-                    "zip2   v23.8h, v24.8h, v25.8h      \n"
-                    "zip1   v24.8h, v26.8h, v27.8h      \n"
-                    "zip2   v25.8h, v26.8h, v27.8h      \n"
-                    "zip1   v26.8h, v28.8h, v29.8h      \n"
-                    "zip2   v27.8h, v28.8h, v29.8h      \n"
-                    "zip1   v28.8h, v30.8h, v31.8h      \n"
-                    "zip2   v29.8h, v30.8h, v31.8h      \n"
-
-                    "zip1   v16.4s, v22.4s, v24.4s      \n"
-                    "zip2   v17.4s, v22.4s, v24.4s      \n"
-                    "zip1   v18.4s, v23.4s, v25.4s      \n"
-                    "zip2   v19.4s, v23.4s, v25.4s      \n"
-                    "zip1   v20.4s, v26.4s, v28.4s      \n"
-                    "zip2   v21.4s, v26.4s, v28.4s      \n"
-                    "zip1   v22.4s, v27.4s, v29.4s      \n"
-                    "zip2   v23.4s, v27.4s, v29.4s      \n"
-
-                    "zip1   v24.2d, v16.2d, v20.2d      \n"
-                    "zip2   v25.2d, v16.2d, v20.2d      \n"
-                    "zip1   v26.2d, v17.2d, v21.2d      \n"
-                    "zip2   v27.2d, v17.2d, v21.2d      \n"
-                    "zip1   v28.2d, v18.2d, v22.2d      \n"
-                    "zip2   v29.2d, v18.2d, v22.2d      \n"
-                    "zip1   v30.2d, v19.2d, v23.2d      \n"
-                    "zip2   v31.2d, v19.2d, v23.2d      \n"
-
-                    "add    x4, %3, %w13, sxtw 1        \n"
-                    "st1    {v24.8h}, [%3], #16         \n"
-                    "st1    {v25.8h}, [x4]              \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v26.8h}, [x4]              \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v27.8h}, [x4]              \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v28.8h}, [x4]              \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v29.8h}, [x4]              \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v30.8h}, [x4]              \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v31.8h}, [x4]              \n"
-
-                    "10:                                \n"
-                    "add    %0, %0, #128                \n"
-                    "b      12f                         \n"
-
-                    "11:                                \n"
-                    "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                    "12:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            else
-            {
-                asm volatile(
-                    "cbz    %w10, 0f                    \n"
-
-                    "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                    "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0]      \n"
-                    "subs   %0, %0, #64                 \n"
-                    "b      3f                          \n"
-
-                    "0:                                 \n"
-                    // if pC
-                    "cbz    %8, 1f                      \n"
-
-                    "ld1    {v24.8h}, [%8]              \n"
-                    "b      2f                          \n"
-
-                    // else
-                    "1:                                 \n"
-                    "eor    v24.16b, v24.16b, v24.16b   \n"
-
-                    "2:                                 \n"
-                    "mov    v25.16b, v24.16b            \n"
-                    "mov    v26.16b, v24.16b            \n"
-                    "mov    v27.16b, v24.16b            \n"
-                    "mov    v28.16b, v24.16b            \n"
-                    "mov    v29.16b, v24.16b            \n"
-                    "mov    v30.16b, v24.16b            \n"
-                    "mov    v31.16b, v24.16b            \n"
-
-                    "3:                                 \n"
-                    "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                    "cmp    w4, #0                      \n"
-                    "beq    5f                          \n"
-
-                    "4:                                 \n"
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%2], #64 \n"
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%1], #64 \n"
-                    "fmla   v24.8h, v4.8h, v0.h[0]      \n"
-                    "fmla   v25.8h, v4.8h, v0.h[1]      \n"
-                    "fmla   v26.8h, v4.8h, v0.h[2]      \n"
-                    "fmla   v27.8h, v4.8h, v0.h[3]      \n"
-                    "fmla   v28.8h, v4.8h, v0.h[4]      \n"
-                    "fmla   v29.8h, v4.8h, v0.h[5]      \n"
-                    "fmla   v30.8h, v4.8h, v0.h[6]      \n"
-                    "fmla   v31.8h, v4.8h, v0.h[7]      \n"
-                    "fmla   v24.8h, v5.8h, v1.h[0]      \n"
-                    "fmla   v25.8h, v5.8h, v1.h[1]      \n"
-                    "fmla   v26.8h, v5.8h, v1.h[2]      \n"
-                    "fmla   v27.8h, v5.8h, v1.h[3]      \n"
-                    "fmla   v28.8h, v5.8h, v1.h[4]      \n"
-                    "fmla   v29.8h, v5.8h, v1.h[5]      \n"
-                    "fmla   v30.8h, v5.8h, v1.h[6]      \n"
-                    "fmla   v31.8h, v5.8h, v1.h[7]      \n"
-                    "fmla   v24.8h, v6.8h, v2.h[0]      \n"
-                    "fmla   v25.8h, v6.8h, v2.h[1]      \n"
-                    "fmla   v26.8h, v6.8h, v2.h[2]      \n"
-                    "fmla   v27.8h, v6.8h, v2.h[3]      \n"
-                    "fmla   v28.8h, v6.8h, v2.h[4]      \n"
-                    "fmla   v29.8h, v6.8h, v2.h[5]      \n"
-                    "fmla   v30.8h, v6.8h, v2.h[6]      \n"
-                    "fmla   v31.8h, v6.8h, v2.h[7]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v24.8h, v7.8h, v3.h[0]      \n"
-                    "fmla   v25.8h, v7.8h, v3.h[1]      \n"
-                    "fmla   v26.8h, v7.8h, v3.h[2]      \n"
-                    "fmla   v27.8h, v7.8h, v3.h[3]      \n"
-                    "fmla   v28.8h, v7.8h, v3.h[4]      \n"
-                    "fmla   v29.8h, v7.8h, v3.h[5]      \n"
-                    "fmla   v30.8h, v7.8h, v3.h[6]      \n"
-                    "fmla   v31.8h, v7.8h, v3.h[7]      \n"
-                    "bne    4b                          \n"
-
-                    "5:                                 \n"
-                    "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                    "cmp    w4, #0                      \n"
-                    "beq    7f                          \n"
-
-                    "6:                                 \n"
-                    "ld1    {v0.8h}, [%2], #16          \n"
-                    "ld1    {v4.8h}, [%1], #16          \n"
-                    "fmla   v24.8h, v4.8h, v0.h[0]      \n"
-                    "fmla   v25.8h, v4.8h, v0.h[1]      \n"
-                    "fmla   v26.8h, v4.8h, v0.h[2]      \n"
-                    "fmla   v27.8h, v4.8h, v0.h[3]      \n"
-                    "subs   w4, w4, #1                  \n"
-                    "fmla   v28.8h, v4.8h, v0.h[4]      \n"
-                    "fmla   v29.8h, v4.8h, v0.h[5]      \n"
-                    "fmla   v30.8h, v4.8h, v0.h[6]      \n"
-                    "fmla   v31.8h, v4.8h, v0.h[7]      \n"
-                    "bne    6b                          \n"
-
-                    "7:                                 \n"
-                    "tst    %w11, #255                  \n"
-                    "beq    11f                         \n"
-
-                    // if out_elempack == 8
-                    "cmp    %w12, #8                    \n"
-                    "bne    8f                          \n"
-
-                    "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%3], #64 \n"
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%3], #64 \n"
-                    "b      10f                         \n"
-
-                    // if out_elempack == 4
-                    "8:                                 \n"
-                    "cmp    %w12, #4                    \n"
-                    "bne    9f                          \n"
-
-                    "zip1   v16.2d, v24.2d, v25.2d      \n"
-                    "zip2   v20.2d, v24.2d, v25.2d      \n"
-                    "zip1   v17.2d, v26.2d, v27.2d      \n"
-                    "zip2   v21.2d, v26.2d, v27.2d      \n"
-                    "zip1   v18.2d, v28.2d, v29.2d      \n"
-                    "zip2   v22.2d, v28.2d, v29.2d      \n"
-                    "zip1   v19.2d, v30.2d, v31.2d      \n"
-                    "zip2   v23.2d, v30.2d, v31.2d      \n"
-
-                    "lsl    w4, %w13, #2                \n"
-                    "add    x4, %3, w4, sxtw 1          \n"
-                    "st1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%3], #64 \n"
-                    "st1    {v20.8h, v21.8h, v22.8h, v23.8h}, [x4] \n"
-                    "b      10f                         \n"
-
-                    // if out_elempack == 1
-                    "9:                                 \n"
-                    // transpose8x8
-                    "zip1   v22.8h, v24.8h, v25.8h      \n"
-                    "zip2   v23.8h, v24.8h, v25.8h      \n"
-                    "zip1   v24.8h, v26.8h, v27.8h      \n"
-                    "zip2   v25.8h, v26.8h, v27.8h      \n"
-                    "zip1   v26.8h, v28.8h, v29.8h      \n"
-                    "zip2   v27.8h, v28.8h, v29.8h      \n"
-                    "zip1   v28.8h, v30.8h, v31.8h      \n"
-                    "zip2   v29.8h, v30.8h, v31.8h      \n"
-
-                    "zip1   v16.4s, v22.4s, v24.4s      \n"
-                    "zip2   v17.4s, v22.4s, v24.4s      \n"
-                    "zip1   v18.4s, v23.4s, v25.4s      \n"
-                    "zip2   v19.4s, v23.4s, v25.4s      \n"
-                    "zip1   v20.4s, v26.4s, v28.4s      \n"
-                    "zip2   v21.4s, v26.4s, v28.4s      \n"
-                    "zip1   v22.4s, v27.4s, v29.4s      \n"
-                    "zip2   v23.4s, v27.4s, v29.4s      \n"
-
-                    "zip1   v24.2d, v16.2d, v20.2d      \n"
-                    "zip2   v25.2d, v16.2d, v20.2d      \n"
-                    "zip1   v26.2d, v17.2d, v21.2d      \n"
-                    "zip2   v27.2d, v17.2d, v21.2d      \n"
-                    "zip1   v28.2d, v18.2d, v22.2d      \n"
-                    "zip2   v29.2d, v18.2d, v22.2d      \n"
-                    "zip1   v30.2d, v19.2d, v23.2d      \n"
-                    "zip2   v31.2d, v19.2d, v23.2d      \n"
-
-                    "add    x4, %3, %w13, sxtw 1        \n"
-                    "st1    {v24.8h}, [%3], #16         \n"
-                    "st1    {v25.8h}, [x4]              \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v26.8h}, [x4]              \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v27.8h}, [x4]              \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v28.8h}, [x4]              \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v29.8h}, [x4]              \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v30.8h}, [x4]              \n"
-                    "add    x4, x4, %w13, sxtw 1        \n"
-                    "st1    {v31.8h}, [x4]              \n"
-
-                    "10:                                \n"
-                    "add    %0, %0, #128                \n"
-                    "b      12f                         \n"
-
-                    "11:                                \n"
-                    "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                    "12:                                \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(pA),     // %1
-                    "=r"(pB),     // %2
-                    "=r"(outptr0) // %3
-                    : "0"(outptr),
-                    "1"(pA),
-                    "2"(pB),
-                    "3"(outptr0),
-                    "r"(pC),           // %8
-                    "r"(max_kk),       // %9
-                    "r"(k),            // %10
-                    "r"(k_end),        // %11
-                    "r"(out_elempack), // %12
-                    "r"(out_hstep)     // %13
-                    : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-#else  // NCNN_GNU_INLINE_ASM
-            float16x8_t _sum0;
-            float16x8_t _sum1;
-            float16x8_t _sum2;
-            float16x8_t _sum3;
-            float16x8_t _sum4;
-            float16x8_t _sum5;
-            float16x8_t _sum6;
-            float16x8_t _sum7;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1q_f16(pC);
-                    _sum1 = _sum0;
-                    _sum2 = _sum0;
-                    _sum3 = _sum0;
-                    _sum4 = _sum0;
-                    _sum5 = _sum0;
-                    _sum6 = _sum0;
-                    _sum7 = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f16(0.f);
-                    _sum1 = vdupq_n_f16(0.f);
-                    _sum2 = vdupq_n_f16(0.f);
-                    _sum3 = vdupq_n_f16(0.f);
-                    _sum4 = vdupq_n_f16(0.f);
-                    _sum5 = vdupq_n_f16(0.f);
-                    _sum6 = vdupq_n_f16(0.f);
-                    _sum7 = vdupq_n_f16(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f16(outptr);
-                _sum1 = vld1q_f16(outptr + 8);
-                _sum2 = vld1q_f16(outptr + 8 * 2);
-                _sum3 = vld1q_f16(outptr + 8 * 3);
-                _sum4 = vld1q_f16(outptr + 8 * 4);
-                _sum5 = vld1q_f16(outptr + 8 * 5);
-                _sum6 = vld1q_f16(outptr + 8 * 6);
-                _sum7 = vld1q_f16(outptr + 8 * 7);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x8_t _pA = vld1q_f16(pA);
-
-                float16x8_t _pB = vld1q_f16(pB);
-
-                _sum0 = vfmaq_laneq_f16(_sum0, _pA, _pB, 0);
-                _sum1 = vfmaq_laneq_f16(_sum1, _pA, _pB, 1);
-                _sum2 = vfmaq_laneq_f16(_sum2, _pA, _pB, 2);
-                _sum3 = vfmaq_laneq_f16(_sum3, _pA, _pB, 3);
-                _sum4 = vfmaq_laneq_f16(_sum4, _pA, _pB, 4);
-                _sum5 = vfmaq_laneq_f16(_sum5, _pA, _pB, 5);
-                _sum6 = vfmaq_laneq_f16(_sum6, _pA, _pB, 6);
-                _sum7 = vfmaq_laneq_f16(_sum7, _pA, _pB, 7);
-
-                pA += 8;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 8)
-                {
-                    vst1q_f16(outptr0, _sum0);
-                    vst1q_f16(outptr0 + 8, _sum1);
-                    vst1q_f16(outptr0 + 8 * 2, _sum2);
-                    vst1q_f16(outptr0 + 8 * 3, _sum3);
-                    vst1q_f16(outptr0 + 8 * 4, _sum4);
-                    vst1q_f16(outptr0 + 8 * 5, _sum5);
-                    vst1q_f16(outptr0 + 8 * 6, _sum6);
-                    vst1q_f16(outptr0 + 8 * 7, _sum7);
-                    outptr0 += 64;
-                }
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr0, vget_low_f16(_sum0));
-                    vst1_f16(outptr0 + 4, vget_low_f16(_sum1));
-                    vst1_f16(outptr0 + 4 * 2, vget_low_f16(_sum2));
-                    vst1_f16(outptr0 + 4 * 3, vget_low_f16(_sum3));
-                    vst1_f16(outptr0 + 4 * 4, vget_low_f16(_sum4));
-                    vst1_f16(outptr0 + 4 * 5, vget_low_f16(_sum5));
-                    vst1_f16(outptr0 + 4 * 6, vget_low_f16(_sum6));
-                    vst1_f16(outptr0 + 4 * 7, vget_low_f16(_sum7));
-
-                    vst1_f16(outptr0 + out_hstep * 4, vget_high_f16(_sum0));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4, vget_high_f16(_sum1));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 2, vget_high_f16(_sum2));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 3, vget_high_f16(_sum3));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 4, vget_high_f16(_sum4));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 5, vget_high_f16(_sum5));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 6, vget_high_f16(_sum6));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 7, vget_high_f16(_sum7));
-
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose8x8_ph(_sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7);
-
-                    vst1q_f16(outptr0, _sum0);
-                    vst1q_f16(outptr0 + out_hstep, _sum1);
-                    vst1q_f16(outptr0 + out_hstep * 2, _sum2);
-                    vst1q_f16(outptr0 + out_hstep * 3, _sum3);
-                    vst1q_f16(outptr0 + out_hstep * 4, _sum4);
-                    vst1q_f16(outptr0 + out_hstep * 5, _sum5);
-                    vst1q_f16(outptr0 + out_hstep * 6, _sum6);
-                    vst1q_f16(outptr0 + out_hstep * 7, _sum7);
-
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1q_f16(outptr, _sum0);
-                vst1q_f16(outptr + 8, _sum1);
-                vst1q_f16(outptr + 8 * 2, _sum2);
-                vst1q_f16(outptr + 8 * 3, _sum3);
-                vst1q_f16(outptr + 8 * 4, _sum4);
-                vst1q_f16(outptr + 8 * 5, _sum5);
-                vst1q_f16(outptr + 8 * 6, _sum6);
-                vst1q_f16(outptr + 8 * 7, _sum7);
-            }
-
-            outptr += 64;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            const __fp16* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0]      \n"
-                "b      3f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "ld1    {v28.8h}, [%8]              \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-
-                "2:                                 \n"
-                "mov    v29.16b, v28.16b            \n"
-                "mov    v30.16b, v28.16b            \n"
-                "mov    v31.16b, v28.16b            \n"
-
-                "3:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                "4:                                 \n"
-                "prfm   pldl1keep, [%2, #256]       \n"
-                "ld1    {v0.8h, v1.8h}, [%2], #32   \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%1], #64 \n"
-                "fmla   v28.8h, v4.8h, v0.h[0]      \n"
-                "fmla   v29.8h, v4.8h, v0.h[1]      \n"
-                "fmla   v30.8h, v4.8h, v0.h[2]      \n"
-                "fmla   v31.8h, v4.8h, v0.h[3]      \n"
-                "fmla   v28.8h, v5.8h, v0.h[4]      \n"
-                "fmla   v29.8h, v5.8h, v0.h[5]      \n"
-                "fmla   v30.8h, v5.8h, v0.h[6]      \n"
-                "fmla   v31.8h, v5.8h, v0.h[7]      \n"
-                "fmla   v28.8h, v6.8h, v1.h[0]      \n"
-                "fmla   v29.8h, v6.8h, v1.h[1]      \n"
-                "fmla   v30.8h, v6.8h, v1.h[2]      \n"
-                "fmla   v31.8h, v6.8h, v1.h[3]      \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v28.8h, v7.8h, v1.h[4]      \n"
-                "fmla   v29.8h, v7.8h, v1.h[5]      \n"
-                "fmla   v30.8h, v7.8h, v1.h[6]      \n"
-                "fmla   v31.8h, v7.8h, v1.h[7]      \n"
-                "bne    4b                          \n"
-
-                "5:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    7f                          \n"
-
-                "6:                                 \n"
-                "ld1    {v0.4h}, [%2], #8           \n"
-                "ld1    {v4.8h}, [%1], #16          \n"
-                "fmla   v28.8h, v4.8h, v0.h[0]      \n"
-                "fmla   v29.8h, v4.8h, v0.h[1]      \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v30.8h, v4.8h, v0.h[2]      \n"
-                "fmla   v31.8h, v4.8h, v0.h[3]      \n"
-                "bne    6b                          \n"
-
-                "7:                                 \n"
-                "tst    %w11, #255                  \n"
-                "beq    11f                         \n"
-
-                // if out_elempack == 8
-                "cmp    %w12, #8                    \n"
-                "bne    8f                          \n"
-
-                "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%3], #64 \n"
-                "b      10f                         \n"
-
-                // if out_elempack == 4
-                "8:                                 \n"
-                "cmp    %w12, #4                    \n"
-                "bne    9f                          \n"
-
-                "zip1   v24.2d, v28.2d, v29.2d      \n"
-                "zip2   v26.2d, v28.2d, v29.2d      \n"
-                "zip1   v25.2d, v30.2d, v31.2d      \n"
-                "zip2   v27.2d, v30.2d, v31.2d      \n"
-
-                "lsl    w4, %w13, #2                \n"
-                "add    x4, %3, w4, sxtw 1          \n"
-                "st1    {v24.8h, v25.8h}, [%3], #32 \n"
-                "st1    {v26.8h, v27.8h}, [x4]      \n"
-                "b      10f                         \n"
-
-                // if out_elempack == 1
-                "9:                                 \n"
-                // transpose8x4
-                "zip1   v24.8h, v28.8h, v29.8h      \n"
-                "zip2   v25.8h, v28.8h, v29.8h      \n"
-                "zip1   v26.8h, v30.8h, v31.8h      \n"
-                "zip2   v27.8h, v30.8h, v31.8h      \n"
-
-                "zip1   v20.4s, v24.4s, v26.4s      \n"
-                "zip2   v22.4s, v24.4s, v26.4s      \n"
-                "zip1   v24.4s, v25.4s, v27.4s      \n"
-                "zip2   v26.4s, v25.4s, v27.4s      \n"
-
-                "mov    v21.d[0], v20.d[1]          \n"
-                "mov    v23.d[0], v22.d[1]          \n"
-                "mov    v25.d[0], v24.d[1]          \n"
-                "mov    v27.d[0], v26.d[1]          \n"
-
-                "add    x4, %3, %w13, sxtw 1        \n"
-                "st1    {v20.4h}, [%3], #8          \n"
-                "st1    {v21.4h}, [x4]              \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v22.4h}, [x4]              \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v23.4h}, [x4]              \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v24.4h}, [x4]              \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v25.4h}, [x4]              \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v26.4h}, [x4]              \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v27.4h}, [x4]              \n"
-
-                "10:                                \n"
-                "add    %0, %0, #64                 \n"
-                "b      12f                         \n"
-
-                "11:                                \n"
-                "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                "12:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v1", "v4", "v5", "v6", "v7", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-            float16x8_t _sum0;
-            float16x8_t _sum1;
-            float16x8_t _sum2;
-            float16x8_t _sum3;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1q_f16(pC);
-                    _sum1 = _sum0;
-                    _sum2 = _sum0;
-                    _sum3 = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f16(0.f);
-                    _sum1 = vdupq_n_f16(0.f);
-                    _sum2 = vdupq_n_f16(0.f);
-                    _sum3 = vdupq_n_f16(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f16(outptr);
-                _sum1 = vld1q_f16(outptr + 8);
-                _sum2 = vld1q_f16(outptr + 8 * 2);
-                _sum3 = vld1q_f16(outptr + 8 * 3);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x8_t _pA = vld1q_f16(pA);
-
-                float16x4_t _pB = vld1_f16(pB);
-
-                _sum0 = vfmaq_lane_f16(_sum0, _pA, _pB, 0);
-                _sum1 = vfmaq_lane_f16(_sum1, _pA, _pB, 1);
-                _sum2 = vfmaq_lane_f16(_sum2, _pA, _pB, 2);
-                _sum3 = vfmaq_lane_f16(_sum3, _pA, _pB, 3);
-
-                pA += 8;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 8)
-                {
-                    vst1q_f16(outptr0, _sum0);
-                    vst1q_f16(outptr0 + 8, _sum1);
-                    vst1q_f16(outptr0 + 8 * 2, _sum2);
-                    vst1q_f16(outptr0 + 8 * 3, _sum3);
-                    outptr0 += 32;
-                }
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr0, vget_low_f16(_sum0));
-                    vst1_f16(outptr0 + 4, vget_low_f16(_sum1));
-                    vst1_f16(outptr0 + 4 * 2, vget_low_f16(_sum2));
-                    vst1_f16(outptr0 + 4 * 3, vget_low_f16(_sum3));
-
-                    vst1_f16(outptr0 + out_hstep * 4, vget_high_f16(_sum0));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4, vget_high_f16(_sum1));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 2, vget_high_f16(_sum2));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4 * 3, vget_high_f16(_sum3));
-
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose8x4_ph(_sum0, _sum1, _sum2, _sum3);
-
-                    vst1_f16(outptr0, vget_low_f16(_sum0));
-                    vst1_f16(outptr0 + out_hstep * 1, vget_high_f16(_sum0));
-                    vst1_f16(outptr0 + out_hstep * 2, vget_low_f16(_sum1));
-                    vst1_f16(outptr0 + out_hstep * 3, vget_high_f16(_sum1));
-                    vst1_f16(outptr0 + out_hstep * 4, vget_low_f16(_sum2));
-                    vst1_f16(outptr0 + out_hstep * 5, vget_high_f16(_sum2));
-                    vst1_f16(outptr0 + out_hstep * 6, vget_low_f16(_sum3));
-                    vst1_f16(outptr0 + out_hstep * 7, vget_high_f16(_sum3));
-
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1q_f16(outptr, _sum0);
-                vst1q_f16(outptr + 8, _sum1);
-                vst1q_f16(outptr + 8 * 2, _sum2);
-                vst1q_f16(outptr + 8 * 3, _sum3);
-            }
-
-            outptr += 32;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            const __fp16* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            asm volatile(
-                "cbz    %w10, 0f                    \n"
-
-                "ld1    {v30.8h, v31.8h}, [%0]      \n"
-                "b      3f                          \n"
-
-                "0:                                 \n"
-                // if pC
-                "cbz    %8, 1f                      \n"
-
-                "ld1    {v30.8h}, [%8]              \n"
-                "b      2f                          \n"
-
-                // else
-                "1:                                 \n"
-                "eor    v30.16b, v30.16b, v30.16b   \n"
-
-                "2:                                 \n"
-                "mov    v31.16b, v30.16b            \n"
-
-                "3:                                 \n"
-                "lsr    w4, %w9, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-                "eor    v29.16b, v29.16b, v29.16b   \n"
-                "4:                                 \n"
-                "prfm   pldl1keep, [%2, #128]       \n"
-                "ld1    {v0.8h}, [%2], #16          \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%1], #64 \n"
-                "fmla   v28.8h, v4.8h, v0.h[0]      \n"
-                "fmla   v29.8h, v4.8h, v0.h[1]      \n"
-                "fmla   v30.8h, v5.8h, v0.h[2]      \n"
-                "fmla   v31.8h, v5.8h, v0.h[3]      \n"
-                "fmla   v28.8h, v6.8h, v0.h[4]      \n"
-                "fmla   v29.8h, v6.8h, v0.h[5]      \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v30.8h, v7.8h, v0.h[6]      \n"
-                "fmla   v31.8h, v7.8h, v0.h[7]      \n"
-                "bne    4b                          \n"
-                "fadd   v30.8h, v30.8h, v28.8h      \n"
-                "fadd   v31.8h, v31.8h, v29.8h      \n"
-
-                "5:                                 \n"
-                "and    w4, %w9, #3                 \n" // w4 = remain = max_kk & 3
-                "cmp    w4, #0                      \n"
-                "beq    7f                          \n"
-
-                "6:                                 \n"
-                "ld1    {v0.s}[0], [%2], #4         \n"
-                "ld1    {v4.8h}, [%1], #16          \n"
-                "subs   w4, w4, #1                  \n"
-                "fmla   v30.8h, v4.8h, v0.h[0]      \n"
-                "fmla   v31.8h, v4.8h, v0.h[1]      \n"
-                "bne    6b                          \n"
-
-                "7:                                 \n"
-                "tst    %w11, #255                  \n"
-                "beq    11f                         \n"
-
-                // if out_elempack == 8
-                "cmp    %w12, #8                    \n"
-                "bne    8f                          \n"
-
-                "st1    {v30.8h, v31.8h}, [%3], #32 \n"
-                "b      10f                         \n"
-
-                // if out_elempack == 4
-                "8:                                 \n"
-                "cmp    %w12, #4                    \n"
-                "bne    9f                          \n"
-
-                "zip1   v28.2d, v30.2d, v31.2d      \n"
-                "zip2   v29.2d, v30.2d, v31.2d      \n"
-
-                "lsl    w4, %w13, #2                \n"
-                "add    x4, %3, w4, sxtw 1          \n"
-                "st1    {v28.8h}, [%3], #16         \n"
-                "st1    {v29.8h}, [x4]              \n"
-                "b      10f                         \n"
-
-                // if out_elempack == 1
-                "9:                                 \n"
-                // transpose8x2
-                "zip1   v28.8h, v30.8h, v31.8h      \n"
-                "zip2   v29.8h, v30.8h, v31.8h      \n"
-
-                "add    x4, %3, %w13, sxtw 1        \n"
-                "st1    {v28.s}[0], [%3], #4        \n"
-                "st1    {v28.s}[1], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v28.s}[2], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v28.s}[3], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v29.s}[0], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v29.s}[1], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v29.s}[2], [x4]            \n"
-                "add    x4, x4, %w13, sxtw 1        \n"
-                "st1    {v29.s}[3], [x4]            \n"
-
-                "10:                                \n"
-                "add    %0, %0, #64                 \n"
-                "b      12f                         \n"
-
-                "11:                                \n"
-                "st1    {v30.8h, v31.8h}, [%0], #32 \n"
-
-                "12:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(pC),           // %8
-                "r"(max_kk),       // %9
-                "r"(k),            // %10
-                "r"(k_end),        // %11
-                "r"(out_elempack), // %12
-                "r"(out_hstep)     // %13
-                : "cc", "memory", "x4", "v0", "v4", "v5", "v6", "v7", "v28", "v29", "v30", "v31");
-#else  // NCNN_GNU_INLINE_ASM
-            float16x8_t _sum0;
-            float16x8_t _sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1q_f16(pC);
-                    _sum1 = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f16(0.f);
-                    _sum1 = vdupq_n_f16(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f16(outptr);
-                _sum1 = vld1q_f16(outptr + 8);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x8_t _pA = vld1q_f16(pA);
-
-                float16x4_t _pB = vld1_f16(pB);
-
-                _sum0 = vfmaq_lane_f16(_sum0, _pA, _pB, 0);
-                _sum1 = vfmaq_lane_f16(_sum1, _pA, _pB, 1);
-
-                pA += 8;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 8)
-                {
-                    vst1q_f16(outptr0, _sum0);
-                    vst1q_f16(outptr0 + 8, _sum1);
-                    outptr0 += 16;
-                }
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr0, vget_low_f16(_sum0));
-                    vst1_f16(outptr0 + 4, vget_low_f16(_sum1));
-
-                    vst1_f16(outptr0 + out_hstep * 4, vget_high_f16(_sum0));
-                    vst1_f16(outptr0 + out_hstep * 4 + 4, vget_high_f16(_sum1));
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-                    __fp16 sum0[8];
-                    __fp16 sum1[8];
-                    vst1q_f16(sum0, _sum0);
-                    vst1q_f16(sum1, _sum1);
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[out_hstep * 4] = sum0[4];
-                    outptr0[out_hstep * 5] = sum0[5];
-                    outptr0[out_hstep * 6] = sum0[6];
-                    outptr0[out_hstep * 7] = sum0[7];
-
-                    outptr0[1] = sum1[0];
-                    outptr0[out_hstep + 1] = sum1[1];
-                    outptr0[out_hstep * 2 + 1] = sum1[2];
-                    outptr0[out_hstep * 3 + 1] = sum1[3];
-                    outptr0[out_hstep * 4 + 1] = sum1[4];
-                    outptr0[out_hstep * 5 + 1] = sum1[5];
-                    outptr0[out_hstep * 6 + 1] = sum1[6];
-                    outptr0[out_hstep * 7 + 1] = sum1[7];
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                vst1q_f16(outptr, _sum0);
-                vst1q_f16(outptr + 8, _sum1);
-            }
-
-            outptr += 16;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            const __fp16* pA = pAT;
-
-            float16x8_t _sum0;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1q_f16(pC);
-                }
-                else
-                {
-                    _sum0 = vdupq_n_f16(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f16(outptr);
-            }
-
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x8_t _pA = vld1q_f16(pA);
-
-                float16x8_t _pB = vld1q_dup_f16(pB);
-
-                _sum0 = vfmaq_f16(_sum0, _pA, _pB);
-
-                pA += 8;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 8)
-                {
-                    vst1q_f16(outptr0, _sum0);
-                    outptr0 += 8;
-                }
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr0, vget_low_f16(_sum0));
-                    vst1_f16(outptr0 + out_hstep * 4, vget_high_f16(_sum0));
-                    outptr0 += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    __fp16 sum0[8];
-                    vst1q_f16(sum0, _sum0);
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep * 1] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[out_hstep * 4] = sum0[4];
-                    outptr0[out_hstep * 5] = sum0[5];
-                    outptr0[out_hstep * 6] = sum0[6];
-                    outptr0[out_hstep * 7] = sum0[7];
-                    outptr0++;
-                }
-            }
-            else
-            {
-                vst1q_f16(outptr, _sum0);
-            }
-
-            outptr += 8;
-        }
-
-        pAT += max_kk * 8;
-    }
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        __fp16* outptr0 = (__fp16*)top_blob + (i + ii) * out_hstep + j * out_elempack;
-
-        const __fp16* pB = pBT;
-
-        if (pC)
-        {
-            pC = (const __fp16*)CT_tile + i + ii;
-        }
-
-        int jj = 0;
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            float16x4_t _sum0;
-            float16x4_t _sum1;
-            float16x4_t _sum2;
-            float16x4_t _sum3;
-            float16x4_t _sum4;
-            float16x4_t _sum5;
-            float16x4_t _sum6;
-            float16x4_t _sum7;
-            float16x4_t _sum8;
-            float16x4_t _sum9;
-            float16x4_t _suma;
-            float16x4_t _sumb;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1_f16(pC);
-                    _sum1 = _sum0;
-                    _sum2 = _sum0;
-                    _sum3 = _sum0;
-                    _sum4 = _sum0;
-                    _sum5 = _sum0;
-                    _sum6 = _sum0;
-                    _sum7 = _sum0;
-                    _sum8 = _sum0;
-                    _sum9 = _sum0;
-                    _suma = _sum0;
-                    _sumb = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdup_n_f16(0.f);
-                    _sum1 = vdup_n_f16(0.f);
-                    _sum2 = vdup_n_f16(0.f);
-                    _sum3 = vdup_n_f16(0.f);
-                    _sum4 = vdup_n_f16(0.f);
-                    _sum5 = vdup_n_f16(0.f);
-                    _sum6 = vdup_n_f16(0.f);
-                    _sum7 = vdup_n_f16(0.f);
-                    _sum8 = vdup_n_f16(0.f);
-                    _sum9 = vdup_n_f16(0.f);
-                    _suma = vdup_n_f16(0.f);
-                    _sumb = vdup_n_f16(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1_f16(outptr);
-                _sum1 = vld1_f16(outptr + 4 * 1);
-                _sum2 = vld1_f16(outptr + 4 * 2);
-                _sum3 = vld1_f16(outptr + 4 * 3);
-                _sum4 = vld1_f16(outptr + 4 * 4);
-                _sum5 = vld1_f16(outptr + 4 * 5);
-                _sum6 = vld1_f16(outptr + 4 * 6);
-                _sum7 = vld1_f16(outptr + 4 * 7);
-                _sum8 = vld1_f16(outptr + 4 * 8);
-                _sum9 = vld1_f16(outptr + 4 * 9);
-                _suma = vld1_f16(outptr + 4 * 10);
-                _sumb = vld1_f16(outptr + 4 * 11);
-            }
-
-            const __fp16* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x4_t _pA = vld1_f16(pA);
-                float16x4_t _pB0 = vld1_f16(pB);
-                float16x4_t _pB1 = vld1_f16(pB + 4);
-                float16x4_t _pB2 = vld1_f16(pB + 8);
-
-                _sum0 = vfma_lane_f16(_sum0, _pA, _pB0, 0);
-                _sum1 = vfma_lane_f16(_sum1, _pA, _pB0, 1);
-                _sum2 = vfma_lane_f16(_sum2, _pA, _pB0, 2);
-                _sum3 = vfma_lane_f16(_sum3, _pA, _pB0, 3);
-                _sum4 = vfma_lane_f16(_sum4, _pA, _pB1, 0);
-                _sum5 = vfma_lane_f16(_sum5, _pA, _pB1, 1);
-                _sum6 = vfma_lane_f16(_sum6, _pA, _pB1, 2);
-                _sum7 = vfma_lane_f16(_sum7, _pA, _pB1, 3);
-                _sum8 = vfma_lane_f16(_sum8, _pA, _pB2, 0);
-                _sum9 = vfma_lane_f16(_sum9, _pA, _pB2, 1);
-                _suma = vfma_lane_f16(_suma, _pA, _pB2, 2);
-                _sumb = vfma_lane_f16(_sumb, _pA, _pB2, 3);
-
-                pA += 4;
-                pB += 12;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr0, _sum0);
-                    vst1_f16(outptr0 + 4, _sum1);
-                    vst1_f16(outptr0 + 4 * 2, _sum2);
-                    vst1_f16(outptr0 + 4 * 3, _sum3);
-                    vst1_f16(outptr0 + 4 * 4, _sum4);
-                    vst1_f16(outptr0 + 4 * 5, _sum5);
-                    vst1_f16(outptr0 + 4 * 6, _sum6);
-                    vst1_f16(outptr0 + 4 * 7, _sum7);
-                    vst1_f16(outptr0 + 4 * 8, _sum8);
-                    vst1_f16(outptr0 + 4 * 9, _sum9);
-                    vst1_f16(outptr0 + 4 * 10, _suma);
-                    vst1_f16(outptr0 + 4 * 11, _sumb);
-                    outptr0 += 48;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose4x12_ph(_sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7, _sum8, _sum9, _suma, _sumb);
-
-                    vst1_f16(outptr0, _sum0);
-                    vst1_f16(outptr0 + 4, _sum1);
-                    vst1_f16(outptr0 + 8, _sum2);
-                    vst1_f16(outptr0 + out_hstep, _sum3);
-                    vst1_f16(outptr0 + out_hstep + 4, _sum4);
-                    vst1_f16(outptr0 + out_hstep + 8, _sum5);
-                    vst1_f16(outptr0 + out_hstep * 2, _sum6);
-                    vst1_f16(outptr0 + out_hstep * 2 + 4, _sum7);
-                    vst1_f16(outptr0 + out_hstep * 2 + 8, _sum8);
-                    vst1_f16(outptr0 + out_hstep * 3, _sum9);
-                    vst1_f16(outptr0 + out_hstep * 3 + 4, _suma);
-                    vst1_f16(outptr0 + out_hstep * 3 + 8, _sumb);
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                vst1_f16(outptr, _sum0);
-                vst1_f16(outptr + 4, _sum1);
-                vst1_f16(outptr + 4 * 2, _sum2);
-                vst1_f16(outptr + 4 * 3, _sum3);
-                vst1_f16(outptr + 4 * 4, _sum4);
-                vst1_f16(outptr + 4 * 5, _sum5);
-                vst1_f16(outptr + 4 * 6, _sum6);
-                vst1_f16(outptr + 4 * 7, _sum7);
-                vst1_f16(outptr + 4 * 8, _sum8);
-                vst1_f16(outptr + 4 * 9, _sum9);
-                vst1_f16(outptr + 4 * 10, _suma);
-                vst1_f16(outptr + 4 * 11, _sumb);
-            }
-
-            outptr += 48;
-        }
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            float16x4_t _sum0;
-            float16x4_t _sum1;
-            float16x4_t _sum2;
-            float16x4_t _sum3;
-            float16x4_t _sum4;
-            float16x4_t _sum5;
-            float16x4_t _sum6;
-            float16x4_t _sum7;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1_f16(pC);
-                    _sum1 = _sum0;
-                    _sum2 = _sum0;
-                    _sum3 = _sum0;
-                    _sum4 = _sum0;
-                    _sum5 = _sum0;
-                    _sum6 = _sum0;
-                    _sum7 = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdup_n_f16(0.f);
-                    _sum1 = vdup_n_f16(0.f);
-                    _sum2 = vdup_n_f16(0.f);
-                    _sum3 = vdup_n_f16(0.f);
-                    _sum4 = vdup_n_f16(0.f);
-                    _sum5 = vdup_n_f16(0.f);
-                    _sum6 = vdup_n_f16(0.f);
-                    _sum7 = vdup_n_f16(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1_f16(outptr);
-                _sum1 = vld1_f16(outptr + 4 * 1);
-                _sum2 = vld1_f16(outptr + 4 * 2);
-                _sum3 = vld1_f16(outptr + 4 * 3);
-                _sum4 = vld1_f16(outptr + 4 * 4);
-                _sum5 = vld1_f16(outptr + 4 * 5);
-                _sum6 = vld1_f16(outptr + 4 * 6);
-                _sum7 = vld1_f16(outptr + 4 * 7);
-            }
-
-            const __fp16* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x4_t _pA = vld1_f16(pA);
-                float16x4_t _pB0 = vld1_f16(pB);
-                float16x4_t _pB1 = vld1_f16(pB + 4);
-
-                _sum0 = vfma_lane_f16(_sum0, _pA, _pB0, 0);
-                _sum1 = vfma_lane_f16(_sum1, _pA, _pB0, 1);
-                _sum2 = vfma_lane_f16(_sum2, _pA, _pB0, 2);
-                _sum3 = vfma_lane_f16(_sum3, _pA, _pB0, 3);
-                _sum4 = vfma_lane_f16(_sum4, _pA, _pB1, 0);
-                _sum5 = vfma_lane_f16(_sum5, _pA, _pB1, 1);
-                _sum6 = vfma_lane_f16(_sum6, _pA, _pB1, 2);
-                _sum7 = vfma_lane_f16(_sum7, _pA, _pB1, 3);
-
-                pA += 4;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr0, _sum0);
-                    vst1_f16(outptr0 + 4, _sum1);
-                    vst1_f16(outptr0 + 4 * 2, _sum2);
-                    vst1_f16(outptr0 + 4 * 3, _sum3);
-                    vst1_f16(outptr0 + 4 * 4, _sum4);
-                    vst1_f16(outptr0 + 4 * 5, _sum5);
-                    vst1_f16(outptr0 + 4 * 6, _sum6);
-                    vst1_f16(outptr0 + 4 * 7, _sum7);
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose4x8_ph(_sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7);
-
-                    vst1_f16(outptr0, _sum0);
-                    vst1_f16(outptr0 + 4, _sum1);
-                    vst1_f16(outptr0 + out_hstep, _sum2);
-                    vst1_f16(outptr0 + out_hstep + 4, _sum3);
-                    vst1_f16(outptr0 + out_hstep * 2, _sum4);
-                    vst1_f16(outptr0 + out_hstep * 2 + 4, _sum5);
-                    vst1_f16(outptr0 + out_hstep * 3, _sum6);
-                    vst1_f16(outptr0 + out_hstep * 3 + 4, _sum7);
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1_f16(outptr, _sum0);
-                vst1_f16(outptr + 4, _sum1);
-                vst1_f16(outptr + 4 * 2, _sum2);
-                vst1_f16(outptr + 4 * 3, _sum3);
-                vst1_f16(outptr + 4 * 4, _sum4);
-                vst1_f16(outptr + 4 * 5, _sum5);
-                vst1_f16(outptr + 4 * 6, _sum6);
-                vst1_f16(outptr + 4 * 7, _sum7);
-            }
-
-            outptr += 32;
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            float16x4_t _sum0;
-            float16x4_t _sum1;
-            float16x4_t _sum2;
-            float16x4_t _sum3;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1_f16(pC);
-                    _sum1 = _sum0;
-                    _sum2 = _sum0;
-                    _sum3 = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdup_n_f16(0.f);
-                    _sum1 = vdup_n_f16(0.f);
-                    _sum2 = vdup_n_f16(0.f);
-                    _sum3 = vdup_n_f16(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1_f16(outptr);
-                _sum1 = vld1_f16(outptr + 4);
-                _sum2 = vld1_f16(outptr + 4 * 2);
-                _sum3 = vld1_f16(outptr + 4 * 3);
-            }
-
-            const __fp16* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x4_t _pA = vld1_f16(pA);
-                float16x4_t _pB = vld1_f16(pB);
-
-                _sum0 = vfma_lane_f16(_sum0, _pA, _pB, 0);
-                _sum1 = vfma_lane_f16(_sum1, _pA, _pB, 1);
-                _sum2 = vfma_lane_f16(_sum2, _pA, _pB, 2);
-                _sum3 = vfma_lane_f16(_sum3, _pA, _pB, 3);
-
-                pA += 4;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr0, _sum0);
-                    vst1_f16(outptr0 + 4, _sum1);
-                    vst1_f16(outptr0 + 4 * 2, _sum2);
-                    vst1_f16(outptr0 + 4 * 3, _sum3);
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose4x4_ph(_sum0, _sum1, _sum2, _sum3);
-
-                    vst1_f16(outptr0, _sum0);
-                    vst1_f16(outptr0 + out_hstep, _sum1);
-                    vst1_f16(outptr0 + out_hstep * 2, _sum2);
-                    vst1_f16(outptr0 + out_hstep * 3, _sum3);
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1_f16(outptr, _sum0);
-                vst1_f16(outptr + 4, _sum1);
-                vst1_f16(outptr + 4 * 2, _sum2);
-                vst1_f16(outptr + 4 * 3, _sum3);
-            }
-
-            outptr += 16;
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            float16x4_t _sum0;
-            float16x4_t _sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1_f16(pC);
-                    _sum1 = _sum0;
-                }
-                else
-                {
-                    _sum0 = vdup_n_f16(0.f);
-                    _sum1 = vdup_n_f16(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1_f16(outptr);
-                _sum1 = vld1_f16(outptr + 4);
-            }
-
-            const __fp16* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x4_t _pA = vld1_f16(pA);
-
-                _sum0 = vfma_n_f16(_sum0, _pA, pB[0]);
-                _sum1 = vfma_n_f16(_sum1, _pA, pB[1]);
-
-                pA += 4;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr0, _sum0);
-                    vst1_f16(outptr0 + 4, _sum1);
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-                    __fp16 sum0[4];
-                    __fp16 sum1[4];
-                    vst1_f16(sum0, _sum0);
-                    vst1_f16(sum1, _sum1);
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[1] = sum1[0];
-                    outptr0[out_hstep + 1] = sum1[1];
-                    outptr0[out_hstep * 2 + 1] = sum1[2];
-                    outptr0[out_hstep * 3 + 1] = sum1[3];
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                vst1_f16(outptr, _sum0);
-                vst1_f16(outptr + 4, _sum1);
-            }
-
-            outptr += 8;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            float16x4_t _sum0;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vld1_f16(pC);
-                }
-                else
-                {
-                    _sum0 = vdup_n_f16(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1_f16(outptr);
-            }
-
-            const __fp16* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x4_t _pA = vld1_f16(pA);
-                float16x4_t _pB = vdup_n_f16(pB[0]);
-
-                _sum0 = vfma_f16(_sum0, _pA, _pB);
-
-                pA += 4;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr0, _sum0);
-                    outptr0 += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    __fp16 sum0[4];
-                    vst1_f16(sum0, _sum0);
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0++;
-                }
-            }
-            else
-            {
-                vst1_f16(outptr, _sum0);
-            }
-
-            outptr += 4;
-        }
-
-        pAT += max_kk * 4;
-    }
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        __fp16* outptr0 = (__fp16*)top_blob + (i + ii) * out_hstep + j;
-
-        const __fp16* pB = pBT;
-
-        if (pC)
-        {
-            pC = (const __fp16*)CT_tile + i + ii;
-        }
-
-        int jj = 0;
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            float16x4_t _sum00;
-            float16x4_t _sum01;
-            float16x4_t _sum02;
-            float16x4_t _sum10;
-            float16x4_t _sum11;
-            float16x4_t _sum12;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vdup_n_f16(pC[0]);
-                    _sum01 = vdup_n_f16(pC[0]);
-                    _sum02 = vdup_n_f16(pC[0]);
-                    _sum10 = vdup_n_f16(pC[1]);
-                    _sum11 = vdup_n_f16(pC[1]);
-                    _sum12 = vdup_n_f16(pC[1]);
-                }
-                else
-                {
-                    _sum00 = vdup_n_f16(0.f);
-                    _sum01 = vdup_n_f16(0.f);
-                    _sum02 = vdup_n_f16(0.f);
-                    _sum10 = vdup_n_f16(0.f);
-                    _sum11 = vdup_n_f16(0.f);
-                    _sum12 = vdup_n_f16(0.f);
-                }
-            }
-            else
-            {
-                float16x4x2_t _tmp01 = vld2_f16(outptr);
-                float16x4x2_t _tmp23 = vld2_f16(outptr + 8);
-                float16x4x2_t _tmp45 = vld2_f16(outptr + 16);
-                _sum00 = _tmp01.val[0];
-                _sum01 = _tmp23.val[0];
-                _sum02 = _tmp45.val[0];
-                _sum10 = _tmp01.val[1];
-                _sum11 = _tmp23.val[1];
-                _sum12 = _tmp45.val[1];
-            }
-
-            const __fp16* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x4_t _pB0 = vld1_f16(pB);
-                float16x4_t _pB1 = vld1_f16(pB + 4);
-                float16x4_t _pB2 = vld1_f16(pB + 8);
-
-                float16x4_t _pA0 = vld1_dup_f16(pA);
-                float16x4_t _pA1 = vld1_dup_f16(pA + 1);
-
-                _sum00 = vfma_f16(_sum00, _pB0, _pA0);
-                _sum01 = vfma_f16(_sum01, _pB1, _pA0);
-                _sum02 = vfma_f16(_sum02, _pB2, _pA0);
-                _sum10 = vfma_f16(_sum10, _pB0, _pA1);
-                _sum11 = vfma_f16(_sum11, _pB1, _pA1);
-                _sum12 = vfma_f16(_sum12, _pB2, _pA1);
-
-                pA += 2;
-                pB += 12;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_f16(outptr0, _sum00);
-                    vst1_f16(outptr0 + 4, _sum01);
-                    vst1_f16(outptr0 + 8, _sum02);
-                    vst1_f16(outptr0 + out_hstep, _sum10);
-                    vst1_f16(outptr0 + out_hstep + 4, _sum11);
-                    vst1_f16(outptr0 + out_hstep + 8, _sum12);
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                float16x4x2_t _tmp01;
-                _tmp01.val[0] = _sum00;
-                _tmp01.val[1] = _sum10;
-                float16x4x2_t _tmp23;
-                _tmp23.val[0] = _sum01;
-                _tmp23.val[1] = _sum11;
-                float16x4x2_t _tmp45;
-                _tmp45.val[0] = _sum02;
-                _tmp45.val[1] = _sum12;
-                vst2_f16(outptr, _tmp01);
-                vst2_f16(outptr + 8, _tmp23);
-                vst2_f16(outptr + 16, _tmp45);
-            }
-
-            outptr += 24;
-        }
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            float16x4_t _sum00;
-            float16x4_t _sum01;
-            float16x4_t _sum10;
-            float16x4_t _sum11;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum00 = vdup_n_f16(pC[0]);
-                    _sum01 = vdup_n_f16(pC[0]);
-                    _sum10 = vdup_n_f16(pC[1]);
-                    _sum11 = vdup_n_f16(pC[1]);
-                }
-                else
-                {
-                    _sum00 = vdup_n_f16(0.f);
-                    _sum01 = vdup_n_f16(0.f);
-                    _sum10 = vdup_n_f16(0.f);
-                    _sum11 = vdup_n_f16(0.f);
-                }
-            }
-            else
-            {
-                float16x4x2_t _tmp01 = vld2_f16(outptr);
-                float16x4x2_t _tmp23 = vld2_f16(outptr + 8);
-                _sum00 = _tmp01.val[0];
-                _sum01 = _tmp23.val[0];
-                _sum10 = _tmp01.val[1];
-                _sum11 = _tmp23.val[1];
-            }
-
-            const __fp16* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x4_t _pB0 = vld1_f16(pB);
-                float16x4_t _pB1 = vld1_f16(pB + 4);
-
-                float16x4_t _pA0 = vld1_dup_f16(pA);
-                float16x4_t _pA1 = vld1_dup_f16(pA + 1);
-
-                _sum00 = vfma_f16(_sum00, _pB0, _pA0);
-                _sum01 = vfma_f16(_sum01, _pB1, _pA0);
-                _sum10 = vfma_f16(_sum10, _pB0, _pA1);
-                _sum11 = vfma_f16(_sum11, _pB1, _pA1);
-
-                pA += 2;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_f16(outptr0, _sum00);
-                    vst1_f16(outptr0 + 4, _sum01);
-                    vst1_f16(outptr0 + out_hstep, _sum10);
-                    vst1_f16(outptr0 + out_hstep + 4, _sum11);
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                float16x4x2_t _tmp01;
-                _tmp01.val[0] = _sum00;
-                _tmp01.val[1] = _sum10;
-                float16x4x2_t _tmp23;
-                _tmp23.val[0] = _sum01;
-                _tmp23.val[1] = _sum11;
-                vst2_f16(outptr, _tmp01);
-                vst2_f16(outptr + 8, _tmp23);
-            }
-
-            outptr += 16;
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            float16x4_t _sum0;
-            float16x4_t _sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vdup_n_f16(pC[0]);
-                    _sum1 = vdup_n_f16(pC[1]);
-                }
-                else
-                {
-                    _sum0 = vdup_n_f16(0.f);
-                    _sum1 = vdup_n_f16(0.f);
-                }
-            }
-            else
-            {
-                float16x4x2_t _tmp01 = vld2_f16(outptr);
-                _sum0 = _tmp01.val[0];
-                _sum1 = _tmp01.val[1];
-            }
-
-            const __fp16* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x4_t _pB = vld1_f16(pB);
-
-                _sum0 = vfma_n_f16(_sum0, _pB, pA[0]);
-                _sum1 = vfma_n_f16(_sum1, _pB, pA[1]);
-
-                pA += 2;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_f16(outptr0, (_sum0));
-                    vst1_f16(outptr0 + out_hstep, (_sum1));
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                float16x4x2_t _tmp01;
-                _tmp01.val[0] = _sum0;
-                _tmp01.val[1] = _sum1;
-                vst2_f16(outptr, _tmp01);
-            }
-
-            outptr += 8;
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            __fp16 sum00;
-            __fp16 sum01;
-            __fp16 sum10;
-            __fp16 sum11;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    sum00 = pC[0];
-                    sum01 = pC[1];
-                    sum10 = pC[0];
-                    sum11 = pC[1];
-                }
-                else
-                {
-                    sum00 = 0.f;
-                    sum01 = 0.f;
-                    sum10 = 0.f;
-                    sum11 = 0.f;
-                }
-            }
-            else
-            {
-                sum00 = outptr[0];
-                sum01 = outptr[1];
-                sum10 = outptr[2];
-                sum11 = outptr[3];
-            }
-
-            const __fp16* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                sum00 += pA[0] * pB[0];
-                sum01 += pA[1] * pB[0];
-                sum10 += pA[0] * pB[1];
-                sum11 += pA[1] * pB[1];
-
-                pA += 2;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum00;
-                    outptr0[1] = sum10;
-                    outptr0[out_hstep] = sum01;
-                    outptr0[out_hstep + 1] = sum11;
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                outptr[0] = sum00;
-                outptr[1] = sum01;
-                outptr[2] = sum10;
-                outptr[3] = sum11;
-            }
-
-            outptr += 4;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            __fp16 sum0;
-            __fp16 sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    sum0 = pC[0];
-                    sum1 = pC[1];
-                }
-                else
-                {
-                    sum0 = 0.f;
-                    sum1 = 0.f;
-                }
-            }
-            else
-            {
-                sum0 = outptr[0];
-                sum1 = outptr[1];
-            }
-
-            const __fp16* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                sum0 += pA[0] * pB[0];
-                sum1 += pA[1] * pB[0];
-                pA += 2;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum0;
-                    outptr0[out_hstep] = sum1;
-                    outptr0++;
-                }
-            }
-            else
-            {
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-            }
-
-            outptr += 2;
-        }
-
-        pAT += max_kk * 2;
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        __fp16* outptr0 = (__fp16*)top_blob + (i + ii) * out_hstep + j;
-
-        const __fp16* pB = pBT;
-
-        if (pC)
-        {
-            pC = (const __fp16*)CT_tile + i + ii;
-        }
-
-        int jj = 0;
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            float16x4_t _sum0;
-            float16x4_t _sum1;
-            float16x4_t _sum2;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vdup_n_f16(pC[0]);
-                    _sum1 = vdup_n_f16(pC[0]);
-                    _sum2 = vdup_n_f16(pC[0]);
-                }
-                else
-                {
-                    _sum0 = vdup_n_f16(0.f);
-                    _sum1 = vdup_n_f16(0.f);
-                    _sum2 = vdup_n_f16(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1_f16(outptr);
-                _sum1 = vld1_f16(outptr + 4);
-                _sum2 = vld1_f16(outptr + 8);
-            }
-
-            const __fp16* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x4_t _pB0 = vld1_f16(pB);
-                float16x4_t _pB1 = vld1_f16(pB + 4);
-                float16x4_t _pB2 = vld1_f16(pB + 8);
-
-                float16x4_t _pA0 = vdup_n_f16(pA[0]);
-
-                _sum0 = vfma_f16(_sum0, _pA0, _pB0);
-                _sum1 = vfma_f16(_sum1, _pA0, _pB1);
-                _sum2 = vfma_f16(_sum2, _pA0, _pB2);
-
-                pA += 1;
-                pB += 12;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_f16(outptr0, _sum0);
-                    vst1_f16(outptr0 + 4, _sum1);
-                    vst1_f16(outptr0 + 8, _sum2);
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                vst1_f16(outptr, _sum0);
-                vst1_f16(outptr + 4, _sum1);
-                vst1_f16(outptr + 8, _sum2);
-            }
-
-            outptr += 12;
-        }
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            float16x4_t _sum0;
-            float16x4_t _sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum0 = vdup_n_f16(pC[0]);
-                    _sum1 = vdup_n_f16(pC[0]);
-                }
-                else
-                {
-                    _sum0 = vdup_n_f16(0.f);
-                    _sum1 = vdup_n_f16(0.f);
-                }
-            }
-            else
-            {
-                _sum0 = vld1_f16(outptr);
-                _sum1 = vld1_f16(outptr + 4);
-            }
-
-            const __fp16* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x4_t _pB0 = vld1_f16(pB);
-                float16x4_t _pB1 = vld1_f16(pB + 4);
-
-                float16x4_t _pA0 = vdup_n_f16(pA[0]);
-
-                _sum0 = vfma_f16(_sum0, _pA0, _pB0);
-                _sum1 = vfma_f16(_sum1, _pA0, _pB1);
-
-                pA += 1;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_f16(outptr0, _sum0);
-                    vst1_f16(outptr0 + 4, _sum1);
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1_f16(outptr, _sum0);
-                vst1_f16(outptr + 4, _sum1);
-            }
-
-            outptr += 8;
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            float16x4_t _sum;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    _sum = vdup_n_f16(pC[0]);
-                }
-                else
-                {
-                    _sum = vdup_n_f16(0.f);
-                }
-            }
-            else
-            {
-                _sum = vld1_f16(outptr);
-            }
-
-            const __fp16* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float16x4_t _pB = vld1_f16(pB);
-                float16x4_t _pA = vdup_n_f16(pA[0]);
-
-                _sum = vfma_f16(_sum, _pA, _pB);
-
-                pA += 1;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_f16(outptr0, _sum);
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1_f16(outptr, _sum);
-            }
-
-            outptr += 4;
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            __fp16 sum0;
-            __fp16 sum1;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    sum0 = pC[0];
-                    sum1 = pC[0];
-                }
-                else
-                {
-                    sum0 = 0.f;
-                    sum1 = 0.f;
-                }
-            }
-            else
-            {
-                sum0 = outptr[0];
-                sum1 = outptr[1];
-            }
-
-            const __fp16* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                sum0 += pA[0] * pB[0];
-                sum1 += pA[0] * pB[1];
-
-                pA += 1;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum0;
-                    outptr0[1] = sum1;
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-            }
-
-            outptr += 2;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            __fp16 sum;
-
-            if (k == 0)
-            {
-                if (pC)
-                {
-                    sum = pC[0];
-                }
-                else
-                {
-                    sum = 0.f;
-                }
-            }
-            else
-            {
-                sum = outptr[0];
-            }
-
-            const __fp16* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                sum += pA[0] * pB[0];
-
-                pA += 1;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum;
-                    outptr0++;
-                }
-            }
-            else
-            {
-                outptr[0] = sum;
-            }
-
-            outptr += 1;
-        }
-
-        pAT += max_kk;
-    }
-}
-
-static void convolution_im2col_gemm_get_optimal_tile_mnk_fp16sa(int M, int N, int K, int& TILE_M, int& TILE_N, int& TILE_K, int nT)
-{
-    // resolve optimal tile size from cache size
-    const int l2_cache_size_fp16 = (int)(get_cpu_level2_cache_size() / sizeof(unsigned short));
-
-    if (nT == 0)
-        nT = get_physical_big_cpu_count();
-
-    // solve K
-    {
-        // try not to split K
-        int tile_size = (l2_cache_size_fp16 - 32) / 12;
-
-        TILE_K = std::max(8, tile_size / 8 * 8);
-
-        int nn_K = (K + TILE_K - 1) / TILE_K;
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 7) / 8 * 8);
-    }
-
-    // solve M
-    {
-        int nn_M = (M + 63) / 64;
-
-        TILE_M = std::max(8, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-    }
-
-    {
-        TILE_M *= std::min(nT, get_physical_cpu_count());
-
-        int nn_M = (M + TILE_M - 1) / TILE_M;
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-
-        if (nT > 1)
-        {
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 7) / 8 * 8);
-        }
-    }
-
-    if (N > 0)
-    {
-        int tile_size;
-        if (TILE_K >= K)
-        {
-            tile_size = (l2_cache_size_fp16 - TILE_M * TILE_K) / TILE_K;
-        }
-        else
-        {
-            tile_size = (l2_cache_size_fp16 - TILE_M * TILE_K) / (TILE_M + TILE_K);
-        }
-
-        TILE_N = std::max(4, tile_size / 4 * 4);
-
-        int nn_N = (N + TILE_N - 1) / TILE_N;
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-    }
-}
-
-static void convolution_im2col_gemm_transform_kernel_fp16sa(const Mat& kernel, Mat& AT, int inch, int outch, int kernel_w, int kernel_h, const Option& opt)
-{
-    // NCNN_LOGE("convolution_im2col_gemm_transform_kernel_fp16sa %p", kernel.data);
-    const int maxk = kernel_w * kernel_h;
-
-    const int M = outch;
-    const int K = inch * maxk;
-
-    int TILE_M, TILE_N, TILE_K;
-    convolution_im2col_gemm_get_optimal_tile_mnk_fp16sa(M, 0, K, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    int elempack = 1;
-    if (opt.use_packing_layout)
-    {
-        elempack = inch % 8 == 0 ? 8 : inch % 4 == 0 ? 4 : 1;
-    }
-
-    // maxk-inch-outch to pa-maxk-inch/pa-outch
-    Mat A_data;
-    if (maxk == 1)
-    {
-        cast_float32_to_float16(kernel, A_data);
-        A_data = A_data.reshape(maxk * inch, outch);
-    }
-    else
-    {
-        Mat weight_data_r2 = kernel.reshape(maxk, inch, outch);
-
-        A_data.create(maxk * inch, outch, (size_t)2u);
-
-        for (int q = 0; q < outch; q += 1)
-        {
-            __fp16* g00 = A_data.row<__fp16>(q);
-
-            for (int p = 0; p + (elempack - 1) < inch; p += elempack)
-            {
-                for (int k = 0; k < maxk; k++)
-                {
-                    for (int i = 0; i < elempack; i++)
-                    {
-                        const float* k00 = weight_data_r2.channel(q).row(p + i);
-                        g00[0] = (__fp16)k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-
-    AT.create(TILE_K * TILE_M, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M, (size_t)2u);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat AT_tile = AT.channel(i / TILE_M).row_range(k / TILE_K, 1);
-
-            convolution_im2col_pack_A_tile_bf16_fp16(A_data, AT_tile, i, max_ii, k, max_kk);
-        }
-    }
-}
-
-static void convolution_im2col_gemm_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, const Mat& bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt)
-{
-    // NCNN_LOGE("convolution_im2col_gemm_fp16sa %p %p %p %p", bottom_blob.data, top_blob.data, AT.data, bias.data);
-    const int maxk = kernel_w * kernel_h;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = top_blob.w * top_blob.h;
-    const int K = bottom_blob.c * bottom_blob.elempack * maxk;
-
-    int TILE_M, TILE_N, TILE_K;
-    convolution_im2col_gemm_get_optimal_tile_mnk_fp16sa(M, N, K, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 2u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-    {
-        const int ppj = ppjk / nn_K;
-        const int ppk = ppjk % nn_K;
-
-        const int j = ppj * TILE_N;
-        const int k = ppk * TILE_K;
-
-        const int max_jj = std::min((N - j), TILE_N);
-        const int max_kk = std::min((K - k), TILE_K);
-
-        Mat BT_tile = BT.channel(j / TILE_N).row_range(k / TILE_K, 1);
-
-        // im2col
-        convolution_im2col_input_tile_bf16_fp16(bottom_blob, BT_tile, j, max_jj, k, max_kk, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h);
-    }
-
-    Mat topT_tileX;
-    if (K > TILE_K)
-        topT_tileX.create(TILE_N * TILE_M, 1, nT, 2u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat topT_tile;
-        if (K > TILE_K)
-            topT_tile = topT_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).row_range(k / TILE_K, 1);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).row_range(k / TILE_K, 1);
-
-                bool k_end = k + TILE_K >= K;
-
-                convolution_gemm_transB_packed_tile_fp16sa(AT_tile, BT_tile, bias, topT_tile, top_blob, i, max_ii, j, max_jj, k, max_kk, k_end, opt.use_a53_a55_optimized_kernel);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_int8.h
deleted file mode 100644
index 63a1df9c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_im2col_gemm_int8.h
+++ /dev/null
@@ -1,13541 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if !(__ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD)
-#if NCNN_RUNTIME_CPU && NCNN_ARM84I8MM && __aarch64__ && !__ARM_FEATURE_MATMUL_INT8
-void convolution_im2col_gemm_transform_kernel_int8_i8mm(const Mat& kernel, Mat& AT, int inch, int outch, int kernel_w, int kernel_h, const Option& opt);
-void convolution_im2col_gemm_int8_i8mm(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt);
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_ARM82DOT && __aarch64__ && !__ARM_FEATURE_DOTPROD
-void convolution_im2col_gemm_transform_kernel_int8_asimddp(const Mat& kernel, Mat& AT, int inch, int outch, int kernel_w, int kernel_h, const Option& opt);
-void convolution_im2col_gemm_int8_asimddp(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt);
-#endif
-#endif
-
-static void convolution_im2col_pack_A_tile_int8(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
-{
-    // A = (pa, maxk, inch/pa), outch
-    const int A_hstep = A.w;
-
-    signed char* pp = AT;
-
-    int ii = 0;
-#if __ARM_NEON
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        const signed char* p0 = (const signed char*)A + (i + ii) * A_hstep + k;
-        const signed char* p1 = (const signed char*)A + (i + ii + 1) * A_hstep + k;
-        const signed char* p2 = (const signed char*)A + (i + ii + 2) * A_hstep + k;
-        const signed char* p3 = (const signed char*)A + (i + ii + 3) * A_hstep + k;
-        const signed char* p4 = (const signed char*)A + (i + ii + 4) * A_hstep + k;
-        const signed char* p5 = (const signed char*)A + (i + ii + 5) * A_hstep + k;
-        const signed char* p6 = (const signed char*)A + (i + ii + 6) * A_hstep + k;
-        const signed char* p7 = (const signed char*)A + (i + ii + 7) * A_hstep + k;
-
-        int kk = 0;
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-        for (; kk + 15 < max_kk; kk += 16)
-        {
-            int8x16_t _r0 = vld1q_s8(p0);
-            int8x16_t _r1 = vld1q_s8(p1);
-            int8x16_t _r2 = vld1q_s8(p2);
-            int8x16_t _r3 = vld1q_s8(p3);
-            int8x16_t _r4 = vld1q_s8(p4);
-            int8x16_t _r5 = vld1q_s8(p5);
-            int8x16_t _r6 = vld1q_s8(p6);
-            int8x16_t _r7 = vld1q_s8(p7);
-            int8x16_t _t0 = vcombine_s8(vget_low_s8(_r0), vget_low_s8(_r1));
-            int8x16_t _t1 = vcombine_s8(vget_low_s8(_r2), vget_low_s8(_r3));
-            int8x16_t _t2 = vcombine_s8(vget_low_s8(_r4), vget_low_s8(_r5));
-            int8x16_t _t3 = vcombine_s8(vget_low_s8(_r6), vget_low_s8(_r7));
-            int8x16_t _t4 = vcombine_s8(vget_high_s8(_r0), vget_high_s8(_r1));
-            int8x16_t _t5 = vcombine_s8(vget_high_s8(_r2), vget_high_s8(_r3));
-            int8x16_t _t6 = vcombine_s8(vget_high_s8(_r4), vget_high_s8(_r5));
-            int8x16_t _t7 = vcombine_s8(vget_high_s8(_r6), vget_high_s8(_r7));
-            vst1q_s8(pp, _t0);
-            vst1q_s8(pp + 16, _t1);
-            vst1q_s8(pp + 32, _t2);
-            vst1q_s8(pp + 48, _t3);
-            vst1q_s8(pp + 64, _t4);
-            vst1q_s8(pp + 80, _t5);
-            vst1q_s8(pp + 96, _t6);
-            vst1q_s8(pp + 112, _t7);
-            pp += 128;
-            p0 += 16;
-            p1 += 16;
-            p2 += 16;
-            p3 += 16;
-            p4 += 16;
-            p5 += 16;
-            p6 += 16;
-            p7 += 16;
-        }
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            int8x8_t _r0 = vld1_s8(p0);
-            int8x8_t _r1 = vld1_s8(p1);
-            int8x8_t _r2 = vld1_s8(p2);
-            int8x8_t _r3 = vld1_s8(p3);
-            int8x8_t _r4 = vld1_s8(p4);
-            int8x8_t _r5 = vld1_s8(p5);
-            int8x8_t _r6 = vld1_s8(p6);
-            int8x8_t _r7 = vld1_s8(p7);
-            vst1_s8(pp, _r0);
-            vst1_s8(pp + 8, _r1);
-            vst1_s8(pp + 16, _r2);
-            vst1_s8(pp + 24, _r3);
-            vst1_s8(pp + 32, _r4);
-            vst1_s8(pp + 40, _r5);
-            vst1_s8(pp + 48, _r6);
-            vst1_s8(pp + 56, _r7);
-            pp += 64;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-            p4 += 8;
-            p5 += 8;
-            p6 += 8;
-            p7 += 8;
-        }
-#else  // __ARM_FEATURE_MATMUL_INT8
-        for (; kk + 15 < max_kk; kk += 16)
-        {
-            int8x16_t _r0 = vld1q_s8(p0);
-            int8x16_t _r1 = vld1q_s8(p1);
-            int8x16_t _r2 = vld1q_s8(p2);
-            int8x16_t _r3 = vld1q_s8(p3);
-            int8x16_t _r4 = vld1q_s8(p4);
-            int8x16_t _r5 = vld1q_s8(p5);
-            int8x16_t _r6 = vld1q_s8(p6);
-            int8x16_t _r7 = vld1q_s8(p7);
-            int32x4x2_t _r01 = vzipq_s32(vreinterpretq_s32_s8(_r0), vreinterpretq_s32_s8(_r1));
-            int32x4x2_t _r23 = vzipq_s32(vreinterpretq_s32_s8(_r2), vreinterpretq_s32_s8(_r3));
-            int32x4x2_t _r45 = vzipq_s32(vreinterpretq_s32_s8(_r4), vreinterpretq_s32_s8(_r5));
-            int32x4x2_t _r67 = vzipq_s32(vreinterpretq_s32_s8(_r6), vreinterpretq_s32_s8(_r7));
-            _r0 = vreinterpretq_s8_s32(vcombine_s32(vget_low_s32(_r01.val[0]), vget_low_s32(_r23.val[0])));
-            _r1 = vreinterpretq_s8_s32(vcombine_s32(vget_low_s32(_r45.val[0]), vget_low_s32(_r67.val[0])));
-            _r2 = vreinterpretq_s8_s32(vcombine_s32(vget_high_s32(_r01.val[0]), vget_high_s32(_r23.val[0])));
-            _r3 = vreinterpretq_s8_s32(vcombine_s32(vget_high_s32(_r45.val[0]), vget_high_s32(_r67.val[0])));
-            _r4 = vreinterpretq_s8_s32(vcombine_s32(vget_low_s32(_r01.val[1]), vget_low_s32(_r23.val[1])));
-            _r5 = vreinterpretq_s8_s32(vcombine_s32(vget_low_s32(_r45.val[1]), vget_low_s32(_r67.val[1])));
-            _r6 = vreinterpretq_s8_s32(vcombine_s32(vget_high_s32(_r01.val[1]), vget_high_s32(_r23.val[1])));
-            _r7 = vreinterpretq_s8_s32(vcombine_s32(vget_high_s32(_r45.val[1]), vget_high_s32(_r67.val[1])));
-            vst1q_s8(pp, _r0);
-            vst1q_s8(pp + 16, _r1);
-            vst1q_s8(pp + 32, _r2);
-            vst1q_s8(pp + 48, _r3);
-            vst1q_s8(pp + 64, _r4);
-            vst1q_s8(pp + 80, _r5);
-            vst1q_s8(pp + 96, _r6);
-            vst1q_s8(pp + 112, _r7);
-            pp += 128;
-            p0 += 16;
-            p1 += 16;
-            p2 += 16;
-            p3 += 16;
-            p4 += 16;
-            p5 += 16;
-            p6 += 16;
-            p7 += 16;
-        }
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            int8x8_t _r0 = vld1_s8(p0);
-            int8x8_t _r1 = vld1_s8(p1);
-            int8x8_t _r2 = vld1_s8(p2);
-            int8x8_t _r3 = vld1_s8(p3);
-            int8x8_t _r4 = vld1_s8(p4);
-            int8x8_t _r5 = vld1_s8(p5);
-            int8x8_t _r6 = vld1_s8(p6);
-            int8x8_t _r7 = vld1_s8(p7);
-            int32x2x2_t _r01 = vzip_s32(vreinterpret_s32_s8(_r0), vreinterpret_s32_s8(_r1));
-            int32x2x2_t _r23 = vzip_s32(vreinterpret_s32_s8(_r2), vreinterpret_s32_s8(_r3));
-            int32x2x2_t _r45 = vzip_s32(vreinterpret_s32_s8(_r4), vreinterpret_s32_s8(_r5));
-            int32x2x2_t _r67 = vzip_s32(vreinterpret_s32_s8(_r6), vreinterpret_s32_s8(_r7));
-            int8x16_t _t0 = vreinterpretq_s8_s32(vcombine_s32(_r01.val[0], _r23.val[0]));
-            int8x16_t _t1 = vreinterpretq_s8_s32(vcombine_s32(_r45.val[0], _r67.val[0]));
-            int8x16_t _t2 = vreinterpretq_s8_s32(vcombine_s32(_r01.val[1], _r23.val[1]));
-            int8x16_t _t3 = vreinterpretq_s8_s32(vcombine_s32(_r45.val[1], _r67.val[1]));
-            vst1q_s8(pp, _t0);
-            vst1q_s8(pp + 16, _t1);
-            vst1q_s8(pp + 32, _t2);
-            vst1q_s8(pp + 48, _t3);
-            pp += 64;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-            p4 += 8;
-            p5 += 8;
-            p6 += 8;
-            p7 += 8;
-        }
-#endif // __ARM_FEATURE_MATMUL_INT8
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            pp[0] = p0[0];
-            pp[1] = p0[1];
-            pp[2] = p0[2];
-            pp[3] = p0[3];
-            pp[4] = p1[0];
-            pp[5] = p1[1];
-            pp[6] = p1[2];
-            pp[7] = p1[3];
-            pp[8] = p2[0];
-            pp[9] = p2[1];
-            pp[10] = p2[2];
-            pp[11] = p2[3];
-            pp[12] = p3[0];
-            pp[13] = p3[1];
-            pp[14] = p3[2];
-            pp[15] = p3[3];
-            pp[16] = p4[0];
-            pp[17] = p4[1];
-            pp[18] = p4[2];
-            pp[19] = p4[3];
-            pp[20] = p5[0];
-            pp[21] = p5[1];
-            pp[22] = p5[2];
-            pp[23] = p5[3];
-            pp[24] = p6[0];
-            pp[25] = p6[1];
-            pp[26] = p6[2];
-            pp[27] = p6[3];
-            pp[28] = p7[0];
-            pp[29] = p7[1];
-            pp[30] = p7[2];
-            pp[31] = p7[3];
-            pp += 32;
-            p0 += 4;
-            p1 += 4;
-            p2 += 4;
-            p3 += 4;
-            p4 += 4;
-            p5 += 4;
-            p6 += 4;
-            p7 += 4;
-        }
-#else  // __ARM_FEATURE_DOTPROD
-        for (; kk + 15 < max_kk; kk += 16)
-        {
-            int8x16_t _r0 = vld1q_s8(p0);
-            int8x16_t _r1 = vld1q_s8(p1);
-            int8x16_t _r2 = vld1q_s8(p2);
-            int8x16_t _r3 = vld1q_s8(p3);
-            int8x16_t _r4 = vld1q_s8(p4);
-            int8x16_t _r5 = vld1q_s8(p5);
-            int8x16_t _r6 = vld1q_s8(p6);
-            int8x16_t _r7 = vld1q_s8(p7);
-            int16x8x2_t _r01 = vzipq_s16(vreinterpretq_s16_s8(_r0), vreinterpretq_s16_s8(_r1));
-            int16x8x2_t _r23 = vzipq_s16(vreinterpretq_s16_s8(_r2), vreinterpretq_s16_s8(_r3));
-            int16x8x2_t _r45 = vzipq_s16(vreinterpretq_s16_s8(_r4), vreinterpretq_s16_s8(_r5));
-            int16x8x2_t _r67 = vzipq_s16(vreinterpretq_s16_s8(_r6), vreinterpretq_s16_s8(_r7));
-            int32x4x2_t _t0 = vzipq_s32(vreinterpretq_s32_s16(_r01.val[0]), vreinterpretq_s32_s16(_r23.val[0]));
-            int32x4x2_t _t1 = vzipq_s32(vreinterpretq_s32_s16(_r01.val[1]), vreinterpretq_s32_s16(_r23.val[1]));
-            int32x4x2_t _t2 = vzipq_s32(vreinterpretq_s32_s16(_r45.val[0]), vreinterpretq_s32_s16(_r67.val[0]));
-            int32x4x2_t _t3 = vzipq_s32(vreinterpretq_s32_s16(_r45.val[1]), vreinterpretq_s32_s16(_r67.val[1]));
-            _r0 = vreinterpretq_s8_s32(vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t2.val[0])));
-            _r1 = vreinterpretq_s8_s32(vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t2.val[0])));
-            _r2 = vreinterpretq_s8_s32(vcombine_s32(vget_low_s32(_t0.val[1]), vget_low_s32(_t2.val[1])));
-            _r3 = vreinterpretq_s8_s32(vcombine_s32(vget_high_s32(_t0.val[1]), vget_high_s32(_t2.val[1])));
-            _r4 = vreinterpretq_s8_s32(vcombine_s32(vget_low_s32(_t1.val[0]), vget_low_s32(_t3.val[0])));
-            _r5 = vreinterpretq_s8_s32(vcombine_s32(vget_high_s32(_t1.val[0]), vget_high_s32(_t3.val[0])));
-            _r6 = vreinterpretq_s8_s32(vcombine_s32(vget_low_s32(_t1.val[1]), vget_low_s32(_t3.val[1])));
-            _r7 = vreinterpretq_s8_s32(vcombine_s32(vget_high_s32(_t1.val[1]), vget_high_s32(_t3.val[1])));
-            vst1q_s8(pp, _r0);
-            vst1q_s8(pp + 16, _r1);
-            vst1q_s8(pp + 32, _r2);
-            vst1q_s8(pp + 48, _r3);
-            vst1q_s8(pp + 64, _r4);
-            vst1q_s8(pp + 80, _r5);
-            vst1q_s8(pp + 96, _r6);
-            vst1q_s8(pp + 112, _r7);
-            pp += 128;
-            p0 += 16;
-            p1 += 16;
-            p2 += 16;
-            p3 += 16;
-            p4 += 16;
-            p5 += 16;
-            p6 += 16;
-            p7 += 16;
-        }
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            int8x8_t _r0 = vld1_s8(p0);
-            int8x8_t _r1 = vld1_s8(p1);
-            int8x8_t _r2 = vld1_s8(p2);
-            int8x8_t _r3 = vld1_s8(p3);
-            int8x8_t _r4 = vld1_s8(p4);
-            int8x8_t _r5 = vld1_s8(p5);
-            int8x8_t _r6 = vld1_s8(p6);
-            int8x8_t _r7 = vld1_s8(p7);
-            int16x8_t _r04 = vreinterpretq_s16_s8(vcombine_s8(_r0, _r4));
-            int16x8_t _r15 = vreinterpretq_s16_s8(vcombine_s8(_r1, _r5));
-            int16x8_t _r26 = vreinterpretq_s16_s8(vcombine_s8(_r2, _r6));
-            int16x8_t _r37 = vreinterpretq_s16_s8(vcombine_s8(_r3, _r7));
-            int16x8x2_t _t0 = vzipq_s16(_r04, _r15);
-            int16x8x2_t _t1 = vzipq_s16(_r26, _r37);
-            int32x4x2_t _t2 = vzipq_s32(vreinterpretq_s32_s16(_t0.val[0]), vreinterpretq_s32_s16(_t1.val[0]));
-            int32x4x2_t _t3 = vzipq_s32(vreinterpretq_s32_s16(_t0.val[1]), vreinterpretq_s32_s16(_t1.val[1]));
-            int8x16_t _t4 = vreinterpretq_s8_s32(vcombine_s32(vget_low_s32(_t2.val[0]), vget_low_s32(_t3.val[0])));
-            int8x16_t _t5 = vreinterpretq_s8_s32(vcombine_s32(vget_high_s32(_t2.val[0]), vget_high_s32(_t3.val[0])));
-            int8x16_t _t6 = vreinterpretq_s8_s32(vcombine_s32(vget_low_s32(_t2.val[1]), vget_low_s32(_t3.val[1])));
-            int8x16_t _t7 = vreinterpretq_s8_s32(vcombine_s32(vget_high_s32(_t2.val[1]), vget_high_s32(_t3.val[1])));
-            vst1q_s8(pp, _t4);
-            vst1q_s8(pp + 16, _t5);
-            vst1q_s8(pp + 32, _t6);
-            vst1q_s8(pp + 48, _t7);
-            pp += 64;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-            p4 += 8;
-            p5 += 8;
-            p6 += 8;
-            p7 += 8;
-        }
-#endif // __ARM_FEATURE_DOTPROD
-        for (; kk + 1 < max_kk; kk += 2)
-        {
-            pp[0] = p0[0];
-            pp[1] = p0[1];
-            pp[2] = p1[0];
-            pp[3] = p1[1];
-            pp[4] = p2[0];
-            pp[5] = p2[1];
-            pp[6] = p3[0];
-            pp[7] = p3[1];
-            pp[8] = p4[0];
-            pp[9] = p4[1];
-            pp[10] = p5[0];
-            pp[11] = p5[1];
-            pp[12] = p6[0];
-            pp[13] = p6[1];
-            pp[14] = p7[0];
-            pp[15] = p7[1];
-            pp += 16;
-            p0 += 2;
-            p1 += 2;
-            p2 += 2;
-            p3 += 2;
-            p4 += 2;
-            p5 += 2;
-            p6 += 2;
-            p7 += 2;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp[1] = p1[0];
-            pp[2] = p2[0];
-            pp[3] = p3[0];
-            pp[4] = p4[0];
-            pp[5] = p5[0];
-            pp[6] = p6[0];
-            pp[7] = p7[0];
-            pp += 8;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-            p4++;
-            p5++;
-            p6++;
-            p7++;
-        }
-    }
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        const signed char* p0 = (const signed char*)A + (i + ii) * A_hstep + k;
-        const signed char* p1 = (const signed char*)A + (i + ii + 1) * A_hstep + k;
-        const signed char* p2 = (const signed char*)A + (i + ii + 2) * A_hstep + k;
-        const signed char* p3 = (const signed char*)A + (i + ii + 3) * A_hstep + k;
-
-        int kk = 0;
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-        for (; kk + 15 < max_kk; kk += 16)
-        {
-            int64x2x4_t _r0123;
-            _r0123.val[0] = vreinterpretq_s64_s8(vld1q_s8(p0));
-            _r0123.val[1] = vreinterpretq_s64_s8(vld1q_s8(p1));
-            _r0123.val[2] = vreinterpretq_s64_s8(vld1q_s8(p2));
-            _r0123.val[3] = vreinterpretq_s64_s8(vld1q_s8(p3));
-            vst4q_s64((int64_t*)pp, _r0123);
-            pp += 64;
-            p0 += 16;
-            p1 += 16;
-            p2 += 16;
-            p3 += 16;
-        }
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            int8x8_t _r0 = vld1_s8(p0);
-            int8x8_t _r1 = vld1_s8(p1);
-            int8x8_t _r2 = vld1_s8(p2);
-            int8x8_t _r3 = vld1_s8(p3);
-            vst1_s8(pp, _r0);
-            vst1_s8(pp + 8, _r1);
-            vst1_s8(pp + 16, _r2);
-            vst1_s8(pp + 24, _r3);
-            pp += 32;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-        }
-#else  // __ARM_FEATURE_MATMUL_INT8
-        for (; kk + 15 < max_kk; kk += 16)
-        {
-            int32x4x4_t _r0123;
-            _r0123.val[0] = vreinterpretq_s32_s8(vld1q_s8(p0));
-            _r0123.val[1] = vreinterpretq_s32_s8(vld1q_s8(p1));
-            _r0123.val[2] = vreinterpretq_s32_s8(vld1q_s8(p2));
-            _r0123.val[3] = vreinterpretq_s32_s8(vld1q_s8(p3));
-            vst4q_s32((int*)pp, _r0123);
-            pp += 64;
-            p0 += 16;
-            p1 += 16;
-            p2 += 16;
-            p3 += 16;
-        }
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            int32x2x4_t _r0123;
-            _r0123.val[0] = vreinterpret_s32_s8(vld1_s8(p0));
-            _r0123.val[1] = vreinterpret_s32_s8(vld1_s8(p1));
-            _r0123.val[2] = vreinterpret_s32_s8(vld1_s8(p2));
-            _r0123.val[3] = vreinterpret_s32_s8(vld1_s8(p3));
-            vst4_s32((int*)pp, _r0123);
-            pp += 32;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-        }
-#endif // __ARM_FEATURE_MATMUL_INT8
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            pp[0] = p0[0];
-            pp[1] = p0[1];
-            pp[2] = p0[2];
-            pp[3] = p0[3];
-            pp[4] = p1[0];
-            pp[5] = p1[1];
-            pp[6] = p1[2];
-            pp[7] = p1[3];
-            pp[8] = p2[0];
-            pp[9] = p2[1];
-            pp[10] = p2[2];
-            pp[11] = p2[3];
-            pp[12] = p3[0];
-            pp[13] = p3[1];
-            pp[14] = p3[2];
-            pp[15] = p3[3];
-            pp += 16;
-            p0 += 4;
-            p1 += 4;
-            p2 += 4;
-            p3 += 4;
-        }
-#else  // __ARM_FEATURE_DOTPROD
-        for (; kk + 15 < max_kk; kk += 16)
-        {
-            int16x8x4_t _r0123;
-            _r0123.val[0] = vreinterpretq_s16_s8(vld1q_s8(p0));
-            _r0123.val[1] = vreinterpretq_s16_s8(vld1q_s8(p1));
-            _r0123.val[2] = vreinterpretq_s16_s8(vld1q_s8(p2));
-            _r0123.val[3] = vreinterpretq_s16_s8(vld1q_s8(p3));
-            vst4q_s16((short*)pp, _r0123);
-            pp += 64;
-            p0 += 16;
-            p1 += 16;
-            p2 += 16;
-            p3 += 16;
-        }
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            int16x4x4_t _r0123;
-            _r0123.val[0] = vreinterpret_s16_s8(vld1_s8(p0));
-            _r0123.val[1] = vreinterpret_s16_s8(vld1_s8(p1));
-            _r0123.val[2] = vreinterpret_s16_s8(vld1_s8(p2));
-            _r0123.val[3] = vreinterpret_s16_s8(vld1_s8(p3));
-            vst4_s16((short*)pp, _r0123);
-            pp += 32;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-        }
-#endif // __ARM_FEATURE_DOTPROD
-        for (; kk + 1 < max_kk; kk += 2)
-        {
-            pp[0] = p0[0];
-            pp[1] = p0[1];
-            pp[2] = p1[0];
-            pp[3] = p1[1];
-            pp[4] = p2[0];
-            pp[5] = p2[1];
-            pp[6] = p3[0];
-            pp[7] = p3[1];
-            pp += 8;
-            p0 += 2;
-            p1 += 2;
-            p2 += 2;
-            p3 += 2;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp[1] = p1[0];
-            pp[2] = p2[0];
-            pp[3] = p3[0];
-            pp += 4;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        const signed char* p0 = (const signed char*)A + (i + ii) * A_hstep + k;
-        const signed char* p1 = (const signed char*)A + (i + ii + 1) * A_hstep + k;
-
-        int kk = 0;
-#if __ARM_NEON
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-        for (; kk + 15 < max_kk; kk += 16)
-        {
-            int64x2x2_t _r01;
-            _r01.val[0] = vreinterpretq_s64_s8(vld1q_s8(p0));
-            _r01.val[1] = vreinterpretq_s64_s8(vld1q_s8(p1));
-            vst2q_s64((int64_t*)pp, _r01);
-            pp += 32;
-            p0 += 16;
-            p1 += 16;
-        }
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            int8x8_t _r0 = vld1_s8(p0);
-            int8x8_t _r1 = vld1_s8(p1);
-            vst1_s8(pp, _r0);
-            vst1_s8(pp + 8, _r1);
-            pp += 16;
-            p0 += 8;
-            p1 += 8;
-        }
-#else  // __ARM_FEATURE_MATMUL_INT8
-        for (; kk + 15 < max_kk; kk += 16)
-        {
-            int32x4x2_t _r01;
-            _r01.val[0] = vreinterpretq_s32_s8(vld1q_s8(p0));
-            _r01.val[1] = vreinterpretq_s32_s8(vld1q_s8(p1));
-            vst2q_s32((int*)pp, _r01);
-            pp += 32;
-            p0 += 16;
-            p1 += 16;
-        }
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            int32x2x2_t _r01;
-            _r01.val[0] = vreinterpret_s32_s8(vld1_s8(p0));
-            _r01.val[1] = vreinterpret_s32_s8(vld1_s8(p1));
-            vst2_s32((int*)pp, _r01);
-            pp += 16;
-            p0 += 8;
-            p1 += 8;
-        }
-#endif // __ARM_FEATURE_MATMUL_INT8
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            pp[0] = p0[0];
-            pp[1] = p0[1];
-            pp[2] = p0[2];
-            pp[3] = p0[3];
-            pp[4] = p1[0];
-            pp[5] = p1[1];
-            pp[6] = p1[2];
-            pp[7] = p1[3];
-            pp += 8;
-            p0 += 4;
-            p1 += 4;
-        }
-#else  // __ARM_FEATURE_DOTPROD
-        for (; kk + 15 < max_kk; kk += 16)
-        {
-            int16x8x2_t _r01;
-            _r01.val[0] = vreinterpretq_s16_s8(vld1q_s8(p0));
-            _r01.val[1] = vreinterpretq_s16_s8(vld1q_s8(p1));
-            vst2q_s16((short*)pp, _r01);
-            pp += 32;
-            p0 += 16;
-            p1 += 16;
-        }
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            int16x4x2_t _r01;
-            _r01.val[0] = vreinterpret_s16_s8(vld1_s8(p0));
-            _r01.val[1] = vreinterpret_s16_s8(vld1_s8(p1));
-            vst2_s16((short*)pp, _r01);
-            pp += 16;
-            p0 += 8;
-            p1 += 8;
-        }
-#endif // __ARM_FEATURE_DOTPROD
-        for (; kk + 1 < max_kk; kk += 2)
-        {
-            pp[0] = p0[0];
-            pp[1] = p0[1];
-            pp[2] = p1[0];
-            pp[3] = p1[1];
-            pp += 4;
-            p0 += 2;
-            p1 += 2;
-        }
-#endif // __ARM_NEON
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp[1] = p1[0];
-            pp += 2;
-            p0++;
-            p1++;
-        }
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        const signed char* p0 = (const signed char*)A + (i + ii) * A_hstep + k;
-
-        int kk = 0;
-#if __ARM_NEON
-        for (; kk + 15 < max_kk; kk += 16)
-        {
-            vst1q_s8(pp, vld1q_s8(p0));
-            pp += 16;
-            p0 += 16;
-        }
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            vst1_s8(pp, vld1_s8(p0));
-            pp += 8;
-            p0 += 8;
-        }
-#endif // __ARM_NEON
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp += 1;
-            p0++;
-        }
-    }
-}
-
-static void convolution_gemm_transB_packed_tile_int8(const Mat& AT_tile, const Mat& BT_tile, Mat& topT_tile, Mat& top_blob, int i, int max_ii, int j, int max_jj, int k, int max_kk, bool k_end)
-{
-    // NCNN_LOGE("convolution_gemm_transB_packed_tile_int8 %d %d %d %d %d %d", i, max_ii, j, max_jj, k, max_kk);
-
-    const int out_elempack = top_blob.elempack;
-    const size_t out_hstep = top_blob.cstep;
-
-    const signed char* pAT = AT_tile;
-    const signed char* pBT = BT_tile;
-
-    int* outptr = topT_tile;
-
-    int ii = 0;
-#if __ARM_NEON
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        int* outptr0 = (int*)top_blob + (i + ii) * out_hstep + j * out_elempack;
-
-        const signed char* pB = pBT;
-
-        int jj = 0;
-#if __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            const signed char* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            asm volatile(
-#if !__ARM_FEATURE_MATMUL_INT8
-                "cmp    %w9, #0                     \n"
-                "beq    0f                          \n"
-
-                "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                "sub    %0, %0, #192                \n"
-                "b      1f                          \n"
-
-                "0:                                 \n"
-                "eor    v16.16b, v16.16b, v16.16b   \n"
-                "eor    v17.16b, v17.16b, v17.16b   \n"
-                "eor    v18.16b, v18.16b, v18.16b   \n"
-                "eor    v19.16b, v19.16b, v19.16b   \n"
-                "eor    v20.16b, v20.16b, v20.16b   \n"
-                "eor    v21.16b, v21.16b, v21.16b   \n"
-                "eor    v22.16b, v22.16b, v22.16b   \n"
-                "eor    v23.16b, v23.16b, v23.16b   \n"
-                "eor    v24.16b, v24.16b, v24.16b   \n"
-                "eor    v25.16b, v25.16b, v25.16b   \n"
-                "eor    v26.16b, v26.16b, v26.16b   \n"
-                "eor    v27.16b, v27.16b, v27.16b   \n"
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-                "eor    v29.16b, v29.16b, v29.16b   \n"
-                "eor    v30.16b, v30.16b, v30.16b   \n"
-                "eor    v31.16b, v31.16b, v31.16b   \n"
-
-                "1:                                 \n"
-#endif // !__ARM_FEATURE_MATMUL_INT8
-
-#if __ARM_FEATURE_DOTPROD
-                "lsr    w4, %w8, #3                 \n" // w4 = max_kk >> 3
-                "cmp    w4, #0                      \n"
-                "beq    101f                        \n"
-
-#if __ARM_FEATURE_MATMUL_INT8
-                "eor    v0.16b, v0.16b, v0.16b      \n"
-                "eor    v1.16b, v1.16b, v1.16b      \n"
-                "eor    v2.16b, v2.16b, v2.16b      \n"
-                "eor    v3.16b, v3.16b, v3.16b      \n"
-                "eor    v4.16b, v4.16b, v4.16b      \n"
-                "eor    v5.16b, v5.16b, v5.16b      \n"
-                "eor    v6.16b, v6.16b, v6.16b      \n"
-                "eor    v7.16b, v7.16b, v7.16b      \n"
-                "eor    v8.16b, v8.16b, v8.16b      \n"
-                "eor    v9.16b, v9.16b, v9.16b      \n"
-                "eor    v10.16b, v10.16b, v10.16b   \n"
-                "eor    v11.16b, v11.16b, v11.16b   \n"
-                "eor    v12.16b, v12.16b, v12.16b   \n"
-                "eor    v13.16b, v13.16b, v13.16b   \n"
-                "eor    v14.16b, v14.16b, v14.16b   \n"
-                "eor    v15.16b, v15.16b, v15.16b   \n"
-
-                "2:                                 \n"
-                "ld1    {v16.16b, v17.16b, v18.16b, v19.16b}, [%1], #64 \n"
-                "ld1    {v20.16b, v21.16b, v22.16b, v23.16b}, [%2], #64 \n"
-                "smmla  v0.4s, v16.16b, v20.16b     \n"
-                "smmla  v1.4s, v17.16b, v20.16b     \n"
-                "smmla  v2.4s, v16.16b, v21.16b     \n"
-                "smmla  v3.4s, v17.16b, v21.16b     \n"
-                "smmla  v4.4s, v18.16b, v20.16b     \n"
-                "smmla  v5.4s, v19.16b, v20.16b     \n"
-                "smmla  v6.4s, v18.16b, v21.16b     \n"
-                "smmla  v7.4s, v19.16b, v21.16b     \n"
-                "subs   w4, w4, #1                  \n"
-                "smmla  v8.4s, v16.16b, v22.16b     \n"
-                "smmla  v9.4s, v17.16b, v22.16b     \n"
-                "smmla  v10.4s, v16.16b, v23.16b    \n"
-                "smmla  v11.4s, v17.16b, v23.16b    \n"
-                "smmla  v12.4s, v18.16b, v22.16b    \n"
-                "smmla  v13.4s, v19.16b, v22.16b    \n"
-                "smmla  v14.4s, v18.16b, v23.16b    \n"
-                "smmla  v15.4s, v19.16b, v23.16b    \n"
-                "bne    2b                          \n"
-
-                "uzp1   v16.4s, v0.4s, v1.4s        \n"
-                "uzp2   v17.4s, v0.4s, v1.4s        \n"
-                "uzp1   v18.4s, v2.4s, v3.4s        \n"
-                "uzp2   v19.4s, v2.4s, v3.4s        \n"
-                "uzp1   v20.4s, v4.4s, v5.4s        \n"
-                "uzp2   v21.4s, v4.4s, v5.4s        \n"
-                "uzp1   v22.4s, v6.4s, v7.4s        \n"
-                "uzp2   v23.4s, v6.4s, v7.4s        \n"
-                "uzp1   v24.4s, v8.4s, v9.4s        \n"
-                "uzp2   v25.4s, v8.4s, v9.4s        \n"
-                "uzp1   v26.4s, v10.4s, v11.4s      \n"
-                "uzp2   v27.4s, v10.4s, v11.4s      \n"
-                "uzp1   v28.4s, v12.4s, v13.4s      \n"
-                "uzp2   v29.4s, v12.4s, v13.4s      \n"
-                "uzp1   v30.4s, v14.4s, v15.4s      \n"
-                "uzp2   v31.4s, v14.4s, v15.4s      \n"
-
-                "cmp    %w9, #0                     \n"
-                "beq    1f                          \n"
-
-                "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64   \n"
-                "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%0], #64   \n"
-                "ld1    {v8.4s, v9.4s, v10.4s, v11.4s}, [%0], #64 \n"
-                "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%0]    \n"
-                "sub    %0, %0, #192                \n"
-                "add    v16.4s, v16.4s, v0.4s       \n"
-                "add    v17.4s, v17.4s, v1.4s       \n"
-                "add    v18.4s, v18.4s, v2.4s       \n"
-                "add    v19.4s, v19.4s, v3.4s       \n"
-                "add    v20.4s, v20.4s, v4.4s       \n"
-                "add    v21.4s, v21.4s, v5.4s       \n"
-                "add    v22.4s, v22.4s, v6.4s       \n"
-                "add    v23.4s, v23.4s, v7.4s       \n"
-                "add    v24.4s, v24.4s, v8.4s       \n"
-                "add    v25.4s, v25.4s, v9.4s       \n"
-                "add    v26.4s, v26.4s, v10.4s      \n"
-                "add    v27.4s, v27.4s, v11.4s      \n"
-                "add    v28.4s, v28.4s, v12.4s      \n"
-                "add    v29.4s, v29.4s, v13.4s      \n"
-                "add    v30.4s, v30.4s, v14.4s      \n"
-                "add    v31.4s, v31.4s, v15.4s      \n"
-                "b      1f                          \n"
-#else  // __ARM_FEATURE_MATMUL_INT8
-                "2:                                 \n"
-                "ld1    {v0.16b, v1.16b, v2.16b, v3.16b}, [%1], #64 \n"
-                "ld1    {v4.16b, v5.16b, v6.16b, v7.16b}, [%2], #64 \n"
-                "sdot   v16.4s, v0.16b, v4.4b[0]    \n"
-                "sdot   v17.4s, v0.16b, v4.4b[1]    \n"
-                "sdot   v18.4s, v0.16b, v4.4b[2]    \n"
-                "sdot   v19.4s, v0.16b, v4.4b[3]    \n"
-                "sdot   v20.4s, v1.16b, v4.4b[0]    \n"
-                "sdot   v21.4s, v1.16b, v4.4b[1]    \n"
-                "sdot   v22.4s, v1.16b, v4.4b[2]    \n"
-                "sdot   v23.4s, v1.16b, v4.4b[3]    \n"
-                "sdot   v24.4s, v0.16b, v5.4b[0]    \n"
-                "sdot   v25.4s, v0.16b, v5.4b[1]    \n"
-                "sdot   v26.4s, v0.16b, v5.4b[2]    \n"
-                "sdot   v27.4s, v0.16b, v5.4b[3]    \n"
-                "sdot   v28.4s, v1.16b, v5.4b[0]    \n"
-                "sdot   v29.4s, v1.16b, v5.4b[1]    \n"
-                "sdot   v30.4s, v1.16b, v5.4b[2]    \n"
-                "sdot   v31.4s, v1.16b, v5.4b[3]    \n"
-                "subs   w4, w4, #1                  \n"
-                "sdot   v16.4s, v2.16b, v6.4b[0]    \n"
-                "sdot   v17.4s, v2.16b, v6.4b[1]    \n"
-                "sdot   v18.4s, v2.16b, v6.4b[2]    \n"
-                "sdot   v19.4s, v2.16b, v6.4b[3]    \n"
-                "sdot   v20.4s, v3.16b, v6.4b[0]    \n"
-                "sdot   v21.4s, v3.16b, v6.4b[1]    \n"
-                "sdot   v22.4s, v3.16b, v6.4b[2]    \n"
-                "sdot   v23.4s, v3.16b, v6.4b[3]    \n"
-                "sdot   v24.4s, v2.16b, v7.4b[0]    \n"
-                "sdot   v25.4s, v2.16b, v7.4b[1]    \n"
-                "sdot   v26.4s, v2.16b, v7.4b[2]    \n"
-                "sdot   v27.4s, v2.16b, v7.4b[3]    \n"
-                "sdot   v28.4s, v3.16b, v7.4b[0]    \n"
-                "sdot   v29.4s, v3.16b, v7.4b[1]    \n"
-                "sdot   v30.4s, v3.16b, v7.4b[2]    \n"
-                "sdot   v31.4s, v3.16b, v7.4b[3]    \n"
-                "bne    2b                          \n"
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                "101:                               \n"
-#if __ARM_FEATURE_MATMUL_INT8
-                "cmp    %w9, #0                     \n"
-                "beq    0f                          \n"
-
-                "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                "ld1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0]      \n"
-                "sub    %0, %0, #192                \n"
-                "b      1f                          \n"
-
-                "0:                                 \n"
-                "eor    v16.16b, v16.16b, v16.16b   \n"
-                "eor    v17.16b, v17.16b, v17.16b   \n"
-                "eor    v18.16b, v18.16b, v18.16b   \n"
-                "eor    v19.16b, v19.16b, v19.16b   \n"
-                "eor    v20.16b, v20.16b, v20.16b   \n"
-                "eor    v21.16b, v21.16b, v21.16b   \n"
-                "eor    v22.16b, v22.16b, v22.16b   \n"
-                "eor    v23.16b, v23.16b, v23.16b   \n"
-                "eor    v24.16b, v24.16b, v24.16b   \n"
-                "eor    v25.16b, v25.16b, v25.16b   \n"
-                "eor    v26.16b, v26.16b, v26.16b   \n"
-                "eor    v27.16b, v27.16b, v27.16b   \n"
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-                "eor    v29.16b, v29.16b, v29.16b   \n"
-                "eor    v30.16b, v30.16b, v30.16b   \n"
-                "eor    v31.16b, v31.16b, v31.16b   \n"
-                "1:                                 \n"
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                "and    w4, %w8, #4                 \n" // w4 = remain = max_kk & 4
-                "cmp    w4, #0                      \n"
-                "beq    3f                          \n"
-
-                // kk += 4 part
-                "ld1    {v0.16b, v1.16b}, [%1], #32 \n"
-                "ld1    {v2.16b, v3.16b}, [%2], #32 \n"
-                "sdot   v16.4s, v0.16b, v2.4b[0]    \n"
-                "sdot   v17.4s, v0.16b, v2.4b[1]    \n"
-                "sdot   v18.4s, v0.16b, v2.4b[2]    \n"
-                "sdot   v19.4s, v0.16b, v2.4b[3]    \n"
-                "sdot   v20.4s, v1.16b, v2.4b[0]    \n"
-                "sdot   v21.4s, v1.16b, v2.4b[1]    \n"
-                "sdot   v22.4s, v1.16b, v2.4b[2]    \n"
-                "sdot   v23.4s, v1.16b, v2.4b[3]    \n"
-                "sdot   v24.4s, v0.16b, v3.4b[0]    \n"
-                "sdot   v25.4s, v0.16b, v3.4b[1]    \n"
-                "sdot   v26.4s, v0.16b, v3.4b[2]    \n"
-                "sdot   v27.4s, v0.16b, v3.4b[3]    \n"
-                "sdot   v28.4s, v1.16b, v3.4b[0]    \n"
-                "sdot   v29.4s, v1.16b, v3.4b[1]    \n"
-                "sdot   v30.4s, v1.16b, v3.4b[2]    \n"
-                "sdot   v31.4s, v1.16b, v3.4b[3]    \n"
-#else  // __ARM_FEATURE_DOTPROD
-                "lsr    w4, %w8, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    3f                          \n"
-
-                "2:                                 \n"
-                "ld1    {v0.16b, v1.16b}, [%1], #32 \n"
-                "ld1    {v4.16b, v5.16b}, [%2], #32 \n"
-                "smull  v8.8h, v0.8b, v4.8b         \n"
-                "smull2 v9.8h, v0.16b, v4.16b       \n"
-                "rev64  v2.4s, v0.4s                \n"
-                "smull  v10.8h, v2.8b, v4.8b        \n"
-                "smull2 v11.8h, v2.16b, v4.16b      \n"
-                "rev64  v6.8h, v4.8h                \n"
-                "smull  v12.8h, v0.8b, v6.8b        \n"
-                "smull2 v13.8h, v0.16b, v6.16b      \n"
-                "rev64  v3.4s, v1.4s                \n"
-                "smull  v14.8h, v2.8b, v6.8b        \n"
-                "smull2 v15.8h, v2.16b, v6.16b      \n"
-                "rev64  v7.8h, v5.8h                \n"
-                "smlal  v8.8h, v1.8b, v5.8b         \n"
-                "smlal2 v9.8h, v1.16b, v5.16b       \n"
-                "smlal  v10.8h, v3.8b, v5.8b        \n"
-                "smlal2 v11.8h, v3.16b, v5.16b      \n"
-                "smlal  v12.8h, v1.8b, v7.8b        \n"
-                "smlal2 v13.8h, v1.16b, v7.16b      \n"
-                "smlal  v14.8h, v3.8b, v7.8b        \n"
-                "smlal2 v15.8h, v3.16b, v7.16b      \n"
-                "ext    v0.16b, v0.16b, v0.16b, #8  \n"
-                "ext    v2.16b, v2.16b, v2.16b, #8  \n"
-                "sadalp v16.4s, v8.8h               \n"
-                "sadalp v17.4s, v9.8h               \n"
-                "sadalp v20.4s, v10.8h              \n"
-                "sadalp v21.4s, v11.8h              \n"
-                "ext    v1.16b, v1.16b, v1.16b, #8  \n"
-                "ext    v3.16b, v3.16b, v3.16b, #8  \n"
-                "smull  v8.8h, v0.8b, v4.8b         \n"
-                "smull2 v9.8h, v0.16b, v4.16b       \n"
-                "smull  v10.8h, v2.8b, v4.8b        \n"
-                "smull2 v11.8h, v2.16b, v4.16b      \n"
-                "sadalp v24.4s, v12.8h              \n"
-                "sadalp v25.4s, v13.8h              \n"
-                "sadalp v28.4s, v14.8h              \n"
-                "sadalp v29.4s, v15.8h              \n"
-                "smull  v12.8h, v0.8b, v6.8b        \n"
-                "smull2 v13.8h, v0.16b, v6.16b      \n"
-                "smull  v14.8h, v2.8b, v6.8b        \n"
-                "smull2 v15.8h, v2.16b, v6.16b      \n"
-                "smlal  v8.8h, v1.8b, v5.8b         \n"
-                "smlal2 v9.8h, v1.16b, v5.16b       \n"
-                "smlal  v10.8h, v3.8b, v5.8b        \n"
-                "smlal2 v11.8h, v3.16b, v5.16b      \n"
-                "smlal  v12.8h, v1.8b, v7.8b        \n"
-                "smlal2 v13.8h, v1.16b, v7.16b      \n"
-                "smlal  v14.8h, v3.8b, v7.8b        \n"
-                "smlal2 v15.8h, v3.16b, v7.16b      \n"
-                "subs   w4, w4, #1                  \n"
-                "sadalp v18.4s, v8.8h               \n"
-                "sadalp v19.4s, v9.8h               \n"
-                "sadalp v22.4s, v10.8h              \n"
-                "sadalp v23.4s, v11.8h              \n"
-                "sadalp v26.4s, v12.8h              \n"
-                "sadalp v27.4s, v13.8h              \n"
-                "sadalp v30.4s, v14.8h              \n"
-                "sadalp v31.4s, v15.8h              \n"
-                "bne    2b                          \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "3:                                 \n"
-                "and    w4, %w8, #2                 \n" // w4 = remain = max_kk & 2
-                "cmp    w4, #0                      \n"
-                "beq    4f                          \n"
-
-                // kk += 2 part
-#if __ARM_FEATURE_DOTPROD
-                "ld1    {v0.16b}, [%1], #16         \n"
-                "ld1    {v1.16b}, [%2], #16         \n"
-                "dup    v4.8h, v1.h[0]              \n"
-                "dup    v5.8h, v1.h[1]              \n"
-                "dup    v6.8h, v1.h[2]              \n"
-                "dup    v7.8h, v1.h[3]              \n"
-                "smull  v8.8h, v0.8b, v4.8b         \n"
-                "smull  v9.8h, v0.8b, v5.8b         \n"
-                "smull  v10.8h, v0.8b, v6.8b        \n"
-                "smull  v11.8h, v0.8b, v7.8b        \n"
-                "smull2 v12.8h, v0.16b, v4.16b      \n"
-                "smull2 v13.8h, v0.16b, v5.16b      \n"
-                "smull2 v14.8h, v0.16b, v6.16b      \n"
-                "smull2 v15.8h, v0.16b, v7.16b      \n"
-                "sadalp v16.4s, v8.8h               \n"
-                "sadalp v17.4s, v9.8h               \n"
-                "sadalp v18.4s, v10.8h              \n"
-                "sadalp v19.4s, v11.8h              \n"
-                "sadalp v20.4s, v12.8h              \n"
-                "sadalp v21.4s, v13.8h              \n"
-                "sadalp v22.4s, v14.8h              \n"
-                "sadalp v23.4s, v15.8h              \n"
-                "dup    v4.8h, v1.h[4]              \n"
-                "dup    v5.8h, v1.h[5]              \n"
-                "dup    v6.8h, v1.h[6]              \n"
-                "dup    v7.8h, v1.h[7]              \n"
-                "smull  v8.8h, v0.8b, v4.8b         \n"
-                "smull  v9.8h, v0.8b, v5.8b         \n"
-                "smull  v10.8h, v0.8b, v6.8b        \n"
-                "smull  v11.8h, v0.8b, v7.8b        \n"
-                "smull2 v12.8h, v0.16b, v4.16b      \n"
-                "smull2 v13.8h, v0.16b, v5.16b      \n"
-                "smull2 v14.8h, v0.16b, v6.16b      \n"
-                "smull2 v15.8h, v0.16b, v7.16b      \n"
-                "sadalp v24.4s, v8.8h               \n"
-                "sadalp v25.4s, v9.8h               \n"
-                "sadalp v26.4s, v10.8h              \n"
-                "sadalp v27.4s, v11.8h              \n"
-                "sadalp v28.4s, v12.8h              \n"
-                "sadalp v29.4s, v13.8h              \n"
-                "sadalp v30.4s, v14.8h              \n"
-                "sadalp v31.4s, v15.8h              \n"
-#else  // __ARM_FEATURE_DOTPROD
-                "ld1    {v0.16b}, [%1], #16         \n"
-                "ld1    {v2.16b}, [%2], #16         \n"
-                "rev64  v1.4s, v0.4s                \n"
-                "rev64  v3.8h, v2.8h                \n"
-                "smull  v8.8h, v0.8b, v2.8b         \n"
-                "smull2 v9.8h, v0.16b, v2.16b       \n"
-                "smull  v10.8h, v1.8b, v2.8b        \n"
-                "smull2 v11.8h, v1.16b, v2.16b      \n"
-                "smull  v12.8h, v0.8b, v3.8b        \n"
-                "smull2 v13.8h, v0.16b, v3.16b      \n"
-                "smull  v14.8h, v1.8b, v3.8b        \n"
-                "smull2 v15.8h, v1.16b, v3.16b      \n"
-                "sadalp v16.4s, v8.8h               \n"
-                "sadalp v17.4s, v9.8h               \n"
-                "sadalp v20.4s, v10.8h              \n"
-                "sadalp v21.4s, v11.8h              \n"
-                "sadalp v24.4s, v12.8h              \n"
-                "sadalp v25.4s, v13.8h              \n"
-                "sadalp v28.4s, v14.8h              \n"
-                "sadalp v29.4s, v15.8h              \n"
-                "ext    v0.16b, v0.16b, v0.16b, #8  \n"
-                "ext    v1.16b, v1.16b, v1.16b, #8  \n"
-                "smull  v8.8h, v0.8b, v2.8b         \n"
-                "smull2 v9.8h, v0.16b, v2.16b       \n"
-                "smull  v10.8h, v1.8b, v2.8b        \n"
-                "smull2 v11.8h, v1.16b, v2.16b      \n"
-                "smull  v12.8h, v0.8b, v3.8b        \n"
-                "smull2 v13.8h, v0.16b, v3.16b      \n"
-                "smull  v14.8h, v1.8b, v3.8b        \n"
-                "smull2 v15.8h, v1.16b, v3.16b      \n"
-                "sadalp v18.4s, v8.8h               \n"
-                "sadalp v19.4s, v9.8h               \n"
-                "sadalp v22.4s, v10.8h              \n"
-                "sadalp v23.4s, v11.8h              \n"
-                "sadalp v26.4s, v12.8h              \n"
-                "sadalp v27.4s, v13.8h              \n"
-                "sadalp v30.4s, v14.8h              \n"
-                "sadalp v31.4s, v15.8h              \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "4:                                 \n"
-                "and    w4, %w8, #1                 \n" // w4 = remain = max_kk & 1
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                // kk += 1 part
-#if __ARM_FEATURE_DOTPROD
-                "ld1    {v0.8b}, [%1], #8           \n"
-                "ld1    {v1.8b}, [%2], #8           \n"
-                "dup    v8.8b, v1.b[0]              \n"
-                "dup    v9.8b, v1.b[1]              \n"
-                "dup    v10.8b, v1.b[2]             \n"
-                "dup    v11.8b, v1.b[3]             \n"
-                "dup    v12.8b, v1.b[4]             \n"
-                "dup    v13.8b, v1.b[5]             \n"
-                "dup    v14.8b, v1.b[6]             \n"
-                "dup    v15.8b, v1.b[7]             \n"
-                "smull  v8.8h, v0.8b, v8.8b         \n"
-                "smull  v9.8h, v0.8b, v9.8b         \n"
-                "smull  v10.8h, v0.8b, v10.8b       \n"
-                "smull  v11.8h, v0.8b, v11.8b       \n"
-                "smull  v12.8h, v0.8b, v12.8b       \n"
-                "smull  v13.8h, v0.8b, v13.8b       \n"
-                "smull  v14.8h, v0.8b, v14.8b       \n"
-                "smull  v15.8h, v0.8b, v15.8b       \n"
-                "saddw  v16.4s, v16.4s, v8.4h       \n"
-                "saddw  v17.4s, v17.4s, v9.4h       \n"
-                "saddw  v18.4s, v18.4s, v10.4h      \n"
-                "saddw  v19.4s, v19.4s, v11.4h      \n"
-                "saddw2 v20.4s, v20.4s, v8.8h       \n"
-                "saddw2 v21.4s, v21.4s, v9.8h       \n"
-                "saddw2 v22.4s, v22.4s, v10.8h      \n"
-                "saddw2 v23.4s, v23.4s, v11.8h      \n"
-                "saddw  v24.4s, v24.4s, v12.4h      \n"
-                "saddw  v25.4s, v25.4s, v13.4h      \n"
-                "saddw  v26.4s, v26.4s, v14.4h      \n"
-                "saddw  v27.4s, v27.4s, v15.4h      \n"
-                "saddw2 v28.4s, v28.4s, v12.8h      \n"
-                "saddw2 v29.4s, v29.4s, v13.8h      \n"
-                "saddw2 v30.4s, v30.4s, v14.8h      \n"
-                "saddw2 v31.4s, v31.4s, v15.8h      \n"
-#else  // __ARM_FEATURE_DOTPROD
-                "ld1    {v0.8b}, [%1], #8           \n"
-                "ld1    {v4.8b}, [%2], #8           \n"
-                "ext    v1.8b, v0.8b, v0.8b, #4     \n"
-                "rev32  v2.4h, v0.4h                \n"
-                "rev64  v3.4h, v0.4h                \n"
-                "rev32  v5.8b, v4.8b                \n"
-                "smull  v8.8h, v0.8b, v4.8b         \n"
-                "smull  v9.8h, v1.8b, v4.8b         \n"
-                "smull  v10.8h, v2.8b, v4.8b        \n"
-                "smull  v11.8h, v3.8b, v4.8b        \n"
-                "smull  v12.8h, v0.8b, v5.8b        \n"
-                "smull  v13.8h, v1.8b, v5.8b        \n"
-                "smull  v14.8h, v2.8b, v5.8b        \n"
-                "smull  v15.8h, v3.8b, v5.8b        \n"
-                "saddw  v16.4s, v16.4s, v8.4h       \n"
-                "saddw2 v17.4s, v17.4s, v8.8h       \n"
-                "saddw  v18.4s, v18.4s, v9.4h       \n"
-                "saddw2 v19.4s, v19.4s, v9.8h       \n"
-                "saddw  v20.4s, v20.4s, v10.4h      \n"
-                "saddw2 v21.4s, v21.4s, v10.8h      \n"
-                "saddw  v22.4s, v22.4s, v11.4h      \n"
-                "saddw2 v23.4s, v23.4s, v11.8h      \n"
-                "saddw  v24.4s, v24.4s, v12.4h      \n"
-                "saddw2 v25.4s, v25.4s, v12.8h      \n"
-                "saddw  v26.4s, v26.4s, v13.4h      \n"
-                "saddw2 v27.4s, v27.4s, v13.8h      \n"
-                "saddw  v28.4s, v28.4s, v14.4h      \n"
-                "saddw2 v29.4s, v29.4s, v14.8h      \n"
-                "saddw  v30.4s, v30.4s, v15.4h      \n"
-                "saddw2 v31.4s, v31.4s, v15.8h      \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "5:                                 \n"
-                "cmp    %w10, #0                    \n"
-                "beq    10f                         \n"
-
-#if __ARM_FEATURE_DOTPROD
-                // from
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                //      a2 b2 c2 d2
-                //      a3 b3 c3 d3
-                //      e0 f0 g0 h0
-                //      e1 f1 g1 h1
-                //      e2 f2 g2 h2
-                //      e3 f3 g3 h3
-                //      a4 b4 c4 d4
-                //      a5 b5 c5 d5
-                //      a6 b6 c6 d6
-                //      a7 b7 c7 d7
-                //      e4 f4 g4 h4
-                //      e5 f5 g5 h5
-                //      e6 f6 g6 h6
-                //      e7 f7 g7 h7
-                // if out_elempack == 4 / 8
-                "cmp    %w11, #1                    \n"
-                "beq    8f                          \n"
-
-                // if out_elempack == 8
-                "cmp    %w11, #8                    \n"
-                "bne    7f                          \n"
-
-                "st1    {v16.4s}, [%3], #16         \n"
-                "st1    {v20.4s}, [%3], #16         \n"
-                "st1    {v17.4s}, [%3], #16         \n"
-                "st1    {v21.4s}, [%3], #16         \n"
-                "st1    {v18.4s}, [%3], #16         \n"
-                "st1    {v22.4s}, [%3], #16         \n"
-                "st1    {v19.4s}, [%3], #16         \n"
-                "st1    {v23.4s}, [%3], #16         \n"
-                "st1    {v24.4s}, [%3], #16         \n"
-                "st1    {v28.4s}, [%3], #16         \n"
-                "st1    {v25.4s}, [%3], #16         \n"
-                "st1    {v29.4s}, [%3], #16         \n"
-                "st1    {v26.4s}, [%3], #16         \n"
-                "st1    {v30.4s}, [%3], #16         \n"
-                "st1    {v27.4s}, [%3], #16         \n"
-                "st1    {v31.4s}, [%3], #16         \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 4
-                "7:                                 \n"
-                "add    x4, %3, %12, lsl #4         \n"
-                "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%3], #64 \n"
-                "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [x4], #64 \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [x4] \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // to
-                //      a0 a1 a2 a3
-                //      a4 a5 a6 a7
-                //      b0 b1 b2 b3
-                //      b4 b5 b6 b7
-                //      c0 c1 c2 c3
-                //      c4 c5 c6 c7
-                //      d0 d1 d2 d3
-                //      d4 d5 d6 d7
-                //      e0 e1 e2 e3
-                //      e4 e5 e6 e7
-                //      f0 f1 f2 f3
-                //      f4 f5 f6 f7
-                //      g0 g1 g2 g3
-                //      g4 g5 g6 g7
-                //      h0 h1 h2 h3
-                //      h4 h5 h6 h7
-                "zip1   v0.4s, v16.4s, v17.4s       \n"
-                "zip2   v1.4s, v16.4s, v17.4s       \n"
-                "zip1   v2.4s, v18.4s, v19.4s       \n"
-                "zip2   v3.4s, v18.4s, v19.4s       \n"
-                "zip1   v4.4s, v24.4s, v25.4s       \n"
-                "zip2   v5.4s, v24.4s, v25.4s       \n"
-                "zip1   v6.4s, v26.4s, v27.4s       \n"
-                "zip2   v7.4s, v26.4s, v27.4s       \n"
-                "zip1   v8.4s, v20.4s, v21.4s       \n"
-                "zip2   v9.4s, v20.4s, v21.4s       \n"
-                "zip1   v10.4s, v22.4s, v23.4s      \n"
-                "zip2   v11.4s, v22.4s, v23.4s      \n"
-                "zip1   v12.4s, v28.4s, v29.4s      \n"
-                "zip2   v13.4s, v28.4s, v29.4s      \n"
-                "zip1   v14.4s, v30.4s, v31.4s      \n"
-                "zip2   v15.4s, v30.4s, v31.4s      \n"
-                "zip1   v16.2d, v0.2d, v2.2d        \n"
-                "zip1   v17.2d, v4.2d, v6.2d        \n"
-                "zip2   v18.2d, v0.2d, v2.2d        \n"
-                "zip2   v19.2d, v4.2d, v6.2d        \n"
-                "zip1   v20.2d, v1.2d, v3.2d        \n"
-                "zip1   v21.2d, v5.2d, v7.2d        \n"
-                "zip2   v22.2d, v1.2d, v3.2d        \n"
-                "zip2   v23.2d, v5.2d, v7.2d        \n"
-                "zip1   v24.2d, v8.2d, v10.2d       \n"
-                "zip1   v25.2d, v12.2d, v14.2d      \n"
-                "zip2   v26.2d, v8.2d, v10.2d       \n"
-                "zip2   v27.2d, v12.2d, v14.2d      \n"
-                "zip1   v28.2d, v9.2d, v11.2d       \n"
-                "zip1   v29.2d, v13.2d, v15.2d      \n"
-                "zip2   v30.2d, v9.2d, v11.2d       \n"
-                "zip2   v31.2d, v13.2d, v15.2d      \n"
-
-                "add    x4, %3, %12, lsl #2         \n"
-                "st1    {v16.4s, v17.4s}, [%3], #32 \n"
-                "st1    {v18.4s, v19.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v20.4s, v21.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v22.4s, v23.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v24.4s, v25.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v26.4s, v27.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v28.4s, v29.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v30.4s, v31.4s}, [x4]      \n"
-#else  // __ARM_FEATURE_DOTPROD
-
-                // from
-                //      a0 b1 c2 d3
-                //      e4 f5 g6 h7
-                //      e0 f1 g2 h3
-                //      a4 b5 c6 d7
-                //      c0 d1 a2 b3
-                //      g4 h5 e6 f7
-                //      g0 h1 e2 f3
-                //      c4 d5 a6 b7
-                //      a3 b2 c1 d0
-                //      e7 f6 g5 h4
-                //      e3 f2 g1 h0
-                //      a7 b6 c5 d4
-                //      c3 d2 a1 b0
-                //      g7 h6 e5 f4
-                //      g3 h2 e1 f0
-                //      c7 d6 a5 b4
-                // if out_elempack == 4 / 8
-                "cmp    %w11, #1                    \n"
-                "beq    8f                          \n"
-
-                "rev64  v24.4s, v24.4s              \n"
-                "rev64  v25.4s, v25.4s              \n"
-                "rev64  v26.4s, v26.4s              \n"
-                "rev64  v27.4s, v27.4s              \n"
-                "rev64  v28.4s, v28.4s              \n"
-                "rev64  v29.4s, v29.4s              \n"
-                "rev64  v30.4s, v30.4s              \n"
-                "rev64  v31.4s, v31.4s              \n"
-                "ext    v24.16b, v24.16b, v24.16b, #8 \n"
-                "ext    v25.16b, v25.16b, v25.16b, #8 \n"
-                "ext    v26.16b, v26.16b, v26.16b, #8 \n"
-                "ext    v27.16b, v27.16b, v27.16b, #8 \n"
-                "ext    v28.16b, v28.16b, v28.16b, #8 \n"
-                "ext    v29.16b, v29.16b, v29.16b, #8 \n"
-                "ext    v30.16b, v30.16b, v30.16b, #8 \n"
-                "ext    v31.16b, v31.16b, v31.16b, #8 \n"
-                "zip1   v0.4s, v16.4s, v28.4s       \n"
-                "zip2   v1.4s, v16.4s, v28.4s       \n"
-                "zip1   v2.4s, v20.4s, v24.4s       \n"
-                "zip2   v3.4s, v20.4s, v24.4s       \n"
-                "zip1   v4.4s, v18.4s, v30.4s       \n"
-                "zip2   v5.4s, v18.4s, v30.4s       \n"
-                "zip1   v6.4s, v22.4s, v26.4s       \n"
-                "zip2   v7.4s, v22.4s, v26.4s       \n"
-                "zip1   v8.4s, v19.4s, v31.4s       \n"
-                "zip2   v9.4s, v19.4s, v31.4s       \n"
-                "zip1   v10.4s, v23.4s, v27.4s      \n"
-                "zip2   v11.4s, v23.4s, v27.4s      \n"
-                "zip1   v12.4s, v17.4s, v29.4s      \n"
-                "zip2   v13.4s, v17.4s, v29.4s      \n"
-                "zip1   v14.4s, v21.4s, v25.4s      \n"
-                "zip2   v15.4s, v21.4s, v25.4s      \n"
-
-                // if out_elempack == 8
-                "cmp    %w11, #8                    \n"
-                "bne    7f                          \n"
-
-                // to
-                //      a0 b0 c0 d0
-                //      e0 f0 g0 h0
-                //      a1 b1 c1 d1
-                //      e1 f1 g1 h1
-                //      a2 b2 c2 d2
-                //      e2 f2 g2 h2
-                //      a3 b3 c3 d3
-                //      e3 f3 g3 h3
-                //      a4 b4 c4 d4
-                //      e4 f4 g4 h4
-                //      a5 b5 c5 d5
-                //      e5 f5 g5 h5
-                //      a6 b6 c6 d6
-                //      e6 f6 g6 h6
-                //      a7 b7 c7 d7
-                //      e7 f7 g7 h7
-                "zip1   v16.2d, v0.2d, v2.2d        \n"
-                "zip1   v17.2d, v4.2d, v6.2d        \n"
-                "zip2   v18.2d, v0.2d, v2.2d        \n"
-                "zip2   v19.2d, v4.2d, v6.2d        \n"
-                "zip1   v20.2d, v3.2d, v1.2d        \n"
-                "zip1   v21.2d, v7.2d, v5.2d        \n"
-                "zip2   v22.2d, v3.2d, v1.2d        \n"
-                "zip2   v23.2d, v7.2d, v5.2d        \n"
-                "zip1   v24.2d, v8.2d, v10.2d       \n"
-                "zip1   v25.2d, v12.2d, v14.2d      \n"
-                "zip2   v26.2d, v8.2d, v10.2d       \n"
-                "zip2   v27.2d, v12.2d, v14.2d      \n"
-                "zip1   v28.2d, v11.2d, v9.2d       \n"
-                "zip1   v29.2d, v15.2d, v13.2d      \n"
-                "zip2   v30.2d, v11.2d, v9.2d       \n"
-                "zip2   v31.2d, v15.2d, v13.2d      \n"
-                "rev64  v18.4s, v18.4s              \n"
-                "rev64  v19.4s, v19.4s              \n"
-                "rev64  v22.4s, v22.4s              \n"
-                "rev64  v23.4s, v23.4s              \n"
-                "rev64  v26.4s, v26.4s              \n"
-                "rev64  v27.4s, v27.4s              \n"
-                "rev64  v30.4s, v30.4s              \n"
-                "rev64  v31.4s, v31.4s              \n"
-
-                "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%3], #64 \n"
-                "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%3], #64 \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%3], #64 \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 4
-                "7:                                 \n"
-                // to
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                //      a2 b2 c2 d2
-                //      a3 b3 c3 d3
-                //      a4 b4 c4 d4
-                //      a5 b5 c5 d5
-                //      a6 b6 c6 d6
-                //      a7 b7 c7 d7
-                //      e0 f0 g0 h0
-                //      e1 f1 g1 h1
-                //      e2 f2 g2 h2
-                //      e3 f3 g3 h3
-                //      e4 f4 g4 h4
-                //      e5 f5 g5 h5
-                //      e6 f6 g6 h6
-                //      e7 f7 g7 h7
-                "zip1   v16.2d, v0.2d, v2.2d        \n"
-                "zip1   v24.2d, v4.2d, v6.2d        \n"
-                "zip2   v17.2d, v0.2d, v2.2d        \n"
-                "zip2   v25.2d, v4.2d, v6.2d        \n"
-                "zip1   v18.2d, v3.2d, v1.2d        \n"
-                "zip1   v26.2d, v7.2d, v5.2d        \n"
-                "zip2   v19.2d, v3.2d, v1.2d        \n"
-                "zip2   v27.2d, v7.2d, v5.2d        \n"
-                "zip1   v20.2d, v8.2d, v10.2d       \n"
-                "zip1   v28.2d, v12.2d, v14.2d      \n"
-                "zip2   v21.2d, v8.2d, v10.2d       \n"
-                "zip2   v29.2d, v12.2d, v14.2d      \n"
-                "zip1   v22.2d, v11.2d, v9.2d       \n"
-                "zip1   v30.2d, v15.2d, v13.2d      \n"
-                "zip2   v23.2d, v11.2d, v9.2d       \n"
-                "zip2   v31.2d, v15.2d, v13.2d      \n"
-                "rev64  v17.4s, v17.4s              \n"
-                "rev64  v19.4s, v19.4s              \n"
-                "rev64  v21.4s, v21.4s              \n"
-                "rev64  v23.4s, v23.4s              \n"
-                "rev64  v25.4s, v25.4s              \n"
-                "rev64  v27.4s, v27.4s              \n"
-                "rev64  v29.4s, v29.4s              \n"
-                "rev64  v31.4s, v31.4s              \n"
-
-                "add    x4, %3, %12, lsl #4         \n"
-                "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%3], #64 \n"
-                "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [x4], #64 \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [x4] \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // to
-                //      a0 a1 a2 a3
-                //      a4 a5 a6 a7
-                //      b0 b1 b2 b3
-                //      b4 b5 b6 b7
-                //      c0 c1 c2 c3
-                //      c4 c5 c6 c7
-                //      d0 d1 d2 d3
-                //      d4 d5 d6 d7
-                //      e0 e1 e2 e3
-                //      e4 e5 e6 e7
-                //      f0 f1 f2 f3
-                //      f4 f5 f6 f7
-                //      g0 g1 g2 g3
-                //      g4 g5 g6 g7
-                //      h0 h1 h2 h3
-                //      h4 h5 h6 h7
-                "ext    v20.16b, v20.16b, v20.16b, #8 \n"
-                "ext    v21.16b, v21.16b, v21.16b, #8 \n"
-                "ext    v22.16b, v22.16b, v22.16b, #8 \n"
-                "ext    v23.16b, v23.16b, v23.16b, #8 \n"
-                "ext    v28.16b, v28.16b, v28.16b, #8 \n"
-                "ext    v29.16b, v29.16b, v29.16b, #8 \n"
-                "ext    v30.16b, v30.16b, v30.16b, #8 \n"
-                "ext    v31.16b, v31.16b, v31.16b, #8 \n"
-                "zip1   v0.4s, v16.4s, v28.4s       \n"
-                "zip2   v1.4s, v16.4s, v28.4s       \n"
-                "zip1   v2.4s, v20.4s, v24.4s       \n"
-                "zip2   v3.4s, v20.4s, v24.4s       \n"
-                "zip1   v4.4s, v19.4s, v31.4s       \n"
-                "zip2   v5.4s, v19.4s, v31.4s       \n"
-                "zip1   v6.4s, v23.4s, v27.4s       \n"
-                "zip2   v7.4s, v23.4s, v27.4s       \n"
-                "zip1   v8.4s, v18.4s, v30.4s       \n"
-                "zip2   v9.4s, v18.4s, v30.4s       \n"
-                "zip1   v10.4s, v22.4s, v26.4s      \n"
-                "zip2   v11.4s, v22.4s, v26.4s      \n"
-                "zip1   v12.4s, v17.4s, v29.4s      \n"
-                "zip2   v13.4s, v17.4s, v29.4s      \n"
-                "zip1   v14.4s, v21.4s, v25.4s      \n"
-                "zip2   v15.4s, v21.4s, v25.4s      \n"
-                "zip1   v16.2d, v0.2d, v2.2d        \n"
-                "zip1   v17.2d, v4.2d, v6.2d        \n"
-                "zip2   v18.2d, v0.2d, v2.2d        \n"
-                "zip2   v19.2d, v4.2d, v6.2d        \n"
-                "zip1   v20.2d, v3.2d, v1.2d        \n"
-                "zip1   v21.2d, v7.2d, v5.2d        \n"
-                "zip2   v22.2d, v3.2d, v1.2d        \n"
-                "zip2   v23.2d, v7.2d, v5.2d        \n"
-                "zip1   v24.2d, v8.2d, v10.2d       \n"
-                "zip1   v25.2d, v12.2d, v14.2d      \n"
-                "zip2   v26.2d, v8.2d, v10.2d       \n"
-                "zip2   v27.2d, v12.2d, v14.2d      \n"
-                "zip1   v28.2d, v11.2d, v9.2d       \n"
-                "zip1   v29.2d, v15.2d, v13.2d      \n"
-                "zip2   v30.2d, v11.2d, v9.2d       \n"
-                "zip2   v31.2d, v15.2d, v13.2d      \n"
-                "rev64  v18.4s, v18.4s              \n"
-                "rev64  v19.4s, v19.4s              \n"
-                "rev64  v22.4s, v22.4s              \n"
-                "rev64  v23.4s, v23.4s              \n"
-                "rev64  v26.4s, v26.4s              \n"
-                "rev64  v27.4s, v27.4s              \n"
-                "rev64  v30.4s, v30.4s              \n"
-                "rev64  v31.4s, v31.4s              \n"
-
-                "add    x4, %3, %12, lsl #2         \n"
-                "st1    {v16.4s, v17.4s}, [%3], #32 \n"
-                "st1    {v18.4s, v19.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v20.4s, v21.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v22.4s, v23.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v24.4s, v25.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v26.4s, v27.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v28.4s, v29.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v30.4s, v31.4s}, [x4]      \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "9:                                 \n"
-                "add    %0, %0, #256                \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-                "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%0], #64 \n"
-                "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(max_kk),       // %8
-                "r"(k),            // %9
-                "r"(k_end),        // %10
-                "r"(out_elempack), // %11
-                "r"(out_hstep)     // %12
-                : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else // NCNN_GNU_INLINE_ASM
-            int32x4_t _sum0;
-            int32x4_t _sum1;
-            int32x4_t _sum2;
-            int32x4_t _sum3;
-            int32x4_t _sum4;
-            int32x4_t _sum5;
-            int32x4_t _sum6;
-            int32x4_t _sum7;
-            int32x4_t _sum8;
-            int32x4_t _sum9;
-            int32x4_t _suma;
-            int32x4_t _sumb;
-            int32x4_t _sumc;
-            int32x4_t _sumd;
-            int32x4_t _sume;
-            int32x4_t _sumf;
-
-#if __ARM_FEATURE_MATMUL_INT8
-            {
-                _sum0 = vdupq_n_s32(0);
-                _sum1 = vdupq_n_s32(0);
-                _sum2 = vdupq_n_s32(0);
-                _sum3 = vdupq_n_s32(0);
-                _sum4 = vdupq_n_s32(0);
-                _sum5 = vdupq_n_s32(0);
-                _sum6 = vdupq_n_s32(0);
-                _sum7 = vdupq_n_s32(0);
-                _sum8 = vdupq_n_s32(0);
-                _sum9 = vdupq_n_s32(0);
-                _suma = vdupq_n_s32(0);
-                _sumb = vdupq_n_s32(0);
-                _sumc = vdupq_n_s32(0);
-                _sumd = vdupq_n_s32(0);
-                _sume = vdupq_n_s32(0);
-                _sumf = vdupq_n_s32(0);
-            }
-#else  // __ARM_FEATURE_MATMUL_INT8
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_s32(0);
-                _sum1 = vdupq_n_s32(0);
-                _sum2 = vdupq_n_s32(0);
-                _sum3 = vdupq_n_s32(0);
-                _sum4 = vdupq_n_s32(0);
-                _sum5 = vdupq_n_s32(0);
-                _sum6 = vdupq_n_s32(0);
-                _sum7 = vdupq_n_s32(0);
-                _sum8 = vdupq_n_s32(0);
-                _sum9 = vdupq_n_s32(0);
-                _suma = vdupq_n_s32(0);
-                _sumb = vdupq_n_s32(0);
-                _sumc = vdupq_n_s32(0);
-                _sumd = vdupq_n_s32(0);
-                _sume = vdupq_n_s32(0);
-                _sumf = vdupq_n_s32(0);
-            }
-            else
-            {
-                _sum0 = vld1q_s32(outptr);
-                _sum1 = vld1q_s32(outptr + 4);
-                _sum2 = vld1q_s32(outptr + 8);
-                _sum3 = vld1q_s32(outptr + 12);
-                _sum4 = vld1q_s32(outptr + 16);
-                _sum5 = vld1q_s32(outptr + 20);
-                _sum6 = vld1q_s32(outptr + 24);
-                _sum7 = vld1q_s32(outptr + 28);
-                _sum8 = vld1q_s32(outptr + 32);
-                _sum9 = vld1q_s32(outptr + 36);
-                _suma = vld1q_s32(outptr + 40);
-                _sumb = vld1q_s32(outptr + 44);
-                _sumc = vld1q_s32(outptr + 48);
-                _sumd = vld1q_s32(outptr + 52);
-                _sume = vld1q_s32(outptr + 56);
-                _sumf = vld1q_s32(outptr + 60);
-            }
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-            int kk = 0;
-#if __ARM_FEATURE_MATMUL_INT8
-            {
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int8x16_t _pA0 = vld1q_s8(pA);
-                    int8x16_t _pA1 = vld1q_s8(pA + 16);
-                    int8x16_t _pA2 = vld1q_s8(pA + 32);
-                    int8x16_t _pA3 = vld1q_s8(pA + 48);
-                    int8x16_t _pB0 = vld1q_s8(pB);
-                    int8x16_t _pB1 = vld1q_s8(pB + 16);
-                    int8x16_t _pB2 = vld1q_s8(pB + 32);
-                    int8x16_t _pB3 = vld1q_s8(pB + 48);
-
-                    _sum0 = vmmlaq_s32(_sum0, _pA0, _pB0);
-                    _sum1 = vmmlaq_s32(_sum1, _pA1, _pB0);
-                    _sum2 = vmmlaq_s32(_sum2, _pA0, _pB1);
-                    _sum3 = vmmlaq_s32(_sum3, _pA1, _pB1);
-                    _sum4 = vmmlaq_s32(_sum4, _pA2, _pB0);
-                    _sum5 = vmmlaq_s32(_sum5, _pA3, _pB0);
-                    _sum6 = vmmlaq_s32(_sum6, _pA2, _pB1);
-                    _sum7 = vmmlaq_s32(_sum7, _pA3, _pB1);
-                    _sum8 = vmmlaq_s32(_sum8, _pA0, _pB2);
-                    _sum9 = vmmlaq_s32(_sum9, _pA1, _pB2);
-                    _suma = vmmlaq_s32(_suma, _pA0, _pB3);
-                    _sumb = vmmlaq_s32(_sumb, _pA1, _pB3);
-                    _sumc = vmmlaq_s32(_sumc, _pA2, _pB2);
-                    _sumd = vmmlaq_s32(_sumd, _pA3, _pB2);
-                    _sume = vmmlaq_s32(_sume, _pA2, _pB3);
-                    _sumf = vmmlaq_s32(_sumf, _pA3, _pB3);
-
-                    pA += 64;
-                    pB += 64;
-                }
-
-                int32x4x2_t _ss0 = vuzpq_s32(_sum0, _sum1);
-                int32x4x2_t _ss1 = vuzpq_s32(_sum2, _sum3);
-                int32x4x2_t _ss2 = vuzpq_s32(_sum4, _sum5);
-                int32x4x2_t _ss3 = vuzpq_s32(_sum6, _sum7);
-                int32x4x2_t _ss4 = vuzpq_s32(_sum8, _sum9);
-                int32x4x2_t _ss5 = vuzpq_s32(_suma, _sumb);
-                int32x4x2_t _ss6 = vuzpq_s32(_sumc, _sumd);
-                int32x4x2_t _ss7 = vuzpq_s32(_sume, _sumf);
-
-                if (k == 0)
-                {
-                    _sum0 = _ss0.val[0];
-                    _sum1 = _ss0.val[1];
-                    _sum2 = _ss1.val[0];
-                    _sum3 = _ss1.val[1];
-                    _sum4 = _ss2.val[0];
-                    _sum5 = _ss2.val[1];
-                    _sum6 = _ss3.val[0];
-                    _sum7 = _ss3.val[1];
-                    _sum8 = _ss4.val[0];
-                    _sum9 = _ss4.val[1];
-                    _suma = _ss5.val[0];
-                    _sumb = _ss5.val[1];
-                    _sumc = _ss6.val[0];
-                    _sumd = _ss6.val[1];
-                    _sume = _ss7.val[0];
-                    _sumf = _ss7.val[1];
-                }
-                else
-                {
-                    _sum0 = vld1q_s32(outptr);
-                    _sum1 = vld1q_s32(outptr + 4);
-                    _sum2 = vld1q_s32(outptr + 8);
-                    _sum3 = vld1q_s32(outptr + 12);
-                    _sum4 = vld1q_s32(outptr + 16);
-                    _sum5 = vld1q_s32(outptr + 20);
-                    _sum6 = vld1q_s32(outptr + 24);
-                    _sum7 = vld1q_s32(outptr + 28);
-                    _sum8 = vld1q_s32(outptr + 32);
-                    _sum9 = vld1q_s32(outptr + 36);
-                    _suma = vld1q_s32(outptr + 40);
-                    _sumb = vld1q_s32(outptr + 44);
-                    _sumc = vld1q_s32(outptr + 48);
-                    _sumd = vld1q_s32(outptr + 52);
-                    _sume = vld1q_s32(outptr + 56);
-                    _sumf = vld1q_s32(outptr + 60);
-
-                    _sum0 = vaddq_s32(_sum0, _ss0.val[0]);
-                    _sum1 = vaddq_s32(_sum1, _ss0.val[1]);
-                    _sum2 = vaddq_s32(_sum2, _ss1.val[0]);
-                    _sum3 = vaddq_s32(_sum3, _ss1.val[1]);
-                    _sum4 = vaddq_s32(_sum4, _ss2.val[0]);
-                    _sum5 = vaddq_s32(_sum5, _ss2.val[1]);
-                    _sum6 = vaddq_s32(_sum6, _ss3.val[0]);
-                    _sum7 = vaddq_s32(_sum7, _ss3.val[1]);
-                    _sum8 = vaddq_s32(_sum8, _ss4.val[0]);
-                    _sum9 = vaddq_s32(_sum9, _ss4.val[1]);
-                    _suma = vaddq_s32(_suma, _ss5.val[0]);
-                    _sumb = vaddq_s32(_sumb, _ss5.val[1]);
-                    _sumc = vaddq_s32(_sumc, _ss6.val[0]);
-                    _sumd = vaddq_s32(_sumd, _ss6.val[1]);
-                    _sume = vaddq_s32(_sume, _ss7.val[0]);
-                    _sumf = vaddq_s32(_sumf, _ss7.val[1]);
-                }
-            }
-#elif __ARM_FEATURE_DOTPROD
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                int8x16_t _pA0 = vld1q_s8(pA);
-                int8x16_t _pA1 = vld1q_s8(pA + 16);
-                int8x16_t _pA2 = vld1q_s8(pA + 32);
-                int8x16_t _pA3 = vld1q_s8(pA + 48);
-                int8x16_t _pB0 = vld1q_s8(pB);
-                int8x16_t _pB1 = vld1q_s8(pB + 16);
-                int8x16_t _pB2 = vld1q_s8(pB + 32);
-                int8x16_t _pB3 = vld1q_s8(pB + 48);
-
-                // aaaa bbbb cccc dddd    eeee ffff gggg hhhh
-
-                // 0000 1111 2222 3333    4444 5555 6666 7777
-                _sum0 = vdotq_laneq_s32(_sum0, _pA0, _pB0, 0);
-                _sum1 = vdotq_laneq_s32(_sum1, _pA0, _pB0, 1);
-                _sum2 = vdotq_laneq_s32(_sum2, _pA0, _pB0, 2);
-                _sum3 = vdotq_laneq_s32(_sum3, _pA0, _pB0, 3);
-                _sum4 = vdotq_laneq_s32(_sum4, _pA1, _pB0, 0);
-                _sum5 = vdotq_laneq_s32(_sum5, _pA1, _pB0, 1);
-                _sum6 = vdotq_laneq_s32(_sum6, _pA1, _pB0, 2);
-                _sum7 = vdotq_laneq_s32(_sum7, _pA1, _pB0, 3);
-                _sum8 = vdotq_laneq_s32(_sum8, _pA0, _pB1, 0);
-                _sum9 = vdotq_laneq_s32(_sum9, _pA0, _pB1, 1);
-                _suma = vdotq_laneq_s32(_suma, _pA0, _pB1, 2);
-                _sumb = vdotq_laneq_s32(_sumb, _pA0, _pB1, 3);
-                _sumc = vdotq_laneq_s32(_sumc, _pA1, _pB1, 0);
-                _sumd = vdotq_laneq_s32(_sumd, _pA1, _pB1, 1);
-                _sume = vdotq_laneq_s32(_sume, _pA1, _pB1, 2);
-                _sumf = vdotq_laneq_s32(_sumf, _pA1, _pB1, 3);
-
-                _sum0 = vdotq_laneq_s32(_sum0, _pA2, _pB2, 0);
-                _sum1 = vdotq_laneq_s32(_sum1, _pA2, _pB2, 1);
-                _sum2 = vdotq_laneq_s32(_sum2, _pA2, _pB2, 2);
-                _sum3 = vdotq_laneq_s32(_sum3, _pA2, _pB2, 3);
-                _sum4 = vdotq_laneq_s32(_sum4, _pA3, _pB2, 0);
-                _sum5 = vdotq_laneq_s32(_sum5, _pA3, _pB2, 1);
-                _sum6 = vdotq_laneq_s32(_sum6, _pA3, _pB2, 2);
-                _sum7 = vdotq_laneq_s32(_sum7, _pA3, _pB2, 3);
-                _sum8 = vdotq_laneq_s32(_sum8, _pA2, _pB3, 0);
-                _sum9 = vdotq_laneq_s32(_sum9, _pA2, _pB3, 1);
-                _suma = vdotq_laneq_s32(_suma, _pA2, _pB3, 2);
-                _sumb = vdotq_laneq_s32(_sumb, _pA2, _pB3, 3);
-                _sumc = vdotq_laneq_s32(_sumc, _pA3, _pB3, 0);
-                _sumd = vdotq_laneq_s32(_sumd, _pA3, _pB3, 1);
-                _sume = vdotq_laneq_s32(_sume, _pA3, _pB3, 2);
-                _sumf = vdotq_laneq_s32(_sumf, _pA3, _pB3, 3);
-
-                pA += 64;
-                pB += 64;
-            }
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x16_t _pA0 = vld1q_s8(pA);
-                int8x16_t _pA1 = vld1q_s8(pA + 16);
-                int8x16_t _pB0 = vld1q_s8(pB);
-                int8x16_t _pB1 = vld1q_s8(pB + 16);
-
-                // aaaa bbbb cccc dddd    eeee ffff gggg hhhh
-
-                // 0000 1111 2222 3333    4444 5555 6666 7777
-                _sum0 = vdotq_laneq_s32(_sum0, _pA0, _pB0, 0);
-                _sum1 = vdotq_laneq_s32(_sum1, _pA0, _pB0, 1);
-                _sum2 = vdotq_laneq_s32(_sum2, _pA0, _pB0, 2);
-                _sum3 = vdotq_laneq_s32(_sum3, _pA0, _pB0, 3);
-                _sum4 = vdotq_laneq_s32(_sum4, _pA1, _pB0, 0);
-                _sum5 = vdotq_laneq_s32(_sum5, _pA1, _pB0, 1);
-                _sum6 = vdotq_laneq_s32(_sum6, _pA1, _pB0, 2);
-                _sum7 = vdotq_laneq_s32(_sum7, _pA1, _pB0, 3);
-                _sum8 = vdotq_laneq_s32(_sum8, _pA0, _pB1, 0);
-                _sum9 = vdotq_laneq_s32(_sum9, _pA0, _pB1, 1);
-                _suma = vdotq_laneq_s32(_suma, _pA0, _pB1, 2);
-                _sumb = vdotq_laneq_s32(_sumb, _pA0, _pB1, 3);
-                _sumc = vdotq_laneq_s32(_sumc, _pA1, _pB1, 0);
-                _sumd = vdotq_laneq_s32(_sumd, _pA1, _pB1, 1);
-                _sume = vdotq_laneq_s32(_sume, _pA1, _pB1, 2);
-                _sumf = vdotq_laneq_s32(_sumf, _pA1, _pB1, 3);
-
-#else  // __ARM_FEATURE_DOTPROD
-                int8x16_t _pA0 = vld1q_s8(pA);
-                int8x16_t _pA2 = vld1q_s8(pA + 16);
-                int8x16_t _pB0 = vld1q_s8(pB);
-                int8x16_t _pB2 = vld1q_s8(pB + 16);
-
-                // aabbccdd eeffgghh
-                // ccddaabb gghheeff
-
-                int8x16_t _pA1 = vreinterpretq_s8_s32(vrev64q_s32(vreinterpretq_s32_s8(_pA0)));
-
-                // 00112233 44556677
-                // 33221100 77665544
-
-                int8x16_t _pB1 = vreinterpretq_s8_s16(vrev64q_s16(vreinterpretq_s16_s8(_pB0)));
-
-                // aabbccdd eeffgghh
-                // ccddaabb gghheeff
-
-                int8x16_t _pA3 = vreinterpretq_s8_s32(vrev64q_s32(vreinterpretq_s32_s8(_pA2)));
-
-                // 00112233 44556677
-                // 33221100 77665544
-
-                int8x16_t _pB3 = vreinterpretq_s8_s16(vrev64q_s16(vreinterpretq_s16_s8(_pB2)));
-
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA0), vget_low_s8(_pB0));
-                int16x8_t _s1 = vmull_s8(vget_high_s8(_pA0), vget_high_s8(_pB0));
-                int16x8_t _s2 = vmull_s8(vget_high_s8(_pA0), vget_low_s8(_pB0));
-                int16x8_t _s3 = vmull_s8(vget_low_s8(_pA0), vget_high_s8(_pB0));
-                int16x8_t _s4 = vmull_s8(vget_low_s8(_pA1), vget_low_s8(_pB0));
-                int16x8_t _s5 = vmull_s8(vget_high_s8(_pA1), vget_high_s8(_pB0));
-                int16x8_t _s6 = vmull_s8(vget_high_s8(_pA1), vget_low_s8(_pB0));
-                int16x8_t _s7 = vmull_s8(vget_low_s8(_pA1), vget_high_s8(_pB0));
-                int16x8_t _s8 = vmull_s8(vget_low_s8(_pA0), vget_low_s8(_pB1));
-                int16x8_t _s9 = vmull_s8(vget_high_s8(_pA0), vget_high_s8(_pB1));
-                int16x8_t _sa = vmull_s8(vget_high_s8(_pA0), vget_low_s8(_pB1));
-                int16x8_t _sb = vmull_s8(vget_low_s8(_pA0), vget_high_s8(_pB1));
-                int16x8_t _sc = vmull_s8(vget_low_s8(_pA1), vget_low_s8(_pB1));
-                int16x8_t _sd = vmull_s8(vget_high_s8(_pA1), vget_high_s8(_pB1));
-                int16x8_t _se = vmull_s8(vget_high_s8(_pA1), vget_low_s8(_pB1));
-                int16x8_t _sf = vmull_s8(vget_low_s8(_pA1), vget_high_s8(_pB1));
-
-                _s0 = vmlal_s8(_s0, vget_low_s8(_pA2), vget_low_s8(_pB2));
-                _s1 = vmlal_s8(_s1, vget_high_s8(_pA2), vget_high_s8(_pB2));
-                _s2 = vmlal_s8(_s2, vget_high_s8(_pA2), vget_low_s8(_pB2));
-                _s3 = vmlal_s8(_s3, vget_low_s8(_pA2), vget_high_s8(_pB2));
-                _s4 = vmlal_s8(_s4, vget_low_s8(_pA3), vget_low_s8(_pB2));
-                _s5 = vmlal_s8(_s5, vget_high_s8(_pA3), vget_high_s8(_pB2));
-                _s6 = vmlal_s8(_s6, vget_high_s8(_pA3), vget_low_s8(_pB2));
-                _s7 = vmlal_s8(_s7, vget_low_s8(_pA3), vget_high_s8(_pB2));
-                _s8 = vmlal_s8(_s8, vget_low_s8(_pA2), vget_low_s8(_pB3));
-                _s9 = vmlal_s8(_s9, vget_high_s8(_pA2), vget_high_s8(_pB3));
-                _sa = vmlal_s8(_sa, vget_high_s8(_pA2), vget_low_s8(_pB3));
-                _sb = vmlal_s8(_sb, vget_low_s8(_pA2), vget_high_s8(_pB3));
-                _sc = vmlal_s8(_sc, vget_low_s8(_pA3), vget_low_s8(_pB3));
-                _sd = vmlal_s8(_sd, vget_high_s8(_pA3), vget_high_s8(_pB3));
-                _se = vmlal_s8(_se, vget_high_s8(_pA3), vget_low_s8(_pB3));
-                _sf = vmlal_s8(_sf, vget_low_s8(_pA3), vget_high_s8(_pB3));
-
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-                _sum4 = vpadalq_s16(_sum4, _s4);
-                _sum5 = vpadalq_s16(_sum5, _s5);
-                _sum6 = vpadalq_s16(_sum6, _s6);
-                _sum7 = vpadalq_s16(_sum7, _s7);
-                _sum8 = vpadalq_s16(_sum8, _s8);
-                _sum9 = vpadalq_s16(_sum9, _s9);
-                _suma = vpadalq_s16(_suma, _sa);
-                _sumb = vpadalq_s16(_sumb, _sb);
-                _sumc = vpadalq_s16(_sumc, _sc);
-                _sumd = vpadalq_s16(_sumd, _sd);
-                _sume = vpadalq_s16(_sume, _se);
-                _sumf = vpadalq_s16(_sumf, _sf);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 32;
-                pB += 32;
-            }
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x16_t _pA = vld1q_s8(pA);
-                int8x16_t _pB = vld1q_s8(pB);
-
-                // aabbccdd eeffgghh
-
-                // 00112233 44556677
-
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_low_s8(_pB)), 0)));
-                int16x8_t _s1 = vmull_s8(vget_low_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_low_s8(_pB)), 1)));
-                int16x8_t _s2 = vmull_s8(vget_low_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_low_s8(_pB)), 2)));
-                int16x8_t _s3 = vmull_s8(vget_low_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_low_s8(_pB)), 3)));
-                int16x8_t _s4 = vmull_s8(vget_high_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_low_s8(_pB)), 0)));
-                int16x8_t _s5 = vmull_s8(vget_high_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_low_s8(_pB)), 1)));
-                int16x8_t _s6 = vmull_s8(vget_high_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_low_s8(_pB)), 2)));
-                int16x8_t _s7 = vmull_s8(vget_high_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_low_s8(_pB)), 3)));
-                int16x8_t _s8 = vmull_s8(vget_low_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_high_s8(_pB)), 0)));
-                int16x8_t _s9 = vmull_s8(vget_low_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_high_s8(_pB)), 1)));
-                int16x8_t _sa = vmull_s8(vget_low_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_high_s8(_pB)), 2)));
-                int16x8_t _sb = vmull_s8(vget_low_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_high_s8(_pB)), 3)));
-                int16x8_t _sc = vmull_s8(vget_high_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_high_s8(_pB)), 0)));
-                int16x8_t _sd = vmull_s8(vget_high_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_high_s8(_pB)), 1)));
-                int16x8_t _se = vmull_s8(vget_high_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_high_s8(_pB)), 2)));
-                int16x8_t _sf = vmull_s8(vget_high_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_high_s8(_pB)), 3)));
-
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-                _sum4 = vpadalq_s16(_sum4, _s4);
-                _sum5 = vpadalq_s16(_sum5, _s5);
-                _sum6 = vpadalq_s16(_sum6, _s6);
-                _sum7 = vpadalq_s16(_sum7, _s7);
-                _sum8 = vpadalq_s16(_sum8, _s8);
-                _sum9 = vpadalq_s16(_sum9, _s9);
-                _suma = vpadalq_s16(_suma, _sa);
-                _sumb = vpadalq_s16(_sumb, _sb);
-                _sumc = vpadalq_s16(_sumc, _sc);
-                _sumd = vpadalq_s16(_sumd, _sd);
-                _sume = vpadalq_s16(_sume, _se);
-                _sumf = vpadalq_s16(_sumf, _sf);
-#else  // __ARM_FEATURE_DOTPROD
-                int8x16_t _pA0 = vld1q_s8(pA);
-                int8x16_t _pB0 = vld1q_s8(pB);
-
-                // aabbccdd eeffgghh
-
-                // ccddaabb gghheeff
-
-                int8x16_t _pA1 = vreinterpretq_s8_s32(vrev64q_s32(vreinterpretq_s32_s8(_pA0)));
-
-                // 00112233 44556677
-
-                // 33221100 77665544
-
-                int8x16_t _pB1 = vreinterpretq_s8_s16(vrev64q_s16(vreinterpretq_s16_s8(_pB0)));
-
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA0), vget_low_s8(_pB0));
-                int16x8_t _s1 = vmull_s8(vget_high_s8(_pA0), vget_high_s8(_pB0));
-                int16x8_t _s2 = vmull_s8(vget_high_s8(_pA0), vget_low_s8(_pB0));
-                int16x8_t _s3 = vmull_s8(vget_low_s8(_pA0), vget_high_s8(_pB0));
-                int16x8_t _s4 = vmull_s8(vget_low_s8(_pA1), vget_low_s8(_pB0));
-                int16x8_t _s5 = vmull_s8(vget_high_s8(_pA1), vget_high_s8(_pB0));
-                int16x8_t _s6 = vmull_s8(vget_high_s8(_pA1), vget_low_s8(_pB0));
-                int16x8_t _s7 = vmull_s8(vget_low_s8(_pA1), vget_high_s8(_pB0));
-                int16x8_t _s8 = vmull_s8(vget_low_s8(_pA0), vget_low_s8(_pB1));
-                int16x8_t _s9 = vmull_s8(vget_high_s8(_pA0), vget_high_s8(_pB1));
-                int16x8_t _sa = vmull_s8(vget_high_s8(_pA0), vget_low_s8(_pB1));
-                int16x8_t _sb = vmull_s8(vget_low_s8(_pA0), vget_high_s8(_pB1));
-                int16x8_t _sc = vmull_s8(vget_low_s8(_pA1), vget_low_s8(_pB1));
-                int16x8_t _sd = vmull_s8(vget_high_s8(_pA1), vget_high_s8(_pB1));
-                int16x8_t _se = vmull_s8(vget_high_s8(_pA1), vget_low_s8(_pB1));
-                int16x8_t _sf = vmull_s8(vget_low_s8(_pA1), vget_high_s8(_pB1));
-
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-                _sum4 = vpadalq_s16(_sum4, _s4);
-                _sum5 = vpadalq_s16(_sum5, _s5);
-                _sum6 = vpadalq_s16(_sum6, _s6);
-                _sum7 = vpadalq_s16(_sum7, _s7);
-                _sum8 = vpadalq_s16(_sum8, _s8);
-                _sum9 = vpadalq_s16(_sum9, _s9);
-                _suma = vpadalq_s16(_suma, _sa);
-                _sumb = vpadalq_s16(_sumb, _sb);
-                _sumc = vpadalq_s16(_sumc, _sc);
-                _sumd = vpadalq_s16(_sumd, _sd);
-                _sume = vpadalq_s16(_sume, _se);
-                _sumf = vpadalq_s16(_sumf, _sf);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 16;
-                pB += 16;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x8_t _pA = vld1_s8(pA);
-                // int8x8_t _pB0 = vld1_s8(pB);
-
-                // abcd efgh
-                // 0123 4567
-
-                int16x8_t _s01 = vmull_s8(_pA, vdup_n_s8(pB[0]));
-                int16x8_t _s23 = vmull_s8(_pA, vdup_n_s8(pB[1]));
-                int16x8_t _s45 = vmull_s8(_pA, vdup_n_s8(pB[2]));
-                int16x8_t _s67 = vmull_s8(_pA, vdup_n_s8(pB[3]));
-                int16x8_t _s89 = vmull_s8(_pA, vdup_n_s8(pB[4]));
-                int16x8_t _sab = vmull_s8(_pA, vdup_n_s8(pB[5]));
-                int16x8_t _scd = vmull_s8(_pA, vdup_n_s8(pB[6]));
-                int16x8_t _sef = vmull_s8(_pA, vdup_n_s8(pB[7]));
-
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s01));
-                _sum1 = vaddw_s16(_sum1, vget_low_s16(_s23));
-                _sum2 = vaddw_s16(_sum2, vget_low_s16(_s45));
-                _sum3 = vaddw_s16(_sum3, vget_low_s16(_s67));
-                _sum4 = vaddw_s16(_sum4, vget_high_s16(_s01));
-                _sum5 = vaddw_s16(_sum5, vget_high_s16(_s23));
-                _sum6 = vaddw_s16(_sum6, vget_high_s16(_s45));
-                _sum7 = vaddw_s16(_sum7, vget_high_s16(_s67));
-                _sum8 = vaddw_s16(_sum8, vget_low_s16(_s89));
-                _sum9 = vaddw_s16(_sum9, vget_low_s16(_sab));
-                _suma = vaddw_s16(_suma, vget_low_s16(_scd));
-                _sumb = vaddw_s16(_sumb, vget_low_s16(_sef));
-                _sumc = vaddw_s16(_sumc, vget_high_s16(_s89));
-                _sumd = vaddw_s16(_sumd, vget_high_s16(_sab));
-                _sume = vaddw_s16(_sume, vget_high_s16(_scd));
-                _sumf = vaddw_s16(_sumf, vget_high_s16(_sef));
-#else  // __ARM_FEATURE_DOTPROD
-                int8x8_t _pA0 = vld1_s8(pA);
-                int8x8_t _pB0 = vld1_s8(pB);
-
-                // abcd efgh
-                // efgh abcd
-                // cdab ghef
-                // ghef cdab
-
-                // 0123 4567
-                // 3210 7654
-
-                // abcdefgh  ->  ghefcdab  ->  cdabghef
-
-                int8x8_t _pA1 = vext_s8(_pA0, _pA0, 4);
-                int8x8_t _pA2 = vreinterpret_s8_s16(vrev32_s16(vreinterpret_s16_s8(_pA0)));
-                int8x8_t _pA3 = vreinterpret_s8_s16(vrev64_s16(vreinterpret_s16_s8(_pA0)));
-
-                // 01234567  ->  32107654
-
-                int8x8_t _pB1 = vrev32_s8(_pB0);
-
-                int16x8_t _s01 = vmull_s8(_pA0, _pB0);
-                int16x8_t _s23 = vmull_s8(_pA1, _pB0);
-                int16x8_t _s45 = vmull_s8(_pA2, _pB0);
-                int16x8_t _s67 = vmull_s8(_pA3, _pB0);
-                int16x8_t _s89 = vmull_s8(_pA0, _pB1);
-                int16x8_t _sab = vmull_s8(_pA1, _pB1);
-                int16x8_t _scd = vmull_s8(_pA2, _pB1);
-                int16x8_t _sef = vmull_s8(_pA3, _pB1);
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s01));
-                _sum1 = vaddw_s16(_sum1, vget_high_s16(_s01));
-                _sum2 = vaddw_s16(_sum2, vget_low_s16(_s23));
-                _sum3 = vaddw_s16(_sum3, vget_high_s16(_s23));
-                _sum4 = vaddw_s16(_sum4, vget_low_s16(_s45));
-                _sum5 = vaddw_s16(_sum5, vget_high_s16(_s45));
-                _sum6 = vaddw_s16(_sum6, vget_low_s16(_s67));
-                _sum7 = vaddw_s16(_sum7, vget_high_s16(_s67));
-                _sum8 = vaddw_s16(_sum8, vget_low_s16(_s89));
-                _sum9 = vaddw_s16(_sum9, vget_high_s16(_s89));
-                _suma = vaddw_s16(_suma, vget_low_s16(_sab));
-                _sumb = vaddw_s16(_sumb, vget_high_s16(_sab));
-                _sumc = vaddw_s16(_sumc, vget_low_s16(_scd));
-                _sumd = vaddw_s16(_sumd, vget_high_s16(_scd));
-                _sume = vaddw_s16(_sume, vget_low_s16(_sef));
-                _sumf = vaddw_s16(_sumf, vget_high_s16(_sef));
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 8;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-#if __ARM_FEATURE_DOTPROD
-                // from
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                //      a2 b2 c2 d2
-                //      a3 b3 c3 d3
-                //      e0 f0 g0 h0
-                //      e1 f1 g1 h1
-                //      e2 f2 g2 h2
-                //      e3 f3 g3 h3
-                //      a4 b4 c4 d4
-                //      a5 b5 c5 d5
-                //      a6 b6 c6 d6
-                //      a7 b7 c7 d7
-                //      e4 f4 g4 h4
-                //      e5 f5 g5 h5
-                //      e6 f6 g6 h6
-                //      e7 f7 g7 h7
-                if (out_elempack == 8)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum4);
-                    vst1q_s32(outptr0 + 8, _sum1);
-                    vst1q_s32(outptr0 + 12, _sum5);
-                    vst1q_s32(outptr0 + 16, _sum2);
-                    vst1q_s32(outptr0 + 20, _sum6);
-                    vst1q_s32(outptr0 + 24, _sum3);
-                    vst1q_s32(outptr0 + 28, _sum7);
-                    vst1q_s32(outptr0 + 32, _sum8);
-                    vst1q_s32(outptr0 + 36, _sumc);
-                    vst1q_s32(outptr0 + 40, _sum9);
-                    vst1q_s32(outptr0 + 44, _sumd);
-                    vst1q_s32(outptr0 + 48, _suma);
-                    vst1q_s32(outptr0 + 52, _sume);
-                    vst1q_s32(outptr0 + 56, _sumb);
-                    vst1q_s32(outptr0 + 60, _sumf);
-                    outptr0 += 64;
-                }
-                if (out_elempack == 4)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + 8, _sum2);
-                    vst1q_s32(outptr0 + 12, _sum3);
-                    vst1q_s32(outptr0 + 16, _sum8);
-                    vst1q_s32(outptr0 + 20, _sum9);
-                    vst1q_s32(outptr0 + 24, _suma);
-                    vst1q_s32(outptr0 + 28, _sumb);
-                    vst1q_s32(outptr0 + out_hstep * 4, _sum4);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 4, _sum5);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 8, _sum6);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 12, _sum7);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 16, _sumc);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 20, _sumd);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 24, _sume);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 28, _sumf);
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-                    // to
-                    //      a0 a1 a2 a3
-                    //      a4 a5 a6 a7
-                    //      b0 b1 b2 b3
-                    //      b4 b5 b6 b7
-                    //      c0 c1 c2 c3
-                    //      c4 c5 c6 c7
-                    //      d0 d1 d2 d3
-                    //      d4 d5 d6 d7
-                    //      e0 e1 e2 e3
-                    //      e4 e5 e6 e7
-                    //      f0 f1 f2 f3
-                    //      f4 f5 f6 f7
-                    //      g0 g1 g2 g3
-                    //      g4 g5 g6 g7
-                    //      h0 h1 h2 h3
-                    //      h4 h5 h6 h7
-                    {
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sum1);
-                        int32x4x2_t _t1 = vzipq_s32(_sum2, _sum3);
-                        int32x4x2_t _t2 = vzipq_s32(_sum8, _sum9);
-                        int32x4x2_t _t3 = vzipq_s32(_suma, _sumb);
-                        int32x4x2_t _t4 = vzipq_s32(_sum4, _sum5);
-                        int32x4x2_t _t5 = vzipq_s32(_sum6, _sum7);
-                        int32x4x2_t _t6 = vzipq_s32(_sumc, _sumd);
-                        int32x4x2_t _t7 = vzipq_s32(_sume, _sumf);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t1.val[0]));
-                        _sum1 = vcombine_s32(vget_low_s32(_t2.val[0]), vget_low_s32(_t3.val[0]));
-                        _sum2 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t1.val[0]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t2.val[0]), vget_high_s32(_t3.val[0]));
-                        _sum4 = vcombine_s32(vget_low_s32(_t0.val[1]), vget_low_s32(_t1.val[1]));
-                        _sum5 = vcombine_s32(vget_low_s32(_t2.val[1]), vget_low_s32(_t3.val[1]));
-                        _sum6 = vcombine_s32(vget_high_s32(_t0.val[1]), vget_high_s32(_t1.val[1]));
-                        _sum7 = vcombine_s32(vget_high_s32(_t2.val[1]), vget_high_s32(_t3.val[1]));
-                        _sum8 = vcombine_s32(vget_low_s32(_t4.val[0]), vget_low_s32(_t5.val[0]));
-                        _sum9 = vcombine_s32(vget_low_s32(_t6.val[0]), vget_low_s32(_t7.val[0]));
-                        _suma = vcombine_s32(vget_high_s32(_t4.val[0]), vget_high_s32(_t5.val[0]));
-                        _sumb = vcombine_s32(vget_high_s32(_t6.val[0]), vget_high_s32(_t7.val[0]));
-                        _sumc = vcombine_s32(vget_low_s32(_t4.val[1]), vget_low_s32(_t5.val[1]));
-                        _sumd = vcombine_s32(vget_low_s32(_t6.val[1]), vget_low_s32(_t7.val[1]));
-                        _sume = vcombine_s32(vget_high_s32(_t4.val[1]), vget_high_s32(_t5.val[1]));
-                        _sumf = vcombine_s32(vget_high_s32(_t6.val[1]), vget_high_s32(_t7.val[1]));
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + out_hstep, _sum2);
-                    vst1q_s32(outptr0 + out_hstep + 4, _sum3);
-                    vst1q_s32(outptr0 + out_hstep * 2, _sum4);
-                    vst1q_s32(outptr0 + out_hstep * 2 + 4, _sum5);
-                    vst1q_s32(outptr0 + out_hstep * 3, _sum6);
-                    vst1q_s32(outptr0 + out_hstep * 3 + 4, _sum7);
-                    vst1q_s32(outptr0 + out_hstep * 4, _sum8);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 4, _sum9);
-                    vst1q_s32(outptr0 + out_hstep * 5, _suma);
-                    vst1q_s32(outptr0 + out_hstep * 5 + 4, _sumb);
-                    vst1q_s32(outptr0 + out_hstep * 6, _sumc);
-                    vst1q_s32(outptr0 + out_hstep * 6 + 4, _sumd);
-                    vst1q_s32(outptr0 + out_hstep * 7, _sume);
-                    vst1q_s32(outptr0 + out_hstep * 7 + 4, _sumf);
-                    outptr0 += 8;
-                }
-#else  // __ARM_FEATURE_DOTPROD
-
-                // from
-                //      a0 b1 c2 d3
-                //      e4 f5 g6 h7
-                //      e0 f1 g2 h3
-                //      a4 b5 c6 d7
-                //      c0 d1 a2 b3
-                //      g4 h5 e6 f7
-                //      g0 h1 e2 f3
-                //      c4 d5 a6 b7
-                //      a3 b2 c1 d0
-                //      e7 f6 g5 h4
-                //      e3 f2 g1 h0
-                //      a7 b6 c5 d4
-                //      c3 d2 a1 b0
-                //      g7 h6 e5 f4
-                //      g3 h2 e1 f0
-                //      c7 d6 a5 b4
-                if (out_elempack == 8)
-                {
-                    // to
-                    //      a0 b0 c0 d0
-                    //      e0 f0 g0 h0
-                    //      a1 b1 c1 d1
-                    //      e1 f1 g1 h1
-                    //      a2 b2 c2 d2
-                    //      e2 f2 g2 h2
-                    //      a3 b3 c3 d3
-                    //      e3 f3 g3 h3
-                    //      a4 b4 c4 d4
-                    //      e4 f4 g4 h4
-                    //      a5 b5 c5 d5
-                    //      e5 f5 g5 h5
-                    //      a6 b6 c6 d6
-                    //      e6 f6 g6 h6
-                    //      a7 b7 c7 d7
-                    //      e7 f7 g7 h7
-                    {
-                        _sum8 = vrev64q_s32(_sum8);
-                        _sum9 = vrev64q_s32(_sum9);
-                        _suma = vrev64q_s32(_suma);
-                        _sumb = vrev64q_s32(_sumb);
-                        _sumc = vrev64q_s32(_sumc);
-                        _sumd = vrev64q_s32(_sumd);
-                        _sume = vrev64q_s32(_sume);
-                        _sumf = vrev64q_s32(_sumf);
-                        _sum8 = vextq_s32(_sum8, _sum8, 2);
-                        _sum9 = vextq_s32(_sum9, _sum9, 2);
-                        _suma = vextq_s32(_suma, _suma, 2);
-                        _sumb = vextq_s32(_sumb, _sumb, 2);
-                        _sumc = vextq_s32(_sumc, _sumc, 2);
-                        _sumd = vextq_s32(_sumd, _sumd, 2);
-                        _sume = vextq_s32(_sume, _sume, 2);
-                        _sumf = vextq_s32(_sumf, _sumf, 2);
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sumc);
-                        int32x4x2_t _t1 = vzipq_s32(_sum4, _sum8);
-                        int32x4x2_t _t2 = vzipq_s32(_sum2, _sume);
-                        int32x4x2_t _t3 = vzipq_s32(_sum6, _suma);
-                        int32x4x2_t _t4 = vzipq_s32(_sum3, _sumf);
-                        int32x4x2_t _t5 = vzipq_s32(_sum7, _sumb);
-                        int32x4x2_t _t6 = vzipq_s32(_sum1, _sumd);
-                        int32x4x2_t _t7 = vzipq_s32(_sum5, _sum9);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t1.val[0]));
-                        _sum1 = vcombine_s32(vget_low_s32(_t2.val[0]), vget_low_s32(_t3.val[0]));
-                        _sum2 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t1.val[0]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t2.val[0]), vget_high_s32(_t3.val[0]));
-                        _sum4 = vcombine_s32(vget_low_s32(_t1.val[1]), vget_low_s32(_t0.val[1]));
-                        _sum5 = vcombine_s32(vget_low_s32(_t3.val[1]), vget_low_s32(_t2.val[1]));
-                        _sum6 = vcombine_s32(vget_high_s32(_t1.val[1]), vget_high_s32(_t0.val[1]));
-                        _sum7 = vcombine_s32(vget_high_s32(_t3.val[1]), vget_high_s32(_t2.val[1]));
-                        _sum8 = vcombine_s32(vget_low_s32(_t4.val[0]), vget_low_s32(_t5.val[0]));
-                        _sum9 = vcombine_s32(vget_low_s32(_t6.val[0]), vget_low_s32(_t7.val[0]));
-                        _suma = vcombine_s32(vget_high_s32(_t4.val[0]), vget_high_s32(_t5.val[0]));
-                        _sumb = vcombine_s32(vget_high_s32(_t6.val[0]), vget_high_s32(_t7.val[0]));
-                        _sumc = vcombine_s32(vget_low_s32(_t5.val[1]), vget_low_s32(_t4.val[1]));
-                        _sumd = vcombine_s32(vget_low_s32(_t7.val[1]), vget_low_s32(_t6.val[1]));
-                        _sume = vcombine_s32(vget_high_s32(_t5.val[1]), vget_high_s32(_t4.val[1]));
-                        _sumf = vcombine_s32(vget_high_s32(_t7.val[1]), vget_high_s32(_t6.val[1]));
-                        _sum2 = vrev64q_s32(_sum2);
-                        _sum3 = vrev64q_s32(_sum3);
-                        _sum6 = vrev64q_s32(_sum6);
-                        _sum7 = vrev64q_s32(_sum7);
-                        _suma = vrev64q_s32(_suma);
-                        _sumb = vrev64q_s32(_sumb);
-                        _sume = vrev64q_s32(_sume);
-                        _sumf = vrev64q_s32(_sumf);
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + 8, _sum2);
-                    vst1q_s32(outptr0 + 12, _sum3);
-                    vst1q_s32(outptr0 + 16, _sum4);
-                    vst1q_s32(outptr0 + 20, _sum5);
-                    vst1q_s32(outptr0 + 24, _sum6);
-                    vst1q_s32(outptr0 + 28, _sum7);
-                    vst1q_s32(outptr0 + 32, _sum8);
-                    vst1q_s32(outptr0 + 36, _sum9);
-                    vst1q_s32(outptr0 + 40, _suma);
-                    vst1q_s32(outptr0 + 44, _sumb);
-                    vst1q_s32(outptr0 + 48, _sumc);
-                    vst1q_s32(outptr0 + 52, _sumd);
-                    vst1q_s32(outptr0 + 56, _sume);
-                    vst1q_s32(outptr0 + 60, _sumf);
-                    outptr0 += 64;
-                }
-                if (out_elempack == 4)
-                {
-                    // to
-                    //      a0 b0 c0 d0
-                    //      a1 b1 c1 d1
-                    //      a2 b2 c2 d2
-                    //      a3 b3 c3 d3
-                    //      a4 b4 c4 d4
-                    //      a5 b5 c5 d5
-                    //      a6 b6 c6 d6
-                    //      a7 b7 c7 d7
-                    //      e0 f0 g0 h0
-                    //      e1 f1 g1 h1
-                    //      e2 f2 g2 h2
-                    //      e3 f3 g3 h3
-                    //      e4 f4 g4 h4
-                    //      e5 f5 g5 h5
-                    //      e6 f6 g6 h6
-                    //      e7 f7 g7 h7
-                    {
-                        _sum8 = vrev64q_s32(_sum8);
-                        _sum9 = vrev64q_s32(_sum9);
-                        _suma = vrev64q_s32(_suma);
-                        _sumb = vrev64q_s32(_sumb);
-                        _sumc = vrev64q_s32(_sumc);
-                        _sumd = vrev64q_s32(_sumd);
-                        _sume = vrev64q_s32(_sume);
-                        _sumf = vrev64q_s32(_sumf);
-                        _sum8 = vextq_s32(_sum8, _sum8, 2);
-                        _sum9 = vextq_s32(_sum9, _sum9, 2);
-                        _suma = vextq_s32(_suma, _suma, 2);
-                        _sumb = vextq_s32(_sumb, _sumb, 2);
-                        _sumc = vextq_s32(_sumc, _sumc, 2);
-                        _sumd = vextq_s32(_sumd, _sumd, 2);
-                        _sume = vextq_s32(_sume, _sume, 2);
-                        _sumf = vextq_s32(_sumf, _sumf, 2);
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sumc);
-                        int32x4x2_t _t1 = vzipq_s32(_sum4, _sum8);
-                        int32x4x2_t _t2 = vzipq_s32(_sum2, _sume);
-                        int32x4x2_t _t3 = vzipq_s32(_sum6, _suma);
-                        int32x4x2_t _t4 = vzipq_s32(_sum3, _sumf);
-                        int32x4x2_t _t5 = vzipq_s32(_sum7, _sumb);
-                        int32x4x2_t _t6 = vzipq_s32(_sum1, _sumd);
-                        int32x4x2_t _t7 = vzipq_s32(_sum5, _sum9);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t1.val[0]));
-                        _sum1 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t1.val[0]));
-                        _sum2 = vcombine_s32(vget_low_s32(_t1.val[1]), vget_low_s32(_t0.val[1]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t1.val[1]), vget_high_s32(_t0.val[1]));
-                        _sum4 = vcombine_s32(vget_low_s32(_t4.val[0]), vget_low_s32(_t5.val[0]));
-                        _sum5 = vcombine_s32(vget_high_s32(_t4.val[0]), vget_high_s32(_t5.val[0]));
-                        _sum6 = vcombine_s32(vget_low_s32(_t5.val[1]), vget_low_s32(_t4.val[1]));
-                        _sum7 = vcombine_s32(vget_high_s32(_t5.val[1]), vget_high_s32(_t4.val[1]));
-                        _sum8 = vcombine_s32(vget_low_s32(_t2.val[0]), vget_low_s32(_t3.val[0]));
-                        _sum9 = vcombine_s32(vget_high_s32(_t2.val[0]), vget_high_s32(_t3.val[0]));
-                        _suma = vcombine_s32(vget_low_s32(_t3.val[1]), vget_low_s32(_t2.val[1]));
-                        _sumb = vcombine_s32(vget_high_s32(_t3.val[1]), vget_high_s32(_t2.val[1]));
-                        _sumc = vcombine_s32(vget_low_s32(_t6.val[0]), vget_low_s32(_t7.val[0]));
-                        _sumd = vcombine_s32(vget_high_s32(_t6.val[0]), vget_high_s32(_t7.val[0]));
-                        _sume = vcombine_s32(vget_low_s32(_t7.val[1]), vget_low_s32(_t6.val[1]));
-                        _sumf = vcombine_s32(vget_high_s32(_t7.val[1]), vget_high_s32(_t6.val[1]));
-                        _sum1 = vrev64q_s32(_sum1);
-                        _sum3 = vrev64q_s32(_sum3);
-                        _sum5 = vrev64q_s32(_sum5);
-                        _sum7 = vrev64q_s32(_sum7);
-                        _sum9 = vrev64q_s32(_sum9);
-                        _sumb = vrev64q_s32(_sumb);
-                        _sumd = vrev64q_s32(_sumd);
-                        _sumf = vrev64q_s32(_sumf);
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + 8, _sum2);
-                    vst1q_s32(outptr0 + 12, _sum3);
-                    vst1q_s32(outptr0 + 16, _sum4);
-                    vst1q_s32(outptr0 + 20, _sum5);
-                    vst1q_s32(outptr0 + 24, _sum6);
-                    vst1q_s32(outptr0 + 28, _sum7);
-                    vst1q_s32(outptr0 + out_hstep * 4, _sum8);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 4, _sum9);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 8, _suma);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 12, _sumb);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 16, _sumc);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 20, _sumd);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 24, _sume);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 28, _sumf);
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-                    // to
-                    //      a0 a1 a2 a3
-                    //      a4 a5 a6 a7
-                    //      b0 b1 b2 b3
-                    //      b4 b5 b6 b7
-                    //      c0 c1 c2 c3
-                    //      c4 c5 c6 c7
-                    //      d0 d1 d2 d3
-                    //      d4 d5 d6 d7
-                    //      e0 e1 e2 e3
-                    //      e4 e5 e6 e7
-                    //      f0 f1 f2 f3
-                    //      f4 f5 f6 f7
-                    //      g0 g1 g2 g3
-                    //      g4 g5 g6 g7
-                    //      h0 h1 h2 h3
-                    //      h4 h5 h6 h7
-                    {
-                        _sum4 = vextq_s32(_sum4, _sum4, 2);
-                        _sum5 = vextq_s32(_sum5, _sum5, 2);
-                        _sum6 = vextq_s32(_sum6, _sum6, 2);
-                        _sum7 = vextq_s32(_sum7, _sum7, 2);
-                        _sumc = vextq_s32(_sumc, _sumc, 2);
-                        _sumd = vextq_s32(_sumd, _sumd, 2);
-                        _sume = vextq_s32(_sume, _sume, 2);
-                        _sumf = vextq_s32(_sumf, _sumf, 2);
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sumc);
-                        int32x4x2_t _t1 = vzipq_s32(_sum4, _sum8);
-                        int32x4x2_t _t2 = vzipq_s32(_sum3, _sumf);
-                        int32x4x2_t _t3 = vzipq_s32(_sum7, _sumb);
-                        int32x4x2_t _t4 = vzipq_s32(_sum2, _sume);
-                        int32x4x2_t _t5 = vzipq_s32(_sum6, _suma);
-                        int32x4x2_t _t6 = vzipq_s32(_sum1, _sumd);
-                        int32x4x2_t _t7 = vzipq_s32(_sum5, _sum9);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t1.val[0]));
-                        _sum1 = vcombine_s32(vget_low_s32(_t2.val[0]), vget_low_s32(_t3.val[0]));
-                        _sum2 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t1.val[0]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t2.val[0]), vget_high_s32(_t3.val[0]));
-                        _sum4 = vcombine_s32(vget_low_s32(_t1.val[1]), vget_low_s32(_t0.val[1]));
-                        _sum5 = vcombine_s32(vget_low_s32(_t3.val[1]), vget_low_s32(_t2.val[1]));
-                        _sum6 = vcombine_s32(vget_high_s32(_t1.val[1]), vget_high_s32(_t0.val[1]));
-                        _sum7 = vcombine_s32(vget_high_s32(_t3.val[1]), vget_high_s32(_t2.val[1]));
-                        _sum8 = vcombine_s32(vget_low_s32(_t4.val[0]), vget_low_s32(_t5.val[0]));
-                        _sum9 = vcombine_s32(vget_low_s32(_t6.val[0]), vget_low_s32(_t7.val[0]));
-                        _suma = vcombine_s32(vget_high_s32(_t4.val[0]), vget_high_s32(_t5.val[0]));
-                        _sumb = vcombine_s32(vget_high_s32(_t6.val[0]), vget_high_s32(_t7.val[0]));
-                        _sumc = vcombine_s32(vget_low_s32(_t5.val[1]), vget_low_s32(_t4.val[1]));
-                        _sumd = vcombine_s32(vget_low_s32(_t7.val[1]), vget_low_s32(_t6.val[1]));
-                        _sume = vcombine_s32(vget_high_s32(_t5.val[1]), vget_high_s32(_t4.val[1]));
-                        _sumf = vcombine_s32(vget_high_s32(_t7.val[1]), vget_high_s32(_t6.val[1]));
-                        _sum2 = vrev64q_s32(_sum2);
-                        _sum3 = vrev64q_s32(_sum3);
-                        _sum6 = vrev64q_s32(_sum6);
-                        _sum7 = vrev64q_s32(_sum7);
-                        _suma = vrev64q_s32(_suma);
-                        _sumb = vrev64q_s32(_sumb);
-                        _sume = vrev64q_s32(_sume);
-                        _sumf = vrev64q_s32(_sumf);
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + out_hstep, _sum2);
-                    vst1q_s32(outptr0 + out_hstep + 4, _sum3);
-                    vst1q_s32(outptr0 + out_hstep * 2, _sum4);
-                    vst1q_s32(outptr0 + out_hstep * 2 + 4, _sum5);
-                    vst1q_s32(outptr0 + out_hstep * 3, _sum6);
-                    vst1q_s32(outptr0 + out_hstep * 3 + 4, _sum7);
-                    vst1q_s32(outptr0 + out_hstep * 4, _sum8);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 4, _sum9);
-                    vst1q_s32(outptr0 + out_hstep * 5, _suma);
-                    vst1q_s32(outptr0 + out_hstep * 5 + 4, _sumb);
-                    vst1q_s32(outptr0 + out_hstep * 6, _sumc);
-                    vst1q_s32(outptr0 + out_hstep * 6 + 4, _sumd);
-                    vst1q_s32(outptr0 + out_hstep * 7, _sume);
-                    vst1q_s32(outptr0 + out_hstep * 7 + 4, _sumf);
-                    outptr0 += 8;
-                }
-#endif // __ARM_FEATURE_DOTPROD
-            }
-            else
-            {
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-                vst1q_s32(outptr + 16, _sum4);
-                vst1q_s32(outptr + 20, _sum5);
-                vst1q_s32(outptr + 24, _sum6);
-                vst1q_s32(outptr + 28, _sum7);
-                vst1q_s32(outptr + 32, _sum8);
-                vst1q_s32(outptr + 36, _sum9);
-                vst1q_s32(outptr + 40, _suma);
-                vst1q_s32(outptr + 44, _sumb);
-                vst1q_s32(outptr + 48, _sumc);
-                vst1q_s32(outptr + 52, _sumd);
-                vst1q_s32(outptr + 56, _sume);
-                vst1q_s32(outptr + 60, _sumf);
-            }
-
-            outptr += 64;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-#endif // __aarch64__
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            const signed char* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "cmp    %w9, #0                     \n"
-                "beq    0f                          \n"
-
-                "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0] \n"
-                "sub    %0, %0, #64                 \n"
-                "b      1f                          \n"
-
-                "0:                                 \n"
-                "eor    v16.16b, v16.16b, v16.16b   \n"
-                "eor    v17.16b, v17.16b, v17.16b   \n"
-                "eor    v18.16b, v18.16b, v18.16b   \n"
-                "eor    v19.16b, v19.16b, v19.16b   \n"
-                "eor    v20.16b, v20.16b, v20.16b   \n"
-                "eor    v21.16b, v21.16b, v21.16b   \n"
-                "eor    v22.16b, v22.16b, v22.16b   \n"
-                "eor    v23.16b, v23.16b, v23.16b   \n"
-
-                "1:                                 \n"
-#if __ARM_FEATURE_DOTPROD
-                "lsr    w4, %w8, #3                 \n" // w4 = max_kk >> 3
-                "cmp    w4, #0                      \n"
-                "beq    101f                        \n"
-
-#if __ARM_FEATURE_MATMUL_INT8
-                "eor    v24.16b, v24.16b, v24.16b   \n"
-                "eor    v25.16b, v25.16b, v25.16b   \n"
-                "eor    v26.16b, v26.16b, v26.16b   \n"
-                "eor    v27.16b, v27.16b, v27.16b   \n"
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-                "eor    v29.16b, v29.16b, v29.16b   \n"
-                "eor    v30.16b, v30.16b, v30.16b   \n"
-                "eor    v31.16b, v31.16b, v31.16b   \n"
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                "2:                                 \n"
-                "ld1    {v0.16b, v1.16b, v2.16b, v3.16b}, [%1], #64 \n"
-                "ld1    {v4.16b, v5.16b}, [%2], #32 \n"
-
-#if __ARM_FEATURE_MATMUL_INT8
-                "smmla  v24.4s, v0.16b, v4.16b      \n"
-                "smmla  v25.4s, v1.16b, v4.16b      \n"
-                "smmla  v26.4s, v0.16b, v5.16b      \n"
-                "smmla  v27.4s, v1.16b, v5.16b      \n"
-                "subs   w4, w4, #1                  \n"
-                "smmla  v28.4s, v2.16b, v4.16b      \n"
-                "smmla  v29.4s, v3.16b, v4.16b      \n"
-                "smmla  v30.4s, v2.16b, v5.16b      \n"
-                "smmla  v31.4s, v3.16b, v5.16b      \n"
-#else  // __ARM_FEATURE_MATMUL_INT8
-                "sdot   v16.4s, v0.16b, v4.4b[0]    \n"
-                "sdot   v17.4s, v0.16b, v4.4b[1]    \n"
-                "sdot   v18.4s, v0.16b, v4.4b[2]    \n"
-                "sdot   v19.4s, v0.16b, v4.4b[3]    \n"
-                "sdot   v20.4s, v1.16b, v4.4b[0]    \n"
-                "sdot   v21.4s, v1.16b, v4.4b[1]    \n"
-                "sdot   v22.4s, v1.16b, v4.4b[2]    \n"
-                "sdot   v23.4s, v1.16b, v4.4b[3]    \n"
-                "subs   w4, w4, #1                  \n"
-                "sdot   v16.4s, v2.16b, v5.4b[0]    \n"
-                "sdot   v17.4s, v2.16b, v5.4b[1]    \n"
-                "sdot   v18.4s, v2.16b, v5.4b[2]    \n"
-                "sdot   v19.4s, v2.16b, v5.4b[3]    \n"
-                "sdot   v20.4s, v3.16b, v5.4b[0]    \n"
-                "sdot   v21.4s, v3.16b, v5.4b[1]    \n"
-                "sdot   v22.4s, v3.16b, v5.4b[2]    \n"
-                "sdot   v23.4s, v3.16b, v5.4b[3]    \n"
-#endif // __ARM_FEATURE_MATMUL_INT8
-                "bne    2b                          \n"
-
-#if __ARM_FEATURE_MATMUL_INT8
-                "uzp1   v0.4s, v24.4s, v25.4s       \n"
-                "uzp2   v1.4s, v24.4s, v25.4s       \n"
-                "uzp1   v2.4s, v26.4s, v27.4s       \n"
-                "uzp2   v3.4s, v26.4s, v27.4s       \n"
-                "uzp1   v4.4s, v28.4s, v29.4s       \n"
-                "uzp2   v5.4s, v28.4s, v29.4s       \n"
-                "uzp1   v6.4s, v30.4s, v31.4s       \n"
-                "uzp2   v7.4s, v30.4s, v31.4s       \n"
-
-                "add    v16.4s, v16.4s, v0.4s       \n"
-                "add    v17.4s, v17.4s, v1.4s       \n"
-                "add    v18.4s, v18.4s, v2.4s       \n"
-                "add    v19.4s, v19.4s, v3.4s       \n"
-                "add    v20.4s, v20.4s, v4.4s       \n"
-                "add    v21.4s, v21.4s, v5.4s       \n"
-                "add    v22.4s, v22.4s, v6.4s       \n"
-                "add    v23.4s, v23.4s, v7.4s       \n"
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                "101:                               \n"
-                "and    w4, %w8, #4                 \n" // w4 = remain = max_kk & 4
-                "cmp    w4, #0                      \n"
-                "beq    3f                          \n"
-
-                // kk += 4 part
-                "ld1    {v0.16b, v1.16b}, [%1], #32 \n"
-                "ld1    {v2.16b}, [%2], #16         \n"
-                "sdot   v16.4s, v0.16b, v2.4b[0]    \n"
-                "sdot   v17.4s, v0.16b, v2.4b[1]    \n"
-                "sdot   v18.4s, v0.16b, v2.4b[2]    \n"
-                "sdot   v19.4s, v0.16b, v2.4b[3]    \n"
-                "sdot   v20.4s, v1.16b, v2.4b[0]    \n"
-                "sdot   v21.4s, v1.16b, v2.4b[1]    \n"
-                "sdot   v22.4s, v1.16b, v2.4b[2]    \n"
-                "sdot   v23.4s, v1.16b, v2.4b[3]    \n"
-#else  // __ARM_FEATURE_DOTPROD
-                "lsr    w4, %w8, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    3f                          \n"
-
-                "2:                                 \n"
-                "ld1    {v0.16b, v1.16b}, [%1], #32 \n"
-                "ld1    {v4.16b}, [%2], #16         \n"
-                "smull  v8.8h, v0.8b, v4.8b         \n"
-                "rev64  v2.4s, v0.4s                \n"
-                "smull  v10.8h, v2.8b, v4.8b        \n"
-                "ext    v5.16b, v4.16b, v4.16b, #8  \n"
-                "smull2 v9.8h, v0.16b, v5.16b       \n"
-                "rev64  v6.8h, v4.8h                \n"
-                "smull2 v11.8h, v2.16b, v5.16b      \n"
-                "ext    v7.16b, v6.16b, v6.16b, #8  \n"
-                "smull  v12.8h, v0.8b, v6.8b        \n"
-                "smull  v14.8h, v2.8b, v6.8b        \n"
-                "rev64  v3.4s, v1.4s                \n"
-                "smull2 v13.8h, v0.16b, v7.16b      \n"
-                "smull2 v15.8h, v2.16b, v7.16b      \n"
-                "smlal  v8.8h, v1.8b, v5.8b         \n"
-                "smlal  v10.8h, v3.8b, v5.8b        \n"
-                "smlal2 v9.8h, v1.16b, v4.16b       \n"
-                "smlal2 v11.8h, v3.16b, v4.16b      \n"
-                "smlal  v12.8h, v1.8b, v7.8b        \n"
-                "smlal  v14.8h, v3.8b, v7.8b        \n"
-                "smlal2 v13.8h, v1.16b, v6.16b      \n"
-                "smlal2 v15.8h, v3.16b, v6.16b      \n"
-                "subs   w4, w4, #1                  \n"
-                "sadalp v16.4s, v8.8h               \n"
-                "sadalp v18.4s, v10.8h              \n"
-                "sadalp v17.4s, v9.8h               \n"
-                "sadalp v19.4s, v11.8h              \n"
-                "sadalp v20.4s, v12.8h              \n"
-                "sadalp v22.4s, v14.8h              \n"
-                "sadalp v21.4s, v13.8h              \n"
-                "sadalp v23.4s, v15.8h              \n"
-                "bne    2b                          \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "3:                                 \n"
-                "and    w4, %w8, #2                 \n" // w4 = remain = max_kk & 2
-                "cmp    w4, #0                      \n"
-                "beq    4f                          \n"
-
-                // kk += 2 part
-#if __ARM_FEATURE_DOTPROD
-                "ld1    {v0.16b}, [%1], #16         \n"
-                "ld1    {v1.8b}, [%2], #8           \n"
-                "dup    v4.8h, v1.h[0]              \n"
-                "dup    v5.8h, v1.h[1]              \n"
-                "dup    v6.8h, v1.h[2]              \n"
-                "dup    v7.8h, v1.h[3]              \n"
-                "smull  v8.8h, v0.8b, v4.8b         \n"
-                "smull  v9.8h, v0.8b, v5.8b         \n"
-                "smull  v10.8h, v0.8b, v6.8b        \n"
-                "smull  v11.8h, v0.8b, v7.8b        \n"
-                "smull2 v12.8h, v0.16b, v4.16b      \n"
-                "smull2 v13.8h, v0.16b, v5.16b      \n"
-                "smull2 v14.8h, v0.16b, v6.16b      \n"
-                "smull2 v15.8h, v0.16b, v7.16b      \n"
-                "sadalp v16.4s, v8.8h               \n"
-                "sadalp v17.4s, v9.8h               \n"
-                "sadalp v18.4s, v10.8h              \n"
-                "sadalp v19.4s, v11.8h              \n"
-                "sadalp v20.4s, v12.8h              \n"
-                "sadalp v21.4s, v13.8h              \n"
-                "sadalp v22.4s, v14.8h              \n"
-                "sadalp v23.4s, v15.8h              \n"
-#else  // __ARM_FEATURE_DOTPROD
-                "ld1    {v0.16b}, [%1], #16         \n"
-                "ld1r   {v2.2d}, [%2]               \n"
-                "add    %2, %2, #8                  \n"
-                "rev64  v1.4s, v0.4s                \n"
-                "rev64  v3.8h, v2.8h                \n"
-                "smull  v8.8h, v0.8b, v2.8b         \n"
-                "smull2 v9.8h, v0.16b, v2.16b       \n"
-                "smull  v10.8h, v1.8b, v2.8b        \n"
-                "smull2 v11.8h, v1.16b, v2.16b      \n"
-                "smull  v12.8h, v0.8b, v3.8b        \n"
-                "smull2 v13.8h, v0.16b, v3.16b      \n"
-                "smull  v14.8h, v1.8b, v3.8b        \n"
-                "smull2 v15.8h, v1.16b, v3.16b      \n"
-                "sadalp v16.4s, v8.8h               \n"
-                "sadalp v17.4s, v9.8h               \n"
-                "sadalp v18.4s, v10.8h              \n"
-                "sadalp v19.4s, v11.8h              \n"
-                "sadalp v20.4s, v12.8h              \n"
-                "sadalp v21.4s, v13.8h              \n"
-                "sadalp v22.4s, v14.8h              \n"
-                "sadalp v23.4s, v15.8h              \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "4:                                 \n"
-                "and    w4, %w8, #1                 \n" // w4 = remain = max_kk & 1
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                // kk += 1 part
-#if __ARM_FEATURE_DOTPROD
-                "ld1    {v0.8b}, [%1], #8           \n"
-                "ld1    {v1.8b}, [%2]               \n"
-                "add    %2, %2, #4                  \n"
-                "dup    v8.8b, v1.b[0]              \n"
-                "dup    v9.8b, v1.b[1]              \n"
-                "dup    v10.8b, v1.b[2]             \n"
-                "dup    v11.8b, v1.b[3]             \n"
-                "smull  v8.8h, v0.8b, v8.8b         \n"
-                "smull  v9.8h, v0.8b, v9.8b         \n"
-                "smull  v10.8h, v0.8b, v10.8b       \n"
-                "smull  v11.8h, v0.8b, v11.8b       \n"
-                "saddw  v16.4s, v16.4s, v8.4h       \n"
-                "saddw  v17.4s, v17.4s, v9.4h       \n"
-                "saddw  v18.4s, v18.4s, v10.4h      \n"
-                "saddw  v19.4s, v19.4s, v11.4h      \n"
-                "saddw2 v20.4s, v20.4s, v8.8h       \n"
-                "saddw2 v21.4s, v21.4s, v9.8h       \n"
-                "saddw2 v22.4s, v22.4s, v10.8h      \n"
-                "saddw2 v23.4s, v23.4s, v11.8h      \n"
-#else  // __ARM_FEATURE_DOTPROD
-                "ld1    {v0.8b}, [%1], #8           \n"
-                "ld1r   {v4.2s}, [%2]               \n"
-                "add    %2, %2, #4                  \n"
-                "rev32  v1.4h, v0.4h                \n"
-                "rev64  v5.8b, v4.8b                \n"
-                "smull  v8.8h, v0.8b, v4.8b         \n"
-                "smull  v9.8h, v1.8b, v4.8b         \n"
-                "smull  v10.8h, v0.8b, v5.8b        \n"
-                "smull  v11.8h, v1.8b, v5.8b        \n"
-                "saddw  v16.4s, v16.4s, v8.4h       \n"
-                "saddw2 v17.4s, v17.4s, v8.8h       \n"
-                "saddw  v18.4s, v18.4s, v9.4h       \n"
-                "saddw2 v19.4s, v19.4s, v9.8h       \n"
-                "saddw  v20.4s, v20.4s, v10.4h      \n"
-                "saddw2 v21.4s, v21.4s, v10.8h      \n"
-                "saddw  v22.4s, v22.4s, v11.4h      \n"
-                "saddw2 v23.4s, v23.4s, v11.8h      \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "5:                                 \n"
-                "cmp    %w10, #0                    \n"
-                "beq    10f                         \n"
-
-#if __ARM_FEATURE_DOTPROD
-                // from
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                //      a2 b2 c2 d2
-                //      a3 b3 c3 d3
-                //      e0 f0 g0 h0
-                //      e1 f1 g1 h1
-                //      e2 f2 g2 h2
-                //      e3 f3 g3 h3
-                // if out_elempack == 4 / 8
-                "cmp    %w11, #1                    \n"
-                "beq    8f                          \n"
-
-                // if out_elempack == 8
-                "cmp    %w11, #8                    \n"
-                "bne    7f                          \n"
-
-                "st1    {v16.4s}, [%3], #16         \n"
-                "st1    {v20.4s}, [%3], #16         \n"
-                "st1    {v17.4s}, [%3], #16         \n"
-                "st1    {v21.4s}, [%3], #16         \n"
-                "st1    {v18.4s}, [%3], #16         \n"
-                "st1    {v22.4s}, [%3], #16         \n"
-                "st1    {v19.4s}, [%3], #16         \n"
-                "st1    {v23.4s}, [%3], #16         \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 4
-                "7:                                 \n"
-                "add    x4, %3, %12, lsl #4         \n"
-                "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [x4] \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // to
-                //      a0 a1 a2 a3
-                //      b0 b1 b2 b3
-                //      c0 c1 c2 c3
-                //      d0 d1 d2 d3
-                //      e0 e1 e2 e3
-                //      f0 f1 f2 f3
-                //      g0 g1 g2 g3
-                //      h0 h1 h2 h3
-                "zip1   v0.4s, v16.4s, v17.4s       \n"
-                "zip2   v1.4s, v16.4s, v17.4s       \n"
-                "zip1   v2.4s, v18.4s, v19.4s       \n"
-                "zip2   v3.4s, v18.4s, v19.4s       \n"
-                "zip1   v4.4s, v20.4s, v21.4s       \n"
-                "zip2   v5.4s, v20.4s, v21.4s       \n"
-                "zip1   v6.4s, v22.4s, v23.4s       \n"
-                "zip2   v7.4s, v22.4s, v23.4s       \n"
-                "zip1   v16.2d, v0.2d, v2.2d        \n"
-                "zip2   v17.2d, v0.2d, v2.2d        \n"
-                "zip1   v18.2d, v1.2d, v3.2d        \n"
-                "zip2   v19.2d, v1.2d, v3.2d        \n"
-                "zip1   v20.2d, v4.2d, v6.2d        \n"
-                "zip2   v21.2d, v4.2d, v6.2d        \n"
-                "zip1   v22.2d, v5.2d, v7.2d        \n"
-                "zip2   v23.2d, v5.2d, v7.2d        \n"
-
-                "add    x4, %3, %12, lsl #2         \n"
-                "st1    {v16.4s}, [%3], #16         \n"
-                "st1    {v17.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v18.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v19.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v20.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v21.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v22.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v23.4s}, [x4]              \n"
-#else  // __ARM_FEATURE_DOTPROD
-
-                // from
-                //      a0 b1 c2 d3
-                //      e0 f1 g2 h3
-                //      c0 d1 a2 b3
-                //      g0 h1 e2 f3
-                //      a3 b2 c1 d0
-                //      e3 f2 g1 h0
-                //      c3 d2 a1 b0
-                //      g3 h2 e1 f0
-                // if out_elempack == 4 / 8
-                "cmp    %w11, #1                    \n"
-                "beq    8f                          \n"
-
-                "rev64  v20.4s, v20.4s              \n"
-                "rev64  v21.4s, v21.4s              \n"
-                "rev64  v22.4s, v22.4s              \n"
-                "rev64  v23.4s, v23.4s              \n"
-                "ext    v20.16b, v20.16b, v20.16b, #8 \n"
-                "ext    v21.16b, v21.16b, v21.16b, #8 \n"
-                "ext    v22.16b, v22.16b, v22.16b, #8 \n"
-                "ext    v23.16b, v23.16b, v23.16b, #8 \n"
-                "zip1   v0.4s, v16.4s, v22.4s       \n"
-                "zip2   v1.4s, v16.4s, v22.4s       \n"
-                "zip1   v2.4s, v18.4s, v20.4s       \n"
-                "zip2   v3.4s, v18.4s, v20.4s       \n"
-                "zip1   v4.4s, v17.4s, v23.4s       \n"
-                "zip2   v5.4s, v17.4s, v23.4s       \n"
-                "zip1   v6.4s, v19.4s, v21.4s       \n"
-                "zip2   v7.4s, v19.4s, v21.4s       \n"
-
-                // if out_elempack == 8
-                "cmp    %w11, #8                    \n"
-                "bne    7f                          \n"
-
-                // to
-                //      a0 b0 c0 d0
-                //      e0 f0 g0 h0
-                //      a1 b1 c1 d1
-                //      e1 f1 g1 h1
-                //      a2 b2 c2 d2
-                //      e2 f2 g2 h2
-                //      a3 b3 c3 d3
-                //      e3 f3 g3 h3
-                "zip1   v16.2d, v0.2d, v2.2d        \n"
-                "zip1   v17.2d, v4.2d, v6.2d        \n"
-                "zip2   v18.2d, v0.2d, v2.2d        \n"
-                "zip2   v19.2d, v4.2d, v6.2d        \n"
-                "zip1   v20.2d, v3.2d, v1.2d        \n"
-                "zip1   v21.2d, v7.2d, v5.2d        \n"
-                "zip2   v22.2d, v3.2d, v1.2d        \n"
-                "zip2   v23.2d, v7.2d, v5.2d        \n"
-                "rev64  v18.4s, v18.4s              \n"
-                "rev64  v19.4s, v19.4s              \n"
-                "rev64  v22.4s, v22.4s              \n"
-                "rev64  v23.4s, v23.4s              \n"
-
-                "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%3], #64 \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 4
-                "7:                                 \n"
-
-                // to
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                //      a2 b2 c2 d2
-                //      a3 b3 c3 d3
-                //      e0 f0 g0 h0
-                //      e1 f1 g1 h1
-                //      e2 f2 g2 h2
-                //      e3 f3 g3 h3
-                "zip1   v16.2d, v0.2d, v2.2d        \n"
-                "zip1   v24.2d, v4.2d, v6.2d        \n"
-                "zip2   v17.2d, v0.2d, v2.2d        \n"
-                "zip2   v25.2d, v4.2d, v6.2d        \n"
-                "zip1   v18.2d, v3.2d, v1.2d        \n"
-                "zip1   v26.2d, v7.2d, v5.2d        \n"
-                "zip2   v19.2d, v3.2d, v1.2d        \n"
-                "zip2   v27.2d, v7.2d, v5.2d        \n"
-                "rev64  v17.4s, v17.4s              \n"
-                "rev64  v25.4s, v25.4s              \n"
-                "rev64  v19.4s, v19.4s              \n"
-                "rev64  v27.4s, v27.4s              \n"
-
-                "add    x4, %3, %12, lsl #4         \n"
-                "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                "st1    {v24.4s, v25.4s, v26.4s, v27.4s}, [x4] \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-
-                // to
-                //      a0 a1 a2 a3
-                //      b0 b1 b2 b3
-                //      c0 c1 c2 c3
-                //      d0 d1 d2 d3
-                //      e0 e1 e2 e3
-                //      f0 f1 f2 f3
-                //      g0 g1 g2 g3
-                //      h0 h1 h2 h3
-                "ext    v18.16b, v18.16b, v18.16b, #8 \n"
-                "ext    v19.16b, v19.16b, v19.16b, #8 \n"
-                "ext    v22.16b, v22.16b, v22.16b, #8 \n"
-                "ext    v23.16b, v23.16b, v23.16b, #8 \n"
-                "zip1   v0.4s, v16.4s, v22.4s       \n"
-                "zip2   v1.4s, v16.4s, v22.4s       \n"
-                "zip1   v2.4s, v18.4s, v20.4s       \n"
-                "zip2   v3.4s, v18.4s, v20.4s       \n"
-                "zip1   v4.4s, v17.4s, v23.4s       \n"
-                "zip2   v5.4s, v17.4s, v23.4s       \n"
-                "zip1   v6.4s, v19.4s, v21.4s       \n"
-                "zip2   v7.4s, v19.4s, v21.4s       \n"
-                "zip1   v16.2d, v0.2d, v2.2d        \n"
-                "zip2   v17.2d, v0.2d, v2.2d        \n"
-                "zip1   v18.2d, v3.2d, v1.2d        \n"
-                "zip2   v19.2d, v3.2d, v1.2d        \n"
-                "zip1   v20.2d, v4.2d, v6.2d        \n"
-                "zip2   v21.2d, v4.2d, v6.2d        \n"
-                "zip1   v22.2d, v7.2d, v5.2d        \n"
-                "zip2   v23.2d, v7.2d, v5.2d        \n"
-                "rev64  v17.4s, v17.4s              \n"
-                "rev64  v19.4s, v19.4s              \n"
-                "rev64  v21.4s, v21.4s              \n"
-                "rev64  v23.4s, v23.4s              \n"
-
-                "add    x4, %3, %12, lsl #2         \n"
-                "st1    {v16.4s}, [%3], #16         \n"
-                "st1    {v17.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v18.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v19.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v20.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v21.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v22.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v23.4s}, [x4]              \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "9:                                 \n"
-                "add    %0, %0, #128                \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(max_kk),       // %8
-                "r"(k),            // %9
-                "r"(k_end),        // %10
-                "r"(out_elempack), // %11
-                "r"(out_hstep)     // %12
-                : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-            asm volatile(
-                "cmp        %9, #0              \n"
-                "beq        0f                  \n"
-
-                "vldm       %0!, {d16-d23}      \n"
-                "vldm       %0, {d24-d31}       \n"
-                "sub        %0, %0, #64         \n"
-                "b          1f                  \n"
-
-                "0:                             \n"
-                "veor       q8, q8              \n"
-                "veor       q9, q9              \n"
-                "veor       q10, q10            \n"
-                "veor       q11, q11            \n"
-                "veor       q12, q12            \n"
-                "veor       q13, q13            \n"
-                "veor       q14, q14            \n"
-                "veor       q15, q15            \n"
-
-                "1:                             \n"
-                "lsr        r4, %8, #2          \n" // r4 = max_kk >> 2
-                "cmp        r4, #0              \n"
-                "beq        3f                  \n"
-
-                ".align 4                       \n"
-                "2:                             \n"
-                "pld        [%1, #256]          \n"
-                "vld1.s8    {d0-d3}, [%1 :64]!  \n"
-                "pld        [%2, #128]          \n"
-                "vld1.s8    {d4-d5}, [%2]!      \n"
-                "vmull.s8   q4, d0, d4          \n"
-                "vrev64.32  q3, q0              \n"
-                "vmull.s8   q5, d1, d4          \n"
-                "vmull.s8   q6, d6, d4          \n"
-                "vmull.s8   q7, d7, d4          \n"
-                "vrev64.32  q0, q1              \n"
-                "vmlal.s8   q4, d2, d5          \n"
-                "vmlal.s8   q5, d3, d5          \n"
-                "vmlal.s8   q6, d0, d5          \n"
-                "vmlal.s8   q7, d1, d5          \n"
-                "vrev64.16  q2, q2              \n"
-                "vpadal.s16 q8, q4              \n"
-                "vrev64.32  q1, q3              \n"
-                "vpadal.s16 q9, q5              \n"
-                "vmull.s8   q4, d6, d4          \n"
-                "vpadal.s16 q10, q6             \n"
-                "vmull.s8   q5, d7, d4          \n"
-                "vpadal.s16 q11, q7             \n"
-                "vmull.s8   q6, d2, d4          \n"
-                "vmull.s8   q7, d3, d4          \n"
-                "vrev64.32  q3, q0              \n"
-                "vmlal.s8   q4, d0, d5          \n"
-                "vmlal.s8   q5, d1, d5          \n"
-                "vmlal.s8   q6, d6, d5          \n"
-                "vmlal.s8   q7, d7, d5          \n"
-                "subs       r4, r4, #1          \n"
-                "vpadal.s16 q14, q4             \n"
-                "vpadal.s16 q15, q5             \n"
-                "vpadal.s16 q12, q6             \n"
-                "vpadal.s16 q13, q7             \n"
-                "bne        2b                  \n"
-
-                "3:                             \n"
-                "and        r4, %8, #2          \n" // r4 = remain = max_kk & 2
-                "cmp        r4, #0              \n"
-                "beq        4f                  \n"
-
-                // kk += 2 part
-                "vld1.s8    {d0-d1}, [%1 :64]!  \n"
-                "vld1.s8    {d4}, [%2]!         \n"
-                "vrev64.32  q1, q0              \n"
-                "vrev64.16  d5, d4              \n"
-                "vmull.s8   q4, d0, d4          \n"
-                "vmull.s8   q5, d1, d4          \n"
-                "vmull.s8   q6, d2, d4          \n"
-                "vmull.s8   q7, d3, d4          \n"
-                "vpadal.s16 q8, q4              \n"
-                "vpadal.s16 q9, q5              \n"
-                "vpadal.s16 q10, q6             \n"
-                "vpadal.s16 q11, q7             \n"
-                "vmull.s8   q4, d0, d5          \n"
-                "vmull.s8   q5, d1, d5          \n"
-                "vmull.s8   q6, d2, d5          \n"
-                "vmull.s8   q7, d3, d5          \n"
-                "vpadal.s16 q12, q4             \n"
-                "vpadal.s16 q13, q5             \n"
-                "vpadal.s16 q14, q6             \n"
-                "vpadal.s16 q15, q7             \n"
-
-                "4:                             \n"
-                "and        r4, %8, #1          \n" // r4 = remain = max_kk & 1
-                "cmp        r4, #0              \n"
-                "beq        5f                  \n"
-
-                // kk += 1 part
-                "vld1.s8    {d0}, [%1 :64]!     \n"
-                "vld1.s32   {d2[]}, [%2]!       \n"
-                "vrev64.16  d1, d0              \n"
-                "vrev64.8   d3, d2              \n"
-                "vext.s8    d1, d1, #4          \n"
-                "vmull.s8   q4, d0, d2          \n"
-                "vmull.s8   q5, d1, d2          \n"
-                "vmull.s8   q6, d0, d3          \n"
-                "vmull.s8   q7, d1, d3          \n"
-                "vaddw.s16  q8, d8              \n"
-                "vaddw.s16  q9, d9              \n"
-                "vaddw.s16  q10, d10            \n"
-                "vaddw.s16  q11, d11            \n"
-                "vaddw.s16  q12, d12            \n"
-                "vaddw.s16  q13, d13            \n"
-                "vaddw.s16  q14, d14            \n"
-                "vaddw.s16  q15, d15            \n"
-
-                "5:                             \n"
-                "cmp        %10, #0             \n"
-                "beq        10f                 \n"
-
-                // from
-                //      a0 b1 c2 d3
-                //      e0 f1 g2 h3
-                //      c0 d1 a2 b3
-                //      g0 h1 e2 f3
-                //      a3 b2 c1 d0
-                //      e3 f2 g1 h0
-                //      c3 d2 a1 b0
-                //      g3 h2 e1 f0
-                // if out_elempack == 4 / 8
-                "cmp        %11, #1             \n"
-                "beq        8f                  \n"
-
-                "vrev64.32  q12, q12            \n"
-                "vrev64.32  q13, q13            \n"
-                "vrev64.32  q14, q14            \n"
-                "vrev64.32  q15, q15            \n"
-                "vext.32    q12, q12, #2        \n"
-                "vext.32    q13, q13, #2        \n"
-                "vext.32    q14, q14, #2        \n"
-                "vext.32    q15, q15, #2        \n"
-                "vzip.32    q8, q14             \n"
-                "vzip.32    q10, q12            \n"
-                "vzip.32    q9, q15             \n"
-                "vzip.32    q11, q13            \n"
-                "vswp       d17, d20            \n"
-                "vswp       d19, d22            \n"
-                "vswp       d28, d25            \n"
-                "vswp       d30, d27            \n"
-                "vrev64.32  q10, q10            \n"
-                "vrev64.32  q11, q11            \n"
-                "vrev64.32  q14, q14            \n"
-                "vrev64.32  q15, q15            \n"
-
-                // if out_elempack == 8
-                "cmp        %11, #8             \n"
-                "bne        7f                  \n"
-
-                // to
-                //      a0 b0 c0 d0
-                //      e0 f0 g0 h0
-                //      a1 b1 c1 d1
-                //      e1 f1 g1 h1
-                //      a2 b2 c2 d2
-                //      e2 f2 g2 h2
-                //      a3 b3 c3 d3
-                //      e3 f3 g3 h3
-                "vstm       %3!, {d16-d23}      \n"
-                "vstm       %3!, {d24-d31}      \n"
-                "b          9f                  \n"
-
-                // if out_elempack == 4
-                "7:                             \n"
-                // to
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                //      a2 b2 c2 d2
-                //      a3 b3 c3 d3
-                //      e0 f0 g0 h0
-                //      e1 f1 g1 h1
-                //      e2 f2 g2 h2
-                //      e3 f3 g3 h3
-                "vswp       q9, q10             \n"
-                "vswp       q13, q14            \n"
-                "vswp       q10, q12            \n"
-                "vswp       q11, q13            \n"
-
-                "add        r4, %3, %12, lsl #4 \n"
-                "vstm       %3!, {d16-d23}      \n"
-                "vstm       r4, {d24-d31}       \n"
-                "b          9f                  \n"
-
-                // if out_elempack == 1
-                "8:                             \n"
-                // to
-                //      a0 a1 a2 a3
-                //      b0 b1 b2 b3
-                //      c0 c1 c2 c3
-                //      d0 d1 d2 d3
-                //      e0 e1 e2 e3
-                //      f0 f1 f2 f3
-                //      g0 g1 g2 g3
-                //      h0 h1 h2 h3
-                "vext.32    q10, q10, #2        \n"
-                "vext.32    q11, q11, #2        \n"
-                "vext.32    q14, q14, #2        \n"
-                "vext.32    q15, q15, #2        \n"
-                "vzip.32    q8, q14             \n"
-                "vzip.32    q10, q12            \n"
-                "vzip.32    q9, q15             \n"
-                "vzip.32    q11, q13            \n"
-                "vswp       d17, d20            \n"
-                "vswp       d19, d22            \n"
-                "vswp       d28, d25            \n"
-                "vswp       d30, d27            \n"
-                "vrev64.32  q10, q10            \n"
-                "vrev64.32  q11, q11            \n"
-                "vrev64.32  q14, q14            \n"
-                "vrev64.32  q15, q15            \n"
-
-                "add        r4, %3, %12, lsl #2 \n"
-                "vst1.s32   {d16-d17}, [%3]!    \n"
-                "vst1.s32   {d20-d21}, [r4]     \n"
-                "add        r4, r4, %12, lsl #2 \n"
-                "vst1.s32   {d24-d25}, [r4]     \n"
-                "add        r4, r4, %12, lsl #2 \n"
-                "vst1.s32   {d28-d29}, [r4]     \n"
-                "add        r4, r4, %12, lsl #2 \n"
-                "vst1.s32   {d18-d19}, [r4]     \n"
-                "add        r4, r4, %12, lsl #2 \n"
-                "vst1.s32   {d22-d23}, [r4]     \n"
-                "add        r4, r4, %12, lsl #2 \n"
-                "vst1.s32   {d26-d27}, [r4]     \n"
-                "add        r4, r4, %12, lsl #2 \n"
-                "vst1.s32   {d30-d31}, [r4]     \n"
-
-                "9:                             \n"
-                "add        %0, %0, #128        \n"
-                "b          11f                 \n"
-
-                "10:                            \n"
-                "vstm       %0!, {d16-d23}      \n"
-                "vstm       %0!, {d24-d31}      \n"
-
-                "11:                            \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(max_kk),       // %8
-                "r"(k),            // %9
-                "r"(k_end),        // %10
-                "r"(out_elempack), // %11
-                "r"(out_hstep)     // %12
-                : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            int32x4_t _sum0;
-            int32x4_t _sum1;
-            int32x4_t _sum2;
-            int32x4_t _sum3;
-            int32x4_t _sum4;
-            int32x4_t _sum5;
-            int32x4_t _sum6;
-            int32x4_t _sum7;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_s32(0);
-                _sum1 = vdupq_n_s32(0);
-                _sum2 = vdupq_n_s32(0);
-                _sum3 = vdupq_n_s32(0);
-                _sum4 = vdupq_n_s32(0);
-                _sum5 = vdupq_n_s32(0);
-                _sum6 = vdupq_n_s32(0);
-                _sum7 = vdupq_n_s32(0);
-            }
-            else
-            {
-                _sum0 = vld1q_s32(outptr);
-                _sum1 = vld1q_s32(outptr + 4);
-                _sum2 = vld1q_s32(outptr + 8);
-                _sum3 = vld1q_s32(outptr + 12);
-                _sum4 = vld1q_s32(outptr + 16);
-                _sum5 = vld1q_s32(outptr + 20);
-                _sum6 = vld1q_s32(outptr + 24);
-                _sum7 = vld1q_s32(outptr + 28);
-            }
-
-            int kk = 0;
-#if __ARM_FEATURE_DOTPROD
-            {
-#if __ARM_FEATURE_MATMUL_INT8
-                int32x4_t _s0 = vdupq_n_s32(0);
-                int32x4_t _s1 = vdupq_n_s32(0);
-                int32x4_t _s2 = vdupq_n_s32(0);
-                int32x4_t _s3 = vdupq_n_s32(0);
-                int32x4_t _s4 = vdupq_n_s32(0);
-                int32x4_t _s5 = vdupq_n_s32(0);
-                int32x4_t _s6 = vdupq_n_s32(0);
-                int32x4_t _s7 = vdupq_n_s32(0);
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int8x16_t _pA0 = vld1q_s8(pA);
-                    int8x16_t _pA1 = vld1q_s8(pA + 16);
-                    int8x16_t _pA2 = vld1q_s8(pA + 32);
-                    int8x16_t _pA3 = vld1q_s8(pA + 48);
-
-                    int8x16_t _pB0 = vld1q_s8(pB);
-                    int8x16_t _pB1 = vld1q_s8(pB + 16);
-
-#if __ARM_FEATURE_MATMUL_INT8
-                    // aaaaaaaa bbbbbbbb ..... hhhhhhhh
-                    // 00000000 11111111 22222222 33333333
-
-                    _s0 = vmmlaq_s32(_s0, _pA0, _pB0);
-                    _s1 = vmmlaq_s32(_s1, _pA1, _pB0);
-                    _s2 = vmmlaq_s32(_s2, _pA0, _pB1);
-                    _s3 = vmmlaq_s32(_s3, _pA1, _pB1);
-                    _s4 = vmmlaq_s32(_s4, _pA2, _pB0);
-                    _s5 = vmmlaq_s32(_s5, _pA3, _pB0);
-                    _s6 = vmmlaq_s32(_s6, _pA2, _pB1);
-                    _s7 = vmmlaq_s32(_s7, _pA3, _pB1);
-#else  // __ARM_FEATURE_MATMUL_INT8
-                    _sum0 = vdotq_laneq_s32(_sum0, _pA0, _pB0, 0);
-                    _sum1 = vdotq_laneq_s32(_sum1, _pA0, _pB0, 1);
-                    _sum2 = vdotq_laneq_s32(_sum2, _pA0, _pB0, 2);
-                    _sum3 = vdotq_laneq_s32(_sum3, _pA0, _pB0, 3);
-                    _sum4 = vdotq_laneq_s32(_sum4, _pA1, _pB0, 0);
-                    _sum5 = vdotq_laneq_s32(_sum5, _pA1, _pB0, 1);
-                    _sum6 = vdotq_laneq_s32(_sum6, _pA1, _pB0, 2);
-                    _sum7 = vdotq_laneq_s32(_sum7, _pA1, _pB0, 3);
-
-                    _sum0 = vdotq_laneq_s32(_sum0, _pA2, _pB1, 0);
-                    _sum1 = vdotq_laneq_s32(_sum1, _pA2, _pB1, 1);
-                    _sum2 = vdotq_laneq_s32(_sum2, _pA2, _pB1, 2);
-                    _sum3 = vdotq_laneq_s32(_sum3, _pA2, _pB1, 3);
-                    _sum4 = vdotq_laneq_s32(_sum4, _pA3, _pB1, 0);
-                    _sum5 = vdotq_laneq_s32(_sum5, _pA3, _pB1, 1);
-                    _sum6 = vdotq_laneq_s32(_sum6, _pA3, _pB1, 2);
-                    _sum7 = vdotq_laneq_s32(_sum7, _pA3, _pB1, 3);
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                    pA += 64;
-                    pB += 32;
-                }
-#if __ARM_FEATURE_MATMUL_INT8
-                int32x4x2_t _ss0 = vuzpq_s32(_s0, _s1);
-                int32x4x2_t _ss1 = vuzpq_s32(_s2, _s3);
-                int32x4x2_t _ss2 = vuzpq_s32(_s4, _s5);
-                int32x4x2_t _ss3 = vuzpq_s32(_s6, _s7);
-                _sum0 = vaddq_s32(_sum0, _ss0.val[0]);
-                _sum1 = vaddq_s32(_sum1, _ss0.val[1]);
-                _sum2 = vaddq_s32(_sum2, _ss1.val[0]);
-                _sum3 = vaddq_s32(_sum3, _ss1.val[1]);
-                _sum4 = vaddq_s32(_sum4, _ss2.val[0]);
-                _sum5 = vaddq_s32(_sum5, _ss2.val[1]);
-                _sum6 = vaddq_s32(_sum6, _ss3.val[0]);
-                _sum7 = vaddq_s32(_sum7, _ss3.val[1]);
-#endif // __ARM_FEATURE_MATMUL_INT8
-            }
-#endif // __ARM_FEATURE_DOTPROD
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x16_t _pA0 = vld1q_s8(pA);
-                int8x16_t _pA1 = vld1q_s8(pA + 16);
-                int8x16_t _pB = vld1q_s8(pB);
-
-                // aaaa bbbb cccc dddd   eeee ffff gggg hhhh
-
-                // 0000 1111 2222 3333
-
-                _sum0 = vdotq_laneq_s32(_sum0, _pA0, _pB, 0);
-                _sum1 = vdotq_laneq_s32(_sum1, _pA0, _pB, 1);
-                _sum2 = vdotq_laneq_s32(_sum2, _pA0, _pB, 2);
-                _sum3 = vdotq_laneq_s32(_sum3, _pA0, _pB, 3);
-                _sum4 = vdotq_laneq_s32(_sum4, _pA1, _pB, 0);
-                _sum5 = vdotq_laneq_s32(_sum5, _pA1, _pB, 1);
-                _sum6 = vdotq_laneq_s32(_sum6, _pA1, _pB, 2);
-                _sum7 = vdotq_laneq_s32(_sum7, _pA1, _pB, 3);
-#else  // __ARM_FEATURE_DOTPROD
-                int8x16_t _pA0 = vld1q_s8(pA);
-                int8x16_t _pA2 = vld1q_s8(pA + 16);
-                int8x16_t _pB02 = vld1q_s8(pB);
-
-                // aabbccdd eeffgghh
-
-                // ccddaabb gghheeff
-
-                int8x16_t _pA1 = vreinterpretq_s8_s32(vrev64q_s32(vreinterpretq_s32_s8(_pA0)));
-                int8x16_t _pA3 = vreinterpretq_s8_s32(vrev64q_s32(vreinterpretq_s32_s8(_pA2)));
-
-                // 00112233 44556677
-
-                // 33221100 77665544
-
-                int8x16_t _pB13 = vreinterpretq_s8_s16(vrev64q_s16(vreinterpretq_s16_s8(_pB02)));
-
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA0), vget_low_s8(_pB02));
-                int16x8_t _s1 = vmull_s8(vget_high_s8(_pA0), vget_low_s8(_pB02));
-                int16x8_t _s2 = vmull_s8(vget_low_s8(_pA1), vget_low_s8(_pB02));
-                int16x8_t _s3 = vmull_s8(vget_high_s8(_pA1), vget_low_s8(_pB02));
-                int16x8_t _s4 = vmull_s8(vget_low_s8(_pA0), vget_low_s8(_pB13));
-                int16x8_t _s5 = vmull_s8(vget_high_s8(_pA0), vget_low_s8(_pB13));
-                int16x8_t _s6 = vmull_s8(vget_low_s8(_pA1), vget_low_s8(_pB13));
-                int16x8_t _s7 = vmull_s8(vget_high_s8(_pA1), vget_low_s8(_pB13));
-
-                _s0 = vmlal_s8(_s0, vget_low_s8(_pA2), vget_high_s8(_pB02));
-                _s1 = vmlal_s8(_s1, vget_high_s8(_pA2), vget_high_s8(_pB02));
-                _s2 = vmlal_s8(_s2, vget_low_s8(_pA3), vget_high_s8(_pB02));
-                _s3 = vmlal_s8(_s3, vget_high_s8(_pA3), vget_high_s8(_pB02));
-                _s4 = vmlal_s8(_s4, vget_low_s8(_pA2), vget_high_s8(_pB13));
-                _s5 = vmlal_s8(_s5, vget_high_s8(_pA2), vget_high_s8(_pB13));
-                _s6 = vmlal_s8(_s6, vget_low_s8(_pA3), vget_high_s8(_pB13));
-                _s7 = vmlal_s8(_s7, vget_high_s8(_pA3), vget_high_s8(_pB13));
-
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-                _sum4 = vpadalq_s16(_sum4, _s4);
-                _sum5 = vpadalq_s16(_sum5, _s5);
-                _sum6 = vpadalq_s16(_sum6, _s6);
-                _sum7 = vpadalq_s16(_sum7, _s7);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 32;
-                pB += 16;
-            }
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x16_t _pA = vld1q_s8(pA);
-                int8x8_t _pB = vld1_s8(pB);
-
-                // aabbccdd eeffgghh
-
-                // 00112233
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 0)));
-                int16x8_t _s1 = vmull_s8(vget_low_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 1)));
-                int16x8_t _s2 = vmull_s8(vget_low_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 2)));
-                int16x8_t _s3 = vmull_s8(vget_low_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 3)));
-                int16x8_t _s4 = vmull_s8(vget_high_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 0)));
-                int16x8_t _s5 = vmull_s8(vget_high_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 1)));
-                int16x8_t _s6 = vmull_s8(vget_high_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 2)));
-                int16x8_t _s7 = vmull_s8(vget_high_s8(_pA), vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 3)));
-
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-                _sum4 = vpadalq_s16(_sum4, _s4);
-                _sum5 = vpadalq_s16(_sum5, _s5);
-                _sum6 = vpadalq_s16(_sum6, _s6);
-                _sum7 = vpadalq_s16(_sum7, _s7);
-#else  // __ARM_FEATURE_DOTPROD
-                int8x16_t _pA0 = vld1q_s8(pA);
-                int8x8_t _pB0 = vld1_s8(pB);
-
-                // aabbccdd eeffgghh
-
-                // ccddaabb gghheeff
-
-                int8x16_t _pA1 = vreinterpretq_s8_s32(vrev64q_s32(vreinterpretq_s32_s8(_pA0)));
-
-                // 00112233
-
-                // 33221100
-
-                int8x8_t _pB1 = vreinterpret_s8_s16(vrev64_s16(vreinterpret_s16_s8(_pB0)));
-
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA0), _pB0);
-                int16x8_t _s1 = vmull_s8(vget_high_s8(_pA0), _pB0);
-                int16x8_t _s2 = vmull_s8(vget_low_s8(_pA1), _pB0);
-                int16x8_t _s3 = vmull_s8(vget_high_s8(_pA1), _pB0);
-                int16x8_t _s4 = vmull_s8(vget_low_s8(_pA0), _pB1);
-                int16x8_t _s5 = vmull_s8(vget_high_s8(_pA0), _pB1);
-                int16x8_t _s6 = vmull_s8(vget_low_s8(_pA1), _pB1);
-                int16x8_t _s7 = vmull_s8(vget_high_s8(_pA1), _pB1);
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-                _sum4 = vpadalq_s16(_sum4, _s4);
-                _sum5 = vpadalq_s16(_sum5, _s5);
-                _sum6 = vpadalq_s16(_sum6, _s6);
-                _sum7 = vpadalq_s16(_sum7, _s7);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 16;
-                pB += 8;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x8_t _pA0 = vld1_s8(pA);
-                // int8x8_t _pB0 = vreinterpret_s32_s8(vld1_dup_s32(pB));
-
-                // abcdefgh
-
-                // 0123
-
-                int16x8_t _s01 = vmull_s8(_pA0, vdup_n_s8(pB[0]));
-                int16x8_t _s23 = vmull_s8(_pA0, vdup_n_s8(pB[1]));
-                int16x8_t _s45 = vmull_s8(_pA0, vdup_n_s8(pB[2]));
-                int16x8_t _s67 = vmull_s8(_pA0, vdup_n_s8(pB[3]));
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s01));
-                _sum1 = vaddw_s16(_sum1, vget_low_s16(_s23));
-                _sum2 = vaddw_s16(_sum2, vget_low_s16(_s45));
-                _sum3 = vaddw_s16(_sum3, vget_low_s16(_s67));
-                _sum4 = vaddw_s16(_sum4, vget_high_s16(_s01));
-                _sum5 = vaddw_s16(_sum5, vget_high_s16(_s23));
-                _sum6 = vaddw_s16(_sum6, vget_high_s16(_s45));
-                _sum7 = vaddw_s16(_sum7, vget_high_s16(_s67));
-#else  // __ARM_FEATURE_DOTPROD
-                int8x8_t _pA0 = vld1_s8(pA);
-                int8x8_t _pB0 = vreinterpret_s8_s32(vld1_dup_s32((const int*)pB));
-                // int8x8_t _pB0 = vld1_s8(pB);
-                // _pB0 = vreinterpret_s8_s32(vzip_s32(vreinterpret_s32_s8(_pB0), vreinterpret_s32_s8(_pB0)).val[0]);
-
-                // abcdefgh  ->  cdabghef
-                int8x8_t _pA1 = vreinterpret_s8_s16(vrev32_s16(vreinterpret_s16_s8(_pA0)));
-
-                // 01230123  ->  32103210
-                int8x8_t _pB1 = vrev64_s8(_pB0);
-
-                int16x8_t _s01 = vmull_s8(_pA0, _pB0);
-                int16x8_t _s23 = vmull_s8(_pA1, _pB0);
-                int16x8_t _s45 = vmull_s8(_pA0, _pB1);
-                int16x8_t _s67 = vmull_s8(_pA1, _pB1);
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s01));
-                _sum1 = vaddw_s16(_sum1, vget_high_s16(_s01));
-                _sum2 = vaddw_s16(_sum2, vget_low_s16(_s23));
-                _sum3 = vaddw_s16(_sum3, vget_high_s16(_s23));
-                _sum4 = vaddw_s16(_sum4, vget_low_s16(_s45));
-                _sum5 = vaddw_s16(_sum5, vget_high_s16(_s45));
-                _sum6 = vaddw_s16(_sum6, vget_low_s16(_s67));
-                _sum7 = vaddw_s16(_sum7, vget_high_s16(_s67));
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 8;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-#if __ARM_FEATURE_DOTPROD
-                // from
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                //      a2 b2 c2 d2
-                //      a3 b3 c3 d3
-                //      e0 f0 g0 h0
-                //      e1 f1 g1 h1
-                //      e2 f2 g2 h2
-                //      e3 f3 g3 h3
-                if (out_elempack == 8)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum4);
-                    vst1q_s32(outptr0 + 8, _sum1);
-                    vst1q_s32(outptr0 + 12, _sum5);
-                    vst1q_s32(outptr0 + 16, _sum2);
-                    vst1q_s32(outptr0 + 20, _sum6);
-                    vst1q_s32(outptr0 + 24, _sum3);
-                    vst1q_s32(outptr0 + 28, _sum7);
-                    outptr0 += 32;
-                }
-                if (out_elempack == 4)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + 8, _sum2);
-                    vst1q_s32(outptr0 + 12, _sum3);
-                    vst1q_s32(outptr0 + out_hstep * 4, _sum4);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 4, _sum5);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 8, _sum6);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 12, _sum7);
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    // to
-                    //      a0 a1 a2 a3
-                    //      b0 b1 b2 b3
-                    //      c0 c1 c2 c3
-                    //      d0 d1 d2 d3
-                    //      e0 e1 e2 e3
-                    //      f0 f1 f2 f3
-                    //      g0 g1 g2 g3
-                    //      h0 h1 h2 h3
-                    {
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sum1);
-                        int32x4x2_t _t1 = vzipq_s32(_sum2, _sum3);
-                        int32x4x2_t _t2 = vzipq_s32(_sum4, _sum5);
-                        int32x4x2_t _t3 = vzipq_s32(_sum6, _sum7);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t1.val[0]));
-                        _sum1 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t1.val[0]));
-                        _sum2 = vcombine_s32(vget_low_s32(_t0.val[1]), vget_low_s32(_t1.val[1]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t0.val[1]), vget_high_s32(_t1.val[1]));
-                        _sum4 = vcombine_s32(vget_low_s32(_t2.val[0]), vget_low_s32(_t3.val[0]));
-                        _sum5 = vcombine_s32(vget_high_s32(_t2.val[0]), vget_high_s32(_t3.val[0]));
-                        _sum6 = vcombine_s32(vget_low_s32(_t2.val[1]), vget_low_s32(_t3.val[1]));
-                        _sum7 = vcombine_s32(vget_high_s32(_t2.val[1]), vget_high_s32(_t3.val[1]));
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + out_hstep, _sum1);
-                    vst1q_s32(outptr0 + out_hstep * 2, _sum2);
-                    vst1q_s32(outptr0 + out_hstep * 3, _sum3);
-                    vst1q_s32(outptr0 + out_hstep * 4, _sum4);
-                    vst1q_s32(outptr0 + out_hstep * 5, _sum5);
-                    vst1q_s32(outptr0 + out_hstep * 6, _sum6);
-                    vst1q_s32(outptr0 + out_hstep * 7, _sum7);
-                    outptr0 += 4;
-                }
-#else  // __ARM_FEATURE_DOTPROD
-
-                // from
-                //      a0 b1 c2 d3
-                //      e0 f1 g2 h3
-                //      c0 d1 a2 b3
-                //      g0 h1 e2 f3
-                //      a3 b2 c1 d0
-                //      e3 f2 g1 h0
-                //      c3 d2 a1 b0
-                //      g3 h2 e1 f0
-                if (out_elempack == 8)
-                {
-                    // to
-                    //      a0 b0 c0 d0
-                    //      e0 f0 g0 h0
-                    //      a1 b1 c1 d1
-                    //      e1 f1 g1 h1
-                    //      a2 b2 c2 d2
-                    //      e2 f2 g2 h2
-                    //      a3 b3 c3 d3
-                    //      e3 f3 g3 h3
-                    {
-                        _sum4 = vrev64q_s32(_sum4);
-                        _sum5 = vrev64q_s32(_sum5);
-                        _sum6 = vrev64q_s32(_sum6);
-                        _sum7 = vrev64q_s32(_sum7);
-                        _sum4 = vextq_s32(_sum4, _sum4, 2);
-                        _sum5 = vextq_s32(_sum5, _sum5, 2);
-                        _sum6 = vextq_s32(_sum6, _sum6, 2);
-                        _sum7 = vextq_s32(_sum7, _sum7, 2);
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sum6);
-                        int32x4x2_t _t1 = vzipq_s32(_sum2, _sum4);
-                        int32x4x2_t _t2 = vzipq_s32(_sum1, _sum7);
-                        int32x4x2_t _t3 = vzipq_s32(_sum3, _sum5);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t1.val[0]));
-                        _sum1 = vcombine_s32(vget_low_s32(_t2.val[0]), vget_low_s32(_t3.val[0]));
-                        _sum2 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t1.val[0]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t2.val[0]), vget_high_s32(_t3.val[0]));
-                        _sum4 = vcombine_s32(vget_low_s32(_t1.val[1]), vget_low_s32(_t0.val[1]));
-                        _sum5 = vcombine_s32(vget_low_s32(_t3.val[1]), vget_low_s32(_t2.val[1]));
-                        _sum6 = vcombine_s32(vget_high_s32(_t1.val[1]), vget_high_s32(_t0.val[1]));
-                        _sum7 = vcombine_s32(vget_high_s32(_t3.val[1]), vget_high_s32(_t2.val[1]));
-                        _sum2 = vrev64q_s32(_sum2);
-                        _sum3 = vrev64q_s32(_sum3);
-                        _sum6 = vrev64q_s32(_sum6);
-                        _sum7 = vrev64q_s32(_sum7);
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + 8, _sum2);
-                    vst1q_s32(outptr0 + 12, _sum3);
-                    vst1q_s32(outptr0 + 16, _sum4);
-                    vst1q_s32(outptr0 + 20, _sum5);
-                    vst1q_s32(outptr0 + 24, _sum6);
-                    vst1q_s32(outptr0 + 28, _sum7);
-                    outptr0 += 32;
-                }
-                if (out_elempack == 4)
-                {
-                    // to
-                    //      a0 b0 c0 d0
-                    //      a1 b1 c1 d1
-                    //      a2 b2 c2 d2
-                    //      a3 b3 c3 d3
-                    //      e0 f0 g0 h0
-                    //      e1 f1 g1 h1
-                    //      e2 f2 g2 h2
-                    //      e3 f3 g3 h3
-                    {
-                        _sum4 = vrev64q_s32(_sum4);
-                        _sum5 = vrev64q_s32(_sum5);
-                        _sum6 = vrev64q_s32(_sum6);
-                        _sum7 = vrev64q_s32(_sum7);
-                        _sum4 = vextq_s32(_sum4, _sum4, 2);
-                        _sum5 = vextq_s32(_sum5, _sum5, 2);
-                        _sum6 = vextq_s32(_sum6, _sum6, 2);
-                        _sum7 = vextq_s32(_sum7, _sum7, 2);
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sum6);
-                        int32x4x2_t _t1 = vzipq_s32(_sum2, _sum4);
-                        int32x4x2_t _t2 = vzipq_s32(_sum1, _sum7);
-                        int32x4x2_t _t3 = vzipq_s32(_sum3, _sum5);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t1.val[0]));
-                        _sum1 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t1.val[0]));
-                        _sum2 = vcombine_s32(vget_low_s32(_t1.val[1]), vget_low_s32(_t0.val[1]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t1.val[1]), vget_high_s32(_t0.val[1]));
-                        _sum4 = vcombine_s32(vget_low_s32(_t2.val[0]), vget_low_s32(_t3.val[0]));
-                        _sum5 = vcombine_s32(vget_high_s32(_t2.val[0]), vget_high_s32(_t3.val[0]));
-                        _sum6 = vcombine_s32(vget_low_s32(_t3.val[1]), vget_low_s32(_t2.val[1]));
-                        _sum7 = vcombine_s32(vget_high_s32(_t3.val[1]), vget_high_s32(_t2.val[1]));
-                        _sum1 = vrev64q_s32(_sum1);
-                        _sum3 = vrev64q_s32(_sum3);
-                        _sum5 = vrev64q_s32(_sum5);
-                        _sum7 = vrev64q_s32(_sum7);
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + 8, _sum2);
-                    vst1q_s32(outptr0 + 12, _sum3);
-                    vst1q_s32(outptr0 + out_hstep * 4, _sum4);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 4, _sum5);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 8, _sum6);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 12, _sum7);
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    // to
-                    //      a0 a1 a2 a3
-                    //      b0 b1 b2 b3
-                    //      c0 c1 c2 c3
-                    //      d0 d1 d2 d3
-                    //      e0 e1 e2 e3
-                    //      f0 f1 f2 f3
-                    //      g0 g1 g2 g3
-                    //      h0 h1 h2 h3
-                    {
-                        _sum2 = vextq_s32(_sum2, _sum2, 2);
-                        _sum3 = vextq_s32(_sum3, _sum3, 2);
-                        _sum6 = vextq_s32(_sum6, _sum6, 2);
-                        _sum7 = vextq_s32(_sum7, _sum7, 2);
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sum6);
-                        int32x4x2_t _t1 = vzipq_s32(_sum2, _sum4);
-                        int32x4x2_t _t2 = vzipq_s32(_sum1, _sum7);
-                        int32x4x2_t _t3 = vzipq_s32(_sum3, _sum5);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t1.val[0]));
-                        _sum1 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t1.val[0]));
-                        _sum2 = vcombine_s32(vget_low_s32(_t1.val[1]), vget_low_s32(_t0.val[1]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t1.val[1]), vget_high_s32(_t0.val[1]));
-                        _sum4 = vcombine_s32(vget_low_s32(_t2.val[0]), vget_low_s32(_t3.val[0]));
-                        _sum5 = vcombine_s32(vget_high_s32(_t2.val[0]), vget_high_s32(_t3.val[0]));
-                        _sum6 = vcombine_s32(vget_low_s32(_t3.val[1]), vget_low_s32(_t2.val[1]));
-                        _sum7 = vcombine_s32(vget_high_s32(_t3.val[1]), vget_high_s32(_t2.val[1]));
-                        _sum1 = vrev64q_s32(_sum1);
-                        _sum3 = vrev64q_s32(_sum3);
-                        _sum5 = vrev64q_s32(_sum5);
-                        _sum7 = vrev64q_s32(_sum7);
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + out_hstep, _sum1);
-                    vst1q_s32(outptr0 + out_hstep * 2, _sum2);
-                    vst1q_s32(outptr0 + out_hstep * 3, _sum3);
-                    vst1q_s32(outptr0 + out_hstep * 4, _sum4);
-                    vst1q_s32(outptr0 + out_hstep * 5, _sum5);
-                    vst1q_s32(outptr0 + out_hstep * 6, _sum6);
-                    vst1q_s32(outptr0 + out_hstep * 7, _sum7);
-                    outptr0 += 4;
-                }
-#endif // __ARM_FEATURE_DOTPROD
-            }
-            else
-            {
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-                vst1q_s32(outptr + 16, _sum4);
-                vst1q_s32(outptr + 20, _sum5);
-                vst1q_s32(outptr + 24, _sum6);
-                vst1q_s32(outptr + 28, _sum7);
-            }
-
-            outptr += 32;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            const signed char* pA = pAT;
-
-            int32x4_t _sum0;
-            int32x4_t _sum1;
-            int32x4_t _sum2;
-            int32x4_t _sum3;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_s32(0);
-                _sum1 = vdupq_n_s32(0);
-                _sum2 = vdupq_n_s32(0);
-                _sum3 = vdupq_n_s32(0);
-            }
-            else
-            {
-                _sum0 = vld1q_s32(outptr);
-                _sum1 = vld1q_s32(outptr + 4);
-                _sum2 = vld1q_s32(outptr + 8);
-                _sum3 = vld1q_s32(outptr + 12);
-            }
-
-            int kk = 0;
-#if __ARM_FEATURE_DOTPROD
-            {
-#if __ARM_FEATURE_MATMUL_INT8
-                int32x4_t _s0 = vdupq_n_s32(0);
-                int32x4_t _s1 = vdupq_n_s32(0);
-                int32x4_t _s2 = vdupq_n_s32(0);
-                int32x4_t _s3 = vdupq_n_s32(0);
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int8x16_t _pA0 = vld1q_s8(pA);
-                    int8x16_t _pA1 = vld1q_s8(pA + 16);
-                    int8x16_t _pA2 = vld1q_s8(pA + 32);
-                    int8x16_t _pA3 = vld1q_s8(pA + 48);
-
-                    int8x16_t _pB = vld1q_s8(pB);
-
-#if __ARM_FEATURE_MATMUL_INT8
-                    // aaaaaaaa bbbbbbbb ..... hhhhhhhh
-                    // 00000000 11111111
-
-                    _s0 = vmmlaq_s32(_s0, _pA0, _pB);
-                    _s1 = vmmlaq_s32(_s1, _pA1, _pB);
-                    _s2 = vmmlaq_s32(_s2, _pA2, _pB);
-                    _s3 = vmmlaq_s32(_s3, _pA3, _pB);
-#else  // __ARM_FEATURE_MATMUL_INT8
-                    _sum0 = vdotq_laneq_s32(_sum0, _pA0, _pB, 0);
-                    _sum1 = vdotq_laneq_s32(_sum1, _pA0, _pB, 1);
-                    _sum2 = vdotq_laneq_s32(_sum2, _pA1, _pB, 0);
-                    _sum3 = vdotq_laneq_s32(_sum3, _pA1, _pB, 1);
-
-                    _sum0 = vdotq_laneq_s32(_sum0, _pA2, _pB, 2);
-                    _sum1 = vdotq_laneq_s32(_sum1, _pA2, _pB, 3);
-                    _sum2 = vdotq_laneq_s32(_sum2, _pA3, _pB, 2);
-                    _sum3 = vdotq_laneq_s32(_sum3, _pA3, _pB, 3);
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                    pA += 64;
-                    pB += 16;
-                }
-#if __ARM_FEATURE_MATMUL_INT8
-                int32x4x2_t _ss0 = vuzpq_s32(_s0, _s1);
-                int32x4x2_t _ss1 = vuzpq_s32(_s2, _s3);
-                _sum0 = vaddq_s32(_sum0, _ss0.val[0]);
-                _sum1 = vaddq_s32(_sum1, _ss0.val[1]);
-                _sum2 = vaddq_s32(_sum2, _ss1.val[0]);
-                _sum3 = vaddq_s32(_sum3, _ss1.val[1]);
-#endif // __ARM_FEATURE_MATMUL_INT8
-            }
-#endif // __ARM_FEATURE_DOTPROD
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x16_t _pA0 = vld1q_s8(pA);
-                int8x16_t _pA1 = vld1q_s8(pA + 16);
-                int8x8_t _pB = vld1_s8(pB);
-
-                // aaaa bbbb cccc dddd eeee ffff gggg hhhh
-
-                // 0000 1111
-
-                _sum0 = vdotq_lane_s32(_sum0, _pA0, _pB, 0);
-                _sum1 = vdotq_lane_s32(_sum1, _pA0, _pB, 1);
-                _sum2 = vdotq_lane_s32(_sum2, _pA1, _pB, 0);
-                _sum3 = vdotq_lane_s32(_sum3, _pA1, _pB, 1);
-#else  // __ARM_FEATURE_DOTPROD
-                int8x16_t _pA0 = vld1q_s8(pA);
-                int8x16_t _pA2 = vld1q_s8(pA + 16);
-                int8x8_t _pB = vld1_s8(pB);
-
-                // aabbccdd eeffgghh   aabbccdd eeffgghh
-
-                // 00112233 -> 00110011 22332233
-
-                // 11001100 33223322
-
-                int32x2x2_t _pBB = vzip_s32(vreinterpret_s32_s8(_pB), vreinterpret_s32_s8(_pB));
-                int8x16_t _pB02 = vreinterpretq_s8_s32(vcombine_s32(_pBB.val[0], _pBB.val[1]));
-
-                int8x16_t _pB13 = vreinterpretq_s8_s16(vrev64q_s16(vreinterpretq_s16_s8(_pB02)));
-
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA0), vget_low_s8(_pB02));
-                int16x8_t _s1 = vmull_s8(vget_high_s8(_pA0), vget_low_s8(_pB02));
-                int16x8_t _s2 = vmull_s8(vget_low_s8(_pA0), vget_low_s8(_pB13));
-                int16x8_t _s3 = vmull_s8(vget_high_s8(_pA0), vget_low_s8(_pB13));
-                _s0 = vmlal_s8(_s0, vget_low_s8(_pA2), vget_high_s8(_pB02));
-                _s1 = vmlal_s8(_s1, vget_high_s8(_pA2), vget_high_s8(_pB02));
-                _s2 = vmlal_s8(_s2, vget_low_s8(_pA2), vget_high_s8(_pB13));
-                _s3 = vmlal_s8(_s3, vget_high_s8(_pA2), vget_high_s8(_pB13));
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 32;
-                pB += 8;
-            }
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x16_t _pA = vld1q_s8(pA);
-                int16x4_t _pB = vreinterpret_s16_s32(vld1_dup_s32((const int*)pB));
-
-                int16x4x2_t _pB01 = vuzp_s16(_pB, _pB);
-                int8x8_t _pB0 = vreinterpret_s8_s16(_pB01.val[0]);
-                int8x8_t _pB1 = vreinterpret_s8_s16(_pB01.val[1]);
-
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA), _pB0);
-                int16x8_t _s1 = vmull_s8(vget_low_s8(_pA), _pB1);
-                int16x8_t _s2 = vmull_s8(vget_high_s8(_pA), _pB0);
-                int16x8_t _s3 = vmull_s8(vget_high_s8(_pA), _pB1);
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-#else  // __ARM_FEATURE_DOTPROD
-                int8x16_t _pA = vld1q_s8(pA);
-                int8x8_t _pB0 = vreinterpret_s8_s32(vld1_dup_s32((const int*)pB));
-
-                // aabbccdd eeffgghh
-
-                // 00110011
-                // 11001100
-
-                int8x8_t _pB1 = vreinterpret_s8_s16(vrev64_s16(vreinterpret_s16_s8(_pB0)));
-
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA), _pB0);
-                int16x8_t _s1 = vmull_s8(vget_high_s8(_pA), _pB0);
-                int16x8_t _s2 = vmull_s8(vget_low_s8(_pA), _pB1);
-                int16x8_t _s3 = vmull_s8(vget_high_s8(_pA), _pB1);
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 16;
-                pB += 4;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x8_t _pA = vld1_s8(pA);
-                int8x8_t _pB = vreinterpret_s8_s16(vld1_dup_s16((const short*)pB));
-
-                int8x8x2_t _pB01 = vuzp_s8(_pB, _pB);
-
-                int16x8_t _s0 = vmull_s8(_pA, _pB01.val[0]);
-                int16x8_t _s1 = vmull_s8(_pA, _pB01.val[1]);
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s0));
-                _sum1 = vaddw_s16(_sum1, vget_low_s16(_s1));
-                _sum2 = vaddw_s16(_sum2, vget_high_s16(_s0));
-                _sum3 = vaddw_s16(_sum3, vget_high_s16(_s1));
-#else  // __ARM_FEATURE_DOTPROD
-                int8x8_t _pA = vld1_s8(pA);
-                int8x8_t _pB0 = vreinterpret_s8_s16(vld1_dup_s16((const short*)pB));
-
-                // abcdefgh
-
-                // 01010101
-                // 10101010
-                int8x8_t _pB1 = vext_s8(_pB0, _pB0, 1);
-
-                int16x8_t _s0 = vmull_s8(_pA, _pB0);
-                int16x8_t _s1 = vmull_s8(_pA, _pB1);
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s0));
-                _sum1 = vaddw_s16(_sum1, vget_high_s16(_s0));
-                _sum2 = vaddw_s16(_sum2, vget_low_s16(_s1));
-                _sum3 = vaddw_s16(_sum3, vget_high_s16(_s1));
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 8;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-#if __ARM_FEATURE_DOTPROD
-                // from
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                //      e0 f0 g0 h0
-                //      e1 f1 g1 h1
-                if (out_elempack == 8)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum2);
-                    vst1q_s32(outptr0 + 8, _sum1);
-                    vst1q_s32(outptr0 + 12, _sum3);
-                    outptr0 += 16;
-                }
-                if (out_elempack == 4)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + out_hstep * 4, _sum2);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 4, _sum3);
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-                    // to
-                    //      a0 a1 b0 b1
-                    //      c0 c1 d0 d1
-                    //      e0 e1 f0 f1
-                    //      g0 g1 h0 h1
-                    {
-                        int32x4x2_t _sum02 = vzipq_s32(_sum0, _sum1);
-                        int32x4x2_t _sum13 = vzipq_s32(_sum2, _sum3);
-                        _sum0 = _sum02.val[0];
-                        _sum1 = _sum02.val[1];
-                        _sum2 = _sum13.val[0];
-                        _sum3 = _sum13.val[1];
-                    }
-
-                    vst1_s32(outptr0, vget_low_s32(_sum0));
-                    vst1_s32(outptr0 + out_hstep, vget_high_s32(_sum0));
-                    vst1_s32(outptr0 + out_hstep * 2, vget_low_s32(_sum1));
-                    vst1_s32(outptr0 + out_hstep * 3, vget_high_s32(_sum1));
-                    vst1_s32(outptr0 + out_hstep * 4, vget_low_s32(_sum2));
-                    vst1_s32(outptr0 + out_hstep * 5, vget_high_s32(_sum2));
-                    vst1_s32(outptr0 + out_hstep * 6, vget_low_s32(_sum3));
-                    vst1_s32(outptr0 + out_hstep * 7, vget_high_s32(_sum3));
-                    outptr0 += 2;
-                }
-#else  // __ARM_FEATURE_DOTPROD
-
-                // from
-                //      a0 b1 c0 d1
-                //      e0 f1 g0 h1
-                //      a1 b0 c1 d0
-                //      e1 f0 g1 h0
-                if (out_elempack == 8)
-                {
-                    // to
-                    //      a0 b0 c0 d0
-                    //      e0 f0 g0 h0
-                    //      a1 b1 c1 d1
-                    //      e1 f1 g1 h1
-                    {
-                        _sum2 = vrev64q_s32(_sum2);
-                        _sum3 = vrev64q_s32(_sum3);
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sum2);
-                        int32x4x2_t _t1 = vzipq_s32(_sum1, _sum3);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t0.val[1]));
-                        _sum1 = vcombine_s32(vget_low_s32(_t1.val[0]), vget_low_s32(_t1.val[1]));
-                        _sum2 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t0.val[1]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t1.val[0]), vget_high_s32(_t1.val[1]));
-                        _sum2 = vrev64q_s32(_sum2);
-                        _sum3 = vrev64q_s32(_sum3);
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + 8, _sum2);
-                    vst1q_s32(outptr0 + 12, _sum3);
-                    outptr0 += 16;
-                }
-                if (out_elempack == 4)
-                {
-                    // to
-                    //      a0 b0 c0 d0
-                    //      a1 b1 c1 d1
-                    //      e0 f0 g0 h0
-                    //      e1 f1 g1 h1
-                    {
-                        _sum2 = vrev64q_s32(_sum2);
-                        _sum3 = vrev64q_s32(_sum3);
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sum2);
-                        int32x4x2_t _t1 = vzipq_s32(_sum1, _sum3);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t0.val[1]));
-                        _sum1 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t0.val[1]));
-                        _sum2 = vcombine_s32(vget_low_s32(_t1.val[0]), vget_low_s32(_t1.val[1]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t1.val[0]), vget_high_s32(_t1.val[1]));
-                        _sum1 = vrev64q_s32(_sum1);
-                        _sum3 = vrev64q_s32(_sum3);
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + out_hstep * 4, _sum2);
-                    vst1q_s32(outptr0 + out_hstep * 4 + 4, _sum3);
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-                    // to
-                    //      a0 a1 c0 c1
-                    //      b0 b1 d0 d1
-                    //      e0 e1 g0 g1
-                    //      f0 f1 h0 h1
-                    {
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sum2);
-                        int32x4x2_t _t1 = vzipq_s32(_sum1, _sum3);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t0.val[1]));
-                        _sum1 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t0.val[1]));
-                        _sum2 = vcombine_s32(vget_low_s32(_t1.val[0]), vget_low_s32(_t1.val[1]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t1.val[0]), vget_high_s32(_t1.val[1]));
-                        _sum1 = vrev64q_s32(_sum1);
-                        _sum3 = vrev64q_s32(_sum3);
-                    }
-
-                    vst1_s32(outptr0, vget_low_s32(_sum0));
-                    vst1_s32(outptr0 + out_hstep, vget_low_s32(_sum1));
-                    vst1_s32(outptr0 + out_hstep * 2, vget_high_s32(_sum0));
-                    vst1_s32(outptr0 + out_hstep * 3, vget_high_s32(_sum1));
-                    vst1_s32(outptr0 + out_hstep * 4, vget_low_s32(_sum2));
-                    vst1_s32(outptr0 + out_hstep * 5, vget_low_s32(_sum3));
-                    vst1_s32(outptr0 + out_hstep * 6, vget_high_s32(_sum2));
-                    vst1_s32(outptr0 + out_hstep * 7, vget_high_s32(_sum3));
-                    outptr0 += 2;
-                }
-#endif // __ARM_FEATURE_DOTPROD
-            }
-            else
-            {
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-            }
-
-            outptr += 16;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            const signed char* pA = pAT;
-
-            int32x4_t _sum0;
-            int32x4_t _sum1;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_s32(0);
-                _sum1 = vdupq_n_s32(0);
-            }
-            else
-            {
-                _sum0 = vld1q_s32(outptr);
-                _sum1 = vld1q_s32(outptr + 4);
-            }
-
-            int kk = 0;
-#if __ARM_FEATURE_DOTPROD
-            {
-#if __ARM_FEATURE_MATMUL_INT8
-                int32x4_t _s0 = vdupq_n_s32(0);
-                int32x4_t _s1 = vdupq_n_s32(0);
-                int32x4_t _s2 = vdupq_n_s32(0);
-                int32x4_t _s3 = vdupq_n_s32(0);
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int8x16_t _pA0 = vld1q_s8(pA);
-                    int8x16_t _pA1 = vld1q_s8(pA + 16);
-                    int8x16_t _pA2 = vld1q_s8(pA + 32);
-                    int8x16_t _pA3 = vld1q_s8(pA + 48);
-
-                    int8x8_t _pB = vld1_s8(pB);
-
-#if __ARM_FEATURE_MATMUL_INT8
-                    // aaaaaaaa bbbbbbbb ..... hhhhhhhh
-                    // 00000000
-                    int8x16_t _pBB = vcombine_s8(_pB, _pB);
-
-                    _s0 = vdotq_s32(_s0, _pA0, _pBB);
-                    _s1 = vdotq_s32(_s1, _pA1, _pBB);
-                    _s2 = vdotq_s32(_s2, _pA2, _pBB);
-                    _s3 = vdotq_s32(_s3, _pA3, _pBB);
-#else  // __ARM_FEATURE_MATMUL_INT8
-                    _sum0 = vdotq_lane_s32(_sum0, _pA0, _pB, 0);
-                    _sum1 = vdotq_lane_s32(_sum1, _pA1, _pB, 0);
-                    _sum0 = vdotq_lane_s32(_sum0, _pA2, _pB, 1);
-                    _sum1 = vdotq_lane_s32(_sum1, _pA3, _pB, 1);
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                    pA += 64;
-                    pB += 8;
-                }
-#if __ARM_FEATURE_MATMUL_INT8
-                _sum0 = vaddq_s32(_sum0, vpaddq_s32(_s0, _s1));
-                _sum1 = vaddq_s32(_sum1, vpaddq_s32(_s2, _s3));
-#endif // __ARM_FEATURE_MATMUL_INT8
-            }
-#endif // __ARM_FEATURE_DOTPROD
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x16_t _pA0 = vld1q_s8(pA);
-                int8x16_t _pA1 = vld1q_s8(pA + 16);
-
-                int8x8_t _pB = vreinterpret_s8_s32(vld1_dup_s32((const int*)pB));
-
-                // aaaa bbbb cccc dddd eeee ffff gggg hhhh
-
-                // 0000 0000
-
-                _sum0 = vdotq_lane_s32(_sum0, _pA0, _pB, 0);
-                _sum1 = vdotq_lane_s32(_sum1, _pA1, _pB, 0);
-#else  // __ARM_FEATURE_DOTPROD
-                int8x16_t _pA0 = vld1q_s8(pA);
-                int8x16_t _pA2 = vld1q_s8(pA + 16);
-                int8x8_t _pB0 = vreinterpret_s8_s16(vld1_dup_s16((const short*)pB));
-                int8x8_t _pB1 = vreinterpret_s8_s16(vld1_dup_s16((const short*)(pB + 2)));
-
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA0), _pB0);
-                int16x8_t _s1 = vmull_s8(vget_high_s8(_pA0), _pB0);
-                _s0 = vmlal_s8(_s0, vget_low_s8(_pA2), _pB1);
-                _s1 = vmlal_s8(_s1, vget_high_s8(_pA2), _pB1);
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 32;
-                pB += 4;
-            }
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                int8x16_t _pA = vld1q_s8(pA);
-                int8x8_t _pB = vreinterpret_s8_s16(vld1_dup_s16((const short*)pB));
-
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA), _pB);
-                int16x8_t _s1 = vmull_s8(vget_high_s8(_pA), _pB);
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-
-                pA += 16;
-                pB += 2;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                int8x8_t _pA = vld1_s8(pA);
-                int8x8_t _pB = vld1_dup_s8(pB);
-
-                int16x8_t _s0 = vmull_s8(_pA, _pB);
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s0));
-                _sum1 = vaddw_s16(_sum1, vget_high_s16(_s0));
-
-                pA += 8;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 8)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    outptr0 += 8;
-                }
-                if (out_elempack == 4)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + out_hstep * 4, _sum1);
-                    outptr0 += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    outptr0[0] = vgetq_lane_s32(_sum0, 0);
-                    outptr0[out_hstep] = vgetq_lane_s32(_sum0, 1);
-                    outptr0[out_hstep * 2] = vgetq_lane_s32(_sum0, 2);
-                    outptr0[out_hstep * 3] = vgetq_lane_s32(_sum0, 3);
-                    outptr0[out_hstep * 4] = vgetq_lane_s32(_sum1, 0);
-                    outptr0[out_hstep * 5] = vgetq_lane_s32(_sum1, 1);
-                    outptr0[out_hstep * 6] = vgetq_lane_s32(_sum1, 2);
-                    outptr0[out_hstep * 7] = vgetq_lane_s32(_sum1, 3);
-                    outptr0++;
-                }
-            }
-            else
-            {
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-            }
-
-            outptr += 8;
-        }
-
-        pAT += max_kk * 8;
-    }
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        int* outptr0 = (int*)top_blob + (i + ii) * out_hstep + j * out_elempack;
-
-        const signed char* pB = pBT;
-
-        int jj = 0;
-#if __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            const signed char* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-            asm volatile(
-                "cmp    %w9, #0                     \n"
-                "beq    0f                          \n"
-
-                "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0] \n"
-                "sub    %0, %0, #64                 \n"
-                "b      1f                          \n"
-
-                "0:                                 \n"
-                "eor    v16.16b, v16.16b, v16.16b   \n"
-                "eor    v17.16b, v17.16b, v17.16b   \n"
-                "eor    v18.16b, v18.16b, v18.16b   \n"
-                "eor    v19.16b, v19.16b, v19.16b   \n"
-                "eor    v20.16b, v20.16b, v20.16b   \n"
-                "eor    v21.16b, v21.16b, v21.16b   \n"
-                "eor    v22.16b, v22.16b, v22.16b   \n"
-                "eor    v23.16b, v23.16b, v23.16b   \n"
-
-                "1:                                 \n"
-#if __ARM_FEATURE_DOTPROD
-                "lsr    w4, %w8, #3                 \n" // w4 = max_kk >> 3
-                "cmp    w4, #0                      \n"
-                "beq    101f                        \n"
-
-#if __ARM_FEATURE_MATMUL_INT8
-                "eor    v24.16b, v24.16b, v24.16b   \n"
-                "eor    v25.16b, v25.16b, v25.16b   \n"
-                "eor    v26.16b, v26.16b, v26.16b   \n"
-                "eor    v27.16b, v27.16b, v27.16b   \n"
-                "eor    v28.16b, v28.16b, v28.16b   \n"
-                "eor    v29.16b, v29.16b, v29.16b   \n"
-                "eor    v30.16b, v30.16b, v30.16b   \n"
-                "eor    v31.16b, v31.16b, v31.16b   \n"
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                "2:                                 \n"
-                "ld1    {v0.16b, v1.16b}, [%1], #32 \n"
-                "ld1    {v2.16b, v3.16b, v4.16b, v5.16b}, [%2], #64 \n"
-
-#if __ARM_FEATURE_MATMUL_INT8
-                "smmla  v24.4s, v0.16b, v2.16b      \n"
-                "smmla  v25.4s, v1.16b, v2.16b      \n"
-                "smmla  v26.4s, v0.16b, v3.16b      \n"
-                "smmla  v27.4s, v1.16b, v3.16b      \n"
-                "subs   w4, w4, #1                  \n"
-                "smmla  v28.4s, v0.16b, v4.16b      \n"
-                "smmla  v29.4s, v1.16b, v4.16b      \n"
-                "smmla  v30.4s, v0.16b, v5.16b      \n"
-                "smmla  v31.4s, v1.16b, v5.16b      \n"
-#else  // __ARM_FEATURE_MATMUL_INT8
-                "sdot   v16.4s, v0.16b, v2.4b[0]    \n"
-                "sdot   v17.4s, v0.16b, v2.4b[1]    \n"
-                "sdot   v18.4s, v0.16b, v2.4b[2]    \n"
-                "sdot   v19.4s, v0.16b, v2.4b[3]    \n"
-                "sdot   v20.4s, v0.16b, v3.4b[0]    \n"
-                "sdot   v21.4s, v0.16b, v3.4b[1]    \n"
-                "sdot   v22.4s, v0.16b, v3.4b[2]    \n"
-                "sdot   v23.4s, v0.16b, v3.4b[3]    \n"
-                "subs   w4, w4, #1                  \n"
-                "sdot   v16.4s, v1.16b, v4.4b[0]    \n"
-                "sdot   v17.4s, v1.16b, v4.4b[1]    \n"
-                "sdot   v18.4s, v1.16b, v4.4b[2]    \n"
-                "sdot   v19.4s, v1.16b, v4.4b[3]    \n"
-                "sdot   v20.4s, v1.16b, v5.4b[0]    \n"
-                "sdot   v21.4s, v1.16b, v5.4b[1]    \n"
-                "sdot   v22.4s, v1.16b, v5.4b[2]    \n"
-                "sdot   v23.4s, v1.16b, v5.4b[3]    \n"
-#endif // __ARM_FEATURE_MATMUL_INT8
-                "bne    2b                          \n"
-
-#if __ARM_FEATURE_MATMUL_INT8
-                "uzp1   v0.4s, v24.4s, v25.4s       \n"
-                "uzp2   v1.4s, v24.4s, v25.4s       \n"
-                "uzp1   v2.4s, v26.4s, v27.4s       \n"
-                "uzp2   v3.4s, v26.4s, v27.4s       \n"
-                "uzp1   v4.4s, v28.4s, v29.4s       \n"
-                "uzp2   v5.4s, v28.4s, v29.4s       \n"
-                "uzp1   v6.4s, v30.4s, v31.4s       \n"
-                "uzp2   v7.4s, v30.4s, v31.4s       \n"
-
-                "add    v16.4s, v16.4s, v0.4s       \n"
-                "add    v17.4s, v17.4s, v1.4s       \n"
-                "add    v18.4s, v18.4s, v2.4s       \n"
-                "add    v19.4s, v19.4s, v3.4s       \n"
-                "add    v20.4s, v20.4s, v4.4s       \n"
-                "add    v21.4s, v21.4s, v5.4s       \n"
-                "add    v22.4s, v22.4s, v6.4s       \n"
-                "add    v23.4s, v23.4s, v7.4s       \n"
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                "101:                               \n"
-                "and    w4, %w8, #4                 \n" // w4 = remain = max_kk & 4
-                "cmp    w4, #0                      \n"
-                "beq    3f                          \n"
-
-                // kk += 4 part
-                "ld1    {v0.16b}, [%1], #16         \n"
-                "ld1    {v2.16b, v3.16b}, [%2], #32 \n"
-                "sdot   v16.4s, v0.16b, v2.4b[0]    \n"
-                "sdot   v17.4s, v0.16b, v2.4b[1]    \n"
-                "sdot   v18.4s, v0.16b, v2.4b[2]    \n"
-                "sdot   v19.4s, v0.16b, v2.4b[3]    \n"
-                "sdot   v20.4s, v0.16b, v3.4b[0]    \n"
-                "sdot   v21.4s, v0.16b, v3.4b[1]    \n"
-                "sdot   v22.4s, v0.16b, v3.4b[2]    \n"
-                "sdot   v23.4s, v0.16b, v3.4b[3]    \n"
-#else  // __ARM_FEATURE_DOTPROD
-                "lsr    w4, %w8, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    3f                          \n"
-
-                "2:                                 \n"
-                "ld1    {v0.16b}, [%1], #16         \n"
-                "ld1    {v4.16b, v5.16b}, [%2], #32 \n"
-                "smull  v8.8h, v0.8b, v4.8b         \n"
-                "smull2 v9.8h, v0.16b, v5.16b       \n"
-                "rev64  v2.4s, v0.4s                \n"
-                "smull  v10.8h, v2.8b, v4.8b        \n"
-                "smull2 v11.8h, v2.16b, v5.16b      \n"
-                "rev64  v6.8h, v4.8h                \n"
-                "smull  v12.8h, v0.8b, v6.8b        \n"
-                "smull  v14.8h, v2.8b, v6.8b        \n"
-                "rev64  v7.8h, v5.8h                \n"
-                "smull2 v13.8h, v0.16b, v7.16b      \n"
-                "smull2 v15.8h, v2.16b, v7.16b      \n"
-                "ext    v1.16b, v0.16b, v0.16b, #8  \n"
-                "ext    v3.16b, v2.16b, v2.16b, #8  \n"
-                "smlal  v8.8h, v1.8b, v5.8b         \n"
-                "smlal2 v9.8h, v1.16b, v4.16b       \n"
-                "smlal  v10.8h, v3.8b, v5.8b        \n"
-                "smlal2 v11.8h, v3.16b, v4.16b      \n"
-                "smlal  v12.8h, v1.8b, v7.8b        \n"
-                "smlal  v14.8h, v3.8b, v7.8b        \n"
-                "smlal2 v13.8h, v1.16b, v6.16b      \n"
-                "smlal2 v15.8h, v3.16b, v6.16b      \n"
-                "subs   w4, w4, #1                  \n"
-                "sadalp v16.4s, v8.8h               \n"
-                "sadalp v17.4s, v9.8h               \n"
-                "sadalp v18.4s, v10.8h              \n"
-                "sadalp v19.4s, v11.8h              \n"
-                "sadalp v20.4s, v12.8h              \n"
-                "sadalp v22.4s, v14.8h              \n"
-                "sadalp v21.4s, v13.8h              \n"
-                "sadalp v23.4s, v15.8h              \n"
-                "bne    2b                          \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "3:                                 \n"
-                "and    w4, %w8, #2                 \n" // w4 = remain = max_kk & 2
-                "cmp    w4, #0                      \n"
-                "beq    4f                          \n"
-
-                // kk += 2 part
-#if __ARM_FEATURE_DOTPROD
-                "ld1    {v0.8b}, [%1], #8           \n"
-                "ld1    {v1.16b}, [%2], #16         \n"
-                "dup    v4.8h, v1.h[0]              \n"
-                "dup    v5.8h, v1.h[1]              \n"
-                "dup    v6.8h, v1.h[2]              \n"
-                "dup    v7.8h, v1.h[3]              \n"
-                "smull  v8.8h, v0.8b, v4.8b         \n"
-                "smull  v9.8h, v0.8b, v5.8b         \n"
-                "smull  v10.8h, v0.8b, v6.8b        \n"
-                "smull  v11.8h, v0.8b, v7.8b        \n"
-                "dup    v4.8h, v1.h[4]              \n"
-                "dup    v5.8h, v1.h[5]              \n"
-                "dup    v6.8h, v1.h[6]              \n"
-                "dup    v7.8h, v1.h[7]              \n"
-                "smull  v12.8h, v0.8b, v4.8b        \n"
-                "smull  v13.8h, v0.8b, v5.8b        \n"
-                "smull  v14.8h, v0.8b, v6.8b        \n"
-                "smull  v15.8h, v0.8b, v7.8b        \n"
-                "sadalp v16.4s, v8.8h               \n"
-                "sadalp v17.4s, v9.8h               \n"
-                "sadalp v18.4s, v10.8h              \n"
-                "sadalp v19.4s, v11.8h              \n"
-                "sadalp v20.4s, v12.8h              \n"
-                "sadalp v21.4s, v13.8h              \n"
-                "sadalp v22.4s, v14.8h              \n"
-                "sadalp v23.4s, v15.8h              \n"
-#else  // __ARM_FEATURE_DOTPROD
-                "ld1r   {v0.2d}, [%1]               \n"
-                "add    %1, %1, #8                  \n"
-                "ld1    {v2.16b}, [%2], #16         \n"
-                "rev64  v1.4s, v0.4s                \n"
-                "rev64  v3.8h, v2.8h                \n"
-                "smull  v8.8h, v0.8b, v2.8b         \n"
-                "smull2 v9.8h, v0.16b, v2.16b       \n"
-                "smull  v10.8h, v1.8b, v2.8b        \n"
-                "smull2 v11.8h, v1.16b, v2.16b      \n"
-                "smull  v12.8h, v0.8b, v3.8b        \n"
-                "smull2 v13.8h, v0.16b, v3.16b      \n"
-                "smull  v14.8h, v1.8b, v3.8b        \n"
-                "smull2 v15.8h, v1.16b, v3.16b      \n"
-                "sadalp v16.4s, v8.8h               \n"
-                "sadalp v17.4s, v9.8h               \n"
-                "sadalp v18.4s, v10.8h              \n"
-                "sadalp v19.4s, v11.8h              \n"
-                "sadalp v20.4s, v12.8h              \n"
-                "sadalp v21.4s, v13.8h              \n"
-                "sadalp v22.4s, v14.8h              \n"
-                "sadalp v23.4s, v15.8h              \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "4:                                 \n"
-                "and    w4, %w8, #1                 \n" // w4 = remain = max_kk & 1
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                // kk += 1 part
-#if __ARM_FEATURE_DOTPROD
-                "ld1r   {v0.2s}, [%1]               \n"
-                "ld1    {v1.8b}, [%2], #8           \n"
-                "add    %1, %1, #4                  \n"
-                "dup    v8.8h, v1.h[0]              \n"
-                "dup    v9.8h, v1.h[1]              \n"
-                "dup    v10.8h, v1.h[2]             \n"
-                "dup    v11.8h, v1.h[3]             \n"
-                "uzp1   v2.8b, v8.8b, v9.8b         \n"
-                "uzp2   v3.8b, v8.8b, v9.8b         \n"
-                "uzp1   v4.8b, v10.8b, v11.8b       \n"
-                "uzp2   v5.8b, v10.8b, v11.8b       \n"
-                "smull  v8.8h, v0.8b, v2.8b         \n"
-                "smull  v9.8h, v0.8b, v3.8b         \n"
-                "smull  v10.8h, v0.8b, v4.8b        \n"
-                "smull  v11.8h, v0.8b, v5.8b        \n"
-                "saddw  v16.4s, v16.4s, v8.4h       \n"
-                "saddw  v17.4s, v17.4s, v9.4h       \n"
-                "saddw2 v18.4s, v18.4s, v8.8h       \n"
-                "saddw2 v19.4s, v19.4s, v9.8h       \n"
-                "saddw  v20.4s, v20.4s, v10.4h      \n"
-                "saddw  v21.4s, v21.4s, v11.4h      \n"
-                "saddw2 v22.4s, v22.4s, v10.8h      \n"
-                "saddw2 v23.4s, v23.4s, v11.8h      \n"
-#else  // __ARM_FEATURE_DOTPROD
-                "ld1r   {v0.2s}, [%1]               \n"
-                "ld1    {v2.8b}, [%2], #8           \n"
-                "add    %1, %1, #4                  \n"
-                "ext    v1.8b, v0.8b, v0.8b, #2     \n"
-                "rev32  v3.8b, v2.8b                \n"
-                "smull  v8.8h, v0.8b, v2.8b         \n"
-                "smull  v9.8h, v1.8b, v2.8b         \n"
-                "smull  v10.8h, v0.8b, v3.8b        \n"
-                "smull  v11.8h, v1.8b, v3.8b        \n"
-                "saddw  v16.4s, v16.4s, v8.4h       \n"
-                "saddw2 v17.4s, v17.4s, v8.8h       \n"
-                "saddw  v18.4s, v18.4s, v9.4h       \n"
-                "saddw2 v19.4s, v19.4s, v9.8h       \n"
-                "saddw  v20.4s, v20.4s, v10.4h      \n"
-                "saddw2 v21.4s, v21.4s, v10.8h      \n"
-                "saddw  v22.4s, v22.4s, v11.4h      \n"
-                "saddw2 v23.4s, v23.4s, v11.8h      \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "5:                                 \n"
-                "cmp    %w10, #0                    \n"
-                "beq    10f                         \n"
-
-#if __ARM_FEATURE_DOTPROD
-                // from
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                //      a2 b2 c2 d2
-                //      a3 b3 c3 d3
-                //      a4 b4 c4 d4
-                //      a5 b5 c5 d5
-                //      a6 b6 c6 d6
-                //      a7 b7 c7 d7
-                // if out_elempack == 4
-                "cmp    %w11, #1                    \n"
-                "beq    8f                          \n"
-
-                "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%3], #64 \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // to
-                //      a0 a1 a2 a3
-                //      a4 a5 a6 a7
-                //      b0 b1 b2 b3
-                //      b4 b5 b6 b7
-                //      c0 c1 c2 c3
-                //      c4 c5 c6 c7
-                //      d0 d1 d2 d3
-                //      d4 d5 d6 d7
-                "zip1   v0.4s, v16.4s, v17.4s       \n"
-                "zip2   v1.4s, v16.4s, v17.4s       \n"
-                "zip1   v2.4s, v18.4s, v19.4s       \n"
-                "zip2   v3.4s, v18.4s, v19.4s       \n"
-                "zip1   v4.4s, v20.4s, v21.4s       \n"
-                "zip2   v5.4s, v20.4s, v21.4s       \n"
-                "zip1   v6.4s, v22.4s, v23.4s       \n"
-                "zip2   v7.4s, v22.4s, v23.4s       \n"
-                "zip1   v16.2d, v0.2d, v2.2d        \n"
-                "zip1   v17.2d, v4.2d, v6.2d        \n"
-                "zip2   v18.2d, v0.2d, v2.2d        \n"
-                "zip2   v19.2d, v4.2d, v6.2d        \n"
-                "zip1   v20.2d, v1.2d, v3.2d        \n"
-                "zip1   v21.2d, v5.2d, v7.2d        \n"
-                "zip2   v22.2d, v1.2d, v3.2d        \n"
-                "zip2   v23.2d, v5.2d, v7.2d        \n"
-
-                "add    x4, %3, %12, lsl #2         \n"
-                "st1    {v16.4s, v17.4s}, [%3], #32 \n"
-                "st1    {v18.4s, v19.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v20.4s, v21.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v22.4s, v23.4s}, [x4]      \n"
-#else  // __ARM_FEATURE_DOTPROD
-
-                // from
-                //      a0 b1 c2 d3
-                //      a4 b5 c6 d7
-                //      c0 d1 a2 b3
-                //      c4 d5 a6 b7
-                //      a3 b2 c1 d0
-                //      a7 b6 c5 d4
-                //      c3 d2 a1 b0
-                //      c7 d6 a5 b4
-                // if out_elempack == 4
-                "cmp    %w11, #1                    \n"
-                "beq    8f                          \n"
-
-                // to
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                //      a2 b2 c2 d2
-                //      a3 b3 c3 d3
-                //      a4 b4 c4 d4
-                //      a5 b5 c5 d5
-                //      a6 b6 c6 d6
-                //      a7 b7 c7 d7
-                "rev64  v20.4s, v20.4s              \n"
-                "rev64  v21.4s, v21.4s              \n"
-                "rev64  v22.4s, v22.4s              \n"
-                "rev64  v23.4s, v23.4s              \n"
-                "ext    v20.16b, v20.16b, v20.16b, #8 \n"
-                "ext    v21.16b, v21.16b, v21.16b, #8 \n"
-                "ext    v22.16b, v22.16b, v22.16b, #8 \n"
-                "ext    v23.16b, v23.16b, v23.16b, #8 \n"
-                "zip1   v0.4s, v16.4s, v22.4s       \n"
-                "zip2   v1.4s, v16.4s, v22.4s       \n"
-                "zip1   v2.4s, v18.4s, v20.4s       \n"
-                "zip2   v3.4s, v18.4s, v20.4s       \n"
-                "zip1   v4.4s, v17.4s, v23.4s       \n"
-                "zip2   v5.4s, v17.4s, v23.4s       \n"
-                "zip1   v6.4s, v19.4s, v21.4s       \n"
-                "zip2   v7.4s, v19.4s, v21.4s       \n"
-                "zip1   v16.2d, v0.2d, v2.2d        \n"
-                "zip2   v17.2d, v0.2d, v2.2d        \n"
-                "zip1   v18.2d, v3.2d, v1.2d        \n"
-                "zip2   v19.2d, v3.2d, v1.2d        \n"
-                "zip1   v20.2d, v4.2d, v6.2d        \n"
-                "zip2   v21.2d, v4.2d, v6.2d        \n"
-                "zip1   v22.2d, v7.2d, v5.2d        \n"
-                "zip2   v23.2d, v7.2d, v5.2d        \n"
-                "rev64  v17.4s, v17.4s              \n"
-                "rev64  v19.4s, v19.4s              \n"
-                "rev64  v21.4s, v21.4s              \n"
-                "rev64  v23.4s, v23.4s              \n"
-
-                "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%3], #64 \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-
-                // to
-                //      a0 a1 a2 a3
-                //      a4 a5 a6 a7
-                //      b0 b1 b2 b3
-                //      b4 b5 b6 b7
-                //      c0 c1 c2 c3
-                //      c4 c5 c6 c7
-                //      d0 d1 d2 d3
-                //      d4 d5 d6 d7
-                "ext    v18.16b, v18.16b, v18.16b, #8 \n"
-                "ext    v19.16b, v19.16b, v19.16b, #8 \n"
-                "ext    v22.16b, v22.16b, v22.16b, #8 \n"
-                "ext    v23.16b, v23.16b, v23.16b, #8 \n"
-                "zip1   v0.4s, v16.4s, v22.4s       \n"
-                "zip2   v1.4s, v16.4s, v22.4s       \n"
-                "zip1   v2.4s, v18.4s, v20.4s       \n"
-                "zip2   v3.4s, v18.4s, v20.4s       \n"
-                "zip1   v4.4s, v17.4s, v23.4s       \n"
-                "zip2   v5.4s, v17.4s, v23.4s       \n"
-                "zip1   v6.4s, v19.4s, v21.4s       \n"
-                "zip2   v7.4s, v19.4s, v21.4s       \n"
-                "zip1   v16.2d, v0.2d, v2.2d        \n"
-                "zip1   v17.2d, v4.2d, v6.2d        \n"
-                "zip2   v18.2d, v0.2d, v2.2d        \n"
-                "zip2   v19.2d, v4.2d, v6.2d        \n"
-                "zip1   v20.2d, v3.2d, v1.2d        \n"
-                "zip1   v21.2d, v7.2d, v5.2d        \n"
-                "zip2   v22.2d, v3.2d, v1.2d        \n"
-                "zip2   v23.2d, v7.2d, v5.2d        \n"
-                "rev64  v18.4s, v18.4s              \n"
-                "rev64  v19.4s, v19.4s              \n"
-                "rev64  v22.4s, v22.4s              \n"
-                "rev64  v23.4s, v23.4s              \n"
-
-                "add    x4, %3, %12, lsl #2         \n"
-                "st1    {v16.4s, v17.4s}, [%3], #32 \n"
-                "st1    {v18.4s, v19.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v20.4s, v21.4s}, [x4]      \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v22.4s, v23.4s}, [x4]      \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "9:                                 \n"
-                "add    %0, %0, #128                \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(max_kk),       // %8
-                "r"(k),            // %9
-                "r"(k_end),        // %10
-                "r"(out_elempack), // %11
-                "r"(out_hstep)     // %12
-                : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else // NCNN_GNU_INLINE_ASM
-            int32x4_t _sum0;
-            int32x4_t _sum1;
-            int32x4_t _sum2;
-            int32x4_t _sum3;
-            int32x4_t _sum4;
-            int32x4_t _sum5;
-            int32x4_t _sum6;
-            int32x4_t _sum7;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_s32(0);
-                _sum1 = vdupq_n_s32(0);
-                _sum2 = vdupq_n_s32(0);
-                _sum3 = vdupq_n_s32(0);
-                _sum4 = vdupq_n_s32(0);
-                _sum5 = vdupq_n_s32(0);
-                _sum6 = vdupq_n_s32(0);
-                _sum7 = vdupq_n_s32(0);
-            }
-            else
-            {
-                _sum0 = vld1q_s32(outptr);
-                _sum1 = vld1q_s32(outptr + 4);
-                _sum2 = vld1q_s32(outptr + 8);
-                _sum3 = vld1q_s32(outptr + 12);
-                _sum4 = vld1q_s32(outptr + 16);
-                _sum5 = vld1q_s32(outptr + 20);
-                _sum6 = vld1q_s32(outptr + 24);
-                _sum7 = vld1q_s32(outptr + 28);
-            }
-
-            int kk = 0;
-#if __ARM_FEATURE_MATMUL_INT8
-            {
-                int32x4_t _sum00 = vdupq_n_s32(0);
-                int32x4_t _sum01 = vdupq_n_s32(0);
-                int32x4_t _sum10 = vdupq_n_s32(0);
-                int32x4_t _sum11 = vdupq_n_s32(0);
-                int32x4_t _sum20 = vdupq_n_s32(0);
-                int32x4_t _sum21 = vdupq_n_s32(0);
-                int32x4_t _sum30 = vdupq_n_s32(0);
-                int32x4_t _sum31 = vdupq_n_s32(0);
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int8x16_t _pA0 = vld1q_s8(pA);
-                    int8x16_t _pA1 = vld1q_s8(pA + 16);
-                    int8x16_t _pB0 = vld1q_s8(pB);
-                    int8x16_t _pB1 = vld1q_s8(pB + 16);
-                    int8x16_t _pB2 = vld1q_s8(pB + 32);
-                    int8x16_t _pB3 = vld1q_s8(pB + 48);
-
-                    // aaaaaaaa bbbbbbbb cccccccc dddddddd
-
-                    // 00000000 11111111 22222222 33333333
-                    // 44444444 55555555 66666666 77777777
-
-                    _sum00 = vmmlaq_s32(_sum00, _pA0, _pB0);
-                    _sum01 = vmmlaq_s32(_sum01, _pA1, _pB0);
-                    _sum10 = vmmlaq_s32(_sum10, _pA0, _pB1);
-                    _sum11 = vmmlaq_s32(_sum11, _pA1, _pB1);
-                    _sum20 = vmmlaq_s32(_sum20, _pA0, _pB2);
-                    _sum21 = vmmlaq_s32(_sum21, _pA1, _pB2);
-                    _sum30 = vmmlaq_s32(_sum30, _pA0, _pB3);
-                    _sum31 = vmmlaq_s32(_sum31, _pA1, _pB3);
-
-                    // a0 a1 b0 b1
-                    // c0 c1 d0 d1
-                    // a2 a3 b2 b3
-                    // c2 c3 d2 d3
-                    // a4 a5 b4 b5
-                    // c4 c5 d4 d5
-                    // a6 a7 b6 b7
-                    // c6 c7 d6 d7
-
-                    pA += 32;
-                    pB += 64;
-                }
-                int32x4x2_t _ss0 = vuzpq_s32(_sum00, _sum01);
-                int32x4x2_t _ss1 = vuzpq_s32(_sum10, _sum11);
-                int32x4x2_t _ss2 = vuzpq_s32(_sum20, _sum21);
-                int32x4x2_t _ss3 = vuzpq_s32(_sum30, _sum31);
-                _sum0 = vaddq_s32(_sum0, _ss0.val[0]);
-                _sum1 = vaddq_s32(_sum1, _ss0.val[1]);
-                _sum2 = vaddq_s32(_sum2, _ss1.val[0]);
-                _sum3 = vaddq_s32(_sum3, _ss1.val[1]);
-                _sum4 = vaddq_s32(_sum4, _ss2.val[0]);
-                _sum5 = vaddq_s32(_sum5, _ss2.val[1]);
-                _sum6 = vaddq_s32(_sum6, _ss3.val[0]);
-                _sum7 = vaddq_s32(_sum7, _ss3.val[1]);
-            }
-#elif __ARM_FEATURE_DOTPROD
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                int8x16_t _pA0 = vld1q_s8(pA);
-                int8x16_t _pA1 = vld1q_s8(pA + 16);
-                int8x16_t _pB0 = vld1q_s8(pB);
-                int8x16_t _pB1 = vld1q_s8(pB + 16);
-                int8x16_t _pB2 = vld1q_s8(pB + 32);
-                int8x16_t _pB3 = vld1q_s8(pB + 48);
-
-                _sum0 = vdotq_laneq_s32(_sum0, _pA0, _pB0, 0);
-                _sum1 = vdotq_laneq_s32(_sum1, _pA0, _pB0, 1);
-                _sum2 = vdotq_laneq_s32(_sum2, _pA0, _pB0, 2);
-                _sum3 = vdotq_laneq_s32(_sum3, _pA0, _pB0, 3);
-                _sum4 = vdotq_laneq_s32(_sum4, _pA0, _pB1, 0);
-                _sum5 = vdotq_laneq_s32(_sum5, _pA0, _pB1, 1);
-                _sum6 = vdotq_laneq_s32(_sum6, _pA0, _pB1, 2);
-                _sum7 = vdotq_laneq_s32(_sum7, _pA0, _pB1, 3);
-
-                _sum0 = vdotq_laneq_s32(_sum0, _pA1, _pB2, 0);
-                _sum1 = vdotq_laneq_s32(_sum1, _pA1, _pB2, 1);
-                _sum2 = vdotq_laneq_s32(_sum2, _pA1, _pB2, 2);
-                _sum3 = vdotq_laneq_s32(_sum3, _pA1, _pB2, 3);
-                _sum4 = vdotq_laneq_s32(_sum4, _pA1, _pB3, 0);
-                _sum5 = vdotq_laneq_s32(_sum5, _pA1, _pB3, 1);
-                _sum6 = vdotq_laneq_s32(_sum6, _pA1, _pB3, 2);
-                _sum7 = vdotq_laneq_s32(_sum7, _pA1, _pB3, 3);
-
-                pA += 32;
-                pB += 64;
-            }
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x16_t _pA = vld1q_s8(pA);
-                int8x16_t _pB0 = vld1q_s8(pB);
-                int8x16_t _pB1 = vld1q_s8(pB + 16);
-
-                _sum0 = vdotq_laneq_s32(_sum0, _pA, _pB0, 0);
-                _sum1 = vdotq_laneq_s32(_sum1, _pA, _pB0, 1);
-                _sum2 = vdotq_laneq_s32(_sum2, _pA, _pB0, 2);
-                _sum3 = vdotq_laneq_s32(_sum3, _pA, _pB0, 3);
-                _sum4 = vdotq_laneq_s32(_sum4, _pA, _pB1, 0);
-                _sum5 = vdotq_laneq_s32(_sum5, _pA, _pB1, 1);
-                _sum6 = vdotq_laneq_s32(_sum6, _pA, _pB1, 2);
-                _sum7 = vdotq_laneq_s32(_sum7, _pA, _pB1, 3);
-#else  // __ARM_FEATURE_DOTPROD
-                int8x16_t _pA02 = vld1q_s8(pA);
-                int8x16_t _pB0 = vld1q_s8(pB);
-                int8x16_t _pB2 = vld1q_s8(pB + 16);
-
-                int8x16_t _pA13 = vreinterpretq_s8_s32(vrev64q_s32(vreinterpretq_s32_s8(_pA02)));
-
-                int8x16_t _pB1 = vreinterpretq_s8_s16(vrev64q_s16(vreinterpretq_s16_s8(_pB0)));
-                int8x16_t _pB3 = vreinterpretq_s8_s16(vrev64q_s16(vreinterpretq_s16_s8(_pB2)));
-
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA02), vget_low_s8(_pB0));
-                int16x8_t _s1 = vmull_s8(vget_low_s8(_pA02), vget_high_s8(_pB0));
-                int16x8_t _s2 = vmull_s8(vget_low_s8(_pA13), vget_low_s8(_pB0));
-                int16x8_t _s3 = vmull_s8(vget_low_s8(_pA13), vget_high_s8(_pB0));
-                int16x8_t _s4 = vmull_s8(vget_low_s8(_pA02), vget_low_s8(_pB1));
-                int16x8_t _s5 = vmull_s8(vget_low_s8(_pA02), vget_high_s8(_pB1));
-                int16x8_t _s6 = vmull_s8(vget_low_s8(_pA13), vget_low_s8(_pB1));
-                int16x8_t _s7 = vmull_s8(vget_low_s8(_pA13), vget_high_s8(_pB1));
-
-                _s0 = vmlal_s8(_s0, vget_high_s8(_pA02), vget_low_s8(_pB2));
-                _s1 = vmlal_s8(_s1, vget_high_s8(_pA02), vget_high_s8(_pB2));
-                _s2 = vmlal_s8(_s2, vget_high_s8(_pA13), vget_low_s8(_pB2));
-                _s3 = vmlal_s8(_s3, vget_high_s8(_pA13), vget_high_s8(_pB2));
-                _s4 = vmlal_s8(_s4, vget_high_s8(_pA02), vget_low_s8(_pB3));
-                _s5 = vmlal_s8(_s5, vget_high_s8(_pA02), vget_high_s8(_pB3));
-                _s6 = vmlal_s8(_s6, vget_high_s8(_pA13), vget_low_s8(_pB3));
-                _s7 = vmlal_s8(_s7, vget_high_s8(_pA13), vget_high_s8(_pB3));
-
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-                _sum4 = vpadalq_s16(_sum4, _s4);
-                _sum5 = vpadalq_s16(_sum5, _s5);
-                _sum6 = vpadalq_s16(_sum6, _s6);
-                _sum7 = vpadalq_s16(_sum7, _s7);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 16;
-                pB += 32;
-            }
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x8_t _pA0 = vld1_s8(pA);
-                int8x16_t _pB01 = vld1q_s8(pB);
-
-                // aabbccdd
-
-                // 00112233 44556677
-
-                int16x8_t _s0 = vmull_s8(_pA0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_low_s8(_pB01)), 0)));
-                int16x8_t _s1 = vmull_s8(_pA0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_low_s8(_pB01)), 1)));
-                int16x8_t _s2 = vmull_s8(_pA0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_low_s8(_pB01)), 2)));
-                int16x8_t _s3 = vmull_s8(_pA0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_low_s8(_pB01)), 3)));
-                int16x8_t _s4 = vmull_s8(_pA0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_high_s8(_pB01)), 0)));
-                int16x8_t _s5 = vmull_s8(_pA0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_high_s8(_pB01)), 1)));
-                int16x8_t _s6 = vmull_s8(_pA0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_high_s8(_pB01)), 2)));
-                int16x8_t _s7 = vmull_s8(_pA0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vget_high_s8(_pB01)), 3)));
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-                _sum4 = vpadalq_s16(_sum4, _s4);
-                _sum5 = vpadalq_s16(_sum5, _s5);
-                _sum6 = vpadalq_s16(_sum6, _s6);
-                _sum7 = vpadalq_s16(_sum7, _s7);
-#else  // __ARM_FEATURE_DOTPROD
-                int8x8_t _pA0 = vld1_s8(pA);
-                int8x16_t _pB0 = vld1q_s8(pB);
-
-                // aabbccdd
-                // ccddaabb
-
-                int8x8_t _pA1 = vreinterpret_s8_s32(vrev64_s32(vreinterpret_s32_s8(_pA0)));
-
-                // 00112233 44556677
-                // 33221100 77665544
-
-                int8x16_t _pB1 = vreinterpretq_s8_s16(vrev64q_s16(vreinterpretq_s16_s8(_pB0)));
-
-                int16x8_t _s0 = vmull_s8(_pA0, vget_low_s8(_pB0));
-                int16x8_t _s1 = vmull_s8(_pA0, vget_high_s8(_pB0));
-                int16x8_t _s2 = vmull_s8(_pA1, vget_low_s8(_pB0));
-                int16x8_t _s3 = vmull_s8(_pA1, vget_high_s8(_pB0));
-                int16x8_t _s4 = vmull_s8(_pA0, vget_low_s8(_pB1));
-                int16x8_t _s5 = vmull_s8(_pA0, vget_high_s8(_pB1));
-                int16x8_t _s6 = vmull_s8(_pA1, vget_low_s8(_pB1));
-                int16x8_t _s7 = vmull_s8(_pA1, vget_high_s8(_pB1));
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-                _sum4 = vpadalq_s16(_sum4, _s4);
-                _sum5 = vpadalq_s16(_sum5, _s5);
-                _sum6 = vpadalq_s16(_sum6, _s6);
-                _sum7 = vpadalq_s16(_sum7, _s7);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 8;
-                pB += 16;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x8_t _pAA = vreinterpret_s8_s32(vld1_dup_s32((const int*)pA));
-                int8x8_t _pB = vld1_s8(pB);
-
-                // abcdabcd
-                // 01234567  ->  01010101 23232323 45454545 67676767
-                int8x8_t _pB0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 0));
-                int8x8_t _pB2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 1));
-                int8x8_t _pB4 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 2));
-                int8x8_t _pB6 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 3));
-
-                int8x8x2_t _pB0123 = vuzp_s8(_pB0, _pB2);
-                int8x8x2_t _pB4567 = vuzp_s8(_pB4, _pB6);
-
-                int16x8_t _s02 = vmull_s8(_pAA, _pB0123.val[0]);
-                int16x8_t _s13 = vmull_s8(_pAA, _pB0123.val[1]);
-                int16x8_t _s46 = vmull_s8(_pAA, _pB4567.val[0]);
-                int16x8_t _s57 = vmull_s8(_pAA, _pB4567.val[1]);
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s02));
-                _sum1 = vaddw_s16(_sum1, vget_low_s16(_s13));
-                _sum2 = vaddw_s16(_sum2, vget_high_s16(_s02));
-                _sum3 = vaddw_s16(_sum3, vget_high_s16(_s13));
-                _sum4 = vaddw_s16(_sum4, vget_low_s16(_s46));
-                _sum5 = vaddw_s16(_sum5, vget_low_s16(_s57));
-                _sum6 = vaddw_s16(_sum6, vget_high_s16(_s46));
-                _sum7 = vaddw_s16(_sum7, vget_high_s16(_s57));
-#else  // __ARM_FEATURE_DOTPROD
-                int8x8_t _pA0 = vreinterpret_s8_s32(vld1_dup_s32((const int*)pA));
-                int8x8_t _pB0 = vld1_s8(pB);
-
-                // abcd abcd
-                // cdab cdab
-
-                int8x8_t _pA1 = vext_s8(_pA0, _pA0, 2);
-
-                // 0123 4567
-                // 3210 7654
-
-                int8x8_t _pB1 = vrev32_s8(_pB0);
-
-                int16x8_t _s01 = vmull_s8(_pA0, _pB0);
-                int16x8_t _s23 = vmull_s8(_pA1, _pB0);
-                int16x8_t _s45 = vmull_s8(_pA0, _pB1);
-                int16x8_t _s67 = vmull_s8(_pA1, _pB1);
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s01));
-                _sum1 = vaddw_s16(_sum1, vget_high_s16(_s01));
-                _sum2 = vaddw_s16(_sum2, vget_low_s16(_s23));
-                _sum3 = vaddw_s16(_sum3, vget_high_s16(_s23));
-                _sum4 = vaddw_s16(_sum4, vget_low_s16(_s45));
-                _sum5 = vaddw_s16(_sum5, vget_high_s16(_s45));
-                _sum6 = vaddw_s16(_sum6, vget_low_s16(_s67));
-                _sum7 = vaddw_s16(_sum7, vget_high_s16(_s67));
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 4;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-#if __ARM_FEATURE_DOTPROD
-                // from
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                //      a2 b2 c2 d2
-                //      a3 b3 c3 d3
-                //      a4 b4 c4 d4
-                //      a5 b5 c5 d5
-                //      a6 b6 c6 d6
-                //      a7 b7 c7 d7
-                if (out_elempack == 4)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + 8, _sum2);
-                    vst1q_s32(outptr0 + 12, _sum3);
-                    vst1q_s32(outptr0 + 16, _sum4);
-                    vst1q_s32(outptr0 + 20, _sum5);
-                    vst1q_s32(outptr0 + 24, _sum6);
-                    vst1q_s32(outptr0 + 28, _sum7);
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-                    // to
-                    //      a0 a1 a2 a3
-                    //      a4 a5 a6 a7
-                    //      b0 b1 b2 b3
-                    //      b4 b5 b6 b7
-                    //      c0 c1 c2 c3
-                    //      c4 c5 c6 c7
-                    //      d0 d1 d2 d3
-                    //      d4 d5 d6 d7
-                    {
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sum1);
-                        int32x4x2_t _t1 = vzipq_s32(_sum2, _sum3);
-                        int32x4x2_t _t2 = vzipq_s32(_sum4, _sum5);
-                        int32x4x2_t _t3 = vzipq_s32(_sum6, _sum7);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t1.val[0]));
-                        _sum1 = vcombine_s32(vget_low_s32(_t2.val[0]), vget_low_s32(_t3.val[0]));
-                        _sum2 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t1.val[0]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t2.val[0]), vget_high_s32(_t3.val[0]));
-                        _sum4 = vcombine_s32(vget_low_s32(_t0.val[1]), vget_low_s32(_t1.val[1]));
-                        _sum5 = vcombine_s32(vget_low_s32(_t2.val[1]), vget_low_s32(_t3.val[1]));
-                        _sum6 = vcombine_s32(vget_high_s32(_t0.val[1]), vget_high_s32(_t1.val[1]));
-                        _sum7 = vcombine_s32(vget_high_s32(_t2.val[1]), vget_high_s32(_t3.val[1]));
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + out_hstep, _sum2);
-                    vst1q_s32(outptr0 + out_hstep + 4, _sum3);
-                    vst1q_s32(outptr0 + out_hstep * 2, _sum4);
-                    vst1q_s32(outptr0 + out_hstep * 2 + 4, _sum5);
-                    vst1q_s32(outptr0 + out_hstep * 3, _sum6);
-                    vst1q_s32(outptr0 + out_hstep * 3 + 4, _sum7);
-                    outptr0 += 8;
-                }
-#else  // __ARM_FEATURE_DOTPROD
-
-                // from
-                //      a0 b1 c2 d3
-                //      a4 b5 c6 d7
-                //      c0 d1 a2 b3
-                //      c4 d5 a6 b7
-                //      a3 b2 c1 d0
-                //      a7 b6 c5 d4
-                //      c3 d2 a1 b0
-                //      c7 d6 a5 b4
-                if (out_elempack == 4)
-                {
-                    // to
-                    //      a0 b0 c0 d0
-                    //      a1 b1 c1 d1
-                    //      a2 b2 c2 d2
-                    //      a3 b3 c3 d3
-                    //      a4 b4 c4 d4
-                    //      a5 b5 c5 d5
-                    //      a6 b6 c6 d6
-                    //      a7 b7 c7 d7
-                    {
-                        _sum4 = vrev64q_s32(_sum4);
-                        _sum5 = vrev64q_s32(_sum5);
-                        _sum6 = vrev64q_s32(_sum6);
-                        _sum7 = vrev64q_s32(_sum7);
-                        _sum4 = vextq_s32(_sum4, _sum4, 2);
-                        _sum5 = vextq_s32(_sum5, _sum5, 2);
-                        _sum6 = vextq_s32(_sum6, _sum6, 2);
-                        _sum7 = vextq_s32(_sum7, _sum7, 2);
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sum6);
-                        int32x4x2_t _t1 = vzipq_s32(_sum2, _sum4);
-                        int32x4x2_t _t2 = vzipq_s32(_sum1, _sum7);
-                        int32x4x2_t _t3 = vzipq_s32(_sum3, _sum5);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t1.val[0]));
-                        _sum1 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t1.val[0]));
-                        _sum2 = vcombine_s32(vget_low_s32(_t1.val[1]), vget_low_s32(_t0.val[1]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t1.val[1]), vget_high_s32(_t0.val[1]));
-                        _sum4 = vcombine_s32(vget_low_s32(_t2.val[0]), vget_low_s32(_t3.val[0]));
-                        _sum5 = vcombine_s32(vget_high_s32(_t2.val[0]), vget_high_s32(_t3.val[0]));
-                        _sum6 = vcombine_s32(vget_low_s32(_t3.val[1]), vget_low_s32(_t2.val[1]));
-                        _sum7 = vcombine_s32(vget_high_s32(_t3.val[1]), vget_high_s32(_t2.val[1]));
-                        _sum1 = vrev64q_s32(_sum1);
-                        _sum3 = vrev64q_s32(_sum3);
-                        _sum5 = vrev64q_s32(_sum5);
-                        _sum7 = vrev64q_s32(_sum7);
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + 8, _sum2);
-                    vst1q_s32(outptr0 + 12, _sum3);
-                    vst1q_s32(outptr0 + 16, _sum4);
-                    vst1q_s32(outptr0 + 20, _sum5);
-                    vst1q_s32(outptr0 + 24, _sum6);
-                    vst1q_s32(outptr0 + 28, _sum7);
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-                    // to
-                    //      a0 a1 a2 a3
-                    //      a4 a5 a6 a7
-                    //      b0 b1 b2 b3
-                    //      b4 b5 b6 b7
-                    //      c0 c1 c2 c3
-                    //      c4 c5 c6 c7
-                    //      d0 d1 d2 d3
-                    //      d4 d5 d6 d7
-                    {
-                        _sum2 = vextq_s32(_sum2, _sum2, 2);
-                        _sum3 = vextq_s32(_sum3, _sum3, 2);
-                        _sum6 = vextq_s32(_sum6, _sum6, 2);
-                        _sum7 = vextq_s32(_sum7, _sum7, 2);
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sum6);
-                        int32x4x2_t _t1 = vzipq_s32(_sum2, _sum4);
-                        int32x4x2_t _t2 = vzipq_s32(_sum1, _sum7);
-                        int32x4x2_t _t3 = vzipq_s32(_sum3, _sum5);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t1.val[0]));
-                        _sum1 = vcombine_s32(vget_low_s32(_t2.val[0]), vget_low_s32(_t3.val[0]));
-                        _sum2 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t1.val[0]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t2.val[0]), vget_high_s32(_t3.val[0]));
-                        _sum4 = vcombine_s32(vget_low_s32(_t1.val[1]), vget_low_s32(_t0.val[1]));
-                        _sum5 = vcombine_s32(vget_low_s32(_t3.val[1]), vget_low_s32(_t2.val[1]));
-                        _sum6 = vcombine_s32(vget_high_s32(_t1.val[1]), vget_high_s32(_t0.val[1]));
-                        _sum7 = vcombine_s32(vget_high_s32(_t3.val[1]), vget_high_s32(_t2.val[1]));
-                        _sum2 = vrev64q_s32(_sum2);
-                        _sum3 = vrev64q_s32(_sum3);
-                        _sum6 = vrev64q_s32(_sum6);
-                        _sum7 = vrev64q_s32(_sum7);
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + out_hstep, _sum2);
-                    vst1q_s32(outptr0 + out_hstep + 4, _sum3);
-                    vst1q_s32(outptr0 + out_hstep * 2, _sum4);
-                    vst1q_s32(outptr0 + out_hstep * 2 + 4, _sum5);
-                    vst1q_s32(outptr0 + out_hstep * 3, _sum6);
-                    vst1q_s32(outptr0 + out_hstep * 3 + 4, _sum7);
-                    outptr0 += 8;
-                }
-#endif // __ARM_FEATURE_DOTPROD
-            }
-            else
-            {
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-                vst1q_s32(outptr + 16, _sum4);
-                vst1q_s32(outptr + 20, _sum5);
-                vst1q_s32(outptr + 24, _sum6);
-                vst1q_s32(outptr + 28, _sum7);
-            }
-
-            outptr += 32;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-#endif // __aarch64__
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            const signed char* pA = pAT;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-            asm volatile(
-                "cmp    %w9, #0                     \n"
-                "beq    0f                          \n"
-
-                "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0] \n"
-                "b      1f                          \n"
-
-                "0:                                 \n"
-                "eor    v16.16b, v16.16b, v16.16b   \n"
-                "eor    v17.16b, v17.16b, v17.16b   \n"
-                "eor    v18.16b, v18.16b, v18.16b   \n"
-                "eor    v19.16b, v19.16b, v19.16b   \n"
-
-                "1:                                 \n"
-#if __ARM_FEATURE_DOTPROD
-                "lsr    w4, %w8, #3                 \n" // w4 = max_kk >> 3
-                "cmp    w4, #0                      \n"
-                "beq    101f                        \n"
-
-#if __ARM_FEATURE_MATMUL_INT8
-                "eor    v24.16b, v24.16b, v24.16b   \n"
-                "eor    v25.16b, v25.16b, v25.16b   \n"
-                "eor    v26.16b, v26.16b, v26.16b   \n"
-                "eor    v27.16b, v27.16b, v27.16b   \n"
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                "2:                                 \n"
-                "ld1    {v0.16b, v1.16b}, [%1], #32 \n"
-                "ld1    {v4.16b, v5.16b}, [%2], #32 \n"
-
-#if __ARM_FEATURE_MATMUL_INT8
-                "smmla  v24.4s, v0.16b, v4.16b      \n"
-                "smmla  v25.4s, v1.16b, v4.16b      \n"
-                "subs   w4, w4, #1                  \n"
-                "smmla  v26.4s, v0.16b, v5.16b      \n"
-                "smmla  v27.4s, v1.16b, v5.16b      \n"
-#else  // __ARM_FEATURE_MATMUL_INT8
-                "sdot   v16.4s, v0.16b, v4.4b[0]    \n"
-                "sdot   v17.4s, v0.16b, v4.4b[1]    \n"
-                "sdot   v18.4s, v0.16b, v4.4b[2]    \n"
-                "sdot   v19.4s, v0.16b, v4.4b[3]    \n"
-                "subs   w4, w4, #1                  \n"
-                "sdot   v16.4s, v1.16b, v5.4b[0]    \n"
-                "sdot   v17.4s, v1.16b, v5.4b[1]    \n"
-                "sdot   v18.4s, v1.16b, v5.4b[2]    \n"
-                "sdot   v19.4s, v1.16b, v5.4b[3]    \n"
-#endif // __ARM_FEATURE_MATMUL_INT8
-                "bne    2b                          \n"
-
-#if __ARM_FEATURE_MATMUL_INT8
-                "uzp1   v0.4s, v24.4s, v25.4s       \n"
-                "uzp2   v1.4s, v24.4s, v25.4s       \n"
-                "uzp1   v2.4s, v26.4s, v27.4s       \n"
-                "uzp2   v3.4s, v26.4s, v27.4s       \n"
-
-                "add    v16.4s, v16.4s, v0.4s       \n"
-                "add    v17.4s, v17.4s, v1.4s       \n"
-                "add    v18.4s, v18.4s, v2.4s       \n"
-                "add    v19.4s, v19.4s, v3.4s       \n"
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                "101:                               \n"
-                "and    w4, %w8, #4                 \n" // w4 = remain = max_kk & 4
-                "cmp    w4, #0                      \n"
-                "beq    3f                          \n"
-
-                // kk += 4 part
-                "ld1    {v0.16b}, [%1], #16         \n"
-                "ld1    {v2.16b}, [%2], #16         \n"
-                "sdot   v16.4s, v0.16b, v2.4b[0]    \n"
-                "sdot   v17.4s, v0.16b, v2.4b[1]    \n"
-                "sdot   v18.4s, v0.16b, v2.4b[2]    \n"
-                "sdot   v19.4s, v0.16b, v2.4b[3]    \n"
-#else  // __ARM_FEATURE_DOTPROD
-                "lsr    w4, %w8, #2                 \n" // w4 = max_kk >> 2
-                "cmp    w4, #0                      \n"
-                "beq    3f                          \n"
-
-                "2:                                 \n"
-                "ld1    {v0.16b}, [%1], #16         \n"
-                "ld1    {v4.16b}, [%2], #16         \n"
-                "smull  v8.8h, v0.8b, v4.8b         \n"
-                "rev64  v1.4s, v0.4s                \n"
-                "smull  v9.8h, v1.8b, v4.8b         \n"
-                "rev64  v5.8h, v4.8h                \n"
-                "smull  v10.8h, v0.8b, v5.8b        \n"
-                "smull  v11.8h, v1.8b, v5.8b        \n"
-                "smlal2 v8.8h, v0.16b, v4.16b       \n"
-                "smlal2 v9.8h, v1.16b, v4.16b       \n"
-                "smlal2 v10.8h, v0.16b, v5.16b      \n"
-                "smlal2 v11.8h, v1.16b, v5.16b      \n"
-                "subs   w4, w4, #1                  \n"
-                "sadalp v16.4s, v8.8h               \n"
-                "sadalp v17.4s, v9.8h               \n"
-                "sadalp v18.4s, v10.8h              \n"
-                "sadalp v19.4s, v11.8h              \n"
-                "bne    2b                          \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "3:                                 \n"
-                "and    w4, %w8, #2                 \n" // w4 = remain = max_kk & 2
-                "cmp    w4, #0                      \n"
-                "beq    4f                          \n"
-
-                // kk += 2 part
-#if __ARM_FEATURE_DOTPROD
-                "ld1    {v0.8b}, [%1], #8           \n"
-                "ld1    {v1.8b}, [%2], #8           \n"
-                "dup    v4.4h, v1.h[0]              \n"
-                "dup    v5.4h, v1.h[1]              \n"
-                "dup    v6.4h, v1.h[2]              \n"
-                "dup    v7.4h, v1.h[3]              \n"
-                "smull  v8.8h, v0.8b, v4.8b         \n"
-                "smull  v9.8h, v0.8b, v5.8b         \n"
-                "smull  v10.8h, v0.8b, v6.8b        \n"
-                "smull  v11.8h, v0.8b, v7.8b        \n"
-                "sadalp v16.4s, v8.8h               \n"
-                "sadalp v17.4s, v9.8h               \n"
-                "sadalp v18.4s, v10.8h              \n"
-                "sadalp v19.4s, v11.8h              \n"
-#else  // __ARM_FEATURE_DOTPROD
-                "ld1    {v0.8b}, [%1], #8           \n"
-                "ld1    {v2.8b}, [%2], #8           \n"
-                "ext    v1.8b, v0.8b, v0.8b, #4     \n"
-                "rev64  v3.4h, v2.4h                \n"
-                "smull  v8.8h, v0.8b, v2.8b         \n"
-                "smull  v9.8h, v1.8b, v2.8b         \n"
-                "smull  v10.8h, v0.8b, v3.8b        \n"
-                "smull  v11.8h, v1.8b, v3.8b        \n"
-                "sadalp v16.4s, v8.8h               \n"
-                "sadalp v17.4s, v9.8h               \n"
-                "sadalp v18.4s, v10.8h              \n"
-                "sadalp v19.4s, v11.8h              \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "4:                                 \n"
-                "and    w4, %w8, #1                 \n" // w4 = remain = max_kk & 1
-                "cmp    w4, #0                      \n"
-                "beq    5f                          \n"
-
-                // kk += 1 part
-#if __ARM_FEATURE_DOTPROD
-                "ld1r   {v0.2s}, [%1]               \n"
-                "ld1r   {v1.2s}, [%2]               \n"
-                "add    %1, %1, #4                  \n"
-                "add    %2, %2, #4                  \n"
-                "zip1   v1.8b, v1.8b, v1.8b         \n"
-                "zip1   v2.4h, v1.4h, v1.4h         \n"
-                "zip2   v3.4h, v1.4h, v1.4h         \n"
-                "smull  v8.8h, v0.8b, v2.8b         \n"
-                "smull  v9.8h, v0.8b, v3.8b         \n"
-                "saddw  v16.4s, v16.4s, v8.4h       \n"
-                "saddw2 v17.4s, v17.4s, v8.8h       \n"
-                "saddw  v18.4s, v18.4s, v9.4h       \n"
-                "saddw2 v19.4s, v19.4s, v9.8h       \n"
-#else  // __ARM_FEATURE_DOTPROD
-                "ld1    {v0.8b}, [%1]               \n"
-                "ld1r   {v4.2s}, [%2]               \n"
-                "add    %1, %1, #4                  \n"
-                "add    %2, %2, #4                  \n"
-                "rev32  v1.4h, v0.4h                \n"
-                "zip1   v0.2s, v0.2s, v1.2s         \n"
-                "rev32  v5.8b, v4.8b                \n"
-                "smull  v8.8h, v0.8b, v4.8b         \n"
-                "smull  v9.8h, v0.8b, v5.8b         \n"
-                "saddw  v16.4s, v16.4s, v8.4h       \n"
-                "saddw2 v17.4s, v17.4s, v8.8h       \n"
-                "saddw  v18.4s, v18.4s, v9.4h       \n"
-                "saddw2 v19.4s, v19.4s, v9.8h       \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "5:                                 \n"
-                "cmp    %w10, #0                    \n"
-                "beq    10f                         \n"
-
-#if __ARM_FEATURE_DOTPROD
-                // from
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                //      a2 b2 c2 d2
-                //      a3 b3 c3 d3
-                // if out_elempack == 4
-                "cmp    %w11, #1                    \n"
-                "beq    8f                          \n"
-
-                "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-                // to
-                //      a0 a1 a2 a3
-                //      b0 b1 b2 b3
-                //      c0 c1 c2 c3
-                //      d0 d1 d2 d3
-                "zip1   v0.4s, v16.4s, v17.4s       \n"
-                "zip2   v1.4s, v16.4s, v17.4s       \n"
-                "zip1   v2.4s, v18.4s, v19.4s       \n"
-                "zip2   v3.4s, v18.4s, v19.4s       \n"
-                "zip1   v16.2d, v0.2d, v2.2d        \n"
-                "zip2   v17.2d, v0.2d, v2.2d        \n"
-                "zip1   v18.2d, v1.2d, v3.2d        \n"
-                "zip2   v19.2d, v1.2d, v3.2d        \n"
-
-                "add    x4, %3, %12, lsl #2         \n"
-                "st1    {v16.4s}, [%3], #16         \n"
-                "st1    {v17.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v18.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v19.4s}, [x4]              \n"
-#else  // __ARM_FEATURE_DOTPROD
-
-                // from
-                //      a0 b1 c2 d3
-                //      c0 d1 a2 b3
-                //      a3 b2 c1 d0
-                //      c3 d2 a1 b0
-                // if out_elempack == 4
-                "cmp    %w11, #1                    \n"
-                "beq    8f                          \n"
-
-                // to
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                //      a2 b2 c2 d2
-                //      a3 b3 c3 d3
-                "rev64  v18.4s, v18.4s              \n"
-                "rev64  v19.4s, v19.4s              \n"
-                "ext    v18.16b, v18.16b, v18.16b, #8 \n"
-                "ext    v19.16b, v19.16b, v19.16b, #8 \n"
-                "zip1   v0.4s, v16.4s, v19.4s       \n"
-                "zip2   v1.4s, v16.4s, v19.4s       \n"
-                "zip1   v2.4s, v17.4s, v18.4s       \n"
-                "zip2   v3.4s, v17.4s, v18.4s       \n"
-                "zip1   v16.2d, v0.2d, v2.2d        \n"
-                "zip2   v17.2d, v0.2d, v2.2d        \n"
-                "zip1   v18.2d, v3.2d, v1.2d        \n"
-                "zip2   v19.2d, v3.2d, v1.2d        \n"
-                "rev64  v17.4s, v17.4s              \n"
-                "rev64  v19.4s, v19.4s              \n"
-
-                "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%3], #64 \n"
-                "b      9f                          \n"
-
-                // if out_elempack == 1
-                "8:                                 \n"
-
-                // to
-                //      a0 a1 a2 a3
-                //      b0 b1 b2 b3
-                //      c0 c1 c2 c3
-                //      d0 d1 d2 d3
-                "ext    v17.16b, v17.16b, v17.16b, #8 \n"
-                "ext    v19.16b, v19.16b, v19.16b, #8 \n"
-                "zip1   v0.4s, v16.4s, v19.4s       \n"
-                "zip2   v1.4s, v16.4s, v19.4s       \n"
-                "zip1   v2.4s, v17.4s, v18.4s       \n"
-                "zip2   v3.4s, v17.4s, v18.4s       \n"
-                "zip1   v16.2d, v0.2d, v2.2d        \n"
-                "zip2   v17.2d, v0.2d, v2.2d        \n"
-                "zip1   v18.2d, v3.2d, v1.2d        \n"
-                "zip2   v19.2d, v3.2d, v1.2d        \n"
-                "rev64  v17.4s, v17.4s              \n"
-                "rev64  v19.4s, v19.4s              \n"
-
-                "add    x4, %3, %12, lsl #2         \n"
-                "st1    {v16.4s}, [%3], #16         \n"
-                "st1    {v17.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v18.4s}, [x4]              \n"
-                "add    x4, x4, %12, lsl #2         \n"
-                "st1    {v19.4s}, [x4]              \n"
-#endif // __ARM_FEATURE_DOTPROD
-
-                "9:                                 \n"
-                "add    %0, %0, #64                 \n"
-                "b      11f                         \n"
-
-                "10:                                \n"
-                "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-
-                "11:                                \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(max_kk),       // %8
-                "r"(k),            // %9
-                "r"(k_end),        // %10
-                "r"(out_elempack), // %11
-                "r"(out_hstep)     // %12
-                : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-            asm volatile(
-                "cmp        %9, #0              \n"
-                "beq        0f                  \n"
-
-                "vldm       %0, {d16-d23}       \n"
-                "b          1f                  \n"
-
-                "0:                             \n"
-                "veor       q8, q8              \n"
-                "veor       q9, q9              \n"
-                "veor       q10, q10            \n"
-                "veor       q11, q11            \n"
-
-                "1:                             \n"
-                "lsr        r4, %8, #2          \n" // r4 = max_kk >> 2
-                "cmp        r4, #0              \n"
-                "beq        3f                  \n"
-
-                ".align 4                       \n"
-                "2:                             \n"
-                "pld        [%1, #256]          \n"
-                "vld1.s8    {d0-d1}, [%1 :64]!  \n"
-                "pld        [%2, #128]          \n"
-                "vld1.s8    {d4-d5}, [%2]!      \n"
-                "vrev64.32  q1, q0              \n"
-                "vmull.s8   q4, d0, d4          \n"
-                "vrev64.16  q3, q2              \n"
-                "vmull.s8   q5, d2, d4          \n"
-                "vmull.s8   q6, d0, d6          \n"
-                "vmull.s8   q7, d2, d6          \n"
-                "vmlal.s8   q4, d1, d5          \n"
-                "vmlal.s8   q5, d3, d5          \n"
-                "vmlal.s8   q6, d1, d7          \n"
-                "vmlal.s8   q7, d3, d7          \n"
-                "subs       r4, r4, #1          \n"
-                "vpadal.s16 q8, q4              \n"
-                "vpadal.s16 q9, q5              \n"
-                "vpadal.s16 q10, q6             \n"
-                "vpadal.s16 q11, q7             \n"
-                "bne        2b                  \n"
-
-                "3:                             \n"
-                "and        r4, %8, #2          \n" // r4 = remain = max_kk & 2
-                "cmp        r4, #0              \n"
-                "beq        4f                  \n"
-
-                // kk += 2 part
-                "vld1.s8    {d0}, [%1 :64]!     \n"
-                "vld1.s8    {d4}, [%2]!         \n"
-                "vext.8     d1, d0, d0, #4      \n"
-                "vrev64.16  d5, d4              \n"
-                "vmull.s8   q4, d0, d4          \n"
-                "vmull.s8   q5, d1, d4          \n"
-                "vmull.s8   q6, d0, d5          \n"
-                "vmull.s8   q7, d1, d5          \n"
-                "vpadal.s16 q8, q4              \n"
-                "vpadal.s16 q9, q5              \n"
-                "vpadal.s16 q10, q6             \n"
-                "vpadal.s16 q11, q7             \n"
-
-                "4:                             \n"
-                "and        r4, %8, #1          \n" // r4 = remain = max_kk & 1
-                "cmp        r4, #0              \n"
-                "beq        5f                  \n"
-
-                // kk += 1 part
-                "vld1.s32   {d0[0]}, [%1]!      \n"
-                "vld1.s32   {d2[]}, [%2]!       \n"
-                "vrev32.16  d1, d0              \n"
-                "vrev32.s8  d3, d2              \n"
-                "vzip.32    d0, d1              \n"
-                "vmull.s8   q4, d0, d2          \n"
-                "vmull.s8   q5, d0, d3          \n"
-                "vaddw.s16  q8, d8              \n"
-                "vaddw.s16  q9, d9              \n"
-                "vaddw.s16  q10, d10            \n"
-                "vaddw.s16  q11, d11            \n"
-
-                "5:                             \n"
-                "cmp        %10, #0             \n"
-                "beq        10f                 \n"
-
-                // from
-                //      a0 b1 c2 d3
-                //      c0 d1 a2 b3
-                //      a3 b2 c1 d0
-                //      c3 d2 a1 b0
-                // if out_elempack == 4
-                "cmp        %11, #1             \n"
-                "beq        8f                  \n"
-
-                // to
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                //      a2 b2 c2 d2
-                //      a3 b3 c3 d3
-                "vrev64.32  q10, q10            \n"
-                "vrev64.32  q11, q11            \n"
-                "vext.32    q10, q10, #2        \n"
-                "vext.32    q11, q11, #2        \n"
-                "vzip.32    q8, q11             \n"
-                "vzip.32    q9, q10             \n"
-                "vswp       d17, d18            \n"
-                "vswp       d21, d22            \n"
-                "vrev64.32  q9, q9              \n"
-                "vrev64.32  q11, q11            \n"
-
-                "vstm       %3!, {d16-d23}      \n"
-                "b          9f                  \n"
-
-                // if out_elempack == 1
-                "8:                             \n"
-                // to
-                //      a0 a1 a2 a3
-                //      b0 b1 b2 b3
-                //      c0 c1 c2 c3
-                //      d0 d1 d2 d3
-                "vext.32    q9, q9, #2          \n"
-                "vext.32    q11, q11, #2        \n"
-                "vzip.32    q8, q11             \n"
-                "vzip.32    q9, q10             \n"
-                "vswp       d17, d18            \n"
-                "vswp       d21, d22            \n"
-                "vrev64.32  q9, q9              \n"
-                "vrev64.32  q11, q11            \n"
-
-                "add        r4, %3, %12, lsl #2 \n"
-                "vst1.s32   {d16-d17}, [%3]!    \n"
-                "vst1.s32   {d18-d19}, [r4]     \n"
-                "add        r4, r4, %12, lsl #2 \n"
-                "vst1.s32   {d20-d21}, [r4]     \n"
-                "add        r4, r4, %12, lsl #2 \n"
-                "vst1.s32   {d22-d23}, [r4]     \n"
-
-                "9:                             \n"
-                "add        %0, %0, #64         \n"
-                "b          11f                 \n"
-
-                "10:                            \n"
-                "vstm       %0!, {d16-d23}      \n"
-
-                "11:                            \n"
-
-                : "=r"(outptr), // %0
-                "=r"(pA),     // %1
-                "=r"(pB),     // %2
-                "=r"(outptr0) // %3
-                : "0"(outptr),
-                "1"(pA),
-                "2"(pB),
-                "3"(outptr0),
-                "r"(max_kk),       // %8
-                "r"(k),            // %9
-                "r"(k_end),        // %10
-                "r"(out_elempack), // %11
-                "r"(out_hstep)     // %12
-                : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-            int32x4_t _sum0;
-            int32x4_t _sum1;
-            int32x4_t _sum2;
-            int32x4_t _sum3;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_s32(0);
-                _sum1 = vdupq_n_s32(0);
-                _sum2 = vdupq_n_s32(0);
-                _sum3 = vdupq_n_s32(0);
-            }
-            else
-            {
-                _sum0 = vld1q_s32(outptr);
-                _sum1 = vld1q_s32(outptr + 4);
-                _sum2 = vld1q_s32(outptr + 8);
-                _sum3 = vld1q_s32(outptr + 12);
-            }
-
-            int kk = 0;
-#if __ARM_FEATURE_MATMUL_INT8
-            {
-                int32x4_t _sum00 = vdupq_n_s32(0);
-                int32x4_t _sum01 = vdupq_n_s32(0);
-                int32x4_t _sum10 = vdupq_n_s32(0);
-                int32x4_t _sum11 = vdupq_n_s32(0);
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int8x16_t _pA0 = vld1q_s8(pA);
-                    int8x16_t _pA1 = vld1q_s8(pA + 16);
-                    int8x16_t _pB0 = vld1q_s8(pB);
-                    int8x16_t _pB1 = vld1q_s8(pB + 16);
-
-                    // aaaaaaaa bbbbbbbb cccccccc dddddddd
-
-                    // 00000000 11111111 22222222 33333333
-
-                    _sum00 = vmmlaq_s32(_sum00, _pA0, _pB0);
-                    _sum01 = vmmlaq_s32(_sum01, _pA1, _pB0);
-                    _sum10 = vmmlaq_s32(_sum10, _pA0, _pB1);
-                    _sum11 = vmmlaq_s32(_sum11, _pA1, _pB1);
-
-                    // a0 a1 b0 b1
-                    // c0 c1 d0 d1
-                    // a2 a3 b2 b3
-                    // c2 c3 d2 d3
-
-                    pA += 32;
-                    pB += 32;
-                }
-                int32x4x2_t _ss0 = vuzpq_s32(_sum00, _sum01);
-                int32x4x2_t _ss1 = vuzpq_s32(_sum10, _sum11);
-                _sum0 = vaddq_s32(_sum0, _ss0.val[0]);
-                _sum1 = vaddq_s32(_sum1, _ss0.val[1]);
-                _sum2 = vaddq_s32(_sum2, _ss1.val[0]);
-                _sum3 = vaddq_s32(_sum3, _ss1.val[1]);
-            }
-#elif __ARM_FEATURE_DOTPROD
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                int8x16_t _pA0 = vld1q_s8(pA);
-                int8x16_t _pA1 = vld1q_s8(pA + 16);
-                int8x16_t _pB0 = vld1q_s8(pB);
-                int8x16_t _pB1 = vld1q_s8(pB + 16);
-
-                _sum0 = vdotq_laneq_s32(_sum0, _pA0, _pB0, 0);
-                _sum1 = vdotq_laneq_s32(_sum1, _pA0, _pB0, 1);
-                _sum2 = vdotq_laneq_s32(_sum2, _pA0, _pB0, 2);
-                _sum3 = vdotq_laneq_s32(_sum3, _pA0, _pB0, 3);
-
-                _sum0 = vdotq_laneq_s32(_sum0, _pA1, _pB1, 0);
-                _sum1 = vdotq_laneq_s32(_sum1, _pA1, _pB1, 1);
-                _sum2 = vdotq_laneq_s32(_sum2, _pA1, _pB1, 2);
-                _sum3 = vdotq_laneq_s32(_sum3, _pA1, _pB1, 3);
-
-                pA += 32;
-                pB += 32;
-            }
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x16_t _pA = vld1q_s8(pA);
-                int8x16_t _pB = vld1q_s8(pB);
-
-                _sum0 = vdotq_laneq_s32(_sum0, _pA, _pB, 0);
-                _sum1 = vdotq_laneq_s32(_sum1, _pA, _pB, 1);
-                _sum2 = vdotq_laneq_s32(_sum2, _pA, _pB, 2);
-                _sum3 = vdotq_laneq_s32(_sum3, _pA, _pB, 3);
-#else  // __ARM_FEATURE_DOTPROD
-                int8x16_t _pA02 = vld1q_s8(pA);
-                int8x16_t _pB02 = vld1q_s8(pB);
-
-                // aabbccdd eeffgghh
-                // ccddaabb gghheeff
-
-                int8x16_t _pA13 = vreinterpretq_s8_s32(vrev64q_s32(vreinterpretq_s32_s8(_pA02)));
-
-                // 00112233 44556677
-                // 33221100 77665544
-
-                int8x16_t _pB13 = vreinterpretq_s8_s16(vrev64q_s16(vreinterpretq_s16_s8(_pB02)));
-
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA02), vget_low_s8(_pB02));
-                int16x8_t _s1 = vmull_s8(vget_low_s8(_pA13), vget_low_s8(_pB02));
-                int16x8_t _s2 = vmull_s8(vget_low_s8(_pA02), vget_low_s8(_pB13));
-                int16x8_t _s3 = vmull_s8(vget_low_s8(_pA13), vget_low_s8(_pB13));
-
-                _s0 = vmlal_s8(_s0, vget_high_s8(_pA02), vget_high_s8(_pB02));
-                _s1 = vmlal_s8(_s1, vget_high_s8(_pA13), vget_high_s8(_pB02));
-                _s2 = vmlal_s8(_s2, vget_high_s8(_pA02), vget_high_s8(_pB13));
-                _s3 = vmlal_s8(_s3, vget_high_s8(_pA13), vget_high_s8(_pB13));
-
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 16;
-                pB += 16;
-            }
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x8_t _pA = vld1_s8(pA);
-                int8x8_t _pB = vld1_s8(pB);
-
-                int16x8_t _s0 = vmull_s8(_pA, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 0)));
-                int16x8_t _s1 = vmull_s8(_pA, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 1)));
-                int16x8_t _s2 = vmull_s8(_pA, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 2)));
-                int16x8_t _s3 = vmull_s8(_pA, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 3)));
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-#else  // __ARM_FEATURE_DOTPROD
-                int8x8_t _pA0 = vld1_s8(pA);
-                int8x8_t _pB0 = vld1_s8(pB);
-
-                // aabbccdd
-                // ccddaabb
-
-                int8x8_t _pA1 = vext_s8(_pA0, _pA0, 4);
-
-                // 00112233
-                // 33221100
-
-                int8x8_t _pB1 = vreinterpret_s8_s16(vrev64_s16(vreinterpret_s16_s8(_pB0)));
-
-                int16x8_t _s0 = vmull_s8(_pA0, _pB0);
-                int16x8_t _s1 = vmull_s8(_pA1, _pB0);
-                int16x8_t _s2 = vmull_s8(_pA0, _pB1);
-                int16x8_t _s3 = vmull_s8(_pA1, _pB1);
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 8;
-                pB += 8;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x8_t _pA = vreinterpret_s8_s32(vld1_dup_s32((const int*)pA));
-                int8x8_t _pB = vreinterpret_s8_s32(vld1_dup_s32((const int*)pB));
-
-                _pB = vzip_s8(_pB, _pB).val[0];
-                int16x4x2_t _pB0123 = vzip_s16(vreinterpret_s16_s8(_pB), vreinterpret_s16_s8(_pB));
-
-                int16x8_t _s01 = vmull_s8(_pA, vreinterpret_s8_s16(_pB0123.val[0]));
-                int16x8_t _s23 = vmull_s8(_pA, vreinterpret_s8_s16(_pB0123.val[1]));
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s01));
-                _sum1 = vaddw_s16(_sum1, vget_high_s16(_s01));
-                _sum2 = vaddw_s16(_sum2, vget_low_s16(_s23));
-                _sum3 = vaddw_s16(_sum3, vget_high_s16(_s23));
-#else  // __ARM_FEATURE_DOTPROD
-
-                int8x8_t _pA0 = vld1_s8(pA);
-                int8x8_t _pB0 = vreinterpret_s8_s32(vld1_dup_s32((const int*)pB));
-
-                // abcd.... -> cdab.... -> abcdcdab
-                int8x8_t _pA1 = vreinterpret_s8_s16(vrev32_s16(vreinterpret_s16_s8(_pA0)));
-                int8x8_t _pA01 = vreinterpret_s8_s32(vzip_s32(vreinterpret_s32_s8(_pA0), vreinterpret_s32_s8(_pA1)).val[0]);
-
-                // 01230123 -> 32103210
-                int8x8_t _pB1 = vrev32_s8(_pB0);
-
-                int16x8_t _s01 = vmull_s8(_pA01, _pB0);
-                int16x8_t _s23 = vmull_s8(_pA01, _pB1);
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s01));
-                _sum1 = vaddw_s16(_sum1, vget_high_s16(_s01));
-                _sum2 = vaddw_s16(_sum2, vget_low_s16(_s23));
-                _sum3 = vaddw_s16(_sum3, vget_high_s16(_s23));
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 4;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-#if __ARM_FEATURE_DOTPROD
-                // from
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                //      a2 b2 c2 d2
-                //      a3 b3 c3 d3
-                if (out_elempack == 4)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + 8, _sum2);
-                    vst1q_s32(outptr0 + 12, _sum3);
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    // to
-                    //      a0 a1 a2 a3
-                    //      b0 b1 b2 b3
-                    //      c0 c1 c2 c3
-                    //      d0 d1 d2 d3
-                    {
-                        int32x4x2_t _r01 = vzipq_s32(_sum0, _sum1);
-                        int32x4x2_t _r23 = vzipq_s32(_sum2, _sum3);
-                        _sum0 = vcombine_s32(vget_low_s32(_r01.val[0]), vget_low_s32(_r23.val[0]));
-                        _sum1 = vcombine_s32(vget_high_s32(_r01.val[0]), vget_high_s32(_r23.val[0]));
-                        _sum2 = vcombine_s32(vget_low_s32(_r01.val[1]), vget_low_s32(_r23.val[1]));
-                        _sum3 = vcombine_s32(vget_high_s32(_r01.val[1]), vget_high_s32(_r23.val[1]));
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + out_hstep, _sum1);
-                    vst1q_s32(outptr0 + out_hstep * 2, _sum2);
-                    vst1q_s32(outptr0 + out_hstep * 3, _sum3);
-                    outptr0 += 4;
-                }
-#else  // __ARM_FEATURE_DOTPROD
-
-                // from
-                //      a0 b1 c2 d3
-                //      c0 d1 a2 b3
-                //      a3 b2 c1 d0
-                //      c3 d2 a1 b0
-                if (out_elempack == 4)
-                {
-                    // to
-                    //      a0 b0 c0 d0
-                    //      a1 b1 c1 d1
-                    //      a2 b2 c2 d2
-                    //      a3 b3 c3 d3
-                    {
-                        _sum2 = vrev64q_s32(_sum2);
-                        _sum3 = vrev64q_s32(_sum3);
-                        _sum2 = vextq_s32(_sum2, _sum2, 2);
-                        _sum3 = vextq_s32(_sum3, _sum3, 2);
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sum3);
-                        int32x4x2_t _t1 = vzipq_s32(_sum1, _sum2);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t1.val[0]));
-                        _sum1 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t1.val[0]));
-                        _sum2 = vcombine_s32(vget_low_s32(_t1.val[1]), vget_low_s32(_t0.val[1]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t1.val[1]), vget_high_s32(_t0.val[1]));
-                        _sum1 = vrev64q_s32(_sum1);
-                        _sum3 = vrev64q_s32(_sum3);
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + 8, _sum2);
-                    vst1q_s32(outptr0 + 12, _sum3);
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    // to
-                    //      a0 a1 a2 a3
-                    //      b0 b1 b2 b3
-                    //      c0 c1 c2 c3
-                    //      d0 d1 d2 d3
-                    {
-                        _sum1 = vextq_s32(_sum1, _sum1, 2);
-                        _sum3 = vextq_s32(_sum3, _sum3, 2);
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sum3);
-                        int32x4x2_t _t1 = vzipq_s32(_sum1, _sum2);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t1.val[0]));
-                        _sum1 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t1.val[0]));
-                        _sum2 = vcombine_s32(vget_low_s32(_t1.val[1]), vget_low_s32(_t0.val[1]));
-                        _sum3 = vcombine_s32(vget_high_s32(_t1.val[1]), vget_high_s32(_t0.val[1]));
-                        _sum1 = vrev64q_s32(_sum1);
-                        _sum3 = vrev64q_s32(_sum3);
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + out_hstep, _sum1);
-                    vst1q_s32(outptr0 + out_hstep * 2, _sum2);
-                    vst1q_s32(outptr0 + out_hstep * 3, _sum3);
-                    outptr0 += 4;
-                }
-#endif // __ARM_FEATURE_DOTPROD
-            }
-            else
-            {
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-            }
-
-            outptr += 16;
-#endif // NCNN_GNU_INLINE_ASM
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            const signed char* pA = pAT;
-
-            int32x4_t _sum0;
-            int32x4_t _sum1;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_s32(0);
-                _sum1 = vdupq_n_s32(0);
-            }
-            else
-            {
-                _sum0 = vld1q_s32(outptr);
-                _sum1 = vld1q_s32(outptr + 4);
-            }
-
-            int kk = 0;
-#if __ARM_FEATURE_DOTPROD
-            {
-#if __ARM_FEATURE_MATMUL_INT8
-                int32x4_t _sum00 = vdupq_n_s32(0);
-                int32x4_t _sum01 = vdupq_n_s32(0);
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int8x16_t _pA0 = vld1q_s8(pA);
-                    int8x16_t _pA1 = vld1q_s8(pA + 16);
-                    int8x16_t _pB = vld1q_s8(pB);
-
-#if __ARM_FEATURE_MATMUL_INT8
-                    // aaaaaaaa bbbbbbbb cccccccc dddddddd
-
-                    // 00000000 11111111
-
-                    _sum00 = vmmlaq_s32(_sum00, _pA0, _pB);
-                    _sum01 = vmmlaq_s32(_sum01, _pA1, _pB);
-#else  // __ARM_FEATURE_MATMUL_INT8
-                    _sum0 = vdotq_laneq_s32(_sum0, _pA0, _pB, 0);
-                    _sum1 = vdotq_laneq_s32(_sum1, _pA0, _pB, 1);
-                    _sum0 = vdotq_laneq_s32(_sum0, _pA1, _pB, 2);
-                    _sum1 = vdotq_laneq_s32(_sum1, _pA1, _pB, 3);
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                    pA += 32;
-                    pB += 16;
-                }
-#if __ARM_FEATURE_MATMUL_INT8
-                int32x4x2_t _ss = vuzpq_s32(_sum00, _sum01);
-                _sum0 = vaddq_s32(_sum0, _ss.val[0]);
-                _sum1 = vaddq_s32(_sum1, _ss.val[1]);
-#endif // __ARM_FEATURE_MATMUL_INT8
-            }
-#endif // __ARM_FEATURE_DOTPROD
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x16_t _pA = vld1q_s8(pA);
-                int8x8_t _pB = vld1_s8(pB);
-
-                _sum0 = vdotq_lane_s32(_sum0, _pA, _pB, 0);
-                _sum1 = vdotq_lane_s32(_sum1, _pA, _pB, 1);
-#else  // __ARM_FEATURE_DOTPROD
-                int8x16_t _pA = vld1q_s8(pA);
-                int8x8_t _pB = vld1_s8(pB);
-
-                // aabbccdd eeffgghh
-
-                // 00112233 -> 00110011 22332233
-                // 11001100 33223322
-
-                int32x2x2_t _pBB = vzip_s32(vreinterpret_s32_s8(_pB), vreinterpret_s32_s8(_pB));
-                int8x16_t _pB02 = vreinterpretq_s8_s32(vcombine_s32(_pBB.val[0], _pBB.val[1]));
-
-                int8x16_t _pB13 = vreinterpretq_s8_s16(vrev64q_s16(vreinterpretq_s16_s8(_pB02)));
-
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA), vget_low_s8(_pB02));
-                int16x8_t _s1 = vmull_s8(vget_low_s8(_pA), vget_low_s8(_pB13));
-                _s0 = vmlal_s8(_s0, vget_high_s8(_pA), vget_high_s8(_pB02));
-                _s1 = vmlal_s8(_s1, vget_high_s8(_pA), vget_high_s8(_pB13));
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 16;
-                pB += 8;
-            }
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x8_t _pA = vld1_s8(pA);
-                int8x8_t _pB = vld1_s8(pB);
-                // aabbccdd
-                // 0011....
-                int16x8_t _s0 = vmull_s8(_pA, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 0)));
-                int16x8_t _s1 = vmull_s8(_pA, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pB), 1)));
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-#else  // __ARM_FEATURE_DOTPROD
-                int8x8_t _pA = vld1_s8(pA);
-                int8x8_t _pB0 = vreinterpret_s8_s32(vld1_dup_s32((const int*)pB));
-
-                // aabbccdd
-
-                // 00110011
-                // 11001100
-                int8x8_t _pB1 = vext_s8(_pB0, _pB0, 2);
-
-                int16x8_t _s0 = vmull_s8(_pA, _pB0);
-                int16x8_t _s1 = vmull_s8(_pA, _pB1);
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 8;
-                pB += 4;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x8_t _pA = vreinterpret_s8_s32(vld1_dup_s32((const int*)pA));
-                int8x8_t _pB = vreinterpret_s8_s16(vld1_dup_s16((const short*)pB));
-
-                // abcdabcd
-
-                // 01010101 -> 00001111
-                _pB = vuzp_s8(_pB, vext_s8(_pB, _pB, 1)).val[0];
-
-                int16x8_t _s0 = vmull_s8(_pA, _pB);
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s0));
-                _sum1 = vaddw_s16(_sum1, vget_high_s16(_s0));
-#else  // __ARM_FEATURE_DOTPROD
-                int8x8_t _pA = vreinterpret_s8_s32(vld1_dup_s32((const int*)pA));
-                int8x8_t _pB0 = vreinterpret_s8_s16(vld1_dup_s16((const short*)pB));
-
-                // abcd abcd
-
-                // 0101 0101 -> 0101 1010
-
-                int8x8_t _pB1 = vext_s8(_pB0, _pB0, 1);
-                int8x8_t _pB = vreinterpret_s8_s32(vzip_s32(vreinterpret_s32_s8(_pB0), vreinterpret_s32_s8(_pB1)).val[0]);
-
-                int16x8_t _s0 = vmull_s8(_pA, _pB);
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s0));
-                _sum1 = vaddw_s16(_sum1, vget_high_s16(_s0));
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 4;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-#if __ARM_FEATURE_DOTPROD
-                // from
-                //      a0 b0 c0 d0
-                //      a1 b1 c1 d1
-                if (out_elempack == 4)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-                    // to
-                    //      a0 a1 b0 b1
-                    //      c0 c1 d0 d1
-                    {
-                        int32x4x2_t _sum01 = vzipq_s32(_sum0, _sum1);
-                        _sum0 = _sum01.val[0];
-                        _sum1 = _sum01.val[1];
-                    }
-
-                    vst1_s32(outptr0, vget_low_s32(_sum0));
-                    vst1_s32(outptr0 + out_hstep, vget_high_s32(_sum0));
-                    vst1_s32(outptr0 + out_hstep * 2, vget_low_s32(_sum1));
-                    vst1_s32(outptr0 + out_hstep * 3, vget_high_s32(_sum1));
-                    outptr0 += 2;
-                }
-#else  // __ARM_FEATURE_DOTPROD
-
-                // from
-                //      a0 b1 c0 d1
-                //      a1 b0 c1 d0
-                if (out_elempack == 4)
-                {
-                    // to
-                    //      a0 b0 c0 d0
-                    //      a1 b1 c1 d1
-                    {
-                        _sum1 = vrev64q_s32(_sum1);
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sum1);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t0.val[1]));
-                        _sum1 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t0.val[1]));
-                        _sum1 = vrev64q_s32(_sum1);
-                    }
-
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-                    // to
-                    //      a0 a1 c0 c1
-                    //      b0 b1 d0 d1
-                    {
-                        int32x4x2_t _t0 = vzipq_s32(_sum0, _sum1);
-                        _sum0 = vcombine_s32(vget_low_s32(_t0.val[0]), vget_low_s32(_t0.val[1]));
-                        _sum1 = vcombine_s32(vget_high_s32(_t0.val[0]), vget_high_s32(_t0.val[1]));
-                        _sum1 = vrev64q_s32(_sum1);
-                    }
-
-                    vst1_s32(outptr0, vget_low_s32(_sum0));
-                    vst1_s32(outptr0 + out_hstep, vget_low_s32(_sum1));
-                    vst1_s32(outptr0 + out_hstep * 2, vget_high_s32(_sum0));
-                    vst1_s32(outptr0 + out_hstep * 3, vget_high_s32(_sum1));
-                    outptr0 += 2;
-                }
-#endif // __ARM_FEATURE_DOTPROD
-            }
-            else
-            {
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-            }
-
-            outptr += 8;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            const signed char* pA = pAT;
-
-            int32x4_t _sum0;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_s32(0);
-            }
-            else
-            {
-                _sum0 = vld1q_s32(outptr);
-            }
-
-            int kk = 0;
-#if __ARM_FEATURE_DOTPROD
-            {
-#if __ARM_FEATURE_MATMUL_INT8
-                int32x4_t _sum01 = vdupq_n_s32(0);
-                int32x4_t _sum23 = vdupq_n_s32(0);
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int8x16_t _pA0 = vld1q_s8(pA);
-                    int8x16_t _pA1 = vld1q_s8(pA + 16);
-                    int8x8_t _pB = vld1_s8(pB);
-
-#if __ARM_FEATURE_MATMUL_INT8
-                    // aaaaaaaa bbbbbbbb cccccccc dddddddd
-
-                    // 00000000
-
-                    int8x16_t _pBB = vcombine_s8(_pB, _pB);
-
-                    _sum01 = vdotq_s32(_sum01, _pA0, _pBB);
-                    _sum23 = vdotq_s32(_sum23, _pA1, _pBB);
-#else  // __ARM_FEATURE_MATMUL_INT8
-                    _sum0 = vdotq_lane_s32(_sum0, _pA0, _pB, 0);
-                    _sum0 = vdotq_lane_s32(_sum0, _pA1, _pB, 1);
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                    pA += 32;
-                    pB += 8;
-                }
-#if __ARM_FEATURE_MATMUL_INT8
-                _sum0 = vaddq_s32(_sum0, vpaddq_s32(_sum01, _sum23));
-#endif // __ARM_FEATURE_MATMUL_INT8
-            }
-#endif // __ARM_FEATURE_DOTPROD
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-#if __ARM_FEATURE_DOTPROD
-                int8x16_t _pA = vld1q_s8(pA);
-                int8x8_t _pB = vld1_s8(pB);
-
-                _sum0 = vdotq_lane_s32(_sum0, _pA, _pB, 0);
-#else  // __ARM_FEATURE_DOTPROD
-                int8x16_t _pA = vld1q_s8(pA);
-                int8x8_t _pB0 = vreinterpret_s8_s16(vld1_dup_s16((const short*)pB));
-                int8x8_t _pB1 = vreinterpret_s8_s16(vld1_dup_s16((const short*)(pB + 2)));
-
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA), _pB0);
-                _s0 = vmlal_s8(_s0, vget_high_s8(_pA), _pB1);
-                _sum0 = vpadalq_s16(_sum0, _s0);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 16;
-                pB += 4;
-            }
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                int8x8_t _pA = vld1_s8(pA);
-                int8x8_t _pB = vreinterpret_s8_s16(vld1_dup_s16((const short*)pB));
-
-                int16x8_t _s0 = vmull_s8(_pA, _pB);
-                _sum0 = vpadalq_s16(_sum0, _s0);
-
-                pA += 8;
-                pB += 2;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                int8x8_t _pA = vreinterpret_s8_s32(vld1_dup_s32((const int*)pA));
-                int8x8_t _pB = vld1_dup_s8(pB);
-
-                int16x8_t _s0 = vmull_s8(_pA, _pB);
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s0));
-
-                pA += 4;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    outptr0 += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    outptr0[0] = vgetq_lane_s32(_sum0, 0);
-                    outptr0[out_hstep] = vgetq_lane_s32(_sum0, 1);
-                    outptr0[out_hstep * 2] = vgetq_lane_s32(_sum0, 2);
-                    outptr0[out_hstep * 3] = vgetq_lane_s32(_sum0, 3);
-                    outptr0++;
-                }
-            }
-            else
-            {
-                vst1q_s32(outptr, _sum0);
-            }
-
-            outptr += 4;
-        }
-
-        pAT += max_kk * 4;
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        int* outptr0 = (int*)top_blob + (i + ii) * out_hstep + j;
-
-        const signed char* pB = pBT;
-
-        int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            int32x4_t _sum0;
-            int32x4_t _sum1;
-            int32x4_t _sum2;
-            int32x4_t _sum3;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_s32(0);
-                _sum1 = vdupq_n_s32(0);
-                _sum2 = vdupq_n_s32(0);
-                _sum3 = vdupq_n_s32(0);
-            }
-            else
-            {
-                _sum0 = vld1q_s32(outptr);
-                _sum1 = vld1q_s32(outptr + 4);
-                _sum2 = vld1q_s32(outptr + 8);
-                _sum3 = vld1q_s32(outptr + 12);
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-#if __ARM_FEATURE_DOTPROD
-            {
-#if __ARM_FEATURE_MATMUL_INT8
-                int32x4_t _sum01 = vdupq_n_s32(0);
-                int32x4_t _sum23 = vdupq_n_s32(0);
-                int32x4_t _sum45 = vdupq_n_s32(0);
-                int32x4_t _sum67 = vdupq_n_s32(0);
-#else  // __ARM_FEATURE_MATMUL_INT8
-                int32x2_t _sum00 = vdup_n_s32(0);
-                int32x2_t _sum01 = vdup_n_s32(0);
-                int32x2_t _sum10 = vdup_n_s32(0);
-                int32x2_t _sum11 = vdup_n_s32(0);
-                int32x2_t _sum20 = vdup_n_s32(0);
-                int32x2_t _sum21 = vdup_n_s32(0);
-                int32x2_t _sum30 = vdup_n_s32(0);
-                int32x2_t _sum31 = vdup_n_s32(0);
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int8x16_t _pA = vld1q_s8(pA);
-                    int8x16_t _pB0 = vld1q_s8(pB);
-                    int8x16_t _pB1 = vld1q_s8(pB + 16);
-                    int8x16_t _pB2 = vld1q_s8(pB + 32);
-                    int8x16_t _pB3 = vld1q_s8(pB + 48);
-
-#if __ARM_FEATURE_MATMUL_INT8
-                    _sum01 = vmmlaq_s32(_sum01, _pA, _pB0);
-                    _sum23 = vmmlaq_s32(_sum23, _pA, _pB1);
-                    _sum45 = vmmlaq_s32(_sum45, _pA, _pB2);
-                    _sum67 = vmmlaq_s32(_sum67, _pA, _pB3);
-#else  // __ARM_FEATURE_MATMUL_INT8
-                    _sum00 = vdot_laneq_s32(_sum00, vget_low_s8(_pA), _pB0, 0);
-                    _sum01 = vdot_laneq_s32(_sum01, vget_low_s8(_pA), _pB0, 1);
-                    _sum10 = vdot_laneq_s32(_sum10, vget_low_s8(_pA), _pB0, 2);
-                    _sum11 = vdot_laneq_s32(_sum11, vget_low_s8(_pA), _pB0, 3);
-                    _sum20 = vdot_laneq_s32(_sum20, vget_low_s8(_pA), _pB1, 0);
-                    _sum21 = vdot_laneq_s32(_sum21, vget_low_s8(_pA), _pB1, 1);
-                    _sum30 = vdot_laneq_s32(_sum30, vget_low_s8(_pA), _pB1, 2);
-                    _sum31 = vdot_laneq_s32(_sum31, vget_low_s8(_pA), _pB1, 3);
-                    _sum00 = vdot_laneq_s32(_sum00, vget_high_s8(_pA), _pB2, 0);
-                    _sum01 = vdot_laneq_s32(_sum01, vget_high_s8(_pA), _pB2, 1);
-                    _sum10 = vdot_laneq_s32(_sum10, vget_high_s8(_pA), _pB2, 2);
-                    _sum11 = vdot_laneq_s32(_sum11, vget_high_s8(_pA), _pB2, 3);
-                    _sum20 = vdot_laneq_s32(_sum20, vget_high_s8(_pA), _pB3, 0);
-                    _sum21 = vdot_laneq_s32(_sum21, vget_high_s8(_pA), _pB3, 1);
-                    _sum30 = vdot_laneq_s32(_sum30, vget_high_s8(_pA), _pB3, 2);
-                    _sum31 = vdot_laneq_s32(_sum31, vget_high_s8(_pA), _pB3, 3);
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                    pA += 16;
-                    pB += 64;
-                }
-#if __ARM_FEATURE_MATMUL_INT8
-                _sum0 = vaddq_s32(_sum0, vcombine_s32(vget_low_s32(_sum01), vget_low_s32(_sum23)));
-                _sum1 = vaddq_s32(_sum1, vcombine_s32(vget_low_s32(_sum45), vget_low_s32(_sum67)));
-                _sum2 = vaddq_s32(_sum2, vcombine_s32(vget_high_s32(_sum01), vget_high_s32(_sum23)));
-                _sum3 = vaddq_s32(_sum3, vcombine_s32(vget_high_s32(_sum45), vget_high_s32(_sum67)));
-#else  // __ARM_FEATURE_MATMUL_INT8
-                int32x2x2_t _sum0x = vzip_s32(_sum00, _sum01);
-                int32x2x2_t _sum1x = vzip_s32(_sum10, _sum11);
-                int32x2x2_t _sum2x = vzip_s32(_sum20, _sum21);
-                int32x2x2_t _sum3x = vzip_s32(_sum30, _sum31);
-                _sum0 = vaddq_s32(_sum0, vcombine_s32(_sum0x.val[0], _sum1x.val[0]));
-                _sum1 = vaddq_s32(_sum1, vcombine_s32(_sum2x.val[0], _sum3x.val[0]));
-                _sum2 = vaddq_s32(_sum2, vcombine_s32(_sum0x.val[1], _sum1x.val[1]));
-                _sum3 = vaddq_s32(_sum3, vcombine_s32(_sum2x.val[1], _sum3x.val[1]));
-#endif // __ARM_FEATURE_MATMUL_INT8
-            }
-#endif // __ARM_FEATURE_DOTPROD
-            {
-#if __ARM_FEATURE_DOTPROD
-                int32x2_t _sum00 = vdup_n_s32(0);
-                int32x2_t _sum01 = vdup_n_s32(0);
-                int32x2_t _sum10 = vdup_n_s32(0);
-                int32x2_t _sum11 = vdup_n_s32(0);
-                int32x2_t _sum20 = vdup_n_s32(0);
-                int32x2_t _sum21 = vdup_n_s32(0);
-                int32x2_t _sum30 = vdup_n_s32(0);
-                int32x2_t _sum31 = vdup_n_s32(0);
-#endif // __ARM_FEATURE_DOTPROD
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int8x8_t _pA = vld1_s8(pA);
-                    int8x16_t _pB0 = vld1q_s8(pB);
-                    int8x16_t _pB1 = vld1q_s8(pB + 16);
-
-#if __ARM_FEATURE_DOTPROD
-                    _sum00 = vdot_laneq_s32(_sum00, _pA, _pB0, 0);
-                    _sum01 = vdot_laneq_s32(_sum01, _pA, _pB0, 1);
-                    _sum10 = vdot_laneq_s32(_sum10, _pA, _pB0, 2);
-                    _sum11 = vdot_laneq_s32(_sum11, _pA, _pB0, 3);
-                    _sum20 = vdot_laneq_s32(_sum20, _pA, _pB1, 0);
-                    _sum21 = vdot_laneq_s32(_sum21, _pA, _pB1, 1);
-                    _sum30 = vdot_laneq_s32(_sum30, _pA, _pB1, 2);
-                    _sum31 = vdot_laneq_s32(_sum31, _pA, _pB1, 3);
-#else  // __ARM_FEATURE_DOTPROD
-                    int8x8_t _pA0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pA), 0));
-                    int8x8_t _pA1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pA), 1));
-                    int8x8_t _pA2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pA), 2));
-                    int8x8_t _pA3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pA), 3));
-
-                    int16x8_t _s0 = vmull_s8(_pA0, vget_low_s8(_pB0));
-                    int16x8_t _s1 = vmull_s8(_pA0, vget_high_s8(_pB0));
-                    int16x8_t _s2 = vmull_s8(_pA1, vget_low_s8(_pB0));
-                    int16x8_t _s3 = vmull_s8(_pA1, vget_high_s8(_pB0));
-                    _s0 = vmlal_s8(_s0, _pA2, vget_low_s8(_pB1));
-                    _s1 = vmlal_s8(_s1, _pA2, vget_high_s8(_pB1));
-                    _s2 = vmlal_s8(_s2, _pA3, vget_low_s8(_pB1));
-                    _s3 = vmlal_s8(_s3, _pA3, vget_high_s8(_pB1));
-                    _sum0 = vpadalq_s16(_sum0, _s0);
-                    _sum1 = vpadalq_s16(_sum1, _s1);
-                    _sum2 = vpadalq_s16(_sum2, _s2);
-                    _sum3 = vpadalq_s16(_sum3, _s3);
-#endif // __ARM_FEATURE_DOTPROD
-
-                    pA += 8;
-                    pB += 32;
-                }
-#if __ARM_FEATURE_DOTPROD
-                int32x2x2_t _sum0x = vzip_s32(_sum00, _sum01);
-                int32x2x2_t _sum1x = vzip_s32(_sum10, _sum11);
-                int32x2x2_t _sum2x = vzip_s32(_sum20, _sum21);
-                int32x2x2_t _sum3x = vzip_s32(_sum30, _sum31);
-                _sum0 = vaddq_s32(_sum0, vcombine_s32(_sum0x.val[0], _sum1x.val[0]));
-                _sum1 = vaddq_s32(_sum1, vcombine_s32(_sum2x.val[0], _sum3x.val[0]));
-                _sum2 = vaddq_s32(_sum2, vcombine_s32(_sum0x.val[1], _sum1x.val[1]));
-                _sum3 = vaddq_s32(_sum3, vcombine_s32(_sum2x.val[1], _sum3x.val[1]));
-#endif // __ARM_FEATURE_DOTPROD
-            }
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                int16x4_t _pA = vreinterpret_s16_s32(vld1_dup_s32((const int*)pA));
-                int8x16_t _pB = vld1q_s8(pB);
-
-                int16x4x2_t _pA01 = vuzp_s16(_pA, _pA);
-                int8x8_t _pA0 = vreinterpret_s8_s16(_pA01.val[0]);
-                int8x8_t _pA1 = vreinterpret_s8_s16(_pA01.val[1]);
-
-                int16x8_t _s0 = vmull_s8(_pA0, vget_low_s8(_pB));
-                int16x8_t _s1 = vmull_s8(_pA0, vget_high_s8(_pB));
-                int16x8_t _s2 = vmull_s8(_pA1, vget_low_s8(_pB));
-                int16x8_t _s3 = vmull_s8(_pA1, vget_high_s8(_pB));
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-                _sum2 = vpadalq_s16(_sum2, _s2);
-                _sum3 = vpadalq_s16(_sum3, _s3);
-
-                pA += 4;
-                pB += 16;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                int8x8_t _pA = vreinterpret_s8_s16(vld1_dup_s16((const short*)pA));
-                int8x8_t _pB = vld1_s8(pB);
-
-                int8x8x2_t _pA01 = vuzp_s8(_pA, _pA);
-
-                int16x8_t _s0 = vmull_s8(_pA01.val[0], _pB);
-                int16x8_t _s1 = vmull_s8(_pA01.val[1], _pB);
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s0));
-                _sum1 = vaddw_s16(_sum1, vget_high_s16(_s0));
-                _sum2 = vaddw_s16(_sum2, vget_low_s16(_s1));
-                _sum3 = vaddw_s16(_sum3, vget_high_s16(_s1));
-
-                pA += 2;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    vst1q_s32(outptr0 + out_hstep, _sum2);
-                    vst1q_s32(outptr0 + out_hstep + 4, _sum3);
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-            }
-
-            outptr += 16;
-        }
-#endif // __aarch64__
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            int32x4_t _sum0;
-            int32x4_t _sum1;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_s32(0);
-                _sum1 = vdupq_n_s32(0);
-            }
-            else
-            {
-                _sum0 = vld1q_s32(outptr);
-                _sum1 = vld1q_s32(outptr + 4);
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-#if __ARM_FEATURE_DOTPROD
-            {
-#if __ARM_FEATURE_MATMUL_INT8
-                int32x4_t _sum01 = vdupq_n_s32(0);
-                int32x4_t _sum23 = vdupq_n_s32(0);
-#else  // __ARM_FEATURE_MATMUL_INT8
-                int32x2_t _sum00 = vdup_n_s32(0);
-                int32x2_t _sum01 = vdup_n_s32(0);
-                int32x2_t _sum10 = vdup_n_s32(0);
-                int32x2_t _sum11 = vdup_n_s32(0);
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int8x16_t _pA = vld1q_s8(pA);
-                    int8x16_t _pB0 = vld1q_s8(pB);
-                    int8x16_t _pB1 = vld1q_s8(pB + 16);
-
-#if __ARM_FEATURE_MATMUL_INT8
-                    _sum01 = vmmlaq_s32(_sum01, _pA, _pB0);
-                    _sum23 = vmmlaq_s32(_sum23, _pA, _pB1);
-#else  // __ARM_FEATURE_MATMUL_INT8
-                    _sum00 = vdot_laneq_s32(_sum00, vget_low_s8(_pA), _pB0, 0);
-                    _sum01 = vdot_laneq_s32(_sum01, vget_low_s8(_pA), _pB0, 1);
-                    _sum10 = vdot_laneq_s32(_sum10, vget_low_s8(_pA), _pB0, 2);
-                    _sum11 = vdot_laneq_s32(_sum11, vget_low_s8(_pA), _pB0, 3);
-                    _sum00 = vdot_laneq_s32(_sum00, vget_high_s8(_pA), _pB1, 0);
-                    _sum01 = vdot_laneq_s32(_sum01, vget_high_s8(_pA), _pB1, 1);
-                    _sum10 = vdot_laneq_s32(_sum10, vget_high_s8(_pA), _pB1, 2);
-                    _sum11 = vdot_laneq_s32(_sum11, vget_high_s8(_pA), _pB1, 3);
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                    pA += 16;
-                    pB += 32;
-                }
-#if __ARM_FEATURE_MATMUL_INT8
-                _sum0 = vaddq_s32(_sum0, vcombine_s32(vget_low_s32(_sum01), vget_low_s32(_sum23)));
-                _sum1 = vaddq_s32(_sum1, vcombine_s32(vget_high_s32(_sum01), vget_high_s32(_sum23)));
-#else  // __ARM_FEATURE_MATMUL_INT8
-                int32x2x2_t _sum0x = vzip_s32(_sum00, _sum01);
-                int32x2x2_t _sum1x = vzip_s32(_sum10, _sum11);
-                _sum0 = vaddq_s32(_sum0, vcombine_s32(_sum0x.val[0], _sum1x.val[0]));
-                _sum1 = vaddq_s32(_sum1, vcombine_s32(_sum0x.val[1], _sum1x.val[1]));
-#endif // __ARM_FEATURE_MATMUL_INT8
-            }
-#endif // __ARM_FEATURE_DOTPROD
-            {
-#if __ARM_FEATURE_DOTPROD
-                int32x2_t _sum00 = vdup_n_s32(0);
-                int32x2_t _sum01 = vdup_n_s32(0);
-                int32x2_t _sum10 = vdup_n_s32(0);
-                int32x2_t _sum11 = vdup_n_s32(0);
-#endif // __ARM_FEATURE_DOTPROD
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int8x8_t _pA = vld1_s8(pA);
-                    int8x16_t _pB = vld1q_s8(pB);
-
-#if __ARM_FEATURE_DOTPROD
-                    _sum00 = vdot_laneq_s32(_sum00, _pA, _pB, 0);
-                    _sum01 = vdot_laneq_s32(_sum01, _pA, _pB, 1);
-                    _sum10 = vdot_laneq_s32(_sum10, _pA, _pB, 2);
-                    _sum11 = vdot_laneq_s32(_sum11, _pA, _pB, 3);
-#else  // __ARM_FEATURE_DOTPROD
-                    int8x8_t _pA0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pA), 0));
-                    int8x8_t _pA1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pA), 1));
-                    int8x8_t _pA2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pA), 2));
-                    int8x8_t _pA3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pA), 3));
-
-                    int16x8_t _s0 = vmull_s8(_pA0, vget_low_s8(_pB));
-                    int16x8_t _s1 = vmull_s8(_pA1, vget_low_s8(_pB));
-                    _s0 = vmlal_s8(_s0, _pA2, vget_high_s8(_pB));
-                    _s1 = vmlal_s8(_s1, _pA3, vget_high_s8(_pB));
-                    _sum0 = vpadalq_s16(_sum0, _s0);
-                    _sum1 = vpadalq_s16(_sum1, _s1);
-#endif // __ARM_FEATURE_DOTPROD
-
-                    pA += 8;
-                    pB += 16;
-                }
-#if __ARM_FEATURE_DOTPROD
-                int32x2x2_t _sum0x = vzip_s32(_sum00, _sum01);
-                int32x2x2_t _sum1x = vzip_s32(_sum10, _sum11);
-                _sum0 = vaddq_s32(_sum0, vcombine_s32(_sum0x.val[0], _sum1x.val[0]));
-                _sum1 = vaddq_s32(_sum1, vcombine_s32(_sum0x.val[1], _sum1x.val[1]));
-#endif // __ARM_FEATURE_DOTPROD
-            }
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                int16x4_t _pA = vreinterpret_s16_s32(vdup_lane_s32(vreinterpret_s32_s8(vld1_s8(pA)), 0));
-                int8x8_t _pB = vld1_s8(pB);
-
-                int16x4x2_t _pA01 = vuzp_s16(_pA, _pA);
-                int8x8_t _pA0 = vreinterpret_s8_s16(_pA01.val[0]);
-                int8x8_t _pA1 = vreinterpret_s8_s16(_pA01.val[1]);
-
-                int16x8_t _s0 = vmull_s8(_pA0, _pB);
-                int16x8_t _s1 = vmull_s8(_pA1, _pB);
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-
-                pA += 4;
-                pB += 8;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                int8x8_t _pA = vreinterpret_s8_s16(vld1_dup_s16((const short*)pA));
-                int8x8_t _pB = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(vld1_s8(pB)), 0));
-
-                _pA = vzip_s8(_pA, _pA).val[0];
-                _pA = vreinterpret_s8_s16(vzip_s16(vreinterpret_s16_s8(_pA), vreinterpret_s16_s8(_pA)).val[0]);
-
-                int16x8_t _s0 = vmull_s8(_pA, _pB);
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s0));
-                _sum1 = vaddw_s16(_sum1, vget_high_s16(_s0));
-
-                pA += 2;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + out_hstep, _sum1);
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-            }
-
-            outptr += 8;
-        }
-#endif // __ARM_NEON
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-#if __ARM_NEON
-            int32x4_t _sum;
-
-            if (k == 0)
-            {
-                _sum = vdupq_n_s32(0);
-            }
-            else
-            {
-                _sum = vld1q_s32(outptr);
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-
-#if __ARM_FEATURE_DOTPROD
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                int8x16_t _pA = vld1q_s8(pA);
-                int8x16_t _pB = vld1q_s8(pB);
-
-#if __ARM_FEATURE_MATMUL_INT8
-                _sum = vmmlaq_s32(_sum, _pA, _pB);
-#else  // __ARM_FEATURE_MATMUL_INT8
-                int32x4x2_t _pAA = vzipq_s32(vreinterpretq_s32_s8(_pA), vreinterpretq_s32_s8(_pA));
-                int8x16_t _pA01 = vreinterpretq_s8_s32(_pAA.val[0]);
-                int8x16_t _pA23 = vreinterpretq_s8_s32(_pAA.val[1]);
-                int8x16_t _pB01 = vcombine_s8(vget_low_s8(_pB), vget_low_s8(_pB));
-                int8x16_t _pB23 = vcombine_s8(vget_high_s8(_pB), vget_high_s8(_pB));
-
-                _sum = vdotq_s32(_sum, _pA01, _pB01);
-                _sum = vdotq_s32(_sum, _pA23, _pB23);
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                pA += 16;
-                pB += 16;
-            }
-#endif // __ARM_FEATURE_DOTPROD
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                int8x8_t _pA = vld1_s8(pA);
-                int8x8_t _pB = vld1_s8(pB);
-
-#if __ARM_FEATURE_DOTPROD
-                int32x2x2_t _pAA = vzip_s32(vreinterpret_s32_s8(_pA), vreinterpret_s32_s8(_pA));
-                int8x16_t _pA01 = vreinterpretq_s8_s32(vcombine_s32(_pAA.val[0], _pAA.val[1]));
-
-                int8x16_t _pB01 = vcombine_s8(_pB, _pB);
-
-                _sum = vdotq_s32(_sum, _pA01, _pB01);
-#else  // __ARM_FEATURE_DOTPROD
-                int16x4x2_t _pA01 = vzip_s16(vreinterpret_s16_s8(_pA), vreinterpret_s16_s8(_pA));
-                int32x2x2_t _pB01 = vzip_s32(vreinterpret_s32_s8(_pB), vreinterpret_s32_s8(_pB));
-
-                int16x8_t _s0 = vmull_s8(vreinterpret_s8_s16(_pA01.val[0]), vreinterpret_s8_s32(_pB01.val[0]));
-                _s0 = vmlal_s8(_s0, vreinterpret_s8_s16(_pA01.val[1]), vreinterpret_s8_s32(_pB01.val[1]));
-                _sum = vpadalq_s16(_sum, _s0);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 8;
-                pB += 8;
-            }
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                int8x8_t _pA = vld1_s8(pA);
-                int8x8_t _pB = vld1_s8(pB);
-
-                _pA = vreinterpret_s8_s16(vzip_s16(vreinterpret_s16_s8(_pA), vreinterpret_s16_s8(_pA)).val[0]);
-                _pB = vreinterpret_s8_s32(vzip_s32(vreinterpret_s32_s8(_pB), vreinterpret_s32_s8(_pB)).val[0]);
-
-                int16x8_t _s0 = vmull_s8(_pA, _pB);
-                _sum = vpadalq_s16(_sum, _s0);
-
-                pA += 4;
-                pB += 4;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                int8x8_t _pA = vreinterpret_s8_s16(vld1_dup_s16((const short*)pA));
-                int8x8_t _pB = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vld1_s8(pB)), 0));
-
-                _pA = vzip_s8(_pA, _pA).val[0];
-
-                int16x8_t _s0 = vmull_s8(_pA, _pB);
-                _sum = vaddw_s16(_sum, vget_low_s16(_s0));
-
-                pA += 2;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_s32(outptr0, vget_low_s32(_sum));
-                    vst1_s32(outptr0 + out_hstep, vget_high_s32(_sum));
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                vst1q_s32(outptr, _sum);
-            }
-
-            outptr += 4;
-#else // __ARM_NEON
-            int sum00;
-            int sum10;
-            int sum01;
-            int sum11;
-
-            if (k == 0)
-            {
-                sum00 = 0;
-                sum10 = 0;
-                sum01 = 0;
-                sum11 = 0;
-            }
-            else
-            {
-                sum00 = outptr[0];
-                sum10 = outptr[1];
-                sum01 = outptr[2];
-                sum11 = outptr[3];
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-#if __ARM_FEATURE_SIMD32 && NCNN_GNU_INLINE_ASM
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                // fomit-frame-pointer implied in optimized flag spare one register
-                // let us stay away from error: ‘asm’ operand has impossible constraints   --- nihui
-#if __OPTIMIZE__
-                asm volatile(
-                    "ldr    r2, [%0], #4    \n" // int8x4_t _pA = *((int8x4_t*)pA); pA += 4;
-                    "ldr    r4, [%1], #4    \n" // int8x4_t _pB = *((int8x4_t*)pB); pB += 4;
-                    "ror    r3, r2, #8      \n" // int8x4_t _pA_r8 = __ror(_pA, 8);
-                    "ror    r5, r4, #8      \n" // int8x4_t _pB_r8 = __ror(_pB, 8);
-                    "sxtb16 r2, r2          \n" // int16x2_t _pA0 = __sxtb16(_pA);
-                    "sxtb16 r4, r4          \n" // int16x2_t _pA1 = __sxtb16(_pA_r8);
-                    "sxtb16 r3, r3          \n" // int16x2_t _pB0 = __sxtb16(_pB);
-                    "sxtb16 r5, r5          \n" // int16x2_t _pB1 = __sxtb16(_pB_r8);
-                    "smlad  %2, r2, r4, %2  \n" // sum00 = __smlad(_pA0, _pB0, sum00);
-                    "smlad  %3, r3, r4, %3  \n" // sum10 = __smlad(_pA1, _pB0, sum10);
-                    "smlad  %4, r2, r5, %4  \n" // sum01 = __smlad(_pA0, _pB1, sum01);
-                    "smlad  %5, r3, r5, %5  \n" // sum11 = __smlad(_pA1, _pB1, sum11);
-                    : "=r"(pA),
-                    "=r"(pB),
-                    "=r"(sum00),
-                    "=r"(sum10),
-                    "=r"(sum01),
-                    "=r"(sum11)
-                    : "0"(pA),
-                    "1"(pB),
-                    "2"(sum00),
-                    "3"(sum10),
-                    "4"(sum01),
-                    "5"(sum11)
-                    : "memory", "r2", "r3", "r4", "r5");
-#else
-                int _pA0 = *((int*)pA);
-                int _pB0 = *((int*)pB);
-                int _pA1;
-                int _pB1;
-                asm volatile("ror %0, %1, #8"
-                             : "=r"(_pA1)
-                             : "r"(_pA0)
-                             :);
-                asm volatile("ror %0, %1, #8"
-                             : "=r"(_pB1)
-                             : "r"(_pB0)
-                             :);
-                asm volatile("sxtb16 %0, %0"
-                             : "=r"(_pA0)
-                             : "0"(_pA0)
-                             :);
-                asm volatile("sxtb16 %0, %0"
-                             : "=r"(_pA1)
-                             : "0"(_pA1)
-                             :);
-                asm volatile("sxtb16 %0, %0"
-                             : "=r"(_pB0)
-                             : "0"(_pB0)
-                             :);
-                asm volatile("sxtb16 %0, %0"
-                             : "=r"(_pB1)
-                             : "0"(_pB1)
-                             :);
-                asm volatile("smlad %0, %2, %3, %0"
-                             : "=r"(sum00)
-                             : "0"(sum00), "r"(_pA0), "r"(_pB0)
-                             :);
-                asm volatile("smlad %0, %2, %3, %0"
-                             : "=r"(sum10)
-                             : "0"(sum10), "r"(_pA1), "r"(_pB0)
-                             :);
-                asm volatile("smlad %0, %2, %3, %0"
-                             : "=r"(sum01)
-                             : "0"(sum01), "r"(_pA0), "r"(_pB1)
-                             :);
-                asm volatile("smlad %0, %2, %3, %0"
-                             : "=r"(sum11)
-                             : "0"(sum11), "r"(_pA1), "r"(_pB1)
-                             :);
-                pA += 4;
-                pB += 4;
-#endif
-            }
-#endif // __ARM_FEATURE_SIMD32 && NCNN_GNU_INLINE_ASM
-            for (; kk < max_kk; kk += 1)
-            {
-                sum00 += pA[0] * pB[0];
-                sum10 += pA[1] * pB[0];
-                sum01 += pA[0] * pB[1];
-                sum11 += pA[1] * pB[1];
-
-                pA += 2;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum00;
-                    outptr0[1] = sum01;
-                    outptr0[out_hstep] = sum10;
-                    outptr0[out_hstep + 1] = sum11;
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                outptr[0] = sum00;
-                outptr[1] = sum10;
-                outptr[2] = sum01;
-                outptr[3] = sum11;
-            }
-
-            outptr += 4;
-#endif // __ARM_NEON
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-#if __ARM_NEON
-            int32x2_t _sum;
-
-            if (k == 0)
-            {
-                _sum = vdup_n_s32(0);
-            }
-            else
-            {
-                _sum = vld1_s32(outptr);
-            }
-#else  // __ARM_NEON
-            int sum0;
-            int sum1;
-
-            if (k == 0)
-            {
-                sum0 = 0;
-                sum1 = 0;
-            }
-            else
-            {
-                sum0 = outptr[0];
-                sum1 = outptr[1];
-            }
-#endif // __ARM_NEON
-
-            const signed char* pA = pAT;
-            int kk = 0;
-#if __ARM_NEON
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-            {
-                int32x4_t _sum0 = vdupq_n_s32(0);
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int8x16_t _pA = vld1q_s8(pA);
-                    int8x8_t _pB = vld1_s8(pB);
-
-                    int8x16_t _pBB = vcombine_s8(_pB, _pB);
-
-                    _sum0 = vdotq_s32(_sum0, _pA, _pBB);
-
-                    pA += 16;
-                    pB += 8;
-                }
-                int32x2_t _ss = vpadd_s32(vget_low_s32(_sum0), vget_high_s32(_sum0));
-                _sum = vadd_s32(_sum, _ss);
-            }
-#else  // __ARM_FEATURE_MATMUL_INT8
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                int8x16_t _pA = vld1q_s8(pA);
-                int8x8_t _pB = vld1_s8(pB);
-
-                _sum = vdot_lane_s32(_sum, vget_low_s8(_pA), _pB, 0);
-                _sum = vdot_lane_s32(_sum, vget_high_s8(_pA), _pB, 1);
-
-                pA += 16;
-                pB += 8;
-            }
-#endif // __ARM_FEATURE_MATMUL_INT8
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                int8x8_t _pA = vld1_s8(pA);
-                int8x8_t _pB = vreinterpret_s8_s32(vld1_dup_s32((const int*)pB));
-
-                _sum = vdot_s32(_sum, _pA, _pB);
-
-                pA += 8;
-                pB += 4;
-            }
-#else  // __ARM_FEATURE_DOTPROD
-            {
-                int32x4_t _sum0 = vdupq_n_s32(0);
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int8x8_t _pA = vld1_s8(pA);
-                    int8x8_t _pB = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(vld1_s8(pB)), 0));
-
-                    _pB = vreinterpret_s8_s16(vzip_s16(vreinterpret_s16_s8(_pB), vreinterpret_s16_s8(_pB)).val[0]);
-
-                    int16x8_t _s0 = vmull_s8(_pA, _pB);
-                    _sum0 = vpadalq_s16(_sum0, _s0);
-
-                    pA += 8;
-                    pB += 4;
-                }
-                int32x2_t _ss = vadd_s32(vget_low_s32(_sum0), vget_high_s32(_sum0));
-                _sum = vadd_s32(_sum, _ss);
-            }
-#endif // __ARM_FEATURE_DOTPROD
-            int sum0 = vget_lane_s32(_sum, 0);
-            int sum1 = vget_lane_s32(_sum, 1);
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                sum0 += pA[0] * pB[0];
-                sum0 += pA[1] * pB[1];
-                sum1 += pA[2] * pB[0];
-                sum1 += pA[3] * pB[1];
-                pA += 4;
-                pB += 2;
-            }
-#endif // __ARM_NEON
-            for (; kk < max_kk; kk += 1)
-            {
-                sum0 += pA[0] * pB[0];
-                sum1 += pA[1] * pB[0];
-                pA += 2;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum0;
-                    outptr0[out_hstep] = sum1;
-                    outptr0++;
-                }
-            }
-            else
-            {
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-            }
-
-            outptr += 2;
-        }
-
-        pAT += max_kk * 2;
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        int* outptr0 = (int*)top_blob + (i + ii) * out_hstep + j;
-
-        const signed char* pB = pBT;
-
-        int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            int32x4_t _sum0;
-            int32x4_t _sum1;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_s32(0);
-                _sum1 = vdupq_n_s32(0);
-            }
-            else
-            {
-                _sum0 = vld1q_s32(outptr);
-                _sum1 = vld1q_s32(outptr + 4);
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-#if __ARM_FEATURE_DOTPROD
-            {
-#if __ARM_FEATURE_MATMUL_INT8
-                int32x4_t _sum00 = vdupq_n_s32(0);
-                int32x4_t _sum01 = vdupq_n_s32(0);
-                int32x4_t _sum10 = vdupq_n_s32(0);
-                int32x4_t _sum11 = vdupq_n_s32(0);
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int8x8_t _pA = vld1_s8(pA);
-                    int8x16_t _pB0 = vld1q_s8(pB);
-                    int8x16_t _pB1 = vld1q_s8(pB + 16);
-                    int8x16_t _pB2 = vld1q_s8(pB + 32);
-                    int8x16_t _pB3 = vld1q_s8(pB + 48);
-
-#if __ARM_FEATURE_MATMUL_INT8
-                    int8x16_t _pAA = vcombine_s8(_pA, _pA);
-                    _sum00 = vdotq_s32(_sum00, _pAA, _pB0);
-                    _sum01 = vdotq_s32(_sum01, _pAA, _pB1);
-                    _sum10 = vdotq_s32(_sum10, _pAA, _pB2);
-                    _sum11 = vdotq_s32(_sum11, _pAA, _pB3);
-#else  // __ARM_FEATURE_MATMUL_INT8
-                    _sum0 = vdotq_lane_s32(_sum0, _pB0, _pA, 0);
-                    _sum1 = vdotq_lane_s32(_sum1, _pB1, _pA, 0);
-                    _sum0 = vdotq_lane_s32(_sum0, _pB2, _pA, 1);
-                    _sum1 = vdotq_lane_s32(_sum1, _pB3, _pA, 1);
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                    pA += 8;
-                    pB += 64;
-                }
-#if __ARM_FEATURE_MATMUL_INT8
-                _sum0 = vaddq_s32(_sum0, vpaddq_s32(_sum00, _sum01));
-                _sum1 = vaddq_s32(_sum1, vpaddq_s32(_sum10, _sum11));
-#endif // __ARM_FEATURE_MATMUL_INT8
-            }
-#endif // __ARM_FEATURE_DOTPROD
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                int8x8_t _pA = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(vld1_s8(pA)), 0));
-                int8x16_t _pB0 = vld1q_s8(pB);
-                int8x16_t _pB1 = vld1q_s8(pB + 16);
-
-#if __ARM_FEATURE_DOTPROD
-                _sum0 = vdotq_lane_s32(_sum0, _pB0, _pA, 0);
-                _sum1 = vdotq_lane_s32(_sum1, _pB1, _pA, 0);
-#else  // __ARM_FEATURE_DOTPROD
-                int8x8_t _pA0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pA), 0));
-                int8x8_t _pA1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pA), 1));
-                int16x8_t _s0 = vmull_s8(_pA0, vget_low_s8(_pB0));
-                int16x8_t _s1 = vmull_s8(_pA0, vget_high_s8(_pB0));
-                _s0 = vmlal_s8(_s0, _pA1, vget_low_s8(_pB1));
-                _s1 = vmlal_s8(_s1, _pA1, vget_high_s8(_pB1));
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 4;
-                pB += 32;
-            }
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                int8x8_t _pA = vreinterpret_s8_s16(vld1_dup_s16((const short*)pA));
-                int8x16_t _pB = vld1q_s8(pB);
-
-                int16x8_t _s0 = vmull_s8(_pA, vget_low_s8(_pB));
-                int16x8_t _s1 = vmull_s8(_pA, vget_high_s8(_pB));
-                _sum0 = vpadalq_s16(_sum0, _s0);
-                _sum1 = vpadalq_s16(_sum1, _s1);
-
-                pA += 2;
-                pB += 16;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                int8x8_t _pA = vld1_dup_s8(pA);
-                int8x8_t _pB = vld1_s8(pB);
-
-                int16x8_t _s0 = vmull_s8(_pA, _pB);
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s0));
-                _sum1 = vaddw_s16(_sum1, vget_high_s16(_s0));
-
-                pA += 1;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    vst1q_s32(outptr0 + 4, _sum1);
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-            }
-
-            outptr += 8;
-        }
-#endif // __aarch64__
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            int32x4_t _sum0;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_s32(0);
-            }
-            else
-            {
-                _sum0 = vld1q_s32(outptr);
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-#if __ARM_FEATURE_DOTPROD
-            {
-#if __ARM_FEATURE_MATMUL_INT8
-                int32x4_t _sum00 = vdupq_n_s32(0);
-                int32x4_t _sum01 = vdupq_n_s32(0);
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int8x8_t _pA = vld1_s8(pA);
-                    int8x16_t _pB0 = vld1q_s8(pB);
-                    int8x16_t _pB1 = vld1q_s8(pB + 16);
-
-#if __ARM_FEATURE_MATMUL_INT8
-                    int8x16_t _pAA = vcombine_s8(_pA, _pA);
-                    _sum00 = vdotq_s32(_sum00, _pAA, _pB0);
-                    _sum01 = vdotq_s32(_sum01, _pAA, _pB1);
-#else  // __ARM_FEATURE_MATMUL_INT8
-                    _sum0 = vdotq_lane_s32(_sum0, _pB0, _pA, 0);
-                    _sum0 = vdotq_lane_s32(_sum0, _pB1, _pA, 1);
-#endif // __ARM_FEATURE_MATMUL_INT8
-
-                    pA += 8;
-                    pB += 32;
-                }
-#if __ARM_FEATURE_MATMUL_INT8
-                _sum0 = vaddq_s32(_sum0, vpaddq_s32(_sum00, _sum01));
-#endif // __ARM_FEATURE_MATMUL_INT8
-            }
-#endif // __ARM_FEATURE_DOTPROD
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                int8x8_t _pA = vld1_s8(pA);
-                int8x16_t _pB = vld1q_s8(pB);
-
-#if __ARM_FEATURE_DOTPROD
-                _sum0 = vdotq_lane_s32(_sum0, _pB, _pA, 0);
-#else  // __ARM_FEATURE_DOTPROD
-                int8x8_t _pA0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pA), 0));
-                int8x8_t _pA1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(_pA), 1));
-                int16x8_t _s0 = vmull_s8(_pA0, vget_low_s8(_pB));
-                _s0 = vmlal_s8(_s0, _pA1, vget_high_s8(_pB));
-                _sum0 = vpadalq_s16(_sum0, _s0);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 4;
-                pB += 16;
-            }
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                int8x8_t _pA = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(vld1_s8(pA)), 0));
-                int8x8_t _pB = vld1_s8(pB);
-
-                int16x8_t _s0 = vmull_s8(_pA, _pB);
-                _sum0 = vpadalq_s16(_sum0, _s0);
-
-                pA += 2;
-                pB += 8;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                int8x8_t _pA = vld1_dup_s8(pA);
-                int8x8_t _pB = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(vld1_s8(pB)), 0));
-
-                int16x8_t _s0 = vmull_s8(_pA, _pB);
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s0));
-
-                pA += 1;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1q_s32(outptr0, _sum0);
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1q_s32(outptr, _sum0);
-            }
-
-            outptr += 4;
-        }
-#endif // __ARM_NEON
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-#if __ARM_NEON
-            int32x2_t _sum;
-
-            if (k == 0)
-            {
-                _sum = vdup_n_s32(0);
-            }
-            else
-            {
-                _sum = vld1_s32(outptr);
-            }
-#else  // __ARM_NEON
-            int sum0;
-            int sum1;
-
-            if (k == 0)
-            {
-                sum0 = 0;
-                sum1 = 0;
-            }
-            else
-            {
-                sum0 = outptr[0];
-                sum1 = outptr[1];
-            }
-#endif // __ARM_NEON
-
-            const signed char* pA = pAT;
-            int kk = 0;
-#if __ARM_NEON
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-            {
-                int32x4_t _sum0 = vdupq_n_s32(0);
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int8x8_t _pA = vld1_s8(pA);
-                    int8x16_t _pB = vld1q_s8(pB);
-
-                    int8x16_t _pAA = vcombine_s8(_pA, _pA);
-
-                    _sum0 = vdotq_s32(_sum0, _pAA, _pB);
-
-                    pA += 8;
-                    pB += 16;
-                }
-                int32x2_t _ss = vpadd_s32(vget_low_s32(_sum0), vget_high_s32(_sum0));
-                _sum = vadd_s32(_sum, _ss);
-            }
-#else  // __ARM_FEATURE_MATMUL_INT8
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                int8x8_t _pA = vld1_s8(pA);
-                int8x16_t _pB = vld1q_s8(pB);
-
-                _sum = vdot_lane_s32(_sum, vget_low_s8(_pB), _pA, 0);
-                _sum = vdot_lane_s32(_sum, vget_high_s8(_pB), _pA, 1);
-
-                pA += 8;
-                pB += 16;
-            }
-#endif // __ARM_FEATURE_MATMUL_INT8
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                int8x8_t _pA = vreinterpret_s8_s32(vld1_dup_s32((const int*)pA));
-                int8x8_t _pB = vld1_s8(pB);
-
-                _sum = vdot_s32(_sum, _pA, _pB);
-
-                pA += 4;
-                pB += 8;
-            }
-#else  // __ARM_FEATURE_DOTPROD
-            {
-                int32x4_t _sum0 = vdupq_n_s32(0);
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int8x8_t _pA = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(vld1_s8(pA)), 0));
-                    int8x8_t _pB = vld1_s8(pB);
-
-                    _pA = vreinterpret_s8_s16(vzip_s16(vreinterpret_s16_s8(_pA), vreinterpret_s16_s8(_pA)).val[0]);
-
-                    int16x8_t _s0 = vmull_s8(_pA, _pB);
-                    _sum0 = vpadalq_s16(_sum0, _s0);
-
-                    pA += 4;
-                    pB += 8;
-                }
-                int32x2_t _ss = vadd_s32(vget_low_s32(_sum0), vget_high_s32(_sum0));
-                _sum = vadd_s32(_sum, _ss);
-            }
-#endif // __ARM_FEATURE_DOTPROD
-            int sum0 = vget_lane_s32(_sum, 0);
-            int sum1 = vget_lane_s32(_sum, 1);
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                sum0 += pA[0] * pB[0];
-                sum0 += pA[1] * pB[1];
-                sum1 += pA[0] * pB[2];
-                sum1 += pA[1] * pB[3];
-                pA += 2;
-                pB += 4;
-            }
-#endif // __ARM_NEON
-            for (; kk < max_kk; kk += 1)
-            {
-                sum0 += pA[0] * pB[0];
-                sum1 += pA[0] * pB[1];
-                pA += 1;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum0;
-                    outptr0[1] = sum1;
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-            }
-
-            outptr += 2;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            int sum;
-
-            if (k == 0)
-            {
-                sum = 0;
-            }
-            else
-            {
-                sum = outptr[0];
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-#if __ARM_NEON
-            int32x4_t _sum = vdupq_n_s32(0);
-            for (; kk + 15 < max_kk; kk += 16)
-            {
-                int8x16_t _pA = vld1q_s8(pA);
-                int8x16_t _pB = vld1q_s8(pB);
-
-#if __ARM_FEATURE_DOTPROD
-                _sum = vdotq_s32(_sum, _pA, _pB);
-#else  // __ARM_FEATURE_DOTPROD
-                int16x8_t _s0 = vmull_s8(vget_low_s8(_pA), vget_low_s8(_pB));
-                _s0 = vmlal_s8(_s0, vget_high_s8(_pA), vget_high_s8(_pB));
-                _sum = vpadalq_s16(_sum, _s0);
-#endif // __ARM_FEATURE_DOTPROD
-
-                pA += 16;
-                pB += 16;
-            }
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                int8x8_t _pA = vld1_s8(pA);
-                int8x8_t _pB = vld1_s8(pB);
-
-                int16x8_t _s0 = vmull_s8(_pA, _pB);
-                _sum = vpadalq_s16(_sum, _s0);
-
-                pA += 8;
-                pB += 8;
-            }
-#if __aarch64__
-            sum += vaddvq_s32(_sum);
-#else
-            int32x2_t _ss = vadd_s32(vget_low_s32(_sum), vget_high_s32(_sum));
-            _ss = vpadd_s32(_ss, _ss);
-            sum += vget_lane_s32(_ss, 0);
-#endif
-#endif // __ARM_NEON
-            for (; kk < max_kk; kk += 1)
-            {
-                sum += pA[0] * pB[0];
-                pA += 1;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum;
-                    outptr0++;
-                }
-            }
-            else
-            {
-                outptr[0] = sum;
-            }
-
-            outptr += 1;
-        }
-
-        pAT += max_kk;
-    }
-}
-
-static void convolution_im2col_gemm_get_optimal_tile_mnk_int8(int M, int N, int K, int& TILE_M, int& TILE_N, int& TILE_K, int nT)
-{
-    // resolve optimal tile size from cache size
-    const size_t l2_cache_size_int8 = (int)(get_cpu_level2_cache_size() / sizeof(signed char));
-
-    if (nT == 0)
-        nT = get_physical_big_cpu_count();
-
-    // solve K
-    {
-        // try not to split K
-#if __ARM_NEON
-        int tile_size = (l2_cache_size_int8 - 16) / 8;
-#else
-        int tile_size = (l2_cache_size_int8 - 2) / 3;
-#endif
-
-#if __ARM_NEON
-        TILE_K = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_K = std::max(2, tile_size / 2 * 2);
-#endif
-
-        int nn_K = (K + TILE_K - 1) / TILE_K;
-#if __ARM_NEON
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 3) / 4 * 4);
-#else
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 1) / 2 * 2);
-#endif
-    }
-
-    // solve M
-    {
-#if __ARM_NEON
-        int nn_M = (M + 31) / 32;
-#else
-        int nn_M = (M + 7) / 8;
-#endif
-
-#if __ARM_NEON
-        TILE_M = std::max(8, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-#else
-        TILE_M = std::max(2, ((M + nn_M - 1) / nn_M + 1) / 2 * 2);
-#endif
-    }
-
-    {
-        TILE_M *= std::min(nT, get_physical_cpu_count());
-
-        int nn_M = (M + TILE_M - 1) / TILE_M;
-#if __ARM_NEON
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-#else
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 1) / 2 * 2);
-#endif
-
-        if (nT > 1)
-        {
-#if __ARM_NEON
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 7) / 8 * 8);
-#else
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 1) / 2 * 2);
-#endif
-        }
-    }
-
-    if (N > 0)
-    {
-        int tile_size;
-        if (TILE_K >= K)
-        {
-            tile_size = (l2_cache_size_int8 - TILE_M * TILE_K) / TILE_K;
-        }
-        else
-        {
-            tile_size = (l2_cache_size_int8 - TILE_M * TILE_K) / (TILE_M * 4 + TILE_K);
-        }
-
-#if __ARM_NEON
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_N = std::max(1, tile_size);
-#endif
-
-        int nn_N = (N + TILE_N - 1) / TILE_N;
-#if __ARM_NEON
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#else
-        TILE_N = std::min(TILE_N, (N + nn_N - 1) / nn_N);
-#endif
-    }
-}
-
-static void convolution_im2col_input_tile_conv1x1s1d1_int8(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk)
-{
-    const int elempack = bottom_blob.elempack;
-
-    signed char* pp = B;
-
-    int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-        if (elempack == 8)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k / 8) + (j + jj) * 8;
-            const size_t cstep = bottom_blob.cstep * 8;
-
-            int kk = 0;
-#if __ARM_FEATURE_MATMUL_INT8
-            for (; kk < max_kk / 8; kk++)
-            {
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], %4 \n"
-                    "st1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp),
-                    "r"(cstep)
-                    : "memory", "v0", "v1", "v2", "v3");
-#else  // NCNN_GNU_INLINE_ASM
-                int8x16_t _r01 = vld1q_s8(p0);
-                int8x16_t _r23 = vld1q_s8(p0 + 16);
-                int8x16_t _r45 = vld1q_s8(p0 + 32);
-                int8x16_t _r67 = vld1q_s8(p0 + 48);
-                vst1q_s8(pp, _r01);
-                vst1q_s8(pp + 16, _r23);
-                vst1q_s8(pp + 32, _r45);
-                vst1q_s8(pp + 48, _r67);
-                pp += 64;
-                p0 += cstep;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-#elif __ARM_FEATURE_DOTPROD
-            for (; kk < max_kk / 8; kk++)
-            {
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], %4 \n"
-                    "uzp1   v4.4s, v0.4s, v1.4s         \n"
-                    "uzp2   v6.4s, v0.4s, v1.4s         \n"
-                    "uzp1   v5.4s, v2.4s, v3.4s         \n"
-                    "uzp2   v7.4s, v2.4s, v3.4s         \n"
-                    "st1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp),
-                    "r"(cstep)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
-#else  // NCNN_GNU_INLINE_ASM
-                int32x4x2_t _r0246 = vld2q_s32((const int*)p0);
-                int32x4x2_t _r1357 = vld2q_s32((const int*)(p0 + 32));
-                vst1q_s32((int*)pp, _r0246.val[0]);
-                vst1q_s32((int*)(pp + 16), _r1357.val[0]);
-                vst1q_s32((int*)(pp + 32), _r0246.val[1]);
-                vst1q_s32((int*)(pp + 48), _r1357.val[1]);
-                pp += 64;
-                p0 += cstep;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-            for (; kk < max_kk / 8; kk++)
-            {
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #512]       \n"
-                    "ld4    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0], %4 \n"
-                    "st1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp),
-                    "r"(cstep)
-                    : "memory", "v0", "v1", "v2", "v3");
-#else  // NCNN_GNU_INLINE_ASM
-                int16x8x4_t _r0 = vld4q_s16((const short*)p0);
-                vst1q_s16((short*)pp, _r0.val[0]);
-                vst1q_s16((short*)(pp + 16), _r0.val[1]);
-                vst1q_s16((short*)(pp + 32), _r0.val[2]);
-                vst1q_s16((short*)(pp + 48), _r0.val[3]);
-                pp += 64;
-                p0 += cstep;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-        }
-
-        if (elempack == 1)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k) + (j + jj);
-            const size_t cstep = bottom_blob.cstep;
-
-            int kk = 0;
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #64]        \n"
-                    "ld1    {v0.8b}, [%0], %4           \n"
-                    "prfm   pldl1keep, [%0, #64]        \n"
-                    "ld1    {v1.8b}, [%0], %4           \n"
-                    "prfm   pldl1keep, [%0, #64]        \n"
-                    "ld1    {v0.d}[1], [%0], %4         \n"
-                    "prfm   pldl1keep, [%0, #64]        \n"
-                    "ld1    {v1.d}[1], [%0], %4         \n"
-                    "prfm   pldl1keep, [%0, #64]        \n"
-                    "ld1    {v2.8b}, [%0], %4           \n"
-                    "prfm   pldl1keep, [%0, #64]        \n"
-                    "ld1    {v3.8b}, [%0], %4           \n"
-                    "prfm   pldl1keep, [%0, #64]        \n"
-                    "ld1    {v2.d}[1], [%0], %4         \n"
-                    "prfm   pldl1keep, [%0, #64]        \n"
-                    "ld1    {v3.d}[1], [%0], %4         \n"
-                    "zip1   v4.16b, v0.16b, v1.16b      \n"
-                    "zip2   v5.16b, v0.16b, v1.16b      \n"
-                    "zip1   v6.16b, v2.16b, v3.16b      \n"
-                    "zip2   v7.16b, v2.16b, v3.16b      \n"
-                    "st4    {v4.8h, v5.8h, v6.8h, v7.8h}, [%1], #64 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp),
-                    "r"(cstep)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
-#else  // NCNN_GNU_INLINE_ASM
-                int8x8_t _r0 = vld1_s8(p0);
-                int8x8_t _r1 = vld1_s8(p0 + cstep);
-                int8x8_t _r2 = vld1_s8(p0 + cstep * 2);
-                int8x8_t _r3 = vld1_s8(p0 + cstep * 3);
-                int8x8_t _r4 = vld1_s8(p0 + cstep * 4);
-                int8x8_t _r5 = vld1_s8(p0 + cstep * 5);
-                int8x8_t _r6 = vld1_s8(p0 + cstep * 6);
-                int8x8_t _r7 = vld1_s8(p0 + cstep * 7);
-                // save as transpose8x8
-                int8x8x2_t _r01 = vzip_s8(_r0, _r1);
-                int8x8x2_t _r23 = vzip_s8(_r2, _r3);
-                int8x8x2_t _r45 = vzip_s8(_r4, _r5);
-                int8x8x2_t _r67 = vzip_s8(_r6, _r7);
-                int16x8x4_t _r0246;
-                _r0246.val[0] = vreinterpretq_s16_s8(vcombine_s8(_r01.val[0], _r01.val[1]));
-                _r0246.val[1] = vreinterpretq_s16_s8(vcombine_s8(_r23.val[0], _r23.val[1]));
-                _r0246.val[2] = vreinterpretq_s16_s8(vcombine_s8(_r45.val[0], _r45.val[1]));
-                _r0246.val[3] = vreinterpretq_s16_s8(vcombine_s8(_r67.val[0], _r67.val[1]));
-                vst4q_s16((short*)pp, _r0246);
-                pp += 64;
-                p0 += cstep * 8;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-#endif // __ARM_FEATURE_MATMUL_INT8
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #64]        \n"
-                    "ld1    {v0.8b}, [%0], %4           \n"
-                    "prfm   pldl1keep, [%0, #64]        \n"
-                    "ld1    {v1.8b}, [%0], %4           \n"
-                    "prfm   pldl1keep, [%0, #64]        \n"
-                    "ld1    {v2.8b}, [%0], %4           \n"
-                    "prfm   pldl1keep, [%0, #64]        \n"
-                    "ld1    {v3.8b}, [%0], %4           \n"
-                    "st4    {v0.8b, v1.8b, v2.8b, v3.8b}, [%1], #32 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp),
-                    "r"(cstep)
-                    : "memory", "v0", "v1", "v2", "v3");
-#else  // NCNN_GNU_INLINE_ASM
-                int8x8x4_t _r0123;
-                _r0123.val[0] = vld1_s8(p0);
-                _r0123.val[1] = vld1_s8(p0 + cstep);
-                _r0123.val[2] = vld1_s8(p0 + cstep * 2);
-                _r0123.val[3] = vld1_s8(p0 + cstep * 3);
-                vst4_s8(pp, _r0123);
-                pp += 32;
-                p0 += cstep * 4;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-#endif // __ARM_FEATURE_DOTPROD
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #64]        \n"
-                    "ld1    {v0.8b}, [%0], %4           \n"
-                    "prfm   pldl1keep, [%0, #64]        \n"
-                    "ld1    {v1.8b}, [%0], %4           \n"
-                    "st2    {v0.8b, v1.8b}, [%1], #16   \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp),
-                    "r"(cstep)
-                    : "memory", "v0", "v1");
-#else  // NCNN_GNU_INLINE_ASM
-                int8x8x2_t _r01;
-                _r01.val[0] = vld1_s8(p0);
-                _r01.val[1] = vld1_s8(p0 + cstep);
-                vst2_s8(pp, _r01);
-                pp += 16;
-                p0 += cstep * 2;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-            for (; kk < max_kk; kk++)
-            {
-                vst1_s8(pp, vld1_s8(p0));
-                pp += 8;
-                p0 += cstep;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-        if (elempack == 8)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k / 8) + (j + jj) * 8;
-            const size_t cstep = bottom_blob.cstep * 8;
-
-            int kk = 0;
-#if __ARM_FEATURE_MATMUL_INT8
-            for (; kk < max_kk / 8; kk++)
-            {
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #256]       \n"
-                    "ld1    {v0.16b, v1.16b}, [%0], %4  \n"
-                    "st1    {v0.16b, v1.16b}, [%1], #32 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp),
-                    "r"(cstep)
-                    : "memory", "v0", "v1");
-#else  // NCNN_GNU_INLINE_ASM
-                int8x16_t _r01 = vld1q_s8(p0);
-                int8x16_t _r23 = vld1q_s8(p0 + 16);
-                vst1q_s8(pp, _r01);
-                vst1q_s8(pp + 16, _r23);
-                pp += 32;
-                p0 += cstep;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-#elif __ARM_FEATURE_DOTPROD
-            for (; kk < max_kk / 8; kk++)
-            {
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #256]       \n"
-                    "ld1    {v0.8b, v1.8b, v2.8b, v3.8b}, [%0], %4 \n"
-                    "st4    {v0.2s, v1.2s, v2.2s, v3.2s}, [%1], #32 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp),
-                    "r"(cstep)
-                    : "memory", "v0", "v1", "v2", "v3");
-#else  // NCNN_GNU_INLINE_ASM
-                int32x2x4_t _r0123;
-                _r0123.val[0] = vreinterpret_s32_s8(vld1_s8(p0));
-                _r0123.val[1] = vreinterpret_s32_s8(vld1_s8(p0 + 8));
-                _r0123.val[2] = vreinterpret_s32_s8(vld1_s8(p0 + 16));
-                _r0123.val[3] = vreinterpret_s32_s8(vld1_s8(p0 + 24));
-                vst4_s32((int*)pp, _r0123);
-                pp += 32;
-                p0 += cstep;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-            for (; kk < max_kk / 8; kk++)
-            {
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #256]       \n"
-                    "ld1    {v0.8b, v1.8b, v2.8b, v3.8b}, [%0], %4 \n"
-                    "st4    {v0.4h, v1.4h, v2.4h, v3.4h}, [%1], #32 \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp),
-                    "r"(cstep)
-                    : "memory", "v0", "v1", "v2", "v3");
-#else
-                asm volatile(
-                    "pld        [%0, #256]          \n"
-                    "vld1.s8    {d0-d3}, [%0], %4   \n"
-                    "vst4.s16   {d0-d3}, [%1 :64]!  \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp),
-                    "r"(cstep)
-                    : "memory", "q0", "q1");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                int16x4x4_t _r0123;
-                _r0123.val[0] = vreinterpret_s16_s8(vld1_s8(p0));
-                _r0123.val[1] = vreinterpret_s16_s8(vld1_s8(p0 + 8));
-                _r0123.val[2] = vreinterpret_s16_s8(vld1_s8(p0 + 16));
-                _r0123.val[3] = vreinterpret_s16_s8(vld1_s8(p0 + 24));
-                vst4_s16((short*)pp, _r0123);
-                pp += 32;
-                p0 += cstep;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-        }
-
-        if (elempack == 1)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k) + (j + jj);
-            const size_t cstep = bottom_blob.cstep;
-
-            int kk = 0;
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[cstep + 0];
-                pp[2] = p0[cstep * 2 + 0];
-                pp[3] = p0[cstep * 3 + 0];
-                pp[4] = p0[cstep * 4 + 0];
-                pp[5] = p0[cstep * 5 + 0];
-                pp[6] = p0[cstep * 6 + 0];
-                pp[7] = p0[cstep * 7 + 0];
-                pp[8] = p0[1];
-                pp[9] = p0[cstep + 1];
-                pp[10] = p0[cstep * 2 + 1];
-                pp[11] = p0[cstep * 3 + 1];
-                pp[12] = p0[cstep * 4 + 1];
-                pp[13] = p0[cstep * 5 + 1];
-                pp[14] = p0[cstep * 6 + 1];
-                pp[15] = p0[cstep * 7 + 1];
-                pp[16] = p0[2];
-                pp[17] = p0[cstep + 2];
-                pp[18] = p0[cstep * 2 + 2];
-                pp[19] = p0[cstep * 3 + 2];
-                pp[20] = p0[cstep * 4 + 2];
-                pp[21] = p0[cstep * 5 + 2];
-                pp[22] = p0[cstep * 6 + 2];
-                pp[23] = p0[cstep * 7 + 2];
-                pp[24] = p0[3];
-                pp[25] = p0[cstep + 3];
-                pp[26] = p0[cstep * 2 + 3];
-                pp[27] = p0[cstep * 3 + 3];
-                pp[28] = p0[cstep * 4 + 3];
-                pp[29] = p0[cstep * 5 + 3];
-                pp[30] = p0[cstep * 6 + 3];
-                pp[31] = p0[cstep * 7 + 3];
-                pp += 32;
-                p0 += cstep * 8;
-            }
-#endif // __ARM_FEATURE_MATMUL_INT8
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[cstep + 0];
-                pp[2] = p0[cstep * 2 + 0];
-                pp[3] = p0[cstep * 3 + 0];
-                pp[4] = p0[1];
-                pp[5] = p0[cstep + 1];
-                pp[6] = p0[cstep * 2 + 1];
-                pp[7] = p0[cstep * 3 + 1];
-                pp[8] = p0[2];
-                pp[9] = p0[cstep + 2];
-                pp[10] = p0[cstep * 2 + 2];
-                pp[11] = p0[cstep * 3 + 2];
-                pp[12] = p0[3];
-                pp[13] = p0[cstep + 3];
-                pp[14] = p0[cstep * 2 + 3];
-                pp[15] = p0[cstep * 3 + 3];
-                pp += 16;
-                p0 += cstep * 4;
-            }
-#endif // __ARM_FEATURE_DOTPROD
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[cstep + 0];
-                pp[2] = p0[1];
-                pp[3] = p0[cstep + 1];
-                pp[4] = p0[2];
-                pp[5] = p0[cstep + 2];
-                pp[6] = p0[3];
-                pp[7] = p0[cstep + 3];
-                pp += 8;
-                p0 += cstep * 2;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                pp[2] = p0[2];
-                pp[3] = p0[3];
-                pp += 4;
-                p0 += cstep;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-#if __ARM_NEON
-        if (elempack == 8)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k / 8) + (j + jj) * 8;
-            const size_t cstep = bottom_blob.cstep * 8;
-
-            int kk = 0;
-#if __ARM_FEATURE_MATMUL_INT8
-            for (; kk < max_kk / 8; kk++)
-            {
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #128]       \n"
-                    "ld1    {v0.16b}, [%0], %4          \n"
-                    "st1    {v0.16b}, [%1], #16         \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp),
-                    "r"(cstep)
-                    : "memory", "v0");
-#else  // NCNN_GNU_INLINE_ASM
-                vst1q_s8(pp, vld1q_s8(p0));
-                pp += 16;
-                p0 += cstep;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-#elif __ARM_FEATURE_DOTPROD
-            for (; kk < max_kk / 8; kk++)
-            {
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #128]       \n"
-                    "ld1    {v0.8b, v1.8b}, [%0], %4    \n"
-                    "st2    {v0.2s, v1.2s}, [%1], #16   \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp),
-                    "r"(cstep)
-                    : "memory", "v0", "v1");
-#else  // NCNN_GNU_INLINE_ASM
-                int32x2x2_t _r01;
-                _r01.val[0] = vreinterpret_s32_s8(vld1_s8(p0));
-                _r01.val[1] = vreinterpret_s32_s8(vld1_s8(p0 + 8));
-                vst2_s32((int*)pp, _r01);
-                pp += 16;
-                p0 += cstep;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-            for (; kk < max_kk / 8; kk++)
-            {
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%0, #128]       \n"
-                    "ld1    {v0.8b, v1.8b}, [%0], %4    \n"
-                    "st2    {v0.4h, v1.4h}, [%1], #16   \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp),
-                    "r"(cstep)
-                    : "memory", "v0", "v1");
-#else
-                asm volatile(
-                    "pld        [%0, #128]          \n"
-                    "vld1.s8    {d0-d1}, [%0], %4   \n"
-                    "vst2.s16   {d0-d1}, [%1 :64]!  \n"
-                    : "=r"(p0), // %0
-                    "=r"(pp)  // %1
-                    : "0"(p0),
-                    "1"(pp),
-                    "r"(cstep)
-                    : "memory", "q0");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-                int16x4x2_t _r01;
-                _r01.val[0] = vreinterpret_s16_s8(vld1_s8(p0));
-                _r01.val[1] = vreinterpret_s16_s8(vld1_s8(p0 + 8));
-                vst2_s16((short*)pp, _r01);
-                pp += 16;
-                p0 += cstep;
-#endif // NCNN_GNU_INLINE_ASM
-            }
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-        }
-#endif // __ARM_NEON
-
-        if (elempack == 1)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k) + (j + jj);
-            const size_t cstep = bottom_blob.cstep;
-
-            int kk = 0;
-#if __ARM_NEON
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[cstep];
-                pp[2] = p0[cstep * 2];
-                pp[3] = p0[cstep * 3];
-                pp[4] = p0[cstep * 4];
-                pp[5] = p0[cstep * 5];
-                pp[6] = p0[cstep * 6];
-                pp[7] = p0[cstep * 7];
-                pp[8] = p0[1];
-                pp[9] = p0[cstep + 1];
-                pp[10] = p0[cstep * 2 + 1];
-                pp[11] = p0[cstep * 3 + 1];
-                pp[12] = p0[cstep * 4 + 1];
-                pp[13] = p0[cstep * 5 + 1];
-                pp[14] = p0[cstep * 6 + 1];
-                pp[15] = p0[cstep * 7 + 1];
-                pp += 16;
-                p0 += cstep * 8;
-            }
-#endif // __ARM_FEATURE_MATMUL_INT8
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[cstep];
-                pp[2] = p0[cstep * 2];
-                pp[3] = p0[cstep * 3];
-                pp[4] = p0[1];
-                pp[5] = p0[cstep + 1];
-                pp[6] = p0[cstep * 2 + 1];
-                pp[7] = p0[cstep * 3 + 1];
-                pp += 8;
-                p0 += cstep * 4;
-            }
-#endif // __ARM_FEATURE_DOTPROD
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[cstep];
-                pp[2] = p0[1];
-                pp[3] = p0[cstep + 1];
-                pp += 4;
-                p0 += cstep * 2;
-            }
-#endif // __ARM_NEON
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                pp += 2;
-                p0 += cstep;
-            }
-        }
-    }
-    for (; jj < max_jj; jj++)
-    {
-#if __ARM_NEON
-        if (elempack == 8)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k / 8) + (j + jj) * 8;
-            const size_t cstep = bottom_blob.cstep * 8;
-
-            int kk = 0;
-            for (; kk < max_kk / 8; kk++)
-            {
-                vst1_s8(pp, vld1_s8(p0));
-                pp += 8;
-                p0 += cstep;
-            }
-        }
-#endif // __ARM_NEON
-
-        if (elempack == 1)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k) + (j + jj);
-            const size_t cstep = bottom_blob.cstep;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp += 1;
-                p0 += cstep;
-            }
-        }
-    }
-}
-
-template<int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h>
-#if __ARM_FEATURE_MATMUL_INT8
-void convolution_im2col_input_tile_int8_i8mm(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk)
-#elif __ARM_FEATURE_DOTPROD
-void convolution_im2col_input_tile_int8_asimddp(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk)
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-void convolution_im2col_input_tile_int8(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk)
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-{
-    const int w = bottom_blob.w;
-    // const int channels = bottom_blob.c;
-    const int elempack = bottom_blob.elempack;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int outw = (w - kernel_extent_w) / stride_w + 1;
-
-    // j max_jj     outw*outh    split w and h
-
-    // k max_kk     pa*maxk*(inch/pa)    split inch
-
-    // k/max_kk shall be multiple of maxk
-
-    const int maxk = kernel_w * kernel_h;
-
-    signed char* pp = B;
-
-    int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-        int dy0 = (j + jj) / outw * stride_h;
-        int dy1 = (j + jj + 1) / outw * stride_h;
-        int dy2 = (j + jj + 2) / outw * stride_h;
-        int dy3 = (j + jj + 3) / outw * stride_h;
-        int dy4 = (j + jj + 4) / outw * stride_h;
-        int dy5 = (j + jj + 5) / outw * stride_h;
-        int dy6 = (j + jj + 6) / outw * stride_h;
-        int dy7 = (j + jj + 7) / outw * stride_h;
-        int dx0 = (j + jj) % outw * stride_w;
-        int dx1 = (j + jj + 1) % outw * stride_w;
-        int dx2 = (j + jj + 2) % outw * stride_w;
-        int dx3 = (j + jj + 3) % outw * stride_w;
-        int dx4 = (j + jj + 4) % outw * stride_w;
-        int dx5 = (j + jj + 5) % outw * stride_w;
-        int dx6 = (j + jj + 6) % outw * stride_w;
-        int dx7 = (j + jj + 7) % outw * stride_w;
-
-        if (dy0 == dy7)
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int p4 = (k + kk + 4) / maxk;
-                    int p5 = (k + kk + 5) / maxk;
-                    int p6 = (k + kk + 6) / maxk;
-                    int p7 = (k + kk + 7) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int uv4 = (k + kk + 4) % maxk;
-                    int uv5 = (k + kk + 5) % maxk;
-                    int uv6 = (k + kk + 6) % maxk;
-                    int uv7 = (k + kk + 7) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int u4 = uv4 / kernel_w;
-                    int u5 = uv5 / kernel_w;
-                    int u6 = uv6 / kernel_w;
-                    int u7 = uv7 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-                    int v4 = uv4 % kernel_w;
-                    int v5 = uv5 % kernel_w;
-                    int v6 = uv6 % kernel_w;
-                    int v7 = uv7 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-                    const Mat img4 = bottom_blob.channel(p4);
-                    const Mat img5 = bottom_blob.channel(p5);
-                    const Mat img6 = bottom_blob.channel(p6);
-                    const Mat img7 = bottom_blob.channel(p7);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-
-                    int x40 = dx0 + dilation_w * v4;
-                    int y40 = dy0 + dilation_h * u4;
-
-                    int x50 = dx0 + dilation_w * v5;
-                    int y50 = dy0 + dilation_h * u5;
-
-                    int x60 = dx0 + dilation_w * v6;
-                    int y60 = dy0 + dilation_h * u6;
-
-                    int x70 = dx0 + dilation_w * v7;
-                    int y70 = dy0 + dilation_h * u7;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr2 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr3 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr4 = img4.row<const signed char>(y40) + x40;
-                    const signed char* sptr5 = img5.row<const signed char>(y50) + x50;
-                    const signed char* sptr6 = img6.row<const signed char>(y60) + x60;
-                    const signed char* sptr7 = img7.row<const signed char>(y70) + x70;
-
-                    if (stride_w == 1)
-                    {
-                        int8x8_t _r0 = vld1_s8(sptr0);
-                        int8x8_t _r1 = vld1_s8(sptr1);
-                        int8x8_t _r2 = vld1_s8(sptr2);
-                        int8x8_t _r3 = vld1_s8(sptr3);
-                        int8x8_t _r4 = vld1_s8(sptr4);
-                        int8x8_t _r5 = vld1_s8(sptr5);
-                        int8x8_t _r6 = vld1_s8(sptr6);
-                        int8x8_t _r7 = vld1_s8(sptr7);
-                        // save as transpose8x8
-                        int8x8x2_t _r01 = vzip_s8(_r0, _r1);
-                        int8x8x2_t _r23 = vzip_s8(_r2, _r3);
-                        int8x8x2_t _r45 = vzip_s8(_r4, _r5);
-                        int8x8x2_t _r67 = vzip_s8(_r6, _r7);
-                        int16x8x4_t _r0246;
-                        _r0246.val[0] = vreinterpretq_s16_s8(vcombine_s8(_r01.val[0], _r01.val[1]));
-                        _r0246.val[1] = vreinterpretq_s16_s8(vcombine_s8(_r23.val[0], _r23.val[1]));
-                        _r0246.val[2] = vreinterpretq_s16_s8(vcombine_s8(_r45.val[0], _r45.val[1]));
-                        _r0246.val[3] = vreinterpretq_s16_s8(vcombine_s8(_r67.val[0], _r67.val[1]));
-                        vst4q_s16((short*)pp, _r0246);
-                        pp += 64;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        int8x16_t _r0 = vld1q_s8(sptr0);
-                        int8x16_t _r1 = vld1q_s8(sptr1);
-                        int8x16_t _r2 = vld1q_s8(sptr2);
-                        int8x16_t _r3 = vld1q_s8(sptr3);
-                        int8x16_t _r4 = vld1q_s8(sptr4);
-                        int8x16_t _r5 = vld1q_s8(sptr5);
-                        int8x16_t _r6 = vld1q_s8(sptr6);
-                        int8x16_t _r7 = vld1q_s8(sptr7);
-                        int8x16_t _r01 = vtrnq_s8(_r0, _r1).val[0];
-                        int8x16_t _r23 = vtrnq_s8(_r2, _r3).val[0];
-                        int8x16_t _r45 = vtrnq_s8(_r4, _r5).val[0];
-                        int8x16_t _r67 = vtrnq_s8(_r6, _r7).val[0];
-                        int16x8x4_t _r0123;
-                        _r0123.val[0] = vreinterpretq_s16_s8(_r01);
-                        _r0123.val[1] = vreinterpretq_s16_s8(_r23);
-                        _r0123.val[2] = vreinterpretq_s16_s8(_r45);
-                        _r0123.val[3] = vreinterpretq_s16_s8(_r67);
-                        vst4q_s16((short*)pp, _r0123);
-                        pp += 64;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr2[0];
-                        pp[3] = sptr3[0];
-                        pp[4] = sptr4[0];
-                        pp[5] = sptr5[0];
-                        pp[6] = sptr6[0];
-                        pp[7] = sptr7[0];
-                        pp[8] = sptr0[stride_w];
-                        pp[9] = sptr1[stride_w];
-                        pp[10] = sptr2[stride_w];
-                        pp[11] = sptr3[stride_w];
-                        pp[12] = sptr4[stride_w];
-                        pp[13] = sptr5[stride_w];
-                        pp[14] = sptr6[stride_w];
-                        pp[15] = sptr7[stride_w];
-                        pp[16] = sptr0[stride_w * 2];
-                        pp[17] = sptr1[stride_w * 2];
-                        pp[18] = sptr2[stride_w * 2];
-                        pp[19] = sptr3[stride_w * 2];
-                        pp[20] = sptr4[stride_w * 2];
-                        pp[21] = sptr5[stride_w * 2];
-                        pp[22] = sptr6[stride_w * 2];
-                        pp[23] = sptr7[stride_w * 2];
-                        pp[24] = sptr0[stride_w * 3];
-                        pp[25] = sptr1[stride_w * 3];
-                        pp[26] = sptr2[stride_w * 3];
-                        pp[27] = sptr3[stride_w * 3];
-                        pp[28] = sptr4[stride_w * 3];
-                        pp[29] = sptr5[stride_w * 3];
-                        pp[30] = sptr6[stride_w * 3];
-                        pp[31] = sptr7[stride_w * 3];
-                        pp[32] = sptr0[stride_w * 4];
-                        pp[33] = sptr1[stride_w * 4];
-                        pp[34] = sptr2[stride_w * 4];
-                        pp[35] = sptr3[stride_w * 4];
-                        pp[36] = sptr4[stride_w * 4];
-                        pp[37] = sptr5[stride_w * 4];
-                        pp[38] = sptr6[stride_w * 4];
-                        pp[39] = sptr7[stride_w * 4];
-                        pp[40] = sptr0[stride_w * 5];
-                        pp[41] = sptr1[stride_w * 5];
-                        pp[42] = sptr2[stride_w * 5];
-                        pp[43] = sptr3[stride_w * 5];
-                        pp[44] = sptr4[stride_w * 5];
-                        pp[45] = sptr5[stride_w * 5];
-                        pp[46] = sptr6[stride_w * 5];
-                        pp[47] = sptr7[stride_w * 5];
-                        pp[48] = sptr0[stride_w * 6];
-                        pp[49] = sptr1[stride_w * 6];
-                        pp[50] = sptr2[stride_w * 6];
-                        pp[51] = sptr3[stride_w * 6];
-                        pp[52] = sptr4[stride_w * 6];
-                        pp[53] = sptr5[stride_w * 6];
-                        pp[54] = sptr6[stride_w * 6];
-                        pp[55] = sptr7[stride_w * 6];
-                        pp[56] = sptr0[stride_w * 7];
-                        pp[57] = sptr1[stride_w * 7];
-                        pp[58] = sptr2[stride_w * 7];
-                        pp[59] = sptr3[stride_w * 7];
-                        pp[60] = sptr4[stride_w * 7];
-                        pp[61] = sptr5[stride_w * 7];
-                        pp[62] = sptr6[stride_w * 7];
-                        pp[63] = sptr7[stride_w * 7];
-                        pp += 64;
-                    }
-                }
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr2 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr3 = img3.row<const signed char>(y30) + x30;
-
-                    if (stride_w == 1)
-                    {
-                        int8x8x4_t _r01;
-                        _r01.val[0] = vld1_s8(sptr0);
-                        _r01.val[1] = vld1_s8(sptr1);
-                        _r01.val[2] = vld1_s8(sptr2);
-                        _r01.val[3] = vld1_s8(sptr3);
-                        vst4_s8(pp, _r01);
-                        pp += 32;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        int8x16_t _r0 = vld1q_s8(sptr0);
-                        int8x16_t _r1 = vld1q_s8(sptr1);
-                        int8x16_t _r2 = vld1q_s8(sptr2);
-                        int8x16_t _r3 = vld1q_s8(sptr3);
-                        int8x16_t _r01 = vtrnq_s8(_r0, _r1).val[0];
-                        int8x16_t _r23 = vtrnq_s8(_r2, _r3).val[0];
-                        int16x8x2_t _r0123;
-                        _r0123.val[0] = vreinterpretq_s16_s8(_r01);
-                        _r0123.val[1] = vreinterpretq_s16_s8(_r23);
-                        vst2q_s16((short*)pp, _r0123);
-                        pp += 32;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr2[0];
-                        pp[3] = sptr3[0];
-                        pp[4] = sptr0[stride_w];
-                        pp[5] = sptr1[stride_w];
-                        pp[6] = sptr2[stride_w];
-                        pp[7] = sptr3[stride_w];
-                        pp[8] = sptr0[stride_w * 2];
-                        pp[9] = sptr1[stride_w * 2];
-                        pp[10] = sptr2[stride_w * 2];
-                        pp[11] = sptr3[stride_w * 2];
-                        pp[12] = sptr0[stride_w * 3];
-                        pp[13] = sptr1[stride_w * 3];
-                        pp[14] = sptr2[stride_w * 3];
-                        pp[15] = sptr3[stride_w * 3];
-                        pp[16] = sptr0[stride_w * 4];
-                        pp[17] = sptr1[stride_w * 4];
-                        pp[18] = sptr2[stride_w * 4];
-                        pp[19] = sptr3[stride_w * 4];
-                        pp[20] = sptr0[stride_w * 5];
-                        pp[21] = sptr1[stride_w * 5];
-                        pp[22] = sptr2[stride_w * 5];
-                        pp[23] = sptr3[stride_w * 5];
-                        pp[24] = sptr0[stride_w * 6];
-                        pp[25] = sptr1[stride_w * 6];
-                        pp[26] = sptr2[stride_w * 6];
-                        pp[27] = sptr3[stride_w * 6];
-                        pp[28] = sptr0[stride_w * 7];
-                        pp[29] = sptr1[stride_w * 7];
-                        pp[30] = sptr2[stride_w * 7];
-                        pp[31] = sptr3[stride_w * 7];
-                        pp += 32;
-                    }
-                }
-#endif // __ARM_FEATURE_DOTPROD
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-
-                    if (stride_w == 1)
-                    {
-                        int8x8x2_t _r01;
-                        _r01.val[0] = vld1_s8(sptr0);
-                        _r01.val[1] = vld1_s8(sptr1);
-                        vst2_s8(pp, _r01);
-                        pp += 16;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        int8x16_t _r0 = vld1q_s8(sptr0);
-                        int8x16_t _r1 = vld1q_s8(sptr1);
-                        int8x16_t _r01 = vtrnq_s8(_r0, _r1).val[0];
-                        vst1q_s8(pp, _r01);
-                        pp += 16;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr0[stride_w];
-                        pp[3] = sptr1[stride_w];
-                        pp[4] = sptr0[stride_w * 2];
-                        pp[5] = sptr1[stride_w * 2];
-                        pp[6] = sptr0[stride_w * 3];
-                        pp[7] = sptr1[stride_w * 3];
-                        pp[8] = sptr0[stride_w * 4];
-                        pp[9] = sptr1[stride_w * 4];
-                        pp[10] = sptr0[stride_w * 5];
-                        pp[11] = sptr1[stride_w * 5];
-                        pp[12] = sptr0[stride_w * 6];
-                        pp[13] = sptr1[stride_w * 6];
-                        pp[14] = sptr0[stride_w * 7];
-                        pp[15] = sptr1[stride_w * 7];
-                        pp += 16;
-                    }
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = dx0 + dilation_w * v;
-                int y0 = dy0 + dilation_h * u;
-
-                const signed char* sptr = img.row<const signed char>(y0) + x0 * elempack;
-
-                if (elempack == 8)
-                {
-#if __ARM_FEATURE_MATMUL_INT8
-                    int8x8_t _r0 = vld1_s8(sptr);
-                    int8x8_t _r1 = vld1_s8(sptr + stride_w * 8);
-                    int8x8_t _r2 = vld1_s8(sptr + stride_w * 16);
-                    int8x8_t _r3 = vld1_s8(sptr + stride_w * 24);
-                    int8x8_t _r4 = vld1_s8(sptr + stride_w * 32);
-                    int8x8_t _r5 = vld1_s8(sptr + stride_w * 40);
-                    int8x8_t _r6 = vld1_s8(sptr + stride_w * 48);
-                    int8x8_t _r7 = vld1_s8(sptr + stride_w * 56);
-                    vst1_s8(pp, _r0);
-                    vst1_s8(pp + 8, _r1);
-                    vst1_s8(pp + 16, _r2);
-                    vst1_s8(pp + 24, _r3);
-                    vst1_s8(pp + 32, _r4);
-                    vst1_s8(pp + 40, _r5);
-                    vst1_s8(pp + 48, _r6);
-                    vst1_s8(pp + 56, _r7);
-                    pp += 64;
-#elif __ARM_FEATURE_DOTPROD
-                    int32x2_t _r0 = vreinterpret_s32_s8(vld1_s8(sptr));
-                    int32x2_t _r1 = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 8));
-                    int32x2_t _r2 = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 16));
-                    int32x2_t _r3 = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 24));
-                    int32x2_t _r4 = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 32));
-                    int32x2_t _r5 = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 40));
-                    int32x2_t _r6 = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 48));
-                    int32x2_t _r7 = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 56));
-                    int32x2x2_t _r01 = vzip_s32(_r0, _r1);
-                    int32x2x2_t _r23 = vzip_s32(_r2, _r3);
-                    int32x2x2_t _r45 = vzip_s32(_r4, _r5);
-                    int32x2x2_t _r67 = vzip_s32(_r6, _r7);
-                    vst1_s32((int*)pp, _r01.val[0]);
-                    vst1_s32((int*)(pp + 8), _r23.val[0]);
-                    vst1_s32((int*)(pp + 16), _r45.val[0]);
-                    vst1_s32((int*)(pp + 24), _r67.val[0]);
-                    vst1_s32((int*)(pp + 32), _r01.val[1]);
-                    vst1_s32((int*)(pp + 40), _r23.val[1]);
-                    vst1_s32((int*)(pp + 48), _r45.val[1]);
-                    vst1_s32((int*)(pp + 56), _r67.val[1]);
-                    pp += 64;
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                    int16x4_t _r0 = vreinterpret_s16_s8(vld1_s8(sptr));
-                    int16x4_t _r1 = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 8));
-                    int16x4_t _r2 = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 16));
-                    int16x4_t _r3 = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 24));
-                    int16x4_t _r4 = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 32));
-                    int16x4_t _r5 = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 40));
-                    int16x4_t _r6 = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 48));
-                    int16x4_t _r7 = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 56));
-                    int16x4x2_t _r01 = vzip_s16(_r0, _r1);
-                    int16x4x2_t _r23 = vzip_s16(_r2, _r3);
-                    int16x4x2_t _r45 = vzip_s16(_r4, _r5);
-                    int16x4x2_t _r67 = vzip_s16(_r6, _r7);
-                    int32x4x4_t _r0123;
-                    _r0123.val[0] = vreinterpretq_s32_s16(vcombine_s16(_r01.val[0], _r01.val[1]));
-                    _r0123.val[1] = vreinterpretq_s32_s16(vcombine_s16(_r23.val[0], _r23.val[1]));
-                    _r0123.val[2] = vreinterpretq_s32_s16(vcombine_s16(_r45.val[0], _r45.val[1]));
-                    _r0123.val[3] = vreinterpretq_s32_s16(vcombine_s16(_r67.val[0], _r67.val[1]));
-                    vst4q_s32((int*)pp, _r0123);
-                    pp += 64;
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr[0];
-                    pp[1] = sptr[stride_w];
-                    pp[2] = sptr[stride_w * 2];
-                    pp[3] = sptr[stride_w * 3];
-                    pp[4] = sptr[stride_w * 4];
-                    pp[5] = sptr[stride_w * 5];
-                    pp[6] = sptr[stride_w * 6];
-                    pp[7] = sptr[stride_w * 7];
-                    pp += 8;
-                }
-            }
-        }
-        else
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int p4 = (k + kk + 4) / maxk;
-                    int p5 = (k + kk + 5) / maxk;
-                    int p6 = (k + kk + 6) / maxk;
-                    int p7 = (k + kk + 7) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int uv4 = (k + kk + 4) % maxk;
-                    int uv5 = (k + kk + 5) % maxk;
-                    int uv6 = (k + kk + 6) % maxk;
-                    int uv7 = (k + kk + 7) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int u4 = uv4 / kernel_w;
-                    int u5 = uv5 / kernel_w;
-                    int u6 = uv6 / kernel_w;
-                    int u7 = uv7 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-                    int v4 = uv4 % kernel_w;
-                    int v5 = uv5 % kernel_w;
-                    int v6 = uv6 % kernel_w;
-                    int v7 = uv7 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-                    const Mat img4 = bottom_blob.channel(p4);
-                    const Mat img5 = bottom_blob.channel(p5);
-                    const Mat img6 = bottom_blob.channel(p6);
-                    const Mat img7 = bottom_blob.channel(p7);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int x02 = dx2 + dilation_w * v0;
-                    int x03 = dx3 + dilation_w * v0;
-                    int x04 = dx4 + dilation_w * v0;
-                    int x05 = dx5 + dilation_w * v0;
-                    int x06 = dx6 + dilation_w * v0;
-                    int x07 = dx7 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int y02 = dy2 + dilation_h * u0;
-                    int y03 = dy3 + dilation_h * u0;
-                    int y04 = dy4 + dilation_h * u0;
-                    int y05 = dy5 + dilation_h * u0;
-                    int y06 = dy6 + dilation_h * u0;
-                    int y07 = dy7 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int x12 = dx2 + dilation_w * v1;
-                    int x13 = dx3 + dilation_w * v1;
-                    int x14 = dx4 + dilation_w * v1;
-                    int x15 = dx5 + dilation_w * v1;
-                    int x16 = dx6 + dilation_w * v1;
-                    int x17 = dx7 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-                    int y12 = dy2 + dilation_h * u1;
-                    int y13 = dy3 + dilation_h * u1;
-                    int y14 = dy4 + dilation_h * u1;
-                    int y15 = dy5 + dilation_h * u1;
-                    int y16 = dy6 + dilation_h * u1;
-                    int y17 = dy7 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int x21 = dx1 + dilation_w * v2;
-                    int x22 = dx2 + dilation_w * v2;
-                    int x23 = dx3 + dilation_w * v2;
-                    int x24 = dx4 + dilation_w * v2;
-                    int x25 = dx5 + dilation_w * v2;
-                    int x26 = dx6 + dilation_w * v2;
-                    int x27 = dx7 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int y21 = dy1 + dilation_h * u2;
-                    int y22 = dy2 + dilation_h * u2;
-                    int y23 = dy3 + dilation_h * u2;
-                    int y24 = dy4 + dilation_h * u2;
-                    int y25 = dy5 + dilation_h * u2;
-                    int y26 = dy6 + dilation_h * u2;
-                    int y27 = dy7 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int x31 = dx1 + dilation_w * v3;
-                    int x32 = dx2 + dilation_w * v3;
-                    int x33 = dx3 + dilation_w * v3;
-                    int x34 = dx4 + dilation_w * v3;
-                    int x35 = dx5 + dilation_w * v3;
-                    int x36 = dx6 + dilation_w * v3;
-                    int x37 = dx7 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-                    int y31 = dy1 + dilation_h * u3;
-                    int y32 = dy2 + dilation_h * u3;
-                    int y33 = dy3 + dilation_h * u3;
-                    int y34 = dy4 + dilation_h * u3;
-                    int y35 = dy5 + dilation_h * u3;
-                    int y36 = dy6 + dilation_h * u3;
-                    int y37 = dy7 + dilation_h * u3;
-
-                    int x40 = dx0 + dilation_w * v4;
-                    int x41 = dx1 + dilation_w * v4;
-                    int x42 = dx2 + dilation_w * v4;
-                    int x43 = dx3 + dilation_w * v4;
-                    int x44 = dx4 + dilation_w * v4;
-                    int x45 = dx5 + dilation_w * v4;
-                    int x46 = dx6 + dilation_w * v4;
-                    int x47 = dx7 + dilation_w * v4;
-                    int y40 = dy0 + dilation_h * u4;
-                    int y41 = dy1 + dilation_h * u4;
-                    int y42 = dy2 + dilation_h * u4;
-                    int y43 = dy3 + dilation_h * u4;
-                    int y44 = dy4 + dilation_h * u4;
-                    int y45 = dy5 + dilation_h * u4;
-                    int y46 = dy6 + dilation_h * u4;
-                    int y47 = dy7 + dilation_h * u4;
-
-                    int x50 = dx0 + dilation_w * v5;
-                    int x51 = dx1 + dilation_w * v5;
-                    int x52 = dx2 + dilation_w * v5;
-                    int x53 = dx3 + dilation_w * v5;
-                    int x54 = dx4 + dilation_w * v5;
-                    int x55 = dx5 + dilation_w * v5;
-                    int x56 = dx6 + dilation_w * v5;
-                    int x57 = dx7 + dilation_w * v5;
-                    int y50 = dy0 + dilation_h * u5;
-                    int y51 = dy1 + dilation_h * u5;
-                    int y52 = dy2 + dilation_h * u5;
-                    int y53 = dy3 + dilation_h * u5;
-                    int y54 = dy4 + dilation_h * u5;
-                    int y55 = dy5 + dilation_h * u5;
-                    int y56 = dy6 + dilation_h * u5;
-                    int y57 = dy7 + dilation_h * u5;
-
-                    int x60 = dx0 + dilation_w * v6;
-                    int x61 = dx1 + dilation_w * v6;
-                    int x62 = dx2 + dilation_w * v6;
-                    int x63 = dx3 + dilation_w * v6;
-                    int x64 = dx4 + dilation_w * v6;
-                    int x65 = dx5 + dilation_w * v6;
-                    int x66 = dx6 + dilation_w * v6;
-                    int x67 = dx7 + dilation_w * v6;
-                    int y60 = dy0 + dilation_h * u6;
-                    int y61 = dy1 + dilation_h * u6;
-                    int y62 = dy2 + dilation_h * u6;
-                    int y63 = dy3 + dilation_h * u6;
-                    int y64 = dy4 + dilation_h * u6;
-                    int y65 = dy5 + dilation_h * u6;
-                    int y66 = dy6 + dilation_h * u6;
-                    int y67 = dy7 + dilation_h * u6;
-
-                    int x70 = dx0 + dilation_w * v7;
-                    int x71 = dx1 + dilation_w * v7;
-                    int x72 = dx2 + dilation_w * v7;
-                    int x73 = dx3 + dilation_w * v7;
-                    int x74 = dx4 + dilation_w * v7;
-                    int x75 = dx5 + dilation_w * v7;
-                    int x76 = dx6 + dilation_w * v7;
-                    int x77 = dx7 + dilation_w * v7;
-                    int y70 = dy0 + dilation_h * u7;
-                    int y71 = dy1 + dilation_h * u7;
-                    int y72 = dy2 + dilation_h * u7;
-                    int y73 = dy3 + dilation_h * u7;
-                    int y74 = dy4 + dilation_h * u7;
-                    int y75 = dy5 + dilation_h * u7;
-                    int y76 = dy6 + dilation_h * u7;
-                    int y77 = dy7 + dilation_h * u7;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-                    const signed char* sptr04 = img0.row<const signed char>(y04) + x04;
-                    const signed char* sptr05 = img0.row<const signed char>(y05) + x05;
-                    const signed char* sptr06 = img0.row<const signed char>(y06) + x06;
-                    const signed char* sptr07 = img0.row<const signed char>(y07) + x07;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-                    const signed char* sptr14 = img1.row<const signed char>(y14) + x14;
-                    const signed char* sptr15 = img1.row<const signed char>(y15) + x15;
-                    const signed char* sptr16 = img1.row<const signed char>(y16) + x16;
-                    const signed char* sptr17 = img1.row<const signed char>(y17) + x17;
-
-                    const signed char* sptr20 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr21 = img2.row<const signed char>(y21) + x21;
-                    const signed char* sptr22 = img2.row<const signed char>(y22) + x22;
-                    const signed char* sptr23 = img2.row<const signed char>(y23) + x23;
-                    const signed char* sptr24 = img2.row<const signed char>(y24) + x24;
-                    const signed char* sptr25 = img2.row<const signed char>(y25) + x25;
-                    const signed char* sptr26 = img2.row<const signed char>(y26) + x26;
-                    const signed char* sptr27 = img2.row<const signed char>(y27) + x27;
-
-                    const signed char* sptr30 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr31 = img3.row<const signed char>(y31) + x31;
-                    const signed char* sptr32 = img3.row<const signed char>(y32) + x32;
-                    const signed char* sptr33 = img3.row<const signed char>(y33) + x33;
-                    const signed char* sptr34 = img3.row<const signed char>(y34) + x34;
-                    const signed char* sptr35 = img3.row<const signed char>(y35) + x35;
-                    const signed char* sptr36 = img3.row<const signed char>(y36) + x36;
-                    const signed char* sptr37 = img3.row<const signed char>(y37) + x37;
-
-                    const signed char* sptr40 = img4.row<const signed char>(y40) + x40;
-                    const signed char* sptr41 = img4.row<const signed char>(y41) + x41;
-                    const signed char* sptr42 = img4.row<const signed char>(y42) + x42;
-                    const signed char* sptr43 = img4.row<const signed char>(y43) + x43;
-                    const signed char* sptr44 = img4.row<const signed char>(y44) + x44;
-                    const signed char* sptr45 = img4.row<const signed char>(y45) + x45;
-                    const signed char* sptr46 = img4.row<const signed char>(y46) + x46;
-                    const signed char* sptr47 = img4.row<const signed char>(y47) + x47;
-
-                    const signed char* sptr50 = img5.row<const signed char>(y50) + x50;
-                    const signed char* sptr51 = img5.row<const signed char>(y51) + x51;
-                    const signed char* sptr52 = img5.row<const signed char>(y52) + x52;
-                    const signed char* sptr53 = img5.row<const signed char>(y53) + x53;
-                    const signed char* sptr54 = img5.row<const signed char>(y54) + x54;
-                    const signed char* sptr55 = img5.row<const signed char>(y55) + x55;
-                    const signed char* sptr56 = img5.row<const signed char>(y56) + x56;
-                    const signed char* sptr57 = img5.row<const signed char>(y57) + x57;
-
-                    const signed char* sptr60 = img6.row<const signed char>(y60) + x60;
-                    const signed char* sptr61 = img6.row<const signed char>(y61) + x61;
-                    const signed char* sptr62 = img6.row<const signed char>(y62) + x62;
-                    const signed char* sptr63 = img6.row<const signed char>(y63) + x63;
-                    const signed char* sptr64 = img6.row<const signed char>(y64) + x64;
-                    const signed char* sptr65 = img6.row<const signed char>(y65) + x65;
-                    const signed char* sptr66 = img6.row<const signed char>(y66) + x66;
-                    const signed char* sptr67 = img6.row<const signed char>(y67) + x67;
-
-                    const signed char* sptr70 = img7.row<const signed char>(y70) + x70;
-                    const signed char* sptr71 = img7.row<const signed char>(y71) + x71;
-                    const signed char* sptr72 = img7.row<const signed char>(y72) + x72;
-                    const signed char* sptr73 = img7.row<const signed char>(y73) + x73;
-                    const signed char* sptr74 = img7.row<const signed char>(y74) + x74;
-                    const signed char* sptr75 = img7.row<const signed char>(y75) + x75;
-                    const signed char* sptr76 = img7.row<const signed char>(y76) + x76;
-                    const signed char* sptr77 = img7.row<const signed char>(y77) + x77;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr20[0];
-                    pp[3] = sptr30[0];
-                    pp[4] = sptr40[0];
-                    pp[5] = sptr50[0];
-                    pp[6] = sptr60[0];
-                    pp[7] = sptr70[0];
-                    pp[8] = sptr01[0];
-                    pp[9] = sptr11[0];
-                    pp[10] = sptr21[0];
-                    pp[11] = sptr31[0];
-                    pp[12] = sptr41[0];
-                    pp[13] = sptr51[0];
-                    pp[14] = sptr61[0];
-                    pp[15] = sptr71[0];
-                    pp[16] = sptr02[0];
-                    pp[17] = sptr12[0];
-                    pp[18] = sptr22[0];
-                    pp[19] = sptr32[0];
-                    pp[20] = sptr42[0];
-                    pp[21] = sptr52[0];
-                    pp[22] = sptr62[0];
-                    pp[23] = sptr72[0];
-                    pp[24] = sptr03[0];
-                    pp[25] = sptr13[0];
-                    pp[26] = sptr23[0];
-                    pp[27] = sptr33[0];
-                    pp[28] = sptr43[0];
-                    pp[29] = sptr53[0];
-                    pp[30] = sptr63[0];
-                    pp[31] = sptr73[0];
-                    pp[32] = sptr04[0];
-                    pp[33] = sptr14[0];
-                    pp[34] = sptr24[0];
-                    pp[35] = sptr34[0];
-                    pp[36] = sptr44[0];
-                    pp[37] = sptr54[0];
-                    pp[38] = sptr64[0];
-                    pp[39] = sptr74[0];
-                    pp[40] = sptr05[0];
-                    pp[41] = sptr15[0];
-                    pp[42] = sptr25[0];
-                    pp[43] = sptr35[0];
-                    pp[44] = sptr45[0];
-                    pp[45] = sptr55[0];
-                    pp[46] = sptr65[0];
-                    pp[47] = sptr75[0];
-                    pp[48] = sptr06[0];
-                    pp[49] = sptr16[0];
-                    pp[50] = sptr26[0];
-                    pp[51] = sptr36[0];
-                    pp[52] = sptr46[0];
-                    pp[53] = sptr56[0];
-                    pp[54] = sptr66[0];
-                    pp[55] = sptr76[0];
-                    pp[56] = sptr07[0];
-                    pp[57] = sptr17[0];
-                    pp[58] = sptr27[0];
-                    pp[59] = sptr37[0];
-                    pp[60] = sptr47[0];
-                    pp[61] = sptr57[0];
-                    pp[62] = sptr67[0];
-                    pp[63] = sptr77[0];
-                    pp += 64;
-                }
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int x02 = dx2 + dilation_w * v0;
-                    int x03 = dx3 + dilation_w * v0;
-                    int x04 = dx4 + dilation_w * v0;
-                    int x05 = dx5 + dilation_w * v0;
-                    int x06 = dx6 + dilation_w * v0;
-                    int x07 = dx7 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int y02 = dy2 + dilation_h * u0;
-                    int y03 = dy3 + dilation_h * u0;
-                    int y04 = dy4 + dilation_h * u0;
-                    int y05 = dy5 + dilation_h * u0;
-                    int y06 = dy6 + dilation_h * u0;
-                    int y07 = dy7 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int x12 = dx2 + dilation_w * v1;
-                    int x13 = dx3 + dilation_w * v1;
-                    int x14 = dx4 + dilation_w * v1;
-                    int x15 = dx5 + dilation_w * v1;
-                    int x16 = dx6 + dilation_w * v1;
-                    int x17 = dx7 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-                    int y12 = dy2 + dilation_h * u1;
-                    int y13 = dy3 + dilation_h * u1;
-                    int y14 = dy4 + dilation_h * u1;
-                    int y15 = dy5 + dilation_h * u1;
-                    int y16 = dy6 + dilation_h * u1;
-                    int y17 = dy7 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int x21 = dx1 + dilation_w * v2;
-                    int x22 = dx2 + dilation_w * v2;
-                    int x23 = dx3 + dilation_w * v2;
-                    int x24 = dx4 + dilation_w * v2;
-                    int x25 = dx5 + dilation_w * v2;
-                    int x26 = dx6 + dilation_w * v2;
-                    int x27 = dx7 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int y21 = dy1 + dilation_h * u2;
-                    int y22 = dy2 + dilation_h * u2;
-                    int y23 = dy3 + dilation_h * u2;
-                    int y24 = dy4 + dilation_h * u2;
-                    int y25 = dy5 + dilation_h * u2;
-                    int y26 = dy6 + dilation_h * u2;
-                    int y27 = dy7 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int x31 = dx1 + dilation_w * v3;
-                    int x32 = dx2 + dilation_w * v3;
-                    int x33 = dx3 + dilation_w * v3;
-                    int x34 = dx4 + dilation_w * v3;
-                    int x35 = dx5 + dilation_w * v3;
-                    int x36 = dx6 + dilation_w * v3;
-                    int x37 = dx7 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-                    int y31 = dy1 + dilation_h * u3;
-                    int y32 = dy2 + dilation_h * u3;
-                    int y33 = dy3 + dilation_h * u3;
-                    int y34 = dy4 + dilation_h * u3;
-                    int y35 = dy5 + dilation_h * u3;
-                    int y36 = dy6 + dilation_h * u3;
-                    int y37 = dy7 + dilation_h * u3;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-                    const signed char* sptr04 = img0.row<const signed char>(y04) + x04;
-                    const signed char* sptr05 = img0.row<const signed char>(y05) + x05;
-                    const signed char* sptr06 = img0.row<const signed char>(y06) + x06;
-                    const signed char* sptr07 = img0.row<const signed char>(y07) + x07;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-                    const signed char* sptr14 = img1.row<const signed char>(y14) + x14;
-                    const signed char* sptr15 = img1.row<const signed char>(y15) + x15;
-                    const signed char* sptr16 = img1.row<const signed char>(y16) + x16;
-                    const signed char* sptr17 = img1.row<const signed char>(y17) + x17;
-
-                    const signed char* sptr20 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr21 = img2.row<const signed char>(y21) + x21;
-                    const signed char* sptr22 = img2.row<const signed char>(y22) + x22;
-                    const signed char* sptr23 = img2.row<const signed char>(y23) + x23;
-                    const signed char* sptr24 = img2.row<const signed char>(y24) + x24;
-                    const signed char* sptr25 = img2.row<const signed char>(y25) + x25;
-                    const signed char* sptr26 = img2.row<const signed char>(y26) + x26;
-                    const signed char* sptr27 = img2.row<const signed char>(y27) + x27;
-
-                    const signed char* sptr30 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr31 = img3.row<const signed char>(y31) + x31;
-                    const signed char* sptr32 = img3.row<const signed char>(y32) + x32;
-                    const signed char* sptr33 = img3.row<const signed char>(y33) + x33;
-                    const signed char* sptr34 = img3.row<const signed char>(y34) + x34;
-                    const signed char* sptr35 = img3.row<const signed char>(y35) + x35;
-                    const signed char* sptr36 = img3.row<const signed char>(y36) + x36;
-                    const signed char* sptr37 = img3.row<const signed char>(y37) + x37;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr20[0];
-                    pp[3] = sptr30[0];
-                    pp[4] = sptr01[0];
-                    pp[5] = sptr11[0];
-                    pp[6] = sptr21[0];
-                    pp[7] = sptr31[0];
-                    pp[8] = sptr02[0];
-                    pp[9] = sptr12[0];
-                    pp[10] = sptr22[0];
-                    pp[11] = sptr32[0];
-                    pp[12] = sptr03[0];
-                    pp[13] = sptr13[0];
-                    pp[14] = sptr23[0];
-                    pp[15] = sptr33[0];
-                    pp[16] = sptr04[0];
-                    pp[17] = sptr14[0];
-                    pp[18] = sptr24[0];
-                    pp[19] = sptr34[0];
-                    pp[20] = sptr05[0];
-                    pp[21] = sptr15[0];
-                    pp[22] = sptr25[0];
-                    pp[23] = sptr35[0];
-                    pp[24] = sptr06[0];
-                    pp[25] = sptr16[0];
-                    pp[26] = sptr26[0];
-                    pp[27] = sptr36[0];
-                    pp[28] = sptr07[0];
-                    pp[29] = sptr17[0];
-                    pp[30] = sptr27[0];
-                    pp[31] = sptr37[0];
-                    pp += 32;
-                }
-#endif // __ARM_FEATURE_DOTPROD
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int x02 = dx2 + dilation_w * v0;
-                    int x03 = dx3 + dilation_w * v0;
-                    int x04 = dx4 + dilation_w * v0;
-                    int x05 = dx5 + dilation_w * v0;
-                    int x06 = dx6 + dilation_w * v0;
-                    int x07 = dx7 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int y02 = dy2 + dilation_h * u0;
-                    int y03 = dy3 + dilation_h * u0;
-                    int y04 = dy4 + dilation_h * u0;
-                    int y05 = dy5 + dilation_h * u0;
-                    int y06 = dy6 + dilation_h * u0;
-                    int y07 = dy7 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int x12 = dx2 + dilation_w * v1;
-                    int x13 = dx3 + dilation_w * v1;
-                    int x14 = dx4 + dilation_w * v1;
-                    int x15 = dx5 + dilation_w * v1;
-                    int x16 = dx6 + dilation_w * v1;
-                    int x17 = dx7 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-                    int y12 = dy2 + dilation_h * u1;
-                    int y13 = dy3 + dilation_h * u1;
-                    int y14 = dy4 + dilation_h * u1;
-                    int y15 = dy5 + dilation_h * u1;
-                    int y16 = dy6 + dilation_h * u1;
-                    int y17 = dy7 + dilation_h * u1;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-                    const signed char* sptr04 = img0.row<const signed char>(y04) + x04;
-                    const signed char* sptr05 = img0.row<const signed char>(y05) + x05;
-                    const signed char* sptr06 = img0.row<const signed char>(y06) + x06;
-                    const signed char* sptr07 = img0.row<const signed char>(y07) + x07;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-                    const signed char* sptr14 = img1.row<const signed char>(y14) + x14;
-                    const signed char* sptr15 = img1.row<const signed char>(y15) + x15;
-                    const signed char* sptr16 = img1.row<const signed char>(y16) + x16;
-                    const signed char* sptr17 = img1.row<const signed char>(y17) + x17;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr01[0];
-                    pp[3] = sptr11[0];
-                    pp[4] = sptr02[0];
-                    pp[5] = sptr12[0];
-                    pp[6] = sptr03[0];
-                    pp[7] = sptr13[0];
-                    pp[8] = sptr04[0];
-                    pp[9] = sptr14[0];
-                    pp[10] = sptr05[0];
-                    pp[11] = sptr15[0];
-                    pp[12] = sptr06[0];
-                    pp[13] = sptr16[0];
-                    pp[14] = sptr07[0];
-                    pp[15] = sptr17[0];
-                    pp += 16;
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = dx0 + dilation_w * v;
-                int x1 = dx1 + dilation_w * v;
-                int x2 = dx2 + dilation_w * v;
-                int x3 = dx3 + dilation_w * v;
-                int x4 = dx4 + dilation_w * v;
-                int x5 = dx5 + dilation_w * v;
-                int x6 = dx6 + dilation_w * v;
-                int x7 = dx7 + dilation_w * v;
-                int y0 = dy0 + dilation_h * u;
-                int y1 = dy1 + dilation_h * u;
-                int y2 = dy2 + dilation_h * u;
-                int y3 = dy3 + dilation_h * u;
-                int y4 = dy4 + dilation_h * u;
-                int y5 = dy5 + dilation_h * u;
-                int y6 = dy6 + dilation_h * u;
-                int y7 = dy7 + dilation_h * u;
-
-                const signed char* sptr0 = img.row<const signed char>(y0) + x0 * elempack;
-                const signed char* sptr1 = img.row<const signed char>(y1) + x1 * elempack;
-                const signed char* sptr2 = img.row<const signed char>(y2) + x2 * elempack;
-                const signed char* sptr3 = img.row<const signed char>(y3) + x3 * elempack;
-                const signed char* sptr4 = img.row<const signed char>(y4) + x4 * elempack;
-                const signed char* sptr5 = img.row<const signed char>(y5) + x5 * elempack;
-                const signed char* sptr6 = img.row<const signed char>(y6) + x6 * elempack;
-                const signed char* sptr7 = img.row<const signed char>(y7) + x7 * elempack;
-
-                if (elempack == 8)
-                {
-#if __ARM_FEATURE_MATMUL_INT8
-                    int8x8_t _r0 = vld1_s8(sptr0);
-                    int8x8_t _r1 = vld1_s8(sptr1);
-                    int8x8_t _r2 = vld1_s8(sptr2);
-                    int8x8_t _r3 = vld1_s8(sptr3);
-                    int8x8_t _r4 = vld1_s8(sptr4);
-                    int8x8_t _r5 = vld1_s8(sptr5);
-                    int8x8_t _r6 = vld1_s8(sptr6);
-                    int8x8_t _r7 = vld1_s8(sptr7);
-                    vst1_s8(pp, _r0);
-                    vst1_s8(pp + 8, _r1);
-                    vst1_s8(pp + 16, _r2);
-                    vst1_s8(pp + 24, _r3);
-                    vst1_s8(pp + 32, _r4);
-                    vst1_s8(pp + 40, _r5);
-                    vst1_s8(pp + 48, _r6);
-                    vst1_s8(pp + 56, _r7);
-                    pp += 64;
-#elif __ARM_FEATURE_DOTPROD
-                    int32x2_t _r0 = vreinterpret_s32_s8(vld1_s8(sptr0));
-                    int32x2_t _r1 = vreinterpret_s32_s8(vld1_s8(sptr1));
-                    int32x2_t _r2 = vreinterpret_s32_s8(vld1_s8(sptr2));
-                    int32x2_t _r3 = vreinterpret_s32_s8(vld1_s8(sptr3));
-                    int32x2_t _r4 = vreinterpret_s32_s8(vld1_s8(sptr4));
-                    int32x2_t _r5 = vreinterpret_s32_s8(vld1_s8(sptr5));
-                    int32x2_t _r6 = vreinterpret_s32_s8(vld1_s8(sptr6));
-                    int32x2_t _r7 = vreinterpret_s32_s8(vld1_s8(sptr7));
-                    int32x2x2_t _r01 = vzip_s32(_r0, _r1);
-                    int32x2x2_t _r23 = vzip_s32(_r2, _r3);
-                    int32x2x2_t _r45 = vzip_s32(_r4, _r5);
-                    int32x2x2_t _r67 = vzip_s32(_r6, _r7);
-                    vst1_s32((int*)pp, _r01.val[0]);
-                    vst1_s32((int*)(pp + 8), _r23.val[0]);
-                    vst1_s32((int*)(pp + 16), _r45.val[0]);
-                    vst1_s32((int*)(pp + 24), _r67.val[0]);
-                    vst1_s32((int*)(pp + 32), _r01.val[1]);
-                    vst1_s32((int*)(pp + 40), _r23.val[1]);
-                    vst1_s32((int*)(pp + 48), _r45.val[1]);
-                    vst1_s32((int*)(pp + 56), _r67.val[1]);
-                    pp += 64;
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                    int16x4_t _r0 = vreinterpret_s16_s8(vld1_s8(sptr0));
-                    int16x4_t _r1 = vreinterpret_s16_s8(vld1_s8(sptr1));
-                    int16x4_t _r2 = vreinterpret_s16_s8(vld1_s8(sptr2));
-                    int16x4_t _r3 = vreinterpret_s16_s8(vld1_s8(sptr3));
-                    int16x4_t _r4 = vreinterpret_s16_s8(vld1_s8(sptr4));
-                    int16x4_t _r5 = vreinterpret_s16_s8(vld1_s8(sptr5));
-                    int16x4_t _r6 = vreinterpret_s16_s8(vld1_s8(sptr6));
-                    int16x4_t _r7 = vreinterpret_s16_s8(vld1_s8(sptr7));
-                    int16x4x2_t _r01 = vzip_s16(_r0, _r1);
-                    int16x4x2_t _r23 = vzip_s16(_r2, _r3);
-                    int16x4x2_t _r45 = vzip_s16(_r4, _r5);
-                    int16x4x2_t _r67 = vzip_s16(_r6, _r7);
-                    int32x4x4_t _r0123;
-                    _r0123.val[0] = vreinterpretq_s32_s16(vcombine_s16(_r01.val[0], _r01.val[1]));
-                    _r0123.val[1] = vreinterpretq_s32_s16(vcombine_s16(_r23.val[0], _r23.val[1]));
-                    _r0123.val[2] = vreinterpretq_s32_s16(vcombine_s16(_r45.val[0], _r45.val[1]));
-                    _r0123.val[3] = vreinterpretq_s32_s16(vcombine_s16(_r67.val[0], _r67.val[1]));
-                    vst4q_s32((int*)pp, _r0123);
-                    pp += 64;
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr2[0];
-                    pp[3] = sptr3[0];
-                    pp[4] = sptr4[0];
-                    pp[5] = sptr5[0];
-                    pp[6] = sptr6[0];
-                    pp[7] = sptr7[0];
-                    pp += 8;
-                }
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-        int dy0 = (j + jj) / outw * stride_h;
-        int dy1 = (j + jj + 1) / outw * stride_h;
-        int dy2 = (j + jj + 2) / outw * stride_h;
-        int dy3 = (j + jj + 3) / outw * stride_h;
-        int dx0 = (j + jj) % outw * stride_w;
-        int dx1 = (j + jj + 1) % outw * stride_w;
-        int dx2 = (j + jj + 2) % outw * stride_w;
-        int dx3 = (j + jj + 3) % outw * stride_w;
-
-        if (dy0 == dy3)
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int p4 = (k + kk + 4) / maxk;
-                    int p5 = (k + kk + 5) / maxk;
-                    int p6 = (k + kk + 6) / maxk;
-                    int p7 = (k + kk + 7) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int uv4 = (k + kk + 4) % maxk;
-                    int uv5 = (k + kk + 5) % maxk;
-                    int uv6 = (k + kk + 6) % maxk;
-                    int uv7 = (k + kk + 7) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int u4 = uv4 / kernel_w;
-                    int u5 = uv5 / kernel_w;
-                    int u6 = uv6 / kernel_w;
-                    int u7 = uv7 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-                    int v4 = uv4 % kernel_w;
-                    int v5 = uv5 % kernel_w;
-                    int v6 = uv6 % kernel_w;
-                    int v7 = uv7 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-                    const Mat img4 = bottom_blob.channel(p4);
-                    const Mat img5 = bottom_blob.channel(p5);
-                    const Mat img6 = bottom_blob.channel(p6);
-                    const Mat img7 = bottom_blob.channel(p7);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-
-                    int x40 = dx0 + dilation_w * v4;
-                    int y40 = dy0 + dilation_h * u4;
-
-                    int x50 = dx0 + dilation_w * v5;
-                    int y50 = dy0 + dilation_h * u5;
-
-                    int x60 = dx0 + dilation_w * v6;
-                    int y60 = dy0 + dilation_h * u6;
-
-                    int x70 = dx0 + dilation_w * v7;
-                    int y70 = dy0 + dilation_h * u7;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr2 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr3 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr4 = img4.row<const signed char>(y40) + x40;
-                    const signed char* sptr5 = img5.row<const signed char>(y50) + x50;
-                    const signed char* sptr6 = img6.row<const signed char>(y60) + x60;
-                    const signed char* sptr7 = img7.row<const signed char>(y70) + x70;
-
-                    if (stride_w == 1)
-                    {
-                        int8x8_t _r0 = vld1_s8(sptr0);
-                        int8x8_t _r1 = vld1_s8(sptr1);
-                        int8x8_t _r2 = vld1_s8(sptr2);
-                        int8x8_t _r3 = vld1_s8(sptr3);
-                        int8x8_t _r4 = vld1_s8(sptr4);
-                        int8x8_t _r5 = vld1_s8(sptr5);
-                        int8x8_t _r6 = vld1_s8(sptr6);
-                        int8x8_t _r7 = vld1_s8(sptr7);
-                        int16x4x4_t _r0123;
-                        _r0123.val[0] = vreinterpret_s16_s8(vzip_s8(_r0, _r1).val[0]);
-                        _r0123.val[1] = vreinterpret_s16_s8(vzip_s8(_r2, _r3).val[0]);
-                        _r0123.val[2] = vreinterpret_s16_s8(vzip_s8(_r4, _r5).val[0]);
-                        _r0123.val[3] = vreinterpret_s16_s8(vzip_s8(_r6, _r7).val[0]);
-                        vst4_s16((short*)pp, _r0123);
-                        pp += 32;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        int8x8_t _r0 = vld1_s8(sptr0);
-                        int8x8_t _r1 = vld1_s8(sptr1);
-                        int8x8_t _r2 = vld1_s8(sptr2);
-                        int8x8_t _r3 = vld1_s8(sptr3);
-                        int8x8_t _r4 = vld1_s8(sptr4);
-                        int8x8_t _r5 = vld1_s8(sptr5);
-                        int8x8_t _r6 = vld1_s8(sptr6);
-                        int8x8_t _r7 = vld1_s8(sptr7);
-                        int8x8_t _r01 = vtrn_s8(_r0, _r1).val[0];
-                        int8x8_t _r23 = vtrn_s8(_r2, _r3).val[0];
-                        int8x8_t _r45 = vtrn_s8(_r4, _r5).val[0];
-                        int8x8_t _r67 = vtrn_s8(_r6, _r7).val[0];
-                        int16x4x4_t _r0123;
-                        _r0123.val[0] = vreinterpret_s16_s8(_r01);
-                        _r0123.val[1] = vreinterpret_s16_s8(_r23);
-                        _r0123.val[2] = vreinterpret_s16_s8(_r45);
-                        _r0123.val[3] = vreinterpret_s16_s8(_r67);
-                        vst4_s16((short*)pp, _r0123);
-                        pp += 32;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr2[0];
-                        pp[3] = sptr3[0];
-                        pp[4] = sptr4[0];
-                        pp[5] = sptr5[0];
-                        pp[6] = sptr6[0];
-                        pp[7] = sptr7[0];
-                        pp[8] = sptr0[stride_w];
-                        pp[9] = sptr1[stride_w];
-                        pp[10] = sptr2[stride_w];
-                        pp[11] = sptr3[stride_w];
-                        pp[12] = sptr4[stride_w];
-                        pp[13] = sptr5[stride_w];
-                        pp[14] = sptr6[stride_w];
-                        pp[15] = sptr7[stride_w];
-                        pp[16] = sptr0[stride_w * 2];
-                        pp[17] = sptr1[stride_w * 2];
-                        pp[18] = sptr2[stride_w * 2];
-                        pp[19] = sptr3[stride_w * 2];
-                        pp[20] = sptr4[stride_w * 2];
-                        pp[21] = sptr5[stride_w * 2];
-                        pp[22] = sptr6[stride_w * 2];
-                        pp[23] = sptr7[stride_w * 2];
-                        pp[24] = sptr0[stride_w * 3];
-                        pp[25] = sptr1[stride_w * 3];
-                        pp[26] = sptr2[stride_w * 3];
-                        pp[27] = sptr3[stride_w * 3];
-                        pp[28] = sptr4[stride_w * 3];
-                        pp[29] = sptr5[stride_w * 3];
-                        pp[30] = sptr6[stride_w * 3];
-                        pp[31] = sptr7[stride_w * 3];
-                        pp += 32;
-                    }
-                }
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr2 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr3 = img3.row<const signed char>(y30) + x30;
-
-                    if (stride_w == 1)
-                    {
-                        int8x8_t _r0 = vld1_s8(sptr0);
-                        int8x8_t _r1 = vld1_s8(sptr1);
-                        int8x8_t _r2 = vld1_s8(sptr2);
-                        int8x8_t _r3 = vld1_s8(sptr3);
-                        int16x4x2_t _r01;
-                        _r01.val[0] = vreinterpret_s16_s8(vzip_s8(_r0, _r1).val[0]);
-                        _r01.val[1] = vreinterpret_s16_s8(vzip_s8(_r2, _r3).val[0]);
-                        vst2_s16((short*)pp, _r01);
-                        pp += 16;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        int8x8_t _r0 = vld1_s8(sptr0);
-                        int8x8_t _r1 = vld1_s8(sptr1);
-                        int8x8_t _r2 = vld1_s8(sptr2);
-                        int8x8_t _r3 = vld1_s8(sptr3);
-                        int8x8_t _r01 = vtrn_s8(_r0, _r1).val[0];
-                        int8x8_t _r23 = vtrn_s8(_r2, _r3).val[0];
-                        int16x4x2_t _r0123;
-                        _r0123.val[0] = vreinterpret_s16_s8(_r01);
-                        _r0123.val[1] = vreinterpret_s16_s8(_r23);
-                        vst2_s16((short*)pp, _r0123);
-                        pp += 16;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr2[0];
-                        pp[3] = sptr3[0];
-                        pp[4] = sptr0[stride_w];
-                        pp[5] = sptr1[stride_w];
-                        pp[6] = sptr2[stride_w];
-                        pp[7] = sptr3[stride_w];
-                        pp[8] = sptr0[stride_w * 2];
-                        pp[9] = sptr1[stride_w * 2];
-                        pp[10] = sptr2[stride_w * 2];
-                        pp[11] = sptr3[stride_w * 2];
-                        pp[12] = sptr0[stride_w * 3];
-                        pp[13] = sptr1[stride_w * 3];
-                        pp[14] = sptr2[stride_w * 3];
-                        pp[15] = sptr3[stride_w * 3];
-                        pp += 16;
-                    }
-                }
-#endif // __ARM_FEATURE_DOTPROD
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-
-                    if (stride_w == 1)
-                    {
-                        int8x8_t _r0 = vld1_s8(sptr0);
-                        int8x8_t _r1 = vld1_s8(sptr1);
-                        int8x8_t _r01 = vzip_s8(_r0, _r1).val[0];
-                        vst1_s8(pp, _r01);
-                        pp += 8;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        int8x8_t _r0 = vld1_s8(sptr0);
-                        int8x8_t _r1 = vld1_s8(sptr1);
-                        int8x8_t _r01 = vtrn_s8(_r0, _r1).val[0];
-                        vst1_s8(pp, _r01);
-                        pp += 8;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr0[stride_w];
-                        pp[3] = sptr1[stride_w];
-                        pp[4] = sptr0[stride_w * 2];
-                        pp[5] = sptr1[stride_w * 2];
-                        pp[6] = sptr0[stride_w * 3];
-                        pp[7] = sptr1[stride_w * 3];
-                        pp += 8;
-                    }
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = dx0 + dilation_w * v;
-                int y0 = dy0 + dilation_h * u;
-
-                const signed char* sptr = img.row<const signed char>(y0) + x0 * elempack;
-
-                if (elempack == 8)
-                {
-#if __ARM_FEATURE_MATMUL_INT8
-                    int8x8_t _r0 = vld1_s8(sptr);
-                    int8x8_t _r1 = vld1_s8(sptr + stride_w * 8);
-                    int8x8_t _r2 = vld1_s8(sptr + stride_w * 16);
-                    int8x8_t _r3 = vld1_s8(sptr + stride_w * 24);
-                    vst1_s8(pp, _r0);
-                    vst1_s8(pp + 8, _r1);
-                    vst1_s8(pp + 16, _r2);
-                    vst1_s8(pp + 24, _r3);
-                    pp += 32;
-#elif __ARM_FEATURE_DOTPROD
-                    int32x2x4_t _r0123;
-                    _r0123.val[0] = vreinterpret_s32_s8(vld1_s8(sptr));
-                    _r0123.val[1] = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 8));
-                    _r0123.val[2] = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 16));
-                    _r0123.val[3] = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 24));
-                    vst4_s32((int*)pp, _r0123);
-                    pp += 32;
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                    int16x4x4_t _r0123;
-                    _r0123.val[0] = vreinterpret_s16_s8(vld1_s8(sptr));
-                    _r0123.val[1] = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 8));
-                    _r0123.val[2] = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 16));
-                    _r0123.val[3] = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 24));
-                    vst4_s16((short*)pp, _r0123);
-                    pp += 32;
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr[0];
-                    pp[1] = sptr[stride_w];
-                    pp[2] = sptr[stride_w * 2];
-                    pp[3] = sptr[stride_w * 3];
-                    pp += 4;
-                }
-            }
-        }
-        else
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int p4 = (k + kk + 4) / maxk;
-                    int p5 = (k + kk + 5) / maxk;
-                    int p6 = (k + kk + 6) / maxk;
-                    int p7 = (k + kk + 7) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int uv4 = (k + kk + 4) % maxk;
-                    int uv5 = (k + kk + 5) % maxk;
-                    int uv6 = (k + kk + 6) % maxk;
-                    int uv7 = (k + kk + 7) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int u4 = uv4 / kernel_w;
-                    int u5 = uv5 / kernel_w;
-                    int u6 = uv6 / kernel_w;
-                    int u7 = uv7 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-                    int v4 = uv4 % kernel_w;
-                    int v5 = uv5 % kernel_w;
-                    int v6 = uv6 % kernel_w;
-                    int v7 = uv7 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-                    const Mat img4 = bottom_blob.channel(p4);
-                    const Mat img5 = bottom_blob.channel(p5);
-                    const Mat img6 = bottom_blob.channel(p6);
-                    const Mat img7 = bottom_blob.channel(p7);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int x02 = dx2 + dilation_w * v0;
-                    int x03 = dx3 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int y02 = dy2 + dilation_h * u0;
-                    int y03 = dy3 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int x12 = dx2 + dilation_w * v1;
-                    int x13 = dx3 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-                    int y12 = dy2 + dilation_h * u1;
-                    int y13 = dy3 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int x21 = dx1 + dilation_w * v2;
-                    int x22 = dx2 + dilation_w * v2;
-                    int x23 = dx3 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int y21 = dy1 + dilation_h * u2;
-                    int y22 = dy2 + dilation_h * u2;
-                    int y23 = dy3 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int x31 = dx1 + dilation_w * v3;
-                    int x32 = dx2 + dilation_w * v3;
-                    int x33 = dx3 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-                    int y31 = dy1 + dilation_h * u3;
-                    int y32 = dy2 + dilation_h * u3;
-                    int y33 = dy3 + dilation_h * u3;
-
-                    int x40 = dx0 + dilation_w * v4;
-                    int x41 = dx1 + dilation_w * v4;
-                    int x42 = dx2 + dilation_w * v4;
-                    int x43 = dx3 + dilation_w * v4;
-                    int y40 = dy0 + dilation_h * u4;
-                    int y41 = dy1 + dilation_h * u4;
-                    int y42 = dy2 + dilation_h * u4;
-                    int y43 = dy3 + dilation_h * u4;
-
-                    int x50 = dx0 + dilation_w * v5;
-                    int x51 = dx1 + dilation_w * v5;
-                    int x52 = dx2 + dilation_w * v5;
-                    int x53 = dx3 + dilation_w * v5;
-                    int y50 = dy0 + dilation_h * u5;
-                    int y51 = dy1 + dilation_h * u5;
-                    int y52 = dy2 + dilation_h * u5;
-                    int y53 = dy3 + dilation_h * u5;
-
-                    int x60 = dx0 + dilation_w * v6;
-                    int x61 = dx1 + dilation_w * v6;
-                    int x62 = dx2 + dilation_w * v6;
-                    int x63 = dx3 + dilation_w * v6;
-                    int y60 = dy0 + dilation_h * u6;
-                    int y61 = dy1 + dilation_h * u6;
-                    int y62 = dy2 + dilation_h * u6;
-                    int y63 = dy3 + dilation_h * u6;
-
-                    int x70 = dx0 + dilation_w * v7;
-                    int x71 = dx1 + dilation_w * v7;
-                    int x72 = dx2 + dilation_w * v7;
-                    int x73 = dx3 + dilation_w * v7;
-                    int y70 = dy0 + dilation_h * u7;
-                    int y71 = dy1 + dilation_h * u7;
-                    int y72 = dy2 + dilation_h * u7;
-                    int y73 = dy3 + dilation_h * u7;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-
-                    const signed char* sptr20 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr21 = img2.row<const signed char>(y21) + x21;
-                    const signed char* sptr22 = img2.row<const signed char>(y22) + x22;
-                    const signed char* sptr23 = img2.row<const signed char>(y23) + x23;
-
-                    const signed char* sptr30 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr31 = img3.row<const signed char>(y31) + x31;
-                    const signed char* sptr32 = img3.row<const signed char>(y32) + x32;
-                    const signed char* sptr33 = img3.row<const signed char>(y33) + x33;
-
-                    const signed char* sptr40 = img4.row<const signed char>(y40) + x40;
-                    const signed char* sptr41 = img4.row<const signed char>(y41) + x41;
-                    const signed char* sptr42 = img4.row<const signed char>(y42) + x42;
-                    const signed char* sptr43 = img4.row<const signed char>(y43) + x43;
-
-                    const signed char* sptr50 = img5.row<const signed char>(y50) + x50;
-                    const signed char* sptr51 = img5.row<const signed char>(y51) + x51;
-                    const signed char* sptr52 = img5.row<const signed char>(y52) + x52;
-                    const signed char* sptr53 = img5.row<const signed char>(y53) + x53;
-
-                    const signed char* sptr60 = img6.row<const signed char>(y60) + x60;
-                    const signed char* sptr61 = img6.row<const signed char>(y61) + x61;
-                    const signed char* sptr62 = img6.row<const signed char>(y62) + x62;
-                    const signed char* sptr63 = img6.row<const signed char>(y63) + x63;
-
-                    const signed char* sptr70 = img7.row<const signed char>(y70) + x70;
-                    const signed char* sptr71 = img7.row<const signed char>(y71) + x71;
-                    const signed char* sptr72 = img7.row<const signed char>(y72) + x72;
-                    const signed char* sptr73 = img7.row<const signed char>(y73) + x73;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr20[0];
-                    pp[3] = sptr30[0];
-                    pp[4] = sptr40[0];
-                    pp[5] = sptr50[0];
-                    pp[6] = sptr60[0];
-                    pp[7] = sptr70[0];
-                    pp[8] = sptr01[0];
-                    pp[9] = sptr11[0];
-                    pp[10] = sptr21[0];
-                    pp[11] = sptr31[0];
-                    pp[12] = sptr41[0];
-                    pp[13] = sptr51[0];
-                    pp[14] = sptr61[0];
-                    pp[15] = sptr71[0];
-                    pp[16] = sptr02[0];
-                    pp[17] = sptr12[0];
-                    pp[18] = sptr22[0];
-                    pp[19] = sptr32[0];
-                    pp[20] = sptr42[0];
-                    pp[21] = sptr52[0];
-                    pp[22] = sptr62[0];
-                    pp[23] = sptr72[0];
-                    pp[24] = sptr03[0];
-                    pp[25] = sptr13[0];
-                    pp[26] = sptr23[0];
-                    pp[27] = sptr33[0];
-                    pp[28] = sptr43[0];
-                    pp[29] = sptr53[0];
-                    pp[30] = sptr63[0];
-                    pp[31] = sptr73[0];
-                    pp += 32;
-                }
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int x02 = dx2 + dilation_w * v0;
-                    int x03 = dx3 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int y02 = dy2 + dilation_h * u0;
-                    int y03 = dy3 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int x12 = dx2 + dilation_w * v1;
-                    int x13 = dx3 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-                    int y12 = dy2 + dilation_h * u1;
-                    int y13 = dy3 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int x21 = dx1 + dilation_w * v2;
-                    int x22 = dx2 + dilation_w * v2;
-                    int x23 = dx3 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int y21 = dy1 + dilation_h * u2;
-                    int y22 = dy2 + dilation_h * u2;
-                    int y23 = dy3 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int x31 = dx1 + dilation_w * v3;
-                    int x32 = dx2 + dilation_w * v3;
-                    int x33 = dx3 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-                    int y31 = dy1 + dilation_h * u3;
-                    int y32 = dy2 + dilation_h * u3;
-                    int y33 = dy3 + dilation_h * u3;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-
-                    const signed char* sptr20 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr21 = img2.row<const signed char>(y21) + x21;
-                    const signed char* sptr22 = img2.row<const signed char>(y22) + x22;
-                    const signed char* sptr23 = img2.row<const signed char>(y23) + x23;
-
-                    const signed char* sptr30 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr31 = img3.row<const signed char>(y31) + x31;
-                    const signed char* sptr32 = img3.row<const signed char>(y32) + x32;
-                    const signed char* sptr33 = img3.row<const signed char>(y33) + x33;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr20[0];
-                    pp[3] = sptr30[0];
-                    pp[4] = sptr01[0];
-                    pp[5] = sptr11[0];
-                    pp[6] = sptr21[0];
-                    pp[7] = sptr31[0];
-                    pp[8] = sptr02[0];
-                    pp[9] = sptr12[0];
-                    pp[10] = sptr22[0];
-                    pp[11] = sptr32[0];
-                    pp[12] = sptr03[0];
-                    pp[13] = sptr13[0];
-                    pp[14] = sptr23[0];
-                    pp[15] = sptr33[0];
-                    pp += 16;
-                }
-#endif // __ARM_FEATURE_DOTPROD
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int x02 = dx2 + dilation_w * v0;
-                    int x03 = dx3 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int y02 = dy2 + dilation_h * u0;
-                    int y03 = dy3 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int x12 = dx2 + dilation_w * v1;
-                    int x13 = dx3 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-                    int y12 = dy2 + dilation_h * u1;
-                    int y13 = dy3 + dilation_h * u1;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr01[0];
-                    pp[3] = sptr11[0];
-                    pp[4] = sptr02[0];
-                    pp[5] = sptr12[0];
-                    pp[6] = sptr03[0];
-                    pp[7] = sptr13[0];
-                    pp += 8;
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = dx0 + dilation_w * v;
-                int x1 = dx1 + dilation_w * v;
-                int x2 = dx2 + dilation_w * v;
-                int x3 = dx3 + dilation_w * v;
-                int y0 = dy0 + dilation_h * u;
-                int y1 = dy1 + dilation_h * u;
-                int y2 = dy2 + dilation_h * u;
-                int y3 = dy3 + dilation_h * u;
-
-                const signed char* sptr0 = img.row<const signed char>(y0) + x0 * elempack;
-                const signed char* sptr1 = img.row<const signed char>(y1) + x1 * elempack;
-                const signed char* sptr2 = img.row<const signed char>(y2) + x2 * elempack;
-                const signed char* sptr3 = img.row<const signed char>(y3) + x3 * elempack;
-
-                if (elempack == 8)
-                {
-#if __ARM_FEATURE_MATMUL_INT8
-                    int8x8_t _r0 = vld1_s8(sptr0);
-                    int8x8_t _r1 = vld1_s8(sptr1);
-                    int8x8_t _r2 = vld1_s8(sptr2);
-                    int8x8_t _r3 = vld1_s8(sptr3);
-                    vst1_s8(pp, _r0);
-                    vst1_s8(pp + 8, _r1);
-                    vst1_s8(pp + 16, _r2);
-                    vst1_s8(pp + 24, _r3);
-                    pp += 32;
-#elif __ARM_FEATURE_DOTPROD
-                    int32x2x4_t _r0123;
-                    _r0123.val[0] = vreinterpret_s32_s8(vld1_s8(sptr0));
-                    _r0123.val[1] = vreinterpret_s32_s8(vld1_s8(sptr1));
-                    _r0123.val[2] = vreinterpret_s32_s8(vld1_s8(sptr2));
-                    _r0123.val[3] = vreinterpret_s32_s8(vld1_s8(sptr3));
-                    vst4_s32((int*)pp, _r0123);
-                    pp += 32;
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                    int16x4x4_t _r0123;
-                    _r0123.val[0] = vreinterpret_s16_s8(vld1_s8(sptr0));
-                    _r0123.val[1] = vreinterpret_s16_s8(vld1_s8(sptr1));
-                    _r0123.val[2] = vreinterpret_s16_s8(vld1_s8(sptr2));
-                    _r0123.val[3] = vreinterpret_s16_s8(vld1_s8(sptr3));
-                    vst4_s16((short*)pp, _r0123);
-                    pp += 32;
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr2[0];
-                    pp[3] = sptr3[0];
-                    pp += 4;
-                }
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-        int dy0 = (j + jj) / outw * stride_h;
-        int dy1 = (j + jj + 1) / outw * stride_h;
-        int dx0 = (j + jj) % outw * stride_w;
-        int dx1 = (j + jj + 1) % outw * stride_w;
-
-        if (dy0 == dy1)
-        {
-            int kk = 0;
-#if __ARM_NEON
-            if (elempack == 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int p4 = (k + kk + 4) / maxk;
-                    int p5 = (k + kk + 5) / maxk;
-                    int p6 = (k + kk + 6) / maxk;
-                    int p7 = (k + kk + 7) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int uv4 = (k + kk + 4) % maxk;
-                    int uv5 = (k + kk + 5) % maxk;
-                    int uv6 = (k + kk + 6) % maxk;
-                    int uv7 = (k + kk + 7) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int u4 = uv4 / kernel_w;
-                    int u5 = uv5 / kernel_w;
-                    int u6 = uv6 / kernel_w;
-                    int u7 = uv7 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-                    int v4 = uv4 % kernel_w;
-                    int v5 = uv5 % kernel_w;
-                    int v6 = uv6 % kernel_w;
-                    int v7 = uv7 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-                    const Mat img4 = bottom_blob.channel(p4);
-                    const Mat img5 = bottom_blob.channel(p5);
-                    const Mat img6 = bottom_blob.channel(p6);
-                    const Mat img7 = bottom_blob.channel(p7);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int x30 = dx0 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-
-                    int x40 = dx0 + dilation_w * v4;
-                    int y40 = dy0 + dilation_h * u4;
-                    int x50 = dx0 + dilation_w * v5;
-                    int y50 = dy0 + dilation_h * u5;
-
-                    int x60 = dx0 + dilation_w * v6;
-                    int y60 = dy0 + dilation_h * u6;
-                    int x70 = dx0 + dilation_w * v7;
-                    int y70 = dy0 + dilation_h * u7;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr2 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr3 = img3.row<const signed char>(y30) + x30;
-
-                    const signed char* sptr4 = img4.row<const signed char>(y40) + x40;
-                    const signed char* sptr5 = img5.row<const signed char>(y50) + x50;
-                    const signed char* sptr6 = img6.row<const signed char>(y60) + x60;
-                    const signed char* sptr7 = img7.row<const signed char>(y70) + x70;
-
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr2[0];
-                    pp[3] = sptr3[0];
-                    pp[4] = sptr4[0];
-                    pp[5] = sptr5[0];
-                    pp[6] = sptr6[0];
-                    pp[7] = sptr7[0];
-                    pp[8] = sptr0[stride_w];
-                    pp[9] = sptr1[stride_w];
-                    pp[10] = sptr2[stride_w];
-                    pp[11] = sptr3[stride_w];
-                    pp[12] = sptr4[stride_w];
-                    pp[13] = sptr5[stride_w];
-                    pp[14] = sptr6[stride_w];
-                    pp[15] = sptr7[stride_w];
-                    pp += 16;
-                }
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int x20 = dx0 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int x30 = dx0 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr2 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr3 = img3.row<const signed char>(y30) + x30;
-
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr2[0];
-                    pp[3] = sptr3[0];
-                    pp[4] = sptr0[stride_w];
-                    pp[5] = sptr1[stride_w];
-                    pp[6] = sptr2[stride_w];
-                    pp[7] = sptr3[stride_w];
-                    pp += 8;
-                }
-#endif // __ARM_FEATURE_DOTPROD
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr0[stride_w];
-                    pp[3] = sptr1[stride_w];
-                    pp += 4;
-                }
-            }
-#endif // __ARM_NEON
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = dx0 + dilation_w * v;
-                int y0 = dy0 + dilation_h * u;
-
-                const signed char* sptr = img.row<const signed char>(y0) + x0 * elempack;
-
-#if __ARM_NEON
-                if (elempack == 8)
-                {
-#if __ARM_FEATURE_MATMUL_INT8
-                    int8x8_t _r0 = vld1_s8(sptr);
-                    int8x8_t _r1 = vld1_s8(sptr + stride_w * 8);
-                    vst1_s8(pp, _r0);
-                    vst1_s8(pp + 8, _r1);
-                    pp += 16;
-#elif __ARM_FEATURE_DOTPROD
-                    int32x2x2_t _r01;
-                    _r01.val[0] = vreinterpret_s32_s8(vld1_s8(sptr));
-                    _r01.val[1] = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 8));
-                    vst2_s32((int*)pp, _r01);
-                    pp += 16;
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                    int16x4x2_t _r01;
-                    _r01.val[0] = vreinterpret_s16_s8(vld1_s8(sptr));
-                    _r01.val[1] = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 8));
-                    vst2_s16((short*)pp, _r01);
-                    pp += 16;
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                }
-#endif // __ARM_NEON
-                if (elempack == 1)
-                {
-                    pp[0] = sptr[0];
-                    pp[1] = sptr[stride_w];
-                    pp += 2;
-                }
-            }
-        }
-        else
-        {
-            int kk = 0;
-#if __ARM_NEON
-            if (elempack == 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int p4 = (k + kk + 4) / maxk;
-                    int p5 = (k + kk + 5) / maxk;
-                    int p6 = (k + kk + 6) / maxk;
-                    int p7 = (k + kk + 7) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int uv4 = (k + kk + 4) % maxk;
-                    int uv5 = (k + kk + 5) % maxk;
-                    int uv6 = (k + kk + 6) % maxk;
-                    int uv7 = (k + kk + 7) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int u4 = uv4 / kernel_w;
-                    int u5 = uv5 / kernel_w;
-                    int u6 = uv6 / kernel_w;
-                    int u7 = uv7 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-                    int v4 = uv4 % kernel_w;
-                    int v5 = uv5 % kernel_w;
-                    int v6 = uv6 % kernel_w;
-                    int v7 = uv7 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-                    const Mat img4 = bottom_blob.channel(p4);
-                    const Mat img5 = bottom_blob.channel(p5);
-                    const Mat img6 = bottom_blob.channel(p6);
-                    const Mat img7 = bottom_blob.channel(p7);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int x21 = dx1 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int y21 = dy1 + dilation_h * u2;
-                    int x30 = dx0 + dilation_w * v3;
-                    int x31 = dx1 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-                    int y31 = dy1 + dilation_h * u3;
-
-                    int x40 = dx0 + dilation_w * v4;
-                    int x41 = dx1 + dilation_w * v4;
-                    int y40 = dy0 + dilation_h * u4;
-                    int y41 = dy1 + dilation_h * u4;
-                    int x50 = dx0 + dilation_w * v5;
-                    int x51 = dx1 + dilation_w * v5;
-                    int y50 = dy0 + dilation_h * u5;
-                    int y51 = dy1 + dilation_h * u5;
-
-                    int x60 = dx0 + dilation_w * v6;
-                    int x61 = dx1 + dilation_w * v6;
-                    int y60 = dy0 + dilation_h * u6;
-                    int y61 = dy1 + dilation_h * u6;
-                    int x70 = dx0 + dilation_w * v7;
-                    int x71 = dx1 + dilation_w * v7;
-                    int y70 = dy0 + dilation_h * u7;
-                    int y71 = dy1 + dilation_h * u7;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr20 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr21 = img2.row<const signed char>(y21) + x21;
-                    const signed char* sptr30 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr31 = img3.row<const signed char>(y31) + x31;
-
-                    const signed char* sptr40 = img4.row<const signed char>(y40) + x40;
-                    const signed char* sptr41 = img4.row<const signed char>(y41) + x41;
-                    const signed char* sptr50 = img5.row<const signed char>(y50) + x50;
-                    const signed char* sptr51 = img5.row<const signed char>(y51) + x51;
-                    const signed char* sptr60 = img6.row<const signed char>(y60) + x60;
-                    const signed char* sptr61 = img6.row<const signed char>(y61) + x61;
-                    const signed char* sptr70 = img7.row<const signed char>(y70) + x70;
-                    const signed char* sptr71 = img7.row<const signed char>(y71) + x71;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr20[0];
-                    pp[3] = sptr30[0];
-                    pp[4] = sptr40[0];
-                    pp[5] = sptr50[0];
-                    pp[6] = sptr60[0];
-                    pp[7] = sptr70[0];
-                    pp[8] = sptr01[0];
-                    pp[9] = sptr11[0];
-                    pp[10] = sptr21[0];
-                    pp[11] = sptr31[0];
-                    pp[12] = sptr41[0];
-                    pp[13] = sptr51[0];
-                    pp[14] = sptr61[0];
-                    pp[15] = sptr71[0];
-                    pp += 16;
-                }
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-                    int x20 = dx0 + dilation_w * v2;
-                    int x21 = dx1 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int y21 = dy1 + dilation_h * u2;
-                    int x30 = dx0 + dilation_w * v3;
-                    int x31 = dx1 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-                    int y31 = dy1 + dilation_h * u3;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr20 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr21 = img2.row<const signed char>(y21) + x21;
-                    const signed char* sptr30 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr31 = img3.row<const signed char>(y31) + x31;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr20[0];
-                    pp[3] = sptr30[0];
-                    pp[4] = sptr01[0];
-                    pp[5] = sptr11[0];
-                    pp[6] = sptr21[0];
-                    pp[7] = sptr31[0];
-                    pp += 8;
-                }
-#endif // __ARM_FEATURE_DOTPROD
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr01[0];
-                    pp[3] = sptr11[0];
-                    pp += 4;
-                }
-            }
-#endif // __ARM_NEON
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = dx0 + dilation_w * v;
-                int x1 = dx1 + dilation_w * v;
-                int y0 = dy0 + dilation_h * u;
-                int y1 = dy1 + dilation_h * u;
-
-                const signed char* sptr0 = img.row<const signed char>(y0) + x0 * elempack;
-                const signed char* sptr1 = img.row<const signed char>(y1) + x1 * elempack;
-
-#if __ARM_NEON
-                if (elempack == 8)
-                {
-#if __ARM_FEATURE_MATMUL_INT8
-                    int8x8_t _r0 = vld1_s8(sptr0);
-                    int8x8_t _r1 = vld1_s8(sptr1);
-                    vst1_s8(pp, _r0);
-                    vst1_s8(pp + 8, _r1);
-                    pp += 16;
-#elif __ARM_FEATURE_DOTPROD
-                    int32x2x2_t _r01;
-                    _r01.val[0] = vreinterpret_s32_s8(vld1_s8(sptr0));
-                    _r01.val[1] = vreinterpret_s32_s8(vld1_s8(sptr1));
-                    vst2_s32((int*)pp, _r01);
-                    pp += 16;
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                    int16x4x2_t _r01;
-                    _r01.val[0] = vreinterpret_s16_s8(vld1_s8(sptr0));
-                    _r01.val[1] = vreinterpret_s16_s8(vld1_s8(sptr1));
-                    vst2_s16((short*)pp, _r01);
-                    pp += 16;
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                }
-#endif // __ARM_NEON
-                if (elempack == 1)
-                {
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp += 2;
-                }
-            }
-        }
-    }
-    for (; jj < max_jj; jj++)
-    {
-        int dy = (j + jj) / outw * stride_h;
-        int dx = (j + jj) % outw * stride_w;
-
-        int kk = 0;
-        for (; kk < max_kk / elempack; kk++)
-        {
-            int p = (k / elempack + kk) / maxk;
-            int uv = (k / elempack + kk) % maxk;
-            int u = uv / kernel_w;
-            int v = uv % kernel_w;
-
-            const Mat img = bottom_blob.channel(p);
-
-            int x = dx + dilation_w * v;
-            int y = dy + dilation_h * u;
-
-            const signed char* sptr = img.row<const signed char>(y) + x * elempack;
-
-#if __ARM_NEON
-            if (elempack == 8)
-            {
-                vst1_s8(pp, vld1_s8(sptr));
-                pp += 8;
-            }
-#endif // __ARM_NEON
-            if (elempack == 1)
-            {
-                pp[0] = sptr[0];
-                pp += 1;
-            }
-        }
-    }
-}
-
-#if __ARM_FEATURE_MATMUL_INT8
-template void convolution_im2col_input_tile_int8_i8mm<1, 1, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8_i8mm<3, 3, 1, 1, 1, 1>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8_i8mm<3, 3, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8_i8mm<5, 5, 1, 1, 1, 1>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8_i8mm<5, 5, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8_i8mm<7, 7, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-#elif __ARM_FEATURE_DOTPROD
-template void convolution_im2col_input_tile_int8_asimddp<1, 1, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8_asimddp<3, 3, 1, 1, 1, 1>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8_asimddp<3, 3, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8_asimddp<5, 5, 1, 1, 1, 1>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8_asimddp<5, 5, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8_asimddp<7, 7, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-template void convolution_im2col_input_tile_int8<1, 1, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8<3, 3, 1, 1, 1, 1>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8<3, 3, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8<5, 5, 1, 1, 1, 1>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8<5, 5, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8<7, 7, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-
-static void convolution_im2col_input_tile_int8(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h)
-{
-    if (kernel_w == 1 && kernel_h == 1 && stride_w == 1 && stride_h == 1)
-    {
-        convolution_im2col_input_tile_conv1x1s1d1_int8(bottom_blob, B, j, max_jj, k, max_kk);
-        return;
-    }
-
-    if (kernel_w == 1 && kernel_h == 1 && stride_w == 2 && stride_h == 2)
-    {
-#if __ARM_FEATURE_MATMUL_INT8
-        convolution_im2col_input_tile_int8_i8mm<1, 1, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#elif __ARM_FEATURE_DOTPROD
-        convolution_im2col_input_tile_int8_asimddp<1, 1, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-        convolution_im2col_input_tile_int8<1, 1, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-        return;
-    }
-
-    if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
-    {
-#if __ARM_FEATURE_MATMUL_INT8
-        convolution_im2col_input_tile_int8_i8mm<3, 3, 1, 1, 1, 1>(bottom_blob, B, j, max_jj, k, max_kk);
-#elif __ARM_FEATURE_DOTPROD
-        convolution_im2col_input_tile_int8_asimddp<3, 3, 1, 1, 1, 1>(bottom_blob, B, j, max_jj, k, max_kk);
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-        convolution_im2col_input_tile_int8<3, 3, 1, 1, 1, 1>(bottom_blob, B, j, max_jj, k, max_kk);
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-        return;
-    }
-
-    if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
-    {
-#if __ARM_FEATURE_MATMUL_INT8
-        convolution_im2col_input_tile_int8_i8mm<3, 3, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#elif __ARM_FEATURE_DOTPROD
-        convolution_im2col_input_tile_int8_asimddp<3, 3, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-        convolution_im2col_input_tile_int8<3, 3, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-        return;
-    }
-
-    if (kernel_w == 5 && kernel_h == 5 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
-    {
-#if __ARM_FEATURE_MATMUL_INT8
-        convolution_im2col_input_tile_int8_i8mm<5, 5, 1, 1, 1, 1>(bottom_blob, B, j, max_jj, k, max_kk);
-#elif __ARM_FEATURE_DOTPROD
-        convolution_im2col_input_tile_int8_asimddp<5, 5, 1, 1, 1, 1>(bottom_blob, B, j, max_jj, k, max_kk);
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-        convolution_im2col_input_tile_int8<5, 5, 1, 1, 1, 1>(bottom_blob, B, j, max_jj, k, max_kk);
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-        return;
-    }
-
-    if (kernel_w == 5 && kernel_h == 5 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
-    {
-#if __ARM_FEATURE_MATMUL_INT8
-        convolution_im2col_input_tile_int8_i8mm<5, 5, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#elif __ARM_FEATURE_DOTPROD
-        convolution_im2col_input_tile_int8_asimddp<5, 5, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-        convolution_im2col_input_tile_int8<5, 5, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-        return;
-    }
-
-    if (kernel_w == 7 && kernel_h == 7 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
-    {
-#if __ARM_FEATURE_MATMUL_INT8
-        convolution_im2col_input_tile_int8_i8mm<7, 7, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#elif __ARM_FEATURE_DOTPROD
-        convolution_im2col_input_tile_int8_asimddp<7, 7, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-        convolution_im2col_input_tile_int8<7, 7, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-        return;
-    }
-
-    const int w = bottom_blob.w;
-    // const int channels = bottom_blob.c;
-    const int elempack = bottom_blob.elempack;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int outw = (w - kernel_extent_w) / stride_w + 1;
-
-    // j max_jj     outw*outh    split w and h
-
-    // k max_kk     pa*maxk*(inch/pa)    split inch
-
-    // k/max_kk shall be multiple of maxk
-
-    const int maxk = kernel_w * kernel_h;
-
-    signed char* pp = B;
-
-    int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-        int dy0 = (j + jj) / outw * stride_h;
-        int dy1 = (j + jj + 1) / outw * stride_h;
-        int dy2 = (j + jj + 2) / outw * stride_h;
-        int dy3 = (j + jj + 3) / outw * stride_h;
-        int dy4 = (j + jj + 4) / outw * stride_h;
-        int dy5 = (j + jj + 5) / outw * stride_h;
-        int dy6 = (j + jj + 6) / outw * stride_h;
-        int dy7 = (j + jj + 7) / outw * stride_h;
-        int dx0 = (j + jj) % outw * stride_w;
-        int dx1 = (j + jj + 1) % outw * stride_w;
-        int dx2 = (j + jj + 2) % outw * stride_w;
-        int dx3 = (j + jj + 3) % outw * stride_w;
-        int dx4 = (j + jj + 4) % outw * stride_w;
-        int dx5 = (j + jj + 5) % outw * stride_w;
-        int dx6 = (j + jj + 6) % outw * stride_w;
-        int dx7 = (j + jj + 7) % outw * stride_w;
-
-        if (dy0 == dy7)
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int p4 = (k + kk + 4) / maxk;
-                    int p5 = (k + kk + 5) / maxk;
-                    int p6 = (k + kk + 6) / maxk;
-                    int p7 = (k + kk + 7) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int uv4 = (k + kk + 4) % maxk;
-                    int uv5 = (k + kk + 5) % maxk;
-                    int uv6 = (k + kk + 6) % maxk;
-                    int uv7 = (k + kk + 7) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int u4 = uv4 / kernel_w;
-                    int u5 = uv5 / kernel_w;
-                    int u6 = uv6 / kernel_w;
-                    int u7 = uv7 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-                    int v4 = uv4 % kernel_w;
-                    int v5 = uv5 % kernel_w;
-                    int v6 = uv6 % kernel_w;
-                    int v7 = uv7 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-                    const Mat img4 = bottom_blob.channel(p4);
-                    const Mat img5 = bottom_blob.channel(p5);
-                    const Mat img6 = bottom_blob.channel(p6);
-                    const Mat img7 = bottom_blob.channel(p7);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-
-                    int x40 = dx0 + dilation_w * v4;
-                    int y40 = dy0 + dilation_h * u4;
-
-                    int x50 = dx0 + dilation_w * v5;
-                    int y50 = dy0 + dilation_h * u5;
-
-                    int x60 = dx0 + dilation_w * v6;
-                    int y60 = dy0 + dilation_h * u6;
-
-                    int x70 = dx0 + dilation_w * v7;
-                    int y70 = dy0 + dilation_h * u7;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr2 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr3 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr4 = img4.row<const signed char>(y40) + x40;
-                    const signed char* sptr5 = img5.row<const signed char>(y50) + x50;
-                    const signed char* sptr6 = img6.row<const signed char>(y60) + x60;
-                    const signed char* sptr7 = img7.row<const signed char>(y70) + x70;
-
-                    if (stride_w == 1)
-                    {
-                        int8x8_t _r0 = vld1_s8(sptr0);
-                        int8x8_t _r1 = vld1_s8(sptr1);
-                        int8x8_t _r2 = vld1_s8(sptr2);
-                        int8x8_t _r3 = vld1_s8(sptr3);
-                        int8x8_t _r4 = vld1_s8(sptr4);
-                        int8x8_t _r5 = vld1_s8(sptr5);
-                        int8x8_t _r6 = vld1_s8(sptr6);
-                        int8x8_t _r7 = vld1_s8(sptr7);
-                        // save as transpose8x8
-                        int8x8x2_t _r01 = vzip_s8(_r0, _r1);
-                        int8x8x2_t _r23 = vzip_s8(_r2, _r3);
-                        int8x8x2_t _r45 = vzip_s8(_r4, _r5);
-                        int8x8x2_t _r67 = vzip_s8(_r6, _r7);
-                        int16x8x4_t _r0246;
-                        _r0246.val[0] = vreinterpretq_s16_s8(vcombine_s8(_r01.val[0], _r01.val[1]));
-                        _r0246.val[1] = vreinterpretq_s16_s8(vcombine_s8(_r23.val[0], _r23.val[1]));
-                        _r0246.val[2] = vreinterpretq_s16_s8(vcombine_s8(_r45.val[0], _r45.val[1]));
-                        _r0246.val[3] = vreinterpretq_s16_s8(vcombine_s8(_r67.val[0], _r67.val[1]));
-                        vst4q_s16((short*)pp, _r0246);
-                        pp += 64;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        int8x16_t _r0 = vld1q_s8(sptr0);
-                        int8x16_t _r1 = vld1q_s8(sptr1);
-                        int8x16_t _r2 = vld1q_s8(sptr2);
-                        int8x16_t _r3 = vld1q_s8(sptr3);
-                        int8x16_t _r4 = vld1q_s8(sptr4);
-                        int8x16_t _r5 = vld1q_s8(sptr5);
-                        int8x16_t _r6 = vld1q_s8(sptr6);
-                        int8x16_t _r7 = vld1q_s8(sptr7);
-                        int8x16_t _r01 = vtrnq_s8(_r0, _r1).val[0];
-                        int8x16_t _r23 = vtrnq_s8(_r2, _r3).val[0];
-                        int8x16_t _r45 = vtrnq_s8(_r4, _r5).val[0];
-                        int8x16_t _r67 = vtrnq_s8(_r6, _r7).val[0];
-                        int16x8x4_t _r0123;
-                        _r0123.val[0] = vreinterpretq_s16_s8(_r01);
-                        _r0123.val[1] = vreinterpretq_s16_s8(_r23);
-                        _r0123.val[2] = vreinterpretq_s16_s8(_r45);
-                        _r0123.val[3] = vreinterpretq_s16_s8(_r67);
-                        vst4q_s16((short*)pp, _r0123);
-                        pp += 64;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr2[0];
-                        pp[3] = sptr3[0];
-                        pp[4] = sptr4[0];
-                        pp[5] = sptr5[0];
-                        pp[6] = sptr6[0];
-                        pp[7] = sptr7[0];
-                        pp[8] = sptr0[stride_w];
-                        pp[9] = sptr1[stride_w];
-                        pp[10] = sptr2[stride_w];
-                        pp[11] = sptr3[stride_w];
-                        pp[12] = sptr4[stride_w];
-                        pp[13] = sptr5[stride_w];
-                        pp[14] = sptr6[stride_w];
-                        pp[15] = sptr7[stride_w];
-                        pp[16] = sptr0[stride_w * 2];
-                        pp[17] = sptr1[stride_w * 2];
-                        pp[18] = sptr2[stride_w * 2];
-                        pp[19] = sptr3[stride_w * 2];
-                        pp[20] = sptr4[stride_w * 2];
-                        pp[21] = sptr5[stride_w * 2];
-                        pp[22] = sptr6[stride_w * 2];
-                        pp[23] = sptr7[stride_w * 2];
-                        pp[24] = sptr0[stride_w * 3];
-                        pp[25] = sptr1[stride_w * 3];
-                        pp[26] = sptr2[stride_w * 3];
-                        pp[27] = sptr3[stride_w * 3];
-                        pp[28] = sptr4[stride_w * 3];
-                        pp[29] = sptr5[stride_w * 3];
-                        pp[30] = sptr6[stride_w * 3];
-                        pp[31] = sptr7[stride_w * 3];
-                        pp[32] = sptr0[stride_w * 4];
-                        pp[33] = sptr1[stride_w * 4];
-                        pp[34] = sptr2[stride_w * 4];
-                        pp[35] = sptr3[stride_w * 4];
-                        pp[36] = sptr4[stride_w * 4];
-                        pp[37] = sptr5[stride_w * 4];
-                        pp[38] = sptr6[stride_w * 4];
-                        pp[39] = sptr7[stride_w * 4];
-                        pp[40] = sptr0[stride_w * 5];
-                        pp[41] = sptr1[stride_w * 5];
-                        pp[42] = sptr2[stride_w * 5];
-                        pp[43] = sptr3[stride_w * 5];
-                        pp[44] = sptr4[stride_w * 5];
-                        pp[45] = sptr5[stride_w * 5];
-                        pp[46] = sptr6[stride_w * 5];
-                        pp[47] = sptr7[stride_w * 5];
-                        pp[48] = sptr0[stride_w * 6];
-                        pp[49] = sptr1[stride_w * 6];
-                        pp[50] = sptr2[stride_w * 6];
-                        pp[51] = sptr3[stride_w * 6];
-                        pp[52] = sptr4[stride_w * 6];
-                        pp[53] = sptr5[stride_w * 6];
-                        pp[54] = sptr6[stride_w * 6];
-                        pp[55] = sptr7[stride_w * 6];
-                        pp[56] = sptr0[stride_w * 7];
-                        pp[57] = sptr1[stride_w * 7];
-                        pp[58] = sptr2[stride_w * 7];
-                        pp[59] = sptr3[stride_w * 7];
-                        pp[60] = sptr4[stride_w * 7];
-                        pp[61] = sptr5[stride_w * 7];
-                        pp[62] = sptr6[stride_w * 7];
-                        pp[63] = sptr7[stride_w * 7];
-                        pp += 64;
-                    }
-                }
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr2 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr3 = img3.row<const signed char>(y30) + x30;
-
-                    if (stride_w == 1)
-                    {
-                        int8x8x4_t _r01;
-                        _r01.val[0] = vld1_s8(sptr0);
-                        _r01.val[1] = vld1_s8(sptr1);
-                        _r01.val[2] = vld1_s8(sptr2);
-                        _r01.val[3] = vld1_s8(sptr3);
-                        vst4_s8(pp, _r01);
-                        pp += 32;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        int8x16_t _r0 = vld1q_s8(sptr0);
-                        int8x16_t _r1 = vld1q_s8(sptr1);
-                        int8x16_t _r2 = vld1q_s8(sptr2);
-                        int8x16_t _r3 = vld1q_s8(sptr3);
-                        int8x16_t _r01 = vtrnq_s8(_r0, _r1).val[0];
-                        int8x16_t _r23 = vtrnq_s8(_r2, _r3).val[0];
-                        int16x8x2_t _r0123;
-                        _r0123.val[0] = vreinterpretq_s16_s8(_r01);
-                        _r0123.val[1] = vreinterpretq_s16_s8(_r23);
-                        vst2q_s16((short*)pp, _r0123);
-                        pp += 32;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr2[0];
-                        pp[3] = sptr3[0];
-                        pp[4] = sptr0[stride_w];
-                        pp[5] = sptr1[stride_w];
-                        pp[6] = sptr2[stride_w];
-                        pp[7] = sptr3[stride_w];
-                        pp[8] = sptr0[stride_w * 2];
-                        pp[9] = sptr1[stride_w * 2];
-                        pp[10] = sptr2[stride_w * 2];
-                        pp[11] = sptr3[stride_w * 2];
-                        pp[12] = sptr0[stride_w * 3];
-                        pp[13] = sptr1[stride_w * 3];
-                        pp[14] = sptr2[stride_w * 3];
-                        pp[15] = sptr3[stride_w * 3];
-                        pp[16] = sptr0[stride_w * 4];
-                        pp[17] = sptr1[stride_w * 4];
-                        pp[18] = sptr2[stride_w * 4];
-                        pp[19] = sptr3[stride_w * 4];
-                        pp[20] = sptr0[stride_w * 5];
-                        pp[21] = sptr1[stride_w * 5];
-                        pp[22] = sptr2[stride_w * 5];
-                        pp[23] = sptr3[stride_w * 5];
-                        pp[24] = sptr0[stride_w * 6];
-                        pp[25] = sptr1[stride_w * 6];
-                        pp[26] = sptr2[stride_w * 6];
-                        pp[27] = sptr3[stride_w * 6];
-                        pp[28] = sptr0[stride_w * 7];
-                        pp[29] = sptr1[stride_w * 7];
-                        pp[30] = sptr2[stride_w * 7];
-                        pp[31] = sptr3[stride_w * 7];
-                        pp += 32;
-                    }
-                }
-#endif // __ARM_FEATURE_DOTPROD
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-
-                    if (stride_w == 1)
-                    {
-                        int8x8x2_t _r01;
-                        _r01.val[0] = vld1_s8(sptr0);
-                        _r01.val[1] = vld1_s8(sptr1);
-                        vst2_s8(pp, _r01);
-                        pp += 16;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        int8x16_t _r0 = vld1q_s8(sptr0);
-                        int8x16_t _r1 = vld1q_s8(sptr1);
-                        int8x16_t _r01 = vtrnq_s8(_r0, _r1).val[0];
-                        vst1q_s8(pp, _r01);
-                        pp += 16;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr0[stride_w];
-                        pp[3] = sptr1[stride_w];
-                        pp[4] = sptr0[stride_w * 2];
-                        pp[5] = sptr1[stride_w * 2];
-                        pp[6] = sptr0[stride_w * 3];
-                        pp[7] = sptr1[stride_w * 3];
-                        pp[8] = sptr0[stride_w * 4];
-                        pp[9] = sptr1[stride_w * 4];
-                        pp[10] = sptr0[stride_w * 5];
-                        pp[11] = sptr1[stride_w * 5];
-                        pp[12] = sptr0[stride_w * 6];
-                        pp[13] = sptr1[stride_w * 6];
-                        pp[14] = sptr0[stride_w * 7];
-                        pp[15] = sptr1[stride_w * 7];
-                        pp += 16;
-                    }
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = dx0 + dilation_w * v;
-                int y0 = dy0 + dilation_h * u;
-
-                const signed char* sptr = img.row<const signed char>(y0) + x0 * elempack;
-
-                if (elempack == 8)
-                {
-#if __ARM_FEATURE_MATMUL_INT8
-                    int8x8_t _r0 = vld1_s8(sptr);
-                    int8x8_t _r1 = vld1_s8(sptr + stride_w * 8);
-                    int8x8_t _r2 = vld1_s8(sptr + stride_w * 16);
-                    int8x8_t _r3 = vld1_s8(sptr + stride_w * 24);
-                    int8x8_t _r4 = vld1_s8(sptr + stride_w * 32);
-                    int8x8_t _r5 = vld1_s8(sptr + stride_w * 40);
-                    int8x8_t _r6 = vld1_s8(sptr + stride_w * 48);
-                    int8x8_t _r7 = vld1_s8(sptr + stride_w * 56);
-                    vst1_s8(pp, _r0);
-                    vst1_s8(pp + 8, _r1);
-                    vst1_s8(pp + 16, _r2);
-                    vst1_s8(pp + 24, _r3);
-                    vst1_s8(pp + 32, _r4);
-                    vst1_s8(pp + 40, _r5);
-                    vst1_s8(pp + 48, _r6);
-                    vst1_s8(pp + 56, _r7);
-                    pp += 64;
-#elif __ARM_FEATURE_DOTPROD
-                    int32x2_t _r0 = vreinterpret_s32_s8(vld1_s8(sptr));
-                    int32x2_t _r1 = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 8));
-                    int32x2_t _r2 = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 16));
-                    int32x2_t _r3 = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 24));
-                    int32x2_t _r4 = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 32));
-                    int32x2_t _r5 = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 40));
-                    int32x2_t _r6 = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 48));
-                    int32x2_t _r7 = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 56));
-                    int32x2x2_t _r01 = vzip_s32(_r0, _r1);
-                    int32x2x2_t _r23 = vzip_s32(_r2, _r3);
-                    int32x2x2_t _r45 = vzip_s32(_r4, _r5);
-                    int32x2x2_t _r67 = vzip_s32(_r6, _r7);
-                    vst1_s32((int*)pp, _r01.val[0]);
-                    vst1_s32((int*)(pp + 8), _r23.val[0]);
-                    vst1_s32((int*)(pp + 16), _r45.val[0]);
-                    vst1_s32((int*)(pp + 24), _r67.val[0]);
-                    vst1_s32((int*)(pp + 32), _r01.val[1]);
-                    vst1_s32((int*)(pp + 40), _r23.val[1]);
-                    vst1_s32((int*)(pp + 48), _r45.val[1]);
-                    vst1_s32((int*)(pp + 56), _r67.val[1]);
-                    pp += 64;
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                    int16x4_t _r0 = vreinterpret_s16_s8(vld1_s8(sptr));
-                    int16x4_t _r1 = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 8));
-                    int16x4_t _r2 = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 16));
-                    int16x4_t _r3 = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 24));
-                    int16x4_t _r4 = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 32));
-                    int16x4_t _r5 = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 40));
-                    int16x4_t _r6 = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 48));
-                    int16x4_t _r7 = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 56));
-                    int16x4x2_t _r01 = vzip_s16(_r0, _r1);
-                    int16x4x2_t _r23 = vzip_s16(_r2, _r3);
-                    int16x4x2_t _r45 = vzip_s16(_r4, _r5);
-                    int16x4x2_t _r67 = vzip_s16(_r6, _r7);
-                    int32x4x4_t _r0123;
-                    _r0123.val[0] = vreinterpretq_s32_s16(vcombine_s16(_r01.val[0], _r01.val[1]));
-                    _r0123.val[1] = vreinterpretq_s32_s16(vcombine_s16(_r23.val[0], _r23.val[1]));
-                    _r0123.val[2] = vreinterpretq_s32_s16(vcombine_s16(_r45.val[0], _r45.val[1]));
-                    _r0123.val[3] = vreinterpretq_s32_s16(vcombine_s16(_r67.val[0], _r67.val[1]));
-                    vst4q_s32((int*)pp, _r0123);
-                    pp += 64;
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr[0];
-                    pp[1] = sptr[stride_w];
-                    pp[2] = sptr[stride_w * 2];
-                    pp[3] = sptr[stride_w * 3];
-                    pp[4] = sptr[stride_w * 4];
-                    pp[5] = sptr[stride_w * 5];
-                    pp[6] = sptr[stride_w * 6];
-                    pp[7] = sptr[stride_w * 7];
-                    pp += 8;
-                }
-            }
-        }
-        else
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int p4 = (k + kk + 4) / maxk;
-                    int p5 = (k + kk + 5) / maxk;
-                    int p6 = (k + kk + 6) / maxk;
-                    int p7 = (k + kk + 7) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int uv4 = (k + kk + 4) % maxk;
-                    int uv5 = (k + kk + 5) % maxk;
-                    int uv6 = (k + kk + 6) % maxk;
-                    int uv7 = (k + kk + 7) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int u4 = uv4 / kernel_w;
-                    int u5 = uv5 / kernel_w;
-                    int u6 = uv6 / kernel_w;
-                    int u7 = uv7 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-                    int v4 = uv4 % kernel_w;
-                    int v5 = uv5 % kernel_w;
-                    int v6 = uv6 % kernel_w;
-                    int v7 = uv7 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-                    const Mat img4 = bottom_blob.channel(p4);
-                    const Mat img5 = bottom_blob.channel(p5);
-                    const Mat img6 = bottom_blob.channel(p6);
-                    const Mat img7 = bottom_blob.channel(p7);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int x02 = dx2 + dilation_w * v0;
-                    int x03 = dx3 + dilation_w * v0;
-                    int x04 = dx4 + dilation_w * v0;
-                    int x05 = dx5 + dilation_w * v0;
-                    int x06 = dx6 + dilation_w * v0;
-                    int x07 = dx7 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int y02 = dy2 + dilation_h * u0;
-                    int y03 = dy3 + dilation_h * u0;
-                    int y04 = dy4 + dilation_h * u0;
-                    int y05 = dy5 + dilation_h * u0;
-                    int y06 = dy6 + dilation_h * u0;
-                    int y07 = dy7 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int x12 = dx2 + dilation_w * v1;
-                    int x13 = dx3 + dilation_w * v1;
-                    int x14 = dx4 + dilation_w * v1;
-                    int x15 = dx5 + dilation_w * v1;
-                    int x16 = dx6 + dilation_w * v1;
-                    int x17 = dx7 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-                    int y12 = dy2 + dilation_h * u1;
-                    int y13 = dy3 + dilation_h * u1;
-                    int y14 = dy4 + dilation_h * u1;
-                    int y15 = dy5 + dilation_h * u1;
-                    int y16 = dy6 + dilation_h * u1;
-                    int y17 = dy7 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int x21 = dx1 + dilation_w * v2;
-                    int x22 = dx2 + dilation_w * v2;
-                    int x23 = dx3 + dilation_w * v2;
-                    int x24 = dx4 + dilation_w * v2;
-                    int x25 = dx5 + dilation_w * v2;
-                    int x26 = dx6 + dilation_w * v2;
-                    int x27 = dx7 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int y21 = dy1 + dilation_h * u2;
-                    int y22 = dy2 + dilation_h * u2;
-                    int y23 = dy3 + dilation_h * u2;
-                    int y24 = dy4 + dilation_h * u2;
-                    int y25 = dy5 + dilation_h * u2;
-                    int y26 = dy6 + dilation_h * u2;
-                    int y27 = dy7 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int x31 = dx1 + dilation_w * v3;
-                    int x32 = dx2 + dilation_w * v3;
-                    int x33 = dx3 + dilation_w * v3;
-                    int x34 = dx4 + dilation_w * v3;
-                    int x35 = dx5 + dilation_w * v3;
-                    int x36 = dx6 + dilation_w * v3;
-                    int x37 = dx7 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-                    int y31 = dy1 + dilation_h * u3;
-                    int y32 = dy2 + dilation_h * u3;
-                    int y33 = dy3 + dilation_h * u3;
-                    int y34 = dy4 + dilation_h * u3;
-                    int y35 = dy5 + dilation_h * u3;
-                    int y36 = dy6 + dilation_h * u3;
-                    int y37 = dy7 + dilation_h * u3;
-
-                    int x40 = dx0 + dilation_w * v4;
-                    int x41 = dx1 + dilation_w * v4;
-                    int x42 = dx2 + dilation_w * v4;
-                    int x43 = dx3 + dilation_w * v4;
-                    int x44 = dx4 + dilation_w * v4;
-                    int x45 = dx5 + dilation_w * v4;
-                    int x46 = dx6 + dilation_w * v4;
-                    int x47 = dx7 + dilation_w * v4;
-                    int y40 = dy0 + dilation_h * u4;
-                    int y41 = dy1 + dilation_h * u4;
-                    int y42 = dy2 + dilation_h * u4;
-                    int y43 = dy3 + dilation_h * u4;
-                    int y44 = dy4 + dilation_h * u4;
-                    int y45 = dy5 + dilation_h * u4;
-                    int y46 = dy6 + dilation_h * u4;
-                    int y47 = dy7 + dilation_h * u4;
-
-                    int x50 = dx0 + dilation_w * v5;
-                    int x51 = dx1 + dilation_w * v5;
-                    int x52 = dx2 + dilation_w * v5;
-                    int x53 = dx3 + dilation_w * v5;
-                    int x54 = dx4 + dilation_w * v5;
-                    int x55 = dx5 + dilation_w * v5;
-                    int x56 = dx6 + dilation_w * v5;
-                    int x57 = dx7 + dilation_w * v5;
-                    int y50 = dy0 + dilation_h * u5;
-                    int y51 = dy1 + dilation_h * u5;
-                    int y52 = dy2 + dilation_h * u5;
-                    int y53 = dy3 + dilation_h * u5;
-                    int y54 = dy4 + dilation_h * u5;
-                    int y55 = dy5 + dilation_h * u5;
-                    int y56 = dy6 + dilation_h * u5;
-                    int y57 = dy7 + dilation_h * u5;
-
-                    int x60 = dx0 + dilation_w * v6;
-                    int x61 = dx1 + dilation_w * v6;
-                    int x62 = dx2 + dilation_w * v6;
-                    int x63 = dx3 + dilation_w * v6;
-                    int x64 = dx4 + dilation_w * v6;
-                    int x65 = dx5 + dilation_w * v6;
-                    int x66 = dx6 + dilation_w * v6;
-                    int x67 = dx7 + dilation_w * v6;
-                    int y60 = dy0 + dilation_h * u6;
-                    int y61 = dy1 + dilation_h * u6;
-                    int y62 = dy2 + dilation_h * u6;
-                    int y63 = dy3 + dilation_h * u6;
-                    int y64 = dy4 + dilation_h * u6;
-                    int y65 = dy5 + dilation_h * u6;
-                    int y66 = dy6 + dilation_h * u6;
-                    int y67 = dy7 + dilation_h * u6;
-
-                    int x70 = dx0 + dilation_w * v7;
-                    int x71 = dx1 + dilation_w * v7;
-                    int x72 = dx2 + dilation_w * v7;
-                    int x73 = dx3 + dilation_w * v7;
-                    int x74 = dx4 + dilation_w * v7;
-                    int x75 = dx5 + dilation_w * v7;
-                    int x76 = dx6 + dilation_w * v7;
-                    int x77 = dx7 + dilation_w * v7;
-                    int y70 = dy0 + dilation_h * u7;
-                    int y71 = dy1 + dilation_h * u7;
-                    int y72 = dy2 + dilation_h * u7;
-                    int y73 = dy3 + dilation_h * u7;
-                    int y74 = dy4 + dilation_h * u7;
-                    int y75 = dy5 + dilation_h * u7;
-                    int y76 = dy6 + dilation_h * u7;
-                    int y77 = dy7 + dilation_h * u7;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-                    const signed char* sptr04 = img0.row<const signed char>(y04) + x04;
-                    const signed char* sptr05 = img0.row<const signed char>(y05) + x05;
-                    const signed char* sptr06 = img0.row<const signed char>(y06) + x06;
-                    const signed char* sptr07 = img0.row<const signed char>(y07) + x07;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-                    const signed char* sptr14 = img1.row<const signed char>(y14) + x14;
-                    const signed char* sptr15 = img1.row<const signed char>(y15) + x15;
-                    const signed char* sptr16 = img1.row<const signed char>(y16) + x16;
-                    const signed char* sptr17 = img1.row<const signed char>(y17) + x17;
-
-                    const signed char* sptr20 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr21 = img2.row<const signed char>(y21) + x21;
-                    const signed char* sptr22 = img2.row<const signed char>(y22) + x22;
-                    const signed char* sptr23 = img2.row<const signed char>(y23) + x23;
-                    const signed char* sptr24 = img2.row<const signed char>(y24) + x24;
-                    const signed char* sptr25 = img2.row<const signed char>(y25) + x25;
-                    const signed char* sptr26 = img2.row<const signed char>(y26) + x26;
-                    const signed char* sptr27 = img2.row<const signed char>(y27) + x27;
-
-                    const signed char* sptr30 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr31 = img3.row<const signed char>(y31) + x31;
-                    const signed char* sptr32 = img3.row<const signed char>(y32) + x32;
-                    const signed char* sptr33 = img3.row<const signed char>(y33) + x33;
-                    const signed char* sptr34 = img3.row<const signed char>(y34) + x34;
-                    const signed char* sptr35 = img3.row<const signed char>(y35) + x35;
-                    const signed char* sptr36 = img3.row<const signed char>(y36) + x36;
-                    const signed char* sptr37 = img3.row<const signed char>(y37) + x37;
-
-                    const signed char* sptr40 = img4.row<const signed char>(y40) + x40;
-                    const signed char* sptr41 = img4.row<const signed char>(y41) + x41;
-                    const signed char* sptr42 = img4.row<const signed char>(y42) + x42;
-                    const signed char* sptr43 = img4.row<const signed char>(y43) + x43;
-                    const signed char* sptr44 = img4.row<const signed char>(y44) + x44;
-                    const signed char* sptr45 = img4.row<const signed char>(y45) + x45;
-                    const signed char* sptr46 = img4.row<const signed char>(y46) + x46;
-                    const signed char* sptr47 = img4.row<const signed char>(y47) + x47;
-
-                    const signed char* sptr50 = img5.row<const signed char>(y50) + x50;
-                    const signed char* sptr51 = img5.row<const signed char>(y51) + x51;
-                    const signed char* sptr52 = img5.row<const signed char>(y52) + x52;
-                    const signed char* sptr53 = img5.row<const signed char>(y53) + x53;
-                    const signed char* sptr54 = img5.row<const signed char>(y54) + x54;
-                    const signed char* sptr55 = img5.row<const signed char>(y55) + x55;
-                    const signed char* sptr56 = img5.row<const signed char>(y56) + x56;
-                    const signed char* sptr57 = img5.row<const signed char>(y57) + x57;
-
-                    const signed char* sptr60 = img6.row<const signed char>(y60) + x60;
-                    const signed char* sptr61 = img6.row<const signed char>(y61) + x61;
-                    const signed char* sptr62 = img6.row<const signed char>(y62) + x62;
-                    const signed char* sptr63 = img6.row<const signed char>(y63) + x63;
-                    const signed char* sptr64 = img6.row<const signed char>(y64) + x64;
-                    const signed char* sptr65 = img6.row<const signed char>(y65) + x65;
-                    const signed char* sptr66 = img6.row<const signed char>(y66) + x66;
-                    const signed char* sptr67 = img6.row<const signed char>(y67) + x67;
-
-                    const signed char* sptr70 = img7.row<const signed char>(y70) + x70;
-                    const signed char* sptr71 = img7.row<const signed char>(y71) + x71;
-                    const signed char* sptr72 = img7.row<const signed char>(y72) + x72;
-                    const signed char* sptr73 = img7.row<const signed char>(y73) + x73;
-                    const signed char* sptr74 = img7.row<const signed char>(y74) + x74;
-                    const signed char* sptr75 = img7.row<const signed char>(y75) + x75;
-                    const signed char* sptr76 = img7.row<const signed char>(y76) + x76;
-                    const signed char* sptr77 = img7.row<const signed char>(y77) + x77;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr20[0];
-                    pp[3] = sptr30[0];
-                    pp[4] = sptr40[0];
-                    pp[5] = sptr50[0];
-                    pp[6] = sptr60[0];
-                    pp[7] = sptr70[0];
-                    pp[8] = sptr01[0];
-                    pp[9] = sptr11[0];
-                    pp[10] = sptr21[0];
-                    pp[11] = sptr31[0];
-                    pp[12] = sptr41[0];
-                    pp[13] = sptr51[0];
-                    pp[14] = sptr61[0];
-                    pp[15] = sptr71[0];
-                    pp[16] = sptr02[0];
-                    pp[17] = sptr12[0];
-                    pp[18] = sptr22[0];
-                    pp[19] = sptr32[0];
-                    pp[20] = sptr42[0];
-                    pp[21] = sptr52[0];
-                    pp[22] = sptr62[0];
-                    pp[23] = sptr72[0];
-                    pp[24] = sptr03[0];
-                    pp[25] = sptr13[0];
-                    pp[26] = sptr23[0];
-                    pp[27] = sptr33[0];
-                    pp[28] = sptr43[0];
-                    pp[29] = sptr53[0];
-                    pp[30] = sptr63[0];
-                    pp[31] = sptr73[0];
-                    pp[32] = sptr04[0];
-                    pp[33] = sptr14[0];
-                    pp[34] = sptr24[0];
-                    pp[35] = sptr34[0];
-                    pp[36] = sptr44[0];
-                    pp[37] = sptr54[0];
-                    pp[38] = sptr64[0];
-                    pp[39] = sptr74[0];
-                    pp[40] = sptr05[0];
-                    pp[41] = sptr15[0];
-                    pp[42] = sptr25[0];
-                    pp[43] = sptr35[0];
-                    pp[44] = sptr45[0];
-                    pp[45] = sptr55[0];
-                    pp[46] = sptr65[0];
-                    pp[47] = sptr75[0];
-                    pp[48] = sptr06[0];
-                    pp[49] = sptr16[0];
-                    pp[50] = sptr26[0];
-                    pp[51] = sptr36[0];
-                    pp[52] = sptr46[0];
-                    pp[53] = sptr56[0];
-                    pp[54] = sptr66[0];
-                    pp[55] = sptr76[0];
-                    pp[56] = sptr07[0];
-                    pp[57] = sptr17[0];
-                    pp[58] = sptr27[0];
-                    pp[59] = sptr37[0];
-                    pp[60] = sptr47[0];
-                    pp[61] = sptr57[0];
-                    pp[62] = sptr67[0];
-                    pp[63] = sptr77[0];
-                    pp += 64;
-                }
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int x02 = dx2 + dilation_w * v0;
-                    int x03 = dx3 + dilation_w * v0;
-                    int x04 = dx4 + dilation_w * v0;
-                    int x05 = dx5 + dilation_w * v0;
-                    int x06 = dx6 + dilation_w * v0;
-                    int x07 = dx7 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int y02 = dy2 + dilation_h * u0;
-                    int y03 = dy3 + dilation_h * u0;
-                    int y04 = dy4 + dilation_h * u0;
-                    int y05 = dy5 + dilation_h * u0;
-                    int y06 = dy6 + dilation_h * u0;
-                    int y07 = dy7 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int x12 = dx2 + dilation_w * v1;
-                    int x13 = dx3 + dilation_w * v1;
-                    int x14 = dx4 + dilation_w * v1;
-                    int x15 = dx5 + dilation_w * v1;
-                    int x16 = dx6 + dilation_w * v1;
-                    int x17 = dx7 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-                    int y12 = dy2 + dilation_h * u1;
-                    int y13 = dy3 + dilation_h * u1;
-                    int y14 = dy4 + dilation_h * u1;
-                    int y15 = dy5 + dilation_h * u1;
-                    int y16 = dy6 + dilation_h * u1;
-                    int y17 = dy7 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int x21 = dx1 + dilation_w * v2;
-                    int x22 = dx2 + dilation_w * v2;
-                    int x23 = dx3 + dilation_w * v2;
-                    int x24 = dx4 + dilation_w * v2;
-                    int x25 = dx5 + dilation_w * v2;
-                    int x26 = dx6 + dilation_w * v2;
-                    int x27 = dx7 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int y21 = dy1 + dilation_h * u2;
-                    int y22 = dy2 + dilation_h * u2;
-                    int y23 = dy3 + dilation_h * u2;
-                    int y24 = dy4 + dilation_h * u2;
-                    int y25 = dy5 + dilation_h * u2;
-                    int y26 = dy6 + dilation_h * u2;
-                    int y27 = dy7 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int x31 = dx1 + dilation_w * v3;
-                    int x32 = dx2 + dilation_w * v3;
-                    int x33 = dx3 + dilation_w * v3;
-                    int x34 = dx4 + dilation_w * v3;
-                    int x35 = dx5 + dilation_w * v3;
-                    int x36 = dx6 + dilation_w * v3;
-                    int x37 = dx7 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-                    int y31 = dy1 + dilation_h * u3;
-                    int y32 = dy2 + dilation_h * u3;
-                    int y33 = dy3 + dilation_h * u3;
-                    int y34 = dy4 + dilation_h * u3;
-                    int y35 = dy5 + dilation_h * u3;
-                    int y36 = dy6 + dilation_h * u3;
-                    int y37 = dy7 + dilation_h * u3;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-                    const signed char* sptr04 = img0.row<const signed char>(y04) + x04;
-                    const signed char* sptr05 = img0.row<const signed char>(y05) + x05;
-                    const signed char* sptr06 = img0.row<const signed char>(y06) + x06;
-                    const signed char* sptr07 = img0.row<const signed char>(y07) + x07;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-                    const signed char* sptr14 = img1.row<const signed char>(y14) + x14;
-                    const signed char* sptr15 = img1.row<const signed char>(y15) + x15;
-                    const signed char* sptr16 = img1.row<const signed char>(y16) + x16;
-                    const signed char* sptr17 = img1.row<const signed char>(y17) + x17;
-
-                    const signed char* sptr20 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr21 = img2.row<const signed char>(y21) + x21;
-                    const signed char* sptr22 = img2.row<const signed char>(y22) + x22;
-                    const signed char* sptr23 = img2.row<const signed char>(y23) + x23;
-                    const signed char* sptr24 = img2.row<const signed char>(y24) + x24;
-                    const signed char* sptr25 = img2.row<const signed char>(y25) + x25;
-                    const signed char* sptr26 = img2.row<const signed char>(y26) + x26;
-                    const signed char* sptr27 = img2.row<const signed char>(y27) + x27;
-
-                    const signed char* sptr30 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr31 = img3.row<const signed char>(y31) + x31;
-                    const signed char* sptr32 = img3.row<const signed char>(y32) + x32;
-                    const signed char* sptr33 = img3.row<const signed char>(y33) + x33;
-                    const signed char* sptr34 = img3.row<const signed char>(y34) + x34;
-                    const signed char* sptr35 = img3.row<const signed char>(y35) + x35;
-                    const signed char* sptr36 = img3.row<const signed char>(y36) + x36;
-                    const signed char* sptr37 = img3.row<const signed char>(y37) + x37;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr20[0];
-                    pp[3] = sptr30[0];
-                    pp[4] = sptr01[0];
-                    pp[5] = sptr11[0];
-                    pp[6] = sptr21[0];
-                    pp[7] = sptr31[0];
-                    pp[8] = sptr02[0];
-                    pp[9] = sptr12[0];
-                    pp[10] = sptr22[0];
-                    pp[11] = sptr32[0];
-                    pp[12] = sptr03[0];
-                    pp[13] = sptr13[0];
-                    pp[14] = sptr23[0];
-                    pp[15] = sptr33[0];
-                    pp[16] = sptr04[0];
-                    pp[17] = sptr14[0];
-                    pp[18] = sptr24[0];
-                    pp[19] = sptr34[0];
-                    pp[20] = sptr05[0];
-                    pp[21] = sptr15[0];
-                    pp[22] = sptr25[0];
-                    pp[23] = sptr35[0];
-                    pp[24] = sptr06[0];
-                    pp[25] = sptr16[0];
-                    pp[26] = sptr26[0];
-                    pp[27] = sptr36[0];
-                    pp[28] = sptr07[0];
-                    pp[29] = sptr17[0];
-                    pp[30] = sptr27[0];
-                    pp[31] = sptr37[0];
-                    pp += 32;
-                }
-#endif // __ARM_FEATURE_DOTPROD
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int x02 = dx2 + dilation_w * v0;
-                    int x03 = dx3 + dilation_w * v0;
-                    int x04 = dx4 + dilation_w * v0;
-                    int x05 = dx5 + dilation_w * v0;
-                    int x06 = dx6 + dilation_w * v0;
-                    int x07 = dx7 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int y02 = dy2 + dilation_h * u0;
-                    int y03 = dy3 + dilation_h * u0;
-                    int y04 = dy4 + dilation_h * u0;
-                    int y05 = dy5 + dilation_h * u0;
-                    int y06 = dy6 + dilation_h * u0;
-                    int y07 = dy7 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int x12 = dx2 + dilation_w * v1;
-                    int x13 = dx3 + dilation_w * v1;
-                    int x14 = dx4 + dilation_w * v1;
-                    int x15 = dx5 + dilation_w * v1;
-                    int x16 = dx6 + dilation_w * v1;
-                    int x17 = dx7 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-                    int y12 = dy2 + dilation_h * u1;
-                    int y13 = dy3 + dilation_h * u1;
-                    int y14 = dy4 + dilation_h * u1;
-                    int y15 = dy5 + dilation_h * u1;
-                    int y16 = dy6 + dilation_h * u1;
-                    int y17 = dy7 + dilation_h * u1;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-                    const signed char* sptr04 = img0.row<const signed char>(y04) + x04;
-                    const signed char* sptr05 = img0.row<const signed char>(y05) + x05;
-                    const signed char* sptr06 = img0.row<const signed char>(y06) + x06;
-                    const signed char* sptr07 = img0.row<const signed char>(y07) + x07;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-                    const signed char* sptr14 = img1.row<const signed char>(y14) + x14;
-                    const signed char* sptr15 = img1.row<const signed char>(y15) + x15;
-                    const signed char* sptr16 = img1.row<const signed char>(y16) + x16;
-                    const signed char* sptr17 = img1.row<const signed char>(y17) + x17;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr01[0];
-                    pp[3] = sptr11[0];
-                    pp[4] = sptr02[0];
-                    pp[5] = sptr12[0];
-                    pp[6] = sptr03[0];
-                    pp[7] = sptr13[0];
-                    pp[8] = sptr04[0];
-                    pp[9] = sptr14[0];
-                    pp[10] = sptr05[0];
-                    pp[11] = sptr15[0];
-                    pp[12] = sptr06[0];
-                    pp[13] = sptr16[0];
-                    pp[14] = sptr07[0];
-                    pp[15] = sptr17[0];
-                    pp += 16;
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = dx0 + dilation_w * v;
-                int x1 = dx1 + dilation_w * v;
-                int x2 = dx2 + dilation_w * v;
-                int x3 = dx3 + dilation_w * v;
-                int x4 = dx4 + dilation_w * v;
-                int x5 = dx5 + dilation_w * v;
-                int x6 = dx6 + dilation_w * v;
-                int x7 = dx7 + dilation_w * v;
-                int y0 = dy0 + dilation_h * u;
-                int y1 = dy1 + dilation_h * u;
-                int y2 = dy2 + dilation_h * u;
-                int y3 = dy3 + dilation_h * u;
-                int y4 = dy4 + dilation_h * u;
-                int y5 = dy5 + dilation_h * u;
-                int y6 = dy6 + dilation_h * u;
-                int y7 = dy7 + dilation_h * u;
-
-                const signed char* sptr0 = img.row<const signed char>(y0) + x0 * elempack;
-                const signed char* sptr1 = img.row<const signed char>(y1) + x1 * elempack;
-                const signed char* sptr2 = img.row<const signed char>(y2) + x2 * elempack;
-                const signed char* sptr3 = img.row<const signed char>(y3) + x3 * elempack;
-                const signed char* sptr4 = img.row<const signed char>(y4) + x4 * elempack;
-                const signed char* sptr5 = img.row<const signed char>(y5) + x5 * elempack;
-                const signed char* sptr6 = img.row<const signed char>(y6) + x6 * elempack;
-                const signed char* sptr7 = img.row<const signed char>(y7) + x7 * elempack;
-
-                if (elempack == 8)
-                {
-#if __ARM_FEATURE_MATMUL_INT8
-                    int8x8_t _r0 = vld1_s8(sptr0);
-                    int8x8_t _r1 = vld1_s8(sptr1);
-                    int8x8_t _r2 = vld1_s8(sptr2);
-                    int8x8_t _r3 = vld1_s8(sptr3);
-                    int8x8_t _r4 = vld1_s8(sptr4);
-                    int8x8_t _r5 = vld1_s8(sptr5);
-                    int8x8_t _r6 = vld1_s8(sptr6);
-                    int8x8_t _r7 = vld1_s8(sptr7);
-                    vst1_s8(pp, _r0);
-                    vst1_s8(pp + 8, _r1);
-                    vst1_s8(pp + 16, _r2);
-                    vst1_s8(pp + 24, _r3);
-                    vst1_s8(pp + 32, _r4);
-                    vst1_s8(pp + 40, _r5);
-                    vst1_s8(pp + 48, _r6);
-                    vst1_s8(pp + 56, _r7);
-                    pp += 64;
-#elif __ARM_FEATURE_DOTPROD
-                    int32x2_t _r0 = vreinterpret_s32_s8(vld1_s8(sptr0));
-                    int32x2_t _r1 = vreinterpret_s32_s8(vld1_s8(sptr1));
-                    int32x2_t _r2 = vreinterpret_s32_s8(vld1_s8(sptr2));
-                    int32x2_t _r3 = vreinterpret_s32_s8(vld1_s8(sptr3));
-                    int32x2_t _r4 = vreinterpret_s32_s8(vld1_s8(sptr4));
-                    int32x2_t _r5 = vreinterpret_s32_s8(vld1_s8(sptr5));
-                    int32x2_t _r6 = vreinterpret_s32_s8(vld1_s8(sptr6));
-                    int32x2_t _r7 = vreinterpret_s32_s8(vld1_s8(sptr7));
-                    int32x2x2_t _r01 = vzip_s32(_r0, _r1);
-                    int32x2x2_t _r23 = vzip_s32(_r2, _r3);
-                    int32x2x2_t _r45 = vzip_s32(_r4, _r5);
-                    int32x2x2_t _r67 = vzip_s32(_r6, _r7);
-                    vst1_s32((int*)pp, _r01.val[0]);
-                    vst1_s32((int*)(pp + 8), _r23.val[0]);
-                    vst1_s32((int*)(pp + 16), _r45.val[0]);
-                    vst1_s32((int*)(pp + 24), _r67.val[0]);
-                    vst1_s32((int*)(pp + 32), _r01.val[1]);
-                    vst1_s32((int*)(pp + 40), _r23.val[1]);
-                    vst1_s32((int*)(pp + 48), _r45.val[1]);
-                    vst1_s32((int*)(pp + 56), _r67.val[1]);
-                    pp += 64;
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                    int16x4_t _r0 = vreinterpret_s16_s8(vld1_s8(sptr0));
-                    int16x4_t _r1 = vreinterpret_s16_s8(vld1_s8(sptr1));
-                    int16x4_t _r2 = vreinterpret_s16_s8(vld1_s8(sptr2));
-                    int16x4_t _r3 = vreinterpret_s16_s8(vld1_s8(sptr3));
-                    int16x4_t _r4 = vreinterpret_s16_s8(vld1_s8(sptr4));
-                    int16x4_t _r5 = vreinterpret_s16_s8(vld1_s8(sptr5));
-                    int16x4_t _r6 = vreinterpret_s16_s8(vld1_s8(sptr6));
-                    int16x4_t _r7 = vreinterpret_s16_s8(vld1_s8(sptr7));
-                    int16x4x2_t _r01 = vzip_s16(_r0, _r1);
-                    int16x4x2_t _r23 = vzip_s16(_r2, _r3);
-                    int16x4x2_t _r45 = vzip_s16(_r4, _r5);
-                    int16x4x2_t _r67 = vzip_s16(_r6, _r7);
-                    int32x4x4_t _r0123;
-                    _r0123.val[0] = vreinterpretq_s32_s16(vcombine_s16(_r01.val[0], _r01.val[1]));
-                    _r0123.val[1] = vreinterpretq_s32_s16(vcombine_s16(_r23.val[0], _r23.val[1]));
-                    _r0123.val[2] = vreinterpretq_s32_s16(vcombine_s16(_r45.val[0], _r45.val[1]));
-                    _r0123.val[3] = vreinterpretq_s32_s16(vcombine_s16(_r67.val[0], _r67.val[1]));
-                    vst4q_s32((int*)pp, _r0123);
-                    pp += 64;
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr2[0];
-                    pp[3] = sptr3[0];
-                    pp[4] = sptr4[0];
-                    pp[5] = sptr5[0];
-                    pp[6] = sptr6[0];
-                    pp[7] = sptr7[0];
-                    pp += 8;
-                }
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-        int dy0 = (j + jj) / outw * stride_h;
-        int dy1 = (j + jj + 1) / outw * stride_h;
-        int dy2 = (j + jj + 2) / outw * stride_h;
-        int dy3 = (j + jj + 3) / outw * stride_h;
-        int dx0 = (j + jj) % outw * stride_w;
-        int dx1 = (j + jj + 1) % outw * stride_w;
-        int dx2 = (j + jj + 2) % outw * stride_w;
-        int dx3 = (j + jj + 3) % outw * stride_w;
-
-        if (dy0 == dy3)
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int p4 = (k + kk + 4) / maxk;
-                    int p5 = (k + kk + 5) / maxk;
-                    int p6 = (k + kk + 6) / maxk;
-                    int p7 = (k + kk + 7) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int uv4 = (k + kk + 4) % maxk;
-                    int uv5 = (k + kk + 5) % maxk;
-                    int uv6 = (k + kk + 6) % maxk;
-                    int uv7 = (k + kk + 7) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int u4 = uv4 / kernel_w;
-                    int u5 = uv5 / kernel_w;
-                    int u6 = uv6 / kernel_w;
-                    int u7 = uv7 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-                    int v4 = uv4 % kernel_w;
-                    int v5 = uv5 % kernel_w;
-                    int v6 = uv6 % kernel_w;
-                    int v7 = uv7 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-                    const Mat img4 = bottom_blob.channel(p4);
-                    const Mat img5 = bottom_blob.channel(p5);
-                    const Mat img6 = bottom_blob.channel(p6);
-                    const Mat img7 = bottom_blob.channel(p7);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-
-                    int x40 = dx0 + dilation_w * v4;
-                    int y40 = dy0 + dilation_h * u4;
-
-                    int x50 = dx0 + dilation_w * v5;
-                    int y50 = dy0 + dilation_h * u5;
-
-                    int x60 = dx0 + dilation_w * v6;
-                    int y60 = dy0 + dilation_h * u6;
-
-                    int x70 = dx0 + dilation_w * v7;
-                    int y70 = dy0 + dilation_h * u7;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr2 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr3 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr4 = img4.row<const signed char>(y40) + x40;
-                    const signed char* sptr5 = img5.row<const signed char>(y50) + x50;
-                    const signed char* sptr6 = img6.row<const signed char>(y60) + x60;
-                    const signed char* sptr7 = img7.row<const signed char>(y70) + x70;
-
-                    if (stride_w == 1)
-                    {
-                        int8x8_t _r0 = vld1_s8(sptr0);
-                        int8x8_t _r1 = vld1_s8(sptr1);
-                        int8x8_t _r2 = vld1_s8(sptr2);
-                        int8x8_t _r3 = vld1_s8(sptr3);
-                        int8x8_t _r4 = vld1_s8(sptr4);
-                        int8x8_t _r5 = vld1_s8(sptr5);
-                        int8x8_t _r6 = vld1_s8(sptr6);
-                        int8x8_t _r7 = vld1_s8(sptr7);
-                        int16x4x4_t _r0123;
-                        _r0123.val[0] = vreinterpret_s16_s8(vzip_s8(_r0, _r1).val[0]);
-                        _r0123.val[1] = vreinterpret_s16_s8(vzip_s8(_r2, _r3).val[0]);
-                        _r0123.val[2] = vreinterpret_s16_s8(vzip_s8(_r4, _r5).val[0]);
-                        _r0123.val[3] = vreinterpret_s16_s8(vzip_s8(_r6, _r7).val[0]);
-                        vst4_s16((short*)pp, _r0123);
-                        pp += 32;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        int8x8_t _r0 = vld1_s8(sptr0);
-                        int8x8_t _r1 = vld1_s8(sptr1);
-                        int8x8_t _r2 = vld1_s8(sptr2);
-                        int8x8_t _r3 = vld1_s8(sptr3);
-                        int8x8_t _r4 = vld1_s8(sptr4);
-                        int8x8_t _r5 = vld1_s8(sptr5);
-                        int8x8_t _r6 = vld1_s8(sptr6);
-                        int8x8_t _r7 = vld1_s8(sptr7);
-                        int8x8_t _r01 = vtrn_s8(_r0, _r1).val[0];
-                        int8x8_t _r23 = vtrn_s8(_r2, _r3).val[0];
-                        int8x8_t _r45 = vtrn_s8(_r4, _r5).val[0];
-                        int8x8_t _r67 = vtrn_s8(_r6, _r7).val[0];
-                        int16x4x4_t _r0123;
-                        _r0123.val[0] = vreinterpret_s16_s8(_r01);
-                        _r0123.val[1] = vreinterpret_s16_s8(_r23);
-                        _r0123.val[2] = vreinterpret_s16_s8(_r45);
-                        _r0123.val[3] = vreinterpret_s16_s8(_r67);
-                        vst4_s16((short*)pp, _r0123);
-                        pp += 32;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr2[0];
-                        pp[3] = sptr3[0];
-                        pp[4] = sptr4[0];
-                        pp[5] = sptr5[0];
-                        pp[6] = sptr6[0];
-                        pp[7] = sptr7[0];
-                        pp[8] = sptr0[stride_w];
-                        pp[9] = sptr1[stride_w];
-                        pp[10] = sptr2[stride_w];
-                        pp[11] = sptr3[stride_w];
-                        pp[12] = sptr4[stride_w];
-                        pp[13] = sptr5[stride_w];
-                        pp[14] = sptr6[stride_w];
-                        pp[15] = sptr7[stride_w];
-                        pp[16] = sptr0[stride_w * 2];
-                        pp[17] = sptr1[stride_w * 2];
-                        pp[18] = sptr2[stride_w * 2];
-                        pp[19] = sptr3[stride_w * 2];
-                        pp[20] = sptr4[stride_w * 2];
-                        pp[21] = sptr5[stride_w * 2];
-                        pp[22] = sptr6[stride_w * 2];
-                        pp[23] = sptr7[stride_w * 2];
-                        pp[24] = sptr0[stride_w * 3];
-                        pp[25] = sptr1[stride_w * 3];
-                        pp[26] = sptr2[stride_w * 3];
-                        pp[27] = sptr3[stride_w * 3];
-                        pp[28] = sptr4[stride_w * 3];
-                        pp[29] = sptr5[stride_w * 3];
-                        pp[30] = sptr6[stride_w * 3];
-                        pp[31] = sptr7[stride_w * 3];
-                        pp += 32;
-                    }
-                }
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr2 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr3 = img3.row<const signed char>(y30) + x30;
-
-                    if (stride_w == 1)
-                    {
-                        int8x8_t _r0 = vld1_s8(sptr0);
-                        int8x8_t _r1 = vld1_s8(sptr1);
-                        int8x8_t _r2 = vld1_s8(sptr2);
-                        int8x8_t _r3 = vld1_s8(sptr3);
-                        int16x4x2_t _r01;
-                        _r01.val[0] = vreinterpret_s16_s8(vzip_s8(_r0, _r1).val[0]);
-                        _r01.val[1] = vreinterpret_s16_s8(vzip_s8(_r2, _r3).val[0]);
-                        vst2_s16((short*)pp, _r01);
-                        pp += 16;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        int8x8_t _r0 = vld1_s8(sptr0);
-                        int8x8_t _r1 = vld1_s8(sptr1);
-                        int8x8_t _r2 = vld1_s8(sptr2);
-                        int8x8_t _r3 = vld1_s8(sptr3);
-                        int8x8_t _r01 = vtrn_s8(_r0, _r1).val[0];
-                        int8x8_t _r23 = vtrn_s8(_r2, _r3).val[0];
-                        int16x4x2_t _r0123;
-                        _r0123.val[0] = vreinterpret_s16_s8(_r01);
-                        _r0123.val[1] = vreinterpret_s16_s8(_r23);
-                        vst2_s16((short*)pp, _r0123);
-                        pp += 16;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr2[0];
-                        pp[3] = sptr3[0];
-                        pp[4] = sptr0[stride_w];
-                        pp[5] = sptr1[stride_w];
-                        pp[6] = sptr2[stride_w];
-                        pp[7] = sptr3[stride_w];
-                        pp[8] = sptr0[stride_w * 2];
-                        pp[9] = sptr1[stride_w * 2];
-                        pp[10] = sptr2[stride_w * 2];
-                        pp[11] = sptr3[stride_w * 2];
-                        pp[12] = sptr0[stride_w * 3];
-                        pp[13] = sptr1[stride_w * 3];
-                        pp[14] = sptr2[stride_w * 3];
-                        pp[15] = sptr3[stride_w * 3];
-                        pp += 16;
-                    }
-                }
-#endif // __ARM_FEATURE_DOTPROD
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-
-                    if (stride_w == 1)
-                    {
-                        int8x8_t _r0 = vld1_s8(sptr0);
-                        int8x8_t _r1 = vld1_s8(sptr1);
-                        int8x8_t _r01 = vzip_s8(_r0, _r1).val[0];
-                        vst1_s8(pp, _r01);
-                        pp += 8;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        int8x8_t _r0 = vld1_s8(sptr0);
-                        int8x8_t _r1 = vld1_s8(sptr1);
-                        int8x8_t _r01 = vtrn_s8(_r0, _r1).val[0];
-                        vst1_s8(pp, _r01);
-                        pp += 8;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr0[stride_w];
-                        pp[3] = sptr1[stride_w];
-                        pp[4] = sptr0[stride_w * 2];
-                        pp[5] = sptr1[stride_w * 2];
-                        pp[6] = sptr0[stride_w * 3];
-                        pp[7] = sptr1[stride_w * 3];
-                        pp += 8;
-                    }
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = dx0 + dilation_w * v;
-                int y0 = dy0 + dilation_h * u;
-
-                const signed char* sptr = img.row<const signed char>(y0) + x0 * elempack;
-
-                if (elempack == 8)
-                {
-#if __ARM_FEATURE_MATMUL_INT8
-                    int8x8_t _r0 = vld1_s8(sptr);
-                    int8x8_t _r1 = vld1_s8(sptr + stride_w * 8);
-                    int8x8_t _r2 = vld1_s8(sptr + stride_w * 16);
-                    int8x8_t _r3 = vld1_s8(sptr + stride_w * 24);
-                    vst1_s8(pp, _r0);
-                    vst1_s8(pp + 8, _r1);
-                    vst1_s8(pp + 16, _r2);
-                    vst1_s8(pp + 24, _r3);
-                    pp += 32;
-#elif __ARM_FEATURE_DOTPROD
-                    int32x2x4_t _r0123;
-                    _r0123.val[0] = vreinterpret_s32_s8(vld1_s8(sptr));
-                    _r0123.val[1] = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 8));
-                    _r0123.val[2] = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 16));
-                    _r0123.val[3] = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 24));
-                    vst4_s32((int*)pp, _r0123);
-                    pp += 32;
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                    int16x4x4_t _r0123;
-                    _r0123.val[0] = vreinterpret_s16_s8(vld1_s8(sptr));
-                    _r0123.val[1] = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 8));
-                    _r0123.val[2] = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 16));
-                    _r0123.val[3] = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 24));
-                    vst4_s16((short*)pp, _r0123);
-                    pp += 32;
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr[0];
-                    pp[1] = sptr[stride_w];
-                    pp[2] = sptr[stride_w * 2];
-                    pp[3] = sptr[stride_w * 3];
-                    pp += 4;
-                }
-            }
-        }
-        else
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int p4 = (k + kk + 4) / maxk;
-                    int p5 = (k + kk + 5) / maxk;
-                    int p6 = (k + kk + 6) / maxk;
-                    int p7 = (k + kk + 7) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int uv4 = (k + kk + 4) % maxk;
-                    int uv5 = (k + kk + 5) % maxk;
-                    int uv6 = (k + kk + 6) % maxk;
-                    int uv7 = (k + kk + 7) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int u4 = uv4 / kernel_w;
-                    int u5 = uv5 / kernel_w;
-                    int u6 = uv6 / kernel_w;
-                    int u7 = uv7 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-                    int v4 = uv4 % kernel_w;
-                    int v5 = uv5 % kernel_w;
-                    int v6 = uv6 % kernel_w;
-                    int v7 = uv7 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-                    const Mat img4 = bottom_blob.channel(p4);
-                    const Mat img5 = bottom_blob.channel(p5);
-                    const Mat img6 = bottom_blob.channel(p6);
-                    const Mat img7 = bottom_blob.channel(p7);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int x02 = dx2 + dilation_w * v0;
-                    int x03 = dx3 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int y02 = dy2 + dilation_h * u0;
-                    int y03 = dy3 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int x12 = dx2 + dilation_w * v1;
-                    int x13 = dx3 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-                    int y12 = dy2 + dilation_h * u1;
-                    int y13 = dy3 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int x21 = dx1 + dilation_w * v2;
-                    int x22 = dx2 + dilation_w * v2;
-                    int x23 = dx3 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int y21 = dy1 + dilation_h * u2;
-                    int y22 = dy2 + dilation_h * u2;
-                    int y23 = dy3 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int x31 = dx1 + dilation_w * v3;
-                    int x32 = dx2 + dilation_w * v3;
-                    int x33 = dx3 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-                    int y31 = dy1 + dilation_h * u3;
-                    int y32 = dy2 + dilation_h * u3;
-                    int y33 = dy3 + dilation_h * u3;
-
-                    int x40 = dx0 + dilation_w * v4;
-                    int x41 = dx1 + dilation_w * v4;
-                    int x42 = dx2 + dilation_w * v4;
-                    int x43 = dx3 + dilation_w * v4;
-                    int y40 = dy0 + dilation_h * u4;
-                    int y41 = dy1 + dilation_h * u4;
-                    int y42 = dy2 + dilation_h * u4;
-                    int y43 = dy3 + dilation_h * u4;
-
-                    int x50 = dx0 + dilation_w * v5;
-                    int x51 = dx1 + dilation_w * v5;
-                    int x52 = dx2 + dilation_w * v5;
-                    int x53 = dx3 + dilation_w * v5;
-                    int y50 = dy0 + dilation_h * u5;
-                    int y51 = dy1 + dilation_h * u5;
-                    int y52 = dy2 + dilation_h * u5;
-                    int y53 = dy3 + dilation_h * u5;
-
-                    int x60 = dx0 + dilation_w * v6;
-                    int x61 = dx1 + dilation_w * v6;
-                    int x62 = dx2 + dilation_w * v6;
-                    int x63 = dx3 + dilation_w * v6;
-                    int y60 = dy0 + dilation_h * u6;
-                    int y61 = dy1 + dilation_h * u6;
-                    int y62 = dy2 + dilation_h * u6;
-                    int y63 = dy3 + dilation_h * u6;
-
-                    int x70 = dx0 + dilation_w * v7;
-                    int x71 = dx1 + dilation_w * v7;
-                    int x72 = dx2 + dilation_w * v7;
-                    int x73 = dx3 + dilation_w * v7;
-                    int y70 = dy0 + dilation_h * u7;
-                    int y71 = dy1 + dilation_h * u7;
-                    int y72 = dy2 + dilation_h * u7;
-                    int y73 = dy3 + dilation_h * u7;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-
-                    const signed char* sptr20 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr21 = img2.row<const signed char>(y21) + x21;
-                    const signed char* sptr22 = img2.row<const signed char>(y22) + x22;
-                    const signed char* sptr23 = img2.row<const signed char>(y23) + x23;
-
-                    const signed char* sptr30 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr31 = img3.row<const signed char>(y31) + x31;
-                    const signed char* sptr32 = img3.row<const signed char>(y32) + x32;
-                    const signed char* sptr33 = img3.row<const signed char>(y33) + x33;
-
-                    const signed char* sptr40 = img4.row<const signed char>(y40) + x40;
-                    const signed char* sptr41 = img4.row<const signed char>(y41) + x41;
-                    const signed char* sptr42 = img4.row<const signed char>(y42) + x42;
-                    const signed char* sptr43 = img4.row<const signed char>(y43) + x43;
-
-                    const signed char* sptr50 = img5.row<const signed char>(y50) + x50;
-                    const signed char* sptr51 = img5.row<const signed char>(y51) + x51;
-                    const signed char* sptr52 = img5.row<const signed char>(y52) + x52;
-                    const signed char* sptr53 = img5.row<const signed char>(y53) + x53;
-
-                    const signed char* sptr60 = img6.row<const signed char>(y60) + x60;
-                    const signed char* sptr61 = img6.row<const signed char>(y61) + x61;
-                    const signed char* sptr62 = img6.row<const signed char>(y62) + x62;
-                    const signed char* sptr63 = img6.row<const signed char>(y63) + x63;
-
-                    const signed char* sptr70 = img7.row<const signed char>(y70) + x70;
-                    const signed char* sptr71 = img7.row<const signed char>(y71) + x71;
-                    const signed char* sptr72 = img7.row<const signed char>(y72) + x72;
-                    const signed char* sptr73 = img7.row<const signed char>(y73) + x73;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr20[0];
-                    pp[3] = sptr30[0];
-                    pp[4] = sptr40[0];
-                    pp[5] = sptr50[0];
-                    pp[6] = sptr60[0];
-                    pp[7] = sptr70[0];
-                    pp[8] = sptr01[0];
-                    pp[9] = sptr11[0];
-                    pp[10] = sptr21[0];
-                    pp[11] = sptr31[0];
-                    pp[12] = sptr41[0];
-                    pp[13] = sptr51[0];
-                    pp[14] = sptr61[0];
-                    pp[15] = sptr71[0];
-                    pp[16] = sptr02[0];
-                    pp[17] = sptr12[0];
-                    pp[18] = sptr22[0];
-                    pp[19] = sptr32[0];
-                    pp[20] = sptr42[0];
-                    pp[21] = sptr52[0];
-                    pp[22] = sptr62[0];
-                    pp[23] = sptr72[0];
-                    pp[24] = sptr03[0];
-                    pp[25] = sptr13[0];
-                    pp[26] = sptr23[0];
-                    pp[27] = sptr33[0];
-                    pp[28] = sptr43[0];
-                    pp[29] = sptr53[0];
-                    pp[30] = sptr63[0];
-                    pp[31] = sptr73[0];
-                    pp += 32;
-                }
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int x02 = dx2 + dilation_w * v0;
-                    int x03 = dx3 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int y02 = dy2 + dilation_h * u0;
-                    int y03 = dy3 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int x12 = dx2 + dilation_w * v1;
-                    int x13 = dx3 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-                    int y12 = dy2 + dilation_h * u1;
-                    int y13 = dy3 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int x21 = dx1 + dilation_w * v2;
-                    int x22 = dx2 + dilation_w * v2;
-                    int x23 = dx3 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int y21 = dy1 + dilation_h * u2;
-                    int y22 = dy2 + dilation_h * u2;
-                    int y23 = dy3 + dilation_h * u2;
-
-                    int x30 = dx0 + dilation_w * v3;
-                    int x31 = dx1 + dilation_w * v3;
-                    int x32 = dx2 + dilation_w * v3;
-                    int x33 = dx3 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-                    int y31 = dy1 + dilation_h * u3;
-                    int y32 = dy2 + dilation_h * u3;
-                    int y33 = dy3 + dilation_h * u3;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-
-                    const signed char* sptr20 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr21 = img2.row<const signed char>(y21) + x21;
-                    const signed char* sptr22 = img2.row<const signed char>(y22) + x22;
-                    const signed char* sptr23 = img2.row<const signed char>(y23) + x23;
-
-                    const signed char* sptr30 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr31 = img3.row<const signed char>(y31) + x31;
-                    const signed char* sptr32 = img3.row<const signed char>(y32) + x32;
-                    const signed char* sptr33 = img3.row<const signed char>(y33) + x33;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr20[0];
-                    pp[3] = sptr30[0];
-                    pp[4] = sptr01[0];
-                    pp[5] = sptr11[0];
-                    pp[6] = sptr21[0];
-                    pp[7] = sptr31[0];
-                    pp[8] = sptr02[0];
-                    pp[9] = sptr12[0];
-                    pp[10] = sptr22[0];
-                    pp[11] = sptr32[0];
-                    pp[12] = sptr03[0];
-                    pp[13] = sptr13[0];
-                    pp[14] = sptr23[0];
-                    pp[15] = sptr33[0];
-                    pp += 16;
-                }
-#endif // __ARM_FEATURE_DOTPROD
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int x02 = dx2 + dilation_w * v0;
-                    int x03 = dx3 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int y02 = dy2 + dilation_h * u0;
-                    int y03 = dy3 + dilation_h * u0;
-
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int x12 = dx2 + dilation_w * v1;
-                    int x13 = dx3 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-                    int y12 = dy2 + dilation_h * u1;
-                    int y13 = dy3 + dilation_h * u1;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr01[0];
-                    pp[3] = sptr11[0];
-                    pp[4] = sptr02[0];
-                    pp[5] = sptr12[0];
-                    pp[6] = sptr03[0];
-                    pp[7] = sptr13[0];
-                    pp += 8;
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = dx0 + dilation_w * v;
-                int x1 = dx1 + dilation_w * v;
-                int x2 = dx2 + dilation_w * v;
-                int x3 = dx3 + dilation_w * v;
-                int y0 = dy0 + dilation_h * u;
-                int y1 = dy1 + dilation_h * u;
-                int y2 = dy2 + dilation_h * u;
-                int y3 = dy3 + dilation_h * u;
-
-                const signed char* sptr0 = img.row<const signed char>(y0) + x0 * elempack;
-                const signed char* sptr1 = img.row<const signed char>(y1) + x1 * elempack;
-                const signed char* sptr2 = img.row<const signed char>(y2) + x2 * elempack;
-                const signed char* sptr3 = img.row<const signed char>(y3) + x3 * elempack;
-
-                if (elempack == 8)
-                {
-#if __ARM_FEATURE_MATMUL_INT8
-                    int8x8_t _r0 = vld1_s8(sptr0);
-                    int8x8_t _r1 = vld1_s8(sptr1);
-                    int8x8_t _r2 = vld1_s8(sptr2);
-                    int8x8_t _r3 = vld1_s8(sptr3);
-                    vst1_s8(pp, _r0);
-                    vst1_s8(pp + 8, _r1);
-                    vst1_s8(pp + 16, _r2);
-                    vst1_s8(pp + 24, _r3);
-                    pp += 32;
-#elif __ARM_FEATURE_DOTPROD
-                    int32x2x4_t _r0123;
-                    _r0123.val[0] = vreinterpret_s32_s8(vld1_s8(sptr0));
-                    _r0123.val[1] = vreinterpret_s32_s8(vld1_s8(sptr1));
-                    _r0123.val[2] = vreinterpret_s32_s8(vld1_s8(sptr2));
-                    _r0123.val[3] = vreinterpret_s32_s8(vld1_s8(sptr3));
-                    vst4_s32((int*)pp, _r0123);
-                    pp += 32;
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                    int16x4x4_t _r0123;
-                    _r0123.val[0] = vreinterpret_s16_s8(vld1_s8(sptr0));
-                    _r0123.val[1] = vreinterpret_s16_s8(vld1_s8(sptr1));
-                    _r0123.val[2] = vreinterpret_s16_s8(vld1_s8(sptr2));
-                    _r0123.val[3] = vreinterpret_s16_s8(vld1_s8(sptr3));
-                    vst4_s16((short*)pp, _r0123);
-                    pp += 32;
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr2[0];
-                    pp[3] = sptr3[0];
-                    pp += 4;
-                }
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-        int dy0 = (j + jj) / outw * stride_h;
-        int dy1 = (j + jj + 1) / outw * stride_h;
-        int dx0 = (j + jj) % outw * stride_w;
-        int dx1 = (j + jj + 1) % outw * stride_w;
-
-        if (dy0 == dy1)
-        {
-            int kk = 0;
-#if __ARM_NEON
-            if (elempack == 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int p4 = (k + kk + 4) / maxk;
-                    int p5 = (k + kk + 5) / maxk;
-                    int p6 = (k + kk + 6) / maxk;
-                    int p7 = (k + kk + 7) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int uv4 = (k + kk + 4) % maxk;
-                    int uv5 = (k + kk + 5) % maxk;
-                    int uv6 = (k + kk + 6) % maxk;
-                    int uv7 = (k + kk + 7) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int u4 = uv4 / kernel_w;
-                    int u5 = uv5 / kernel_w;
-                    int u6 = uv6 / kernel_w;
-                    int u7 = uv7 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-                    int v4 = uv4 % kernel_w;
-                    int v5 = uv5 % kernel_w;
-                    int v6 = uv6 % kernel_w;
-                    int v7 = uv7 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-                    const Mat img4 = bottom_blob.channel(p4);
-                    const Mat img5 = bottom_blob.channel(p5);
-                    const Mat img6 = bottom_blob.channel(p6);
-                    const Mat img7 = bottom_blob.channel(p7);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int x30 = dx0 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-
-                    int x40 = dx0 + dilation_w * v4;
-                    int y40 = dy0 + dilation_h * u4;
-                    int x50 = dx0 + dilation_w * v5;
-                    int y50 = dy0 + dilation_h * u5;
-
-                    int x60 = dx0 + dilation_w * v6;
-                    int y60 = dy0 + dilation_h * u6;
-                    int x70 = dx0 + dilation_w * v7;
-                    int y70 = dy0 + dilation_h * u7;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr2 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr3 = img3.row<const signed char>(y30) + x30;
-
-                    const signed char* sptr4 = img4.row<const signed char>(y40) + x40;
-                    const signed char* sptr5 = img5.row<const signed char>(y50) + x50;
-                    const signed char* sptr6 = img6.row<const signed char>(y60) + x60;
-                    const signed char* sptr7 = img7.row<const signed char>(y70) + x70;
-
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr2[0];
-                    pp[3] = sptr3[0];
-                    pp[4] = sptr4[0];
-                    pp[5] = sptr5[0];
-                    pp[6] = sptr6[0];
-                    pp[7] = sptr7[0];
-                    pp[8] = sptr0[stride_w];
-                    pp[9] = sptr1[stride_w];
-                    pp[10] = sptr2[stride_w];
-                    pp[11] = sptr3[stride_w];
-                    pp[12] = sptr4[stride_w];
-                    pp[13] = sptr5[stride_w];
-                    pp[14] = sptr6[stride_w];
-                    pp[15] = sptr7[stride_w];
-                    pp += 16;
-                }
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int x20 = dx0 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int x30 = dx0 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr2 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr3 = img3.row<const signed char>(y30) + x30;
-
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr2[0];
-                    pp[3] = sptr3[0];
-                    pp[4] = sptr0[stride_w];
-                    pp[5] = sptr1[stride_w];
-                    pp[6] = sptr2[stride_w];
-                    pp[7] = sptr3[stride_w];
-                    pp += 8;
-                }
-#endif // __ARM_FEATURE_DOTPROD
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int x10 = dx0 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr0[stride_w];
-                    pp[3] = sptr1[stride_w];
-                    pp += 4;
-                }
-            }
-#endif // __ARM_NEON
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = dx0 + dilation_w * v;
-                int y0 = dy0 + dilation_h * u;
-
-                const signed char* sptr = img.row<const signed char>(y0) + x0 * elempack;
-
-#if __ARM_NEON
-                if (elempack == 8)
-                {
-#if __ARM_FEATURE_MATMUL_INT8
-                    int8x8_t _r0 = vld1_s8(sptr);
-                    int8x8_t _r1 = vld1_s8(sptr + stride_w * 8);
-                    vst1_s8(pp, _r0);
-                    vst1_s8(pp + 8, _r1);
-                    pp += 16;
-#elif __ARM_FEATURE_DOTPROD
-                    int32x2x2_t _r01;
-                    _r01.val[0] = vreinterpret_s32_s8(vld1_s8(sptr));
-                    _r01.val[1] = vreinterpret_s32_s8(vld1_s8(sptr + stride_w * 8));
-                    vst2_s32((int*)pp, _r01);
-                    pp += 16;
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                    int16x4x2_t _r01;
-                    _r01.val[0] = vreinterpret_s16_s8(vld1_s8(sptr));
-                    _r01.val[1] = vreinterpret_s16_s8(vld1_s8(sptr + stride_w * 8));
-                    vst2_s16((short*)pp, _r01);
-                    pp += 16;
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                }
-#endif // __ARM_NEON
-                if (elempack == 1)
-                {
-                    pp[0] = sptr[0];
-                    pp[1] = sptr[stride_w];
-                    pp += 2;
-                }
-            }
-        }
-        else
-        {
-            int kk = 0;
-#if __ARM_NEON
-            if (elempack == 1)
-            {
-#if __ARM_FEATURE_DOTPROD
-#if __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 7 < max_kk; kk += 8)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int p4 = (k + kk + 4) / maxk;
-                    int p5 = (k + kk + 5) / maxk;
-                    int p6 = (k + kk + 6) / maxk;
-                    int p7 = (k + kk + 7) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int uv4 = (k + kk + 4) % maxk;
-                    int uv5 = (k + kk + 5) % maxk;
-                    int uv6 = (k + kk + 6) % maxk;
-                    int uv7 = (k + kk + 7) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int u4 = uv4 / kernel_w;
-                    int u5 = uv5 / kernel_w;
-                    int u6 = uv6 / kernel_w;
-                    int u7 = uv7 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-                    int v4 = uv4 % kernel_w;
-                    int v5 = uv5 % kernel_w;
-                    int v6 = uv6 % kernel_w;
-                    int v7 = uv7 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-                    const Mat img4 = bottom_blob.channel(p4);
-                    const Mat img5 = bottom_blob.channel(p5);
-                    const Mat img6 = bottom_blob.channel(p6);
-                    const Mat img7 = bottom_blob.channel(p7);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-
-                    int x20 = dx0 + dilation_w * v2;
-                    int x21 = dx1 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int y21 = dy1 + dilation_h * u2;
-                    int x30 = dx0 + dilation_w * v3;
-                    int x31 = dx1 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-                    int y31 = dy1 + dilation_h * u3;
-
-                    int x40 = dx0 + dilation_w * v4;
-                    int x41 = dx1 + dilation_w * v4;
-                    int y40 = dy0 + dilation_h * u4;
-                    int y41 = dy1 + dilation_h * u4;
-                    int x50 = dx0 + dilation_w * v5;
-                    int x51 = dx1 + dilation_w * v5;
-                    int y50 = dy0 + dilation_h * u5;
-                    int y51 = dy1 + dilation_h * u5;
-
-                    int x60 = dx0 + dilation_w * v6;
-                    int x61 = dx1 + dilation_w * v6;
-                    int y60 = dy0 + dilation_h * u6;
-                    int y61 = dy1 + dilation_h * u6;
-                    int x70 = dx0 + dilation_w * v7;
-                    int x71 = dx1 + dilation_w * v7;
-                    int y70 = dy0 + dilation_h * u7;
-                    int y71 = dy1 + dilation_h * u7;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr20 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr21 = img2.row<const signed char>(y21) + x21;
-                    const signed char* sptr30 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr31 = img3.row<const signed char>(y31) + x31;
-
-                    const signed char* sptr40 = img4.row<const signed char>(y40) + x40;
-                    const signed char* sptr41 = img4.row<const signed char>(y41) + x41;
-                    const signed char* sptr50 = img5.row<const signed char>(y50) + x50;
-                    const signed char* sptr51 = img5.row<const signed char>(y51) + x51;
-                    const signed char* sptr60 = img6.row<const signed char>(y60) + x60;
-                    const signed char* sptr61 = img6.row<const signed char>(y61) + x61;
-                    const signed char* sptr70 = img7.row<const signed char>(y70) + x70;
-                    const signed char* sptr71 = img7.row<const signed char>(y71) + x71;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr20[0];
-                    pp[3] = sptr30[0];
-                    pp[4] = sptr40[0];
-                    pp[5] = sptr50[0];
-                    pp[6] = sptr60[0];
-                    pp[7] = sptr70[0];
-                    pp[8] = sptr01[0];
-                    pp[9] = sptr11[0];
-                    pp[10] = sptr21[0];
-                    pp[11] = sptr31[0];
-                    pp[12] = sptr41[0];
-                    pp[13] = sptr51[0];
-                    pp[14] = sptr61[0];
-                    pp[15] = sptr71[0];
-                    pp += 16;
-                }
-#endif // __ARM_FEATURE_MATMUL_INT8
-                for (; kk + 3 < max_kk; kk += 4)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int p2 = (k + kk + 2) / maxk;
-                    int p3 = (k + kk + 3) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int uv2 = (k + kk + 2) % maxk;
-                    int uv3 = (k + kk + 3) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int u2 = uv2 / kernel_w;
-                    int u3 = uv3 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-                    int v2 = uv2 % kernel_w;
-                    int v3 = uv3 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-                    const Mat img2 = bottom_blob.channel(p2);
-                    const Mat img3 = bottom_blob.channel(p3);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-                    int x20 = dx0 + dilation_w * v2;
-                    int x21 = dx1 + dilation_w * v2;
-                    int y20 = dy0 + dilation_h * u2;
-                    int y21 = dy1 + dilation_h * u2;
-                    int x30 = dx0 + dilation_w * v3;
-                    int x31 = dx1 + dilation_w * v3;
-                    int y30 = dy0 + dilation_h * u3;
-                    int y31 = dy1 + dilation_h * u3;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr20 = img2.row<const signed char>(y20) + x20;
-                    const signed char* sptr21 = img2.row<const signed char>(y21) + x21;
-                    const signed char* sptr30 = img3.row<const signed char>(y30) + x30;
-                    const signed char* sptr31 = img3.row<const signed char>(y31) + x31;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr20[0];
-                    pp[3] = sptr30[0];
-                    pp[4] = sptr01[0];
-                    pp[5] = sptr11[0];
-                    pp[6] = sptr21[0];
-                    pp[7] = sptr31[0];
-                    pp += 8;
-                }
-#endif // __ARM_FEATURE_DOTPROD
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = dx0 + dilation_w * v0;
-                    int x01 = dx1 + dilation_w * v0;
-                    int y00 = dy0 + dilation_h * u0;
-                    int y01 = dy1 + dilation_h * u0;
-                    int x10 = dx0 + dilation_w * v1;
-                    int x11 = dx1 + dilation_w * v1;
-                    int y10 = dy0 + dilation_h * u1;
-                    int y11 = dy1 + dilation_h * u1;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr01[0];
-                    pp[3] = sptr11[0];
-                    pp += 4;
-                }
-            }
-#endif // __ARM_NEON
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = dx0 + dilation_w * v;
-                int x1 = dx1 + dilation_w * v;
-                int y0 = dy0 + dilation_h * u;
-                int y1 = dy1 + dilation_h * u;
-
-                const signed char* sptr0 = img.row<const signed char>(y0) + x0 * elempack;
-                const signed char* sptr1 = img.row<const signed char>(y1) + x1 * elempack;
-
-#if __ARM_NEON
-                if (elempack == 8)
-                {
-#if __ARM_FEATURE_MATMUL_INT8
-                    int8x8_t _r0 = vld1_s8(sptr0);
-                    int8x8_t _r1 = vld1_s8(sptr1);
-                    vst1_s8(pp, _r0);
-                    vst1_s8(pp + 8, _r1);
-                    pp += 16;
-#elif __ARM_FEATURE_DOTPROD
-                    int32x2x2_t _r01;
-                    _r01.val[0] = vreinterpret_s32_s8(vld1_s8(sptr0));
-                    _r01.val[1] = vreinterpret_s32_s8(vld1_s8(sptr1));
-                    vst2_s32((int*)pp, _r01);
-                    pp += 16;
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                    int16x4x2_t _r01;
-                    _r01.val[0] = vreinterpret_s16_s8(vld1_s8(sptr0));
-                    _r01.val[1] = vreinterpret_s16_s8(vld1_s8(sptr1));
-                    vst2_s16((short*)pp, _r01);
-                    pp += 16;
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                }
-#endif // __ARM_NEON
-                if (elempack == 1)
-                {
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp += 2;
-                }
-            }
-        }
-    }
-    for (; jj < max_jj; jj++)
-    {
-        int dy = (j + jj) / outw * stride_h;
-        int dx = (j + jj) % outw * stride_w;
-
-        int kk = 0;
-        for (; kk < max_kk / elempack; kk++)
-        {
-            int p = (k / elempack + kk) / maxk;
-            int uv = (k / elempack + kk) % maxk;
-            int u = uv / kernel_w;
-            int v = uv % kernel_w;
-
-            const Mat img = bottom_blob.channel(p);
-
-            int x = dx + dilation_w * v;
-            int y = dy + dilation_h * u;
-
-            const signed char* sptr = img.row<const signed char>(y) + x * elempack;
-
-#if __ARM_NEON
-            if (elempack == 8)
-            {
-                vst1_s8(pp, vld1_s8(sptr));
-                pp += 8;
-            }
-#endif // __ARM_NEON
-            if (elempack == 1)
-            {
-                pp[0] = sptr[0];
-                pp += 1;
-            }
-        }
-    }
-}
-
-static void convolution_im2col_gemm_transform_kernel_int8(const Mat& kernel, Mat& AT, int inch, int outch, int kernel_w, int kernel_h, const Option& opt)
-{
-#if !(__ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD)
-#if NCNN_RUNTIME_CPU && NCNN_ARM84I8MM && __aarch64__ && !__ARM_FEATURE_MATMUL_INT8
-    if (ncnn::cpu_support_arm_i8mm())
-    {
-        convolution_im2col_gemm_transform_kernel_int8_i8mm(kernel, AT, inch, outch, kernel_w, kernel_h, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_ARM82DOT && __aarch64__ && !__ARM_FEATURE_DOTPROD
-    if (ncnn::cpu_support_arm_asimddp())
-    {
-        convolution_im2col_gemm_transform_kernel_int8_asimddp(kernel, AT, inch, outch, kernel_w, kernel_h, opt);
-        return;
-    }
-#endif
-#endif
-
-    // NCNN_LOGE("convolution_im2col_gemm_transform_kernel");
-    const int maxk = kernel_w * kernel_h;
-
-    const int M = outch;
-    const int K = inch * maxk;
-
-    int TILE_M, TILE_N, TILE_K;
-    convolution_im2col_gemm_get_optimal_tile_mnk_int8(M, 0, K, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    int elempack = 1;
-#if __ARM_NEON
-    if (opt.use_packing_layout)
-    {
-        elempack = inch % 8 == 0 ? 8 : 1;
-    }
-#endif // __ARM_NEON
-
-    // maxk-inch-outch to pa-maxk-inch/pa-outch
-    Mat A_data;
-    if (maxk == 1)
-    {
-        A_data = kernel.reshape(maxk * inch, outch);
-    }
-    else
-    {
-        Mat weight_data_r2 = kernel.reshape(maxk, inch, outch);
-
-        A_data.create(maxk * inch, outch, (size_t)1u, 1);
-
-        for (int q = 0; q < outch; q += 1)
-        {
-            signed char* g00 = A_data.row<signed char>(q);
-
-            for (int p = 0; p + (elempack - 1) < inch; p += elempack)
-            {
-                for (int k = 0; k < maxk; k++)
-                {
-                    for (int i = 0; i < elempack; i++)
-                    {
-                        const signed char* k00 = weight_data_r2.channel(q).row<const signed char>(p + i);
-                        g00[0] = k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-
-    AT.create(TILE_K * TILE_M, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M, (size_t)1u, 1);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat AT_tile = AT.channel(i / TILE_M).row_range(k / TILE_K, 1);
-
-            convolution_im2col_pack_A_tile_int8(A_data, AT_tile, i, max_ii, k, max_kk);
-        }
-    }
-}
-
-static void convolution_im2col_gemm_int8(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt)
-{
-#if !(__ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD)
-#if NCNN_RUNTIME_CPU && NCNN_ARM84I8MM && __aarch64__ && !__ARM_FEATURE_MATMUL_INT8
-    if (ncnn::cpu_support_arm_i8mm())
-    {
-        convolution_im2col_gemm_int8_i8mm(bottom_blob, top_blob, AT, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, nT, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_ARM82DOT && __aarch64__ && !__ARM_FEATURE_DOTPROD
-    if (ncnn::cpu_support_arm_asimddp())
-    {
-        convolution_im2col_gemm_int8_asimddp(bottom_blob, top_blob, AT, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, nT, opt);
-        return;
-    }
-#endif
-#endif
-
-    const int maxk = kernel_w * kernel_h;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = top_blob.w * top_blob.h;
-    const int K = bottom_blob.c * bottom_blob.elempack * maxk;
-
-    int TILE_M, TILE_N, TILE_K;
-    convolution_im2col_gemm_get_optimal_tile_mnk_int8(M, N, K, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 1u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-    {
-        const int ppj = ppjk / nn_K;
-        const int ppk = ppjk % nn_K;
-
-        const int j = ppj * TILE_N;
-        const int k = ppk * TILE_K;
-
-        const int max_jj = std::min((N - j), TILE_N);
-        const int max_kk = std::min((K - k), TILE_K);
-
-        Mat BT_tile = BT.channel(j / TILE_N).row_range(k / TILE_K, 1);
-
-        // im2col
-        convolution_im2col_input_tile_int8(bottom_blob, BT_tile, j, max_jj, k, max_kk, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h);
-    }
-
-    Mat topT_tileX;
-    if (K > TILE_K)
-        topT_tileX.create(TILE_N * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat topT_tile;
-        if (K > TILE_K)
-            topT_tile = topT_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).row_range(k / TILE_K, 1);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).row_range(k / TILE_K, 1);
-
-                bool k_end = k + TILE_K >= K;
-
-                convolution_gemm_transB_packed_tile_int8(AT_tile, BT_tile, topT_tile, top_blob, i, max_ii, j, max_jj, k, max_kk, k_end);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed.h
deleted file mode 100644
index 0be5c1d4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed.h
+++ /dev/null
@@ -1,1321 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_transform_kernel_packed(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // src = kw-kh-inch-outch
-    // dst = pb-pa-kw-kh-inch/pa-outch/pb
-
-    // clang-format off
-    // *INDENT-OFF*
-#if __ARM_NEON
-#if __aarch64__
-    if (outch >= 8)
-    {
-        if (inch >= 8)
-            kernel_tm.create(8 * 8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2);
-        else if (inch >= 4)
-            kernel_tm.create(8 * 4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2);
-        else if (inch >= 2)
-            kernel_tm.create(8 * 2 * maxk, inch / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2);
-        else
-            kernel_tm.create(8 * maxk, inch, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2);
-    }
-    else
-#endif // __aarch64__
-    if (outch >= 4)
-    {
-#if __aarch64__
-        if (inch >= 8)
-            kernel_tm.create(4 * 8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2);
-        else
-#endif // __aarch64__
-        if (inch >= 4)
-            kernel_tm.create(4 * 4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2);
-        else if (inch >= 2)
-            kernel_tm.create(4 * 2 * maxk, inch / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2);
-        else
-            kernel_tm.create(4 * maxk, inch, outch / 4 + (outch % 4) / 2 + outch % 2);
-    }
-    else
-#endif // __ARM_NEON
-    if (outch >= 2)
-    {
-#if __ARM_NEON
-#if __aarch64__
-        if (inch >= 8)
-            kernel_tm.create(2 * 8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 2 + outch % 2);
-        else
-#endif // __aarch64__
-        if (inch >= 4)
-            kernel_tm.create(2 * 4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch / 2 + outch % 2);
-        else
-#endif // __ARM_NEON
-        if (inch >= 2)
-            kernel_tm.create(2 * 2 * maxk, inch / 2 + inch % 2, outch / 2 + outch % 2);
-        else
-            kernel_tm.create(2 * maxk, inch, outch / 2 + outch % 2);
-    }
-    else
-    {
-#if __ARM_NEON
-#if __aarch64__
-        if (inch >= 8)
-            kernel_tm.create(8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch);
-        else
-#endif // __aarch64__
-        if (inch >= 4)
-            kernel_tm.create(4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch);
-        else
-#endif // __ARM_NEON
-        if (inch >= 2)
-            kernel_tm.create(2 * maxk, inch / 2 + inch % 2, outch);
-        else
-            kernel_tm.create(maxk, inch, outch);
-    }
-    // *INDENT-ON*
-    // clang-format on
-
-    int q = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; q + 7 < outch; q += 8)
-    {
-        const float* kptr0 = (const float*)kernel + q * inch * maxk;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inch * maxk;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inch * maxk;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inch * maxk;
-        const float* kptr4 = (const float*)kernel + (q + 4) * inch * maxk;
-        const float* kptr5 = (const float*)kernel + (q + 5) * inch * maxk;
-        const float* kptr6 = (const float*)kernel + (q + 6) * inch * maxk;
-        const float* kptr7 = (const float*)kernel + (q + 7) * inch * maxk;
-
-        float* g00 = kernel_tm.channel(q / 8);
-
-        int p = 0;
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-                const float* k4 = kptr4 + p * maxk;
-                const float* k5 = kptr5 + p * maxk;
-                const float* k6 = kptr6 + p * maxk;
-                const float* k7 = kptr7 + p * maxk;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[k];
-                    g00[1] = k1[k];
-                    g00[2] = k2[k];
-                    g00[3] = k3[k];
-                    g00[4] = k4[k];
-                    g00[5] = k5[k];
-                    g00[6] = k6[k];
-                    g00[7] = k7[k];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    k4 += maxk;
-                    k5 += maxk;
-                    k6 += maxk;
-                    k7 += maxk;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-                const float* k4 = kptr4 + p * maxk;
-                const float* k5 = kptr5 + p * maxk;
-                const float* k6 = kptr6 + p * maxk;
-                const float* k7 = kptr7 + p * maxk;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k0[k];
-                    g00[1] = k1[k];
-                    g00[2] = k2[k];
-                    g00[3] = k3[k];
-                    g00[4] = k4[k];
-                    g00[5] = k5[k];
-                    g00[6] = k6[k];
-                    g00[7] = k7[k];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    k4 += maxk;
-                    k5 += maxk;
-                    k6 += maxk;
-                    k7 += maxk;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-                const float* k4 = kptr4 + p * maxk;
-                const float* k5 = kptr5 + p * maxk;
-                const float* k6 = kptr6 + p * maxk;
-                const float* k7 = kptr7 + p * maxk;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = k0[k];
-                    g00[1] = k1[k];
-                    g00[2] = k2[k];
-                    g00[3] = k3[k];
-                    g00[4] = k4[k];
-                    g00[5] = k5[k];
-                    g00[6] = k6[k];
-                    g00[7] = k7[k];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    k4 += maxk;
-                    k5 += maxk;
-                    k6 += maxk;
-                    k7 += maxk;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            const float* k0 = kptr0 + p * maxk;
-            const float* k1 = kptr1 + p * maxk;
-            const float* k2 = kptr2 + p * maxk;
-            const float* k3 = kptr3 + p * maxk;
-            const float* k4 = kptr4 + p * maxk;
-            const float* k5 = kptr5 + p * maxk;
-            const float* k6 = kptr6 + p * maxk;
-            const float* k7 = kptr7 + p * maxk;
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k0[k];
-                g00[1] = k1[k];
-                g00[2] = k2[k];
-                g00[3] = k3[k];
-                g00[4] = k4[k];
-                g00[5] = k5[k];
-                g00[6] = k6[k];
-                g00[7] = k7[k];
-                g00 += 8;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; q + 3 < outch; q += 4)
-    {
-        const float* kptr0 = (const float*)kernel + q * inch * maxk;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inch * maxk;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inch * maxk;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inch * maxk;
-
-#if __aarch64__
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4);
-#else
-        float* g00 = kernel_tm.channel(q / 4);
-#endif
-
-        int p = 0;
-#if __aarch64__
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[k];
-                    g00[1] = k1[k];
-                    g00[2] = k2[k];
-                    g00[3] = k3[k];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    g00 += 4;
-                }
-            }
-        }
-#endif // __aarch64__
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k0[k];
-                    g00[1] = k1[k];
-                    g00[2] = k2[k];
-                    g00[3] = k3[k];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    g00 += 4;
-                }
-            }
-        }
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = k0[k];
-                    g00[1] = k1[k];
-                    g00[2] = k2[k];
-                    g00[3] = k3[k];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    g00 += 4;
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            const float* k0 = kptr0 + p * maxk;
-            const float* k1 = kptr1 + p * maxk;
-            const float* k2 = kptr2 + p * maxk;
-            const float* k3 = kptr3 + p * maxk;
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k0[k];
-                g00[1] = k1[k];
-                g00[2] = k2[k];
-                g00[3] = k3[k];
-                g00 += 4;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; q + 1 < outch; q += 2)
-    {
-        const float* kptr0 = (const float*)kernel + q * inch * maxk;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inch * maxk;
-
-#if __aarch64__
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2);
-#elif __ARM_NEON
-        float* g00 = kernel_tm.channel(q / 4 + (q % 4) / 2);
-#else
-        float* g00 = kernel_tm.channel(q / 2);
-#endif
-
-        int p = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk + k;
-                const float* k1 = kptr1 + p * maxk + k;
-
-                g00[0] = k0[0];
-                g00[1] = k0[maxk];
-                g00[2] = k0[maxk * 2];
-                g00[3] = k0[maxk * 3];
-                g00[4] = k0[maxk * 4];
-                g00[5] = k0[maxk * 5];
-                g00[6] = k0[maxk * 6];
-                g00[7] = k0[maxk * 7];
-                g00[8] = k1[0];
-                g00[9] = k1[maxk];
-                g00[10] = k1[maxk * 2];
-                g00[11] = k1[maxk * 3];
-                g00[12] = k1[maxk * 4];
-                g00[13] = k1[maxk * 5];
-                g00[14] = k1[maxk * 6];
-                g00[15] = k1[maxk * 7];
-                g00 += 16;
-            }
-        }
-#endif // __aarch64__
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk + k;
-                const float* k1 = kptr1 + p * maxk + k;
-
-                g00[0] = k0[0];
-                g00[1] = k0[maxk];
-                g00[2] = k0[maxk * 2];
-                g00[3] = k0[maxk * 3];
-                g00[4] = k1[0];
-                g00[5] = k1[maxk];
-                g00[6] = k1[maxk * 2];
-                g00[7] = k1[maxk * 3];
-                g00 += 8;
-            }
-        }
-#endif // __ARM_NEON
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = k0[k];
-                    g00[1] = k1[k];
-                    k0 += maxk;
-                    k1 += maxk;
-                    g00 += 2;
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            const float* k0 = kptr0 + p * maxk;
-            const float* k1 = kptr1 + p * maxk;
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k0[k];
-                g00[1] = k1[k];
-                g00 += 2;
-            }
-        }
-    }
-    for (; q < outch; q++)
-    {
-        const float* kptr = (const float*)kernel + q * inch * maxk;
-
-#if __aarch64__
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2 + q % 2);
-#elif __ARM_NEON
-        float* g00 = kernel_tm.channel(q / 4 + (q % 4) / 2 + q % 2);
-#else
-        float* g00 = kernel_tm.channel(q / 2 + q % 2);
-#endif
-
-        int p = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr + p * maxk;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[k];
-                    k0 += maxk;
-                    g00 += 1;
-                }
-            }
-        }
-#endif // __aarch64__
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr + p * maxk;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k0[k];
-                    k0 += maxk;
-                    g00 += 1;
-                }
-            }
-        }
-#endif // __ARM_NEON
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr + p * maxk;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = k0[k];
-                    k0 += maxk;
-                    g00 += 1;
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            const float* k0 = kptr + p * maxk;
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k0[k];
-                g00++;
-            }
-        }
-    }
-}
-
-static void convolution_packed(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int elempack = bottom_blob.elempack;
-    const int inch = bottom_blob.c * elempack;
-
-    const int N = bottom_blob.cstep * elempack;
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int outch = top_blob.c * out_elempack;
-
-    const int M = top_blob.cstep * out_elempack;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2 * elempack;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    int nn_outch = 0;
-    int remain_outch_start = 0;
-#if __ARM_NEON
-#if __aarch64__
-    nn_outch = (outch - remain_outch_start) / 8;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 8;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int out_elempack = top_blob.elempack;
-
-        float* outptr = top_blob.channel(p / out_elempack);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                float32x4_t _sum2 = vdupq_n_f32(0.f);
-                float32x4_t _sum3 = vdupq_n_f32(0.f);
-                float32x4_t _sum4 = vdupq_n_f32(0.f);
-                float32x4_t _sum5 = vdupq_n_f32(0.f);
-                float32x4_t _sum6 = vdupq_n_f32(0.f);
-                float32x4_t _sum7 = vdupq_n_f32(0.f);
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = vld1q_f32(bias_data_ptr + p);
-                    _sum1 = vld1q_f32(bias_data_ptr + p + 4);
-                }
-
-                const float* kptr = weight_data_tm.channel(p / 8);
-
-                int q = 0;
-                for (; q + 7 < inch; q += 8)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        float32x4_t _r1;
-                        if (elempack == 4)
-                        {
-                            _r0 = vld1q_f32(r0 + sok);
-                            _r1 = vld1q_f32(r0 + sok + N);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float32x4_t();
-                            _r1 = float32x4_t();
-                            _r0 = vsetq_lane_f32(r0[sok], _r0, 0);
-                            _r0 = vsetq_lane_f32(r0[sok + N], _r0, 1);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 2], _r0, 2);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 3], _r0, 3);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 4], _r1, 0);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 5], _r1, 1);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 6], _r1, 2);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 7], _r1, 3);
-                        }
-
-                        float32x4_t _w0 = vld1q_f32(kptr);
-                        float32x4_t _w1 = vld1q_f32(kptr + 4);
-                        float32x4_t _w2 = vld1q_f32(kptr + 4 * 2);
-                        float32x4_t _w3 = vld1q_f32(kptr + 4 * 3);
-                        float32x4_t _w4 = vld1q_f32(kptr + 4 * 4);
-                        float32x4_t _w5 = vld1q_f32(kptr + 4 * 5);
-                        float32x4_t _w6 = vld1q_f32(kptr + 4 * 6);
-                        float32x4_t _w7 = vld1q_f32(kptr + 4 * 7);
-                        float32x4_t _w8 = vld1q_f32(kptr + 4 * 8);
-                        float32x4_t _w9 = vld1q_f32(kptr + 4 * 9);
-                        float32x4_t _wa = vld1q_f32(kptr + 4 * 10);
-                        float32x4_t _wb = vld1q_f32(kptr + 4 * 11);
-                        float32x4_t _wc = vld1q_f32(kptr + 4 * 12);
-                        float32x4_t _wd = vld1q_f32(kptr + 4 * 13);
-                        float32x4_t _we = vld1q_f32(kptr + 4 * 14);
-                        float32x4_t _wf = vld1q_f32(kptr + 4 * 15);
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 0);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 1);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 1);
-                        _sum4 = vfmaq_laneq_f32(_sum4, _w4, _r0, 2);
-                        _sum5 = vfmaq_laneq_f32(_sum5, _w5, _r0, 2);
-                        _sum6 = vfmaq_laneq_f32(_sum6, _w6, _r0, 3);
-                        _sum7 = vfmaq_laneq_f32(_sum7, _w7, _r0, 3);
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w8, _r1, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w9, _r1, 0);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _wa, _r1, 1);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _wb, _r1, 1);
-                        _sum4 = vfmaq_laneq_f32(_sum4, _wc, _r1, 2);
-                        _sum5 = vfmaq_laneq_f32(_sum5, _wd, _r1, 2);
-                        _sum6 = vfmaq_laneq_f32(_sum6, _we, _r1, 3);
-                        _sum7 = vfmaq_laneq_f32(_sum7, _wf, _r1, 3);
-
-                        kptr += 64;
-                    }
-                }
-                for (; q + 3 < inch; q += 4)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = vld1q_f32(r0 + sok);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float32x4_t();
-                            _r0 = vsetq_lane_f32(r0[sok], _r0, 0);
-                            _r0 = vsetq_lane_f32(r0[sok + N], _r0, 1);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 2], _r0, 2);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 3], _r0, 3);
-                        }
-
-                        float32x4_t _w0 = vld1q_f32(kptr);
-                        float32x4_t _w1 = vld1q_f32(kptr + 4);
-                        float32x4_t _w2 = vld1q_f32(kptr + 4 * 2);
-                        float32x4_t _w3 = vld1q_f32(kptr + 4 * 3);
-                        float32x4_t _w4 = vld1q_f32(kptr + 4 * 4);
-                        float32x4_t _w5 = vld1q_f32(kptr + 4 * 5);
-                        float32x4_t _w6 = vld1q_f32(kptr + 4 * 6);
-                        float32x4_t _w7 = vld1q_f32(kptr + 4 * 7);
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 0);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 1);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 1);
-                        _sum4 = vfmaq_laneq_f32(_sum4, _w4, _r0, 2);
-                        _sum5 = vfmaq_laneq_f32(_sum5, _w5, _r0, 2);
-                        _sum6 = vfmaq_laneq_f32(_sum6, _w6, _r0, 3);
-                        _sum7 = vfmaq_laneq_f32(_sum7, _w7, _r0, 3);
-
-                        kptr += 32;
-                    }
-                }
-                for (; q + 1 < inch; q += 2)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float val0;
-                        float val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = r0[sok];
-                            val1 = r0[sok + N];
-                        }
-
-                        float32x4_t _w0 = vld1q_f32(kptr);
-                        float32x4_t _w1 = vld1q_f32(kptr + 4);
-                        float32x4_t _w2 = vld1q_f32(kptr + 8);
-                        float32x4_t _w3 = vld1q_f32(kptr + 12);
-                        _sum0 = vfmaq_n_f32(_sum0, _w0, val0);
-                        _sum1 = vfmaq_n_f32(_sum1, _w1, val0);
-                        _sum2 = vfmaq_n_f32(_sum2, _w2, val1);
-                        _sum3 = vfmaq_n_f32(_sum3, _w3, val1);
-
-                        kptr += 16;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float32x4_t _val;
-                        // if (elempack == 1)
-                        {
-                            _val = vdupq_n_f32(r0[space_ofs[k]]);
-                        }
-
-                        float32x4_t _w0 = vld1q_f32(kptr);
-                        float32x4_t _w1 = vld1q_f32(kptr + 4);
-                        _sum0 = vfmaq_f32(_sum0, _w0, _val);
-                        _sum1 = vfmaq_f32(_sum1, _w1, _val);
-
-                        kptr += 8;
-                    }
-                }
-
-                _sum0 = vaddq_f32(_sum0, _sum2);
-                _sum1 = vaddq_f32(_sum1, _sum3);
-                _sum4 = vaddq_f32(_sum4, _sum6);
-                _sum5 = vaddq_f32(_sum5, _sum7);
-                _sum0 = vaddq_f32(_sum0, _sum4);
-                _sum1 = vaddq_f32(_sum1, _sum5);
-
-                _sum0 = activation_ps(_sum0, activation_type, activation_params);
-                _sum1 = activation_ps(_sum1, activation_type, activation_params);
-
-                if (out_elempack == 4)
-                {
-                    vst1q_f32(outptr, _sum0);
-                    vst1q_f32(outptr + M, _sum1);
-                    outptr += 4;
-                }
-                else // if (out_elempack == 1)
-                {
-                    outptr[0] = vgetq_lane_f32(_sum0, 0);
-                    outptr[M] = vgetq_lane_f32(_sum0, 1);
-                    outptr[M * 2] = vgetq_lane_f32(_sum0, 2);
-                    outptr[M * 3] = vgetq_lane_f32(_sum0, 3);
-                    outptr[M * 4] = vgetq_lane_f32(_sum1, 0);
-                    outptr[M * 5] = vgetq_lane_f32(_sum1, 1);
-                    outptr[M * 6] = vgetq_lane_f32(_sum1, 2);
-                    outptr[M * 7] = vgetq_lane_f32(_sum1, 3);
-                    outptr += 1;
-                }
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 8;
-    nn_outch = (outch - remain_outch_start) / 4;
-#else // __aarch64__
-    nn_outch = (outch - remain_outch_start) / 4;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __aarch64__
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 4;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int out_elempack = top_blob.elempack;
-
-        float* outptr = top_blob.channel(p / out_elempack);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                float32x4_t _sum2 = vdupq_n_f32(0.f);
-                float32x4_t _sum3 = vdupq_n_f32(0.f);
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = vld1q_f32(bias_data_ptr + p);
-                }
-
-#if __aarch64__
-                const float* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4);
-#else
-                const float* kptr = weight_data_tm.channel(p / 4);
-#endif
-
-                int q = 0;
-#if __aarch64__
-                for (; q + 7 < inch; q += 8)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        float32x4_t _r1;
-                        if (elempack == 4)
-                        {
-                            _r0 = vld1q_f32(r0 + sok);
-                            _r1 = vld1q_f32(r0 + sok + N);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float32x4_t();
-                            _r1 = float32x4_t();
-                            _r0 = vsetq_lane_f32(r0[sok], _r0, 0);
-                            _r0 = vsetq_lane_f32(r0[sok + N], _r0, 1);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 2], _r0, 2);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 3], _r0, 3);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 4], _r1, 0);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 5], _r1, 1);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 6], _r1, 2);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 7], _r1, 3);
-                        }
-
-                        float32x4_t _w0 = vld1q_f32(kptr);
-                        float32x4_t _w1 = vld1q_f32(kptr + 4);
-                        float32x4_t _w2 = vld1q_f32(kptr + 8);
-                        float32x4_t _w3 = vld1q_f32(kptr + 12);
-                        float32x4_t _w4 = vld1q_f32(kptr + 16);
-                        float32x4_t _w5 = vld1q_f32(kptr + 20);
-                        float32x4_t _w6 = vld1q_f32(kptr + 24);
-                        float32x4_t _w7 = vld1q_f32(kptr + 28);
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 1);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 2);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 3);
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w4, _r1, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w5, _r1, 1);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _w6, _r1, 2);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _w7, _r1, 3);
-
-                        kptr += 32;
-                    }
-                }
-#endif // __aarch64__
-                for (; q + 3 < inch; q += 4)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = vld1q_f32(r0 + sok);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float32x4_t();
-                            _r0 = vsetq_lane_f32(r0[sok], _r0, 0);
-                            _r0 = vsetq_lane_f32(r0[sok + N], _r0, 1);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 2], _r0, 2);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 3], _r0, 3);
-                        }
-
-                        float32x4_t _w0 = vld1q_f32(kptr);
-                        float32x4_t _w1 = vld1q_f32(kptr + 4);
-                        float32x4_t _w2 = vld1q_f32(kptr + 8);
-                        float32x4_t _w3 = vld1q_f32(kptr + 12);
-#if __aarch64__
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 1);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 2);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 3);
-#else
-                        _sum0 = vmlaq_lane_f32(_sum0, _w0, vget_low_f32(_r0), 0);
-                        _sum1 = vmlaq_lane_f32(_sum1, _w1, vget_low_f32(_r0), 1);
-                        _sum2 = vmlaq_lane_f32(_sum2, _w2, vget_high_f32(_r0), 0);
-                        _sum3 = vmlaq_lane_f32(_sum3, _w3, vget_high_f32(_r0), 1);
-#endif
-
-                        kptr += 16;
-                    }
-                }
-                for (; q + 1 < inch; q += 2)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float val0;
-                        float val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = r0[sok];
-                            val1 = r0[sok + N];
-                        }
-
-                        float32x4_t _w0 = vld1q_f32(kptr);
-                        float32x4_t _w1 = vld1q_f32(kptr + 4);
-#if __aarch64__
-                        _sum0 = vfmaq_n_f32(_sum0, _w0, val0);
-                        _sum1 = vfmaq_n_f32(_sum1, _w1, val1);
-#else
-                        _sum0 = vmlaq_n_f32(_sum0, _w0, val0);
-                        _sum1 = vmlaq_n_f32(_sum1, _w1, val1);
-#endif
-
-                        kptr += 8;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float32x4_t _val;
-                        // if (elempack == 1)
-                        {
-                            _val = vdupq_n_f32(r0[space_ofs[k]]);
-                        }
-
-                        float32x4_t _w = vld1q_f32(kptr);
-#if __aarch64__
-                        _sum0 = vfmaq_f32(_sum0, _val, _w);
-#else
-                        _sum0 = vmlaq_f32(_sum0, _val, _w);
-#endif
-
-                        kptr += 4;
-                    }
-                }
-
-                _sum0 = vaddq_f32(_sum0, _sum1);
-                _sum2 = vaddq_f32(_sum2, _sum3);
-                _sum0 = vaddq_f32(_sum0, _sum2);
-
-                _sum0 = activation_ps(_sum0, activation_type, activation_params);
-
-                if (out_elempack == 4)
-                {
-                    vst1q_f32(outptr, _sum0);
-                    outptr += 4;
-                }
-                else // if (out_elempack == 1)
-                {
-                    outptr[0] = vgetq_lane_f32(_sum0, 0);
-                    outptr[M] = vgetq_lane_f32(_sum0, 1);
-                    outptr[M * 2] = vgetq_lane_f32(_sum0, 2);
-                    outptr[M * 3] = vgetq_lane_f32(_sum0, 3);
-                    outptr += 1;
-                }
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 4;
-    nn_outch = (outch - remain_outch_start) / 2;
-#else // __ARM_NEON
-    nn_outch = (outch - remain_outch_start) / 2;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __ARM_NEON
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 2;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        float* outptr0 = top_blob.channel(p);
-        float* outptr1 = top_blob.channel(p + 1);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum0 = bias_data_ptr[p];
-                    sum1 = bias_data_ptr[p + 1];
-                }
-
-#if __aarch64__
-                const float* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2);
-#elif __ARM_NEON
-                const float* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2);
-#else
-                const float* kptr = weight_data_tm.channel(p / 2);
-#endif
-
-                int q = 0;
-#if __ARM_NEON
-#if __aarch64__
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                float32x4_t _sum2 = vdupq_n_f32(0.f);
-                float32x4_t _sum3 = vdupq_n_f32(0.f);
-                for (; q + 7 < inch; q += 8)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        float32x4_t _r1;
-                        if (elempack == 4)
-                        {
-                            _r0 = vld1q_f32(r0 + sok);
-                            _r1 = vld1q_f32(r0 + sok + N);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float32x4_t();
-                            _r1 = float32x4_t();
-                            _r0 = vsetq_lane_f32(r0[sok], _r0, 0);
-                            _r0 = vsetq_lane_f32(r0[sok + N], _r0, 1);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 2], _r0, 2);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 3], _r0, 3);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 4], _r1, 0);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 5], _r1, 1);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 6], _r1, 2);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 7], _r1, 3);
-                        }
-
-                        float32x4_t _w0 = vld1q_f32(kptr);
-                        float32x4_t _w1 = vld1q_f32(kptr + 4);
-                        float32x4_t _w2 = vld1q_f32(kptr + 8);
-                        float32x4_t _w3 = vld1q_f32(kptr + 12);
-                        _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                        _sum1 = vfmaq_f32(_sum1, _r1, _w1);
-                        _sum2 = vfmaq_f32(_sum2, _r0, _w2);
-                        _sum3 = vfmaq_f32(_sum3, _r1, _w3);
-
-                        kptr += 16;
-                    }
-                }
-                _sum0 = vaddq_f32(_sum0, _sum1);
-                _sum2 = vaddq_f32(_sum2, _sum3);
-                sum0 += vaddvq_f32(_sum0);
-                sum1 += vaddvq_f32(_sum2);
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-#else  // __aarch64__
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-#endif // __aarch64__
-                for (; q + 3 < inch; q += 4)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = vld1q_f32(r0 + sok);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float32x4_t();
-                            _r0 = vsetq_lane_f32(r0[sok], _r0, 0);
-                            _r0 = vsetq_lane_f32(r0[sok + N], _r0, 1);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 2], _r0, 2);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 3], _r0, 3);
-                        }
-
-                        float32x4_t _w0 = vld1q_f32(kptr);
-                        float32x4_t _w1 = vld1q_f32(kptr + 4);
-#if __aarch64__
-                        _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                        _sum1 = vfmaq_f32(_sum1, _r0, _w1);
-#else
-                        _sum0 = vmlaq_f32(_sum0, _r0, _w0);
-                        _sum1 = vmlaq_f32(_sum1, _r0, _w1);
-#endif
-
-                        kptr += 8;
-                    }
-                }
-#if __aarch64__
-                sum0 += vaddvq_f32(_sum0);
-                sum1 += vaddvq_f32(_sum1);
-#else
-                float32x2_t _ss0 = vadd_f32(vget_low_f32(_sum0), vget_high_f32(_sum0));
-                float32x2_t _ss1 = vadd_f32(vget_low_f32(_sum1), vget_high_f32(_sum1));
-                float32x2_t _ss = vpadd_f32(_ss0, _ss1);
-                sum0 += vget_lane_f32(_ss, 0);
-                sum1 += vget_lane_f32(_ss, 1);
-#endif
-#endif // __ARM_NEON
-                for (; q + 1 < inch; q += 2)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float val0;
-                        float val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = r0[sok];
-                            val1 = r0[sok + N];
-                        }
-
-                        sum0 += val0 * kptr[0];
-                        sum1 += val0 * kptr[1];
-                        sum0 += val1 * kptr[2];
-                        sum1 += val1 * kptr[3];
-
-                        kptr += 4;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float val;
-                        // if (elempack == 1)
-                        {
-                            val = r0[space_ofs[k]];
-                        }
-
-                        sum0 += val * kptr[0];
-                        sum1 += val * kptr[1];
-
-                        kptr += 2;
-                    }
-                }
-
-                sum0 = activation_ss(sum0, activation_type, activation_params);
-                sum1 = activation_ss(sum1, activation_type, activation_params);
-
-                outptr0[0] = sum0;
-                outptr1[0] = sum1;
-                outptr0 += 1;
-                outptr1 += 1;
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 2;
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-#if __aarch64__
-                const float* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#elif __ARM_NEON
-                const float* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2 + p % 2);
-#else
-                const float* kptr = weight_data_tm.channel(p / 2 + p % 2);
-#endif
-
-                int q = 0;
-#if __ARM_NEON
-#if __aarch64__
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                for (; q + 7 < inch; q += 8)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        float32x4_t _r1;
-                        if (elempack == 4)
-                        {
-                            _r0 = vld1q_f32(r0 + sok);
-                            _r1 = vld1q_f32(r0 + sok + N);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float32x4_t();
-                            _r1 = float32x4_t();
-                            _r0 = vsetq_lane_f32(r0[sok], _r0, 0);
-                            _r0 = vsetq_lane_f32(r0[sok + N], _r0, 1);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 2], _r0, 2);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 3], _r0, 3);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 4], _r1, 0);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 5], _r1, 1);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 6], _r1, 2);
-                            _r1 = vsetq_lane_f32(r0[sok + N * 7], _r1, 3);
-                        }
-
-                        float32x4_t _w0 = vld1q_f32(kptr);
-                        float32x4_t _w1 = vld1q_f32(kptr + 4);
-                        _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                        _sum1 = vfmaq_f32(_sum1, _r1, _w1);
-
-                        kptr += 8;
-                    }
-                }
-                _sum0 = vaddq_f32(_sum0, _sum1);
-                sum += vaddvq_f32(_sum0);
-#endif // __aarch64__
-                float32x4_t _sum = vdupq_n_f32(0.f);
-                for (; q + 3 < inch; q += 4)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = vld1q_f32(r0 + sok);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float32x4_t();
-                            _r0 = vsetq_lane_f32(r0[sok], _r0, 0);
-                            _r0 = vsetq_lane_f32(r0[sok + N], _r0, 1);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 2], _r0, 2);
-                            _r0 = vsetq_lane_f32(r0[sok + N * 3], _r0, 3);
-                        }
-
-                        float32x4_t _w = vld1q_f32(kptr);
-#if __aarch64__
-                        _sum = vfmaq_f32(_sum, _r0, _w);
-#else
-                        _sum = vmlaq_f32(_sum, _r0, _w);
-#endif
-
-                        kptr += 4;
-                    }
-                }
-#if __aarch64__
-                sum += vaddvq_f32(_sum);
-#else
-                float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-                _ss = vpadd_f32(_ss, _ss);
-                sum += vget_lane_f32(_ss, 0);
-#endif
-#endif // __ARM_NEON
-                for (; q + 1 < inch; q += 2)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float val0;
-                        float val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = r0[sok];
-                            val1 = r0[sok + N];
-                        }
-
-                        sum += val0 * kptr[0];
-                        sum += val1 * kptr[1];
-
-                        kptr += 2;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float val;
-                        // if (elempack == 1)
-                        {
-                            val = r0[space_ofs[k]];
-                        }
-
-                        sum += val * kptr[0];
-
-                        kptr += 1;
-                    }
-                }
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[0] = sum;
-                outptr += 1;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed_bf16s.h
deleted file mode 100644
index b4000da9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed_bf16s.h
+++ /dev/null
@@ -1,1358 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_transform_kernel_packed_bf16s(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // src = kw-kh-inch-outch
-    // dst = pb-pa-kw-kh-inch/pa-outch/pb
-
-    // clang-format off
-    // *INDENT-OFF*
-#if __ARM_NEON
-#if __aarch64__
-    if (outch >= 8)
-    {
-        if (inch >= 8)
-            kernel_tm.create(8 * 8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-        else if (inch >= 4)
-            kernel_tm.create(8 * 4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-        else if (inch >= 2)
-            kernel_tm.create(8 * 2 * maxk, inch / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-        else
-            kernel_tm.create(8 * maxk, inch, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-    }
-    else
-#endif // __aarch64__
-    if (outch >= 4)
-    {
-#if __aarch64__
-        if (inch >= 8)
-            kernel_tm.create(4 * 8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-        else
-#endif // __aarch64__
-        if (inch >= 4)
-            kernel_tm.create(4 * 4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-        else if (inch >= 2)
-            kernel_tm.create(4 * 2 * maxk, inch / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-        else
-            kernel_tm.create(4 * maxk, inch, outch / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-    }
-    else
-#endif // __ARM_NEON
-    if (outch >= 2)
-    {
-#if __ARM_NEON
-#if __aarch64__
-        if (inch >= 8)
-            kernel_tm.create(2 * 8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 2 + outch % 2, (size_t)2u);
-        else
-#endif // __aarch64__
-        if (inch >= 4)
-            kernel_tm.create(2 * 4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch / 2 + outch % 2, (size_t)2u);
-        else
-#endif // __ARM_NEON
-        if (inch >= 2)
-            kernel_tm.create(2 * 2 * maxk, inch / 2 + inch % 2, outch / 2 + outch % 2, (size_t)2u);
-        else
-            kernel_tm.create(2 * maxk, inch, outch / 2 + outch % 2, (size_t)2u);
-    }
-    else
-    {
-#if __ARM_NEON
-#if __aarch64__
-        if (inch >= 8)
-            kernel_tm.create(8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch, (size_t)2u);
-        else
-#endif // __aarch64__
-        if (inch >= 4)
-            kernel_tm.create(4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch, (size_t)2u);
-        else
-#endif // __ARM_NEON
-        if (inch >= 2)
-            kernel_tm.create(2 * maxk, inch / 2 + inch % 2, outch, (size_t)2u);
-        else
-            kernel_tm.create(maxk, inch, outch, (size_t)2u);
-    }
-    // *INDENT-ON*
-    // clang-format on
-
-    int q = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; q + 7 < outch; q += 8)
-    {
-        const float* kptr0 = (const float*)kernel + q * inch * maxk;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inch * maxk;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inch * maxk;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inch * maxk;
-        const float* kptr4 = (const float*)kernel + (q + 4) * inch * maxk;
-        const float* kptr5 = (const float*)kernel + (q + 5) * inch * maxk;
-        const float* kptr6 = (const float*)kernel + (q + 6) * inch * maxk;
-        const float* kptr7 = (const float*)kernel + (q + 7) * inch * maxk;
-
-        unsigned short* g00 = kernel_tm.channel(q / 8);
-
-        int p = 0;
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-                const float* k4 = kptr4 + p * maxk;
-                const float* k5 = kptr5 + p * maxk;
-                const float* k6 = kptr6 + p * maxk;
-                const float* k7 = kptr7 + p * maxk;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    g00[1] = float32_to_bfloat16(k1[k]);
-                    g00[2] = float32_to_bfloat16(k2[k]);
-                    g00[3] = float32_to_bfloat16(k3[k]);
-                    g00[4] = float32_to_bfloat16(k4[k]);
-                    g00[5] = float32_to_bfloat16(k5[k]);
-                    g00[6] = float32_to_bfloat16(k6[k]);
-                    g00[7] = float32_to_bfloat16(k7[k]);
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    k4 += maxk;
-                    k5 += maxk;
-                    k6 += maxk;
-                    k7 += maxk;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-                const float* k4 = kptr4 + p * maxk;
-                const float* k5 = kptr5 + p * maxk;
-                const float* k6 = kptr6 + p * maxk;
-                const float* k7 = kptr7 + p * maxk;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    g00[1] = float32_to_bfloat16(k1[k]);
-                    g00[2] = float32_to_bfloat16(k2[k]);
-                    g00[3] = float32_to_bfloat16(k3[k]);
-                    g00[4] = float32_to_bfloat16(k4[k]);
-                    g00[5] = float32_to_bfloat16(k5[k]);
-                    g00[6] = float32_to_bfloat16(k6[k]);
-                    g00[7] = float32_to_bfloat16(k7[k]);
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    k4 += maxk;
-                    k5 += maxk;
-                    k6 += maxk;
-                    k7 += maxk;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-                const float* k4 = kptr4 + p * maxk;
-                const float* k5 = kptr5 + p * maxk;
-                const float* k6 = kptr6 + p * maxk;
-                const float* k7 = kptr7 + p * maxk;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    g00[1] = float32_to_bfloat16(k1[k]);
-                    g00[2] = float32_to_bfloat16(k2[k]);
-                    g00[3] = float32_to_bfloat16(k3[k]);
-                    g00[4] = float32_to_bfloat16(k4[k]);
-                    g00[5] = float32_to_bfloat16(k5[k]);
-                    g00[6] = float32_to_bfloat16(k6[k]);
-                    g00[7] = float32_to_bfloat16(k7[k]);
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    k4 += maxk;
-                    k5 += maxk;
-                    k6 += maxk;
-                    k7 += maxk;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            const float* k0 = kptr0 + p * maxk;
-            const float* k1 = kptr1 + p * maxk;
-            const float* k2 = kptr2 + p * maxk;
-            const float* k3 = kptr3 + p * maxk;
-            const float* k4 = kptr4 + p * maxk;
-            const float* k5 = kptr5 + p * maxk;
-            const float* k6 = kptr6 + p * maxk;
-            const float* k7 = kptr7 + p * maxk;
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = float32_to_bfloat16(k0[k]);
-                g00[1] = float32_to_bfloat16(k1[k]);
-                g00[2] = float32_to_bfloat16(k2[k]);
-                g00[3] = float32_to_bfloat16(k3[k]);
-                g00[4] = float32_to_bfloat16(k4[k]);
-                g00[5] = float32_to_bfloat16(k5[k]);
-                g00[6] = float32_to_bfloat16(k6[k]);
-                g00[7] = float32_to_bfloat16(k7[k]);
-                g00 += 8;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; q + 3 < outch; q += 4)
-    {
-        const float* kptr0 = (const float*)kernel + q * inch * maxk;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inch * maxk;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inch * maxk;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inch * maxk;
-
-#if __aarch64__
-        unsigned short* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4);
-#else
-        unsigned short* g00 = kernel_tm.channel(q / 4);
-#endif
-
-        int p = 0;
-#if __aarch64__
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    g00[1] = float32_to_bfloat16(k1[k]);
-                    g00[2] = float32_to_bfloat16(k2[k]);
-                    g00[3] = float32_to_bfloat16(k3[k]);
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    g00 += 4;
-                }
-            }
-        }
-#endif // __aarch64__
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    g00[1] = float32_to_bfloat16(k1[k]);
-                    g00[2] = float32_to_bfloat16(k2[k]);
-                    g00[3] = float32_to_bfloat16(k3[k]);
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    g00 += 4;
-                }
-            }
-        }
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    g00[1] = float32_to_bfloat16(k1[k]);
-                    g00[2] = float32_to_bfloat16(k2[k]);
-                    g00[3] = float32_to_bfloat16(k3[k]);
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    g00 += 4;
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            const float* k0 = kptr0 + p * maxk;
-            const float* k1 = kptr1 + p * maxk;
-            const float* k2 = kptr2 + p * maxk;
-            const float* k3 = kptr3 + p * maxk;
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = float32_to_bfloat16(k0[k]);
-                g00[1] = float32_to_bfloat16(k1[k]);
-                g00[2] = float32_to_bfloat16(k2[k]);
-                g00[3] = float32_to_bfloat16(k3[k]);
-                g00 += 4;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; q + 1 < outch; q += 2)
-    {
-        const float* kptr0 = (const float*)kernel + q * inch * maxk;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inch * maxk;
-
-#if __aarch64__
-        unsigned short* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2);
-#elif __ARM_NEON
-        unsigned short* g00 = kernel_tm.channel(q / 4 + (q % 4) / 2);
-#else
-        unsigned short* g00 = kernel_tm.channel(q / 2);
-#endif
-
-        int p = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk + k;
-                const float* k1 = kptr1 + p * maxk + k;
-
-                g00[0] = float32_to_bfloat16(k0[0]);
-                g00[1] = float32_to_bfloat16(k0[maxk]);
-                g00[2] = float32_to_bfloat16(k0[maxk * 2]);
-                g00[3] = float32_to_bfloat16(k0[maxk * 3]);
-                g00[4] = float32_to_bfloat16(k0[maxk * 4]);
-                g00[5] = float32_to_bfloat16(k0[maxk * 5]);
-                g00[6] = float32_to_bfloat16(k0[maxk * 6]);
-                g00[7] = float32_to_bfloat16(k0[maxk * 7]);
-                g00[8] = float32_to_bfloat16(k1[0]);
-                g00[9] = float32_to_bfloat16(k1[maxk]);
-                g00[10] = float32_to_bfloat16(k1[maxk * 2]);
-                g00[11] = float32_to_bfloat16(k1[maxk * 3]);
-                g00[12] = float32_to_bfloat16(k1[maxk * 4]);
-                g00[13] = float32_to_bfloat16(k1[maxk * 5]);
-                g00[14] = float32_to_bfloat16(k1[maxk * 6]);
-                g00[15] = float32_to_bfloat16(k1[maxk * 7]);
-                g00 += 16;
-            }
-        }
-#endif // __aarch64__
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk + k;
-                const float* k1 = kptr1 + p * maxk + k;
-
-                g00[0] = float32_to_bfloat16(k0[0]);
-                g00[1] = float32_to_bfloat16(k0[maxk]);
-                g00[2] = float32_to_bfloat16(k0[maxk * 2]);
-                g00[3] = float32_to_bfloat16(k0[maxk * 3]);
-                g00[4] = float32_to_bfloat16(k1[0]);
-                g00[5] = float32_to_bfloat16(k1[maxk]);
-                g00[6] = float32_to_bfloat16(k1[maxk * 2]);
-                g00[7] = float32_to_bfloat16(k1[maxk * 3]);
-                g00 += 8;
-            }
-        }
-#endif // __ARM_NEON
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    g00[1] = float32_to_bfloat16(k1[k]);
-                    k0 += maxk;
-                    k1 += maxk;
-                    g00 += 2;
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            const float* k0 = kptr0 + p * maxk;
-            const float* k1 = kptr1 + p * maxk;
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = float32_to_bfloat16(k0[k]);
-                g00[1] = float32_to_bfloat16(k1[k]);
-                g00 += 2;
-            }
-        }
-    }
-    for (; q < outch; q++)
-    {
-        const float* kptr = (const float*)kernel + q * inch * maxk;
-
-#if __aarch64__
-        unsigned short* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2 + q % 2);
-#elif __ARM_NEON
-        unsigned short* g00 = kernel_tm.channel(q / 4 + (q % 4) / 2 + q % 2);
-#else
-        unsigned short* g00 = kernel_tm.channel(q / 2 + q % 2);
-#endif
-
-        int p = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr + p * maxk;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    k0 += maxk;
-                    g00 += 1;
-                }
-            }
-        }
-#endif // __aarch64__
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr + p * maxk;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    k0 += maxk;
-                    g00 += 1;
-                }
-            }
-        }
-#endif // __ARM_NEON
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr + p * maxk;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = float32_to_bfloat16(k0[k]);
-                    k0 += maxk;
-                    g00 += 1;
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            const float* k0 = kptr + p * maxk;
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = float32_to_bfloat16(k0[k]);
-                g00++;
-            }
-        }
-    }
-}
-
-static void convolution_packed_bf16s(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int elempack = bottom_blob.elempack;
-    const int inch = bottom_blob.c * elempack;
-
-    const int N = bottom_blob.cstep * elempack;
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int outch = top_blob.c * out_elempack;
-
-    const int M = top_blob.cstep * out_elempack;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2 * elempack;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    int nn_outch = 0;
-    int remain_outch_start = 0;
-#if __ARM_NEON
-#if __aarch64__
-    nn_outch = (outch - remain_outch_start) / 8;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 8;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int out_elempack = top_blob.elempack;
-
-        unsigned short* outptr = top_blob.channel(p / out_elempack);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                float32x4_t _sum2 = vdupq_n_f32(0.f);
-                float32x4_t _sum3 = vdupq_n_f32(0.f);
-                float32x4_t _sum4 = vdupq_n_f32(0.f);
-                float32x4_t _sum5 = vdupq_n_f32(0.f);
-                float32x4_t _sum6 = vdupq_n_f32(0.f);
-                float32x4_t _sum7 = vdupq_n_f32(0.f);
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = vld1q_f32(bias_data_ptr + p);
-                    _sum1 = vld1q_f32(bias_data_ptr + p + 4);
-                }
-
-                const unsigned short* kptr = weight_data_tm.channel(p / 8);
-
-                int q = 0;
-                for (; q + 7 < inch; q += 8)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q / elempack).row<const unsigned short>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        float32x4_t _r1;
-                        if (elempack == 4)
-                        {
-                            _r0 = bfloat2float(vld1_u16(r0 + sok));
-                            _r1 = bfloat2float(vld1_u16(r0 + sok + N));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            uint16x8_t _r_u16 = uint16x8_t();
-                            _r_u16 = vsetq_lane_u16(r0[sok], _r_u16, 0);
-                            _r_u16 = vsetq_lane_u16(r0[sok + N], _r_u16, 1);
-                            _r_u16 = vsetq_lane_u16(r0[sok + N * 2], _r_u16, 2);
-                            _r_u16 = vsetq_lane_u16(r0[sok + N * 3], _r_u16, 3);
-                            _r_u16 = vsetq_lane_u16(r0[sok + N * 4], _r_u16, 4);
-                            _r_u16 = vsetq_lane_u16(r0[sok + N * 5], _r_u16, 5);
-                            _r_u16 = vsetq_lane_u16(r0[sok + N * 6], _r_u16, 6);
-                            _r_u16 = vsetq_lane_u16(r0[sok + N * 7], _r_u16, 7);
-                            _r0 = bfloat2float(vget_low_u16(_r_u16));
-                            _r1 = bfloat2float(vget_high_u16(_r_u16));
-                        }
-
-                        uint16x8_t _w01 = vld1q_u16(kptr);
-                        uint16x8_t _w23 = vld1q_u16(kptr + 8);
-                        uint16x8_t _w45 = vld1q_u16(kptr + 16);
-                        uint16x8_t _w67 = vld1q_u16(kptr + 24);
-                        uint16x8_t _w89 = vld1q_u16(kptr + 32);
-                        uint16x8_t _wab = vld1q_u16(kptr + 40);
-                        uint16x8_t _wcd = vld1q_u16(kptr + 48);
-                        uint16x8_t _wef = vld1q_u16(kptr + 56);
-                        float32x4_t _w0 = bfloat2float(vget_low_u16(_w01));
-                        float32x4_t _w1 = bfloat2float(vget_high_u16(_w01));
-                        float32x4_t _w2 = bfloat2float(vget_low_u16(_w23));
-                        float32x4_t _w3 = bfloat2float(vget_high_u16(_w23));
-                        float32x4_t _w4 = bfloat2float(vget_low_u16(_w45));
-                        float32x4_t _w5 = bfloat2float(vget_high_u16(_w45));
-                        float32x4_t _w6 = bfloat2float(vget_low_u16(_w67));
-                        float32x4_t _w7 = bfloat2float(vget_high_u16(_w67));
-                        float32x4_t _w8 = bfloat2float(vget_low_u16(_w89));
-                        float32x4_t _w9 = bfloat2float(vget_high_u16(_w89));
-                        float32x4_t _wa = bfloat2float(vget_low_u16(_wab));
-                        float32x4_t _wb = bfloat2float(vget_high_u16(_wab));
-                        float32x4_t _wc = bfloat2float(vget_low_u16(_wcd));
-                        float32x4_t _wd = bfloat2float(vget_high_u16(_wcd));
-                        float32x4_t _we = bfloat2float(vget_low_u16(_wef));
-                        float32x4_t _wf = bfloat2float(vget_high_u16(_wef));
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 0);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 1);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 1);
-                        _sum4 = vfmaq_laneq_f32(_sum4, _w4, _r0, 2);
-                        _sum5 = vfmaq_laneq_f32(_sum5, _w5, _r0, 2);
-                        _sum6 = vfmaq_laneq_f32(_sum6, _w6, _r0, 3);
-                        _sum7 = vfmaq_laneq_f32(_sum7, _w7, _r0, 3);
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w8, _r1, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w9, _r1, 0);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _wa, _r1, 1);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _wb, _r1, 1);
-                        _sum4 = vfmaq_laneq_f32(_sum4, _wc, _r1, 2);
-                        _sum5 = vfmaq_laneq_f32(_sum5, _wd, _r1, 2);
-                        _sum6 = vfmaq_laneq_f32(_sum6, _we, _r1, 3);
-                        _sum7 = vfmaq_laneq_f32(_sum7, _wf, _r1, 3);
-
-                        kptr += 64;
-                    }
-                }
-                for (; q + 3 < inch; q += 4)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q / elempack).row<const unsigned short>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = bfloat2float(vld1_u16(r0 + sok));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            uint16x4_t _r_u16 = uint16x4_t();
-                            _r_u16 = vset_lane_u16(r0[sok], _r_u16, 0);
-                            _r_u16 = vset_lane_u16(r0[sok + N], _r_u16, 1);
-                            _r_u16 = vset_lane_u16(r0[sok + N * 2], _r_u16, 2);
-                            _r_u16 = vset_lane_u16(r0[sok + N * 3], _r_u16, 3);
-                            _r0 = bfloat2float(_r_u16);
-                        }
-
-                        uint16x8_t _w01 = vld1q_u16(kptr);
-                        uint16x8_t _w23 = vld1q_u16(kptr + 8);
-                        uint16x8_t _w45 = vld1q_u16(kptr + 16);
-                        uint16x8_t _w67 = vld1q_u16(kptr + 24);
-                        float32x4_t _w0 = bfloat2float(vget_low_u16(_w01));
-                        float32x4_t _w1 = bfloat2float(vget_high_u16(_w01));
-                        float32x4_t _w2 = bfloat2float(vget_low_u16(_w23));
-                        float32x4_t _w3 = bfloat2float(vget_high_u16(_w23));
-                        float32x4_t _w4 = bfloat2float(vget_low_u16(_w45));
-                        float32x4_t _w5 = bfloat2float(vget_high_u16(_w45));
-                        float32x4_t _w6 = bfloat2float(vget_low_u16(_w67));
-                        float32x4_t _w7 = bfloat2float(vget_high_u16(_w67));
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 0);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 1);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 1);
-                        _sum4 = vfmaq_laneq_f32(_sum4, _w4, _r0, 2);
-                        _sum5 = vfmaq_laneq_f32(_sum5, _w5, _r0, 2);
-                        _sum6 = vfmaq_laneq_f32(_sum6, _w6, _r0, 3);
-                        _sum7 = vfmaq_laneq_f32(_sum7, _w7, _r0, 3);
-
-                        kptr += 32;
-                    }
-                }
-                for (; q + 1 < inch; q += 2)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q).row<const unsigned short>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float val0;
-                        float val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = bfloat16_to_float32(r0[sok]);
-                            val1 = bfloat16_to_float32(r0[sok + N]);
-                        }
-
-                        uint16x8_t _w01 = vld1q_u16(kptr);
-                        uint16x8_t _w23 = vld1q_u16(kptr + 8);
-                        float32x4_t _w0 = bfloat2float(vget_low_u16(_w01));
-                        float32x4_t _w1 = bfloat2float(vget_high_u16(_w01));
-                        float32x4_t _w2 = bfloat2float(vget_low_u16(_w23));
-                        float32x4_t _w3 = bfloat2float(vget_high_u16(_w23));
-                        _sum0 = vfmaq_n_f32(_sum0, _w0, val0);
-                        _sum1 = vfmaq_n_f32(_sum1, _w1, val0);
-                        _sum2 = vfmaq_n_f32(_sum2, _w2, val1);
-                        _sum3 = vfmaq_n_f32(_sum3, _w3, val1);
-
-                        kptr += 16;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q).row<const unsigned short>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float32x4_t _val;
-                        // if (elempack == 1)
-                        {
-                            _val = bfloat2float(vdup_n_u16(r0[space_ofs[k]]));
-                        }
-
-                        uint16x8_t _w = vld1q_u16(kptr);
-                        float32x4_t _w0 = bfloat2float(vget_low_u16(_w));
-                        float32x4_t _w1 = bfloat2float(vget_high_u16(_w));
-                        _sum0 = vfmaq_f32(_sum0, _w0, _val);
-                        _sum1 = vfmaq_f32(_sum1, _w1, _val);
-
-                        kptr += 8;
-                    }
-                }
-
-                _sum0 = vaddq_f32(_sum0, _sum2);
-                _sum1 = vaddq_f32(_sum1, _sum3);
-                _sum4 = vaddq_f32(_sum4, _sum6);
-                _sum5 = vaddq_f32(_sum5, _sum7);
-                _sum0 = vaddq_f32(_sum0, _sum4);
-                _sum1 = vaddq_f32(_sum1, _sum5);
-
-                _sum0 = activation_ps(_sum0, activation_type, activation_params);
-                _sum1 = activation_ps(_sum1, activation_type, activation_params);
-
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr, float2bfloat(_sum0));
-                    vst1_u16(outptr + M, float2bfloat(_sum1));
-                    outptr += 4;
-                }
-                else // if (out_elempack == 1)
-                {
-                    uint16x4_t _sum0_u16 = float2bfloat(_sum0);
-                    uint16x4_t _sum1_u16 = float2bfloat(_sum1);
-                    outptr[0] = vget_lane_u16(_sum0_u16, 0);
-                    outptr[M] = vget_lane_u16(_sum0_u16, 1);
-                    outptr[M * 2] = vget_lane_u16(_sum0_u16, 2);
-                    outptr[M * 3] = vget_lane_u16(_sum0_u16, 3);
-                    outptr[M * 4] = vget_lane_u16(_sum1_u16, 0);
-                    outptr[M * 5] = vget_lane_u16(_sum1_u16, 1);
-                    outptr[M * 6] = vget_lane_u16(_sum1_u16, 2);
-                    outptr[M * 7] = vget_lane_u16(_sum1_u16, 3);
-                    outptr += 1;
-                }
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 8;
-    nn_outch = (outch - remain_outch_start) / 4;
-#else // __aarch64__
-    nn_outch = (outch - remain_outch_start) / 4;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __aarch64__
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 4;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int out_elempack = top_blob.elempack;
-
-        unsigned short* outptr = top_blob.channel(p / out_elempack);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                float32x4_t _sum2 = vdupq_n_f32(0.f);
-                float32x4_t _sum3 = vdupq_n_f32(0.f);
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = vld1q_f32(bias_data_ptr + p);
-                }
-
-#if __aarch64__
-                const unsigned short* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4);
-#else
-                const unsigned short* kptr = weight_data_tm.channel(p / 4);
-#endif
-
-                int q = 0;
-#if __aarch64__
-                for (; q + 7 < inch; q += 8)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q / elempack).row<const unsigned short>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        float32x4_t _r1;
-                        if (elempack == 4)
-                        {
-                            _r0 = bfloat2float(vld1_u16(r0 + sok));
-                            _r1 = bfloat2float(vld1_u16(r0 + sok + N));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            uint16x8_t _r_u16 = uint16x8_t();
-                            _r_u16 = vsetq_lane_u16(r0[sok], _r_u16, 0);
-                            _r_u16 = vsetq_lane_u16(r0[sok + N], _r_u16, 1);
-                            _r_u16 = vsetq_lane_u16(r0[sok + N * 2], _r_u16, 2);
-                            _r_u16 = vsetq_lane_u16(r0[sok + N * 3], _r_u16, 3);
-                            _r_u16 = vsetq_lane_u16(r0[sok + N * 4], _r_u16, 4);
-                            _r_u16 = vsetq_lane_u16(r0[sok + N * 5], _r_u16, 5);
-                            _r_u16 = vsetq_lane_u16(r0[sok + N * 6], _r_u16, 6);
-                            _r_u16 = vsetq_lane_u16(r0[sok + N * 7], _r_u16, 7);
-                            _r0 = bfloat2float(vget_low_u16(_r_u16));
-                            _r1 = bfloat2float(vget_high_u16(_r_u16));
-                        }
-
-                        uint16x8_t _w01 = vld1q_u16(kptr);
-                        uint16x8_t _w23 = vld1q_u16(kptr + 8);
-                        uint16x8_t _w45 = vld1q_u16(kptr + 16);
-                        uint16x8_t _w67 = vld1q_u16(kptr + 24);
-                        float32x4_t _w0 = bfloat2float(vget_low_u16(_w01));
-                        float32x4_t _w1 = bfloat2float(vget_high_u16(_w01));
-                        float32x4_t _w2 = bfloat2float(vget_low_u16(_w23));
-                        float32x4_t _w3 = bfloat2float(vget_high_u16(_w23));
-                        float32x4_t _w4 = bfloat2float(vget_low_u16(_w45));
-                        float32x4_t _w5 = bfloat2float(vget_high_u16(_w45));
-                        float32x4_t _w6 = bfloat2float(vget_low_u16(_w67));
-                        float32x4_t _w7 = bfloat2float(vget_high_u16(_w67));
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 1);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 2);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 3);
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w4, _r1, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w5, _r1, 1);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _w6, _r1, 2);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _w7, _r1, 3);
-
-                        kptr += 32;
-                    }
-                }
-#endif // __aarch64__
-                for (; q + 3 < inch; q += 4)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q / elempack).row<const unsigned short>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = bfloat2float(vld1_u16(r0 + sok));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            uint16x4_t _r_u16 = uint16x4_t();
-                            _r_u16 = vset_lane_u16(r0[sok], _r_u16, 0);
-                            _r_u16 = vset_lane_u16(r0[sok + N], _r_u16, 1);
-                            _r_u16 = vset_lane_u16(r0[sok + N * 2], _r_u16, 2);
-                            _r_u16 = vset_lane_u16(r0[sok + N * 3], _r_u16, 3);
-                            _r0 = bfloat2float(_r_u16);
-                        }
-
-                        uint16x8_t _w01 = vld1q_u16(kptr);
-                        uint16x8_t _w23 = vld1q_u16(kptr + 8);
-                        float32x4_t _w0 = bfloat2float(vget_low_u16(_w01));
-                        float32x4_t _w1 = bfloat2float(vget_high_u16(_w01));
-                        float32x4_t _w2 = bfloat2float(vget_low_u16(_w23));
-                        float32x4_t _w3 = bfloat2float(vget_high_u16(_w23));
-#if __aarch64__
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 1);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 2);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 3);
-#else
-                        _sum0 = vmlaq_lane_f32(_sum0, _w0, vget_low_f32(_r0), 0);
-                        _sum1 = vmlaq_lane_f32(_sum1, _w1, vget_low_f32(_r0), 1);
-                        _sum2 = vmlaq_lane_f32(_sum2, _w2, vget_high_f32(_r0), 0);
-                        _sum3 = vmlaq_lane_f32(_sum3, _w3, vget_high_f32(_r0), 1);
-#endif
-
-                        kptr += 16;
-                    }
-                }
-                for (; q + 1 < inch; q += 2)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q).row<const unsigned short>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float val0;
-                        float val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = bfloat16_to_float32(r0[sok]);
-                            val1 = bfloat16_to_float32(r0[sok + N]);
-                        }
-
-                        uint16x8_t _w = vld1q_u16(kptr);
-                        float32x4_t _w0 = bfloat2float(vget_low_u16(_w));
-                        float32x4_t _w1 = bfloat2float(vget_high_u16(_w));
-#if __aarch64__
-                        _sum0 = vfmaq_n_f32(_sum0, _w0, val0);
-                        _sum1 = vfmaq_n_f32(_sum1, _w1, val1);
-#else
-                        _sum0 = vmlaq_n_f32(_sum0, _w0, val0);
-                        _sum1 = vmlaq_n_f32(_sum1, _w1, val1);
-#endif
-
-                        kptr += 8;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q).row<const unsigned short>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float32x4_t _val;
-                        // if (elempack == 1)
-                        {
-                            _val = bfloat2float(vdup_n_u16(r0[space_ofs[k]]));
-                        }
-
-                        float32x4_t _w = bfloat2float(vld1_u16(kptr));
-#if __aarch64__
-                        _sum0 = vfmaq_f32(_sum0, _val, _w);
-#else
-                        _sum0 = vmlaq_f32(_sum0, _val, _w);
-#endif
-
-                        kptr += 4;
-                    }
-                }
-
-                _sum0 = vaddq_f32(_sum0, _sum1);
-                _sum2 = vaddq_f32(_sum2, _sum3);
-                _sum0 = vaddq_f32(_sum0, _sum2);
-
-                _sum0 = activation_ps(_sum0, activation_type, activation_params);
-
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr, float2bfloat(_sum0));
-                    outptr += 4;
-                }
-                else // if (out_elempack == 1)
-                {
-                    uint16x4_t _sum0_u16 = float2bfloat(_sum0);
-                    outptr[0] = vget_lane_u16(_sum0_u16, 0);
-                    outptr[M] = vget_lane_u16(_sum0_u16, 1);
-                    outptr[M * 2] = vget_lane_u16(_sum0_u16, 2);
-                    outptr[M * 3] = vget_lane_u16(_sum0_u16, 3);
-                    outptr += 1;
-                }
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 4;
-    nn_outch = (outch - remain_outch_start) / 2;
-#else // __ARM_NEON
-    nn_outch = (outch - remain_outch_start) / 2;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __ARM_NEON
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 2;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        unsigned short* outptr0 = top_blob.channel(p);
-        unsigned short* outptr1 = top_blob.channel(p + 1);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum0 = bias_data_ptr[p];
-                    sum1 = bias_data_ptr[p + 1];
-                }
-
-#if __aarch64__
-                const unsigned short* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2);
-#elif __ARM_NEON
-                const unsigned short* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2);
-#else
-                const unsigned short* kptr = weight_data_tm.channel(p / 2);
-#endif
-
-                int q = 0;
-#if __ARM_NEON
-#if __aarch64__
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                float32x4_t _sum2 = vdupq_n_f32(0.f);
-                float32x4_t _sum3 = vdupq_n_f32(0.f);
-                for (; q + 7 < inch; q += 8)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q / elempack).row<const unsigned short>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        float32x4_t _r1;
-                        if (elempack == 4)
-                        {
-                            _r0 = bfloat2float(vld1_u16(r0 + sok));
-                            _r1 = bfloat2float(vld1_u16(r0 + sok + N));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            uint16x8_t _r01_u16 = uint16x8_t();
-                            _r01_u16 = vsetq_lane_u16(r0[sok], _r01_u16, 0);
-                            _r01_u16 = vsetq_lane_u16(r0[sok + N], _r01_u16, 1);
-                            _r01_u16 = vsetq_lane_u16(r0[sok + N * 2], _r01_u16, 2);
-                            _r01_u16 = vsetq_lane_u16(r0[sok + N * 3], _r01_u16, 3);
-                            _r01_u16 = vsetq_lane_u16(r0[sok + N * 4], _r01_u16, 4);
-                            _r01_u16 = vsetq_lane_u16(r0[sok + N * 5], _r01_u16, 5);
-                            _r01_u16 = vsetq_lane_u16(r0[sok + N * 6], _r01_u16, 6);
-                            _r01_u16 = vsetq_lane_u16(r0[sok + N * 7], _r01_u16, 7);
-                            _r0 = bfloat2float(vget_low_u16(_r01_u16));
-                            _r1 = bfloat2float(vget_high_u16(_r01_u16));
-                        }
-
-                        uint16x8_t _w01 = vld1q_u16(kptr);
-                        uint16x8_t _w23 = vld1q_u16(kptr + 8);
-                        float32x4_t _w0 = bfloat2float(vget_low_u16(_w01));
-                        float32x4_t _w1 = bfloat2float(vget_high_u16(_w01));
-                        float32x4_t _w2 = bfloat2float(vget_low_u16(_w23));
-                        float32x4_t _w3 = bfloat2float(vget_high_u16(_w23));
-                        _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                        _sum1 = vfmaq_f32(_sum1, _r1, _w1);
-                        _sum2 = vfmaq_f32(_sum2, _r0, _w2);
-                        _sum3 = vfmaq_f32(_sum3, _r1, _w3);
-
-                        kptr += 16;
-                    }
-                }
-                _sum0 = vaddq_f32(_sum0, _sum1);
-                _sum2 = vaddq_f32(_sum2, _sum3);
-                sum0 += vaddvq_f32(_sum0);
-                sum1 += vaddvq_f32(_sum2);
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-#else  // __aarch64__
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-#endif // __aarch64__
-                for (; q + 3 < inch; q += 4)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q / elempack).row<const unsigned short>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = bfloat2float(vld1_u16(r0 + sok));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            uint16x4_t _r0_u16 = uint16x4_t();
-                            _r0_u16 = vset_lane_u16(r0[sok], _r0_u16, 0);
-                            _r0_u16 = vset_lane_u16(r0[sok + N], _r0_u16, 1);
-                            _r0_u16 = vset_lane_u16(r0[sok + N * 2], _r0_u16, 2);
-                            _r0_u16 = vset_lane_u16(r0[sok + N * 3], _r0_u16, 3);
-                            _r0 = bfloat2float(_r0_u16);
-                        }
-
-                        uint16x8_t _w = vld1q_u16(kptr);
-                        float32x4_t _w0 = bfloat2float(vget_low_u16(_w));
-                        float32x4_t _w1 = bfloat2float(vget_high_u16(_w));
-#if __aarch64__
-                        _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                        _sum1 = vfmaq_f32(_sum1, _r0, _w1);
-#else
-                        _sum0 = vmlaq_f32(_sum0, _r0, _w0);
-                        _sum1 = vmlaq_f32(_sum1, _r0, _w1);
-#endif
-
-                        kptr += 8;
-                    }
-                }
-#if __aarch64__
-                sum0 += vaddvq_f32(_sum0);
-                sum1 += vaddvq_f32(_sum1);
-#else
-                float32x2_t _ss0 = vadd_f32(vget_low_f32(_sum0), vget_high_f32(_sum0));
-                float32x2_t _ss1 = vadd_f32(vget_low_f32(_sum1), vget_high_f32(_sum1));
-                float32x2_t _ss = vpadd_f32(_ss0, _ss1);
-                sum0 += vget_lane_f32(_ss, 0);
-                sum1 += vget_lane_f32(_ss, 1);
-#endif
-#endif // __ARM_NEON
-                for (; q + 1 < inch; q += 2)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q).row<const unsigned short>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float val0;
-                        float val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = bfloat16_to_float32(r0[sok]);
-                            val1 = bfloat16_to_float32(r0[sok + N]);
-                        }
-
-                        sum0 += val0 * bfloat16_to_float32(kptr[0]);
-                        sum1 += val0 * bfloat16_to_float32(kptr[1]);
-                        sum0 += val1 * bfloat16_to_float32(kptr[2]);
-                        sum1 += val1 * bfloat16_to_float32(kptr[3]);
-
-                        kptr += 4;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q).row<const unsigned short>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float val;
-                        // if (elempack == 1)
-                        {
-                            val = bfloat16_to_float32(r0[space_ofs[k]]);
-                        }
-
-                        sum0 += val * bfloat16_to_float32(kptr[0]);
-                        sum1 += val * bfloat16_to_float32(kptr[1]);
-
-                        kptr += 2;
-                    }
-                }
-
-                sum0 = activation_ss(sum0, activation_type, activation_params);
-                sum1 = activation_ss(sum1, activation_type, activation_params);
-
-                outptr0[0] = float32_to_bfloat16(sum0);
-                outptr1[0] = float32_to_bfloat16(sum1);
-                outptr0 += 1;
-                outptr1 += 1;
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 2;
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        unsigned short* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-#if __aarch64__
-                const unsigned short* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#elif __ARM_NEON
-                const unsigned short* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2 + p % 2);
-#else
-                const unsigned short* kptr = weight_data_tm.channel(p / 2 + p % 2);
-#endif
-
-                int q = 0;
-#if __ARM_NEON
-#if __aarch64__
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                for (; q + 7 < inch; q += 8)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q / elempack).row<const unsigned short>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        float32x4_t _r1;
-                        if (elempack == 4)
-                        {
-                            _r0 = bfloat2float(vld1_u16(r0 + sok));
-                            _r1 = bfloat2float(vld1_u16(r0 + sok + N));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            uint16x8_t _r01_u16 = uint16x8_t();
-                            _r01_u16 = vsetq_lane_u16(r0[sok], _r01_u16, 0);
-                            _r01_u16 = vsetq_lane_u16(r0[sok + N], _r01_u16, 1);
-                            _r01_u16 = vsetq_lane_u16(r0[sok + N * 2], _r01_u16, 2);
-                            _r01_u16 = vsetq_lane_u16(r0[sok + N * 3], _r01_u16, 3);
-                            _r01_u16 = vsetq_lane_u16(r0[sok + N * 4], _r01_u16, 4);
-                            _r01_u16 = vsetq_lane_u16(r0[sok + N * 5], _r01_u16, 5);
-                            _r01_u16 = vsetq_lane_u16(r0[sok + N * 6], _r01_u16, 6);
-                            _r01_u16 = vsetq_lane_u16(r0[sok + N * 7], _r01_u16, 7);
-                            _r0 = bfloat2float(vget_low_u16(_r01_u16));
-                            _r1 = bfloat2float(vget_high_u16(_r01_u16));
-                        }
-
-                        uint16x8_t _w = vld1q_u16(kptr);
-                        float32x4_t _w0 = bfloat2float(vget_low_u16(_w));
-                        float32x4_t _w1 = bfloat2float(vget_high_u16(_w));
-                        _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                        _sum1 = vfmaq_f32(_sum1, _r1, _w1);
-
-                        kptr += 8;
-                    }
-                }
-                _sum0 = vaddq_f32(_sum0, _sum1);
-                sum += vaddvq_f32(_sum0);
-#endif // __aarch64__
-                float32x4_t _sum = vdupq_n_f32(0.f);
-                for (; q + 3 < inch; q += 4)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q / elempack).row<const unsigned short>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = bfloat2float(vld1_u16(r0 + sok));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            uint16x4_t _r0_u16 = uint16x4_t();
-                            _r0_u16 = vset_lane_u16(r0[sok], _r0_u16, 0);
-                            _r0_u16 = vset_lane_u16(r0[sok + N], _r0_u16, 1);
-                            _r0_u16 = vset_lane_u16(r0[sok + N * 2], _r0_u16, 2);
-                            _r0_u16 = vset_lane_u16(r0[sok + N * 3], _r0_u16, 3);
-                            _r0 = bfloat2float(_r0_u16);
-                        }
-
-                        float32x4_t _w = bfloat2float(vld1_u16(kptr));
-#if __aarch64__
-                        _sum = vfmaq_f32(_sum, _r0, _w);
-#else
-                        _sum = vmlaq_f32(_sum, _r0, _w);
-#endif
-
-                        kptr += 4;
-                    }
-                }
-#if __aarch64__
-                sum += vaddvq_f32(_sum);
-#else
-                float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-                _ss = vpadd_f32(_ss, _ss);
-                sum += vget_lane_f32(_ss, 0);
-#endif
-#endif // __ARM_NEON
-                for (; q + 1 < inch; q += 2)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q).row<const unsigned short>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float val0;
-                        float val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = bfloat16_to_float32(r0[sok]);
-                            val1 = bfloat16_to_float32(r0[sok + N]);
-                        }
-
-                        sum += val0 * bfloat16_to_float32(kptr[0]);
-                        sum += val1 * bfloat16_to_float32(kptr[1]);
-
-                        kptr += 2;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const unsigned short* r0 = bottom_blob.channel(q).row<const unsigned short>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float val;
-                        // if (elempack == 1)
-                        {
-                            val = bfloat16_to_float32(r0[space_ofs[k]]);
-                        }
-
-                        sum += val * bfloat16_to_float32(kptr[0]);
-
-                        kptr += 1;
-                    }
-                }
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[0] = float32_to_bfloat16(sum);
-                outptr += 1;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed_fp16s.h
deleted file mode 100644
index d565fb23..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed_fp16s.h
+++ /dev/null
@@ -1,1933 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_transform_kernel_packed_fp16s(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // src = kw-kh-inch-outch
-    // dst = pb-pa-kw-kh-inch/pa-outch/pb
-
-    // clang-format off
-    // *INDENT-OFF*
-    if (outch >= 8)
-    {
-        if (inch >= 8)
-            kernel_tm.create(8 * 8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-        else if (inch >= 4)
-            kernel_tm.create(8 * 4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-        else if (inch >= 2)
-            kernel_tm.create(8 * 2 * maxk, inch / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-        else
-            kernel_tm.create(8 * maxk, inch, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-    }
-    else if (outch >= 4)
-    {
-        if (inch >= 8)
-            kernel_tm.create(4 * 8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-        else if (inch >= 4)
-            kernel_tm.create(4 * 4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-        else if (inch >= 2)
-            kernel_tm.create(4 * 2 * maxk, inch / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-        else
-            kernel_tm.create(4 * maxk, inch, outch / 4 + (outch % 4) / 2 + outch % 2, (size_t)2u);
-    }
-    else if (outch >= 2)
-    {
-        if (inch >= 8)
-            kernel_tm.create(2 * 8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 2 + outch % 2, (size_t)2u);
-        else if (inch >= 4)
-            kernel_tm.create(2 * 4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch / 2 + outch % 2, (size_t)2u);
-        else if (inch >= 2)
-            kernel_tm.create(2 * 2 * maxk, inch / 2 + inch % 2, outch / 2 + outch % 2, (size_t)2u);
-        else
-            kernel_tm.create(2 * maxk, inch, outch / 2 + outch % 2, (size_t)2u);
-    }
-    else
-    {
-        if (inch >= 8)
-            kernel_tm.create(8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch, (size_t)2u);
-        else if (inch >= 4)
-            kernel_tm.create(4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch, (size_t)2u);
-        else if (inch >= 2)
-            kernel_tm.create(2 * maxk, inch / 2 + inch % 2, outch, (size_t)2u);
-        else
-            kernel_tm.create(maxk, inch, outch, (size_t)2u);
-    }
-    // *INDENT-ON*
-    // clang-format on
-
-    int q = 0;
-    for (; q + 7 < outch; q += 8)
-    {
-        const float* kptr0 = (const float*)kernel + q * inch * maxk;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inch * maxk;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inch * maxk;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inch * maxk;
-        const float* kptr4 = (const float*)kernel + (q + 4) * inch * maxk;
-        const float* kptr5 = (const float*)kernel + (q + 5) * inch * maxk;
-        const float* kptr6 = (const float*)kernel + (q + 6) * inch * maxk;
-        const float* kptr7 = (const float*)kernel + (q + 7) * inch * maxk;
-
-        __fp16* g00 = kernel_tm.channel(q / 8);
-
-        int p = 0;
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-                const float* k4 = kptr4 + p * maxk;
-                const float* k5 = kptr5 + p * maxk;
-                const float* k6 = kptr6 + p * maxk;
-                const float* k7 = kptr7 + p * maxk;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    g00[1] = (__fp16)k1[k];
-                    g00[2] = (__fp16)k2[k];
-                    g00[3] = (__fp16)k3[k];
-                    g00[4] = (__fp16)k4[k];
-                    g00[5] = (__fp16)k5[k];
-                    g00[6] = (__fp16)k6[k];
-                    g00[7] = (__fp16)k7[k];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    k4 += maxk;
-                    k5 += maxk;
-                    k6 += maxk;
-                    k7 += maxk;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-                const float* k4 = kptr4 + p * maxk;
-                const float* k5 = kptr5 + p * maxk;
-                const float* k6 = kptr6 + p * maxk;
-                const float* k7 = kptr7 + p * maxk;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    g00[1] = (__fp16)k1[k];
-                    g00[2] = (__fp16)k2[k];
-                    g00[3] = (__fp16)k3[k];
-                    g00[4] = (__fp16)k4[k];
-                    g00[5] = (__fp16)k5[k];
-                    g00[6] = (__fp16)k6[k];
-                    g00[7] = (__fp16)k7[k];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    k4 += maxk;
-                    k5 += maxk;
-                    k6 += maxk;
-                    k7 += maxk;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-                const float* k4 = kptr4 + p * maxk;
-                const float* k5 = kptr5 + p * maxk;
-                const float* k6 = kptr6 + p * maxk;
-                const float* k7 = kptr7 + p * maxk;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    g00[1] = (__fp16)k1[k];
-                    g00[2] = (__fp16)k2[k];
-                    g00[3] = (__fp16)k3[k];
-                    g00[4] = (__fp16)k4[k];
-                    g00[5] = (__fp16)k5[k];
-                    g00[6] = (__fp16)k6[k];
-                    g00[7] = (__fp16)k7[k];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    k4 += maxk;
-                    k5 += maxk;
-                    k6 += maxk;
-                    k7 += maxk;
-                    g00 += 8;
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            const float* k0 = kptr0 + p * maxk;
-            const float* k1 = kptr1 + p * maxk;
-            const float* k2 = kptr2 + p * maxk;
-            const float* k3 = kptr3 + p * maxk;
-            const float* k4 = kptr4 + p * maxk;
-            const float* k5 = kptr5 + p * maxk;
-            const float* k6 = kptr6 + p * maxk;
-            const float* k7 = kptr7 + p * maxk;
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = (__fp16)k0[k];
-                g00[1] = (__fp16)k1[k];
-                g00[2] = (__fp16)k2[k];
-                g00[3] = (__fp16)k3[k];
-                g00[4] = (__fp16)k4[k];
-                g00[5] = (__fp16)k5[k];
-                g00[6] = (__fp16)k6[k];
-                g00[7] = (__fp16)k7[k];
-                g00 += 8;
-            }
-        }
-    }
-    for (; q + 3 < outch; q += 4)
-    {
-        const float* kptr0 = (const float*)kernel + q * inch * maxk;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inch * maxk;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inch * maxk;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inch * maxk;
-
-        __fp16* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4);
-
-        int p = 0;
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    g00[1] = (__fp16)k1[k];
-                    g00[2] = (__fp16)k2[k];
-                    g00[3] = (__fp16)k3[k];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    g00 += 4;
-                }
-            }
-        }
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    g00[1] = (__fp16)k1[k];
-                    g00[2] = (__fp16)k2[k];
-                    g00[3] = (__fp16)k3[k];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    g00 += 4;
-                }
-            }
-        }
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-                const float* k2 = kptr2 + p * maxk;
-                const float* k3 = kptr3 + p * maxk;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    g00[1] = (__fp16)k1[k];
-                    g00[2] = (__fp16)k2[k];
-                    g00[3] = (__fp16)k3[k];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    g00 += 4;
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            const float* k0 = kptr0 + p * maxk;
-            const float* k1 = kptr1 + p * maxk;
-            const float* k2 = kptr2 + p * maxk;
-            const float* k3 = kptr3 + p * maxk;
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = (__fp16)k0[k];
-                g00[1] = (__fp16)k1[k];
-                g00[2] = (__fp16)k2[k];
-                g00[3] = (__fp16)k3[k];
-                g00 += 4;
-            }
-        }
-    }
-    for (; q + 1 < outch; q += 2)
-    {
-        const float* kptr0 = (const float*)kernel + q * inch * maxk;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inch * maxk;
-
-        __fp16* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2);
-
-        int p = 0;
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk + k;
-                const float* k1 = kptr1 + p * maxk + k;
-
-                g00[0] = (__fp16)k0[0];
-                g00[1] = (__fp16)k0[maxk];
-                g00[2] = (__fp16)k0[maxk * 2];
-                g00[3] = (__fp16)k0[maxk * 3];
-                g00[4] = (__fp16)k0[maxk * 4];
-                g00[5] = (__fp16)k0[maxk * 5];
-                g00[6] = (__fp16)k0[maxk * 6];
-                g00[7] = (__fp16)k0[maxk * 7];
-                g00[8] = (__fp16)k1[0];
-                g00[9] = (__fp16)k1[maxk];
-                g00[10] = (__fp16)k1[maxk * 2];
-                g00[11] = (__fp16)k1[maxk * 3];
-                g00[12] = (__fp16)k1[maxk * 4];
-                g00[13] = (__fp16)k1[maxk * 5];
-                g00[14] = (__fp16)k1[maxk * 6];
-                g00[15] = (__fp16)k1[maxk * 7];
-                g00 += 16;
-            }
-        }
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk + k;
-                const float* k1 = kptr1 + p * maxk + k;
-
-                g00[0] = (__fp16)k0[0];
-                g00[1] = (__fp16)k0[maxk];
-                g00[2] = (__fp16)k0[maxk * 2];
-                g00[3] = (__fp16)k0[maxk * 3];
-                g00[4] = (__fp16)k1[0];
-                g00[5] = (__fp16)k1[maxk];
-                g00[6] = (__fp16)k1[maxk * 2];
-                g00[7] = (__fp16)k1[maxk * 3];
-                g00 += 8;
-            }
-        }
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + p * maxk;
-                const float* k1 = kptr1 + p * maxk;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    g00[1] = (__fp16)k1[k];
-                    k0 += maxk;
-                    k1 += maxk;
-                    g00 += 2;
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            const float* k0 = kptr0 + p * maxk;
-            const float* k1 = kptr1 + p * maxk;
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = (__fp16)k0[k];
-                g00[1] = (__fp16)k1[k];
-                g00 += 2;
-            }
-        }
-    }
-    for (; q < outch; q++)
-    {
-        const float* kptr = (const float*)kernel + q * inch * maxk;
-
-        __fp16* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2 + q % 2);
-
-        int p = 0;
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr + p * maxk;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    k0 += maxk;
-                    g00 += 1;
-                }
-            }
-        }
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr + p * maxk;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    k0 += maxk;
-                    g00 += 1;
-                }
-            }
-        }
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr + p * maxk;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = (__fp16)k0[k];
-                    k0 += maxk;
-                    g00 += 1;
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            const float* k0 = kptr + p * maxk;
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = (__fp16)k0[k];
-                g00++;
-            }
-        }
-    }
-}
-
-static void convolution_packed_fp16s(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int elempack = bottom_blob.elempack;
-    const int inch = bottom_blob.c * elempack;
-
-    const int N = bottom_blob.cstep * elempack;
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int outch = top_blob.c * out_elempack;
-
-    const int M = top_blob.cstep * out_elempack;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2 * elempack;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    int nn_outch = 0;
-    int remain_outch_start = 0;
-    nn_outch = (outch - remain_outch_start) / 8;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 8;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int out_elempack = top_blob.elempack;
-
-        __fp16* outptr = top_blob.channel(p / out_elempack);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                float32x4_t _sum2 = vdupq_n_f32(0.f);
-                float32x4_t _sum3 = vdupq_n_f32(0.f);
-                float32x4_t _sum4 = vdupq_n_f32(0.f);
-                float32x4_t _sum5 = vdupq_n_f32(0.f);
-                float32x4_t _sum6 = vdupq_n_f32(0.f);
-                float32x4_t _sum7 = vdupq_n_f32(0.f);
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = vld1q_f32(bias_data_ptr + p);
-                    _sum1 = vld1q_f32(bias_data_ptr + p + 4);
-                }
-
-                const __fp16* kptr = weight_data_tm.channel(p / 8);
-
-                int q = 0;
-                for (; q + 7 < inch; q += 8)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        float32x4_t _r1;
-                        if (elempack == 4)
-                        {
-                            _r0 = vcvt_f32_f16(vld1_f16(r0 + sok));
-                            _r1 = vcvt_f32_f16(vld1_f16(r0 + sok + N));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            float16x8_t _r_f16 = float16x8_t();
-                            _r_f16 = vsetq_lane_f16(r0[sok], _r_f16, 0);
-                            _r_f16 = vsetq_lane_f16(r0[sok + N], _r_f16, 1);
-                            _r_f16 = vsetq_lane_f16(r0[sok + N * 2], _r_f16, 2);
-                            _r_f16 = vsetq_lane_f16(r0[sok + N * 3], _r_f16, 3);
-                            _r_f16 = vsetq_lane_f16(r0[sok + N * 4], _r_f16, 4);
-                            _r_f16 = vsetq_lane_f16(r0[sok + N * 5], _r_f16, 5);
-                            _r_f16 = vsetq_lane_f16(r0[sok + N * 6], _r_f16, 6);
-                            _r_f16 = vsetq_lane_f16(r0[sok + N * 7], _r_f16, 7);
-                            _r0 = vcvt_f32_f16(vget_low_f16(_r_f16));
-                            _r1 = vcvt_f32_f16(vget_high_f16(_r_f16));
-                        }
-
-                        float16x8_t _w01 = vld1q_f16(kptr);
-                        float16x8_t _w23 = vld1q_f16(kptr + 8);
-                        float16x8_t _w45 = vld1q_f16(kptr + 16);
-                        float16x8_t _w67 = vld1q_f16(kptr + 24);
-                        float16x8_t _w89 = vld1q_f16(kptr + 32);
-                        float16x8_t _wab = vld1q_f16(kptr + 40);
-                        float16x8_t _wcd = vld1q_f16(kptr + 48);
-                        float16x8_t _wef = vld1q_f16(kptr + 56);
-                        float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01));
-                        float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01));
-                        float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23));
-                        float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23));
-                        float32x4_t _w4 = vcvt_f32_f16(vget_low_f16(_w45));
-                        float32x4_t _w5 = vcvt_f32_f16(vget_high_f16(_w45));
-                        float32x4_t _w6 = vcvt_f32_f16(vget_low_f16(_w67));
-                        float32x4_t _w7 = vcvt_f32_f16(vget_high_f16(_w67));
-                        float32x4_t _w8 = vcvt_f32_f16(vget_low_f16(_w89));
-                        float32x4_t _w9 = vcvt_f32_f16(vget_high_f16(_w89));
-                        float32x4_t _wa = vcvt_f32_f16(vget_low_f16(_wab));
-                        float32x4_t _wb = vcvt_f32_f16(vget_high_f16(_wab));
-                        float32x4_t _wc = vcvt_f32_f16(vget_low_f16(_wcd));
-                        float32x4_t _wd = vcvt_f32_f16(vget_high_f16(_wcd));
-                        float32x4_t _we = vcvt_f32_f16(vget_low_f16(_wef));
-                        float32x4_t _wf = vcvt_f32_f16(vget_high_f16(_wef));
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 0);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 1);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 1);
-                        _sum4 = vfmaq_laneq_f32(_sum4, _w4, _r0, 2);
-                        _sum5 = vfmaq_laneq_f32(_sum5, _w5, _r0, 2);
-                        _sum6 = vfmaq_laneq_f32(_sum6, _w6, _r0, 3);
-                        _sum7 = vfmaq_laneq_f32(_sum7, _w7, _r0, 3);
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w8, _r1, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w9, _r1, 0);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _wa, _r1, 1);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _wb, _r1, 1);
-                        _sum4 = vfmaq_laneq_f32(_sum4, _wc, _r1, 2);
-                        _sum5 = vfmaq_laneq_f32(_sum5, _wd, _r1, 2);
-                        _sum6 = vfmaq_laneq_f32(_sum6, _we, _r1, 3);
-                        _sum7 = vfmaq_laneq_f32(_sum7, _wf, _r1, 3);
-
-                        kptr += 64;
-                    }
-                }
-                for (; q + 3 < inch; q += 4)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = vcvt_f32_f16(vld1_f16(r0 + sok));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            float16x4_t _r_f16 = float16x4_t();
-                            _r_f16 = vset_lane_f16(r0[sok], _r_f16, 0);
-                            _r_f16 = vset_lane_f16(r0[sok + N], _r_f16, 1);
-                            _r_f16 = vset_lane_f16(r0[sok + N * 2], _r_f16, 2);
-                            _r_f16 = vset_lane_f16(r0[sok + N * 3], _r_f16, 3);
-                            _r0 = vcvt_f32_f16(_r_f16);
-                        }
-
-                        float16x8_t _w01 = vld1q_f16(kptr);
-                        float16x8_t _w23 = vld1q_f16(kptr + 8);
-                        float16x8_t _w45 = vld1q_f16(kptr + 16);
-                        float16x8_t _w67 = vld1q_f16(kptr + 24);
-                        float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01));
-                        float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01));
-                        float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23));
-                        float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23));
-                        float32x4_t _w4 = vcvt_f32_f16(vget_low_f16(_w45));
-                        float32x4_t _w5 = vcvt_f32_f16(vget_high_f16(_w45));
-                        float32x4_t _w6 = vcvt_f32_f16(vget_low_f16(_w67));
-                        float32x4_t _w7 = vcvt_f32_f16(vget_high_f16(_w67));
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 0);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 1);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 1);
-                        _sum4 = vfmaq_laneq_f32(_sum4, _w4, _r0, 2);
-                        _sum5 = vfmaq_laneq_f32(_sum5, _w5, _r0, 2);
-                        _sum6 = vfmaq_laneq_f32(_sum6, _w6, _r0, 3);
-                        _sum7 = vfmaq_laneq_f32(_sum7, _w7, _r0, 3);
-
-                        kptr += 32;
-                    }
-                }
-                for (; q + 1 < inch; q += 2)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float val0;
-                        float val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = (float)(r0[sok]);
-                            val1 = (float)(r0[sok + N]);
-                        }
-
-                        float16x8_t _w01 = vld1q_f16(kptr);
-                        float16x8_t _w23 = vld1q_f16(kptr + 8);
-                        float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01));
-                        float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01));
-                        float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23));
-                        float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23));
-                        _sum0 = vfmaq_n_f32(_sum0, _w0, val0);
-                        _sum1 = vfmaq_n_f32(_sum1, _w1, val0);
-                        _sum2 = vfmaq_n_f32(_sum2, _w2, val1);
-                        _sum3 = vfmaq_n_f32(_sum3, _w3, val1);
-
-                        kptr += 16;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float32x4_t _val;
-                        // if (elempack == 1)
-                        {
-                            _val = vcvt_f32_f16(vdup_n_f16(r0[space_ofs[k]]));
-                        }
-
-                        float16x8_t _w = vld1q_f16(kptr);
-                        float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w));
-                        float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w));
-                        _sum0 = vfmaq_f32(_sum0, _w0, _val);
-                        _sum1 = vfmaq_f32(_sum1, _w1, _val);
-
-                        kptr += 8;
-                    }
-                }
-
-                _sum0 = vaddq_f32(_sum0, _sum2);
-                _sum1 = vaddq_f32(_sum1, _sum3);
-                _sum4 = vaddq_f32(_sum4, _sum6);
-                _sum5 = vaddq_f32(_sum5, _sum7);
-                _sum0 = vaddq_f32(_sum0, _sum4);
-                _sum1 = vaddq_f32(_sum1, _sum5);
-
-                _sum0 = activation_ps(_sum0, activation_type, activation_params);
-                _sum1 = activation_ps(_sum1, activation_type, activation_params);
-
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr, vcvt_f16_f32(_sum0));
-                    vst1_f16(outptr + M, vcvt_f16_f32(_sum1));
-                    outptr += 4;
-                }
-                else // if (out_elempack == 1)
-                {
-                    float16x4_t _sum0_f16 = vcvt_f16_f32(_sum0);
-                    float16x4_t _sum1_f16 = vcvt_f16_f32(_sum1);
-                    outptr[0] = vget_lane_f16(_sum0_f16, 0);
-                    outptr[M] = vget_lane_f16(_sum0_f16, 1);
-                    outptr[M * 2] = vget_lane_f16(_sum0_f16, 2);
-                    outptr[M * 3] = vget_lane_f16(_sum0_f16, 3);
-                    outptr[M * 4] = vget_lane_f16(_sum1_f16, 0);
-                    outptr[M * 5] = vget_lane_f16(_sum1_f16, 1);
-                    outptr[M * 6] = vget_lane_f16(_sum1_f16, 2);
-                    outptr[M * 7] = vget_lane_f16(_sum1_f16, 3);
-                    outptr += 1;
-                }
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 8;
-    nn_outch = (outch - remain_outch_start) / 4;
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 4;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int out_elempack = top_blob.elempack;
-
-        __fp16* outptr = top_blob.channel(p / out_elempack);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                float32x4_t _sum2 = vdupq_n_f32(0.f);
-                float32x4_t _sum3 = vdupq_n_f32(0.f);
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = vld1q_f32(bias_data_ptr + p);
-                }
-
-                const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4);
-
-                int q = 0;
-                for (; q + 7 < inch; q += 8)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        float32x4_t _r1;
-                        if (elempack == 4)
-                        {
-                            _r0 = vcvt_f32_f16(vld1_f16(r0 + sok));
-                            _r1 = vcvt_f32_f16(vld1_f16(r0 + sok + N));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            float16x8_t _r_f16 = float16x8_t();
-                            _r_f16 = vsetq_lane_f16(r0[sok], _r_f16, 0);
-                            _r_f16 = vsetq_lane_f16(r0[sok + N], _r_f16, 1);
-                            _r_f16 = vsetq_lane_f16(r0[sok + N * 2], _r_f16, 2);
-                            _r_f16 = vsetq_lane_f16(r0[sok + N * 3], _r_f16, 3);
-                            _r_f16 = vsetq_lane_f16(r0[sok + N * 4], _r_f16, 4);
-                            _r_f16 = vsetq_lane_f16(r0[sok + N * 5], _r_f16, 5);
-                            _r_f16 = vsetq_lane_f16(r0[sok + N * 6], _r_f16, 6);
-                            _r_f16 = vsetq_lane_f16(r0[sok + N * 7], _r_f16, 7);
-                            _r0 = vcvt_f32_f16(vget_low_f16(_r_f16));
-                            _r1 = vcvt_f32_f16(vget_high_f16(_r_f16));
-                        }
-
-                        float16x8_t _w01 = vld1q_f16(kptr);
-                        float16x8_t _w23 = vld1q_f16(kptr + 8);
-                        float16x8_t _w45 = vld1q_f16(kptr + 16);
-                        float16x8_t _w67 = vld1q_f16(kptr + 24);
-                        float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01));
-                        float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01));
-                        float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23));
-                        float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23));
-                        float32x4_t _w4 = vcvt_f32_f16(vget_low_f16(_w45));
-                        float32x4_t _w5 = vcvt_f32_f16(vget_high_f16(_w45));
-                        float32x4_t _w6 = vcvt_f32_f16(vget_low_f16(_w67));
-                        float32x4_t _w7 = vcvt_f32_f16(vget_high_f16(_w67));
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 1);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 2);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 3);
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w4, _r1, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w5, _r1, 1);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _w6, _r1, 2);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _w7, _r1, 3);
-
-                        kptr += 32;
-                    }
-                }
-                for (; q + 3 < inch; q += 4)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = vcvt_f32_f16(vld1_f16(r0 + sok));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            float16x4_t _r_f16 = float16x4_t();
-                            _r_f16 = vset_lane_f16(r0[sok], _r_f16, 0);
-                            _r_f16 = vset_lane_f16(r0[sok + N], _r_f16, 1);
-                            _r_f16 = vset_lane_f16(r0[sok + N * 2], _r_f16, 2);
-                            _r_f16 = vset_lane_f16(r0[sok + N * 3], _r_f16, 3);
-                            _r0 = vcvt_f32_f16(_r_f16);
-                        }
-
-                        float16x8_t _w01 = vld1q_f16(kptr);
-                        float16x8_t _w23 = vld1q_f16(kptr + 8);
-                        float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01));
-                        float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01));
-                        float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23));
-                        float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23));
-                        _sum0 = vfmaq_laneq_f32(_sum0, _w0, _r0, 0);
-                        _sum1 = vfmaq_laneq_f32(_sum1, _w1, _r0, 1);
-                        _sum2 = vfmaq_laneq_f32(_sum2, _w2, _r0, 2);
-                        _sum3 = vfmaq_laneq_f32(_sum3, _w3, _r0, 3);
-
-                        kptr += 16;
-                    }
-                }
-                for (; q + 1 < inch; q += 2)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float val0;
-                        float val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = (float)(r0[sok]);
-                            val1 = (float)(r0[sok + N]);
-                        }
-
-                        float16x8_t _w = vld1q_f16(kptr);
-                        float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w));
-                        float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w));
-                        _sum0 = vfmaq_n_f32(_sum0, _w0, val0);
-                        _sum1 = vfmaq_n_f32(_sum1, _w1, val1);
-
-                        kptr += 8;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float32x4_t _val;
-                        // if (elempack == 1)
-                        {
-                            _val = vcvt_f32_f16(vdup_n_f16(r0[space_ofs[k]]));
-                        }
-
-                        float32x4_t _w = vcvt_f32_f16(vld1_f16(kptr));
-                        _sum0 = vfmaq_f32(_sum0, _val, _w);
-
-                        kptr += 4;
-                    }
-                }
-
-                _sum0 = vaddq_f32(_sum0, _sum1);
-                _sum2 = vaddq_f32(_sum2, _sum3);
-                _sum0 = vaddq_f32(_sum0, _sum2);
-
-                _sum0 = activation_ps(_sum0, activation_type, activation_params);
-
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr, vcvt_f16_f32(_sum0));
-                    outptr += 4;
-                }
-                else // if (out_elempack == 1)
-                {
-                    float16x4_t _sum0_f16 = vcvt_f16_f32(_sum0);
-                    outptr[0] = vget_lane_f16(_sum0_f16, 0);
-                    outptr[M] = vget_lane_f16(_sum0_f16, 1);
-                    outptr[M * 2] = vget_lane_f16(_sum0_f16, 2);
-                    outptr[M * 3] = vget_lane_f16(_sum0_f16, 3);
-                    outptr += 1;
-                }
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 4;
-    nn_outch = (outch - remain_outch_start) / 2;
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 2;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        __fp16* outptr0 = top_blob.channel(p);
-        __fp16* outptr1 = top_blob.channel(p + 1);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum0 = bias_data_ptr[p];
-                    sum1 = bias_data_ptr[p + 1];
-                }
-
-                const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2);
-
-                int q = 0;
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                float32x4_t _sum2 = vdupq_n_f32(0.f);
-                float32x4_t _sum3 = vdupq_n_f32(0.f);
-                for (; q + 7 < inch; q += 8)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        float32x4_t _r1;
-                        if (elempack == 4)
-                        {
-                            _r0 = vcvt_f32_f16(vld1_f16(r0 + sok));
-                            _r1 = vcvt_f32_f16(vld1_f16(r0 + sok + N));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            float16x8_t _r01_f16 = float16x8_t();
-                            _r01_f16 = vsetq_lane_f16(r0[sok], _r01_f16, 0);
-                            _r01_f16 = vsetq_lane_f16(r0[sok + N], _r01_f16, 1);
-                            _r01_f16 = vsetq_lane_f16(r0[sok + N * 2], _r01_f16, 2);
-                            _r01_f16 = vsetq_lane_f16(r0[sok + N * 3], _r01_f16, 3);
-                            _r01_f16 = vsetq_lane_f16(r0[sok + N * 4], _r01_f16, 4);
-                            _r01_f16 = vsetq_lane_f16(r0[sok + N * 5], _r01_f16, 5);
-                            _r01_f16 = vsetq_lane_f16(r0[sok + N * 6], _r01_f16, 6);
-                            _r01_f16 = vsetq_lane_f16(r0[sok + N * 7], _r01_f16, 7);
-                            _r0 = vcvt_f32_f16(vget_low_f16(_r01_f16));
-                            _r1 = vcvt_f32_f16(vget_high_f16(_r01_f16));
-                        }
-
-                        float16x8_t _w01 = vld1q_f16(kptr);
-                        float16x8_t _w23 = vld1q_f16(kptr + 8);
-                        float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01));
-                        float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01));
-                        float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23));
-                        float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23));
-                        _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                        _sum1 = vfmaq_f32(_sum1, _r1, _w1);
-                        _sum2 = vfmaq_f32(_sum2, _r0, _w2);
-                        _sum3 = vfmaq_f32(_sum3, _r1, _w3);
-
-                        kptr += 16;
-                    }
-                }
-                _sum0 = vaddq_f32(_sum0, _sum1);
-                _sum2 = vaddq_f32(_sum2, _sum3);
-                sum0 += vaddvq_f32(_sum0);
-                sum1 += vaddvq_f32(_sum2);
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-                for (; q + 3 < inch; q += 4)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = vcvt_f32_f16(vld1_f16(r0 + sok));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            float16x4_t _r0_f16 = float16x4_t();
-                            _r0_f16 = vset_lane_f16(r0[sok], _r0_f16, 0);
-                            _r0_f16 = vset_lane_f16(r0[sok + N], _r0_f16, 1);
-                            _r0_f16 = vset_lane_f16(r0[sok + N * 2], _r0_f16, 2);
-                            _r0_f16 = vset_lane_f16(r0[sok + N * 3], _r0_f16, 3);
-                            _r0 = vcvt_f32_f16(_r0_f16);
-                        }
-
-                        float16x8_t _w = vld1q_f16(kptr);
-                        float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w));
-                        float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w));
-                        _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                        _sum1 = vfmaq_f32(_sum1, _r0, _w1);
-
-                        kptr += 8;
-                    }
-                }
-                sum0 += vaddvq_f32(_sum0);
-                sum1 += vaddvq_f32(_sum1);
-                for (; q + 1 < inch; q += 2)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float val0;
-                        float val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = (float)(r0[sok]);
-                            val1 = (float)(r0[sok + N]);
-                        }
-
-                        sum0 += val0 * (float)(kptr[0]);
-                        sum1 += val0 * (float)(kptr[1]);
-                        sum0 += val1 * (float)(kptr[2]);
-                        sum1 += val1 * (float)(kptr[3]);
-
-                        kptr += 4;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float val;
-                        // if (elempack == 1)
-                        {
-                            val = (float)(r0[space_ofs[k]]);
-                        }
-
-                        sum0 += val * (float)(kptr[0]);
-                        sum1 += val * (float)(kptr[1]);
-
-                        kptr += 2;
-                    }
-                }
-
-                sum0 = activation_ss(sum0, activation_type, activation_params);
-                sum1 = activation_ss(sum1, activation_type, activation_params);
-
-                outptr0[0] = (__fp16)(sum0);
-                outptr1[0] = (__fp16)(sum1);
-                outptr0 += 1;
-                outptr1 += 1;
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 2;
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-
-                int q = 0;
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                for (; q + 7 < inch; q += 8)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        float32x4_t _r1;
-                        if (elempack == 4)
-                        {
-                            _r0 = vcvt_f32_f16(vld1_f16(r0 + sok));
-                            _r1 = vcvt_f32_f16(vld1_f16(r0 + sok + N));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            float16x8_t _r01_f16 = float16x8_t();
-                            _r01_f16 = vsetq_lane_f16(r0[sok], _r01_f16, 0);
-                            _r01_f16 = vsetq_lane_f16(r0[sok + N], _r01_f16, 1);
-                            _r01_f16 = vsetq_lane_f16(r0[sok + N * 2], _r01_f16, 2);
-                            _r01_f16 = vsetq_lane_f16(r0[sok + N * 3], _r01_f16, 3);
-                            _r01_f16 = vsetq_lane_f16(r0[sok + N * 4], _r01_f16, 4);
-                            _r01_f16 = vsetq_lane_f16(r0[sok + N * 5], _r01_f16, 5);
-                            _r01_f16 = vsetq_lane_f16(r0[sok + N * 6], _r01_f16, 6);
-                            _r01_f16 = vsetq_lane_f16(r0[sok + N * 7], _r01_f16, 7);
-                            _r0 = vcvt_f32_f16(vget_low_f16(_r01_f16));
-                            _r1 = vcvt_f32_f16(vget_high_f16(_r01_f16));
-                        }
-
-                        float16x8_t _w = vld1q_f16(kptr);
-                        float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w));
-                        float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w));
-                        _sum0 = vfmaq_f32(_sum0, _r0, _w0);
-                        _sum1 = vfmaq_f32(_sum1, _r1, _w1);
-
-                        kptr += 8;
-                    }
-                }
-                _sum0 = vaddq_f32(_sum0, _sum1);
-                sum += vaddvq_f32(_sum0);
-                float32x4_t _sum = vdupq_n_f32(0.f);
-                for (; q + 3 < inch; q += 4)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float32x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = vcvt_f32_f16(vld1_f16(r0 + sok));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            float16x4_t _r0_f16 = float16x4_t();
-                            _r0_f16 = vset_lane_f16(r0[sok], _r0_f16, 0);
-                            _r0_f16 = vset_lane_f16(r0[sok + N], _r0_f16, 1);
-                            _r0_f16 = vset_lane_f16(r0[sok + N * 2], _r0_f16, 2);
-                            _r0_f16 = vset_lane_f16(r0[sok + N * 3], _r0_f16, 3);
-                            _r0 = vcvt_f32_f16(_r0_f16);
-                        }
-
-                        float32x4_t _w = vcvt_f32_f16(vld1_f16(kptr));
-                        _sum = vfmaq_f32(_sum, _r0, _w);
-
-                        kptr += 4;
-                    }
-                }
-                sum += vaddvq_f32(_sum);
-                for (; q + 1 < inch; q += 2)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float val0;
-                        float val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = (float)(r0[sok]);
-                            val1 = (float)(r0[sok + N]);
-                        }
-
-                        sum += val0 * (float)(kptr[0]);
-                        sum += val1 * (float)(kptr[1]);
-
-                        kptr += 2;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float val;
-                        // if (elempack == 1)
-                        {
-                            val = (float)(r0[space_ofs[k]]);
-                        }
-
-                        sum += val * (float)(kptr[0]);
-
-                        kptr += 1;
-                    }
-                }
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[0] = (__fp16)(sum);
-                outptr += 1;
-            }
-        }
-    }
-}
-
-static void convolution_packed_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int elempack = bottom_blob.elempack;
-    const int inch = bottom_blob.c * elempack;
-
-    const int N = bottom_blob.cstep * elempack;
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int outch = top_blob.c * out_elempack;
-
-    const int M = top_blob.cstep * out_elempack;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2 * elempack;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const __fp16* bias_data_ptr = bias_data;
-
-    int nn_outch = 0;
-    int remain_outch_start = 0;
-    nn_outch = (outch - remain_outch_start) / 8;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 8;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int out_elempack = top_blob.elempack;
-
-        __fp16* outptr = top_blob.channel(p / out_elempack);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float16x8_t _sum0 = vdupq_n_f16(0.f);
-                float16x8_t _sum1 = vdupq_n_f16(0.f);
-                float16x8_t _sum2 = vdupq_n_f16(0.f);
-                float16x8_t _sum3 = vdupq_n_f16(0.f);
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = vld1q_f16(bias_data_ptr + p);
-                }
-
-                const __fp16* kptr = weight_data_tm.channel(p / 8);
-
-                int q = 0;
-                for (; q + 7 < inch; q += 8)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float16x8_t _r0;
-                        if (elempack == 8)
-                        {
-                            _r0 = vld1q_f16(r0 + sok);
-                        }
-                        else if (elempack == 4)
-                        {
-                            _r0 = vcombine_f16(vld1_f16(r0 + sok), vld1_f16(r0 + sok + N));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float16x8_t();
-                            _r0 = vsetq_lane_f16(r0[sok], _r0, 0);
-                            _r0 = vsetq_lane_f16(r0[sok + N], _r0, 1);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 2], _r0, 2);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 3], _r0, 3);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 4], _r0, 4);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 5], _r0, 5);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 6], _r0, 6);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 7], _r0, 7);
-                        }
-
-                        float16x8_t _w0 = vld1q_f16(kptr);
-                        float16x8_t _w1 = vld1q_f16(kptr + 8);
-                        float16x8_t _w2 = vld1q_f16(kptr + 8 * 2);
-                        float16x8_t _w3 = vld1q_f16(kptr + 8 * 3);
-                        float16x8_t _w4 = vld1q_f16(kptr + 8 * 4);
-                        float16x8_t _w5 = vld1q_f16(kptr + 8 * 5);
-                        float16x8_t _w6 = vld1q_f16(kptr + 8 * 6);
-                        float16x8_t _w7 = vld1q_f16(kptr + 8 * 7);
-                        _sum0 = vfmaq_laneq_f16(_sum0, _w0, _r0, 0);
-                        _sum1 = vfmaq_laneq_f16(_sum1, _w1, _r0, 1);
-                        _sum2 = vfmaq_laneq_f16(_sum2, _w2, _r0, 2);
-                        _sum3 = vfmaq_laneq_f16(_sum3, _w3, _r0, 3);
-                        _sum0 = vfmaq_laneq_f16(_sum0, _w4, _r0, 4);
-                        _sum1 = vfmaq_laneq_f16(_sum1, _w5, _r0, 5);
-                        _sum2 = vfmaq_laneq_f16(_sum2, _w6, _r0, 6);
-                        _sum3 = vfmaq_laneq_f16(_sum3, _w7, _r0, 7);
-
-                        kptr += 64;
-                    }
-                }
-                for (; q + 3 < inch; q += 4)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float16x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = vld1_f16(r0 + sok);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float16x4_t();
-                            _r0 = vset_lane_f16(r0[sok], _r0, 0);
-                            _r0 = vset_lane_f16(r0[sok + N], _r0, 1);
-                            _r0 = vset_lane_f16(r0[sok + N * 2], _r0, 2);
-                            _r0 = vset_lane_f16(r0[sok + N * 3], _r0, 3);
-                        }
-
-                        float16x8_t _w0 = vld1q_f16(kptr);
-                        float16x8_t _w1 = vld1q_f16(kptr + 8);
-                        float16x8_t _w2 = vld1q_f16(kptr + 8 * 2);
-                        float16x8_t _w3 = vld1q_f16(kptr + 8 * 3);
-                        _sum0 = vfmaq_lane_f16(_sum0, _w0, _r0, 0);
-                        _sum1 = vfmaq_lane_f16(_sum1, _w1, _r0, 1);
-                        _sum2 = vfmaq_lane_f16(_sum2, _w2, _r0, 2);
-                        _sum3 = vfmaq_lane_f16(_sum3, _w3, _r0, 3);
-
-                        kptr += 32;
-                    }
-                }
-                for (; q + 1 < inch; q += 2)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        __fp16 val0;
-                        __fp16 val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = r0[sok];
-                            val1 = r0[sok + N];
-                        }
-
-                        float16x8_t _w0 = vld1q_f16(kptr);
-                        float16x8_t _w1 = vld1q_f16(kptr + 8);
-                        _sum0 = vfmaq_n_f16(_sum0, _w0, val0);
-                        _sum1 = vfmaq_n_f16(_sum1, _w1, val1);
-
-                        kptr += 16;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float16x8_t _val;
-                        // if (elempack == 1)
-                        {
-                            _val = vdupq_n_f16(r0[space_ofs[k]]);
-                        }
-
-                        float16x8_t _w0 = vld1q_f16(kptr);
-                        _sum0 = vfmaq_f16(_sum0, _w0, _val);
-
-                        kptr += 8;
-                    }
-                }
-
-                _sum0 = vaddq_f16(_sum0, _sum1);
-                _sum2 = vaddq_f16(_sum2, _sum3);
-                _sum0 = vaddq_f16(_sum0, _sum2);
-
-                _sum0 = activation_ps_f16(_sum0, activation_type, activation_params);
-
-                if (out_elempack == 8)
-                {
-                    vst1q_f16(outptr, _sum0);
-                    outptr += 8;
-                }
-                else if (out_elempack == 4)
-                {
-                    vst1_f16(outptr, vget_low_f16(_sum0));
-                    vst1_f16(outptr + M, vget_high_f16(_sum0));
-                    outptr += 4;
-                }
-                else // if (out_elempack == 1)
-                {
-                    outptr[0] = vgetq_lane_f16(_sum0, 0);
-                    outptr[M] = vgetq_lane_f16(_sum0, 1);
-                    outptr[M * 2] = vgetq_lane_f16(_sum0, 2);
-                    outptr[M * 3] = vgetq_lane_f16(_sum0, 3);
-                    outptr[M * 4] = vgetq_lane_f16(_sum0, 4);
-                    outptr[M * 5] = vgetq_lane_f16(_sum0, 5);
-                    outptr[M * 6] = vgetq_lane_f16(_sum0, 6);
-                    outptr[M * 7] = vgetq_lane_f16(_sum0, 7);
-                    outptr += 1;
-                }
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 8;
-    nn_outch = (outch - remain_outch_start) / 4;
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 4;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int out_elempack = top_blob.elempack;
-
-        __fp16* outptr = top_blob.channel(p / out_elempack);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float16x4_t _sum0 = vdup_n_f16(0.f);
-                float16x4_t _sum1 = vdup_n_f16(0.f);
-                float16x4_t _sum2 = vdup_n_f16(0.f);
-                float16x4_t _sum3 = vdup_n_f16(0.f);
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = vld1_f16(bias_data_ptr + p);
-                }
-
-                const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4);
-
-                int q = 0;
-                for (; q + 7 < inch; q += 8)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float16x4_t _r0;
-                        float16x4_t _r1;
-                        if (elempack == 8)
-                        {
-                            float16x8_t _r01 = vld1q_f16(r0 + sok);
-                            _r0 = vget_low_f16(_r01);
-                            _r1 = vget_high_f16(_r01);
-                        }
-                        else if (elempack == 4)
-                        {
-                            _r0 = vld1_f16(r0 + sok);
-                            _r1 = vld1_f16(r0 + sok + N);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float16x4_t();
-                            _r1 = float16x4_t();
-                            _r0 = vset_lane_f16(r0[sok], _r0, 0);
-                            _r0 = vset_lane_f16(r0[sok + N], _r0, 1);
-                            _r0 = vset_lane_f16(r0[sok + N * 2], _r0, 2);
-                            _r0 = vset_lane_f16(r0[sok + N * 3], _r0, 3);
-                            _r1 = vset_lane_f16(r0[sok + N * 4], _r1, 0);
-                            _r1 = vset_lane_f16(r0[sok + N * 5], _r1, 1);
-                            _r1 = vset_lane_f16(r0[sok + N * 6], _r1, 2);
-                            _r1 = vset_lane_f16(r0[sok + N * 7], _r1, 3);
-                        }
-
-                        float16x4_t _w0 = vld1_f16(kptr);
-                        float16x4_t _w1 = vld1_f16(kptr + 4);
-                        float16x4_t _w2 = vld1_f16(kptr + 8);
-                        float16x4_t _w3 = vld1_f16(kptr + 12);
-                        float16x4_t _w4 = vld1_f16(kptr + 16);
-                        float16x4_t _w5 = vld1_f16(kptr + 20);
-                        float16x4_t _w6 = vld1_f16(kptr + 24);
-                        float16x4_t _w7 = vld1_f16(kptr + 28);
-                        _sum0 = vfma_lane_f16(_sum0, _w0, _r0, 0);
-                        _sum1 = vfma_lane_f16(_sum1, _w1, _r0, 1);
-                        _sum2 = vfma_lane_f16(_sum2, _w2, _r0, 2);
-                        _sum3 = vfma_lane_f16(_sum3, _w3, _r0, 3);
-                        _sum0 = vfma_lane_f16(_sum0, _w4, _r1, 0);
-                        _sum1 = vfma_lane_f16(_sum1, _w5, _r1, 1);
-                        _sum2 = vfma_lane_f16(_sum2, _w6, _r1, 2);
-                        _sum3 = vfma_lane_f16(_sum3, _w7, _r1, 3);
-
-                        kptr += 32;
-                    }
-                }
-                for (; q + 3 < inch; q += 4)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float16x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = vld1_f16(r0 + sok);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float16x4_t();
-                            _r0 = vset_lane_f16(r0[sok], _r0, 0);
-                            _r0 = vset_lane_f16(r0[sok + N], _r0, 1);
-                            _r0 = vset_lane_f16(r0[sok + N * 2], _r0, 2);
-                            _r0 = vset_lane_f16(r0[sok + N * 3], _r0, 3);
-                        }
-
-                        float16x4_t _w0 = vld1_f16(kptr);
-                        float16x4_t _w1 = vld1_f16(kptr + 4);
-                        float16x4_t _w2 = vld1_f16(kptr + 8);
-                        float16x4_t _w3 = vld1_f16(kptr + 12);
-                        _sum0 = vfma_lane_f16(_sum0, _w0, _r0, 0);
-                        _sum1 = vfma_lane_f16(_sum1, _w1, _r0, 1);
-                        _sum2 = vfma_lane_f16(_sum2, _w2, _r0, 2);
-                        _sum3 = vfma_lane_f16(_sum3, _w3, _r0, 3);
-
-                        kptr += 16;
-                    }
-                }
-                for (; q + 1 < inch; q += 2)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        __fp16 val0;
-                        __fp16 val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = r0[sok];
-                            val1 = r0[sok + N];
-                        }
-
-                        float16x4_t _w0 = vld1_f16(kptr);
-                        float16x4_t _w1 = vld1_f16(kptr + 4);
-                        _sum0 = vfma_n_f16(_sum0, _w0, val0);
-                        _sum1 = vfma_n_f16(_sum1, _w1, val1);
-
-                        kptr += 8;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float16x4_t _val;
-                        // if (elempack == 1)
-                        {
-                            _val = vdup_n_f16(r0[space_ofs[k]]);
-                        }
-
-                        float16x4_t _w = vld1_f16(kptr);
-                        _sum0 = vfma_f16(_sum0, _val, _w);
-
-                        kptr += 4;
-                    }
-                }
-
-                _sum0 = vadd_f16(_sum0, _sum1);
-                _sum2 = vadd_f16(_sum2, _sum3);
-                _sum0 = vadd_f16(_sum0, _sum2);
-
-                _sum0 = activation_ps_f16(_sum0, activation_type, activation_params);
-
-                if (out_elempack == 4)
-                {
-                    vst1_f16(outptr, _sum0);
-                    outptr += 4;
-                }
-                else // if (out_elempack == 1)
-                {
-                    outptr[0] = vget_lane_f16(_sum0, 0);
-                    outptr[M] = vget_lane_f16(_sum0, 1);
-                    outptr[M * 2] = vget_lane_f16(_sum0, 2);
-                    outptr[M * 3] = vget_lane_f16(_sum0, 3);
-                    outptr += 1;
-                }
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 4;
-    nn_outch = (outch - remain_outch_start) / 2;
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 2;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        __fp16* outptr0 = top_blob.channel(p);
-        __fp16* outptr1 = top_blob.channel(p + 1);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __fp16 sum0 = 0.f;
-                __fp16 sum1 = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum0 = bias_data_ptr[p];
-                    sum1 = bias_data_ptr[p + 1];
-                }
-
-                const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2);
-
-                int q = 0;
-                float16x8_t _sum0 = vdupq_n_f16(0.f);
-                float16x8_t _sum1 = vdupq_n_f16(0.f);
-                for (; q + 7 < inch; q += 8)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float16x8_t _r0;
-                        if (elempack == 8)
-                        {
-                            _r0 = vld1q_f16(r0 + sok);
-                        }
-                        else if (elempack == 4)
-                        {
-                            _r0 = vcombine_f16(vld1_f16(r0 + sok), vld1_f16(r0 + sok + N));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float16x8_t();
-                            _r0 = vsetq_lane_f16(r0[sok], _r0, 0);
-                            _r0 = vsetq_lane_f16(r0[sok + N], _r0, 1);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 2], _r0, 2);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 3], _r0, 3);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 4], _r0, 4);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 5], _r0, 5);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 6], _r0, 6);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 7], _r0, 7);
-                        }
-
-                        float16x8_t _w0 = vld1q_f16(kptr);
-                        float16x8_t _w1 = vld1q_f16(kptr + 8);
-                        _sum0 = vfmaq_f16(_sum0, _r0, _w0);
-                        _sum1 = vfmaq_f16(_sum1, _r0, _w1);
-
-                        kptr += 16;
-                    }
-                }
-                for (; q + 3 < inch; q += 4)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float16x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = vld1_f16(r0 + sok);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float16x4_t();
-                            _r0 = vset_lane_f16(r0[sok], _r0, 0);
-                            _r0 = vset_lane_f16(r0[sok + N], _r0, 1);
-                            _r0 = vset_lane_f16(r0[sok + N * 2], _r0, 2);
-                            _r0 = vset_lane_f16(r0[sok + N * 3], _r0, 3);
-                        }
-
-                        float16x4_t _w0 = vld1_f16(kptr);
-                        float16x4_t _w1 = vld1_f16(kptr + 4);
-                        _sum0 = vcombine_f16(vfma_f16(vget_low_f16(_sum0), _r0, _w0), vget_high_f16(_sum0));
-                        _sum1 = vcombine_f16(vfma_f16(vget_low_f16(_sum1), _r0, _w1), vget_high_f16(_sum1));
-
-                        kptr += 8;
-                    }
-                }
-                for (; q + 1 < inch; q += 2)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        __fp16 val0;
-                        __fp16 val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = r0[sok];
-                            val1 = r0[sok + N];
-                        }
-
-                        sum0 += val0 * kptr[0];
-                        sum1 += val0 * kptr[1];
-                        sum0 += val1 * kptr[2];
-                        sum1 += val1 * kptr[3];
-
-                        kptr += 4;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        __fp16 val;
-                        // if (elempack == 1)
-                        {
-                            val = r0[space_ofs[k]];
-                        }
-
-                        sum0 += val * kptr[0];
-                        sum1 += val * kptr[1];
-
-                        kptr += 2;
-                    }
-                }
-
-                float16x4_t _ss0 = vadd_f16(vget_low_f16(_sum0), vget_high_f16(_sum0));
-                float16x4_t _ss1 = vadd_f16(vget_low_f16(_sum1), vget_high_f16(_sum1));
-                float16x4_t _ss = vpadd_f16(_ss0, _ss1);
-                _ss = vpadd_f16(_ss, _ss);
-                sum0 += vget_lane_f16(_ss, 0);
-                sum1 += vget_lane_f16(_ss, 1);
-
-                sum0 = activation_ss_f16(sum0, activation_type, activation_params);
-                sum1 = activation_ss_f16(sum1, activation_type, activation_params);
-
-                outptr0[0] = sum0;
-                outptr1[0] = sum1;
-                outptr0 += 1;
-                outptr1 += 1;
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 2;
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __fp16 sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                const __fp16* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-
-                int q = 0;
-                float16x8_t _sum = vdupq_n_f16(0.f);
-                for (; q + 7 < inch; q += 8)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float16x8_t _r0;
-                        if (elempack == 8)
-                        {
-                            _r0 = vld1q_f16(r0 + sok);
-                        }
-                        else if (elempack == 4)
-                        {
-                            _r0 = vcombine_f16(vld1_f16(r0 + sok), vld1_f16(r0 + sok + N));
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float16x8_t();
-                            _r0 = vsetq_lane_f16(r0[sok], _r0, 0);
-                            _r0 = vsetq_lane_f16(r0[sok + N], _r0, 1);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 2], _r0, 2);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 3], _r0, 3);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 4], _r0, 4);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 5], _r0, 5);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 6], _r0, 6);
-                            _r0 = vsetq_lane_f16(r0[sok + N * 7], _r0, 7);
-                        }
-
-                        float16x8_t _w0 = vld1q_f16(kptr);
-                        _sum = vfmaq_f16(_sum, _r0, _w0);
-
-                        kptr += 8;
-                    }
-                }
-                for (; q + 3 < inch; q += 4)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q / elempack).row<const __fp16>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        float16x4_t _r0;
-                        if (elempack == 4)
-                        {
-                            _r0 = vld1_f16(r0 + sok);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            _r0 = float16x4_t();
-                            _r0 = vset_lane_f16(r0[sok], _r0, 0);
-                            _r0 = vset_lane_f16(r0[sok + N], _r0, 1);
-                            _r0 = vset_lane_f16(r0[sok + N * 2], _r0, 2);
-                            _r0 = vset_lane_f16(r0[sok + N * 3], _r0, 3);
-                        }
-
-                        float16x4_t _w = vld1_f16(kptr);
-                        _sum = vcombine_f16(vfma_f16(vget_low_f16(_sum), _r0, _w), vget_high_f16(_sum));
-
-                        kptr += 4;
-                    }
-                }
-                for (; q + 1 < inch; q += 2)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const int sok = space_ofs[k];
-                        __fp16 val0;
-                        __fp16 val1;
-                        // if (elempack == 1)
-                        {
-                            val0 = r0[sok];
-                            val1 = r0[sok + N];
-                        }
-
-                        sum += val0 * kptr[0];
-                        sum += val1 * kptr[1];
-
-                        kptr += 2;
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const __fp16* r0 = bottom_blob.channel(q).row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        __fp16 val;
-                        // if (elempack == 1)
-                        {
-                            val = r0[space_ofs[k]];
-                        }
-
-                        sum += val * kptr[0];
-
-                        kptr += 1;
-                    }
-                }
-
-                float16x4_t _ss = vadd_f16(vget_low_f16(_sum), vget_high_f16(_sum));
-                _ss = vpadd_f16(_ss, _ss);
-                _ss = vpadd_f16(_ss, _ss);
-                sum += vget_lane_f16(_ss, 0);
-
-                sum = activation_ss_f16(sum, activation_type, activation_params);
-
-                outptr[0] = sum;
-                outptr += 1;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed_int8.h
deleted file mode 100644
index 19342a0d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolution_packed_int8.h
+++ /dev/null
@@ -1,1573 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if !(__ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD)
-#if NCNN_RUNTIME_CPU && NCNN_ARM84I8MM && __aarch64__ && !__ARM_FEATURE_MATMUL_INT8
-void convolution_transform_kernel_packed_int8_i8mm(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h);
-void convolution_packed_int8_i8mm(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt);
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_ARM82DOT && __aarch64__ && !__ARM_FEATURE_DOTPROD
-void convolution_transform_kernel_packed_int8_asimddp(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h);
-void convolution_packed_int8_asimddp(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt);
-#endif
-#endif
-
-static void convolution_transform_kernel_packed_int8(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-#if !(__ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD)
-#if NCNN_RUNTIME_CPU && NCNN_ARM84I8MM && __aarch64__ && !__ARM_FEATURE_MATMUL_INT8
-    if (ncnn::cpu_support_arm_i8mm())
-    {
-        convolution_transform_kernel_packed_int8_i8mm(kernel, kernel_tm, inch, outch, kernel_w, kernel_h);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_ARM82DOT && __aarch64__ && !__ARM_FEATURE_DOTPROD
-    if (ncnn::cpu_support_arm_asimddp())
-    {
-        convolution_transform_kernel_packed_int8_asimddp(kernel, kernel_tm, inch, outch, kernel_w, kernel_h);
-        return;
-    }
-#endif
-#endif
-
-    const int maxk = kernel_w * kernel_h;
-
-    // src = kw-kh-inch-outch
-    // dst = pb-pa-kw-kh-inch/pa-outch/pb
-
-    // clang-format off
-    // *INDENT-OFF*
-#if __ARM_NEON
-    if (outch >= 8)
-    {
-        if (inch >= 8)
-            kernel_tm.create(maxk, inch / 8 + inch % 8, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)64u, 64);
-        else
-            kernel_tm.create(maxk, inch, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)8u, 8);
-    }
-    else if (outch >= 4)
-    {
-        if (inch >= 8)
-            kernel_tm.create(maxk, inch / 8 + inch % 8, outch / 4 + (outch % 4) / 2 + outch % 2, (size_t)32u, 32);
-        else
-            kernel_tm.create(maxk, inch, outch / 4 + (outch % 4) / 2 + outch % 2, (size_t)4u, 4);
-    }
-    else
-#endif // __ARM_NEON
-    if (outch >= 2)
-    {
-#if __ARM_NEON
-        if (inch >= 8)
-            kernel_tm.create(maxk, inch / 8 + inch % 8, outch / 2 + outch % 2, (size_t)16u, 16);
-        else
-#endif // __ARM_NEON
-            kernel_tm.create(maxk, inch, outch / 2 + outch % 2, (size_t)2u, 2);
-    }
-    else
-    {
-#if __ARM_NEON
-        if (inch >= 8)
-            kernel_tm.create(maxk, inch / 8 + inch % 8, outch, (size_t)8u, 8);
-        else
-#endif // __ARM_NEON
-            kernel_tm.create(maxk, inch, outch, (size_t)1u, 1);
-    }
-    // *INDENT-ON*
-    // clang-format on
-
-    int q = 0;
-#if __ARM_NEON
-    for (; q + 7 < outch; q += 8)
-    {
-        const signed char* kptr0 = (const signed char*)kernel + q * inch * maxk;
-        const signed char* kptr1 = (const signed char*)kernel + (q + 1) * inch * maxk;
-        const signed char* kptr2 = (const signed char*)kernel + (q + 2) * inch * maxk;
-        const signed char* kptr3 = (const signed char*)kernel + (q + 3) * inch * maxk;
-        const signed char* kptr4 = (const signed char*)kernel + (q + 4) * inch * maxk;
-        const signed char* kptr5 = (const signed char*)kernel + (q + 5) * inch * maxk;
-        const signed char* kptr6 = (const signed char*)kernel + (q + 6) * inch * maxk;
-        const signed char* kptr7 = (const signed char*)kernel + (q + 7) * inch * maxk;
-
-        signed char* g00 = kernel_tm.channel(q / 8);
-
-        int p = 0;
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k0 = kptr0 + k;
-                const signed char* k1 = kptr1 + k;
-                const signed char* k2 = kptr2 + k;
-                const signed char* k3 = kptr3 + k;
-                const signed char* k4 = kptr4 + k;
-                const signed char* k5 = kptr5 + k;
-                const signed char* k6 = kptr6 + k;
-                const signed char* k7 = kptr7 + k;
-
-#if __ARM_FEATURE_MATMUL_INT8
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[0];
-                    g00 += 1;
-                    k0 += maxk;
-                }
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k1[0];
-                    g00 += 1;
-                    k1 += maxk;
-                }
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k2[0];
-                    g00 += 1;
-                    k2 += maxk;
-                }
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k3[0];
-                    g00 += 1;
-                    k3 += maxk;
-                }
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k4[0];
-                    g00 += 1;
-                    k4 += maxk;
-                }
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k5[0];
-                    g00 += 1;
-                    k5 += maxk;
-                }
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k6[0];
-                    g00 += 1;
-                    k6 += maxk;
-                }
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k7[0];
-                    g00 += 1;
-                    k7 += maxk;
-                }
-#elif __ARM_FEATURE_DOTPROD
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = k0[0];
-                    g00[1] = k0[maxk];
-                    g00[2] = k0[maxk * 2];
-                    g00[3] = k0[maxk * 3];
-                    g00[4] = k1[0];
-                    g00[5] = k1[maxk];
-                    g00[6] = k1[maxk * 2];
-                    g00[7] = k1[maxk * 3];
-                    g00[8] = k2[0];
-                    g00[9] = k2[maxk];
-                    g00[10] = k2[maxk * 2];
-                    g00[11] = k2[maxk * 3];
-                    g00[12] = k3[0];
-                    g00[13] = k3[maxk];
-                    g00[14] = k3[maxk * 2];
-                    g00[15] = k3[maxk * 3];
-                    g00[16] = k4[0];
-                    g00[17] = k4[maxk];
-                    g00[18] = k4[maxk * 2];
-                    g00[19] = k4[maxk * 3];
-                    g00[20] = k5[0];
-                    g00[21] = k5[maxk];
-                    g00[22] = k5[maxk * 2];
-                    g00[23] = k5[maxk * 3];
-                    g00[24] = k6[0];
-                    g00[25] = k6[maxk];
-                    g00[26] = k6[maxk * 2];
-                    g00[27] = k6[maxk * 3];
-                    g00[28] = k7[0];
-                    g00[29] = k7[maxk];
-                    g00[30] = k7[maxk * 2];
-                    g00[31] = k7[maxk * 3];
-                    g00 += 32;
-                    k0 += maxk * 4;
-                    k1 += maxk * 4;
-                    k2 += maxk * 4;
-                    k3 += maxk * 4;
-                    k4 += maxk * 4;
-                    k5 += maxk * 4;
-                    k6 += maxk * 4;
-                    k7 += maxk * 4;
-                }
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k0[0];
-                    g00[1] = k0[maxk];
-                    g00[2] = k1[0];
-                    g00[3] = k1[maxk];
-                    g00[4] = k2[0];
-                    g00[5] = k2[maxk];
-                    g00[6] = k3[0];
-                    g00[7] = k3[maxk];
-                    g00[8] = k4[0];
-                    g00[9] = k4[maxk];
-                    g00[10] = k5[0];
-                    g00[11] = k5[maxk];
-                    g00[12] = k6[0];
-                    g00[13] = k6[maxk];
-                    g00[14] = k7[0];
-                    g00[15] = k7[maxk];
-                    g00 += 16;
-                    k0 += maxk * 2;
-                    k1 += maxk * 2;
-                    k2 += maxk * 2;
-                    k3 += maxk * 2;
-                    k4 += maxk * 2;
-                    k5 += maxk * 2;
-                    k6 += maxk * 2;
-                    k7 += maxk * 2;
-                }
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-            }
-
-            kptr0 += maxk * 8;
-            kptr1 += maxk * 8;
-            kptr2 += maxk * 8;
-            kptr3 += maxk * 8;
-            kptr4 += maxk * 8;
-            kptr5 += maxk * 8;
-            kptr6 += maxk * 8;
-            kptr7 += maxk * 8;
-        }
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k0 = kptr0 + k;
-                const signed char* k1 = kptr1 + k;
-                const signed char* k2 = kptr2 + k;
-                const signed char* k3 = kptr3 + k;
-                const signed char* k4 = kptr4 + k;
-                const signed char* k5 = kptr5 + k;
-                const signed char* k6 = kptr6 + k;
-                const signed char* k7 = kptr7 + k;
-
-                g00[0] = k0[0];
-                g00[1] = k1[0];
-                g00[2] = k2[0];
-                g00[3] = k3[0];
-                g00[4] = k4[0];
-                g00[5] = k5[0];
-                g00[6] = k6[0];
-                g00[7] = k7[0];
-                g00 += 8;
-            }
-
-            kptr0 += maxk;
-            kptr1 += maxk;
-            kptr2 += maxk;
-            kptr3 += maxk;
-            kptr4 += maxk;
-            kptr5 += maxk;
-            kptr6 += maxk;
-            kptr7 += maxk;
-        }
-    }
-    for (; q + 3 < outch; q += 4)
-    {
-        const signed char* kptr0 = (const signed char*)kernel + q * inch * maxk;
-        const signed char* kptr1 = (const signed char*)kernel + (q + 1) * inch * maxk;
-        const signed char* kptr2 = (const signed char*)kernel + (q + 2) * inch * maxk;
-        const signed char* kptr3 = (const signed char*)kernel + (q + 3) * inch * maxk;
-
-        signed char* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4);
-
-        int p = 0;
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k0 = kptr0 + k;
-                const signed char* k1 = kptr1 + k;
-                const signed char* k2 = kptr2 + k;
-                const signed char* k3 = kptr3 + k;
-
-#if __ARM_FEATURE_MATMUL_INT8
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[0];
-                    g00 += 1;
-                    k0 += maxk;
-                }
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k1[0];
-                    g00 += 1;
-                    k1 += maxk;
-                }
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k2[0];
-                    g00 += 1;
-                    k2 += maxk;
-                }
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k3[0];
-                    g00 += 1;
-                    k3 += maxk;
-                }
-#elif __ARM_FEATURE_DOTPROD
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = k0[0];
-                    g00[1] = k0[maxk];
-                    g00[2] = k0[maxk * 2];
-                    g00[3] = k0[maxk * 3];
-                    g00[4] = k1[0];
-                    g00[5] = k1[maxk];
-                    g00[6] = k1[maxk * 2];
-                    g00[7] = k1[maxk * 3];
-                    g00[8] = k2[0];
-                    g00[9] = k2[maxk];
-                    g00[10] = k2[maxk * 2];
-                    g00[11] = k2[maxk * 3];
-                    g00[12] = k3[0];
-                    g00[13] = k3[maxk];
-                    g00[14] = k3[maxk * 2];
-                    g00[15] = k3[maxk * 3];
-                    g00 += 16;
-                    k0 += maxk * 4;
-                    k1 += maxk * 4;
-                    k2 += maxk * 4;
-                    k3 += maxk * 4;
-                }
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k0[0];
-                    g00[1] = k0[maxk];
-                    g00[2] = k1[0];
-                    g00[3] = k1[maxk];
-                    g00[4] = k2[0];
-                    g00[5] = k2[maxk];
-                    g00[6] = k3[0];
-                    g00[7] = k3[maxk];
-                    g00 += 8;
-                    k0 += maxk * 2;
-                    k1 += maxk * 2;
-                    k2 += maxk * 2;
-                    k3 += maxk * 2;
-                }
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-            }
-
-            kptr0 += maxk * 8;
-            kptr1 += maxk * 8;
-            kptr2 += maxk * 8;
-            kptr3 += maxk * 8;
-        }
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k0 = kptr0 + k;
-                const signed char* k1 = kptr1 + k;
-                const signed char* k2 = kptr2 + k;
-                const signed char* k3 = kptr3 + k;
-
-                g00[0] = k0[0];
-                g00[1] = k1[0];
-                g00[2] = k2[0];
-                g00[3] = k3[0];
-                g00 += 4;
-            }
-
-            kptr0 += maxk;
-            kptr1 += maxk;
-            kptr2 += maxk;
-            kptr3 += maxk;
-        }
-    }
-#endif // __ARM_NEON
-    for (; q + 1 < outch; q += 2)
-    {
-        const signed char* kptr0 = (const signed char*)kernel + q * inch * maxk;
-        const signed char* kptr1 = (const signed char*)kernel + (q + 1) * inch * maxk;
-
-#if __ARM_NEON
-        signed char* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2);
-#else
-        signed char* g00 = kernel_tm.channel(q / 2);
-#endif
-
-        int p = 0;
-#if __ARM_NEON
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k0 = kptr0 + k;
-                const signed char* k1 = kptr1 + k;
-
-#if __ARM_FEATURE_DOTPROD
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[0];
-                    k0 += maxk;
-                    g00 += 1;
-                }
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k1[0];
-                    k1 += maxk;
-                    g00 += 1;
-                }
-#else  // __ARM_FEATURE_DOTPROD
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k0[0];
-                    k0 += maxk;
-                    g00 += 1;
-                }
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k1[0];
-                    k1 += maxk;
-                    g00 += 1;
-                }
-
-                for (int i = 4; i < 8; i++)
-                {
-                    g00[0] = k0[0];
-                    k0 += maxk;
-                    g00 += 1;
-                }
-                for (int i = 4; i < 8; i++)
-                {
-                    g00[0] = k1[0];
-                    k1 += maxk;
-                    g00 += 1;
-                }
-#endif // __ARM_FEATURE_DOTPROD
-            }
-
-            kptr0 += maxk * 8;
-            kptr1 += maxk * 8;
-        }
-#endif // __ARM_NEON
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k0 = kptr0 + k;
-                const signed char* k1 = kptr1 + k;
-
-                g00[0] = k0[0];
-                g00[1] = k1[0];
-                g00 += 2;
-            }
-
-            kptr0 += maxk;
-            kptr1 += maxk;
-        }
-    }
-    for (; q < outch; q++)
-    {
-        const signed char* kptr = (const signed char*)kernel + q * inch * maxk;
-
-#if __ARM_NEON
-        signed char* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2 + q % 2);
-#else
-        signed char* g00 = kernel_tm.channel(q / 2 + q % 2);
-#endif
-
-        int p = 0;
-#if __ARM_NEON
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k0 = kptr + k;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[0];
-                    k0 += maxk;
-                    g00 += 1;
-                }
-            }
-
-            kptr += maxk * 8;
-        }
-#endif // __ARM_NEON
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k0 = kptr + k;
-                g00[0] = k0[0];
-                g00++;
-            }
-
-            kptr += maxk;
-        }
-    }
-}
-
-static void convolution_packed_int8(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-#if !(__ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD)
-#if NCNN_RUNTIME_CPU && NCNN_ARM84I8MM && __aarch64__ && !__ARM_FEATURE_MATMUL_INT8
-    if (ncnn::cpu_support_arm_i8mm())
-    {
-        convolution_packed_int8_i8mm(bottom_blob, top_blob, weight_data_tm, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_ARM82DOT && __aarch64__ && !__ARM_FEATURE_DOTPROD
-    if (ncnn::cpu_support_arm_asimddp())
-    {
-        convolution_packed_int8_asimddp(bottom_blob, top_blob, weight_data_tm, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
-        return;
-    }
-#endif
-#endif
-
-    const int w = bottom_blob.w;
-    const int elempack = bottom_blob.elempack;
-    const int inch = bottom_blob.c * elempack;
-
-    const int N = bottom_blob.cstep * elempack;
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int outch = top_blob.c * out_elempack;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2 * elempack;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    int nn_outch = 0;
-    int remain_outch_start = 0;
-#if __ARM_NEON
-    nn_outch = (outch - remain_outch_start) / 8;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 8;
-
-        // shadowed variable for less openmp task args
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int N = bottom_blob.cstep * elempack;
-        const int M = top_blob.cstep * out_elempack;
-
-        int* outptr = top_blob.channel(p / out_elempack);
-
-        int ij = 0;
-        for (; ij + 1 < outw * outh; ij += 2)
-        {
-            const int i0 = ij / outw;
-            const int i1 = (ij + 1) / outw;
-            const int j0 = ij % outw;
-            const int j1 = (ij + 1) % outw;
-
-            int32x4_t _sum0 = vdupq_n_s32(0);
-            int32x4_t _sum1 = vdupq_n_s32(0);
-            int32x4_t _sum2 = vdupq_n_s32(0);
-            int32x4_t _sum3 = vdupq_n_s32(0);
-
-            const signed char* kptr = weight_data_tm.channel(p / 8);
-
-            int q = 0;
-            {
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-
-                        int8x8_t _r0;
-                        int8x8_t _r1;
-                        if (elempack == 8)
-                        {
-                            _r0 = vld1_s8(r0s);
-                            _r1 = vld1_s8(r1s);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            signed char tmp0[8] = {r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7]};
-                            signed char tmp1[8] = {r1s[0], r1s[N], r1s[N * 2], r1s[N * 3], r1s[N * 4], r1s[N * 5], r1s[N * 6], r1s[N * 7]};
-                            _r0 = vld1_s8(tmp0);
-                            _r1 = vld1_s8(tmp1);
-                        }
-
-                        int8x16_t _w0 = vld1q_s8(kptr);
-                        int8x16_t _w1 = vld1q_s8(kptr + 16);
-                        int8x16_t _w2 = vld1q_s8(kptr + 32);
-                        int8x16_t _w3 = vld1q_s8(kptr + 48);
-
-#if __ARM_FEATURE_MATMUL_INT8
-                        int8x16_t _r01 = vcombine_s8(_r0, _r1);
-                        _sum0 = vmmlaq_s32(_sum0, _r01, _w0);
-                        _sum1 = vmmlaq_s32(_sum1, _r01, _w1);
-                        _sum2 = vmmlaq_s32(_sum2, _r01, _w2);
-                        _sum3 = vmmlaq_s32(_sum3, _r01, _w3);
-#elif __ARM_FEATURE_DOTPROD
-                        _sum0 = vdotq_lane_s32(_sum0, _w0, _r0, 0);
-                        _sum1 = vdotq_lane_s32(_sum1, _w1, _r0, 0);
-                        _sum2 = vdotq_lane_s32(_sum2, _w0, _r1, 0);
-                        _sum3 = vdotq_lane_s32(_sum3, _w1, _r1, 0);
-                        _sum0 = vdotq_lane_s32(_sum0, _w2, _r0, 1);
-                        _sum1 = vdotq_lane_s32(_sum1, _w3, _r0, 1);
-                        _sum2 = vdotq_lane_s32(_sum2, _w2, _r1, 1);
-                        _sum3 = vdotq_lane_s32(_sum3, _w3, _r1, 1);
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                        int16x4_t _rr0 = vreinterpret_s16_s8(_r0);
-                        int16x4_t _rr1 = vreinterpret_s16_s8(_r1);
-
-                        int8x8_t _r0ll = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 0));
-                        int8x8_t _r1ll = vreinterpret_s8_s16(vdup_lane_s16(_rr1, 0));
-                        int8x8_t _r0hl = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 2));
-                        int8x8_t _r1hl = vreinterpret_s8_s16(vdup_lane_s16(_rr1, 2));
-
-                        int16x8_t _s0l = vmull_s8(_r0ll, vget_low_s8(_w0));
-                        int16x8_t _s1l = vmull_s8(_r0ll, vget_high_s8(_w0));
-                        int16x8_t _s2l = vmull_s8(_r1ll, vget_low_s8(_w0));
-                        int16x8_t _s3l = vmull_s8(_r1ll, vget_high_s8(_w0));
-                        _s0l = vmlal_s8(_s0l, _r0hl, vget_low_s8(_w2));
-                        _s1l = vmlal_s8(_s1l, _r0hl, vget_high_s8(_w2));
-                        _s2l = vmlal_s8(_s2l, _r1hl, vget_low_s8(_w2));
-                        _s3l = vmlal_s8(_s3l, _r1hl, vget_high_s8(_w2));
-
-                        _sum0 = vpadalq_s16(_sum0, _s0l);
-                        _sum1 = vpadalq_s16(_sum1, _s1l);
-                        _sum2 = vpadalq_s16(_sum2, _s2l);
-                        _sum3 = vpadalq_s16(_sum3, _s3l);
-
-                        int8x8_t _r0lh = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 1));
-                        int8x8_t _r1lh = vreinterpret_s8_s16(vdup_lane_s16(_rr1, 1));
-                        int8x8_t _r0hh = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 3));
-                        int8x8_t _r1hh = vreinterpret_s8_s16(vdup_lane_s16(_rr1, 3));
-
-                        int16x8_t _s0h = vmull_s8(_r0lh, vget_low_s8(_w1));
-                        int16x8_t _s1h = vmull_s8(_r0lh, vget_high_s8(_w1));
-                        int16x8_t _s2h = vmull_s8(_r1lh, vget_low_s8(_w1));
-                        int16x8_t _s3h = vmull_s8(_r1lh, vget_high_s8(_w1));
-                        _s0h = vmlal_s8(_s0h, _r0hh, vget_low_s8(_w3));
-                        _s1h = vmlal_s8(_s1h, _r0hh, vget_high_s8(_w3));
-                        _s2h = vmlal_s8(_s2h, _r1hh, vget_low_s8(_w3));
-                        _s3h = vmlal_s8(_s3h, _r1hh, vget_high_s8(_w3));
-
-                        _sum0 = vpadalq_s16(_sum0, _s0h);
-                        _sum1 = vpadalq_s16(_sum1, _s1h);
-                        _sum2 = vpadalq_s16(_sum2, _s2h);
-                        _sum3 = vpadalq_s16(_sum3, _s3h);
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-
-                        kptr += 64;
-                    }
-                }
-#if __ARM_FEATURE_MATMUL_INT8
-                {
-                    int32x4_t _tmp0 = vcombine_s32(vget_low_s32(_sum0), vget_low_s32(_sum1));
-                    int32x4_t _tmp1 = vcombine_s32(vget_low_s32(_sum2), vget_low_s32(_sum3));
-                    int32x4_t _tmp2 = vcombine_s32(vget_high_s32(_sum0), vget_high_s32(_sum1));
-                    int32x4_t _tmp3 = vcombine_s32(vget_high_s32(_sum2), vget_high_s32(_sum3));
-                    _sum0 = _tmp0;
-                    _sum1 = _tmp1;
-                    _sum2 = _tmp2;
-                    _sum3 = _tmp3;
-                }
-#endif
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        int8x8_t _r0 = vdup_n_s8(r0s[0]);
-                        int8x8_t _r1 = vdup_n_s8(r1s[0]);
-                        int8x8_t _w = vld1_s8(kptr);
-                        int16x8_t _s0 = vmull_s8(_r0, _w);
-                        int16x8_t _s1 = vmull_s8(_r1, _w);
-                        _sum0 = vaddw_s16(_sum0, vget_low_s16(_s0));
-                        _sum1 = vaddw_s16(_sum1, vget_high_s16(_s0));
-                        _sum2 = vaddw_s16(_sum2, vget_low_s16(_s1));
-                        _sum3 = vaddw_s16(_sum3, vget_high_s16(_s1));
-
-                        kptr += 8;
-                    }
-                }
-            }
-
-            if (out_elempack == 8)
-            {
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                vst1q_s32(outptr + 8, _sum2);
-                vst1q_s32(outptr + 12, _sum3);
-                outptr += 16;
-            }
-            if (out_elempack == 4)
-            {
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum2);
-                vst1q_s32(outptr + M, _sum1);
-                vst1q_s32(outptr + M + 4, _sum3);
-                outptr += 8;
-            }
-            if (out_elempack == 1)
-            {
-                outptr[0] = vgetq_lane_s32(_sum0, 0);
-                outptr[1] = vgetq_lane_s32(_sum2, 0);
-                outptr[M] = vgetq_lane_s32(_sum0, 1);
-                outptr[M + 1] = vgetq_lane_s32(_sum2, 1);
-                outptr[M * 2] = vgetq_lane_s32(_sum0, 2);
-                outptr[M * 2 + 1] = vgetq_lane_s32(_sum2, 2);
-                outptr[M * 3] = vgetq_lane_s32(_sum0, 3);
-                outptr[M * 3 + 1] = vgetq_lane_s32(_sum2, 3);
-                outptr[M * 4] = vgetq_lane_s32(_sum1, 0);
-                outptr[M * 4 + 1] = vgetq_lane_s32(_sum3, 0);
-                outptr[M * 5] = vgetq_lane_s32(_sum1, 1);
-                outptr[M * 5 + 1] = vgetq_lane_s32(_sum3, 1);
-                outptr[M * 6] = vgetq_lane_s32(_sum1, 2);
-                outptr[M * 6 + 1] = vgetq_lane_s32(_sum3, 2);
-                outptr[M * 7] = vgetq_lane_s32(_sum1, 3);
-                outptr[M * 7 + 1] = vgetq_lane_s32(_sum3, 3);
-                outptr += 2;
-            }
-        }
-        for (; ij < outw * outh; ij++)
-        {
-            const int i = ij / outw;
-            const int j = ij % outw;
-
-            int32x4_t _sum0 = vdupq_n_s32(0);
-            int32x4_t _sum1 = vdupq_n_s32(0);
-            int32x4_t _sum2 = vdupq_n_s32(0);
-            int32x4_t _sum3 = vdupq_n_s32(0);
-
-            const signed char* kptr = weight_data_tm.channel(p / 8);
-
-            int q = 0;
-            {
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-
-                        int8x8_t _r0;
-                        if (elempack == 8)
-                        {
-                            _r0 = vld1_s8(r0s);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            signed char tmp[8] = {r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7]};
-                            _r0 = vld1_s8(tmp);
-                        }
-
-                        int8x16_t _w0 = vld1q_s8(kptr);
-                        int8x16_t _w1 = vld1q_s8(kptr + 16);
-                        int8x16_t _w2 = vld1q_s8(kptr + 32);
-                        int8x16_t _w3 = vld1q_s8(kptr + 48);
-
-#if __ARM_FEATURE_MATMUL_INT8
-                        int8x16_t _r00 = vcombine_s8(_r0, _r0);
-                        _sum0 = vdotq_s32(_sum0, _r00, _w0);
-                        _sum1 = vdotq_s32(_sum1, _r00, _w1);
-                        _sum2 = vdotq_s32(_sum2, _r00, _w2);
-                        _sum3 = vdotq_s32(_sum3, _r00, _w3);
-#elif __ARM_FEATURE_DOTPROD
-                        _sum0 = vdotq_lane_s32(_sum0, _w0, _r0, 0);
-                        _sum1 = vdotq_lane_s32(_sum1, _w1, _r0, 0);
-                        _sum2 = vdotq_lane_s32(_sum2, _w2, _r0, 1);
-                        _sum3 = vdotq_lane_s32(_sum3, _w3, _r0, 1);
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                        int16x4_t _rr0 = vreinterpret_s16_s8(_r0);
-                        int8x8_t _r0ll = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 0));
-                        int8x8_t _r0lh = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 1));
-                        int8x8_t _r0hl = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 2));
-                        int8x8_t _r0hh = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 3));
-
-                        int16x8_t _s0l = vmull_s8(_r0ll, vget_low_s8(_w0));
-                        int16x8_t _s1l = vmull_s8(_r0ll, vget_high_s8(_w0));
-                        int16x8_t _s0h = vmull_s8(_r0lh, vget_low_s8(_w1));
-                        int16x8_t _s1h = vmull_s8(_r0lh, vget_high_s8(_w1));
-                        _s0l = vmlal_s8(_s0l, _r0hl, vget_low_s8(_w2));
-                        _s1l = vmlal_s8(_s1l, _r0hl, vget_high_s8(_w2));
-                        _s0h = vmlal_s8(_s0h, _r0hh, vget_low_s8(_w3));
-                        _s1h = vmlal_s8(_s1h, _r0hh, vget_high_s8(_w3));
-
-                        _sum0 = vpadalq_s16(_sum0, _s0l);
-                        _sum1 = vpadalq_s16(_sum1, _s1l);
-                        _sum2 = vpadalq_s16(_sum2, _s0h);
-                        _sum3 = vpadalq_s16(_sum3, _s1h);
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-
-                        kptr += 64;
-                    }
-                }
-#if __ARM_FEATURE_MATMUL_INT8
-                {
-                    _sum0 = vpaddq_s32(_sum0, _sum1);
-                    _sum1 = vpaddq_s32(_sum2, _sum3);
-                }
-#else
-                {
-                    _sum0 = vaddq_s32(_sum0, _sum2);
-                    _sum1 = vaddq_s32(_sum1, _sum3);
-                }
-#endif
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i * stride_h) + j * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        int8x8_t _val = vdup_n_s8(r0s[0]);
-                        int8x8_t _w = vld1_s8(kptr);
-                        int16x8_t _s0 = vmull_s8(_val, _w);
-                        _sum0 = vaddw_s16(_sum0, vget_low_s16(_s0));
-                        _sum1 = vaddw_s16(_sum1, vget_high_s16(_s0));
-
-                        kptr += 8;
-                    }
-                }
-            }
-
-            if (out_elempack == 8)
-            {
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                outptr += 8;
-            }
-            if (out_elempack == 4)
-            {
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + M, _sum1);
-                outptr += 4;
-            }
-            if (out_elempack == 1)
-            {
-                outptr[0] = vgetq_lane_s32(_sum0, 0);
-                outptr[M] = vgetq_lane_s32(_sum0, 1);
-                outptr[M * 2] = vgetq_lane_s32(_sum0, 2);
-                outptr[M * 3] = vgetq_lane_s32(_sum0, 3);
-                outptr[M * 4] = vgetq_lane_s32(_sum1, 0);
-                outptr[M * 5] = vgetq_lane_s32(_sum1, 1);
-                outptr[M * 6] = vgetq_lane_s32(_sum1, 2);
-                outptr[M * 7] = vgetq_lane_s32(_sum1, 3);
-                outptr += 1;
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 8;
-    nn_outch = (outch - remain_outch_start) / 4;
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 4;
-
-        // shadowed variable for less openmp task args
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int N = bottom_blob.cstep * elempack;
-        const int M = top_blob.cstep * out_elempack;
-
-        int* outptr = top_blob.channel(p / out_elempack);
-
-        int ij = 0;
-        for (; ij + 1 < outw * outh; ij += 2)
-        {
-            const int i0 = ij / outw;
-            const int i1 = (ij + 1) / outw;
-            const int j0 = ij % outw;
-            const int j1 = (ij + 1) % outw;
-
-            int32x4_t _sum0 = vdupq_n_s32(0);
-            int32x4_t _sum1 = vdupq_n_s32(0);
-
-            const signed char* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4);
-
-            int q = 0;
-            {
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-
-                        int8x8_t _r0;
-                        int8x8_t _r1;
-                        if (elempack == 8)
-                        {
-                            _r0 = vld1_s8(r0s);
-                            _r1 = vld1_s8(r1s);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            signed char tmp0[8] = {r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7]};
-                            signed char tmp1[8] = {r1s[0], r1s[N], r1s[N * 2], r1s[N * 3], r1s[N * 4], r1s[N * 5], r1s[N * 6], r1s[N * 7]};
-                            _r0 = vld1_s8(tmp0);
-                            _r1 = vld1_s8(tmp1);
-                        }
-
-                        int8x16_t _w0 = vld1q_s8(kptr);
-                        int8x16_t _w1 = vld1q_s8(kptr + 16);
-
-#if __ARM_FEATURE_MATMUL_INT8
-                        int8x16_t _r01 = vcombine_s8(_r0, _r1);
-                        _sum0 = vmmlaq_s32(_sum0, _r01, _w0);
-                        _sum1 = vmmlaq_s32(_sum1, _r01, _w1);
-#elif __ARM_FEATURE_DOTPROD
-                        _sum0 = vdotq_lane_s32(_sum0, _w0, _r0, 0);
-                        _sum1 = vdotq_lane_s32(_sum1, _w0, _r1, 0);
-                        _sum0 = vdotq_lane_s32(_sum0, _w1, _r0, 1);
-                        _sum1 = vdotq_lane_s32(_sum1, _w1, _r1, 1);
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                        int16x4_t _rr0 = vreinterpret_s16_s8(_r0);
-                        int16x4_t _rr1 = vreinterpret_s16_s8(_r1);
-
-                        int8x8_t _r0ll = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 0));
-                        int8x8_t _r1ll = vreinterpret_s8_s16(vdup_lane_s16(_rr1, 0));
-                        int8x8_t _r0lh = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 1));
-                        int8x8_t _r1lh = vreinterpret_s8_s16(vdup_lane_s16(_rr1, 1));
-
-                        int16x8_t _s0l = vmull_s8(_r0ll, vget_low_s8(_w0));
-                        int16x8_t _s1l = vmull_s8(_r1ll, vget_low_s8(_w0));
-                        int16x8_t _s0h = vmull_s8(_r0lh, vget_high_s8(_w0));
-                        int16x8_t _s1h = vmull_s8(_r1lh, vget_high_s8(_w0));
-
-                        int8x8_t _r0hl = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 2));
-                        int8x8_t _r1hl = vreinterpret_s8_s16(vdup_lane_s16(_rr1, 2));
-                        int8x8_t _r0hh = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 3));
-                        int8x8_t _r1hh = vreinterpret_s8_s16(vdup_lane_s16(_rr1, 3));
-
-                        _s0l = vmlal_s8(_s0l, _r0hl, vget_low_s8(_w1));
-                        _s1l = vmlal_s8(_s1l, _r1hl, vget_low_s8(_w1));
-                        _s0h = vmlal_s8(_s0h, _r0hh, vget_high_s8(_w1));
-                        _s1h = vmlal_s8(_s1h, _r1hh, vget_high_s8(_w1));
-
-                        _sum0 = vpadalq_s16(_sum0, _s0l);
-                        _sum1 = vpadalq_s16(_sum1, _s1l);
-                        _sum0 = vpadalq_s16(_sum0, _s0h);
-                        _sum1 = vpadalq_s16(_sum1, _s1h);
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-
-                        kptr += 32;
-                    }
-                }
-#if __ARM_FEATURE_MATMUL_INT8
-                {
-                    int32x4_t _tmp0 = vcombine_s32(vget_low_s32(_sum0), vget_low_s32(_sum1));
-                    int32x4_t _tmp1 = vcombine_s32(vget_high_s32(_sum0), vget_high_s32(_sum1));
-                    _sum0 = _tmp0;
-                    _sum1 = _tmp1;
-                }
-#endif
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        int8x8_t _r0 = vdup_n_s8(r0s[0]);
-                        int8x8_t _r1 = vdup_n_s8(r1s[0]);
-                        int8x8_t _r01 = vreinterpret_s8_s32(vzip_s32(vreinterpret_s32_s8(_r0), vreinterpret_s32_s8(_r1)).val[0]);
-                        int8x8_t _w = vld1_s8(kptr);
-                        int8x8_t _ww = vreinterpret_s8_s32(vzip_s32(vreinterpret_s32_s8(_w), vreinterpret_s32_s8(_w)).val[0]);
-                        int16x8_t _s01 = vmull_s8(_r01, _ww);
-                        _sum0 = vaddw_s16(_sum0, vget_low_s16(_s01));
-                        _sum1 = vaddw_s16(_sum1, vget_high_s16(_s01));
-
-                        kptr += 4;
-                    }
-                }
-            }
-
-            if (out_elempack == 4)
-            {
-                vst1q_s32(outptr, _sum0);
-                vst1q_s32(outptr + 4, _sum1);
-                outptr += 8;
-            }
-            if (out_elempack == 1)
-            {
-                int32x4x2_t _sum01 = vzipq_s32(_sum0, _sum1);
-                vst1_s32(outptr, vget_low_s32(_sum01.val[0]));
-                vst1_s32(outptr + M, vget_high_s32(_sum01.val[0]));
-                vst1_s32(outptr + M * 2, vget_low_s32(_sum01.val[1]));
-                vst1_s32(outptr + M * 3, vget_high_s32(_sum01.val[1]));
-                outptr += 2;
-            }
-        }
-        for (; ij < outw * outh; ij++)
-        {
-            const int i = ij / outw;
-            const int j = ij % outw;
-
-            int32x4_t _sum0 = vdupq_n_s32(0);
-            int32x4_t _sum1 = vdupq_n_s32(0);
-
-            const signed char* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4);
-
-            int q = 0;
-            {
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-
-                        int8x8_t _r0;
-                        if (elempack == 8)
-                        {
-                            _r0 = vld1_s8(r0s);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            signed char tmp[8] = {r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7]};
-                            _r0 = vld1_s8(tmp);
-                        }
-
-                        int8x16_t _w0 = vld1q_s8(kptr);
-                        int8x16_t _w1 = vld1q_s8(kptr + 16);
-
-#if __ARM_FEATURE_MATMUL_INT8
-                        int8x16_t _r00 = vcombine_s8(_r0, _r0);
-                        _sum0 = vdotq_s32(_sum0, _r00, _w0);
-                        _sum1 = vdotq_s32(_sum1, _r00, _w1);
-#elif __ARM_FEATURE_DOTPROD
-                        _sum0 = vdotq_lane_s32(_sum0, _w0, _r0, 0);
-                        _sum1 = vdotq_lane_s32(_sum1, _w1, _r0, 1);
-#else  // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-                        int16x4_t _rr0 = vreinterpret_s16_s8(_r0);
-                        int8x8_t _r0ll = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 0));
-                        int8x8_t _r0lh = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 1));
-                        int8x8_t _r0hl = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 2));
-                        int8x8_t _r0hh = vreinterpret_s8_s16(vdup_lane_s16(_rr0, 3));
-
-                        int16x8_t _sl = vmull_s8(_r0ll, vget_low_s8(_w0));
-                        int16x8_t _sh = vmull_s8(_r0lh, vget_high_s8(_w0));
-                        _sl = vmlal_s8(_sl, _r0hl, vget_low_s8(_w1));
-                        _sh = vmlal_s8(_sh, _r0hh, vget_high_s8(_w1));
-
-                        _sum0 = vpadalq_s16(_sum0, _sl);
-                        _sum1 = vpadalq_s16(_sum1, _sh);
-#endif // __ARM_FEATURE_MATMUL_INT8 || __ARM_FEATURE_DOTPROD
-
-                        kptr += 32;
-                    }
-                }
-#if __ARM_FEATURE_MATMUL_INT8
-                {
-                    _sum0 = vpaddq_s32(_sum0, _sum1);
-                }
-#else
-                {
-                    _sum0 = vaddq_s32(_sum0, _sum1);
-                }
-#endif
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i * stride_h) + j * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        int8x8_t _val = vdup_n_s8(r0s[0]);
-                        int8x8_t _w = vld1_s8(kptr);
-                        int16x8_t _s0 = vmull_s8(_val, _w);
-                        _sum0 = vaddw_s16(_sum0, vget_low_s16(_s0));
-
-                        kptr += 4;
-                    }
-                }
-            }
-
-            if (out_elempack == 4)
-            {
-                vst1q_s32(outptr, _sum0);
-                outptr += 4;
-            }
-            if (out_elempack == 1)
-            {
-                outptr[0] = vgetq_lane_s32(_sum0, 0);
-                outptr[M] = vgetq_lane_s32(_sum0, 1);
-                outptr[M * 2] = vgetq_lane_s32(_sum0, 2);
-                outptr[M * 3] = vgetq_lane_s32(_sum0, 3);
-                outptr += 1;
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 4;
-    nn_outch = (outch - remain_outch_start) / 2;
-#else // __ARM_NEON
-    nn_outch = (outch - remain_outch_start) / 2;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __ARM_NEON
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 2;
-
-        // shadowed variable for less openmp task args
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int N = bottom_blob.cstep * elempack;
-
-        int* outptr0 = top_blob.channel(p);
-        int* outptr1 = top_blob.channel(p + 1);
-
-        int ij = 0;
-        for (; ij + 1 < outw * outh; ij += 2)
-        {
-            const int i0 = ij / outw;
-            const int i1 = (ij + 1) / outw;
-            const int j0 = ij % outw;
-            const int j1 = (ij + 1) % outw;
-
-            int sum00 = 0;
-            int sum01 = 0;
-            int sum10 = 0;
-            int sum11 = 0;
-
-#if __ARM_NEON
-            const signed char* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 2);
-#endif
-
-            int q = 0;
-#if __ARM_NEON
-            {
-                int32x4_t _sum01 = vdupq_n_s32(0);
-                int32x4_t _sum23 = vdupq_n_s32(0);
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-
-                        int8x8_t _r0;
-                        int8x8_t _r1;
-                        if (elempack == 8)
-                        {
-                            _r0 = vld1_s8(r0s);
-                            _r1 = vld1_s8(r1s);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            signed char tmp0[8] = {r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7]};
-                            signed char tmp1[8] = {r1s[0], r1s[N], r1s[N * 2], r1s[N * 3], r1s[N * 4], r1s[N * 5], r1s[N * 6], r1s[N * 7]};
-                            _r0 = vld1_s8(tmp0);
-                            _r1 = vld1_s8(tmp1);
-                        }
-
-                        int8x16_t _w0 = vld1q_s8(kptr);
-
-#if __ARM_FEATURE_DOTPROD
-                        int8x16_t _r00 = vcombine_s8(_r0, _r0);
-                        int8x16_t _r11 = vcombine_s8(_r1, _r1);
-                        _sum01 = vdotq_s32(_sum01, _r00, _w0);
-                        _sum23 = vdotq_s32(_sum23, _r11, _w0);
-#else  // __ARM_FEATURE_DOTPROD
-                        int32x2x2_t _rr0 = vzip_s32(vreinterpret_s32_s8(_r0), vreinterpret_s32_s8(_r0));
-                        int32x2x2_t _rr1 = vzip_s32(vreinterpret_s32_s8(_r1), vreinterpret_s32_s8(_r1));
-                        int8x8_t _r0l = vreinterpret_s8_s32(_rr0.val[0]);
-                        int8x8_t _r0h = vreinterpret_s8_s32(_rr0.val[1]);
-                        int8x8_t _r1l = vreinterpret_s8_s32(_rr1.val[0]);
-                        int8x8_t _r1h = vreinterpret_s8_s32(_rr1.val[1]);
-
-                        int16x8_t _s01 = vmull_s8(_r0l, vget_low_s8(_w0));
-                        int16x8_t _s23 = vmull_s8(_r1l, vget_low_s8(_w0));
-                        _s01 = vmlal_s8(_s01, _r0h, vget_high_s8(_w0));
-                        _s23 = vmlal_s8(_s23, _r1h, vget_high_s8(_w0));
-
-                        _sum01 = vpadalq_s16(_sum01, _s01);
-                        _sum23 = vpadalq_s16(_sum23, _s23);
-#endif // __ARM_FEATURE_DOTPROD
-
-                        kptr += 16;
-                    }
-                }
-                int32x2_t _s0 = vpadd_s32(vget_low_s32(_sum01), vget_high_s32(_sum01));
-                int32x2_t _s1 = vpadd_s32(vget_low_s32(_sum23), vget_high_s32(_sum23));
-                sum00 += vget_lane_s32(_s0, 0);
-                sum01 += vget_lane_s32(_s1, 0);
-                sum10 += vget_lane_s32(_s0, 1);
-                sum11 += vget_lane_s32(_s1, 1);
-            }
-#endif // __ARM_NEON
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        sum00 += r0s[0] * kptr[0];
-                        sum01 += r1s[0] * kptr[0];
-                        sum10 += r0s[0] * kptr[1];
-                        sum11 += r1s[0] * kptr[1];
-
-                        kptr += 2;
-                    }
-                }
-            }
-
-            outptr0[0] = sum00;
-            outptr0[1] = sum01;
-            outptr1[0] = sum10;
-            outptr1[1] = sum11;
-            outptr0 += 2;
-            outptr1 += 2;
-        }
-        for (; ij < outw * outh; ij++)
-        {
-            const int i = ij / outw;
-            const int j = ij % outw;
-
-            int sum0 = 0;
-            int sum1 = 0;
-
-#if __ARM_NEON
-            const signed char* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 2);
-#endif
-
-            int q = 0;
-#if __ARM_NEON
-            {
-                int32x4_t _sum01 = vdupq_n_s32(0);
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-
-                        int8x8_t _r0;
-                        if (elempack == 8)
-                        {
-                            _r0 = vld1_s8(r0s);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            signed char tmp[8] = {r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7]};
-                            _r0 = vld1_s8(tmp);
-                        }
-
-                        int8x16_t _w0 = vld1q_s8(kptr);
-
-#if __ARM_FEATURE_DOTPROD
-                        int8x16_t _r00 = vcombine_s8(_r0, _r0);
-                        _sum01 = vdotq_s32(_sum01, _r00, _w0);
-#else  // __ARM_FEATURE_DOTPROD
-                        int32x2x2_t _rr0 = vzip_s32(vreinterpret_s32_s8(_r0), vreinterpret_s32_s8(_r0));
-                        int8x8_t _r0l = vreinterpret_s8_s32(_rr0.val[0]);
-                        int8x8_t _r0h = vreinterpret_s8_s32(_rr0.val[1]);
-
-                        int16x8_t _s01 = vmull_s8(_r0l, vget_low_s8(_w0));
-                        _s01 = vmlal_s8(_s01, _r0h, vget_high_s8(_w0));
-
-                        _sum01 = vpadalq_s16(_sum01, _s01);
-#endif // __ARM_FEATURE_DOTPROD
-
-                        kptr += 16;
-                    }
-                }
-                int32x2_t _s0 = vpadd_s32(vget_low_s32(_sum01), vget_high_s32(_sum01));
-                sum0 += vget_lane_s32(_s0, 0);
-                sum1 += vget_lane_s32(_s0, 1);
-            }
-#endif // __ARM_NEON
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i * stride_h) + j * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        sum0 += r0s[0] * kptr[0];
-                        sum1 += r0s[0] * kptr[1];
-
-                        kptr += 2;
-                    }
-                }
-            }
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-            outptr0 += 1;
-            outptr1 += 1;
-        }
-    }
-    remain_outch_start += nn_outch * 2;
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        int* outptr = top_blob.channel(p);
-
-        int ij = 0;
-        for (; ij + 1 < outw * outh; ij += 2)
-        {
-            const int i0 = ij / outw;
-            const int i1 = (ij + 1) / outw;
-            const int j0 = ij % outw;
-            const int j1 = (ij + 1) % outw;
-
-            int sum0 = 0;
-            int sum1 = 0;
-
-#if __ARM_NEON
-            const signed char* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 2 + p % 2);
-#endif
-
-            int q = 0;
-#if __ARM_NEON
-            {
-                int32x4_t _sum0 = vdupq_n_s32(0);
-                int32x4_t _sum1 = vdupq_n_s32(0);
-                int32x4_t _sum2 = vdupq_n_s32(0);
-                int32x4_t _sum3 = vdupq_n_s32(0);
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-
-                        int8x8_t _r0;
-                        int8x8_t _r1;
-                        if (elempack == 8)
-                        {
-                            _r0 = vld1_s8(r0s);
-                            _r1 = vld1_s8(r1s);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            signed char tmp0[8] = {r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7]};
-                            signed char tmp1[8] = {r1s[0], r1s[N], r1s[N * 2], r1s[N * 3], r1s[N * 4], r1s[N * 5], r1s[N * 6], r1s[N * 7]};
-                            _r0 = vld1_s8(tmp0);
-                            _r1 = vld1_s8(tmp1);
-                        }
-
-                        int8x8_t _w = vld1_s8(kptr);
-
-                        int16x8_t _s0 = vmull_s8(_r0, _w);
-                        int16x8_t _s1 = vmull_s8(_r1, _w);
-
-                        _sum0 = vaddw_s16(_sum0, vget_low_s16(_s0));
-                        _sum1 = vaddw_s16(_sum1, vget_high_s16(_s0));
-                        _sum2 = vaddw_s16(_sum2, vget_low_s16(_s1));
-                        _sum3 = vaddw_s16(_sum3, vget_high_s16(_s1));
-
-                        kptr += 8;
-                    }
-                }
-                _sum0 = vaddq_s32(_sum0, _sum1);
-                _sum2 = vaddq_s32(_sum2, _sum3);
-#if __aarch64__
-                sum0 += vaddvq_s32(_sum0);
-                sum1 += vaddvq_s32(_sum2);
-#else
-                int32x2_t _ss0 = vadd_s32(vget_low_s32(_sum0), vget_high_s32(_sum0));
-                int32x2_t _ss2 = vadd_s32(vget_low_s32(_sum2), vget_high_s32(_sum2));
-                _ss0 = vpadd_s32(_ss0, _ss2);
-                sum0 += vget_lane_s32(_ss0, 0);
-                sum1 += vget_lane_s32(_ss0, 1);
-#endif
-            }
-#endif // __ARM_NEON
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        sum0 += r0s[0] * kptr[0];
-                        sum1 += r1s[0] * kptr[0];
-
-                        kptr += 1;
-                    }
-                }
-            }
-
-            outptr[0] = sum0;
-            outptr[1] = sum1;
-            outptr += 2;
-        }
-        for (; ij < outw * outh; ij++)
-        {
-            const int i = ij / outw;
-            const int j = ij % outw;
-
-            int sum = 0;
-
-#if __ARM_NEON
-            const signed char* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 2 + p % 2);
-#endif
-
-            int q = 0;
-#if __ARM_NEON
-            {
-                int32x4_t _sum0 = vdupq_n_s32(0);
-                int32x4_t _sum1 = vdupq_n_s32(0);
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-
-                        int8x8_t _r0;
-                        if (elempack == 8)
-                        {
-                            _r0 = vld1_s8(r0s);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            signed char tmp[8] = {r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7]};
-                            _r0 = vld1_s8(tmp);
-                        }
-
-                        int8x8_t _w = vld1_s8(kptr);
-
-                        int16x8_t _s0 = vmull_s8(_r0, _w);
-
-                        _sum0 = vaddw_s16(_sum0, vget_low_s16(_s0));
-                        _sum1 = vaddw_s16(_sum1, vget_high_s16(_s0));
-
-                        kptr += 8;
-                    }
-                }
-                int32x4_t _sum = vaddq_s32(_sum0, _sum1);
-#if __aarch64__
-                sum += vaddvq_s32(_sum);
-#else
-                int32x2_t _ss = vadd_s32(vget_low_s32(_sum), vget_high_s32(_sum));
-                _ss = vpadd_s32(_ss, _ss);
-                sum += vget_lane_s32(_ss, 0);
-#endif
-            }
-#endif // __ARM_NEON
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i * stride_h) + j * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        sum += r0s[0] * kptr[0];
-
-                        kptr += 1;
-                    }
-                }
-            }
-
-            outptr[0] = sum;
-            outptr += 1;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3.h
deleted file mode 100644
index 8260d80e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3.h
+++ /dev/null
@@ -1,1076 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const float bias0 = bias ? bias[g] : 0.f;
-
-        const float* kernel0 = kernel + g * 9;
-
-        float* outptr = out;
-        float* outptr2 = outptr + outw;
-
-        const float* img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0;
-        const float* r1 = img0 + w;
-        const float* r2 = img0 + w * 2;
-        const float* r3 = img0 + w * 3;
-
-#if __ARM_NEON
-        float32x4_t _k012x = vld1q_f32(kernel0);
-        float32x4_t _k345x = vld1q_f32(kernel0 + 3);
-        float32x4_t _k678x = vld1q_f32(kernel0 + 6);
-
-        _k012x = vsetq_lane_f32(0.f, _k012x, 3);
-        _k345x = vsetq_lane_f32(0.f, _k345x, 3);
-        _k678x = vsetq_lane_f32(0.f, _k678x, 3);
-
-        float32x4_t _bias0 = vdupq_n_f32(bias0);
-#else
-        const float* k0 = kernel0;
-        const float* k1 = kernel0 + 3;
-        const float* k2 = kernel0 + 6;
-#endif // __ARM_NEON
-
-        int i = 0;
-
-        for (; i + 1 < outh; i += 2)
-        {
-#if __ARM_NEON
-#if __aarch64__
-            int nn = outw >> 3;
-            int remain = outw & 7;
-#else
-            int nn = outw >> 2;
-            int remain = outw & 3;
-#endif // __aarch64__
-#else
-            int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%3, #384]           \n"
-                    "ld1    {v8.4s, v9.4s, v10.4s}, [%3]    \n" // r0
-                    "add    %3, %3, #32                     \n"
-
-                    "ext    v11.16b, v8.16b, v9.16b, #4     \n"
-                    "ext    v13.16b, v9.16b, v10.16b, #4    \n"
-
-                    "ext    v12.16b, v8.16b, v9.16b, #8     \n"
-                    "ext    v14.16b, v9.16b, v10.16b, #8    \n"
-
-                    "0:                                     \n"
-
-                    "and    v4.16b, %17.16b, %17.16b        \n" // v4 = _bias0
-                    "and    v5.16b, %17.16b, %17.16b        \n" // v5 = _bias0
-
-                    "prfm   pldl1keep, [%6, #384]           \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s}, [%6]  \n" // r3
-                    "add    %6, %6, #32                     \n"
-
-                    "and    v6.16b, %17.16b, %17.16b        \n" // v6 = _bias0
-                    "and    v7.16b, %17.16b, %17.16b        \n" // v7 = _bias0
-
-                    "ext    v15.16b, v16.16b, v17.16b, #4   \n"
-
-                    "fmla   v4.4s, v8.4s, %14.s[0]          \n"
-                    "fmla   v5.4s, v9.4s, %14.s[0]          \n"
-
-                    "ext    v20.16b, v17.16b, v18.16b, #4   \n"
-
-                    "fmla   v6.4s, v16.4s, %16.s[0]         \n"
-                    "fmla   v7.4s, v17.4s, %16.s[0]         \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #8   \n"
-
-                    "fmla   v4.4s, v11.4s, %14.s[1]         \n"
-                    "fmla   v5.4s, v13.4s, %14.s[1]         \n"
-
-                    "ext    v21.16b, v17.16b, v18.16b, #8   \n"
-
-                    "fmla   v6.4s, v15.4s, %16.s[1]         \n"
-                    "fmla   v7.4s, v20.4s, %16.s[1]         \n"
-
-                    "prfm   pldl1keep, [%4, #384]           \n"
-                    "ld1    {v22.4s, v23.4s, v24.4s}, [%4]  \n" // r1
-
-                    "fmla   v4.4s, v12.4s, %14.s[2]         \n"
-                    "fmla   v5.4s, v14.4s, %14.s[2]         \n"
-
-                    "add    %4, %4, #32                     \n"
-
-                    "fmla   v6.4s, v19.4s, %16.s[2]         \n"
-                    "fmla   v7.4s, v21.4s, %16.s[2]         \n"
-
-                    "ext    v25.16b, v22.16b, v23.16b, #4   \n"
-
-                    "fmla   v4.4s, v22.4s, %15.s[0]         \n"
-                    "fmla   v5.4s, v23.4s, %15.s[0]         \n"
-
-                    "ext    v27.16b, v23.16b, v24.16b, #4   \n"
-
-                    "fmla   v6.4s, v22.4s, %14.s[0]         \n"
-                    "fmla   v7.4s, v23.4s, %14.s[0]         \n"
-
-                    "ext    v26.16b, v22.16b, v23.16b, #8   \n"
-
-                    "fmla   v4.4s, v25.4s, %15.s[1]         \n"
-                    "fmla   v5.4s, v27.4s, %15.s[1]         \n"
-
-                    "ext    v28.16b, v23.16b, v24.16b, #8   \n"
-
-                    "fmla   v6.4s, v25.4s, %14.s[1]         \n"
-                    "fmla   v7.4s, v27.4s, %14.s[1]         \n"
-
-                    "prfm   pldl1keep, [%5, #384]           \n"
-                    "ld1    {v8.4s, v9.4s, v10.4s}, [%5]    \n" // r2
-
-                    "fmla   v4.4s, v26.4s, %15.s[2]         \n"
-                    "fmla   v5.4s, v28.4s, %15.s[2]         \n"
-
-                    "add    %5, %5, #32                     \n"
-
-                    "fmla   v6.4s, v26.4s, %14.s[2]         \n"
-                    "fmla   v7.4s, v28.4s, %14.s[2]         \n"
-
-                    "ext    v11.16b, v8.16b, v9.16b, #4     \n"
-
-                    "fmla   v4.4s, v8.4s, %16.s[0]          \n"
-                    "fmla   v5.4s, v9.4s, %16.s[0]          \n"
-
-                    "ext    v13.16b, v9.16b, v10.16b, #4    \n"
-
-                    "fmla   v6.4s, v8.4s, %15.s[0]          \n"
-                    "fmla   v7.4s, v9.4s, %15.s[0]          \n"
-
-                    "ext    v12.16b, v8.16b, v9.16b, #8     \n"
-
-                    "fmla   v4.4s, v11.4s, %16.s[1]         \n"
-                    "fmla   v5.4s, v13.4s, %16.s[1]         \n"
-
-                    "ext    v14.16b, v9.16b, v10.16b, #8    \n"
-
-                    "fmla   v6.4s, v11.4s, %15.s[1]         \n"
-                    "fmla   v7.4s, v13.4s, %15.s[1]         \n"
-
-                    "prfm   pldl1keep, [%3, #384]           \n"
-                    "ld1    {v8.4s, v9.4s, v10.4s}, [%3]    \n" // r0 next loop
-
-                    "fmla   v4.4s, v12.4s, %16.s[2]         \n"
-                    "fmla   v5.4s, v14.4s, %16.s[2]         \n"
-
-                    "add    %3, %3, #32                     \n"
-                    "ext    v11.16b, v8.16b, v9.16b, #4     \n"
-
-                    "fmla   v6.4s, v12.4s, %15.s[2]         \n"
-                    "fmla   v7.4s, v14.4s, %15.s[2]         \n"
-
-                    "ext    v13.16b, v9.16b, v10.16b, #4    \n"
-                    "ext    v12.16b, v8.16b, v9.16b, #8     \n"
-
-                    "st1    {v4.4s, v5.4s}, [%1], #32       \n"
-
-                    "ext    v14.16b, v9.16b, v10.16b, #8    \n"
-
-                    "subs   %w0, %w0, #1                    \n"
-
-                    "st1    {v6.4s, v7.4s}, [%2], #32       \n"
-
-                    "bne    0b                              \n"
-                    "sub    %3, %3, #32                     \n"
-                    : "=r"(nn),      // %0
-                    "=r"(outptr),  // %1
-                    "=r"(outptr2), // %2
-                    "=r"(r0),      // %3
-                    "=r"(r1),      // %4
-                    "=r"(r2),      // %5
-                    "=r"(r3)       // %6
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(outptr2),
-                    "3"(r0),
-                    "4"(r1),
-                    "5"(r2),
-                    "6"(r3),
-                    "w"(_k012x), // %14
-                    "w"(_k345x), // %15
-                    "w"(_k678x), // %16
-                    "w"(_bias0)  // %17
-                    : "cc", "memory", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28");
-            }
-
-            if (remain >= 4)
-            {
-                remain -= 4;
-
-                asm volatile(
-                    "prfm   pldl1keep, [%2, #256]           \n"
-                    "ld1    {v8.4s, v9.4s}, [%2]            \n" // r0
-                    "add    %2, %2, #16                     \n"
-
-                    "and    v4.16b, %15.16b, %15.16b        \n" // v4 = _bias0
-                    "and    v6.16b, %15.16b, %15.16b        \n" // v6 = _bias0
-
-                    "prfm   pldl1keep, [%5, #256]           \n"
-                    "ld1    {v16.4s, v17.4s}, [%5]          \n" // r3
-                    "add    %5, %5, #16                     \n"
-
-                    "ext    v11.16b, v8.16b, v9.16b, #4     \n"
-                    "ext    v15.16b, v16.16b, v17.16b, #4   \n"
-
-                    "fmla   v4.4s, v8.4s, %12.s[0]          \n"
-                    "fmla   v6.4s, v16.4s, %14.s[0]         \n"
-
-                    "ext    v12.16b, v8.16b, v9.16b, #8     \n"
-                    "ext    v19.16b, v16.16b, v17.16b, #8   \n"
-
-                    "fmla   v4.4s, v11.4s, %12.s[1]         \n"
-                    "fmla   v6.4s, v15.4s, %14.s[1]         \n"
-
-                    "prfm   pldl1keep, [%3, #256]           \n"
-                    "ld1    {v22.4s, v23.4s}, [%3]          \n" // r1
-
-                    "fmla   v4.4s, v12.4s, %12.s[2]         \n"
-
-                    "add    %3, %3, #16                     \n"
-
-                    "fmla   v6.4s, v19.4s, %14.s[2]         \n"
-
-                    "ext    v25.16b, v22.16b, v23.16b, #4   \n"
-
-                    "fmla   v4.4s, v22.4s, %13.s[0]         \n"
-                    "fmla   v6.4s, v22.4s, %12.s[0]         \n"
-
-                    "ext    v26.16b, v22.16b, v23.16b, #8   \n"
-
-                    "fmla   v4.4s, v25.4s, %13.s[1]         \n"
-                    "fmla   v6.4s, v25.4s, %12.s[1]         \n"
-
-                    "prfm   pldl1keep, [%4, #256]           \n"
-                    "ld1    {v8.4s, v9.4s}, [%4]            \n" // r2
-
-                    "fmla   v4.4s, v26.4s, %13.s[2]         \n"
-
-                    "add    %4, %4, #16                     \n"
-
-                    "fmla   v6.4s, v26.4s, %12.s[2]         \n"
-
-                    "ext    v11.16b, v8.16b, v9.16b, #4     \n"
-
-                    "fmla   v4.4s, v8.4s, %14.s[0]          \n"
-                    "fmla   v6.4s, v8.4s, %13.s[0]          \n"
-
-                    "ext    v12.16b, v8.16b, v9.16b, #8     \n"
-
-                    "fmla   v4.4s, v11.4s, %14.s[1]         \n"
-                    "fmla   v6.4s, v11.4s, %13.s[1]         \n"
-
-                    "fmla   v4.4s, v12.4s, %14.s[2]         \n"
-                    "fmla   v6.4s, v12.4s, %13.s[2]         \n"
-
-                    "st1    {v4.4s}, [%0], #16              \n"
-                    "st1    {v6.4s}, [%1], #16              \n"
-
-                    : "=r"(outptr),  // %0
-                    "=r"(outptr2), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3)       // %5
-                    : "0"(outptr),
-                    "1"(outptr2),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "w"(_k012x), // %12
-                    "w"(_k345x), // %13
-                    "w"(_k678x), // %14
-                    "w"(_bias0)  // %15
-                    : "cc", "memory", "v4", "v6", "v8", "v9", "v11", "v12", "v15", "v16", "v17", "v18", "v19", "v22", "v23", "v25", "v26");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    "pld        [%3, #192]          \n"
-                    "vld1.f32   {d18-d20}, [%3 :64] \n" // r0
-                    "add        %3, #16             \n"
-
-                    "vext.32    q11, q9, q10, #1    \n"
-                    "vext.32    q12, q9, q10, #2    \n"
-
-                    "0:                             \n"
-
-                    "vmul.f32   q7, q9, %e14[0]     \n"
-
-                    "vand       q13, %q17, %q17     \n" // q13 = _bias0
-                    "vmul.f32   q6, q11, %e14[1]    \n"
-                    "vmla.f32   q13, q12, %f14[0]   \n"
-
-                    "pld        [%4, #192]          \n"
-                    "vld1.f32   {d18-d20}, [%4]     \n" // r1
-                    "add        %4, #16             \n"
-
-                    "vmla.f32   q7, q9, %e15[0]     \n"
-
-                    "vext.32    q11, q9, q10, #1    \n"
-                    "vext.32    q12, q9, q10, #2    \n"
-
-                    "vmla.f32   q6, q11, %e15[1]    \n"
-                    "vmla.f32   q13, q12, %f15[0]   \n"
-
-                    "vmul.f32   q8, q9, %e14[0]     \n"
-
-                    "vand       q15, %q17, %q17     \n" // q15 = _bias0
-                    "vmul.f32   q14, q11, %e14[1]   \n"
-                    "vmla.f32   q15, q12, %f14[0]   \n"
-
-                    "pld        [%5, #192]          \n"
-                    "vld1.f32   {d18-d20}, [%5 :64] \n" // r2
-                    "add        %5, #16             \n"
-
-                    "vmla.f32   q7, q9, %e16[0]     \n"
-
-                    "vext.32    q11, q9, q10, #1    \n"
-                    "vext.32    q12, q9, q10, #2    \n"
-
-                    "vmla.f32   q6, q11, %e16[1]    \n"
-                    "vmla.f32   q13, q12, %f16[0]   \n"
-
-                    "vmla.f32   q8, q9, %e15[0]     \n"
-                    "vmla.f32   q14, q11, %e15[1]   \n"
-                    "vmla.f32   q15, q12, %f15[0]   \n"
-
-                    "pld        [%6, #192]          \n"
-                    "vld1.f32   {d18-d20}, [%6]     \n" // r3
-                    "add        %6, #16             \n"
-
-                    "vmla.f32   q8, q9, %e16[0]     \n"
-
-                    "vext.32    q11, q9, q10, #1    \n"
-                    "vext.32    q12, q9, q10, #2    \n"
-
-                    "vmla.f32   q14, q11, %e16[1]   \n"
-                    "vmla.f32   q15, q12, %f16[0]   \n"
-
-                    "vadd.f32   q7, q7, q6          \n"
-
-                    "pld        [%3, #192]          \n"
-                    "vld1.f32   {d18-d20}, [%3 :64] \n" // r0
-
-                    "vadd.f32   q8, q8, q14         \n"
-                    "vadd.f32   q7, q7, q13         \n"
-                    "vadd.f32   q8, q8, q15         \n"
-
-                    "vext.32    q11, q9, q10, #1    \n"
-                    "vext.32    q12, q9, q10, #2    \n"
-
-                    "add        %3, #16             \n"
-
-                    "vst1.f32   {d14-d15}, [%1]!    \n"
-                    "vst1.f32   {d16-d17}, [%2]!    \n"
-
-                    "subs       %0, #1              \n"
-                    "bne        0b                  \n"
-
-                    "sub        %3, #16             \n"
-                    : "=r"(nn),      // %0
-                    "=r"(outptr),  // %1
-                    "=r"(outptr2), // %2
-                    "=r"(r0),      // %3
-                    "=r"(r1),      // %4
-                    "=r"(r2),      // %5
-                    "=r"(r3)       // %6
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(outptr2),
-                    "3"(r0),
-                    "4"(r1),
-                    "5"(r2),
-                    "6"(r3),
-                    "w"(_k012x), // %14
-                    "w"(_k345x), // %15
-                    "w"(_k678x), // %16
-                    "w"(_bias0)  // %17
-                    : "cc", "memory", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-#if __ARM_NEON
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r30 = vld1q_f32(r3);
-
-                float32x4_t _sum = vmulq_f32(_r00, _k012x);
-                _sum = vmlaq_f32(_sum, _r10, _k345x);
-                _sum = vmlaq_f32(_sum, _r20, _k678x);
-
-                float32x4_t _sum2 = vmulq_f32(_r10, _k012x);
-                _sum2 = vmlaq_f32(_sum2, _r20, _k345x);
-                _sum2 = vmlaq_f32(_sum2, _r30, _k678x);
-
-                _sum = vsetq_lane_f32(bias0, _sum, 3);
-                _sum2 = vsetq_lane_f32(bias0, _sum2, 3);
-#if __aarch64__
-                *outptr = vaddvq_f32(_sum);
-                *outptr2 = vaddvq_f32(_sum2);
-#else
-                float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-                float32x2_t _ss2 = vadd_f32(vget_low_f32(_sum2), vget_high_f32(_sum2));
-
-                float32x2_t _sss2 = vpadd_f32(_ss, _ss2);
-
-                *outptr = vget_lane_f32(_sss2, 0);
-                *outptr2 = vget_lane_f32(_sss2, 1);
-#endif // __aarch64__
-#else
-                float sum = bias0;
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-
-                float sum2 = bias0;
-                sum2 += r1[0] * k0[0];
-                sum2 += r1[1] * k0[1];
-                sum2 += r1[2] * k0[2];
-                sum2 += r2[0] * k1[0];
-                sum2 += r2[1] * k1[1];
-                sum2 += r2[2] * k1[2];
-                sum2 += r3[0] * k2[0];
-                sum2 += r3[1] * k2[1];
-                sum2 += r3[2] * k2[2];
-
-                *outptr = sum;
-                *outptr2 = sum2;
-#endif
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                outptr++;
-                outptr2++;
-            }
-
-            r0 += 2 + w;
-            r1 += 2 + w;
-            r2 += 2 + w;
-            r3 += 2 + w;
-
-            outptr += outw;
-            outptr2 += outw;
-        }
-
-        for (; i < outh; i++)
-        {
-#if __ARM_NEON
-#if __aarch64__
-            int nn = outw >> 3;
-            int remain = outw & 7;
-#else
-            int nn = outw >> 2;
-            int remain = outw & 3;
-#endif // __aarch64__
-#else
-            int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%2, #384]           \n"
-                    "ld1    {v8.4s, v9.4s, v10.4s}, [%2]    \n" // r0
-                    "add    %2, %2, #32                     \n"
-
-                    "ext    v12.16b, v8.16b, v9.16b, #4     \n"
-                    "ext    v14.16b, v9.16b, v10.16b, #4    \n"
-
-                    "0:                                     \n"
-
-                    "fmul   v6.4s, v8.4s, %10.s[0]          \n"
-
-                    "and    v4.16b, %13.16b, %13.16b        \n" // v4 = _bias0
-
-                    "fmul   v7.4s, v9.4s, %10.s[0]          \n"
-
-                    "and    v5.16b, %13.16b, %13.16b        \n" // v5 = _bias0
-
-                    "fmla   v4.4s, v12.4s, %10.s[1]         \n"
-
-                    "ext    v13.16b, v8.16b, v9.16b, #8     \n"
-
-                    "fmla   v5.4s, v14.4s, %10.s[1]         \n"
-
-                    "ext    v15.16b, v9.16b, v10.16b, #8    \n"
-
-                    "fmla   v6.4s, v13.4s, %10.s[2]         \n"
-
-                    "prfm   pldl1keep, [%3, #384]           \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s}, [%3]  \n" // r1
-
-                    "fmla   v7.4s, v15.4s, %10.s[2]         \n"
-
-                    "add    %3, %3, #32                     \n"
-
-                    "fmla   v4.4s, v16.4s, %11.s[0]         \n"
-
-                    "ext    v20.16b, v16.16b, v17.16b, #4   \n"
-
-                    "fmla   v5.4s, v17.4s, %11.s[0]         \n"
-
-                    "ext    v22.16b, v17.16b, v18.16b, #4   \n"
-
-                    "fmla   v6.4s, v20.4s, %11.s[1]         \n"
-
-                    "ext    v21.16b, v16.16b, v17.16b, #8   \n"
-
-                    "fmla   v7.4s, v22.4s, %11.s[1]         \n"
-
-                    "ext    v23.16b, v17.16b, v18.16b, #8   \n"
-
-                    "fmla   v4.4s, v21.4s, %11.s[2]         \n"
-
-                    "prfm   pldl1keep, [%4, #384]           \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s}, [%4]  \n" // r2
-
-                    "fmla   v5.4s, v23.4s, %11.s[2]         \n"
-
-                    "add    %4, %4, #32                     \n"
-
-                    "fmla   v6.4s, v24.4s, %12.s[0]         \n"
-
-                    "ext    v12.16b, v24.16b, v25.16b, #4   \n"
-
-                    "fmla   v7.4s, v25.4s, %12.s[0]         \n"
-
-                    "ext    v14.16b, v25.16b, v26.16b, #4   \n"
-
-                    "fmla   v4.4s, v12.4s, %12.s[1]         \n"
-
-                    "ext    v13.16b, v24.16b, v25.16b, #8   \n"
-
-                    "fmla   v5.4s, v14.4s, %12.s[1]         \n"
-
-                    "ext    v15.16b, v25.16b, v26.16b, #8   \n"
-
-                    "fmla   v6.4s, v13.4s, %12.s[2]         \n"
-                    "fmla   v7.4s, v15.4s, %12.s[2]         \n"
-
-                    "prfm   pldl1keep, [%2, #384]           \n"
-                    "ld1    {v8.4s, v9.4s, v10.4s}, [%2]    \n" // r0 next loop
-
-                    "fadd   v4.4s, v4.4s, v6.4s             \n"
-
-                    "add    %2, %2, #32                     \n"
-
-                    "fadd   v5.4s, v5.4s, v7.4s             \n"
-
-                    "ext    v12.16b, v8.16b, v9.16b, #4     \n"
-                    "ext    v14.16b, v9.16b, v10.16b, #4    \n"
-
-                    "subs   %w0, %w0, #1                    \n"
-
-                    "st1    {v4.4s, v5.4s}, [%1], #32       \n"
-
-                    "bne    0b                              \n"
-                    "sub    %2, %2, #32                     \n"
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0),     // %2
-                    "=r"(r1),     // %3
-                    "=r"(r2)      // %4
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "w"(_k012x), // %10
-                    "w"(_k345x), // %11
-                    "w"(_k678x), // %12
-                    "w"(_bias0)  // %13
-                    : "cc", "memory", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v20", "v21", "v22", "v23", "v24", "v25", "v26");
-            }
-
-            if (remain >= 4)
-            {
-                remain -= 4;
-
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #192]           \n"
-                    "ld1    {v8.4s, v9.4s}, [%1]            \n" // r0
-                    "add    %1, %1, #16                     \n"
-
-                    "and    v4.16b, %11.16b, %11.16b        \n" // v4 = _bias0
-
-                    "ext    v12.16b, v8.16b, v9.16b, #4     \n"
-
-                    "fmul   v6.4s, v8.4s, %8.s[0]           \n"
-
-                    "ext    v13.16b, v8.16b, v9.16b, #8     \n"
-
-                    "fmla   v4.4s, v12.4s, %8.s[1]          \n"
-
-                    "prfm   pldl1keep, [%2, #192]           \n"
-                    "ld1    {v16.4s, v17.4s}, [%2]          \n" // r1
-                    "add    %2, %2, #16                     \n"
-
-                    "fmla   v6.4s, v13.4s, %8.s[2]          \n"
-
-                    "ext    v20.16b, v16.16b, v17.16b, #4   \n"
-
-                    "fmla   v4.4s, v16.4s, %9.s[0]          \n"
-
-                    "ext    v21.16b, v16.16b, v17.16b, #8   \n"
-
-                    "fmla   v6.4s, v20.4s, %9.s[1]          \n"
-
-                    "prfm   pldl1keep, [%3, #192]           \n"
-                    "ld1    {v24.4s, v25.4s}, [%3]          \n" // r2
-                    "add    %3, %3, #16                     \n"
-
-                    "fmla   v4.4s, v21.4s, %9.s[2]          \n"
-
-                    "ext    v12.16b, v24.16b, v25.16b, #4   \n"
-
-                    "fmla   v6.4s, v24.4s, %10.s[0]         \n"
-
-                    "ext    v13.16b, v24.16b, v25.16b, #8   \n"
-
-                    "fmla   v4.4s, v12.4s, %10.s[1]         \n"
-
-                    "fmla   v6.4s, v13.4s, %10.s[2]         \n"
-
-                    "fadd   v4.4s, v4.4s, v6.4s             \n"
-
-                    "st1    {v4.4s}, [%0], #16              \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(r0),     // %1
-                    "=r"(r1),     // %2
-                    "=r"(r2)      // %3
-                    : "0"(outptr),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k012x), // %8
-                    "w"(_k345x), // %9
-                    "w"(_k678x), // %10
-                    "w"(_bias0)  // %11
-                    : "cc", "memory", "v4", "v6", "v8", "v9", "v12", "v13", "v16", "v17", "v20", "v21", "v24", "v25");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    "pld        [%2, #192]          \n"
-                    "vld1.f32   {d16-d18}, [%2]     \n" // r0
-                    "add        %2, #16             \n"
-
-                    "vext.32    q10, q8, q9, #1     \n"
-                    "vext.32    q11, q8, q9, #2     \n"
-
-                    "0:                             \n"
-
-                    "vmul.f32   q7, q8, %e10[0]     \n"
-
-                    "vand       q14, %q13, %q13     \n" // q14 = _bias0
-                    "vmul.f32   q13, q10, %e10[1]   \n"
-                    "vmla.f32   q14, q11, %f10[0]   \n"
-
-                    "pld        [%3, #192]          \n"
-                    "vld1.f32   {d16-d18}, [%3]     \n" // r1
-                    "add        %3, #16             \n"
-
-                    "vmla.f32   q7, q8, %e11[0]     \n"
-
-                    "vext.32    q10, q8, q9, #1     \n"
-                    "vext.32    q11, q8, q9, #2     \n"
-
-                    "vmla.f32   q13, q10, %e11[1]   \n"
-                    "vmla.f32   q14, q11, %f11[0]   \n"
-
-                    "pld        [%4, #192]          \n"
-                    "vld1.f32   {d16-d18}, [%4]     \n" // r2
-                    "add        %4, #16             \n"
-
-                    "vmla.f32   q7, q8, %e12[0]     \n"
-
-                    "vext.32    q10, q8, q9, #1     \n"
-                    "vext.32    q11, q8, q9, #2     \n"
-
-                    "vmla.f32   q13, q10, %e12[1]   \n"
-                    "vmla.f32   q14, q11, %f12[0]   \n"
-
-                    "pld        [%2, #192]          \n"
-                    "vld1.f32   {d16-d18}, [%2]     \n" // r0
-                    "add        %2, #16             \n"
-
-                    "vadd.f32   q7, q7, q13         \n"
-                    "vadd.f32   q7, q7, q14         \n"
-
-                    "vext.32    q10, q8, q9, #1     \n"
-                    "vext.32    q11, q8, q9, #2     \n"
-
-                    "vst1.f32   {d14-d15}, [%1]!    \n"
-
-                    "subs       %0, #1              \n"
-                    "bne        0b                  \n"
-
-                    "sub        %2, #16             \n"
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0),     // %2
-                    "=r"(r1),     // %3
-                    "=r"(r2)      // %4
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "w"(_k012x), // %10
-                    "w"(_k345x), // %11
-                    "w"(_k678x), // %12
-                    "w"(_bias0)  // %13
-                    : "cc", "memory", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-#if __ARM_NEON
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r20 = vld1q_f32(r2);
-
-                float32x4_t _sum = vmulq_f32(_r00, _k012x);
-                _sum = vmlaq_f32(_sum, _r10, _k345x);
-                _sum = vmlaq_f32(_sum, _r20, _k678x);
-
-                _sum = vsetq_lane_f32(bias0, _sum, 3);
-#if __aarch64__
-                *outptr = vaddvq_f32(_sum);
-#else
-                float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-                _ss = vpadd_f32(_ss, _ss);
-
-                *outptr = vget_lane_f32(_ss, 0);
-#endif // __aarch64__
-#else
-                float sum = bias0;
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-
-                *outptr = sum;
-#endif
-                r0++;
-                r1++;
-                r2++;
-                outptr++;
-            }
-
-            r0 += 2;
-            r1 += 2;
-            r2 += 2;
-        }
-    }
-}
-
-static void convdw3x3s2_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const float bias0 = bias ? bias[g] : 0.f;
-
-        const float* kernel0 = kernel + g * 9;
-
-        float* outptr = out;
-
-        const float* img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0;
-        const float* r1 = img0 + w;
-        const float* r2 = img0 + w * 2;
-
-#if __ARM_NEON
-        float32x4_t _k012x = vld1q_f32(kernel0);
-        float32x4_t _k345x = vld1q_f32(kernel0 + 3);
-        float32x4_t _k678x = vld1q_f32(kernel0 + 6);
-
-        _k012x = vsetq_lane_f32(0.f, _k012x, 3);
-        _k345x = vsetq_lane_f32(0.f, _k345x, 3);
-        _k678x = vsetq_lane_f32(0.f, _k678x, 3);
-
-        float32x4_t _bias0 = vdupq_n_f32(bias0);
-#else
-        const float* k0 = kernel0;
-        const float* k1 = kernel0 + 3;
-        const float* k2 = kernel0 + 6;
-#endif // __ARM_NEON
-
-        int i = 0;
-
-        for (; i < outh; i++)
-        {
-#if __ARM_NEON
-            int nn = outw >> 2;
-            int remain = outw & 3;
-#else
-            int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    "prfm       pldl1keep, [%2, #256]          \n"
-                    "ld2        {v2.4s, v3.4s}, [%2], #32      \n"
-
-                    "and        v11.16b, %13.16b, %13.16b      \n" // v11 = _bias0
-
-                    "0:                                        \n"
-                    "fmul       v0.4s,  v2.4s, %10.s[0]        \n"
-                    "fmul       v10.4s, v3.4s, %10.s[1]        \n"
-
-                    "prfm       pldl1keep, [%2, #256]          \n"
-                    "ld2        {v8.4s, v9.4s}, [%2]           \n"
-                    "ext        v1.16b, v2.16b, v8.16b, #4     \n"
-
-                    "fmla       v11.4s, v1.4s, %10.s[2]        \n"
-
-                    "prfm       pldl1keep, [%3, #256]          \n"
-                    "ld2        {v2.4s, v3.4s}, [%3], #32      \n"
-
-                    "fmla       v0.4s,  v2.4s, %11.s[0]        \n"
-                    "fmla       v10.4s, v3.4s, %11.s[1]        \n"
-
-                    "prfm       pldl1keep, [%3, #256]          \n"
-                    "ld2        {v8.4s, v9.4s}, [%3]           \n"
-                    "ext        v1.16b, v2.16b, v8.16b, #4     \n"
-
-                    "fmla       v11.4s, v1.4s, %11.s[2]        \n"
-
-                    "prfm       pldl1keep, [%4, #256]          \n"
-                    "ld2        {v2.4s, v3.4s}, [%4], #32      \n"
-
-                    "fmla       v0.4s,  v2.4s, %12.s[0]        \n"
-                    "fmla       v10.4s, v3.4s, %12.s[1]        \n"
-
-                    "prfm       pldl1keep, [%4, #256]          \n"
-                    "ld2        {v8.4s, v9.4s}, [%4]           \n"
-                    "ext        v1.16b, v2.16b, v8.16b, #4     \n"
-
-                    "fmla       v11.4s, v1.4s, %12.s[2]        \n"
-
-                    "prfm       pldl1keep, [%2, #256]          \n"
-                    "ld2        {v2.4s, v3.4s}, [%2], #32      \n"
-
-                    "fadd       v0.4s, v0.4s, v10.4s           \n"
-                    "fadd       v0.4s, v0.4s, v11.4s           \n"
-
-                    "and        v11.16b, %13.16b, %13.16b      \n" // v11 = _bias0
-
-                    "subs       %w0, %w0, #1                   \n"
-                    "st1        {v0.4s}, [%1], #16             \n"
-                    "bne        0b                             \n"
-                    "sub        %2, %2, #32                    \n"
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0),     // %2
-                    "=r"(r1),     // %3
-                    "=r"(r2)      // %4
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "w"(_k012x), // %10
-                    "w"(_k345x), // %11
-                    "w"(_k678x), // %12
-                    "w"(_bias0)  // %13
-                    : "cc", "memory", "v0", "v1", "v2", "v3", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    "pld        [%2, #256]          \n"
-                    "vld2.f32   {d4-d7}, [%2]!      \n"
-
-                    "vand       q11, %q13, %q13     \n"
-
-                    "0:                             \n"
-                    "vmul.f32   q0, q2, %e10[0]     \n"
-                    "vmul.f32   q10, q3, %e10[1]    \n"
-
-                    "pld        [%2, #128]          \n"
-                    "vld2.f32   {d16-d17}, [%2]     \n"
-                    "vext.32    q1, q2, q8, #1      \n"
-
-                    "vmla.f32   q11, q1, %f10[0]    \n"
-
-                    "pld        [%3, #256]          \n"
-                    "vld2.f32   {d4-d7}, [%3]!      \n"
-
-                    "vmla.f32   q0, q2, %e11[0]     \n"
-                    "vmla.f32   q10, q3, %e11[1]    \n"
-
-                    "pld        [%3, #128]          \n"
-                    "vld2.f32   {d16-d17}, [%3]     \n"
-                    "vext.32    q1, q2, q8, #1      \n"
-
-                    "vmla.f32   q11, q1, %f11[0]    \n"
-
-                    "pld        [%4, #256]          \n"
-                    "vld2.f32   {d4-d7}, [%4]!      \n"
-
-                    "vmla.f32   q0, q2, %e12[0]     \n"
-                    "vmla.f32   q10, q3, %e12[1]    \n"
-
-                    "pld        [%4, #128]          \n"
-                    "vld2.f32   {d16-d17}, [%4]     \n"
-                    "vext.32    q1, q2, q8, #1      \n"
-
-                    "vmla.f32   q11, q1, %f12[0]    \n"
-
-                    "pld        [%2, #256]          \n"
-                    "vld2.f32   {d4-d7}, [%2]!      \n"
-
-                    "vadd.f32   q0, q0, q10         \n"
-                    "vadd.f32   q0, q0, q11         \n"
-
-                    "vand       q11, %q13, %q13     \n"
-
-                    "subs       %0, #1              \n"
-                    "vst1.f32   {d0-d1}, [%1]!      \n"
-                    "bne        0b                  \n"
-                    "sub        %2, #32             \n"
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0),     // %2
-                    "=r"(r1),     // %3
-                    "=r"(r2)      // %4
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "w"(_k012x), // %10
-                    "w"(_k345x), // %11
-                    "w"(_k678x), // %12
-                    "w"(_bias0)  // %13
-                    : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-#if __ARM_NEON
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r20 = vld1q_f32(r2);
-
-                float32x4_t _sum = vmulq_f32(_r00, _k012x);
-                _sum = vmlaq_f32(_sum, _r10, _k345x);
-                _sum = vmlaq_f32(_sum, _r20, _k678x);
-
-                _sum = vsetq_lane_f32(bias0, _sum, 3);
-#if __aarch64__
-                *outptr = vaddvq_f32(_sum);
-#else
-                float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-                _ss = vpadd_f32(_ss, _ss);
-
-                *outptr = vget_lane_f32(_ss, 0);
-#endif // __aarch64__
-#else
-                float sum = bias0;
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-
-                *outptr = sum;
-#endif // __ARM_NEON
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-                outptr++;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_fp16s.h
deleted file mode 100644
index 118914c9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_fp16s.h
+++ /dev/null
@@ -1,457 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const __fp16* kernel = _kernel;
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const __fp16 bias0 = bias ? bias[g] : 0.f;
-
-        const __fp16* kernel0 = kernel + g * 9;
-
-        __fp16* outptr0 = out;
-        __fp16* outptr1 = outptr0 + outw;
-
-        const __fp16* img0 = bottom_blob.channel(g);
-
-        const __fp16* r0 = img0;
-        const __fp16* r1 = img0 + w;
-        const __fp16* r2 = img0 + w * 2;
-        const __fp16* r3 = img0 + w * 3;
-
-        float16x4_t _k012x = vld1_f16(kernel0);
-        float16x4_t _k345x = vld1_f16(kernel0 + 3);
-        float16x4_t _k678x = vld1_f16(kernel0 + 6);
-
-        _k012x = vset_lane_f16(0.f, _k012x, 3);
-        _k345x = vset_lane_f16(0.f, _k345x, 3);
-        _k678x = vset_lane_f16(0.f, _k678x, 3);
-
-        float16x8_t _bias0 = vdupq_n_f16(bias0);
-
-        int i = 0;
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-            for (; j + 7 < outw; j += 8)
-            {
-                float16x8_t _r00 = vld1q_f16(r0);
-                float16x8_t _r10 = vld1q_f16(r1);
-                float16x8_t _r20 = vld1q_f16(r2);
-                float16x8_t _r30 = vld1q_f16(r3);
-
-                float16x8_t _r0n = vld1q_f16(r0 + 8);
-                float16x8_t _r1n = vld1q_f16(r1 + 8);
-                float16x8_t _r2n = vld1q_f16(r2 + 8);
-                float16x8_t _r3n = vld1q_f16(r3 + 8);
-
-                float16x8_t _r01 = vextq_f16(_r00, _r0n, 1);
-                float16x8_t _r11 = vextq_f16(_r10, _r1n, 1);
-                float16x8_t _r21 = vextq_f16(_r20, _r2n, 1);
-                float16x8_t _r31 = vextq_f16(_r30, _r3n, 1);
-
-                float16x8_t _r02 = vextq_f16(_r00, _r0n, 2);
-                float16x8_t _r12 = vextq_f16(_r10, _r1n, 2);
-                float16x8_t _r22 = vextq_f16(_r20, _r2n, 2);
-                float16x8_t _r32 = vextq_f16(_r30, _r3n, 2);
-
-                float16x8_t _sum0 = _bias0;
-                float16x8_t _sum1 = _bias0;
-
-                _sum0 = vfmaq_lane_f16(_sum0, _r00, _k012x, 0);
-                _sum0 = vfmaq_lane_f16(_sum0, _r01, _k012x, 1);
-                _sum0 = vfmaq_lane_f16(_sum0, _r02, _k012x, 2);
-                _sum1 = vfmaq_lane_f16(_sum1, _r10, _k012x, 0);
-                _sum1 = vfmaq_lane_f16(_sum1, _r11, _k012x, 1);
-                _sum1 = vfmaq_lane_f16(_sum1, _r12, _k012x, 2);
-
-                _sum0 = vfmaq_lane_f16(_sum0, _r10, _k345x, 0);
-                _sum0 = vfmaq_lane_f16(_sum0, _r11, _k345x, 1);
-                _sum0 = vfmaq_lane_f16(_sum0, _r12, _k345x, 2);
-                _sum1 = vfmaq_lane_f16(_sum1, _r20, _k345x, 0);
-                _sum1 = vfmaq_lane_f16(_sum1, _r21, _k345x, 1);
-                _sum1 = vfmaq_lane_f16(_sum1, _r22, _k345x, 2);
-
-                _sum0 = vfmaq_lane_f16(_sum0, _r20, _k678x, 0);
-                _sum0 = vfmaq_lane_f16(_sum0, _r21, _k678x, 1);
-                _sum0 = vfmaq_lane_f16(_sum0, _r22, _k678x, 2);
-                _sum1 = vfmaq_lane_f16(_sum1, _r30, _k678x, 0);
-                _sum1 = vfmaq_lane_f16(_sum1, _r31, _k678x, 1);
-                _sum1 = vfmaq_lane_f16(_sum1, _r32, _k678x, 2);
-
-                vst1q_f16(outptr0, _sum0);
-                vst1q_f16(outptr1, _sum1);
-
-                r0 += 8;
-                r1 += 8;
-                r2 += 8;
-                r3 += 8;
-                outptr0 += 8;
-                outptr1 += 8;
-            }
-            for (; j + 3 < outw; j += 4)
-            {
-                float16x4_t _r00 = vld1_f16(r0);
-                float16x4_t _r10 = vld1_f16(r1);
-                float16x4_t _r20 = vld1_f16(r2);
-                float16x4_t _r30 = vld1_f16(r3);
-
-                float16x4_t _r0n = vld1_f16(r0 + 4);
-                float16x4_t _r1n = vld1_f16(r1 + 4);
-                float16x4_t _r2n = vld1_f16(r2 + 4);
-                float16x4_t _r3n = vld1_f16(r3 + 4);
-
-                float16x4_t _r01 = vext_f16(_r00, _r0n, 1);
-                float16x4_t _r11 = vext_f16(_r10, _r1n, 1);
-                float16x4_t _r21 = vext_f16(_r20, _r2n, 1);
-                float16x4_t _r31 = vext_f16(_r30, _r3n, 1);
-
-                float16x4_t _r02 = vext_f16(_r00, _r0n, 2);
-                float16x4_t _r12 = vext_f16(_r10, _r1n, 2);
-                float16x4_t _r22 = vext_f16(_r20, _r2n, 2);
-                float16x4_t _r32 = vext_f16(_r30, _r3n, 2);
-
-                float16x4_t _sum0 = vget_low_f16(_bias0);
-                float16x4_t _sum1 = vget_low_f16(_bias0);
-
-                _sum0 = vfma_lane_f16(_sum0, _r00, _k012x, 0);
-                _sum0 = vfma_lane_f16(_sum0, _r01, _k012x, 1);
-                _sum0 = vfma_lane_f16(_sum0, _r02, _k012x, 2);
-                _sum1 = vfma_lane_f16(_sum1, _r10, _k012x, 0);
-                _sum1 = vfma_lane_f16(_sum1, _r11, _k012x, 1);
-                _sum1 = vfma_lane_f16(_sum1, _r12, _k012x, 2);
-
-                _sum0 = vfma_lane_f16(_sum0, _r10, _k345x, 0);
-                _sum0 = vfma_lane_f16(_sum0, _r11, _k345x, 1);
-                _sum0 = vfma_lane_f16(_sum0, _r12, _k345x, 2);
-                _sum1 = vfma_lane_f16(_sum1, _r20, _k345x, 0);
-                _sum1 = vfma_lane_f16(_sum1, _r21, _k345x, 1);
-                _sum1 = vfma_lane_f16(_sum1, _r22, _k345x, 2);
-
-                _sum0 = vfma_lane_f16(_sum0, _r20, _k678x, 0);
-                _sum0 = vfma_lane_f16(_sum0, _r21, _k678x, 1);
-                _sum0 = vfma_lane_f16(_sum0, _r22, _k678x, 2);
-                _sum1 = vfma_lane_f16(_sum1, _r30, _k678x, 0);
-                _sum1 = vfma_lane_f16(_sum1, _r31, _k678x, 1);
-                _sum1 = vfma_lane_f16(_sum1, _r32, _k678x, 2);
-
-                vst1_f16(outptr0, _sum0);
-                vst1_f16(outptr1, _sum1);
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                r3 += 4;
-                outptr0 += 4;
-                outptr1 += 4;
-            }
-            for (; j < outw; j++)
-            {
-                float16x4_t _r0 = vld1_f16(r0);
-                float16x4_t _r1 = vld1_f16(r1);
-                float16x4_t _r2 = vld1_f16(r2);
-                float16x4_t _r3 = vld1_f16(r3);
-
-                float16x4_t _sum0 = vmul_f16(_r0, _k012x);
-                _sum0 = vfma_f16(_sum0, _r1, _k345x);
-                _sum0 = vfma_f16(_sum0, _r2, _k678x);
-
-                float16x4_t _sum1 = vmul_f16(_r1, _k012x);
-                _sum1 = vfma_f16(_sum1, _r2, _k345x);
-                _sum1 = vfma_f16(_sum1, _r3, _k678x);
-
-                _sum0 = vset_lane_f16(bias0, _sum0, 3);
-                _sum1 = vset_lane_f16(bias0, _sum1, 3);
-
-                *outptr0 = (__fp16)vaddvq_f32(vcvt_f32_f16(_sum0));
-                *outptr1 = (__fp16)vaddvq_f32(vcvt_f32_f16(_sum1));
-
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                outptr0++;
-                outptr1++;
-            }
-
-            r0 += 2 + w;
-            r1 += 2 + w;
-            r2 += 2 + w;
-            r3 += 2 + w;
-
-            outptr0 += outw;
-            outptr1 += outw;
-        }
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 7 < outw; j += 8)
-            {
-                float16x8_t _r00 = vld1q_f16(r0);
-                float16x8_t _r10 = vld1q_f16(r1);
-                float16x8_t _r20 = vld1q_f16(r2);
-
-                float16x8_t _r0n = vld1q_f16(r0 + 8);
-                float16x8_t _r1n = vld1q_f16(r1 + 8);
-                float16x8_t _r2n = vld1q_f16(r2 + 8);
-
-                float16x8_t _r01 = vextq_f16(_r00, _r0n, 1);
-                float16x8_t _r11 = vextq_f16(_r10, _r1n, 1);
-                float16x8_t _r21 = vextq_f16(_r20, _r2n, 1);
-
-                float16x8_t _r02 = vextq_f16(_r00, _r0n, 2);
-                float16x8_t _r12 = vextq_f16(_r10, _r1n, 2);
-                float16x8_t _r22 = vextq_f16(_r20, _r2n, 2);
-
-                float16x8_t _sum0 = _bias0;
-
-                _sum0 = vfmaq_lane_f16(_sum0, _r00, _k012x, 0);
-                _sum0 = vfmaq_lane_f16(_sum0, _r01, _k012x, 1);
-                _sum0 = vfmaq_lane_f16(_sum0, _r02, _k012x, 2);
-
-                _sum0 = vfmaq_lane_f16(_sum0, _r10, _k345x, 0);
-                _sum0 = vfmaq_lane_f16(_sum0, _r11, _k345x, 1);
-                _sum0 = vfmaq_lane_f16(_sum0, _r12, _k345x, 2);
-
-                _sum0 = vfmaq_lane_f16(_sum0, _r20, _k678x, 0);
-                _sum0 = vfmaq_lane_f16(_sum0, _r21, _k678x, 1);
-                _sum0 = vfmaq_lane_f16(_sum0, _r22, _k678x, 2);
-
-                vst1q_f16(outptr0, _sum0);
-
-                r0 += 8;
-                r1 += 8;
-                r2 += 8;
-                outptr0 += 8;
-            }
-            for (; j + 3 < outw; j += 4)
-            {
-                float16x4_t _r00 = vld1_f16(r0);
-                float16x4_t _r10 = vld1_f16(r1);
-                float16x4_t _r20 = vld1_f16(r2);
-
-                float16x4_t _r0n = vld1_f16(r0 + 4);
-                float16x4_t _r1n = vld1_f16(r1 + 4);
-                float16x4_t _r2n = vld1_f16(r2 + 4);
-
-                float16x4_t _r01 = vext_f16(_r00, _r0n, 1);
-                float16x4_t _r11 = vext_f16(_r10, _r1n, 1);
-                float16x4_t _r21 = vext_f16(_r20, _r2n, 1);
-
-                float16x4_t _r02 = vext_f16(_r00, _r0n, 2);
-                float16x4_t _r12 = vext_f16(_r10, _r1n, 2);
-                float16x4_t _r22 = vext_f16(_r20, _r2n, 2);
-
-                float16x4_t _sum0 = vget_low_f16(_bias0);
-
-                _sum0 = vfma_lane_f16(_sum0, _r00, _k012x, 0);
-                _sum0 = vfma_lane_f16(_sum0, _r01, _k012x, 1);
-                _sum0 = vfma_lane_f16(_sum0, _r02, _k012x, 2);
-
-                _sum0 = vfma_lane_f16(_sum0, _r10, _k345x, 0);
-                _sum0 = vfma_lane_f16(_sum0, _r11, _k345x, 1);
-                _sum0 = vfma_lane_f16(_sum0, _r12, _k345x, 2);
-
-                _sum0 = vfma_lane_f16(_sum0, _r20, _k678x, 0);
-                _sum0 = vfma_lane_f16(_sum0, _r21, _k678x, 1);
-                _sum0 = vfma_lane_f16(_sum0, _r22, _k678x, 2);
-
-                vst1_f16(outptr0, _sum0);
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                outptr0 += 4;
-            }
-            for (; j < outw; j++)
-            {
-                float16x4_t _r0 = vld1_f16(r0);
-                float16x4_t _r1 = vld1_f16(r1);
-                float16x4_t _r2 = vld1_f16(r2);
-
-                float16x4_t _sum = vmul_f16(_r0, _k012x);
-                _sum = vfma_f16(_sum, _r1, _k345x);
-                _sum = vfma_f16(_sum, _r2, _k678x);
-
-                _sum = vset_lane_f16(bias0, _sum, 3);
-
-                *outptr0 = (__fp16)vaddvq_f32(vcvt_f32_f16(_sum));
-
-                r0++;
-                r1++;
-                r2++;
-                outptr0++;
-            }
-
-            r0 += 2;
-            r1 += 2;
-            r2 += 2;
-        }
-    }
-}
-
-static void convdw3x3s2_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const __fp16* kernel = _kernel;
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const __fp16 bias0 = bias ? bias[g] : 0.f;
-
-        const __fp16* kernel0 = kernel + g * 9;
-
-        __fp16* outptr = out;
-
-        const __fp16* img0 = bottom_blob.channel(g);
-
-        const __fp16* r0 = img0;
-        const __fp16* r1 = img0 + w;
-        const __fp16* r2 = img0 + w * 2;
-
-        float16x4_t _k012x = vld1_f16(kernel0);
-        float16x4_t _k345x = vld1_f16(kernel0 + 3);
-        float16x4_t _k678x = vld1_f16(kernel0 + 6);
-
-        _k012x = vset_lane_f16(0.f, _k012x, 3);
-        _k345x = vset_lane_f16(0.f, _k345x, 3);
-        _k678x = vset_lane_f16(0.f, _k678x, 3);
-
-        float16x8_t _bias0 = vdupq_n_f16(bias0);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 7 < outw; j += 8)
-            {
-                float16x8x2_t _r00 = vld2q_f16(r0);
-                float16x8x2_t _r10 = vld2q_f16(r1);
-                float16x8x2_t _r20 = vld2q_f16(r2);
-
-                float16x8x2_t _r0n = vld2q_f16(r0 + 16);
-                float16x8x2_t _r1n = vld2q_f16(r1 + 16);
-                float16x8x2_t _r2n = vld2q_f16(r2 + 16);
-
-                float16x8_t _r02 = vextq_f16(_r00.val[0], _r0n.val[0], 1);
-                float16x8_t _r12 = vextq_f16(_r10.val[0], _r1n.val[0], 1);
-                float16x8_t _r22 = vextq_f16(_r20.val[0], _r2n.val[0], 1);
-
-                float16x8_t _sum = _bias0;
-
-                _sum = vfmaq_lane_f16(_sum, _r00.val[0], _k012x, 0);
-                _sum = vfmaq_lane_f16(_sum, _r00.val[1], _k012x, 1);
-                _sum = vfmaq_lane_f16(_sum, _r02, _k012x, 2);
-
-                _sum = vfmaq_lane_f16(_sum, _r10.val[0], _k345x, 0);
-                _sum = vfmaq_lane_f16(_sum, _r10.val[1], _k345x, 1);
-                _sum = vfmaq_lane_f16(_sum, _r12, _k345x, 2);
-
-                _sum = vfmaq_lane_f16(_sum, _r20.val[0], _k678x, 0);
-                _sum = vfmaq_lane_f16(_sum, _r20.val[1], _k678x, 1);
-                _sum = vfmaq_lane_f16(_sum, _r22, _k678x, 2);
-
-                vst1q_f16(outptr, _sum);
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                outptr += 8;
-            }
-            for (; j + 3 < outw; j += 4)
-            {
-                float16x4x2_t _r00 = vld2_f16(r0);
-                float16x4x2_t _r10 = vld2_f16(r1);
-                float16x4x2_t _r20 = vld2_f16(r2);
-
-                float16x4x2_t _r0n = vld2_f16(r0 + 8);
-                float16x4x2_t _r1n = vld2_f16(r1 + 8);
-                float16x4x2_t _r2n = vld2_f16(r2 + 8);
-
-                float16x4_t _r02 = vext_f16(_r00.val[0], _r0n.val[0], 1);
-                float16x4_t _r12 = vext_f16(_r10.val[0], _r1n.val[0], 1);
-                float16x4_t _r22 = vext_f16(_r20.val[0], _r2n.val[0], 1);
-
-                float16x4_t _sum = vget_low_f16(_bias0);
-
-                _sum = vfma_lane_f16(_sum, _r00.val[0], _k012x, 0);
-                _sum = vfma_lane_f16(_sum, _r00.val[1], _k012x, 1);
-                _sum = vfma_lane_f16(_sum, _r02, _k012x, 2);
-
-                _sum = vfma_lane_f16(_sum, _r10.val[0], _k345x, 0);
-                _sum = vfma_lane_f16(_sum, _r10.val[1], _k345x, 1);
-                _sum = vfma_lane_f16(_sum, _r12, _k345x, 2);
-
-                _sum = vfma_lane_f16(_sum, _r20.val[0], _k678x, 0);
-                _sum = vfma_lane_f16(_sum, _r20.val[1], _k678x, 1);
-                _sum = vfma_lane_f16(_sum, _r22, _k678x, 2);
-
-                vst1_f16(outptr, _sum);
-
-                r0 += 8;
-                r1 += 8;
-                r2 += 8;
-                outptr += 4;
-            }
-            for (; j < outw; j++)
-            {
-                float16x4_t _r0 = vld1_f16(r0);
-                float16x4_t _r1 = vld1_f16(r1);
-                float16x4_t _r2 = vld1_f16(r2);
-
-                float16x4_t _sum = vmul_f16(_r0, _k012x);
-                _sum = vfma_f16(_sum, _r1, _k345x);
-                _sum = vfma_f16(_sum, _r2, _k678x);
-
-                _sum = vset_lane_f16(bias0, _sum, 3);
-
-                *outptr = (__fp16)vaddvq_f32(vcvt_f32_f16(_sum));
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-                outptr++;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_int8.h
deleted file mode 100644
index 1d7dd1b7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_int8.h
+++ /dev/null
@@ -1,1798 +0,0 @@
-// BUG1989 is pleased to support the open source community by supporting ncnn available.
-//
-// Copyright (C) 2019 BUG1989. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_int8_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const signed char* kernel = (const signed char*)_kernel + p * 9;
-
-        int* outptr0 = out;
-        int* outptr0n = outptr0 + outw;
-
-        const signed char* img0 = bottom_blob.channel(p);
-
-        const signed char* r0 = img0;
-        const signed char* r1 = img0 + w;
-        const signed char* r2 = img0 + w * 2;
-        const signed char* r3 = img0 + w * 3;
-
-        int i = 0;
-
-#if __ARM_NEON
-        int8x16_t _k0123456789x = vld1q_s8(kernel);
-        int16x8_t _k_s16 = vmovl_s8(vget_low_s8(_k0123456789x));
-        int16x8_t _kn_s16 = vmovl_s8(vget_high_s8(_k0123456789x));
-
-        int16x4_t _k0123 = vget_low_s16(_k_s16);
-        int16x4_t _k4567 = vget_high_s16(_k_s16);
-        int16x4_t _k8xxx = vget_low_s16(_kn_s16);
-#endif // __ARM_NEON
-
-        for (; i + 1 < outh; i += 2)
-        {
-#if __ARM_NEON
-            int nn = outw >> 3;
-            int remain = outw & 7;
-#else
-            int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    "0:                                   \n"
-                    "ld1    {v4.8b, v5.8b}, [%3]          \n"
-                    "ld1    {v6.8b, v7.8b}, [%4]          \n"
-                    "ld1    {v8.8b, v9.8b}, [%5]          \n"
-                    "ld1    {v10.8b, v11.8b}, [%6]        \n"
-                    "add    %3, %3, #8                    \n"
-                    "add    %4, %4, #8                    \n"
-                    "add    %5, %5, #8                    \n"
-                    "add    %6, %6, #8                    \n"
-
-                    "ext    v12.8b, v4.8b, v5.8b, #1      \n"
-                    "ext    v13.8b, v4.8b, v5.8b, #2      \n"
-                    "ext    v14.8b, v6.8b, v7.8b, #1      \n"
-                    "ext    v15.8b, v6.8b, v7.8b, #2      \n"
-                    "ext    v16.8b, v8.8b, v9.8b, #1      \n"
-                    "ext    v17.8b, v8.8b, v9.8b, #2      \n"
-                    "ext    v18.8b, v10.8b, v11.8b, #1    \n"
-                    "ext    v19.8b, v10.8b, v11.8b, #2    \n"
-
-                    "sshll  v4.8h, v4.8b, #0              \n" // r00
-                    "sshll  v12.8h, v12.8b, #0            \n" // r01
-                    "sshll  v13.8h, v13.8b, #0            \n" // r02
-                    "sshll  v6.8h, v6.8b, #0              \n" // r10
-                    "sshll  v14.8h, v14.8b, #0            \n" // r11
-                    "sshll  v15.8h, v15.8b, #0            \n" // r12
-                    "sshll  v8.8h, v8.8b, #0              \n" // r20
-                    "sshll  v16.8h, v16.8b, #0            \n" // r21
-                    "sshll  v17.8h, v17.8b, #0            \n" // r22
-                    "sshll  v10.8h, v10.8b, #0            \n" // r30
-                    "sshll  v18.8h, v18.8b, #0            \n" // r31
-                    "sshll  v19.8h, v19.8b, #0            \n" // r32
-
-                    // r0
-                    "smull  v20.4s, v4.4h, %14.h[0]       \n" // (r00 - r07) * k00
-                    "smull2  v21.4s, v4.8h, %14.h[0]      \n"
-                    "smull  v22.4s, v12.4h, %14.h[1]      \n" // (r01 - r08) * k01
-                    "smull2  v23.4s, v12.8h, %14.h[1]     \n"
-                    "smull  v24.4s, v13.4h, %14.h[2]      \n" // (r02 - r09) * k02
-                    "smull2  v25.4s, v13.8h, %14.h[2]     \n"
-
-                    // r1
-                    "smull  v26.4s, v6.4h, %14.h[0]       \n" // (r10 - r17) * k00
-                    "smull2  v27.4s, v6.8h, %14.h[0]      \n"
-                    "smull  v28.4s, v14.4h, %14.h[1]      \n" // (r11 - r18) * k01
-                    "smull2  v29.4s, v14.8h, %14.h[1]     \n"
-                    "smull  v30.4s, v15.4h, %14.h[2]      \n" // (r12 - r19) * k02
-                    "smull2  v31.4s, v15.8h, %14.h[2]     \n"
-
-                    "smlal  v20.4s, v6.4h, %14.h[3]       \n" // (r10 - r17) * k03
-                    "smlal2  v21.4s, v6.8h, %14.h[3]      \n"
-                    "smlal  v22.4s, v14.4h, %15.h[0]      \n" // (r11 - r18) * k04
-                    "smlal2  v23.4s, v14.8h, %15.h[0]     \n"
-                    "smlal  v24.4s, v15.4h, %15.h[1]      \n" // (r12 - r19) * k05
-                    "smlal2  v25.4s, v15.8h, %15.h[1]     \n"
-
-                    // r2
-                    "smlal  v26.4s, v8.4h, %14.h[3]       \n" // (r20 - r27) * k03
-                    "smlal2  v27.4s, v8.8h, %14.h[3]      \n"
-                    "smlal  v28.4s, v16.4h, %15.h[0]      \n" // (r21 - r28) * k04
-                    "smlal2  v29.4s, v16.8h, %15.h[0]     \n"
-                    "smlal  v30.4s, v17.4h, %15.h[1]      \n" // (r22 - r29) * k05
-                    "smlal2  v31.4s, v17.8h, %15.h[1]     \n"
-
-                    "smlal  v20.4s, v8.4h, %15.h[2]       \n" // (r20 - r27) * k06
-                    "smlal2  v21.4s, v8.8h, %15.h[2]      \n"
-                    "smlal  v22.4s, v16.4h, %15.h[3]      \n" // (r21 - r28) * k07
-                    "smlal2  v23.4s, v16.8h, %15.h[3]     \n"
-                    "smlal  v24.4s, v17.4h, %16.h[0]      \n" // (r22 - r29) * k08
-                    "smlal2  v25.4s, v17.8h, %16.h[0]     \n"
-
-                    // r3
-                    "smlal  v26.4s, v10.4h, %15.h[2]      \n" // (r30 - r37) * k06
-                    "smlal2  v27.4s, v10.8h, %15.h[2]     \n"
-                    "smlal  v28.4s, v18.4h, %15.h[3]      \n" // (r31 - r38) * k07
-                    "smlal2  v29.4s, v18.8h, %15.h[3]     \n"
-                    "smlal  v30.4s, v19.4h, %16.h[0]      \n" // (r32 - r39) * k08
-                    "smlal2  v31.4s, v19.8h, %16.h[0]     \n"
-
-                    // add and save
-                    "add    v20.4s, v20.4s, v22.4s        \n"
-                    "add    v21.4s, v21.4s, v23.4s        \n"
-                    "add    v26.4s, v26.4s, v28.4s        \n"
-                    "add    v27.4s, v27.4s, v29.4s        \n"
-                    "add    v20.4s, v20.4s, v24.4s        \n"
-                    "add    v21.4s, v21.4s, v25.4s        \n"
-                    "add    v26.4s, v26.4s, v30.4s        \n"
-                    "add    v27.4s, v27.4s, v31.4s        \n"
-
-                    "st1    {v20.4s, v21.4s}, [%1], #32   \n"
-                    "st1    {v26.4s, v27.4s}, [%2], #32   \n"
-
-                    "subs   %w0, %w0, #1                  \n"
-                    "bne    0b                            \n"
-
-                    : "=r"(nn),       // %0
-                    "=r"(outptr0),  // %1
-                    "=r"(outptr0n), // %2
-                    "=r"(r0),       // %3
-                    "=r"(r1),       // %4
-                    "=r"(r2),       // %5
-                    "=r"(r3)        // %6
-                    : "0"(nn),
-                    "1"(outptr0),
-                    "2"(outptr0n),
-                    "3"(r0),
-                    "4"(r1),
-                    "5"(r2),
-                    "6"(r3),
-                    "w"(_k0123), // %14
-                    "w"(_k4567), // %15
-                    "w"(_k8xxx)  // %16
-                    : "cc", "memory", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    "0:                              \n"
-                    // r0
-                    "vld1.s8    {d30-d31}, [%3]      \n" // r0
-                    "add    %3, %3, #8               \n"
-
-                    "vext.s8    d10, d30, d31, #1    \n"
-                    "vext.s8    d12, d30, d31, #2    \n"
-
-                    "vmovl.s8    q15, d30            \n" // r00
-                    "vmovl.s8    q5, d10             \n" // r01
-                    "vmovl.s8    q6, d12             \n" // r02
-                    // sum0
-                    "vmull.s16  q7, d30, %P14[0]     \n" // (r00 - r07) * k00
-                    "vmull.s16  q8, d31, %P14[0]     \n"
-                    "vmull.s16  q9, d10, %P14[1]     \n" // (r01 - r08) * k01
-                    "vmull.s16  q10, d11, %P14[1]    \n"
-                    "vmlal.s16  q7, d12, %P14[2]     \n" // (r02 - r09) * k02
-                    "vmlal.s16  q8, d13, %P14[2]     \n"
-
-                    // r1
-                    "vld1.s8    {d30-d31}, [%4]      \n" // r1
-                    "add    %4, %4, #8               \n"
-
-                    "vext.s8    d10, d30, d31, #1    \n"
-                    "vext.s8    d12, d30, d31, #2    \n"
-
-                    "vmovl.s8    q15, d30            \n" // r10
-                    "vmovl.s8    q5, d10             \n" // r11
-                    "vmovl.s8    q6, d12             \n" // r12
-                    // sum0
-                    "vmlal.s16  q7, d30, %P14[3]     \n" // (r10 - r17) * k03
-                    "vmlal.s16  q8, d31, %P14[3]     \n"
-                    "vmlal.s16  q9, d10, %P15[0]     \n" // (r11 - r18) * k04
-                    "vmlal.s16  q10, d11, %P15[0]    \n"
-                    "vmlal.s16  q7, d12, %P15[1]     \n" // (r12 - r19) * k05
-                    "vmlal.s16  q8, d13, %P15[1]     \n"
-                    // sum1
-                    "vmull.s16  q11, d30, %P14[0]    \n" // (r10 - r17) * k00
-                    "vmull.s16  q12, d31, %P14[0]    \n"
-                    "vmull.s16  q13, d10, %P14[1]    \n" // (r11 - r18) * k01
-                    "vmull.s16  q14, d11, %P14[1]    \n"
-                    "vmlal.s16  q11, d12, %P14[2]    \n" // (r12 - r19) * k02
-                    "vmlal.s16  q12, d13, %P14[2]    \n"
-
-                    // r2
-                    "vld1.s8    {d30-d31}, [%5]      \n" // r2
-                    "add    %5, %5, #8               \n"
-
-                    "vext.s8    d10, d30, d31, #1    \n"
-                    "vext.s8    d12, d30, d31, #2    \n"
-
-                    "vmovl.s8    q15, d30            \n" // r20
-                    "vmovl.s8    q5, d10             \n" // r21
-                    "vmovl.s8    q6, d12             \n" // r22
-
-                    // sum0
-                    "vmlal.s16  q7, d30, %P15[2]     \n" // (r20 - r27) * k06
-                    "vmlal.s16  q8, d31, %P15[2]     \n"
-                    "vmlal.s16  q9, d10, %P15[3]     \n" // (r21 - r28) * k07
-                    "vmlal.s16  q10, d11, %P15[3]    \n"
-                    "vmlal.s16  q7, d12, %P16[0]     \n" // (r22 - r29) * k08
-                    "vmlal.s16  q8, d13, %P16[0]     \n"
-                    // sum1
-                    "vmlal.s16  q11, d30, %P14[3]    \n" // (r20 - r27) * k03
-                    "vmlal.s16  q12, d31, %P14[3]    \n"
-                    "vmlal.s16  q13, d10, %P15[0]    \n" // (r21 - r28) * k04
-                    "vmlal.s16  q14, d11, %P15[0]    \n"
-                    "vmlal.s16  q11, d12, %P15[1]    \n" // (r22 - r29) * k05
-                    "vmlal.s16  q12, d13, %P15[1]    \n"
-
-                    // r3
-                    "vld1.s8    {d30-d31}, [%6]      \n" // r3
-                    "add    %6, %6, #8               \n"
-
-                    "vext.s8    d10, d30, d31, #1    \n"
-                    "vext.s8    d12, d30, d31, #2    \n"
-
-                    "vmovl.s8    q15, d30            \n" // r30
-                    "vmovl.s8    q5, d10             \n" // r31
-                    "vmovl.s8    q6, d12             \n" // r32
-
-                    // sum1
-                    "vmlal.s16  q11, d30, %P15[2]    \n" // (r30 - r37) * k06
-                    "vmlal.s16  q12, d31, %P15[2]    \n"
-                    "vmlal.s16  q13, d10, %P15[3]    \n" // (r31 - r38) * k07
-                    "vmlal.s16  q14, d11, %P15[3]    \n"
-                    "vmlal.s16  q11, d12, %P16[0]    \n" // (r32 - r39) * k08
-                    "vmlal.s16  q12, d13, %P16[0]    \n"
-
-                    "subs   %0, %0, #1               \n"
-
-                    // add and save
-                    "vadd.s32    q7, q7, q9          \n"
-                    "vadd.s32    q8, q8, q10         \n"
-                    "vadd.s32    q11, q11, q13       \n"
-                    "vadd.s32    q12, q12, q14       \n"
-
-                    "vst1.s32    {d14-d17}, [%1]!    \n"
-                    "vst1.s32    {d22-d25}, [%2]!    \n"
-
-                    "bne    0b                       \n"
-
-                    : "=r"(nn),       // %0
-                    "=r"(outptr0),  // %1
-                    "=r"(outptr0n), // %2
-                    "=r"(r0),       // %3
-                    "=r"(r1),       // %4
-                    "=r"(r2),       // %5
-                    "=r"(r3)        // %6
-                    : "0"(nn),
-                    "1"(outptr0),
-                    "2"(outptr0n),
-                    "3"(r0),
-                    "4"(r1),
-                    "5"(r2),
-                    "6"(r3),
-                    "w"(_k0123), // %14
-                    "w"(_k4567), // %15
-                    "w"(_k8xxx)  // %16
-                    : "cc", "memory", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-                // TODO NEON
-                int sum0 = 0;
-                int sum0n = 0;
-
-                sum0 += (int)r0[0] * kernel[0];
-                sum0 += (int)r0[1] * kernel[1];
-                sum0 += (int)r0[2] * kernel[2];
-                sum0 += (int)r1[0] * kernel[3];
-                sum0 += (int)r1[1] * kernel[4];
-                sum0 += (int)r1[2] * kernel[5];
-                sum0 += (int)r2[0] * kernel[6];
-                sum0 += (int)r2[1] * kernel[7];
-                sum0 += (int)r2[2] * kernel[8];
-
-                sum0n += (int)r1[0] * kernel[0];
-                sum0n += (int)r1[1] * kernel[1];
-                sum0n += (int)r1[2] * kernel[2];
-                sum0n += (int)r2[0] * kernel[3];
-                sum0n += (int)r2[1] * kernel[4];
-                sum0n += (int)r2[2] * kernel[5];
-                sum0n += (int)r3[0] * kernel[6];
-                sum0n += (int)r3[1] * kernel[7];
-                sum0n += (int)r3[2] * kernel[8];
-
-                *outptr0 = sum0;
-                *outptr0n = sum0n;
-
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                outptr0++;
-                outptr0n++;
-            }
-
-            r0 += 2 + w;
-            r1 += 2 + w;
-            r2 += 2 + w;
-            r3 += 2 + w;
-
-            outptr0 += outw;
-            outptr0n += outw;
-        }
-
-        for (; i < outh; i++)
-        {
-#if __ARM_NEON
-            int nn = outw >> 3;
-            int remain = outw & 7;
-#else
-            int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    "0:                                   \n"
-                    "ld1    {v4.8b, v5.8b}, [%2]          \n"
-                    "ld1    {v6.8b, v7.8b}, [%3]          \n"
-                    "ld1    {v8.8b, v9.8b}, [%4]          \n"
-                    "add    %2, %2, #8                    \n"
-                    "add    %3, %3, #8                    \n"
-                    "add    %4, %4, #8                    \n"
-
-                    "ext    v12.8b, v4.8b, v5.8b, #1      \n"
-                    "ext    v13.8b, v4.8b, v5.8b, #2      \n"
-                    "ext    v14.8b, v6.8b, v7.8b, #1      \n"
-                    "ext    v15.8b, v6.8b, v7.8b, #2      \n"
-                    "ext    v16.8b, v8.8b, v9.8b, #1      \n"
-                    "ext    v17.8b, v8.8b, v9.8b, #2      \n"
-
-                    "sshll  v4.8h, v4.8b, #0              \n" // r00
-                    "sshll  v12.8h, v12.8b, #0            \n" // r01
-                    "sshll  v13.8h, v13.8b, #0            \n" // r02
-                    "sshll  v6.8h, v6.8b, #0              \n" // r10
-                    "sshll  v14.8h, v14.8b, #0            \n" // r11
-                    "sshll  v15.8h, v15.8b, #0            \n" // r12
-                    "sshll  v8.8h, v8.8b, #0              \n" // r20
-                    "sshll  v16.8h, v16.8b, #0            \n" // r21
-                    "sshll  v17.8h, v17.8b, #0            \n" // r22
-
-                    // r0
-                    "smull  v20.4s, v4.4h, %10.h[0]       \n" // (r00 - r07) * k00
-                    "smull2  v21.4s, v4.8h, %10.h[0]      \n"
-                    "smull  v22.4s, v12.4h, %10.h[1]      \n" // (r01 - r08) * k01
-                    "smull2  v23.4s, v12.8h, %10.h[1]     \n"
-                    "smull  v24.4s, v13.4h, %10.h[2]      \n" // (r02 - r09) * k02
-                    "smull2  v25.4s, v13.8h, %10.h[2]     \n"
-
-                    // r1
-                    "smlal  v20.4s, v6.4h, %10.h[3]       \n" // (r10 - r17) * k03
-                    "smlal2  v21.4s, v6.8h, %10.h[3]      \n"
-                    "smlal  v22.4s, v14.4h, %11.h[0]      \n" // (r11 - r18) * k04
-                    "smlal2  v23.4s, v14.8h, %11.h[0]     \n"
-                    "smlal  v24.4s, v15.4h, %11.h[1]      \n" // (r12 - r19) * k05
-                    "smlal2  v25.4s, v15.8h, %11.h[1]     \n"
-
-                    // r2
-                    "smlal  v20.4s, v8.4h, %11.h[2]       \n" // (r20 - r27) * k06
-                    "smlal2  v21.4s, v8.8h, %11.h[2]      \n"
-                    "smlal  v22.4s, v16.4h, %11.h[3]      \n" // (r21 - r28) * k07
-                    "smlal2  v23.4s, v16.8h, %11.h[3]     \n"
-                    "smlal  v24.4s, v17.4h, %12.h[0]      \n" // (r22 - r29) * k08
-                    "smlal2  v25.4s, v17.8h, %12.h[0]     \n"
-
-                    // add and save
-                    "add    v20.4s, v20.4s, v22.4s        \n"
-                    "add    v21.4s, v21.4s, v23.4s        \n"
-                    "add    v20.4s, v20.4s, v24.4s        \n"
-                    "add    v21.4s, v21.4s, v25.4s        \n"
-
-                    "st1    {v20.4s, v21.4s}, [%1], #32   \n"
-
-                    "subs   %w0, %w0, #1                  \n"
-                    "bne    0b                            \n"
-
-                    : "=r"(nn),      // %0
-                    "=r"(outptr0), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2)       // %4
-                    : "0"(nn),
-                    "1"(outptr0),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "w"(_k0123), // %10
-                    "w"(_k4567), // %11
-                    "w"(_k8xxx)  // %12
-                    : "cc", "memory", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    "0:                              \n"
-                    // r0
-                    "vld1.s8    {d30-d31}, [%2]        \n" // r0
-                    "add    %2, %2, #8               \n"
-
-                    "vext.s8    d10, d30, d31, #1      \n"
-                    "vext.s8    d12, d30, d31, #2      \n"
-
-                    "vmovl.s8    q15, d30              \n" // r00
-                    "vmovl.s8    q5, d10             \n"   // r01
-                    "vmovl.s8    q6, d12             \n"   // r02
-                    // sum0
-                    "vmull.s16  q7, d30, %P10[0]      \n" // (r00 - r07) * k00
-                    "vmull.s16  q8, d31, %P10[0]      \n"
-                    "vmull.s16  q9, d10, %P10[1]     \n" // (r01 - r08) * k01
-                    "vmull.s16  q10, d11, %P10[1]    \n"
-                    "vmlal.s16  q7, d12, %P10[2]     \n" // (r02 - r09) * k02
-                    "vmlal.s16  q8, d13, %P10[2]     \n"
-
-                    // r1
-                    "vld1.s8    {d30-d31}, [%3]        \n" // r1
-                    "add    %3, %3, #8               \n"
-
-                    "vext.s8    d10, d30, d31, #1      \n"
-                    "vext.s8    d12, d30, d31, #2      \n"
-
-                    "vmovl.s8    q15, d30              \n" // r10
-                    "vmovl.s8    q5, d10             \n"   // r11
-                    "vmovl.s8    q6, d12             \n"   // r12
-                    // sum0
-                    "vmlal.s16  q7, d30, %P10[3]      \n" // (r10 - r17) * k03
-                    "vmlal.s16  q8, d31, %P10[3]      \n"
-                    "vmlal.s16  q9, d10, %P11[0]     \n" // (r11 - r18) * k04
-                    "vmlal.s16  q10, d11, %P11[0]    \n"
-                    "vmlal.s16  q7, d12, %P11[1]     \n" // (r12 - r19) * k05
-                    "vmlal.s16  q8, d13, %P11[1]     \n"
-
-                    // r2
-                    "vld1.s8    {d30-d31}, [%4]        \n" // r2
-                    "add    %4, %4, #8               \n"
-
-                    "vext.s8    d10, d30, d31, #1      \n"
-                    "vext.s8    d12, d30, d31, #2      \n"
-
-                    "vmovl.s8    q15, d30              \n" // r20
-                    "vmovl.s8    q5, d10             \n"   // r21
-                    "vmovl.s8    q6, d12             \n"   // r22
-
-                    // sum0
-                    "vmlal.s16  q7, d30, %P11[2]      \n" // (r20 - r27) * k06
-                    "vmlal.s16  q8, d31, %P11[2]      \n"
-                    "vmlal.s16  q9, d10, %P11[3]     \n" // (r21 - r28) * k07
-                    "vmlal.s16  q10, d11, %P11[3]    \n"
-                    "vmlal.s16  q7, d12, %P12[0]     \n" // (r22 - r29) * k08
-                    "vmlal.s16  q8, d13, %P12[0]     \n"
-
-                    "subs   %0, %0, #1               \n"
-
-                    // add and save
-                    "vadd.s32    q7, q7, q9          \n"
-                    "vadd.s32    q8, q8, q10         \n"
-
-                    "vst1.s32    {d14-d17}, [%1]!    \n"
-
-                    "bne    0b                       \n"
-
-                    : "=r"(nn),      // %0
-                    "=r"(outptr0), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2)       // %4
-                    : "0"(nn),
-                    "1"(outptr0),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "w"(_k0123), // %10
-                    "w"(_k4567), // %11
-                    "w"(_k8xxx)  // %12
-                    : "cc", "memory", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-                int sum = 0;
-
-                sum += (int)r0[0] * kernel[0];
-                sum += (int)r0[1] * kernel[1];
-                sum += (int)r0[2] * kernel[2];
-                sum += (int)r1[0] * kernel[3];
-                sum += (int)r1[1] * kernel[4];
-                sum += (int)r1[2] * kernel[5];
-                sum += (int)r2[0] * kernel[6];
-                sum += (int)r2[1] * kernel[7];
-                sum += (int)r2[2] * kernel[8];
-
-                *outptr0 = sum;
-
-                r0++;
-                r1++;
-                r2++;
-                outptr0++;
-            }
-
-            r0 += 2;
-            r1 += 2;
-            r2 += 2;
-        }
-    }
-}
-
-static void convdw3x3s2_int8_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const signed char* kernel = (const signed char*)_kernel + p * 9;
-
-        int* outptr = out;
-
-        const signed char* img = bottom_blob.channel(p);
-
-        const signed char* r0 = img;
-        const signed char* r1 = img + w;
-        const signed char* r2 = img + w * 2;
-
-        int i = 0;
-#if __ARM_NEON
-        int8x16_t _k0123456789x = vld1q_s8(kernel);
-        int16x8_t _k_s16 = vmovl_s8(vget_low_s8(_k0123456789x));
-        int16x8_t _kn_s16 = vmovl_s8(vget_high_s8(_k0123456789x));
-
-        int16x4_t _k0123 = vget_low_s16(_k_s16);
-        int16x4_t _k4567 = vget_high_s16(_k_s16);
-        int16x4_t _k8xxx = vget_low_s16(_kn_s16);
-#endif // __ARM_NEON
-        for (; i < outh; i++)
-        {
-#if __ARM_NEON
-            int nn = outw >> 3;
-            int remain = outw & 7;
-#else
-            int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    "0:                                   \n"
-                    "ld2    {v4.8b, v5.8b}, [%2], #16     \n"
-                    "ld2    {v6.8b, v7.8b}, [%2]          \n"
-                    "ld2    {v8.8b, v9.8b}, [%3], #16     \n"
-                    "ld2    {v10.8b, v11.8b}, [%3]        \n"
-                    "ld2    {v12.8b, v13.8b}, [%4], #16   \n"
-                    "ld2    {v14.8b, v15.8b}, [%4]        \n"
-
-                    "ext    v6.8b, v4.8b, v6.8b, #1       \n"
-                    "ext    v10.8b, v8.8b, v10.8b, #1     \n"
-                    "ext    v14.8b, v12.8b, v14.8b, #1    \n"
-
-                    "sshll  v4.8h, v4.8b, #0              \n" // r00
-                    "sshll  v5.8h, v5.8b, #0              \n" // r01
-                    "sshll  v6.8h, v6.8b, #0              \n" // r02
-                    "sshll  v8.8h, v8.8b, #0              \n" // r10
-                    "sshll  v9.8h, v9.8b, #0              \n" // r11
-                    "sshll  v10.8h, v10.8b, #0            \n" // r12
-                    "sshll  v12.8h, v12.8b, #0            \n" // r20
-                    "sshll  v13.8h, v13.8b, #0            \n" // r21
-                    "sshll  v14.8h, v14.8b, #0            \n" // r22
-
-                    // r0
-                    "smull  v20.4s, v4.4h, %10.h[0]       \n" // (r00 - r07) * k00
-                    "smull2  v21.4s, v4.8h, %10.h[0]      \n"
-                    "smull  v22.4s, v5.4h, %10.h[1]       \n" // (r01 - r08) * k01
-                    "smull2  v23.4s, v5.8h, %10.h[1]      \n"
-                    "smull  v24.4s, v6.4h, %10.h[2]       \n" // (r02 - r09) * k02
-                    "smull2  v25.4s, v6.8h, %10.h[2]      \n"
-
-                    // r1
-                    "smlal  v20.4s, v8.4h, %10.h[3]       \n" // (r10 - r17) * k03
-                    "smlal2  v21.4s, v8.8h, %10.h[3]      \n"
-                    "smlal  v22.4s, v9.4h, %11.h[0]       \n" // (r11 - r18) * k04
-                    "smlal2  v23.4s, v9.8h, %11.h[0]      \n"
-                    "smlal  v24.4s, v10.4h, %11.h[1]      \n" // (r12 - r19) * k05
-                    "smlal2  v25.4s, v10.8h, %11.h[1]     \n"
-
-                    // r2
-                    "smlal  v20.4s, v12.4h, %11.h[2]      \n" // (r20 - r27) * k06
-                    "smlal2  v21.4s, v12.8h, %11.h[2]     \n"
-                    "smlal  v22.4s, v13.4h, %11.h[3]      \n" // (r21 - r28) * k07
-                    "smlal2  v23.4s, v13.8h, %11.h[3]     \n"
-                    "smlal  v24.4s, v14.4h, %12.h[0]      \n" // (r22 - r29) * k08
-                    "smlal2  v25.4s, v14.8h, %12.h[0]     \n"
-
-                    // add and save
-                    "add    v20.4s, v20.4s, v22.4s        \n"
-                    "add    v21.4s, v21.4s, v23.4s        \n"
-                    "add    v20.4s, v20.4s, v24.4s        \n"
-                    "add    v21.4s, v21.4s, v25.4s        \n"
-
-                    "st1    {v20.4s, v21.4s}, [%1], #32   \n"
-
-                    "subs   %w0, %w0, #1                  \n"
-                    "bne    0b                            \n"
-
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0),     // %2
-                    "=r"(r1),     // %3
-                    "=r"(r2)      // %4
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "w"(_k0123), // %10
-                    "w"(_k4567), // %11
-                    "w"(_k8xxx)  // %12
-                    : "cc", "memory", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    "0:                              \n"
-                    // r0
-                    "vld2.s8    {d30-d31}, [%2]!     \n" // r0
-                    "vld2.s8    {d10-d11}, [%2]      \n"
-                    "vext.s8    d12, d30, d10, #1    \n"
-
-                    "vmovl.s8    q5, d31             \n" // r01
-                    "vmovl.s8    q15, d30            \n" // r00
-                    "vmovl.s8    q6, d12             \n" // r02
-                    // sum0
-                    "vmull.s16  q7, d30, %P10[0]     \n" // (r00 - r07) * k00
-                    "vmull.s16  q8, d31, %P10[0]     \n"
-                    "vmull.s16  q9, d10, %P10[1]     \n" // (r01 - r08) * k01
-                    "vmull.s16  q10, d11, %P10[1]    \n"
-                    "vmlal.s16  q7, d12, %P10[2]     \n" // (r02 - r09) * k02
-                    "vmlal.s16  q8, d13, %P10[2]     \n"
-
-                    // r1
-                    "vld2.s8    {d30-d31}, [%3]!     \n" // r1
-                    "vld2.s8    {d10-d11}, [%3]      \n"
-                    "vext.s8    d12, d30, d10, #1    \n"
-
-                    "vmovl.s8    q5, d31             \n" // r11
-                    "vmovl.s8    q15, d30            \n" // r10
-                    "vmovl.s8    q6, d12             \n" // r12
-                    // sum0
-                    "vmlal.s16  q7, d30, %P10[3]     \n" // (r10 - r17) * k03
-                    "vmlal.s16  q8, d31, %P10[3]     \n"
-                    "vmlal.s16  q9, d10, %P11[0]     \n" // (r11 - r18) * k04
-                    "vmlal.s16  q10, d11, %P11[0]    \n"
-                    "vmlal.s16  q7, d12, %P11[1]     \n" // (r12 - r19) * k05
-                    "vmlal.s16  q8, d13, %P11[1]     \n"
-
-                    // r2
-                    "vld2.s8    {d30-d31}, [%4]!     \n" // r2
-                    "vld2.s8    {d10-d11}, [%4]      \n"
-                    "vext.s8    d12, d30, d10, #1    \n"
-
-                    "vmovl.s8    q5, d31             \n" // r21
-                    "vmovl.s8    q15, d30            \n" // r20
-                    "vmovl.s8    q6, d12             \n" // r22
-
-                    // sum0
-                    "vmlal.s16  q7, d30, %P11[2]     \n" // (r20 - r27) * k06
-                    "vmlal.s16  q8, d31, %P11[2]     \n"
-                    "vmlal.s16  q9, d10, %P11[3]     \n" // (r21 - r28) * k07
-                    "vmlal.s16  q10, d11, %P11[3]    \n"
-                    "vmlal.s16  q7, d12, %P12[0]     \n" // (r22 - r29) * k08
-                    "vmlal.s16  q8, d13, %P12[0]     \n"
-
-                    "subs   %0, %0, #1               \n"
-
-                    // add and save
-                    "vadd.s32    q7, q7, q9          \n"
-                    "vadd.s32    q8, q8, q10         \n"
-
-                    "vst1.s32    {d14-d17}, [%1]!    \n"
-
-                    "bne    0b                       \n"
-
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0),     // %2
-                    "=r"(r1),     // %3
-                    "=r"(r2)      // %4
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "w"(_k0123), // %10
-                    "w"(_k4567), // %11
-                    "w"(_k8xxx)  // %12
-                    : "cc", "memory", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-                int sum = 0;
-
-                sum += (int)r0[0] * kernel[0];
-                sum += (int)r0[1] * kernel[1];
-                sum += (int)r0[2] * kernel[2];
-                sum += (int)r1[0] * kernel[3];
-                sum += (int)r1[1] * kernel[4];
-                sum += (int)r1[2] * kernel[5];
-                sum += (int)r2[0] * kernel[6];
-                sum += (int)r2[1] * kernel[7];
-                sum += (int)r2[2] * kernel[8];
-
-                *outptr = sum;
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-                outptr++;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
-
-static void convdw3x3s1_int8_requant_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, std::vector<float> scales_requant, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        const float scale_requant_in = scales_requant[2 * p];
-        const float scale_requant_out = scales_requant[2 * p + 1];
-
-        const signed char* kernel = (const signed char*)_kernel + p * 9;
-
-        signed char* outptr0 = out;
-        signed char* outptr0n = outptr0 + outw;
-
-        const signed char* img0 = bottom_blob.channel(p);
-
-        const signed char* r0 = img0;
-        const signed char* r1 = img0 + w;
-        const signed char* r2 = img0 + w * 2;
-        const signed char* r3 = img0 + w * 3;
-
-        int i = 0;
-
-#if __ARM_NEON
-        int8x16_t _k0123456789x = vld1q_s8(kernel);
-        int16x8_t _k_s16 = vmovl_s8(vget_low_s8(_k0123456789x));
-        int16x8_t _kn_s16 = vmovl_s8(vget_high_s8(_k0123456789x));
-
-        int16x4_t _k0123 = vget_low_s16(_k_s16);
-        int16x4_t _k4567 = vget_high_s16(_k_s16);
-        int16x4_t _k8xxx = vget_low_s16(_kn_s16);
-#endif // __ARM_NEON
-
-        for (; i + 1 < outh; i += 2)
-        {
-#if __ARM_NEON
-            int nn = outw >> 3;
-            int remain = outw & 7;
-#else
-            int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    "0:                                   \n"
-                    "ld1    {v4.8b, v5.8b}, [%3]          \n"
-                    "ld1    {v6.8b, v7.8b}, [%4]          \n"
-                    "ld1    {v8.8b, v9.8b}, [%5]          \n"
-                    "ld1    {v10.8b, v11.8b}, [%6]        \n"
-                    "add    %3, %3, #8                    \n"
-                    "add    %4, %4, #8                    \n"
-                    "add    %5, %5, #8                    \n"
-                    "add    %6, %6, #8                    \n"
-
-                    "ext    v12.8b, v4.8b, v5.8b, #1      \n"
-                    "ext    v13.8b, v4.8b, v5.8b, #2      \n"
-                    "ext    v14.8b, v6.8b, v7.8b, #1      \n"
-                    "ext    v15.8b, v6.8b, v7.8b, #2      \n"
-                    "ext    v16.8b, v8.8b, v9.8b, #1      \n"
-                    "ext    v17.8b, v8.8b, v9.8b, #2      \n"
-                    "ext    v18.8b, v10.8b, v11.8b, #1    \n"
-                    "ext    v19.8b, v10.8b, v11.8b, #2    \n"
-
-                    "sshll  v4.8h, v4.8b, #0              \n" // r00
-                    "sshll  v12.8h, v12.8b, #0            \n" // r01
-                    "sshll  v13.8h, v13.8b, #0            \n" // r02
-                    "sshll  v6.8h, v6.8b, #0              \n" // r10
-                    "sshll  v14.8h, v14.8b, #0            \n" // r11
-                    "sshll  v15.8h, v15.8b, #0            \n" // r12
-                    "sshll  v8.8h, v8.8b, #0              \n" // r20
-                    "sshll  v16.8h, v16.8b, #0            \n" // r21
-                    "sshll  v17.8h, v17.8b, #0            \n" // r22
-                    "sshll  v10.8h, v10.8b, #0            \n" // r30
-                    "sshll  v18.8h, v18.8b, #0            \n" // r31
-                    "sshll  v19.8h, v19.8b, #0            \n" // r32
-
-                    // r0
-                    "smull  v20.4s, v4.4h, %14.h[0]       \n" // (r00 - r07) * k00
-                    "smull2  v21.4s, v4.8h, %14.h[0]      \n"
-                    "smull  v22.4s, v12.4h, %14.h[1]      \n" // (r01 - r08) * k01
-                    "smull2  v23.4s, v12.8h, %14.h[1]     \n"
-                    "smull  v24.4s, v13.4h, %14.h[2]      \n" // (r02 - r09) * k02
-                    "smull2  v25.4s, v13.8h, %14.h[2]     \n"
-
-                    // r1
-                    "smull  v26.4s, v6.4h, %14.h[0]       \n" // (r10 - r17) * k00
-                    "smull2  v27.4s, v6.8h, %14.h[0]      \n"
-                    "smull  v28.4s, v14.4h, %14.h[1]      \n" // (r11 - r18) * k01
-                    "smull2  v29.4s, v14.8h, %14.h[1]     \n"
-                    "smull  v30.4s, v15.4h, %14.h[2]      \n" // (r12 - r19) * k02
-                    "smull2  v31.4s, v15.8h, %14.h[2]     \n"
-
-                    "smlal  v20.4s, v6.4h, %14.h[3]       \n" // (r10 - r17) * k03
-                    "smlal2  v21.4s, v6.8h, %14.h[3]      \n"
-                    "smlal  v22.4s, v14.4h, %15.h[0]      \n" // (r11 - r18) * k04
-                    "smlal2  v23.4s, v14.8h, %15.h[0]     \n"
-                    "smlal  v24.4s, v15.4h, %15.h[1]      \n" // (r12 - r19) * k05
-                    "smlal2  v25.4s, v15.8h, %15.h[1]     \n"
-
-                    // r2
-                    "smlal  v26.4s, v8.4h, %14.h[3]       \n" // (r20 - r27) * k03
-                    "smlal2  v27.4s, v8.8h, %14.h[3]      \n"
-                    "smlal  v28.4s, v16.4h, %15.h[0]      \n" // (r21 - r28) * k04
-                    "smlal2  v29.4s, v16.8h, %15.h[0]     \n"
-                    "smlal  v30.4s, v17.4h, %15.h[1]      \n" // (r22 - r29) * k05
-                    "smlal2  v31.4s, v17.8h, %15.h[1]     \n"
-
-                    "smlal  v20.4s, v8.4h, %15.h[2]       \n" // (r20 - r27) * k06
-                    "smlal2  v21.4s, v8.8h, %15.h[2]      \n"
-                    "smlal  v22.4s, v16.4h, %15.h[3]      \n" // (r21 - r28) * k07
-                    "smlal2  v23.4s, v16.8h, %15.h[3]     \n"
-                    "smlal  v24.4s, v17.4h, %16.h[0]      \n" // (r22 - r29) * k08
-                    "smlal2  v25.4s, v17.8h, %16.h[0]     \n"
-
-                    // r3
-                    "smlal  v26.4s, v10.4h, %15.h[2]      \n" // (r30 - r37) * k06
-                    "smlal2  v27.4s, v10.8h, %15.h[2]     \n"
-                    "smlal  v28.4s, v18.4h, %15.h[3]      \n" // (r31 - r38) * k07
-                    "smlal2  v29.4s, v18.8h, %15.h[3]     \n"
-                    "smlal  v30.4s, v19.4h, %16.h[0]      \n" // (r32 - r39) * k08
-                    "smlal2  v31.4s, v19.8h, %16.h[0]     \n"
-
-                    // add and save
-                    "add    v20.4s, v20.4s, v22.4s        \n"
-                    "add    v21.4s, v21.4s, v23.4s        \n"
-                    "add    v26.4s, v26.4s, v28.4s        \n"
-                    "add    v27.4s, v27.4s, v29.4s        \n"
-                    "add    v20.4s, v20.4s, v24.4s        \n"
-                    "add    v21.4s, v21.4s, v25.4s        \n"
-                    "add    v26.4s, v26.4s, v30.4s        \n"
-                    "add    v27.4s, v27.4s, v31.4s        \n"
-
-                    "dup    v4.4s, %w17                   \n" // bias
-                    "dup    v5.4s, %w18                   \n" // scale_in
-                    "dup    v6.4s, %w19                   \n" // scale_out
-
-                    // top_s32 -> top_f32
-                    "scvtf  v20.4s, v20.4s                 \n"
-                    "scvtf  v21.4s, v21.4s                 \n"
-                    "scvtf  v26.4s, v26.4s                 \n"
-                    "scvtf  v27.4s, v27.4s                 \n"
-
-                    // top_f32 = top_f32 * scale_in
-                    "fmul   v20.4s, v20.4s, v5.4s          \n"
-                    "fmul   v21.4s, v21.4s, v5.4s          \n"
-                    "fmul   v26.4s, v26.4s, v5.4s          \n"
-                    "fmul   v27.4s, v27.4s, v5.4s          \n"
-                    // top_f32 = top_f32 + bias
-                    "fadd   v20.4s, v20.4s, v4.4s          \n"
-                    "fadd   v21.4s, v21.4s, v4.4s          \n"
-                    "fadd   v26.4s, v26.4s, v4.4s          \n"
-                    "fadd   v27.4s, v27.4s, v4.4s          \n"
-                    // top_f32 = top_f32 * scale_out
-                    "fmul   v20.4s, v20.4s, v6.4s          \n"
-                    "fmul   v21.4s, v21.4s, v6.4s          \n"
-                    "fmul   v26.4s, v26.4s, v6.4s          \n"
-                    "fmul   v27.4s, v27.4s, v6.4s          \n"
-                    // top_f32 -> top_s32
-                    "fcvtas v20.4s, v20.4s                 \n"
-                    "fcvtas v21.4s, v21.4s                 \n"
-                    "fcvtas v26.4s, v26.4s                 \n"
-                    "fcvtas v27.4s, v27.4s                 \n"
-                    // top_s32 -> top_s16
-                    "sqxtn  v7.4h, v20.4s                 \n"
-                    "sqxtn  v9.4h, v26.4s                 \n"
-                    "sqxtn2 v7.8h, v21.4s                 \n"
-                    "sqxtn2 v9.8h, v27.4s                 \n"
-                    // top_s16 -> top_s8
-                    "sqxtn  v8.8b, v7.8h                  \n"
-                    "sqxtn  v10.8b, v9.8h                 \n"
-                    // save top_s8
-                    "st1    {v8.8b}, [%1], #8             \n"
-                    "st1    {v10.8b}, [%2], #8            \n"
-
-                    "subs   %w0, %w0, #1                  \n"
-                    "bne    0b                            \n"
-
-                    : "=r"(nn),       // %0
-                    "=r"(outptr0),  // %1
-                    "=r"(outptr0n), // %2
-                    "=r"(r0),       // %3
-                    "=r"(r1),       // %4
-                    "=r"(r2),       // %5
-                    "=r"(r3)        // %6
-                    : "0"(nn),
-                    "1"(outptr0),
-                    "2"(outptr0n),
-                    "3"(r0),
-                    "4"(r1),
-                    "5"(r2),
-                    "6"(r3),
-                    "w"(_k0123),           // %14
-                    "w"(_k4567),           // %15
-                    "w"(_k8xxx),           // %16
-                    "r"(bias0),            // %17
-                    "r"(scale_requant_in), // %18
-                    "r"(scale_requant_out) // %19
-                    : "cc", "memory", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    "0:                              \n"
-                    // r0
-                    "vld1.s8    {d30-d31}, [%3]      \n" // r0
-                    "add    %3, %3, #8               \n"
-
-                    "vext.s8    d10, d30, d31, #1    \n"
-                    "vext.s8    d12, d30, d31, #2    \n"
-
-                    "vmovl.s8    q15, d30            \n" // r00
-                    "vmovl.s8    q5, d10             \n" // r01
-                    "vmovl.s8    q6, d12             \n" // r02
-                    // sum0
-                    "vmull.s16  q7, d30, %P14[0]     \n" // (r00 - r07) * k00
-                    "vmull.s16  q8, d31, %P14[0]     \n"
-                    "vmull.s16  q9, d10, %P14[1]     \n" // (r01 - r08) * k01
-                    "vmull.s16  q10, d11, %P14[1]    \n"
-                    "vmlal.s16  q7, d12, %P14[2]     \n" // (r02 - r09) * k02
-                    "vmlal.s16  q8, d13, %P14[2]     \n"
-
-                    // r1
-                    "vld1.s8    {d30-d31}, [%4]      \n" // r1
-                    "add    %4, %4, #8               \n"
-
-                    "vext.s8    d10, d30, d31, #1    \n"
-                    "vext.s8    d12, d30, d31, #2    \n"
-
-                    "vmovl.s8    q15, d30            \n" // r10
-                    "vmovl.s8    q5, d10             \n" // r11
-                    "vmovl.s8    q6, d12             \n" // r12
-                    // sum0
-                    "vmlal.s16  q7, d30, %P14[3]     \n" // (r10 - r17) * k03
-                    "vmlal.s16  q8, d31, %P14[3]     \n"
-                    "vmlal.s16  q9, d10, %P15[0]     \n" // (r11 - r18) * k04
-                    "vmlal.s16  q10, d11, %P15[0]    \n"
-                    "vmlal.s16  q7, d12, %P15[1]     \n" // (r12 - r19) * k05
-                    "vmlal.s16  q8, d13, %P15[1]     \n"
-                    // sum1
-                    "vmull.s16  q11, d30, %P14[0]    \n" // (r10 - r17) * k00
-                    "vmull.s16  q12, d31, %P14[0]    \n"
-                    "vmull.s16  q13, d10, %P14[1]    \n" // (r11 - r18) * k01
-                    "vmull.s16  q14, d11, %P14[1]    \n"
-                    "vmlal.s16  q11, d12, %P14[2]    \n" // (r12 - r19) * k02
-                    "vmlal.s16  q12, d13, %P14[2]    \n"
-
-                    // r2
-                    "vld1.s8    {d30-d31}, [%5]      \n" // r2
-                    "add    %5, %5, #8               \n"
-
-                    "vext.s8    d10, d30, d31, #1    \n"
-                    "vext.s8    d12, d30, d31, #2    \n"
-
-                    "vmovl.s8    q15, d30            \n" // r20
-                    "vmovl.s8    q5, d10             \n" // r21
-                    "vmovl.s8    q6, d12             \n" // r22
-
-                    // sum0
-                    "vmlal.s16  q7, d30, %P15[2]     \n" // (r20 - r27) * k06
-                    "vmlal.s16  q8, d31, %P15[2]     \n"
-                    "vmlal.s16  q9, d10, %P15[3]     \n" // (r21 - r28) * k07
-                    "vmlal.s16  q10, d11, %P15[3]    \n"
-                    "vmlal.s16  q7, d12, %P16[0]     \n" // (r22 - r29) * k08
-                    "vmlal.s16  q8, d13, %P16[0]     \n"
-                    // sum1
-                    "vmlal.s16  q11, d30, %P14[3]    \n" // (r20 - r27) * k03
-                    "vmlal.s16  q12, d31, %P14[3]    \n"
-                    "vmlal.s16  q13, d10, %P15[0]    \n" // (r21 - r28) * k04
-                    "vmlal.s16  q14, d11, %P15[0]    \n"
-                    "vmlal.s16  q11, d12, %P15[1]    \n" // (r22 - r29) * k05
-                    "vmlal.s16  q12, d13, %P15[1]    \n"
-
-                    // r3
-                    "vld1.s8    {d30-d31}, [%6]      \n" // r3
-                    "add    %6, %6, #8               \n"
-
-                    "vext.s8    d10, d30, d31, #1    \n"
-                    "vext.s8    d12, d30, d31, #2    \n"
-
-                    "vmovl.s8    q15, d30            \n" // r30
-                    "vmovl.s8    q5, d10             \n" // r31
-                    "vmovl.s8    q6, d12             \n" // r32
-
-                    // sum1
-                    "vmlal.s16  q11, d30, %P15[2]    \n" // (r30 - r37) * k06
-                    "vmlal.s16  q12, d31, %P15[2]    \n"
-                    "vmlal.s16  q13, d10, %P15[3]    \n" // (r31 - r38) * k07
-                    "vmlal.s16  q14, d11, %P15[3]    \n"
-                    "vmlal.s16  q11, d12, %P16[0]    \n" // (r32 - r39) * k08
-                    "vmlal.s16  q12, d13, %P16[0]    \n"
-
-                    "subs   %0, %0, #1               \n"
-
-                    // add and save
-                    "vadd.s32    q7, q7, q9          \n"
-                    "vadd.s32    q8, q8, q10         \n"
-                    "vadd.s32    q11, q11, q13       \n"
-                    "vadd.s32    q12, q12, q14       \n"
-
-                    "vdup.f32   q13, %17             \n" // bias
-                    "vdup.f32   q14, %18             \n" // scale_in
-                    "vdup.f32   q15, %19             \n" // scale_out
-
-                    // top_s32 -> top_f32
-                    "vcvt.f32.s32 q7, q7            \n"
-                    "vcvt.f32.s32 q8, q8            \n"
-                    // top_f32 = top_f32 * scale_int
-                    "vmul.f32   q0, q7, q14         \n"
-                    "vmul.f32   q4, q8, q14         \n"
-                    // top_f32 = top_f32 + bias
-                    "vadd.f32   q0, q0, q13         \n"
-                    "vadd.f32   q4, q4, q13         \n"
-                    // top_f32 = top_f32 * scale_out
-                    "vmul.f32   q0, q0, q15         \n"
-                    "vmul.f32   q4, q4, q15         \n"
-                    // top_f32 -> top_s32
-                    "vcvtr.s32.f32 s0, s0           \n"
-                    "vcvtr.s32.f32 s1, s1           \n"
-                    "vcvtr.s32.f32 s2, s2           \n"
-                    "vcvtr.s32.f32 s3, s3           \n"
-                    "vcvtr.s32.f32 s16, s16           \n"
-                    "vcvtr.s32.f32 s17, s17           \n"
-                    "vcvtr.s32.f32 s18, s18           \n"
-                    "vcvtr.s32.f32 s19, s19           \n"
-                    // top_s32 -> top_s16
-                    "vqmovn.s32 d14, q0             \n"
-                    "vqmovn.s32 d15, q4             \n"
-                    // top_s16 -> top_s8
-                    "vqmovn.s16   d14, q7           \n"
-                    // save top_s8
-                    "vst1.8     {d14}, [%1]!        \n"
-
-                    // top_s32 -> top_f32
-                    "vcvt.f32.s32 q11, q11          \n"
-                    "vcvt.f32.s32 q12, q12          \n"
-                    // top_f32 = top_f32 * scale_int
-                    "vmul.f32   q0, q11, q14        \n"
-                    "vmul.f32   q4, q12, q14        \n"
-                    // top_f32 = top_f32 + bias
-                    "vadd.f32   q0, q0, q13         \n"
-                    "vadd.f32   q4, q4, q13         \n"
-                    // top_f32 = top_f32 * scale_out
-                    "vmul.f32   q0, q0, q15         \n"
-                    "vmul.f32   q4, q4, q15         \n"
-                    // top_f32 -> top_s32
-                    "vcvtr.s32.f32 s0, s0           \n"
-                    "vcvtr.s32.f32 s1, s1           \n"
-                    "vcvtr.s32.f32 s2, s2           \n"
-                    "vcvtr.s32.f32 s3, s3           \n"
-                    "vcvtr.s32.f32 s16, s16           \n"
-                    "vcvtr.s32.f32 s17, s17           \n"
-                    "vcvtr.s32.f32 s18, s18           \n"
-                    "vcvtr.s32.f32 s19, s19           \n"
-                    // top_s32 -> top_s16
-                    "vqmovn.s32 d14, q0             \n"
-                    "vqmovn.s32 d15, q4             \n"
-                    // top_s16 -> top_s8
-                    "vqmovn.s16   d14, q7           \n"
-                    // save top_s8
-                    "vst1.8     {d14}, [%2]!        \n"
-
-                    "bne    0b                      \n"
-
-                    : "=r"(nn),       // %0
-                    "=r"(outptr0),  // %1
-                    "=r"(outptr0n), // %2
-                    "=r"(r0),       // %3
-                    "=r"(r1),       // %4
-                    "=r"(r2),       // %5
-                    "=r"(r3)        // %6
-                    : "0"(nn),
-                    "1"(outptr0),
-                    "2"(outptr0n),
-                    "3"(r0),
-                    "4"(r1),
-                    "5"(r2),
-                    "6"(r3),
-                    "w"(_k0123),           // %14
-                    "w"(_k4567),           // %15
-                    "w"(_k8xxx),           // %16
-                    "r"(bias0),            // %17
-                    "r"(scale_requant_in), // %18
-                    "r"(scale_requant_out) // %19
-                    : "cc", "memory", "q0", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-                // TODO NEON
-                int sum0 = 0;
-                int sum0n = 0;
-
-                sum0 += (int)r0[0] * kernel[0];
-                sum0 += (int)r0[1] * kernel[1];
-                sum0 += (int)r0[2] * kernel[2];
-                sum0 += (int)r1[0] * kernel[3];
-                sum0 += (int)r1[1] * kernel[4];
-                sum0 += (int)r1[2] * kernel[5];
-                sum0 += (int)r2[0] * kernel[6];
-                sum0 += (int)r2[1] * kernel[7];
-                sum0 += (int)r2[2] * kernel[8];
-
-                sum0n += (int)r1[0] * kernel[0];
-                sum0n += (int)r1[1] * kernel[1];
-                sum0n += (int)r1[2] * kernel[2];
-                sum0n += (int)r2[0] * kernel[3];
-                sum0n += (int)r2[1] * kernel[4];
-                sum0n += (int)r2[2] * kernel[5];
-                sum0n += (int)r3[0] * kernel[6];
-                sum0n += (int)r3[1] * kernel[7];
-                sum0n += (int)r3[2] * kernel[8];
-
-                *outptr0 = float2int8(((float)sum0 * scale_requant_in + bias0) * scale_requant_out);
-                *outptr0n = float2int8(((float)sum0n * scale_requant_in + bias0) * scale_requant_out);
-
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                outptr0++;
-                outptr0n++;
-            }
-
-            r0 += 2 + w;
-            r1 += 2 + w;
-            r2 += 2 + w;
-            r3 += 2 + w;
-
-            outptr0 += outw;
-            outptr0n += outw;
-        }
-
-        for (; i < outh; i++)
-        {
-#if __ARM_NEON
-            int nn = outw >> 3;
-            int remain = outw & 7;
-#else
-            int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    "dup    v26.4s, %w13                  \n"
-                    "dup    v27.4s, %w14                  \n"
-                    "dup    v28.4s, %w15                  \n"
-
-                    "0:                                   \n"
-                    "ld1    {v4.8b, v5.8b}, [%2]          \n"
-                    "ld1    {v6.8b, v7.8b}, [%3]          \n"
-                    "ld1    {v8.8b, v9.8b}, [%4]          \n"
-                    "add    %2, %2, #8                    \n"
-                    "add    %3, %3, #8                    \n"
-                    "add    %4, %4, #8                    \n"
-
-                    "ext    v12.8b, v4.8b, v5.8b, #1      \n"
-                    "ext    v13.8b, v4.8b, v5.8b, #2      \n"
-                    "ext    v14.8b, v6.8b, v7.8b, #1      \n"
-                    "ext    v15.8b, v6.8b, v7.8b, #2      \n"
-                    "ext    v16.8b, v8.8b, v9.8b, #1      \n"
-                    "ext    v17.8b, v8.8b, v9.8b, #2      \n"
-
-                    "sshll  v4.8h, v4.8b, #0              \n" // r00
-                    "sshll  v12.8h, v12.8b, #0            \n" // r01
-                    "sshll  v13.8h, v13.8b, #0            \n" // r02
-                    "sshll  v6.8h, v6.8b, #0              \n" // r10
-                    "sshll  v14.8h, v14.8b, #0            \n" // r11
-                    "sshll  v15.8h, v15.8b, #0            \n" // r12
-                    "sshll  v8.8h, v8.8b, #0              \n" // r20
-                    "sshll  v16.8h, v16.8b, #0            \n" // r21
-                    "sshll  v17.8h, v17.8b, #0            \n" // r22
-
-                    // r0
-                    "smull  v20.4s, v4.4h, %10.h[0]       \n" // (r00 - r07) * k00
-                    "smull2  v21.4s, v4.8h, %10.h[0]      \n"
-                    "smull  v22.4s, v12.4h, %10.h[1]      \n" // (r01 - r08) * k01
-                    "smull2  v23.4s, v12.8h, %10.h[1]     \n"
-                    "smull  v24.4s, v13.4h, %10.h[2]      \n" // (r02 - r09) * k02
-                    "smull2  v25.4s, v13.8h, %10.h[2]     \n"
-
-                    // r1
-                    "smlal  v20.4s, v6.4h, %10.h[3]       \n" // (r10 - r17) * k03
-                    "smlal2  v21.4s, v6.8h, %10.h[3]      \n"
-                    "smlal  v22.4s, v14.4h, %11.h[0]      \n" // (r11 - r18) * k04
-                    "smlal2  v23.4s, v14.8h, %11.h[0]     \n"
-                    "smlal  v24.4s, v15.4h, %11.h[1]      \n" // (r12 - r19) * k05
-                    "smlal2  v25.4s, v15.8h, %11.h[1]     \n"
-
-                    // r2
-                    "smlal  v20.4s, v8.4h, %11.h[2]       \n" // (r20 - r27) * k06
-                    "smlal2  v21.4s, v8.8h, %11.h[2]      \n"
-                    "smlal  v22.4s, v16.4h, %11.h[3]      \n" // (r21 - r28) * k07
-                    "smlal2  v23.4s, v16.8h, %11.h[3]     \n"
-                    "smlal  v24.4s, v17.4h, %12.h[0]      \n" // (r22 - r29) * k08
-                    "smlal2  v25.4s, v17.8h, %12.h[0]     \n"
-
-                    // add and save
-                    "add    v20.4s, v20.4s, v22.4s        \n"
-                    "add    v21.4s, v21.4s, v23.4s        \n"
-                    "add    v20.4s, v20.4s, v24.4s        \n"
-                    "add    v21.4s, v21.4s, v25.4s        \n"
-
-                    // top_s32 -> top_f32
-                    "scvtf  v20.4s, v20.4s                \n"
-                    "scvtf  v21.4s, v21.4s                \n"
-                    // top_f32 = top_f32 * scale_in
-                    "fmul   v20.4s, v20.4s, v27.4s        \n"
-                    "fmul   v21.4s, v21.4s, v27.4s        \n"
-                    // top_f32 = top_f32 + bias
-                    "fadd   v20.4s, v20.4s, v26.4s        \n"
-                    "fadd   v21.4s, v21.4s, v26.4s        \n"
-                    // top_f32 = top_f32 * scale_out
-                    "fmul   v20.4s, v20.4s, v28.4s        \n"
-                    "fmul   v21.4s, v21.4s, v28.4s        \n"
-                    // top_f32 -> top_s32
-                    "fcvtas v20.4s, v20.4s                \n"
-                    "fcvtas v21.4s, v21.4s                \n"
-                    // top_s32 -> top_s16
-                    "sqxtn  v7.4h, v20.4s                 \n"
-                    "sqxtn2 v7.8h, v21.4s                 \n"
-                    // top_s16 -> top_s8
-                    "sqxtn  v8.8b, v7.8h                  \n"
-                    // save top_s8
-                    "st1    {v8.8b}, [%1], #8             \n"
-
-                    "subs   %w0, %w0, #1                  \n"
-                    "bne    0b                            \n"
-
-                    : "=r"(nn),      // %0
-                    "=r"(outptr0), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2)       // %4
-                    : "0"(nn),
-                    "1"(outptr0),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "w"(_k0123),           // %10
-                    "w"(_k4567),           // %11
-                    "w"(_k8xxx),           // %12
-                    "r"(bias0),            // %13
-                    "r"(scale_requant_in), // %14
-                    "r"(scale_requant_out) // %15
-                    : "cc", "memory", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    "0:                              \n"
-                    // r0
-                    "vld1.s8    {d30-d31}, [%2]      \n" // r0
-                    "add    %2, %2, #8               \n"
-
-                    "vext.s8    d10, d30, d31, #1    \n"
-                    "vext.s8    d12, d30, d31, #2    \n"
-
-                    "vmovl.s8    q15, d30            \n" // r00
-                    "vmovl.s8    q5, d10             \n" // r01
-                    "vmovl.s8    q6, d12             \n" // r02
-                    // sum0
-                    "vmull.s16  q7, d30, %P10[0]     \n" // (r00 - r07) * k00
-                    "vmull.s16  q8, d31, %P10[0]     \n"
-                    "vmull.s16  q9, d10, %P10[1]     \n" // (r01 - r08) * k01
-                    "vmull.s16  q10, d11, %P10[1]    \n"
-                    "vmlal.s16  q7, d12, %P10[2]     \n" // (r02 - r09) * k02
-                    "vmlal.s16  q8, d13, %P10[2]     \n"
-
-                    // r1
-                    "vld1.s8    {d30-d31}, [%3]      \n" // r1
-                    "add    %3, %3, #8               \n"
-
-                    "vext.s8    d10, d30, d31, #1    \n"
-                    "vext.s8    d12, d30, d31, #2    \n"
-
-                    "vmovl.s8    q15, d30            \n" // r10
-                    "vmovl.s8    q5, d10             \n" // r11
-                    "vmovl.s8    q6, d12             \n" // r12
-                    // sum0
-                    "vmlal.s16  q7, d30, %P10[3]     \n" // (r10 - r17) * k03
-                    "vmlal.s16  q8, d31, %P10[3]     \n"
-                    "vmlal.s16  q9, d10, %P11[0]     \n" // (r11 - r18) * k04
-                    "vmlal.s16  q10, d11, %P11[0]    \n"
-                    "vmlal.s16  q7, d12, %P11[1]     \n" // (r12 - r19) * k05
-                    "vmlal.s16  q8, d13, %P11[1]     \n"
-
-                    // r2
-                    "vld1.s8    {d30-d31}, [%4]      \n" // r2
-                    "add    %4, %4, #8               \n"
-
-                    "vext.s8    d10, d30, d31, #1    \n"
-                    "vext.s8    d12, d30, d31, #2    \n"
-
-                    "vmovl.s8    q15, d30            \n" // r20
-                    "vmovl.s8    q5, d10             \n" // r21
-                    "vmovl.s8    q6, d12             \n" // r22
-
-                    // sum0
-                    "vmlal.s16  q7, d30, %P11[2]     \n" // (r20 - r27) * k06
-                    "vmlal.s16  q8, d31, %P11[2]     \n"
-                    "vmlal.s16  q9, d10, %P11[3]     \n" // (r21 - r28) * k07
-                    "vmlal.s16  q10, d11, %P11[3]    \n"
-                    "vmlal.s16  q7, d12, %P12[0]     \n" // (r22 - r29) * k08
-                    "vmlal.s16  q8, d13, %P12[0]     \n"
-
-                    "subs   %0, %0, #1               \n"
-
-                    // add and save
-                    "vadd.s32    q7, q7, q9          \n"
-                    "vadd.s32    q8, q8, q10         \n"
-
-                    "vdup.f32   q13, %13             \n" // bias
-                    "vdup.f32   q14, %14             \n" // scale_in
-                    "vdup.f32   q15, %15             \n" // scale_out
-
-                    // top_s32 -> top_f32
-                    "vcvt.f32.s32 q7, q7            \n"
-                    "vcvt.f32.s32 q8, q8            \n"
-                    // top_f32 = top_f32 * scale_int
-                    "vmul.f32   q0, q7, q14         \n"
-                    "vmul.f32   q4, q8, q14         \n"
-                    // top_f32 = top_f32 + bias
-                    "vadd.f32   q0, q0, q13         \n"
-                    "vadd.f32   q4, q4, q13         \n"
-                    // top_f32 = top_f32 * scale_out
-                    "vmul.f32   q0, q0, q15         \n"
-                    "vmul.f32   q4, q4, q15         \n"
-                    // top_f32 -> top_s32
-                    "vcvtr.s32.f32 s0, s0           \n"
-                    "vcvtr.s32.f32 s1, s1           \n"
-                    "vcvtr.s32.f32 s2, s2           \n"
-                    "vcvtr.s32.f32 s3, s3           \n"
-                    "vcvtr.s32.f32 s16, s16           \n"
-                    "vcvtr.s32.f32 s17, s17           \n"
-                    "vcvtr.s32.f32 s18, s18           \n"
-                    "vcvtr.s32.f32 s19, s19           \n"
-                    // top_s32 -> top_s16
-                    "vqmovn.s32 d14, q0             \n"
-                    "vqmovn.s32 d15, q4             \n"
-                    // top_s16 -> top_s8
-                    "vqmovn.s16   d14, q7           \n"
-                    // save top_s8
-                    "vst1.8     {d14}, [%1]!        \n"
-
-                    "bne    0b                      \n"
-
-                    : "=r"(nn),      // %0
-                    "=r"(outptr0), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2)       // %4
-                    : "0"(nn),
-                    "1"(outptr0),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "w"(_k0123),           // %10
-                    "w"(_k4567),           // %11
-                    "w"(_k8xxx),           // %12
-                    "r"(bias0),            // %13
-                    "r"(scale_requant_in), // %14
-                    "r"(scale_requant_out) // %15
-                    : "cc", "memory", "q0", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-                int sum = 0;
-
-                sum += (int)r0[0] * kernel[0];
-                sum += (int)r0[1] * kernel[1];
-                sum += (int)r0[2] * kernel[2];
-                sum += (int)r1[0] * kernel[3];
-                sum += (int)r1[1] * kernel[4];
-                sum += (int)r1[2] * kernel[5];
-                sum += (int)r2[0] * kernel[6];
-                sum += (int)r2[1] * kernel[7];
-                sum += (int)r2[2] * kernel[8];
-
-                *outptr0 = float2int8(((float)sum * scale_requant_in + bias0) * scale_requant_out);
-
-                r0++;
-                r1++;
-                r2++;
-                outptr0++;
-            }
-
-            r0 += 2;
-            r1 += 2;
-            r2 += 2;
-        }
-    }
-}
-
-static void convdw3x3s2_int8_requant_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, std::vector<float> scales_requant, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    const int tailstep = w - 2 * outw + w;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        const float scale_requant_in = scales_requant[2 * p];
-        const float scale_requant_out = scales_requant[2 * p + 1];
-
-        const signed char* kernel = (const signed char*)_kernel + p * 9;
-
-        signed char* outptr = out;
-
-        const signed char* img = bottom_blob.channel(p);
-
-        const signed char* r0 = img;
-        const signed char* r1 = img + w;
-        const signed char* r2 = img + w * 2;
-
-        int i = 0;
-#if __ARM_NEON
-        int8x16_t _k0123456789x = vld1q_s8(kernel);
-        int16x8_t _k_s16 = vmovl_s8(vget_low_s8(_k0123456789x));
-        int16x8_t _kn_s16 = vmovl_s8(vget_high_s8(_k0123456789x));
-
-        int16x4_t _k0123 = vget_low_s16(_k_s16);
-        int16x4_t _k4567 = vget_high_s16(_k_s16);
-        int16x4_t _k8xxx = vget_low_s16(_kn_s16);
-#endif // __ARM_NEON
-        for (; i < outh; i++)
-        {
-#if __ARM_NEON
-            int nn = outw >> 3;
-            int remain = outw & 7;
-#else
-            int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    "dup    v26.4s, %w13                  \n"
-                    "dup    v27.4s, %w14                  \n"
-                    "dup    v28.4s, %w15                  \n"
-                    "0:                                   \n"
-                    "ld2    {v4.8b, v5.8b}, [%2], #16     \n"
-                    "ld2    {v6.8b, v7.8b}, [%2]          \n"
-                    "ld2    {v8.8b, v9.8b}, [%3], #16     \n"
-                    "ld2    {v10.8b, v11.8b}, [%3]        \n"
-                    "ld2    {v12.8b, v13.8b}, [%4], #16   \n"
-                    "ld2    {v14.8b, v15.8b}, [%4]        \n"
-
-                    "ext    v6.8b, v4.8b, v6.8b, #1       \n"
-                    "ext    v10.8b, v8.8b, v10.8b, #1     \n"
-                    "ext    v14.8b, v12.8b, v14.8b, #1    \n"
-
-                    "sshll  v4.8h, v4.8b, #0              \n" // r00
-                    "sshll  v5.8h, v5.8b, #0              \n" // r01
-                    "sshll  v6.8h, v6.8b, #0              \n" // r02
-                    "sshll  v8.8h, v8.8b, #0              \n" // r10
-                    "sshll  v9.8h, v9.8b, #0              \n" // r11
-                    "sshll  v10.8h, v10.8b, #0            \n" // r12
-                    "sshll  v12.8h, v12.8b, #0            \n" // r20
-                    "sshll  v13.8h, v13.8b, #0            \n" // r21
-                    "sshll  v14.8h, v14.8b, #0            \n" // r22
-
-                    // r0
-                    "smull  v20.4s, v4.4h, %10.h[0]       \n" // (r00 - r07) * k00
-                    "smull2  v21.4s, v4.8h, %10.h[0]      \n"
-                    "smull  v22.4s, v5.4h, %10.h[1]       \n" // (r01 - r08) * k01
-                    "smull2  v23.4s, v5.8h, %10.h[1]      \n"
-                    "smull  v24.4s, v6.4h, %10.h[2]       \n" // (r02 - r09) * k02
-                    "smull2  v25.4s, v6.8h, %10.h[2]      \n"
-
-                    // r1
-                    "smlal  v20.4s, v8.4h, %10.h[3]       \n" // (r10 - r17) * k03
-                    "smlal2  v21.4s, v8.8h, %10.h[3]      \n"
-                    "smlal  v22.4s, v9.4h, %11.h[0]       \n" // (r11 - r18) * k04
-                    "smlal2  v23.4s, v9.8h, %11.h[0]      \n"
-                    "smlal  v24.4s, v10.4h, %11.h[1]      \n" // (r12 - r19) * k05
-                    "smlal2  v25.4s, v10.8h, %11.h[1]     \n"
-
-                    // r2
-                    "smlal  v20.4s, v12.4h, %11.h[2]      \n" // (r20 - r27) * k06
-                    "smlal2  v21.4s, v12.8h, %11.h[2]     \n"
-                    "smlal  v22.4s, v13.4h, %11.h[3]      \n" // (r21 - r28) * k07
-                    "smlal2  v23.4s, v13.8h, %11.h[3]     \n"
-                    "smlal  v24.4s, v14.4h, %12.h[0]      \n" // (r22 - r29) * k08
-                    "smlal2  v25.4s, v14.8h, %12.h[0]     \n"
-
-                    // add and save
-                    "add    v20.4s, v20.4s, v22.4s        \n"
-                    "add    v21.4s, v21.4s, v23.4s        \n"
-                    "add    v20.4s, v20.4s, v24.4s        \n"
-                    "add    v21.4s, v21.4s, v25.4s        \n"
-
-                    // top_s32 -> top_f32
-                    "scvtf  v20.4s, v20.4s                \n"
-                    "scvtf  v21.4s, v21.4s                \n"
-                    // top_f32 = top_f32 * scale_in
-                    "fmul   v20.4s, v20.4s, v27.4s        \n"
-                    "fmul   v21.4s, v21.4s, v27.4s        \n"
-                    // top_f32 = top_f32 + bias
-                    "fadd   v20.4s, v20.4s, v26.4s        \n"
-                    "fadd   v21.4s, v21.4s, v26.4s        \n"
-                    // top_f32 = top_f32 * scale_out
-                    "fmul   v20.4s, v20.4s, v28.4s        \n"
-                    "fmul   v21.4s, v21.4s, v28.4s        \n"
-                    // top_f32 -> top_s32
-                    "fcvtas v20.4s, v20.4s                \n"
-                    "fcvtas v21.4s, v21.4s                \n"
-                    // top_s32 -> top_s16
-                    "sqxtn  v7.4h, v20.4s                 \n"
-                    "sqxtn2 v7.8h, v21.4s                 \n"
-                    // top_s16 -> top_s8
-                    "sqxtn  v8.8b, v7.8h                  \n"
-                    // save top_s8
-                    "st1    {v8.8b}, [%1], #8             \n"
-
-                    "subs   %w0, %w0, #1                  \n"
-                    "bne    0b                            \n"
-
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0),     // %2
-                    "=r"(r1),     // %3
-                    "=r"(r2)      // %4
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "w"(_k0123),           // %10
-                    "w"(_k4567),           // %11
-                    "w"(_k8xxx),           // %12
-                    "r"(bias0),            // %13
-                    "r"(scale_requant_in), // %14
-                    "r"(scale_requant_out) // %15
-                    : "cc", "memory", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    "0:                              \n"
-                    // r0
-                    "vld2.s8    {d30-d31}, [%2]!     \n" // r0
-                    "vld2.s8    {d10-d11}, [%2]      \n"
-                    "vext.s8    d12, d30, d10, #1    \n"
-
-                    "vmovl.s8    q5, d31             \n" // r01
-                    "vmovl.s8    q15, d30            \n" // r00
-                    "vmovl.s8    q6, d12             \n" // r02
-                    // sum0
-                    "vmull.s16  q7, d30, %P10[0]     \n" // (r00 - r07) * k00
-                    "vmull.s16  q8, d31, %P10[0]     \n"
-                    "vmull.s16  q9, d10, %P10[1]     \n" // (r01 - r08) * k01
-                    "vmull.s16  q10, d11, %P10[1]    \n"
-                    "vmlal.s16  q7, d12, %P10[2]     \n" // (r02 - r09) * k02
-                    "vmlal.s16  q8, d13, %P10[2]     \n"
-
-                    // r1
-                    "vld2.s8    {d30-d31}, [%3]!     \n" // r1
-                    "vld2.s8    {d10-d11}, [%3]      \n"
-                    "vext.s8    d12, d30, d10, #1    \n"
-
-                    "vmovl.s8    q5, d31             \n" // r11
-                    "vmovl.s8    q15, d30            \n" // r10
-                    "vmovl.s8    q6, d12             \n" // r12
-                    // sum0
-                    "vmlal.s16  q7, d30, %P10[3]     \n" // (r10 - r17) * k03
-                    "vmlal.s16  q8, d31, %P10[3]     \n"
-                    "vmlal.s16  q9, d10, %P11[0]     \n" // (r11 - r18) * k04
-                    "vmlal.s16  q10, d11, %P11[0]    \n"
-                    "vmlal.s16  q7, d12, %P11[1]     \n" // (r12 - r19) * k05
-                    "vmlal.s16  q8, d13, %P11[1]     \n"
-
-                    // r2
-                    "vld2.s8    {d30-d31}, [%4]!     \n" // r2
-                    "vld2.s8    {d10-d11}, [%4]      \n"
-                    "vext.s8    d12, d30, d10, #1    \n"
-
-                    "vmovl.s8    q5, d31             \n" // r21
-                    "vmovl.s8    q15, d30            \n" // r20
-                    "vmovl.s8    q6, d12             \n" // r22
-
-                    // sum0
-                    "vmlal.s16  q7, d30, %P11[2]     \n" // (r20 - r27) * k06
-                    "vmlal.s16  q8, d31, %P11[2]     \n"
-                    "vmlal.s16  q9, d10, %P11[3]     \n" // (r21 - r28) * k07
-                    "vmlal.s16  q10, d11, %P11[3]    \n"
-                    "vmlal.s16  q7, d12, %P12[0]     \n" // (r22 - r29) * k08
-                    "vmlal.s16  q8, d13, %P12[0]     \n"
-
-                    "subs   %0, %0, #1               \n"
-
-                    // add and save
-                    "vadd.s32    q7, q7, q9          \n"
-                    "vadd.s32    q8, q8, q10         \n"
-
-                    "vdup.f32   q11, %13             \n" // bias
-                    "vdup.f32   q12, %14             \n" // scale_in
-                    "vdup.f32   q13, %15             \n" // scale_out
-
-                    // top_s32 -> top_f32
-                    "vcvt.f32.s32 q7, q7             \n"
-                    "vcvt.f32.s32 q8, q8             \n"
-                    // top_f32 = top_f32 * scale_int
-                    "vmul.f32   q0, q7, q12          \n"
-                    "vmul.f32   q4, q8, q12          \n"
-                    // top_f32 = top_f32 + bias
-                    "vadd.f32   q0, q0, q11          \n"
-                    "vadd.f32   q4, q4, q11          \n"
-                    // top_f32 = top_f32 * scale_out
-                    "vmul.f32   q0, q0, q13          \n"
-                    "vmul.f32   q4, q4, q13          \n"
-                    // top_f32 -> top_s32
-                    "vcvtr.s32.f32 s0, s0            \n"
-                    "vcvtr.s32.f32 s1, s1            \n"
-                    "vcvtr.s32.f32 s2, s2            \n"
-                    "vcvtr.s32.f32 s3, s3            \n"
-                    "vcvtr.s32.f32 s16, s16            \n"
-                    "vcvtr.s32.f32 s17, s17            \n"
-                    "vcvtr.s32.f32 s18, s18            \n"
-                    "vcvtr.s32.f32 s19, s19            \n"
-                    // top_s32 -> top_s16
-                    "vqmovn.s32 d14, q0              \n"
-                    "vqmovn.s32 d15, q4              \n"
-                    // top_s16 -> top_s8
-                    "vqmovn.s16   d14, q7            \n"
-                    // save top_s8
-                    "vst1.8     {d14}, [%1]!         \n"
-
-                    "bne    0b                       \n"
-
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0),     // %2
-                    "=r"(r1),     // %3
-                    "=r"(r2)      // %4
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "w"(_k0123),           // %10
-                    "w"(_k4567),           // %11
-                    "w"(_k8xxx),           // %12
-                    "r"(bias0),            // %13
-                    "r"(scale_requant_in), // %14
-                    "r"(scale_requant_out) // %15
-                    : "cc", "memory", "q0", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-                int sum = 0;
-
-                sum += (int)r0[0] * kernel[0];
-                sum += (int)r0[1] * kernel[1];
-                sum += (int)r0[2] * kernel[2];
-                sum += (int)r1[0] * kernel[3];
-                sum += (int)r1[1] * kernel[4];
-                sum += (int)r1[2] * kernel[5];
-                sum += (int)r2[0] * kernel[6];
-                sum += (int)r2[1] * kernel[7];
-                sum += (int)r2[2] * kernel[8];
-
-                *outptr = float2int8(((float)sum * scale_requant_in + bias0) * scale_requant_out);
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-                outptr++;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack4.h
deleted file mode 100644
index c3089111..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack4.h
+++ /dev/null
@@ -1,1290 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_pack4_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-#if __aarch64__
-    const int w = bottom_blob.w;
-#endif
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + g * 4) : vdupq_n_f32(0.f);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-
-        float32x4_t _k00 = vld1q_f32(k0);
-        float32x4_t _k01 = vld1q_f32(k0 + 4);
-        float32x4_t _k02 = vld1q_f32(k0 + 8);
-        float32x4_t _k10 = vld1q_f32(k0 + 12);
-        float32x4_t _k11 = vld1q_f32(k0 + 16);
-        float32x4_t _k12 = vld1q_f32(k0 + 20);
-        float32x4_t _k20 = vld1q_f32(k0 + 24);
-        float32x4_t _k21 = vld1q_f32(k0 + 28);
-        float32x4_t _k22 = vld1q_f32(k0 + 32);
-
-        int i = 0;
-
-#if __aarch64__
-        float* outptr1 = out.row(1);
-        const float* r3 = img0.row(3);
-
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-
-            for (; j + 3 < outw; j += 4)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v10.4s, v11.4s}, [%3], #32 \n" // r10 r11
-
-                    "mov    v16.16b, %21.16b            \n" // sum00
-                    "mov    v17.16b, %21.16b            \n" // sum01
-                    "mov    v18.16b, %21.16b            \n" // sum02
-                    "mov    v19.16b, %21.16b            \n" // sum03
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%3] \n" // r12 r13 r14 r15
-
-                    "mov    v20.16b, %21.16b            \n" // sum10
-                    "mov    v21.16b, %21.16b            \n" // sum11
-                    "mov    v22.16b, %21.16b            \n" // sum12
-                    "mov    v23.16b, %21.16b            \n" // sum13
-
-                    "fmla   v16.4s, %15.4s, v10.4s      \n"
-                    "fmla   v17.4s, %15.4s, v11.4s      \n"
-                    "fmla   v18.4s, %15.4s, v12.4s      \n"
-                    "fmla   v19.4s, %15.4s, v13.4s      \n"
-                    "fmla   v20.4s, %12.4s, v10.4s      \n"
-                    "fmla   v21.4s, %12.4s, v11.4s      \n"
-                    "fmla   v22.4s, %12.4s, v12.4s      \n"
-                    "fmla   v23.4s, %12.4s, v13.4s      \n"
-
-                    "add    %3, %3, #32                 \n"
-
-                    "fmla   v16.4s, %16.4s, v11.4s      \n"
-                    "fmla   v17.4s, %16.4s, v12.4s      \n"
-                    "fmla   v18.4s, %16.4s, v13.4s      \n"
-                    "fmla   v19.4s, %16.4s, v14.4s      \n"
-                    "fmla   v20.4s, %13.4s, v11.4s      \n"
-                    "fmla   v21.4s, %13.4s, v12.4s      \n"
-                    "fmla   v22.4s, %13.4s, v13.4s      \n"
-                    "fmla   v23.4s, %13.4s, v14.4s      \n"
-
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v10.4s, v11.4s}, [%4], #32 \n" // r20 r21
-
-                    "fmla   v16.4s, %17.4s, v12.4s      \n"
-                    "fmla   v17.4s, %17.4s, v13.4s      \n"
-                    "fmla   v18.4s, %17.4s, v14.4s      \n"
-                    "fmla   v19.4s, %17.4s, v15.4s      \n"
-                    "fmla   v20.4s, %14.4s, v12.4s      \n"
-                    "fmla   v21.4s, %14.4s, v13.4s      \n"
-                    "fmla   v22.4s, %14.4s, v14.4s      \n"
-                    "fmla   v23.4s, %14.4s, v15.4s      \n"
-
-                    "prfm   pldl1keep, [%4, #512]       \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%4] \n" // r22 r23 r24 r25
-
-                    "fmla   v16.4s, %18.4s, v10.4s      \n"
-                    "fmla   v17.4s, %18.4s, v11.4s      \n"
-                    "fmla   v18.4s, %18.4s, v12.4s      \n"
-                    "fmla   v19.4s, %18.4s, v13.4s      \n"
-                    "fmla   v20.4s, %15.4s, v10.4s      \n"
-                    "fmla   v21.4s, %15.4s, v11.4s      \n"
-                    "fmla   v22.4s, %15.4s, v12.4s      \n"
-                    "fmla   v23.4s, %15.4s, v13.4s      \n"
-
-                    "add    %4, %4, #32                 \n"
-
-                    "fmla   v16.4s, %19.4s, v11.4s      \n"
-                    "fmla   v17.4s, %19.4s, v12.4s      \n"
-                    "fmla   v18.4s, %19.4s, v13.4s      \n"
-                    "fmla   v19.4s, %19.4s, v14.4s      \n"
-                    "fmla   v20.4s, %16.4s, v11.4s      \n"
-                    "fmla   v21.4s, %16.4s, v12.4s      \n"
-                    "fmla   v22.4s, %16.4s, v13.4s      \n"
-                    "fmla   v23.4s, %16.4s, v14.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v10.4s, v11.4s}, [%2], #32 \n" // r00 r01
-
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v24.4s, v25.4s}, [%5], #32 \n" // r30 r31
-
-                    "fmla   v16.4s, %20.4s, v12.4s      \n"
-                    "fmla   v17.4s, %20.4s, v13.4s      \n"
-                    "fmla   v18.4s, %20.4s, v14.4s      \n"
-                    "fmla   v19.4s, %20.4s, v15.4s      \n"
-                    "fmla   v20.4s, %17.4s, v12.4s      \n"
-                    "fmla   v21.4s, %17.4s, v13.4s      \n"
-                    "fmla   v22.4s, %17.4s, v14.4s      \n"
-                    "fmla   v23.4s, %17.4s, v15.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%2] \n" // r02 r03 r04 r05
-
-                    "prfm   pldl1keep, [%5, #512]       \n"
-                    "ld1    {v26.4s, v27.4s, v28.4s, v29.4s}, [%5] \n" // r32 r33 r34 r35
-
-                    "fmla   v16.4s, %12.4s, v10.4s      \n"
-                    "fmla   v17.4s, %12.4s, v11.4s      \n"
-                    "fmla   v18.4s, %12.4s, v12.4s      \n"
-                    "fmla   v19.4s, %12.4s, v13.4s      \n"
-                    "fmla   v20.4s, %18.4s, v24.4s      \n"
-                    "fmla   v21.4s, %18.4s, v25.4s      \n"
-                    "fmla   v22.4s, %18.4s, v26.4s      \n"
-                    "fmla   v23.4s, %18.4s, v27.4s      \n"
-
-                    "add    %2, %2, #32                 \n"
-
-                    "fmla   v16.4s, %13.4s, v11.4s      \n"
-                    "fmla   v17.4s, %13.4s, v12.4s      \n"
-                    "fmla   v18.4s, %13.4s, v13.4s      \n"
-                    "fmla   v19.4s, %13.4s, v14.4s      \n"
-                    "fmla   v20.4s, %19.4s, v25.4s      \n"
-                    "fmla   v21.4s, %19.4s, v26.4s      \n"
-                    "fmla   v22.4s, %19.4s, v27.4s      \n"
-                    "fmla   v23.4s, %19.4s, v28.4s      \n"
-
-                    "add    %5, %5, #32                 \n"
-
-                    "fmla   v16.4s, %14.4s, v12.4s      \n"
-                    "fmla   v17.4s, %14.4s, v13.4s      \n"
-                    "fmla   v18.4s, %14.4s, v14.4s      \n"
-                    "fmla   v19.4s, %14.4s, v15.4s      \n"
-                    "fmla   v20.4s, %20.4s, v26.4s      \n"
-                    "fmla   v21.4s, %20.4s, v27.4s      \n"
-                    "fmla   v22.4s, %20.4s, v28.4s      \n"
-                    "fmla   v23.4s, %20.4s, v29.4s      \n"
-
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%1], #64 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3)       // %5
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "w"(_k00),  // %12
-                    "w"(_k01),  // %13
-                    "w"(_k02),  // %14
-                    "w"(_k10),  // %15
-                    "w"(_k11),  // %16
-                    "w"(_k12),  // %17
-                    "w"(_k20),  // %18
-                    "w"(_k21),  // %19
-                    "w"(_k22),  // %20
-                    "w"(_bias0) // %21
-                    : "memory", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29");
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%3] \n" // r10 r11 r12 r13
-
-                    "mov    v16.16b, %21.16b            \n" // sum00
-                    "mov    v17.16b, %21.16b            \n" // sum01
-                    "mov    v18.16b, %21.16b            \n" // sum10
-                    "mov    v19.16b, %21.16b            \n" // sum11
-
-                    "fmla   v16.4s, %15.4s, v10.4s      \n"
-                    "fmla   v17.4s, %15.4s, v11.4s      \n"
-                    "fmla   v18.4s, %12.4s, v10.4s      \n"
-                    "fmla   v19.4s, %12.4s, v11.4s      \n"
-
-                    "add    %3, %3, #32                 \n"
-
-                    "fmla   v16.4s, %16.4s, v11.4s      \n"
-                    "fmla   v17.4s, %16.4s, v12.4s      \n"
-                    "fmla   v18.4s, %13.4s, v11.4s      \n"
-                    "fmla   v19.4s, %13.4s, v12.4s      \n"
-
-                    "prfm   pldl1keep, [%4, #512]       \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%4] \n" // r20 r21 r22 r23
-
-                    "fmla   v16.4s, %17.4s, v12.4s      \n"
-                    "fmla   v17.4s, %17.4s, v13.4s      \n"
-                    "fmla   v18.4s, %14.4s, v12.4s      \n"
-                    "fmla   v19.4s, %14.4s, v13.4s      \n"
-
-                    "add    %4, %4, #32                 \n"
-
-                    "fmla   v16.4s, %18.4s, v20.4s      \n"
-                    "fmla   v17.4s, %18.4s, v21.4s      \n"
-                    "fmla   v18.4s, %15.4s, v20.4s      \n"
-                    "fmla   v19.4s, %15.4s, v21.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%2] \n" // r00 r01 r02 r03
-
-                    "fmla   v16.4s, %19.4s, v21.4s      \n"
-                    "fmla   v17.4s, %19.4s, v22.4s      \n"
-                    "fmla   v18.4s, %16.4s, v21.4s      \n"
-                    "fmla   v19.4s, %16.4s, v22.4s      \n"
-
-                    "prfm   pldl1keep, [%5, #512]       \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%5] \n" // r30 r31 r32 r33
-
-                    "fmla   v16.4s, %20.4s, v22.4s      \n"
-                    "fmla   v17.4s, %20.4s, v23.4s      \n"
-                    "fmla   v18.4s, %17.4s, v22.4s      \n"
-                    "fmla   v19.4s, %17.4s, v23.4s      \n"
-
-                    "add    %2, %2, #32                 \n"
-
-                    "fmla   v16.4s, %12.4s, v10.4s      \n"
-                    "fmla   v17.4s, %12.4s, v11.4s      \n"
-                    "fmla   v18.4s, %18.4s, v24.4s      \n"
-                    "fmla   v19.4s, %18.4s, v25.4s      \n"
-
-                    "add    %5, %5, #32                 \n"
-
-                    "fmla   v16.4s, %13.4s, v11.4s      \n"
-                    "fmla   v17.4s, %13.4s, v12.4s      \n"
-                    "fmla   v18.4s, %19.4s, v25.4s      \n"
-                    "fmla   v19.4s, %19.4s, v26.4s      \n"
-
-                    "fmla   v16.4s, %14.4s, v12.4s      \n"
-                    "fmla   v17.4s, %14.4s, v13.4s      \n"
-                    "fmla   v18.4s, %20.4s, v26.4s      \n"
-                    "fmla   v19.4s, %20.4s, v27.4s      \n"
-
-                    "st1    {v16.4s, v17.4s}, [%0], #32 \n"
-                    "st1    {v18.4s, v19.4s}, [%1], #32 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3)       // %5
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "w"(_k00),  // %12
-                    "w"(_k01),  // %13
-                    "w"(_k02),  // %14
-                    "w"(_k10),  // %15
-                    "w"(_k11),  // %16
-                    "w"(_k12),  // %17
-                    "w"(_k20),  // %18
-                    "w"(_k21),  // %19
-                    "w"(_k22),  // %20
-                    "w"(_bias0) // %21
-                    : "memory", "v10", "v11", "v12", "v13", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-            }
-            for (; j < outw; j++)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%3, #384]       \n"
-                    "ld1    {v10.4s, v11.4s, v12.4s}, [%3] \n" // r10 r11 r12
-
-                    "mov    v16.16b, %21.16b            \n" // sum0
-                    "mov    v17.16b, %21.16b            \n" // sum1
-
-                    "fmla   v16.4s, %15.4s, v10.4s      \n"
-                    "fmla   v17.4s, %12.4s, v10.4s      \n"
-
-                    "add    %3, %3, #16                 \n"
-
-                    "fmla   v16.4s, %16.4s, v11.4s      \n"
-                    "fmla   v17.4s, %13.4s, v11.4s      \n"
-
-                    "prfm   pldl1keep, [%4, #384]       \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s}, [%4] \n" // r20 r21 r22
-
-                    "fmla   v16.4s, %17.4s, v12.4s      \n"
-                    "fmla   v17.4s, %14.4s, v12.4s      \n"
-
-                    "add    %4, %4, #16                 \n"
-
-                    "fmla   v16.4s, %18.4s, v20.4s      \n"
-                    "fmla   v17.4s, %15.4s, v20.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #384]       \n"
-                    "ld1    {v10.4s, v11.4s, v12.4s}, [%2] \n" // r00 r01 r02
-
-                    "fmla   v16.4s, %19.4s, v21.4s      \n"
-                    "fmla   v17.4s, %16.4s, v21.4s      \n"
-
-                    "prfm   pldl1keep, [%5, #384]       \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s}, [%5] \n" // r30 r31 r32
-
-                    "fmla   v16.4s, %20.4s, v22.4s      \n"
-                    "fmla   v17.4s, %17.4s, v22.4s      \n"
-
-                    "add    %2, %2, #16                 \n"
-
-                    "fmla   v16.4s, %12.4s, v10.4s      \n"
-                    "fmla   v17.4s, %18.4s, v24.4s      \n"
-
-                    "add    %5, %5, #16                 \n"
-
-                    "fmla   v16.4s, %13.4s, v11.4s      \n"
-                    "fmla   v17.4s, %19.4s, v25.4s      \n"
-
-                    "fmla   v16.4s, %14.4s, v12.4s      \n"
-                    "fmla   v17.4s, %20.4s, v26.4s      \n"
-
-                    "st1    {v16.4s}, [%0], #16         \n"
-                    "st1    {v17.4s}, [%1], #16         \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3)       // %5
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "w"(_k00),  // %12
-                    "w"(_k01),  // %13
-                    "w"(_k02),  // %14
-                    "w"(_k10),  // %15
-                    "w"(_k11),  // %16
-                    "w"(_k12),  // %17
-                    "w"(_k20),  // %18
-                    "w"(_k21),  // %19
-                    "w"(_k22),  // %20
-                    "w"(_bias0) // %21
-                    : "memory", "v10", "v11", "v12", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v24", "v25", "v26");
-            }
-
-            r0 += 2 * 4 + w * 4;
-            r1 += 2 * 4 + w * 4;
-            r2 += 2 * 4 + w * 4;
-            r3 += 2 * 4 + w * 4;
-
-            outptr0 += outw * 4;
-            outptr1 += outw * 4;
-        }
-#endif // __aarch64__
-        for (; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j + 3 < outw; j += 4)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v10.4s, v11.4s}, [%1], #32 \n" // r00 r01
-
-                    "mov    v16.16b, %17.16b            \n" // sum00
-                    "mov    v17.16b, %17.16b            \n" // sum01
-                    "mov    v18.16b, %17.16b            \n" // sum02
-                    "mov    v19.16b, %17.16b            \n" // sum03
-
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%1] \n" // r02 r03 r04 r05
-
-                    "fmla   v16.4s, %8.4s, v10.4s       \n"
-                    "fmla   v17.4s, %8.4s, v11.4s       \n"
-                    "fmla   v18.4s, %8.4s, v12.4s       \n"
-                    "fmla   v19.4s, %8.4s, v13.4s       \n"
-
-                    "add    %1, %1, #32                 \n"
-
-                    "fmla   v16.4s, %9.4s, v11.4s       \n"
-                    "fmla   v17.4s, %9.4s, v12.4s       \n"
-                    "fmla   v18.4s, %9.4s, v13.4s       \n"
-                    "fmla   v19.4s, %9.4s, v14.4s       \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v10.4s, v11.4s}, [%2], #32 \n" // r10 r11
-
-                    "fmla   v16.4s, %10.4s, v12.4s      \n"
-                    "fmla   v17.4s, %10.4s, v13.4s      \n"
-                    "fmla   v18.4s, %10.4s, v14.4s      \n"
-                    "fmla   v19.4s, %10.4s, v15.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%2] \n" // r12 r13 r14 r15
-
-                    "fmla   v16.4s, %11.4s, v10.4s      \n"
-                    "fmla   v17.4s, %11.4s, v11.4s      \n"
-                    "fmla   v18.4s, %11.4s, v12.4s      \n"
-                    "fmla   v19.4s, %11.4s, v13.4s      \n"
-
-                    "add    %2, %2, #32                 \n"
-
-                    "fmla   v16.4s, %12.4s, v11.4s      \n"
-                    "fmla   v17.4s, %12.4s, v12.4s      \n"
-                    "fmla   v18.4s, %12.4s, v13.4s      \n"
-                    "fmla   v19.4s, %12.4s, v14.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v10.4s, v11.4s}, [%3], #32 \n" // r20 r21
-
-                    "fmla   v16.4s, %13.4s, v12.4s      \n"
-                    "fmla   v17.4s, %13.4s, v13.4s      \n"
-                    "fmla   v18.4s, %13.4s, v14.4s      \n"
-                    "fmla   v19.4s, %13.4s, v15.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%3] \n" // r22 r23 r24 r25
-
-                    "fmla   v16.4s, %14.4s, v10.4s      \n"
-                    "fmla   v17.4s, %14.4s, v11.4s      \n"
-                    "fmla   v18.4s, %14.4s, v12.4s      \n"
-                    "fmla   v19.4s, %14.4s, v13.4s      \n"
-
-                    "add    %3, %3, #32                 \n"
-
-                    "fmla   v16.4s, %15.4s, v11.4s      \n"
-                    "fmla   v17.4s, %15.4s, v12.4s      \n"
-                    "fmla   v18.4s, %15.4s, v13.4s      \n"
-                    "fmla   v19.4s, %15.4s, v14.4s      \n"
-
-                    "fmla   v16.4s, %16.4s, v12.4s      \n"
-                    "fmla   v17.4s, %16.4s, v13.4s      \n"
-                    "fmla   v18.4s, %16.4s, v14.4s      \n"
-                    "fmla   v19.4s, %16.4s, v15.4s      \n"
-
-                    "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19");
-#else
-                asm volatile(
-                    "pld        [%1, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%1 :128]! \n" // r00 r01
-
-                    "vmov       q10, %q17       \n" // sum00
-                    "vmov       q11, %q17       \n" // sum01
-
-                    "vmla.f32   q10, %q8, q14   \n"
-                    "vmla.f32   q11, %q8, q15   \n"
-                    "vmla.f32   q10, %q9, q15   \n"
-
-                    "pld        [%1, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%1 :128]! \n" // r02 r03
-
-                    "vmov       q12, %q17       \n" // sum02
-                    "vmov       q13, %q17       \n" // sum03
-
-                    "vmla.f32   q12, %q8, q14   \n"
-                    "vmla.f32   q11, %q9, q14   \n"
-                    "vmla.f32   q13, %q8, q15   \n"
-                    "vmla.f32   q10, %q10, q14  \n"
-                    "vmla.f32   q12, %q9, q15   \n"
-                    "vmla.f32   q11, %q10, q15  \n"
-
-                    //                     "pld        [%1, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%1 :128] \n" // r04 r05
-
-                    "vmla.f32   q13, %q9, q14   \n"
-                    "vmla.f32   q12, %q10, q14  \n"
-                    "vmla.f32   q13, %q10, q15  \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%2 :128]! \n" // r10 r11
-
-                    "vmla.f32   q10, %q11, q14  \n"
-                    "vmla.f32   q11, %q11, q15  \n"
-                    "vmla.f32   q10, %q12, q15  \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%2 :128]! \n" // r12 r13
-
-                    "vmla.f32   q12, %q11, q14  \n"
-                    "vmla.f32   q11, %q12, q14  \n"
-                    "vmla.f32   q13, %q11, q15  \n"
-                    "vmla.f32   q10, %q13, q14  \n"
-                    "vmla.f32   q12, %q12, q15  \n"
-                    "vmla.f32   q11, %q13, q15  \n"
-
-                    //                     "pld        [%2, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%2 :128] \n" // r14 r15
-
-                    "vmla.f32   q13, %q12, q14  \n"
-                    "vmla.f32   q12, %q13, q14  \n"
-                    "vmla.f32   q13, %q13, q15  \n"
-
-                    "pld        [%3, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%3 :128]! \n" // r20 r21
-
-                    "vmla.f32   q10, %q14, q14  \n"
-                    "vmla.f32   q11, %q14, q15  \n"
-                    "vmla.f32   q10, %q15, q15  \n"
-
-                    "pld        [%3, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%3 :128]! \n" // r22 r23
-
-                    "vmla.f32   q12, %q14, q14  \n"
-                    "vmla.f32   q11, %q15, q14  \n"
-                    "vmla.f32   q13, %q14, q15  \n"
-                    "vmla.f32   q10, %q16, q14  \n"
-                    "vmla.f32   q12, %q15, q15  \n"
-                    "vmla.f32   q11, %q16, q15  \n"
-
-                    //                     "pld        [%3, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%3 :128] \n" // r24 r25
-
-                    "vmla.f32   q13, %q15, q14  \n"
-                    "vmla.f32   q12, %q16, q14  \n"
-                    "vmla.f32   q13, %q16, q15  \n"
-
-                    "vstm       %0!, {d20-d27}  \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%1] \n" // r00 r01 r02 r03
-
-                    "mov    v16.16b, %17.16b            \n" // sum00
-                    "mov    v17.16b, %17.16b            \n" // sum01
-
-                    "eor    v18.16b, v18.16b, v18.16b   \n"
-                    "eor    v19.16b, v19.16b, v19.16b   \n"
-
-                    "fmla   v16.4s, %8.4s, v12.4s       \n"
-                    "fmla   v17.4s, %8.4s, v13.4s       \n"
-
-                    "add    %1, %1, #32                 \n"
-
-                    "fmla   v18.4s, %9.4s, v13.4s       \n"
-                    "fmla   v19.4s, %9.4s, v14.4s       \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%2] \n" // r10 r11 r12 r13
-
-                    "fmla   v16.4s, %10.4s, v14.4s      \n"
-                    "fmla   v17.4s, %10.4s, v15.4s      \n"
-
-                    "add    %2, %2, #32                 \n"
-
-                    "fmla   v18.4s, %11.4s, v20.4s      \n"
-                    "fmla   v19.4s, %11.4s, v21.4s      \n"
-
-                    "fmla   v16.4s, %12.4s, v21.4s      \n"
-                    "fmla   v17.4s, %12.4s, v22.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s, v15.4s}, [%3] \n" // r20 r21 r22 r23
-
-                    "fmla   v18.4s, %13.4s, v22.4s      \n"
-                    "fmla   v19.4s, %13.4s, v23.4s      \n"
-
-                    "fmla   v16.4s, %14.4s, v12.4s      \n"
-                    "fmla   v17.4s, %14.4s, v13.4s      \n"
-
-                    "fmla   v18.4s, %15.4s, v13.4s      \n"
-                    "fmla   v19.4s, %15.4s, v14.4s      \n"
-
-                    "fmla   v16.4s, %16.4s, v14.4s      \n"
-                    "fmla   v17.4s, %16.4s, v15.4s      \n"
-
-                    "add    %3, %3, #32                 \n"
-
-                    "fadd   v16.4s, v16.4s, v18.4s      \n"
-                    "fadd   v17.4s, v17.4s, v19.4s      \n"
-
-                    "st1    {v16.4s, v17.4s}, [%0], #32 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-#else
-                asm volatile(
-                    "pld        [%1, #256]      \n"
-                    "vld1.f32   {d24-d27}, [%1 :128]! \n" // r00 r01
-
-                    "vmov       q10, %q17       \n" // sum00
-                    "vmov       q11, %q17       \n" // sum01
-
-                    "vmla.f32   q10, %q8, q12   \n"
-                    "vmla.f32   q11, %q8, q13   \n"
-
-                    "pld        [%1, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%1 :128] \n" // r02 r03
-
-                    "vmla.f32   q10, %q9, q13   \n"
-
-                    "vmla.f32   q11, %q9, q14   \n"
-                    "vmla.f32   q10, %q10, q14  \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.f32   {d24-d27}, [%2 :128]! \n" // r10 r11
-
-                    "vmla.f32   q11, %q10, q15  \n"
-
-                    "vmla.f32   q10, %q11, q12  \n"
-                    "vmla.f32   q11, %q11, q13  \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%2 :128] \n" // r12 r13
-
-                    "vmla.f32   q10, %q12, q13  \n"
-
-                    "vmla.f32   q11, %q12, q14  \n"
-                    "vmla.f32   q10, %q13, q14  \n"
-
-                    "pld        [%3, #256]      \n"
-                    "vld1.f32   {d24-d27}, [%3 :128]! \n" // r20 r21
-
-                    "vmla.f32   q11, %q13, q15  \n"
-
-                    "vmla.f32   q10, %q14, q12  \n"
-                    "vmla.f32   q11, %q14, q13  \n"
-
-                    "pld        [%3, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%3 :128] \n" // r22 r23
-
-                    "vmla.f32   q10, %q15, q13  \n"
-
-                    "vmla.f32   q11, %q15, q14  \n"
-                    "vmla.f32   q10, %q16, q14  \n"
-                    "vmla.f32   q11, %q16, q15  \n"
-
-                    "vst1.f32   {d20-d23}, [%0 :128]! \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif
-            }
-            for (; j < outw; j++)
-            {
-                float32x4_t _sum0 = _bias0;
-
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r02 = vld1q_f32(r0 + 8);
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r12 = vld1q_f32(r1 + 8);
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r22 = vld1q_f32(r2 + 8);
-
-                _sum0 = vmlaq_f32(_sum0, _k00, _r00);
-                _sum0 = vmlaq_f32(_sum0, _k01, _r01);
-                _sum0 = vmlaq_f32(_sum0, _k02, _r02);
-                _sum0 = vmlaq_f32(_sum0, _k10, _r10);
-                _sum0 = vmlaq_f32(_sum0, _k11, _r11);
-                _sum0 = vmlaq_f32(_sum0, _k12, _r12);
-                _sum0 = vmlaq_f32(_sum0, _k20, _r20);
-                _sum0 = vmlaq_f32(_sum0, _k21, _r21);
-                _sum0 = vmlaq_f32(_sum0, _k22, _r22);
-
-                vst1q_f32(outptr0, _sum0);
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                outptr0 += 4;
-            }
-
-            r0 += 2 * 4;
-            r1 += 2 * 4;
-            r2 += 2 * 4;
-        }
-    }
-}
-
-static void convdw3x3s2_pack4_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + g * 4) : vdupq_n_f32(0.f);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out;
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-
-        float32x4_t _k00 = vld1q_f32(k0);
-        float32x4_t _k01 = vld1q_f32(k0 + 4);
-        float32x4_t _k02 = vld1q_f32(k0 + 8);
-        float32x4_t _k10 = vld1q_f32(k0 + 12);
-        float32x4_t _k11 = vld1q_f32(k0 + 16);
-        float32x4_t _k12 = vld1q_f32(k0 + 20);
-        float32x4_t _k20 = vld1q_f32(k0 + 24);
-        float32x4_t _k21 = vld1q_f32(k0 + 28);
-        float32x4_t _k22 = vld1q_f32(k0 + 32);
-
-        int i = 0;
-
-        for (; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j + 3 < outw; j += 4)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%1], #64 \n" // r00 r01 r02 r03
-
-                    "mov    v28.16b, %17.16b            \n" // sum00
-                    "mov    v29.16b, %17.16b            \n" // sum01
-                    "mov    v30.16b, %17.16b            \n" // sum02
-                    "mov    v31.16b, %17.16b            \n" // sum03
-
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v14.4s, v15.4s, v16.4s, v17.4s}, [%1], #64 \n" // r04 r05 r06 r07
-
-                    "fmla   v28.4s, %8.4s, v10.4s       \n"
-                    "fmla   v29.4s, %8.4s, v12.4s       \n"
-                    "fmla   v30.4s, %8.4s, v14.4s       \n"
-                    "fmla   v31.4s, %8.4s, v16.4s       \n"
-
-                    "prfm   pldl1keep, [%1, #128]       \n"
-                    "ld1    {v18.4s}, [%1]              \n" // r08
-
-                    "fmla   v28.4s, %9.4s, v11.4s       \n"
-                    "fmla   v29.4s, %9.4s, v13.4s       \n"
-                    "fmla   v30.4s, %9.4s, v15.4s       \n"
-                    "fmla   v31.4s, %9.4s, v17.4s       \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%2], #64 \n" // r10 r11 r12 r13
-
-                    "fmla   v28.4s, %10.4s, v12.4s      \n"
-                    "fmla   v29.4s, %10.4s, v14.4s      \n"
-                    "fmla   v30.4s, %10.4s, v16.4s      \n"
-                    "fmla   v31.4s, %10.4s, v18.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v24.4s, v25.4s, v26.4s, v27.4s}, [%2], #64 \n" // r14 r15 r16 r17
-
-                    "fmla   v28.4s, %11.4s, v20.4s      \n"
-                    "fmla   v29.4s, %11.4s, v22.4s      \n"
-                    "fmla   v30.4s, %11.4s, v24.4s      \n"
-                    "fmla   v31.4s, %11.4s, v26.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "ld1    {v19.4s}, [%2]              \n" // r18
-
-                    "fmla   v28.4s, %12.4s, v21.4s      \n"
-                    "fmla   v29.4s, %12.4s, v23.4s      \n"
-                    "fmla   v30.4s, %12.4s, v25.4s      \n"
-                    "fmla   v31.4s, %12.4s, v27.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%3], #64 \n" // r20 r21 r22 r23
-
-                    "fmla   v28.4s, %13.4s, v22.4s      \n"
-                    "fmla   v29.4s, %13.4s, v24.4s      \n"
-                    "fmla   v30.4s, %13.4s, v26.4s      \n"
-                    "fmla   v31.4s, %13.4s, v19.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v14.4s, v15.4s, v16.4s, v17.4s}, [%3], #64 \n" // r24 r25 r26 r27
-
-                    "fmla   v28.4s, %14.4s, v10.4s      \n"
-                    "fmla   v29.4s, %14.4s, v12.4s      \n"
-                    "fmla   v30.4s, %14.4s, v14.4s      \n"
-                    "fmla   v31.4s, %14.4s, v16.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #128]       \n"
-                    "ld1    {v18.4s}, [%3]              \n" // r28
-
-                    "fmla   v28.4s, %15.4s, v11.4s      \n"
-                    "fmla   v29.4s, %15.4s, v13.4s      \n"
-                    "fmla   v30.4s, %15.4s, v15.4s      \n"
-                    "fmla   v31.4s, %15.4s, v17.4s      \n"
-
-                    "fmla   v28.4s, %16.4s, v12.4s      \n"
-                    "fmla   v29.4s, %16.4s, v14.4s      \n"
-                    "fmla   v30.4s, %16.4s, v16.4s      \n"
-                    "fmla   v31.4s, %16.4s, v18.4s      \n"
-
-                    "st1    {v28.4s, v29.4s, v30.4s, v31.4s}, [%0], #64 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else
-                asm volatile(
-                    "pld        [%1, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%1 :128]! \n" // r00 r01
-
-                    "vmov       q10, %q17       \n" // sum00
-
-                    "vmla.f32   q10, %q8, q14   \n"
-
-                    "vmov       q11, %q17       \n" // sum01
-
-                    "vmla.f32   q10, %q9, q15   \n"
-
-                    "pld        [%1, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%1 :128]! \n" // r02 r03
-
-                    "vmla.f32   q11, %q8, q14   \n"
-                    "vmla.f32   q10, %q10, q14  \n"
-
-                    "vmov       q12, %q17       \n" // sum02
-
-                    "vmla.f32   q11, %q9, q15   \n"
-
-                    "pld        [%1, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%1 :128]! \n" // r04 r05
-
-                    "vmla.f32   q12, %q8, q14   \n"
-                    "vmla.f32   q11, %q10, q14  \n"
-
-                    "vmla.f32   q12, %q9, q15   \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%2 :128]! \n" // r10 r11
-
-                    "vmla.f32   q10, %q11, q14  \n"
-
-                    "vmov       q13, %q17       \n" // sum03
-
-                    "vmla.f32   q10, %q12, q15  \n"
-
-                    "pld        [%1, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%1 :128]! \n" // r06 r07
-
-                    "vmla.f32   q13, %q8, q14   \n"
-                    "vmla.f32   q12, %q10, q14  \n"
-
-                    "vmla.f32   q13, %q9, q15   \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%2 :128]! \n" // r12 r13
-
-                    "vmla.f32   q11, %q11, q14  \n"
-                    "vmla.f32   q10, %q13, q14  \n"
-
-                    "vmla.f32   q11, %q12, q15  \n"
-
-                    "vld1.f32   {d28-d29}, [%1 :128] \n" // r08
-
-                    "vmla.f32   q13, %q10, q14  \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%2 :128]! \n" // r14 r15
-
-                    "vmla.f32   q12, %q11, q14  \n"
-                    "vmla.f32   q11, %q13, q14  \n"
-
-                    "vmla.f32   q12, %q12, q15  \n"
-
-                    "pld        [%3, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%3 :128]! \n" // r20 r21
-
-                    "vmla.f32   q10, %q14, q14  \n"
-                    "vmla.f32   q10, %q15, q15  \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%2 :128]! \n" // r16 r17
-
-                    "vmla.f32   q13, %q11, q14  \n"
-                    "vmla.f32   q12, %q13, q14  \n"
-
-                    "vmla.f32   q13, %q12, q15  \n"
-
-                    "pld        [%3, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%3 :128]! \n" // r22 r23
-
-                    "vmla.f32   q11, %q14, q14  \n"
-                    "vmla.f32   q10, %q16, q14  \n"
-
-                    "vmla.f32   q11, %q15, q15  \n"
-
-                    "vld1.f32   {d28-d29}, [%2 :128] \n" // r18
-
-                    "vmla.f32   q13, %q13, q14  \n"
-
-                    "pld        [%3, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%3 :128]! \n" // r24 r25
-
-                    "vmla.f32   q12, %q14, q14  \n"
-                    "vmla.f32   q11, %q16, q14  \n"
-
-                    "vmla.f32   q12, %q15, q15  \n"
-
-                    "pld        [%3, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%3 :128]! \n" // r26 r27
-
-                    "vmla.f32   q13, %q14, q14  \n"
-                    "vmla.f32   q12, %q16, q14  \n"
-
-                    "vmla.f32   q13, %q15, q15  \n"
-
-                    "vld1.f32   {d28-d29}, [%3 :128] \n" // r28
-
-                    "vmla.f32   q13, %q16, q14  \n"
-
-                    "vstm       %0!, {d20-d27}  \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%1], #64 \n" // r00 r01 r02 r03
-
-                    "mov    v20.16b, %17.16b            \n" // sum00
-                    "mov    v21.16b, %17.16b            \n" // sum01
-
-                    "eor    v22.16b, v22.16b, v22.16b   \n"
-                    "eor    v23.16b, v23.16b, v23.16b   \n"
-
-                    "fmla   v20.4s, %8.4s, v10.4s       \n"
-                    "fmla   v21.4s, %8.4s, v12.4s       \n"
-
-                    "prfm   pldl1keep, [%1, #128]       \n"
-                    "ld1    {v14.4s}, [%1]              \n" // r04
-
-                    "fmla   v22.4s, %9.4s, v11.4s       \n"
-                    "fmla   v23.4s, %9.4s, v13.4s       \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%2], #64 \n" // r10 r11 r12 r13
-
-                    "fmla   v20.4s, %10.4s, v12.4s      \n"
-                    "fmla   v21.4s, %10.4s, v14.4s      \n"
-
-                    "fmla   v22.4s, %11.4s, v16.4s      \n"
-                    "fmla   v23.4s, %11.4s, v18.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "ld1    {v15.4s}, [%2]              \n" // r14
-
-                    "fmla   v20.4s, %12.4s, v17.4s      \n"
-                    "fmla   v21.4s, %12.4s, v19.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v10.4s, v11.4s, v12.4s, v13.4s}, [%3], #64 \n" // r20 r21 r22 r23
-
-                    "fmla   v22.4s, %13.4s, v18.4s      \n"
-                    "fmla   v23.4s, %13.4s, v15.4s      \n"
-
-                    "fmla   v20.4s, %14.4s, v10.4s      \n"
-                    "fmla   v21.4s, %14.4s, v12.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #128]       \n"
-                    "ld1    {v14.4s}, [%3]              \n" // r24
-
-                    "fmla   v22.4s, %15.4s, v11.4s      \n"
-                    "fmla   v23.4s, %15.4s, v13.4s      \n"
-
-                    "fmla   v20.4s, %16.4s, v12.4s      \n"
-                    "fmla   v21.4s, %16.4s, v14.4s      \n"
-
-                    "fadd   v20.4s, v20.4s, v22.4s      \n"
-                    "fadd   v21.4s, v21.4s, v23.4s      \n"
-
-                    "st1    {v20.4s, v21.4s}, [%0], #32 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-#else
-                asm volatile(
-                    "pld        [%1, #256]      \n"
-                    "vld1.f32   {d24-d27}, [%1 :128]! \n" // r00 r01
-
-                    "vmov       q10, %q17       \n" // sum00
-                    "vmov       q11, %q17       \n" // sum01
-
-                    "vmla.f32   q10, %q8, q12   \n"
-
-                    "pld        [%1, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%1 :128]! \n" // r02 r03
-
-                    "vmla.f32   q10, %q9, q13   \n"
-
-                    "vmla.f32   q11, %q8, q14   \n"
-                    "vmla.f32   q10, %q10, q14  \n"
-
-                    "vld1.f32   {d24-d25}, [%1 :128] \n" // r04
-
-                    "vmla.f32   q11, %q9, q15   \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%2 :128]! \n" // r10 r11
-
-                    "vmla.f32   q11, %q10, q12  \n"
-
-                    "vmla.f32   q10, %q11, q14  \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.f32   {d24-d27}, [%2 :128]! \n" // r12 r13
-
-                    "vmla.f32   q10, %q12, q15  \n"
-
-                    "vmla.f32   q11, %q11, q12  \n"
-                    "vmla.f32   q10, %q13, q12  \n"
-
-                    "vld1.f32   {d28-d29}, [%2 :128] \n" // r14
-
-                    "vmla.f32   q11, %q12, q13  \n"
-
-                    "pld        [%3, #256]      \n"
-                    "vld1.f32   {d24-d27}, [%3 :128]! \n" // r20 r21
-
-                    "vmla.f32   q11, %q13, q14  \n"
-
-                    "vmla.f32   q10, %q14, q12  \n"
-
-                    "pld        [%3, #256]      \n"
-                    "vld1.f32   {d28-d31}, [%3 :128]! \n" // r22 r23
-
-                    "vmla.f32   q10, %q15, q13  \n"
-
-                    "vmla.f32   q11, %q14, q14  \n"
-                    "vmla.f32   q10, %q16, q14  \n"
-
-                    "vld1.f32   {d24-d25}, [%3 :128] \n" // r24
-
-                    "vmla.f32   q11, %q15, q15  \n"
-
-                    "vmla.f32   q11, %q16, q12  \n"
-
-                    "vst1.f32   {d20-d23}, [%0 :128]! \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif
-            }
-            for (; j < outw; j++)
-            {
-                float32x4_t _sum0 = _bias0;
-
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r02 = vld1q_f32(r0 + 8);
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r12 = vld1q_f32(r1 + 8);
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r22 = vld1q_f32(r2 + 8);
-
-                _sum0 = vmlaq_f32(_sum0, _k00, _r00);
-                _sum0 = vmlaq_f32(_sum0, _k01, _r01);
-                _sum0 = vmlaq_f32(_sum0, _k02, _r02);
-                _sum0 = vmlaq_f32(_sum0, _k10, _r10);
-                _sum0 = vmlaq_f32(_sum0, _k11, _r11);
-                _sum0 = vmlaq_f32(_sum0, _k12, _r12);
-                _sum0 = vmlaq_f32(_sum0, _k20, _r20);
-                _sum0 = vmlaq_f32(_sum0, _k21, _r21);
-                _sum0 = vmlaq_f32(_sum0, _k22, _r22);
-
-                vst1q_f32(outptr0, _sum0);
-
-                r0 += 2 * 4;
-                r1 += 2 * 4;
-                r2 += 2 * 4;
-                outptr0 += 4;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack4_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack4_bf16s.h
deleted file mode 100644
index 8c778132..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack4_bf16s.h
+++ /dev/null
@@ -1,1501 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_pack4_bf16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-#if __aarch64__
-    const int w = bottom_blob.w;
-#endif
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + g * 4) : vdupq_n_f32(0.f);
-
-        const unsigned short* k0 = kernel.row<const unsigned short>(g);
-
-        unsigned short* outptr0 = out.row<unsigned short>(0);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const unsigned short* r0 = img0.row<const unsigned short>(0);
-        const unsigned short* r1 = img0.row<const unsigned short>(1);
-        const unsigned short* r2 = img0.row<const unsigned short>(2);
-
-        float32x4_t _k00 = bfloat2float(vld1_u16(k0));
-        float32x4_t _k01 = bfloat2float(vld1_u16(k0 + 4));
-        float32x4_t _k02 = bfloat2float(vld1_u16(k0 + 8));
-        float32x4_t _k10 = bfloat2float(vld1_u16(k0 + 12));
-        float32x4_t _k11 = bfloat2float(vld1_u16(k0 + 16));
-        float32x4_t _k12 = bfloat2float(vld1_u16(k0 + 20));
-        float32x4_t _k20 = bfloat2float(vld1_u16(k0 + 24));
-        float32x4_t _k21 = bfloat2float(vld1_u16(k0 + 28));
-        float32x4_t _k22 = bfloat2float(vld1_u16(k0 + 32));
-
-        int i = 0;
-
-#if __aarch64__
-        unsigned short* outptr1 = out.row<unsigned short>(1);
-        const unsigned short* r3 = img0.row<const unsigned short>(3);
-
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-
-            for (; j + 3 < outw; j += 4)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v10.4h, v11.4h, v12.4h, v13.4h}, [%3], #32 \n" // r10 r11 r12 r13
-
-                    "mov    v16.16b, %21.16b            \n" // sum00
-                    "mov    v17.16b, %21.16b            \n" // sum01
-
-                    "prfm   pldl1keep, [%3, #128]       \n"
-                    "ld1    {v28.4h, v29.4h}, [%3]      \n" // r14 r15
-
-                    "shll   v10.4s, v10.4h, #16         \n"
-                    "shll   v11.4s, v11.4h, #16         \n"
-
-                    "mov    v18.16b, %21.16b            \n" // sum02
-                    "mov    v19.16b, %21.16b            \n" // sum03
-
-                    "shll   v12.4s, v12.4h, #16         \n"
-                    "shll   v13.4s, v13.4h, #16         \n"
-
-                    "mov    v20.16b, %21.16b            \n" // sum10
-
-                    "fmla   v16.4s, %15.4s, v10.4s      \n"
-                    "fmla   v17.4s, %15.4s, v11.4s      \n"
-
-                    "mov    v21.16b, %21.16b            \n" // sum11
-
-                    "fmla   v18.4s, %15.4s, v12.4s      \n"
-                    "fmla   v19.4s, %15.4s, v13.4s      \n"
-
-                    "mov    v22.16b, %21.16b            \n" // sum12
-
-                    "fmla   v20.4s, %12.4s, v10.4s      \n"
-                    "fmla   v21.4s, %12.4s, v11.4s      \n"
-
-                    "mov    v23.16b, %21.16b            \n" // sum13
-
-                    "fmla   v22.4s, %12.4s, v12.4s      \n"
-                    "fmla   v23.4s, %12.4s, v13.4s      \n"
-
-                    "shll   v28.4s, v28.4h, #16         \n"
-
-                    "fmla   v16.4s, %16.4s, v11.4s      \n"
-                    "fmla   v17.4s, %16.4s, v12.4s      \n"
-
-                    "shll   v29.4s, v29.4h, #16         \n"
-
-                    "fmla   v18.4s, %16.4s, v13.4s      \n"
-                    "fmla   v19.4s, %16.4s, v28.4s      \n"
-
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%4], #32 \n" // r20 r21 r22 r23
-
-                    "fmla   v20.4s, %13.4s, v11.4s      \n"
-                    "fmla   v21.4s, %13.4s, v12.4s      \n"
-                    "fmla   v22.4s, %13.4s, v13.4s      \n"
-                    "fmla   v23.4s, %13.4s, v28.4s      \n"
-
-                    "prfm   pldl1keep, [%4, #128]       \n"
-                    "ld1    {v14.4h, v15.4h}, [%4]      \n" // r24 r25
-
-                    "fmla   v16.4s, %17.4s, v12.4s      \n"
-                    "fmla   v17.4s, %17.4s, v13.4s      \n"
-
-                    "shll   v24.4s, v24.4h, #16         \n"
-
-                    "fmla   v18.4s, %17.4s, v28.4s      \n"
-                    "fmla   v19.4s, %17.4s, v29.4s      \n"
-
-                    "shll   v25.4s, v25.4h, #16         \n"
-
-                    "fmla   v20.4s, %14.4s, v12.4s      \n"
-                    "fmla   v21.4s, %14.4s, v13.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v10.4h, v11.4h, v12.4h, v13.4h}, [%2], #32 \n" // r00 r01 r02 r03
-
-                    "fmla   v22.4s, %14.4s, v28.4s      \n"
-                    "fmla   v23.4s, %14.4s, v29.4s      \n"
-
-                    "shll   v26.4s, v26.4h, #16         \n"
-
-                    "fmla   v16.4s, %18.4s, v24.4s      \n"
-                    "fmla   v17.4s, %18.4s, v25.4s      \n"
-
-                    "shll   v27.4s, v27.4h, #16         \n"
-
-                    "fmla   v18.4s, %18.4s, v26.4s      \n"
-                    "fmla   v19.4s, %18.4s, v27.4s      \n"
-
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v28.4h, v29.4h, v30.4h, v31.4h}, [%5], #32 \n" // r30 r31 r32 r33
-
-                    "fmla   v20.4s, %15.4s, v24.4s      \n"
-                    "fmla   v21.4s, %15.4s, v25.4s      \n"
-
-                    "shll   v14.4s, v14.4h, #16         \n"
-
-                    "fmla   v22.4s, %15.4s, v26.4s      \n"
-                    "fmla   v23.4s, %15.4s, v27.4s      \n"
-
-                    "shll   v15.4s, v15.4h, #16         \n"
-
-                    "fmla   v16.4s, %19.4s, v25.4s      \n"
-                    "fmla   v17.4s, %19.4s, v26.4s      \n"
-
-                    "fmla   v18.4s, %19.4s, v27.4s      \n"
-                    "fmla   v19.4s, %19.4s, v14.4s      \n"
-
-                    "fmla   v20.4s, %16.4s, v25.4s      \n"
-                    "fmla   v21.4s, %16.4s, v26.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "ld1    {v24.4h, v25.4h}, [%2]      \n" // r04 r05
-
-                    "fmla   v22.4s, %16.4s, v27.4s      \n"
-                    "fmla   v23.4s, %16.4s, v14.4s      \n"
-
-                    "shll   v10.4s, v10.4h, #16         \n"
-                    "shll   v11.4s, v11.4h, #16         \n"
-
-                    "fmla   v16.4s, %20.4s, v26.4s      \n"
-                    "fmla   v17.4s, %20.4s, v27.4s      \n"
-
-                    "shll   v12.4s, v12.4h, #16         \n"
-
-                    "fmla   v18.4s, %20.4s, v14.4s      \n"
-                    "fmla   v19.4s, %20.4s, v15.4s      \n"
-
-                    "shll   v13.4s, v13.4h, #16         \n"
-
-                    "fmla   v20.4s, %17.4s, v26.4s      \n"
-                    "fmla   v21.4s, %17.4s, v27.4s      \n"
-
-                    "prfm   pldl1keep, [%5, #128]       \n"
-                    "ld1    {v26.4h, v27.4h}, [%5]      \n" // r34 r35
-
-                    "fmla   v22.4s, %17.4s, v14.4s      \n"
-                    "fmla   v23.4s, %17.4s, v15.4s      \n"
-
-                    "shll   v28.4s, v28.4h, #16         \n"
-
-                    "fmla   v16.4s, %12.4s, v10.4s      \n"
-                    "fmla   v17.4s, %12.4s, v11.4s      \n"
-
-                    "shll   v29.4s, v29.4h, #16         \n"
-
-                    "fmla   v18.4s, %12.4s, v12.4s      \n"
-                    "fmla   v19.4s, %12.4s, v13.4s      \n"
-
-                    "shll   v30.4s, v30.4h, #16         \n"
-
-                    "fmla   v20.4s, %18.4s, v28.4s      \n"
-                    "fmla   v21.4s, %18.4s, v29.4s      \n"
-
-                    "shll   v31.4s, v31.4h, #16         \n"
-
-                    "fmla   v22.4s, %18.4s, v30.4s      \n"
-                    "fmla   v23.4s, %18.4s, v31.4s      \n"
-
-                    "shll   v24.4s, v24.4h, #16         \n"
-
-                    "fmla   v16.4s, %13.4s, v11.4s      \n"
-                    "fmla   v17.4s, %13.4s, v12.4s      \n"
-                    "fmla   v18.4s, %13.4s, v13.4s      \n"
-                    "fmla   v19.4s, %13.4s, v24.4s      \n"
-
-                    "shll   v26.4s, v26.4h, #16         \n"
-
-                    "fmla   v20.4s, %19.4s, v29.4s      \n"
-                    "fmla   v21.4s, %19.4s, v30.4s      \n"
-                    "fmla   v22.4s, %19.4s, v31.4s      \n"
-                    "fmla   v23.4s, %19.4s, v26.4s      \n"
-
-                    "shll   v25.4s, v25.4h, #16         \n"
-
-                    "fmla   v16.4s, %14.4s, v12.4s      \n"
-                    "fmla   v17.4s, %14.4s, v13.4s      \n"
-                    "fmla   v18.4s, %14.4s, v24.4s      \n"
-                    "fmla   v19.4s, %14.4s, v25.4s      \n"
-
-                    "shll   v27.4s, v27.4h, #16         \n"
-
-                    "fmla   v20.4s, %20.4s, v30.4s      \n"
-                    "fmla   v21.4s, %20.4s, v31.4s      \n"
-                    "fmla   v22.4s, %20.4s, v26.4s      \n"
-                    "fmla   v23.4s, %20.4s, v27.4s      \n"
-
-                    "shrn   v16.4h, v16.4s, #16         \n"
-                    "shrn   v17.4h, v17.4s, #16         \n"
-                    "shrn   v18.4h, v18.4s, #16         \n"
-                    "shrn   v19.4h, v19.4s, #16         \n"
-                    "shrn   v20.4h, v20.4s, #16         \n"
-                    "shrn   v21.4h, v21.4s, #16         \n"
-                    "shrn   v22.4h, v22.4s, #16         \n"
-                    "shrn   v23.4h, v23.4s, #16         \n"
-
-                    "st1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%0], #32 \n"
-                    "st1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%1], #32 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3)       // %5
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "w"(_k00),  // %12
-                    "w"(_k01),  // %13
-                    "w"(_k02),  // %14
-                    "w"(_k10),  // %15
-                    "w"(_k11),  // %16
-                    "w"(_k12),  // %17
-                    "w"(_k20),  // %18
-                    "w"(_k21),  // %19
-                    "w"(_k22),  // %20
-                    "w"(_bias0) // %21
-                    : "memory", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v10.4h, v11.4h, v12.4h, v13.4h}, [%3] \n" // r10 r11 r12 r13
-
-                    "mov    v16.16b, %21.16b            \n" // sum00
-                    "mov    v17.16b, %21.16b            \n" // sum01
-
-                    "shll   v10.4s, v10.4h, #16         \n"
-                    "shll   v11.4s, v11.4h, #16         \n"
-
-                    "mov    v18.16b, %21.16b            \n" // sum10
-                    "mov    v19.16b, %21.16b            \n" // sum11
-
-                    "fmla   v16.4s, %15.4s, v10.4s      \n"
-                    "fmla   v17.4s, %15.4s, v11.4s      \n"
-
-                    "shll   v12.4s, v12.4h, #16         \n"
-
-                    "fmla   v18.4s, %12.4s, v10.4s      \n"
-                    "fmla   v19.4s, %12.4s, v11.4s      \n"
-
-                    "shll   v13.4s, v13.4h, #16         \n"
-
-                    "fmla   v16.4s, %16.4s, v11.4s      \n"
-                    "fmla   v17.4s, %16.4s, v12.4s      \n"
-
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%4] \n" // r20 r21 r22 r23
-
-                    "fmla   v18.4s, %13.4s, v11.4s      \n"
-                    "fmla   v19.4s, %13.4s, v12.4s      \n"
-
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "fmla   v16.4s, %17.4s, v12.4s      \n"
-                    "fmla   v17.4s, %17.4s, v13.4s      \n"
-
-                    "shll   v21.4s, v21.4h, #16         \n"
-
-                    "fmla   v18.4s, %14.4s, v12.4s      \n"
-                    "fmla   v19.4s, %14.4s, v13.4s      \n"
-
-                    "shll   v22.4s, v22.4h, #16         \n"
-
-                    "fmla   v16.4s, %18.4s, v20.4s      \n"
-                    "fmla   v17.4s, %18.4s, v21.4s      \n"
-
-                    "shll   v23.4s, v23.4h, #16         \n"
-
-                    "fmla   v18.4s, %15.4s, v20.4s      \n"
-                    "fmla   v19.4s, %15.4s, v21.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v10.4h, v11.4h, v12.4h, v13.4h}, [%2] \n" // r00 r01 r02 r03
-
-                    "fmla   v16.4s, %19.4s, v21.4s      \n"
-                    "fmla   v17.4s, %19.4s, v22.4s      \n"
-
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%5] \n" // r30 r31 r32 r33
-
-                    "fmla   v18.4s, %16.4s, v21.4s      \n"
-                    "fmla   v19.4s, %16.4s, v22.4s      \n"
-
-                    "shll   v10.4s, v10.4h, #16         \n"
-
-                    "fmla   v16.4s, %20.4s, v22.4s      \n"
-                    "fmla   v17.4s, %20.4s, v23.4s      \n"
-
-                    "shll   v24.4s, v24.4h, #16         \n"
-
-                    "fmla   v18.4s, %17.4s, v22.4s      \n"
-                    "fmla   v19.4s, %17.4s, v23.4s      \n"
-
-                    "shll   v11.4s, v11.4h, #16         \n"
-                    "shll   v25.4s, v25.4h, #16         \n"
-
-                    "fmla   v16.4s, %12.4s, v10.4s      \n"
-                    "fmla   v17.4s, %12.4s, v11.4s      \n"
-
-                    "shll   v12.4s, v12.4h, #16         \n"
-
-                    "fmla   v18.4s, %18.4s, v24.4s      \n"
-                    "fmla   v19.4s, %18.4s, v25.4s      \n"
-
-                    "shll   v26.4s, v26.4h, #16         \n"
-
-                    "fmla   v16.4s, %13.4s, v11.4s      \n"
-                    "fmla   v17.4s, %13.4s, v12.4s      \n"
-
-                    "shll   v13.4s, v13.4h, #16         \n"
-
-                    "fmla   v18.4s, %19.4s, v25.4s      \n"
-                    "fmla   v19.4s, %19.4s, v26.4s      \n"
-
-                    "shll   v27.4s, v27.4h, #16         \n"
-
-                    "fmla   v16.4s, %14.4s, v12.4s      \n"
-                    "fmla   v17.4s, %14.4s, v13.4s      \n"
-
-                    "add    %3, %3, #16                 \n"
-
-                    "fmla   v18.4s, %20.4s, v26.4s      \n"
-                    "fmla   v19.4s, %20.4s, v27.4s      \n"
-
-                    "add    %4, %4, #16                 \n"
-
-                    "shrn   v16.4h, v16.4s, #16         \n"
-                    "shrn   v17.4h, v17.4s, #16         \n"
-
-                    "add    %2, %2, #16                 \n"
-
-                    "shrn   v18.4h, v18.4s, #16         \n"
-                    "shrn   v19.4h, v19.4s, #16         \n"
-
-                    "add    %5, %5, #16                 \n"
-
-                    "st1    {v16.4h, v17.4h}, [%0], #16 \n"
-                    "st1    {v18.4h, v19.4h}, [%1], #16 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3)       // %5
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "w"(_k00),  // %12
-                    "w"(_k01),  // %13
-                    "w"(_k02),  // %14
-                    "w"(_k10),  // %15
-                    "w"(_k11),  // %16
-                    "w"(_k12),  // %17
-                    "w"(_k20),  // %18
-                    "w"(_k21),  // %19
-                    "w"(_k22),  // %20
-                    "w"(_bias0) // %21
-                    : "memory", "v10", "v11", "v12", "v13", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-            }
-            for (; j < outw; j++)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%3, #192]       \n"
-                    "ld1    {v10.4h, v11.4h, v12.4h}, [%3] \n" // r10 r11 r12
-
-                    "mov    v18.16b, %21.16b            \n" // sum0
-                    "mov    v19.16b, %21.16b            \n" // sum1
-
-                    "shll   v10.4s, v10.4h, #16         \n"
-                    "shll   v11.4s, v11.4h, #16         \n"
-
-                    "fmul   v16.4s, %15.4s, v10.4s      \n"
-                    "fmul   v17.4s, %12.4s, v10.4s      \n"
-
-                    "shll   v12.4s, v12.4h, #16         \n"
-
-                    "fmla   v18.4s, %16.4s, v11.4s      \n"
-                    "fmla   v19.4s, %13.4s, v11.4s      \n"
-
-                    "prfm   pldl1keep, [%4, #192]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h}, [%4] \n" // r20 r21 r22
-
-                    "fmla   v16.4s, %17.4s, v12.4s      \n"
-                    "fmla   v17.4s, %14.4s, v12.4s      \n"
-
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-
-                    "fmla   v18.4s, %18.4s, v20.4s      \n"
-                    "fmla   v19.4s, %15.4s, v20.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #192]       \n"
-                    "ld1    {v10.4h, v11.4h, v12.4h}, [%2] \n" // r00 r01 r02
-
-                    "shll   v22.4s, v22.4h, #16         \n"
-
-                    "prfm   pldl1keep, [%5, #192]       \n"
-                    "ld1    {v24.4h, v25.4h, v26.4h}, [%5] \n" // r30 r31 r32
-
-                    "fmla   v16.4s, %19.4s, v21.4s      \n"
-                    "fmla   v17.4s, %16.4s, v21.4s      \n"
-
-                    "shll   v10.4s, v10.4h, #16         \n"
-                    "shll   v24.4s, v24.4h, #16         \n"
-
-                    "fmla   v18.4s, %20.4s, v22.4s      \n"
-                    "fmla   v19.4s, %17.4s, v22.4s      \n"
-
-                    "shll   v11.4s, v11.4h, #16         \n"
-                    "shll   v25.4s, v25.4h, #16         \n"
-
-                    "fmla   v16.4s, %12.4s, v10.4s      \n"
-                    "fmla   v17.4s, %18.4s, v24.4s      \n"
-
-                    "shll   v12.4s, v12.4h, #16         \n"
-                    "shll   v26.4s, v26.4h, #16         \n"
-
-                    "fmla   v18.4s, %13.4s, v11.4s      \n"
-                    "fmla   v19.4s, %19.4s, v25.4s      \n"
-
-                    "add    %3, %3, #8                  \n"
-
-                    "fmla   v16.4s, %14.4s, v12.4s      \n"
-                    "fmla   v17.4s, %20.4s, v26.4s      \n"
-
-                    "add    %4, %4, #8                  \n"
-
-                    "fadd   v18.4s, v18.4s, v16.4s      \n"
-                    "fadd   v19.4s, v19.4s, v17.4s      \n"
-
-                    "add    %2, %2, #8                  \n"
-
-                    "shrn   v18.4h, v18.4s, #16         \n"
-                    "shrn   v19.4h, v19.4s, #16         \n"
-
-                    "add    %5, %5, #8                  \n"
-
-                    "st1    {v18.4h}, [%0], #8          \n"
-                    "st1    {v19.4h}, [%1], #8          \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3)       // %5
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "w"(_k00),  // %12
-                    "w"(_k01),  // %13
-                    "w"(_k02),  // %14
-                    "w"(_k10),  // %15
-                    "w"(_k11),  // %16
-                    "w"(_k12),  // %17
-                    "w"(_k20),  // %18
-                    "w"(_k21),  // %19
-                    "w"(_k22),  // %20
-                    "w"(_bias0) // %21
-                    : "memory", "v10", "v11", "v12", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v24", "v25", "v26");
-            }
-
-            r0 += 2 * 4 + w * 4;
-            r1 += 2 * 4 + w * 4;
-            r2 += 2 * 4 + w * 4;
-            r3 += 2 * 4 + w * 4;
-
-            outptr0 += outw * 4;
-            outptr1 += outw * 4;
-        }
-#endif // __aarch64__
-        for (; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j + 3 < outw; j += 4)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v10.4h, v11.4h, v12.4h, v13.4h}, [%1], #32 \n" // r00 r01 r02 r03
-
-                    "mov    v16.16b, %17.16b            \n" // sum00
-                    "mov    v17.16b, %17.16b            \n" // sum01
-                    "mov    v18.16b, %17.16b            \n" // sum02
-                    "mov    v19.16b, %17.16b            \n" // sum03
-
-                    "shll   v10.4s, v10.4h, #16         \n"
-                    "shll   v11.4s, v11.4h, #16         \n"
-
-                    "fmla   v16.4s, %8.4s, v10.4s       \n"
-                    "fmla   v17.4s, %8.4s, v11.4s       \n"
-
-                    "shll   v12.4s, v12.4h, #16         \n"
-                    "shll   v13.4s, v13.4h, #16         \n"
-
-                    "fmla   v18.4s, %8.4s, v12.4s       \n"
-                    "fmla   v19.4s, %8.4s, v13.4s       \n"
-
-                    "prfm   pldl1keep, [%1, #128]       \n"
-                    "ld1    {v14.4h, v15.4h}, [%1]      \n" // r04 r05
-
-                    "fmla   v16.4s, %9.4s, v11.4s       \n"
-                    "fmla   v17.4s, %9.4s, v12.4s       \n"
-
-                    "shll   v14.4s, v14.4h, #16         \n"
-
-                    "fmla   v18.4s, %9.4s, v13.4s       \n"
-                    "fmla   v19.4s, %9.4s, v14.4s       \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%2], #32 \n" // r10 r11 r12 r13
-
-                    "fmla   v16.4s, %10.4s, v12.4s      \n"
-                    "fmla   v17.4s, %10.4s, v13.4s      \n"
-
-                    "shll   v15.4s, v15.4h, #16         \n"
-
-                    "fmla   v18.4s, %10.4s, v14.4s      \n"
-                    "fmla   v19.4s, %10.4s, v15.4s      \n"
-
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-
-                    "fmla   v16.4s, %11.4s, v20.4s      \n"
-                    "fmla   v17.4s, %11.4s, v21.4s      \n"
-
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-
-                    "fmla   v18.4s, %11.4s, v22.4s      \n"
-                    "fmla   v19.4s, %11.4s, v23.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "ld1    {v14.4h, v15.4h}, [%2]      \n" // r14 r15
-
-                    "fmla   v16.4s, %12.4s, v21.4s      \n"
-                    "fmla   v17.4s, %12.4s, v22.4s      \n"
-
-                    "shll   v14.4s, v14.4h, #16         \n"
-
-                    "fmla   v18.4s, %12.4s, v23.4s      \n"
-                    "fmla   v19.4s, %12.4s, v14.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v10.4h, v11.4h, v12.4h, v13.4h}, [%3], #32 \n" // r20 r21 r22 r23
-
-                    "fmla   v16.4s, %13.4s, v22.4s      \n"
-                    "fmla   v17.4s, %13.4s, v23.4s      \n"
-
-                    "shll   v15.4s, v15.4h, #16         \n"
-
-                    "fmla   v18.4s, %13.4s, v14.4s      \n"
-                    "fmla   v19.4s, %13.4s, v15.4s      \n"
-
-                    "shll   v10.4s, v10.4h, #16         \n"
-                    "shll   v11.4s, v11.4h, #16         \n"
-
-                    "fmla   v16.4s, %14.4s, v10.4s      \n"
-                    "fmla   v17.4s, %14.4s, v11.4s      \n"
-
-                    "shll   v12.4s, v12.4h, #16         \n"
-                    "shll   v13.4s, v13.4h, #16         \n"
-
-                    "fmla   v18.4s, %14.4s, v12.4s      \n"
-                    "fmla   v19.4s, %14.4s, v13.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #128]       \n"
-                    "ld1    {v14.4h, v15.4h}, [%3]      \n" // r24 r25
-
-                    "fmla   v16.4s, %15.4s, v11.4s      \n"
-                    "fmla   v17.4s, %15.4s, v12.4s      \n"
-
-                    "shll   v14.4s, v14.4h, #16         \n"
-
-                    "fmla   v18.4s, %15.4s, v13.4s      \n"
-                    "fmla   v19.4s, %15.4s, v14.4s      \n"
-
-                    "fmla   v16.4s, %16.4s, v12.4s      \n"
-                    "fmla   v17.4s, %16.4s, v13.4s      \n"
-
-                    "shll   v15.4s, v15.4h, #16         \n"
-
-                    "fmla   v18.4s, %16.4s, v14.4s      \n"
-                    "fmla   v19.4s, %16.4s, v15.4s      \n"
-
-                    "shrn   v16.4h, v16.4s, #16         \n"
-                    "shrn   v17.4h, v17.4s, #16         \n"
-                    "shrn   v18.4h, v18.4s, #16         \n"
-                    "shrn   v19.4h, v19.4s, #16         \n"
-
-                    "st1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%0], #32 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-#else
-                asm volatile(
-                    "pld        [%1, #128]      \n"
-                    "vld1.u16   {d30-d31}, [%1 :64]! \n" // r00 r01
-
-                    "vmov       q10, %q17       \n" // sum00
-                    "vmov       q11, %q17       \n" // sum01
-
-                    "vshll.u16  q14, d30, #16   \n"
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmla.f32   q10, %q8, q14   \n"
-                    "vmla.f32   q11, %q8, q15   \n"
-                    "vmla.f32   q10, %q9, q15   \n"
-
-                    "pld        [%1, #128]      \n"
-                    "vld1.u16   {d30-d31}, [%1 :64]! \n" // r02 r03
-
-                    "vmov       q12, %q17       \n" // sum02
-                    "vmov       q13, %q17       \n" // sum03
-
-                    "vshll.u16  q14, d30, #16   \n"
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmla.f32   q12, %q8, q14   \n"
-                    "vmla.f32   q11, %q9, q14   \n"
-                    "vmla.f32   q13, %q8, q15   \n"
-                    "vmla.f32   q10, %q10, q14  \n"
-                    "vmla.f32   q12, %q9, q15   \n"
-                    "vmla.f32   q11, %q10, q15  \n"
-
-                    //                     "pld        [%1, #128]      \n"
-                    "vld1.u16   {d30-d31}, [%1 :64] \n" // r04 r05
-
-                    "vshll.u16  q14, d30, #16   \n"
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmla.f32   q13, %q9, q14   \n"
-                    "vmla.f32   q12, %q10, q14  \n"
-                    "vmla.f32   q13, %q10, q15  \n"
-
-                    "pld        [%2, #128]      \n"
-                    "vld1.u16   {d30-d31}, [%2 :64]! \n" // r10 r11
-
-                    "vshll.u16  q14, d30, #16   \n"
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmla.f32   q10, %q11, q14  \n"
-                    "vmla.f32   q11, %q11, q15  \n"
-                    "vmla.f32   q10, %q12, q15  \n"
-
-                    "pld        [%2, #128]      \n"
-                    "vld1.u16   {d30-d31}, [%2 :64]! \n" // r12 r13
-
-                    "vshll.u16  q14, d30, #16   \n"
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmla.f32   q12, %q11, q14  \n"
-                    "vmla.f32   q11, %q12, q14  \n"
-                    "vmla.f32   q13, %q11, q15  \n"
-                    "vmla.f32   q10, %q13, q14  \n"
-                    "vmla.f32   q12, %q12, q15  \n"
-                    "vmla.f32   q11, %q13, q15  \n"
-
-                    //                     "pld        [%2, #128]      \n"
-                    "vld1.u16   {d30-d31}, [%2 :64] \n" // r14 r15
-
-                    "vshll.u16  q14, d30, #16   \n"
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmla.f32   q13, %q12, q14  \n"
-                    "vmla.f32   q12, %q13, q14  \n"
-                    "vmla.f32   q13, %q13, q15  \n"
-
-                    "pld        [%3, #128]      \n"
-                    "vld1.u16   {d30-d31}, [%3 :64]! \n" // r20 r21
-
-                    "vshll.u16  q14, d30, #16   \n"
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmla.f32   q10, %q14, q14  \n"
-                    "vmla.f32   q11, %q14, q15  \n"
-                    "vmla.f32   q10, %q15, q15  \n"
-
-                    "pld        [%3, #128]      \n"
-                    "vld1.u16   {d30-d31}, [%3 :64]! \n" // r22 r23
-
-                    "vshll.u16  q14, d30, #16   \n"
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmla.f32   q12, %q14, q14  \n"
-                    "vmla.f32   q11, %q15, q14  \n"
-                    "vmla.f32   q13, %q14, q15  \n"
-                    "vmla.f32   q10, %q16, q14  \n"
-                    "vmla.f32   q12, %q15, q15  \n"
-                    "vmla.f32   q11, %q16, q15  \n"
-
-                    //                     "pld        [%3, #128]      \n"
-                    "vld1.u16   {d30-d31}, [%3 :64] \n" // r24 r25
-
-                    "vshll.u16  q14, d30, #16   \n"
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmla.f32   q13, %q15, q14  \n"
-                    "vmla.f32   q12, %q16, q14  \n"
-                    "vmla.f32   q13, %q16, q15  \n"
-
-                    "vshrn.u32  d20, q10, #16   \n"
-                    "vshrn.u32  d21, q11, #16   \n"
-                    "vshrn.u32  d22, q12, #16   \n"
-                    "vshrn.u32  d23, q13, #16   \n"
-
-                    "vst1.u16   {d20-d23}, [%0 :64]! \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v12.4h, v13.4h, v14.4h, v15.4h}, [%1] \n" // r00 r01 r02 r03
-
-                    "mov    v18.16b, %17.16b            \n" // sum00
-                    "mov    v19.16b, %17.16b            \n" // sum01
-
-                    "shll   v12.4s, v12.4h, #16         \n"
-                    "shll   v13.4s, v13.4h, #16         \n"
-
-                    "fmul   v16.4s, %8.4s, v12.4s       \n"
-                    "fmul   v17.4s, %8.4s, v13.4s       \n"
-
-                    "shll   v14.4s, v14.4h, #16         \n"
-                    "shll   v15.4s, v15.4h, #16         \n"
-
-                    "fmla   v18.4s, %9.4s, v13.4s       \n"
-                    "fmla   v19.4s, %9.4s, v14.4s       \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%2] \n" // r10 r11 r12 r13
-
-                    "fmla   v16.4s, %10.4s, v14.4s      \n"
-                    "fmla   v17.4s, %10.4s, v15.4s      \n"
-
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-
-                    "fmla   v18.4s, %11.4s, v20.4s      \n"
-                    "fmla   v19.4s, %11.4s, v21.4s      \n"
-
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-
-                    "fmla   v16.4s, %12.4s, v21.4s      \n"
-                    "fmla   v17.4s, %12.4s, v22.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v12.4h, v13.4h, v14.4h, v15.4h}, [%3] \n" // r20 r21 r22 r23
-
-                    "fmla   v18.4s, %13.4s, v22.4s      \n"
-                    "fmla   v19.4s, %13.4s, v23.4s      \n"
-
-                    "shll   v12.4s, v12.4h, #16         \n"
-                    "shll   v13.4s, v13.4h, #16         \n"
-
-                    "fmla   v16.4s, %14.4s, v12.4s      \n"
-                    "fmla   v17.4s, %14.4s, v13.4s      \n"
-
-                    "shll   v14.4s, v14.4h, #16         \n"
-                    "shll   v15.4s, v15.4h, #16         \n"
-
-                    "fmla   v18.4s, %15.4s, v13.4s      \n"
-                    "fmla   v19.4s, %15.4s, v14.4s      \n"
-
-                    "add    %1, %1, #16                 \n"
-
-                    "fmla   v16.4s, %16.4s, v14.4s      \n"
-                    "fmla   v17.4s, %16.4s, v15.4s      \n"
-
-                    "add    %2, %2, #16                 \n"
-
-                    "fadd   v18.4s, v18.4s, v16.4s      \n"
-                    "fadd   v19.4s, v19.4s, v17.4s      \n"
-
-                    "add    %3, %3, #16                 \n"
-
-                    "shrn   v18.4h, v18.4s, #16         \n"
-                    "shrn   v19.4h, v19.4s, #16         \n"
-
-                    "st1    {v18.4h, v19.4h}, [%0], #16 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-#else
-                asm volatile(
-                    "pld        [%1, #256]      \n"
-                    "vld1.u16   {d28-d31}, [%1 :64] \n" // r00 r01 r02 r03
-
-                    "vmov       q10, %q17       \n" // sum00
-                    "vmov       q11, %q17       \n" // sum01
-
-                    "vshll.u16  q12, d28, #16   \n"
-                    "vshll.u16  q13, d29, #16   \n"
-
-                    "vmla.f32   q10, %q8, q12   \n"
-                    "vmla.f32   q11, %q8, q13   \n"
-
-                    "vshll.u16  q14, d30, #16   \n"
-
-                    "vmla.f32   q10, %q9, q13   \n"
-                    "vmla.f32   q11, %q9, q14   \n"
-
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmla.f32   q10, %q10, q14  \n"
-                    "vmla.f32   q11, %q10, q15  \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.u16   {d28-d31}, [%2 :64] \n" // r10 r11 r12 r13
-
-                    "vshll.u16  q12, d28, #16   \n"
-                    "vshll.u16  q13, d29, #16   \n"
-
-                    "vmla.f32   q10, %q11, q12  \n"
-                    "vmla.f32   q11, %q11, q13  \n"
-
-                    "vshll.u16  q14, d30, #16   \n"
-
-                    "vmla.f32   q10, %q12, q13  \n"
-                    "vmla.f32   q11, %q12, q14  \n"
-
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmla.f32   q10, %q13, q14  \n"
-                    "vmla.f32   q11, %q13, q15  \n"
-
-                    "pld        [%3, #256]      \n"
-                    "vld1.u16   {d28-d31}, [%3 :64] \n" // r20 r21 r22 r23
-
-                    "vshll.u16  q12, d28, #16   \n"
-                    "vshll.u16  q13, d29, #16   \n"
-
-                    "vmla.f32   q10, %q14, q12  \n"
-                    "vmla.f32   q11, %q14, q13  \n"
-
-                    "vshll.u16  q14, d30, #16   \n"
-
-                    "vmla.f32   q10, %q15, q13  \n"
-                    "vmla.f32   q11, %q15, q14  \n"
-
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmla.f32   q10, %q16, q14  \n"
-                    "vmla.f32   q11, %q16, q15  \n"
-
-                    "add        %1, %1, #16     \n"
-                    "add        %2, %2, #16     \n"
-
-                    "vshrn.u32  d20, q10, #16   \n"
-                    "vshrn.u32  d21, q11, #16   \n"
-
-                    "add        %3, %3, #16     \n"
-
-                    "vst1.u16   {d20-d21}, [%0 :64]! \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif
-            }
-            for (; j < outw; j++)
-            {
-                float32x4_t _sum0 = _bias0;
-
-                float32x4_t _r00 = bfloat2float(vld1_u16(r0));
-                float32x4_t _r01 = bfloat2float(vld1_u16(r0 + 4));
-                float32x4_t _r02 = bfloat2float(vld1_u16(r0 + 8));
-                float32x4_t _r10 = bfloat2float(vld1_u16(r1));
-                float32x4_t _r11 = bfloat2float(vld1_u16(r1 + 4));
-                float32x4_t _r12 = bfloat2float(vld1_u16(r1 + 8));
-                float32x4_t _r20 = bfloat2float(vld1_u16(r2));
-                float32x4_t _r21 = bfloat2float(vld1_u16(r2 + 4));
-                float32x4_t _r22 = bfloat2float(vld1_u16(r2 + 8));
-
-                _sum0 = vmlaq_f32(_sum0, _k00, _r00);
-                _sum0 = vmlaq_f32(_sum0, _k01, _r01);
-                _sum0 = vmlaq_f32(_sum0, _k02, _r02);
-                _sum0 = vmlaq_f32(_sum0, _k10, _r10);
-                _sum0 = vmlaq_f32(_sum0, _k11, _r11);
-                _sum0 = vmlaq_f32(_sum0, _k12, _r12);
-                _sum0 = vmlaq_f32(_sum0, _k20, _r20);
-                _sum0 = vmlaq_f32(_sum0, _k21, _r21);
-                _sum0 = vmlaq_f32(_sum0, _k22, _r22);
-
-                vst1_u16(outptr0, float2bfloat(_sum0));
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                outptr0 += 4;
-            }
-
-            r0 += 2 * 4;
-            r1 += 2 * 4;
-            r2 += 2 * 4;
-        }
-    }
-}
-
-static void convdw3x3s2_pack4_bf16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + g * 4) : vdupq_n_f32(0.f);
-
-        const unsigned short* k0 = kernel.row<const unsigned short>(g);
-
-        unsigned short* outptr0 = out;
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const unsigned short* r0 = img0.row<const unsigned short>(0);
-        const unsigned short* r1 = img0.row<const unsigned short>(1);
-        const unsigned short* r2 = img0.row<const unsigned short>(2);
-
-        float32x4_t _k00 = bfloat2float(vld1_u16(k0));
-        float32x4_t _k01 = bfloat2float(vld1_u16(k0 + 4));
-        float32x4_t _k02 = bfloat2float(vld1_u16(k0 + 8));
-        float32x4_t _k10 = bfloat2float(vld1_u16(k0 + 12));
-        float32x4_t _k11 = bfloat2float(vld1_u16(k0 + 16));
-        float32x4_t _k12 = bfloat2float(vld1_u16(k0 + 20));
-        float32x4_t _k20 = bfloat2float(vld1_u16(k0 + 24));
-        float32x4_t _k21 = bfloat2float(vld1_u16(k0 + 28));
-        float32x4_t _k22 = bfloat2float(vld1_u16(k0 + 32));
-
-        int i = 0;
-
-        for (; i < outh; i++)
-        {
-            int j = 0;
-
-#if __aarch64__
-            for (; j + 3 < outw; j += 4)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v10.4h, v11.4h, v12.4h, v13.4h}, [%1], #32 \n" // r00 r01 r02 r03
-
-                    "mov    v28.16b, %17.16b            \n" // sum00
-                    "mov    v29.16b, %17.16b            \n" // sum01
-                    "mov    v30.16b, %17.16b            \n" // sum02
-                    "mov    v31.16b, %17.16b            \n" // sum03
-
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v14.4h, v15.4h, v16.4h, v17.4h}, [%1], #32 \n" // r04 r05 r06 r07
-
-                    "shll   v10.4s, v10.4h, #16         \n"
-                    "shll   v11.4s, v11.4h, #16         \n"
-                    "shll   v12.4s, v12.4h, #16         \n"
-                    "shll   v13.4s, v13.4h, #16         \n"
-
-                    "prfm   pldl1keep, [%1, #64]        \n"
-                    "ld1    {v18.4h}, [%1]              \n" // r08
-
-                    "shll   v14.4s, v14.4h, #16         \n"
-                    "shll   v15.4s, v15.4h, #16         \n"
-
-                    "fmla   v28.4s, %8.4s, v10.4s       \n"
-                    "fmla   v29.4s, %8.4s, v12.4s       \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-
-                    "fmla   v30.4s, %8.4s, v14.4s       \n"
-                    "fmla   v31.4s, %8.4s, v16.4s       \n"
-
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "fmla   v28.4s, %9.4s, v11.4s       \n"
-                    "fmla   v29.4s, %9.4s, v13.4s       \n"
-                    "fmla   v30.4s, %9.4s, v15.4s       \n"
-                    "fmla   v31.4s, %9.4s, v17.4s       \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%2], #32 \n" // r10 r11 r12 r13
-
-                    "fmla   v28.4s, %10.4s, v12.4s      \n"
-                    "fmla   v29.4s, %10.4s, v14.4s      \n"
-
-                    "shll   v18.4s, v18.4h, #16         \n"
-
-                    "fmla   v30.4s, %10.4s, v16.4s      \n"
-                    "fmla   v31.4s, %10.4s, v18.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%2], #32 \n" // r14 r15 r16 r17
-
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-
-                    "prfm   pldl1keep, [%2, #64]        \n"
-                    "ld1    {v19.4h}, [%2]              \n" // r18
-
-                    "shll   v24.4s, v24.4h, #16         \n"
-                    "shll   v25.4s, v25.4h, #16         \n"
-
-                    "fmla   v28.4s, %11.4s, v20.4s      \n"
-                    "fmla   v29.4s, %11.4s, v22.4s      \n"
-
-                    "shll   v26.4s, v26.4h, #16         \n"
-
-                    "fmla   v30.4s, %11.4s, v24.4s      \n"
-                    "fmla   v31.4s, %11.4s, v26.4s      \n"
-
-                    "shll   v27.4s, v27.4h, #16         \n"
-
-                    "fmla   v28.4s, %12.4s, v21.4s      \n"
-                    "fmla   v29.4s, %12.4s, v23.4s      \n"
-                    "fmla   v30.4s, %12.4s, v25.4s      \n"
-                    "fmla   v31.4s, %12.4s, v27.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v10.4h, v11.4h, v12.4h, v13.4h}, [%3], #32 \n" // r20 r21 r22 r23
-
-                    "fmla   v28.4s, %13.4s, v22.4s      \n"
-                    "fmla   v29.4s, %13.4s, v24.4s      \n"
-
-                    "shll   v19.4s, v19.4h, #16         \n"
-
-                    "fmla   v30.4s, %13.4s, v26.4s      \n"
-                    "fmla   v31.4s, %13.4s, v19.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v14.4h, v15.4h, v16.4h, v17.4h}, [%3], #32 \n" // r24 r25 r26 r27
-
-                    "shll   v10.4s, v10.4h, #16         \n"
-                    "shll   v11.4s, v11.4h, #16         \n"
-                    "shll   v12.4s, v12.4h, #16         \n"
-                    "shll   v13.4s, v13.4h, #16         \n"
-
-                    "prfm   pldl1keep, [%3, #64]        \n"
-                    "ld1    {v18.4h}, [%3]              \n" // r28
-
-                    "shll   v14.4s, v14.4h, #16         \n"
-                    "shll   v15.4s, v15.4h, #16         \n"
-
-                    "fmla   v28.4s, %14.4s, v10.4s      \n"
-                    "fmla   v29.4s, %14.4s, v12.4s      \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-
-                    "fmla   v30.4s, %14.4s, v14.4s      \n"
-                    "fmla   v31.4s, %14.4s, v16.4s      \n"
-
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "fmla   v28.4s, %15.4s, v11.4s      \n"
-                    "fmla   v29.4s, %15.4s, v13.4s      \n"
-                    "fmla   v30.4s, %15.4s, v15.4s      \n"
-                    "fmla   v31.4s, %15.4s, v17.4s      \n"
-
-                    "fmla   v28.4s, %16.4s, v12.4s      \n"
-                    "fmla   v29.4s, %16.4s, v14.4s      \n"
-
-                    "shll   v18.4s, v18.4h, #16         \n"
-
-                    "fmla   v30.4s, %16.4s, v16.4s      \n"
-                    "fmla   v31.4s, %16.4s, v18.4s      \n"
-
-                    "shrn   v28.4h, v28.4s, #16         \n"
-                    "shrn   v29.4h, v29.4s, #16         \n"
-                    "shrn   v30.4h, v30.4s, #16         \n"
-                    "shrn   v31.4h, v31.4s, #16         \n"
-
-                    "st1    {v28.4h, v29.4h, v30.4h, v31.4h}, [%0], #32 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-#endif // __aarch64__
-            for (; j + 1 < outw; j += 2)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v10.4h, v11.4h, v12.4h, v13.4h}, [%1], #32 \n" // r00 r01 r02 r03
-
-                    "mov    v22.16b, %17.16b            \n" // sum00
-                    "mov    v23.16b, %17.16b            \n" // sum01
-
-                    "shll   v10.4s, v10.4h, #16         \n"
-                    "shll   v11.4s, v11.4h, #16         \n"
-
-                    "fmul   v20.4s, %8.4s, v10.4s       \n"
-
-                    "shll   v12.4s, v12.4h, #16         \n"
-                    "shll   v13.4s, v13.4h, #16         \n"
-
-                    "fmul   v21.4s, %8.4s, v12.4s       \n"
-
-                    "prfm   pldl1keep, [%1, #64]        \n"
-                    "ld1    {v14.4h}, [%1]              \n" // r04
-
-                    "fmla   v22.4s, %9.4s, v11.4s       \n"
-                    "fmla   v23.4s, %9.4s, v13.4s       \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%2], #32 \n" // r10 r11 r12 r13
-
-                    "shll   v14.4s, v14.4h, #16         \n"
-
-                    "fmla   v20.4s, %10.4s, v12.4s      \n"
-                    "fmla   v21.4s, %10.4s, v14.4s      \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "fmla   v22.4s, %11.4s, v16.4s      \n"
-
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-
-                    "fmla   v23.4s, %11.4s, v18.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #64]        \n"
-                    "ld1    {v15.4h}, [%2]              \n" // r14
-
-                    "fmla   v20.4s, %12.4s, v17.4s      \n"
-                    "fmla   v21.4s, %12.4s, v19.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v10.4h, v11.4h, v12.4h, v13.4h}, [%3], #32 \n" // r20 r21 r22 r23
-
-                    "shll   v15.4s, v15.4h, #16         \n"
-
-                    "fmla   v22.4s, %13.4s, v18.4s      \n"
-                    "fmla   v23.4s, %13.4s, v15.4s      \n"
-
-                    "shll   v10.4s, v10.4h, #16         \n"
-                    "shll   v11.4s, v11.4h, #16         \n"
-
-                    "fmla   v20.4s, %14.4s, v10.4s      \n"
-
-                    "shll   v12.4s, v12.4h, #16         \n"
-                    "shll   v13.4s, v13.4h, #16         \n"
-
-                    "fmla   v21.4s, %14.4s, v12.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #64]        \n"
-                    "ld1    {v14.4h}, [%3]              \n" // r24
-
-                    "fmla   v22.4s, %15.4s, v11.4s      \n"
-                    "fmla   v23.4s, %15.4s, v13.4s      \n"
-
-                    "shll   v14.4s, v14.4h, #16         \n"
-
-                    "fmla   v20.4s, %16.4s, v12.4s      \n"
-                    "fmla   v21.4s, %16.4s, v14.4s      \n"
-
-                    "fadd   v22.4s, v20.4s, v22.4s      \n"
-                    "fadd   v23.4s, v21.4s, v23.4s      \n"
-
-                    "shrn   v22.4h, v22.4s, #16         \n"
-                    "shrn   v23.4h, v23.4s, #16         \n"
-
-                    "st1    {v22.4h, v23.4h}, [%0], #16 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-#else
-                asm volatile(
-                    "pld        [%1, #256]      \n"
-                    "vld1.u16   {d28-d31}, [%1 :64]! \n" // r00 r01 r02 r03
-
-                    "vmov       q10, %q17       \n" // sum00
-                    "vmov       q11, %q17       \n" // sum01
-
-                    "vshll.u16  q12, d28, #16   \n"
-                    "vshll.u16  q13, d29, #16   \n"
-
-                    "vmla.f32   q10, %q8, q12   \n"
-
-                    "vshll.u16  q14, d30, #16   \n"
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmla.f32   q11, %q8, q14   \n"
-
-                    "vld1.u16   {d25}, [%1]     \n" // r04
-
-                    "vmla.f32   q10, %q9, q13   \n"
-                    "vmla.f32   q11, %q9, q15   \n"
-
-                    "vshll.u16  q12, d25, #16   \n"
-
-                    "vmla.f32   q10, %q10, q14  \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.u16   {d28-d31}, [%2 :64]! \n" // r10 r11 r12 r13
-
-                    "vmla.f32   q11, %q10, q12  \n"
-
-                    "vshll.u16  q12, d28, #16   \n"
-                    "vshll.u16  q13, d29, #16   \n"
-
-                    "vmla.f32   q10, %q11, q12  \n"
-
-                    "vshll.u16  q14, d30, #16   \n"
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmla.f32   q11, %q11, q14  \n"
-
-                    "vld1.u16   {d25}, [%2]     \n" // r14
-
-                    "vmla.f32   q10, %q12, q13  \n"
-                    "vmla.f32   q11, %q12, q15  \n"
-
-                    "vshll.u16  q12, d25, #16   \n"
-
-                    "vmla.f32   q10, %q13, q14  \n"
-
-                    "pld        [%3, #256]      \n"
-                    "vld1.u16   {d28-d31}, [%3 :64]! \n" // r20 r21 r22 r23
-
-                    "vmla.f32   q11, %q13, q12  \n"
-
-                    "vshll.u16  q12, d28, #16   \n"
-                    "vshll.u16  q13, d29, #16   \n"
-
-                    "vmla.f32   q10, %q14, q12  \n"
-
-                    "vshll.u16  q14, d30, #16   \n"
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmla.f32   q11, %q14, q14  \n"
-
-                    "vld1.u16   {d25}, [%3]     \n" // r24
-
-                    "vmla.f32   q10, %q15, q13  \n"
-                    "vmla.f32   q11, %q15, q15  \n"
-
-                    "vshll.u16  q12, d25, #16   \n"
-
-                    "vmla.f32   q10, %q16, q14  \n"
-                    "vmla.f32   q11, %q16, q12  \n"
-
-                    "vshrn.u32  d20, q10, #16   \n"
-                    "vshrn.u32  d21, q11, #16   \n"
-
-                    "vst1.u16   {d20-d21}, [%0 :64]! \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif
-            }
-            for (; j < outw; j++)
-            {
-                float32x4_t _sum0 = _bias0;
-
-                float32x4_t _r00 = bfloat2float(vld1_u16(r0));
-                float32x4_t _r01 = bfloat2float(vld1_u16(r0 + 4));
-                float32x4_t _r02 = bfloat2float(vld1_u16(r0 + 8));
-                float32x4_t _r10 = bfloat2float(vld1_u16(r1));
-                float32x4_t _r11 = bfloat2float(vld1_u16(r1 + 4));
-                float32x4_t _r12 = bfloat2float(vld1_u16(r1 + 8));
-                float32x4_t _r20 = bfloat2float(vld1_u16(r2));
-                float32x4_t _r21 = bfloat2float(vld1_u16(r2 + 4));
-                float32x4_t _r22 = bfloat2float(vld1_u16(r2 + 8));
-
-                _sum0 = vmlaq_f32(_sum0, _k00, _r00);
-                _sum0 = vmlaq_f32(_sum0, _k01, _r01);
-                _sum0 = vmlaq_f32(_sum0, _k02, _r02);
-                _sum0 = vmlaq_f32(_sum0, _k10, _r10);
-                _sum0 = vmlaq_f32(_sum0, _k11, _r11);
-                _sum0 = vmlaq_f32(_sum0, _k12, _r12);
-                _sum0 = vmlaq_f32(_sum0, _k20, _r20);
-                _sum0 = vmlaq_f32(_sum0, _k21, _r21);
-                _sum0 = vmlaq_f32(_sum0, _k22, _r22);
-
-                vst1_u16(outptr0, float2bfloat(_sum0));
-
-                r0 += 2 * 4;
-                r1 += 2 * 4;
-                r2 += 2 * 4;
-                outptr0 += 4;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack8_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack8_fp16s.h
deleted file mode 100644
index 05ca6af9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack8_fp16s.h
+++ /dev/null
@@ -1,909 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_pack8_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        float16x8_t _bias0 = bias ? vld1q_f16(bias + g * 8) : vdupq_n_f16((__fp16)0.f);
-
-        const __fp16* k0 = kernel.row<const __fp16>(g);
-
-        __fp16* outptr0 = out.row<__fp16>(0);
-        __fp16* outptr1 = out.row<__fp16>(1);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const __fp16* r0 = img0.row<const __fp16>(0);
-        const __fp16* r1 = img0.row<const __fp16>(1);
-        const __fp16* r2 = img0.row<const __fp16>(2);
-        const __fp16* r3 = img0.row<const __fp16>(3);
-
-        float16x8_t _k00 = vld1q_f16(k0);
-        float16x8_t _k01 = vld1q_f16(k0 + 8);
-        float16x8_t _k02 = vld1q_f16(k0 + 16);
-        float16x8_t _k10 = vld1q_f16(k0 + 24);
-        float16x8_t _k11 = vld1q_f16(k0 + 32);
-        float16x8_t _k12 = vld1q_f16(k0 + 40);
-        float16x8_t _k20 = vld1q_f16(k0 + 48);
-        float16x8_t _k21 = vld1q_f16(k0 + 56);
-        float16x8_t _k22 = vld1q_f16(k0 + 64);
-
-        int i = 0;
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%3], #64 \n" // r10 r11 r12 r13
-
-                    "mov    v24.16b, %21.16b            \n" // sum00
-                    "mov    v25.16b, %21.16b            \n" // sum01
-                    "mov    v26.16b, %21.16b            \n" // sum02
-                    "mov    v27.16b, %21.16b            \n" // sum03
-
-                    "fmla   v24.8h, %15.8h, v12.8h      \n"
-                    "fmla   v25.8h, %15.8h, v13.8h      \n"
-
-                    "mov    v28.16b, %21.16b            \n" // sum10
-                    "mov    v29.16b, %21.16b            \n" // sum11
-                    "mov    v30.16b, %21.16b            \n" // sum12
-                    "mov    v31.16b, %21.16b            \n" // sum13
-
-                    "fmla   v26.8h, %15.8h, v14.8h      \n"
-                    "fmla   v27.8h, %15.8h, v15.8h      \n"
-
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v16.8h, v17.8h}, [%3]      \n" // r14 r15
-
-                    "fmla   v28.8h, %12.8h, v12.8h      \n"
-                    "fmla   v29.8h, %12.8h, v13.8h      \n"
-                    "fmla   v30.8h, %12.8h, v14.8h      \n"
-                    "fmla   v31.8h, %12.8h, v15.8h      \n"
-
-                    "fmla   v24.8h, %16.8h, v13.8h      \n"
-                    "fmla   v25.8h, %16.8h, v14.8h      \n"
-                    "fmla   v26.8h, %16.8h, v15.8h      \n"
-                    "fmla   v27.8h, %16.8h, v16.8h      \n"
-
-                    "fmla   v28.8h, %13.8h, v13.8h      \n"
-                    "fmla   v29.8h, %13.8h, v14.8h      \n"
-                    "fmla   v30.8h, %13.8h, v15.8h      \n"
-                    "fmla   v31.8h, %13.8h, v16.8h      \n"
-
-                    "prfm   pldl1keep, [%4, #512]       \n"
-                    "ld1    {v18.8h, v19.8h, v20.8h, v21.8h}, [%4], #64 \n" // r20 r21 r22 r23
-
-                    "fmla   v24.8h, %17.8h, v14.8h      \n"
-                    "fmla   v25.8h, %17.8h, v15.8h      \n"
-                    "fmla   v26.8h, %17.8h, v16.8h      \n"
-                    "fmla   v27.8h, %17.8h, v17.8h      \n"
-
-                    "fmla   v28.8h, %14.8h, v14.8h      \n"
-                    "fmla   v29.8h, %14.8h, v15.8h      \n"
-                    "fmla   v30.8h, %14.8h, v16.8h      \n"
-                    "fmla   v31.8h, %14.8h, v17.8h      \n"
-
-                    "fmla   v24.8h, %18.8h, v18.8h      \n"
-                    "fmla   v25.8h, %18.8h, v19.8h      \n"
-                    "fmla   v26.8h, %18.8h, v20.8h      \n"
-                    "fmla   v27.8h, %18.8h, v21.8h      \n"
-
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v22.8h, v23.8h}, [%4]      \n" // r24 r25
-
-                    "fmla   v28.8h, %15.8h, v18.8h      \n"
-                    "fmla   v29.8h, %15.8h, v19.8h      \n"
-                    "fmla   v30.8h, %15.8h, v20.8h      \n"
-                    "fmla   v31.8h, %15.8h, v21.8h      \n"
-
-                    "fmla   v24.8h, %19.8h, v19.8h      \n"
-                    "fmla   v25.8h, %19.8h, v20.8h      \n"
-                    "fmla   v26.8h, %19.8h, v21.8h      \n"
-                    "fmla   v27.8h, %19.8h, v22.8h      \n"
-
-                    "fmla   v28.8h, %16.8h, v19.8h      \n"
-                    "fmla   v29.8h, %16.8h, v20.8h      \n"
-                    "fmla   v30.8h, %16.8h, v21.8h      \n"
-                    "fmla   v31.8h, %16.8h, v22.8h      \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%2], #64 \n" // r00 r01 r02 r03
-
-                    "fmla   v24.8h, %20.8h, v20.8h      \n"
-                    "fmla   v25.8h, %20.8h, v21.8h      \n"
-                    "fmla   v26.8h, %20.8h, v22.8h      \n"
-                    "fmla   v27.8h, %20.8h, v23.8h      \n"
-
-                    "fmla   v28.8h, %17.8h, v20.8h      \n"
-                    "fmla   v29.8h, %17.8h, v21.8h      \n"
-                    "fmla   v30.8h, %17.8h, v22.8h      \n"
-                    "fmla   v31.8h, %17.8h, v23.8h      \n"
-
-                    "prfm   pldl1keep, [%5, #512]       \n"
-                    "ld1    {v18.8h, v19.8h, v20.8h, v21.8h}, [%5], #64 \n" // r30 r31 r32 r33
-
-                    "fmla   v24.8h, %12.8h, v12.8h      \n"
-                    "fmla   v25.8h, %12.8h, v13.8h      \n"
-                    "fmla   v26.8h, %12.8h, v14.8h      \n"
-                    "fmla   v27.8h, %12.8h, v15.8h      \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v16.8h, v17.8h}, [%2]      \n" // r04 r05
-
-                    "fmla   v28.8h, %18.8h, v18.8h      \n"
-                    "fmla   v29.8h, %18.8h, v19.8h      \n"
-                    "fmla   v30.8h, %18.8h, v20.8h      \n"
-                    "fmla   v31.8h, %18.8h, v21.8h      \n"
-
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v22.8h, v23.8h}, [%5]      \n" // r34 r35
-
-                    "fmla   v24.8h, %13.8h, v13.8h      \n"
-                    "fmla   v25.8h, %13.8h, v14.8h      \n"
-                    "fmla   v26.8h, %13.8h, v15.8h      \n"
-                    "fmla   v27.8h, %13.8h, v16.8h      \n"
-
-                    "fmla   v28.8h, %19.8h, v19.8h      \n"
-                    "fmla   v29.8h, %19.8h, v20.8h      \n"
-                    "fmla   v30.8h, %19.8h, v21.8h      \n"
-                    "fmla   v31.8h, %19.8h, v22.8h      \n"
-
-                    "fmla   v24.8h, %14.8h, v14.8h      \n"
-                    "fmla   v25.8h, %14.8h, v15.8h      \n"
-                    "fmla   v26.8h, %14.8h, v16.8h      \n"
-                    "fmla   v27.8h, %14.8h, v17.8h      \n"
-
-                    "fmla   v28.8h, %20.8h, v20.8h      \n"
-                    "fmla   v29.8h, %20.8h, v21.8h      \n"
-                    "fmla   v30.8h, %20.8h, v22.8h      \n"
-                    "fmla   v31.8h, %20.8h, v23.8h      \n"
-
-                    "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%1], #64 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3)       // %5
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "w"(_k00),  // %12
-                    "w"(_k01),  // %13
-                    "w"(_k02),  // %14
-                    "w"(_k10),  // %15
-                    "w"(_k11),  // %16
-                    "w"(_k12),  // %17
-                    "w"(_k20),  // %18
-                    "w"(_k21),  // %19
-                    "w"(_k22),  // %20
-                    "w"(_bias0) // %21
-                    : "memory", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%3] \n" // r10 r11 r12 r13
-
-                    "mov    v28.16b, %21.16b            \n" // sum00
-                    "mov    v29.16b, %21.16b            \n" // sum01
-                    "mov    v30.16b, %21.16b            \n" // sum10
-                    "mov    v31.16b, %21.16b            \n" // sum11
-
-                    "fmla   v28.8h, %15.8h, v16.8h      \n"
-                    "fmla   v30.8h, %12.8h, v16.8h      \n"
-                    "fmla   v29.8h, %15.8h, v17.8h      \n"
-                    "fmla   v31.8h, %12.8h, v17.8h      \n"
-
-                    "prfm   pldl1keep, [%4, #512]       \n"
-                    "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4] \n" // r20 r21 r22 r23
-
-                    "fmla   v28.8h, %16.8h, v17.8h      \n"
-                    "fmla   v30.8h, %13.8h, v17.8h      \n"
-                    "fmla   v29.8h, %16.8h, v18.8h      \n"
-                    "fmla   v31.8h, %13.8h, v18.8h      \n"
-
-                    "fmla   v28.8h, %17.8h, v18.8h      \n"
-                    "fmla   v30.8h, %14.8h, v18.8h      \n"
-                    "fmla   v29.8h, %17.8h, v19.8h      \n"
-                    "fmla   v31.8h, %14.8h, v19.8h      \n"
-
-                    "fmla   v28.8h, %18.8h, v20.8h      \n"
-                    "fmla   v30.8h, %15.8h, v20.8h      \n"
-                    "fmla   v29.8h, %18.8h, v21.8h      \n"
-                    "fmla   v31.8h, %15.8h, v21.8h      \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%2] \n" // r00 r01 r02 r03
-
-                    "fmla   v28.8h, %19.8h, v21.8h      \n"
-                    "fmla   v30.8h, %16.8h, v21.8h      \n"
-                    "fmla   v29.8h, %19.8h, v22.8h      \n"
-                    "fmla   v31.8h, %16.8h, v22.8h      \n"
-
-                    "prfm   pldl1keep, [%5, #512]       \n"
-                    "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%5] \n" // r30 r31 r32 r33
-
-                    "fmla   v28.8h, %20.8h, v22.8h      \n"
-                    "fmla   v30.8h, %17.8h, v22.8h      \n"
-                    "fmla   v29.8h, %20.8h, v23.8h      \n"
-                    "fmla   v31.8h, %17.8h, v23.8h      \n"
-
-                    "fmla   v28.8h, %12.8h, v12.8h      \n"
-                    "fmla   v30.8h, %18.8h, v24.8h      \n"
-                    "fmla   v29.8h, %12.8h, v13.8h      \n"
-                    "fmla   v31.8h, %18.8h, v25.8h      \n"
-                    "fmla   v28.8h, %13.8h, v13.8h      \n"
-                    "fmla   v30.8h, %19.8h, v25.8h      \n"
-                    "fmla   v29.8h, %13.8h, v14.8h      \n"
-                    "fmla   v31.8h, %19.8h, v26.8h      \n"
-                    "fmla   v28.8h, %14.8h, v14.8h      \n"
-                    "fmla   v30.8h, %20.8h, v26.8h      \n"
-                    "fmla   v29.8h, %14.8h, v15.8h      \n"
-                    "fmla   v31.8h, %20.8h, v27.8h      \n"
-
-                    "add    %2, %2, #32                 \n"
-                    "add    %3, %3, #32                 \n"
-                    "add    %4, %4, #32                 \n"
-                    "add    %5, %5, #32                 \n"
-
-                    "st1    {v28.8h, v29.8h}, [%0], #32 \n"
-                    "st1    {v30.8h, v31.8h}, [%1], #32 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3)       // %5
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "w"(_k00),  // %12
-                    "w"(_k01),  // %13
-                    "w"(_k02),  // %14
-                    "w"(_k10),  // %15
-                    "w"(_k11),  // %16
-                    "w"(_k12),  // %17
-                    "w"(_k20),  // %18
-                    "w"(_k21),  // %19
-                    "w"(_k22),  // %20
-                    "w"(_bias0) // %21
-                    : "memory", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            for (; j < outw; j++)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%3, #384]       \n"
-                    "ld1    {v15.8h, v16.8h, v17.8h}, [%3] \n" // r10 r11 r12
-
-                    "mov    v28.16b, %21.16b            \n" // sum00
-                    "mov    v30.16b, %21.16b            \n" // sum10
-
-                    "fmul   v29.8h, %15.8h, v15.8h      \n"
-                    "fmul   v31.8h, %12.8h, v15.8h      \n"
-
-                    "prfm   pldl1keep, [%4, #384]       \n"
-                    "ld1    {v18.8h, v19.8h, v20.8h}, [%4] \n" // r20 r21 r22
-
-                    "fmla   v28.8h, %16.8h, v16.8h      \n"
-                    "fmla   v30.8h, %13.8h, v16.8h      \n"
-
-                    "fmla   v29.8h, %17.8h, v17.8h      \n"
-                    "fmla   v31.8h, %14.8h, v17.8h      \n"
-
-                    "prfm   pldl1keep, [%2, #384]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h}, [%2] \n" // r00 r01 r02
-
-                    "fmla   v28.8h, %18.8h, v18.8h      \n"
-                    "fmla   v30.8h, %15.8h, v18.8h      \n"
-
-                    "fmla   v29.8h, %19.8h, v19.8h      \n"
-                    "fmla   v31.8h, %16.8h, v19.8h      \n"
-
-                    "prfm   pldl1keep, [%5, #384]       \n"
-                    "ld1    {v21.8h, v22.8h, v23.8h}, [%5] \n" // r30 r31 r32
-
-                    "fmla   v28.8h, %20.8h, v20.8h      \n"
-                    "fmla   v30.8h, %17.8h, v20.8h      \n"
-
-                    "fmla   v29.8h, %12.8h, v12.8h      \n"
-                    "fmla   v31.8h, %18.8h, v21.8h      \n"
-                    "fmla   v28.8h, %13.8h, v13.8h      \n"
-                    "fmla   v30.8h, %19.8h, v22.8h      \n"
-                    "fmla   v29.8h, %14.8h, v14.8h      \n"
-                    "fmla   v31.8h, %20.8h, v23.8h      \n"
-
-                    "add    %2, %2, #16                 \n"
-                    "add    %3, %3, #16                 \n"
-
-                    "fadd   v28.8h, v28.8h, v29.8h      \n"
-                    "fadd   v30.8h, v30.8h, v31.8h      \n"
-
-                    "add    %4, %4, #16                 \n"
-                    "add    %5, %5, #16                 \n"
-
-                    "st1    {v28.8h}, [%0], #16         \n"
-                    "st1    {v30.8h}, [%1], #16         \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3)       // %5
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "w"(_k00),  // %12
-                    "w"(_k01),  // %13
-                    "w"(_k02),  // %14
-                    "w"(_k10),  // %15
-                    "w"(_k11),  // %16
-                    "w"(_k12),  // %17
-                    "w"(_k20),  // %18
-                    "w"(_k21),  // %19
-                    "w"(_k22),  // %20
-                    "w"(_bias0) // %21
-                    : "memory", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-            }
-
-            r0 += 2 * 8 + w * 8;
-            r1 += 2 * 8 + w * 8;
-            r2 += 2 * 8 + w * 8;
-            r3 += 2 * 8 + w * 8;
-
-            outptr0 += outw * 8;
-            outptr1 += outw * 8;
-        }
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%1], #64 \n" // r00 r01 r02 r03
-
-                    "mov    v28.16b, %17.16b            \n" // sum00
-                    "mov    v29.16b, %17.16b            \n" // sum01
-                    "mov    v30.16b, %17.16b            \n" // sum02
-                    "mov    v31.16b, %17.16b            \n" // sum03
-
-                    "fmla   v28.8h, %8.8h, v12.8h       \n"
-                    "fmla   v29.8h, %8.8h, v13.8h       \n"
-                    "fmla   v30.8h, %8.8h, v14.8h       \n"
-                    "fmla   v31.8h, %8.8h, v15.8h       \n"
-
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v16.8h, v17.8h}, [%1]      \n" // r04 r05
-
-                    "fmla   v28.8h, %9.8h, v13.8h       \n"
-                    "fmla   v29.8h, %9.8h, v14.8h       \n"
-                    "fmla   v30.8h, %9.8h, v15.8h       \n"
-                    "fmla   v31.8h, %9.8h, v16.8h       \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v18.8h, v19.8h, v20.8h, v21.8h}, [%2], #64 \n" // r10 r11 r12 r13
-
-                    "fmla   v28.8h, %10.8h, v14.8h      \n"
-                    "fmla   v29.8h, %10.8h, v15.8h      \n"
-                    "fmla   v30.8h, %10.8h, v16.8h      \n"
-                    "fmla   v31.8h, %10.8h, v17.8h      \n"
-
-                    "fmla   v28.8h, %11.8h, v18.8h      \n"
-                    "fmla   v29.8h, %11.8h, v19.8h      \n"
-                    "fmla   v30.8h, %11.8h, v20.8h      \n"
-                    "fmla   v31.8h, %11.8h, v21.8h      \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v22.8h, v23.8h}, [%2]      \n" // r14 r15
-
-                    "fmla   v28.8h, %12.8h, v19.8h      \n"
-                    "fmla   v29.8h, %12.8h, v20.8h      \n"
-                    "fmla   v30.8h, %12.8h, v21.8h      \n"
-                    "fmla   v31.8h, %12.8h, v22.8h      \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%3], #64 \n" // r20 r21 r22 r23
-
-                    "fmla   v28.8h, %13.8h, v20.8h      \n"
-                    "fmla   v29.8h, %13.8h, v21.8h      \n"
-                    "fmla   v30.8h, %13.8h, v22.8h      \n"
-                    "fmla   v31.8h, %13.8h, v23.8h      \n"
-
-                    "fmla   v28.8h, %14.8h, v12.8h      \n"
-                    "fmla   v29.8h, %14.8h, v13.8h      \n"
-                    "fmla   v30.8h, %14.8h, v14.8h      \n"
-                    "fmla   v31.8h, %14.8h, v15.8h      \n"
-
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v16.8h, v17.8h}, [%3]      \n" // r24 r25
-
-                    "fmla   v28.8h, %15.8h, v13.8h      \n"
-                    "fmla   v29.8h, %15.8h, v14.8h      \n"
-                    "fmla   v30.8h, %15.8h, v15.8h      \n"
-                    "fmla   v31.8h, %15.8h, v16.8h      \n"
-
-                    "fmla   v28.8h, %16.8h, v14.8h      \n"
-                    "fmla   v29.8h, %16.8h, v15.8h      \n"
-                    "fmla   v30.8h, %16.8h, v16.8h      \n"
-                    "fmla   v31.8h, %16.8h, v17.8h      \n"
-
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%1] \n" // r00 r01 r02 r03
-
-                    "mov    v28.16b, %17.16b            \n" // sum00
-                    "mov    v29.16b, %17.16b            \n" // sum01
-
-                    "fmul   v30.8h, %8.8h, v12.8h       \n"
-                    "fmul   v31.8h, %8.8h, v13.8h       \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%2] \n" // r10 r11 r12 r13
-
-                    "fmla   v28.8h, %9.8h, v13.8h       \n"
-                    "fmla   v29.8h, %9.8h, v14.8h       \n"
-                    "fmla   v30.8h, %10.8h, v14.8h      \n"
-                    "fmla   v31.8h, %10.8h, v15.8h      \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%3] \n" // r20 r21 r22 r23
-
-                    "fmla   v28.8h, %11.8h, v16.8h      \n"
-                    "fmla   v29.8h, %11.8h, v17.8h      \n"
-                    "fmla   v30.8h, %12.8h, v17.8h      \n"
-                    "fmla   v31.8h, %12.8h, v18.8h      \n"
-                    "fmla   v28.8h, %13.8h, v18.8h      \n"
-                    "fmla   v29.8h, %13.8h, v19.8h      \n"
-
-                    "fmla   v30.8h, %14.8h, v20.8h      \n"
-                    "fmla   v31.8h, %14.8h, v21.8h      \n"
-                    "fmla   v28.8h, %15.8h, v21.8h      \n"
-                    "fmla   v29.8h, %15.8h, v22.8h      \n"
-                    "fmla   v30.8h, %16.8h, v22.8h      \n"
-                    "fmla   v31.8h, %16.8h, v23.8h      \n"
-
-                    "add    %1, %1, #32                 \n"
-
-                    "fadd   v28.8h, v28.8h, v30.8h      \n"
-                    "fadd   v29.8h, v29.8h, v31.8h      \n"
-
-                    "add    %2, %2, #32                 \n"
-                    "add    %3, %3, #32                 \n"
-
-                    "st1    {v28.8h, v29.8h}, [%0], #32 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v28", "v29", "v30", "v31");
-            }
-            for (; j < outw; j++)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #384]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h}, [%1] \n" // r00 r01 r02
-
-                    "mov    v28.16b, %17.16b            \n" // sum00
-
-                    "fmul   v29.8h, %8.8h, v12.8h       \n"
-
-                    "prfm   pldl1keep, [%2, #384]       \n"
-                    "ld1    {v15.8h, v16.8h, v17.8h}, [%2] \n" // r10 r11 r12
-
-                    "fmul   v30.8h, %9.8h, v13.8h       \n"
-                    "fmla   v28.8h, %10.8h, v14.8h      \n"
-
-                    "prfm   pldl1keep, [%3, #384]       \n"
-                    "ld1    {v18.8h, v19.8h, v20.8h}, [%3] \n" // r20 r21 r22
-
-                    "fmla   v29.8h, %11.8h, v15.8h      \n"
-                    "fmla   v30.8h, %12.8h, v16.8h      \n"
-                    "fmla   v28.8h, %13.8h, v17.8h      \n"
-
-                    "fmla   v29.8h, %14.8h, v18.8h      \n"
-                    "fmla   v30.8h, %15.8h, v19.8h      \n"
-                    "fmla   v28.8h, %16.8h, v20.8h      \n"
-
-                    "add    %1, %1, #16                 \n"
-
-                    "fadd   v29.8h, v29.8h, v30.8h      \n"
-                    "fadd   v28.8h, v28.8h, v29.8h      \n"
-
-                    "add    %2, %2, #16                 \n"
-                    "add    %3, %3, #16                 \n"
-
-                    "st1    {v28.8h}, [%0], #16         \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v28", "v29", "v30");
-            }
-
-            r0 += 2 * 8;
-            r1 += 2 * 8;
-            r2 += 2 * 8;
-        }
-    }
-}
-
-static void convdw3x3s2_pack8_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 8;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        float16x8_t _bias0 = bias ? vld1q_f16(bias + g * 8) : vdupq_n_f16((__fp16)0.f);
-
-        const __fp16* k0 = kernel.row<const __fp16>(g);
-
-        __fp16* outptr0 = out;
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const __fp16* r0 = img0.row<const __fp16>(0);
-        const __fp16* r1 = img0.row<const __fp16>(1);
-        const __fp16* r2 = img0.row<const __fp16>(2);
-
-        float16x8_t _k00 = vld1q_f16(k0);
-        float16x8_t _k01 = vld1q_f16(k0 + 8);
-        float16x8_t _k02 = vld1q_f16(k0 + 16);
-        float16x8_t _k10 = vld1q_f16(k0 + 24);
-        float16x8_t _k11 = vld1q_f16(k0 + 32);
-        float16x8_t _k12 = vld1q_f16(k0 + 40);
-        float16x8_t _k20 = vld1q_f16(k0 + 48);
-        float16x8_t _k21 = vld1q_f16(k0 + 56);
-        float16x8_t _k22 = vld1q_f16(k0 + 64);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%1], #64 \n" // r00 r01 r02 r03
-
-                    "mov    v28.16b, %17.16b            \n" // sum00
-                    "mov    v29.16b, %17.16b            \n" // sum01
-                    "mov    v30.16b, %17.16b            \n" // sum02
-                    "mov    v31.16b, %17.16b            \n" // sum03
-
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%1], #64 \n" // r04 r05 r06 r07
-
-                    "fmla   v28.8h, %8.8h, v0.8h        \n"
-                    "fmla   v29.8h, %8.8h, v2.8h        \n"
-                    "fmla   v30.8h, %8.8h, v4.8h        \n"
-                    "fmla   v31.8h, %8.8h, v6.8h        \n"
-
-                    "prfm   pldl1keep, [%1, #128]       \n"
-                    "ld1    {v8.8h}, [%1]               \n" // r08
-
-                    "fmla   v28.8h, %9.8h, v1.8h        \n"
-                    "fmla   v29.8h, %9.8h, v3.8h        \n"
-                    "fmla   v30.8h, %9.8h, v5.8h        \n"
-                    "fmla   v31.8h, %9.8h, v7.8h        \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%2], #64 \n" // r10 r11 r12 r13
-
-                    "fmla   v28.8h, %10.8h, v2.8h       \n"
-                    "fmla   v29.8h, %10.8h, v4.8h       \n"
-                    "fmla   v30.8h, %10.8h, v6.8h       \n"
-                    "fmla   v31.8h, %10.8h, v8.8h       \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%2], #64 \n" // r14 r15 r16 r17
-
-                    "fmla   v28.8h, %11.8h, v16.8h      \n"
-                    "fmla   v29.8h, %11.8h, v18.8h      \n"
-                    "fmla   v30.8h, %11.8h, v20.8h      \n"
-                    "fmla   v31.8h, %11.8h, v22.8h      \n"
-
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "ld1    {v24.8h}, [%2]              \n" // r18
-
-                    "fmla   v28.8h, %12.8h, v17.8h      \n"
-                    "fmla   v29.8h, %12.8h, v19.8h      \n"
-                    "fmla   v30.8h, %12.8h, v21.8h      \n"
-                    "fmla   v31.8h, %12.8h, v23.8h      \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%3], #64 \n" // r20 r21 r22 r23
-
-                    "fmla   v28.8h, %13.8h, v18.8h      \n"
-                    "fmla   v29.8h, %13.8h, v20.8h      \n"
-                    "fmla   v30.8h, %13.8h, v22.8h      \n"
-                    "fmla   v31.8h, %13.8h, v24.8h      \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%3], #64 \n" // r24 r25 r26 r27
-
-                    "fmla   v28.8h, %14.8h, v0.8h       \n"
-                    "fmla   v29.8h, %14.8h, v2.8h       \n"
-                    "fmla   v30.8h, %14.8h, v4.8h       \n"
-                    "fmla   v31.8h, %14.8h, v6.8h       \n"
-
-                    "prfm   pldl1keep, [%3, #128]       \n"
-                    "ld1    {v8.8h}, [%3]               \n" // r28
-
-                    "fmla   v28.8h, %15.8h, v1.8h       \n"
-                    "fmla   v29.8h, %15.8h, v3.8h       \n"
-                    "fmla   v30.8h, %15.8h, v5.8h       \n"
-                    "fmla   v31.8h, %15.8h, v7.8h       \n"
-
-                    "fmla   v28.8h, %16.8h, v2.8h       \n"
-                    "fmla   v29.8h, %16.8h, v4.8h       \n"
-                    "fmla   v30.8h, %16.8h, v6.8h       \n"
-                    "fmla   v31.8h, %16.8h, v8.8h       \n"
-
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v28", "v29", "v30", "v31");
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%1], #64 \n" // r00 r01 r02 r03
-
-                    "mov    v28.16b, %17.16b            \n" // sum00
-                    "mov    v29.16b, %17.16b            \n" // sum01
-
-                    "fmul   v30.8h, %8.8h, v12.8h       \n"
-                    "fmul   v31.8h, %8.8h, v14.8h       \n"
-
-                    "prfm   pldl1keep, [%1, #128]       \n"
-                    "ld1    {v16.8h}, [%1]              \n" // r04
-
-                    "fmla   v28.8h, %9.8h, v13.8h       \n"
-                    "fmla   v29.8h, %9.8h, v15.8h       \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v17.8h, v18.8h, v19.8h, v20.8h}, [%2], #64 \n" // r10 r11 r12 r13
-
-                    "fmla   v30.8h, %10.8h, v14.8h      \n"
-                    "fmla   v31.8h, %10.8h, v16.8h      \n"
-
-                    "prfm   pldl1keep, [%1, #128]       \n"
-                    "ld1    {v21.8h}, [%1]              \n" // r14
-
-                    "fmla   v28.8h, %11.8h, v17.8h      \n"
-                    "fmla   v29.8h, %11.8h, v19.8h      \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v22.8h, v23.8h, v24.8h, v25.8h}, [%3], #64 \n" // r20 r21 r22 r23
-
-                    "fmla   v30.8h, %12.8h, v18.8h      \n"
-                    "fmla   v31.8h, %12.8h, v20.8h      \n"
-
-                    "fmla   v28.8h, %13.8h, v19.8h      \n"
-                    "fmla   v29.8h, %13.8h, v21.8h      \n"
-
-                    "prfm   pldl1keep, [%1, #128]       \n"
-                    "ld1    {v26.8h}, [%1]              \n" // r24
-
-                    "fmla   v30.8h, %14.8h, v22.8h      \n"
-                    "fmla   v31.8h, %14.8h, v24.8h      \n"
-
-                    "fmla   v28.8h, %15.8h, v23.8h      \n"
-                    "fmla   v29.8h, %15.8h, v25.8h      \n"
-                    "fmla   v30.8h, %16.8h, v24.8h      \n"
-                    "fmla   v31.8h, %16.8h, v26.8h      \n"
-
-                    "fadd   v28.8h, v28.8h, v30.8h      \n"
-                    "fadd   v29.8h, v29.8h, v31.8h      \n"
-
-                    "st1    {v28.8h, v29.8h}, [%0], #32 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v28", "v29", "v30", "v31");
-            }
-            for (; j < outw; j++)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #384]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h}, [%1] \n" // r00 r01 r02
-
-                    "mov    v28.16b, %17.16b            \n" // sum00
-
-                    "fmul   v29.8h, %8.8h, v12.8h       \n"
-
-                    "prfm   pldl1keep, [%2, #384]       \n"
-                    "ld1    {v15.8h, v16.8h, v17.8h}, [%2] \n" // r10 r11 r12
-
-                    "fmul   v30.8h, %9.8h, v13.8h       \n"
-                    "fmla   v28.8h, %10.8h, v14.8h      \n"
-
-                    "prfm   pldl1keep, [%3, #384]       \n"
-                    "ld1    {v18.8h, v19.8h, v20.8h}, [%3] \n" // r20 r21 r22
-
-                    "fmla   v29.8h, %11.8h, v15.8h      \n"
-                    "fmla   v30.8h, %12.8h, v16.8h      \n"
-                    "fmla   v28.8h, %13.8h, v17.8h      \n"
-
-                    "fmla   v29.8h, %14.8h, v18.8h      \n"
-                    "fmla   v30.8h, %15.8h, v19.8h      \n"
-                    "fmla   v28.8h, %16.8h, v20.8h      \n"
-
-                    "add    %1, %1, #32                 \n"
-
-                    "fadd   v29.8h, v29.8h, v30.8h      \n"
-                    "fadd   v28.8h, v28.8h, v29.8h      \n"
-
-                    "add    %2, %2, #32                 \n"
-                    "add    %3, %3, #32                 \n"
-
-                    "st1    {v28.8h}, [%0], #16         \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2)       // %3
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "w"(_k00),  // %8
-                    "w"(_k01),  // %9
-                    "w"(_k02),  // %10
-                    "w"(_k10),  // %11
-                    "w"(_k11),  // %12
-                    "w"(_k12),  // %13
-                    "w"(_k20),  // %14
-                    "w"(_k21),  // %15
-                    "w"(_k22),  // %16
-                    "w"(_bias0) // %17
-                    : "memory", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v28", "v29", "v30");
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack8_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack8_int8.h
deleted file mode 100644
index 47714a43..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_3x3_pack8_int8.h
+++ /dev/null
@@ -1,482 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_pack8_int8_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const signed char* k0 = kernel.row<const signed char>(g);
-
-        int* outptr0 = out.row<int>(0);
-        int* outptr1 = out.row<int>(1);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const signed char* r0 = img0.row<const signed char>(0);
-        const signed char* r1 = img0.row<const signed char>(1);
-        const signed char* r2 = img0.row<const signed char>(2);
-        const signed char* r3 = img0.row<const signed char>(3);
-
-        int8x8_t _k00 = vld1_s8(k0);
-        int8x8_t _k01 = vld1_s8(k0 + 8);
-        int8x8_t _k02 = vld1_s8(k0 + 16);
-        int8x8_t _k10 = vld1_s8(k0 + 24);
-        int8x8_t _k11 = vld1_s8(k0 + 32);
-        int8x8_t _k12 = vld1_s8(k0 + 40);
-        int8x8_t _k20 = vld1_s8(k0 + 48);
-        int8x8_t _k21 = vld1_s8(k0 + 56);
-        int8x8_t _k22 = vld1_s8(k0 + 64);
-
-        int i = 0;
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                int8x16_t _r0001 = vld1q_s8(r0);
-                int8x16_t _r0203 = vld1q_s8(r0 + 16);
-                int8x16_t _r1011 = vld1q_s8(r1);
-                int8x16_t _r1213 = vld1q_s8(r1 + 16);
-                int8x16_t _r2021 = vld1q_s8(r2);
-                int8x16_t _r2223 = vld1q_s8(r2 + 16);
-                int8x16_t _r3031 = vld1q_s8(r3);
-                int8x16_t _r3233 = vld1q_s8(r3 + 16);
-
-                int16x8_t _s00 = vmull_s8(vget_low_s8(_r0001), _k00);
-                int16x8_t _s01 = vmull_s8(vget_high_s8(_r0001), _k01);
-                int16x8_t _s02 = vmull_s8(vget_low_s8(_r0203), _k02);
-                int16x8_t _s03 = vmull_s8(vget_low_s8(_r1011), _k10);
-                int16x8_t _s10 = vmull_s8(vget_high_s8(_r0001), _k00);
-                int16x8_t _s11 = vmull_s8(vget_low_s8(_r0203), _k01);
-                int16x8_t _s12 = vmull_s8(vget_high_s8(_r0203), _k02);
-                int16x8_t _s13 = vmull_s8(vget_high_s8(_r1011), _k10);
-
-                int16x8_t _s20 = vmull_s8(vget_low_s8(_r1011), _k00);
-                int16x8_t _s21 = vmull_s8(vget_high_s8(_r1011), _k01);
-                int16x8_t _s22 = vmull_s8(vget_low_s8(_r1213), _k02);
-                int16x8_t _s23 = vmull_s8(vget_low_s8(_r2021), _k10);
-                int16x8_t _s30 = vmull_s8(vget_high_s8(_r1011), _k00);
-                int16x8_t _s31 = vmull_s8(vget_low_s8(_r1213), _k01);
-                int16x8_t _s32 = vmull_s8(vget_high_s8(_r1213), _k02);
-                int16x8_t _s33 = vmull_s8(vget_high_s8(_r2021), _k10);
-
-                _s00 = vmlal_s8(_s00, vget_high_s8(_r1011), _k11);
-                _s01 = vmlal_s8(_s01, vget_low_s8(_r1213), _k12);
-                _s02 = vmlal_s8(_s02, vget_low_s8(_r2021), _k20);
-                _s03 = vmlal_s8(_s03, vget_high_s8(_r2021), _k21);
-                _s10 = vmlal_s8(_s10, vget_low_s8(_r1213), _k11);
-                _s11 = vmlal_s8(_s11, vget_high_s8(_r1213), _k12);
-                _s12 = vmlal_s8(_s12, vget_high_s8(_r2021), _k20);
-                _s13 = vmlal_s8(_s13, vget_low_s8(_r2223), _k21);
-
-                _s20 = vmlal_s8(_s20, vget_high_s8(_r2021), _k11);
-                _s21 = vmlal_s8(_s21, vget_low_s8(_r2223), _k12);
-                _s22 = vmlal_s8(_s22, vget_low_s8(_r3031), _k20);
-                _s23 = vmlal_s8(_s23, vget_high_s8(_r3031), _k21);
-                _s30 = vmlal_s8(_s30, vget_low_s8(_r2223), _k11);
-                _s31 = vmlal_s8(_s31, vget_high_s8(_r2223), _k12);
-                _s32 = vmlal_s8(_s32, vget_high_s8(_r3031), _k20);
-                _s33 = vmlal_s8(_s33, vget_low_s8(_r3233), _k21);
-
-                int16x8_t _s08 = vmull_s8(vget_low_s8(_r2223), _k22);
-                int16x8_t _s18 = vmull_s8(vget_high_s8(_r2223), _k22);
-                int16x8_t _s28 = vmull_s8(vget_low_s8(_r3233), _k22);
-                int16x8_t _s38 = vmull_s8(vget_high_s8(_r3233), _k22);
-
-                int32x4_t _sum00 = vaddl_s16(vget_low_s16(_s00), vget_low_s16(_s01));
-                int32x4_t _sum01 = vaddl_s16(vget_high_s16(_s00), vget_high_s16(_s01));
-                int32x4_t _sum02 = vaddl_s16(vget_low_s16(_s02), vget_low_s16(_s03));
-                int32x4_t _sum03 = vaddl_s16(vget_high_s16(_s02), vget_high_s16(_s03));
-                int32x4_t _sum10 = vaddl_s16(vget_low_s16(_s10), vget_low_s16(_s11));
-                int32x4_t _sum11 = vaddl_s16(vget_high_s16(_s10), vget_high_s16(_s11));
-                int32x4_t _sum12 = vaddl_s16(vget_low_s16(_s12), vget_low_s16(_s13));
-                int32x4_t _sum13 = vaddl_s16(vget_high_s16(_s12), vget_high_s16(_s13));
-                int32x4_t _sum20 = vaddl_s16(vget_low_s16(_s20), vget_low_s16(_s21));
-                int32x4_t _sum21 = vaddl_s16(vget_high_s16(_s20), vget_high_s16(_s21));
-                int32x4_t _sum22 = vaddl_s16(vget_low_s16(_s22), vget_low_s16(_s23));
-                int32x4_t _sum23 = vaddl_s16(vget_high_s16(_s22), vget_high_s16(_s23));
-                int32x4_t _sum30 = vaddl_s16(vget_low_s16(_s30), vget_low_s16(_s31));
-                int32x4_t _sum31 = vaddl_s16(vget_high_s16(_s30), vget_high_s16(_s31));
-                int32x4_t _sum32 = vaddl_s16(vget_low_s16(_s32), vget_low_s16(_s33));
-                int32x4_t _sum33 = vaddl_s16(vget_high_s16(_s32), vget_high_s16(_s33));
-                _sum00 = vaddw_s16(_sum00, vget_low_s16(_s08));
-                _sum01 = vaddw_s16(_sum01, vget_high_s16(_s08));
-                _sum10 = vaddw_s16(_sum10, vget_low_s16(_s18));
-                _sum11 = vaddw_s16(_sum11, vget_high_s16(_s18));
-                _sum20 = vaddw_s16(_sum20, vget_low_s16(_s28));
-                _sum21 = vaddw_s16(_sum21, vget_high_s16(_s28));
-                _sum30 = vaddw_s16(_sum30, vget_low_s16(_s38));
-                _sum31 = vaddw_s16(_sum31, vget_high_s16(_s38));
-                _sum00 = vaddq_s32(_sum00, _sum02);
-                _sum01 = vaddq_s32(_sum01, _sum03);
-                _sum10 = vaddq_s32(_sum10, _sum12);
-                _sum11 = vaddq_s32(_sum11, _sum13);
-                _sum20 = vaddq_s32(_sum20, _sum22);
-                _sum21 = vaddq_s32(_sum21, _sum23);
-                _sum30 = vaddq_s32(_sum30, _sum32);
-                _sum31 = vaddq_s32(_sum31, _sum33);
-
-                vst1q_s32(outptr0, _sum00);
-                vst1q_s32(outptr0 + 4, _sum01);
-                vst1q_s32(outptr0 + 8, _sum10);
-                vst1q_s32(outptr0 + 12, _sum11);
-                vst1q_s32(outptr1, _sum20);
-                vst1q_s32(outptr1 + 4, _sum21);
-                vst1q_s32(outptr1 + 8, _sum30);
-                vst1q_s32(outptr1 + 12, _sum31);
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                r3 += 16;
-                outptr0 += 16;
-                outptr1 += 16;
-            }
-            for (; j < outw; j++)
-            {
-                int8x8_t _r00 = vld1_s8(r0);
-                int8x8_t _r01 = vld1_s8(r0 + 8);
-                int8x8_t _r02 = vld1_s8(r0 + 16);
-                int8x8_t _r10 = vld1_s8(r1);
-                int8x8_t _r11 = vld1_s8(r1 + 8);
-                int8x8_t _r12 = vld1_s8(r1 + 16);
-                int8x8_t _r20 = vld1_s8(r2);
-                int8x8_t _r21 = vld1_s8(r2 + 8);
-                int8x8_t _r22 = vld1_s8(r2 + 16);
-                int8x8_t _r30 = vld1_s8(r3);
-                int8x8_t _r31 = vld1_s8(r3 + 8);
-                int8x8_t _r32 = vld1_s8(r3 + 16);
-
-                int16x8_t _s00 = vmull_s8(_r00, _k00);
-                int16x8_t _s01 = vmull_s8(_r01, _k01);
-                int16x8_t _s02 = vmull_s8(_r02, _k02);
-                int16x8_t _s03 = vmull_s8(_r10, _k10);
-                int16x8_t _s10 = vmull_s8(_r10, _k00);
-                int16x8_t _s11 = vmull_s8(_r11, _k01);
-                int16x8_t _s12 = vmull_s8(_r12, _k02);
-                int16x8_t _s13 = vmull_s8(_r20, _k10);
-                _s00 = vmlal_s8(_s00, _r11, _k11);
-                _s01 = vmlal_s8(_s01, _r12, _k12);
-                _s02 = vmlal_s8(_s02, _r20, _k20);
-                _s03 = vmlal_s8(_s03, _r21, _k21);
-                _s10 = vmlal_s8(_s10, _r21, _k11);
-                _s11 = vmlal_s8(_s11, _r22, _k12);
-                _s12 = vmlal_s8(_s12, _r30, _k20);
-                _s13 = vmlal_s8(_s13, _r31, _k21);
-                int16x8_t _s08 = vmull_s8(_r22, _k22);
-                int16x8_t _s18 = vmull_s8(_r32, _k22);
-
-                int32x4_t _sum00 = vaddl_s16(vget_low_s16(_s00), vget_low_s16(_s01));
-                int32x4_t _sum01 = vaddl_s16(vget_high_s16(_s00), vget_high_s16(_s01));
-                int32x4_t _sum02 = vaddl_s16(vget_low_s16(_s02), vget_low_s16(_s03));
-                int32x4_t _sum03 = vaddl_s16(vget_high_s16(_s02), vget_high_s16(_s03));
-                int32x4_t _sum10 = vaddl_s16(vget_low_s16(_s10), vget_low_s16(_s11));
-                int32x4_t _sum11 = vaddl_s16(vget_high_s16(_s10), vget_high_s16(_s11));
-                int32x4_t _sum12 = vaddl_s16(vget_low_s16(_s12), vget_low_s16(_s13));
-                int32x4_t _sum13 = vaddl_s16(vget_high_s16(_s12), vget_high_s16(_s13));
-                _sum00 = vaddw_s16(_sum00, vget_low_s16(_s08));
-                _sum01 = vaddw_s16(_sum01, vget_high_s16(_s08));
-                _sum10 = vaddw_s16(_sum10, vget_low_s16(_s18));
-                _sum11 = vaddw_s16(_sum11, vget_high_s16(_s18));
-                _sum00 = vaddq_s32(_sum00, _sum02);
-                _sum01 = vaddq_s32(_sum01, _sum03);
-                _sum10 = vaddq_s32(_sum10, _sum12);
-                _sum11 = vaddq_s32(_sum11, _sum13);
-
-                vst1q_s32(outptr0, _sum00);
-                vst1q_s32(outptr0 + 4, _sum01);
-                vst1q_s32(outptr1, _sum10);
-                vst1q_s32(outptr1 + 4, _sum11);
-
-                r0 += 8;
-                r1 += 8;
-                r2 += 8;
-                r3 += 8;
-                outptr0 += 8;
-                outptr1 += 8;
-            }
-
-            r0 += 2 * 8 + w * 8;
-            r1 += 2 * 8 + w * 8;
-            r2 += 2 * 8 + w * 8;
-            r3 += 2 * 8 + w * 8;
-
-            outptr0 += outw * 8;
-            outptr1 += outw * 8;
-        }
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                int8x16_t _r0001 = vld1q_s8(r0);
-                int8x16_t _r0203 = vld1q_s8(r0 + 16);
-                int8x16_t _r1011 = vld1q_s8(r1);
-                int8x16_t _r1213 = vld1q_s8(r1 + 16);
-                int8x16_t _r2021 = vld1q_s8(r2);
-                int8x16_t _r2223 = vld1q_s8(r2 + 16);
-
-                int16x8_t _s00 = vmull_s8(vget_low_s8(_r0001), _k00);
-                int16x8_t _s01 = vmull_s8(vget_high_s8(_r0001), _k01);
-                int16x8_t _s02 = vmull_s8(vget_low_s8(_r0203), _k02);
-                int16x8_t _s03 = vmull_s8(vget_low_s8(_r1011), _k10);
-                int16x8_t _s10 = vmull_s8(vget_high_s8(_r0001), _k00);
-                int16x8_t _s11 = vmull_s8(vget_low_s8(_r0203), _k01);
-                int16x8_t _s12 = vmull_s8(vget_high_s8(_r0203), _k02);
-                int16x8_t _s13 = vmull_s8(vget_high_s8(_r1011), _k10);
-                _s00 = vmlal_s8(_s00, vget_high_s8(_r1011), _k11);
-                _s01 = vmlal_s8(_s01, vget_low_s8(_r1213), _k12);
-                _s02 = vmlal_s8(_s02, vget_low_s8(_r2021), _k20);
-                _s03 = vmlal_s8(_s03, vget_high_s8(_r2021), _k21);
-                _s10 = vmlal_s8(_s10, vget_low_s8(_r1213), _k11);
-                _s11 = vmlal_s8(_s11, vget_high_s8(_r1213), _k12);
-                _s12 = vmlal_s8(_s12, vget_high_s8(_r2021), _k20);
-                _s13 = vmlal_s8(_s13, vget_low_s8(_r2223), _k21);
-                int16x8_t _s08 = vmull_s8(vget_low_s8(_r2223), _k22);
-                int16x8_t _s18 = vmull_s8(vget_high_s8(_r2223), _k22);
-
-                int32x4_t _sum00 = vaddl_s16(vget_low_s16(_s00), vget_low_s16(_s01));
-                int32x4_t _sum01 = vaddl_s16(vget_high_s16(_s00), vget_high_s16(_s01));
-                int32x4_t _sum02 = vaddl_s16(vget_low_s16(_s02), vget_low_s16(_s03));
-                int32x4_t _sum03 = vaddl_s16(vget_high_s16(_s02), vget_high_s16(_s03));
-                int32x4_t _sum10 = vaddl_s16(vget_low_s16(_s10), vget_low_s16(_s11));
-                int32x4_t _sum11 = vaddl_s16(vget_high_s16(_s10), vget_high_s16(_s11));
-                int32x4_t _sum12 = vaddl_s16(vget_low_s16(_s12), vget_low_s16(_s13));
-                int32x4_t _sum13 = vaddl_s16(vget_high_s16(_s12), vget_high_s16(_s13));
-                _sum00 = vaddw_s16(_sum00, vget_low_s16(_s08));
-                _sum01 = vaddw_s16(_sum01, vget_high_s16(_s08));
-                _sum10 = vaddw_s16(_sum10, vget_low_s16(_s18));
-                _sum11 = vaddw_s16(_sum11, vget_high_s16(_s18));
-                _sum00 = vaddq_s32(_sum00, _sum02);
-                _sum01 = vaddq_s32(_sum01, _sum03);
-                _sum10 = vaddq_s32(_sum10, _sum12);
-                _sum11 = vaddq_s32(_sum11, _sum13);
-
-                vst1q_s32(outptr0, _sum00);
-                vst1q_s32(outptr0 + 4, _sum01);
-                vst1q_s32(outptr0 + 8, _sum10);
-                vst1q_s32(outptr0 + 12, _sum11);
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                outptr0 += 16;
-            }
-            for (; j < outw; j++)
-            {
-                int8x8_t _r00 = vld1_s8(r0);
-                int8x8_t _r01 = vld1_s8(r0 + 8);
-                int8x8_t _r02 = vld1_s8(r0 + 16);
-                int8x8_t _r10 = vld1_s8(r1);
-                int8x8_t _r11 = vld1_s8(r1 + 8);
-                int8x8_t _r12 = vld1_s8(r1 + 16);
-                int8x8_t _r20 = vld1_s8(r2);
-                int8x8_t _r21 = vld1_s8(r2 + 8);
-                int8x8_t _r22 = vld1_s8(r2 + 16);
-
-                int16x8_t _s0 = vmull_s8(_r00, _k00);
-                int16x8_t _s1 = vmull_s8(_r01, _k01);
-                int16x8_t _s2 = vmull_s8(_r02, _k02);
-                int16x8_t _s3 = vmull_s8(_r10, _k10);
-                _s0 = vmlal_s8(_s0, _r11, _k11);
-                _s1 = vmlal_s8(_s1, _r12, _k12);
-                _s2 = vmlal_s8(_s2, _r20, _k20);
-                _s3 = vmlal_s8(_s3, _r21, _k21);
-                int16x8_t _s4 = vmull_s8(_r22, _k22);
-
-                int32x4_t _sum0 = vaddl_s16(vget_low_s16(_s0), vget_low_s16(_s1));
-                int32x4_t _sum1 = vaddl_s16(vget_high_s16(_s0), vget_high_s16(_s1));
-                int32x4_t _sum2 = vaddl_s16(vget_low_s16(_s2), vget_low_s16(_s3));
-                int32x4_t _sum3 = vaddl_s16(vget_high_s16(_s2), vget_high_s16(_s3));
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s4));
-                _sum1 = vaddw_s16(_sum1, vget_high_s16(_s4));
-                _sum0 = vaddq_s32(_sum0, _sum2);
-                _sum1 = vaddq_s32(_sum1, _sum3);
-
-                vst1q_s32(outptr0, _sum0);
-                vst1q_s32(outptr0 + 4, _sum1);
-
-                r0 += 8;
-                r1 += 8;
-                r2 += 8;
-                outptr0 += 8;
-            }
-
-            r0 += 2 * 8;
-            r1 += 2 * 8;
-            r2 += 2 * 8;
-        }
-    }
-}
-
-static void convdw3x3s2_pack8_int8_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 8;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const signed char* k0 = kernel.row<const signed char>(g);
-
-        int* outptr0 = out;
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const signed char* r0 = img0.row<const signed char>(0);
-        const signed char* r1 = img0.row<const signed char>(1);
-        const signed char* r2 = img0.row<const signed char>(2);
-
-        int8x8_t _k00 = vld1_s8(k0);
-        int8x8_t _k01 = vld1_s8(k0 + 8);
-        int8x8_t _k02 = vld1_s8(k0 + 16);
-        int8x8_t _k10 = vld1_s8(k0 + 24);
-        int8x8_t _k11 = vld1_s8(k0 + 32);
-        int8x8_t _k12 = vld1_s8(k0 + 40);
-        int8x8_t _k20 = vld1_s8(k0 + 48);
-        int8x8_t _k21 = vld1_s8(k0 + 56);
-        int8x8_t _k22 = vld1_s8(k0 + 64);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                int8x8_t _r00 = vld1_s8(r0);
-                int8x8_t _r01 = vld1_s8(r0 + 8);
-                int8x8_t _r02 = vld1_s8(r0 + 16);
-                int8x8_t _r03 = vld1_s8(r0 + 24);
-                int8x8_t _r04 = vld1_s8(r0 + 32);
-                int8x8_t _r10 = vld1_s8(r1);
-                int8x8_t _r11 = vld1_s8(r1 + 8);
-                int8x8_t _r12 = vld1_s8(r1 + 16);
-                int8x8_t _r13 = vld1_s8(r1 + 24);
-                int8x8_t _r14 = vld1_s8(r1 + 32);
-                int8x8_t _r20 = vld1_s8(r2);
-                int8x8_t _r21 = vld1_s8(r2 + 8);
-                int8x8_t _r22 = vld1_s8(r2 + 16);
-                int8x8_t _r23 = vld1_s8(r2 + 24);
-                int8x8_t _r24 = vld1_s8(r2 + 32);
-
-                int16x8_t _s00 = vmull_s8(_r00, _k00);
-                int16x8_t _s01 = vmull_s8(_r01, _k01);
-                int16x8_t _s02 = vmull_s8(_r02, _k02);
-                int16x8_t _s03 = vmull_s8(_r10, _k10);
-                int16x8_t _s10 = vmull_s8(_r02, _k00);
-                int16x8_t _s11 = vmull_s8(_r03, _k01);
-                int16x8_t _s12 = vmull_s8(_r04, _k02);
-                int16x8_t _s13 = vmull_s8(_r12, _k10);
-                _s00 = vmlal_s8(_s00, _r11, _k11);
-                _s01 = vmlal_s8(_s01, _r12, _k12);
-                _s02 = vmlal_s8(_s02, _r20, _k20);
-                _s03 = vmlal_s8(_s03, _r21, _k21);
-                _s10 = vmlal_s8(_s10, _r13, _k11);
-                _s11 = vmlal_s8(_s11, _r14, _k12);
-                _s12 = vmlal_s8(_s12, _r22, _k20);
-                _s13 = vmlal_s8(_s13, _r23, _k21);
-                int16x8_t _s08 = vmull_s8(_r22, _k22);
-                int16x8_t _s18 = vmull_s8(_r24, _k22);
-
-                int32x4_t _sum00 = vaddl_s16(vget_low_s16(_s00), vget_low_s16(_s01));
-                int32x4_t _sum01 = vaddl_s16(vget_high_s16(_s00), vget_high_s16(_s01));
-                int32x4_t _sum02 = vaddl_s16(vget_low_s16(_s02), vget_low_s16(_s03));
-                int32x4_t _sum03 = vaddl_s16(vget_high_s16(_s02), vget_high_s16(_s03));
-                int32x4_t _sum10 = vaddl_s16(vget_low_s16(_s10), vget_low_s16(_s11));
-                int32x4_t _sum11 = vaddl_s16(vget_high_s16(_s10), vget_high_s16(_s11));
-                int32x4_t _sum12 = vaddl_s16(vget_low_s16(_s12), vget_low_s16(_s13));
-                int32x4_t _sum13 = vaddl_s16(vget_high_s16(_s12), vget_high_s16(_s13));
-                _sum00 = vaddw_s16(_sum00, vget_low_s16(_s08));
-                _sum01 = vaddw_s16(_sum01, vget_high_s16(_s08));
-                _sum10 = vaddw_s16(_sum10, vget_low_s16(_s18));
-                _sum11 = vaddw_s16(_sum11, vget_high_s16(_s18));
-                _sum00 = vaddq_s32(_sum00, _sum02);
-                _sum01 = vaddq_s32(_sum01, _sum03);
-                _sum10 = vaddq_s32(_sum10, _sum12);
-                _sum11 = vaddq_s32(_sum11, _sum13);
-
-                vst1q_s32(outptr0, _sum00);
-                vst1q_s32(outptr0 + 4, _sum01);
-                vst1q_s32(outptr0 + 8, _sum10);
-                vst1q_s32(outptr0 + 12, _sum11);
-
-                r0 += 32;
-                r1 += 32;
-                r2 += 32;
-                outptr0 += 16;
-            }
-            for (; j < outw; j++)
-            {
-                int8x8_t _r00 = vld1_s8(r0);
-                int8x8_t _r01 = vld1_s8(r0 + 8);
-                int8x8_t _r02 = vld1_s8(r0 + 16);
-                int8x8_t _r10 = vld1_s8(r1);
-                int8x8_t _r11 = vld1_s8(r1 + 8);
-                int8x8_t _r12 = vld1_s8(r1 + 16);
-                int8x8_t _r20 = vld1_s8(r2);
-                int8x8_t _r21 = vld1_s8(r2 + 8);
-                int8x8_t _r22 = vld1_s8(r2 + 16);
-
-                int16x8_t _s0 = vmull_s8(_r00, _k00);
-                int16x8_t _s1 = vmull_s8(_r01, _k01);
-                int16x8_t _s2 = vmull_s8(_r02, _k02);
-                int16x8_t _s3 = vmull_s8(_r10, _k10);
-                _s0 = vmlal_s8(_s0, _r11, _k11);
-                _s1 = vmlal_s8(_s1, _r12, _k12);
-                _s2 = vmlal_s8(_s2, _r20, _k20);
-                _s3 = vmlal_s8(_s3, _r21, _k21);
-                int16x8_t _s4 = vmull_s8(_r22, _k22);
-
-                int32x4_t _sum0 = vaddl_s16(vget_low_s16(_s0), vget_low_s16(_s1));
-                int32x4_t _sum1 = vaddl_s16(vget_high_s16(_s0), vget_high_s16(_s1));
-                int32x4_t _sum2 = vaddl_s16(vget_low_s16(_s2), vget_low_s16(_s3));
-                int32x4_t _sum3 = vaddl_s16(vget_high_s16(_s2), vget_high_s16(_s3));
-                _sum0 = vaddw_s16(_sum0, vget_low_s16(_s4));
-                _sum1 = vaddw_s16(_sum1, vget_high_s16(_s4));
-                _sum0 = vaddq_s32(_sum0, _sum2);
-                _sum1 = vaddq_s32(_sum1, _sum3);
-
-                vst1q_s32(outptr0, _sum0);
-                vst1q_s32(outptr0 + 4, _sum1);
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                outptr0 += 8;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5.h
deleted file mode 100644
index ef9b1e78..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5.h
+++ /dev/null
@@ -1,2250 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2018 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw5x5s1_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const float bias0 = bias ? bias[g] : 0.f;
-
-        const float* kernel0 = kernel + g * 25;
-
-        float* outptr = out;
-        float* outptr2 = outptr + outw;
-
-        const float* img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0;
-        const float* r1 = img0 + w;
-        const float* r2 = img0 + w * 2;
-        const float* r3 = img0 + w * 3;
-        const float* r4 = img0 + w * 4;
-        const float* r5 = img0 + w * 5;
-
-        const float* k0 = kernel0;
-        const float* k1 = kernel0 + 5;
-        const float* k2 = kernel0 + 10;
-        const float* k3 = kernel0 + 15;
-        const float* k4 = kernel0 + 20;
-
-#if __ARM_NEON
-        float32x4_t _k0123 = vld1q_f32(kernel0);
-        float32x4_t _k4567 = vld1q_f32(kernel0 + 4);
-        float32x4_t _k891011 = vld1q_f32(kernel0 + 8);
-        float32x4_t _k12131415 = vld1q_f32(kernel0 + 12);
-        float32x4_t _k16171819 = vld1q_f32(kernel0 + 16);
-        float32x4_t _k20212223 = vld1q_f32(kernel0 + 20);
-        float32x4_t _k24242424 = vdupq_n_f32(kernel0[24]);
-
-        float32x4_t _bias0 = vdupq_n_f32(bias0);
-#endif // __ARM_NEON
-
-        int i = 0;
-
-        for (; i + 1 < outh; i += 2)
-        {
-#if __ARM_NEON
-#if __aarch64__
-            int nn = outw >> 3;
-            int remain = outw & 7;
-#else
-            int nn = outw >> 2;
-            int remain = outw & 3;
-#endif // __aarch64__
-#else
-            int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    // r1
-                    "prfm   pldl1keep, [%4, #384]           \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s}, [%4]  \n" // v16 v17 v18 = r10 r14 r18
-
-                    "mov    v8.16b, %25.16b                 \n" // v8 = _bias0
-                    "mov    v9.16b, %25.16b                 \n" // v9 = _bias0
-
-                    "0:                                     \n"
-
-                    "mov    v10.16b, %25.16b                \n" // v10 = _bias0
-                    "mov    v11.16b, %25.16b                \n" // v11 = _bias0
-
-                    "fmla   v8.4s, v16.4s, %19.s[1]         \n"
-                    "fmla   v10.4s, v16.4s, %18.s[0]        \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #4   \n" // r11
-
-                    "fmla   v9.4s, v17.4s, %19.s[1]         \n"
-                    "fmla   v11.4s, v17.4s, %18.s[0]        \n"
-
-                    "ext    v20.16b, v17.16b, v18.16b, #4   \n" // r15
-
-                    "fmla   v8.4s, v17.4s, %20.s[1]         \n"
-                    "fmla   v10.4s, v17.4s, %19.s[0]        \n"
-
-                    "ext    v21.16b, v16.16b, v17.16b, #8   \n" // r12
-
-                    "fmla   v9.4s, v18.4s, %20.s[1]         \n"
-                    "fmla   v11.4s, v18.4s, %19.s[0]        \n"
-
-                    "ext    v22.16b, v17.16b, v18.16b, #8   \n" // r16
-
-                    "fmla   v8.4s, v19.4s, %19.s[2]         \n"
-                    "fmla   v10.4s, v19.4s, %18.s[1]        \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #12  \n" // r13
-
-                    "fmla   v9.4s, v20.4s, %19.s[2]         \n"
-                    "fmla   v11.4s, v20.4s, %18.s[1]        \n"
-
-                    "ext    v20.16b, v17.16b, v18.16b, #12  \n" // r17
-
-                    "fmla   v8.4s, v21.4s, %19.s[3]         \n"
-                    "fmla   v10.4s, v21.4s, %18.s[2]        \n"
-
-                    "add    %4, %4, #32                     \n"
-
-                    "fmla   v9.4s, v22.4s, %19.s[3]         \n"
-                    "fmla   v11.4s, v22.4s, %18.s[2]        \n"
-
-                    // r2
-                    "prfm   pldl1keep, [%5, #384]           \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s}, [%5]  \n" // v12 v13 v14 = r20 r24 r28
-
-                    "fmla   v8.4s, v19.4s, %20.s[0]         \n"
-                    "fmla   v10.4s, v19.4s, %18.s[3]        \n"
-                    "fmla   v9.4s, v20.4s, %20.s[0]         \n"
-                    "fmla   v11.4s, v20.4s, %18.s[3]        \n"
-
-                    "add    %5, %5, #32                     \n"
-
-                    "fmla   v8.4s, v12.4s, %20.s[2]         \n"
-                    "fmla   v10.4s, v12.4s, %19.s[1]        \n"
-
-                    "ext    v21.16b, v12.16b, v13.16b, #4   \n" // r21
-
-                    "fmla   v9.4s, v13.4s, %20.s[2]         \n"
-                    "fmla   v11.4s, v13.4s, %19.s[1]        \n"
-
-                    "ext    v22.16b, v13.16b, v14.16b, #4   \n" // r25
-
-                    "fmla   v8.4s, v13.4s, %21.s[2]         \n"
-                    "fmla   v10.4s, v13.4s, %20.s[1]        \n"
-
-                    "ext    v19.16b, v12.16b, v13.16b, #8   \n" // r22
-
-                    "fmla   v9.4s, v14.4s, %21.s[2]         \n"
-                    "fmla   v11.4s, v14.4s, %20.s[1]        \n"
-
-                    "ext    v20.16b, v13.16b, v14.16b, #8   \n" // r26
-
-                    "fmla   v8.4s, v21.4s, %20.s[3]         \n"
-                    "fmla   v10.4s, v21.4s, %19.s[2]        \n"
-
-                    "ext    v21.16b, v12.16b, v13.16b, #12  \n" // r23
-
-                    "fmla   v9.4s, v22.4s, %20.s[3]         \n"
-                    "fmla   v11.4s, v22.4s, %19.s[2]        \n"
-
-                    "ext    v22.16b, v13.16b, v14.16b, #12  \n" // r27
-
-                    "fmla   v8.4s, v19.4s, %21.s[0]         \n"
-                    "fmla   v10.4s, v19.4s, %19.s[3]        \n"
-                    "fmla   v9.4s, v20.4s, %21.s[0]         \n"
-                    "fmla   v11.4s, v20.4s, %19.s[3]        \n"
-
-                    // r3
-                    "prfm   pldl1keep, [%6, #384]           \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s}, [%6]  \n" // v16 v17 v18 = r30 r34 r38
-
-                    "fmla   v8.4s, v21.4s, %21.s[1]         \n"
-                    "fmla   v10.4s, v21.4s, %20.s[0]        \n"
-                    "fmla   v9.4s, v22.4s, %21.s[1]         \n"
-                    "fmla   v11.4s, v22.4s, %20.s[0]        \n"
-
-                    "add    %6, %6, #32                     \n"
-
-                    "fmla   v8.4s, v16.4s, %21.s[3]         \n"
-                    "fmla   v10.4s, v16.4s, %20.s[2]        \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #4   \n" // r31
-
-                    "fmla   v9.4s, v17.4s, %21.s[3]         \n"
-                    "fmla   v11.4s, v17.4s, %20.s[2]        \n"
-
-                    "ext    v20.16b, v17.16b, v18.16b, #4   \n" // r35
-
-                    "fmla   v8.4s, v17.4s, %22.s[3]         \n"
-                    "fmla   v10.4s, v17.4s, %21.s[2]        \n"
-
-                    "ext    v21.16b, v16.16b, v17.16b, #8   \n" // r32
-
-                    "fmla   v9.4s, v18.4s, %22.s[3]         \n"
-                    "fmla   v11.4s, v18.4s, %21.s[2]        \n"
-
-                    "ext    v22.16b, v17.16b, v18.16b, #8   \n" // r36
-
-                    "fmla   v8.4s, v19.4s, %22.s[0]         \n"
-                    "fmla   v10.4s, v19.4s, %20.s[3]        \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #12  \n" // r33
-
-                    "fmla   v9.4s, v20.4s, %22.s[0]         \n"
-                    "fmla   v11.4s, v20.4s, %20.s[3]        \n"
-
-                    "ext    v20.16b, v17.16b, v18.16b, #12  \n" // r37
-
-                    "fmla   v8.4s, v21.4s, %22.s[1]         \n"
-                    "fmla   v10.4s, v21.4s, %21.s[0]        \n"
-                    "fmla   v9.4s, v22.4s, %22.s[1]         \n"
-                    "fmla   v11.4s, v22.4s, %21.s[0]        \n"
-
-                    // r4
-                    "prfm   pldl1keep, [%7, #384]           \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s}, [%7]  \n" // v12 v13 v14 = r40 r44 r48
-
-                    "fmla   v8.4s, v19.4s, %22.s[2]         \n"
-                    "fmla   v10.4s, v19.4s, %21.s[1]        \n"
-
-                    "add    %7, %7, #32                     \n"
-
-                    "fmla   v9.4s, v20.4s, %22.s[2]         \n"
-                    "fmla   v11.4s, v20.4s, %21.s[1]        \n"
-
-                    "ext    v21.16b, v12.16b, v13.16b, #4   \n" // r41
-
-                    "fmla   v8.4s, v12.4s, %23.s[0]         \n"
-                    "fmla   v10.4s, v12.4s, %21.s[3]        \n"
-
-                    "ext    v22.16b, v13.16b, v14.16b, #4   \n" // r45
-
-                    "fmla   v9.4s, v13.4s, %23.s[0]         \n"
-                    "fmla   v11.4s, v13.4s, %21.s[3]        \n"
-
-                    "ext    v19.16b, v12.16b, v13.16b, #8   \n" // r42
-
-                    "fmla   v8.4s, v13.4s, %24.s[0]         \n"
-                    "fmla   v10.4s, v13.4s, %22.s[3]        \n"
-
-                    "ext    v20.16b, v13.16b, v14.16b, #8   \n" // r46
-
-                    "fmla   v9.4s, v14.4s, %24.s[0]         \n"
-                    "fmla   v11.4s, v14.4s, %22.s[3]        \n"
-
-                    // r0 and r5
-                    "prfm   pldl1keep, [%3, #384]           \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s}, [%3]  \n" // v16 v17 v18 = r00 r04 r08
-
-                    "fmla   v8.4s, v21.4s, %23.s[1]         \n"
-                    "fmla   v10.4s, v21.4s, %22.s[0]        \n"
-
-                    "ext    v21.16b, v12.16b, v13.16b, #12  \n" // r43
-
-                    "fmla   v9.4s, v22.4s, %23.s[1]         \n"
-                    "fmla   v11.4s, v22.4s, %22.s[0]        \n"
-
-                    "ext    v22.16b, v13.16b, v14.16b, #12  \n" // r47
-
-                    "fmla   v8.4s, v19.4s, %23.s[2]         \n"
-                    "fmla   v10.4s, v19.4s, %22.s[1]        \n"
-
-                    "prfm   pldl1keep, [%8, #384]           \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s}, [%8]  \n" // v12 v13 v14 = r50 r54 r58
-
-                    "fmla   v9.4s, v20.4s, %23.s[2]         \n"
-                    "fmla   v11.4s, v20.4s, %22.s[1]        \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #4   \n" // r01
-
-                    "fmla   v8.4s, v21.4s, %23.s[3]         \n"
-                    "fmla   v10.4s, v21.4s, %22.s[2]        \n"
-
-                    "ext    v23.16b, v12.16b, v13.16b, #4   \n" // r51
-
-                    "fmla   v9.4s, v22.4s, %23.s[3]         \n"
-                    "fmla   v11.4s, v22.4s, %22.s[2]        \n"
-
-                    "ext    v20.16b, v17.16b, v18.16b, #4   \n" // r05
-
-                    "fmla   v8.4s, v16.4s, %18.s[0]         \n"
-                    "fmla   v10.4s, v12.4s, %23.s[0]        \n"
-
-                    "ext    v24.16b, v13.16b, v14.16b, #4   \n" // r55
-
-                    "fmla   v9.4s, v17.4s, %18.s[0]         \n"
-                    "fmla   v11.4s, v13.4s, %23.s[0]        \n"
-
-                    "ext    v21.16b, v16.16b, v17.16b, #8   \n" // r02
-
-                    "fmla   v8.4s, v17.4s, %19.s[0]         \n"
-                    "fmla   v10.4s, v13.4s, %24.s[0]        \n"
-
-                    "ext    v25.16b, v12.16b, v13.16b, #8   \n" // r52
-
-                    "fmla   v9.4s, v18.4s, %19.s[0]         \n"
-                    "fmla   v11.4s, v14.4s, %24.s[0]        \n"
-
-                    "ext    v22.16b, v17.16b, v18.16b, #8   \n" // r06
-
-                    "fmla   v8.4s, v19.4s, %18.s[1]         \n"
-                    "fmla   v10.4s, v23.4s, %23.s[1]        \n"
-
-                    "ext    v26.16b, v13.16b, v14.16b, #8   \n" // r56
-
-                    "fmla   v9.4s, v20.4s, %18.s[1]         \n"
-                    "fmla   v11.4s, v24.4s, %23.s[1]        \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #12  \n" // r03
-
-                    "fmla   v8.4s, v21.4s, %18.s[2]         \n"
-                    "fmla   v10.4s, v25.4s, %23.s[2]        \n"
-
-                    "ext    v23.16b, v12.16b, v13.16b, #12  \n" // r53
-
-                    "fmla   v9.4s, v22.4s, %18.s[2]         \n"
-                    "fmla   v11.4s, v26.4s, %23.s[2]        \n"
-
-                    "ext    v20.16b, v17.16b, v18.16b, #12  \n" // r07
-
-                    "fmla   v8.4s, v19.4s, %18.s[3]         \n"
-                    "fmla   v10.4s, v23.4s, %23.s[3]        \n"
-
-                    "ext    v24.16b, v13.16b, v14.16b, #12  \n" // r57
-
-                    "fmla   v9.4s, v20.4s, %18.s[3]         \n"
-
-                    "add    %3, %3, #32                     \n"
-
-                    "fmla   v11.4s, v24.4s, %23.s[3]        \n"
-
-                    "add    %8, %8, #32                     \n"
-
-                    // r1
-                    "prfm   pldl1keep, [%4, #384]           \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s}, [%4]  \n" // v16 v17 v18 = r10 r14 r18
-
-                    "subs   %w0, %w0, #1                    \n"
-
-                    "st1    {v8.4s, v9.4s}, [%1], #32       \n"
-
-                    "mov    v8.16b, %25.16b                 \n" // v8 = _bias0
-                    "mov    v9.16b, %25.16b                 \n" // v9 = _bias0
-
-                    "st1    {v10.4s, v11.4s}, [%2], #32     \n"
-
-                    "bne    0b                              \n"
-                    : "=r"(nn),      // %0
-                    "=r"(outptr),  // %1
-                    "=r"(outptr2), // %2
-                    "=r"(r0),      // %3
-                    "=r"(r1),      // %4
-                    "=r"(r2),      // %5
-                    "=r"(r3),      // %6
-                    "=r"(r4),      // %7
-                    "=r"(r5)       // %8
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(outptr2),
-                    "3"(r0),
-                    "4"(r1),
-                    "5"(r2),
-                    "6"(r3),
-                    "7"(r4),
-                    "8"(r5),
-                    "w"(_k0123),     // %18
-                    "w"(_k4567),     // %19
-                    "w"(_k891011),   // %20
-                    "w"(_k12131415), // %21
-                    "w"(_k16171819), // %22
-                    "w"(_k20212223), // %23
-                    "w"(_k24242424), // %24
-                    "w"(_bias0)      // %25
-                    : "cc", "memory", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26");
-            }
-
-            if (remain >= 4)
-            {
-                remain -= 4;
-                asm volatile(
-                    // r1
-                    "prfm   pldl1keep, [%3, #256]           \n"
-                    "ld1    {v12.4s, v13.4s}, [%3]          \n" // v12 v13 = r10 r14
-
-                    "mov    v8.16b, %23.16b                 \n" // v8 = _bias0
-                    "mov    v9.16b, %23.16b                 \n" // v9 = _bias0
-
-                    "fmul   v10.4s, v12.4s, %17.s[1]        \n"
-                    "fmul   v11.4s, v12.4s, %16.s[0]        \n"
-
-                    "ext    v21.16b, v12.16b, v13.16b, #4   \n" // r11
-
-                    "fmla   v8.4s, v13.4s, %18.s[1]         \n"
-                    "fmla   v9.4s, v13.4s, %17.s[0]         \n"
-
-                    "ext    v22.16b, v12.16b, v13.16b, #8   \n" // r12
-
-                    "fmla   v10.4s, v21.4s, %17.s[2]        \n"
-                    "fmla   v11.4s, v21.4s, %16.s[1]        \n"
-
-                    "ext    v23.16b, v12.16b, v13.16b, #12  \n" // r13
-
-                    "fmla   v8.4s, v22.4s, %17.s[3]         \n"
-                    "fmla   v9.4s, v22.4s, %16.s[2]         \n"
-
-                    // r2
-                    "prfm   pldl1keep, [%4, #256]           \n"
-                    "ld1    {v16.4s, v17.4s}, [%4]          \n" // v16 v17 = r20 r24
-
-                    "fmla   v10.4s, v23.4s, %18.s[0]        \n"
-                    "fmla   v11.4s, v23.4s, %16.s[3]        \n"
-
-                    "add    %4, %4, #16                     \n"
-
-                    "fmla   v8.4s, v16.4s, %18.s[2]         \n"
-                    "fmla   v9.4s, v16.4s, %17.s[1]         \n"
-
-                    "ext    v18.16b, v16.16b, v17.16b, #4   \n" // r21
-
-                    "fmla   v10.4s, v17.4s, %19.s[2]        \n"
-                    "fmla   v11.4s, v17.4s, %18.s[1]        \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #8   \n" // r22
-
-                    "fmla   v8.4s, v18.4s, %18.s[3]         \n"
-                    "fmla   v9.4s, v18.4s, %17.s[2]         \n"
-
-                    "ext    v20.16b, v16.16b, v17.16b, #12  \n" // r23
-
-                    "fmla   v10.4s, v19.4s, %19.s[0]        \n"
-                    "fmla   v11.4s, v19.4s, %17.s[3]        \n"
-
-                    // r3
-                    "prfm   pldl1keep, [%5, #256]           \n"
-                    "ld1    {v12.4s, v13.4s}, [%5]          \n" // v12 v13 = r30 r34
-
-                    "fmla   v8.4s, v20.4s, %19.s[1]         \n"
-                    "fmla   v9.4s, v20.4s, %18.s[0]         \n"
-
-                    "add    %5, %5, #16                     \n"
-
-                    "fmla   v10.4s, v12.4s, %19.s[3]        \n"
-                    "fmla   v11.4s, v12.4s, %18.s[2]        \n"
-
-                    "ext    v21.16b, v12.16b, v13.16b, #4   \n" // r31
-
-                    "fmla   v8.4s, v13.4s, %20.s[3]         \n"
-                    "fmla   v9.4s, v13.4s, %19.s[2]         \n"
-
-                    "ext    v22.16b, v12.16b, v13.16b, #8   \n" // r32
-
-                    "fmla   v10.4s, v21.4s, %20.s[0]        \n"
-                    "fmla   v11.4s, v21.4s, %18.s[3]        \n"
-
-                    "ext    v23.16b, v12.16b, v13.16b, #12  \n" // r33
-
-                    "fmla   v8.4s, v22.4s, %20.s[1]         \n"
-                    "fmla   v9.4s, v22.4s, %19.s[0]         \n"
-
-                    // r4
-                    "prfm   pldl1keep, [%6, #256]           \n"
-                    "ld1    {v16.4s, v17.4s}, [%6]          \n" // v16 v17 = r40 r44
-
-                    "fmla   v10.4s, v23.4s, %20.s[2]        \n"
-                    "fmla   v11.4s, v23.4s, %19.s[1]        \n"
-
-                    "add    %6, %6, #16                     \n"
-
-                    "fmla   v8.4s, v16.4s, %21.s[0]         \n"
-                    "fmla   v9.4s, v16.4s, %19.s[3]         \n"
-
-                    "ext    v18.16b, v16.16b, v17.16b, #4   \n" // r41
-
-                    "fmla   v10.4s, v17.4s, %22.s[0]        \n"
-                    "fmla   v11.4s, v17.4s, %20.s[3]        \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #8   \n" // r42
-
-                    "fmla   v8.4s, v18.4s, %21.s[1]         \n"
-                    "fmla   v9.4s, v18.4s, %20.s[0]         \n"
-
-                    "ext    v20.16b, v16.16b, v17.16b, #12  \n" // r43
-
-                    "fmla   v10.4s, v19.4s, %21.s[2]        \n"
-                    "fmla   v11.4s, v19.4s, %20.s[1]        \n"
-
-                    // r0
-                    "prfm   pldl1keep, [%2, #256]           \n"
-                    "ld1    {v16.4s, v17.4s}, [%2]          \n" // v16 v17 = r00 r04
-
-                    "fmla   v8.4s, v20.4s, %21.s[3]         \n"
-                    "fmla   v9.4s, v20.4s, %20.s[2]         \n"
-
-                    // r5
-                    "prfm   pldl1keep, [%7, #256]           \n"
-                    "ld1    {v12.4s, v13.4s}, [%7]          \n" // v12 v13 = r50 r54
-
-                    "fmla   v10.4s, v16.4s, %16.s[0]        \n"
-                    "fmla   v11.4s, v12.4s, %21.s[0]        \n"
-
-                    "ext    v18.16b, v16.16b, v17.16b, #4   \n" // r01
-
-                    "fmla   v8.4s, v17.4s, %17.s[0]         \n"
-
-                    "ext    v21.16b, v12.16b, v13.16b, #4   \n" // r51
-
-                    "fmla   v9.4s, v13.4s, %22.s[0]         \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #8   \n" // r02
-
-                    "fmla   v10.4s, v18.4s, %16.s[1]        \n"
-
-                    "ext    v22.16b, v12.16b, v13.16b, #8   \n" // r52
-
-                    "fmla   v11.4s, v21.4s, %21.s[1]        \n"
-
-                    "ext    v20.16b, v16.16b, v17.16b, #12  \n" // r03
-
-                    "fmla   v8.4s, v19.4s, %16.s[2]         \n"
-
-                    "ext    v23.16b, v12.16b, v13.16b, #12  \n" // r53
-
-                    "fmla   v9.4s, v22.4s, %21.s[2]         \n"
-
-                    "add    %3, %3, #16                     \n"
-
-                    "fmla   v10.4s, v20.4s, %16.s[3]        \n"
-                    "fmla   v11.4s, v23.4s, %21.s[3]        \n"
-
-                    "add    %2, %2, #16                     \n"
-
-                    "fadd   v8.4s, v8.4s, v10.4s            \n"
-                    "fadd   v9.4s, v9.4s, v11.4s            \n"
-
-                    "add    %7, %7, #16                     \n"
-
-                    "st1    {v8.4s}, [%0], #16              \n"
-                    "st1    {v9.4s}, [%1], #16              \n"
-
-                    : "=r"(outptr),  // %0
-                    "=r"(outptr2), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3),      // %5
-                    "=r"(r4),      // %6
-                    "=r"(r5)       // %7
-                    : "0"(outptr),
-                    "1"(outptr2),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "7"(r5),
-                    "w"(_k0123),     // %16
-                    "w"(_k4567),     // %17
-                    "w"(_k891011),   // %18
-                    "w"(_k12131415), // %19
-                    "w"(_k16171819), // %20
-                    "w"(_k20212223), // %21
-                    "w"(_k24242424), // %22
-                    "w"(_bias0)      // %23
-                    : "cc", "memory", "v8", "v9", "v10", "v11", "v12", "v13", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    // r1
-                    "pld        [%4, #256]          \n"
-                    "vld1.f32   {d28-d31}, [%4]     \n" // q14 q15 = r10 r14
-
-                    "vmov       q8, %q25            \n" // q8 = _bias0
-
-                    "0:                             \n"
-
-                    "vmov       q9, %q25            \n" // q9 = _bias0
-
-                    "vmla.f32   q8, q14, %e19[1]    \n"
-                    "vmla.f32   q9, q14, %e18[0]    \n"
-
-                    "vext.32    q12, q14, q15, #1   \n" // r11
-
-                    "vmla.f32   q8, q15, %e20[1]    \n"
-                    "vmla.f32   q9, q15, %e19[0]    \n"
-
-                    "vext.32    q13, q14, q15, #2   \n" // r12
-
-                    "vmla.f32   q8, q12, %f19[0]    \n"
-                    "vmla.f32   q9, q12, %e18[1]    \n"
-
-                    "vext.32    q12, q14, q15, #3   \n" // r13
-
-                    "vmla.f32   q8, q13, %f19[1]    \n"
-                    "vmla.f32   q9, q13, %f18[0]    \n"
-
-                    // r2
-                    "pld        [%5, #256]          \n"
-                    "vld1.f32   {d20-d23}, [%5]     \n" // q10 q11 = r20 r24
-
-                    "vmla.f32   q8, q12, %e20[0]    \n"
-                    "vmla.f32   q9, q12, %f18[1]    \n"
-
-                    "add        %5, #16             \n"
-
-                    "vmla.f32   q8, q10, %f20[0]    \n"
-                    "vmla.f32   q9, q10, %e19[1]    \n"
-
-                    "vext.32    q12, q10, q11, #1   \n" // r21
-
-                    "vmla.f32   q8, q11, %f21[0]    \n"
-                    "vmla.f32   q9, q11, %e20[1]    \n"
-
-                    "vext.32    q13, q10, q11, #2   \n" // r22
-
-                    "vmla.f32   q8, q12, %f20[1]    \n"
-                    "vmla.f32   q9, q12, %f19[0]    \n"
-
-                    "vext.32    q12, q10, q11, #3   \n" // r23
-
-                    "vmla.f32   q8, q13, %e21[0]    \n"
-                    "vmla.f32   q9, q13, %f19[1]    \n"
-
-                    // r3
-                    "pld        [%6, #256]          \n"
-                    "vld1.f32   {d28-d31}, [%6]     \n" // q14 q15 = r30 r34
-
-                    "vmla.f32   q8, q12, %e21[1]    \n"
-                    "vmla.f32   q9, q12, %e20[0]    \n"
-
-                    "add        %6, #16             \n"
-
-                    "vmla.f32   q8, q14, %f21[1]    \n"
-                    "vmla.f32   q9, q14, %f20[0]    \n"
-
-                    "vext.32    q12, q14, q15, #1   \n" // r31
-
-                    "vmla.f32   q8, q15, %f22[1]    \n"
-                    "vmla.f32   q9, q15, %f21[0]    \n"
-
-                    "vext.32    q13, q14, q15, #2   \n" // r32
-
-                    "vmla.f32   q8, q12, %e22[0]    \n"
-                    "vmla.f32   q9, q12, %f20[1]    \n"
-
-                    "vext.32    q12, q14, q15, #3   \n" // r33
-
-                    "vmla.f32   q8, q13, %e22[1]    \n"
-                    "vmla.f32   q9, q13, %e21[0]    \n"
-
-                    // r4
-                    "pld        [%7, #256]          \n"
-                    "vld1.f32   {d20-d23}, [%7]     \n" // q10 q11 = r40 r44
-
-                    "vmla.f32   q8, q12, %f22[0]    \n"
-                    "vmla.f32   q9, q12, %e21[1]    \n"
-
-                    "add        %7, #16             \n"
-
-                    "vmla.f32   q8, q10, %e23[0]    \n"
-                    "vmla.f32   q9, q10, %f21[1]    \n"
-
-                    "vext.32    q12, q10, q11, #1   \n" // r41
-
-                    "vmla.f32   q8, q11, %e24[0]    \n"
-                    "vmla.f32   q9, q11, %f22[1]    \n"
-
-                    "vext.32    q13, q10, q11, #2   \n" // r42
-
-                    "vmla.f32   q8, q12, %e23[1]    \n"
-                    "vmla.f32   q9, q12, %e22[0]    \n"
-
-                    "vext.32    q12, q10, q11, #3   \n" // r43
-
-                    "vmla.f32   q8, q13, %f23[0]    \n"
-                    "vmla.f32   q9, q13, %e22[1]    \n"
-
-                    // r0 and r5
-                    "pld        [%3, #256]          \n"
-                    "vld1.f32   {d20-d23}, [%3]     \n" // q10 q11 = r00 r04
-
-                    "vmla.f32   q8, q12, %f23[1]    \n"
-                    "vmla.f32   q9, q12, %f22[0]    \n"
-
-                    // r5
-                    "pld        [%8, #256]          \n"
-                    "vld1.f32   {d28-d31}, [%8]     \n" // q14 q15 = r50 r54
-
-                    "vmla.f32   q8, q10, %e18[0]    \n"
-                    "vmla.f32   q9, q14, %e23[0]    \n"
-
-                    "vext.32    q12, q10, q11, #1   \n" // r01
-
-                    "vmla.f32   q8, q11, %e19[0]    \n"
-                    "vmla.f32   q9, q15, %e24[0]    \n"
-
-                    "vext.32    q13, q14, q15, #1   \n" // r51
-
-                    "vmla.f32   q8, q12, %e18[1]    \n"
-
-                    "vext.32    q12, q10, q11, #2   \n" // r02
-
-                    "vmla.f32   q9, q13, %e23[1]    \n"
-
-                    "vext.32    q13, q14, q15, #2   \n" // r52
-
-                    "vmla.f32   q8, q12, %f18[0]    \n"
-
-                    "vext.32    q12, q10, q11, #3   \n" // r03
-
-                    "vmla.f32   q9, q13, %f23[0]    \n"
-
-                    "vext.32    q13, q14, q15, #3   \n" // r33
-
-                    "vmla.f32   q8, q12, %f18[1]    \n"
-
-                    "add        %3, #16             \n"
-
-                    "vmla.f32   q9, q13, %f23[1]    \n"
-
-                    "add        %4, #16             \n"
-
-                    // r1
-                    "pld        [%4, #256]          \n"
-                    "vld1.f32   {d28-d31}, [%4]     \n" // q14 q15 = r10 r14
-
-                    "add        %8, #16             \n"
-
-                    "vst1.f32   {d16-d17}, [%1]!    \n"
-
-                    "vmov       q8, %q25            \n" // q8 = _bias0
-
-                    "subs       %0, #1              \n"
-
-                    "vst1.f32   {d18-d19}, [%2]!    \n"
-
-                    "bne        0b                  \n"
-                    : "=r"(nn),      // %0
-                    "=r"(outptr),  // %1
-                    "=r"(outptr2), // %2
-                    "=r"(r0),      // %3
-                    "=r"(r1),      // %4
-                    "=r"(r2),      // %5
-                    "=r"(r3),      // %6
-                    "=r"(r4),      // %7
-                    "=r"(r5)       // %8
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(outptr2),
-                    "3"(r0),
-                    "4"(r1),
-                    "5"(r2),
-                    "6"(r3),
-                    "7"(r4),
-                    "8"(r5),
-                    "w"(_k0123),     // %18
-                    "w"(_k4567),     // %19
-                    "w"(_k891011),   // %20
-                    "w"(_k12131415), // %21
-                    "w"(_k16171819), // %22
-                    "w"(_k20212223), // %23
-                    "w"(_k24242424), // %24
-                    "w"(_bias0)      // %25
-                    : "cc", "memory", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-                float sum = bias0;
-                float sum2 = bias0;
-#if __ARM_NEON
-                // TODO neon assembly optimize
-                float32x4_t _r1 = vld1q_f32(r1);
-                float32x4_t _k1 = vld1q_f32(k1);
-                float32x4_t _sum = vmulq_f32(_r1, _k1);
-                float32x4_t _sum2 = vmulq_f32(_r1, _k0123);
-
-                float32x4_t _r2 = vld1q_f32(r2);
-                float32x4_t _k2 = vld1q_f32(k2);
-                _sum = vmlaq_f32(_sum, _r2, _k2);
-                _sum2 = vmlaq_f32(_sum2, _r2, _k1);
-
-                float32x4_t _r3 = vld1q_f32(r3);
-                float32x4_t _k3 = vld1q_f32(k3);
-                _sum = vmlaq_f32(_sum, _r3, _k3);
-                _sum2 = vmlaq_f32(_sum2, _r3, _k2);
-
-                float32x4_t _r4 = vld1q_f32(r4);
-                _sum = vmlaq_f32(_sum, _r4, _k20212223);
-                _sum2 = vmlaq_f32(_sum2, _r4, _k3);
-
-                float32x4_t _r0 = vld1q_f32(r0);
-                _sum = vmlaq_f32(_sum, _r0, _k0123);
-                float32x4_t _r5 = vld1q_f32(r5);
-                _sum2 = vmlaq_f32(_sum2, _r5, _k20212223);
-
-                float32x4_t _k_t4 = {};
-
-                _k_t4 = vsetq_lane_f32(k0[4], _k_t4, 0);
-                _k_t4 = vsetq_lane_f32(k1[4], _k_t4, 1);
-                _k_t4 = vsetq_lane_f32(k2[4], _k_t4, 2);
-                _k_t4 = vsetq_lane_f32(k3[4], _k_t4, 3);
-
-                float32x4_t _r_t4 = {};
-
-                _r_t4 = vsetq_lane_f32(r0[4], _r_t4, 0);
-                _r_t4 = vsetq_lane_f32(r1[4], _r_t4, 1);
-                _r_t4 = vsetq_lane_f32(r2[4], _r_t4, 2);
-                _r_t4 = vsetq_lane_f32(r3[4], _r_t4, 3);
-                _sum = vmlaq_f32(_sum, _r_t4, _k_t4);
-
-                sum += r4[4] * k4[4];
-
-                _r_t4 = vextq_f32(_r_t4, _r_t4, 1);
-                _r_t4 = vsetq_lane_f32(r4[4], _r_t4, 3);
-                _sum2 = vmlaq_f32(_sum2, _r_t4, _k_t4);
-
-                sum2 += r5[4] * k4[4];
-
-                float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-                float32x2_t _ss2 = vadd_f32(vget_low_f32(_sum2), vget_high_f32(_sum2));
-                float32x2_t _ss_ss2 = vpadd_f32(_ss, _ss2);
-
-                sum += vget_lane_f32(_ss_ss2, 0);
-                sum2 += vget_lane_f32(_ss_ss2, 1);
-#else
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum += r0[3] * k0[3];
-                sum += r0[4] * k0[4];
-
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum += r1[3] * k1[3];
-                sum += r1[4] * k1[4];
-
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-                sum += r2[3] * k2[3];
-                sum += r2[4] * k2[4];
-
-                sum += r3[0] * k3[0];
-                sum += r3[1] * k3[1];
-                sum += r3[2] * k3[2];
-                sum += r3[3] * k3[3];
-                sum += r3[4] * k3[4];
-
-                sum += r4[0] * k4[0];
-                sum += r4[1] * k4[1];
-                sum += r4[2] * k4[2];
-                sum += r4[3] * k4[3];
-                sum += r4[4] * k4[4];
-
-                sum2 += r1[0] * k0[0];
-                sum2 += r1[1] * k0[1];
-                sum2 += r1[2] * k0[2];
-                sum2 += r1[3] * k0[3];
-                sum2 += r1[4] * k0[4];
-
-                sum2 += r2[0] * k1[0];
-                sum2 += r2[1] * k1[1];
-                sum2 += r2[2] * k1[2];
-                sum2 += r2[3] * k1[3];
-                sum2 += r2[4] * k1[4];
-
-                sum2 += r3[0] * k2[0];
-                sum2 += r3[1] * k2[1];
-                sum2 += r3[2] * k2[2];
-                sum2 += r3[3] * k2[3];
-                sum2 += r3[4] * k2[4];
-
-                sum2 += r4[0] * k3[0];
-                sum2 += r4[1] * k3[1];
-                sum2 += r4[2] * k3[2];
-                sum2 += r4[3] * k3[3];
-                sum2 += r4[4] * k3[4];
-
-                sum2 += r5[0] * k4[0];
-                sum2 += r5[1] * k4[1];
-                sum2 += r5[2] * k4[2];
-                sum2 += r5[3] * k4[3];
-                sum2 += r5[4] * k4[4];
-#endif // __ARM_NEON
-                *outptr = sum;
-                *outptr2 = sum2;
-
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                r4++;
-                r5++;
-                outptr++;
-                outptr2++;
-            }
-
-            r0 += 4 + w;
-            r1 += 4 + w;
-            r2 += 4 + w;
-            r3 += 4 + w;
-            r4 += 4 + w;
-            r5 += 4 + w;
-
-            outptr += outw;
-            outptr2 += outw;
-        }
-
-        for (; i < outh; i++)
-        {
-#if __ARM_NEON
-#if __aarch64__
-            int nn = outw >> 3;
-            int remain = outw & 7;
-#else
-            int nn = outw >> 2;
-            int remain = outw & 3;
-#endif // __aarch64__
-#else
-            int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    // v10 v11
-                    // r0
-                    "prfm   pldl1keep, [%2, #384]           \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s}, [%2]  \n" // v16 v17 v18 = r00 r04 r08
-
-                    "mov    v8.16b, %21.16b                 \n" // v8 = _bias0
-                    "mov    v9.16b, %21.16b                 \n" // v9 = _bias0
-
-                    "0:                                     \n"
-
-                    "fmul   v10.4s, v16.4s, %14.s[0]         \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #4   \n" // r01
-
-                    "fmul   v11.4s, v17.4s, %14.s[0]         \n"
-
-                    "ext    v20.16b, v17.16b, v18.16b, #4   \n" // r05
-
-                    "fmla   v8.4s, v17.4s, %15.s[0]         \n"
-
-                    "ext    v21.16b, v16.16b, v17.16b, #8   \n" // r02
-
-                    "fmla   v9.4s, v18.4s, %15.s[0]         \n"
-
-                    "ext    v22.16b, v17.16b, v18.16b, #8   \n" // r06
-
-                    "fmla   v10.4s, v19.4s, %14.s[1]         \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #12  \n" // r03
-
-                    "fmla   v11.4s, v20.4s, %14.s[1]         \n"
-
-                    "ext    v20.16b, v17.16b, v18.16b, #12  \n" // r07
-
-                    "fmla   v8.4s, v21.4s, %14.s[2]         \n"
-                    "fmla   v9.4s, v22.4s, %14.s[2]         \n"
-
-                    // r1
-                    "prfm   pldl1keep, [%3, #384]           \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s}, [%3]  \n" // v12 v13 v14 = r10 r14 r18
-
-                    "fmla   v10.4s, v19.4s, %14.s[3]         \n"
-                    "fmla   v11.4s, v20.4s, %14.s[3]         \n"
-
-                    "fmla   v8.4s, v12.4s, %15.s[1]         \n"
-
-                    "ext    v19.16b, v12.16b, v13.16b, #4   \n" // r11
-
-                    "fmla   v9.4s, v13.4s, %15.s[1]         \n"
-
-                    "ext    v20.16b, v13.16b, v14.16b, #4   \n" // r15
-
-                    "fmla   v10.4s, v13.4s, %16.s[1]         \n"
-
-                    "ext    v21.16b, v12.16b, v13.16b, #8   \n" // r12
-
-                    "fmla   v11.4s, v14.4s, %16.s[1]         \n"
-
-                    "ext    v22.16b, v13.16b, v14.16b, #8   \n" // r16
-
-                    "fmla   v8.4s, v19.4s, %15.s[2]         \n"
-
-                    "ext    v19.16b, v12.16b, v13.16b, #12  \n" // r13
-
-                    "fmla   v9.4s, v20.4s, %15.s[2]         \n"
-
-                    "ext    v20.16b, v13.16b, v14.16b, #12  \n" // r17
-
-                    "fmla   v10.4s, v21.4s, %15.s[3]         \n"
-                    "fmla   v11.4s, v22.4s, %15.s[3]         \n"
-
-                    // r2
-                    "prfm   pldl1keep, [%4, #384]           \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s}, [%4]  \n" // v16 v17 v18 = r20 r24 r28
-
-                    "fmla   v8.4s, v19.4s, %16.s[0]         \n"
-                    "fmla   v9.4s, v20.4s, %16.s[0]         \n"
-
-                    "fmla   v10.4s, v16.4s, %16.s[2]         \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #4   \n" // r21
-
-                    "fmla   v11.4s, v17.4s, %16.s[2]         \n"
-
-                    "ext    v20.16b, v17.16b, v18.16b, #4   \n" // r25
-
-                    "fmla   v8.4s, v17.4s, %17.s[2]         \n"
-
-                    "ext    v21.16b, v16.16b, v17.16b, #8   \n" // r22
-
-                    "fmla   v9.4s, v18.4s, %17.s[2]         \n"
-
-                    "ext    v22.16b, v17.16b, v18.16b, #8   \n" // r26
-
-                    "fmla   v10.4s, v19.4s, %16.s[3]         \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #12  \n" // r23
-
-                    "fmla   v11.4s, v20.4s, %16.s[3]         \n"
-
-                    "ext    v20.16b, v17.16b, v18.16b, #12  \n" // r27
-
-                    "fmla   v8.4s, v21.4s, %17.s[0]         \n"
-                    "fmla   v9.4s, v22.4s, %17.s[0]         \n"
-
-                    // r3
-                    "prfm   pldl1keep, [%5, #384]           \n"
-                    "ld1    {v12.4s, v13.4s, v14.4s}, [%5]  \n" // v12 v13 v14 = r30 r34 r38
-
-                    "fmla   v10.4s, v19.4s, %17.s[1]         \n"
-                    "fmla   v11.4s, v20.4s, %17.s[1]         \n"
-
-                    "fmla   v8.4s, v12.4s, %17.s[3]         \n"
-
-                    "ext    v19.16b, v12.16b, v13.16b, #4   \n" // r11
-
-                    "fmla   v9.4s, v13.4s, %17.s[3]         \n"
-
-                    "ext    v20.16b, v13.16b, v14.16b, #4   \n" // r15
-
-                    "fmla   v10.4s, v13.4s, %18.s[3]         \n"
-
-                    "ext    v21.16b, v12.16b, v13.16b, #8   \n" // r12
-
-                    "fmla   v11.4s, v14.4s, %18.s[3]         \n"
-
-                    "ext    v22.16b, v13.16b, v14.16b, #8   \n" // r16
-
-                    "fmla   v8.4s, v19.4s, %18.s[0]         \n"
-
-                    "ext    v19.16b, v12.16b, v13.16b, #12  \n" // r13
-
-                    "fmla   v9.4s, v20.4s, %18.s[0]         \n"
-
-                    "ext    v20.16b, v13.16b, v14.16b, #12  \n" // r17
-
-                    "fmla   v10.4s, v21.4s, %18.s[1]         \n"
-                    "fmla   v11.4s, v22.4s, %18.s[1]         \n"
-
-                    // r4
-                    "prfm   pldl1keep, [%6, #384]           \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s}, [%6]  \n" // v16 v17 v18 = r40 r44 r48
-
-                    "fmla   v8.4s, v19.4s, %18.s[2]         \n"
-                    "fmla   v9.4s, v20.4s, %18.s[2]         \n"
-
-                    "fmla   v10.4s, v16.4s, %19.s[0]         \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #4   \n" // r41
-
-                    "fmla   v11.4s, v17.4s, %19.s[0]         \n"
-
-                    "ext    v20.16b, v17.16b, v18.16b, #4   \n" // r45
-
-                    "fmla   v8.4s, v17.4s, %20.s[0]         \n"
-
-                    "ext    v21.16b, v16.16b, v17.16b, #8   \n" // r42
-
-                    "fmla   v9.4s, v18.4s, %20.s[0]         \n"
-
-                    "ext    v22.16b, v17.16b, v18.16b, #8   \n" // r46
-
-                    "fmla   v10.4s, v19.4s, %19.s[1]         \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #12  \n" // r43
-
-                    "fmla   v11.4s, v20.4s, %19.s[1]         \n"
-
-                    "ext    v20.16b, v17.16b, v18.16b, #12  \n" // r47
-
-                    "fmla   v8.4s, v21.4s, %19.s[2]         \n"
-
-                    "add    %2, %2, #32                     \n"
-
-                    "fmla   v9.4s, v22.4s, %19.s[2]         \n"
-
-                    "add    %3, %3, #32                     \n"
-
-                    "fmla   v10.4s, v19.4s, %19.s[3]         \n"
-
-                    "add    %4, %4, #32                     \n"
-
-                    "fmla   v11.4s, v20.4s, %19.s[3]         \n"
-
-                    // r0
-                    "prfm   pldl1keep, [%2, #384]           \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s}, [%2]  \n" // v16 v17 v18 = r00 r04 r08
-
-                    "add    %5, %5, #32                     \n"
-
-                    "fadd   v10.4s, v8.4s, v10.4s           \n"
-
-                    "add    %6, %6, #32                     \n"
-
-                    "fadd   v11.4s, v9.4s, v11.4s           \n"
-
-                    "mov    v8.16b, %21.16b                 \n" // v8 = _bias0
-                    "mov    v9.16b, %21.16b                 \n" // v9 = _bias0
-
-                    "subs   %w0, %w0, #1                    \n"
-
-                    "st1    {v10.4s, v11.4s}, [%1], #32     \n"
-
-                    "bne    0b                              \n"
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0),     // %2
-                    "=r"(r1),     // %3
-                    "=r"(r2),     // %4
-                    "=r"(r3),     // %5
-                    "=r"(r4)      // %6
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "w"(_k0123),     // %14
-                    "w"(_k4567),     // %15
-                    "w"(_k891011),   // %16
-                    "w"(_k12131415), // %17
-                    "w"(_k16171819), // %18
-                    "w"(_k20212223), // %19
-                    "w"(_k24242424), // %20
-                    "w"(_bias0)      // %21
-                    : "cc", "memory", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v16", "v17", "v18", "v19", "v20", "v21", "v22");
-            }
-
-            if (remain >= 4)
-            {
-                remain -= 4;
-                asm volatile(
-                    // r0
-                    "prfm   pldl1keep, [%1, #256]           \n"
-                    "ld1    {v16.4s, v17.4s}, [%1]          \n" // v16 v17 = r00 r04
-
-                    "mov    v8.16b, %19.16b                 \n" // v8 = _bias0
-
-                    "add    %1, %1, #16                     \n"
-
-                    "fmul   v9.4s, v16.4s, %12.s[0]         \n"
-
-                    "ext    v18.16b, v16.16b, v17.16b, #4   \n" // r01
-
-                    "fmla   v8.4s, v17.4s, %13.s[0]         \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #8   \n" // r02
-
-                    "fmla   v9.4s, v18.4s, %12.s[1]         \n"
-
-                    "ext    v20.16b, v16.16b, v17.16b, #12  \n" // r03
-
-                    "fmla   v8.4s, v19.4s, %12.s[2]         \n"
-
-                    // r1
-                    "prfm   pldl1keep, [%2, #256]           \n"
-                    "ld1    {v10.4s, v11.4s}, [%2]          \n" // v10 v11 = r10 r14
-
-                    "fmla   v9.4s, v20.4s, %12.s[3]         \n"
-
-                    "add    %2, %2, #16                     \n"
-
-                    "fmla   v8.4s, v10.4s, %13.s[1]         \n"
-
-                    "ext    v12.16b, v10.16b, v11.16b, #4   \n" // r11
-
-                    "fmla   v9.4s, v11.4s, %14.s[1]         \n"
-
-                    "ext    v13.16b, v10.16b, v11.16b, #8   \n" // r12
-
-                    "fmla   v8.4s, v12.4s, %13.s[2]         \n"
-
-                    "ext    v14.16b, v10.16b, v11.16b, #12  \n" // r13
-
-                    "fmla   v9.4s, v13.4s, %13.s[3]         \n"
-
-                    // r2
-                    "prfm   pldl1keep, [%3, #256]           \n"
-                    "ld1    {v16.4s, v17.4s}, [%3]          \n" // v16 v17 = r20 r24
-
-                    "fmla   v8.4s, v14.4s, %14.s[0]         \n"
-
-                    "add    %3, %3, #16                     \n"
-
-                    "fmla   v9.4s, v16.4s, %14.s[2]         \n"
-
-                    "ext    v18.16b, v16.16b, v17.16b, #4   \n" // r21
-
-                    "fmla   v8.4s, v17.4s, %15.s[2]         \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #8   \n" // r22
-
-                    "fmla   v9.4s, v18.4s, %14.s[3]         \n"
-
-                    "ext    v20.16b, v16.16b, v17.16b, #12  \n" // r23
-
-                    "fmla   v8.4s, v19.4s, %15.s[0]         \n"
-
-                    // r3
-                    "prfm   pldl1keep, [%4, #256]           \n"
-                    "ld1    {v10.4s, v11.4s}, [%4]          \n" // v10 v11 = r30 r34
-
-                    "fmla   v9.4s, v20.4s, %15.s[1]         \n"
-
-                    "add    %4, %4, #16                     \n"
-
-                    "fmla   v8.4s, v10.4s, %15.s[3]         \n"
-
-                    "ext    v12.16b, v10.16b, v11.16b, #4   \n" // r31
-
-                    "fmla   v9.4s, v11.4s, %16.s[3]         \n"
-
-                    "ext    v13.16b, v10.16b, v11.16b, #8   \n" // r32
-
-                    "fmla   v8.4s, v12.4s, %16.s[0]         \n"
-
-                    "ext    v14.16b, v10.16b, v11.16b, #12  \n" // r33
-
-                    "fmla   v9.4s, v13.4s, %16.s[1]         \n"
-
-                    // r4
-                    "prfm   pldl1keep, [%5, #256]           \n"
-                    "ld1    {v16.4s, v17.4s}, [%5]          \n" // v16 v17 = r40 r44
-
-                    "fmla   v8.4s, v14.4s, %16.s[2]         \n"
-
-                    "add    %5, %5, #16                     \n"
-
-                    "fmla   v9.4s, v16.4s, %17.s[0]         \n"
-
-                    "ext    v18.16b, v16.16b, v17.16b, #4   \n" // r41
-
-                    "fmla   v8.4s, v17.4s, %18.s[0]         \n"
-
-                    "ext    v19.16b, v16.16b, v17.16b, #8   \n" // r42
-
-                    "fmla   v9.4s, v18.4s, %17.s[1]         \n"
-
-                    "ext    v20.16b, v16.16b, v17.16b, #12  \n" // r43
-
-                    "fmla   v8.4s, v19.4s, %17.s[2]         \n"
-
-                    "fmla   v9.4s, v20.4s, %17.s[3]         \n"
-
-                    "fadd   v8.4s, v8.4s, v9.4s             \n"
-
-                    "st1    {v8.4s}, [%0], #16              \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(r0),     // %1
-                    "=r"(r1),     // %2
-                    "=r"(r2),     // %3
-                    "=r"(r3),     // %4
-                    "=r"(r4)      // %5
-                    : "0"(outptr),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "4"(r3),
-                    "5"(r4),
-                    "w"(_k0123),     // %12
-                    "w"(_k4567),     // %13
-                    "w"(_k891011),   // %14
-                    "w"(_k12131415), // %15
-                    "w"(_k16171819), // %16
-                    "w"(_k20212223), // %17
-                    "w"(_k24242424), // %18
-                    "w"(_bias0)      // %19
-                    : "cc", "memory", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v16", "v17", "v18", "v19", "v20");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    // r0
-                    "pld        [%2, #256]          \n"
-                    "vld1.f32   {d20-d23}, [%2]     \n" // q10 q11 = r00 r04
-
-                    "vmov       q8, %q21            \n" // q8 = _bias0
-
-                    "0:                             \n"
-
-                    "vmul.f32   q9, q10, %e14[0]    \n"
-
-                    "vext.32    q12, q10, q11, #1   \n" // r01
-
-                    "vmla.f32   q8, q11, %e15[0]    \n"
-
-                    "vext.32    q13, q10, q11, #2   \n" // r02
-
-                    "vmla.f32   q9, q12, %e14[1]    \n"
-
-                    "vext.32    q12, q10, q11, #3   \n" // r03
-
-                    "vmla.f32   q8, q13, %f14[0]    \n"
-
-                    // r1
-                    "pld        [%3, #256]          \n"
-                    "vld1.f32   {d28-d31}, [%3]     \n" // q14 q15 = r10 r14
-
-                    "vmla.f32   q9, q12, %f14[1]    \n"
-
-                    "add        %3, #16             \n"
-
-                    "vmla.f32   q8, q14, %e15[1]    \n"
-
-                    "vext.32    q12, q14, q15, #1   \n" // r11
-
-                    "vmla.f32   q9, q15, %e16[1]    \n"
-
-                    "vext.32    q13, q14, q15, #2   \n" // r12
-
-                    "vmla.f32   q8, q12, %f15[0]    \n"
-
-                    "vext.32    q12, q14, q15, #3   \n" // r13
-
-                    "vmla.f32   q9, q13, %f15[1]    \n"
-
-                    // r2
-                    "pld        [%4, #256]          \n"
-                    "vld1.f32   {d20-d23}, [%4]     \n" // q10 q11 = r20 r24
-
-                    "vmla.f32   q8, q12, %e16[0]    \n"
-
-                    "add        %4, #16             \n"
-
-                    "vmla.f32   q9, q10, %f16[0]    \n"
-
-                    "vext.32    q12, q10, q11, #1   \n" // r21
-
-                    "vmla.f32   q8, q11, %f17[0]    \n"
-
-                    "vext.32    q13, q10, q11, #2   \n" // r22
-
-                    "vmla.f32   q9, q12, %f16[1]    \n"
-
-                    "vext.32    q12, q10, q11, #3   \n" // r23
-
-                    "vmla.f32   q8, q13, %e17[0]    \n"
-
-                    // r3
-                    "pld        [%5, #256]          \n"
-                    "vld1.f32   {d28-d31}, [%5]     \n" // q14 q15 = r30 r34
-
-                    "vmla.f32   q9, q12, %e17[1]    \n"
-
-                    "add        %5, #16             \n"
-
-                    "vmla.f32   q8, q14, %f17[1]    \n"
-
-                    "vext.32    q12, q14, q15, #1   \n" // r31
-
-                    "vmla.f32   q9, q15, %f18[1]    \n"
-
-                    "vext.32    q13, q14, q15, #2   \n" // r32
-
-                    "vmla.f32   q8, q12, %e18[0]    \n"
-
-                    "vext.32    q12, q14, q15, #3   \n" // r33
-
-                    "vmla.f32   q9, q13, %e18[1]    \n"
-
-                    // r4
-                    "pld        [%6, #256]          \n"
-                    "vld1.f32   {d20-d23}, [%6]     \n" // q10 q11 = r40 r44
-
-                    "vmla.f32   q8, q12, %f18[0]    \n"
-
-                    "add        %6, #16             \n"
-
-                    "vmla.f32   q9, q10, %e19[0]    \n"
-
-                    "vext.32    q12, q10, q11, #1   \n" // r41
-
-                    "vmla.f32   q8, q11, %e20[0]    \n"
-
-                    "vext.32    q13, q10, q11, #2   \n" // r42
-
-                    "vmla.f32   q9, q12, %e19[1]    \n"
-
-                    "vext.32    q12, q10, q11, #3   \n" // r43
-
-                    "vmla.f32   q8, q13, %f19[0]    \n"
-
-                    "add        %2, #16             \n"
-
-                    "vmla.f32   q9, q12, %f19[1]    \n"
-
-                    // r0
-                    "pld        [%2, #256]          \n"
-                    "vld1.f32   {d20-d23}, [%2]     \n" // q10 q11 = r00 r04
-
-                    "vadd.f32   q9, q9, q8          \n"
-
-                    "vmov       q8, %q21            \n" // q8 = _bias0
-
-                    "subs       %0, #1              \n"
-
-                    "vst1.f32   {d18-d19}, [%1]!    \n"
-
-                    "bne        0b                  \n"
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0),     // %2
-                    "=r"(r1),     // %3
-                    "=r"(r2),     // %4
-                    "=r"(r3),     // %5
-                    "=r"(r4)      // %6
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "w"(_k0123),     // %14
-                    "w"(_k4567),     // %15
-                    "w"(_k891011),   // %16
-                    "w"(_k12131415), // %17
-                    "w"(_k16171819), // %18
-                    "w"(_k20212223), // %19
-                    "w"(_k24242424), // %20
-                    "w"(_bias0)      // %21
-                    : "cc", "memory", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-#if __ARM_NEON
-#if __aarch64__
-                // TODO neon assembly optimize
-                float sum = bias0;
-
-                float32x4_t _r0 = vld1q_f32(r0);
-                float32x4_t _sum = vmulq_f32(_r0, _k0123);
-
-                float32x4_t _r1 = vld1q_f32(r1);
-                _sum = vmlaq_f32(_sum, _r1, vld1q_f32(k1));
-
-                float32x4_t _r2 = vld1q_f32(r2);
-                _sum = vmlaq_f32(_sum, _r2, vld1q_f32(k2));
-
-                float32x4_t _r3 = vld1q_f32(r3);
-                _sum = vmlaq_f32(_sum, _r3, vld1q_f32(k3));
-
-                float32x4_t _r4 = vld1q_f32(r4);
-                _sum = vmlaq_f32(_sum, _r4, _k20212223);
-
-                float32x4_t _k_t4 = {};
-
-                _k_t4 = vsetq_lane_f32(k0[4], _k_t4, 0);
-                _k_t4 = vsetq_lane_f32(k1[4], _k_t4, 1);
-                _k_t4 = vsetq_lane_f32(k2[4], _k_t4, 2);
-                _k_t4 = vsetq_lane_f32(k3[4], _k_t4, 3);
-
-                float32x4_t _r_t4 = {};
-
-                _r_t4 = vsetq_lane_f32(r0[4], _r_t4, 0);
-                _r_t4 = vsetq_lane_f32(r1[4], _r_t4, 1);
-                _r_t4 = vsetq_lane_f32(r2[4], _r_t4, 2);
-                _r_t4 = vsetq_lane_f32(r3[4], _r_t4, 3);
-                _sum = vmlaq_f32(_sum, _r_t4, _k_t4);
-
-                sum += r4[4] * k4[4];
-
-                float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-                _ss = vpadd_f32(_ss, _ss);
-
-                sum += vget_lane_f32(_ss, 0);
-
-                *outptr = sum;
-
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                r4++;
-                outptr++;
-#else
-                // TODO neon assembly optimize
-                asm volatile(
-                    "veor       q14, q14            \n"
-                    "vext.32    q14, %q19, q14, #3  \n" // q14 = bias0 0 0 0
-
-                    "vld1.f32   {d16-d17}, [%1]     \n" // q8 = r00 r01 r02 r03
-
-                    "vld1.f32   {d18-d19}, [%2]     \n" // q9 = r10 r11 r12 r13(X)
-                    "add        r4, %1, #16         \n"
-                    "vld1.f32   {d19[1]}, [r4]      \n"
-                    "vext.32    q9, q9, q9, #3      \n" // q9 = r04 r10 r11 r12
-
-                    "vmla.f32   q14, q8, %q12       \n"
-
-                    "add        r4, %2, #12         \n"
-                    "vld1.f32   {d20}, [r4]         \n" // d20 = r13 r14
-                    "vld1.f32   {d21}, [%3]         \n" // d21 = r20 r21
-
-                    "vmla.f32   q14, q9, %q13       \n"
-
-                    "add        r4, %3, #8          \n"
-                    "vld1.f32   {d22-d23}, [r4]     \n" // q11 = r22 r23 r24 X
-                    "vld1.f32   {d23[1]}, [%4]      \n" // q11 = r22 r23 r24 r30
-
-                    "vmla.f32   q14, q10, %q14      \n"
-
-                    "add        r4, %4, #4          \n"
-                    "vld1.f32   {d24-d25}, [r4]     \n" // q12 = r31 r32 r33 r34
-
-                    "vmla.f32   q14, q11, %q15      \n"
-
-                    "vld1.f32   {d26-d27}, [%5]     \n" // q13 = r40 r41 r42 r43
-
-                    "vmla.f32   q14, q12, %q16      \n"
-
-                    "veor       d30, d30            \n"
-                    "add        r4, %5, #16         \n"
-                    "vld1.f32   {d30[0]}, [r4]      \n" // d30 = r44 0
-
-                    "vmla.f32   q14, q13, %q17      \n"
-
-                    "vmla.f32   d28, d30, %e18      \n"
-
-                    "add        %1, #4              \n"
-
-                    // h-sum
-                    "vadd.f32   d28, d28, d29       \n"
-
-                    "add        %2, #4              \n"
-                    "add        %3, #4              \n"
-
-                    "vpadd.f32  d28, d28, d28       \n"
-
-                    "add        %4, #4              \n"
-                    "add        %5, #4              \n"
-
-                    "vst1.f32   {d28[0]}, [%0]!     \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(r0),     // %1
-                    "=r"(r1),     // %2
-                    "=r"(r2),     // %3
-                    "=r"(r3),     // %4
-                    "=r"(r4)      // %5
-                    : "0"(outptr),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "4"(r3),
-                    "5"(r4),
-                    "w"(_k0123),     // %12
-                    "w"(_k4567),     // %13
-                    "w"(_k891011),   // %14
-                    "w"(_k12131415), // %15
-                    "w"(_k16171819), // %16
-                    "w"(_k20212223), // %17
-                    "w"(_k24242424), // %18
-                    "w"(_bias0)      // %19
-                    : "cc", "memory", "r4", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else
-                float sum = bias0;
-
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum += r0[3] * k0[3];
-                sum += r0[4] * k0[4];
-
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum += r1[3] * k1[3];
-                sum += r1[4] * k1[4];
-
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-                sum += r2[3] * k2[3];
-                sum += r2[4] * k2[4];
-
-                sum += r3[0] * k3[0];
-                sum += r3[1] * k3[1];
-                sum += r3[2] * k3[2];
-                sum += r3[3] * k3[3];
-                sum += r3[4] * k3[4];
-
-                sum += r4[0] * k4[0];
-                sum += r4[1] * k4[1];
-                sum += r4[2] * k4[2];
-                sum += r4[3] * k4[3];
-                sum += r4[4] * k4[4];
-
-                *outptr = sum;
-
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                r4++;
-                outptr++;
-#endif
-            }
-
-            r0 += 4;
-            r1 += 4;
-            r2 += 4;
-            r3 += 4;
-            r4 += 4;
-        }
-    }
-}
-
-static void convdw5x5s2_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    //int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    //int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const int group = bottom_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const float bias0 = bias ? bias[g] : 0.f;
-
-        const float* kernel0 = kernel + g * 25;
-
-        float* outptr = out;
-
-        const float* img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0;
-        const float* r1 = img0 + w;
-        const float* r2 = img0 + w * 2;
-        const float* r3 = img0 + w * 3;
-        const float* r4 = img0 + w * 4;
-
-        const float* k0 = kernel0;
-        const float* k1 = kernel0 + 5;
-        const float* k2 = kernel0 + 10;
-        const float* k3 = kernel0 + 15;
-        const float* k4 = kernel0 + 20;
-
-#if __ARM_NEON
-        float32x4_t _k0123 = vld1q_f32(kernel0);
-        float32x4_t _k4567 = vld1q_f32(kernel0 + 4);
-        float32x4_t _k891011 = vld1q_f32(kernel0 + 8);
-        float32x4_t _k12131415 = vld1q_f32(kernel0 + 12);
-        float32x4_t _k16171819 = vld1q_f32(kernel0 + 16);
-        float32x4_t _k20212223 = vld1q_f32(kernel0 + 20);
-        float32x4_t _k24242424 = vdupq_n_f32(kernel0[24]);
-
-        float32x4_t _bias0 = vdupq_n_f32(bias0);
-#endif // __ARM_NEON
-
-        int i = 0;
-
-        // NOTE unroll outh 2 results somewhat speed drop :| (about -4%)
-        // so we do not implement it here
-
-        for (; i < outh; i++)
-        {
-#if __ARM_NEON
-#if __aarch64__
-            int nn = outw >> 3;
-            int remain = outw & 7;
-#else
-            int nn = outw >> 2;
-            int remain = outw & 3;
-#endif // __aarch64__
-#else
-            int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    // r0
-                    "prfm   pldl1keep, [%2, #256]           \n"
-                    "ld2    {v16.4s, v17.4s}, [%2], #32     \n" // v16 v17 = r00 r01
-
-                    "mov    v8.16b, %21.16b                 \n" // v8 = _bias0
-                    "mov    v9.16b, %21.16b                 \n" // v9 = _bias0
-
-                    "prfm   pldl1keep, [%2, #256]           \n"
-                    "ld2    {v18.4s, v19.4s}, [%2], #32     \n" // v18 v19 = r08 r09
-
-                    "0:                                     \n"
-
-                    "fmul   v10.4s, v16.4s, %14.s[0]        \n"
-
-                    "prfm   pldl1keep, [%2, #256]           \n"
-                    "ld2    {v20.4s, v21.4s}, [%2]          \n" // v20 v21 = r016 r017
-
-                    "fmul   v11.4s, v18.4s, %14.s[0]        \n"
-
-                    "ext    v22.16b, v16.16b, v18.16b, #4   \n" // v22 = r02
-
-                    "fmla   v8.4s, v17.4s, %14.s[1]         \n"
-
-                    "ext    v25.16b, v18.16b, v20.16b, #4   \n" // v25 = r010
-
-                    "fmla   v9.4s, v19.4s, %14.s[1]         \n"
-
-                    "ext    v23.16b, v17.16b, v19.16b, #4   \n" // v23 = r03
-
-                    "fmla   v10.4s, v22.4s, %14.s[2]        \n"
-
-                    "ext    v26.16b, v19.16b, v21.16b, #4   \n" // v26 = r011
-
-                    "fmla   v11.4s, v25.4s, %14.s[2]        \n"
-
-                    "ext    v24.16b, v16.16b, v18.16b, #8   \n" // v24 = r04
-
-                    "fmla   v8.4s, v23.4s, %14.s[3]         \n"
-
-                    "ext    v27.16b, v18.16b, v20.16b, #8   \n" // v27 = r012
-
-                    "fmla   v9.4s, v26.4s, %14.s[3]         \n"
-
-                    // r1
-                    "prfm   pldl1keep, [%3, #256]           \n"
-                    "ld2    {v12.4s, v13.4s}, [%3], #32     \n" // v12 v13 = r10 r11
-
-                    "fmla   v10.4s, v24.4s, %15.s[0]        \n"
-
-                    "prfm   pldl1keep, [%3, #256]           \n"
-                    "ld2    {v14.4s, v15.4s}, [%3], #32     \n" // v14 v15 = r18 r19
-
-                    "fmla   v11.4s, v27.4s, %15.s[0]        \n"
-
-                    "fmla   v8.4s, v12.4s, %15.s[1]         \n"
-
-                    "prfm   pldl1keep, [%3, #256]           \n"
-                    "ld2    {v20.4s, v21.4s}, [%3]          \n" // v20 v21 = r116 r117
-
-                    "fmla   v9.4s, v14.4s, %15.s[1]         \n"
-
-                    "ext    v22.16b, v12.16b, v14.16b, #4   \n" // v22 = r12
-
-                    "fmla   v10.4s, v13.4s, %15.s[2]        \n"
-
-                    "ext    v25.16b, v14.16b, v20.16b, #4   \n" // v25 = r110
-
-                    "fmla   v11.4s, v15.4s, %15.s[2]        \n"
-
-                    "ext    v23.16b, v13.16b, v15.16b, #4   \n" // v23 = r13
-
-                    "fmla   v8.4s, v22.4s, %15.s[3]         \n"
-
-                    "ext    v26.16b, v15.16b, v21.16b, #4   \n" // v26 = r111
-
-                    "fmla   v9.4s, v25.4s, %15.s[3]         \n"
-
-                    "ext    v24.16b, v12.16b, v14.16b, #8   \n" // v24 = r14
-
-                    "fmla   v10.4s, v23.4s, %16.s[0]        \n"
-
-                    "ext    v27.16b, v14.16b, v20.16b, #8   \n" // v27 = r112
-
-                    "fmla   v11.4s, v26.4s, %16.s[0]        \n"
-
-                    // r2
-                    "prfm   pldl1keep, [%4, #256]           \n"
-                    "ld2    {v16.4s, v17.4s}, [%4], #32     \n" // v16 v17 = r20 r21
-
-                    "fmla   v8.4s, v24.4s, %16.s[1]         \n"
-
-                    "prfm   pldl1keep, [%4, #256]           \n"
-                    "ld2    {v18.4s, v19.4s}, [%4], #32     \n" // v18 v19 = r28 r29
-
-                    "fmla   v9.4s, v27.4s, %16.s[1]         \n"
-
-                    "fmla   v10.4s, v16.4s, %16.s[2]        \n"
-
-                    "prfm   pldl1keep, [%4, #256]           \n"
-                    "ld2    {v20.4s, v21.4s}, [%4]          \n" // v20 v21 = r216 r217
-
-                    "fmla   v11.4s, v18.4s, %16.s[2]        \n"
-
-                    "ext    v22.16b, v16.16b, v18.16b, #4   \n" // v22 = r22
-
-                    "fmla   v8.4s, v17.4s, %16.s[3]         \n"
-
-                    "ext    v25.16b, v18.16b, v20.16b, #4   \n" // v25 = r210
-
-                    "fmla   v9.4s, v19.4s, %16.s[3]         \n"
-
-                    "ext    v23.16b, v17.16b, v19.16b, #4   \n" // v23 = r23
-
-                    "fmla   v10.4s, v22.4s, %17.s[0]        \n"
-
-                    "ext    v26.16b, v19.16b, v21.16b, #4   \n" // v26 = r211
-
-                    "fmla   v11.4s, v25.4s, %17.s[0]        \n"
-
-                    "ext    v24.16b, v16.16b, v18.16b, #8   \n" // v24 = r24
-
-                    "fmla   v8.4s, v23.4s, %17.s[1]         \n"
-
-                    "ext    v27.16b, v18.16b, v20.16b, #8   \n" // v27 = r212
-
-                    "fmla   v9.4s, v26.4s, %17.s[1]         \n"
-
-                    // r3
-                    "prfm   pldl1keep, [%5, #256]           \n"
-                    "ld2    {v12.4s, v13.4s}, [%5], #32     \n" // v12 v13 = r30 r31
-
-                    "fmla   v10.4s, v24.4s, %17.s[2]        \n"
-
-                    "prfm   pldl1keep, [%5, #256]           \n"
-                    "ld2    {v14.4s, v15.4s}, [%5], #32     \n" // v14 v15 = r38 r39
-
-                    "fmla   v11.4s, v27.4s, %17.s[2]        \n"
-
-                    "fmla   v8.4s, v12.4s, %17.s[3]         \n"
-
-                    "prfm   pldl1keep, [%5, #256]           \n"
-                    "ld2    {v20.4s, v21.4s}, [%5]          \n" // v20 v21 = r316 r317
-
-                    "fmla   v9.4s, v14.4s, %17.s[3]         \n"
-
-                    "ext    v22.16b, v12.16b, v14.16b, #4   \n" // v22 = r32
-
-                    "fmla   v10.4s, v13.4s, %18.s[0]        \n"
-
-                    "ext    v25.16b, v14.16b, v20.16b, #4   \n" // v25 = r310
-
-                    "fmla   v11.4s, v15.4s, %18.s[0]        \n"
-
-                    "ext    v23.16b, v13.16b, v15.16b, #4   \n" // v23 = r33
-
-                    "fmla   v8.4s, v22.4s, %18.s[1]         \n"
-
-                    "ext    v26.16b, v15.16b, v21.16b, #4   \n" // v26 = r311
-
-                    "fmla   v9.4s, v25.4s, %18.s[1]         \n"
-
-                    "ext    v24.16b, v12.16b, v14.16b, #8   \n" // v24 = r34
-
-                    "fmla   v10.4s, v23.4s, %18.s[2]        \n"
-
-                    "ext    v27.16b, v14.16b, v20.16b, #8   \n" // v27 = r312
-
-                    "fmla   v11.4s, v26.4s, %18.s[2]        \n"
-
-                    // r4
-                    "prfm   pldl1keep, [%6, #256]           \n"
-                    "ld2    {v16.4s, v17.4s}, [%6], #32     \n" // v16 v17 = r40 r41
-
-                    "fmla   v8.4s, v24.4s, %18.s[3]         \n"
-
-                    "prfm   pldl1keep, [%6, #256]           \n"
-                    "ld2    {v18.4s, v19.4s}, [%6], #32     \n" // v18 v19 = r48 r49
-
-                    "fmla   v9.4s, v27.4s, %18.s[3]         \n"
-
-                    "fmla   v10.4s, v16.4s, %19.s[0]        \n"
-
-                    "prfm   pldl1keep, [%6, #256]           \n"
-                    "ld2    {v20.4s, v21.4s}, [%6]          \n" // v20 v21 = r416 r417
-
-                    "fmla   v11.4s, v18.4s, %19.s[0]        \n"
-
-                    "ext    v22.16b, v16.16b, v18.16b, #4   \n" // v22 = r42
-
-                    "fmla   v8.4s, v17.4s, %19.s[1]         \n"
-
-                    "ext    v25.16b, v18.16b, v20.16b, #4   \n" // v25 = r410
-
-                    "fmla   v9.4s, v19.4s, %19.s[1]         \n"
-
-                    "ext    v23.16b, v17.16b, v19.16b, #4   \n" // v23 = r43
-
-                    "fmla   v10.4s, v22.4s, %19.s[2]        \n"
-
-                    "ext    v26.16b, v19.16b, v21.16b, #4   \n" // v26 = r411
-
-                    "fmla   v11.4s, v25.4s, %19.s[2]        \n"
-
-                    "ext    v24.16b, v16.16b, v18.16b, #8   \n" // v24 = r44
-
-                    "fmla   v8.4s, v23.4s, %19.s[3]         \n"
-
-                    "ext    v27.16b, v18.16b, v20.16b, #8   \n" // v27 = r412
-
-                    "fmla   v9.4s, v26.4s, %19.s[3]         \n"
-                    "fmla   v10.4s, v24.4s, %20.s[0]        \n"
-
-                    // r0
-                    "prfm   pldl1keep, [%2, #256]           \n"
-                    "ld2    {v16.4s, v17.4s}, [%2], #32     \n" // v16 v17 = r00 r01
-
-                    "fmla   v11.4s, v27.4s, %20.s[0]        \n"
-
-                    "prfm   pldl1keep, [%2, #256]           \n"
-                    "ld2    {v18.4s, v19.4s}, [%2], #32     \n" // v18 v19 = r08 r09
-
-                    "fadd   v10.4s, v8.4s, v10.4s           \n"
-                    "fadd   v11.4s, v9.4s, v11.4s           \n"
-
-                    "subs   %w0, %w0, #1                    \n"
-
-                    "mov    v8.16b, %21.16b                 \n" // v8 = _bias0
-                    "mov    v9.16b, %21.16b                 \n" // v9 = _bias0
-
-                    "st1    {v10.4s, v11.4s}, [%1], #32     \n"
-
-                    "bne    0b                              \n"
-                    "sub    %2, %2, #64                     \n"
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0),     // %2
-                    "=r"(r1),     // %3
-                    "=r"(r2),     // %4
-                    "=r"(r3),     // %5
-                    "=r"(r4)      // %6
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "w"(_k0123),     // %14
-                    "w"(_k4567),     // %15
-                    "w"(_k891011),   // %16
-                    "w"(_k12131415), // %17
-                    "w"(_k16171819), // %18
-                    "w"(_k20212223), // %19
-                    "w"(_k24242424), // %20
-                    "w"(_bias0)      // %21
-                    : "cc", "memory", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    // r0
-                    "pld        [%2, #256]          \n"
-                    "vld2.f32   {d20-d23}, [%2]!    \n" // q10 q11 = r00 r01
-
-                    "vmov       q8, %q21            \n"
-
-                    "pld        [%2, #128]          \n"
-                    "vld2.f32   {d24-d25}, [%2]     \n" // q12 = r08 x x
-
-                    "0:                             \n"
-
-                    "vmul.f32   q9, q10, %e14[0]    \n"
-
-                    "vmov       d26, d25            \n" // q13 = r09 x x
-
-                    "vext.32    q14, q10, q12, #1   \n" // q14 = r02
-
-                    "vmla.f32   q8, q11, %e14[1]    \n"
-
-                    "vext.32    q15, q11, q13, #1   \n" // q15 = r03
-
-                    "vmla.f32   q9, q14, %f14[0]    \n"
-
-                    "vext.32    q14, q10, q12, #2   \n" // q14 = r04
-
-                    "vmla.f32   q8, q15, %f14[1]    \n"
-
-                    // r1
-                    "pld        [%3, #256]          \n"
-                    "vld2.f32   {d20-d23}, [%3]!    \n" // q10 q11 = r10 r11
-
-                    "vmla.f32   q9, q14, %e15[0]    \n"
-
-                    "pld        [%3, #128]          \n"
-                    "vld2.f32   {d24-d25}, [%3]     \n" // q12 = r18 x x
-
-                    "vmla.f32   q8, q10, %e15[1]    \n"
-
-                    "vmov       d26, d25            \n" // q13 = r19 x x
-
-                    "vext.32    q14, q10, q12, #1   \n" // q14 = r12
-
-                    "vmla.f32   q9, q11, %f15[0]    \n"
-
-                    "vext.32    q15, q11, q13, #1   \n" // q15 = r13
-
-                    "vmla.f32   q8, q14, %f15[1]    \n"
-
-                    "vext.32    q14, q10, q12, #2   \n" // q14 = r14
-
-                    "vmla.f32   q9, q15, %e16[0]    \n"
-
-                    // r2
-                    "pld        [%4, #256]          \n"
-                    "vld2.f32   {d20-d23}, [%4]!    \n" // q10 q11 = r20 r21
-
-                    "vmla.f32   q8, q14, %e16[1]    \n"
-
-                    "pld        [%4, #128]          \n"
-                    "vld2.f32   {d24-d25}, [%4]     \n" // q12 = r28 x x
-
-                    "vmla.f32   q9, q10, %f16[0]    \n"
-
-                    "vmov       d26, d25            \n" // q13 = r29 x x
-
-                    "vext.32    q14, q10, q12, #1   \n" // q14 = r22
-
-                    "vmla.f32   q8, q11, %f16[1]    \n"
-
-                    "vext.32    q15, q11, q13, #1   \n" // q15 = r23
-
-                    "vmla.f32   q9, q14, %e17[0]    \n"
-
-                    "vext.32    q14, q10, q12, #2   \n" // q14 = r24
-
-                    "vmla.f32   q8, q15, %e17[1]    \n"
-
-                    // r3
-                    "pld        [%5, #256]          \n"
-                    "vld2.f32   {d20-d23}, [%5]!    \n" // q10 q11 = r30 r31
-
-                    "vmla.f32   q9, q14, %f17[0]    \n"
-
-                    "pld        [%5, #128]          \n"
-                    "vld2.f32   {d24-d25}, [%5]     \n" // q12 = r38 x x
-
-                    "vmla.f32   q8, q10, %f17[1]    \n"
-
-                    "vmov       d26, d25            \n" // q13 = r39 x x
-
-                    "vext.32    q14, q10, q12, #1   \n" // q14 = r32
-
-                    "vmla.f32   q9, q11, %e18[0]    \n"
-
-                    "vext.32    q15, q11, q13, #1   \n" // q15 = r33
-
-                    "vmla.f32   q8, q14, %e18[1]    \n"
-
-                    "vext.32    q14, q10, q12, #2   \n" // q14 = r34
-
-                    "vmla.f32   q9, q15, %f18[0]    \n"
-
-                    // r4
-                    "pld        [%6, #256]          \n"
-                    "vld2.f32   {d20-d23}, [%6]!    \n" // q10 q11 = r40 r41
-
-                    "vmla.f32   q8, q14, %f18[1]    \n"
-
-                    "pld        [%6, #128]          \n"
-                    "vld2.f32   {d24-d25}, [%6]     \n" // q12 = r48 x x
-
-                    "vmla.f32   q9, q10, %e19[0]    \n"
-
-                    "vmov       d26, d25            \n" // q13 = r49 x x
-
-                    "vext.32    q14, q10, q12, #1   \n" // q14 = r42
-
-                    "vmla.f32   q8, q11, %e19[1]    \n"
-
-                    "vext.32    q15, q11, q13, #1   \n" // q15 = r43
-
-                    "vmla.f32   q9, q14, %f19[0]    \n"
-
-                    "vext.32    q14, q10, q12, #2   \n" // q14 = r44
-
-                    "vmla.f32   q8, q15, %f19[1]    \n"
-
-                    // r0
-                    "pld        [%2, #256]          \n"
-                    "vld2.f32   {d20-d23}, [%2]!    \n" // q10 q11 = r00 r01
-
-                    "vmla.f32   q9, q14, %e20[0]    \n"
-
-                    "pld        [%2, #128]          \n"
-                    "vld2.f32   {d24-d25}, [%2]     \n" // q12 = r08 x x
-
-                    "vadd.f32   q9, q8, q9          \n"
-
-                    "vmov       q8, %q21            \n"
-
-                    "subs       %0, #1              \n"
-
-                    "vst1.f32   {d18-d19}, [%1]!    \n"
-
-                    "bne        0b                  \n"
-                    "sub        %2, #32             \n"
-
-                    : "=r"(nn),     // %0
-                    "=r"(outptr), // %1
-                    "=r"(r0),     // %2
-                    "=r"(r1),     // %3
-                    "=r"(r2),     // %4
-                    "=r"(r3),     // %5
-                    "=r"(r4)      // %6
-                    : "0"(nn),
-                    "1"(outptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "w"(_k0123),     // %14
-                    "w"(_k4567),     // %15
-                    "w"(_k891011),   // %16
-                    "w"(_k12131415), // %17
-                    "w"(_k16171819), // %18
-                    "w"(_k20212223), // %19
-                    "w"(_k24242424), // %20
-                    "w"(_bias0)      // %21
-                    : "cc", "memory", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-                float sum = bias0;
-#if __ARM_NEON
-                // TODO neon assembly optimize
-                float32x4_t _r0 = vld1q_f32(r0);
-                float32x4_t _sum = vmulq_f32(_r0, _k0123);
-
-                float32x4_t _r1 = vld1q_f32(r1);
-                _sum = vmlaq_f32(_sum, _r1, vld1q_f32(k1));
-
-                float32x4_t _r2 = vld1q_f32(r2);
-                _sum = vmlaq_f32(_sum, _r2, vld1q_f32(k2));
-
-                float32x4_t _r3 = vld1q_f32(r3);
-                _sum = vmlaq_f32(_sum, _r3, vld1q_f32(k3));
-
-                float32x4_t _r4 = vld1q_f32(r4);
-                _sum = vmlaq_f32(_sum, _r4, _k20212223);
-
-                sum += r0[4] * k0[4];
-                sum += r1[4] * k1[4];
-                sum += r2[4] * k2[4];
-                sum += r3[4] * k3[4];
-                sum += r4[4] * k4[4];
-
-                float32x2_t _ss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-                _ss = vpadd_f32(_ss, _ss);
-
-                sum += vget_lane_f32(_ss, 0);
-#else
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum += r0[3] * k0[3];
-                sum += r0[4] * k0[4];
-
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum += r1[3] * k1[3];
-                sum += r1[4] * k1[4];
-
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-                sum += r2[3] * k2[3];
-                sum += r2[4] * k2[4];
-
-                sum += r3[0] * k3[0];
-                sum += r3[1] * k3[1];
-                sum += r3[2] * k3[2];
-                sum += r3[3] * k3[3];
-                sum += r3[4] * k3[4];
-
-                sum += r4[0] * k4[0];
-                sum += r4[1] * k4[1];
-                sum += r4[2] * k4[2];
-                sum += r4[3] * k4[3];
-                sum += r4[4] * k4[4];
-#endif
-                *outptr = sum;
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-                r3 += 2;
-                r4 += 2;
-                outptr++;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-            r3 += tailstep;
-            r4 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5_pack4.h
deleted file mode 100644
index 6f9ab51d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5_pack4.h
+++ /dev/null
@@ -1,1737 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw5x5s1_pack4_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-#if __aarch64__
-    const int w = bottom_blob.w;
-#endif
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + g * 4) : vdupq_n_f32(0.f);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-        const float* r4 = img0.row(4);
-
-        int i = 0;
-
-#if __aarch64__
-        float* outptr1 = out.row(1);
-        const float* r5 = img0.row(5);
-
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-
-            for (; j + 3 < outw; j += 4)
-            {
-                float32x4_t _sum00 = _bias0;
-                float32x4_t _sum01 = _bias0;
-                float32x4_t _sum02 = _bias0;
-                float32x4_t _sum03 = _bias0;
-                float32x4_t _sum10 = _bias0;
-                float32x4_t _sum11 = _bias0;
-                float32x4_t _sum12 = _bias0;
-                float32x4_t _sum13 = _bias0;
-
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r02 = vld1q_f32(r0 + 8);
-                float32x4_t _r03 = vld1q_f32(r0 + 12);
-                float32x4_t _r04 = vld1q_f32(r0 + 16);
-                float32x4_t _r05 = vld1q_f32(r0 + 20);
-                float32x4_t _r06 = vld1q_f32(r0 + 24);
-                float32x4_t _r07 = vld1q_f32(r0 + 28);
-
-                float32x4_t _k00 = vld1q_f32(k0);
-                float32x4_t _k01 = vld1q_f32(k0 + 4);
-                float32x4_t _k02 = vld1q_f32(k0 + 8);
-                float32x4_t _k03 = vld1q_f32(k0 + 12);
-                float32x4_t _k04 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum00 = vmlaq_f32(_sum00, _k00, _r00);
-                _sum00 = vmlaq_f32(_sum00, _k01, _r01);
-                _sum00 = vmlaq_f32(_sum00, _k02, _r02);
-                _sum00 = vmlaq_f32(_sum00, _k03, _r03);
-                _sum00 = vmlaq_f32(_sum00, _k04, _r04);
-                _sum01 = vmlaq_f32(_sum01, _k00, _r01);
-                _sum01 = vmlaq_f32(_sum01, _k01, _r02);
-                _sum01 = vmlaq_f32(_sum01, _k02, _r03);
-                _sum01 = vmlaq_f32(_sum01, _k03, _r04);
-                _sum01 = vmlaq_f32(_sum01, _k04, _r05);
-                _sum02 = vmlaq_f32(_sum02, _k00, _r02);
-                _sum02 = vmlaq_f32(_sum02, _k01, _r03);
-                _sum02 = vmlaq_f32(_sum02, _k02, _r04);
-                _sum02 = vmlaq_f32(_sum02, _k03, _r05);
-                _sum02 = vmlaq_f32(_sum02, _k04, _r06);
-                _sum03 = vmlaq_f32(_sum03, _k00, _r03);
-                _sum03 = vmlaq_f32(_sum03, _k01, _r04);
-                _sum03 = vmlaq_f32(_sum03, _k02, _r05);
-                _sum03 = vmlaq_f32(_sum03, _k03, _r06);
-                _sum03 = vmlaq_f32(_sum03, _k04, _r07);
-
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r12 = vld1q_f32(r1 + 8);
-                float32x4_t _r13 = vld1q_f32(r1 + 12);
-                float32x4_t _r14 = vld1q_f32(r1 + 16);
-                float32x4_t _r15 = vld1q_f32(r1 + 20);
-                float32x4_t _r16 = vld1q_f32(r1 + 24);
-                float32x4_t _r17 = vld1q_f32(r1 + 28);
-
-                float32x4_t _k10 = vld1q_f32(k0);
-                float32x4_t _k11 = vld1q_f32(k0 + 4);
-                float32x4_t _k12 = vld1q_f32(k0 + 8);
-                float32x4_t _k13 = vld1q_f32(k0 + 12);
-                float32x4_t _k14 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum10 = vmlaq_f32(_sum10, _k00, _r10);
-                _sum10 = vmlaq_f32(_sum10, _k01, _r11);
-                _sum10 = vmlaq_f32(_sum10, _k02, _r12);
-                _sum10 = vmlaq_f32(_sum10, _k03, _r13);
-                _sum10 = vmlaq_f32(_sum10, _k04, _r14);
-                _sum11 = vmlaq_f32(_sum11, _k00, _r11);
-                _sum11 = vmlaq_f32(_sum11, _k01, _r12);
-                _sum11 = vmlaq_f32(_sum11, _k02, _r13);
-                _sum11 = vmlaq_f32(_sum11, _k03, _r14);
-                _sum11 = vmlaq_f32(_sum11, _k04, _r15);
-                _sum12 = vmlaq_f32(_sum12, _k00, _r12);
-                _sum12 = vmlaq_f32(_sum12, _k01, _r13);
-                _sum12 = vmlaq_f32(_sum12, _k02, _r14);
-                _sum12 = vmlaq_f32(_sum12, _k03, _r15);
-                _sum12 = vmlaq_f32(_sum12, _k04, _r16);
-                _sum13 = vmlaq_f32(_sum13, _k00, _r13);
-                _sum13 = vmlaq_f32(_sum13, _k01, _r14);
-                _sum13 = vmlaq_f32(_sum13, _k02, _r15);
-                _sum13 = vmlaq_f32(_sum13, _k03, _r16);
-                _sum13 = vmlaq_f32(_sum13, _k04, _r17);
-
-                _sum00 = vmlaq_f32(_sum00, _k10, _r10);
-                _sum00 = vmlaq_f32(_sum00, _k11, _r11);
-                _sum00 = vmlaq_f32(_sum00, _k12, _r12);
-                _sum00 = vmlaq_f32(_sum00, _k13, _r13);
-                _sum00 = vmlaq_f32(_sum00, _k14, _r14);
-                _sum01 = vmlaq_f32(_sum01, _k10, _r11);
-                _sum01 = vmlaq_f32(_sum01, _k11, _r12);
-                _sum01 = vmlaq_f32(_sum01, _k12, _r13);
-                _sum01 = vmlaq_f32(_sum01, _k13, _r14);
-                _sum01 = vmlaq_f32(_sum01, _k14, _r15);
-                _sum02 = vmlaq_f32(_sum02, _k10, _r12);
-                _sum02 = vmlaq_f32(_sum02, _k11, _r13);
-                _sum02 = vmlaq_f32(_sum02, _k12, _r14);
-                _sum02 = vmlaq_f32(_sum02, _k13, _r15);
-                _sum02 = vmlaq_f32(_sum02, _k14, _r16);
-                _sum03 = vmlaq_f32(_sum03, _k10, _r13);
-                _sum03 = vmlaq_f32(_sum03, _k11, _r14);
-                _sum03 = vmlaq_f32(_sum03, _k12, _r15);
-                _sum03 = vmlaq_f32(_sum03, _k13, _r16);
-                _sum03 = vmlaq_f32(_sum03, _k14, _r17);
-
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r22 = vld1q_f32(r2 + 8);
-                float32x4_t _r23 = vld1q_f32(r2 + 12);
-                float32x4_t _r24 = vld1q_f32(r2 + 16);
-                float32x4_t _r25 = vld1q_f32(r2 + 20);
-                float32x4_t _r26 = vld1q_f32(r2 + 24);
-                float32x4_t _r27 = vld1q_f32(r2 + 28);
-
-                float32x4_t _k20 = vld1q_f32(k0);
-                float32x4_t _k21 = vld1q_f32(k0 + 4);
-                float32x4_t _k22 = vld1q_f32(k0 + 8);
-                float32x4_t _k23 = vld1q_f32(k0 + 12);
-                float32x4_t _k24 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum10 = vmlaq_f32(_sum10, _k10, _r20);
-                _sum10 = vmlaq_f32(_sum10, _k11, _r21);
-                _sum10 = vmlaq_f32(_sum10, _k12, _r22);
-                _sum10 = vmlaq_f32(_sum10, _k13, _r23);
-                _sum10 = vmlaq_f32(_sum10, _k14, _r24);
-                _sum11 = vmlaq_f32(_sum11, _k10, _r21);
-                _sum11 = vmlaq_f32(_sum11, _k11, _r22);
-                _sum11 = vmlaq_f32(_sum11, _k12, _r23);
-                _sum11 = vmlaq_f32(_sum11, _k13, _r24);
-                _sum11 = vmlaq_f32(_sum11, _k14, _r25);
-                _sum12 = vmlaq_f32(_sum12, _k10, _r22);
-                _sum12 = vmlaq_f32(_sum12, _k11, _r23);
-                _sum12 = vmlaq_f32(_sum12, _k12, _r24);
-                _sum12 = vmlaq_f32(_sum12, _k13, _r25);
-                _sum12 = vmlaq_f32(_sum12, _k14, _r26);
-                _sum13 = vmlaq_f32(_sum13, _k10, _r23);
-                _sum13 = vmlaq_f32(_sum13, _k11, _r24);
-                _sum13 = vmlaq_f32(_sum13, _k12, _r25);
-                _sum13 = vmlaq_f32(_sum13, _k13, _r26);
-                _sum13 = vmlaq_f32(_sum13, _k14, _r27);
-
-                _sum00 = vmlaq_f32(_sum00, _k20, _r20);
-                _sum00 = vmlaq_f32(_sum00, _k21, _r21);
-                _sum00 = vmlaq_f32(_sum00, _k22, _r22);
-                _sum00 = vmlaq_f32(_sum00, _k23, _r23);
-                _sum00 = vmlaq_f32(_sum00, _k24, _r24);
-                _sum01 = vmlaq_f32(_sum01, _k20, _r21);
-                _sum01 = vmlaq_f32(_sum01, _k21, _r22);
-                _sum01 = vmlaq_f32(_sum01, _k22, _r23);
-                _sum01 = vmlaq_f32(_sum01, _k23, _r24);
-                _sum01 = vmlaq_f32(_sum01, _k24, _r25);
-                _sum02 = vmlaq_f32(_sum02, _k20, _r22);
-                _sum02 = vmlaq_f32(_sum02, _k21, _r23);
-                _sum02 = vmlaq_f32(_sum02, _k22, _r24);
-                _sum02 = vmlaq_f32(_sum02, _k23, _r25);
-                _sum02 = vmlaq_f32(_sum02, _k24, _r26);
-                _sum03 = vmlaq_f32(_sum03, _k20, _r23);
-                _sum03 = vmlaq_f32(_sum03, _k21, _r24);
-                _sum03 = vmlaq_f32(_sum03, _k22, _r25);
-                _sum03 = vmlaq_f32(_sum03, _k23, _r26);
-                _sum03 = vmlaq_f32(_sum03, _k24, _r27);
-
-                float32x4_t _r30 = vld1q_f32(r3);
-                float32x4_t _r31 = vld1q_f32(r3 + 4);
-                float32x4_t _r32 = vld1q_f32(r3 + 8);
-                float32x4_t _r33 = vld1q_f32(r3 + 12);
-                float32x4_t _r34 = vld1q_f32(r3 + 16);
-                float32x4_t _r35 = vld1q_f32(r3 + 20);
-                float32x4_t _r36 = vld1q_f32(r3 + 24);
-                float32x4_t _r37 = vld1q_f32(r3 + 28);
-
-                float32x4_t _k30 = vld1q_f32(k0);
-                float32x4_t _k31 = vld1q_f32(k0 + 4);
-                float32x4_t _k32 = vld1q_f32(k0 + 8);
-                float32x4_t _k33 = vld1q_f32(k0 + 12);
-                float32x4_t _k34 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum10 = vmlaq_f32(_sum10, _k20, _r30);
-                _sum10 = vmlaq_f32(_sum10, _k21, _r31);
-                _sum10 = vmlaq_f32(_sum10, _k22, _r32);
-                _sum10 = vmlaq_f32(_sum10, _k23, _r33);
-                _sum10 = vmlaq_f32(_sum10, _k24, _r34);
-                _sum11 = vmlaq_f32(_sum11, _k20, _r31);
-                _sum11 = vmlaq_f32(_sum11, _k21, _r32);
-                _sum11 = vmlaq_f32(_sum11, _k22, _r33);
-                _sum11 = vmlaq_f32(_sum11, _k23, _r34);
-                _sum11 = vmlaq_f32(_sum11, _k24, _r35);
-                _sum12 = vmlaq_f32(_sum12, _k20, _r32);
-                _sum12 = vmlaq_f32(_sum12, _k21, _r33);
-                _sum12 = vmlaq_f32(_sum12, _k22, _r34);
-                _sum12 = vmlaq_f32(_sum12, _k23, _r35);
-                _sum12 = vmlaq_f32(_sum12, _k24, _r36);
-                _sum13 = vmlaq_f32(_sum13, _k20, _r33);
-                _sum13 = vmlaq_f32(_sum13, _k21, _r34);
-                _sum13 = vmlaq_f32(_sum13, _k22, _r35);
-                _sum13 = vmlaq_f32(_sum13, _k23, _r36);
-                _sum13 = vmlaq_f32(_sum13, _k24, _r37);
-
-                _sum00 = vmlaq_f32(_sum00, _k30, _r30);
-                _sum00 = vmlaq_f32(_sum00, _k31, _r31);
-                _sum00 = vmlaq_f32(_sum00, _k32, _r32);
-                _sum00 = vmlaq_f32(_sum00, _k33, _r33);
-                _sum00 = vmlaq_f32(_sum00, _k34, _r34);
-                _sum01 = vmlaq_f32(_sum01, _k30, _r31);
-                _sum01 = vmlaq_f32(_sum01, _k31, _r32);
-                _sum01 = vmlaq_f32(_sum01, _k32, _r33);
-                _sum01 = vmlaq_f32(_sum01, _k33, _r34);
-                _sum01 = vmlaq_f32(_sum01, _k34, _r35);
-                _sum02 = vmlaq_f32(_sum02, _k30, _r32);
-                _sum02 = vmlaq_f32(_sum02, _k31, _r33);
-                _sum02 = vmlaq_f32(_sum02, _k32, _r34);
-                _sum02 = vmlaq_f32(_sum02, _k33, _r35);
-                _sum02 = vmlaq_f32(_sum02, _k34, _r36);
-                _sum03 = vmlaq_f32(_sum03, _k30, _r33);
-                _sum03 = vmlaq_f32(_sum03, _k31, _r34);
-                _sum03 = vmlaq_f32(_sum03, _k32, _r35);
-                _sum03 = vmlaq_f32(_sum03, _k33, _r36);
-                _sum03 = vmlaq_f32(_sum03, _k34, _r37);
-
-                float32x4_t _r40 = vld1q_f32(r4);
-                float32x4_t _r41 = vld1q_f32(r4 + 4);
-                float32x4_t _r42 = vld1q_f32(r4 + 8);
-                float32x4_t _r43 = vld1q_f32(r4 + 12);
-                float32x4_t _r44 = vld1q_f32(r4 + 16);
-                float32x4_t _r45 = vld1q_f32(r4 + 20);
-                float32x4_t _r46 = vld1q_f32(r4 + 24);
-                float32x4_t _r47 = vld1q_f32(r4 + 28);
-
-                float32x4_t _k40 = vld1q_f32(k0);
-                float32x4_t _k41 = vld1q_f32(k0 + 4);
-                float32x4_t _k42 = vld1q_f32(k0 + 8);
-                float32x4_t _k43 = vld1q_f32(k0 + 12);
-                float32x4_t _k44 = vld1q_f32(k0 + 16);
-                k0 -= 80;
-
-                _sum10 = vmlaq_f32(_sum10, _k30, _r40);
-                _sum10 = vmlaq_f32(_sum10, _k31, _r41);
-                _sum10 = vmlaq_f32(_sum10, _k32, _r42);
-                _sum10 = vmlaq_f32(_sum10, _k33, _r43);
-                _sum10 = vmlaq_f32(_sum10, _k34, _r44);
-                _sum11 = vmlaq_f32(_sum11, _k30, _r41);
-                _sum11 = vmlaq_f32(_sum11, _k31, _r42);
-                _sum11 = vmlaq_f32(_sum11, _k32, _r43);
-                _sum11 = vmlaq_f32(_sum11, _k33, _r44);
-                _sum11 = vmlaq_f32(_sum11, _k34, _r45);
-                _sum12 = vmlaq_f32(_sum12, _k30, _r42);
-                _sum12 = vmlaq_f32(_sum12, _k31, _r43);
-                _sum12 = vmlaq_f32(_sum12, _k32, _r44);
-                _sum12 = vmlaq_f32(_sum12, _k33, _r45);
-                _sum12 = vmlaq_f32(_sum12, _k34, _r46);
-                _sum13 = vmlaq_f32(_sum13, _k30, _r43);
-                _sum13 = vmlaq_f32(_sum13, _k31, _r44);
-                _sum13 = vmlaq_f32(_sum13, _k32, _r45);
-                _sum13 = vmlaq_f32(_sum13, _k33, _r46);
-                _sum13 = vmlaq_f32(_sum13, _k34, _r47);
-
-                _sum00 = vmlaq_f32(_sum00, _k40, _r40);
-                _sum00 = vmlaq_f32(_sum00, _k41, _r41);
-                _sum00 = vmlaq_f32(_sum00, _k42, _r42);
-                _sum00 = vmlaq_f32(_sum00, _k43, _r43);
-                _sum00 = vmlaq_f32(_sum00, _k44, _r44);
-                _sum01 = vmlaq_f32(_sum01, _k40, _r41);
-                _sum01 = vmlaq_f32(_sum01, _k41, _r42);
-                _sum01 = vmlaq_f32(_sum01, _k42, _r43);
-                _sum01 = vmlaq_f32(_sum01, _k43, _r44);
-                _sum01 = vmlaq_f32(_sum01, _k44, _r45);
-                _sum02 = vmlaq_f32(_sum02, _k40, _r42);
-                _sum02 = vmlaq_f32(_sum02, _k41, _r43);
-                _sum02 = vmlaq_f32(_sum02, _k42, _r44);
-                _sum02 = vmlaq_f32(_sum02, _k43, _r45);
-                _sum02 = vmlaq_f32(_sum02, _k44, _r46);
-                _sum03 = vmlaq_f32(_sum03, _k40, _r43);
-                _sum03 = vmlaq_f32(_sum03, _k41, _r44);
-                _sum03 = vmlaq_f32(_sum03, _k42, _r45);
-                _sum03 = vmlaq_f32(_sum03, _k43, _r46);
-                _sum03 = vmlaq_f32(_sum03, _k44, _r47);
-
-                float32x4_t _r50 = vld1q_f32(r5);
-                float32x4_t _r51 = vld1q_f32(r5 + 4);
-                float32x4_t _r52 = vld1q_f32(r5 + 8);
-                float32x4_t _r53 = vld1q_f32(r5 + 12);
-                float32x4_t _r54 = vld1q_f32(r5 + 16);
-                float32x4_t _r55 = vld1q_f32(r5 + 20);
-                float32x4_t _r56 = vld1q_f32(r5 + 24);
-                float32x4_t _r57 = vld1q_f32(r5 + 28);
-
-                _sum10 = vmlaq_f32(_sum10, _k40, _r50);
-                _sum10 = vmlaq_f32(_sum10, _k41, _r51);
-                _sum10 = vmlaq_f32(_sum10, _k42, _r52);
-                _sum10 = vmlaq_f32(_sum10, _k43, _r53);
-                _sum10 = vmlaq_f32(_sum10, _k44, _r54);
-                _sum11 = vmlaq_f32(_sum11, _k40, _r51);
-                _sum11 = vmlaq_f32(_sum11, _k41, _r52);
-                _sum11 = vmlaq_f32(_sum11, _k42, _r53);
-                _sum11 = vmlaq_f32(_sum11, _k43, _r54);
-                _sum11 = vmlaq_f32(_sum11, _k44, _r55);
-                _sum12 = vmlaq_f32(_sum12, _k40, _r52);
-                _sum12 = vmlaq_f32(_sum12, _k41, _r53);
-                _sum12 = vmlaq_f32(_sum12, _k42, _r54);
-                _sum12 = vmlaq_f32(_sum12, _k43, _r55);
-                _sum12 = vmlaq_f32(_sum12, _k44, _r56);
-                _sum13 = vmlaq_f32(_sum13, _k40, _r53);
-                _sum13 = vmlaq_f32(_sum13, _k41, _r54);
-                _sum13 = vmlaq_f32(_sum13, _k42, _r55);
-                _sum13 = vmlaq_f32(_sum13, _k43, _r56);
-                _sum13 = vmlaq_f32(_sum13, _k44, _r57);
-
-                vst1q_f32(outptr0, _sum00);
-                vst1q_f32(outptr0 + 4, _sum01);
-                vst1q_f32(outptr0 + 8, _sum02);
-                vst1q_f32(outptr0 + 12, _sum03);
-                vst1q_f32(outptr1, _sum10);
-                vst1q_f32(outptr1 + 4, _sum11);
-                vst1q_f32(outptr1 + 8, _sum12);
-                vst1q_f32(outptr1 + 12, _sum13);
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                r3 += 16;
-                r4 += 16;
-                r5 += 16;
-                outptr0 += 16;
-                outptr1 += 16;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                float32x4_t _sum00 = _bias0;
-                float32x4_t _sum01 = _bias0;
-                float32x4_t _sum10 = _bias0;
-                float32x4_t _sum11 = _bias0;
-
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r02 = vld1q_f32(r0 + 8);
-                float32x4_t _r03 = vld1q_f32(r0 + 12);
-                float32x4_t _r04 = vld1q_f32(r0 + 16);
-                float32x4_t _r05 = vld1q_f32(r0 + 20);
-
-                float32x4_t _k00 = vld1q_f32(k0);
-                float32x4_t _k01 = vld1q_f32(k0 + 4);
-                float32x4_t _k02 = vld1q_f32(k0 + 8);
-                float32x4_t _k03 = vld1q_f32(k0 + 12);
-                float32x4_t _k04 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum00 = vmlaq_f32(_sum00, _k00, _r00);
-                _sum00 = vmlaq_f32(_sum00, _k01, _r01);
-                _sum00 = vmlaq_f32(_sum00, _k02, _r02);
-                _sum00 = vmlaq_f32(_sum00, _k03, _r03);
-                _sum00 = vmlaq_f32(_sum00, _k04, _r04);
-                _sum01 = vmlaq_f32(_sum01, _k00, _r01);
-                _sum01 = vmlaq_f32(_sum01, _k01, _r02);
-                _sum01 = vmlaq_f32(_sum01, _k02, _r03);
-                _sum01 = vmlaq_f32(_sum01, _k03, _r04);
-                _sum01 = vmlaq_f32(_sum01, _k04, _r05);
-
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r12 = vld1q_f32(r1 + 8);
-                float32x4_t _r13 = vld1q_f32(r1 + 12);
-                float32x4_t _r14 = vld1q_f32(r1 + 16);
-                float32x4_t _r15 = vld1q_f32(r1 + 20);
-
-                float32x4_t _k10 = vld1q_f32(k0);
-                float32x4_t _k11 = vld1q_f32(k0 + 4);
-                float32x4_t _k12 = vld1q_f32(k0 + 8);
-                float32x4_t _k13 = vld1q_f32(k0 + 12);
-                float32x4_t _k14 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum10 = vmlaq_f32(_sum10, _k00, _r10);
-                _sum10 = vmlaq_f32(_sum10, _k01, _r11);
-                _sum10 = vmlaq_f32(_sum10, _k02, _r12);
-                _sum10 = vmlaq_f32(_sum10, _k03, _r13);
-                _sum10 = vmlaq_f32(_sum10, _k04, _r14);
-                _sum11 = vmlaq_f32(_sum11, _k00, _r11);
-                _sum11 = vmlaq_f32(_sum11, _k01, _r12);
-                _sum11 = vmlaq_f32(_sum11, _k02, _r13);
-                _sum11 = vmlaq_f32(_sum11, _k03, _r14);
-                _sum11 = vmlaq_f32(_sum11, _k04, _r15);
-
-                _sum00 = vmlaq_f32(_sum00, _k10, _r10);
-                _sum00 = vmlaq_f32(_sum00, _k11, _r11);
-                _sum00 = vmlaq_f32(_sum00, _k12, _r12);
-                _sum00 = vmlaq_f32(_sum00, _k13, _r13);
-                _sum00 = vmlaq_f32(_sum00, _k14, _r14);
-                _sum01 = vmlaq_f32(_sum01, _k10, _r11);
-                _sum01 = vmlaq_f32(_sum01, _k11, _r12);
-                _sum01 = vmlaq_f32(_sum01, _k12, _r13);
-                _sum01 = vmlaq_f32(_sum01, _k13, _r14);
-                _sum01 = vmlaq_f32(_sum01, _k14, _r15);
-
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r22 = vld1q_f32(r2 + 8);
-                float32x4_t _r23 = vld1q_f32(r2 + 12);
-                float32x4_t _r24 = vld1q_f32(r2 + 16);
-                float32x4_t _r25 = vld1q_f32(r2 + 20);
-
-                float32x4_t _k20 = vld1q_f32(k0);
-                float32x4_t _k21 = vld1q_f32(k0 + 4);
-                float32x4_t _k22 = vld1q_f32(k0 + 8);
-                float32x4_t _k23 = vld1q_f32(k0 + 12);
-                float32x4_t _k24 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum10 = vmlaq_f32(_sum10, _k10, _r20);
-                _sum10 = vmlaq_f32(_sum10, _k11, _r21);
-                _sum10 = vmlaq_f32(_sum10, _k12, _r22);
-                _sum10 = vmlaq_f32(_sum10, _k13, _r23);
-                _sum10 = vmlaq_f32(_sum10, _k14, _r24);
-                _sum11 = vmlaq_f32(_sum11, _k10, _r21);
-                _sum11 = vmlaq_f32(_sum11, _k11, _r22);
-                _sum11 = vmlaq_f32(_sum11, _k12, _r23);
-                _sum11 = vmlaq_f32(_sum11, _k13, _r24);
-                _sum11 = vmlaq_f32(_sum11, _k14, _r25);
-
-                _sum00 = vmlaq_f32(_sum00, _k20, _r20);
-                _sum00 = vmlaq_f32(_sum00, _k21, _r21);
-                _sum00 = vmlaq_f32(_sum00, _k22, _r22);
-                _sum00 = vmlaq_f32(_sum00, _k23, _r23);
-                _sum00 = vmlaq_f32(_sum00, _k24, _r24);
-                _sum01 = vmlaq_f32(_sum01, _k20, _r21);
-                _sum01 = vmlaq_f32(_sum01, _k21, _r22);
-                _sum01 = vmlaq_f32(_sum01, _k22, _r23);
-                _sum01 = vmlaq_f32(_sum01, _k23, _r24);
-                _sum01 = vmlaq_f32(_sum01, _k24, _r25);
-
-                float32x4_t _r30 = vld1q_f32(r3);
-                float32x4_t _r31 = vld1q_f32(r3 + 4);
-                float32x4_t _r32 = vld1q_f32(r3 + 8);
-                float32x4_t _r33 = vld1q_f32(r3 + 12);
-                float32x4_t _r34 = vld1q_f32(r3 + 16);
-                float32x4_t _r35 = vld1q_f32(r3 + 20);
-
-                float32x4_t _k30 = vld1q_f32(k0);
-                float32x4_t _k31 = vld1q_f32(k0 + 4);
-                float32x4_t _k32 = vld1q_f32(k0 + 8);
-                float32x4_t _k33 = vld1q_f32(k0 + 12);
-                float32x4_t _k34 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum10 = vmlaq_f32(_sum10, _k20, _r30);
-                _sum10 = vmlaq_f32(_sum10, _k21, _r31);
-                _sum10 = vmlaq_f32(_sum10, _k22, _r32);
-                _sum10 = vmlaq_f32(_sum10, _k23, _r33);
-                _sum10 = vmlaq_f32(_sum10, _k24, _r34);
-                _sum11 = vmlaq_f32(_sum11, _k20, _r31);
-                _sum11 = vmlaq_f32(_sum11, _k21, _r32);
-                _sum11 = vmlaq_f32(_sum11, _k22, _r33);
-                _sum11 = vmlaq_f32(_sum11, _k23, _r34);
-                _sum11 = vmlaq_f32(_sum11, _k24, _r35);
-
-                _sum00 = vmlaq_f32(_sum00, _k30, _r30);
-                _sum00 = vmlaq_f32(_sum00, _k31, _r31);
-                _sum00 = vmlaq_f32(_sum00, _k32, _r32);
-                _sum00 = vmlaq_f32(_sum00, _k33, _r33);
-                _sum00 = vmlaq_f32(_sum00, _k34, _r34);
-                _sum01 = vmlaq_f32(_sum01, _k30, _r31);
-                _sum01 = vmlaq_f32(_sum01, _k31, _r32);
-                _sum01 = vmlaq_f32(_sum01, _k32, _r33);
-                _sum01 = vmlaq_f32(_sum01, _k33, _r34);
-                _sum01 = vmlaq_f32(_sum01, _k34, _r35);
-
-                float32x4_t _r40 = vld1q_f32(r4);
-                float32x4_t _r41 = vld1q_f32(r4 + 4);
-                float32x4_t _r42 = vld1q_f32(r4 + 8);
-                float32x4_t _r43 = vld1q_f32(r4 + 12);
-                float32x4_t _r44 = vld1q_f32(r4 + 16);
-                float32x4_t _r45 = vld1q_f32(r4 + 20);
-
-                float32x4_t _k40 = vld1q_f32(k0);
-                float32x4_t _k41 = vld1q_f32(k0 + 4);
-                float32x4_t _k42 = vld1q_f32(k0 + 8);
-                float32x4_t _k43 = vld1q_f32(k0 + 12);
-                float32x4_t _k44 = vld1q_f32(k0 + 16);
-                k0 -= 80;
-
-                _sum10 = vmlaq_f32(_sum10, _k30, _r40);
-                _sum10 = vmlaq_f32(_sum10, _k31, _r41);
-                _sum10 = vmlaq_f32(_sum10, _k32, _r42);
-                _sum10 = vmlaq_f32(_sum10, _k33, _r43);
-                _sum10 = vmlaq_f32(_sum10, _k34, _r44);
-                _sum11 = vmlaq_f32(_sum11, _k30, _r41);
-                _sum11 = vmlaq_f32(_sum11, _k31, _r42);
-                _sum11 = vmlaq_f32(_sum11, _k32, _r43);
-                _sum11 = vmlaq_f32(_sum11, _k33, _r44);
-                _sum11 = vmlaq_f32(_sum11, _k34, _r45);
-
-                _sum00 = vmlaq_f32(_sum00, _k40, _r40);
-                _sum00 = vmlaq_f32(_sum00, _k41, _r41);
-                _sum00 = vmlaq_f32(_sum00, _k42, _r42);
-                _sum00 = vmlaq_f32(_sum00, _k43, _r43);
-                _sum00 = vmlaq_f32(_sum00, _k44, _r44);
-                _sum01 = vmlaq_f32(_sum01, _k40, _r41);
-                _sum01 = vmlaq_f32(_sum01, _k41, _r42);
-                _sum01 = vmlaq_f32(_sum01, _k42, _r43);
-                _sum01 = vmlaq_f32(_sum01, _k43, _r44);
-                _sum01 = vmlaq_f32(_sum01, _k44, _r45);
-
-                float32x4_t _r50 = vld1q_f32(r5);
-                float32x4_t _r51 = vld1q_f32(r5 + 4);
-                float32x4_t _r52 = vld1q_f32(r5 + 8);
-                float32x4_t _r53 = vld1q_f32(r5 + 12);
-                float32x4_t _r54 = vld1q_f32(r5 + 16);
-                float32x4_t _r55 = vld1q_f32(r5 + 20);
-
-                _sum10 = vmlaq_f32(_sum10, _k40, _r50);
-                _sum10 = vmlaq_f32(_sum10, _k41, _r51);
-                _sum10 = vmlaq_f32(_sum10, _k42, _r52);
-                _sum10 = vmlaq_f32(_sum10, _k43, _r53);
-                _sum10 = vmlaq_f32(_sum10, _k44, _r54);
-                _sum11 = vmlaq_f32(_sum11, _k40, _r51);
-                _sum11 = vmlaq_f32(_sum11, _k41, _r52);
-                _sum11 = vmlaq_f32(_sum11, _k42, _r53);
-                _sum11 = vmlaq_f32(_sum11, _k43, _r54);
-                _sum11 = vmlaq_f32(_sum11, _k44, _r55);
-
-                vst1q_f32(outptr0, _sum00);
-                vst1q_f32(outptr0 + 4, _sum01);
-                vst1q_f32(outptr1, _sum10);
-                vst1q_f32(outptr1 + 4, _sum11);
-
-                r0 += 8;
-                r1 += 8;
-                r2 += 8;
-                r3 += 8;
-                r4 += 8;
-                r5 += 8;
-                outptr0 += 8;
-                outptr1 += 8;
-            }
-            for (; j < outw; j++)
-            {
-                float32x4_t _sum0 = _bias0;
-                float32x4_t _sum1 = _bias0;
-
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r02 = vld1q_f32(r0 + 8);
-                float32x4_t _r03 = vld1q_f32(r0 + 12);
-                float32x4_t _r04 = vld1q_f32(r0 + 16);
-
-                float32x4_t _k00 = vld1q_f32(k0);
-                float32x4_t _k01 = vld1q_f32(k0 + 4);
-                float32x4_t _k02 = vld1q_f32(k0 + 8);
-                float32x4_t _k03 = vld1q_f32(k0 + 12);
-                float32x4_t _k04 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k00, _r00);
-                _sum0 = vmlaq_f32(_sum0, _k01, _r01);
-                _sum0 = vmlaq_f32(_sum0, _k02, _r02);
-                _sum0 = vmlaq_f32(_sum0, _k03, _r03);
-                _sum0 = vmlaq_f32(_sum0, _k04, _r04);
-
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r12 = vld1q_f32(r1 + 8);
-                float32x4_t _r13 = vld1q_f32(r1 + 12);
-                float32x4_t _r14 = vld1q_f32(r1 + 16);
-
-                float32x4_t _k10 = vld1q_f32(k0);
-                float32x4_t _k11 = vld1q_f32(k0 + 4);
-                float32x4_t _k12 = vld1q_f32(k0 + 8);
-                float32x4_t _k13 = vld1q_f32(k0 + 12);
-                float32x4_t _k14 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum1 = vmlaq_f32(_sum1, _k00, _r10);
-                _sum1 = vmlaq_f32(_sum1, _k01, _r11);
-                _sum1 = vmlaq_f32(_sum1, _k02, _r12);
-                _sum1 = vmlaq_f32(_sum1, _k03, _r13);
-                _sum1 = vmlaq_f32(_sum1, _k04, _r14);
-
-                _sum0 = vmlaq_f32(_sum0, _k10, _r10);
-                _sum0 = vmlaq_f32(_sum0, _k11, _r11);
-                _sum0 = vmlaq_f32(_sum0, _k12, _r12);
-                _sum0 = vmlaq_f32(_sum0, _k13, _r13);
-                _sum0 = vmlaq_f32(_sum0, _k14, _r14);
-
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r22 = vld1q_f32(r2 + 8);
-                float32x4_t _r23 = vld1q_f32(r2 + 12);
-                float32x4_t _r24 = vld1q_f32(r2 + 16);
-
-                float32x4_t _k20 = vld1q_f32(k0);
-                float32x4_t _k21 = vld1q_f32(k0 + 4);
-                float32x4_t _k22 = vld1q_f32(k0 + 8);
-                float32x4_t _k23 = vld1q_f32(k0 + 12);
-                float32x4_t _k24 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum1 = vmlaq_f32(_sum1, _k10, _r20);
-                _sum1 = vmlaq_f32(_sum1, _k11, _r21);
-                _sum1 = vmlaq_f32(_sum1, _k12, _r22);
-                _sum1 = vmlaq_f32(_sum1, _k13, _r23);
-                _sum1 = vmlaq_f32(_sum1, _k14, _r24);
-
-                _sum0 = vmlaq_f32(_sum0, _k20, _r20);
-                _sum0 = vmlaq_f32(_sum0, _k21, _r21);
-                _sum0 = vmlaq_f32(_sum0, _k22, _r22);
-                _sum0 = vmlaq_f32(_sum0, _k23, _r23);
-                _sum0 = vmlaq_f32(_sum0, _k24, _r24);
-
-                float32x4_t _r30 = vld1q_f32(r3);
-                float32x4_t _r31 = vld1q_f32(r3 + 4);
-                float32x4_t _r32 = vld1q_f32(r3 + 8);
-                float32x4_t _r33 = vld1q_f32(r3 + 12);
-                float32x4_t _r34 = vld1q_f32(r3 + 16);
-
-                float32x4_t _k30 = vld1q_f32(k0);
-                float32x4_t _k31 = vld1q_f32(k0 + 4);
-                float32x4_t _k32 = vld1q_f32(k0 + 8);
-                float32x4_t _k33 = vld1q_f32(k0 + 12);
-                float32x4_t _k34 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum1 = vmlaq_f32(_sum1, _k20, _r30);
-                _sum1 = vmlaq_f32(_sum1, _k21, _r31);
-                _sum1 = vmlaq_f32(_sum1, _k22, _r32);
-                _sum1 = vmlaq_f32(_sum1, _k23, _r33);
-                _sum1 = vmlaq_f32(_sum1, _k24, _r34);
-
-                _sum0 = vmlaq_f32(_sum0, _k30, _r30);
-                _sum0 = vmlaq_f32(_sum0, _k31, _r31);
-                _sum0 = vmlaq_f32(_sum0, _k32, _r32);
-                _sum0 = vmlaq_f32(_sum0, _k33, _r33);
-                _sum0 = vmlaq_f32(_sum0, _k34, _r34);
-
-                float32x4_t _r40 = vld1q_f32(r4);
-                float32x4_t _r41 = vld1q_f32(r4 + 4);
-                float32x4_t _r42 = vld1q_f32(r4 + 8);
-                float32x4_t _r43 = vld1q_f32(r4 + 12);
-                float32x4_t _r44 = vld1q_f32(r4 + 16);
-
-                float32x4_t _k40 = vld1q_f32(k0);
-                float32x4_t _k41 = vld1q_f32(k0 + 4);
-                float32x4_t _k42 = vld1q_f32(k0 + 8);
-                float32x4_t _k43 = vld1q_f32(k0 + 12);
-                float32x4_t _k44 = vld1q_f32(k0 + 16);
-                k0 -= 80;
-
-                _sum1 = vmlaq_f32(_sum1, _k30, _r40);
-                _sum1 = vmlaq_f32(_sum1, _k31, _r41);
-                _sum1 = vmlaq_f32(_sum1, _k32, _r42);
-                _sum1 = vmlaq_f32(_sum1, _k33, _r43);
-                _sum1 = vmlaq_f32(_sum1, _k34, _r44);
-
-                _sum0 = vmlaq_f32(_sum0, _k40, _r40);
-                _sum0 = vmlaq_f32(_sum0, _k41, _r41);
-                _sum0 = vmlaq_f32(_sum0, _k42, _r42);
-                _sum0 = vmlaq_f32(_sum0, _k43, _r43);
-                _sum0 = vmlaq_f32(_sum0, _k44, _r44);
-
-                float32x4_t _r50 = vld1q_f32(r5);
-                float32x4_t _r51 = vld1q_f32(r5 + 4);
-                float32x4_t _r52 = vld1q_f32(r5 + 8);
-                float32x4_t _r53 = vld1q_f32(r5 + 12);
-                float32x4_t _r54 = vld1q_f32(r5 + 16);
-
-                _sum1 = vmlaq_f32(_sum1, _k40, _r50);
-                _sum1 = vmlaq_f32(_sum1, _k41, _r51);
-                _sum1 = vmlaq_f32(_sum1, _k42, _r52);
-                _sum1 = vmlaq_f32(_sum1, _k43, _r53);
-                _sum1 = vmlaq_f32(_sum1, _k44, _r54);
-
-                vst1q_f32(outptr0, _sum0);
-                vst1q_f32(outptr1, _sum1);
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                r3 += 4;
-                r4 += 4;
-                r5 += 4;
-                outptr0 += 4;
-                outptr1 += 4;
-            }
-
-            r0 += 4 * 4 + w * 4;
-            r1 += 4 * 4 + w * 4;
-            r2 += 4 * 4 + w * 4;
-            r3 += 4 * 4 + w * 4;
-            r4 += 4 * 4 + w * 4;
-            r5 += 4 * 4 + w * 4;
-
-            outptr0 += outw * 4;
-            outptr1 += outw * 4;
-        }
-#endif // __aarch64__
-        for (; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j + 3 < outw; j += 4)
-            {
-                float32x4_t _sum0 = _bias0;
-                float32x4_t _sum1 = _bias0;
-                float32x4_t _sum2 = _bias0;
-                float32x4_t _sum3 = _bias0;
-
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r02 = vld1q_f32(r0 + 8);
-                float32x4_t _r03 = vld1q_f32(r0 + 12);
-                float32x4_t _r04 = vld1q_f32(r0 + 16);
-                float32x4_t _r05 = vld1q_f32(r0 + 20);
-                float32x4_t _r06 = vld1q_f32(r0 + 24);
-                float32x4_t _r07 = vld1q_f32(r0 + 28);
-
-                float32x4_t _k00 = vld1q_f32(k0);
-                float32x4_t _k01 = vld1q_f32(k0 + 4);
-                float32x4_t _k02 = vld1q_f32(k0 + 8);
-                float32x4_t _k03 = vld1q_f32(k0 + 12);
-                float32x4_t _k04 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k00, _r00);
-                _sum0 = vmlaq_f32(_sum0, _k01, _r01);
-                _sum0 = vmlaq_f32(_sum0, _k02, _r02);
-                _sum0 = vmlaq_f32(_sum0, _k03, _r03);
-                _sum0 = vmlaq_f32(_sum0, _k04, _r04);
-                _sum1 = vmlaq_f32(_sum1, _k00, _r01);
-                _sum1 = vmlaq_f32(_sum1, _k01, _r02);
-                _sum1 = vmlaq_f32(_sum1, _k02, _r03);
-                _sum1 = vmlaq_f32(_sum1, _k03, _r04);
-                _sum1 = vmlaq_f32(_sum1, _k04, _r05);
-                _sum2 = vmlaq_f32(_sum2, _k00, _r02);
-                _sum2 = vmlaq_f32(_sum2, _k01, _r03);
-                _sum2 = vmlaq_f32(_sum2, _k02, _r04);
-                _sum2 = vmlaq_f32(_sum2, _k03, _r05);
-                _sum2 = vmlaq_f32(_sum2, _k04, _r06);
-                _sum3 = vmlaq_f32(_sum3, _k00, _r03);
-                _sum3 = vmlaq_f32(_sum3, _k01, _r04);
-                _sum3 = vmlaq_f32(_sum3, _k02, _r05);
-                _sum3 = vmlaq_f32(_sum3, _k03, _r06);
-                _sum3 = vmlaq_f32(_sum3, _k04, _r07);
-
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r12 = vld1q_f32(r1 + 8);
-                float32x4_t _r13 = vld1q_f32(r1 + 12);
-                float32x4_t _r14 = vld1q_f32(r1 + 16);
-                float32x4_t _r15 = vld1q_f32(r1 + 20);
-                float32x4_t _r16 = vld1q_f32(r1 + 24);
-                float32x4_t _r17 = vld1q_f32(r1 + 28);
-
-                float32x4_t _k10 = vld1q_f32(k0);
-                float32x4_t _k11 = vld1q_f32(k0 + 4);
-                float32x4_t _k12 = vld1q_f32(k0 + 8);
-                float32x4_t _k13 = vld1q_f32(k0 + 12);
-                float32x4_t _k14 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k10, _r10);
-                _sum0 = vmlaq_f32(_sum0, _k11, _r11);
-                _sum0 = vmlaq_f32(_sum0, _k12, _r12);
-                _sum0 = vmlaq_f32(_sum0, _k13, _r13);
-                _sum0 = vmlaq_f32(_sum0, _k14, _r14);
-                _sum1 = vmlaq_f32(_sum1, _k10, _r11);
-                _sum1 = vmlaq_f32(_sum1, _k11, _r12);
-                _sum1 = vmlaq_f32(_sum1, _k12, _r13);
-                _sum1 = vmlaq_f32(_sum1, _k13, _r14);
-                _sum1 = vmlaq_f32(_sum1, _k14, _r15);
-                _sum2 = vmlaq_f32(_sum2, _k10, _r12);
-                _sum2 = vmlaq_f32(_sum2, _k11, _r13);
-                _sum2 = vmlaq_f32(_sum2, _k12, _r14);
-                _sum2 = vmlaq_f32(_sum2, _k13, _r15);
-                _sum2 = vmlaq_f32(_sum2, _k14, _r16);
-                _sum3 = vmlaq_f32(_sum3, _k10, _r13);
-                _sum3 = vmlaq_f32(_sum3, _k11, _r14);
-                _sum3 = vmlaq_f32(_sum3, _k12, _r15);
-                _sum3 = vmlaq_f32(_sum3, _k13, _r16);
-                _sum3 = vmlaq_f32(_sum3, _k14, _r17);
-
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r22 = vld1q_f32(r2 + 8);
-                float32x4_t _r23 = vld1q_f32(r2 + 12);
-                float32x4_t _r24 = vld1q_f32(r2 + 16);
-                float32x4_t _r25 = vld1q_f32(r2 + 20);
-                float32x4_t _r26 = vld1q_f32(r2 + 24);
-                float32x4_t _r27 = vld1q_f32(r2 + 28);
-
-                float32x4_t _k20 = vld1q_f32(k0);
-                float32x4_t _k21 = vld1q_f32(k0 + 4);
-                float32x4_t _k22 = vld1q_f32(k0 + 8);
-                float32x4_t _k23 = vld1q_f32(k0 + 12);
-                float32x4_t _k24 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k20, _r20);
-                _sum0 = vmlaq_f32(_sum0, _k21, _r21);
-                _sum0 = vmlaq_f32(_sum0, _k22, _r22);
-                _sum0 = vmlaq_f32(_sum0, _k23, _r23);
-                _sum0 = vmlaq_f32(_sum0, _k24, _r24);
-                _sum1 = vmlaq_f32(_sum1, _k20, _r21);
-                _sum1 = vmlaq_f32(_sum1, _k21, _r22);
-                _sum1 = vmlaq_f32(_sum1, _k22, _r23);
-                _sum1 = vmlaq_f32(_sum1, _k23, _r24);
-                _sum1 = vmlaq_f32(_sum1, _k24, _r25);
-                _sum2 = vmlaq_f32(_sum2, _k20, _r22);
-                _sum2 = vmlaq_f32(_sum2, _k21, _r23);
-                _sum2 = vmlaq_f32(_sum2, _k22, _r24);
-                _sum2 = vmlaq_f32(_sum2, _k23, _r25);
-                _sum2 = vmlaq_f32(_sum2, _k24, _r26);
-                _sum3 = vmlaq_f32(_sum3, _k20, _r23);
-                _sum3 = vmlaq_f32(_sum3, _k21, _r24);
-                _sum3 = vmlaq_f32(_sum3, _k22, _r25);
-                _sum3 = vmlaq_f32(_sum3, _k23, _r26);
-                _sum3 = vmlaq_f32(_sum3, _k24, _r27);
-
-                float32x4_t _r30 = vld1q_f32(r3);
-                float32x4_t _r31 = vld1q_f32(r3 + 4);
-                float32x4_t _r32 = vld1q_f32(r3 + 8);
-                float32x4_t _r33 = vld1q_f32(r3 + 12);
-                float32x4_t _r34 = vld1q_f32(r3 + 16);
-                float32x4_t _r35 = vld1q_f32(r3 + 20);
-                float32x4_t _r36 = vld1q_f32(r3 + 24);
-                float32x4_t _r37 = vld1q_f32(r3 + 28);
-
-                float32x4_t _k30 = vld1q_f32(k0);
-                float32x4_t _k31 = vld1q_f32(k0 + 4);
-                float32x4_t _k32 = vld1q_f32(k0 + 8);
-                float32x4_t _k33 = vld1q_f32(k0 + 12);
-                float32x4_t _k34 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k30, _r30);
-                _sum0 = vmlaq_f32(_sum0, _k31, _r31);
-                _sum0 = vmlaq_f32(_sum0, _k32, _r32);
-                _sum0 = vmlaq_f32(_sum0, _k33, _r33);
-                _sum0 = vmlaq_f32(_sum0, _k34, _r34);
-                _sum1 = vmlaq_f32(_sum1, _k30, _r31);
-                _sum1 = vmlaq_f32(_sum1, _k31, _r32);
-                _sum1 = vmlaq_f32(_sum1, _k32, _r33);
-                _sum1 = vmlaq_f32(_sum1, _k33, _r34);
-                _sum1 = vmlaq_f32(_sum1, _k34, _r35);
-                _sum2 = vmlaq_f32(_sum2, _k30, _r32);
-                _sum2 = vmlaq_f32(_sum2, _k31, _r33);
-                _sum2 = vmlaq_f32(_sum2, _k32, _r34);
-                _sum2 = vmlaq_f32(_sum2, _k33, _r35);
-                _sum2 = vmlaq_f32(_sum2, _k34, _r36);
-                _sum3 = vmlaq_f32(_sum3, _k30, _r33);
-                _sum3 = vmlaq_f32(_sum3, _k31, _r34);
-                _sum3 = vmlaq_f32(_sum3, _k32, _r35);
-                _sum3 = vmlaq_f32(_sum3, _k33, _r36);
-                _sum3 = vmlaq_f32(_sum3, _k34, _r37);
-
-                float32x4_t _r40 = vld1q_f32(r4);
-                float32x4_t _r41 = vld1q_f32(r4 + 4);
-                float32x4_t _r42 = vld1q_f32(r4 + 8);
-                float32x4_t _r43 = vld1q_f32(r4 + 12);
-                float32x4_t _r44 = vld1q_f32(r4 + 16);
-                float32x4_t _r45 = vld1q_f32(r4 + 20);
-                float32x4_t _r46 = vld1q_f32(r4 + 24);
-                float32x4_t _r47 = vld1q_f32(r4 + 28);
-
-                float32x4_t _k40 = vld1q_f32(k0);
-                float32x4_t _k41 = vld1q_f32(k0 + 4);
-                float32x4_t _k42 = vld1q_f32(k0 + 8);
-                float32x4_t _k43 = vld1q_f32(k0 + 12);
-                float32x4_t _k44 = vld1q_f32(k0 + 16);
-                k0 -= 80;
-
-                _sum0 = vmlaq_f32(_sum0, _k40, _r40);
-                _sum0 = vmlaq_f32(_sum0, _k41, _r41);
-                _sum0 = vmlaq_f32(_sum0, _k42, _r42);
-                _sum0 = vmlaq_f32(_sum0, _k43, _r43);
-                _sum0 = vmlaq_f32(_sum0, _k44, _r44);
-                _sum1 = vmlaq_f32(_sum1, _k40, _r41);
-                _sum1 = vmlaq_f32(_sum1, _k41, _r42);
-                _sum1 = vmlaq_f32(_sum1, _k42, _r43);
-                _sum1 = vmlaq_f32(_sum1, _k43, _r44);
-                _sum1 = vmlaq_f32(_sum1, _k44, _r45);
-                _sum2 = vmlaq_f32(_sum2, _k40, _r42);
-                _sum2 = vmlaq_f32(_sum2, _k41, _r43);
-                _sum2 = vmlaq_f32(_sum2, _k42, _r44);
-                _sum2 = vmlaq_f32(_sum2, _k43, _r45);
-                _sum2 = vmlaq_f32(_sum2, _k44, _r46);
-                _sum3 = vmlaq_f32(_sum3, _k40, _r43);
-                _sum3 = vmlaq_f32(_sum3, _k41, _r44);
-                _sum3 = vmlaq_f32(_sum3, _k42, _r45);
-                _sum3 = vmlaq_f32(_sum3, _k43, _r46);
-                _sum3 = vmlaq_f32(_sum3, _k44, _r47);
-
-                vst1q_f32(outptr0, _sum0);
-                vst1q_f32(outptr0 + 4, _sum1);
-                vst1q_f32(outptr0 + 8, _sum2);
-                vst1q_f32(outptr0 + 12, _sum3);
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                r3 += 16;
-                r4 += 16;
-                outptr0 += 16;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                float32x4_t _sum0 = _bias0;
-                float32x4_t _sum1 = _bias0;
-
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r02 = vld1q_f32(r0 + 8);
-                float32x4_t _r03 = vld1q_f32(r0 + 12);
-                float32x4_t _r04 = vld1q_f32(r0 + 16);
-                float32x4_t _r05 = vld1q_f32(r0 + 20);
-
-                float32x4_t _k00 = vld1q_f32(k0);
-                float32x4_t _k01 = vld1q_f32(k0 + 4);
-                float32x4_t _k02 = vld1q_f32(k0 + 8);
-                float32x4_t _k03 = vld1q_f32(k0 + 12);
-                float32x4_t _k04 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k00, _r00);
-                _sum0 = vmlaq_f32(_sum0, _k01, _r01);
-                _sum0 = vmlaq_f32(_sum0, _k02, _r02);
-                _sum0 = vmlaq_f32(_sum0, _k03, _r03);
-                _sum0 = vmlaq_f32(_sum0, _k04, _r04);
-                _sum1 = vmlaq_f32(_sum1, _k00, _r01);
-                _sum1 = vmlaq_f32(_sum1, _k01, _r02);
-                _sum1 = vmlaq_f32(_sum1, _k02, _r03);
-                _sum1 = vmlaq_f32(_sum1, _k03, _r04);
-                _sum1 = vmlaq_f32(_sum1, _k04, _r05);
-
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r12 = vld1q_f32(r1 + 8);
-                float32x4_t _r13 = vld1q_f32(r1 + 12);
-                float32x4_t _r14 = vld1q_f32(r1 + 16);
-                float32x4_t _r15 = vld1q_f32(r1 + 20);
-
-                float32x4_t _k10 = vld1q_f32(k0);
-                float32x4_t _k11 = vld1q_f32(k0 + 4);
-                float32x4_t _k12 = vld1q_f32(k0 + 8);
-                float32x4_t _k13 = vld1q_f32(k0 + 12);
-                float32x4_t _k14 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k10, _r10);
-                _sum0 = vmlaq_f32(_sum0, _k11, _r11);
-                _sum0 = vmlaq_f32(_sum0, _k12, _r12);
-                _sum0 = vmlaq_f32(_sum0, _k13, _r13);
-                _sum0 = vmlaq_f32(_sum0, _k14, _r14);
-                _sum1 = vmlaq_f32(_sum1, _k10, _r11);
-                _sum1 = vmlaq_f32(_sum1, _k11, _r12);
-                _sum1 = vmlaq_f32(_sum1, _k12, _r13);
-                _sum1 = vmlaq_f32(_sum1, _k13, _r14);
-                _sum1 = vmlaq_f32(_sum1, _k14, _r15);
-
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r22 = vld1q_f32(r2 + 8);
-                float32x4_t _r23 = vld1q_f32(r2 + 12);
-                float32x4_t _r24 = vld1q_f32(r2 + 16);
-                float32x4_t _r25 = vld1q_f32(r2 + 20);
-
-                float32x4_t _k20 = vld1q_f32(k0);
-                float32x4_t _k21 = vld1q_f32(k0 + 4);
-                float32x4_t _k22 = vld1q_f32(k0 + 8);
-                float32x4_t _k23 = vld1q_f32(k0 + 12);
-                float32x4_t _k24 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k20, _r20);
-                _sum0 = vmlaq_f32(_sum0, _k21, _r21);
-                _sum0 = vmlaq_f32(_sum0, _k22, _r22);
-                _sum0 = vmlaq_f32(_sum0, _k23, _r23);
-                _sum0 = vmlaq_f32(_sum0, _k24, _r24);
-                _sum1 = vmlaq_f32(_sum1, _k20, _r21);
-                _sum1 = vmlaq_f32(_sum1, _k21, _r22);
-                _sum1 = vmlaq_f32(_sum1, _k22, _r23);
-                _sum1 = vmlaq_f32(_sum1, _k23, _r24);
-                _sum1 = vmlaq_f32(_sum1, _k24, _r25);
-
-                float32x4_t _r30 = vld1q_f32(r3);
-                float32x4_t _r31 = vld1q_f32(r3 + 4);
-                float32x4_t _r32 = vld1q_f32(r3 + 8);
-                float32x4_t _r33 = vld1q_f32(r3 + 12);
-                float32x4_t _r34 = vld1q_f32(r3 + 16);
-                float32x4_t _r35 = vld1q_f32(r3 + 20);
-
-                float32x4_t _k30 = vld1q_f32(k0);
-                float32x4_t _k31 = vld1q_f32(k0 + 4);
-                float32x4_t _k32 = vld1q_f32(k0 + 8);
-                float32x4_t _k33 = vld1q_f32(k0 + 12);
-                float32x4_t _k34 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k30, _r30);
-                _sum0 = vmlaq_f32(_sum0, _k31, _r31);
-                _sum0 = vmlaq_f32(_sum0, _k32, _r32);
-                _sum0 = vmlaq_f32(_sum0, _k33, _r33);
-                _sum0 = vmlaq_f32(_sum0, _k34, _r34);
-                _sum1 = vmlaq_f32(_sum1, _k30, _r31);
-                _sum1 = vmlaq_f32(_sum1, _k31, _r32);
-                _sum1 = vmlaq_f32(_sum1, _k32, _r33);
-                _sum1 = vmlaq_f32(_sum1, _k33, _r34);
-                _sum1 = vmlaq_f32(_sum1, _k34, _r35);
-
-                float32x4_t _r40 = vld1q_f32(r4);
-                float32x4_t _r41 = vld1q_f32(r4 + 4);
-                float32x4_t _r42 = vld1q_f32(r4 + 8);
-                float32x4_t _r43 = vld1q_f32(r4 + 12);
-                float32x4_t _r44 = vld1q_f32(r4 + 16);
-                float32x4_t _r45 = vld1q_f32(r4 + 20);
-
-                float32x4_t _k40 = vld1q_f32(k0);
-                float32x4_t _k41 = vld1q_f32(k0 + 4);
-                float32x4_t _k42 = vld1q_f32(k0 + 8);
-                float32x4_t _k43 = vld1q_f32(k0 + 12);
-                float32x4_t _k44 = vld1q_f32(k0 + 16);
-                k0 -= 80;
-
-                _sum0 = vmlaq_f32(_sum0, _k40, _r40);
-                _sum0 = vmlaq_f32(_sum0, _k41, _r41);
-                _sum0 = vmlaq_f32(_sum0, _k42, _r42);
-                _sum0 = vmlaq_f32(_sum0, _k43, _r43);
-                _sum0 = vmlaq_f32(_sum0, _k44, _r44);
-                _sum1 = vmlaq_f32(_sum1, _k40, _r41);
-                _sum1 = vmlaq_f32(_sum1, _k41, _r42);
-                _sum1 = vmlaq_f32(_sum1, _k42, _r43);
-                _sum1 = vmlaq_f32(_sum1, _k43, _r44);
-                _sum1 = vmlaq_f32(_sum1, _k44, _r45);
-
-                vst1q_f32(outptr0, _sum0);
-                vst1q_f32(outptr0 + 4, _sum1);
-
-                r0 += 8;
-                r1 += 8;
-                r2 += 8;
-                r3 += 8;
-                r4 += 8;
-                outptr0 += 8;
-            }
-            for (; j < outw; j++)
-            {
-                float32x4_t _sum0 = _bias0;
-
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r02 = vld1q_f32(r0 + 8);
-                float32x4_t _r03 = vld1q_f32(r0 + 12);
-                float32x4_t _r04 = vld1q_f32(r0 + 16);
-
-                float32x4_t _k00 = vld1q_f32(k0);
-                float32x4_t _k01 = vld1q_f32(k0 + 4);
-                float32x4_t _k02 = vld1q_f32(k0 + 8);
-                float32x4_t _k03 = vld1q_f32(k0 + 12);
-                float32x4_t _k04 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k00, _r00);
-                _sum0 = vmlaq_f32(_sum0, _k01, _r01);
-                _sum0 = vmlaq_f32(_sum0, _k02, _r02);
-                _sum0 = vmlaq_f32(_sum0, _k03, _r03);
-                _sum0 = vmlaq_f32(_sum0, _k04, _r04);
-
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r12 = vld1q_f32(r1 + 8);
-                float32x4_t _r13 = vld1q_f32(r1 + 12);
-                float32x4_t _r14 = vld1q_f32(r1 + 16);
-
-                float32x4_t _k10 = vld1q_f32(k0);
-                float32x4_t _k11 = vld1q_f32(k0 + 4);
-                float32x4_t _k12 = vld1q_f32(k0 + 8);
-                float32x4_t _k13 = vld1q_f32(k0 + 12);
-                float32x4_t _k14 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k10, _r10);
-                _sum0 = vmlaq_f32(_sum0, _k11, _r11);
-                _sum0 = vmlaq_f32(_sum0, _k12, _r12);
-                _sum0 = vmlaq_f32(_sum0, _k13, _r13);
-                _sum0 = vmlaq_f32(_sum0, _k14, _r14);
-
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r22 = vld1q_f32(r2 + 8);
-                float32x4_t _r23 = vld1q_f32(r2 + 12);
-                float32x4_t _r24 = vld1q_f32(r2 + 16);
-
-                float32x4_t _k20 = vld1q_f32(k0);
-                float32x4_t _k21 = vld1q_f32(k0 + 4);
-                float32x4_t _k22 = vld1q_f32(k0 + 8);
-                float32x4_t _k23 = vld1q_f32(k0 + 12);
-                float32x4_t _k24 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k20, _r20);
-                _sum0 = vmlaq_f32(_sum0, _k21, _r21);
-                _sum0 = vmlaq_f32(_sum0, _k22, _r22);
-                _sum0 = vmlaq_f32(_sum0, _k23, _r23);
-                _sum0 = vmlaq_f32(_sum0, _k24, _r24);
-
-                float32x4_t _r30 = vld1q_f32(r3);
-                float32x4_t _r31 = vld1q_f32(r3 + 4);
-                float32x4_t _r32 = vld1q_f32(r3 + 8);
-                float32x4_t _r33 = vld1q_f32(r3 + 12);
-                float32x4_t _r34 = vld1q_f32(r3 + 16);
-
-                float32x4_t _k30 = vld1q_f32(k0);
-                float32x4_t _k31 = vld1q_f32(k0 + 4);
-                float32x4_t _k32 = vld1q_f32(k0 + 8);
-                float32x4_t _k33 = vld1q_f32(k0 + 12);
-                float32x4_t _k34 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k30, _r30);
-                _sum0 = vmlaq_f32(_sum0, _k31, _r31);
-                _sum0 = vmlaq_f32(_sum0, _k32, _r32);
-                _sum0 = vmlaq_f32(_sum0, _k33, _r33);
-                _sum0 = vmlaq_f32(_sum0, _k34, _r34);
-
-                float32x4_t _r40 = vld1q_f32(r4);
-                float32x4_t _r41 = vld1q_f32(r4 + 4);
-                float32x4_t _r42 = vld1q_f32(r4 + 8);
-                float32x4_t _r43 = vld1q_f32(r4 + 12);
-                float32x4_t _r44 = vld1q_f32(r4 + 16);
-
-                float32x4_t _k40 = vld1q_f32(k0);
-                float32x4_t _k41 = vld1q_f32(k0 + 4);
-                float32x4_t _k42 = vld1q_f32(k0 + 8);
-                float32x4_t _k43 = vld1q_f32(k0 + 12);
-                float32x4_t _k44 = vld1q_f32(k0 + 16);
-                k0 -= 80;
-
-                _sum0 = vmlaq_f32(_sum0, _k40, _r40);
-                _sum0 = vmlaq_f32(_sum0, _k41, _r41);
-                _sum0 = vmlaq_f32(_sum0, _k42, _r42);
-                _sum0 = vmlaq_f32(_sum0, _k43, _r43);
-                _sum0 = vmlaq_f32(_sum0, _k44, _r44);
-
-                vst1q_f32(outptr0, _sum0);
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                r3 += 4;
-                r4 += 4;
-                outptr0 += 4;
-            }
-
-            r0 += 4 * 4;
-            r1 += 4 * 4;
-            r2 += 4 * 4;
-            r3 += 4 * 4;
-            r4 += 4 * 4;
-        }
-    }
-}
-
-static void convdw5x5s2_pack4_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + g * 4) : vdupq_n_f32(0.f);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out;
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-        const float* r4 = img0.row(4);
-
-        int i = 0;
-
-        for (; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j + 3 < outw; j += 4)
-            {
-                float32x4_t _sum0 = _bias0;
-                float32x4_t _sum1 = _bias0;
-                float32x4_t _sum2 = _bias0;
-                float32x4_t _sum3 = _bias0;
-
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r02 = vld1q_f32(r0 + 8);
-                float32x4_t _r03 = vld1q_f32(r0 + 12);
-                float32x4_t _r04 = vld1q_f32(r0 + 16);
-                float32x4_t _r05 = vld1q_f32(r0 + 20);
-                float32x4_t _r06 = vld1q_f32(r0 + 24);
-                float32x4_t _r07 = vld1q_f32(r0 + 28);
-                float32x4_t _r08 = vld1q_f32(r0 + 32);
-                float32x4_t _r09 = vld1q_f32(r0 + 36);
-                float32x4_t _r010 = vld1q_f32(r0 + 40);
-
-                float32x4_t _k00 = vld1q_f32(k0);
-                float32x4_t _k01 = vld1q_f32(k0 + 4);
-                float32x4_t _k02 = vld1q_f32(k0 + 8);
-                float32x4_t _k03 = vld1q_f32(k0 + 12);
-                float32x4_t _k04 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k00, _r00);
-                _sum0 = vmlaq_f32(_sum0, _k01, _r01);
-                _sum0 = vmlaq_f32(_sum0, _k02, _r02);
-                _sum0 = vmlaq_f32(_sum0, _k03, _r03);
-                _sum0 = vmlaq_f32(_sum0, _k04, _r04);
-                _sum1 = vmlaq_f32(_sum1, _k00, _r02);
-                _sum1 = vmlaq_f32(_sum1, _k01, _r03);
-                _sum1 = vmlaq_f32(_sum1, _k02, _r04);
-                _sum1 = vmlaq_f32(_sum1, _k03, _r05);
-                _sum1 = vmlaq_f32(_sum1, _k04, _r06);
-                _sum2 = vmlaq_f32(_sum2, _k00, _r04);
-                _sum2 = vmlaq_f32(_sum2, _k01, _r05);
-                _sum2 = vmlaq_f32(_sum2, _k02, _r06);
-                _sum2 = vmlaq_f32(_sum2, _k03, _r07);
-                _sum2 = vmlaq_f32(_sum2, _k04, _r08);
-                _sum3 = vmlaq_f32(_sum3, _k00, _r06);
-                _sum3 = vmlaq_f32(_sum3, _k01, _r07);
-                _sum3 = vmlaq_f32(_sum3, _k02, _r08);
-                _sum3 = vmlaq_f32(_sum3, _k03, _r09);
-                _sum3 = vmlaq_f32(_sum3, _k04, _r010);
-
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r12 = vld1q_f32(r1 + 8);
-                float32x4_t _r13 = vld1q_f32(r1 + 12);
-                float32x4_t _r14 = vld1q_f32(r1 + 16);
-                float32x4_t _r15 = vld1q_f32(r1 + 20);
-                float32x4_t _r16 = vld1q_f32(r1 + 24);
-                float32x4_t _r17 = vld1q_f32(r1 + 28);
-                float32x4_t _r18 = vld1q_f32(r1 + 32);
-                float32x4_t _r19 = vld1q_f32(r1 + 36);
-                float32x4_t _r110 = vld1q_f32(r1 + 40);
-
-                float32x4_t _k10 = vld1q_f32(k0);
-                float32x4_t _k11 = vld1q_f32(k0 + 4);
-                float32x4_t _k12 = vld1q_f32(k0 + 8);
-                float32x4_t _k13 = vld1q_f32(k0 + 12);
-                float32x4_t _k14 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k10, _r10);
-                _sum0 = vmlaq_f32(_sum0, _k11, _r11);
-                _sum0 = vmlaq_f32(_sum0, _k12, _r12);
-                _sum0 = vmlaq_f32(_sum0, _k13, _r13);
-                _sum0 = vmlaq_f32(_sum0, _k14, _r14);
-                _sum1 = vmlaq_f32(_sum1, _k10, _r12);
-                _sum1 = vmlaq_f32(_sum1, _k11, _r13);
-                _sum1 = vmlaq_f32(_sum1, _k12, _r14);
-                _sum1 = vmlaq_f32(_sum1, _k13, _r15);
-                _sum1 = vmlaq_f32(_sum1, _k14, _r16);
-                _sum2 = vmlaq_f32(_sum2, _k10, _r14);
-                _sum2 = vmlaq_f32(_sum2, _k11, _r15);
-                _sum2 = vmlaq_f32(_sum2, _k12, _r16);
-                _sum2 = vmlaq_f32(_sum2, _k13, _r17);
-                _sum2 = vmlaq_f32(_sum2, _k14, _r18);
-                _sum3 = vmlaq_f32(_sum3, _k10, _r16);
-                _sum3 = vmlaq_f32(_sum3, _k11, _r17);
-                _sum3 = vmlaq_f32(_sum3, _k12, _r18);
-                _sum3 = vmlaq_f32(_sum3, _k13, _r19);
-                _sum3 = vmlaq_f32(_sum3, _k14, _r110);
-
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r22 = vld1q_f32(r2 + 8);
-                float32x4_t _r23 = vld1q_f32(r2 + 12);
-                float32x4_t _r24 = vld1q_f32(r2 + 16);
-                float32x4_t _r25 = vld1q_f32(r2 + 20);
-                float32x4_t _r26 = vld1q_f32(r2 + 24);
-                float32x4_t _r27 = vld1q_f32(r2 + 28);
-                float32x4_t _r28 = vld1q_f32(r2 + 32);
-                float32x4_t _r29 = vld1q_f32(r2 + 36);
-                float32x4_t _r210 = vld1q_f32(r2 + 40);
-
-                float32x4_t _k20 = vld1q_f32(k0);
-                float32x4_t _k21 = vld1q_f32(k0 + 4);
-                float32x4_t _k22 = vld1q_f32(k0 + 8);
-                float32x4_t _k23 = vld1q_f32(k0 + 12);
-                float32x4_t _k24 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k20, _r20);
-                _sum0 = vmlaq_f32(_sum0, _k21, _r21);
-                _sum0 = vmlaq_f32(_sum0, _k22, _r22);
-                _sum0 = vmlaq_f32(_sum0, _k23, _r23);
-                _sum0 = vmlaq_f32(_sum0, _k24, _r24);
-                _sum1 = vmlaq_f32(_sum1, _k20, _r22);
-                _sum1 = vmlaq_f32(_sum1, _k21, _r23);
-                _sum1 = vmlaq_f32(_sum1, _k22, _r24);
-                _sum1 = vmlaq_f32(_sum1, _k23, _r25);
-                _sum1 = vmlaq_f32(_sum1, _k24, _r26);
-                _sum2 = vmlaq_f32(_sum2, _k20, _r24);
-                _sum2 = vmlaq_f32(_sum2, _k21, _r25);
-                _sum2 = vmlaq_f32(_sum2, _k22, _r26);
-                _sum2 = vmlaq_f32(_sum2, _k23, _r27);
-                _sum2 = vmlaq_f32(_sum2, _k24, _r28);
-                _sum3 = vmlaq_f32(_sum3, _k20, _r26);
-                _sum3 = vmlaq_f32(_sum3, _k21, _r27);
-                _sum3 = vmlaq_f32(_sum3, _k22, _r28);
-                _sum3 = vmlaq_f32(_sum3, _k23, _r29);
-                _sum3 = vmlaq_f32(_sum3, _k24, _r210);
-
-                float32x4_t _r30 = vld1q_f32(r3);
-                float32x4_t _r31 = vld1q_f32(r3 + 4);
-                float32x4_t _r32 = vld1q_f32(r3 + 8);
-                float32x4_t _r33 = vld1q_f32(r3 + 12);
-                float32x4_t _r34 = vld1q_f32(r3 + 16);
-                float32x4_t _r35 = vld1q_f32(r3 + 20);
-                float32x4_t _r36 = vld1q_f32(r3 + 24);
-                float32x4_t _r37 = vld1q_f32(r3 + 28);
-                float32x4_t _r38 = vld1q_f32(r3 + 32);
-                float32x4_t _r39 = vld1q_f32(r3 + 36);
-                float32x4_t _r310 = vld1q_f32(r3 + 40);
-
-                float32x4_t _k30 = vld1q_f32(k0);
-                float32x4_t _k31 = vld1q_f32(k0 + 4);
-                float32x4_t _k32 = vld1q_f32(k0 + 8);
-                float32x4_t _k33 = vld1q_f32(k0 + 12);
-                float32x4_t _k34 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k30, _r30);
-                _sum0 = vmlaq_f32(_sum0, _k31, _r31);
-                _sum0 = vmlaq_f32(_sum0, _k32, _r32);
-                _sum0 = vmlaq_f32(_sum0, _k33, _r33);
-                _sum0 = vmlaq_f32(_sum0, _k34, _r34);
-                _sum1 = vmlaq_f32(_sum1, _k30, _r32);
-                _sum1 = vmlaq_f32(_sum1, _k31, _r33);
-                _sum1 = vmlaq_f32(_sum1, _k32, _r34);
-                _sum1 = vmlaq_f32(_sum1, _k33, _r35);
-                _sum1 = vmlaq_f32(_sum1, _k34, _r36);
-                _sum2 = vmlaq_f32(_sum2, _k30, _r34);
-                _sum2 = vmlaq_f32(_sum2, _k31, _r35);
-                _sum2 = vmlaq_f32(_sum2, _k32, _r36);
-                _sum2 = vmlaq_f32(_sum2, _k33, _r37);
-                _sum2 = vmlaq_f32(_sum2, _k34, _r38);
-                _sum3 = vmlaq_f32(_sum3, _k30, _r36);
-                _sum3 = vmlaq_f32(_sum3, _k31, _r37);
-                _sum3 = vmlaq_f32(_sum3, _k32, _r38);
-                _sum3 = vmlaq_f32(_sum3, _k33, _r39);
-                _sum3 = vmlaq_f32(_sum3, _k34, _r310);
-
-                float32x4_t _r40 = vld1q_f32(r4);
-                float32x4_t _r41 = vld1q_f32(r4 + 4);
-                float32x4_t _r42 = vld1q_f32(r4 + 8);
-                float32x4_t _r43 = vld1q_f32(r4 + 12);
-                float32x4_t _r44 = vld1q_f32(r4 + 16);
-                float32x4_t _r45 = vld1q_f32(r4 + 20);
-                float32x4_t _r46 = vld1q_f32(r4 + 24);
-                float32x4_t _r47 = vld1q_f32(r4 + 28);
-                float32x4_t _r48 = vld1q_f32(r4 + 32);
-                float32x4_t _r49 = vld1q_f32(r4 + 36);
-                float32x4_t _r410 = vld1q_f32(r4 + 40);
-
-                float32x4_t _k40 = vld1q_f32(k0);
-                float32x4_t _k41 = vld1q_f32(k0 + 4);
-                float32x4_t _k42 = vld1q_f32(k0 + 8);
-                float32x4_t _k43 = vld1q_f32(k0 + 12);
-                float32x4_t _k44 = vld1q_f32(k0 + 16);
-                k0 -= 80;
-
-                _sum0 = vmlaq_f32(_sum0, _k40, _r40);
-                _sum0 = vmlaq_f32(_sum0, _k41, _r41);
-                _sum0 = vmlaq_f32(_sum0, _k42, _r42);
-                _sum0 = vmlaq_f32(_sum0, _k43, _r43);
-                _sum0 = vmlaq_f32(_sum0, _k44, _r44);
-                _sum1 = vmlaq_f32(_sum1, _k40, _r42);
-                _sum1 = vmlaq_f32(_sum1, _k41, _r43);
-                _sum1 = vmlaq_f32(_sum1, _k42, _r44);
-                _sum1 = vmlaq_f32(_sum1, _k43, _r45);
-                _sum1 = vmlaq_f32(_sum1, _k44, _r46);
-                _sum2 = vmlaq_f32(_sum2, _k40, _r44);
-                _sum2 = vmlaq_f32(_sum2, _k41, _r45);
-                _sum2 = vmlaq_f32(_sum2, _k42, _r46);
-                _sum2 = vmlaq_f32(_sum2, _k43, _r47);
-                _sum2 = vmlaq_f32(_sum2, _k44, _r48);
-                _sum3 = vmlaq_f32(_sum3, _k40, _r46);
-                _sum3 = vmlaq_f32(_sum3, _k41, _r47);
-                _sum3 = vmlaq_f32(_sum3, _k42, _r48);
-                _sum3 = vmlaq_f32(_sum3, _k43, _r49);
-                _sum3 = vmlaq_f32(_sum3, _k44, _r410);
-
-                vst1q_f32(outptr0, _sum0);
-                vst1q_f32(outptr0 + 4, _sum1);
-                vst1q_f32(outptr0 + 8, _sum2);
-                vst1q_f32(outptr0 + 12, _sum3);
-
-                r0 += 8 * 4;
-                r1 += 8 * 4;
-                r2 += 8 * 4;
-                r3 += 8 * 4;
-                r4 += 8 * 4;
-                outptr0 += 16;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                float32x4_t _sum0 = _bias0;
-                float32x4_t _sum1 = _bias0;
-
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r02 = vld1q_f32(r0 + 8);
-                float32x4_t _r03 = vld1q_f32(r0 + 12);
-                float32x4_t _r04 = vld1q_f32(r0 + 16);
-                float32x4_t _r05 = vld1q_f32(r0 + 20);
-                float32x4_t _r06 = vld1q_f32(r0 + 24);
-
-                float32x4_t _k00 = vld1q_f32(k0);
-                float32x4_t _k01 = vld1q_f32(k0 + 4);
-                float32x4_t _k02 = vld1q_f32(k0 + 8);
-                float32x4_t _k03 = vld1q_f32(k0 + 12);
-                float32x4_t _k04 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k00, _r00);
-                _sum0 = vmlaq_f32(_sum0, _k01, _r01);
-                _sum0 = vmlaq_f32(_sum0, _k02, _r02);
-                _sum0 = vmlaq_f32(_sum0, _k03, _r03);
-                _sum0 = vmlaq_f32(_sum0, _k04, _r04);
-                _sum1 = vmlaq_f32(_sum1, _k00, _r02);
-                _sum1 = vmlaq_f32(_sum1, _k01, _r03);
-                _sum1 = vmlaq_f32(_sum1, _k02, _r04);
-                _sum1 = vmlaq_f32(_sum1, _k03, _r05);
-                _sum1 = vmlaq_f32(_sum1, _k04, _r06);
-
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r12 = vld1q_f32(r1 + 8);
-                float32x4_t _r13 = vld1q_f32(r1 + 12);
-                float32x4_t _r14 = vld1q_f32(r1 + 16);
-                float32x4_t _r15 = vld1q_f32(r1 + 20);
-                float32x4_t _r16 = vld1q_f32(r1 + 24);
-
-                float32x4_t _k10 = vld1q_f32(k0);
-                float32x4_t _k11 = vld1q_f32(k0 + 4);
-                float32x4_t _k12 = vld1q_f32(k0 + 8);
-                float32x4_t _k13 = vld1q_f32(k0 + 12);
-                float32x4_t _k14 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k10, _r10);
-                _sum0 = vmlaq_f32(_sum0, _k11, _r11);
-                _sum0 = vmlaq_f32(_sum0, _k12, _r12);
-                _sum0 = vmlaq_f32(_sum0, _k13, _r13);
-                _sum0 = vmlaq_f32(_sum0, _k14, _r14);
-                _sum1 = vmlaq_f32(_sum1, _k10, _r12);
-                _sum1 = vmlaq_f32(_sum1, _k11, _r13);
-                _sum1 = vmlaq_f32(_sum1, _k12, _r14);
-                _sum1 = vmlaq_f32(_sum1, _k13, _r15);
-                _sum1 = vmlaq_f32(_sum1, _k14, _r16);
-
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r22 = vld1q_f32(r2 + 8);
-                float32x4_t _r23 = vld1q_f32(r2 + 12);
-                float32x4_t _r24 = vld1q_f32(r2 + 16);
-                float32x4_t _r25 = vld1q_f32(r2 + 20);
-                float32x4_t _r26 = vld1q_f32(r2 + 24);
-
-                float32x4_t _k20 = vld1q_f32(k0);
-                float32x4_t _k21 = vld1q_f32(k0 + 4);
-                float32x4_t _k22 = vld1q_f32(k0 + 8);
-                float32x4_t _k23 = vld1q_f32(k0 + 12);
-                float32x4_t _k24 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k20, _r20);
-                _sum0 = vmlaq_f32(_sum0, _k21, _r21);
-                _sum0 = vmlaq_f32(_sum0, _k22, _r22);
-                _sum0 = vmlaq_f32(_sum0, _k23, _r23);
-                _sum0 = vmlaq_f32(_sum0, _k24, _r24);
-                _sum1 = vmlaq_f32(_sum1, _k20, _r22);
-                _sum1 = vmlaq_f32(_sum1, _k21, _r23);
-                _sum1 = vmlaq_f32(_sum1, _k22, _r24);
-                _sum1 = vmlaq_f32(_sum1, _k23, _r25);
-                _sum1 = vmlaq_f32(_sum1, _k24, _r26);
-
-                float32x4_t _r30 = vld1q_f32(r3);
-                float32x4_t _r31 = vld1q_f32(r3 + 4);
-                float32x4_t _r32 = vld1q_f32(r3 + 8);
-                float32x4_t _r33 = vld1q_f32(r3 + 12);
-                float32x4_t _r34 = vld1q_f32(r3 + 16);
-                float32x4_t _r35 = vld1q_f32(r3 + 20);
-                float32x4_t _r36 = vld1q_f32(r3 + 24);
-
-                float32x4_t _k30 = vld1q_f32(k0);
-                float32x4_t _k31 = vld1q_f32(k0 + 4);
-                float32x4_t _k32 = vld1q_f32(k0 + 8);
-                float32x4_t _k33 = vld1q_f32(k0 + 12);
-                float32x4_t _k34 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k30, _r30);
-                _sum0 = vmlaq_f32(_sum0, _k31, _r31);
-                _sum0 = vmlaq_f32(_sum0, _k32, _r32);
-                _sum0 = vmlaq_f32(_sum0, _k33, _r33);
-                _sum0 = vmlaq_f32(_sum0, _k34, _r34);
-                _sum1 = vmlaq_f32(_sum1, _k30, _r32);
-                _sum1 = vmlaq_f32(_sum1, _k31, _r33);
-                _sum1 = vmlaq_f32(_sum1, _k32, _r34);
-                _sum1 = vmlaq_f32(_sum1, _k33, _r35);
-                _sum1 = vmlaq_f32(_sum1, _k34, _r36);
-
-                float32x4_t _r40 = vld1q_f32(r4);
-                float32x4_t _r41 = vld1q_f32(r4 + 4);
-                float32x4_t _r42 = vld1q_f32(r4 + 8);
-                float32x4_t _r43 = vld1q_f32(r4 + 12);
-                float32x4_t _r44 = vld1q_f32(r4 + 16);
-                float32x4_t _r45 = vld1q_f32(r4 + 20);
-                float32x4_t _r46 = vld1q_f32(r4 + 24);
-
-                float32x4_t _k40 = vld1q_f32(k0);
-                float32x4_t _k41 = vld1q_f32(k0 + 4);
-                float32x4_t _k42 = vld1q_f32(k0 + 8);
-                float32x4_t _k43 = vld1q_f32(k0 + 12);
-                float32x4_t _k44 = vld1q_f32(k0 + 16);
-                k0 -= 80;
-
-                _sum0 = vmlaq_f32(_sum0, _k40, _r40);
-                _sum0 = vmlaq_f32(_sum0, _k41, _r41);
-                _sum0 = vmlaq_f32(_sum0, _k42, _r42);
-                _sum0 = vmlaq_f32(_sum0, _k43, _r43);
-                _sum0 = vmlaq_f32(_sum0, _k44, _r44);
-                _sum1 = vmlaq_f32(_sum1, _k40, _r42);
-                _sum1 = vmlaq_f32(_sum1, _k41, _r43);
-                _sum1 = vmlaq_f32(_sum1, _k42, _r44);
-                _sum1 = vmlaq_f32(_sum1, _k43, _r45);
-                _sum1 = vmlaq_f32(_sum1, _k44, _r46);
-
-                vst1q_f32(outptr0, _sum0);
-                vst1q_f32(outptr0 + 4, _sum1);
-
-                r0 += 4 * 4;
-                r1 += 4 * 4;
-                r2 += 4 * 4;
-                r3 += 4 * 4;
-                r4 += 4 * 4;
-                outptr0 += 8;
-            }
-            for (; j < outw; j++)
-            {
-                float32x4_t _sum0 = _bias0;
-
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r02 = vld1q_f32(r0 + 8);
-                float32x4_t _r03 = vld1q_f32(r0 + 12);
-                float32x4_t _r04 = vld1q_f32(r0 + 16);
-
-                float32x4_t _k00 = vld1q_f32(k0);
-                float32x4_t _k01 = vld1q_f32(k0 + 4);
-                float32x4_t _k02 = vld1q_f32(k0 + 8);
-                float32x4_t _k03 = vld1q_f32(k0 + 12);
-                float32x4_t _k04 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k00, _r00);
-                _sum0 = vmlaq_f32(_sum0, _k01, _r01);
-                _sum0 = vmlaq_f32(_sum0, _k02, _r02);
-                _sum0 = vmlaq_f32(_sum0, _k03, _r03);
-                _sum0 = vmlaq_f32(_sum0, _k04, _r04);
-
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r12 = vld1q_f32(r1 + 8);
-                float32x4_t _r13 = vld1q_f32(r1 + 12);
-                float32x4_t _r14 = vld1q_f32(r1 + 16);
-
-                float32x4_t _k10 = vld1q_f32(k0);
-                float32x4_t _k11 = vld1q_f32(k0 + 4);
-                float32x4_t _k12 = vld1q_f32(k0 + 8);
-                float32x4_t _k13 = vld1q_f32(k0 + 12);
-                float32x4_t _k14 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k10, _r10);
-                _sum0 = vmlaq_f32(_sum0, _k11, _r11);
-                _sum0 = vmlaq_f32(_sum0, _k12, _r12);
-                _sum0 = vmlaq_f32(_sum0, _k13, _r13);
-                _sum0 = vmlaq_f32(_sum0, _k14, _r14);
-
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r22 = vld1q_f32(r2 + 8);
-                float32x4_t _r23 = vld1q_f32(r2 + 12);
-                float32x4_t _r24 = vld1q_f32(r2 + 16);
-
-                float32x4_t _k20 = vld1q_f32(k0);
-                float32x4_t _k21 = vld1q_f32(k0 + 4);
-                float32x4_t _k22 = vld1q_f32(k0 + 8);
-                float32x4_t _k23 = vld1q_f32(k0 + 12);
-                float32x4_t _k24 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k20, _r20);
-                _sum0 = vmlaq_f32(_sum0, _k21, _r21);
-                _sum0 = vmlaq_f32(_sum0, _k22, _r22);
-                _sum0 = vmlaq_f32(_sum0, _k23, _r23);
-                _sum0 = vmlaq_f32(_sum0, _k24, _r24);
-
-                float32x4_t _r30 = vld1q_f32(r3);
-                float32x4_t _r31 = vld1q_f32(r3 + 4);
-                float32x4_t _r32 = vld1q_f32(r3 + 8);
-                float32x4_t _r33 = vld1q_f32(r3 + 12);
-                float32x4_t _r34 = vld1q_f32(r3 + 16);
-
-                float32x4_t _k30 = vld1q_f32(k0);
-                float32x4_t _k31 = vld1q_f32(k0 + 4);
-                float32x4_t _k32 = vld1q_f32(k0 + 8);
-                float32x4_t _k33 = vld1q_f32(k0 + 12);
-                float32x4_t _k34 = vld1q_f32(k0 + 16);
-                k0 += 20;
-
-                _sum0 = vmlaq_f32(_sum0, _k30, _r30);
-                _sum0 = vmlaq_f32(_sum0, _k31, _r31);
-                _sum0 = vmlaq_f32(_sum0, _k32, _r32);
-                _sum0 = vmlaq_f32(_sum0, _k33, _r33);
-                _sum0 = vmlaq_f32(_sum0, _k34, _r34);
-
-                float32x4_t _r40 = vld1q_f32(r4);
-                float32x4_t _r41 = vld1q_f32(r4 + 4);
-                float32x4_t _r42 = vld1q_f32(r4 + 8);
-                float32x4_t _r43 = vld1q_f32(r4 + 12);
-                float32x4_t _r44 = vld1q_f32(r4 + 16);
-
-                float32x4_t _k40 = vld1q_f32(k0);
-                float32x4_t _k41 = vld1q_f32(k0 + 4);
-                float32x4_t _k42 = vld1q_f32(k0 + 8);
-                float32x4_t _k43 = vld1q_f32(k0 + 12);
-                float32x4_t _k44 = vld1q_f32(k0 + 16);
-                k0 -= 80;
-
-                _sum0 = vmlaq_f32(_sum0, _k40, _r40);
-                _sum0 = vmlaq_f32(_sum0, _k41, _r41);
-                _sum0 = vmlaq_f32(_sum0, _k42, _r42);
-                _sum0 = vmlaq_f32(_sum0, _k43, _r43);
-                _sum0 = vmlaq_f32(_sum0, _k44, _r44);
-
-                vst1q_f32(outptr0, _sum0);
-
-                r0 += 2 * 4;
-                r1 += 2 * 4;
-                r2 += 2 * 4;
-                r3 += 2 * 4;
-                r4 += 2 * 4;
-                outptr0 += 4;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-            r3 += tailstep;
-            r4 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5_pack4_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5_pack4_bf16s.h
deleted file mode 100644
index f65c3ab5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5_pack4_bf16s.h
+++ /dev/null
@@ -1,3818 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw5x5s1_pack4_bf16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-#if __aarch64__
-    const int w = bottom_blob.w;
-#endif
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const unsigned short* kptr = kernel.row<const unsigned short>(g);
-
-        unsigned short* outptr0 = out.row<unsigned short>(0);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const unsigned short* r0 = img0.row<const unsigned short>(0);
-        const unsigned short* r1 = img0.row<const unsigned short>(1);
-        const unsigned short* r2 = img0.row<const unsigned short>(2);
-        const unsigned short* r3 = img0.row<const unsigned short>(3);
-        const unsigned short* r4 = img0.row<const unsigned short>(4);
-
-#if __aarch64__
-        unsigned short* outptr1 = out.row<unsigned short>(1);
-        const unsigned short* r5 = img0.row<const unsigned short>(5);
-
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + g * 4) : vdupq_n_f32(0.f);
-
-        // 4 * 25
-        uint16x8_t _k00_01 = vld1q_u16(kptr);
-        uint16x8_t _k02_03 = vld1q_u16(kptr + 8);
-        uint16x8_t _k04_10 = vld1q_u16(kptr + 16);
-        uint16x8_t _k11_12 = vld1q_u16(kptr + 24);
-        uint16x8_t _k13_14 = vld1q_u16(kptr + 32);
-        uint16x8_t _k20_21 = vld1q_u16(kptr + 40);
-        uint16x8_t _k22_23 = vld1q_u16(kptr + 48);
-        uint16x8_t _k24_30 = vld1q_u16(kptr + 56);
-        uint16x8_t _k31_32 = vld1q_u16(kptr + 64);
-        uint16x8_t _k33_34 = vld1q_u16(kptr + 72);
-        uint16x8_t _k40_41 = vld1q_u16(kptr + 80);
-        uint16x8_t _k42_43 = vld1q_u16(kptr + 88);
-        uint16x4_t _k44 = vld1_u16(kptr + 96);
-#else  // __aarch64__
-        float bias0_data[4];
-        if (bias)
-        {
-            bias0_data[0] = bias[g * 4 + 0];
-            bias0_data[1] = bias[g * 4 + 1];
-            bias0_data[2] = bias[g * 4 + 2];
-            bias0_data[3] = bias[g * 4 + 3];
-        }
-        else
-        {
-            bias0_data[0] = 0.f;
-            bias0_data[1] = 0.f;
-            bias0_data[2] = 0.f;
-            bias0_data[3] = 0.f;
-        }
-        const float* bias0_data_ptr = bias0_data;
-#endif // __aarch64__
-
-        int i = 0;
-#if __aarch64__
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-
-            for (; j + 3 < outw; j += 4)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%3], #32 \n" // r10 r11 r12 r13
-
-                    "shll2  v14.4s, %18.8h, #16         \n"
-
-                    "mov    v24.16b, %29.16b            \n" // sum00
-                    "mov    v25.16b, %29.16b            \n" // sum01
-                    "mov    v26.16b, %29.16b            \n" // sum02
-                    "mov    v27.16b, %29.16b            \n" // sum03
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "mov    v28.16b, %29.16b            \n" // sum10
-                    "mov    v29.16b, %29.16b            \n" // sum11
-                    "mov    v30.16b, %29.16b            \n" // sum12
-                    "mov    v31.16b, %29.16b            \n" // sum13
-
-                    "shll   v15.4s, %16.4h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v25.4s, v14.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%3]      \n" // r14 r15 r16 r17
-                    "fmla   v27.4s, v14.4s, v19.4s      \n"
-
-                    "shll   v14.4s, %19.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v16.4s      \n"
-                    "fmla   v29.4s, v15.4s, v17.4s      \n"
-                    "fmla   v30.4s, v15.4s, v18.4s      \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-
-                    "shll2  v15.4s, %16.8h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v17.4s      \n"
-                    "fmla   v25.4s, v14.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v19.4s      \n"
-                    "fmla   v27.4s, v14.4s, v20.4s      \n"
-
-                    "shll2  v14.4s, %19.8h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v17.4s      \n"
-                    "fmla   v29.4s, v15.4s, v18.4s      \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-
-                    "shll   v15.4s, %17.4h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v18.4s      \n"
-                    "fmla   v25.4s, v14.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v20.4s      \n"
-                    "fmla   v27.4s, v14.4s, v21.4s      \n"
-
-                    "shll   v14.4s, %20.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v18.4s      \n"
-                    "fmla   v29.4s, v15.4s, v19.4s      \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-
-                    "shll2  v15.4s, %17.8h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v19.4s      \n"
-                    "fmla   v25.4s, v14.4s, v20.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v21.4s      \n"
-                    "fmla   v27.4s, v14.4s, v22.4s      \n"
-
-                    "shll2  v14.4s, %20.8h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v19.4s      \n"
-                    "fmla   v29.4s, v15.4s, v20.4s      \n"
-                    "fmla   v30.4s, v15.4s, v21.4s      \n"
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%4], #32 \n" // r20 r21 r22 r23
-                    "fmla   v31.4s, v15.4s, v22.4s      \n"
-
-                    "shll   v15.4s, %18.4h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v20.4s      \n"
-                    "fmla   v25.4s, v14.4s, v21.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v22.4s      \n"
-                    "fmla   v27.4s, v14.4s, v23.4s      \n"
-
-                    "shll   v14.4s, %21.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v20.4s      \n"
-                    "fmla   v29.4s, v15.4s, v21.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v22.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v23.4s      \n"
-
-                    "shll2  v15.4s, %18.8h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v25.4s, v14.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%4]      \n" // r24 r25 r26 r27
-                    "fmla   v27.4s, v14.4s, v19.4s      \n"
-
-                    "shll2  v14.4s, %21.8h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v16.4s      \n"
-                    "fmla   v29.4s, v15.4s, v17.4s      \n"
-                    "fmla   v30.4s, v15.4s, v18.4s      \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-
-                    "shll   v15.4s, %19.4h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v17.4s      \n"
-                    "fmla   v25.4s, v14.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v19.4s      \n"
-                    "fmla   v27.4s, v14.4s, v20.4s      \n"
-
-                    "shll   v14.4s, %22.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v17.4s      \n"
-                    "fmla   v29.4s, v15.4s, v18.4s      \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-
-                    "shll2  v15.4s, %19.8h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v18.4s      \n"
-                    "fmla   v25.4s, v14.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v20.4s      \n"
-                    "fmla   v27.4s, v14.4s, v21.4s      \n"
-
-                    "shll2  v14.4s, %22.8h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v18.4s      \n"
-                    "fmla   v29.4s, v15.4s, v19.4s      \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-
-                    "shll   v15.4s, %20.4h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v19.4s      \n"
-                    "fmla   v25.4s, v14.4s, v20.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v21.4s      \n"
-                    "fmla   v27.4s, v14.4s, v22.4s      \n"
-
-                    "shll   v14.4s, %23.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v19.4s      \n"
-                    "fmla   v29.4s, v15.4s, v20.4s      \n"
-                    "fmla   v30.4s, v15.4s, v21.4s      \n"
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%5], #32 \n" // r30 r31 r32 r33
-                    "fmla   v31.4s, v15.4s, v22.4s      \n"
-
-                    "shll2  v15.4s, %20.8h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v20.4s      \n"
-                    "fmla   v25.4s, v14.4s, v21.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v22.4s      \n"
-                    "fmla   v27.4s, v14.4s, v23.4s      \n"
-
-                    "shll2  v14.4s, %23.8h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v20.4s      \n"
-                    "fmla   v29.4s, v15.4s, v21.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v22.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v23.4s      \n"
-
-                    "shll   v15.4s, %21.4h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v25.4s, v14.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%5]      \n" // r34 r35 r36 r37
-                    "fmla   v27.4s, v14.4s, v19.4s      \n"
-
-                    "shll   v14.4s, %24.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v16.4s      \n"
-                    "fmla   v29.4s, v15.4s, v17.4s      \n"
-                    "fmla   v30.4s, v15.4s, v18.4s      \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-
-                    "shll2  v15.4s, %21.8h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v17.4s      \n"
-                    "fmla   v25.4s, v14.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v19.4s      \n"
-                    "fmla   v27.4s, v14.4s, v20.4s      \n"
-
-                    "shll2  v14.4s, %24.8h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v17.4s      \n"
-                    "fmla   v29.4s, v15.4s, v18.4s      \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-
-                    "shll   v15.4s, %22.4h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v18.4s      \n"
-                    "fmla   v25.4s, v14.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v20.4s      \n"
-                    "fmla   v27.4s, v14.4s, v21.4s      \n"
-
-                    "shll   v14.4s, %25.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v18.4s      \n"
-                    "fmla   v29.4s, v15.4s, v19.4s      \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-                    "shll2  v15.4s, %22.8h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v19.4s      \n"
-                    "fmla   v25.4s, v14.4s, v20.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v21.4s      \n"
-                    "fmla   v27.4s, v14.4s, v22.4s      \n"
-
-                    "shll2  v14.4s, %25.8h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v19.4s      \n"
-                    "fmla   v29.4s, v15.4s, v20.4s      \n"
-                    "fmla   v30.4s, v15.4s, v21.4s      \n"
-                    "prfm   pldl1keep, [%6, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%6], #32 \n" // r40 r41 r42 r43
-                    "fmla   v31.4s, v15.4s, v22.4s      \n"
-
-                    "shll   v15.4s, %23.4h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v20.4s      \n"
-                    "fmla   v25.4s, v14.4s, v21.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v22.4s      \n"
-                    "fmla   v27.4s, v14.4s, v23.4s      \n"
-                    "shll   v14.4s, %26.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v20.4s      \n"
-                    "fmla   v29.4s, v15.4s, v21.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v22.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v23.4s      \n"
-
-                    "shll2  v15.4s, %23.8h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v25.4s, v14.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%6, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%6]      \n" // r44 r45 r46 r47
-                    "fmla   v27.4s, v14.4s, v19.4s      \n"
-
-                    "shll2  v14.4s, %26.8h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v16.4s      \n"
-                    "fmla   v29.4s, v15.4s, v17.4s      \n"
-                    "fmla   v30.4s, v15.4s, v18.4s      \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "shll   v15.4s, %24.4h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v17.4s      \n"
-                    "fmla   v25.4s, v14.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v19.4s      \n"
-                    "fmla   v27.4s, v14.4s, v20.4s      \n"
-                    "shll   v14.4s, %27.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v17.4s      \n"
-                    "fmla   v29.4s, v15.4s, v18.4s      \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-                    "shll2  v15.4s, %24.8h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v18.4s      \n"
-                    "fmla   v25.4s, v14.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v20.4s      \n"
-                    "fmla   v27.4s, v14.4s, v21.4s      \n"
-                    "shll2  v14.4s, %27.8h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v18.4s      \n"
-                    "fmla   v29.4s, v15.4s, v19.4s      \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-                    "shll   v15.4s, %25.4h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v19.4s      \n"
-                    "fmla   v25.4s, v14.4s, v20.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v21.4s      \n"
-                    "fmla   v27.4s, v14.4s, v22.4s      \n"
-                    "shll   v14.4s, %28.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v19.4s      \n"
-                    "fmla   v29.4s, v15.4s, v20.4s      \n"
-                    "fmla   v30.4s, v15.4s, v21.4s      \n"
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%2], #32 \n" // r00 r01 r02 r03
-                    "fmla   v31.4s, v15.4s, v22.4s      \n"
-
-                    "shll2  v15.4s, %25.8h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v20.4s      \n"
-                    "fmla   v25.4s, v14.4s, v21.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v22.4s      \n"
-                    "fmla   v27.4s, v14.4s, v23.4s      \n"
-
-                    "shll   v14.4s, %16.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v21.4s      \n"
-                    "fmla   v30.4s, v15.4s, v22.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v23.4s      \n"
-
-                    "shll2  v15.4s, %16.8h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v25.4s, v14.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%2]      \n" // r04 r05 r06 r07
-                    "fmla   v27.4s, v14.4s, v19.4s      \n"
-
-                    "shll   v14.4s, %17.4h, #16         \n"
-
-                    "fmla   v24.4s, v15.4s, v17.4s      \n"
-                    "fmla   v25.4s, v15.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v26.4s, v15.4s, v19.4s      \n"
-                    "fmla   v27.4s, v15.4s, v20.4s      \n"
-
-                    "shll2  v15.4s, %17.8h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v18.4s      \n"
-                    "fmla   v25.4s, v14.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v20.4s      \n"
-                    "fmla   v27.4s, v14.4s, v21.4s      \n"
-
-                    "shll   v14.4s, %18.4h, #16         \n"
-
-                    "fmla   v24.4s, v15.4s, v19.4s      \n"
-                    "fmla   v25.4s, v15.4s, v20.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v26.4s, v15.4s, v21.4s      \n"
-                    "prfm   pldl1keep, [%7, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%7], #32 \n" // r50 r51 r52 r53
-                    "fmla   v27.4s, v15.4s, v22.4s      \n"
-
-                    "shll   v15.4s, %26.4h, #16         \n"
-
-                    "fmla   v24.4s, v14.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v25.4s, v14.4s, v21.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v26.4s, v14.4s, v22.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v27.4s, v14.4s, v23.4s      \n"
-
-                    "shll2  v14.4s, %26.8h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%7, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%7]      \n" // r54 r55 r56 r57
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-
-                    "shll   v15.4s, %27.4h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "fmla   v29.4s, v14.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v19.4s      \n"
-                    "fmla   v31.4s, v14.4s, v20.4s      \n"
-
-                    "shll2  v14.4s, %27.8h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v18.4s      \n"
-                    "fmla   v29.4s, v15.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-
-                    "shll   v15.4s, %28.4h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v21.4s      \n"
-                    "fmla   v31.4s, v14.4s, v22.4s      \n"
-
-                    "fmla   v28.4s, v15.4s, v20.4s      \n"
-                    "fmla   v29.4s, v15.4s, v21.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v22.4s      \n"
-                    "fmla   v31.4s, v15.4s, v23.4s      \n"
-
-                    "shrn   v24.4h, v24.4s, #16         \n"
-                    "shrn   v25.4h, v25.4s, #16         \n"
-                    "shrn   v26.4h, v26.4s, #16         \n"
-                    "shrn   v27.4h, v27.4s, #16         \n"
-                    "shrn   v28.4h, v28.4s, #16         \n"
-                    "shrn   v29.4h, v29.4s, #16         \n"
-                    "shrn   v30.4h, v30.4s, #16         \n"
-                    "shrn   v31.4h, v31.4s, #16         \n"
-
-                    "st1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%0], #32 \n"
-                    "st1    {v28.4h, v29.4h, v30.4h, v31.4h}, [%1], #32 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3),      // %5
-                    "=r"(r4),      // %6
-                    "=r"(r5)       // %7
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "7"(r5),
-                    "w"(_k00_01), // %16
-                    "w"(_k02_03), // %17
-                    "w"(_k04_10), // %18
-                    "w"(_k11_12), // %19
-                    "w"(_k13_14), // %20
-                    "w"(_k20_21), // %21
-                    "w"(_k22_23), // %22
-                    "w"(_k24_30), // %23
-                    "w"(_k31_32), // %24
-                    "w"(_k33_34), // %25
-                    "w"(_k40_41), // %26
-                    "w"(_k42_43), // %27
-                    "w"(_k44),    // %28
-                    "w"(_bias0)   // %29
-                    : "memory", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%3, #128]       \n"
-                    "ld1    {v16.4h, v17.4h}, [%3], #16 \n" // r10 r11
-
-                    "shll2  v14.4s, %18.8h, #16         \n"
-
-                    "mov    v28.16b, %29.16b            \n" // sum00
-                    "mov    v29.16b, %29.16b            \n" // sum01
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "mov    v30.16b, %29.16b            \n" // sum10
-                    "mov    v31.16b, %29.16b            \n" // sum11
-
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v18.4h, v19.4h, v20.4h, v21.4h}, [%3] \n" // r12 r13 r14 r15
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v15.4s, %16.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v17.4s      \n"
-                    "shll   v14.4s, %19.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v17.4s      \n"
-                    "shll2  v15.4s, %16.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "fmla   v29.4s, v14.4s, v18.4s      \n"
-                    "shll2  v14.4s, %19.8h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v18.4s      \n"
-                    "shll   v15.4s, %17.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%4, #128]       \n"
-                    "ld1    {v16.4h, v17.4h}, [%4], #16 \n" // r20 r21
-
-                    "fmla   v29.4s, v14.4s, v19.4s      \n"
-                    "shll   v14.4s, %20.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "shll2  v15.4s, %17.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll2  v14.4s, %20.8h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-                    "shll   v15.4s, %18.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v21.4s      \n"
-                    "shll   v14.4s, %21.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v18.4h, v19.4h, v20.4h, v21.4h}, [%4] \n" // r22 r23 r24 r25
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll2  v15.4s, %18.8h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v17.4s      \n"
-                    "shll2  v14.4s, %21.8h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v17.4s      \n"
-                    "shll   v15.4s, %19.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "fmla   v29.4s, v14.4s, v18.4s      \n"
-                    "shll   v14.4s, %22.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v18.4s      \n"
-                    "shll2  v15.4s, %19.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%5, #128]       \n"
-                    "ld1    {v16.4h, v17.4h}, [%5], #16 \n" // r30 r31
-
-                    "fmla   v29.4s, v14.4s, v19.4s      \n"
-                    "shll2  v14.4s, %22.8h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "shll   v15.4s, %20.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll   v14.4s, %23.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-                    "shll2  v15.4s, %20.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v21.4s      \n"
-                    "shll2  v14.4s, %23.8h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v18.4h, v19.4h, v20.4h, v21.4h}, [%5] \n" // r32 r33 r34 r35
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v15.4s, %21.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v17.4s      \n"
-                    "shll   v14.4s, %24.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v17.4s      \n"
-                    "shll2  v15.4s, %21.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "fmla   v29.4s, v14.4s, v18.4s      \n"
-                    "shll2  v14.4s, %24.8h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v18.4s      \n"
-                    "shll   v15.4s, %22.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%6, #128]       \n"
-                    "ld1    {v16.4h, v17.4h}, [%6], #16 \n" // r40 r41
-
-                    "fmla   v29.4s, v14.4s, v19.4s      \n"
-                    "shll   v14.4s, %25.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "shll2  v15.4s, %22.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll2  v14.4s, %25.8h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-                    "shll   v15.4s, %23.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v21.4s      \n"
-                    "shll   v14.4s, %26.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-
-                    "prfm   pldl1keep, [%6, #256]       \n"
-                    "ld1    {v18.4h, v19.4h, v20.4h, v21.4h}, [%6] \n" // r42 r43 r44 r45
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll2  v15.4s, %23.8h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v17.4s      \n"
-                    "shll2  v14.4s, %26.8h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v17.4s      \n"
-                    "shll   v15.4s, %24.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "fmla   v29.4s, v14.4s, v18.4s      \n"
-                    "shll   v14.4s, %27.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v18.4s      \n"
-                    "shll2  v15.4s, %24.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "ld1    {v16.4h, v17.4h}, [%2], #16 \n" // r00 r01
-
-                    "fmla   v29.4s, v14.4s, v19.4s      \n"
-                    "shll2  v14.4s, %27.8h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "shll   v15.4s, %25.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "prfm   pldl1keep, [%7, #128]       \n"
-                    "ld1    {v22.4h, v23.4h}, [%7], #16 \n" // r50 r51
-                    "shll   v16.4s, v16.4h, #16         \n"
-
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll   v14.4s, %28.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-                    "shll2  v15.4s, %25.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v21.4s      \n"
-                    "shll   v14.4s, %16.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v18.4h, v19.4h, v20.4h, v21.4h}, [%2] \n" // r02 r03 r04 r05
-                    "shll   v23.4s, v23.4h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v15.4s, %26.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v17.4s      \n"
-                    "prfm   pldl1keep, [%7, #256]       \n"
-                    "ld1    {v24.4h, v25.4h, v26.4h, v27.4h}, [%7] \n" // r52 r53 r54 r55
-
-                    "shll2  v14.4s, %16.8h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v22.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v23.4s      \n"
-                    "shll2  v15.4s, %26.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "shll   v24.4s, v24.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v18.4s      \n"
-                    "shll   v14.4s, %17.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v23.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v24.4s      \n"
-                    "shll   v15.4s, %27.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "shll   v25.4s, v25.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v19.4s      \n"
-                    "shll2  v14.4s, %17.8h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v24.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v25.4s      \n"
-                    "shll2  v15.4s, %27.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "shll   v26.4s, v26.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll   v14.4s, %18.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v25.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v26.4s      \n"
-                    "shll   v15.4s, %28.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "fmla   v29.4s, v14.4s, v21.4s      \n"
-                    "shll   v27.4s, v27.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v26.4s      \n"
-                    "fmla   v31.4s, v15.4s, v27.4s      \n"
-
-                    "shrn   v28.4h, v28.4s, #16         \n"
-                    "shrn   v29.4h, v29.4s, #16         \n"
-                    "shrn   v30.4h, v30.4s, #16         \n"
-                    "shrn   v31.4h, v31.4s, #16         \n"
-
-                    "st1    {v28.4h, v29.4h}, [%0], #16 \n"
-                    "st1    {v30.4h, v31.4h}, [%1], #16 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3),      // %5
-                    "=r"(r4),      // %6
-                    "=r"(r5)       // %7
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "7"(r5),
-                    "w"(_k00_01), // %16
-                    "w"(_k02_03), // %17
-                    "w"(_k04_10), // %18
-                    "w"(_k11_12), // %19
-                    "w"(_k13_14), // %20
-                    "w"(_k20_21), // %21
-                    "w"(_k22_23), // %22
-                    "w"(_k24_30), // %23
-                    "w"(_k31_32), // %24
-                    "w"(_k33_34), // %25
-                    "w"(_k40_41), // %26
-                    "w"(_k42_43), // %27
-                    "w"(_k44),    // %28
-                    "w"(_bias0)   // %29
-                    : "memory", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            for (; j < outw; j++)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%3, #64]        \n"
-                    "ld1    {v16.4h}, [%3], #8          \n" // r10
-
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v17.4h, v18.4h, v19.4h, v20.4h}, [%3] \n" // r11 r12 r13 r14
-
-                    "mov    v30.16b, %29.16b            \n" // sum00
-                    "mov    v31.16b, %29.16b            \n" // sum10
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "shll2  v14.4s, %18.8h, #16         \n"
-                    "shll   v15.4s, %16.4h, #16         \n"
-                    "fmul   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v14.4s, %19.4h, #16         \n"
-                    "fmul   v29.4s, v15.4s, v16.4s      \n"
-                    "shll2  v15.4s, %16.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v17.4s      \n"
-                    "shll2  v14.4s, %19.8h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v17.4s      \n"
-                    "shll   v15.4s, %17.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "shll   v14.4s, %20.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v18.4s      \n"
-                    "shll2  v15.4s, %17.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v19.4s      \n"
-                    "shll2  v14.4s, %20.8h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "shll   v15.4s, %18.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "shll   v14.4s, %21.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v20.4s      \n"
-
-                    "prfm   pldl1keep, [%4, #64]        \n"
-                    "ld1    {v16.4h}, [%4], #8          \n" // r20
-
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v17.4h, v18.4h, v19.4h, v20.4h}, [%4] \n" // r21 r22 r23 r24
-
-                    "shll2  v15.4s, %18.8h, #16         \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "fmla   v30.4s, v14.4s, v16.4s      \n"
-                    "shll2  v14.4s, %21.8h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v16.4s      \n"
-                    "shll   v15.4s, %19.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "shll   v14.4s, %22.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v17.4s      \n"
-                    "shll2  v15.4s, %19.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v18.4s      \n"
-                    "shll2  v14.4s, %22.8h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v18.4s      \n"
-                    "shll   v15.4s, %20.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "shll   v14.4s, %23.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v19.4s      \n"
-                    "shll2  v15.4s, %20.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v20.4s      \n"
-                    "shll2  v14.4s, %23.8h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-
-                    "prfm   pldl1keep, [%5, #64]        \n"
-                    "ld1    {v16.4h}, [%5], #8          \n" // r30
-
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v17.4h, v18.4h, v19.4h, v20.4h}, [%5] \n" // r31 r32 r33 r34
-
-                    "shll   v15.4s, %21.4h, #16         \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v14.4s, %24.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v16.4s      \n"
-                    "shll2  v15.4s, %21.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v17.4s      \n"
-                    "shll2  v14.4s, %24.8h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v17.4s      \n"
-                    "shll   v15.4s, %22.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "shll   v14.4s, %25.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v18.4s      \n"
-                    "shll2  v15.4s, %22.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v19.4s      \n"
-                    "shll2  v14.4s, %25.8h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "shll   v15.4s, %23.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "shll   v14.4s, %26.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v20.4s      \n"
-
-                    "prfm   pldl1keep, [%6, #64]        \n"
-                    "ld1    {v16.4h}, [%6], #8          \n" // r40
-
-                    "prfm   pldl1keep, [%6, #256]       \n"
-                    "ld1    {v17.4h, v18.4h, v19.4h, v20.4h}, [%6] \n" // r41 r42 r43 r44
-
-                    "shll2  v15.4s, %23.8h, #16         \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "fmla   v30.4s, v14.4s, v16.4s      \n"
-                    "shll2  v14.4s, %26.8h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v16.4s      \n"
-                    "shll   v15.4s, %24.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "shll   v14.4s, %27.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v17.4s      \n"
-                    "shll2  v15.4s, %24.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v18.4s      \n"
-                    "shll2  v14.4s, %27.8h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v18.4s      \n"
-                    "shll   v15.4s, %25.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "shll   v14.4s, %28.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v19.4s      \n"
-                    "shll2  v15.4s, %25.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v20.4s      \n"
-                    "shll   v14.4s, %16.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #64]        \n"
-                    "ld1    {v16.4h}, [%2], #8          \n" // r00
-
-                    "prfm   pldl1keep, [%7, #64]        \n"
-                    "ld1    {v21.4h}, [%7], #8          \n" // r50
-
-                    "prfm   pldl1keep, [%7, #256]       \n"
-                    "ld1    {v22.4h, v23.4h, v24.4h, v25.4h}, [%7] \n" // r51 r52 r53 r54
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v17.4h, v18.4h, v19.4h, v20.4h}, [%2] \n" // r01 r02 r03 r04
-
-                    "shll   v15.4s, %26.4h, #16         \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "shll   v24.4s, v24.4h, #16         \n"
-                    "shll   v25.4s, v25.4h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll2  v14.4s, %16.8h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v21.4s      \n"
-                    "shll2  v15.4s, %26.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v17.4s      \n"
-                    "shll   v14.4s, %17.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v22.4s      \n"
-                    "shll   v15.4s, %27.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "shll2  v14.4s, %17.8h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v23.4s      \n"
-                    "shll2  v15.4s, %27.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v19.4s      \n"
-                    "shll   v14.4s, %18.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v24.4s      \n"
-                    "shll   v15.4s, %28.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "fmla   v29.4s, v15.4s, v25.4s      \n"
-
-                    "fadd   v30.4s, v30.4s, v28.4s      \n"
-                    "fadd   v31.4s, v31.4s, v29.4s      \n"
-
-                    "shrn   v30.4h, v30.4s, #16         \n"
-                    "shrn   v31.4h, v31.4s, #16         \n"
-
-                    "st1    {v30.4h}, [%0], #8          \n"
-                    "st1    {v31.4h}, [%1], #8          \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3),      // %5
-                    "=r"(r4),      // %6
-                    "=r"(r5)       // %7
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "7"(r5),
-                    "w"(_k00_01), // %16
-                    "w"(_k02_03), // %17
-                    "w"(_k04_10), // %18
-                    "w"(_k11_12), // %19
-                    "w"(_k13_14), // %20
-                    "w"(_k20_21), // %21
-                    "w"(_k22_23), // %22
-                    "w"(_k24_30), // %23
-                    "w"(_k31_32), // %24
-                    "w"(_k33_34), // %25
-                    "w"(_k40_41), // %26
-                    "w"(_k42_43), // %27
-                    "w"(_k44),    // %28
-                    "w"(_bias0)   // %29
-                    : "memory", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-
-            r0 += 4 * 4 + w * 4;
-            r1 += 4 * 4 + w * 4;
-            r2 += 4 * 4 + w * 4;
-            r3 += 4 * 4 + w * 4;
-            r4 += 4 * 4 + w * 4;
-            r5 += 4 * 4 + w * 4;
-
-            outptr0 += outw * 4;
-            outptr1 += outw * 4;
-        }
-#endif // __aarch64__
-        for (; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j + 3 < outw; j += 4)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%1], #32 \n" // r00 r01 r02 r03
-
-                    "shll   v14.4s, %12.4h, #16         \n"
-
-                    "mov    v28.16b, %25.16b            \n" // sum00
-                    "mov    v29.16b, %25.16b            \n" // sum01
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "mov    v30.16b, %25.16b            \n" // sum02
-                    "mov    v31.16b, %25.16b            \n" // sum03
-
-                    "shll2  v15.4s, %12.8h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%1]      \n" // r04 r05 r06 r07
-                    "fmla   v31.4s, v14.4s, v19.4s      \n"
-
-                    "shll   v14.4s, %13.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v17.4s      \n"
-                    "fmla   v29.4s, v15.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-
-                    "shll2  v15.4s, %13.8h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "fmla   v29.4s, v14.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v20.4s      \n"
-                    "fmla   v31.4s, v14.4s, v21.4s      \n"
-
-                    "shll   v14.4s, %14.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v19.4s      \n"
-                    "fmla   v29.4s, v15.4s, v20.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v21.4s      \n"
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%2], #32 \n" // r10 r11 r12 r13
-                    "fmla   v31.4s, v15.4s, v22.4s      \n"
-
-                    "shll2  v15.4s, %14.8h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v21.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v22.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v31.4s, v14.4s, v23.4s      \n"
-
-                    "shll   v14.4s, %15.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%2]      \n" // r14 r15 r16 r17
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-
-                    "shll2  v15.4s, %15.8h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "fmla   v29.4s, v14.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v19.4s      \n"
-                    "fmla   v31.4s, v14.4s, v20.4s      \n"
-
-                    "shll   v14.4s, %16.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v18.4s      \n"
-                    "fmla   v29.4s, v15.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-
-                    "shll2  v15.4s, %16.8h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v21.4s      \n"
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%3], #32 \n" // r20 r21 r22 r23
-                    "fmla   v31.4s, v14.4s, v22.4s      \n"
-
-                    "shll   v14.4s, %17.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v21.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v22.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v23.4s      \n"
-
-                    "shll2  v15.4s, %17.8h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%3]      \n" // r24 r25 r26 r27
-                    "fmla   v31.4s, v14.4s, v19.4s      \n"
-
-                    "shll   v14.4s, %18.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v17.4s      \n"
-                    "fmla   v29.4s, v15.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-
-                    "shll2  v15.4s, %18.8h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "fmla   v29.4s, v14.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v20.4s      \n"
-                    "fmla   v31.4s, v14.4s, v21.4s      \n"
-
-                    "shll   v14.4s, %19.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v19.4s      \n"
-                    "fmla   v29.4s, v15.4s, v20.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v21.4s      \n"
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%4], #32 \n" // r30 r31 r32 r33
-                    "fmla   v31.4s, v15.4s, v22.4s      \n"
-
-                    "shll2  v15.4s, %19.8h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v21.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v22.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v31.4s, v14.4s, v23.4s      \n"
-
-                    "shll   v14.4s, %20.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%4]      \n" // r34 r35 r36 r37
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-
-                    "shll2  v15.4s, %20.8h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "fmla   v29.4s, v14.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v19.4s      \n"
-                    "fmla   v31.4s, v14.4s, v20.4s      \n"
-
-                    "shll   v14.4s, %21.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v18.4s      \n"
-                    "fmla   v29.4s, v15.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-
-                    "shll2  v15.4s, %21.8h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v21.4s      \n"
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%5], #32 \n" // r40 r41 r42 r43
-                    "fmla   v31.4s, v14.4s, v22.4s      \n"
-
-                    "shll   v14.4s, %22.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v21.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v22.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v23.4s      \n"
-
-                    "shll2  v15.4s, %22.8h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%5]      \n" // r44 r45 r46 r47
-                    "fmla   v31.4s, v14.4s, v19.4s      \n"
-
-                    "shll   v14.4s, %23.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v17.4s      \n"
-                    "fmla   v29.4s, v15.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-
-                    "shll2  v15.4s, %23.8h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "fmla   v29.4s, v14.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v20.4s      \n"
-                    "fmla   v31.4s, v14.4s, v21.4s      \n"
-
-                    "shll   v14.4s, %24.4h, #16         \n"
-
-                    "fmla   v28.4s, v15.4s, v19.4s      \n"
-                    "fmla   v29.4s, v15.4s, v20.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v21.4s      \n"
-                    "fmla   v31.4s, v15.4s, v22.4s      \n"
-
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "fmla   v29.4s, v14.4s, v21.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v22.4s      \n"
-                    "fmla   v31.4s, v14.4s, v23.4s      \n"
-
-                    "shrn   v28.4h, v28.4s, #16         \n"
-                    "shrn   v29.4h, v29.4s, #16         \n"
-                    "shrn   v30.4h, v30.4s, #16         \n"
-                    "shrn   v31.4h, v31.4s, #16         \n"
-
-                    "st1    {v28.4h, v29.4h, v30.4h, v31.4h}, [%0], #32 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2),      // %3
-                    "=r"(r3),      // %4
-                    "=r"(r4)       // %5
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "4"(r3),
-                    "5"(r4),
-                    "w"(_k00_01), // %12
-                    "w"(_k02_03), // %13
-                    "w"(_k04_10), // %14
-                    "w"(_k11_12), // %15
-                    "w"(_k13_14), // %16
-                    "w"(_k20_21), // %17
-                    "w"(_k22_23), // %18
-                    "w"(_k24_30), // %19
-                    "w"(_k31_32), // %20
-                    "w"(_k33_34), // %21
-                    "w"(_k40_41), // %22
-                    "w"(_k42_43), // %23
-                    "w"(_k44),    // %24
-                    "w"(_bias0)   // %25
-                    : "memory", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                    "pld        [%2, #256]          \n"
-                    "vld1.u16   {d4-d7}, [%2 :64]!  \n" // r00 r01 r02 r03
-
-                    "vshll.u16  q8, d20, #16        \n" // k00
-
-                    "pld        [%1, #128]          \n"
-                    "vld1.f32   {d24-d25}, [%1]     \n"
-
-                    "vshll.u16  q0, d4, #16         \n"
-                    "vshll.u16  q1, d5, #16         \n"
-
-                    "vmov       q13, q12            \n" // sum0 sum1
-                    "vmov       q14, q12            \n"
-
-                    "vshll.u16  q9, d21, #16        \n" // k01
-
-                    "vmov       q15, q12            \n" // sum2 sum3
-
-                    "vmla.f32   q12, q8, q0         \n"
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vmla.f32   q13, q8, q1         \n"
-                    "vshll.u16  q3, d7, #16         \n"
-                    "vmla.f32   q14, q8, q2         \n"
-                    "pld        [%2, #256]          \n"
-                    "vld1.u16   {d12-d15}, [%2 :64] \n" // r04 r05 r06 r07
-
-                    "vmla.f32   q15, q8, q3         \n"
-
-                    "vshll.u16  q10, d22, #16       \n" // k02
-                    "vmla.f32   q12, q9, q1         \n"
-                    "vmla.f32   q13, q9, q2         \n"
-                    "vshll.u16  q4, d12, #16        \n"
-                    "vmla.f32   q14, q9, q3         \n"
-                    "vmla.f32   q15, q9, q4         \n"
-
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                    "vshll.u16  q11, d23, #16       \n" // k03
-                    "vmla.f32   q12, q10, q2        \n"
-                    "vmla.f32   q13, q10, q3        \n"
-                    "vshll.u16  q5, d13, #16        \n"
-                    "vmla.f32   q14, q10, q4        \n"
-                    "vmla.f32   q15, q10, q5        \n"
-
-                    "vshll.u16  q10, d16, #16       \n" // k04
-                    "vmla.f32   q12, q11, q3        \n"
-                    "vmla.f32   q13, q11, q4        \n"
-                    "vshll.u16  q6, d14, #16        \n"
-                    "vmla.f32   q14, q11, q5        \n"
-                    "pld        [%3, #256]          \n"
-                    "vld1.u16   {d4-d7}, [%3 :64]!  \n" // r10 r11 r12 r13
-
-                    "vmla.f32   q15, q11, q6        \n"
-
-                    "vshll.u16  q11, d17, #16       \n" // k10
-
-                    "vmla.f32   q12, q10, q4        \n"
-                    "vshll.u16  q0, d4, #16         \n"
-                    "vmla.f32   q13, q10, q5        \n"
-                    "vshll.u16  q7, d15, #16        \n"
-                    "vmla.f32   q14, q10, q6        \n"
-                    "vshll.u16  q1, d5, #16         \n"
-                    "vmla.f32   q15, q10, q7        \n"
-
-                    "vshll.u16  q8, d18, #16        \n" // k11
-                    "vmla.f32   q12, q11, q0        \n"
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vmla.f32   q13, q11, q1        \n"
-                    "vshll.u16  q3, d7, #16         \n"
-                    "vmla.f32   q14, q11, q2        \n"
-                    "pld        [%3, #256]          \n"
-                    "vld1.u16   {d12-d15}, [%3 :64] \n" // r14 r15 r16 r17
-
-                    "vmla.f32   q15, q11, q3        \n"
-
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                    "vshll.u16  q9, d19, #16        \n" // k12
-                    "vmla.f32   q12, q8, q1         \n"
-                    "vmla.f32   q13, q8, q2         \n"
-                    "vshll.u16  q4, d12, #16        \n"
-                    "vmla.f32   q14, q8, q3         \n"
-                    "vmla.f32   q15, q8, q4         \n"
-
-                    "vshll.u16  q8, d20, #16        \n" // k13
-                    "vmla.f32   q12, q9, q2         \n"
-                    "vmla.f32   q13, q9, q3         \n"
-                    "vshll.u16  q5, d13, #16        \n"
-                    "vmla.f32   q14, q9, q4         \n"
-                    "vmla.f32   q15, q9, q5         \n"
-
-                    "vshll.u16  q9, d21, #16        \n" // k14
-
-                    "vmla.f32   q12, q8, q3         \n"
-                    "vmla.f32   q13, q8, q4         \n"
-                    "vshll.u16  q6, d14, #16        \n"
-                    "vmla.f32   q14, q8, q5         \n"
-                    "pld        [%4, #256]          \n"
-                    "vld1.u16   {d4-d7}, [%4 :64]!  \n" // r20 r21 r22 r23
-
-                    "vmla.f32   q15, q8, q6         \n"
-
-                    "vshll.u16  q10, d22, #16       \n" // k20
-                    "vmla.f32   q12, q9, q4         \n"
-                    "vshll.u16  q0, d4, #16         \n"
-                    "vmla.f32   q13, q9, q5         \n"
-                    "vshll.u16  q7, d15, #16        \n"
-                    "vmla.f32   q14, q9, q6         \n"
-                    "vshll.u16  q1, d5, #16         \n"
-                    "vmla.f32   q15, q9, q7         \n"
-
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                    "vshll.u16  q11, d23, #16       \n" // k21
-                    "vmla.f32   q12, q10, q0        \n"
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vmla.f32   q13, q10, q1        \n"
-                    "vshll.u16  q3, d7, #16         \n"
-                    "vmla.f32   q14, q10, q2        \n"
-                    "pld        [%4, #256]          \n"
-                    "vld1.u16   {d12-d15}, [%4 :64] \n" // r24 r25 r26 r27
-
-                    "vmla.f32   q15, q10, q3        \n"
-
-                    "vshll.u16  q10, d16, #16       \n" // k22
-                    "vmla.f32   q12, q11, q1        \n"
-                    "vmla.f32   q13, q11, q2        \n"
-                    "vshll.u16  q4, d12, #16        \n"
-                    "vmla.f32   q14, q11, q3        \n"
-                    "vmla.f32   q15, q11, q4        \n"
-
-                    "vshll.u16  q11, d17, #16       \n" // k23
-
-                    "vmla.f32   q12, q10, q2        \n"
-                    "vmla.f32   q13, q10, q3        \n"
-                    "vshll.u16  q5, d13, #16        \n"
-                    "vmla.f32   q14, q10, q4        \n"
-                    "vmla.f32   q15, q10, q5        \n"
-
-                    "vshll.u16  q8, d18, #16        \n" // k24
-                    "vmla.f32   q12, q11, q3        \n"
-                    "vmla.f32   q13, q11, q4        \n"
-                    "vshll.u16  q6, d14, #16        \n"
-                    "vmla.f32   q14, q11, q5        \n"
-                    "pld        [%5, #256]          \n"
-                    "vld1.u16   {d4-d7}, [%5 :64]!  \n" // r30 r31 r32 r33
-
-                    "vmla.f32   q15, q11, q6        \n"
-
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                    "vshll.u16  q9, d19, #16        \n" // k30
-                    "vmla.f32   q12, q8, q4         \n"
-                    "vshll.u16  q0, d4, #16         \n"
-                    "vmla.f32   q13, q8, q5         \n"
-                    "vshll.u16  q7, d15, #16        \n"
-                    "vmla.f32   q14, q8, q6         \n"
-                    "vshll.u16  q1, d5, #16         \n"
-                    "vmla.f32   q15, q8, q7         \n"
-
-                    "vshll.u16  q8, d20, #16        \n" // k31
-                    "vmla.f32   q12, q9, q0         \n"
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vmla.f32   q13, q9, q1         \n"
-                    "vshll.u16  q3, d7, #16         \n"
-                    "vmla.f32   q14, q9, q2         \n"
-                    "pld        [%5, #256]          \n"
-                    "vld1.u16   {d12-d15}, [%5 :64] \n" // r34 r35 r36 r37
-
-                    "vmla.f32   q15, q9, q3         \n"
-
-                    "vshll.u16  q9, d21, #16        \n" // k32
-
-                    "vmla.f32   q12, q8, q1         \n"
-                    "vmla.f32   q13, q8, q2         \n"
-                    "vshll.u16  q4, d12, #16        \n"
-                    "vmla.f32   q14, q8, q3         \n"
-                    "vmla.f32   q15, q8, q4         \n"
-
-                    "vshll.u16  q10, d22, #16       \n" // k33
-                    "vmla.f32   q12, q9, q2         \n"
-                    "vmla.f32   q13, q9, q3         \n"
-                    "vshll.u16  q5, d13, #16        \n"
-                    "vmla.f32   q14, q9, q4         \n"
-                    "vmla.f32   q15, q9, q5         \n"
-
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                    "vmla.f32   q12, q10, q3        \n"
-                    "vshll.u16  q11, d23, #16       \n" // k34
-                    "vmla.f32   q13, q10, q4        \n"
-                    "vshll.u16  q6, d14, #16        \n"
-                    "vmla.f32   q14, q10, q5        \n"
-                    "pld        [%6, #256]          \n"
-                    "vld1.u16   {d4-d7}, [%6 :64]!  \n" // r40 r41 r42 r43
-
-                    "vmla.f32   q15, q10, q6        \n"
-
-                    "vshll.u16  q10, d16, #16       \n" // k40
-                    "vmla.f32   q12, q11, q4        \n"
-                    "vshll.u16  q0, d4, #16         \n"
-                    "vmla.f32   q13, q11, q5        \n"
-                    "vshll.u16  q7, d15, #16        \n"
-                    "vmla.f32   q14, q11, q6        \n"
-                    "vshll.u16  q1, d5, #16         \n"
-                    "vmla.f32   q15, q11, q7        \n"
-
-                    "vshll.u16  q11, d17, #16       \n" // k41
-
-                    "vmla.f32   q12, q10, q0        \n"
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vmla.f32   q13, q10, q1        \n"
-                    "vshll.u16  q3, d7, #16         \n"
-                    "vmla.f32   q14, q10, q2        \n"
-                    "pld        [%6, #256]          \n"
-                    "vld1.u16   {d12-d15}, [%6 :64] \n" // r44 r45 r46 r47
-
-                    "vmla.f32   q15, q10, q3        \n"
-
-                    "vshll.u16  q8, d18, #16        \n" // k42
-                    "vmla.f32   q12, q11, q1        \n"
-                    "vmla.f32   q13, q11, q2        \n"
-                    "vshll.u16  q4, d12, #16        \n"
-                    "vmla.f32   q14, q11, q3        \n"
-                    "vmla.f32   q15, q11, q4        \n"
-
-                    "pld        [%7, #64]           \n"
-                    "vld1.u16   {d20}, [%7 :64]     \n"
-
-                    "vmla.f32   q12, q8, q2         \n"
-                    "vshll.u16  q9, d19, #16        \n" // k43
-                    "vmla.f32   q13, q8, q3         \n"
-                    "vshll.u16  q5, d13, #16        \n"
-                    "vmla.f32   q14, q8, q4         \n"
-                    "vmla.f32   q15, q8, q5         \n"
-
-                    "vshll.u16  q8, d20, #16        \n" // k44
-
-                    "vmla.f32   q12, q9, q3         \n"
-                    "vmla.f32   q13, q9, q4         \n"
-                    "vshll.u16  q6, d14, #16        \n"
-                    "vmla.f32   q14, q9, q5         \n"
-                    "vmla.f32   q15, q9, q6         \n"
-
-                    "vmla.f32   q12, q8, q4         \n"
-                    "vmla.f32   q13, q8, q5         \n"
-                    "vshll.u16  q7, d15, #16        \n"
-                    "vmla.f32   q14, q8, q6         \n"
-                    "vmla.f32   q15, q8, q7         \n"
-
-                    "sub        %7, %7, #192        \n" // kptr -= 24 * 4;
-
-                    "vshrn.u32  d24, q12, #16       \n"
-                    "vshrn.u32  d25, q13, #16       \n"
-                    "vshrn.u32  d26, q14, #16       \n"
-                    "vshrn.u32  d27, q15, #16       \n"
-
-                    "vst1.u16   {d24-d27}, [%0 :64]! \n"
-
-                    : "=r"(outptr0),        // %0
-                    "=r"(bias0_data_ptr), // %1
-                    "=r"(r0),             // %2
-                    "=r"(r1),             // %3
-                    "=r"(r2),             // %4
-                    "=r"(r3),             // %5
-                    "=r"(r4),             // %6
-                    "=r"(kptr)            // %7
-                    : "0"(outptr0),
-                    "1"(bias0_data_ptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "7"(kptr)
-                    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #128]       \n"
-                    "ld1    {v16.4h, v17.4h}, [%1], #16 \n" // r00 r01
-
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v18.4h, v19.4h, v20.4h, v21.4h}, [%1] \n" // r02 r03 r04 r05
-
-                    "shll   v14.4s, %12.4h, #16         \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "mov    v30.16b, %25.16b            \n" // sum01
-                    "mov    v31.16b, %25.16b            \n" // sum02
-
-                    "shll2  v15.4s, %12.8h, #16         \n"
-                    "fmul   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmul   v29.4s, v14.4s, v17.4s      \n"
-                    "shll   v14.4s, %13.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "ld1    {v16.4h, v17.4h}, [%2], #16 \n" // r10 r11
-                    "shll2  v15.4s, %13.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v19.4s      \n"
-                    "shll   v14.4s, %14.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-                    "shll2  v15.4s, %14.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v21.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v18.4h, v19.4h, v20.4h, v21.4h}, [%2] \n" // r12 r13 r14 r15
-
-                    "shll   v14.4s, %15.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "fmla   v30.4s, v15.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v17.4s      \n"
-                    "shll2  v15.4s, %15.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%3, #128]       \n"
-                    "ld1    {v16.4h, v17.4h}, [%3], #16 \n" // r20 r21
-                    "shll   v14.4s, %16.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "shll2  v15.4s, %16.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll   v14.4s, %17.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v18.4h, v19.4h, v20.4h, v21.4h}, [%3] \n" // r22 r23 r24 r25
-
-                    "shll2  v15.4s, %17.8h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v17.4s      \n"
-                    "shll   v14.4s, %18.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%4, #128]       \n"
-                    "ld1    {v16.4h, v17.4h}, [%4], #16 \n" // r30 r31
-                    "shll2  v15.4s, %18.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v19.4s      \n"
-                    "shll   v14.4s, %19.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-                    "shll2  v15.4s, %19.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v21.4s      \n"
-
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v18.4h, v19.4h, v20.4h, v21.4h}, [%4] \n" // r32 r33 r34 r35
-
-                    "shll   v14.4s, %20.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "fmla   v30.4s, v15.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v17.4s      \n"
-                    "shll2  v15.4s, %20.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v18.4s      \n"
-                    "prfm   pldl1keep, [%5, #128]       \n"
-                    "ld1    {v16.4h, v17.4h}, [%5], #16 \n" // r40 r41
-                    "shll   v14.4s, %21.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "shll2  v15.4s, %21.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll   v14.4s, %22.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v18.4h, v19.4h, v20.4h, v21.4h}, [%5] \n" // r42 r43 r44 r45
-
-                    "shll2  v15.4s, %22.8h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v17.4s      \n"
-                    "shll   v14.4s, %23.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v17.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v18.4s      \n"
-                    "shll2  v15.4s, %23.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v19.4s      \n"
-                    "shll   v14.4s, %24.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "fmla   v29.4s, v14.4s, v21.4s      \n"
-
-                    "fadd   v30.4s, v30.4s, v28.4s      \n"
-                    "fadd   v31.4s, v31.4s, v29.4s      \n"
-
-                    "shrn   v30.4h, v30.4s, #16         \n"
-                    "shrn   v31.4h, v31.4s, #16         \n"
-
-                    "st1    {v30.4h, v31.4h}, [%0], #16 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2),      // %3
-                    "=r"(r3),      // %4
-                    "=r"(r4)       // %5
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "4"(r3),
-                    "5"(r4),
-                    "w"(_k00_01), // %12
-                    "w"(_k02_03), // %13
-                    "w"(_k04_10), // %14
-                    "w"(_k11_12), // %15
-                    "w"(_k13_14), // %16
-                    "w"(_k20_21), // %17
-                    "w"(_k22_23), // %18
-                    "w"(_k24_30), // %19
-                    "w"(_k31_32), // %20
-                    "w"(_k33_34), // %21
-                    "w"(_k40_41), // %22
-                    "w"(_k42_43), // %23
-                    "w"(_k44),    // %24
-                    "w"(_bias0)   // %25
-                    : "memory", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                    "pld        [%2, #128]          \n"
-                    "vld1.u16   {d2-d3}, [%2 :64]!  \n" // r00 r01
-
-                    "vshll.u16  q8, d20, #16        \n" // k00
-
-                    "pld        [%1, #128]          \n"
-                    "vld1.f32   {d24-d25}, [%1]     \n"
-
-                    "pld        [%2, #256]          \n"
-                    "vld1.u16   {d8-d11}, [%2 :64]  \n" // r02 r03 r04 r05
-
-                    "vshll.u16  q0, d2, #16         \n"
-
-                    "vmov       q13, q12            \n" // sum0 sum1
-
-                    "vshll.u16  q1, d3, #16         \n"
-                    "vshll.u16  q9, d21, #16        \n" // k01
-                    "vmul.f32   q14, q8, q0         \n"
-                    "vshll.u16  q2, d8, #16         \n"
-                    "vmul.f32   q15, q8, q1         \n"
-                    "vshll.u16  q10, d22, #16       \n" // k02
-                    "vmla.f32   q12, q9, q1         \n"
-                    "pld        [%3, #128]          \n"
-                    "vld1.u16   {d2-d3}, [%3 :64]!  \n" // r10 r11
-
-                    "vshll.u16  q3, d9, #16         \n"
-                    "vmla.f32   q13, q9, q2         \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-                    "vshll.u16  q11, d23, #16       \n" // k03
-                    "vmla.f32   q14, q10, q2        \n"
-                    "vshll.u16  q4, d10, #16        \n"
-                    "vmla.f32   q15, q10, q3        \n"
-                    "vshll.u16  q10, d16, #16       \n" // k04
-                    "vmla.f32   q12, q11, q3        \n"
-                    "vshll.u16  q5, d11, #16        \n"
-                    "vmla.f32   q13, q11, q4        \n"
-                    "vshll.u16  q11, d17, #16       \n" // k10
-                    "vmla.f32   q14, q10, q4        \n"
-                    "vshll.u16  q0, d2, #16         \n"
-                    "vmla.f32   q15, q10, q5        \n"
-
-                    "pld        [%3, #256]          \n"
-                    "vld1.u16   {d8-d11}, [%3 :64]  \n" // r12 r13 r14 r15
-
-                    "vshll.u16  q1, d3, #16         \n"
-                    "vshll.u16  q8, d18, #16        \n" // k11
-                    "vmla.f32   q12, q11, q0        \n"
-                    "vshll.u16  q2, d8, #16         \n"
-                    "vmla.f32   q13, q11, q1        \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-                    "vshll.u16  q9, d19, #16        \n" // k12
-                    "vmla.f32   q14, q8, q1         \n"
-                    "pld        [%4, #128]          \n"
-                    "vld1.u16   {d2-d3}, [%4 :64]!  \n" // r20 r21
-
-                    "vshll.u16  q3, d9, #16         \n"
-                    "vmla.f32   q15, q8, q2         \n"
-                    "vshll.u16  q8, d20, #16        \n" // k13
-                    "vmla.f32   q12, q9, q2         \n"
-                    "vshll.u16  q4, d10, #16        \n"
-                    "vmla.f32   q13, q9, q3         \n"
-                    "vshll.u16  q9, d21, #16        \n" // k14
-                    "vmla.f32   q14, q8, q3         \n"
-                    "vshll.u16  q5, d11, #16        \n"
-                    "vmla.f32   q15, q8, q4         \n"
-                    "vshll.u16  q10, d22, #16       \n" // k20
-                    "vmla.f32   q12, q9, q4         \n"
-                    "vshll.u16  q0, d2, #16         \n"
-                    "vmla.f32   q13, q9, q5         \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                    "pld        [%4, #256]          \n"
-                    "vld1.u16   {d8-d11}, [%4 :64]  \n" // r22 r23 r24 r25
-
-                    "vshll.u16  q1, d3, #16         \n"
-                    "vshll.u16  q11, d23, #16       \n" // k21
-                    "vmla.f32   q14, q10, q0        \n"
-                    "vshll.u16  q2, d8, #16         \n"
-                    "vmla.f32   q15, q10, q1        \n"
-                    "vshll.u16  q10, d16, #16       \n" // k22
-                    "vmla.f32   q12, q11, q1        \n"
-                    "pld        [%5, #128]          \n"
-                    "vld1.u16   {d2-d3}, [%5 :64]!  \n" // r30 r31
-
-                    "vshll.u16  q3, d9, #16         \n"
-                    "vmla.f32   q13, q11, q2        \n"
-                    "vshll.u16  q11, d17, #16       \n" // k23
-                    "vmla.f32   q14, q10, q2        \n"
-                    "vshll.u16  q4, d10, #16        \n"
-                    "vmla.f32   q15, q10, q3        \n"
-                    "vshll.u16  q8, d18, #16        \n" // k24
-                    "vmla.f32   q12, q11, q3        \n"
-                    "vshll.u16  q5, d11, #16        \n"
-                    "vmla.f32   q13, q11, q4        \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-                    "vshll.u16  q9, d19, #16        \n" // k30
-                    "vmla.f32   q14, q8, q4         \n"
-                    "vshll.u16  q0, d2, #16         \n"
-                    "vmla.f32   q15, q8, q5         \n"
-
-                    "pld        [%5, #256]          \n"
-                    "vld1.u16   {d8-d11}, [%5 :64]  \n" // r32 r33 r34 r35
-
-                    "vshll.u16  q1, d3, #16         \n"
-                    "vshll.u16  q8, d20, #16        \n" // k31
-                    "vmla.f32   q12, q9, q0         \n"
-                    "vshll.u16  q2, d8, #16         \n"
-                    "vmla.f32   q13, q9, q1         \n"
-                    "vshll.u16  q9, d21, #16        \n" // k32
-                    "vmla.f32   q14, q8, q1         \n"
-                    "pld        [%6, #128]          \n"
-                    "vld1.u16   {d2-d3}, [%6 :64]!  \n" // r40 r41
-
-                    "vshll.u16  q3, d9, #16         \n"
-                    "vmla.f32   q15, q8, q2         \n"
-                    "vshll.u16  q10, d22, #16       \n" // k33
-                    "vmla.f32   q12, q9, q2         \n"
-                    "vshll.u16  q4, d10, #16        \n"
-                    "vmla.f32   q13, q9, q3         \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-                    "vshll.u16  q11, d23, #16       \n" // k34
-                    "vmla.f32   q14, q10, q3        \n"
-                    "vshll.u16  q5, d11, #16        \n"
-                    "vmla.f32   q15, q10, q4        \n"
-                    "vshll.u16  q10, d16, #16       \n" // k40
-                    "vmla.f32   q12, q11, q4        \n"
-                    "vshll.u16  q0, d2, #16         \n"
-                    "vmla.f32   q13, q11, q5        \n"
-
-                    "pld        [%6, #256]          \n"
-                    "vld1.u16   {d8-d11}, [%6 :64]  \n" // r42 r43 r44 r45
-
-                    "vshll.u16  q1, d3, #16         \n"
-                    "vshll.u16  q11, d17, #16       \n" // k41
-                    "vmla.f32   q14, q10, q0        \n"
-                    "vshll.u16  q2, d8, #16         \n"
-                    "vmla.f32   q15, q10, q1        \n"
-                    "vshll.u16  q8, d18, #16        \n" // k42
-                    "vmla.f32   q12, q11, q1        \n"
-                    "vshll.u16  q3, d9, #16         \n"
-                    "vmla.f32   q13, q11, q2        \n"
-                    "pld        [%7, #64]           \n"
-                    "vld1.u16   {d20}, [%7 :64]     \n"
-                    "vshll.u16  q9, d19, #16        \n" // k43
-                    "vmla.f32   q14, q8, q2         \n"
-                    "vshll.u16  q4, d10, #16        \n"
-                    "vmla.f32   q15, q8, q3         \n"
-                    "vshll.u16  q8, d20, #16        \n" // k44
-                    "vmla.f32   q12, q9, q3         \n"
-                    "vmla.f32   q13, q9, q4         \n"
-                    "vshll.u16  q5, d11, #16        \n"
-                    "vmla.f32   q14, q8, q4         \n"
-                    "vmla.f32   q15, q8, q5         \n"
-
-                    "vadd.f32   q12, q12, q14       \n"
-                    "vadd.f32   q13, q13, q15       \n"
-
-                    "sub        %7, %7, #192        \n" // kptr -= 24 * 4;
-
-                    "vshrn.u32  d24, q12, #16       \n"
-                    "vshrn.u32  d25, q13, #16       \n"
-
-                    "vst1.u16   {d24-d25}, [%0 :64]! \n"
-
-                    : "=r"(outptr0),        // %0
-                    "=r"(bias0_data_ptr), // %1
-                    "=r"(r0),             // %2
-                    "=r"(r1),             // %3
-                    "=r"(r2),             // %4
-                    "=r"(r3),             // %5
-                    "=r"(r4),             // %6
-                    "=r"(kptr)            // %7
-                    : "0"(outptr0),
-                    "1"(bias0_data_ptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "7"(kptr)
-                    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-            }
-            for (; j < outw; j++)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #64]        \n"
-                    "ld1    {v16.4h}, [%1], #8          \n" // r00
-
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v17.4h, v18.4h, v19.4h, v20.4h}, [%1] \n" // r01 r02 r03 r04
-
-                    "shll   v14.4s, %12.4h, #16         \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "mov    v31.16b, %25.16b            \n" // sum01
-
-                    "shll2  v15.4s, %12.8h, #16         \n"
-                    "fmul   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v14.4s, %13.4h, #16         \n"
-                    "fmul   v29.4s, v15.4s, v17.4s      \n"
-                    "shll2  v15.4s, %13.8h, #16         \n"
-                    "fmul   v30.4s, v14.4s, v18.4s      \n"
-                    "shll   v14.4s, %14.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "shll2  v15.4s, %14.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #64]        \n"
-                    "ld1    {v16.4h}, [%2], #8          \n" // r10
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v17.4h, v18.4h, v19.4h, v20.4h}, [%2] \n" // r11 r12 r13 r14
-
-                    "shll   v14.4s, %15.4h, #16         \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "fmla   v29.4s, v15.4s, v16.4s      \n"
-                    "shll2  v15.4s, %15.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v17.4s      \n"
-                    "shll   v14.4s, %16.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v18.4s      \n"
-                    "shll2  v15.4s, %16.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "shll   v14.4s, %17.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v20.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #64]        \n"
-                    "ld1    {v16.4h}, [%3], #8          \n" // r20
-
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v17.4h, v18.4h, v19.4h, v20.4h}, [%3] \n" // r21 r22 r23 r24
-
-                    "shll2  v15.4s, %17.8h, #16         \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "fmla   v30.4s, v14.4s, v16.4s      \n"
-                    "shll   v14.4s, %18.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v17.4s      \n"
-                    "shll2  v15.4s, %18.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "shll   v14.4s, %19.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v19.4s      \n"
-                    "shll2  v15.4s, %19.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v20.4s      \n"
-
-                    "prfm   pldl1keep, [%4, #64]        \n"
-                    "ld1    {v16.4h}, [%4], #8          \n" // r30
-
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v17.4h, v18.4h, v19.4h, v20.4h}, [%4] \n" // r31 r32 r33 r34
-
-                    "shll   v14.4s, %20.4h, #16         \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "fmla   v31.4s, v15.4s, v16.4s      \n"
-                    "shll2  v15.4s, %20.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "shll   v14.4s, %21.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v18.4s      \n"
-                    "shll2  v15.4s, %21.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v19.4s      \n"
-                    "shll   v14.4s, %22.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-
-                    "prfm   pldl1keep, [%5, #64]        \n"
-                    "ld1    {v16.4h}, [%5], #8          \n" // r40
-
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v17.4h, v18.4h, v19.4h, v20.4h}, [%5] \n" // r41 r42 r43 r44
-
-                    "shll2  v15.4s, %22.8h, #16         \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v14.4s, %23.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v17.4s      \n"
-                    "shll2  v15.4s, %23.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v18.4s      \n"
-                    "shll   v14.4s, %24.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-
-                    "fadd   v29.4s, v29.4s, v30.4s      \n"
-                    "fadd   v31.4s, v31.4s, v28.4s      \n"
-                    "fadd   v31.4s, v31.4s, v29.4s      \n"
-
-                    "shrn   v31.4h, v31.4s, #16         \n"
-
-                    "st1    {v31.4h}, [%0], #8          \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2),      // %3
-                    "=r"(r3),      // %4
-                    "=r"(r4)       // %5
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "4"(r3),
-                    "5"(r4),
-                    "w"(_k00_01), // %12
-                    "w"(_k02_03), // %13
-                    "w"(_k04_10), // %14
-                    "w"(_k11_12), // %15
-                    "w"(_k13_14), // %16
-                    "w"(_k20_21), // %17
-                    "w"(_k22_23), // %18
-                    "w"(_k24_30), // %19
-                    "w"(_k31_32), // %20
-                    "w"(_k33_34), // %21
-                    "w"(_k40_41), // %22
-                    "w"(_k42_43), // %23
-                    "w"(_k44),    // %24
-                    "w"(_bias0)   // %25
-                    : "memory", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                    "pld        [%1, #128]          \n"
-                    "vld1.f32   {d24-d25}, [%1]     \n" // sum0
-
-                    "pld        [%2, #64]           \n"
-                    "vld1.u16   {d1}, [%2 :64]!     \n" // r00
-
-                    "vshll.u16  q8, d20, #16        \n" // k00
-
-                    "pld        [%2, #256]          \n"
-                    "vld1.u16   {d6-d9}, [%2 :64]   \n" // r01 r02 r03 r04
-
-                    "vshll.u16  q0, d1, #16         \n"
-
-                    "vshll.u16  q9, d21, #16        \n" // k01
-                    "vshll.u16  q1, d6, #16         \n"
-                    "vmul.f32   q13, q8, q0         \n"
-                    "pld        [%3, #64]           \n"
-                    "vld1.u16   {d1}, [%3 :64]!     \n" // r10
-
-                    "vshll.u16  q2, d7, #16         \n"
-                    "vshll.u16  q10, d22, #16       \n" // k02
-                    "vmul.f32   q14, q9, q1         \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-                    "vshll.u16  q3, d8, #16         \n"
-                    "vshll.u16  q11, d23, #16       \n" // k03
-                    "vmul.f32   q15, q10, q2        \n"
-                    "vshll.u16  q4, d9, #16         \n"
-                    "vshll.u16  q10, d16, #16       \n" // k04
-                    "vmla.f32   q12, q11, q3        \n"
-                    "vshll.u16  q0, d1, #16         \n"
-                    "vshll.u16  q11, d17, #16       \n" // k10
-                    "vmla.f32   q13, q10, q4        \n"
-
-                    "pld        [%3, #256]          \n"
-                    "vld1.u16   {d6-d9}, [%3 :64]   \n" // r11 r12 r13 r14
-
-                    "vshll.u16  q8, d18, #16        \n" // k11
-                    "vshll.u16  q1, d6, #16         \n"
-                    "vmla.f32   q14, q11, q0        \n"
-                    "pld        [%4, #64]           \n"
-                    "vld1.u16   {d1}, [%4 :64]!     \n" // r20
-
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-                    "vshll.u16  q2, d7, #16         \n"
-                    "vshll.u16  q9, d19, #16        \n" // k12
-                    "vmla.f32   q15, q8, q1         \n"
-                    "vshll.u16  q3, d8, #16         \n"
-                    "vshll.u16  q8, d20, #16        \n" // k13
-                    "vmla.f32   q12, q9, q2         \n"
-                    "vshll.u16  q4, d9, #16         \n"
-                    "vshll.u16  q9, d21, #16        \n" // k14
-                    "vmla.f32   q13, q8, q3         \n"
-                    "vshll.u16  q0, d1, #16         \n"
-                    "vshll.u16  q10, d22, #16       \n" // k20
-                    "vmla.f32   q14, q9, q4         \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                    "pld        [%4, #256]          \n"
-                    "vld1.u16   {d6-d9}, [%4 :64]   \n" // r21 r22 r23 r24
-
-                    "vshll.u16  q11, d23, #16       \n" // k21
-                    "vshll.u16  q1, d6, #16         \n"
-                    "vmla.f32   q15, q10, q0        \n"
-                    "pld        [%5, #64]           \n"
-                    "vld1.u16   {d1}, [%5 :64]!     \n" // r30
-
-                    "vshll.u16  q2, d7, #16         \n"
-                    "vshll.u16  q10, d16, #16       \n" // k22
-                    "vmla.f32   q12, q11, q1        \n"
-                    "vshll.u16  q3, d8, #16         \n"
-                    "vshll.u16  q11, d17, #16       \n" // k23
-                    "vmla.f32   q13, q10, q2        \n"
-                    "vshll.u16  q4, d9, #16         \n"
-                    "vshll.u16  q8, d18, #16        \n" // k24
-                    "vmla.f32   q14, q11, q3        \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-                    "vshll.u16  q0, d1, #16         \n"
-                    "vshll.u16  q9, d19, #16        \n" // k30
-                    "vmla.f32   q15, q8, q4         \n"
-
-                    "pld        [%5, #256]          \n"
-                    "vld1.u16   {d6-d9}, [%5 :64]   \n" // r31 r32 r33 r34
-
-                    "vshll.u16  q8, d20, #16        \n" // k31
-                    "vshll.u16  q1, d6, #16         \n"
-                    "vmla.f32   q12, q9, q0         \n"
-                    "pld        [%6, #64]           \n"
-                    "vld1.u16   {d1}, [%6 :64]!     \n" // r40
-
-                    "vshll.u16  q2, d7, #16         \n"
-                    "vshll.u16  q9, d21, #16        \n" // k32
-                    "vmla.f32   q13, q8, q1         \n"
-                    "vshll.u16  q3, d8, #16         \n"
-                    "vshll.u16  q10, d22, #16       \n" // k33
-                    "vmla.f32   q14, q9, q2         \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-                    "vshll.u16  q4, d9, #16         \n"
-                    "vshll.u16  q11, d23, #16       \n" // k34
-                    "vmla.f32   q15, q10, q3        \n"
-                    "vshll.u16  q0, d1, #16         \n"
-                    "vshll.u16  q10, d16, #16       \n" // k40
-                    "vmla.f32   q12, q11, q4        \n"
-
-                    "pld        [%6, #256]          \n"
-                    "vld1.u16   {d6-d9}, [%6 :64]   \n" // r41 r42 r43 r44
-
-                    "vshll.u16  q11, d17, #16       \n" // k41
-                    "vshll.u16  q1, d6, #16         \n"
-                    "vmla.f32   q13, q10, q0        \n"
-                    "vshll.u16  q2, d7, #16         \n"
-                    "vshll.u16  q8, d18, #16        \n" // k42
-                    "vmla.f32   q14, q11, q1        \n"
-                    "pld        [%7, #64]           \n"
-                    "vld1.u16   {d20}, [%7 :64]     \n"
-                    "vshll.u16  q3, d8, #16         \n"
-                    "vshll.u16  q9, d19, #16        \n" // k43
-                    "vmla.f32   q15, q8, q2         \n"
-                    "vshll.u16  q4, d9, #16         \n"
-                    "vshll.u16  q8, d20, #16        \n" // k44
-                    "vmla.f32   q12, q9, q3         \n"
-                    "vmla.f32   q13, q8, q4         \n"
-
-                    "vadd.f32   q14, q14, q15       \n"
-                    "vadd.f32   q12, q12, q13       \n"
-                    "vadd.f32   q12, q12, q14       \n"
-
-                    "sub        %7, %7, #192        \n" // kptr -= 24 * 4;
-
-                    "vshrn.u32  d24, q12, #16       \n"
-
-                    "vst1.u16   {d24}, [%0 :64]!    \n"
-
-                    : "=r"(outptr0),        // %0
-                    "=r"(bias0_data_ptr), // %1
-                    "=r"(r0),             // %2
-                    "=r"(r1),             // %3
-                    "=r"(r2),             // %4
-                    "=r"(r3),             // %5
-                    "=r"(r4),             // %6
-                    "=r"(kptr)            // %7
-                    : "0"(outptr0),
-                    "1"(bias0_data_ptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "7"(kptr)
-                    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-            }
-
-            r0 += 4 * 4;
-            r1 += 4 * 4;
-            r2 += 4 * 4;
-            r3 += 4 * 4;
-            r4 += 4 * 4;
-        }
-    }
-}
-
-static void convdw5x5s2_pack4_bf16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-#if __aarch64__
-        float32x4_t _bias0 = bias ? vld1q_f32((const float*)bias + g * 4) : vdupq_n_f32(0.f);
-#endif // __aarch64__
-
-        const unsigned short* kptr = kernel.row<const unsigned short>(g);
-
-        unsigned short* outptr0 = out;
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const unsigned short* r0 = img0.row<const unsigned short>(0);
-        const unsigned short* r1 = img0.row<const unsigned short>(1);
-        const unsigned short* r2 = img0.row<const unsigned short>(2);
-        const unsigned short* r3 = img0.row<const unsigned short>(3);
-        const unsigned short* r4 = img0.row<const unsigned short>(4);
-
-#if __aarch64__
-        // 4 * 25
-        uint16x8_t _k00_01 = vld1q_u16(kptr);
-        uint16x8_t _k02_03 = vld1q_u16(kptr + 8);
-        uint16x8_t _k04_10 = vld1q_u16(kptr + 16);
-        uint16x8_t _k11_12 = vld1q_u16(kptr + 24);
-        uint16x8_t _k13_14 = vld1q_u16(kptr + 32);
-        uint16x8_t _k20_21 = vld1q_u16(kptr + 40);
-        uint16x8_t _k22_23 = vld1q_u16(kptr + 48);
-        uint16x8_t _k24_30 = vld1q_u16(kptr + 56);
-        uint16x8_t _k31_32 = vld1q_u16(kptr + 64);
-        uint16x8_t _k33_34 = vld1q_u16(kptr + 72);
-        uint16x8_t _k40_41 = vld1q_u16(kptr + 80);
-        uint16x8_t _k42_43 = vld1q_u16(kptr + 88);
-        uint16x4_t _k44 = vld1_u16(kptr + 96);
-#else  // __aarch64__
-        float bias0_data[4];
-        if (bias)
-        {
-            bias0_data[0] = bias[g * 4 + 0];
-            bias0_data[1] = bias[g * 4 + 1];
-            bias0_data[2] = bias[g * 4 + 2];
-            bias0_data[3] = bias[g * 4 + 3];
-        }
-        else
-        {
-            bias0_data[0] = 0.f;
-            bias0_data[1] = 0.f;
-            bias0_data[2] = 0.f;
-            bias0_data[3] = 0.f;
-        }
-        const float* bias0_data_ptr = bias0_data;
-#endif // __aarch64__
-
-        int i = 0;
-
-        for (; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j + 3 < outw; j += 4)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%1], #32 \n" // r00 r01 r02 r03
-
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%1], #32 \n" // r04 r05 r06 r07
-
-                    "shll   v14.4s, %12.4h, #16         \n"
-
-                    "mov    v28.16b, %25.16b            \n" // sum00
-                    "mov    v29.16b, %25.16b            \n" // sum01
-                    "mov    v30.16b, %25.16b            \n" // sum02
-                    "mov    v31.16b, %25.16b            \n" // sum03
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-
-                    "prfm   pldl1keep, [%1, #192]       \n"
-                    "ld1    {v24.4h, v25.4h, v26.4h}, [%1] \n" // r08 r09 r010
-
-                    "shll2  v15.4s, %12.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v18.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v20.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v31.4s, v14.4s, v22.4s      \n"
-                    "shll   v14.4s, %13.4h, #16         \n"
-                    "fmla   v28.4s, v15.4s, v17.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v19.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v21.4s      \n"
-                    "fmla   v31.4s, v15.4s, v23.4s      \n"
-                    "shll2  v15.4s, %13.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll   v24.4s, v24.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v22.4s      \n"
-                    "fmla   v31.4s, v14.4s, v24.4s      \n"
-                    "shll   v14.4s, %14.4h, #16         \n"
-                    "fmla   v28.4s, v15.4s, v19.4s      \n"
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%2], #32 \n" // r10 r11 r12 r13
-
-                    "fmla   v29.4s, v15.4s, v21.4s      \n"
-                    "shll   v25.4s, v25.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v23.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v25.4s      \n"
-                    "shll2  v15.4s, %14.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v22.4s      \n"
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%2], #32 \n" // r14 r15 r16 r17
-
-                    "shll   v26.4s, v26.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v24.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v31.4s, v14.4s, v26.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #192]       \n"
-                    "ld1    {v24.4h, v25.4h, v26.4h}, [%2] \n" // r18 r19 r110
-
-                    "shll   v14.4s, %15.4h, #16         \n"
-                    "fmla   v28.4s, v15.4s, v16.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v18.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v22.4s      \n"
-                    "shll2  v15.4s, %15.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v19.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v21.4s      \n"
-                    "fmla   v31.4s, v14.4s, v23.4s      \n"
-                    "shll   v14.4s, %16.4h, #16         \n"
-                    "fmla   v28.4s, v15.4s, v18.4s      \n"
-                    "fmla   v29.4s, v15.4s, v20.4s      \n"
-                    "shll   v24.4s, v24.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v22.4s      \n"
-                    "fmla   v31.4s, v15.4s, v24.4s      \n"
-                    "shll2  v15.4s, %16.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%3], #32 \n" // r20 r21 r22 r23
-
-                    "fmla   v29.4s, v14.4s, v21.4s      \n"
-                    "shll   v25.4s, v25.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v23.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v31.4s, v14.4s, v25.4s      \n"
-                    "shll   v14.4s, %17.4h, #16         \n"
-                    "fmla   v28.4s, v15.4s, v20.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v22.4s      \n"
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%3], #32 \n" // r24 r25 r26 r27
-
-                    "shll   v26.4s, v26.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v24.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v26.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #192]       \n"
-                    "ld1    {v24.4h, v25.4h, v26.4h}, [%3] \n" // r28 r29 r210
-
-                    "shll2  v15.4s, %17.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v18.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v20.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v31.4s, v14.4s, v22.4s      \n"
-                    "shll   v14.4s, %18.4h, #16         \n"
-                    "fmla   v28.4s, v15.4s, v17.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v19.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v21.4s      \n"
-                    "fmla   v31.4s, v15.4s, v23.4s      \n"
-                    "shll2  v15.4s, %18.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll   v24.4s, v24.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v22.4s      \n"
-                    "fmla   v31.4s, v14.4s, v24.4s      \n"
-                    "shll   v14.4s, %19.4h, #16         \n"
-                    "fmla   v28.4s, v15.4s, v19.4s      \n"
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%4], #32 \n" // r30 r31 r32 r33
-
-                    "fmla   v29.4s, v15.4s, v21.4s      \n"
-                    "shll   v25.4s, v25.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v23.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v25.4s      \n"
-                    "shll2  v15.4s, %19.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v22.4s      \n"
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%4], #32 \n" // r34 r35 r36 r37
-
-                    "shll   v26.4s, v26.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v24.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v31.4s, v14.4s, v26.4s      \n"
-
-                    "prfm   pldl1keep, [%4, #192]       \n"
-                    "ld1    {v24.4h, v25.4h, v26.4h}, [%4] \n" // r38 r39 r310
-
-                    "shll   v14.4s, %20.4h, #16         \n"
-                    "fmla   v28.4s, v15.4s, v16.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v18.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v22.4s      \n"
-                    "shll2  v15.4s, %20.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v19.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v21.4s      \n"
-                    "fmla   v31.4s, v14.4s, v23.4s      \n"
-                    "shll   v14.4s, %21.4h, #16         \n"
-                    "fmla   v28.4s, v15.4s, v18.4s      \n"
-                    "fmla   v29.4s, v15.4s, v20.4s      \n"
-                    "shll   v24.4s, v24.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v22.4s      \n"
-                    "fmla   v31.4s, v15.4s, v24.4s      \n"
-                    "shll2  v15.4s, %21.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%5], #32 \n" // r40 r41 r42 r43
-
-                    "fmla   v29.4s, v14.4s, v21.4s      \n"
-                    "shll   v25.4s, v25.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v23.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v31.4s, v14.4s, v25.4s      \n"
-                    "shll   v14.4s, %22.4h, #16         \n"
-                    "fmla   v28.4s, v15.4s, v20.4s      \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v22.4s      \n"
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%5], #32 \n" // r44 r45 r46 r47
-
-                    "shll   v26.4s, v26.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v24.4s      \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v26.4s      \n"
-
-                    "prfm   pldl1keep, [%5, #192]       \n"
-                    "ld1    {v24.4h, v25.4h, v26.4h}, [%5] \n" // r48 r49 r410
-
-                    "shll2  v15.4s, %22.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v18.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v20.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v31.4s, v14.4s, v22.4s      \n"
-                    "shll   v14.4s, %23.4h, #16         \n"
-                    "fmla   v28.4s, v15.4s, v17.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v19.4s      \n"
-                    "shll   v23.4s, v23.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v21.4s      \n"
-                    "fmla   v31.4s, v15.4s, v23.4s      \n"
-                    "shll2  v15.4s, %23.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll   v24.4s, v24.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v22.4s      \n"
-                    "fmla   v31.4s, v14.4s, v24.4s      \n"
-                    "shll   v14.4s, %24.4h, #16         \n"
-                    "fmla   v28.4s, v15.4s, v19.4s      \n"
-                    "fmla   v29.4s, v15.4s, v21.4s      \n"
-                    "shll   v25.4s, v25.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v23.4s      \n"
-                    "fmla   v31.4s, v15.4s, v25.4s      \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "fmla   v29.4s, v14.4s, v22.4s      \n"
-                    "shll   v26.4s, v26.4h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v24.4s      \n"
-                    "fmla   v31.4s, v14.4s, v26.4s      \n"
-
-                    "shrn   v28.4h, v28.4s, #16         \n"
-                    "shrn   v29.4h, v29.4s, #16         \n"
-                    "shrn   v30.4h, v30.4s, #16         \n"
-                    "shrn   v31.4h, v31.4s, #16         \n"
-
-                    "st1    {v28.4h, v29.4h, v30.4h, v31.4h}, [%0], #32 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2),      // %3
-                    "=r"(r3),      // %4
-                    "=r"(r4)       // %5
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "4"(r3),
-                    "5"(r4),
-                    "w"(_k00_01), // %12
-                    "w"(_k02_03), // %13
-                    "w"(_k04_10), // %14
-                    "w"(_k11_12), // %15
-                    "w"(_k13_14), // %16
-                    "w"(_k20_21), // %17
-                    "w"(_k22_23), // %18
-                    "w"(_k24_30), // %19
-                    "w"(_k31_32), // %20
-                    "w"(_k33_34), // %21
-                    "w"(_k40_41), // %22
-                    "w"(_k42_43), // %23
-                    "w"(_k44),    // %24
-                    "w"(_bias0)   // %25
-                    : "memory", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                    "pld        [%1, #128]          \n"
-                    "vld1.f32   {d24-d25}, [%1]     \n"
-                    "vmov       q13, q12            \n" // sum0 sum1
-
-                    "vshll.u16  q8, d20, #16        \n" // k00
-
-                    "pld        [%2, #256]          \n"
-                    "vld1.u16   {d4-d7}, [%2 :64]!  \n" // r00 r01 r02 r03
-
-                    "vmov       q14, q12            \n"
-                    "vmov       q15, q12            \n" // sum2 sum3
-
-                    "vshll.u16  q9, d21, #16        \n" // k01
-
-                    "pld        [%2, #256]          \n"
-                    "vld1.u16   {d12-d15}, [%2 :64]! \n" // r04 r05 r06 r07
-
-                    "vshll.u16  q0, d4, #16         \n"
-                    "vshll.u16  q1, d5, #16         \n"
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vshll.u16  q3, d7, #16         \n"
-
-                    "vshll.u16  q4, d12, #16        \n"
-                    "vshll.u16  q5, d13, #16        \n"
-
-                    "vmla.f32   q12, q8, q0         \n"
-                    "vmla.f32   q13, q8, q2         \n"
-                    "vshll.u16  q6, d14, #16        \n"
-                    "vmla.f32   q14, q8, q4         \n"
-                    "vmla.f32   q15, q8, q6         \n"
-
-                    "vshll.u16  q10, d22, #16       \n" // k02
-
-                    "vmla.f32   q12, q9, q1         \n"
-                    "vmla.f32   q13, q9, q3         \n"
-                    "vshll.u16  q7, d15, #16        \n"
-                    "vmla.f32   q14, q9, q5         \n"
-                    "vmla.f32   q15, q9, q7         \n"
-
-                    "pld        [%2, #128]          \n"
-                    "vld1.u16   {d2-d3}, [%2 :64]!  \n" // r08 r09
-
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                    "vmla.f32   q12, q10, q2        \n"
-                    "vshll.u16  q11, d23, #16       \n" // k03
-                    "vmla.f32   q13, q10, q4        \n"
-                    "vshll.u16  q0, d2, #16         \n"
-                    "vmla.f32   q14, q10, q6        \n"
-                    "vmla.f32   q15, q10, q0        \n"
-
-                    "vshll.u16  q10, d16, #16       \n" // k04
-
-                    "vmla.f32   q12, q11, q3        \n"
-                    "vmla.f32   q13, q11, q5        \n"
-                    "vshll.u16  q1, d3, #16         \n"
-                    "vmla.f32   q14, q11, q7        \n"
-                    "vmla.f32   q15, q11, q1        \n"
-
-                    "pld        [%2, #64]           \n"
-                    "vld1.u16   {d5}, [%2 :64]      \n" // r010
-
-                    "vmla.f32   q12, q10, q4        \n"
-                    "vshll.u16  q11, d17, #16       \n" // k10
-                    "vmla.f32   q13, q10, q6        \n"
-                    "vshll.u16  q2, d5, #16         \n"
-                    "vmla.f32   q14, q10, q0        \n"
-                    "pld        [%3, #256]          \n"
-                    "vld1.u16   {d12-d15}, [%3 :64]! \n" // r10 r11 r12 r13
-
-                    "vmla.f32   q15, q10, q2        \n"
-
-                    "vshll.u16  q8, d18, #16        \n" // k11
-
-                    "pld        [%3, #256]          \n"
-                    "vld1.u16   {d4-d7}, [%3 :64]!  \n" // r14 r15 r16 r17
-
-                    "vshll.u16  q4, d12, #16        \n"
-                    "vshll.u16  q5, d13, #16        \n"
-                    "vshll.u16  q6, d14, #16        \n"
-                    "vshll.u16  q7, d15, #16        \n"
-
-                    "vshll.u16  q0, d4, #16         \n"
-                    "vshll.u16  q1, d5, #16         \n"
-
-                    "vmla.f32   q12, q11, q4        \n"
-                    "vmla.f32   q13, q11, q6        \n"
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vmla.f32   q14, q11, q0        \n"
-                    "vmla.f32   q15, q11, q2        \n"
-
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                    "vmla.f32   q12, q8, q5         \n"
-                    "vshll.u16  q9, d19, #16        \n" // k12
-                    "vmla.f32   q13, q8, q7         \n"
-                    "vshll.u16  q3, d7, #16         \n"
-                    "vmla.f32   q14, q8, q1         \n"
-                    "vmla.f32   q15, q8, q3         \n"
-
-                    "pld        [%3, #128]          \n"
-                    "vld1.u16   {d10-d11}, [%3 :64]! \n" // r18 r19
-
-                    "vmla.f32   q12, q9, q6         \n"
-                    "vshll.u16  q8, d20, #16        \n" // k13
-                    "vmla.f32   q13, q9, q0         \n"
-                    "vshll.u16  q4, d10, #16        \n"
-                    "vmla.f32   q14, q9, q2         \n"
-                    "vmla.f32   q15, q9, q4         \n"
-
-                    "vshll.u16  q9, d21, #16        \n" // k14
-
-                    "vmla.f32   q12, q8, q7         \n"
-                    "vmla.f32   q13, q8, q1         \n"
-                    "vshll.u16  q5, d11, #16        \n"
-                    "vmla.f32   q14, q8, q3         \n"
-                    "vmla.f32   q15, q8, q5         \n"
-
-                    "pld        [%3, #64]           \n"
-                    "vld1.u16   {d13}, [%3 :64]     \n" // r110
-
-                    "vmla.f32   q12, q9, q0         \n"
-                    "vshll.u16  q10, d22, #16       \n" // k20
-                    "vmla.f32   q13, q9, q2         \n"
-                    "vshll.u16  q6, d13, #16        \n"
-                    "vmla.f32   q14, q9, q4         \n"
-                    "pld        [%4, #256]          \n"
-                    "vld1.u16   {d4-d7}, [%4 :64]!  \n" // r20 r21 r22 r23
-
-                    "vmla.f32   q15, q9, q6         \n"
-
-                    "vshll.u16  q11, d23, #16       \n" // k21
-
-                    "pld        [%4, #256]          \n"
-                    "vld1.u16   {d12-d15}, [%4 :64]! \n" // r24 r25 r26 r27
-
-                    "vshll.u16  q0, d4, #16         \n"
-                    "vshll.u16  q1, d5, #16         \n"
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vshll.u16  q3, d7, #16         \n"
-
-                    "vshll.u16  q4, d12, #16        \n"
-                    "vshll.u16  q5, d13, #16        \n"
-
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                    "vmla.f32   q12, q10, q0        \n"
-                    "vmla.f32   q13, q10, q2        \n"
-                    "vshll.u16  q6, d14, #16        \n"
-                    "vmla.f32   q14, q10, q4        \n"
-                    "vmla.f32   q15, q10, q6        \n"
-
-                    "vshll.u16  q10, d16, #16       \n" // k22
-
-                    "vmla.f32   q12, q11, q1        \n"
-                    "vmla.f32   q13, q11, q3        \n"
-                    "vshll.u16  q7, d15, #16        \n"
-                    "vmla.f32   q14, q11, q5        \n"
-                    "vmla.f32   q15, q11, q7        \n"
-
-                    "pld        [%4, #128]          \n"
-                    "vld1.u16   {d2-d3}, [%4 :64]!  \n" // r28 r29
-
-                    "vmla.f32   q12, q10, q2        \n"
-                    "vshll.u16  q11, d17, #16       \n" // k23
-                    "vmla.f32   q13, q10, q4        \n"
-                    "vshll.u16  q0, d2, #16         \n"
-                    "vmla.f32   q14, q10, q6        \n"
-                    "vmla.f32   q15, q10, q0        \n"
-
-                    "vshll.u16  q8, d18, #16        \n" // k24
-
-                    "vmla.f32   q12, q11, q3        \n"
-                    "vmla.f32   q13, q11, q5        \n"
-                    "vshll.u16  q1, d3, #16         \n"
-                    "vmla.f32   q14, q11, q7        \n"
-                    "vmla.f32   q15, q11, q1        \n"
-
-                    "pld        [%4, #64]           \n"
-                    "vld1.u16   {d5}, [%4 :64]      \n" // r210
-
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                    "vmla.f32   q12, q8, q4         \n"
-                    "vshll.u16  q9, d19, #16        \n" // k30
-                    "vmla.f32   q13, q8, q6         \n"
-                    "vshll.u16  q2, d5, #16         \n"
-                    "vmla.f32   q14, q8, q0         \n"
-                    "pld        [%5, #256]          \n"
-                    "vld1.u16   {d12-d15}, [%5 :64]! \n" // r30 r31 r32 r33
-
-                    "vmla.f32   q15, q8, q2         \n"
-
-                    "vshll.u16  q8, d20, #16        \n" // k31
-
-                    "pld        [%5, #256]          \n"
-                    "vld1.u16   {d4-d7}, [%5 :64]!  \n" // r34 r35 r36 r37
-
-                    "vshll.u16  q4, d12, #16        \n"
-                    "vshll.u16  q5, d13, #16        \n"
-                    "vshll.u16  q6, d14, #16        \n"
-                    "vshll.u16  q7, d15, #16        \n"
-
-                    "vshll.u16  q0, d4, #16         \n"
-                    "vshll.u16  q1, d5, #16         \n"
-
-                    "vmla.f32   q12, q9, q4         \n"
-                    "vmla.f32   q13, q9, q6         \n"
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vmla.f32   q14, q9, q0         \n"
-                    "vmla.f32   q15, q9, q2         \n"
-
-                    "vshll.u16  q9, d21, #16        \n" // k32
-
-                    "vmla.f32   q12, q8, q5         \n"
-                    "vmla.f32   q13, q8, q7         \n"
-                    "vshll.u16  q3, d7, #16         \n"
-                    "vmla.f32   q14, q8, q1         \n"
-                    "vmla.f32   q15, q8, q3         \n"
-
-                    "pld        [%5, #128]          \n"
-                    "vld1.u16   {d10-d11}, [%5 :64]! \n" // r38 r39
-
-                    "vmla.f32   q12, q9, q6         \n"
-                    "vshll.u16  q10, d22, #16       \n" // k33
-                    "vmla.f32   q13, q9, q0         \n"
-                    "vshll.u16  q4, d10, #16        \n"
-                    "vmla.f32   q14, q9, q2         \n"
-                    "vmla.f32   q15, q9, q4         \n"
-
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                    "vmla.f32   q12, q10, q7        \n"
-                    "vshll.u16  q11, d23, #16       \n" // k34
-                    "vmla.f32   q13, q10, q1        \n"
-                    "vshll.u16  q5, d11, #16        \n"
-                    "vmla.f32   q14, q10, q3        \n"
-                    "vmla.f32   q15, q10, q5        \n"
-
-                    "pld        [%5, #64]           \n"
-                    "vld1.u16   {d13}, [%5 :64]     \n" // r310
-
-                    "vmla.f32   q12, q11, q0        \n"
-                    "vshll.u16  q10, d16, #16       \n" // k40
-                    "vmla.f32   q13, q11, q2        \n"
-                    "vshll.u16  q6, d13, #16        \n"
-                    "vmla.f32   q14, q11, q4        \n"
-                    "pld        [%6, #256]          \n"
-                    "vld1.u16   {d4-d7}, [%6 :64]!  \n" // r40 r41 r42 r43
-
-                    "vmla.f32   q15, q11, q6        \n"
-
-                    "vshll.u16  q11, d17, #16       \n" // k41
-
-                    "pld        [%6, #256]          \n"
-                    "vld1.u16   {d12-d15}, [%6 :64]! \n" // r44 r45 r46 r47
-
-                    "vshll.u16  q0, d4, #16         \n"
-                    "vshll.u16  q1, d5, #16         \n"
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vshll.u16  q3, d7, #16         \n"
-
-                    "vshll.u16  q4, d12, #16        \n"
-                    "vshll.u16  q5, d13, #16        \n"
-
-                    "vmla.f32   q12, q10, q0        \n"
-                    "vmla.f32   q13, q10, q2        \n"
-                    "vshll.u16  q6, d14, #16        \n"
-                    "vmla.f32   q14, q10, q4        \n"
-                    "vmla.f32   q15, q10, q6        \n"
-
-                    "vshll.u16  q8, d18, #16        \n" // k42
-
-                    "vmla.f32   q12, q11, q1        \n"
-                    "vmla.f32   q13, q11, q3        \n"
-                    "vshll.u16  q7, d15, #16        \n"
-                    "vmla.f32   q14, q11, q5        \n"
-                    "pld        [%7, #64]           \n"
-                    "vld1.u16   {d20}, [%7 :64]     \n"
-
-                    "vmla.f32   q15, q11, q7        \n"
-
-                    "pld        [%6, #128]          \n"
-                    "vld1.u16   {d2-d3}, [%6 :64]!  \n" // r48 r49
-
-                    "vmla.f32   q12, q8, q2         \n"
-                    "vshll.u16  q9, d19, #16        \n" // k43
-                    "vmla.f32   q13, q8, q4         \n"
-                    "vshll.u16  q0, d2, #16         \n"
-                    "vmla.f32   q14, q8, q6         \n"
-                    "vmla.f32   q15, q8, q0         \n"
-
-                    "vshll.u16  q8, d20, #16        \n" // k44
-
-                    "vmla.f32   q12, q9, q3         \n"
-                    "vmla.f32   q13, q9, q5         \n"
-                    "vshll.u16  q1, d3, #16         \n"
-                    "vmla.f32   q14, q9, q7         \n"
-                    "vmla.f32   q15, q9, q1         \n"
-
-                    "pld        [%6, #64]           \n"
-                    "vld1.u16   {d5}, [%6 :64]      \n" // r410
-
-                    "vmla.f32   q12, q8, q4         \n"
-                    "vmla.f32   q13, q8, q6         \n"
-                    "vshll.u16  q2, d5, #16         \n"
-                    "vmla.f32   q14, q8, q0         \n"
-                    "vmla.f32   q15, q8, q2         \n"
-
-                    "sub        %7, %7, #192        \n" // kptr -= 24 * 4;
-
-                    "sub        %2, %2, #16         \n"
-                    "sub        %3, %3, #16         \n"
-                    "sub        %4, %4, #16         \n"
-                    "sub        %5, %5, #16         \n"
-                    "sub        %6, %6, #16         \n"
-
-                    "vshrn.u32  d24, q12, #16       \n"
-                    "vshrn.u32  d25, q13, #16       \n"
-                    "vshrn.u32  d26, q14, #16       \n"
-                    "vshrn.u32  d27, q15, #16       \n"
-
-                    "vst1.u16   {d24-d27}, [%0 :64]! \n"
-
-                    : "=r"(outptr0),        // %0
-                    "=r"(bias0_data_ptr), // %1
-                    "=r"(r0),             // %2
-                    "=r"(r1),             // %3
-                    "=r"(r2),             // %4
-                    "=r"(r3),             // %5
-                    "=r"(r4),             // %6
-                    "=r"(kptr)            // %7
-                    : "0"(outptr0),
-                    "1"(bias0_data_ptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "7"(kptr)
-                    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%1], #32 \n" // r00 r01 r02 r03
-
-                    "prfm   pldl1keep, [%1, #192]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h}, [%1] \n" // r04 r05 r06
-
-                    "shll   v14.4s, %12.4h, #16         \n"
-                    "shll2  v15.4s, %12.8h, #16         \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-
-                    "mov    v30.16b, %25.16b            \n" // sum00
-                    "mov    v31.16b, %25.16b            \n" // sum01
-
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-
-                    "fmul   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmul   v29.4s, v14.4s, v18.4s      \n"
-                    "shll   v14.4s, %13.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v17.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "shll2  v15.4s, %13.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll   v14.4s, %14.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%2], #32 \n" // r10 r11 r12 r13
-
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-                    "shll2  v15.4s, %14.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v22.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #192]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h}, [%2] \n" // r14 r15 r16
-
-                    "shll   v14.4s, %15.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-
-                    "fmla   v30.4s, v15.4s, v16.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v18.4s      \n"
-                    "shll2  v15.4s, %15.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v19.4s      \n"
-                    "shll   v14.4s, %16.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v18.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-                    "shll2  v15.4s, %16.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%3], #32 \n" // r20 r21 r22 r23
-
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v21.4s      \n"
-                    "shll   v14.4s, %17.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v22.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #192]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h}, [%3] \n" // r24 r25 r26
-
-                    "shll2  v15.4s, %17.8h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v18.4s      \n"
-                    "shll   v14.4s, %18.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v17.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "shll2  v15.4s, %18.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll   v14.4s, %19.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%4], #32 \n" // r30 r31 r32 r33
-
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-                    "shll2  v15.4s, %19.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v22.4s      \n"
-
-                    "prfm   pldl1keep, [%4, #192]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h}, [%4] \n" // r34 r35 r36
-
-                    "shll   v14.4s, %20.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-
-                    "fmla   v30.4s, v15.4s, v16.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v18.4s      \n"
-                    "shll2  v15.4s, %20.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v19.4s      \n"
-                    "shll   v14.4s, %21.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v18.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-                    "shll2  v15.4s, %21.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%5], #32 \n" // r40 r41 r42 r43
-
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v21.4s      \n"
-                    "shll   v14.4s, %22.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v20.4s      \n"
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v22.4s      \n"
-
-                    "prfm   pldl1keep, [%5, #192]       \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h}, [%5] \n" // r44 r45 r46
-
-                    "shll2  v15.4s, %22.8h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-                    "shll   v18.4s, v18.4h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v18.4s      \n"
-                    "shll   v14.4s, %23.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v17.4s      \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "shll2  v15.4s, %23.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "shll   v21.4s, v21.4h, #16         \n"
-                    "fmla   v29.4s, v14.4s, v20.4s      \n"
-                    "shll   v14.4s, %24.4h, #16         \n"
-                    "fmla   v30.4s, v15.4s, v19.4s      \n"
-                    "fmla   v31.4s, v15.4s, v21.4s      \n"
-                    "shll   v22.4s, v22.4h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-                    "fmla   v29.4s, v14.4s, v22.4s      \n"
-
-                    "fadd   v30.4s, v30.4s, v28.4s      \n"
-                    "fadd   v31.4s, v31.4s, v29.4s      \n"
-
-                    "shrn   v30.4h, v30.4s, #16         \n"
-                    "shrn   v31.4h, v31.4s, #16         \n"
-
-                    "st1    {v30.4h, v31.4h}, [%0], #16 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2),      // %3
-                    "=r"(r3),      // %4
-                    "=r"(r4)       // %5
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "4"(r3),
-                    "5"(r4),
-                    "w"(_k00_01), // %12
-                    "w"(_k02_03), // %13
-                    "w"(_k04_10), // %14
-                    "w"(_k11_12), // %15
-                    "w"(_k13_14), // %16
-                    "w"(_k20_21), // %17
-                    "w"(_k22_23), // %18
-                    "w"(_k24_30), // %19
-                    "w"(_k31_32), // %20
-                    "w"(_k33_34), // %21
-                    "w"(_k40_41), // %22
-                    "w"(_k42_43), // %23
-                    "w"(_k44),    // %24
-                    "w"(_bias0)   // %25
-                    : "memory", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                    "pld        [%1, #128]          \n"
-                    "vld1.f32   {d24-d25}, [%1]     \n"
-
-                    "pld        [%2, #256]          \n"
-                    "vld1.u16   {d4-d7}, [%2 :64]!  \n" // r00 r01 r02 r03
-
-                    "vshll.u16  q8, d20, #16        \n" // k00
-
-                    "pld        [%2, #256]          \n"
-                    "vld1.u16   {d10-d12}, [%2 :64] \n" // r04 r05 r06
-
-                    "vmov       q13, q12            \n" // sum0 sum1
-
-                    "vshll.u16  q0, d4, #16         \n"
-                    "vshll.u16  q1, d5, #16         \n"
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vshll.u16  q3, d7, #16         \n"
-
-                    "vshll.u16  q9, d21, #16        \n" // k01
-                    "vmul.f32   q14, q8, q0         \n"
-                    "vshll.u16  q4, d10, #16        \n"
-                    "vmul.f32   q15, q8, q2         \n"
-                    "vshll.u16  q10, d22, #16       \n" // k02
-                    "vmla.f32   q12, q9, q1         \n"
-                    "vshll.u16  q5, d11, #16        \n"
-                    "vmla.f32   q13, q9, q3         \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-                    "vmla.f32   q14, q10, q2        \n"
-                    "vshll.u16  q11, d23, #16       \n" // k03
-                    "vmla.f32   q15, q10, q4        \n"
-                    "vshll.u16  q10, d16, #16       \n" // k04
-                    "vmla.f32   q12, q11, q3        \n"
-                    "vshll.u16  q6, d12, #16        \n"
-                    "vmla.f32   q13, q11, q5        \n"
-                    "vshll.u16  q11, d17, #16       \n" // k10
-                    "vmla.f32   q14, q10, q4        \n"
-                    "vmla.f32   q15, q10, q6        \n"
-
-                    "pld        [%3, #256]          \n"
-                    "vld1.u16   {d4-d7}, [%3 :64]!  \n" // r10 r11 r12 r13
-
-                    "vshll.u16  q8, d18, #16        \n" // k11
-
-                    "pld        [%3, #256]          \n"
-                    "vld1.u16   {d10-d12}, [%3 :64] \n" // r14 r15 r16
-
-                    "vshll.u16  q0, d4, #16         \n"
-                    "vshll.u16  q1, d5, #16         \n"
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vshll.u16  q3, d7, #16         \n"
-
-                    "vmla.f32   q12, q11, q0        \n"
-                    "vshll.u16  q4, d10, #16        \n"
-                    "vmla.f32   q13, q11, q2        \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-                    "vmla.f32   q14, q8, q1         \n"
-                    "vshll.u16  q9, d19, #16        \n" // k12
-                    "vmla.f32   q15, q8, q3         \n"
-                    "vshll.u16  q8, d20, #16        \n" // k13
-                    "vmla.f32   q12, q9, q2         \n"
-                    "vshll.u16  q5, d11, #16        \n"
-                    "vmla.f32   q13, q9, q4         \n"
-                    "vshll.u16  q9, d21, #16        \n" // k14
-                    "vmla.f32   q14, q8, q3         \n"
-                    "vshll.u16  q6, d12, #16        \n"
-                    "vmla.f32   q15, q8, q5         \n"
-                    "vshll.u16  q10, d22, #16       \n" // k20
-                    "vmla.f32   q12, q9, q4         \n"
-                    "vmla.f32   q13, q9, q6         \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-
-                    "pld        [%4, #256]          \n"
-                    "vld1.u16   {d4-d7}, [%4 :64]!  \n" // r20 r21 r22 r23
-
-                    "vshll.u16  q11, d23, #16       \n" // k21
-
-                    "pld        [%4, #256]          \n"
-                    "vld1.u16   {d10-d12}, [%4 :64] \n" // r24 r25 r26
-
-                    "vshll.u16  q0, d4, #16         \n"
-                    "vshll.u16  q1, d5, #16         \n"
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vshll.u16  q3, d7, #16         \n"
-
-                    "vmla.f32   q14, q10, q0        \n"
-                    "vmla.f32   q15, q10, q2        \n"
-                    "vshll.u16  q10, d16, #16       \n" // k22
-                    "vmla.f32   q12, q11, q1        \n"
-                    "vshll.u16  q4, d10, #16        \n"
-                    "vmla.f32   q13, q11, q3        \n"
-                    "vshll.u16  q11, d17, #16       \n" // k23
-                    "vmla.f32   q14, q10, q2        \n"
-                    "vshll.u16  q5, d11, #16        \n"
-                    "vmla.f32   q15, q10, q4        \n"
-                    "vshll.u16  q8, d18, #16        \n" // k24
-                    "vmla.f32   q12, q11, q3        \n"
-                    "vshll.u16  q6, d12, #16        \n"
-                    "vmla.f32   q13, q11, q5        \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-                    "vmla.f32   q14, q8, q4         \n"
-                    "vshll.u16  q9, d19, #16        \n" // k30
-                    "vmla.f32   q15, q8, q6         \n"
-
-                    "pld        [%5, #256]          \n"
-                    "vld1.u16   {d4-d7}, [%5 :64]!  \n" // r30 r31 r32 r33
-
-                    "vshll.u16  q8, d20, #16        \n" // k31
-
-                    "pld        [%5, #256]          \n"
-                    "vld1.u16   {d10-d12}, [%5 :64] \n" // r34 r35 r36
-
-                    "vshll.u16  q0, d4, #16         \n"
-                    "vshll.u16  q1, d5, #16         \n"
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vshll.u16  q3, d7, #16         \n"
-
-                    "vmla.f32   q12, q9, q0         \n"
-                    "vshll.u16  q4, d10, #16        \n"
-                    "vmla.f32   q13, q9, q2         \n"
-                    "vshll.u16  q9, d21, #16        \n" // k32
-                    "vmla.f32   q14, q8, q1         \n"
-                    "vshll.u16  q5, d11, #16        \n"
-                    "vmla.f32   q15, q8, q3         \n"
-                    "vshll.u16  q10, d22, #16       \n" // k33
-                    "vmla.f32   q12, q9, q2         \n"
-                    "vshll.u16  q6, d12, #16        \n"
-                    "vmla.f32   q13, q9, q4         \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-                    "vmla.f32   q14, q10, q3        \n"
-                    "vshll.u16  q11, d23, #16       \n" // k34
-                    "vmla.f32   q15, q10, q5        \n"
-                    "vshll.u16  q10, d16, #16       \n" // k40
-                    "vmla.f32   q12, q11, q4        \n"
-                    "vmla.f32   q13, q11, q6        \n"
-
-                    "pld        [%6, #256]          \n"
-                    "vld1.u16   {d4-d7}, [%6 :64]!  \n" // r40 r41 r42 r43
-
-                    "vshll.u16  q11, d17, #16       \n" // k41
-
-                    "pld        [%6, #256]          \n"
-                    "vld1.u16   {d10-d12}, [%6 :64] \n" // r44 r45 r46
-
-                    "vshll.u16  q0, d4, #16         \n"
-                    "vshll.u16  q1, d5, #16         \n"
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vshll.u16  q3, d7, #16         \n"
-
-                    "vmla.f32   q14, q10, q0        \n"
-                    "vshll.u16  q4, d10, #16        \n"
-                    "vmla.f32   q15, q10, q2        \n"
-                    "vshll.u16  q8, d18, #16        \n" // k42
-                    "vmla.f32   q12, q11, q1        \n"
-                    "vshll.u16  q5, d11, #16        \n"
-                    "vmla.f32   q13, q11, q3        \n"
-                    "pld        [%7, #64]           \n"
-                    "vld1.u16   {d20}, [%7 :64]     \n"
-                    "vmla.f32   q14, q8, q2         \n"
-                    "vshll.u16  q9, d19, #16        \n" // k43
-                    "vmla.f32   q15, q8, q4         \n"
-                    "vshll.u16  q8, d20, #16        \n" // k44
-                    "vmla.f32   q12, q9, q3         \n"
-                    "vshll.u16  q6, d12, #16        \n"
-                    "vmla.f32   q13, q9, q5         \n"
-                    "vmla.f32   q14, q8, q4         \n"
-                    "vmla.f32   q15, q8, q6         \n"
-
-                    "vadd.f32   q12, q12, q14       \n"
-                    "vadd.f32   q13, q13, q15       \n"
-
-                    "sub        %7, %7, #192        \n" // kptr -= 24 * 4;
-
-                    "vshrn.u32  d24, q12, #16       \n"
-                    "vshrn.u32  d25, q13, #16       \n"
-
-                    "vst1.u16   {d24-d25}, [%0 :64]! \n"
-
-                    : "=r"(outptr0),        // %0
-                    "=r"(bias0_data_ptr), // %1
-                    "=r"(r0),             // %2
-                    "=r"(r1),             // %3
-                    "=r"(r2),             // %4
-                    "=r"(r3),             // %5
-                    "=r"(r4),             // %6
-                    "=r"(kptr)            // %7
-                    : "0"(outptr0),
-                    "1"(bias0_data_ptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "7"(kptr)
-                    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-            }
-            for (; j < outw; j++)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #128]       \n"
-                    "ld1    {v16.4h, v17.4h}, [%1], #16 \n" // r00 r01
-
-                    "prfm   pldl1keep, [%1, #192]       \n"
-                    "ld1    {v18.4h, v19.4h, v20.4h}, [%1] \n" // r02 r03 r04
-
-                    "shll   v14.4s, %12.4h, #16         \n"
-
-                    "mov    v31.16b, %25.16b            \n" // sum00
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "shll2  v15.4s, %12.8h, #16         \n"
-                    "fmul   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v14.4s, %13.4h, #16         \n"
-                    "fmul   v29.4s, v15.4s, v17.4s      \n"
-                    "shll2  v15.4s, %13.8h, #16         \n"
-                    "fmul   v30.4s, v14.4s, v18.4s      \n"
-                    "shll   v14.4s, %14.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "shll2  v15.4s, %14.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "ld1    {v16.4h, v17.4h}, [%2], #16 \n" // r10 r11
-
-                    "prfm   pldl1keep, [%2, #192]       \n"
-                    "ld1    {v18.4h, v19.4h, v20.4h}, [%2] \n" // r12 r13 r14
-
-                    "shll   v14.4s, %15.4h, #16         \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "fmla   v29.4s, v15.4s, v16.4s      \n"
-                    "shll2  v15.4s, %15.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v17.4s      \n"
-                    "shll   v14.4s, %16.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v18.4s      \n"
-                    "shll2  v15.4s, %16.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v19.4s      \n"
-                    "shll   v14.4s, %17.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v20.4s      \n"
-
-                    "prfm   pldl1keep, [%3, #128]       \n"
-                    "ld1    {v16.4h, v17.4h}, [%3], #16 \n" // r20 r21
-
-                    "prfm   pldl1keep, [%3, #192]       \n"
-                    "ld1    {v18.4h, v19.4h, v20.4h}, [%3] \n" // r22 r23 r24
-
-                    "shll2  v15.4s, %17.8h, #16         \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "fmla   v30.4s, v14.4s, v16.4s      \n"
-                    "shll   v14.4s, %18.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v17.4s      \n"
-                    "shll2  v15.4s, %18.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v18.4s      \n"
-                    "shll   v14.4s, %19.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v19.4s      \n"
-                    "shll2  v15.4s, %19.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v20.4s      \n"
-
-                    "prfm   pldl1keep, [%4, #128]       \n"
-                    "ld1    {v16.4h, v17.4h}, [%4], #16 \n" // r30 r31
-
-                    "prfm   pldl1keep, [%4, #192]       \n"
-                    "ld1    {v18.4h, v19.4h, v20.4h}, [%4] \n" // r32 r33 r34
-
-                    "shll   v14.4s, %20.4h, #16         \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "fmla   v31.4s, v15.4s, v16.4s      \n"
-                    "shll2  v15.4s, %20.8h, #16         \n"
-                    "fmla   v28.4s, v14.4s, v17.4s      \n"
-                    "shll   v14.4s, %21.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v18.4s      \n"
-                    "shll2  v15.4s, %21.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v19.4s      \n"
-                    "shll   v14.4s, %22.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v20.4s      \n"
-
-                    "prfm   pldl1keep, [%5, #128]       \n"
-                    "ld1    {v16.4h, v17.4h}, [%5], #16 \n" // r40 r41
-
-                    "prfm   pldl1keep, [%5, #192]       \n"
-                    "ld1    {v18.4h, v19.4h, v20.4h}, [%5] \n" // r42 r43 r44
-
-                    "shll2  v15.4s, %22.8h, #16         \n"
-
-                    "shll   v16.4s, v16.4h, #16         \n"
-                    "shll   v17.4s, v17.4h, #16         \n"
-
-                    "shll   v18.4s, v18.4h, #16         \n"
-                    "shll   v19.4s, v19.4h, #16         \n"
-                    "shll   v20.4s, v20.4h, #16         \n"
-
-                    "fmla   v28.4s, v14.4s, v16.4s      \n"
-                    "shll   v14.4s, %23.4h, #16         \n"
-                    "fmla   v29.4s, v15.4s, v17.4s      \n"
-                    "shll2  v15.4s, %23.8h, #16         \n"
-                    "fmla   v30.4s, v14.4s, v18.4s      \n"
-                    "shll   v14.4s, %24.4h, #16         \n"
-                    "fmla   v31.4s, v15.4s, v19.4s      \n"
-                    "fmla   v28.4s, v14.4s, v20.4s      \n"
-
-                    "fadd   v29.4s, v29.4s, v30.4s      \n"
-                    "fadd   v31.4s, v31.4s, v28.4s      \n"
-                    "fadd   v31.4s, v31.4s, v29.4s      \n"
-
-                    "shrn   v31.4h, v31.4s, #16         \n"
-
-                    "st1    {v31.4h}, [%0], #8          \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2),      // %3
-                    "=r"(r3),      // %4
-                    "=r"(r4)       // %5
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "4"(r3),
-                    "5"(r4),
-                    "w"(_k00_01), // %12
-                    "w"(_k02_03), // %13
-                    "w"(_k04_10), // %14
-                    "w"(_k11_12), // %15
-                    "w"(_k13_14), // %16
-                    "w"(_k20_21), // %17
-                    "w"(_k22_23), // %18
-                    "w"(_k24_30), // %19
-                    "w"(_k31_32), // %20
-                    "w"(_k33_34), // %21
-                    "w"(_k40_41), // %22
-                    "w"(_k42_43), // %23
-                    "w"(_k44),    // %24
-                    "w"(_bias0)   // %25
-                    : "memory", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%2, #128]          \n"
-                    "vld1.u16   {d2-d3}, [%2 :64]!  \n" // r00 r01
-
-                    "pld        [%2, #192]          \n"
-                    "vld1.u16   {d6-d8}, [%2 :64]   \n" // r02 r03 r04
-
-                    "vshll.u16  q0, d2, #16         \n"
-                    "vshll.u16  q1, d3, #16         \n"
-
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-
-                    "vshll.u16  q8, d20, #16        \n" // k00
-
-                    "pld        [%1, #128]          \n"
-                    "vld1.f32   {d24-d25}, [%1]     \n" // sum0
-
-                    "vshll.u16  q9, d21, #16        \n" // k01
-                    "vmul.f32   q13, q8, q0         \n"
-                    "vshll.u16  q10, d22, #16       \n" // k02
-                    "vmul.f32   q14, q9, q1         \n"
-
-                    "pld        [%3, #128]          \n"
-                    "vld1.u16   {d14-d15}, [%3 :64]! \n" // r10 r11
-
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vshll.u16  q3, d7, #16         \n"
-                    "vshll.u16  q4, d8, #16         \n"
-
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-                    "vshll.u16  q11, d23, #16       \n" // k03
-                    "vmul.f32   q15, q10, q2        \n"
-                    "vshll.u16  q10, d16, #16       \n" // k04
-                    "vmla.f32   q12, q11, q3        \n"
-                    "vshll.u16  q11, d17, #16       \n" // k10
-                    "vmla.f32   q13, q10, q4        \n"
-
-                    "pld        [%3, #192]          \n"
-                    "vld1.u16   {d8-d10}, [%3 :64]  \n" // r12 r13 r14
-
-                    "vshll.u16  q6, d14, #16        \n"
-                    "vshll.u16  q7, d15, #16        \n"
-
-                    "vshll.u16  q8, d18, #16        \n" // k11
-                    "vmla.f32   q14, q11, q6        \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-                    "vshll.u16  q9, d19, #16        \n" // k12
-                    "vmla.f32   q15, q8, q7         \n"
-
-                    "pld        [%4, #128]          \n"
-                    "vld1.u16   {d2-d3}, [%4 :64]!  \n" // r20 r21
-
-                    "vshll.u16  q3, d8, #16         \n"
-                    "vshll.u16  q4, d9, #16         \n"
-                    "vshll.u16  q5, d10, #16        \n"
-
-                    "vshll.u16  q8, d20, #16        \n" // k13
-                    "vmla.f32   q12, q9, q3         \n"
-                    "vshll.u16  q9, d21, #16        \n" // k14
-                    "vmla.f32   q13, q8, q4         \n"
-                    "vshll.u16  q10, d22, #16       \n" // k20
-                    "vmla.f32   q14, q9, q5         \n"
-
-                    "pld        [%4, #192]          \n"
-                    "vld1.u16   {d6-d8}, [%4 :64]   \n" // r22 r23 r24
-
-                    "vshll.u16  q0, d2, #16         \n"
-                    "vshll.u16  q1, d3, #16         \n"
-
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-                    "vshll.u16  q11, d23, #16       \n" // k21
-                    "vmla.f32   q15, q10, q0        \n"
-                    "vshll.u16  q10, d16, #16       \n" // k22
-                    "vmla.f32   q12, q11, q1        \n"
-
-                    "pld        [%5, #128]          \n"
-                    "vld1.u16   {d14-d15}, [%5 :64]! \n" // r30 r31
-
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vshll.u16  q3, d7, #16         \n"
-                    "vshll.u16  q4, d8, #16         \n"
-
-                    "vshll.u16  q11, d17, #16       \n" // k23
-                    "vmla.f32   q13, q10, q2        \n"
-                    "vshll.u16  q8, d18, #16        \n" // k24
-                    "vmla.f32   q14, q11, q3        \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d20-d23}, [%7 :64]! \n"
-                    "vshll.u16  q9, d19, #16        \n" // k30
-                    "vmla.f32   q15, q8, q4         \n"
-
-                    "pld        [%5, #192]          \n"
-                    "vld1.u16   {d8-d10}, [%5 :64]  \n" // r32 r33 r34
-
-                    "vshll.u16  q6, d14, #16        \n"
-                    "vshll.u16  q7, d15, #16        \n"
-
-                    "vshll.u16  q8, d20, #16        \n" // k31
-                    "vmla.f32   q12, q9, q6         \n"
-
-                    "vshll.u16  q9, d21, #16        \n" // k32
-                    "vmla.f32   q13, q8, q7         \n"
-
-                    "pld        [%6, #128]          \n"
-                    "vld1.u16   {d2-d3}, [%6 :64]!  \n" // r40 r41
-
-                    "vshll.u16  q3, d8, #16         \n"
-                    "vshll.u16  q4, d9, #16         \n"
-                    "vshll.u16  q5, d10, #16        \n"
-
-                    "vshll.u16  q10, d22, #16       \n" // k33
-                    "vmla.f32   q14, q9, q3         \n"
-                    "pld        [%7, #256]          \n"
-                    "vld1.u16   {d16-d19}, [%7 :64]! \n"
-                    "vshll.u16  q11, d23, #16       \n" // k34
-                    "vmla.f32   q15, q10, q4        \n"
-                    "vshll.u16  q10, d16, #16       \n" // k40
-                    "vmla.f32   q12, q11, q5        \n"
-
-                    "pld        [%6, #192]          \n"
-                    "vld1.u16   {d6-d8}, [%6 :64]   \n" // r42 r43 r44
-
-                    "vshll.u16  q0, d2, #16         \n"
-                    "vshll.u16  q1, d3, #16         \n"
-
-                    "vshll.u16  q11, d17, #16       \n" // k41
-                    "vmla.f32   q13, q10, q0        \n"
-                    "vshll.u16  q8, d18, #16        \n" // k42
-                    "vmla.f32   q14, q11, q1        \n"
-
-                    "vshll.u16  q2, d6, #16         \n"
-                    "vshll.u16  q3, d7, #16         \n"
-                    "vshll.u16  q4, d8, #16         \n"
-
-                    "pld        [%7, #64]           \n"
-                    "vld1.u16   {d20}, [%7 :64]     \n"
-                    "vshll.u16  q9, d19, #16        \n" // k43
-                    "vmla.f32   q15, q8, q2         \n"
-                    "vshll.u16  q8, d20, #16        \n" // k44
-                    "vmla.f32   q12, q9, q3         \n"
-
-                    "vmla.f32   q13, q8, q4         \n"
-
-                    "vadd.f32   q14, q14, q15       \n"
-                    "vadd.f32   q12, q12, q13       \n"
-
-                    "sub        %7, %7, #192        \n" // kptr -= 24 * 4;
-
-                    "vadd.f32   q12, q12, q14       \n"
-
-                    "vshrn.u32  d24, q12, #16       \n"
-
-                    "vst1.u16   {d24}, [%0 :64]!    \n"
-
-                    : "=r"(outptr0),        // %0
-                    "=r"(bias0_data_ptr), // %1
-                    "=r"(r0),             // %2
-                    "=r"(r1),             // %3
-                    "=r"(r2),             // %4
-                    "=r"(r3),             // %5
-                    "=r"(r4),             // %6
-                    "=r"(kptr)            // %7
-                    : "0"(outptr0),
-                    "1"(bias0_data_ptr),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "7"(kptr)
-                    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-            r3 += tailstep;
-            r4 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5_pack8_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5_pack8_fp16s.h
deleted file mode 100644
index 81e0a199..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_5x5_pack8_fp16s.h
+++ /dev/null
@@ -1,1426 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw5x5s1_pack8_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __fp16 bias0_data[8] = {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
-
-        const __fp16* k0 = kernel.row<const __fp16>(g);
-
-        __fp16* outptr0 = out.row<__fp16>(0);
-        __fp16* outptr1 = out.row<__fp16>(1);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const __fp16* r0 = img0.row<const __fp16>(0);
-        const __fp16* r1 = img0.row<const __fp16>(1);
-        const __fp16* r2 = img0.row<const __fp16>(2);
-        const __fp16* r3 = img0.row<const __fp16>(3);
-        const __fp16* r4 = img0.row<const __fp16>(4);
-        const __fp16* r5 = img0.row<const __fp16>(5);
-
-        int i = 0;
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                const __fp16* bias0_data_ptr = bias ? bias + g * 8 : bias0_data;
-
-                asm volatile(
-                    "prfm   pldl1keep, [%18, #512]      \n"
-                    "ld1    {v31.8h}, [%18]             \n" // sum13
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%2], #64 \n" // r0_0123
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%8], #64 \n" // w0_0123
-
-                    "mov    v24.16b, v31.16b            \n" // sum00
-                    "mov    v25.16b, v31.16b            \n" // sum01
-                    "mov    v26.16b, v31.16b            \n" // sum02
-                    "mov    v27.16b, v31.16b            \n" // sum03
-
-                    "fmla   v24.8h, v16.8h, v0.8h       \n"
-                    "fmla   v25.8h, v17.8h, v0.8h       \n"
-                    "fmla   v26.8h, v18.8h, v0.8h       \n"
-                    "fmla   v27.8h, v19.8h, v0.8h       \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%2] \n" // r0_4567
-
-                    "fmla   v24.8h, v17.8h, v1.8h       \n"
-                    "fmla   v25.8h, v18.8h, v1.8h       \n"
-                    "fmla   v26.8h, v19.8h, v1.8h       \n"
-                    "fmla   v27.8h, v20.8h, v1.8h       \n"
-
-                    "mov    v28.16b, v31.16b            \n" // sum10
-
-                    "fmla   v24.8h, v18.8h, v2.8h       \n"
-                    "fmla   v25.8h, v19.8h, v2.8h       \n"
-                    "fmla   v26.8h, v20.8h, v2.8h       \n"
-                    "fmla   v27.8h, v21.8h, v2.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%8], #64 \n" // w04 w1_012
-
-                    "fmla   v24.8h, v19.8h, v3.8h       \n"
-                    "fmla   v25.8h, v20.8h, v3.8h       \n"
-                    "fmla   v26.8h, v21.8h, v3.8h       \n"
-                    "fmla   v27.8h, v22.8h, v3.8h       \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v8.8h, v9.8h, v10.8h, v11.8h}, [%3], #64 \n" // r1_0123
-
-                    "fmla   v24.8h, v20.8h, v4.8h       \n"
-                    "fmla   v25.8h, v21.8h, v4.8h       \n"
-                    "fmla   v26.8h, v22.8h, v4.8h       \n"
-                    "fmla   v27.8h, v23.8h, v4.8h       \n"
-
-                    "mov    v29.16b, v31.16b            \n" // sum11
-                    "mov    v30.16b, v31.16b            \n" // sum12
-
-                    "fmla   v28.8h, v8.8h, v0.8h        \n"
-                    "fmla   v29.8h, v9.8h, v0.8h        \n"
-                    "fmla   v30.8h, v10.8h, v0.8h       \n"
-                    "fmla   v31.8h, v11.8h, v0.8h       \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%3] \n" // r1_4567
-
-                    "fmla   v28.8h, v9.8h, v1.8h        \n"
-                    "fmla   v29.8h, v10.8h, v1.8h       \n"
-                    "fmla   v30.8h, v11.8h, v1.8h       \n"
-                    "fmla   v31.8h, v12.8h, v1.8h       \n"
-
-                    "fmla   v28.8h, v10.8h, v2.8h       \n"
-                    "fmla   v29.8h, v11.8h, v2.8h       \n"
-                    "fmla   v30.8h, v12.8h, v2.8h       \n"
-                    "fmla   v31.8h, v13.8h, v2.8h       \n"
-
-                    "fmla   v28.8h, v11.8h, v3.8h       \n"
-                    "fmla   v29.8h, v12.8h, v3.8h       \n"
-                    "fmla   v30.8h, v13.8h, v3.8h       \n"
-                    "fmla   v31.8h, v14.8h, v3.8h       \n"
-
-                    "fmla   v28.8h, v12.8h, v4.8h       \n"
-                    "fmla   v29.8h, v13.8h, v4.8h       \n"
-                    "fmla   v30.8h, v14.8h, v4.8h       \n"
-                    "fmla   v31.8h, v15.8h, v4.8h       \n"
-
-                    "fmla   v24.8h, v8.8h, v5.8h        \n"
-                    "fmla   v25.8h, v9.8h, v5.8h        \n"
-                    "fmla   v26.8h, v10.8h, v5.8h       \n"
-                    "fmla   v27.8h, v11.8h, v5.8h       \n"
-
-                    "fmla   v24.8h, v9.8h, v6.8h        \n"
-                    "fmla   v25.8h, v10.8h, v6.8h       \n"
-                    "fmla   v26.8h, v11.8h, v6.8h       \n"
-                    "fmla   v27.8h, v12.8h, v6.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%8], #64 \n" // w1_34 w2_01
-
-                    "fmla   v24.8h, v10.8h, v7.8h       \n"
-                    "fmla   v25.8h, v11.8h, v7.8h       \n"
-                    "fmla   v26.8h, v12.8h, v7.8h       \n"
-                    "fmla   v27.8h, v13.8h, v7.8h       \n"
-
-                    "fmla   v24.8h, v11.8h, v0.8h       \n"
-                    "fmla   v25.8h, v12.8h, v0.8h       \n"
-                    "fmla   v26.8h, v13.8h, v0.8h       \n"
-                    "fmla   v27.8h, v14.8h, v0.8h       \n"
-
-                    "prfm   pldl1keep, [%4, #512]       \n"
-                    "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%4], #64 \n" // r2_0123
-
-                    "fmla   v24.8h, v12.8h, v1.8h       \n"
-                    "fmla   v25.8h, v13.8h, v1.8h       \n"
-                    "fmla   v26.8h, v14.8h, v1.8h       \n"
-                    "fmla   v27.8h, v15.8h, v1.8h       \n"
-
-                    "fmla   v28.8h, v16.8h, v5.8h       \n"
-                    "fmla   v29.8h, v17.8h, v5.8h       \n"
-                    "fmla   v30.8h, v18.8h, v5.8h       \n"
-                    "fmla   v31.8h, v19.8h, v5.8h       \n"
-
-                    "prfm   pldl1keep, [%4, #512]       \n"
-                    "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4] \n" // r2_4567
-
-                    "fmla   v28.8h, v17.8h, v6.8h       \n"
-                    "fmla   v29.8h, v18.8h, v6.8h       \n"
-                    "fmla   v30.8h, v19.8h, v6.8h       \n"
-                    "fmla   v31.8h, v20.8h, v6.8h       \n"
-
-                    "fmla   v28.8h, v18.8h, v7.8h       \n"
-                    "fmla   v29.8h, v19.8h, v7.8h       \n"
-                    "fmla   v30.8h, v20.8h, v7.8h       \n"
-                    "fmla   v31.8h, v21.8h, v7.8h       \n"
-
-                    "fmla   v28.8h, v19.8h, v0.8h       \n"
-                    "fmla   v29.8h, v20.8h, v0.8h       \n"
-                    "fmla   v30.8h, v21.8h, v0.8h       \n"
-                    "fmla   v31.8h, v22.8h, v0.8h       \n"
-
-                    "fmla   v28.8h, v20.8h, v1.8h       \n"
-                    "fmla   v29.8h, v21.8h, v1.8h       \n"
-                    "fmla   v30.8h, v22.8h, v1.8h       \n"
-                    "fmla   v31.8h, v23.8h, v1.8h       \n"
-
-                    "fmla   v24.8h, v16.8h, v2.8h       \n"
-                    "fmla   v25.8h, v17.8h, v2.8h       \n"
-                    "fmla   v26.8h, v18.8h, v2.8h       \n"
-                    "fmla   v27.8h, v19.8h, v2.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%8], #64 \n" // w2_234 w30
-
-                    "fmla   v24.8h, v17.8h, v3.8h       \n"
-                    "fmla   v25.8h, v18.8h, v3.8h       \n"
-                    "fmla   v26.8h, v19.8h, v3.8h       \n"
-                    "fmla   v27.8h, v20.8h, v3.8h       \n"
-
-                    "fmla   v24.8h, v18.8h, v4.8h       \n"
-                    "fmla   v25.8h, v19.8h, v4.8h       \n"
-                    "fmla   v26.8h, v20.8h, v4.8h       \n"
-                    "fmla   v27.8h, v21.8h, v4.8h       \n"
-
-                    "fmla   v24.8h, v19.8h, v5.8h       \n"
-                    "fmla   v25.8h, v20.8h, v5.8h       \n"
-                    "fmla   v26.8h, v21.8h, v5.8h       \n"
-                    "fmla   v27.8h, v22.8h, v5.8h       \n"
-
-                    "prfm   pldl1keep, [%5, #512]       \n"
-                    "ld1    {v8.8h, v9.8h, v10.8h, v11.8h}, [%5], #64 \n" // r3_0123
-
-                    "fmla   v24.8h, v20.8h, v6.8h       \n"
-                    "fmla   v25.8h, v21.8h, v6.8h       \n"
-                    "fmla   v26.8h, v22.8h, v6.8h       \n"
-                    "fmla   v27.8h, v23.8h, v6.8h       \n"
-
-                    "fmla   v28.8h, v8.8h, v2.8h        \n"
-                    "fmla   v29.8h, v9.8h, v2.8h        \n"
-                    "fmla   v30.8h, v10.8h, v2.8h       \n"
-                    "fmla   v31.8h, v11.8h, v2.8h       \n"
-
-                    "prfm   pldl1keep, [%5, #512]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%5] \n" // r3_4567
-
-                    "fmla   v28.8h, v9.8h, v3.8h        \n"
-                    "fmla   v29.8h, v10.8h, v3.8h       \n"
-                    "fmla   v30.8h, v11.8h, v3.8h       \n"
-                    "fmla   v31.8h, v12.8h, v3.8h       \n"
-
-                    "fmla   v28.8h, v10.8h, v4.8h       \n"
-                    "fmla   v29.8h, v11.8h, v4.8h       \n"
-                    "fmla   v30.8h, v12.8h, v4.8h       \n"
-                    "fmla   v31.8h, v13.8h, v4.8h       \n"
-
-                    "fmla   v28.8h, v11.8h, v5.8h       \n"
-                    "fmla   v29.8h, v12.8h, v5.8h       \n"
-                    "fmla   v30.8h, v13.8h, v5.8h       \n"
-                    "fmla   v31.8h, v14.8h, v5.8h       \n"
-
-                    "fmla   v28.8h, v12.8h, v6.8h       \n"
-                    "fmla   v29.8h, v13.8h, v6.8h       \n"
-                    "fmla   v30.8h, v14.8h, v6.8h       \n"
-                    "fmla   v31.8h, v15.8h, v6.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%8], #64 \n" // w3_1234
-
-                    "fmla   v24.8h, v8.8h, v7.8h        \n"
-                    "fmla   v25.8h, v9.8h, v7.8h        \n"
-                    "fmla   v26.8h, v10.8h, v7.8h       \n"
-                    "fmla   v27.8h, v11.8h, v7.8h       \n"
-
-                    "fmla   v24.8h, v9.8h, v0.8h        \n"
-                    "fmla   v25.8h, v10.8h, v0.8h       \n"
-                    "fmla   v26.8h, v11.8h, v0.8h       \n"
-                    "fmla   v27.8h, v12.8h, v0.8h       \n"
-
-                    "fmla   v24.8h, v10.8h, v1.8h       \n"
-                    "fmla   v25.8h, v11.8h, v1.8h       \n"
-                    "fmla   v26.8h, v12.8h, v1.8h       \n"
-                    "fmla   v27.8h, v13.8h, v1.8h       \n"
-
-                    "fmla   v24.8h, v11.8h, v2.8h       \n"
-                    "fmla   v25.8h, v12.8h, v2.8h       \n"
-                    "fmla   v26.8h, v13.8h, v2.8h       \n"
-                    "fmla   v27.8h, v14.8h, v2.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%6], #64 \n" // r4_0123
-
-                    "fmla   v24.8h, v12.8h, v3.8h       \n"
-                    "fmla   v25.8h, v13.8h, v3.8h       \n"
-                    "fmla   v26.8h, v14.8h, v3.8h       \n"
-                    "fmla   v27.8h, v15.8h, v3.8h       \n"
-
-                    "fmla   v28.8h, v16.8h, v7.8h       \n"
-                    "fmla   v29.8h, v17.8h, v7.8h       \n"
-                    "fmla   v30.8h, v18.8h, v7.8h       \n"
-                    "fmla   v31.8h, v19.8h, v7.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%6] \n" // r4_4567
-
-                    "fmla   v28.8h, v17.8h, v0.8h       \n"
-                    "fmla   v29.8h, v18.8h, v0.8h       \n"
-                    "fmla   v30.8h, v19.8h, v0.8h       \n"
-                    "fmla   v31.8h, v20.8h, v0.8h       \n"
-
-                    "fmla   v28.8h, v18.8h, v1.8h       \n"
-                    "fmla   v29.8h, v19.8h, v1.8h       \n"
-                    "fmla   v30.8h, v20.8h, v1.8h       \n"
-                    "fmla   v31.8h, v21.8h, v1.8h       \n"
-
-                    "fmla   v28.8h, v19.8h, v2.8h       \n"
-                    "fmla   v29.8h, v20.8h, v2.8h       \n"
-                    "fmla   v30.8h, v21.8h, v2.8h       \n"
-                    "fmla   v31.8h, v22.8h, v2.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%8], #64 \n" // w4_0123
-
-                    "fmla   v28.8h, v20.8h, v3.8h       \n"
-                    "fmla   v29.8h, v21.8h, v3.8h       \n"
-                    "fmla   v30.8h, v22.8h, v3.8h       \n"
-                    "fmla   v31.8h, v23.8h, v3.8h       \n"
-
-                    "fmla   v24.8h, v16.8h, v4.8h       \n"
-                    "fmla   v25.8h, v17.8h, v4.8h       \n"
-                    "fmla   v26.8h, v18.8h, v4.8h       \n"
-                    "fmla   v27.8h, v19.8h, v4.8h       \n"
-
-                    "fmla   v24.8h, v17.8h, v5.8h       \n"
-                    "fmla   v25.8h, v18.8h, v5.8h       \n"
-                    "fmla   v26.8h, v19.8h, v5.8h       \n"
-                    "fmla   v27.8h, v20.8h, v5.8h       \n"
-
-                    "fmla   v24.8h, v18.8h, v6.8h       \n"
-                    "fmla   v25.8h, v19.8h, v6.8h       \n"
-                    "fmla   v26.8h, v20.8h, v6.8h       \n"
-                    "fmla   v27.8h, v21.8h, v6.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #128]       \n"
-                    "ld1    {v0.8h}, [%8]               \n" // w44
-
-                    "fmla   v24.8h, v19.8h, v7.8h       \n"
-                    "fmla   v25.8h, v20.8h, v7.8h       \n"
-                    "fmla   v26.8h, v21.8h, v7.8h       \n"
-                    "fmla   v27.8h, v22.8h, v7.8h       \n"
-
-                    "prfm   pldl1keep, [%7, #512]       \n"
-                    "ld1    {v8.8h, v9.8h, v10.8h, v11.8h}, [%7], #64 \n" // r5_0123
-
-                    "fmla   v24.8h, v20.8h, v0.8h       \n"
-                    "fmla   v25.8h, v21.8h, v0.8h       \n"
-                    "fmla   v26.8h, v22.8h, v0.8h       \n"
-                    "fmla   v27.8h, v23.8h, v0.8h       \n"
-
-                    "fmla   v28.8h, v8.8h, v4.8h        \n"
-                    "fmla   v29.8h, v9.8h, v4.8h        \n"
-                    "fmla   v30.8h, v10.8h, v4.8h       \n"
-                    "fmla   v31.8h, v11.8h, v4.8h       \n"
-
-                    "prfm   pldl1keep, [%7, #512]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%7] \n" // r5_4567
-
-                    "fmla   v28.8h, v9.8h, v5.8h        \n"
-                    "fmla   v29.8h, v10.8h, v5.8h       \n"
-                    "fmla   v30.8h, v11.8h, v5.8h       \n"
-                    "fmla   v31.8h, v12.8h, v5.8h       \n"
-
-                    "fmla   v28.8h, v10.8h, v6.8h       \n"
-                    "fmla   v29.8h, v11.8h, v6.8h       \n"
-                    "fmla   v30.8h, v12.8h, v6.8h       \n"
-                    "fmla   v31.8h, v13.8h, v6.8h       \n"
-
-                    "fmla   v28.8h, v11.8h, v7.8h       \n"
-                    "fmla   v29.8h, v12.8h, v7.8h       \n"
-                    "fmla   v30.8h, v13.8h, v7.8h       \n"
-                    "fmla   v31.8h, v14.8h, v7.8h       \n"
-
-                    "fmla   v28.8h, v12.8h, v0.8h       \n"
-                    "fmla   v29.8h, v13.8h, v0.8h       \n"
-                    "fmla   v30.8h, v14.8h, v0.8h       \n"
-                    "fmla   v31.8h, v15.8h, v0.8h       \n"
-
-                    "sub    %8, %8, #384                \n" // k0 -= 24 * 8
-
-                    "st1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%0], #64 \n"
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%1], #64 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3),      // %5
-                    "=r"(r4),      // %6
-                    "=r"(r5),      // %7
-                    "=r"(k0)       // %8
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "7"(r5),
-                    "8"(k0),
-                    "r"(bias0_data_ptr) // %18
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-
-            float16x8_t _bias0 = bias ? vld1q_f16(bias + g * 8) : vdupq_n_f16((__fp16)0.f);
-
-            for (; j + 1 < outw; j += 2)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v16.8h, v17.8h}, [%2], #32 \n" // r0_01
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%8], #64 \n" // w0_0123
-
-                    "mov    v28.16b, %18.16b            \n" // sum00
-                    "mov    v29.16b, %18.16b            \n" // sum01
-
-                    "fmla   v28.8h, v16.8h, v0.8h       \n"
-                    "fmla   v29.8h, v17.8h, v0.8h       \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v18.8h, v19.8h, v20.8h, v21.8h}, [%2] \n" // r0_2345
-
-                    "mov    v30.16b, %18.16b            \n" // sum10
-                    "mov    v31.16b, %18.16b            \n" // sum11
-
-                    "fmla   v28.8h, v17.8h, v1.8h       \n"
-                    "fmla   v29.8h, v18.8h, v1.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%8], #64 \n" // w04 w1_012
-
-                    "fmla   v28.8h, v18.8h, v2.8h       \n"
-                    "fmla   v29.8h, v19.8h, v2.8h       \n"
-                    "fmla   v28.8h, v19.8h, v3.8h       \n"
-                    "fmla   v29.8h, v20.8h, v3.8h       \n"
-
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v22.8h, v23.8h}, [%3], #32 \n" // r1_01
-
-                    "fmla   v28.8h, v20.8h, v4.8h       \n"
-                    "fmla   v29.8h, v21.8h, v4.8h       \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%3] \n" // r1_2345
-
-                    "fmla   v30.8h, v22.8h, v0.8h       \n"
-                    "fmla   v31.8h, v23.8h, v0.8h       \n"
-                    "fmla   v30.8h, v23.8h, v1.8h       \n"
-                    "fmla   v31.8h, v24.8h, v1.8h       \n"
-                    "fmla   v30.8h, v24.8h, v2.8h       \n"
-                    "fmla   v31.8h, v25.8h, v2.8h       \n"
-                    "fmla   v30.8h, v25.8h, v3.8h       \n"
-                    "fmla   v31.8h, v26.8h, v3.8h       \n"
-                    "fmla   v30.8h, v26.8h, v4.8h       \n"
-                    "fmla   v31.8h, v27.8h, v4.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%8], #64 \n" // w1_34 w2_01
-
-                    "fmla   v28.8h, v22.8h, v5.8h       \n"
-                    "fmla   v29.8h, v23.8h, v5.8h       \n"
-                    "fmla   v28.8h, v23.8h, v6.8h       \n"
-                    "fmla   v29.8h, v24.8h, v6.8h       \n"
-                    "fmla   v28.8h, v24.8h, v7.8h       \n"
-                    "fmla   v29.8h, v25.8h, v7.8h       \n"
-                    "fmla   v28.8h, v25.8h, v0.8h       \n"
-                    "fmla   v29.8h, v26.8h, v0.8h       \n"
-
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v16.8h, v17.8h}, [%4], #32 \n" // r2_01
-
-                    "fmla   v28.8h, v26.8h, v1.8h       \n"
-                    "fmla   v29.8h, v27.8h, v1.8h       \n"
-
-                    "prfm   pldl1keep, [%4, #512]       \n"
-                    "ld1    {v18.8h, v19.8h, v20.8h, v21.8h}, [%4] \n" // r2_2345
-
-                    "fmla   v30.8h, v16.8h, v5.8h       \n"
-                    "fmla   v31.8h, v17.8h, v5.8h       \n"
-                    "fmla   v30.8h, v17.8h, v6.8h       \n"
-                    "fmla   v31.8h, v18.8h, v6.8h       \n"
-                    "fmla   v30.8h, v18.8h, v7.8h       \n"
-                    "fmla   v31.8h, v19.8h, v7.8h       \n"
-                    "fmla   v30.8h, v19.8h, v0.8h       \n"
-                    "fmla   v31.8h, v20.8h, v0.8h       \n"
-                    "fmla   v30.8h, v20.8h, v1.8h       \n"
-                    "fmla   v31.8h, v21.8h, v1.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%8], #64 \n" // w2_234 w30
-
-                    "fmla   v28.8h, v16.8h, v2.8h       \n"
-                    "fmla   v29.8h, v17.8h, v2.8h       \n"
-                    "fmla   v28.8h, v17.8h, v3.8h       \n"
-                    "fmla   v29.8h, v18.8h, v3.8h       \n"
-                    "fmla   v28.8h, v18.8h, v4.8h       \n"
-                    "fmla   v29.8h, v19.8h, v4.8h       \n"
-                    "fmla   v28.8h, v19.8h, v5.8h       \n"
-                    "fmla   v29.8h, v20.8h, v5.8h       \n"
-
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v22.8h, v23.8h}, [%5], #32 \n" // r3_01
-
-                    "fmla   v28.8h, v20.8h, v6.8h       \n"
-                    "fmla   v29.8h, v21.8h, v6.8h       \n"
-
-                    "prfm   pldl1keep, [%5, #512]       \n"
-                    "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%5] \n" // r3_2345
-
-                    "fmla   v30.8h, v22.8h, v2.8h       \n"
-                    "fmla   v31.8h, v23.8h, v2.8h       \n"
-                    "fmla   v30.8h, v23.8h, v3.8h       \n"
-                    "fmla   v31.8h, v24.8h, v3.8h       \n"
-                    "fmla   v30.8h, v24.8h, v4.8h       \n"
-                    "fmla   v31.8h, v25.8h, v4.8h       \n"
-                    "fmla   v30.8h, v25.8h, v5.8h       \n"
-                    "fmla   v31.8h, v26.8h, v5.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%8], #64 \n" // w3_1234
-
-                    "fmla   v30.8h, v26.8h, v6.8h       \n"
-                    "fmla   v31.8h, v27.8h, v6.8h       \n"
-
-                    "fmla   v28.8h, v22.8h, v7.8h       \n"
-                    "fmla   v29.8h, v23.8h, v7.8h       \n"
-                    "fmla   v28.8h, v23.8h, v0.8h       \n"
-                    "fmla   v29.8h, v24.8h, v0.8h       \n"
-                    "fmla   v28.8h, v24.8h, v1.8h       \n"
-                    "fmla   v29.8h, v25.8h, v1.8h       \n"
-                    "fmla   v28.8h, v25.8h, v2.8h       \n"
-                    "fmla   v29.8h, v26.8h, v2.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #256]       \n"
-                    "ld1    {v16.8h, v17.8h}, [%6], #32 \n" // r4_01
-
-                    "fmla   v28.8h, v26.8h, v3.8h       \n"
-                    "fmla   v29.8h, v27.8h, v3.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v18.8h, v19.8h, v20.8h, v21.8h}, [%6] \n" // r4_2345
-
-                    "fmla   v30.8h, v16.8h, v7.8h       \n"
-                    "fmla   v31.8h, v17.8h, v7.8h       \n"
-                    "fmla   v30.8h, v17.8h, v0.8h       \n"
-                    "fmla   v31.8h, v18.8h, v0.8h       \n"
-                    "fmla   v30.8h, v18.8h, v1.8h       \n"
-                    "fmla   v31.8h, v19.8h, v1.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%8], #64 \n" // w4_0123
-
-                    "fmla   v30.8h, v19.8h, v2.8h       \n"
-                    "fmla   v31.8h, v20.8h, v2.8h       \n"
-                    "fmla   v30.8h, v20.8h, v3.8h       \n"
-                    "fmla   v31.8h, v21.8h, v3.8h       \n"
-
-                    "fmla   v28.8h, v16.8h, v4.8h       \n"
-                    "fmla   v29.8h, v17.8h, v4.8h       \n"
-                    "fmla   v28.8h, v17.8h, v5.8h       \n"
-                    "fmla   v29.8h, v18.8h, v5.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #128]       \n"
-                    "ld1    {v0.8h}, [%8]               \n" // w44
-
-                    "fmla   v28.8h, v18.8h, v6.8h       \n"
-                    "fmla   v29.8h, v19.8h, v6.8h       \n"
-                    "fmla   v28.8h, v19.8h, v7.8h       \n"
-                    "fmla   v29.8h, v20.8h, v7.8h       \n"
-
-                    "prfm   pldl1keep, [%7, #256]       \n"
-                    "ld1    {v22.8h, v23.8h}, [%7], #32 \n" // r5_01
-
-                    "fmla   v28.8h, v20.8h, v0.8h       \n"
-                    "fmla   v29.8h, v21.8h, v0.8h       \n"
-
-                    "prfm   pldl1keep, [%7, #512]       \n"
-                    "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%7] \n" // r5_2345
-
-                    "fmla   v30.8h, v22.8h, v4.8h       \n"
-                    "fmla   v31.8h, v23.8h, v4.8h       \n"
-                    "fmla   v30.8h, v23.8h, v5.8h       \n"
-                    "fmla   v31.8h, v24.8h, v5.8h       \n"
-                    "fmla   v30.8h, v24.8h, v6.8h       \n"
-                    "fmla   v31.8h, v25.8h, v6.8h       \n"
-                    "fmla   v30.8h, v25.8h, v7.8h       \n"
-                    "fmla   v31.8h, v26.8h, v7.8h       \n"
-                    "fmla   v30.8h, v26.8h, v0.8h       \n"
-                    "fmla   v31.8h, v27.8h, v0.8h       \n"
-
-                    "sub    %8, %8, #384                \n" // k0 -= 24 * 8
-
-                    "st1    {v28.8h, v29.8h}, [%0], #32 \n"
-                    "st1    {v30.8h, v31.8h}, [%1], #32 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3),      // %5
-                    "=r"(r4),      // %6
-                    "=r"(r5),      // %7
-                    "=r"(k0)       // %8
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "7"(r5),
-                    "8"(k0),
-                    "w"(_bias0) // %18
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            for (; j < outw; j++)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "ld1    {v16.8h}, [%2], #16         \n" // r0_0
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v17.8h, v18.8h, v19.8h, v20.8h}, [%2] \n" // r0_1234
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%8], #64 \n" // w0_0123
-
-                    "mov    v30.16b, %18.16b            \n" // sum00
-                    "mov    v31.16b, %18.16b            \n" // sum10
-
-                    "fmla   v30.8h, v16.8h, v0.8h       \n"
-                    "fmla   v30.8h, v17.8h, v1.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%8], #64 \n" // w04 w1_012
-
-                    "fmla   v30.8h, v18.8h, v2.8h       \n"
-
-                    "prfm   pldl1keep, [%3, #128]       \n"
-                    "ld1    {v21.8h}, [%3], #16         \n" // r1_0
-
-                    "fmla   v30.8h, v19.8h, v3.8h       \n"
-                    "fmla   v30.8h, v20.8h, v4.8h       \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v22.8h, v23.8h, v24.8h, v25.8h}, [%3] \n" // r1_1234
-
-                    "fmla   v31.8h, v21.8h, v0.8h       \n"
-                    "fmla   v31.8h, v22.8h, v1.8h       \n"
-                    "fmla   v31.8h, v23.8h, v2.8h       \n"
-                    "fmla   v31.8h, v24.8h, v3.8h       \n"
-                    "fmla   v31.8h, v25.8h, v4.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%8], #64 \n" // w1_34 w2_01
-
-                    "fmla   v30.8h, v21.8h, v5.8h       \n"
-                    "fmla   v30.8h, v22.8h, v6.8h       \n"
-                    "fmla   v30.8h, v23.8h, v7.8h       \n"
-
-                    "prfm   pldl1keep, [%4, #128]       \n"
-                    "ld1    {v16.8h}, [%4], #16         \n" // r2_0
-
-                    "fmla   v30.8h, v24.8h, v0.8h       \n"
-                    "fmla   v30.8h, v25.8h, v1.8h       \n"
-
-                    "prfm   pldl1keep, [%4, #512]       \n"
-                    "ld1    {v17.8h, v18.8h, v19.8h, v20.8h}, [%4] \n" // r2_1234
-
-                    "fmla   v31.8h, v16.8h, v5.8h       \n"
-                    "fmla   v31.8h, v17.8h, v6.8h       \n"
-                    "fmla   v31.8h, v18.8h, v7.8h       \n"
-                    "fmla   v31.8h, v19.8h, v0.8h       \n"
-                    "fmla   v31.8h, v20.8h, v1.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%8], #64 \n" // w2_234 w30
-
-                    "fmla   v30.8h, v16.8h, v2.8h       \n"
-                    "fmla   v30.8h, v17.8h, v3.8h       \n"
-                    "fmla   v30.8h, v18.8h, v4.8h       \n"
-
-                    "prfm   pldl1keep, [%5, #128]       \n"
-                    "ld1    {v21.8h}, [%5], #16         \n" // r3_0
-
-                    "fmla   v30.8h, v19.8h, v5.8h       \n"
-                    "fmla   v30.8h, v20.8h, v6.8h       \n"
-
-                    "prfm   pldl1keep, [%5, #512]       \n"
-                    "ld1    {v22.8h, v23.8h, v24.8h, v25.8h}, [%5] \n" // r3_1234
-
-                    "fmla   v31.8h, v21.8h, v2.8h       \n"
-                    "fmla   v31.8h, v22.8h, v3.8h       \n"
-                    "fmla   v31.8h, v23.8h, v4.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%8], #64 \n" // w3_1234
-
-                    "fmla   v31.8h, v24.8h, v5.8h       \n"
-                    "fmla   v31.8h, v25.8h, v6.8h       \n"
-
-                    "fmla   v30.8h, v21.8h, v7.8h       \n"
-                    "fmla   v30.8h, v22.8h, v0.8h       \n"
-                    "fmla   v30.8h, v23.8h, v1.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #128]       \n"
-                    "ld1    {v16.8h}, [%6], #16         \n" // r4_0
-
-                    "fmla   v30.8h, v24.8h, v2.8h       \n"
-                    "fmla   v30.8h, v25.8h, v3.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v17.8h, v18.8h, v19.8h, v20.8h}, [%6] \n" // r4_1234
-
-                    "fmla   v31.8h, v16.8h, v7.8h       \n"
-                    "fmla   v31.8h, v17.8h, v0.8h       \n"
-                    "fmla   v31.8h, v18.8h, v1.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%8], #64 \n" // w4_0123
-
-                    "fmla   v31.8h, v19.8h, v2.8h       \n"
-                    "fmla   v31.8h, v20.8h, v3.8h       \n"
-
-                    "fmla   v30.8h, v16.8h, v4.8h       \n"
-                    "fmla   v30.8h, v17.8h, v5.8h       \n"
-
-                    "prfm   pldl1keep, [%8, #128]       \n"
-                    "ld1    {v0.8h}, [%8]               \n" // w44
-
-                    "fmla   v30.8h, v18.8h, v6.8h       \n"
-
-                    "prfm   pldl1keep, [%7, #128]       \n"
-                    "ld1    {v21.8h}, [%7], #16         \n" // r5_0
-
-                    "fmla   v30.8h, v19.8h, v7.8h       \n"
-                    "fmla   v30.8h, v20.8h, v0.8h       \n"
-
-                    "prfm   pldl1keep, [%7, #512]       \n"
-                    "ld1    {v22.8h, v23.8h, v24.8h, v25.8h}, [%7] \n" // r5_1234
-
-                    "fmla   v31.8h, v21.8h, v4.8h       \n"
-                    "fmla   v31.8h, v22.8h, v5.8h       \n"
-                    "fmla   v31.8h, v23.8h, v6.8h       \n"
-                    "fmla   v31.8h, v24.8h, v7.8h       \n"
-                    "fmla   v31.8h, v25.8h, v0.8h       \n"
-
-                    "sub    %8, %8, #384                \n" // k0 -= 24 * 8
-
-                    "st1    {v30.8h}, [%0], #16         \n"
-                    "st1    {v31.8h}, [%1], #16         \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(outptr1), // %1
-                    "=r"(r0),      // %2
-                    "=r"(r1),      // %3
-                    "=r"(r2),      // %4
-                    "=r"(r3),      // %5
-                    "=r"(r4),      // %6
-                    "=r"(r5),      // %7
-                    "=r"(k0)       // %8
-                    : "0"(outptr0),
-                    "1"(outptr1),
-                    "2"(r0),
-                    "3"(r1),
-                    "4"(r2),
-                    "5"(r3),
-                    "6"(r4),
-                    "7"(r5),
-                    "8"(k0),
-                    "w"(_bias0) // %18
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v30", "v31");
-            }
-
-            r0 += 4 * 8 + w * 8;
-            r1 += 4 * 8 + w * 8;
-            r2 += 4 * 8 + w * 8;
-            r3 += 4 * 8 + w * 8;
-            r4 += 4 * 8 + w * 8;
-            r5 += 4 * 8 + w * 8;
-
-            outptr0 += outw * 8;
-            outptr1 += outw * 8;
-        }
-
-        float16x8_t _bias0 = bias ? vld1q_f16(bias + g * 8) : vdupq_n_f16((__fp16)0.f);
-
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%1], #64 \n" // r0_0123
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%6], #64 \n" // w0_0123
-
-                    "mov    v28.16b, %14.16b            \n" // sum00
-                    "mov    v29.16b, %14.16b            \n" // sum01
-                    "mov    v30.16b, %14.16b            \n" // sum02
-                    "mov    v31.16b, %14.16b            \n" // sum03
-
-                    "fmla   v28.8h, v12.8h, v0.8h       \n"
-                    "fmla   v29.8h, v13.8h, v0.8h       \n"
-                    "fmla   v30.8h, v14.8h, v0.8h       \n"
-                    "fmla   v31.8h, v15.8h, v0.8h       \n"
-
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%1] \n" // r0_4567
-
-                    "fmla   v28.8h, v13.8h, v1.8h       \n"
-                    "fmla   v29.8h, v14.8h, v1.8h       \n"
-                    "fmla   v30.8h, v15.8h, v1.8h       \n"
-                    "fmla   v31.8h, v16.8h, v1.8h       \n"
-
-                    "fmla   v28.8h, v14.8h, v2.8h       \n"
-                    "fmla   v29.8h, v15.8h, v2.8h       \n"
-                    "fmla   v30.8h, v16.8h, v2.8h       \n"
-                    "fmla   v31.8h, v17.8h, v2.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%6], #64 \n" // w04 w1_012
-
-                    "fmla   v28.8h, v15.8h, v3.8h       \n"
-                    "fmla   v29.8h, v16.8h, v3.8h       \n"
-                    "fmla   v30.8h, v17.8h, v3.8h       \n"
-                    "fmla   v31.8h, v18.8h, v3.8h       \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%2], #64 \n" // r1_0123
-
-                    "fmla   v28.8h, v16.8h, v4.8h       \n"
-                    "fmla   v29.8h, v17.8h, v4.8h       \n"
-                    "fmla   v30.8h, v18.8h, v4.8h       \n"
-                    "fmla   v31.8h, v19.8h, v4.8h       \n"
-
-                    "fmla   v28.8h, v20.8h, v5.8h       \n"
-                    "fmla   v29.8h, v21.8h, v5.8h       \n"
-                    "fmla   v30.8h, v22.8h, v5.8h       \n"
-                    "fmla   v31.8h, v23.8h, v5.8h       \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%2] \n" // r1_4567
-
-                    "fmla   v28.8h, v21.8h, v6.8h       \n"
-                    "fmla   v29.8h, v22.8h, v6.8h       \n"
-                    "fmla   v30.8h, v23.8h, v6.8h       \n"
-                    "fmla   v31.8h, v24.8h, v6.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%6], #64 \n" // w1_34 w2_01
-
-                    "fmla   v28.8h, v22.8h, v7.8h       \n"
-                    "fmla   v29.8h, v23.8h, v7.8h       \n"
-                    "fmla   v30.8h, v24.8h, v7.8h       \n"
-                    "fmla   v31.8h, v25.8h, v7.8h       \n"
-
-                    "fmla   v28.8h, v23.8h, v0.8h       \n"
-                    "fmla   v29.8h, v24.8h, v0.8h       \n"
-                    "fmla   v30.8h, v25.8h, v0.8h       \n"
-                    "fmla   v31.8h, v26.8h, v0.8h       \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%3], #64 \n" // r2_0123
-
-                    "fmla   v28.8h, v24.8h, v1.8h       \n"
-                    "fmla   v29.8h, v25.8h, v1.8h       \n"
-                    "fmla   v30.8h, v26.8h, v1.8h       \n"
-                    "fmla   v31.8h, v27.8h, v1.8h       \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%3] \n" // r2_4567
-
-                    "fmla   v28.8h, v12.8h, v2.8h       \n"
-                    "fmla   v29.8h, v13.8h, v2.8h       \n"
-                    "fmla   v30.8h, v14.8h, v2.8h       \n"
-                    "fmla   v31.8h, v15.8h, v2.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%6], #64 \n" // w2_234 w30
-
-                    "fmla   v28.8h, v13.8h, v3.8h       \n"
-                    "fmla   v29.8h, v14.8h, v3.8h       \n"
-                    "fmla   v30.8h, v15.8h, v3.8h       \n"
-                    "fmla   v31.8h, v16.8h, v3.8h       \n"
-
-                    "fmla   v28.8h, v14.8h, v4.8h       \n"
-                    "fmla   v29.8h, v15.8h, v4.8h       \n"
-                    "fmla   v30.8h, v16.8h, v4.8h       \n"
-                    "fmla   v31.8h, v17.8h, v4.8h       \n"
-
-                    "fmla   v28.8h, v15.8h, v5.8h       \n"
-                    "fmla   v29.8h, v16.8h, v5.8h       \n"
-                    "fmla   v30.8h, v17.8h, v5.8h       \n"
-                    "fmla   v31.8h, v18.8h, v5.8h       \n"
-
-                    "prfm   pldl1keep, [%4, #512]       \n"
-                    "ld1    {v20.8h, v21.8h, v22.8h, v23.8h}, [%4], #64 \n" // r3_0123
-
-                    "fmla   v28.8h, v16.8h, v6.8h       \n"
-                    "fmla   v29.8h, v17.8h, v6.8h       \n"
-                    "fmla   v30.8h, v18.8h, v6.8h       \n"
-                    "fmla   v31.8h, v19.8h, v6.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%6], #64 \n" // w3_1234
-
-                    "fmla   v28.8h, v20.8h, v7.8h       \n"
-                    "fmla   v29.8h, v21.8h, v7.8h       \n"
-                    "fmla   v30.8h, v22.8h, v7.8h       \n"
-                    "fmla   v31.8h, v23.8h, v7.8h       \n"
-
-                    "prfm   pldl1keep, [%4, #512]       \n"
-                    "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%4] \n" // r3_4567
-
-                    "fmla   v28.8h, v21.8h, v0.8h       \n"
-                    "fmla   v29.8h, v22.8h, v0.8h       \n"
-                    "fmla   v30.8h, v23.8h, v0.8h       \n"
-                    "fmla   v31.8h, v24.8h, v0.8h       \n"
-
-                    "fmla   v28.8h, v22.8h, v1.8h       \n"
-                    "fmla   v29.8h, v23.8h, v1.8h       \n"
-                    "fmla   v30.8h, v24.8h, v1.8h       \n"
-                    "fmla   v31.8h, v25.8h, v1.8h       \n"
-
-                    "prfm   pldl1keep, [%5, #512]       \n"
-                    "ld1    {v12.8h, v13.8h, v14.8h, v15.8h}, [%5], #64 \n" // r4_0123
-
-                    "fmla   v28.8h, v23.8h, v2.8h       \n"
-                    "fmla   v29.8h, v24.8h, v2.8h       \n"
-                    "fmla   v30.8h, v25.8h, v2.8h       \n"
-                    "fmla   v31.8h, v26.8h, v2.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%6], #64 \n" // w4_0123
-
-                    "fmla   v28.8h, v24.8h, v3.8h       \n"
-                    "fmla   v29.8h, v25.8h, v3.8h       \n"
-                    "fmla   v30.8h, v26.8h, v3.8h       \n"
-                    "fmla   v31.8h, v27.8h, v3.8h       \n"
-
-                    "fmla   v28.8h, v12.8h, v4.8h       \n"
-                    "fmla   v29.8h, v13.8h, v4.8h       \n"
-                    "fmla   v30.8h, v14.8h, v4.8h       \n"
-                    "fmla   v31.8h, v15.8h, v4.8h       \n"
-
-                    "prfm   pldl1keep, [%5, #512]       \n"
-                    "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%5] \n" // r4_4567
-
-                    "fmla   v28.8h, v13.8h, v5.8h       \n"
-                    "fmla   v29.8h, v14.8h, v5.8h       \n"
-                    "fmla   v30.8h, v15.8h, v5.8h       \n"
-                    "fmla   v31.8h, v16.8h, v5.8h       \n"
-
-                    "fmla   v28.8h, v14.8h, v6.8h       \n"
-                    "fmla   v29.8h, v15.8h, v6.8h       \n"
-                    "fmla   v30.8h, v16.8h, v6.8h       \n"
-                    "fmla   v31.8h, v17.8h, v6.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #128]       \n"
-                    "ld1    {v0.8h}, [%6]               \n" // w44
-
-                    "fmla   v28.8h, v15.8h, v7.8h       \n"
-                    "fmla   v29.8h, v16.8h, v7.8h       \n"
-                    "fmla   v30.8h, v17.8h, v7.8h       \n"
-                    "fmla   v31.8h, v18.8h, v7.8h       \n"
-
-                    "fmla   v28.8h, v16.8h, v0.8h       \n"
-                    "fmla   v29.8h, v17.8h, v0.8h       \n"
-                    "fmla   v30.8h, v18.8h, v0.8h       \n"
-                    "fmla   v31.8h, v19.8h, v0.8h       \n"
-
-                    "sub    %6, %6, #384                \n" // k0 -= 24 * 8
-
-                    "st1    {v28.8h, v29.8h, v30.8h, v31.8h}, [%0], #64 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2),      // %3
-                    "=r"(r3),      // %4
-                    "=r"(r4),      // %5
-                    "=r"(k0)       // %6
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "4"(r3),
-                    "5"(r4),
-                    "6"(k0),
-                    "w"(_bias0) // %14
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #256]       \n"
-                    "ld1    {v16.8h, v17.8h}, [%1], #32 \n" // r0_01
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%6], #64 \n" // w0_0123
-
-                    "mov    v30.16b, %14.16b            \n" // sum00
-                    "mov    v31.16b, %14.16b            \n" // sum01
-
-                    "fmla   v30.8h, v16.8h, v0.8h       \n"
-                    "fmla   v31.8h, v17.8h, v0.8h       \n"
-
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v18.8h, v19.8h, v20.8h, v21.8h}, [%1] \n" // r0_2345
-
-                    "fmla   v30.8h, v17.8h, v1.8h       \n"
-                    "fmla   v31.8h, v18.8h, v1.8h       \n"
-                    "fmla   v30.8h, v18.8h, v2.8h       \n"
-                    "fmla   v31.8h, v19.8h, v2.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%6], #64 \n" // w04 w1_012
-
-                    "fmla   v30.8h, v19.8h, v3.8h       \n"
-                    "fmla   v31.8h, v20.8h, v3.8h       \n"
-
-                    "prfm   pldl1keep, [%2, #256]       \n"
-                    "ld1    {v22.8h, v23.8h}, [%2], #32 \n" // r1_01
-
-                    "fmla   v30.8h, v20.8h, v4.8h       \n"
-                    "fmla   v31.8h, v21.8h, v4.8h       \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%2] \n" // r1_2345
-
-                    "fmla   v30.8h, v22.8h, v5.8h       \n"
-                    "fmla   v31.8h, v23.8h, v5.8h       \n"
-                    "fmla   v30.8h, v23.8h, v6.8h       \n"
-                    "fmla   v31.8h, v24.8h, v6.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%6], #64 \n" // w1_34 w2_01
-
-                    "fmla   v30.8h, v24.8h, v7.8h       \n"
-                    "fmla   v31.8h, v25.8h, v7.8h       \n"
-                    "fmla   v30.8h, v25.8h, v0.8h       \n"
-                    "fmla   v31.8h, v26.8h, v0.8h       \n"
-
-                    "prfm   pldl1keep, [%3, #256]       \n"
-                    "ld1    {v16.8h, v17.8h}, [%3], #32 \n" // r2_01
-
-                    "fmla   v30.8h, v26.8h, v1.8h       \n"
-                    "fmla   v31.8h, v27.8h, v1.8h       \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v18.8h, v19.8h, v20.8h, v21.8h}, [%3] \n" // r2_2345
-
-                    "fmla   v30.8h, v16.8h, v2.8h       \n"
-                    "fmla   v31.8h, v17.8h, v2.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%6], #64 \n" // w2_234 w30
-
-                    "fmla   v30.8h, v17.8h, v3.8h       \n"
-                    "fmla   v31.8h, v18.8h, v3.8h       \n"
-                    "fmla   v30.8h, v18.8h, v4.8h       \n"
-                    "fmla   v31.8h, v19.8h, v4.8h       \n"
-
-                    "prfm   pldl1keep, [%4, #256]       \n"
-                    "ld1    {v22.8h, v23.8h}, [%4], #32 \n" // r3_01
-
-                    "fmla   v30.8h, v19.8h, v5.8h       \n"
-                    "fmla   v31.8h, v20.8h, v5.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%6], #64 \n" // w3_1234
-
-                    "fmla   v30.8h, v20.8h, v6.8h       \n"
-                    "fmla   v31.8h, v21.8h, v6.8h       \n"
-
-                    "prfm   pldl1keep, [%4, #512]       \n"
-                    "ld1    {v24.8h, v25.8h, v26.8h, v27.8h}, [%4] \n" // r3_2345
-
-                    "fmla   v30.8h, v22.8h, v7.8h       \n"
-                    "fmla   v31.8h, v23.8h, v7.8h       \n"
-                    "fmla   v30.8h, v23.8h, v0.8h       \n"
-                    "fmla   v31.8h, v24.8h, v0.8h       \n"
-                    "fmla   v30.8h, v24.8h, v1.8h       \n"
-                    "fmla   v31.8h, v25.8h, v1.8h       \n"
-
-                    "prfm   pldl1keep, [%5, #256]       \n"
-                    "ld1    {v16.8h, v17.8h}, [%5], #32 \n" // r4_01
-
-                    "fmla   v30.8h, v25.8h, v2.8h       \n"
-                    "fmla   v31.8h, v26.8h, v2.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%6], #64 \n" // w4_0123
-
-                    "fmla   v30.8h, v26.8h, v3.8h       \n"
-                    "fmla   v31.8h, v27.8h, v3.8h       \n"
-
-                    "prfm   pldl1keep, [%5, #512]       \n"
-                    "ld1    {v18.8h, v19.8h, v20.8h, v21.8h}, [%5] \n" // r4_2345
-
-                    "fmla   v30.8h, v16.8h, v4.8h       \n"
-                    "fmla   v31.8h, v17.8h, v4.8h       \n"
-                    "fmla   v30.8h, v17.8h, v5.8h       \n"
-                    "fmla   v31.8h, v18.8h, v5.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #128]       \n"
-                    "ld1    {v0.8h}, [%6]               \n" // w44
-
-                    "fmla   v30.8h, v18.8h, v6.8h       \n"
-                    "fmla   v31.8h, v19.8h, v6.8h       \n"
-                    "fmla   v30.8h, v19.8h, v7.8h       \n"
-                    "fmla   v31.8h, v20.8h, v7.8h       \n"
-                    "fmla   v30.8h, v20.8h, v0.8h       \n"
-                    "fmla   v31.8h, v21.8h, v0.8h       \n"
-
-                    "sub    %6, %6, #384                \n" // k0 -= 24 * 8
-
-                    "st1    {v30.8h, v31.8h}, [%0], #32 \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2),      // %3
-                    "=r"(r3),      // %4
-                    "=r"(r4),      // %5
-                    "=r"(k0)       // %6
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "4"(r3),
-                    "5"(r4),
-                    "6"(k0),
-                    "w"(_bias0) // %14
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v30", "v31");
-            }
-            for (; j < outw; j++)
-            {
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #128]       \n"
-                    "ld1    {v16.8h}, [%1], #16         \n" // r0_0
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%6], #64 \n" // w0_0123
-
-                    "mov    v30.16b, %14.16b            \n" // sum00
-
-                    "prfm   pldl1keep, [%1, #512]       \n"
-                    "ld1    {v17.8h, v18.8h, v19.8h, v20.8h}, [%1] \n" // r0_1234
-
-                    "fmla   v30.8h, v16.8h, v0.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%6], #64 \n" // w04 w1_012
-
-                    "fmla   v30.8h, v17.8h, v1.8h       \n"
-
-                    "fmla   v30.8h, v18.8h, v2.8h       \n"
-
-                    "prfm   pldl1keep, [%2, #128]       \n"
-                    "ld1    {v21.8h}, [%2], #16         \n" // r1_0
-
-                    "fmla   v30.8h, v19.8h, v3.8h       \n"
-
-                    "prfm   pldl1keep, [%2, #512]       \n"
-                    "ld1    {v22.8h, v23.8h, v24.8h, v25.8h}, [%2] \n" // r1_1234
-
-                    "fmla   v30.8h, v20.8h, v4.8h       \n"
-
-                    "fmla   v30.8h, v21.8h, v5.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%6], #64 \n" // w1_34 w2_01
-
-                    "fmla   v30.8h, v22.8h, v6.8h       \n"
-
-                    "fmla   v30.8h, v23.8h, v7.8h       \n"
-
-                    "prfm   pldl1keep, [%3, #128]       \n"
-                    "ld1    {v16.8h}, [%3], #16         \n" // r2_0
-
-                    "fmla   v30.8h, v24.8h, v0.8h       \n"
-
-                    "prfm   pldl1keep, [%3, #512]       \n"
-                    "ld1    {v17.8h, v18.8h, v19.8h, v20.8h}, [%3] \n" // r2_1234
-
-                    "fmla   v30.8h, v25.8h, v1.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%6], #64 \n" // w2_234 w30
-
-                    "fmla   v30.8h, v16.8h, v2.8h       \n"
-                    "fmla   v30.8h, v17.8h, v3.8h       \n"
-
-                    "prfm   pldl1keep, [%4, #128]       \n"
-                    "ld1    {v21.8h}, [%4], #16         \n" // r3_0
-
-                    "fmla   v30.8h, v18.8h, v4.8h       \n"
-
-                    "prfm   pldl1keep, [%4, #512]       \n"
-                    "ld1    {v22.8h, v23.8h, v24.8h, v25.8h}, [%4] \n" // r3_1234
-
-                    "fmla   v30.8h, v19.8h, v5.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%6], #64 \n" // w3_1234
-
-                    "fmla   v30.8h, v20.8h, v6.8h       \n"
-
-                    "fmla   v30.8h, v21.8h, v7.8h       \n"
-                    "fmla   v30.8h, v22.8h, v0.8h       \n"
-
-                    "prfm   pldl1keep, [%5, #128]       \n"
-                    "ld1    {v16.8h}, [%5], #16         \n" // r4_0
-
-                    "fmla   v30.8h, v23.8h, v1.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #512]       \n"
-                    "ld1    {v4.8h, v5.8h, v6.8h, v7.8h}, [%6], #64 \n" // w4_0123
-
-                    "fmla   v30.8h, v24.8h, v2.8h       \n"
-
-                    "prfm   pldl1keep, [%5, #512]       \n"
-                    "ld1    {v17.8h, v18.8h, v19.8h, v20.8h}, [%5] \n" // r4_1234
-
-                    "fmla   v30.8h, v25.8h, v3.8h       \n"
-
-                    "fmla   v30.8h, v16.8h, v4.8h       \n"
-                    "fmla   v30.8h, v17.8h, v5.8h       \n"
-
-                    "prfm   pldl1keep, [%6, #128]       \n"
-                    "ld1    {v0.8h}, [%6]               \n" // w44
-
-                    "fmla   v30.8h, v18.8h, v6.8h       \n"
-                    "fmla   v30.8h, v19.8h, v7.8h       \n"
-                    "fmla   v30.8h, v20.8h, v0.8h       \n"
-
-                    "sub    %6, %6, #384                \n" // k0 -= 24 * 8
-
-                    "st1    {v30.8h}, [%0], #16         \n"
-
-                    : "=r"(outptr0), // %0
-                    "=r"(r0),      // %1
-                    "=r"(r1),      // %2
-                    "=r"(r2),      // %3
-                    "=r"(r3),      // %4
-                    "=r"(r4),      // %5
-                    "=r"(k0)       // %6
-                    : "0"(outptr0),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "4"(r3),
-                    "5"(r4),
-                    "6"(k0),
-                    "w"(_bias0) // %14
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v30");
-            }
-
-            r0 += 4 * 8;
-            r1 += 4 * 8;
-            r2 += 4 * 8;
-            r3 += 4 * 8;
-            r4 += 4 * 8;
-        }
-    }
-}
-
-static void convdw5x5s2_pack8_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 8;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        float16x8_t _bias0 = bias ? vld1q_f16(bias + g * 8) : vdupq_n_f16((__fp16)0.f);
-
-        const __fp16* k0 = kernel.row<const __fp16>(g);
-
-        __fp16* outptr0 = out.row<__fp16>(0);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const __fp16* r0 = img0.row<const __fp16>(0);
-        const __fp16* r1 = img0.row<const __fp16>(1);
-        const __fp16* r2 = img0.row<const __fp16>(2);
-        const __fp16* r3 = img0.row<const __fp16>(3);
-        const __fp16* r4 = img0.row<const __fp16>(4);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                float16x8_t _sum0 = _bias0;
-
-                float16x8_t _r00 = vld1q_f16(r0);
-                float16x8_t _r01 = vld1q_f16(r0 + 8);
-                float16x8_t _r02 = vld1q_f16(r0 + 16);
-                float16x8_t _r03 = vld1q_f16(r0 + 24);
-                float16x8_t _r04 = vld1q_f16(r0 + 32);
-
-                float16x8_t _k00 = vld1q_f16(k0);
-                float16x8_t _k01 = vld1q_f16(k0 + 8);
-                float16x8_t _k02 = vld1q_f16(k0 + 16);
-                float16x8_t _k03 = vld1q_f16(k0 + 24);
-                float16x8_t _k04 = vld1q_f16(k0 + 32);
-                k0 += 40;
-
-                _sum0 = vfmaq_f16(_sum0, _k00, _r00);
-                _sum0 = vfmaq_f16(_sum0, _k01, _r01);
-                _sum0 = vfmaq_f16(_sum0, _k02, _r02);
-                _sum0 = vfmaq_f16(_sum0, _k03, _r03);
-                _sum0 = vfmaq_f16(_sum0, _k04, _r04);
-
-                float16x8_t _r10 = vld1q_f16(r1);
-                float16x8_t _r11 = vld1q_f16(r1 + 8);
-                float16x8_t _r12 = vld1q_f16(r1 + 16);
-                float16x8_t _r13 = vld1q_f16(r1 + 24);
-                float16x8_t _r14 = vld1q_f16(r1 + 32);
-
-                float16x8_t _k10 = vld1q_f16(k0);
-                float16x8_t _k11 = vld1q_f16(k0 + 8);
-                float16x8_t _k12 = vld1q_f16(k0 + 16);
-                float16x8_t _k13 = vld1q_f16(k0 + 24);
-                float16x8_t _k14 = vld1q_f16(k0 + 32);
-                k0 += 40;
-
-                _sum0 = vfmaq_f16(_sum0, _k10, _r10);
-                _sum0 = vfmaq_f16(_sum0, _k11, _r11);
-                _sum0 = vfmaq_f16(_sum0, _k12, _r12);
-                _sum0 = vfmaq_f16(_sum0, _k13, _r13);
-                _sum0 = vfmaq_f16(_sum0, _k14, _r14);
-
-                float16x8_t _r20 = vld1q_f16(r2);
-                float16x8_t _r21 = vld1q_f16(r2 + 8);
-                float16x8_t _r22 = vld1q_f16(r2 + 16);
-                float16x8_t _r23 = vld1q_f16(r2 + 24);
-                float16x8_t _r24 = vld1q_f16(r2 + 32);
-
-                float16x8_t _k20 = vld1q_f16(k0);
-                float16x8_t _k21 = vld1q_f16(k0 + 8);
-                float16x8_t _k22 = vld1q_f16(k0 + 16);
-                float16x8_t _k23 = vld1q_f16(k0 + 24);
-                float16x8_t _k24 = vld1q_f16(k0 + 32);
-                k0 += 40;
-
-                _sum0 = vfmaq_f16(_sum0, _k20, _r20);
-                _sum0 = vfmaq_f16(_sum0, _k21, _r21);
-                _sum0 = vfmaq_f16(_sum0, _k22, _r22);
-                _sum0 = vfmaq_f16(_sum0, _k23, _r23);
-                _sum0 = vfmaq_f16(_sum0, _k24, _r24);
-
-                float16x8_t _r30 = vld1q_f16(r3);
-                float16x8_t _r31 = vld1q_f16(r3 + 8);
-                float16x8_t _r32 = vld1q_f16(r3 + 16);
-                float16x8_t _r33 = vld1q_f16(r3 + 24);
-                float16x8_t _r34 = vld1q_f16(r3 + 32);
-
-                float16x8_t _k30 = vld1q_f16(k0);
-                float16x8_t _k31 = vld1q_f16(k0 + 8);
-                float16x8_t _k32 = vld1q_f16(k0 + 16);
-                float16x8_t _k33 = vld1q_f16(k0 + 24);
-                float16x8_t _k34 = vld1q_f16(k0 + 32);
-                k0 += 40;
-
-                _sum0 = vfmaq_f16(_sum0, _k30, _r30);
-                _sum0 = vfmaq_f16(_sum0, _k31, _r31);
-                _sum0 = vfmaq_f16(_sum0, _k32, _r32);
-                _sum0 = vfmaq_f16(_sum0, _k33, _r33);
-                _sum0 = vfmaq_f16(_sum0, _k34, _r34);
-
-                float16x8_t _r40 = vld1q_f16(r4);
-                float16x8_t _r41 = vld1q_f16(r4 + 8);
-                float16x8_t _r42 = vld1q_f16(r4 + 16);
-                float16x8_t _r43 = vld1q_f16(r4 + 24);
-                float16x8_t _r44 = vld1q_f16(r4 + 32);
-
-                float16x8_t _k40 = vld1q_f16(k0);
-                float16x8_t _k41 = vld1q_f16(k0 + 8);
-                float16x8_t _k42 = vld1q_f16(k0 + 16);
-                float16x8_t _k43 = vld1q_f16(k0 + 24);
-                float16x8_t _k44 = vld1q_f16(k0 + 32);
-                k0 -= 160;
-
-                _sum0 = vfmaq_f16(_sum0, _k40, _r40);
-                _sum0 = vfmaq_f16(_sum0, _k41, _r41);
-                _sum0 = vfmaq_f16(_sum0, _k42, _r42);
-                _sum0 = vfmaq_f16(_sum0, _k43, _r43);
-                _sum0 = vfmaq_f16(_sum0, _k44, _r44);
-
-                vst1q_f16(outptr0, _sum0);
-
-                outptr0 += 8;
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                r3 += 16;
-                r4 += 16;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-            r3 += tailstep;
-            r4 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_arm.h
deleted file mode 100644
index 8536c081..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/convolutiondepthwise_arm.h
+++ /dev/null
@@ -1,61 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTIONDEPTHWISE_ARM_H
-#define LAYER_CONVOLUTIONDEPTHWISE_ARM_H
-
-#include "convolutiondepthwise.h"
-
-namespace ncnn {
-
-class ConvolutionDepthWise_arm : public ConvolutionDepthWise
-{
-public:
-    ConvolutionDepthWise_arm();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    int create_group_ops(const Option& opt);
-#if NCNN_ARM82
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_INT8
-    int create_pipeline_int8_arm(const Option& opt);
-    int forward_int8_arm(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Layer* activation;
-    std::vector<ncnn::Layer*> group_ops;
-
-    Mat weight_data_tm;
-
-    // fp16
-    Mat bias_data_fp16;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTIONDEPTHWISE_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/crop_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/crop_arm.h
deleted file mode 100644
index 9f2bea6e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/crop_arm.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CROP_ARM_H
-#define LAYER_CROP_ARM_H
-
-#include "crop.h"
-
-namespace ncnn {
-
-class Crop_arm : public Crop
-{
-public:
-    Crop_arm();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CROP_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_3x3.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_3x3.h
deleted file mode 100644
index 30df240f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_3x3.h
+++ /dev/null
@@ -1,267 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconv3x3s1_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outch = top_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch * 9 + q * 9;
-
-            const float* r0 = img0;
-
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-#if __ARM_NEON
-            float32x4_t _k0 = vld1q_f32(k0);
-            float32x4_t _k1 = vld1q_f32(k1);
-            float32x4_t _k2 = vld1q_f32(k2);
-#endif // __ARM_NEON
-
-            for (int i = 0; i < h; i++)
-            {
-                float* outptr = out.row(i);
-
-                float* outptr0 = outptr;
-                float* outptr1 = outptr + outw;
-                float* outptr2 = outptr + outw * 2;
-
-                int j = 0;
-
-#if __ARM_NEON
-                for (; j + 3 < w; j += 4)
-                {
-                    float32x4_t _v = vld1q_f32(r0);
-
-                    //
-                    float32x4_t _out00 = vld1q_f32(outptr0 + 0);
-                    _out00 = vmlaq_lane_f32(_out00, _v, vget_low_f32(_k0), 0);
-                    vst1q_f32(outptr0 + 0, _out00);
-
-                    float32x4_t _out01 = vld1q_f32(outptr0 + 1);
-                    _out01 = vmlaq_lane_f32(_out01, _v, vget_low_f32(_k0), 1);
-                    vst1q_f32(outptr0 + 1, _out01);
-
-                    float32x4_t _out02 = vld1q_f32(outptr0 + 2);
-                    _out02 = vmlaq_lane_f32(_out02, _v, vget_high_f32(_k0), 0);
-                    vst1q_f32(outptr0 + 2, _out02);
-
-                    //
-                    float32x4_t _out10 = vld1q_f32(outptr1 + 0);
-                    _out10 = vmlaq_lane_f32(_out10, _v, vget_low_f32(_k1), 0);
-                    vst1q_f32(outptr1 + 0, _out10);
-
-                    float32x4_t _out11 = vld1q_f32(outptr1 + 1);
-                    _out11 = vmlaq_lane_f32(_out11, _v, vget_low_f32(_k1), 1);
-                    vst1q_f32(outptr1 + 1, _out11);
-
-                    float32x4_t _out12 = vld1q_f32(outptr1 + 2);
-                    _out12 = vmlaq_lane_f32(_out12, _v, vget_high_f32(_k1), 0);
-                    vst1q_f32(outptr1 + 2, _out12);
-
-                    //
-                    float32x4_t _out20 = vld1q_f32(outptr2 + 0);
-                    _out20 = vmlaq_lane_f32(_out20, _v, vget_low_f32(_k2), 0);
-                    vst1q_f32(outptr2 + 0, _out20);
-
-                    float32x4_t _out21 = vld1q_f32(outptr2 + 1);
-                    _out21 = vmlaq_lane_f32(_out21, _v, vget_low_f32(_k2), 1);
-                    vst1q_f32(outptr2 + 1, _out21);
-
-                    float32x4_t _out22 = vld1q_f32(outptr2 + 2);
-                    _out22 = vmlaq_lane_f32(_out22, _v, vget_high_f32(_k2), 0);
-                    vst1q_f32(outptr2 + 2, _out22);
-
-                    r0 += 4;
-                    outptr0 += 4;
-                    outptr1 += 4;
-                    outptr2 += 4;
-                }
-#endif // __ARM_NEON
-
-                for (; j < w; j++)
-                {
-                    float val = r0[0];
-
-                    outptr0[0] += val * k0[0];
-                    outptr0[1] += val * k0[1];
-                    outptr0[2] += val * k0[2];
-
-                    outptr1[0] += val * k1[0];
-                    outptr1[1] += val * k1[1];
-                    outptr1[2] += val * k1[2];
-
-                    outptr2[0] += val * k2[0];
-                    outptr2[1] += val * k2[1];
-                    outptr2[2] += val * k2[2];
-
-                    r0++;
-                    outptr0++;
-                    outptr1++;
-                    outptr2++;
-                }
-            }
-        }
-    }
-}
-
-static void deconv3x3s2_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outch = top_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch * 9 + q * 9;
-
-            const float* r0 = img0;
-
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-#if __ARM_NEON
-            float32x4_t _k0 = vld1q_f32(k0);
-            float32x4_t _k1 = vld1q_f32(k1);
-            float32x4_t _k2 = vld1q_f32(k2);
-#endif // __ARM_NEON
-
-            for (int i = 0; i < h; i++)
-            {
-                float* outptr = out.row(i * 2);
-
-                float* outptr0 = outptr;
-                float* outptr1 = outptr0 + outw;
-                float* outptr2 = outptr1 + outw;
-
-                int j = 0;
-#if __ARM_NEON
-                for (; j + 3 < w; j += 4)
-                {
-                    float32x4_t _v = vld1q_f32(r0);
-
-                    // out row 0
-                    float32x4_t _out00 = vmulq_lane_f32(_v, vget_low_f32(_k0), 0);  // 0,2,4,6
-                    float32x4_t _out01 = vmulq_lane_f32(_v, vget_low_f32(_k0), 1);  // 1,3,5,7
-                    float32x4_t _out02 = vmulq_lane_f32(_v, vget_high_f32(_k0), 0); // 2,4,6,8
-
-                    float32x4x2_t _out0 = vld2q_f32(outptr0);
-                    _out0.val[0] = vaddq_f32(_out0.val[0], _out00); // 0,2,4,6
-                    _out0.val[1] = vaddq_f32(_out0.val[1], _out01); // 1,3,5,7
-                    vst2q_f32(outptr0, _out0);
-
-                    _out0 = vld2q_f32(outptr0 + 2);
-                    _out0.val[0] = vaddq_f32(_out0.val[0], _out02); // 2,4,6,8
-                    vst2q_f32(outptr0 + 2, _out0);
-
-                    // out row 1
-                    float32x4_t _out10 = vmulq_lane_f32(_v, vget_low_f32(_k1), 0);  // 0,2,4,6
-                    float32x4_t _out11 = vmulq_lane_f32(_v, vget_low_f32(_k1), 1);  // 1,3,5,7
-                    float32x4_t _out12 = vmulq_lane_f32(_v, vget_high_f32(_k1), 0); // 2,4,6,8
-
-                    float32x4x2_t _out1 = vld2q_f32(outptr1);
-                    _out1.val[0] = vaddq_f32(_out1.val[0], _out10); // 0,2,4,6
-                    _out1.val[1] = vaddq_f32(_out1.val[1], _out11); // 1,3,5,7
-                    vst2q_f32(outptr1, _out1);
-
-                    _out1 = vld2q_f32(outptr1 + 2);
-                    _out1.val[0] = vaddq_f32(_out1.val[0], _out12); // 2,4,6,8
-                    vst2q_f32(outptr1 + 2, _out1);
-
-                    // out row 2
-                    float32x4_t _out20 = vmulq_lane_f32(_v, vget_low_f32(_k2), 0);  // 0,2,4,6
-                    float32x4_t _out21 = vmulq_lane_f32(_v, vget_low_f32(_k2), 1);  // 1,3,5,7
-                    float32x4_t _out22 = vmulq_lane_f32(_v, vget_high_f32(_k2), 0); // 2,4,6,8
-
-                    float32x4x2_t _out2 = vld2q_f32(outptr2);
-                    _out2.val[0] = vaddq_f32(_out2.val[0], _out20); // 0,2,4,6
-                    _out2.val[1] = vaddq_f32(_out2.val[1], _out21); // 1,3,5,7
-                    vst2q_f32(outptr2, _out2);
-
-                    _out2 = vld2q_f32(outptr2 + 2);
-                    _out2.val[0] = vaddq_f32(_out2.val[0], _out22); // 2,4,6,8
-                    vst2q_f32(outptr2 + 2, _out2);
-
-                    r0 += 4;
-                    outptr0 += 8;
-                    outptr1 += 8;
-                    outptr2 += 8;
-                }
-#endif // __ARM_NEON
-
-                for (; j < w; j++)
-                {
-                    float val = r0[0];
-
-                    outptr0[0] += val * k0[0];
-                    outptr0[1] += val * k0[1];
-                    outptr0[2] += val * k0[2];
-
-                    outptr1[0] += val * k1[0];
-                    outptr1[1] += val * k1[1];
-                    outptr1[2] += val * k1[2];
-
-                    outptr2[0] += val * k2[0];
-                    outptr2[1] += val * k2[1];
-                    outptr2[2] += val * k2[2];
-
-                    r0++;
-                    outptr0 += 2;
-                    outptr1 += 2;
-                    outptr2 += 2;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_4x4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_4x4.h
deleted file mode 100644
index 4d1c5d82..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_4x4.h
+++ /dev/null
@@ -1,334 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconv4x4s1_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outch = top_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch * 16 + q * 16;
-
-            const float* r0 = img0;
-
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 4;
-            const float* k2 = kernel0 + 8;
-            const float* k3 = kernel0 + 12;
-
-#if __ARM_NEON
-            float32x4_t _k0 = vld1q_f32(k0);
-            float32x4_t _k1 = vld1q_f32(k1);
-            float32x4_t _k2 = vld1q_f32(k2);
-            float32x4_t _k3 = vld1q_f32(k3);
-#endif // __ARM_NEON
-
-            for (int i = 0; i < h; i++)
-            {
-                float* outptr = out.row(i);
-
-                float* outptr0 = outptr;
-                float* outptr1 = outptr0 + outw;
-                float* outptr2 = outptr1 + outw;
-                float* outptr3 = outptr2 + outw;
-
-                int j = 0;
-
-#if __ARM_NEON
-                for (; j + 3 < w; j += 4)
-                {
-                    float32x4_t _v = vld1q_f32(r0);
-
-                    //
-                    float32x4_t _out00 = vld1q_f32(outptr0 + 0);
-                    _out00 = vmlaq_lane_f32(_out00, _v, vget_low_f32(_k0), 0);
-                    vst1q_f32(outptr0 + 0, _out00);
-
-                    float32x4_t _out01 = vld1q_f32(outptr0 + 1);
-                    _out01 = vmlaq_lane_f32(_out01, _v, vget_low_f32(_k0), 1);
-                    vst1q_f32(outptr0 + 1, _out01);
-
-                    float32x4_t _out02 = vld1q_f32(outptr0 + 2);
-                    _out02 = vmlaq_lane_f32(_out02, _v, vget_high_f32(_k0), 0);
-                    vst1q_f32(outptr0 + 2, _out02);
-
-                    float32x4_t _out03 = vld1q_f32(outptr0 + 3);
-                    _out03 = vmlaq_lane_f32(_out03, _v, vget_high_f32(_k0), 1);
-                    vst1q_f32(outptr0 + 3, _out03);
-
-                    //
-                    float32x4_t _out10 = vld1q_f32(outptr1 + 0);
-                    _out10 = vmlaq_lane_f32(_out10, _v, vget_low_f32(_k1), 0);
-                    vst1q_f32(outptr1 + 0, _out10);
-
-                    float32x4_t _out11 = vld1q_f32(outptr1 + 1);
-                    _out11 = vmlaq_lane_f32(_out11, _v, vget_low_f32(_k1), 1);
-                    vst1q_f32(outptr1 + 1, _out11);
-
-                    float32x4_t _out12 = vld1q_f32(outptr1 + 2);
-                    _out12 = vmlaq_lane_f32(_out12, _v, vget_high_f32(_k1), 0);
-                    vst1q_f32(outptr1 + 2, _out12);
-
-                    float32x4_t _out13 = vld1q_f32(outptr1 + 3);
-                    _out13 = vmlaq_lane_f32(_out13, _v, vget_high_f32(_k1), 1);
-                    vst1q_f32(outptr1 + 3, _out13);
-
-                    //
-                    float32x4_t _out20 = vld1q_f32(outptr2 + 0);
-                    _out20 = vmlaq_lane_f32(_out20, _v, vget_low_f32(_k2), 0);
-                    vst1q_f32(outptr2 + 0, _out20);
-
-                    float32x4_t _out21 = vld1q_f32(outptr2 + 1);
-                    _out21 = vmlaq_lane_f32(_out21, _v, vget_low_f32(_k2), 1);
-                    vst1q_f32(outptr2 + 1, _out21);
-
-                    float32x4_t _out22 = vld1q_f32(outptr2 + 2);
-                    _out22 = vmlaq_lane_f32(_out22, _v, vget_high_f32(_k2), 0);
-                    vst1q_f32(outptr2 + 2, _out22);
-
-                    float32x4_t _out23 = vld1q_f32(outptr2 + 3);
-                    _out23 = vmlaq_lane_f32(_out23, _v, vget_high_f32(_k2), 1);
-                    vst1q_f32(outptr2 + 3, _out23);
-
-                    //
-                    float32x4_t _out30 = vld1q_f32(outptr3 + 0);
-                    _out30 = vmlaq_lane_f32(_out30, _v, vget_low_f32(_k3), 0);
-                    vst1q_f32(outptr3 + 0, _out30);
-
-                    float32x4_t _out31 = vld1q_f32(outptr3 + 1);
-                    _out31 = vmlaq_lane_f32(_out31, _v, vget_low_f32(_k3), 1);
-                    vst1q_f32(outptr3 + 1, _out31);
-
-                    float32x4_t _out32 = vld1q_f32(outptr3 + 2);
-                    _out32 = vmlaq_lane_f32(_out32, _v, vget_high_f32(_k3), 0);
-                    vst1q_f32(outptr3 + 2, _out32);
-
-                    float32x4_t _out33 = vld1q_f32(outptr3 + 3);
-                    _out33 = vmlaq_lane_f32(_out33, _v, vget_high_f32(_k3), 1);
-                    vst1q_f32(outptr3 + 3, _out33);
-
-                    r0 += 4;
-                    outptr0 += 4;
-                    outptr1 += 4;
-                    outptr2 += 4;
-                    outptr3 += 4;
-                }
-
-#endif // __ARM_NEON
-
-                for (; j < w; j++)
-                {
-                    float val = r0[0];
-
-                    outptr0[0] += val * k0[0];
-                    outptr0[1] += val * k0[1];
-                    outptr0[2] += val * k0[2];
-                    outptr0[3] += val * k0[3];
-
-                    outptr1[0] += val * k1[0];
-                    outptr1[1] += val * k1[1];
-                    outptr1[2] += val * k1[2];
-                    outptr1[3] += val * k1[3];
-
-                    outptr2[0] += val * k2[0];
-                    outptr2[1] += val * k2[1];
-                    outptr2[2] += val * k2[2];
-                    outptr2[3] += val * k2[3];
-
-                    outptr3[0] += val * k3[0];
-                    outptr3[1] += val * k3[1];
-                    outptr3[2] += val * k3[2];
-                    outptr3[3] += val * k3[3];
-
-                    r0++;
-                    outptr0++;
-                    outptr1++;
-                    outptr2++;
-                    outptr3++;
-                }
-            }
-        }
-    }
-}
-
-static void deconv4x4s2_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outch = top_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch * 16 + q * 16;
-
-            const float* r0 = img0;
-
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 4;
-            const float* k2 = kernel0 + 8;
-            const float* k3 = kernel0 + 12;
-
-#if __ARM_NEON
-            float32x4_t _k0 = vld1q_f32(k0);
-            float32x4_t _k1 = vld1q_f32(k1);
-            float32x4_t _k2 = vld1q_f32(k2);
-            float32x4_t _k3 = vld1q_f32(k3);
-#endif // __ARM_NEON
-
-            for (int i = 0; i < h; i++)
-            {
-                float* outptr = out.row(i * 2);
-
-                float* outptr0 = outptr;
-                float* outptr1 = outptr0 + outw;
-                float* outptr2 = outptr1 + outw;
-                float* outptr3 = outptr2 + outw;
-
-                int j = 0;
-#if __ARM_NEON
-                for (; j + 3 < w; j += 4)
-                {
-                    float32x4_t _v = vld1q_f32(r0);
-
-                    // row 0
-                    float32x4x2_t _out0 = vld2q_f32(outptr0);
-                    // 0,2,4,6
-                    _out0.val[0] = vmlaq_lane_f32(_out0.val[0], _v, vget_low_f32(_k0), 0);
-                    // 1,3,5,7
-                    _out0.val[1] = vmlaq_lane_f32(_out0.val[1], _v, vget_low_f32(_k0), 1);
-                    vst2q_f32(outptr0, _out0);
-
-                    _out0 = vld2q_f32(outptr0 + 2);
-                    // 2,4,6,8
-                    _out0.val[0] = vmlaq_lane_f32(_out0.val[0], _v, vget_high_f32(_k0), 0);
-                    // 3,5,7,9
-                    _out0.val[1] = vmlaq_lane_f32(_out0.val[1], _v, vget_high_f32(_k0), 1);
-                    vst2q_f32(outptr0 + 2, _out0);
-
-                    // row 1
-                    float32x4x2_t _out1 = vld2q_f32(outptr1);
-                    // 0,2,4,6
-                    _out1.val[0] = vmlaq_lane_f32(_out1.val[0], _v, vget_low_f32(_k1), 0);
-                    // 1,3,5,7
-                    _out1.val[1] = vmlaq_lane_f32(_out1.val[1], _v, vget_low_f32(_k1), 1);
-                    vst2q_f32(outptr1, _out1);
-
-                    _out1 = vld2q_f32(outptr1 + 2);
-                    // 2,4,6,8
-                    _out1.val[0] = vmlaq_lane_f32(_out1.val[0], _v, vget_high_f32(_k1), 0);
-                    // 3,5,7,9
-                    _out1.val[1] = vmlaq_lane_f32(_out1.val[1], _v, vget_high_f32(_k1), 1);
-                    vst2q_f32(outptr1 + 2, _out1);
-
-                    // row 2
-                    float32x4x2_t _out2 = vld2q_f32(outptr2);
-                    _out2.val[0] = vmlaq_lane_f32(_out2.val[0], _v, vget_low_f32(_k2), 0);
-                    _out2.val[1] = vmlaq_lane_f32(_out2.val[1], _v, vget_low_f32(_k2), 1);
-                    vst2q_f32(outptr2, _out2);
-
-                    _out2 = vld2q_f32(outptr2 + 2);
-                    _out2.val[0] = vmlaq_lane_f32(_out2.val[0], _v, vget_high_f32(_k2), 0);
-                    _out2.val[1] = vmlaq_lane_f32(_out2.val[1], _v, vget_high_f32(_k2), 1);
-                    vst2q_f32(outptr2 + 2, _out2);
-
-                    // row 3
-                    float32x4x2_t _out3 = vld2q_f32(outptr3);
-                    _out3.val[0] = vmlaq_lane_f32(_out3.val[0], _v, vget_low_f32(_k3), 0);
-                    _out3.val[1] = vmlaq_lane_f32(_out3.val[1], _v, vget_low_f32(_k3), 1);
-                    vst2q_f32(outptr3, _out3);
-
-                    _out3 = vld2q_f32(outptr3 + 2);
-                    _out3.val[0] = vmlaq_lane_f32(_out3.val[0], _v, vget_high_f32(_k3), 0);
-                    _out3.val[1] = vmlaq_lane_f32(_out3.val[1], _v, vget_high_f32(_k3), 1);
-                    vst2q_f32(outptr3 + 2, _out3);
-
-                    r0 += 4;
-                    outptr0 += 8;
-                    outptr1 += 8;
-                    outptr2 += 8;
-                    outptr3 += 8;
-                }
-
-#endif // __ARM_NEON
-
-                for (; j < w; j++)
-                {
-                    float val = r0[0];
-
-                    outptr0[0] += val * k0[0];
-                    outptr0[1] += val * k0[1];
-                    outptr0[2] += val * k0[2];
-                    outptr0[3] += val * k0[3];
-
-                    outptr1[0] += val * k1[0];
-                    outptr1[1] += val * k1[1];
-                    outptr1[2] += val * k1[2];
-                    outptr1[3] += val * k1[3];
-
-                    outptr2[0] += val * k2[0];
-                    outptr2[1] += val * k2[1];
-                    outptr2[2] += val * k2[2];
-                    outptr2[3] += val * k2[3];
-
-                    outptr3[0] += val * k3[0];
-                    outptr3[1] += val * k3[1];
-                    outptr3[2] += val * k3[2];
-                    outptr3[3] += val * k3[3];
-
-                    r0++;
-                    outptr0 += 2;
-                    outptr1 += 2;
-                    outptr2 += 2;
-                    outptr3 += 2;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_4x4_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_4x4_fp16s.h
deleted file mode 100644
index 03e1d5f3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_4x4_fp16s.h
+++ /dev/null
@@ -1,159 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconv4x4s2_fp16sa_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outch = top_blob.c;
-
-    const __fp16* kernel = _kernel;
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const __fp16 bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            const __fp16* img0 = bottom_blob.channel(q);
-
-            const __fp16* kernel0 = kernel + p * inch * 16 + q * 16;
-
-            const __fp16* r0 = img0;
-
-            const __fp16* k0 = kernel0;
-            const __fp16* k1 = kernel0 + 4;
-            const __fp16* k2 = kernel0 + 8;
-            const __fp16* k3 = kernel0 + 12;
-
-            float16x4_t _k0 = vld1_f16(k0);
-            float16x4_t _k1 = vld1_f16(k1);
-            float16x4_t _k2 = vld1_f16(k2);
-            float16x4_t _k3 = vld1_f16(k3);
-
-            for (int i = 0; i < h; i++)
-            {
-                __fp16* outptr = out.row<__fp16>(i * 2);
-
-                __fp16* outptr0 = outptr;
-                __fp16* outptr1 = outptr0 + outw;
-                __fp16* outptr2 = outptr1 + outw;
-                __fp16* outptr3 = outptr2 + outw;
-
-                int j = 0;
-                for (; j + 3 < w; j += 4)
-                {
-                    float16x4_t _v = vld1_f16(r0);
-
-                    // row 0
-                    float16x4x2_t _out0 = vld2_f16(outptr0);
-                    // 0,2,4,6
-                    _out0.val[0] = vfma_lane_f16(_out0.val[0], _v, _k0, 0);
-                    // 1,3,5,7
-                    _out0.val[1] = vfma_lane_f16(_out0.val[1], _v, _k0, 1);
-                    vst2_f16(outptr0, _out0);
-
-                    _out0 = vld2_f16(outptr0 + 2);
-                    // 2,4,6,8
-                    _out0.val[0] = vfma_lane_f16(_out0.val[0], _v, _k0, 2);
-                    // 3,5,7,9
-                    _out0.val[1] = vfma_lane_f16(_out0.val[1], _v, _k0, 3);
-                    vst2_f16(outptr0 + 2, _out0);
-
-                    // row 1
-                    float16x4x2_t _out1 = vld2_f16(outptr1);
-                    // 0,2,4,6
-                    _out1.val[0] = vfma_lane_f16(_out1.val[0], _v, _k1, 0);
-                    // 1,3,5,7
-                    _out1.val[1] = vfma_lane_f16(_out1.val[1], _v, _k1, 1);
-                    vst2_f16(outptr1, _out1);
-
-                    _out1 = vld2_f16(outptr1 + 2);
-                    // 2,4,6,8
-                    _out1.val[0] = vfma_lane_f16(_out1.val[0], _v, _k1, 2);
-                    // 3,5,7,9
-                    _out1.val[1] = vfma_lane_f16(_out1.val[1], _v, _k1, 3);
-                    vst2_f16(outptr1 + 2, _out1);
-
-                    // row 2
-                    float16x4x2_t _out2 = vld2_f16(outptr2);
-                    _out2.val[0] = vfma_lane_f16(_out2.val[0], _v, _k2, 0);
-                    _out2.val[1] = vfma_lane_f16(_out2.val[1], _v, _k2, 1);
-                    vst2_f16(outptr2, _out2);
-
-                    _out2 = vld2_f16(outptr2 + 2);
-                    _out2.val[0] = vfma_lane_f16(_out2.val[0], _v, _k2, 2);
-                    _out2.val[1] = vfma_lane_f16(_out2.val[1], _v, _k2, 3);
-                    vst2_f16(outptr2 + 2, _out2);
-
-                    // row 3
-                    float16x4x2_t _out3 = vld2_f16(outptr3);
-                    _out3.val[0] = vfma_lane_f16(_out3.val[0], _v, _k3, 0);
-                    _out3.val[1] = vfma_lane_f16(_out3.val[1], _v, _k3, 1);
-                    vst2_f16(outptr3, _out3);
-
-                    _out3 = vld2_f16(outptr3 + 2);
-                    _out3.val[0] = vfma_lane_f16(_out3.val[0], _v, _k3, 2);
-                    _out3.val[1] = vfma_lane_f16(_out3.val[1], _v, _k3, 3);
-                    vst2_f16(outptr3 + 2, _out3);
-
-                    r0 += 4;
-                    outptr0 += 8;
-                    outptr1 += 8;
-                    outptr2 += 8;
-                    outptr3 += 8;
-                }
-                for (; j < w; j++)
-                {
-                    __fp16 val = r0[0];
-
-                    outptr0[0] += val * k0[0];
-                    outptr0[1] += val * k0[1];
-                    outptr0[2] += val * k0[2];
-                    outptr0[3] += val * k0[3];
-
-                    outptr1[0] += val * k1[0];
-                    outptr1[1] += val * k1[1];
-                    outptr1[2] += val * k1[2];
-                    outptr1[3] += val * k1[3];
-
-                    outptr2[0] += val * k2[0];
-                    outptr2[1] += val * k2[1];
-                    outptr2[2] += val * k2[2];
-                    outptr2[3] += val * k2[3];
-
-                    outptr3[0] += val * k3[0];
-                    outptr3[1] += val * k3[1];
-                    outptr3[2] += val * k3[2];
-                    outptr3[3] += val * k3[3];
-
-                    r0++;
-                    outptr0 += 2;
-                    outptr1 += 2;
-                    outptr2 += 2;
-                    outptr3 += 2;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_arm.h
deleted file mode 100644
index b4cdcbe0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolution_arm.h
+++ /dev/null
@@ -1,57 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTION_ARM_H
-#define LAYER_DECONVOLUTION_ARM_H
-
-#include "deconvolution.h"
-
-namespace ncnn {
-
-class Deconvolution_arm : public Deconvolution
-{
-public:
-    Deconvolution_arm();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int create_pipeline_bf16s(const Option& opt);
-    int forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Layer* activation;
-    Layer* gemm;
-
-    Mat weight_data_tm;
-
-    // fp16
-    Mat bias_data_fp16;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTION_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolutiondepthwise_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolutiondepthwise_arm.h
deleted file mode 100644
index a7ef393d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/deconvolutiondepthwise_arm.h
+++ /dev/null
@@ -1,55 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTIONDEPTHWISE_ARM_H
-#define LAYER_DECONVOLUTIONDEPTHWISE_ARM_H
-
-#include "deconvolutiondepthwise.h"
-
-namespace ncnn {
-
-class DeconvolutionDepthWise_arm : public DeconvolutionDepthWise
-{
-public:
-    DeconvolutionDepthWise_arm();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    std::vector<ncnn::Layer*> group_ops;
-
-    Mat weight_data_tm;
-
-    // fp16
-    Mat bias_data_fp16;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTIONDEPTHWISE_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/dequantize_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/dequantize_arm.h
deleted file mode 100644
index 677c731d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/dequantize_arm.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2018 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DEQUANTIZE_ARM_H
-#define LAYER_DEQUANTIZE_ARM_H
-
-#include "dequantize.h"
-
-namespace ncnn {
-
-class Dequantize_arm : public Dequantize
-{
-public:
-    Dequantize_arm();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DEQUANTIZE_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/dropout_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/dropout_arm.h
deleted file mode 100644
index 9a970525..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/dropout_arm.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DROPOUT_ARM_H
-#define LAYER_DROPOUT_ARM_H
-
-#include "dropout.h"
-
-namespace ncnn {
-
-class Dropout_arm : public Dropout
-{
-public:
-    Dropout_arm();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DROPOUT_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/eltwise_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/eltwise_arm.h
deleted file mode 100644
index 6bd91f5d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/eltwise_arm.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ELTWISE_ARM_H
-#define LAYER_ELTWISE_ARM_H
-
-#include "eltwise.h"
-
-namespace ncnn {
-
-class Eltwise_arm : public Eltwise
-{
-public:
-    Eltwise_arm();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_fp16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-    int forward_fp16sa(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_bf16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ELTWISE_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/flatten_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/flatten_arm.h
deleted file mode 100644
index 9bc9a0d1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/flatten_arm.h
+++ /dev/null
@@ -1,36 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_FLATTEN_ARM_H
-#define LAYER_FLATTEN_ARM_H
-
-#include "flatten.h"
-
-namespace ncnn {
-
-class Flatten_arm : public Flatten
-{
-public:
-    Flatten_arm();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_bf16s_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_FLATTEN_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gelu_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gelu_arm.h
deleted file mode 100644
index 5be9fc4d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gelu_arm.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GELU_ARM_H
-#define LAYER_GELU_ARM_H
-
-#include "gelu.h"
-
-namespace ncnn {
-
-class GELU_arm : public GELU
-{
-public:
-    GELU_arm();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_inplace_bf16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GELU_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_arm.h
deleted file mode 100644
index 0c1eab10..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_arm.h
+++ /dev/null
@@ -1,54 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GEMM_ARM_H
-#define LAYER_GEMM_ARM_H
-
-#include "gemm.h"
-
-namespace ncnn {
-
-class Gemm_arm : public Gemm
-{
-public:
-    Gemm_arm();
-
-    virtual int create_pipeline(const Option& opt);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if NCNN_VFPV4
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-#if NCNN_ARM82
-    int create_pipeline_fp16sa(const Option& opt);
-    int forward_fp16sa(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-#endif
-#endif
-#if NCNN_BF16
-    int create_pipeline_bf16s(const Option& opt);
-    int forward_bf16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-#endif
-
-public:
-    int nT;
-    Mat AT_data;
-    Mat BT_data;
-    Mat CT_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GEMM_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_bf16s.h
deleted file mode 100644
index 06a1b8f5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_bf16s.h
+++ /dev/null
@@ -1,3530 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pack_A_tile_fp32_to_bf16(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
-{
-    const int A_hstep = A.dims == 3 ? (int)A.cstep : A.w;
-
-    unsigned short* pp = AT;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        const float* p0 = (const float*)A + (i + ii) * A_hstep + k;
-        const float* p1 = (const float*)A + (i + ii + 1) * A_hstep + k;
-        const float* p2 = (const float*)A + (i + ii + 2) * A_hstep + k;
-        const float* p3 = (const float*)A + (i + ii + 3) * A_hstep + k;
-        const float* p4 = (const float*)A + (i + ii + 4) * A_hstep + k;
-        const float* p5 = (const float*)A + (i + ii + 5) * A_hstep + k;
-        const float* p6 = (const float*)A + (i + ii + 6) * A_hstep + k;
-        const float* p7 = (const float*)A + (i + ii + 7) * A_hstep + k;
-
-        int kk = 0;
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8_t _r0 = vcombine_u16(float2bfloat(vld1q_f32(p0)), float2bfloat(vld1q_f32(p0 + 4)));
-            uint16x8_t _r1 = vcombine_u16(float2bfloat(vld1q_f32(p1)), float2bfloat(vld1q_f32(p1 + 4)));
-            uint16x8_t _r2 = vcombine_u16(float2bfloat(vld1q_f32(p2)), float2bfloat(vld1q_f32(p2 + 4)));
-            uint16x8_t _r3 = vcombine_u16(float2bfloat(vld1q_f32(p3)), float2bfloat(vld1q_f32(p3 + 4)));
-            uint16x8_t _r4 = vcombine_u16(float2bfloat(vld1q_f32(p4)), float2bfloat(vld1q_f32(p4 + 4)));
-            uint16x8_t _r5 = vcombine_u16(float2bfloat(vld1q_f32(p5)), float2bfloat(vld1q_f32(p5 + 4)));
-            uint16x8_t _r6 = vcombine_u16(float2bfloat(vld1q_f32(p6)), float2bfloat(vld1q_f32(p6 + 4)));
-            uint16x8_t _r7 = vcombine_u16(float2bfloat(vld1q_f32(p7)), float2bfloat(vld1q_f32(p7 + 4)));
-            transpose8x8_u16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-            vst1q_u16(pp, _r0);
-            vst1q_u16(pp + 8, _r1);
-            vst1q_u16(pp + 8 * 2, _r2);
-            vst1q_u16(pp + 8 * 3, _r3);
-            vst1q_u16(pp + 8 * 4, _r4);
-            vst1q_u16(pp + 8 * 5, _r5);
-            vst1q_u16(pp + 8 * 6, _r6);
-            vst1q_u16(pp + 8 * 7, _r7);
-            pp += 64;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-            p4 += 8;
-            p5 += 8;
-            p6 += 8;
-            p7 += 8;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_bfloat16(p0[0]);
-            pp[1] = float32_to_bfloat16(p1[0]);
-            pp[2] = float32_to_bfloat16(p2[0]);
-            pp[3] = float32_to_bfloat16(p3[0]);
-            pp[4] = float32_to_bfloat16(p4[0]);
-            pp[5] = float32_to_bfloat16(p5[0]);
-            pp[6] = float32_to_bfloat16(p6[0]);
-            pp[7] = float32_to_bfloat16(p7[0]);
-            pp += 8;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-            p4++;
-            p5++;
-            p6++;
-            p7++;
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        const float* p0 = (const float*)A + (i + ii) * A_hstep + k;
-        const float* p1 = (const float*)A + (i + ii + 1) * A_hstep + k;
-        const float* p2 = (const float*)A + (i + ii + 2) * A_hstep + k;
-        const float* p3 = (const float*)A + (i + ii + 3) * A_hstep + k;
-
-        int kk = 0;
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8x4_t _r0123;
-            _r0123.val[0] = vcombine_u16(float2bfloat(vld1q_f32(p0)), float2bfloat(vld1q_f32(p0 + 4)));
-            _r0123.val[1] = vcombine_u16(float2bfloat(vld1q_f32(p1)), float2bfloat(vld1q_f32(p1 + 4)));
-            _r0123.val[2] = vcombine_u16(float2bfloat(vld1q_f32(p2)), float2bfloat(vld1q_f32(p2 + 4)));
-            _r0123.val[3] = vcombine_u16(float2bfloat(vld1q_f32(p3)), float2bfloat(vld1q_f32(p3 + 4)));
-            vst4q_u16(pp, _r0123);
-            pp += 32;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4x4_t _r0123;
-            _r0123.val[0] = float2bfloat(vld1q_f32(p0));
-            _r0123.val[1] = float2bfloat(vld1q_f32(p1));
-            _r0123.val[2] = float2bfloat(vld1q_f32(p2));
-            _r0123.val[3] = float2bfloat(vld1q_f32(p3));
-            vst4_u16(pp, _r0123);
-            pp += 16;
-            p0 += 4;
-            p1 += 4;
-            p2 += 4;
-            p3 += 4;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_bfloat16(p0[0]);
-            pp[1] = float32_to_bfloat16(p1[0]);
-            pp[2] = float32_to_bfloat16(p2[0]);
-            pp[3] = float32_to_bfloat16(p3[0]);
-            pp += 4;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        const float* p0 = (const float*)A + (i + ii) * A_hstep + k;
-        const float* p1 = (const float*)A + (i + ii + 1) * A_hstep + k;
-
-        int kk = 0;
-#if __ARM_NEON
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8x2_t _r01;
-            _r01.val[0] = vcombine_u16(float2bfloat(vld1q_f32(p0)), float2bfloat(vld1q_f32(p0 + 4)));
-            _r01.val[1] = vcombine_u16(float2bfloat(vld1q_f32(p1)), float2bfloat(vld1q_f32(p1 + 4)));
-            vst2q_u16(pp, _r01);
-            pp += 16;
-            p0 += 8;
-            p1 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4x2_t _r01;
-            _r01.val[0] = float2bfloat(vld1q_f32(p0));
-            _r01.val[1] = float2bfloat(vld1q_f32(p1));
-            vst2_u16(pp, _r01);
-            pp += 8;
-            p0 += 4;
-            p1 += 4;
-        }
-#endif // __ARM_NEON
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_bfloat16(p0[0]);
-            pp[1] = float32_to_bfloat16(p1[0]);
-            pp += 2;
-            p0++;
-            p1++;
-        }
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        const float* p0 = (const float*)A + (i + ii) * A_hstep + k;
-
-        int kk = 0;
-#if __ARM_NEON
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8_t _r0 = vcombine_u16(float2bfloat(vld1q_f32(p0)), float2bfloat(vld1q_f32(p0 + 4)));
-            vst1q_u16(pp, _r0);
-            pp += 8;
-            p0 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4_t _r0 = float2bfloat(vld1q_f32(p0));
-            vst1_u16(pp, _r0);
-            pp += 4;
-            p0 += 4;
-        }
-#endif // __ARM_NEON
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_bfloat16(p0[0]);
-            pp += 1;
-            p0++;
-        }
-    }
-}
-
-static void transpose_pack_A_tile_fp32_to_bf16(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
-{
-    const int A_hstep = A.dims == 3 ? (int)A.cstep : A.w;
-
-    unsigned short* pp = AT;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        const float* p0 = (const float*)A + k * A_hstep + (i + ii);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            uint16x8_t _r0 = vcombine_u16(float2bfloat(vld1q_f32(p0)), float2bfloat(vld1q_f32(p0 + 4)));
-            vst1q_u16(pp, _r0);
-            pp += 8;
-            p0 += A_hstep;
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        const float* p0 = (const float*)A + k * A_hstep + (i + ii);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            uint16x4_t _r0 = float2bfloat(vld1q_f32(p0));
-            vst1_u16(pp, _r0);
-            pp += 4;
-            p0 += A_hstep;
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        const float* p0 = (const float*)A + k * A_hstep + (i + ii);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_bfloat16(p0[0]);
-            pp[1] = float32_to_bfloat16(p0[1]);
-            pp += 2;
-            p0 += A_hstep;
-        }
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        const float* p0 = (const float*)A + k * A_hstep + (i + ii);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_bfloat16(p0[0]);
-            pp += 1;
-            p0 += A_hstep;
-        }
-    }
-}
-
-static void pack_B_tile_fp32_to_bf16(const Mat& B, Mat& BT, int j, int max_jj, int k, int max_kk)
-{
-    const int B_hstep = B.dims == 3 ? (int)B.cstep : B.w;
-
-    unsigned short* pp = BT;
-
-    int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; jj + 11 < max_jj; jj += 12)
-    {
-        const float* p0 = (const float*)B + (j + jj) * B_hstep + k;
-        const float* p1 = (const float*)B + (j + jj + 1) * B_hstep + k;
-        const float* p2 = (const float*)B + (j + jj + 2) * B_hstep + k;
-        const float* p3 = (const float*)B + (j + jj + 3) * B_hstep + k;
-        const float* p4 = (const float*)B + (j + jj + 4) * B_hstep + k;
-        const float* p5 = (const float*)B + (j + jj + 5) * B_hstep + k;
-        const float* p6 = (const float*)B + (j + jj + 6) * B_hstep + k;
-        const float* p7 = (const float*)B + (j + jj + 7) * B_hstep + k;
-        const float* p8 = (const float*)B + (j + jj + 8) * B_hstep + k;
-        const float* p9 = (const float*)B + (j + jj + 9) * B_hstep + k;
-        const float* pa = (const float*)B + (j + jj + 10) * B_hstep + k;
-        const float* pb = (const float*)B + (j + jj + 11) * B_hstep + k;
-
-        int kk = 0;
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4_t _r0 = float2bfloat(vld1q_f32(p0));
-            uint16x4_t _r1 = float2bfloat(vld1q_f32(p1));
-            uint16x4_t _r2 = float2bfloat(vld1q_f32(p2));
-            uint16x4_t _r3 = float2bfloat(vld1q_f32(p3));
-            uint16x4_t _r4 = float2bfloat(vld1q_f32(p4));
-            uint16x4_t _r5 = float2bfloat(vld1q_f32(p5));
-            uint16x4_t _r6 = float2bfloat(vld1q_f32(p6));
-            uint16x4_t _r7 = float2bfloat(vld1q_f32(p7));
-            uint16x4_t _r8 = float2bfloat(vld1q_f32(p8));
-            uint16x4_t _r9 = float2bfloat(vld1q_f32(p9));
-            uint16x4_t _ra = float2bfloat(vld1q_f32(pa));
-            uint16x4_t _rb = float2bfloat(vld1q_f32(pb));
-
-            transpose4x4_u16(_r0, _r1, _r2, _r3);
-            transpose4x4_u16(_r4, _r5, _r6, _r7);
-            transpose4x4_u16(_r8, _r9, _ra, _rb);
-
-            vst1_u16(pp, _r0);
-            vst1_u16(pp + 4, _r4);
-            vst1_u16(pp + 4 * 2, _r8);
-            vst1_u16(pp + 4 * 3, _r1);
-            vst1_u16(pp + 4 * 4, _r5);
-            vst1_u16(pp + 4 * 5, _r9);
-            vst1_u16(pp + 4 * 6, _r2);
-            vst1_u16(pp + 4 * 7, _r6);
-            vst1_u16(pp + 4 * 8, _ra);
-            vst1_u16(pp + 4 * 9, _r3);
-            vst1_u16(pp + 4 * 10, _r7);
-            vst1_u16(pp + 4 * 11, _rb);
-            pp += 48;
-            p0 += 4;
-            p1 += 4;
-            p2 += 4;
-            p3 += 4;
-            p4 += 4;
-            p5 += 4;
-            p6 += 4;
-            p7 += 4;
-            p8 += 4;
-            p9 += 4;
-            pa += 4;
-            pb += 4;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_bfloat16(p0[0]);
-            pp[1] = float32_to_bfloat16(p1[0]);
-            pp[2] = float32_to_bfloat16(p2[0]);
-            pp[3] = float32_to_bfloat16(p3[0]);
-            pp[4] = float32_to_bfloat16(p4[0]);
-            pp[5] = float32_to_bfloat16(p5[0]);
-            pp[6] = float32_to_bfloat16(p6[0]);
-            pp[7] = float32_to_bfloat16(p7[0]);
-            pp[8] = float32_to_bfloat16(p8[0]);
-            pp[9] = float32_to_bfloat16(p9[0]);
-            pp[10] = float32_to_bfloat16(pa[0]);
-            pp[11] = float32_to_bfloat16(pb[0]);
-            pp += 12;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-            p4++;
-            p5++;
-            p6++;
-            p7++;
-            p8++;
-            p9++;
-            pa++;
-            pb++;
-        }
-    }
-#endif // __aarch64__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-        const float* p0 = (const float*)B + (j + jj) * B_hstep + k;
-        const float* p1 = (const float*)B + (j + jj + 1) * B_hstep + k;
-        const float* p2 = (const float*)B + (j + jj + 2) * B_hstep + k;
-        const float* p3 = (const float*)B + (j + jj + 3) * B_hstep + k;
-        const float* p4 = (const float*)B + (j + jj + 4) * B_hstep + k;
-        const float* p5 = (const float*)B + (j + jj + 5) * B_hstep + k;
-        const float* p6 = (const float*)B + (j + jj + 6) * B_hstep + k;
-        const float* p7 = (const float*)B + (j + jj + 7) * B_hstep + k;
-
-        int kk = 0;
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8_t _r0 = vcombine_u16(float2bfloat(vld1q_f32(p0)), float2bfloat(vld1q_f32(p0 + 4)));
-            uint16x8_t _r1 = vcombine_u16(float2bfloat(vld1q_f32(p1)), float2bfloat(vld1q_f32(p1 + 4)));
-            uint16x8_t _r2 = vcombine_u16(float2bfloat(vld1q_f32(p2)), float2bfloat(vld1q_f32(p2 + 4)));
-            uint16x8_t _r3 = vcombine_u16(float2bfloat(vld1q_f32(p3)), float2bfloat(vld1q_f32(p3 + 4)));
-            uint16x8_t _r4 = vcombine_u16(float2bfloat(vld1q_f32(p4)), float2bfloat(vld1q_f32(p4 + 4)));
-            uint16x8_t _r5 = vcombine_u16(float2bfloat(vld1q_f32(p5)), float2bfloat(vld1q_f32(p5 + 4)));
-            uint16x8_t _r6 = vcombine_u16(float2bfloat(vld1q_f32(p6)), float2bfloat(vld1q_f32(p6 + 4)));
-            uint16x8_t _r7 = vcombine_u16(float2bfloat(vld1q_f32(p7)), float2bfloat(vld1q_f32(p7 + 4)));
-            transpose8x8_u16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-            vst1q_u16(pp, _r0);
-            vst1q_u16(pp + 8, _r1);
-            vst1q_u16(pp + 8 * 2, _r2);
-            vst1q_u16(pp + 8 * 3, _r3);
-            vst1q_u16(pp + 8 * 4, _r4);
-            vst1q_u16(pp + 8 * 5, _r5);
-            vst1q_u16(pp + 8 * 6, _r6);
-            vst1q_u16(pp + 8 * 7, _r7);
-            pp += 64;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-            p4 += 8;
-            p5 += 8;
-            p6 += 8;
-            p7 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4_t _r0 = float2bfloat(vld1q_f32(p0));
-            uint16x4_t _r1 = float2bfloat(vld1q_f32(p1));
-            uint16x4_t _r2 = float2bfloat(vld1q_f32(p2));
-            uint16x4_t _r3 = float2bfloat(vld1q_f32(p3));
-            uint16x4_t _r4 = float2bfloat(vld1q_f32(p4));
-            uint16x4_t _r5 = float2bfloat(vld1q_f32(p5));
-            uint16x4_t _r6 = float2bfloat(vld1q_f32(p6));
-            uint16x4_t _r7 = float2bfloat(vld1q_f32(p7));
-
-            transpose4x4_u16(_r0, _r1, _r2, _r3);
-            transpose4x4_u16(_r4, _r5, _r6, _r7);
-
-            vst1_u16(pp, _r0);
-            vst1_u16(pp + 4, _r4);
-            vst1_u16(pp + 4 * 2, _r1);
-            vst1_u16(pp + 4 * 3, _r5);
-            vst1_u16(pp + 4 * 4, _r2);
-            vst1_u16(pp + 4 * 5, _r6);
-            vst1_u16(pp + 4 * 6, _r3);
-            vst1_u16(pp + 4 * 7, _r7);
-            pp += 32;
-            p0 += 4;
-            p1 += 4;
-            p2 += 4;
-            p3 += 4;
-            p4 += 4;
-            p5 += 4;
-            p6 += 4;
-            p7 += 4;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_bfloat16(p0[0]);
-            pp[1] = float32_to_bfloat16(p1[0]);
-            pp[2] = float32_to_bfloat16(p2[0]);
-            pp[3] = float32_to_bfloat16(p3[0]);
-            pp[4] = float32_to_bfloat16(p4[0]);
-            pp[5] = float32_to_bfloat16(p5[0]);
-            pp[6] = float32_to_bfloat16(p6[0]);
-            pp[7] = float32_to_bfloat16(p7[0]);
-            pp += 8;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-            p4++;
-            p5++;
-            p6++;
-            p7++;
-        }
-    }
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-        const float* p0 = (const float*)B + (j + jj) * B_hstep + k;
-        const float* p1 = (const float*)B + (j + jj + 1) * B_hstep + k;
-        const float* p2 = (const float*)B + (j + jj + 2) * B_hstep + k;
-        const float* p3 = (const float*)B + (j + jj + 3) * B_hstep + k;
-
-        int kk = 0;
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8x4_t _r0123;
-            _r0123.val[0] = vcombine_u16(float2bfloat(vld1q_f32(p0)), float2bfloat(vld1q_f32(p0 + 4)));
-            _r0123.val[1] = vcombine_u16(float2bfloat(vld1q_f32(p1)), float2bfloat(vld1q_f32(p1 + 4)));
-            _r0123.val[2] = vcombine_u16(float2bfloat(vld1q_f32(p2)), float2bfloat(vld1q_f32(p2 + 4)));
-            _r0123.val[3] = vcombine_u16(float2bfloat(vld1q_f32(p3)), float2bfloat(vld1q_f32(p3 + 4)));
-            vst4q_u16(pp, _r0123);
-            pp += 32;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4x4_t _r0123;
-            _r0123.val[0] = float2bfloat(vld1q_f32(p0));
-            _r0123.val[1] = float2bfloat(vld1q_f32(p1));
-            _r0123.val[2] = float2bfloat(vld1q_f32(p2));
-            _r0123.val[3] = float2bfloat(vld1q_f32(p3));
-            vst4_u16(pp, _r0123);
-            pp += 16;
-            p0 += 4;
-            p1 += 4;
-            p2 += 4;
-            p3 += 4;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_bfloat16(p0[0]);
-            pp[1] = float32_to_bfloat16(p1[0]);
-            pp[2] = float32_to_bfloat16(p2[0]);
-            pp[3] = float32_to_bfloat16(p3[0]);
-            pp += 4;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-        }
-    }
-#endif // __ARM_NEON
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-        const float* p0 = (const float*)B + (j + jj) * B_hstep + k;
-        const float* p1 = (const float*)B + (j + jj + 1) * B_hstep + k;
-
-        int kk = 0;
-#if __ARM_NEON
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8x2_t _r01;
-            _r01.val[0] = vcombine_u16(float2bfloat(vld1q_f32(p0)), float2bfloat(vld1q_f32(p0 + 4)));
-            _r01.val[1] = vcombine_u16(float2bfloat(vld1q_f32(p1)), float2bfloat(vld1q_f32(p1 + 4)));
-            vst2q_u16(pp, _r01);
-            pp += 16;
-            p0 += 8;
-            p1 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4x2_t _r01;
-            _r01.val[0] = float2bfloat(vld1q_f32(p0));
-            _r01.val[1] = float2bfloat(vld1q_f32(p1));
-            vst2_u16(pp, _r01);
-            pp += 8;
-            p0 += 4;
-            p1 += 4;
-        }
-#endif // __ARM_NEON
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_bfloat16(p0[0]);
-            pp[1] = float32_to_bfloat16(p1[0]);
-            pp += 2;
-            p0++;
-            p1++;
-        }
-    }
-    for (; jj < max_jj; jj += 1)
-    {
-        const float* p0 = (const float*)B + (j + jj) * B_hstep + k;
-
-        int kk = 0;
-#if __ARM_NEON
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8_t _r0 = vcombine_u16(float2bfloat(vld1q_f32(p0)), float2bfloat(vld1q_f32(p0 + 4)));
-            vst1q_u16(pp, _r0);
-            pp += 8;
-            p0 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4_t _r0 = float2bfloat(vld1q_f32(p0));
-            vst1_u16(pp, _r0);
-            pp += 4;
-            p0 += 4;
-        }
-#endif // __ARM_NEON
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_bfloat16(p0[0]);
-            pp += 1;
-            p0++;
-        }
-    }
-}
-
-static void transpose_pack_B_tile_fp32_to_bf16(const Mat& B, Mat& BT, int j, int max_jj, int k, int max_kk)
-{
-    const int B_hstep = B.dims == 3 ? (int)B.cstep : B.w;
-
-    unsigned short* pp = BT;
-
-    int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; jj + 11 < max_jj; jj += 12)
-    {
-        const float* p0 = (const float*)B + k * B_hstep + (j + jj);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            vst1_u16(pp, float2bfloat(vld1q_f32(p0)));
-            vst1_u16(pp + 4, float2bfloat(vld1q_f32(p0 + 4)));
-            vst1_u16(pp + 8, float2bfloat(vld1q_f32(p0 + 8)));
-            pp += 12;
-            p0 += B_hstep;
-        }
-    }
-#endif // __aarch64__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-        const float* p0 = (const float*)B + k * B_hstep + (j + jj);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            uint16x8_t _r0 = vcombine_u16(float2bfloat(vld1q_f32(p0)), float2bfloat(vld1q_f32(p0 + 4)));
-            vst1q_u16(pp, _r0);
-            pp += 8;
-            p0 += B_hstep;
-        }
-    }
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-        const float* p0 = (const float*)B + k * B_hstep + (j + jj);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            uint16x4_t _r0 = float2bfloat(vld1q_f32(p0));
-            vst1_u16(pp, _r0);
-            pp += 4;
-            p0 += B_hstep;
-        }
-    }
-#endif // __ARM_NEON
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-        const float* p0 = (const float*)B + k * B_hstep + (j + jj);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_bfloat16(p0[0]);
-            pp[1] = float32_to_bfloat16(p0[1]);
-            pp += 2;
-            p0 += B_hstep;
-        }
-    }
-    for (; jj < max_jj; jj += 1)
-    {
-        const float* p0 = (const float*)B + k * B_hstep + (j + jj);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_bfloat16(p0[0]);
-            pp += 1;
-            p0 += B_hstep;
-        }
-    }
-}
-
-static void transpose_unpack_output_tile_fp32_to_bf16(const Mat& topT, Mat& top_blob, int i, int max_ii, int j, int max_jj)
-{
-    const int out_elempack = top_blob.elempack;
-    const int out_hstep = top_blob.dims == 3 ? (int)top_blob.cstep : top_blob.w;
-
-    const float* pp = topT;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        if (out_elempack == 4)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii) * 4;
-
-            for (int jj = 0; jj + 3 < max_jj; jj += 4)
-            {
-                uint16x8x4_t _r0;
-                _r0.val[0] = vcombine_u16(float2bfloat(vld1q_f32(pp)), float2bfloat(vld1q_f32(pp + 4)));
-                _r0.val[1] = vcombine_u16(float2bfloat(vld1q_f32(pp + 8)), float2bfloat(vld1q_f32(pp + 12)));
-                _r0.val[2] = vcombine_u16(float2bfloat(vld1q_f32(pp + 16)), float2bfloat(vld1q_f32(pp + 20)));
-                _r0.val[3] = vcombine_u16(float2bfloat(vld1q_f32(pp + 24)), float2bfloat(vld1q_f32(pp + 28)));
-                vst4q_u16(p0, _r0);
-                pp += 32;
-                p0 += out_hstep * 4;
-            }
-        }
-        if (out_elempack == 1)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii);
-
-            for (int jj = 0; jj < max_jj; jj += 1)
-            {
-                uint16x8_t _r0 = vcombine_u16(float2bfloat(vld1q_f32(pp)), float2bfloat(vld1q_f32(pp + 4)));
-                vst1q_u16(p0, _r0);
-                pp += 8;
-                p0 += out_hstep;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        if (out_elempack == 4)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii) * 4;
-
-            for (int jj = 0; jj + 3 < max_jj; jj += 4)
-            {
-                uint16x4x4_t _r0123;
-                _r0123.val[0] = float2bfloat(vld1q_f32(pp));
-                _r0123.val[1] = float2bfloat(vld1q_f32(pp + 4));
-                _r0123.val[2] = float2bfloat(vld1q_f32(pp + 8));
-                _r0123.val[3] = float2bfloat(vld1q_f32(pp + 12));
-                vst4_u16(p0, _r0123);
-                pp += 16;
-                p0 += out_hstep * 4;
-            }
-        }
-        if (out_elempack == 1)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii);
-
-            for (int jj = 0; jj < max_jj; jj += 1)
-            {
-                uint16x4_t _r0 = float2bfloat(vld1q_f32(pp));
-                vst1_u16(p0, _r0);
-                pp += 4;
-                p0 += out_hstep;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-#if __ARM_NEON
-        if (out_elempack == 4)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii) * 4;
-
-            for (int jj = 0; jj + 3 < max_jj; jj += 4)
-            {
-                p0[0] = float32_to_bfloat16(pp[0]);
-                p0[1] = float32_to_bfloat16(pp[2]);
-                p0[2] = float32_to_bfloat16(pp[4]);
-                p0[3] = float32_to_bfloat16(pp[6]);
-                p0[4] = float32_to_bfloat16(pp[1]);
-                p0[5] = float32_to_bfloat16(pp[3]);
-                p0[6] = float32_to_bfloat16(pp[5]);
-                p0[7] = float32_to_bfloat16(pp[7]);
-                pp += 8;
-                p0 += out_hstep * 4;
-            }
-        }
-#endif // __ARM_NEON
-        if (out_elempack == 1)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii);
-
-            for (int jj = 0; jj < max_jj; jj += 1)
-            {
-                p0[0] = float32_to_bfloat16(pp[0]);
-                p0[1] = float32_to_bfloat16(pp[1]);
-                pp += 2;
-                p0 += out_hstep;
-            }
-        }
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-#if __ARM_NEON
-        if (out_elempack == 4)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii) * 4;
-
-            for (int jj = 0; jj + 3 < max_jj; jj += 4)
-            {
-                uint16x4_t _r0 = float2bfloat(vld1q_f32(pp));
-                vst1_u16(p0, _r0);
-                pp += 4;
-                p0 += out_hstep * 4;
-            }
-        }
-#endif // __ARM_NEON
-        if (out_elempack == 1)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii);
-
-            for (int jj = 0; jj < max_jj; jj += 1)
-            {
-                p0[0] = float32_to_bfloat16(pp[0]);
-                pp += 1;
-                p0 += out_hstep;
-            }
-        }
-    }
-}
-
-static void gemm_transB_packed_tile_bf16s(const Mat& AT_tile, const Mat& BT_tile, const Mat& CT_tile, Mat& topT_tile, Mat& top_blob, int broadcast_type_C, float alpha, int i, int max_ii, int j, int max_jj, int k, int max_kk, bool k_end)
-{
-    const int out_elempack = top_blob.elempack;
-    const int out_hstep = top_blob.dims == 3 ? (int)top_blob.cstep : top_blob.w;
-
-    const unsigned short* pAT = AT_tile;
-    const unsigned short* pBT = BT_tile;
-    const float* pC = CT_tile;
-
-    float* outptr = topT_tile;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        unsigned short* outptr0 = (unsigned short*)top_blob + (i + ii) * out_hstep + j * out_elempack;
-
-        const unsigned short* pB = pBT;
-
-        if (pC)
-        {
-            if (broadcast_type_C == 1 || broadcast_type_C == 2)
-            {
-                pC = (const float*)CT_tile + i + ii;
-            }
-            if (broadcast_type_C == 4)
-            {
-                pC = (const float*)CT_tile + j;
-            }
-        }
-
-        int jj = 0;
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum20;
-            float32x4_t _sum21;
-            float32x4_t _sum30;
-            float32x4_t _sum31;
-            float32x4_t _sum40;
-            float32x4_t _sum41;
-            float32x4_t _sum50;
-            float32x4_t _sum51;
-            float32x4_t _sum60;
-            float32x4_t _sum61;
-            float32x4_t _sum70;
-            float32x4_t _sum71;
-            float32x4_t _sum80;
-            float32x4_t _sum81;
-            float32x4_t _sum90;
-            float32x4_t _sum91;
-            float32x4_t _suma0;
-            float32x4_t _suma1;
-            float32x4_t _sumb0;
-            float32x4_t _sumb1;
-
-            if (k == 0)
-            {
-                _sum00 = vdupq_n_f32(0.f);
-                _sum01 = vdupq_n_f32(0.f);
-                _sum10 = vdupq_n_f32(0.f);
-                _sum11 = vdupq_n_f32(0.f);
-                _sum20 = vdupq_n_f32(0.f);
-                _sum21 = vdupq_n_f32(0.f);
-                _sum30 = vdupq_n_f32(0.f);
-                _sum31 = vdupq_n_f32(0.f);
-                _sum40 = vdupq_n_f32(0.f);
-                _sum41 = vdupq_n_f32(0.f);
-                _sum50 = vdupq_n_f32(0.f);
-                _sum51 = vdupq_n_f32(0.f);
-                _sum60 = vdupq_n_f32(0.f);
-                _sum61 = vdupq_n_f32(0.f);
-                _sum70 = vdupq_n_f32(0.f);
-                _sum71 = vdupq_n_f32(0.f);
-                _sum80 = vdupq_n_f32(0.f);
-                _sum81 = vdupq_n_f32(0.f);
-                _sum90 = vdupq_n_f32(0.f);
-                _sum91 = vdupq_n_f32(0.f);
-                _suma0 = vdupq_n_f32(0.f);
-                _suma1 = vdupq_n_f32(0.f);
-                _sumb0 = vdupq_n_f32(0.f);
-                _sumb1 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum10 = _sum00;
-                        _sum11 = _sum00;
-                        _sum20 = _sum00;
-                        _sum21 = _sum00;
-                        _sum30 = _sum00;
-                        _sum31 = _sum00;
-                        _sum40 = _sum00;
-                        _sum41 = _sum00;
-                        _sum50 = _sum00;
-                        _sum51 = _sum00;
-                        _sum60 = _sum00;
-                        _sum61 = _sum00;
-                        _sum70 = _sum00;
-                        _sum71 = _sum00;
-                        _sum80 = _sum00;
-                        _sum81 = _sum00;
-                        _sum90 = _sum00;
-                        _sum91 = _sum00;
-                        _suma0 = _sum00;
-                        _suma1 = _sum00;
-                        _sumb0 = _sum00;
-                        _sumb1 = _sum00;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                        _sum10 = _sum00;
-                        _sum11 = _sum01;
-                        _sum20 = _sum00;
-                        _sum21 = _sum01;
-                        _sum30 = _sum00;
-                        _sum31 = _sum01;
-                        _sum40 = _sum00;
-                        _sum41 = _sum01;
-                        _sum50 = _sum00;
-                        _sum51 = _sum01;
-                        _sum60 = _sum00;
-                        _sum61 = _sum01;
-                        _sum70 = _sum00;
-                        _sum71 = _sum01;
-                        _sum80 = _sum00;
-                        _sum81 = _sum01;
-                        _sum90 = _sum00;
-                        _sum91 = _sum01;
-                        _suma0 = _sum00;
-                        _suma1 = _sum01;
-                        _sumb0 = _sum00;
-                        _sumb1 = _sum01;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4 * 1);
-                        _sum10 = vld1q_f32(pC + 4 * 2);
-                        _sum11 = vld1q_f32(pC + 4 * 3);
-                        _sum20 = vld1q_f32(pC + 4 * 4);
-                        _sum21 = vld1q_f32(pC + 4 * 5);
-                        _sum30 = vld1q_f32(pC + 4 * 6);
-                        _sum31 = vld1q_f32(pC + 4 * 7);
-                        _sum40 = vld1q_f32(pC + 4 * 8);
-                        _sum41 = vld1q_f32(pC + 4 * 9);
-                        _sum50 = vld1q_f32(pC + 4 * 10);
-                        _sum51 = vld1q_f32(pC + 4 * 11);
-                        _sum60 = vld1q_f32(pC + 4 * 12);
-                        _sum61 = vld1q_f32(pC + 4 * 13);
-                        _sum70 = vld1q_f32(pC + 4 * 14);
-                        _sum71 = vld1q_f32(pC + 4 * 15);
-                        _sum80 = vld1q_f32(pC + 4 * 16);
-                        _sum81 = vld1q_f32(pC + 4 * 17);
-                        _sum90 = vld1q_f32(pC + 4 * 18);
-                        _sum91 = vld1q_f32(pC + 4 * 19);
-                        _suma0 = vld1q_f32(pC + 4 * 20);
-                        _suma1 = vld1q_f32(pC + 4 * 21);
-                        _sumb0 = vld1q_f32(pC + 4 * 22);
-                        _sumb1 = vld1q_f32(pC + 4 * 23);
-                        pC += 96;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum10 = vdupq_n_f32(pC[1]);
-                        _sum20 = vdupq_n_f32(pC[2]);
-                        _sum30 = vdupq_n_f32(pC[3]);
-                        _sum40 = vdupq_n_f32(pC[4]);
-                        _sum50 = vdupq_n_f32(pC[5]);
-                        _sum60 = vdupq_n_f32(pC[6]);
-                        _sum70 = vdupq_n_f32(pC[7]);
-                        _sum80 = vdupq_n_f32(pC[8]);
-                        _sum90 = vdupq_n_f32(pC[9]);
-                        _suma0 = vdupq_n_f32(pC[10]);
-                        _sumb0 = vdupq_n_f32(pC[11]);
-                        _sum01 = _sum00;
-                        _sum11 = _sum10;
-                        _sum21 = _sum20;
-                        _sum31 = _sum30;
-                        _sum41 = _sum40;
-                        _sum51 = _sum50;
-                        _sum61 = _sum60;
-                        _sum71 = _sum70;
-                        _sum81 = _sum80;
-                        _sum91 = _sum90;
-                        _suma1 = _suma0;
-                        _sumb1 = _sumb0;
-                        pC += 12;
-                    }
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-                _sum20 = vld1q_f32(outptr + 4 * 4);
-                _sum21 = vld1q_f32(outptr + 4 * 5);
-                _sum30 = vld1q_f32(outptr + 4 * 6);
-                _sum31 = vld1q_f32(outptr + 4 * 7);
-                _sum40 = vld1q_f32(outptr + 4 * 8);
-                _sum41 = vld1q_f32(outptr + 4 * 9);
-                _sum50 = vld1q_f32(outptr + 4 * 10);
-                _sum51 = vld1q_f32(outptr + 4 * 11);
-                _sum60 = vld1q_f32(outptr + 4 * 12);
-                _sum61 = vld1q_f32(outptr + 4 * 13);
-                _sum70 = vld1q_f32(outptr + 4 * 14);
-                _sum71 = vld1q_f32(outptr + 4 * 15);
-                _sum80 = vld1q_f32(outptr + 4 * 16);
-                _sum81 = vld1q_f32(outptr + 4 * 17);
-                _sum90 = vld1q_f32(outptr + 4 * 18);
-                _sum91 = vld1q_f32(outptr + 4 * 19);
-                _suma0 = vld1q_f32(outptr + 4 * 20);
-                _suma1 = vld1q_f32(outptr + 4 * 21);
-                _sumb0 = vld1q_f32(outptr + 4 * 22);
-                _sumb1 = vld1q_f32(outptr + 4 * 23);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                uint16x8_t _pA = vld1q_u16(pA);
-                float32x4_t _pA0 = bfloat2float(vget_low_u16(_pA));
-                float32x4_t _pA1 = bfloat2float(vget_high_u16(_pA));
-
-                float32x4_t _pB0 = bfloat2float(vld1_u16(pB));
-                float32x4_t _pB1 = bfloat2float(vld1_u16(pB + 4));
-                float32x4_t _pB2 = bfloat2float(vld1_u16(pB + 8));
-
-                _sum00 = vfmaq_laneq_f32(_sum00, _pA0, _pB0, 0);
-                _sum01 = vfmaq_laneq_f32(_sum01, _pA1, _pB0, 0);
-                _sum10 = vfmaq_laneq_f32(_sum10, _pA0, _pB0, 1);
-                _sum11 = vfmaq_laneq_f32(_sum11, _pA1, _pB0, 1);
-                _sum20 = vfmaq_laneq_f32(_sum20, _pA0, _pB0, 2);
-                _sum21 = vfmaq_laneq_f32(_sum21, _pA1, _pB0, 2);
-                _sum30 = vfmaq_laneq_f32(_sum30, _pA0, _pB0, 3);
-                _sum31 = vfmaq_laneq_f32(_sum31, _pA1, _pB0, 3);
-                _sum40 = vfmaq_laneq_f32(_sum40, _pA0, _pB1, 0);
-                _sum41 = vfmaq_laneq_f32(_sum41, _pA1, _pB1, 0);
-                _sum50 = vfmaq_laneq_f32(_sum50, _pA0, _pB1, 1);
-                _sum51 = vfmaq_laneq_f32(_sum51, _pA1, _pB1, 1);
-                _sum60 = vfmaq_laneq_f32(_sum60, _pA0, _pB1, 2);
-                _sum61 = vfmaq_laneq_f32(_sum61, _pA1, _pB1, 2);
-                _sum70 = vfmaq_laneq_f32(_sum70, _pA0, _pB1, 3);
-                _sum71 = vfmaq_laneq_f32(_sum71, _pA1, _pB1, 3);
-                _sum80 = vfmaq_laneq_f32(_sum80, _pA0, _pB2, 0);
-                _sum81 = vfmaq_laneq_f32(_sum81, _pA1, _pB2, 0);
-                _sum90 = vfmaq_laneq_f32(_sum90, _pA0, _pB2, 1);
-                _sum91 = vfmaq_laneq_f32(_sum91, _pA1, _pB2, 1);
-                _suma0 = vfmaq_laneq_f32(_suma0, _pA0, _pB2, 2);
-                _suma1 = vfmaq_laneq_f32(_suma1, _pA1, _pB2, 2);
-                _sumb0 = vfmaq_laneq_f32(_sumb0, _pA0, _pB2, 3);
-                _sumb1 = vfmaq_laneq_f32(_sumb1, _pA1, _pB2, 3);
-
-                pA += 8;
-                pB += 12;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum00 = vmulq_f32(_sum00, _alpha);
-                _sum01 = vmulq_f32(_sum01, _alpha);
-                _sum10 = vmulq_f32(_sum10, _alpha);
-                _sum11 = vmulq_f32(_sum11, _alpha);
-                _sum20 = vmulq_f32(_sum20, _alpha);
-                _sum21 = vmulq_f32(_sum21, _alpha);
-                _sum30 = vmulq_f32(_sum30, _alpha);
-                _sum31 = vmulq_f32(_sum31, _alpha);
-                _sum40 = vmulq_f32(_sum40, _alpha);
-                _sum41 = vmulq_f32(_sum41, _alpha);
-                _sum50 = vmulq_f32(_sum50, _alpha);
-                _sum51 = vmulq_f32(_sum51, _alpha);
-                _sum60 = vmulq_f32(_sum60, _alpha);
-                _sum61 = vmulq_f32(_sum61, _alpha);
-                _sum70 = vmulq_f32(_sum70, _alpha);
-                _sum71 = vmulq_f32(_sum71, _alpha);
-                _sum80 = vmulq_f32(_sum80, _alpha);
-                _sum81 = vmulq_f32(_sum81, _alpha);
-                _sum90 = vmulq_f32(_sum90, _alpha);
-                _sum91 = vmulq_f32(_sum91, _alpha);
-                _suma0 = vmulq_f32(_suma0, _alpha);
-                _suma1 = vmulq_f32(_suma1, _alpha);
-                _sumb0 = vmulq_f32(_sumb0, _alpha);
-                _sumb1 = vmulq_f32(_sumb1, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum10));
-                    vst1_u16(outptr0 + 4 * 2, float2bfloat(_sum20));
-                    vst1_u16(outptr0 + 4 * 3, float2bfloat(_sum30));
-                    vst1_u16(outptr0 + 4 * 4, float2bfloat(_sum40));
-                    vst1_u16(outptr0 + 4 * 5, float2bfloat(_sum50));
-                    vst1_u16(outptr0 + 4 * 6, float2bfloat(_sum60));
-                    vst1_u16(outptr0 + 4 * 7, float2bfloat(_sum70));
-                    vst1_u16(outptr0 + 4 * 8, float2bfloat(_sum80));
-                    vst1_u16(outptr0 + 4 * 9, float2bfloat(_sum90));
-                    vst1_u16(outptr0 + 4 * 10, float2bfloat(_suma0));
-                    vst1_u16(outptr0 + 4 * 11, float2bfloat(_sumb0));
-
-                    vst1_u16(outptr0 + out_hstep * 4, float2bfloat(_sum01));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, float2bfloat(_sum11));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 2, float2bfloat(_sum21));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 3, float2bfloat(_sum31));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 4, float2bfloat(_sum41));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 5, float2bfloat(_sum51));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 6, float2bfloat(_sum61));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 7, float2bfloat(_sum71));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 8, float2bfloat(_sum81));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 9, float2bfloat(_sum91));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 10, float2bfloat(_suma1));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 11, float2bfloat(_sumb1));
-
-                    outptr0 += 48;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose8x12_ps(_sum00, _sum01, _sum10, _sum11, _sum20, _sum21, _sum30, _sum31, _sum40, _sum41, _sum50, _sum51, _sum60, _sum61, _sum70, _sum71, _sum80, _sum81, _sum90, _sum91, _suma0, _suma1, _sumb0, _sumb1);
-
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum01));
-                    vst1_u16(outptr0 + 8, float2bfloat(_sum10));
-                    vst1_u16(outptr0 + out_hstep, float2bfloat(_sum11));
-                    vst1_u16(outptr0 + out_hstep + 4, float2bfloat(_sum20));
-                    vst1_u16(outptr0 + out_hstep + 8, float2bfloat(_sum21));
-                    vst1_u16(outptr0 + out_hstep * 2, float2bfloat(_sum30));
-                    vst1_u16(outptr0 + out_hstep * 2 + 4, float2bfloat(_sum31));
-                    vst1_u16(outptr0 + out_hstep * 2 + 8, float2bfloat(_sum40));
-                    vst1_u16(outptr0 + out_hstep * 3, float2bfloat(_sum41));
-                    vst1_u16(outptr0 + out_hstep * 3 + 4, float2bfloat(_sum50));
-                    vst1_u16(outptr0 + out_hstep * 3 + 8, float2bfloat(_sum51));
-                    vst1_u16(outptr0 + out_hstep * 4, float2bfloat(_sum60));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, float2bfloat(_sum61));
-                    vst1_u16(outptr0 + out_hstep * 4 + 8, float2bfloat(_sum70));
-                    vst1_u16(outptr0 + out_hstep * 5, float2bfloat(_sum71));
-                    vst1_u16(outptr0 + out_hstep * 5 + 4, float2bfloat(_sum80));
-                    vst1_u16(outptr0 + out_hstep * 5 + 8, float2bfloat(_sum81));
-                    vst1_u16(outptr0 + out_hstep * 6, float2bfloat(_sum90));
-                    vst1_u16(outptr0 + out_hstep * 6 + 4, float2bfloat(_sum91));
-                    vst1_u16(outptr0 + out_hstep * 6 + 8, float2bfloat(_suma0));
-                    vst1_u16(outptr0 + out_hstep * 7, float2bfloat(_suma1));
-                    vst1_u16(outptr0 + out_hstep * 7 + 4, float2bfloat(_sumb0));
-                    vst1_u16(outptr0 + out_hstep * 7 + 8, float2bfloat(_sumb1));
-
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                vst1q_f32(outptr + 4 * 4, _sum20);
-                vst1q_f32(outptr + 4 * 5, _sum21);
-                vst1q_f32(outptr + 4 * 6, _sum30);
-                vst1q_f32(outptr + 4 * 7, _sum31);
-                vst1q_f32(outptr + 4 * 8, _sum40);
-                vst1q_f32(outptr + 4 * 9, _sum41);
-                vst1q_f32(outptr + 4 * 10, _sum50);
-                vst1q_f32(outptr + 4 * 11, _sum51);
-                vst1q_f32(outptr + 4 * 12, _sum60);
-                vst1q_f32(outptr + 4 * 13, _sum61);
-                vst1q_f32(outptr + 4 * 14, _sum70);
-                vst1q_f32(outptr + 4 * 15, _sum71);
-                vst1q_f32(outptr + 4 * 16, _sum80);
-                vst1q_f32(outptr + 4 * 17, _sum81);
-                vst1q_f32(outptr + 4 * 18, _sum90);
-                vst1q_f32(outptr + 4 * 19, _sum91);
-                vst1q_f32(outptr + 4 * 20, _suma0);
-                vst1q_f32(outptr + 4 * 21, _suma1);
-                vst1q_f32(outptr + 4 * 22, _sumb0);
-                vst1q_f32(outptr + 4 * 23, _sumb1);
-            }
-
-            outptr += 96;
-        }
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum20;
-            float32x4_t _sum21;
-            float32x4_t _sum30;
-            float32x4_t _sum31;
-            float32x4_t _sum40;
-            float32x4_t _sum41;
-            float32x4_t _sum50;
-            float32x4_t _sum51;
-            float32x4_t _sum60;
-            float32x4_t _sum61;
-            float32x4_t _sum70;
-            float32x4_t _sum71;
-
-            if (k == 0)
-            {
-                _sum00 = vdupq_n_f32(0.f);
-                _sum01 = vdupq_n_f32(0.f);
-                _sum10 = vdupq_n_f32(0.f);
-                _sum11 = vdupq_n_f32(0.f);
-                _sum20 = vdupq_n_f32(0.f);
-                _sum21 = vdupq_n_f32(0.f);
-                _sum30 = vdupq_n_f32(0.f);
-                _sum31 = vdupq_n_f32(0.f);
-                _sum40 = vdupq_n_f32(0.f);
-                _sum41 = vdupq_n_f32(0.f);
-                _sum50 = vdupq_n_f32(0.f);
-                _sum51 = vdupq_n_f32(0.f);
-                _sum60 = vdupq_n_f32(0.f);
-                _sum61 = vdupq_n_f32(0.f);
-                _sum70 = vdupq_n_f32(0.f);
-                _sum71 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum10 = _sum00;
-                        _sum11 = _sum00;
-                        _sum20 = _sum00;
-                        _sum21 = _sum00;
-                        _sum30 = _sum00;
-                        _sum31 = _sum00;
-                        _sum40 = _sum00;
-                        _sum41 = _sum00;
-                        _sum50 = _sum00;
-                        _sum51 = _sum00;
-                        _sum60 = _sum00;
-                        _sum61 = _sum00;
-                        _sum70 = _sum00;
-                        _sum71 = _sum00;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                        _sum10 = _sum00;
-                        _sum11 = _sum01;
-                        _sum20 = _sum00;
-                        _sum21 = _sum01;
-                        _sum30 = _sum00;
-                        _sum31 = _sum01;
-                        _sum40 = _sum00;
-                        _sum41 = _sum01;
-                        _sum50 = _sum00;
-                        _sum51 = _sum01;
-                        _sum60 = _sum00;
-                        _sum61 = _sum01;
-                        _sum70 = _sum00;
-                        _sum71 = _sum01;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4 * 1);
-                        _sum10 = vld1q_f32(pC + 4 * 2);
-                        _sum11 = vld1q_f32(pC + 4 * 3);
-                        _sum20 = vld1q_f32(pC + 4 * 4);
-                        _sum21 = vld1q_f32(pC + 4 * 5);
-                        _sum30 = vld1q_f32(pC + 4 * 6);
-                        _sum31 = vld1q_f32(pC + 4 * 7);
-                        _sum40 = vld1q_f32(pC + 4 * 8);
-                        _sum41 = vld1q_f32(pC + 4 * 9);
-                        _sum50 = vld1q_f32(pC + 4 * 10);
-                        _sum51 = vld1q_f32(pC + 4 * 11);
-                        _sum60 = vld1q_f32(pC + 4 * 12);
-                        _sum61 = vld1q_f32(pC + 4 * 13);
-                        _sum70 = vld1q_f32(pC + 4 * 14);
-                        _sum71 = vld1q_f32(pC + 4 * 15);
-                        pC += 64;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum10 = vdupq_n_f32(pC[1]);
-                        _sum20 = vdupq_n_f32(pC[2]);
-                        _sum30 = vdupq_n_f32(pC[3]);
-                        _sum40 = vdupq_n_f32(pC[4]);
-                        _sum50 = vdupq_n_f32(pC[5]);
-                        _sum60 = vdupq_n_f32(pC[6]);
-                        _sum70 = vdupq_n_f32(pC[7]);
-                        _sum01 = _sum00;
-                        _sum11 = _sum10;
-                        _sum21 = _sum20;
-                        _sum31 = _sum30;
-                        _sum41 = _sum40;
-                        _sum51 = _sum50;
-                        _sum61 = _sum60;
-                        _sum71 = _sum70;
-                        pC += 8;
-                    }
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-                _sum20 = vld1q_f32(outptr + 4 * 4);
-                _sum21 = vld1q_f32(outptr + 4 * 5);
-                _sum30 = vld1q_f32(outptr + 4 * 6);
-                _sum31 = vld1q_f32(outptr + 4 * 7);
-                _sum40 = vld1q_f32(outptr + 4 * 8);
-                _sum41 = vld1q_f32(outptr + 4 * 9);
-                _sum50 = vld1q_f32(outptr + 4 * 10);
-                _sum51 = vld1q_f32(outptr + 4 * 11);
-                _sum60 = vld1q_f32(outptr + 4 * 12);
-                _sum61 = vld1q_f32(outptr + 4 * 13);
-                _sum70 = vld1q_f32(outptr + 4 * 14);
-                _sum71 = vld1q_f32(outptr + 4 * 15);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                uint16x8_t _pA = vld1q_u16(pA);
-                float32x4_t _pA0 = bfloat2float(vget_low_u16(_pA));
-                float32x4_t _pA1 = bfloat2float(vget_high_u16(_pA));
-
-                float32x4_t _pB0 = bfloat2float(vld1_u16(pB));
-                float32x4_t _pB1 = bfloat2float(vld1_u16(pB + 4));
-
-                _sum00 = vfmaq_laneq_f32(_sum00, _pA0, _pB0, 0);
-                _sum01 = vfmaq_laneq_f32(_sum01, _pA1, _pB0, 0);
-                _sum10 = vfmaq_laneq_f32(_sum10, _pA0, _pB0, 1);
-                _sum11 = vfmaq_laneq_f32(_sum11, _pA1, _pB0, 1);
-                _sum20 = vfmaq_laneq_f32(_sum20, _pA0, _pB0, 2);
-                _sum21 = vfmaq_laneq_f32(_sum21, _pA1, _pB0, 2);
-                _sum30 = vfmaq_laneq_f32(_sum30, _pA0, _pB0, 3);
-                _sum31 = vfmaq_laneq_f32(_sum31, _pA1, _pB0, 3);
-                _sum40 = vfmaq_laneq_f32(_sum40, _pA0, _pB1, 0);
-                _sum41 = vfmaq_laneq_f32(_sum41, _pA1, _pB1, 0);
-                _sum50 = vfmaq_laneq_f32(_sum50, _pA0, _pB1, 1);
-                _sum51 = vfmaq_laneq_f32(_sum51, _pA1, _pB1, 1);
-                _sum60 = vfmaq_laneq_f32(_sum60, _pA0, _pB1, 2);
-                _sum61 = vfmaq_laneq_f32(_sum61, _pA1, _pB1, 2);
-                _sum70 = vfmaq_laneq_f32(_sum70, _pA0, _pB1, 3);
-                _sum71 = vfmaq_laneq_f32(_sum71, _pA1, _pB1, 3);
-
-                pA += 8;
-                pB += 8;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum00 = vmulq_f32(_sum00, _alpha);
-                _sum01 = vmulq_f32(_sum01, _alpha);
-                _sum10 = vmulq_f32(_sum10, _alpha);
-                _sum11 = vmulq_f32(_sum11, _alpha);
-                _sum20 = vmulq_f32(_sum20, _alpha);
-                _sum21 = vmulq_f32(_sum21, _alpha);
-                _sum30 = vmulq_f32(_sum30, _alpha);
-                _sum31 = vmulq_f32(_sum31, _alpha);
-                _sum40 = vmulq_f32(_sum40, _alpha);
-                _sum41 = vmulq_f32(_sum41, _alpha);
-                _sum50 = vmulq_f32(_sum50, _alpha);
-                _sum51 = vmulq_f32(_sum51, _alpha);
-                _sum60 = vmulq_f32(_sum60, _alpha);
-                _sum61 = vmulq_f32(_sum61, _alpha);
-                _sum70 = vmulq_f32(_sum70, _alpha);
-                _sum71 = vmulq_f32(_sum71, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum10));
-                    vst1_u16(outptr0 + 4 * 2, float2bfloat(_sum20));
-                    vst1_u16(outptr0 + 4 * 3, float2bfloat(_sum30));
-                    vst1_u16(outptr0 + 4 * 4, float2bfloat(_sum40));
-                    vst1_u16(outptr0 + 4 * 5, float2bfloat(_sum50));
-                    vst1_u16(outptr0 + 4 * 6, float2bfloat(_sum60));
-                    vst1_u16(outptr0 + 4 * 7, float2bfloat(_sum70));
-
-                    vst1_u16(outptr0 + out_hstep * 4, float2bfloat(_sum01));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, float2bfloat(_sum11));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 2, float2bfloat(_sum21));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 3, float2bfloat(_sum31));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 4, float2bfloat(_sum41));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 5, float2bfloat(_sum51));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 6, float2bfloat(_sum61));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 7, float2bfloat(_sum71));
-
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose8x8_ps(_sum00, _sum01, _sum10, _sum11, _sum20, _sum21, _sum30, _sum31, _sum40, _sum41, _sum50, _sum51, _sum60, _sum61, _sum70, _sum71);
-
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum01));
-                    vst1_u16(outptr0 + out_hstep, float2bfloat(_sum10));
-                    vst1_u16(outptr0 + out_hstep + 4, float2bfloat(_sum11));
-                    vst1_u16(outptr0 + out_hstep * 2, float2bfloat(_sum20));
-                    vst1_u16(outptr0 + out_hstep * 2 + 4, float2bfloat(_sum21));
-                    vst1_u16(outptr0 + out_hstep * 3, float2bfloat(_sum30));
-                    vst1_u16(outptr0 + out_hstep * 3 + 4, float2bfloat(_sum31));
-                    vst1_u16(outptr0 + out_hstep * 4, float2bfloat(_sum40));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, float2bfloat(_sum41));
-                    vst1_u16(outptr0 + out_hstep * 5, float2bfloat(_sum50));
-                    vst1_u16(outptr0 + out_hstep * 5 + 4, float2bfloat(_sum51));
-                    vst1_u16(outptr0 + out_hstep * 6, float2bfloat(_sum60));
-                    vst1_u16(outptr0 + out_hstep * 6 + 4, float2bfloat(_sum61));
-                    vst1_u16(outptr0 + out_hstep * 7, float2bfloat(_sum70));
-                    vst1_u16(outptr0 + out_hstep * 7 + 4, float2bfloat(_sum71));
-
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                vst1q_f32(outptr + 4 * 4, _sum20);
-                vst1q_f32(outptr + 4 * 5, _sum21);
-                vst1q_f32(outptr + 4 * 6, _sum30);
-                vst1q_f32(outptr + 4 * 7, _sum31);
-                vst1q_f32(outptr + 4 * 8, _sum40);
-                vst1q_f32(outptr + 4 * 9, _sum41);
-                vst1q_f32(outptr + 4 * 10, _sum50);
-                vst1q_f32(outptr + 4 * 11, _sum51);
-                vst1q_f32(outptr + 4 * 12, _sum60);
-                vst1q_f32(outptr + 4 * 13, _sum61);
-                vst1q_f32(outptr + 4 * 14, _sum70);
-                vst1q_f32(outptr + 4 * 15, _sum71);
-            }
-
-            outptr += 64;
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum20;
-            float32x4_t _sum21;
-            float32x4_t _sum30;
-            float32x4_t _sum31;
-
-            if (k == 0)
-            {
-                _sum00 = vdupq_n_f32(0.f);
-                _sum01 = vdupq_n_f32(0.f);
-                _sum10 = vdupq_n_f32(0.f);
-                _sum11 = vdupq_n_f32(0.f);
-                _sum20 = vdupq_n_f32(0.f);
-                _sum21 = vdupq_n_f32(0.f);
-                _sum30 = vdupq_n_f32(0.f);
-                _sum31 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum10 = _sum00;
-                        _sum11 = _sum00;
-                        _sum20 = _sum00;
-                        _sum21 = _sum00;
-                        _sum30 = _sum00;
-                        _sum31 = _sum00;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                        _sum10 = _sum00;
-                        _sum11 = _sum01;
-                        _sum20 = _sum00;
-                        _sum21 = _sum01;
-                        _sum30 = _sum00;
-                        _sum31 = _sum01;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4 * 1);
-                        _sum10 = vld1q_f32(pC + 4 * 2);
-                        _sum11 = vld1q_f32(pC + 4 * 3);
-                        _sum20 = vld1q_f32(pC + 4 * 4);
-                        _sum21 = vld1q_f32(pC + 4 * 5);
-                        _sum30 = vld1q_f32(pC + 4 * 6);
-                        _sum31 = vld1q_f32(pC + 4 * 7);
-                        pC += 32;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum10 = vdupq_n_f32(pC[1]);
-                        _sum20 = vdupq_n_f32(pC[2]);
-                        _sum30 = vdupq_n_f32(pC[3]);
-                        _sum01 = _sum00;
-                        _sum11 = _sum10;
-                        _sum21 = _sum20;
-                        _sum31 = _sum30;
-                        pC += 4;
-                    }
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-                _sum20 = vld1q_f32(outptr + 4 * 4);
-                _sum21 = vld1q_f32(outptr + 4 * 5);
-                _sum30 = vld1q_f32(outptr + 4 * 6);
-                _sum31 = vld1q_f32(outptr + 4 * 7);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                uint16x8_t _pA = vld1q_u16(pA);
-                float32x4_t _pA0 = bfloat2float(vget_low_u16(_pA));
-                float32x4_t _pA1 = bfloat2float(vget_high_u16(_pA));
-
-                float32x4_t _pB0 = bfloat2float(vld1_u16(pB));
-
-                _sum00 = vfmaq_laneq_f32(_sum00, _pA0, _pB0, 0);
-                _sum01 = vfmaq_laneq_f32(_sum01, _pA1, _pB0, 0);
-                _sum10 = vfmaq_laneq_f32(_sum10, _pA0, _pB0, 1);
-                _sum11 = vfmaq_laneq_f32(_sum11, _pA1, _pB0, 1);
-                _sum20 = vfmaq_laneq_f32(_sum20, _pA0, _pB0, 2);
-                _sum21 = vfmaq_laneq_f32(_sum21, _pA1, _pB0, 2);
-                _sum30 = vfmaq_laneq_f32(_sum30, _pA0, _pB0, 3);
-                _sum31 = vfmaq_laneq_f32(_sum31, _pA1, _pB0, 3);
-
-                pA += 8;
-                pB += 4;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum00 = vmulq_f32(_sum00, _alpha);
-                _sum01 = vmulq_f32(_sum01, _alpha);
-                _sum10 = vmulq_f32(_sum10, _alpha);
-                _sum11 = vmulq_f32(_sum11, _alpha);
-                _sum20 = vmulq_f32(_sum20, _alpha);
-                _sum21 = vmulq_f32(_sum21, _alpha);
-                _sum30 = vmulq_f32(_sum30, _alpha);
-                _sum31 = vmulq_f32(_sum31, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum10));
-                    vst1_u16(outptr0 + 4 * 2, float2bfloat(_sum20));
-                    vst1_u16(outptr0 + 4 * 3, float2bfloat(_sum30));
-
-                    vst1_u16(outptr0 + out_hstep * 4, float2bfloat(_sum01));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, float2bfloat(_sum11));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 2, float2bfloat(_sum21));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 3, float2bfloat(_sum31));
-
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose8x4_ps(_sum00, _sum01, _sum10, _sum11, _sum20, _sum21, _sum30, _sum31);
-
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + out_hstep * 1, float2bfloat(_sum01));
-                    vst1_u16(outptr0 + out_hstep * 2, float2bfloat(_sum10));
-                    vst1_u16(outptr0 + out_hstep * 3, float2bfloat(_sum11));
-                    vst1_u16(outptr0 + out_hstep * 4, float2bfloat(_sum20));
-                    vst1_u16(outptr0 + out_hstep * 5, float2bfloat(_sum21));
-                    vst1_u16(outptr0 + out_hstep * 6, float2bfloat(_sum30));
-                    vst1_u16(outptr0 + out_hstep * 7, float2bfloat(_sum31));
-
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                vst1q_f32(outptr + 4 * 4, _sum20);
-                vst1q_f32(outptr + 4 * 5, _sum21);
-                vst1q_f32(outptr + 4 * 6, _sum30);
-                vst1q_f32(outptr + 4 * 7, _sum31);
-            }
-
-            outptr += 32;
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-
-            if (k == 0)
-            {
-                _sum00 = vdupq_n_f32(0.f);
-                _sum01 = vdupq_n_f32(0.f);
-                _sum10 = vdupq_n_f32(0.f);
-                _sum11 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum10 = _sum00;
-                        _sum11 = _sum00;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                        _sum10 = _sum00;
-                        _sum11 = _sum01;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4 * 1);
-                        _sum10 = vld1q_f32(pC + 4 * 2);
-                        _sum11 = vld1q_f32(pC + 4 * 3);
-                        pC += 16;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum10 = vdupq_n_f32(pC[1]);
-                        _sum01 = _sum00;
-                        _sum11 = _sum10;
-                        pC += 2;
-                    }
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                uint16x8_t _pA = vld1q_u16(pA);
-                float32x4_t _pA0 = bfloat2float(vget_low_u16(_pA));
-                float32x4_t _pA1 = bfloat2float(vget_high_u16(_pA));
-
-                float32x4_t _pB0 = bfloat2float(vdup_n_u16(pB[0]));
-                float32x4_t _pB1 = bfloat2float(vdup_n_u16(pB[1]));
-
-                _sum00 = vfmaq_f32(_sum00, _pA0, _pB0);
-                _sum01 = vfmaq_f32(_sum01, _pA1, _pB0);
-                _sum10 = vfmaq_f32(_sum10, _pA0, _pB1);
-                _sum11 = vfmaq_f32(_sum11, _pA1, _pB1);
-
-                pA += 8;
-                pB += 2;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum00 = vmulq_f32(_sum00, _alpha);
-                _sum01 = vmulq_f32(_sum01, _alpha);
-                _sum10 = vmulq_f32(_sum10, _alpha);
-                _sum11 = vmulq_f32(_sum11, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum10));
-
-                    vst1_u16(outptr0 + out_hstep * 4, float2bfloat(_sum01));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, float2bfloat(_sum11));
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short sum0[8];
-                    unsigned short sum1[8];
-                    vst1_u16(sum0, float2bfloat(_sum00));
-                    vst1_u16(sum0 + 4, float2bfloat(_sum01));
-                    vst1_u16(sum1, float2bfloat(_sum10));
-                    vst1_u16(sum1 + 4, float2bfloat(_sum11));
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[out_hstep * 4] = sum0[4];
-                    outptr0[out_hstep * 5] = sum0[5];
-                    outptr0[out_hstep * 6] = sum0[6];
-                    outptr0[out_hstep * 7] = sum0[7];
-
-                    outptr0[1] = sum1[0];
-                    outptr0[out_hstep + 1] = sum1[1];
-                    outptr0[out_hstep * 2 + 1] = sum1[2];
-                    outptr0[out_hstep * 3 + 1] = sum1[3];
-                    outptr0[out_hstep * 4 + 1] = sum1[4];
-                    outptr0[out_hstep * 5 + 1] = sum1[5];
-                    outptr0[out_hstep * 6 + 1] = sum1[6];
-                    outptr0[out_hstep * 7 + 1] = sum1[7];
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-            }
-
-            outptr += 16;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-
-            if (k == 0)
-            {
-                _sum00 = vdupq_n_f32(0.f);
-                _sum01 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                        pC += 8;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        pC += 1;
-                    }
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                uint16x8_t _pA = vld1q_u16(pA);
-                float32x4_t _pA0 = bfloat2float(vget_low_u16(_pA));
-                float32x4_t _pA1 = bfloat2float(vget_high_u16(_pA));
-
-                float32x4_t _pB = bfloat2float(vld1_dup_u16(pB));
-
-                _sum00 = vfmaq_f32(_sum00, _pA0, _pB);
-                _sum01 = vfmaq_f32(_sum01, _pA1, _pB);
-
-                pA += 8;
-                pB += 1;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum00 = vmulq_f32(_sum00, _alpha);
-                _sum01 = vmulq_f32(_sum01, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + out_hstep * 4, float2bfloat(_sum01));
-                    outptr0 += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short sum0[8];
-                    vst1_u16(sum0, float2bfloat(_sum00));
-                    vst1_u16(sum0 + 4, float2bfloat(_sum01));
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep * 1] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[out_hstep * 4] = sum0[4];
-                    outptr0[out_hstep * 5] = sum0[5];
-                    outptr0[out_hstep * 6] = sum0[6];
-                    outptr0[out_hstep * 7] = sum0[7];
-                    outptr0++;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-            }
-
-            outptr += 8;
-        }
-
-        pAT += max_kk * 8;
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        unsigned short* outptr0 = (unsigned short*)top_blob + (i + ii) * out_hstep + j * out_elempack;
-
-        const unsigned short* pB = pBT;
-
-        if (pC)
-        {
-            if (broadcast_type_C == 1 || broadcast_type_C == 2)
-            {
-                pC = (const float*)CT_tile + i + ii;
-            }
-            if (broadcast_type_C == 4)
-            {
-                pC = (const float*)CT_tile + j;
-            }
-        }
-
-        int jj = 0;
-#if __aarch64__
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-            float32x4_t _sum3;
-            float32x4_t _sum4;
-            float32x4_t _sum5;
-            float32x4_t _sum6;
-            float32x4_t _sum7;
-            float32x4_t _sum8;
-            float32x4_t _sum9;
-            float32x4_t _suma;
-            float32x4_t _sumb;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-                _sum2 = vdupq_n_f32(0.f);
-                _sum3 = vdupq_n_f32(0.f);
-                _sum4 = vdupq_n_f32(0.f);
-                _sum5 = vdupq_n_f32(0.f);
-                _sum6 = vdupq_n_f32(0.f);
-                _sum7 = vdupq_n_f32(0.f);
-                _sum8 = vdupq_n_f32(0.f);
-                _sum9 = vdupq_n_f32(0.f);
-                _suma = vdupq_n_f32(0.f);
-                _sumb = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = _sum0;
-                        _sum2 = _sum0;
-                        _sum3 = _sum0;
-                        _sum4 = _sum0;
-                        _sum5 = _sum0;
-                        _sum6 = _sum0;
-                        _sum7 = _sum0;
-                        _sum8 = _sum0;
-                        _sum9 = _sum0;
-                        _suma = _sum0;
-                        _sumb = _sum0;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = _sum0;
-                        _sum2 = _sum0;
-                        _sum3 = _sum0;
-                        _sum4 = _sum0;
-                        _sum5 = _sum0;
-                        _sum6 = _sum0;
-                        _sum7 = _sum0;
-                        _sum8 = _sum0;
-                        _sum9 = _sum0;
-                        _suma = _sum0;
-                        _sumb = _sum0;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = vld1q_f32(pC + 4);
-                        _sum2 = vld1q_f32(pC + 8);
-                        _sum3 = vld1q_f32(pC + 12);
-                        _sum4 = vld1q_f32(pC + 16);
-                        _sum5 = vld1q_f32(pC + 20);
-                        _sum6 = vld1q_f32(pC + 24);
-                        _sum7 = vld1q_f32(pC + 28);
-                        _sum8 = vld1q_f32(pC + 32);
-                        _sum9 = vld1q_f32(pC + 36);
-                        _suma = vld1q_f32(pC + 40);
-                        _sumb = vld1q_f32(pC + 44);
-                        pC += 48;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = vdupq_n_f32(pC[1]);
-                        _sum2 = vdupq_n_f32(pC[2]);
-                        _sum3 = vdupq_n_f32(pC[3]);
-                        _sum4 = vdupq_n_f32(pC[4]);
-                        _sum5 = vdupq_n_f32(pC[5]);
-                        _sum6 = vdupq_n_f32(pC[6]);
-                        _sum7 = vdupq_n_f32(pC[7]);
-                        _sum8 = vdupq_n_f32(pC[8]);
-                        _sum9 = vdupq_n_f32(pC[9]);
-                        _suma = vdupq_n_f32(pC[10]);
-                        _sumb = vdupq_n_f32(pC[11]);
-                        pC += 12;
-                    }
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4 * 1);
-                _sum2 = vld1q_f32(outptr + 4 * 2);
-                _sum3 = vld1q_f32(outptr + 4 * 3);
-                _sum4 = vld1q_f32(outptr + 4 * 4);
-                _sum5 = vld1q_f32(outptr + 4 * 5);
-                _sum6 = vld1q_f32(outptr + 4 * 6);
-                _sum7 = vld1q_f32(outptr + 4 * 7);
-                _sum8 = vld1q_f32(outptr + 4 * 8);
-                _sum9 = vld1q_f32(outptr + 4 * 9);
-                _suma = vld1q_f32(outptr + 4 * 10);
-                _sumb = vld1q_f32(outptr + 4 * 11);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA = bfloat2float(vld1_u16(pA));
-                float32x4_t _pB0 = bfloat2float(vld1_u16(pB));
-                float32x4_t _pB1 = bfloat2float(vld1_u16(pB + 4));
-                float32x4_t _pB2 = bfloat2float(vld1_u16(pB + 8));
-
-                _sum0 = vfmaq_laneq_f32(_sum0, _pA, _pB0, 0);
-                _sum1 = vfmaq_laneq_f32(_sum1, _pA, _pB0, 1);
-                _sum2 = vfmaq_laneq_f32(_sum2, _pA, _pB0, 2);
-                _sum3 = vfmaq_laneq_f32(_sum3, _pA, _pB0, 3);
-                _sum4 = vfmaq_laneq_f32(_sum4, _pA, _pB1, 0);
-                _sum5 = vfmaq_laneq_f32(_sum5, _pA, _pB1, 1);
-                _sum6 = vfmaq_laneq_f32(_sum6, _pA, _pB1, 2);
-                _sum7 = vfmaq_laneq_f32(_sum7, _pA, _pB1, 3);
-                _sum8 = vfmaq_laneq_f32(_sum8, _pA, _pB2, 0);
-                _sum9 = vfmaq_laneq_f32(_sum9, _pA, _pB2, 1);
-                _suma = vfmaq_laneq_f32(_suma, _pA, _pB2, 2);
-                _sumb = vfmaq_laneq_f32(_sumb, _pA, _pB2, 3);
-
-                pA += 4;
-                pB += 12;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-                _sum1 = vmulq_f32(_sum1, _alpha);
-                _sum2 = vmulq_f32(_sum2, _alpha);
-                _sum3 = vmulq_f32(_sum3, _alpha);
-                _sum4 = vmulq_f32(_sum4, _alpha);
-                _sum5 = vmulq_f32(_sum5, _alpha);
-                _sum6 = vmulq_f32(_sum6, _alpha);
-                _sum7 = vmulq_f32(_sum7, _alpha);
-                _sum8 = vmulq_f32(_sum8, _alpha);
-                _sum9 = vmulq_f32(_sum9, _alpha);
-                _suma = vmulq_f32(_suma, _alpha);
-                _sumb = vmulq_f32(_sumb, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum1));
-                    vst1_u16(outptr0 + 4 * 2, float2bfloat(_sum2));
-                    vst1_u16(outptr0 + 4 * 3, float2bfloat(_sum3));
-                    vst1_u16(outptr0 + 4 * 4, float2bfloat(_sum4));
-                    vst1_u16(outptr0 + 4 * 5, float2bfloat(_sum5));
-                    vst1_u16(outptr0 + 4 * 6, float2bfloat(_sum6));
-                    vst1_u16(outptr0 + 4 * 7, float2bfloat(_sum7));
-                    vst1_u16(outptr0 + 4 * 8, float2bfloat(_sum8));
-                    vst1_u16(outptr0 + 4 * 9, float2bfloat(_sum9));
-                    vst1_u16(outptr0 + 4 * 10, float2bfloat(_suma));
-                    vst1_u16(outptr0 + 4 * 11, float2bfloat(_sumb));
-                    outptr0 += 48;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose4x12_ps(_sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7, _sum8, _sum9, _suma, _sumb);
-
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum1));
-                    vst1_u16(outptr0 + 8, float2bfloat(_sum2));
-                    vst1_u16(outptr0 + out_hstep, float2bfloat(_sum3));
-                    vst1_u16(outptr0 + out_hstep + 4, float2bfloat(_sum4));
-                    vst1_u16(outptr0 + out_hstep + 8, float2bfloat(_sum5));
-                    vst1_u16(outptr0 + out_hstep * 2, float2bfloat(_sum6));
-                    vst1_u16(outptr0 + out_hstep * 2 + 4, float2bfloat(_sum7));
-                    vst1_u16(outptr0 + out_hstep * 2 + 8, float2bfloat(_sum8));
-                    vst1_u16(outptr0 + out_hstep * 3, float2bfloat(_sum9));
-                    vst1_u16(outptr0 + out_hstep * 3 + 4, float2bfloat(_suma));
-                    vst1_u16(outptr0 + out_hstep * 3 + 8, float2bfloat(_sumb));
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 4 * 2, _sum2);
-                vst1q_f32(outptr + 4 * 3, _sum3);
-                vst1q_f32(outptr + 4 * 4, _sum4);
-                vst1q_f32(outptr + 4 * 5, _sum5);
-                vst1q_f32(outptr + 4 * 6, _sum6);
-                vst1q_f32(outptr + 4 * 7, _sum7);
-                vst1q_f32(outptr + 4 * 8, _sum8);
-                vst1q_f32(outptr + 4 * 9, _sum9);
-                vst1q_f32(outptr + 4 * 10, _suma);
-                vst1q_f32(outptr + 4 * 11, _sumb);
-            }
-
-            outptr += 48;
-        }
-#endif // __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-            float32x4_t _sum3;
-            float32x4_t _sum4;
-            float32x4_t _sum5;
-            float32x4_t _sum6;
-            float32x4_t _sum7;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-                _sum2 = vdupq_n_f32(0.f);
-                _sum3 = vdupq_n_f32(0.f);
-                _sum4 = vdupq_n_f32(0.f);
-                _sum5 = vdupq_n_f32(0.f);
-                _sum6 = vdupq_n_f32(0.f);
-                _sum7 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = _sum0;
-                        _sum2 = _sum0;
-                        _sum3 = _sum0;
-                        _sum4 = _sum0;
-                        _sum5 = _sum0;
-                        _sum6 = _sum0;
-                        _sum7 = _sum0;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = _sum0;
-                        _sum2 = _sum0;
-                        _sum3 = _sum0;
-                        _sum4 = _sum0;
-                        _sum5 = _sum0;
-                        _sum6 = _sum0;
-                        _sum7 = _sum0;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = vld1q_f32(pC + 4);
-                        _sum2 = vld1q_f32(pC + 8);
-                        _sum3 = vld1q_f32(pC + 12);
-                        _sum4 = vld1q_f32(pC + 16);
-                        _sum5 = vld1q_f32(pC + 20);
-                        _sum6 = vld1q_f32(pC + 24);
-                        _sum7 = vld1q_f32(pC + 28);
-                        pC += 32;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = vdupq_n_f32(pC[1]);
-                        _sum2 = vdupq_n_f32(pC[2]);
-                        _sum3 = vdupq_n_f32(pC[3]);
-                        _sum4 = vdupq_n_f32(pC[4]);
-                        _sum5 = vdupq_n_f32(pC[5]);
-                        _sum6 = vdupq_n_f32(pC[6]);
-                        _sum7 = vdupq_n_f32(pC[7]);
-                        pC += 8;
-                    }
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4 * 1);
-                _sum2 = vld1q_f32(outptr + 4 * 2);
-                _sum3 = vld1q_f32(outptr + 4 * 3);
-                _sum4 = vld1q_f32(outptr + 4 * 4);
-                _sum5 = vld1q_f32(outptr + 4 * 5);
-                _sum6 = vld1q_f32(outptr + 4 * 6);
-                _sum7 = vld1q_f32(outptr + 4 * 7);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA = bfloat2float(vld1_u16(pA));
-                float32x4_t _pB0 = bfloat2float(vld1_u16(pB));
-                float32x4_t _pB1 = bfloat2float(vld1_u16(pB + 4));
-
-#if __aarch64__
-                _sum0 = vfmaq_laneq_f32(_sum0, _pA, _pB0, 0);
-                _sum1 = vfmaq_laneq_f32(_sum1, _pA, _pB0, 1);
-                _sum2 = vfmaq_laneq_f32(_sum2, _pA, _pB0, 2);
-                _sum3 = vfmaq_laneq_f32(_sum3, _pA, _pB0, 3);
-                _sum4 = vfmaq_laneq_f32(_sum4, _pA, _pB1, 0);
-                _sum5 = vfmaq_laneq_f32(_sum5, _pA, _pB1, 1);
-                _sum6 = vfmaq_laneq_f32(_sum6, _pA, _pB1, 2);
-                _sum7 = vfmaq_laneq_f32(_sum7, _pA, _pB1, 3);
-#else
-                _sum0 = vmlaq_lane_f32(_sum0, _pA, vget_low_f32(_pB0), 0);
-                _sum1 = vmlaq_lane_f32(_sum1, _pA, vget_low_f32(_pB0), 1);
-                _sum2 = vmlaq_lane_f32(_sum2, _pA, vget_high_f32(_pB0), 0);
-                _sum3 = vmlaq_lane_f32(_sum3, _pA, vget_high_f32(_pB0), 1);
-                _sum4 = vmlaq_lane_f32(_sum4, _pA, vget_low_f32(_pB1), 0);
-                _sum5 = vmlaq_lane_f32(_sum5, _pA, vget_low_f32(_pB1), 1);
-                _sum6 = vmlaq_lane_f32(_sum6, _pA, vget_high_f32(_pB1), 0);
-                _sum7 = vmlaq_lane_f32(_sum7, _pA, vget_high_f32(_pB1), 1);
-#endif
-
-                pA += 4;
-                pB += 8;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-                _sum1 = vmulq_f32(_sum1, _alpha);
-                _sum2 = vmulq_f32(_sum2, _alpha);
-                _sum3 = vmulq_f32(_sum3, _alpha);
-                _sum4 = vmulq_f32(_sum4, _alpha);
-                _sum5 = vmulq_f32(_sum5, _alpha);
-                _sum6 = vmulq_f32(_sum6, _alpha);
-                _sum7 = vmulq_f32(_sum7, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum1));
-                    vst1_u16(outptr0 + 4 * 2, float2bfloat(_sum2));
-                    vst1_u16(outptr0 + 4 * 3, float2bfloat(_sum3));
-                    vst1_u16(outptr0 + 4 * 4, float2bfloat(_sum4));
-                    vst1_u16(outptr0 + 4 * 5, float2bfloat(_sum5));
-                    vst1_u16(outptr0 + 4 * 6, float2bfloat(_sum6));
-                    vst1_u16(outptr0 + 4 * 7, float2bfloat(_sum7));
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose4x8_ps(_sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7);
-
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum1));
-                    vst1_u16(outptr0 + out_hstep, float2bfloat(_sum2));
-                    vst1_u16(outptr0 + out_hstep + 4, float2bfloat(_sum3));
-                    vst1_u16(outptr0 + out_hstep * 2, float2bfloat(_sum4));
-                    vst1_u16(outptr0 + out_hstep * 2 + 4, float2bfloat(_sum5));
-                    vst1_u16(outptr0 + out_hstep * 3, float2bfloat(_sum6));
-                    vst1_u16(outptr0 + out_hstep * 3 + 4, float2bfloat(_sum7));
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 4 * 2, _sum2);
-                vst1q_f32(outptr + 4 * 3, _sum3);
-                vst1q_f32(outptr + 4 * 4, _sum4);
-                vst1q_f32(outptr + 4 * 5, _sum5);
-                vst1q_f32(outptr + 4 * 6, _sum6);
-                vst1q_f32(outptr + 4 * 7, _sum7);
-            }
-
-            outptr += 32;
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-            float32x4_t _sum3;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-                _sum2 = vdupq_n_f32(0.f);
-                _sum3 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = _sum0;
-                        _sum2 = _sum0;
-                        _sum3 = _sum0;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = _sum0;
-                        _sum2 = _sum0;
-                        _sum3 = _sum0;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = vld1q_f32(pC + 4);
-                        _sum2 = vld1q_f32(pC + 8);
-                        _sum3 = vld1q_f32(pC + 12);
-                        pC += 16;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = vdupq_n_f32(pC[1]);
-                        _sum2 = vdupq_n_f32(pC[2]);
-                        _sum3 = vdupq_n_f32(pC[3]);
-                        pC += 4;
-                    }
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4 * 1);
-                _sum2 = vld1q_f32(outptr + 4 * 2);
-                _sum3 = vld1q_f32(outptr + 4 * 3);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA = bfloat2float(vld1_u16(pA));
-                float32x4_t _pB = bfloat2float(vld1_u16(pB));
-
-#if __aarch64__
-                _sum0 = vfmaq_laneq_f32(_sum0, _pA, _pB, 0);
-                _sum1 = vfmaq_laneq_f32(_sum1, _pA, _pB, 1);
-                _sum2 = vfmaq_laneq_f32(_sum2, _pA, _pB, 2);
-                _sum3 = vfmaq_laneq_f32(_sum3, _pA, _pB, 3);
-#else
-                _sum0 = vmlaq_lane_f32(_sum0, _pA, vget_low_f32(_pB), 0);
-                _sum1 = vmlaq_lane_f32(_sum1, _pA, vget_low_f32(_pB), 1);
-                _sum2 = vmlaq_lane_f32(_sum2, _pA, vget_high_f32(_pB), 0);
-                _sum3 = vmlaq_lane_f32(_sum3, _pA, vget_high_f32(_pB), 1);
-#endif
-
-                pA += 4;
-                pB += 4;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-                _sum1 = vmulq_f32(_sum1, _alpha);
-                _sum2 = vmulq_f32(_sum2, _alpha);
-                _sum3 = vmulq_f32(_sum3, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum1));
-                    vst1_u16(outptr0 + 4 * 2, float2bfloat(_sum2));
-                    vst1_u16(outptr0 + 4 * 3, float2bfloat(_sum3));
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose4x4_ps(_sum0, _sum1, _sum2, _sum3);
-
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + out_hstep, float2bfloat(_sum1));
-                    vst1_u16(outptr0 + out_hstep * 2, float2bfloat(_sum2));
-                    vst1_u16(outptr0 + out_hstep * 3, float2bfloat(_sum3));
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 4 * 2, _sum2);
-                vst1q_f32(outptr + 4 * 3, _sum3);
-            }
-
-            outptr += 16;
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = _sum0;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = _sum0;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = vld1q_f32(pC + 4);
-                        pC += 8;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = vdupq_n_f32(pC[1]);
-                        pC += 2;
-                    }
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA = bfloat2float(vld1_u16(pA));
-                float32x4_t _pB0 = bfloat2float(vdup_n_u16(pB[0]));
-                float32x4_t _pB1 = bfloat2float(vdup_n_u16(pB[1]));
-
-#if __aarch64__
-                _sum0 = vfmaq_f32(_sum0, _pA, _pB0);
-                _sum1 = vfmaq_f32(_sum1, _pA, _pB1);
-#else
-                _sum0 = vmlaq_f32(_sum0, _pA, _pB0);
-                _sum1 = vmlaq_f32(_sum1, _pA, _pB1);
-#endif
-
-                pA += 4;
-                pB += 2;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-                _sum1 = vmulq_f32(_sum1, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum1));
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short sum0[4];
-                    unsigned short sum1[4];
-                    vst1_u16(sum0, float2bfloat(_sum0));
-                    vst1_u16(sum1, float2bfloat(_sum1));
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[1] = sum1[0];
-                    outptr0[out_hstep + 1] = sum1[1];
-                    outptr0[out_hstep * 2 + 1] = sum1[2];
-                    outptr0[out_hstep * 3 + 1] = sum1[3];
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-            }
-
-            outptr += 8;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            float32x4_t _sum0;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        pC += 4;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        pC += 1;
-                    }
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pA = bfloat2float(vld1_u16(pA));
-                float32x4_t _pB = bfloat2float(vdup_n_u16(pB[0]));
-
-#if __aarch64__
-                _sum0 = vfmaq_f32(_sum0, _pA, _pB);
-#else
-                _sum0 = vmlaq_f32(_sum0, _pA, _pB);
-#endif
-
-                pA += 4;
-                pB += 1;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    outptr0 += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short sum0[4];
-                    vst1_u16(sum0, float2bfloat(_sum0));
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0++;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-            }
-
-            outptr += 4;
-        }
-
-        pAT += max_kk * 4;
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        unsigned short* outptr0 = (unsigned short*)top_blob + (i + ii) * out_hstep + j;
-
-        const unsigned short* pB = pBT;
-
-        if (pC)
-        {
-            if (broadcast_type_C == 1 || broadcast_type_C == 2)
-            {
-                pC = (const float*)CT_tile + i + ii;
-            }
-            if (broadcast_type_C == 4)
-            {
-                pC = (const float*)CT_tile + j;
-            }
-        }
-
-        int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum02;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum12;
-
-            if (k == 0)
-            {
-                _sum00 = vdupq_n_f32(0.f);
-                _sum01 = vdupq_n_f32(0.f);
-                _sum02 = vdupq_n_f32(0.f);
-                _sum10 = vdupq_n_f32(0.f);
-                _sum11 = vdupq_n_f32(0.f);
-                _sum12 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum02 = _sum00;
-                        _sum10 = _sum00;
-                        _sum11 = _sum00;
-                        _sum12 = _sum00;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum02 = _sum00;
-                        _sum10 = vdupq_n_f32(pC[1]);
-                        _sum11 = _sum10;
-                        _sum12 = _sum10;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        float32x4x2_t _tmp01 = vld2q_f32(pC);
-                        float32x4x2_t _tmp23 = vld2q_f32(pC + 8);
-                        float32x4x2_t _tmp45 = vld2q_f32(pC + 16);
-                        _sum00 = _tmp01.val[0];
-                        _sum01 = _tmp23.val[0];
-                        _sum02 = _tmp45.val[0];
-                        _sum10 = _tmp01.val[1];
-                        _sum11 = _tmp23.val[1];
-                        _sum12 = _tmp45.val[1];
-                        pC += 24;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                        _sum02 = vld1q_f32(pC + 8);
-                        _sum10 = _sum00;
-                        _sum11 = _sum01;
-                        _sum12 = _sum02;
-                        pC += 12;
-                    }
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01 = vld2q_f32(outptr);
-                float32x4x2_t _tmp23 = vld2q_f32(outptr + 8);
-                float32x4x2_t _tmp45 = vld2q_f32(outptr + 16);
-                _sum00 = _tmp01.val[0];
-                _sum01 = _tmp23.val[0];
-                _sum02 = _tmp45.val[0];
-                _sum10 = _tmp01.val[1];
-                _sum11 = _tmp23.val[1];
-                _sum12 = _tmp45.val[1];
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                uint16x8_t _pB = vld1q_u16(pB);
-                float32x4_t _pB0 = bfloat2float(vget_low_u16(_pB));
-                float32x4_t _pB1 = bfloat2float(vget_high_u16(_pB));
-                float32x4_t _pB2 = bfloat2float(vld1_u16(pB + 8));
-
-                float32x4_t _pA0 = bfloat2float(vdup_n_u16(pA[0]));
-                float32x4_t _pA1 = bfloat2float(vdup_n_u16(pA[1]));
-
-                _sum00 = vfmaq_f32(_sum00, _pB0, _pA0);
-                _sum01 = vfmaq_f32(_sum01, _pB1, _pA0);
-                _sum02 = vfmaq_f32(_sum02, _pB2, _pA0);
-                _sum10 = vfmaq_f32(_sum10, _pB0, _pA1);
-                _sum11 = vfmaq_f32(_sum11, _pB1, _pA1);
-                _sum12 = vfmaq_f32(_sum12, _pB2, _pA1);
-
-                pA += 2;
-                pB += 12;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum00 = vmulq_f32(_sum00, _alpha);
-                _sum01 = vmulq_f32(_sum01, _alpha);
-                _sum02 = vmulq_f32(_sum02, _alpha);
-                _sum10 = vmulq_f32(_sum10, _alpha);
-                _sum11 = vmulq_f32(_sum11, _alpha);
-                _sum12 = vmulq_f32(_sum12, _alpha);
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum01));
-                    vst1_u16(outptr0 + 8, float2bfloat(_sum02));
-                    vst1_u16(outptr0 + out_hstep, float2bfloat(_sum10));
-                    vst1_u16(outptr0 + out_hstep + 4, float2bfloat(_sum11));
-                    vst1_u16(outptr0 + out_hstep + 8, float2bfloat(_sum12));
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01;
-                _tmp01.val[0] = _sum00;
-                _tmp01.val[1] = _sum10;
-                float32x4x2_t _tmp23;
-                _tmp23.val[0] = _sum01;
-                _tmp23.val[1] = _sum11;
-                float32x4x2_t _tmp45;
-                _tmp45.val[0] = _sum02;
-                _tmp45.val[1] = _sum12;
-                vst2q_f32(outptr, _tmp01);
-                vst2q_f32(outptr + 8, _tmp23);
-                vst2q_f32(outptr + 16, _tmp45);
-            }
-
-            outptr += 24;
-        }
-#endif // __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-
-            if (k == 0)
-            {
-                _sum00 = vdupq_n_f32(0.f);
-                _sum01 = vdupq_n_f32(0.f);
-                _sum10 = vdupq_n_f32(0.f);
-                _sum11 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum10 = _sum00;
-                        _sum11 = _sum00;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum10 = vdupq_n_f32(pC[1]);
-                        _sum11 = _sum10;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        float32x4x2_t _tmp01 = vld2q_f32(pC);
-                        float32x4x2_t _tmp23 = vld2q_f32(pC + 8);
-                        _sum00 = _tmp01.val[0];
-                        _sum01 = _tmp23.val[0];
-                        _sum10 = _tmp01.val[1];
-                        _sum11 = _tmp23.val[1];
-                        pC += 16;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                        _sum10 = _sum00;
-                        _sum11 = _sum01;
-                        pC += 8;
-                    }
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01 = vld2q_f32(outptr);
-                float32x4x2_t _tmp23 = vld2q_f32(outptr + 8);
-                _sum00 = _tmp01.val[0];
-                _sum01 = _tmp23.val[0];
-                _sum10 = _tmp01.val[1];
-                _sum11 = _tmp23.val[1];
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                uint16x8_t _pB = vld1q_u16(pB);
-                float32x4_t _pB0 = bfloat2float(vget_low_u16(_pB));
-                float32x4_t _pB1 = bfloat2float(vget_high_u16(_pB));
-
-                float32x4_t _pA0 = bfloat2float(vdup_n_u16(pA[0]));
-                float32x4_t _pA1 = bfloat2float(vdup_n_u16(pA[1]));
-#if __aarch64__
-                _sum00 = vfmaq_f32(_sum00, _pB0, _pA0);
-                _sum01 = vfmaq_f32(_sum01, _pB1, _pA0);
-                _sum10 = vfmaq_f32(_sum10, _pB0, _pA1);
-                _sum11 = vfmaq_f32(_sum11, _pB1, _pA1);
-#else
-                _sum00 = vmlaq_f32(_sum00, _pB0, _pA0);
-                _sum01 = vmlaq_f32(_sum01, _pB1, _pA0);
-                _sum10 = vmlaq_f32(_sum10, _pB0, _pA1);
-                _sum11 = vmlaq_f32(_sum11, _pB1, _pA1);
-#endif
-
-                pA += 2;
-                pB += 8;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum00 = vmulq_f32(_sum00, _alpha);
-                _sum01 = vmulq_f32(_sum01, _alpha);
-                _sum10 = vmulq_f32(_sum10, _alpha);
-                _sum11 = vmulq_f32(_sum11, _alpha);
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum00));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum01));
-                    vst1_u16(outptr0 + out_hstep, float2bfloat(_sum10));
-                    vst1_u16(outptr0 + out_hstep + 4, float2bfloat(_sum11));
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01;
-                _tmp01.val[0] = _sum00;
-                _tmp01.val[1] = _sum10;
-                float32x4x2_t _tmp23;
-                _tmp23.val[0] = _sum01;
-                _tmp23.val[1] = _sum11;
-                vst2q_f32(outptr, _tmp01);
-                vst2q_f32(outptr + 8, _tmp23);
-            }
-
-            outptr += 16;
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = _sum0;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = vdupq_n_f32(pC[1]);
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        float32x4x2_t _tmp01 = vld2q_f32(pC);
-                        _sum0 = _tmp01.val[0];
-                        _sum1 = _tmp01.val[1];
-                        pC += 8;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = _sum0;
-                        pC += 4;
-                    }
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01 = vld2q_f32(outptr);
-                _sum0 = _tmp01.val[0];
-                _sum1 = _tmp01.val[1];
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pB = bfloat2float(vld1_u16(pB));
-
-                float32x4_t _pA0 = bfloat2float(vdup_n_u16(pA[0]));
-                float32x4_t _pA1 = bfloat2float(vdup_n_u16(pA[1]));
-#if __aarch64__
-                _sum0 = vfmaq_f32(_sum0, _pB, _pA0);
-                _sum1 = vfmaq_f32(_sum1, _pB, _pA1);
-#else
-                _sum0 = vmlaq_f32(_sum0, _pB, _pA0);
-                _sum1 = vmlaq_f32(_sum1, _pB, _pA1);
-#endif
-
-                pA += 2;
-                pB += 4;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-                _sum1 = vmulq_f32(_sum1, _alpha);
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + out_hstep, float2bfloat(_sum1));
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01;
-                _tmp01.val[0] = _sum0;
-                _tmp01.val[1] = _sum1;
-                vst2q_f32(outptr, _tmp01);
-            }
-
-            outptr += 8;
-        }
-#endif // __ARM_NEON
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            float sum00;
-            float sum01;
-            float sum10;
-            float sum11;
-
-            if (k == 0)
-            {
-                sum00 = 0.f;
-                sum01 = 0.f;
-                sum10 = 0.f;
-                sum11 = 0.f;
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        sum00 = pC[0];
-                        sum01 = pC[0];
-                        sum10 = pC[0];
-                        sum11 = pC[0];
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        sum00 = pC[0];
-                        sum01 = pC[1];
-                        sum10 = pC[0];
-                        sum11 = pC[1];
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        sum00 = pC[0];
-                        sum01 = pC[1];
-                        sum10 = pC[2];
-                        sum11 = pC[3];
-                        pC += 4;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        sum00 = pC[0];
-                        sum01 = pC[0];
-                        sum10 = pC[1];
-                        sum11 = pC[1];
-                        pC += 2;
-                    }
-                }
-            }
-            else
-            {
-                sum00 = outptr[0];
-                sum01 = outptr[1];
-                sum10 = outptr[2];
-                sum11 = outptr[3];
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-#if __ARM_NEON
-            // clang 15.0.1 on aarch64 auto vectorization produces wrong result on this loop
-            // we have to teach it a bit  :$   --- nihui
-            float32x4_t _sum0 = vdupq_n_f32(0.f);
-            float32x4_t _sum1 = vdupq_n_f32(0.f);
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                float32x4_t _pA0123 = bfloat2float(vld1_u16(pA));
-                float32x4_t _pB0123 = bfloat2float(vld1_u16(pB));
-
-                float32x4x2_t _pB0213 = vtrnq_f32(_pB0123, _pB0123);
-
-#if __aarch64__
-                _sum0 = vfmaq_f32(_sum0, _pA0123, _pB0213.val[0]);
-                _sum1 = vfmaq_f32(_sum1, _pA0123, _pB0213.val[1]);
-#else
-                _sum0 = vmlaq_f32(_sum0, _pA0123, _pB0213.val[0]);
-                _sum1 = vmlaq_f32(_sum1, _pA0123, _pB0213.val[1]);
-#endif
-
-                pA += 4;
-                pB += 4;
-            }
-            sum00 += vgetq_lane_f32(_sum0, 0) + vgetq_lane_f32(_sum0, 2);
-            sum01 += vgetq_lane_f32(_sum0, 1) + vgetq_lane_f32(_sum0, 3);
-            sum10 += vgetq_lane_f32(_sum1, 0) + vgetq_lane_f32(_sum1, 2);
-            sum11 += vgetq_lane_f32(_sum1, 1) + vgetq_lane_f32(_sum1, 3);
-#endif // __ARM_NEON
-            for (; kk < max_kk; kk += 1)
-            {
-                float pA0 = bfloat16_to_float32(pA[0]);
-                float pA1 = bfloat16_to_float32(pA[1]);
-                float pB0 = bfloat16_to_float32(pB[0]);
-                float pB1 = bfloat16_to_float32(pB[1]);
-
-                sum00 += pA0 * pB0;
-                sum01 += pA1 * pB0;
-                sum10 += pA0 * pB1;
-                sum11 += pA1 * pB1;
-
-                pA += 2;
-                pB += 2;
-            }
-
-            if (alpha != 1.f)
-            {
-                sum00 *= alpha;
-                sum01 *= alpha;
-                sum10 *= alpha;
-                sum11 *= alpha;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = float32_to_bfloat16(sum00);
-                    outptr0[1] = float32_to_bfloat16(sum10);
-                    outptr0[out_hstep] = float32_to_bfloat16(sum01);
-                    outptr0[out_hstep + 1] = float32_to_bfloat16(sum11);
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                outptr[0] = sum00;
-                outptr[1] = sum01;
-                outptr[2] = sum10;
-                outptr[3] = sum11;
-            }
-
-            outptr += 4;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            float sum0;
-            float sum1;
-
-            if (k == 0)
-            {
-                sum0 = 0.f;
-                sum1 = 0.f;
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        sum0 = pC[0];
-                        sum1 = pC[0];
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        sum0 = pC[0];
-                        sum1 = pC[1];
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        sum0 = pC[0];
-                        sum1 = pC[1];
-                        pC += 2;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        sum0 = pC[0];
-                        sum1 = pC[0];
-                        pC += 1;
-                    }
-                }
-            }
-            else
-            {
-                sum0 = outptr[0];
-                sum1 = outptr[1];
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float pA0 = bfloat16_to_float32(pA[0]);
-                float pA1 = bfloat16_to_float32(pA[1]);
-                float pB0 = bfloat16_to_float32(pB[0]);
-
-                sum0 += pA0 * pB0;
-                sum1 += pA1 * pB0;
-                pA += 2;
-                pB += 1;
-            }
-
-            if (alpha != 1.f)
-            {
-                sum0 *= alpha;
-                sum1 *= alpha;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = float32_to_bfloat16(sum0);
-                    outptr0[out_hstep] = float32_to_bfloat16(sum1);
-                    outptr0++;
-                }
-            }
-            else
-            {
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-            }
-
-            outptr += 2;
-        }
-
-        pAT += max_kk * 2;
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        unsigned short* outptr0 = (unsigned short*)top_blob + (i + ii) * out_hstep + j;
-
-        const unsigned short* pB = pBT;
-
-        if (pC)
-        {
-            if (broadcast_type_C == 1 || broadcast_type_C == 2)
-            {
-                pC = (const float*)CT_tile + i + ii;
-            }
-            if (broadcast_type_C == 4)
-            {
-                pC = (const float*)CT_tile + j;
-            }
-        }
-
-        int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-                _sum2 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0 || broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = vdupq_n_f32(pC[0]);
-                        _sum2 = vdupq_n_f32(pC[0]);
-                    }
-                    if (broadcast_type_C == 3 || broadcast_type_C == 4)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = vld1q_f32(pC + 4);
-                        _sum2 = vld1q_f32(pC + 8);
-                        pC += 12;
-                    }
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4);
-                _sum2 = vld1q_f32(outptr + 8);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                uint16x8_t _pB = vld1q_u16(pB);
-                float32x4_t _pB0 = bfloat2float(vget_low_u16(_pB));
-                float32x4_t _pB1 = bfloat2float(vget_high_u16(_pB));
-                float32x4_t _pB2 = bfloat2float(vld1_u16(pB + 8));
-
-                float32x4_t _pA0 = bfloat2float(vdup_n_u16(pA[0]));
-
-                _sum0 = vfmaq_f32(_sum0, _pA0, _pB0);
-                _sum1 = vfmaq_f32(_sum1, _pA0, _pB1);
-                _sum2 = vfmaq_f32(_sum2, _pA0, _pB2);
-
-                pA += 1;
-                pB += 12;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-                _sum1 = vmulq_f32(_sum1, _alpha);
-                _sum2 = vmulq_f32(_sum2, _alpha);
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum1));
-                    vst1_u16(outptr0 + 8, float2bfloat(_sum2));
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 8, _sum2);
-            }
-
-            outptr += 12;
-        }
-#endif // __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0 || broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = vdupq_n_f32(pC[0]);
-                    }
-                    if (broadcast_type_C == 3 || broadcast_type_C == 4)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = vld1q_f32(pC + 4);
-                        pC += 8;
-                    }
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                uint16x8_t _pB = vld1q_u16(pB);
-                float32x4_t _pB0 = bfloat2float(vget_low_u16(_pB));
-                float32x4_t _pB1 = bfloat2float(vget_high_u16(_pB));
-
-                float32x4_t _pA0 = bfloat2float(vdup_n_u16(pA[0]));
-#if __aarch64__
-                _sum0 = vfmaq_f32(_sum0, _pA0, _pB0);
-                _sum1 = vfmaq_f32(_sum1, _pA0, _pB1);
-#else
-                _sum0 = vmlaq_f32(_sum0, _pA0, _pB0);
-                _sum1 = vmlaq_f32(_sum1, _pA0, _pB1);
-#endif
-
-                pA += 1;
-                pB += 8;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-                _sum1 = vmulq_f32(_sum1, _alpha);
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum0));
-                    vst1_u16(outptr0 + 4, float2bfloat(_sum1));
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-            }
-
-            outptr += 8;
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            float32x4_t _sum;
-
-            if (k == 0)
-            {
-                _sum = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0 || broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum = vdupq_n_f32(pC[0]);
-                    }
-                    if (broadcast_type_C == 3 || broadcast_type_C == 4)
-                    {
-                        _sum = vld1q_f32(pC);
-                        pC += 4;
-                    }
-                }
-            }
-            else
-            {
-                _sum = vld1q_f32(outptr);
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float32x4_t _pB = bfloat2float(vld1_u16(pB));
-                float32x4_t _pA = bfloat2float(vdup_n_u16(pA[0]));
-#if __aarch64__
-                _sum = vfmaq_f32(_sum, _pA, _pB);
-#else
-                _sum = vmlaq_f32(_sum, _pA, _pB);
-#endif
-
-                pA += 1;
-                pB += 4;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum = vmulq_f32(_sum, _alpha);
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, float2bfloat(_sum));
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum);
-            }
-
-            outptr += 4;
-        }
-#endif // __ARM_NEON
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            float sum0;
-            float sum1;
-
-            if (k == 0)
-            {
-                sum0 = 0.f;
-                sum1 = 0.f;
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0 || broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        sum0 = pC[0];
-                        sum1 = pC[0];
-                    }
-                    if (broadcast_type_C == 3 || broadcast_type_C == 4)
-                    {
-                        sum0 = pC[0];
-                        sum1 = pC[1];
-                        pC += 2;
-                    }
-                }
-            }
-            else
-            {
-                sum0 = outptr[0];
-                sum1 = outptr[1];
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float pA0 = bfloat16_to_float32(pA[0]);
-                float pB0 = bfloat16_to_float32(pB[0]);
-                float pB1 = bfloat16_to_float32(pB[1]);
-
-                sum0 += pA0 * pB0;
-                sum1 += pA0 * pB1;
-
-                pA += 1;
-                pB += 2;
-            }
-
-            if (alpha != 1.f)
-            {
-                sum0 *= alpha;
-                sum1 *= alpha;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = float32_to_bfloat16(sum0);
-                    outptr0[1] = float32_to_bfloat16(sum1);
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-            }
-
-            outptr += 2;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            float sum;
-
-            if (k == 0)
-            {
-                sum = 0.f;
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0 || broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        sum = pC[0];
-                    }
-                    if (broadcast_type_C == 3 || broadcast_type_C == 4)
-                    {
-                        sum = pC[0];
-                        pC += 1;
-                    }
-                }
-            }
-            else
-            {
-                sum = outptr[0];
-            }
-
-            const unsigned short* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                float pA0 = bfloat16_to_float32(pA[0]);
-                float pB0 = bfloat16_to_float32(pB[0]);
-
-                sum += pA0 * pB0;
-                pA += 1;
-                pB += 1;
-            }
-
-            if (alpha != 1.f)
-            {
-                sum *= alpha;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = float32_to_bfloat16(sum);
-                    outptr0++;
-                }
-            }
-            else
-            {
-                outptr[0] = sum;
-            }
-
-            outptr += 1;
-        }
-
-        pAT += max_kk;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_bf16s_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_bf16s_fp16s.h
deleted file mode 100644
index 7f93d003..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_bf16s_fp16s.h
+++ /dev/null
@@ -1,1584 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pack_A_tile_bf16_fp16(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
-{
-    const int elempack = A.elempack;
-    const int A_hstep = A.dims == 3 ? (int)A.cstep : A.w;
-
-    unsigned short* pp = AT;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + (i + ii) * A_hstep + k * 8;
-
-            for (int kk = 0; kk < max_kk; kk++)
-            {
-                vst1q_u16(pp, vld1q_u16(p0));
-                pp += 8;
-                p0 += 8;
-            }
-        }
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + (i + ii) * A_hstep + k * 4;
-            const unsigned short* p1 = (const unsigned short*)A + (i + ii + 4) * A_hstep + k * 4;
-
-            for (int kk = 0; kk < max_kk; kk++)
-            {
-                uint16x8_t _r0 = vcombine_u16(vld1_u16(p0), vld1_u16(p1));
-                vst1q_u16(pp, _r0);
-                pp += 8;
-                p0 += 4;
-                p1 += 4;
-            }
-        }
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + (i + ii) * A_hstep + k;
-            const unsigned short* p1 = (const unsigned short*)A + (i + ii + 1) * A_hstep + k;
-            const unsigned short* p2 = (const unsigned short*)A + (i + ii + 2) * A_hstep + k;
-            const unsigned short* p3 = (const unsigned short*)A + (i + ii + 3) * A_hstep + k;
-            const unsigned short* p4 = (const unsigned short*)A + (i + ii + 4) * A_hstep + k;
-            const unsigned short* p5 = (const unsigned short*)A + (i + ii + 5) * A_hstep + k;
-            const unsigned short* p6 = (const unsigned short*)A + (i + ii + 6) * A_hstep + k;
-            const unsigned short* p7 = (const unsigned short*)A + (i + ii + 7) * A_hstep + k;
-
-            int kk = 0;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                uint16x8_t _r0 = vld1q_u16(p0);
-                uint16x8_t _r1 = vld1q_u16(p1);
-                uint16x8_t _r2 = vld1q_u16(p2);
-                uint16x8_t _r3 = vld1q_u16(p3);
-                uint16x8_t _r4 = vld1q_u16(p4);
-                uint16x8_t _r5 = vld1q_u16(p5);
-                uint16x8_t _r6 = vld1q_u16(p6);
-                uint16x8_t _r7 = vld1q_u16(p7);
-                transpose8x8_u16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-                vst1q_u16(pp, _r0);
-                vst1q_u16(pp + 8, _r1);
-                vst1q_u16(pp + 8 * 2, _r2);
-                vst1q_u16(pp + 8 * 3, _r3);
-                vst1q_u16(pp + 8 * 4, _r4);
-                vst1q_u16(pp + 8 * 5, _r5);
-                vst1q_u16(pp + 8 * 6, _r6);
-                vst1q_u16(pp + 8 * 7, _r7);
-                pp += 64;
-                p0 += 8;
-                p1 += 8;
-                p2 += 8;
-                p3 += 8;
-                p4 += 8;
-                p5 += 8;
-                p6 += 8;
-                p7 += 8;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p1[0];
-                pp[2] = p2[0];
-                pp[3] = p3[0];
-                pp[4] = p4[0];
-                pp[5] = p5[0];
-                pp[6] = p6[0];
-                pp[7] = p7[0];
-                pp += 8;
-                p0++;
-                p1++;
-                p2++;
-                p3++;
-                p4++;
-                p5++;
-                p6++;
-                p7++;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + (i + ii) * A_hstep + k * 4;
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                vst1q_u16(pp, vld1q_u16(p0));
-                pp += 8;
-                p0 += 8;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                vst1_u16(pp, vld1_u16(p0));
-                pp += 4;
-                p0 += 4;
-            }
-        }
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + (i + ii) * A_hstep + k;
-            const unsigned short* p1 = (const unsigned short*)A + (i + ii + 1) * A_hstep + k;
-            const unsigned short* p2 = (const unsigned short*)A + (i + ii + 2) * A_hstep + k;
-            const unsigned short* p3 = (const unsigned short*)A + (i + ii + 3) * A_hstep + k;
-
-            int kk = 0;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                uint16x8x4_t _r0123;
-                _r0123.val[0] = vld1q_u16(p0);
-                _r0123.val[1] = vld1q_u16(p1);
-                _r0123.val[2] = vld1q_u16(p2);
-                _r0123.val[3] = vld1q_u16(p3);
-                vst4q_u16(pp, _r0123);
-                pp += 32;
-                p0 += 8;
-                p1 += 8;
-                p2 += 8;
-                p3 += 8;
-            }
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                uint16x4x4_t _r0123;
-                _r0123.val[0] = vld1_u16(p0);
-                _r0123.val[1] = vld1_u16(p1);
-                _r0123.val[2] = vld1_u16(p2);
-                _r0123.val[3] = vld1_u16(p3);
-                vst4_u16(pp, _r0123);
-                pp += 16;
-                p0 += 4;
-                p1 += 4;
-                p2 += 4;
-                p3 += 4;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p1[0];
-                pp[2] = p2[0];
-                pp[3] = p3[0];
-                pp += 4;
-                p0++;
-                p1++;
-                p2++;
-                p3++;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        // if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + (i + ii) * A_hstep + k;
-            const unsigned short* p1 = (const unsigned short*)A + (i + ii + 1) * A_hstep + k;
-
-            int kk = 0;
-#if __ARM_NEON
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                uint16x8x2_t _r01;
-                _r01.val[0] = vld1q_u16(p0);
-                _r01.val[1] = vld1q_u16(p1);
-                vst2q_u16(pp, _r01);
-                pp += 16;
-                p0 += 8;
-                p1 += 8;
-            }
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                uint16x4x2_t _r01;
-                _r01.val[0] = vld1_u16(p0);
-                _r01.val[1] = vld1_u16(p1);
-                vst2_u16(pp, _r01);
-                pp += 8;
-                p0 += 4;
-                p1 += 4;
-            }
-#endif // __ARM_NEON
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p1[0];
-                pp += 2;
-                p0++;
-                p1++;
-            }
-        }
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        // if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + (i + ii) * A_hstep + k;
-
-            int kk = 0;
-#if __ARM_NEON
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                vst1q_u16(pp, vld1q_u16(p0));
-                pp += 8;
-                p0 += 8;
-            }
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                vst1_u16(pp, vld1_u16(p0));
-                pp += 4;
-                p0 += 4;
-            }
-#endif // __ARM_NEON
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = (unsigned short)p0[0];
-                pp += 1;
-                p0++;
-            }
-        }
-    }
-}
-
-static void transpose_pack_A_tile_bf16_fp16(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
-{
-    const int elempack = A.elempack;
-    const int A_hstep = A.dims == 3 ? (int)A.cstep : A.w;
-
-    unsigned short* pp = AT;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + k * A_hstep + (i + ii) * 8;
-
-            int kk = 0;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                uint16x8x4_t _r0123 = vld4q_u16(p0);
-                uint16x8x4_t _r4567 = vld4q_u16(p0 + 32);
-                uint16x8x2_t _r04 = vuzpq_u16(_r0123.val[0], _r4567.val[0]);
-                uint16x8x2_t _r15 = vuzpq_u16(_r0123.val[1], _r4567.val[1]);
-                uint16x8x2_t _r26 = vuzpq_u16(_r0123.val[2], _r4567.val[2]);
-                uint16x8x2_t _r37 = vuzpq_u16(_r0123.val[3], _r4567.val[3]);
-                vst1q_u16(pp, _r04.val[0]);
-                vst1q_u16(pp + 8, _r15.val[0]);
-                vst1q_u16(pp + 16, _r26.val[0]);
-                vst1q_u16(pp + 24, _r37.val[0]);
-                vst1q_u16(pp + 32, _r04.val[1]);
-                vst1q_u16(pp + 40, _r15.val[1]);
-                vst1q_u16(pp + 48, _r26.val[1]);
-                vst1q_u16(pp + 56, _r37.val[1]);
-                pp += 64;
-                p0 += A_hstep * 8;
-            }
-        }
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + k * A_hstep + (i + ii) * 4;
-
-            int kk = 0;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                uint16x8x4_t _r0123 = vld4q_u16(p0);
-                vst1q_u16(pp, _r0123.val[0]);
-                vst1q_u16(pp + 8, _r0123.val[1]);
-                vst1q_u16(pp + 16, _r0123.val[2]);
-                vst1q_u16(pp + 24, _r0123.val[3]);
-                pp += 32;
-                p0 += A_hstep * 4;
-            }
-        }
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + k * A_hstep + (i + ii);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                vst1q_u16(pp, vld1q_u16(p0));
-                pp += 8;
-                p0 += A_hstep;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + k * A_hstep + (i + ii) * 8;
-
-            int kk = 0;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                uint16x8x4_t _r0123;
-                _r0123.val[0] = vld1q_u16(p0);
-                _r0123.val[1] = vld1q_u16(p0 + 8);
-                _r0123.val[2] = vld1q_u16(p0 + 16);
-                _r0123.val[3] = vld1q_u16(p0 + 24);
-                vst4q_u16(pp, _r0123);
-                pp += 32;
-                p0 += A_hstep * 8;
-            }
-        }
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + k * A_hstep + (i + ii) * 4;
-
-            int kk = 0;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                uint16x4x4_t _r0123 = vld4_u16(p0);
-                vst1q_u16(pp, vcombine_u16(_r0123.val[0], _r0123.val[1]));
-                vst1q_u16(pp + 8, vcombine_u16(_r0123.val[2], _r0123.val[3]));
-                pp += 16;
-                p0 += A_hstep * 4;
-            }
-        }
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + k * A_hstep + (i + ii);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                vst1_u16(pp, vld1_u16(p0));
-                pp += 4;
-                p0 += A_hstep;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-#if __ARM_NEON
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + k * A_hstep + (i + ii) * 8;
-
-            int kk = 0;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                uint16x8x2_t _r01;
-                _r01.val[0] = vld1q_u16(p0);
-                _r01.val[1] = vld1q_u16(p0 + 8);
-                vst2q_u16(pp, _r01);
-                pp += 16;
-                p0 += A_hstep * 8;
-            }
-        }
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + k * A_hstep + (i + ii) * 4;
-
-            int kk = 0;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                uint16x4x2_t _r01;
-                _r01.val[0] = vld1_u16(p0);
-                _r01.val[1] = vld1_u16(p0 + 4);
-                vst2_u16(pp, _r01);
-                pp += 8;
-                p0 += A_hstep * 4;
-            }
-        }
-#endif // __ARM_NEON
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + k * A_hstep + (i + ii);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                pp += 2;
-                p0 += A_hstep;
-            }
-        }
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-#if __ARM_NEON
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + k * A_hstep + (i + ii) * 8;
-
-            int kk = 0;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                vst1q_u16(pp, vld1q_u16(p0));
-                pp += 8;
-                p0 += A_hstep * 8;
-            }
-        }
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + k * A_hstep + (i + ii) * 4;
-
-            int kk = 0;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                vst1_u16(pp, vld1_u16(p0));
-                pp += 4;
-                p0 += A_hstep * 4;
-            }
-        }
-#endif // __ARM_NEON
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)A + k * A_hstep + (i + ii);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp += 1;
-                p0 += A_hstep;
-            }
-        }
-    }
-}
-
-static void pack_B_tile_bf16_fp16(const Mat& B, Mat& BT, int j, int max_jj, int k, int max_kk)
-{
-    const int elempack = B.elempack;
-    const int B_hstep = B.dims == 3 ? (int)B.cstep : B.w;
-
-    unsigned short* pp = BT;
-
-    int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; jj + 11 < max_jj; jj += 12)
-    {
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + (j + jj) / 8 * 8 * B_hstep + k * 8;
-            const unsigned short* p1 = (const unsigned short*)B + (j + jj + 8) / 8 * 8 * B_hstep + k * 8;
-
-            if ((j + jj) % 8 == 0)
-            {
-                for (int kk = 0; kk < max_kk; kk++)
-                {
-                    vst1q_u16(pp, vld1q_u16(p0));
-                    vst1_u16(pp + 8, vld1_u16(p1));
-                    pp += 12;
-                    p0 += 8;
-                    p1 += 8;
-                }
-            }
-            if ((j + jj) % 8 == 4)
-            {
-                for (int kk = 0; kk < max_kk; kk++)
-                {
-                    vst1_u16(pp, vld1_u16(p0 + 4));
-                    vst1q_u16(pp + 4, vld1q_u16(p1));
-                    pp += 12;
-                    p0 += 8;
-                    p1 += 8;
-                }
-            }
-        }
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + (j + jj) * B_hstep + k * 4;
-            const unsigned short* p1 = (const unsigned short*)B + (j + jj + 4) * B_hstep + k * 4;
-            const unsigned short* p2 = (const unsigned short*)B + (j + jj + 8) * B_hstep + k * 4;
-
-            for (int kk = 0; kk < max_kk; kk++)
-            {
-                vst1_u16(pp, vld1_u16(p0));
-                vst1_u16(pp + 4, vld1_u16(p1));
-                vst1_u16(pp + 8, vld1_u16(p2));
-                pp += 12;
-                p0 += 4;
-                p1 += 4;
-                p2 += 4;
-            }
-        }
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + (j + jj) * B_hstep + k;
-            const unsigned short* p1 = (const unsigned short*)B + (j + jj + 1) * B_hstep + k;
-            const unsigned short* p2 = (const unsigned short*)B + (j + jj + 2) * B_hstep + k;
-            const unsigned short* p3 = (const unsigned short*)B + (j + jj + 3) * B_hstep + k;
-            const unsigned short* p4 = (const unsigned short*)B + (j + jj + 4) * B_hstep + k;
-            const unsigned short* p5 = (const unsigned short*)B + (j + jj + 5) * B_hstep + k;
-            const unsigned short* p6 = (const unsigned short*)B + (j + jj + 6) * B_hstep + k;
-            const unsigned short* p7 = (const unsigned short*)B + (j + jj + 7) * B_hstep + k;
-            const unsigned short* p8 = (const unsigned short*)B + (j + jj + 8) * B_hstep + k;
-            const unsigned short* p9 = (const unsigned short*)B + (j + jj + 9) * B_hstep + k;
-            const unsigned short* pa = (const unsigned short*)B + (j + jj + 10) * B_hstep + k;
-            const unsigned short* pb = (const unsigned short*)B + (j + jj + 11) * B_hstep + k;
-
-            int kk = 0;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                uint16x4_t _r0 = vld1_u16(p0);
-                uint16x4_t _r1 = vld1_u16(p1);
-                uint16x4_t _r2 = vld1_u16(p2);
-                uint16x4_t _r3 = vld1_u16(p3);
-                uint16x4_t _r4 = vld1_u16(p4);
-                uint16x4_t _r5 = vld1_u16(p5);
-                uint16x4_t _r6 = vld1_u16(p6);
-                uint16x4_t _r7 = vld1_u16(p7);
-                uint16x4_t _r8 = vld1_u16(p8);
-                uint16x4_t _r9 = vld1_u16(p9);
-                uint16x4_t _ra = vld1_u16(pa);
-                uint16x4_t _rb = vld1_u16(pb);
-
-                transpose4x4_u16(_r0, _r1, _r2, _r3);
-                transpose4x4_u16(_r4, _r5, _r6, _r7);
-                transpose4x4_u16(_r8, _r9, _ra, _rb);
-
-                vst1_u16(pp, _r0);
-                vst1_u16(pp + 4, _r4);
-                vst1_u16(pp + 4 * 2, _r8);
-                vst1_u16(pp + 4 * 3, _r1);
-                vst1_u16(pp + 4 * 4, _r5);
-                vst1_u16(pp + 4 * 5, _r9);
-                vst1_u16(pp + 4 * 6, _r2);
-                vst1_u16(pp + 4 * 7, _r6);
-                vst1_u16(pp + 4 * 8, _ra);
-                vst1_u16(pp + 4 * 9, _r3);
-                vst1_u16(pp + 4 * 10, _r7);
-                vst1_u16(pp + 4 * 11, _rb);
-                pp += 48;
-                p0 += 4;
-                p1 += 4;
-                p2 += 4;
-                p3 += 4;
-                p4 += 4;
-                p5 += 4;
-                p6 += 4;
-                p7 += 4;
-                p8 += 4;
-                p9 += 4;
-                pa += 4;
-                pb += 4;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p1[0];
-                pp[2] = p2[0];
-                pp[3] = p3[0];
-                pp[4] = p4[0];
-                pp[5] = p5[0];
-                pp[6] = p6[0];
-                pp[7] = p7[0];
-                pp[8] = p8[0];
-                pp[9] = p9[0];
-                pp[10] = pa[0];
-                pp[11] = pb[0];
-                pp += 12;
-                p0++;
-                p1++;
-                p2++;
-                p3++;
-                p4++;
-                p5++;
-                p6++;
-                p7++;
-                p8++;
-                p9++;
-                pa++;
-                pb++;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + (j + jj) / 8 * 8 * B_hstep + k * 8;
-            const unsigned short* p1 = (const unsigned short*)B + (j + jj + 8) / 8 * 8 * B_hstep + k * 8;
-
-            if ((j + jj) % 8 == 0)
-            {
-                for (int kk = 0; kk < max_kk; kk++)
-                {
-                    vst1q_u16(pp, vld1q_u16(p0));
-                    pp += 8;
-                    p0 += 8;
-                }
-            }
-            if ((j + jj) % 8 == 4)
-            {
-                for (int kk = 0; kk < max_kk; kk++)
-                {
-                    vst1q_u16(pp, vcombine_u16(vld1_u16(p0 + 4), vld1_u16(p1)));
-                    pp += 8;
-                    p0 += 8;
-                    p1 += 8;
-                }
-            }
-        }
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + (j + jj) * B_hstep + k * 4;
-            const unsigned short* p1 = (const unsigned short*)B + (j + jj + 4) * B_hstep + k * 4;
-
-            for (int kk = 0; kk < max_kk; kk++)
-            {
-                uint16x8_t _r0 = vcombine_u16(vld1_u16(p0), vld1_u16(p1));
-                vst1q_u16(pp, _r0);
-                pp += 8;
-                p0 += 4;
-                p1 += 4;
-            }
-        }
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + (j + jj) * B_hstep + k;
-            const unsigned short* p1 = (const unsigned short*)B + (j + jj + 1) * B_hstep + k;
-            const unsigned short* p2 = (const unsigned short*)B + (j + jj + 2) * B_hstep + k;
-            const unsigned short* p3 = (const unsigned short*)B + (j + jj + 3) * B_hstep + k;
-            const unsigned short* p4 = (const unsigned short*)B + (j + jj + 4) * B_hstep + k;
-            const unsigned short* p5 = (const unsigned short*)B + (j + jj + 5) * B_hstep + k;
-            const unsigned short* p6 = (const unsigned short*)B + (j + jj + 6) * B_hstep + k;
-            const unsigned short* p7 = (const unsigned short*)B + (j + jj + 7) * B_hstep + k;
-
-            int kk = 0;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                uint16x8_t _r0 = vld1q_u16(p0);
-                uint16x8_t _r1 = vld1q_u16(p1);
-                uint16x8_t _r2 = vld1q_u16(p2);
-                uint16x8_t _r3 = vld1q_u16(p3);
-                uint16x8_t _r4 = vld1q_u16(p4);
-                uint16x8_t _r5 = vld1q_u16(p5);
-                uint16x8_t _r6 = vld1q_u16(p6);
-                uint16x8_t _r7 = vld1q_u16(p7);
-                transpose8x8_u16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-                vst1q_u16(pp, _r0);
-                vst1q_u16(pp + 8, _r1);
-                vst1q_u16(pp + 8 * 2, _r2);
-                vst1q_u16(pp + 8 * 3, _r3);
-                vst1q_u16(pp + 8 * 4, _r4);
-                vst1q_u16(pp + 8 * 5, _r5);
-                vst1q_u16(pp + 8 * 6, _r6);
-                vst1q_u16(pp + 8 * 7, _r7);
-                pp += 64;
-                p0 += 8;
-                p1 += 8;
-                p2 += 8;
-                p3 += 8;
-                p4 += 8;
-                p5 += 8;
-                p6 += 8;
-                p7 += 8;
-            }
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                uint16x4_t _r0 = vld1_u16(p0);
-                uint16x4_t _r1 = vld1_u16(p1);
-                uint16x4_t _r2 = vld1_u16(p2);
-                uint16x4_t _r3 = vld1_u16(p3);
-                uint16x4_t _r4 = vld1_u16(p4);
-                uint16x4_t _r5 = vld1_u16(p5);
-                uint16x4_t _r6 = vld1_u16(p6);
-                uint16x4_t _r7 = vld1_u16(p7);
-
-                transpose4x4_u16(_r0, _r1, _r2, _r3);
-                transpose4x4_u16(_r4, _r5, _r6, _r7);
-
-                vst1_u16(pp, _r0);
-                vst1_u16(pp + 4, _r4);
-                vst1_u16(pp + 4 * 2, _r1);
-                vst1_u16(pp + 4 * 3, _r5);
-                vst1_u16(pp + 4 * 4, _r2);
-                vst1_u16(pp + 4 * 5, _r6);
-                vst1_u16(pp + 4 * 6, _r3);
-                vst1_u16(pp + 4 * 7, _r7);
-                pp += 32;
-                p0 += 4;
-                p1 += 4;
-                p2 += 4;
-                p3 += 4;
-                p4 += 4;
-                p5 += 4;
-                p6 += 4;
-                p7 += 4;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p1[0];
-                pp[2] = p2[0];
-                pp[3] = p3[0];
-                pp[4] = p4[0];
-                pp[5] = p5[0];
-                pp[6] = p6[0];
-                pp[7] = p7[0];
-                pp += 8;
-                p0++;
-                p1++;
-                p2++;
-                p3++;
-                p4++;
-                p5++;
-                p6++;
-                p7++;
-            }
-        }
-    }
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + (j + jj) / 8 * 8 * B_hstep + k * 8;
-
-            if ((j + jj) % 8 == 0)
-            {
-                for (int kk = 0; kk < max_kk; kk++)
-                {
-                    vst1_u16(pp, vld1_u16(p0));
-                    pp += 4;
-                    p0 += 8;
-                }
-            }
-            if ((j + jj) % 8 == 4)
-            {
-                for (int kk = 0; kk < max_kk; kk++)
-                {
-                    vst1_u16(pp, vld1_u16(p0 + 4));
-                    pp += 4;
-                    p0 += 8;
-                }
-            }
-        }
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + (j + jj) * B_hstep + k * 4;
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                vst1q_u16(pp, vld1q_u16(p0));
-                pp += 8;
-                p0 += 8;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                vst1_u16(pp, vld1_u16(p0));
-                pp += 4;
-                p0 += 4;
-            }
-        }
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + (j + jj) * B_hstep + k;
-            const unsigned short* p1 = (const unsigned short*)B + (j + jj + 1) * B_hstep + k;
-            const unsigned short* p2 = (const unsigned short*)B + (j + jj + 2) * B_hstep + k;
-            const unsigned short* p3 = (const unsigned short*)B + (j + jj + 3) * B_hstep + k;
-
-            int kk = 0;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                uint16x8x4_t _r0123;
-                _r0123.val[0] = vld1q_u16(p0);
-                _r0123.val[1] = vld1q_u16(p1);
-                _r0123.val[2] = vld1q_u16(p2);
-                _r0123.val[3] = vld1q_u16(p3);
-                vst4q_u16(pp, _r0123);
-                pp += 32;
-                p0 += 8;
-                p1 += 8;
-                p2 += 8;
-                p3 += 8;
-            }
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                uint16x4x4_t _r0123;
-                _r0123.val[0] = vld1_u16(p0);
-                _r0123.val[1] = vld1_u16(p1);
-                _r0123.val[2] = vld1_u16(p2);
-                _r0123.val[3] = vld1_u16(p3);
-                vst4_u16(pp, _r0123);
-                pp += 16;
-                p0 += 4;
-                p1 += 4;
-                p2 += 4;
-                p3 += 4;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p1[0];
-                pp[2] = p2[0];
-                pp[3] = p3[0];
-                pp += 4;
-                p0++;
-                p1++;
-                p2++;
-                p3++;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-        // if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + (j + jj) * B_hstep + k;
-            const unsigned short* p1 = (const unsigned short*)B + (j + jj + 1) * B_hstep + k;
-
-            int kk = 0;
-#if __ARM_NEON
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                uint16x8x2_t _r01;
-                _r01.val[0] = vld1q_u16(p0);
-                _r01.val[1] = vld1q_u16(p1);
-                vst2q_u16(pp, _r01);
-                pp += 16;
-                p0 += 8;
-                p1 += 8;
-            }
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                uint16x4x2_t _r01;
-                _r01.val[0] = vld1_u16(p0);
-                _r01.val[1] = vld1_u16(p1);
-                vst2_u16(pp, _r01);
-                pp += 8;
-                p0 += 4;
-                p1 += 4;
-            }
-#endif // __ARM_NEON
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p1[0];
-                pp += 2;
-                p0++;
-                p1++;
-            }
-        }
-    }
-    for (; jj < max_jj; jj += 1)
-    {
-        // if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + (j + jj) * B_hstep + k;
-
-            int kk = 0;
-#if __ARM_NEON
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                vst1q_u16(pp, vld1q_u16(p0));
-                pp += 8;
-                p0 += 8;
-            }
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                vst1_u16(pp, vld1_u16(p0));
-                pp += 4;
-                p0 += 4;
-            }
-#endif // __ARM_NEON
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp += 1;
-                p0++;
-            }
-        }
-    }
-}
-
-static void transpose_pack_B_tile_bf16_fp16(const Mat& B, Mat& BT, int j, int max_jj, int k, int max_kk)
-{
-    const int elempack = B.elempack;
-    const int B_hstep = B.dims == 3 ? (int)B.cstep : B.w;
-
-    unsigned short* pp = BT;
-
-    int jj = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; jj + 11 < max_jj; jj += 12)
-    {
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + k * B_hstep + (j + jj) * 8;
-
-            int kk = 0;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                uint16x8x4_t _r0123 = vld4q_u16(p0);
-                uint16x8x4_t _r4567 = vld4q_u16(p0 + 32);
-                uint16x8x4_t _r89ab = vld4q_u16(p0 + 64);
-                uint16x8x2_t _r04 = vuzpq_u16(_r0123.val[0], _r4567.val[0]);
-                uint16x8x2_t _r15 = vuzpq_u16(_r0123.val[1], _r4567.val[1]);
-                uint16x8x2_t _r26 = vuzpq_u16(_r0123.val[2], _r4567.val[2]);
-                uint16x8x2_t _r37 = vuzpq_u16(_r0123.val[3], _r4567.val[3]);
-                uint16x4x2_t _r04_1 = vuzp_u16(vget_low_u16(_r89ab.val[0]), vget_high_u16(_r89ab.val[0]));
-                uint16x4x2_t _r15_1 = vuzp_u16(vget_low_u16(_r89ab.val[1]), vget_high_u16(_r89ab.val[1]));
-                uint16x4x2_t _r26_1 = vuzp_u16(vget_low_u16(_r89ab.val[2]), vget_high_u16(_r89ab.val[2]));
-                uint16x4x2_t _r37_1 = vuzp_u16(vget_low_u16(_r89ab.val[3]), vget_high_u16(_r89ab.val[3]));
-                vst1q_u16(pp, _r04.val[0]);
-                vst1_u16(pp + 8, _r04_1.val[0]);
-                vst1q_u16(pp + 12, _r15.val[0]);
-                vst1_u16(pp + 20, _r15_1.val[0]);
-                vst1q_u16(pp + 24, _r26.val[0]);
-                vst1_u16(pp + 32, _r26_1.val[0]);
-                vst1q_u16(pp + 36, _r37.val[0]);
-                vst1_u16(pp + 44, _r37_1.val[0]);
-                vst1q_u16(pp + 48, _r04.val[1]);
-                vst1_u16(pp + 56, _r04_1.val[1]);
-                vst1q_u16(pp + 60, _r15.val[1]);
-                vst1_u16(pp + 68, _r15_1.val[1]);
-                vst1q_u16(pp + 72, _r26.val[1]);
-                vst1_u16(pp + 80, _r26_1.val[1]);
-                vst1q_u16(pp + 84, _r37.val[1]);
-                vst1_u16(pp + 92, _r37_1.val[1]);
-                pp += 96;
-                p0 += B_hstep * 8;
-            }
-        }
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + k * B_hstep + (j + jj) * 4;
-
-            int kk = 0;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                uint16x8x4_t _r0123 = vld4q_u16(p0);
-                uint16x4x4_t _r89ab = vld4_u16(p0 + 32);
-                vst1q_u16(pp, _r0123.val[0]);
-                vst1_u16(pp + 8, _r89ab.val[0]);
-                vst1q_u16(pp + 12, _r0123.val[1]);
-                vst1_u16(pp + 20, _r89ab.val[1]);
-                vst1q_u16(pp + 24, _r0123.val[2]);
-                vst1_u16(pp + 32, _r89ab.val[2]);
-                vst1q_u16(pp + 36, _r0123.val[3]);
-                vst1_u16(pp + 44, _r89ab.val[3]);
-                pp += 48;
-                p0 += B_hstep * 4;
-            }
-        }
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + k * B_hstep + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                vst1q_u16(pp, vld1q_u16(p0));
-                vst1_u16(pp + 8, vld1_u16(p0 + 8));
-                pp += 12;
-                p0 += B_hstep;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + k * B_hstep + (j + jj) * 8;
-
-            int kk = 0;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                uint16x8x4_t _r0123 = vld4q_u16(p0);
-                uint16x8x4_t _r4567 = vld4q_u16(p0 + 32);
-                uint16x8x2_t _r04 = vuzpq_u16(_r0123.val[0], _r4567.val[0]);
-                uint16x8x2_t _r15 = vuzpq_u16(_r0123.val[1], _r4567.val[1]);
-                uint16x8x2_t _r26 = vuzpq_u16(_r0123.val[2], _r4567.val[2]);
-                uint16x8x2_t _r37 = vuzpq_u16(_r0123.val[3], _r4567.val[3]);
-                vst1q_u16(pp, _r04.val[0]);
-                vst1q_u16(pp + 8, _r15.val[0]);
-                vst1q_u16(pp + 16, _r26.val[0]);
-                vst1q_u16(pp + 24, _r37.val[0]);
-                vst1q_u16(pp + 32, _r04.val[1]);
-                vst1q_u16(pp + 40, _r15.val[1]);
-                vst1q_u16(pp + 48, _r26.val[1]);
-                vst1q_u16(pp + 56, _r37.val[1]);
-                pp += 64;
-                p0 += B_hstep * 8;
-            }
-        }
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + k * B_hstep + (j + jj) * 4;
-
-            int kk = 0;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                uint16x8x4_t _r0123 = vld4q_u16(p0);
-                vst1q_u16(pp, _r0123.val[0]);
-                vst1q_u16(pp + 8, _r0123.val[1]);
-                vst1q_u16(pp + 16, _r0123.val[2]);
-                vst1q_u16(pp + 24, _r0123.val[3]);
-                pp += 32;
-                p0 += B_hstep * 4;
-            }
-        }
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + k * B_hstep + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                vst1q_u16(pp, vld1q_u16(p0));
-                pp += 8;
-                p0 += B_hstep;
-            }
-        }
-    }
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + k * B_hstep + (j + jj) * 8;
-
-            int kk = 0;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                uint16x8x4_t _r0123;
-                _r0123.val[0] = vld1q_u16(p0);
-                _r0123.val[1] = vld1q_u16(p0 + 8);
-                _r0123.val[2] = vld1q_u16(p0 + 16);
-                _r0123.val[3] = vld1q_u16(p0 + 24);
-                vst4q_u16(pp, _r0123);
-                pp += 32;
-                p0 += B_hstep * 8;
-            }
-        }
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + k * B_hstep + (j + jj) * 4;
-
-            int kk = 0;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                uint16x4x4_t _r0123 = vld4_u16(p0);
-                vst1q_u16(pp, vcombine_u16(_r0123.val[0], _r0123.val[1]));
-                vst1q_u16(pp + 8, vcombine_u16(_r0123.val[2], _r0123.val[3]));
-                pp += 16;
-                p0 += B_hstep * 4;
-            }
-        }
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + k * B_hstep + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                vst1_u16(pp, vld1_u16(p0));
-                pp += 4;
-                p0 += B_hstep;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-#if __ARM_NEON
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + k * B_hstep + (j + jj) * 8;
-
-            int kk = 0;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                uint16x8x2_t _r01;
-                _r01.val[0] = vld1q_u16(p0);
-                _r01.val[1] = vld1q_u16(p0 + 8);
-                vst2q_u16(pp, _r01);
-                pp += 16;
-                p0 += B_hstep * 8;
-            }
-        }
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + k * B_hstep + (j + jj) * 4;
-
-            int kk = 0;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                uint16x4x2_t _r01;
-                _r01.val[0] = vld1_u16(p0);
-                _r01.val[1] = vld1_u16(p0 + 4);
-                vst2_u16(pp, _r01);
-                pp += 8;
-                p0 += B_hstep * 4;
-            }
-        }
-#endif // __ARM_NEON
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + k * B_hstep + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                pp += 2;
-                p0 += B_hstep;
-            }
-        }
-    }
-    for (; jj < max_jj; jj += 1)
-    {
-#if __ARM_NEON
-        if (elempack == 8)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + k * B_hstep + (j + jj) * 8;
-
-            int kk = 0;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                vst1q_u16(pp, vld1q_u16(p0));
-                pp += 8;
-                p0 += B_hstep * 8;
-            }
-        }
-        if (elempack == 4)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + k * B_hstep + (j + jj) * 4;
-
-            int kk = 0;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                vst1_u16(pp, vld1_u16(p0));
-                pp += 4;
-                p0 += B_hstep * 4;
-            }
-        }
-#endif // __ARM_NEON
-        if (elempack == 1)
-        {
-            const unsigned short* p0 = (const unsigned short*)B + k * B_hstep + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp += 1;
-                p0 += B_hstep;
-            }
-        }
-    }
-}
-
-static void transpose_unpack_output_tile_bf16_fp16(const Mat& topT, Mat& top_blob, int i, int max_ii, int j, int max_jj)
-{
-    const int out_elempack = top_blob.elempack;
-    const int out_hstep = top_blob.dims == 3 ? (int)top_blob.cstep : top_blob.w;
-
-    const unsigned short* pp = topT;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        if (out_elempack == 8)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + (j / 8 * 8) * out_hstep + (i + ii) * 8;
-
-            int jj = 0;
-            if (j % 8 == 4)
-            {
-                uint16x8_t _r0 = vld1q_u16(pp);
-                uint16x8_t _r1 = vld1q_u16(pp + 8);
-                uint16x8_t _r2 = vld1q_u16(pp + 8 * 2);
-                uint16x8_t _r3 = vld1q_u16(pp + 8 * 3);
-                transpose8x4_u16(_r0, _r1, _r2, _r3);
-                vst1_u16(p0 + 4, vget_low_u16(_r0));
-                vst1_u16(p0 + 8 + 4, vget_high_u16(_r0));
-                vst1_u16(p0 + 8 * 2 + 4, vget_low_u16(_r1));
-                vst1_u16(p0 + 8 * 3 + 4, vget_high_u16(_r1));
-                vst1_u16(p0 + 8 * 4 + 4, vget_low_u16(_r2));
-                vst1_u16(p0 + 8 * 5 + 4, vget_high_u16(_r2));
-                vst1_u16(p0 + 8 * 6 + 4, vget_low_u16(_r3));
-                vst1_u16(p0 + 8 * 7 + 4, vget_high_u16(_r3));
-                pp += 32;
-                p0 += out_hstep * 8;
-                jj += 4;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                uint16x8_t _r0 = vld1q_u16(pp);
-                uint16x8_t _r1 = vld1q_u16(pp + 8);
-                uint16x8_t _r2 = vld1q_u16(pp + 8 * 2);
-                uint16x8_t _r3 = vld1q_u16(pp + 8 * 3);
-                uint16x8_t _r4 = vld1q_u16(pp + 8 * 4);
-                uint16x8_t _r5 = vld1q_u16(pp + 8 * 5);
-                uint16x8_t _r6 = vld1q_u16(pp + 8 * 6);
-                uint16x8_t _r7 = vld1q_u16(pp + 8 * 7);
-                transpose8x8_u16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-                vst1q_u16(p0, _r0);
-                vst1q_u16(p0 + 8, _r1);
-                vst1q_u16(p0 + 8 * 2, _r2);
-                vst1q_u16(p0 + 8 * 3, _r3);
-                vst1q_u16(p0 + 8 * 4, _r4);
-                vst1q_u16(p0 + 8 * 5, _r5);
-                vst1q_u16(p0 + 8 * 6, _r6);
-                vst1q_u16(p0 + 8 * 7, _r7);
-                pp += 64;
-                p0 += out_hstep * 8;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                uint16x8_t _r0 = vld1q_u16(pp);
-                uint16x8_t _r1 = vld1q_u16(pp + 8);
-                uint16x8_t _r2 = vld1q_u16(pp + 8 * 2);
-                uint16x8_t _r3 = vld1q_u16(pp + 8 * 3);
-                transpose8x4_u16(_r0, _r1, _r2, _r3);
-                vst1_u16(p0, vget_low_u16(_r0));
-                vst1_u16(p0 + 8, vget_high_u16(_r0));
-                vst1_u16(p0 + 8 * 2, vget_low_u16(_r1));
-                vst1_u16(p0 + 8 * 3, vget_high_u16(_r1));
-                vst1_u16(p0 + 8 * 4, vget_low_u16(_r2));
-                vst1_u16(p0 + 8 * 5, vget_high_u16(_r2));
-                vst1_u16(p0 + 8 * 6, vget_low_u16(_r3));
-                vst1_u16(p0 + 8 * 7, vget_high_u16(_r3));
-                pp += 32;
-                p0 += out_hstep * 8;
-            }
-        }
-        if (out_elempack == 4)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii) * 4;
-
-            for (int jj = 0; jj + 3 < max_jj; jj += 4)
-            {
-                uint16x8x4_t _r0123;
-                _r0123.val[0] = vld1q_u16(pp);
-                _r0123.val[1] = vld1q_u16(pp + 8);
-                _r0123.val[2] = vld1q_u16(pp + 8 * 2);
-                _r0123.val[3] = vld1q_u16(pp + 8 * 3);
-                vst4q_u16(p0, _r0123);
-                pp += 32;
-                p0 += out_hstep * 4;
-            }
-        }
-        if (out_elempack == 1)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii);
-
-            for (int jj = 0; jj < max_jj; jj += 1)
-            {
-                uint16x8_t _r0 = vld1q_u16(pp);
-                vst1q_u16(p0, _r0);
-                pp += 8;
-                p0 += out_hstep;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-#if __aarch64__
-        if (out_elempack == 8)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + (j / 8 * 8) * out_hstep + (i + ii) * 8;
-
-            int jj = 0;
-            if (j % 8 == 4)
-            {
-                uint16x4x4_t _r0123 = vld4_u16(pp);
-                vst1_u16(p0 + 4, _r0123.val[0]);
-                vst1_u16(p0 + 8 + 4, _r0123.val[1]);
-                vst1_u16(p0 + 16 + 4, _r0123.val[2]);
-                vst1_u16(p0 + 24 + 4, _r0123.val[3]);
-                pp += 16;
-                p0 += out_hstep * 8;
-                jj += 4;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                uint16x8x4_t _r0123 = vld4q_u16(pp);
-                vst1q_u16(p0, _r0123.val[0]);
-                vst1q_u16(p0 + 8, _r0123.val[1]);
-                vst1q_u16(p0 + 16, _r0123.val[2]);
-                vst1q_u16(p0 + 24, _r0123.val[3]);
-                pp += 32;
-                p0 += out_hstep * 8;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                uint16x4x4_t _r0123 = vld4_u16(pp);
-                vst1_u16(p0, _r0123.val[0]);
-                vst1_u16(p0 + 8, _r0123.val[1]);
-                vst1_u16(p0 + 16, _r0123.val[2]);
-                vst1_u16(p0 + 24, _r0123.val[3]);
-                pp += 16;
-                p0 += out_hstep * 8;
-            }
-        }
-#endif // __aarch64__
-        if (out_elempack == 4)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii) * 4;
-
-            for (int jj = 0; jj + 3 < max_jj; jj += 4)
-            {
-                uint16x4x4_t _r0123;
-                _r0123.val[0] = vld1_u16(pp);
-                _r0123.val[1] = vld1_u16(pp + 4);
-                _r0123.val[2] = vld1_u16(pp + 8);
-                _r0123.val[3] = vld1_u16(pp + 12);
-                vst4_u16(p0, _r0123);
-                pp += 16;
-                p0 += out_hstep * 4;
-            }
-        }
-        if (out_elempack == 1)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii);
-
-            for (int jj = 0; jj < max_jj; jj += 1)
-            {
-                uint16x4_t _r0 = vld1_u16(pp);
-                vst1_u16(p0, _r0);
-                pp += 4;
-                p0 += out_hstep;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-#if __ARM_NEON
-#if __aarch64__
-        if (out_elempack == 8)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + (j / 8 * 8) * out_hstep + (i + ii) * 8;
-
-            int jj = 0;
-            if (j % 8 == 4)
-            {
-                p0[0 + 4] = pp[0];
-                p0[1 + 4] = pp[2];
-                p0[2 + 4] = pp[4];
-                p0[3 + 4] = pp[6];
-                p0[8 + 4] = pp[1];
-                p0[9 + 4] = pp[3];
-                p0[10 + 4] = pp[5];
-                p0[11 + 4] = pp[7];
-                pp += 8;
-                p0 += out_hstep * 8;
-                jj += 4;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                p0[0] = pp[0];
-                p0[1] = pp[2];
-                p0[2] = pp[4];
-                p0[3] = pp[6];
-                p0[4] = pp[8];
-                p0[5] = pp[10];
-                p0[6] = pp[12];
-                p0[7] = pp[14];
-                p0[8] = pp[1];
-                p0[9] = pp[3];
-                p0[10] = pp[5];
-                p0[11] = pp[7];
-                p0[12] = pp[9];
-                p0[13] = pp[11];
-                p0[14] = pp[13];
-                p0[15] = pp[15];
-                pp += 16;
-                p0 += out_hstep * 8;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                p0[0] = pp[0];
-                p0[1] = pp[2];
-                p0[2] = pp[4];
-                p0[3] = pp[6];
-                p0[8] = pp[1];
-                p0[9] = pp[3];
-                p0[10] = pp[5];
-                p0[11] = pp[7];
-                pp += 8;
-                p0 += out_hstep * 8;
-            }
-        }
-#endif // __aarch64__
-        if (out_elempack == 4)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii) * 4;
-
-            for (int jj = 0; jj + 3 < max_jj; jj += 4)
-            {
-                p0[0] = pp[0];
-                p0[1] = pp[2];
-                p0[2] = pp[4];
-                p0[3] = pp[6];
-                p0[4] = pp[1];
-                p0[5] = pp[3];
-                p0[6] = pp[5];
-                p0[7] = pp[7];
-                pp += 8;
-                p0 += out_hstep * 4;
-            }
-        }
-#endif // __ARM_NEON
-        if (out_elempack == 1)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii);
-
-            for (int jj = 0; jj < max_jj; jj += 1)
-            {
-                p0[0] = pp[0];
-                p0[1] = pp[1];
-                pp += 2;
-                p0 += out_hstep;
-            }
-        }
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-#if __ARM_NEON
-#if __aarch64__
-        if (out_elempack == 8)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + (j / 8 * 8) * out_hstep + (i + ii) * 8;
-
-            int jj = 0;
-            if (j % 8 == 4)
-            {
-                uint16x4_t _r0 = vld1_u16(pp);
-                vst1_u16(p0 + 4, _r0);
-                pp += 4;
-                p0 += out_hstep * 8;
-                jj += 4;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                uint16x8_t _r0 = vld1q_u16(pp);
-                vst1q_u16(p0, _r0);
-                pp += 8;
-                p0 += out_hstep * 8;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                uint16x4_t _r0 = vld1_u16(pp);
-                vst1_u16(p0, _r0);
-                pp += 4;
-                p0 += out_hstep * 8;
-            }
-        }
-#endif // __aarch64__
-        if (out_elempack == 4)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii) * 4;
-
-            for (int jj = 0; jj + 3 < max_jj; jj += 4)
-            {
-                uint16x4_t _r0 = vld1_u16(pp);
-                vst1_u16(p0, _r0);
-                pp += 4;
-                p0 += out_hstep * 4;
-            }
-        }
-#endif // __ARM_NEON
-        if (out_elempack == 1)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii);
-
-            for (int jj = 0; jj < max_jj; jj += 1)
-            {
-                p0[0] = pp[0];
-                pp += 1;
-                p0 += out_hstep;
-            }
-        }
-    }
-}
-
-static void get_optimal_tile_mnk_bf16s_fp16s(int M, int N, int K, int constant_TILE_M, int constant_TILE_N, int constant_TILE_K, int& TILE_M, int& TILE_N, int& TILE_K, int nT)
-{
-    // resolve optimal tile size from cache size
-    const size_t l2_cache_size = get_cpu_level2_cache_size();
-
-    if (nT == 0)
-        nT = get_physical_big_cpu_count();
-
-    int tile_size = (int)sqrtf((float)l2_cache_size / (2 * sizeof(unsigned short) + sizeof(float)));
-
-    TILE_M = std::max(8, tile_size / 8 * 8);
-    TILE_N = std::max(4, tile_size / 4 * 4);
-    TILE_K = std::max(8, tile_size / 8 * 8);
-
-    if (K > 0)
-    {
-        int nn_K = (K + TILE_K - 1) / TILE_K;
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 7) / 8 * 8);
-
-        if (nn_K == 1)
-        {
-            tile_size = (int)((float)l2_cache_size / 2 / sizeof(unsigned short) / TILE_K);
-
-            TILE_M = std::max(8, tile_size / 8 * 8);
-            TILE_N = std::max(4, tile_size / 4 * 4);
-        }
-    }
-
-    TILE_M *= std::min(nT, get_physical_cpu_count());
-
-    if (M > 0)
-    {
-        int nn_M = (M + TILE_M - 1) / TILE_M;
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-    }
-
-    if (N > 0)
-    {
-        int nn_N = (N + TILE_N - 1) / TILE_N;
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-    }
-
-    if (nT > 1)
-    {
-        TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 7) / 8 * 8);
-    }
-
-    // always take constant TILE_M/N/K value when provided
-    if (constant_TILE_M > 0)
-    {
-        TILE_M = (constant_TILE_M + 7) / 8 * 8;
-    }
-
-    if (constant_TILE_N > 0)
-    {
-        TILE_N = (constant_TILE_N + 3) / 4 * 4;
-    }
-
-    if (constant_TILE_K > 0)
-    {
-        TILE_K = (constant_TILE_K + 7) / 8 * 8;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_fp16s.h
deleted file mode 100644
index 6905c637..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gemm_fp16s.h
+++ /dev/null
@@ -1,3945 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if NCNN_RUNTIME_CPU && NCNN_ARM82FP16FML && __aarch64__ && !__ARM_FEATURE_FP16_FML
-void gemm_transB_packed_tile_fp16s_asimdfhm(const Mat& AT_tile, const Mat& BT_tile, const Mat& CT_tile, Mat& topT_tile, Mat& top_blob, int broadcast_type_C, float alpha, int i, int max_ii, int j, int max_jj, int k, int max_kk, bool k_end);
-#endif
-
-static void pack_A_tile_fp32_to_fp16(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
-{
-    const int A_hstep = A.dims == 3 ? (int)A.cstep : A.w;
-
-    unsigned short* pp = AT;
-
-    int ii = 0;
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        const float* p0 = (const float*)A + (i + ii) * A_hstep + k;
-        const float* p1 = (const float*)A + (i + ii + 1) * A_hstep + k;
-        const float* p2 = (const float*)A + (i + ii + 2) * A_hstep + k;
-        const float* p3 = (const float*)A + (i + ii + 3) * A_hstep + k;
-        const float* p4 = (const float*)A + (i + ii + 4) * A_hstep + k;
-        const float* p5 = (const float*)A + (i + ii + 5) * A_hstep + k;
-        const float* p6 = (const float*)A + (i + ii + 6) * A_hstep + k;
-        const float* p7 = (const float*)A + (i + ii + 7) * A_hstep + k;
-
-        int kk = 0;
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8_t _r0 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p0)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0 + 4)));
-            uint16x8_t _r1 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p1)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p1 + 4)));
-            uint16x8_t _r2 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p2)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p2 + 4)));
-            uint16x8_t _r3 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p3)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p3 + 4)));
-            uint16x8_t _r4 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p4)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p4 + 4)));
-            uint16x8_t _r5 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p5)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p5 + 4)));
-            uint16x8_t _r6 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p6)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p6 + 4)));
-            uint16x8_t _r7 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p7)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p7 + 4)));
-            transpose8x8_u16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-            vst1q_u16(pp, _r0);
-            vst1q_u16(pp + 8, _r1);
-            vst1q_u16(pp + 8 * 2, _r2);
-            vst1q_u16(pp + 8 * 3, _r3);
-            vst1q_u16(pp + 8 * 4, _r4);
-            vst1q_u16(pp + 8 * 5, _r5);
-            vst1q_u16(pp + 8 * 6, _r6);
-            vst1q_u16(pp + 8 * 7, _r7);
-            pp += 64;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-            p4 += 8;
-            p5 += 8;
-            p6 += 8;
-            p7 += 8;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_float16(p0[0]);
-            pp[1] = float32_to_float16(p1[0]);
-            pp[2] = float32_to_float16(p2[0]);
-            pp[3] = float32_to_float16(p3[0]);
-            pp[4] = float32_to_float16(p4[0]);
-            pp[5] = float32_to_float16(p5[0]);
-            pp[6] = float32_to_float16(p6[0]);
-            pp[7] = float32_to_float16(p7[0]);
-            pp += 8;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-            p4++;
-            p5++;
-            p6++;
-            p7++;
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        const float* p0 = (const float*)A + (i + ii) * A_hstep + k;
-        const float* p1 = (const float*)A + (i + ii + 1) * A_hstep + k;
-        const float* p2 = (const float*)A + (i + ii + 2) * A_hstep + k;
-        const float* p3 = (const float*)A + (i + ii + 3) * A_hstep + k;
-
-        int kk = 0;
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8x4_t _r0123;
-            _r0123.val[0] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p0)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0 + 4)));
-            _r0123.val[1] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p1)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p1 + 4)));
-            _r0123.val[2] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p2)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p2 + 4)));
-            _r0123.val[3] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p3)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p3 + 4)));
-            vst4q_u16(pp, _r0123);
-            pp += 32;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4x4_t _r0123;
-            _r0123.val[0] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0));
-            _r0123.val[1] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p1));
-            _r0123.val[2] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p2));
-            _r0123.val[3] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p3));
-            vst4_u16(pp, _r0123);
-            pp += 16;
-            p0 += 4;
-            p1 += 4;
-            p2 += 4;
-            p3 += 4;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_float16(p0[0]);
-            pp[1] = float32_to_float16(p1[0]);
-            pp[2] = float32_to_float16(p2[0]);
-            pp[3] = float32_to_float16(p3[0]);
-            pp += 4;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-        }
-    }
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        const float* p0 = (const float*)A + (i + ii) * A_hstep + k;
-        const float* p1 = (const float*)A + (i + ii + 1) * A_hstep + k;
-
-        int kk = 0;
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8x2_t _r01;
-            _r01.val[0] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p0)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0 + 4)));
-            _r01.val[1] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p1)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p1 + 4)));
-            vst2q_u16(pp, _r01);
-            pp += 16;
-            p0 += 8;
-            p1 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4x2_t _r01;
-            _r01.val[0] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0));
-            _r01.val[1] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p1));
-            vst2_u16(pp, _r01);
-            pp += 8;
-            p0 += 4;
-            p1 += 4;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_float16(p0[0]);
-            pp[1] = float32_to_float16(p1[0]);
-            pp += 2;
-            p0++;
-            p1++;
-        }
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        const float* p0 = (const float*)A + (i + ii) * A_hstep + k;
-
-        int kk = 0;
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8_t _r0 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p0)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0 + 4)));
-            vst1q_u16(pp, _r0);
-            pp += 8;
-            p0 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4_t _r0 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0));
-            vst1_u16(pp, _r0);
-            pp += 4;
-            p0 += 4;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_float16(p0[0]);
-            pp += 1;
-            p0++;
-        }
-    }
-}
-
-static void transpose_pack_A_tile_fp32_to_fp16(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
-{
-    const int A_hstep = A.dims == 3 ? (int)A.cstep : A.w;
-
-    unsigned short* pp = AT;
-
-    int ii = 0;
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        const float* p0 = (const float*)A + k * A_hstep + (i + ii);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            uint16x8_t _r0 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p0)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0 + 4)));
-            vst1q_u16(pp, _r0);
-            pp += 8;
-            p0 += A_hstep;
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        const float* p0 = (const float*)A + k * A_hstep + (i + ii);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            uint16x4_t _r0 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0));
-            vst1_u16(pp, _r0);
-            pp += 4;
-            p0 += A_hstep;
-        }
-    }
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        const float* p0 = (const float*)A + k * A_hstep + (i + ii);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_float16(p0[0]);
-            pp[1] = float32_to_float16(p0[1]);
-            pp += 2;
-            p0 += A_hstep;
-        }
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        const float* p0 = (const float*)A + k * A_hstep + (i + ii);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_float16(p0[0]);
-            pp += 1;
-            p0 += A_hstep;
-        }
-    }
-}
-
-static void pack_B_tile_fp32_to_fp16(const Mat& B, Mat& BT, int j, int max_jj, int k, int max_kk)
-{
-    const int B_hstep = B.dims == 3 ? (int)B.cstep : B.w;
-
-    unsigned short* pp = BT;
-
-    int jj = 0;
-#if __aarch64__
-    for (; jj + 11 < max_jj; jj += 12)
-    {
-        const float* p0 = (const float*)B + (j + jj) * B_hstep + k;
-        const float* p1 = (const float*)B + (j + jj + 1) * B_hstep + k;
-        const float* p2 = (const float*)B + (j + jj + 2) * B_hstep + k;
-        const float* p3 = (const float*)B + (j + jj + 3) * B_hstep + k;
-        const float* p4 = (const float*)B + (j + jj + 4) * B_hstep + k;
-        const float* p5 = (const float*)B + (j + jj + 5) * B_hstep + k;
-        const float* p6 = (const float*)B + (j + jj + 6) * B_hstep + k;
-        const float* p7 = (const float*)B + (j + jj + 7) * B_hstep + k;
-        const float* p8 = (const float*)B + (j + jj + 8) * B_hstep + k;
-        const float* p9 = (const float*)B + (j + jj + 9) * B_hstep + k;
-        const float* pa = (const float*)B + (j + jj + 10) * B_hstep + k;
-        const float* pb = (const float*)B + (j + jj + 11) * B_hstep + k;
-
-        int kk = 0;
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4_t _r0 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0));
-            uint16x4_t _r1 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p1));
-            uint16x4_t _r2 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p2));
-            uint16x4_t _r3 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p3));
-            uint16x4_t _r4 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p4));
-            uint16x4_t _r5 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p5));
-            uint16x4_t _r6 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p6));
-            uint16x4_t _r7 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p7));
-            uint16x4_t _r8 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p8));
-            uint16x4_t _r9 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p9));
-            uint16x4_t _ra = (uint16x4_t)vcvt_f16_f32(vld1q_f32(pa));
-            uint16x4_t _rb = (uint16x4_t)vcvt_f16_f32(vld1q_f32(pb));
-
-            transpose4x4_u16(_r0, _r1, _r2, _r3);
-            transpose4x4_u16(_r4, _r5, _r6, _r7);
-            transpose4x4_u16(_r8, _r9, _ra, _rb);
-
-            vst1_u16(pp, _r0);
-            vst1_u16(pp + 4, _r4);
-            vst1_u16(pp + 4 * 2, _r8);
-            vst1_u16(pp + 4 * 3, _r1);
-            vst1_u16(pp + 4 * 4, _r5);
-            vst1_u16(pp + 4 * 5, _r9);
-            vst1_u16(pp + 4 * 6, _r2);
-            vst1_u16(pp + 4 * 7, _r6);
-            vst1_u16(pp + 4 * 8, _ra);
-            vst1_u16(pp + 4 * 9, _r3);
-            vst1_u16(pp + 4 * 10, _r7);
-            vst1_u16(pp + 4 * 11, _rb);
-            pp += 48;
-            p0 += 4;
-            p1 += 4;
-            p2 += 4;
-            p3 += 4;
-            p4 += 4;
-            p5 += 4;
-            p6 += 4;
-            p7 += 4;
-            p8 += 4;
-            p9 += 4;
-            pa += 4;
-            pb += 4;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_float16(p0[0]);
-            pp[1] = float32_to_float16(p1[0]);
-            pp[2] = float32_to_float16(p2[0]);
-            pp[3] = float32_to_float16(p3[0]);
-            pp[4] = float32_to_float16(p4[0]);
-            pp[5] = float32_to_float16(p5[0]);
-            pp[6] = float32_to_float16(p6[0]);
-            pp[7] = float32_to_float16(p7[0]);
-            pp[8] = float32_to_float16(p8[0]);
-            pp[9] = float32_to_float16(p9[0]);
-            pp[10] = float32_to_float16(pa[0]);
-            pp[11] = float32_to_float16(pb[0]);
-            pp += 12;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-            p4++;
-            p5++;
-            p6++;
-            p7++;
-            p8++;
-            p9++;
-            pa++;
-            pb++;
-        }
-    }
-#endif // __aarch64__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-        const float* p0 = (const float*)B + (j + jj) * B_hstep + k;
-        const float* p1 = (const float*)B + (j + jj + 1) * B_hstep + k;
-        const float* p2 = (const float*)B + (j + jj + 2) * B_hstep + k;
-        const float* p3 = (const float*)B + (j + jj + 3) * B_hstep + k;
-        const float* p4 = (const float*)B + (j + jj + 4) * B_hstep + k;
-        const float* p5 = (const float*)B + (j + jj + 5) * B_hstep + k;
-        const float* p6 = (const float*)B + (j + jj + 6) * B_hstep + k;
-        const float* p7 = (const float*)B + (j + jj + 7) * B_hstep + k;
-
-        int kk = 0;
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8_t _r0 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p0)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0 + 4)));
-            uint16x8_t _r1 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p1)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p1 + 4)));
-            uint16x8_t _r2 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p2)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p2 + 4)));
-            uint16x8_t _r3 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p3)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p3 + 4)));
-            uint16x8_t _r4 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p4)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p4 + 4)));
-            uint16x8_t _r5 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p5)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p5 + 4)));
-            uint16x8_t _r6 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p6)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p6 + 4)));
-            uint16x8_t _r7 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p7)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p7 + 4)));
-            transpose8x8_u16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-            vst1q_u16(pp, _r0);
-            vst1q_u16(pp + 8, _r1);
-            vst1q_u16(pp + 8 * 2, _r2);
-            vst1q_u16(pp + 8 * 3, _r3);
-            vst1q_u16(pp + 8 * 4, _r4);
-            vst1q_u16(pp + 8 * 5, _r5);
-            vst1q_u16(pp + 8 * 6, _r6);
-            vst1q_u16(pp + 8 * 7, _r7);
-            pp += 64;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-            p4 += 8;
-            p5 += 8;
-            p6 += 8;
-            p7 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4_t _r0 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0));
-            uint16x4_t _r1 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p1));
-            uint16x4_t _r2 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p2));
-            uint16x4_t _r3 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p3));
-            uint16x4_t _r4 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p4));
-            uint16x4_t _r5 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p5));
-            uint16x4_t _r6 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p6));
-            uint16x4_t _r7 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p7));
-
-            transpose4x4_u16(_r0, _r1, _r2, _r3);
-            transpose4x4_u16(_r4, _r5, _r6, _r7);
-
-            vst1_u16(pp, _r0);
-            vst1_u16(pp + 4, _r4);
-            vst1_u16(pp + 4 * 2, _r1);
-            vst1_u16(pp + 4 * 3, _r5);
-            vst1_u16(pp + 4 * 4, _r2);
-            vst1_u16(pp + 4 * 5, _r6);
-            vst1_u16(pp + 4 * 6, _r3);
-            vst1_u16(pp + 4 * 7, _r7);
-            pp += 32;
-            p0 += 4;
-            p1 += 4;
-            p2 += 4;
-            p3 += 4;
-            p4 += 4;
-            p5 += 4;
-            p6 += 4;
-            p7 += 4;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_float16(p0[0]);
-            pp[1] = float32_to_float16(p1[0]);
-            pp[2] = float32_to_float16(p2[0]);
-            pp[3] = float32_to_float16(p3[0]);
-            pp[4] = float32_to_float16(p4[0]);
-            pp[5] = float32_to_float16(p5[0]);
-            pp[6] = float32_to_float16(p6[0]);
-            pp[7] = float32_to_float16(p7[0]);
-            pp += 8;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-            p4++;
-            p5++;
-            p6++;
-            p7++;
-        }
-    }
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-        const float* p0 = (const float*)B + (j + jj) * B_hstep + k;
-        const float* p1 = (const float*)B + (j + jj + 1) * B_hstep + k;
-        const float* p2 = (const float*)B + (j + jj + 2) * B_hstep + k;
-        const float* p3 = (const float*)B + (j + jj + 3) * B_hstep + k;
-
-        int kk = 0;
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8x4_t _r0123;
-            _r0123.val[0] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p0)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0 + 4)));
-            _r0123.val[1] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p1)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p1 + 4)));
-            _r0123.val[2] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p2)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p2 + 4)));
-            _r0123.val[3] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p3)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p3 + 4)));
-            vst4q_u16(pp, _r0123);
-            pp += 32;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4x4_t _r0123;
-            _r0123.val[0] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0));
-            _r0123.val[1] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p1));
-            _r0123.val[2] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p2));
-            _r0123.val[3] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p3));
-            vst4_u16(pp, _r0123);
-            pp += 16;
-            p0 += 4;
-            p1 += 4;
-            p2 += 4;
-            p3 += 4;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_float16(p0[0]);
-            pp[1] = float32_to_float16(p1[0]);
-            pp[2] = float32_to_float16(p2[0]);
-            pp[3] = float32_to_float16(p3[0]);
-            pp += 4;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-        }
-    }
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-        const float* p0 = (const float*)B + (j + jj) * B_hstep + k;
-        const float* p1 = (const float*)B + (j + jj + 1) * B_hstep + k;
-
-        int kk = 0;
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8x2_t _r01;
-            _r01.val[0] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p0)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0 + 4)));
-            _r01.val[1] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p1)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p1 + 4)));
-            vst2q_u16(pp, _r01);
-            pp += 16;
-            p0 += 8;
-            p1 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4x2_t _r01;
-            _r01.val[0] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0));
-            _r01.val[1] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p1));
-            vst2_u16(pp, _r01);
-            pp += 8;
-            p0 += 4;
-            p1 += 4;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_float16(p0[0]);
-            pp[1] = float32_to_float16(p1[0]);
-            pp += 2;
-            p0++;
-            p1++;
-        }
-    }
-    for (; jj < max_jj; jj += 1)
-    {
-        const float* p0 = (const float*)B + (j + jj) * B_hstep + k;
-
-        int kk = 0;
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            uint16x8_t _r0 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p0)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0 + 4)));
-            vst1q_u16(pp, _r0);
-            pp += 8;
-            p0 += 8;
-        }
-        for (; kk + 3 < max_kk; kk += 4)
-        {
-            uint16x4_t _r0 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0));
-            vst1_u16(pp, _r0);
-            pp += 4;
-            p0 += 4;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_float16(p0[0]);
-            pp += 1;
-            p0++;
-        }
-    }
-}
-
-static void transpose_pack_B_tile_fp32_to_fp16(const Mat& B, Mat& BT, int j, int max_jj, int k, int max_kk)
-{
-    const int B_hstep = B.dims == 3 ? (int)B.cstep : B.w;
-
-    unsigned short* pp = BT;
-
-    int jj = 0;
-#if __aarch64__
-    for (; jj + 11 < max_jj; jj += 12)
-    {
-        const float* p0 = (const float*)B + k * B_hstep + (j + jj);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            vst1_u16(pp, (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0)));
-            vst1_u16(pp + 4, (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0 + 4)));
-            vst1_u16(pp + 8, (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0 + 8)));
-            pp += 12;
-            p0 += B_hstep;
-        }
-    }
-#endif // __aarch64__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-        const float* p0 = (const float*)B + k * B_hstep + (j + jj);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            uint16x8_t _r0 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(p0)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0 + 4)));
-            vst1q_u16(pp, _r0);
-            pp += 8;
-            p0 += B_hstep;
-        }
-    }
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-        const float* p0 = (const float*)B + k * B_hstep + (j + jj);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            uint16x4_t _r0 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(p0));
-            vst1_u16(pp, _r0);
-            pp += 4;
-            p0 += B_hstep;
-        }
-    }
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-        const float* p0 = (const float*)B + k * B_hstep + (j + jj);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_float16(p0[0]);
-            pp[1] = float32_to_float16(p0[1]);
-            pp += 2;
-            p0 += B_hstep;
-        }
-    }
-    for (; jj < max_jj; jj += 1)
-    {
-        const float* p0 = (const float*)B + k * B_hstep + (j + jj);
-
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = float32_to_float16(p0[0]);
-            pp += 1;
-            p0 += B_hstep;
-        }
-    }
-}
-
-static void transpose_unpack_output_tile_fp32_to_fp16(const Mat& topT, Mat& top_blob, int i, int max_ii, int j, int max_jj)
-{
-    const int out_elempack = top_blob.elempack;
-    const int out_hstep = top_blob.dims == 3 ? (int)top_blob.cstep : top_blob.w;
-
-    const float* pp = topT;
-
-    int ii = 0;
-#if __ARM_NEON
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        if (out_elempack == 4)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii) * 4;
-
-            for (int jj = 0; jj + 3 < max_jj; jj += 4)
-            {
-                uint16x8x4_t _r0;
-                _r0.val[0] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(pp)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(pp + 4)));
-                _r0.val[1] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(pp + 8)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(pp + 12)));
-                _r0.val[2] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(pp + 16)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(pp + 20)));
-                _r0.val[3] = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(pp + 24)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(pp + 28)));
-                vst4q_u16(p0, _r0);
-                pp += 32;
-                p0 += out_hstep * 4;
-            }
-        }
-        if (out_elempack == 1)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii);
-
-            for (int jj = 0; jj < max_jj; jj += 1)
-            {
-                uint16x8_t _r0 = vcombine_u16((uint16x4_t)vcvt_f16_f32(vld1q_f32(pp)), (uint16x4_t)vcvt_f16_f32(vld1q_f32(pp + 4)));
-                vst1q_u16(p0, _r0);
-                pp += 8;
-                p0 += out_hstep;
-            }
-        }
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        if (out_elempack == 4)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii) * 4;
-
-            for (int jj = 0; jj + 3 < max_jj; jj += 4)
-            {
-                uint16x4x4_t _r0123;
-                _r0123.val[0] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(pp));
-                _r0123.val[1] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(pp + 4));
-                _r0123.val[2] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(pp + 8));
-                _r0123.val[3] = (uint16x4_t)vcvt_f16_f32(vld1q_f32(pp + 12));
-                vst4_u16(p0, _r0123);
-                pp += 16;
-                p0 += out_hstep * 4;
-            }
-        }
-        if (out_elempack == 1)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii);
-
-            for (int jj = 0; jj < max_jj; jj += 1)
-            {
-                uint16x4_t _r0 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(pp));
-                vst1_u16(p0, _r0);
-                pp += 4;
-                p0 += out_hstep;
-            }
-        }
-    }
-#endif // __ARM_NEON
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-#if __ARM_NEON
-        if (out_elempack == 4)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii) * 4;
-
-            for (int jj = 0; jj + 3 < max_jj; jj += 4)
-            {
-                p0[0] = float32_to_float16(pp[0]);
-                p0[1] = float32_to_float16(pp[2]);
-                p0[2] = float32_to_float16(pp[4]);
-                p0[3] = float32_to_float16(pp[6]);
-                p0[4] = float32_to_float16(pp[1]);
-                p0[5] = float32_to_float16(pp[3]);
-                p0[6] = float32_to_float16(pp[5]);
-                p0[7] = float32_to_float16(pp[7]);
-                pp += 8;
-                p0 += out_hstep * 4;
-            }
-        }
-#endif // __ARM_NEON
-        if (out_elempack == 1)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii);
-
-            for (int jj = 0; jj < max_jj; jj += 1)
-            {
-                p0[0] = float32_to_float16(pp[0]);
-                p0[1] = float32_to_float16(pp[1]);
-                pp += 2;
-                p0 += out_hstep;
-            }
-        }
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-#if __ARM_NEON
-        if (out_elempack == 4)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii) * 4;
-
-            for (int jj = 0; jj + 3 < max_jj; jj += 4)
-            {
-                uint16x4_t _r0 = (uint16x4_t)vcvt_f16_f32(vld1q_f32(pp));
-                vst1_u16(p0, _r0);
-                pp += 4;
-                p0 += out_hstep * 4;
-            }
-        }
-#endif // __ARM_NEON
-        if (out_elempack == 1)
-        {
-            unsigned short* p0 = (unsigned short*)top_blob + j * out_hstep + (i + ii);
-
-            for (int jj = 0; jj < max_jj; jj += 1)
-            {
-                p0[0] = float32_to_float16(pp[0]);
-                pp += 1;
-                p0 += out_hstep;
-            }
-        }
-    }
-}
-
-static void gemm_transB_packed_tile_fp16s(const Mat& AT_tile, const Mat& BT_tile, const Mat& CT_tile, Mat& topT_tile, Mat& top_blob, int broadcast_type_C, float alpha, int i, int max_ii, int j, int max_jj, int k, int max_kk, bool k_end)
-{
-#if NCNN_RUNTIME_CPU && NCNN_ARM82FP16FML && __aarch64__ && !__ARM_FEATURE_FP16_FML
-    if (ncnn::cpu_support_arm_asimdfhm())
-    {
-        gemm_transB_packed_tile_fp16s_asimdfhm(AT_tile, BT_tile, CT_tile, topT_tile, top_blob, broadcast_type_C, alpha, i, max_ii, j, max_jj, k, max_kk, k_end);
-        return;
-    }
-#endif
-
-    const int out_elempack = top_blob.elempack;
-    const int out_hstep = top_blob.dims == 3 ? (int)top_blob.cstep : top_blob.w;
-
-#if __ARM_FEATURE_FP16_FML
-    const __fp16* pAT = AT_tile;
-    const __fp16* pBT = BT_tile;
-#else
-    const unsigned short* pAT = AT_tile;
-    const unsigned short* pBT = BT_tile;
-#endif
-    const float* pC = CT_tile;
-
-    float* outptr = topT_tile;
-
-    int ii = 0;
-#if __aarch64__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        unsigned short* outptr0 = (unsigned short*)top_blob + (i + ii) * out_hstep + j * out_elempack;
-
-#if __ARM_FEATURE_FP16_FML
-        const __fp16* pB = pBT;
-#else
-        const unsigned short* pB = pBT;
-#endif
-
-        if (pC)
-        {
-            if (broadcast_type_C == 1 || broadcast_type_C == 2)
-            {
-                pC = (const float*)CT_tile + i + ii;
-            }
-            if (broadcast_type_C == 4)
-            {
-                pC = (const float*)CT_tile + j;
-            }
-        }
-
-        int jj = 0;
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum20;
-            float32x4_t _sum21;
-            float32x4_t _sum30;
-            float32x4_t _sum31;
-            float32x4_t _sum40;
-            float32x4_t _sum41;
-            float32x4_t _sum50;
-            float32x4_t _sum51;
-            float32x4_t _sum60;
-            float32x4_t _sum61;
-            float32x4_t _sum70;
-            float32x4_t _sum71;
-            float32x4_t _sum80;
-            float32x4_t _sum81;
-            float32x4_t _sum90;
-            float32x4_t _sum91;
-            float32x4_t _suma0;
-            float32x4_t _suma1;
-            float32x4_t _sumb0;
-            float32x4_t _sumb1;
-
-            if (k == 0)
-            {
-                _sum00 = vdupq_n_f32(0.f);
-                _sum01 = vdupq_n_f32(0.f);
-                _sum10 = vdupq_n_f32(0.f);
-                _sum11 = vdupq_n_f32(0.f);
-                _sum20 = vdupq_n_f32(0.f);
-                _sum21 = vdupq_n_f32(0.f);
-                _sum30 = vdupq_n_f32(0.f);
-                _sum31 = vdupq_n_f32(0.f);
-                _sum40 = vdupq_n_f32(0.f);
-                _sum41 = vdupq_n_f32(0.f);
-                _sum50 = vdupq_n_f32(0.f);
-                _sum51 = vdupq_n_f32(0.f);
-                _sum60 = vdupq_n_f32(0.f);
-                _sum61 = vdupq_n_f32(0.f);
-                _sum70 = vdupq_n_f32(0.f);
-                _sum71 = vdupq_n_f32(0.f);
-                _sum80 = vdupq_n_f32(0.f);
-                _sum81 = vdupq_n_f32(0.f);
-                _sum90 = vdupq_n_f32(0.f);
-                _sum91 = vdupq_n_f32(0.f);
-                _suma0 = vdupq_n_f32(0.f);
-                _suma1 = vdupq_n_f32(0.f);
-                _sumb0 = vdupq_n_f32(0.f);
-                _sumb1 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum10 = _sum00;
-                        _sum11 = _sum00;
-                        _sum20 = _sum00;
-                        _sum21 = _sum00;
-                        _sum30 = _sum00;
-                        _sum31 = _sum00;
-                        _sum40 = _sum00;
-                        _sum41 = _sum00;
-                        _sum50 = _sum00;
-                        _sum51 = _sum00;
-                        _sum60 = _sum00;
-                        _sum61 = _sum00;
-                        _sum70 = _sum00;
-                        _sum71 = _sum00;
-                        _sum80 = _sum00;
-                        _sum81 = _sum00;
-                        _sum90 = _sum00;
-                        _sum91 = _sum00;
-                        _suma0 = _sum00;
-                        _suma1 = _sum00;
-                        _sumb0 = _sum00;
-                        _sumb1 = _sum00;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                        _sum10 = _sum00;
-                        _sum11 = _sum01;
-                        _sum20 = _sum00;
-                        _sum21 = _sum01;
-                        _sum30 = _sum00;
-                        _sum31 = _sum01;
-                        _sum40 = _sum00;
-                        _sum41 = _sum01;
-                        _sum50 = _sum00;
-                        _sum51 = _sum01;
-                        _sum60 = _sum00;
-                        _sum61 = _sum01;
-                        _sum70 = _sum00;
-                        _sum71 = _sum01;
-                        _sum80 = _sum00;
-                        _sum81 = _sum01;
-                        _sum90 = _sum00;
-                        _sum91 = _sum01;
-                        _suma0 = _sum00;
-                        _suma1 = _sum01;
-                        _sumb0 = _sum00;
-                        _sumb1 = _sum01;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4 * 1);
-                        _sum10 = vld1q_f32(pC + 4 * 2);
-                        _sum11 = vld1q_f32(pC + 4 * 3);
-                        _sum20 = vld1q_f32(pC + 4 * 4);
-                        _sum21 = vld1q_f32(pC + 4 * 5);
-                        _sum30 = vld1q_f32(pC + 4 * 6);
-                        _sum31 = vld1q_f32(pC + 4 * 7);
-                        _sum40 = vld1q_f32(pC + 4 * 8);
-                        _sum41 = vld1q_f32(pC + 4 * 9);
-                        _sum50 = vld1q_f32(pC + 4 * 10);
-                        _sum51 = vld1q_f32(pC + 4 * 11);
-                        _sum60 = vld1q_f32(pC + 4 * 12);
-                        _sum61 = vld1q_f32(pC + 4 * 13);
-                        _sum70 = vld1q_f32(pC + 4 * 14);
-                        _sum71 = vld1q_f32(pC + 4 * 15);
-                        _sum80 = vld1q_f32(pC + 4 * 16);
-                        _sum81 = vld1q_f32(pC + 4 * 17);
-                        _sum90 = vld1q_f32(pC + 4 * 18);
-                        _sum91 = vld1q_f32(pC + 4 * 19);
-                        _suma0 = vld1q_f32(pC + 4 * 20);
-                        _suma1 = vld1q_f32(pC + 4 * 21);
-                        _sumb0 = vld1q_f32(pC + 4 * 22);
-                        _sumb1 = vld1q_f32(pC + 4 * 23);
-                        pC += 96;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum10 = vdupq_n_f32(pC[1]);
-                        _sum20 = vdupq_n_f32(pC[2]);
-                        _sum30 = vdupq_n_f32(pC[3]);
-                        _sum40 = vdupq_n_f32(pC[4]);
-                        _sum50 = vdupq_n_f32(pC[5]);
-                        _sum60 = vdupq_n_f32(pC[6]);
-                        _sum70 = vdupq_n_f32(pC[7]);
-                        _sum80 = vdupq_n_f32(pC[8]);
-                        _sum90 = vdupq_n_f32(pC[9]);
-                        _suma0 = vdupq_n_f32(pC[10]);
-                        _sumb0 = vdupq_n_f32(pC[11]);
-                        _sum01 = _sum00;
-                        _sum11 = _sum10;
-                        _sum21 = _sum20;
-                        _sum31 = _sum30;
-                        _sum41 = _sum40;
-                        _sum51 = _sum50;
-                        _sum61 = _sum60;
-                        _sum71 = _sum70;
-                        _sum81 = _sum80;
-                        _sum91 = _sum90;
-                        _suma1 = _suma0;
-                        _sumb1 = _sumb0;
-                        pC += 12;
-                    }
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-                _sum20 = vld1q_f32(outptr + 4 * 4);
-                _sum21 = vld1q_f32(outptr + 4 * 5);
-                _sum30 = vld1q_f32(outptr + 4 * 6);
-                _sum31 = vld1q_f32(outptr + 4 * 7);
-                _sum40 = vld1q_f32(outptr + 4 * 8);
-                _sum41 = vld1q_f32(outptr + 4 * 9);
-                _sum50 = vld1q_f32(outptr + 4 * 10);
-                _sum51 = vld1q_f32(outptr + 4 * 11);
-                _sum60 = vld1q_f32(outptr + 4 * 12);
-                _sum61 = vld1q_f32(outptr + 4 * 13);
-                _sum70 = vld1q_f32(outptr + 4 * 14);
-                _sum71 = vld1q_f32(outptr + 4 * 15);
-                _sum80 = vld1q_f32(outptr + 4 * 16);
-                _sum81 = vld1q_f32(outptr + 4 * 17);
-                _sum90 = vld1q_f32(outptr + 4 * 18);
-                _sum91 = vld1q_f32(outptr + 4 * 19);
-                _suma0 = vld1q_f32(outptr + 4 * 20);
-                _suma1 = vld1q_f32(outptr + 4 * 21);
-                _sumb0 = vld1q_f32(outptr + 4 * 22);
-                _sumb1 = vld1q_f32(outptr + 4 * 23);
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x8_t _pA = vld1q_f16(pA);
-
-                float16x4_t _pB0 = vld1_f16(pB);
-                float16x4_t _pB1 = vld1_f16(pB + 4);
-                float16x4_t _pB2 = vld1_f16(pB + 8);
-
-                _sum00 = vfmlalq_lane_low_f16(_sum00, _pA, _pB0, 0);
-                _sum01 = vfmlalq_lane_high_f16(_sum01, _pA, _pB0, 0);
-                _sum10 = vfmlalq_lane_low_f16(_sum10, _pA, _pB0, 1);
-                _sum11 = vfmlalq_lane_high_f16(_sum11, _pA, _pB0, 1);
-                _sum20 = vfmlalq_lane_low_f16(_sum20, _pA, _pB0, 2);
-                _sum21 = vfmlalq_lane_high_f16(_sum21, _pA, _pB0, 2);
-                _sum30 = vfmlalq_lane_low_f16(_sum30, _pA, _pB0, 3);
-                _sum31 = vfmlalq_lane_high_f16(_sum31, _pA, _pB0, 3);
-                _sum40 = vfmlalq_lane_low_f16(_sum40, _pA, _pB1, 0);
-                _sum41 = vfmlalq_lane_high_f16(_sum41, _pA, _pB1, 0);
-                _sum50 = vfmlalq_lane_low_f16(_sum50, _pA, _pB1, 1);
-                _sum51 = vfmlalq_lane_high_f16(_sum51, _pA, _pB1, 1);
-                _sum60 = vfmlalq_lane_low_f16(_sum60, _pA, _pB1, 2);
-                _sum61 = vfmlalq_lane_high_f16(_sum61, _pA, _pB1, 2);
-                _sum70 = vfmlalq_lane_low_f16(_sum70, _pA, _pB1, 3);
-                _sum71 = vfmlalq_lane_high_f16(_sum71, _pA, _pB1, 3);
-                _sum80 = vfmlalq_lane_low_f16(_sum80, _pA, _pB2, 0);
-                _sum81 = vfmlalq_lane_high_f16(_sum81, _pA, _pB2, 0);
-                _sum90 = vfmlalq_lane_low_f16(_sum90, _pA, _pB2, 1);
-                _sum91 = vfmlalq_lane_high_f16(_sum91, _pA, _pB2, 1);
-                _suma0 = vfmlalq_lane_low_f16(_suma0, _pA, _pB2, 2);
-                _suma1 = vfmlalq_lane_high_f16(_suma1, _pA, _pB2, 2);
-                _sumb0 = vfmlalq_lane_low_f16(_sumb0, _pA, _pB2, 3);
-                _sumb1 = vfmlalq_lane_high_f16(_sumb1, _pA, _pB2, 3);
-#else
-                uint16x8_t _pA = vld1q_u16(pA);
-                float32x4_t _pA0 = vcvt_f32_f16((float16x4_t)vget_low_u16(_pA));
-                float32x4_t _pA1 = vcvt_f32_f16((float16x4_t)vget_high_u16(_pA));
-
-                float32x4_t _pB0 = vcvt_f32_f16((float16x4_t)vld1_u16(pB));
-                float32x4_t _pB1 = vcvt_f32_f16((float16x4_t)vld1_u16(pB + 4));
-                float32x4_t _pB2 = vcvt_f32_f16((float16x4_t)vld1_u16(pB + 8));
-
-                _sum00 = vfmaq_laneq_f32(_sum00, _pA0, _pB0, 0);
-                _sum01 = vfmaq_laneq_f32(_sum01, _pA1, _pB0, 0);
-                _sum10 = vfmaq_laneq_f32(_sum10, _pA0, _pB0, 1);
-                _sum11 = vfmaq_laneq_f32(_sum11, _pA1, _pB0, 1);
-                _sum20 = vfmaq_laneq_f32(_sum20, _pA0, _pB0, 2);
-                _sum21 = vfmaq_laneq_f32(_sum21, _pA1, _pB0, 2);
-                _sum30 = vfmaq_laneq_f32(_sum30, _pA0, _pB0, 3);
-                _sum31 = vfmaq_laneq_f32(_sum31, _pA1, _pB0, 3);
-                _sum40 = vfmaq_laneq_f32(_sum40, _pA0, _pB1, 0);
-                _sum41 = vfmaq_laneq_f32(_sum41, _pA1, _pB1, 0);
-                _sum50 = vfmaq_laneq_f32(_sum50, _pA0, _pB1, 1);
-                _sum51 = vfmaq_laneq_f32(_sum51, _pA1, _pB1, 1);
-                _sum60 = vfmaq_laneq_f32(_sum60, _pA0, _pB1, 2);
-                _sum61 = vfmaq_laneq_f32(_sum61, _pA1, _pB1, 2);
-                _sum70 = vfmaq_laneq_f32(_sum70, _pA0, _pB1, 3);
-                _sum71 = vfmaq_laneq_f32(_sum71, _pA1, _pB1, 3);
-                _sum80 = vfmaq_laneq_f32(_sum80, _pA0, _pB2, 0);
-                _sum81 = vfmaq_laneq_f32(_sum81, _pA1, _pB2, 0);
-                _sum90 = vfmaq_laneq_f32(_sum90, _pA0, _pB2, 1);
-                _sum91 = vfmaq_laneq_f32(_sum91, _pA1, _pB2, 1);
-                _suma0 = vfmaq_laneq_f32(_suma0, _pA0, _pB2, 2);
-                _suma1 = vfmaq_laneq_f32(_suma1, _pA1, _pB2, 2);
-                _sumb0 = vfmaq_laneq_f32(_sumb0, _pA0, _pB2, 3);
-                _sumb1 = vfmaq_laneq_f32(_sumb1, _pA1, _pB2, 3);
-#endif
-
-                pA += 8;
-                pB += 12;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum00 = vmulq_f32(_sum00, _alpha);
-                _sum01 = vmulq_f32(_sum01, _alpha);
-                _sum10 = vmulq_f32(_sum10, _alpha);
-                _sum11 = vmulq_f32(_sum11, _alpha);
-                _sum20 = vmulq_f32(_sum20, _alpha);
-                _sum21 = vmulq_f32(_sum21, _alpha);
-                _sum30 = vmulq_f32(_sum30, _alpha);
-                _sum31 = vmulq_f32(_sum31, _alpha);
-                _sum40 = vmulq_f32(_sum40, _alpha);
-                _sum41 = vmulq_f32(_sum41, _alpha);
-                _sum50 = vmulq_f32(_sum50, _alpha);
-                _sum51 = vmulq_f32(_sum51, _alpha);
-                _sum60 = vmulq_f32(_sum60, _alpha);
-                _sum61 = vmulq_f32(_sum61, _alpha);
-                _sum70 = vmulq_f32(_sum70, _alpha);
-                _sum71 = vmulq_f32(_sum71, _alpha);
-                _sum80 = vmulq_f32(_sum80, _alpha);
-                _sum81 = vmulq_f32(_sum81, _alpha);
-                _sum90 = vmulq_f32(_sum90, _alpha);
-                _sum91 = vmulq_f32(_sum91, _alpha);
-                _suma0 = vmulq_f32(_suma0, _alpha);
-                _suma1 = vmulq_f32(_suma1, _alpha);
-                _sumb0 = vmulq_f32(_sumb0, _alpha);
-                _sumb1 = vmulq_f32(_sumb1, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum00));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum10));
-                    vst1_u16(outptr0 + 4 * 2, (uint16x4_t)vcvt_f16_f32(_sum20));
-                    vst1_u16(outptr0 + 4 * 3, (uint16x4_t)vcvt_f16_f32(_sum30));
-                    vst1_u16(outptr0 + 4 * 4, (uint16x4_t)vcvt_f16_f32(_sum40));
-                    vst1_u16(outptr0 + 4 * 5, (uint16x4_t)vcvt_f16_f32(_sum50));
-                    vst1_u16(outptr0 + 4 * 6, (uint16x4_t)vcvt_f16_f32(_sum60));
-                    vst1_u16(outptr0 + 4 * 7, (uint16x4_t)vcvt_f16_f32(_sum70));
-                    vst1_u16(outptr0 + 4 * 8, (uint16x4_t)vcvt_f16_f32(_sum80));
-                    vst1_u16(outptr0 + 4 * 9, (uint16x4_t)vcvt_f16_f32(_sum90));
-                    vst1_u16(outptr0 + 4 * 10, (uint16x4_t)vcvt_f16_f32(_suma0));
-                    vst1_u16(outptr0 + 4 * 11, (uint16x4_t)vcvt_f16_f32(_sumb0));
-
-                    vst1_u16(outptr0 + out_hstep * 4, (uint16x4_t)vcvt_f16_f32(_sum01));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, (uint16x4_t)vcvt_f16_f32(_sum11));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 2, (uint16x4_t)vcvt_f16_f32(_sum21));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 3, (uint16x4_t)vcvt_f16_f32(_sum31));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 4, (uint16x4_t)vcvt_f16_f32(_sum41));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 5, (uint16x4_t)vcvt_f16_f32(_sum51));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 6, (uint16x4_t)vcvt_f16_f32(_sum61));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 7, (uint16x4_t)vcvt_f16_f32(_sum71));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 8, (uint16x4_t)vcvt_f16_f32(_sum81));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 9, (uint16x4_t)vcvt_f16_f32(_sum91));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 10, (uint16x4_t)vcvt_f16_f32(_suma1));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 11, (uint16x4_t)vcvt_f16_f32(_sumb1));
-
-                    outptr0 += 48;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose8x12_ps(_sum00, _sum01, _sum10, _sum11, _sum20, _sum21, _sum30, _sum31, _sum40, _sum41, _sum50, _sum51, _sum60, _sum61, _sum70, _sum71, _sum80, _sum81, _sum90, _sum91, _suma0, _suma1, _sumb0, _sumb1);
-
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum00));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum01));
-                    vst1_u16(outptr0 + 8, (uint16x4_t)vcvt_f16_f32(_sum10));
-                    vst1_u16(outptr0 + out_hstep, (uint16x4_t)vcvt_f16_f32(_sum11));
-                    vst1_u16(outptr0 + out_hstep + 4, (uint16x4_t)vcvt_f16_f32(_sum20));
-                    vst1_u16(outptr0 + out_hstep + 8, (uint16x4_t)vcvt_f16_f32(_sum21));
-                    vst1_u16(outptr0 + out_hstep * 2, (uint16x4_t)vcvt_f16_f32(_sum30));
-                    vst1_u16(outptr0 + out_hstep * 2 + 4, (uint16x4_t)vcvt_f16_f32(_sum31));
-                    vst1_u16(outptr0 + out_hstep * 2 + 8, (uint16x4_t)vcvt_f16_f32(_sum40));
-                    vst1_u16(outptr0 + out_hstep * 3, (uint16x4_t)vcvt_f16_f32(_sum41));
-                    vst1_u16(outptr0 + out_hstep * 3 + 4, (uint16x4_t)vcvt_f16_f32(_sum50));
-                    vst1_u16(outptr0 + out_hstep * 3 + 8, (uint16x4_t)vcvt_f16_f32(_sum51));
-                    vst1_u16(outptr0 + out_hstep * 4, (uint16x4_t)vcvt_f16_f32(_sum60));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, (uint16x4_t)vcvt_f16_f32(_sum61));
-                    vst1_u16(outptr0 + out_hstep * 4 + 8, (uint16x4_t)vcvt_f16_f32(_sum70));
-                    vst1_u16(outptr0 + out_hstep * 5, (uint16x4_t)vcvt_f16_f32(_sum71));
-                    vst1_u16(outptr0 + out_hstep * 5 + 4, (uint16x4_t)vcvt_f16_f32(_sum80));
-                    vst1_u16(outptr0 + out_hstep * 5 + 8, (uint16x4_t)vcvt_f16_f32(_sum81));
-                    vst1_u16(outptr0 + out_hstep * 6, (uint16x4_t)vcvt_f16_f32(_sum90));
-                    vst1_u16(outptr0 + out_hstep * 6 + 4, (uint16x4_t)vcvt_f16_f32(_sum91));
-                    vst1_u16(outptr0 + out_hstep * 6 + 8, (uint16x4_t)vcvt_f16_f32(_suma0));
-                    vst1_u16(outptr0 + out_hstep * 7, (uint16x4_t)vcvt_f16_f32(_suma1));
-                    vst1_u16(outptr0 + out_hstep * 7 + 4, (uint16x4_t)vcvt_f16_f32(_sumb0));
-                    vst1_u16(outptr0 + out_hstep * 7 + 8, (uint16x4_t)vcvt_f16_f32(_sumb1));
-
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                vst1q_f32(outptr + 4 * 4, _sum20);
-                vst1q_f32(outptr + 4 * 5, _sum21);
-                vst1q_f32(outptr + 4 * 6, _sum30);
-                vst1q_f32(outptr + 4 * 7, _sum31);
-                vst1q_f32(outptr + 4 * 8, _sum40);
-                vst1q_f32(outptr + 4 * 9, _sum41);
-                vst1q_f32(outptr + 4 * 10, _sum50);
-                vst1q_f32(outptr + 4 * 11, _sum51);
-                vst1q_f32(outptr + 4 * 12, _sum60);
-                vst1q_f32(outptr + 4 * 13, _sum61);
-                vst1q_f32(outptr + 4 * 14, _sum70);
-                vst1q_f32(outptr + 4 * 15, _sum71);
-                vst1q_f32(outptr + 4 * 16, _sum80);
-                vst1q_f32(outptr + 4 * 17, _sum81);
-                vst1q_f32(outptr + 4 * 18, _sum90);
-                vst1q_f32(outptr + 4 * 19, _sum91);
-                vst1q_f32(outptr + 4 * 20, _suma0);
-                vst1q_f32(outptr + 4 * 21, _suma1);
-                vst1q_f32(outptr + 4 * 22, _sumb0);
-                vst1q_f32(outptr + 4 * 23, _sumb1);
-            }
-
-            outptr += 96;
-        }
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum20;
-            float32x4_t _sum21;
-            float32x4_t _sum30;
-            float32x4_t _sum31;
-            float32x4_t _sum40;
-            float32x4_t _sum41;
-            float32x4_t _sum50;
-            float32x4_t _sum51;
-            float32x4_t _sum60;
-            float32x4_t _sum61;
-            float32x4_t _sum70;
-            float32x4_t _sum71;
-
-            if (k == 0)
-            {
-                _sum00 = vdupq_n_f32(0.f);
-                _sum01 = vdupq_n_f32(0.f);
-                _sum10 = vdupq_n_f32(0.f);
-                _sum11 = vdupq_n_f32(0.f);
-                _sum20 = vdupq_n_f32(0.f);
-                _sum21 = vdupq_n_f32(0.f);
-                _sum30 = vdupq_n_f32(0.f);
-                _sum31 = vdupq_n_f32(0.f);
-                _sum40 = vdupq_n_f32(0.f);
-                _sum41 = vdupq_n_f32(0.f);
-                _sum50 = vdupq_n_f32(0.f);
-                _sum51 = vdupq_n_f32(0.f);
-                _sum60 = vdupq_n_f32(0.f);
-                _sum61 = vdupq_n_f32(0.f);
-                _sum70 = vdupq_n_f32(0.f);
-                _sum71 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum10 = _sum00;
-                        _sum11 = _sum00;
-                        _sum20 = _sum00;
-                        _sum21 = _sum00;
-                        _sum30 = _sum00;
-                        _sum31 = _sum00;
-                        _sum40 = _sum00;
-                        _sum41 = _sum00;
-                        _sum50 = _sum00;
-                        _sum51 = _sum00;
-                        _sum60 = _sum00;
-                        _sum61 = _sum00;
-                        _sum70 = _sum00;
-                        _sum71 = _sum00;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                        _sum10 = _sum00;
-                        _sum11 = _sum01;
-                        _sum20 = _sum00;
-                        _sum21 = _sum01;
-                        _sum30 = _sum00;
-                        _sum31 = _sum01;
-                        _sum40 = _sum00;
-                        _sum41 = _sum01;
-                        _sum50 = _sum00;
-                        _sum51 = _sum01;
-                        _sum60 = _sum00;
-                        _sum61 = _sum01;
-                        _sum70 = _sum00;
-                        _sum71 = _sum01;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4 * 1);
-                        _sum10 = vld1q_f32(pC + 4 * 2);
-                        _sum11 = vld1q_f32(pC + 4 * 3);
-                        _sum20 = vld1q_f32(pC + 4 * 4);
-                        _sum21 = vld1q_f32(pC + 4 * 5);
-                        _sum30 = vld1q_f32(pC + 4 * 6);
-                        _sum31 = vld1q_f32(pC + 4 * 7);
-                        _sum40 = vld1q_f32(pC + 4 * 8);
-                        _sum41 = vld1q_f32(pC + 4 * 9);
-                        _sum50 = vld1q_f32(pC + 4 * 10);
-                        _sum51 = vld1q_f32(pC + 4 * 11);
-                        _sum60 = vld1q_f32(pC + 4 * 12);
-                        _sum61 = vld1q_f32(pC + 4 * 13);
-                        _sum70 = vld1q_f32(pC + 4 * 14);
-                        _sum71 = vld1q_f32(pC + 4 * 15);
-                        pC += 64;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum10 = vdupq_n_f32(pC[1]);
-                        _sum20 = vdupq_n_f32(pC[2]);
-                        _sum30 = vdupq_n_f32(pC[3]);
-                        _sum40 = vdupq_n_f32(pC[4]);
-                        _sum50 = vdupq_n_f32(pC[5]);
-                        _sum60 = vdupq_n_f32(pC[6]);
-                        _sum70 = vdupq_n_f32(pC[7]);
-                        _sum01 = _sum00;
-                        _sum11 = _sum10;
-                        _sum21 = _sum20;
-                        _sum31 = _sum30;
-                        _sum41 = _sum40;
-                        _sum51 = _sum50;
-                        _sum61 = _sum60;
-                        _sum71 = _sum70;
-                        pC += 8;
-                    }
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-                _sum20 = vld1q_f32(outptr + 4 * 4);
-                _sum21 = vld1q_f32(outptr + 4 * 5);
-                _sum30 = vld1q_f32(outptr + 4 * 6);
-                _sum31 = vld1q_f32(outptr + 4 * 7);
-                _sum40 = vld1q_f32(outptr + 4 * 8);
-                _sum41 = vld1q_f32(outptr + 4 * 9);
-                _sum50 = vld1q_f32(outptr + 4 * 10);
-                _sum51 = vld1q_f32(outptr + 4 * 11);
-                _sum60 = vld1q_f32(outptr + 4 * 12);
-                _sum61 = vld1q_f32(outptr + 4 * 13);
-                _sum70 = vld1q_f32(outptr + 4 * 14);
-                _sum71 = vld1q_f32(outptr + 4 * 15);
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x8_t _pA = vld1q_f16(pA);
-                float16x8_t _pB = vld1q_f16(pB);
-
-                _sum00 = vfmlalq_laneq_low_f16(_sum00, _pA, _pB, 0);
-                _sum01 = vfmlalq_laneq_high_f16(_sum01, _pA, _pB, 0);
-                _sum10 = vfmlalq_laneq_low_f16(_sum10, _pA, _pB, 1);
-                _sum11 = vfmlalq_laneq_high_f16(_sum11, _pA, _pB, 1);
-                _sum20 = vfmlalq_laneq_low_f16(_sum20, _pA, _pB, 2);
-                _sum21 = vfmlalq_laneq_high_f16(_sum21, _pA, _pB, 2);
-                _sum30 = vfmlalq_laneq_low_f16(_sum30, _pA, _pB, 3);
-                _sum31 = vfmlalq_laneq_high_f16(_sum31, _pA, _pB, 3);
-                _sum40 = vfmlalq_laneq_low_f16(_sum40, _pA, _pB, 4);
-                _sum41 = vfmlalq_laneq_high_f16(_sum41, _pA, _pB, 4);
-                _sum50 = vfmlalq_laneq_low_f16(_sum50, _pA, _pB, 5);
-                _sum51 = vfmlalq_laneq_high_f16(_sum51, _pA, _pB, 5);
-                _sum60 = vfmlalq_laneq_low_f16(_sum60, _pA, _pB, 6);
-                _sum61 = vfmlalq_laneq_high_f16(_sum61, _pA, _pB, 6);
-                _sum70 = vfmlalq_laneq_low_f16(_sum70, _pA, _pB, 7);
-                _sum71 = vfmlalq_laneq_high_f16(_sum71, _pA, _pB, 7);
-#else
-                uint16x8_t _pA = vld1q_u16(pA);
-                float32x4_t _pA0 = vcvt_f32_f16((float16x4_t)vget_low_u16(_pA));
-                float32x4_t _pA1 = vcvt_f32_f16((float16x4_t)vget_high_u16(_pA));
-
-                float32x4_t _pB0 = vcvt_f32_f16((float16x4_t)vld1_u16(pB));
-                float32x4_t _pB1 = vcvt_f32_f16((float16x4_t)vld1_u16(pB + 4));
-
-                _sum00 = vfmaq_laneq_f32(_sum00, _pA0, _pB0, 0);
-                _sum01 = vfmaq_laneq_f32(_sum01, _pA1, _pB0, 0);
-                _sum10 = vfmaq_laneq_f32(_sum10, _pA0, _pB0, 1);
-                _sum11 = vfmaq_laneq_f32(_sum11, _pA1, _pB0, 1);
-                _sum20 = vfmaq_laneq_f32(_sum20, _pA0, _pB0, 2);
-                _sum21 = vfmaq_laneq_f32(_sum21, _pA1, _pB0, 2);
-                _sum30 = vfmaq_laneq_f32(_sum30, _pA0, _pB0, 3);
-                _sum31 = vfmaq_laneq_f32(_sum31, _pA1, _pB0, 3);
-                _sum40 = vfmaq_laneq_f32(_sum40, _pA0, _pB1, 0);
-                _sum41 = vfmaq_laneq_f32(_sum41, _pA1, _pB1, 0);
-                _sum50 = vfmaq_laneq_f32(_sum50, _pA0, _pB1, 1);
-                _sum51 = vfmaq_laneq_f32(_sum51, _pA1, _pB1, 1);
-                _sum60 = vfmaq_laneq_f32(_sum60, _pA0, _pB1, 2);
-                _sum61 = vfmaq_laneq_f32(_sum61, _pA1, _pB1, 2);
-                _sum70 = vfmaq_laneq_f32(_sum70, _pA0, _pB1, 3);
-                _sum71 = vfmaq_laneq_f32(_sum71, _pA1, _pB1, 3);
-#endif
-
-                pA += 8;
-                pB += 8;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum00 = vmulq_f32(_sum00, _alpha);
-                _sum01 = vmulq_f32(_sum01, _alpha);
-                _sum10 = vmulq_f32(_sum10, _alpha);
-                _sum11 = vmulq_f32(_sum11, _alpha);
-                _sum20 = vmulq_f32(_sum20, _alpha);
-                _sum21 = vmulq_f32(_sum21, _alpha);
-                _sum30 = vmulq_f32(_sum30, _alpha);
-                _sum31 = vmulq_f32(_sum31, _alpha);
-                _sum40 = vmulq_f32(_sum40, _alpha);
-                _sum41 = vmulq_f32(_sum41, _alpha);
-                _sum50 = vmulq_f32(_sum50, _alpha);
-                _sum51 = vmulq_f32(_sum51, _alpha);
-                _sum60 = vmulq_f32(_sum60, _alpha);
-                _sum61 = vmulq_f32(_sum61, _alpha);
-                _sum70 = vmulq_f32(_sum70, _alpha);
-                _sum71 = vmulq_f32(_sum71, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum00));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum10));
-                    vst1_u16(outptr0 + 4 * 2, (uint16x4_t)vcvt_f16_f32(_sum20));
-                    vst1_u16(outptr0 + 4 * 3, (uint16x4_t)vcvt_f16_f32(_sum30));
-                    vst1_u16(outptr0 + 4 * 4, (uint16x4_t)vcvt_f16_f32(_sum40));
-                    vst1_u16(outptr0 + 4 * 5, (uint16x4_t)vcvt_f16_f32(_sum50));
-                    vst1_u16(outptr0 + 4 * 6, (uint16x4_t)vcvt_f16_f32(_sum60));
-                    vst1_u16(outptr0 + 4 * 7, (uint16x4_t)vcvt_f16_f32(_sum70));
-
-                    vst1_u16(outptr0 + out_hstep * 4, (uint16x4_t)vcvt_f16_f32(_sum01));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, (uint16x4_t)vcvt_f16_f32(_sum11));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 2, (uint16x4_t)vcvt_f16_f32(_sum21));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 3, (uint16x4_t)vcvt_f16_f32(_sum31));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 4, (uint16x4_t)vcvt_f16_f32(_sum41));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 5, (uint16x4_t)vcvt_f16_f32(_sum51));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 6, (uint16x4_t)vcvt_f16_f32(_sum61));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 7, (uint16x4_t)vcvt_f16_f32(_sum71));
-
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose8x8_ps(_sum00, _sum01, _sum10, _sum11, _sum20, _sum21, _sum30, _sum31, _sum40, _sum41, _sum50, _sum51, _sum60, _sum61, _sum70, _sum71);
-
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum00));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum01));
-                    vst1_u16(outptr0 + out_hstep, (uint16x4_t)vcvt_f16_f32(_sum10));
-                    vst1_u16(outptr0 + out_hstep + 4, (uint16x4_t)vcvt_f16_f32(_sum11));
-                    vst1_u16(outptr0 + out_hstep * 2, (uint16x4_t)vcvt_f16_f32(_sum20));
-                    vst1_u16(outptr0 + out_hstep * 2 + 4, (uint16x4_t)vcvt_f16_f32(_sum21));
-                    vst1_u16(outptr0 + out_hstep * 3, (uint16x4_t)vcvt_f16_f32(_sum30));
-                    vst1_u16(outptr0 + out_hstep * 3 + 4, (uint16x4_t)vcvt_f16_f32(_sum31));
-                    vst1_u16(outptr0 + out_hstep * 4, (uint16x4_t)vcvt_f16_f32(_sum40));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, (uint16x4_t)vcvt_f16_f32(_sum41));
-                    vst1_u16(outptr0 + out_hstep * 5, (uint16x4_t)vcvt_f16_f32(_sum50));
-                    vst1_u16(outptr0 + out_hstep * 5 + 4, (uint16x4_t)vcvt_f16_f32(_sum51));
-                    vst1_u16(outptr0 + out_hstep * 6, (uint16x4_t)vcvt_f16_f32(_sum60));
-                    vst1_u16(outptr0 + out_hstep * 6 + 4, (uint16x4_t)vcvt_f16_f32(_sum61));
-                    vst1_u16(outptr0 + out_hstep * 7, (uint16x4_t)vcvt_f16_f32(_sum70));
-                    vst1_u16(outptr0 + out_hstep * 7 + 4, (uint16x4_t)vcvt_f16_f32(_sum71));
-
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                vst1q_f32(outptr + 4 * 4, _sum20);
-                vst1q_f32(outptr + 4 * 5, _sum21);
-                vst1q_f32(outptr + 4 * 6, _sum30);
-                vst1q_f32(outptr + 4 * 7, _sum31);
-                vst1q_f32(outptr + 4 * 8, _sum40);
-                vst1q_f32(outptr + 4 * 9, _sum41);
-                vst1q_f32(outptr + 4 * 10, _sum50);
-                vst1q_f32(outptr + 4 * 11, _sum51);
-                vst1q_f32(outptr + 4 * 12, _sum60);
-                vst1q_f32(outptr + 4 * 13, _sum61);
-                vst1q_f32(outptr + 4 * 14, _sum70);
-                vst1q_f32(outptr + 4 * 15, _sum71);
-            }
-
-            outptr += 64;
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum20;
-            float32x4_t _sum21;
-            float32x4_t _sum30;
-            float32x4_t _sum31;
-
-            if (k == 0)
-            {
-                _sum00 = vdupq_n_f32(0.f);
-                _sum01 = vdupq_n_f32(0.f);
-                _sum10 = vdupq_n_f32(0.f);
-                _sum11 = vdupq_n_f32(0.f);
-                _sum20 = vdupq_n_f32(0.f);
-                _sum21 = vdupq_n_f32(0.f);
-                _sum30 = vdupq_n_f32(0.f);
-                _sum31 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum10 = _sum00;
-                        _sum11 = _sum00;
-                        _sum20 = _sum00;
-                        _sum21 = _sum00;
-                        _sum30 = _sum00;
-                        _sum31 = _sum00;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                        _sum10 = _sum00;
-                        _sum11 = _sum01;
-                        _sum20 = _sum00;
-                        _sum21 = _sum01;
-                        _sum30 = _sum00;
-                        _sum31 = _sum01;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4 * 1);
-                        _sum10 = vld1q_f32(pC + 4 * 2);
-                        _sum11 = vld1q_f32(pC + 4 * 3);
-                        _sum20 = vld1q_f32(pC + 4 * 4);
-                        _sum21 = vld1q_f32(pC + 4 * 5);
-                        _sum30 = vld1q_f32(pC + 4 * 6);
-                        _sum31 = vld1q_f32(pC + 4 * 7);
-                        pC += 32;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum10 = vdupq_n_f32(pC[1]);
-                        _sum20 = vdupq_n_f32(pC[2]);
-                        _sum30 = vdupq_n_f32(pC[3]);
-                        _sum01 = _sum00;
-                        _sum11 = _sum10;
-                        _sum21 = _sum20;
-                        _sum31 = _sum30;
-                        pC += 4;
-                    }
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-                _sum20 = vld1q_f32(outptr + 4 * 4);
-                _sum21 = vld1q_f32(outptr + 4 * 5);
-                _sum30 = vld1q_f32(outptr + 4 * 6);
-                _sum31 = vld1q_f32(outptr + 4 * 7);
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x8_t _pA = vld1q_f16(pA);
-                float16x4_t _pB = vld1_f16(pB);
-
-                _sum00 = vfmlalq_lane_low_f16(_sum00, _pA, _pB, 0);
-                _sum01 = vfmlalq_lane_high_f16(_sum01, _pA, _pB, 0);
-                _sum10 = vfmlalq_lane_low_f16(_sum10, _pA, _pB, 1);
-                _sum11 = vfmlalq_lane_high_f16(_sum11, _pA, _pB, 1);
-                _sum20 = vfmlalq_lane_low_f16(_sum20, _pA, _pB, 2);
-                _sum21 = vfmlalq_lane_high_f16(_sum21, _pA, _pB, 2);
-                _sum30 = vfmlalq_lane_low_f16(_sum30, _pA, _pB, 3);
-                _sum31 = vfmlalq_lane_high_f16(_sum31, _pA, _pB, 3);
-#else
-                uint16x8_t _pA = vld1q_u16(pA);
-                float32x4_t _pA0 = vcvt_f32_f16((float16x4_t)vget_low_u16(_pA));
-                float32x4_t _pA1 = vcvt_f32_f16((float16x4_t)vget_high_u16(_pA));
-
-                float32x4_t _pB0 = vcvt_f32_f16((float16x4_t)vld1_u16(pB));
-
-                _sum00 = vfmaq_laneq_f32(_sum00, _pA0, _pB0, 0);
-                _sum01 = vfmaq_laneq_f32(_sum01, _pA1, _pB0, 0);
-                _sum10 = vfmaq_laneq_f32(_sum10, _pA0, _pB0, 1);
-                _sum11 = vfmaq_laneq_f32(_sum11, _pA1, _pB0, 1);
-                _sum20 = vfmaq_laneq_f32(_sum20, _pA0, _pB0, 2);
-                _sum21 = vfmaq_laneq_f32(_sum21, _pA1, _pB0, 2);
-                _sum30 = vfmaq_laneq_f32(_sum30, _pA0, _pB0, 3);
-                _sum31 = vfmaq_laneq_f32(_sum31, _pA1, _pB0, 3);
-#endif
-
-                pA += 8;
-                pB += 4;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum00 = vmulq_f32(_sum00, _alpha);
-                _sum01 = vmulq_f32(_sum01, _alpha);
-                _sum10 = vmulq_f32(_sum10, _alpha);
-                _sum11 = vmulq_f32(_sum11, _alpha);
-                _sum20 = vmulq_f32(_sum20, _alpha);
-                _sum21 = vmulq_f32(_sum21, _alpha);
-                _sum30 = vmulq_f32(_sum30, _alpha);
-                _sum31 = vmulq_f32(_sum31, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum00));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum10));
-                    vst1_u16(outptr0 + 4 * 2, (uint16x4_t)vcvt_f16_f32(_sum20));
-                    vst1_u16(outptr0 + 4 * 3, (uint16x4_t)vcvt_f16_f32(_sum30));
-
-                    vst1_u16(outptr0 + out_hstep * 4, (uint16x4_t)vcvt_f16_f32(_sum01));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, (uint16x4_t)vcvt_f16_f32(_sum11));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 2, (uint16x4_t)vcvt_f16_f32(_sum21));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4 * 3, (uint16x4_t)vcvt_f16_f32(_sum31));
-
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose8x4_ps(_sum00, _sum01, _sum10, _sum11, _sum20, _sum21, _sum30, _sum31);
-
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum00));
-                    vst1_u16(outptr0 + out_hstep * 1, (uint16x4_t)vcvt_f16_f32(_sum01));
-                    vst1_u16(outptr0 + out_hstep * 2, (uint16x4_t)vcvt_f16_f32(_sum10));
-                    vst1_u16(outptr0 + out_hstep * 3, (uint16x4_t)vcvt_f16_f32(_sum11));
-                    vst1_u16(outptr0 + out_hstep * 4, (uint16x4_t)vcvt_f16_f32(_sum20));
-                    vst1_u16(outptr0 + out_hstep * 5, (uint16x4_t)vcvt_f16_f32(_sum21));
-                    vst1_u16(outptr0 + out_hstep * 6, (uint16x4_t)vcvt_f16_f32(_sum30));
-                    vst1_u16(outptr0 + out_hstep * 7, (uint16x4_t)vcvt_f16_f32(_sum31));
-
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-                vst1q_f32(outptr + 4 * 4, _sum20);
-                vst1q_f32(outptr + 4 * 5, _sum21);
-                vst1q_f32(outptr + 4 * 6, _sum30);
-                vst1q_f32(outptr + 4 * 7, _sum31);
-            }
-
-            outptr += 32;
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-
-            if (k == 0)
-            {
-                _sum00 = vdupq_n_f32(0.f);
-                _sum01 = vdupq_n_f32(0.f);
-                _sum10 = vdupq_n_f32(0.f);
-                _sum11 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum10 = _sum00;
-                        _sum11 = _sum00;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                        _sum10 = _sum00;
-                        _sum11 = _sum01;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4 * 1);
-                        _sum10 = vld1q_f32(pC + 4 * 2);
-                        _sum11 = vld1q_f32(pC + 4 * 3);
-                        pC += 16;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum10 = vdupq_n_f32(pC[1]);
-                        _sum01 = _sum00;
-                        _sum11 = _sum10;
-                        pC += 2;
-                    }
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4 * 1);
-                _sum10 = vld1q_f32(outptr + 4 * 2);
-                _sum11 = vld1q_f32(outptr + 4 * 3);
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x8_t _pA = vld1q_f16(pA);
-                float16x4_t _pB0 = vdup_n_f16(pB[0]);
-                float16x4_t _pB1 = vdup_n_f16(pB[1]);
-                float16x8_t _pB01 = vcombine_f16(_pB0, _pB1);
-                float16x8_t _pB10 = vcombine_f16(_pB1, _pB0);
-
-                _sum00 = vfmlalq_low_f16(_sum00, _pA, _pB01);
-                _sum01 = vfmlalq_high_f16(_sum01, _pA, _pB10);
-                _sum10 = vfmlalq_low_f16(_sum10, _pA, _pB10);
-                _sum11 = vfmlalq_high_f16(_sum11, _pA, _pB01);
-#else
-                uint16x8_t _pA = vld1q_u16(pA);
-                float32x4_t _pA0 = vcvt_f32_f16((float16x4_t)vget_low_u16(_pA));
-                float32x4_t _pA1 = vcvt_f32_f16((float16x4_t)vget_high_u16(_pA));
-
-                float32x4_t _pB0 = vcvt_f32_f16((float16x4_t)vdup_n_u16(pB[0]));
-                float32x4_t _pB1 = vcvt_f32_f16((float16x4_t)vdup_n_u16(pB[1]));
-
-                _sum00 = vfmaq_f32(_sum00, _pA0, _pB0);
-                _sum01 = vfmaq_f32(_sum01, _pA1, _pB0);
-                _sum10 = vfmaq_f32(_sum10, _pA0, _pB1);
-                _sum11 = vfmaq_f32(_sum11, _pA1, _pB1);
-#endif
-
-                pA += 8;
-                pB += 2;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum00 = vmulq_f32(_sum00, _alpha);
-                _sum01 = vmulq_f32(_sum01, _alpha);
-                _sum10 = vmulq_f32(_sum10, _alpha);
-                _sum11 = vmulq_f32(_sum11, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum00));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum10));
-
-                    vst1_u16(outptr0 + out_hstep * 4, (uint16x4_t)vcvt_f16_f32(_sum01));
-                    vst1_u16(outptr0 + out_hstep * 4 + 4, (uint16x4_t)vcvt_f16_f32(_sum11));
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short sum0[8];
-                    unsigned short sum1[8];
-                    vst1_u16(sum0, (uint16x4_t)vcvt_f16_f32(_sum00));
-                    vst1_u16(sum0 + 4, (uint16x4_t)vcvt_f16_f32(_sum01));
-                    vst1_u16(sum1, (uint16x4_t)vcvt_f16_f32(_sum10));
-                    vst1_u16(sum1 + 4, (uint16x4_t)vcvt_f16_f32(_sum11));
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[out_hstep * 4] = sum0[4];
-                    outptr0[out_hstep * 5] = sum0[5];
-                    outptr0[out_hstep * 6] = sum0[6];
-                    outptr0[out_hstep * 7] = sum0[7];
-
-                    outptr0[1] = sum1[0];
-                    outptr0[out_hstep + 1] = sum1[1];
-                    outptr0[out_hstep * 2 + 1] = sum1[2];
-                    outptr0[out_hstep * 3 + 1] = sum1[3];
-                    outptr0[out_hstep * 4 + 1] = sum1[4];
-                    outptr0[out_hstep * 5 + 1] = sum1[5];
-                    outptr0[out_hstep * 6 + 1] = sum1[6];
-                    outptr0[out_hstep * 7 + 1] = sum1[7];
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-                vst1q_f32(outptr + 4 * 2, _sum10);
-                vst1q_f32(outptr + 4 * 3, _sum11);
-            }
-
-            outptr += 16;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-
-            if (k == 0)
-            {
-                _sum00 = vdupq_n_f32(0.f);
-                _sum01 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                        pC += 8;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        pC += 1;
-                    }
-                }
-            }
-            else
-            {
-                _sum00 = vld1q_f32(outptr);
-                _sum01 = vld1q_f32(outptr + 4);
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x8_t _pA = vld1q_f16(pA);
-                float16x8_t _pB = vdupq_n_f16(pB[0]);
-
-                _sum00 = vfmlalq_low_f16(_sum00, _pA, _pB);
-                _sum01 = vfmlalq_high_f16(_sum01, _pA, _pB);
-#else
-                uint16x8_t _pA = vld1q_u16(pA);
-                float32x4_t _pA0 = vcvt_f32_f16((float16x4_t)vget_low_u16(_pA));
-                float32x4_t _pA1 = vcvt_f32_f16((float16x4_t)vget_high_u16(_pA));
-
-                float32x4_t _pB = vcvt_f32_f16((float16x4_t)vld1_dup_u16(pB));
-
-                _sum00 = vfmaq_f32(_sum00, _pA0, _pB);
-                _sum01 = vfmaq_f32(_sum01, _pA1, _pB);
-#endif
-
-                pA += 8;
-                pB += 1;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum00 = vmulq_f32(_sum00, _alpha);
-                _sum01 = vmulq_f32(_sum01, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum00));
-                    vst1_u16(outptr0 + out_hstep * 4, (uint16x4_t)vcvt_f16_f32(_sum01));
-                    outptr0 += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short sum0[8];
-                    vst1_u16(sum0, (uint16x4_t)vcvt_f16_f32(_sum00));
-                    vst1_u16(sum0 + 4, (uint16x4_t)vcvt_f16_f32(_sum01));
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep * 1] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[out_hstep * 4] = sum0[4];
-                    outptr0[out_hstep * 5] = sum0[5];
-                    outptr0[out_hstep * 6] = sum0[6];
-                    outptr0[out_hstep * 7] = sum0[7];
-                    outptr0++;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum00);
-                vst1q_f32(outptr + 4, _sum01);
-            }
-
-            outptr += 8;
-        }
-
-        pAT += max_kk * 8;
-    }
-#endif // __aarch64__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        unsigned short* outptr0 = (unsigned short*)top_blob + (i + ii) * out_hstep + j * out_elempack;
-
-#if __ARM_FEATURE_FP16_FML
-        const __fp16* pB = pBT;
-#else
-        const unsigned short* pB = pBT;
-#endif
-
-        if (pC)
-        {
-            if (broadcast_type_C == 1 || broadcast_type_C == 2)
-            {
-                pC = (const float*)CT_tile + i + ii;
-            }
-            if (broadcast_type_C == 4)
-            {
-                pC = (const float*)CT_tile + j;
-            }
-        }
-
-        int jj = 0;
-#if __aarch64__
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-            float32x4_t _sum3;
-            float32x4_t _sum4;
-            float32x4_t _sum5;
-            float32x4_t _sum6;
-            float32x4_t _sum7;
-            float32x4_t _sum8;
-            float32x4_t _sum9;
-            float32x4_t _suma;
-            float32x4_t _sumb;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-                _sum2 = vdupq_n_f32(0.f);
-                _sum3 = vdupq_n_f32(0.f);
-                _sum4 = vdupq_n_f32(0.f);
-                _sum5 = vdupq_n_f32(0.f);
-                _sum6 = vdupq_n_f32(0.f);
-                _sum7 = vdupq_n_f32(0.f);
-                _sum8 = vdupq_n_f32(0.f);
-                _sum9 = vdupq_n_f32(0.f);
-                _suma = vdupq_n_f32(0.f);
-                _sumb = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = _sum0;
-                        _sum2 = _sum0;
-                        _sum3 = _sum0;
-                        _sum4 = _sum0;
-                        _sum5 = _sum0;
-                        _sum6 = _sum0;
-                        _sum7 = _sum0;
-                        _sum8 = _sum0;
-                        _sum9 = _sum0;
-                        _suma = _sum0;
-                        _sumb = _sum0;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = _sum0;
-                        _sum2 = _sum0;
-                        _sum3 = _sum0;
-                        _sum4 = _sum0;
-                        _sum5 = _sum0;
-                        _sum6 = _sum0;
-                        _sum7 = _sum0;
-                        _sum8 = _sum0;
-                        _sum9 = _sum0;
-                        _suma = _sum0;
-                        _sumb = _sum0;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = vld1q_f32(pC + 4);
-                        _sum2 = vld1q_f32(pC + 8);
-                        _sum3 = vld1q_f32(pC + 12);
-                        _sum4 = vld1q_f32(pC + 16);
-                        _sum5 = vld1q_f32(pC + 20);
-                        _sum6 = vld1q_f32(pC + 24);
-                        _sum7 = vld1q_f32(pC + 28);
-                        _sum8 = vld1q_f32(pC + 32);
-                        _sum9 = vld1q_f32(pC + 36);
-                        _suma = vld1q_f32(pC + 40);
-                        _sumb = vld1q_f32(pC + 44);
-                        pC += 48;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = vdupq_n_f32(pC[1]);
-                        _sum2 = vdupq_n_f32(pC[2]);
-                        _sum3 = vdupq_n_f32(pC[3]);
-                        _sum4 = vdupq_n_f32(pC[4]);
-                        _sum5 = vdupq_n_f32(pC[5]);
-                        _sum6 = vdupq_n_f32(pC[6]);
-                        _sum7 = vdupq_n_f32(pC[7]);
-                        _sum8 = vdupq_n_f32(pC[8]);
-                        _sum9 = vdupq_n_f32(pC[9]);
-                        _suma = vdupq_n_f32(pC[10]);
-                        _sumb = vdupq_n_f32(pC[11]);
-                        pC += 12;
-                    }
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4 * 1);
-                _sum2 = vld1q_f32(outptr + 4 * 2);
-                _sum3 = vld1q_f32(outptr + 4 * 3);
-                _sum4 = vld1q_f32(outptr + 4 * 4);
-                _sum5 = vld1q_f32(outptr + 4 * 5);
-                _sum6 = vld1q_f32(outptr + 4 * 6);
-                _sum7 = vld1q_f32(outptr + 4 * 7);
-                _sum8 = vld1q_f32(outptr + 4 * 8);
-                _sum9 = vld1q_f32(outptr + 4 * 9);
-                _suma = vld1q_f32(outptr + 4 * 10);
-                _sumb = vld1q_f32(outptr + 4 * 11);
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x4_t _pA = vld1_f16(pA);
-                float16x8_t _pAA = vcombine_f16(_pA, _pA);
-
-                float16x4_t _pB0 = vld1_f16(pB);
-                float16x4_t _pB1 = vld1_f16(pB + 4);
-                float16x4_t _pB2 = vld1_f16(pB + 8);
-
-                _sum0 = vfmlalq_lane_low_f16(_sum0, _pAA, _pB0, 0);
-                _sum1 = vfmlalq_lane_low_f16(_sum1, _pAA, _pB0, 1);
-                _sum2 = vfmlalq_lane_low_f16(_sum2, _pAA, _pB0, 2);
-                _sum3 = vfmlalq_lane_low_f16(_sum3, _pAA, _pB0, 3);
-                _sum4 = vfmlalq_lane_low_f16(_sum4, _pAA, _pB1, 0);
-                _sum5 = vfmlalq_lane_low_f16(_sum5, _pAA, _pB1, 1);
-                _sum6 = vfmlalq_lane_low_f16(_sum6, _pAA, _pB1, 2);
-                _sum7 = vfmlalq_lane_low_f16(_sum7, _pAA, _pB1, 3);
-                _sum8 = vfmlalq_lane_low_f16(_sum8, _pAA, _pB2, 0);
-                _sum9 = vfmlalq_lane_low_f16(_sum9, _pAA, _pB2, 1);
-                _suma = vfmlalq_lane_low_f16(_suma, _pAA, _pB2, 2);
-                _sumb = vfmlalq_lane_low_f16(_sumb, _pAA, _pB2, 3);
-
-                pA += 4;
-                pB += 12;
-#else
-#if __aarch64__
-                float32x4_t _pA = vcvt_f32_f16((float16x4_t)vld1_u16(pA));
-                float32x4_t _pB0 = vcvt_f32_f16((float16x4_t)vld1_u16(pB));
-                float32x4_t _pB1 = vcvt_f32_f16((float16x4_t)vld1_u16(pB + 4));
-                float32x4_t _pB2 = vcvt_f32_f16((float16x4_t)vld1_u16(pB + 8));
-
-                _sum0 = vfmaq_laneq_f32(_sum0, _pA, _pB0, 0);
-                _sum1 = vfmaq_laneq_f32(_sum1, _pA, _pB0, 1);
-                _sum2 = vfmaq_laneq_f32(_sum2, _pA, _pB0, 2);
-                _sum3 = vfmaq_laneq_f32(_sum3, _pA, _pB0, 3);
-                _sum4 = vfmaq_laneq_f32(_sum4, _pA, _pB1, 0);
-                _sum5 = vfmaq_laneq_f32(_sum5, _pA, _pB1, 1);
-                _sum6 = vfmaq_laneq_f32(_sum6, _pA, _pB1, 2);
-                _sum7 = vfmaq_laneq_f32(_sum7, _pA, _pB1, 3);
-                _sum8 = vfmaq_laneq_f32(_sum8, _pA, _pB2, 0);
-                _sum9 = vfmaq_laneq_f32(_sum9, _pA, _pB2, 1);
-                _suma = vfmaq_laneq_f32(_suma, _pA, _pB2, 2);
-                _sumb = vfmaq_laneq_f32(_sumb, _pA, _pB2, 3);
-
-                pA += 4;
-                pB += 12;
-#else // __aarch64__
-#if NCNN_GNU_INLINE_ASM
-                asm volatile(
-                    "pld        [%0, #64]       \n"
-                    "pld        [%1, #192]      \n"
-                    "vld1.u16   {d6}, [%0 :64]! \n"
-                    "vld1.u16   {d2-d4}, [%1 :64]! \n"
-                    "vcvt.f32.f16 q3, d6        \n"
-                    "vcvt.f32.f16 q0, d2        \n"
-                    "vcvt.f32.f16 q1, d3        \n"
-                    "vcvt.f32.f16 q2, d4        \n"
-                    "vmla.f32   %q2, q3, d0[0]  \n"
-                    "vmla.f32   %q3, q3, d0[1]  \n"
-                    "vmla.f32   %q4, q3, d1[0]  \n"
-                    "vmla.f32   %q5, q3, d1[1]  \n"
-                    "vmla.f32   %q6, q3, d2[0]  \n"
-                    "vmla.f32   %q7, q3, d2[1]  \n"
-                    "vmla.f32   %q8, q3, d3[0]  \n"
-                    "vmla.f32   %q9, q3, d3[1]  \n"
-                    "vmla.f32   %q10, q3, d4[0] \n"
-                    "vmla.f32   %q11, q3, d4[1] \n"
-                    "vmla.f32   %q12, q3, d5[0] \n"
-                    "vmla.f32   %q13, q3, d5[1] \n"
-                    : "=r"(pA),
-                    "=r"(pB),
-                    "=w"(_sum0),
-                    "=w"(_sum1),
-                    "=w"(_sum2),
-                    "=w"(_sum3),
-                    "=w"(_sum4),
-                    "=w"(_sum5),
-                    "=w"(_sum6),
-                    "=w"(_sum7),
-                    "=w"(_sum8),
-                    "=w"(_sum9),
-                    "=w"(_suma),
-                    "=w"(_sumb)
-                    : "0"(pA),
-                    "1"(pB),
-                    "2"(_sum0),
-                    "3"(_sum1),
-                    "4"(_sum2),
-                    "5"(_sum3),
-                    "6"(_sum4),
-                    "7"(_sum5),
-                    "8"(_sum6),
-                    "9"(_sum7),
-                    "10"(_sum8),
-                    "11"(_sum9),
-                    "12"(_suma),
-                    "13"(_sumb)
-                    : "memory", "q0", "q1", "q2", "q3");
-#else
-                float32x4_t _pA = vcvt_f32_f16((float16x4_t)vld1_u16(pA));
-                float32x4_t _pB0 = vcvt_f32_f16((float16x4_t)vld1_u16(pB));
-                float32x4_t _pB1 = vcvt_f32_f16((float16x4_t)vld1_u16(pB + 4));
-                float32x4_t _pB2 = vcvt_f32_f16((float16x4_t)vld1_u16(pB + 8));
-
-                _sum0 = vmlaq_lane_f32(_sum0, _pA, vget_low_f32(_pB0), 0);
-                _sum1 = vmlaq_lane_f32(_sum1, _pA, vget_low_f32(_pB0), 1);
-                _sum2 = vmlaq_lane_f32(_sum2, _pA, vget_high_f32(_pB0), 0);
-                _sum3 = vmlaq_lane_f32(_sum3, _pA, vget_high_f32(_pB0), 1);
-                _sum4 = vmlaq_lane_f32(_sum4, _pA, vget_low_f32(_pB1), 0);
-                _sum5 = vmlaq_lane_f32(_sum5, _pA, vget_low_f32(_pB1), 1);
-                _sum6 = vmlaq_lane_f32(_sum6, _pA, vget_high_f32(_pB1), 0);
-                _sum7 = vmlaq_lane_f32(_sum7, _pA, vget_high_f32(_pB1), 1);
-                _sum8 = vmlaq_lane_f32(_sum8, _pA, vget_low_f32(_pB2), 0);
-                _sum9 = vmlaq_lane_f32(_sum9, _pA, vget_low_f32(_pB2), 1);
-                _suma = vmlaq_lane_f32(_suma, _pA, vget_high_f32(_pB2), 0);
-                _sumb = vmlaq_lane_f32(_sumb, _pA, vget_high_f32(_pB2), 1);
-
-                pA += 4;
-                pB += 12;
-#endif
-#endif // __aarch64__
-#endif
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-                _sum1 = vmulq_f32(_sum1, _alpha);
-                _sum2 = vmulq_f32(_sum2, _alpha);
-                _sum3 = vmulq_f32(_sum3, _alpha);
-                _sum4 = vmulq_f32(_sum4, _alpha);
-                _sum5 = vmulq_f32(_sum5, _alpha);
-                _sum6 = vmulq_f32(_sum6, _alpha);
-                _sum7 = vmulq_f32(_sum7, _alpha);
-                _sum8 = vmulq_f32(_sum8, _alpha);
-                _sum9 = vmulq_f32(_sum9, _alpha);
-                _suma = vmulq_f32(_suma, _alpha);
-                _sumb = vmulq_f32(_sumb, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum0));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum1));
-                    vst1_u16(outptr0 + 4 * 2, (uint16x4_t)vcvt_f16_f32(_sum2));
-                    vst1_u16(outptr0 + 4 * 3, (uint16x4_t)vcvt_f16_f32(_sum3));
-                    vst1_u16(outptr0 + 4 * 4, (uint16x4_t)vcvt_f16_f32(_sum4));
-                    vst1_u16(outptr0 + 4 * 5, (uint16x4_t)vcvt_f16_f32(_sum5));
-                    vst1_u16(outptr0 + 4 * 6, (uint16x4_t)vcvt_f16_f32(_sum6));
-                    vst1_u16(outptr0 + 4 * 7, (uint16x4_t)vcvt_f16_f32(_sum7));
-                    vst1_u16(outptr0 + 4 * 8, (uint16x4_t)vcvt_f16_f32(_sum8));
-                    vst1_u16(outptr0 + 4 * 9, (uint16x4_t)vcvt_f16_f32(_sum9));
-                    vst1_u16(outptr0 + 4 * 10, (uint16x4_t)vcvt_f16_f32(_suma));
-                    vst1_u16(outptr0 + 4 * 11, (uint16x4_t)vcvt_f16_f32(_sumb));
-                    outptr0 += 48;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose4x12_ps(_sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7, _sum8, _sum9, _suma, _sumb);
-
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum0));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum1));
-                    vst1_u16(outptr0 + 8, (uint16x4_t)vcvt_f16_f32(_sum2));
-                    vst1_u16(outptr0 + out_hstep, (uint16x4_t)vcvt_f16_f32(_sum3));
-                    vst1_u16(outptr0 + out_hstep + 4, (uint16x4_t)vcvt_f16_f32(_sum4));
-                    vst1_u16(outptr0 + out_hstep + 8, (uint16x4_t)vcvt_f16_f32(_sum5));
-                    vst1_u16(outptr0 + out_hstep * 2, (uint16x4_t)vcvt_f16_f32(_sum6));
-                    vst1_u16(outptr0 + out_hstep * 2 + 4, (uint16x4_t)vcvt_f16_f32(_sum7));
-                    vst1_u16(outptr0 + out_hstep * 2 + 8, (uint16x4_t)vcvt_f16_f32(_sum8));
-                    vst1_u16(outptr0 + out_hstep * 3, (uint16x4_t)vcvt_f16_f32(_sum9));
-                    vst1_u16(outptr0 + out_hstep * 3 + 4, (uint16x4_t)vcvt_f16_f32(_suma));
-                    vst1_u16(outptr0 + out_hstep * 3 + 8, (uint16x4_t)vcvt_f16_f32(_sumb));
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 4 * 2, _sum2);
-                vst1q_f32(outptr + 4 * 3, _sum3);
-                vst1q_f32(outptr + 4 * 4, _sum4);
-                vst1q_f32(outptr + 4 * 5, _sum5);
-                vst1q_f32(outptr + 4 * 6, _sum6);
-                vst1q_f32(outptr + 4 * 7, _sum7);
-                vst1q_f32(outptr + 4 * 8, _sum8);
-                vst1q_f32(outptr + 4 * 9, _sum9);
-                vst1q_f32(outptr + 4 * 10, _suma);
-                vst1q_f32(outptr + 4 * 11, _sumb);
-            }
-
-            outptr += 48;
-        }
-#endif // __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-            float32x4_t _sum3;
-            float32x4_t _sum4;
-            float32x4_t _sum5;
-            float32x4_t _sum6;
-            float32x4_t _sum7;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-                _sum2 = vdupq_n_f32(0.f);
-                _sum3 = vdupq_n_f32(0.f);
-                _sum4 = vdupq_n_f32(0.f);
-                _sum5 = vdupq_n_f32(0.f);
-                _sum6 = vdupq_n_f32(0.f);
-                _sum7 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = _sum0;
-                        _sum2 = _sum0;
-                        _sum3 = _sum0;
-                        _sum4 = _sum0;
-                        _sum5 = _sum0;
-                        _sum6 = _sum0;
-                        _sum7 = _sum0;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = _sum0;
-                        _sum2 = _sum0;
-                        _sum3 = _sum0;
-                        _sum4 = _sum0;
-                        _sum5 = _sum0;
-                        _sum6 = _sum0;
-                        _sum7 = _sum0;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = vld1q_f32(pC + 4);
-                        _sum2 = vld1q_f32(pC + 8);
-                        _sum3 = vld1q_f32(pC + 12);
-                        _sum4 = vld1q_f32(pC + 16);
-                        _sum5 = vld1q_f32(pC + 20);
-                        _sum6 = vld1q_f32(pC + 24);
-                        _sum7 = vld1q_f32(pC + 28);
-                        pC += 32;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = vdupq_n_f32(pC[1]);
-                        _sum2 = vdupq_n_f32(pC[2]);
-                        _sum3 = vdupq_n_f32(pC[3]);
-                        _sum4 = vdupq_n_f32(pC[4]);
-                        _sum5 = vdupq_n_f32(pC[5]);
-                        _sum6 = vdupq_n_f32(pC[6]);
-                        _sum7 = vdupq_n_f32(pC[7]);
-                        pC += 8;
-                    }
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4 * 1);
-                _sum2 = vld1q_f32(outptr + 4 * 2);
-                _sum3 = vld1q_f32(outptr + 4 * 3);
-                _sum4 = vld1q_f32(outptr + 4 * 4);
-                _sum5 = vld1q_f32(outptr + 4 * 5);
-                _sum6 = vld1q_f32(outptr + 4 * 6);
-                _sum7 = vld1q_f32(outptr + 4 * 7);
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x4_t _pA = vld1_f16(pA);
-                float16x8_t _pAA = vcombine_f16(_pA, _pA);
-
-                float16x4_t _pB0 = vld1_f16(pB);
-                float16x4_t _pB1 = vld1_f16(pB + 4);
-
-                _sum0 = vfmlalq_lane_low_f16(_sum0, _pAA, _pB0, 0);
-                _sum1 = vfmlalq_lane_low_f16(_sum1, _pAA, _pB0, 1);
-                _sum2 = vfmlalq_lane_low_f16(_sum2, _pAA, _pB0, 2);
-                _sum3 = vfmlalq_lane_low_f16(_sum3, _pAA, _pB0, 3);
-                _sum4 = vfmlalq_lane_low_f16(_sum4, _pAA, _pB1, 0);
-                _sum5 = vfmlalq_lane_low_f16(_sum5, _pAA, _pB1, 1);
-                _sum6 = vfmlalq_lane_low_f16(_sum6, _pAA, _pB1, 2);
-                _sum7 = vfmlalq_lane_low_f16(_sum7, _pAA, _pB1, 3);
-#else
-                float32x4_t _pA = vcvt_f32_f16((float16x4_t)vld1_u16(pA));
-                float32x4_t _pB0 = vcvt_f32_f16((float16x4_t)vld1_u16(pB));
-                float32x4_t _pB1 = vcvt_f32_f16((float16x4_t)vld1_u16(pB + 4));
-
-#if __aarch64__
-                _sum0 = vfmaq_laneq_f32(_sum0, _pA, _pB0, 0);
-                _sum1 = vfmaq_laneq_f32(_sum1, _pA, _pB0, 1);
-                _sum2 = vfmaq_laneq_f32(_sum2, _pA, _pB0, 2);
-                _sum3 = vfmaq_laneq_f32(_sum3, _pA, _pB0, 3);
-                _sum4 = vfmaq_laneq_f32(_sum4, _pA, _pB1, 0);
-                _sum5 = vfmaq_laneq_f32(_sum5, _pA, _pB1, 1);
-                _sum6 = vfmaq_laneq_f32(_sum6, _pA, _pB1, 2);
-                _sum7 = vfmaq_laneq_f32(_sum7, _pA, _pB1, 3);
-#else
-                _sum0 = vmlaq_lane_f32(_sum0, _pA, vget_low_f32(_pB0), 0);
-                _sum1 = vmlaq_lane_f32(_sum1, _pA, vget_low_f32(_pB0), 1);
-                _sum2 = vmlaq_lane_f32(_sum2, _pA, vget_high_f32(_pB0), 0);
-                _sum3 = vmlaq_lane_f32(_sum3, _pA, vget_high_f32(_pB0), 1);
-                _sum4 = vmlaq_lane_f32(_sum4, _pA, vget_low_f32(_pB1), 0);
-                _sum5 = vmlaq_lane_f32(_sum5, _pA, vget_low_f32(_pB1), 1);
-                _sum6 = vmlaq_lane_f32(_sum6, _pA, vget_high_f32(_pB1), 0);
-                _sum7 = vmlaq_lane_f32(_sum7, _pA, vget_high_f32(_pB1), 1);
-#endif
-#endif
-
-                pA += 4;
-                pB += 8;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-                _sum1 = vmulq_f32(_sum1, _alpha);
-                _sum2 = vmulq_f32(_sum2, _alpha);
-                _sum3 = vmulq_f32(_sum3, _alpha);
-                _sum4 = vmulq_f32(_sum4, _alpha);
-                _sum5 = vmulq_f32(_sum5, _alpha);
-                _sum6 = vmulq_f32(_sum6, _alpha);
-                _sum7 = vmulq_f32(_sum7, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum0));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum1));
-                    vst1_u16(outptr0 + 4 * 2, (uint16x4_t)vcvt_f16_f32(_sum2));
-                    vst1_u16(outptr0 + 4 * 3, (uint16x4_t)vcvt_f16_f32(_sum3));
-                    vst1_u16(outptr0 + 4 * 4, (uint16x4_t)vcvt_f16_f32(_sum4));
-                    vst1_u16(outptr0 + 4 * 5, (uint16x4_t)vcvt_f16_f32(_sum5));
-                    vst1_u16(outptr0 + 4 * 6, (uint16x4_t)vcvt_f16_f32(_sum6));
-                    vst1_u16(outptr0 + 4 * 7, (uint16x4_t)vcvt_f16_f32(_sum7));
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose4x8_ps(_sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7);
-
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum0));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum1));
-                    vst1_u16(outptr0 + out_hstep, (uint16x4_t)vcvt_f16_f32(_sum2));
-                    vst1_u16(outptr0 + out_hstep + 4, (uint16x4_t)vcvt_f16_f32(_sum3));
-                    vst1_u16(outptr0 + out_hstep * 2, (uint16x4_t)vcvt_f16_f32(_sum4));
-                    vst1_u16(outptr0 + out_hstep * 2 + 4, (uint16x4_t)vcvt_f16_f32(_sum5));
-                    vst1_u16(outptr0 + out_hstep * 3, (uint16x4_t)vcvt_f16_f32(_sum6));
-                    vst1_u16(outptr0 + out_hstep * 3 + 4, (uint16x4_t)vcvt_f16_f32(_sum7));
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 4 * 2, _sum2);
-                vst1q_f32(outptr + 4 * 3, _sum3);
-                vst1q_f32(outptr + 4 * 4, _sum4);
-                vst1q_f32(outptr + 4 * 5, _sum5);
-                vst1q_f32(outptr + 4 * 6, _sum6);
-                vst1q_f32(outptr + 4 * 7, _sum7);
-            }
-
-            outptr += 32;
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-            float32x4_t _sum3;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-                _sum2 = vdupq_n_f32(0.f);
-                _sum3 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = _sum0;
-                        _sum2 = _sum0;
-                        _sum3 = _sum0;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = _sum0;
-                        _sum2 = _sum0;
-                        _sum3 = _sum0;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = vld1q_f32(pC + 4);
-                        _sum2 = vld1q_f32(pC + 8);
-                        _sum3 = vld1q_f32(pC + 12);
-                        pC += 16;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = vdupq_n_f32(pC[1]);
-                        _sum2 = vdupq_n_f32(pC[2]);
-                        _sum3 = vdupq_n_f32(pC[3]);
-                        pC += 4;
-                    }
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4 * 1);
-                _sum2 = vld1q_f32(outptr + 4 * 2);
-                _sum3 = vld1q_f32(outptr + 4 * 3);
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x4_t _pA = vld1_f16(pA);
-                float16x8_t _pAA = vcombine_f16(_pA, _pA);
-
-                float16x4_t _pB0 = vld1_f16(pB);
-
-                _sum0 = vfmlalq_lane_low_f16(_sum0, _pAA, _pB0, 0);
-                _sum1 = vfmlalq_lane_low_f16(_sum1, _pAA, _pB0, 1);
-                _sum2 = vfmlalq_lane_low_f16(_sum2, _pAA, _pB0, 2);
-                _sum3 = vfmlalq_lane_low_f16(_sum3, _pAA, _pB0, 3);
-#else
-                float32x4_t _pA = vcvt_f32_f16((float16x4_t)vld1_u16(pA));
-                float32x4_t _pB = vcvt_f32_f16((float16x4_t)vld1_u16(pB));
-
-#if __aarch64__
-                _sum0 = vfmaq_laneq_f32(_sum0, _pA, _pB, 0);
-                _sum1 = vfmaq_laneq_f32(_sum1, _pA, _pB, 1);
-                _sum2 = vfmaq_laneq_f32(_sum2, _pA, _pB, 2);
-                _sum3 = vfmaq_laneq_f32(_sum3, _pA, _pB, 3);
-#else
-                _sum0 = vmlaq_lane_f32(_sum0, _pA, vget_low_f32(_pB), 0);
-                _sum1 = vmlaq_lane_f32(_sum1, _pA, vget_low_f32(_pB), 1);
-                _sum2 = vmlaq_lane_f32(_sum2, _pA, vget_high_f32(_pB), 0);
-                _sum3 = vmlaq_lane_f32(_sum3, _pA, vget_high_f32(_pB), 1);
-#endif
-#endif
-
-                pA += 4;
-                pB += 4;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-                _sum1 = vmulq_f32(_sum1, _alpha);
-                _sum2 = vmulq_f32(_sum2, _alpha);
-                _sum3 = vmulq_f32(_sum3, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum0));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum1));
-                    vst1_u16(outptr0 + 4 * 2, (uint16x4_t)vcvt_f16_f32(_sum2));
-                    vst1_u16(outptr0 + 4 * 3, (uint16x4_t)vcvt_f16_f32(_sum3));
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    transpose4x4_ps(_sum0, _sum1, _sum2, _sum3);
-
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum0));
-                    vst1_u16(outptr0 + out_hstep, (uint16x4_t)vcvt_f16_f32(_sum1));
-                    vst1_u16(outptr0 + out_hstep * 2, (uint16x4_t)vcvt_f16_f32(_sum2));
-                    vst1_u16(outptr0 + out_hstep * 3, (uint16x4_t)vcvt_f16_f32(_sum3));
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 4 * 2, _sum2);
-                vst1q_f32(outptr + 4 * 3, _sum3);
-            }
-
-            outptr += 16;
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = _sum0;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = _sum0;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = vld1q_f32(pC + 4);
-                        pC += 8;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = vdupq_n_f32(pC[1]);
-                        pC += 2;
-                    }
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4);
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x4_t _pA = vld1_f16(pA);
-                float16x8_t _pAA = vcombine_f16(_pA, _pA);
-
-                float16x4_t _pB0 = vdup_n_f16(pB[0]);
-                float16x4_t _pB1 = vdup_n_f16(pB[1]);
-                float16x8_t _pB01 = vcombine_f16(_pB0, _pB1);
-
-                _sum0 = vfmlalq_low_f16(_sum0, _pAA, _pB01);
-                _sum1 = vfmlalq_high_f16(_sum1, _pAA, _pB01);
-#else
-                float32x4_t _pA = vcvt_f32_f16((float16x4_t)vld1_u16(pA));
-                float32x4_t _pB0 = vcvt_f32_f16((float16x4_t)vdup_n_u16(pB[0]));
-                float32x4_t _pB1 = vcvt_f32_f16((float16x4_t)vdup_n_u16(pB[1]));
-
-#if __aarch64__
-                _sum0 = vfmaq_f32(_sum0, _pA, _pB0);
-                _sum1 = vfmaq_f32(_sum1, _pA, _pB1);
-#else
-                _sum0 = vmlaq_f32(_sum0, _pA, _pB0);
-                _sum1 = vmlaq_f32(_sum1, _pA, _pB1);
-#endif
-#endif
-
-                pA += 4;
-                pB += 2;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-                _sum1 = vmulq_f32(_sum1, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum0));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum1));
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short sum0[4];
-                    unsigned short sum1[4];
-                    vst1_u16(sum0, (uint16x4_t)vcvt_f16_f32(_sum0));
-                    vst1_u16(sum1, (uint16x4_t)vcvt_f16_f32(_sum1));
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[1] = sum1[0];
-                    outptr0[out_hstep + 1] = sum1[1];
-                    outptr0[out_hstep * 2 + 1] = sum1[2];
-                    outptr0[out_hstep * 3 + 1] = sum1[3];
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-            }
-
-            outptr += 8;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            float32x4_t _sum0;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        pC += 4;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        pC += 1;
-                    }
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x4_t _pA = vld1_f16(pA);
-                float16x8_t _pAA = vcombine_f16(_pA, _pA);
-
-                float16x8_t _pB = vdupq_n_f16(pB[0]);
-
-                _sum0 = vfmlalq_low_f16(_sum0, _pAA, _pB);
-#else
-                float32x4_t _pA = vcvt_f32_f16((float16x4_t)vld1_u16(pA));
-                float32x4_t _pB = vcvt_f32_f16((float16x4_t)vdup_n_u16(pB[0]));
-
-#if __aarch64__
-                _sum0 = vfmaq_f32(_sum0, _pA, _pB);
-#else
-                _sum0 = vmlaq_f32(_sum0, _pA, _pB);
-#endif
-#endif
-
-                pA += 4;
-                pB += 1;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum0));
-                    outptr0 += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    unsigned short sum0[4];
-                    vst1_u16(sum0, (uint16x4_t)vcvt_f16_f32(_sum0));
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0++;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-            }
-
-            outptr += 4;
-        }
-
-        pAT += max_kk * 4;
-    }
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        unsigned short* outptr0 = (unsigned short*)top_blob + (i + ii) * out_hstep + j;
-
-#if __ARM_FEATURE_FP16_FML
-        const __fp16* pB = pBT;
-#else
-        const unsigned short* pB = pBT;
-#endif
-
-        if (pC)
-        {
-            if (broadcast_type_C == 1 || broadcast_type_C == 2)
-            {
-                pC = (const float*)CT_tile + i + ii;
-            }
-            if (broadcast_type_C == 4)
-            {
-                pC = (const float*)CT_tile + j;
-            }
-        }
-
-        int jj = 0;
-#if __aarch64__
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum02;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-            float32x4_t _sum12;
-
-            if (k == 0)
-            {
-                _sum00 = vdupq_n_f32(0.f);
-                _sum01 = vdupq_n_f32(0.f);
-                _sum02 = vdupq_n_f32(0.f);
-                _sum10 = vdupq_n_f32(0.f);
-                _sum11 = vdupq_n_f32(0.f);
-                _sum12 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum02 = _sum00;
-                        _sum10 = _sum00;
-                        _sum11 = _sum00;
-                        _sum12 = _sum00;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum02 = _sum00;
-                        _sum10 = vdupq_n_f32(pC[1]);
-                        _sum11 = _sum10;
-                        _sum12 = _sum10;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        float32x4x2_t _tmp01 = vld2q_f32(pC);
-                        float32x4x2_t _tmp23 = vld2q_f32(pC + 8);
-                        float32x4x2_t _tmp45 = vld2q_f32(pC + 16);
-                        _sum00 = _tmp01.val[0];
-                        _sum01 = _tmp23.val[0];
-                        _sum02 = _tmp45.val[0];
-                        _sum10 = _tmp01.val[1];
-                        _sum11 = _tmp23.val[1];
-                        _sum12 = _tmp45.val[1];
-                        pC += 24;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                        _sum02 = vld1q_f32(pC + 8);
-                        _sum10 = _sum00;
-                        _sum11 = _sum01;
-                        _sum12 = _sum02;
-                        pC += 12;
-                    }
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01 = vld2q_f32(outptr);
-                float32x4x2_t _tmp23 = vld2q_f32(outptr + 8);
-                float32x4x2_t _tmp45 = vld2q_f32(outptr + 16);
-                _sum00 = _tmp01.val[0];
-                _sum01 = _tmp23.val[0];
-                _sum02 = _tmp45.val[0];
-                _sum10 = _tmp01.val[1];
-                _sum11 = _tmp23.val[1];
-                _sum12 = _tmp45.val[1];
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x8_t _pB01 = vld1q_f16(pB);
-                float16x4_t _pB2 = vld1_f16(pB + 8);
-                float16x8_t _pB22 = vcombine_f16(_pB2, _pB2);
-
-                float16x4_t _pA0 = vdup_n_f16(pA[0]);
-                float16x4_t _pA1 = vdup_n_f16(pA[1]);
-                float16x8_t _pA01 = vcombine_f16(_pA0, _pA1);
-                float16x8_t _pA10 = vcombine_f16(_pA1, _pA0);
-
-                _sum00 = vfmlalq_low_f16(_sum00, _pB01, _pA01);
-                _sum01 = vfmlalq_high_f16(_sum01, _pB01, _pA10);
-                _sum02 = vfmlalq_low_f16(_sum02, _pB22, _pA01);
-                _sum10 = vfmlalq_low_f16(_sum10, _pB01, _pA10);
-                _sum11 = vfmlalq_high_f16(_sum11, _pB01, _pA01);
-                _sum12 = vfmlalq_low_f16(_sum12, _pB22, _pA10);
-#else
-                uint16x8_t _pB = vld1q_u16(pB);
-                float32x4_t _pB0 = vcvt_f32_f16((float16x4_t)vget_low_u16(_pB));
-                float32x4_t _pB1 = vcvt_f32_f16((float16x4_t)vget_high_u16(_pB));
-                float32x4_t _pB2 = vcvt_f32_f16((float16x4_t)vld1_u16(pB + 8));
-
-                float32x4_t _pA0 = vcvt_f32_f16((float16x4_t)vdup_n_u16(pA[0]));
-                float32x4_t _pA1 = vcvt_f32_f16((float16x4_t)vdup_n_u16(pA[1]));
-#if __aarch64__
-                _sum00 = vfmaq_f32(_sum00, _pB0, _pA0);
-                _sum01 = vfmaq_f32(_sum01, _pB1, _pA0);
-                _sum02 = vfmaq_f32(_sum02, _pB2, _pA0);
-                _sum10 = vfmaq_f32(_sum10, _pB0, _pA1);
-                _sum11 = vfmaq_f32(_sum11, _pB1, _pA1);
-                _sum12 = vfmaq_f32(_sum12, _pB2, _pA1);
-#else
-                _sum00 = vmlaq_f32(_sum00, _pB0, _pA0);
-                _sum01 = vmlaq_f32(_sum01, _pB1, _pA0);
-                _sum02 = vmlaq_f32(_sum02, _pB2, _pA0);
-                _sum10 = vmlaq_f32(_sum10, _pB0, _pA1);
-                _sum11 = vmlaq_f32(_sum11, _pB1, _pA1);
-                _sum12 = vmlaq_f32(_sum12, _pB2, _pA1);
-#endif
-#endif
-
-                pA += 2;
-                pB += 12;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum00 = vmulq_f32(_sum00, _alpha);
-                _sum01 = vmulq_f32(_sum01, _alpha);
-                _sum02 = vmulq_f32(_sum02, _alpha);
-                _sum10 = vmulq_f32(_sum10, _alpha);
-                _sum11 = vmulq_f32(_sum11, _alpha);
-                _sum12 = vmulq_f32(_sum12, _alpha);
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum00));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum01));
-                    vst1_u16(outptr0 + 8, (uint16x4_t)vcvt_f16_f32(_sum02));
-                    vst1_u16(outptr0 + out_hstep, (uint16x4_t)vcvt_f16_f32(_sum10));
-                    vst1_u16(outptr0 + out_hstep + 4, (uint16x4_t)vcvt_f16_f32(_sum11));
-                    vst1_u16(outptr0 + out_hstep + 8, (uint16x4_t)vcvt_f16_f32(_sum12));
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01;
-                _tmp01.val[0] = _sum00;
-                _tmp01.val[1] = _sum10;
-                float32x4x2_t _tmp23;
-                _tmp23.val[0] = _sum01;
-                _tmp23.val[1] = _sum11;
-                float32x4x2_t _tmp45;
-                _tmp45.val[0] = _sum02;
-                _tmp45.val[1] = _sum12;
-                vst2q_f32(outptr, _tmp01);
-                vst2q_f32(outptr + 8, _tmp23);
-                vst2q_f32(outptr + 16, _tmp45);
-            }
-
-            outptr += 24;
-        }
-#endif // __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            float32x4_t _sum00;
-            float32x4_t _sum01;
-            float32x4_t _sum10;
-            float32x4_t _sum11;
-
-            if (k == 0)
-            {
-                _sum00 = vdupq_n_f32(0.f);
-                _sum01 = vdupq_n_f32(0.f);
-                _sum10 = vdupq_n_f32(0.f);
-                _sum11 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum10 = _sum00;
-                        _sum11 = _sum00;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum00 = vdupq_n_f32(pC[0]);
-                        _sum01 = _sum00;
-                        _sum10 = vdupq_n_f32(pC[1]);
-                        _sum11 = _sum10;
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        float32x4x2_t _tmp01 = vld2q_f32(pC);
-                        float32x4x2_t _tmp23 = vld2q_f32(pC + 8);
-                        _sum00 = _tmp01.val[0];
-                        _sum01 = _tmp23.val[0];
-                        _sum10 = _tmp01.val[1];
-                        _sum11 = _tmp23.val[1];
-                        pC += 16;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum00 = vld1q_f32(pC);
-                        _sum01 = vld1q_f32(pC + 4);
-                        _sum10 = _sum00;
-                        _sum11 = _sum01;
-                        pC += 8;
-                    }
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01 = vld2q_f32(outptr);
-                float32x4x2_t _tmp23 = vld2q_f32(outptr + 8);
-                _sum00 = _tmp01.val[0];
-                _sum01 = _tmp23.val[0];
-                _sum10 = _tmp01.val[1];
-                _sum11 = _tmp23.val[1];
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x8_t _pB01 = vld1q_f16(pB);
-
-                float16x4_t _pA0 = vdup_n_f16(pA[0]);
-                float16x4_t _pA1 = vdup_n_f16(pA[1]);
-                float16x8_t _pA01 = vcombine_f16(_pA0, _pA1);
-                float16x8_t _pA10 = vcombine_f16(_pA1, _pA0);
-
-                _sum00 = vfmlalq_low_f16(_sum00, _pB01, _pA01);
-                _sum01 = vfmlalq_high_f16(_sum01, _pB01, _pA10);
-                _sum10 = vfmlalq_low_f16(_sum10, _pB01, _pA10);
-                _sum11 = vfmlalq_high_f16(_sum11, _pB01, _pA01);
-#else
-                uint16x8_t _pB = vld1q_u16(pB);
-                float32x4_t _pB0 = vcvt_f32_f16((float16x4_t)vget_low_u16(_pB));
-                float32x4_t _pB1 = vcvt_f32_f16((float16x4_t)vget_high_u16(_pB));
-
-                float32x4_t _pA0 = vcvt_f32_f16((float16x4_t)vdup_n_u16(pA[0]));
-                float32x4_t _pA1 = vcvt_f32_f16((float16x4_t)vdup_n_u16(pA[1]));
-#if __aarch64__
-                _sum00 = vfmaq_f32(_sum00, _pB0, _pA0);
-                _sum01 = vfmaq_f32(_sum01, _pB1, _pA0);
-                _sum10 = vfmaq_f32(_sum10, _pB0, _pA1);
-                _sum11 = vfmaq_f32(_sum11, _pB1, _pA1);
-#else
-                _sum00 = vmlaq_f32(_sum00, _pB0, _pA0);
-                _sum01 = vmlaq_f32(_sum01, _pB1, _pA0);
-                _sum10 = vmlaq_f32(_sum10, _pB0, _pA1);
-                _sum11 = vmlaq_f32(_sum11, _pB1, _pA1);
-#endif
-#endif
-
-                pA += 2;
-                pB += 8;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum00 = vmulq_f32(_sum00, _alpha);
-                _sum01 = vmulq_f32(_sum01, _alpha);
-                _sum10 = vmulq_f32(_sum10, _alpha);
-                _sum11 = vmulq_f32(_sum11, _alpha);
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum00));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum01));
-                    vst1_u16(outptr0 + out_hstep, (uint16x4_t)vcvt_f16_f32(_sum10));
-                    vst1_u16(outptr0 + out_hstep + 4, (uint16x4_t)vcvt_f16_f32(_sum11));
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01;
-                _tmp01.val[0] = _sum00;
-                _tmp01.val[1] = _sum10;
-                float32x4x2_t _tmp23;
-                _tmp23.val[0] = _sum01;
-                _tmp23.val[1] = _sum11;
-                vst2q_f32(outptr, _tmp01);
-                vst2q_f32(outptr + 8, _tmp23);
-            }
-
-            outptr += 16;
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = _sum0;
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = vdupq_n_f32(pC[1]);
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        float32x4x2_t _tmp01 = vld2q_f32(pC);
-                        _sum0 = _tmp01.val[0];
-                        _sum1 = _tmp01.val[1];
-                        pC += 8;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = _sum0;
-                        pC += 4;
-                    }
-                }
-            }
-            else
-            {
-                float32x4_t _tmp0 = vld1q_f32(outptr);
-                float32x4_t _tmp1 = vld1q_f32(outptr + 4);
-                float32x4x2_t _tmp01 = vuzpq_f32(_tmp0, _tmp1);
-                _sum0 = _tmp01.val[0];
-                _sum1 = _tmp01.val[1];
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x4_t _pB = vld1_f16(pB);
-                float16x8_t _pBB = vcombine_f16(_pB, _pB);
-
-                float16x4_t _pA0 = vdup_n_f16(pA[0]);
-                float16x4_t _pA1 = vdup_n_f16(pA[1]);
-                float16x8_t _pA01 = vcombine_f16(_pA0, _pA1);
-
-                _sum0 = vfmlalq_low_f16(_sum0, _pBB, _pA01);
-                _sum1 = vfmlalq_high_f16(_sum1, _pBB, _pA01);
-#else
-                float32x4_t _pB = vcvt_f32_f16((float16x4_t)vld1_u16(pB));
-
-                float32x4_t _pA0 = vcvt_f32_f16((float16x4_t)vdup_n_u16(pA[0]));
-                float32x4_t _pA1 = vcvt_f32_f16((float16x4_t)vdup_n_u16(pA[1]));
-#if __aarch64__
-                _sum0 = vfmaq_f32(_sum0, _pB, _pA0);
-                _sum1 = vfmaq_f32(_sum1, _pB, _pA1);
-#else
-                _sum0 = vmlaq_f32(_sum0, _pB, _pA0);
-                _sum1 = vmlaq_f32(_sum1, _pB, _pA1);
-#endif
-#endif
-
-                pA += 2;
-                pB += 4;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-                _sum1 = vmulq_f32(_sum1, _alpha);
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum0));
-                    vst1_u16(outptr0 + out_hstep, (uint16x4_t)vcvt_f16_f32(_sum1));
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                float32x4x2_t _tmp01;
-                _tmp01.val[0] = _sum0;
-                _tmp01.val[1] = _sum1;
-                vst2q_f32(outptr, _tmp01);
-            }
-
-            outptr += 8;
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            float sum00;
-            float sum01;
-            float sum10;
-            float sum11;
-
-            if (k == 0)
-            {
-                sum00 = 0.f;
-                sum01 = 0.f;
-                sum10 = 0.f;
-                sum11 = 0.f;
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        sum00 = pC[0];
-                        sum01 = pC[0];
-                        sum10 = pC[0];
-                        sum11 = pC[0];
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        sum00 = pC[0];
-                        sum01 = pC[1];
-                        sum10 = pC[0];
-                        sum11 = pC[1];
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        sum00 = pC[0];
-                        sum01 = pC[1];
-                        sum10 = pC[2];
-                        sum11 = pC[3];
-                        pC += 4;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        sum00 = pC[0];
-                        sum01 = pC[0];
-                        sum10 = pC[1];
-                        sum11 = pC[1];
-                        pC += 2;
-                    }
-                }
-            }
-            else
-            {
-                sum00 = outptr[0];
-                sum01 = outptr[1];
-                sum10 = outptr[2];
-                sum11 = outptr[3];
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                __fp16 pA0 = pA[0];
-                __fp16 pA1 = pA[1];
-                __fp16 pB0 = pB[0];
-                __fp16 pB1 = pB[1];
-#else
-                float pA0 = float16_to_float32(pA[0]);
-                float pA1 = float16_to_float32(pA[1]);
-                float pB0 = float16_to_float32(pB[0]);
-                float pB1 = float16_to_float32(pB[1]);
-#endif
-
-                sum00 += pA0 * pB0;
-                sum01 += pA1 * pB0;
-                sum10 += pA0 * pB1;
-                sum11 += pA1 * pB1;
-
-                pA += 2;
-                pB += 2;
-            }
-
-            if (alpha != 1.f)
-            {
-                sum00 *= alpha;
-                sum01 *= alpha;
-                sum10 *= alpha;
-                sum11 *= alpha;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = float32_to_float16(sum00);
-                    outptr0[1] = float32_to_float16(sum10);
-                    outptr0[out_hstep] = float32_to_float16(sum01);
-                    outptr0[out_hstep + 1] = float32_to_float16(sum11);
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                outptr[0] = sum00;
-                outptr[1] = sum01;
-                outptr[2] = sum10;
-                outptr[3] = sum11;
-            }
-
-            outptr += 4;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            float sum0;
-            float sum1;
-
-            if (k == 0)
-            {
-                sum0 = 0.f;
-                sum1 = 0.f;
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0)
-                    {
-                        sum0 = pC[0];
-                        sum1 = pC[0];
-                    }
-                    if (broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        sum0 = pC[0];
-                        sum1 = pC[1];
-                    }
-                    if (broadcast_type_C == 3)
-                    {
-                        sum0 = pC[0];
-                        sum1 = pC[1];
-                        pC += 2;
-                    }
-                    if (broadcast_type_C == 4)
-                    {
-                        sum0 = pC[0];
-                        sum1 = pC[0];
-                        pC += 1;
-                    }
-                }
-            }
-            else
-            {
-                sum0 = outptr[0];
-                sum1 = outptr[1];
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                __fp16 pA0 = pA[0];
-                __fp16 pA1 = pA[1];
-                __fp16 pB0 = pB[0];
-#else
-                float pA0 = float16_to_float32(pA[0]);
-                float pA1 = float16_to_float32(pA[1]);
-                float pB0 = float16_to_float32(pB[0]);
-#endif
-
-                sum0 += pA0 * pB0;
-                sum1 += pA1 * pB0;
-                pA += 2;
-                pB += 1;
-            }
-
-            if (alpha != 1.f)
-            {
-                sum0 *= alpha;
-                sum1 *= alpha;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = float32_to_float16(sum0);
-                    outptr0[out_hstep] = float32_to_float16(sum1);
-                    outptr0++;
-                }
-            }
-            else
-            {
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-            }
-
-            outptr += 2;
-        }
-
-        pAT += max_kk * 2;
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        unsigned short* outptr0 = (unsigned short*)top_blob + (i + ii) * out_hstep + j;
-
-#if __ARM_FEATURE_FP16_FML
-        const __fp16* pB = pBT;
-#else
-        const unsigned short* pB = pBT;
-#endif
-
-        if (pC)
-        {
-            if (broadcast_type_C == 1 || broadcast_type_C == 2)
-            {
-                pC = (const float*)CT_tile + i + ii;
-            }
-            if (broadcast_type_C == 4)
-            {
-                pC = (const float*)CT_tile + j;
-            }
-        }
-
-        int jj = 0;
-#if __aarch64__
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-            float32x4_t _sum2;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-                _sum2 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0 || broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = vdupq_n_f32(pC[0]);
-                        _sum2 = vdupq_n_f32(pC[0]);
-                    }
-                    if (broadcast_type_C == 3 || broadcast_type_C == 4)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = vld1q_f32(pC + 4);
-                        _sum2 = vld1q_f32(pC + 8);
-                        pC += 12;
-                    }
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4);
-                _sum2 = vld1q_f32(outptr + 8);
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x8_t _pB01 = vld1q_f16(pB);
-                float16x4_t _pB2 = vld1_f16(pB + 8);
-                float16x8_t _pB22 = vcombine_f16(_pB2, _pB2);
-
-                float16x8_t _pA = vdupq_n_f16(pA[0]);
-
-                _sum0 = vfmlalq_low_f16(_sum0, _pA, _pB01);
-                _sum1 = vfmlalq_high_f16(_sum1, _pA, _pB01);
-                _sum2 = vfmlalq_low_f16(_sum2, _pA, _pB22);
-#else
-                uint16x8_t _pB = vld1q_u16(pB);
-                float32x4_t _pB0 = vcvt_f32_f16((float16x4_t)vget_low_u16(_pB));
-                float32x4_t _pB1 = vcvt_f32_f16((float16x4_t)vget_high_u16(_pB));
-                float32x4_t _pB2 = vcvt_f32_f16((float16x4_t)vld1_u16(pB + 8));
-
-                float32x4_t _pA0 = vcvt_f32_f16((float16x4_t)vdup_n_u16(pA[0]));
-#if __aarch64__
-                _sum0 = vfmaq_f32(_sum0, _pA0, _pB0);
-                _sum1 = vfmaq_f32(_sum1, _pA0, _pB1);
-                _sum2 = vfmaq_f32(_sum2, _pA0, _pB2);
-#else
-                _sum0 = vmlaq_f32(_sum0, _pA0, _pB0);
-                _sum1 = vmlaq_f32(_sum1, _pA0, _pB1);
-                _sum2 = vmlaq_f32(_sum2, _pA0, _pB2);
-#endif
-#endif
-
-                pA += 1;
-                pB += 12;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-                _sum1 = vmulq_f32(_sum1, _alpha);
-                _sum2 = vmulq_f32(_sum2, _alpha);
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum0));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum1));
-                    vst1_u16(outptr0 + 8, (uint16x4_t)vcvt_f16_f32(_sum2));
-                    outptr0 += 12;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 8, _sum2);
-            }
-
-            outptr += 12;
-        }
-#endif // __aarch64__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            float32x4_t _sum0;
-            float32x4_t _sum1;
-
-            if (k == 0)
-            {
-                _sum0 = vdupq_n_f32(0.f);
-                _sum1 = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0 || broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum0 = vdupq_n_f32(pC[0]);
-                        _sum1 = vdupq_n_f32(pC[0]);
-                    }
-                    if (broadcast_type_C == 3 || broadcast_type_C == 4)
-                    {
-                        _sum0 = vld1q_f32(pC);
-                        _sum1 = vld1q_f32(pC + 4);
-                        pC += 8;
-                    }
-                }
-            }
-            else
-            {
-                _sum0 = vld1q_f32(outptr);
-                _sum1 = vld1q_f32(outptr + 4);
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x8_t _pB01 = vld1q_f16(pB);
-                float16x8_t _pA = vdupq_n_f16(pA[0]);
-
-                _sum0 = vfmlalq_low_f16(_sum0, _pA, _pB01);
-                _sum1 = vfmlalq_high_f16(_sum1, _pA, _pB01);
-#else
-                uint16x8_t _pB = vld1q_u16(pB);
-                float32x4_t _pB0 = vcvt_f32_f16((float16x4_t)vget_low_u16(_pB));
-                float32x4_t _pB1 = vcvt_f32_f16((float16x4_t)vget_high_u16(_pB));
-
-                float32x4_t _pA0 = vcvt_f32_f16((float16x4_t)vdup_n_u16(pA[0]));
-#if __aarch64__
-                _sum0 = vfmaq_f32(_sum0, _pA0, _pB0);
-                _sum1 = vfmaq_f32(_sum1, _pA0, _pB1);
-#else
-                _sum0 = vmlaq_f32(_sum0, _pA0, _pB0);
-                _sum1 = vmlaq_f32(_sum1, _pA0, _pB1);
-#endif
-#endif
-
-                pA += 1;
-                pB += 8;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum0 = vmulq_f32(_sum0, _alpha);
-                _sum1 = vmulq_f32(_sum1, _alpha);
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum0));
-                    vst1_u16(outptr0 + 4, (uint16x4_t)vcvt_f16_f32(_sum1));
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-            }
-
-            outptr += 8;
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            float32x4_t _sum;
-
-            if (k == 0)
-            {
-                _sum = vdupq_n_f32(0.f);
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0 || broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        _sum = vdupq_n_f32(pC[0]);
-                    }
-                    if (broadcast_type_C == 3 || broadcast_type_C == 4)
-                    {
-                        _sum = vld1q_f32(pC);
-                        pC += 4;
-                    }
-                }
-            }
-            else
-            {
-                _sum = vld1q_f32(outptr);
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                float16x4_t _pB = vld1_f16(pB);
-                float16x8_t _pBB = vcombine_f16(_pB, _pB);
-                float16x8_t _pA = vdupq_n_f16(pA[0]);
-
-                _sum = vfmlalq_low_f16(_sum, _pA, _pBB);
-#else
-                float32x4_t _pB = vcvt_f32_f16((float16x4_t)vld1_u16(pB));
-                float32x4_t _pA = vcvt_f32_f16((float16x4_t)vdup_n_u16(pA[0]));
-#if __aarch64__
-                _sum = vfmaq_f32(_sum, _pA, _pB);
-#else
-                _sum = vmlaq_f32(_sum, _pA, _pB);
-#endif
-#endif
-
-                pA += 1;
-                pB += 4;
-            }
-
-            if (alpha != 1.f)
-            {
-                float32x4_t _alpha = vdupq_n_f32(alpha);
-                _sum = vmulq_f32(_sum, _alpha);
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    vst1_u16(outptr0, (uint16x4_t)vcvt_f16_f32(_sum));
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                vst1q_f32(outptr, _sum);
-            }
-
-            outptr += 4;
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            float sum0;
-            float sum1;
-
-            if (k == 0)
-            {
-                sum0 = 0.f;
-                sum1 = 0.f;
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0 || broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        sum0 = pC[0];
-                        sum1 = pC[0];
-                    }
-                    if (broadcast_type_C == 3 || broadcast_type_C == 4)
-                    {
-                        sum0 = pC[0];
-                        sum1 = pC[1];
-                        pC += 2;
-                    }
-                }
-            }
-            else
-            {
-                sum0 = outptr[0];
-                sum1 = outptr[1];
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                __fp16 pA0 = pA[0];
-                __fp16 pB0 = pB[0];
-                __fp16 pB1 = pB[1];
-#else
-                float pA0 = float16_to_float32(pA[0]);
-                float pB0 = float16_to_float32(pB[0]);
-                float pB1 = float16_to_float32(pB[1]);
-#endif
-
-                sum0 += pA0 * pB0;
-                sum1 += pA0 * pB1;
-
-                pA += 1;
-                pB += 2;
-            }
-
-            if (alpha != 1.f)
-            {
-                sum0 *= alpha;
-                sum1 *= alpha;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = float32_to_float16(sum0);
-                    outptr0[1] = float32_to_float16(sum1);
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-            }
-
-            outptr += 2;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            float sum;
-
-            if (k == 0)
-            {
-                sum = 0.f;
-
-                if (pC)
-                {
-                    if (broadcast_type_C == 0 || broadcast_type_C == 1 || broadcast_type_C == 2)
-                    {
-                        sum = pC[0];
-                    }
-                    if (broadcast_type_C == 3 || broadcast_type_C == 4)
-                    {
-                        sum = pC[0];
-                        pC += 1;
-                    }
-                }
-            }
-            else
-            {
-                sum = outptr[0];
-            }
-
-#if __ARM_FEATURE_FP16_FML
-            const __fp16* pA = pAT;
-#else
-            const unsigned short* pA = pAT;
-#endif
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-#if __ARM_FEATURE_FP16_FML
-                __fp16 pA0 = pA[0];
-                __fp16 pB0 = pB[0];
-#else
-                float pA0 = float16_to_float32(pA[0]);
-                float pB0 = float16_to_float32(pB[0]);
-#endif
-
-                sum += pA0 * pB0;
-                pA += 1;
-                pB += 1;
-            }
-
-            if (alpha != 1.f)
-            {
-                sum *= alpha;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = float32_to_float16(sum);
-                    outptr0++;
-                }
-            }
-            else
-            {
-                outptr[0] = sum;
-            }
-
-            outptr += 1;
-        }
-
-        pAT += max_kk;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gru_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gru_arm.h
deleted file mode 100644
index 6eae1656..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/gru_arm.h
+++ /dev/null
@@ -1,53 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GRU_ARM_H
-#define LAYER_GRU_ARM_H
-
-#include "gru.h"
-
-namespace ncnn {
-
-class GRU_arm : public GRU
-{
-public:
-    GRU_arm();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int create_pipeline_bf16s(const Option& opt);
-    int forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_bf16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-#endif
-
-public:
-    Mat weight_xc_data_packed;
-    Mat bias_c_data_packed;
-    Mat weight_hc_data_packed;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GRU_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/hardsigmoid_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/hardsigmoid_arm.h
deleted file mode 100644
index 13783ff1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/hardsigmoid_arm.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_HARDSIGMOID_ARM_H
-#define LAYER_HARDSIGMOID_ARM_H
-
-#include "hardsigmoid.h"
-
-namespace ncnn {
-
-class HardSigmoid_arm : public HardSigmoid
-{
-public:
-    HardSigmoid_arm();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_inplace_bf16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_HARDSIGMOID_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/hardswish_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/hardswish_arm.h
deleted file mode 100644
index a534ceb1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/hardswish_arm.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_HARDSWISH_ARM_H
-#define LAYER_HARDSWISH_ARM_H
-
-#include "hardswish.h"
-
-namespace ncnn {
-
-class HardSwish_arm : public HardSwish
-{
-public:
-    HardSwish_arm();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_inplace_bf16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_HARDSWISH_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/innerproduct_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/innerproduct_arm.h
deleted file mode 100644
index 70a54533..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/innerproduct_arm.h
+++ /dev/null
@@ -1,64 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INNERPRODUCT_ARM_H
-#define LAYER_INNERPRODUCT_ARM_H
-
-#include "innerproduct.h"
-
-namespace ncnn {
-
-class InnerProduct_arm : public InnerProduct
-{
-public:
-    InnerProduct_arm();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_VFPV4
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_ARM82
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int create_pipeline_bf16s(const Option& opt);
-    int forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_INT8
-    int create_pipeline_int8_arm(const Option& opt);
-    int forward_int8_arm(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Layer* flatten;
-
-    Mat weight_data_tm;
-
-    // fp16
-    Mat bias_data_fp16;
-
-#if NCNN_INT8
-    Mat scale_in_data;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INNERPRODUCT_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/innerproduct_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/innerproduct_fp16s.h
deleted file mode 100644
index d4fa9950..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/innerproduct_fp16s.h
+++ /dev/null
@@ -1,697 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if !(__ARM_FEATURE_FP16_FML || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
-#if NCNN_RUNTIME_CPU && NCNN_ARM82FP16FML && __aarch64__ && !__ARM_FEATURE_FP16_FML
-void innerproduct_pack4_fp16s_neon_asimdfhm(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt);
-void innerproduct_fp16s_neon_asimdfhm(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt);
-void innerproduct_transform_kernel_fp16s_neon_asimdfhm(const Mat& weight_data, Mat& weight_data_tm, int num_input, int num_output, const Option& opt);
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-void innerproduct_pack4_fp16s_neon_asimdhp(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt);
-void innerproduct_fp16s_neon_asimdhp(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt);
-void innerproduct_transform_kernel_fp16s_neon_asimdhp(const Mat& weight_data, Mat& weight_data_tm, int num_input, int num_output, const Option& opt);
-#endif
-#endif
-
-static void innerproduct_pack4_fp16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt)
-{
-#if !(__ARM_FEATURE_FP16_FML || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
-#if NCNN_RUNTIME_CPU && NCNN_ARM82FP16FML && __aarch64__ && !__ARM_FEATURE_FP16_FML
-    if (ncnn::cpu_support_arm_asimdfhm())
-    {
-        innerproduct_pack4_fp16s_neon_asimdfhm(bottom_blob, top_blob, weight_data_fp16, bias_data, activation_type, activation_params, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-    if (ncnn::cpu_support_arm_asimdhp())
-    {
-        innerproduct_pack4_fp16s_neon_asimdhp(bottom_blob, top_blob, weight_data_fp16, bias_data, activation_type, activation_params, opt);
-        return;
-    }
-#endif
-#endif
-
-    const int num_input = bottom_blob.w * bottom_blob.elempack;
-    const int num_output = top_blob.w;
-
-    const float* bias_data_ptr = bias_data;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < num_output; p++)
-    {
-        float32x4_t _sum0 = vdupq_n_f32(0.f);
-
-        if (bias_data_ptr)
-        {
-            _sum0 = vld1q_f32(bias_data_ptr + p * 4);
-        }
-
-        float32x4_t _sum1 = vdupq_n_f32(0.f);
-        float32x4_t _sum2 = vdupq_n_f32(0.f);
-        float32x4_t _sum3 = vdupq_n_f32(0.f);
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        const __fp16* sptr = bottom_blob;
-        const __fp16* kptr = weight_data_fp16.row<const __fp16>(p);
-#else
-        const float* sptr = bottom_blob;
-        const unsigned short* kptr = weight_data_fp16.row<const unsigned short>(p);
-#endif
-
-        int i = 0;
-#if NCNN_GNU_INLINE_ASM
-        for (; i + 7 < num_input; i += 8)
-        {
-#if __aarch64__
-#if __ARM_FEATURE_FP16_FML
-            asm volatile(
-                "prfm   pldl1keep, [%0, #128]       \n"
-                "ld1    {v0.8h}, [%0], #16          \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v2.8h, v3.8h, v4.8h, v5.8h}, [%1], #64 \n"
-                "fmlal  %2.4s, v2.4h, v0.h[0]       \n"
-                "fmlal2 %3.4s, v2.4h, v0.h[1]       \n"
-                "fmlal  %4.4s, v3.4h, v0.h[2]       \n"
-                "fmlal2 %5.4s, v3.4h, v0.h[3]       \n"
-                "fmlal  %2.4s, v4.4h, v0.h[4]       \n"
-                "fmlal2 %3.4s, v4.4h, v0.h[5]       \n"
-                "fmlal  %4.4s, v5.4h, v0.h[6]       \n"
-                "fmlal2 %5.4s, v5.4h, v0.h[7]       \n"
-                : "=r"(sptr),  // %0
-                "=r"(kptr),  // %1
-                "=w"(_sum0), // %2
-                "=w"(_sum1), // %3
-                "=w"(_sum2), // %4
-                "=w"(_sum3)  // %5
-                : "0"(sptr),
-                "1"(kptr),
-                "2"(_sum0),
-                "3"(_sum1),
-                "4"(_sum2),
-                "5"(_sum3)
-                : "cc", "memory", "v0", "v2", "v3", "v4", "v5");
-#elif __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            asm volatile(
-                "prfm   pldl1keep, [%0, #128]       \n"
-                "ld1    {v1.8h}, [%0], #16          \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v6.8h, v7.8h, v8.8h, v9.8h}, [%1], #64 \n"
-                "fcvtl  v0.4s, v1.4h                \n"
-                "fcvtl2 v1.4s, v1.8h                \n"
-                "fcvtl  v2.4s, v6.4h                \n"
-                "fcvtl2 v3.4s, v6.8h                \n"
-                "fcvtl  v4.4s, v7.4h                \n"
-                "fcvtl2 v5.4s, v7.8h                \n"
-                "fcvtl  v6.4s, v8.4h                \n"
-                "fcvtl2 v7.4s, v8.8h                \n"
-                "fcvtl  v8.4s, v9.4h                \n"
-                "fcvtl2 v9.4s, v9.8h                \n"
-                "fmla   %2.4s, v2.4s, v0.s[0]       \n"
-                "fmla   %3.4s, v3.4s, v0.s[1]       \n"
-                "fmla   %4.4s, v4.4s, v0.s[2]       \n"
-                "fmla   %5.4s, v5.4s, v0.s[3]       \n"
-                "fmla   %2.4s, v6.4s, v1.s[0]       \n"
-                "fmla   %3.4s, v7.4s, v1.s[1]       \n"
-                "fmla   %4.4s, v8.4s, v1.s[2]       \n"
-                "fmla   %5.4s, v9.4s, v1.s[3]       \n"
-                : "=r"(sptr),  // %0
-                "=r"(kptr),  // %1
-                "=w"(_sum0), // %2
-                "=w"(_sum1), // %3
-                "=w"(_sum2), // %4
-                "=w"(_sum3)  // %5
-                : "0"(sptr),
-                "1"(kptr),
-                "2"(_sum0),
-                "3"(_sum1),
-                "4"(_sum2),
-                "5"(_sum3)
-                : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9");
-#else  // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            asm volatile(
-                "prfm   pldl1keep, [%0, #256]       \n"
-                "ld1    {v0.4s, v1.4s}, [%0], #32   \n"
-                "prfm   pldl1keep, [%1, #512]       \n"
-                "ld1    {v6.8h, v7.8h, v8.8h, v9.8h}, [%1], #64 \n"
-                "fcvtl  v2.4s, v6.4h                \n"
-                "fcvtl2 v3.4s, v6.8h                \n"
-                "fcvtl  v4.4s, v7.4h                \n"
-                "fcvtl2 v5.4s, v7.8h                \n"
-                "fcvtl  v6.4s, v8.4h                \n"
-                "fcvtl2 v7.4s, v8.8h                \n"
-                "fcvtl  v8.4s, v9.4h                \n"
-                "fcvtl2 v9.4s, v9.8h                \n"
-                "fmla   %2.4s, v2.4s, v0.s[0]       \n"
-                "fmla   %3.4s, v3.4s, v0.s[1]       \n"
-                "fmla   %4.4s, v4.4s, v0.s[2]       \n"
-                "fmla   %5.4s, v5.4s, v0.s[3]       \n"
-                "fmla   %2.4s, v6.4s, v1.s[0]       \n"
-                "fmla   %3.4s, v7.4s, v1.s[1]       \n"
-                "fmla   %4.4s, v8.4s, v1.s[2]       \n"
-                "fmla   %5.4s, v9.4s, v1.s[3]       \n"
-                : "=r"(sptr),  // %0
-                "=r"(kptr),  // %1
-                "=w"(_sum0), // %2
-                "=w"(_sum1), // %3
-                "=w"(_sum2), // %4
-                "=w"(_sum3)  // %5
-                : "0"(sptr),
-                "1"(kptr),
-                "2"(_sum0),
-                "3"(_sum1),
-                "4"(_sum2),
-                "5"(_sum3)
-                : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9");
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-#else  // __aarch64__
-            asm volatile(
-                "pld        [%0, #256]          \n"
-                "vld1.f32   {d0-d3}, [%0 :128]! \n"
-                "pld        [%1, #512]          \n"
-                "vldm       %1!, {d12-d19}      \n"
-                "vcvt.f32.f16 q2, d12           \n"
-                "vcvt.f32.f16 q3, d13           \n"
-                "vcvt.f32.f16 q4, d14           \n"
-                "vcvt.f32.f16 q5, d15           \n"
-                "vcvt.f32.f16 q6, d16           \n"
-                "vcvt.f32.f16 q7, d17           \n"
-                "vcvt.f32.f16 q8, d18           \n"
-                "vcvt.f32.f16 q9, d19           \n"
-                "vmla.f32   %q2, q2, d0[0]      \n"
-                "vmla.f32   %q3, q3, d0[1]      \n"
-                "vmla.f32   %q4, q4, d1[0]      \n"
-                "vmla.f32   %q5, q5, d1[1]      \n"
-                "vmla.f32   %q2, q6, d2[0]      \n"
-                "vmla.f32   %q3, q7, d2[1]      \n"
-                "vmla.f32   %q4, q8, d3[0]      \n"
-                "vmla.f32   %q5, q9, d3[1]      \n"
-                : "=r"(sptr),  // %0
-                "=r"(kptr),  // %1
-                "=w"(_sum0), // %2
-                "=w"(_sum1), // %3
-                "=w"(_sum2), // %4
-                "=w"(_sum3)  // %5
-                : "0"(sptr),
-                "1"(kptr),
-                "2"(_sum0),
-                "3"(_sum1),
-                "4"(_sum2),
-                "5"(_sum3)
-                : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9");
-#endif // __aarch64__
-        }
-#endif // NCNN_GNU_INLINE_ASM
-        for (; i + 3 < num_input; i += 4)
-        {
-#if __ARM_FEATURE_FP16_FML
-            float16x4_t _val = vld1_f16(sptr);
-            float16x8_t _w01 = vld1q_f16(kptr);
-            float16x8_t _w23 = vld1q_f16(kptr + 8);
-
-            _sum0 = vfmlalq_lane_low_f16(_sum0, _w01, _val, 0);
-            _sum1 = vfmlalq_lane_high_f16(_sum1, _w01, _val, 1);
-            _sum2 = vfmlalq_lane_low_f16(_sum2, _w23, _val, 2);
-            _sum3 = vfmlalq_lane_high_f16(_sum3, _w23, _val, 3);
-#else // __ARM_FEATURE_FP16_FML
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            float32x4_t _val = vcvt_f32_f16(vld1_f16(sptr));
-            float16x8_t _w01 = vld1q_f16(kptr);
-            float16x8_t _w23 = vld1q_f16(kptr + 8);
-            float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01));
-            float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01));
-            float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23));
-            float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23));
-#else
-            float32x4_t _val = vld1q_f32(sptr);
-            uint16x8_t _w01 = vld1q_u16(kptr);
-            uint16x8_t _w23 = vld1q_u16(kptr + 8);
-            float32x4_t _w0 = vcvt_f32_f16((float16x4_t)(vget_low_u16(_w01)));
-            float32x4_t _w1 = vcvt_f32_f16((float16x4_t)(vget_high_u16(_w01)));
-            float32x4_t _w2 = vcvt_f32_f16((float16x4_t)(vget_low_u16(_w23)));
-            float32x4_t _w3 = vcvt_f32_f16((float16x4_t)(vget_high_u16(_w23)));
-#endif
-
-#if __aarch64__
-            _sum0 = vfmaq_laneq_f32(_sum0, _w0, _val, 0);
-            _sum1 = vfmaq_laneq_f32(_sum1, _w1, _val, 1);
-            _sum2 = vfmaq_laneq_f32(_sum2, _w2, _val, 2);
-            _sum3 = vfmaq_laneq_f32(_sum3, _w3, _val, 3);
-#else
-            _sum0 = vmlaq_lane_f32(_sum0, _w0, vget_low_f32(_val), 0);
-            _sum1 = vmlaq_lane_f32(_sum1, _w1, vget_low_f32(_val), 1);
-            _sum2 = vmlaq_lane_f32(_sum2, _w2, vget_high_f32(_val), 0);
-            _sum3 = vmlaq_lane_f32(_sum3, _w3, vget_high_f32(_val), 1);
-#endif
-#endif // __ARM_FEATURE_FP16_FML
-
-            sptr += 4;
-            kptr += 16;
-        }
-        for (; i < num_input; i++)
-        {
-            float32x4_t _val = vdupq_n_f32((float)sptr[0]);
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            float32x4_t _w = vcvt_f32_f16(vld1_f16(kptr));
-#else
-            float32x4_t _w = vcvt_f32_f16((float16x4_t)(vld1_u16(kptr)));
-#endif
-            _sum0 = vfmaq_f32(_sum0, _val, _w);
-
-            sptr += 1;
-            kptr += 4;
-        }
-
-        _sum0 = vaddq_f32(_sum0, _sum1);
-        _sum2 = vaddq_f32(_sum2, _sum3);
-        _sum0 = vaddq_f32(_sum0, _sum2);
-
-        _sum0 = activation_ps(_sum0, activation_type, activation_params);
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        __fp16* outptr = (__fp16*)top_blob;
-        vst1_f16(outptr + p * 4, vcvt_f16_f32(_sum0));
-#else
-        float* outptr = top_blob;
-        vst1q_f32(outptr + p * 4, _sum0);
-#endif
-    }
-}
-
-static void innerproduct_fp16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt)
-{
-#if !(__ARM_FEATURE_FP16_FML || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
-#if NCNN_RUNTIME_CPU && NCNN_ARM82FP16FML && __aarch64__ && !__ARM_FEATURE_FP16_FML
-    if (ncnn::cpu_support_arm_asimdfhm())
-    {
-        innerproduct_fp16s_neon_asimdfhm(bottom_blob, top_blob, weight_data_fp16, bias_data, activation_type, activation_params, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-    if (ncnn::cpu_support_arm_asimdhp())
-    {
-        innerproduct_fp16s_neon_asimdhp(bottom_blob, top_blob, weight_data_fp16, bias_data, activation_type, activation_params, opt);
-        return;
-    }
-#endif
-#endif
-
-    const int num_input = bottom_blob.w * bottom_blob.elempack;
-    const int num_output = top_blob.w;
-
-    const float* bias_data_ptr = bias_data;
-
-    int nn_num_output = num_output >> 2;
-    int remain_num_output_start = nn_num_output << 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_num_output; pp++)
-    {
-        int p = pp * 4;
-
-        float sums[4] = {0.0f};
-        if (bias_data_ptr)
-        {
-            sums[0] = bias_data_ptr[p];
-            sums[1] = bias_data_ptr[p + 1];
-            sums[2] = bias_data_ptr[p + 2];
-            sums[3] = bias_data_ptr[p + 3];
-        }
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        const __fp16* sptr = bottom_blob;
-        const __fp16* kptr0 = weight_data_fp16.row<const __fp16>(p);
-        const __fp16* kptr1 = weight_data_fp16.row<const __fp16>(p + 1);
-        const __fp16* kptr2 = weight_data_fp16.row<const __fp16>(p + 2);
-        const __fp16* kptr3 = weight_data_fp16.row<const __fp16>(p + 3);
-#else
-        const float* sptr = bottom_blob;
-        const unsigned short* kptr0 = weight_data_fp16.row<const unsigned short>(p);
-        const unsigned short* kptr1 = weight_data_fp16.row<const unsigned short>(p + 1);
-        const unsigned short* kptr2 = weight_data_fp16.row<const unsigned short>(p + 2);
-        const unsigned short* kptr3 = weight_data_fp16.row<const unsigned short>(p + 3);
-#endif
-
-        int i = 0;
-
-        float32x4_t _sum0 = vdupq_n_f32(0.f);
-        float32x4_t _sum1 = vdupq_n_f32(0.f);
-        float32x4_t _sum2 = vdupq_n_f32(0.f);
-        float32x4_t _sum3 = vdupq_n_f32(0.f);
-        for (; i + 3 < num_input; i += 4)
-        {
-#if __ARM_FEATURE_FP16_FML
-            float16x4_t _val = vld1_f16(sptr);
-            float16x4_t _w0 = vld1_f16(kptr0);
-            float16x4_t _w1 = vld1_f16(kptr1);
-            float16x4_t _w2 = vld1_f16(kptr2);
-            float16x4_t _w3 = vld1_f16(kptr3);
-            float16x8_t _w01 = vcombine_f16(_w0, _w1);
-            float16x8_t _w23 = vcombine_f16(_w2, _w3);
-            float16x8_t _valval = vcombine_f16(_val, _val);
-
-            _sum0 = vfmlalq_low_f16(_sum0, _w01, _valval);
-            _sum1 = vfmlalq_high_f16(_sum1, _w01, _valval);
-            _sum2 = vfmlalq_low_f16(_sum2, _w23, _valval);
-            _sum3 = vfmlalq_high_f16(_sum3, _w23, _valval);
-#else // __ARM_FEATURE_FP16_FML
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            float32x4_t _val = vcvt_f32_f16(vld1_f16(sptr));
-            float32x4_t _w0 = vcvt_f32_f16(vld1_f16(kptr0));
-            float32x4_t _w1 = vcvt_f32_f16(vld1_f16(kptr1));
-            float32x4_t _w2 = vcvt_f32_f16(vld1_f16(kptr2));
-            float32x4_t _w3 = vcvt_f32_f16(vld1_f16(kptr3));
-#else
-            float32x4_t _val = vld1q_f32(sptr);
-            float32x4_t _w0 = vcvt_f32_f16((float16x4_t)(vld1_u16(kptr0)));
-            float32x4_t _w1 = vcvt_f32_f16((float16x4_t)(vld1_u16(kptr1)));
-            float32x4_t _w2 = vcvt_f32_f16((float16x4_t)(vld1_u16(kptr2)));
-            float32x4_t _w3 = vcvt_f32_f16((float16x4_t)(vld1_u16(kptr3)));
-#endif
-
-            _sum0 = vfmaq_f32(_sum0, _val, _w0);
-            _sum1 = vfmaq_f32(_sum1, _val, _w1);
-            _sum2 = vfmaq_f32(_sum2, _val, _w2);
-            _sum3 = vfmaq_f32(_sum3, _val, _w3);
-#endif // __ARM_FEATURE_FP16_FML
-
-            sptr += 4;
-            kptr0 += 4;
-            kptr1 += 4;
-            kptr2 += 4;
-            kptr3 += 4;
-        }
-
-#if __aarch64__
-        sums[0] += vaddvq_f32(_sum0);
-        sums[1] += vaddvq_f32(_sum1);
-        sums[2] += vaddvq_f32(_sum2);
-        sums[3] += vaddvq_f32(_sum3);
-#else
-        float32x2_t _sum0ss = vadd_f32(vget_low_f32(_sum0), vget_high_f32(_sum0));
-        float32x2_t _sum1ss = vadd_f32(vget_low_f32(_sum1), vget_high_f32(_sum1));
-        float32x2_t _sum2ss = vadd_f32(vget_low_f32(_sum2), vget_high_f32(_sum2));
-        float32x2_t _sum3ss = vadd_f32(vget_low_f32(_sum3), vget_high_f32(_sum3));
-        float32x2_t _sum01ss = vpadd_f32(_sum0ss, _sum1ss);
-        float32x2_t _sum23ss = vpadd_f32(_sum2ss, _sum3ss);
-        sums[0] += vget_lane_f32(_sum01ss, 0);
-        sums[1] += vget_lane_f32(_sum01ss, 1);
-        sums[2] += vget_lane_f32(_sum23ss, 0);
-        sums[3] += vget_lane_f32(_sum23ss, 1);
-#endif
-
-        for (; i < num_input; i++)
-        {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            sums[0] += (float)(*sptr) * (float)(*kptr0);
-            sums[1] += (float)(*sptr) * (float)(*kptr1);
-            sums[2] += (float)(*sptr) * (float)(*kptr2);
-            sums[3] += (float)(*sptr) * (float)(*kptr3);
-#else
-            sums[0] += *sptr * float16_to_float32(*kptr0);
-            sums[1] += *sptr * float16_to_float32(*kptr1);
-            sums[2] += *sptr * float16_to_float32(*kptr2);
-            sums[3] += *sptr * float16_to_float32(*kptr3);
-#endif
-
-            sptr++;
-            kptr0++;
-            kptr1++;
-            kptr2++;
-            kptr3++;
-        }
-
-        float32x4_t _sum = vld1q_f32(sums);
-
-        _sum = activation_ps(_sum, activation_type, activation_params);
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        __fp16* outptr = (__fp16*)top_blob;
-        vst1_f16(outptr + p, vcvt_f16_f32(_sum));
-#else
-        float* outptr = top_blob;
-        vst1q_f32(outptr + p, _sum);
-#endif
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_num_output_start; p < num_output; p++)
-    {
-        float sum = 0.f;
-
-        if (bias_data_ptr)
-            sum = bias_data_ptr[p];
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        const __fp16* sptr = bottom_blob;
-        const __fp16* kptr = weight_data_fp16.row<const __fp16>(p);
-#else
-        const float* sptr = bottom_blob;
-        const unsigned short* kptr = weight_data_fp16.row<const unsigned short>(p);
-#endif
-
-        int i = 0;
-
-        float32x4_t _sum = vdupq_n_f32(0.f);
-        for (; i + 3 < num_input; i += 4)
-        {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            float32x4_t _val = vcvt_f32_f16(vld1_f16(sptr));
-            float32x4_t _w = vcvt_f32_f16(vld1_f16(kptr));
-#else
-            float32x4_t _val = vld1q_f32(sptr);
-            float32x4_t _w = vcvt_f32_f16((float16x4_t)(vld1_u16(kptr)));
-#endif
-            _sum = vfmaq_f32(_sum, _val, _w);
-
-            sptr += 4;
-            kptr += 4;
-        }
-        for (; i < num_input; i++)
-        {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            sum += (float)(*sptr) * (float)(*kptr);
-#else
-            sum += *sptr * float16_to_float32(*kptr);
-#endif
-            sptr++;
-            kptr++;
-        }
-
-#if __aarch64__
-        sum += vaddvq_f32(_sum);
-#else
-        float32x2_t _sumss = vadd_f32(vget_low_f32(_sum), vget_high_f32(_sum));
-        _sumss = vpadd_f32(_sumss, _sumss);
-        sum += vget_lane_f32(_sumss, 0);
-#endif // __aarch64__
-
-        sum = activation_ss(sum, activation_type, activation_params);
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        __fp16* outptr = (__fp16*)top_blob;
-        outptr[p] = (__fp16)sum;
-#else
-        float* outptr = top_blob;
-        outptr[p] = sum;
-#endif
-    }
-}
-
-static void innerproduct_transform_kernel_fp16s_neon(const Mat& weight_data, Mat& weight_data_tm, int num_input, int num_output, const Option& opt)
-{
-#if !(__ARM_FEATURE_FP16_FML || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
-#if NCNN_RUNTIME_CPU && NCNN_ARM82FP16FML && __aarch64__ && !__ARM_FEATURE_FP16_FML
-    if (ncnn::cpu_support_arm_asimdfhm())
-    {
-        innerproduct_transform_kernel_fp16s_neon_asimdfhm(weight_data, weight_data_tm, num_input, num_output, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-    if (ncnn::cpu_support_arm_asimdhp())
-    {
-        innerproduct_transform_kernel_fp16s_neon_asimdhp(weight_data, weight_data_tm, num_input, num_output, opt);
-        return;
-    }
-#endif
-#endif
-
-    int out_elempack = 1;
-    if (opt.use_packing_layout)
-    {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        out_elempack = opt.use_fp16_arithmetic && num_output % 8 == 0 ? 8 : num_output % 4 == 0 ? 4 : 1;
-#else
-        out_elempack = num_output % 4 == 0 ? 4 : 1;
-#endif
-    }
-
-    // src = inch-outch
-    // dst = pb-inch-outch/pb
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-    if (out_elempack == 8)
-    {
-        Mat weight_data_r2 = weight_data.reshape(num_input, num_output);
-
-        weight_data_tm.create(num_input, num_output / 8, (size_t)16u, 8);
-
-        for (int q = 0; q + 7 < num_output; q += 8)
-        {
-            unsigned short* g0 = weight_data_tm.row<unsigned short>(q / 8);
-
-            const float* k0 = weight_data_r2.row(q);
-            const float* k1 = weight_data_r2.row(q + 1);
-            const float* k2 = weight_data_r2.row(q + 2);
-            const float* k3 = weight_data_r2.row(q + 3);
-            const float* k4 = weight_data_r2.row(q + 4);
-            const float* k5 = weight_data_r2.row(q + 5);
-            const float* k6 = weight_data_r2.row(q + 6);
-            const float* k7 = weight_data_r2.row(q + 7);
-
-            int p = 0;
-#if NCNN_GNU_INLINE_ASM
-            for (; p + 7 < num_input; p += 8)
-            {
-                // transpose 8x8
-                asm volatile(
-                    "ld1    {v0.4s, v1.4s}, [%0], #32   \n"
-                    "ld1    {v2.4s, v3.4s}, [%1], #32   \n"
-                    "ld1    {v4.4s, v5.4s}, [%2], #32   \n"
-                    "ld1    {v6.4s, v7.4s}, [%3], #32   \n"
-                    "ld1    {v8.4s, v9.4s}, [%4], #32   \n"
-                    "ld1    {v10.4s, v11.4s}, [%5], #32 \n"
-                    "ld1    {v12.4s, v13.4s}, [%6], #32 \n"
-                    "ld1    {v14.4s, v15.4s}, [%7], #32 \n"
-
-                    "fcvtn  v0.4h, v0.4s            \n"
-                    "fcvtn2 v0.8h, v1.4s            \n"
-                    "fcvtn  v1.4h, v2.4s            \n"
-                    "fcvtn2 v1.8h, v3.4s            \n"
-                    "fcvtn  v2.4h, v4.4s            \n"
-                    "fcvtn2 v2.8h, v5.4s            \n"
-                    "fcvtn  v3.4h, v6.4s            \n"
-                    "fcvtn2 v3.8h, v7.4s            \n"
-                    "fcvtn  v4.4h, v8.4s            \n"
-                    "fcvtn2 v4.8h, v9.4s            \n"
-                    "fcvtn  v5.4h, v10.4s           \n"
-                    "fcvtn2 v5.8h, v11.4s           \n"
-                    "fcvtn  v6.4h, v12.4s           \n"
-                    "fcvtn2 v6.8h, v13.4s           \n"
-                    "fcvtn  v7.4h, v14.4s           \n"
-                    "fcvtn2 v7.8h, v15.4s           \n"
-
-                    "zip1   v16.8h, v0.8h, v4.8h    \n"
-                    "zip2   v20.8h, v0.8h, v4.8h    \n"
-                    "zip1   v17.8h, v1.8h, v5.8h    \n"
-                    "zip2   v21.8h, v1.8h, v5.8h    \n"
-                    "zip1   v18.8h, v2.8h, v6.8h    \n"
-                    "zip2   v22.8h, v2.8h, v6.8h    \n"
-                    "zip1   v19.8h, v3.8h, v7.8h    \n"
-                    "zip2   v23.8h, v3.8h, v7.8h    \n"
-
-                    "st4    {v16.8h, v17.8h, v18.8h, v19.8h}, [%8], #64 \n"
-                    "st4    {v20.8h, v21.8h, v22.8h, v23.8h}, [%8], #64 \n"
-                    : "=r"(k0), // %0
-                    "=r"(k1), // %1
-                    "=r"(k2), // %2
-                    "=r"(k3), // %3
-                    "=r"(k4), // %4
-                    "=r"(k5), // %5
-                    "=r"(k6), // %6
-                    "=r"(k7), // %7
-                    "=r"(g0)  // %8
-                    : "0"(k0),
-                    "1"(k1),
-                    "2"(k2),
-                    "3"(k3),
-                    "4"(k4),
-                    "5"(k5),
-                    "6"(k6),
-                    "7"(k7),
-                    "8"(g0)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-            }
-#endif // NCNN_GNU_INLINE_ASM
-            for (; p < num_input; p++)
-            {
-                g0[0] = float32_to_float16(*k0++);
-                g0[1] = float32_to_float16(*k1++);
-                g0[2] = float32_to_float16(*k2++);
-                g0[3] = float32_to_float16(*k3++);
-                g0[4] = float32_to_float16(*k4++);
-                g0[5] = float32_to_float16(*k5++);
-                g0[6] = float32_to_float16(*k6++);
-                g0[7] = float32_to_float16(*k7++);
-                g0 += 8;
-            }
-        }
-    }
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-
-    if (out_elempack == 4)
-    {
-        Mat weight_data_r2 = weight_data.reshape(num_input, num_output);
-
-        weight_data_tm.create(num_input, num_output / 4, (size_t)8u, 4);
-
-        for (int q = 0; q + 3 < num_output; q += 4)
-        {
-            unsigned short* g0 = weight_data_tm.row<unsigned short>(q / 4);
-
-            const float* k0 = weight_data_r2.row(q);
-            const float* k1 = weight_data_r2.row(q + 1);
-            const float* k2 = weight_data_r2.row(q + 2);
-            const float* k3 = weight_data_r2.row(q + 3);
-
-            int p = 0;
-            for (; p + 3 < num_input; p += 4)
-            {
-                // transpose 4x4
-                uint16x4x4_t _p;
-                _p.val[0] = (uint16x4_t)(vcvt_f16_f32(vld1q_f32(k0)));
-                _p.val[1] = (uint16x4_t)(vcvt_f16_f32(vld1q_f32(k1)));
-                _p.val[2] = (uint16x4_t)(vcvt_f16_f32(vld1q_f32(k2)));
-                _p.val[3] = (uint16x4_t)(vcvt_f16_f32(vld1q_f32(k3)));
-                vst4_u16(g0, _p);
-
-                k0 += 4;
-                k1 += 4;
-                k2 += 4;
-                k3 += 4;
-                g0 += 16;
-            }
-            for (; p < num_input; p++)
-            {
-                g0[0] = float32_to_float16(*k0++);
-                g0[1] = float32_to_float16(*k1++);
-                g0[2] = float32_to_float16(*k2++);
-                g0[3] = float32_to_float16(*k3++);
-                g0 += 4;
-            }
-        }
-    }
-
-    if (out_elempack == 1)
-    {
-        Mat weight_data_r2 = weight_data.reshape(num_input, num_output);
-        ncnn::cast_float32_to_float16(weight_data_r2, weight_data_tm, opt);
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/innerproduct_gemm_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/innerproduct_gemm_fp16s.h
deleted file mode 100644
index 1724cfdf..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/innerproduct_gemm_fp16s.h
+++ /dev/null
@@ -1,458 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if !(__ARM_FEATURE_FP16_FML || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
-#if NCNN_RUNTIME_CPU && NCNN_ARM82FP16FML && __aarch64__ && !__ARM_FEATURE_FP16_FML
-void innerproduct_gemm_fp16s_neon_asimdfhm(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt);
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-void innerproduct_gemm_fp16s_neon_asimdhp(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt);
-#endif
-#endif
-
-static void innerproduct_gemm_fp16s_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt)
-{
-#if !(__ARM_FEATURE_FP16_FML || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
-#if NCNN_RUNTIME_CPU && NCNN_ARM82FP16FML && __aarch64__ && !__ARM_FEATURE_FP16_FML
-    if (ncnn::cpu_support_arm_asimdfhm())
-    {
-        innerproduct_gemm_fp16s_neon_asimdfhm(bottom_blob, top_blob, weight_data_fp16, bias_data, activation_type, activation_params, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_ARM82 && __aarch64__ && !__ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-    if (ncnn::cpu_support_arm_asimdhp())
-    {
-        innerproduct_gemm_fp16s_neon_asimdhp(bottom_blob, top_blob, weight_data_fp16, bias_data, activation_type, activation_params, opt);
-        return;
-    }
-#endif
-#endif
-
-    const int num_input = bottom_blob.w;
-    const int elempack = bottom_blob.elempack;
-    const int num_output = top_blob.w;
-    const int h = bottom_blob.h;
-
-    const float* bias_data_ptr = bias_data;
-
-    int num_output_elempack = 1;
-    if (opt.use_packing_layout)
-    {
-        num_output_elempack = num_output % 4 == 0 ? 4 : 1;
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int j = 0; j < h; j++)
-    {
-        if (elempack == 4 && num_output_elempack == 4)
-        {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            __fp16* outptr = top_blob.row<__fp16>(j);
-#else
-            float* outptr = top_blob.row(j);
-#endif
-
-            for (int p = 0; p < num_output / num_output_elempack; p++)
-            {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                const __fp16* m = bottom_blob.row<const __fp16>(j);
-                const __fp16* kptr = weight_data_fp16.row<const __fp16>(p);
-#else
-                const float* m = bottom_blob.row(j);
-                const unsigned short* kptr = weight_data_fp16.row<const unsigned short>(p);
-#endif
-
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                float32x4_t _sum2 = vdupq_n_f32(0.f);
-                float32x4_t _sum3 = vdupq_n_f32(0.f);
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = vdupq_n_f32(bias_data_ptr[p * 4 + 0]);
-                    _sum1 = vdupq_n_f32(bias_data_ptr[p * 4 + 1]);
-                    _sum2 = vdupq_n_f32(bias_data_ptr[p * 4 + 2]);
-                    _sum3 = vdupq_n_f32(bias_data_ptr[p * 4 + 3]);
-                }
-
-                int i = 0;
-                for (; i < num_input; i++)
-                {
-#if __ARM_FEATURE_FP16_FML
-                    float16x4_t _val = vld1_f16(m);
-                    float16x4_t _w = vld1_f16(kptr);
-                    float16x8_t _valval = vcombine_f16(_val, _val);
-
-                    _sum0 = vfmlalq_lane_low_f16(_sum0, _valval, _w, 0);
-                    _sum1 = vfmlalq_lane_low_f16(_sum1, _valval, _w, 1);
-                    _sum2 = vfmlalq_lane_low_f16(_sum2, _valval, _w, 2);
-                    _sum3 = vfmlalq_lane_low_f16(_sum3, _valval, _w, 3);
-#else // __ARM_FEATURE_FP16_FML
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                    float32x4_t _val = vcvt_f32_f16(vld1_f16(m));
-                    float32x4_t _w = vcvt_f32_f16(vld1_f16(kptr));
-#else
-                    float32x4_t _val = vld1q_f32(m);
-                    float32x4_t _w = vcvt_f32_f16((float16x4_t)(vld1_u16(kptr)));
-#endif
-
-#if __aarch64__
-                    _sum0 = vfmaq_laneq_f32(_sum0, _val, _w, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _val, _w, 1);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _val, _w, 2);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _val, _w, 3);
-#else
-                    _sum0 = vmlaq_lane_f32(_sum0, _val, vget_low_f32(_w), 0);
-                    _sum1 = vmlaq_lane_f32(_sum1, _val, vget_low_f32(_w), 1);
-                    _sum2 = vmlaq_lane_f32(_sum2, _val, vget_high_f32(_w), 0);
-                    _sum3 = vmlaq_lane_f32(_sum3, _val, vget_high_f32(_w), 1);
-#endif
-#endif // __ARM_FEATURE_FP16_FML
-
-                    m += 4;
-                    kptr += 4;
-                }
-
-                _sum0 = activation_ps(_sum0, activation_type, activation_params);
-                _sum1 = activation_ps(_sum1, activation_type, activation_params);
-                _sum2 = activation_ps(_sum2, activation_type, activation_params);
-                _sum3 = activation_ps(_sum3, activation_type, activation_params);
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                vst1_f16(outptr, vcvt_f16_f32(_sum0));
-                vst1_f16(outptr + 4, vcvt_f16_f32(_sum1));
-                vst1_f16(outptr + 8, vcvt_f16_f32(_sum2));
-                vst1_f16(outptr + 12, vcvt_f16_f32(_sum3));
-#else
-                vst1q_f32(outptr, _sum0);
-                vst1q_f32(outptr + 4, _sum1);
-                vst1q_f32(outptr + 8, _sum2);
-                vst1q_f32(outptr + 12, _sum3);
-#endif
-                outptr += 16;
-            }
-        }
-
-        if (elempack == 1 && num_output_elempack == 4)
-        {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            __fp16* outptr = top_blob.row<__fp16>(j);
-#else
-            float* outptr = top_blob.row(j);
-#endif
-
-            for (int p = 0; p < num_output / num_output_elempack; p++)
-            {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                const __fp16* m = bottom_blob.row<const __fp16>(j);
-                const __fp16* kptr = weight_data_fp16.row<const __fp16>(p);
-#else
-                const float* m = bottom_blob.row(j);
-                const unsigned short* kptr = weight_data_fp16.row<const unsigned short>(p);
-#endif
-
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = vld1q_f32(bias_data_ptr + p * 4);
-                }
-
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                float32x4_t _sum2 = vdupq_n_f32(0.f);
-                float32x4_t _sum3 = vdupq_n_f32(0.f);
-
-                int i = 0;
-                for (; i + 3 < num_input; i += 4)
-                {
-#if __ARM_FEATURE_FP16_FML
-                    float16x4_t _val = vld1_f16(m);
-                    float16x8_t _w01 = vld1q_f16(kptr);
-                    float16x8_t _w23 = vld1q_f16(kptr + 8);
-
-                    _sum0 = vfmlalq_lane_low_f16(_sum0, _w01, _val, 0);
-                    _sum1 = vfmlalq_lane_high_f16(_sum1, _w01, _val, 1);
-                    _sum2 = vfmlalq_lane_low_f16(_sum2, _w23, _val, 2);
-                    _sum3 = vfmlalq_lane_high_f16(_sum3, _w23, _val, 3);
-#else // __ARM_FEATURE_FP16_FML
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                    float32x4_t _val = vcvt_f32_f16(vld1_f16(m));
-                    float16x8_t _w01 = vld1q_f16(kptr);
-                    float16x8_t _w23 = vld1q_f16(kptr + 8);
-                    float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01));
-                    float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01));
-                    float32x4_t _w2 = vcvt_f32_f16(vget_low_f16(_w23));
-                    float32x4_t _w3 = vcvt_f32_f16(vget_high_f16(_w23));
-#else
-                    float32x4_t _val = vld1q_f32(m);
-                    uint16x8_t _w01 = vld1q_u16(kptr);
-                    uint16x8_t _w23 = vld1q_u16(kptr + 8);
-                    float32x4_t _w0 = vcvt_f32_f16((float16x4_t)(vget_low_u16(_w01)));
-                    float32x4_t _w1 = vcvt_f32_f16((float16x4_t)(vget_high_u16(_w01)));
-                    float32x4_t _w2 = vcvt_f32_f16((float16x4_t)(vget_low_u16(_w23)));
-                    float32x4_t _w3 = vcvt_f32_f16((float16x4_t)(vget_high_u16(_w23)));
-#endif
-
-#if __aarch64__
-                    _sum0 = vfmaq_laneq_f32(_sum0, _w0, _val, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _w1, _val, 1);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _w2, _val, 2);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _w3, _val, 3);
-#else
-                    _sum0 = vmlaq_lane_f32(_sum0, _w0, vget_low_f32(_val), 0);
-                    _sum1 = vmlaq_lane_f32(_sum1, _w1, vget_low_f32(_val), 1);
-                    _sum2 = vmlaq_lane_f32(_sum2, _w2, vget_high_f32(_val), 0);
-                    _sum3 = vmlaq_lane_f32(_sum3, _w3, vget_high_f32(_val), 1);
-#endif
-#endif // __ARM_FEATURE_FP16_FML
-
-                    m += 4;
-                    kptr += 16;
-                }
-                for (; i < num_input; i++)
-                {
-                    float32x4_t _val = vdupq_n_f32((float)m[0]);
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                    float32x4_t _w = vcvt_f32_f16(vld1_f16(kptr));
-#else
-                    float32x4_t _w = vcvt_f32_f16((float16x4_t)(vld1_u16(kptr)));
-#endif
-                    _sum0 = vfmaq_f32(_sum0, _val, _w);
-
-                    m += 1;
-                    kptr += 4;
-                }
-
-                _sum0 = vaddq_f32(_sum0, _sum1);
-                _sum2 = vaddq_f32(_sum2, _sum3);
-                _sum0 = vaddq_f32(_sum0, _sum2);
-
-                _sum0 = activation_ps(_sum0, activation_type, activation_params);
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                vst1_f16(outptr, vcvt_f16_f32(_sum0));
-#else
-                vst1q_f32(outptr, _sum0);
-#endif
-                outptr += 4;
-            }
-        }
-
-        if (elempack == 4 && num_output_elempack == 1)
-        {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            __fp16* outptr = top_blob.row<__fp16>(j);
-#else
-            float* outptr = top_blob.row(j);
-#endif
-
-            for (int p = 0; p < num_output; p++)
-            {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                const __fp16* m = bottom_blob.row<const __fp16>(j);
-                const __fp16* kptr = weight_data_fp16.row<const __fp16>(p);
-#else
-                const float* m = bottom_blob.row(j);
-                const unsigned short* kptr = weight_data_fp16.row<const unsigned short>(p);
-#endif
-
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                float32x4_t _sum2 = vdupq_n_f32(0.f);
-                float32x4_t _sum3 = vdupq_n_f32(0.f);
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = vdupq_n_f32(bias_data_ptr[p]);
-                }
-
-                int i = 0;
-                for (; i + 3 < num_input; i += 4)
-                {
-#if __ARM_FEATURE_FP16_FML
-                    float16x8_t _val01 = vld1q_f16(m);
-                    float16x8_t _val23 = vld1q_f16(m + 8);
-                    float16x4_t _w = vld1_f16(kptr);
-
-                    _sum0 = vfmlalq_lane_low_f16(_sum0, _val01, _w, 0);
-                    _sum1 = vfmlalq_lane_high_f16(_sum1, _val01, _w, 1);
-                    _sum2 = vfmlalq_lane_low_f16(_sum2, _val23, _w, 2);
-                    _sum3 = vfmlalq_lane_high_f16(_sum3, _val23, _w, 3);
-#else // __ARM_FEATURE_FP16_FML
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                    float32x4_t _val0 = vcvt_f32_f16(vld1_f16(m));
-                    float32x4_t _val1 = vcvt_f32_f16(vld1_f16(m + 4));
-                    float32x4_t _val2 = vcvt_f32_f16(vld1_f16(m + 8));
-                    float32x4_t _val3 = vcvt_f32_f16(vld1_f16(m + 12));
-                    float32x4_t _w = vcvt_f32_f16(vld1_f16(kptr));
-#else
-                    float32x4_t _val0 = vld1q_f32(m);
-                    float32x4_t _val1 = vld1q_f32(m + 4);
-                    float32x4_t _val2 = vld1q_f32(m + 8);
-                    float32x4_t _val3 = vld1q_f32(m + 12);
-                    float32x4_t _w = vcvt_f32_f16((float16x4_t)(vld1_u16(kptr)));
-#endif
-
-#if __aarch64__
-                    _sum0 = vfmaq_laneq_f32(_sum0, _val0, _w, 0);
-                    _sum1 = vfmaq_laneq_f32(_sum1, _val1, _w, 1);
-                    _sum2 = vfmaq_laneq_f32(_sum2, _val2, _w, 2);
-                    _sum3 = vfmaq_laneq_f32(_sum3, _val3, _w, 3);
-#else
-                    _sum0 = vmlaq_lane_f32(_sum0, _val0, vget_low_f32(_w), 0);
-                    _sum1 = vmlaq_lane_f32(_sum1, _val1, vget_low_f32(_w), 1);
-                    _sum2 = vmlaq_lane_f32(_sum2, _val2, vget_high_f32(_w), 0);
-                    _sum3 = vmlaq_lane_f32(_sum3, _val3, vget_high_f32(_w), 1);
-#endif
-#endif // __ARM_FEATURE_FP16_FML
-
-                    m += 16;
-                    kptr += 4;
-                }
-                for (; i < num_input; i++)
-                {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                    float32x4_t _val = vcvt_f32_f16(vld1_f16(m));
-                    float32x4_t _k = vdupq_n_f32((float)(kptr[0]));
-#else
-                    float32x4_t _val = vld1q_f32(m);
-                    float32x4_t _k = vdupq_n_f32(float16_to_float32(kptr[0]));
-#endif
-                    _sum0 = vfmaq_f32(_sum0, _val, _k);
-
-                    m += 4;
-                    kptr += 1;
-                }
-
-                _sum0 = vaddq_f32(_sum0, _sum1);
-                _sum2 = vaddq_f32(_sum2, _sum3);
-                _sum0 = vaddq_f32(_sum0, _sum2);
-
-                _sum0 = activation_ps(_sum0, activation_type, activation_params);
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                vst1_f16(outptr, vcvt_f16_f32(_sum0));
-#else
-                vst1q_f32(outptr, _sum0);
-#endif
-                outptr += 4;
-            }
-        }
-
-        if (elempack == 1 && num_output_elempack == 1)
-        {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-            __fp16* outptr = top_blob.row<__fp16>(j);
-#else
-            float* outptr = top_blob.row(j);
-#endif
-
-            for (int p = 0; p < num_output; p++)
-            {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                const __fp16* m = bottom_blob.row<const __fp16>(j);
-                const __fp16* kptr = weight_data_fp16.row<const __fp16>(p);
-#else
-                const float* m = bottom_blob.row(j);
-                const unsigned short* kptr = weight_data_fp16.row<const unsigned short>(p);
-#endif
-
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                int i = 0;
-                float32x4_t _sum0 = vdupq_n_f32(0.f);
-                float32x4_t _sum1 = vdupq_n_f32(0.f);
-                for (; i + 7 < num_input; i += 8)
-                {
-#if __ARM_FEATURE_FP16_FML
-                    float16x8_t _val01 = vld1q_f16(m);
-                    float16x8_t _w01 = vld1q_f16(kptr);
-
-                    _sum0 = vfmlalq_low_f16(_sum0, _val01, _w01);
-                    _sum1 = vfmlalq_high_f16(_sum1, _val01, _w01);
-#else // __ARM_FEATURE_FP16_FML
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                    float16x8_t _val01 = vld1q_f16(m);
-                    float16x8_t _w01 = vld1q_f16(kptr);
-                    float32x4_t _val0 = vcvt_f32_f16(vget_low_f16(_val01));
-                    float32x4_t _val1 = vcvt_f32_f16(vget_high_f16(_val01));
-                    float32x4_t _w0 = vcvt_f32_f16(vget_low_f16(_w01));
-                    float32x4_t _w1 = vcvt_f32_f16(vget_high_f16(_w01));
-#else
-                    float32x4_t _val0 = vld1q_f32(m);
-                    float32x4_t _val1 = vld1q_f32(m + 4);
-                    uint16x8_t _w01 = vld1q_u16(kptr);
-                    float32x4_t _w0 = vcvt_f32_f16((float16x4_t)(vget_low_u16(_w01)));
-                    float32x4_t _w1 = vcvt_f32_f16((float16x4_t)(vget_high_u16(_w01)));
-#endif
-
-                    _sum0 = vfmaq_f32(_sum0, _val0, _w0);
-                    _sum1 = vfmaq_f32(_sum1, _val1, _w1);
-#endif // __ARM_FEATURE_FP16_FML
-
-                    m += 8;
-                    kptr += 8;
-                }
-                _sum0 = vaddq_f32(_sum0, _sum1);
-                for (; i + 3 < num_input; i += 4)
-                {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                    float32x4_t _val = vcvt_f32_f16(vld1_f16(m));
-                    float32x4_t _w = vcvt_f32_f16(vld1_f16(kptr));
-#else
-                    float32x4_t _val = vld1q_f32(m);
-                    float32x4_t _w = vcvt_f32_f16((float16x4_t)(vld1_u16(kptr)));
-#endif
-
-                    _sum0 = vfmaq_f32(_sum0, _val, _w);
-
-                    m += 4;
-                    kptr += 4;
-                }
-#if __aarch64__
-                sum += vaddvq_f32(_sum0);
-#else
-                float32x2_t _ss = vadd_f32(vget_low_f32(_sum0), vget_high_f32(_sum0));
-                _ss = vpadd_f32(_ss, _ss);
-                sum += vget_lane_f32(_ss, 0);
-#endif
-                for (; i < num_input; i++)
-                {
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                    sum += (float)(*m++) * (float)(*kptr++);
-#else
-                    sum += *m++ * float16_to_float32(*kptr++);
-#endif
-                }
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-                outptr[0] = (__fp16)sum;
-#else
-                outptr[0] = sum;
-#endif
-                outptr += 1;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/instancenorm_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/instancenorm_arm.h
deleted file mode 100644
index 98dec71a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/instancenorm_arm.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INSTANCENORM_ARM_H
-#define LAYER_INSTANCENORM_ARM_H
-
-#include "instancenorm.h"
-
-namespace ncnn {
-
-class InstanceNorm_arm : public InstanceNorm
-{
-public:
-    InstanceNorm_arm();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_inplace_bf16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INSTANCENORM_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_arm.h
deleted file mode 100644
index 6c15c738..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_arm.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INTERP_ARM_H
-#define LAYER_INTERP_ARM_H
-
-#include "interp.h"
-
-namespace ncnn {
-
-class Interp_arm : public Interp
-{
-public:
-    Interp_arm();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_fp16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-    int forward_fp16sa(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_bf16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INTERP_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic.h
deleted file mode 100644
index efff2d4d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic.h
+++ /dev/null
@@ -1,261 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static inline void interpolate_cubic(float fx, float* coeffs)
-{
-    const float A = -0.75f;
-
-    float fx0 = fx + 1;
-    float fx1 = fx;
-    float fx2 = 1 - fx;
-    // float fx3 = 2 - fx;
-
-    coeffs[0] = A * fx0 * fx0 * fx0 - 5 * A * fx0 * fx0 + 8 * A * fx0 - 4 * A;
-    coeffs[1] = (A + 2) * fx1 * fx1 * fx1 - (A + 3) * fx1 * fx1 + 1;
-    coeffs[2] = (A + 2) * fx2 * fx2 * fx2 - (A + 3) * fx2 * fx2 + 1;
-    coeffs[3] = 1.f - coeffs[0] - coeffs[1] - coeffs[2];
-}
-
-static void cubic_coeffs(int w, int outw, int* xofs, float* alpha, int align_corner)
-{
-    double scale = (double)w / outw;
-    if (align_corner)
-    {
-        scale = (double)(w - 1) / (outw - 1);
-    }
-
-    for (int dx = 0; dx < outw; dx++)
-    {
-        float fx = (float)((dx + 0.5) * scale - 0.5);
-        if (align_corner)
-        {
-            fx = (float)(dx * scale);
-        }
-
-        int sx = static_cast<int>(floor(fx));
-        fx -= sx;
-
-        interpolate_cubic(fx, alpha + dx * 4);
-
-        if (sx <= -1)
-        {
-            sx = 1;
-            alpha[dx * 4 + 0] = 1.f - alpha[dx * 4 + 3];
-            alpha[dx * 4 + 1] = alpha[dx * 4 + 3];
-            alpha[dx * 4 + 2] = 0.f;
-            alpha[dx * 4 + 3] = 0.f;
-        }
-        if (sx == 0)
-        {
-            sx = 1;
-            alpha[dx * 4 + 0] = alpha[dx * 4 + 0] + alpha[dx * 4 + 1];
-            alpha[dx * 4 + 1] = alpha[dx * 4 + 2];
-            alpha[dx * 4 + 2] = alpha[dx * 4 + 3];
-            alpha[dx * 4 + 3] = 0.f;
-        }
-        if (sx == w - 2)
-        {
-            sx = w - 3;
-            alpha[dx * 4 + 3] = alpha[dx * 4 + 2] + alpha[dx * 4 + 3];
-            alpha[dx * 4 + 2] = alpha[dx * 4 + 1];
-            alpha[dx * 4 + 1] = alpha[dx * 4 + 0];
-            alpha[dx * 4 + 0] = 0.f;
-        }
-        if (sx >= w - 1)
-        {
-            sx = w - 3;
-            alpha[dx * 4 + 3] = 1.f - alpha[dx * 4 + 0];
-            alpha[dx * 4 + 2] = alpha[dx * 4 + 0];
-            alpha[dx * 4 + 1] = 0.f;
-            alpha[dx * 4 + 0] = 0.f;
-        }
-
-        xofs[dx] = sx;
-    }
-}
-
-static void resize_bicubic_image(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    Mat rowsbuf2(w);
-    Mat rowsbuf3(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows1p[dx] = S1p[-1] * a0 + S1p[0] * a1 + S1p[1] * a2 + S1p[2] * a3;
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const float* S0 = src.row(sy - 1);
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows0p[dx] = S0p[-1] * a0 + S0p[0] * a1 + S0p[1] * a2 + S0p[2] * a3;
-                rows1p[dx] = S1p[-1] * a0 + S1p[0] * a1 + S1p[1] * a2 + S1p[2] * a3;
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-        float b2 = beta[2];
-        float b3 = beta[3];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        float* Dp = dst.row(dy);
-        for (int dx = 0; dx < w; dx++)
-        {
-            //             D[x] = rows0[x]*b0 + rows1[x]*b1 + rows2[x]*b2 + rows3[x]*b3;
-            *Dp++ = *rows0p++ * b0 + *rows1p++ * b1 + *rows2p++ * b2 + *rows3p++ * b3;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_bf16s.h
deleted file mode 100644
index bd7d1f49..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_bf16s.h
+++ /dev/null
@@ -1,188 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bicubic_image_bf16s(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    Mat rowsbuf2(w);
-    Mat rowsbuf3(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const unsigned short* S3 = src.row<const unsigned short>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const unsigned short* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows3p[dx] = bfloat16_to_float32(S3p[-1]) * a0 + bfloat16_to_float32(S3p[0]) * a1 + bfloat16_to_float32(S3p[1]) * a2 + bfloat16_to_float32(S3p[2]) * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const unsigned short* S2 = src.row<const unsigned short>(sy + 1);
-            const unsigned short* S3 = src.row<const unsigned short>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const unsigned short* S2p = S2 + sx;
-                const unsigned short* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows2p[dx] = bfloat16_to_float32(S2p[-1]) * a0 + bfloat16_to_float32(S2p[0]) * a1 + bfloat16_to_float32(S2p[1]) * a2 + bfloat16_to_float32(S2p[2]) * a3;
-                rows3p[dx] = bfloat16_to_float32(S3p[-1]) * a0 + bfloat16_to_float32(S3p[0]) * a1 + bfloat16_to_float32(S3p[1]) * a2 + bfloat16_to_float32(S3p[2]) * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const unsigned short* S1 = src.row<const unsigned short>(sy);
-            const unsigned short* S2 = src.row<const unsigned short>(sy + 1);
-            const unsigned short* S3 = src.row<const unsigned short>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const unsigned short* S1p = S1 + sx;
-                const unsigned short* S2p = S2 + sx;
-                const unsigned short* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows1p[dx] = bfloat16_to_float32(S1p[-1]) * a0 + bfloat16_to_float32(S1p[0]) * a1 + bfloat16_to_float32(S1p[1]) * a2 + bfloat16_to_float32(S1p[2]) * a3;
-                rows2p[dx] = bfloat16_to_float32(S2p[-1]) * a0 + bfloat16_to_float32(S2p[0]) * a1 + bfloat16_to_float32(S2p[1]) * a2 + bfloat16_to_float32(S2p[2]) * a3;
-                rows3p[dx] = bfloat16_to_float32(S3p[-1]) * a0 + bfloat16_to_float32(S3p[0]) * a1 + bfloat16_to_float32(S3p[1]) * a2 + bfloat16_to_float32(S3p[2]) * a3;
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const unsigned short* S0 = src.row<const unsigned short>(sy - 1);
-            const unsigned short* S1 = src.row<const unsigned short>(sy);
-            const unsigned short* S2 = src.row<const unsigned short>(sy + 1);
-            const unsigned short* S3 = src.row<const unsigned short>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const unsigned short* S0p = S0 + sx;
-                const unsigned short* S1p = S1 + sx;
-                const unsigned short* S2p = S2 + sx;
-                const unsigned short* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows0p[dx] = bfloat16_to_float32(S0p[-1]) * a0 + bfloat16_to_float32(S0p[0]) * a1 + bfloat16_to_float32(S0p[1]) * a2 + bfloat16_to_float32(S0p[2]) * a3;
-                rows1p[dx] = bfloat16_to_float32(S1p[-1]) * a0 + bfloat16_to_float32(S1p[0]) * a1 + bfloat16_to_float32(S1p[1]) * a2 + bfloat16_to_float32(S1p[2]) * a3;
-                rows2p[dx] = bfloat16_to_float32(S2p[-1]) * a0 + bfloat16_to_float32(S2p[0]) * a1 + bfloat16_to_float32(S2p[1]) * a2 + bfloat16_to_float32(S2p[2]) * a3;
-                rows3p[dx] = bfloat16_to_float32(S3p[-1]) * a0 + bfloat16_to_float32(S3p[0]) * a1 + bfloat16_to_float32(S3p[1]) * a2 + bfloat16_to_float32(S3p[2]) * a3;
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-        float b2 = beta[2];
-        float b3 = beta[3];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        unsigned short* Dp = dst.row<unsigned short>(dy);
-        for (int dx = 0; dx < w; dx++)
-        {
-            //             D[x] = rows0[x]*b0 + rows1[x]*b1 + rows2[x]*b2 + rows3[x]*b3;
-            *Dp++ = float32_to_bfloat16(*rows0p++ * b0 + *rows1p++ * b1 + *rows2p++ * b2 + *rows3p++ * b3);
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_fp16s.h
deleted file mode 100644
index 18b2e567..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_fp16s.h
+++ /dev/null
@@ -1,436 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static inline void interpolate_cubic_fp16sa(float fx, __fp16* coeffs)
-{
-    const float A = -0.75f;
-
-    float fx0 = fx + 1;
-    float fx1 = fx;
-    float fx2 = 1 - fx;
-    // float fx3 = 2 - fx;
-
-    coeffs[0] = (__fp16)(A * fx0 * fx0 * fx0 - 5 * A * fx0 * fx0 + 8 * A * fx0 - 4 * A);
-    coeffs[1] = (__fp16)((A + 2) * fx1 * fx1 * fx1 - (A + 3) * fx1 * fx1 + 1);
-    coeffs[2] = (__fp16)((A + 2) * fx2 * fx2 * fx2 - (A + 3) * fx2 * fx2 + 1);
-    coeffs[3] = (__fp16)((__fp16)1.f - coeffs[0] - coeffs[1] - coeffs[2]);
-}
-
-static void cubic_coeffs_fp16sa(int w, int outw, int* xofs, __fp16* alpha, int align_corner)
-{
-    double scale = (double)w / outw;
-    if (align_corner)
-    {
-        scale = (double)(w - 1) / (outw - 1);
-    }
-
-    for (int dx = 0; dx < outw; dx++)
-    {
-        float fx = (float)((dx + 0.5) * scale - 0.5);
-        if (align_corner)
-        {
-            fx = static_cast<float>(dx * scale);
-        }
-
-        int sx = static_cast<int>(floor(fx));
-        fx -= sx;
-
-        interpolate_cubic_fp16sa(fx, alpha + dx * 4);
-
-        if (sx <= -1)
-        {
-            sx = 1;
-            alpha[dx * 4 + 0] = (__fp16)((__fp16)1.f - alpha[dx * 4 + 3]);
-            alpha[dx * 4 + 1] = (__fp16)alpha[dx * 4 + 3];
-            alpha[dx * 4 + 2] = (__fp16)0.f;
-            alpha[dx * 4 + 3] = (__fp16)0.f;
-        }
-        if (sx == 0)
-        {
-            sx = 1;
-            alpha[dx * 4 + 0] = (__fp16)(alpha[dx * 4 + 0] + alpha[dx * 4 + 1]);
-            alpha[dx * 4 + 1] = (__fp16)alpha[dx * 4 + 2];
-            alpha[dx * 4 + 2] = (__fp16)alpha[dx * 4 + 3];
-            alpha[dx * 4 + 3] = (__fp16)0.f;
-        }
-        if (sx == w - 2)
-        {
-            sx = w - 3;
-            alpha[dx * 4 + 3] = (__fp16)(alpha[dx * 4 + 2] + alpha[dx * 4 + 3]);
-            alpha[dx * 4 + 2] = (__fp16)alpha[dx * 4 + 1];
-            alpha[dx * 4 + 1] = (__fp16)alpha[dx * 4 + 0];
-            alpha[dx * 4 + 0] = (__fp16)0.f;
-        }
-        if (sx >= w - 1)
-        {
-            sx = w - 3;
-            alpha[dx * 4 + 3] = (__fp16)((__fp16)1.f - alpha[dx * 4 + 0]);
-            alpha[dx * 4 + 2] = (__fp16)(alpha[dx * 4 + 0]);
-            alpha[dx * 4 + 1] = (__fp16)0.f;
-            alpha[dx * 4 + 0] = (__fp16)0.f;
-        }
-
-        xofs[dx] = sx;
-    }
-}
-
-static void resize_bicubic_image_fp16s(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    Mat rowsbuf2(w);
-    Mat rowsbuf3(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows3p[dx] = (float)S3p[-1] * a0 + (float)S3p[0] * a1 + (float)S3p[1] * a2 + (float)S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows2p[dx] = (float)S2p[-1] * a0 + (float)S2p[0] * a1 + (float)S2p[1] * a2 + (float)S2p[2] * a3;
-                rows3p[dx] = (float)S3p[-1] * a0 + (float)S3p[0] * a1 + (float)S3p[1] * a2 + (float)S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows1p[dx] = (float)S1p[-1] * a0 + (float)S1p[0] * a1 + (float)S1p[1] * a2 + (float)S1p[2] * a3;
-                rows2p[dx] = (float)S2p[-1] * a0 + (float)S2p[0] * a1 + (float)S2p[1] * a2 + (float)S2p[2] * a3;
-                rows3p[dx] = (float)S3p[-1] * a0 + (float)S3p[0] * a1 + (float)S3p[1] * a2 + (float)S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const __fp16* S0 = src.row<const __fp16>(sy - 1);
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows0p[dx] = (float)S0p[-1] * a0 + (float)S0p[0] * a1 + (float)S0p[1] * a2 + (float)S0p[2] * a3;
-                rows1p[dx] = (float)S1p[-1] * a0 + (float)S1p[0] * a1 + (float)S1p[1] * a2 + (float)S1p[2] * a3;
-                rows2p[dx] = (float)S2p[-1] * a0 + (float)S2p[0] * a1 + (float)S2p[1] * a2 + (float)S2p[2] * a3;
-                rows3p[dx] = (float)S3p[-1] * a0 + (float)S3p[0] * a1 + (float)S3p[1] * a2 + (float)S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-        float b2 = beta[2];
-        float b3 = beta[3];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        __fp16* Dp = dst.row<__fp16>(dy);
-        for (int dx = 0; dx < w; dx++)
-        {
-            // D[x] = rows0[x]*b0 + rows1[x]*b1 + rows2[x]*b2 + rows3[x]*b3;
-            *Dp++ = (__fp16)(*rows0p++ * b0 + *rows1p++ * b1 + *rows2p++ * b2 + *rows3p++ * b3);
-        }
-
-        beta += 4;
-    }
-}
-
-static void resize_bicubic_image_fp16sa(const Mat& src, Mat& dst, __fp16* alpha, int* xofs, __fp16* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)2u);
-    Mat rowsbuf1(w, (size_t)2u);
-    Mat rowsbuf2(w, (size_t)2u);
-    Mat rowsbuf3(w, (size_t)2u);
-    __fp16* rows0 = rowsbuf0;
-    __fp16* rows1 = rowsbuf1;
-    __fp16* rows2 = rowsbuf2;
-    __fp16* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            __fp16* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S3p = S3 + sx;
-
-                __fp16 a0 = alphap[0];
-                __fp16 a1 = alphap[1];
-                __fp16 a2 = alphap[2];
-                __fp16 a3 = alphap[3];
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            __fp16* rows0_old = rows0;
-            __fp16* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows2p = rows2;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                __fp16 a0 = alphap[0];
-                __fp16 a1 = alphap[1];
-                __fp16 a2 = alphap[2];
-                __fp16 a3 = alphap[3];
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            __fp16* rows0_old = rows0;
-            __fp16* rows1_old = rows1;
-            __fp16* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows1p = rows1;
-            __fp16* rows2p = rows2;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                __fp16 a0 = alphap[0];
-                __fp16 a1 = alphap[1];
-                __fp16 a2 = alphap[2];
-                __fp16 a3 = alphap[3];
-                rows1p[dx] = S1p[-1] * a0 + S1p[0] * a1 + S1p[1] * a2 + S1p[2] * a3;
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const __fp16* S0 = src.row<const __fp16>(sy - 1);
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows0p = rows0;
-            __fp16* rows1p = rows1;
-            __fp16* rows2p = rows2;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                __fp16 a0 = alphap[0];
-                __fp16 a1 = alphap[1];
-                __fp16 a2 = alphap[2];
-                __fp16 a3 = alphap[3];
-                rows0p[dx] = S0p[-1] * a0 + S0p[0] * a1 + S0p[1] * a2 + S0p[2] * a3;
-                rows1p[dx] = S1p[-1] * a0 + S1p[0] * a1 + S1p[1] * a2 + S1p[2] * a3;
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        __fp16 b0 = beta[0];
-        __fp16 b1 = beta[1];
-        __fp16 b2 = beta[2];
-        __fp16 b3 = beta[3];
-
-        __fp16* rows0p = rows0;
-        __fp16* rows1p = rows1;
-        __fp16* rows2p = rows2;
-        __fp16* rows3p = rows3;
-        __fp16* Dp = dst.row<__fp16>(dy);
-        for (int dx = 0; dx < w; dx++)
-        {
-            // D[x] = rows0[x]*b0 + rows1[x]*b1 + rows2[x]*b2 + rows3[x]*b3;
-            *Dp++ = (*rows0p++ * b0 + *rows1p++ * b1 + *rows2p++ * b2 + *rows3p++ * b3);
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack4.h
deleted file mode 100644
index ab0d5612..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack4.h
+++ /dev/null
@@ -1,272 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bicubic_image_pack4(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf1(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf2(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf3(w, (size_t)4 * 4u, 4);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S3p = S3 + sx;
-
-                float32x4_t _a0123 = vld1q_f32(alphap);
-
-                float32x4_t _S30 = vld1q_f32(S3p - 4);
-                float32x4_t _S31 = vld1q_f32(S3p + 0);
-                float32x4_t _S32 = vld1q_f32(S3p + 4);
-                float32x4_t _S33 = vld1q_f32(S3p + 8);
-                float32x4_t _rows3 = vmulq_lane_f32(_S30, vget_low_f32(_a0123), 0);
-                _rows3 = vmlaq_lane_f32(_rows3, _S31, vget_low_f32(_a0123), 1);
-                _rows3 = vmlaq_lane_f32(_rows3, _S32, vget_high_f32(_a0123), 0);
-                _rows3 = vmlaq_lane_f32(_rows3, _S33, vget_high_f32(_a0123), 1);
-                vst1q_f32(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float32x4_t _a0123 = vld1q_f32(alphap);
-
-                float32x4_t _S20 = vld1q_f32(S2p - 4);
-                float32x4_t _S21 = vld1q_f32(S2p + 0);
-                float32x4_t _S22 = vld1q_f32(S2p + 4);
-                float32x4_t _S23 = vld1q_f32(S2p + 8);
-                float32x4_t _S30 = vld1q_f32(S3p - 4);
-                float32x4_t _S31 = vld1q_f32(S3p + 0);
-                float32x4_t _S32 = vld1q_f32(S3p + 4);
-                float32x4_t _S33 = vld1q_f32(S3p + 8);
-                float32x4_t _rows2 = vmulq_lane_f32(_S20, vget_low_f32(_a0123), 0);
-                float32x4_t _rows3 = vmulq_lane_f32(_S30, vget_low_f32(_a0123), 0);
-                _rows2 = vmlaq_lane_f32(_rows2, _S21, vget_low_f32(_a0123), 1);
-                _rows3 = vmlaq_lane_f32(_rows3, _S31, vget_low_f32(_a0123), 1);
-                _rows2 = vmlaq_lane_f32(_rows2, _S22, vget_high_f32(_a0123), 0);
-                _rows3 = vmlaq_lane_f32(_rows3, _S32, vget_high_f32(_a0123), 0);
-                _rows2 = vmlaq_lane_f32(_rows2, _S23, vget_high_f32(_a0123), 1);
-                _rows3 = vmlaq_lane_f32(_rows3, _S33, vget_high_f32(_a0123), 1);
-                vst1q_f32(rows2p + dx * 4, _rows2);
-                vst1q_f32(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float32x4_t _a0123 = vld1q_f32(alphap);
-
-                float32x4_t _S10 = vld1q_f32(S1p - 4);
-                float32x4_t _S11 = vld1q_f32(S1p + 0);
-                float32x4_t _S12 = vld1q_f32(S1p + 4);
-                float32x4_t _S13 = vld1q_f32(S1p + 8);
-                float32x4_t _S20 = vld1q_f32(S2p - 4);
-                float32x4_t _S21 = vld1q_f32(S2p + 0);
-                float32x4_t _S22 = vld1q_f32(S2p + 4);
-                float32x4_t _S23 = vld1q_f32(S2p + 8);
-                float32x4_t _S30 = vld1q_f32(S3p - 4);
-                float32x4_t _S31 = vld1q_f32(S3p + 0);
-                float32x4_t _S32 = vld1q_f32(S3p + 4);
-                float32x4_t _S33 = vld1q_f32(S3p + 8);
-                float32x4_t _rows1 = vmulq_lane_f32(_S10, vget_low_f32(_a0123), 0);
-                float32x4_t _rows2 = vmulq_lane_f32(_S20, vget_low_f32(_a0123), 0);
-                float32x4_t _rows3 = vmulq_lane_f32(_S30, vget_low_f32(_a0123), 0);
-                _rows1 = vmlaq_lane_f32(_rows1, _S11, vget_low_f32(_a0123), 1);
-                _rows2 = vmlaq_lane_f32(_rows2, _S21, vget_low_f32(_a0123), 1);
-                _rows3 = vmlaq_lane_f32(_rows3, _S31, vget_low_f32(_a0123), 1);
-                _rows1 = vmlaq_lane_f32(_rows1, _S12, vget_high_f32(_a0123), 0);
-                _rows2 = vmlaq_lane_f32(_rows2, _S22, vget_high_f32(_a0123), 0);
-                _rows3 = vmlaq_lane_f32(_rows3, _S32, vget_high_f32(_a0123), 0);
-                _rows1 = vmlaq_lane_f32(_rows1, _S13, vget_high_f32(_a0123), 1);
-                _rows2 = vmlaq_lane_f32(_rows2, _S23, vget_high_f32(_a0123), 1);
-                _rows3 = vmlaq_lane_f32(_rows3, _S33, vget_high_f32(_a0123), 1);
-                vst1q_f32(rows1p + dx * 4, _rows1);
-                vst1q_f32(rows2p + dx * 4, _rows2);
-                vst1q_f32(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const float* S0 = src.row(sy - 1);
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float32x4_t _a0123 = vld1q_f32(alphap);
-
-                // TODO check the generated assembly on armv7
-                float32x4_t _S00 = vld1q_f32(S0p - 4);
-                float32x4_t _S01 = vld1q_f32(S0p + 0);
-                float32x4_t _S02 = vld1q_f32(S0p + 4);
-                float32x4_t _S03 = vld1q_f32(S0p + 8);
-                float32x4_t _S10 = vld1q_f32(S1p - 4);
-                float32x4_t _S11 = vld1q_f32(S1p + 0);
-                float32x4_t _S12 = vld1q_f32(S1p + 4);
-                float32x4_t _S13 = vld1q_f32(S1p + 8);
-                float32x4_t _S20 = vld1q_f32(S2p - 4);
-                float32x4_t _S21 = vld1q_f32(S2p + 0);
-                float32x4_t _S22 = vld1q_f32(S2p + 4);
-                float32x4_t _S23 = vld1q_f32(S2p + 8);
-                float32x4_t _S30 = vld1q_f32(S3p - 4);
-                float32x4_t _S31 = vld1q_f32(S3p + 0);
-                float32x4_t _S32 = vld1q_f32(S3p + 4);
-                float32x4_t _S33 = vld1q_f32(S3p + 8);
-                float32x4_t _rows0 = vmulq_lane_f32(_S00, vget_low_f32(_a0123), 0);
-                float32x4_t _rows1 = vmulq_lane_f32(_S10, vget_low_f32(_a0123), 0);
-                float32x4_t _rows2 = vmulq_lane_f32(_S20, vget_low_f32(_a0123), 0);
-                float32x4_t _rows3 = vmulq_lane_f32(_S30, vget_low_f32(_a0123), 0);
-                _rows0 = vmlaq_lane_f32(_rows0, _S01, vget_low_f32(_a0123), 1);
-                _rows1 = vmlaq_lane_f32(_rows1, _S11, vget_low_f32(_a0123), 1);
-                _rows2 = vmlaq_lane_f32(_rows2, _S21, vget_low_f32(_a0123), 1);
-                _rows3 = vmlaq_lane_f32(_rows3, _S31, vget_low_f32(_a0123), 1);
-                _rows0 = vmlaq_lane_f32(_rows0, _S02, vget_high_f32(_a0123), 0);
-                _rows1 = vmlaq_lane_f32(_rows1, _S12, vget_high_f32(_a0123), 0);
-                _rows2 = vmlaq_lane_f32(_rows2, _S22, vget_high_f32(_a0123), 0);
-                _rows3 = vmlaq_lane_f32(_rows3, _S32, vget_high_f32(_a0123), 0);
-                _rows0 = vmlaq_lane_f32(_rows0, _S03, vget_high_f32(_a0123), 1);
-                _rows1 = vmlaq_lane_f32(_rows1, _S13, vget_high_f32(_a0123), 1);
-                _rows2 = vmlaq_lane_f32(_rows2, _S23, vget_high_f32(_a0123), 1);
-                _rows3 = vmlaq_lane_f32(_rows3, _S33, vget_high_f32(_a0123), 1);
-                vst1q_f32(rows0p + dx * 4, _rows0);
-                vst1q_f32(rows1p + dx * 4, _rows1);
-                vst1q_f32(rows2p + dx * 4, _rows2);
-                vst1q_f32(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float32x4_t _b0123 = vld1q_f32(beta);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        float* Dp = dst.row(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            float32x4_t _rows0 = vld1q_f32(rows0p);
-            float32x4_t _rows1 = vld1q_f32(rows1p);
-            float32x4_t _rows2 = vld1q_f32(rows2p);
-            float32x4_t _rows3 = vld1q_f32(rows3p);
-            float32x4_t _Dp = vmulq_lane_f32(_rows0, vget_low_f32(_b0123), 0);
-            _Dp = vmlaq_lane_f32(_Dp, _rows1, vget_low_f32(_b0123), 1);
-            _Dp = vmlaq_lane_f32(_Dp, _rows2, vget_high_f32(_b0123), 0);
-            _Dp = vmlaq_lane_f32(_Dp, _rows3, vget_high_f32(_b0123), 1);
-            vst1q_f32(Dp, _Dp);
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-            rows2p += 4;
-            rows3p += 4;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack4_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack4_bf16s.h
deleted file mode 100644
index 6546accf..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack4_bf16s.h
+++ /dev/null
@@ -1,272 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bicubic_image_pack4_bf16s(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf1(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf2(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf3(w, (size_t)4 * 4u, 4);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const unsigned short* S3 = src.row<const unsigned short>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const unsigned short* S3p = S3 + sx;
-
-                float32x4_t _a0123 = vld1q_f32(alphap);
-
-                float32x4_t _S30 = bfloat2float(vld1_u16(S3p - 4));
-                float32x4_t _S31 = bfloat2float(vld1_u16(S3p + 0));
-                float32x4_t _S32 = bfloat2float(vld1_u16(S3p + 4));
-                float32x4_t _S33 = bfloat2float(vld1_u16(S3p + 8));
-                float32x4_t _rows3 = vmulq_lane_f32(_S30, vget_low_f32(_a0123), 0);
-                _rows3 = vmlaq_lane_f32(_rows3, _S31, vget_low_f32(_a0123), 1);
-                _rows3 = vmlaq_lane_f32(_rows3, _S32, vget_high_f32(_a0123), 0);
-                _rows3 = vmlaq_lane_f32(_rows3, _S33, vget_high_f32(_a0123), 1);
-                vst1q_f32(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const unsigned short* S2 = src.row<const unsigned short>(sy + 1);
-            const unsigned short* S3 = src.row<const unsigned short>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const unsigned short* S2p = S2 + sx;
-                const unsigned short* S3p = S3 + sx;
-
-                float32x4_t _a0123 = vld1q_f32(alphap);
-
-                float32x4_t _S20 = bfloat2float(vld1_u16(S2p - 4));
-                float32x4_t _S21 = bfloat2float(vld1_u16(S2p + 0));
-                float32x4_t _S22 = bfloat2float(vld1_u16(S2p + 4));
-                float32x4_t _S23 = bfloat2float(vld1_u16(S2p + 8));
-                float32x4_t _S30 = bfloat2float(vld1_u16(S3p - 4));
-                float32x4_t _S31 = bfloat2float(vld1_u16(S3p + 0));
-                float32x4_t _S32 = bfloat2float(vld1_u16(S3p + 4));
-                float32x4_t _S33 = bfloat2float(vld1_u16(S3p + 8));
-                float32x4_t _rows2 = vmulq_lane_f32(_S20, vget_low_f32(_a0123), 0);
-                float32x4_t _rows3 = vmulq_lane_f32(_S30, vget_low_f32(_a0123), 0);
-                _rows2 = vmlaq_lane_f32(_rows2, _S21, vget_low_f32(_a0123), 1);
-                _rows3 = vmlaq_lane_f32(_rows3, _S31, vget_low_f32(_a0123), 1);
-                _rows2 = vmlaq_lane_f32(_rows2, _S22, vget_high_f32(_a0123), 0);
-                _rows3 = vmlaq_lane_f32(_rows3, _S32, vget_high_f32(_a0123), 0);
-                _rows2 = vmlaq_lane_f32(_rows2, _S23, vget_high_f32(_a0123), 1);
-                _rows3 = vmlaq_lane_f32(_rows3, _S33, vget_high_f32(_a0123), 1);
-                vst1q_f32(rows2p + dx * 4, _rows2);
-                vst1q_f32(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const unsigned short* S1 = src.row<const unsigned short>(sy);
-            const unsigned short* S2 = src.row<const unsigned short>(sy + 1);
-            const unsigned short* S3 = src.row<const unsigned short>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const unsigned short* S1p = S1 + sx;
-                const unsigned short* S2p = S2 + sx;
-                const unsigned short* S3p = S3 + sx;
-
-                float32x4_t _a0123 = vld1q_f32(alphap);
-
-                float32x4_t _S10 = bfloat2float(vld1_u16(S1p - 4));
-                float32x4_t _S11 = bfloat2float(vld1_u16(S1p + 0));
-                float32x4_t _S12 = bfloat2float(vld1_u16(S1p + 4));
-                float32x4_t _S13 = bfloat2float(vld1_u16(S1p + 8));
-                float32x4_t _S20 = bfloat2float(vld1_u16(S2p - 4));
-                float32x4_t _S21 = bfloat2float(vld1_u16(S2p + 0));
-                float32x4_t _S22 = bfloat2float(vld1_u16(S2p + 4));
-                float32x4_t _S23 = bfloat2float(vld1_u16(S2p + 8));
-                float32x4_t _S30 = bfloat2float(vld1_u16(S3p - 4));
-                float32x4_t _S31 = bfloat2float(vld1_u16(S3p + 0));
-                float32x4_t _S32 = bfloat2float(vld1_u16(S3p + 4));
-                float32x4_t _S33 = bfloat2float(vld1_u16(S3p + 8));
-                float32x4_t _rows1 = vmulq_lane_f32(_S10, vget_low_f32(_a0123), 0);
-                float32x4_t _rows2 = vmulq_lane_f32(_S20, vget_low_f32(_a0123), 0);
-                float32x4_t _rows3 = vmulq_lane_f32(_S30, vget_low_f32(_a0123), 0);
-                _rows1 = vmlaq_lane_f32(_rows1, _S11, vget_low_f32(_a0123), 1);
-                _rows2 = vmlaq_lane_f32(_rows2, _S21, vget_low_f32(_a0123), 1);
-                _rows3 = vmlaq_lane_f32(_rows3, _S31, vget_low_f32(_a0123), 1);
-                _rows1 = vmlaq_lane_f32(_rows1, _S12, vget_high_f32(_a0123), 0);
-                _rows2 = vmlaq_lane_f32(_rows2, _S22, vget_high_f32(_a0123), 0);
-                _rows3 = vmlaq_lane_f32(_rows3, _S32, vget_high_f32(_a0123), 0);
-                _rows1 = vmlaq_lane_f32(_rows1, _S13, vget_high_f32(_a0123), 1);
-                _rows2 = vmlaq_lane_f32(_rows2, _S23, vget_high_f32(_a0123), 1);
-                _rows3 = vmlaq_lane_f32(_rows3, _S33, vget_high_f32(_a0123), 1);
-                vst1q_f32(rows1p + dx * 4, _rows1);
-                vst1q_f32(rows2p + dx * 4, _rows2);
-                vst1q_f32(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const unsigned short* S0 = src.row<const unsigned short>(sy - 1);
-            const unsigned short* S1 = src.row<const unsigned short>(sy);
-            const unsigned short* S2 = src.row<const unsigned short>(sy + 1);
-            const unsigned short* S3 = src.row<const unsigned short>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const unsigned short* S0p = S0 + sx;
-                const unsigned short* S1p = S1 + sx;
-                const unsigned short* S2p = S2 + sx;
-                const unsigned short* S3p = S3 + sx;
-
-                float32x4_t _a0123 = vld1q_f32(alphap);
-
-                // TODO check the generated assembly on armv7
-                float32x4_t _S00 = bfloat2float(vld1_u16(S0p - 4));
-                float32x4_t _S01 = bfloat2float(vld1_u16(S0p + 0));
-                float32x4_t _S02 = bfloat2float(vld1_u16(S0p + 4));
-                float32x4_t _S03 = bfloat2float(vld1_u16(S0p + 8));
-                float32x4_t _S10 = bfloat2float(vld1_u16(S1p - 4));
-                float32x4_t _S11 = bfloat2float(vld1_u16(S1p + 0));
-                float32x4_t _S12 = bfloat2float(vld1_u16(S1p + 4));
-                float32x4_t _S13 = bfloat2float(vld1_u16(S1p + 8));
-                float32x4_t _S20 = bfloat2float(vld1_u16(S2p - 4));
-                float32x4_t _S21 = bfloat2float(vld1_u16(S2p + 0));
-                float32x4_t _S22 = bfloat2float(vld1_u16(S2p + 4));
-                float32x4_t _S23 = bfloat2float(vld1_u16(S2p + 8));
-                float32x4_t _S30 = bfloat2float(vld1_u16(S3p - 4));
-                float32x4_t _S31 = bfloat2float(vld1_u16(S3p + 0));
-                float32x4_t _S32 = bfloat2float(vld1_u16(S3p + 4));
-                float32x4_t _S33 = bfloat2float(vld1_u16(S3p + 8));
-                float32x4_t _rows0 = vmulq_lane_f32(_S00, vget_low_f32(_a0123), 0);
-                float32x4_t _rows1 = vmulq_lane_f32(_S10, vget_low_f32(_a0123), 0);
-                float32x4_t _rows2 = vmulq_lane_f32(_S20, vget_low_f32(_a0123), 0);
-                float32x4_t _rows3 = vmulq_lane_f32(_S30, vget_low_f32(_a0123), 0);
-                _rows0 = vmlaq_lane_f32(_rows0, _S01, vget_low_f32(_a0123), 1);
-                _rows1 = vmlaq_lane_f32(_rows1, _S11, vget_low_f32(_a0123), 1);
-                _rows2 = vmlaq_lane_f32(_rows2, _S21, vget_low_f32(_a0123), 1);
-                _rows3 = vmlaq_lane_f32(_rows3, _S31, vget_low_f32(_a0123), 1);
-                _rows0 = vmlaq_lane_f32(_rows0, _S02, vget_high_f32(_a0123), 0);
-                _rows1 = vmlaq_lane_f32(_rows1, _S12, vget_high_f32(_a0123), 0);
-                _rows2 = vmlaq_lane_f32(_rows2, _S22, vget_high_f32(_a0123), 0);
-                _rows3 = vmlaq_lane_f32(_rows3, _S32, vget_high_f32(_a0123), 0);
-                _rows0 = vmlaq_lane_f32(_rows0, _S03, vget_high_f32(_a0123), 1);
-                _rows1 = vmlaq_lane_f32(_rows1, _S13, vget_high_f32(_a0123), 1);
-                _rows2 = vmlaq_lane_f32(_rows2, _S23, vget_high_f32(_a0123), 1);
-                _rows3 = vmlaq_lane_f32(_rows3, _S33, vget_high_f32(_a0123), 1);
-                vst1q_f32(rows0p + dx * 4, _rows0);
-                vst1q_f32(rows1p + dx * 4, _rows1);
-                vst1q_f32(rows2p + dx * 4, _rows2);
-                vst1q_f32(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float32x4_t _b0123 = vld1q_f32(beta);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        unsigned short* Dp = dst.row<unsigned short>(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            float32x4_t _rows0 = vld1q_f32(rows0p);
-            float32x4_t _rows1 = vld1q_f32(rows1p);
-            float32x4_t _rows2 = vld1q_f32(rows2p);
-            float32x4_t _rows3 = vld1q_f32(rows3p);
-            float32x4_t _Dp = vmulq_lane_f32(_rows0, vget_low_f32(_b0123), 0);
-            _Dp = vmlaq_lane_f32(_Dp, _rows1, vget_low_f32(_b0123), 1);
-            _Dp = vmlaq_lane_f32(_Dp, _rows2, vget_high_f32(_b0123), 0);
-            _Dp = vmlaq_lane_f32(_Dp, _rows3, vget_high_f32(_b0123), 1);
-            vst1_u16(Dp, float2bfloat(_Dp));
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-            rows2p += 4;
-            rows3p += 4;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack4_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack4_fp16s.h
deleted file mode 100644
index 7f823148..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack4_fp16s.h
+++ /dev/null
@@ -1,529 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bicubic_image_pack4_fp16s(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf1(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf2(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf3(w, (size_t)4 * 4u, 4);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const __fp16* S3p = S3 + sx;
-
-                float32x4_t _a0123 = vld1q_f32(alphap);
-
-                float32x4_t _S30 = vcvt_f32_f16(vld1_f16(S3p - 4));
-                float32x4_t _S31 = vcvt_f32_f16(vld1_f16(S3p + 0));
-                float32x4_t _S32 = vcvt_f32_f16(vld1_f16(S3p + 4));
-                float32x4_t _S33 = vcvt_f32_f16(vld1_f16(S3p + 8));
-                float32x4_t _rows3 = vmulq_laneq_f32(_S30, _a0123, 0);
-                _rows3 = vfmaq_laneq_f32(_rows3, _S31, _a0123, 1);
-                _rows3 = vfmaq_laneq_f32(_rows3, _S32, _a0123, 2);
-                _rows3 = vfmaq_laneq_f32(_rows3, _S33, _a0123, 3);
-                vst1q_f32(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                float32x4_t _a0123 = vld1q_f32(alphap);
-
-                float32x4_t _S20 = vcvt_f32_f16(vld1_f16(S2p - 4));
-                float32x4_t _S21 = vcvt_f32_f16(vld1_f16(S2p + 0));
-                float32x4_t _S22 = vcvt_f32_f16(vld1_f16(S2p + 4));
-                float32x4_t _S23 = vcvt_f32_f16(vld1_f16(S2p + 8));
-                float32x4_t _S30 = vcvt_f32_f16(vld1_f16(S3p - 4));
-                float32x4_t _S31 = vcvt_f32_f16(vld1_f16(S3p + 0));
-                float32x4_t _S32 = vcvt_f32_f16(vld1_f16(S3p + 4));
-                float32x4_t _S33 = vcvt_f32_f16(vld1_f16(S3p + 8));
-                float32x4_t _rows2 = vmulq_laneq_f32(_S20, _a0123, 0);
-                float32x4_t _rows3 = vmulq_laneq_f32(_S30, _a0123, 0);
-                _rows2 = vfmaq_laneq_f32(_rows2, _S21, _a0123, 1);
-                _rows3 = vfmaq_laneq_f32(_rows3, _S31, _a0123, 1);
-                _rows2 = vfmaq_laneq_f32(_rows2, _S22, _a0123, 2);
-                _rows3 = vfmaq_laneq_f32(_rows3, _S32, _a0123, 2);
-                _rows2 = vfmaq_laneq_f32(_rows2, _S23, _a0123, 3);
-                _rows3 = vfmaq_laneq_f32(_rows3, _S33, _a0123, 3);
-                vst1q_f32(rows2p + dx * 4, _rows2);
-                vst1q_f32(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                float32x4_t _a0123 = vld1q_f32(alphap);
-
-                float32x4_t _S10 = vcvt_f32_f16(vld1_f16(S1p - 4));
-                float32x4_t _S11 = vcvt_f32_f16(vld1_f16(S1p + 0));
-                float32x4_t _S12 = vcvt_f32_f16(vld1_f16(S1p + 4));
-                float32x4_t _S13 = vcvt_f32_f16(vld1_f16(S1p + 8));
-                float32x4_t _S20 = vcvt_f32_f16(vld1_f16(S2p - 4));
-                float32x4_t _S21 = vcvt_f32_f16(vld1_f16(S2p + 0));
-                float32x4_t _S22 = vcvt_f32_f16(vld1_f16(S2p + 4));
-                float32x4_t _S23 = vcvt_f32_f16(vld1_f16(S2p + 8));
-                float32x4_t _S30 = vcvt_f32_f16(vld1_f16(S3p - 4));
-                float32x4_t _S31 = vcvt_f32_f16(vld1_f16(S3p + 0));
-                float32x4_t _S32 = vcvt_f32_f16(vld1_f16(S3p + 4));
-                float32x4_t _S33 = vcvt_f32_f16(vld1_f16(S3p + 8));
-                float32x4_t _rows1 = vmulq_laneq_f32(_S10, _a0123, 0);
-                float32x4_t _rows2 = vmulq_laneq_f32(_S20, _a0123, 0);
-                float32x4_t _rows3 = vmulq_laneq_f32(_S30, _a0123, 0);
-                _rows1 = vfmaq_laneq_f32(_rows1, _S11, _a0123, 1);
-                _rows2 = vfmaq_laneq_f32(_rows2, _S21, _a0123, 1);
-                _rows3 = vfmaq_laneq_f32(_rows3, _S31, _a0123, 1);
-                _rows1 = vfmaq_laneq_f32(_rows1, _S12, _a0123, 2);
-                _rows2 = vfmaq_laneq_f32(_rows2, _S22, _a0123, 2);
-                _rows3 = vfmaq_laneq_f32(_rows3, _S32, _a0123, 2);
-                _rows1 = vfmaq_laneq_f32(_rows1, _S13, _a0123, 3);
-                _rows2 = vfmaq_laneq_f32(_rows2, _S23, _a0123, 3);
-                _rows3 = vfmaq_laneq_f32(_rows3, _S33, _a0123, 3);
-                vst1q_f32(rows1p + dx * 4, _rows1);
-                vst1q_f32(rows2p + dx * 4, _rows2);
-                vst1q_f32(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const __fp16* S0 = src.row<const __fp16>(sy - 1);
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                float32x4_t _a0123 = vld1q_f32(alphap);
-
-                float32x4_t _S00 = vcvt_f32_f16(vld1_f16(S0p - 4));
-                float32x4_t _S01 = vcvt_f32_f16(vld1_f16(S0p + 0));
-                float32x4_t _S02 = vcvt_f32_f16(vld1_f16(S0p + 4));
-                float32x4_t _S03 = vcvt_f32_f16(vld1_f16(S0p + 8));
-                float32x4_t _S10 = vcvt_f32_f16(vld1_f16(S1p - 4));
-                float32x4_t _S11 = vcvt_f32_f16(vld1_f16(S1p + 0));
-                float32x4_t _S12 = vcvt_f32_f16(vld1_f16(S1p + 4));
-                float32x4_t _S13 = vcvt_f32_f16(vld1_f16(S1p + 8));
-                float32x4_t _S20 = vcvt_f32_f16(vld1_f16(S2p - 4));
-                float32x4_t _S21 = vcvt_f32_f16(vld1_f16(S2p + 0));
-                float32x4_t _S22 = vcvt_f32_f16(vld1_f16(S2p + 4));
-                float32x4_t _S23 = vcvt_f32_f16(vld1_f16(S2p + 8));
-                float32x4_t _S30 = vcvt_f32_f16(vld1_f16(S3p - 4));
-                float32x4_t _S31 = vcvt_f32_f16(vld1_f16(S3p + 0));
-                float32x4_t _S32 = vcvt_f32_f16(vld1_f16(S3p + 4));
-                float32x4_t _S33 = vcvt_f32_f16(vld1_f16(S3p + 8));
-                float32x4_t _rows0 = vmulq_laneq_f32(_S00, _a0123, 0);
-                float32x4_t _rows1 = vmulq_laneq_f32(_S10, _a0123, 0);
-                float32x4_t _rows2 = vmulq_laneq_f32(_S20, _a0123, 0);
-                float32x4_t _rows3 = vmulq_laneq_f32(_S30, _a0123, 0);
-                _rows0 = vfmaq_laneq_f32(_rows0, _S01, _a0123, 1);
-                _rows1 = vfmaq_laneq_f32(_rows1, _S11, _a0123, 1);
-                _rows2 = vfmaq_laneq_f32(_rows2, _S21, _a0123, 1);
-                _rows3 = vfmaq_laneq_f32(_rows3, _S31, _a0123, 1);
-                _rows0 = vfmaq_laneq_f32(_rows0, _S02, _a0123, 2);
-                _rows1 = vfmaq_laneq_f32(_rows1, _S12, _a0123, 2);
-                _rows2 = vfmaq_laneq_f32(_rows2, _S22, _a0123, 2);
-                _rows3 = vfmaq_laneq_f32(_rows3, _S32, _a0123, 2);
-                _rows0 = vfmaq_laneq_f32(_rows0, _S03, _a0123, 3);
-                _rows1 = vfmaq_laneq_f32(_rows1, _S13, _a0123, 3);
-                _rows2 = vfmaq_laneq_f32(_rows2, _S23, _a0123, 3);
-                _rows3 = vfmaq_laneq_f32(_rows3, _S33, _a0123, 3);
-                vst1q_f32(rows0p + dx * 4, _rows0);
-                vst1q_f32(rows1p + dx * 4, _rows1);
-                vst1q_f32(rows2p + dx * 4, _rows2);
-                vst1q_f32(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float32x4_t _b0123 = vld1q_f32(beta);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        __fp16* Dp = dst.row<__fp16>(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            float32x4_t _rows0 = vld1q_f32(rows0p);
-            float32x4_t _rows1 = vld1q_f32(rows1p);
-            float32x4_t _rows2 = vld1q_f32(rows2p);
-            float32x4_t _rows3 = vld1q_f32(rows3p);
-            float32x4_t _Dp = vmulq_laneq_f32(_rows0, _b0123, 0);
-            _Dp = vfmaq_laneq_f32(_Dp, _rows1, _b0123, 1);
-            _Dp = vfmaq_laneq_f32(_Dp, _rows2, _b0123, 2);
-            _Dp = vfmaq_laneq_f32(_Dp, _rows3, _b0123, 3);
-            vst1_f16(Dp, vcvt_f16_f32(_Dp));
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-            rows2p += 4;
-            rows3p += 4;
-        }
-
-        beta += 4;
-    }
-}
-
-static void resize_bicubic_image_pack4_fp16sa(const Mat& src, Mat& dst, __fp16* alpha, int* xofs, __fp16* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)4 * 2u, 4);
-    Mat rowsbuf1(w, (size_t)4 * 2u, 4);
-    Mat rowsbuf2(w, (size_t)4 * 2u, 4);
-    Mat rowsbuf3(w, (size_t)4 * 2u, 4);
-    __fp16* rows0 = rowsbuf0;
-    __fp16* rows1 = rowsbuf1;
-    __fp16* rows2 = rowsbuf2;
-    __fp16* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            __fp16* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const __fp16* S3p = S3 + sx;
-
-                float16x4_t _a0123 = vld1_f16(alphap);
-
-                float16x4_t _S30 = vld1_f16(S3p - 4);
-                float16x4_t _S31 = vld1_f16(S3p + 0);
-                float16x4_t _S32 = vld1_f16(S3p + 4);
-                float16x4_t _S33 = vld1_f16(S3p + 8);
-                float16x4_t _rows3 = vmul_lane_f16(_S30, _a0123, 0);
-                _rows3 = vfma_lane_f16(_rows3, _S31, _a0123, 1);
-                _rows3 = vfma_lane_f16(_rows3, _S32, _a0123, 2);
-                _rows3 = vfma_lane_f16(_rows3, _S33, _a0123, 3);
-                vst1_f16(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            __fp16* rows0_old = rows0;
-            __fp16* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows2p = rows2;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                float16x4_t _a0123 = vld1_f16(alphap);
-
-                float16x4_t _S20 = vld1_f16(S2p - 4);
-                float16x4_t _S21 = vld1_f16(S2p + 0);
-                float16x4_t _S22 = vld1_f16(S2p + 4);
-                float16x4_t _S23 = vld1_f16(S2p + 8);
-                float16x4_t _S30 = vld1_f16(S3p - 4);
-                float16x4_t _S31 = vld1_f16(S3p + 0);
-                float16x4_t _S32 = vld1_f16(S3p + 4);
-                float16x4_t _S33 = vld1_f16(S3p + 8);
-                float16x4_t _rows2 = vmul_lane_f16(_S20, _a0123, 0);
-                float16x4_t _rows3 = vmul_lane_f16(_S30, _a0123, 0);
-                _rows2 = vfma_lane_f16(_rows2, _S21, _a0123, 1);
-                _rows3 = vfma_lane_f16(_rows3, _S31, _a0123, 1);
-                _rows2 = vfma_lane_f16(_rows2, _S22, _a0123, 2);
-                _rows3 = vfma_lane_f16(_rows3, _S32, _a0123, 2);
-                _rows2 = vfma_lane_f16(_rows2, _S23, _a0123, 3);
-                _rows3 = vfma_lane_f16(_rows3, _S33, _a0123, 3);
-                vst1_f16(rows2p + dx * 4, _rows2);
-                vst1_f16(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            __fp16* rows0_old = rows0;
-            __fp16* rows1_old = rows1;
-            __fp16* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows1p = rows1;
-            __fp16* rows2p = rows2;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                float16x4_t _a0123 = vld1_f16(alphap);
-
-                float16x4_t _S10 = vld1_f16(S1p - 4);
-                float16x4_t _S11 = vld1_f16(S1p + 0);
-                float16x4_t _S12 = vld1_f16(S1p + 4);
-                float16x4_t _S13 = vld1_f16(S1p + 8);
-                float16x4_t _S20 = vld1_f16(S2p - 4);
-                float16x4_t _S21 = vld1_f16(S2p + 0);
-                float16x4_t _S22 = vld1_f16(S2p + 4);
-                float16x4_t _S23 = vld1_f16(S2p + 8);
-                float16x4_t _S30 = vld1_f16(S3p - 4);
-                float16x4_t _S31 = vld1_f16(S3p + 0);
-                float16x4_t _S32 = vld1_f16(S3p + 4);
-                float16x4_t _S33 = vld1_f16(S3p + 8);
-                float16x4_t _rows1 = vmul_lane_f16(_S10, _a0123, 0);
-                float16x4_t _rows2 = vmul_lane_f16(_S20, _a0123, 0);
-                float16x4_t _rows3 = vmul_lane_f16(_S30, _a0123, 0);
-                _rows1 = vfma_lane_f16(_rows1, _S11, _a0123, 1);
-                _rows2 = vfma_lane_f16(_rows2, _S21, _a0123, 1);
-                _rows3 = vfma_lane_f16(_rows3, _S31, _a0123, 1);
-                _rows1 = vfma_lane_f16(_rows1, _S12, _a0123, 2);
-                _rows2 = vfma_lane_f16(_rows2, _S22, _a0123, 2);
-                _rows3 = vfma_lane_f16(_rows3, _S32, _a0123, 2);
-                _rows1 = vfma_lane_f16(_rows1, _S13, _a0123, 3);
-                _rows2 = vfma_lane_f16(_rows2, _S23, _a0123, 3);
-                _rows3 = vfma_lane_f16(_rows3, _S33, _a0123, 3);
-                vst1_f16(rows1p + dx * 4, _rows1);
-                vst1_f16(rows2p + dx * 4, _rows2);
-                vst1_f16(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const __fp16* S0 = src.row<const __fp16>(sy - 1);
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows0p = rows0;
-            __fp16* rows1p = rows1;
-            __fp16* rows2p = rows2;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                float16x4_t _a0123 = vld1_f16(alphap);
-
-                float16x4_t _S00 = vld1_f16(S0p - 4);
-                float16x4_t _S01 = vld1_f16(S0p + 0);
-                float16x4_t _S02 = vld1_f16(S0p + 4);
-                float16x4_t _S03 = vld1_f16(S0p + 8);
-                float16x4_t _S10 = vld1_f16(S1p - 4);
-                float16x4_t _S11 = vld1_f16(S1p + 0);
-                float16x4_t _S12 = vld1_f16(S1p + 4);
-                float16x4_t _S13 = vld1_f16(S1p + 8);
-                float16x4_t _S20 = vld1_f16(S2p - 4);
-                float16x4_t _S21 = vld1_f16(S2p + 0);
-                float16x4_t _S22 = vld1_f16(S2p + 4);
-                float16x4_t _S23 = vld1_f16(S2p + 8);
-                float16x4_t _S30 = vld1_f16(S3p - 4);
-                float16x4_t _S31 = vld1_f16(S3p + 0);
-                float16x4_t _S32 = vld1_f16(S3p + 4);
-                float16x4_t _S33 = vld1_f16(S3p + 8);
-                float16x4_t _rows0 = vmul_lane_f16(_S00, _a0123, 0);
-                float16x4_t _rows1 = vmul_lane_f16(_S10, _a0123, 0);
-                float16x4_t _rows2 = vmul_lane_f16(_S20, _a0123, 0);
-                float16x4_t _rows3 = vmul_lane_f16(_S30, _a0123, 0);
-                _rows0 = vfma_lane_f16(_rows0, _S01, _a0123, 1);
-                _rows1 = vfma_lane_f16(_rows1, _S11, _a0123, 1);
-                _rows2 = vfma_lane_f16(_rows2, _S21, _a0123, 1);
-                _rows3 = vfma_lane_f16(_rows3, _S31, _a0123, 1);
-                _rows0 = vfma_lane_f16(_rows0, _S02, _a0123, 2);
-                _rows1 = vfma_lane_f16(_rows1, _S12, _a0123, 2);
-                _rows2 = vfma_lane_f16(_rows2, _S22, _a0123, 2);
-                _rows3 = vfma_lane_f16(_rows3, _S32, _a0123, 2);
-                _rows0 = vfma_lane_f16(_rows0, _S03, _a0123, 3);
-                _rows1 = vfma_lane_f16(_rows1, _S13, _a0123, 3);
-                _rows2 = vfma_lane_f16(_rows2, _S23, _a0123, 3);
-                _rows3 = vfma_lane_f16(_rows3, _S33, _a0123, 3);
-                vst1_f16(rows0p + dx * 4, _rows0);
-                vst1_f16(rows1p + dx * 4, _rows1);
-                vst1_f16(rows2p + dx * 4, _rows2);
-                vst1_f16(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float16x4_t _b0123 = vld1_f16(beta);
-
-        __fp16* rows0p = rows0;
-        __fp16* rows1p = rows1;
-        __fp16* rows2p = rows2;
-        __fp16* rows3p = rows3;
-        __fp16* Dp = dst.row<__fp16>(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            float16x4_t _rows0 = vld1_f16(rows0p);
-            float16x4_t _rows1 = vld1_f16(rows1p);
-            float16x4_t _rows2 = vld1_f16(rows2p);
-            float16x4_t _rows3 = vld1_f16(rows3p);
-            float16x4_t _Dp = vmul_lane_f16(_rows0, _b0123, 0);
-            _Dp = vfma_lane_f16(_Dp, _rows1, _b0123, 1);
-            _Dp = vfma_lane_f16(_Dp, _rows2, _b0123, 2);
-            _Dp = vfma_lane_f16(_Dp, _rows3, _b0123, 3);
-            vst1_f16(Dp, _Dp);
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-            rows2p += 4;
-            rows3p += 4;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack8_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack8_fp16s.h
deleted file mode 100644
index 75157b6d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bicubic_pack8_fp16s.h
+++ /dev/null
@@ -1,271 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bicubic_image_pack8_fp16sa(const Mat& src, Mat& dst, __fp16* alpha, int* xofs, __fp16* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)8 * 2u, 8);
-    Mat rowsbuf1(w, (size_t)8 * 2u, 8);
-    Mat rowsbuf2(w, (size_t)8 * 2u, 8);
-    Mat rowsbuf3(w, (size_t)8 * 2u, 8);
-    __fp16* rows0 = rowsbuf0;
-    __fp16* rows1 = rowsbuf1;
-    __fp16* rows2 = rowsbuf2;
-    __fp16* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            __fp16* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 8;
-                const __fp16* S3p = S3 + sx;
-
-                float16x4_t _a0123 = vld1_f16(alphap);
-
-                float16x8_t _S30 = vld1q_f16(S3p - 8);
-                float16x8_t _S31 = vld1q_f16(S3p + 0);
-                float16x8_t _S32 = vld1q_f16(S3p + 8);
-                float16x8_t _S33 = vld1q_f16(S3p + 16);
-                float16x8_t _rows3 = vmulq_lane_f16(_S30, _a0123, 0);
-                _rows3 = vfmaq_lane_f16(_rows3, _S31, _a0123, 1);
-                _rows3 = vfmaq_lane_f16(_rows3, _S32, _a0123, 2);
-                _rows3 = vfmaq_lane_f16(_rows3, _S33, _a0123, 3);
-                vst1q_f16(rows3p + dx * 8, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            __fp16* rows0_old = rows0;
-            __fp16* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows2p = rows2;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 8;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                float16x4_t _a0123 = vld1_f16(alphap);
-
-                float16x8_t _S20 = vld1q_f16(S2p - 8);
-                float16x8_t _S21 = vld1q_f16(S2p + 0);
-                float16x8_t _S22 = vld1q_f16(S2p + 8);
-                float16x8_t _S23 = vld1q_f16(S2p + 16);
-                float16x8_t _S30 = vld1q_f16(S3p - 8);
-                float16x8_t _S31 = vld1q_f16(S3p + 0);
-                float16x8_t _S32 = vld1q_f16(S3p + 8);
-                float16x8_t _S33 = vld1q_f16(S3p + 16);
-                float16x8_t _rows2 = vmulq_lane_f16(_S20, _a0123, 0);
-                float16x8_t _rows3 = vmulq_lane_f16(_S30, _a0123, 0);
-                _rows2 = vfmaq_lane_f16(_rows2, _S21, _a0123, 1);
-                _rows3 = vfmaq_lane_f16(_rows3, _S31, _a0123, 1);
-                _rows2 = vfmaq_lane_f16(_rows2, _S22, _a0123, 2);
-                _rows3 = vfmaq_lane_f16(_rows3, _S32, _a0123, 2);
-                _rows2 = vfmaq_lane_f16(_rows2, _S23, _a0123, 3);
-                _rows3 = vfmaq_lane_f16(_rows3, _S33, _a0123, 3);
-                vst1q_f16(rows2p + dx * 8, _rows2);
-                vst1q_f16(rows3p + dx * 8, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            __fp16* rows0_old = rows0;
-            __fp16* rows1_old = rows1;
-            __fp16* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows1p = rows1;
-            __fp16* rows2p = rows2;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 8;
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                float16x4_t _a0123 = vld1_f16(alphap);
-
-                float16x8_t _S10 = vld1q_f16(S1p - 8);
-                float16x8_t _S11 = vld1q_f16(S1p + 0);
-                float16x8_t _S12 = vld1q_f16(S1p + 8);
-                float16x8_t _S13 = vld1q_f16(S1p + 16);
-                float16x8_t _S20 = vld1q_f16(S2p - 8);
-                float16x8_t _S21 = vld1q_f16(S2p + 0);
-                float16x8_t _S22 = vld1q_f16(S2p + 8);
-                float16x8_t _S23 = vld1q_f16(S2p + 16);
-                float16x8_t _S30 = vld1q_f16(S3p - 8);
-                float16x8_t _S31 = vld1q_f16(S3p + 0);
-                float16x8_t _S32 = vld1q_f16(S3p + 8);
-                float16x8_t _S33 = vld1q_f16(S3p + 16);
-                float16x8_t _rows1 = vmulq_lane_f16(_S10, _a0123, 0);
-                float16x8_t _rows2 = vmulq_lane_f16(_S20, _a0123, 0);
-                float16x8_t _rows3 = vmulq_lane_f16(_S30, _a0123, 0);
-                _rows1 = vfmaq_lane_f16(_rows1, _S11, _a0123, 1);
-                _rows2 = vfmaq_lane_f16(_rows2, _S21, _a0123, 1);
-                _rows3 = vfmaq_lane_f16(_rows3, _S31, _a0123, 1);
-                _rows1 = vfmaq_lane_f16(_rows1, _S12, _a0123, 2);
-                _rows2 = vfmaq_lane_f16(_rows2, _S22, _a0123, 2);
-                _rows3 = vfmaq_lane_f16(_rows3, _S32, _a0123, 2);
-                _rows1 = vfmaq_lane_f16(_rows1, _S13, _a0123, 3);
-                _rows2 = vfmaq_lane_f16(_rows2, _S23, _a0123, 3);
-                _rows3 = vfmaq_lane_f16(_rows3, _S33, _a0123, 3);
-                vst1q_f16(rows1p + dx * 8, _rows1);
-                vst1q_f16(rows2p + dx * 8, _rows2);
-                vst1q_f16(rows3p + dx * 8, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const __fp16* S0 = src.row<const __fp16>(sy - 1);
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows0p = rows0;
-            __fp16* rows1p = rows1;
-            __fp16* rows2p = rows2;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 8;
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                float16x4_t _a0123 = vld1_f16(alphap);
-
-                float16x8_t _S00 = vld1q_f16(S0p - 8);
-                float16x8_t _S01 = vld1q_f16(S0p + 0);
-                float16x8_t _S02 = vld1q_f16(S0p + 8);
-                float16x8_t _S03 = vld1q_f16(S0p + 16);
-                float16x8_t _S10 = vld1q_f16(S1p - 8);
-                float16x8_t _S11 = vld1q_f16(S1p + 0);
-                float16x8_t _S12 = vld1q_f16(S1p + 8);
-                float16x8_t _S13 = vld1q_f16(S1p + 16);
-                float16x8_t _S20 = vld1q_f16(S2p - 8);
-                float16x8_t _S21 = vld1q_f16(S2p + 0);
-                float16x8_t _S22 = vld1q_f16(S2p + 8);
-                float16x8_t _S23 = vld1q_f16(S2p + 16);
-                float16x8_t _S30 = vld1q_f16(S3p - 8);
-                float16x8_t _S31 = vld1q_f16(S3p + 0);
-                float16x8_t _S32 = vld1q_f16(S3p + 8);
-                float16x8_t _S33 = vld1q_f16(S3p + 16);
-                float16x8_t _rows0 = vmulq_lane_f16(_S00, _a0123, 0);
-                float16x8_t _rows1 = vmulq_lane_f16(_S10, _a0123, 0);
-                float16x8_t _rows2 = vmulq_lane_f16(_S20, _a0123, 0);
-                float16x8_t _rows3 = vmulq_lane_f16(_S30, _a0123, 0);
-                _rows0 = vfmaq_lane_f16(_rows0, _S01, _a0123, 1);
-                _rows1 = vfmaq_lane_f16(_rows1, _S11, _a0123, 1);
-                _rows2 = vfmaq_lane_f16(_rows2, _S21, _a0123, 1);
-                _rows3 = vfmaq_lane_f16(_rows3, _S31, _a0123, 1);
-                _rows0 = vfmaq_lane_f16(_rows0, _S02, _a0123, 2);
-                _rows1 = vfmaq_lane_f16(_rows1, _S12, _a0123, 2);
-                _rows2 = vfmaq_lane_f16(_rows2, _S22, _a0123, 2);
-                _rows3 = vfmaq_lane_f16(_rows3, _S32, _a0123, 2);
-                _rows0 = vfmaq_lane_f16(_rows0, _S03, _a0123, 3);
-                _rows1 = vfmaq_lane_f16(_rows1, _S13, _a0123, 3);
-                _rows2 = vfmaq_lane_f16(_rows2, _S23, _a0123, 3);
-                _rows3 = vfmaq_lane_f16(_rows3, _S33, _a0123, 3);
-                vst1q_f16(rows0p + dx * 8, _rows0);
-                vst1q_f16(rows1p + dx * 8, _rows1);
-                vst1q_f16(rows2p + dx * 8, _rows2);
-                vst1q_f16(rows3p + dx * 8, _rows3);
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float16x4_t _b0123 = vld1_f16(beta);
-
-        __fp16* rows0p = rows0;
-        __fp16* rows1p = rows1;
-        __fp16* rows2p = rows2;
-        __fp16* rows3p = rows3;
-        __fp16* Dp = dst.row<__fp16>(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            float16x8_t _rows0 = vld1q_f16(rows0p);
-            float16x8_t _rows1 = vld1q_f16(rows1p);
-            float16x8_t _rows2 = vld1q_f16(rows2p);
-            float16x8_t _rows3 = vld1q_f16(rows3p);
-            float16x8_t _Dp = vmulq_lane_f16(_rows0, _b0123, 0);
-            _Dp = vfmaq_lane_f16(_Dp, _rows1, _b0123, 1);
-            _Dp = vfmaq_lane_f16(_Dp, _rows2, _b0123, 2);
-            _Dp = vfmaq_lane_f16(_Dp, _rows3, _b0123, 3);
-            vst1q_f16(Dp, _Dp);
-
-            Dp += 8;
-            rows0p += 8;
-            rows1p += 8;
-            rows2p += 8;
-            rows3p += 8;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear.h
deleted file mode 100644
index 38f00207..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear.h
+++ /dev/null
@@ -1,222 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void linear_coeffs(int w, int outw, int* xofs, float* alpha, int align_corner)
-{
-    double scale = (double)w / outw;
-    if (align_corner)
-    {
-        scale = (double)(w - 1) / (outw - 1);
-    }
-
-    for (int dx = 0; dx < outw; dx++)
-    {
-        float fx = (float)((dx + 0.5) * scale - 0.5);
-        if (align_corner)
-        {
-            fx = (float)(dx * scale);
-        }
-
-        int sx = floor(fx);
-        fx -= sx;
-
-        if (sx < 0)
-        {
-            sx = 0;
-            fx = 0.f;
-        }
-        if (sx >= w - 1)
-        {
-            sx = w - 2;
-            fx = 1.f;
-        }
-
-        xofs[dx] = sx;
-
-        alpha[dx * 2] = 1.f - fx;
-        alpha[dx * 2 + 1] = fx;
-    }
-}
-
-static void resize_bilinear_image(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-#if __ARM_NEON
-            for (; dx + 1 < w; dx += 2)
-            {
-                int sx = xofs[dx];
-                int sxn = xofs[dx + 1];
-                const float* S1p = S1 + sx;
-                const float* S1np = S1 + sxn;
-
-                float32x4_t _a = vld1q_f32(alphap);
-                float32x2_t _S1 = vld1_f32(S1p);
-                float32x2_t _S1n = vld1_f32(S1np);
-
-                float32x4_t _S1S1n = vcombine_f32(_S1, _S1n);
-                float32x4_t _ms1 = vmulq_f32(_S1S1n, _a);
-                float32x2_t _rows1 = vpadd_f32(vget_low_f32(_ms1), vget_high_f32(_ms1));
-
-                vst1_f32(rows1p + dx, _rows1);
-
-                alphap += 4;
-            }
-#endif // __ARM_NEON
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows1p[dx] = S1p[0] * a0 + S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const float* S0 = src.row(sy);
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-#if __ARM_NEON
-            for (; dx + 1 < w; dx += 2)
-            {
-                int sx = xofs[dx];
-                int sxn = xofs[dx + 1];
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-                const float* S0np = S0 + sxn;
-                const float* S1np = S1 + sxn;
-
-                float32x4_t _a = vld1q_f32(alphap);
-                float32x2_t _S0 = vld1_f32(S0p);
-                float32x2_t _S1 = vld1_f32(S1p);
-                float32x2_t _S0n = vld1_f32(S0np);
-                float32x2_t _S1n = vld1_f32(S1np);
-
-                float32x4_t _S0S0n = vcombine_f32(_S0, _S0n);
-                float32x4_t _S1S1n = vcombine_f32(_S1, _S1n);
-                float32x4_t _ms0 = vmulq_f32(_S0S0n, _a);
-                float32x4_t _ms1 = vmulq_f32(_S1S1n, _a);
-                float32x2_t _rows0 = vpadd_f32(vget_low_f32(_ms0), vget_high_f32(_ms0));
-                float32x2_t _rows1 = vpadd_f32(vget_low_f32(_ms1), vget_high_f32(_ms1));
-
-                vst1_f32(rows0p + dx, _rows0);
-                vst1_f32(rows1p + dx, _rows1);
-
-                alphap += 4;
-            }
-#endif // __ARM_NEON
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows0p[dx] = S0p[0] * a0 + S0p[1] * a1;
-                rows1p[dx] = S1p[0] * a0 + S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* Dp = dst.row(dy);
-
-#if __ARM_NEON
-        int nn = w >> 3;
-#else
-        int nn = 0;
-#endif
-        int remain = w - (nn << 3);
-
-#if __ARM_NEON
-        float32x4_t _b0 = vdupq_n_f32(b0);
-        float32x4_t _b1 = vdupq_n_f32(b1);
-        for (; nn > 0; nn--)
-        {
-            float32x4_t _rows0 = vld1q_f32(rows0p);
-            float32x4_t _rows1 = vld1q_f32(rows1p);
-
-            float32x4_t _Dp = vmulq_f32(_rows0, _b0);
-            _Dp = vmlaq_f32(_Dp, _rows1, _b1);
-
-            vst1q_f32(Dp, _Dp);
-
-            float32x4_t _rows0n = vld1q_f32(rows0p + 4);
-            float32x4_t _rows1n = vld1q_f32(rows1p + 4);
-
-            float32x4_t _Dpn = vmulq_f32(_rows0n, _b0);
-            _Dpn = vmlaq_f32(_Dpn, _rows1n, _b1);
-
-            vst1q_f32(Dp + 4, _Dpn);
-
-            Dp += 8;
-            rows0p += 8;
-            rows1p += 8;
-        }
-#endif // __ARM_NEON
-        for (; remain; --remain)
-        {
-            //             D[x] = rows0[x]*b0 + rows1[x]*b1;
-            *Dp++ = *rows0p++ * b0 + *rows1p++ * b1;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_bf16s.h
deleted file mode 100644
index 84583f83..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_bf16s.h
+++ /dev/null
@@ -1,135 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bilinear_image_bf16s(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const unsigned short* S1 = src.row<const unsigned short>(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const unsigned short* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows1p[dx] = bfloat16_to_float32(S1p[0]) * a0 + bfloat16_to_float32(S1p[1]) * a1;
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const unsigned short* S0 = src.row<const unsigned short>(sy);
-            const unsigned short* S1 = src.row<const unsigned short>(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const unsigned short* S0p = S0 + sx;
-                const unsigned short* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows0p[dx] = bfloat16_to_float32(S0p[0]) * a0 + bfloat16_to_float32(S0p[1]) * a1;
-                rows1p[dx] = bfloat16_to_float32(S1p[0]) * a0 + bfloat16_to_float32(S1p[1]) * a1;
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        unsigned short* Dp = dst.row<unsigned short>(dy);
-
-#if __ARM_NEON
-        int nn = w >> 3;
-#else
-        int nn = 0;
-#endif
-        int remain = w - (nn << 3);
-
-#if __ARM_NEON
-        float32x4_t _b0 = vdupq_n_f32(b0);
-        float32x4_t _b1 = vdupq_n_f32(b1);
-        for (; nn > 0; nn--)
-        {
-            float32x4_t _rows0 = vld1q_f32(rows0p);
-            float32x4_t _rows1 = vld1q_f32(rows1p);
-
-            float32x4_t _Dp = vmulq_f32(_rows0, _b0);
-            _Dp = vmlaq_f32(_Dp, _rows1, _b1);
-
-            vst1_u16(Dp, float2bfloat(_Dp));
-
-            float32x4_t _rows0n = vld1q_f32(rows0p + 4);
-            float32x4_t _rows1n = vld1q_f32(rows1p + 4);
-
-            float32x4_t _Dpn = vmulq_f32(_rows0n, _b0);
-            _Dpn = vmlaq_f32(_Dpn, _rows1n, _b1);
-
-            vst1_u16(Dp + 4, float2bfloat(_Dpn));
-
-            Dp += 8;
-            rows0p += 8;
-            rows1p += 8;
-        }
-#endif // __ARM_NEON
-        for (; remain; --remain)
-        {
-            //             D[x] = rows0[x]*b0 + rows1[x]*b1;
-            *Dp++ = float32_to_bfloat16(*rows0p++ * b0 + *rows1p++ * b1);
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_fp16s.h
deleted file mode 100644
index 33314e7a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_fp16s.h
+++ /dev/null
@@ -1,274 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void linear_coeffs_fp16sa(int w, int outw, int* xofs, __fp16* alpha, int align_corner)
-{
-    double scale = (double)w / outw;
-    if (align_corner)
-    {
-        scale = (double)(w - 1) / (outw - 1);
-    }
-
-    for (int dx = 0; dx < outw; dx++)
-    {
-        float fx = (float)((dx + 0.5) * scale - 0.5);
-        if (align_corner)
-        {
-            fx = static_cast<float>(dx * scale);
-        }
-
-        int sx = floor(fx);
-        fx -= sx;
-
-        if (sx < 0)
-        {
-            sx = 0;
-            fx = 0.f;
-        }
-        if (sx >= w - 1)
-        {
-            sx = w - 2;
-            fx = 1.f;
-        }
-
-        xofs[dx] = sx;
-
-        alpha[dx * 2] = (__fp16)(1.f - fx);
-        alpha[dx * 2 + 1] = (__fp16)fx;
-    }
-}
-
-static void resize_bilinear_image_fp16s(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows1p[dx] = (float)S1p[0] * a0 + (float)S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const __fp16* S0 = src.row<const __fp16>(sy);
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows0p[dx] = (float)S0p[0] * a0 + (float)S0p[1] * a1;
-                rows1p[dx] = (float)S1p[0] * a0 + (float)S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        __fp16* Dp = dst.row<__fp16>(dy);
-
-        int nn = w >> 3;
-        int remain = w - (nn << 3);
-
-        float32x4_t _b0 = vdupq_n_f32(b0);
-        float32x4_t _b1 = vdupq_n_f32(b1);
-        for (; nn > 0; nn--)
-        {
-            float32x4_t _rows0 = vld1q_f32(rows0p);
-            float32x4_t _rows1 = vld1q_f32(rows1p);
-
-            float32x4_t _Dp = vmulq_f32(_rows0, _b0);
-            _Dp = vfmaq_f32(_Dp, _rows1, _b1);
-
-            vst1_f16(Dp, vcvt_f16_f32(_Dp));
-
-            float32x4_t _rows0n = vld1q_f32(rows0p + 4);
-            float32x4_t _rows1n = vld1q_f32(rows1p + 4);
-
-            float32x4_t _Dn = vmulq_f32(_rows0n, _b0);
-            _Dn = vfmaq_f32(_Dn, _rows1n, _b1);
-
-            vst1_f16(Dp + 4, vcvt_f16_f32(_Dn));
-
-            Dp += 8;
-            rows0p += 8;
-            rows1p += 8;
-        }
-        for (; remain; --remain)
-        {
-            // D[x] = rows0[x]*b0 + rows1[x]*b1;
-            *Dp++ = (__fp16)(*rows0p++ * b0 + *rows1p++ * b1);
-        }
-
-        beta += 2;
-    }
-}
-
-static void resize_bilinear_image_fp16sa(const Mat& src, Mat& dst, __fp16* alpha, int* xofs, __fp16* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)2u);
-    Mat rowsbuf1(w, (size_t)2u);
-    __fp16* rows0 = rowsbuf0;
-    __fp16* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            __fp16* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S1p = S1 + sx;
-
-                __fp16 a0 = alphap[0];
-                __fp16 a1 = alphap[1];
-                rows1p[dx] = S1p[0] * a0 + S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const __fp16* S0 = src.row<const __fp16>(sy);
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows0p = rows0;
-            __fp16* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-
-                __fp16 a0 = alphap[0];
-                __fp16 a1 = alphap[1];
-                rows0p[dx] = S0p[0] * a0 + S0p[1] * a1;
-                rows1p[dx] = S1p[0] * a0 + S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        __fp16 b0 = beta[0];
-        __fp16 b1 = beta[1];
-
-        __fp16* rows0p = rows0;
-        __fp16* rows1p = rows1;
-        __fp16* Dp = dst.row<__fp16>(dy);
-
-        int nn = w >> 3;
-        int remain = w - (nn << 3);
-
-        float16x8_t _b0 = vdupq_n_f16(b0);
-        float16x8_t _b1 = vdupq_n_f16(b1);
-        for (; nn > 0; nn--)
-        {
-            float16x8_t _rows0 = vld1q_f16(rows0p);
-            float16x8_t _rows1 = vld1q_f16(rows1p);
-
-            float16x8_t _Dp = vmulq_f16(_rows0, _b0);
-            _Dp = vfmaq_f16(_Dp, _rows1, _b1);
-
-            vst1q_f16(Dp, _Dp);
-
-            Dp += 8;
-            rows0p += 8;
-            rows1p += 8;
-        }
-        for (; remain; --remain)
-        {
-            // D[x] = rows0[x]*b0 + rows1[x]*b1;
-            *Dp++ = (__fp16)(*rows0p++ * b0 + *rows1p++ * b1);
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack4.h
deleted file mode 100644
index 46f20485..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack4.h
+++ /dev/null
@@ -1,120 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bilinear_image_pack4(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf1(w, (size_t)4 * 4u, 4);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S1p = S1 + sx;
-
-                float32x2_t _a01 = vld1_f32(alphap);
-
-                float32x4_t _S10 = vld1q_f32(S1p);
-                float32x4_t _S11 = vld1q_f32(S1p + 4);
-                float32x4_t _rows1 = vmulq_lane_f32(_S10, _a01, 0);
-                _rows1 = vmlaq_lane_f32(_rows1, _S11, _a01, 1);
-                vst1q_f32(rows1p + dx * 4, _rows1);
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const float* S0 = src.row(sy);
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-
-                float32x2_t _a01 = vld1_f32(alphap);
-
-                float32x4_t _S00 = vld1q_f32(S0p);
-                float32x4_t _S01 = vld1q_f32(S0p + 4);
-                float32x4_t _S10 = vld1q_f32(S1p);
-                float32x4_t _S11 = vld1q_f32(S1p + 4);
-                float32x4_t _rows0 = vmulq_lane_f32(_S00, _a01, 0);
-                float32x4_t _rows1 = vmulq_lane_f32(_S10, _a01, 0);
-                _rows0 = vmlaq_lane_f32(_rows0, _S01, _a01, 1);
-                _rows1 = vmlaq_lane_f32(_rows1, _S11, _a01, 1);
-                vst1q_f32(rows0p + dx * 4, _rows0);
-                vst1q_f32(rows1p + dx * 4, _rows1);
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float32x2_t _b01 = vld1_f32(beta);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* Dp = dst.row(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            float32x4_t _rows0 = vld1q_f32(rows0p);
-            float32x4_t _rows1 = vld1q_f32(rows1p);
-            float32x4_t _Dp = vmulq_lane_f32(_rows0, _b01, 0);
-            _Dp = vmlaq_lane_f32(_Dp, _rows1, _b01, 1);
-            vst1q_f32(Dp, _Dp);
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack4_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack4_bf16s.h
deleted file mode 100644
index 87ebac50..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack4_bf16s.h
+++ /dev/null
@@ -1,120 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bilinear_image_pack4_bf16s(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf1(w, (size_t)4 * 4u, 4);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const unsigned short* S1 = src.row<const unsigned short>(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const unsigned short* S1p = S1 + sx;
-
-                float32x2_t _a01 = vld1_f32(alphap);
-
-                float32x4_t _S10 = bfloat2float(vld1_u16(S1p));
-                float32x4_t _S11 = bfloat2float(vld1_u16(S1p + 4));
-                float32x4_t _rows1 = vmulq_lane_f32(_S10, _a01, 0);
-                _rows1 = vmlaq_lane_f32(_rows1, _S11, _a01, 1);
-                vst1q_f32(rows1p + dx * 4, _rows1);
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const unsigned short* S0 = src.row<const unsigned short>(sy);
-            const unsigned short* S1 = src.row<const unsigned short>(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const unsigned short* S0p = S0 + sx;
-                const unsigned short* S1p = S1 + sx;
-
-                float32x2_t _a01 = vld1_f32(alphap);
-
-                float32x4_t _S00 = bfloat2float(vld1_u16(S0p));
-                float32x4_t _S01 = bfloat2float(vld1_u16(S0p + 4));
-                float32x4_t _S10 = bfloat2float(vld1_u16(S1p));
-                float32x4_t _S11 = bfloat2float(vld1_u16(S1p + 4));
-                float32x4_t _rows0 = vmulq_lane_f32(_S00, _a01, 0);
-                float32x4_t _rows1 = vmulq_lane_f32(_S10, _a01, 0);
-                _rows0 = vmlaq_lane_f32(_rows0, _S01, _a01, 1);
-                _rows1 = vmlaq_lane_f32(_rows1, _S11, _a01, 1);
-                vst1q_f32(rows0p + dx * 4, _rows0);
-                vst1q_f32(rows1p + dx * 4, _rows1);
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float32x2_t _b01 = vld1_f32(beta);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        unsigned short* Dp = dst.row<unsigned short>(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            float32x4_t _rows0 = vld1q_f32(rows0p);
-            float32x4_t _rows1 = vld1q_f32(rows1p);
-            float32x4_t _Dp = vmulq_lane_f32(_rows0, _b01, 0);
-            _Dp = vmlaq_lane_f32(_Dp, _rows1, _b01, 1);
-            vst1_u16(Dp, float2bfloat(_Dp));
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack4_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack4_fp16s.h
deleted file mode 100644
index 0b6c6de0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack4_fp16s.h
+++ /dev/null
@@ -1,227 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bilinear_image_pack4_fp16s(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf1(w, (size_t)4 * 4u, 4);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const __fp16* S1p = S1 + sx;
-
-                float32x2_t _a01 = vld1_f32(alphap);
-
-                float32x4_t _S10 = vcvt_f32_f16(vld1_f16(S1p));
-                float32x4_t _S11 = vcvt_f32_f16(vld1_f16(S1p + 4));
-                float32x4_t _rows1 = vmulq_lane_f32(_S10, _a01, 0);
-                _rows1 = vmlaq_lane_f32(_rows1, _S11, _a01, 1);
-                vst1q_f32(rows1p + dx * 4, _rows1);
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const __fp16* S0 = src.row<const __fp16>(sy);
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-
-                float32x2_t _a01 = vld1_f32(alphap);
-
-                float32x4_t _S00 = vcvt_f32_f16(vld1_f16(S0p));
-                float32x4_t _S01 = vcvt_f32_f16(vld1_f16(S0p + 4));
-                float32x4_t _S10 = vcvt_f32_f16(vld1_f16(S1p));
-                float32x4_t _S11 = vcvt_f32_f16(vld1_f16(S1p + 4));
-                float32x4_t _rows0 = vmulq_lane_f32(_S00, _a01, 0);
-                float32x4_t _rows1 = vmulq_lane_f32(_S10, _a01, 0);
-                _rows0 = vmlaq_lane_f32(_rows0, _S01, _a01, 1);
-                _rows1 = vmlaq_lane_f32(_rows1, _S11, _a01, 1);
-                vst1q_f32(rows0p + dx * 4, _rows0);
-                vst1q_f32(rows1p + dx * 4, _rows1);
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float32x2_t _b01 = vld1_f32(beta);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        __fp16* Dp = dst.row<__fp16>(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            float32x4_t _rows0 = vld1q_f32(rows0p);
-            float32x4_t _rows1 = vld1q_f32(rows1p);
-            float32x4_t _Dp = vmulq_lane_f32(_rows0, _b01, 0);
-            _Dp = vmlaq_lane_f32(_Dp, _rows1, _b01, 1);
-            vst1_f16(Dp, vcvt_f16_f32(_Dp));
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-        }
-
-        beta += 2;
-    }
-}
-
-static void resize_bilinear_image_pack4_fp16sa(const Mat& src, Mat& dst, __fp16* alpha, int* xofs, __fp16* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)4 * 2u, 4);
-    Mat rowsbuf1(w, (size_t)4 * 2u, 4);
-    __fp16* rows0 = rowsbuf0;
-    __fp16* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            __fp16* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const __fp16* S1p = S1 + sx;
-
-                float16x4_t _a01 = vld1_f16(alphap);
-
-                float16x4_t _S10 = vld1_f16(S1p);
-                float16x4_t _S11 = vld1_f16(S1p + 4);
-                float16x4_t _rows1 = vmul_lane_f16(_S10, _a01, 0);
-                _rows1 = vfma_lane_f16(_rows1, _S11, _a01, 1);
-                vst1_f16(rows1p + dx * 4, _rows1);
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const __fp16* S0 = src.row<const __fp16>(sy);
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows0p = rows0;
-            __fp16* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-
-                float16x4_t _a01 = vld1_f16(alphap);
-
-                float16x4_t _S00 = vld1_f16(S0p);
-                float16x4_t _S01 = vld1_f16(S0p + 4);
-                float16x4_t _S10 = vld1_f16(S1p);
-                float16x4_t _S11 = vld1_f16(S1p + 4);
-                float16x4_t _rows0 = vmul_lane_f16(_S00, _a01, 0);
-                float16x4_t _rows1 = vmul_lane_f16(_S10, _a01, 0);
-                _rows0 = vfma_lane_f16(_rows0, _S01, _a01, 1);
-                _rows1 = vfma_lane_f16(_rows1, _S11, _a01, 1);
-                vst1_f16(rows0p + dx * 4, _rows0);
-                vst1_f16(rows1p + dx * 4, _rows1);
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float16x4_t _b01 = vld1_f16(beta);
-
-        __fp16* rows0p = rows0;
-        __fp16* rows1p = rows1;
-        __fp16* Dp = dst.row<__fp16>(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            float16x4_t _rows0 = vld1_f16(rows0p);
-            float16x4_t _rows1 = vld1_f16(rows1p);
-            float16x4_t _Dp = vmul_lane_f16(_rows0, _b01, 0);
-            _Dp = vfma_lane_f16(_Dp, _rows1, _b01, 1);
-            vst1_f16(Dp, _Dp);
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack8_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack8_fp16s.h
deleted file mode 100644
index c0f84b01..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/interp_bilinear_pack8_fp16s.h
+++ /dev/null
@@ -1,120 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bilinear_image_pack8_fp16sa(const Mat& src, Mat& dst, __fp16* alpha, int* xofs, __fp16* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)8 * 2u, 8);
-    Mat rowsbuf1(w, (size_t)8 * 2u, 8);
-    __fp16* rows0 = rowsbuf0;
-    __fp16* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            __fp16* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 8;
-                const __fp16* S1p = S1 + sx;
-
-                float16x4_t _a01 = vld1_f16(alphap);
-
-                float16x8_t _S10 = vld1q_f16(S1p);
-                float16x8_t _S11 = vld1q_f16(S1p + 8);
-                float16x8_t _rows1 = vmulq_lane_f16(_S10, _a01, 0);
-                _rows1 = vfmaq_lane_f16(_rows1, _S11, _a01, 1);
-                vst1q_f16(rows1p + dx * 8, _rows1);
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const __fp16* S0 = src.row<const __fp16>(sy);
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows0p = rows0;
-            __fp16* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 8;
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-
-                float16x4_t _a01 = vld1_f16(alphap);
-
-                float16x8_t _S00 = vld1q_f16(S0p);
-                float16x8_t _S01 = vld1q_f16(S0p + 8);
-                float16x8_t _S10 = vld1q_f16(S1p);
-                float16x8_t _S11 = vld1q_f16(S1p + 8);
-                float16x8_t _rows0 = vmulq_lane_f16(_S00, _a01, 0);
-                float16x8_t _rows1 = vmulq_lane_f16(_S10, _a01, 0);
-                _rows0 = vfmaq_lane_f16(_rows0, _S01, _a01, 1);
-                _rows1 = vfmaq_lane_f16(_rows1, _S11, _a01, 1);
-                vst1q_f16(rows0p + dx * 8, _rows0);
-                vst1q_f16(rows1p + dx * 8, _rows1);
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float16x4_t _b01 = vld1_f16(beta);
-
-        __fp16* rows0p = rows0;
-        __fp16* rows1p = rows1;
-        __fp16* Dp = dst.row<__fp16>(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            float16x8_t _rows0 = vld1q_f16(rows0p);
-            float16x8_t _rows1 = vld1q_f16(rows1p);
-            float16x8_t _Dp = vmulq_lane_f16(_rows0, _b01, 0);
-            _Dp = vfmaq_lane_f16(_Dp, _rows1, _b01, 1);
-            vst1q_f16(Dp, _Dp);
-
-            Dp += 8;
-            rows0p += 8;
-            rows1p += 8;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/lrn_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/lrn_arm.h
deleted file mode 100644
index f2c43ba0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/lrn_arm.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_LRN_ARM_H
-#define LAYER_LRN_ARM_H
-
-#include "lrn.h"
-
-namespace ncnn {
-
-class LRN_arm : public LRN
-{
-public:
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_LRN_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/lstm_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/lstm_arm.h
deleted file mode 100644
index b5ee1092..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/lstm_arm.h
+++ /dev/null
@@ -1,53 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_LSTM_ARM_H
-#define LAYER_LSTM_ARM_H
-
-#include "lstm.h"
-
-namespace ncnn {
-
-class LSTM_arm : public LSTM
-{
-public:
-    LSTM_arm();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int create_pipeline_bf16s(const Option& opt);
-    int forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_bf16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-#endif
-
-public:
-    Mat weight_xc_data_packed;
-    Mat bias_c_data_packed;
-    Mat weight_hc_data_packed;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_LSTM_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/matmul_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/matmul_arm.h
deleted file mode 100644
index a4537300..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/matmul_arm.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MATMUL_ARM_H
-#define LAYER_MATMUL_ARM_H
-
-#include "matmul.h"
-
-namespace ncnn {
-
-class MatMul_arm : public MatMul
-{
-public:
-    MatMul_arm();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    Layer* gemm;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MATMUL_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/mish_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/mish_arm.h
deleted file mode 100644
index 9f99a7a1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/mish_arm.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MISH_ARM_H
-#define LAYER_MISH_ARM_H
-
-#include "mish.h"
-
-namespace ncnn {
-
-class Mish_arm : public Mish
-{
-public:
-    Mish_arm();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_inplace_bf16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MISH_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/multiheadattention_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/multiheadattention_arm.h
deleted file mode 100644
index f1b721f2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/multiheadattention_arm.h
+++ /dev/null
@@ -1,46 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MULTIHEADATTENTION_ARM_H
-#define LAYER_MULTIHEADATTENTION_ARM_H
-
-#include "multiheadattention.h"
-
-namespace ncnn {
-
-class MultiHeadAttention_arm : public MultiHeadAttention
-{
-public:
-    MultiHeadAttention_arm();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    Layer* q_gemm;
-    Layer* k_gemm;
-    Layer* v_gemm;
-    Layer* o_gemm;
-
-    Layer* qk_gemm;
-    Layer* qkv_gemm;
-
-    Layer* qk_softmax;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MULTIHEADATTENTION_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/neon_mathfun.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/neon_mathfun.h
deleted file mode 100644
index 2b4094a9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/neon_mathfun.h
+++ /dev/null
@@ -1,399 +0,0 @@
-/* NEON implementation of sin, cos, exp and log
- *
- *   Inspired by Intel Approximate Math library, and based on the
- *   corresponding algorithms of the cephes math library
- */
-
-/* Copyright (C) 2011  Julien Pommier
- *
- *  This software is provided 'as-is', without any express or implied
- *  warranty.  In no event will the authors be held liable for any damages
- *  arising from the use of this software.
- *
- *  Permission is granted to anyone to use this software for any purpose,
- *  including commercial applications, and to alter it and redistribute it
- *  freely, subject to the following restrictions:
- *
- *  1. The origin of this software must not be misrepresented; you must not
- *     claim that you wrote the original software. If you use this software
- *     in a product, an acknowledgment in the product documentation would be
- *     appreciated but is not required.
- *  2. Altered source versions must be plainly marked as such, and must not be
- *     misrepresented as being the original software.
- *  3. This notice may not be removed or altered from any source distribution.
- *
- *  (this is the zlib license)
- */
-
-#ifndef NEON_MATHFUN_H
-#define NEON_MATHFUN_H
-
-#include <arm_neon.h>
-
-#define c_inv_mant_mask ~0x7f800000u
-#define c_cephes_SQRTHF 0.707106781186547524
-#define c_cephes_log_p0 7.0376836292E-2
-#define c_cephes_log_p1 -1.1514610310E-1
-#define c_cephes_log_p2 1.1676998740E-1
-#define c_cephes_log_p3 -1.2420140846E-1
-#define c_cephes_log_p4 +1.4249322787E-1
-#define c_cephes_log_p5 -1.6668057665E-1
-#define c_cephes_log_p6 +2.0000714765E-1
-#define c_cephes_log_p7 -2.4999993993E-1
-#define c_cephes_log_p8 +3.3333331174E-1
-#define c_cephes_log_q1 -2.12194440e-4
-#define c_cephes_log_q2 0.693359375
-
-/* natural logarithm computed for 4 simultaneous float
- *   return NaN for x <= 0
- */
-static inline float32x4_t log_ps(float32x4_t x)
-{
-    float32x4_t one = vdupq_n_f32(1);
-
-    x = vmaxq_f32(x, vdupq_n_f32(0)); /* force flush to zero on denormal values */
-    uint32x4_t invalid_mask = vcleq_f32(x, vdupq_n_f32(0));
-
-    int32x4_t ux = vreinterpretq_s32_f32(x);
-
-    int32x4_t emm0 = vshrq_n_s32(ux, 23);
-
-    /* keep only the fractional part */
-    ux = vandq_s32(ux, vdupq_n_s32(c_inv_mant_mask));
-    ux = vorrq_s32(ux, vreinterpretq_s32_f32(vdupq_n_f32(0.5f)));
-    x = vreinterpretq_f32_s32(ux);
-
-    emm0 = vsubq_s32(emm0, vdupq_n_s32(0x7f));
-    float32x4_t e = vcvtq_f32_s32(emm0);
-
-    e = vaddq_f32(e, one);
-
-    /* part2:
-     *     if( x < SQRTHF ) {
-     *       e -= 1;
-     *       x = x + x - 1.0;
-     *     } else { x = x - 1.0; }
-     */
-    uint32x4_t mask = vcltq_f32(x, vdupq_n_f32(c_cephes_SQRTHF));
-    float32x4_t tmp = vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(x), mask));
-    x = vsubq_f32(x, one);
-    e = vsubq_f32(e, vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(one), mask)));
-    x = vaddq_f32(x, tmp);
-
-    float32x4_t z = vmulq_f32(x, x);
-
-    float32x4_t y = vdupq_n_f32(c_cephes_log_p0);
-    y = vmlaq_f32(vdupq_n_f32(c_cephes_log_p1), y, x);
-    y = vmlaq_f32(vdupq_n_f32(c_cephes_log_p2), y, x);
-    y = vmlaq_f32(vdupq_n_f32(c_cephes_log_p3), y, x);
-    y = vmlaq_f32(vdupq_n_f32(c_cephes_log_p4), y, x);
-    y = vmlaq_f32(vdupq_n_f32(c_cephes_log_p5), y, x);
-    y = vmlaq_f32(vdupq_n_f32(c_cephes_log_p6), y, x);
-    y = vmlaq_f32(vdupq_n_f32(c_cephes_log_p7), y, x);
-    y = vmlaq_f32(vdupq_n_f32(c_cephes_log_p8), y, x);
-    y = vmulq_f32(y, x);
-
-    y = vmulq_f32(y, z);
-
-    y = vmlaq_f32(y, e, vdupq_n_f32(c_cephes_log_q1));
-
-    y = vmlsq_f32(y, z, vdupq_n_f32(0.5f));
-
-    x = vaddq_f32(x, y);
-    x = vmlaq_f32(x, e, vdupq_n_f32(c_cephes_log_q2));
-    x = vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(x), invalid_mask)); // negative arg will be NAN
-    return x;
-}
-
-#define c_exp_hi 88.3762626647949f
-#define c_exp_lo -88.3762626647949f
-
-#define c_cephes_LOG2EF 1.44269504088896341
-#define c_cephes_exp_C1 0.693359375
-#define c_cephes_exp_C2 -2.12194440e-4
-
-#define c_cephes_exp_p0 1.9875691500E-4
-#define c_cephes_exp_p1 1.3981999507E-3
-#define c_cephes_exp_p2 8.3334519073E-3
-#define c_cephes_exp_p3 4.1665795894E-2
-#define c_cephes_exp_p4 1.6666665459E-1
-#define c_cephes_exp_p5 5.0000001201E-1
-
-/* exp() computed for 4 float at once */
-static inline float32x4_t exp_ps(float32x4_t x)
-{
-    float32x4_t tmp, fx;
-
-    float32x4_t one = vdupq_n_f32(1);
-    x = vminq_f32(x, vdupq_n_f32(c_exp_hi));
-    x = vmaxq_f32(x, vdupq_n_f32(c_exp_lo));
-
-    /* express exp(x) as exp(g + n*log(2)) */
-    fx = vmlaq_f32(vdupq_n_f32(0.5f), x, vdupq_n_f32(c_cephes_LOG2EF));
-
-    /* perform a floorf */
-    tmp = vcvtq_f32_s32(vcvtq_s32_f32(fx));
-
-    /* if greater, substract 1 */
-    uint32x4_t mask = vcgtq_f32(tmp, fx);
-    mask = vandq_u32(mask, vreinterpretq_u32_f32(one));
-
-    fx = vsubq_f32(tmp, vreinterpretq_f32_u32(mask));
-
-    tmp = vmulq_f32(fx, vdupq_n_f32(c_cephes_exp_C1));
-    float32x4_t z = vmulq_f32(fx, vdupq_n_f32(c_cephes_exp_C2));
-    x = vsubq_f32(x, tmp);
-    x = vsubq_f32(x, z);
-
-    z = vmulq_f32(x, x);
-
-    float32x4_t y = vdupq_n_f32(c_cephes_exp_p0);
-    y = vmlaq_f32(vdupq_n_f32(c_cephes_exp_p1), y, x);
-    y = vmlaq_f32(vdupq_n_f32(c_cephes_exp_p2), y, x);
-    y = vmlaq_f32(vdupq_n_f32(c_cephes_exp_p3), y, x);
-    y = vmlaq_f32(vdupq_n_f32(c_cephes_exp_p4), y, x);
-    y = vmlaq_f32(vdupq_n_f32(c_cephes_exp_p5), y, x);
-
-    y = vmlaq_f32(x, y, z);
-    y = vaddq_f32(y, one);
-
-    /* build 2^n */
-    int32x4_t mm;
-    mm = vcvtq_s32_f32(fx);
-    mm = vaddq_s32(mm, vdupq_n_s32(0x7f));
-    mm = vshlq_n_s32(mm, 23);
-    float32x4_t pow2n = vreinterpretq_f32_s32(mm);
-
-    y = vmulq_f32(y, pow2n);
-    return y;
-}
-
-#define c_minus_cephes_DP1 -0.78515625
-#define c_minus_cephes_DP2 -2.4187564849853515625e-4
-#define c_minus_cephes_DP3 -3.77489497744594108e-8
-#define c_sincof_p0        -1.9515295891E-4
-#define c_sincof_p1        8.3321608736E-3
-#define c_sincof_p2        -1.6666654611E-1
-#define c_coscof_p0        2.443315711809948E-005
-#define c_coscof_p1        -1.388731625493765E-003
-#define c_coscof_p2        4.166664568298827E-002
-#define c_cephes_FOPI      1.27323954473516 // 4 / M_PI
-
-/* evaluation of 4 sines & cosines at once.
- *
- *   The code is the exact rewriting of the cephes sinf function.
- *   Precision is excellent as long as x < 8192 (I did not bother to
- *   take into account the special handling they have for greater values
- *   -- it does not return garbage for arguments over 8192, though, but
- *   the extra precision is missing).
- *
- *   Note that it is such that sinf((float)M_PI) = 8.74e-8, which is the
- *   surprising but correct result.
- *
- *   Note also that when you compute sin(x), cos(x) is available at
- *   almost no extra price so both sin_ps and cos_ps make use of
- *   sincos_ps..
- */
-static inline void sincos_ps(float32x4_t x, float32x4_t* ysin, float32x4_t* ycos)
-{
-    // any x
-    float32x4_t y;
-
-    uint32x4_t emm2;
-
-    uint32x4_t sign_mask_sin, sign_mask_cos;
-    sign_mask_sin = vcltq_f32(x, vdupq_n_f32(0));
-    x = vabsq_f32(x);
-
-    /* scale by 4/Pi */
-    y = vmulq_f32(x, vdupq_n_f32(c_cephes_FOPI));
-
-    /* store the integer part of y in mm0 */
-    emm2 = vcvtq_u32_f32(y);
-    /* j=(j+1) & (~1) (see the cephes sources) */
-    emm2 = vaddq_u32(emm2, vdupq_n_u32(1));
-    emm2 = vandq_u32(emm2, vdupq_n_u32(~1));
-    y = vcvtq_f32_u32(emm2);
-
-    /* get the polynom selection mask
-     *     there is one polynom for 0 <= x <= Pi/4
-     *     and another one for Pi/4<x<=Pi/2
-     *
-     *     Both branches will be computed.
-     */
-    uint32x4_t poly_mask = vtstq_u32(emm2, vdupq_n_u32(2));
-
-    /* The magic pass: "Extended precision modular arithmetic"
-     *     x = ((x - y * DP1) - y * DP2) - y * DP3; */
-    x = vmlaq_f32(x, y, vdupq_n_f32(c_minus_cephes_DP1));
-    x = vmlaq_f32(x, y, vdupq_n_f32(c_minus_cephes_DP2));
-    x = vmlaq_f32(x, y, vdupq_n_f32(c_minus_cephes_DP3));
-
-    sign_mask_sin = veorq_u32(sign_mask_sin, vtstq_u32(emm2, vdupq_n_u32(4)));
-    sign_mask_cos = vtstq_u32(vsubq_u32(emm2, vdupq_n_u32(2)), vdupq_n_u32(4));
-
-    /* Evaluate the first polynom  (0 <= x <= Pi/4) in y1,
-     *     and the second polynom      (Pi/4 <= x <= 0) in y2 */
-    float32x4_t z = vmulq_f32(x, x);
-    float32x4_t y1, y2;
-
-    y1 = vmlaq_f32(vdupq_n_f32(c_coscof_p1), z, vdupq_n_f32(c_coscof_p0));
-    y2 = vmlaq_f32(vdupq_n_f32(c_sincof_p1), z, vdupq_n_f32(c_sincof_p0));
-    y1 = vmlaq_f32(vdupq_n_f32(c_coscof_p2), y1, z);
-    y2 = vmlaq_f32(vdupq_n_f32(c_sincof_p2), y2, z);
-    y1 = vmulq_f32(y1, z);
-    y2 = vmulq_f32(y2, z);
-    y1 = vmulq_f32(y1, z);
-    y1 = vmlsq_f32(y1, z, vdupq_n_f32(0.5f));
-    y2 = vmlaq_f32(x, y2, x);
-    y1 = vaddq_f32(y1, vdupq_n_f32(1));
-
-    /* select the correct result from the two polynoms */
-    float32x4_t ys = vbslq_f32(poly_mask, y1, y2);
-    float32x4_t yc = vbslq_f32(poly_mask, y2, y1);
-    *ysin = vbslq_f32(sign_mask_sin, vnegq_f32(ys), ys);
-    *ycos = vbslq_f32(sign_mask_cos, yc, vnegq_f32(yc));
-}
-
-static inline float32x4_t sin_ps(float32x4_t x)
-{
-    float32x4_t ysin, ycos;
-    sincos_ps(x, &ysin, &ycos);
-    return ysin;
-}
-
-static inline float32x4_t cos_ps(float32x4_t x)
-{
-    float32x4_t ysin, ycos;
-    sincos_ps(x, &ysin, &ycos);
-    return ycos;
-}
-
-static inline float32x4_t div_ps(float32x4_t a, float32x4_t b)
-{
-#if __aarch64__
-    return vdivq_f32(a, b);
-#else
-    float32x4_t reciprocal = vrecpeq_f32(b);
-    reciprocal = vmulq_f32(vrecpsq_f32(b, reciprocal), reciprocal);
-    // reciprocal = vmulq_f32(vrecpsq_f32(b, reciprocal), reciprocal);
-    return vmulq_f32(a, reciprocal);
-#endif
-}
-
-static inline float32x4_t tan_ps(float32x4_t x)
-{
-    float32x4_t ysin, ycos;
-    sincos_ps(x, &ysin, &ycos);
-    float32x4_t ytan = div_ps(ysin, ycos);
-    return ytan;
-}
-
-static inline float32x4_t pow_ps(float32x4_t a, float32x4_t b)
-{
-    // pow(x, m) = exp(m * log(x))
-    return exp_ps(vmulq_f32(b, log_ps(a)));
-}
-
-static inline float32x4_t sigmoid_ps(float32x4_t _v)
-{
-    float32x4_t _one = vdupq_n_f32(1.f);
-    _v = vnegq_f32(_v);
-    _v = exp_ps(_v);
-    _v = vaddq_f32(_v, _one);
-    float32x4_t _outp = vrecpeq_f32(_v);
-    // _outp = vmulq_f32(vrecpsq_f32(_v, _outp), _outp);
-    return vmulq_f32(vrecpsq_f32(_v, _outp), _outp);
-}
-
-static const float asinf_lut[7] = {
-    1.5707961728,
-    -0.2145852647,
-    0.0887556286,
-    -0.0488025043,
-    0.0268999482,
-    -0.0111462294,
-    0.0022959648
-};
-
-static inline void asincos_ps(float32x4_t x, float32x4_t* yasin, float32x4_t* yacos)
-{
-    int i = 0;
-    float32x4_t one = vdupq_n_f32(1);
-    float32x4_t negone = vdupq_n_f32(-1);
-    float32x4_t lut[7];
-    float32x4_t xv[5];
-    float32x4_t a0, a1, a2, a3;
-    float32x4_t phx;
-    float32x4_t arcsinx, arcnsinx;
-    float32x4_t sat = vdupq_n_f32(0.9999999f);
-    float32x4_t m_pi_2 = vdupq_n_f32(1.570796326);
-    for (i = 0; i <= 6; i++)
-    {
-        lut[i] = vdupq_n_f32(asinf_lut[i]);
-    }
-
-    uint32x4_t sign_mask_asin, saturate;
-    sign_mask_asin = vcltq_f32(x, vdupq_n_f32(0));
-    x = vabsq_f32(x);
-    saturate = vcgeq_f32(x, one);
-    x = vbslq_f32(saturate, sat, x);
-    float32x4_t y = vsubq_f32(one, x);
-
-#if __aarch64__
-    y = vsqrtq_f32(y);
-#else
-    float32x4_t _reciprocal = vrsqrteq_f32(y);
-    _reciprocal = vmulq_f32(vrsqrtsq_f32(vmulq_f32(y, _reciprocal), _reciprocal), _reciprocal);
-    y = vmulq_f32(y, _reciprocal);
-#endif
-
-    xv[0] = vmulq_f32(x, x);
-    for (i = 1; i < 5; i++)
-    {
-        xv[i] = vmulq_f32(xv[i - 1], x);
-    }
-
-    a0 = vaddq_f32(lut[0], vmulq_f32(lut[1], x));
-    a1 = vaddq_f32(vmulq_f32(lut[2], xv[0]), vmulq_f32(lut[3], xv[1]));
-    a2 = vaddq_f32(vmulq_f32(lut[4], xv[2]), vmulq_f32(lut[5], xv[3]));
-    a3 = vmulq_f32(lut[6], xv[4]);
-    phx = vaddq_f32(vaddq_f32(a0, vaddq_f32(a1, a2)), a3);
-
-    arcsinx = vmulq_f32(y, phx);
-    arcsinx = vsubq_f32(m_pi_2, arcsinx);
-    arcnsinx = vmulq_f32(negone, arcsinx);
-    arcsinx = vbslq_f32(sign_mask_asin, arcnsinx, arcsinx);
-
-    *yasin = arcsinx;
-    *yacos = vsubq_f32(m_pi_2, arcsinx);
-}
-
-static inline float32x4_t asin_ps(float32x4_t x)
-{
-    float32x4_t yasin, yacos;
-    asincos_ps(x, &yasin, &yacos);
-    return yasin;
-}
-
-static inline float32x4_t acos_ps(float32x4_t x)
-{
-    float32x4_t yasin, yacos;
-    asincos_ps(x, &yasin, &yacos);
-    return yacos;
-}
-
-static inline float32x4_t atan2_ps(float32x4_t a, float32x4_t b)
-{
-    //TODO neon optimize
-    float tmpx[4];
-    float tmpy[4];
-    vst1q_f32(tmpx, a);
-    vst1q_f32(tmpy, b);
-    for (int i = 0; i < 4; i++)
-        tmpx[i] = atan2f(tmpx[i], tmpy[i]);
-    return vld1q_f32(tmpx);
-}
-
-#include "neon_mathfun_tanh.h"
-#endif // NEON_MATHFUN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/neon_mathfun_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/neon_mathfun_fp16s.h
deleted file mode 100644
index 2f4864c1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/neon_mathfun_fp16s.h
+++ /dev/null
@@ -1,633 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-/* NEON implementation of sin, cos, exp and log
- *
- *   Inspired by Intel Approximate Math library, and based on the
- *   corresponding algorithms of the cephes math library
- */
-
-/* Copyright (C) 2011  Julien Pommier
- *
- *  This software is provided 'as-is', without any express or implied
- *  warranty.  In no event will the authors be held liable for any damages
- *  arising from the use of this software.
- *
- *  Permission is granted to anyone to use this software for any purpose,
- *  including commercial applications, and to alter it and redistribute it
- *  freely, subject to the following restrictions:
- *
- *  1. The origin of this software must not be misrepresented; you must not
- *     claim that you wrote the original software. If you use this software
- *     in a product, an acknowledgment in the product documentation would be
- *     appreciated but is not required.
- *  2. Altered source versions must be plainly marked as such, and must not be
- *     misrepresented as being the original software.
- *  3. This notice may not be removed or altered from any source distribution.
- *
- *  (this is the zlib license)
- */
-
-#ifndef NEON_MATHFUN_FP16S_H
-#define NEON_MATHFUN_FP16S_H
-
-#include <arm_neon.h>
-
-#define c_inv_mant_mask_f16 -31745 // ~0x7c00u
-#define c_cephes_SQRTHF     0.707106781186547524
-#define c_cephes_log_p0     7.0376836292E-2
-#define c_cephes_log_p1     -1.1514610310E-1
-#define c_cephes_log_p2     1.1676998740E-1
-#define c_cephes_log_p3     -1.2420140846E-1
-#define c_cephes_log_p4     +1.4249322787E-1
-#define c_cephes_log_p5     -1.6668057665E-1
-#define c_cephes_log_p6     +2.0000714765E-1
-#define c_cephes_log_p7     -2.4999993993E-1
-#define c_cephes_log_p8     +3.3333331174E-1
-#define c_cephes_log_q1     -2.12194440e-4
-#define c_cephes_log_q2     0.693359375
-
-/* natural logarithm computed for 4 simultaneous float
- *   return NaN for x <= 0
- */
-static inline float16x4_t log_ps_f16(float16x4_t x)
-{
-    float16x4_t one = vdup_n_f16(1);
-
-    x = vmax_f16(x, vdup_n_f16(0)); /* force flush to zero on denormal values */
-    uint16x4_t invalid_mask = vcle_f16(x, vdup_n_f16(0));
-
-    int16x4_t ux = vreinterpret_s16_f16(x);
-
-    int16x4_t emm0 = vshr_n_s16(ux, 10);
-
-    /* keep only the fractional part */
-    ux = vand_s16(ux, vdup_n_s16(c_inv_mant_mask_f16));
-    ux = vorr_s16(ux, vreinterpret_s16_f16(vdup_n_f16(0.5f)));
-    x = vreinterpret_f16_s16(ux);
-
-    emm0 = vsub_s16(emm0, vdup_n_s16(0xf));
-    float16x4_t e = vcvt_f16_s16(emm0);
-
-    e = vadd_f16(e, one);
-
-    /* part2:
-     *     if( x < SQRTHF ) {
-     *       e -= 1;
-     *       x = x + x - 1.0;
-     *     } else { x = x - 1.0; }
-     */
-    uint16x4_t mask = vclt_f16(x, vdup_n_f16(c_cephes_SQRTHF));
-    float16x4_t tmp = (float16x4_t)(vand_u16((uint16x4_t)(x), mask));
-    x = vsub_f16(x, one);
-    e = vsub_f16(e, (float16x4_t)(vand_u16((uint16x4_t)(one), mask)));
-    x = vadd_f16(x, tmp);
-
-    float16x4_t z = vmul_f16(x, x);
-
-    float16x4_t y = vdup_n_f16(c_cephes_log_p0);
-    y = vfma_f16(vdup_n_f16(c_cephes_log_p1), y, x);
-    y = vfma_f16(vdup_n_f16(c_cephes_log_p2), y, x);
-    y = vfma_f16(vdup_n_f16(c_cephes_log_p3), y, x);
-    y = vfma_f16(vdup_n_f16(c_cephes_log_p4), y, x);
-    y = vfma_f16(vdup_n_f16(c_cephes_log_p5), y, x);
-    y = vfma_f16(vdup_n_f16(c_cephes_log_p6), y, x);
-    y = vfma_f16(vdup_n_f16(c_cephes_log_p7), y, x);
-    y = vfma_f16(vdup_n_f16(c_cephes_log_p8), y, x);
-    y = vmul_f16(y, x);
-
-    y = vmul_f16(y, z);
-
-    y = vfma_f16(y, e, vdup_n_f16(c_cephes_log_q1));
-
-    y = vfms_f16(y, z, vdup_n_f16(0.5f));
-
-    x = vadd_f16(x, y);
-    x = vfma_f16(x, e, vdup_n_f16(c_cephes_log_q2));
-    x = (float16x4_t)(vorr_u16((uint16x4_t)(x), invalid_mask)); // negative arg will be NAN
-    return x;
-}
-
-static inline float16x8_t log_ps_f16(float16x8_t x)
-{
-    float16x8_t one = vdupq_n_f16(1);
-
-    x = vmaxq_f16(x, vdupq_n_f16(0)); /* force flush to zero on denormal values */
-    uint16x8_t invalid_mask = vcleq_f16(x, vdupq_n_f16(0));
-
-    int16x8_t ux = vreinterpretq_s16_f16(x);
-
-    int16x8_t emm0 = vshrq_n_s16(ux, 10);
-
-    /* keep only the fractional part */
-    ux = vandq_s16(ux, vdupq_n_s16(c_inv_mant_mask_f16));
-    ux = vorrq_s16(ux, vreinterpretq_s16_f16(vdupq_n_f16(0.5f)));
-    x = vreinterpretq_f16_s16(ux);
-
-    emm0 = vsubq_s16(emm0, vdupq_n_s16(0xf));
-    float16x8_t e = vcvtq_f16_s16(emm0);
-
-    e = vaddq_f16(e, one);
-
-    /* part2:
-     *     if( x < SQRTHF ) {
-     *       e -= 1;
-     *       x = x + x - 1.0;
-     *     } else { x = x - 1.0; }
-     */
-    uint16x8_t mask = vcltq_f16(x, vdupq_n_f16(c_cephes_SQRTHF));
-    float16x8_t tmp = vreinterpretq_f16_u16(vandq_u16(vreinterpretq_u16_f16(x), mask));
-    x = vsubq_f16(x, one);
-    e = vsubq_f16(e, vreinterpretq_f16_u16(vandq_u16(vreinterpretq_u16_f16(one), mask)));
-    x = vaddq_f16(x, tmp);
-
-    float16x8_t z = vmulq_f16(x, x);
-
-    float16x8_t y = vdupq_n_f16(c_cephes_log_p0);
-    y = vfmaq_f16(vdupq_n_f16(c_cephes_log_p1), y, x);
-    y = vfmaq_f16(vdupq_n_f16(c_cephes_log_p2), y, x);
-    y = vfmaq_f16(vdupq_n_f16(c_cephes_log_p3), y, x);
-    y = vfmaq_f16(vdupq_n_f16(c_cephes_log_p4), y, x);
-    y = vfmaq_f16(vdupq_n_f16(c_cephes_log_p5), y, x);
-    y = vfmaq_f16(vdupq_n_f16(c_cephes_log_p6), y, x);
-    y = vfmaq_f16(vdupq_n_f16(c_cephes_log_p7), y, x);
-    y = vfmaq_f16(vdupq_n_f16(c_cephes_log_p8), y, x);
-    y = vmulq_f16(y, x);
-
-    y = vmulq_f16(y, z);
-
-    y = vfmaq_f16(y, e, vdupq_n_f16(c_cephes_log_q1));
-
-    y = vfmsq_f16(y, z, vdupq_n_f16(0.5f));
-
-    x = vaddq_f16(x, y);
-    x = vfmaq_f16(x, e, vdupq_n_f16(c_cephes_log_q2));
-    x = vreinterpretq_f16_u16(vorrq_u16(vreinterpretq_u16_f16(x), invalid_mask)); // negative arg will be NAN
-    return x;
-}
-
-#define c_exp_hi_f16 10.7421875f
-#define c_exp_lo_f16 -10.7421875f
-
-#define c_cephes_LOG2EF 1.44269504088896341
-#define c_cephes_exp_C1 0.693359375
-#define c_cephes_exp_C2 -2.12194440e-4
-
-#define c_cephes_exp_p0 1.9875691500E-4
-#define c_cephes_exp_p1 1.3981999507E-3
-#define c_cephes_exp_p2 8.3334519073E-3
-#define c_cephes_exp_p3 4.1665795894E-2
-#define c_cephes_exp_p4 1.6666665459E-1
-#define c_cephes_exp_p5 5.0000001201E-1
-
-/* exp() computed for 4 float at once */
-static inline float16x4_t exp_ps_f16(float16x4_t x)
-{
-    float16x4_t tmp, fx;
-
-    float16x4_t one = vdup_n_f16(1);
-    x = vmin_f16(x, vdup_n_f16(c_exp_hi_f16));
-    x = vmax_f16(x, vdup_n_f16(c_exp_lo_f16));
-
-    /* express exp(x) as exp(g + n*log(2)) */
-#if _MSC_VER
-    fx = vfma_f16(vdup_n_f16(0.5f), x, vcvt_f16_f32(vdupq_n_f32(c_cephes_LOG2EF)));
-#else
-    fx = vfma_f16(vdup_n_f16(0.5f), x, vdup_n_f16(c_cephes_LOG2EF));
-#endif
-
-    /* perform a floorf */
-    tmp = vcvt_f16_s16(vcvt_s16_f16(fx));
-
-    /* if greater, substract 1 */
-    uint16x4_t mask = vcgt_f16(tmp, fx);
-    mask = vand_u16(mask, (uint16x4_t)(one));
-
-    fx = vsub_f16(tmp, (float16x4_t)(mask));
-
-#if _MSC_VER
-    tmp = vmul_f16(fx, vcvt_f16_f32(vdupq_n_f32(c_cephes_exp_C1)));
-    float16x4_t z = vmul_f16(fx, vcvt_f16_f32(vdupq_n_f32(c_cephes_exp_C2)));
-#else
-    tmp = vmul_f16(fx, vdup_n_f16(c_cephes_exp_C1));
-    float16x4_t z = vmul_f16(fx, vdup_n_f16(c_cephes_exp_C2));
-#endif
-    x = vsub_f16(x, tmp);
-    x = vsub_f16(x, z);
-
-    z = vmul_f16(x, x);
-
-    float16x4_t y = vdup_n_f16(c_cephes_exp_p0);
-    y = vfma_f16(vdup_n_f16(c_cephes_exp_p1), y, x);
-    y = vfma_f16(vdup_n_f16(c_cephes_exp_p2), y, x);
-    y = vfma_f16(vdup_n_f16(c_cephes_exp_p3), y, x);
-    y = vfma_f16(vdup_n_f16(c_cephes_exp_p4), y, x);
-    y = vfma_f16(vdup_n_f16(c_cephes_exp_p5), y, x);
-
-    y = vfma_f16(x, y, z);
-    y = vadd_f16(y, one);
-
-    /* build 2^n */
-    int16x4_t mm;
-    mm = vcvt_s16_f16(fx);
-    mm = vadd_s16(mm, vdup_n_s16(0xf));
-    mm = vshl_n_s16(mm, 10);
-    float16x4_t pow2n = vreinterpret_f16_s16(mm);
-
-    y = vmul_f16(y, pow2n);
-    return y;
-}
-
-static inline float16x8_t exp_ps_f16(float16x8_t x)
-{
-    float16x8_t tmp, fx;
-
-    float16x8_t one = vdupq_n_f16(1);
-    x = vminq_f16(x, vdupq_n_f16(c_exp_hi_f16));
-    x = vmaxq_f16(x, vdupq_n_f16(c_exp_lo_f16));
-
-    /* express exp(x) as exp(g + n*log(2)) */
-#if _MSC_VER
-    float16x4_t _c_cephes_LOG2EF = vcvt_f16_f32(vdupq_n_f32(c_cephes_LOG2EF));
-    fx = vfmaq_f16(vdupq_n_f16(0.5f), x, vcombine_f16(_c_cephes_LOG2EF, _c_cephes_LOG2EF));
-#else
-    fx = vfmaq_f16(vdupq_n_f16(0.5f), x, vdupq_n_f16(c_cephes_LOG2EF));
-#endif
-
-    /* perform a floorf */
-    tmp = vcvtq_f16_s16(vcvtq_s16_f16(fx));
-
-    /* if greater, substract 1 */
-    uint16x8_t mask = vcgtq_f16(tmp, fx);
-    mask = vandq_u16(mask, vreinterpretq_u16_f16(one));
-
-    fx = vsubq_f16(tmp, vreinterpretq_f16_u16(mask));
-
-#if _MSC_VER
-    float16x4_t _c_cephes_exp_C1 = vcvt_f16_f32(vdupq_n_f32(c_cephes_exp_C1));
-    tmp = vmulq_f16(fx, vcombine_f16(_c_cephes_exp_C1, _c_cephes_exp_C1));
-    float16x4_t _c_cephes_exp_C2 = vcvt_f16_f32(vdupq_n_f32(c_cephes_exp_C2));
-    float16x8_t z = vmulq_f16(fx, vcombine_f16(_c_cephes_exp_C2, _c_cephes_exp_C2));
-#else
-    tmp = vmulq_f16(fx, vdupq_n_f16(c_cephes_exp_C1));
-    float16x8_t z = vmulq_f16(fx, vdupq_n_f16(c_cephes_exp_C2));
-#endif
-    x = vsubq_f16(x, tmp);
-    x = vsubq_f16(x, z);
-
-    z = vmulq_f16(x, x);
-
-    float16x8_t y = vdupq_n_f16(c_cephes_exp_p0);
-    y = vfmaq_f16(vdupq_n_f16(c_cephes_exp_p1), y, x);
-    y = vfmaq_f16(vdupq_n_f16(c_cephes_exp_p2), y, x);
-    y = vfmaq_f16(vdupq_n_f16(c_cephes_exp_p3), y, x);
-    y = vfmaq_f16(vdupq_n_f16(c_cephes_exp_p4), y, x);
-    y = vfmaq_f16(vdupq_n_f16(c_cephes_exp_p5), y, x);
-
-    y = vfmaq_f16(x, y, z);
-    y = vaddq_f16(y, one);
-
-    /* build 2^n */
-    int16x8_t mm;
-    mm = vcvtq_s16_f16(fx);
-    mm = vaddq_s16(mm, vdupq_n_s16(0xf));
-    mm = vshlq_n_s16(mm, 10);
-    float16x8_t pow2n = vreinterpretq_f16_s16(mm);
-
-    y = vmulq_f16(y, pow2n);
-    return y;
-}
-
-#define c_minus_cephes_DP1 -0.78515625
-#define c_minus_cephes_DP2 -2.4187564849853515625e-4
-#define c_minus_cephes_DP3 -3.77489497744594108e-8
-#define c_sincof_p0        -1.9515295891E-4
-#define c_sincof_p1        8.3321608736E-3
-#define c_sincof_p2        -1.6666654611E-1
-#define c_coscof_p0        2.443315711809948E-005
-#define c_coscof_p1        -1.388731625493765E-003
-#define c_coscof_p2        4.166664568298827E-002
-#define c_cephes_FOPI      1.27323954473516 // 4 / M_PI
-
-/* evaluation of 4 sines & cosines at once.
- *
- *   The code is the exact rewriting of the cephes sinf function.
- *   Precision is excellent as long as x < 8192 (I did not bother to
- *   take into account the special handling they have for greater values
- *   -- it does not return garbage for arguments over 8192, though, but
- *   the extra precision is missing).
- *
- *   Note that it is such that sinf((float)M_PI) = 8.74e-8, which is the
- *   surprising but correct result.
- *
- *   Note also that when you compute sin(x), cos(x) is available at
- *   almost no extra price so both sin_ps and cos_ps make use of
- *   sincos_ps..
- */
-static inline void sincos_ps_f16(float16x4_t x, float16x4_t* ysin, float16x4_t* ycos)
-{
-    // any x
-    float16x4_t y;
-
-    uint16x4_t emm2;
-
-    uint16x4_t sign_mask_sin, sign_mask_cos;
-    sign_mask_sin = vclt_f16(x, vdup_n_f16(0));
-    x = vabs_f16(x);
-
-    /* scale by 4/Pi */
-#if _MSC_VER
-    float16x4_t _c_cephes_FOPI = vcvt_f16_f32(vdupq_n_f32(c_cephes_FOPI));
-    y = vmul_f16(x, _c_cephes_FOPI);
-#else
-    y = vmul_f16(x, vdup_n_f16(c_cephes_FOPI));
-#endif
-
-    /* store the integer part of y in mm0 */
-    emm2 = vcvt_u16_f16(y);
-    /* j=(j+1) & (~1) (see the cephes sources) */
-    emm2 = vadd_u16(emm2, vdup_n_u16(1));
-    emm2 = vand_u16(emm2, vdup_n_u16(~1));
-    y = vcvt_f16_u16(emm2);
-
-    /* get the polynom selection mask
-     *     there is one polynom for 0 <= x <= Pi/4
-     *     and another one for Pi/4<x<=Pi/2
-     *
-     *     Both branches will be computed.
-     */
-    uint16x4_t poly_mask = vtst_u16(emm2, vdup_n_u16(2));
-
-    /* The magic pass: "Extended precision modular arithmetic"
-     *     x = ((x - y * DP1) - y * DP2) - y * DP3; */
-#if _MSC_VER
-    float16x4_t _c_minus_cephes_DP1 = vcvt_f16_f32(vdupq_n_f32(c_minus_cephes_DP1));
-    float16x4_t _c_minus_cephes_DP2 = vcvt_f16_f32(vdupq_n_f32(c_minus_cephes_DP2));
-    float16x4_t _c_minus_cephes_DP3 = vcvt_f16_f32(vdupq_n_f32(c_minus_cephes_DP3));
-    x = vfma_f16(x, y, _c_minus_cephes_DP1);
-    x = vfma_f16(x, y, _c_minus_cephes_DP2);
-    x = vfma_f16(x, y, _c_minus_cephes_DP3);
-#else
-    x = vfma_f16(x, y, vdup_n_f16(c_minus_cephes_DP1));
-    x = vfma_f16(x, y, vdup_n_f16(c_minus_cephes_DP2));
-    x = vfma_f16(x, y, vdup_n_f16(c_minus_cephes_DP3));
-#endif
-
-    sign_mask_sin = veor_u16(sign_mask_sin, vtst_u16(emm2, vdup_n_u16(4)));
-    sign_mask_cos = vtst_u16(vsub_u16(emm2, vdup_n_u16(2)), vdup_n_u16(4));
-
-    /* Evaluate the first polynom  (0 <= x <= Pi/4) in y1,
-     *     and the second polynom      (Pi/4 <= x <= 0) in y2 */
-    float16x4_t z = vmul_f16(x, x);
-    float16x4_t y1, y2;
-
-    y1 = vfma_f16(vdup_n_f16(c_coscof_p1), z, vdup_n_f16(c_coscof_p0));
-    y2 = vfma_f16(vdup_n_f16(c_sincof_p1), z, vdup_n_f16(c_sincof_p0));
-    y1 = vfma_f16(vdup_n_f16(c_coscof_p2), y1, z);
-    y2 = vfma_f16(vdup_n_f16(c_sincof_p2), y2, z);
-    y1 = vmul_f16(y1, z);
-    y2 = vmul_f16(y2, z);
-    y1 = vmul_f16(y1, z);
-    y1 = vfms_f16(y1, z, vdup_n_f16(0.5f));
-    y2 = vfma_f16(x, y2, x);
-    y1 = vadd_f16(y1, vdup_n_f16(1));
-
-    /* select the correct result from the two polynoms */
-    float16x4_t ys = vbsl_f16(poly_mask, y1, y2);
-    float16x4_t yc = vbsl_f16(poly_mask, y2, y1);
-    *ysin = vbsl_f16(sign_mask_sin, vneg_f16(ys), ys);
-    *ycos = vbsl_f16(sign_mask_cos, yc, vneg_f16(yc));
-}
-
-static inline void sincos_ps_f16(float16x8_t x, float16x8_t* ysin, float16x8_t* ycos)
-{
-    // any x
-    float16x8_t y;
-
-    uint16x8_t emm2;
-
-    uint16x8_t sign_mask_sin, sign_mask_cos;
-    sign_mask_sin = vcltq_f16(x, vdupq_n_f16(0));
-    x = vabsq_f16(x);
-
-    /* scale by 4/Pi */
-#if _MSC_VER
-    float16x4_t _c_cephes_FOPI = vcvt_f16_f32(vdupq_n_f32(c_cephes_FOPI));
-    y = vmulq_f16(x, vcombine_f16(_c_cephes_FOPI, _c_cephes_FOPI));
-#else
-    y = vmulq_f16(x, vdupq_n_f16(c_cephes_FOPI));
-#endif
-
-    /* store the integer part of y in mm0 */
-    emm2 = vcvtq_u16_f16(y);
-    /* j=(j+1) & (~1) (see the cephes sources) */
-    emm2 = vaddq_u16(emm2, vdupq_n_u16(1));
-    emm2 = vandq_u16(emm2, vdupq_n_u16(~1));
-    y = vcvtq_f16_u16(emm2);
-
-    /* get the polynom selection mask
-     *     there is one polynom for 0 <= x <= Pi/4
-     *     and another one for Pi/4<x<=Pi/2
-     *
-     *     Both branches will be computed.
-     */
-    uint16x8_t poly_mask = vtstq_u16(emm2, vdupq_n_u16(2));
-
-    /* The magic pass: "Extended precision modular arithmetic"
-     *     x = ((x - y * DP1) - y * DP2) - y * DP3; */
-#if _MSC_VER
-    float16x4_t _c_minus_cephes_DP1 = vcvt_f16_f32(vdupq_n_f32(c_minus_cephes_DP1));
-    float16x4_t _c_minus_cephes_DP2 = vcvt_f16_f32(vdupq_n_f32(c_minus_cephes_DP2));
-    float16x4_t _c_minus_cephes_DP3 = vcvt_f16_f32(vdupq_n_f32(c_minus_cephes_DP3));
-    x = vfmaq_f16(x, y, vcombine_f16(_c_minus_cephes_DP1, _c_minus_cephes_DP1));
-    x = vfmaq_f16(x, y, vcombine_f16(_c_minus_cephes_DP2, _c_minus_cephes_DP2));
-    x = vfmaq_f16(x, y, vcombine_f16(_c_minus_cephes_DP3, _c_minus_cephes_DP3));
-#else
-    x = vfmaq_f16(x, y, vdupq_n_f16(c_minus_cephes_DP1));
-    x = vfmaq_f16(x, y, vdupq_n_f16(c_minus_cephes_DP2));
-    x = vfmaq_f16(x, y, vdupq_n_f16(c_minus_cephes_DP3));
-#endif
-
-    sign_mask_sin = veorq_u16(sign_mask_sin, vtstq_u16(emm2, vdupq_n_u16(4)));
-    sign_mask_cos = vtstq_u16(vsubq_u16(emm2, vdupq_n_u16(2)), vdupq_n_u16(4));
-
-    /* Evaluate the first polynom  (0 <= x <= Pi/4) in y1,
-     *     and the second polynom      (Pi/4 <= x <= 0) in y2 */
-    float16x8_t z = vmulq_f16(x, x);
-    float16x8_t y1, y2;
-
-    y1 = vfmaq_f16(vdupq_n_f16(c_coscof_p1), z, vdupq_n_f16(c_coscof_p0));
-    y2 = vfmaq_f16(vdupq_n_f16(c_sincof_p1), z, vdupq_n_f16(c_sincof_p0));
-    y1 = vfmaq_f16(vdupq_n_f16(c_coscof_p2), y1, z);
-    y2 = vfmaq_f16(vdupq_n_f16(c_sincof_p2), y2, z);
-    y1 = vmulq_f16(y1, z);
-    y2 = vmulq_f16(y2, z);
-    y1 = vmulq_f16(y1, z);
-    y1 = vfmsq_f16(y1, z, vdupq_n_f16(0.5f));
-    y2 = vfmaq_f16(x, y2, x);
-    y1 = vaddq_f16(y1, vdupq_n_f16(1));
-
-    /* select the correct result from the two polynoms */
-    float16x8_t ys = vbslq_f16(poly_mask, y1, y2);
-    float16x8_t yc = vbslq_f16(poly_mask, y2, y1);
-    *ysin = vbslq_f16(sign_mask_sin, vnegq_f16(ys), ys);
-    *ycos = vbslq_f16(sign_mask_cos, yc, vnegq_f16(yc));
-}
-
-static inline float16x4_t sin_ps_f16(float16x4_t x)
-{
-    float16x4_t ysin, ycos;
-    sincos_ps_f16(x, &ysin, &ycos);
-    return ysin;
-}
-
-static inline float16x8_t sin_ps_f16(float16x8_t x)
-{
-    float16x8_t ysin, ycos;
-    sincos_ps_f16(x, &ysin, &ycos);
-    return ysin;
-}
-
-static inline float16x4_t cos_ps_f16(float16x4_t x)
-{
-    float16x4_t ysin, ycos;
-    sincos_ps_f16(x, &ysin, &ycos);
-    return ycos;
-}
-
-static inline float16x8_t cos_ps_f16(float16x8_t x)
-{
-    float16x8_t ysin, ycos;
-    sincos_ps_f16(x, &ysin, &ycos);
-    return ycos;
-}
-
-#define c_tanh_tiny 1e-4f
-#define c_tanh_hi   9.0f
-// The monomial coefficients of the numerator polynomial (odd).
-#define c_tanh_alpha_1  4.89352455891786e-3f
-#define c_tanh_alpha_3  6.37261928875436e-4f
-#define c_tanh_alpha_5  1.48572235717979e-5f
-#define c_tanh_alpha_7  5.12229709037114e-8f
-#define c_tanh_alpha_9  -8.60467152213735e-11f
-#define c_tanh_alpha_11 2.00018790482477e-13f
-#define c_tanh_alpha_13 -2.76076847742355e-16f
-// The monomial coefficients of the denominator polynomial (even).
-#define c_tanh_beta_0 4.89352518554385e-3f
-#define c_tanh_beta_2 2.26843463243900e-3f
-#define c_tanh_beta_4 1.18534705686654e-4f
-#define c_tanh_beta_6 1.19825839466702e-6f
-
-/* Single precision hyperbolic tangent computed for 4 simultaneous float */
-static inline float16x4_t tanh_ps_f16(float16x4_t x)
-{
-    float16x4_t x2 = vabs_f16(x);
-
-    uint16x4_t tiny_mask = vcge_f16(x2, vdup_n_f16(c_tanh_tiny));
-
-    // clamp the inputs to the range [-9, 9] since anything outside
-    // this range is -/+1.0f in single-precision.
-    x2 = (float16x4_t)(vbsl_u16(vcge_f16(vdup_n_f16(c_tanh_hi), x2), (uint16x4_t)(x2), (uint16x4_t)(vdup_n_f16(c_tanh_hi))));
-
-    // since the polynomials are odd/even, we need x**2.
-    float16x4_t z = vmul_f16(x2, x2);
-
-    // evaluate the numerator polynomial y.
-    float16x4_t y = vdup_n_f16(c_tanh_alpha_13);
-    y = vfma_f16(vdup_n_f16(c_tanh_alpha_11), y, z);
-    y = vfma_f16(vdup_n_f16(c_tanh_alpha_9), y, z);
-    y = vfma_f16(vdup_n_f16(c_tanh_alpha_7), y, z);
-    y = vfma_f16(vdup_n_f16(c_tanh_alpha_5), y, z);
-    y = vfma_f16(vdup_n_f16(c_tanh_alpha_3), y, z);
-    y = vfma_f16(vdup_n_f16(c_tanh_alpha_1), y, z);
-    y = vmul_f16(y, x2);
-
-    // evaluate the denominator polynomial w.
-    float16x4_t w = vdup_n_f16(c_tanh_beta_6);
-    w = vfma_f16(vdup_n_f16(c_tanh_beta_4), w, z);
-    w = vfma_f16(vdup_n_f16(c_tanh_beta_2), w, z);
-    w = vfma_f16(vdup_n_f16(c_tanh_beta_0), w, z);
-
-    // divide the numerator by the denominator.
-    y = vdiv_f16(y, w);
-
-    // reinstate the sign.
-    y = (float16x4_t)(vbsl_u16(vdup_n_u16(1u << 15), (uint16x4_t)(x), (uint16x4_t)(y)));
-
-    // when the argument is very small in magnitude it's more accurate to just return it.
-    y = (float16x4_t)(vbsl_u16(tiny_mask, (uint16x4_t)(y), (uint16x4_t)(x)));
-
-    return y;
-}
-
-static inline float16x8_t tanh_ps_f16(float16x8_t x)
-{
-    float16x8_t x2 = vabsq_f16(x);
-
-    uint16x8_t tiny_mask = vcgeq_f16(x2, vdupq_n_f16(c_tanh_tiny));
-
-    // clamp the inputs to the range [-9, 9] since anything outside
-    // this range is -/+1.0f in single-precision.
-    x2 = vreinterpretq_f16_u16(vbslq_u16(vcgeq_f16(vdupq_n_f16(c_tanh_hi), x2), vreinterpretq_u16_f16(x2), vreinterpretq_u16_f16(vdupq_n_f16(c_tanh_hi))));
-
-    // since the polynomials are odd/even, we need x**2.
-    float16x8_t z = vmulq_f16(x2, x2);
-
-    // evaluate the numerator polynomial y.
-    float16x8_t y = vdupq_n_f16(c_tanh_alpha_13);
-    y = vfmaq_f16(vdupq_n_f16(c_tanh_alpha_11), y, z);
-    y = vfmaq_f16(vdupq_n_f16(c_tanh_alpha_9), y, z);
-    y = vfmaq_f16(vdupq_n_f16(c_tanh_alpha_7), y, z);
-    y = vfmaq_f16(vdupq_n_f16(c_tanh_alpha_5), y, z);
-    y = vfmaq_f16(vdupq_n_f16(c_tanh_alpha_3), y, z);
-    y = vfmaq_f16(vdupq_n_f16(c_tanh_alpha_1), y, z);
-    y = vmulq_f16(y, x2);
-
-    // evaluate the denominator polynomial w.
-    float16x8_t w = vdupq_n_f16(c_tanh_beta_6);
-    w = vfmaq_f16(vdupq_n_f16(c_tanh_beta_4), w, z);
-    w = vfmaq_f16(vdupq_n_f16(c_tanh_beta_2), w, z);
-    w = vfmaq_f16(vdupq_n_f16(c_tanh_beta_0), w, z);
-
-    // divide the numerator by the denominator.
-    y = vdivq_f16(y, w);
-
-    // reinstate the sign.
-    y = vreinterpretq_f16_u16(vbslq_u16(vdupq_n_u16(1u << 15), vreinterpretq_u16_f16(x), vreinterpretq_u16_f16(y)));
-
-    // when the argument is very small in magnitude it's more accurate to just return it.
-    y = vreinterpretq_f16_u16(vbslq_u16(tiny_mask, vreinterpretq_u16_f16(y), vreinterpretq_u16_f16(x)));
-
-    return y;
-}
-
-static inline float16x4_t sigmoid_ps_f16(float16x4_t _v)
-{
-    float16x4_t _one = vdup_n_f16(1.f);
-    _v = vneg_f16(_v);
-    _v = exp_ps_f16(_v);
-    _v = vadd_f16(_v, _one);
-    return vdiv_f16(_one, _v);
-}
-
-static inline float16x8_t sigmoid_ps_f16(float16x8_t _v)
-{
-    float16x8_t _one = vdupq_n_f16(1.f);
-    _v = vnegq_f16(_v);
-    _v = exp_ps_f16(_v);
-    _v = vaddq_f16(_v, _one);
-    return vdivq_f16(_one, _v);
-}
-
-#endif // NEON_MATHFUN_FP16S_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/neon_mathfun_tanh.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/neon_mathfun_tanh.h
deleted file mode 100644
index 1f9073dd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/neon_mathfun_tanh.h
+++ /dev/null
@@ -1,82 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NEON_MATHFUN_TANH_H
-#define NEON_MATHFUN_TANH_H
-
-#include <arm_neon.h>
-
-#define c_tanh_tiny 1e-4f
-#define c_tanh_hi   9.0f
-// The monomial coefficients of the numerator polynomial (odd).
-#define c_tanh_alpha_1  4.89352455891786e-3f
-#define c_tanh_alpha_3  6.37261928875436e-4f
-#define c_tanh_alpha_5  1.48572235717979e-5f
-#define c_tanh_alpha_7  5.12229709037114e-8f
-#define c_tanh_alpha_9  -8.60467152213735e-11f
-#define c_tanh_alpha_11 2.00018790482477e-13f
-#define c_tanh_alpha_13 -2.76076847742355e-16f
-// The monomial coefficients of the denominator polynomial (even).
-#define c_tanh_beta_0 4.89352518554385e-3f
-#define c_tanh_beta_2 2.26843463243900e-3f
-#define c_tanh_beta_4 1.18534705686654e-4f
-#define c_tanh_beta_6 1.19825839466702e-6f
-
-/* Single precision hyperbolic tangent computed for 4 simultaneous float */
-static inline float32x4_t tanh_ps(float32x4_t x)
-{
-    float32x4_t x2 = vabsq_f32(x);
-
-    uint32x4_t tiny_mask = vcgeq_f32(x2, vdupq_n_f32(c_tanh_tiny));
-
-    // clamp the inputs to the range [-9, 9] since anything outside
-    // this range is -/+1.0f in single-precision.
-    x2 = vreinterpretq_f32_u32(vbslq_u32(vcgeq_f32(vdupq_n_f32(c_tanh_hi), x2), vreinterpretq_u32_f32(x2), vreinterpretq_u32_f32(vdupq_n_f32(c_tanh_hi))));
-
-    // since the polynomials are odd/even, we need x**2.
-    float32x4_t z = vmulq_f32(x2, x2);
-
-    // evaluate the numerator polynomial y.
-    float32x4_t y = vdupq_n_f32(c_tanh_alpha_13);
-    y = vmlaq_f32(vdupq_n_f32(c_tanh_alpha_11), y, z);
-    y = vmlaq_f32(vdupq_n_f32(c_tanh_alpha_9), y, z);
-    y = vmlaq_f32(vdupq_n_f32(c_tanh_alpha_7), y, z);
-    y = vmlaq_f32(vdupq_n_f32(c_tanh_alpha_5), y, z);
-    y = vmlaq_f32(vdupq_n_f32(c_tanh_alpha_3), y, z);
-    y = vmlaq_f32(vdupq_n_f32(c_tanh_alpha_1), y, z);
-    y = vmulq_f32(y, x2);
-
-    // evaluate the denominator polynomial w.
-    float32x4_t w = vdupq_n_f32(c_tanh_beta_6);
-    w = vmlaq_f32(vdupq_n_f32(c_tanh_beta_4), w, z);
-    w = vmlaq_f32(vdupq_n_f32(c_tanh_beta_2), w, z);
-    w = vmlaq_f32(vdupq_n_f32(c_tanh_beta_0), w, z);
-
-    // divide the numerator by the denominator.
-#if __aarch64__
-    y = vdivq_f32(y, w);
-#else
-    y = div_ps(y, w);
-#endif
-
-    // reinstate the sign.
-    y = vreinterpretq_f32_u32(vbslq_u32(vdupq_n_u32(1u << 31), vreinterpretq_u32_f32(x), vreinterpretq_u32_f32(y)));
-
-    // when the argument is very small in magnitude it's more accurate to just return it.
-    y = vreinterpretq_f32_u32(vbslq_u32(tiny_mask, vreinterpretq_u32_f32(y), vreinterpretq_u32_f32(x)));
-
-    return y;
-}
-
-#endif // NEON_MATHFUN_TANH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/packing_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/packing_arm.h
deleted file mode 100644
index 17c64854..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/packing_arm.h
+++ /dev/null
@@ -1,36 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PACKING_ARM_H
-#define LAYER_PACKING_ARM_H
-
-#include "packing.h"
-
-namespace ncnn {
-
-class Packing_arm : public Packing
-{
-public:
-    Packing_arm();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_bf16s_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PACKING_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_arm.h
deleted file mode 100644
index 164cfe4c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_arm.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PADDING_ARM_H
-#define LAYER_PADDING_ARM_H
-
-#include "padding.h"
-
-namespace ncnn {
-
-class Padding_arm : public Padding
-{
-public:
-    Padding_arm();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_bf16s_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-#if NCNN_BF16
-    // bf16
-    unsigned short value_bf16;
-    Mat per_channel_pad_data_bf16;
-#endif
-
-    // fp16
-    unsigned short value_fp16;
-    Mat per_channel_pad_data_fp16;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PADDING_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack4.h
deleted file mode 100644
index d52e991b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack4.h
+++ /dev/null
@@ -1,562 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void padding_constant_pack4_neon(const Mat& src, Mat& dst, int top, int bottom, int left, int right, float32x4_t v)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-
-    int w = src.w;
-    int h = src.h;
-
-    int top_size = top * dst.w;
-    int bottom_size = bottom * dst.w;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-    asm volatile(
-        "mov    v0.16b, %10.16b         \n"
-        "mov    v1.16b, %10.16b         \n"
-        "mov    v2.16b, %10.16b         \n"
-        "mov    v3.16b, %10.16b         \n"
-
-        // fill top
-        "lsr    w4, %w8, #3             \n" // w4 = nn = top_size >> 3
-        "cmp    w4, #0                  \n"
-        "beq    1f                      \n"
-
-        "0:                             \n"
-        "st1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-        "subs   w4, w4, #1              \n"
-        "st1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-        "bne    0b                      \n"
-
-        "1:                             \n"
-
-        // fill top remain
-        "and    w4, %w8, #7             \n" // w4 = remain = top_size & 7
-
-        "cmp    w4, #4                  \n" // w4 >= 4
-        "blt    2f                      \n"
-        "sub    w4, w4, #4              \n"
-        "st1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-        "2:                             \n"
-
-        "cmp    w4, #2                  \n" // w4 >= 2
-        "blt    3f                      \n"
-        "sub    w4, w4, #2              \n"
-        "st1    {v0.4s, v1.4s}, [%0], #32 \n"
-        "3:                             \n"
-
-        "cmp    w4, #0                  \n" // w4 > 0
-        "beq    4f                      \n"
-        "st1    {v0.4s}, [%0], #16      \n"
-        "4:                             \n"
-
-        // fill center h loop
-        "cmp    %w5, #0                 \n"
-        "beq    15f                     \n"
-        "5:                             \n"
-
-        // fill left
-        "mov    w4, %w6                 \n" // w4 = left
-        "cmp    w4, #0                  \n"
-        "beq    7f                      \n"
-
-        "6:                             \n"
-        "st1    {v0.4s}, [%0], #16      \n"
-        "subs   w4, w4, #1              \n"
-        "bne    6b                      \n"
-
-        "7:                             \n"
-
-        // fill middle
-        "lsr    w4, %w4, #3             \n" // w4 = nn = w >> 3
-        "cmp    w4, #0                  \n"
-        "beq    9f                      \n"
-
-        "8:                             \n"
-        "prfm   pldl1keep, [%1, #512]   \n"
-        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%1], #64 \n"
-        "prfm   pldl1keep, [%1, #512]   \n"
-        "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%1], #64 \n"
-        "subs   w4, w4, #1              \n"
-        "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-        "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-        "bne    8b                      \n"
-
-        "9:                             \n"
-
-        "and    w4, %w4, #7             \n" // w4 = remain = w & 7
-
-        "cmp    w4, #4                  \n" // w4 >= 4
-        "blt    10f                     \n"
-        "prfm   pldl1keep, [%1, #512]   \n"
-        "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%1], #64 \n"
-        "sub    w4, w4, #4              \n"
-        "st1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%0], #64 \n"
-        "10:                            \n"
-
-        "cmp    w4, #2                  \n" // w4 >= 2
-        "blt    11f                     \n"
-        "prfm   pldl1keep, [%1, #256]   \n"
-        "ld1    {v16.4s, v17.4s}, [%1], #32 \n"
-        "sub    w4, w4, #2              \n"
-        "st1    {v16.4s, v17.4s}, [%0], #32 \n"
-        "11:                            \n"
-
-        "cmp    w4, #0                  \n" // w4 > 0
-        "beq    12f                     \n"
-        "prfm   pldl1keep, [%1, #128]   \n"
-        "ld1    {v16.4s}, [%1], #16     \n"
-        "st1    {v16.4s}, [%0], #16     \n"
-        "12:                            \n"
-
-        // fill right
-        "mov    w4, %w7                 \n" // w4 = right
-        "cmp    w4, #0                  \n"
-        "beq    14f                     \n"
-
-        "13:                            \n"
-        "subs   w4, w4, #1              \n"
-        "st1    {v0.4s}, [%0], #16      \n"
-        "bne    13b                     \n"
-        "14:                            \n"
-
-        "subs   %w5, %w5, #1            \n"
-        "bne    5b                      \n"
-
-        "15:                            \n"
-
-        // fill bottom
-        "lsr    w4, %w9, #3             \n" // w4 = nn = bottom_size >> 3
-        "cmp    w4, #0                  \n"
-        "beq    17f                     \n"
-
-        "16:                            \n"
-        "st1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-        "subs   w4, w4, #1              \n"
-        "st1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-        "bne    16b                     \n"
-        "17:                            \n"
-
-        // fill bottom remain
-        "and    w4, %w9, #7             \n" // w4 = remain = bottom_size & 7
-
-        "cmp    w4, #4                  \n" // w4 >= 4
-        "blt    18f                     \n"
-        "sub    w4, w4, #4              \n"
-        "st1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-        "18:                            \n"
-
-        "cmp    w4, #2                  \n" // w4 >= 2
-        "blt    19f                     \n"
-        "sub    w4, w4, #2              \n"
-        "st1    {v0.4s, v1.4s}, [%0], #32 \n"
-        "19:                            \n"
-
-        "cmp    w4, #0                  \n" // w4 > 0
-        "beq    20f                     \n"
-        "st1    {v0.4s}, [%0], #16      \n"
-        "20:                            \n"
-
-        : "=r"(outptr), // %0
-        "=r"(ptr)     // %1
-        : "0"(outptr),
-        "1"(ptr),
-        "r"(w),           // %4
-        "r"(h),           // %5
-        "r"(left),        // %6
-        "r"(right),       // %7
-        "r"(top_size),    // %8
-        "r"(bottom_size), // %9
-        "w"(v)            // %10
-        : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-#else  // __aarch64__
-    asm volatile(
-        "vmov       q0, %q10            \n"
-        "vmov       q1, %q10            \n"
-        "vmov       q2, %q10            \n"
-        "vmov       q3, %q10            \n"
-
-        // fill top
-        "lsr        r4, %8, #3          \n" // r4 = nn = top_size >> 3
-        "cmp        r4, #0              \n"
-        "beq        1f                  \n"
-
-        "0:                             \n"
-        "vstm       %0!, {d0-d7}        \n"
-        "subs       r4, r4, #1          \n"
-        "vstm       %0!, {d0-d7}        \n"
-        "bne        0b                  \n"
-
-        "1:                             \n"
-
-        // fill top remain
-        "and        r4, %8, #7          \n" // r4 = remain = top_size & 7
-
-        "cmp        r4, #4              \n" // r4 >= 4
-        "blt        2f                  \n"
-        "sub        r4, r4, #4          \n"
-        "vstm       %0!, {d0-d7}        \n"
-        "2:                             \n"
-
-        "cmp        r4, #2              \n" // r4 >= 2
-        "blt        3f                  \n"
-        "sub        r4, r4, #2          \n"
-        "vst1.f32   {d0-d3}, [%0 :128]! \n"
-        "3:                             \n"
-
-        "cmp        r4, #0              \n" // r4 > 0
-        "beq        4f                  \n"
-        "vst1.f32   {d0-d1}, [%0 :128]! \n"
-        "4:                             \n"
-
-        // fill center h loop
-        "cmp        %5, #0              \n"
-        "beq        15f                 \n"
-        "5:                             \n"
-
-        // fill left
-        "mov        r4, %6              \n" // r4 = left
-        "cmp        r4, #0              \n"
-        "beq        7f                  \n"
-
-        "6:                             \n"
-        "vst1.f32   {d0-d1}, [%0 :128]! \n"
-        "subs       r4, r4, #1          \n"
-        "bne        6b                  \n"
-
-        "7:                             \n"
-
-        // fill middle
-        "lsr        r4, %4, #3          \n" // r4 = nn = w >> 3
-        "cmp        r4, #0              \n"
-        "beq        9f                  \n"
-
-        "8:                             \n"
-        "pld        [%1, #512]          \n"
-        "vldm       %1!, {d16-d23}      \n"
-        "pld        [%1, #512]          \n"
-        "vldm       %1!, {d24-d31}      \n"
-        "subs       r4, r4, #1          \n"
-        "vstm       %0!, {d16-d23}      \n"
-        "vstm       %0!, {d24-d31}      \n"
-        "bne        8b                  \n"
-
-        "9:                             \n"
-
-        "and        r4, %4, #7          \n" // r4 = remain = w & 7
-
-        "cmp        r4, #4              \n" // r4 >= 4
-        "blt        10f                 \n"
-        "pld        [%1, #512]          \n"
-        "vldm       %1!, {d16-d23}      \n"
-        "sub        r4, r4, #4          \n"
-        "vstm       %0!, {d16-d23}      \n"
-        "10:                            \n"
-
-        "cmp        r4, #2              \n" // r4 >= 2
-        "blt        11f                 \n"
-        "pld        [%1, #256]          \n"
-        "vld1.f32   {d16-d19}, [%1 :128]! \n"
-        "sub        r4, r4, #2          \n"
-        "vst1.f32   {d16-d19}, [%0 :128]! \n"
-        "11:                            \n"
-
-        "cmp        r4, #0              \n" // r4 > 0
-        "beq        12f                 \n"
-        "pld        [%1, #128]          \n"
-        "vld1.f32   {d16-d17}, [%1 :128]! \n"
-        "vst1.f32   {d16-d17}, [%0 :128]! \n"
-        "12:                            \n"
-
-        // fill right
-        "mov        r4, %7              \n" // r4 = right
-        "cmp        r4, #0              \n"
-        "beq        14f                 \n"
-
-        "13:                            \n"
-        "subs       r4, r4, #1          \n"
-        "vst1.f32   {d0-d1}, [%0 :128]! \n"
-        "bne        13b                 \n"
-        "14:                            \n"
-
-        "subs       %5, %5, #1          \n"
-        "bne        5b                  \n"
-
-        "15:                            \n"
-
-        // fill bottom
-        "lsr        r4, %9, #3          \n" // r4 = nn = bottom_size >> 3
-        "cmp        r4, #0              \n"
-        "beq        17f                 \n"
-
-        "16:                            \n"
-        "vstm       %0!, {d0-d7}        \n"
-        "subs       r4, r4, #1          \n"
-        "vstm       %0!, {d0-d7}        \n"
-        "bne        16b                 \n"
-        "17:                            \n"
-
-        // fill bottom remain
-        "and        r4, %9, #7          \n" // r4 = remain = bottom_size & 7
-
-        "cmp        r4, #4              \n" // r4 >= 4
-        "blt        18f                 \n"
-        "sub        r4, r4, #4          \n"
-        "vstm       %0!, {d0-d7}        \n"
-        "18:                            \n"
-
-        "cmp        r4, #2              \n" // r4 >= 2
-        "blt        19f                 \n"
-        "sub        r4, r4, #2          \n"
-        "vst1.f32   {d0-d3}, [%0 :128]! \n"
-        "19:                            \n"
-
-        "cmp        r4, #0              \n" // r4 > 0
-        "beq        20f                 \n"
-        "vst1.f32   {d0-d1}, [%0 :128]! \n"
-        "20:                            \n"
-
-        : "=r"(outptr), // %0
-        "=r"(ptr)     // %1
-        : "0"(outptr),
-        "1"(ptr),
-        "r"(w),           // %4
-        "r"(h),           // %5
-        "r"(left),        // %6
-        "r"(right),       // %7
-        "r"(top_size),    // %8
-        "r"(bottom_size), // %9
-        "w"(v)            // %10
-        : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-
-    // fill top
-    {
-        int x = 0;
-        for (; x + 3 < top_size; x += 4)
-        {
-            vst1q_f32(outptr, v);
-            vst1q_f32(outptr + 4, v);
-            vst1q_f32(outptr + 8, v);
-            vst1q_f32(outptr + 12, v);
-            outptr += 16;
-        }
-        for (; x < top_size; x++)
-        {
-            vst1q_f32(outptr, v);
-            outptr += 4;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            vst1q_f32(outptr, v);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            float32x4_t _p = vld1q_f32(ptr);
-            vst1q_f32(outptr, _p);
-            ptr += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1q_f32(outptr, v);
-            outptr += 4;
-        }
-    }
-    // fill bottom
-    {
-        int x = 0;
-        for (; x + 3 < bottom_size; x += 4)
-        {
-            vst1q_f32(outptr, v);
-            vst1q_f32(outptr + 4, v);
-            vst1q_f32(outptr + 8, v);
-            vst1q_f32(outptr + 12, v);
-            outptr += 16;
-        }
-        for (; x < bottom_size; x++)
-        {
-            vst1q_f32(outptr, v);
-            outptr += 4;
-        }
-    }
-#endif // NCNN_GNU_INLINE_ASM
-}
-
-static void padding_replicate_pack4_neon(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-
-    // fill top
-    for (int y = 0; y < top; y++)
-    {
-        const float* ptr0 = ptr;
-        float32x4_t _p = vld1q_f32(ptr0);
-        for (int x = 0; x < left; x++)
-        {
-            vst1q_f32(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = vld1q_f32(ptr0);
-            vst1q_f32(outptr, _p);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1q_f32(outptr, _p);
-            outptr += 4;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        float32x4_t _p = vld1q_f32(ptr);
-        for (int x = 0; x < left; x++)
-        {
-            vst1q_f32(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = vld1q_f32(ptr);
-            vst1q_f32(outptr, _p);
-            ptr += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1q_f32(outptr, _p);
-            outptr += 4;
-        }
-    }
-    // fill bottom
-    ptr -= src.w * 4;
-    for (int y = 0; y < bottom; y++)
-    {
-        const float* ptr0 = ptr;
-        float32x4_t _p = vld1q_f32(ptr0);
-        for (int x = 0; x < left; x++)
-        {
-            vst1q_f32(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = vld1q_f32(ptr0);
-            vst1q_f32(outptr, _p);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1q_f32(outptr, _p);
-            outptr += 4;
-        }
-    }
-}
-
-static void padding_reflect_pack4_neon(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-
-    // fill top
-    ptr += top * src.w * 4;
-    for (int y = 0; y < top; y++)
-    {
-        const float* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            float32x4_t _p = vld1q_f32(ptr0 + (left - x) * 4);
-            vst1q_f32(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            float32x4_t _p = vld1q_f32(ptr0);
-            vst1q_f32(outptr, _p);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            float32x4_t _p = vld1q_f32(ptr0 - 8 - x * 4);
-            vst1q_f32(outptr, _p);
-            outptr += 4;
-        }
-        ptr -= src.w * 4;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            float32x4_t _p = vld1q_f32(ptr + (left - x) * 4);
-            vst1q_f32(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            float32x4_t _p = vld1q_f32(ptr);
-            vst1q_f32(outptr, _p);
-            ptr += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            float32x4_t _p = vld1q_f32(ptr - 8 - x * 4);
-            vst1q_f32(outptr, _p);
-            outptr += 4;
-        }
-    }
-    // fill bottom
-    ptr -= 2 * src.w * 4;
-    for (int y = 0; y < bottom; y++)
-    {
-        const float* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            float32x4_t _p = vld1q_f32(ptr0 + (left - x) * 4);
-            vst1q_f32(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            float32x4_t _p = vld1q_f32(ptr0);
-            vst1q_f32(outptr, _p);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            float32x4_t _p = vld1q_f32(ptr0 - 8 - x * 4);
-            vst1q_f32(outptr, _p);
-            outptr += 4;
-        }
-        ptr -= src.w * 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack4_bf16s_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack4_bf16s_fp16s.h
deleted file mode 100644
index 3ca2b7fe..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack4_bf16s_fp16s.h
+++ /dev/null
@@ -1,548 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void padding_constant_pack4_bf16_fp16s_neon(const Mat& src, Mat& dst, int top, int bottom, int left, int right, uint16x8_t v)
-{
-    const unsigned short* ptr = src;
-    unsigned short* outptr = dst;
-
-    int w = src.w;
-    int h = src.h;
-
-    int top_size = top * dst.w;
-    int bottom_size = bottom * dst.w;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-    asm volatile(
-        "mov    v0.16b, %10.16b         \n"
-        "mov    v1.16b, %10.16b         \n"
-        "mov    v2.16b, %10.16b         \n"
-        "mov    v3.16b, %10.16b         \n"
-
-        // fill top
-        "lsr    w4, %w8, #3             \n" // w4 = nn = top_size >> 3
-        "cmp    w4, #0                  \n"
-        "beq    1f                      \n"
-
-        "0:                             \n"
-        "st1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0], #64 \n"
-        "subs   w4, w4, #1              \n"
-        "bne    0b                      \n"
-
-        "1:                             \n"
-
-        // fill top remain
-        "and    w4, %w8, #7             \n" // w4 = remain = top_size & 7
-
-        "cmp    w4, #4                  \n" // w4 >= 4
-        "blt    2f                      \n"
-        "sub    w4, w4, #4              \n"
-        "st1    {v0.8h, v1.8h}, [%0], #32 \n"
-        "2:                             \n"
-
-        "cmp    w4, #2                  \n" // w4 >= 2
-        "blt    3f                      \n"
-        "sub    w4, w4, #2              \n"
-        "st1    {v0.8h}, [%0], #16      \n"
-        "3:                             \n"
-
-        "cmp    w4, #0                  \n" // w4 > 0
-        "beq    4f                      \n"
-        "st1    {v0.4h}, [%0], #8       \n"
-        "4:                             \n"
-
-        // fill center h loop
-        "cmp    %w5, #0                 \n"
-        "beq    15f                     \n"
-        "5:                             \n"
-
-        // fill left
-        "mov    w4, %w6                 \n" // w4 = left
-        "cmp    w4, #0                  \n"
-        "beq    7f                      \n"
-
-        "6:                             \n"
-        "st1    {v0.4h}, [%0], #8       \n"
-        "subs   w4, w4, #1              \n"
-        "bne    6b                      \n"
-
-        "7:                             \n"
-
-        // fill middle
-        "lsr    w4, %w4, #3             \n" // w4 = nn = w >> 3
-        "cmp    w4, #0                  \n"
-        "beq    9f                      \n"
-
-        "8:                             \n"
-        "prfm   pldl1keep, [%1, #256]   \n"
-        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%1], #64 \n"
-        "subs   w4, w4, #1              \n"
-        "st1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%0], #64 \n"
-        "bne    8b                      \n"
-
-        "9:                             \n"
-
-        "and    w4, %w4, #7             \n" // w4 = remain = w & 7
-
-        "cmp    w4, #4                  \n" // w4 >= 4
-        "blt    10f                     \n"
-        "prfm   pldl1keep, [%1, #256]   \n"
-        "ld1    {v16.8h, v17.8h}, [%1], #32 \n"
-        "sub    w4, w4, #4              \n"
-        "st1    {v16.8h, v17.8h}, [%0], #32 \n"
-        "10:                            \n"
-
-        "cmp    w4, #2                  \n" // w4 >= 2
-        "blt    11f                     \n"
-        "prfm   pldl1keep, [%1, #128]   \n"
-        "ld1    {v16.8h}, [%1], #16     \n"
-        "sub    w4, w4, #2              \n"
-        "st1    {v16.8h}, [%0], #16     \n"
-        "11:                            \n"
-
-        "cmp    w4, #0                  \n" // w4 > 0
-        "beq    12f                     \n"
-        "prfm   pldl1keep, [%1, #64]    \n"
-        "ld1    {v16.4h}, [%1], #8      \n"
-        "st1    {v16.4h}, [%0], #8      \n"
-        "12:                            \n"
-
-        // fill right
-        "mov    w4, %w7                 \n" // w4 = right
-        "cmp    w4, #0                  \n"
-        "beq    14f                     \n"
-
-        "13:                            \n"
-        "subs   w4, w4, #1              \n"
-        "st1    {v0.4h}, [%0], #8       \n"
-        "bne    13b                     \n"
-        "14:                            \n"
-
-        "subs   %w5, %w5, #1            \n"
-        "bne    5b                      \n"
-
-        "15:                            \n"
-
-        // fill bottom
-        "lsr    w4, %w9, #3             \n" // w4 = nn = bottom_size >> 3
-        "cmp    w4, #0                  \n"
-        "beq    17f                     \n"
-
-        "16:                            \n"
-        "st1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0], #64 \n"
-        "subs   w4, w4, #1              \n"
-        "bne    16b                     \n"
-        "17:                            \n"
-
-        // fill bottom remain
-        "and    w4, %w9, #7             \n" // w4 = remain = bottom_size & 7
-
-        "cmp    w4, #4                  \n" // w4 >= 4
-        "blt    18f                     \n"
-        "sub    w4, w4, #4              \n"
-        "st1    {v0.8h, v1.8h}, [%0], #32 \n"
-        "18:                            \n"
-
-        "cmp    w4, #2                  \n" // w4 >= 2
-        "blt    19f                     \n"
-        "sub    w4, w4, #2              \n"
-        "st1    {v0.8h}, [%0], #16      \n"
-        "19:                            \n"
-
-        "cmp    w4, #0                  \n" // w4 > 0
-        "beq    20f                     \n"
-        "st1    {v0.4h}, [%0], #8       \n"
-        "20:                            \n"
-
-        : "=r"(outptr), // %0
-        "=r"(ptr)     // %1
-        : "0"(outptr),
-        "1"(ptr),
-        "r"(w),           // %4
-        "r"(h),           // %5
-        "r"(left),        // %6
-        "r"(right),       // %7
-        "r"(top_size),    // %8
-        "r"(bottom_size), // %9
-        "w"(v)            // %10
-        : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
-#else  // __aarch64__
-    asm volatile(
-        "vmov       q0, %q10            \n"
-        "vmov       q1, %q10            \n"
-        "vmov       q2, %q10            \n"
-        "vmov       q3, %q10            \n"
-
-        // fill top
-        "lsr        r4, %8, #3          \n" // r4 = nn = top_size >> 3
-        "cmp        r4, #0              \n"
-        "beq        1f                  \n"
-
-        "0:                             \n"
-        "vstm       %0!, {d0-d7}        \n"
-        "subs       r4, r4, #1          \n"
-        "bne        0b                  \n"
-
-        "1:                             \n"
-
-        // fill top remain
-        "and        r4, %8, #7          \n" // r4 = remain = top_size & 7
-
-        "cmp        r4, #4              \n" // r4 >= 4
-        "blt        2f                  \n"
-        "sub        r4, r4, #4          \n"
-        "vst1.u16   {d0-d3}, [%0 :64]!  \n"
-        "2:                             \n"
-
-        "cmp        r4, #2              \n" // r4 >= 2
-        "blt        3f                  \n"
-        "sub        r4, r4, #2          \n"
-        "vst1.u16   {d0-d1}, [%0 :64]!  \n"
-        "3:                             \n"
-
-        "cmp        r4, #0              \n" // r4 > 0
-        "beq        4f                  \n"
-        "vst1.u16   {d0}, [%0 :64]!     \n"
-        "4:                             \n"
-
-        // fill center h loop
-        "cmp        %5, #0              \n"
-        "beq        15f                 \n"
-        "5:                             \n"
-
-        // fill left
-        "mov        r4, %6              \n" // r4 = left
-        "cmp        r4, #0              \n"
-        "beq        7f                  \n"
-
-        "6:                             \n"
-        "vst1.u16   {d0}, [%0 :64]!     \n"
-        "subs       r4, r4, #1          \n"
-        "bne        6b                  \n"
-
-        "7:                             \n"
-
-        // fill middle
-        "lsr        r4, %4, #3          \n" // r4 = nn = w >> 3
-        "cmp        r4, #0              \n"
-        "beq        9f                  \n"
-
-        "8:                             \n"
-        "pld        [%1, #512]          \n"
-        "vldm       %1!, {d16-d23}      \n"
-        "subs       r4, r4, #1          \n"
-        "vstm       %0!, {d16-d23}      \n"
-        "bne        8b                  \n"
-
-        "9:                             \n"
-
-        "and        r4, %4, #7          \n" // r4 = remain = w & 7
-
-        "cmp        r4, #4              \n" // r4 >= 4
-        "blt        10f                 \n"
-        "pld        [%1, #256]          \n"
-        "vld1.u16   {d16-d19}, [%1 :64]! \n"
-        "sub        r4, r4, #4          \n"
-        "vst1.u16   {d16-d19}, [%0 :64]! \n"
-        "10:                            \n"
-
-        "cmp        r4, #2              \n" // r4 >= 2
-        "blt        11f                 \n"
-        "pld        [%1, #128]          \n"
-        "vld1.u16   {d16-d17}, [%1 :64]! \n"
-        "sub        r4, r4, #2          \n"
-        "vst1.u16   {d16-d17}, [%0 :64]! \n"
-        "11:                            \n"
-
-        "cmp        r4, #0              \n" // r4 > 0
-        "beq        12f                 \n"
-        "pld        [%1, #64]           \n"
-        "vld1.u16   {d16}, [%1 :64]!    \n"
-        "vst1.u16   {d16}, [%0 :64]!    \n"
-        "12:                            \n"
-
-        // fill right
-        "mov        r4, %7              \n" // r4 = right
-        "cmp        r4, #0              \n"
-        "beq        14f                 \n"
-
-        "13:                            \n"
-        "subs       r4, r4, #1          \n"
-        "vst1.u16   {d0}, [%0 :64]!     \n"
-        "bne        13b                 \n"
-        "14:                            \n"
-
-        "subs       %5, %5, #1          \n"
-        "bne        5b                  \n"
-
-        "15:                            \n"
-
-        // fill bottom
-        "lsr        r4, %9, #3          \n" // r4 = nn = bottom_size >> 3
-        "cmp        r4, #0              \n"
-        "beq        17f                 \n"
-
-        "16:                            \n"
-        "vstm       %0!, {d0-d7}        \n"
-        "subs       r4, r4, #1          \n"
-        "bne        16b                 \n"
-        "17:                            \n"
-
-        // fill bottom remain
-        "and        r4, %9, #7          \n" // r4 = remain = bottom_size & 7
-
-        "cmp        r4, #4              \n" // r4 >= 4
-        "blt        18f                 \n"
-        "sub        r4, r4, #4          \n"
-        "vst1.u16   {d0-d3}, [%0 :64]!  \n"
-        "18:                            \n"
-
-        "cmp        r4, #2              \n" // r4 >= 2
-        "blt        19f                 \n"
-        "sub        r4, r4, #2          \n"
-        "vst1.u16   {d0-d1}, [%0 :64]!  \n"
-        "19:                            \n"
-
-        "cmp        r4, #0              \n" // r4 > 0
-        "beq        20f                 \n"
-        "vst1.u16   {d0}, [%0 :64]!     \n"
-        "20:                            \n"
-
-        : "=r"(outptr), // %0
-        "=r"(ptr)     // %1
-        : "0"(outptr),
-        "1"(ptr),
-        "r"(w),           // %4
-        "r"(h),           // %5
-        "r"(left),        // %6
-        "r"(right),       // %7
-        "r"(top_size),    // %8
-        "r"(bottom_size), // %9
-        "w"(v)            // %10
-        : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-
-    // fill top
-    {
-        int x = 0;
-        for (; x + 3 < top_size; x += 4)
-        {
-            vst1q_u16(outptr, v);
-            vst1q_u16(outptr + 8, v);
-            outptr += 16;
-        }
-        for (; x < top_size; x++)
-        {
-            vst1_u16(outptr, vget_low_u16(v));
-            outptr += 4;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            vst1_u16(outptr, vget_low_u16(v));
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            uint16x4_t _p = vld1_u16(ptr);
-            vst1_u16(outptr, _p);
-            ptr += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1_u16(outptr, vget_low_u16(v));
-            outptr += 4;
-        }
-    }
-    // fill bottom
-    {
-        int x = 0;
-        for (; x + 3 < bottom_size; x += 4)
-        {
-            vst1q_u16(outptr, v);
-            vst1q_u16(outptr + 8, v);
-            outptr += 16;
-        }
-        for (; x < bottom_size; x++)
-        {
-            vst1_u16(outptr, vget_low_u16(v));
-            outptr += 4;
-        }
-    }
-#endif // NCNN_GNU_INLINE_ASM
-}
-
-static void padding_replicate_pack4_bf16_fp16s_neon(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const unsigned short* ptr = src;
-    unsigned short* outptr = dst;
-
-    // fill top
-    for (int y = 0; y < top; y++)
-    {
-        const unsigned short* ptr0 = ptr;
-        uint16x4_t _p = vld1_u16(ptr0);
-        for (int x = 0; x < left; x++)
-        {
-            vst1_u16(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = vld1_u16(ptr0);
-            vst1_u16(outptr, _p);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1_u16(outptr, _p);
-            outptr += 4;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        uint16x4_t _p = vld1_u16(ptr);
-        for (int x = 0; x < left; x++)
-        {
-            vst1_u16(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = vld1_u16(ptr);
-            vst1_u16(outptr, _p);
-            ptr += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1_u16(outptr, _p);
-            outptr += 4;
-        }
-    }
-    // fill bottom
-    ptr -= src.w * 4;
-    for (int y = 0; y < bottom; y++)
-    {
-        const unsigned short* ptr0 = ptr;
-        uint16x4_t _p = vld1_u16(ptr0);
-        for (int x = 0; x < left; x++)
-        {
-            vst1_u16(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = vld1_u16(ptr0);
-            vst1_u16(outptr, _p);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1_u16(outptr, _p);
-            outptr += 4;
-        }
-    }
-}
-
-static void padding_reflect_pack4_bf16_fp16s_neon(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const unsigned short* ptr = src;
-    unsigned short* outptr = dst;
-
-    // fill top
-    ptr += top * src.w * 4;
-    for (int y = 0; y < top; y++)
-    {
-        const unsigned short* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            uint16x4_t _p = vld1_u16(ptr0 + (left - x) * 4);
-            vst1_u16(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            uint16x4_t _p = vld1_u16(ptr0);
-            vst1_u16(outptr, _p);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            uint16x4_t _p = vld1_u16(ptr0 - 8 - x * 4);
-            vst1_u16(outptr, _p);
-            outptr += 4;
-        }
-        ptr -= src.w * 4;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            uint16x4_t _p = vld1_u16(ptr + (left - x) * 4);
-            vst1_u16(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            uint16x4_t _p = vld1_u16(ptr);
-            vst1_u16(outptr, _p);
-            ptr += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            uint16x4_t _p = vld1_u16(ptr - 8 - x * 4);
-            vst1_u16(outptr, _p);
-            outptr += 4;
-        }
-    }
-    // fill bottom
-    ptr -= 2 * src.w * 4;
-    for (int y = 0; y < bottom; y++)
-    {
-        const unsigned short* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            uint16x4_t _p = vld1_u16(ptr0 + (left - x) * 4);
-            vst1_u16(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            uint16x4_t _p = vld1_u16(ptr0);
-            vst1_u16(outptr, _p);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            uint16x4_t _p = vld1_u16(ptr0 - 8 - x * 4);
-            vst1_u16(outptr, _p);
-            outptr += 4;
-        }
-        ptr -= src.w * 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack8_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack8_fp16s.h
deleted file mode 100644
index 7116821f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack8_fp16s.h
+++ /dev/null
@@ -1,376 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void padding_constant_pack8_fp16s_neon(const Mat& src, Mat& dst, int top, int bottom, int left, int right, uint16x8_t v)
-{
-    const unsigned short* ptr = src;
-    unsigned short* outptr = dst;
-
-    int w = src.w;
-    int h = src.h;
-
-    int top_size = top * dst.w;
-    int bottom_size = bottom * dst.w;
-
-#if NCNN_GNU_INLINE_ASM
-    asm volatile(
-        "mov    v0.16b, %10.16b         \n"
-        "mov    v1.16b, %10.16b         \n"
-        "mov    v2.16b, %10.16b         \n"
-        "mov    v3.16b, %10.16b         \n"
-
-        // fill top
-        "lsr    w4, %w8, #2             \n" // w4 = nn = top_size >> 2
-        "cmp    w4, #0                  \n"
-        "beq    1f                      \n"
-
-        "0:                             \n"
-        "st1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0], #64 \n"
-        "subs   w4, w4, #1              \n"
-        "bne    0b                      \n"
-
-        "1:                             \n"
-
-        // fill top remain
-        "and    w4, %w8, #3             \n" // w4 = remain = top_size & 3
-
-        "cmp    w4, #2                  \n" // w4 >= 2
-        "blt    2f                      \n"
-        "sub    w4, w4, #2              \n"
-        "st1    {v0.8h, v1.8h}, [%0], #32 \n"
-        "2:                             \n"
-
-        "cmp    w4, #0                  \n" // w4 > 0
-        "beq    3f                      \n"
-        "st1    {v0.8h}, [%0], #16      \n"
-        "3:                             \n"
-
-        // fill center h loop
-        "cmp    %w5, #0                 \n"
-        "beq    13f                     \n"
-        "4:                             \n"
-
-        // fill left
-        "mov    w4, %w6                 \n" // w4 = left
-        "cmp    w4, #0                  \n"
-        "beq    6f                      \n"
-
-        "5:                             \n"
-        "st1    {v0.8h}, [%0], #16      \n"
-        "subs   w4, w4, #1              \n"
-        "bne    5b                      \n"
-
-        "6:                             \n"
-
-        // fill middle
-        "lsr    w4, %w4, #2             \n" // w4 = nn = w >> 2
-        "cmp    w4, #0                  \n"
-        "beq    8f                      \n"
-
-        "7:                             \n"
-        "prfm   pldl1keep, [%1, #512]   \n"
-        "ld1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%1], #64 \n"
-        "subs   w4, w4, #1              \n"
-        "st1    {v16.8h, v17.8h, v18.8h, v19.8h}, [%0], #64 \n"
-        "bne    7b                      \n"
-
-        "8:                             \n"
-
-        "and    w4, %w4, #3             \n" // w4 = remain = w & 3
-
-        "cmp    w4, #2                  \n" // w4 >= 2
-        "blt    9f                      \n"
-        "prfm   pldl1keep, [%1, #256]   \n"
-        "ld1    {v16.8h, v17.8h}, [%1], #32 \n"
-        "sub    w4, w4, #2              \n"
-        "st1    {v16.8h, v17.8h}, [%0], #32 \n"
-        "9:                             \n"
-
-        "cmp    w4, #0                  \n" // w4 > 0
-        "beq    10f                     \n"
-        "prfm   pldl1keep, [%1, #128]   \n"
-        "ld1    {v16.8h}, [%1], #16     \n"
-        "st1    {v16.8h}, [%0], #16     \n"
-        "10:                            \n"
-
-        // fill right
-        "mov    w4, %w7                 \n" // w4 = right
-        "cmp    w4, #0                  \n"
-        "beq    12f                     \n"
-
-        "11:                            \n"
-        "subs   w4, w4, #1              \n"
-        "st1    {v0.8h}, [%0], #16      \n"
-        "bne    11b                     \n"
-        "12:                            \n"
-
-        "subs   %w5, %w5, #1            \n"
-        "bne    4b                      \n"
-
-        "13:                            \n"
-
-        // fill bottom
-        "lsr    w4, %w9, #2             \n" // w4 = nn = bottom_size >> 2
-        "cmp    w4, #0                  \n"
-        "beq    15f                     \n"
-
-        "14:                            \n"
-        "st1    {v0.8h, v1.8h, v2.8h, v3.8h}, [%0], #64 \n"
-        "subs   w4, w4, #1              \n"
-        "bne    14b                     \n"
-        "15:                            \n"
-
-        // fill bottom remain
-        "and    w4, %w9, #3             \n" // w4 = remain = bottom_size & 3
-
-        "cmp    w4, #2                  \n" // w4 >= 2
-        "blt    16f                     \n"
-        "sub    w4, w4, #2              \n"
-        "st1    {v0.8h, v1.8h}, [%0], #32 \n"
-        "16:                            \n"
-
-        "cmp    w4, #0                  \n" // w4 > 0
-        "beq    17f                     \n"
-        "st1    {v0.8h}, [%0], #16      \n"
-        "17:                            \n"
-
-        : "=r"(outptr), // %0
-        "=r"(ptr)     // %1
-        : "0"(outptr),
-        "1"(ptr),
-        "r"(w),           // %4
-        "r"(h),           // %5
-        "r"(left),        // %6
-        "r"(right),       // %7
-        "r"(top_size),    // %8
-        "r"(bottom_size), // %9
-        "w"(v)            // %10
-        : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
-#else  // NCNN_GNU_INLINE_ASM
-
-    // fill top
-    {
-        int x = 0;
-        for (; x + 3 < top_size; x += 4)
-        {
-            vst1q_u16(outptr, v);
-            vst1q_u16(outptr + 8, v);
-            vst1q_u16(outptr + 16, v);
-            vst1q_u16(outptr + 24, v);
-            outptr += 32;
-        }
-        for (; x < top_size; x++)
-        {
-            vst1q_u16(outptr, v);
-            outptr += 8;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            vst1q_u16(outptr, v);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            uint16x8_t _p = vld1q_u16(ptr);
-            vst1q_u16(outptr, _p);
-            ptr += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1q_u16(outptr, v);
-            outptr += 8;
-        }
-    }
-    // fill bottom
-    {
-        int x = 0;
-        for (; x + 3 < bottom_size; x += 4)
-        {
-            vst1q_u16(outptr, v);
-            vst1q_u16(outptr + 8, v);
-            vst1q_u16(outptr + 16, v);
-            vst1q_u16(outptr + 24, v);
-            outptr += 32;
-        }
-        for (; x < bottom_size; x++)
-        {
-            vst1q_u16(outptr, v);
-            outptr += 8;
-        }
-    }
-#endif // NCNN_GNU_INLINE_ASM
-}
-
-static void padding_replicate_pack8_fp16s_neon(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const unsigned short* ptr = src;
-    unsigned short* outptr = dst;
-
-    // fill top
-    for (int y = 0; y < top; y++)
-    {
-        const unsigned short* ptr0 = ptr;
-        uint16x8_t _p = vld1q_u16(ptr0);
-        for (int x = 0; x < left; x++)
-        {
-            vst1q_u16(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = vld1q_u16(ptr0);
-            vst1q_u16(outptr, _p);
-            ptr0 += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1q_u16(outptr, _p);
-            outptr += 8;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        uint16x8_t _p = vld1q_u16(ptr);
-        for (int x = 0; x < left; x++)
-        {
-            vst1q_u16(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = vld1q_u16(ptr);
-            vst1q_u16(outptr, _p);
-            ptr += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1q_u16(outptr, _p);
-            outptr += 8;
-        }
-    }
-    // fill bottom
-    ptr -= src.w * 8;
-    for (int y = 0; y < bottom; y++)
-    {
-        const unsigned short* ptr0 = ptr;
-        uint16x8_t _p = vld1q_u16(ptr0);
-        for (int x = 0; x < left; x++)
-        {
-            vst1q_u16(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = vld1q_u16(ptr0);
-            vst1q_u16(outptr, _p);
-            ptr0 += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1q_u16(outptr, _p);
-            outptr += 8;
-        }
-    }
-}
-
-static void padding_reflect_pack8_fp16s_neon(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const unsigned short* ptr = src;
-    unsigned short* outptr = dst;
-
-    // fill top
-    ptr += top * src.w * 8;
-    for (int y = 0; y < top; y++)
-    {
-        const unsigned short* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            uint16x8_t _p = vld1q_u16(ptr0 + (left - x) * 8);
-            vst1q_u16(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            uint16x8_t _p = vld1q_u16(ptr0);
-            vst1q_u16(outptr, _p);
-            ptr0 += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            uint16x8_t _p = vld1q_u16(ptr0 - 16 - x * 8);
-            vst1q_u16(outptr, _p);
-            outptr += 8;
-        }
-        ptr -= src.w * 8;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            uint16x8_t _p = vld1q_u16(ptr + (left - x) * 8);
-            vst1q_u16(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            uint16x8_t _p = vld1q_u16(ptr);
-            vst1q_u16(outptr, _p);
-            ptr += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            uint16x8_t _p = vld1q_u16(ptr - 16 - x * 8);
-            vst1q_u16(outptr, _p);
-            outptr += 8;
-        }
-    }
-    // fill bottom
-    ptr -= 2 * src.w * 8;
-    for (int y = 0; y < bottom; y++)
-    {
-        const unsigned short* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            uint16x8_t _p = vld1q_u16(ptr0 + (left - x) * 8);
-            vst1q_u16(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            uint16x8_t _p = vld1q_u16(ptr0);
-            vst1q_u16(outptr, _p);
-            ptr0 += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            uint16x8_t _p = vld1q_u16(ptr0 - 16 - x * 8);
-            vst1q_u16(outptr, _p);
-            outptr += 8;
-        }
-        ptr -= src.w * 8;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack8_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack8_int8.h
deleted file mode 100644
index 02a5b030..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/padding_pack8_int8.h
+++ /dev/null
@@ -1,554 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void padding_constant_pack8_int8_neon(const Mat& src, Mat& dst, int top, int bottom, int left, int right, int8x8_t v)
-{
-    const signed char* ptr = src;
-    signed char* outptr = dst;
-
-    int w = src.w;
-    int h = src.h;
-
-    int top_size = top * dst.w;
-    int bottom_size = bottom * dst.w;
-
-#if NCNN_GNU_INLINE_ASM
-#if __aarch64__
-    asm volatile(
-        "mov    v0.8b, %10.8b           \n"
-        "mov    v0.d[1], v0.d[0]        \n"
-        "mov    v1.16b, v0.16b          \n"
-        "mov    v2.16b, v0.16b          \n"
-        "mov    v3.16b, v0.16b          \n"
-
-        // fill top
-        "lsr    w4, %w8, #3             \n" // w4 = nn = top_size >> 3
-        "cmp    w4, #0                  \n"
-        "beq    1f                      \n"
-
-        "0:                             \n"
-        "st1    {v0.16b, v1.16b, v2.16b, v3.16b}, [%0], #64 \n"
-        "subs   w4, w4, #1              \n"
-        "bne    0b                      \n"
-
-        "1:                             \n"
-
-        // fill top remain
-        "and    w4, %w8, #7             \n" // w4 = remain = top_size & 7
-
-        "cmp    w4, #4                  \n" // w4 >= 4
-        "blt    2f                      \n"
-        "sub    w4, w4, #4              \n"
-        "st1    {v0.16b, v1.16b}, [%0], #32 \n"
-        "2:                             \n"
-
-        "cmp    w4, #2                  \n" // w4 >= 2
-        "blt    3f                      \n"
-        "sub    w4, w4, #2              \n"
-        "st1    {v0.16b}, [%0], #16     \n"
-        "3:                             \n"
-
-        "cmp    w4, #0                  \n" // w4 > 0
-        "beq    4f                      \n"
-        "st1    {v0.8b}, [%0], #8       \n"
-        "4:                             \n"
-
-        // fill center h loop
-        "cmp    %w5, #0                 \n"
-        "beq    15f                     \n"
-        "5:                             \n"
-
-        // fill left
-        "mov    w4, %w6                 \n" // w4 = left
-        "cmp    w4, #0                  \n"
-        "beq    7f                      \n"
-
-        "6:                             \n"
-        "st1    {v0.8b}, [%0], #8       \n"
-        "subs   w4, w4, #1              \n"
-        "bne    6b                      \n"
-
-        "7:                             \n"
-
-        // fill middle
-        "lsr    w4, %w4, #3             \n" // w4 = nn = w >> 3
-        "cmp    w4, #0                  \n"
-        "beq    9f                      \n"
-
-        "8:                             \n"
-        "prfm   pldl1keep, [%1, #512]   \n"
-        "ld1    {v16.16b, v17.16b, v18.16b, v19.16b}, [%1], #64 \n"
-        "subs   w4, w4, #1              \n"
-        "st1    {v16.16b, v17.16b, v18.16b, v19.16b}, [%0], #64 \n"
-        "bne    8b                      \n"
-
-        "9:                             \n"
-
-        "and    w4, %w4, #7             \n" // w4 = remain = w & 7
-
-        "cmp    w4, #4                  \n" // w4 >= 4
-        "blt    10f                     \n"
-        "prfm   pldl1keep, [%1, #256]   \n"
-        "ld1    {v16.16b, v17.16b}, [%1], #32 \n"
-        "sub    w4, w4, #4              \n"
-        "st1    {v16.16b, v17.16b}, [%0], #32 \n"
-        "10:                            \n"
-
-        "cmp    w4, #2                  \n" // w4 >= 2
-        "blt    11f                     \n"
-        "prfm   pldl1keep, [%1, #128]   \n"
-        "ld1    {v16.16b}, [%1], #16    \n"
-        "sub    w4, w4, #2              \n"
-        "st1    {v16.16b}, [%0], #16    \n"
-        "11:                            \n"
-
-        "cmp    w4, #0                  \n" // w4 > 0
-        "beq    12f                     \n"
-        "prfm   pldl1keep, [%1, #64]    \n"
-        "ld1    {v16.8b}, [%1], #8      \n"
-        "st1    {v16.8b}, [%0], #8      \n"
-        "12:                            \n"
-
-        // fill right
-        "mov    w4, %w7                 \n" // w4 = right
-        "cmp    w4, #0                  \n"
-        "beq    14f                     \n"
-
-        "13:                            \n"
-        "subs   w4, w4, #1              \n"
-        "st1    {v0.8b}, [%0], #8       \n"
-        "bne    13b                     \n"
-        "14:                            \n"
-
-        "subs   %w5, %w5, #1            \n"
-        "bne    5b                      \n"
-
-        "15:                            \n"
-
-        // fill bottom
-        "lsr    w4, %w9, #3             \n" // w4 = nn = bottom_size >> 3
-        "cmp    w4, #0                  \n"
-        "beq    17f                     \n"
-
-        "16:                            \n"
-        "st1    {v0.16b, v1.16b, v2.16b, v3.16b}, [%0], #64 \n"
-        "subs   w4, w4, #1              \n"
-        "bne    16b                     \n"
-        "17:                            \n"
-
-        // fill bottom remain
-        "and    w4, %w9, #7             \n" // w4 = remain = bottom_size & 7
-
-        "cmp    w4, #4                  \n" // w4 >= 4
-        "blt    18f                     \n"
-        "sub    w4, w4, #4              \n"
-        "st1    {v0.16b, v1.16b}, [%0], #32 \n"
-        "18:                            \n"
-
-        "cmp    w4, #2                  \n" // w4 >= 2
-        "blt    19f                     \n"
-        "sub    w4, w4, #2              \n"
-        "st1    {v0.16b}, [%0], #16     \n"
-        "19:                            \n"
-
-        "cmp    w4, #0                  \n" // w4 > 0
-        "beq    20f                     \n"
-        "st1    {v0.8b}, [%0], #8       \n"
-        "20:                            \n"
-
-        : "=r"(outptr), // %0
-        "=r"(ptr)     // %1
-        : "0"(outptr),
-        "1"(ptr),
-        "r"(w),           // %4
-        "r"(h),           // %5
-        "r"(left),        // %6
-        "r"(right),       // %7
-        "r"(top_size),    // %8
-        "r"(bottom_size), // %9
-        "w"(v)            // %10
-        : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
-#else  // __aarch64__
-    asm volatile(
-        "vmov       d0, %P10            \n"
-        "vmov       d1, d0              \n"
-        "vmov       q1, q0              \n"
-        "vmov       q2, q0              \n"
-        "vmov       q3, q0              \n"
-
-        // fill top
-        "lsr        r4, %8, #3          \n" // r4 = nn = top_size >> 3
-        "cmp        r4, #0              \n"
-        "beq        1f                  \n"
-
-        "0:                             \n"
-        "vstm       %0!, {d0-d7}        \n"
-        "subs       r4, r4, #1          \n"
-        "bne        0b                  \n"
-
-        "1:                             \n"
-
-        // fill top remain
-        "and        r4, %8, #7          \n" // r4 = remain = top_size & 7
-
-        "cmp        r4, #4              \n" // r4 >= 4
-        "blt        2f                  \n"
-        "sub        r4, r4, #4          \n"
-        "vst1.s8    {d0-d3}, [%0 :128]! \n"
-        "2:                             \n"
-
-        "cmp        r4, #2              \n" // r4 >= 2
-        "blt        3f                  \n"
-        "sub        r4, r4, #2          \n"
-        "vst1.s8    {d0-d1}, [%0 :128]! \n"
-        "3:                             \n"
-
-        "cmp        r4, #0              \n" // r4 > 0
-        "beq        4f                  \n"
-        "vst1.s8    {d0}, [%0 :64]!     \n"
-        "4:                             \n"
-
-        // fill center h loop
-        "cmp        %5, #0              \n"
-        "beq        15f                 \n"
-        "5:                             \n"
-
-        // fill left
-        "mov        r4, %6              \n" // r4 = left
-        "cmp        r4, #0              \n"
-        "beq        7f                  \n"
-
-        "6:                             \n"
-        "vst1.s8    {d0}, [%0 :64]!     \n"
-        "subs       r4, r4, #1          \n"
-        "bne        6b                  \n"
-
-        "7:                             \n"
-
-        // fill middle
-        "lsr        r4, %4, #3          \n" // r4 = nn = w >> 3
-        "cmp        r4, #0              \n"
-        "beq        9f                  \n"
-
-        "8:                             \n"
-        "pld        [%1, #512]          \n"
-        "vldm       %1!, {d16-d23}      \n"
-        "subs       r4, r4, #1          \n"
-        "vstm       %0!, {d16-d23}      \n"
-        "bne        8b                  \n"
-
-        "9:                             \n"
-
-        "and        r4, %4, #7          \n" // r4 = remain = w & 7
-
-        "cmp        r4, #4              \n" // r4 >= 4
-        "blt        10f                 \n"
-        "pld        [%1, #256]          \n"
-        "vld1.s8    {d16-d19}, [%1 :64]! \n"
-        "sub        r4, r4, #4          \n"
-        "vst1.s8    {d16-d19}, [%0 :64]! \n"
-        "10:                            \n"
-
-        "cmp        r4, #2              \n" // r4 >= 2
-        "blt        11f                 \n"
-        "pld        [%1, #128]          \n"
-        "vld1.s8    {d16-d17}, [%1 :64]! \n"
-        "sub        r4, r4, #2          \n"
-        "vst1.s8    {d16-d17}, [%0 :64]! \n"
-        "11:                            \n"
-
-        "cmp        r4, #0              \n" // r4 > 0
-        "beq        12f                 \n"
-        "pld        [%1, #64]           \n"
-        "vld1.s8    {d16}, [%1 :64]!    \n"
-        "vst1.s8    {d16}, [%0 :64]!    \n"
-        "12:                            \n"
-
-        // fill right
-        "mov        r4, %7              \n" // r4 = right
-        "cmp        r4, #0              \n"
-        "beq        14f                 \n"
-
-        "13:                            \n"
-        "subs       r4, r4, #1          \n"
-        "vst1.s8    {d0}, [%0 :64]!     \n"
-        "bne        13b                 \n"
-        "14:                            \n"
-
-        "subs       %5, %5, #1          \n"
-        "bne        5b                  \n"
-
-        "15:                            \n"
-
-        // fill bottom
-        "lsr        r4, %9, #3          \n" // r4 = nn = bottom_size >> 3
-        "cmp        r4, #0              \n"
-        "beq        17f                 \n"
-
-        "16:                            \n"
-        "vstm       %0!, {d0-d7}        \n"
-        "subs       r4, r4, #1          \n"
-        "bne        16b                 \n"
-        "17:                            \n"
-
-        // fill bottom remain
-        "and        r4, %9, #7          \n" // r4 = remain = bottom_size & 7
-
-        "cmp        r4, #4              \n" // r4 >= 4
-        "blt        18f                 \n"
-        "sub        r4, r4, #4          \n"
-        "vst1.s8    {d0-d3}, [%0 :64]!  \n"
-        "18:                            \n"
-
-        "cmp        r4, #2              \n" // r4 >= 2
-        "blt        19f                 \n"
-        "sub        r4, r4, #2          \n"
-        "vst1.s8    {d0-d1}, [%0 :64]!  \n"
-        "19:                            \n"
-
-        "cmp        r4, #0              \n" // r4 > 0
-        "beq        20f                 \n"
-        "vst1.s8    {d0}, [%0 :64]!     \n"
-        "20:                            \n"
-
-        : "=r"(outptr), // %0
-        "=r"(ptr)     // %1
-        : "0"(outptr),
-        "1"(ptr),
-        "r"(w),           // %4
-        "r"(h),           // %5
-        "r"(left),        // %6
-        "r"(right),       // %7
-        "r"(top_size),    // %8
-        "r"(bottom_size), // %9
-        "w"(v)            // %10
-        : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
-#endif // __aarch64__
-#else  // NCNN_GNU_INLINE_ASM
-
-    // fill top
-    {
-        int x = 0;
-        for (; x + 3 < top_size; x += 4)
-        {
-            vst1_s8(outptr, v);
-            vst1_s8(outptr + 8, v);
-            vst1_s8(outptr + 16, v);
-            vst1_s8(outptr + 24, v);
-            outptr += 32;
-        }
-        for (; x < top_size; x++)
-        {
-            vst1_s8(outptr, v);
-            outptr += 8;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            vst1_s8(outptr, v);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            int8x8_t _p = vld1_s8(ptr);
-            vst1_s8(outptr, _p);
-            ptr += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1_s8(outptr, v);
-            outptr += 8;
-        }
-    }
-    // fill bottom
-    {
-        int x = 0;
-        for (; x + 3 < bottom_size; x += 4)
-        {
-            vst1_s8(outptr, v);
-            vst1_s8(outptr + 8, v);
-            vst1_s8(outptr + 16, v);
-            vst1_s8(outptr + 24, v);
-            outptr += 32;
-        }
-        for (; x < bottom_size; x++)
-        {
-            vst1_s8(outptr, v);
-            outptr += 8;
-        }
-    }
-#endif // NCNN_GNU_INLINE_ASM
-}
-
-static void padding_replicate_pack8_int8_neon(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const signed char* ptr = src;
-    signed char* outptr = dst;
-
-    // fill top
-    for (int y = 0; y < top; y++)
-    {
-        const signed char* ptr0 = ptr;
-        int8x8_t _p = vld1_s8(ptr0);
-        for (int x = 0; x < left; x++)
-        {
-            vst1_s8(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = vld1_s8(ptr0);
-            vst1_s8(outptr, _p);
-            ptr0 += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1_s8(outptr, _p);
-            outptr += 8;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        int8x8_t _p = vld1_s8(ptr);
-        for (int x = 0; x < left; x++)
-        {
-            vst1_s8(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = vld1_s8(ptr);
-            vst1_s8(outptr, _p);
-            ptr += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1_s8(outptr, _p);
-            outptr += 8;
-        }
-    }
-    // fill bottom
-    ptr -= src.w * 8;
-    for (int y = 0; y < bottom; y++)
-    {
-        const signed char* ptr0 = ptr;
-        int8x8_t _p = vld1_s8(ptr0);
-        for (int x = 0; x < left; x++)
-        {
-            vst1_s8(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = vld1_s8(ptr0);
-            vst1_s8(outptr, _p);
-            ptr0 += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            vst1_s8(outptr, _p);
-            outptr += 8;
-        }
-    }
-}
-
-static void padding_reflect_pack8_int8_neon(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const signed char* ptr = src;
-    signed char* outptr = dst;
-
-    // fill top
-    ptr += top * src.w * 8;
-    for (int y = 0; y < top; y++)
-    {
-        const signed char* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            int8x8_t _p = vld1_s8(ptr0 + (left - x) * 8);
-            vst1_s8(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            int8x8_t _p = vld1_s8(ptr0);
-            vst1_s8(outptr, _p);
-            ptr0 += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            int8x8_t _p = vld1_s8(ptr0 - 16 - x * 8);
-            vst1_s8(outptr, _p);
-            outptr += 8;
-        }
-        ptr -= src.w * 8;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            int8x8_t _p = vld1_s8(ptr + (left - x) * 8);
-            vst1_s8(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            int8x8_t _p = vld1_s8(ptr);
-            vst1_s8(outptr, _p);
-            ptr += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            int8x8_t _p = vld1_s8(ptr - 16 - x * 8);
-            vst1_s8(outptr, _p);
-            outptr += 8;
-        }
-    }
-    // fill bottom
-    ptr -= 2 * src.w * 8;
-    for (int y = 0; y < bottom; y++)
-    {
-        const signed char* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            int8x8_t _p = vld1_s8(ptr0 + (left - x) * 8);
-            vst1_s8(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            int8x8_t _p = vld1_s8(ptr0);
-            vst1_s8(outptr, _p);
-            ptr0 += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            int8x8_t _p = vld1_s8(ptr0 - 16 - x * 8);
-            vst1_s8(outptr, _p);
-            outptr += 8;
-        }
-        ptr -= src.w * 8;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pixelshuffle_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pixelshuffle_arm.h
deleted file mode 100644
index a2d714c9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pixelshuffle_arm.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PIXELSHUFFLE_ARM_H
-#define LAYER_PIXELSHUFFLE_ARM_H
-
-#include "pixelshuffle.h"
-
-namespace ncnn {
-
-class PixelShuffle_arm : public PixelShuffle
-{
-public:
-    PixelShuffle_arm();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_bf16s_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PIXELSHUFFLE_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_2x2.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_2x2.h
deleted file mode 100644
index 34482de4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_2x2.h
+++ /dev/null
@@ -1,113 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pooling2x2s2_max_neon(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = w - 2 * outw + w;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const float* img0 = bottom_blob.channel(q);
-        float* outptr = top_blob.channel(q);
-
-        const float* r0 = img0;
-        const float* r1 = img0 + w;
-
-        for (int i = 0; i < outh; i++)
-        {
-#if __ARM_NEON
-            int nn = outw >> 2;
-            int remain = outw - (nn << 2);
-#else
-            int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    "0:                                   \n"
-                    "prfm       pldl1keep, [%1, #256]     \n"
-                    "prfm       pldl1keep, [%2, #256]     \n"
-                    "ld1        {v0.4s, v1.4s}, [%1], #32 \n"
-                    "ld1        {v2.4s, v3.4s}, [%2], #32 \n"
-                    "fmax       v0.4s, v0.4s, v2.4s       \n"
-                    "fmax       v1.4s, v1.4s, v3.4s       \n"
-                    "fmaxp      v2.4s, v0.4s, v1.4s       \n"
-                    "subs       %w0, %w0, #1              \n"
-                    "st1        {v2.4s}, [%3], #16        \n"
-                    "bne        0b                        \n"
-                    : "=r"(nn),    // %0
-                    "=r"(r0),    // %1
-                    "=r"(r1),    // %2
-                    "=r"(outptr) // %3
-                    : "0"(nn),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(outptr)
-                    : "cc", "memory", "v0", "v1", "v2", "v3");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    "0:                             \n"
-                    "pld        [%1, #256]          \n"
-                    "pld        [%2, #256]          \n"
-                    "vld1.f32   {d0-d3}, [%1]!      \n"
-                    "vld1.f32   {d4-d7}, [%2]!      \n"
-                    "vmax.f32   q0, q0, q2          \n"
-                    "vmax.f32   q1, q1, q3          \n"
-                    "vpmax.f32  d4, d0, d1          \n"
-                    "vpmax.f32  d5, d2, d3          \n"
-                    "subs       %0, #1              \n"
-                    "vst1.f32   {d4-d5}, [%3]!      \n"
-                    "bne        0b                  \n"
-                    : "=r"(nn),    // %0
-                    "=r"(r0),    // %1
-                    "=r"(r1),    // %2
-                    "=r"(outptr) // %3
-                    : "0"(nn),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(outptr)
-                    : "cc", "memory", "q0", "q1", "q2", "q3");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-                float max0 = std::max(r0[0], r0[1]);
-                float max1 = std::max(r1[0], r1[1]);
-
-                *outptr = std::max(max0, max1);
-
-                r0 += 2;
-                r1 += 2;
-                outptr++;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_2x2_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_2x2_pack4.h
deleted file mode 100644
index 881e249c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_2x2_pack4.h
+++ /dev/null
@@ -1,144 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pooling2x2s2_max_pack4_neon(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        float* outptr = top_blob.channel(q);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j + 3 < outw; j += 4)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]   \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n"
-
-                    "fmax   v0.4s, v0.4s, v1.4s     \n"
-                    "fmax   v2.4s, v2.4s, v3.4s     \n"
-
-                    "prfm   pldl1keep, [%1, #512]   \n"
-                    "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n"
-
-                    "fmax   v4.4s, v4.4s, v5.4s     \n"
-                    "fmax   v6.4s, v6.4s, v7.4s     \n"
-
-                    "prfm   pldl1keep, [%2, #512]   \n"
-                    "ld1    {v16.4s, v17.4s, v18.4s, v19.4s}, [%2], #64 \n"
-
-                    "fmax   v16.4s, v16.4s, v17.4s  \n"
-                    "fmax   v18.4s, v18.4s, v19.4s  \n"
-
-                    "prfm   pldl1keep, [%2, #512]   \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%2], #64 \n"
-
-                    "fmax   v20.4s, v20.4s, v21.4s  \n"
-                    "fmax   v22.4s, v22.4s, v23.4s  \n"
-
-                    "fmax   v0.4s, v0.4s, v16.4s    \n"
-                    "fmax   v1.4s, v2.4s, v18.4s    \n"
-                    "fmax   v2.4s, v4.4s, v20.4s    \n"
-                    "fmax   v3.4s, v6.4s, v22.4s    \n"
-
-                    "st1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%0], #64 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(r0),     // %1
-                    "=r"(r1)      // %2
-                    : "0"(outptr),
-                    "1"(r0),
-                    "2"(r1)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%1, #512]      \n"
-                    "vldm       %1!, {d0-d7}    \n"
-
-                    "vmax.f32   q0, q0, q1      \n"
-                    "vmax.f32   q2, q2, q3      \n"
-
-                    "pld        [%1, #512]      \n"
-                    "vldm       %1!, {d8-d15}   \n"
-
-                    "vmax.f32   q4, q4, q5      \n"
-                    "vmax.f32   q6, q6, q7      \n"
-
-                    "pld        [%2, #512]      \n"
-                    "vldm       %2!, {d16-d23}  \n"
-
-                    "vmax.f32   q8, q8, q9      \n"
-                    "vmax.f32   q10, q10, q11   \n"
-
-                    "pld        [%2, #512]      \n"
-                    "vldm       %2!, {d24-d31}  \n"
-
-                    "vmax.f32   q12, q12, q13   \n"
-                    "vmax.f32   q14, q14, q15   \n"
-
-                    "vmax.f32   q0, q0, q8      \n"
-                    "vmax.f32   q1, q2, q10     \n"
-                    "vmax.f32   q2, q4, q12     \n"
-                    "vmax.f32   q3, q6, q14     \n"
-
-                    "vstm       %0!, {d0-d7}    \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(r0),     // %1
-                    "=r"(r1)      // %2
-                    : "0"(outptr),
-                    "1"(r0),
-                    "2"(r1)
-                    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-            }
-            for (; j < outw; j++)
-            {
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-
-                float32x4_t _max0 = vmaxq_f32(_r00, _r01);
-                float32x4_t _max1 = vmaxq_f32(_r10, _r11);
-                float32x4_t _max = vmaxq_f32(_max0, _max1);
-
-                vst1q_f32(outptr, _max);
-
-                r0 += 8;
-                r1 += 8;
-                outptr += 4;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_2x2_pack4_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_2x2_pack4_bf16s.h
deleted file mode 100644
index de3f14f4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_2x2_pack4_bf16s.h
+++ /dev/null
@@ -1,194 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pooling2x2s2_max_pack4_bf16s_neon(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        unsigned short* outptr = top_blob.channel(q);
-
-        const unsigned short* r0 = img0.row<const unsigned short>(0);
-        const unsigned short* r1 = img0.row<const unsigned short>(1);
-
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j + 3 < outw; j += 4)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #256]   \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%1], #32 \n"
-
-                    "shll   v0.4s, v0.4h, #16       \n"
-                    "shll   v1.4s, v1.4h, #16       \n"
-                    "shll   v2.4s, v2.4h, #16       \n"
-                    "shll   v3.4s, v3.4h, #16       \n"
-
-                    "fmax   v0.4s, v0.4s, v1.4s     \n"
-                    "fmax   v2.4s, v2.4s, v3.4s     \n"
-
-                    "prfm   pldl1keep, [%1, #256]   \n"
-                    "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%1], #32 \n"
-
-                    "shll   v4.4s, v4.4h, #16       \n"
-                    "shll   v5.4s, v5.4h, #16       \n"
-                    "shll   v6.4s, v6.4h, #16       \n"
-                    "shll   v7.4s, v7.4h, #16       \n"
-
-                    "fmax   v4.4s, v4.4s, v5.4s     \n"
-                    "fmax   v6.4s, v6.4s, v7.4s     \n"
-
-                    "prfm   pldl1keep, [%2, #256]   \n"
-                    "ld1    {v16.4h, v17.4h, v18.4h, v19.4h}, [%2], #32 \n"
-
-                    "shll   v16.4s, v16.4h, #16     \n"
-                    "shll   v17.4s, v17.4h, #16     \n"
-                    "shll   v18.4s, v18.4h, #16     \n"
-                    "shll   v19.4s, v19.4h, #16     \n"
-
-                    "fmax   v16.4s, v16.4s, v17.4s  \n"
-                    "fmax   v18.4s, v18.4s, v19.4s  \n"
-
-                    "prfm   pldl1keep, [%2, #256]   \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%2], #32 \n"
-
-                    "shll   v20.4s, v20.4h, #16     \n"
-                    "shll   v21.4s, v21.4h, #16     \n"
-                    "shll   v22.4s, v22.4h, #16     \n"
-                    "shll   v23.4s, v23.4h, #16     \n"
-
-                    "fmax   v20.4s, v20.4s, v21.4s  \n"
-                    "fmax   v22.4s, v22.4s, v23.4s  \n"
-
-                    "fmax   v0.4s, v0.4s, v16.4s    \n"
-                    "fmax   v1.4s, v2.4s, v18.4s    \n"
-                    "fmax   v2.4s, v4.4s, v20.4s    \n"
-                    "fmax   v3.4s, v6.4s, v22.4s    \n"
-
-                    "shrn   v0.4h, v0.4s, #16       \n"
-                    "shrn   v1.4h, v1.4s, #16       \n"
-                    "shrn   v2.4h, v2.4s, #16       \n"
-                    "shrn   v3.4h, v3.4s, #16       \n"
-
-                    "st1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%0], #32 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(r0),     // %1
-                    "=r"(r1)      // %2
-                    : "0"(outptr),
-                    "1"(r0),
-                    "2"(r1)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%1, #256]      \n"
-                    "vld1.u16   {d4-d7}, [%1]!  \n"
-
-                    "vshll.u16  q0, d4, #16     \n"
-                    "vshll.u16  q1, d5, #16     \n"
-                    "vshll.u16  q2, d6, #16     \n"
-                    "vshll.u16  q3, d7, #16     \n"
-
-                    "vmax.f32   q0, q0, q1      \n"
-                    "vmax.f32   q2, q2, q3      \n"
-
-                    "pld        [%1, #256]      \n"
-                    "vld1.u16   {d12-d15}, [%1]! \n"
-
-                    "vshll.u16  q4, d12, #16    \n"
-                    "vshll.u16  q5, d13, #16    \n"
-                    "vshll.u16  q6, d14, #16    \n"
-                    "vshll.u16  q7, d15, #16    \n"
-
-                    "vmax.f32   q4, q4, q5      \n"
-                    "vmax.f32   q6, q6, q7      \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.u16   {d20-d23}, [%2]! \n"
-
-                    "vshll.u16  q8, d20, #16    \n"
-                    "vshll.u16  q9, d21, #16    \n"
-                    "vshll.u16  q10, d22, #16   \n"
-                    "vshll.u16  q11, d23, #16   \n"
-
-                    "vmax.f32   q8, q8, q9      \n"
-                    "vmax.f32   q10, q10, q11   \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.u16   {d28-d31}, [%2]! \n"
-
-                    "vshll.u16  q12, d28, #16   \n"
-                    "vshll.u16  q13, d29, #16   \n"
-                    "vshll.u16  q14, d30, #16   \n"
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmax.f32   q12, q12, q13   \n"
-                    "vmax.f32   q14, q14, q15   \n"
-
-                    "vmax.f32   q0, q0, q8      \n"
-                    "vmax.f32   q1, q2, q10     \n"
-                    "vmax.f32   q2, q4, q12     \n"
-                    "vmax.f32   q3, q6, q14     \n"
-
-                    "vshrn.u32  d0, q0, #16     \n"
-                    "vshrn.u32  d1, q1, #16     \n"
-                    "vshrn.u32  d2, q2, #16     \n"
-                    "vshrn.u32  d3, q3, #16     \n"
-
-                    "vst1.u16   {d0-d3}, [%0]!  \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(r0),     // %1
-                    "=r"(r1)      // %2
-                    : "0"(outptr),
-                    "1"(r0),
-                    "2"(r1)
-                    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-            }
-            for (; j < outw; j++)
-            {
-                float32x4_t _r00 = bfloat2float(vld1_u16(r0));
-                float32x4_t _r01 = bfloat2float(vld1_u16(r0 + 4));
-                float32x4_t _r10 = bfloat2float(vld1_u16(r1));
-                float32x4_t _r11 = bfloat2float(vld1_u16(r1 + 4));
-
-                float32x4_t _max0 = vmaxq_f32(_r00, _r01);
-                float32x4_t _max1 = vmaxq_f32(_r10, _r11);
-                float32x4_t _max = vmaxq_f32(_max0, _max1);
-
-                vst1_u16(outptr, float2bfloat(_max));
-
-                r0 += 8;
-                r1 += 8;
-                outptr += 4;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_3x3.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_3x3.h
deleted file mode 100644
index f105aeea..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_3x3.h
+++ /dev/null
@@ -1,191 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pooling3x3s2_max_neon(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = w - 2 * outw + w;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const float* img0 = bottom_blob.channel(q);
-        float* outptr = top_blob.channel(q);
-
-        const float* r0 = img0;
-        const float* r1 = img0 + w;
-        const float* r2 = img0 + w * 2;
-
-        for (int i = 0; i < outh; i++)
-        {
-#if __ARM_NEON
-            int nn = outw >> 2;
-            int remain = outw - (nn << 2);
-#else
-            int remain = outw;
-#endif // __ARM_NEON
-
-#if __ARM_NEON
-#if __aarch64__
-            if (nn > 0)
-            {
-                asm volatile(
-                    "prfm       pldl1keep, [%1, #256]       \n"
-                    "ld2        {v0.4s, v1.4s}, [%1], #32   \n"
-                    "prfm       pldl1keep, [%2, #256]       \n"
-                    "ld2        {v2.4s, v3.4s}, [%2], #32   \n"
-                    "prfm       pldl1keep, [%3, #256]       \n"
-                    "ld2        {v4.4s, v5.4s}, [%3], #32   \n"
-                    "0:                                     \n"
-
-                    "prfm       pldl1keep, [%1, #256]       \n"
-                    "ld2        {v6.4s, v7.4s}, [%1], #32   \n"
-
-                    "fmax       v12.4s, v0.4s, v1.4s        \n"
-                    "fmax       v13.4s, v2.4s, v3.4s        \n"
-
-                    "prfm       pldl1keep, [%2, #256]       \n"
-                    "ld2        {v8.4s, v9.4s}, [%2], #32   \n"
-
-                    "fmax       v14.4s, v4.4s, v5.4s        \n"
-                    "ext        v0.16b, v0.16b, v6.16b, #4  \n"
-
-                    "prfm       pldl1keep, [%3, #256]       \n"
-                    "ld2        {v10.4s, v11.4s}, [%3], #32 \n"
-
-                    "ext        v2.16b, v2.16b, v8.16b, #4  \n"
-
-                    "fmax       v12.4s, v12.4s, v0.4s       \n"
-                    "ext        v4.16b, v4.16b, v10.16b, #4 \n"
-
-                    "fmax       v13.4s, v13.4s, v2.4s       \n"
-                    "fmax       v14.4s, v14.4s, v4.4s       \n"
-                    "fmax       v12.4s, v12.4s, v13.4s      \n"
-
-                    "orr        v0.16b, v6.16b, v6.16b      \n"
-                    "orr        v1.16b, v7.16b, v7.16b      \n"
-                    "fmax       v12.4s, v12.4s, v14.4s      \n"
-
-                    "orr        v2.16b, v8.16b, v8.16b      \n"
-                    "orr        v3.16b, v9.16b, v9.16b      \n"
-                    "orr        v4.16b, v10.16b, v10.16b    \n"
-                    "orr        v5.16b, v11.16b, v11.16b    \n"
-
-                    "subs       %w0, %w0, #1                \n"
-                    "st1        {v12.4s}, [%4], #16         \n"
-                    "bne        0b                          \n"
-                    "sub        %1, %1, #32                 \n"
-                    "sub        %2, %2, #32                 \n"
-                    "sub        %3, %3, #32                 \n"
-                    : "=r"(nn),    // %0
-                    "=r"(r0),    // %1
-                    "=r"(r1),    // %2
-                    "=r"(r2),    // %3
-                    "=r"(outptr) // %4
-                    : "0"(nn),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "4"(outptr)
-                    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14");
-            }
-#else
-            if (nn > 0)
-            {
-                asm volatile(
-                    "pld        [%1, #256]          \n"
-                    "vld2.f32   {d0-d3}, [%1]!      \n" // q0 = 0 2 4 6  q1 = 1 3 5 7
-                    "pld        [%2, #256]          \n"
-                    "vld2.f32   {d4-d7}, [%2]!      \n"
-                    "pld        [%3, #256]          \n"
-                    "vld2.f32   {d8-d11}, [%3]!     \n"
-                    "0:                             \n"
-                    "pld        [%1, #256]          \n"
-                    "vld2.f32   {d12-d15}, [%1]!    \n" // q6 = 8 10 12 14  q7 = 9 11 13 15
-
-                    "vmax.f32   q12, q0, q1         \n"
-                    "vmax.f32   q13, q2, q3         \n"
-
-                    "pld        [%2, #256]          \n"
-                    "vld2.f32   {d16-d19}, [%2]!    \n"
-
-                    "vmax.f32   q14, q4, q5         \n"
-                    "vext.32    q0, q0, q6, #1      \n"
-
-                    "pld        [%3, #256]          \n"
-                    "vld2.f32   {d20-d23}, [%3]!    \n"
-
-                    "vext.32    q2, q2, q8, #1      \n"
-
-                    "vmax.f32   q12, q12, q0        \n"
-                    "vext.32    q4, q4, q10, #1     \n"
-
-                    "vmax.f32   q13, q13, q2        \n"
-                    "vmax.f32   q14, q14, q4        \n"
-                    "vmax.f32   q12, q12, q13       \n"
-
-                    "vorr       q0, q6, q6          \n"
-                    "vorr       q1, q7, q7          \n"
-                    "vmax.f32   q12, q12, q14       \n"
-
-                    "vorr       q2, q8, q8          \n"
-                    "vorr       q3, q9, q9          \n"
-                    "vorr       q4, q10, q10        \n"
-                    "vorr       q5, q11, q11        \n"
-
-                    "subs       %0, #1              \n"
-                    "vst1.f32   {d24-d25}, [%4]!    \n"
-                    "bne        0b                  \n"
-                    "sub        %1, #32             \n"
-                    "sub        %2, #32             \n"
-                    "sub        %3, #32             \n"
-                    : "=r"(nn),    // %0
-                    "=r"(r0),    // %1
-                    "=r"(r1),    // %2
-                    "=r"(r2),    // %3
-                    "=r"(outptr) // %4
-                    : "0"(nn),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2),
-                    "4"(outptr)
-                    : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14");
-            }
-#endif // __aarch64__
-#endif // __ARM_NEON
-            for (; remain > 0; remain--)
-            {
-                float max0 = std::max(std::max(r0[0], r0[1]), r0[2]);
-                float max1 = std::max(std::max(r1[0], r1[1]), r1[2]);
-                float max2 = std::max(std::max(r2[0], r2[1]), r2[2]);
-
-                *outptr = std::max(std::max(max0, max1), max2);
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-                outptr++;
-            }
-
-            r0 += tailstep; //1 + w;
-            r1 += tailstep; //1 + w;
-            r2 += tailstep; //1 + w;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_3x3_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_3x3_pack4.h
deleted file mode 100644
index c0d50945..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_3x3_pack4.h
+++ /dev/null
@@ -1,333 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pooling3x3s2_max_pack4_neon(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        float* outptr = top_blob.channel(q);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j + 3 < outw; j += 4)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]   \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n"
-
-                    "fmax   v16.4s, v0.4s, v1.4s    \n"
-                    "fmax   v17.4s, v2.4s, v3.4s    \n"
-
-                    "prfm   pldl1keep, [%1, #512]   \n"
-                    "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%1], #64 \n"
-
-                    "fmax   v18.4s, v4.4s, v5.4s    \n"
-                    "fmax   v19.4s, v6.4s, v7.4s    \n"
-
-                    "ld1    {v8.4s}, [%1]           \n"
-
-                    "fmax   v20.4s, v16.4s, v2.4s   \n"
-                    "fmax   v21.4s, v17.4s, v4.4s   \n"
-
-                    "prfm   pldl1keep, [%2, #512]   \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%2], #64 \n"
-
-                    "fmax   v22.4s, v18.4s, v6.4s   \n"
-                    "fmax   v23.4s, v19.4s, v8.4s   \n"
-
-                    "prfm   pldl1keep, [%2, #512]   \n"
-                    "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%2], #64 \n"
-
-                    "fmax   v16.4s, v0.4s, v1.4s    \n"
-                    "fmax   v17.4s, v2.4s, v3.4s    \n"
-
-                    "fmax   v18.4s, v4.4s, v5.4s    \n"
-                    "fmax   v19.4s, v6.4s, v7.4s    \n"
-
-                    "ld1    {v8.4s}, [%2]           \n"
-
-                    "fmax   v24.4s, v16.4s, v2.4s   \n"
-                    "fmax   v25.4s, v17.4s, v4.4s   \n"
-
-                    "prfm   pldl1keep, [%3, #512]   \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%3], #64 \n"
-
-                    "fmax   v26.4s, v18.4s, v6.4s   \n"
-                    "fmax   v27.4s, v19.4s, v8.4s   \n"
-
-                    "prfm   pldl1keep, [%3, #512]   \n"
-                    "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%3], #64 \n"
-
-                    "fmax   v16.4s, v0.4s, v1.4s    \n"
-                    "fmax   v17.4s, v2.4s, v3.4s    \n"
-
-                    "fmax   v18.4s, v4.4s, v5.4s    \n"
-                    "fmax   v19.4s, v6.4s, v7.4s    \n"
-
-                    "ld1    {v8.4s}, [%3]           \n"
-
-                    "fmax   v28.4s, v16.4s, v2.4s   \n"
-                    "fmax   v29.4s, v17.4s, v4.4s   \n"
-                    "fmax   v30.4s, v18.4s, v6.4s   \n"
-                    "fmax   v31.4s, v19.4s, v8.4s   \n"
-
-                    "fmax   v20.4s, v20.4s, v24.4s  \n"
-                    "fmax   v21.4s, v21.4s, v25.4s  \n"
-                    "fmax   v22.4s, v22.4s, v26.4s  \n"
-                    "fmax   v23.4s, v23.4s, v27.4s  \n"
-
-                    "fmax   v20.4s, v20.4s, v28.4s  \n"
-                    "fmax   v21.4s, v21.4s, v29.4s  \n"
-                    "fmax   v22.4s, v22.4s, v30.4s  \n"
-                    "fmax   v23.4s, v23.4s, v31.4s  \n"
-
-                    "st1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%0], #64 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(r0),     // %1
-                    "=r"(r1),     // %2
-                    "=r"(r2)      // %3
-                    : "0"(outptr),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%1, #512]      \n"
-                    "vldm       %1!, {d0-d7}    \n"
-
-                    "pld        [%2, #512]      \n"
-                    "vldm       %2!, {d8-d15}   \n"
-
-                    "vmax.f32   q0, q0, q4      \n"
-                    "vmax.f32   q1, q1, q5      \n"
-
-                    "pld        [%3, #512]      \n"
-                    "vldm       %3!, {d16-d23}  \n"
-
-                    "vmax.f32   q2, q2, q6      \n"
-                    "vmax.f32   q3, q3, q7      \n"
-
-                    "vmax.f32   q0, q0, q8      \n"
-                    "vmax.f32   q1, q1, q9      \n"
-
-                    "pld        [%1, #512]      \n"
-                    "vldm       %1!, {d8-d15}   \n"
-
-                    "vmax.f32   q2, q2, q10     \n"
-                    "vmax.f32   q3, q3, q11     \n"
-
-                    "pld        [%2, #512]      \n"
-                    "vldm       %2!, {d16-d23}  \n"
-
-                    "vmax.f32   q4, q4, q8      \n"
-                    "vmax.f32   q5, q5, q9      \n"
-
-                    "pld        [%3, #512]      \n"
-                    "vldm       %3!, {d24-d31}  \n"
-
-                    "vmax.f32   q6, q6, q10     \n"
-                    "vmax.f32   q7, q7, q11     \n"
-
-                    "vmax.f32   q4, q4, q12     \n"
-                    "vmax.f32   q5, q5, q13     \n"
-
-                    "vld1.f32   {d24-d25}, [%1 :128] \n"
-                    "vld1.f32   {d26-d27}, [%2 :128] \n"
-
-                    "vmax.f32   q6, q6, q14     \n"
-                    "vmax.f32   q7, q7, q15     \n"
-
-                    "vld1.f32   {d28-d29}, [%3 :128] \n"
-
-                    "vmax.f32   q8, q12, q13    \n"
-                    "vmax.f32   q8, q8, q14     \n"
-
-                    "vmax.f32   q12, q0, q1     \n"
-                    "vmax.f32   q13, q2, q3     \n"
-                    "vmax.f32   q14, q4, q5     \n"
-                    "vmax.f32   q15, q6, q7     \n"
-
-                    "vmax.f32   q12, q12, q2    \n"
-                    "vmax.f32   q13, q13, q4    \n"
-                    "vmax.f32   q14, q14, q6    \n"
-                    "vmax.f32   q15, q15, q8    \n"
-
-                    "vstm       %0!, {d24-d31}  \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(r0),     // %1
-                    "=r"(r1),     // %2
-                    "=r"(r2)      // %3
-                    : "0"(outptr),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2)
-                    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #512]   \n"
-                    "ld1    {v0.4s, v1.4s, v2.4s, v3.4s}, [%1], #64 \n"
-
-                    "prfm   pldl1keep, [%2, #512]   \n"
-                    "ld1    {v4.4s, v5.4s, v6.4s, v7.4s}, [%2], #64 \n"
-
-                    "fmax   v16.4s, v0.4s, v4.4s    \n"
-                    "fmax   v17.4s, v1.4s, v5.4s    \n"
-
-                    "prfm   pldl1keep, [%3, #512]   \n"
-                    "ld1    {v20.4s, v21.4s, v22.4s, v23.4s}, [%3], #64 \n"
-
-                    "fmax   v18.4s, v2.4s, v6.4s    \n"
-                    "fmax   v19.4s, v3.4s, v7.4s    \n"
-
-                    "ld1    {v0.4s}, [%1]           \n"
-
-                    "fmax   v16.4s, v16.4s, v20.4s  \n"
-                    "fmax   v17.4s, v17.4s, v21.4s  \n"
-
-                    "ld1    {v1.4s}, [%2]           \n"
-
-                    "fmax   v18.4s, v18.4s, v22.4s  \n"
-                    "fmax   v19.4s, v19.4s, v23.4s  \n"
-
-                    "ld1    {v2.4s}, [%3]           \n"
-
-                    "fmax   v3.4s, v0.4s, v1.4s     \n"
-
-                    "fmax   v20.4s, v16.4s, v17.4s  \n"
-                    "fmax   v21.4s, v18.4s, v19.4s  \n"
-
-                    "fmax   v3.4s, v3.4s, v2.4s     \n"
-
-                    "fmax   v20.4s, v20.4s, v18.4s  \n"
-                    "fmax   v21.4s, v21.4s, v3.4s   \n"
-
-                    "st1    {v20.4s, v21.4s}, [%0], #32 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(r0),     // %1
-                    "=r"(r1),     // %2
-                    "=r"(r2)      // %3
-                    : "0"(outptr),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%1, #512]      \n"
-                    "vldm       %1!, {d0-d7}    \n"
-
-                    "pld        [%2, #512]      \n"
-                    "vldm       %2!, {d8-d15}   \n"
-
-                    "vmax.f32   q12, q0, q4     \n"
-                    "vmax.f32   q13, q1, q5     \n"
-
-                    "pld        [%3, #512]      \n"
-                    "vldm       %3!, {d16-d23}  \n"
-
-                    "vmax.f32   q14, q2, q6     \n"
-                    "vmax.f32   q15, q3, q7     \n"
-
-                    "vld1.f32   {d0-d1}, [%1 :128] \n"
-
-                    "vmax.f32   q12, q12, q8    \n"
-                    "vmax.f32   q13, q13, q9    \n"
-
-                    "vld1.f32   {d2-d3}, [%2 :128] \n"
-
-                    "vmax.f32   q14, q14, q10   \n"
-                    "vmax.f32   q15, q15, q11   \n"
-
-                    "vld1.f32   {d4-d5}, [%3 :128] \n"
-
-                    "vmax.f32   q3, q0, q1      \n"
-
-                    "vmax.f32   q4, q12, q13    \n"
-                    "vmax.f32   q5, q14, q15    \n"
-
-                    "vmax.f32   q3, q3, q2      \n"
-
-                    "vmax.f32   q4, q4, q14     \n"
-                    "vmax.f32   q5, q5, q3      \n"
-
-                    "vst1.f32   {d8-d11}, [%0 :128]! \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(r0),     // %1
-                    "=r"(r1),     // %2
-                    "=r"(r2)      // %3
-                    : "0"(outptr),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2)
-                    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-            }
-            for (; j < outw; j++)
-            {
-                float32x4_t _r00 = vld1q_f32(r0);
-                float32x4_t _r01 = vld1q_f32(r0 + 4);
-                float32x4_t _r02 = vld1q_f32(r0 + 8);
-                float32x4_t _r10 = vld1q_f32(r1);
-                float32x4_t _r11 = vld1q_f32(r1 + 4);
-                float32x4_t _r12 = vld1q_f32(r1 + 8);
-                float32x4_t _r20 = vld1q_f32(r2);
-                float32x4_t _r21 = vld1q_f32(r2 + 4);
-                float32x4_t _r22 = vld1q_f32(r2 + 8);
-
-                float32x4_t _max0 = vmaxq_f32(vmaxq_f32(_r00, _r01), _r02);
-                float32x4_t _max1 = vmaxq_f32(vmaxq_f32(_r10, _r11), _r12);
-                float32x4_t _max2 = vmaxq_f32(vmaxq_f32(_r20, _r21), _r22);
-
-                float32x4_t _max = vmaxq_f32(vmaxq_f32(_max0, _max1), _max2);
-
-                vst1q_f32(outptr, _max);
-
-                r0 += 8;
-                r1 += 8;
-                r2 += 8;
-                outptr += 4;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_3x3_pack4_bf16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_3x3_pack4_bf16s.h
deleted file mode 100644
index b5f1a4c5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_3x3_pack4_bf16s.h
+++ /dev/null
@@ -1,448 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pooling3x3s2_max_pack4_bf16s_neon(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        unsigned short* outptr = top_blob.channel(q);
-
-        const unsigned short* r0 = img0.row<const unsigned short>(0);
-        const unsigned short* r1 = img0.row<const unsigned short>(1);
-        const unsigned short* r2 = img0.row<const unsigned short>(2);
-
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j + 3 < outw; j += 4)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #256]   \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%1], #32 \n"
-
-                    "shll   v0.4s, v0.4h, #16       \n"
-                    "shll   v1.4s, v1.4h, #16       \n"
-                    "shll   v2.4s, v2.4h, #16       \n"
-                    "shll   v3.4s, v3.4h, #16       \n"
-
-                    "fmax   v16.4s, v0.4s, v1.4s    \n"
-                    "fmax   v17.4s, v2.4s, v3.4s    \n"
-
-                    "prfm   pldl1keep, [%1, #256]   \n"
-                    "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%1], #32 \n"
-
-                    "shll   v4.4s, v4.4h, #16       \n"
-                    "shll   v5.4s, v5.4h, #16       \n"
-                    "shll   v6.4s, v6.4h, #16       \n"
-                    "shll   v7.4s, v7.4h, #16       \n"
-
-                    "fmax   v18.4s, v4.4s, v5.4s    \n"
-                    "fmax   v19.4s, v6.4s, v7.4s    \n"
-
-                    "ld1    {v8.4h}, [%1]           \n"
-                    "shll   v8.4s, v8.4h, #16       \n"
-
-                    "fmax   v20.4s, v16.4s, v2.4s   \n"
-                    "fmax   v21.4s, v17.4s, v4.4s   \n"
-
-                    "prfm   pldl1keep, [%2, #256]   \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%2], #32 \n"
-
-                    "shll   v0.4s, v0.4h, #16       \n"
-                    "shll   v1.4s, v1.4h, #16       \n"
-                    "shll   v2.4s, v2.4h, #16       \n"
-                    "shll   v3.4s, v3.4h, #16       \n"
-
-                    "fmax   v22.4s, v18.4s, v6.4s   \n"
-                    "fmax   v23.4s, v19.4s, v8.4s   \n"
-
-                    "prfm   pldl1keep, [%2, #256]   \n"
-                    "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%2], #32 \n"
-
-                    "shll   v4.4s, v4.4h, #16       \n"
-                    "shll   v5.4s, v5.4h, #16       \n"
-                    "shll   v6.4s, v6.4h, #16       \n"
-                    "shll   v7.4s, v7.4h, #16       \n"
-
-                    "fmax   v16.4s, v0.4s, v1.4s    \n"
-                    "fmax   v17.4s, v2.4s, v3.4s    \n"
-
-                    "fmax   v18.4s, v4.4s, v5.4s    \n"
-                    "fmax   v19.4s, v6.4s, v7.4s    \n"
-
-                    "ld1    {v8.4h}, [%2]           \n"
-                    "shll   v8.4s, v8.4h, #16       \n"
-
-                    "fmax   v24.4s, v16.4s, v2.4s   \n"
-                    "fmax   v25.4s, v17.4s, v4.4s   \n"
-
-                    "prfm   pldl1keep, [%3, #256]   \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%3], #32 \n"
-
-                    "shll   v0.4s, v0.4h, #16       \n"
-                    "shll   v1.4s, v1.4h, #16       \n"
-                    "shll   v2.4s, v2.4h, #16       \n"
-                    "shll   v3.4s, v3.4h, #16       \n"
-
-                    "fmax   v26.4s, v18.4s, v6.4s   \n"
-                    "fmax   v27.4s, v19.4s, v8.4s   \n"
-
-                    "prfm   pldl1keep, [%3, #256]   \n"
-                    "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%3], #32 \n"
-
-                    "shll   v4.4s, v4.4h, #16       \n"
-                    "shll   v5.4s, v5.4h, #16       \n"
-                    "shll   v6.4s, v6.4h, #16       \n"
-                    "shll   v7.4s, v7.4h, #16       \n"
-
-                    "fmax   v16.4s, v0.4s, v1.4s    \n"
-                    "fmax   v17.4s, v2.4s, v3.4s    \n"
-
-                    "fmax   v18.4s, v4.4s, v5.4s    \n"
-                    "fmax   v19.4s, v6.4s, v7.4s    \n"
-
-                    "ld1    {v8.4h}, [%3]           \n"
-                    "shll   v8.4s, v8.4h, #16       \n"
-
-                    "fmax   v28.4s, v16.4s, v2.4s   \n"
-                    "fmax   v29.4s, v17.4s, v4.4s   \n"
-                    "fmax   v30.4s, v18.4s, v6.4s   \n"
-                    "fmax   v31.4s, v19.4s, v8.4s   \n"
-
-                    "fmax   v20.4s, v20.4s, v24.4s  \n"
-                    "fmax   v21.4s, v21.4s, v25.4s  \n"
-                    "fmax   v22.4s, v22.4s, v26.4s  \n"
-                    "fmax   v23.4s, v23.4s, v27.4s  \n"
-
-                    "fmax   v20.4s, v20.4s, v28.4s  \n"
-                    "fmax   v21.4s, v21.4s, v29.4s  \n"
-                    "fmax   v22.4s, v22.4s, v30.4s  \n"
-                    "fmax   v23.4s, v23.4s, v31.4s  \n"
-
-                    "shrn   v20.4h, v20.4s, #16     \n"
-                    "shrn   v21.4h, v21.4s, #16     \n"
-                    "shrn   v22.4h, v22.4s, #16     \n"
-                    "shrn   v23.4h, v23.4s, #16     \n"
-
-                    "st1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%0], #32 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(r0),     // %1
-                    "=r"(r1),     // %2
-                    "=r"(r2)      // %3
-                    : "0"(outptr),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%1, #256]      \n"
-                    "vld1.u16   {d4-d7}, [%1]!  \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.u16   {d12-d15}, [%2]! \n"
-
-                    "vshll.u16  q0, d4, #16     \n"
-                    "vshll.u16  q1, d5, #16     \n"
-                    "vshll.u16  q2, d6, #16     \n"
-                    "vshll.u16  q3, d7, #16     \n"
-
-                    "vshll.u16  q4, d12, #16    \n"
-                    "vshll.u16  q5, d13, #16    \n"
-                    "vshll.u16  q6, d14, #16    \n"
-                    "vshll.u16  q7, d15, #16    \n"
-
-                    "vmax.f32   q0, q0, q4      \n"
-                    "vmax.f32   q1, q1, q5      \n"
-
-                    "pld        [%3, #256]      \n"
-                    "vld1.u16   {d20-d23}, [%3]! \n"
-
-                    "vshll.u16  q8, d20, #16    \n"
-                    "vshll.u16  q9, d21, #16    \n"
-                    "vshll.u16  q10, d22, #16   \n"
-                    "vshll.u16  q11, d23, #16   \n"
-
-                    "vmax.f32   q2, q2, q6      \n"
-                    "vmax.f32   q3, q3, q7      \n"
-
-                    "vmax.f32   q0, q0, q8      \n"
-                    "vmax.f32   q1, q1, q9      \n"
-
-                    "pld        [%1, #256]      \n"
-                    "vld1.u16   {d12-d15}, [%1]! \n"
-
-                    "vshll.u16  q4, d12, #16    \n"
-                    "vshll.u16  q5, d13, #16    \n"
-                    "vshll.u16  q6, d14, #16    \n"
-                    "vshll.u16  q7, d15, #16    \n"
-
-                    "vmax.f32   q2, q2, q10     \n"
-                    "vmax.f32   q3, q3, q11     \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.u16   {d20-d23}, [%2]! \n"
-
-                    "vshll.u16  q8, d20, #16    \n"
-                    "vshll.u16  q9, d21, #16    \n"
-                    "vshll.u16  q10, d22, #16   \n"
-                    "vshll.u16  q11, d23, #16   \n"
-
-                    "vmax.f32   q4, q4, q8      \n"
-                    "vmax.f32   q5, q5, q9      \n"
-
-                    "pld        [%3, #256]      \n"
-                    "vld1.u16   {d28-d31}, [%3]! \n"
-
-                    "vshll.u16  q12, d28, #16   \n"
-                    "vshll.u16  q13, d29, #16   \n"
-                    "vshll.u16  q14, d30, #16   \n"
-                    "vshll.u16  q15, d31, #16   \n"
-
-                    "vmax.f32   q6, q6, q10     \n"
-                    "vmax.f32   q7, q7, q11     \n"
-
-                    "vmax.f32   q4, q4, q12     \n"
-                    "vmax.f32   q5, q5, q13     \n"
-
-                    "vld1.u16   {d25}, [%1]     \n"
-                    "vld1.u16   {d27}, [%2]     \n"
-                    "vshll.u16  q12, d25, #16   \n"
-                    "vshll.u16  q13, d27, #16   \n"
-
-                    "vmax.f32   q6, q6, q14     \n"
-                    "vmax.f32   q7, q7, q15     \n"
-
-                    "vld1.u16   {d29}, [%3]     \n"
-                    "vshll.u16  q14, d29, #16   \n"
-
-                    "vmax.f32   q8, q12, q13    \n"
-                    "vmax.f32   q8, q8, q14     \n"
-
-                    "vmax.f32   q12, q0, q1     \n"
-                    "vmax.f32   q13, q2, q3     \n"
-                    "vmax.f32   q14, q4, q5     \n"
-                    "vmax.f32   q15, q6, q7     \n"
-
-                    "vmax.f32   q12, q12, q2    \n"
-                    "vmax.f32   q13, q13, q4    \n"
-                    "vmax.f32   q14, q14, q6    \n"
-                    "vmax.f32   q15, q15, q8    \n"
-
-                    "vshrn.u32  d24, q12, #16   \n"
-                    "vshrn.u32  d25, q13, #16   \n"
-                    "vshrn.u32  d26, q14, #16   \n"
-                    "vshrn.u32  d27, q15, #16   \n"
-
-                    "vst1.u16   {d24-d27}, [%0]! \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(r0),     // %1
-                    "=r"(r1),     // %2
-                    "=r"(r2)      // %3
-                    : "0"(outptr),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2)
-                    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-#if __aarch64__
-                asm volatile(
-                    "prfm   pldl1keep, [%1, #256]   \n"
-                    "ld1    {v0.4h, v1.4h, v2.4h, v3.4h}, [%1], #32 \n"
-
-                    "prfm   pldl1keep, [%2, #256]   \n"
-                    "ld1    {v4.4h, v5.4h, v6.4h, v7.4h}, [%2], #32 \n"
-
-                    "shll   v0.4s, v0.4h, #16       \n"
-                    "shll   v1.4s, v1.4h, #16       \n"
-                    "shll   v2.4s, v2.4h, #16       \n"
-                    "shll   v3.4s, v3.4h, #16       \n"
-
-                    "shll   v4.4s, v4.4h, #16       \n"
-                    "shll   v5.4s, v5.4h, #16       \n"
-                    "shll   v6.4s, v6.4h, #16       \n"
-                    "shll   v7.4s, v7.4h, #16       \n"
-
-                    "fmax   v16.4s, v0.4s, v4.4s    \n"
-                    "fmax   v17.4s, v1.4s, v5.4s    \n"
-
-                    "prfm   pldl1keep, [%3, #256]   \n"
-                    "ld1    {v20.4h, v21.4h, v22.4h, v23.4h}, [%3], #32 \n"
-
-                    "shll   v20.4s, v20.4h, #16     \n"
-                    "shll   v21.4s, v21.4h, #16     \n"
-                    "shll   v22.4s, v22.4h, #16     \n"
-                    "shll   v23.4s, v23.4h, #16     \n"
-
-                    "fmax   v18.4s, v2.4s, v6.4s    \n"
-                    "fmax   v19.4s, v3.4s, v7.4s    \n"
-
-                    "ld1    {v0.4s}, [%1]           \n"
-
-                    "fmax   v16.4s, v16.4s, v20.4s  \n"
-                    "fmax   v17.4s, v17.4s, v21.4s  \n"
-
-                    "ld1    {v1.4s}, [%2]           \n"
-
-                    "fmax   v18.4s, v18.4s, v22.4s  \n"
-                    "fmax   v19.4s, v19.4s, v23.4s  \n"
-
-                    "ld1    {v2.4s}, [%3]           \n"
-
-                    "fmax   v3.4s, v0.4s, v1.4s     \n"
-
-                    "fmax   v20.4s, v16.4s, v17.4s  \n"
-                    "fmax   v21.4s, v18.4s, v19.4s  \n"
-
-                    "fmax   v3.4s, v3.4s, v2.4s     \n"
-
-                    "fmax   v20.4s, v20.4s, v18.4s  \n"
-                    "fmax   v21.4s, v21.4s, v3.4s   \n"
-
-                    "shrn   v20.4h, v20.4s, #16     \n"
-                    "shrn   v21.4h, v21.4s, #16     \n"
-
-                    "st1    {v20.4h, v21.4h}, [%0], #16 \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(r0),     // %1
-                    "=r"(r1),     // %2
-                    "=r"(r2)      // %3
-                    : "0"(outptr),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2)
-                    : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
-#else  // __aarch64__
-                asm volatile(
-                    "pld        [%1, #256]      \n"
-                    "vld1.u16   {d4-d7}, [%1]!  \n"
-
-                    "pld        [%2, #256]      \n"
-                    "vld1.u16   {d12-d15}, [%2]! \n"
-
-                    "vshll.u16  q0, d4, #16     \n"
-                    "vshll.u16  q1, d5, #16     \n"
-                    "vshll.u16  q2, d6, #16     \n"
-                    "vshll.u16  q3, d7, #16     \n"
-
-                    "vshll.u16  q4, d12, #16    \n"
-                    "vshll.u16  q5, d13, #16    \n"
-                    "vshll.u16  q6, d14, #16    \n"
-                    "vshll.u16  q7, d15, #16    \n"
-
-                    "vmax.f32   q12, q0, q4     \n"
-                    "vmax.f32   q13, q1, q5     \n"
-
-                    "pld        [%3, #256]      \n"
-                    "vld1.u16   {d20-d23}, [%3]! \n"
-
-                    "vshll.u16  q8, d20, #16    \n"
-                    "vshll.u16  q9, d21, #16    \n"
-                    "vshll.u16  q10, d22, #16   \n"
-                    "vshll.u16  q11, d23, #16   \n"
-
-                    "vmax.f32   q14, q2, q6     \n"
-                    "vmax.f32   q15, q3, q7     \n"
-
-                    "vld1.u16   {d1}, [%1]      \n"
-                    "vshll.u16  q0, d1, #16     \n"
-
-                    "vmax.f32   q12, q12, q8    \n"
-                    "vmax.f32   q13, q13, q9    \n"
-
-                    "vld1.u16   {d3}, [%2]      \n"
-                    "vshll.u16  q1, d3, #16     \n"
-
-                    "vmax.f32   q14, q14, q10   \n"
-                    "vmax.f32   q15, q15, q11   \n"
-
-                    "vld1.u16   {d5}, [%3]      \n"
-                    "vshll.u16  q2, d5, #16     \n"
-
-                    "vmax.f32   q3, q0, q1      \n"
-
-                    "vmax.f32   q4, q12, q13    \n"
-                    "vmax.f32   q5, q14, q15    \n"
-
-                    "vmax.f32   q3, q3, q2      \n"
-
-                    "vmax.f32   q4, q4, q14     \n"
-                    "vmax.f32   q5, q5, q3      \n"
-
-                    "vshrn.u32  d8, q4, #16     \n"
-                    "vshrn.u32  d9, q5, #16     \n"
-
-                    "vst1.u16   {d8-d9}, [%0]!  \n"
-
-                    : "=r"(outptr), // %0
-                    "=r"(r0),     // %1
-                    "=r"(r1),     // %2
-                    "=r"(r2)      // %3
-                    : "0"(outptr),
-                    "1"(r0),
-                    "2"(r1),
-                    "3"(r2)
-                    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
-#endif // __aarch64__
-            }
-            for (; j < outw; j++)
-            {
-                float32x4_t _r00 = bfloat2float(vld1_u16(r0));
-                float32x4_t _r01 = bfloat2float(vld1_u16(r0 + 4));
-                float32x4_t _r02 = bfloat2float(vld1_u16(r0 + 8));
-                float32x4_t _r10 = bfloat2float(vld1_u16(r1));
-                float32x4_t _r11 = bfloat2float(vld1_u16(r1 + 4));
-                float32x4_t _r12 = bfloat2float(vld1_u16(r1 + 8));
-                float32x4_t _r20 = bfloat2float(vld1_u16(r2));
-                float32x4_t _r21 = bfloat2float(vld1_u16(r2 + 4));
-                float32x4_t _r22 = bfloat2float(vld1_u16(r2 + 8));
-
-                float32x4_t _max0 = vmaxq_f32(vmaxq_f32(_r00, _r01), _r02);
-                float32x4_t _max1 = vmaxq_f32(vmaxq_f32(_r10, _r11), _r12);
-                float32x4_t _max2 = vmaxq_f32(vmaxq_f32(_r20, _r21), _r22);
-
-                float32x4_t _max = vmaxq_f32(vmaxq_f32(_max0, _max1), _max2);
-
-                vst1_u16(outptr, float2bfloat(_max));
-
-                r0 += 8;
-                r1 += 8;
-                r2 += 8;
-                outptr += 4;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_arm.h
deleted file mode 100644
index ead9270c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/pooling_arm.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_POOLING_ARM_H
-#define LAYER_POOLING_ARM_H
-
-#include "pooling.h"
-
-namespace ncnn {
-
-class Pooling_arm : public Pooling
-{
-public:
-    Pooling_arm();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_POOLING_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/prelu_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/prelu_arm.h
deleted file mode 100644
index 9354be74..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/prelu_arm.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PRELU_ARM_H
-#define LAYER_PRELU_ARM_H
-
-#include "prelu.h"
-
-namespace ncnn {
-
-class PReLU_arm : public PReLU
-{
-public:
-    PReLU_arm();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_inplace_bf16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PRELU_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/quantize_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/quantize_arm.h
deleted file mode 100644
index 60a71619..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/quantize_arm.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2018 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_QUANTIZE_ARM_H
-#define LAYER_QUANTIZE_ARM_H
-
-#include "quantize.h"
-
-namespace ncnn {
-
-class Quantize_arm : public Quantize
-{
-public:
-    Quantize_arm();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_QUANTIZE_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/relu_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/relu_arm.h
deleted file mode 100644
index c2212513..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/relu_arm.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_RELU_ARM_H
-#define LAYER_RELU_ARM_H
-
-#include "relu.h"
-
-namespace ncnn {
-
-class ReLU_arm : public ReLU
-{
-public:
-    ReLU_arm();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_inplace_bf16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-    int forward_inplace_int8(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_RELU_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_arm.h
deleted file mode 100644
index c6fc184a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_arm.h
+++ /dev/null
@@ -1,33 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 BUG1989. All rights reserved.
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_REQUANTIZE_ARM_H
-#define LAYER_REQUANTIZE_ARM_H
-
-#include "requantize.h"
-
-namespace ncnn {
-
-class Requantize_arm : public Requantize
-{
-public:
-    Requantize_arm();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_REQUANTIZE_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_leakyrelu_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_leakyrelu_pack4.h
deleted file mode 100644
index 98482ac0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_leakyrelu_pack4.h
+++ /dev/null
@@ -1,275 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_leakyrelu_pack4_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, float slope, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-    int size = w * h;
-    int outc = top_blob.c;
-    int out_elempack = top_blob.elempack;
-
-    int scale_in_data_size = scale_in_data.w;
-    int scale_out_data_size = scale_out_data.w;
-    int bias_data_size = bias_data.w;
-
-    // int8(leakyrelu(v * scale_in, slope) * scale_out)
-    // int8_leakyrelu(v * (scale_in * scale_out), slope)
-
-    // int8(leakyrelu(v * scale_in + bias, slope) * scale_out)
-    // int8_leakyrelu(v * (scale_in * scale_out) + (bias * scale_out), slope)
-
-    if (out_elempack == 8)
-    {
-        if (bias_data_size == 0)
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < outc; q++)
-            {
-                const int* intptr0 = bottom_blob.channel(q * 2);
-                const int* intptr1 = bottom_blob.channel(q * 2 + 1);
-                signed char* ptr = top_blob.channel(q);
-
-                float32x4_t _scale_in0 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8);
-                float32x4_t _scale_in1 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8 + 4);
-                float32x4_t _scale_out0 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8);
-                float32x4_t _scale_out1 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8 + 4);
-
-                float32x4_t _scale0 = vmulq_f32(_scale_in0, _scale_out0);
-                float32x4_t _scale1 = vmulq_f32(_scale_in1, _scale_out1);
-                float32x4_t _slope = vdupq_n_f32(slope);
-
-                int i = 0;
-#if __aarch64__
-                for (; i + 3 < size; i += 4)
-                {
-                    float32x4_t _v00 = vcvtq_f32_s32(vld1q_s32(intptr0));
-                    float32x4_t _v01 = vcvtq_f32_s32(vld1q_s32(intptr0 + 4));
-                    float32x4_t _v02 = vcvtq_f32_s32(vld1q_s32(intptr0 + 8));
-                    float32x4_t _v03 = vcvtq_f32_s32(vld1q_s32(intptr0 + 12));
-                    float32x4_t _v10 = vcvtq_f32_s32(vld1q_s32(intptr1));
-                    float32x4_t _v11 = vcvtq_f32_s32(vld1q_s32(intptr1 + 4));
-                    float32x4_t _v12 = vcvtq_f32_s32(vld1q_s32(intptr1 + 8));
-                    float32x4_t _v13 = vcvtq_f32_s32(vld1q_s32(intptr1 + 12));
-                    _v00 = vmulq_f32(_v00, _scale0);
-                    _v01 = vmulq_f32(_v01, _scale0);
-                    _v02 = vmulq_f32(_v02, _scale0);
-                    _v03 = vmulq_f32(_v03, _scale0);
-                    _v10 = vmulq_f32(_v10, _scale1);
-                    _v11 = vmulq_f32(_v11, _scale1);
-                    _v12 = vmulq_f32(_v12, _scale1);
-                    _v13 = vmulq_f32(_v13, _scale1);
-                    vst1_s8(ptr, float2int8leakyrelu(_v00, _v10, _slope));
-                    vst1_s8(ptr + 8, float2int8leakyrelu(_v01, _v11, _slope));
-                    vst1_s8(ptr + 16, float2int8leakyrelu(_v02, _v12, _slope));
-                    vst1_s8(ptr + 24, float2int8leakyrelu(_v03, _v13, _slope));
-
-                    intptr0 += 16;
-                    intptr1 += 16;
-                    ptr += 32;
-                }
-#endif // __aarch64__
-                for (; i < size; i++)
-                {
-                    float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr0));
-                    float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr1));
-                    _v0 = vmulq_f32(_v0, _scale0);
-                    _v1 = vmulq_f32(_v1, _scale1);
-                    vst1_s8(ptr, float2int8leakyrelu(_v0, _v1, _slope));
-
-                    intptr0 += 4;
-                    intptr1 += 4;
-                    ptr += 8;
-                }
-            }
-        }
-        else
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < outc; q++)
-            {
-                const int* intptr0 = bottom_blob.channel(q * 2);
-                const int* intptr1 = bottom_blob.channel(q * 2 + 1);
-                signed char* ptr = top_blob.channel(q);
-
-                float32x4_t _scale_in0 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8);
-                float32x4_t _scale_in1 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8 + 4);
-                float32x4_t _scale_out0 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8);
-                float32x4_t _scale_out1 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8 + 4);
-                float32x4_t _bias0 = bias_data_size == 1 ? vdupq_n_f32(bias_data[0]) : vld1q_f32((const float*)bias_data + q * 8);
-                float32x4_t _bias1 = bias_data_size == 1 ? vdupq_n_f32(bias_data[0]) : vld1q_f32((const float*)bias_data + q * 8 + 4);
-
-                float32x4_t _scale0 = vmulq_f32(_scale_in0, _scale_out0);
-                float32x4_t _scale1 = vmulq_f32(_scale_in1, _scale_out1);
-                _bias0 = vmulq_f32(_bias0, _scale_out0);
-                _bias1 = vmulq_f32(_bias1, _scale_out1);
-                float32x4_t _slope = vdupq_n_f32(slope);
-
-                int i = 0;
-#if __aarch64__
-                for (; i + 3 < size; i += 4)
-                {
-                    float32x4_t _v00 = vcvtq_f32_s32(vld1q_s32(intptr0));
-                    float32x4_t _v01 = vcvtq_f32_s32(vld1q_s32(intptr0 + 4));
-                    float32x4_t _v02 = vcvtq_f32_s32(vld1q_s32(intptr0 + 8));
-                    float32x4_t _v03 = vcvtq_f32_s32(vld1q_s32(intptr0 + 12));
-                    float32x4_t _v10 = vcvtq_f32_s32(vld1q_s32(intptr1));
-                    float32x4_t _v11 = vcvtq_f32_s32(vld1q_s32(intptr1 + 4));
-                    float32x4_t _v12 = vcvtq_f32_s32(vld1q_s32(intptr1 + 8));
-                    float32x4_t _v13 = vcvtq_f32_s32(vld1q_s32(intptr1 + 12));
-                    _v00 = vfmaq_f32(_bias0, _v00, _scale0);
-                    _v01 = vfmaq_f32(_bias0, _v01, _scale0);
-                    _v02 = vfmaq_f32(_bias0, _v02, _scale0);
-                    _v03 = vfmaq_f32(_bias0, _v03, _scale0);
-                    _v10 = vfmaq_f32(_bias1, _v10, _scale1);
-                    _v11 = vfmaq_f32(_bias1, _v11, _scale1);
-                    _v12 = vfmaq_f32(_bias1, _v12, _scale1);
-                    _v13 = vfmaq_f32(_bias1, _v13, _scale1);
-                    vst1_s8(ptr, float2int8leakyrelu(_v00, _v10, _slope));
-                    vst1_s8(ptr + 8, float2int8leakyrelu(_v01, _v11, _slope));
-                    vst1_s8(ptr + 16, float2int8leakyrelu(_v02, _v12, _slope));
-                    vst1_s8(ptr + 24, float2int8leakyrelu(_v03, _v13, _slope));
-
-                    intptr0 += 16;
-                    intptr1 += 16;
-                    ptr += 32;
-                }
-#endif // __aarch64__
-                for (; i + 1 < size; i += 2)
-                {
-                    float32x4_t _v00 = vcvtq_f32_s32(vld1q_s32(intptr0));
-                    float32x4_t _v01 = vcvtq_f32_s32(vld1q_s32(intptr0 + 4));
-                    float32x4_t _v10 = vcvtq_f32_s32(vld1q_s32(intptr1));
-                    float32x4_t _v11 = vcvtq_f32_s32(vld1q_s32(intptr1 + 4));
-#if __aarch64__
-                    _v00 = vfmaq_f32(_bias0, _v00, _scale0);
-                    _v01 = vfmaq_f32(_bias0, _v01, _scale0);
-                    _v10 = vfmaq_f32(_bias1, _v10, _scale1);
-                    _v11 = vfmaq_f32(_bias1, _v11, _scale1);
-#else  // __aarch64__
-                    _v00 = vmlaq_f32(_bias0, _v00, _scale0);
-                    _v01 = vmlaq_f32(_bias0, _v01, _scale0);
-                    _v10 = vmlaq_f32(_bias1, _v10, _scale1);
-                    _v11 = vmlaq_f32(_bias1, _v11, _scale1);
-#endif // __aarch64__
-                    vst1_s8(ptr, float2int8leakyrelu(_v00, _v10, _slope));
-                    vst1_s8(ptr + 8, float2int8leakyrelu(_v01, _v11, _slope));
-
-                    intptr0 += 8;
-                    intptr1 += 8;
-                    ptr += 16;
-                }
-                for (; i < size; i++)
-                {
-                    float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr0));
-                    float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr1));
-#if __aarch64__
-                    _v0 = vfmaq_f32(_bias0, _v0, _scale0);
-                    _v1 = vfmaq_f32(_bias1, _v1, _scale1);
-#else  // __aarch64__
-                    _v0 = vmlaq_f32(_bias0, _v0, _scale0);
-                    _v1 = vmlaq_f32(_bias1, _v1, _scale1);
-#endif // __aarch64__
-                    vst1_s8(ptr, float2int8leakyrelu(_v0, _v1, _slope));
-
-                    intptr0 += 4;
-                    intptr1 += 4;
-                    ptr += 8;
-                }
-            }
-        }
-    }
-    if (out_elempack == 1)
-    {
-        if (bias_data_size == 0)
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < channels; q++)
-            {
-                const int* intptr = bottom_blob.channel(q);
-                signed char* ptr0 = top_blob.channel(q * 4);
-                signed char* ptr1 = top_blob.channel(q * 4 + 1);
-                signed char* ptr2 = top_blob.channel(q * 4 + 2);
-                signed char* ptr3 = top_blob.channel(q * 4 + 3);
-
-                float32x4_t _scale_in = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 4);
-                float32x4_t _scale_out = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 4);
-
-                float32x4_t _scale = vmulq_f32(_scale_in, _scale_out);
-                float32x4_t _slope = vdupq_n_f32(slope);
-
-                int i = 0;
-                for (; i < size; i++)
-                {
-                    float32x4_t _v = vcvtq_f32_s32(vld1q_s32(intptr));
-                    _v = vmulq_f32(_v, _scale);
-                    int8x8_t v = float2int8leakyrelu(_v, _v, _slope);
-                    ptr0[0] = vget_lane_s8(v, 0);
-                    ptr1[0] = vget_lane_s8(v, 1);
-                    ptr2[0] = vget_lane_s8(v, 2);
-                    ptr3[0] = vget_lane_s8(v, 3);
-
-                    intptr += 4;
-                    ptr0 += 1;
-                    ptr1 += 1;
-                    ptr2 += 1;
-                    ptr3 += 1;
-                }
-            }
-        }
-        else
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < channels; q++)
-            {
-                const int* intptr = bottom_blob.channel(q);
-                signed char* ptr0 = top_blob.channel(q * 4);
-                signed char* ptr1 = top_blob.channel(q * 4 + 1);
-                signed char* ptr2 = top_blob.channel(q * 4 + 2);
-                signed char* ptr3 = top_blob.channel(q * 4 + 3);
-
-                float32x4_t _scale_in = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 4);
-                float32x4_t _scale_out = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 4);
-                float32x4_t _bias = bias_data_size == 1 ? vdupq_n_f32(bias_data[0]) : vld1q_f32((const float*)bias_data + q * 4);
-
-                float32x4_t _scale = vmulq_f32(_scale_in, _scale_out);
-                _bias = vmulq_f32(_bias, _scale_out);
-                float32x4_t _slope = vdupq_n_f32(slope);
-
-                int i = 0;
-                for (; i < size; i++)
-                {
-                    float32x4_t _v = vcvtq_f32_s32(vld1q_s32(intptr));
-#if __aarch64__
-                    _v = vfmaq_f32(_bias, _v, _scale);
-#else
-                    _v = vmlaq_f32(_bias, _v, _scale);
-#endif
-                    int8x8_t v = float2int8leakyrelu(_v, _v, _slope);
-                    ptr0[0] = vget_lane_s8(v, 0);
-                    ptr1[0] = vget_lane_s8(v, 1);
-                    ptr2[0] = vget_lane_s8(v, 2);
-                    ptr3[0] = vget_lane_s8(v, 3);
-
-                    intptr += 4;
-                    ptr0 += 1;
-                    ptr1 += 1;
-                    ptr2 += 1;
-                    ptr3 += 1;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_leakyrelu_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_leakyrelu_pack8.h
deleted file mode 100644
index 9efee4dd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_leakyrelu_pack8.h
+++ /dev/null
@@ -1,202 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_leakyrelu_pack8_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, float slope, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-    int size = w * h;
-
-    int scale_in_data_size = scale_in_data.w;
-    int scale_out_data_size = scale_out_data.w;
-    int bias_data_size = bias_data.w;
-
-    // int8(leakyrelu(v * scale_in, slope) * scale_out)
-    // int8_leakyrelu(v * (scale_in * scale_out), slope)
-
-    // int8(leakyrelu(v * scale_in + bias, slope) * scale_out)
-    // int8_leakyrelu(v * (scale_in * scale_out) + (bias * scale_out), slope)
-
-    if (bias_data_size == 0)
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int q = 0; q < channels; q++)
-        {
-            const int* intptr = bottom_blob.channel(q);
-            signed char* ptr = top_blob.channel(q);
-
-            float32x4_t _scale_in0 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8);
-            float32x4_t _scale_in1 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8 + 4);
-            float32x4_t _scale_out0 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8);
-            float32x4_t _scale_out1 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8 + 4);
-
-            float32x4_t _scale0 = vmulq_f32(_scale_in0, _scale_out0);
-            float32x4_t _scale1 = vmulq_f32(_scale_in1, _scale_out1);
-            float32x4_t _slope = vdupq_n_f32(slope);
-
-            int i = 0;
-#if __aarch64__
-            for (; i + 3 < size; i += 4)
-            {
-                float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr));
-                float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr + 4));
-                float32x4_t _v2 = vcvtq_f32_s32(vld1q_s32(intptr + 8));
-                float32x4_t _v3 = vcvtq_f32_s32(vld1q_s32(intptr + 12));
-                float32x4_t _v4 = vcvtq_f32_s32(vld1q_s32(intptr + 16));
-                float32x4_t _v5 = vcvtq_f32_s32(vld1q_s32(intptr + 20));
-                float32x4_t _v6 = vcvtq_f32_s32(vld1q_s32(intptr + 24));
-                float32x4_t _v7 = vcvtq_f32_s32(vld1q_s32(intptr + 28));
-                _v0 = vmulq_f32(_v0, _scale0);
-                _v1 = vmulq_f32(_v1, _scale1);
-                _v2 = vmulq_f32(_v2, _scale0);
-                _v3 = vmulq_f32(_v3, _scale1);
-                _v4 = vmulq_f32(_v4, _scale0);
-                _v5 = vmulq_f32(_v5, _scale1);
-                _v6 = vmulq_f32(_v6, _scale0);
-                _v7 = vmulq_f32(_v7, _scale1);
-                vst1_s8(ptr, float2int8leakyrelu(_v0, _v1, _slope));
-                vst1_s8(ptr + 8, float2int8leakyrelu(_v2, _v3, _slope));
-                vst1_s8(ptr + 16, float2int8leakyrelu(_v4, _v5, _slope));
-                vst1_s8(ptr + 24, float2int8leakyrelu(_v6, _v7, _slope));
-
-                intptr += 32;
-                ptr += 32;
-            }
-#endif // __aarch64__
-            for (; i + 1 < size; i += 2)
-            {
-                float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr));
-                float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr + 4));
-                float32x4_t _v2 = vcvtq_f32_s32(vld1q_s32(intptr + 8));
-                float32x4_t _v3 = vcvtq_f32_s32(vld1q_s32(intptr + 12));
-                _v0 = vmulq_f32(_v0, _scale0);
-                _v1 = vmulq_f32(_v1, _scale1);
-                _v2 = vmulq_f32(_v2, _scale0);
-                _v3 = vmulq_f32(_v3, _scale1);
-                vst1_s8(ptr, float2int8leakyrelu(_v0, _v1, _slope));
-                vst1_s8(ptr + 8, float2int8leakyrelu(_v2, _v3, _slope));
-
-                intptr += 16;
-                ptr += 16;
-            }
-            for (; i < size; i++)
-            {
-                float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr));
-                float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr + 4));
-                _v0 = vmulq_f32(_v0, _scale0);
-                _v1 = vmulq_f32(_v1, _scale1);
-                vst1_s8(ptr, float2int8leakyrelu(_v0, _v1, _slope));
-
-                intptr += 8;
-                ptr += 8;
-            }
-        }
-    }
-    else
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int q = 0; q < channels; q++)
-        {
-            const int* intptr = bottom_blob.channel(q);
-            signed char* ptr = top_blob.channel(q);
-
-            float32x4_t _scale_in0 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8);
-            float32x4_t _scale_in1 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8 + 4);
-            float32x4_t _scale_out0 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8);
-            float32x4_t _scale_out1 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8 + 4);
-            float32x4_t _bias0 = bias_data_size == 1 ? vdupq_n_f32(bias_data[0]) : vld1q_f32((const float*)bias_data + q * 8);
-            float32x4_t _bias1 = bias_data_size == 1 ? vdupq_n_f32(bias_data[0]) : vld1q_f32((const float*)bias_data + q * 8 + 4);
-
-            float32x4_t _scale0 = vmulq_f32(_scale_in0, _scale_out0);
-            float32x4_t _scale1 = vmulq_f32(_scale_in1, _scale_out1);
-            _bias0 = vmulq_f32(_bias0, _scale_out0);
-            _bias1 = vmulq_f32(_bias1, _scale_out1);
-            float32x4_t _slope = vdupq_n_f32(slope);
-
-            int i = 0;
-#if __aarch64__
-            for (; i + 3 < size; i += 4)
-            {
-                float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr));
-                float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr + 4));
-                float32x4_t _v2 = vcvtq_f32_s32(vld1q_s32(intptr + 8));
-                float32x4_t _v3 = vcvtq_f32_s32(vld1q_s32(intptr + 12));
-                float32x4_t _v4 = vcvtq_f32_s32(vld1q_s32(intptr + 16));
-                float32x4_t _v5 = vcvtq_f32_s32(vld1q_s32(intptr + 20));
-                float32x4_t _v6 = vcvtq_f32_s32(vld1q_s32(intptr + 24));
-                float32x4_t _v7 = vcvtq_f32_s32(vld1q_s32(intptr + 28));
-
-                _v0 = vfmaq_f32(_bias0, _v0, _scale0);
-                _v1 = vfmaq_f32(_bias1, _v1, _scale1);
-                _v2 = vfmaq_f32(_bias0, _v2, _scale0);
-                _v3 = vfmaq_f32(_bias1, _v3, _scale1);
-                _v4 = vfmaq_f32(_bias0, _v4, _scale0);
-                _v5 = vfmaq_f32(_bias1, _v5, _scale1);
-                _v6 = vfmaq_f32(_bias0, _v6, _scale0);
-                _v7 = vfmaq_f32(_bias1, _v7, _scale1);
-
-                vst1_s8(ptr, float2int8leakyrelu(_v0, _v1, _slope));
-                vst1_s8(ptr + 8, float2int8leakyrelu(_v2, _v3, _slope));
-                vst1_s8(ptr + 16, float2int8leakyrelu(_v4, _v5, _slope));
-                vst1_s8(ptr + 24, float2int8leakyrelu(_v6, _v7, _slope));
-
-                intptr += 32;
-                ptr += 32;
-            }
-#endif // __aarch64__
-            for (; i + 1 < size; i += 2)
-            {
-                float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr));
-                float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr + 4));
-                float32x4_t _v2 = vcvtq_f32_s32(vld1q_s32(intptr + 8));
-                float32x4_t _v3 = vcvtq_f32_s32(vld1q_s32(intptr + 12));
-
-#if __aarch64__
-                _v0 = vfmaq_f32(_bias0, _v0, _scale0);
-                _v1 = vfmaq_f32(_bias1, _v1, _scale1);
-                _v2 = vfmaq_f32(_bias0, _v2, _scale0);
-                _v3 = vfmaq_f32(_bias1, _v3, _scale1);
-#else  // __aarch64__
-                _v0 = vmlaq_f32(_bias0, _v0, _scale0);
-                _v1 = vmlaq_f32(_bias1, _v1, _scale1);
-                _v2 = vmlaq_f32(_bias0, _v2, _scale0);
-                _v3 = vmlaq_f32(_bias1, _v3, _scale1);
-#endif // __aarch64__
-
-                vst1_s8(ptr, float2int8leakyrelu(_v0, _v1, _slope));
-                vst1_s8(ptr + 8, float2int8leakyrelu(_v2, _v3, _slope));
-
-                intptr += 16;
-                ptr += 16;
-            }
-            for (; i < size; i++)
-            {
-                float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr));
-                float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr + 4));
-#if __aarch64__
-                _v0 = vfmaq_f32(_bias0, _v0, _scale0);
-                _v1 = vfmaq_f32(_bias1, _v1, _scale1);
-#else  // __aarch64__
-                _v0 = vmlaq_f32(_bias0, _v0, _scale0);
-                _v1 = vmlaq_f32(_bias1, _v1, _scale1);
-#endif // __aarch64__
-                vst1_s8(ptr, float2int8leakyrelu(_v0, _v1, _slope));
-
-                intptr += 8;
-                ptr += 8;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_relu_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_relu_pack4.h
deleted file mode 100644
index 961562a3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_relu_pack4.h
+++ /dev/null
@@ -1,271 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_relu_pack4_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-    int size = w * h;
-    int outc = top_blob.c;
-    int out_elempack = top_blob.elempack;
-
-    int scale_in_data_size = scale_in_data.w;
-    int scale_out_data_size = scale_out_data.w;
-    int bias_data_size = bias_data.w;
-
-    // int8(relu(v * scale_in) * scale_out)
-    // int8_relu(v * (scale_in * scale_out))
-
-    // int8(relu(v * scale_in + bias) * scale_out)
-    // int8_relu(v * (scale_in * scale_out) + (bias * scale_out))
-
-    if (out_elempack == 8)
-    {
-        if (bias_data_size == 0)
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < outc; q++)
-            {
-                const int* intptr0 = bottom_blob.channel(q * 2);
-                const int* intptr1 = bottom_blob.channel(q * 2 + 1);
-                signed char* ptr = top_blob.channel(q);
-
-                float32x4_t _scale_in0 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8);
-                float32x4_t _scale_in1 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8 + 4);
-                float32x4_t _scale_out0 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8);
-                float32x4_t _scale_out1 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8 + 4);
-
-                float32x4_t _scale0 = vmulq_f32(_scale_in0, _scale_out0);
-                float32x4_t _scale1 = vmulq_f32(_scale_in1, _scale_out1);
-
-                int i = 0;
-#if __aarch64__
-                for (; i + 3 < size; i += 4)
-                {
-                    float32x4_t _v00 = vcvtq_f32_s32(vld1q_s32(intptr0));
-                    float32x4_t _v01 = vcvtq_f32_s32(vld1q_s32(intptr0 + 4));
-                    float32x4_t _v02 = vcvtq_f32_s32(vld1q_s32(intptr0 + 8));
-                    float32x4_t _v03 = vcvtq_f32_s32(vld1q_s32(intptr0 + 12));
-                    float32x4_t _v10 = vcvtq_f32_s32(vld1q_s32(intptr1));
-                    float32x4_t _v11 = vcvtq_f32_s32(vld1q_s32(intptr1 + 4));
-                    float32x4_t _v12 = vcvtq_f32_s32(vld1q_s32(intptr1 + 8));
-                    float32x4_t _v13 = vcvtq_f32_s32(vld1q_s32(intptr1 + 12));
-                    _v00 = vmulq_f32(_v00, _scale0);
-                    _v01 = vmulq_f32(_v01, _scale0);
-                    _v02 = vmulq_f32(_v02, _scale0);
-                    _v03 = vmulq_f32(_v03, _scale0);
-                    _v10 = vmulq_f32(_v10, _scale1);
-                    _v11 = vmulq_f32(_v11, _scale1);
-                    _v12 = vmulq_f32(_v12, _scale1);
-                    _v13 = vmulq_f32(_v13, _scale1);
-                    vst1_s8(ptr, float2int8relu(_v00, _v10));
-                    vst1_s8(ptr + 8, float2int8relu(_v01, _v11));
-                    vst1_s8(ptr + 16, float2int8relu(_v02, _v12));
-                    vst1_s8(ptr + 24, float2int8relu(_v03, _v13));
-
-                    intptr0 += 16;
-                    intptr1 += 16;
-                    ptr += 32;
-                }
-#endif // __aarch64__
-                for (; i < size; i++)
-                {
-                    float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr0));
-                    float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr1));
-                    _v0 = vmulq_f32(_v0, _scale0);
-                    _v1 = vmulq_f32(_v1, _scale1);
-                    vst1_s8(ptr, float2int8relu(_v0, _v1));
-
-                    intptr0 += 4;
-                    intptr1 += 4;
-                    ptr += 8;
-                }
-            }
-        }
-        else
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < outc; q++)
-            {
-                const int* intptr0 = bottom_blob.channel(q * 2);
-                const int* intptr1 = bottom_blob.channel(q * 2 + 1);
-                signed char* ptr = top_blob.channel(q);
-
-                float32x4_t _scale_in0 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8);
-                float32x4_t _scale_in1 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8 + 4);
-                float32x4_t _scale_out0 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8);
-                float32x4_t _scale_out1 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8 + 4);
-                float32x4_t _bias0 = bias_data_size == 1 ? vdupq_n_f32(bias_data[0]) : vld1q_f32((const float*)bias_data + q * 8);
-                float32x4_t _bias1 = bias_data_size == 1 ? vdupq_n_f32(bias_data[0]) : vld1q_f32((const float*)bias_data + q * 8 + 4);
-
-                float32x4_t _scale0 = vmulq_f32(_scale_in0, _scale_out0);
-                float32x4_t _scale1 = vmulq_f32(_scale_in1, _scale_out1);
-                _bias0 = vmulq_f32(_bias0, _scale_out0);
-                _bias1 = vmulq_f32(_bias1, _scale_out1);
-
-                int i = 0;
-#if __aarch64__
-                for (; i + 3 < size; i += 4)
-                {
-                    float32x4_t _v00 = vcvtq_f32_s32(vld1q_s32(intptr0));
-                    float32x4_t _v01 = vcvtq_f32_s32(vld1q_s32(intptr0 + 4));
-                    float32x4_t _v02 = vcvtq_f32_s32(vld1q_s32(intptr0 + 8));
-                    float32x4_t _v03 = vcvtq_f32_s32(vld1q_s32(intptr0 + 12));
-                    float32x4_t _v10 = vcvtq_f32_s32(vld1q_s32(intptr1));
-                    float32x4_t _v11 = vcvtq_f32_s32(vld1q_s32(intptr1 + 4));
-                    float32x4_t _v12 = vcvtq_f32_s32(vld1q_s32(intptr1 + 8));
-                    float32x4_t _v13 = vcvtq_f32_s32(vld1q_s32(intptr1 + 12));
-                    _v00 = vfmaq_f32(_bias0, _v00, _scale0);
-                    _v01 = vfmaq_f32(_bias0, _v01, _scale0);
-                    _v02 = vfmaq_f32(_bias0, _v02, _scale0);
-                    _v03 = vfmaq_f32(_bias0, _v03, _scale0);
-                    _v10 = vfmaq_f32(_bias1, _v10, _scale1);
-                    _v11 = vfmaq_f32(_bias1, _v11, _scale1);
-                    _v12 = vfmaq_f32(_bias1, _v12, _scale1);
-                    _v13 = vfmaq_f32(_bias1, _v13, _scale1);
-                    vst1_s8(ptr, float2int8relu(_v00, _v10));
-                    vst1_s8(ptr + 8, float2int8relu(_v01, _v11));
-                    vst1_s8(ptr + 16, float2int8relu(_v02, _v12));
-                    vst1_s8(ptr + 24, float2int8relu(_v03, _v13));
-
-                    intptr0 += 16;
-                    intptr1 += 16;
-                    ptr += 32;
-                }
-#endif // __aarch64__
-                for (; i + 1 < size; i += 2)
-                {
-                    float32x4_t _v00 = vcvtq_f32_s32(vld1q_s32(intptr0));
-                    float32x4_t _v01 = vcvtq_f32_s32(vld1q_s32(intptr0 + 4));
-                    float32x4_t _v10 = vcvtq_f32_s32(vld1q_s32(intptr1));
-                    float32x4_t _v11 = vcvtq_f32_s32(vld1q_s32(intptr1 + 4));
-#if __aarch64__
-                    _v00 = vfmaq_f32(_bias0, _v00, _scale0);
-                    _v01 = vfmaq_f32(_bias0, _v01, _scale0);
-                    _v10 = vfmaq_f32(_bias1, _v10, _scale1);
-                    _v11 = vfmaq_f32(_bias1, _v11, _scale1);
-#else  // __aarch64__
-                    _v00 = vmlaq_f32(_bias0, _v00, _scale0);
-                    _v01 = vmlaq_f32(_bias0, _v01, _scale0);
-                    _v10 = vmlaq_f32(_bias1, _v10, _scale1);
-                    _v11 = vmlaq_f32(_bias1, _v11, _scale1);
-#endif // __aarch64__
-                    vst1_s8(ptr, float2int8relu(_v00, _v10));
-                    vst1_s8(ptr + 8, float2int8relu(_v01, _v11));
-
-                    intptr0 += 8;
-                    intptr1 += 8;
-                    ptr += 16;
-                }
-                for (; i < size; i++)
-                {
-                    float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr0));
-                    float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr1));
-#if __aarch64__
-                    _v0 = vfmaq_f32(_bias0, _v0, _scale0);
-                    _v1 = vfmaq_f32(_bias1, _v1, _scale1);
-#else  // __aarch64__
-                    _v0 = vmlaq_f32(_bias0, _v0, _scale0);
-                    _v1 = vmlaq_f32(_bias1, _v1, _scale1);
-#endif // __aarch64__
-                    vst1_s8(ptr, float2int8relu(_v0, _v1));
-
-                    intptr0 += 4;
-                    intptr1 += 4;
-                    ptr += 8;
-                }
-            }
-        }
-    }
-    if (out_elempack == 1)
-    {
-        if (bias_data_size == 0)
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < channels; q++)
-            {
-                const int* intptr = bottom_blob.channel(q);
-                signed char* ptr0 = top_blob.channel(q * 4);
-                signed char* ptr1 = top_blob.channel(q * 4 + 1);
-                signed char* ptr2 = top_blob.channel(q * 4 + 2);
-                signed char* ptr3 = top_blob.channel(q * 4 + 3);
-
-                float32x4_t _scale_in = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 4);
-                float32x4_t _scale_out = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 4);
-
-                float32x4_t _scale = vmulq_f32(_scale_in, _scale_out);
-
-                int i = 0;
-                for (; i < size; i++)
-                {
-                    float32x4_t _v = vcvtq_f32_s32(vld1q_s32(intptr));
-                    _v = vmulq_f32(_v, _scale);
-                    int8x8_t v = float2int8relu(_v, _v);
-                    ptr0[0] = vget_lane_s8(v, 0);
-                    ptr1[0] = vget_lane_s8(v, 1);
-                    ptr2[0] = vget_lane_s8(v, 2);
-                    ptr3[0] = vget_lane_s8(v, 3);
-
-                    intptr += 4;
-                    ptr0 += 1;
-                    ptr1 += 1;
-                    ptr2 += 1;
-                    ptr3 += 1;
-                }
-            }
-        }
-        else
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < channels; q++)
-            {
-                const int* intptr = bottom_blob.channel(q);
-                signed char* ptr0 = top_blob.channel(q * 4);
-                signed char* ptr1 = top_blob.channel(q * 4 + 1);
-                signed char* ptr2 = top_blob.channel(q * 4 + 2);
-                signed char* ptr3 = top_blob.channel(q * 4 + 3);
-
-                float32x4_t _scale_in = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 4);
-                float32x4_t _scale_out = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 4);
-                float32x4_t _bias = bias_data_size == 1 ? vdupq_n_f32(bias_data[0]) : vld1q_f32((const float*)bias_data + q * 4);
-
-                float32x4_t _scale = vmulq_f32(_scale_in, _scale_out);
-                _bias = vmulq_f32(_bias, _scale_out);
-
-                int i = 0;
-                for (; i < size; i++)
-                {
-                    float32x4_t _v = vcvtq_f32_s32(vld1q_s32(intptr));
-#if __aarch64__
-                    _v = vfmaq_f32(_bias, _v, _scale);
-#else
-                    _v = vmlaq_f32(_bias, _v, _scale);
-#endif
-                    int8x8_t v = float2int8relu(_v, _v);
-                    ptr0[0] = vget_lane_s8(v, 0);
-                    ptr1[0] = vget_lane_s8(v, 1);
-                    ptr2[0] = vget_lane_s8(v, 2);
-                    ptr3[0] = vget_lane_s8(v, 3);
-
-                    intptr += 4;
-                    ptr0 += 1;
-                    ptr1 += 1;
-                    ptr2 += 1;
-                    ptr3 += 1;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_relu_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_relu_pack8.h
deleted file mode 100644
index e93ec83b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/requantize_relu_pack8.h
+++ /dev/null
@@ -1,200 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_relu_pack8_neon(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-    int size = w * h;
-
-    int scale_in_data_size = scale_in_data.w;
-    int scale_out_data_size = scale_out_data.w;
-    int bias_data_size = bias_data.w;
-
-    // int8(relu(v * scale_in) * scale_out)
-    // int8_relu(v * (scale_in * scale_out))
-
-    // int8(relu(v * scale_in + bias) * scale_out)
-    // int8_relu(v * (scale_in * scale_out) + (bias * scale_out))
-
-    if (bias_data_size == 0)
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int q = 0; q < channels; q++)
-        {
-            const int* intptr = bottom_blob.channel(q);
-            signed char* ptr = top_blob.channel(q);
-
-            float32x4_t _scale_in0 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8);
-            float32x4_t _scale_in1 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8 + 4);
-            float32x4_t _scale_out0 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8);
-            float32x4_t _scale_out1 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8 + 4);
-
-            float32x4_t _scale0 = vmulq_f32(_scale_in0, _scale_out0);
-            float32x4_t _scale1 = vmulq_f32(_scale_in1, _scale_out1);
-
-            int i = 0;
-#if __aarch64__
-            for (; i + 3 < size; i += 4)
-            {
-                float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr));
-                float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr + 4));
-                float32x4_t _v2 = vcvtq_f32_s32(vld1q_s32(intptr + 8));
-                float32x4_t _v3 = vcvtq_f32_s32(vld1q_s32(intptr + 12));
-                float32x4_t _v4 = vcvtq_f32_s32(vld1q_s32(intptr + 16));
-                float32x4_t _v5 = vcvtq_f32_s32(vld1q_s32(intptr + 20));
-                float32x4_t _v6 = vcvtq_f32_s32(vld1q_s32(intptr + 24));
-                float32x4_t _v7 = vcvtq_f32_s32(vld1q_s32(intptr + 28));
-                _v0 = vmulq_f32(_v0, _scale0);
-                _v1 = vmulq_f32(_v1, _scale1);
-                _v2 = vmulq_f32(_v2, _scale0);
-                _v3 = vmulq_f32(_v3, _scale1);
-                _v4 = vmulq_f32(_v4, _scale0);
-                _v5 = vmulq_f32(_v5, _scale1);
-                _v6 = vmulq_f32(_v6, _scale0);
-                _v7 = vmulq_f32(_v7, _scale1);
-                vst1_s8(ptr, float2int8relu(_v0, _v1));
-                vst1_s8(ptr + 8, float2int8relu(_v2, _v3));
-                vst1_s8(ptr + 16, float2int8relu(_v4, _v5));
-                vst1_s8(ptr + 24, float2int8relu(_v6, _v7));
-
-                intptr += 32;
-                ptr += 32;
-            }
-#endif // __aarch64__
-            for (; i + 1 < size; i += 2)
-            {
-                float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr));
-                float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr + 4));
-                float32x4_t _v2 = vcvtq_f32_s32(vld1q_s32(intptr + 8));
-                float32x4_t _v3 = vcvtq_f32_s32(vld1q_s32(intptr + 12));
-                _v0 = vmulq_f32(_v0, _scale0);
-                _v1 = vmulq_f32(_v1, _scale1);
-                _v2 = vmulq_f32(_v2, _scale0);
-                _v3 = vmulq_f32(_v3, _scale1);
-                vst1_s8(ptr, float2int8relu(_v0, _v1));
-                vst1_s8(ptr + 8, float2int8relu(_v2, _v3));
-
-                intptr += 16;
-                ptr += 16;
-            }
-            for (; i < size; i++)
-            {
-                float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr));
-                float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr + 4));
-                _v0 = vmulq_f32(_v0, _scale0);
-                _v1 = vmulq_f32(_v1, _scale1);
-                vst1_s8(ptr, float2int8relu(_v0, _v1));
-
-                intptr += 8;
-                ptr += 8;
-            }
-        }
-    }
-    else
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int q = 0; q < channels; q++)
-        {
-            const int* intptr = bottom_blob.channel(q);
-            signed char* ptr = top_blob.channel(q);
-
-            float32x4_t _scale_in0 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8);
-            float32x4_t _scale_in1 = scale_in_data_size == 1 ? vdupq_n_f32(scale_in_data[0]) : vld1q_f32((const float*)scale_in_data + q * 8 + 4);
-            float32x4_t _scale_out0 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8);
-            float32x4_t _scale_out1 = scale_out_data_size == 1 ? vdupq_n_f32(scale_out_data[0]) : vld1q_f32((const float*)scale_out_data + q * 8 + 4);
-            float32x4_t _bias0 = bias_data_size == 1 ? vdupq_n_f32(bias_data[0]) : vld1q_f32((const float*)bias_data + q * 8);
-            float32x4_t _bias1 = bias_data_size == 1 ? vdupq_n_f32(bias_data[0]) : vld1q_f32((const float*)bias_data + q * 8 + 4);
-
-            float32x4_t _scale0 = vmulq_f32(_scale_in0, _scale_out0);
-            float32x4_t _scale1 = vmulq_f32(_scale_in1, _scale_out1);
-            _bias0 = vmulq_f32(_bias0, _scale_out0);
-            _bias1 = vmulq_f32(_bias1, _scale_out1);
-
-            int i = 0;
-#if __aarch64__
-            for (; i + 3 < size; i += 4)
-            {
-                float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr));
-                float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr + 4));
-                float32x4_t _v2 = vcvtq_f32_s32(vld1q_s32(intptr + 8));
-                float32x4_t _v3 = vcvtq_f32_s32(vld1q_s32(intptr + 12));
-                float32x4_t _v4 = vcvtq_f32_s32(vld1q_s32(intptr + 16));
-                float32x4_t _v5 = vcvtq_f32_s32(vld1q_s32(intptr + 20));
-                float32x4_t _v6 = vcvtq_f32_s32(vld1q_s32(intptr + 24));
-                float32x4_t _v7 = vcvtq_f32_s32(vld1q_s32(intptr + 28));
-
-                _v0 = vfmaq_f32(_bias0, _v0, _scale0);
-                _v1 = vfmaq_f32(_bias1, _v1, _scale1);
-                _v2 = vfmaq_f32(_bias0, _v2, _scale0);
-                _v3 = vfmaq_f32(_bias1, _v3, _scale1);
-                _v4 = vfmaq_f32(_bias0, _v4, _scale0);
-                _v5 = vfmaq_f32(_bias1, _v5, _scale1);
-                _v6 = vfmaq_f32(_bias0, _v6, _scale0);
-                _v7 = vfmaq_f32(_bias1, _v7, _scale1);
-
-                vst1_s8(ptr, float2int8relu(_v0, _v1));
-                vst1_s8(ptr + 8, float2int8relu(_v2, _v3));
-                vst1_s8(ptr + 16, float2int8relu(_v4, _v5));
-                vst1_s8(ptr + 24, float2int8relu(_v6, _v7));
-
-                intptr += 32;
-                ptr += 32;
-            }
-#endif // __aarch64__
-            for (; i + 1 < size; i += 2)
-            {
-                float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr));
-                float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr + 4));
-                float32x4_t _v2 = vcvtq_f32_s32(vld1q_s32(intptr + 8));
-                float32x4_t _v3 = vcvtq_f32_s32(vld1q_s32(intptr + 12));
-
-#if __aarch64__
-                _v0 = vfmaq_f32(_bias0, _v0, _scale0);
-                _v1 = vfmaq_f32(_bias1, _v1, _scale1);
-                _v2 = vfmaq_f32(_bias0, _v2, _scale0);
-                _v3 = vfmaq_f32(_bias1, _v3, _scale1);
-#else  // __aarch64__
-                _v0 = vmlaq_f32(_bias0, _v0, _scale0);
-                _v1 = vmlaq_f32(_bias1, _v1, _scale1);
-                _v2 = vmlaq_f32(_bias0, _v2, _scale0);
-                _v3 = vmlaq_f32(_bias1, _v3, _scale1);
-#endif // __aarch64__
-
-                vst1_s8(ptr, float2int8relu(_v0, _v1));
-                vst1_s8(ptr + 8, float2int8relu(_v2, _v3));
-
-                intptr += 16;
-                ptr += 16;
-            }
-            for (; i < size; i++)
-            {
-                float32x4_t _v0 = vcvtq_f32_s32(vld1q_s32(intptr));
-                float32x4_t _v1 = vcvtq_f32_s32(vld1q_s32(intptr + 4));
-#if __aarch64__
-                _v0 = vfmaq_f32(_bias0, _v0, _scale0);
-                _v1 = vfmaq_f32(_bias1, _v1, _scale1);
-#else  // __aarch64__
-                _v0 = vmlaq_f32(_bias0, _v0, _scale0);
-                _v1 = vmlaq_f32(_bias1, _v1, _scale1);
-#endif // __aarch64__
-                vst1_s8(ptr, float2int8relu(_v0, _v1));
-
-                intptr += 8;
-                ptr += 8;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/reshape_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/reshape_arm.h
deleted file mode 100644
index 85466ecf..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/reshape_arm.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_RESHAPE_ARM_H
-#define LAYER_RESHAPE_ARM_H
-
-#include "reshape.h"
-
-namespace ncnn {
-
-class Reshape_arm : public Reshape
-{
-public:
-    Reshape_arm();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_bf16s_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_RESHAPE_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/rnn_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/rnn_arm.h
deleted file mode 100644
index 18e75642..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/rnn_arm.h
+++ /dev/null
@@ -1,53 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_RNN_ARM_H
-#define LAYER_RNN_ARM_H
-
-#include "rnn.h"
-
-namespace ncnn {
-
-class RNN_arm : public RNN
-{
-public:
-    RNN_arm();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int create_pipeline_bf16s(const Option& opt);
-    int forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_bf16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-#endif
-
-public:
-    Mat weight_xc_data_packed;
-    Mat bias_c_data_packed;
-    Mat weight_hc_data_packed;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_RNN_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/scale_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/scale_arm.h
deleted file mode 100644
index c540cdd6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/scale_arm.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SCALE_ARM_H
-#define LAYER_SCALE_ARM_H
-
-#include "scale.h"
-
-namespace ncnn {
-
-class Scale_arm : public Scale
-{
-public:
-    Scale_arm();
-
-    virtual int forward_inplace(std::vector<Mat>& bottom_top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SCALE_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/selu_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/selu_arm.h
deleted file mode 100644
index d951804d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/selu_arm.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SELU_ARM_H
-#define LAYER_SELU_ARM_H
-
-#include "selu.h"
-
-namespace ncnn {
-
-class SELU_arm : public SELU
-{
-public:
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SELU_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/shufflechannel_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/shufflechannel_arm.h
deleted file mode 100644
index dcdbf760..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/shufflechannel_arm.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SHUFFLECHANNEL_ARM_H
-#define LAYER_SHUFFLECHANNEL_ARM_H
-
-#include "shufflechannel.h"
-
-namespace ncnn {
-
-class ShuffleChannel_arm : public ShuffleChannel
-{
-public:
-    ShuffleChannel_arm();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_bf16s_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SHUFFLECHANNEL_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/sigmoid_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/sigmoid_arm.h
deleted file mode 100644
index 4c3901ab..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/sigmoid_arm.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SIGMOID_ARM_H
-#define LAYER_SIGMOID_ARM_H
-
-#include "sigmoid.h"
-
-namespace ncnn {
-
-class Sigmoid_arm : public Sigmoid
-{
-public:
-    Sigmoid_arm();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_inplace_bf16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SIGMOID_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/slice_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/slice_arm.h
deleted file mode 100644
index c3b558b9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/slice_arm.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SLICE_ARM_H
-#define LAYER_SLICE_ARM_H
-
-#include "slice.h"
-
-namespace ncnn {
-
-class Slice_arm : public Slice
-{
-public:
-    Slice_arm();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    int forward_bf16s_fp16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SLICE_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/softmax_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/softmax_arm.h
deleted file mode 100644
index 78c54084..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/softmax_arm.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SOFTMAX_ARM_H
-#define LAYER_SOFTMAX_ARM_H
-
-#include "softmax.h"
-
-namespace ncnn {
-
-class Softmax_arm : public Softmax
-{
-public:
-    Softmax_arm();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_inplace_bf16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SOFTMAX_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/swish_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/swish_arm.h
deleted file mode 100644
index 907d7970..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/swish_arm.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SWISH_ARM_H
-#define LAYER_SWISH_ARM_H
-
-#include "swish.h"
-
-namespace ncnn {
-
-class Swish_arm : public Swish
-{
-public:
-    Swish_arm();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_inplace_bf16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SWISH_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/tanh_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/tanh_arm.h
deleted file mode 100644
index db62f117..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/tanh_arm.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_TANH_ARM_H
-#define LAYER_TANH_ARM_H
-
-#include "tanh.h"
-
-namespace ncnn {
-
-class TanH_arm : public TanH
-{
-public:
-    TanH_arm();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_inplace_bf16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_TANH_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/unaryop_arm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/unaryop_arm.h
deleted file mode 100644
index ab4b23c0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/arm/unaryop_arm.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_UNARYOP_ARM_H
-#define LAYER_UNARYOP_ARM_H
-
-#include "unaryop.h"
-
-namespace ncnn {
-
-class UnaryOp_arm : public UnaryOp
-{
-public:
-    UnaryOp_arm();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_ARM82
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-#if NCNN_BF16
-    int forward_inplace_bf16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_UNARYOP_ARM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/batchnorm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/batchnorm.h
deleted file mode 100644
index e2b56cda..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/batchnorm.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BATCHNORM_H
-#define LAYER_BATCHNORM_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class BatchNorm : public Layer
-{
-public:
-    BatchNorm();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    // param
-    int channels;
-    float eps;
-
-    // model
-    Mat slope_data;
-    Mat mean_data;
-    Mat var_data;
-    Mat bias_data;
-
-    Mat a_data;
-    Mat b_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BATCHNORM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/bias.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/bias.h
deleted file mode 100644
index f540d34e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/bias.h
+++ /dev/null
@@ -1,43 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BIAS_H
-#define LAYER_BIAS_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Bias : public Layer
-{
-public:
-    Bias();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    // param
-    int bias_data_size;
-
-    // model
-    Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BIAS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/binaryop.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/binaryop.h
deleted file mode 100644
index 5fc06918..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/binaryop.h
+++ /dev/null
@@ -1,60 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BINARYOP_H
-#define LAYER_BINARYOP_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class BinaryOp : public Layer
-{
-public:
-    BinaryOp();
-
-    virtual int load_param(const ParamDict& pd);
-
-    using Layer::forward;
-    using Layer::forward_inplace;
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-    enum OperationType
-    {
-        Operation_ADD = 0,
-        Operation_SUB = 1,
-        Operation_MUL = 2,
-        Operation_DIV = 3,
-        Operation_MAX = 4,
-        Operation_MIN = 5,
-        Operation_POW = 6,
-        Operation_RSUB = 7,
-        Operation_RDIV = 8,
-        Operation_RPOW = 9,
-        Operation_ATAN2 = 10,
-        Operation_RATAN2 = 11
-    };
-
-public:
-    // param
-    int op_type;
-    int with_scalar;
-    float b;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BINARYOP_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/bnll.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/bnll.h
deleted file mode 100644
index e1293aa9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/bnll.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BNLL_H
-#define LAYER_BNLL_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class BNLL : public Layer
-{
-public:
-    BNLL();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BNLL_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/cast.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/cast.h
deleted file mode 100644
index e178e5bc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/cast.h
+++ /dev/null
@@ -1,44 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CAST_H
-#define LAYER_CAST_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Cast : public Layer
-{
-public:
-    Cast();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    // element type
-    // 0 = auto
-    // 1 = float32
-    // 2 = float16
-    // 3 = int8
-    // 4 = bfloat16
-    int type_from;
-    int type_to;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CAST_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/celu.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/celu.h
deleted file mode 100644
index 5e4b9c87..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/celu.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CELU_H
-#define LAYER_CELU_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class CELU : public Layer
-{
-public:
-    CELU();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    float alpha;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CELU_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/clip.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/clip.h
deleted file mode 100644
index 608964c0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/clip.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2018 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CLIP_H
-#define LAYER_CLIP_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Clip : public Layer
-{
-public:
-    Clip();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    float min;
-    float max;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CLIP_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/concat.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/concat.h
deleted file mode 100644
index bf2fe77e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/concat.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONCAT_H
-#define LAYER_CONCAT_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Concat : public Layer
-{
-public:
-    Concat();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    int axis;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONCAT_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolution.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolution.h
deleted file mode 100644
index 7af0735f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolution.h
+++ /dev/null
@@ -1,82 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTION_H
-#define LAYER_CONVOLUTION_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Convolution : public Layer
-{
-public:
-    Convolution();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    void make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, const Option& opt) const;
-    void make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, int kernel_w, int kernel_h, const Option& opt) const;
-
-#if NCNN_INT8
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    // param
-    int num_output;
-    int kernel_w;
-    int kernel_h;
-    int dilation_w;
-    int dilation_h;
-    int stride_w;
-    int stride_h;
-    int pad_left; // -233=SAME_UPPER -234=SAME_LOWER
-    int pad_right;
-    int pad_top;
-    int pad_bottom;
-    float pad_value;
-    int bias_term;
-
-    int weight_data_size;
-
-    int int8_scale_term;
-
-    // 0=none 1=relu 2=leakyrelu 3=clip 4=sigmoid
-    int activation_type;
-    Mat activation_params;
-
-    int dynamic_weight;
-
-    // model
-    Mat weight_data;
-    Mat bias_data;
-
-#if NCNN_INT8
-    Mat weight_data_int8_scales;
-    Mat bottom_blob_int8_scales;
-    Mat top_blob_int8_scales;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTION_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolution1d.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolution1d.h
deleted file mode 100644
index d87099e2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolution1d.h
+++ /dev/null
@@ -1,65 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTION1D_H
-#define LAYER_CONVOLUTION1D_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Convolution1D : public Layer
-{
-public:
-    Convolution1D();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    void make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, const Option& opt) const;
-    void make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, int kernel_w, const Option& opt) const;
-
-public:
-    // param
-    int num_output;
-    int kernel_w;
-    int dilation_w;
-    int stride_w;
-    int pad_left; // -233=SAME_UPPER -234=SAME_LOWER
-    int pad_right;
-    float pad_value;
-    int bias_term;
-
-    int weight_data_size;
-
-    // 0=none 1=relu 2=leakyrelu 3=clip 4=sigmoid
-    int activation_type;
-    Mat activation_params;
-
-    int dynamic_weight;
-
-    // model
-    Mat weight_data;
-    Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTION1D_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolution3d.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolution3d.h
deleted file mode 100644
index aee37742..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolution3d.h
+++ /dev/null
@@ -1,68 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTION3D_H
-#define LAYER_CONVOLUTION3D_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Convolution3D : public Layer
-{
-public:
-    Convolution3D();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    void make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, const Option& opt) const;
-
-public:
-    // param
-    int num_output;
-    int kernel_w;
-    int kernel_h;
-    int kernel_d;
-    int dilation_w;
-    int dilation_h;
-    int dilation_d;
-    int stride_w;
-    int stride_h;
-    int stride_d;
-    int pad_left; // -233=SAME_UPPER -234=SAME_LOWER
-    int pad_right;
-    int pad_top;
-    int pad_bottom;
-    int pad_front;
-    int pad_behind;
-    float pad_value;
-    int bias_term;
-
-    int weight_data_size;
-
-    int activation_type;
-    Mat activation_params;
-
-    Mat weight_data;
-    Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif //LAYER_CONVOLUTION3D_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolutiondepthwise.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolutiondepthwise.h
deleted file mode 100644
index 8a955dbd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolutiondepthwise.h
+++ /dev/null
@@ -1,83 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTIONDEPTHWISE_H
-#define LAYER_CONVOLUTIONDEPTHWISE_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class ConvolutionDepthWise : public Layer
-{
-public:
-    ConvolutionDepthWise();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    void make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, const Option& opt) const;
-    void make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, int kernel_w, int kernel_h, const Option& opt) const;
-
-#if NCNN_INT8
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    // param
-    int num_output;
-    int kernel_w;
-    int kernel_h;
-    int dilation_w;
-    int dilation_h;
-    int stride_w;
-    int stride_h;
-    int pad_left; // -233=SAME_UPPER -234=SAME_LOWER
-    int pad_right;
-    int pad_top;
-    int pad_bottom;
-    float pad_value;
-    int bias_term;
-
-    int weight_data_size;
-    int group;
-
-    int int8_scale_term;
-
-    // 0=none 1=relu 2=leakyrelu 3=clip 4=sigmoid
-    int activation_type;
-    Mat activation_params;
-
-    int dynamic_weight;
-
-    // model
-    Mat weight_data;
-    Mat bias_data;
-
-#if NCNN_INT8
-    Mat weight_data_int8_scales;
-    Mat bottom_blob_int8_scales;
-    Mat top_blob_int8_scales;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTIONDEPTHWISE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolutiondepthwise1d.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolutiondepthwise1d.h
deleted file mode 100644
index 6026f049..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolutiondepthwise1d.h
+++ /dev/null
@@ -1,66 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTIONDEPTHWISE1D_H
-#define LAYER_CONVOLUTIONDEPTHWISE1D_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class ConvolutionDepthWise1D : public Layer
-{
-public:
-    ConvolutionDepthWise1D();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    void make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, const Option& opt) const;
-    void make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, int kernel_w, const Option& opt) const;
-
-public:
-    // param
-    int num_output;
-    int kernel_w;
-    int dilation_w;
-    int stride_w;
-    int pad_left; // -233=SAME_UPPER -234=SAME_LOWER
-    int pad_right;
-    float pad_value;
-    int bias_term;
-
-    int weight_data_size;
-    int group;
-
-    // 0=none 1=relu 2=leakyrelu 3=clip 4=sigmoid
-    int activation_type;
-    Mat activation_params;
-
-    int dynamic_weight;
-
-    // model
-    Mat weight_data;
-    Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTIONDEPTHWISE1D_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolutiondepthwise3d.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolutiondepthwise3d.h
deleted file mode 100644
index e8aa1db0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/convolutiondepthwise3d.h
+++ /dev/null
@@ -1,69 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTIONDEPTHWISE3D_H
-#define LAYER_CONVOLUTIONDEPTHWISE3D_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class ConvolutionDepthWise3D : public Layer
-{
-public:
-    ConvolutionDepthWise3D();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    void make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, const Option& opt) const;
-
-public:
-    // param
-    int num_output;
-    int kernel_w;
-    int kernel_h;
-    int kernel_d;
-    int dilation_w;
-    int dilation_h;
-    int dilation_d;
-    int stride_w;
-    int stride_h;
-    int stride_d;
-    int pad_left; // -233=SAME_UPPER -234=SAME_LOWER
-    int pad_right;
-    int pad_top;
-    int pad_bottom;
-    int pad_front;
-    int pad_behind;
-    float pad_value;
-    int bias_term;
-
-    int weight_data_size;
-    int group;
-
-    int activation_type;
-    Mat activation_params;
-
-    Mat weight_data;
-    Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif //LAYER_CONVOLUTIONDEPTHWISE3D_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/copyto.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/copyto.h
deleted file mode 100644
index f1a6eb3a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/copyto.h
+++ /dev/null
@@ -1,48 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_COPYTO_H
-#define LAYER_COPYTO_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class CopyTo : public Layer
-{
-public:
-    CopyTo();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    void resolve_copyto_offset(const Mat& self_blob, int& woffset, int& hoffset, int& doffset, int& coffset) const;
-
-public:
-    int woffset;
-    int hoffset;
-    int doffset;
-    int coffset;
-
-    // numpy-style slice
-    // if provided, all the above attributes will be ignored
-    Mat starts;
-    Mat axes;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_COPYTO_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/crop.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/crop.h
deleted file mode 100644
index 826fe1bf..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/crop.h
+++ /dev/null
@@ -1,67 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CROP_H
-#define LAYER_CROP_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Crop : public Layer
-{
-public:
-    Crop();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    void resolve_crop_roi(const Mat& bottom_blob, int& woffset, int& hoffset, int& doffset, int& coffset, int& outw, int& outh, int& outd, int& outc) const;
-    void resolve_crop_roi(const Mat& bottom_blob, const Mat& reference_blob, int& woffset, int& hoffset, int& doffset, int& coffset, int& outw, int& outh, int& outd, int& outc) const;
-    void resolve_crop_roi(const Mat& bottom_blob, const int* param_data, int& woffset, int& hoffset, int& doffset, int& coffset, int& outw, int& outh, int& outd, int& outc) const;
-
-public:
-    // -233 = dynamic offset from reference blob
-    int woffset;
-    int hoffset;
-    int doffset;
-    int coffset;
-
-    // -233 = remaining
-    int outw;
-    int outh;
-    int outd;
-    int outc;
-
-    // tail offset for cropping, ignored if reference blob is provided
-    // woffset is aka left, and woffset2 is aka right
-    int woffset2;
-    int hoffset2;
-    int doffset2;
-    int coffset2;
-
-    // numpy-style slice
-    // if provided, all the above attributes will be ignored
-    Mat starts;
-    Mat ends;
-    Mat axes;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CROP_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/cumulativesum.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/cumulativesum.h
deleted file mode 100644
index e50f370b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/cumulativesum.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Copyright (c) 2023 Xiaomi Corp.        (author: Fangjun Kuang)
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this
-// file except in compliance with the License. You may obtain a copy of the
-// License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
-// WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
-// License for the specific language governing permissions and limitations under
-// the License.
-
-#ifndef LAYER_CUMULATIVESUM_H
-#define LAYER_CUMULATIVESUM_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class CumulativeSum : public Layer
-{
-public:
-    CumulativeSum();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    int axis;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CUMULATIVESUM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolution.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolution.h
deleted file mode 100644
index ece5e1e5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolution.h
+++ /dev/null
@@ -1,72 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTION_H
-#define LAYER_DECONVOLUTION_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Deconvolution : public Layer
-{
-public:
-    Deconvolution();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    void cut_padding(const Mat& top_blob_bordered, Mat& top_blob, const Option& opt) const;
-
-public:
-    // param
-    int num_output;
-    int kernel_w;
-    int kernel_h;
-    int dilation_w;
-    int dilation_h;
-    int stride_w;
-    int stride_h;
-    int pad_left;
-    int pad_right;
-    int pad_top;
-    int pad_bottom;
-    int output_pad_right;
-    int output_pad_bottom;
-    int output_w;
-    int output_h;
-    int bias_term;
-
-    int weight_data_size;
-
-    // 0=none 1=relu 2=leakyrelu 3=clip 4=sigmoid
-    int activation_type;
-    Mat activation_params;
-
-    int dynamic_weight;
-
-    // model
-    Mat weight_data;
-    Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTION_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolution1d.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolution1d.h
deleted file mode 100644
index c3b64dd5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolution1d.h
+++ /dev/null
@@ -1,64 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTION1D_H
-#define LAYER_DECONVOLUTION1D_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Deconvolution1D : public Layer
-{
-public:
-    Deconvolution1D();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    void cut_padding(const Mat& top_blob_bordered, Mat& top_blob, const Option& opt) const;
-
-public:
-    // param
-    int num_output;
-    int kernel_w;
-    int dilation_w;
-    int stride_w;
-    int pad_left;
-    int pad_right;
-    int output_pad_right;
-    int output_w;
-    int bias_term;
-
-    int weight_data_size;
-
-    int activation_type;
-    Mat activation_params;
-
-    int dynamic_weight;
-
-    // model
-    Mat weight_data;
-    Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTION1D_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolution3d.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolution3d.h
deleted file mode 100644
index 3545c040..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolution3d.h
+++ /dev/null
@@ -1,74 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTION3D_H
-#define LAYER_DECONVOLUTION3D_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Deconvolution3D : public Layer
-{
-public:
-    Deconvolution3D();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    void cut_padding(const Mat& top_blob_bordered, Mat& top_blob, const Option& opt) const;
-
-public:
-    // param
-    int num_output;
-    int kernel_w;
-    int kernel_h;
-    int kernel_d;
-    int dilation_w;
-    int dilation_h;
-    int dilation_d;
-    int stride_w;
-    int stride_h;
-    int stride_d;
-    int pad_left;
-    int pad_right;
-    int pad_top;
-    int pad_bottom;
-    int pad_front;
-    int pad_behind;
-    int output_pad_right;
-    int output_pad_bottom;
-    int output_pad_behind;
-    int output_w;
-    int output_h;
-    int output_d;
-    int bias_term;
-
-    int weight_data_size;
-
-    int activation_type;
-    Mat activation_params;
-
-    // model
-    Mat weight_data;
-    Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTION3D_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolutiondepthwise.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolutiondepthwise.h
deleted file mode 100644
index 8e6507d6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolutiondepthwise.h
+++ /dev/null
@@ -1,73 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTIONDEPTHWISE_H
-#define LAYER_DECONVOLUTIONDEPTHWISE_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class DeconvolutionDepthWise : public Layer
-{
-public:
-    DeconvolutionDepthWise();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    void cut_padding(const Mat& top_blob_bordered, Mat& top_blob, const Option& opt) const;
-
-public:
-    // param
-    int num_output;
-    int kernel_w;
-    int kernel_h;
-    int dilation_w;
-    int dilation_h;
-    int stride_w;
-    int stride_h;
-    int pad_left;
-    int pad_right;
-    int pad_top;
-    int pad_bottom;
-    int output_pad_right;
-    int output_pad_bottom;
-    int output_w;
-    int output_h;
-    int bias_term;
-
-    int weight_data_size;
-    int group;
-
-    // 0=none 1=relu 2=leakyrelu 3=clip 4=sigmoid
-    int activation_type;
-    Mat activation_params;
-
-    int dynamic_weight;
-
-    // model
-    Mat weight_data;
-    Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTIONDEPTHWISE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolutiondepthwise1d.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolutiondepthwise1d.h
deleted file mode 100644
index 1ed1513d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolutiondepthwise1d.h
+++ /dev/null
@@ -1,65 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTIONDEPTHWISE1D_H
-#define LAYER_DECONVOLUTIONDEPTHWISE1D_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class DeconvolutionDepthWise1D : public Layer
-{
-public:
-    DeconvolutionDepthWise1D();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    void cut_padding(const Mat& top_blob_bordered, Mat& top_blob, const Option& opt) const;
-
-public:
-    // param
-    int num_output;
-    int kernel_w;
-    int dilation_w;
-    int stride_w;
-    int pad_left;
-    int pad_right;
-    int output_pad_right;
-    int output_w;
-    int bias_term;
-
-    int weight_data_size;
-    int group;
-
-    int activation_type;
-    Mat activation_params;
-
-    int dynamic_weight;
-
-    // model
-    Mat weight_data;
-    Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTIONDEPTHWISE1D_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolutiondepthwise3d.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolutiondepthwise3d.h
deleted file mode 100644
index 7820b1bb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deconvolutiondepthwise3d.h
+++ /dev/null
@@ -1,75 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTIONDEPTHWISE3D_H
-#define LAYER_DECONVOLUTIONDEPTHWISE3D_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class DeconvolutionDepthWise3D : public Layer
-{
-public:
-    DeconvolutionDepthWise3D();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    void cut_padding(const Mat& top_blob_bordered, Mat& top_blob, const Option& opt) const;
-
-public:
-    // param
-    int num_output;
-    int kernel_w;
-    int kernel_h;
-    int kernel_d;
-    int dilation_w;
-    int dilation_h;
-    int dilation_d;
-    int stride_w;
-    int stride_h;
-    int stride_d;
-    int pad_left;
-    int pad_right;
-    int pad_top;
-    int pad_bottom;
-    int pad_front;
-    int pad_behind;
-    int output_pad_right;
-    int output_pad_bottom;
-    int output_pad_behind;
-    int output_w;
-    int output_h;
-    int output_d;
-    int bias_term;
-
-    int weight_data_size;
-    int group;
-
-    int activation_type;
-    Mat activation_params;
-
-    // model
-    Mat weight_data;
-    Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTIONDEPTHWISE3D_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deepcopy.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deepcopy.h
deleted file mode 100644
index 4e9015f2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deepcopy.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DEEPCOPY_H
-#define LAYER_DEEPCOPY_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class DeepCopy : public Layer
-{
-public:
-    DeepCopy();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DEEPCOPY_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deformableconv2d.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deformableconv2d.h
deleted file mode 100644
index d7315fc3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/deformableconv2d.h
+++ /dev/null
@@ -1,61 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DEFORMABLECONV2D_H
-#define LAYER_DEFORMABLECONV2D_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class DeformableConv2D : public Layer
-{
-public:
-    DeformableConv2D();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    // param
-    int num_output;
-    int kernel_w;
-    int kernel_h;
-    int dilation_w;
-    int dilation_h;
-    int stride_w;
-    int stride_h;
-    int pad_left; // -233=SAME_UPPER -234=SAME_LOWER
-    int pad_right;
-    int pad_top;
-    int pad_bottom;
-    int bias_term;
-
-    int weight_data_size;
-
-    // 0=none 1=relu 2=leakyrelu 3=clip 4=sigmoid
-    int activation_type;
-    Mat activation_params;
-
-    // model
-    Mat weight_data;
-    Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DEFORMABLECONV2D_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/dequantize.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/dequantize.h
deleted file mode 100644
index 39cd209e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/dequantize.h
+++ /dev/null
@@ -1,43 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2018 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DEQUANTIZE_H
-#define LAYER_DEQUANTIZE_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Dequantize : public Layer
-{
-public:
-    Dequantize();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int scale_data_size;
-    int bias_data_size;
-
-    Mat scale_data;
-    Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DEQUANTIZE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/detectionoutput.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/detectionoutput.h
deleted file mode 100644
index 02347da2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/detectionoutput.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DETECTIONOUTPUT_H
-#define LAYER_DETECTIONOUTPUT_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class DetectionOutput : public Layer
-{
-public:
-    DetectionOutput();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    int num_class;
-    float nms_threshold;
-    int nms_top_k;
-    int keep_top_k;
-    float confidence_threshold;
-    float variances[4];
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DETECTIONOUTPUT_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/diag.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/diag.h
deleted file mode 100644
index 8dd0babe..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/diag.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DIAG_H
-#define LAYER_DIAG_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Diag : public Layer
-{
-public:
-    Diag();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int diagonal;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DIAG_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/dropout.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/dropout.h
deleted file mode 100644
index 396740ab..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/dropout.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DROPOUT_H
-#define LAYER_DROPOUT_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Dropout : public Layer
-{
-public:
-    Dropout();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    float scale;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DROPOUT_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/einsum.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/einsum.h
deleted file mode 100644
index 07f484c4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/einsum.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_EINSUM_H
-#define LAYER_EINSUM_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Einsum : public Layer
-{
-public:
-    Einsum();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    // equation tokens
-    std::vector<std::string> lhs_tokens;
-    std::string rhs_token;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_EINSUM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/eltwise.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/eltwise.h
deleted file mode 100644
index 4256bc30..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/eltwise.h
+++ /dev/null
@@ -1,46 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ELTWISE_H
-#define LAYER_ELTWISE_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Eltwise : public Layer
-{
-public:
-    Eltwise();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-    enum OperationType
-    {
-        Operation_PROD = 0,
-        Operation_SUM = 1,
-        Operation_MAX = 2
-    };
-
-public:
-    // param
-    int op_type;
-    Mat coeffs;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ELTWISE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/elu.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/elu.h
deleted file mode 100644
index 8d44124c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/elu.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ELU_H
-#define LAYER_ELU_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class ELU : public Layer
-{
-public:
-    ELU();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    float alpha;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ELU_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/embed.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/embed.h
deleted file mode 100644
index 8e236656..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/embed.h
+++ /dev/null
@@ -1,48 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_EMBED_H
-#define LAYER_EMBED_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Embed : public Layer
-{
-public:
-    Embed();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    // param
-    int num_output;
-    int input_dim;
-    int bias_term;
-
-    int weight_data_size;
-
-    // model
-    Mat weight_data;
-    Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_EMBED_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/erf.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/erf.h
deleted file mode 100644
index 31a7058b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/erf.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ERF_H
-#define LAYER_ERF_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Erf : public Layer
-{
-public:
-    Erf();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ERF_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/exp.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/exp.h
deleted file mode 100644
index df1c53f8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/exp.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_EXP_H
-#define LAYER_EXP_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Exp : public Layer
-{
-public:
-    Exp();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    float base;
-    float scale;
-    float shift;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_EXP_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/expanddims.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/expanddims.h
deleted file mode 100644
index 4c8c990f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/expanddims.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_EXPANDDIMS_H
-#define LAYER_EXPANDDIMS_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class ExpandDims : public Layer
-{
-public:
-    ExpandDims();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int expand_w;
-    int expand_h;
-    int expand_d;
-    int expand_c;
-    Mat axes;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_EXPANDDIMS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/flatten.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/flatten.h
deleted file mode 100644
index 22667a45..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/flatten.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_FLATTEN_H
-#define LAYER_FLATTEN_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Flatten : public Layer
-{
-public:
-    Flatten();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_FLATTEN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/fold.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/fold.h
deleted file mode 100644
index e6eccad5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/fold.h
+++ /dev/null
@@ -1,48 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_FOLD_H
-#define LAYER_FOLD_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Fold : public Layer
-{
-public:
-    Fold();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int kernel_w;
-    int kernel_h;
-    int dilation_w;
-    int dilation_h;
-    int stride_w;
-    int stride_h;
-    int pad_left; // -233=SAME_UPPER -234=SAME_LOWER
-    int pad_right;
-    int pad_top;
-    int pad_bottom;
-    int output_w;
-    int output_h;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_FOLD_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/fused_activation.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/fused_activation.h
deleted file mode 100644
index d5919257..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/fused_activation.h
+++ /dev/null
@@ -1,139 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef FUSED_ACTIVATION_H
-#define FUSED_ACTIVATION_H
-
-#include "mat.h"
-#include "layer_type.h"
-
-static NCNN_FORCEINLINE float activation_ss(float v, int activation_type, const ncnn::Mat& activation_params)
-{
-    switch (activation_type)
-    {
-    case 1:
-    {
-        v = fmaxf(v, 0.f);
-        break;
-    }
-    case 2:
-    {
-        float slope = activation_params[0];
-        v = v > 0.f ? v : v * slope;
-        break;
-    }
-    case 3:
-    {
-        float min = activation_params[0];
-        float max = activation_params[1];
-        if (v < min)
-            v = min;
-        if (v > max)
-            v = max;
-        break;
-    }
-    case 4:
-    {
-        v = std::min(v, 88.3762626647949f);
-        v = std::max(v, -88.3762626647949f);
-        v = 1.f / (1.f + expf(-v));
-        break;
-    }
-    case 5:
-    {
-        v = v * tanhf(logf(expf(v) + 1.f));
-        break;
-    }
-    case 6:
-    {
-        float alpha = activation_params[0];
-        float beta = activation_params[1];
-        float lower = -beta / alpha;
-        float upper = (1.f / alpha) + lower;
-        if (v < lower)
-            v = 0.f;
-        else if (v > upper)
-            ;
-        else
-            v = v * (v * alpha + beta);
-        break;
-    }
-    }
-
-    return v;
-}
-
-static ncnn::Layer* create_activation_layer(int activation_type, const ncnn::Mat& activation_params, const ncnn::Option& opt)
-{
-    ncnn::Layer* activation = 0;
-
-    if (activation_type == 1)
-    {
-        activation = ncnn::create_layer_cpu(ncnn::LayerType::ReLU);
-
-        ncnn::ParamDict pd;
-        activation->load_param(pd);
-    }
-    else if (activation_type == 2)
-    {
-        activation = ncnn::create_layer_cpu(ncnn::LayerType::ReLU);
-
-        ncnn::ParamDict pd;
-        pd.set(0, activation_params[0]); // slope
-        activation->load_param(pd);
-    }
-    else if (activation_type == 3)
-    {
-        activation = ncnn::create_layer_cpu(ncnn::LayerType::Clip);
-
-        ncnn::ParamDict pd;
-        pd.set(0, activation_params[0]); // min
-        pd.set(1, activation_params[1]); // max
-
-        activation->load_param(pd);
-    }
-    else if (activation_type == 4)
-    {
-        activation = ncnn::create_layer_cpu(ncnn::LayerType::Sigmoid);
-
-        ncnn::ParamDict pd;
-        activation->load_param(pd);
-    }
-    else if (activation_type == 5)
-    {
-        activation = ncnn::create_layer_cpu(ncnn::LayerType::Mish);
-
-        ncnn::ParamDict pd;
-        activation->load_param(pd);
-    }
-    else if (activation_type == 6)
-    {
-        activation = ncnn::create_layer_cpu(ncnn::LayerType::HardSwish);
-
-        ncnn::ParamDict pd;
-        pd.set(0, activation_params[0]); // alpha
-        pd.set(1, activation_params[1]); // beta
-
-        activation->load_param(pd);
-    }
-
-    if (activation)
-    {
-        activation->create_pipeline(opt);
-    }
-
-    return activation;
-}
-
-#endif // FUSED_ACTIVATION_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gelu.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gelu.h
deleted file mode 100644
index a4fb37a6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gelu.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GELU_H
-#define LAYER_GELU_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class GELU : public Layer
-{
-public:
-    GELU();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    int fast_gelu;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GELU_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gemm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gemm.h
deleted file mode 100644
index e006114a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gemm.h
+++ /dev/null
@@ -1,65 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GEMM_H
-#define LAYER_GEMM_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Gemm : public Layer
-{
-public:
-    Gemm();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    float alpha;
-    float beta;
-    int transA;
-    int transB;
-
-    int constantA;
-    int constantB;
-    int constantC;
-    int constantM;
-    int constantN;
-    int constantK;
-    int constant_broadcast_type_C;
-    int output_N1M;
-    int output_elempack;
-    int output_elemtype; // 0=auto 1=fp32
-    int output_transpose;
-
-    int constant_TILE_M;
-    int constant_TILE_N;
-    int constant_TILE_K;
-
-    // constant A / B / C
-    Mat A_data;
-    Mat B_data;
-    Mat C_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GEMM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/glu.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/glu.h
deleted file mode 100644
index 00368295..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/glu.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Copyright (c) 2022 Xiaomi Corp.        (author: Fangjun Kuang)
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this
-// file except in compliance with the License. You may obtain a copy of the
-// License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
-// WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
-// License for the specific language governing permissions and limitations under
-// the License.
-
-#ifndef LAYER_GLU_H
-#define LAYER_GLU_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class GLU : public Layer
-{
-public:
-    GLU();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob,
-                        const Option& opt) const;
-
-public:
-    int axis;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GLU_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gridsample.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gridsample.h
deleted file mode 100644
index f6e17c9d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gridsample.h
+++ /dev/null
@@ -1,56 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GRIDSAMPLE_H
-#define LAYER_GRIDSAMPLE_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class GridSample : public Layer
-{
-public:
-    GridSample();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-    enum InterpolationMode // 1=bilinear  2=nearest  3=bicubic
-    {
-        Interpolation_BILINEAR = 1,
-        Interpolation_NEAREST = 2,
-        Interpolation_BICUBIC = 3
-    };
-
-    enum PaddingMode // 1=zeros     2=border   3=reflection
-    {
-        Padding_ZEROS = 1,
-        Padding_BORDER = 2,
-        Padding_REFLECTION = 3
-    };
-
-public:
-    // param
-    int sample_type;  // 1=bilinear  2=nearest  3=bicubic
-    int padding_mode; // 1=zeros     2=border   3=reflection
-    int align_corner;
-
-    int permute_fusion;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GRIDSAMPLE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/groupnorm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/groupnorm.h
deleted file mode 100644
index b6bd5a6d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/groupnorm.h
+++ /dev/null
@@ -1,47 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GROUPNORM_H
-#define LAYER_GROUPNORM_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class GroupNorm : public Layer
-{
-public:
-    GroupNorm();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    // param
-    int group;
-    int channels;
-    float eps;
-    int affine;
-
-    // model
-    Mat gamma_data;
-    Mat beta_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GROUPNORM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gru.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gru.h
deleted file mode 100644
index 976550f7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/gru.h
+++ /dev/null
@@ -1,47 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GRU_H
-#define LAYER_GRU_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class GRU : public Layer
-{
-public:
-    GRU();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    int num_output;
-    int weight_data_size;
-    int direction; // 0=forward 1=reverse 2=bidirectional
-
-    Mat weight_hc_data;
-    Mat weight_xc_data;
-    Mat bias_c_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GRU_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/hardsigmoid.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/hardsigmoid.h
deleted file mode 100644
index 57b7074d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/hardsigmoid.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_HARDSIGMOID_H
-#define LAYER_HARDSIGMOID_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class HardSigmoid : public Layer
-{
-public:
-    HardSigmoid();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    float alpha, beta, lower, upper;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_HARDSIGMOID_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/hardswish.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/hardswish.h
deleted file mode 100644
index 021db2f8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/hardswish.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_HARDSWISH_H
-#define LAYER_HARDSWISH_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class HardSwish : public Layer
-{
-public:
-    HardSwish();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    float alpha, beta, lower, upper;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_HARDSWISH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/innerproduct.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/innerproduct.h
deleted file mode 100644
index becf7b1d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/innerproduct.h
+++ /dev/null
@@ -1,63 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INNERPRODUCT_H
-#define LAYER_INNERPRODUCT_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class InnerProduct : public Layer
-{
-public:
-    InnerProduct();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_INT8
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    // param
-    int num_output;
-    int bias_term;
-
-    int weight_data_size;
-
-    int int8_scale_term;
-
-    // 0=none 1=relu 2=leakyrelu 3=clip 4=sigmoid
-    int activation_type;
-    Mat activation_params;
-
-    // model
-    Mat weight_data;
-    Mat bias_data;
-
-#if NCNN_INT8
-    Mat weight_data_int8_scales;
-    Mat bottom_blob_int8_scales;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INNERPRODUCT_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/input.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/input.h
deleted file mode 100644
index 2a90b018..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/input.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INPUT_H
-#define LAYER_INPUT_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Input : public Layer
-{
-public:
-    Input();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-#if NCNN_VULKAN
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-#endif // NCNN_VULKAN
-
-public:
-    int w;
-    int h;
-    int d;
-    int c;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INPUT_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/instancenorm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/instancenorm.h
deleted file mode 100644
index 9a37932a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/instancenorm.h
+++ /dev/null
@@ -1,46 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INSTANCENORM_H
-#define LAYER_INSTANCENORM_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class InstanceNorm : public Layer
-{
-public:
-    InstanceNorm();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    // param
-    int channels;
-    float eps;
-    int affine;
-
-    // model
-    Mat gamma_data;
-    Mat beta_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INSTANCENORM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/interp.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/interp.h
deleted file mode 100644
index 21ef776a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/interp.h
+++ /dev/null
@@ -1,46 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INTERP_H
-#define LAYER_INTERP_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Interp : public Layer
-{
-public:
-    Interp();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    // param
-    int resize_type; //1=nearest  2=bilinear  3=bicubic
-    float width_scale;
-    float height_scale;
-    int output_width;
-    int output_height;
-    int dynamic_target_size;
-    int align_corner;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INTERP_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/layernorm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/layernorm.h
deleted file mode 100644
index 375c0eb4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/layernorm.h
+++ /dev/null
@@ -1,46 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_LAYERNORM_H
-#define LAYER_LAYERNORM_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class LayerNorm : public Layer
-{
-public:
-    LayerNorm();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    // param
-    int affine_size;
-    float eps;
-    int affine;
-
-    // model
-    Mat gamma_data;
-    Mat beta_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_LAYERNORM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/log.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/log.h
deleted file mode 100644
index 185a1407..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/log.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_LOG_H
-#define LAYER_LOG_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Log : public Layer
-{
-public:
-    Log();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    float base;
-    float scale;
-    float shift;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_LOG_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/absval_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/absval_loongarch.h
deleted file mode 100644
index 855f959c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/absval_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ABSVAL_LOONGARCH_H
-#define LAYER_ABSVAL_LOONGARCH_H
-
-#include "absval.h"
-
-namespace ncnn {
-
-class AbsVal_loongarch : public AbsVal
-{
-public:
-    AbsVal_loongarch();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ABSVAL_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/batchnorm_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/batchnorm_loongarch.h
deleted file mode 100644
index fb477a9a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/batchnorm_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BATCHNORM_LOONGARCH_H
-#define LAYER_BATCHNORM_LOONGARCH_H
-
-#include "batchnorm.h"
-
-namespace ncnn {
-
-class BatchNorm_loongarch : public BatchNorm
-{
-public:
-    BatchNorm_loongarch();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BATCHNORM_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/bias_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/bias_loongarch.h
deleted file mode 100644
index 35824997..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/bias_loongarch.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BIAS_LOONGARCH_H
-#define LAYER_BIAS_LOONGARCH_H
-
-#include "bias.h"
-
-namespace ncnn {
-
-class Bias_loongarch : public Bias
-{
-public:
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BIAS_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/binaryop_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/binaryop_loongarch.h
deleted file mode 100644
index 2fc401ad..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/binaryop_loongarch.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BINARYOP_LOONGARCH_H
-#define LAYER_BINARYOP_LOONGARCH_H
-
-#include "binaryop.h"
-
-namespace ncnn {
-
-class BinaryOp_loongarch : public BinaryOp
-{
-public:
-    BinaryOp_loongarch();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BINARYOP_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/cast_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/cast_loongarch.h
deleted file mode 100644
index 8925f242..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/cast_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CAST_LOONGARCH_H
-#define LAYER_CAST_LOONGARCH_H
-
-#include "cast.h"
-
-namespace ncnn {
-
-class Cast_loongarch : public Cast
-{
-public:
-    Cast_loongarch();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CAST_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/clip_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/clip_loongarch.h
deleted file mode 100644
index 1ebeee2a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/clip_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CLIP_LOONGARCH_H
-#define LAYER_CLIP_LOONGARCH_H
-
-#include "clip.h"
-
-namespace ncnn {
-
-class Clip_loongarch : public Clip
-{
-public:
-    Clip_loongarch();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CLIP_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/concat_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/concat_loongarch.h
deleted file mode 100644
index 91b32ef2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/concat_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONCAT_LOONGARCH_H
-#define LAYER_CONCAT_LOONGARCH_H
-
-#include "concat.h"
-
-namespace ncnn {
-
-class Concat_loongarch : public Concat
-{
-public:
-    Concat_loongarch();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONCAT_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution1d_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution1d_loongarch.h
deleted file mode 100644
index 922fae59..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution1d_loongarch.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTION1D_LOONGARCH_H
-#define LAYER_CONVOLUTION1D_LOONGARCH_H
-
-#include "convolution1d.h"
-
-namespace ncnn {
-
-class Convolution1D_loongarch : public Convolution1D
-{
-public:
-    Convolution1D_loongarch();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    // packn
-    Mat weight_data_packed;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTION1D_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1.h
deleted file mode 100644
index 83d37784..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1.h
+++ /dev/null
@@ -1,26 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_lsx(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_int8.h
deleted file mode 100644
index 08f439c4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_int8.h
+++ /dev/null
@@ -1,83 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_int8_lsx(bottom_im2col, top_blob, kernel, opt);
-}
-
-static void conv1x1s2_sgemm_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = w - 2 * outw + w;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const signed char* r0 = bottom_blob.channel(p);
-        signed char* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                outptr[0] = r0[0];
-                outptr[1] = r0[2];
-                outptr[2] = r0[4];
-                outptr[3] = r0[6];
-
-                r0 += 8;
-                outptr += 4;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                outptr[0] = r0[0];
-                outptr[1] = r0[2];
-
-                r0 += 4;
-                outptr += 2;
-            }
-            for (; j < outw; j++)
-            {
-                outptr[0] = r0[0];
-
-                r0 += 2;
-                outptr += 1;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_int8_lsx(bottom_blob_shrinked, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack1to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack1to4_int8.h
deleted file mode 100644
index 00e1e258..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack1to4_int8.h
+++ /dev/null
@@ -1,83 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_pack1to4_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_pack1to4_int8_lsx(bottom_im2col, top_blob, kernel, opt);
-}
-
-static void conv1x1s2_sgemm_pack1to4_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = w - 2 * outw + w;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const signed char* r0 = bottom_blob.channel(p);
-        signed char* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                outptr[0] = r0[0];
-                outptr[1] = r0[2];
-                outptr[2] = r0[4];
-                outptr[3] = r0[6];
-
-                r0 += 8;
-                outptr += 4;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                outptr[0] = r0[0];
-                outptr[1] = r0[2];
-
-                r0 += 4;
-                outptr += 2;
-            }
-            for (; j < outw; j++)
-            {
-                outptr[0] = r0[0];
-
-                r0 += 2;
-                outptr += 1;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_pack1to4_int8_lsx(bottom_blob_shrinked, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack4.h
deleted file mode 100644
index cf5a5b8e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack4.h
+++ /dev/null
@@ -1,65 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_pack4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_pack4_lsx(bottom_im2col, top_blob, kernel, _bias, opt);
-}
-
-static void conv1x1s2_sgemm_pack4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const float* r0 = bottom_blob.channel(p);
-        float* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m128 _val = (__m128)__lsx_vld(r0, 0);
-                __lsx_vst(_val, outptr, 0);
-
-                r0 += 4 * 2;
-                outptr += 4;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_pack4_lsx(bottom_blob_shrinked, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack4to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack4to1.h
deleted file mode 100644
index b8712909..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack4to1.h
+++ /dev/null
@@ -1,65 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_pack4to1_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_pack4to1_lsx(bottom_im2col, top_blob, kernel, _bias, opt);
-}
-
-static void conv1x1s2_sgemm_pack4to1_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const float* r0 = bottom_blob.channel(p);
-        float* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m128 _val = (__m128)__lsx_vld(r0, 0);
-                __lsx_vst(_val, outptr, 0);
-
-                r0 += 4 * 2;
-                outptr += 4;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_pack4to1_lsx(bottom_blob_shrinked, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack8to1_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack8to1_int8.h
deleted file mode 100644
index 8df0e128..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack8to1_int8.h
+++ /dev/null
@@ -1,65 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_pack8to1_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_pack8to1_int8_lsx(bottom_im2col, top_blob, kernel, opt);
-}
-
-static void conv1x1s2_sgemm_pack8to1_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = w - 2 * outw + w;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const int64_t* r0 = bottom_blob.channel(p);
-        int64_t* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                outptr[0] = r0[0];
-
-                r0 += 2;
-                outptr += 1;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_pack8to1_int8_lsx(bottom_blob_shrinked, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack8to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack8to4_int8.h
deleted file mode 100644
index 6aaa720d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_1x1_pack8to4_int8.h
+++ /dev/null
@@ -1,65 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_pack8to4_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_pack8to4_int8_lsx(bottom_im2col, top_blob, kernel, opt);
-}
-
-static void conv1x1s2_sgemm_pack8to4_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = w - 2 * outw + w;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const int64_t* r0 = bottom_blob.channel(p);
-        int64_t* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                outptr[0] = r0[0];
-
-                r0 += 2;
-                outptr += 1;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_pack8to4_int8_lsx(bottom_blob_shrinked, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3.h
deleted file mode 100644
index 66e10106..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3.h
+++ /dev/null
@@ -1,412 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd23_transform_kernel_lsx(const Mat& kernel, Mat& kernel_tm2, int inch, int outch, const Option& opt)
-{
-    Mat kernel_tm(4 * 4, inch, outch);
-
-    // G
-    const float ktm[4][3] = {
-        {1.0f, 0.0f, 0.0f},
-        {1.0f / 2, 1.0f / 2, 1.0f / 2},
-        {1.0f / 2, -1.0f / 2, 1.0f / 2},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[4][3];
-            for (int i = 0; i < 4; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 4; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 4; i++)
-                {
-                    kernel_tm0[j * 4 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 16-inch-outch
-    // dst = inch-16-outch
-#if __loongarch_sx
-    kernel_tm2.create(8 * inch, 16, outch / 8 + (outch % 8) / 4 + outch % 4);
-#else
-    kernel_tm2.create(2 * inch, 16, outch / 2 + outch % 2);
-#endif
-
-    int q = 0;
-#if __loongarch_sx
-    for (; q + 7 < outch; q += 8)
-    {
-        Mat g0 = kernel_tm2.channel(q / 8);
-
-        for (int k = 0; k < 16; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 8; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-    for (; q + 3 < outch; q += 4)
-    {
-        Mat g0 = kernel_tm2.channel(q / 8 + (q % 8) / 4);
-
-        for (int k = 0; k < 16; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#else  // __loongarch_sx
-    for (; q + 1 < outch; q += 2)
-    {
-        Mat g0 = kernel_tm2.channel(q / 2);
-
-        for (int k = 0; k < 16; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 2; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#endif // __loongarch_sx
-    for (; q < outch; q++)
-    {
-#if __loongarch_sx
-        Mat g0 = kernel_tm2.channel(q / 8 + (q % 8) / 4 + q % 4);
-#else
-        Mat g0 = kernel_tm2.channel(q / 2 + q % 2);
-#endif
-
-        for (int k = 0; k < 16; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                const float* k00 = kernel_tm.channel(q).row(p);
-                g00[0] = k00[k];
-                g00++;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 2n+2, winograd F(2,3)
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 1) / 2 * 2;
-    outh = (outh + 1) / 2 * 2;
-
-    w = outw + 2;
-    h = outh + 2;
-    Option opt_b = opt;
-    opt_b.blob_allocator = opt.workspace_allocator;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, 0, 0.f, opt_b);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 2;
-        int h_tiles = outh / 2;
-        int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 16, inch, 4u, opt.workspace_allocator);
-        conv3x3s1_winograd23_transform_input_lsx(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_lsx(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, 4u, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd23_transform_output_lsx(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
-
-static void conv3x3s1_winograd43_transform_kernel_lsx(const Mat& kernel, Mat& kernel_tm2, int inch, int outch, const Option& opt)
-{
-    Mat kernel_tm(6 * 6, inch, outch);
-
-    // G
-    const float ktm[6][3] = {
-        {1.0f / 4, 0.0f, 0.0f},
-        {-1.0f / 6, -1.0f / 6, -1.0f / 6},
-        {-1.0f / 6, 1.0f / 6, -1.0f / 6},
-        {1.0f / 24, 1.0f / 12, 1.0f / 6},
-        {1.0f / 24, -1.0f / 12, 1.0f / 6},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[6][3];
-            for (int i = 0; i < 6; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 6; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 6; i++)
-                {
-                    kernel_tm0[j * 6 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 36-inch-outch
-    // dst = inch-36-outch
-#if __loongarch_sx
-    kernel_tm2.create(8 * inch, 36, outch / 8 + (outch % 8) / 4 + outch % 4);
-#else
-    kernel_tm2.create(2 * inch, 36, outch / 2 + outch % 2);
-#endif
-
-    int q = 0;
-#if __loongarch_sx
-    for (; q + 7 < outch; q += 8)
-    {
-        Mat g0 = kernel_tm2.channel(q / 8);
-
-        for (int k = 0; k < 36; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 8; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-    for (; q + 3 < outch; q += 4)
-    {
-        Mat g0 = kernel_tm2.channel(q / 8 + (q % 8) / 4);
-
-        for (int k = 0; k < 36; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#else  // __loongarch_sx
-    for (; q + 1 < outch; q += 2)
-    {
-        Mat g0 = kernel_tm2.channel(q / 2);
-
-        for (int k = 0; k < 36; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 2; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#endif // __loongarch_sx
-    for (; q < outch; q++)
-    {
-#if __loongarch_sx
-        Mat g0 = kernel_tm2.channel(q / 8 + (q % 8) / 4 + q % 4);
-#else
-        Mat g0 = kernel_tm2.channel(q / 2 + q % 2);
-#endif
-
-        for (int k = 0; k < 36; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                const float* k00 = kernel_tm.channel(q).row(p);
-                g00[0] = k00[k];
-                g00++;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 4n+2, winograd F(4,3)
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 3) / 4 * 4;
-    outh = (outh + 3) / 4 * 4;
-
-    w = outw + 2;
-    h = outh + 2;
-
-    Option opt_b = opt;
-    opt_b.blob_allocator = opt.workspace_allocator;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, 0, 0.f, opt_b);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 4;
-        int h_tiles = outh / 4;
-        int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 36, inch, 4u, opt.workspace_allocator);
-        conv3x3s1_winograd43_transform_input_lsx(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_lsx(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, 4u, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd43_transform_output_lsx(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_int8.h
deleted file mode 100644
index 3ea28dd0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_int8.h
+++ /dev/null
@@ -1,252 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd43_transform_kernel_int8_lsx(const Mat& kernel, Mat& kernel_tm_packed, int inch, int outch, const Option& opt)
-{
-    // winograd43 transform kernel
-    Mat kernel_tm(6 * 6, inch, outch, (size_t)2u);
-
-    const short ktm[6][3] = {
-        {6, 0, 0},
-        {-4, -4, -4},
-        {-4, 4, -4},
-        {1, 2, 4},
-        {1, -2, 4},
-        {0, 0, 6}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const signed char* kernel0 = (const signed char*)kernel + p * inch * 9 + q * 9;
-            short* kernel_tm0 = kernel_tm.channel(p).row<short>(q);
-
-            // transform kernel
-            const signed char* k0 = kernel0;
-            const signed char* k1 = kernel0 + 3;
-            const signed char* k2 = kernel0 + 6;
-
-            // h
-            short tmp[6][3];
-            for (int i = 0; i < 6; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 6; j++)
-            {
-                short* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 6; i++)
-                {
-                    kernel_tm0[j * 6 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 36-inch-outch
-    // dst = 2b-inch-36-outch/2b
-#if __loongarch_sx
-    if (outch >= 4)
-    {
-        if (inch >= 4)
-            kernel_tm_packed.create(inch / 4 + inch % 4, 36, outch / 4 + outch % 4, (size_t)2u * 16, 16);
-        else
-            kernel_tm_packed.create(inch, 36, outch / 4 + outch % 4, (size_t)2u * 4, 4);
-    }
-#else  // __loongarch_sx
-    if (outch >= 2)
-    {
-        kernel_tm_packed.create(inch, 36, outch / 2 + outch % 2, (size_t)2u * 2, 2);
-    }
-#endif // __loongarch_sx
-    else
-    {
-#if __loongarch_sx
-        if (inch >= 4)
-            kernel_tm_packed.create(inch / 4 + inch % 4, 36, outch, (size_t)2u * 4, 4);
-        else
-#endif // __loongarch_sx
-        {
-            kernel_tm_packed.create(inch, 36, outch, (size_t)2u, 1);
-        }
-    }
-
-    int p = 0;
-#if __loongarch_sx
-    for (; p + 3 < outch; p += 4)
-    {
-        const Mat k0 = kernel_tm.channel(p);
-        const Mat k1 = kernel_tm.channel(p + 1);
-        const Mat k2 = kernel_tm.channel(p + 2);
-        const Mat k3 = kernel_tm.channel(p + 3);
-
-        Mat g0 = kernel_tm_packed.channel(p / 4);
-
-        for (int k = 0; k < 36; k++)
-        {
-            short* g00 = g0.row<short>(k);
-
-            int q = 0;
-            for (; q + 3 < inch; q += 4)
-            {
-                g00[0] = k0.row<const short>(q)[k];
-                g00[1] = k0.row<const short>(q + 1)[k];
-                g00[2] = k0.row<const short>(q + 2)[k];
-                g00[3] = k0.row<const short>(q + 3)[k];
-                g00[4] = k1.row<const short>(q)[k];
-                g00[5] = k1.row<const short>(q + 1)[k];
-                g00[6] = k1.row<const short>(q + 2)[k];
-                g00[7] = k1.row<const short>(q + 3)[k];
-                g00[8] = k2.row<const short>(q)[k];
-                g00[9] = k2.row<const short>(q + 1)[k];
-                g00[10] = k2.row<const short>(q + 2)[k];
-                g00[11] = k2.row<const short>(q + 3)[k];
-                g00[12] = k3.row<const short>(q)[k];
-                g00[13] = k3.row<const short>(q + 1)[k];
-                g00[14] = k3.row<const short>(q + 2)[k];
-                g00[15] = k3.row<const short>(q + 3)[k];
-                g00 += 16;
-            }
-            for (; q < inch; q++)
-            {
-                g00[0] = k0.row<const short>(q)[k];
-                g00[1] = k1.row<const short>(q)[k];
-                g00[2] = k2.row<const short>(q)[k];
-                g00[3] = k3.row<const short>(q)[k];
-                g00 += 4;
-            }
-        }
-    }
-#else  // __loongarch_sx
-    for (; p + 1 < outch; p += 2)
-    {
-        const Mat k0 = kernel_tm.channel(p);
-        const Mat k1 = kernel_tm.channel(p + 1);
-
-        Mat g0 = kernel_tm_packed.channel(p / 2);
-
-        for (int k = 0; k < 36; k++)
-        {
-            short* g00 = g0.row<short>(k);
-
-            int q = 0;
-            for (; q < inch; q++)
-            {
-                g00[0] = k0.row<const short>(q)[k];
-                g00[1] = k1.row<const short>(q)[k];
-                g00 += 2;
-            }
-        }
-    }
-#endif // __loongarch_sx
-    for (; p < outch; p++)
-    {
-        const Mat k0 = kernel_tm.channel(p);
-
-#if __loongarch_sx
-        Mat g0 = kernel_tm_packed.channel(p / 4 + p % 4);
-#else
-        Mat g0 = kernel_tm_packed.channel(p / 2 + p % 2);
-#endif
-
-        for (int k = 0; k < 36; k++)
-        {
-            short* g00 = g0.row<short>(k);
-
-            int q = 0;
-#if __loongarch_sx
-            for (; q + 3 < inch; q += 4)
-            {
-                g00[0] = k0.row<const short>(q)[k];
-                g00[1] = k0.row<const short>(q + 1)[k];
-                g00[2] = k0.row<const short>(q + 2)[k];
-                g00[3] = k0.row<const short>(q + 3)[k];
-                g00 += 4;
-            }
-#endif // __loongarch_sx
-            for (; q < inch; q++)
-            {
-                g00[0] = k0.row<const short>(q)[k];
-                g00 += 1;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    //     size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 4n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 3) / 4 * 4;
-    outh = (outh + 3) / 4 * 4;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 4;
-        int h_tiles = outh / 4;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 36, inch, 2u * elempack, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd43_transform_input_int8_lsx(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_int8_lsx(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, 4u, 1, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd43_transform_output_int8_lsx(top_blob_tm, top_blob_bordered, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack1to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack1to4.h
deleted file mode 100644
index 2bcb0ce1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack1to4.h
+++ /dev/null
@@ -1,812 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack1to4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        __m128 _bias0 = bias ? (__m128)__lsx_vld(bias + p * 4, 0) : (__m128)__lsx_vreplgr2vr_w(0);
-        out0.fill(_bias0);
-
-        const float* k0 = kernel.channel(p);
-
-        int q = 0;
-        for (; q < inch; q++)
-        {
-            float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            __m128 _k00 = (__m128)__lsx_vld(k0, 0);
-            __m128 _k01 = (__m128)__lsx_vld(k0 + 4, 0);
-            __m128 _k02 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-            __m128 _k10 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-            __m128 _k11 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-            __m128 _k12 = (__m128)__lsx_vld(k0 + 4 * 5, 0);
-            __m128 _k20 = (__m128)__lsx_vld(k0 + 4 * 6, 0);
-            __m128 _k21 = (__m128)__lsx_vld(k0 + 4 * 7, 0);
-            __m128 _k22 = (__m128)__lsx_vld(k0 + 4 * 8, 0);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    __m128 _sum0 = (__m128)__lsx_vld(outptr0, 0);
-                    __m128 _sum1 = (__m128)__lsx_vld(outptr0 + 4, 0);
-                    __m128 _sum2 = (__m128)__lsx_vld(outptr0 + 4 * 2, 0);
-                    __m128 _sum3 = (__m128)__lsx_vld(outptr0 + 4 * 3, 0);
-                    __m128 _sum4 = (__m128)__lsx_vld(outptr0 + 4 * 4, 0);
-                    __m128 _sum5 = (__m128)__lsx_vld(outptr0 + 4 * 5, 0);
-                    __m128 _sum6 = (__m128)__lsx_vld(outptr0 + 4 * 6, 0);
-                    __m128 _sum7 = (__m128)__lsx_vld(outptr0 + 4 * 7, 0);
-
-                    __m128i _r0 = __lsx_vld(r0, 0);
-                    __m128i _r0n = __lsx_vld(r0 + 4, 0);
-                    __m128i _r0nn = __lsx_vld(r0 + 8, 0);
-
-                    __m128 _r00 = (__m128)__lsx_vreplvei_w(_r0, 0);
-                    __m128 _r01 = (__m128)__lsx_vreplvei_w(_r0, 1);
-                    __m128 _r02 = (__m128)__lsx_vreplvei_w(_r0, 2);
-                    __m128 _r03 = (__m128)__lsx_vreplvei_w(_r0, 3);
-                    __m128 _r04 = (__m128)__lsx_vreplvei_w(_r0n, 0);
-                    __m128 _r05 = (__m128)__lsx_vreplvei_w(_r0n, 1);
-                    __m128 _r06 = (__m128)__lsx_vreplvei_w(_r0n, 2);
-                    __m128 _r07 = (__m128)__lsx_vreplvei_w(_r0n, 3);
-                    __m128 _r08 = (__m128)__lsx_vreplvei_w(_r0nn, 0);
-                    __m128 _r09 = (__m128)__lsx_vreplvei_w(_r0nn, 1);
-
-                    _sum0 = __lsx_vfmadd_s(_k00, _r00, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k00, _r01, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k00, _r02, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k00, _r03, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k00, _r04, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k00, _r05, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k00, _r06, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k00, _r07, _sum7);
-                    _sum0 = __lsx_vfmadd_s(_k01, _r01, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k01, _r02, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k01, _r03, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k01, _r04, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k01, _r05, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k01, _r06, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k01, _r07, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k01, _r08, _sum7);
-                    _sum0 = __lsx_vfmadd_s(_k02, _r02, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k02, _r03, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k02, _r04, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k02, _r05, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k02, _r06, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k02, _r07, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k02, _r08, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k02, _r09, _sum7);
-
-                    __m128i _r1 = __lsx_vld(r1, 0);
-                    __m128i _r1n = __lsx_vld(r1 + 4, 0);
-                    __m128i _r1nn = __lsx_vld(r1 + 8, 0);
-
-                    __m128 _r10 = (__m128)__lsx_vreplvei_w(_r1, 0);
-                    __m128 _r11 = (__m128)__lsx_vreplvei_w(_r1, 1);
-                    __m128 _r12 = (__m128)__lsx_vreplvei_w(_r1, 2);
-                    __m128 _r13 = (__m128)__lsx_vreplvei_w(_r1, 3);
-                    __m128 _r14 = (__m128)__lsx_vreplvei_w(_r1n, 0);
-                    __m128 _r15 = (__m128)__lsx_vreplvei_w(_r1n, 1);
-                    __m128 _r16 = (__m128)__lsx_vreplvei_w(_r1n, 2);
-                    __m128 _r17 = (__m128)__lsx_vreplvei_w(_r1n, 3);
-                    __m128 _r18 = (__m128)__lsx_vreplvei_w(_r1nn, 0);
-                    __m128 _r19 = (__m128)__lsx_vreplvei_w(_r1nn, 1);
-
-                    _sum0 = __lsx_vfmadd_s(_k10, _r10, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k10, _r11, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k10, _r12, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k10, _r13, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k10, _r14, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k10, _r15, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k10, _r16, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k10, _r17, _sum7);
-                    _sum0 = __lsx_vfmadd_s(_k11, _r11, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k11, _r12, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k11, _r13, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k11, _r14, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k11, _r15, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k11, _r16, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k11, _r17, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k11, _r18, _sum7);
-                    _sum0 = __lsx_vfmadd_s(_k12, _r12, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k12, _r13, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k12, _r14, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k12, _r15, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k12, _r16, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k12, _r17, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k12, _r18, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k12, _r19, _sum7);
-
-                    __m128i _r2 = __lsx_vld(r2, 0);
-                    __m128i _r2n = __lsx_vld(r2 + 4, 0);
-                    __m128i _r2nn = __lsx_vld(r2 + 8, 0);
-
-                    __m128 _r20 = (__m128)__lsx_vreplvei_w(_r2, 0);
-                    __m128 _r21 = (__m128)__lsx_vreplvei_w(_r2, 1);
-                    __m128 _r22 = (__m128)__lsx_vreplvei_w(_r2, 2);
-                    __m128 _r23 = (__m128)__lsx_vreplvei_w(_r2, 3);
-                    __m128 _r24 = (__m128)__lsx_vreplvei_w(_r2n, 0);
-                    __m128 _r25 = (__m128)__lsx_vreplvei_w(_r2n, 1);
-                    __m128 _r26 = (__m128)__lsx_vreplvei_w(_r2n, 2);
-                    __m128 _r27 = (__m128)__lsx_vreplvei_w(_r2n, 3);
-                    __m128 _r28 = (__m128)__lsx_vreplvei_w(_r2nn, 0);
-                    __m128 _r29 = (__m128)__lsx_vreplvei_w(_r2nn, 1);
-
-                    _sum0 = __lsx_vfmadd_s(_k20, _r20, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k20, _r21, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k20, _r22, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k20, _r23, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k20, _r24, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k20, _r25, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k20, _r26, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k20, _r27, _sum7);
-                    _sum0 = __lsx_vfmadd_s(_k21, _r21, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k21, _r22, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k21, _r23, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k21, _r24, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k21, _r25, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k21, _r26, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k21, _r27, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k21, _r28, _sum7);
-                    _sum0 = __lsx_vfmadd_s(_k22, _r22, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k22, _r23, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k22, _r24, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k22, _r25, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k22, _r26, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k22, _r27, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k22, _r28, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k22, _r29, _sum7);
-
-                    __lsx_vst(_sum0, outptr0, 0);
-                    __lsx_vst(_sum1, outptr0 + 4, 0);
-                    __lsx_vst(_sum2, outptr0 + 4 * 2, 0);
-                    __lsx_vst(_sum3, outptr0 + 4 * 3, 0);
-                    __lsx_vst(_sum4, outptr0 + 4 * 4, 0);
-                    __lsx_vst(_sum5, outptr0 + 4 * 5, 0);
-                    __lsx_vst(_sum6, outptr0 + 4 * 6, 0);
-                    __lsx_vst(_sum7, outptr0 + 4 * 7, 0);
-
-                    outptr0 += 4 * 8;
-
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                }
-                for (; j + 3 < outw; j += 4)
-                {
-                    __m128 _sum0 = (__m128)__lsx_vld(outptr0, 0);
-                    __m128 _sum1 = (__m128)__lsx_vld(outptr0 + 4, 0);
-                    __m128 _sum2 = (__m128)__lsx_vld(outptr0 + 4 * 2, 0);
-                    __m128 _sum3 = (__m128)__lsx_vld(outptr0 + 4 * 3, 0);
-
-                    __m128i _r0 = __lsx_vld(r0, 0);
-                    __m128i _r0n = __lsx_vld(r0 + 4, 0);
-
-                    __m128 _r00 = (__m128)__lsx_vreplvei_w(_r0, 0);
-                    __m128 _r01 = (__m128)__lsx_vreplvei_w(_r0, 1);
-                    __m128 _r02 = (__m128)__lsx_vreplvei_w(_r0, 2);
-                    __m128 _r03 = (__m128)__lsx_vreplvei_w(_r0, 3);
-                    __m128 _r04 = (__m128)__lsx_vreplvei_w(_r0n, 0);
-                    __m128 _r05 = (__m128)__lsx_vreplvei_w(_r0n, 1);
-
-                    _sum0 = __lsx_vfmadd_s(_k00, _r00, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k00, _r01, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k00, _r02, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k00, _r03, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k01, _r01, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k01, _r02, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k01, _r03, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k01, _r04, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k02, _r02, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k02, _r03, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k02, _r04, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k02, _r05, _sum3);
-
-                    __m128i _r1 = __lsx_vld(r1, 0);
-                    __m128i _r1n = __lsx_vld(r1 + 4, 0);
-
-                    __m128 _r10 = (__m128)__lsx_vreplvei_w(_r1, 0);
-                    __m128 _r11 = (__m128)__lsx_vreplvei_w(_r1, 1);
-                    __m128 _r12 = (__m128)__lsx_vreplvei_w(_r1, 2);
-                    __m128 _r13 = (__m128)__lsx_vreplvei_w(_r1, 3);
-                    __m128 _r14 = (__m128)__lsx_vreplvei_w(_r1n, 0);
-                    __m128 _r15 = (__m128)__lsx_vreplvei_w(_r1n, 1);
-
-                    _sum0 = __lsx_vfmadd_s(_k10, _r10, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k10, _r11, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k10, _r12, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k10, _r13, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k11, _r11, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k11, _r12, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k11, _r13, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k11, _r14, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k12, _r12, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k12, _r13, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k12, _r14, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k12, _r15, _sum3);
-
-                    __m128i _r2 = __lsx_vld(r2, 0);
-                    __m128i _r2n = __lsx_vld(r2 + 4, 0);
-
-                    __m128 _r20 = (__m128)__lsx_vreplvei_w(_r2, 0);
-                    __m128 _r21 = (__m128)__lsx_vreplvei_w(_r2, 1);
-                    __m128 _r22 = (__m128)__lsx_vreplvei_w(_r2, 2);
-                    __m128 _r23 = (__m128)__lsx_vreplvei_w(_r2, 3);
-                    __m128 _r24 = (__m128)__lsx_vreplvei_w(_r2n, 0);
-                    __m128 _r25 = (__m128)__lsx_vreplvei_w(_r2n, 1);
-
-                    _sum0 = __lsx_vfmadd_s(_k20, _r20, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k20, _r21, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k20, _r22, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k20, _r23, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k21, _r21, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k21, _r22, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k21, _r23, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k21, _r24, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k22, _r22, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k22, _r23, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k22, _r24, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k22, _r25, _sum3);
-
-                    __lsx_vst(_sum0, outptr0, 0);
-                    __lsx_vst(_sum1, outptr0 + 4, 0);
-                    __lsx_vst(_sum2, outptr0 + 4 * 2, 0);
-                    __lsx_vst(_sum3, outptr0 + 4 * 3, 0);
-
-                    outptr0 += 4 * 4;
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    __m128 _sum0 = (__m128)__lsx_vld(outptr0, 0);
-                    __m128 _sum1 = (__m128)__lsx_vld(outptr0 + 4, 0);
-
-                    __m128i _r0 = __lsx_vld(r0, 0);
-                    __m128 _r00 = (__m128)__lsx_vreplvei_w(_r0, 0);
-                    __m128 _r01 = (__m128)__lsx_vreplvei_w(_r0, 1);
-                    __m128 _r02 = (__m128)__lsx_vreplvei_w(_r0, 2);
-                    __m128 _r03 = (__m128)__lsx_vreplvei_w(_r0, 3);
-
-                    _sum0 = __lsx_vfmadd_s(_k00, _r00, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k00, _r01, _sum1);
-                    _sum0 = __lsx_vfmadd_s(_k01, _r01, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k01, _r02, _sum1);
-                    _sum0 = __lsx_vfmadd_s(_k02, _r02, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k02, _r03, _sum1);
-
-                    __m128i _r1 = __lsx_vld(r1, 0);
-                    __m128 _r10 = (__m128)__lsx_vreplvei_w(_r1, 0);
-                    __m128 _r11 = (__m128)__lsx_vreplvei_w(_r1, 1);
-                    __m128 _r12 = (__m128)__lsx_vreplvei_w(_r1, 2);
-                    __m128 _r13 = (__m128)__lsx_vreplvei_w(_r1, 3);
-
-                    _sum0 = __lsx_vfmadd_s(_k10, _r10, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k10, _r11, _sum1);
-                    _sum0 = __lsx_vfmadd_s(_k11, _r11, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k11, _r12, _sum1);
-                    _sum0 = __lsx_vfmadd_s(_k12, _r12, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k12, _r13, _sum1);
-
-                    __m128i _r2 = __lsx_vld(r2, 0);
-                    __m128 _r20 = (__m128)__lsx_vreplvei_w(_r2, 0);
-                    __m128 _r21 = (__m128)__lsx_vreplvei_w(_r2, 1);
-                    __m128 _r22 = (__m128)__lsx_vreplvei_w(_r2, 2);
-                    __m128 _r23 = (__m128)__lsx_vreplvei_w(_r2, 3);
-
-                    _sum0 = __lsx_vfmadd_s(_k20, _r20, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k20, _r21, _sum1);
-                    _sum0 = __lsx_vfmadd_s(_k21, _r21, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k21, _r22, _sum1);
-                    _sum0 = __lsx_vfmadd_s(_k22, _r22, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k22, _r23, _sum1);
-
-                    __lsx_vst(_sum0, outptr0, 0);
-                    __lsx_vst(_sum1, outptr0 + 4, 0);
-
-                    outptr0 += 4 * 2;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                }
-                for (; j < outw; j++)
-                {
-                    __m128 _sum0 = (__m128)__lsx_vld(outptr0, 0);
-
-                    __m128i _r0 = __lsx_vld(r0, 0);
-                    __m128 _r00 = (__m128)__lsx_vreplvei_w(_r0, 0);
-                    __m128 _r01 = (__m128)__lsx_vreplvei_w(_r0, 1);
-                    __m128 _r02 = (__m128)__lsx_vreplvei_w(_r0, 2);
-
-                    _sum0 = __lsx_vfmadd_s(_k00, _r00, _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k01, _r01, _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k02, _r02, _sum0);
-
-                    __m128i _r1 = __lsx_vld(r1, 0);
-                    __m128 _r10 = (__m128)__lsx_vreplvei_w(_r1, 0);
-                    __m128 _r11 = (__m128)__lsx_vreplvei_w(_r1, 1);
-                    __m128 _r12 = (__m128)__lsx_vreplvei_w(_r1, 2);
-
-                    _sum0 = __lsx_vfmadd_s(_k10, _r10, _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k11, _r11, _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k12, _r12, _sum0);
-
-                    __m128i _r2 = __lsx_vld(r2, 0);
-                    __m128 _r20 = (__m128)__lsx_vreplvei_w(_r2, 0);
-                    __m128 _r21 = (__m128)__lsx_vreplvei_w(_r2, 1);
-                    __m128 _r22 = (__m128)__lsx_vreplvei_w(_r2, 2);
-
-                    _sum0 = __lsx_vfmadd_s(_k20, _r20, _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k21, _r21, _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k22, _r22, _sum0);
-
-                    __lsx_vst(_sum0, outptr0, 0);
-
-                    outptr0 += 4;
-
-                    r0 += 1;
-                    r1 += 1;
-                    r2 += 1;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * 4;
-        }
-    }
-}
-
-static void conv3x3s2_pack1to4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        __m128 _bias0 = bias ? (__m128)__lsx_vld(bias + p * 4, 0) : (__m128)__lsx_vreplgr2vr_w(0);
-        out0.fill(_bias0);
-
-        const float* k0 = kernel.channel(p);
-
-        int q = 0;
-        for (; q < inch; q++)
-        {
-            float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            __m128 _k00 = (__m128)__lsx_vld(k0, 0);
-            __m128 _k01 = (__m128)__lsx_vld(k0 + 4, 0);
-            __m128 _k02 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-            __m128 _k10 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-            __m128 _k11 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-            __m128 _k12 = (__m128)__lsx_vld(k0 + 4 * 5, 0);
-            __m128 _k20 = (__m128)__lsx_vld(k0 + 4 * 6, 0);
-            __m128 _k21 = (__m128)__lsx_vld(k0 + 4 * 7, 0);
-            __m128 _k22 = (__m128)__lsx_vld(k0 + 4 * 8, 0);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    __m128 _sum0 = (__m128)__lsx_vld(outptr0, 0);
-                    __m128 _sum1 = (__m128)__lsx_vld(outptr0 + 4, 0);
-                    __m128 _sum2 = (__m128)__lsx_vld(outptr0 + 4 * 2, 0);
-                    __m128 _sum3 = (__m128)__lsx_vld(outptr0 + 4 * 3, 0);
-                    __m128 _sum4 = (__m128)__lsx_vld(outptr0 + 4 * 4, 0);
-                    __m128 _sum5 = (__m128)__lsx_vld(outptr0 + 4 * 5, 0);
-                    __m128 _sum6 = (__m128)__lsx_vld(outptr0 + 4 * 6, 0);
-                    __m128 _sum7 = (__m128)__lsx_vld(outptr0 + 4 * 7, 0);
-
-                    __m128i _r0 = __lsx_vld(r0, 0);
-                    __m128i _r0n = __lsx_vld(r0 + 4, 0);
-                    __m128i _r0nn = __lsx_vld(r0 + 8, 0);
-                    __m128i _r0nnn = __lsx_vld(r0 + 12, 0);
-
-                    __m128 _r00 = (__m128)__lsx_vreplvei_w(_r0, 0);
-                    __m128 _r01 = (__m128)__lsx_vreplvei_w(_r0, 1);
-                    __m128 _r02 = (__m128)__lsx_vreplvei_w(_r0, 2);
-                    __m128 _r03 = (__m128)__lsx_vreplvei_w(_r0, 3);
-                    __m128 _r04 = (__m128)__lsx_vreplvei_w(_r0n, 0);
-                    __m128 _r05 = (__m128)__lsx_vreplvei_w(_r0n, 1);
-                    __m128 _r06 = (__m128)__lsx_vreplvei_w(_r0n, 2);
-                    __m128 _r07 = (__m128)__lsx_vreplvei_w(_r0n, 3);
-                    __m128 _r08 = (__m128)__lsx_vreplvei_w(_r0nn, 0);
-                    __m128 _r09 = (__m128)__lsx_vreplvei_w(_r0nn, 1);
-                    __m128 _r0a = (__m128)__lsx_vreplvei_w(_r0nn, 2);
-                    __m128 _r0b = (__m128)__lsx_vreplvei_w(_r0nn, 3);
-                    __m128 _r0c = (__m128)__lsx_vreplvei_w(_r0nnn, 0);
-                    __m128 _r0d = (__m128)__lsx_vreplvei_w(_r0nnn, 1);
-                    __m128 _r0e = (__m128)__lsx_vreplvei_w(_r0nnn, 2);
-                    __m128 _r0f = (__m128)__lsx_vreplvei_w(_r0nnn, 3);
-                    __m128 _r0g = __lsx_vreplfr2vr_s(r0[16]);
-
-                    _sum0 = __lsx_vfmadd_s(_k00, _r00, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k00, _r02, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k00, _r04, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k00, _r06, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k00, _r08, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k00, _r0a, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k00, _r0c, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k00, _r0e, _sum7);
-                    _sum0 = __lsx_vfmadd_s(_k01, _r01, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k01, _r03, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k01, _r05, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k01, _r07, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k01, _r09, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k01, _r0b, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k01, _r0d, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k01, _r0f, _sum7);
-                    _sum0 = __lsx_vfmadd_s(_k02, _r02, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k02, _r04, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k02, _r06, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k02, _r08, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k02, _r0a, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k02, _r0c, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k02, _r0e, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k02, _r0g, _sum7);
-
-                    __m128i _r1 = __lsx_vld(r1, 0);
-                    __m128i _r1n = __lsx_vld(r1 + 4, 0);
-                    __m128i _r1nn = __lsx_vld(r1 + 8, 0);
-                    __m128i _r1nnn = __lsx_vld(r1 + 12, 0);
-
-                    __m128 _r10 = (__m128)__lsx_vreplvei_w(_r1, 0);
-                    __m128 _r11 = (__m128)__lsx_vreplvei_w(_r1, 1);
-                    __m128 _r12 = (__m128)__lsx_vreplvei_w(_r1, 2);
-                    __m128 _r13 = (__m128)__lsx_vreplvei_w(_r1, 3);
-                    __m128 _r14 = (__m128)__lsx_vreplvei_w(_r1n, 0);
-                    __m128 _r15 = (__m128)__lsx_vreplvei_w(_r1n, 1);
-                    __m128 _r16 = (__m128)__lsx_vreplvei_w(_r1n, 2);
-                    __m128 _r17 = (__m128)__lsx_vreplvei_w(_r1n, 3);
-                    __m128 _r18 = (__m128)__lsx_vreplvei_w(_r1nn, 0);
-                    __m128 _r19 = (__m128)__lsx_vreplvei_w(_r1nn, 1);
-                    __m128 _r1a = (__m128)__lsx_vreplvei_w(_r1nn, 2);
-                    __m128 _r1b = (__m128)__lsx_vreplvei_w(_r1nn, 3);
-                    __m128 _r1c = (__m128)__lsx_vreplvei_w(_r1nnn, 0);
-                    __m128 _r1d = (__m128)__lsx_vreplvei_w(_r1nnn, 1);
-                    __m128 _r1e = (__m128)__lsx_vreplvei_w(_r1nnn, 2);
-                    __m128 _r1f = (__m128)__lsx_vreplvei_w(_r1nnn, 3);
-                    __m128 _r1g = __lsx_vreplfr2vr_s(r1[16]);
-
-                    _sum0 = __lsx_vfmadd_s(_k10, _r10, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k10, _r12, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k10, _r14, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k10, _r16, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k10, _r18, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k10, _r1a, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k10, _r1c, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k10, _r1e, _sum7);
-                    _sum0 = __lsx_vfmadd_s(_k11, _r11, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k11, _r13, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k11, _r15, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k11, _r17, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k11, _r19, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k11, _r1b, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k11, _r1d, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k11, _r1f, _sum7);
-                    _sum0 = __lsx_vfmadd_s(_k12, _r12, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k12, _r14, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k12, _r16, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k12, _r18, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k12, _r1a, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k12, _r1c, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k12, _r1e, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k12, _r1g, _sum7);
-
-                    __m128i _r2 = __lsx_vld(r2, 0);
-                    __m128i _r2n = __lsx_vld(r2 + 4, 0);
-                    __m128i _r2nn = __lsx_vld(r2 + 8, 0);
-                    __m128i _r2nnn = __lsx_vld(r2 + 12, 0);
-
-                    __m128 _r20 = (__m128)__lsx_vreplvei_w(_r2, 0);
-                    __m128 _r21 = (__m128)__lsx_vreplvei_w(_r2, 1);
-                    __m128 _r22 = (__m128)__lsx_vreplvei_w(_r2, 2);
-                    __m128 _r23 = (__m128)__lsx_vreplvei_w(_r2, 3);
-                    __m128 _r24 = (__m128)__lsx_vreplvei_w(_r2n, 0);
-                    __m128 _r25 = (__m128)__lsx_vreplvei_w(_r2n, 1);
-                    __m128 _r26 = (__m128)__lsx_vreplvei_w(_r2n, 2);
-                    __m128 _r27 = (__m128)__lsx_vreplvei_w(_r2n, 3);
-                    __m128 _r28 = (__m128)__lsx_vreplvei_w(_r2nn, 0);
-                    __m128 _r29 = (__m128)__lsx_vreplvei_w(_r2nn, 1);
-                    __m128 _r2a = (__m128)__lsx_vreplvei_w(_r2nn, 2);
-                    __m128 _r2b = (__m128)__lsx_vreplvei_w(_r2nn, 3);
-                    __m128 _r2c = (__m128)__lsx_vreplvei_w(_r2nnn, 0);
-                    __m128 _r2d = (__m128)__lsx_vreplvei_w(_r2nnn, 1);
-                    __m128 _r2e = (__m128)__lsx_vreplvei_w(_r2nnn, 2);
-                    __m128 _r2f = (__m128)__lsx_vreplvei_w(_r2nnn, 3);
-                    __m128 _r2g = __lsx_vreplfr2vr_s(r2[16]);
-
-                    _sum0 = __lsx_vfmadd_s(_k20, _r20, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k20, _r22, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k20, _r24, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k20, _r26, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k20, _r28, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k20, _r2a, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k20, _r2c, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k20, _r2e, _sum7);
-                    _sum0 = __lsx_vfmadd_s(_k21, _r21, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k21, _r23, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k21, _r25, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k21, _r27, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k21, _r29, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k21, _r2b, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k21, _r2d, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k21, _r2f, _sum7);
-                    _sum0 = __lsx_vfmadd_s(_k22, _r22, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k22, _r24, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k22, _r26, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k22, _r28, _sum3);
-                    _sum4 = __lsx_vfmadd_s(_k22, _r2a, _sum4);
-                    _sum5 = __lsx_vfmadd_s(_k22, _r2c, _sum5);
-                    _sum6 = __lsx_vfmadd_s(_k22, _r2e, _sum6);
-                    _sum7 = __lsx_vfmadd_s(_k22, _r2g, _sum7);
-
-                    __lsx_vst(_sum0, outptr0, 0);
-                    __lsx_vst(_sum1, outptr0 + 4, 0);
-                    __lsx_vst(_sum2, outptr0 + 4 * 2, 0);
-                    __lsx_vst(_sum3, outptr0 + 4 * 3, 0);
-                    __lsx_vst(_sum4, outptr0 + 4 * 4, 0);
-                    __lsx_vst(_sum5, outptr0 + 4 * 5, 0);
-                    __lsx_vst(_sum6, outptr0 + 4 * 6, 0);
-                    __lsx_vst(_sum7, outptr0 + 4 * 7, 0);
-
-                    outptr0 += 4 * 8;
-
-                    r0 += 16;
-                    r1 += 16;
-                    r2 += 16;
-                }
-                for (; j + 3 < outw; j += 4)
-                {
-                    __m128 _sum0 = (__m128)__lsx_vld(outptr0, 0);
-                    __m128 _sum1 = (__m128)__lsx_vld(outptr0 + 4, 0);
-                    __m128 _sum2 = (__m128)__lsx_vld(outptr0 + 4 * 2, 0);
-                    __m128 _sum3 = (__m128)__lsx_vld(outptr0 + 4 * 3, 0);
-
-                    __m128i _r0 = __lsx_vld(r0, 0);
-                    __m128i _r0n = __lsx_vld(r0 + 4, 0);
-
-                    __m128 _r00 = (__m128)__lsx_vreplvei_w(_r0, 0);
-                    __m128 _r01 = (__m128)__lsx_vreplvei_w(_r0, 1);
-                    __m128 _r02 = (__m128)__lsx_vreplvei_w(_r0, 2);
-                    __m128 _r03 = (__m128)__lsx_vreplvei_w(_r0, 3);
-                    __m128 _r04 = (__m128)__lsx_vreplvei_w(_r0n, 0);
-                    __m128 _r05 = (__m128)__lsx_vreplvei_w(_r0n, 1);
-                    __m128 _r06 = (__m128)__lsx_vreplvei_w(_r0n, 2);
-                    __m128 _r07 = (__m128)__lsx_vreplvei_w(_r0n, 3);
-                    __m128 _r08 = __lsx_vreplfr2vr_s(r0[8]);
-
-                    _sum0 = __lsx_vfmadd_s(_k00, _r00, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k00, _r02, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k00, _r04, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k00, _r06, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k01, _r01, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k01, _r03, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k01, _r05, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k01, _r07, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k02, _r02, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k02, _r04, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k02, _r06, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k02, _r08, _sum3);
-
-                    __m128i _r1 = __lsx_vld(r1, 0);
-                    __m128i _r1n = __lsx_vld(r1 + 4, 0);
-
-                    __m128 _r10 = (__m128)__lsx_vreplvei_w(_r1, 0);
-                    __m128 _r11 = (__m128)__lsx_vreplvei_w(_r1, 1);
-                    __m128 _r12 = (__m128)__lsx_vreplvei_w(_r1, 2);
-                    __m128 _r13 = (__m128)__lsx_vreplvei_w(_r1, 3);
-                    __m128 _r14 = (__m128)__lsx_vreplvei_w(_r1n, 0);
-                    __m128 _r15 = (__m128)__lsx_vreplvei_w(_r1n, 1);
-                    __m128 _r16 = (__m128)__lsx_vreplvei_w(_r1n, 2);
-                    __m128 _r17 = (__m128)__lsx_vreplvei_w(_r1n, 3);
-                    __m128 _r18 = __lsx_vreplfr2vr_s(r1[8]);
-
-                    _sum0 = __lsx_vfmadd_s(_k10, _r10, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k10, _r12, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k10, _r14, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k10, _r16, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k11, _r11, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k11, _r13, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k11, _r15, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k11, _r17, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k12, _r12, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k12, _r14, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k12, _r16, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k12, _r18, _sum3);
-
-                    __m128i _r2 = __lsx_vld(r2, 0);
-                    __m128i _r2n = __lsx_vld(r2 + 4, 0);
-
-                    __m128 _r20 = (__m128)__lsx_vreplvei_w(_r2, 0);
-                    __m128 _r21 = (__m128)__lsx_vreplvei_w(_r2, 1);
-                    __m128 _r22 = (__m128)__lsx_vreplvei_w(_r2, 2);
-                    __m128 _r23 = (__m128)__lsx_vreplvei_w(_r2, 3);
-                    __m128 _r24 = (__m128)__lsx_vreplvei_w(_r2n, 0);
-                    __m128 _r25 = (__m128)__lsx_vreplvei_w(_r2n, 1);
-                    __m128 _r26 = (__m128)__lsx_vreplvei_w(_r2n, 2);
-                    __m128 _r27 = (__m128)__lsx_vreplvei_w(_r2n, 3);
-                    __m128 _r28 = __lsx_vreplfr2vr_s(r2[8]);
-
-                    _sum0 = __lsx_vfmadd_s(_k20, _r20, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k20, _r22, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k20, _r24, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k20, _r26, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k21, _r21, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k21, _r23, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k21, _r25, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k21, _r27, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k22, _r22, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k22, _r24, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k22, _r26, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k22, _r28, _sum3);
-
-                    __lsx_vst(_sum0, outptr0, 0);
-                    __lsx_vst(_sum1, outptr0 + 4, 0);
-                    __lsx_vst(_sum2, outptr0 + 4 * 2, 0);
-                    __lsx_vst(_sum3, outptr0 + 4 * 3, 0);
-
-                    outptr0 += 4 * 4;
-
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    __m128 _sum0 = (__m128)__lsx_vld(outptr0, 0);
-                    __m128 _sum1 = (__m128)__lsx_vld(outptr0 + 4, 0);
-
-                    __m128i _r0 = __lsx_vld(r0, 0);
-                    __m128 _r00 = (__m128)__lsx_vreplvei_w(_r0, 0);
-                    __m128 _r01 = (__m128)__lsx_vreplvei_w(_r0, 1);
-                    __m128 _r02 = (__m128)__lsx_vreplvei_w(_r0, 2);
-                    __m128 _r03 = (__m128)__lsx_vreplvei_w(_r0, 3);
-                    __m128 _r04 = __lsx_vreplfr2vr_s(r0[4]);
-
-                    _sum0 = __lsx_vfmadd_s(_k00, _r00, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k00, _r02, _sum1);
-                    _sum0 = __lsx_vfmadd_s(_k01, _r01, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k01, _r03, _sum1);
-                    _sum0 = __lsx_vfmadd_s(_k02, _r02, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k02, _r04, _sum1);
-
-                    __m128i _r1 = __lsx_vld(r1, 0);
-                    __m128 _r10 = (__m128)__lsx_vreplvei_w(_r1, 0);
-                    __m128 _r11 = (__m128)__lsx_vreplvei_w(_r1, 1);
-                    __m128 _r12 = (__m128)__lsx_vreplvei_w(_r1, 2);
-                    __m128 _r13 = (__m128)__lsx_vreplvei_w(_r1, 3);
-                    __m128 _r14 = __lsx_vreplfr2vr_s(r1[4]);
-
-                    _sum0 = __lsx_vfmadd_s(_k10, _r10, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k10, _r12, _sum1);
-                    _sum0 = __lsx_vfmadd_s(_k11, _r11, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k11, _r13, _sum1);
-                    _sum0 = __lsx_vfmadd_s(_k12, _r12, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k12, _r14, _sum1);
-
-                    __m128i _r2 = __lsx_vld(r2, 0);
-                    __m128 _r20 = (__m128)__lsx_vreplvei_w(_r2, 0);
-                    __m128 _r21 = (__m128)__lsx_vreplvei_w(_r2, 1);
-                    __m128 _r22 = (__m128)__lsx_vreplvei_w(_r2, 2);
-                    __m128 _r23 = (__m128)__lsx_vreplvei_w(_r2, 3);
-                    __m128 _r24 = __lsx_vreplfr2vr_s(r2[4]);
-
-                    _sum0 = __lsx_vfmadd_s(_k20, _r20, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k20, _r22, _sum1);
-                    _sum0 = __lsx_vfmadd_s(_k21, _r21, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k21, _r23, _sum1);
-                    _sum0 = __lsx_vfmadd_s(_k22, _r22, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k22, _r24, _sum1);
-
-                    __lsx_vst(_sum0, outptr0, 0);
-                    __lsx_vst(_sum1, outptr0 + 4, 0);
-
-                    outptr0 += 4 * 2;
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                }
-                for (; j < outw; j++)
-                {
-                    __m128 _sum0 = (__m128)__lsx_vld(outptr0, 0);
-
-                    __m128i _r0 = __lsx_vld(r0, 0);
-                    __m128 _r00 = (__m128)__lsx_vreplvei_w(_r0, 0);
-                    __m128 _r01 = (__m128)__lsx_vreplvei_w(_r0, 1);
-                    __m128 _r02 = (__m128)__lsx_vreplvei_w(_r0, 2);
-
-                    _sum0 = __lsx_vfmadd_s(_k00, _r00, _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k01, _r01, _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k02, _r02, _sum0);
-
-                    __m128i _r1 = __lsx_vld(r1, 0);
-                    __m128 _r10 = (__m128)__lsx_vreplvei_w(_r1, 0);
-                    __m128 _r11 = (__m128)__lsx_vreplvei_w(_r1, 1);
-                    __m128 _r12 = (__m128)__lsx_vreplvei_w(_r1, 2);
-
-                    _sum0 = __lsx_vfmadd_s(_k10, _r10, _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k11, _r11, _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k12, _r12, _sum0);
-
-                    __m128i _r2 = __lsx_vld(r2, 0);
-                    __m128 _r20 = (__m128)__lsx_vreplvei_w(_r2, 0);
-                    __m128 _r21 = (__m128)__lsx_vreplvei_w(_r2, 1);
-                    __m128 _r22 = (__m128)__lsx_vreplvei_w(_r2, 2);
-
-                    _sum0 = __lsx_vfmadd_s(_k20, _r20, _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k21, _r21, _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k22, _r22, _sum0);
-
-                    __lsx_vst(_sum0, outptr0, 0);
-
-                    outptr0 += 4;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack4.h
deleted file mode 100644
index f06bb7e9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack4.h
+++ /dev/null
@@ -1,425 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd63_transform_kernel_pack4_lsx(const Mat& kernel, Mat& kernel_tm_pack4, int inch, int outch, const Option& opt)
-{
-    // winograd63 transform kernel
-    Mat kernel_tm;
-    kernel_tm.create(8 * 8, inch, outch);
-
-    const float ktm[8][3] = {
-        {1.0f, 0.0f, 0.0f},
-        {-2.0f / 9, -2.0f / 9, -2.0f / 9},
-        {-2.0f / 9, 2.0f / 9, -2.0f / 9},
-        {1.0f / 90, 1.0f / 45, 2.0f / 45},
-        {1.0f / 90, -1.0f / 45, 2.0f / 45},
-        {1.0f / 45, 1.0f / 90, 1.0f / 180},
-        {1.0f / 45, -1.0f / 90, 1.0f / 180},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel, transposed
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[8][3];
-            for (int i = 0; i < 8; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // v
-            for (int j = 0; j < 8; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 8; i++)
-                {
-                    kernel_tm0[j * 8 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 64-inch-outch
-    // dst = pb-pa-inch/pa-64-outch/pb
-    kernel_tm_pack4.create(inch / 4, 64, outch / 4, (size_t)4u * 4 * 4, 4 * 4);
-
-    for (int q = 0; q + 3 < outch; q += 4)
-    {
-        Mat g0 = kernel_tm_pack4.channel(q / 4);
-
-        for (int k = 0; k < 64; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p + 3 < inch; p += 4)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 4; j++)
-                    {
-                        const float* k00 = kernel_tm.channel(q + j).row(p + i);
-                        g00[0] = k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd63_pack4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 6n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 5) / 6 * 6;
-    outh = (outh + 5) / 6 * 6;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 6;
-        int h_tiles = outh / 6;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 64, inch, elemsize, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd63_transform_input_pack4_lsx(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_pack4_lsx(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, elemsize, elempack, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd63_transform_output_pack4_lsx(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
-
-static void conv3x3s1_winograd43_transform_kernel_pack4_lsx(const Mat& kernel, Mat& kernel_tm_pack4, int inch, int outch, const Option& opt)
-{
-    // winograd43 transform kernel
-    Mat kernel_tm(6 * 6, inch, outch);
-
-    const float ktm[6][3] = {
-        {1.0f / 4, 0.0f, 0.0f},
-        {-1.0f / 6, -1.0f / 6, -1.0f / 6},
-        {-1.0f / 6, 1.0f / 6, -1.0f / 6},
-        {1.0f / 24, 1.0f / 12, 1.0f / 6},
-        {1.0f / 24, -1.0f / 12, 1.0f / 6},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[6][3];
-            for (int i = 0; i < 6; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 6; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 6; i++)
-                {
-                    kernel_tm0[j * 6 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 36-inch-outch
-    // dst = pb-pa-inch/pa-36-outch/pb
-    kernel_tm_pack4.create(inch / 4, 36, outch / 4, (size_t)4u * 4 * 4, 4 * 4);
-
-    for (int q = 0; q + 3 < outch; q += 4)
-    {
-        Mat g0 = kernel_tm_pack4.channel(q / 4);
-
-        for (int k = 0; k < 36; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p + 3 < inch; p += 4)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 4; j++)
-                    {
-                        const float* k00 = kernel_tm.channel(q + j).row(p + i);
-                        g00[0] = k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_pack4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 4n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 3) / 4 * 4;
-    outh = (outh + 3) / 4 * 4;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 4;
-        int h_tiles = outh / 4;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 36, inch, elemsize, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd43_transform_input_pack4_lsx(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_pack4_lsx(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, elemsize, elempack, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd43_transform_output_pack4_lsx(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
-
-static void conv3x3s1_winograd23_transform_kernel_pack4_lsx(const Mat& kernel, Mat& kernel_tm_pack4, int inch, int outch, const Option& opt)
-{
-    // winograd23 transform kernel
-    Mat kernel_tm(4 * 4, inch, outch);
-
-    const float ktm[4][3] = {
-        {1.0f, 0.0f, 0.0f},
-        {1.0f / 2, 1.0f / 2, 1.0f / 2},
-        {1.0f / 2, -1.0f / 2, 1.0f / 2},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[4][3];
-            for (int i = 0; i < 4; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 4; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 4; i++)
-                {
-                    kernel_tm0[j * 4 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 16-inch-outch
-    // dst = pb-pa-inch/pa-16-outch/pb
-    kernel_tm_pack4.create(inch / 4, 16, outch / 4, (size_t)4u * 4 * 4, 4 * 4);
-
-    for (int q = 0; q + 3 < outch; q += 4)
-    {
-        Mat g0 = kernel_tm_pack4.channel(q / 4);
-
-        for (int k = 0; k < 16; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p + 3 < inch; p += 4)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 4; j++)
-                    {
-                        const float* k00 = kernel_tm.channel(q + j).row(p + i);
-                        g00[0] = k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_pack4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 2n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 1) / 2 * 2;
-    outh = (outh + 1) / 2 * 2;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 2;
-        int h_tiles = outh / 2;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 16, inch, elemsize, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd23_transform_input_pack4_lsx(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_pack4_lsx(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, elemsize, elempack, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd23_transform_output_pack4_lsx(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack8to1_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack8to1_int8.h
deleted file mode 100644
index 3c4f9718..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack8to1_int8.h
+++ /dev/null
@@ -1,177 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd43_transform_kernel_pack8to1_int8_lsx(const Mat& kernel, Mat& kernel_tm_pack8to1, int inch, int outch, const Option& opt)
-{
-    // winograd43 transform kernel
-    Mat kernel_tm(6 * 6, inch, outch, (size_t)2u);
-
-    const short ktm[6][3] = {
-        {6, 0, 0},
-        {-4, -4, -4},
-        {-4, 4, -4},
-        {1, 2, 4},
-        {1, -2, 4},
-        {0, 0, 6}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const signed char* kernel0 = (const signed char*)kernel + p * inch * 9 + q * 9;
-            short* kernel_tm0 = kernel_tm.channel(p).row<short>(q);
-
-            // transform kernel
-            const signed char* k0 = kernel0;
-            const signed char* k1 = kernel0 + 3;
-            const signed char* k2 = kernel0 + 6;
-
-            // h
-            short tmp[6][3];
-            for (int i = 0; i < 6; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 6; j++)
-            {
-                short* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 6; i++)
-                {
-                    kernel_tm0[j * 6 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 36-inch-outch
-    // dst = 4b-8a-inch/8a-36-outch/4b
-    kernel_tm_pack8to1.create(8 * inch / 8, 36, outch / 4 + outch % 4, (size_t)2u * 4, 4);
-
-    int p = 0;
-    for (; p + 3 < outch; p += 4)
-    {
-        const Mat k0 = kernel_tm.channel(p);
-        const Mat k1 = kernel_tm.channel(p + 1);
-        const Mat k2 = kernel_tm.channel(p + 2);
-        const Mat k3 = kernel_tm.channel(p + 3);
-
-        Mat g0 = kernel_tm_pack8to1.channel(p / 4);
-
-        for (int k = 0; k < 36; k++)
-        {
-            short* g00 = g0.row<short>(k);
-
-            for (int q = 0; q + 7 < inch; q += 8)
-            {
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0.row<const short>(q + i)[k];
-                    g00[1] = k1.row<const short>(q + i)[k];
-                    g00[2] = k2.row<const short>(q + i)[k];
-                    g00[3] = k3.row<const short>(q + i)[k];
-
-                    g00 += 4;
-                }
-            }
-        }
-    }
-    for (; p < outch; p++)
-    {
-        const Mat k0 = kernel_tm.channel(p);
-
-        Mat g0 = kernel_tm_pack8to1.channel(p / 4 + p % 4);
-
-        for (int k = 0; k < 36; k++)
-        {
-            short* g00 = g0.row<short>(k);
-
-            for (int q = 0; q + 7 < inch; q += 8)
-            {
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0.row<const short>(q + i)[k];
-
-                    g00 += 1;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_pack8to1_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    //     size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 4n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 3) / 4 * 4;
-    outh = (outh + 3) / 4 * 4;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 4;
-        int h_tiles = outh / 4;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 36, inch, 2u * elempack, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd43_transform_input_pack8_int8_lsx(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_pack8to1_int8_lsx(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, 4u, 1, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd43_transform_output_int8_lsx(top_blob_tm, top_blob_bordered, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack8to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack8to4_int8.h
deleted file mode 100644
index bf328cee..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_3x3_pack8to4_int8.h
+++ /dev/null
@@ -1,161 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd43_transform_kernel_pack8to4_int8_lsx(const Mat& kernel, Mat& kernel_tm_pack8, int inch, int outch, const Option& opt)
-{
-    // winograd43 transform kernel
-    Mat kernel_tm(6 * 6, inch, outch, (size_t)2u);
-
-    const short ktm[6][3] = {
-        {6, 0, 0},
-        {-4, -4, -4},
-        {-4, 4, -4},
-        {1, 2, 4},
-        {1, -2, 4},
-        {0, 0, 6}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const signed char* kernel0 = (const signed char*)kernel + p * inch * 9 + q * 9;
-            short* kernel_tm0 = kernel_tm.channel(p).row<short>(q);
-
-            // transform kernel
-            const signed char* k0 = kernel0;
-            const signed char* k1 = kernel0 + 3;
-            const signed char* k2 = kernel0 + 6;
-
-            // h
-            short tmp[6][3];
-            for (int i = 0; i < 6; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 6; j++)
-            {
-                short* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 6; i++)
-                {
-                    kernel_tm0[j * 6 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 36-inch-outch
-    // dst = 4b-8a-inch/8a-36-outch/4b
-    kernel_tm_pack8.create(inch / 8, 36, outch / 4, (size_t)2u * 32, 32);
-
-    int q = 0;
-    for (; q + 3 < outch; q += 4)
-    {
-        const Mat k0 = kernel_tm.channel(q);
-        const Mat k1 = kernel_tm.channel(q + 1);
-        const Mat k2 = kernel_tm.channel(q + 2);
-        const Mat k3 = kernel_tm.channel(q + 3);
-
-        Mat kernel_tm = kernel_tm_pack8.channel(q / 4);
-
-        for (int k = 0; k < 36; k++)
-        {
-            short* g00 = kernel_tm.row<short>(k);
-
-            for (int p = 0; p + 7 < inch; p += 8)
-            {
-                for (int i = 0; i < 8; i++)
-                {
-                    const short* k00 = k0.row<const short>(p + i);
-                    const short* k10 = k1.row<const short>(p + i);
-                    const short* k20 = k2.row<const short>(p + i);
-                    const short* k30 = k3.row<const short>(p + i);
-
-                    g00[0] = k00[k];
-                    g00[1] = k10[k];
-                    g00[2] = k20[k];
-                    g00[3] = k30[k];
-
-                    g00 += 4;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_pack8to4_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    //     size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 4n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 3) / 4 * 4;
-    outh = (outh + 3) / 4 * 4;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 4;
-        int h_tiles = outh / 4;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 36, inch, 2u * elempack, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd43_transform_input_pack8_int8_lsx(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_pack8to4_int8_lsx(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, 4u * 4, 4, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd43_transform_output_pack4_int8_lsx(top_blob_tm, top_blob_bordered, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_7x7_pack1to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_7x7_pack1to4.h
deleted file mode 100644
index f57923b5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_7x7_pack1to4.h
+++ /dev/null
@@ -1,652 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv7x7s2_pack1to4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        __m128 _bias0 = bias ? (__m128)__lsx_vld(bias + p * 4, 0) : (__m128)__lsx_vreplgr2vr_w(0);
-        out0.fill(_bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-            const float* r3 = img0.row(3);
-            const float* r4 = img0.row(4);
-            const float* r5 = img0.row(5);
-            const float* r6 = img0.row(6);
-
-            const float* kptr = kernel.channel(p).row(q);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-                    __m128 _sum0 = (__m128)__lsx_vld(outptr0, 0);
-                    __m128 _sum1 = (__m128)__lsx_vld(outptr0 + 4, 0);
-                    __m128 _sum2 = (__m128)__lsx_vld(outptr0 + 4 * 2, 0);
-                    __m128 _sum3 = (__m128)__lsx_vld(outptr0 + 4 * 3, 0);
-
-                    __m128 _k00 = (__m128)__lsx_vld(kptr, 0);
-                    __m128 _k01 = (__m128)__lsx_vld(kptr + 4, 0);
-                    __m128 _k02 = (__m128)__lsx_vld(kptr + 4 * 2, 0);
-                    __m128 _k03 = (__m128)__lsx_vld(kptr + 4 * 3, 0);
-                    __m128 _k04 = (__m128)__lsx_vld(kptr + 4 * 4, 0);
-                    __m128 _k05 = (__m128)__lsx_vld(kptr + 4 * 5, 0);
-                    __m128 _k06 = (__m128)__lsx_vld(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    __m128i _r0 = __lsx_vld(r0, 0);
-                    __m128i _r0n = __lsx_vld(r0 + 4, 0);
-                    __m128i _r0nn = __lsx_vld(r0 + 8, 0);
-
-                    __m128 _r00 = (__m128)__lsx_vreplvei_w(_r0, 0);
-                    __m128 _r01 = (__m128)__lsx_vreplvei_w(_r0, 1);
-                    __m128 _r02 = (__m128)__lsx_vreplvei_w(_r0, 2);
-                    __m128 _r03 = (__m128)__lsx_vreplvei_w(_r0, 3);
-                    __m128 _r04 = (__m128)__lsx_vreplvei_w(_r0n, 0);
-                    __m128 _r05 = (__m128)__lsx_vreplvei_w(_r0n, 1);
-                    __m128 _r06 = (__m128)__lsx_vreplvei_w(_r0n, 2);
-                    __m128 _r07 = (__m128)__lsx_vreplvei_w(_r0n, 3);
-                    __m128 _r08 = (__m128)__lsx_vreplvei_w(_r0nn, 0);
-                    __m128 _r09 = (__m128)__lsx_vreplvei_w(_r0nn, 1);
-                    __m128 _r0a = (__m128)__lsx_vreplvei_w(_r0nn, 2);
-                    __m128 _r0b = (__m128)__lsx_vreplvei_w(_r0nn, 3);
-                    __m128 _r0c = __lsx_vreplfr2vr_s(r0[12]);
-
-                    _sum0 = __lsx_vfmadd_s(_k00, _r00, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k00, _r02, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k00, _r04, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k00, _r06, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k01, _r01, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k01, _r03, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k01, _r05, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k01, _r07, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k02, _r02, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k02, _r04, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k02, _r06, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k02, _r08, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k03, _r03, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k03, _r05, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k03, _r07, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k03, _r09, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k04, _r04, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k04, _r06, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k04, _r08, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k04, _r0a, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k05, _r05, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k05, _r07, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k05, _r09, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k05, _r0b, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k06, _r06, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k06, _r08, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k06, _r0a, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k06, _r0c, _sum3);
-
-                    __m128 _k10 = (__m128)__lsx_vld(kptr, 0);
-                    __m128 _k11 = (__m128)__lsx_vld(kptr + 4, 0);
-                    __m128 _k12 = (__m128)__lsx_vld(kptr + 4 * 2, 0);
-                    __m128 _k13 = (__m128)__lsx_vld(kptr + 4 * 3, 0);
-                    __m128 _k14 = (__m128)__lsx_vld(kptr + 4 * 4, 0);
-                    __m128 _k15 = (__m128)__lsx_vld(kptr + 4 * 5, 0);
-                    __m128 _k16 = (__m128)__lsx_vld(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    __m128i _r1 = __lsx_vld(r1, 0);
-                    __m128i _r1n = __lsx_vld(r1 + 4, 0);
-                    __m128i _r1nn = __lsx_vld(r1 + 8, 0);
-
-                    __m128 _r10 = (__m128)__lsx_vreplvei_w(_r1, 0);
-                    __m128 _r11 = (__m128)__lsx_vreplvei_w(_r1, 1);
-                    __m128 _r12 = (__m128)__lsx_vreplvei_w(_r1, 2);
-                    __m128 _r13 = (__m128)__lsx_vreplvei_w(_r1, 3);
-                    __m128 _r14 = (__m128)__lsx_vreplvei_w(_r1n, 0);
-                    __m128 _r15 = (__m128)__lsx_vreplvei_w(_r1n, 1);
-                    __m128 _r16 = (__m128)__lsx_vreplvei_w(_r1n, 2);
-                    __m128 _r17 = (__m128)__lsx_vreplvei_w(_r1n, 3);
-                    __m128 _r18 = (__m128)__lsx_vreplvei_w(_r1nn, 0);
-                    __m128 _r19 = (__m128)__lsx_vreplvei_w(_r1nn, 1);
-                    __m128 _r1a = (__m128)__lsx_vreplvei_w(_r1nn, 2);
-                    __m128 _r1b = (__m128)__lsx_vreplvei_w(_r1nn, 3);
-                    __m128 _r1c = __lsx_vreplfr2vr_s(r1[12]);
-
-                    _sum0 = __lsx_vfmadd_s(_k10, _r10, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k10, _r12, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k10, _r14, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k10, _r16, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k11, _r11, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k11, _r13, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k11, _r15, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k11, _r17, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k12, _r12, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k12, _r14, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k12, _r16, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k12, _r18, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k13, _r13, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k13, _r15, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k13, _r17, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k13, _r19, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k14, _r14, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k14, _r16, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k14, _r18, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k14, _r1a, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k15, _r15, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k15, _r17, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k15, _r19, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k15, _r1b, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k16, _r16, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k16, _r18, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k16, _r1a, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k16, _r1c, _sum3);
-
-                    __m128 _k20 = (__m128)__lsx_vld(kptr, 0);
-                    __m128 _k21 = (__m128)__lsx_vld(kptr + 4, 0);
-                    __m128 _k22 = (__m128)__lsx_vld(kptr + 4 * 2, 0);
-                    __m128 _k23 = (__m128)__lsx_vld(kptr + 4 * 3, 0);
-                    __m128 _k24 = (__m128)__lsx_vld(kptr + 4 * 4, 0);
-                    __m128 _k25 = (__m128)__lsx_vld(kptr + 4 * 5, 0);
-                    __m128 _k26 = (__m128)__lsx_vld(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    __m128i _r2 = __lsx_vld(r2, 0);
-                    __m128i _r2n = __lsx_vld(r2 + 4, 0);
-                    __m128i _r2nn = __lsx_vld(r2 + 8, 0);
-
-                    __m128 _r20 = (__m128)__lsx_vreplvei_w(_r2, 0);
-                    __m128 _r21 = (__m128)__lsx_vreplvei_w(_r2, 1);
-                    __m128 _r22 = (__m128)__lsx_vreplvei_w(_r2, 2);
-                    __m128 _r23 = (__m128)__lsx_vreplvei_w(_r2, 3);
-                    __m128 _r24 = (__m128)__lsx_vreplvei_w(_r2n, 0);
-                    __m128 _r25 = (__m128)__lsx_vreplvei_w(_r2n, 1);
-                    __m128 _r26 = (__m128)__lsx_vreplvei_w(_r2n, 2);
-                    __m128 _r27 = (__m128)__lsx_vreplvei_w(_r2n, 3);
-                    __m128 _r28 = (__m128)__lsx_vreplvei_w(_r2nn, 0);
-                    __m128 _r29 = (__m128)__lsx_vreplvei_w(_r2nn, 1);
-                    __m128 _r2a = (__m128)__lsx_vreplvei_w(_r2nn, 2);
-                    __m128 _r2b = (__m128)__lsx_vreplvei_w(_r2nn, 3);
-                    __m128 _r2c = __lsx_vreplfr2vr_s(r2[12]);
-
-                    _sum0 = __lsx_vfmadd_s(_k20, _r20, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k20, _r22, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k20, _r24, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k20, _r26, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k21, _r21, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k21, _r23, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k21, _r25, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k21, _r27, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k22, _r22, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k22, _r24, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k22, _r26, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k22, _r28, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k23, _r23, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k23, _r25, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k23, _r27, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k23, _r29, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k24, _r24, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k24, _r26, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k24, _r28, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k24, _r2a, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k25, _r25, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k25, _r27, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k25, _r29, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k25, _r2b, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k26, _r26, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k26, _r28, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k26, _r2a, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k26, _r2c, _sum3);
-
-                    __m128 _k30 = (__m128)__lsx_vld(kptr, 0);
-                    __m128 _k31 = (__m128)__lsx_vld(kptr + 4, 0);
-                    __m128 _k32 = (__m128)__lsx_vld(kptr + 4 * 2, 0);
-                    __m128 _k33 = (__m128)__lsx_vld(kptr + 4 * 3, 0);
-                    __m128 _k34 = (__m128)__lsx_vld(kptr + 4 * 4, 0);
-                    __m128 _k35 = (__m128)__lsx_vld(kptr + 4 * 5, 0);
-                    __m128 _k36 = (__m128)__lsx_vld(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    __m128i _r3 = __lsx_vld(r3, 0);
-                    __m128i _r3n = __lsx_vld(r3 + 4, 0);
-                    __m128i _r3nn = __lsx_vld(r3 + 8, 0);
-
-                    __m128 _r30 = (__m128)__lsx_vreplvei_w(_r3, 0);
-                    __m128 _r31 = (__m128)__lsx_vreplvei_w(_r3, 1);
-                    __m128 _r32 = (__m128)__lsx_vreplvei_w(_r3, 2);
-                    __m128 _r33 = (__m128)__lsx_vreplvei_w(_r3, 3);
-                    __m128 _r34 = (__m128)__lsx_vreplvei_w(_r3n, 0);
-                    __m128 _r35 = (__m128)__lsx_vreplvei_w(_r3n, 1);
-                    __m128 _r36 = (__m128)__lsx_vreplvei_w(_r3n, 2);
-                    __m128 _r37 = (__m128)__lsx_vreplvei_w(_r3n, 3);
-                    __m128 _r38 = (__m128)__lsx_vreplvei_w(_r3nn, 0);
-                    __m128 _r39 = (__m128)__lsx_vreplvei_w(_r3nn, 1);
-                    __m128 _r3a = (__m128)__lsx_vreplvei_w(_r3nn, 2);
-                    __m128 _r3b = (__m128)__lsx_vreplvei_w(_r3nn, 3);
-                    __m128 _r3c = __lsx_vreplfr2vr_s(r3[12]);
-
-                    _sum0 = __lsx_vfmadd_s(_k30, _r30, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k30, _r32, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k30, _r34, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k30, _r36, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k31, _r31, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k31, _r33, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k31, _r35, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k31, _r37, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k32, _r32, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k32, _r34, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k32, _r36, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k32, _r38, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k33, _r33, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k33, _r35, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k33, _r37, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k33, _r39, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k34, _r34, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k34, _r36, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k34, _r38, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k34, _r3a, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k35, _r35, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k35, _r37, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k35, _r39, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k35, _r3b, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k36, _r36, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k36, _r38, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k36, _r3a, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k36, _r3c, _sum3);
-
-                    __m128 _k40 = (__m128)__lsx_vld(kptr, 0);
-                    __m128 _k41 = (__m128)__lsx_vld(kptr + 4, 0);
-                    __m128 _k42 = (__m128)__lsx_vld(kptr + 4 * 2, 0);
-                    __m128 _k43 = (__m128)__lsx_vld(kptr + 4 * 3, 0);
-                    __m128 _k44 = (__m128)__lsx_vld(kptr + 4 * 4, 0);
-                    __m128 _k45 = (__m128)__lsx_vld(kptr + 4 * 5, 0);
-                    __m128 _k46 = (__m128)__lsx_vld(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    __m128i _r4 = __lsx_vld(r4, 0);
-                    __m128i _r4n = __lsx_vld(r4 + 4, 0);
-                    __m128i _r4nn = __lsx_vld(r4 + 8, 0);
-
-                    __m128 _r40 = (__m128)__lsx_vreplvei_w(_r4, 0);
-                    __m128 _r41 = (__m128)__lsx_vreplvei_w(_r4, 1);
-                    __m128 _r42 = (__m128)__lsx_vreplvei_w(_r4, 2);
-                    __m128 _r43 = (__m128)__lsx_vreplvei_w(_r4, 3);
-                    __m128 _r44 = (__m128)__lsx_vreplvei_w(_r4n, 0);
-                    __m128 _r45 = (__m128)__lsx_vreplvei_w(_r4n, 1);
-                    __m128 _r46 = (__m128)__lsx_vreplvei_w(_r4n, 2);
-                    __m128 _r47 = (__m128)__lsx_vreplvei_w(_r4n, 3);
-                    __m128 _r48 = (__m128)__lsx_vreplvei_w(_r4nn, 0);
-                    __m128 _r49 = (__m128)__lsx_vreplvei_w(_r4nn, 1);
-                    __m128 _r4a = (__m128)__lsx_vreplvei_w(_r4nn, 2);
-                    __m128 _r4b = (__m128)__lsx_vreplvei_w(_r4nn, 3);
-                    __m128 _r4c = __lsx_vreplfr2vr_s(r4[12]);
-
-                    _sum0 = __lsx_vfmadd_s(_k40, _r40, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k40, _r42, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k40, _r44, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k40, _r46, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k41, _r41, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k41, _r43, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k41, _r45, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k41, _r47, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k42, _r42, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k42, _r44, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k42, _r46, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k42, _r48, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k43, _r43, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k43, _r45, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k43, _r47, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k43, _r49, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k44, _r44, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k44, _r46, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k44, _r48, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k44, _r4a, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k45, _r45, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k45, _r47, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k45, _r49, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k45, _r4b, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k46, _r46, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k46, _r48, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k46, _r4a, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k46, _r4c, _sum3);
-
-                    __m128 _k50 = (__m128)__lsx_vld(kptr, 0);
-                    __m128 _k51 = (__m128)__lsx_vld(kptr + 4, 0);
-                    __m128 _k52 = (__m128)__lsx_vld(kptr + 4 * 2, 0);
-                    __m128 _k53 = (__m128)__lsx_vld(kptr + 4 * 3, 0);
-                    __m128 _k54 = (__m128)__lsx_vld(kptr + 4 * 4, 0);
-                    __m128 _k55 = (__m128)__lsx_vld(kptr + 4 * 5, 0);
-                    __m128 _k56 = (__m128)__lsx_vld(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    __m128i _r5 = __lsx_vld(r5, 0);
-                    __m128i _r5n = __lsx_vld(r5 + 4, 0);
-                    __m128i _r5nn = __lsx_vld(r5 + 8, 0);
-
-                    __m128 _r50 = (__m128)__lsx_vreplvei_w(_r5, 0);
-                    __m128 _r51 = (__m128)__lsx_vreplvei_w(_r5, 1);
-                    __m128 _r52 = (__m128)__lsx_vreplvei_w(_r5, 2);
-                    __m128 _r53 = (__m128)__lsx_vreplvei_w(_r5, 3);
-                    __m128 _r54 = (__m128)__lsx_vreplvei_w(_r5n, 0);
-                    __m128 _r55 = (__m128)__lsx_vreplvei_w(_r5n, 1);
-                    __m128 _r56 = (__m128)__lsx_vreplvei_w(_r5n, 2);
-                    __m128 _r57 = (__m128)__lsx_vreplvei_w(_r5n, 3);
-                    __m128 _r58 = (__m128)__lsx_vreplvei_w(_r5nn, 0);
-                    __m128 _r59 = (__m128)__lsx_vreplvei_w(_r5nn, 1);
-                    __m128 _r5a = (__m128)__lsx_vreplvei_w(_r5nn, 2);
-                    __m128 _r5b = (__m128)__lsx_vreplvei_w(_r5nn, 3);
-                    __m128 _r5c = __lsx_vreplfr2vr_s(r5[12]);
-
-                    _sum0 = __lsx_vfmadd_s(_k50, _r50, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k50, _r52, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k50, _r54, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k50, _r56, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k51, _r51, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k51, _r53, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k51, _r55, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k51, _r57, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k52, _r52, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k52, _r54, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k52, _r56, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k52, _r58, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k53, _r53, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k53, _r55, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k53, _r57, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k53, _r59, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k54, _r54, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k54, _r56, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k54, _r58, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k54, _r5a, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k55, _r55, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k55, _r57, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k55, _r59, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k55, _r5b, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k56, _r56, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k56, _r58, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k56, _r5a, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k56, _r5c, _sum3);
-
-                    __m128 _k60 = (__m128)__lsx_vld(kptr, 0);
-                    __m128 _k61 = (__m128)__lsx_vld(kptr + 4, 0);
-                    __m128 _k62 = (__m128)__lsx_vld(kptr + 4 * 2, 0);
-                    __m128 _k63 = (__m128)__lsx_vld(kptr + 4 * 3, 0);
-                    __m128 _k64 = (__m128)__lsx_vld(kptr + 4 * 4, 0);
-                    __m128 _k65 = (__m128)__lsx_vld(kptr + 4 * 5, 0);
-                    __m128 _k66 = (__m128)__lsx_vld(kptr + 4 * 6, 0);
-
-                    kptr -= 4 * 42;
-
-                    __m128i _r6 = __lsx_vld(r6, 0);
-                    __m128i _r6n = __lsx_vld(r6 + 4, 0);
-                    __m128i _r6nn = __lsx_vld(r6 + 8, 0);
-
-                    __m128 _r60 = (__m128)__lsx_vreplvei_w(_r6, 0);
-                    __m128 _r61 = (__m128)__lsx_vreplvei_w(_r6, 1);
-                    __m128 _r62 = (__m128)__lsx_vreplvei_w(_r6, 2);
-                    __m128 _r63 = (__m128)__lsx_vreplvei_w(_r6, 3);
-                    __m128 _r64 = (__m128)__lsx_vreplvei_w(_r6n, 0);
-                    __m128 _r65 = (__m128)__lsx_vreplvei_w(_r6n, 1);
-                    __m128 _r66 = (__m128)__lsx_vreplvei_w(_r6n, 2);
-                    __m128 _r67 = (__m128)__lsx_vreplvei_w(_r6n, 3);
-                    __m128 _r68 = (__m128)__lsx_vreplvei_w(_r6nn, 0);
-                    __m128 _r69 = (__m128)__lsx_vreplvei_w(_r6nn, 1);
-                    __m128 _r6a = (__m128)__lsx_vreplvei_w(_r6nn, 2);
-                    __m128 _r6b = (__m128)__lsx_vreplvei_w(_r6nn, 3);
-                    __m128 _r6c = __lsx_vreplfr2vr_s(r6[12]);
-
-                    _sum0 = __lsx_vfmadd_s(_k60, _r60, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k60, _r62, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k60, _r64, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k60, _r66, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k61, _r61, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k61, _r63, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k61, _r65, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k61, _r67, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k62, _r62, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k62, _r64, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k62, _r66, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k62, _r68, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k63, _r63, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k63, _r65, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k63, _r67, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k63, _r69, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k64, _r64, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k64, _r66, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k64, _r68, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k64, _r6a, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k65, _r65, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k65, _r67, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k65, _r69, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k65, _r6b, _sum3);
-                    _sum0 = __lsx_vfmadd_s(_k66, _r66, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_k66, _r68, _sum1);
-                    _sum2 = __lsx_vfmadd_s(_k66, _r6a, _sum2);
-                    _sum3 = __lsx_vfmadd_s(_k66, _r6c, _sum3);
-
-                    __lsx_vst(_sum0, outptr0, 0);
-                    __lsx_vst(_sum1, outptr0 + 4, 0);
-                    __lsx_vst(_sum2, outptr0 + 4 * 2, 0);
-                    __lsx_vst(_sum3, outptr0 + 4 * 3, 0);
-
-                    outptr0 += 4 * 4;
-
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                    r3 += 8;
-                    r4 += 8;
-                    r5 += 8;
-                    r6 += 8;
-                }
-                for (; j < outw; j++)
-                {
-                    __m128 _sum0 = (__m128)__lsx_vld(outptr0, 0);
-
-                    __m128 _k00 = (__m128)__lsx_vld(kptr, 0);
-                    __m128 _k01 = (__m128)__lsx_vld(kptr + 4, 0);
-                    __m128 _k02 = (__m128)__lsx_vld(kptr + 4 * 2, 0);
-                    __m128 _k03 = (__m128)__lsx_vld(kptr + 4 * 3, 0);
-                    __m128 _k04 = (__m128)__lsx_vld(kptr + 4 * 4, 0);
-                    __m128 _k05 = (__m128)__lsx_vld(kptr + 4 * 5, 0);
-                    __m128 _k06 = (__m128)__lsx_vld(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    __m128i _r0 = __lsx_vld(r0, 0);
-                    __m128i _r0n = __lsx_vld(r0 + 4, 0);
-
-                    _sum0 = __lsx_vfmadd_s(_k00, (__m128)__lsx_vreplvei_w(_r0, 0), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k01, (__m128)__lsx_vreplvei_w(_r0, 1), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k02, (__m128)__lsx_vreplvei_w(_r0, 2), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k03, (__m128)__lsx_vreplvei_w(_r0, 3), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k04, (__m128)__lsx_vreplvei_w(_r0n, 0), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k05, (__m128)__lsx_vreplvei_w(_r0n, 1), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k06, (__m128)__lsx_vreplvei_w(_r0n, 2), _sum0);
-
-                    __m128 _k10 = (__m128)__lsx_vld(kptr, 0);
-                    __m128 _k11 = (__m128)__lsx_vld(kptr + 4, 0);
-                    __m128 _k12 = (__m128)__lsx_vld(kptr + 4 * 2, 0);
-                    __m128 _k13 = (__m128)__lsx_vld(kptr + 4 * 3, 0);
-                    __m128 _k14 = (__m128)__lsx_vld(kptr + 4 * 4, 0);
-                    __m128 _k15 = (__m128)__lsx_vld(kptr + 4 * 5, 0);
-                    __m128 _k16 = (__m128)__lsx_vld(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    __m128i _r1 = __lsx_vld(r1, 0);
-                    __m128i _r1n = __lsx_vld(r1 + 4, 0);
-
-                    _sum0 = __lsx_vfmadd_s(_k10, (__m128)__lsx_vreplvei_w(_r1, 0), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k11, (__m128)__lsx_vreplvei_w(_r1, 1), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k12, (__m128)__lsx_vreplvei_w(_r1, 2), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k13, (__m128)__lsx_vreplvei_w(_r1, 3), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k14, (__m128)__lsx_vreplvei_w(_r1n, 0), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k15, (__m128)__lsx_vreplvei_w(_r1n, 1), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k16, (__m128)__lsx_vreplvei_w(_r1n, 2), _sum0);
-
-                    __m128 _k20 = (__m128)__lsx_vld(kptr, 0);
-                    __m128 _k21 = (__m128)__lsx_vld(kptr + 4, 0);
-                    __m128 _k22 = (__m128)__lsx_vld(kptr + 4 * 2, 0);
-                    __m128 _k23 = (__m128)__lsx_vld(kptr + 4 * 3, 0);
-                    __m128 _k24 = (__m128)__lsx_vld(kptr + 4 * 4, 0);
-                    __m128 _k25 = (__m128)__lsx_vld(kptr + 4 * 5, 0);
-                    __m128 _k26 = (__m128)__lsx_vld(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    __m128i _r2 = __lsx_vld(r2, 0);
-                    __m128i _r2n = __lsx_vld(r2 + 4, 0);
-
-                    _sum0 = __lsx_vfmadd_s(_k20, (__m128)__lsx_vreplvei_w(_r2, 0), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k21, (__m128)__lsx_vreplvei_w(_r2, 1), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k22, (__m128)__lsx_vreplvei_w(_r2, 2), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k23, (__m128)__lsx_vreplvei_w(_r2, 3), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k24, (__m128)__lsx_vreplvei_w(_r2n, 0), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k25, (__m128)__lsx_vreplvei_w(_r2n, 1), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k26, (__m128)__lsx_vreplvei_w(_r2n, 2), _sum0);
-
-                    __m128 _k30 = (__m128)__lsx_vld(kptr, 0);
-                    __m128 _k31 = (__m128)__lsx_vld(kptr + 4, 0);
-                    __m128 _k32 = (__m128)__lsx_vld(kptr + 4 * 2, 0);
-                    __m128 _k33 = (__m128)__lsx_vld(kptr + 4 * 3, 0);
-                    __m128 _k34 = (__m128)__lsx_vld(kptr + 4 * 4, 0);
-                    __m128 _k35 = (__m128)__lsx_vld(kptr + 4 * 5, 0);
-                    __m128 _k36 = (__m128)__lsx_vld(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    __m128i _r3 = __lsx_vld(r3, 0);
-                    __m128i _r3n = __lsx_vld(r3 + 4, 0);
-
-                    _sum0 = __lsx_vfmadd_s(_k30, (__m128)__lsx_vreplvei_w(_r3, 0), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k31, (__m128)__lsx_vreplvei_w(_r3, 1), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k32, (__m128)__lsx_vreplvei_w(_r3, 2), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k33, (__m128)__lsx_vreplvei_w(_r3, 3), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k34, (__m128)__lsx_vreplvei_w(_r3n, 0), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k35, (__m128)__lsx_vreplvei_w(_r3n, 1), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k36, (__m128)__lsx_vreplvei_w(_r3n, 2), _sum0);
-
-                    __m128 _k40 = (__m128)__lsx_vld(kptr, 0);
-                    __m128 _k41 = (__m128)__lsx_vld(kptr + 4, 0);
-                    __m128 _k42 = (__m128)__lsx_vld(kptr + 4 * 2, 0);
-                    __m128 _k43 = (__m128)__lsx_vld(kptr + 4 * 3, 0);
-                    __m128 _k44 = (__m128)__lsx_vld(kptr + 4 * 4, 0);
-                    __m128 _k45 = (__m128)__lsx_vld(kptr + 4 * 5, 0);
-                    __m128 _k46 = (__m128)__lsx_vld(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    __m128i _r4 = __lsx_vld(r4, 0);
-                    __m128i _r4n = __lsx_vld(r4 + 4, 0);
-
-                    _sum0 = __lsx_vfmadd_s(_k40, (__m128)__lsx_vreplvei_w(_r4, 0), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k41, (__m128)__lsx_vreplvei_w(_r4, 1), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k42, (__m128)__lsx_vreplvei_w(_r4, 2), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k43, (__m128)__lsx_vreplvei_w(_r4, 3), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k44, (__m128)__lsx_vreplvei_w(_r4n, 0), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k45, (__m128)__lsx_vreplvei_w(_r4n, 1), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k46, (__m128)__lsx_vreplvei_w(_r4n, 2), _sum0);
-
-                    __m128 _k50 = (__m128)__lsx_vld(kptr, 0);
-                    __m128 _k51 = (__m128)__lsx_vld(kptr + 4, 0);
-                    __m128 _k52 = (__m128)__lsx_vld(kptr + 4 * 2, 0);
-                    __m128 _k53 = (__m128)__lsx_vld(kptr + 4 * 3, 0);
-                    __m128 _k54 = (__m128)__lsx_vld(kptr + 4 * 4, 0);
-                    __m128 _k55 = (__m128)__lsx_vld(kptr + 4 * 5, 0);
-                    __m128 _k56 = (__m128)__lsx_vld(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    __m128i _r5 = __lsx_vld(r5, 0);
-                    __m128i _r5n = __lsx_vld(r5 + 4, 0);
-
-                    _sum0 = __lsx_vfmadd_s(_k50, (__m128)__lsx_vreplvei_w(_r5, 0), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k51, (__m128)__lsx_vreplvei_w(_r5, 1), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k52, (__m128)__lsx_vreplvei_w(_r5, 2), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k53, (__m128)__lsx_vreplvei_w(_r5, 3), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k54, (__m128)__lsx_vreplvei_w(_r5n, 0), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k55, (__m128)__lsx_vreplvei_w(_r5n, 1), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k56, (__m128)__lsx_vreplvei_w(_r5n, 2), _sum0);
-
-                    __m128 _k60 = (__m128)__lsx_vld(kptr, 0);
-                    __m128 _k61 = (__m128)__lsx_vld(kptr + 4, 0);
-                    __m128 _k62 = (__m128)__lsx_vld(kptr + 4 * 2, 0);
-                    __m128 _k63 = (__m128)__lsx_vld(kptr + 4 * 3, 0);
-                    __m128 _k64 = (__m128)__lsx_vld(kptr + 4 * 4, 0);
-                    __m128 _k65 = (__m128)__lsx_vld(kptr + 4 * 5, 0);
-                    __m128 _k66 = (__m128)__lsx_vld(kptr + 4 * 6, 0);
-
-                    kptr -= 4 * 42;
-
-                    __m128i _r6 = __lsx_vld(r6, 0);
-                    __m128i _r6n = __lsx_vld(r6 + 4, 0);
-
-                    _sum0 = __lsx_vfmadd_s(_k60, (__m128)__lsx_vreplvei_w(_r6, 0), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k61, (__m128)__lsx_vreplvei_w(_r6, 1), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k62, (__m128)__lsx_vreplvei_w(_r6, 2), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k63, (__m128)__lsx_vreplvei_w(_r6, 3), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k64, (__m128)__lsx_vreplvei_w(_r6n, 0), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k65, (__m128)__lsx_vreplvei_w(_r6n, 1), _sum0);
-                    _sum0 = __lsx_vfmadd_s(_k66, (__m128)__lsx_vreplvei_w(_r6n, 2), _sum0);
-
-                    __lsx_vst(_sum0, outptr0, 0);
-
-                    outptr0 += 4;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    r3 += 2;
-                    r4 += 2;
-                    r5 += 2;
-                    r6 += 2;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-                r4 += tailstep;
-                r5 += tailstep;
-                r6 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_int8.h
deleted file mode 100644
index 22c7a8cc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_int8.h
+++ /dev/null
@@ -1,82 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_int8(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_int8, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        int* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                int sum = 0;
-
-                //                 const signed char* kptr = weight_data_int8.channel(p);
-                const signed char* kptr = (const signed char*)weight_data_int8 + maxk * channels * p;
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const signed char* sptr = m.row<signed char>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        signed char val = sptr[space_ofs[k]];
-                        signed char w = kptr[k];
-                        sum += val * w;
-                    }
-
-                    kptr += maxk;
-                }
-
-                outptr[j] = sum;
-            }
-
-            outptr += outw;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_loongarch.h
deleted file mode 100644
index 7807f43f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_loongarch.h
+++ /dev/null
@@ -1,56 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTION_LOONGARCH_H
-#define LAYER_CONVOLUTION_LOONGARCH_H
-
-#include "convolution.h"
-
-namespace ncnn {
-
-class Convolution_loongarch : public Convolution
-{
-public:
-    Convolution_loongarch();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if NCNN_INT8
-    int create_pipeline_int8_loongarch(const Option& opt);
-    int forward_int8_loongarch(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Layer* activation;
-
-    Mat weight_data_tm;
-    Mat weight_sgemm_data;
-    Mat weight_winograd23_data;
-    Mat weight_winograd43_data;
-    Mat weight_winograd63_data;
-
-#if NCNN_INT8
-    Mat scale_in_data;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTION_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack1to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack1to4.h
deleted file mode 100644
index b7e0123d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack1to4.h
+++ /dev/null
@@ -1,90 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_pack1to4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_pack1ton, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m128 _sum = (__m128)__lsx_vreplgr2vr_w(0);
-
-                if (bias_data_ptr)
-                {
-                    _sum = (__m128)__lsx_vld(bias_data_ptr + p * 4, 0);
-                }
-
-                const float* kptr = (const float*)weight_data_pack1ton + maxk * channels * p * 4;
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const float* sptr = m.row(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++) // 29.23
-                    {
-                        __m128 _val = __lsx_vreplfr2vr_s(sptr[space_ofs[k]]);
-                        __m128 _w = (__m128)__lsx_vld(kptr, 0);
-                        _sum = __lsx_vfmadd_s(_w, _val, _sum);
-
-                        kptr += 4;
-                    }
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params);
-
-                __lsx_vst(_sum, outptr + j * 4, 0);
-            }
-
-            outptr += outw * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack1to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack1to4_int8.h
deleted file mode 100644
index b043503c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack1to4_int8.h
+++ /dev/null
@@ -1,87 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_pack1to4_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_int8, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        int* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m128i _sum = __lsx_vreplgr2vr_w(0);
-
-                const signed char* kptr = weight_data_int8.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const signed char* sptr = m.row<const signed char>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        __m128i _val = __lsx_vreplgr2vr_h((short)sptr[space_ofs[k]]);
-
-                        __m128i _w = __lsx_vld(kptr, 0);
-                        __m128i _w16 = __lsx_vilvl_b(__lsx_vslti_b(_w, 0), _w);
-
-                        __m128i _s0 = __lsx_vmul_h(_val, _w16);
-                        __m128i _s032 = __lsx_vilvl_h(__lsx_vslti_h(_s0, 0), _s0);
-
-                        _sum = __lsx_vadd_w(_sum, _s032);
-
-                        kptr += 4;
-                    }
-                }
-
-                __lsx_vst(_sum, outptr + j * 4, 0);
-            }
-
-            outptr += outw * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack4.h
deleted file mode 100644
index 66a7863f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack4.h
+++ /dev/null
@@ -1,102 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_pack4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_pack4, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m128 _sum = (__m128)__lsx_vreplgr2vr_w(0);
-
-                if (bias_data_ptr)
-                {
-                    _sum = (__m128)__lsx_vld(bias_data_ptr + p * 4, 0);
-                }
-
-                const float* kptr = (const float*)weight_data_pack4.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const float* sptr = m.row(i * stride_h) + j * stride_w * 4;
-
-                    for (int k = 0; k < maxk; k++) // 29.23
-                    {
-                        const float* slptr = sptr + space_ofs[k] * 4;
-
-                        __m128 _val0 = __lsx_vreplfr2vr_s(slptr[0]);
-                        __m128 _val1 = __lsx_vreplfr2vr_s(slptr[1]);
-                        __m128 _val2 = __lsx_vreplfr2vr_s(slptr[2]);
-                        __m128 _val3 = __lsx_vreplfr2vr_s(slptr[3]);
-
-                        __m128 _w0 = (__m128)__lsx_vld(kptr, 0);
-                        __m128 _w1 = (__m128)__lsx_vld(kptr + 4, 0);
-                        __m128 _w2 = (__m128)__lsx_vld(kptr + 8, 0);
-                        __m128 _w3 = (__m128)__lsx_vld(kptr + 12, 0);
-
-                        _sum = __lsx_vfmadd_s(_w0, _val0, _sum);
-                        _sum = __lsx_vfmadd_s(_w1, _val1, _sum);
-                        _sum = __lsx_vfmadd_s(_w2, _val2, _sum);
-                        _sum = __lsx_vfmadd_s(_w3, _val3, _sum);
-
-                        kptr += 16;
-                    }
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params);
-
-                __lsx_vst(_sum, outptr + j * 4, 0);
-            }
-
-            outptr += outw * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack4to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack4to1.h
deleted file mode 100644
index 872759fc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack4to1.h
+++ /dev/null
@@ -1,94 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_pack4to1_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_pack4to1, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                __m128 _sum = (__m128)__lsx_vreplgr2vr_w(0);
-
-                const float* kptr = (const float*)weight_data_pack4to1.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const float* sptr = m.row(i * stride_h) + j * stride_w * 4;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        __m128 _val = (__m128)__lsx_vld(sptr + space_ofs[k] * 4, 0);
-                        __m128 _w = (__m128)__lsx_vld(kptr, 0);
-                        _sum = __lsx_vfmadd_s(_w, _val, _sum);
-
-                        kptr += 4;
-                    }
-                }
-
-                sum += __lsx_reduce_fadd_s(_sum);
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[j] = sum;
-            }
-
-            outptr += outw;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack8to1_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack8to1_int8.h
deleted file mode 100644
index c7463a47..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack8to1_int8.h
+++ /dev/null
@@ -1,87 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_pack8to1_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_int8, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        int* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m128i _sum = __lsx_vreplgr2vr_w(0);
-
-                const signed char* kptr = weight_data_int8.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const signed char* sptr = m.row<const signed char>(i * stride_h) + j * stride_w * 8;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        __m128i _val = __lsx_vld(sptr + space_ofs[k] * 8, 0);
-                        __m128i _val16 = __lsx_vilvl_b(__lsx_vslti_b(_val, 0), _val);
-
-                        __m128i _w = __lsx_vld(kptr, 0);
-                        __m128i _w16 = __lsx_vilvl_b(__lsx_vslti_b(_w, 0), _w);
-
-                        __m128i _s0 = __lsx_vmul_h(_val16, _w16);
-
-                        _sum = __lsx_vadd_w(_sum, __lsx_vhaddw_w_h(_s0, _s0));
-
-                        kptr += 8;
-                    }
-                }
-
-                outptr[j] = __lsx_reduce_add_w(_sum);
-            }
-
-            outptr += outw;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack8to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack8to4_int8.h
deleted file mode 100644
index 00d90387..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_pack8to4_int8.h
+++ /dev/null
@@ -1,120 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_pack8to4_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_int8, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        int* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m128i _sum0 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum1 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum2 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum3 = __lsx_vreplgr2vr_w(0);
-
-                const signed char* kptr = weight_data_int8.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const signed char* sptr = m.row<signed char>(i * stride_h) + j * stride_w * 8;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        __m128i _val = __lsx_vld(sptr + space_ofs[k] * 8, 0);
-                        __m128i _val16 = __lsx_vilvl_b(__lsx_vslti_b(_val, 0), _val);
-
-                        __m128i _w01 = __lsx_vld(kptr, 0);
-                        __m128i _w23 = __lsx_vld(kptr + 16, 0);
-                        __m128i _extw01 = __lsx_vslti_b(_w01, 0);
-                        __m128i _extw23 = __lsx_vslti_b(_w23, 0);
-                        __m128i _w0 = __lsx_vilvl_b(_extw01, _w01);
-                        __m128i _w1 = __lsx_vilvh_b(_extw01, _w01);
-                        __m128i _w2 = __lsx_vilvl_b(_extw23, _w23);
-                        __m128i _w3 = __lsx_vilvh_b(_extw23, _w23);
-
-                        __m128i _s0 = __lsx_vmul_h(_val16, _w0);
-                        __m128i _s1 = __lsx_vmul_h(_val16, _w1);
-                        __m128i _s2 = __lsx_vmul_h(_val16, _w2);
-                        __m128i _s3 = __lsx_vmul_h(_val16, _w3);
-
-                        _sum0 = __lsx_vadd_w(_sum0, __lsx_vhaddw_w_h(_s0, _s0));
-                        _sum1 = __lsx_vadd_w(_sum1, __lsx_vhaddw_w_h(_s1, _s1));
-                        _sum2 = __lsx_vadd_w(_sum2, __lsx_vhaddw_w_h(_s2, _s2));
-                        _sum3 = __lsx_vadd_w(_sum3, __lsx_vhaddw_w_h(_s3, _s3));
-
-                        kptr += 32;
-                    }
-                }
-
-                // transpose 4x4
-                {
-                    __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                    _tmp0 = __lsx_vilvl_w(_sum1, _sum0);
-                    _tmp1 = __lsx_vilvl_w(_sum3, _sum2);
-                    _tmp2 = __lsx_vilvh_w(_sum1, _sum0);
-                    _tmp3 = __lsx_vilvh_w(_sum3, _sum2);
-                    _sum0 = __lsx_vilvl_d(_tmp1, _tmp0);
-                    _sum1 = __lsx_vilvh_d(_tmp1, _tmp0);
-                    _sum2 = __lsx_vilvl_d(_tmp3, _tmp2);
-                    _sum3 = __lsx_vilvh_d(_tmp3, _tmp2);
-                }
-
-                _sum0 = __lsx_vadd_w(_sum0, _sum1);
-                _sum2 = __lsx_vadd_w(_sum2, _sum3);
-
-                _sum0 = __lsx_vadd_w(_sum0, _sum2);
-
-                __lsx_vst(_sum0, outptr + j * 4, 0);
-            }
-
-            outptr += outw * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm.h
deleted file mode 100644
index 7b74ceac..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm.h
+++ /dev/null
@@ -1,650 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_lsx(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    // Mat bottom_im2col(size, maxk, inch, 4u, 1, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    // permute
-    Mat tmp;
-    if (size >= 4)
-        tmp.create(4 * maxk, inch, size / 4 + size % 4, 4u, 1, opt.workspace_allocator);
-    else
-        tmp.create(maxk, inch, size, 4u, 1, opt.workspace_allocator);
-    {
-        int nn_size = size / 4;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = ii * 4;
-
-            float* tmpptr = tmp.channel(i / 4);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-#if __loongarch_sx
-                    __lsx_vst(__lsx_vld(img0, 0), tmpptr, 0);
-#else
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img0[1];
-                    tmpptr[2] = img0[2];
-                    tmpptr[3] = img0[3];
-#endif
-                    img0 += size;
-                    tmpptr += 4;
-                }
-            }
-        }
-
-        int remain_size_start = nn_size * 4;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            float* tmpptr = tmp.channel(i / 4 + i % 4);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    img0 += size;
-                    tmpptr += 1;
-                }
-            }
-        }
-    }
-
-#if __loongarch_sx
-    int nn_outch = outch >> 3;
-    int remain_outch_start = nn_outch << 3;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 8;
-
-        float* outptr0 = top_blob.channel(p);
-        float* outptr1 = top_blob.channel(p + 1);
-        float* outptr2 = top_blob.channel(p + 2);
-        float* outptr3 = top_blob.channel(p + 3);
-        float* outptr4 = top_blob.channel(p + 4);
-        float* outptr5 = top_blob.channel(p + 5);
-        float* outptr6 = top_blob.channel(p + 6);
-        float* outptr7 = top_blob.channel(p + 7);
-
-        const float zeros[8] = {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
-        const float* biasptr = bias ? bias + p : zeros;
-
-        int i = 0;
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 4);
-            const float* kptr = kernel.channel(p / 8);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __m128 _sum0 = __lsx_vreplfr2vr_s(biasptr[0]);
-            __m128 _sum1 = __lsx_vreplfr2vr_s(biasptr[1]);
-            __m128 _sum2 = __lsx_vreplfr2vr_s(biasptr[2]);
-            __m128 _sum3 = __lsx_vreplfr2vr_s(biasptr[3]);
-            __m128 _sum4 = __lsx_vreplfr2vr_s(biasptr[4]);
-            __m128 _sum5 = __lsx_vreplfr2vr_s(biasptr[5]);
-            __m128 _sum6 = __lsx_vreplfr2vr_s(biasptr[6]);
-            __m128 _sum7 = __lsx_vreplfr2vr_s(biasptr[7]);
-
-            for (int q = 0; q < nn; q++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr + 32);
-                __m128 _val = (__m128)__lsx_vld(tmpptr, 0);
-                __m128i _w0123 = __lsx_vld(kptr, 0);
-                __m128i _w4567 = __lsx_vld(kptr + 4, 0);
-                _sum0 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 0), _val, _sum0);
-                _sum1 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 1), _val, _sum1);
-                _sum2 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 2), _val, _sum2);
-                _sum3 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 3), _val, _sum3);
-                _sum4 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w4567, 0), _val, _sum4);
-                _sum5 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w4567, 1), _val, _sum5);
-                _sum6 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w4567, 2), _val, _sum6);
-                _sum7 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w4567, 3), _val, _sum7);
-
-                tmpptr += 4;
-                kptr += 8;
-            }
-
-            __lsx_vst(_sum0, outptr0, 0);
-            __lsx_vst(_sum1, outptr1, 0);
-            __lsx_vst(_sum2, outptr2, 0);
-            __lsx_vst(_sum3, outptr3, 0);
-            __lsx_vst(_sum4, outptr4, 0);
-            __lsx_vst(_sum5, outptr5, 0);
-            __lsx_vst(_sum6, outptr6, 0);
-            __lsx_vst(_sum7, outptr7, 0);
-
-            outptr0 += 4;
-            outptr1 += 4;
-            outptr2 += 4;
-            outptr3 += 4;
-            outptr4 += 4;
-            outptr5 += 4;
-            outptr6 += 4;
-            outptr7 += 4;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 4 + i % 4);
-            const float* kptr = kernel.channel(p / 8);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = biasptr[0];
-            float sum1 = biasptr[1];
-            float sum2 = biasptr[2];
-            float sum3 = biasptr[3];
-            float sum4 = biasptr[4];
-            float sum5 = biasptr[5];
-            float sum6 = biasptr[6];
-            float sum7 = biasptr[7];
-
-            for (int q = 0; q < nn; q++)
-            {
-                sum0 += tmpptr[0] * kptr[0];
-                sum1 += tmpptr[0] * kptr[1];
-                sum2 += tmpptr[0] * kptr[2];
-                sum3 += tmpptr[0] * kptr[3];
-                sum4 += tmpptr[0] * kptr[4];
-                sum5 += tmpptr[0] * kptr[5];
-                sum6 += tmpptr[0] * kptr[6];
-                sum7 += tmpptr[0] * kptr[7];
-                tmpptr++;
-                kptr += 8;
-            }
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-            outptr2[0] = sum2;
-            outptr3[0] = sum3;
-            outptr4[0] = sum4;
-            outptr5[0] = sum5;
-            outptr6[0] = sum6;
-            outptr7[0] = sum7;
-
-            outptr0++;
-            outptr1++;
-            outptr2++;
-            outptr3++;
-            outptr4++;
-            outptr5++;
-            outptr6++;
-            outptr7++;
-        }
-    }
-
-    nn_outch = (outch - remain_outch_start) >> 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = remain_outch_start + pp * 4;
-
-        float* outptr0 = top_blob.channel(p);
-        float* outptr1 = top_blob.channel(p + 1);
-        float* outptr2 = top_blob.channel(p + 2);
-        float* outptr3 = top_blob.channel(p + 3);
-
-        const float zeros[4] = {0.f, 0.f, 0.f, 0.f};
-        const float* biasptr = bias ? bias + p : zeros;
-
-        int i = 0;
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 4);
-            const float* kptr = kernel.channel(p / 8 + (p % 8) / 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __m128 _sum0 = __lsx_vreplfr2vr_s(biasptr[0]);
-            __m128 _sum1 = __lsx_vreplfr2vr_s(biasptr[1]);
-            __m128 _sum2 = __lsx_vreplfr2vr_s(biasptr[2]);
-            __m128 _sum3 = __lsx_vreplfr2vr_s(biasptr[3]);
-
-            for (int q = 0; q < nn; q++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr + 16);
-                __m128 _val = (__m128)__lsx_vld(tmpptr, 0);
-                __m128i _w0123 = __lsx_vld(kptr, 0);
-                _sum0 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 0), _val, _sum0);
-                _sum1 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 1), _val, _sum1);
-                _sum2 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 2), _val, _sum2);
-                _sum3 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 3), _val, _sum3);
-
-                tmpptr += 4;
-                kptr += 4;
-            }
-
-            __lsx_vst(_sum0, outptr0, 0);
-            __lsx_vst(_sum1, outptr1, 0);
-            __lsx_vst(_sum2, outptr2, 0);
-            __lsx_vst(_sum3, outptr3, 0);
-
-            outptr0 += 4;
-            outptr1 += 4;
-            outptr2 += 4;
-            outptr3 += 4;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 4 + i % 4);
-            const float* kptr = kernel.channel(p / 8 + (p % 8) / 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = biasptr[0];
-            float sum1 = biasptr[1];
-            float sum2 = biasptr[2];
-            float sum3 = biasptr[3];
-
-            for (int q = 0; q < nn; q++)
-            {
-                sum0 += tmpptr[0] * kptr[0];
-                sum1 += tmpptr[0] * kptr[1];
-                sum2 += tmpptr[0] * kptr[2];
-                sum3 += tmpptr[0] * kptr[3];
-                tmpptr++;
-                kptr += 4;
-            }
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-            outptr2[0] = sum2;
-            outptr3[0] = sum3;
-
-            outptr0++;
-            outptr1++;
-            outptr2++;
-            outptr3++;
-        }
-    }
-
-    remain_outch_start += nn_outch << 2;
-#else // __loongarch_sx
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        float* outptr0 = top_blob.channel(p);
-        float* outptr1 = top_blob.channel(p + 1);
-
-        const float zeros[2] = {0.f, 0.f};
-        const float* biasptr = bias ? bias + p : zeros;
-
-        int i = 0;
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 4);
-            const float* kptr = kernel.channel(p / 2);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum00 = biasptr[0];
-            float sum01 = biasptr[0];
-            float sum02 = biasptr[0];
-            float sum03 = biasptr[0];
-            float sum10 = biasptr[1];
-            float sum11 = biasptr[1];
-            float sum12 = biasptr[1];
-            float sum13 = biasptr[1];
-
-            for (int q = 0; q < nn; q++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr + 8);
-                float k0 = kptr[0];
-                float k1 = kptr[1];
-                sum00 += tmpptr[0] * k0;
-                sum01 += tmpptr[1] * k0;
-                sum02 += tmpptr[2] * k0;
-                sum03 += tmpptr[3] * k0;
-                sum10 += tmpptr[0] * k1;
-                sum11 += tmpptr[1] * k1;
-                sum12 += tmpptr[2] * k1;
-                sum13 += tmpptr[3] * k1;
-                tmpptr += 4;
-                kptr += 2;
-            }
-
-            outptr0[0] = sum00;
-            outptr0[1] = sum01;
-            outptr0[2] = sum02;
-            outptr0[3] = sum03;
-            outptr1[0] = sum10;
-            outptr1[1] = sum11;
-            outptr1[2] = sum12;
-            outptr1[3] = sum13;
-
-            outptr0 += 4;
-            outptr1 += 4;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 4 + i % 4);
-            const float* kptr = kernel.channel(p / 2);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = biasptr[0];
-            float sum1 = biasptr[1];
-
-            for (int q = 0; q < nn; q++)
-            {
-                __builtin_prefetch(tmpptr + 4);
-                __builtin_prefetch(kptr + 8);
-                sum0 += tmpptr[0] * kptr[0];
-                sum1 += tmpptr[0] * kptr[1];
-                tmpptr++;
-                kptr += 2;
-            }
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-
-            outptr0++;
-            outptr1++;
-        }
-    }
-#endif // __loongarch_sx
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        float* outptr0 = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        int i = 0;
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 4);
-#if __loongarch_sx
-            const float* kptr = kernel.channel(p / 8 + (p % 8) / 4 + p % 4);
-#else
-            const float* kptr = kernel.channel(p / 2 + p % 2);
-#endif
-
-            int nn = inch * maxk; // inch always > 0
-
-#if __loongarch_sx
-            __m128 _sum0 = __lsx_vreplfr2vr_s(bias0);
-
-            for (int q = 0; q < nn; q++)
-            {
-                _sum0 = __lsx_vfmadd_s((__m128)__lsx_vld(tmpptr, 0), __lsx_vreplfr2vr_s(kptr[0]), _sum0);
-                tmpptr += 4;
-                kptr++;
-            }
-
-            __lsx_vst(_sum0, outptr0, 0);
-
-            outptr0 += 4;
-#else
-            float sum0 = bias0;
-            float sum1 = bias0;
-            float sum2 = bias0;
-            float sum3 = bias0;
-
-            for (int q = 0; q < nn; q++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr + 4);
-                sum0 += tmpptr[0] * kptr[0];
-                sum1 += tmpptr[1] * kptr[0];
-                sum2 += tmpptr[2] * kptr[0];
-                sum3 += tmpptr[3] * kptr[0];
-                tmpptr += 4;
-                kptr++;
-            }
-
-            outptr0[0] = sum0;
-            outptr0[1] = sum1;
-            outptr0[2] = sum2;
-            outptr0[3] = sum3;
-
-            outptr0 += 4;
-#endif // __loongarch_sx
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 4 + i % 4);
-#if __loongarch_sx
-            const float* kptr = kernel.channel(p / 8 + (p % 8) / 4 + p % 4);
-#else
-            const float* kptr = kernel.channel(p / 2 + p % 2);
-#endif
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = bias0;
-
-            for (int q = 0; q < nn; q++)
-            {
-                sum0 += tmpptr[0] * kptr[0];
-                tmpptr++;
-                kptr++;
-            }
-
-            outptr0[0] = sum0;
-
-            outptr0++;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_transform_kernel_lsx(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = 8b-maxk-inch-outch/8b
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-#if __loongarch_sx
-    kernel_tm.create(8 * maxk, inch, outch / 8 + (outch % 8) / 4 + outch % 4);
-#else
-    kernel_tm.create(2 * maxk, inch, outch / 2 + outch % 2);
-#endif
-
-    int q = 0;
-#if __loongarch_sx
-    for (; q + 7 < outch; q += 8)
-    {
-        const Mat k0 = kernel.channel(q);
-        const Mat k1 = kernel.channel(q + 1);
-        const Mat k2 = kernel.channel(q + 2);
-        const Mat k3 = kernel.channel(q + 3);
-        const Mat k4 = kernel.channel(q + 4);
-        const Mat k5 = kernel.channel(q + 5);
-        const Mat k6 = kernel.channel(q + 6);
-        const Mat k7 = kernel.channel(q + 7);
-
-        float* g00 = kernel_tm.channel(q / 8);
-
-        for (int p = 0; p < inch; p++)
-        {
-            const float* k00 = k0.row(p);
-            const float* k10 = k1.row(p);
-            const float* k20 = k2.row(p);
-            const float* k30 = k3.row(p);
-            const float* k40 = k4.row(p);
-            const float* k50 = k5.row(p);
-            const float* k60 = k6.row(p);
-            const float* k70 = k7.row(p);
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k00[k];
-                g00[1] = k10[k];
-                g00[2] = k20[k];
-                g00[3] = k30[k];
-                g00[4] = k40[k];
-                g00[5] = k50[k];
-                g00[6] = k60[k];
-                g00[7] = k70[k];
-
-                g00 += 8;
-            }
-        }
-    }
-    for (; q + 3 < outch; q += 4)
-    {
-        const Mat k0 = kernel.channel(q);
-        const Mat k1 = kernel.channel(q + 1);
-        const Mat k2 = kernel.channel(q + 2);
-        const Mat k3 = kernel.channel(q + 3);
-
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4);
-
-        for (int p = 0; p < inch; p++)
-        {
-            const float* k00 = k0.row(p);
-            const float* k10 = k1.row(p);
-            const float* k20 = k2.row(p);
-            const float* k30 = k3.row(p);
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k00[k];
-                g00[1] = k10[k];
-                g00[2] = k20[k];
-                g00[3] = k30[k];
-
-                g00 += 4;
-            }
-        }
-    }
-#else
-    for (; q + 1 < outch; q += 2)
-    {
-        const Mat k0 = kernel.channel(q);
-        const Mat k1 = kernel.channel(q + 1);
-
-        float* g00 = kernel_tm.channel(q / 2);
-
-        for (int p = 0; p < inch; p++)
-        {
-            const float* k00 = k0.row(p);
-            const float* k10 = k1.row(p);
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k00[k];
-                g00[1] = k10[k];
-
-                g00 += 2;
-            }
-        }
-    }
-#endif // __loongarch_sx
-    for (; q < outch; q++)
-    {
-        const Mat k0 = kernel.channel(q);
-
-#if __loongarch_sx
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + q % 4);
-#else
-        float* g00 = kernel_tm.channel(q / 2 + q % 2);
-#endif
-
-        for (int p = 0; p < inch; p++)
-        {
-            const float* k00 = k0.row(p);
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k00[k];
-
-                g00 += 1;
-            }
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 4u, 1, opt.workspace_allocator);
-    {
-        const int gap = w * stride_h - outw * stride_w;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            float* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const float* sptr = img.row<const float>(dilation_h * u) + dilation_w * v;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            ptr[0] = sptr[0];
-
-                            sptr += stride_w;
-                            ptr += 1;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_lsx(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_int8.h
deleted file mode 100644
index 98f47760..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_int8.h
+++ /dev/null
@@ -1,800 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_int8_lsx(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    // Mat bottom_im2col(size, maxk, inch, 8u, 8, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    // permute
-    Mat tmp;
-#if __loongarch_sx
-    if (inch >= 4)
-    {
-        if (size >= 2)
-            tmp.create(2 * maxk, inch / 4 + inch % 4, size / 2 + size % 2, 4u, 4, opt.workspace_allocator);
-        else
-            tmp.create(maxk, inch / 4 + inch % 4, size, 4u, 4, opt.workspace_allocator);
-    }
-    else
-#endif // __loongarch_sx
-    {
-        if (size >= 2)
-            tmp.create(2 * maxk, inch, size / 2 + size % 2, 1u, 1, opt.workspace_allocator);
-        else
-            tmp.create(maxk, inch, size, 1u, 1, opt.workspace_allocator);
-    }
-    {
-        int remain_size_start = 0;
-        int nn_size = (size - remain_size_start) >> 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 2;
-
-            signed char* tmpptr = tmp.channel(i / 2);
-
-            int q = 0;
-#if __loongarch_sx
-            for (; q + 3 < inch; q += 4)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-                const signed char* img1 = (const signed char*)bottom_im2col.channel(q + 1) + i;
-                const signed char* img2 = (const signed char*)bottom_im2col.channel(q + 2) + i;
-                const signed char* img3 = (const signed char*)bottom_im2col.channel(q + 3) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img1[0];
-                    tmpptr[2] = img2[0];
-                    tmpptr[3] = img3[0];
-                    tmpptr[4] = img0[1];
-                    tmpptr[5] = img1[1];
-                    tmpptr[6] = img2[1];
-                    tmpptr[7] = img3[1];
-                    tmpptr += 8;
-
-                    img0 += size;
-                    img1 += size;
-                    img2 += size;
-                    img3 += size;
-                }
-            }
-#endif // __loongarch_sx
-            for (; q < inch; q++)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img0[1];
-
-                    tmpptr += 2;
-
-                    img0 += size;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-
-            int q = 0;
-#if __loongarch_sx
-            for (; q + 3 < inch; q += 4)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-                const signed char* img1 = (const signed char*)bottom_im2col.channel(q + 1) + i;
-                const signed char* img2 = (const signed char*)bottom_im2col.channel(q + 2) + i;
-                const signed char* img3 = (const signed char*)bottom_im2col.channel(q + 3) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img1[0];
-                    tmpptr[2] = img2[0];
-                    tmpptr[3] = img3[0];
-                    tmpptr += 4;
-
-                    img0 += size;
-                    img1 += size;
-                    img2 += size;
-                    img3 += size;
-                }
-            }
-#endif // __loongarch_sx
-            for (; q < inch; q++)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-
-                    tmpptr += 1;
-
-                    img0 += size;
-                }
-            }
-        }
-    }
-
-#if __loongarch_sx
-    int nn_outch = outch >> 2;
-    int remain_outch_start = nn_outch << 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 4;
-
-        int* outptr0 = top_blob.channel(p);
-        int* outptr1 = top_blob.channel(p + 1);
-        int* outptr2 = top_blob.channel(p + 2);
-        int* outptr3 = top_blob.channel(p + 3);
-
-        int i = 0;
-        for (; i + 1 < size; i += 2)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2);
-            const signed char* kptr = kernel.channel(p / 4);
-
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            __m128i _sum00 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum10 = __lsx_vreplgr2vr_w(0);
-
-            if (nn4 > 0)
-            {
-                __m128i _sum01 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum02 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum03 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum11 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum12 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum13 = __lsx_vreplgr2vr_w(0);
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    __m128i _val = __lsx_vld(tmpptr, 0);
-                    __m128i _val01 = __lsx_vilvl_b(__lsx_vslti_b(_val, 0), _val);
-
-                    __m128i _val0 = __lsx_vilvl_d(_val01, _val01);
-                    __m128i _val1 = __lsx_vilvh_d(_val01, _val01);
-
-                    __m128i _w01 = __lsx_vld(kptr, 0);
-                    __m128i _extw01 = __lsx_vslti_b(_w01, 0);
-                    __m128i _w0 = __lsx_vilvl_b(_extw01, _w01);
-                    __m128i _w1 = __lsx_vilvh_b(_extw01, _w01);
-
-                    __m128i _s00 = __lsx_vmul_h(_val0, _w0);
-                    __m128i _s01 = __lsx_vmul_h(_val0, _w1);
-                    __m128i _s10 = __lsx_vmul_h(_val1, _w0);
-                    __m128i _s11 = __lsx_vmul_h(_val1, _w1);
-
-                    __m128i _exts00 = __lsx_vslti_h(_s00, 0);
-                    __m128i _exts01 = __lsx_vslti_h(_s01, 0);
-                    __m128i _exts10 = __lsx_vslti_h(_s10, 0);
-                    __m128i _exts11 = __lsx_vslti_h(_s11, 0);
-                    __m128i _s00l = __lsx_vilvl_h(_exts00, _s00);
-                    __m128i _s00h = __lsx_vilvh_h(_exts00, _s00);
-                    __m128i _s01l = __lsx_vilvl_h(_exts01, _s01);
-                    __m128i _s01h = __lsx_vilvh_h(_exts01, _s01);
-                    __m128i _s10l = __lsx_vilvl_h(_exts10, _s10);
-                    __m128i _s10h = __lsx_vilvh_h(_exts10, _s10);
-                    __m128i _s11l = __lsx_vilvl_h(_exts11, _s11);
-                    __m128i _s11h = __lsx_vilvh_h(_exts11, _s11);
-
-                    _sum00 = __lsx_vadd_w(_sum00, _s00l);
-                    _sum01 = __lsx_vadd_w(_sum01, _s00h);
-                    _sum02 = __lsx_vadd_w(_sum02, _s01l);
-                    _sum03 = __lsx_vadd_w(_sum03, _s01h);
-                    _sum10 = __lsx_vadd_w(_sum10, _s10l);
-                    _sum11 = __lsx_vadd_w(_sum11, _s10h);
-                    _sum12 = __lsx_vadd_w(_sum12, _s11l);
-                    _sum13 = __lsx_vadd_w(_sum13, _s11h);
-
-                    tmpptr += 8;
-                    kptr += 16;
-                }
-
-                // transpose 4x4
-                {
-                    __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                    _tmp0 = __lsx_vilvl_w(_sum01, _sum00);
-                    _tmp1 = __lsx_vilvl_w(_sum03, _sum02);
-                    _tmp2 = __lsx_vilvh_w(_sum01, _sum00);
-                    _tmp3 = __lsx_vilvh_w(_sum03, _sum02);
-                    _sum00 = __lsx_vilvl_d(_tmp1, _tmp0);
-                    _sum01 = __lsx_vilvh_d(_tmp1, _tmp0);
-                    _sum02 = __lsx_vilvl_d(_tmp3, _tmp2);
-                    _sum03 = __lsx_vilvh_d(_tmp3, _tmp2);
-                }
-                {
-                    __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                    _tmp0 = __lsx_vilvl_w(_sum11, _sum10);
-                    _tmp1 = __lsx_vilvl_w(_sum13, _sum12);
-                    _tmp2 = __lsx_vilvh_w(_sum11, _sum10);
-                    _tmp3 = __lsx_vilvh_w(_sum13, _sum12);
-                    _sum10 = __lsx_vilvl_d(_tmp1, _tmp0);
-                    _sum11 = __lsx_vilvh_d(_tmp1, _tmp0);
-                    _sum12 = __lsx_vilvl_d(_tmp3, _tmp2);
-                    _sum13 = __lsx_vilvh_d(_tmp3, _tmp2);
-                }
-
-                _sum00 = __lsx_vadd_w(_sum00, _sum01);
-                _sum02 = __lsx_vadd_w(_sum02, _sum03);
-                _sum10 = __lsx_vadd_w(_sum10, _sum11);
-                _sum12 = __lsx_vadd_w(_sum12, _sum13);
-
-                _sum00 = __lsx_vadd_w(_sum00, _sum02);
-                _sum10 = __lsx_vadd_w(_sum10, _sum12);
-            }
-
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                __m128i _val0 = __lsx_vreplgr2vr_h(tmpptr[0]);
-                __m128i _val1 = __lsx_vreplgr2vr_h(tmpptr[1]);
-                __m128i _val = __lsx_vilvl_d(_val1, _val0);
-
-                __m128i _w = __lsx_vld(kptr, 0);
-                __m128i _w16 = __lsx_vilvl_b(__lsx_vslti_b(_w, 0), _w);
-
-                _w16 = __lsx_vilvl_d(_w16, _w16);
-
-                __m128i _s0 = __lsx_vmul_h(_val, _w16);
-                __m128i _exts0 = __lsx_vslti_h(_s0, 0);
-                __m128i _s0l = __lsx_vilvl_h(_exts0, _s0);
-                __m128i _s0h = __lsx_vilvh_h(_exts0, _s0);
-
-                _sum00 = __lsx_vadd_w(_sum00, _s0l);
-                _sum10 = __lsx_vadd_w(_sum10, _s0h);
-
-                tmpptr += 2;
-                kptr += 4;
-            }
-
-            int sum[8];
-            __lsx_vst(_sum00, sum, 0);
-            __lsx_vst(_sum10, sum + 4, 0);
-
-            outptr0[0] = sum[0];
-            outptr1[0] = sum[1];
-            outptr2[0] = sum[2];
-            outptr3[0] = sum[3];
-            outptr0[1] = sum[4];
-            outptr1[1] = sum[5];
-            outptr2[1] = sum[6];
-            outptr3[1] = sum[7];
-            outptr0 += 2;
-            outptr1 += 2;
-            outptr2 += 2;
-            outptr3 += 2;
-        }
-        for (; i < size; i++)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-            const signed char* kptr = kernel.channel(p / 4);
-
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            __m128i _sum0 = __lsx_vreplgr2vr_w(0);
-
-            if (nn4 > 0)
-            {
-                __m128i _sum1 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum2 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum3 = __lsx_vreplgr2vr_w(0);
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    __m128i _val = __lsx_vld(tmpptr, 0);
-                    __m128i _val16 = __lsx_vilvl_b(__lsx_vslti_b(_val, 0), _val);
-
-                    _val16 = __lsx_vilvl_d(_val16, _val16);
-
-                    __m128i _w01 = __lsx_vld(kptr, 0);
-                    __m128i _extw01 = __lsx_vslti_b(_w01, 0);
-                    __m128i _w0 = __lsx_vilvl_b(_extw01, _w01);
-                    __m128i _w1 = __lsx_vilvh_b(_extw01, _w01);
-
-                    __m128i _s0 = __lsx_vmul_h(_val16, _w0);
-                    __m128i _s1 = __lsx_vmul_h(_val16, _w1);
-
-                    __m128i _exts0 = __lsx_vslti_h(_s0, 0);
-                    __m128i _exts1 = __lsx_vslti_h(_s1, 0);
-                    __m128i _s0l = __lsx_vilvl_h(_exts0, _s0);
-                    __m128i _s0h = __lsx_vilvh_h(_exts0, _s0);
-                    __m128i _s1l = __lsx_vilvl_h(_exts1, _s1);
-                    __m128i _s1h = __lsx_vilvh_h(_exts1, _s1);
-
-                    _sum0 = __lsx_vadd_w(_sum0, _s0l);
-                    _sum1 = __lsx_vadd_w(_sum1, _s0h);
-                    _sum2 = __lsx_vadd_w(_sum2, _s1l);
-                    _sum3 = __lsx_vadd_w(_sum3, _s1h);
-
-                    tmpptr += 4;
-                    kptr += 16;
-                }
-
-                // transpose 4x4
-                {
-                    __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                    _tmp0 = __lsx_vilvl_w(_sum1, _sum0);
-                    _tmp1 = __lsx_vilvl_w(_sum3, _sum2);
-                    _tmp2 = __lsx_vilvh_w(_sum1, _sum0);
-                    _tmp3 = __lsx_vilvh_w(_sum3, _sum2);
-                    _sum0 = __lsx_vilvl_d(_tmp1, _tmp0);
-                    _sum1 = __lsx_vilvh_d(_tmp1, _tmp0);
-                    _sum2 = __lsx_vilvl_d(_tmp3, _tmp2);
-                    _sum3 = __lsx_vilvh_d(_tmp3, _tmp2);
-                }
-
-                _sum0 = __lsx_vadd_w(_sum0, _sum1);
-                _sum2 = __lsx_vadd_w(_sum2, _sum3);
-                _sum0 = __lsx_vadd_w(_sum0, _sum2);
-            }
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                __m128i _val = __lsx_vreplgr2vr_h(tmpptr[0]);
-
-                __m128i _w = __lsx_vld(kptr, 0);
-                __m128i _w16 = __lsx_vilvl_b(__lsx_vslti_b(_w, 0), _w);
-
-                __m128i _s0 = __lsx_vmul_h(_val, _w16);
-                __m128i _s032 = __lsx_vilvl_h(__lsx_vslti_h(_s0, 0), _s0);
-
-                _sum0 = __lsx_vadd_w(_sum0, _s032);
-
-                tmpptr += 1;
-                kptr += 4;
-            }
-
-            int sum[4];
-            __lsx_vst(_sum0, sum, 0);
-
-            outptr0[0] = sum[0];
-            outptr1[0] = sum[1];
-            outptr2[0] = sum[2];
-            outptr3[0] = sum[3];
-            outptr0 += 1;
-            outptr1 += 1;
-            outptr2 += 1;
-            outptr3 += 1;
-        }
-    }
-#else // __loongarch_sx
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        int* outptr0 = top_blob.channel(p);
-        int* outptr1 = top_blob.channel(p + 1);
-
-        int i = 0;
-        for (; i + 1 < size; i += 2)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2);
-            const signed char* kptr = kernel.channel(p / 2);
-
-            int sum00 = 0;
-            int sum01 = 0;
-            int sum10 = 0;
-            int sum11 = 0;
-
-            int nn1 = inch * maxk;
-
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                signed char val0 = tmpptr[0];
-                signed char val1 = tmpptr[1];
-                signed char w0 = kptr[0];
-                signed char w1 = kptr[1];
-
-                sum00 += val0 * w0;
-                sum01 += val1 * w0;
-                sum10 += val0 * w1;
-                sum11 += val1 * w1;
-
-                tmpptr += 2;
-                kptr += 2;
-            }
-
-            outptr0[0] = sum00;
-            outptr0[1] = sum01;
-            outptr1[0] = sum10;
-            outptr1[1] = sum11;
-            outptr0 += 2;
-            outptr1 += 2;
-        }
-        for (; i < size; i++)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-            const signed char* kptr = kernel.channel(p / 2);
-
-            int sum00 = 0;
-            int sum10 = 0;
-
-            int nn1 = inch * maxk;
-
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                signed char val0 = tmpptr[0];
-                signed char w0 = kptr[0];
-                signed char w1 = kptr[1];
-
-                sum00 += val0 * w0;
-                sum10 += val0 * w1;
-
-                tmpptr += 1;
-                kptr += 2;
-            }
-
-            outptr0[0] = sum00;
-            outptr1[0] = sum10;
-            outptr0 += 1;
-            outptr1 += 1;
-        }
-    }
-#endif // __loongarch_sx
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        int* outptr0 = top_blob.channel(p);
-
-        int i = 0;
-        for (; i + 1 < size; i += 2)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2);
-#if __loongarch_sx
-            const signed char* kptr = kernel.channel(p / 4 + p % 4);
-#else
-            const signed char* kptr = kernel.channel(p / 2 + p % 2);
-#endif
-
-            int sum0 = 0;
-            int sum1 = 0;
-
-#if __loongarch_sx
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            if (nn4 > 0)
-            {
-                __m128i _sum0 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum1 = __lsx_vreplgr2vr_w(0);
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    __m128i _val = __lsx_vld(tmpptr, 0);
-                    __m128i _val16 = __lsx_vilvl_b(__lsx_vslti_b(_val, 0), _val);
-
-                    __m128i _w = __lsx_vld(kptr, 0);
-                    __m128i _w16 = __lsx_vilvl_b(__lsx_vslti_b(_w, 0), _w);
-
-                    _w16 = __lsx_vilvl_d(_w16, _w16);
-
-                    __m128i _s0 = __lsx_vmul_h(_val16, _w16);
-                    __m128i _exts0 = __lsx_vslti_h(_s0, 0);
-                    __m128i _s0l = __lsx_vilvl_h(_exts0, _s0);
-                    __m128i _s0h = __lsx_vilvh_h(_exts0, _s0);
-
-                    _sum0 = __lsx_vadd_w(_sum0, _s0l);
-                    _sum1 = __lsx_vadd_w(_sum1, _s0h);
-
-                    tmpptr += 8;
-                    kptr += 4;
-                }
-
-                sum0 = __lsx_reduce_add_w(_sum0);
-                sum1 = __lsx_reduce_add_w(_sum1);
-            }
-#else
-            int nn1 = inch * maxk;
-#endif // __loongarch_sx
-
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                signed char val0 = tmpptr[0];
-                signed char val1 = tmpptr[1];
-                signed char w = kptr[0];
-
-                sum0 += val0 * w;
-                sum1 += val1 * w;
-
-                tmpptr += 2;
-                kptr += 1;
-            }
-
-            outptr0[0] = sum0;
-            outptr0[1] = sum1;
-            outptr0 += 2;
-        }
-        for (; i < size; i++)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-#if __loongarch_sx
-            const signed char* kptr = kernel.channel(p / 4 + p % 4);
-#else
-            const signed char* kptr = kernel.channel(p / 2 + p % 2);
-#endif
-
-            int sum = 0;
-
-#if __loongarch_sx
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            if (nn4 > 0)
-            {
-                __m128i _sum = __lsx_vreplgr2vr_w(0);
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    __m128i _val = __lsx_vld(tmpptr, 0);
-                    __m128i _val16 = __lsx_vilvl_b(__lsx_vslti_b(_val, 0), _val);
-
-                    __m128i _w = __lsx_vld(kptr, 0);
-                    __m128i _w16 = __lsx_vilvl_b(__lsx_vslti_b(_w, 0), _w);
-
-                    __m128i _s0 = __lsx_vmul_h(_val16, _w16);
-                    __m128i _s032 = __lsx_vilvl_h(__lsx_vslti_h(_s0, 0), _s0);
-
-                    _sum = __lsx_vadd_w(_sum, _s032);
-
-                    tmpptr += 4;
-                    kptr += 4;
-                }
-
-                sum = __lsx_reduce_add_w(_sum);
-            }
-#else
-            int nn1 = inch * maxk;
-#endif // __loongarch_sx
-
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                signed char val = tmpptr[0];
-                signed char w = kptr[0];
-
-                sum += val * w;
-
-                tmpptr += 1;
-                kptr += 1;
-            }
-
-            outptr0[0] = sum;
-            outptr0 += 1;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_transform_kernel_int8_lsx(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = 4a-4b-maxk-inch/4a-outch/4b
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-#if __loongarch_sx
-    if (outch >= 4)
-    {
-        if (inch >= 4)
-            kernel_tm.create(16 * maxk, inch / 4 + inch % 4, outch / 4 + outch % 4, (size_t)1u);
-        else
-            kernel_tm.create(4 * maxk, inch, outch / 4 + outch % 4, (size_t)1u);
-    }
-#else
-    if (outch >= 2)
-    {
-        kernel_tm.create(2 * maxk, inch, outch / 2 + outch % 2, (size_t)1u);
-    }
-#endif // __loongarch_sx
-    else
-    {
-#if __loongarch_sx
-        if (inch >= 4)
-            kernel_tm.create(4 * maxk, inch / 4 + inch % 4, outch, (size_t)1u);
-        else
-#endif // __loongarch_sx
-        {
-            kernel_tm.create(1 * maxk, inch, outch, (size_t)1u);
-        }
-    }
-
-    int q = 0;
-#if __loongarch_sx
-    for (; q + 3 < outch; q += 4)
-    {
-        signed char* g00 = kernel_tm.channel(q / 4);
-
-        int p = 0;
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 4; j++)
-                    {
-                        const signed char* k00 = kernel.channel(q + i).row<const signed char>(p + j);
-                        g00[0] = k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    const signed char* k00 = kernel.channel(q + i).row<const signed char>(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#else  // __loongarch_sx
-    for (; q + 1 < outch; q += 2)
-    {
-        signed char* g00 = kernel_tm.channel(q / 2);
-
-        int p = 0;
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 2; i++)
-                {
-                    const signed char* k00 = kernel.channel(q + i).row<const signed char>(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#endif // __loongarch_sx
-    for (; q < outch; q++)
-    {
-#if __loongarch_sx
-        signed char* g00 = kernel_tm.channel(q / 4 + q % 4);
-#else
-        signed char* g00 = kernel_tm.channel(q / 2 + q % 2);
-#endif
-
-        int p = 0;
-#if __loongarch_sx
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int j = 0; j < 4; j++)
-                {
-                    const signed char* k00 = kernel.channel(q).row<const signed char>(p + j);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-#endif // __loongarch_sx
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k00 = kernel.channel(q).row<const signed char>(p);
-                g00[0] = k00[k];
-                g00++;
-            }
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 1u, 1, opt.workspace_allocator);
-    {
-        const int gap = w * stride_h - outw * stride_w;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            signed char* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const signed char* sptr = img.row<const signed char>(dilation_h * u) + dilation_w * v;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j + 3 < outw; j += 4)
-                        {
-                            ptr[0] = sptr[0];
-                            ptr[1] = sptr[stride_w];
-                            ptr[2] = sptr[stride_w * 2];
-                            ptr[3] = sptr[stride_w * 3];
-
-                            sptr += stride_w * 4;
-                            ptr += 4;
-                        }
-                        for (; j + 1 < outw; j += 2)
-                        {
-                            ptr[0] = sptr[0];
-                            ptr[1] = sptr[stride_w];
-
-                            sptr += stride_w * 2;
-                            ptr += 2;
-                        }
-                        for (; j < outw; j++)
-                        {
-                            ptr[0] = sptr[0];
-
-                            sptr += stride_w;
-                            ptr += 1;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_int8_lsx(bottom_im2col, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack1to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack1to4_int8.h
deleted file mode 100644
index 3429bfae..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack1to4_int8.h
+++ /dev/null
@@ -1,481 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_pack1to4_int8_lsx(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    // Mat bottom_im2col(size, maxk, inch, 8u, 8, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    // permute
-    Mat tmp;
-    if (inch >= 4)
-    {
-        if (size >= 2)
-            tmp.create(2 * maxk, inch / 4 + inch % 4, size / 2 + size % 2, 4u, 4, opt.workspace_allocator);
-        else
-            tmp.create(maxk, inch / 4 + inch % 4, size, 4u, 4, opt.workspace_allocator);
-    }
-    else
-    {
-        if (size >= 2)
-            tmp.create(2 * maxk, inch, size / 2 + size % 2, 1u, 1, opt.workspace_allocator);
-        else
-            tmp.create(maxk, inch, size, 1u, 1, opt.workspace_allocator);
-    }
-    {
-        int remain_size_start = 0;
-        int nn_size = (size - remain_size_start) >> 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 2;
-
-            signed char* tmpptr = tmp.channel(i / 2);
-
-            int q = 0;
-            for (; q + 3 < inch; q += 4)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-                const signed char* img1 = (const signed char*)bottom_im2col.channel(q + 1) + i;
-                const signed char* img2 = (const signed char*)bottom_im2col.channel(q + 2) + i;
-                const signed char* img3 = (const signed char*)bottom_im2col.channel(q + 3) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img1[0];
-                    tmpptr[2] = img2[0];
-                    tmpptr[3] = img3[0];
-                    tmpptr[4] = img0[1];
-                    tmpptr[5] = img1[1];
-                    tmpptr[6] = img2[1];
-                    tmpptr[7] = img3[1];
-                    tmpptr += 8;
-
-                    img0 += size;
-                    img1 += size;
-                    img2 += size;
-                    img3 += size;
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img0[1];
-
-                    tmpptr += 2;
-
-                    img0 += size;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-
-            int q = 0;
-            for (; q + 3 < inch; q += 4)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-                const signed char* img1 = (const signed char*)bottom_im2col.channel(q + 1) + i;
-                const signed char* img2 = (const signed char*)bottom_im2col.channel(q + 2) + i;
-                const signed char* img3 = (const signed char*)bottom_im2col.channel(q + 3) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img1[0];
-                    tmpptr[2] = img2[0];
-                    tmpptr[3] = img3[0];
-                    tmpptr += 4;
-
-                    img0 += size;
-                    img1 += size;
-                    img2 += size;
-                    img3 += size;
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-
-                    tmpptr += 1;
-
-                    img0 += size;
-                }
-            }
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        int* outptr0 = top_blob.channel(p);
-
-        int i = 0;
-        for (; i + 1 < size; i += 2)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2);
-            const signed char* kptr = kernel.channel(p);
-
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            __m128i _sum00 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum10 = __lsx_vreplgr2vr_w(0);
-
-            if (nn4 > 0)
-            {
-                __m128i _sum01 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum02 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum03 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum11 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum12 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum13 = __lsx_vreplgr2vr_w(0);
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    __builtin_prefetch(tmpptr + 32);
-                    __builtin_prefetch(kptr + 64);
-                    __m128i _val = __lsx_vld(tmpptr, 0);
-                    __m128i _val01 = __lsx_vilvl_b(__lsx_vslti_b(_val, 0), _val);
-
-                    __m128i _val0 = __lsx_vilvl_d(_val01, _val01);
-                    __m128i _val1 = __lsx_vilvh_d(_val01, _val01);
-
-                    __m128i _w01 = __lsx_vld(kptr, 0);
-                    __m128i _extw01 = __lsx_vslti_b(_w01, 0);
-                    __m128i _w0 = __lsx_vilvl_b(_extw01, _w01);
-                    __m128i _w1 = __lsx_vilvh_b(_extw01, _w01);
-
-                    __m128i _s00 = __lsx_vmul_h(_val0, _w0);
-                    __m128i _s01 = __lsx_vmul_h(_val0, _w1);
-                    __m128i _s10 = __lsx_vmul_h(_val1, _w0);
-                    __m128i _s11 = __lsx_vmul_h(_val1, _w1);
-
-                    __m128i _exts00 = __lsx_vslti_h(_s00, 0);
-                    __m128i _exts01 = __lsx_vslti_h(_s01, 0);
-                    __m128i _exts10 = __lsx_vslti_h(_s10, 0);
-                    __m128i _exts11 = __lsx_vslti_h(_s11, 0);
-                    __m128i _s00l = __lsx_vilvl_h(_exts00, _s00);
-                    __m128i _s00h = __lsx_vilvh_h(_exts00, _s00);
-                    __m128i _s01l = __lsx_vilvl_h(_exts01, _s01);
-                    __m128i _s01h = __lsx_vilvh_h(_exts01, _s01);
-                    __m128i _s10l = __lsx_vilvl_h(_exts10, _s10);
-                    __m128i _s10h = __lsx_vilvh_h(_exts10, _s10);
-                    __m128i _s11l = __lsx_vilvl_h(_exts11, _s11);
-                    __m128i _s11h = __lsx_vilvh_h(_exts11, _s11);
-
-                    _sum00 = __lsx_vadd_w(_sum00, _s00l);
-                    _sum01 = __lsx_vadd_w(_sum01, _s00h);
-                    _sum02 = __lsx_vadd_w(_sum02, _s01l);
-                    _sum03 = __lsx_vadd_w(_sum03, _s01h);
-                    _sum10 = __lsx_vadd_w(_sum10, _s10l);
-                    _sum11 = __lsx_vadd_w(_sum11, _s10h);
-                    _sum12 = __lsx_vadd_w(_sum12, _s11l);
-                    _sum13 = __lsx_vadd_w(_sum13, _s11h);
-
-                    tmpptr += 8;
-                    kptr += 16;
-                }
-
-                // transpose 4x4
-                {
-                    __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                    _tmp0 = __lsx_vilvl_w(_sum01, _sum00);
-                    _tmp1 = __lsx_vilvl_w(_sum03, _sum02);
-                    _tmp2 = __lsx_vilvh_w(_sum01, _sum00);
-                    _tmp3 = __lsx_vilvh_w(_sum03, _sum02);
-                    _sum00 = __lsx_vilvl_d(_tmp1, _tmp0);
-                    _sum01 = __lsx_vilvh_d(_tmp1, _tmp0);
-                    _sum02 = __lsx_vilvl_d(_tmp3, _tmp2);
-                    _sum03 = __lsx_vilvh_d(_tmp3, _tmp2);
-                }
-                {
-                    __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                    _tmp0 = __lsx_vilvl_w(_sum11, _sum10);
-                    _tmp1 = __lsx_vilvl_w(_sum13, _sum12);
-                    _tmp2 = __lsx_vilvh_w(_sum11, _sum10);
-                    _tmp3 = __lsx_vilvh_w(_sum13, _sum12);
-                    _sum10 = __lsx_vilvl_d(_tmp1, _tmp0);
-                    _sum11 = __lsx_vilvh_d(_tmp1, _tmp0);
-                    _sum12 = __lsx_vilvl_d(_tmp3, _tmp2);
-                    _sum13 = __lsx_vilvh_d(_tmp3, _tmp2);
-                }
-
-                _sum00 = __lsx_vadd_w(_sum00, _sum01);
-                _sum02 = __lsx_vadd_w(_sum02, _sum03);
-                _sum10 = __lsx_vadd_w(_sum10, _sum11);
-                _sum12 = __lsx_vadd_w(_sum12, _sum13);
-
-                _sum00 = __lsx_vadd_w(_sum00, _sum02);
-                _sum10 = __lsx_vadd_w(_sum10, _sum12);
-            }
-
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                __m128i _val0 = __lsx_vreplgr2vr_h(tmpptr[0]);
-                __m128i _val1 = __lsx_vreplgr2vr_h(tmpptr[1]);
-                __m128i _val = __lsx_vilvl_d(_val1, _val0);
-
-                __m128i _w = __lsx_vld(kptr, 0);
-                __m128i _w16 = __lsx_vilvl_b(__lsx_vslti_b(_w, 0), _w);
-
-                _w16 = __lsx_vilvl_d(_w16, _w16);
-
-                __m128i _s0 = __lsx_vmul_h(_val, _w16);
-                __m128i _exts0 = __lsx_vslti_h(_s0, 0);
-                __m128i _s0l = __lsx_vilvl_h(_exts0, _s0);
-                __m128i _s0h = __lsx_vilvh_h(_exts0, _s0);
-
-                _sum00 = __lsx_vadd_w(_sum00, _s0l);
-                _sum10 = __lsx_vadd_w(_sum10, _s0h);
-
-                tmpptr += 2;
-                kptr += 4;
-            }
-
-            __lsx_vst(_sum00, outptr0, 0);
-            __lsx_vst(_sum10, outptr0 + 4, 0);
-            outptr0 += 8;
-        }
-        for (; i < size; i++)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-            const signed char* kptr = kernel.channel(p);
-
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            __m128i _sum0 = __lsx_vreplgr2vr_w(0);
-
-            if (nn4 > 0)
-            {
-                __m128i _sum1 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum2 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum3 = __lsx_vreplgr2vr_w(0);
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    __builtin_prefetch(tmpptr + 16);
-                    __builtin_prefetch(kptr + 64);
-                    __m128i _val = __lsx_vld(tmpptr, 0);
-                    __m128i _val16 = __lsx_vilvl_b(__lsx_vslti_b(_val, 0), _val);
-
-                    _val16 = __lsx_vilvl_d(_val16, _val16);
-
-                    __m128i _w01 = __lsx_vld(kptr, 0);
-                    __m128i _extw01 = __lsx_vslti_b(_w01, 0);
-                    __m128i _w0 = __lsx_vilvl_b(_extw01, _w01);
-                    __m128i _w1 = __lsx_vilvh_b(_extw01, _w01);
-
-                    __m128i _s0 = __lsx_vmul_h(_val16, _w0);
-                    __m128i _s1 = __lsx_vmul_h(_val16, _w1);
-
-                    __m128i _exts0 = __lsx_vslti_h(_s0, 0);
-                    __m128i _exts1 = __lsx_vslti_h(_s1, 0);
-                    __m128i _s0l = __lsx_vilvl_h(_exts0, _s0);
-                    __m128i _s0h = __lsx_vilvh_h(_exts0, _s0);
-                    __m128i _s1l = __lsx_vilvl_h(_exts1, _s1);
-                    __m128i _s1h = __lsx_vilvh_h(_exts1, _s1);
-
-                    _sum0 = __lsx_vadd_w(_sum0, _s0l);
-                    _sum1 = __lsx_vadd_w(_sum1, _s0h);
-                    _sum2 = __lsx_vadd_w(_sum2, _s1l);
-                    _sum3 = __lsx_vadd_w(_sum3, _s1h);
-
-                    tmpptr += 4;
-                    kptr += 16;
-                }
-
-                // transpose 4x4
-                {
-                    __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                    _tmp0 = __lsx_vilvl_w(_sum1, _sum0);
-                    _tmp1 = __lsx_vilvl_w(_sum3, _sum2);
-                    _tmp2 = __lsx_vilvh_w(_sum1, _sum0);
-                    _tmp3 = __lsx_vilvh_w(_sum3, _sum2);
-                    _sum0 = __lsx_vilvl_d(_tmp1, _tmp0);
-                    _sum1 = __lsx_vilvh_d(_tmp1, _tmp0);
-                    _sum2 = __lsx_vilvl_d(_tmp3, _tmp2);
-                    _sum3 = __lsx_vilvh_d(_tmp3, _tmp2);
-                }
-
-                _sum0 = __lsx_vadd_w(_sum0, _sum1);
-                _sum2 = __lsx_vadd_w(_sum2, _sum3);
-                _sum0 = __lsx_vadd_w(_sum0, _sum2);
-            }
-
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                __m128i _val = __lsx_vreplgr2vr_h(tmpptr[0]);
-
-                __m128i _w = __lsx_vld(kptr, 0);
-                __m128i _w16 = __lsx_vilvl_b(__lsx_vslti_b(_w, 0), _w);
-
-                __m128i _s0 = __lsx_vmul_h(_val, _w16);
-                __m128i _s032 = __lsx_vilvl_h(__lsx_vslti_h(_s0, 0), _s0);
-
-                _sum0 = __lsx_vadd_w(_sum0, _s032);
-
-                tmpptr += 1;
-                kptr += 4;
-            }
-
-            __lsx_vst(_sum0, outptr0, 0);
-            outptr0 += 4;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_transform_kernel_pack1to4_int8_lsx(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = 4a-4b-maxk-inch/4a-outch/4b
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-    if (inch >= 4)
-        kernel_tm.create(16 * maxk, inch / 4 + inch % 4, outch / 4, (size_t)1u);
-    else
-        kernel_tm.create(4 * maxk, inch, outch / 4, (size_t)1u);
-
-    for (int q = 0; q + 3 < outch; q += 4)
-    {
-        signed char* g00 = kernel_tm.channel(q / 4);
-
-        int p = 0;
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 4; j++)
-                    {
-                        const signed char* k00 = kernel.channel(q + i).row<const signed char>(p + j);
-
-                        g00[0] = k00[k];
-
-                        g00++;
-                    }
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    const signed char* k00 = kernel.channel(q + i).row<const signed char>(p);
-
-                    g00[0] = k00[k];
-
-                    g00++;
-                }
-            }
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_pack1to4_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 1u, 1, opt.workspace_allocator);
-    {
-        const int gap = w * stride_h - outw * stride_w;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            signed char* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const signed char* sptr = img.row<const signed char>(dilation_h * u) + dilation_w * v;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j + 3 < outw; j += 4)
-                        {
-                            ptr[0] = sptr[0];
-                            ptr[1] = sptr[stride_w];
-                            ptr[2] = sptr[stride_w * 2];
-                            ptr[3] = sptr[stride_w * 3];
-
-                            sptr += stride_w * 4;
-                            ptr += 4;
-                        }
-                        for (; j + 1 < outw; j += 2)
-                        {
-                            ptr[0] = sptr[0];
-                            ptr[1] = sptr[stride_w];
-
-                            sptr += stride_w * 2;
-                            ptr += 2;
-                        }
-                        for (; j < outw; j++)
-                        {
-                            ptr[0] = sptr[0];
-
-                            sptr += stride_w;
-                            ptr += 1;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_pack1to4_int8_lsx(bottom_im2col, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack4.h
deleted file mode 100644
index e3e7279a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack4.h
+++ /dev/null
@@ -1,519 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_pack4_lsx(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    // Mat bottom_im2col(size, maxk, inch, 4u * 4, 4, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    // permute
-    Mat tmp;
-    if (size >= 12)
-        tmp.create(12 * maxk, inch, size / 12 + (size % 12) / 8 + (size % 12 % 8) / 4 + (size % 12 % 4) / 2 + size % 12 % 2, 4u * 4, 4, opt.workspace_allocator);
-    else if (size >= 8)
-        tmp.create(8 * maxk, inch, size / 8 + (size % 8) / 4 + (size % 4) / 2 + size % 2, 4u * 4, 4, opt.workspace_allocator);
-    else if (size >= 4)
-        tmp.create(4 * maxk, inch, size / 4 + (size % 4) / 2 + size % 2, 4u * 4, 4, opt.workspace_allocator);
-    else if (size >= 2)
-        tmp.create(2 * maxk, inch, size / 2 + size % 2, 4u * 4, 4, opt.workspace_allocator);
-    else
-        tmp.create(maxk, inch, size, 4u * 4, 4, opt.workspace_allocator);
-    {
-        int remain_size_start = 0;
-        int nn_size = size / 12;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 12;
-
-            float* tmpptr = tmp.channel(i / 12);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    // transpose 4x12
-                    __m128i _r0 = __lsx_vld(img0, 0);
-                    __m128i _r1 = __lsx_vld(img0 + 4, 0);
-                    __m128i _r2 = __lsx_vld(img0 + 4 * 2, 0);
-                    __m128i _r3 = __lsx_vld(img0 + 4 * 3, 0);
-                    __m128i _r4 = __lsx_vld(img0 + 4 * 4, 0);
-                    __m128i _r5 = __lsx_vld(img0 + 4 * 5, 0);
-                    __m128i _r6 = __lsx_vld(img0 + 4 * 6, 0);
-                    __m128i _r7 = __lsx_vld(img0 + 4 * 7, 0);
-                    __m128i _r8 = __lsx_vld(img0 + 4 * 8, 0);
-                    __m128i _r9 = __lsx_vld(img0 + 4 * 9, 0);
-                    __m128i _ra = __lsx_vld(img0 + 4 * 10, 0);
-                    __m128i _rb = __lsx_vld(img0 + 4 * 11, 0);
-
-                    __m128i _r01r = __lsx_vilvl_w(_r1, _r0);
-                    __m128i _r01l = __lsx_vilvh_w(_r1, _r0);
-                    __m128i _r23r = __lsx_vilvl_w(_r3, _r2);
-                    __m128i _r23l = __lsx_vilvh_w(_r3, _r2);
-                    __m128i _r45r = __lsx_vilvl_w(_r5, _r4);
-                    __m128i _r45l = __lsx_vilvh_w(_r5, _r4);
-                    __m128i _r67r = __lsx_vilvl_w(_r7, _r6);
-                    __m128i _r67l = __lsx_vilvh_w(_r7, _r6);
-                    __m128i _r89r = __lsx_vilvl_w(_r9, _r8);
-                    __m128i _r89l = __lsx_vilvh_w(_r9, _r8);
-                    __m128i _rabr = __lsx_vilvl_w(_rb, _ra);
-                    __m128i _rabl = __lsx_vilvh_w(_rb, _ra);
-                    __m128i _r0123_0 = __lsx_vilvl_d(_r23r, _r01r);
-                    __m128i _r0123_1 = __lsx_vilvh_d(_r23r, _r01r);
-                    __m128i _r0123_2 = __lsx_vilvl_d(_r23l, _r01l);
-                    __m128i _r0123_3 = __lsx_vilvh_d(_r23l, _r01l);
-                    __m128i _r4567_0 = __lsx_vilvl_d(_r67r, _r45r);
-                    __m128i _r4567_1 = __lsx_vilvh_d(_r67r, _r45r);
-                    __m128i _r4567_2 = __lsx_vilvl_d(_r67l, _r45l);
-                    __m128i _r4567_3 = __lsx_vilvh_d(_r67l, _r45l);
-                    __m128i _r89ab_0 = __lsx_vilvl_d(_rabr, _r89r);
-                    __m128i _r89ab_1 = __lsx_vilvh_d(_rabr, _r89r);
-                    __m128i _r89ab_2 = __lsx_vilvl_d(_rabl, _r89l);
-                    __m128i _r89ab_3 = __lsx_vilvh_d(_rabl, _r89l);
-
-                    __lsx_vst(_r0123_0, tmpptr, 0);
-                    __lsx_vst(_r4567_0, tmpptr + 4, 0);
-                    __lsx_vst(_r89ab_0, tmpptr + 4 * 2, 0);
-                    __lsx_vst(_r0123_1, tmpptr + 4 * 3, 0);
-                    __lsx_vst(_r4567_1, tmpptr + 4 * 4, 0);
-                    __lsx_vst(_r89ab_1, tmpptr + 4 * 5, 0);
-                    __lsx_vst(_r0123_2, tmpptr + 4 * 6, 0);
-                    __lsx_vst(_r4567_2, tmpptr + 4 * 7, 0);
-                    __lsx_vst(_r89ab_2, tmpptr + 4 * 8, 0);
-                    __lsx_vst(_r0123_3, tmpptr + 4 * 9, 0);
-                    __lsx_vst(_r4567_3, tmpptr + 4 * 10, 0);
-                    __lsx_vst(_r89ab_3, tmpptr + 4 * 11, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 48;
-                }
-            }
-        }
-
-        remain_size_start += nn_size * 12;
-        nn_size = (size - remain_size_start) >> 3;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 8;
-
-            float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    // transpose 4x8
-                    __m128i _r0 = __lsx_vld(img0, 0);
-                    __m128i _r1 = __lsx_vld(img0 + 4, 0);
-                    __m128i _r2 = __lsx_vld(img0 + 4 * 2, 0);
-                    __m128i _r3 = __lsx_vld(img0 + 4 * 3, 0);
-                    __m128i _r4 = __lsx_vld(img0 + 4 * 4, 0);
-                    __m128i _r5 = __lsx_vld(img0 + 4 * 5, 0);
-                    __m128i _r6 = __lsx_vld(img0 + 4 * 6, 0);
-                    __m128i _r7 = __lsx_vld(img0 + 4 * 7, 0);
-
-                    __m128i _r01r = __lsx_vilvl_w(_r1, _r0);
-                    __m128i _r01l = __lsx_vilvh_w(_r1, _r0);
-                    __m128i _r23r = __lsx_vilvl_w(_r3, _r2);
-                    __m128i _r23l = __lsx_vilvh_w(_r3, _r2);
-                    __m128i _r45r = __lsx_vilvl_w(_r5, _r4);
-                    __m128i _r45l = __lsx_vilvh_w(_r5, _r4);
-                    __m128i _r67r = __lsx_vilvl_w(_r7, _r6);
-                    __m128i _r67l = __lsx_vilvh_w(_r7, _r6);
-                    __m128i _r0123_0 = __lsx_vilvl_d(_r23r, _r01r);
-                    __m128i _r0123_1 = __lsx_vilvh_d(_r23r, _r01r);
-                    __m128i _r0123_2 = __lsx_vilvl_d(_r23l, _r01l);
-                    __m128i _r0123_3 = __lsx_vilvh_d(_r23l, _r01l);
-                    __m128i _r4567_0 = __lsx_vilvl_d(_r67r, _r45r);
-                    __m128i _r4567_1 = __lsx_vilvh_d(_r67r, _r45r);
-                    __m128i _r4567_2 = __lsx_vilvl_d(_r67l, _r45l);
-                    __m128i _r4567_3 = __lsx_vilvh_d(_r67l, _r45l);
-
-                    __lsx_vst(_r0123_0, tmpptr, 0);
-                    __lsx_vst(_r4567_0, tmpptr + 4, 0);
-                    __lsx_vst(_r0123_1, tmpptr + 4 * 2, 0);
-                    __lsx_vst(_r4567_1, tmpptr + 4 * 3, 0);
-                    __lsx_vst(_r0123_2, tmpptr + 4 * 4, 0);
-                    __lsx_vst(_r4567_2, tmpptr + 4 * 5, 0);
-                    __lsx_vst(_r0123_3, tmpptr + 4 * 6, 0);
-                    __lsx_vst(_r4567_3, tmpptr + 4 * 7, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 32;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 3;
-        nn_size = (size - remain_size_start) >> 2;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 4;
-
-            float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    // transpose 4x4
-                    __m128i _r0 = __lsx_vld(img0, 0);
-                    __m128i _r1 = __lsx_vld(img0 + 4, 0);
-                    __m128i _r2 = __lsx_vld(img0 + 4 * 2, 0);
-                    __m128i _r3 = __lsx_vld(img0 + 4 * 3, 0);
-
-                    __m128i _r01r = __lsx_vilvl_w(_r1, _r0);
-                    __m128i _r01l = __lsx_vilvh_w(_r1, _r0);
-                    __m128i _r23r = __lsx_vilvl_w(_r3, _r2);
-                    __m128i _r23l = __lsx_vilvh_w(_r3, _r2);
-                    __m128i _r0123_0 = __lsx_vilvl_d(_r23r, _r01r);
-                    __m128i _r0123_1 = __lsx_vilvh_d(_r23r, _r01r);
-                    __m128i _r0123_2 = __lsx_vilvl_d(_r23l, _r01l);
-                    __m128i _r0123_3 = __lsx_vilvh_d(_r23l, _r01l);
-
-                    __lsx_vst(_r0123_0, tmpptr, 0);
-                    __lsx_vst(_r0123_1, tmpptr + 4, 0);
-                    __lsx_vst(_r0123_2, tmpptr + 4 * 2, 0);
-                    __lsx_vst(_r0123_3, tmpptr + 4 * 3, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 16;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 2;
-        nn_size = (size - remain_size_start) >> 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 2;
-
-            float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    // transpose 4x2
-                    __m128i _r0 = __lsx_vld(img0, 0);
-                    __m128i _r1 = __lsx_vld(img0 + 4, 0);
-
-                    __m128i _r01_0 = __lsx_vilvl_w(_r1, _r0);
-                    __m128i _r01_1 = __lsx_vilvh_w(_r1, _r0);
-
-                    __lsx_vst(_r01_0, tmpptr, 0);
-                    __lsx_vst(_r01_1, tmpptr + 4, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 8;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2 + i % 12 % 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    __m128i _val = __lsx_vld(img0, 0);
-                    __lsx_vst(_val, tmpptr, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 4;
-                }
-            }
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr0 = top_blob.channel(p);
-
-        int i = 0;
-        for (; i + 11 < size; i += 12)
-        {
-            const float* tmpptr = tmp.channel(i / 12);
-            const float* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            __m128 _sum0 = bias ? (__m128)__lsx_vld(bias + p * 4, 0) : (__m128)__lsx_vreplgr2vr_w(0);
-            __m128 _sum1 = _sum0;
-            __m128 _sum2 = _sum0;
-            __m128 _sum3 = _sum0;
-            __m128 _sum4 = _sum0;
-            __m128 _sum5 = _sum0;
-            __m128 _sum6 = _sum0;
-            __m128 _sum7 = _sum0;
-            __m128 _sum8 = _sum0;
-            __m128 _sum9 = _sum0;
-            __m128 _suma = _sum0;
-            __m128 _sumb = _sum0;
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 48);
-                __builtin_prefetch(kptr0 + 16);
-                __m128i _val0123 = __lsx_vld(tmpptr, 0);
-                __m128i _val4567 = __lsx_vld(tmpptr + 4, 0);
-                __m128i _val89ab = __lsx_vld(tmpptr + 8, 0);
-                __m128 _w0 = (__m128)__lsx_vld(kptr0, 0);
-                _sum0 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 0), _sum0);
-                _sum1 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 1), _sum1);
-                _sum2 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 2), _sum2);
-                _sum3 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 3), _sum3);
-                _sum4 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 0), _sum4);
-                _sum5 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 1), _sum5);
-                _sum6 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 2), _sum6);
-                _sum7 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 3), _sum7);
-                _sum8 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val89ab, 0), _sum8);
-                _sum9 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val89ab, 1), _sum9);
-                _suma = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val89ab, 2), _suma);
-                _sumb = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val89ab, 3), _sumb);
-
-                tmpptr += 12;
-                kptr0 += 4;
-            }
-
-            __lsx_vst(_sum0, outptr0, 0);
-            __lsx_vst(_sum1, outptr0 + 4, 0);
-            __lsx_vst(_sum2, outptr0 + 4 * 2, 0);
-            __lsx_vst(_sum3, outptr0 + 4 * 3, 0);
-            __lsx_vst(_sum4, outptr0 + 4 * 4, 0);
-            __lsx_vst(_sum5, outptr0 + 4 * 5, 0);
-            __lsx_vst(_sum6, outptr0 + 4 * 6, 0);
-            __lsx_vst(_sum7, outptr0 + 4 * 7, 0);
-            __lsx_vst(_sum8, outptr0 + 4 * 8, 0);
-            __lsx_vst(_sum9, outptr0 + 4 * 9, 0);
-            __lsx_vst(_suma, outptr0 + 4 * 10, 0);
-            __lsx_vst(_sumb, outptr0 + 4 * 11, 0);
-
-            outptr0 += 4 * 12;
-        }
-        for (; i + 7 < size; i += 8)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8);
-            const float* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            __m128 _sum0 = bias ? (__m128)__lsx_vld(bias + p * 4, 0) : (__m128)__lsx_vreplgr2vr_w(0);
-            __m128 _sum1 = _sum0;
-            __m128 _sum2 = _sum0;
-            __m128 _sum3 = _sum0;
-            __m128 _sum4 = _sum0;
-            __m128 _sum5 = _sum0;
-            __m128 _sum6 = _sum0;
-            __m128 _sum7 = _sum0;
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 32);
-                __builtin_prefetch(kptr0 + 16);
-                __m128i _val0123 = __lsx_vld(tmpptr, 0);
-                __m128i _val4567 = __lsx_vld(tmpptr + 4, 0);
-                __m128 _w0 = (__m128)__lsx_vld(kptr0, 0);
-                _sum0 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 0), _sum0);
-                _sum1 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 1), _sum1);
-                _sum2 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 2), _sum2);
-                _sum3 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 3), _sum3);
-                _sum4 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 0), _sum4);
-                _sum5 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 1), _sum5);
-                _sum6 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 2), _sum6);
-                _sum7 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 3), _sum7);
-
-                tmpptr += 8;
-                kptr0 += 4;
-            }
-
-            __lsx_vst(_sum0, outptr0, 0);
-            __lsx_vst(_sum1, outptr0 + 4, 0);
-            __lsx_vst(_sum2, outptr0 + 4 * 2, 0);
-            __lsx_vst(_sum3, outptr0 + 4 * 3, 0);
-            __lsx_vst(_sum4, outptr0 + 4 * 4, 0);
-            __lsx_vst(_sum5, outptr0 + 4 * 5, 0);
-            __lsx_vst(_sum6, outptr0 + 4 * 6, 0);
-            __lsx_vst(_sum7, outptr0 + 4 * 7, 0);
-
-            outptr0 += 4 * 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4);
-            const float* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            __m128 _sum0 = bias ? (__m128)__lsx_vld(bias + p * 4, 0) : (__m128)__lsx_vreplgr2vr_w(0);
-            __m128 _sum1 = _sum0;
-            __m128 _sum2 = _sum0;
-            __m128 _sum3 = _sum0;
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr0 + 16);
-                __m128i _val0123 = __lsx_vld(tmpptr, 0);
-                __m128 _w0 = (__m128)__lsx_vld(kptr0, 0);
-                _sum0 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 0), _sum0);
-                _sum1 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 1), _sum1);
-                _sum2 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 2), _sum2);
-                _sum3 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 3), _sum3);
-
-                tmpptr += 4;
-                kptr0 += 4;
-            }
-
-            __lsx_vst(_sum0, outptr0, 0);
-            __lsx_vst(_sum1, outptr0 + 4, 0);
-            __lsx_vst(_sum2, outptr0 + 4 * 2, 0);
-            __lsx_vst(_sum3, outptr0 + 4 * 3, 0);
-
-            outptr0 += 4 * 4;
-        }
-        for (; i + 1 < size; i += 2)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2);
-            const float* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            __m128 _sum0 = bias ? (__m128)__lsx_vld(bias + p * 4, 0) : (__m128)__lsx_vreplgr2vr_w(0);
-            __m128 _sum1 = _sum0;
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 8);
-                __builtin_prefetch(kptr0 + 16);
-                __m128 _val0 = __lsx_vreplfr2vr_s(*tmpptr++);
-                __m128 _val1 = __lsx_vreplfr2vr_s(*tmpptr++);
-                __m128 _w0 = (__m128)__lsx_vld(kptr0, 0);
-                _sum0 = __lsx_vfmadd_s(_w0, _val0, _sum0);
-                _sum1 = __lsx_vfmadd_s(_w0, _val1, _sum1);
-
-                kptr0 += 4;
-            }
-
-            __lsx_vst(_sum0, outptr0, 0);
-            __lsx_vst(_sum1, outptr0 + 4, 0);
-
-            outptr0 += 4 * 2;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2 + i % 12 % 2);
-            const float* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            __m128 _sum = bias ? (__m128)__lsx_vld(bias + p * 4, 0) : (__m128)__lsx_vreplgr2vr_w(0);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 4);
-                __builtin_prefetch(kptr0 + 16);
-                __m128 _val0 = __lsx_vreplfr2vr_s(*tmpptr++);
-                __m128 _w0 = (__m128)__lsx_vld(kptr0, 0);
-                _sum = __lsx_vfmadd_s(_w0, _val0, _sum);
-
-                kptr0 += 4;
-            }
-
-            __lsx_vst(_sum, outptr0, 0);
-
-            outptr0 += 4;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_pack4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 4u * 4, 4, opt.workspace_allocator);
-    {
-        const int gap = (w * stride_h - outw * stride_w) * 4;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            float* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const float* sptr = img.row<const float>(dilation_h * u) + dilation_w * v * 4;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            __m128 _val = (__m128)__lsx_vld(sptr, 0);
-                            __lsx_vst(_val, ptr, 0);
-
-                            sptr += stride_w * 4;
-                            ptr += 4;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_pack4_lsx(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack4to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack4to1.h
deleted file mode 100644
index 3748645b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack4to1.h
+++ /dev/null
@@ -1,667 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_pack4to1_lsx(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    // Mat bottom_im2col(size, maxk, inch, 4u * 4, 4, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    Mat tmp;
-    if (size >= 12)
-        tmp.create(12 * maxk, inch, size / 12 + (size % 12) / 8 + (size % 12 % 8) / 4 + size % 12 % 4, 4u * 4, 4, opt.workspace_allocator);
-    else if (size >= 8)
-        tmp.create(8 * maxk, inch, size / 8 + (size % 8) / 4 + size % 4, 4u * 4, 4, opt.workspace_allocator);
-    else if (size >= 4)
-        tmp.create(4 * maxk, inch, size / 4 + size % 4, 4u * 4, 4, opt.workspace_allocator);
-    else
-        tmp.create(maxk, inch, size, 4u * 4, 4, opt.workspace_allocator);
-    {
-        int remain_size_start = 0;
-        int nn_size = size / 12;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 12;
-
-            float* tmpptr = tmp.channel(i / 12);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    // transpose 4x12
-                    __m128i _r0 = __lsx_vld(img0, 0);
-                    __m128i _r1 = __lsx_vld(img0 + 4, 0);
-                    __m128i _r2 = __lsx_vld(img0 + 4 * 2, 0);
-                    __m128i _r3 = __lsx_vld(img0 + 4 * 3, 0);
-                    __m128i _r4 = __lsx_vld(img0 + 4 * 4, 0);
-                    __m128i _r5 = __lsx_vld(img0 + 4 * 5, 0);
-                    __m128i _r6 = __lsx_vld(img0 + 4 * 6, 0);
-                    __m128i _r7 = __lsx_vld(img0 + 4 * 7, 0);
-                    __m128i _r8 = __lsx_vld(img0 + 4 * 8, 0);
-                    __m128i _r9 = __lsx_vld(img0 + 4 * 9, 0);
-                    __m128i _ra = __lsx_vld(img0 + 4 * 10, 0);
-                    __m128i _rb = __lsx_vld(img0 + 4 * 11, 0);
-
-                    __m128i _r01r = __lsx_vilvl_w(_r1, _r0);
-                    __m128i _r01l = __lsx_vilvh_w(_r1, _r0);
-                    __m128i _r23r = __lsx_vilvl_w(_r3, _r2);
-                    __m128i _r23l = __lsx_vilvh_w(_r3, _r2);
-                    __m128i _r45r = __lsx_vilvl_w(_r5, _r4);
-                    __m128i _r45l = __lsx_vilvh_w(_r5, _r4);
-                    __m128i _r67r = __lsx_vilvl_w(_r7, _r6);
-                    __m128i _r67l = __lsx_vilvh_w(_r7, _r6);
-                    __m128i _r89r = __lsx_vilvl_w(_r9, _r8);
-                    __m128i _r89l = __lsx_vilvh_w(_r9, _r8);
-                    __m128i _rabr = __lsx_vilvl_w(_rb, _ra);
-                    __m128i _rabl = __lsx_vilvh_w(_rb, _ra);
-                    __m128i _r0123_0 = __lsx_vilvl_d(_r23r, _r01r);
-                    __m128i _r0123_1 = __lsx_vilvh_d(_r23r, _r01r);
-                    __m128i _r0123_2 = __lsx_vilvl_d(_r23l, _r01l);
-                    __m128i _r0123_3 = __lsx_vilvh_d(_r23l, _r01l);
-                    __m128i _r4567_0 = __lsx_vilvl_d(_r67r, _r45r);
-                    __m128i _r4567_1 = __lsx_vilvh_d(_r67r, _r45r);
-                    __m128i _r4567_2 = __lsx_vilvl_d(_r67l, _r45l);
-                    __m128i _r4567_3 = __lsx_vilvh_d(_r67l, _r45l);
-                    __m128i _r89ab_0 = __lsx_vilvl_d(_rabr, _r89r);
-                    __m128i _r89ab_1 = __lsx_vilvh_d(_rabr, _r89r);
-                    __m128i _r89ab_2 = __lsx_vilvl_d(_rabl, _r89l);
-                    __m128i _r89ab_3 = __lsx_vilvh_d(_rabl, _r89l);
-
-                    __lsx_vst(_r0123_0, tmpptr, 0);
-                    __lsx_vst(_r4567_0, tmpptr + 4, 0);
-                    __lsx_vst(_r89ab_0, tmpptr + 4 * 2, 0);
-                    __lsx_vst(_r0123_1, tmpptr + 4 * 3, 0);
-                    __lsx_vst(_r4567_1, tmpptr + 4 * 4, 0);
-                    __lsx_vst(_r89ab_1, tmpptr + 4 * 5, 0);
-                    __lsx_vst(_r0123_2, tmpptr + 4 * 6, 0);
-                    __lsx_vst(_r4567_2, tmpptr + 4 * 7, 0);
-                    __lsx_vst(_r89ab_2, tmpptr + 4 * 8, 0);
-                    __lsx_vst(_r0123_3, tmpptr + 4 * 9, 0);
-                    __lsx_vst(_r4567_3, tmpptr + 4 * 10, 0);
-                    __lsx_vst(_r89ab_3, tmpptr + 4 * 11, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 48;
-                }
-            }
-        }
-
-        remain_size_start += nn_size * 12;
-        nn_size = (size - remain_size_start) >> 3;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 8;
-
-            float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    // transpose 4x8
-                    __m128i _r0 = __lsx_vld(img0, 0);
-                    __m128i _r1 = __lsx_vld(img0 + 4, 0);
-                    __m128i _r2 = __lsx_vld(img0 + 4 * 2, 0);
-                    __m128i _r3 = __lsx_vld(img0 + 4 * 3, 0);
-                    __m128i _r4 = __lsx_vld(img0 + 4 * 4, 0);
-                    __m128i _r5 = __lsx_vld(img0 + 4 * 5, 0);
-                    __m128i _r6 = __lsx_vld(img0 + 4 * 6, 0);
-                    __m128i _r7 = __lsx_vld(img0 + 4 * 7, 0);
-
-                    __m128i _r01r = __lsx_vilvl_w(_r1, _r0);
-                    __m128i _r01l = __lsx_vilvh_w(_r1, _r0);
-                    __m128i _r23r = __lsx_vilvl_w(_r3, _r2);
-                    __m128i _r23l = __lsx_vilvh_w(_r3, _r2);
-                    __m128i _r45r = __lsx_vilvl_w(_r5, _r4);
-                    __m128i _r45l = __lsx_vilvh_w(_r5, _r4);
-                    __m128i _r67r = __lsx_vilvl_w(_r7, _r6);
-                    __m128i _r67l = __lsx_vilvh_w(_r7, _r6);
-                    __m128i _r0123_0 = __lsx_vilvl_d(_r23r, _r01r);
-                    __m128i _r0123_1 = __lsx_vilvh_d(_r23r, _r01r);
-                    __m128i _r0123_2 = __lsx_vilvl_d(_r23l, _r01l);
-                    __m128i _r0123_3 = __lsx_vilvh_d(_r23l, _r01l);
-                    __m128i _r4567_0 = __lsx_vilvl_d(_r67r, _r45r);
-                    __m128i _r4567_1 = __lsx_vilvh_d(_r67r, _r45r);
-                    __m128i _r4567_2 = __lsx_vilvl_d(_r67l, _r45l);
-                    __m128i _r4567_3 = __lsx_vilvh_d(_r67l, _r45l);
-
-                    __lsx_vst(_r0123_0, tmpptr, 0);
-                    __lsx_vst(_r4567_0, tmpptr + 4, 0);
-                    __lsx_vst(_r0123_1, tmpptr + 4 * 2, 0);
-                    __lsx_vst(_r4567_1, tmpptr + 4 * 3, 0);
-                    __lsx_vst(_r0123_2, tmpptr + 4 * 4, 0);
-                    __lsx_vst(_r4567_2, tmpptr + 4 * 5, 0);
-                    __lsx_vst(_r0123_3, tmpptr + 4 * 6, 0);
-                    __lsx_vst(_r4567_3, tmpptr + 4 * 7, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 32;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 3;
-        nn_size = (size - remain_size_start) >> 2;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 4;
-
-            float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    // transpose 4x4
-                    __m128i _r0 = __lsx_vld(img0, 0);
-                    __m128i _r1 = __lsx_vld(img0 + 4, 0);
-                    __m128i _r2 = __lsx_vld(img0 + 4 * 2, 0);
-                    __m128i _r3 = __lsx_vld(img0 + 4 * 3, 0);
-
-                    __m128i _r01r = __lsx_vilvl_w(_r1, _r0);
-                    __m128i _r01l = __lsx_vilvh_w(_r1, _r0);
-                    __m128i _r23r = __lsx_vilvl_w(_r3, _r2);
-                    __m128i _r23l = __lsx_vilvh_w(_r3, _r2);
-                    __m128i _r0123_0 = __lsx_vilvl_d(_r23r, _r01r);
-                    __m128i _r0123_1 = __lsx_vilvh_d(_r23r, _r01r);
-                    __m128i _r0123_2 = __lsx_vilvl_d(_r23l, _r01l);
-                    __m128i _r0123_3 = __lsx_vilvh_d(_r23l, _r01l);
-
-                    __lsx_vst(_r0123_0, tmpptr, 0);
-                    __lsx_vst(_r0123_1, tmpptr + 4, 0);
-                    __lsx_vst(_r0123_2, tmpptr + 4 * 2, 0);
-                    __lsx_vst(_r0123_3, tmpptr + 4 * 3, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 16;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 2;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + i % 12 % 4);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    __m128 _val = (__m128)__lsx_vld(img0, 0);
-                    __lsx_vst(_val, tmpptr, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 4;
-                }
-            }
-        }
-    }
-
-    int nn_outch = outch / 4;
-    int remain_outch_start = nn_outch * 4;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 4;
-
-        float* outptr0 = top_blob.channel(p);
-        float* outptr1 = top_blob.channel(p + 1);
-        float* outptr2 = top_blob.channel(p + 2);
-        float* outptr3 = top_blob.channel(p + 3);
-
-        const float zeros[4] = {0.f};
-        const float* biasptr = bias ? bias + p : zeros;
-
-        int i = 0;
-        for (; i + 11 < size; i += 12)
-        {
-            const float* tmpptr = tmp.channel(i / 12);
-            const float* kptr0 = kernel.channel(p / 4);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            __m128i _bias = __lsx_vld(biasptr, 0);
-            __m128 _sum0 = (__m128)__lsx_vreplvei_w(_bias, 0);
-            __m128 _sum1 = (__m128)__lsx_vreplvei_w(_bias, 0);
-            __m128 _sum2 = (__m128)__lsx_vreplvei_w(_bias, 0);
-            __m128 _sum3 = (__m128)__lsx_vreplvei_w(_bias, 1);
-            __m128 _sum4 = (__m128)__lsx_vreplvei_w(_bias, 1);
-            __m128 _sum5 = (__m128)__lsx_vreplvei_w(_bias, 1);
-            __m128 _sum6 = (__m128)__lsx_vreplvei_w(_bias, 2);
-            __m128 _sum7 = (__m128)__lsx_vreplvei_w(_bias, 2);
-            __m128 _sum8 = (__m128)__lsx_vreplvei_w(_bias, 2);
-            __m128 _sum9 = (__m128)__lsx_vreplvei_w(_bias, 3);
-            __m128 _suma = (__m128)__lsx_vreplvei_w(_bias, 3);
-            __m128 _sumb = (__m128)__lsx_vreplvei_w(_bias, 3);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 48);
-                __builtin_prefetch(kptr0 + 16);
-                __m128 _val0 = (__m128)__lsx_vld(tmpptr, 0);
-                __m128 _val1 = (__m128)__lsx_vld(tmpptr + 4, 0);
-                __m128 _val2 = (__m128)__lsx_vld(tmpptr + 8, 0);
-                __m128i _w0123 = __lsx_vld(kptr0, 0);
-                _sum0 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 0), _val0, _sum0);
-                _sum1 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 0), _val1, _sum1);
-                _sum2 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 0), _val2, _sum2);
-                _sum3 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 1), _val0, _sum3);
-                _sum4 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 1), _val1, _sum4);
-                _sum5 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 1), _val2, _sum5);
-                _sum6 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 2), _val0, _sum6);
-                _sum7 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 2), _val1, _sum7);
-                _sum8 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 2), _val2, _sum8);
-                _sum9 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 3), _val0, _sum9);
-                _suma = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 3), _val1, _suma);
-                _sumb = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 3), _val2, _sumb);
-
-                tmpptr += 12;
-                kptr0 += 4;
-            }
-
-            __lsx_vst(_sum0, outptr0, 0);
-            __lsx_vst(_sum1, outptr0 + 4, 0);
-            __lsx_vst(_sum2, outptr0 + 8, 0);
-            __lsx_vst(_sum3, outptr1, 0);
-            __lsx_vst(_sum4, outptr1 + 4, 0);
-            __lsx_vst(_sum5, outptr1 + 8, 0);
-            __lsx_vst(_sum6, outptr2, 0);
-            __lsx_vst(_sum7, outptr2 + 4, 0);
-            __lsx_vst(_sum8, outptr2 + 8, 0);
-            __lsx_vst(_sum9, outptr3, 0);
-            __lsx_vst(_suma, outptr3 + 4, 0);
-            __lsx_vst(_sumb, outptr3 + 8, 0);
-
-            outptr0 += 12;
-            outptr1 += 12;
-            outptr2 += 12;
-            outptr3 += 12;
-        }
-        for (; i + 7 < size; i += 8)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8);
-            const float* kptr0 = kernel.channel(p / 4);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            __m128i _bias = __lsx_vld(biasptr, 0);
-            __m128 _sum0 = (__m128)__lsx_vreplvei_w(_bias, 0);
-            __m128 _sum1 = (__m128)__lsx_vreplvei_w(_bias, 0);
-            __m128 _sum2 = (__m128)__lsx_vreplvei_w(_bias, 1);
-            __m128 _sum3 = (__m128)__lsx_vreplvei_w(_bias, 1);
-            __m128 _sum4 = (__m128)__lsx_vreplvei_w(_bias, 2);
-            __m128 _sum5 = (__m128)__lsx_vreplvei_w(_bias, 2);
-            __m128 _sum6 = (__m128)__lsx_vreplvei_w(_bias, 3);
-            __m128 _sum7 = (__m128)__lsx_vreplvei_w(_bias, 3);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 32);
-                __builtin_prefetch(kptr0 + 16);
-                __m128 _val0 = (__m128)__lsx_vld(tmpptr, 0);
-                __m128 _val1 = (__m128)__lsx_vld(tmpptr + 4, 0);
-                __m128i _w0123 = __lsx_vld(kptr0, 0);
-                _sum0 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 0), _val0, _sum0);
-                _sum1 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 0), _val1, _sum1);
-                _sum2 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 1), _val0, _sum2);
-                _sum3 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 1), _val1, _sum3);
-                _sum4 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 2), _val0, _sum4);
-                _sum5 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 2), _val1, _sum5);
-                _sum6 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 3), _val0, _sum6);
-                _sum7 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 3), _val1, _sum7);
-
-                tmpptr += 8;
-                kptr0 += 4;
-            }
-
-            __lsx_vst(_sum0, outptr0, 0);
-            __lsx_vst(_sum1, outptr0 + 4, 0);
-            __lsx_vst(_sum2, outptr1, 0);
-            __lsx_vst(_sum3, outptr1 + 4, 0);
-            __lsx_vst(_sum4, outptr2, 0);
-            __lsx_vst(_sum5, outptr2 + 4, 0);
-            __lsx_vst(_sum6, outptr3, 0);
-            __lsx_vst(_sum7, outptr3 + 4, 0);
-
-            outptr0 += 8;
-            outptr1 += 8;
-            outptr2 += 8;
-            outptr3 += 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4);
-            const float* kptr0 = kernel.channel(p / 4);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            __m128i _bias = __lsx_vld(biasptr, 0);
-            __m128 _sum0 = (__m128)__lsx_vreplvei_w(_bias, 0);
-            __m128 _sum1 = (__m128)__lsx_vreplvei_w(_bias, 1);
-            __m128 _sum2 = (__m128)__lsx_vreplvei_w(_bias, 2);
-            __m128 _sum3 = (__m128)__lsx_vreplvei_w(_bias, 3);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr0 + 16);
-                __m128 _val0 = (__m128)__lsx_vld(tmpptr, 0);
-                __m128i _w0123 = __lsx_vld(kptr0, 0);
-                _sum0 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 0), _val0, _sum0);
-                _sum1 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 1), _val0, _sum1);
-                _sum2 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 2), _val0, _sum2);
-                _sum3 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 3), _val0, _sum3);
-
-                tmpptr += 4;
-                kptr0 += 4;
-            }
-
-            __lsx_vst(_sum0, outptr0, 0);
-            __lsx_vst(_sum1, outptr1, 0);
-            __lsx_vst(_sum2, outptr2, 0);
-            __lsx_vst(_sum3, outptr3, 0);
-
-            outptr0 += 4;
-            outptr1 += 4;
-            outptr2 += 4;
-            outptr3 += 4;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + i % 12 % 4);
-            const float* kptr0 = kernel.channel(p / 4);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            __m128 _sum = (__m128)__lsx_vld(biasptr, 0);
-            float* _sum_p = (float*)&_sum;
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 4);
-                __builtin_prefetch(kptr0 + 16);
-                __m128 _val0 = __lsx_vreplfr2vr_s(*tmpptr++);
-                __m128 _w0 = (__m128)__lsx_vld(kptr0, 0);
-                _sum = __lsx_vfmadd_s(_w0, _val0, _sum);
-
-                kptr0 += 4;
-            }
-
-            outptr0[0] = _sum_p[0];
-            outptr1[0] = _sum_p[1];
-            outptr2[0] = _sum_p[2];
-            outptr3[0] = _sum_p[3];
-
-            outptr0 += 1;
-            outptr1 += 1;
-            outptr2 += 1;
-            outptr3 += 1;
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        float* outptr0 = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        int i = 0;
-        for (; i + 11 < size; i += 12)
-        {
-            const float* tmpptr = tmp.channel(i / 12);
-            const float* kptr0 = kernel.channel(p / 4 + p % 4);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            __m128 _sum0 = __lsx_vreplfr2vr_s(bias0);
-            __m128 _sum1 = __lsx_vreplfr2vr_s(bias0);
-            __m128 _sum2 = __lsx_vreplfr2vr_s(bias0);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 48);
-                __builtin_prefetch(kptr0 + 4);
-                __m128 _val0 = (__m128)__lsx_vld(tmpptr, 0);
-                __m128 _val1 = (__m128)__lsx_vld(tmpptr + 4, 0);
-                __m128 _val2 = (__m128)__lsx_vld(tmpptr + 8, 0);
-                __m128 _w0 = __lsx_vreplfr2vr_s(*kptr0);
-                _sum0 = __lsx_vfmadd_s(_val0, _w0, _sum0);
-                _sum1 = __lsx_vfmadd_s(_val1, _w0, _sum1);
-                _sum2 = __lsx_vfmadd_s(_val2, _w0, _sum2);
-
-                tmpptr += 12;
-                kptr0 += 1;
-            }
-
-            __lsx_vst(_sum0, outptr0, 0);
-            __lsx_vst(_sum1, outptr0 + 4, 0);
-            __lsx_vst(_sum2, outptr0 + 8, 0);
-
-            outptr0 += 12;
-        }
-        for (; i + 7 < size; i += 8)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8);
-            const float* kptr0 = kernel.channel(p / 4 + p % 4);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            __m128 _sum0 = __lsx_vreplfr2vr_s(bias0);
-            __m128 _sum1 = __lsx_vreplfr2vr_s(bias0);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 32);
-                __builtin_prefetch(kptr0 + 4);
-                __m128 _val0 = (__m128)__lsx_vld(tmpptr, 0);
-                __m128 _val1 = (__m128)__lsx_vld(tmpptr + 4, 0);
-                __m128 _w0 = __lsx_vreplfr2vr_s(*kptr0);
-                _sum0 = __lsx_vfmadd_s(_val0, _w0, _sum0);
-                _sum1 = __lsx_vfmadd_s(_val1, _w0, _sum1);
-
-                tmpptr += 8;
-                kptr0 += 1;
-            }
-
-            __lsx_vst(_sum0, outptr0, 0);
-            __lsx_vst(_sum1, outptr0 + 4, 0);
-
-            outptr0 += 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4);
-            const float* kptr0 = kernel.channel(p / 4 + p % 4);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            __m128 _sum0 = __lsx_vreplfr2vr_s(bias0);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr0 + 4);
-                __m128 _val0 = (__m128)__lsx_vld(tmpptr, 0);
-                __m128 _w0 = __lsx_vreplfr2vr_s(*kptr0);
-                _sum0 = __lsx_vfmadd_s(_val0, _w0, _sum0);
-
-                tmpptr += 4;
-                kptr0 += 1;
-            }
-
-            __lsx_vst(_sum0, outptr0, 0);
-
-            outptr0 += 4;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + i % 12 % 4);
-            const float* kptr0 = kernel.channel(p / 4 + p % 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = bias0;
-
-            __m128 _sum0 = (__m128)__lsx_vreplgr2vr_w(0);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr0 + 16);
-                __m128 _val0 = (__m128)__lsx_vld(tmpptr, 0);
-                __m128 _w0 = (__m128)__lsx_vld(kptr0, 0);
-                _sum0 = __lsx_vfmadd_s(_w0, _val0, _sum0);
-                tmpptr += 4;
-                kptr0 += 4;
-            }
-
-            sum0 += __lsx_reduce_fadd_s(_sum0);
-
-            outptr0[0] = sum0;
-
-            outptr0 += 1;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_transform_kernel_pack4to1_lsx(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = pb-pa-maxk-inch/pa-outch/pb
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-    kernel_tm.create(4 * 4 * maxk, inch / 4, outch / 4 + outch % 4);
-
-    int q = 0;
-    for (; q + 3 < outch; q += 4)
-    {
-        float* g00 = kernel_tm.channel(q / 4);
-
-        for (int p = 0; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 4; j++)
-                    {
-                        const float* k00 = kernel.channel(q + j).row(p + i);
-
-                        g00[0] = k00[k];
-
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-    for (; q < outch; q++)
-    {
-        const Mat k0 = kernel.channel(q);
-
-        float* g00 = kernel_tm.channel(q / 4 + q % 4);
-
-        for (int p = 0; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int j = 0; j < 4; j++)
-                {
-                    const float* k00 = k0.row(p + j);
-
-                    g00[0] = k00[k];
-
-                    g00++;
-                }
-            }
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_pack4to1_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 4u * 4, 4, opt.workspace_allocator);
-    {
-        const int gap = (w * stride_h - outw * stride_w) * 4;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            float* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const float* sptr = img.row(dilation_h * u) + dilation_w * v * 4;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            __m128 _val = (__m128)__lsx_vld(sptr, 0);
-                            __lsx_vst(_val, ptr, 0);
-
-                            sptr += stride_w * 4;
-                            ptr += 4;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_pack4to1_lsx(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack8to1_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack8to1_int8.h
deleted file mode 100644
index 98d11a57..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack8to1_int8.h
+++ /dev/null
@@ -1,458 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_pack8to1_int8_lsx(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    // Mat bottom_im2col(size, maxk, inch, 8u, 8, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    // permute
-    Mat tmp;
-    if (size >= 2)
-        tmp.create(2 * maxk, inch, size / 2 + size % 2, 8u, 8, opt.workspace_allocator);
-    else
-        tmp.create(maxk, inch, size, 8u, 8, opt.workspace_allocator);
-    {
-        int remain_size_start = 0;
-        int nn_size = (size - remain_size_start) >> 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 2;
-
-            int64_t* tmpptr = tmp.channel(i / 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const int64_t* img0 = (const int64_t*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    __m128i _v = __lsx_vld(img0, 0);
-                    __lsx_vst(_v, tmpptr, 0);
-                    tmpptr += 2;
-                    img0 += size;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            int64_t* tmpptr = tmp.channel(i / 2 + i % 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const int64_t* img0 = (const int64_t*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr += 1;
-                    img0 += size;
-                }
-            }
-        }
-    }
-
-    int nn_outch = 0;
-    int remain_outch_start = 0;
-
-    nn_outch = outch >> 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 4;
-
-        int* outptr0 = top_blob.channel(p);
-        int* outptr1 = top_blob.channel(p + 1);
-        int* outptr2 = top_blob.channel(p + 2);
-        int* outptr3 = top_blob.channel(p + 3);
-
-        int i = 0;
-        for (; i + 1 < size; i += 2)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2);
-            const signed char* kptr = kernel.channel(p / 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __m128i _sum00 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum01 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum02 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum03 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum10 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum11 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum12 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum13 = __lsx_vreplgr2vr_w(0);
-
-            int j = 0;
-            for (; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 64);
-                __builtin_prefetch(kptr + 128);
-                __m128i _val01 = __lsx_vld(tmpptr, 0);
-                __m128i _extval01 = __lsx_vslti_b(_val01, 0);
-                __m128i _val0 = __lsx_vilvl_b(_extval01, _val01);
-                __m128i _val1 = __lsx_vilvh_b(_extval01, _val01);
-
-                __m128i _w01 = __lsx_vld(kptr, 0);
-                __m128i _w23 = __lsx_vld(kptr + 16, 0);
-                __m128i _extw01 = __lsx_vslti_b(_w01, 0);
-                __m128i _extw23 = __lsx_vslti_b(_w23, 0);
-                __m128i _w0 = __lsx_vilvl_b(_extw01, _w01);
-                __m128i _w1 = __lsx_vilvh_b(_extw01, _w01);
-                __m128i _w2 = __lsx_vilvl_b(_extw23, _w23);
-                __m128i _w3 = __lsx_vilvh_b(_extw23, _w23);
-
-                __m128i _s00 = __lsx_vmul_h(_val0, _w0);
-                __m128i _s01 = __lsx_vmul_h(_val0, _w1);
-                __m128i _s02 = __lsx_vmul_h(_val0, _w2);
-                __m128i _s03 = __lsx_vmul_h(_val0, _w3);
-                __m128i _s10 = __lsx_vmul_h(_val1, _w0);
-                __m128i _s11 = __lsx_vmul_h(_val1, _w1);
-                __m128i _s12 = __lsx_vmul_h(_val1, _w2);
-                __m128i _s13 = __lsx_vmul_h(_val1, _w3);
-
-                _sum00 = __lsx_vadd_w(_sum00, __lsx_vhaddw_w_h(_s00, _s00));
-                _sum01 = __lsx_vadd_w(_sum01, __lsx_vhaddw_w_h(_s01, _s01));
-                _sum02 = __lsx_vadd_w(_sum02, __lsx_vhaddw_w_h(_s02, _s02));
-                _sum03 = __lsx_vadd_w(_sum03, __lsx_vhaddw_w_h(_s03, _s03));
-                _sum10 = __lsx_vadd_w(_sum10, __lsx_vhaddw_w_h(_s10, _s10));
-                _sum11 = __lsx_vadd_w(_sum11, __lsx_vhaddw_w_h(_s11, _s11));
-                _sum12 = __lsx_vadd_w(_sum12, __lsx_vhaddw_w_h(_s12, _s12));
-                _sum13 = __lsx_vadd_w(_sum13, __lsx_vhaddw_w_h(_s13, _s13));
-
-                tmpptr += 16;
-                kptr += 32;
-            }
-
-            // transpose 4x4
-            {
-                __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                _tmp0 = __lsx_vilvl_w(_sum01, _sum00);
-                _tmp1 = __lsx_vilvl_w(_sum03, _sum02);
-                _tmp2 = __lsx_vilvh_w(_sum01, _sum00);
-                _tmp3 = __lsx_vilvh_w(_sum03, _sum02);
-                _sum00 = __lsx_vilvl_d(_tmp1, _tmp0);
-                _sum01 = __lsx_vilvh_d(_tmp1, _tmp0);
-                _sum02 = __lsx_vilvl_d(_tmp3, _tmp2);
-                _sum03 = __lsx_vilvh_d(_tmp3, _tmp2);
-            }
-            {
-                __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                _tmp0 = __lsx_vilvl_w(_sum11, _sum10);
-                _tmp1 = __lsx_vilvl_w(_sum13, _sum12);
-                _tmp2 = __lsx_vilvh_w(_sum11, _sum10);
-                _tmp3 = __lsx_vilvh_w(_sum13, _sum12);
-                _sum10 = __lsx_vilvl_d(_tmp1, _tmp0);
-                _sum11 = __lsx_vilvh_d(_tmp1, _tmp0);
-                _sum12 = __lsx_vilvl_d(_tmp3, _tmp2);
-                _sum13 = __lsx_vilvh_d(_tmp3, _tmp2);
-            }
-
-            _sum00 = __lsx_vadd_w(_sum00, _sum01);
-            _sum02 = __lsx_vadd_w(_sum02, _sum03);
-            _sum10 = __lsx_vadd_w(_sum10, _sum11);
-            _sum12 = __lsx_vadd_w(_sum12, _sum13);
-
-            _sum00 = __lsx_vadd_w(_sum00, _sum02);
-            _sum10 = __lsx_vadd_w(_sum10, _sum12);
-
-            int sum[8];
-            __lsx_vst(_sum00, sum, 0);
-            __lsx_vst(_sum10, sum + 4, 0);
-
-            outptr0[0] = sum[0];
-            outptr1[0] = sum[1];
-            outptr2[0] = sum[2];
-            outptr3[0] = sum[3];
-            outptr0[1] = sum[4];
-            outptr1[1] = sum[5];
-            outptr2[1] = sum[6];
-            outptr3[1] = sum[7];
-            outptr0 += 2;
-            outptr1 += 2;
-            outptr2 += 2;
-            outptr3 += 2;
-        }
-        for (; i < size; i++)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-            const signed char* kptr = kernel.channel(p / 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __m128i _sum0 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum1 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum2 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum3 = __lsx_vreplgr2vr_w(0);
-
-            int j = 0;
-            for (; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 32);
-                __builtin_prefetch(kptr + 128);
-                __m128i _val = __lsx_vld(tmpptr, 0);
-                __m128i _val16 = __lsx_vilvl_b(__lsx_vslti_b(_val, 0), _val);
-
-                __m128i _w01 = __lsx_vld(kptr, 0);
-                __m128i _w23 = __lsx_vld(kptr + 16, 0);
-                __m128i _extw01 = __lsx_vslti_b(_w01, 0);
-                __m128i _extw23 = __lsx_vslti_b(_w23, 0);
-                __m128i _w0 = __lsx_vilvl_b(_extw01, _w01);
-                __m128i _w1 = __lsx_vilvh_b(_extw01, _w01);
-                __m128i _w2 = __lsx_vilvl_b(_extw23, _w23);
-                __m128i _w3 = __lsx_vilvh_b(_extw23, _w23);
-
-                __m128i _s0 = __lsx_vmul_h(_val16, _w0);
-                __m128i _s1 = __lsx_vmul_h(_val16, _w1);
-                __m128i _s2 = __lsx_vmul_h(_val16, _w2);
-                __m128i _s3 = __lsx_vmul_h(_val16, _w3);
-
-                _sum0 = __lsx_vadd_w(_sum0, __lsx_vhaddw_w_h(_s0, _s0));
-                _sum1 = __lsx_vadd_w(_sum1, __lsx_vhaddw_w_h(_s1, _s1));
-                _sum2 = __lsx_vadd_w(_sum2, __lsx_vhaddw_w_h(_s2, _s2));
-                _sum3 = __lsx_vadd_w(_sum3, __lsx_vhaddw_w_h(_s3, _s3));
-
-                tmpptr += 8;
-                kptr += 32;
-            }
-
-            // transpose 4x4
-            {
-                __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                _tmp0 = __lsx_vilvl_w(_sum1, _sum0);
-                _tmp1 = __lsx_vilvl_w(_sum3, _sum2);
-                _tmp2 = __lsx_vilvh_w(_sum1, _sum0);
-                _tmp3 = __lsx_vilvh_w(_sum3, _sum2);
-                _sum0 = __lsx_vilvl_d(_tmp1, _tmp0);
-                _sum1 = __lsx_vilvh_d(_tmp1, _tmp0);
-                _sum2 = __lsx_vilvl_d(_tmp3, _tmp2);
-                _sum3 = __lsx_vilvh_d(_tmp3, _tmp2);
-            }
-
-            _sum0 = __lsx_vadd_w(_sum0, _sum1);
-            _sum2 = __lsx_vadd_w(_sum2, _sum3);
-
-            _sum0 = __lsx_vadd_w(_sum0, _sum2);
-
-            int sum[4];
-            __lsx_vst(_sum0, sum, 0);
-
-            outptr0[0] = sum[0];
-            outptr1[0] = sum[1];
-            outptr2[0] = sum[2];
-            outptr3[0] = sum[3];
-            outptr0 += 1;
-            outptr1 += 1;
-            outptr2 += 1;
-            outptr3 += 1;
-        }
-    }
-
-    remain_outch_start += nn_outch << 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        int* outptr0 = top_blob.channel(p);
-
-        int i = 0;
-        for (; i + 1 < size; i += 2)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2);
-            const signed char* kptr = kernel.channel(p / 4 + p % 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __m128i _sum0 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum1 = __lsx_vreplgr2vr_w(0);
-
-            int j = 0;
-            for (; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 64);
-                __builtin_prefetch(kptr + 32);
-                __m128i _val01 = __lsx_vld(tmpptr, 0);
-                __m128i _extval01 = __lsx_vslti_b(_val01, 0);
-                __m128i _val0 = __lsx_vilvl_b(_extval01, _val01);
-                __m128i _val1 = __lsx_vilvh_b(_extval01, _val01);
-
-                __m128i _w = __lsx_vld(kptr, 0);
-                __m128i _w16 = __lsx_vilvl_b(__lsx_vslti_b(_w, 0), _w);
-
-                __m128i _s0 = __lsx_vmul_h(_val0, _w16);
-                __m128i _s1 = __lsx_vmul_h(_val1, _w16);
-
-                _sum0 = __lsx_vadd_w(_sum0, __lsx_vhaddw_w_h(_s0, _s0));
-                _sum1 = __lsx_vadd_w(_sum1, __lsx_vhaddw_w_h(_s1, _s1));
-
-                tmpptr += 16;
-                kptr += 8;
-            }
-
-            outptr0[0] = __lsx_reduce_add_w(_sum0);
-            outptr0[1] = __lsx_reduce_add_w(_sum1);
-            outptr0 += 2;
-        }
-        for (; i < size; i++)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-            const signed char* kptr = kernel.channel(p / 4 + p % 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __m128i _sum = __lsx_vreplgr2vr_w(0);
-
-            int j = 0;
-            for (; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 32);
-                __builtin_prefetch(kptr + 32);
-                __m128i _val = __lsx_vld(tmpptr, 0);
-                __m128i _val16 = __lsx_vilvl_b(__lsx_vslti_b(_val, 0), _val);
-
-                __m128i _w = __lsx_vld(kptr, 0);
-                __m128i _w16 = __lsx_vilvl_b(__lsx_vslti_b(_w, 0), _w);
-
-                __m128i _s0 = __lsx_vmul_h(_val16, _w16);
-
-                _sum = __lsx_vadd_w(_sum, __lsx_vhaddw_w_h(_s0, _s0));
-
-                tmpptr += 8;
-                kptr += 8;
-            }
-
-            outptr0[0] = __lsx_reduce_add_w(_sum);
-            outptr0 += 1;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_transform_kernel_pack8to1_int8_lsx(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = 8a-4b-maxk-inch/8a-outch/4b
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-    if (outch >= 4)
-        kernel_tm.create(32 * maxk, inch / 8, outch / 4 + outch % 4, (size_t)1u);
-    else
-        kernel_tm.create(8 * maxk, inch / 8, outch, (size_t)1u);
-
-    int q = 0;
-    for (; q + 3 < outch; q += 4)
-    {
-        signed char* g00 = kernel_tm.channel(q / 4);
-
-        for (int p = 0; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 8; j++)
-                    {
-                        const signed char* k00 = kernel.channel(q + i).row<const signed char>(p + j);
-
-                        g00[0] = k00[k];
-
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-    // TODO unroll 2
-    for (; q < outch; q++)
-    {
-        signed char* g00 = kernel_tm.channel(q / 4 + q % 4);
-
-        for (int p = 0; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int j = 0; j < 8; j++)
-                {
-                    const signed char* k00 = kernel.channel(q).row<const signed char>(p + j);
-
-                    g00[0] = k00[k];
-
-                    g00++;
-                }
-            }
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_pack8to1_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 8u, 8, opt.workspace_allocator);
-    {
-        const int gap = w * stride_h - outw * stride_w;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            int64_t* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const int64_t* sptr = img.row<const int64_t>(dilation_h * u) + dilation_w * v;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            ptr[0] = sptr[0];
-
-                            sptr += stride_w;
-                            ptr += 1;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_pack8to1_int8_lsx(bottom_im2col, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack8to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack8to4_int8.h
deleted file mode 100644
index ae9090c9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_sgemm_pack8to4_int8.h
+++ /dev/null
@@ -1,324 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_pack8to4_int8_lsx(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    // Mat bottom_im2col(size, maxk, inch, 8u, 8, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    // permute
-    Mat tmp;
-    if (size >= 2)
-        tmp.create(2 * maxk, inch, size / 2 + size % 2, 8u, 8, opt.workspace_allocator);
-    else
-        tmp.create(maxk, inch, size, 8u, 8, opt.workspace_allocator);
-    {
-        int remain_size_start = 0;
-        int nn_size = size >> 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 2;
-
-            int64_t* tmpptr = tmp.channel(i / 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const int64_t* img0 = (const int64_t*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    __m128i _v = __lsx_vld(img0, 0);
-                    __lsx_vst(_v, tmpptr, 0);
-                    tmpptr += 2;
-                    img0 += size;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            int64_t* tmpptr = tmp.channel(i / 2 + i % 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const int64_t* img0 = (const int64_t*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr += 1;
-                    img0 += size;
-                }
-            }
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        int* outptr0 = top_blob.channel(p);
-
-        int i = 0;
-        for (; i + 1 < size; i += 2)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2);
-            const signed char* kptr = kernel.channel(p);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __m128i _sum00 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum01 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum02 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum03 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum10 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum11 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum12 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum13 = __lsx_vreplgr2vr_w(0);
-
-            int j = 0;
-            for (; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 64);
-                __builtin_prefetch(kptr + 128);
-                __m128i _val01 = __lsx_vld(tmpptr, 0);
-                __m128i _extval01 = __lsx_vslti_b(_val01, 0);
-                __m128i _val0 = __lsx_vilvl_b(_extval01, _val01);
-                __m128i _val1 = __lsx_vilvh_b(_extval01, _val01);
-
-                __m128i _w01 = __lsx_vld(kptr, 0);
-                __m128i _w23 = __lsx_vld(kptr + 16, 0);
-                __m128i _extw01 = __lsx_vslti_b(_w01, 0);
-                __m128i _extw23 = __lsx_vslti_b(_w23, 0);
-                __m128i _w0 = __lsx_vilvl_b(_extw01, _w01);
-                __m128i _w1 = __lsx_vilvh_b(_extw01, _w01);
-                __m128i _w2 = __lsx_vilvl_b(_extw23, _w23);
-                __m128i _w3 = __lsx_vilvh_b(_extw23, _w23);
-
-                __m128i _s00 = __lsx_vmul_h(_val0, _w0);
-                __m128i _s01 = __lsx_vmul_h(_val0, _w1);
-                __m128i _s02 = __lsx_vmul_h(_val0, _w2);
-                __m128i _s03 = __lsx_vmul_h(_val0, _w3);
-                __m128i _s10 = __lsx_vmul_h(_val1, _w0);
-                __m128i _s11 = __lsx_vmul_h(_val1, _w1);
-                __m128i _s12 = __lsx_vmul_h(_val1, _w2);
-                __m128i _s13 = __lsx_vmul_h(_val1, _w3);
-
-                _sum00 = __lsx_vadd_w(_sum00, __lsx_vhaddw_w_h(_s00, _s00));
-                _sum01 = __lsx_vadd_w(_sum01, __lsx_vhaddw_w_h(_s01, _s01));
-                _sum02 = __lsx_vadd_w(_sum02, __lsx_vhaddw_w_h(_s02, _s02));
-                _sum03 = __lsx_vadd_w(_sum03, __lsx_vhaddw_w_h(_s03, _s03));
-                _sum10 = __lsx_vadd_w(_sum10, __lsx_vhaddw_w_h(_s10, _s10));
-                _sum11 = __lsx_vadd_w(_sum11, __lsx_vhaddw_w_h(_s11, _s11));
-                _sum12 = __lsx_vadd_w(_sum12, __lsx_vhaddw_w_h(_s12, _s12));
-                _sum13 = __lsx_vadd_w(_sum13, __lsx_vhaddw_w_h(_s13, _s13));
-
-                tmpptr += 16;
-                kptr += 32;
-            }
-
-            // transpose 4x4
-            {
-                __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                _tmp0 = __lsx_vilvl_w(_sum01, _sum00);
-                _tmp1 = __lsx_vilvl_w(_sum03, _sum02);
-                _tmp2 = __lsx_vilvh_w(_sum01, _sum00);
-                _tmp3 = __lsx_vilvh_w(_sum03, _sum02);
-                _sum00 = __lsx_vilvl_d(_tmp1, _tmp0);
-                _sum01 = __lsx_vilvh_d(_tmp1, _tmp0);
-                _sum02 = __lsx_vilvl_d(_tmp3, _tmp2);
-                _sum03 = __lsx_vilvh_d(_tmp3, _tmp2);
-            }
-            {
-                __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                _tmp0 = __lsx_vilvl_w(_sum11, _sum10);
-                _tmp1 = __lsx_vilvl_w(_sum13, _sum12);
-                _tmp2 = __lsx_vilvh_w(_sum11, _sum10);
-                _tmp3 = __lsx_vilvh_w(_sum13, _sum12);
-                _sum10 = __lsx_vilvl_d(_tmp1, _tmp0);
-                _sum11 = __lsx_vilvh_d(_tmp1, _tmp0);
-                _sum12 = __lsx_vilvl_d(_tmp3, _tmp2);
-                _sum13 = __lsx_vilvh_d(_tmp3, _tmp2);
-            }
-
-            _sum00 = __lsx_vadd_w(_sum00, _sum01);
-            _sum02 = __lsx_vadd_w(_sum02, _sum03);
-            _sum10 = __lsx_vadd_w(_sum10, _sum11);
-            _sum12 = __lsx_vadd_w(_sum12, _sum13);
-
-            _sum00 = __lsx_vadd_w(_sum00, _sum02);
-            _sum10 = __lsx_vadd_w(_sum10, _sum12);
-
-            __lsx_vst(_sum00, outptr0, 0);
-            __lsx_vst(_sum10, outptr0 + 4, 0);
-            outptr0 += 8;
-        }
-        for (; i < size; i++)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-            const signed char* kptr = kernel.channel(p);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __m128i _sum0 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum1 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum2 = __lsx_vreplgr2vr_w(0);
-            __m128i _sum3 = __lsx_vreplgr2vr_w(0);
-
-            int j = 0;
-            for (; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 32);
-                __builtin_prefetch(kptr + 128);
-                __m128i _val = __lsx_vld(tmpptr, 0);
-                __m128i _val16 = __lsx_vilvl_b(__lsx_vslti_b(_val, 0), _val);
-
-                __m128i _w01 = __lsx_vld(kptr, 0);
-                __m128i _w23 = __lsx_vld(kptr + 16, 0);
-                __m128i _extw01 = __lsx_vslti_b(_w01, 0);
-                __m128i _extw23 = __lsx_vslti_b(_w23, 0);
-                __m128i _w0 = __lsx_vilvl_b(_extw01, _w01);
-                __m128i _w1 = __lsx_vilvh_b(_extw01, _w01);
-                __m128i _w2 = __lsx_vilvl_b(_extw23, _w23);
-                __m128i _w3 = __lsx_vilvh_b(_extw23, _w23);
-
-                __m128i _s0 = __lsx_vmul_h(_val16, _w0);
-                __m128i _s1 = __lsx_vmul_h(_val16, _w1);
-                __m128i _s2 = __lsx_vmul_h(_val16, _w2);
-                __m128i _s3 = __lsx_vmul_h(_val16, _w3);
-
-                _sum0 = __lsx_vadd_w(_sum0, __lsx_vhaddw_w_h(_s0, _s0));
-                _sum1 = __lsx_vadd_w(_sum1, __lsx_vhaddw_w_h(_s1, _s1));
-                _sum2 = __lsx_vadd_w(_sum2, __lsx_vhaddw_w_h(_s2, _s2));
-                _sum3 = __lsx_vadd_w(_sum3, __lsx_vhaddw_w_h(_s3, _s3));
-
-                tmpptr += 8;
-                kptr += 32;
-            }
-
-            // transpose 4x4
-            {
-                __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                _tmp0 = __lsx_vilvl_w(_sum1, _sum0);
-                _tmp1 = __lsx_vilvl_w(_sum3, _sum2);
-                _tmp2 = __lsx_vilvh_w(_sum1, _sum0);
-                _tmp3 = __lsx_vilvh_w(_sum3, _sum2);
-                _sum0 = __lsx_vilvl_d(_tmp1, _tmp0);
-                _sum1 = __lsx_vilvh_d(_tmp1, _tmp0);
-                _sum2 = __lsx_vilvl_d(_tmp3, _tmp2);
-                _sum3 = __lsx_vilvh_d(_tmp3, _tmp2);
-            }
-
-            _sum0 = __lsx_vadd_w(_sum0, _sum1);
-            _sum2 = __lsx_vadd_w(_sum2, _sum3);
-
-            _sum0 = __lsx_vadd_w(_sum0, _sum2);
-
-            __lsx_vst(_sum0, outptr0, 0);
-            outptr0 += 4;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_transform_kernel_pack8to4_int8_lsx(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = 8a-4b-maxk-inch/8a-outch/4b
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-    kernel_tm.create(32 * maxk, inch / 8, outch / 4, (size_t)1u);
-
-    for (int q = 0; q + 3 < outch; q += 4)
-    {
-        signed char* g00 = kernel_tm.channel(q / 4);
-
-        for (int p = 0; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 8; j++)
-                    {
-                        const signed char* k00 = kernel.channel(q + i).row<const signed char>(p + j);
-
-                        g00[0] = k00[k];
-
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_pack8to4_int8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 8u, 8, opt.workspace_allocator);
-    {
-        const int gap = w * stride_h - outw * stride_w;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            int64_t* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const int64_t* sptr = img.row<const int64_t>(dilation_h * u) + dilation_w * v;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            ptr[0] = sptr[0];
-
-                            sptr += stride_w;
-                            ptr += 1;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_pack8to4_int8_lsx(bottom_im2col, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot.h
deleted file mode 100644
index 9dbbe495..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot.h
+++ /dev/null
@@ -1,495 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_winograd_dot_lsx(Mat& bottom_blob_tm, int outch, const Mat& kernel_tm, Mat& top_blob_tm, const Option& opt)
-{
-    // Mat bottom_blob_tm(tiles, 16/36/64, inch, 4u, opt.workspace_allocator);
-
-    const int tiles = bottom_blob_tm.w;
-    const int batch = bottom_blob_tm.h;
-    const int inch = bottom_blob_tm.c;
-
-    // permute
-    Mat bottom_blob_tm2;
-    if (tiles >= 4)
-        bottom_blob_tm2.create(4 * inch, tiles / 4 + tiles % 4, batch, 4u, opt.workspace_allocator);
-    else
-        bottom_blob_tm2.create(1 * inch, tiles, batch, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int r = 0; r < batch; r++)
-    {
-        Mat tm2 = bottom_blob_tm2.channel(r);
-
-        // tile
-        int i = 0;
-        for (; i + 3 < tiles; i += 4)
-        {
-            float* tmpptr = tm2.row(i / 4);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i);
-
-            for (int q = 0; q < inch; q++)
-            {
-#if __loongarch_sx
-                __lsx_vst(__lsx_vld(r0, 0), tmpptr, 0);
-#else
-                tmpptr[0] = r0[0];
-                tmpptr[1] = r0[1];
-                tmpptr[2] = r0[2];
-                tmpptr[3] = r0[3];
-#endif
-
-                r0 += bottom_blob_tm.cstep;
-                tmpptr += 4;
-            }
-        }
-        for (; i < tiles; i++)
-        {
-            float* tmpptr = tm2.row(i / 4 + i % 4);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i);
-
-            for (int q = 0; q < inch; q++)
-            {
-                tmpptr[0] = r0[0];
-
-                r0 += bottom_blob_tm.cstep;
-                tmpptr += 1;
-            }
-        }
-    }
-
-    bottom_blob_tm = Mat();
-    // permute end
-
-    top_blob_tm.create(tiles, batch, outch, 4u, opt.workspace_allocator);
-
-#if __loongarch_sx
-    int nn_outch = outch >> 3;
-    int remain_outch_start = nn_outch << 3;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 8;
-
-        float* output0_tm = top_blob_tm.channel(p);
-        float* output1_tm = top_blob_tm.channel(p + 1);
-        float* output2_tm = top_blob_tm.channel(p + 2);
-        float* output3_tm = top_blob_tm.channel(p + 3);
-        float* output4_tm = top_blob_tm.channel(p + 4);
-        float* output5_tm = top_blob_tm.channel(p + 5);
-        float* output6_tm = top_blob_tm.channel(p + 6);
-        float* output7_tm = top_blob_tm.channel(p + 7);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 8);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 3 < tiles; i += 4)
-            {
-                const float* r0 = bb2.row(i / 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                __m128 _sum0 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum1 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum2 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum3 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum4 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum5 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum6 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum7 = (__m128)__lsx_vreplgr2vr_w(0);
-
-                int j = 0;
-                for (; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 16);
-                    __builtin_prefetch(k0 + 32);
-                    __m128 _val = (__m128)__lsx_vld(r0, 0);
-                    __m128i _w0123 = __lsx_vld(k0, 0);
-                    __m128i _w4567 = __lsx_vld(k0 + 4, 0);
-                    _sum0 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 0), _val, _sum0);
-                    _sum1 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 1), _val, _sum1);
-                    _sum2 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 2), _val, _sum2);
-                    _sum3 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 3), _val, _sum3);
-                    _sum4 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w4567, 0), _val, _sum4);
-                    _sum5 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w4567, 1), _val, _sum5);
-                    _sum6 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w4567, 2), _val, _sum6);
-                    _sum7 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w4567, 3), _val, _sum7);
-
-                    r0 += 4;
-                    k0 += 8;
-                }
-
-                __lsx_vst(_sum0, output0_tm, 0);
-                __lsx_vst(_sum1, output1_tm, 0);
-                __lsx_vst(_sum2, output2_tm, 0);
-                __lsx_vst(_sum3, output3_tm, 0);
-                __lsx_vst(_sum4, output4_tm, 0);
-                __lsx_vst(_sum5, output5_tm, 0);
-                __lsx_vst(_sum6, output6_tm, 0);
-                __lsx_vst(_sum7, output7_tm, 0);
-
-                output0_tm += 4;
-                output1_tm += 4;
-                output2_tm += 4;
-                output3_tm += 4;
-                output4_tm += 4;
-                output5_tm += 4;
-                output6_tm += 4;
-                output7_tm += 4;
-            }
-            for (; i < tiles; i++)
-            {
-                const float* r0 = bb2.row(i / 4 + i % 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-                float sum2 = 0.f;
-                float sum3 = 0.f;
-                float sum4 = 0.f;
-                float sum5 = 0.f;
-                float sum6 = 0.f;
-                float sum7 = 0.f;
-
-                int j = 0;
-                for (; j < nn; j++)
-                {
-                    sum0 += r0[0] * k0[0];
-                    sum1 += r0[0] * k0[1];
-                    sum2 += r0[0] * k0[2];
-                    sum3 += r0[0] * k0[3];
-                    sum4 += r0[0] * k0[4];
-                    sum5 += r0[0] * k0[5];
-                    sum6 += r0[0] * k0[6];
-                    sum7 += r0[0] * k0[7];
-
-                    r0 += 1;
-                    k0 += 8;
-                }
-
-                output0_tm[0] = sum0;
-                output1_tm[0] = sum1;
-                output2_tm[0] = sum2;
-                output3_tm[0] = sum3;
-                output4_tm[0] = sum4;
-                output5_tm[0] = sum5;
-                output6_tm[0] = sum6;
-                output7_tm[0] = sum7;
-
-                output0_tm++;
-                output1_tm++;
-                output2_tm++;
-                output3_tm++;
-                output4_tm++;
-                output5_tm++;
-                output6_tm++;
-                output7_tm++;
-            }
-        }
-    }
-
-    nn_outch = (outch - remain_outch_start) >> 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = remain_outch_start + pp * 4;
-
-        float* output0_tm = top_blob_tm.channel(p);
-        float* output1_tm = top_blob_tm.channel(p + 1);
-        float* output2_tm = top_blob_tm.channel(p + 2);
-        float* output3_tm = top_blob_tm.channel(p + 3);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 8 + (p % 8) / 4);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 3 < tiles; i += 4)
-            {
-                const float* r0 = bb2.row(i / 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                __m128 _sum0 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum1 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum2 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum3 = (__m128)__lsx_vreplgr2vr_w(0);
-
-                int j = 0;
-                for (; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 16);
-                    __builtin_prefetch(k0 + 16);
-                    __m128 _val = (__m128)__lsx_vld(r0, 0);
-                    __m128i _w0123 = __lsx_vld(k0, 0);
-                    _sum0 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 0), _val, _sum0);
-                    _sum1 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 1), _val, _sum1);
-                    _sum2 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 2), _val, _sum2);
-                    _sum3 = __lsx_vfmadd_s((__m128)__lsx_vreplvei_w(_w0123, 3), _val, _sum3);
-
-                    r0 += 4;
-                    k0 += 4;
-                }
-
-                __lsx_vst(_sum0, output0_tm, 0);
-                __lsx_vst(_sum1, output1_tm, 0);
-                __lsx_vst(_sum2, output2_tm, 0);
-                __lsx_vst(_sum3, output3_tm, 0);
-
-                output0_tm += 4;
-                output1_tm += 4;
-                output2_tm += 4;
-                output3_tm += 4;
-            }
-            for (; i < tiles; i++)
-            {
-                const float* r0 = bb2.row(i / 4 + i % 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-                float sum2 = 0.f;
-                float sum3 = 0.f;
-
-                int j = 0;
-                for (; j < nn; j++)
-                {
-                    sum0 += r0[0] * k0[0];
-                    sum1 += r0[0] * k0[1];
-                    sum2 += r0[0] * k0[2];
-                    sum3 += r0[0] * k0[3];
-
-                    r0 += 1;
-                    k0 += 4;
-                }
-
-                output0_tm[0] = sum0;
-                output1_tm[0] = sum1;
-                output2_tm[0] = sum2;
-                output3_tm[0] = sum3;
-
-                output0_tm++;
-                output1_tm++;
-                output2_tm++;
-                output3_tm++;
-            }
-        }
-    }
-
-    remain_outch_start += nn_outch << 2;
-#else
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        float* output0_tm = top_blob_tm.channel(p);
-        float* output1_tm = top_blob_tm.channel(p + 1);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 2);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 3 < tiles; i += 4)
-            {
-                const float* r0 = bb2.row(i / 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum00 = 0.f;
-                float sum01 = 0.f;
-                float sum02 = 0.f;
-                float sum03 = 0.f;
-                float sum10 = 0.f;
-                float sum11 = 0.f;
-                float sum12 = 0.f;
-                float sum13 = 0.f;
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 16);
-                    __builtin_prefetch(k0 + 8);
-                    float w0 = k0[0];
-                    float w1 = k0[1];
-                    sum00 += r0[0] * w0;
-                    sum01 += r0[1] * w0;
-                    sum02 += r0[2] * w0;
-                    sum03 += r0[3] * w0;
-                    sum10 += r0[0] * w1;
-                    sum11 += r0[1] * w1;
-                    sum12 += r0[2] * w1;
-                    sum13 += r0[3] * w1;
-
-                    r0 += 4;
-                    k0 += 2;
-                }
-
-                output0_tm[0] = sum00;
-                output0_tm[1] = sum01;
-                output0_tm[2] = sum02;
-                output0_tm[3] = sum03;
-                output1_tm[0] = sum10;
-                output1_tm[1] = sum11;
-                output1_tm[2] = sum12;
-                output1_tm[3] = sum13;
-
-                output0_tm += 4;
-                output1_tm += 4;
-            }
-            for (; i < tiles; i++)
-            {
-                const float* r0 = bb2.row(i / 4 + i % 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum00 = 0.f;
-                float sum10 = 0.f;
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 4);
-                    __builtin_prefetch(k0 + 8);
-                    float val0 = r0[0];
-                    sum00 += val0 * k0[0];
-                    sum10 += val0 * k0[1];
-
-                    r0 += 1;
-                    k0 += 2;
-                }
-
-                output0_tm[0] = sum00;
-                output1_tm[0] = sum10;
-                output0_tm++;
-                output1_tm++;
-            }
-        }
-    }
-#endif
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        float* output0_tm = top_blob_tm.channel(p);
-
-#if __loongarch_sx
-        const Mat kernel0_tm = kernel_tm.channel(p / 8 + (p % 8) / 4 + p % 4);
-#else
-        const Mat kernel0_tm = kernel_tm.channel(p / 2 + p % 2);
-#endif
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 3 < tiles; i += 4)
-            {
-                const float* r0 = bb2.row(i / 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                int j = 0;
-#if __loongarch_sx
-                __m128 _sum0 = (__m128)__lsx_vreplgr2vr_w(0);
-
-                for (; j < nn; j++)
-                {
-                    _sum0 = __lsx_vfmadd_s((__m128)__lsx_vld(r0, 0), __lsx_vreplfr2vr_s(k0[0]), _sum0);
-                    r0 += 4;
-                    k0++;
-                }
-
-                __lsx_vst(_sum0, output0_tm, 0);
-                output0_tm += 4;
-#else  // __loongarch_sx
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-                float sum2 = 0.f;
-                float sum3 = 0.f;
-
-                for (; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 16);
-                    __builtin_prefetch(k0 + 4);
-                    float w0 = k0[0];
-                    sum0 += r0[0] * w0;
-                    sum1 += r0[1] * w0;
-                    sum2 += r0[2] * w0;
-                    sum3 += r0[3] * w0;
-
-                    r0 += 4;
-                    k0++;
-                }
-
-                output0_tm[0] = sum0;
-                output0_tm[1] = sum1;
-                output0_tm[2] = sum2;
-                output0_tm[3] = sum3;
-                output0_tm += 4;
-#endif // __loongarch_sx
-            }
-            for (; i < tiles; i++)
-            {
-                const float* r0 = bb2.row(i / 4 + i % 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum = 0.f;
-
-                for (int j = 0; j < nn; j++)
-                {
-                    float w0 = k0[0];
-                    float val0 = r0[0];
-                    sum += val0 * w0;
-
-                    r0 += 1;
-                    k0 += 1;
-                }
-
-                output0_tm[0] = sum;
-                output0_tm += 1;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_int8.h
deleted file mode 100644
index 2ae5ce4f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_int8.h
+++ /dev/null
@@ -1,594 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_winograd_dot_int8_lsx(Mat& bottom_blob_tm, int outch, const Mat& kernel_tm, Mat& top_blob_tm, const Option& opt)
-{
-    // Mat bottom_blob_tm(tiles, 16/36/64, inch, 2u, 1, opt.workspace_allocator);
-
-    const int tiles = bottom_blob_tm.w;
-    const int batch = bottom_blob_tm.h;
-    const int inch = bottom_blob_tm.c;
-
-    // permute
-    Mat bottom_blob_tm2;
-#if __loongarch_sx
-    if (inch >= 4)
-    {
-        if (tiles >= 2)
-            bottom_blob_tm2.create(inch / 4 + inch % 4, tiles / 2 + tiles % 2, batch, 16u, 8, opt.workspace_allocator);
-        else // if (tiles >= 1)
-            bottom_blob_tm2.create(inch / 4 + inch % 4, tiles, batch, 8u, 4, opt.workspace_allocator);
-    }
-    else
-#endif // __loongarch_sx
-    {
-        if (tiles >= 2)
-            bottom_blob_tm2.create(inch, tiles / 2 + tiles % 2, batch, 4u, 2, opt.workspace_allocator);
-        else // if (tiles >= 1)
-            bottom_blob_tm2.create(inch, tiles, batch, 2u, 1, opt.workspace_allocator);
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int r = 0; r < batch; r++)
-    {
-        Mat tm2 = bottom_blob_tm2.channel(r);
-
-        // tile
-        int i = 0;
-        for (; i + 1 < tiles; i += 2)
-        {
-            short* tmpptr = tm2.row<short>(i / 2);
-
-            const short* r0 = (const short*)bottom_blob_tm + r * tiles + i;
-
-            int q = 0;
-#if __loongarch_sx
-            const short* r1 = (const short*)bottom_blob_tm.channel(1) + r * tiles + i;
-            const short* r2 = (const short*)bottom_blob_tm.channel(2) + r * tiles + i;
-            const short* r3 = (const short*)bottom_blob_tm.channel(3) + r * tiles + i;
-            for (; q + 3 < inch; q += 4)
-            {
-                tmpptr[0] = r0[0];
-                tmpptr[1] = r1[0];
-                tmpptr[2] = r2[0];
-                tmpptr[3] = r3[0];
-                tmpptr[4] = r0[1];
-                tmpptr[5] = r1[1];
-                tmpptr[6] = r2[1];
-                tmpptr[7] = r3[1];
-                r0 += bottom_blob_tm.cstep * 4;
-                r1 += bottom_blob_tm.cstep * 4;
-                r2 += bottom_blob_tm.cstep * 4;
-                r3 += bottom_blob_tm.cstep * 4;
-                tmpptr += 8;
-            }
-#endif // __loongarch_sx
-            for (; q < inch; q++)
-            {
-                tmpptr[0] = r0[0];
-                tmpptr[1] = r0[1];
-                r0 += bottom_blob_tm.cstep;
-                tmpptr += 2;
-            }
-        }
-        for (; i < tiles; i++)
-        {
-            short* tmpptr = tm2.row<short>(i / 2 + i % 2);
-
-            const short* r0 = (const short*)bottom_blob_tm + r * tiles + i;
-
-            int q = 0;
-#if __loongarch_sx
-            const short* r1 = (const short*)bottom_blob_tm.channel(1) + r * tiles + i;
-            const short* r2 = (const short*)bottom_blob_tm.channel(2) + r * tiles + i;
-            const short* r3 = (const short*)bottom_blob_tm.channel(3) + r * tiles + i;
-            for (; q + 3 < inch; q += 4)
-            {
-                tmpptr[0] = r0[0];
-                tmpptr[1] = r1[0];
-                tmpptr[2] = r2[0];
-                tmpptr[3] = r3[0];
-                r0 += bottom_blob_tm.cstep * 4;
-                r1 += bottom_blob_tm.cstep * 4;
-                r2 += bottom_blob_tm.cstep * 4;
-                r3 += bottom_blob_tm.cstep * 4;
-                tmpptr += 4;
-            }
-#endif // __loongarch_sx
-            for (; q < inch; q++)
-            {
-                tmpptr[0] = r0[0];
-                r0 += bottom_blob_tm.cstep;
-                tmpptr += 1;
-            }
-        }
-    }
-
-    bottom_blob_tm = Mat();
-    // permute end
-
-    top_blob_tm.create(tiles, batch, outch, 4u, 1, opt.workspace_allocator);
-
-#if __loongarch_sx
-    int nn_outch = outch >> 2;
-    int remain_outch_start = nn_outch << 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 4;
-
-        int* output0_tm = top_blob_tm.channel(p);
-        int* output1_tm = top_blob_tm.channel(p + 1);
-        int* output2_tm = top_blob_tm.channel(p + 2);
-        int* output3_tm = top_blob_tm.channel(p + 3);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 4);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 1 < tiles; i += 2)
-            {
-                const short* r0 = bb2.row<const short>(i / 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int nn4 = inch / 4;
-                int nn1 = inch % 4;
-
-                __m128i _sum00 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum10 = __lsx_vreplgr2vr_w(0);
-
-                if (nn4 > 0)
-                {
-                    __m128i _sum01 = __lsx_vreplgr2vr_w(0);
-                    __m128i _sum02 = __lsx_vreplgr2vr_w(0);
-                    __m128i _sum03 = __lsx_vreplgr2vr_w(0);
-                    __m128i _sum11 = __lsx_vreplgr2vr_w(0);
-                    __m128i _sum12 = __lsx_vreplgr2vr_w(0);
-                    __m128i _sum13 = __lsx_vreplgr2vr_w(0);
-
-                    int j = 0;
-                    for (; j < nn4; j++)
-                    {
-                        __m128i _val01 = __lsx_vld(r0, 0);
-
-                        __m128i _val0 = __lsx_vilvl_d(_val01, _val01);
-                        __m128i _val1 = __lsx_vilvh_d(_val01, _val01);
-
-                        __m128i _w0 = __lsx_vld(k0, 0);
-                        __m128i _w1 = __lsx_vld(k0 + 8, 0);
-
-                        __m128i _extval0 = __lsx_vslti_h(_val0, 0);
-                        __m128i _extval1 = __lsx_vslti_h(_val1, 0);
-                        __m128i _extw0 = __lsx_vslti_h(_w0, 0);
-                        __m128i _extw1 = __lsx_vslti_h(_w1, 0);
-
-                        __m128i _val0l = __lsx_vilvl_h(_extval0, _val0);
-                        __m128i _val0h = __lsx_vilvh_h(_extval0, _val0);
-                        __m128i _val1l = __lsx_vilvl_h(_extval1, _val1);
-                        __m128i _val1h = __lsx_vilvh_h(_extval1, _val1);
-
-                        __m128i _w0l = __lsx_vilvl_h(_extw0, _w0);
-                        __m128i _w0h = __lsx_vilvh_h(_extw0, _w0);
-                        __m128i _w1l = __lsx_vilvl_h(_extw1, _w1);
-                        __m128i _w1h = __lsx_vilvh_h(_extw1, _w1);
-
-                        _sum00 = __lsx_vmadd_w(_sum00, _val0l, _w0l);
-                        _sum01 = __lsx_vmadd_w(_sum01, _val0h, _w0h);
-                        _sum02 = __lsx_vmadd_w(_sum02, _val0l, _w1l);
-                        _sum03 = __lsx_vmadd_w(_sum03, _val0h, _w1h);
-                        _sum10 = __lsx_vmadd_w(_sum10, _val1l, _w0l);
-                        _sum11 = __lsx_vmadd_w(_sum11, _val1h, _w0h);
-                        _sum12 = __lsx_vmadd_w(_sum12, _val1l, _w1l);
-                        _sum13 = __lsx_vmadd_w(_sum13, _val1h, _w1h);
-
-                        r0 += 8;
-                        k0 += 16;
-                    }
-
-                    // transpose 4x4
-                    {
-                        __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                        _tmp0 = __lsx_vilvl_w(_sum01, _sum00);
-                        _tmp1 = __lsx_vilvl_w(_sum03, _sum02);
-                        _tmp2 = __lsx_vilvh_w(_sum01, _sum00);
-                        _tmp3 = __lsx_vilvh_w(_sum03, _sum02);
-                        _sum00 = __lsx_vilvl_d(_tmp1, _tmp0);
-                        _sum01 = __lsx_vilvh_d(_tmp1, _tmp0);
-                        _sum02 = __lsx_vilvl_d(_tmp3, _tmp2);
-                        _sum03 = __lsx_vilvh_d(_tmp3, _tmp2);
-                    }
-                    {
-                        __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                        _tmp0 = __lsx_vilvl_w(_sum11, _sum10);
-                        _tmp1 = __lsx_vilvl_w(_sum13, _sum12);
-                        _tmp2 = __lsx_vilvh_w(_sum11, _sum10);
-                        _tmp3 = __lsx_vilvh_w(_sum13, _sum12);
-                        _sum10 = __lsx_vilvl_d(_tmp1, _tmp0);
-                        _sum11 = __lsx_vilvh_d(_tmp1, _tmp0);
-                        _sum12 = __lsx_vilvl_d(_tmp3, _tmp2);
-                        _sum13 = __lsx_vilvh_d(_tmp3, _tmp2);
-                    }
-
-                    _sum00 = __lsx_vadd_w(_sum00, _sum01);
-                    _sum02 = __lsx_vadd_w(_sum02, _sum03);
-                    _sum10 = __lsx_vadd_w(_sum10, _sum11);
-                    _sum12 = __lsx_vadd_w(_sum12, _sum13);
-
-                    _sum00 = __lsx_vadd_w(_sum00, _sum02);
-                    _sum10 = __lsx_vadd_w(_sum10, _sum12);
-                }
-
-                for (int j = 0; j < nn1; j++)
-                {
-                    __m128i _val0 = __lsx_vreplgr2vr_h(r0[0]);
-                    __m128i _val1 = __lsx_vreplgr2vr_h(r0[1]);
-                    __m128i _val = __lsx_vilvl_d(_val1, _val0);
-
-                    __m128i _w16 = __lsx_vld(k0, 0);
-
-                    _w16 = __lsx_vilvl_d(_w16, _w16);
-
-                    __m128i _extval = __lsx_vslti_h(_val, 0);
-                    __m128i _extw16 = __lsx_vslti_h(_w16, 0);
-
-                    __m128i _vall = __lsx_vilvl_h(_extval, _val);
-                    __m128i _valh = __lsx_vilvh_h(_extval, _val);
-                    __m128i _w0l = __lsx_vilvl_h(_extw16, _w16);
-                    __m128i _w0h = __lsx_vilvh_h(_extw16, _w16);
-
-                    _sum00 = __lsx_vmadd_w(_sum00, _vall, _w0l);
-                    _sum10 = __lsx_vmadd_w(_sum10, _valh, _w0h);
-
-                    r0 += 2;
-                    k0 += 4;
-                }
-
-                int sum[8];
-                __lsx_vst(_sum00, sum, 0);
-                __lsx_vst(_sum10, sum + 4, 0);
-
-                output0_tm[0] = sum[0];
-                output1_tm[0] = sum[1];
-                output2_tm[0] = sum[2];
-                output3_tm[0] = sum[3];
-                output0_tm[1] = sum[4];
-                output1_tm[1] = sum[5];
-                output2_tm[1] = sum[6];
-                output3_tm[1] = sum[7];
-                output0_tm += 2;
-                output1_tm += 2;
-                output2_tm += 2;
-                output3_tm += 2;
-            }
-            for (; i < tiles; i++)
-            {
-                const short* r0 = bb2.row<const short>(i / 2 + i % 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int nn4 = inch / 4;
-                int nn1 = inch % 4;
-
-                __m128i _sum0 = __lsx_vreplgr2vr_w(0);
-
-                if (nn4 > 0)
-                {
-                    __m128i _sum1 = __lsx_vreplgr2vr_w(0);
-                    __m128i _sum2 = __lsx_vreplgr2vr_w(0);
-                    __m128i _sum3 = __lsx_vreplgr2vr_w(0);
-
-                    int j = 0;
-                    for (; j < nn4; j++)
-                    {
-                        __m128i _val16 = __lsx_vld(r0, 0);
-
-                        _val16 = __lsx_vilvl_d(_val16, _val16);
-
-                        __m128i _w0 = __lsx_vld(k0, 0);
-                        __m128i _w1 = __lsx_vld(k0 + 8, 0);
-
-                        __m128i _extval16 = __lsx_vslti_h(_val16, 0);
-                        __m128i _extw0 = __lsx_vslti_h(_w0, 0);
-                        __m128i _extw1 = __lsx_vslti_h(_w1, 0);
-
-                        __m128i _val0l = __lsx_vilvl_h(_extval16, _val16);
-                        __m128i _val0h = __lsx_vilvh_h(_extval16, _val16);
-
-                        __m128i _w0l = __lsx_vilvl_h(_extw0, _w0);
-                        __m128i _w0h = __lsx_vilvh_h(_extw0, _w0);
-                        __m128i _w1l = __lsx_vilvl_h(_extw1, _w1);
-                        __m128i _w1h = __lsx_vilvh_h(_extw1, _w1);
-
-                        _sum0 = __lsx_vmadd_w(_sum0, _val0l, _w0l);
-                        _sum1 = __lsx_vmadd_w(_sum1, _val0h, _w0h);
-                        _sum2 = __lsx_vmadd_w(_sum2, _val0l, _w1l);
-                        _sum3 = __lsx_vmadd_w(_sum3, _val0h, _w1h);
-
-                        r0 += 4;
-                        k0 += 16;
-                    }
-
-                    // transpose 4x4
-                    {
-                        __m128i _tmp0, _tmp1, _tmp2, _tmp3;
-                        _tmp0 = __lsx_vilvl_w(_sum1, _sum0);
-                        _tmp1 = __lsx_vilvl_w(_sum3, _sum2);
-                        _tmp2 = __lsx_vilvh_w(_sum1, _sum0);
-                        _tmp3 = __lsx_vilvh_w(_sum3, _sum2);
-                        _sum0 = __lsx_vilvl_d(_tmp1, _tmp0);
-                        _sum1 = __lsx_vilvh_d(_tmp1, _tmp0);
-                        _sum2 = __lsx_vilvl_d(_tmp3, _tmp2);
-                        _sum3 = __lsx_vilvh_d(_tmp3, _tmp2);
-                    }
-
-                    _sum0 = __lsx_vadd_w(_sum0, _sum1);
-                    _sum2 = __lsx_vadd_w(_sum2, _sum3);
-                    _sum0 = __lsx_vadd_w(_sum0, _sum2);
-                }
-
-                for (int j = 0; j < nn1; j++)
-                {
-                    __m128i _val = __lsx_vreplgr2vr_w(r0[0]);
-                    __m128i _w16 = __lsx_vld(k0, 0);
-
-                    __m128i _extw16 = __lsx_vslti_h(_w16, 0);
-                    __m128i _w0l = __lsx_vilvl_h(_extw16, _w16);
-
-                    _sum0 = __lsx_vmadd_w(_sum0, _val, _w0l);
-
-                    r0 += 1;
-                    k0 += 4;
-                }
-
-                int sum[4];
-                __lsx_vst(_sum0, sum, 0);
-
-                output0_tm[0] = sum[0];
-                output1_tm[0] = sum[1];
-                output2_tm[0] = sum[2];
-                output3_tm[0] = sum[3];
-                output0_tm += 1;
-                output1_tm += 1;
-                output2_tm += 1;
-                output3_tm += 1;
-            }
-        }
-    }
-#else // __loongarch_sx
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        int* output0_tm = top_blob_tm.channel(p);
-        int* output1_tm = top_blob_tm.channel(p + 1);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 2);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 1 < tiles; i += 2)
-            {
-                const short* r0 = bb2.row<const short>(i / 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int sum00 = 0;
-                int sum01 = 0;
-                int sum10 = 0;
-                int sum11 = 0;
-
-                int nn1 = inch;
-
-                for (int j = 0; j < nn1; j++)
-                {
-                    signed short val0 = r0[0];
-                    signed short val1 = r0[1];
-                    signed short w0 = k0[0];
-                    signed short w1 = k0[1];
-
-                    sum00 += val0 * w0;
-                    sum01 += val1 * w0;
-                    sum10 += val0 * w1;
-                    sum11 += val1 * w1;
-
-                    r0 += 2;
-                    k0 += 2;
-                }
-
-                output0_tm[0] = sum00;
-                output0_tm[1] = sum01;
-                output1_tm[0] = sum10;
-                output1_tm[1] = sum11;
-                output0_tm += 2;
-                output1_tm += 2;
-            }
-            for (; i < tiles; i++)
-            {
-                const short* r0 = bb2.row<const short>(i / 2 + i % 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int sum0 = 0;
-                int sum1 = 0;
-
-                int nn1 = inch;
-
-                for (int j = 0; j < nn1; j++)
-                {
-                    signed short val0 = r0[0];
-                    signed short w0 = k0[0];
-                    signed short w1 = k0[1];
-
-                    sum0 += val0 * w0;
-                    sum1 += val0 * w1;
-
-                    r0 += 1;
-                    k0 += 2;
-                }
-
-                output0_tm[0] = sum0;
-                output1_tm[0] = sum1;
-                output0_tm += 1;
-                output1_tm += 1;
-            }
-        }
-    }
-#endif // __loongarch_sx
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        int* output0_tm = top_blob_tm.channel(p);
-
-#if __loongarch_sx
-        const Mat kernel0_tm = kernel_tm.channel(p / 4 + p % 4);
-#else
-        const Mat kernel0_tm = kernel_tm.channel(p / 2 + p % 2);
-#endif
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 1 < tiles; i += 2)
-            {
-                const short* r0 = bb2.row<const short>(i / 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int sum0 = 0;
-                int sum1 = 0;
-
-#if __loongarch_sx
-                int nn4 = inch / 4;
-                int nn1 = inch % 4;
-
-                if (nn4 > 0)
-                {
-                    __m128i _sum0 = __lsx_vreplgr2vr_w(0);
-                    __m128i _sum1 = __lsx_vreplgr2vr_w(0);
-
-                    int j = 0;
-                    for (; j < nn4; j++)
-                    {
-                        __m128i _val16 = __lsx_vld(r0, 0);
-
-                        __m128i _w16 = __lsx_vld(k0, 0);
-
-                        _w16 = __lsx_vilvl_d(_w16, _w16);
-
-                        __m128i _extval16 = __lsx_vslti_h(_val16, 0);
-                        __m128i _extw16 = __lsx_vslti_h(_w16, 0);
-
-                        __m128i _val0l = __lsx_vilvl_h(_extval16, _val16);
-                        __m128i _val0h = __lsx_vilvh_h(_extval16, _val16);
-
-                        __m128i _w0l = __lsx_vilvl_h(_extw16, _w16);
-                        __m128i _w0h = __lsx_vilvh_h(_extw16, _w16);
-
-                        _sum0 = __lsx_vmadd_w(_sum0, _val0l, _w0l);
-                        _sum1 = __lsx_vmadd_w(_sum1, _val0h, _w0h);
-
-                        r0 += 8;
-                        k0 += 4;
-                    }
-
-                    sum0 = __lsx_reduce_add_w(_sum0);
-                    sum1 = __lsx_reduce_add_w(_sum1);
-                }
-#else  // __loongarch_sx
-                int nn1 = inch;
-#endif // __loongarch_sx
-
-                for (int q = 0; q < nn1; q++)
-                {
-                    signed short val0 = r0[0];
-                    signed short val1 = r0[1];
-                    signed short w = k0[0];
-
-                    sum0 += val0 * w;
-                    sum1 += val1 * w;
-
-                    k0 += 1;
-                    r0 += 2;
-                }
-
-                output0_tm[0] = sum0;
-                output0_tm[1] = sum1;
-                output0_tm += 2;
-            }
-            for (; i < tiles; i++)
-            {
-                const short* r0 = bb2.row<const short>(i / 2 + i % 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int sum = 0;
-
-#if __loongarch_sx
-                int nn4 = inch / 4;
-                int nn1 = inch % 4;
-
-                if (nn4 > 0)
-                {
-                    __m128i _sum = __lsx_vreplgr2vr_w(0);
-
-                    int j = 0;
-                    for (; j < nn4; j++)
-                    {
-                        __m128i _val16 = __lsx_vld(r0, 0);
-                        __m128i _w16 = __lsx_vld(k0, 0);
-
-                        __m128i _extval16 = __lsx_vslti_h(_val16, 0);
-                        __m128i _extw16 = __lsx_vslti_h(_w16, 0);
-
-                        __m128i _val0l = __lsx_vilvl_h(_extval16, _val16);
-                        __m128i _w0l = __lsx_vilvl_h(_extw16, _w16);
-
-                        _sum = __lsx_vmadd_w(_sum, _val0l, _w0l);
-
-                        r0 += 4;
-                        k0 += 4;
-                    }
-
-                    sum = __lsx_reduce_add_w(_sum);
-                }
-#else  // __loongarch_sx
-                int nn1 = inch;
-#endif // __loongarch_sx
-
-                for (int q = 0; q < nn1; q++)
-                {
-                    signed short val = r0[0];
-                    signed short w = k0[0];
-
-                    sum += val * w;
-
-                    k0 += 1;
-                    r0 += 1;
-                }
-
-                output0_tm[0] = sum;
-                output0_tm++;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_pack4.h
deleted file mode 100644
index 66002a62..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_pack4.h
+++ /dev/null
@@ -1,448 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_winograd_dot_pack4_lsx(Mat& bottom_blob_tm, int outch, const Mat& kernel_tm, Mat& top_blob_tm, const Option& opt)
-{
-    // Mat bottom_blob_tm(tiles, 16/36/64, inch, 16u, 4, opt.workspace_allocator);
-
-    const int tiles = bottom_blob_tm.w;
-    const int batch = bottom_blob_tm.h;
-    const int inch = bottom_blob_tm.c;
-
-    // permute
-    Mat bottom_blob_tm2;
-    if (tiles >= 12)
-        bottom_blob_tm2.create(12 * inch, tiles / 12 + (tiles % 12) / 8 + (tiles % 12 % 8) / 4 + (tiles % 12 % 4) / 2 + tiles % 12 % 2, batch, 16u, 4, opt.workspace_allocator);
-    else if (tiles >= 8)
-        bottom_blob_tm2.create(8 * inch, tiles / 8 + (tiles % 8) / 4 + (tiles % 4) / 2 + tiles % 2, batch, 16u, 4, opt.workspace_allocator);
-    else if (tiles >= 4)
-        bottom_blob_tm2.create(4 * inch, tiles / 4 + (tiles % 4) / 2 + tiles % 2, batch, 16u, 4, opt.workspace_allocator);
-    else if (tiles >= 2)
-        bottom_blob_tm2.create(2 * inch, tiles / 2 + tiles % 2, batch, 16u, 4, opt.workspace_allocator);
-    else // if (tiles >= 1)
-        bottom_blob_tm2.create(1 * inch, tiles, batch, 16u, 4, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int r = 0; r < batch; r++)
-    {
-        Mat tm2 = bottom_blob_tm2.channel(r);
-
-        // tile
-        int i = 0;
-        for (; i + 11 < tiles; i += 12)
-        {
-            float* tmpptr = tm2.row(i / 12);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 4;
-
-            for (int q = 0; q < inch; q++)
-            {
-                // transpose 4x8
-                __m128i _r0 = __lsx_vld(r0, 0);
-                __m128i _r1 = __lsx_vld(r0 + 4, 0);
-                __m128i _r2 = __lsx_vld(r0 + 4 * 2, 0);
-                __m128i _r3 = __lsx_vld(r0 + 4 * 3, 0);
-                __m128i _r4 = __lsx_vld(r0 + 4 * 4, 0);
-                __m128i _r5 = __lsx_vld(r0 + 4 * 5, 0);
-                __m128i _r6 = __lsx_vld(r0 + 4 * 6, 0);
-                __m128i _r7 = __lsx_vld(r0 + 4 * 7, 0);
-                __m128i _r8 = __lsx_vld(r0 + 4 * 8, 0);
-                __m128i _r9 = __lsx_vld(r0 + 4 * 9, 0);
-                __m128i _ra = __lsx_vld(r0 + 4 * 10, 0);
-                __m128i _rb = __lsx_vld(r0 + 4 * 11, 0);
-
-                __m128i _r01r = __lsx_vilvl_w(_r1, _r0);
-                __m128i _r01l = __lsx_vilvh_w(_r1, _r0);
-                __m128i _r23r = __lsx_vilvl_w(_r3, _r2);
-                __m128i _r23l = __lsx_vilvh_w(_r3, _r2);
-                __m128i _r45r = __lsx_vilvl_w(_r5, _r4);
-                __m128i _r45l = __lsx_vilvh_w(_r5, _r4);
-                __m128i _r67r = __lsx_vilvl_w(_r7, _r6);
-                __m128i _r67l = __lsx_vilvh_w(_r7, _r6);
-                __m128i _r89r = __lsx_vilvl_w(_r9, _r8);
-                __m128i _r89l = __lsx_vilvh_w(_r9, _r8);
-                __m128i _rabr = __lsx_vilvl_w(_rb, _ra);
-                __m128i _rabl = __lsx_vilvh_w(_rb, _ra);
-                __m128i _r0123_0 = __lsx_vilvl_d(_r23r, _r01r);
-                __m128i _r0123_1 = __lsx_vilvh_d(_r23r, _r01r);
-                __m128i _r0123_2 = __lsx_vilvl_d(_r23l, _r01l);
-                __m128i _r0123_3 = __lsx_vilvh_d(_r23l, _r01l);
-                __m128i _r4567_0 = __lsx_vilvl_d(_r67r, _r45r);
-                __m128i _r4567_1 = __lsx_vilvh_d(_r67r, _r45r);
-                __m128i _r4567_2 = __lsx_vilvl_d(_r67l, _r45l);
-                __m128i _r4567_3 = __lsx_vilvh_d(_r67l, _r45l);
-                __m128i _r89ab_0 = __lsx_vilvl_d(_rabr, _r89r);
-                __m128i _r89ab_1 = __lsx_vilvh_d(_rabr, _r89r);
-                __m128i _r89ab_2 = __lsx_vilvl_d(_rabl, _r89l);
-                __m128i _r89ab_3 = __lsx_vilvh_d(_rabl, _r89l);
-
-                __lsx_vst(_r0123_0, tmpptr, 0);
-                __lsx_vst(_r4567_0, tmpptr + 4, 0);
-                __lsx_vst(_r89ab_0, tmpptr + 4 * 2, 0);
-                __lsx_vst(_r0123_1, tmpptr + 4 * 3, 0);
-                __lsx_vst(_r4567_1, tmpptr + 4 * 4, 0);
-                __lsx_vst(_r89ab_1, tmpptr + 4 * 5, 0);
-                __lsx_vst(_r0123_2, tmpptr + 4 * 6, 0);
-                __lsx_vst(_r4567_2, tmpptr + 4 * 7, 0);
-                __lsx_vst(_r89ab_2, tmpptr + 4 * 8, 0);
-                __lsx_vst(_r0123_3, tmpptr + 4 * 9, 0);
-                __lsx_vst(_r4567_3, tmpptr + 4 * 10, 0);
-                __lsx_vst(_r89ab_3, tmpptr + 4 * 11, 0);
-
-                r0 += bottom_blob_tm.cstep * 4;
-                tmpptr += 48;
-            }
-        }
-        for (; i + 7 < tiles; i += 8)
-        {
-            float* tmpptr = tm2.row(i / 12 + (i % 12) / 8);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 4;
-
-            for (int q = 0; q < inch; q++)
-            {
-                // transpose 4x8
-                __m128i _r0 = __lsx_vld(r0, 0);
-                __m128i _r1 = __lsx_vld(r0 + 4, 0);
-                __m128i _r2 = __lsx_vld(r0 + 4 * 2, 0);
-                __m128i _r3 = __lsx_vld(r0 + 4 * 3, 0);
-                __m128i _r4 = __lsx_vld(r0 + 4 * 4, 0);
-                __m128i _r5 = __lsx_vld(r0 + 4 * 5, 0);
-                __m128i _r6 = __lsx_vld(r0 + 4 * 6, 0);
-                __m128i _r7 = __lsx_vld(r0 + 4 * 7, 0);
-
-                __m128i _r01r = __lsx_vilvl_w(_r1, _r0);
-                __m128i _r01l = __lsx_vilvh_w(_r1, _r0);
-                __m128i _r23r = __lsx_vilvl_w(_r3, _r2);
-                __m128i _r23l = __lsx_vilvh_w(_r3, _r2);
-                __m128i _r45r = __lsx_vilvl_w(_r5, _r4);
-                __m128i _r45l = __lsx_vilvh_w(_r5, _r4);
-                __m128i _r67r = __lsx_vilvl_w(_r7, _r6);
-                __m128i _r67l = __lsx_vilvh_w(_r7, _r6);
-                __m128i _r0123_0 = __lsx_vilvl_d(_r23r, _r01r);
-                __m128i _r0123_1 = __lsx_vilvh_d(_r23r, _r01r);
-                __m128i _r0123_2 = __lsx_vilvl_d(_r23l, _r01l);
-                __m128i _r0123_3 = __lsx_vilvh_d(_r23l, _r01l);
-                __m128i _r4567_0 = __lsx_vilvl_d(_r67r, _r45r);
-                __m128i _r4567_1 = __lsx_vilvh_d(_r67r, _r45r);
-                __m128i _r4567_2 = __lsx_vilvl_d(_r67l, _r45l);
-                __m128i _r4567_3 = __lsx_vilvh_d(_r67l, _r45l);
-
-                __lsx_vst(_r0123_0, tmpptr, 0);
-                __lsx_vst(_r4567_0, tmpptr + 4, 0);
-                __lsx_vst(_r0123_1, tmpptr + 4 * 2, 0);
-                __lsx_vst(_r4567_1, tmpptr + 4 * 3, 0);
-                __lsx_vst(_r0123_2, tmpptr + 4 * 4, 0);
-                __lsx_vst(_r4567_2, tmpptr + 4 * 5, 0);
-                __lsx_vst(_r0123_3, tmpptr + 4 * 6, 0);
-                __lsx_vst(_r4567_3, tmpptr + 4 * 7, 0);
-
-                r0 += bottom_blob_tm.cstep * 4;
-                tmpptr += 32;
-            }
-        }
-        for (; i + 3 < tiles; i += 4)
-        {
-            float* tmpptr = tm2.row(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 4;
-
-            for (int q = 0; q < inch; q++)
-            {
-                // transpose 4x4
-                __m128i _r0 = __lsx_vld(r0, 0);
-                __m128i _r1 = __lsx_vld(r0 + 4, 0);
-                __m128i _r2 = __lsx_vld(r0 + 4 * 2, 0);
-                __m128i _r3 = __lsx_vld(r0 + 4 * 3, 0);
-
-                __m128i _r01r = __lsx_vilvl_w(_r1, _r0);
-                __m128i _r01l = __lsx_vilvh_w(_r1, _r0);
-                __m128i _r23r = __lsx_vilvl_w(_r3, _r2);
-                __m128i _r23l = __lsx_vilvh_w(_r3, _r2);
-                __m128i _r0123_0 = __lsx_vilvl_d(_r23r, _r01r);
-                __m128i _r0123_1 = __lsx_vilvh_d(_r23r, _r01r);
-                __m128i _r0123_2 = __lsx_vilvl_d(_r23l, _r01l);
-                __m128i _r0123_3 = __lsx_vilvh_d(_r23l, _r01l);
-
-                __lsx_vst(_r0123_0, tmpptr, 0);
-                __lsx_vst(_r0123_1, tmpptr + 4, 0);
-                __lsx_vst(_r0123_2, tmpptr + 4 * 2, 0);
-                __lsx_vst(_r0123_3, tmpptr + 4 * 3, 0);
-
-                r0 += bottom_blob_tm.cstep * 4;
-                tmpptr += 16;
-            }
-        }
-        for (; i + 1 < tiles; i += 2)
-        {
-            float* tmpptr = tm2.row(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 4;
-
-            for (int q = 0; q < inch; q++)
-            {
-                // transpose 4x2
-                __m128i _r0 = __lsx_vld(r0, 0);
-                __m128i _r1 = __lsx_vld(r0 + 4, 0);
-
-                __m128i _r01_0 = __lsx_vilvl_w(_r1, _r0);
-                __m128i _r01_1 = __lsx_vilvh_w(_r1, _r0);
-
-                __lsx_vst(_r01_0, tmpptr, 0);
-                __lsx_vst(_r01_1, tmpptr + 4, 0);
-
-                r0 += bottom_blob_tm.cstep * 4;
-                tmpptr += 8;
-            }
-        }
-        for (; i < tiles; i++)
-        {
-            float* tmpptr = tm2.row(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2 + i % 12 % 2);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 4;
-
-            for (int q = 0; q < inch; q++)
-            {
-                __m128i _val = __lsx_vld(r0, 0);
-                __lsx_vst(_val, tmpptr, 0);
-
-                r0 += bottom_blob_tm.cstep * 4;
-                tmpptr += 4;
-            }
-        }
-    }
-
-    bottom_blob_tm = Mat();
-    // permute end
-
-    top_blob_tm.create(tiles, batch, outch, 16u, 4, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* output0_tm = top_blob_tm.channel(p);
-
-        const Mat kernel0_tm = kernel_tm.channel(p);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 11 < tiles; i += 12)
-            {
-                const float* r0 = bb2.row(i / 12);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch * 4; // inch always > 0
-
-                __m128 _sum0 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum1 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum2 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum3 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum4 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum5 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum6 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum7 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum8 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum9 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _suma = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sumb = (__m128)__lsx_vreplgr2vr_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 48);
-                    __builtin_prefetch(k0 + 16);
-                    __m128i _val0123 = __lsx_vld(r0, 0);
-                    __m128i _val4567 = __lsx_vld(r0 + 4, 0);
-                    __m128i _val89ab = __lsx_vld(r0 + 8, 0);
-                    __m128 _w0 = (__m128)__lsx_vld(k0, 0);
-                    _sum0 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 0), _sum0);
-                    _sum1 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 1), _sum1);
-                    _sum2 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 2), _sum2);
-                    _sum3 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 3), _sum3);
-                    _sum4 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 0), _sum4);
-                    _sum5 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 1), _sum5);
-                    _sum6 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 2), _sum6);
-                    _sum7 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 3), _sum7);
-                    _sum8 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val89ab, 0), _sum8);
-                    _sum9 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val89ab, 1), _sum9);
-                    _suma = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val89ab, 2), _suma);
-                    _sumb = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val89ab, 3), _sumb);
-
-                    r0 += 12;
-                    k0 += 4;
-                }
-
-                __lsx_vst(_sum0, output0_tm, 0);
-                __lsx_vst(_sum1, output0_tm + 4, 0);
-                __lsx_vst(_sum2, output0_tm + 4 * 2, 0);
-                __lsx_vst(_sum3, output0_tm + 4 * 3, 0);
-                __lsx_vst(_sum4, output0_tm + 4 * 4, 0);
-                __lsx_vst(_sum5, output0_tm + 4 * 5, 0);
-                __lsx_vst(_sum6, output0_tm + 4 * 6, 0);
-                __lsx_vst(_sum7, output0_tm + 4 * 7, 0);
-                __lsx_vst(_sum8, output0_tm + 4 * 8, 0);
-                __lsx_vst(_sum9, output0_tm + 4 * 9, 0);
-                __lsx_vst(_suma, output0_tm + 4 * 10, 0);
-                __lsx_vst(_sumb, output0_tm + 4 * 11, 0);
-
-                output0_tm += 4 * 12;
-            }
-            for (; i + 7 < tiles; i += 8)
-            {
-                const float* r0 = bb2.row(i / 12 + (i % 12) / 8);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch * 4; // inch always > 0
-
-                __m128 _sum0 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum1 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum2 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum3 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum4 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum5 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum6 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum7 = (__m128)__lsx_vreplgr2vr_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 32);
-                    __builtin_prefetch(k0 + 16);
-                    __m128i _val0123 = __lsx_vld(r0, 0);
-                    __m128i _val4567 = __lsx_vld(r0 + 4, 0);
-                    __m128 _w0 = (__m128)__lsx_vld(k0, 0);
-                    _sum0 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 0), _sum0);
-                    _sum1 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 1), _sum1);
-                    _sum2 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 2), _sum2);
-                    _sum3 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 3), _sum3);
-                    _sum4 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 0), _sum4);
-                    _sum5 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 1), _sum5);
-                    _sum6 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 2), _sum6);
-                    _sum7 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val4567, 3), _sum7);
-
-                    r0 += 8;
-                    k0 += 4;
-                }
-
-                __lsx_vst(_sum0, output0_tm, 0);
-                __lsx_vst(_sum1, output0_tm + 4, 0);
-                __lsx_vst(_sum2, output0_tm + 4 * 2, 0);
-                __lsx_vst(_sum3, output0_tm + 4 * 3, 0);
-                __lsx_vst(_sum4, output0_tm + 4 * 4, 0);
-                __lsx_vst(_sum5, output0_tm + 4 * 5, 0);
-                __lsx_vst(_sum6, output0_tm + 4 * 6, 0);
-                __lsx_vst(_sum7, output0_tm + 4 * 7, 0);
-
-                output0_tm += 4 * 8;
-            }
-            for (; i + 3 < tiles; i += 4)
-            {
-                const float* r0 = bb2.row(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch * 4; // inch always > 0
-
-                __m128 _sum0 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum1 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum2 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum3 = (__m128)__lsx_vreplgr2vr_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 16);
-                    __builtin_prefetch(k0 + 16);
-                    __m128i _val0123 = __lsx_vld(r0, 0);
-                    __m128 _w0 = (__m128)__lsx_vld(k0, 0);
-                    _sum0 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 0), _sum0);
-                    _sum1 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 1), _sum1);
-                    _sum2 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 2), _sum2);
-                    _sum3 = __lsx_vfmadd_s(_w0, (__m128)__lsx_vreplvei_w(_val0123, 3), _sum3);
-
-                    r0 += 4;
-                    k0 += 4;
-                }
-
-                __lsx_vst(_sum0, output0_tm, 0);
-                __lsx_vst(_sum1, output0_tm + 4, 0);
-                __lsx_vst(_sum2, output0_tm + 4 * 2, 0);
-                __lsx_vst(_sum3, output0_tm + 4 * 3, 0);
-
-                output0_tm += 4 * 4;
-            }
-            for (; i + 1 < tiles; i += 2)
-            {
-                const float* r0 = bb2.row(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch * 4; // inch always > 0
-
-                __m128 _sum0 = (__m128)__lsx_vreplgr2vr_w(0);
-                __m128 _sum1 = (__m128)__lsx_vreplgr2vr_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 8);
-                    __builtin_prefetch(k0 + 16);
-                    __m128 _val0 = __lsx_vreplfr2vr_s(*r0++);
-                    __m128 _val1 = __lsx_vreplfr2vr_s(*r0++);
-                    __m128 _w0 = (__m128)__lsx_vld(k0, 0);
-                    _sum0 = __lsx_vfmadd_s(_w0, _val0, _sum0);
-                    _sum1 = __lsx_vfmadd_s(_w0, _val1, _sum1);
-
-                    k0 += 4;
-                }
-
-                __lsx_vst(_sum0, output0_tm, 0);
-                __lsx_vst(_sum1, output0_tm + 4, 0);
-
-                output0_tm += 4 * 2;
-            }
-            for (; i < tiles; i++)
-            {
-                const float* r0 = bb2.row(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2 + i % 12 % 2);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch * 4; // inch always > 0
-
-                __m128 _sum = (__m128)__lsx_vreplgr2vr_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 4);
-                    __builtin_prefetch(k0 + 16);
-                    __m128 _val0 = __lsx_vreplfr2vr_s(*r0++);
-                    __m128 _w0 = (__m128)__lsx_vld(k0, 0);
-                    _sum = __lsx_vfmadd_s(_w0, _val0, _sum);
-
-                    k0 += 4;
-                }
-
-                __lsx_vst(_sum, output0_tm, 0);
-
-                output0_tm += 4;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_pack8to1_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_pack8to1_int8.h
deleted file mode 100644
index f87aa9ef..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_pack8to1_int8.h
+++ /dev/null
@@ -1,363 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_winograd_dot_pack8to1_int8_lsx(Mat& bottom_blob_tm, int outch, const Mat& kernel_tm, Mat& top_blob_tm, const Option& opt)
-{
-    // Mat bottom_blob_tm(tiles, 16/36/64, inch, 16u, 8, opt.workspace_allocator);
-
-    const int tiles = bottom_blob_tm.w;
-    const int batch = bottom_blob_tm.h;
-    const int inch = bottom_blob_tm.c;
-
-    // permute
-    Mat bottom_blob_tm2;
-    if (tiles >= 2)
-        bottom_blob_tm2.create(2 * inch, tiles / 2 + tiles % 2, batch, 16u, 8, opt.workspace_allocator);
-    else // if (tiles >= 1)
-        bottom_blob_tm2.create(1 * inch, tiles, batch, 16u, 8, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int r = 0; r < batch; r++)
-    {
-        Mat tm2 = bottom_blob_tm2.channel(r);
-
-        // tile
-        int i = 0;
-        for (; i + 1 < tiles; i += 2)
-        {
-            short* tmpptr = tm2.row<short>(i / 2);
-
-            const short* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 8;
-
-            for (int q = 0; q < inch; q++)
-            {
-                __m128i _r0 = __lsx_vld(r0, 0);
-                __m128i _r1 = __lsx_vld(r0 + 8, 0);
-                __lsx_vst(_r0, tmpptr, 0);
-                __lsx_vst(_r1, tmpptr + 8, 0);
-                r0 += bottom_blob_tm.cstep * 8;
-                tmpptr += 16;
-            }
-        }
-        for (; i < tiles; i++)
-        {
-            short* tmpptr = tm2.row<short>(i / 2 + i % 2);
-
-            const short* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 8;
-
-            for (int q = 0; q < inch; q++)
-            {
-                __m128i _r0 = __lsx_vld(r0, 0);
-                __lsx_vst(_r0, tmpptr, 0);
-                r0 += bottom_blob_tm.cstep * 8;
-                tmpptr += 8;
-            }
-        }
-    }
-
-    bottom_blob_tm = Mat();
-    // permute end
-
-    top_blob_tm.create(tiles, batch, outch, 4u, 1, opt.workspace_allocator);
-
-    int nn_outch = 0;
-    int remain_outch_start = 0;
-
-    nn_outch = outch >> 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 4;
-
-        int* output0_tm = top_blob_tm.channel(p);
-        int* output1_tm = top_blob_tm.channel(p + 1);
-        int* output2_tm = top_blob_tm.channel(p + 2);
-        int* output3_tm = top_blob_tm.channel(p + 3);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 4);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 1 < tiles; i += 2)
-            {
-                const short* r0 = bb2.row<const short>(i / 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int nn = inch; // inch always > 0
-
-                __m128i _sum0 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum1 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum2 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum3 = __lsx_vreplgr2vr_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 64);
-                    __builtin_prefetch(k0 + 128);
-                    __m128i _w0 = __lsx_vld(k0, 0);
-                    __m128i _w1 = __lsx_vld(k0 + 8, 0);
-                    __m128i _w2 = __lsx_vld(k0 + 16, 0);
-                    __m128i _w3 = __lsx_vld(k0 + 24, 0);
-
-                    __m128i _extw0 = __lsx_vslti_h(_w0, 0);
-                    __m128i _extw1 = __lsx_vslti_h(_w1, 0);
-                    __m128i _extw2 = __lsx_vslti_h(_w2, 0);
-                    __m128i _extw3 = __lsx_vslti_h(_w3, 0);
-
-                    __m128i _w0l = __lsx_vilvl_h(_extw0, _w0);
-                    __m128i _w0h = __lsx_vilvh_h(_extw0, _w0);
-                    __m128i _w1l = __lsx_vilvl_h(_extw1, _w1);
-                    __m128i _w1h = __lsx_vilvh_h(_extw1, _w1);
-                    __m128i _w2l = __lsx_vilvl_h(_extw2, _w2);
-                    __m128i _w2h = __lsx_vilvh_h(_extw2, _w2);
-                    __m128i _w3l = __lsx_vilvl_h(_extw3, _w3);
-                    __m128i _w3h = __lsx_vilvh_h(_extw3, _w3);
-
-                    __m128i _val0_0 = __lsx_vreplgr2vr_w(r0[0]);
-                    __m128i _val0_1 = __lsx_vreplgr2vr_w(r0[1]);
-                    __m128i _val0_2 = __lsx_vreplgr2vr_w(r0[2]);
-                    __m128i _val0_3 = __lsx_vreplgr2vr_w(r0[3]);
-                    __m128i _val0_4 = __lsx_vreplgr2vr_w(r0[4]);
-                    __m128i _val0_5 = __lsx_vreplgr2vr_w(r0[5]);
-                    __m128i _val0_6 = __lsx_vreplgr2vr_w(r0[6]);
-                    __m128i _val0_7 = __lsx_vreplgr2vr_w(r0[7]);
-                    __m128i _val1_0 = __lsx_vreplgr2vr_w(r0[8]);
-                    __m128i _val1_1 = __lsx_vreplgr2vr_w(r0[9]);
-                    __m128i _val1_2 = __lsx_vreplgr2vr_w(r0[10]);
-                    __m128i _val1_3 = __lsx_vreplgr2vr_w(r0[11]);
-                    __m128i _val1_4 = __lsx_vreplgr2vr_w(r0[12]);
-                    __m128i _val1_5 = __lsx_vreplgr2vr_w(r0[13]);
-                    __m128i _val1_6 = __lsx_vreplgr2vr_w(r0[14]);
-                    __m128i _val1_7 = __lsx_vreplgr2vr_w(r0[15]);
-
-                    _sum0 = __lsx_vmadd_w(_sum0, _w0l, _val0_0);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w0h, _val0_1);
-                    _sum2 = __lsx_vmadd_w(_sum2, _w0l, _val1_0);
-                    _sum3 = __lsx_vmadd_w(_sum3, _w0h, _val1_1);
-                    _sum0 = __lsx_vmadd_w(_sum0, _w1l, _val0_2);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w1h, _val0_3);
-                    _sum2 = __lsx_vmadd_w(_sum2, _w1l, _val1_2);
-                    _sum3 = __lsx_vmadd_w(_sum3, _w1h, _val1_3);
-                    _sum0 = __lsx_vmadd_w(_sum0, _w2l, _val0_4);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w2h, _val0_5);
-                    _sum2 = __lsx_vmadd_w(_sum2, _w2l, _val1_4);
-                    _sum3 = __lsx_vmadd_w(_sum3, _w2h, _val1_5);
-                    _sum0 = __lsx_vmadd_w(_sum0, _w3l, _val0_6);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w3h, _val0_7);
-                    _sum2 = __lsx_vmadd_w(_sum2, _w3l, _val1_6);
-                    _sum3 = __lsx_vmadd_w(_sum3, _w3h, _val1_7);
-
-                    r0 += 16;
-                    k0 += 32;
-                }
-
-                _sum0 = __lsx_vadd_w(_sum0, _sum1);
-                _sum2 = __lsx_vadd_w(_sum2, _sum3);
-
-                int sum[8];
-                __lsx_vst(_sum0, sum, 0);
-                __lsx_vst(_sum2, sum + 4, 0);
-
-                output0_tm[0] = sum[0];
-                output1_tm[0] = sum[1];
-                output2_tm[0] = sum[2];
-                output3_tm[0] = sum[3];
-                output0_tm[1] = sum[4];
-                output1_tm[1] = sum[5];
-                output2_tm[1] = sum[6];
-                output3_tm[1] = sum[7];
-                output0_tm += 2;
-                output1_tm += 2;
-                output2_tm += 2;
-                output3_tm += 2;
-            }
-            for (; i < tiles; i++)
-            {
-                const short* r0 = bb2.row<const short>(i / 2 + i % 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int nn = inch; // inch always > 0
-
-                __m128i _sum0 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum1 = __lsx_vreplgr2vr_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 32);
-                    __builtin_prefetch(k0 + 128);
-                    __m128i _w0 = __lsx_vld(k0, 0);
-                    __m128i _w1 = __lsx_vld(k0 + 8, 0);
-                    __m128i _w2 = __lsx_vld(k0 + 16, 0);
-                    __m128i _w3 = __lsx_vld(k0 + 24, 0);
-
-                    __m128i _extw0 = __lsx_vslti_h(_w0, 0);
-                    __m128i _extw1 = __lsx_vslti_h(_w1, 0);
-                    __m128i _extw2 = __lsx_vslti_h(_w2, 0);
-                    __m128i _extw3 = __lsx_vslti_h(_w3, 0);
-
-                    __m128i _w0l = __lsx_vilvl_h(_extw0, _w0);
-                    __m128i _w0h = __lsx_vilvh_h(_extw0, _w0);
-                    __m128i _w1l = __lsx_vilvl_h(_extw1, _w1);
-                    __m128i _w1h = __lsx_vilvh_h(_extw1, _w1);
-                    __m128i _w2l = __lsx_vilvl_h(_extw2, _w2);
-                    __m128i _w2h = __lsx_vilvh_h(_extw2, _w2);
-                    __m128i _w3l = __lsx_vilvl_h(_extw3, _w3);
-                    __m128i _w3h = __lsx_vilvh_h(_extw3, _w3);
-
-                    __m128i _val0 = __lsx_vreplgr2vr_w(r0[0]);
-                    __m128i _val1 = __lsx_vreplgr2vr_w(r0[1]);
-                    __m128i _val2 = __lsx_vreplgr2vr_w(r0[2]);
-                    __m128i _val3 = __lsx_vreplgr2vr_w(r0[3]);
-                    __m128i _val4 = __lsx_vreplgr2vr_w(r0[4]);
-                    __m128i _val5 = __lsx_vreplgr2vr_w(r0[5]);
-                    __m128i _val6 = __lsx_vreplgr2vr_w(r0[6]);
-                    __m128i _val7 = __lsx_vreplgr2vr_w(r0[7]);
-
-                    _sum0 = __lsx_vmadd_w(_sum0, _w0l, _val0);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w0h, _val1);
-                    _sum0 = __lsx_vmadd_w(_sum0, _w1l, _val2);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w1h, _val3);
-                    _sum0 = __lsx_vmadd_w(_sum0, _w2l, _val4);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w2h, _val5);
-                    _sum0 = __lsx_vmadd_w(_sum0, _w3l, _val6);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w3h, _val7);
-
-                    r0 += 8;
-                    k0 += 32;
-                }
-
-                _sum0 = __lsx_vadd_w(_sum0, _sum1);
-
-                int sum[4];
-                __lsx_vst(_sum0, sum, 0);
-
-                output0_tm[0] = sum[0];
-                output1_tm[0] = sum[1];
-                output2_tm[0] = sum[2];
-                output3_tm[0] = sum[3];
-                output0_tm += 1;
-                output1_tm += 1;
-                output2_tm += 1;
-                output3_tm += 1;
-            }
-        }
-    }
-
-    remain_outch_start += nn_outch << 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        int* output0_tm = top_blob_tm.channel(p);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 4 + p % 4);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 1 < tiles; i += 2)
-            {
-                const short* r0 = bb2.row<const short>(i / 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                __m128i _sum0 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum1 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum2 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum3 = __lsx_vreplgr2vr_w(0);
-
-                for (int q = 0; q < inch; q++)
-                {
-                    __builtin_prefetch(r0 + 32);
-                    __builtin_prefetch(k0 + 64);
-                    __m128i _val0 = __lsx_vld(r0, 0);
-                    __m128i _val1 = __lsx_vld(r0 + 8, 0);
-
-                    __m128i _extval0 = __lsx_vslti_h(_val0, 0);
-                    __m128i _extval1 = __lsx_vslti_h(_val1, 0);
-                    __m128i _val0l = __lsx_vilvl_h(_extval0, _val0);
-                    __m128i _val0h = __lsx_vilvh_h(_extval0, _val0);
-                    __m128i _val1l = __lsx_vilvl_h(_extval1, _val1);
-                    __m128i _val1h = __lsx_vilvh_h(_extval1, _val1);
-
-                    __m128i _w0 = __lsx_vld(k0, 0);
-
-                    __m128i _extw0 = __lsx_vslti_h(_w0, 0);
-                    __m128i _w0l = __lsx_vilvl_h(_extw0, _w0);
-                    __m128i _w0h = __lsx_vilvh_h(_extw0, _w0);
-
-                    _sum0 = __lsx_vmadd_w(_sum0, _w0l, _val0l);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w0h, _val0h);
-                    _sum2 = __lsx_vmadd_w(_sum2, _w0l, _val1l);
-                    _sum3 = __lsx_vmadd_w(_sum3, _w0h, _val1h);
-
-                    k0 += 8;
-                    r0 += 16;
-                }
-
-                _sum0 = __lsx_vadd_w(_sum0, _sum1);
-                _sum2 = __lsx_vadd_w(_sum2, _sum3);
-
-                output0_tm[0] = __lsx_reduce_add_w(_sum0);
-                output0_tm[1] = __lsx_reduce_add_w(_sum2);
-                output0_tm += 2;
-            }
-            for (; i < tiles; i++)
-            {
-                const short* r0 = bb2.row<const short>(i / 2 + i % 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                __m128i _sum0 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum1 = __lsx_vreplgr2vr_w(0);
-
-                for (int q = 0; q < inch; q++)
-                {
-                    __builtin_prefetch(r0 + 32);
-                    __builtin_prefetch(k0 + 32);
-                    __m128i _val = __lsx_vld(r0, 0);
-
-                    __m128i _extval = __lsx_vslti_h(_val, 0);
-                    __m128i _vall = __lsx_vilvl_h(_extval, _val);
-                    __m128i _valh = __lsx_vilvh_h(_extval, _val);
-
-                    __m128i _w0 = __lsx_vld(k0, 0);
-
-                    __m128i _extw0 = __lsx_vslti_h(_w0, 0);
-                    __m128i _w0l = __lsx_vilvl_h(_extw0, _w0);
-                    __m128i _w0h = __lsx_vilvh_h(_extw0, _w0);
-
-                    _sum0 = __lsx_vmadd_w(_sum0, _w0l, _vall);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w0h, _valh);
-
-                    k0 += 8;
-                    r0 += 8;
-                }
-
-                _sum0 = __lsx_vadd_w(_sum0, _sum1);
-
-                output0_tm[0] = __lsx_reduce_add_w(_sum0);
-                output0_tm++;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_pack8to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_pack8to4_int8.h
deleted file mode 100644
index c20400cb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_dot_pack8to4_int8.h
+++ /dev/null
@@ -1,233 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_winograd_dot_pack8to4_int8_lsx(Mat& bottom_blob_tm, int outch, const Mat& kernel_tm, Mat& top_blob_tm, const Option& opt)
-{
-    // Mat bottom_blob_tm(tiles, 16/36/64, inch, 16u, 8, opt.workspace_allocator);
-
-    const int tiles = bottom_blob_tm.w;
-    const int batch = bottom_blob_tm.h;
-    const int inch = bottom_blob_tm.c;
-
-    // permute
-    Mat bottom_blob_tm2;
-    if (tiles >= 2)
-        bottom_blob_tm2.create(2 * inch, tiles / 2 + tiles % 2, batch, 16u, 8, opt.workspace_allocator);
-    else // if (tiles >= 1)
-        bottom_blob_tm2.create(1 * inch, tiles, batch, 16u, 8, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int r = 0; r < batch; r++)
-    {
-        Mat tm2 = bottom_blob_tm2.channel(r);
-
-        // tile
-        int i = 0;
-        for (; i + 1 < tiles; i += 2)
-        {
-            short* tmpptr = tm2.row<short>(i / 2);
-
-            const short* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 8;
-
-            for (int q = 0; q < inch; q++)
-            {
-                __m128i _r0 = __lsx_vld(r0, 0);
-                __m128i _r1 = __lsx_vld(r0 + 8, 0);
-                __lsx_vst(_r0, tmpptr, 0);
-                __lsx_vst(_r1, tmpptr + 8, 0);
-                r0 += bottom_blob_tm.cstep * 8;
-                tmpptr += 16;
-            }
-        }
-        for (; i < tiles; i++)
-        {
-            short* tmpptr = tm2.row<short>(i / 2 + i % 2);
-
-            const short* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 8;
-
-            for (int q = 0; q < inch; q++)
-            {
-                __m128i _r0 = __lsx_vld(r0, 0);
-                __lsx_vst(_r0, tmpptr, 0);
-                r0 += bottom_blob_tm.cstep * 8;
-                tmpptr += 8;
-            }
-        }
-    }
-
-    bottom_blob_tm = Mat();
-    // permute end
-
-    top_blob_tm.create(tiles, batch, outch, 16u, 4, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        int* output0_tm = top_blob_tm.channel(p);
-
-        const Mat kernel0_tm = kernel_tm.channel(p);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 1 < tiles; i += 2)
-            {
-                const short* r0 = bb2.row<const short>(i / 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int nn = inch; // inch always > 0
-
-                __m128i _sum0 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum1 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum2 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum3 = __lsx_vreplgr2vr_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 64);
-                    __builtin_prefetch(k0 + 128);
-                    __m128i _w0 = __lsx_vld(k0, 0);
-                    __m128i _w1 = __lsx_vld(k0 + 8, 0);
-                    __m128i _w2 = __lsx_vld(k0 + 16, 0);
-                    __m128i _w3 = __lsx_vld(k0 + 24, 0);
-
-                    __m128i _extw0 = __lsx_vslti_h(_w0, 0);
-                    __m128i _extw1 = __lsx_vslti_h(_w1, 0);
-                    __m128i _extw2 = __lsx_vslti_h(_w2, 0);
-                    __m128i _extw3 = __lsx_vslti_h(_w3, 0);
-
-                    __m128i _w0l = __lsx_vilvl_h(_extw0, _w0);
-                    __m128i _w0h = __lsx_vilvh_h(_extw0, _w0);
-                    __m128i _w1l = __lsx_vilvl_h(_extw1, _w1);
-                    __m128i _w1h = __lsx_vilvh_h(_extw1, _w1);
-                    __m128i _w2l = __lsx_vilvl_h(_extw2, _w2);
-                    __m128i _w2h = __lsx_vilvh_h(_extw2, _w2);
-                    __m128i _w3l = __lsx_vilvl_h(_extw3, _w3);
-                    __m128i _w3h = __lsx_vilvh_h(_extw3, _w3);
-
-                    __m128i _val0_0 = __lsx_vreplgr2vr_w(r0[0]);
-                    __m128i _val0_1 = __lsx_vreplgr2vr_w(r0[1]);
-                    __m128i _val0_2 = __lsx_vreplgr2vr_w(r0[2]);
-                    __m128i _val0_3 = __lsx_vreplgr2vr_w(r0[3]);
-                    __m128i _val0_4 = __lsx_vreplgr2vr_w(r0[4]);
-                    __m128i _val0_5 = __lsx_vreplgr2vr_w(r0[5]);
-                    __m128i _val0_6 = __lsx_vreplgr2vr_w(r0[6]);
-                    __m128i _val0_7 = __lsx_vreplgr2vr_w(r0[7]);
-                    __m128i _val1_0 = __lsx_vreplgr2vr_w(r0[8]);
-                    __m128i _val1_1 = __lsx_vreplgr2vr_w(r0[9]);
-                    __m128i _val1_2 = __lsx_vreplgr2vr_w(r0[10]);
-                    __m128i _val1_3 = __lsx_vreplgr2vr_w(r0[11]);
-                    __m128i _val1_4 = __lsx_vreplgr2vr_w(r0[12]);
-                    __m128i _val1_5 = __lsx_vreplgr2vr_w(r0[13]);
-                    __m128i _val1_6 = __lsx_vreplgr2vr_w(r0[14]);
-                    __m128i _val1_7 = __lsx_vreplgr2vr_w(r0[15]);
-
-                    _sum0 = __lsx_vmadd_w(_sum0, _w0l, _val0_0);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w0h, _val0_1);
-                    _sum2 = __lsx_vmadd_w(_sum2, _w0l, _val1_0);
-                    _sum3 = __lsx_vmadd_w(_sum3, _w0h, _val1_1);
-                    _sum0 = __lsx_vmadd_w(_sum0, _w1l, _val0_2);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w1h, _val0_3);
-                    _sum2 = __lsx_vmadd_w(_sum2, _w1l, _val1_2);
-                    _sum3 = __lsx_vmadd_w(_sum3, _w1h, _val1_3);
-                    _sum0 = __lsx_vmadd_w(_sum0, _w2l, _val0_4);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w2h, _val0_5);
-                    _sum2 = __lsx_vmadd_w(_sum2, _w2l, _val1_4);
-                    _sum3 = __lsx_vmadd_w(_sum3, _w2h, _val1_5);
-                    _sum0 = __lsx_vmadd_w(_sum0, _w3l, _val0_6);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w3h, _val0_7);
-                    _sum2 = __lsx_vmadd_w(_sum2, _w3l, _val1_6);
-                    _sum3 = __lsx_vmadd_w(_sum3, _w3h, _val1_7);
-
-                    r0 += 16;
-                    k0 += 32;
-                }
-
-                _sum0 = __lsx_vadd_w(_sum0, _sum1);
-                _sum2 = __lsx_vadd_w(_sum2, _sum3);
-
-                __lsx_vst(_sum0, output0_tm, 0);
-                __lsx_vst(_sum2, output0_tm + 4, 0);
-
-                output0_tm += 8;
-            }
-            for (; i < tiles; i++)
-            {
-                const short* r0 = bb2.row<const short>(i / 2 + i % 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int nn = inch; // inch always > 0
-
-                __m128i _sum0 = __lsx_vreplgr2vr_w(0);
-                __m128i _sum1 = __lsx_vreplgr2vr_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 32);
-                    __builtin_prefetch(k0 + 128);
-                    __m128i _w0 = __lsx_vld(k0, 0);
-                    __m128i _w1 = __lsx_vld(k0 + 8, 0);
-                    __m128i _w2 = __lsx_vld(k0 + 16, 0);
-                    __m128i _w3 = __lsx_vld(k0 + 24, 0);
-
-                    __m128i _extw0 = __lsx_vslti_h(_w0, 0);
-                    __m128i _extw1 = __lsx_vslti_h(_w1, 0);
-                    __m128i _extw2 = __lsx_vslti_h(_w2, 0);
-                    __m128i _extw3 = __lsx_vslti_h(_w3, 0);
-
-                    __m128i _w0l = __lsx_vilvl_h(_extw0, _w0);
-                    __m128i _w0h = __lsx_vilvh_h(_extw0, _w0);
-                    __m128i _w1l = __lsx_vilvl_h(_extw1, _w1);
-                    __m128i _w1h = __lsx_vilvh_h(_extw1, _w1);
-                    __m128i _w2l = __lsx_vilvl_h(_extw2, _w2);
-                    __m128i _w2h = __lsx_vilvh_h(_extw2, _w2);
-                    __m128i _w3l = __lsx_vilvl_h(_extw3, _w3);
-                    __m128i _w3h = __lsx_vilvh_h(_extw3, _w3);
-
-                    __m128i _val0 = __lsx_vreplgr2vr_w(r0[0]);
-                    __m128i _val1 = __lsx_vreplgr2vr_w(r0[1]);
-                    __m128i _val2 = __lsx_vreplgr2vr_w(r0[2]);
-                    __m128i _val3 = __lsx_vreplgr2vr_w(r0[3]);
-                    __m128i _val4 = __lsx_vreplgr2vr_w(r0[4]);
-                    __m128i _val5 = __lsx_vreplgr2vr_w(r0[5]);
-                    __m128i _val6 = __lsx_vreplgr2vr_w(r0[6]);
-                    __m128i _val7 = __lsx_vreplgr2vr_w(r0[7]);
-
-                    _sum0 = __lsx_vmadd_w(_sum0, _w0l, _val0);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w0h, _val1);
-                    _sum0 = __lsx_vmadd_w(_sum0, _w1l, _val2);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w1h, _val3);
-                    _sum0 = __lsx_vmadd_w(_sum0, _w2l, _val4);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w2h, _val5);
-                    _sum0 = __lsx_vmadd_w(_sum0, _w3l, _val6);
-                    _sum1 = __lsx_vmadd_w(_sum1, _w3h, _val7);
-
-                    r0 += 8;
-                    k0 += 32;
-                }
-
-                _sum0 = __lsx_vadd_w(_sum0, _sum1);
-
-                __lsx_vst(_sum0, output0_tm, 0);
-                output0_tm += 4;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform.h
deleted file mode 100644
index 624600e9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform.h
+++ /dev/null
@@ -1,405 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd43_transform_input_lsx(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 4;
-    const int h_tiles = (h - 2) / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[6][6] = {
-    //     {4.0f, 0.0f, -5.0f, 0.0f, 1.0f, 0.0f},
-    //     {0.0f,-4.0f, -4.0f, 1.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f, -4.0f,-1.0f, 1.0f, 0.0f},
-    //     {0.0f,-2.0f, -1.0f, 2.0f, 1.0f, 0.0f},
-    //     {0.0f, 2.0f, -1.0f,-2.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f,  0.0f,-5.0f, 0.0f, 1.0f}
-    // };
-
-    // 0 =  4 * r00 - 5 * r02 + r04
-    // 1 = -4 * (r01 + r02) + r04 + r03
-    // 2 =  4 * (r01 - r02) + r04 - r03
-    // 3 = -2 * (r01 - r03) + r04 - r02
-    // 4 =  2 * (r01 - r03) + r04 - r02
-    // 5 =  4 * r01 - 5 * r03 + r05
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        float tmp[6][6];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* r0 = img0.row(i * 4) + (j * 4);
-
-                for (int m = 0; m < 6; m++)
-                {
-                    float r00 = r0[0];
-                    float r01 = r0[1];
-                    float r02 = r0[2];
-                    float r03 = r0[3];
-                    float r04 = r0[4];
-                    float r05 = r0[5];
-
-                    float tmp0m = 4 * r00 - 5 * r02 + r04;
-                    float tmp1m = -4 * (r01 + r02) + r04 + r03;
-                    float tmp2m = 4 * (r01 - r02) + r04 - r03;
-                    float tmp3m = -2 * (r01 - r03) + r04 - r02;
-                    float tmp4m = 2 * (r01 - r03) + r04 - r02;
-                    float tmp5m = 4 * r01 - 5 * r03 + r05;
-
-                    tmp[0][m] = tmp0m;
-                    tmp[1][m] = tmp1m;
-                    tmp[2][m] = tmp2m;
-                    tmp[3][m] = tmp3m;
-                    tmp[4][m] = tmp4m;
-                    tmp[5][m] = tmp5m;
-
-                    r0 += w;
-                }
-
-                float* r0_tm_0 = (float*)img0_tm + (i * w_tiles + j);
-                float* r0_tm_1 = r0_tm_0 + tiles;
-                float* r0_tm_2 = r0_tm_0 + tiles * 2;
-                float* r0_tm_3 = r0_tm_0 + tiles * 3;
-                float* r0_tm_4 = r0_tm_0 + tiles * 4;
-                float* r0_tm_5 = r0_tm_0 + tiles * 5;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    float tmp00 = tmp[m][0];
-                    float tmp01 = tmp[m][1];
-                    float tmp02 = tmp[m][2];
-                    float tmp03 = tmp[m][3];
-                    float tmp04 = tmp[m][4];
-                    float tmp05 = tmp[m][5];
-
-                    float r0tm0 = 4 * tmp00 - 5 * tmp02 + tmp04;
-                    float r0tm1 = -4 * (tmp01 + tmp02) + tmp04 + tmp03;
-                    float r0tm2 = 4 * (tmp01 - tmp02) + tmp04 - tmp03;
-                    float r0tm3 = -2 * (tmp01 - tmp03) + tmp04 - tmp02;
-                    float r0tm4 = 2 * (tmp01 - tmp03) + tmp04 - tmp02;
-                    float r0tm5 = 4 * tmp01 - 5 * tmp03 + tmp05;
-
-                    r0_tm_0[0] = r0tm0;
-                    r0_tm_1[0] = r0tm1;
-                    r0_tm_2[0] = r0tm2;
-                    r0_tm_3[0] = r0tm3;
-                    r0_tm_4[0] = r0tm4;
-                    r0_tm_5[0] = r0tm5;
-
-                    r0_tm_0 += tiles * 6;
-                    r0_tm_1 += tiles * 6;
-                    r0_tm_2 += tiles * 6;
-                    r0_tm_3 += tiles * 6;
-                    r0_tm_4 += tiles * 6;
-                    r0_tm_5 += tiles * 6;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_output_lsx(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 4;
-    const int h_tiles = outh / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    const float* biasptr = bias;
-
-    // const float otm[4][6] = {
-    //     {1.0f, 1.0f,  1.0f, 1.0f,  1.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 2.0f, -2.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f, 4.0f,  4.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 8.0f, -8.0f, 1.0f}
-    // };
-
-    // 0 = r00 + (r01 + r02) + (r03 + r04)
-    // 1 =       (r01 - r02) + (r03 - r04) * 2
-    // 2 =       (r01 + r02) + (r03 + r04) * 4
-    // 3 = r05 + (r01 - r02) + (r03 - r04) * 8
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        float bias0 = biasptr ? biasptr[p] : 0.f;
-
-        float tmp[4][6];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* output0_tm_0 = (const float*)out0_tm + (i * w_tiles + j);
-                const float* output0_tm_1 = output0_tm_0 + tiles;
-                const float* output0_tm_2 = output0_tm_0 + tiles * 2;
-                const float* output0_tm_3 = output0_tm_0 + tiles * 3;
-                const float* output0_tm_4 = output0_tm_0 + tiles * 4;
-                const float* output0_tm_5 = output0_tm_0 + tiles * 5;
-
-                float* output0 = out0.row(i * 4) + (j * 4);
-
-                for (int m = 0; m < 6; m++)
-                {
-                    float out0tm0 = output0_tm_0[0];
-                    float out0tm1 = output0_tm_1[0];
-                    float out0tm2 = output0_tm_2[0];
-                    float out0tm3 = output0_tm_3[0];
-                    float out0tm4 = output0_tm_4[0];
-                    float out0tm5 = output0_tm_5[0];
-
-                    float tmp02a = out0tm1 + out0tm2;
-                    float tmp13a = out0tm1 - out0tm2;
-
-                    float tmp02b = out0tm3 + out0tm4;
-                    float tmp13b = out0tm3 - out0tm4;
-
-                    float tmp0m = out0tm0 + tmp02a + tmp02b;
-                    float tmp1m = tmp13a + tmp13b * 2;
-                    float tmp2m = tmp02a + tmp02b * 4;
-                    float tmp3m = out0tm5 + tmp13a + tmp13b * 8;
-
-                    tmp[0][m] = tmp0m;
-                    tmp[1][m] = tmp1m;
-                    tmp[2][m] = tmp2m;
-                    tmp[3][m] = tmp3m;
-
-                    output0_tm_0 += tiles * 6;
-                    output0_tm_1 += tiles * 6;
-                    output0_tm_2 += tiles * 6;
-                    output0_tm_3 += tiles * 6;
-                    output0_tm_4 += tiles * 6;
-                    output0_tm_5 += tiles * 6;
-                }
-
-                for (int m = 0; m < 4; m++)
-                {
-                    float tmp00 = tmp[m][0];
-                    float tmp01 = tmp[m][1];
-                    float tmp02 = tmp[m][2];
-                    float tmp03 = tmp[m][3];
-                    float tmp04 = tmp[m][4];
-                    float tmp05 = tmp[m][5];
-
-                    float tmp02a = tmp01 + tmp02;
-                    float tmp13a = tmp01 - tmp02;
-
-                    float tmp02b = tmp03 + tmp04;
-                    float tmp13b = tmp03 - tmp04;
-
-                    float out00 = bias0 + tmp00 + tmp02a + tmp02b;
-                    float out01 = bias0 + tmp13a + tmp13b * 2;
-                    float out02 = bias0 + tmp02a + tmp02b * 4;
-                    float out03 = bias0 + tmp05 + tmp13a + tmp13b * 8;
-
-                    output0[0] = out00;
-                    output0[1] = out01;
-                    output0[2] = out02;
-                    output0[3] = out03;
-
-                    output0 += outw;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_input_lsx(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 2;
-    const int h_tiles = (h - 2) / 2;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[4][4] = {
-    //     {1.0f,  0.0f, -1.0f,  0.0f},
-    //     {0.0f,  1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  0.00f, 1.0f}
-    // };
-
-    // 0 = r00 - r02
-    // 1 = r01 + r02
-    // 2 = r02 - r01
-    // 3 = r03 - r01
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        float tmp[4][4];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* r0 = img0.row(i * 2) + (j * 2);
-
-                for (int m = 0; m < 4; m++)
-                {
-                    float r00 = r0[0];
-                    float r01 = r0[1];
-                    float r02 = r0[2];
-                    float r03 = r0[3];
-
-                    float tmp0m = r00 - r02;
-                    float tmp1m = r01 + r02;
-                    float tmp2m = r02 - r01;
-                    float tmp3m = r03 - r01;
-
-                    tmp[0][m] = tmp0m;
-                    tmp[1][m] = tmp1m;
-                    tmp[2][m] = tmp2m;
-                    tmp[3][m] = tmp3m;
-
-                    r0 += w;
-                }
-
-                float* r0_tm_0 = (float*)img0_tm + (i * w_tiles + j);
-                float* r0_tm_1 = r0_tm_0 + tiles;
-                float* r0_tm_2 = r0_tm_0 + tiles * 2;
-                float* r0_tm_3 = r0_tm_0 + tiles * 3;
-
-                for (int m = 0; m < 4; m++)
-                {
-                    float tmp00 = tmp[m][0];
-                    float tmp01 = tmp[m][1];
-                    float tmp02 = tmp[m][2];
-                    float tmp03 = tmp[m][3];
-
-                    float r0tm0 = tmp00 - tmp02;
-                    float r0tm1 = tmp01 + tmp02;
-                    float r0tm2 = tmp02 - tmp01;
-                    float r0tm3 = tmp03 - tmp01;
-
-                    r0_tm_0[0] = r0tm0;
-                    r0_tm_1[0] = r0tm1;
-                    r0_tm_2[0] = r0tm2;
-                    r0_tm_3[0] = r0tm3;
-
-                    r0_tm_0 += tiles * 4;
-                    r0_tm_1 += tiles * 4;
-                    r0_tm_2 += tiles * 4;
-                    r0_tm_3 += tiles * 4;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_output_lsx(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 2;
-    const int h_tiles = outh / 2;
-    const int tiles = w_tiles * h_tiles;
-
-    const float* biasptr = bias;
-
-    // const float otm[2][4] = {
-    //     {1.0f,  1.0f,  1.0f,  0.0f},
-    //     {0.0f,  1.0f, -1.0f,  1.0f}
-    // };
-
-    // 0 = r00 + r01 + r02
-    // 1 = r01 - r02 + r03
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        float bias0 = biasptr ? biasptr[p] : 0.f;
-
-        float tmp[2][4];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* output0_tm_0 = (const float*)out0_tm + (i * w_tiles + j);
-                const float* output0_tm_1 = output0_tm_0 + tiles;
-                const float* output0_tm_2 = output0_tm_0 + tiles * 2;
-                const float* output0_tm_3 = output0_tm_0 + tiles * 3;
-
-                float* output0 = out0.row(i * 2) + (j * 2);
-
-                for (int m = 0; m < 4; m++)
-                {
-                    float out0tm0 = output0_tm_0[0];
-                    float out0tm1 = output0_tm_1[0];
-                    float out0tm2 = output0_tm_2[0];
-                    float out0tm3 = output0_tm_3[0];
-
-                    float tmp0m = out0tm0 + out0tm1 + out0tm2;
-                    float tmp1m = out0tm1 - out0tm2 + out0tm3;
-
-                    tmp[0][m] = tmp0m;
-                    tmp[1][m] = tmp1m;
-
-                    output0_tm_0 += tiles * 4;
-                    output0_tm_1 += tiles * 4;
-                    output0_tm_2 += tiles * 4;
-                    output0_tm_3 += tiles * 4;
-                }
-
-                for (int m = 0; m < 2; m++)
-                {
-                    float tmp00 = tmp[m][0];
-                    float tmp01 = tmp[m][1];
-                    float tmp02 = tmp[m][2];
-                    float tmp03 = tmp[m][3];
-
-                    float out00 = bias0 + tmp00 + tmp01 + tmp02;
-                    float out01 = bias0 + tmp01 - tmp02 + tmp03;
-
-                    output0[0] = out00;
-                    output0[1] = out01;
-
-                    output0 += outw;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_int8.h
deleted file mode 100644
index 09ef669e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_int8.h
+++ /dev/null
@@ -1,229 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd43_transform_input_int8_lsx(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 4;
-    const int h_tiles = (h - 2) / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[6][6] = {
-    //     {4.0f, 0.0f, -5.0f, 0.0f, 1.0f, 0.0f},
-    //     {0.0f,-4.0f, -4.0f, 1.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f, -4.0f,-1.0f, 1.0f, 0.0f},
-    //     {0.0f,-2.0f, -1.0f, 2.0f, 1.0f, 0.0f},
-    //     {0.0f, 2.0f, -1.0f,-2.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f,  0.0f,-5.0f, 0.0f, 1.0f}
-    // };
-
-    // 0 =  4 * r00 - 5 * r02 + r04
-    // 1 = -4 * (r01 + r02) + r04 + r03
-    // 2 =  4 * (r01 - r02) + r04 - r03
-    // 3 = -2 * (r01 - r03) + r04 - r02
-    // 4 =  2 * (r01 - r03) + r04 - r02
-    // 5 =  4 * r01 - 5 * r03 + r05
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        short tmp[6][6];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const signed char* r0 = img0.row<const signed char>(i * 4) + (j * 4);
-
-                for (int m = 0; m < 6; m++)
-                {
-                    signed char r00 = r0[0];
-                    signed char r01 = r0[1];
-                    signed char r02 = r0[2];
-                    signed char r03 = r0[3];
-                    signed char r04 = r0[4];
-                    signed char r05 = r0[5];
-
-                    short tmp0m = 4 * r00 - 5 * r02 + r04;
-                    short tmp1m = -4 * (r01 + r02) + r04 + r03;
-                    short tmp2m = 4 * (r01 - r02) + r04 - r03;
-                    short tmp3m = -2 * (r01 - r03) + r04 - r02;
-                    short tmp4m = 2 * (r01 - r03) + r04 - r02;
-                    short tmp5m = 4 * r01 - 5 * r03 + r05;
-
-                    tmp[0][m] = tmp0m;
-                    tmp[1][m] = tmp1m;
-                    tmp[2][m] = tmp2m;
-                    tmp[3][m] = tmp3m;
-                    tmp[4][m] = tmp4m;
-                    tmp[5][m] = tmp5m;
-
-                    r0 += w;
-                }
-
-                short* r0_tm_0 = (short*)img0_tm + (i * w_tiles + j);
-                short* r0_tm_1 = r0_tm_0 + tiles;
-                short* r0_tm_2 = r0_tm_0 + tiles * 2;
-                short* r0_tm_3 = r0_tm_0 + tiles * 3;
-                short* r0_tm_4 = r0_tm_0 + tiles * 4;
-                short* r0_tm_5 = r0_tm_0 + tiles * 5;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    short tmp00 = tmp[m][0];
-                    short tmp01 = tmp[m][1];
-                    short tmp02 = tmp[m][2];
-                    short tmp03 = tmp[m][3];
-                    short tmp04 = tmp[m][4];
-                    short tmp05 = tmp[m][5];
-
-                    short r0tm0 = 4 * tmp00 - 5 * tmp02 + tmp04;
-                    short r0tm1 = -4 * (tmp01 + tmp02) + tmp04 + tmp03;
-                    short r0tm2 = 4 * (tmp01 - tmp02) + tmp04 - tmp03;
-                    short r0tm3 = -2 * (tmp01 - tmp03) + tmp04 - tmp02;
-                    short r0tm4 = 2 * (tmp01 - tmp03) + tmp04 - tmp02;
-                    short r0tm5 = 4 * tmp01 - 5 * tmp03 + tmp05;
-
-                    r0_tm_0[0] = r0tm0;
-                    r0_tm_1[0] = r0tm1;
-                    r0_tm_2[0] = r0tm2;
-                    r0_tm_3[0] = r0tm3;
-                    r0_tm_4[0] = r0tm4;
-                    r0_tm_5[0] = r0tm5;
-
-                    r0_tm_0 += tiles * 6;
-                    r0_tm_1 += tiles * 6;
-                    r0_tm_2 += tiles * 6;
-                    r0_tm_3 += tiles * 6;
-                    r0_tm_4 += tiles * 6;
-                    r0_tm_5 += tiles * 6;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_output_int8_lsx(const Mat& top_blob_tm, Mat& top_blob, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 4;
-    const int h_tiles = outh / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float otm[4][6] = {
-    //     {1.0f, 1.0f,  1.0f, 1.0f,  1.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 2.0f, -2.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f, 4.0f,  4.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 8.0f, -8.0f, 1.0f}
-    // };
-
-    // 0 = r00 + (r01 + r02) + (r03 + r04)
-    // 1 =       (r01 - r02) + (r03 - r04) * 2
-    // 2 =       (r01 + r02) + (r03 + r04) * 4
-    // 3 = r05 + (r01 - r02) + (r03 - r04) * 8
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        int tmp[4][6];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const int* output0_tm_0 = (const int*)out0_tm + (i * w_tiles + j) * 1;
-                const int* output0_tm_1 = output0_tm_0 + tiles * 1;
-                const int* output0_tm_2 = output0_tm_0 + tiles * 2;
-                const int* output0_tm_3 = output0_tm_0 + tiles * 3;
-                const int* output0_tm_4 = output0_tm_0 + tiles * 4;
-                const int* output0_tm_5 = output0_tm_0 + tiles * 5;
-
-                int* output0 = out0.row<int>(i * 4) + j * 4;
-
-                for (int m = 0; m < 5; m++)
-                {
-                    int tmp02a = output0_tm_1[0] + output0_tm_2[0];
-                    int tmp13a = output0_tm_1[0] - output0_tm_2[0];
-
-                    int tmp02b = output0_tm_3[0] + output0_tm_4[0];
-                    int tmp13b = output0_tm_3[0] - output0_tm_4[0];
-
-                    tmp[0][m] = output0_tm_0[0] + tmp02a + tmp02b;
-                    tmp[1][m] = tmp13a + tmp13b * 2;
-                    tmp[2][m] = tmp02a + tmp02b * 4;
-                    tmp[3][m] = output0_tm_5[0] * 4 + tmp13a + tmp13b * 8;
-
-                    output0_tm_0 += tiles * 6;
-                    output0_tm_1 += tiles * 6;
-                    output0_tm_2 += tiles * 6;
-                    output0_tm_3 += tiles * 6;
-                    output0_tm_4 += tiles * 6;
-                    output0_tm_5 += tiles * 6;
-                }
-                for (int m = 5; m < 6; m++)
-                {
-                    int tmp02a = output0_tm_1[0] + output0_tm_2[0];
-                    int tmp13a = output0_tm_1[0] - output0_tm_2[0];
-
-                    int tmp02b = output0_tm_3[0] + output0_tm_4[0];
-                    int tmp13b = output0_tm_3[0] - output0_tm_4[0];
-
-                    tmp[0][m] = (output0_tm_0[0] + tmp02a + tmp02b) * 4;
-                    tmp[1][m] = (tmp13a + tmp13b * 2) * 4;
-                    tmp[2][m] = (tmp02a + tmp02b * 4) * 4;
-                    tmp[3][m] = (output0_tm_5[0] * 4 + tmp13a + tmp13b * 8) * 4;
-
-                    output0_tm_0 += tiles * 6;
-                    output0_tm_1 += tiles * 6;
-                    output0_tm_2 += tiles * 6;
-                    output0_tm_3 += tiles * 6;
-                    output0_tm_4 += tiles * 6;
-                    output0_tm_5 += tiles * 6;
-                }
-
-                for (int m = 0; m < 4; m++)
-                {
-                    const int* tmp0 = tmp[m];
-
-                    int tmp02a = tmp0[1] + tmp0[2];
-                    int tmp13a = tmp0[1] - tmp0[2];
-
-                    int tmp02b = tmp0[3] + tmp0[4];
-                    int tmp13b = tmp0[3] - tmp0[4];
-
-                    output0[0] = (tmp0[0] + tmp02a + tmp02b) / 576;
-                    output0[1] = (tmp13a + tmp13b * 2) / 576;
-                    output0[2] = (tmp02a + tmp02b * 4) / 576;
-                    output0[3] = (tmp0[5] + tmp13a + tmp13b * 8) / 576;
-
-                    output0 += outw;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_pack4.h
deleted file mode 100644
index 3969e59c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_pack4.h
+++ /dev/null
@@ -1,730 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd63_transform_input_pack4_lsx(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 6;
-    const int h_tiles = (h - 2) / 6;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[8][8] = {
-    //     {1.0f,  0.0f, -5.25f,  0.00f,  5.25f,  0.00f, -1.0f, 0.0f},
-    //
-    //     {0.0f,  1.0f,  1.00f, -4.25f, -4.25f,  1.00f,  1.0f, 0.0f},
-    //     {0.0f, -1.0f,  1.00f,  4.25f, -4.25f, -1.00f,  1.0f, 0.0f},
-    //
-    //     {0.0f,  0.5f,  0.25f, -2.50f, -1.25f,  2.00f,  1.0f, 0.0f},
-    //     {0.0f, -0.5f,  0.25f,  2.50f, -1.25f, -2.00f,  1.0f, 0.0f},
-    //
-    //     {0.0f,  2.0f,  4.00f, -2.50f, -5.00f,  0.50f,  1.0f, 0.0f},
-    //     {0.0f, -2.0f,  4.00f,  2.50f, -5.00f, -0.50f,  1.0f, 0.0f},
-    //
-    //     {0.0f, -1.0f,  0.00f,  5.25f,  0.00f, -5.25f,  0.0f, 1.0f}
-    // };
-
-    // 0 = r00 - r06 + (r04 - r02) * 5.25
-    // 7 = r07 - r01 + (r03 - r05) * 5.25
-
-    // 1 = (r02 + r06 - r04 * 4.25) + (r01 - r03 * 4.25 + r05)
-    // 2 = (r02 + r06 - r04 * 4.25) - (r01 - r03 * 4.25 + r05)
-
-    // 3 = (r06 + r02 * 0.25 - r04 * 1.25) + (r01 * 0.5 - r03 * 2.5 + r05 * 2)
-    // 4 = (r06 + r02 * 0.25 - r04 * 1.25) - (r01 * 0.5 - r03 * 2.5 + r05 * 2)
-
-    // reuse r04 * 1.25
-    // reuse r03 * 2.5
-    // 5 = (r06 + (r02 - r04 * 1.25) * 4) + (r01 * 2 - r03 * 2.5 + r05 * 0.5)
-    // 6 = (r06 + (r02 - r04 * 1.25) * 4) - (r01 * 2 - r03 * 2.5 + r05 * 0.5)
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        float tmp[8][8][4];
-
-        __m128 _v5_25 = __lsx_vreplfr2vr_s(5.25f);
-        __m128 _vm4_25 = __lsx_vreplfr2vr_s(-4.25f);
-        __m128 _vm1_25 = __lsx_vreplfr2vr_s(-1.25f);
-        __m128 _v0_25 = __lsx_vreplfr2vr_s(0.25f);
-        __m128 _vm2_5 = __lsx_vreplfr2vr_s(-2.5f);
-        __m128 _v0_5 = __lsx_vreplfr2vr_s(0.5f);
-        __m128 _v2 = __lsx_vreplfr2vr_s(2.f);
-        __m128 _v4 = __lsx_vreplfr2vr_s(4.f);
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* r0 = img0.row(i * 6) + (j * 6) * 4;
-
-                for (int m = 0; m < 8; m++)
-                {
-                    __m128 _r00 = (__m128)__lsx_vld(r0, 0);
-                    __m128 _r01 = (__m128)__lsx_vld(r0 + 4, 0);
-                    __m128 _r02 = (__m128)__lsx_vld(r0 + 4 * 2, 0);
-                    __m128 _r03 = (__m128)__lsx_vld(r0 + 4 * 3, 0);
-                    __m128 _r04 = (__m128)__lsx_vld(r0 + 4 * 4, 0);
-                    __m128 _r05 = (__m128)__lsx_vld(r0 + 4 * 5, 0);
-                    __m128 _r06 = (__m128)__lsx_vld(r0 + 4 * 6, 0);
-                    __m128 _r07 = (__m128)__lsx_vld(r0 + 4 * 7, 0);
-
-                    __m128 _tmp0m = __lsx_vfmadd_s(__lsx_vfsub_s(_r04, _r02), _v5_25, __lsx_vfsub_s(_r00, _r06));
-                    __m128 _tmp7m = __lsx_vfmadd_s(__lsx_vfsub_s(_r03, _r05), _v5_25, __lsx_vfsub_s(_r07, _r01));
-                    __lsx_vst(_tmp0m, tmp[0][m], 0);
-                    __lsx_vst(_tmp7m, tmp[7][m], 0);
-
-                    __m128 _tmp12a = __lsx_vfmadd_s(_r04, _vm4_25, __lsx_vfadd_s(_r02, _r06));
-                    __m128 _tmp12b = __lsx_vfmadd_s(_r03, _vm4_25, __lsx_vfadd_s(_r01, _r05));
-
-                    __m128 _tmp1m = __lsx_vfadd_s(_tmp12a, _tmp12b);
-                    __m128 _tmp2m = __lsx_vfsub_s(_tmp12a, _tmp12b);
-                    __lsx_vst(_tmp1m, tmp[1][m], 0);
-                    __lsx_vst(_tmp2m, tmp[2][m], 0);
-
-                    __m128 _tmp34a = __lsx_vfmadd_s(_r04, _vm1_25, __lsx_vfmadd_s(_r02, _v0_25, _r06));
-                    __m128 _tmp34b = __lsx_vfmadd_s(_r05, _v2, __lsx_vfmadd_s(_r03, _vm2_5, __lsx_vfmul_s(_r01, _v0_5)));
-
-                    __m128 _tmp3m = __lsx_vfadd_s(_tmp34a, _tmp34b);
-                    __m128 _tmp4m = __lsx_vfsub_s(_tmp34a, _tmp34b);
-                    __lsx_vst(_tmp3m, tmp[3][m], 0);
-                    __lsx_vst(_tmp4m, tmp[4][m], 0);
-
-                    __m128 _tmp56a = __lsx_vfmadd_s(__lsx_vfmadd_s(_r04, _vm1_25, _r02), _v4, _r06);
-                    __m128 _tmp56b = __lsx_vfmadd_s(_r05, _v0_5, __lsx_vfmadd_s(_r03, _vm2_5, __lsx_vfmul_s(_r01, _v2)));
-
-                    __m128 _tmp5m = __lsx_vfadd_s(_tmp56a, _tmp56b);
-                    __m128 _tmp6m = __lsx_vfsub_s(_tmp56a, _tmp56b);
-                    __lsx_vst(_tmp5m, tmp[5][m], 0);
-                    __lsx_vst(_tmp6m, tmp[6][m], 0);
-
-                    r0 += w * 4;
-                }
-
-                float* r0_tm_0 = (float*)img0_tm + (i * w_tiles + j) * 4;
-                float* r0_tm_1 = r0_tm_0 + tiles * 4;
-                float* r0_tm_2 = r0_tm_0 + tiles * 4 * 2;
-                float* r0_tm_3 = r0_tm_0 + tiles * 4 * 3;
-                float* r0_tm_4 = r0_tm_0 + tiles * 4 * 4;
-                float* r0_tm_5 = r0_tm_0 + tiles * 4 * 5;
-                float* r0_tm_6 = r0_tm_0 + tiles * 4 * 6;
-                float* r0_tm_7 = r0_tm_0 + tiles * 4 * 7;
-
-                for (int m = 0; m < 8; m++)
-                {
-                    __m128 _tmp00 = (__m128)__lsx_vld(tmp[m][0], 0);
-                    __m128 _tmp01 = (__m128)__lsx_vld(tmp[m][1], 0);
-                    __m128 _tmp02 = (__m128)__lsx_vld(tmp[m][2], 0);
-                    __m128 _tmp03 = (__m128)__lsx_vld(tmp[m][3], 0);
-                    __m128 _tmp04 = (__m128)__lsx_vld(tmp[m][4], 0);
-                    __m128 _tmp05 = (__m128)__lsx_vld(tmp[m][5], 0);
-                    __m128 _tmp06 = (__m128)__lsx_vld(tmp[m][6], 0);
-                    __m128 _tmp07 = (__m128)__lsx_vld(tmp[m][7], 0);
-
-                    __m128 _r0tm0 = __lsx_vfmadd_s(__lsx_vfsub_s(_tmp04, _tmp02), _v5_25, __lsx_vfsub_s(_tmp00, _tmp06));
-                    __m128 _r0tm7 = __lsx_vfmadd_s(__lsx_vfsub_s(_tmp03, _tmp05), _v5_25, __lsx_vfsub_s(_tmp07, _tmp01));
-
-                    __m128 _tmp12a = __lsx_vfmadd_s(_tmp04, _vm4_25, __lsx_vfadd_s(_tmp02, _tmp06));
-                    __m128 _tmp12b = __lsx_vfmadd_s(_tmp03, _vm4_25, __lsx_vfadd_s(_tmp01, _tmp05));
-
-                    __m128 _r0tm1 = __lsx_vfadd_s(_tmp12a, _tmp12b);
-                    __m128 _r0tm2 = __lsx_vfsub_s(_tmp12a, _tmp12b);
-
-                    __m128 _tmp34a = __lsx_vfmadd_s(_tmp04, _vm1_25, __lsx_vfmadd_s(_tmp02, _v0_25, _tmp06));
-                    __m128 _tmp34b = __lsx_vfmadd_s(_tmp05, _v2, __lsx_vfmadd_s(_tmp03, _vm2_5, __lsx_vfmul_s(_tmp01, _v0_5)));
-
-                    __m128 _r0tm3 = __lsx_vfadd_s(_tmp34a, _tmp34b);
-                    __m128 _r0tm4 = __lsx_vfsub_s(_tmp34a, _tmp34b);
-
-                    __m128 _tmp56a = __lsx_vfmadd_s(__lsx_vfmadd_s(_tmp04, _vm1_25, _tmp02), _v4, _tmp06);
-                    __m128 _tmp56b = __lsx_vfmadd_s(_tmp05, _v0_5, __lsx_vfmadd_s(_tmp03, _vm2_5, __lsx_vfmul_s(_tmp01, _v2)));
-
-                    __m128 _r0tm5 = __lsx_vfadd_s(_tmp56a, _tmp56b);
-                    __m128 _r0tm6 = __lsx_vfsub_s(_tmp56a, _tmp56b);
-
-                    __lsx_vst(_r0tm0, r0_tm_0, 0);
-                    __lsx_vst(_r0tm1, r0_tm_1, 0);
-                    __lsx_vst(_r0tm2, r0_tm_2, 0);
-                    __lsx_vst(_r0tm3, r0_tm_3, 0);
-                    __lsx_vst(_r0tm4, r0_tm_4, 0);
-                    __lsx_vst(_r0tm5, r0_tm_5, 0);
-                    __lsx_vst(_r0tm6, r0_tm_6, 0);
-                    __lsx_vst(_r0tm7, r0_tm_7, 0);
-
-                    r0_tm_0 += tiles * 4 * 8;
-                    r0_tm_1 += tiles * 4 * 8;
-                    r0_tm_2 += tiles * 4 * 8;
-                    r0_tm_3 += tiles * 4 * 8;
-                    r0_tm_4 += tiles * 4 * 8;
-                    r0_tm_5 += tiles * 4 * 8;
-                    r0_tm_6 += tiles * 4 * 8;
-                    r0_tm_7 += tiles * 4 * 8;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd63_transform_output_pack4_lsx(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 6;
-    const int h_tiles = outh / 6;
-    const int tiles = w_tiles * h_tiles;
-
-    const float* biasptr = bias;
-
-    // const float otm[6][8] = {
-    //     {1.0f,  1.0f,   1.0f,   1.0f,   1.0f,  32.0f, 32.0f, 0.0f},
-    //     {0.0f,  1.0f,  -1.0f,   2.0f,  -2.0f,  16.0f,-16.0f, 0.0f},
-    //     {0.0f,  1.0f,   1.0f,   4.0f,   4.0f,   8.0f,  8.0f, 0.0f},
-    //     {0.0f,  1.0f,  -1.0f,   8.0f,  -8.0f,   4.0f, -4.0f, 0.0f},
-    //     {0.0f,  1.0f,   1.0f,  16.0f,  16.0f,   2.0f,  2.0f, 0.0f},
-    //     {0.0f,  1.0f,  -1.0f,  32.0f, -32.0f,   1.0f, -1.0f, 1.0f}
-    // };
-
-    // 0 = r0 + (r1 + r2) + (r3 + r4)     + (r5 + r6) * 32
-    // 1 =      (r1 - r2) + (r3 - r4) * 2 + (r5 - r6) * 16
-    // 2 =      (r1 + r2) + (r3 + r4) * 4 + (r5 + r6) * 8
-    // 3 =      (r1 - r2) + (r3 - r4) * 8 + (r5 - r6) * 4
-    // 4 =      (r1 + r2) + (r3 + r4) * 16+ (r5 + r6) * 2
-    // 5 = r7 + (r1 - r2) + (r3 - r4) * 32+ (r5 - r6)
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        __m128 _bias0 = biasptr ? (__m128)__lsx_vld(biasptr + p * 4, 0) : (__m128)__lsx_vreplgr2vr_w(0);
-
-        float tmp[6][8][4];
-
-        __m128 _v32 = __lsx_vreplfr2vr_s(32.f);
-        __m128 _v16 = __lsx_vreplfr2vr_s(16.f);
-        __m128 _v8 = __lsx_vreplfr2vr_s(8.f);
-        __m128 _v4 = __lsx_vreplfr2vr_s(4.f);
-        __m128 _v2 = __lsx_vreplfr2vr_s(2.f);
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* output0_tm_0 = (const float*)out0_tm + (i * w_tiles + j) * 4;
-                const float* output0_tm_1 = output0_tm_0 + tiles * 4;
-                const float* output0_tm_2 = output0_tm_0 + tiles * 4 * 2;
-                const float* output0_tm_3 = output0_tm_0 + tiles * 4 * 3;
-                const float* output0_tm_4 = output0_tm_0 + tiles * 4 * 4;
-                const float* output0_tm_5 = output0_tm_0 + tiles * 4 * 5;
-                const float* output0_tm_6 = output0_tm_0 + tiles * 4 * 6;
-                const float* output0_tm_7 = output0_tm_0 + tiles * 4 * 7;
-
-                float* output0 = out0.row<float>(i * 6) + (j * 6) * 4;
-
-                for (int m = 0; m < 8; m++)
-                {
-                    __m128 _out0tm0 = (__m128)__lsx_vld(output0_tm_0, 0);
-                    __m128 _out0tm1 = (__m128)__lsx_vld(output0_tm_1, 0);
-                    __m128 _out0tm2 = (__m128)__lsx_vld(output0_tm_2, 0);
-                    __m128 _out0tm3 = (__m128)__lsx_vld(output0_tm_3, 0);
-                    __m128 _out0tm4 = (__m128)__lsx_vld(output0_tm_4, 0);
-                    __m128 _out0tm5 = (__m128)__lsx_vld(output0_tm_5, 0);
-                    __m128 _out0tm6 = (__m128)__lsx_vld(output0_tm_6, 0);
-                    __m128 _out0tm7 = (__m128)__lsx_vld(output0_tm_7, 0);
-
-                    __m128 _tmp024a = __lsx_vfadd_s(_out0tm1, _out0tm2);
-                    __m128 _tmp135a = __lsx_vfsub_s(_out0tm1, _out0tm2);
-
-                    __m128 _tmp024b = __lsx_vfadd_s(_out0tm3, _out0tm4);
-                    __m128 _tmp135b = __lsx_vfsub_s(_out0tm3, _out0tm4);
-
-                    __m128 _tmp024c = __lsx_vfadd_s(_out0tm5, _out0tm6);
-                    __m128 _tmp135c = __lsx_vfsub_s(_out0tm5, _out0tm6);
-
-                    __m128 _tmp0m = __lsx_vfadd_s(__lsx_vfadd_s(_out0tm0, _tmp024a), __lsx_vfmadd_s(_tmp024c, _v32, _tmp024b));
-                    __m128 _tmp2m = __lsx_vfmadd_s(_tmp024c, _v8, __lsx_vfmadd_s(_tmp024b, _v4, _tmp024a));
-                    __m128 _tmp4m = __lsx_vfmadd_s(_tmp024c, _v2, __lsx_vfmadd_s(_tmp024b, _v16, _tmp024a));
-                    __lsx_vst(_tmp0m, tmp[0][m], 0);
-                    __lsx_vst(_tmp2m, tmp[2][m], 0);
-                    __lsx_vst(_tmp4m, tmp[4][m], 0);
-
-                    __m128 _tmp1m = __lsx_vfmadd_s(_tmp135c, _v16, __lsx_vfmadd_s(_tmp135b, _v2, _tmp135a));
-                    __m128 _tmp3m = __lsx_vfmadd_s(_tmp135c, _v4, __lsx_vfmadd_s(_tmp135b, _v8, _tmp135a));
-                    __m128 _tmp5m = __lsx_vfadd_s(__lsx_vfadd_s(_out0tm7, _tmp135a), __lsx_vfmadd_s(_tmp135b, _v32, _tmp135c));
-                    __lsx_vst(_tmp1m, tmp[1][m], 0);
-                    __lsx_vst(_tmp3m, tmp[3][m], 0);
-                    __lsx_vst(_tmp5m, tmp[5][m], 0);
-
-                    output0_tm_0 += tiles * 4 * 8;
-                    output0_tm_1 += tiles * 4 * 8;
-                    output0_tm_2 += tiles * 4 * 8;
-                    output0_tm_3 += tiles * 4 * 8;
-                    output0_tm_4 += tiles * 4 * 8;
-                    output0_tm_5 += tiles * 4 * 8;
-                    output0_tm_6 += tiles * 4 * 8;
-                    output0_tm_7 += tiles * 4 * 8;
-                }
-
-                for (int m = 0; m < 6; m++)
-                {
-                    __m128 _tmp00 = (__m128)__lsx_vld(tmp[m][0], 0);
-                    __m128 _tmp01 = (__m128)__lsx_vld(tmp[m][1], 0);
-                    __m128 _tmp02 = (__m128)__lsx_vld(tmp[m][2], 0);
-                    __m128 _tmp03 = (__m128)__lsx_vld(tmp[m][3], 0);
-                    __m128 _tmp04 = (__m128)__lsx_vld(tmp[m][4], 0);
-                    __m128 _tmp05 = (__m128)__lsx_vld(tmp[m][5], 0);
-                    __m128 _tmp06 = (__m128)__lsx_vld(tmp[m][6], 0);
-                    __m128 _tmp07 = (__m128)__lsx_vld(tmp[m][7], 0);
-
-                    __m128 _tmp024a = __lsx_vfadd_s(_tmp01, _tmp02);
-                    __m128 _tmp135a = __lsx_vfsub_s(_tmp01, _tmp02);
-
-                    __m128 _tmp024b = __lsx_vfadd_s(_tmp03, _tmp04);
-                    __m128 _tmp135b = __lsx_vfsub_s(_tmp03, _tmp04);
-
-                    __m128 _tmp024c = __lsx_vfadd_s(_tmp05, _tmp06);
-                    __m128 _tmp135c = __lsx_vfsub_s(_tmp05, _tmp06);
-
-                    __m128 _out00 = __lsx_vfadd_s(_bias0, __lsx_vfadd_s(__lsx_vfadd_s(_tmp00, _tmp024a), __lsx_vfmadd_s(_tmp024c, _v32, _tmp024b)));
-                    __m128 _out02 = __lsx_vfadd_s(_bias0, __lsx_vfmadd_s(_tmp024c, _v8, __lsx_vfmadd_s(_tmp024b, _v4, _tmp024a)));
-                    __m128 _out04 = __lsx_vfadd_s(_bias0, __lsx_vfmadd_s(_tmp024c, _v2, __lsx_vfmadd_s(_tmp024b, _v16, _tmp024a)));
-                    __lsx_vst(_out00, output0, 0);
-                    __lsx_vst(_out02, output0 + 4 * 2, 0);
-                    __lsx_vst(_out04, output0 + 4 * 4, 0);
-
-                    __m128 _out01 = __lsx_vfadd_s(_bias0, __lsx_vfmadd_s(_tmp135c, _v16, __lsx_vfmadd_s(_tmp135b, _v2, _tmp135a)));
-                    __m128 _out03 = __lsx_vfadd_s(_bias0, __lsx_vfmadd_s(_tmp135c, _v4, __lsx_vfmadd_s(_tmp135b, _v8, _tmp135a)));
-                    __m128 _out05 = __lsx_vfadd_s(_bias0, __lsx_vfadd_s(__lsx_vfadd_s(_tmp07, _tmp135a), __lsx_vfmadd_s(_tmp135b, _v32, _tmp135c)));
-                    __lsx_vst(_out01, output0 + 4, 0);
-                    __lsx_vst(_out03, output0 + 4 * 3, 0);
-                    __lsx_vst(_out05, output0 + 4 * 5, 0);
-
-                    output0 += outw * 4;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_input_pack4_lsx(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 4;
-    const int h_tiles = (h - 2) / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[6][6] = {
-    //     {4.0f, 0.0f, -5.0f, 0.0f, 1.0f, 0.0f},
-    //     {0.0f,-4.0f, -4.0f, 1.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f, -4.0f,-1.0f, 1.0f, 0.0f},
-    //     {0.0f,-2.0f, -1.0f, 2.0f, 1.0f, 0.0f},
-    //     {0.0f, 2.0f, -1.0f,-2.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f,  0.0f,-5.0f, 0.0f, 1.0f}
-    // };
-
-    // 0 =  4 * r00 - 5 * r02 + r04
-    // 1 = -4 * (r01 + r02) + r04 + r03
-    // 2 =  4 * (r01 - r02) + r04 - r03
-    // 3 = -2 * (r01 - r03) + r04 - r02
-    // 4 =  2 * (r01 - r03) + r04 - r02
-    // 5 =  4 * r01 - 5 * r03 + r05
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        float tmp[6][6][4];
-
-        __m128 _vm5 = __lsx_vreplfr2vr_s(-5.f);
-        __m128 _vm4 = __lsx_vreplfr2vr_s(-4.f);
-        __m128 _v4 = __lsx_vreplfr2vr_s(4.f);
-        __m128 _vm2 = __lsx_vreplfr2vr_s(-2.f);
-        __m128 _v2 = __lsx_vreplfr2vr_s(2.f);
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* r0 = img0.row(i * 4) + (j * 4) * 4;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    __m128 _r00 = (__m128)__lsx_vld(r0, 0);
-                    __m128 _r01 = (__m128)__lsx_vld(r0 + 4, 0);
-                    __m128 _r02 = (__m128)__lsx_vld(r0 + 4 * 2, 0);
-                    __m128 _r03 = (__m128)__lsx_vld(r0 + 4 * 3, 0);
-                    __m128 _r04 = (__m128)__lsx_vld(r0 + 4 * 4, 0);
-                    __m128 _r05 = (__m128)__lsx_vld(r0 + 4 * 5, 0);
-
-                    __m128 _tmp0m = __lsx_vfmadd_s(_r02, _vm5, __lsx_vfmadd_s(_r00, _v4, _r04));
-                    __m128 _tmp1m = __lsx_vfmadd_s(__lsx_vfadd_s(_r01, _r02), _vm4, __lsx_vfadd_s(_r04, _r03));
-                    __m128 _tmp2m = __lsx_vfmadd_s(__lsx_vfsub_s(_r01, _r02), _v4, __lsx_vfsub_s(_r04, _r03));
-                    __m128 _tmp3m = __lsx_vfmadd_s(__lsx_vfsub_s(_r01, _r03), _vm2, __lsx_vfsub_s(_r04, _r02));
-                    __m128 _tmp4m = __lsx_vfmadd_s(__lsx_vfsub_s(_r01, _r03), _v2, __lsx_vfsub_s(_r04, _r02));
-                    __m128 _tmp5m = __lsx_vfmadd_s(_r03, _vm5, __lsx_vfmadd_s(_r01, _v4, _r05));
-
-                    __lsx_vst(_tmp0m, tmp[0][m], 0);
-                    __lsx_vst(_tmp1m, tmp[1][m], 0);
-                    __lsx_vst(_tmp2m, tmp[2][m], 0);
-                    __lsx_vst(_tmp3m, tmp[3][m], 0);
-                    __lsx_vst(_tmp4m, tmp[4][m], 0);
-                    __lsx_vst(_tmp5m, tmp[5][m], 0);
-
-                    r0 += w * 4;
-                }
-
-                float* r0_tm_0 = (float*)img0_tm + (i * w_tiles + j) * 4;
-                float* r0_tm_1 = r0_tm_0 + tiles * 4;
-                float* r0_tm_2 = r0_tm_0 + tiles * 4 * 2;
-                float* r0_tm_3 = r0_tm_0 + tiles * 4 * 3;
-                float* r0_tm_4 = r0_tm_0 + tiles * 4 * 4;
-                float* r0_tm_5 = r0_tm_0 + tiles * 4 * 5;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    __m128 _tmp00 = (__m128)__lsx_vld(tmp[m][0], 0);
-                    __m128 _tmp01 = (__m128)__lsx_vld(tmp[m][1], 0);
-                    __m128 _tmp02 = (__m128)__lsx_vld(tmp[m][2], 0);
-                    __m128 _tmp03 = (__m128)__lsx_vld(tmp[m][3], 0);
-                    __m128 _tmp04 = (__m128)__lsx_vld(tmp[m][4], 0);
-                    __m128 _tmp05 = (__m128)__lsx_vld(tmp[m][5], 0);
-
-                    __m128 _r0tm0 = __lsx_vfmadd_s(_tmp02, _vm5, __lsx_vfmadd_s(_tmp00, _v4, _tmp04));
-                    __m128 _r0tm1 = __lsx_vfmadd_s(__lsx_vfadd_s(_tmp01, _tmp02), _vm4, __lsx_vfadd_s(_tmp04, _tmp03));
-                    __m128 _r0tm2 = __lsx_vfmadd_s(__lsx_vfsub_s(_tmp01, _tmp02), _v4, __lsx_vfsub_s(_tmp04, _tmp03));
-                    __m128 _r0tm3 = __lsx_vfmadd_s(__lsx_vfsub_s(_tmp01, _tmp03), _vm2, __lsx_vfsub_s(_tmp04, _tmp02));
-                    __m128 _r0tm4 = __lsx_vfmadd_s(__lsx_vfsub_s(_tmp01, _tmp03), _v2, __lsx_vfsub_s(_tmp04, _tmp02));
-                    __m128 _r0tm5 = __lsx_vfmadd_s(_tmp03, _vm5, __lsx_vfmadd_s(_tmp01, _v4, _tmp05));
-
-                    __lsx_vst(_r0tm0, r0_tm_0, 0);
-                    __lsx_vst(_r0tm1, r0_tm_1, 0);
-                    __lsx_vst(_r0tm2, r0_tm_2, 0);
-                    __lsx_vst(_r0tm3, r0_tm_3, 0);
-                    __lsx_vst(_r0tm4, r0_tm_4, 0);
-                    __lsx_vst(_r0tm5, r0_tm_5, 0);
-
-                    r0_tm_0 += tiles * 4 * 6;
-                    r0_tm_1 += tiles * 4 * 6;
-                    r0_tm_2 += tiles * 4 * 6;
-                    r0_tm_3 += tiles * 4 * 6;
-                    r0_tm_4 += tiles * 4 * 6;
-                    r0_tm_5 += tiles * 4 * 6;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_output_pack4_lsx(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 4;
-    const int h_tiles = outh / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    const float* biasptr = bias;
-
-    // const float otm[4][6] = {
-    //     {1.0f, 1.0f,  1.0f, 1.0f,  1.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 2.0f, -2.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f, 4.0f,  4.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 8.0f, -8.0f, 1.0f}
-    // };
-
-    // 0 = r00 + (r01 + r02) + (r03 + r04)
-    // 1 =       (r01 - r02) + (r03 - r04) * 2
-    // 2 =       (r01 + r02) + (r03 + r04) * 4
-    // 3 = r05 + (r01 - r02) + (r03 - r04) * 8
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        __m128 _bias0 = biasptr ? (__m128)__lsx_vld(biasptr + p * 4, 0) : (__m128)__lsx_vreplgr2vr_w(0);
-
-        float tmp[4][6][4];
-
-        __m128 _v2 = __lsx_vreplfr2vr_s(2.f);
-        __m128 _v4 = __lsx_vreplfr2vr_s(4.f);
-        __m128 _v8 = __lsx_vreplfr2vr_s(8.f);
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* output0_tm_0 = (const float*)out0_tm + (i * w_tiles + j) * 4;
-                const float* output0_tm_1 = output0_tm_0 + tiles * 4;
-                const float* output0_tm_2 = output0_tm_0 + tiles * 4 * 2;
-                const float* output0_tm_3 = output0_tm_0 + tiles * 4 * 3;
-                const float* output0_tm_4 = output0_tm_0 + tiles * 4 * 4;
-                const float* output0_tm_5 = output0_tm_0 + tiles * 4 * 5;
-
-                float* output0 = out0.row<float>(i * 4) + (j * 4) * 4;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    __m128 _out0tm0 = (__m128)__lsx_vld(output0_tm_0, 0);
-                    __m128 _out0tm1 = (__m128)__lsx_vld(output0_tm_1, 0);
-                    __m128 _out0tm2 = (__m128)__lsx_vld(output0_tm_2, 0);
-                    __m128 _out0tm3 = (__m128)__lsx_vld(output0_tm_3, 0);
-                    __m128 _out0tm4 = (__m128)__lsx_vld(output0_tm_4, 0);
-                    __m128 _out0tm5 = (__m128)__lsx_vld(output0_tm_5, 0);
-
-                    __m128 _tmp02a = __lsx_vfadd_s(_out0tm1, _out0tm2);
-                    __m128 _tmp13a = __lsx_vfsub_s(_out0tm1, _out0tm2);
-
-                    __m128 _tmp02b = __lsx_vfadd_s(_out0tm3, _out0tm4);
-                    __m128 _tmp13b = __lsx_vfsub_s(_out0tm3, _out0tm4);
-
-                    __m128 _tmp0m = __lsx_vfadd_s(__lsx_vfadd_s(_out0tm0, _tmp02a), _tmp02b);
-                    __m128 _tmp1m = __lsx_vfmadd_s(_tmp13b, _v2, _tmp13a);
-                    __m128 _tmp2m = __lsx_vfmadd_s(_tmp02b, _v4, _tmp02a);
-                    __m128 _tmp3m = __lsx_vfmadd_s(_tmp13b, _v8, __lsx_vfadd_s(_out0tm5, _tmp13a));
-
-                    __lsx_vst(_tmp0m, tmp[0][m], 0);
-                    __lsx_vst(_tmp1m, tmp[1][m], 0);
-                    __lsx_vst(_tmp2m, tmp[2][m], 0);
-                    __lsx_vst(_tmp3m, tmp[3][m], 0);
-
-                    output0_tm_0 += tiles * 4 * 6;
-                    output0_tm_1 += tiles * 4 * 6;
-                    output0_tm_2 += tiles * 4 * 6;
-                    output0_tm_3 += tiles * 4 * 6;
-                    output0_tm_4 += tiles * 4 * 6;
-                    output0_tm_5 += tiles * 4 * 6;
-                }
-
-                for (int m = 0; m < 4; m++)
-                {
-                    __m128 _tmp00 = (__m128)__lsx_vld(tmp[m][0], 0);
-                    __m128 _tmp01 = (__m128)__lsx_vld(tmp[m][1], 0);
-                    __m128 _tmp02 = (__m128)__lsx_vld(tmp[m][2], 0);
-                    __m128 _tmp03 = (__m128)__lsx_vld(tmp[m][3], 0);
-                    __m128 _tmp04 = (__m128)__lsx_vld(tmp[m][4], 0);
-                    __m128 _tmp05 = (__m128)__lsx_vld(tmp[m][5], 0);
-
-                    __m128 _tmp02a = __lsx_vfadd_s(_tmp01, _tmp02);
-                    __m128 _tmp13a = __lsx_vfsub_s(_tmp01, _tmp02);
-
-                    __m128 _tmp02b = __lsx_vfadd_s(_tmp03, _tmp04);
-                    __m128 _tmp13b = __lsx_vfsub_s(_tmp03, _tmp04);
-
-                    __m128 _out00 = __lsx_vfadd_s(_bias0, __lsx_vfadd_s(__lsx_vfadd_s(_tmp00, _tmp02a), _tmp02b));
-                    __m128 _out01 = __lsx_vfadd_s(_bias0, __lsx_vfmadd_s(_tmp13b, _v2, _tmp13a));
-                    __m128 _out02 = __lsx_vfadd_s(_bias0, __lsx_vfmadd_s(_tmp02b, _v4, _tmp02a));
-                    __m128 _out03 = __lsx_vfadd_s(_bias0, __lsx_vfmadd_s(_tmp13b, _v8, __lsx_vfadd_s(_tmp05, _tmp13a)));
-
-                    __lsx_vst(_out00, output0, 0);
-                    __lsx_vst(_out01, output0 + 4, 0);
-                    __lsx_vst(_out02, output0 + 4 * 2, 0);
-                    __lsx_vst(_out03, output0 + 4 * 3, 0);
-
-                    output0 += outw * 4;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_input_pack4_lsx(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 2;
-    const int h_tiles = (h - 2) / 2;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[4][4] = {
-    //     {1.0f,  0.0f, -1.0f,  0.0f},
-    //     {0.0f,  1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  0.00f, 1.0f}
-    // };
-
-    // 0 = r00 - r02
-    // 1 = r01 + r02
-    // 2 = r02 - r01
-    // 3 = r03 - r01
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        float tmp[4][4][4];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* r0 = img0.row(i * 2) + (j * 2) * 4;
-
-                for (int m = 0; m < 4; m++)
-                {
-                    __m128 _r00 = (__m128)__lsx_vld(r0, 0);
-                    __m128 _r01 = (__m128)__lsx_vld(r0 + 4, 0);
-                    __m128 _r02 = (__m128)__lsx_vld(r0 + 4 * 2, 0);
-                    __m128 _r03 = (__m128)__lsx_vld(r0 + 4 * 3, 0);
-
-                    __m128 _tmp0m = __lsx_vfsub_s(_r00, _r02);
-                    __m128 _tmp1m = __lsx_vfadd_s(_r01, _r02);
-                    __m128 _tmp2m = __lsx_vfsub_s(_r02, _r01);
-                    __m128 _tmp3m = __lsx_vfsub_s(_r03, _r01);
-
-                    __lsx_vst(_tmp0m, tmp[0][m], 0);
-                    __lsx_vst(_tmp1m, tmp[1][m], 0);
-                    __lsx_vst(_tmp2m, tmp[2][m], 0);
-                    __lsx_vst(_tmp3m, tmp[3][m], 0);
-
-                    r0 += w * 4;
-                }
-
-                float* r0_tm_0 = (float*)img0_tm + (i * w_tiles + j) * 4;
-                float* r0_tm_1 = r0_tm_0 + tiles * 4;
-                float* r0_tm_2 = r0_tm_0 + tiles * 4 * 2;
-                float* r0_tm_3 = r0_tm_0 + tiles * 4 * 3;
-
-                for (int m = 0; m < 4; m++)
-                {
-                    __m128 _tmp00 = (__m128)__lsx_vld(tmp[m][0], 0);
-                    __m128 _tmp01 = (__m128)__lsx_vld(tmp[m][1], 0);
-                    __m128 _tmp02 = (__m128)__lsx_vld(tmp[m][2], 0);
-                    __m128 _tmp03 = (__m128)__lsx_vld(tmp[m][3], 0);
-
-                    __m128 _r0tm0 = __lsx_vfsub_s(_tmp00, _tmp02);
-                    __m128 _r0tm1 = __lsx_vfadd_s(_tmp01, _tmp02);
-                    __m128 _r0tm2 = __lsx_vfsub_s(_tmp02, _tmp01);
-                    __m128 _r0tm3 = __lsx_vfsub_s(_tmp03, _tmp01);
-
-                    __lsx_vst(_r0tm0, r0_tm_0, 0);
-                    __lsx_vst(_r0tm1, r0_tm_1, 0);
-                    __lsx_vst(_r0tm2, r0_tm_2, 0);
-                    __lsx_vst(_r0tm3, r0_tm_3, 0);
-
-                    r0_tm_0 += tiles * 4 * 4;
-                    r0_tm_1 += tiles * 4 * 4;
-                    r0_tm_2 += tiles * 4 * 4;
-                    r0_tm_3 += tiles * 4 * 4;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_output_pack4_lsx(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 2;
-    const int h_tiles = outh / 2;
-    const int tiles = w_tiles * h_tiles;
-
-    const float* biasptr = bias;
-
-    // const float otm[2][4] = {
-    //     {1.0f,  1.0f,  1.0f,  0.0f},
-    //     {0.0f,  1.0f, -1.0f,  1.0f}
-    // };
-
-    // 0 = r00 + r01 + r02
-    // 1 = r01 - r02 + r03
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        __m128 _bias0 = biasptr ? (__m128)__lsx_vld(biasptr + p * 4, 0) : (__m128)__lsx_vreplgr2vr_w(0);
-
-        float tmp[2][4][4];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* output0_tm_0 = (const float*)out0_tm + (i * w_tiles + j) * 4;
-                const float* output0_tm_1 = output0_tm_0 + tiles * 4;
-                const float* output0_tm_2 = output0_tm_0 + tiles * 4 * 2;
-                const float* output0_tm_3 = output0_tm_0 + tiles * 4 * 3;
-
-                float* output0 = out0.row<float>(i * 2) + (j * 2) * 4;
-
-                for (int m = 0; m < 4; m++)
-                {
-                    __m128 _out0tm0 = (__m128)__lsx_vld(output0_tm_0, 0);
-                    __m128 _out0tm1 = (__m128)__lsx_vld(output0_tm_1, 0);
-                    __m128 _out0tm2 = (__m128)__lsx_vld(output0_tm_2, 0);
-                    __m128 _out0tm3 = (__m128)__lsx_vld(output0_tm_3, 0);
-
-                    __m128 _tmp0m = __lsx_vfadd_s(__lsx_vfadd_s(_out0tm0, _out0tm1), _out0tm2);
-                    __m128 _tmp1m = __lsx_vfadd_s(__lsx_vfsub_s(_out0tm1, _out0tm2), _out0tm3);
-
-                    __lsx_vst(_tmp0m, tmp[0][m], 0);
-                    __lsx_vst(_tmp1m, tmp[1][m], 0);
-
-                    output0_tm_0 += tiles * 4 * 4;
-                    output0_tm_1 += tiles * 4 * 4;
-                    output0_tm_2 += tiles * 4 * 4;
-                    output0_tm_3 += tiles * 4 * 4;
-                }
-
-                for (int m = 0; m < 2; m++)
-                {
-                    __m128 _tmp00 = (__m128)__lsx_vld(tmp[m][0], 0);
-                    __m128 _tmp01 = (__m128)__lsx_vld(tmp[m][1], 0);
-                    __m128 _tmp02 = (__m128)__lsx_vld(tmp[m][2], 0);
-                    __m128 _tmp03 = (__m128)__lsx_vld(tmp[m][3], 0);
-
-                    __m128 _out00 = __lsx_vfadd_s(_bias0, __lsx_vfadd_s(__lsx_vfadd_s(_tmp00, _tmp01), _tmp02));
-                    __m128 _out01 = __lsx_vfadd_s(_bias0, __lsx_vfadd_s(__lsx_vfsub_s(_tmp01, _tmp02), _tmp03));
-
-                    __lsx_vst(_out00, output0, 0);
-                    __lsx_vst(_out01, output0 + 4, 0);
-
-                    output0 += outw * 4;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_pack4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_pack4_int8.h
deleted file mode 100644
index 8b31ce97..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_pack4_int8.h
+++ /dev/null
@@ -1,166 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd43_transform_output_pack4_int8_lsx(const Mat& top_blob_tm, Mat& top_blob, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 4;
-    const int h_tiles = outh / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float otm[4][6] = {
-    //     {1.0f, 1.0f,  1.0f, 1.0f,  1.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 2.0f, -2.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f, 4.0f,  4.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 8.0f, -8.0f, 1.0f}
-    // };
-
-    // 0 = r00 + (r01 + r02) + (r03 + r04)
-    // 1 =       (r01 - r02) + (r03 - r04) * 2
-    // 2 =       (r01 + r02) + (r03 + r04) * 4
-    // 3 = r05 + (r01 - r02) + (r03 - r04) * 8
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        int tmp[4][6][4];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const int* output0_tm_0 = (const int*)out0_tm + (i * w_tiles + j) * 4;
-                const int* output0_tm_1 = output0_tm_0 + tiles * 4;
-                const int* output0_tm_2 = output0_tm_0 + tiles * 8;
-                const int* output0_tm_3 = output0_tm_0 + tiles * 12;
-                const int* output0_tm_4 = output0_tm_0 + tiles * 16;
-                const int* output0_tm_5 = output0_tm_0 + tiles * 20;
-
-                int* output0 = out0.row<int>(i * 4) + (j * 4) * 4;
-
-                for (int m = 0; m < 5; m++)
-                {
-                    __m128i _out0tm0 = __lsx_vld(output0_tm_0, 0);
-                    __m128i _out0tm1 = __lsx_vld(output0_tm_1, 0);
-                    __m128i _out0tm2 = __lsx_vld(output0_tm_2, 0);
-                    __m128i _out0tm3 = __lsx_vld(output0_tm_3, 0);
-                    __m128i _out0tm4 = __lsx_vld(output0_tm_4, 0);
-                    __m128i _out0tm5 = __lsx_vld(output0_tm_5, 0);
-
-                    __m128i _tmp02a = __lsx_vadd_w(_out0tm1, _out0tm2);
-                    __m128i _tmp13a = __lsx_vsub_w(_out0tm1, _out0tm2);
-
-                    __m128i _tmp02b = __lsx_vadd_w(_out0tm3, _out0tm4);
-                    __m128i _tmp13b = __lsx_vsub_w(_out0tm3, _out0tm4);
-
-                    __m128i _tmp0m = __lsx_vadd_w(__lsx_vadd_w(_out0tm0, _tmp02a), _tmp02b);
-                    __m128i _tmp1m = __lsx_vadd_w(_tmp13a, __lsx_vslli_w(_tmp13b, 1));
-                    __m128i _tmp2m = __lsx_vadd_w(_tmp02a, __lsx_vslli_w(_tmp02b, 2));
-                    __m128i _tmp3m = __lsx_vadd_w(__lsx_vadd_w(_tmp13a, __lsx_vslli_w(_out0tm5, 2)), __lsx_vslli_w(_tmp13b, 3));
-
-                    __lsx_vst(_tmp0m, tmp[0][m], 0);
-                    __lsx_vst(_tmp1m, tmp[1][m], 0);
-                    __lsx_vst(_tmp2m, tmp[2][m], 0);
-                    __lsx_vst(_tmp3m, tmp[3][m], 0);
-
-                    output0_tm_0 += tiles * 24;
-                    output0_tm_1 += tiles * 24;
-                    output0_tm_2 += tiles * 24;
-                    output0_tm_3 += tiles * 24;
-                    output0_tm_4 += tiles * 24;
-                    output0_tm_5 += tiles * 24;
-                }
-                for (int m = 5; m < 6; m++)
-                {
-                    __m128i _out0tm0 = __lsx_vld(output0_tm_0, 0);
-                    __m128i _out0tm1 = __lsx_vld(output0_tm_1, 0);
-                    __m128i _out0tm2 = __lsx_vld(output0_tm_2, 0);
-                    __m128i _out0tm3 = __lsx_vld(output0_tm_3, 0);
-                    __m128i _out0tm4 = __lsx_vld(output0_tm_4, 0);
-                    __m128i _out0tm5 = __lsx_vld(output0_tm_5, 0);
-
-                    __m128i _tmp02a = __lsx_vadd_w(_out0tm1, _out0tm2);
-                    __m128i _tmp13a = __lsx_vsub_w(_out0tm1, _out0tm2);
-
-                    __m128i _tmp02b = __lsx_vadd_w(_out0tm3, _out0tm4);
-                    __m128i _tmp13b = __lsx_vsub_w(_out0tm3, _out0tm4);
-
-                    __m128i _tmp0m = __lsx_vadd_w(__lsx_vadd_w(_out0tm0, _tmp02a), _tmp02b);
-                    __m128i _tmp1m = __lsx_vadd_w(_tmp13a, __lsx_vslli_w(_tmp13b, 1));
-                    __m128i _tmp2m = __lsx_vadd_w(_tmp02a, __lsx_vslli_w(_tmp02b, 2));
-                    __m128i _tmp3m = __lsx_vadd_w(__lsx_vadd_w(_tmp13a, __lsx_vslli_w(_out0tm5, 2)), __lsx_vslli_w(_tmp13b, 3));
-
-                    _tmp0m = __lsx_vslli_w(_tmp0m, 2);
-                    _tmp1m = __lsx_vslli_w(_tmp1m, 2);
-                    _tmp2m = __lsx_vslli_w(_tmp2m, 2);
-                    _tmp3m = __lsx_vslli_w(_tmp3m, 2);
-
-                    __lsx_vst(_tmp0m, tmp[0][m], 0);
-                    __lsx_vst(_tmp1m, tmp[1][m], 0);
-                    __lsx_vst(_tmp2m, tmp[2][m], 0);
-                    __lsx_vst(_tmp3m, tmp[3][m], 0);
-
-                    output0_tm_0 += tiles * 24;
-                    output0_tm_1 += tiles * 24;
-                    output0_tm_2 += tiles * 24;
-                    output0_tm_3 += tiles * 24;
-                    output0_tm_4 += tiles * 24;
-                    output0_tm_5 += tiles * 24;
-                }
-
-                for (int m = 0; m < 4; m++)
-                {
-                    __m128i _tmp00 = __lsx_vld(tmp[m][0], 0);
-                    __m128i _tmp01 = __lsx_vld(tmp[m][1], 0);
-                    __m128i _tmp02 = __lsx_vld(tmp[m][2], 0);
-                    __m128i _tmp03 = __lsx_vld(tmp[m][3], 0);
-                    __m128i _tmp04 = __lsx_vld(tmp[m][4], 0);
-                    __m128i _tmp05 = __lsx_vld(tmp[m][5], 0);
-
-                    __m128i _tmp02a = __lsx_vadd_w(_tmp01, _tmp02);
-                    __m128i _tmp13a = __lsx_vsub_w(_tmp01, _tmp02);
-
-                    __m128i _tmp02b = __lsx_vadd_w(_tmp03, _tmp04);
-                    __m128i _tmp13b = __lsx_vsub_w(_tmp03, _tmp04);
-
-                    __m128i _out00 = __lsx_vadd_w(__lsx_vadd_w(_tmp00, _tmp02a), _tmp02b);
-                    __m128i _out01 = __lsx_vadd_w(_tmp13a, __lsx_vslli_w(_tmp13b, 1));
-                    __m128i _out02 = __lsx_vadd_w(_tmp02a, __lsx_vslli_w(_tmp02b, 2));
-                    __m128i _out03 = __lsx_vadd_w(__lsx_vadd_w(_tmp05, _tmp13a), __lsx_vslli_w(_tmp13b, 3));
-
-                    // TODO use integer trick for division by 576
-                    __m128 _v576 = __lsx_vreplfr2vr_s(1.0 / 576);
-                    _out00 = __lsx_vftint_w_s(__lsx_vfmul_s(__lsx_vffint_s_w(_out00), _v576));
-                    _out01 = __lsx_vftint_w_s(__lsx_vfmul_s(__lsx_vffint_s_w(_out01), _v576));
-                    _out02 = __lsx_vftint_w_s(__lsx_vfmul_s(__lsx_vffint_s_w(_out02), _v576));
-                    _out03 = __lsx_vftint_w_s(__lsx_vfmul_s(__lsx_vffint_s_w(_out03), _v576));
-
-                    __lsx_vst(_out00, output0, 0);
-                    __lsx_vst(_out01, output0 + 4, 0);
-                    __lsx_vst(_out02, output0 + 8, 0);
-                    __lsx_vst(_out03, output0 + 12, 0);
-
-                    output0 += outw * 4;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_pack8_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_pack8_int8.h
deleted file mode 100644
index 5e49a876..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolution_winograd_transform_pack8_int8.h
+++ /dev/null
@@ -1,132 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd43_transform_input_pack8_int8_lsx(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 4;
-    const int h_tiles = (h - 2) / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[6][6] = {
-    //     {4.0f, 0.0f, -5.0f, 0.0f, 1.0f, 0.0f},
-    //     {0.0f,-4.0f, -4.0f, 1.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f, -4.0f,-1.0f, 1.0f, 0.0f},
-    //     {0.0f,-2.0f, -1.0f, 2.0f, 1.0f, 0.0f},
-    //     {0.0f, 2.0f, -1.0f,-2.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f,  0.0f,-5.0f, 0.0f, 1.0f}
-    // };
-
-    // 0 =  4 * r00 - 5 * r02 + r04
-    // 1 = -4 * (r01 + r02) + r04 + r03
-    // 2 =  4 * (r01 - r02) + r04 - r03
-    // 3 = -2 * (r01 - r03) + r04 - r02
-    // 4 =  2 * (r01 - r03) + r04 - r02
-    // 5 =  4 * r01 - 5 * r03 + r05
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        short tmp[6][6][8];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const signed char* r0 = img0.row<const signed char>(i * 4) + (j * 4) * 8;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    __m128i _r00_01 = __lsx_vld(r0, 0);
-                    __m128i _r02_03 = __lsx_vld(r0 + 16, 0);
-                    __m128i _r04_05 = __lsx_vld(r0 + 32, 0);
-                    __m128i _extr0001 = __lsx_vslti_b(_r00_01, 0);
-                    __m128i _extr0203 = __lsx_vslti_b(_r02_03, 0);
-                    __m128i _extr0405 = __lsx_vslti_b(_r04_05, 0);
-                    __m128i _r00 = __lsx_vilvl_b(_extr0001, _r00_01);
-                    __m128i _r01 = __lsx_vilvh_b(_extr0001, _r00_01);
-                    __m128i _r02 = __lsx_vilvl_b(_extr0203, _r02_03);
-                    __m128i _r03 = __lsx_vilvh_b(_extr0203, _r02_03);
-                    __m128i _r04 = __lsx_vilvl_b(_extr0405, _r04_05);
-                    __m128i _r05 = __lsx_vilvh_b(_extr0405, _r04_05);
-
-                    __m128i _v5 = __lsx_vreplgr2vr_h(5);
-
-                    __m128i _tmp0m = __lsx_vsub_h(__lsx_vadd_h(__lsx_vslli_h(_r00, 2), _r04), __lsx_vmul_h(_r02, _v5));
-                    __m128i _tmp1m = __lsx_vsub_h(__lsx_vadd_h(_r04, _r03), __lsx_vslli_h(__lsx_vadd_h(_r01, _r02), 2));
-                    __m128i _tmp2m = __lsx_vadd_h(__lsx_vsub_h(_r04, _r03), __lsx_vslli_h(__lsx_vsub_h(_r01, _r02), 2));
-                    __m128i _tmp3m = __lsx_vsub_h(__lsx_vsub_h(_r04, _r02), __lsx_vslli_h(__lsx_vsub_h(_r01, _r03), 1));
-                    __m128i _tmp4m = __lsx_vadd_h(__lsx_vsub_h(_r04, _r02), __lsx_vslli_h(__lsx_vsub_h(_r01, _r03), 1));
-                    __m128i _tmp5m = __lsx_vsub_h(__lsx_vadd_h(__lsx_vslli_h(_r01, 2), _r05), __lsx_vmul_h(_r03, _v5));
-
-                    __lsx_vst(_tmp0m, tmp[0][m], 0);
-                    __lsx_vst(_tmp1m, tmp[1][m], 0);
-                    __lsx_vst(_tmp2m, tmp[2][m], 0);
-                    __lsx_vst(_tmp3m, tmp[3][m], 0);
-                    __lsx_vst(_tmp4m, tmp[4][m], 0);
-                    __lsx_vst(_tmp5m, tmp[5][m], 0);
-
-                    r0 += w * 8;
-                }
-
-                short* r0_tm_0 = (short*)img0_tm + (i * w_tiles + j) * 8;
-                short* r0_tm_1 = r0_tm_0 + tiles * 8;
-                short* r0_tm_2 = r0_tm_0 + tiles * 16;
-                short* r0_tm_3 = r0_tm_0 + tiles * 24;
-                short* r0_tm_4 = r0_tm_0 + tiles * 32;
-                short* r0_tm_5 = r0_tm_0 + tiles * 40;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    __m128i _tmp00 = __lsx_vld(tmp[m][0], 0);
-                    __m128i _tmp01 = __lsx_vld(tmp[m][1], 0);
-                    __m128i _tmp02 = __lsx_vld(tmp[m][2], 0);
-                    __m128i _tmp03 = __lsx_vld(tmp[m][3], 0);
-                    __m128i _tmp04 = __lsx_vld(tmp[m][4], 0);
-                    __m128i _tmp05 = __lsx_vld(tmp[m][5], 0);
-
-                    __m128i _v5 = __lsx_vreplgr2vr_h(5);
-
-                    __m128i _r0tm0 = __lsx_vsub_h(__lsx_vadd_h(__lsx_vslli_h(_tmp00, 2), _tmp04), __lsx_vmul_h(_tmp02, _v5));
-                    __m128i _r0tm1 = __lsx_vsub_h(__lsx_vadd_h(_tmp04, _tmp03), __lsx_vslli_h(__lsx_vadd_h(_tmp01, _tmp02), 2));
-                    __m128i _r0tm2 = __lsx_vadd_h(__lsx_vsub_h(_tmp04, _tmp03), __lsx_vslli_h(__lsx_vsub_h(_tmp01, _tmp02), 2));
-                    __m128i _r0tm3 = __lsx_vsub_h(__lsx_vsub_h(_tmp04, _tmp02), __lsx_vslli_h(__lsx_vsub_h(_tmp01, _tmp03), 1));
-                    __m128i _r0tm4 = __lsx_vadd_h(__lsx_vsub_h(_tmp04, _tmp02), __lsx_vslli_h(__lsx_vsub_h(_tmp01, _tmp03), 1));
-                    __m128i _r0tm5 = __lsx_vsub_h(__lsx_vadd_h(__lsx_vslli_h(_tmp01, 2), _tmp05), __lsx_vmul_h(_tmp03, _v5));
-
-                    __lsx_vst(_r0tm0, r0_tm_0, 0);
-                    __lsx_vst(_r0tm1, r0_tm_1, 0);
-                    __lsx_vst(_r0tm2, r0_tm_2, 0);
-                    __lsx_vst(_r0tm3, r0_tm_3, 0);
-                    __lsx_vst(_r0tm4, r0_tm_4, 0);
-                    __lsx_vst(_r0tm5, r0_tm_5, 0);
-
-                    r0_tm_0 += tiles * 48;
-                    r0_tm_1 += tiles * 48;
-                    r0_tm_2 += tiles * 48;
-                    r0_tm_3 += tiles * 48;
-                    r0_tm_4 += tiles * 48;
-                    r0_tm_5 += tiles * 48;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_3x3.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_3x3.h
deleted file mode 100644
index 1c37f778..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_3x3.h
+++ /dev/null
@@ -1,193 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const float bias0 = bias ? bias[g] : 0.f;
-
-        const float* kernel0 = kernel + g * 9;
-
-        float* outptr0 = out;
-        float* outptr1 = outptr0 + outw;
-
-        const float* img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0;
-        const float* r1 = img0 + w;
-        const float* r2 = img0 + w * 2;
-        const float* r3 = img0 + w * 3;
-
-        const float* k0 = kernel0;
-        const float* k1 = kernel0 + 3;
-        const float* k2 = kernel0 + 6;
-
-        int i = 0;
-
-        for (; i + 1 < outh; i += 2)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = bias0;
-                float sum2 = bias0;
-
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum2 += r1[0] * k0[0];
-                sum2 += r1[1] * k0[1];
-                sum2 += r1[2] * k0[2];
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum2 += r2[0] * k1[0];
-                sum2 += r2[1] * k1[1];
-                sum2 += r2[2] * k1[2];
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-                sum2 += r3[0] * k2[0];
-                sum2 += r3[1] * k2[1];
-                sum2 += r3[2] * k2[2];
-
-                *outptr0 = sum;
-                *outptr1 = sum2;
-
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                outptr0++;
-                outptr1++;
-            }
-
-            r0 += 2 + w;
-            r1 += 2 + w;
-            r2 += 2 + w;
-            r3 += 2 + w;
-
-            outptr0 += outw;
-            outptr1 += outw;
-        }
-
-        for (; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = bias0;
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-
-                *outptr0 = sum;
-
-                r0++;
-                r1++;
-                r2++;
-                outptr0++;
-            }
-
-            r0 += 2;
-            r1 += 2;
-            r2 += 2;
-        }
-    }
-}
-
-static void convdw3x3s2_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const float bias0 = bias ? bias[g] : 0.f;
-
-        const float* kernel0 = kernel + g * 9;
-
-        float* outptr = out;
-
-        const float* img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0;
-        const float* r1 = img0 + w;
-        const float* r2 = img0 + w * 2;
-
-        const float* k0 = kernel0;
-        const float* k1 = kernel0 + 3;
-        const float* k2 = kernel0 + 6;
-
-        int i = 0;
-
-        for (; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = bias0;
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-
-                *outptr = sum;
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-                outptr++;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_3x3_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_3x3_pack4.h
deleted file mode 100644
index 48ae6641..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_3x3_pack4.h
+++ /dev/null
@@ -1,464 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_pack4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m128 _bias0 = bias ? (__m128)__lsx_vld(bias + g * 4, 0) : (__m128)__lsx_vreplgr2vr_w(0);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-        float* outptr1 = out.row(1);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-
-        __m128 _k00 = (__m128)__lsx_vld(k0, 0);
-        __m128 _k01 = (__m128)__lsx_vld(k0 + 4, 0);
-        __m128 _k02 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-        __m128 _k10 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-        __m128 _k11 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-        __m128 _k12 = (__m128)__lsx_vld(k0 + 4 * 5, 0);
-        __m128 _k20 = (__m128)__lsx_vld(k0 + 4 * 6, 0);
-        __m128 _k21 = (__m128)__lsx_vld(k0 + 4 * 7, 0);
-        __m128 _k22 = (__m128)__lsx_vld(k0 + 4 * 8, 0);
-
-        int i = 0;
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                __builtin_prefetch(r0 + 32);
-                __builtin_prefetch(r1 + 32);
-                __builtin_prefetch(r2 + 32);
-                __builtin_prefetch(r3 + 32);
-
-                __m128 _sum00 = _bias0;
-                __m128 _sum01 = _bias0;
-                __m128 _sum10 = _bias0;
-                __m128 _sum11 = _bias0;
-
-                __m128 _r00 = (__m128)__lsx_vld(r0, 0);
-                __m128 _r01 = (__m128)__lsx_vld(r0 + 4, 0);
-                __m128 _r02 = (__m128)__lsx_vld(r0 + 4 * 2, 0);
-                __m128 _r03 = (__m128)__lsx_vld(r0 + 4 * 3, 0);
-
-                _sum00 = __lsx_vfmadd_s(_r00, _k00, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r01, _k01, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r02, _k02, _sum00);
-                _sum01 = __lsx_vfmadd_s(_r01, _k00, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r02, _k01, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r03, _k02, _sum01);
-
-                __m128 _r10 = (__m128)__lsx_vld(r1, 0);
-                __m128 _r11 = (__m128)__lsx_vld(r1 + 4, 0);
-                __m128 _r12 = (__m128)__lsx_vld(r1 + 4 * 2, 0);
-                __m128 _r13 = (__m128)__lsx_vld(r1 + 4 * 3, 0);
-
-                _sum00 = __lsx_vfmadd_s(_r10, _k10, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r11, _k11, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r12, _k12, _sum00);
-                _sum01 = __lsx_vfmadd_s(_r11, _k10, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r12, _k11, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r13, _k12, _sum01);
-                _sum10 = __lsx_vfmadd_s(_r10, _k00, _sum10);
-                _sum10 = __lsx_vfmadd_s(_r11, _k01, _sum10);
-                _sum10 = __lsx_vfmadd_s(_r12, _k02, _sum10);
-                _sum11 = __lsx_vfmadd_s(_r11, _k00, _sum11);
-                _sum11 = __lsx_vfmadd_s(_r12, _k01, _sum11);
-                _sum11 = __lsx_vfmadd_s(_r13, _k02, _sum11);
-
-                __m128 _r20 = (__m128)__lsx_vld(r2, 0);
-                __m128 _r21 = (__m128)__lsx_vld(r2 + 4, 0);
-                __m128 _r22 = (__m128)__lsx_vld(r2 + 4 * 2, 0);
-                __m128 _r23 = (__m128)__lsx_vld(r2 + 4 * 3, 0);
-
-                _sum00 = __lsx_vfmadd_s(_r20, _k20, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r21, _k21, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r22, _k22, _sum00);
-                _sum01 = __lsx_vfmadd_s(_r21, _k20, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r22, _k21, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r23, _k22, _sum01);
-                _sum10 = __lsx_vfmadd_s(_r20, _k10, _sum10);
-                _sum10 = __lsx_vfmadd_s(_r21, _k11, _sum10);
-                _sum10 = __lsx_vfmadd_s(_r22, _k12, _sum10);
-                _sum11 = __lsx_vfmadd_s(_r21, _k10, _sum11);
-                _sum11 = __lsx_vfmadd_s(_r22, _k11, _sum11);
-                _sum11 = __lsx_vfmadd_s(_r23, _k12, _sum11);
-
-                __m128 _r30 = (__m128)__lsx_vld(r3, 0);
-                __m128 _r31 = (__m128)__lsx_vld(r3 + 4, 0);
-                __m128 _r32 = (__m128)__lsx_vld(r3 + 4 * 2, 0);
-                __m128 _r33 = (__m128)__lsx_vld(r3 + 4 * 3, 0);
-
-                _sum10 = __lsx_vfmadd_s(_r30, _k20, _sum10);
-                _sum10 = __lsx_vfmadd_s(_r31, _k21, _sum10);
-                _sum10 = __lsx_vfmadd_s(_r32, _k22, _sum10);
-                _sum11 = __lsx_vfmadd_s(_r31, _k20, _sum11);
-                _sum11 = __lsx_vfmadd_s(_r32, _k21, _sum11);
-                _sum11 = __lsx_vfmadd_s(_r33, _k22, _sum11);
-
-                __lsx_vst(_sum00, outptr0, 0);
-                __lsx_vst(_sum01, outptr0 + 4, 0);
-                __lsx_vst(_sum10, outptr1, 0);
-                __lsx_vst(_sum11, outptr1 + 4, 0);
-
-                outptr0 += 4 * 2;
-                outptr1 += 4 * 2;
-
-                r0 += 4 * 2;
-                r1 += 4 * 2;
-                r2 += 4 * 2;
-                r3 += 4 * 2;
-            }
-            for (; j < outw; j++)
-            {
-                __builtin_prefetch(r0 + 16);
-                __builtin_prefetch(r1 + 16);
-                __builtin_prefetch(r2 + 16);
-                __builtin_prefetch(r3 + 16);
-
-                __m128 _sum0 = _bias0;
-                __m128 _sum1 = _bias0;
-
-                __m128 _r00 = (__m128)__lsx_vld(r0, 0);
-                __m128 _r01 = (__m128)__lsx_vld(r0 + 4, 0);
-                __m128 _r02 = (__m128)__lsx_vld(r0 + 4 * 2, 0);
-
-                _sum0 = __lsx_vfmadd_s(_r00, _k00, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r01, _k01, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r02, _k02, _sum0);
-
-                __m128 _r10 = (__m128)__lsx_vld(r1, 0);
-                __m128 _r11 = (__m128)__lsx_vld(r1 + 4, 0);
-                __m128 _r12 = (__m128)__lsx_vld(r1 + 4 * 2, 0);
-
-                _sum0 = __lsx_vfmadd_s(_r10, _k10, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r11, _k11, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r12, _k12, _sum0);
-                _sum1 = __lsx_vfmadd_s(_r10, _k00, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r11, _k01, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r12, _k02, _sum1);
-
-                __m128 _r20 = (__m128)__lsx_vld(r2, 0);
-                __m128 _r21 = (__m128)__lsx_vld(r2 + 4, 0);
-                __m128 _r22 = (__m128)__lsx_vld(r2 + 4 * 2, 0);
-
-                _sum0 = __lsx_vfmadd_s(_r20, _k20, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r21, _k21, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r22, _k22, _sum0);
-                _sum1 = __lsx_vfmadd_s(_r20, _k10, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r21, _k11, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r22, _k12, _sum1);
-
-                __m128 _r30 = (__m128)__lsx_vld(r3, 0);
-                __m128 _r31 = (__m128)__lsx_vld(r3 + 4, 0);
-                __m128 _r32 = (__m128)__lsx_vld(r3 + 4 * 2, 0);
-
-                _sum1 = __lsx_vfmadd_s(_r30, _k20, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r31, _k21, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r32, _k22, _sum1);
-
-                __lsx_vst(_sum0, outptr0, 0);
-                __lsx_vst(_sum1, outptr1, 0);
-
-                outptr0 += 4;
-                outptr1 += 4;
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                r3 += 4;
-            }
-
-            r0 += 2 * 4 + w * 4;
-            r1 += 2 * 4 + w * 4;
-            r2 += 2 * 4 + w * 4;
-            r3 += 2 * 4 + w * 4;
-
-            outptr0 += outw * 4;
-            outptr1 += outw * 4;
-        }
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                __builtin_prefetch(r0 + 32);
-                __builtin_prefetch(r1 + 32);
-                __builtin_prefetch(r2 + 32);
-
-                __m128 _sum00 = _bias0;
-                __m128 _sum01 = _bias0;
-
-                __m128 _r00 = (__m128)__lsx_vld(r0, 0);
-                __m128 _r01 = (__m128)__lsx_vld(r0 + 4, 0);
-                __m128 _r02 = (__m128)__lsx_vld(r0 + 4 * 2, 0);
-                __m128 _r03 = (__m128)__lsx_vld(r0 + 4 * 3, 0);
-
-                _sum00 = __lsx_vfmadd_s(_r00, _k00, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r01, _k01, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r02, _k02, _sum00);
-                _sum01 = __lsx_vfmadd_s(_r01, _k00, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r02, _k01, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r03, _k02, _sum01);
-
-                __m128 _r10 = (__m128)__lsx_vld(r1, 0);
-                __m128 _r11 = (__m128)__lsx_vld(r1 + 4, 0);
-                __m128 _r12 = (__m128)__lsx_vld(r1 + 4 * 2, 0);
-                __m128 _r13 = (__m128)__lsx_vld(r1 + 4 * 3, 0);
-
-                _sum00 = __lsx_vfmadd_s(_r10, _k10, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r11, _k11, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r12, _k12, _sum00);
-                _sum01 = __lsx_vfmadd_s(_r11, _k10, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r12, _k11, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r13, _k12, _sum01);
-
-                __m128 _r20 = (__m128)__lsx_vld(r2, 0);
-                __m128 _r21 = (__m128)__lsx_vld(r2 + 4, 0);
-                __m128 _r22 = (__m128)__lsx_vld(r2 + 4 * 2, 0);
-                __m128 _r23 = (__m128)__lsx_vld(r2 + 4 * 3, 0);
-
-                _sum00 = __lsx_vfmadd_s(_r20, _k20, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r21, _k21, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r22, _k22, _sum00);
-                _sum01 = __lsx_vfmadd_s(_r21, _k20, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r22, _k21, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r23, _k22, _sum01);
-
-                __lsx_vst(_sum00, outptr0, 0);
-                __lsx_vst(_sum01, outptr0 + 4, 0);
-
-                outptr0 += 4 * 2;
-
-                r0 += 4 * 2;
-                r1 += 4 * 2;
-                r2 += 4 * 2;
-            }
-            for (; j < outw; j++)
-            {
-                __builtin_prefetch(r0 + 16);
-                __builtin_prefetch(r1 + 16);
-                __builtin_prefetch(r2 + 16);
-
-                __m128 _sum0 = _bias0;
-
-                __m128 _r00 = (__m128)__lsx_vld(r0, 0);
-                __m128 _r01 = (__m128)__lsx_vld(r0 + 4, 0);
-                __m128 _r02 = (__m128)__lsx_vld(r0 + 4 * 2, 0);
-
-                _sum0 = __lsx_vfmadd_s(_r00, _k00, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r01, _k01, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r02, _k02, _sum0);
-
-                __m128 _r10 = (__m128)__lsx_vld(r1, 0);
-                __m128 _r11 = (__m128)__lsx_vld(r1 + 4, 0);
-                __m128 _r12 = (__m128)__lsx_vld(r1 + 4 * 2, 0);
-
-                _sum0 = __lsx_vfmadd_s(_r10, _k10, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r11, _k11, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r12, _k12, _sum0);
-
-                __m128 _r20 = (__m128)__lsx_vld(r2, 0);
-                __m128 _r21 = (__m128)__lsx_vld(r2 + 4, 0);
-                __m128 _r22 = (__m128)__lsx_vld(r2 + 4 * 2, 0);
-
-                _sum0 = __lsx_vfmadd_s(_r20, _k20, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r21, _k21, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r22, _k22, _sum0);
-
-                __lsx_vst(_sum0, outptr0, 0);
-
-                outptr0 += 4;
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-            }
-
-            r0 += 2 * 4;
-            r1 += 2 * 4;
-            r2 += 2 * 4;
-        }
-    }
-}
-
-static void convdw3x3s2_pack4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m128 _bias0 = bias ? (__m128)__lsx_vld(bias + g * 4, 0) : (__m128)__lsx_vreplgr2vr_w(0);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out;
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-
-        __m128 _k00 = (__m128)__lsx_vld(k0, 0);
-        __m128 _k01 = (__m128)__lsx_vld(k0 + 4, 0);
-        __m128 _k02 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-        __m128 _k10 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-        __m128 _k11 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-        __m128 _k12 = (__m128)__lsx_vld(k0 + 4 * 5, 0);
-        __m128 _k20 = (__m128)__lsx_vld(k0 + 4 * 6, 0);
-        __m128 _k21 = (__m128)__lsx_vld(k0 + 4 * 7, 0);
-        __m128 _k22 = (__m128)__lsx_vld(k0 + 4 * 8, 0);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                __builtin_prefetch(r0 + 64);
-                __builtin_prefetch(r1 + 64);
-                __builtin_prefetch(r2 + 64);
-
-                __m128 _sum00 = _bias0;
-                __m128 _sum01 = _bias0;
-
-                __m128 _r00 = (__m128)__lsx_vld(r0, 0);
-                __m128 _r01 = (__m128)__lsx_vld(r0 + 4, 0);
-                __m128 _r02 = (__m128)__lsx_vld(r0 + 4 * 2, 0);
-                __m128 _r03 = (__m128)__lsx_vld(r0 + 4 * 3, 0);
-                __m128 _r04 = (__m128)__lsx_vld(r0 + 4 * 4, 0);
-
-                _sum00 = __lsx_vfmadd_s(_r00, _k00, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r01, _k01, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r02, _k02, _sum00);
-                _sum01 = __lsx_vfmadd_s(_r02, _k00, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r03, _k01, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r04, _k02, _sum01);
-
-                __m128 _r10 = (__m128)__lsx_vld(r1, 0);
-                __m128 _r11 = (__m128)__lsx_vld(r1 + 4, 0);
-                __m128 _r12 = (__m128)__lsx_vld(r1 + 4 * 2, 0);
-                __m128 _r13 = (__m128)__lsx_vld(r1 + 4 * 3, 0);
-                __m128 _r14 = (__m128)__lsx_vld(r1 + 4 * 4, 0);
-
-                _sum00 = __lsx_vfmadd_s(_r10, _k10, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r11, _k11, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r12, _k12, _sum00);
-                _sum01 = __lsx_vfmadd_s(_r12, _k10, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r13, _k11, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r14, _k12, _sum01);
-
-                __m128 _r20 = (__m128)__lsx_vld(r2, 0);
-                __m128 _r21 = (__m128)__lsx_vld(r2 + 4, 0);
-                __m128 _r22 = (__m128)__lsx_vld(r2 + 4 * 2, 0);
-                __m128 _r23 = (__m128)__lsx_vld(r2 + 4 * 3, 0);
-                __m128 _r24 = (__m128)__lsx_vld(r2 + 4 * 4, 0);
-
-                _sum00 = __lsx_vfmadd_s(_r20, _k20, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r21, _k21, _sum00);
-                _sum00 = __lsx_vfmadd_s(_r22, _k22, _sum00);
-                _sum01 = __lsx_vfmadd_s(_r22, _k20, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r23, _k21, _sum01);
-                _sum01 = __lsx_vfmadd_s(_r24, _k22, _sum01);
-
-                __lsx_vst(_sum00, outptr0, 0);
-                __lsx_vst(_sum01, outptr0 + 4, 0);
-
-                outptr0 += 4 * 2;
-
-                r0 += 4 * 4;
-                r1 += 4 * 4;
-                r2 += 4 * 4;
-            }
-            for (; j < outw; j++)
-            {
-                __builtin_prefetch(r0 + 32);
-                __builtin_prefetch(r1 + 32);
-                __builtin_prefetch(r2 + 32);
-
-                __m128 _sum0 = _bias0;
-
-                __m128 _r00 = (__m128)__lsx_vld(r0, 0);
-                __m128 _r01 = (__m128)__lsx_vld(r0 + 4, 0);
-                __m128 _r02 = (__m128)__lsx_vld(r0 + 4 * 2, 0);
-
-                _sum0 = __lsx_vfmadd_s(_r00, _k00, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r01, _k01, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r02, _k02, _sum0);
-
-                __m128 _r10 = (__m128)__lsx_vld(r1, 0);
-                __m128 _r11 = (__m128)__lsx_vld(r1 + 4, 0);
-                __m128 _r12 = (__m128)__lsx_vld(r1 + 4 * 2, 0);
-
-                _sum0 = __lsx_vfmadd_s(_r10, _k10, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r11, _k11, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r12, _k12, _sum0);
-
-                __m128 _r20 = (__m128)__lsx_vld(r2, 0);
-                __m128 _r21 = (__m128)__lsx_vld(r2 + 4, 0);
-                __m128 _r22 = (__m128)__lsx_vld(r2 + 4 * 2, 0);
-
-                _sum0 = __lsx_vfmadd_s(_r20, _k20, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r21, _k21, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r22, _k22, _sum0);
-
-                __lsx_vst(_sum0, outptr0, 0);
-
-                outptr0 += 4;
-
-                r0 += 4 * 2;
-                r1 += 4 * 2;
-                r2 += 4 * 2;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_5x5_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_5x5_pack4.h
deleted file mode 100644
index 4f94c5e6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_5x5_pack4.h
+++ /dev/null
@@ -1,511 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw5x5s1_pack4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m128 _bias0 = bias ? (__m128)__lsx_vld(bias + g * 4, 0) : (__m128)__lsx_vreplgr2vr_w(0);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-        float* outptr1 = out.row(1);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-        const float* r4 = img0.row(4);
-        const float* r5 = img0.row(5);
-
-        int i = 0;
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                __builtin_prefetch(r0 + 16);
-                __builtin_prefetch(r1 + 16);
-                __builtin_prefetch(r2 + 16);
-                __builtin_prefetch(r3 + 16);
-                __builtin_prefetch(r4 + 16);
-                __builtin_prefetch(r5 + 16);
-
-                __builtin_prefetch(k0 + 400);
-
-                __m128 _sum0 = _bias0;
-                __m128 _sum1 = _bias0;
-
-                __m128 _r00 = (__m128)__lsx_vld(r0, 0);
-                __m128 _r01 = (__m128)__lsx_vld(r0 + 4, 0);
-                __m128 _r02 = (__m128)__lsx_vld(r0 + 4 * 2, 0);
-                __m128 _r03 = (__m128)__lsx_vld(r0 + 4 * 3, 0);
-                __m128 _r04 = (__m128)__lsx_vld(r0 + 4 * 4, 0);
-
-                __m128 _k00 = (__m128)__lsx_vld(k0, 0);
-                __m128 _k01 = (__m128)__lsx_vld(k0 + 4, 0);
-                __m128 _k02 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-                __m128 _k03 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-                __m128 _k04 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __lsx_vfmadd_s(_r00, _k00, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r01, _k01, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r02, _k02, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r03, _k03, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r04, _k04, _sum0);
-
-                __m128 _r10 = (__m128)__lsx_vld(r1, 0);
-                __m128 _r11 = (__m128)__lsx_vld(r1 + 4, 0);
-                __m128 _r12 = (__m128)__lsx_vld(r1 + 4 * 2, 0);
-                __m128 _r13 = (__m128)__lsx_vld(r1 + 4 * 3, 0);
-                __m128 _r14 = (__m128)__lsx_vld(r1 + 4 * 4, 0);
-
-                _sum1 = __lsx_vfmadd_s(_r10, _k00, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r11, _k01, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r12, _k02, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r13, _k03, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r14, _k04, _sum1);
-
-                __m128 _k10 = (__m128)__lsx_vld(k0, 0);
-                __m128 _k11 = (__m128)__lsx_vld(k0 + 4, 0);
-                __m128 _k12 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-                __m128 _k13 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-                __m128 _k14 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __lsx_vfmadd_s(_r10, _k10, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r11, _k11, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r12, _k12, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r13, _k13, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r14, _k14, _sum0);
-
-                __m128 _r20 = (__m128)__lsx_vld(r2, 0);
-                __m128 _r21 = (__m128)__lsx_vld(r2 + 4, 0);
-                __m128 _r22 = (__m128)__lsx_vld(r2 + 4 * 2, 0);
-                __m128 _r23 = (__m128)__lsx_vld(r2 + 4 * 3, 0);
-                __m128 _r24 = (__m128)__lsx_vld(r2 + 4 * 4, 0);
-
-                _sum1 = __lsx_vfmadd_s(_r20, _k10, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r21, _k11, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r22, _k12, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r23, _k13, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r24, _k14, _sum1);
-
-                __m128 _k20 = (__m128)__lsx_vld(k0, 0);
-                __m128 _k21 = (__m128)__lsx_vld(k0 + 4, 0);
-                __m128 _k22 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-                __m128 _k23 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-                __m128 _k24 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __lsx_vfmadd_s(_r20, _k20, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r21, _k21, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r22, _k22, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r23, _k23, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r24, _k24, _sum0);
-
-                __m128 _r30 = (__m128)__lsx_vld(r3, 0);
-                __m128 _r31 = (__m128)__lsx_vld(r3 + 4, 0);
-                __m128 _r32 = (__m128)__lsx_vld(r3 + 4 * 2, 0);
-                __m128 _r33 = (__m128)__lsx_vld(r3 + 4 * 3, 0);
-                __m128 _r34 = (__m128)__lsx_vld(r3 + 4 * 4, 0);
-
-                _sum1 = __lsx_vfmadd_s(_r30, _k20, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r31, _k21, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r32, _k22, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r33, _k23, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r34, _k24, _sum1);
-
-                __m128 _k30 = (__m128)__lsx_vld(k0, 0);
-                __m128 _k31 = (__m128)__lsx_vld(k0 + 4, 0);
-                __m128 _k32 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-                __m128 _k33 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-                __m128 _k34 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __lsx_vfmadd_s(_r30, _k30, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r31, _k31, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r32, _k32, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r33, _k33, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r34, _k34, _sum0);
-
-                __m128 _r40 = (__m128)__lsx_vld(r4, 0);
-                __m128 _r41 = (__m128)__lsx_vld(r4 + 4, 0);
-                __m128 _r42 = (__m128)__lsx_vld(r4 + 4 * 2, 0);
-                __m128 _r43 = (__m128)__lsx_vld(r4 + 4 * 3, 0);
-                __m128 _r44 = (__m128)__lsx_vld(r4 + 4 * 4, 0);
-
-                _sum1 = __lsx_vfmadd_s(_r40, _k30, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r41, _k31, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r42, _k32, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r43, _k33, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r44, _k34, _sum1);
-
-                __m128 _k40 = (__m128)__lsx_vld(k0, 0);
-                __m128 _k41 = (__m128)__lsx_vld(k0 + 4, 0);
-                __m128 _k42 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-                __m128 _k43 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-                __m128 _k44 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-                k0 -= 4 * 20;
-
-                _sum0 = __lsx_vfmadd_s(_r40, _k40, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r41, _k41, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r42, _k42, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r43, _k43, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r44, _k44, _sum0);
-
-                __m128 _r50 = (__m128)__lsx_vld(r5, 0);
-                __m128 _r51 = (__m128)__lsx_vld(r5 + 4, 0);
-                __m128 _r52 = (__m128)__lsx_vld(r5 + 4 * 2, 0);
-                __m128 _r53 = (__m128)__lsx_vld(r5 + 4 * 3, 0);
-                __m128 _r54 = (__m128)__lsx_vld(r5 + 4 * 4, 0);
-
-                _sum1 = __lsx_vfmadd_s(_r50, _k40, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r51, _k41, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r52, _k42, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r53, _k43, _sum1);
-                _sum1 = __lsx_vfmadd_s(_r54, _k44, _sum1);
-
-                __lsx_vst(_sum0, outptr0, 0);
-                __lsx_vst(_sum1, outptr1, 0);
-
-                outptr0 += 4;
-                outptr1 += 4;
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                r3 += 4;
-                r4 += 4;
-                r5 += 4;
-            }
-
-            r0 += 4 * 4 + w * 4;
-            r1 += 4 * 4 + w * 4;
-            r2 += 4 * 4 + w * 4;
-            r3 += 4 * 4 + w * 4;
-            r4 += 4 * 4 + w * 4;
-            r5 += 4 * 4 + w * 4;
-
-            outptr0 += outw * 4;
-            outptr1 += outw * 4;
-        }
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                __builtin_prefetch(r0 + 16);
-                __builtin_prefetch(r1 + 16);
-                __builtin_prefetch(r2 + 16);
-                __builtin_prefetch(r3 + 16);
-                __builtin_prefetch(r4 + 16);
-
-                __builtin_prefetch(k0 + 400);
-
-                __m128 _sum0 = _bias0;
-
-                __m128 _r00 = (__m128)__lsx_vld(r0, 0);
-                __m128 _r01 = (__m128)__lsx_vld(r0 + 4, 0);
-                __m128 _r02 = (__m128)__lsx_vld(r0 + 4 * 2, 0);
-                __m128 _r03 = (__m128)__lsx_vld(r0 + 4 * 3, 0);
-                __m128 _r04 = (__m128)__lsx_vld(r0 + 4 * 4, 0);
-
-                __m128 _k00 = (__m128)__lsx_vld(k0, 0);
-                __m128 _k01 = (__m128)__lsx_vld(k0 + 4, 0);
-                __m128 _k02 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-                __m128 _k03 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-                __m128 _k04 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __lsx_vfmadd_s(_r00, _k00, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r01, _k01, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r02, _k02, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r03, _k03, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r04, _k04, _sum0);
-
-                __m128 _r10 = (__m128)__lsx_vld(r1, 0);
-                __m128 _r11 = (__m128)__lsx_vld(r1 + 4, 0);
-                __m128 _r12 = (__m128)__lsx_vld(r1 + 4 * 2, 0);
-                __m128 _r13 = (__m128)__lsx_vld(r1 + 4 * 3, 0);
-                __m128 _r14 = (__m128)__lsx_vld(r1 + 4 * 4, 0);
-
-                __m128 _k10 = (__m128)__lsx_vld(k0, 0);
-                __m128 _k11 = (__m128)__lsx_vld(k0 + 4, 0);
-                __m128 _k12 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-                __m128 _k13 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-                __m128 _k14 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __lsx_vfmadd_s(_r10, _k10, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r11, _k11, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r12, _k12, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r13, _k13, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r14, _k14, _sum0);
-
-                __m128 _r20 = (__m128)__lsx_vld(r2, 0);
-                __m128 _r21 = (__m128)__lsx_vld(r2 + 4, 0);
-                __m128 _r22 = (__m128)__lsx_vld(r2 + 4 * 2, 0);
-                __m128 _r23 = (__m128)__lsx_vld(r2 + 4 * 3, 0);
-                __m128 _r24 = (__m128)__lsx_vld(r2 + 4 * 4, 0);
-
-                __m128 _k20 = (__m128)__lsx_vld(k0, 0);
-                __m128 _k21 = (__m128)__lsx_vld(k0 + 4, 0);
-                __m128 _k22 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-                __m128 _k23 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-                __m128 _k24 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __lsx_vfmadd_s(_r20, _k20, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r21, _k21, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r22, _k22, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r23, _k23, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r24, _k24, _sum0);
-
-                __m128 _r30 = (__m128)__lsx_vld(r3, 0);
-                __m128 _r31 = (__m128)__lsx_vld(r3 + 4, 0);
-                __m128 _r32 = (__m128)__lsx_vld(r3 + 4 * 2, 0);
-                __m128 _r33 = (__m128)__lsx_vld(r3 + 4 * 3, 0);
-                __m128 _r34 = (__m128)__lsx_vld(r3 + 4 * 4, 0);
-
-                __m128 _k30 = (__m128)__lsx_vld(k0, 0);
-                __m128 _k31 = (__m128)__lsx_vld(k0 + 4, 0);
-                __m128 _k32 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-                __m128 _k33 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-                __m128 _k34 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __lsx_vfmadd_s(_r30, _k30, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r31, _k31, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r32, _k32, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r33, _k33, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r34, _k34, _sum0);
-
-                __m128 _r40 = (__m128)__lsx_vld(r4, 0);
-                __m128 _r41 = (__m128)__lsx_vld(r4 + 4, 0);
-                __m128 _r42 = (__m128)__lsx_vld(r4 + 4 * 2, 0);
-                __m128 _r43 = (__m128)__lsx_vld(r4 + 4 * 3, 0);
-                __m128 _r44 = (__m128)__lsx_vld(r4 + 4 * 4, 0);
-
-                __m128 _k40 = (__m128)__lsx_vld(k0, 0);
-                __m128 _k41 = (__m128)__lsx_vld(k0 + 4, 0);
-                __m128 _k42 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-                __m128 _k43 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-                __m128 _k44 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-                k0 -= 4 * 20;
-
-                _sum0 = __lsx_vfmadd_s(_r40, _k40, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r41, _k41, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r42, _k42, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r43, _k43, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r44, _k44, _sum0);
-
-                __lsx_vst(_sum0, outptr0, 0);
-
-                outptr0 += 4;
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                r3 += 4;
-                r4 += 4;
-            }
-
-            r0 += 4 * 4;
-            r1 += 4 * 4;
-            r2 += 4 * 4;
-            r3 += 4 * 4;
-            r4 += 4 * 4;
-        }
-    }
-}
-
-static void convdw5x5s2_pack4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m128 _bias0 = bias ? (__m128)__lsx_vld(bias + g * 4, 0) : (__m128)__lsx_vreplgr2vr_w(0);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out;
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-        const float* r4 = img0.row(4);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                __builtin_prefetch(r0 + 32);
-                __builtin_prefetch(r1 + 32);
-                __builtin_prefetch(r2 + 32);
-                __builtin_prefetch(r3 + 32);
-                __builtin_prefetch(r4 + 32);
-
-                __builtin_prefetch(k0 + 400);
-
-                __m128 _sum0 = _bias0;
-
-                __m128 _r00 = (__m128)__lsx_vld(r0, 0);
-                __m128 _r01 = (__m128)__lsx_vld(r0 + 4, 0);
-                __m128 _r02 = (__m128)__lsx_vld(r0 + 4 * 2, 0);
-                __m128 _r03 = (__m128)__lsx_vld(r0 + 4 * 3, 0);
-                __m128 _r04 = (__m128)__lsx_vld(r0 + 4 * 4, 0);
-
-                __m128 _k00 = (__m128)__lsx_vld(k0, 0);
-                __m128 _k01 = (__m128)__lsx_vld(k0 + 4, 0);
-                __m128 _k02 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-                __m128 _k03 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-                __m128 _k04 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __lsx_vfmadd_s(_r00, _k00, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r01, _k01, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r02, _k02, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r03, _k03, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r04, _k04, _sum0);
-
-                __m128 _r10 = (__m128)__lsx_vld(r1, 0);
-                __m128 _r11 = (__m128)__lsx_vld(r1 + 4, 0);
-                __m128 _r12 = (__m128)__lsx_vld(r1 + 4 * 2, 0);
-                __m128 _r13 = (__m128)__lsx_vld(r1 + 4 * 3, 0);
-                __m128 _r14 = (__m128)__lsx_vld(r1 + 4 * 4, 0);
-
-                __m128 _k10 = (__m128)__lsx_vld(k0, 0);
-                __m128 _k11 = (__m128)__lsx_vld(k0 + 4, 0);
-                __m128 _k12 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-                __m128 _k13 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-                __m128 _k14 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __lsx_vfmadd_s(_r10, _k10, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r11, _k11, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r12, _k12, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r13, _k13, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r14, _k14, _sum0);
-
-                __m128 _r20 = (__m128)__lsx_vld(r2, 0);
-                __m128 _r21 = (__m128)__lsx_vld(r2 + 4, 0);
-                __m128 _r22 = (__m128)__lsx_vld(r2 + 4 * 2, 0);
-                __m128 _r23 = (__m128)__lsx_vld(r2 + 4 * 3, 0);
-                __m128 _r24 = (__m128)__lsx_vld(r2 + 4 * 4, 0);
-
-                __m128 _k20 = (__m128)__lsx_vld(k0, 0);
-                __m128 _k21 = (__m128)__lsx_vld(k0 + 4, 0);
-                __m128 _k22 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-                __m128 _k23 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-                __m128 _k24 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __lsx_vfmadd_s(_r20, _k20, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r21, _k21, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r22, _k22, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r23, _k23, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r24, _k24, _sum0);
-
-                __m128 _r30 = (__m128)__lsx_vld(r3, 0);
-                __m128 _r31 = (__m128)__lsx_vld(r3 + 4, 0);
-                __m128 _r32 = (__m128)__lsx_vld(r3 + 4 * 2, 0);
-                __m128 _r33 = (__m128)__lsx_vld(r3 + 4 * 3, 0);
-                __m128 _r34 = (__m128)__lsx_vld(r3 + 4 * 4, 0);
-
-                __m128 _k30 = (__m128)__lsx_vld(k0, 0);
-                __m128 _k31 = (__m128)__lsx_vld(k0 + 4, 0);
-                __m128 _k32 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-                __m128 _k33 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-                __m128 _k34 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __lsx_vfmadd_s(_r30, _k30, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r31, _k31, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r32, _k32, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r33, _k33, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r34, _k34, _sum0);
-
-                __m128 _r40 = (__m128)__lsx_vld(r4, 0);
-                __m128 _r41 = (__m128)__lsx_vld(r4 + 4, 0);
-                __m128 _r42 = (__m128)__lsx_vld(r4 + 4 * 2, 0);
-                __m128 _r43 = (__m128)__lsx_vld(r4 + 4 * 3, 0);
-                __m128 _r44 = (__m128)__lsx_vld(r4 + 4 * 4, 0);
-
-                __m128 _k40 = (__m128)__lsx_vld(k0, 0);
-                __m128 _k41 = (__m128)__lsx_vld(k0 + 4, 0);
-                __m128 _k42 = (__m128)__lsx_vld(k0 + 4 * 2, 0);
-                __m128 _k43 = (__m128)__lsx_vld(k0 + 4 * 3, 0);
-                __m128 _k44 = (__m128)__lsx_vld(k0 + 4 * 4, 0);
-                k0 -= 4 * 20;
-
-                _sum0 = __lsx_vfmadd_s(_r40, _k40, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r41, _k41, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r42, _k42, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r43, _k43, _sum0);
-                _sum0 = __lsx_vfmadd_s(_r44, _k44, _sum0);
-
-                __lsx_vst(_sum0, outptr0, 0);
-
-                outptr0 += 4;
-
-                r0 += 4 * 2;
-                r1 += 4 * 2;
-                r2 += 4 * 2;
-                r3 += 4 * 2;
-                r4 += 4 * 2;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-            r3 += tailstep;
-            r4 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_loongarch.h
deleted file mode 100644
index 35cdd8f0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/convolutiondepthwise_loongarch.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTIONDEPTHWISE_LOONGARCH_H
-#define LAYER_CONVOLUTIONDEPTHWISE_LOONGARCH_H
-
-#include "convolutiondepthwise.h"
-
-namespace ncnn {
-
-class ConvolutionDepthWise_loongarch : public ConvolutionDepthWise
-{
-public:
-    ConvolutionDepthWise_loongarch();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    int create_group_ops(const Option& opt);
-#if NCNN_INT8
-    int create_pipeline_int8_loongarch(const Option& opt);
-    int forward_int8_loongarch(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Layer* activation;
-    std::vector<ncnn::Layer*> group_ops;
-
-    Mat weight_data_tm;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTIONDEPTHWISE_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/crop_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/crop_loongarch.h
deleted file mode 100644
index cfb4ff35..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/crop_loongarch.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CROP_LOONGARCH_H
-#define LAYER_CROP_LOONGARCH_H
-
-#include "crop.h"
-
-namespace ncnn {
-
-class Crop_loongarch : public Crop
-{
-public:
-    Crop_loongarch();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CROP_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_loongarch.h
deleted file mode 100644
index 00ddf67e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_loongarch.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTION_LOONGARCH_H
-#define LAYER_DECONVOLUTION_LOONGARCH_H
-
-#include "deconvolution.h"
-
-namespace ncnn {
-
-class Deconvolution_loongarch : public Deconvolution
-{
-public:
-    Deconvolution_loongarch();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    Mat weight_data_tm;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTION_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_pack1to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_pack1to4.h
deleted file mode 100644
index ee1f932b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_pack1to4.h
+++ /dev/null
@@ -1,99 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack1to4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_pack1ton, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m128 _sum = (__m128)__lsx_vreplgr2vr_w(0);
-
-                if (bias_data_ptr)
-                {
-                    _sum = (__m128)__lsx_vld((const float*)bias_data_ptr + p * 4, 0);
-                }
-
-                const float* kptr = (const float*)weight_data_pack1ton + maxk * channels * p * 4;
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        const float* sptr = m.row(sy);
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            float val = sptr[sx];
-
-                            int k = y * kernel_w + x;
-
-                            __m128 _val = (__m128)__lsx_vreplfr2vr_s(val);
-                            __m128 _w = (__m128)__lsx_vld(kptr + k * 4, 0);
-                            _sum = __lsx_vfmadd_s(_w, _val, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 4;
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params);
-
-                __lsx_vst(_sum, outptr + j * 4, 0);
-            }
-
-            outptr += outw * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_pack4.h
deleted file mode 100644
index 179a4103..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_pack4.h
+++ /dev/null
@@ -1,106 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_pack4, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m128 _sum = (__m128)__lsx_vreplgr2vr_w(0);
-
-                if (bias_data_ptr)
-                {
-                    _sum = (__m128)__lsx_vld((const float*)bias_data_ptr + p * 4, 0);
-                }
-
-                const float* kptr = (const float*)weight_data_pack4.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 4;
-
-                            int k = (y * kernel_w + x) * 16;
-
-                            __m128 _val0 = (__m128)__lsx_vreplfr2vr_s(*sptr++);
-                            __m128 _val1 = (__m128)__lsx_vreplfr2vr_s(*sptr++);
-                            __m128 _val2 = (__m128)__lsx_vreplfr2vr_s(*sptr++);
-                            __m128 _val3 = (__m128)__lsx_vreplfr2vr_s(*sptr++);
-                            __m128 _w0 = (__m128)__lsx_vld(kptr + k, 0);
-                            __m128 _w1 = (__m128)__lsx_vld(kptr + k + 4, 0);
-                            __m128 _w2 = (__m128)__lsx_vld(kptr + k + 8, 0);
-                            __m128 _w3 = (__m128)__lsx_vld(kptr + k + 12, 0);
-                            _sum = __lsx_vfmadd_s(_w0, _val0, _sum);
-                            _sum = __lsx_vfmadd_s(_w1, _val1, _sum);
-                            _sum = __lsx_vfmadd_s(_w2, _val2, _sum);
-                            _sum = __lsx_vfmadd_s(_w3, _val3, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 16;
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params);
-
-                __lsx_vst(_sum, outptr + j * 4, 0);
-            }
-
-            outptr += outw * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_pack4to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_pack4to1.h
deleted file mode 100644
index e13721c2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolution_pack4to1.h
+++ /dev/null
@@ -1,101 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack4to1_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_pack4to1, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                __m128 _sum = (__m128)__lsx_vreplgr2vr_w(0);
-
-                const float* kptr = (const float*)weight_data_pack4to1 + maxk * channels * p * 4;
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 4;
-
-                            int k = y * kernel_w + x;
-
-                            __m128 _val = (__m128)__lsx_vld(sptr, 0);
-                            __m128 _w = (__m128)__lsx_vld(kptr + k * 4, 0);
-                            _sum = __lsx_vfmadd_s(_w, _val, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 4;
-                }
-
-                sum += __lsx_reduce_fadd_s(_sum);
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[j] = sum;
-            }
-
-            outptr += outw;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolutiondepthwise_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolutiondepthwise_loongarch.h
deleted file mode 100644
index 87c5351f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/deconvolutiondepthwise_loongarch.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTIONDEPTHWISE_LOONGARCH_H
-#define LAYER_DECONVOLUTIONDEPTHWISE_LOONGARCH_H
-
-#include "deconvolutiondepthwise.h"
-
-namespace ncnn {
-
-class DeconvolutionDepthWise_loongarch : public DeconvolutionDepthWise
-{
-public:
-    DeconvolutionDepthWise_loongarch();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    int create_group_ops(const Option& opt);
-
-public:
-    std::vector<ncnn::Layer*> group_ops;
-
-    Mat weight_data_tm;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTIONDEPTHWISE_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/dequantize_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/dequantize_loongarch.h
deleted file mode 100644
index ae7d3fe6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/dequantize_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DEQUANTIZE_LOONGARCH_H
-#define LAYER_DEQUANTIZE_LOONGARCH_H
-
-#include "dequantize.h"
-
-namespace ncnn {
-
-class Dequantize_loongarch : public Dequantize
-{
-public:
-    Dequantize_loongarch();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DEQUANTIZE_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/dropout_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/dropout_loongarch.h
deleted file mode 100644
index f9beff05..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/dropout_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DROPOUT_LOONGARCH_H
-#define LAYER_DROPOUT_LOONGARCH_H
-
-#include "dropout.h"
-
-namespace ncnn {
-
-class Dropout_loongarch : public Dropout
-{
-public:
-    Dropout_loongarch();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DROPOUT_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/eltwise_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/eltwise_loongarch.h
deleted file mode 100644
index f523132b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/eltwise_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ELTWISE_LOONGARCH_H
-#define LAYER_ELTWISE_LOONGARCH_H
-
-#include "eltwise.h"
-
-namespace ncnn {
-
-class Eltwise_loongarch : public Eltwise
-{
-public:
-    Eltwise_loongarch();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ELTWISE_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/flatten_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/flatten_loongarch.h
deleted file mode 100644
index da75fd12..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/flatten_loongarch.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_FLATTEN_LOONGARCH_H
-#define LAYER_FLATTEN_LOONGARCH_H
-
-#include "flatten.h"
-
-namespace ncnn {
-
-class Flatten_loongarch : public Flatten
-{
-public:
-    Flatten_loongarch();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_FLATTEN_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/hardsigmoid_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/hardsigmoid_loongarch.h
deleted file mode 100644
index 519a4ba9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/hardsigmoid_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_HARDSIGMOID_LOONGARCH_H
-#define LAYER_HARDSIGMOID_LOONGARCH_H
-
-#include "hardsigmoid.h"
-
-namespace ncnn {
-
-class HardSigmoid_loongarch : public HardSigmoid
-{
-public:
-    HardSigmoid_loongarch();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_HARDSIGMOID_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/hardswish_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/hardswish_loongarch.h
deleted file mode 100644
index ef69cb05..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/hardswish_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_HARDSWISH_LOONGARCH_H
-#define LAYER_HARDSWISH_LOONGARCH_H
-
-#include "hardswish.h"
-
-namespace ncnn {
-
-class HardSwish_loongarch : public HardSwish
-{
-public:
-    HardSwish_loongarch();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_HARDSWISH_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/innerproduct_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/innerproduct_loongarch.h
deleted file mode 100644
index 2ae1a1e5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/innerproduct_loongarch.h
+++ /dev/null
@@ -1,54 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INNERPRODUCT_LOONGARCH_H
-#define LAYER_INNERPRODUCT_LOONGARCH_H
-
-#include "innerproduct.h"
-
-namespace ncnn {
-
-class InnerProduct_loongarch : public InnerProduct
-{
-public:
-    InnerProduct_loongarch();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-#if __loongarch_sx
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_INT8
-    int create_pipeline_int8_loongarch(const Option& opt);
-    int forward_int8_loongarch(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Layer* flatten;
-
-    Mat weight_data_tm;
-
-#if NCNN_INT8
-    Mat scale_in_data;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INNERPRODUCT_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bicubic.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bicubic.h
deleted file mode 100644
index e52ba81d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bicubic.h
+++ /dev/null
@@ -1,261 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static inline void interpolate_cubic(float fx, float* coeffs)
-{
-    const float A = -0.75f;
-
-    float fx0 = fx + 1;
-    float fx1 = fx;
-    float fx2 = 1 - fx;
-    // float fx3 = 2 - fx;
-
-    coeffs[0] = A * fx0 * fx0 * fx0 - 5 * A * fx0 * fx0 + 8 * A * fx0 - 4 * A;
-    coeffs[1] = (A + 2) * fx1 * fx1 * fx1 - (A + 3) * fx1 * fx1 + 1;
-    coeffs[2] = (A + 2) * fx2 * fx2 * fx2 - (A + 3) * fx2 * fx2 + 1;
-    coeffs[3] = 1.f - coeffs[0] - coeffs[1] - coeffs[2];
-}
-
-static void cubic_coeffs(int w, int outw, int* xofs, float* alpha, int align_corner)
-{
-    double scale = (double)w / outw;
-    if (align_corner)
-    {
-        scale = (double)(w - 1) / (outw - 1);
-    }
-
-    for (int dx = 0; dx < outw; dx++)
-    {
-        float fx = (float)((dx + 0.5) * scale - 0.5);
-        if (align_corner)
-        {
-            fx = (float)(dx * scale);
-        }
-
-        int sx = static_cast<int>(floor(fx));
-        fx -= sx;
-
-        interpolate_cubic(fx, alpha + dx * 4);
-
-        if (sx <= -1)
-        {
-            sx = 1;
-            alpha[dx * 4 + 0] = 1.f - alpha[dx * 4 + 3];
-            alpha[dx * 4 + 1] = alpha[dx * 4 + 3];
-            alpha[dx * 4 + 2] = 0.f;
-            alpha[dx * 4 + 3] = 0.f;
-        }
-        if (sx == 0)
-        {
-            sx = 1;
-            alpha[dx * 4 + 0] = alpha[dx * 4 + 0] + alpha[dx * 4 + 1];
-            alpha[dx * 4 + 1] = alpha[dx * 4 + 2];
-            alpha[dx * 4 + 2] = alpha[dx * 4 + 3];
-            alpha[dx * 4 + 3] = 0.f;
-        }
-        if (sx == w - 2)
-        {
-            sx = w - 3;
-            alpha[dx * 4 + 3] = alpha[dx * 4 + 2] + alpha[dx * 4 + 3];
-            alpha[dx * 4 + 2] = alpha[dx * 4 + 1];
-            alpha[dx * 4 + 1] = alpha[dx * 4 + 0];
-            alpha[dx * 4 + 0] = 0.f;
-        }
-        if (sx >= w - 1)
-        {
-            sx = w - 3;
-            alpha[dx * 4 + 3] = 1.f - alpha[dx * 4 + 0];
-            alpha[dx * 4 + 2] = alpha[dx * 4 + 0];
-            alpha[dx * 4 + 1] = 0.f;
-            alpha[dx * 4 + 0] = 0.f;
-        }
-
-        xofs[dx] = sx;
-    }
-}
-
-static void resize_bicubic_image(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    Mat rowsbuf2(w);
-    Mat rowsbuf3(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows1p[dx] = S1p[-1] * a0 + S1p[0] * a1 + S1p[1] * a2 + S1p[2] * a3;
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const float* S0 = src.row(sy - 1);
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows0p[dx] = S0p[-1] * a0 + S0p[0] * a1 + S0p[1] * a2 + S0p[2] * a3;
-                rows1p[dx] = S1p[-1] * a0 + S1p[0] * a1 + S1p[1] * a2 + S1p[2] * a3;
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-        float b2 = beta[2];
-        float b3 = beta[3];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        float* Dp = dst.row(dy);
-        for (int dx = 0; dx < w; dx++)
-        {
-            //             D[x] = rows0[x]*b0 + rows1[x]*b1 + rows2[x]*b2 + rows3[x]*b3;
-            *Dp++ = *rows0p++ * b0 + *rows1p++ * b1 + *rows2p++ * b2 + *rows3p++ * b3;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bicubic_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bicubic_pack4.h
deleted file mode 100644
index 6e52dc1e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bicubic_pack4.h
+++ /dev/null
@@ -1,286 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bicubic_image_pack4(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf1(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf2(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf3(w, (size_t)4 * 4u, 4);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S3p = S3 + sx;
-
-                __m128 _a0 = __lsx_vreplfr2vr_s(alphap[0]);
-                __m128 _a1 = __lsx_vreplfr2vr_s(alphap[1]);
-                __m128 _a2 = __lsx_vreplfr2vr_s(alphap[2]);
-                __m128 _a3 = __lsx_vreplfr2vr_s(alphap[3]);
-
-                __m128 _S30 = (__m128)__lsx_vld(S3p - 4, 0);
-                __m128 _S31 = (__m128)__lsx_vld(S3p + 0, 0);
-                __m128 _S32 = (__m128)__lsx_vld(S3p + 4, 0);
-                __m128 _S33 = (__m128)__lsx_vld(S3p + 8, 0);
-                __m128 _rows3 = __lsx_vfmul_s(_S30, _a0);
-                _rows3 = __lsx_vfmadd_s(_a1, _S31, _rows3);
-                _rows3 = __lsx_vfmadd_s(_a2, _S32, _rows3);
-                _rows3 = __lsx_vfmadd_s(_a3, _S33, _rows3);
-                __lsx_vst(_rows3, rows3p + dx * 4, 0);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                __m128 _a0 = __lsx_vreplfr2vr_s(alphap[0]);
-                __m128 _a1 = __lsx_vreplfr2vr_s(alphap[1]);
-                __m128 _a2 = __lsx_vreplfr2vr_s(alphap[2]);
-                __m128 _a3 = __lsx_vreplfr2vr_s(alphap[3]);
-
-                __m128 _S20 = (__m128)__lsx_vld(S2p - 4, 0);
-                __m128 _S21 = (__m128)__lsx_vld(S2p + 0, 0);
-                __m128 _S22 = (__m128)__lsx_vld(S2p + 4, 0);
-                __m128 _S23 = (__m128)__lsx_vld(S2p + 8, 0);
-                __m128 _S30 = (__m128)__lsx_vld(S3p - 4, 0);
-                __m128 _S31 = (__m128)__lsx_vld(S3p + 0, 0);
-                __m128 _S32 = (__m128)__lsx_vld(S3p + 4, 0);
-                __m128 _S33 = (__m128)__lsx_vld(S3p + 8, 0);
-                __m128 _rows2 = __lsx_vfmul_s(_S20, _a0);
-                __m128 _rows3 = __lsx_vfmul_s(_S30, _a0);
-                _rows2 = __lsx_vfmadd_s(_a1, _S21, _rows2);
-                _rows3 = __lsx_vfmadd_s(_a1, _S31, _rows3);
-                _rows2 = __lsx_vfmadd_s(_a2, _S22, _rows2);
-                _rows3 = __lsx_vfmadd_s(_a2, _S32, _rows3);
-                _rows2 = __lsx_vfmadd_s(_a3, _S23, _rows2);
-                _rows3 = __lsx_vfmadd_s(_a3, _S33, _rows3);
-                __lsx_vst(_rows2, rows2p + dx * 4, 0);
-                __lsx_vst(_rows3, rows3p + dx * 4, 0);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                __m128 _a0 = __lsx_vreplfr2vr_s(alphap[0]);
-                __m128 _a1 = __lsx_vreplfr2vr_s(alphap[1]);
-                __m128 _a2 = __lsx_vreplfr2vr_s(alphap[2]);
-                __m128 _a3 = __lsx_vreplfr2vr_s(alphap[3]);
-
-                __m128 _S10 = (__m128)__lsx_vld(S1p - 4, 0);
-                __m128 _S11 = (__m128)__lsx_vld(S1p + 0, 0);
-                __m128 _S12 = (__m128)__lsx_vld(S1p + 4, 0);
-                __m128 _S13 = (__m128)__lsx_vld(S1p + 8, 0);
-                __m128 _S20 = (__m128)__lsx_vld(S2p - 4, 0);
-                __m128 _S21 = (__m128)__lsx_vld(S2p + 0, 0);
-                __m128 _S22 = (__m128)__lsx_vld(S2p + 4, 0);
-                __m128 _S23 = (__m128)__lsx_vld(S2p + 8, 0);
-                __m128 _S30 = (__m128)__lsx_vld(S3p - 4, 0);
-                __m128 _S31 = (__m128)__lsx_vld(S3p + 0, 0);
-                __m128 _S32 = (__m128)__lsx_vld(S3p + 4, 0);
-                __m128 _S33 = (__m128)__lsx_vld(S3p + 8, 0);
-                __m128 _rows1 = __lsx_vfmul_s(_S10, _a0);
-                __m128 _rows2 = __lsx_vfmul_s(_S20, _a0);
-                __m128 _rows3 = __lsx_vfmul_s(_S30, _a0);
-                _rows1 = __lsx_vfmadd_s(_a1, _S11, _rows1);
-                _rows2 = __lsx_vfmadd_s(_a1, _S21, _rows2);
-                _rows3 = __lsx_vfmadd_s(_a1, _S31, _rows3);
-                _rows1 = __lsx_vfmadd_s(_a2, _S12, _rows1);
-                _rows2 = __lsx_vfmadd_s(_a2, _S22, _rows2);
-                _rows3 = __lsx_vfmadd_s(_a2, _S32, _rows3);
-                _rows1 = __lsx_vfmadd_s(_a3, _S13, _rows1);
-                _rows2 = __lsx_vfmadd_s(_a3, _S23, _rows2);
-                _rows3 = __lsx_vfmadd_s(_a3, _S33, _rows3);
-                __lsx_vst(_rows1, rows1p + dx * 4, 0);
-                __lsx_vst(_rows2, rows2p + dx * 4, 0);
-                __lsx_vst(_rows3, rows3p + dx * 4, 0);
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const float* S0 = src.row(sy - 1);
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                __m128 _a0 = __lsx_vreplfr2vr_s(alphap[0]);
-                __m128 _a1 = __lsx_vreplfr2vr_s(alphap[1]);
-                __m128 _a2 = __lsx_vreplfr2vr_s(alphap[2]);
-                __m128 _a3 = __lsx_vreplfr2vr_s(alphap[3]);
-
-                __m128 _S00 = (__m128)__lsx_vld(S0p - 4, 0);
-                __m128 _S01 = (__m128)__lsx_vld(S0p + 0, 0);
-                __m128 _S02 = (__m128)__lsx_vld(S0p + 4, 0);
-                __m128 _S03 = (__m128)__lsx_vld(S0p + 8, 0);
-                __m128 _S10 = (__m128)__lsx_vld(S1p - 4, 0);
-                __m128 _S11 = (__m128)__lsx_vld(S1p + 0, 0);
-                __m128 _S12 = (__m128)__lsx_vld(S1p + 4, 0);
-                __m128 _S13 = (__m128)__lsx_vld(S1p + 8, 0);
-                __m128 _S20 = (__m128)__lsx_vld(S2p - 4, 0);
-                __m128 _S21 = (__m128)__lsx_vld(S2p + 0, 0);
-                __m128 _S22 = (__m128)__lsx_vld(S2p + 4, 0);
-                __m128 _S23 = (__m128)__lsx_vld(S2p + 8, 0);
-                __m128 _S30 = (__m128)__lsx_vld(S3p - 4, 0);
-                __m128 _S31 = (__m128)__lsx_vld(S3p + 0, 0);
-                __m128 _S32 = (__m128)__lsx_vld(S3p + 4, 0);
-                __m128 _S33 = (__m128)__lsx_vld(S3p + 8, 0);
-                __m128 _rows0 = __lsx_vfmul_s(_S00, _a0);
-                __m128 _rows1 = __lsx_vfmul_s(_S10, _a0);
-                __m128 _rows2 = __lsx_vfmul_s(_S20, _a0);
-                __m128 _rows3 = __lsx_vfmul_s(_S30, _a0);
-                _rows0 = __lsx_vfmadd_s(_a1, _S01, _rows0);
-                _rows1 = __lsx_vfmadd_s(_a1, _S11, _rows1);
-                _rows2 = __lsx_vfmadd_s(_a1, _S21, _rows2);
-                _rows3 = __lsx_vfmadd_s(_a1, _S31, _rows3);
-                _rows0 = __lsx_vfmadd_s(_a2, _S02, _rows0);
-                _rows1 = __lsx_vfmadd_s(_a2, _S12, _rows1);
-                _rows2 = __lsx_vfmadd_s(_a2, _S22, _rows2);
-                _rows3 = __lsx_vfmadd_s(_a2, _S32, _rows3);
-                _rows0 = __lsx_vfmadd_s(_a3, _S03, _rows0);
-                _rows1 = __lsx_vfmadd_s(_a3, _S13, _rows1);
-                _rows2 = __lsx_vfmadd_s(_a3, _S23, _rows2);
-                _rows3 = __lsx_vfmadd_s(_a3, _S33, _rows3);
-                __lsx_vst(_rows0, rows0p + dx * 4, 0);
-                __lsx_vst(_rows1, rows1p + dx * 4, 0);
-                __lsx_vst(_rows2, rows2p + dx * 4, 0);
-                __lsx_vst(_rows3, rows3p + dx * 4, 0);
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        __m128 _b0 = __lsx_vreplfr2vr_s(beta[0]);
-        __m128 _b1 = __lsx_vreplfr2vr_s(beta[1]);
-        __m128 _b2 = __lsx_vreplfr2vr_s(beta[2]);
-        __m128 _b3 = __lsx_vreplfr2vr_s(beta[3]);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        float* Dp = dst.row(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            __m128 _rows0 = (__m128)__lsx_vld(rows0p, 0);
-            __m128 _rows1 = (__m128)__lsx_vld(rows1p, 0);
-            __m128 _rows2 = (__m128)__lsx_vld(rows2p, 0);
-            __m128 _rows3 = (__m128)__lsx_vld(rows3p, 0);
-            __m128 _Dp = __lsx_vfmul_s(_rows0, _b0);
-            _Dp = __lsx_vfmadd_s(_b1, _rows1, _Dp);
-            _Dp = __lsx_vfmadd_s(_b2, _rows2, _Dp);
-            _Dp = __lsx_vfmadd_s(_b3, _rows3, _Dp);
-            __lsx_vst(_Dp, Dp, 0);
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-            rows2p += 4;
-            rows3p += 4;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bilinear.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bilinear.h
deleted file mode 100644
index 8ace1868..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bilinear.h
+++ /dev/null
@@ -1,172 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void linear_coeffs(int w, int outw, int* xofs, float* alpha, int align_corner)
-{
-    double scale = (double)w / outw;
-    if (align_corner)
-    {
-        scale = (double)(w - 1) / (outw - 1);
-    }
-
-    for (int dx = 0; dx < outw; dx++)
-    {
-        float fx = (float)((dx + 0.5) * scale - 0.5);
-        if (align_corner)
-        {
-            fx = (float)(dx * scale);
-        }
-
-        int sx = floor(fx);
-        fx -= sx;
-
-        if (sx < 0)
-        {
-            sx = 0;
-            fx = 0.f;
-        }
-        if (sx >= w - 1)
-        {
-            sx = w - 2;
-            fx = 1.f;
-        }
-
-        xofs[dx] = sx;
-
-        alpha[dx * 2] = 1.f - fx;
-        alpha[dx * 2 + 1] = fx;
-    }
-}
-
-static void resize_bilinear_image(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows1p[dx] = S1p[0] * a0 + S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const float* S0 = src.row(sy);
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows0p[dx] = S0p[0] * a0 + S0p[1] * a1;
-                rows1p[dx] = S1p[0] * a0 + S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* Dp = dst.row(dy);
-
-#if __loongarch_sx
-        int nn = w >> 3;
-#else
-        int nn = 0;
-#endif
-        int remain = w - (nn << 3);
-
-#if __loongarch_sx
-        __m128 _b0 = __lsx_vreplfr2vr_s(b0);
-        __m128 _b1 = __lsx_vreplfr2vr_s(b1);
-        for (; nn > 0; nn--)
-        {
-            __m128 _rows0 = (__m128)__lsx_vld(rows0p, 0);
-            __m128 _rows1 = (__m128)__lsx_vld(rows1p, 0);
-
-            __m128 _Dp = __lsx_vfmul_s(_rows0, _b0);
-            _Dp = __lsx_vfmadd_s(_b1, _rows1, _Dp);
-
-            __lsx_vst(_Dp, Dp, 0);
-
-            __m128 _rows0n = (__m128)__lsx_vld(rows0p + 4, 0);
-            __m128 _rows1n = (__m128)__lsx_vld(rows1p + 4, 0);
-
-            __m128 _Dpn = __lsx_vfmul_s(_rows0n, _b0);
-            _Dpn = __lsx_vfmadd_s(_b1, _rows1n, _Dpn);
-
-            __lsx_vst(_Dpn, Dp + 4, 0);
-
-            Dp += 8;
-            rows0p += 8;
-            rows1p += 8;
-        }
-#endif // __loongarch_sx
-        for (; remain; --remain)
-        {
-            //             D[x] = rows0[x]*b0 + rows1[x]*b1;
-            *Dp++ = *rows0p++ * b0 + *rows1p++ * b1;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bilinear_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bilinear_pack4.h
deleted file mode 100644
index 3702b257..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_bilinear_pack4.h
+++ /dev/null
@@ -1,123 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bilinear_image_pack4(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf1(w, (size_t)4 * 4u, 4);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S1p = S1 + sx;
-
-                __m128 _a0 = __lsx_vreplfr2vr_s(alphap[0]);
-                __m128 _a1 = __lsx_vreplfr2vr_s(alphap[1]);
-
-                __m128 _S10 = (__m128)__lsx_vld(S1p, 0);
-                __m128 _S11 = (__m128)__lsx_vld(S1p + 4, 0);
-                __m128 _rows1 = __lsx_vfmul_s(_S10, _a0);
-                _rows1 = __lsx_vfmadd_s(_a1, _S11, _rows1);
-                __lsx_vst(_rows1, rows1p + dx * 4, 0);
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const float* S0 = src.row(sy);
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-
-                __m128 _a0 = __lsx_vreplfr2vr_s(alphap[0]);
-                __m128 _a1 = __lsx_vreplfr2vr_s(alphap[1]);
-
-                __m128 _S00 = (__m128)__lsx_vld(S0p, 0);
-                __m128 _S01 = (__m128)__lsx_vld(S0p + 4, 0);
-                __m128 _S10 = (__m128)__lsx_vld(S1p, 0);
-                __m128 _S11 = (__m128)__lsx_vld(S1p + 4, 0);
-                __m128 _rows0 = __lsx_vfmul_s(_S00, _a0);
-                __m128 _rows1 = __lsx_vfmul_s(_S10, _a0);
-                _rows0 = __lsx_vfmadd_s(_a1, _S01, _rows0);
-                _rows1 = __lsx_vfmadd_s(_a1, _S11, _rows1);
-                __lsx_vst(_rows0, rows0p + dx * 4, 0);
-                __lsx_vst(_rows1, rows1p + dx * 4, 0);
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        __m128 _b0 = __lsx_vreplfr2vr_s(beta[0]);
-        __m128 _b1 = __lsx_vreplfr2vr_s(beta[1]);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* Dp = dst.row(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            __m128 _rows0 = (__m128)__lsx_vld(rows0p, 0);
-            __m128 _rows1 = (__m128)__lsx_vld(rows1p, 0);
-            __m128 _Dp = __lsx_vfmul_s(_rows0, _b0);
-            _Dp = __lsx_vfmadd_s(_b1, _rows1, _Dp);
-            __lsx_vst(_Dp, Dp, 0);
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_loongarch.h
deleted file mode 100644
index f1fa8070..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/interp_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INTERP_LOONGARCH_H
-#define LAYER_INTERP_LOONGARCH_H
-
-#include "interp.h"
-
-namespace ncnn {
-
-class Interp_loongarch : public Interp
-{
-public:
-    Interp_loongarch();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INTERP_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/loongarch_activation.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/loongarch_activation.h
deleted file mode 100644
index abb268f4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/loongarch_activation.h
+++ /dev/null
@@ -1,70 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LOONGARCH_ACTIVATION_H
-#define LOONGARCH_ACTIVATION_H
-
-#include "fused_activation.h"
-
-#if __loongarch_sx
-#include <lsxintrin.h>
-#include "lsx_mathfun.h"
-
-static inline __m128 activation_ps(__m128 _v, int activation_type, const ncnn::Mat& activation_params)
-{
-    if (activation_type == 1)
-    {
-        __m128 _zero = (__m128)__lsx_vreplgr2vr_w(0);
-        _v = __lsx_vfmax_s(_v, _zero);
-    }
-    else if (activation_type == 2)
-    {
-        __m128 _zero = (__m128)__lsx_vreplgr2vr_w(0);
-        __m128 _slope = (__m128)__lsx_vreplfr2vr_s(activation_params[0]);
-        __m128i _lemask = __lsx_vfcmp_cle_s(_v, _zero);
-        __m128 _ps = __lsx_vfmul_s(_v, _slope);
-        _v = (__m128)__lsx_vbitsel_v((__m128i)_v, (__m128i)_ps, (__m128i)_lemask);
-    }
-    else if (activation_type == 3)
-    {
-        __m128 _min = (__m128)__lsx_vreplfr2vr_s(activation_params[0]);
-        __m128 _max = (__m128)__lsx_vreplfr2vr_s(activation_params[1]);
-        _v = __lsx_vfmax_s(_v, _min);
-        _v = __lsx_vfmin_s(_v, _max);
-    }
-    else if (activation_type == 4)
-    {
-        _v = sigmoid_ps(_v);
-    }
-    else if (activation_type == 5)
-    {
-        _v = __lsx_vfmul_s(_v, tanh_ps(log_ps(__lsx_vfadd_s(exp_ps(_v), (__m128)__lsx_vreplfr2vr_s(1.f)))));
-    }
-    else if (activation_type == 6)
-    {
-        __m128 _alpha = (__m128)__lsx_vreplfr2vr_s(activation_params[0]);
-        __m128 _beta = (__m128)__lsx_vreplfr2vr_s(activation_params[1]);
-        __m128 _zero = (__m128)__lsx_vreplgr2vr_w(0);
-        __m128 _one = (__m128)__lsx_vreplfr2vr_s(1.f);
-        __m128 _outp = __lsx_vfmadd_s(_alpha, _v, _beta);
-        _outp = __lsx_vfmax_s(_outp, _zero);
-        _outp = __lsx_vfmin_s(_outp, _one);
-        _v = __lsx_vfmul_s(_outp, _v);
-    }
-
-    return _v;
-}
-#endif // __loongarch_sx
-
-#endif // LOONGARCH_ACTIVATION_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/loongarch_usability.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/loongarch_usability.h
deleted file mode 100644
index 0cd82e8f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/loongarch_usability.h
+++ /dev/null
@@ -1,235 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LOONGARCH_USABILITY_H
-#define LOONGARCH_USABILITY_H
-
-#if __loongarch_sx
-#include <lsxintrin.h>
-#endif // __loongarch_sx
-
-#include <stdint.h>
-
-namespace ncnn {
-
-typedef union
-{
-    int32_t i;
-    float f;
-} FloatInt;
-
-} // namespace ncnn
-
-#if __loongarch_sx
-/* declare some loongarch constants with union */
-#define _LOONGARCH_FLOAT_CONST(Name, Val) \
-    static const ncnn::FloatInt Name = {.f = Val}
-
-/* float type data load instructions */
-static NCNN_FORCEINLINE __m128 __lsx_vreplfr2vr_s(float val)
-{
-    ncnn::FloatInt fi_tmpval = {.f = val};
-    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
-}
-
-static NCNN_FORCEINLINE float __lsx_reduce_fadd_s(__m128 _v)
-{
-    // TODO find a more efficient way
-    float* _v_p = (float*)&_v;
-    return _v_p[0] + _v_p[1] + _v_p[2] + _v_p[3];
-}
-
-static NCNN_FORCEINLINE int __lsx_reduce_add_w(__m128i _v)
-{
-    // TODO find a more efficient way
-    int* _v_p = (int*)&_v;
-    return _v_p[0] + _v_p[1] + _v_p[2] + _v_p[3];
-}
-
-#endif // __loongarch_sx
-
-static NCNN_FORCEINLINE signed char float2int8(float v)
-{
-    int int32 = round(v);
-    if (int32 > 127) return 127;
-    if (int32 < -127) return -127;
-    return (signed char)int32;
-}
-
-#if __loongarch_sx
-static NCNN_FORCEINLINE __m128i float2int8(__m128 _v)
-{
-    // simulate round to nearest via +/-0.5
-    __m128 _p5 = (__m128)__lsx_vreplfr2vr_s(0.5f);
-    __m128i _signmask = __lsx_vreplgr2vr_w(1 << 31);
-
-    __m128i _sign = __lsx_vand_v((__m128i)_v, _signmask);
-    __m128 _p5s = (__m128)__lsx_vor_v((__m128i)_p5, (__m128i)_sign);
-    __m128 _v5 = __lsx_vfadd_s(_v, _p5s);
-    __m128i _v32 = __lsx_vftintrz_w_s(_v5);
-
-    __m128i _v32_16 = __lsx_vsat_w(_v32, 15);
-    __m128i _v16 = __lsx_vpickev_h(_v32_16, _v32_16);
-    _v16 = __lsx_vmax_h(_v16, __lsx_vreplgr2vr_h(-127));
-    __m128i _v16_8 = __lsx_vsat_h(_v16, 7);
-    __m128i _v8 = __lsx_vpickev_b(_v16_8, _v16_8);
-
-    return _v8;
-}
-
-static NCNN_FORCEINLINE int64_t float2int8(__m128 _vlow, __m128 _vhigh)
-{
-    // simulate round to nearest via +/-0.5
-    __m128 _p5 = (__m128)__lsx_vreplfr2vr_s(0.5f);
-    __m128i _signmask = __lsx_vreplgr2vr_w(1 << 31);
-
-    __m128i _signlow = __lsx_vand_v((__m128i)_vlow, _signmask);
-    __m128i _signhigh = __lsx_vand_v((__m128i)_vhigh, _signmask);
-    __m128 _p5low = (__m128)__lsx_vor_v((__m128i)_p5, _signlow);
-    __m128 _p5high = (__m128)__lsx_vor_v((__m128i)_p5, _signhigh);
-    __m128 _vlow5 = __lsx_vfadd_s(_vlow, _p5low);
-    __m128 _vhigh5 = __lsx_vfadd_s(_vhigh, _p5high);
-    __m128i _vlow32 = __lsx_vftintrz_w_s(_vlow5);
-    __m128i _vhigh32 = __lsx_vftintrz_w_s(_vhigh5);
-
-    __m128i _vlow32_16 = __lsx_vsat_w(_vlow32, 15);
-    __m128i _vhigh32_16 = __lsx_vsat_w(_vhigh32, 15);
-    __m128i _v16 = __lsx_vpickev_h(_vhigh32_16, _vlow32_16);
-    _v16 = __lsx_vmax_h(_v16, __lsx_vreplgr2vr_h(-127));
-    __m128i _v16_8 = __lsx_vsat_h(_v16, 7);
-    __m128i _v8 = __lsx_vpickev_b(_v16_8, _v16_8);
-
-    return _v8[0];
-}
-
-static NCNN_FORCEINLINE __m128i float2int8relu(__m128 _v)
-{
-    // simulate round to nearest via +/-0.5
-    __m128 _p5 = (__m128)__lsx_vreplfr2vr_s(0.5f);
-    __m128i _signmask = __lsx_vreplgr2vr_w(1 << 31);
-
-    __m128i _sign = __lsx_vand_v((__m128i)_v, _signmask);
-    __m128 _p5s = (__m128)__lsx_vor_v((__m128i)_p5, _sign);
-    __m128 _v5 = __lsx_vfadd_s(_v, _p5s);
-    __m128i _v32 = __lsx_vftintrz_w_s(_v5);
-
-    __m128i _v32_16 = __lsx_vsat_w(_v32, 15);
-    __m128i _v16 = __lsx_vpickev_h(_v32_16, _v32_16);
-    _v16 = __lsx_vmaxi_h(_v16, 0);
-    __m128i _v16_8 = __lsx_vsat_h(_v16, 7);
-    __m128i _v8 = __lsx_vpickev_b(_v16_8, _v16_8);
-
-    return _v8;
-}
-
-static NCNN_FORCEINLINE int64_t float2int8relu(__m128 _vlow, __m128 _vhigh)
-{
-    // simulate round to nearest via +/-0.5
-    __m128 _p5 = (__m128)__lsx_vreplfr2vr_s(0.5f);
-    __m128i _signmask = __lsx_vreplgr2vr_w(1 << 31);
-
-    __m128i _signlow = __lsx_vand_v((__m128i)_vlow, _signmask);
-    __m128i _signhigh = __lsx_vand_v((__m128i)_vhigh, _signmask);
-    __m128 _p5low = (__m128)__lsx_vor_v((__m128i)_p5, _signlow);
-    __m128 _p5high = (__m128)__lsx_vor_v((__m128i)_p5, _signhigh);
-    __m128 _vlow5 = __lsx_vfadd_s(_vlow, _p5low);
-    __m128 _vhigh5 = __lsx_vfadd_s(_vhigh, _p5high);
-    __m128i _vlow32 = __lsx_vftintrz_w_s(_vlow5);
-    __m128i _vhigh32 = __lsx_vftintrz_w_s(_vhigh5);
-
-    __m128i _vlow32_16 = __lsx_vsat_w(_vlow32, 15);
-    __m128i _vhigh32_16 = __lsx_vsat_w(_vhigh32, 15);
-    __m128i _v16 = __lsx_vpickev_h(_vhigh32_16, _vlow32_16);
-    _v16 = __lsx_vmaxi_h(_v16, 0);
-    __m128i _v16_8 = __lsx_vsat_h(_v16, 7);
-    __m128i _v8 = __lsx_vpickev_b(_v16_8, _v16_8);
-
-    return _v8[0];
-}
-
-static NCNN_FORCEINLINE __m128i float2int8leakyrelu(__m128 _v, __m128 _slope)
-{
-    __m128 _v_leaky = __lsx_vfmul_s(_v, _slope);
-
-    // simulate round to nearest via +/-0.5
-    __m128 _p5 = (__m128)__lsx_vreplfr2vr_s(0.5f);
-    __m128i _signmask = __lsx_vreplgr2vr_w(1 << 31);
-
-    __m128i _sign = __lsx_vand_v((__m128i)_v, _signmask);
-    __m128 _p5s = (__m128)__lsx_vor_v((__m128i)_p5, _sign);
-    __m128 _v5 = __lsx_vfadd_s(_v, _p5s);
-    __m128i _v32 = __lsx_vftintrz_w_s(_v5);
-
-    __m128i _sign_leaky = __lsx_vand_v((__m128i)_v_leaky, _signmask);
-    __m128 _p5_leaky = (__m128)__lsx_vor_v((__m128i)_p5, _sign_leaky);
-    __m128 _v5_leaky = __lsx_vfadd_s(_v_leaky, _p5_leaky);
-    __m128i _v32_leaky = __lsx_vftintrz_w_s(_v5_leaky);
-
-    __m128i _v32_16 = __lsx_vsat_w(_v32, 15);
-    __m128i _v16 = __lsx_vpickev_h(_v32_16, _v32_16);
-
-    __m128i _v32_16_leaky = __lsx_vsat_w(_v32_leaky, 15);
-    __m128i _v16_leaky = __lsx_vpickev_h(_v32_16_leaky, _v32_16_leaky);
-
-    _v16 = __lsx_vmax_h(_v16, _v16_leaky);
-    __m128i _v16_8 = __lsx_vsat_h(_v16, 7);
-    __m128i _v8 = __lsx_vpickev_b(_v16_8, _v16_8);
-
-    return _v8;
-}
-
-static NCNN_FORCEINLINE int64_t float2int8leakyrelu(__m128 _vlow, __m128 _vhigh, __m128 _slope)
-{
-    __m128 _vlow_leaky = __lsx_vfmul_s(_vlow, _slope);
-    __m128 _vhigh_leaky = __lsx_vfmul_s(_vhigh, _slope);
-
-    // simulate round to nearest via +/-0.5
-    __m128i _p5 = (__m128i)__lsx_vreplfr2vr_s(0.5f);
-    __m128i _signmask = __lsx_vreplgr2vr_w(1 << 31);
-
-    __m128i _signlow = __lsx_vand_v((__m128i)_vlow, _signmask);
-    __m128i _signhigh = __lsx_vand_v((__m128i)_vhigh, _signmask);
-    __m128 _p5low = (__m128)__lsx_vor_v(_p5, _signlow);
-    __m128 _p5high = (__m128)__lsx_vor_v(_p5, _signhigh);
-    __m128 _vlow5 = __lsx_vfadd_s(_vlow, _p5low);
-    __m128 _vhigh5 = __lsx_vfadd_s(_vhigh, _p5high);
-    __m128i _vlow32 = __lsx_vftintrz_w_s(_vlow5);
-    __m128i _vhigh32 = __lsx_vftintrz_w_s(_vhigh5);
-
-    __m128i _signlow_leaky = __lsx_vand_v((__m128i)_vlow_leaky, _signmask);
-    __m128i _signhigh_leaky = __lsx_vand_v((__m128i)_vhigh_leaky, _signmask);
-    __m128 _p5low_leaky = (__m128)__lsx_vor_v(_p5, _signlow_leaky);
-    __m128 _p5high_leaky = (__m128)__lsx_vor_v(_p5, _signhigh_leaky);
-    __m128 _vlow5_leaky = __lsx_vfadd_s(_vlow_leaky, _p5low_leaky);
-    __m128 _vhigh5_leaky = __lsx_vfadd_s(_vhigh_leaky, _p5high_leaky);
-    __m128i _vlow32_leaky = __lsx_vftintrz_w_s(_vlow5_leaky);
-    __m128i _vhigh32_leaky = __lsx_vftintrz_w_s(_vhigh5_leaky);
-
-    __m128i _vlow32_16 = __lsx_vsat_w(_vlow32, 15);
-    __m128i _vhigh32_16 = __lsx_vsat_w(_vhigh32, 15);
-    __m128i _v16 = __lsx_vpickev_h(_vhigh32_16, _vlow32_16);
-
-    __m128i _vlow32_16_leaky = __lsx_vsat_w(_vlow32_leaky, 15);
-    __m128i _vhigh32_16_leaky = __lsx_vsat_w(_vhigh32_leaky, 15);
-    __m128i _v16_leaky = __lsx_vpickev_h(_vhigh32_16_leaky, _vlow32_16_leaky);
-
-    _v16 = __lsx_vmax_h(_v16, _v16_leaky);
-    __m128i _v16_8 = __lsx_vsat_h(_v16, 7);
-    __m128i _v8 = __lsx_vpickev_b(_v16_8, _v16_8);
-
-    return _v8[0];
-}
-#endif // __loongarch_sx
-
-#endif // LOONGARCH_USABILITY_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/lsx_mathfun.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/lsx_mathfun.h
deleted file mode 100644
index 194f63be..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/lsx_mathfun.h
+++ /dev/null
@@ -1,272 +0,0 @@
-/* LOONGARCH implementation of exp
- *
- *   Inspired by Intel Approximate Math library, and based on the
- *   corresponding algorithms of the cephes math library
- *   Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
- */
-
-/*
- *  This software is provided 'as-is', without any express or implied
- *  warranty.  In no event will the authors be held liable for any damages
- *  arising from the use of this software.
- *
- *  Permission is granted to anyone to use this software for any purpose,
- *  including commercial applications, and to alter it and redistribute it
- *  freely, subject to the following restrictions:
- *
- *  1. The origin of this software must not be misrepresented; you must not
- *     claim that you wrote the original software. If you use this software
- *     in a product, an acknowledgment in the product documentation would be
- *     appreciated but is not required.
- *  2. Altered source versions must be plainly marked as such, and must not be
- *     misrepresented as being the original software.
- *  3. This notice may not be removed or altered from any source distribution.
- *
- *  (this is the zlib license)
- */
-
-#ifndef LSX_MATHFUN_H
-#define LSX_MATHFUN_H
-
-#include "loongarch_usability.h"
-
-#include <lsxintrin.h>
-
-_LOONGARCH_FLOAT_CONST(c_1, 1.0f);
-_LOONGARCH_FLOAT_CONST(c_2, 2.0f);
-_LOONGARCH_FLOAT_CONST(c_n1, -1.0f);
-_LOONGARCH_FLOAT_CONST(c_0p5, 0.5f);
-
-#define c_inv_mant_mask ~0x7f800000u
-_LOONGARCH_FLOAT_CONST(c_cephes_SQRTHF, 0.707106781186547524);
-_LOONGARCH_FLOAT_CONST(c_cephes_log_p0, 7.0376836292E-2);
-_LOONGARCH_FLOAT_CONST(c_cephes_log_p1, -1.1514610310E-1);
-_LOONGARCH_FLOAT_CONST(c_cephes_log_p2, 1.1676998740E-1);
-_LOONGARCH_FLOAT_CONST(c_cephes_log_p3, -1.2420140846E-1);
-_LOONGARCH_FLOAT_CONST(c_cephes_log_p4, +1.4249322787E-1);
-_LOONGARCH_FLOAT_CONST(c_cephes_log_p5, -1.6668057665E-1);
-_LOONGARCH_FLOAT_CONST(c_cephes_log_p6, +2.0000714765E-1);
-_LOONGARCH_FLOAT_CONST(c_cephes_log_p7, -2.4999993993E-1);
-_LOONGARCH_FLOAT_CONST(c_cephes_log_p8, +3.3333331174E-1);
-_LOONGARCH_FLOAT_CONST(c_cephes_log_q1, -2.12194440e-4);
-_LOONGARCH_FLOAT_CONST(c_cephes_log_q2, 0.693359375);
-
-/* natural logarithm computed for 4 simultaneous float
- *   return NaN for x <= 0
- */
-static inline __m128 log_ps(__m128 x)
-{
-    __m128 one = (__m128)__lsx_vreplgr2vr_w(c_1.i);
-
-    x = __lsx_vfmax_s(x, (__m128)__lsx_vreplgr2vr_w(0)); /* force flush to zero on denormal values */
-    __m128i invalid_mask = __lsx_vfcmp_cle_s(x, (__m128)__lsx_vreplgr2vr_w(0));
-
-    __m128i ux = (__m128i)(x);
-
-    __m128i emm0 = __lsx_vsrl_w(ux, __lsx_vreplgr2vr_w(23));
-
-    /* keep only the fractional part */
-    ux = __lsx_vand_v(ux, __lsx_vreplgr2vr_w(c_inv_mant_mask));
-    ux = __lsx_vor_v(ux, __lsx_vreplgr2vr_w(c_0p5.i));
-    x = (__m128)(ux);
-
-    emm0 = __lsx_vsub_w(emm0, __lsx_vreplgr2vr_w(0x7f));
-    __m128 e = __lsx_vffint_s_w(emm0);
-
-    e = __lsx_vfadd_s(e, one);
-
-    /* part2:
-     *     if( x < SQRTHF ) {
-     *       e -= 1;
-     *       x = x + x - 1.0;
-     *     } else { x = x - 1.0; }
-     */
-    __m128i mask = __lsx_vfcmp_clt_s((__m128)x, (__m128)__lsx_vreplgr2vr_w(c_cephes_SQRTHF.i));
-    __m128 tmp = (__m128)(__lsx_vand_v((__m128i)(x), (__m128i)mask));
-    x = __lsx_vfsub_s(x, one);
-    e = __lsx_vfsub_s(e, (__m128)(__lsx_vand_v((__m128i)(one), (__m128i)mask)));
-    x = __lsx_vfadd_s(x, tmp);
-
-    __m128 z = __lsx_vfmul_s(x, x);
-
-    __m128 y = (__m128)__lsx_vreplgr2vr_w(c_cephes_log_p0.i);
-
-    y = __lsx_vfmadd_s(x, y, (__m128)__lsx_vreplgr2vr_w(c_cephes_log_p1.i));
-    y = __lsx_vfmadd_s(x, y, (__m128)__lsx_vreplgr2vr_w(c_cephes_log_p2.i));
-    y = __lsx_vfmadd_s(x, y, (__m128)__lsx_vreplgr2vr_w(c_cephes_log_p3.i));
-    y = __lsx_vfmadd_s(x, y, (__m128)__lsx_vreplgr2vr_w(c_cephes_log_p4.i));
-    y = __lsx_vfmadd_s(x, y, (__m128)__lsx_vreplgr2vr_w(c_cephes_log_p5.i));
-    y = __lsx_vfmadd_s(x, y, (__m128)__lsx_vreplgr2vr_w(c_cephes_log_p6.i));
-    y = __lsx_vfmadd_s(x, y, (__m128)__lsx_vreplgr2vr_w(c_cephes_log_p7.i));
-    y = __lsx_vfmadd_s(x, y, (__m128)__lsx_vreplgr2vr_w(c_cephes_log_p8.i));
-    y = __lsx_vfmul_s(y, x);
-
-    y = __lsx_vfmul_s(y, z);
-
-    tmp = __lsx_vfmul_s(e, (__m128)__lsx_vreplgr2vr_w(c_cephes_log_q1.i));
-    y = __lsx_vfadd_s(y, tmp);
-
-    tmp = __lsx_vfmul_s(z, (__m128)__lsx_vreplgr2vr_w(c_0p5.i));
-    y = __lsx_vfsub_s(y, tmp);
-
-    tmp = __lsx_vfmul_s(e, (__m128)__lsx_vreplgr2vr_w(c_cephes_log_q2.i));
-    x = __lsx_vfadd_s(x, y);
-    x = __lsx_vfadd_s(x, tmp);
-    x = (__m128)(__lsx_vor_v((__m128i)(x), (__m128i)invalid_mask)); // negative arg will be NAN
-    return x;
-}
-
-_LOONGARCH_FLOAT_CONST(c_exp_hi, 88.3762626647949f);
-_LOONGARCH_FLOAT_CONST(c_exp_lo, -88.3762626647949f);
-
-_LOONGARCH_FLOAT_CONST(c_cephes_LOG2EF, 1.44269504088896341);
-_LOONGARCH_FLOAT_CONST(c_cephes_exp_C1, 0.693359375);
-_LOONGARCH_FLOAT_CONST(c_cephes_exp_C2, -2.12194440e-4);
-
-_LOONGARCH_FLOAT_CONST(c_cephes_exp_p0, 1.9875691500E-4);
-_LOONGARCH_FLOAT_CONST(c_cephes_exp_p1, 1.3981999507E-3);
-_LOONGARCH_FLOAT_CONST(c_cephes_exp_p2, 8.3334519073E-3);
-_LOONGARCH_FLOAT_CONST(c_cephes_exp_p3, 4.1665795894E-2);
-_LOONGARCH_FLOAT_CONST(c_cephes_exp_p4, 1.6666665459E-1);
-_LOONGARCH_FLOAT_CONST(c_cephes_exp_p5, 5.0000001201E-1);
-
-/* exp() computed for 4 float at once */
-static inline __m128 exp_ps(__m128 x)
-{
-    __m128 tmp, fx;
-
-    __m128 one = (__m128)__lsx_vreplgr2vr_w(c_1.i);
-    x = __lsx_vfmin_s(x, (__m128)__lsx_vreplgr2vr_w(c_exp_hi.i));
-    x = __lsx_vfmax_s(x, (__m128)__lsx_vreplgr2vr_w(c_exp_lo.i));
-
-    /* express exp(x) as exp(g + n*log(2)) */
-    fx = __lsx_vfmul_s(x, (__m128)__lsx_vreplgr2vr_w(c_cephes_LOG2EF.i));
-    fx = __lsx_vfadd_s(fx, (__m128)__lsx_vreplgr2vr_w(c_0p5.i));
-
-    /* perform a floorf */
-    tmp = __lsx_vffint_s_w(__lsx_vftint_w_s(fx));
-
-    /* if greater, substract 1 */
-    __m128i mask = __lsx_vfcmp_clt_s(fx, tmp);
-    mask = __lsx_vand_v(mask, (__m128i)one);
-
-    fx = __lsx_vfsub_s(tmp, (__m128)mask);
-
-    tmp = __lsx_vfmul_s(fx, (__m128)__lsx_vreplgr2vr_w(c_cephes_exp_C1.i));
-    __m128 z = __lsx_vfmul_s(fx, (__m128)__lsx_vreplgr2vr_w(c_cephes_exp_C2.i));
-    x = __lsx_vfsub_s(x, tmp);
-    x = __lsx_vfsub_s(x, z);
-
-    __m128 y = (__m128)__lsx_vreplgr2vr_w(c_cephes_exp_p0.i);
-
-    z = __lsx_vfmul_s(x, x);
-
-    y = __lsx_vfmadd_s(x, y, (__m128)__lsx_vreplgr2vr_w(c_cephes_exp_p1.i));
-    y = __lsx_vfmadd_s(x, y, (__m128)__lsx_vreplgr2vr_w(c_cephes_exp_p2.i));
-    y = __lsx_vfmadd_s(x, y, (__m128)__lsx_vreplgr2vr_w(c_cephes_exp_p3.i));
-    y = __lsx_vfmadd_s(x, y, (__m128)__lsx_vreplgr2vr_w(c_cephes_exp_p4.i));
-    y = __lsx_vfmadd_s(x, y, (__m128)__lsx_vreplgr2vr_w(c_cephes_exp_p5.i));
-
-    y = __lsx_vfmul_s(y, z);
-    y = __lsx_vfadd_s(y, x);
-    y = __lsx_vfadd_s(y, one);
-
-    /* build 2^n */
-    __m128i mm;
-    mm = __lsx_vftintrz_w_s(fx);
-    mm = __lsx_vadd_w(mm, __lsx_vreplgr2vr_w(0x7f));
-    mm = __lsx_vsll_w(mm, __lsx_vreplgr2vr_w(23));
-
-    y = __lsx_vfmul_s(y, (__m128)mm);
-    return y;
-}
-
-_LOONGARCH_FLOAT_CONST(c_tanh_tiny, 1e-4f);
-_LOONGARCH_FLOAT_CONST(c_tanh_hi, 9.0f);
-// The monomial coefficients of the numerator polynomial (odd).
-_LOONGARCH_FLOAT_CONST(c_tanh_alpha_1, 4.89352455891786e-3f);
-_LOONGARCH_FLOAT_CONST(c_tanh_alpha_3, 6.37261928875436e-4f);
-_LOONGARCH_FLOAT_CONST(c_tanh_alpha_5, 1.48572235717979e-5f);
-_LOONGARCH_FLOAT_CONST(c_tanh_alpha_7, 5.12229709037114e-8f);
-_LOONGARCH_FLOAT_CONST(c_tanh_alpha_9, -8.60467152213735e-11f);
-_LOONGARCH_FLOAT_CONST(c_tanh_alpha_11, 2.00018790482477e-13f);
-_LOONGARCH_FLOAT_CONST(c_tanh_alpha_13, -2.76076847742355e-16f);
-// The monomial coefficients of the denominator polynomial (even).
-_LOONGARCH_FLOAT_CONST(c_tanh_beta_0, 4.89352518554385e-3f);
-_LOONGARCH_FLOAT_CONST(c_tanh_beta_2, 2.26843463243900e-3f);
-_LOONGARCH_FLOAT_CONST(c_tanh_beta_4, 1.18534705686654e-4f);
-_LOONGARCH_FLOAT_CONST(c_tanh_beta_6, 1.19825839466702e-6f);
-
-/* tanh() computed for 4 float at once */
-static inline __m128 tanh_ps(__m128 x)
-{
-    __m128 x2 = (__m128)__lsx_vbitclri_w((__m128i)x, 31);
-    __m128i tiny_mask = __lsx_vfcmp_clt_s((__m128)x2, (__m128)(__m128)__lsx_vreplgr2vr_w(c_tanh_tiny.i));
-    __m128i sig_mask = __lsx_vreplgr2vr_w(1 << 31);
-    __m128i sig_save = __lsx_vand_v((__m128i)x, sig_mask);
-
-    // clamp the inputs to the range [-9, 9] since anything outside
-    // this range is -/+1.0f in single-precision.
-    x2 = (__m128)__lsx_vbitsel_v((__m128i)x2, (__m128i)__lsx_vreplgr2vr_w(c_tanh_hi.i), (__m128i)__lsx_vfcmp_clt_s((__m128)__lsx_vreplgr2vr_w(c_tanh_hi.i), (__m128)x2));
-
-    // since the polynomials are odd/even, we need x**2.
-    __m128 z = __lsx_vfmul_s(x2, x2);
-
-    // evaluate the numerator polynomial y.
-    __m128 y = (__m128)__lsx_vreplgr2vr_w(c_tanh_alpha_13.i);
-    y = __lsx_vfmadd_s(z, y, (__m128)__lsx_vreplgr2vr_w(c_tanh_alpha_11.i));
-    y = __lsx_vfmadd_s(z, y, (__m128)__lsx_vreplgr2vr_w(c_tanh_alpha_9.i));
-    y = __lsx_vfmadd_s(z, y, (__m128)__lsx_vreplgr2vr_w(c_tanh_alpha_7.i));
-    y = __lsx_vfmadd_s(z, y, (__m128)__lsx_vreplgr2vr_w(c_tanh_alpha_5.i));
-    y = __lsx_vfmadd_s(z, y, (__m128)__lsx_vreplgr2vr_w(c_tanh_alpha_3.i));
-    y = __lsx_vfmadd_s(z, y, (__m128)__lsx_vreplgr2vr_w(c_tanh_alpha_1.i));
-    y = __lsx_vfmul_s(y, x2);
-
-    // evaluate the denominator polynomial w.
-    __m128 w = (__m128)__lsx_vreplgr2vr_w(c_tanh_beta_6.i);
-    w = __lsx_vfmadd_s(z, w, (__m128)__lsx_vreplgr2vr_w(c_tanh_beta_4.i));
-    w = __lsx_vfmadd_s(z, w, (__m128)__lsx_vreplgr2vr_w(c_tanh_beta_2.i));
-    w = __lsx_vfmadd_s(z, w, (__m128)__lsx_vreplgr2vr_w(c_tanh_beta_0.i));
-
-    // divide the numerator by the denominator.
-    y = __lsx_vfdiv_s(y, w);
-
-    // reinstate the sign.
-    y = (__m128)__lsx_vor_v((__m128i)y, sig_save);
-
-    // when the argument is very small in magnitude it's more accurate to just return it.
-    y = (__m128)__lsx_vbitsel_v((__m128i)y, (__m128i)x, (__m128i)tiny_mask);
-
-    return y;
-}
-
-static inline __m128 pow_ps(__m128 a, __m128 b)
-{
-    // pow(x, m) = exp(m * log(x))
-    return exp_ps(__lsx_vfmul_s(b, log_ps(a)));
-}
-
-static inline __m128 sigmoid_ps(__m128 _v)
-{
-    __m128 _one = __lsx_vreplfr2vr_s(1.f);
-    _v = (__m128)__lsx_vbitrevi_w((__m128i)_v, 31);
-    _v = exp_ps(_v);
-    _v = __lsx_vfadd_s(_v, _one);
-    return __lsx_vfdiv_s(_one, _v);
-}
-
-static inline __m128 atan2_ps(__m128 a, __m128 b)
-{
-    //TODO lsx optimize
-    float tmpx[4];
-    float tmpy[4];
-    __lsx_vst(a, tmpx, 0);
-    __lsx_vst(b, tmpy, 0);
-    tmpx[0] = atan2(tmpx[0], tmpy[0]);
-    tmpx[1] = atan2(tmpx[1], tmpy[1]);
-    tmpx[2] = atan2(tmpx[2], tmpy[2]);
-    tmpx[3] = atan2(tmpx[3], tmpy[3]);
-    return (__m128)__lsx_vld(tmpx, 0);
-}
-
-#endif // LSX_MATHFUN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/mish_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/mish_loongarch.h
deleted file mode 100644
index 0c796758..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/mish_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MISH_LOONGARCH_H
-#define LAYER_MISH_LOONGARCH_H
-
-#include "mish.h"
-
-namespace ncnn {
-
-class Mish_loongarch : public Mish
-{
-public:
-    Mish_loongarch();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MISH_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/packing_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/packing_loongarch.h
deleted file mode 100644
index 476ebd33..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/packing_loongarch.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PACKING_LOONGARCH_H
-#define LAYER_PACKING_LOONGARCH_H
-
-#include "packing.h"
-
-namespace ncnn {
-
-class Packing_loongarch : public Packing
-{
-public:
-    Packing_loongarch();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PACKING_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/padding_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/padding_loongarch.h
deleted file mode 100644
index de416464..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/padding_loongarch.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PADDING_LOONGARCH_H
-#define LAYER_PADDING_LOONGARCH_H
-
-#include "padding.h"
-
-namespace ncnn {
-
-class Padding_loongarch : public Padding
-{
-public:
-    Padding_loongarch();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PADDING_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/padding_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/padding_pack4.h
deleted file mode 100644
index d040ce77..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/padding_pack4.h
+++ /dev/null
@@ -1,213 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void padding_constant_pack4_lsx(const Mat& src, Mat& dst, int top, int bottom, int left, int right, __m128 v)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-    int top_size = top * dst.w;
-    int bottom_size = bottom * dst.w;
-
-    // fill top
-    for (int y = 0; y < top_size; y++)
-    {
-        __lsx_vst(v, outptr, 0);
-        outptr += 4;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            __lsx_vst(v, outptr, 0);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            __builtin_prefetch(ptr + 32);
-            __lsx_vst(__lsx_vld(ptr, 0), outptr, 0);
-            ptr += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __lsx_vst(v, outptr, 0);
-            outptr += 4;
-        }
-    }
-    // fill top
-    for (int y = 0; y < bottom_size; y++)
-    {
-        __lsx_vst(v, outptr, 0);
-        outptr += 4;
-    }
-}
-
-static void padding_replicate_pack4_lsx(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-
-    // fill top
-    for (int y = 0; y < top; y++)
-    {
-        const float* ptr0 = ptr;
-        __m128 _p = (__m128)__lsx_vld(ptr0, 0);
-        for (int x = 0; x < left; x++)
-        {
-            __lsx_vst(_p, outptr, 0);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = (__m128)__lsx_vld(ptr0, 0);
-            __lsx_vst(_p, outptr, 0);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __lsx_vst(_p, outptr, 0);
-            outptr += 4;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        __m128 _p = (__m128)__lsx_vld(ptr, 0);
-        for (int x = 0; x < left; x++)
-        {
-            __lsx_vst(_p, outptr, 0);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = (__m128)__lsx_vld(ptr, 0);
-            __lsx_vst(_p, outptr, 0);
-            ptr += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __lsx_vst(_p, outptr, 0);
-            outptr += 4;
-        }
-    }
-    // fill bottom
-    ptr -= src.w * 4;
-    for (int y = 0; y < bottom; y++)
-    {
-        const float* ptr0 = ptr;
-        __m128 _p = (__m128)__lsx_vld(ptr0, 0);
-        for (int x = 0; x < left; x++)
-        {
-            __lsx_vst(_p, outptr, 0);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = (__m128)__lsx_vld(ptr0, 0);
-            __lsx_vst(_p, outptr, 0);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __lsx_vst(_p, outptr, 0);
-            outptr += 4;
-        }
-    }
-}
-
-static void padding_reflect_pack4_lsx(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-
-    // fill top
-    ptr += top * src.w * 4;
-    for (int y = 0; y < top; y++)
-    {
-        const float* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            __m128 _p = (__m128)__lsx_vld(ptr0 + (left - x) * 4, 0);
-            __lsx_vst(_p, outptr, 0);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            __m128 _p = (__m128)__lsx_vld(ptr0, 0);
-            __lsx_vst(_p, outptr, 0);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __m128 _p = (__m128)__lsx_vld(ptr0 - 8 - x * 4, 0);
-            __lsx_vst(_p, outptr, 0);
-            outptr += 4;
-        }
-        ptr -= src.w * 4;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            __m128 _p = (__m128)__lsx_vld(ptr + (left - x) * 4, 0);
-            __lsx_vst(_p, outptr, 0);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            __m128 _p = (__m128)__lsx_vld(ptr, 0);
-            __lsx_vst(_p, outptr, 0);
-            ptr += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __m128 _p = (__m128)__lsx_vld(ptr - 8 - x * 4, 0);
-            __lsx_vst(_p, outptr, 0);
-            outptr += 4;
-        }
-    }
-    // fill bottom
-    ptr -= 2 * src.w * 4;
-    for (int y = 0; y < bottom; y++)
-    {
-        const float* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            __m128 _p = (__m128)__lsx_vld(ptr0 + (left - x) * 4, 0);
-            __lsx_vst(_p, outptr, 0);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            __m128 _p = (__m128)__lsx_vld(ptr0, 0);
-            __lsx_vst(_p, outptr, 0);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __m128 _p = (__m128)__lsx_vld(ptr0 - 8 - x * 4, 0);
-            __lsx_vst(_p, outptr, 0);
-            outptr += 4;
-        }
-        ptr -= src.w * 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/padding_pack8_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/padding_pack8_int8.h
deleted file mode 100644
index 4c6586c6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/padding_pack8_int8.h
+++ /dev/null
@@ -1,171 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void padding_constant_pack8_int8_lsx(const Mat& src, Mat& dst, int top, int bottom, int left, int right, int64_t _v)
-{
-    const int64_t* ptr = src;
-    int64_t* outptr = dst;
-
-    // fill top
-    for (int y = 0; y < top; y++)
-    {
-        for (int x = 0; x < dst.w; x++)
-        {
-            *outptr++ = _v;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = _v;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = _v;
-        }
-    }
-    // fill bottom
-    for (int y = 0; y < bottom; y++)
-    {
-        for (int x = 0; x < dst.w; x++)
-        {
-            *outptr++ = _v;
-        }
-    }
-}
-
-static void padding_replicate_pack8_int8_lsx(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const int64_t* ptr = src;
-    int64_t* outptr = dst;
-
-    // fill top
-    for (int y = 0; y < top; y++)
-    {
-        const int64_t* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = *ptr0;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr0++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr0[-1];
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = *ptr;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr[-1];
-        }
-    }
-    // fill bottom
-    ptr -= src.w;
-    for (int y = 0; y < bottom; y++)
-    {
-        const int64_t* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = *ptr0;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr0++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr0[-1];
-        }
-    }
-}
-
-static void padding_reflect_pack8_int8_lsx(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const int64_t* ptr = src;
-    int64_t* outptr = dst;
-
-    // fill top
-    ptr += top * src.w;
-    for (int y = 0; y < top; y++)
-    {
-        const int64_t* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = ptr0[left - x];
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr0++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr0[-2 - x];
-        }
-        ptr -= src.w;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = ptr[left - x];
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr[-2 - x];
-        }
-    }
-    // fill bottom
-    ptr -= 2 * src.w;
-    for (int y = 0; y < bottom; y++)
-    {
-        const int64_t* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = ptr0[left - x];
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr0++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr0[-2 - x];
-        }
-        ptr -= src.w;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/pooling_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/pooling_loongarch.h
deleted file mode 100644
index 646b1094..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/pooling_loongarch.h
+++ /dev/null
@@ -1,33 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_POOLING_LOONGARCH_H
-#define LAYER_POOLING_LOONGARCH_H
-
-#include "pooling.h"
-
-namespace ncnn {
-
-class Pooling_loongarch : public Pooling
-{
-public:
-    Pooling_loongarch();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_POOLING_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/prelu_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/prelu_loongarch.h
deleted file mode 100644
index bafd7ac4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/prelu_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PRELU_LOONGARCH_H
-#define LAYER_PRELU_LOONGARCH_H
-
-#include "prelu.h"
-
-namespace ncnn {
-
-class PReLU_loongarch : public PReLU
-{
-public:
-    PReLU_loongarch();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PRELU_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/quantize_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/quantize_loongarch.h
deleted file mode 100644
index dcc0d8e0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/quantize_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_QUANTIZE_LOONGARCH_H
-#define LAYER_QUANTIZE_LOONGARCH_H
-
-#include "quantize.h"
-
-namespace ncnn {
-
-class Quantize_loongarch : public Quantize
-{
-public:
-    Quantize_loongarch();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_QUANTIZE_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/relu_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/relu_loongarch.h
deleted file mode 100644
index 6ee6684f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/relu_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_RELU_LOONGARCH_H
-#define LAYER_RELU_LOONGARCH_H
-
-#include "relu.h"
-
-namespace ncnn {
-
-class ReLU_loongarch : public ReLU
-{
-public:
-    ReLU_loongarch();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_RELU_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_leakyrelu_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_leakyrelu_pack4.h
deleted file mode 100644
index d6b49942..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_leakyrelu_pack4.h
+++ /dev/null
@@ -1,271 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_leakyrelu_pack4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, float slope, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-    int size = w * h;
-    int outc = top_blob.c;
-    int out_elempack = top_blob.elempack;
-
-    int scale_in_data_size = scale_in_data.w;
-    int scale_out_data_size = scale_out_data.w;
-    int bias_data_size = bias_data.w;
-
-    // int8(leakyrelu(v * scale_in, slope) * scale_out)
-    // int8_leakyrelu(v * (scale_in * scale_out), slope)
-
-    // int8(leakyrelu(v * scale_in + bias, slope) * scale_out)
-    // int8_leakyrelu(v * (scale_in * scale_out) + (bias * scale_out), slope)
-
-    if (out_elempack == 8)
-    {
-        if (bias_data_size == 0)
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < outc; q++)
-            {
-                const int* intptr0 = bottom_blob.channel(q * 2);
-                const int* intptr1 = bottom_blob.channel(q * 2 + 1);
-                signed char* ptr = top_blob.channel(q);
-
-                __m128 _scale_in0 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8, 0);
-                __m128 _scale_in1 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8 + 4, 0);
-                __m128 _scale_out0 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8, 0);
-                __m128 _scale_out1 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8 + 4, 0);
-
-                __m128 _scale0 = __lsx_vfmul_s(_scale_in0, _scale_out0);
-                __m128 _scale1 = __lsx_vfmul_s(_scale_in1, _scale_out1);
-                __m128 _slope = (__m128)__lsx_vreplfr2vr_s(slope);
-
-                int i = 0;
-                for (; i + 3 < size; i += 4)
-                {
-                    __builtin_prefetch(intptr0 + 64);
-                    __builtin_prefetch(intptr1 + 64);
-                    __m128 _v00 = __lsx_vffint_s_w(__lsx_vld(intptr0, 0));
-                    __m128 _v01 = __lsx_vffint_s_w(__lsx_vld(intptr0 + 4, 0));
-                    __m128 _v02 = __lsx_vffint_s_w(__lsx_vld(intptr0 + 8, 0));
-                    __m128 _v03 = __lsx_vffint_s_w(__lsx_vld(intptr0 + 12, 0));
-                    __m128 _v10 = __lsx_vffint_s_w(__lsx_vld(intptr1, 0));
-                    __m128 _v11 = __lsx_vffint_s_w(__lsx_vld(intptr1 + 4, 0));
-                    __m128 _v12 = __lsx_vffint_s_w(__lsx_vld(intptr1 + 8, 0));
-                    __m128 _v13 = __lsx_vffint_s_w(__lsx_vld(intptr1 + 12, 0));
-                    _v00 = __lsx_vfmul_s(_v00, _scale0);
-                    _v01 = __lsx_vfmul_s(_v01, _scale0);
-                    _v02 = __lsx_vfmul_s(_v02, _scale0);
-                    _v03 = __lsx_vfmul_s(_v03, _scale0);
-                    _v10 = __lsx_vfmul_s(_v10, _scale1);
-                    _v11 = __lsx_vfmul_s(_v11, _scale1);
-                    _v12 = __lsx_vfmul_s(_v12, _scale1);
-                    _v13 = __lsx_vfmul_s(_v13, _scale1);
-                    *((int64_t*)ptr) = float2int8leakyrelu(_v00, _v10, _slope);
-                    *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v01, _v11, _slope);
-                    *((int64_t*)(ptr + 16)) = float2int8leakyrelu(_v02, _v12, _slope);
-                    *((int64_t*)(ptr + 24)) = float2int8leakyrelu(_v03, _v13, _slope);
-
-                    intptr0 += 16;
-                    intptr1 += 16;
-                    ptr += 32;
-                }
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr0 + 16);
-                    __builtin_prefetch(intptr1 + 16);
-                    __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr0, 0));
-                    __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr1, 0));
-                    _v0 = __lsx_vfmul_s(_v0, _scale0);
-                    _v1 = __lsx_vfmul_s(_v1, _scale1);
-                    *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-
-                    intptr0 += 4;
-                    intptr1 += 4;
-                    ptr += 8;
-                }
-            }
-        }
-        else
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < outc; q++)
-            {
-                const int* intptr0 = bottom_blob.channel(q * 2);
-                const int* intptr1 = bottom_blob.channel(q * 2 + 1);
-                signed char* ptr = top_blob.channel(q);
-
-                __m128 _scale_in0 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8, 0);
-                __m128 _scale_in1 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8 + 4, 0);
-                __m128 _scale_out0 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8, 0);
-                __m128 _scale_out1 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8 + 4, 0);
-                __m128 _bias0 = bias_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(bias_data[0]) : (__m128)__lsx_vld((const float*)bias_data + q * 8, 0);
-                __m128 _bias1 = bias_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(bias_data[0]) : (__m128)__lsx_vld((const float*)bias_data + q * 8 + 4, 0);
-
-                __m128 _scale0 = __lsx_vfmul_s(_scale_in0, _scale_out0);
-                __m128 _scale1 = __lsx_vfmul_s(_scale_in1, _scale_out1);
-                _bias0 = __lsx_vfmul_s(_bias0, _scale_out0);
-                _bias1 = __lsx_vfmul_s(_bias1, _scale_out1);
-                __m128 _slope = (__m128)__lsx_vreplfr2vr_s(slope);
-
-                int i = 0;
-                for (; i + 3 < size; i += 4)
-                {
-                    __builtin_prefetch(intptr0 + 64);
-                    __builtin_prefetch(intptr1 + 64);
-                    __m128 _v00 = __lsx_vffint_s_w(__lsx_vld(intptr0, 0));
-                    __m128 _v01 = __lsx_vffint_s_w(__lsx_vld(intptr0 + 4, 0));
-                    __m128 _v02 = __lsx_vffint_s_w(__lsx_vld(intptr0 + 8, 0));
-                    __m128 _v03 = __lsx_vffint_s_w(__lsx_vld(intptr0 + 12, 0));
-                    __m128 _v10 = __lsx_vffint_s_w(__lsx_vld(intptr1, 0));
-                    __m128 _v11 = __lsx_vffint_s_w(__lsx_vld(intptr1 + 4, 0));
-                    __m128 _v12 = __lsx_vffint_s_w(__lsx_vld(intptr1 + 8, 0));
-                    __m128 _v13 = __lsx_vffint_s_w(__lsx_vld(intptr1 + 12, 0));
-                    _v00 = __lsx_vfmadd_s(_scale0, _v00, _bias0);
-                    _v01 = __lsx_vfmadd_s(_scale0, _v01, _bias0);
-                    _v02 = __lsx_vfmadd_s(_scale0, _v02, _bias0);
-                    _v03 = __lsx_vfmadd_s(_scale0, _v03, _bias0);
-                    _v10 = __lsx_vfmadd_s(_scale1, _v10, _bias1);
-                    _v11 = __lsx_vfmadd_s(_scale1, _v11, _bias1);
-                    _v12 = __lsx_vfmadd_s(_scale1, _v12, _bias1);
-                    _v13 = __lsx_vfmadd_s(_scale1, _v13, _bias1);
-                    *((int64_t*)ptr) = float2int8leakyrelu(_v00, _v10, _slope);
-                    *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v01, _v11, _slope);
-                    *((int64_t*)(ptr + 16)) = float2int8leakyrelu(_v02, _v12, _slope);
-                    *((int64_t*)(ptr + 24)) = float2int8leakyrelu(_v03, _v13, _slope);
-
-                    intptr0 += 16;
-                    intptr1 += 16;
-                    ptr += 32;
-                }
-                for (; i + 1 < size; i += 2)
-                {
-                    __builtin_prefetch(intptr0 + 32);
-                    __builtin_prefetch(intptr1 + 32);
-                    __m128 _v00 = __lsx_vffint_s_w(__lsx_vld(intptr0, 0));
-                    __m128 _v01 = __lsx_vffint_s_w(__lsx_vld(intptr0 + 4, 0));
-                    __m128 _v10 = __lsx_vffint_s_w(__lsx_vld(intptr1, 0));
-                    __m128 _v11 = __lsx_vffint_s_w(__lsx_vld(intptr1 + 4, 0));
-                    _v00 = __lsx_vfmadd_s(_scale0, _v00, _bias0);
-                    _v01 = __lsx_vfmadd_s(_scale0, _v01, _bias0);
-                    _v10 = __lsx_vfmadd_s(_scale1, _v10, _bias1);
-                    _v11 = __lsx_vfmadd_s(_scale1, _v11, _bias1);
-                    *((int64_t*)ptr) = float2int8leakyrelu(_v00, _v10, _slope);
-                    *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v01, _v11, _slope);
-
-                    intptr0 += 8;
-                    intptr1 += 8;
-                    ptr += 16;
-                }
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr0 + 16);
-                    __builtin_prefetch(intptr1 + 16);
-                    __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr0, 0));
-                    __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr1, 0));
-                    _v0 = __lsx_vfmadd_s(_scale0, _v0, _bias0);
-                    _v1 = __lsx_vfmadd_s(_scale1, _v1, _bias1);
-                    *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-
-                    intptr0 += 4;
-                    intptr1 += 4;
-                    ptr += 8;
-                }
-            }
-        }
-    }
-    if (out_elempack == 1)
-    {
-        if (bias_data_size == 0)
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < channels; q++)
-            {
-                const int* intptr = bottom_blob.channel(q);
-                signed char* ptr0 = top_blob.channel(q * 4);
-                signed char* ptr1 = top_blob.channel(q * 4 + 1);
-                signed char* ptr2 = top_blob.channel(q * 4 + 2);
-                signed char* ptr3 = top_blob.channel(q * 4 + 3);
-                signed char* vp;
-
-                __m128 _scale_in = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 4, 0);
-                __m128 _scale_out = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 4, 0);
-
-                __m128 _scale = __lsx_vfmul_s(_scale_in, _scale_out);
-                __m128 _slope = (__m128)__lsx_vreplfr2vr_s(slope);
-
-                int i = 0;
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr + 16);
-                    __m128 _v = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                    _v = __lsx_vfmul_s(_v, _scale);
-                    __m128i v = float2int8leakyrelu(_v, _slope);
-                    vp = (signed char*)&v;
-                    ptr0[0] = vp[0];
-                    ptr1[0] = vp[1];
-                    ptr2[0] = vp[2];
-                    ptr3[0] = vp[3];
-
-                    intptr += 4;
-                    ptr0 += 1;
-                    ptr1 += 1;
-                    ptr2 += 1;
-                    ptr3 += 1;
-                }
-            }
-        }
-        else
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < channels; q++)
-            {
-                const int* intptr = bottom_blob.channel(q);
-                signed char* ptr0 = top_blob.channel(q * 4);
-                signed char* ptr1 = top_blob.channel(q * 4 + 1);
-                signed char* ptr2 = top_blob.channel(q * 4 + 2);
-                signed char* ptr3 = top_blob.channel(q * 4 + 3);
-                signed char* vp;
-
-                __m128 _scale_in = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 4, 0);
-                __m128 _scale_out = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 4, 0);
-                __m128 _bias = bias_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(bias_data[0]) : (__m128)__lsx_vld((const float*)bias_data + q * 4, 0);
-
-                __m128 _scale = __lsx_vfmul_s(_scale_in, _scale_out);
-                _bias = __lsx_vfmul_s(_bias, _scale_out);
-                __m128 _slope = (__m128)__lsx_vreplfr2vr_s(slope);
-
-                int i = 0;
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr + 16);
-                    __m128 _v = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                    _v = __lsx_vfmadd_s(_scale, _v, _bias);
-                    __m128i v = float2int8leakyrelu(_v, _slope);
-                    vp = (signed char*)&v;
-                    ptr0[0] = vp[0];
-                    ptr1[0] = vp[1];
-                    ptr2[0] = vp[2];
-                    ptr3[0] = vp[3];
-
-                    intptr += 4;
-                    ptr0 += 1;
-                    ptr1 += 1;
-                    ptr2 += 1;
-                    ptr3 += 1;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_leakyrelu_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_leakyrelu_pack8.h
deleted file mode 100644
index a2c4faed..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_leakyrelu_pack8.h
+++ /dev/null
@@ -1,188 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_leakyrelu_pack8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, float slope, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-    int size = w * h;
-
-    int scale_in_data_size = scale_in_data.w;
-    int scale_out_data_size = scale_out_data.w;
-    int bias_data_size = bias_data.w;
-
-    // int8(leakyrelu(v * scale_in, slope) * scale_out)
-    // int8_leakyrelu(v * (scale_in * scale_out), slope)
-
-    // int8(leakyrelu(v * scale_in + bias, slope) * scale_out)
-    // int8_leakyrelu(v * (scale_in * scale_out) + (bias * scale_out), slope)
-
-    if (bias_data_size == 0)
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int q = 0; q < channels; q++)
-        {
-            const int* intptr = bottom_blob.channel(q);
-            signed char* ptr = top_blob.channel(q);
-
-            __m128 _scale_in0 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8, 0);
-            __m128 _scale_in1 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8 + 4, 0);
-            __m128 _scale_out0 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8, 0);
-            __m128 _scale_out1 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8 + 4, 0);
-
-            __m128 _scale0 = __lsx_vfmul_s(_scale_in0, _scale_out0);
-            __m128 _scale1 = __lsx_vfmul_s(_scale_in1, _scale_out1);
-            __m128 _slope = (__m128)__lsx_vreplfr2vr_s(slope);
-
-            int i = 0;
-            for (; i + 3 < size; i += 4)
-            {
-                __builtin_prefetch(intptr + 128);
-                __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr + 4, 0));
-                __m128 _v2 = __lsx_vffint_s_w(__lsx_vld(intptr + 8, 0));
-                __m128 _v3 = __lsx_vffint_s_w(__lsx_vld(intptr + 12, 0));
-                __m128 _v4 = __lsx_vffint_s_w(__lsx_vld(intptr + 16, 0));
-                __m128 _v5 = __lsx_vffint_s_w(__lsx_vld(intptr + 20, 0));
-                __m128 _v6 = __lsx_vffint_s_w(__lsx_vld(intptr + 24, 0));
-                __m128 _v7 = __lsx_vffint_s_w(__lsx_vld(intptr + 28, 0));
-                _v0 = __lsx_vfmul_s(_v0, _scale0);
-                _v1 = __lsx_vfmul_s(_v1, _scale1);
-                _v2 = __lsx_vfmul_s(_v2, _scale0);
-                _v3 = __lsx_vfmul_s(_v3, _scale1);
-                _v4 = __lsx_vfmul_s(_v4, _scale0);
-                _v5 = __lsx_vfmul_s(_v5, _scale1);
-                _v6 = __lsx_vfmul_s(_v6, _scale0);
-                _v7 = __lsx_vfmul_s(_v7, _scale1);
-                *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-                *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v2, _v3, _slope);
-                *((int64_t*)(ptr + 16)) = float2int8leakyrelu(_v4, _v5, _slope);
-                *((int64_t*)(ptr + 24)) = float2int8leakyrelu(_v6, _v7, _slope);
-
-                intptr += 32;
-                ptr += 32;
-            }
-            for (; i + 1 < size; i += 2)
-            {
-                __builtin_prefetch(intptr + 64);
-                __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr + 4, 0));
-                __m128 _v2 = __lsx_vffint_s_w(__lsx_vld(intptr + 8, 0));
-                __m128 _v3 = __lsx_vffint_s_w(__lsx_vld(intptr + 12, 0));
-                _v0 = __lsx_vfmul_s(_v0, _scale0);
-                _v1 = __lsx_vfmul_s(_v1, _scale1);
-                _v2 = __lsx_vfmul_s(_v2, _scale0);
-                _v3 = __lsx_vfmul_s(_v3, _scale1);
-                *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-                *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v2, _v3, _slope);
-
-                intptr += 16;
-                ptr += 16;
-            }
-            for (; i < size; i++)
-            {
-                __builtin_prefetch(intptr + 32);
-                __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr + 4, 0));
-                _v0 = __lsx_vfmul_s(_v0, _scale0);
-                _v1 = __lsx_vfmul_s(_v1, _scale1);
-                *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-
-                intptr += 8;
-                ptr += 8;
-            }
-        }
-    }
-    else
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int q = 0; q < channels; q++)
-        {
-            const int* intptr = bottom_blob.channel(q);
-            signed char* ptr = top_blob.channel(q);
-
-            __m128 _scale_in0 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8, 0);
-            __m128 _scale_in1 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8 + 4, 0);
-            __m128 _scale_out0 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8, 0);
-            __m128 _scale_out1 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8 + 4, 0);
-            __m128 _bias0 = bias_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(bias_data[0]) : (__m128)__lsx_vld((const float*)bias_data + q * 8, 0);
-            __m128 _bias1 = bias_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(bias_data[0]) : (__m128)__lsx_vld((const float*)bias_data + q * 8 + 4, 0);
-
-            __m128 _scale0 = __lsx_vfmul_s(_scale_in0, _scale_out0);
-            __m128 _scale1 = __lsx_vfmul_s(_scale_in1, _scale_out1);
-            _bias0 = __lsx_vfmul_s(_bias0, _scale_out0);
-            _bias1 = __lsx_vfmul_s(_bias1, _scale_out1);
-            __m128 _slope = (__m128)__lsx_vreplfr2vr_s(slope);
-
-            int i = 0;
-            for (; i + 3 < size; i += 4)
-            {
-                __builtin_prefetch(intptr + 128);
-                __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr + 4, 0));
-                __m128 _v2 = __lsx_vffint_s_w(__lsx_vld(intptr + 8, 0));
-                __m128 _v3 = __lsx_vffint_s_w(__lsx_vld(intptr + 12, 0));
-                __m128 _v4 = __lsx_vffint_s_w(__lsx_vld(intptr + 16, 0));
-                __m128 _v5 = __lsx_vffint_s_w(__lsx_vld(intptr + 20, 0));
-                __m128 _v6 = __lsx_vffint_s_w(__lsx_vld(intptr + 24, 0));
-                __m128 _v7 = __lsx_vffint_s_w(__lsx_vld(intptr + 28, 0));
-                _v0 = __lsx_vfmadd_s(_scale0, _v0, _bias0);
-                _v1 = __lsx_vfmadd_s(_scale1, _v1, _bias1);
-                _v2 = __lsx_vfmadd_s(_scale0, _v2, _bias0);
-                _v3 = __lsx_vfmadd_s(_scale1, _v3, _bias1);
-                _v4 = __lsx_vfmadd_s(_scale0, _v4, _bias0);
-                _v5 = __lsx_vfmadd_s(_scale1, _v5, _bias1);
-                _v6 = __lsx_vfmadd_s(_scale0, _v6, _bias0);
-                _v7 = __lsx_vfmadd_s(_scale1, _v7, _bias1);
-                *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-                *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v2, _v3, _slope);
-                *((int64_t*)(ptr + 16)) = float2int8leakyrelu(_v4, _v5, _slope);
-                *((int64_t*)(ptr + 24)) = float2int8leakyrelu(_v6, _v7, _slope);
-
-                intptr += 32;
-                ptr += 32;
-            }
-            for (; i + 1 < size; i += 2)
-            {
-                __builtin_prefetch(intptr + 64);
-                __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr + 4, 0));
-                __m128 _v2 = __lsx_vffint_s_w(__lsx_vld(intptr + 8, 0));
-                __m128 _v3 = __lsx_vffint_s_w(__lsx_vld(intptr + 12, 0));
-                _v0 = __lsx_vfmadd_s(_scale0, _v0, _bias0);
-                _v1 = __lsx_vfmadd_s(_scale1, _v1, _bias1);
-                _v2 = __lsx_vfmadd_s(_scale0, _v2, _bias0);
-                _v3 = __lsx_vfmadd_s(_scale1, _v3, _bias1);
-                *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-                *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v2, _v3, _slope);
-
-                intptr += 16;
-                ptr += 16;
-            }
-            for (; i < size; i++)
-            {
-                __builtin_prefetch(intptr + 32);
-                __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr + 4, 0));
-                _v0 = __lsx_vfmadd_s(_scale0, _v0, _bias0);
-                _v1 = __lsx_vfmadd_s(_scale1, _v1, _bias1);
-                *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-
-                intptr += 8;
-                ptr += 8;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_loongarch.h
deleted file mode 100644
index 4afaf9df..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_REQUANTIZE_LOONGARCH_H
-#define LAYER_REQUANTIZE_LOONGARCH_H
-
-#include "requantize.h"
-
-namespace ncnn {
-
-class Requantize_loongarch : public Requantize
-{
-public:
-    Requantize_loongarch();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_REQUANTIZE_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_relu_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_relu_pack4.h
deleted file mode 100644
index 2fba8dfc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_relu_pack4.h
+++ /dev/null
@@ -1,267 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_relu_pack4_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-    int size = w * h;
-    int outc = top_blob.c;
-    int out_elempack = top_blob.elempack;
-
-    int scale_in_data_size = scale_in_data.w;
-    int scale_out_data_size = scale_out_data.w;
-    int bias_data_size = bias_data.w;
-
-    // int8(relu(v * scale_in) * scale_out)
-    // int8_relu(v * (scale_in * scale_out))
-
-    // int8(relu(v * scale_in + bias) * scale_out)
-    // int8_relu(v * (scale_in * scale_out) + (bias * scale_out))
-
-    if (out_elempack == 8)
-    {
-        if (bias_data_size == 0)
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < outc; q++)
-            {
-                const int* intptr0 = bottom_blob.channel(q * 2);
-                const int* intptr1 = bottom_blob.channel(q * 2 + 1);
-                signed char* ptr = top_blob.channel(q);
-
-                __m128 _scale_in0 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8, 0);
-                __m128 _scale_in1 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8 + 4, 0);
-                __m128 _scale_out0 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8, 0);
-                __m128 _scale_out1 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8 + 4, 0);
-
-                __m128 _scale0 = __lsx_vfmul_s(_scale_in0, _scale_out0);
-                __m128 _scale1 = __lsx_vfmul_s(_scale_in1, _scale_out1);
-
-                int i = 0;
-                for (; i + 3 < size; i += 4)
-                {
-                    __builtin_prefetch(intptr0 + 64);
-                    __builtin_prefetch(intptr1 + 64);
-                    __m128 _v00 = __lsx_vffint_s_w(__lsx_vld(intptr0, 0));
-                    __m128 _v01 = __lsx_vffint_s_w(__lsx_vld(intptr0 + 4, 0));
-                    __m128 _v02 = __lsx_vffint_s_w(__lsx_vld(intptr0 + 8, 0));
-                    __m128 _v03 = __lsx_vffint_s_w(__lsx_vld(intptr0 + 12, 0));
-                    __m128 _v10 = __lsx_vffint_s_w(__lsx_vld(intptr1, 0));
-                    __m128 _v11 = __lsx_vffint_s_w(__lsx_vld(intptr1 + 4, 0));
-                    __m128 _v12 = __lsx_vffint_s_w(__lsx_vld(intptr1 + 8, 0));
-                    __m128 _v13 = __lsx_vffint_s_w(__lsx_vld(intptr1 + 12, 0));
-                    _v00 = __lsx_vfmul_s(_v00, _scale0);
-                    _v01 = __lsx_vfmul_s(_v01, _scale0);
-                    _v02 = __lsx_vfmul_s(_v02, _scale0);
-                    _v03 = __lsx_vfmul_s(_v03, _scale0);
-                    _v10 = __lsx_vfmul_s(_v10, _scale1);
-                    _v11 = __lsx_vfmul_s(_v11, _scale1);
-                    _v12 = __lsx_vfmul_s(_v12, _scale1);
-                    _v13 = __lsx_vfmul_s(_v13, _scale1);
-                    *((int64_t*)ptr) = float2int8relu(_v00, _v10);
-                    *((int64_t*)(ptr + 8)) = float2int8relu(_v01, _v11);
-                    *((int64_t*)(ptr + 16)) = float2int8relu(_v02, _v12);
-                    *((int64_t*)(ptr + 24)) = float2int8relu(_v03, _v13);
-
-                    intptr0 += 16;
-                    intptr1 += 16;
-                    ptr += 32;
-                }
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr0 + 16);
-                    __builtin_prefetch(intptr1 + 16);
-                    __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr0, 0));
-                    __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr1, 0));
-                    _v0 = __lsx_vfmul_s(_v0, _scale0);
-                    _v1 = __lsx_vfmul_s(_v1, _scale1);
-                    *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-
-                    intptr0 += 4;
-                    intptr1 += 4;
-                    ptr += 8;
-                }
-            }
-        }
-        else
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < outc; q++)
-            {
-                const int* intptr0 = bottom_blob.channel(q * 2);
-                const int* intptr1 = bottom_blob.channel(q * 2 + 1);
-                signed char* ptr = top_blob.channel(q);
-
-                __m128 _scale_in0 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8, 0);
-                __m128 _scale_in1 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8 + 4, 0);
-                __m128 _scale_out0 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8, 0);
-                __m128 _scale_out1 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8 + 4, 0);
-                __m128 _bias0 = bias_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(bias_data[0]) : (__m128)__lsx_vld((const float*)bias_data + q * 8, 0);
-                __m128 _bias1 = bias_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(bias_data[0]) : (__m128)__lsx_vld((const float*)bias_data + q * 8 + 4, 0);
-
-                __m128 _scale0 = __lsx_vfmul_s(_scale_in0, _scale_out0);
-                __m128 _scale1 = __lsx_vfmul_s(_scale_in1, _scale_out1);
-                _bias0 = __lsx_vfmul_s(_bias0, _scale_out0);
-                _bias1 = __lsx_vfmul_s(_bias1, _scale_out1);
-
-                int i = 0;
-                for (; i + 3 < size; i += 4)
-                {
-                    __builtin_prefetch(intptr0 + 64);
-                    __builtin_prefetch(intptr1 + 64);
-                    __m128 _v00 = __lsx_vffint_s_w(__lsx_vld(intptr0, 0));
-                    __m128 _v01 = __lsx_vffint_s_w(__lsx_vld(intptr0 + 4, 0));
-                    __m128 _v02 = __lsx_vffint_s_w(__lsx_vld(intptr0 + 8, 0));
-                    __m128 _v03 = __lsx_vffint_s_w(__lsx_vld(intptr0 + 12, 0));
-                    __m128 _v10 = __lsx_vffint_s_w(__lsx_vld(intptr1, 0));
-                    __m128 _v11 = __lsx_vffint_s_w(__lsx_vld(intptr1 + 4, 0));
-                    __m128 _v12 = __lsx_vffint_s_w(__lsx_vld(intptr1 + 8, 0));
-                    __m128 _v13 = __lsx_vffint_s_w(__lsx_vld(intptr1 + 12, 0));
-                    _v00 = __lsx_vfmadd_s(_scale0, _v00, _bias0);
-                    _v01 = __lsx_vfmadd_s(_scale0, _v01, _bias0);
-                    _v02 = __lsx_vfmadd_s(_scale0, _v02, _bias0);
-                    _v03 = __lsx_vfmadd_s(_scale0, _v03, _bias0);
-                    _v10 = __lsx_vfmadd_s(_scale1, _v10, _bias1);
-                    _v11 = __lsx_vfmadd_s(_scale1, _v11, _bias1);
-                    _v12 = __lsx_vfmadd_s(_scale1, _v12, _bias1);
-                    _v13 = __lsx_vfmadd_s(_scale1, _v13, _bias1);
-                    *((int64_t*)ptr) = float2int8relu(_v00, _v10);
-                    *((int64_t*)(ptr + 8)) = float2int8relu(_v01, _v11);
-                    *((int64_t*)(ptr + 16)) = float2int8relu(_v02, _v12);
-                    *((int64_t*)(ptr + 24)) = float2int8relu(_v03, _v13);
-
-                    intptr0 += 16;
-                    intptr1 += 16;
-                    ptr += 32;
-                }
-                for (; i + 1 < size; i += 2)
-                {
-                    __builtin_prefetch(intptr0 + 32);
-                    __builtin_prefetch(intptr1 + 32);
-                    __m128 _v00 = __lsx_vffint_s_w(__lsx_vld(intptr0, 0));
-                    __m128 _v01 = __lsx_vffint_s_w(__lsx_vld(intptr0 + 4, 0));
-                    __m128 _v10 = __lsx_vffint_s_w(__lsx_vld(intptr1, 0));
-                    __m128 _v11 = __lsx_vffint_s_w(__lsx_vld(intptr1 + 4, 0));
-                    _v00 = __lsx_vfmadd_s(_scale0, _v00, _bias0);
-                    _v01 = __lsx_vfmadd_s(_scale0, _v01, _bias0);
-                    _v10 = __lsx_vfmadd_s(_scale1, _v10, _bias1);
-                    _v11 = __lsx_vfmadd_s(_scale1, _v11, _bias1);
-                    *((int64_t*)ptr) = float2int8relu(_v00, _v10);
-                    *((int64_t*)(ptr + 8)) = float2int8relu(_v01, _v11);
-
-                    intptr0 += 8;
-                    intptr1 += 8;
-                    ptr += 16;
-                }
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr0 + 16);
-                    __builtin_prefetch(intptr1 + 16);
-                    __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr0, 0));
-                    __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr1, 0));
-                    _v0 = __lsx_vfmadd_s(_scale0, _v0, _bias0);
-                    _v1 = __lsx_vfmadd_s(_scale1, _v1, _bias1);
-                    *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-
-                    intptr0 += 4;
-                    intptr1 += 4;
-                    ptr += 8;
-                }
-            }
-        }
-    }
-    if (out_elempack == 1)
-    {
-        if (bias_data_size == 0)
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < channels; q++)
-            {
-                const int* intptr = bottom_blob.channel(q);
-                signed char* ptr0 = top_blob.channel(q * 4);
-                signed char* ptr1 = top_blob.channel(q * 4 + 1);
-                signed char* ptr2 = top_blob.channel(q * 4 + 2);
-                signed char* ptr3 = top_blob.channel(q * 4 + 3);
-                signed char* vp;
-
-                __m128 _scale_in = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 4, 0);
-                __m128 _scale_out = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 4, 0);
-
-                __m128 _scale = __lsx_vfmul_s(_scale_in, _scale_out);
-
-                int i = 0;
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr + 16);
-                    __m128 _v = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                    _v = __lsx_vfmul_s(_v, _scale);
-                    __m128i v = float2int8relu(_v);
-                    vp = (signed char*)&v;
-                    ptr0[0] = vp[0];
-                    ptr1[0] = vp[1];
-                    ptr2[0] = vp[2];
-                    ptr3[0] = vp[3];
-
-                    intptr += 4;
-                    ptr0 += 1;
-                    ptr1 += 1;
-                    ptr2 += 1;
-                    ptr3 += 1;
-                }
-            }
-        }
-        else
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < channels; q++)
-            {
-                const int* intptr = bottom_blob.channel(q);
-                signed char* ptr0 = top_blob.channel(q * 4);
-                signed char* ptr1 = top_blob.channel(q * 4 + 1);
-                signed char* ptr2 = top_blob.channel(q * 4 + 2);
-                signed char* ptr3 = top_blob.channel(q * 4 + 3);
-                signed char* vp;
-
-                __m128 _scale_in = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 4, 0);
-                __m128 _scale_out = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 4, 0);
-                __m128 _bias = bias_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(bias_data[0]) : (__m128)__lsx_vld((const float*)bias_data + q * 4, 0);
-
-                __m128 _scale = __lsx_vfmul_s(_scale_in, _scale_out);
-                _bias = __lsx_vfmul_s(_bias, _scale_out);
-
-                int i = 0;
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr + 16);
-                    __m128 _v = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                    _v = __lsx_vfmadd_s(_scale, _v, _bias);
-                    __m128i v = float2int8relu(_v);
-                    vp = (signed char*)&v;
-                    ptr0[0] = vp[0];
-                    ptr1[0] = vp[1];
-                    ptr2[0] = vp[2];
-                    ptr3[0] = vp[3];
-
-                    intptr += 4;
-                    ptr0 += 1;
-                    ptr1 += 1;
-                    ptr2 += 1;
-                    ptr3 += 1;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_relu_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_relu_pack8.h
deleted file mode 100644
index 3d2a45b4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/requantize_relu_pack8.h
+++ /dev/null
@@ -1,186 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_relu_pack8_lsx(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-    int size = w * h;
-
-    int scale_in_data_size = scale_in_data.w;
-    int scale_out_data_size = scale_out_data.w;
-    int bias_data_size = bias_data.w;
-
-    // int8(relu(v * scale_in) * scale_out)
-    // int8_relu(v * (scale_in * scale_out))
-
-    // int8(relu(v * scale_in + bias) * scale_out)
-    // int8_relu(v * (scale_in * scale_out) + (bias * scale_out))
-
-    if (bias_data_size == 0)
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int q = 0; q < channels; q++)
-        {
-            const int* intptr = bottom_blob.channel(q);
-            signed char* ptr = top_blob.channel(q);
-
-            __m128 _scale_in0 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8, 0);
-            __m128 _scale_in1 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8 + 4, 0);
-            __m128 _scale_out0 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8, 0);
-            __m128 _scale_out1 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8 + 4, 0);
-
-            __m128 _scale0 = __lsx_vfmul_s(_scale_in0, _scale_out0);
-            __m128 _scale1 = __lsx_vfmul_s(_scale_in1, _scale_out1);
-
-            int i = 0;
-            for (; i + 3 < size; i += 4)
-            {
-                __builtin_prefetch(intptr + 128);
-                __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr + 4, 0));
-                __m128 _v2 = __lsx_vffint_s_w(__lsx_vld(intptr + 8, 0));
-                __m128 _v3 = __lsx_vffint_s_w(__lsx_vld(intptr + 12, 0));
-                __m128 _v4 = __lsx_vffint_s_w(__lsx_vld(intptr + 16, 0));
-                __m128 _v5 = __lsx_vffint_s_w(__lsx_vld(intptr + 20, 0));
-                __m128 _v6 = __lsx_vffint_s_w(__lsx_vld(intptr + 24, 0));
-                __m128 _v7 = __lsx_vffint_s_w(__lsx_vld(intptr + 28, 0));
-                _v0 = __lsx_vfmul_s(_v0, _scale0);
-                _v1 = __lsx_vfmul_s(_v1, _scale1);
-                _v2 = __lsx_vfmul_s(_v2, _scale0);
-                _v3 = __lsx_vfmul_s(_v3, _scale1);
-                _v4 = __lsx_vfmul_s(_v4, _scale0);
-                _v5 = __lsx_vfmul_s(_v5, _scale1);
-                _v6 = __lsx_vfmul_s(_v6, _scale0);
-                _v7 = __lsx_vfmul_s(_v7, _scale1);
-                *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-                *((int64_t*)(ptr + 8)) = float2int8relu(_v2, _v3);
-                *((int64_t*)(ptr + 16)) = float2int8relu(_v4, _v5);
-                *((int64_t*)(ptr + 24)) = float2int8relu(_v6, _v7);
-
-                intptr += 32;
-                ptr += 32;
-            }
-            for (; i + 1 < size; i += 2)
-            {
-                __builtin_prefetch(intptr + 64);
-                __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr + 4, 0));
-                __m128 _v2 = __lsx_vffint_s_w(__lsx_vld(intptr + 8, 0));
-                __m128 _v3 = __lsx_vffint_s_w(__lsx_vld(intptr + 12, 0));
-                _v0 = __lsx_vfmul_s(_v0, _scale0);
-                _v1 = __lsx_vfmul_s(_v1, _scale1);
-                _v2 = __lsx_vfmul_s(_v2, _scale0);
-                _v3 = __lsx_vfmul_s(_v3, _scale1);
-                *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-                *((int64_t*)(ptr + 8)) = float2int8relu(_v2, _v3);
-
-                intptr += 16;
-                ptr += 16;
-            }
-            for (; i < size; i++)
-            {
-                __builtin_prefetch(intptr + 32);
-                __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr + 4, 0));
-                _v0 = __lsx_vfmul_s(_v0, _scale0);
-                _v1 = __lsx_vfmul_s(_v1, _scale1);
-                *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-
-                intptr += 8;
-                ptr += 8;
-            }
-        }
-    }
-    else
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int q = 0; q < channels; q++)
-        {
-            const int* intptr = bottom_blob.channel(q);
-            signed char* ptr = top_blob.channel(q);
-
-            __m128 _scale_in0 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8, 0);
-            __m128 _scale_in1 = scale_in_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_in_data[0]) : (__m128)__lsx_vld((const float*)scale_in_data + q * 8 + 4, 0);
-            __m128 _scale_out0 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8, 0);
-            __m128 _scale_out1 = scale_out_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(scale_out_data[0]) : (__m128)__lsx_vld((const float*)scale_out_data + q * 8 + 4, 0);
-            __m128 _bias0 = bias_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(bias_data[0]) : (__m128)__lsx_vld((const float*)bias_data + q * 8, 0);
-            __m128 _bias1 = bias_data_size == 1 ? (__m128)__lsx_vreplfr2vr_s(bias_data[0]) : (__m128)__lsx_vld((const float*)bias_data + q * 8 + 4, 0);
-
-            __m128 _scale0 = __lsx_vfmul_s(_scale_in0, _scale_out0);
-            __m128 _scale1 = __lsx_vfmul_s(_scale_in1, _scale_out1);
-            _bias0 = __lsx_vfmul_s(_bias0, _scale_out0);
-            _bias1 = __lsx_vfmul_s(_bias1, _scale_out1);
-
-            int i = 0;
-            for (; i + 3 < size; i += 4)
-            {
-                __builtin_prefetch(intptr + 128);
-                __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr + 4, 0));
-                __m128 _v2 = __lsx_vffint_s_w(__lsx_vld(intptr + 8, 0));
-                __m128 _v3 = __lsx_vffint_s_w(__lsx_vld(intptr + 12, 0));
-                __m128 _v4 = __lsx_vffint_s_w(__lsx_vld(intptr + 16, 0));
-                __m128 _v5 = __lsx_vffint_s_w(__lsx_vld(intptr + 20, 0));
-                __m128 _v6 = __lsx_vffint_s_w(__lsx_vld(intptr + 24, 0));
-                __m128 _v7 = __lsx_vffint_s_w(__lsx_vld(intptr + 28, 0));
-                _v0 = __lsx_vfmadd_s(_scale0, _v0, _bias0);
-                _v1 = __lsx_vfmadd_s(_scale1, _v1, _bias1);
-                _v2 = __lsx_vfmadd_s(_scale0, _v2, _bias0);
-                _v3 = __lsx_vfmadd_s(_scale1, _v3, _bias1);
-                _v4 = __lsx_vfmadd_s(_scale0, _v4, _bias0);
-                _v5 = __lsx_vfmadd_s(_scale1, _v5, _bias1);
-                _v6 = __lsx_vfmadd_s(_scale0, _v6, _bias0);
-                _v7 = __lsx_vfmadd_s(_scale1, _v7, _bias1);
-                *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-                *((int64_t*)(ptr + 8)) = float2int8relu(_v2, _v3);
-                *((int64_t*)(ptr + 16)) = float2int8relu(_v4, _v5);
-                *((int64_t*)(ptr + 24)) = float2int8relu(_v6, _v7);
-
-                intptr += 32;
-                ptr += 32;
-            }
-            for (; i + 1 < size; i += 2)
-            {
-                __builtin_prefetch(intptr + 64);
-                __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr + 4, 0));
-                __m128 _v2 = __lsx_vffint_s_w(__lsx_vld(intptr + 8, 0));
-                __m128 _v3 = __lsx_vffint_s_w(__lsx_vld(intptr + 12, 0));
-                _v0 = __lsx_vfmadd_s(_scale0, _v0, _bias0);
-                _v1 = __lsx_vfmadd_s(_scale1, _v1, _bias1);
-                _v2 = __lsx_vfmadd_s(_scale0, _v2, _bias0);
-                _v3 = __lsx_vfmadd_s(_scale1, _v3, _bias1);
-                *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-                *((int64_t*)(ptr + 8)) = float2int8relu(_v2, _v3);
-
-                intptr += 16;
-                ptr += 16;
-            }
-            for (; i < size; i++)
-            {
-                __builtin_prefetch(intptr + 32);
-                __m128 _v0 = __lsx_vffint_s_w(__lsx_vld(intptr, 0));
-                __m128 _v1 = __lsx_vffint_s_w(__lsx_vld(intptr + 4, 0));
-                _v0 = __lsx_vfmadd_s(_scale0, _v0, _bias0);
-                _v1 = __lsx_vfmadd_s(_scale1, _v1, _bias1);
-                *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-
-                intptr += 8;
-                ptr += 8;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/sigmoid_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/sigmoid_loongarch.h
deleted file mode 100644
index 02354d2a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/sigmoid_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SIGMOID_LOONGARCH_H
-#define LAYER_SIGMOID_LOONGARCH_H
-
-#include "sigmoid.h"
-
-namespace ncnn {
-
-class Sigmoid_loongarch : public Sigmoid
-{
-public:
-    Sigmoid_loongarch();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SIGMOID_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/slice_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/slice_loongarch.h
deleted file mode 100644
index 2f5faed8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/slice_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SLICE_LOONGARCH_H
-#define LAYER_SLICE_LOONGARCH_H
-
-#include "slice.h"
-
-namespace ncnn {
-
-class Slice_loongarch : public Slice
-{
-public:
-    Slice_loongarch();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SLICE_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/softmax_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/softmax_loongarch.h
deleted file mode 100644
index baf930fc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/softmax_loongarch.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SOFTMAX_LOONGARCH_H
-#define LAYER_SOFTMAX_LOONGARCH_H
-
-#include "softmax.h"
-
-namespace ncnn {
-
-class Softmax_loongarch : public Softmax
-{
-public:
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SOFTMAX_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/swish_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/swish_loongarch.h
deleted file mode 100644
index 9b7d2ac8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/swish_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SWISH_LOONGARCH_H
-#define LAYER_SWISH_LOONGARCH_H
-
-#include "swish.h"
-
-namespace ncnn {
-
-class Swish_loongarch : public Swish
-{
-public:
-    Swish_loongarch();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SWISH_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/tanh_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/tanh_loongarch.h
deleted file mode 100644
index 74231eb5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/tanh_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_TANH_LOONGARCH_H
-#define LAYER_TANH_LOONGARCH_H
-
-#include "tanh.h"
-
-namespace ncnn {
-
-class TanH_loongarch : public TanH
-{
-public:
-    TanH_loongarch();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_TANH_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/unaryop_loongarch.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/unaryop_loongarch.h
deleted file mode 100644
index f4210aea..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/loongarch/unaryop_loongarch.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// yala is pleased to support the open source community by making ncnn available.
-//
-//
-// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_UNARYOP_LOONGARCH_H
-#define LAYER_UNARYOP_LOONGARCH_H
-
-#include "unaryop.h"
-
-namespace ncnn {
-
-class UnaryOp_loongarch : public UnaryOp
-{
-public:
-    UnaryOp_loongarch();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_UNARYOP_LOONGARCH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/lrn.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/lrn.h
deleted file mode 100644
index 07c5f96f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/lrn.h
+++ /dev/null
@@ -1,48 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_LRN_H
-#define LAYER_LRN_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class LRN : public Layer
-{
-public:
-    LRN();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-    enum NormRegionType
-    {
-        NormRegion_ACROSS_CHANNELS = 0,
-        NormRegion_WITHIN_CHANNEL = 1
-    };
-
-public:
-    // param
-    int region_type;
-    int local_size;
-    float alpha;
-    float beta;
-    float bias;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_LRN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/lstm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/lstm.h
deleted file mode 100644
index 58bd67f9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/lstm.h
+++ /dev/null
@@ -1,49 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_LSTM_H
-#define LAYER_LSTM_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class LSTM : public Layer
-{
-public:
-    LSTM();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    int num_output;
-    int weight_data_size;
-    int direction; // 0=forward 1=reverse 2=bidirectional
-    int hidden_size;
-
-    Mat weight_hc_data;
-    Mat weight_xc_data;
-    Mat bias_c_data;
-    Mat weight_hr_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_LSTM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/matmul.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/matmul.h
deleted file mode 100644
index 15367096..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/matmul.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MATMUL_H
-#define LAYER_MATMUL_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class MatMul : public Layer
-{
-public:
-    MatMul();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    int transB;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MATMUL_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/memorydata.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/memorydata.h
deleted file mode 100644
index d5175ad0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/memorydata.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MEMORYDATA_H
-#define LAYER_MEMORYDATA_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class MemoryData : public Layer
-{
-public:
-    MemoryData();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    int w;
-    int h;
-    int d;
-    int c;
-    int load_type;
-
-    Mat data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MEMORYDATA_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/absval_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/absval_mips.h
deleted file mode 100644
index 95dca4d5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/absval_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Leo is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 Leo <leo@nullptr.com.cn>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ABSVAL_MIPS_H
-#define LAYER_ABSVAL_MIPS_H
-
-#include "absval.h"
-
-namespace ncnn {
-
-class AbsVal_mips : public AbsVal
-{
-public:
-    AbsVal_mips();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ABSVAL_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/batchnorm_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/batchnorm_mips.h
deleted file mode 100644
index 6df49407..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/batchnorm_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BATCHNORM_MIPS_H
-#define LAYER_BATCHNORM_MIPS_H
-
-#include "batchnorm.h"
-
-namespace ncnn {
-
-class BatchNorm_mips : public BatchNorm
-{
-public:
-    BatchNorm_mips();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BATCHNORM_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/bias_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/bias_mips.h
deleted file mode 100644
index dfef2159..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/bias_mips.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// Leo is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 Leo <leo@nullptr.com.cn>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BIAS_MIPS_H
-#define LAYER_BIAS_MIPS_H
-
-#include "bias.h"
-
-namespace ncnn {
-
-class Bias_mips : public Bias
-{
-public:
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BIAS_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/binaryop_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/binaryop_mips.h
deleted file mode 100644
index e682373b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/binaryop_mips.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BINARYOP_MIPS_H
-#define LAYER_BINARYOP_MIPS_H
-
-#include "binaryop.h"
-
-namespace ncnn {
-
-class BinaryOp_mips : public BinaryOp
-{
-public:
-    BinaryOp_mips();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BINARYOP_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/cast_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/cast_mips.h
deleted file mode 100644
index adabee5f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/cast_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CAST_MIPS_H
-#define LAYER_CAST_MIPS_H
-
-#include "cast.h"
-
-namespace ncnn {
-
-class Cast_mips : public Cast
-{
-public:
-    Cast_mips();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CAST_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/clip_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/clip_mips.h
deleted file mode 100644
index 5db94bc5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/clip_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Leo is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 Leo <leo@nullptr.com.cn>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CLIP_MIPS_H
-#define LAYER_CLIP_MIPS_H
-
-#include "clip.h"
-
-namespace ncnn {
-
-class Clip_mips : public Clip
-{
-public:
-    Clip_mips();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CLIP_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/concat_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/concat_mips.h
deleted file mode 100644
index c4ab84f3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/concat_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONCAT_MIPS_H
-#define LAYER_CONCAT_MIPS_H
-
-#include "concat.h"
-
-namespace ncnn {
-
-class Concat_mips : public Concat
-{
-public:
-    Concat_mips();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONCAT_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution1d_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution1d_mips.h
deleted file mode 100644
index dcc9bd4d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution1d_mips.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTION1D_MIPS_H
-#define LAYER_CONVOLUTION1D_MIPS_H
-
-#include "convolution1d.h"
-
-namespace ncnn {
-
-class Convolution1D_mips : public Convolution1D
-{
-public:
-    Convolution1D_mips();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    // packn
-    Mat weight_data_packed;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTION1D_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1.h
deleted file mode 100644
index cb36e999..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1.h
+++ /dev/null
@@ -1,26 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_msa(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_int8.h
deleted file mode 100644
index 8730041f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_int8.h
+++ /dev/null
@@ -1,83 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_int8_msa(bottom_im2col, top_blob, kernel, opt);
-}
-
-static void conv1x1s2_sgemm_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = w - 2 * outw + w;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const signed char* r0 = bottom_blob.channel(p);
-        signed char* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                outptr[0] = r0[0];
-                outptr[1] = r0[2];
-                outptr[2] = r0[4];
-                outptr[3] = r0[6];
-
-                r0 += 8;
-                outptr += 4;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                outptr[0] = r0[0];
-                outptr[1] = r0[2];
-
-                r0 += 4;
-                outptr += 2;
-            }
-            for (; j < outw; j++)
-            {
-                outptr[0] = r0[0];
-
-                r0 += 2;
-                outptr += 1;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_int8_msa(bottom_blob_shrinked, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack1to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack1to4_int8.h
deleted file mode 100644
index 928e1633..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack1to4_int8.h
+++ /dev/null
@@ -1,83 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_pack1to4_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_pack1to4_int8_msa(bottom_im2col, top_blob, kernel, opt);
-}
-
-static void conv1x1s2_sgemm_pack1to4_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = w - 2 * outw + w;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const signed char* r0 = bottom_blob.channel(p);
-        signed char* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                outptr[0] = r0[0];
-                outptr[1] = r0[2];
-                outptr[2] = r0[4];
-                outptr[3] = r0[6];
-
-                r0 += 8;
-                outptr += 4;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                outptr[0] = r0[0];
-                outptr[1] = r0[2];
-
-                r0 += 4;
-                outptr += 2;
-            }
-            for (; j < outw; j++)
-            {
-                outptr[0] = r0[0];
-
-                r0 += 2;
-                outptr += 1;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_pack1to4_int8_msa(bottom_blob_shrinked, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack4.h
deleted file mode 100644
index 0116c9be..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack4.h
+++ /dev/null
@@ -1,65 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_pack4_msa(bottom_im2col, top_blob, kernel, _bias, opt);
-}
-
-static void conv1x1s2_sgemm_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const float* r0 = bottom_blob.channel(p);
-        float* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                v4f32 _val = (v4f32)__msa_ld_w(r0, 0);
-                __msa_st_w((v4i32)_val, outptr, 0);
-
-                r0 += 4 * 2;
-                outptr += 4;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_pack4_msa(bottom_blob_shrinked, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack4to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack4to1.h
deleted file mode 100644
index 72125f08..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack4to1.h
+++ /dev/null
@@ -1,65 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_pack4to1_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_pack4to1_msa(bottom_im2col, top_blob, kernel, _bias, opt);
-}
-
-static void conv1x1s2_sgemm_pack4to1_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const float* r0 = bottom_blob.channel(p);
-        float* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                v4f32 _val = (v4f32)__msa_ld_w(r0, 0);
-                __msa_st_w((v4i32)_val, outptr, 0);
-
-                r0 += 4 * 2;
-                outptr += 4;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_pack4to1_msa(bottom_blob_shrinked, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack8to1_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack8to1_int8.h
deleted file mode 100644
index 398f85e2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack8to1_int8.h
+++ /dev/null
@@ -1,65 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_pack8to1_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_pack8to1_int8_msa(bottom_im2col, top_blob, kernel, opt);
-}
-
-static void conv1x1s2_sgemm_pack8to1_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = w - 2 * outw + w;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const int64_t* r0 = bottom_blob.channel(p);
-        int64_t* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                outptr[0] = r0[0];
-
-                r0 += 2;
-                outptr += 1;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_pack8to1_int8_msa(bottom_blob_shrinked, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack8to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack8to4_int8.h
deleted file mode 100644
index aa38542e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_1x1_pack8to4_int8.h
+++ /dev/null
@@ -1,65 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_pack8to4_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_pack8to4_int8_msa(bottom_im2col, top_blob, kernel, opt);
-}
-
-static void conv1x1s2_sgemm_pack8to4_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = w - 2 * outw + w;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const int64_t* r0 = bottom_blob.channel(p);
-        int64_t* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                outptr[0] = r0[0];
-
-                r0 += 2;
-                outptr += 1;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_pack8to4_int8_msa(bottom_blob_shrinked, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3.h
deleted file mode 100644
index 4b21d20f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3.h
+++ /dev/null
@@ -1,412 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd23_transform_kernel_msa(const Mat& kernel, Mat& kernel_tm2, int inch, int outch, const Option& opt)
-{
-    Mat kernel_tm(4 * 4, inch, outch);
-
-    // G
-    const float ktm[4][3] = {
-        {1.0f, 0.0f, 0.0f},
-        {1.0f / 2, 1.0f / 2, 1.0f / 2},
-        {1.0f / 2, -1.0f / 2, 1.0f / 2},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[4][3];
-            for (int i = 0; i < 4; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 4; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 4; i++)
-                {
-                    kernel_tm0[j * 4 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 16-inch-outch
-    // dst = inch-16-outch
-#if __mips_msa
-    kernel_tm2.create(8 * inch, 16, outch / 8 + (outch % 8) / 4 + outch % 4);
-#else
-    kernel_tm2.create(2 * inch, 16, outch / 2 + outch % 2);
-#endif
-
-    int q = 0;
-#if __mips_msa
-    for (; q + 7 < outch; q += 8)
-    {
-        Mat g0 = kernel_tm2.channel(q / 8);
-
-        for (int k = 0; k < 16; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 8; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-    for (; q + 3 < outch; q += 4)
-    {
-        Mat g0 = kernel_tm2.channel(q / 8 + (q % 8) / 4);
-
-        for (int k = 0; k < 16; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#else  // __mips_msa
-    for (; q + 1 < outch; q += 2)
-    {
-        Mat g0 = kernel_tm2.channel(q / 2);
-
-        for (int k = 0; k < 16; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 2; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#endif // __mips_msa
-    for (; q < outch; q++)
-    {
-#if __mips_msa
-        Mat g0 = kernel_tm2.channel(q / 8 + (q % 8) / 4 + q % 4);
-#else
-        Mat g0 = kernel_tm2.channel(q / 2 + q % 2);
-#endif
-
-        for (int k = 0; k < 16; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                const float* k00 = kernel_tm.channel(q).row(p);
-                g00[0] = k00[k];
-                g00++;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 2n+2, winograd F(2,3)
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 1) / 2 * 2;
-    outh = (outh + 1) / 2 * 2;
-
-    w = outw + 2;
-    h = outh + 2;
-    Option opt_b = opt;
-    opt_b.blob_allocator = opt.workspace_allocator;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, 0, 0.f, opt_b);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 2;
-        int h_tiles = outh / 2;
-        int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 16, inch, 4u, opt.workspace_allocator);
-        conv3x3s1_winograd23_transform_input_msa(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_msa(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, 4u, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd23_transform_output_msa(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
-
-static void conv3x3s1_winograd43_transform_kernel_msa(const Mat& kernel, Mat& kernel_tm2, int inch, int outch, const Option& opt)
-{
-    Mat kernel_tm(6 * 6, inch, outch);
-
-    // G
-    const float ktm[6][3] = {
-        {1.0f / 4, 0.0f, 0.0f},
-        {-1.0f / 6, -1.0f / 6, -1.0f / 6},
-        {-1.0f / 6, 1.0f / 6, -1.0f / 6},
-        {1.0f / 24, 1.0f / 12, 1.0f / 6},
-        {1.0f / 24, -1.0f / 12, 1.0f / 6},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[6][3];
-            for (int i = 0; i < 6; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 6; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 6; i++)
-                {
-                    kernel_tm0[j * 6 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 36-inch-outch
-    // dst = inch-36-outch
-#if __mips_msa
-    kernel_tm2.create(8 * inch, 36, outch / 8 + (outch % 8) / 4 + outch % 4);
-#else
-    kernel_tm2.create(2 * inch, 36, outch / 2 + outch % 2);
-#endif
-
-    int q = 0;
-#if __mips_msa
-    for (; q + 7 < outch; q += 8)
-    {
-        Mat g0 = kernel_tm2.channel(q / 8);
-
-        for (int k = 0; k < 36; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 8; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-    for (; q + 3 < outch; q += 4)
-    {
-        Mat g0 = kernel_tm2.channel(q / 8 + (q % 8) / 4);
-
-        for (int k = 0; k < 36; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#else  // __mips_msa
-    for (; q + 1 < outch; q += 2)
-    {
-        Mat g0 = kernel_tm2.channel(q / 2);
-
-        for (int k = 0; k < 36; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 2; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#endif // __mips_msa
-    for (; q < outch; q++)
-    {
-#if __mips_msa
-        Mat g0 = kernel_tm2.channel(q / 8 + (q % 8) / 4 + q % 4);
-#else
-        Mat g0 = kernel_tm2.channel(q / 2 + q % 2);
-#endif
-
-        for (int k = 0; k < 36; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                const float* k00 = kernel_tm.channel(q).row(p);
-                g00[0] = k00[k];
-                g00++;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 4n+2, winograd F(4,3)
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 3) / 4 * 4;
-    outh = (outh + 3) / 4 * 4;
-
-    w = outw + 2;
-    h = outh + 2;
-
-    Option opt_b = opt;
-    opt_b.blob_allocator = opt.workspace_allocator;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, 0, 0.f, opt_b);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 4;
-        int h_tiles = outh / 4;
-        int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 36, inch, 4u, opt.workspace_allocator);
-        conv3x3s1_winograd43_transform_input_msa(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_msa(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, 4u, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd43_transform_output_msa(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_int8.h
deleted file mode 100644
index 3958b12b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_int8.h
+++ /dev/null
@@ -1,285 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if NCNN_RUNTIME_CPU && NCNN_MMI && !__mips_msa && !__mips_loongson_mmi
-void conv3x3s1_winograd43_transform_kernel_int8_loongson_mmi(const Mat& kernel, Mat& kernel_tm_packed, int inch, int outch, const Option& opt);
-#endif
-
-static void conv3x3s1_winograd43_transform_kernel_int8_msa(const Mat& kernel, Mat& kernel_tm_packed, int inch, int outch, const Option& opt)
-{
-#if NCNN_RUNTIME_CPU && NCNN_MMI && !__mips_msa && !__mips_loongson_mmi
-    if (ncnn::cpu_support_loongson_mmi())
-    {
-        conv3x3s1_winograd43_transform_kernel_int8_loongson_mmi(kernel, kernel_tm_packed, inch, outch, opt);
-        return;
-    }
-#endif
-
-    // winograd43 transform kernel
-    Mat kernel_tm(6 * 6, inch, outch, (size_t)2u);
-
-    const short ktm[6][3] = {
-        {6, 0, 0},
-        {-4, -4, -4},
-        {-4, 4, -4},
-        {1, 2, 4},
-        {1, -2, 4},
-        {0, 0, 6}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const signed char* kernel0 = (const signed char*)kernel + p * inch * 9 + q * 9;
-            short* kernel_tm0 = kernel_tm.channel(p).row<short>(q);
-
-            // transform kernel
-            const signed char* k0 = kernel0;
-            const signed char* k1 = kernel0 + 3;
-            const signed char* k2 = kernel0 + 6;
-
-            // h
-            short tmp[6][3];
-            for (int i = 0; i < 6; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 6; j++)
-            {
-                short* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 6; i++)
-                {
-                    kernel_tm0[j * 6 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 36-inch-outch
-    // dst = 2b-inch-36-outch/2b
-#if __mips_msa
-    if (outch >= 4)
-    {
-        if (inch >= 4)
-            kernel_tm_packed.create(inch / 4 + inch % 4, 36, outch / 4 + outch % 4, (size_t)2u * 16, 16);
-        else
-            kernel_tm_packed.create(inch, 36, outch / 4 + outch % 4, (size_t)2u * 4, 4);
-    }
-#else // __mips_msa
-    if (outch >= 2)
-    {
-#if __mips_loongson_mmi
-        if (inch >= 4)
-            kernel_tm_packed.create(inch / 4 + inch % 4, 36, outch / 2 + outch % 2, (size_t)2u * 8, 8);
-        else
-#endif // __mips_loongson_mmi
-        {
-            kernel_tm_packed.create(inch, 36, outch / 2 + outch % 2, (size_t)2u * 2, 2);
-        }
-    }
-#endif // __mips_msa
-    else
-    {
-#if __mips_msa || __mips_loongson_mmi
-        if (inch >= 4)
-            kernel_tm_packed.create(inch / 4 + inch % 4, 36, outch, (size_t)2u * 4, 4);
-        else
-#endif // __mips_msa || __mips_loongson_mmi
-        {
-            kernel_tm_packed.create(inch, 36, outch, (size_t)2u, 1);
-        }
-    }
-
-    int p = 0;
-#if __mips_msa
-    for (; p + 3 < outch; p += 4)
-    {
-        const Mat k0 = kernel_tm.channel(p);
-        const Mat k1 = kernel_tm.channel(p + 1);
-        const Mat k2 = kernel_tm.channel(p + 2);
-        const Mat k3 = kernel_tm.channel(p + 3);
-
-        Mat g0 = kernel_tm_packed.channel(p / 4);
-
-        for (int k = 0; k < 36; k++)
-        {
-            short* g00 = g0.row<short>(k);
-
-            int q = 0;
-            for (; q + 3 < inch; q += 4)
-            {
-                g00[0] = k0.row<const short>(q)[k];
-                g00[1] = k0.row<const short>(q + 1)[k];
-                g00[2] = k0.row<const short>(q + 2)[k];
-                g00[3] = k0.row<const short>(q + 3)[k];
-                g00[4] = k1.row<const short>(q)[k];
-                g00[5] = k1.row<const short>(q + 1)[k];
-                g00[6] = k1.row<const short>(q + 2)[k];
-                g00[7] = k1.row<const short>(q + 3)[k];
-                g00[8] = k2.row<const short>(q)[k];
-                g00[9] = k2.row<const short>(q + 1)[k];
-                g00[10] = k2.row<const short>(q + 2)[k];
-                g00[11] = k2.row<const short>(q + 3)[k];
-                g00[12] = k3.row<const short>(q)[k];
-                g00[13] = k3.row<const short>(q + 1)[k];
-                g00[14] = k3.row<const short>(q + 2)[k];
-                g00[15] = k3.row<const short>(q + 3)[k];
-                g00 += 16;
-            }
-            for (; q < inch; q++)
-            {
-                g00[0] = k0.row<const short>(q)[k];
-                g00[1] = k1.row<const short>(q)[k];
-                g00[2] = k2.row<const short>(q)[k];
-                g00[3] = k3.row<const short>(q)[k];
-                g00 += 4;
-            }
-        }
-    }
-#else // __mips_msa
-    for (; p + 1 < outch; p += 2)
-    {
-        const Mat k0 = kernel_tm.channel(p);
-        const Mat k1 = kernel_tm.channel(p + 1);
-
-        Mat g0 = kernel_tm_packed.channel(p / 2);
-
-        for (int k = 0; k < 36; k++)
-        {
-            short* g00 = g0.row<short>(k);
-
-            int q = 0;
-#if __mips_loongson_mmi
-            for (; q + 3 < inch; q += 4)
-            {
-                g00[0] = k0.row<const short>(q)[k];
-                g00[1] = k0.row<const short>(q + 1)[k];
-                g00[2] = k1.row<const short>(q)[k];
-                g00[3] = k1.row<const short>(q + 1)[k];
-                g00[4] = k0.row<const short>(q + 2)[k];
-                g00[5] = k0.row<const short>(q + 3)[k];
-                g00[6] = k1.row<const short>(q + 2)[k];
-                g00[7] = k1.row<const short>(q + 3)[k];
-                g00 += 8;
-            }
-#endif // __mips_loongson_mmi
-            for (; q < inch; q++)
-            {
-                g00[0] = k0.row<const short>(q)[k];
-                g00[1] = k1.row<const short>(q)[k];
-                g00 += 2;
-            }
-        }
-    }
-#endif // __mips_msa
-    for (; p < outch; p++)
-    {
-        const Mat k0 = kernel_tm.channel(p);
-
-#if __mips_msa
-        Mat g0 = kernel_tm_packed.channel(p / 4 + p % 4);
-#else
-        Mat g0 = kernel_tm_packed.channel(p / 2 + p % 2);
-#endif
-
-        for (int k = 0; k < 36; k++)
-        {
-            short* g00 = g0.row<short>(k);
-
-            int q = 0;
-#if __mips_msa || __mips_loongson_mmi
-            for (; q + 3 < inch; q += 4)
-            {
-                g00[0] = k0.row<const short>(q)[k];
-                g00[1] = k0.row<const short>(q + 1)[k];
-                g00[2] = k0.row<const short>(q + 2)[k];
-                g00[3] = k0.row<const short>(q + 3)[k];
-                g00 += 4;
-            }
-#endif // __mips_msa || __mips_loongson_mmi
-            for (; q < inch; q++)
-            {
-                g00[0] = k0.row<const short>(q)[k];
-                g00 += 1;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    //     size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 4n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 3) / 4 * 4;
-    outh = (outh + 3) / 4 * 4;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 4;
-        int h_tiles = outh / 4;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 36, inch, 2u * elempack, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd43_transform_input_int8_msa(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_int8_msa(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, 4u, 1, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd43_transform_output_int8_msa(top_blob_tm, top_blob_bordered, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack1to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack1to4.h
deleted file mode 100644
index 409afeaa..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack1to4.h
+++ /dev/null
@@ -1,812 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack1to4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        v4f32 _bias0 = bias ? (v4f32)__msa_ld_w(bias + p * 4, 0) : (v4f32)__msa_fill_w(0);
-        out0.fill(_bias0);
-
-        const float* k0 = kernel.channel(p);
-
-        int q = 0;
-        for (; q < inch; q++)
-        {
-            float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            v4f32 _k00 = (v4f32)__msa_ld_w(k0, 0);
-            v4f32 _k01 = (v4f32)__msa_ld_w(k0 + 4, 0);
-            v4f32 _k02 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-            v4f32 _k10 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-            v4f32 _k11 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-            v4f32 _k12 = (v4f32)__msa_ld_w(k0 + 4 * 5, 0);
-            v4f32 _k20 = (v4f32)__msa_ld_w(k0 + 4 * 6, 0);
-            v4f32 _k21 = (v4f32)__msa_ld_w(k0 + 4 * 7, 0);
-            v4f32 _k22 = (v4f32)__msa_ld_w(k0 + 4 * 8, 0);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    v4f32 _sum0 = (v4f32)__msa_ld_w(outptr0, 0);
-                    v4f32 _sum1 = (v4f32)__msa_ld_w(outptr0 + 4, 0);
-                    v4f32 _sum2 = (v4f32)__msa_ld_w(outptr0 + 4 * 2, 0);
-                    v4f32 _sum3 = (v4f32)__msa_ld_w(outptr0 + 4 * 3, 0);
-                    v4f32 _sum4 = (v4f32)__msa_ld_w(outptr0 + 4 * 4, 0);
-                    v4f32 _sum5 = (v4f32)__msa_ld_w(outptr0 + 4 * 5, 0);
-                    v4f32 _sum6 = (v4f32)__msa_ld_w(outptr0 + 4 * 6, 0);
-                    v4f32 _sum7 = (v4f32)__msa_ld_w(outptr0 + 4 * 7, 0);
-
-                    v4i32 _r0 = __msa_ld_w(r0, 0);
-                    v4i32 _r0n = __msa_ld_w(r0 + 4, 0);
-                    v4i32 _r0nn = __msa_ld_w(r0 + 8, 0);
-
-                    v4f32 _r00 = (v4f32)__msa_splati_w(_r0, 0);
-                    v4f32 _r01 = (v4f32)__msa_splati_w(_r0, 1);
-                    v4f32 _r02 = (v4f32)__msa_splati_w(_r0, 2);
-                    v4f32 _r03 = (v4f32)__msa_splati_w(_r0, 3);
-                    v4f32 _r04 = (v4f32)__msa_splati_w(_r0n, 0);
-                    v4f32 _r05 = (v4f32)__msa_splati_w(_r0n, 1);
-                    v4f32 _r06 = (v4f32)__msa_splati_w(_r0n, 2);
-                    v4f32 _r07 = (v4f32)__msa_splati_w(_r0n, 3);
-                    v4f32 _r08 = (v4f32)__msa_splati_w(_r0nn, 0);
-                    v4f32 _r09 = (v4f32)__msa_splati_w(_r0nn, 1);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r00, _k00);
-                    _sum1 = __msa_fmadd_w(_sum1, _r01, _k00);
-                    _sum2 = __msa_fmadd_w(_sum2, _r02, _k00);
-                    _sum3 = __msa_fmadd_w(_sum3, _r03, _k00);
-                    _sum4 = __msa_fmadd_w(_sum4, _r04, _k00);
-                    _sum5 = __msa_fmadd_w(_sum5, _r05, _k00);
-                    _sum6 = __msa_fmadd_w(_sum6, _r06, _k00);
-                    _sum7 = __msa_fmadd_w(_sum7, _r07, _k00);
-                    _sum0 = __msa_fmadd_w(_sum0, _r01, _k01);
-                    _sum1 = __msa_fmadd_w(_sum1, _r02, _k01);
-                    _sum2 = __msa_fmadd_w(_sum2, _r03, _k01);
-                    _sum3 = __msa_fmadd_w(_sum3, _r04, _k01);
-                    _sum4 = __msa_fmadd_w(_sum4, _r05, _k01);
-                    _sum5 = __msa_fmadd_w(_sum5, _r06, _k01);
-                    _sum6 = __msa_fmadd_w(_sum6, _r07, _k01);
-                    _sum7 = __msa_fmadd_w(_sum7, _r08, _k01);
-                    _sum0 = __msa_fmadd_w(_sum0, _r02, _k02);
-                    _sum1 = __msa_fmadd_w(_sum1, _r03, _k02);
-                    _sum2 = __msa_fmadd_w(_sum2, _r04, _k02);
-                    _sum3 = __msa_fmadd_w(_sum3, _r05, _k02);
-                    _sum4 = __msa_fmadd_w(_sum4, _r06, _k02);
-                    _sum5 = __msa_fmadd_w(_sum5, _r07, _k02);
-                    _sum6 = __msa_fmadd_w(_sum6, _r08, _k02);
-                    _sum7 = __msa_fmadd_w(_sum7, _r09, _k02);
-
-                    v4i32 _r1 = __msa_ld_w(r1, 0);
-                    v4i32 _r1n = __msa_ld_w(r1 + 4, 0);
-                    v4i32 _r1nn = __msa_ld_w(r1 + 8, 0);
-
-                    v4f32 _r10 = (v4f32)__msa_splati_w(_r1, 0);
-                    v4f32 _r11 = (v4f32)__msa_splati_w(_r1, 1);
-                    v4f32 _r12 = (v4f32)__msa_splati_w(_r1, 2);
-                    v4f32 _r13 = (v4f32)__msa_splati_w(_r1, 3);
-                    v4f32 _r14 = (v4f32)__msa_splati_w(_r1n, 0);
-                    v4f32 _r15 = (v4f32)__msa_splati_w(_r1n, 1);
-                    v4f32 _r16 = (v4f32)__msa_splati_w(_r1n, 2);
-                    v4f32 _r17 = (v4f32)__msa_splati_w(_r1n, 3);
-                    v4f32 _r18 = (v4f32)__msa_splati_w(_r1nn, 0);
-                    v4f32 _r19 = (v4f32)__msa_splati_w(_r1nn, 1);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r10, _k10);
-                    _sum1 = __msa_fmadd_w(_sum1, _r11, _k10);
-                    _sum2 = __msa_fmadd_w(_sum2, _r12, _k10);
-                    _sum3 = __msa_fmadd_w(_sum3, _r13, _k10);
-                    _sum4 = __msa_fmadd_w(_sum4, _r14, _k10);
-                    _sum5 = __msa_fmadd_w(_sum5, _r15, _k10);
-                    _sum6 = __msa_fmadd_w(_sum6, _r16, _k10);
-                    _sum7 = __msa_fmadd_w(_sum7, _r17, _k10);
-                    _sum0 = __msa_fmadd_w(_sum0, _r11, _k11);
-                    _sum1 = __msa_fmadd_w(_sum1, _r12, _k11);
-                    _sum2 = __msa_fmadd_w(_sum2, _r13, _k11);
-                    _sum3 = __msa_fmadd_w(_sum3, _r14, _k11);
-                    _sum4 = __msa_fmadd_w(_sum4, _r15, _k11);
-                    _sum5 = __msa_fmadd_w(_sum5, _r16, _k11);
-                    _sum6 = __msa_fmadd_w(_sum6, _r17, _k11);
-                    _sum7 = __msa_fmadd_w(_sum7, _r18, _k11);
-                    _sum0 = __msa_fmadd_w(_sum0, _r12, _k12);
-                    _sum1 = __msa_fmadd_w(_sum1, _r13, _k12);
-                    _sum2 = __msa_fmadd_w(_sum2, _r14, _k12);
-                    _sum3 = __msa_fmadd_w(_sum3, _r15, _k12);
-                    _sum4 = __msa_fmadd_w(_sum4, _r16, _k12);
-                    _sum5 = __msa_fmadd_w(_sum5, _r17, _k12);
-                    _sum6 = __msa_fmadd_w(_sum6, _r18, _k12);
-                    _sum7 = __msa_fmadd_w(_sum7, _r19, _k12);
-
-                    v4i32 _r2 = __msa_ld_w(r2, 0);
-                    v4i32 _r2n = __msa_ld_w(r2 + 4, 0);
-                    v4i32 _r2nn = __msa_ld_w(r2 + 8, 0);
-
-                    v4f32 _r20 = (v4f32)__msa_splati_w(_r2, 0);
-                    v4f32 _r21 = (v4f32)__msa_splati_w(_r2, 1);
-                    v4f32 _r22 = (v4f32)__msa_splati_w(_r2, 2);
-                    v4f32 _r23 = (v4f32)__msa_splati_w(_r2, 3);
-                    v4f32 _r24 = (v4f32)__msa_splati_w(_r2n, 0);
-                    v4f32 _r25 = (v4f32)__msa_splati_w(_r2n, 1);
-                    v4f32 _r26 = (v4f32)__msa_splati_w(_r2n, 2);
-                    v4f32 _r27 = (v4f32)__msa_splati_w(_r2n, 3);
-                    v4f32 _r28 = (v4f32)__msa_splati_w(_r2nn, 0);
-                    v4f32 _r29 = (v4f32)__msa_splati_w(_r2nn, 1);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r20, _k20);
-                    _sum1 = __msa_fmadd_w(_sum1, _r21, _k20);
-                    _sum2 = __msa_fmadd_w(_sum2, _r22, _k20);
-                    _sum3 = __msa_fmadd_w(_sum3, _r23, _k20);
-                    _sum4 = __msa_fmadd_w(_sum4, _r24, _k20);
-                    _sum5 = __msa_fmadd_w(_sum5, _r25, _k20);
-                    _sum6 = __msa_fmadd_w(_sum6, _r26, _k20);
-                    _sum7 = __msa_fmadd_w(_sum7, _r27, _k20);
-                    _sum0 = __msa_fmadd_w(_sum0, _r21, _k21);
-                    _sum1 = __msa_fmadd_w(_sum1, _r22, _k21);
-                    _sum2 = __msa_fmadd_w(_sum2, _r23, _k21);
-                    _sum3 = __msa_fmadd_w(_sum3, _r24, _k21);
-                    _sum4 = __msa_fmadd_w(_sum4, _r25, _k21);
-                    _sum5 = __msa_fmadd_w(_sum5, _r26, _k21);
-                    _sum6 = __msa_fmadd_w(_sum6, _r27, _k21);
-                    _sum7 = __msa_fmadd_w(_sum7, _r28, _k21);
-                    _sum0 = __msa_fmadd_w(_sum0, _r22, _k22);
-                    _sum1 = __msa_fmadd_w(_sum1, _r23, _k22);
-                    _sum2 = __msa_fmadd_w(_sum2, _r24, _k22);
-                    _sum3 = __msa_fmadd_w(_sum3, _r25, _k22);
-                    _sum4 = __msa_fmadd_w(_sum4, _r26, _k22);
-                    _sum5 = __msa_fmadd_w(_sum5, _r27, _k22);
-                    _sum6 = __msa_fmadd_w(_sum6, _r28, _k22);
-                    _sum7 = __msa_fmadd_w(_sum7, _r29, _k22);
-
-                    __msa_st_w((v4i32)_sum0, outptr0, 0);
-                    __msa_st_w((v4i32)_sum1, outptr0 + 4, 0);
-                    __msa_st_w((v4i32)_sum2, outptr0 + 4 * 2, 0);
-                    __msa_st_w((v4i32)_sum3, outptr0 + 4 * 3, 0);
-                    __msa_st_w((v4i32)_sum4, outptr0 + 4 * 4, 0);
-                    __msa_st_w((v4i32)_sum5, outptr0 + 4 * 5, 0);
-                    __msa_st_w((v4i32)_sum6, outptr0 + 4 * 6, 0);
-                    __msa_st_w((v4i32)_sum7, outptr0 + 4 * 7, 0);
-
-                    outptr0 += 4 * 8;
-
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                }
-                for (; j + 3 < outw; j += 4)
-                {
-                    v4f32 _sum0 = (v4f32)__msa_ld_w(outptr0, 0);
-                    v4f32 _sum1 = (v4f32)__msa_ld_w(outptr0 + 4, 0);
-                    v4f32 _sum2 = (v4f32)__msa_ld_w(outptr0 + 4 * 2, 0);
-                    v4f32 _sum3 = (v4f32)__msa_ld_w(outptr0 + 4 * 3, 0);
-
-                    v4i32 _r0 = __msa_ld_w(r0, 0);
-                    v4i32 _r0n = __msa_ld_w(r0 + 4, 0);
-
-                    v4f32 _r00 = (v4f32)__msa_splati_w(_r0, 0);
-                    v4f32 _r01 = (v4f32)__msa_splati_w(_r0, 1);
-                    v4f32 _r02 = (v4f32)__msa_splati_w(_r0, 2);
-                    v4f32 _r03 = (v4f32)__msa_splati_w(_r0, 3);
-                    v4f32 _r04 = (v4f32)__msa_splati_w(_r0n, 0);
-                    v4f32 _r05 = (v4f32)__msa_splati_w(_r0n, 1);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r00, _k00);
-                    _sum1 = __msa_fmadd_w(_sum1, _r01, _k00);
-                    _sum2 = __msa_fmadd_w(_sum2, _r02, _k00);
-                    _sum3 = __msa_fmadd_w(_sum3, _r03, _k00);
-                    _sum0 = __msa_fmadd_w(_sum0, _r01, _k01);
-                    _sum1 = __msa_fmadd_w(_sum1, _r02, _k01);
-                    _sum2 = __msa_fmadd_w(_sum2, _r03, _k01);
-                    _sum3 = __msa_fmadd_w(_sum3, _r04, _k01);
-                    _sum0 = __msa_fmadd_w(_sum0, _r02, _k02);
-                    _sum1 = __msa_fmadd_w(_sum1, _r03, _k02);
-                    _sum2 = __msa_fmadd_w(_sum2, _r04, _k02);
-                    _sum3 = __msa_fmadd_w(_sum3, _r05, _k02);
-
-                    v4i32 _r1 = __msa_ld_w(r1, 0);
-                    v4i32 _r1n = __msa_ld_w(r1 + 4, 0);
-
-                    v4f32 _r10 = (v4f32)__msa_splati_w(_r1, 0);
-                    v4f32 _r11 = (v4f32)__msa_splati_w(_r1, 1);
-                    v4f32 _r12 = (v4f32)__msa_splati_w(_r1, 2);
-                    v4f32 _r13 = (v4f32)__msa_splati_w(_r1, 3);
-                    v4f32 _r14 = (v4f32)__msa_splati_w(_r1n, 0);
-                    v4f32 _r15 = (v4f32)__msa_splati_w(_r1n, 1);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r10, _k10);
-                    _sum1 = __msa_fmadd_w(_sum1, _r11, _k10);
-                    _sum2 = __msa_fmadd_w(_sum2, _r12, _k10);
-                    _sum3 = __msa_fmadd_w(_sum3, _r13, _k10);
-                    _sum0 = __msa_fmadd_w(_sum0, _r11, _k11);
-                    _sum1 = __msa_fmadd_w(_sum1, _r12, _k11);
-                    _sum2 = __msa_fmadd_w(_sum2, _r13, _k11);
-                    _sum3 = __msa_fmadd_w(_sum3, _r14, _k11);
-                    _sum0 = __msa_fmadd_w(_sum0, _r12, _k12);
-                    _sum1 = __msa_fmadd_w(_sum1, _r13, _k12);
-                    _sum2 = __msa_fmadd_w(_sum2, _r14, _k12);
-                    _sum3 = __msa_fmadd_w(_sum3, _r15, _k12);
-
-                    v4i32 _r2 = __msa_ld_w(r2, 0);
-                    v4i32 _r2n = __msa_ld_w(r2 + 4, 0);
-
-                    v4f32 _r20 = (v4f32)__msa_splati_w(_r2, 0);
-                    v4f32 _r21 = (v4f32)__msa_splati_w(_r2, 1);
-                    v4f32 _r22 = (v4f32)__msa_splati_w(_r2, 2);
-                    v4f32 _r23 = (v4f32)__msa_splati_w(_r2, 3);
-                    v4f32 _r24 = (v4f32)__msa_splati_w(_r2n, 0);
-                    v4f32 _r25 = (v4f32)__msa_splati_w(_r2n, 1);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r20, _k20);
-                    _sum1 = __msa_fmadd_w(_sum1, _r21, _k20);
-                    _sum2 = __msa_fmadd_w(_sum2, _r22, _k20);
-                    _sum3 = __msa_fmadd_w(_sum3, _r23, _k20);
-                    _sum0 = __msa_fmadd_w(_sum0, _r21, _k21);
-                    _sum1 = __msa_fmadd_w(_sum1, _r22, _k21);
-                    _sum2 = __msa_fmadd_w(_sum2, _r23, _k21);
-                    _sum3 = __msa_fmadd_w(_sum3, _r24, _k21);
-                    _sum0 = __msa_fmadd_w(_sum0, _r22, _k22);
-                    _sum1 = __msa_fmadd_w(_sum1, _r23, _k22);
-                    _sum2 = __msa_fmadd_w(_sum2, _r24, _k22);
-                    _sum3 = __msa_fmadd_w(_sum3, _r25, _k22);
-
-                    __msa_st_w((v4i32)_sum0, outptr0, 0);
-                    __msa_st_w((v4i32)_sum1, outptr0 + 4, 0);
-                    __msa_st_w((v4i32)_sum2, outptr0 + 4 * 2, 0);
-                    __msa_st_w((v4i32)_sum3, outptr0 + 4 * 3, 0);
-
-                    outptr0 += 4 * 4;
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    v4f32 _sum0 = (v4f32)__msa_ld_w(outptr0, 0);
-                    v4f32 _sum1 = (v4f32)__msa_ld_w(outptr0 + 4, 0);
-
-                    v4i32 _r0 = __msa_ld_w(r0, 0);
-                    v4f32 _r00 = (v4f32)__msa_splati_w(_r0, 0);
-                    v4f32 _r01 = (v4f32)__msa_splati_w(_r0, 1);
-                    v4f32 _r02 = (v4f32)__msa_splati_w(_r0, 2);
-                    v4f32 _r03 = (v4f32)__msa_splati_w(_r0, 3);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r00, _k00);
-                    _sum1 = __msa_fmadd_w(_sum1, _r01, _k00);
-                    _sum0 = __msa_fmadd_w(_sum0, _r01, _k01);
-                    _sum1 = __msa_fmadd_w(_sum1, _r02, _k01);
-                    _sum0 = __msa_fmadd_w(_sum0, _r02, _k02);
-                    _sum1 = __msa_fmadd_w(_sum1, _r03, _k02);
-
-                    v4i32 _r1 = __msa_ld_w(r1, 0);
-                    v4f32 _r10 = (v4f32)__msa_splati_w(_r1, 0);
-                    v4f32 _r11 = (v4f32)__msa_splati_w(_r1, 1);
-                    v4f32 _r12 = (v4f32)__msa_splati_w(_r1, 2);
-                    v4f32 _r13 = (v4f32)__msa_splati_w(_r1, 3);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r10, _k10);
-                    _sum1 = __msa_fmadd_w(_sum1, _r11, _k10);
-                    _sum0 = __msa_fmadd_w(_sum0, _r11, _k11);
-                    _sum1 = __msa_fmadd_w(_sum1, _r12, _k11);
-                    _sum0 = __msa_fmadd_w(_sum0, _r12, _k12);
-                    _sum1 = __msa_fmadd_w(_sum1, _r13, _k12);
-
-                    v4i32 _r2 = __msa_ld_w(r2, 0);
-                    v4f32 _r20 = (v4f32)__msa_splati_w(_r2, 0);
-                    v4f32 _r21 = (v4f32)__msa_splati_w(_r2, 1);
-                    v4f32 _r22 = (v4f32)__msa_splati_w(_r2, 2);
-                    v4f32 _r23 = (v4f32)__msa_splati_w(_r2, 3);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r20, _k20);
-                    _sum1 = __msa_fmadd_w(_sum1, _r21, _k20);
-                    _sum0 = __msa_fmadd_w(_sum0, _r21, _k21);
-                    _sum1 = __msa_fmadd_w(_sum1, _r22, _k21);
-                    _sum0 = __msa_fmadd_w(_sum0, _r22, _k22);
-                    _sum1 = __msa_fmadd_w(_sum1, _r23, _k22);
-
-                    __msa_st_w((v4i32)_sum0, outptr0, 0);
-                    __msa_st_w((v4i32)_sum1, outptr0 + 4, 0);
-
-                    outptr0 += 4 * 2;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                }
-                for (; j < outw; j++)
-                {
-                    v4f32 _sum0 = (v4f32)__msa_ld_w(outptr0, 0);
-
-                    v4i32 _r0 = __msa_ld_w(r0, 0);
-                    v4f32 _r00 = (v4f32)__msa_splati_w(_r0, 0);
-                    v4f32 _r01 = (v4f32)__msa_splati_w(_r0, 1);
-                    v4f32 _r02 = (v4f32)__msa_splati_w(_r0, 2);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r00, _k00);
-                    _sum0 = __msa_fmadd_w(_sum0, _r01, _k01);
-                    _sum0 = __msa_fmadd_w(_sum0, _r02, _k02);
-
-                    v4i32 _r1 = __msa_ld_w(r1, 0);
-                    v4f32 _r10 = (v4f32)__msa_splati_w(_r1, 0);
-                    v4f32 _r11 = (v4f32)__msa_splati_w(_r1, 1);
-                    v4f32 _r12 = (v4f32)__msa_splati_w(_r1, 2);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r10, _k10);
-                    _sum0 = __msa_fmadd_w(_sum0, _r11, _k11);
-                    _sum0 = __msa_fmadd_w(_sum0, _r12, _k12);
-
-                    v4i32 _r2 = __msa_ld_w(r2, 0);
-                    v4f32 _r20 = (v4f32)__msa_splati_w(_r2, 0);
-                    v4f32 _r21 = (v4f32)__msa_splati_w(_r2, 1);
-                    v4f32 _r22 = (v4f32)__msa_splati_w(_r2, 2);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r20, _k20);
-                    _sum0 = __msa_fmadd_w(_sum0, _r21, _k21);
-                    _sum0 = __msa_fmadd_w(_sum0, _r22, _k22);
-
-                    __msa_st_w((v4i32)_sum0, outptr0, 0);
-
-                    outptr0 += 4;
-
-                    r0 += 1;
-                    r1 += 1;
-                    r2 += 1;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * 4;
-        }
-    }
-}
-
-static void conv3x3s2_pack1to4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        v4f32 _bias0 = bias ? (v4f32)__msa_ld_w(bias + p * 4, 0) : (v4f32)__msa_fill_w(0);
-        out0.fill(_bias0);
-
-        const float* k0 = kernel.channel(p);
-
-        int q = 0;
-        for (; q < inch; q++)
-        {
-            float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            v4f32 _k00 = (v4f32)__msa_ld_w(k0, 0);
-            v4f32 _k01 = (v4f32)__msa_ld_w(k0 + 4, 0);
-            v4f32 _k02 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-            v4f32 _k10 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-            v4f32 _k11 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-            v4f32 _k12 = (v4f32)__msa_ld_w(k0 + 4 * 5, 0);
-            v4f32 _k20 = (v4f32)__msa_ld_w(k0 + 4 * 6, 0);
-            v4f32 _k21 = (v4f32)__msa_ld_w(k0 + 4 * 7, 0);
-            v4f32 _k22 = (v4f32)__msa_ld_w(k0 + 4 * 8, 0);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    v4f32 _sum0 = (v4f32)__msa_ld_w(outptr0, 0);
-                    v4f32 _sum1 = (v4f32)__msa_ld_w(outptr0 + 4, 0);
-                    v4f32 _sum2 = (v4f32)__msa_ld_w(outptr0 + 4 * 2, 0);
-                    v4f32 _sum3 = (v4f32)__msa_ld_w(outptr0 + 4 * 3, 0);
-                    v4f32 _sum4 = (v4f32)__msa_ld_w(outptr0 + 4 * 4, 0);
-                    v4f32 _sum5 = (v4f32)__msa_ld_w(outptr0 + 4 * 5, 0);
-                    v4f32 _sum6 = (v4f32)__msa_ld_w(outptr0 + 4 * 6, 0);
-                    v4f32 _sum7 = (v4f32)__msa_ld_w(outptr0 + 4 * 7, 0);
-
-                    v4i32 _r0 = __msa_ld_w(r0, 0);
-                    v4i32 _r0n = __msa_ld_w(r0 + 4, 0);
-                    v4i32 _r0nn = __msa_ld_w(r0 + 8, 0);
-                    v4i32 _r0nnn = __msa_ld_w(r0 + 12, 0);
-
-                    v4f32 _r00 = (v4f32)__msa_splati_w(_r0, 0);
-                    v4f32 _r01 = (v4f32)__msa_splati_w(_r0, 1);
-                    v4f32 _r02 = (v4f32)__msa_splati_w(_r0, 2);
-                    v4f32 _r03 = (v4f32)__msa_splati_w(_r0, 3);
-                    v4f32 _r04 = (v4f32)__msa_splati_w(_r0n, 0);
-                    v4f32 _r05 = (v4f32)__msa_splati_w(_r0n, 1);
-                    v4f32 _r06 = (v4f32)__msa_splati_w(_r0n, 2);
-                    v4f32 _r07 = (v4f32)__msa_splati_w(_r0n, 3);
-                    v4f32 _r08 = (v4f32)__msa_splati_w(_r0nn, 0);
-                    v4f32 _r09 = (v4f32)__msa_splati_w(_r0nn, 1);
-                    v4f32 _r0a = (v4f32)__msa_splati_w(_r0nn, 2);
-                    v4f32 _r0b = (v4f32)__msa_splati_w(_r0nn, 3);
-                    v4f32 _r0c = (v4f32)__msa_splati_w(_r0nnn, 0);
-                    v4f32 _r0d = (v4f32)__msa_splati_w(_r0nnn, 1);
-                    v4f32 _r0e = (v4f32)__msa_splati_w(_r0nnn, 2);
-                    v4f32 _r0f = (v4f32)__msa_splati_w(_r0nnn, 3);
-                    v4f32 _r0g = __msa_fill_w_f32(r0[16]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r00, _k00);
-                    _sum1 = __msa_fmadd_w(_sum1, _r02, _k00);
-                    _sum2 = __msa_fmadd_w(_sum2, _r04, _k00);
-                    _sum3 = __msa_fmadd_w(_sum3, _r06, _k00);
-                    _sum4 = __msa_fmadd_w(_sum4, _r08, _k00);
-                    _sum5 = __msa_fmadd_w(_sum5, _r0a, _k00);
-                    _sum6 = __msa_fmadd_w(_sum6, _r0c, _k00);
-                    _sum7 = __msa_fmadd_w(_sum7, _r0e, _k00);
-                    _sum0 = __msa_fmadd_w(_sum0, _r01, _k01);
-                    _sum1 = __msa_fmadd_w(_sum1, _r03, _k01);
-                    _sum2 = __msa_fmadd_w(_sum2, _r05, _k01);
-                    _sum3 = __msa_fmadd_w(_sum3, _r07, _k01);
-                    _sum4 = __msa_fmadd_w(_sum4, _r09, _k01);
-                    _sum5 = __msa_fmadd_w(_sum5, _r0b, _k01);
-                    _sum6 = __msa_fmadd_w(_sum6, _r0d, _k01);
-                    _sum7 = __msa_fmadd_w(_sum7, _r0f, _k01);
-                    _sum0 = __msa_fmadd_w(_sum0, _r02, _k02);
-                    _sum1 = __msa_fmadd_w(_sum1, _r04, _k02);
-                    _sum2 = __msa_fmadd_w(_sum2, _r06, _k02);
-                    _sum3 = __msa_fmadd_w(_sum3, _r08, _k02);
-                    _sum4 = __msa_fmadd_w(_sum4, _r0a, _k02);
-                    _sum5 = __msa_fmadd_w(_sum5, _r0c, _k02);
-                    _sum6 = __msa_fmadd_w(_sum6, _r0e, _k02);
-                    _sum7 = __msa_fmadd_w(_sum7, _r0g, _k02);
-
-                    v4i32 _r1 = __msa_ld_w(r1, 0);
-                    v4i32 _r1n = __msa_ld_w(r1 + 4, 0);
-                    v4i32 _r1nn = __msa_ld_w(r1 + 8, 0);
-                    v4i32 _r1nnn = __msa_ld_w(r1 + 12, 0);
-
-                    v4f32 _r10 = (v4f32)__msa_splati_w(_r1, 0);
-                    v4f32 _r11 = (v4f32)__msa_splati_w(_r1, 1);
-                    v4f32 _r12 = (v4f32)__msa_splati_w(_r1, 2);
-                    v4f32 _r13 = (v4f32)__msa_splati_w(_r1, 3);
-                    v4f32 _r14 = (v4f32)__msa_splati_w(_r1n, 0);
-                    v4f32 _r15 = (v4f32)__msa_splati_w(_r1n, 1);
-                    v4f32 _r16 = (v4f32)__msa_splati_w(_r1n, 2);
-                    v4f32 _r17 = (v4f32)__msa_splati_w(_r1n, 3);
-                    v4f32 _r18 = (v4f32)__msa_splati_w(_r1nn, 0);
-                    v4f32 _r19 = (v4f32)__msa_splati_w(_r1nn, 1);
-                    v4f32 _r1a = (v4f32)__msa_splati_w(_r1nn, 2);
-                    v4f32 _r1b = (v4f32)__msa_splati_w(_r1nn, 3);
-                    v4f32 _r1c = (v4f32)__msa_splati_w(_r1nnn, 0);
-                    v4f32 _r1d = (v4f32)__msa_splati_w(_r1nnn, 1);
-                    v4f32 _r1e = (v4f32)__msa_splati_w(_r1nnn, 2);
-                    v4f32 _r1f = (v4f32)__msa_splati_w(_r1nnn, 3);
-                    v4f32 _r1g = __msa_fill_w_f32(r1[16]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r10, _k10);
-                    _sum1 = __msa_fmadd_w(_sum1, _r12, _k10);
-                    _sum2 = __msa_fmadd_w(_sum2, _r14, _k10);
-                    _sum3 = __msa_fmadd_w(_sum3, _r16, _k10);
-                    _sum4 = __msa_fmadd_w(_sum4, _r18, _k10);
-                    _sum5 = __msa_fmadd_w(_sum5, _r1a, _k10);
-                    _sum6 = __msa_fmadd_w(_sum6, _r1c, _k10);
-                    _sum7 = __msa_fmadd_w(_sum7, _r1e, _k10);
-                    _sum0 = __msa_fmadd_w(_sum0, _r11, _k11);
-                    _sum1 = __msa_fmadd_w(_sum1, _r13, _k11);
-                    _sum2 = __msa_fmadd_w(_sum2, _r15, _k11);
-                    _sum3 = __msa_fmadd_w(_sum3, _r17, _k11);
-                    _sum4 = __msa_fmadd_w(_sum4, _r19, _k11);
-                    _sum5 = __msa_fmadd_w(_sum5, _r1b, _k11);
-                    _sum6 = __msa_fmadd_w(_sum6, _r1d, _k11);
-                    _sum7 = __msa_fmadd_w(_sum7, _r1f, _k11);
-                    _sum0 = __msa_fmadd_w(_sum0, _r12, _k12);
-                    _sum1 = __msa_fmadd_w(_sum1, _r14, _k12);
-                    _sum2 = __msa_fmadd_w(_sum2, _r16, _k12);
-                    _sum3 = __msa_fmadd_w(_sum3, _r18, _k12);
-                    _sum4 = __msa_fmadd_w(_sum4, _r1a, _k12);
-                    _sum5 = __msa_fmadd_w(_sum5, _r1c, _k12);
-                    _sum6 = __msa_fmadd_w(_sum6, _r1e, _k12);
-                    _sum7 = __msa_fmadd_w(_sum7, _r1g, _k12);
-
-                    v4i32 _r2 = __msa_ld_w(r2, 0);
-                    v4i32 _r2n = __msa_ld_w(r2 + 4, 0);
-                    v4i32 _r2nn = __msa_ld_w(r2 + 8, 0);
-                    v4i32 _r2nnn = __msa_ld_w(r2 + 12, 0);
-
-                    v4f32 _r20 = (v4f32)__msa_splati_w(_r2, 0);
-                    v4f32 _r21 = (v4f32)__msa_splati_w(_r2, 1);
-                    v4f32 _r22 = (v4f32)__msa_splati_w(_r2, 2);
-                    v4f32 _r23 = (v4f32)__msa_splati_w(_r2, 3);
-                    v4f32 _r24 = (v4f32)__msa_splati_w(_r2n, 0);
-                    v4f32 _r25 = (v4f32)__msa_splati_w(_r2n, 1);
-                    v4f32 _r26 = (v4f32)__msa_splati_w(_r2n, 2);
-                    v4f32 _r27 = (v4f32)__msa_splati_w(_r2n, 3);
-                    v4f32 _r28 = (v4f32)__msa_splati_w(_r2nn, 0);
-                    v4f32 _r29 = (v4f32)__msa_splati_w(_r2nn, 1);
-                    v4f32 _r2a = (v4f32)__msa_splati_w(_r2nn, 2);
-                    v4f32 _r2b = (v4f32)__msa_splati_w(_r2nn, 3);
-                    v4f32 _r2c = (v4f32)__msa_splati_w(_r2nnn, 0);
-                    v4f32 _r2d = (v4f32)__msa_splati_w(_r2nnn, 1);
-                    v4f32 _r2e = (v4f32)__msa_splati_w(_r2nnn, 2);
-                    v4f32 _r2f = (v4f32)__msa_splati_w(_r2nnn, 3);
-                    v4f32 _r2g = __msa_fill_w_f32(r2[16]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r20, _k20);
-                    _sum1 = __msa_fmadd_w(_sum1, _r22, _k20);
-                    _sum2 = __msa_fmadd_w(_sum2, _r24, _k20);
-                    _sum3 = __msa_fmadd_w(_sum3, _r26, _k20);
-                    _sum4 = __msa_fmadd_w(_sum4, _r28, _k20);
-                    _sum5 = __msa_fmadd_w(_sum5, _r2a, _k20);
-                    _sum6 = __msa_fmadd_w(_sum6, _r2c, _k20);
-                    _sum7 = __msa_fmadd_w(_sum7, _r2e, _k20);
-                    _sum0 = __msa_fmadd_w(_sum0, _r21, _k21);
-                    _sum1 = __msa_fmadd_w(_sum1, _r23, _k21);
-                    _sum2 = __msa_fmadd_w(_sum2, _r25, _k21);
-                    _sum3 = __msa_fmadd_w(_sum3, _r27, _k21);
-                    _sum4 = __msa_fmadd_w(_sum4, _r29, _k21);
-                    _sum5 = __msa_fmadd_w(_sum5, _r2b, _k21);
-                    _sum6 = __msa_fmadd_w(_sum6, _r2d, _k21);
-                    _sum7 = __msa_fmadd_w(_sum7, _r2f, _k21);
-                    _sum0 = __msa_fmadd_w(_sum0, _r22, _k22);
-                    _sum1 = __msa_fmadd_w(_sum1, _r24, _k22);
-                    _sum2 = __msa_fmadd_w(_sum2, _r26, _k22);
-                    _sum3 = __msa_fmadd_w(_sum3, _r28, _k22);
-                    _sum4 = __msa_fmadd_w(_sum4, _r2a, _k22);
-                    _sum5 = __msa_fmadd_w(_sum5, _r2c, _k22);
-                    _sum6 = __msa_fmadd_w(_sum6, _r2e, _k22);
-                    _sum7 = __msa_fmadd_w(_sum7, _r2g, _k22);
-
-                    __msa_st_w((v4i32)_sum0, outptr0, 0);
-                    __msa_st_w((v4i32)_sum1, outptr0 + 4, 0);
-                    __msa_st_w((v4i32)_sum2, outptr0 + 4 * 2, 0);
-                    __msa_st_w((v4i32)_sum3, outptr0 + 4 * 3, 0);
-                    __msa_st_w((v4i32)_sum4, outptr0 + 4 * 4, 0);
-                    __msa_st_w((v4i32)_sum5, outptr0 + 4 * 5, 0);
-                    __msa_st_w((v4i32)_sum6, outptr0 + 4 * 6, 0);
-                    __msa_st_w((v4i32)_sum7, outptr0 + 4 * 7, 0);
-
-                    outptr0 += 4 * 8;
-
-                    r0 += 16;
-                    r1 += 16;
-                    r2 += 16;
-                }
-                for (; j + 3 < outw; j += 4)
-                {
-                    v4f32 _sum0 = (v4f32)__msa_ld_w(outptr0, 0);
-                    v4f32 _sum1 = (v4f32)__msa_ld_w(outptr0 + 4, 0);
-                    v4f32 _sum2 = (v4f32)__msa_ld_w(outptr0 + 4 * 2, 0);
-                    v4f32 _sum3 = (v4f32)__msa_ld_w(outptr0 + 4 * 3, 0);
-
-                    v4i32 _r0 = __msa_ld_w(r0, 0);
-                    v4i32 _r0n = __msa_ld_w(r0 + 4, 0);
-
-                    v4f32 _r00 = (v4f32)__msa_splati_w(_r0, 0);
-                    v4f32 _r01 = (v4f32)__msa_splati_w(_r0, 1);
-                    v4f32 _r02 = (v4f32)__msa_splati_w(_r0, 2);
-                    v4f32 _r03 = (v4f32)__msa_splati_w(_r0, 3);
-                    v4f32 _r04 = (v4f32)__msa_splati_w(_r0n, 0);
-                    v4f32 _r05 = (v4f32)__msa_splati_w(_r0n, 1);
-                    v4f32 _r06 = (v4f32)__msa_splati_w(_r0n, 2);
-                    v4f32 _r07 = (v4f32)__msa_splati_w(_r0n, 3);
-                    v4f32 _r08 = __msa_fill_w_f32(r0[8]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r00, _k00);
-                    _sum1 = __msa_fmadd_w(_sum1, _r02, _k00);
-                    _sum2 = __msa_fmadd_w(_sum2, _r04, _k00);
-                    _sum3 = __msa_fmadd_w(_sum3, _r06, _k00);
-                    _sum0 = __msa_fmadd_w(_sum0, _r01, _k01);
-                    _sum1 = __msa_fmadd_w(_sum1, _r03, _k01);
-                    _sum2 = __msa_fmadd_w(_sum2, _r05, _k01);
-                    _sum3 = __msa_fmadd_w(_sum3, _r07, _k01);
-                    _sum0 = __msa_fmadd_w(_sum0, _r02, _k02);
-                    _sum1 = __msa_fmadd_w(_sum1, _r04, _k02);
-                    _sum2 = __msa_fmadd_w(_sum2, _r06, _k02);
-                    _sum3 = __msa_fmadd_w(_sum3, _r08, _k02);
-
-                    v4i32 _r1 = __msa_ld_w(r1, 0);
-                    v4i32 _r1n = __msa_ld_w(r1 + 4, 0);
-
-                    v4f32 _r10 = (v4f32)__msa_splati_w(_r1, 0);
-                    v4f32 _r11 = (v4f32)__msa_splati_w(_r1, 1);
-                    v4f32 _r12 = (v4f32)__msa_splati_w(_r1, 2);
-                    v4f32 _r13 = (v4f32)__msa_splati_w(_r1, 3);
-                    v4f32 _r14 = (v4f32)__msa_splati_w(_r1n, 0);
-                    v4f32 _r15 = (v4f32)__msa_splati_w(_r1n, 1);
-                    v4f32 _r16 = (v4f32)__msa_splati_w(_r1n, 2);
-                    v4f32 _r17 = (v4f32)__msa_splati_w(_r1n, 3);
-                    v4f32 _r18 = __msa_fill_w_f32(r1[8]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r10, _k10);
-                    _sum1 = __msa_fmadd_w(_sum1, _r12, _k10);
-                    _sum2 = __msa_fmadd_w(_sum2, _r14, _k10);
-                    _sum3 = __msa_fmadd_w(_sum3, _r16, _k10);
-                    _sum0 = __msa_fmadd_w(_sum0, _r11, _k11);
-                    _sum1 = __msa_fmadd_w(_sum1, _r13, _k11);
-                    _sum2 = __msa_fmadd_w(_sum2, _r15, _k11);
-                    _sum3 = __msa_fmadd_w(_sum3, _r17, _k11);
-                    _sum0 = __msa_fmadd_w(_sum0, _r12, _k12);
-                    _sum1 = __msa_fmadd_w(_sum1, _r14, _k12);
-                    _sum2 = __msa_fmadd_w(_sum2, _r16, _k12);
-                    _sum3 = __msa_fmadd_w(_sum3, _r18, _k12);
-
-                    v4i32 _r2 = __msa_ld_w(r2, 0);
-                    v4i32 _r2n = __msa_ld_w(r2 + 4, 0);
-
-                    v4f32 _r20 = (v4f32)__msa_splati_w(_r2, 0);
-                    v4f32 _r21 = (v4f32)__msa_splati_w(_r2, 1);
-                    v4f32 _r22 = (v4f32)__msa_splati_w(_r2, 2);
-                    v4f32 _r23 = (v4f32)__msa_splati_w(_r2, 3);
-                    v4f32 _r24 = (v4f32)__msa_splati_w(_r2n, 0);
-                    v4f32 _r25 = (v4f32)__msa_splati_w(_r2n, 1);
-                    v4f32 _r26 = (v4f32)__msa_splati_w(_r2n, 2);
-                    v4f32 _r27 = (v4f32)__msa_splati_w(_r2n, 3);
-                    v4f32 _r28 = __msa_fill_w_f32(r2[8]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r20, _k20);
-                    _sum1 = __msa_fmadd_w(_sum1, _r22, _k20);
-                    _sum2 = __msa_fmadd_w(_sum2, _r24, _k20);
-                    _sum3 = __msa_fmadd_w(_sum3, _r26, _k20);
-                    _sum0 = __msa_fmadd_w(_sum0, _r21, _k21);
-                    _sum1 = __msa_fmadd_w(_sum1, _r23, _k21);
-                    _sum2 = __msa_fmadd_w(_sum2, _r25, _k21);
-                    _sum3 = __msa_fmadd_w(_sum3, _r27, _k21);
-                    _sum0 = __msa_fmadd_w(_sum0, _r22, _k22);
-                    _sum1 = __msa_fmadd_w(_sum1, _r24, _k22);
-                    _sum2 = __msa_fmadd_w(_sum2, _r26, _k22);
-                    _sum3 = __msa_fmadd_w(_sum3, _r28, _k22);
-
-                    __msa_st_w((v4i32)_sum0, outptr0, 0);
-                    __msa_st_w((v4i32)_sum1, outptr0 + 4, 0);
-                    __msa_st_w((v4i32)_sum2, outptr0 + 4 * 2, 0);
-                    __msa_st_w((v4i32)_sum3, outptr0 + 4 * 3, 0);
-
-                    outptr0 += 4 * 4;
-
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    v4f32 _sum0 = (v4f32)__msa_ld_w(outptr0, 0);
-                    v4f32 _sum1 = (v4f32)__msa_ld_w(outptr0 + 4, 0);
-
-                    v4i32 _r0 = __msa_ld_w(r0, 0);
-                    v4f32 _r00 = (v4f32)__msa_splati_w(_r0, 0);
-                    v4f32 _r01 = (v4f32)__msa_splati_w(_r0, 1);
-                    v4f32 _r02 = (v4f32)__msa_splati_w(_r0, 2);
-                    v4f32 _r03 = (v4f32)__msa_splati_w(_r0, 3);
-                    v4f32 _r04 = __msa_fill_w_f32(r0[4]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r00, _k00);
-                    _sum1 = __msa_fmadd_w(_sum1, _r02, _k00);
-                    _sum0 = __msa_fmadd_w(_sum0, _r01, _k01);
-                    _sum1 = __msa_fmadd_w(_sum1, _r03, _k01);
-                    _sum0 = __msa_fmadd_w(_sum0, _r02, _k02);
-                    _sum1 = __msa_fmadd_w(_sum1, _r04, _k02);
-
-                    v4i32 _r1 = __msa_ld_w(r1, 0);
-                    v4f32 _r10 = (v4f32)__msa_splati_w(_r1, 0);
-                    v4f32 _r11 = (v4f32)__msa_splati_w(_r1, 1);
-                    v4f32 _r12 = (v4f32)__msa_splati_w(_r1, 2);
-                    v4f32 _r13 = (v4f32)__msa_splati_w(_r1, 3);
-                    v4f32 _r14 = __msa_fill_w_f32(r1[4]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r10, _k10);
-                    _sum1 = __msa_fmadd_w(_sum1, _r12, _k10);
-                    _sum0 = __msa_fmadd_w(_sum0, _r11, _k11);
-                    _sum1 = __msa_fmadd_w(_sum1, _r13, _k11);
-                    _sum0 = __msa_fmadd_w(_sum0, _r12, _k12);
-                    _sum1 = __msa_fmadd_w(_sum1, _r14, _k12);
-
-                    v4i32 _r2 = __msa_ld_w(r2, 0);
-                    v4f32 _r20 = (v4f32)__msa_splati_w(_r2, 0);
-                    v4f32 _r21 = (v4f32)__msa_splati_w(_r2, 1);
-                    v4f32 _r22 = (v4f32)__msa_splati_w(_r2, 2);
-                    v4f32 _r23 = (v4f32)__msa_splati_w(_r2, 3);
-                    v4f32 _r24 = __msa_fill_w_f32(r2[4]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r20, _k20);
-                    _sum1 = __msa_fmadd_w(_sum1, _r22, _k20);
-                    _sum0 = __msa_fmadd_w(_sum0, _r21, _k21);
-                    _sum1 = __msa_fmadd_w(_sum1, _r23, _k21);
-                    _sum0 = __msa_fmadd_w(_sum0, _r22, _k22);
-                    _sum1 = __msa_fmadd_w(_sum1, _r24, _k22);
-
-                    __msa_st_w((v4i32)_sum0, outptr0, 0);
-                    __msa_st_w((v4i32)_sum1, outptr0 + 4, 0);
-
-                    outptr0 += 4 * 2;
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                }
-                for (; j < outw; j++)
-                {
-                    v4f32 _sum0 = (v4f32)__msa_ld_w(outptr0, 0);
-
-                    v4i32 _r0 = __msa_ld_w(r0, 0);
-                    v4f32 _r00 = (v4f32)__msa_splati_w(_r0, 0);
-                    v4f32 _r01 = (v4f32)__msa_splati_w(_r0, 1);
-                    v4f32 _r02 = (v4f32)__msa_splati_w(_r0, 2);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r00, _k00);
-                    _sum0 = __msa_fmadd_w(_sum0, _r01, _k01);
-                    _sum0 = __msa_fmadd_w(_sum0, _r02, _k02);
-
-                    v4i32 _r1 = __msa_ld_w(r1, 0);
-                    v4f32 _r10 = (v4f32)__msa_splati_w(_r1, 0);
-                    v4f32 _r11 = (v4f32)__msa_splati_w(_r1, 1);
-                    v4f32 _r12 = (v4f32)__msa_splati_w(_r1, 2);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r10, _k10);
-                    _sum0 = __msa_fmadd_w(_sum0, _r11, _k11);
-                    _sum0 = __msa_fmadd_w(_sum0, _r12, _k12);
-
-                    v4i32 _r2 = __msa_ld_w(r2, 0);
-                    v4f32 _r20 = (v4f32)__msa_splati_w(_r2, 0);
-                    v4f32 _r21 = (v4f32)__msa_splati_w(_r2, 1);
-                    v4f32 _r22 = (v4f32)__msa_splati_w(_r2, 2);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r20, _k20);
-                    _sum0 = __msa_fmadd_w(_sum0, _r21, _k21);
-                    _sum0 = __msa_fmadd_w(_sum0, _r22, _k22);
-
-                    __msa_st_w((v4i32)_sum0, outptr0, 0);
-
-                    outptr0 += 4;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack4.h
deleted file mode 100644
index 2f257f19..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack4.h
+++ /dev/null
@@ -1,425 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd63_transform_kernel_pack4_msa(const Mat& kernel, Mat& kernel_tm_pack4, int inch, int outch, const Option& opt)
-{
-    // winograd63 transform kernel
-    Mat kernel_tm;
-    kernel_tm.create(8 * 8, inch, outch);
-
-    const float ktm[8][3] = {
-        {1.0f, 0.0f, 0.0f},
-        {-2.0f / 9, -2.0f / 9, -2.0f / 9},
-        {-2.0f / 9, 2.0f / 9, -2.0f / 9},
-        {1.0f / 90, 1.0f / 45, 2.0f / 45},
-        {1.0f / 90, -1.0f / 45, 2.0f / 45},
-        {1.0f / 45, 1.0f / 90, 1.0f / 180},
-        {1.0f / 45, -1.0f / 90, 1.0f / 180},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel, transposed
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[8][3];
-            for (int i = 0; i < 8; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // v
-            for (int j = 0; j < 8; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 8; i++)
-                {
-                    kernel_tm0[j * 8 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 64-inch-outch
-    // dst = pb-pa-inch/pa-64-outch/pb
-    kernel_tm_pack4.create(inch / 4, 64, outch / 4, (size_t)4u * 4 * 4, 4 * 4);
-
-    for (int q = 0; q + 3 < outch; q += 4)
-    {
-        Mat g0 = kernel_tm_pack4.channel(q / 4);
-
-        for (int k = 0; k < 64; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p + 3 < inch; p += 4)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 4; j++)
-                    {
-                        const float* k00 = kernel_tm.channel(q + j).row(p + i);
-                        g00[0] = k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd63_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 6n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 5) / 6 * 6;
-    outh = (outh + 5) / 6 * 6;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 6;
-        int h_tiles = outh / 6;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 64, inch, elemsize, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd63_transform_input_pack4_msa(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_pack4_msa(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, elemsize, elempack, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd63_transform_output_pack4_msa(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
-
-static void conv3x3s1_winograd43_transform_kernel_pack4_msa(const Mat& kernel, Mat& kernel_tm_pack4, int inch, int outch, const Option& opt)
-{
-    // winograd43 transform kernel
-    Mat kernel_tm(6 * 6, inch, outch);
-
-    const float ktm[6][3] = {
-        {1.0f / 4, 0.0f, 0.0f},
-        {-1.0f / 6, -1.0f / 6, -1.0f / 6},
-        {-1.0f / 6, 1.0f / 6, -1.0f / 6},
-        {1.0f / 24, 1.0f / 12, 1.0f / 6},
-        {1.0f / 24, -1.0f / 12, 1.0f / 6},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[6][3];
-            for (int i = 0; i < 6; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 6; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 6; i++)
-                {
-                    kernel_tm0[j * 6 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 36-inch-outch
-    // dst = pb-pa-inch/pa-36-outch/pb
-    kernel_tm_pack4.create(inch / 4, 36, outch / 4, (size_t)4u * 4 * 4, 4 * 4);
-
-    for (int q = 0; q + 3 < outch; q += 4)
-    {
-        Mat g0 = kernel_tm_pack4.channel(q / 4);
-
-        for (int k = 0; k < 36; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p + 3 < inch; p += 4)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 4; j++)
-                    {
-                        const float* k00 = kernel_tm.channel(q + j).row(p + i);
-                        g00[0] = k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 4n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 3) / 4 * 4;
-    outh = (outh + 3) / 4 * 4;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 4;
-        int h_tiles = outh / 4;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 36, inch, elemsize, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd43_transform_input_pack4_msa(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_pack4_msa(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, elemsize, elempack, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd43_transform_output_pack4_msa(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
-
-static void conv3x3s1_winograd23_transform_kernel_pack4_msa(const Mat& kernel, Mat& kernel_tm_pack4, int inch, int outch, const Option& opt)
-{
-    // winograd23 transform kernel
-    Mat kernel_tm(4 * 4, inch, outch);
-
-    const float ktm[4][3] = {
-        {1.0f, 0.0f, 0.0f},
-        {1.0f / 2, 1.0f / 2, 1.0f / 2},
-        {1.0f / 2, -1.0f / 2, 1.0f / 2},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[4][3];
-            for (int i = 0; i < 4; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 4; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 4; i++)
-                {
-                    kernel_tm0[j * 4 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 16-inch-outch
-    // dst = pb-pa-inch/pa-16-outch/pb
-    kernel_tm_pack4.create(inch / 4, 16, outch / 4, (size_t)4u * 4 * 4, 4 * 4);
-
-    for (int q = 0; q + 3 < outch; q += 4)
-    {
-        Mat g0 = kernel_tm_pack4.channel(q / 4);
-
-        for (int k = 0; k < 16; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p + 3 < inch; p += 4)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 4; j++)
-                    {
-                        const float* k00 = kernel_tm.channel(q + j).row(p + i);
-                        g00[0] = k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 2n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 1) / 2 * 2;
-    outh = (outh + 1) / 2 * 2;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 2;
-        int h_tiles = outh / 2;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 16, inch, elemsize, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd23_transform_input_pack4_msa(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_pack4_msa(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, elemsize, elempack, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd23_transform_output_pack4_msa(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack8to1_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack8to1_int8.h
deleted file mode 100644
index 986afc60..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack8to1_int8.h
+++ /dev/null
@@ -1,177 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd43_transform_kernel_pack8to1_int8_msa(const Mat& kernel, Mat& kernel_tm_pack8to1, int inch, int outch, const Option& opt)
-{
-    // winograd43 transform kernel
-    Mat kernel_tm(6 * 6, inch, outch, (size_t)2u);
-
-    const short ktm[6][3] = {
-        {6, 0, 0},
-        {-4, -4, -4},
-        {-4, 4, -4},
-        {1, 2, 4},
-        {1, -2, 4},
-        {0, 0, 6}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const signed char* kernel0 = (const signed char*)kernel + p * inch * 9 + q * 9;
-            short* kernel_tm0 = kernel_tm.channel(p).row<short>(q);
-
-            // transform kernel
-            const signed char* k0 = kernel0;
-            const signed char* k1 = kernel0 + 3;
-            const signed char* k2 = kernel0 + 6;
-
-            // h
-            short tmp[6][3];
-            for (int i = 0; i < 6; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 6; j++)
-            {
-                short* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 6; i++)
-                {
-                    kernel_tm0[j * 6 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 36-inch-outch
-    // dst = 4b-8a-inch/8a-36-outch/4b
-    kernel_tm_pack8to1.create(8 * inch / 8, 36, outch / 4 + outch % 4, (size_t)2u * 4, 4);
-
-    int p = 0;
-    for (; p + 3 < outch; p += 4)
-    {
-        const Mat k0 = kernel_tm.channel(p);
-        const Mat k1 = kernel_tm.channel(p + 1);
-        const Mat k2 = kernel_tm.channel(p + 2);
-        const Mat k3 = kernel_tm.channel(p + 3);
-
-        Mat g0 = kernel_tm_pack8to1.channel(p / 4);
-
-        for (int k = 0; k < 36; k++)
-        {
-            short* g00 = g0.row<short>(k);
-
-            for (int q = 0; q + 7 < inch; q += 8)
-            {
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0.row<const short>(q + i)[k];
-                    g00[1] = k1.row<const short>(q + i)[k];
-                    g00[2] = k2.row<const short>(q + i)[k];
-                    g00[3] = k3.row<const short>(q + i)[k];
-
-                    g00 += 4;
-                }
-            }
-        }
-    }
-    for (; p < outch; p++)
-    {
-        const Mat k0 = kernel_tm.channel(p);
-
-        Mat g0 = kernel_tm_pack8to1.channel(p / 4 + p % 4);
-
-        for (int k = 0; k < 36; k++)
-        {
-            short* g00 = g0.row<short>(k);
-
-            for (int q = 0; q + 7 < inch; q += 8)
-            {
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0.row<const short>(q + i)[k];
-
-                    g00 += 1;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_pack8to1_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    //     size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 4n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 3) / 4 * 4;
-    outh = (outh + 3) / 4 * 4;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 4;
-        int h_tiles = outh / 4;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 36, inch, 2u * elempack, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd43_transform_input_pack8_int8_msa(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_pack8to1_int8_msa(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, 4u, 1, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd43_transform_output_int8_msa(top_blob_tm, top_blob_bordered, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack8to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack8to4_int8.h
deleted file mode 100644
index c77ae4d1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_3x3_pack8to4_int8.h
+++ /dev/null
@@ -1,161 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd43_transform_kernel_pack8to4_int8_msa(const Mat& kernel, Mat& kernel_tm_pack8, int inch, int outch, const Option& opt)
-{
-    // winograd43 transform kernel
-    Mat kernel_tm(6 * 6, inch, outch, (size_t)2u);
-
-    const short ktm[6][3] = {
-        {6, 0, 0},
-        {-4, -4, -4},
-        {-4, 4, -4},
-        {1, 2, 4},
-        {1, -2, 4},
-        {0, 0, 6}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const signed char* kernel0 = (const signed char*)kernel + p * inch * 9 + q * 9;
-            short* kernel_tm0 = kernel_tm.channel(p).row<short>(q);
-
-            // transform kernel
-            const signed char* k0 = kernel0;
-            const signed char* k1 = kernel0 + 3;
-            const signed char* k2 = kernel0 + 6;
-
-            // h
-            short tmp[6][3];
-            for (int i = 0; i < 6; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 6; j++)
-            {
-                short* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 6; i++)
-                {
-                    kernel_tm0[j * 6 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 36-inch-outch
-    // dst = 4b-8a-inch/8a-36-outch/4b
-    kernel_tm_pack8.create(inch / 8, 36, outch / 4, (size_t)2u * 32, 32);
-
-    int q = 0;
-    for (; q + 3 < outch; q += 4)
-    {
-        const Mat k0 = kernel_tm.channel(q);
-        const Mat k1 = kernel_tm.channel(q + 1);
-        const Mat k2 = kernel_tm.channel(q + 2);
-        const Mat k3 = kernel_tm.channel(q + 3);
-
-        Mat kernel_tm = kernel_tm_pack8.channel(q / 4);
-
-        for (int k = 0; k < 36; k++)
-        {
-            short* g00 = kernel_tm.row<short>(k);
-
-            for (int p = 0; p + 7 < inch; p += 8)
-            {
-                for (int i = 0; i < 8; i++)
-                {
-                    const short* k00 = k0.row<const short>(p + i);
-                    const short* k10 = k1.row<const short>(p + i);
-                    const short* k20 = k2.row<const short>(p + i);
-                    const short* k30 = k3.row<const short>(p + i);
-
-                    g00[0] = k00[k];
-                    g00[1] = k10[k];
-                    g00[2] = k20[k];
-                    g00[3] = k30[k];
-
-                    g00 += 4;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_pack8to4_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    //     size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 4n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 3) / 4 * 4;
-    outh = (outh + 3) / 4 * 4;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 4;
-        int h_tiles = outh / 4;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 36, inch, 2u * elempack, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd43_transform_input_pack8_int8_msa(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_pack8to4_int8_msa(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, 4u * 4, 4, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd43_transform_output_pack4_int8_msa(top_blob_tm, top_blob_bordered, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_7x7_pack1to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_7x7_pack1to4.h
deleted file mode 100644
index 8cff5dd0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_7x7_pack1to4.h
+++ /dev/null
@@ -1,652 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv7x7s2_pack1to4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        v4f32 _bias0 = bias ? (v4f32)__msa_ld_w(bias + p * 4, 0) : (v4f32)__msa_fill_w(0);
-        out0.fill(_bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-            const float* r3 = img0.row(3);
-            const float* r4 = img0.row(4);
-            const float* r5 = img0.row(5);
-            const float* r6 = img0.row(6);
-
-            const float* kptr = kernel.channel(p).row(q);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-                    v4f32 _sum0 = (v4f32)__msa_ld_w(outptr0, 0);
-                    v4f32 _sum1 = (v4f32)__msa_ld_w(outptr0 + 4, 0);
-                    v4f32 _sum2 = (v4f32)__msa_ld_w(outptr0 + 4 * 2, 0);
-                    v4f32 _sum3 = (v4f32)__msa_ld_w(outptr0 + 4 * 3, 0);
-
-                    v4f32 _k00 = (v4f32)__msa_ld_w(kptr, 0);
-                    v4f32 _k01 = (v4f32)__msa_ld_w(kptr + 4, 0);
-                    v4f32 _k02 = (v4f32)__msa_ld_w(kptr + 4 * 2, 0);
-                    v4f32 _k03 = (v4f32)__msa_ld_w(kptr + 4 * 3, 0);
-                    v4f32 _k04 = (v4f32)__msa_ld_w(kptr + 4 * 4, 0);
-                    v4f32 _k05 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0);
-                    v4f32 _k06 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    v4i32 _r0 = __msa_ld_w(r0, 0);
-                    v4i32 _r0n = __msa_ld_w(r0 + 4, 0);
-                    v4i32 _r0nn = __msa_ld_w(r0 + 8, 0);
-
-                    v4f32 _r00 = (v4f32)__msa_splati_w(_r0, 0);
-                    v4f32 _r01 = (v4f32)__msa_splati_w(_r0, 1);
-                    v4f32 _r02 = (v4f32)__msa_splati_w(_r0, 2);
-                    v4f32 _r03 = (v4f32)__msa_splati_w(_r0, 3);
-                    v4f32 _r04 = (v4f32)__msa_splati_w(_r0n, 0);
-                    v4f32 _r05 = (v4f32)__msa_splati_w(_r0n, 1);
-                    v4f32 _r06 = (v4f32)__msa_splati_w(_r0n, 2);
-                    v4f32 _r07 = (v4f32)__msa_splati_w(_r0n, 3);
-                    v4f32 _r08 = (v4f32)__msa_splati_w(_r0nn, 0);
-                    v4f32 _r09 = (v4f32)__msa_splati_w(_r0nn, 1);
-                    v4f32 _r0a = (v4f32)__msa_splati_w(_r0nn, 2);
-                    v4f32 _r0b = (v4f32)__msa_splati_w(_r0nn, 3);
-                    v4f32 _r0c = __msa_fill_w_f32(r0[12]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r00, _k00);
-                    _sum1 = __msa_fmadd_w(_sum1, _r02, _k00);
-                    _sum2 = __msa_fmadd_w(_sum2, _r04, _k00);
-                    _sum3 = __msa_fmadd_w(_sum3, _r06, _k00);
-                    _sum0 = __msa_fmadd_w(_sum0, _r01, _k01);
-                    _sum1 = __msa_fmadd_w(_sum1, _r03, _k01);
-                    _sum2 = __msa_fmadd_w(_sum2, _r05, _k01);
-                    _sum3 = __msa_fmadd_w(_sum3, _r07, _k01);
-                    _sum0 = __msa_fmadd_w(_sum0, _r02, _k02);
-                    _sum1 = __msa_fmadd_w(_sum1, _r04, _k02);
-                    _sum2 = __msa_fmadd_w(_sum2, _r06, _k02);
-                    _sum3 = __msa_fmadd_w(_sum3, _r08, _k02);
-                    _sum0 = __msa_fmadd_w(_sum0, _r03, _k03);
-                    _sum1 = __msa_fmadd_w(_sum1, _r05, _k03);
-                    _sum2 = __msa_fmadd_w(_sum2, _r07, _k03);
-                    _sum3 = __msa_fmadd_w(_sum3, _r09, _k03);
-                    _sum0 = __msa_fmadd_w(_sum0, _r04, _k04);
-                    _sum1 = __msa_fmadd_w(_sum1, _r06, _k04);
-                    _sum2 = __msa_fmadd_w(_sum2, _r08, _k04);
-                    _sum3 = __msa_fmadd_w(_sum3, _r0a, _k04);
-                    _sum0 = __msa_fmadd_w(_sum0, _r05, _k05);
-                    _sum1 = __msa_fmadd_w(_sum1, _r07, _k05);
-                    _sum2 = __msa_fmadd_w(_sum2, _r09, _k05);
-                    _sum3 = __msa_fmadd_w(_sum3, _r0b, _k05);
-                    _sum0 = __msa_fmadd_w(_sum0, _r06, _k06);
-                    _sum1 = __msa_fmadd_w(_sum1, _r08, _k06);
-                    _sum2 = __msa_fmadd_w(_sum2, _r0a, _k06);
-                    _sum3 = __msa_fmadd_w(_sum3, _r0c, _k06);
-
-                    v4f32 _k10 = (v4f32)__msa_ld_w(kptr, 0);
-                    v4f32 _k11 = (v4f32)__msa_ld_w(kptr + 4, 0);
-                    v4f32 _k12 = (v4f32)__msa_ld_w(kptr + 4 * 2, 0);
-                    v4f32 _k13 = (v4f32)__msa_ld_w(kptr + 4 * 3, 0);
-                    v4f32 _k14 = (v4f32)__msa_ld_w(kptr + 4 * 4, 0);
-                    v4f32 _k15 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0);
-                    v4f32 _k16 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    v4i32 _r1 = __msa_ld_w(r1, 0);
-                    v4i32 _r1n = __msa_ld_w(r1 + 4, 0);
-                    v4i32 _r1nn = __msa_ld_w(r1 + 8, 0);
-
-                    v4f32 _r10 = (v4f32)__msa_splati_w(_r1, 0);
-                    v4f32 _r11 = (v4f32)__msa_splati_w(_r1, 1);
-                    v4f32 _r12 = (v4f32)__msa_splati_w(_r1, 2);
-                    v4f32 _r13 = (v4f32)__msa_splati_w(_r1, 3);
-                    v4f32 _r14 = (v4f32)__msa_splati_w(_r1n, 0);
-                    v4f32 _r15 = (v4f32)__msa_splati_w(_r1n, 1);
-                    v4f32 _r16 = (v4f32)__msa_splati_w(_r1n, 2);
-                    v4f32 _r17 = (v4f32)__msa_splati_w(_r1n, 3);
-                    v4f32 _r18 = (v4f32)__msa_splati_w(_r1nn, 0);
-                    v4f32 _r19 = (v4f32)__msa_splati_w(_r1nn, 1);
-                    v4f32 _r1a = (v4f32)__msa_splati_w(_r1nn, 2);
-                    v4f32 _r1b = (v4f32)__msa_splati_w(_r1nn, 3);
-                    v4f32 _r1c = __msa_fill_w_f32(r1[12]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r10, _k10);
-                    _sum1 = __msa_fmadd_w(_sum1, _r12, _k10);
-                    _sum2 = __msa_fmadd_w(_sum2, _r14, _k10);
-                    _sum3 = __msa_fmadd_w(_sum3, _r16, _k10);
-                    _sum0 = __msa_fmadd_w(_sum0, _r11, _k11);
-                    _sum1 = __msa_fmadd_w(_sum1, _r13, _k11);
-                    _sum2 = __msa_fmadd_w(_sum2, _r15, _k11);
-                    _sum3 = __msa_fmadd_w(_sum3, _r17, _k11);
-                    _sum0 = __msa_fmadd_w(_sum0, _r12, _k12);
-                    _sum1 = __msa_fmadd_w(_sum1, _r14, _k12);
-                    _sum2 = __msa_fmadd_w(_sum2, _r16, _k12);
-                    _sum3 = __msa_fmadd_w(_sum3, _r18, _k12);
-                    _sum0 = __msa_fmadd_w(_sum0, _r13, _k13);
-                    _sum1 = __msa_fmadd_w(_sum1, _r15, _k13);
-                    _sum2 = __msa_fmadd_w(_sum2, _r17, _k13);
-                    _sum3 = __msa_fmadd_w(_sum3, _r19, _k13);
-                    _sum0 = __msa_fmadd_w(_sum0, _r14, _k14);
-                    _sum1 = __msa_fmadd_w(_sum1, _r16, _k14);
-                    _sum2 = __msa_fmadd_w(_sum2, _r18, _k14);
-                    _sum3 = __msa_fmadd_w(_sum3, _r1a, _k14);
-                    _sum0 = __msa_fmadd_w(_sum0, _r15, _k15);
-                    _sum1 = __msa_fmadd_w(_sum1, _r17, _k15);
-                    _sum2 = __msa_fmadd_w(_sum2, _r19, _k15);
-                    _sum3 = __msa_fmadd_w(_sum3, _r1b, _k15);
-                    _sum0 = __msa_fmadd_w(_sum0, _r16, _k16);
-                    _sum1 = __msa_fmadd_w(_sum1, _r18, _k16);
-                    _sum2 = __msa_fmadd_w(_sum2, _r1a, _k16);
-                    _sum3 = __msa_fmadd_w(_sum3, _r1c, _k16);
-
-                    v4f32 _k20 = (v4f32)__msa_ld_w(kptr, 0);
-                    v4f32 _k21 = (v4f32)__msa_ld_w(kptr + 4, 0);
-                    v4f32 _k22 = (v4f32)__msa_ld_w(kptr + 4 * 2, 0);
-                    v4f32 _k23 = (v4f32)__msa_ld_w(kptr + 4 * 3, 0);
-                    v4f32 _k24 = (v4f32)__msa_ld_w(kptr + 4 * 4, 0);
-                    v4f32 _k25 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0);
-                    v4f32 _k26 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    v4i32 _r2 = __msa_ld_w(r2, 0);
-                    v4i32 _r2n = __msa_ld_w(r2 + 4, 0);
-                    v4i32 _r2nn = __msa_ld_w(r2 + 8, 0);
-
-                    v4f32 _r20 = (v4f32)__msa_splati_w(_r2, 0);
-                    v4f32 _r21 = (v4f32)__msa_splati_w(_r2, 1);
-                    v4f32 _r22 = (v4f32)__msa_splati_w(_r2, 2);
-                    v4f32 _r23 = (v4f32)__msa_splati_w(_r2, 3);
-                    v4f32 _r24 = (v4f32)__msa_splati_w(_r2n, 0);
-                    v4f32 _r25 = (v4f32)__msa_splati_w(_r2n, 1);
-                    v4f32 _r26 = (v4f32)__msa_splati_w(_r2n, 2);
-                    v4f32 _r27 = (v4f32)__msa_splati_w(_r2n, 3);
-                    v4f32 _r28 = (v4f32)__msa_splati_w(_r2nn, 0);
-                    v4f32 _r29 = (v4f32)__msa_splati_w(_r2nn, 1);
-                    v4f32 _r2a = (v4f32)__msa_splati_w(_r2nn, 2);
-                    v4f32 _r2b = (v4f32)__msa_splati_w(_r2nn, 3);
-                    v4f32 _r2c = __msa_fill_w_f32(r2[12]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r20, _k20);
-                    _sum1 = __msa_fmadd_w(_sum1, _r22, _k20);
-                    _sum2 = __msa_fmadd_w(_sum2, _r24, _k20);
-                    _sum3 = __msa_fmadd_w(_sum3, _r26, _k20);
-                    _sum0 = __msa_fmadd_w(_sum0, _r21, _k21);
-                    _sum1 = __msa_fmadd_w(_sum1, _r23, _k21);
-                    _sum2 = __msa_fmadd_w(_sum2, _r25, _k21);
-                    _sum3 = __msa_fmadd_w(_sum3, _r27, _k21);
-                    _sum0 = __msa_fmadd_w(_sum0, _r22, _k22);
-                    _sum1 = __msa_fmadd_w(_sum1, _r24, _k22);
-                    _sum2 = __msa_fmadd_w(_sum2, _r26, _k22);
-                    _sum3 = __msa_fmadd_w(_sum3, _r28, _k22);
-                    _sum0 = __msa_fmadd_w(_sum0, _r23, _k23);
-                    _sum1 = __msa_fmadd_w(_sum1, _r25, _k23);
-                    _sum2 = __msa_fmadd_w(_sum2, _r27, _k23);
-                    _sum3 = __msa_fmadd_w(_sum3, _r29, _k23);
-                    _sum0 = __msa_fmadd_w(_sum0, _r24, _k24);
-                    _sum1 = __msa_fmadd_w(_sum1, _r26, _k24);
-                    _sum2 = __msa_fmadd_w(_sum2, _r28, _k24);
-                    _sum3 = __msa_fmadd_w(_sum3, _r2a, _k24);
-                    _sum0 = __msa_fmadd_w(_sum0, _r25, _k25);
-                    _sum1 = __msa_fmadd_w(_sum1, _r27, _k25);
-                    _sum2 = __msa_fmadd_w(_sum2, _r29, _k25);
-                    _sum3 = __msa_fmadd_w(_sum3, _r2b, _k25);
-                    _sum0 = __msa_fmadd_w(_sum0, _r26, _k26);
-                    _sum1 = __msa_fmadd_w(_sum1, _r28, _k26);
-                    _sum2 = __msa_fmadd_w(_sum2, _r2a, _k26);
-                    _sum3 = __msa_fmadd_w(_sum3, _r2c, _k26);
-
-                    v4f32 _k30 = (v4f32)__msa_ld_w(kptr, 0);
-                    v4f32 _k31 = (v4f32)__msa_ld_w(kptr + 4, 0);
-                    v4f32 _k32 = (v4f32)__msa_ld_w(kptr + 4 * 2, 0);
-                    v4f32 _k33 = (v4f32)__msa_ld_w(kptr + 4 * 3, 0);
-                    v4f32 _k34 = (v4f32)__msa_ld_w(kptr + 4 * 4, 0);
-                    v4f32 _k35 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0);
-                    v4f32 _k36 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    v4i32 _r3 = __msa_ld_w(r3, 0);
-                    v4i32 _r3n = __msa_ld_w(r3 + 4, 0);
-                    v4i32 _r3nn = __msa_ld_w(r3 + 8, 0);
-
-                    v4f32 _r30 = (v4f32)__msa_splati_w(_r3, 0);
-                    v4f32 _r31 = (v4f32)__msa_splati_w(_r3, 1);
-                    v4f32 _r32 = (v4f32)__msa_splati_w(_r3, 2);
-                    v4f32 _r33 = (v4f32)__msa_splati_w(_r3, 3);
-                    v4f32 _r34 = (v4f32)__msa_splati_w(_r3n, 0);
-                    v4f32 _r35 = (v4f32)__msa_splati_w(_r3n, 1);
-                    v4f32 _r36 = (v4f32)__msa_splati_w(_r3n, 2);
-                    v4f32 _r37 = (v4f32)__msa_splati_w(_r3n, 3);
-                    v4f32 _r38 = (v4f32)__msa_splati_w(_r3nn, 0);
-                    v4f32 _r39 = (v4f32)__msa_splati_w(_r3nn, 1);
-                    v4f32 _r3a = (v4f32)__msa_splati_w(_r3nn, 2);
-                    v4f32 _r3b = (v4f32)__msa_splati_w(_r3nn, 3);
-                    v4f32 _r3c = __msa_fill_w_f32(r3[12]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r30, _k30);
-                    _sum1 = __msa_fmadd_w(_sum1, _r32, _k30);
-                    _sum2 = __msa_fmadd_w(_sum2, _r34, _k30);
-                    _sum3 = __msa_fmadd_w(_sum3, _r36, _k30);
-                    _sum0 = __msa_fmadd_w(_sum0, _r31, _k31);
-                    _sum1 = __msa_fmadd_w(_sum1, _r33, _k31);
-                    _sum2 = __msa_fmadd_w(_sum2, _r35, _k31);
-                    _sum3 = __msa_fmadd_w(_sum3, _r37, _k31);
-                    _sum0 = __msa_fmadd_w(_sum0, _r32, _k32);
-                    _sum1 = __msa_fmadd_w(_sum1, _r34, _k32);
-                    _sum2 = __msa_fmadd_w(_sum2, _r36, _k32);
-                    _sum3 = __msa_fmadd_w(_sum3, _r38, _k32);
-                    _sum0 = __msa_fmadd_w(_sum0, _r33, _k33);
-                    _sum1 = __msa_fmadd_w(_sum1, _r35, _k33);
-                    _sum2 = __msa_fmadd_w(_sum2, _r37, _k33);
-                    _sum3 = __msa_fmadd_w(_sum3, _r39, _k33);
-                    _sum0 = __msa_fmadd_w(_sum0, _r34, _k34);
-                    _sum1 = __msa_fmadd_w(_sum1, _r36, _k34);
-                    _sum2 = __msa_fmadd_w(_sum2, _r38, _k34);
-                    _sum3 = __msa_fmadd_w(_sum3, _r3a, _k34);
-                    _sum0 = __msa_fmadd_w(_sum0, _r35, _k35);
-                    _sum1 = __msa_fmadd_w(_sum1, _r37, _k35);
-                    _sum2 = __msa_fmadd_w(_sum2, _r39, _k35);
-                    _sum3 = __msa_fmadd_w(_sum3, _r3b, _k35);
-                    _sum0 = __msa_fmadd_w(_sum0, _r36, _k36);
-                    _sum1 = __msa_fmadd_w(_sum1, _r38, _k36);
-                    _sum2 = __msa_fmadd_w(_sum2, _r3a, _k36);
-                    _sum3 = __msa_fmadd_w(_sum3, _r3c, _k36);
-
-                    v4f32 _k40 = (v4f32)__msa_ld_w(kptr, 0);
-                    v4f32 _k41 = (v4f32)__msa_ld_w(kptr + 4, 0);
-                    v4f32 _k42 = (v4f32)__msa_ld_w(kptr + 4 * 2, 0);
-                    v4f32 _k43 = (v4f32)__msa_ld_w(kptr + 4 * 3, 0);
-                    v4f32 _k44 = (v4f32)__msa_ld_w(kptr + 4 * 4, 0);
-                    v4f32 _k45 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0);
-                    v4f32 _k46 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    v4i32 _r4 = __msa_ld_w(r4, 0);
-                    v4i32 _r4n = __msa_ld_w(r4 + 4, 0);
-                    v4i32 _r4nn = __msa_ld_w(r4 + 8, 0);
-
-                    v4f32 _r40 = (v4f32)__msa_splati_w(_r4, 0);
-                    v4f32 _r41 = (v4f32)__msa_splati_w(_r4, 1);
-                    v4f32 _r42 = (v4f32)__msa_splati_w(_r4, 2);
-                    v4f32 _r43 = (v4f32)__msa_splati_w(_r4, 3);
-                    v4f32 _r44 = (v4f32)__msa_splati_w(_r4n, 0);
-                    v4f32 _r45 = (v4f32)__msa_splati_w(_r4n, 1);
-                    v4f32 _r46 = (v4f32)__msa_splati_w(_r4n, 2);
-                    v4f32 _r47 = (v4f32)__msa_splati_w(_r4n, 3);
-                    v4f32 _r48 = (v4f32)__msa_splati_w(_r4nn, 0);
-                    v4f32 _r49 = (v4f32)__msa_splati_w(_r4nn, 1);
-                    v4f32 _r4a = (v4f32)__msa_splati_w(_r4nn, 2);
-                    v4f32 _r4b = (v4f32)__msa_splati_w(_r4nn, 3);
-                    v4f32 _r4c = __msa_fill_w_f32(r4[12]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r40, _k40);
-                    _sum1 = __msa_fmadd_w(_sum1, _r42, _k40);
-                    _sum2 = __msa_fmadd_w(_sum2, _r44, _k40);
-                    _sum3 = __msa_fmadd_w(_sum3, _r46, _k40);
-                    _sum0 = __msa_fmadd_w(_sum0, _r41, _k41);
-                    _sum1 = __msa_fmadd_w(_sum1, _r43, _k41);
-                    _sum2 = __msa_fmadd_w(_sum2, _r45, _k41);
-                    _sum3 = __msa_fmadd_w(_sum3, _r47, _k41);
-                    _sum0 = __msa_fmadd_w(_sum0, _r42, _k42);
-                    _sum1 = __msa_fmadd_w(_sum1, _r44, _k42);
-                    _sum2 = __msa_fmadd_w(_sum2, _r46, _k42);
-                    _sum3 = __msa_fmadd_w(_sum3, _r48, _k42);
-                    _sum0 = __msa_fmadd_w(_sum0, _r43, _k43);
-                    _sum1 = __msa_fmadd_w(_sum1, _r45, _k43);
-                    _sum2 = __msa_fmadd_w(_sum2, _r47, _k43);
-                    _sum3 = __msa_fmadd_w(_sum3, _r49, _k43);
-                    _sum0 = __msa_fmadd_w(_sum0, _r44, _k44);
-                    _sum1 = __msa_fmadd_w(_sum1, _r46, _k44);
-                    _sum2 = __msa_fmadd_w(_sum2, _r48, _k44);
-                    _sum3 = __msa_fmadd_w(_sum3, _r4a, _k44);
-                    _sum0 = __msa_fmadd_w(_sum0, _r45, _k45);
-                    _sum1 = __msa_fmadd_w(_sum1, _r47, _k45);
-                    _sum2 = __msa_fmadd_w(_sum2, _r49, _k45);
-                    _sum3 = __msa_fmadd_w(_sum3, _r4b, _k45);
-                    _sum0 = __msa_fmadd_w(_sum0, _r46, _k46);
-                    _sum1 = __msa_fmadd_w(_sum1, _r48, _k46);
-                    _sum2 = __msa_fmadd_w(_sum2, _r4a, _k46);
-                    _sum3 = __msa_fmadd_w(_sum3, _r4c, _k46);
-
-                    v4f32 _k50 = (v4f32)__msa_ld_w(kptr, 0);
-                    v4f32 _k51 = (v4f32)__msa_ld_w(kptr + 4, 0);
-                    v4f32 _k52 = (v4f32)__msa_ld_w(kptr + 4 * 2, 0);
-                    v4f32 _k53 = (v4f32)__msa_ld_w(kptr + 4 * 3, 0);
-                    v4f32 _k54 = (v4f32)__msa_ld_w(kptr + 4 * 4, 0);
-                    v4f32 _k55 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0);
-                    v4f32 _k56 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    v4i32 _r5 = __msa_ld_w(r5, 0);
-                    v4i32 _r5n = __msa_ld_w(r5 + 4, 0);
-                    v4i32 _r5nn = __msa_ld_w(r5 + 8, 0);
-
-                    v4f32 _r50 = (v4f32)__msa_splati_w(_r5, 0);
-                    v4f32 _r51 = (v4f32)__msa_splati_w(_r5, 1);
-                    v4f32 _r52 = (v4f32)__msa_splati_w(_r5, 2);
-                    v4f32 _r53 = (v4f32)__msa_splati_w(_r5, 3);
-                    v4f32 _r54 = (v4f32)__msa_splati_w(_r5n, 0);
-                    v4f32 _r55 = (v4f32)__msa_splati_w(_r5n, 1);
-                    v4f32 _r56 = (v4f32)__msa_splati_w(_r5n, 2);
-                    v4f32 _r57 = (v4f32)__msa_splati_w(_r5n, 3);
-                    v4f32 _r58 = (v4f32)__msa_splati_w(_r5nn, 0);
-                    v4f32 _r59 = (v4f32)__msa_splati_w(_r5nn, 1);
-                    v4f32 _r5a = (v4f32)__msa_splati_w(_r5nn, 2);
-                    v4f32 _r5b = (v4f32)__msa_splati_w(_r5nn, 3);
-                    v4f32 _r5c = __msa_fill_w_f32(r5[12]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r50, _k50);
-                    _sum1 = __msa_fmadd_w(_sum1, _r52, _k50);
-                    _sum2 = __msa_fmadd_w(_sum2, _r54, _k50);
-                    _sum3 = __msa_fmadd_w(_sum3, _r56, _k50);
-                    _sum0 = __msa_fmadd_w(_sum0, _r51, _k51);
-                    _sum1 = __msa_fmadd_w(_sum1, _r53, _k51);
-                    _sum2 = __msa_fmadd_w(_sum2, _r55, _k51);
-                    _sum3 = __msa_fmadd_w(_sum3, _r57, _k51);
-                    _sum0 = __msa_fmadd_w(_sum0, _r52, _k52);
-                    _sum1 = __msa_fmadd_w(_sum1, _r54, _k52);
-                    _sum2 = __msa_fmadd_w(_sum2, _r56, _k52);
-                    _sum3 = __msa_fmadd_w(_sum3, _r58, _k52);
-                    _sum0 = __msa_fmadd_w(_sum0, _r53, _k53);
-                    _sum1 = __msa_fmadd_w(_sum1, _r55, _k53);
-                    _sum2 = __msa_fmadd_w(_sum2, _r57, _k53);
-                    _sum3 = __msa_fmadd_w(_sum3, _r59, _k53);
-                    _sum0 = __msa_fmadd_w(_sum0, _r54, _k54);
-                    _sum1 = __msa_fmadd_w(_sum1, _r56, _k54);
-                    _sum2 = __msa_fmadd_w(_sum2, _r58, _k54);
-                    _sum3 = __msa_fmadd_w(_sum3, _r5a, _k54);
-                    _sum0 = __msa_fmadd_w(_sum0, _r55, _k55);
-                    _sum1 = __msa_fmadd_w(_sum1, _r57, _k55);
-                    _sum2 = __msa_fmadd_w(_sum2, _r59, _k55);
-                    _sum3 = __msa_fmadd_w(_sum3, _r5b, _k55);
-                    _sum0 = __msa_fmadd_w(_sum0, _r56, _k56);
-                    _sum1 = __msa_fmadd_w(_sum1, _r58, _k56);
-                    _sum2 = __msa_fmadd_w(_sum2, _r5a, _k56);
-                    _sum3 = __msa_fmadd_w(_sum3, _r5c, _k56);
-
-                    v4f32 _k60 = (v4f32)__msa_ld_w(kptr, 0);
-                    v4f32 _k61 = (v4f32)__msa_ld_w(kptr + 4, 0);
-                    v4f32 _k62 = (v4f32)__msa_ld_w(kptr + 4 * 2, 0);
-                    v4f32 _k63 = (v4f32)__msa_ld_w(kptr + 4 * 3, 0);
-                    v4f32 _k64 = (v4f32)__msa_ld_w(kptr + 4 * 4, 0);
-                    v4f32 _k65 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0);
-                    v4f32 _k66 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0);
-
-                    kptr -= 4 * 42;
-
-                    v4i32 _r6 = __msa_ld_w(r6, 0);
-                    v4i32 _r6n = __msa_ld_w(r6 + 4, 0);
-                    v4i32 _r6nn = __msa_ld_w(r6 + 8, 0);
-
-                    v4f32 _r60 = (v4f32)__msa_splati_w(_r6, 0);
-                    v4f32 _r61 = (v4f32)__msa_splati_w(_r6, 1);
-                    v4f32 _r62 = (v4f32)__msa_splati_w(_r6, 2);
-                    v4f32 _r63 = (v4f32)__msa_splati_w(_r6, 3);
-                    v4f32 _r64 = (v4f32)__msa_splati_w(_r6n, 0);
-                    v4f32 _r65 = (v4f32)__msa_splati_w(_r6n, 1);
-                    v4f32 _r66 = (v4f32)__msa_splati_w(_r6n, 2);
-                    v4f32 _r67 = (v4f32)__msa_splati_w(_r6n, 3);
-                    v4f32 _r68 = (v4f32)__msa_splati_w(_r6nn, 0);
-                    v4f32 _r69 = (v4f32)__msa_splati_w(_r6nn, 1);
-                    v4f32 _r6a = (v4f32)__msa_splati_w(_r6nn, 2);
-                    v4f32 _r6b = (v4f32)__msa_splati_w(_r6nn, 3);
-                    v4f32 _r6c = __msa_fill_w_f32(r6[12]);
-
-                    _sum0 = __msa_fmadd_w(_sum0, _r60, _k60);
-                    _sum1 = __msa_fmadd_w(_sum1, _r62, _k60);
-                    _sum2 = __msa_fmadd_w(_sum2, _r64, _k60);
-                    _sum3 = __msa_fmadd_w(_sum3, _r66, _k60);
-                    _sum0 = __msa_fmadd_w(_sum0, _r61, _k61);
-                    _sum1 = __msa_fmadd_w(_sum1, _r63, _k61);
-                    _sum2 = __msa_fmadd_w(_sum2, _r65, _k61);
-                    _sum3 = __msa_fmadd_w(_sum3, _r67, _k61);
-                    _sum0 = __msa_fmadd_w(_sum0, _r62, _k62);
-                    _sum1 = __msa_fmadd_w(_sum1, _r64, _k62);
-                    _sum2 = __msa_fmadd_w(_sum2, _r66, _k62);
-                    _sum3 = __msa_fmadd_w(_sum3, _r68, _k62);
-                    _sum0 = __msa_fmadd_w(_sum0, _r63, _k63);
-                    _sum1 = __msa_fmadd_w(_sum1, _r65, _k63);
-                    _sum2 = __msa_fmadd_w(_sum2, _r67, _k63);
-                    _sum3 = __msa_fmadd_w(_sum3, _r69, _k63);
-                    _sum0 = __msa_fmadd_w(_sum0, _r64, _k64);
-                    _sum1 = __msa_fmadd_w(_sum1, _r66, _k64);
-                    _sum2 = __msa_fmadd_w(_sum2, _r68, _k64);
-                    _sum3 = __msa_fmadd_w(_sum3, _r6a, _k64);
-                    _sum0 = __msa_fmadd_w(_sum0, _r65, _k65);
-                    _sum1 = __msa_fmadd_w(_sum1, _r67, _k65);
-                    _sum2 = __msa_fmadd_w(_sum2, _r69, _k65);
-                    _sum3 = __msa_fmadd_w(_sum3, _r6b, _k65);
-                    _sum0 = __msa_fmadd_w(_sum0, _r66, _k66);
-                    _sum1 = __msa_fmadd_w(_sum1, _r68, _k66);
-                    _sum2 = __msa_fmadd_w(_sum2, _r6a, _k66);
-                    _sum3 = __msa_fmadd_w(_sum3, _r6c, _k66);
-
-                    __msa_st_w((v4i32)_sum0, outptr0, 0);
-                    __msa_st_w((v4i32)_sum1, outptr0 + 4, 0);
-                    __msa_st_w((v4i32)_sum2, outptr0 + 4 * 2, 0);
-                    __msa_st_w((v4i32)_sum3, outptr0 + 4 * 3, 0);
-
-                    outptr0 += 4 * 4;
-
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                    r3 += 8;
-                    r4 += 8;
-                    r5 += 8;
-                    r6 += 8;
-                }
-                for (; j < outw; j++)
-                {
-                    v4f32 _sum0 = (v4f32)__msa_ld_w(outptr0, 0);
-
-                    v4f32 _k00 = (v4f32)__msa_ld_w(kptr, 0);
-                    v4f32 _k01 = (v4f32)__msa_ld_w(kptr + 4, 0);
-                    v4f32 _k02 = (v4f32)__msa_ld_w(kptr + 4 * 2, 0);
-                    v4f32 _k03 = (v4f32)__msa_ld_w(kptr + 4 * 3, 0);
-                    v4f32 _k04 = (v4f32)__msa_ld_w(kptr + 4 * 4, 0);
-                    v4f32 _k05 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0);
-                    v4f32 _k06 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    v4i32 _r0 = __msa_ld_w(r0, 0);
-                    v4i32 _r0n = __msa_ld_w(r0 + 4, 0);
-
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r0, 0), _k00);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r0, 1), _k01);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r0, 2), _k02);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r0, 3), _k03);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r0n, 0), _k04);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r0n, 1), _k05);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r0n, 2), _k06);
-
-                    v4f32 _k10 = (v4f32)__msa_ld_w(kptr, 0);
-                    v4f32 _k11 = (v4f32)__msa_ld_w(kptr + 4, 0);
-                    v4f32 _k12 = (v4f32)__msa_ld_w(kptr + 4 * 2, 0);
-                    v4f32 _k13 = (v4f32)__msa_ld_w(kptr + 4 * 3, 0);
-                    v4f32 _k14 = (v4f32)__msa_ld_w(kptr + 4 * 4, 0);
-                    v4f32 _k15 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0);
-                    v4f32 _k16 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    v4i32 _r1 = __msa_ld_w(r1, 0);
-                    v4i32 _r1n = __msa_ld_w(r1 + 4, 0);
-
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r1, 0), _k10);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r1, 1), _k11);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r1, 2), _k12);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r1, 3), _k13);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r1n, 0), _k14);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r1n, 1), _k15);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r1n, 2), _k16);
-
-                    v4f32 _k20 = (v4f32)__msa_ld_w(kptr, 0);
-                    v4f32 _k21 = (v4f32)__msa_ld_w(kptr + 4, 0);
-                    v4f32 _k22 = (v4f32)__msa_ld_w(kptr + 4 * 2, 0);
-                    v4f32 _k23 = (v4f32)__msa_ld_w(kptr + 4 * 3, 0);
-                    v4f32 _k24 = (v4f32)__msa_ld_w(kptr + 4 * 4, 0);
-                    v4f32 _k25 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0);
-                    v4f32 _k26 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    v4i32 _r2 = __msa_ld_w(r2, 0);
-                    v4i32 _r2n = __msa_ld_w(r2 + 4, 0);
-
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r2, 0), _k20);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r2, 1), _k21);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r2, 2), _k22);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r2, 3), _k23);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r2n, 0), _k24);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r2n, 1), _k25);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r2n, 2), _k26);
-
-                    v4f32 _k30 = (v4f32)__msa_ld_w(kptr, 0);
-                    v4f32 _k31 = (v4f32)__msa_ld_w(kptr + 4, 0);
-                    v4f32 _k32 = (v4f32)__msa_ld_w(kptr + 4 * 2, 0);
-                    v4f32 _k33 = (v4f32)__msa_ld_w(kptr + 4 * 3, 0);
-                    v4f32 _k34 = (v4f32)__msa_ld_w(kptr + 4 * 4, 0);
-                    v4f32 _k35 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0);
-                    v4f32 _k36 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    v4i32 _r3 = __msa_ld_w(r3, 0);
-                    v4i32 _r3n = __msa_ld_w(r3 + 4, 0);
-
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r3, 0), _k30);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r3, 1), _k31);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r3, 2), _k32);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r3, 3), _k33);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r3n, 0), _k34);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r3n, 1), _k35);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r3n, 2), _k36);
-
-                    v4f32 _k40 = (v4f32)__msa_ld_w(kptr, 0);
-                    v4f32 _k41 = (v4f32)__msa_ld_w(kptr + 4, 0);
-                    v4f32 _k42 = (v4f32)__msa_ld_w(kptr + 4 * 2, 0);
-                    v4f32 _k43 = (v4f32)__msa_ld_w(kptr + 4 * 3, 0);
-                    v4f32 _k44 = (v4f32)__msa_ld_w(kptr + 4 * 4, 0);
-                    v4f32 _k45 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0);
-                    v4f32 _k46 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    v4i32 _r4 = __msa_ld_w(r4, 0);
-                    v4i32 _r4n = __msa_ld_w(r4 + 4, 0);
-
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r4, 0), _k40);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r4, 1), _k41);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r4, 2), _k42);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r4, 3), _k43);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r4n, 0), _k44);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r4n, 1), _k45);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r4n, 2), _k46);
-
-                    v4f32 _k50 = (v4f32)__msa_ld_w(kptr, 0);
-                    v4f32 _k51 = (v4f32)__msa_ld_w(kptr + 4, 0);
-                    v4f32 _k52 = (v4f32)__msa_ld_w(kptr + 4 * 2, 0);
-                    v4f32 _k53 = (v4f32)__msa_ld_w(kptr + 4 * 3, 0);
-                    v4f32 _k54 = (v4f32)__msa_ld_w(kptr + 4 * 4, 0);
-                    v4f32 _k55 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0);
-                    v4f32 _k56 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0);
-
-                    kptr += 4 * 7;
-
-                    v4i32 _r5 = __msa_ld_w(r5, 0);
-                    v4i32 _r5n = __msa_ld_w(r5 + 4, 0);
-
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r5, 0), _k50);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r5, 1), _k51);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r5, 2), _k52);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r5, 3), _k53);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r5n, 0), _k54);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r5n, 1), _k55);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r5n, 2), _k56);
-
-                    v4f32 _k60 = (v4f32)__msa_ld_w(kptr, 0);
-                    v4f32 _k61 = (v4f32)__msa_ld_w(kptr + 4, 0);
-                    v4f32 _k62 = (v4f32)__msa_ld_w(kptr + 4 * 2, 0);
-                    v4f32 _k63 = (v4f32)__msa_ld_w(kptr + 4 * 3, 0);
-                    v4f32 _k64 = (v4f32)__msa_ld_w(kptr + 4 * 4, 0);
-                    v4f32 _k65 = (v4f32)__msa_ld_w(kptr + 4 * 5, 0);
-                    v4f32 _k66 = (v4f32)__msa_ld_w(kptr + 4 * 6, 0);
-
-                    kptr -= 4 * 42;
-
-                    v4i32 _r6 = __msa_ld_w(r6, 0);
-                    v4i32 _r6n = __msa_ld_w(r6 + 4, 0);
-
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r6, 0), _k60);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r6, 1), _k61);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r6, 2), _k62);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r6, 3), _k63);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r6n, 0), _k64);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r6n, 1), _k65);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_r6n, 2), _k66);
-
-                    __msa_st_w((v4i32)_sum0, outptr0, 0);
-
-                    outptr0 += 4;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    r3 += 2;
-                    r4 += 2;
-                    r5 += 2;
-                    r6 += 2;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-                r4 += tailstep;
-                r5 += tailstep;
-                r6 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_int8.h
deleted file mode 100644
index 6b56d11e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_int8.h
+++ /dev/null
@@ -1,82 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_int8(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_int8, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        int* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                int sum = 0;
-
-                //                 const signed char* kptr = weight_data_int8.channel(p);
-                const signed char* kptr = (const signed char*)weight_data_int8 + maxk * channels * p;
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const signed char* sptr = m.row<signed char>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        signed char val = sptr[space_ofs[k]];
-                        signed char w = kptr[k];
-                        sum += val * w;
-                    }
-
-                    kptr += maxk;
-                }
-
-                outptr[j] = sum;
-            }
-
-            outptr += outw;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_mips.h
deleted file mode 100644
index 8401c6df..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_mips.h
+++ /dev/null
@@ -1,56 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTION_MIPS_H
-#define LAYER_CONVOLUTION_MIPS_H
-
-#include "convolution.h"
-
-namespace ncnn {
-
-class Convolution_mips : public Convolution
-{
-public:
-    Convolution_mips();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if NCNN_INT8
-    int create_pipeline_int8_mips(const Option& opt);
-    int forward_int8_mips(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Layer* activation;
-
-    Mat weight_data_tm;
-    Mat weight_sgemm_data;
-    Mat weight_winograd23_data;
-    Mat weight_winograd43_data;
-    Mat weight_winograd63_data;
-
-#if NCNN_INT8
-    Mat scale_in_data;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTION_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack1to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack1to4.h
deleted file mode 100644
index 2d72eaf7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack1to4.h
+++ /dev/null
@@ -1,90 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_pack1to4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_pack1ton, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                v4f32 _sum = (v4f32)__msa_fill_w(0);
-
-                if (bias_data_ptr)
-                {
-                    _sum = (v4f32)__msa_ld_w(bias_data_ptr + p * 4, 0);
-                }
-
-                const float* kptr = (const float*)weight_data_pack1ton + maxk * channels * p * 4;
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const float* sptr = m.row(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++) // 29.23
-                    {
-                        v4f32 _val = __msa_fill_w_f32(sptr[space_ofs[k]]);
-                        v4f32 _w = (v4f32)__msa_ld_w(kptr, 0);
-                        _sum = __msa_fmadd_w(_sum, _val, _w);
-
-                        kptr += 4;
-                    }
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params);
-
-                __msa_st_w((v4i32)_sum, outptr + j * 4, 0);
-            }
-
-            outptr += outw * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack1to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack1to4_int8.h
deleted file mode 100644
index b752bda0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack1to4_int8.h
+++ /dev/null
@@ -1,87 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_pack1to4_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_int8, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        int* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                v4i32 _sum = __msa_fill_w(0);
-
-                const signed char* kptr = weight_data_int8.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const signed char* sptr = m.row<const signed char>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        v8i16 _val = __msa_fill_h((short)sptr[space_ofs[k]]);
-
-                        v16i8 _w = __msa_ld_b(kptr, 0);
-                        v8i16 _w16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_w, 0), _w);
-
-                        v8i16 _s0 = __msa_mulv_h(_val, _w16);
-                        v4i32 _s032 = (v4i32)__msa_ilvr_h(__msa_clti_s_h(_s0, 0), _s0);
-
-                        _sum = __msa_addv_w(_sum, _s032);
-
-                        kptr += 4;
-                    }
-                }
-
-                __msa_st_w(_sum, outptr + j * 4, 0);
-            }
-
-            outptr += outw * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack4.h
deleted file mode 100644
index 2da1a690..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack4.h
+++ /dev/null
@@ -1,102 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_pack4, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                v4f32 _sum = (v4f32)__msa_fill_w(0);
-
-                if (bias_data_ptr)
-                {
-                    _sum = (v4f32)__msa_ld_w(bias_data_ptr + p * 4, 0);
-                }
-
-                const float* kptr = (const float*)weight_data_pack4.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const float* sptr = m.row(i * stride_h) + j * stride_w * 4;
-
-                    for (int k = 0; k < maxk; k++) // 29.23
-                    {
-                        const float* slptr = sptr + space_ofs[k] * 4;
-
-                        v4f32 _val0 = __msa_fill_w_f32(slptr[0]);
-                        v4f32 _val1 = __msa_fill_w_f32(slptr[1]);
-                        v4f32 _val2 = __msa_fill_w_f32(slptr[2]);
-                        v4f32 _val3 = __msa_fill_w_f32(slptr[3]);
-
-                        v4f32 _w0 = (v4f32)__msa_ld_w(kptr, 0);
-                        v4f32 _w1 = (v4f32)__msa_ld_w(kptr + 4, 0);
-                        v4f32 _w2 = (v4f32)__msa_ld_w(kptr + 8, 0);
-                        v4f32 _w3 = (v4f32)__msa_ld_w(kptr + 12, 0);
-
-                        _sum = __msa_fmadd_w(_sum, _val0, _w0);
-                        _sum = __msa_fmadd_w(_sum, _val1, _w1);
-                        _sum = __msa_fmadd_w(_sum, _val2, _w2);
-                        _sum = __msa_fmadd_w(_sum, _val3, _w3);
-
-                        kptr += 16;
-                    }
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params);
-
-                __msa_st_w((v4i32)_sum, outptr + j * 4, 0);
-            }
-
-            outptr += outw * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack4to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack4to1.h
deleted file mode 100644
index 8269860b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack4to1.h
+++ /dev/null
@@ -1,94 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_pack4to1_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_pack4to1, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                v4f32 _sum = (v4f32)__msa_fill_w(0);
-
-                const float* kptr = (const float*)weight_data_pack4to1.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const float* sptr = m.row(i * stride_h) + j * stride_w * 4;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        v4f32 _val = (v4f32)__msa_ld_w(sptr + space_ofs[k] * 4, 0);
-                        v4f32 _w = (v4f32)__msa_ld_w(kptr, 0);
-                        _sum = __msa_fmadd_w(_sum, _val, _w);
-
-                        kptr += 4;
-                    }
-                }
-
-                sum += __msa_reduce_fadd_w(_sum);
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[j] = sum;
-            }
-
-            outptr += outw;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack8to1_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack8to1_int8.h
deleted file mode 100644
index 1f9ad5a6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack8to1_int8.h
+++ /dev/null
@@ -1,87 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_pack8to1_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_int8, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        int* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                v4i32 _sum = __msa_fill_w(0);
-
-                const signed char* kptr = weight_data_int8.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const signed char* sptr = m.row<const signed char>(i * stride_h) + j * stride_w * 8;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        v16i8 _val = __msa_ld_b(sptr + space_ofs[k] * 8, 0);
-                        v8i16 _val16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_val, 0), _val);
-
-                        v16i8 _w = __msa_ld_b(kptr, 0);
-                        v8i16 _w16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_w, 0), _w);
-
-                        v8i16 _s0 = __msa_mulv_h(_val16, _w16);
-
-                        _sum = __msa_addv_w(_sum, __msa_hadd_s_w(_s0, _s0));
-
-                        kptr += 8;
-                    }
-                }
-
-                outptr[j] = __msa_reduce_add_w(_sum);
-            }
-
-            outptr += outw;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack8to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack8to4_int8.h
deleted file mode 100644
index 50636f82..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_pack8to4_int8.h
+++ /dev/null
@@ -1,120 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_pack8to4_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_int8, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        int* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                v4i32 _sum0 = __msa_fill_w(0);
-                v4i32 _sum1 = __msa_fill_w(0);
-                v4i32 _sum2 = __msa_fill_w(0);
-                v4i32 _sum3 = __msa_fill_w(0);
-
-                const signed char* kptr = weight_data_int8.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const signed char* sptr = m.row<signed char>(i * stride_h) + j * stride_w * 8;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        v16i8 _val = __msa_ld_b(sptr + space_ofs[k] * 8, 0);
-                        v8i16 _val16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_val, 0), _val);
-
-                        v16i8 _w01 = __msa_ld_b(kptr, 0);
-                        v16i8 _w23 = __msa_ld_b(kptr + 16, 0);
-                        v16i8 _extw01 = __msa_clti_s_b(_w01, 0);
-                        v16i8 _extw23 = __msa_clti_s_b(_w23, 0);
-                        v8i16 _w0 = (v8i16)__msa_ilvr_b(_extw01, _w01);
-                        v8i16 _w1 = (v8i16)__msa_ilvl_b(_extw01, _w01);
-                        v8i16 _w2 = (v8i16)__msa_ilvr_b(_extw23, _w23);
-                        v8i16 _w3 = (v8i16)__msa_ilvl_b(_extw23, _w23);
-
-                        v8i16 _s0 = __msa_mulv_h(_val16, _w0);
-                        v8i16 _s1 = __msa_mulv_h(_val16, _w1);
-                        v8i16 _s2 = __msa_mulv_h(_val16, _w2);
-                        v8i16 _s3 = __msa_mulv_h(_val16, _w3);
-
-                        _sum0 = __msa_addv_w(_sum0, __msa_hadd_s_w(_s0, _s0));
-                        _sum1 = __msa_addv_w(_sum1, __msa_hadd_s_w(_s1, _s1));
-                        _sum2 = __msa_addv_w(_sum2, __msa_hadd_s_w(_s2, _s2));
-                        _sum3 = __msa_addv_w(_sum3, __msa_hadd_s_w(_s3, _s3));
-
-                        kptr += 32;
-                    }
-                }
-
-                // transpose 4x4
-                {
-                    v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                    _tmp0 = __msa_ilvr_w(_sum1, _sum0);
-                    _tmp1 = __msa_ilvr_w(_sum3, _sum2);
-                    _tmp2 = __msa_ilvl_w(_sum1, _sum0);
-                    _tmp3 = __msa_ilvl_w(_sum3, _sum2);
-                    _sum0 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                    _sum1 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                    _sum2 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                    _sum3 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-                }
-
-                _sum0 = __msa_addv_w(_sum0, _sum1);
-                _sum2 = __msa_addv_w(_sum2, _sum3);
-
-                _sum0 = __msa_addv_w(_sum0, _sum2);
-
-                __msa_st_w(_sum0, outptr + j * 4, 0);
-            }
-
-            outptr += outw * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm.h
deleted file mode 100644
index 2a032491..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm.h
+++ /dev/null
@@ -1,650 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_msa(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    // Mat bottom_im2col(size, maxk, inch, 4u, 1, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    // permute
-    Mat tmp;
-    if (size >= 4)
-        tmp.create(4 * maxk, inch, size / 4 + size % 4, 4u, 1, opt.workspace_allocator);
-    else
-        tmp.create(maxk, inch, size, 4u, 1, opt.workspace_allocator);
-    {
-        int nn_size = size / 4;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = ii * 4;
-
-            float* tmpptr = tmp.channel(i / 4);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-#if __mips_msa
-                    __msa_st_w(__msa_ld_w(img0, 0), tmpptr, 0);
-#else
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img0[1];
-                    tmpptr[2] = img0[2];
-                    tmpptr[3] = img0[3];
-#endif
-                    img0 += size;
-                    tmpptr += 4;
-                }
-            }
-        }
-
-        int remain_size_start = nn_size * 4;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            float* tmpptr = tmp.channel(i / 4 + i % 4);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    img0 += size;
-                    tmpptr += 1;
-                }
-            }
-        }
-    }
-
-#if __mips_msa
-    int nn_outch = outch >> 3;
-    int remain_outch_start = nn_outch << 3;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 8;
-
-        float* outptr0 = top_blob.channel(p);
-        float* outptr1 = top_blob.channel(p + 1);
-        float* outptr2 = top_blob.channel(p + 2);
-        float* outptr3 = top_blob.channel(p + 3);
-        float* outptr4 = top_blob.channel(p + 4);
-        float* outptr5 = top_blob.channel(p + 5);
-        float* outptr6 = top_blob.channel(p + 6);
-        float* outptr7 = top_blob.channel(p + 7);
-
-        const float zeros[8] = {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
-        const float* biasptr = bias ? bias + p : zeros;
-
-        int i = 0;
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 4);
-            const float* kptr = kernel.channel(p / 8);
-
-            int nn = inch * maxk; // inch always > 0
-
-            v4f32 _sum0 = __msa_fill_w_f32(biasptr[0]);
-            v4f32 _sum1 = __msa_fill_w_f32(biasptr[1]);
-            v4f32 _sum2 = __msa_fill_w_f32(biasptr[2]);
-            v4f32 _sum3 = __msa_fill_w_f32(biasptr[3]);
-            v4f32 _sum4 = __msa_fill_w_f32(biasptr[4]);
-            v4f32 _sum5 = __msa_fill_w_f32(biasptr[5]);
-            v4f32 _sum6 = __msa_fill_w_f32(biasptr[6]);
-            v4f32 _sum7 = __msa_fill_w_f32(biasptr[7]);
-
-            for (int q = 0; q < nn; q++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr + 32);
-                v4f32 _val = (v4f32)__msa_ld_w(tmpptr, 0);
-                v4i32 _w0123 = __msa_ld_w(kptr, 0);
-                v4i32 _w4567 = __msa_ld_w(kptr + 4, 0);
-                _sum0 = __msa_fmadd_w(_sum0, _val, (v4f32)__msa_splati_w(_w0123, 0));
-                _sum1 = __msa_fmadd_w(_sum1, _val, (v4f32)__msa_splati_w(_w0123, 1));
-                _sum2 = __msa_fmadd_w(_sum2, _val, (v4f32)__msa_splati_w(_w0123, 2));
-                _sum3 = __msa_fmadd_w(_sum3, _val, (v4f32)__msa_splati_w(_w0123, 3));
-                _sum4 = __msa_fmadd_w(_sum4, _val, (v4f32)__msa_splati_w(_w4567, 0));
-                _sum5 = __msa_fmadd_w(_sum5, _val, (v4f32)__msa_splati_w(_w4567, 1));
-                _sum6 = __msa_fmadd_w(_sum6, _val, (v4f32)__msa_splati_w(_w4567, 2));
-                _sum7 = __msa_fmadd_w(_sum7, _val, (v4f32)__msa_splati_w(_w4567, 3));
-
-                tmpptr += 4;
-                kptr += 8;
-            }
-
-            __msa_st_w((v4i32)_sum0, outptr0, 0);
-            __msa_st_w((v4i32)_sum1, outptr1, 0);
-            __msa_st_w((v4i32)_sum2, outptr2, 0);
-            __msa_st_w((v4i32)_sum3, outptr3, 0);
-            __msa_st_w((v4i32)_sum4, outptr4, 0);
-            __msa_st_w((v4i32)_sum5, outptr5, 0);
-            __msa_st_w((v4i32)_sum6, outptr6, 0);
-            __msa_st_w((v4i32)_sum7, outptr7, 0);
-
-            outptr0 += 4;
-            outptr1 += 4;
-            outptr2 += 4;
-            outptr3 += 4;
-            outptr4 += 4;
-            outptr5 += 4;
-            outptr6 += 4;
-            outptr7 += 4;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 4 + i % 4);
-            const float* kptr = kernel.channel(p / 8);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = biasptr[0];
-            float sum1 = biasptr[1];
-            float sum2 = biasptr[2];
-            float sum3 = biasptr[3];
-            float sum4 = biasptr[4];
-            float sum5 = biasptr[5];
-            float sum6 = biasptr[6];
-            float sum7 = biasptr[7];
-
-            for (int q = 0; q < nn; q++)
-            {
-                sum0 += tmpptr[0] * kptr[0];
-                sum1 += tmpptr[0] * kptr[1];
-                sum2 += tmpptr[0] * kptr[2];
-                sum3 += tmpptr[0] * kptr[3];
-                sum4 += tmpptr[0] * kptr[4];
-                sum5 += tmpptr[0] * kptr[5];
-                sum6 += tmpptr[0] * kptr[6];
-                sum7 += tmpptr[0] * kptr[7];
-                tmpptr++;
-                kptr += 8;
-            }
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-            outptr2[0] = sum2;
-            outptr3[0] = sum3;
-            outptr4[0] = sum4;
-            outptr5[0] = sum5;
-            outptr6[0] = sum6;
-            outptr7[0] = sum7;
-
-            outptr0++;
-            outptr1++;
-            outptr2++;
-            outptr3++;
-            outptr4++;
-            outptr5++;
-            outptr6++;
-            outptr7++;
-        }
-    }
-
-    nn_outch = (outch - remain_outch_start) >> 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = remain_outch_start + pp * 4;
-
-        float* outptr0 = top_blob.channel(p);
-        float* outptr1 = top_blob.channel(p + 1);
-        float* outptr2 = top_blob.channel(p + 2);
-        float* outptr3 = top_blob.channel(p + 3);
-
-        const float zeros[4] = {0.f, 0.f, 0.f, 0.f};
-        const float* biasptr = bias ? bias + p : zeros;
-
-        int i = 0;
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 4);
-            const float* kptr = kernel.channel(p / 8 + (p % 8) / 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            v4f32 _sum0 = __msa_fill_w_f32(biasptr[0]);
-            v4f32 _sum1 = __msa_fill_w_f32(biasptr[1]);
-            v4f32 _sum2 = __msa_fill_w_f32(biasptr[2]);
-            v4f32 _sum3 = __msa_fill_w_f32(biasptr[3]);
-
-            for (int q = 0; q < nn; q++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr + 16);
-                v4f32 _val = (v4f32)__msa_ld_w(tmpptr, 0);
-                v4i32 _w0123 = __msa_ld_w(kptr, 0);
-                _sum0 = __msa_fmadd_w(_sum0, _val, (v4f32)__msa_splati_w(_w0123, 0));
-                _sum1 = __msa_fmadd_w(_sum1, _val, (v4f32)__msa_splati_w(_w0123, 1));
-                _sum2 = __msa_fmadd_w(_sum2, _val, (v4f32)__msa_splati_w(_w0123, 2));
-                _sum3 = __msa_fmadd_w(_sum3, _val, (v4f32)__msa_splati_w(_w0123, 3));
-
-                tmpptr += 4;
-                kptr += 4;
-            }
-
-            __msa_st_w((v4i32)_sum0, outptr0, 0);
-            __msa_st_w((v4i32)_sum1, outptr1, 0);
-            __msa_st_w((v4i32)_sum2, outptr2, 0);
-            __msa_st_w((v4i32)_sum3, outptr3, 0);
-
-            outptr0 += 4;
-            outptr1 += 4;
-            outptr2 += 4;
-            outptr3 += 4;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 4 + i % 4);
-            const float* kptr = kernel.channel(p / 8 + (p % 8) / 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = biasptr[0];
-            float sum1 = biasptr[1];
-            float sum2 = biasptr[2];
-            float sum3 = biasptr[3];
-
-            for (int q = 0; q < nn; q++)
-            {
-                sum0 += tmpptr[0] * kptr[0];
-                sum1 += tmpptr[0] * kptr[1];
-                sum2 += tmpptr[0] * kptr[2];
-                sum3 += tmpptr[0] * kptr[3];
-                tmpptr++;
-                kptr += 4;
-            }
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-            outptr2[0] = sum2;
-            outptr3[0] = sum3;
-
-            outptr0++;
-            outptr1++;
-            outptr2++;
-            outptr3++;
-        }
-    }
-
-    remain_outch_start += nn_outch << 2;
-#else // __mips_msa
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        float* outptr0 = top_blob.channel(p);
-        float* outptr1 = top_blob.channel(p + 1);
-
-        const float zeros[2] = {0.f, 0.f};
-        const float* biasptr = bias ? bias + p : zeros;
-
-        int i = 0;
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 4);
-            const float* kptr = kernel.channel(p / 2);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum00 = biasptr[0];
-            float sum01 = biasptr[0];
-            float sum02 = biasptr[0];
-            float sum03 = biasptr[0];
-            float sum10 = biasptr[1];
-            float sum11 = biasptr[1];
-            float sum12 = biasptr[1];
-            float sum13 = biasptr[1];
-
-            for (int q = 0; q < nn; q++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr + 8);
-                float k0 = kptr[0];
-                float k1 = kptr[1];
-                sum00 += tmpptr[0] * k0;
-                sum01 += tmpptr[1] * k0;
-                sum02 += tmpptr[2] * k0;
-                sum03 += tmpptr[3] * k0;
-                sum10 += tmpptr[0] * k1;
-                sum11 += tmpptr[1] * k1;
-                sum12 += tmpptr[2] * k1;
-                sum13 += tmpptr[3] * k1;
-                tmpptr += 4;
-                kptr += 2;
-            }
-
-            outptr0[0] = sum00;
-            outptr0[1] = sum01;
-            outptr0[2] = sum02;
-            outptr0[3] = sum03;
-            outptr1[0] = sum10;
-            outptr1[1] = sum11;
-            outptr1[2] = sum12;
-            outptr1[3] = sum13;
-
-            outptr0 += 4;
-            outptr1 += 4;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 4 + i % 4);
-            const float* kptr = kernel.channel(p / 2);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = biasptr[0];
-            float sum1 = biasptr[1];
-
-            for (int q = 0; q < nn; q++)
-            {
-                __builtin_prefetch(tmpptr + 4);
-                __builtin_prefetch(kptr + 8);
-                sum0 += tmpptr[0] * kptr[0];
-                sum1 += tmpptr[0] * kptr[1];
-                tmpptr++;
-                kptr += 2;
-            }
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-
-            outptr0++;
-            outptr1++;
-        }
-    }
-#endif // __mips_msa
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        float* outptr0 = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        int i = 0;
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 4);
-#if __mips_msa
-            const float* kptr = kernel.channel(p / 8 + (p % 8) / 4 + p % 4);
-#else
-            const float* kptr = kernel.channel(p / 2 + p % 2);
-#endif
-
-            int nn = inch * maxk; // inch always > 0
-
-#if __mips_msa
-            v4f32 _sum0 = __msa_fill_w_f32(bias0);
-
-            for (int q = 0; q < nn; q++)
-            {
-                _sum0 = __msa_fmadd_w(_sum0, __msa_fill_w_f32(kptr[0]), (v4f32)__msa_ld_w(tmpptr, 0));
-                tmpptr += 4;
-                kptr++;
-            }
-
-            __msa_st_w((v4i32)_sum0, outptr0, 0);
-
-            outptr0 += 4;
-#else
-            float sum0 = bias0;
-            float sum1 = bias0;
-            float sum2 = bias0;
-            float sum3 = bias0;
-
-            for (int q = 0; q < nn; q++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr + 4);
-                sum0 += tmpptr[0] * kptr[0];
-                sum1 += tmpptr[1] * kptr[0];
-                sum2 += tmpptr[2] * kptr[0];
-                sum3 += tmpptr[3] * kptr[0];
-                tmpptr += 4;
-                kptr++;
-            }
-
-            outptr0[0] = sum0;
-            outptr0[1] = sum1;
-            outptr0[2] = sum2;
-            outptr0[3] = sum3;
-
-            outptr0 += 4;
-#endif // __mips_msa
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 4 + i % 4);
-#if __mips_msa
-            const float* kptr = kernel.channel(p / 8 + (p % 8) / 4 + p % 4);
-#else
-            const float* kptr = kernel.channel(p / 2 + p % 2);
-#endif
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = bias0;
-
-            for (int q = 0; q < nn; q++)
-            {
-                sum0 += tmpptr[0] * kptr[0];
-                tmpptr++;
-                kptr++;
-            }
-
-            outptr0[0] = sum0;
-
-            outptr0++;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_transform_kernel_msa(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = 8b-maxk-inch-outch/8b
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-#if __mips_msa
-    kernel_tm.create(8 * maxk, inch, outch / 8 + (outch % 8) / 4 + outch % 4);
-#else
-    kernel_tm.create(2 * maxk, inch, outch / 2 + outch % 2);
-#endif
-
-    int q = 0;
-#if __mips_msa
-    for (; q + 7 < outch; q += 8)
-    {
-        const Mat k0 = kernel.channel(q);
-        const Mat k1 = kernel.channel(q + 1);
-        const Mat k2 = kernel.channel(q + 2);
-        const Mat k3 = kernel.channel(q + 3);
-        const Mat k4 = kernel.channel(q + 4);
-        const Mat k5 = kernel.channel(q + 5);
-        const Mat k6 = kernel.channel(q + 6);
-        const Mat k7 = kernel.channel(q + 7);
-
-        float* g00 = kernel_tm.channel(q / 8);
-
-        for (int p = 0; p < inch; p++)
-        {
-            const float* k00 = k0.row(p);
-            const float* k10 = k1.row(p);
-            const float* k20 = k2.row(p);
-            const float* k30 = k3.row(p);
-            const float* k40 = k4.row(p);
-            const float* k50 = k5.row(p);
-            const float* k60 = k6.row(p);
-            const float* k70 = k7.row(p);
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k00[k];
-                g00[1] = k10[k];
-                g00[2] = k20[k];
-                g00[3] = k30[k];
-                g00[4] = k40[k];
-                g00[5] = k50[k];
-                g00[6] = k60[k];
-                g00[7] = k70[k];
-
-                g00 += 8;
-            }
-        }
-    }
-    for (; q + 3 < outch; q += 4)
-    {
-        const Mat k0 = kernel.channel(q);
-        const Mat k1 = kernel.channel(q + 1);
-        const Mat k2 = kernel.channel(q + 2);
-        const Mat k3 = kernel.channel(q + 3);
-
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4);
-
-        for (int p = 0; p < inch; p++)
-        {
-            const float* k00 = k0.row(p);
-            const float* k10 = k1.row(p);
-            const float* k20 = k2.row(p);
-            const float* k30 = k3.row(p);
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k00[k];
-                g00[1] = k10[k];
-                g00[2] = k20[k];
-                g00[3] = k30[k];
-
-                g00 += 4;
-            }
-        }
-    }
-#else
-    for (; q + 1 < outch; q += 2)
-    {
-        const Mat k0 = kernel.channel(q);
-        const Mat k1 = kernel.channel(q + 1);
-
-        float* g00 = kernel_tm.channel(q / 2);
-
-        for (int p = 0; p < inch; p++)
-        {
-            const float* k00 = k0.row(p);
-            const float* k10 = k1.row(p);
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k00[k];
-                g00[1] = k10[k];
-
-                g00 += 2;
-            }
-        }
-    }
-#endif // __mips_msa
-    for (; q < outch; q++)
-    {
-        const Mat k0 = kernel.channel(q);
-
-#if __mips_msa
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + q % 4);
-#else
-        float* g00 = kernel_tm.channel(q / 2 + q % 2);
-#endif
-
-        for (int p = 0; p < inch; p++)
-        {
-            const float* k00 = k0.row(p);
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k00[k];
-
-                g00 += 1;
-            }
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 4u, 1, opt.workspace_allocator);
-    {
-        const int gap = w * stride_h - outw * stride_w;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            float* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const float* sptr = img.row<const float>(dilation_h * u) + dilation_w * v;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            ptr[0] = sptr[0];
-
-                            sptr += stride_w;
-                            ptr += 1;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_msa(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_int8.h
deleted file mode 100644
index ab1b1a6b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_int8.h
+++ /dev/null
@@ -1,1179 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if NCNN_RUNTIME_CPU && NCNN_MMI && !__mips_msa && !__mips_loongson_mmi
-void im2col_sgemm_int8_loongson_mmi(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt);
-void convolution_im2col_sgemm_transform_kernel_int8_loongson_mmi(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h);
-#endif
-
-static void im2col_sgemm_int8_msa(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-#if NCNN_RUNTIME_CPU && NCNN_MMI && !__mips_msa && !__mips_loongson_mmi
-    if (ncnn::cpu_support_loongson_mmi())
-    {
-        im2col_sgemm_int8_loongson_mmi(bottom_im2col, top_blob, kernel, opt);
-        return;
-    }
-#endif
-
-    // Mat bottom_im2col(size, maxk, inch, 8u, 8, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    // permute
-    Mat tmp;
-#if __mips_msa || __mips_loongson_mmi
-    if (inch >= 4)
-    {
-        if (size >= 2)
-            tmp.create(2 * maxk, inch / 4 + inch % 4, size / 2 + size % 2, 4u, 4, opt.workspace_allocator);
-        else
-            tmp.create(maxk, inch / 4 + inch % 4, size, 4u, 4, opt.workspace_allocator);
-    }
-    else
-#endif // __mips_msa || __mips_loongson_mmi
-    {
-        if (size >= 2)
-            tmp.create(2 * maxk, inch, size / 2 + size % 2, 1u, 1, opt.workspace_allocator);
-        else
-            tmp.create(maxk, inch, size, 1u, 1, opt.workspace_allocator);
-    }
-    {
-        int remain_size_start = 0;
-        int nn_size = (size - remain_size_start) >> 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 2;
-
-            signed char* tmpptr = tmp.channel(i / 2);
-
-            int q = 0;
-#if __mips_msa || __mips_loongson_mmi
-            for (; q + 3 < inch; q += 4)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-                const signed char* img1 = (const signed char*)bottom_im2col.channel(q + 1) + i;
-                const signed char* img2 = (const signed char*)bottom_im2col.channel(q + 2) + i;
-                const signed char* img3 = (const signed char*)bottom_im2col.channel(q + 3) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-#if __mips_loongson_mmi
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img1[0];
-                    tmpptr[2] = img0[1];
-                    tmpptr[3] = img1[1];
-                    tmpptr[4] = img2[0];
-                    tmpptr[5] = img3[0];
-                    tmpptr[6] = img2[1];
-                    tmpptr[7] = img3[1];
-#else  // __mips_loongson_mmi
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img1[0];
-                    tmpptr[2] = img2[0];
-                    tmpptr[3] = img3[0];
-                    tmpptr[4] = img0[1];
-                    tmpptr[5] = img1[1];
-                    tmpptr[6] = img2[1];
-                    tmpptr[7] = img3[1];
-#endif // __mips_loongson_mmi
-                    tmpptr += 8;
-
-                    img0 += size;
-                    img1 += size;
-                    img2 += size;
-                    img3 += size;
-                }
-            }
-#endif // __mips_msa || __mips_loongson_mmi
-            for (; q < inch; q++)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img0[1];
-
-                    tmpptr += 2;
-
-                    img0 += size;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-
-            int q = 0;
-#if __mips_msa || __mips_loongson_mmi
-            for (; q + 3 < inch; q += 4)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-                const signed char* img1 = (const signed char*)bottom_im2col.channel(q + 1) + i;
-                const signed char* img2 = (const signed char*)bottom_im2col.channel(q + 2) + i;
-                const signed char* img3 = (const signed char*)bottom_im2col.channel(q + 3) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img1[0];
-                    tmpptr[2] = img2[0];
-                    tmpptr[3] = img3[0];
-                    tmpptr += 4;
-
-                    img0 += size;
-                    img1 += size;
-                    img2 += size;
-                    img3 += size;
-                }
-            }
-#endif // __mips_msa || __mips_loongson_mmi
-            for (; q < inch; q++)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-
-                    tmpptr += 1;
-
-                    img0 += size;
-                }
-            }
-        }
-    }
-
-#if __mips_msa
-    int nn_outch = outch >> 2;
-    int remain_outch_start = nn_outch << 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 4;
-
-        int* outptr0 = top_blob.channel(p);
-        int* outptr1 = top_blob.channel(p + 1);
-        int* outptr2 = top_blob.channel(p + 2);
-        int* outptr3 = top_blob.channel(p + 3);
-
-        int i = 0;
-        for (; i + 1 < size; i += 2)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2);
-            const signed char* kptr = kernel.channel(p / 4);
-
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            v4i32 _sum00 = __msa_fill_w(0);
-            v4i32 _sum10 = __msa_fill_w(0);
-
-            if (nn4 > 0)
-            {
-                v4i32 _sum01 = __msa_fill_w(0);
-                v4i32 _sum02 = __msa_fill_w(0);
-                v4i32 _sum03 = __msa_fill_w(0);
-                v4i32 _sum11 = __msa_fill_w(0);
-                v4i32 _sum12 = __msa_fill_w(0);
-                v4i32 _sum13 = __msa_fill_w(0);
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    v16i8 _val = __msa_ld_b(tmpptr, 0);
-                    v8i16 _val01 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_val, 0), _val);
-
-                    v8i16 _val0 = (v8i16)__msa_ilvr_d((v2i64)_val01, (v2i64)_val01);
-                    v8i16 _val1 = (v8i16)__msa_ilvl_d((v2i64)_val01, (v2i64)_val01);
-
-                    v16i8 _w01 = __msa_ld_b(kptr, 0);
-                    v16i8 _extw01 = __msa_clti_s_b(_w01, 0);
-                    v8i16 _w0 = (v8i16)__msa_ilvr_b(_extw01, _w01);
-                    v8i16 _w1 = (v8i16)__msa_ilvl_b(_extw01, _w01);
-
-                    v8i16 _s00 = __msa_mulv_h(_val0, _w0);
-                    v8i16 _s01 = __msa_mulv_h(_val0, _w1);
-                    v8i16 _s10 = __msa_mulv_h(_val1, _w0);
-                    v8i16 _s11 = __msa_mulv_h(_val1, _w1);
-
-                    v8i16 _exts00 = __msa_clti_s_h(_s00, 0);
-                    v8i16 _exts01 = __msa_clti_s_h(_s01, 0);
-                    v8i16 _exts10 = __msa_clti_s_h(_s10, 0);
-                    v8i16 _exts11 = __msa_clti_s_h(_s11, 0);
-                    v4i32 _s00l = (v4i32)__msa_ilvr_h(_exts00, _s00);
-                    v4i32 _s00h = (v4i32)__msa_ilvl_h(_exts00, _s00);
-                    v4i32 _s01l = (v4i32)__msa_ilvr_h(_exts01, _s01);
-                    v4i32 _s01h = (v4i32)__msa_ilvl_h(_exts01, _s01);
-                    v4i32 _s10l = (v4i32)__msa_ilvr_h(_exts10, _s10);
-                    v4i32 _s10h = (v4i32)__msa_ilvl_h(_exts10, _s10);
-                    v4i32 _s11l = (v4i32)__msa_ilvr_h(_exts11, _s11);
-                    v4i32 _s11h = (v4i32)__msa_ilvl_h(_exts11, _s11);
-
-                    _sum00 = __msa_addv_w(_sum00, _s00l);
-                    _sum01 = __msa_addv_w(_sum01, _s00h);
-                    _sum02 = __msa_addv_w(_sum02, _s01l);
-                    _sum03 = __msa_addv_w(_sum03, _s01h);
-                    _sum10 = __msa_addv_w(_sum10, _s10l);
-                    _sum11 = __msa_addv_w(_sum11, _s10h);
-                    _sum12 = __msa_addv_w(_sum12, _s11l);
-                    _sum13 = __msa_addv_w(_sum13, _s11h);
-
-                    tmpptr += 8;
-                    kptr += 16;
-                }
-
-                // transpose 4x4
-                {
-                    v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                    _tmp0 = __msa_ilvr_w(_sum01, _sum00);
-                    _tmp1 = __msa_ilvr_w(_sum03, _sum02);
-                    _tmp2 = __msa_ilvl_w(_sum01, _sum00);
-                    _tmp3 = __msa_ilvl_w(_sum03, _sum02);
-                    _sum00 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                    _sum01 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                    _sum02 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                    _sum03 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-                }
-                {
-                    v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                    _tmp0 = __msa_ilvr_w(_sum11, _sum10);
-                    _tmp1 = __msa_ilvr_w(_sum13, _sum12);
-                    _tmp2 = __msa_ilvl_w(_sum11, _sum10);
-                    _tmp3 = __msa_ilvl_w(_sum13, _sum12);
-                    _sum10 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                    _sum11 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                    _sum12 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                    _sum13 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-                }
-
-                _sum00 = __msa_addv_w(_sum00, _sum01);
-                _sum02 = __msa_addv_w(_sum02, _sum03);
-                _sum10 = __msa_addv_w(_sum10, _sum11);
-                _sum12 = __msa_addv_w(_sum12, _sum13);
-
-                _sum00 = __msa_addv_w(_sum00, _sum02);
-                _sum10 = __msa_addv_w(_sum10, _sum12);
-            }
-
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                v8i16 _val0 = __msa_fill_h(tmpptr[0]);
-                v8i16 _val1 = __msa_fill_h(tmpptr[1]);
-                v8i16 _val = (v8i16)__msa_ilvr_d((v2i64)_val1, (v2i64)_val0);
-
-                v16i8 _w = __msa_ld_b(kptr, 0);
-                v8i16 _w16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_w, 0), _w);
-
-                _w16 = (v8i16)__msa_ilvr_d((v2i64)_w16, (v2i64)_w16);
-
-                v8i16 _s0 = __msa_mulv_h(_val, _w16);
-                v8i16 _exts0 = __msa_clti_s_h(_s0, 0);
-                v4i32 _s0l = (v4i32)__msa_ilvr_h(_exts0, _s0);
-                v4i32 _s0h = (v4i32)__msa_ilvl_h(_exts0, _s0);
-
-                _sum00 = __msa_addv_w(_sum00, _s0l);
-                _sum10 = __msa_addv_w(_sum10, _s0h);
-
-                tmpptr += 2;
-                kptr += 4;
-            }
-
-            int sum[8];
-            __msa_st_w(_sum00, sum, 0);
-            __msa_st_w(_sum10, sum + 4, 0);
-
-            outptr0[0] = sum[0];
-            outptr1[0] = sum[1];
-            outptr2[0] = sum[2];
-            outptr3[0] = sum[3];
-            outptr0[1] = sum[4];
-            outptr1[1] = sum[5];
-            outptr2[1] = sum[6];
-            outptr3[1] = sum[7];
-            outptr0 += 2;
-            outptr1 += 2;
-            outptr2 += 2;
-            outptr3 += 2;
-        }
-        for (; i < size; i++)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-            const signed char* kptr = kernel.channel(p / 4);
-
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            v4i32 _sum0 = __msa_fill_w(0);
-
-            if (nn4 > 0)
-            {
-                v4i32 _sum1 = __msa_fill_w(0);
-                v4i32 _sum2 = __msa_fill_w(0);
-                v4i32 _sum3 = __msa_fill_w(0);
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    v16i8 _val = __msa_ld_b(tmpptr, 0);
-                    v8i16 _val16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_val, 0), _val);
-
-                    _val16 = (v8i16)__msa_ilvr_d((v2i64)_val16, (v2i64)_val16);
-
-                    v16i8 _w01 = __msa_ld_b(kptr, 0);
-                    v16i8 _extw01 = __msa_clti_s_b(_w01, 0);
-                    v8i16 _w0 = (v8i16)__msa_ilvr_b(_extw01, _w01);
-                    v8i16 _w1 = (v8i16)__msa_ilvl_b(_extw01, _w01);
-
-                    v8i16 _s0 = __msa_mulv_h(_val16, _w0);
-                    v8i16 _s1 = __msa_mulv_h(_val16, _w1);
-
-                    v8i16 _exts0 = __msa_clti_s_h(_s0, 0);
-                    v8i16 _exts1 = __msa_clti_s_h(_s1, 0);
-                    v4i32 _s0l = (v4i32)__msa_ilvr_h(_exts0, _s0);
-                    v4i32 _s0h = (v4i32)__msa_ilvl_h(_exts0, _s0);
-                    v4i32 _s1l = (v4i32)__msa_ilvr_h(_exts1, _s1);
-                    v4i32 _s1h = (v4i32)__msa_ilvl_h(_exts1, _s1);
-
-                    _sum0 = __msa_addv_w(_sum0, _s0l);
-                    _sum1 = __msa_addv_w(_sum1, _s0h);
-                    _sum2 = __msa_addv_w(_sum2, _s1l);
-                    _sum3 = __msa_addv_w(_sum3, _s1h);
-
-                    tmpptr += 4;
-                    kptr += 16;
-                }
-
-                // transpose 4x4
-                {
-                    v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                    _tmp0 = __msa_ilvr_w(_sum1, _sum0);
-                    _tmp1 = __msa_ilvr_w(_sum3, _sum2);
-                    _tmp2 = __msa_ilvl_w(_sum1, _sum0);
-                    _tmp3 = __msa_ilvl_w(_sum3, _sum2);
-                    _sum0 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                    _sum1 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                    _sum2 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                    _sum3 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-                }
-
-                _sum0 = __msa_addv_w(_sum0, _sum1);
-                _sum2 = __msa_addv_w(_sum2, _sum3);
-                _sum0 = __msa_addv_w(_sum0, _sum2);
-            }
-
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                v8i16 _val = __msa_fill_h(tmpptr[0]);
-
-                v16i8 _w = __msa_ld_b(kptr, 0);
-                v8i16 _w16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_w, 0), _w);
-
-                v8i16 _s0 = __msa_mulv_h(_val, _w16);
-                v4i32 _s032 = (v4i32)__msa_ilvr_h(__msa_clti_s_h(_s0, 0), _s0);
-
-                _sum0 = __msa_addv_w(_sum0, _s032);
-
-                tmpptr += 1;
-                kptr += 4;
-            }
-
-            int sum[4];
-            __msa_st_w(_sum0, sum, 0);
-
-            outptr0[0] = sum[0];
-            outptr1[0] = sum[1];
-            outptr2[0] = sum[2];
-            outptr3[0] = sum[3];
-            outptr0 += 1;
-            outptr1 += 1;
-            outptr2 += 1;
-            outptr3 += 1;
-        }
-    }
-#else // __mips_msa
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        int* outptr0 = top_blob.channel(p);
-        int* outptr1 = top_blob.channel(p + 1);
-
-        int i = 0;
-        for (; i + 1 < size; i += 2)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2);
-            const signed char* kptr = kernel.channel(p / 2);
-
-            int sum00 = 0;
-            int sum01 = 0;
-            int sum10 = 0;
-            int sum11 = 0;
-
-#if __mips_loongson_mmi
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            if (nn4 > 0)
-            {
-                int32x2_t _sum0 = __mmi_pzerow_s();
-                int32x2_t _sum1 = __mmi_pzerow_s();
-
-                double temp0 = 0;
-                double temp1 = 0;
-                double temp2 = 0;
-                double temp3 = 0;
-                double temp4 = 0;
-                double temp5 = 0;
-                double temp6 = 0;
-                double temp7 = 0;
-
-                uint64_t flag_0x44 = 0x44;
-                uint64_t flag_0xee = 0xee;
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    asm volatile(
-                        "ld         $0, 32(%0)      \n" // __builtin_prefetch(tmpptr + 32);
-                        "ld         $0, 32(%1)      \n" // __builtin_prefetch(kptr + 32);
-
-                        "ldc1       %4, 0(%0)       \n" // int8x8_t _v = __mmi_pldb_s(tmpptr);
-                        "ldc1       %6, 0(%1)       \n" // int8x8_t _k = __mmi_pldb_s(kptr);
-
-                        "mtc1       $0, %8          \n" // int8x8_t _zero = __mmi_pzerob_s();
-                        "pcmpgtb    %5, %8, %4      \n" // int8x8_t _extv = __mmi_pcmpgtb_s(_zero, _v);
-                        "pcmpgtb    %7, %8, %6      \n" // int8x8_t _extk = __mmi_pcmpgtb_s(_zero, _k);
-
-                        "punpcklbh  %8, %4, %5      \n" // int16x4_t _v0 = (int16x4_t)__mmi_punpcklbh_s(_v, _extv);
-                        "punpckhbh  %9, %4, %5      \n" // int16x4_t _v1 = (int16x4_t)__mmi_punpckhbh_s(_v, _extv);
-                        "punpcklbh  %10, %6, %7     \n" // int16x4_t _k0 = (int16x4_t)__mmi_punpcklbh_s(_k, _extk);
-                        "punpckhbh  %11, %6, %7     \n" // int16x4_t _k1 = (int16x4_t)__mmi_punpckhbh_s(_k, _extk);
-
-                        "pshufh     %4, %10, %24    \n" // int16x4_t _k0202 = __mmi_pshufh_s(_k0, 0x44);
-                        "pshufh     %5, %10, %25    \n" // int16x4_t _k1313 = __mmi_pshufh_s(_k0, 0xee);
-                        "pshufh     %6, %11, %24    \n" // int16x4_t _k4646 = __mmi_pshufh_s(_k1, 0x44);
-                        "pshufh     %7, %11, %25    \n" // int16x4_t _k5757 = __mmi_pshufh_s(_k1, 0xee);
-
-                        "pmaddhw    %4, %8, %4      \n" // int32x2_t _s0x = __mmi_pmaddhw(_v0, _k0202);
-                        "pmaddhw    %5, %8, %5      \n" // int32x2_t _s1x = __mmi_pmaddhw(_v0, _k1313);
-                        "pmaddhw    %6, %9, %6      \n" // int32x2_t _s0y = __mmi_pmaddhw(_v1, _k4646);
-                        "pmaddhw    %7, %9, %7      \n" // int32x2_t _s1y = __mmi_pmaddhw(_v1, _k5757);
-
-                        "paddw      %2, %2, %4      \n" // _sum0 = __mmi_paddw_s(_sum0, _s0x);
-                        "paddw      %3, %3, %5      \n" // _sum1 = __mmi_paddw_s(_sum1, _s1x);
-                        "paddw      %2, %2, %6      \n" // _sum0 = __mmi_paddw_s(_sum0, _s0y);
-                        "paddw      %3, %3, %7      \n" // _sum1 = __mmi_paddw_s(_sum1, _s1y);
-
-                        : "=r"(tmpptr), // %0
-                        "=r"(kptr),   // %1
-                        "=f"(_sum0),  // %2
-                        "=f"(_sum1),  // %3
-                        "=f"(temp0),  // %4
-                        "=f"(temp1),  // %5
-                        "=f"(temp2),  // %6
-                        "=f"(temp3),  // %7
-                        "=f"(temp4),  // %8
-                        "=f"(temp5),  // %9
-                        "=f"(temp6),  // %10
-                        "=f"(temp7)   // %11
-                        : "0"(tmpptr),
-                        "1"(kptr),
-                        "2"(_sum0),
-                        "3"(_sum1),
-                        "4"(temp0),
-                        "5"(temp1),
-                        "6"(temp2),
-                        "7"(temp3),
-                        "8"(temp4),
-                        "9"(temp5),
-                        "10"(temp6),
-                        "11"(temp7),
-                        "f"(flag_0x44), // %24
-                        "f"(flag_0xee)  // %25
-                        : "memory");
-
-                    tmpptr += 8;
-                    kptr += 8;
-                }
-
-                int sum[4];
-                __mmi_pstw_s(sum, _sum0);
-                __mmi_pstw_s(sum + 2, _sum1);
-
-                sum00 = sum[0];
-                sum01 = sum[1];
-                sum10 = sum[2];
-                sum11 = sum[3];
-            }
-#else  // __mips_loongson_mmi
-            int nn1 = inch * maxk;
-#endif // __mips_loongson_mmi
-
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                signed char val0 = tmpptr[0];
-                signed char val1 = tmpptr[1];
-                signed char w0 = kptr[0];
-                signed char w1 = kptr[1];
-
-                sum00 += val0 * w0;
-                sum01 += val1 * w0;
-                sum10 += val0 * w1;
-                sum11 += val1 * w1;
-
-                tmpptr += 2;
-                kptr += 2;
-            }
-
-            outptr0[0] = sum00;
-            outptr0[1] = sum01;
-            outptr1[0] = sum10;
-            outptr1[1] = sum11;
-            outptr0 += 2;
-            outptr1 += 2;
-        }
-        for (; i < size; i++)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-            const signed char* kptr = kernel.channel(p / 2);
-
-            int sum00 = 0;
-            int sum10 = 0;
-
-#if __mips_loongson_mmi
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            if (nn4 > 0)
-            {
-                int32x2_t _sum01 = __mmi_pzerow_s();
-
-                double temp0 = 0;
-                double temp1 = 0;
-                double temp2 = 0;
-                double temp3 = 0;
-                double temp4 = 0;
-                double temp5 = 0;
-
-                uint64_t flag_0x44 = 0x44;
-                uint64_t flag_0xee = 0xee;
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    asm volatile(
-                        "ld         $0, 16(%0)      \n" // __builtin_prefetch(tmpptr + 16);
-                        "ld         $0, 32(%1)      \n" // __builtin_prefetch(kptr + 32);
-
-                        "ldc1       %3, 0(%0)       \n" // int8x8_t _v = __mmi_pldb_s(tmpptr);
-                        "ldc1       %5, 0(%1)       \n" // int8x8_t _k = __mmi_pldb_s(kptr);
-
-                        "mtc1       $0, %7          \n" // int8x8_t _zero = __mmi_pzerob_s();
-                        "pcmpgtb    %4, %7, %3      \n" // int8x8_t _extv = __mmi_pcmpgtb_s(_zero, _v);
-                        "pcmpgtb    %6, %7, %5      \n" // int8x8_t _extk = __mmi_pcmpgtb_s(_zero, _k);
-
-                        "punpcklbh  %4, %3, %4      \n" // int16x4_t _v0 = (int16x4_t)__mmi_punpcklbh_s(_v, _extv);
-                        "punpcklbh  %7, %5, %6      \n" // int16x4_t _k0 = (int16x4_t)__mmi_punpcklbh_s(_k, _extk);
-                        "punpckhbh  %8, %5, %6      \n" // int16x4_t _k1 = (int16x4_t)__mmi_punpckhbh_s(_k, _extk);
-
-                        "pshufh     %3, %4, %18     \n" // int16x4_t _v0101 = __mmi_pshufh_s(_v0, 0x44);
-                        "pshufh     %4, %4, %19     \n" // int16x4_t _v2323 = __mmi_pshufh_s(_v0, 0xee);
-
-                        "pmaddhw    %3, %3, %7      \n" // int32x2_t _s01x = __mmi_pmaddhw(_v0101, _k0);
-                        "pmaddhw    %4, %4, %8      \n" // int32x2_t _s01y = __mmi_pmaddhw(_v2323, _k1);
-
-                        "paddw      %2, %2, %3      \n" // _sum01 = __mmi_paddw_s(_sum01, _s01x);
-                        "paddw      %2, %2, %4      \n" // _sum01 = __mmi_paddw_s(_sum01, _s01y);
-
-                        : "=r"(tmpptr), // %0
-                        "=r"(kptr),   // %1
-                        "=f"(_sum01), // %2
-                        "=f"(temp0),  // %3
-                        "=f"(temp1),  // %4
-                        "=f"(temp2),  // %5
-                        "=f"(temp3),  // %6
-                        "=f"(temp4),  // %7
-                        "=f"(temp5)   // %8
-                        : "0"(tmpptr),
-                        "1"(kptr),
-                        "2"(_sum01),
-                        "3"(temp0),
-                        "4"(temp1),
-                        "5"(temp2),
-                        "6"(temp3),
-                        "7"(temp4),
-                        "8"(temp5),
-                        "f"(flag_0x44), // %18
-                        "f"(flag_0xee)  // %19
-                        : "memory");
-
-                    tmpptr += 4;
-                    kptr += 8;
-                }
-
-                int sum[2];
-                __mmi_pstw_s(sum, _sum01);
-
-                sum00 = sum[0];
-                sum10 = sum[1];
-            }
-#else  // __mips_loongson_mmi
-            int nn1 = inch * maxk;
-#endif // __mips_loongson_mmi
-
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                signed char val0 = tmpptr[0];
-                signed char w0 = kptr[0];
-                signed char w1 = kptr[1];
-
-                sum00 += val0 * w0;
-                sum10 += val0 * w1;
-
-                tmpptr += 1;
-                kptr += 2;
-            }
-
-            outptr0[0] = sum00;
-            outptr1[0] = sum10;
-            outptr0 += 1;
-            outptr1 += 1;
-        }
-    }
-#endif // __mips_msa
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        int* outptr0 = top_blob.channel(p);
-
-        int i = 0;
-        for (; i + 1 < size; i += 2)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2);
-#if __mips_msa
-            const signed char* kptr = kernel.channel(p / 4 + p % 4);
-#else
-            const signed char* kptr = kernel.channel(p / 2 + p % 2);
-#endif
-
-            int sum0 = 0;
-            int sum1 = 0;
-
-#if __mips_msa
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            if (nn4 > 0)
-            {
-                v4i32 _sum0 = __msa_fill_w(0);
-                v4i32 _sum1 = __msa_fill_w(0);
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    v16i8 _val = __msa_ld_b(tmpptr, 0);
-                    v8i16 _val16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_val, 0), _val);
-
-                    v16i8 _w = __msa_ld_b(kptr, 0);
-                    v8i16 _w16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_w, 0), _w);
-
-                    _w16 = (v8i16)__msa_ilvr_d((v2i64)_w16, (v2i64)_w16);
-
-                    v8i16 _s0 = __msa_mulv_h(_val16, _w16);
-                    v8i16 _exts0 = __msa_clti_s_h(_s0, 0);
-                    v4i32 _s0l = (v4i32)__msa_ilvr_h(_exts0, _s0);
-                    v4i32 _s0h = (v4i32)__msa_ilvl_h(_exts0, _s0);
-
-                    _sum0 = __msa_addv_w(_sum0, _s0l);
-                    _sum1 = __msa_addv_w(_sum1, _s0h);
-
-                    tmpptr += 8;
-                    kptr += 4;
-                }
-
-                sum0 = _sum0[0] + _sum0[1] + _sum0[2] + _sum0[3];
-                sum1 = _sum1[0] + _sum1[1] + _sum1[2] + _sum1[3];
-            }
-#elif __mips_loongson_mmi
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            if (nn4 > 0)
-            {
-                int32x2_t _sum01 = __mmi_pzerow_s();
-
-                double temp0 = 0;
-                double temp1 = 0;
-                double temp2 = 0;
-                double temp3 = 0;
-                double temp4 = 0;
-                double temp5 = 0;
-
-                uint64_t flag_0x44 = 0x44;
-                uint64_t flag_0xee = 0xee;
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    asm volatile(
-                        "ld         $0, 32(%0)      \n" // __builtin_prefetch(tmpptr + 32);
-                        "ld         $0, 16(%1)      \n" // __builtin_prefetch(kptr + 16);
-
-                        "ldc1       %3, 0(%0)       \n" // int8x8_t _v = __mmi_pldb_s(tmpptr);
-                        "ldc1       %5, 0(%1)       \n" // int8x8_t _k = __mmi_pldb_s(kptr);
-
-                        "mtc1       $0, %6          \n" // int8x8_t _zero = __mmi_pzerob_s();
-                        "pcmpgtb    %4, %6, %3      \n" // int8x8_t _extv = __mmi_pcmpgtb_s(_zero, _v);
-                        "pcmpgtb    %6, %6, %5      \n" // int8x8_t _extk = __mmi_pcmpgtb_s(_zero, _k);
-
-                        "punpcklbh  %7, %3, %4      \n" // int16x4_t _v0 = (int16x4_t)__mmi_punpcklbh_s(_v, _extv);
-                        "punpckhbh  %8, %3, %4      \n" // int16x4_t _v1 = (int16x4_t)__mmi_punpckhbh_s(_v, _extv);
-                        "punpcklbh  %5, %5, %6      \n" // int16x4_t _k0 = (int16x4_t)__mmi_punpcklbh_s(_k, _extk);
-
-                        "pshufh     %3, %5, %18     \n" // int16x4_t _k0202 = __mmi_pshufh_s(_k0, 0x44);
-                        "pshufh     %4, %5, %19     \n" // int16x4_t _k1313 = __mmi_pshufh_s(_k0, 0xee);
-
-                        "pmaddhw    %3, %7, %3      \n" // int32x2_t _s01x = __mmi_pmaddhw(_v0, _k0101);
-                        "pmaddhw    %4, %8, %4      \n" // int32x2_t _s01y = __mmi_pmaddhw(_v1, _k2323);
-
-                        "paddw      %2, %2, %3      \n" // _sum01 = __mmi_paddw_s(_sum01, _s01x);
-                        "paddw      %2, %2, %4      \n" // _sum01 = __mmi_paddw_s(_sum01, _s01y);
-
-                        : "=r"(tmpptr), // %0
-                        "=r"(kptr),   // %1
-                        "=f"(_sum01), // %2
-                        "=f"(temp0),  // %3
-                        "=f"(temp1),  // %4
-                        "=f"(temp2),  // %5
-                        "=f"(temp3),  // %6
-                        "=f"(temp4),  // %7
-                        "=f"(temp5)   // %8
-                        : "0"(tmpptr),
-                        "1"(kptr),
-                        "2"(_sum01),
-                        "3"(temp0),
-                        "4"(temp1),
-                        "5"(temp2),
-                        "6"(temp3),
-                        "7"(temp4),
-                        "8"(temp5),
-                        "f"(flag_0x44), // %18
-                        "f"(flag_0xee)  // %19
-                        : "memory");
-
-                    tmpptr += 8;
-                    kptr += 4;
-                }
-
-                int sum[2];
-                __mmi_pstw_s(sum, _sum01);
-
-                sum0 = sum[0];
-                sum1 = sum[1];
-            }
-#else  // __mips_loongson_mmi
-            int nn1 = inch * maxk;
-#endif // __mips_msa || __mips_loongson_mmi
-
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                signed char val0 = tmpptr[0];
-                signed char val1 = tmpptr[1];
-                signed char w = kptr[0];
-
-                sum0 += val0 * w;
-                sum1 += val1 * w;
-
-                tmpptr += 2;
-                kptr += 1;
-            }
-
-            outptr0[0] = sum0;
-            outptr0[1] = sum1;
-            outptr0 += 2;
-        }
-        for (; i < size; i++)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-#if __mips_msa
-            const signed char* kptr = kernel.channel(p / 4 + p % 4);
-#else
-            const signed char* kptr = kernel.channel(p / 2 + p % 2);
-#endif
-
-            int sum = 0;
-
-#if __mips_msa
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            if (nn4 > 0)
-            {
-                v4i32 _sum = __msa_fill_w(0);
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    v16i8 _val = __msa_ld_b(tmpptr, 0);
-                    v8i16 _val16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_val, 0), _val);
-
-                    v16i8 _w = __msa_ld_b(kptr, 0);
-                    v8i16 _w16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_w, 0), _w);
-
-                    v8i16 _s0 = __msa_mulv_h(_val16, _w16);
-                    v4i32 _s032 = (v4i32)__msa_ilvr_h(__msa_clti_s_h(_s0, 0), _s0);
-
-                    _sum = __msa_addv_w(_sum, _s032);
-
-                    tmpptr += 4;
-                    kptr += 4;
-                }
-
-                sum = _sum[0] + _sum[1] + _sum[2] + _sum[3];
-            }
-#elif __mips_loongson_mmi
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            if (nn4 > 0)
-            {
-                int32x2_t _sum0 = __mmi_pzerow_s();
-
-                double temp0 = 0;
-                double temp1 = 0;
-                double temp2 = 0;
-                double temp3 = 0;
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    asm volatile(
-                        "ld         $0, 16(%0)      \n" // __builtin_prefetch(tmpptr + 16);
-                        "ld         $0, 16(%1)      \n" // __builtin_prefetch(kptr + 16);
-
-                        "ldc1       %3, 0(%0)       \n" // int8x8_t _v = __mmi_pldb_s(tmpptr);
-                        "ldc1       %5, 0(%1)       \n" // int8x8_t _k = __mmi_pldb_s(kptr);
-
-                        "mtc1       $0, %6          \n" // int8x8_t _zero = __mmi_pzerob_s();
-                        "pcmpgtb    %4, %6, %3      \n" // int8x8_t _extv = __mmi_pcmpgtb_s(_zero, _v);
-                        "pcmpgtb    %6, %6, %5      \n" // int8x8_t _extk = __mmi_pcmpgtb_s(_zero, _k);
-
-                        "punpcklbh  %3, %3, %4      \n" // int16x4_t _v0 = (int16x4_t)__mmi_punpcklbh_s(_v, _extv);
-                        "punpcklbh  %5, %5, %6      \n" // int16x4_t _k0 = (int16x4_t)__mmi_punpcklbh_s(_k, _extk);
-
-                        "pmaddhw    %3, %3, %5      \n" // int32x2_t _s0x = __mmi_pmaddhw(_v0, _k0);
-                        "paddw      %2, %2, %3      \n" // _sum0 = __mmi_paddw_s(_sum0, _s0x);
-
-                        : "=r"(tmpptr), // %0
-                        "=r"(kptr),   // %1
-                        "=f"(_sum0),  // %2
-                        "=f"(temp0),  // %3
-                        "=f"(temp1),  // %4
-                        "=f"(temp2),  // %5
-                        "=f"(temp3)   // %6
-                        : "0"(tmpptr),
-                        "1"(kptr),
-                        "2"(_sum0),
-                        "3"(temp0),
-                        "4"(temp1),
-                        "5"(temp2),
-                        "6"(temp3)
-                        : "memory");
-
-                    tmpptr += 4;
-                    kptr += 4;
-                }
-
-                int tmp[2];
-                __mmi_pstw_s(tmp, _sum0);
-
-                sum = tmp[0] + tmp[1];
-            }
-#else  // __mips_loongson_mmi
-            int nn1 = inch * maxk;
-#endif // __mips_msa
-
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                signed char val = tmpptr[0];
-                signed char w = kptr[0];
-
-                sum += val * w;
-
-                tmpptr += 1;
-                kptr += 1;
-            }
-
-            outptr0[0] = sum;
-            outptr0 += 1;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_transform_kernel_int8_msa(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-#if NCNN_RUNTIME_CPU && NCNN_MMI && !__mips_msa && !__mips_loongson_mmi
-    if (ncnn::cpu_support_loongson_mmi())
-    {
-        convolution_im2col_sgemm_transform_kernel_int8_loongson_mmi(_kernel, kernel_tm, inch, outch, kernel_w, kernel_h);
-        return;
-    }
-#endif
-
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = 4a-4b-maxk-inch/4a-outch/4b
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-#if __mips_msa
-    if (outch >= 4)
-    {
-        if (inch >= 4)
-            kernel_tm.create(16 * maxk, inch / 4 + inch % 4, outch / 4 + outch % 4, (size_t)1u);
-        else
-            kernel_tm.create(4 * maxk, inch, outch / 4 + outch % 4, (size_t)1u);
-    }
-#else
-    if (outch >= 2)
-    {
-#if __mips_loongson_mmi
-        if (inch >= 4)
-            kernel_tm.create(8 * maxk, inch / 4 + inch % 4, outch / 2 + outch % 2, (size_t)1u);
-        else
-#endif // __mips_loongson_mmi
-        {
-            kernel_tm.create(2 * maxk, inch, outch / 2 + outch % 2, (size_t)1u);
-        }
-    }
-#endif // __mips_msa
-    else
-    {
-#if __mips_msa || __mips_loongson_mmi
-        if (inch >= 4)
-            kernel_tm.create(4 * maxk, inch / 4 + inch % 4, outch, (size_t)1u);
-        else
-#endif // __mips_msa || __mips_loongson_mmi
-        {
-            kernel_tm.create(1 * maxk, inch, outch, (size_t)1u);
-        }
-    }
-
-    int q = 0;
-#if __mips_msa
-    for (; q + 3 < outch; q += 4)
-    {
-        signed char* g00 = kernel_tm.channel(q / 4);
-
-        int p = 0;
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 4; j++)
-                    {
-                        const signed char* k00 = kernel.channel(q + i).row<const signed char>(p + j);
-                        g00[0] = k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    const signed char* k00 = kernel.channel(q + i).row<const signed char>(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#else // __mips_msa
-    for (; q + 1 < outch; q += 2)
-    {
-        signed char* g00 = kernel_tm.channel(q / 2);
-
-        int p = 0;
-#if __mips_loongson_mmi
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k00 = kernel.channel(q).row<const signed char>(p);
-                const signed char* k01 = kernel.channel(q).row<const signed char>(p + 1);
-                const signed char* k02 = kernel.channel(q).row<const signed char>(p + 2);
-                const signed char* k03 = kernel.channel(q).row<const signed char>(p + 3);
-                const signed char* k10 = kernel.channel(q + 1).row<const signed char>(p);
-                const signed char* k11 = kernel.channel(q + 1).row<const signed char>(p + 1);
-                const signed char* k12 = kernel.channel(q + 1).row<const signed char>(p + 2);
-                const signed char* k13 = kernel.channel(q + 1).row<const signed char>(p + 3);
-
-                g00[0] = k00[k];
-                g00[1] = k01[k];
-                g00[2] = k10[k];
-                g00[3] = k11[k];
-                g00[4] = k02[k];
-                g00[5] = k03[k];
-                g00[6] = k12[k];
-                g00[7] = k13[k];
-
-                g00 += 8;
-            }
-        }
-#endif // __mips_loongson_mmi
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 2; i++)
-                {
-                    const signed char* k00 = kernel.channel(q + i).row<const signed char>(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#endif // __mips_msa
-    for (; q < outch; q++)
-    {
-#if __mips_msa
-        signed char* g00 = kernel_tm.channel(q / 4 + q % 4);
-#else
-        signed char* g00 = kernel_tm.channel(q / 2 + q % 2);
-#endif
-
-        int p = 0;
-#if __mips_msa || __mips_loongson_mmi
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int j = 0; j < 4; j++)
-                {
-                    const signed char* k00 = kernel.channel(q).row<const signed char>(p + j);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-#endif // __mips_msa || __mips_loongson_mmi
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k00 = kernel.channel(q).row<const signed char>(p);
-                g00[0] = k00[k];
-                g00++;
-            }
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 1u, 1, opt.workspace_allocator);
-    {
-        const int gap = w * stride_h - outw * stride_w;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            signed char* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const signed char* sptr = img.row<const signed char>(dilation_h * u) + dilation_w * v;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j + 3 < outw; j += 4)
-                        {
-                            ptr[0] = sptr[0];
-                            ptr[1] = sptr[stride_w];
-                            ptr[2] = sptr[stride_w * 2];
-                            ptr[3] = sptr[stride_w * 3];
-
-                            sptr += stride_w * 4;
-                            ptr += 4;
-                        }
-                        for (; j + 1 < outw; j += 2)
-                        {
-                            ptr[0] = sptr[0];
-                            ptr[1] = sptr[stride_w];
-
-                            sptr += stride_w * 2;
-                            ptr += 2;
-                        }
-                        for (; j < outw; j++)
-                        {
-                            ptr[0] = sptr[0];
-
-                            sptr += stride_w;
-                            ptr += 1;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_int8_msa(bottom_im2col, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack1to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack1to4_int8.h
deleted file mode 100644
index 55dd7ecd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack1to4_int8.h
+++ /dev/null
@@ -1,481 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_pack1to4_int8_msa(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    // Mat bottom_im2col(size, maxk, inch, 8u, 8, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    // permute
-    Mat tmp;
-    if (inch >= 4)
-    {
-        if (size >= 2)
-            tmp.create(2 * maxk, inch / 4 + inch % 4, size / 2 + size % 2, 4u, 4, opt.workspace_allocator);
-        else
-            tmp.create(maxk, inch / 4 + inch % 4, size, 4u, 4, opt.workspace_allocator);
-    }
-    else
-    {
-        if (size >= 2)
-            tmp.create(2 * maxk, inch, size / 2 + size % 2, 1u, 1, opt.workspace_allocator);
-        else
-            tmp.create(maxk, inch, size, 1u, 1, opt.workspace_allocator);
-    }
-    {
-        int remain_size_start = 0;
-        int nn_size = (size - remain_size_start) >> 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 2;
-
-            signed char* tmpptr = tmp.channel(i / 2);
-
-            int q = 0;
-            for (; q + 3 < inch; q += 4)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-                const signed char* img1 = (const signed char*)bottom_im2col.channel(q + 1) + i;
-                const signed char* img2 = (const signed char*)bottom_im2col.channel(q + 2) + i;
-                const signed char* img3 = (const signed char*)bottom_im2col.channel(q + 3) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img1[0];
-                    tmpptr[2] = img2[0];
-                    tmpptr[3] = img3[0];
-                    tmpptr[4] = img0[1];
-                    tmpptr[5] = img1[1];
-                    tmpptr[6] = img2[1];
-                    tmpptr[7] = img3[1];
-                    tmpptr += 8;
-
-                    img0 += size;
-                    img1 += size;
-                    img2 += size;
-                    img3 += size;
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img0[1];
-
-                    tmpptr += 2;
-
-                    img0 += size;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-
-            int q = 0;
-            for (; q + 3 < inch; q += 4)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-                const signed char* img1 = (const signed char*)bottom_im2col.channel(q + 1) + i;
-                const signed char* img2 = (const signed char*)bottom_im2col.channel(q + 2) + i;
-                const signed char* img3 = (const signed char*)bottom_im2col.channel(q + 3) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img1[0];
-                    tmpptr[2] = img2[0];
-                    tmpptr[3] = img3[0];
-                    tmpptr += 4;
-
-                    img0 += size;
-                    img1 += size;
-                    img2 += size;
-                    img3 += size;
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* img0 = (const signed char*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-
-                    tmpptr += 1;
-
-                    img0 += size;
-                }
-            }
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        int* outptr0 = top_blob.channel(p);
-
-        int i = 0;
-        for (; i + 1 < size; i += 2)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2);
-            const signed char* kptr = kernel.channel(p);
-
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            v4i32 _sum00 = __msa_fill_w(0);
-            v4i32 _sum10 = __msa_fill_w(0);
-
-            if (nn4 > 0)
-            {
-                v4i32 _sum01 = __msa_fill_w(0);
-                v4i32 _sum02 = __msa_fill_w(0);
-                v4i32 _sum03 = __msa_fill_w(0);
-                v4i32 _sum11 = __msa_fill_w(0);
-                v4i32 _sum12 = __msa_fill_w(0);
-                v4i32 _sum13 = __msa_fill_w(0);
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    __builtin_prefetch(tmpptr + 32);
-                    __builtin_prefetch(kptr + 64);
-                    v16i8 _val = __msa_ld_b(tmpptr, 0);
-                    v8i16 _val01 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_val, 0), _val);
-
-                    v8i16 _val0 = (v8i16)__msa_ilvr_d((v2i64)_val01, (v2i64)_val01);
-                    v8i16 _val1 = (v8i16)__msa_ilvl_d((v2i64)_val01, (v2i64)_val01);
-
-                    v16i8 _w01 = __msa_ld_b(kptr, 0);
-                    v16i8 _extw01 = __msa_clti_s_b(_w01, 0);
-                    v8i16 _w0 = (v8i16)__msa_ilvr_b(_extw01, _w01);
-                    v8i16 _w1 = (v8i16)__msa_ilvl_b(_extw01, _w01);
-
-                    v8i16 _s00 = __msa_mulv_h(_val0, _w0);
-                    v8i16 _s01 = __msa_mulv_h(_val0, _w1);
-                    v8i16 _s10 = __msa_mulv_h(_val1, _w0);
-                    v8i16 _s11 = __msa_mulv_h(_val1, _w1);
-
-                    v8i16 _exts00 = __msa_clti_s_h(_s00, 0);
-                    v8i16 _exts01 = __msa_clti_s_h(_s01, 0);
-                    v8i16 _exts10 = __msa_clti_s_h(_s10, 0);
-                    v8i16 _exts11 = __msa_clti_s_h(_s11, 0);
-                    v4i32 _s00l = (v4i32)__msa_ilvr_h(_exts00, _s00);
-                    v4i32 _s00h = (v4i32)__msa_ilvl_h(_exts00, _s00);
-                    v4i32 _s01l = (v4i32)__msa_ilvr_h(_exts01, _s01);
-                    v4i32 _s01h = (v4i32)__msa_ilvl_h(_exts01, _s01);
-                    v4i32 _s10l = (v4i32)__msa_ilvr_h(_exts10, _s10);
-                    v4i32 _s10h = (v4i32)__msa_ilvl_h(_exts10, _s10);
-                    v4i32 _s11l = (v4i32)__msa_ilvr_h(_exts11, _s11);
-                    v4i32 _s11h = (v4i32)__msa_ilvl_h(_exts11, _s11);
-
-                    _sum00 = __msa_addv_w(_sum00, _s00l);
-                    _sum01 = __msa_addv_w(_sum01, _s00h);
-                    _sum02 = __msa_addv_w(_sum02, _s01l);
-                    _sum03 = __msa_addv_w(_sum03, _s01h);
-                    _sum10 = __msa_addv_w(_sum10, _s10l);
-                    _sum11 = __msa_addv_w(_sum11, _s10h);
-                    _sum12 = __msa_addv_w(_sum12, _s11l);
-                    _sum13 = __msa_addv_w(_sum13, _s11h);
-
-                    tmpptr += 8;
-                    kptr += 16;
-                }
-
-                // transpose 4x4
-                {
-                    v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                    _tmp0 = __msa_ilvr_w(_sum01, _sum00);
-                    _tmp1 = __msa_ilvr_w(_sum03, _sum02);
-                    _tmp2 = __msa_ilvl_w(_sum01, _sum00);
-                    _tmp3 = __msa_ilvl_w(_sum03, _sum02);
-                    _sum00 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                    _sum01 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                    _sum02 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                    _sum03 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-                }
-                {
-                    v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                    _tmp0 = __msa_ilvr_w(_sum11, _sum10);
-                    _tmp1 = __msa_ilvr_w(_sum13, _sum12);
-                    _tmp2 = __msa_ilvl_w(_sum11, _sum10);
-                    _tmp3 = __msa_ilvl_w(_sum13, _sum12);
-                    _sum10 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                    _sum11 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                    _sum12 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                    _sum13 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-                }
-
-                _sum00 = __msa_addv_w(_sum00, _sum01);
-                _sum02 = __msa_addv_w(_sum02, _sum03);
-                _sum10 = __msa_addv_w(_sum10, _sum11);
-                _sum12 = __msa_addv_w(_sum12, _sum13);
-
-                _sum00 = __msa_addv_w(_sum00, _sum02);
-                _sum10 = __msa_addv_w(_sum10, _sum12);
-            }
-
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                v8i16 _val0 = __msa_fill_h(tmpptr[0]);
-                v8i16 _val1 = __msa_fill_h(tmpptr[1]);
-                v8i16 _val = (v8i16)__msa_ilvr_d((v2i64)_val1, (v2i64)_val0);
-
-                v16i8 _w = __msa_ld_b(kptr, 0);
-                v8i16 _w16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_w, 0), _w);
-
-                _w16 = (v8i16)__msa_ilvr_d((v2i64)_w16, (v2i64)_w16);
-
-                v8i16 _s0 = __msa_mulv_h(_val, _w16);
-                v8i16 _exts0 = __msa_clti_s_h(_s0, 0);
-                v4i32 _s0l = (v4i32)__msa_ilvr_h(_exts0, _s0);
-                v4i32 _s0h = (v4i32)__msa_ilvl_h(_exts0, _s0);
-
-                _sum00 = __msa_addv_w(_sum00, _s0l);
-                _sum10 = __msa_addv_w(_sum10, _s0h);
-
-                tmpptr += 2;
-                kptr += 4;
-            }
-
-            __msa_st_w(_sum00, outptr0, 0);
-            __msa_st_w(_sum10, outptr0 + 4, 0);
-            outptr0 += 8;
-        }
-        for (; i < size; i++)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-            const signed char* kptr = kernel.channel(p);
-
-            int nn4 = (inch / 4) * maxk;
-            int nn1 = (inch % 4) * maxk;
-
-            v4i32 _sum0 = __msa_fill_w(0);
-
-            if (nn4 > 0)
-            {
-                v4i32 _sum1 = __msa_fill_w(0);
-                v4i32 _sum2 = __msa_fill_w(0);
-                v4i32 _sum3 = __msa_fill_w(0);
-
-                int j = 0;
-                for (; j < nn4; j++)
-                {
-                    __builtin_prefetch(tmpptr + 16);
-                    __builtin_prefetch(kptr + 64);
-                    v16i8 _val = __msa_ld_b(tmpptr, 0);
-                    v8i16 _val16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_val, 0), _val);
-
-                    _val16 = (v8i16)__msa_ilvr_d((v2i64)_val16, (v2i64)_val16);
-
-                    v16i8 _w01 = __msa_ld_b(kptr, 0);
-                    v16i8 _extw01 = __msa_clti_s_b(_w01, 0);
-                    v8i16 _w0 = (v8i16)__msa_ilvr_b(_extw01, _w01);
-                    v8i16 _w1 = (v8i16)__msa_ilvl_b(_extw01, _w01);
-
-                    v8i16 _s0 = __msa_mulv_h(_val16, _w0);
-                    v8i16 _s1 = __msa_mulv_h(_val16, _w1);
-
-                    v8i16 _exts0 = __msa_clti_s_h(_s0, 0);
-                    v8i16 _exts1 = __msa_clti_s_h(_s1, 0);
-                    v4i32 _s0l = (v4i32)__msa_ilvr_h(_exts0, _s0);
-                    v4i32 _s0h = (v4i32)__msa_ilvl_h(_exts0, _s0);
-                    v4i32 _s1l = (v4i32)__msa_ilvr_h(_exts1, _s1);
-                    v4i32 _s1h = (v4i32)__msa_ilvl_h(_exts1, _s1);
-
-                    _sum0 = __msa_addv_w(_sum0, _s0l);
-                    _sum1 = __msa_addv_w(_sum1, _s0h);
-                    _sum2 = __msa_addv_w(_sum2, _s1l);
-                    _sum3 = __msa_addv_w(_sum3, _s1h);
-
-                    tmpptr += 4;
-                    kptr += 16;
-                }
-
-                // transpose 4x4
-                {
-                    v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                    _tmp0 = __msa_ilvr_w(_sum1, _sum0);
-                    _tmp1 = __msa_ilvr_w(_sum3, _sum2);
-                    _tmp2 = __msa_ilvl_w(_sum1, _sum0);
-                    _tmp3 = __msa_ilvl_w(_sum3, _sum2);
-                    _sum0 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                    _sum1 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                    _sum2 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                    _sum3 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-                }
-
-                _sum0 = __msa_addv_w(_sum0, _sum1);
-                _sum2 = __msa_addv_w(_sum2, _sum3);
-                _sum0 = __msa_addv_w(_sum0, _sum2);
-            }
-
-            int j = 0;
-            for (; j < nn1; j++)
-            {
-                v8i16 _val = __msa_fill_h(tmpptr[0]);
-
-                v16i8 _w = __msa_ld_b(kptr, 0);
-                v8i16 _w16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_w, 0), _w);
-
-                v8i16 _s0 = __msa_mulv_h(_val, _w16);
-                v4i32 _s032 = (v4i32)__msa_ilvr_h(__msa_clti_s_h(_s0, 0), _s0);
-
-                _sum0 = __msa_addv_w(_sum0, _s032);
-
-                tmpptr += 1;
-                kptr += 4;
-            }
-
-            __msa_st_w(_sum0, outptr0, 0);
-            outptr0 += 4;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_transform_kernel_pack1to4_int8_msa(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = 4a-4b-maxk-inch/4a-outch/4b
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-    if (inch >= 4)
-        kernel_tm.create(16 * maxk, inch / 4 + inch % 4, outch / 4, (size_t)1u);
-    else
-        kernel_tm.create(4 * maxk, inch, outch / 4, (size_t)1u);
-
-    for (int q = 0; q + 3 < outch; q += 4)
-    {
-        signed char* g00 = kernel_tm.channel(q / 4);
-
-        int p = 0;
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 4; j++)
-                    {
-                        const signed char* k00 = kernel.channel(q + i).row<const signed char>(p + j);
-
-                        g00[0] = k00[k];
-
-                        g00++;
-                    }
-                }
-            }
-        }
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    const signed char* k00 = kernel.channel(q + i).row<const signed char>(p);
-
-                    g00[0] = k00[k];
-
-                    g00++;
-                }
-            }
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_pack1to4_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 1u, 1, opt.workspace_allocator);
-    {
-        const int gap = w * stride_h - outw * stride_w;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            signed char* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const signed char* sptr = img.row<const signed char>(dilation_h * u) + dilation_w * v;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j + 3 < outw; j += 4)
-                        {
-                            ptr[0] = sptr[0];
-                            ptr[1] = sptr[stride_w];
-                            ptr[2] = sptr[stride_w * 2];
-                            ptr[3] = sptr[stride_w * 3];
-
-                            sptr += stride_w * 4;
-                            ptr += 4;
-                        }
-                        for (; j + 1 < outw; j += 2)
-                        {
-                            ptr[0] = sptr[0];
-                            ptr[1] = sptr[stride_w];
-
-                            sptr += stride_w * 2;
-                            ptr += 2;
-                        }
-                        for (; j < outw; j++)
-                        {
-                            ptr[0] = sptr[0];
-
-                            sptr += stride_w;
-                            ptr += 1;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_pack1to4_int8_msa(bottom_im2col, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack4.h
deleted file mode 100644
index a9cbeb18..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack4.h
+++ /dev/null
@@ -1,519 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_pack4_msa(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    // Mat bottom_im2col(size, maxk, inch, 4u * 4, 4, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    // permute
-    Mat tmp;
-    if (size >= 12)
-        tmp.create(12 * maxk, inch, size / 12 + (size % 12) / 8 + (size % 12 % 8) / 4 + (size % 12 % 4) / 2 + size % 12 % 2, 4u * 4, 4, opt.workspace_allocator);
-    else if (size >= 8)
-        tmp.create(8 * maxk, inch, size / 8 + (size % 8) / 4 + (size % 4) / 2 + size % 2, 4u * 4, 4, opt.workspace_allocator);
-    else if (size >= 4)
-        tmp.create(4 * maxk, inch, size / 4 + (size % 4) / 2 + size % 2, 4u * 4, 4, opt.workspace_allocator);
-    else if (size >= 2)
-        tmp.create(2 * maxk, inch, size / 2 + size % 2, 4u * 4, 4, opt.workspace_allocator);
-    else
-        tmp.create(maxk, inch, size, 4u * 4, 4, opt.workspace_allocator);
-    {
-        int remain_size_start = 0;
-        int nn_size = size / 12;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 12;
-
-            float* tmpptr = tmp.channel(i / 12);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    // transpose 4x12
-                    v4f32 _r0 = (v4f32)__msa_ld_w(img0, 0);
-                    v4f32 _r1 = (v4f32)__msa_ld_w(img0 + 4, 0);
-                    v4f32 _r2 = (v4f32)__msa_ld_w(img0 + 4 * 2, 0);
-                    v4f32 _r3 = (v4f32)__msa_ld_w(img0 + 4 * 3, 0);
-                    v4f32 _r4 = (v4f32)__msa_ld_w(img0 + 4 * 4, 0);
-                    v4f32 _r5 = (v4f32)__msa_ld_w(img0 + 4 * 5, 0);
-                    v4f32 _r6 = (v4f32)__msa_ld_w(img0 + 4 * 6, 0);
-                    v4f32 _r7 = (v4f32)__msa_ld_w(img0 + 4 * 7, 0);
-                    v4f32 _r8 = (v4f32)__msa_ld_w(img0 + 4 * 8, 0);
-                    v4f32 _r9 = (v4f32)__msa_ld_w(img0 + 4 * 9, 0);
-                    v4f32 _ra = (v4f32)__msa_ld_w(img0 + 4 * 10, 0);
-                    v4f32 _rb = (v4f32)__msa_ld_w(img0 + 4 * 11, 0);
-
-                    v4i32 _r01r = __msa_ilvr_w((v4i32)_r1, (v4i32)_r0);
-                    v4i32 _r01l = __msa_ilvl_w((v4i32)_r1, (v4i32)_r0);
-                    v4i32 _r23r = __msa_ilvr_w((v4i32)_r3, (v4i32)_r2);
-                    v4i32 _r23l = __msa_ilvl_w((v4i32)_r3, (v4i32)_r2);
-                    v4i32 _r45r = __msa_ilvr_w((v4i32)_r5, (v4i32)_r4);
-                    v4i32 _r45l = __msa_ilvl_w((v4i32)_r5, (v4i32)_r4);
-                    v4i32 _r67r = __msa_ilvr_w((v4i32)_r7, (v4i32)_r6);
-                    v4i32 _r67l = __msa_ilvl_w((v4i32)_r7, (v4i32)_r6);
-                    v4i32 _r89r = __msa_ilvr_w((v4i32)_r9, (v4i32)_r8);
-                    v4i32 _r89l = __msa_ilvl_w((v4i32)_r9, (v4i32)_r8);
-                    v4i32 _rabr = __msa_ilvr_w((v4i32)_rb, (v4i32)_ra);
-                    v4i32 _rabl = __msa_ilvl_w((v4i32)_rb, (v4i32)_ra);
-                    v2i64 _r0123_0 = __msa_ilvr_d((v2i64)_r23r, (v2i64)_r01r);
-                    v2i64 _r0123_1 = __msa_ilvl_d((v2i64)_r23r, (v2i64)_r01r);
-                    v2i64 _r0123_2 = __msa_ilvr_d((v2i64)_r23l, (v2i64)_r01l);
-                    v2i64 _r0123_3 = __msa_ilvl_d((v2i64)_r23l, (v2i64)_r01l);
-                    v2i64 _r4567_0 = __msa_ilvr_d((v2i64)_r67r, (v2i64)_r45r);
-                    v2i64 _r4567_1 = __msa_ilvl_d((v2i64)_r67r, (v2i64)_r45r);
-                    v2i64 _r4567_2 = __msa_ilvr_d((v2i64)_r67l, (v2i64)_r45l);
-                    v2i64 _r4567_3 = __msa_ilvl_d((v2i64)_r67l, (v2i64)_r45l);
-                    v2i64 _r89ab_0 = __msa_ilvr_d((v2i64)_rabr, (v2i64)_r89r);
-                    v2i64 _r89ab_1 = __msa_ilvl_d((v2i64)_rabr, (v2i64)_r89r);
-                    v2i64 _r89ab_2 = __msa_ilvr_d((v2i64)_rabl, (v2i64)_r89l);
-                    v2i64 _r89ab_3 = __msa_ilvl_d((v2i64)_rabl, (v2i64)_r89l);
-
-                    __msa_st_w((v4i32)_r0123_0, tmpptr, 0);
-                    __msa_st_w((v4i32)_r4567_0, tmpptr + 4, 0);
-                    __msa_st_w((v4i32)_r89ab_0, tmpptr + 4 * 2, 0);
-                    __msa_st_w((v4i32)_r0123_1, tmpptr + 4 * 3, 0);
-                    __msa_st_w((v4i32)_r4567_1, tmpptr + 4 * 4, 0);
-                    __msa_st_w((v4i32)_r89ab_1, tmpptr + 4 * 5, 0);
-                    __msa_st_w((v4i32)_r0123_2, tmpptr + 4 * 6, 0);
-                    __msa_st_w((v4i32)_r4567_2, tmpptr + 4 * 7, 0);
-                    __msa_st_w((v4i32)_r89ab_2, tmpptr + 4 * 8, 0);
-                    __msa_st_w((v4i32)_r0123_3, tmpptr + 4 * 9, 0);
-                    __msa_st_w((v4i32)_r4567_3, tmpptr + 4 * 10, 0);
-                    __msa_st_w((v4i32)_r89ab_3, tmpptr + 4 * 11, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 48;
-                }
-            }
-        }
-
-        remain_size_start += nn_size * 12;
-        nn_size = (size - remain_size_start) >> 3;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 8;
-
-            float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    // transpose 4x8
-                    v4f32 _r0 = (v4f32)__msa_ld_w(img0, 0);
-                    v4f32 _r1 = (v4f32)__msa_ld_w(img0 + 4, 0);
-                    v4f32 _r2 = (v4f32)__msa_ld_w(img0 + 4 * 2, 0);
-                    v4f32 _r3 = (v4f32)__msa_ld_w(img0 + 4 * 3, 0);
-                    v4f32 _r4 = (v4f32)__msa_ld_w(img0 + 4 * 4, 0);
-                    v4f32 _r5 = (v4f32)__msa_ld_w(img0 + 4 * 5, 0);
-                    v4f32 _r6 = (v4f32)__msa_ld_w(img0 + 4 * 6, 0);
-                    v4f32 _r7 = (v4f32)__msa_ld_w(img0 + 4 * 7, 0);
-
-                    v4i32 _r01r = __msa_ilvr_w((v4i32)_r1, (v4i32)_r0);
-                    v4i32 _r01l = __msa_ilvl_w((v4i32)_r1, (v4i32)_r0);
-                    v4i32 _r23r = __msa_ilvr_w((v4i32)_r3, (v4i32)_r2);
-                    v4i32 _r23l = __msa_ilvl_w((v4i32)_r3, (v4i32)_r2);
-                    v4i32 _r45r = __msa_ilvr_w((v4i32)_r5, (v4i32)_r4);
-                    v4i32 _r45l = __msa_ilvl_w((v4i32)_r5, (v4i32)_r4);
-                    v4i32 _r67r = __msa_ilvr_w((v4i32)_r7, (v4i32)_r6);
-                    v4i32 _r67l = __msa_ilvl_w((v4i32)_r7, (v4i32)_r6);
-                    v2i64 _r0123_0 = __msa_ilvr_d((v2i64)_r23r, (v2i64)_r01r);
-                    v2i64 _r0123_1 = __msa_ilvl_d((v2i64)_r23r, (v2i64)_r01r);
-                    v2i64 _r0123_2 = __msa_ilvr_d((v2i64)_r23l, (v2i64)_r01l);
-                    v2i64 _r0123_3 = __msa_ilvl_d((v2i64)_r23l, (v2i64)_r01l);
-                    v2i64 _r4567_0 = __msa_ilvr_d((v2i64)_r67r, (v2i64)_r45r);
-                    v2i64 _r4567_1 = __msa_ilvl_d((v2i64)_r67r, (v2i64)_r45r);
-                    v2i64 _r4567_2 = __msa_ilvr_d((v2i64)_r67l, (v2i64)_r45l);
-                    v2i64 _r4567_3 = __msa_ilvl_d((v2i64)_r67l, (v2i64)_r45l);
-
-                    __msa_st_w((v4i32)_r0123_0, tmpptr, 0);
-                    __msa_st_w((v4i32)_r4567_0, tmpptr + 4, 0);
-                    __msa_st_w((v4i32)_r0123_1, tmpptr + 4 * 2, 0);
-                    __msa_st_w((v4i32)_r4567_1, tmpptr + 4 * 3, 0);
-                    __msa_st_w((v4i32)_r0123_2, tmpptr + 4 * 4, 0);
-                    __msa_st_w((v4i32)_r4567_2, tmpptr + 4 * 5, 0);
-                    __msa_st_w((v4i32)_r0123_3, tmpptr + 4 * 6, 0);
-                    __msa_st_w((v4i32)_r4567_3, tmpptr + 4 * 7, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 32;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 3;
-        nn_size = (size - remain_size_start) >> 2;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 4;
-
-            float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    // transpose 4x4
-                    v4f32 _r0 = (v4f32)__msa_ld_w(img0, 0);
-                    v4f32 _r1 = (v4f32)__msa_ld_w(img0 + 4, 0);
-                    v4f32 _r2 = (v4f32)__msa_ld_w(img0 + 4 * 2, 0);
-                    v4f32 _r3 = (v4f32)__msa_ld_w(img0 + 4 * 3, 0);
-
-                    v4i32 _r01r = __msa_ilvr_w((v4i32)_r1, (v4i32)_r0);
-                    v4i32 _r01l = __msa_ilvl_w((v4i32)_r1, (v4i32)_r0);
-                    v4i32 _r23r = __msa_ilvr_w((v4i32)_r3, (v4i32)_r2);
-                    v4i32 _r23l = __msa_ilvl_w((v4i32)_r3, (v4i32)_r2);
-                    v2i64 _r0123_0 = __msa_ilvr_d((v2i64)_r23r, (v2i64)_r01r);
-                    v2i64 _r0123_1 = __msa_ilvl_d((v2i64)_r23r, (v2i64)_r01r);
-                    v2i64 _r0123_2 = __msa_ilvr_d((v2i64)_r23l, (v2i64)_r01l);
-                    v2i64 _r0123_3 = __msa_ilvl_d((v2i64)_r23l, (v2i64)_r01l);
-
-                    __msa_st_w((v4i32)_r0123_0, tmpptr, 0);
-                    __msa_st_w((v4i32)_r0123_1, tmpptr + 4, 0);
-                    __msa_st_w((v4i32)_r0123_2, tmpptr + 4 * 2, 0);
-                    __msa_st_w((v4i32)_r0123_3, tmpptr + 4 * 3, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 16;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 2;
-        nn_size = (size - remain_size_start) >> 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 2;
-
-            float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    // transpose 4x2
-                    v4f32 _r0 = (v4f32)__msa_ld_w(img0, 0);
-                    v4f32 _r1 = (v4f32)__msa_ld_w(img0 + 4, 0);
-
-                    v4i32 _r01_0 = __msa_ilvr_w((v4i32)_r1, (v4i32)_r0);
-                    v4i32 _r01_1 = __msa_ilvl_w((v4i32)_r1, (v4i32)_r0);
-
-                    __msa_st_w((v4i32)_r01_0, tmpptr, 0);
-                    __msa_st_w((v4i32)_r01_1, tmpptr + 4, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 8;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2 + i % 12 % 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    v4f32 _val = (v4f32)__msa_ld_w(img0, 0);
-                    __msa_st_w((v4i32)_val, tmpptr, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 4;
-                }
-            }
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr0 = top_blob.channel(p);
-
-        int i = 0;
-        for (; i + 11 < size; i += 12)
-        {
-            const float* tmpptr = tmp.channel(i / 12);
-            const float* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            v4f32 _sum0 = bias ? (v4f32)__msa_ld_w(bias + p * 4, 0) : (v4f32)__msa_fill_w(0);
-            v4f32 _sum1 = _sum0;
-            v4f32 _sum2 = _sum0;
-            v4f32 _sum3 = _sum0;
-            v4f32 _sum4 = _sum0;
-            v4f32 _sum5 = _sum0;
-            v4f32 _sum6 = _sum0;
-            v4f32 _sum7 = _sum0;
-            v4f32 _sum8 = _sum0;
-            v4f32 _sum9 = _sum0;
-            v4f32 _suma = _sum0;
-            v4f32 _sumb = _sum0;
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 48);
-                __builtin_prefetch(kptr0 + 16);
-                v4i32 _val0123 = __msa_ld_w(tmpptr, 0);
-                v4i32 _val4567 = __msa_ld_w(tmpptr + 4, 0);
-                v4i32 _val89ab = __msa_ld_w(tmpptr + 8, 0);
-                v4f32 _w0 = (v4f32)__msa_ld_w(kptr0, 0);
-                _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_val0123, 0), _w0);
-                _sum1 = __msa_fmadd_w(_sum1, (v4f32)__msa_splati_w(_val0123, 1), _w0);
-                _sum2 = __msa_fmadd_w(_sum2, (v4f32)__msa_splati_w(_val0123, 2), _w0);
-                _sum3 = __msa_fmadd_w(_sum3, (v4f32)__msa_splati_w(_val0123, 3), _w0);
-                _sum4 = __msa_fmadd_w(_sum4, (v4f32)__msa_splati_w(_val4567, 0), _w0);
-                _sum5 = __msa_fmadd_w(_sum5, (v4f32)__msa_splati_w(_val4567, 1), _w0);
-                _sum6 = __msa_fmadd_w(_sum6, (v4f32)__msa_splati_w(_val4567, 2), _w0);
-                _sum7 = __msa_fmadd_w(_sum7, (v4f32)__msa_splati_w(_val4567, 3), _w0);
-                _sum8 = __msa_fmadd_w(_sum8, (v4f32)__msa_splati_w(_val89ab, 0), _w0);
-                _sum9 = __msa_fmadd_w(_sum9, (v4f32)__msa_splati_w(_val89ab, 1), _w0);
-                _suma = __msa_fmadd_w(_suma, (v4f32)__msa_splati_w(_val89ab, 2), _w0);
-                _sumb = __msa_fmadd_w(_sumb, (v4f32)__msa_splati_w(_val89ab, 3), _w0);
-
-                tmpptr += 12;
-                kptr0 += 4;
-            }
-
-            __msa_st_w((v4i32)_sum0, outptr0, 0);
-            __msa_st_w((v4i32)_sum1, outptr0 + 4, 0);
-            __msa_st_w((v4i32)_sum2, outptr0 + 4 * 2, 0);
-            __msa_st_w((v4i32)_sum3, outptr0 + 4 * 3, 0);
-            __msa_st_w((v4i32)_sum4, outptr0 + 4 * 4, 0);
-            __msa_st_w((v4i32)_sum5, outptr0 + 4 * 5, 0);
-            __msa_st_w((v4i32)_sum6, outptr0 + 4 * 6, 0);
-            __msa_st_w((v4i32)_sum7, outptr0 + 4 * 7, 0);
-            __msa_st_w((v4i32)_sum8, outptr0 + 4 * 8, 0);
-            __msa_st_w((v4i32)_sum9, outptr0 + 4 * 9, 0);
-            __msa_st_w((v4i32)_suma, outptr0 + 4 * 10, 0);
-            __msa_st_w((v4i32)_sumb, outptr0 + 4 * 11, 0);
-
-            outptr0 += 4 * 12;
-        }
-        for (; i + 7 < size; i += 8)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8);
-            const float* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            v4f32 _sum0 = bias ? (v4f32)__msa_ld_w(bias + p * 4, 0) : (v4f32)__msa_fill_w(0);
-            v4f32 _sum1 = _sum0;
-            v4f32 _sum2 = _sum0;
-            v4f32 _sum3 = _sum0;
-            v4f32 _sum4 = _sum0;
-            v4f32 _sum5 = _sum0;
-            v4f32 _sum6 = _sum0;
-            v4f32 _sum7 = _sum0;
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 32);
-                __builtin_prefetch(kptr0 + 16);
-                v4i32 _val0123 = __msa_ld_w(tmpptr, 0);
-                v4i32 _val4567 = __msa_ld_w(tmpptr + 4, 0);
-                v4f32 _w0 = (v4f32)__msa_ld_w(kptr0, 0);
-                _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_val0123, 0), _w0);
-                _sum1 = __msa_fmadd_w(_sum1, (v4f32)__msa_splati_w(_val0123, 1), _w0);
-                _sum2 = __msa_fmadd_w(_sum2, (v4f32)__msa_splati_w(_val0123, 2), _w0);
-                _sum3 = __msa_fmadd_w(_sum3, (v4f32)__msa_splati_w(_val0123, 3), _w0);
-                _sum4 = __msa_fmadd_w(_sum4, (v4f32)__msa_splati_w(_val4567, 0), _w0);
-                _sum5 = __msa_fmadd_w(_sum5, (v4f32)__msa_splati_w(_val4567, 1), _w0);
-                _sum6 = __msa_fmadd_w(_sum6, (v4f32)__msa_splati_w(_val4567, 2), _w0);
-                _sum7 = __msa_fmadd_w(_sum7, (v4f32)__msa_splati_w(_val4567, 3), _w0);
-
-                tmpptr += 8;
-                kptr0 += 4;
-            }
-
-            __msa_st_w((v4i32)_sum0, outptr0, 0);
-            __msa_st_w((v4i32)_sum1, outptr0 + 4, 0);
-            __msa_st_w((v4i32)_sum2, outptr0 + 4 * 2, 0);
-            __msa_st_w((v4i32)_sum3, outptr0 + 4 * 3, 0);
-            __msa_st_w((v4i32)_sum4, outptr0 + 4 * 4, 0);
-            __msa_st_w((v4i32)_sum5, outptr0 + 4 * 5, 0);
-            __msa_st_w((v4i32)_sum6, outptr0 + 4 * 6, 0);
-            __msa_st_w((v4i32)_sum7, outptr0 + 4 * 7, 0);
-
-            outptr0 += 4 * 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4);
-            const float* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            v4f32 _sum0 = bias ? (v4f32)__msa_ld_w(bias + p * 4, 0) : (v4f32)__msa_fill_w(0);
-            v4f32 _sum1 = _sum0;
-            v4f32 _sum2 = _sum0;
-            v4f32 _sum3 = _sum0;
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr0 + 16);
-                v4i32 _val0123 = __msa_ld_w(tmpptr, 0);
-                v4f32 _w0 = (v4f32)__msa_ld_w(kptr0, 0);
-                _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_val0123, 0), _w0);
-                _sum1 = __msa_fmadd_w(_sum1, (v4f32)__msa_splati_w(_val0123, 1), _w0);
-                _sum2 = __msa_fmadd_w(_sum2, (v4f32)__msa_splati_w(_val0123, 2), _w0);
-                _sum3 = __msa_fmadd_w(_sum3, (v4f32)__msa_splati_w(_val0123, 3), _w0);
-
-                tmpptr += 4;
-                kptr0 += 4;
-            }
-
-            __msa_st_w((v4i32)_sum0, outptr0, 0);
-            __msa_st_w((v4i32)_sum1, outptr0 + 4, 0);
-            __msa_st_w((v4i32)_sum2, outptr0 + 4 * 2, 0);
-            __msa_st_w((v4i32)_sum3, outptr0 + 4 * 3, 0);
-
-            outptr0 += 4 * 4;
-        }
-        for (; i + 1 < size; i += 2)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2);
-            const float* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            v4f32 _sum0 = bias ? (v4f32)__msa_ld_w(bias + p * 4, 0) : (v4f32)__msa_fill_w(0);
-            v4f32 _sum1 = _sum0;
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 8);
-                __builtin_prefetch(kptr0 + 16);
-                v4f32 _val0 = __msa_fill_w_f32(*tmpptr++);
-                v4f32 _val1 = __msa_fill_w_f32(*tmpptr++);
-                v4f32 _w0 = (v4f32)__msa_ld_w(kptr0, 0);
-                _sum0 = __msa_fmadd_w(_sum0, _val0, _w0);
-                _sum1 = __msa_fmadd_w(_sum1, _val1, _w0);
-
-                kptr0 += 4;
-            }
-
-            __msa_st_w((v4i32)_sum0, outptr0, 0);
-            __msa_st_w((v4i32)_sum1, outptr0 + 4, 0);
-
-            outptr0 += 4 * 2;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2 + i % 12 % 2);
-            const float* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            v4f32 _sum = bias ? (v4f32)__msa_ld_w(bias + p * 4, 0) : (v4f32)__msa_fill_w(0);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 4);
-                __builtin_prefetch(kptr0 + 16);
-                v4f32 _val0 = __msa_fill_w_f32(*tmpptr++);
-                v4f32 _w0 = (v4f32)__msa_ld_w(kptr0, 0);
-                _sum = __msa_fmadd_w(_sum, _val0, _w0);
-
-                kptr0 += 4;
-            }
-
-            __msa_st_w((v4i32)_sum, outptr0, 0);
-
-            outptr0 += 4;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 4u * 4, 4, opt.workspace_allocator);
-    {
-        const int gap = (w * stride_h - outw * stride_w) * 4;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            float* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const float* sptr = img.row<const float>(dilation_h * u) + dilation_w * v * 4;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            v4f32 _val = (v4f32)__msa_ld_w(sptr, 0);
-                            __msa_st_w((v4i32)_val, ptr, 0);
-
-                            sptr += stride_w * 4;
-                            ptr += 4;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_pack4_msa(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack4to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack4to1.h
deleted file mode 100644
index 49472680..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack4to1.h
+++ /dev/null
@@ -1,666 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_pack4to1_msa(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    // Mat bottom_im2col(size, maxk, inch, 4u * 4, 4, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    Mat tmp;
-    if (size >= 12)
-        tmp.create(12 * maxk, inch, size / 12 + (size % 12) / 8 + (size % 12 % 8) / 4 + size % 12 % 4, 4u * 4, 4, opt.workspace_allocator);
-    else if (size >= 8)
-        tmp.create(8 * maxk, inch, size / 8 + (size % 8) / 4 + size % 4, 4u * 4, 4, opt.workspace_allocator);
-    else if (size >= 4)
-        tmp.create(4 * maxk, inch, size / 4 + size % 4, 4u * 4, 4, opt.workspace_allocator);
-    else
-        tmp.create(maxk, inch, size, 4u * 4, 4, opt.workspace_allocator);
-    {
-        int remain_size_start = 0;
-        int nn_size = size / 12;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 12;
-
-            float* tmpptr = tmp.channel(i / 12);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    // transpose 4x12
-                    v4f32 _r0 = (v4f32)__msa_ld_w(img0, 0);
-                    v4f32 _r1 = (v4f32)__msa_ld_w(img0 + 4, 0);
-                    v4f32 _r2 = (v4f32)__msa_ld_w(img0 + 4 * 2, 0);
-                    v4f32 _r3 = (v4f32)__msa_ld_w(img0 + 4 * 3, 0);
-                    v4f32 _r4 = (v4f32)__msa_ld_w(img0 + 4 * 4, 0);
-                    v4f32 _r5 = (v4f32)__msa_ld_w(img0 + 4 * 5, 0);
-                    v4f32 _r6 = (v4f32)__msa_ld_w(img0 + 4 * 6, 0);
-                    v4f32 _r7 = (v4f32)__msa_ld_w(img0 + 4 * 7, 0);
-                    v4f32 _r8 = (v4f32)__msa_ld_w(img0 + 4 * 8, 0);
-                    v4f32 _r9 = (v4f32)__msa_ld_w(img0 + 4 * 9, 0);
-                    v4f32 _ra = (v4f32)__msa_ld_w(img0 + 4 * 10, 0);
-                    v4f32 _rb = (v4f32)__msa_ld_w(img0 + 4 * 11, 0);
-
-                    v4i32 _r01r = __msa_ilvr_w((v4i32)_r1, (v4i32)_r0);
-                    v4i32 _r01l = __msa_ilvl_w((v4i32)_r1, (v4i32)_r0);
-                    v4i32 _r23r = __msa_ilvr_w((v4i32)_r3, (v4i32)_r2);
-                    v4i32 _r23l = __msa_ilvl_w((v4i32)_r3, (v4i32)_r2);
-                    v4i32 _r45r = __msa_ilvr_w((v4i32)_r5, (v4i32)_r4);
-                    v4i32 _r45l = __msa_ilvl_w((v4i32)_r5, (v4i32)_r4);
-                    v4i32 _r67r = __msa_ilvr_w((v4i32)_r7, (v4i32)_r6);
-                    v4i32 _r67l = __msa_ilvl_w((v4i32)_r7, (v4i32)_r6);
-                    v4i32 _r89r = __msa_ilvr_w((v4i32)_r9, (v4i32)_r8);
-                    v4i32 _r89l = __msa_ilvl_w((v4i32)_r9, (v4i32)_r8);
-                    v4i32 _rabr = __msa_ilvr_w((v4i32)_rb, (v4i32)_ra);
-                    v4i32 _rabl = __msa_ilvl_w((v4i32)_rb, (v4i32)_ra);
-                    v2i64 _r0123_0 = __msa_ilvr_d((v2i64)_r23r, (v2i64)_r01r);
-                    v2i64 _r0123_1 = __msa_ilvl_d((v2i64)_r23r, (v2i64)_r01r);
-                    v2i64 _r0123_2 = __msa_ilvr_d((v2i64)_r23l, (v2i64)_r01l);
-                    v2i64 _r0123_3 = __msa_ilvl_d((v2i64)_r23l, (v2i64)_r01l);
-                    v2i64 _r4567_0 = __msa_ilvr_d((v2i64)_r67r, (v2i64)_r45r);
-                    v2i64 _r4567_1 = __msa_ilvl_d((v2i64)_r67r, (v2i64)_r45r);
-                    v2i64 _r4567_2 = __msa_ilvr_d((v2i64)_r67l, (v2i64)_r45l);
-                    v2i64 _r4567_3 = __msa_ilvl_d((v2i64)_r67l, (v2i64)_r45l);
-                    v2i64 _r89ab_0 = __msa_ilvr_d((v2i64)_rabr, (v2i64)_r89r);
-                    v2i64 _r89ab_1 = __msa_ilvl_d((v2i64)_rabr, (v2i64)_r89r);
-                    v2i64 _r89ab_2 = __msa_ilvr_d((v2i64)_rabl, (v2i64)_r89l);
-                    v2i64 _r89ab_3 = __msa_ilvl_d((v2i64)_rabl, (v2i64)_r89l);
-
-                    __msa_st_w((v4i32)_r0123_0, tmpptr, 0);
-                    __msa_st_w((v4i32)_r4567_0, tmpptr + 4, 0);
-                    __msa_st_w((v4i32)_r89ab_0, tmpptr + 4 * 2, 0);
-                    __msa_st_w((v4i32)_r0123_1, tmpptr + 4 * 3, 0);
-                    __msa_st_w((v4i32)_r4567_1, tmpptr + 4 * 4, 0);
-                    __msa_st_w((v4i32)_r89ab_1, tmpptr + 4 * 5, 0);
-                    __msa_st_w((v4i32)_r0123_2, tmpptr + 4 * 6, 0);
-                    __msa_st_w((v4i32)_r4567_2, tmpptr + 4 * 7, 0);
-                    __msa_st_w((v4i32)_r89ab_2, tmpptr + 4 * 8, 0);
-                    __msa_st_w((v4i32)_r0123_3, tmpptr + 4 * 9, 0);
-                    __msa_st_w((v4i32)_r4567_3, tmpptr + 4 * 10, 0);
-                    __msa_st_w((v4i32)_r89ab_3, tmpptr + 4 * 11, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 48;
-                }
-            }
-        }
-
-        remain_size_start += nn_size * 12;
-        nn_size = (size - remain_size_start) >> 3;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 8;
-
-            float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    // transpose 4x8
-                    v4f32 _r0 = (v4f32)__msa_ld_w(img0, 0);
-                    v4f32 _r1 = (v4f32)__msa_ld_w(img0 + 4, 0);
-                    v4f32 _r2 = (v4f32)__msa_ld_w(img0 + 4 * 2, 0);
-                    v4f32 _r3 = (v4f32)__msa_ld_w(img0 + 4 * 3, 0);
-                    v4f32 _r4 = (v4f32)__msa_ld_w(img0 + 4 * 4, 0);
-                    v4f32 _r5 = (v4f32)__msa_ld_w(img0 + 4 * 5, 0);
-                    v4f32 _r6 = (v4f32)__msa_ld_w(img0 + 4 * 6, 0);
-                    v4f32 _r7 = (v4f32)__msa_ld_w(img0 + 4 * 7, 0);
-
-                    v4i32 _r01r = __msa_ilvr_w((v4i32)_r1, (v4i32)_r0);
-                    v4i32 _r01l = __msa_ilvl_w((v4i32)_r1, (v4i32)_r0);
-                    v4i32 _r23r = __msa_ilvr_w((v4i32)_r3, (v4i32)_r2);
-                    v4i32 _r23l = __msa_ilvl_w((v4i32)_r3, (v4i32)_r2);
-                    v4i32 _r45r = __msa_ilvr_w((v4i32)_r5, (v4i32)_r4);
-                    v4i32 _r45l = __msa_ilvl_w((v4i32)_r5, (v4i32)_r4);
-                    v4i32 _r67r = __msa_ilvr_w((v4i32)_r7, (v4i32)_r6);
-                    v4i32 _r67l = __msa_ilvl_w((v4i32)_r7, (v4i32)_r6);
-                    v2i64 _r0123_0 = __msa_ilvr_d((v2i64)_r23r, (v2i64)_r01r);
-                    v2i64 _r0123_1 = __msa_ilvl_d((v2i64)_r23r, (v2i64)_r01r);
-                    v2i64 _r0123_2 = __msa_ilvr_d((v2i64)_r23l, (v2i64)_r01l);
-                    v2i64 _r0123_3 = __msa_ilvl_d((v2i64)_r23l, (v2i64)_r01l);
-                    v2i64 _r4567_0 = __msa_ilvr_d((v2i64)_r67r, (v2i64)_r45r);
-                    v2i64 _r4567_1 = __msa_ilvl_d((v2i64)_r67r, (v2i64)_r45r);
-                    v2i64 _r4567_2 = __msa_ilvr_d((v2i64)_r67l, (v2i64)_r45l);
-                    v2i64 _r4567_3 = __msa_ilvl_d((v2i64)_r67l, (v2i64)_r45l);
-
-                    __msa_st_w((v4i32)_r0123_0, tmpptr, 0);
-                    __msa_st_w((v4i32)_r4567_0, tmpptr + 4, 0);
-                    __msa_st_w((v4i32)_r0123_1, tmpptr + 4 * 2, 0);
-                    __msa_st_w((v4i32)_r4567_1, tmpptr + 4 * 3, 0);
-                    __msa_st_w((v4i32)_r0123_2, tmpptr + 4 * 4, 0);
-                    __msa_st_w((v4i32)_r4567_2, tmpptr + 4 * 5, 0);
-                    __msa_st_w((v4i32)_r0123_3, tmpptr + 4 * 6, 0);
-                    __msa_st_w((v4i32)_r4567_3, tmpptr + 4 * 7, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 32;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 3;
-        nn_size = (size - remain_size_start) >> 2;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 4;
-
-            float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    // transpose 4x4
-                    v4f32 _r0 = (v4f32)__msa_ld_w(img0, 0);
-                    v4f32 _r1 = (v4f32)__msa_ld_w(img0 + 4, 0);
-                    v4f32 _r2 = (v4f32)__msa_ld_w(img0 + 4 * 2, 0);
-                    v4f32 _r3 = (v4f32)__msa_ld_w(img0 + 4 * 3, 0);
-
-                    v4i32 _r01r = __msa_ilvr_w((v4i32)_r1, (v4i32)_r0);
-                    v4i32 _r01l = __msa_ilvl_w((v4i32)_r1, (v4i32)_r0);
-                    v4i32 _r23r = __msa_ilvr_w((v4i32)_r3, (v4i32)_r2);
-                    v4i32 _r23l = __msa_ilvl_w((v4i32)_r3, (v4i32)_r2);
-                    v2i64 _r0123_0 = __msa_ilvr_d((v2i64)_r23r, (v2i64)_r01r);
-                    v2i64 _r0123_1 = __msa_ilvl_d((v2i64)_r23r, (v2i64)_r01r);
-                    v2i64 _r0123_2 = __msa_ilvr_d((v2i64)_r23l, (v2i64)_r01l);
-                    v2i64 _r0123_3 = __msa_ilvl_d((v2i64)_r23l, (v2i64)_r01l);
-
-                    __msa_st_w((v4i32)_r0123_0, tmpptr, 0);
-                    __msa_st_w((v4i32)_r0123_1, tmpptr + 4, 0);
-                    __msa_st_w((v4i32)_r0123_2, tmpptr + 4 * 2, 0);
-                    __msa_st_w((v4i32)_r0123_3, tmpptr + 4 * 3, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 16;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 2;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + i % 12 % 4);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * 4;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    v4f32 _val = (v4f32)__msa_ld_w(img0, 0);
-                    __msa_st_w((v4i32)_val, tmpptr, 0);
-
-                    img0 += size * 4;
-                    tmpptr += 4;
-                }
-            }
-        }
-    }
-
-    int nn_outch = outch / 4;
-    int remain_outch_start = nn_outch * 4;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 4;
-
-        float* outptr0 = top_blob.channel(p);
-        float* outptr1 = top_blob.channel(p + 1);
-        float* outptr2 = top_blob.channel(p + 2);
-        float* outptr3 = top_blob.channel(p + 3);
-
-        const float zeros[4] = {0.f};
-        const float* biasptr = bias ? bias + p : zeros;
-
-        int i = 0;
-        for (; i + 11 < size; i += 12)
-        {
-            const float* tmpptr = tmp.channel(i / 12);
-            const float* kptr0 = kernel.channel(p / 4);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            v4i32 _bias = __msa_ld_w(biasptr, 0);
-            v4f32 _sum0 = (v4f32)__msa_splati_w(_bias, 0);
-            v4f32 _sum1 = (v4f32)__msa_splati_w(_bias, 0);
-            v4f32 _sum2 = (v4f32)__msa_splati_w(_bias, 0);
-            v4f32 _sum3 = (v4f32)__msa_splati_w(_bias, 1);
-            v4f32 _sum4 = (v4f32)__msa_splati_w(_bias, 1);
-            v4f32 _sum5 = (v4f32)__msa_splati_w(_bias, 1);
-            v4f32 _sum6 = (v4f32)__msa_splati_w(_bias, 2);
-            v4f32 _sum7 = (v4f32)__msa_splati_w(_bias, 2);
-            v4f32 _sum8 = (v4f32)__msa_splati_w(_bias, 2);
-            v4f32 _sum9 = (v4f32)__msa_splati_w(_bias, 3);
-            v4f32 _suma = (v4f32)__msa_splati_w(_bias, 3);
-            v4f32 _sumb = (v4f32)__msa_splati_w(_bias, 3);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 48);
-                __builtin_prefetch(kptr0 + 16);
-                v4f32 _val0 = (v4f32)__msa_ld_w(tmpptr, 0);
-                v4f32 _val1 = (v4f32)__msa_ld_w(tmpptr + 4, 0);
-                v4f32 _val2 = (v4f32)__msa_ld_w(tmpptr + 8, 0);
-                v4i32 _w0123 = __msa_ld_w(kptr0, 0);
-                _sum0 = __msa_fmadd_w(_sum0, _val0, (v4f32)__msa_splati_w(_w0123, 0));
-                _sum1 = __msa_fmadd_w(_sum1, _val1, (v4f32)__msa_splati_w(_w0123, 0));
-                _sum2 = __msa_fmadd_w(_sum2, _val2, (v4f32)__msa_splati_w(_w0123, 0));
-                _sum3 = __msa_fmadd_w(_sum3, _val0, (v4f32)__msa_splati_w(_w0123, 1));
-                _sum4 = __msa_fmadd_w(_sum4, _val1, (v4f32)__msa_splati_w(_w0123, 1));
-                _sum5 = __msa_fmadd_w(_sum5, _val2, (v4f32)__msa_splati_w(_w0123, 1));
-                _sum6 = __msa_fmadd_w(_sum6, _val0, (v4f32)__msa_splati_w(_w0123, 2));
-                _sum7 = __msa_fmadd_w(_sum7, _val1, (v4f32)__msa_splati_w(_w0123, 2));
-                _sum8 = __msa_fmadd_w(_sum8, _val2, (v4f32)__msa_splati_w(_w0123, 2));
-                _sum9 = __msa_fmadd_w(_sum9, _val0, (v4f32)__msa_splati_w(_w0123, 3));
-                _suma = __msa_fmadd_w(_suma, _val1, (v4f32)__msa_splati_w(_w0123, 3));
-                _sumb = __msa_fmadd_w(_sumb, _val2, (v4f32)__msa_splati_w(_w0123, 3));
-
-                tmpptr += 12;
-                kptr0 += 4;
-            }
-
-            __msa_st_w((v4i32)_sum0, outptr0, 0);
-            __msa_st_w((v4i32)_sum1, outptr0 + 4, 0);
-            __msa_st_w((v4i32)_sum2, outptr0 + 8, 0);
-            __msa_st_w((v4i32)_sum3, outptr1, 0);
-            __msa_st_w((v4i32)_sum4, outptr1 + 4, 0);
-            __msa_st_w((v4i32)_sum5, outptr1 + 8, 0);
-            __msa_st_w((v4i32)_sum6, outptr2, 0);
-            __msa_st_w((v4i32)_sum7, outptr2 + 4, 0);
-            __msa_st_w((v4i32)_sum8, outptr2 + 8, 0);
-            __msa_st_w((v4i32)_sum9, outptr3, 0);
-            __msa_st_w((v4i32)_suma, outptr3 + 4, 0);
-            __msa_st_w((v4i32)_sumb, outptr3 + 8, 0);
-
-            outptr0 += 12;
-            outptr1 += 12;
-            outptr2 += 12;
-            outptr3 += 12;
-        }
-        for (; i + 7 < size; i += 8)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8);
-            const float* kptr0 = kernel.channel(p / 4);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            v4i32 _bias = __msa_ld_w(biasptr, 0);
-            v4f32 _sum0 = (v4f32)__msa_splati_w(_bias, 0);
-            v4f32 _sum1 = (v4f32)__msa_splati_w(_bias, 0);
-            v4f32 _sum2 = (v4f32)__msa_splati_w(_bias, 1);
-            v4f32 _sum3 = (v4f32)__msa_splati_w(_bias, 1);
-            v4f32 _sum4 = (v4f32)__msa_splati_w(_bias, 2);
-            v4f32 _sum5 = (v4f32)__msa_splati_w(_bias, 2);
-            v4f32 _sum6 = (v4f32)__msa_splati_w(_bias, 3);
-            v4f32 _sum7 = (v4f32)__msa_splati_w(_bias, 3);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 32);
-                __builtin_prefetch(kptr0 + 16);
-                v4f32 _val0 = (v4f32)__msa_ld_w(tmpptr, 0);
-                v4f32 _val1 = (v4f32)__msa_ld_w(tmpptr + 4, 0);
-                v4i32 _w0123 = __msa_ld_w(kptr0, 0);
-                _sum0 = __msa_fmadd_w(_sum0, _val0, (v4f32)__msa_splati_w(_w0123, 0));
-                _sum1 = __msa_fmadd_w(_sum1, _val1, (v4f32)__msa_splati_w(_w0123, 0));
-                _sum2 = __msa_fmadd_w(_sum2, _val0, (v4f32)__msa_splati_w(_w0123, 1));
-                _sum3 = __msa_fmadd_w(_sum3, _val1, (v4f32)__msa_splati_w(_w0123, 1));
-                _sum4 = __msa_fmadd_w(_sum4, _val0, (v4f32)__msa_splati_w(_w0123, 2));
-                _sum5 = __msa_fmadd_w(_sum5, _val1, (v4f32)__msa_splati_w(_w0123, 2));
-                _sum6 = __msa_fmadd_w(_sum6, _val0, (v4f32)__msa_splati_w(_w0123, 3));
-                _sum7 = __msa_fmadd_w(_sum7, _val1, (v4f32)__msa_splati_w(_w0123, 3));
-
-                tmpptr += 8;
-                kptr0 += 4;
-            }
-
-            __msa_st_w((v4i32)_sum0, outptr0, 0);
-            __msa_st_w((v4i32)_sum1, outptr0 + 4, 0);
-            __msa_st_w((v4i32)_sum2, outptr1, 0);
-            __msa_st_w((v4i32)_sum3, outptr1 + 4, 0);
-            __msa_st_w((v4i32)_sum4, outptr2, 0);
-            __msa_st_w((v4i32)_sum5, outptr2 + 4, 0);
-            __msa_st_w((v4i32)_sum6, outptr3, 0);
-            __msa_st_w((v4i32)_sum7, outptr3 + 4, 0);
-
-            outptr0 += 8;
-            outptr1 += 8;
-            outptr2 += 8;
-            outptr3 += 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4);
-            const float* kptr0 = kernel.channel(p / 4);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            v4i32 _bias = __msa_ld_w(biasptr, 0);
-            v4f32 _sum0 = (v4f32)__msa_splati_w(_bias, 0);
-            v4f32 _sum1 = (v4f32)__msa_splati_w(_bias, 1);
-            v4f32 _sum2 = (v4f32)__msa_splati_w(_bias, 2);
-            v4f32 _sum3 = (v4f32)__msa_splati_w(_bias, 3);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr0 + 16);
-                v4f32 _val0 = (v4f32)__msa_ld_w(tmpptr, 0);
-                v4i32 _w0123 = __msa_ld_w(kptr0, 0);
-                _sum0 = __msa_fmadd_w(_sum0, _val0, (v4f32)__msa_splati_w(_w0123, 0));
-                _sum1 = __msa_fmadd_w(_sum1, _val0, (v4f32)__msa_splati_w(_w0123, 1));
-                _sum2 = __msa_fmadd_w(_sum2, _val0, (v4f32)__msa_splati_w(_w0123, 2));
-                _sum3 = __msa_fmadd_w(_sum3, _val0, (v4f32)__msa_splati_w(_w0123, 3));
-
-                tmpptr += 4;
-                kptr0 += 4;
-            }
-
-            __msa_st_w((v4i32)_sum0, outptr0, 0);
-            __msa_st_w((v4i32)_sum1, outptr1, 0);
-            __msa_st_w((v4i32)_sum2, outptr2, 0);
-            __msa_st_w((v4i32)_sum3, outptr3, 0);
-
-            outptr0 += 4;
-            outptr1 += 4;
-            outptr2 += 4;
-            outptr3 += 4;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + i % 12 % 4);
-            const float* kptr0 = kernel.channel(p / 4);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            v4f32 _sum = (v4f32)__msa_ld_w(biasptr, 0);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 4);
-                __builtin_prefetch(kptr0 + 16);
-                v4f32 _val0 = __msa_fill_w_f32(*tmpptr++);
-                v4f32 _w0 = (v4f32)__msa_ld_w(kptr0, 0);
-                _sum = __msa_fmadd_w(_sum, _val0, _w0);
-
-                kptr0 += 4;
-            }
-
-            outptr0[0] = _sum[0];
-            outptr1[0] = _sum[1];
-            outptr2[0] = _sum[2];
-            outptr3[0] = _sum[3];
-
-            outptr0 += 1;
-            outptr1 += 1;
-            outptr2 += 1;
-            outptr3 += 1;
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        float* outptr0 = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        int i = 0;
-        for (; i + 11 < size; i += 12)
-        {
-            const float* tmpptr = tmp.channel(i / 12);
-            const float* kptr0 = kernel.channel(p / 4 + p % 4);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            v4f32 _sum0 = __msa_fill_w_f32(bias0);
-            v4f32 _sum1 = __msa_fill_w_f32(bias0);
-            v4f32 _sum2 = __msa_fill_w_f32(bias0);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 48);
-                __builtin_prefetch(kptr0 + 4);
-                v4f32 _val0 = (v4f32)__msa_ld_w(tmpptr, 0);
-                v4f32 _val1 = (v4f32)__msa_ld_w(tmpptr + 4, 0);
-                v4f32 _val2 = (v4f32)__msa_ld_w(tmpptr + 8, 0);
-                v4f32 _w0 = __msa_fill_w_f32(*kptr0);
-                _sum0 = __msa_fmadd_w(_sum0, _w0, _val0);
-                _sum1 = __msa_fmadd_w(_sum1, _w0, _val1);
-                _sum2 = __msa_fmadd_w(_sum2, _w0, _val2);
-
-                tmpptr += 12;
-                kptr0 += 1;
-            }
-
-            __msa_st_w((v4i32)_sum0, outptr0, 0);
-            __msa_st_w((v4i32)_sum1, outptr0 + 4, 0);
-            __msa_st_w((v4i32)_sum2, outptr0 + 8, 0);
-
-            outptr0 += 12;
-        }
-        for (; i + 7 < size; i += 8)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8);
-            const float* kptr0 = kernel.channel(p / 4 + p % 4);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            v4f32 _sum0 = __msa_fill_w_f32(bias0);
-            v4f32 _sum1 = __msa_fill_w_f32(bias0);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 32);
-                __builtin_prefetch(kptr0 + 4);
-                v4f32 _val0 = (v4f32)__msa_ld_w(tmpptr, 0);
-                v4f32 _val1 = (v4f32)__msa_ld_w(tmpptr + 4, 0);
-                v4f32 _w0 = __msa_fill_w_f32(*kptr0);
-                _sum0 = __msa_fmadd_w(_sum0, _w0, _val0);
-                _sum1 = __msa_fmadd_w(_sum1, _w0, _val1);
-
-                tmpptr += 8;
-                kptr0 += 1;
-            }
-
-            __msa_st_w((v4i32)_sum0, outptr0, 0);
-            __msa_st_w((v4i32)_sum1, outptr0 + 4, 0);
-
-            outptr0 += 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4);
-            const float* kptr0 = kernel.channel(p / 4 + p % 4);
-
-            int nn = inch * maxk * 4; // inch always > 0
-
-            v4f32 _sum0 = __msa_fill_w_f32(bias0);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr0 + 4);
-                v4f32 _val0 = (v4f32)__msa_ld_w(tmpptr, 0);
-                v4f32 _w0 = __msa_fill_w_f32(*kptr0);
-                _sum0 = __msa_fmadd_w(_sum0, _w0, _val0);
-
-                tmpptr += 4;
-                kptr0 += 1;
-            }
-
-            __msa_st_w((v4i32)_sum0, outptr0, 0);
-
-            outptr0 += 4;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + i % 12 % 4);
-            const float* kptr0 = kernel.channel(p / 4 + p % 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = bias0;
-
-            v4f32 _sum0 = (v4f32)__msa_fill_w(0);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 16);
-                __builtin_prefetch(kptr0 + 16);
-                v4f32 _val0 = (v4f32)__msa_ld_w(tmpptr, 0);
-                v4f32 _w0 = (v4f32)__msa_ld_w(kptr0, 0);
-                _sum0 = __msa_fmadd_w(_sum0, _val0, _w0);
-                tmpptr += 4;
-                kptr0 += 4;
-            }
-
-            sum0 += __msa_reduce_fadd_w(_sum0);
-
-            outptr0[0] = sum0;
-
-            outptr0 += 1;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_transform_kernel_pack4to1_msa(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = pb-pa-maxk-inch/pa-outch/pb
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-    kernel_tm.create(4 * 4 * maxk, inch / 4, outch / 4 + outch % 4);
-
-    int q = 0;
-    for (; q + 3 < outch; q += 4)
-    {
-        float* g00 = kernel_tm.channel(q / 4);
-
-        for (int p = 0; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 4; j++)
-                    {
-                        const float* k00 = kernel.channel(q + j).row(p + i);
-
-                        g00[0] = k00[k];
-
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-    for (; q < outch; q++)
-    {
-        const Mat k0 = kernel.channel(q);
-
-        float* g00 = kernel_tm.channel(q / 4 + q % 4);
-
-        for (int p = 0; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int j = 0; j < 4; j++)
-                {
-                    const float* k00 = k0.row(p + j);
-
-                    g00[0] = k00[k];
-
-                    g00++;
-                }
-            }
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_pack4to1_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 4u * 4, 4, opt.workspace_allocator);
-    {
-        const int gap = (w * stride_h - outw * stride_w) * 4;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            float* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const float* sptr = img.row(dilation_h * u) + dilation_w * v * 4;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            v4f32 _val = (v4f32)__msa_ld_w(sptr, 0);
-                            __msa_st_w((v4i32)_val, ptr, 0);
-
-                            sptr += stride_w * 4;
-                            ptr += 4;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_pack4to1_msa(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack8to1_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack8to1_int8.h
deleted file mode 100644
index 6d8e08eb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack8to1_int8.h
+++ /dev/null
@@ -1,458 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_pack8to1_int8_msa(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    // Mat bottom_im2col(size, maxk, inch, 8u, 8, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    // permute
-    Mat tmp;
-    if (size >= 2)
-        tmp.create(2 * maxk, inch, size / 2 + size % 2, 8u, 8, opt.workspace_allocator);
-    else
-        tmp.create(maxk, inch, size, 8u, 8, opt.workspace_allocator);
-    {
-        int remain_size_start = 0;
-        int nn_size = (size - remain_size_start) >> 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 2;
-
-            int64_t* tmpptr = tmp.channel(i / 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const int64_t* img0 = (const int64_t*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    v16i8 _v = __msa_ld_b(img0, 0);
-                    __msa_st_b(_v, tmpptr, 0);
-                    tmpptr += 2;
-                    img0 += size;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            int64_t* tmpptr = tmp.channel(i / 2 + i % 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const int64_t* img0 = (const int64_t*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr += 1;
-                    img0 += size;
-                }
-            }
-        }
-    }
-
-    int nn_outch = 0;
-    int remain_outch_start = 0;
-
-    nn_outch = outch >> 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 4;
-
-        int* outptr0 = top_blob.channel(p);
-        int* outptr1 = top_blob.channel(p + 1);
-        int* outptr2 = top_blob.channel(p + 2);
-        int* outptr3 = top_blob.channel(p + 3);
-
-        int i = 0;
-        for (; i + 1 < size; i += 2)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2);
-            const signed char* kptr = kernel.channel(p / 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            v4i32 _sum00 = __msa_fill_w(0);
-            v4i32 _sum01 = __msa_fill_w(0);
-            v4i32 _sum02 = __msa_fill_w(0);
-            v4i32 _sum03 = __msa_fill_w(0);
-            v4i32 _sum10 = __msa_fill_w(0);
-            v4i32 _sum11 = __msa_fill_w(0);
-            v4i32 _sum12 = __msa_fill_w(0);
-            v4i32 _sum13 = __msa_fill_w(0);
-
-            int j = 0;
-            for (; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 64);
-                __builtin_prefetch(kptr + 128);
-                v16i8 _val01 = __msa_ld_b(tmpptr, 0);
-                v16i8 _extval01 = __msa_clti_s_b(_val01, 0);
-                v8i16 _val0 = (v8i16)__msa_ilvr_b(_extval01, _val01);
-                v8i16 _val1 = (v8i16)__msa_ilvl_b(_extval01, _val01);
-
-                v16i8 _w01 = __msa_ld_b(kptr, 0);
-                v16i8 _w23 = __msa_ld_b(kptr + 16, 0);
-                v16i8 _extw01 = __msa_clti_s_b(_w01, 0);
-                v16i8 _extw23 = __msa_clti_s_b(_w23, 0);
-                v8i16 _w0 = (v8i16)__msa_ilvr_b(_extw01, _w01);
-                v8i16 _w1 = (v8i16)__msa_ilvl_b(_extw01, _w01);
-                v8i16 _w2 = (v8i16)__msa_ilvr_b(_extw23, _w23);
-                v8i16 _w3 = (v8i16)__msa_ilvl_b(_extw23, _w23);
-
-                v8i16 _s00 = __msa_mulv_h(_val0, _w0);
-                v8i16 _s01 = __msa_mulv_h(_val0, _w1);
-                v8i16 _s02 = __msa_mulv_h(_val0, _w2);
-                v8i16 _s03 = __msa_mulv_h(_val0, _w3);
-                v8i16 _s10 = __msa_mulv_h(_val1, _w0);
-                v8i16 _s11 = __msa_mulv_h(_val1, _w1);
-                v8i16 _s12 = __msa_mulv_h(_val1, _w2);
-                v8i16 _s13 = __msa_mulv_h(_val1, _w3);
-
-                _sum00 = __msa_addv_w(_sum00, __msa_hadd_s_w(_s00, _s00));
-                _sum01 = __msa_addv_w(_sum01, __msa_hadd_s_w(_s01, _s01));
-                _sum02 = __msa_addv_w(_sum02, __msa_hadd_s_w(_s02, _s02));
-                _sum03 = __msa_addv_w(_sum03, __msa_hadd_s_w(_s03, _s03));
-                _sum10 = __msa_addv_w(_sum10, __msa_hadd_s_w(_s10, _s10));
-                _sum11 = __msa_addv_w(_sum11, __msa_hadd_s_w(_s11, _s11));
-                _sum12 = __msa_addv_w(_sum12, __msa_hadd_s_w(_s12, _s12));
-                _sum13 = __msa_addv_w(_sum13, __msa_hadd_s_w(_s13, _s13));
-
-                tmpptr += 16;
-                kptr += 32;
-            }
-
-            // transpose 4x4
-            {
-                v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                _tmp0 = __msa_ilvr_w(_sum01, _sum00);
-                _tmp1 = __msa_ilvr_w(_sum03, _sum02);
-                _tmp2 = __msa_ilvl_w(_sum01, _sum00);
-                _tmp3 = __msa_ilvl_w(_sum03, _sum02);
-                _sum00 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                _sum01 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                _sum02 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                _sum03 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-            }
-            {
-                v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                _tmp0 = __msa_ilvr_w(_sum11, _sum10);
-                _tmp1 = __msa_ilvr_w(_sum13, _sum12);
-                _tmp2 = __msa_ilvl_w(_sum11, _sum10);
-                _tmp3 = __msa_ilvl_w(_sum13, _sum12);
-                _sum10 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                _sum11 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                _sum12 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                _sum13 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-            }
-
-            _sum00 = __msa_addv_w(_sum00, _sum01);
-            _sum02 = __msa_addv_w(_sum02, _sum03);
-            _sum10 = __msa_addv_w(_sum10, _sum11);
-            _sum12 = __msa_addv_w(_sum12, _sum13);
-
-            _sum00 = __msa_addv_w(_sum00, _sum02);
-            _sum10 = __msa_addv_w(_sum10, _sum12);
-
-            int sum[8];
-            __msa_st_w(_sum00, sum, 0);
-            __msa_st_w(_sum10, sum + 4, 0);
-
-            outptr0[0] = sum[0];
-            outptr1[0] = sum[1];
-            outptr2[0] = sum[2];
-            outptr3[0] = sum[3];
-            outptr0[1] = sum[4];
-            outptr1[1] = sum[5];
-            outptr2[1] = sum[6];
-            outptr3[1] = sum[7];
-            outptr0 += 2;
-            outptr1 += 2;
-            outptr2 += 2;
-            outptr3 += 2;
-        }
-        for (; i < size; i++)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-            const signed char* kptr = kernel.channel(p / 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            v4i32 _sum0 = __msa_fill_w(0);
-            v4i32 _sum1 = __msa_fill_w(0);
-            v4i32 _sum2 = __msa_fill_w(0);
-            v4i32 _sum3 = __msa_fill_w(0);
-
-            int j = 0;
-            for (; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 32);
-                __builtin_prefetch(kptr + 128);
-                v16i8 _val = __msa_ld_b(tmpptr, 0);
-                v8i16 _val16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_val, 0), _val);
-
-                v16i8 _w01 = __msa_ld_b(kptr, 0);
-                v16i8 _w23 = __msa_ld_b(kptr + 16, 0);
-                v16i8 _extw01 = __msa_clti_s_b(_w01, 0);
-                v16i8 _extw23 = __msa_clti_s_b(_w23, 0);
-                v8i16 _w0 = (v8i16)__msa_ilvr_b(_extw01, _w01);
-                v8i16 _w1 = (v8i16)__msa_ilvl_b(_extw01, _w01);
-                v8i16 _w2 = (v8i16)__msa_ilvr_b(_extw23, _w23);
-                v8i16 _w3 = (v8i16)__msa_ilvl_b(_extw23, _w23);
-
-                v8i16 _s0 = __msa_mulv_h(_val16, _w0);
-                v8i16 _s1 = __msa_mulv_h(_val16, _w1);
-                v8i16 _s2 = __msa_mulv_h(_val16, _w2);
-                v8i16 _s3 = __msa_mulv_h(_val16, _w3);
-
-                _sum0 = __msa_addv_w(_sum0, __msa_hadd_s_w(_s0, _s0));
-                _sum1 = __msa_addv_w(_sum1, __msa_hadd_s_w(_s1, _s1));
-                _sum2 = __msa_addv_w(_sum2, __msa_hadd_s_w(_s2, _s2));
-                _sum3 = __msa_addv_w(_sum3, __msa_hadd_s_w(_s3, _s3));
-
-                tmpptr += 8;
-                kptr += 32;
-            }
-
-            // transpose 4x4
-            {
-                v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                _tmp0 = __msa_ilvr_w(_sum1, _sum0);
-                _tmp1 = __msa_ilvr_w(_sum3, _sum2);
-                _tmp2 = __msa_ilvl_w(_sum1, _sum0);
-                _tmp3 = __msa_ilvl_w(_sum3, _sum2);
-                _sum0 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                _sum1 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                _sum2 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                _sum3 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-            }
-
-            _sum0 = __msa_addv_w(_sum0, _sum1);
-            _sum2 = __msa_addv_w(_sum2, _sum3);
-
-            _sum0 = __msa_addv_w(_sum0, _sum2);
-
-            int sum[4];
-            __msa_st_w(_sum0, sum, 0);
-
-            outptr0[0] = sum[0];
-            outptr1[0] = sum[1];
-            outptr2[0] = sum[2];
-            outptr3[0] = sum[3];
-            outptr0 += 1;
-            outptr1 += 1;
-            outptr2 += 1;
-            outptr3 += 1;
-        }
-    }
-
-    remain_outch_start += nn_outch << 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        int* outptr0 = top_blob.channel(p);
-
-        int i = 0;
-        for (; i + 1 < size; i += 2)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2);
-            const signed char* kptr = kernel.channel(p / 4 + p % 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            v4i32 _sum0 = __msa_fill_w(0);
-            v4i32 _sum1 = __msa_fill_w(0);
-
-            int j = 0;
-            for (; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 64);
-                __builtin_prefetch(kptr + 32);
-                v16i8 _val01 = __msa_ld_b(tmpptr, 0);
-                v16i8 _extval01 = __msa_clti_s_b(_val01, 0);
-                v8i16 _val0 = (v8i16)__msa_ilvr_b(_extval01, _val01);
-                v8i16 _val1 = (v8i16)__msa_ilvl_b(_extval01, _val01);
-
-                v16i8 _w = __msa_ld_b(kptr, 0);
-                v8i16 _w16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_w, 0), _w);
-
-                v8i16 _s0 = __msa_mulv_h(_val0, _w16);
-                v8i16 _s1 = __msa_mulv_h(_val1, _w16);
-
-                _sum0 = __msa_addv_w(_sum0, __msa_hadd_s_w(_s0, _s0));
-                _sum1 = __msa_addv_w(_sum1, __msa_hadd_s_w(_s1, _s1));
-
-                tmpptr += 16;
-                kptr += 8;
-            }
-
-            outptr0[0] = __msa_reduce_add_w(_sum0);
-            outptr0[1] = __msa_reduce_add_w(_sum1);
-            outptr0 += 2;
-        }
-        for (; i < size; i++)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-            const signed char* kptr = kernel.channel(p / 4 + p % 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            v4i32 _sum = __msa_fill_w(0);
-
-            int j = 0;
-            for (; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 32);
-                __builtin_prefetch(kptr + 32);
-                v16i8 _val = __msa_ld_b(tmpptr, 0);
-                v8i16 _val16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_val, 0), _val);
-
-                v16i8 _w = __msa_ld_b(kptr, 0);
-                v8i16 _w16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_w, 0), _w);
-
-                v8i16 _s0 = __msa_mulv_h(_val16, _w16);
-
-                _sum = __msa_addv_w(_sum, __msa_hadd_s_w(_s0, _s0));
-
-                tmpptr += 8;
-                kptr += 8;
-            }
-
-            outptr0[0] = __msa_reduce_add_w(_sum);
-            outptr0 += 1;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_transform_kernel_pack8to1_int8_msa(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = 8a-4b-maxk-inch/8a-outch/4b
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-    if (outch >= 4)
-        kernel_tm.create(32 * maxk, inch / 8, outch / 4 + outch % 4, (size_t)1u);
-    else
-        kernel_tm.create(8 * maxk, inch / 8, outch, (size_t)1u);
-
-    int q = 0;
-    for (; q + 3 < outch; q += 4)
-    {
-        signed char* g00 = kernel_tm.channel(q / 4);
-
-        for (int p = 0; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 8; j++)
-                    {
-                        const signed char* k00 = kernel.channel(q + i).row<const signed char>(p + j);
-
-                        g00[0] = k00[k];
-
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-    // TODO unroll 2
-    for (; q < outch; q++)
-    {
-        signed char* g00 = kernel_tm.channel(q / 4 + q % 4);
-
-        for (int p = 0; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int j = 0; j < 8; j++)
-                {
-                    const signed char* k00 = kernel.channel(q).row<const signed char>(p + j);
-
-                    g00[0] = k00[k];
-
-                    g00++;
-                }
-            }
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_pack8to1_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 8u, 8, opt.workspace_allocator);
-    {
-        const int gap = w * stride_h - outw * stride_w;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            int64_t* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const int64_t* sptr = img.row<const int64_t>(dilation_h * u) + dilation_w * v;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            ptr[0] = sptr[0];
-
-                            sptr += stride_w;
-                            ptr += 1;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_pack8to1_int8_msa(bottom_im2col, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack8to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack8to4_int8.h
deleted file mode 100644
index 0c24a411..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_sgemm_pack8to4_int8.h
+++ /dev/null
@@ -1,324 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_pack8to4_int8_msa(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
-{
-    // Mat bottom_im2col(size, maxk, inch, 8u, 8, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    // permute
-    Mat tmp;
-    if (size >= 2)
-        tmp.create(2 * maxk, inch, size / 2 + size % 2, 8u, 8, opt.workspace_allocator);
-    else
-        tmp.create(maxk, inch, size, 8u, 8, opt.workspace_allocator);
-    {
-        int remain_size_start = 0;
-        int nn_size = size >> 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 2;
-
-            int64_t* tmpptr = tmp.channel(i / 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const int64_t* img0 = (const int64_t*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    v16i8 _v = __msa_ld_b(img0, 0);
-                    __msa_st_b(_v, tmpptr, 0);
-                    tmpptr += 2;
-                    img0 += size;
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            int64_t* tmpptr = tmp.channel(i / 2 + i % 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const int64_t* img0 = (const int64_t*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr += 1;
-                    img0 += size;
-                }
-            }
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        int* outptr0 = top_blob.channel(p);
-
-        int i = 0;
-        for (; i + 1 < size; i += 2)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2);
-            const signed char* kptr = kernel.channel(p);
-
-            int nn = inch * maxk; // inch always > 0
-
-            v4i32 _sum00 = __msa_fill_w(0);
-            v4i32 _sum01 = __msa_fill_w(0);
-            v4i32 _sum02 = __msa_fill_w(0);
-            v4i32 _sum03 = __msa_fill_w(0);
-            v4i32 _sum10 = __msa_fill_w(0);
-            v4i32 _sum11 = __msa_fill_w(0);
-            v4i32 _sum12 = __msa_fill_w(0);
-            v4i32 _sum13 = __msa_fill_w(0);
-
-            int j = 0;
-            for (; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 64);
-                __builtin_prefetch(kptr + 128);
-                v16i8 _val01 = __msa_ld_b(tmpptr, 0);
-                v16i8 _extval01 = __msa_clti_s_b(_val01, 0);
-                v8i16 _val0 = (v8i16)__msa_ilvr_b(_extval01, _val01);
-                v8i16 _val1 = (v8i16)__msa_ilvl_b(_extval01, _val01);
-
-                v16i8 _w01 = __msa_ld_b(kptr, 0);
-                v16i8 _w23 = __msa_ld_b(kptr + 16, 0);
-                v16i8 _extw01 = __msa_clti_s_b(_w01, 0);
-                v16i8 _extw23 = __msa_clti_s_b(_w23, 0);
-                v8i16 _w0 = (v8i16)__msa_ilvr_b(_extw01, _w01);
-                v8i16 _w1 = (v8i16)__msa_ilvl_b(_extw01, _w01);
-                v8i16 _w2 = (v8i16)__msa_ilvr_b(_extw23, _w23);
-                v8i16 _w3 = (v8i16)__msa_ilvl_b(_extw23, _w23);
-
-                v8i16 _s00 = __msa_mulv_h(_val0, _w0);
-                v8i16 _s01 = __msa_mulv_h(_val0, _w1);
-                v8i16 _s02 = __msa_mulv_h(_val0, _w2);
-                v8i16 _s03 = __msa_mulv_h(_val0, _w3);
-                v8i16 _s10 = __msa_mulv_h(_val1, _w0);
-                v8i16 _s11 = __msa_mulv_h(_val1, _w1);
-                v8i16 _s12 = __msa_mulv_h(_val1, _w2);
-                v8i16 _s13 = __msa_mulv_h(_val1, _w3);
-
-                _sum00 = __msa_addv_w(_sum00, __msa_hadd_s_w(_s00, _s00));
-                _sum01 = __msa_addv_w(_sum01, __msa_hadd_s_w(_s01, _s01));
-                _sum02 = __msa_addv_w(_sum02, __msa_hadd_s_w(_s02, _s02));
-                _sum03 = __msa_addv_w(_sum03, __msa_hadd_s_w(_s03, _s03));
-                _sum10 = __msa_addv_w(_sum10, __msa_hadd_s_w(_s10, _s10));
-                _sum11 = __msa_addv_w(_sum11, __msa_hadd_s_w(_s11, _s11));
-                _sum12 = __msa_addv_w(_sum12, __msa_hadd_s_w(_s12, _s12));
-                _sum13 = __msa_addv_w(_sum13, __msa_hadd_s_w(_s13, _s13));
-
-                tmpptr += 16;
-                kptr += 32;
-            }
-
-            // transpose 4x4
-            {
-                v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                _tmp0 = __msa_ilvr_w(_sum01, _sum00);
-                _tmp1 = __msa_ilvr_w(_sum03, _sum02);
-                _tmp2 = __msa_ilvl_w(_sum01, _sum00);
-                _tmp3 = __msa_ilvl_w(_sum03, _sum02);
-                _sum00 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                _sum01 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                _sum02 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                _sum03 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-            }
-            {
-                v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                _tmp0 = __msa_ilvr_w(_sum11, _sum10);
-                _tmp1 = __msa_ilvr_w(_sum13, _sum12);
-                _tmp2 = __msa_ilvl_w(_sum11, _sum10);
-                _tmp3 = __msa_ilvl_w(_sum13, _sum12);
-                _sum10 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                _sum11 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                _sum12 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                _sum13 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-            }
-
-            _sum00 = __msa_addv_w(_sum00, _sum01);
-            _sum02 = __msa_addv_w(_sum02, _sum03);
-            _sum10 = __msa_addv_w(_sum10, _sum11);
-            _sum12 = __msa_addv_w(_sum12, _sum13);
-
-            _sum00 = __msa_addv_w(_sum00, _sum02);
-            _sum10 = __msa_addv_w(_sum10, _sum12);
-
-            __msa_st_w(_sum00, outptr0, 0);
-            __msa_st_w(_sum10, outptr0 + 4, 0);
-            outptr0 += 8;
-        }
-        for (; i < size; i++)
-        {
-            const signed char* tmpptr = tmp.channel(i / 2 + i % 2);
-            const signed char* kptr = kernel.channel(p);
-
-            int nn = inch * maxk; // inch always > 0
-
-            v4i32 _sum0 = __msa_fill_w(0);
-            v4i32 _sum1 = __msa_fill_w(0);
-            v4i32 _sum2 = __msa_fill_w(0);
-            v4i32 _sum3 = __msa_fill_w(0);
-
-            int j = 0;
-            for (; j < nn; j++)
-            {
-                __builtin_prefetch(tmpptr + 32);
-                __builtin_prefetch(kptr + 128);
-                v16i8 _val = __msa_ld_b(tmpptr, 0);
-                v8i16 _val16 = (v8i16)__msa_ilvr_b(__msa_clti_s_b(_val, 0), _val);
-
-                v16i8 _w01 = __msa_ld_b(kptr, 0);
-                v16i8 _w23 = __msa_ld_b(kptr + 16, 0);
-                v16i8 _extw01 = __msa_clti_s_b(_w01, 0);
-                v16i8 _extw23 = __msa_clti_s_b(_w23, 0);
-                v8i16 _w0 = (v8i16)__msa_ilvr_b(_extw01, _w01);
-                v8i16 _w1 = (v8i16)__msa_ilvl_b(_extw01, _w01);
-                v8i16 _w2 = (v8i16)__msa_ilvr_b(_extw23, _w23);
-                v8i16 _w3 = (v8i16)__msa_ilvl_b(_extw23, _w23);
-
-                v8i16 _s0 = __msa_mulv_h(_val16, _w0);
-                v8i16 _s1 = __msa_mulv_h(_val16, _w1);
-                v8i16 _s2 = __msa_mulv_h(_val16, _w2);
-                v8i16 _s3 = __msa_mulv_h(_val16, _w3);
-
-                _sum0 = __msa_addv_w(_sum0, __msa_hadd_s_w(_s0, _s0));
-                _sum1 = __msa_addv_w(_sum1, __msa_hadd_s_w(_s1, _s1));
-                _sum2 = __msa_addv_w(_sum2, __msa_hadd_s_w(_s2, _s2));
-                _sum3 = __msa_addv_w(_sum3, __msa_hadd_s_w(_s3, _s3));
-
-                tmpptr += 8;
-                kptr += 32;
-            }
-
-            // transpose 4x4
-            {
-                v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                _tmp0 = __msa_ilvr_w(_sum1, _sum0);
-                _tmp1 = __msa_ilvr_w(_sum3, _sum2);
-                _tmp2 = __msa_ilvl_w(_sum1, _sum0);
-                _tmp3 = __msa_ilvl_w(_sum3, _sum2);
-                _sum0 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                _sum1 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                _sum2 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                _sum3 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-            }
-
-            _sum0 = __msa_addv_w(_sum0, _sum1);
-            _sum2 = __msa_addv_w(_sum2, _sum3);
-
-            _sum0 = __msa_addv_w(_sum0, _sum2);
-
-            __msa_st_w(_sum0, outptr0, 0);
-            outptr0 += 4;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_transform_kernel_pack8to4_int8_msa(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = 8a-4b-maxk-inch/8a-outch/4b
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-    kernel_tm.create(32 * maxk, inch / 8, outch / 4, (size_t)1u);
-
-    for (int q = 0; q + 3 < outch; q += 4)
-    {
-        signed char* g00 = kernel_tm.channel(q / 4);
-
-        for (int p = 0; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    for (int j = 0; j < 8; j++)
-                    {
-                        const signed char* k00 = kernel.channel(q + i).row<const signed char>(p + j);
-
-                        g00[0] = k00[k];
-
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_pack8to4_int8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 8u, 8, opt.workspace_allocator);
-    {
-        const int gap = w * stride_h - outw * stride_w;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            int64_t* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const int64_t* sptr = img.row<const int64_t>(dilation_h * u) + dilation_w * v;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            ptr[0] = sptr[0];
-
-                            sptr += stride_w;
-                            ptr += 1;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_pack8to4_int8_msa(bottom_im2col, top_blob, kernel, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot.h
deleted file mode 100644
index c4bede4b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot.h
+++ /dev/null
@@ -1,495 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_winograd_dot_msa(Mat& bottom_blob_tm, int outch, const Mat& kernel_tm, Mat& top_blob_tm, const Option& opt)
-{
-    // Mat bottom_blob_tm(tiles, 16/36/64, inch, 4u, opt.workspace_allocator);
-
-    const int tiles = bottom_blob_tm.w;
-    const int batch = bottom_blob_tm.h;
-    const int inch = bottom_blob_tm.c;
-
-    // permute
-    Mat bottom_blob_tm2;
-    if (tiles >= 4)
-        bottom_blob_tm2.create(4 * inch, tiles / 4 + tiles % 4, batch, 4u, opt.workspace_allocator);
-    else
-        bottom_blob_tm2.create(1 * inch, tiles, batch, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int r = 0; r < batch; r++)
-    {
-        Mat tm2 = bottom_blob_tm2.channel(r);
-
-        // tile
-        int i = 0;
-        for (; i + 3 < tiles; i += 4)
-        {
-            float* tmpptr = tm2.row(i / 4);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i);
-
-            for (int q = 0; q < inch; q++)
-            {
-#if __mips_msa
-                __msa_st_w(__msa_ld_w(r0, 0), tmpptr, 0);
-#else
-                tmpptr[0] = r0[0];
-                tmpptr[1] = r0[1];
-                tmpptr[2] = r0[2];
-                tmpptr[3] = r0[3];
-#endif
-
-                r0 += bottom_blob_tm.cstep;
-                tmpptr += 4;
-            }
-        }
-        for (; i < tiles; i++)
-        {
-            float* tmpptr = tm2.row(i / 4 + i % 4);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i);
-
-            for (int q = 0; q < inch; q++)
-            {
-                tmpptr[0] = r0[0];
-
-                r0 += bottom_blob_tm.cstep;
-                tmpptr += 1;
-            }
-        }
-    }
-
-    bottom_blob_tm = Mat();
-    // permute end
-
-    top_blob_tm.create(tiles, batch, outch, 4u, opt.workspace_allocator);
-
-#if __mips_msa
-    int nn_outch = outch >> 3;
-    int remain_outch_start = nn_outch << 3;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 8;
-
-        float* output0_tm = top_blob_tm.channel(p);
-        float* output1_tm = top_blob_tm.channel(p + 1);
-        float* output2_tm = top_blob_tm.channel(p + 2);
-        float* output3_tm = top_blob_tm.channel(p + 3);
-        float* output4_tm = top_blob_tm.channel(p + 4);
-        float* output5_tm = top_blob_tm.channel(p + 5);
-        float* output6_tm = top_blob_tm.channel(p + 6);
-        float* output7_tm = top_blob_tm.channel(p + 7);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 8);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 3 < tiles; i += 4)
-            {
-                const float* r0 = bb2.row(i / 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                v4f32 _sum0 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum1 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum2 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum3 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum4 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum5 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum6 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum7 = (v4f32)__msa_fill_w(0);
-
-                int j = 0;
-                for (; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 16);
-                    __builtin_prefetch(k0 + 32);
-                    v4f32 _val = (v4f32)__msa_ld_w(r0, 0);
-                    v4i32 _w0123 = __msa_ld_w(k0, 0);
-                    v4i32 _w4567 = __msa_ld_w(k0 + 4, 0);
-                    _sum0 = __msa_fmadd_w(_sum0, _val, (v4f32)__msa_splati_w(_w0123, 0));
-                    _sum1 = __msa_fmadd_w(_sum1, _val, (v4f32)__msa_splati_w(_w0123, 1));
-                    _sum2 = __msa_fmadd_w(_sum2, _val, (v4f32)__msa_splati_w(_w0123, 2));
-                    _sum3 = __msa_fmadd_w(_sum3, _val, (v4f32)__msa_splati_w(_w0123, 3));
-                    _sum4 = __msa_fmadd_w(_sum4, _val, (v4f32)__msa_splati_w(_w4567, 0));
-                    _sum5 = __msa_fmadd_w(_sum5, _val, (v4f32)__msa_splati_w(_w4567, 1));
-                    _sum6 = __msa_fmadd_w(_sum6, _val, (v4f32)__msa_splati_w(_w4567, 2));
-                    _sum7 = __msa_fmadd_w(_sum7, _val, (v4f32)__msa_splati_w(_w4567, 3));
-
-                    r0 += 4;
-                    k0 += 8;
-                }
-
-                __msa_st_w((v4i32)_sum0, output0_tm, 0);
-                __msa_st_w((v4i32)_sum1, output1_tm, 0);
-                __msa_st_w((v4i32)_sum2, output2_tm, 0);
-                __msa_st_w((v4i32)_sum3, output3_tm, 0);
-                __msa_st_w((v4i32)_sum4, output4_tm, 0);
-                __msa_st_w((v4i32)_sum5, output5_tm, 0);
-                __msa_st_w((v4i32)_sum6, output6_tm, 0);
-                __msa_st_w((v4i32)_sum7, output7_tm, 0);
-
-                output0_tm += 4;
-                output1_tm += 4;
-                output2_tm += 4;
-                output3_tm += 4;
-                output4_tm += 4;
-                output5_tm += 4;
-                output6_tm += 4;
-                output7_tm += 4;
-            }
-            for (; i < tiles; i++)
-            {
-                const float* r0 = bb2.row(i / 4 + i % 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-                float sum2 = 0.f;
-                float sum3 = 0.f;
-                float sum4 = 0.f;
-                float sum5 = 0.f;
-                float sum6 = 0.f;
-                float sum7 = 0.f;
-
-                int j = 0;
-                for (; j < nn; j++)
-                {
-                    sum0 += r0[0] * k0[0];
-                    sum1 += r0[0] * k0[1];
-                    sum2 += r0[0] * k0[2];
-                    sum3 += r0[0] * k0[3];
-                    sum4 += r0[0] * k0[4];
-                    sum5 += r0[0] * k0[5];
-                    sum6 += r0[0] * k0[6];
-                    sum7 += r0[0] * k0[7];
-
-                    r0 += 1;
-                    k0 += 8;
-                }
-
-                output0_tm[0] = sum0;
-                output1_tm[0] = sum1;
-                output2_tm[0] = sum2;
-                output3_tm[0] = sum3;
-                output4_tm[0] = sum4;
-                output5_tm[0] = sum5;
-                output6_tm[0] = sum6;
-                output7_tm[0] = sum7;
-
-                output0_tm++;
-                output1_tm++;
-                output2_tm++;
-                output3_tm++;
-                output4_tm++;
-                output5_tm++;
-                output6_tm++;
-                output7_tm++;
-            }
-        }
-    }
-
-    nn_outch = (outch - remain_outch_start) >> 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = remain_outch_start + pp * 4;
-
-        float* output0_tm = top_blob_tm.channel(p);
-        float* output1_tm = top_blob_tm.channel(p + 1);
-        float* output2_tm = top_blob_tm.channel(p + 2);
-        float* output3_tm = top_blob_tm.channel(p + 3);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 8 + (p % 8) / 4);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 3 < tiles; i += 4)
-            {
-                const float* r0 = bb2.row(i / 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                v4f32 _sum0 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum1 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum2 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum3 = (v4f32)__msa_fill_w(0);
-
-                int j = 0;
-                for (; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 16);
-                    __builtin_prefetch(k0 + 16);
-                    v4f32 _val = (v4f32)__msa_ld_w(r0, 0);
-                    v4i32 _w0123 = __msa_ld_w(k0, 0);
-                    _sum0 = __msa_fmadd_w(_sum0, _val, (v4f32)__msa_splati_w(_w0123, 0));
-                    _sum1 = __msa_fmadd_w(_sum1, _val, (v4f32)__msa_splati_w(_w0123, 1));
-                    _sum2 = __msa_fmadd_w(_sum2, _val, (v4f32)__msa_splati_w(_w0123, 2));
-                    _sum3 = __msa_fmadd_w(_sum3, _val, (v4f32)__msa_splati_w(_w0123, 3));
-
-                    r0 += 4;
-                    k0 += 4;
-                }
-
-                __msa_st_w((v4i32)_sum0, output0_tm, 0);
-                __msa_st_w((v4i32)_sum1, output1_tm, 0);
-                __msa_st_w((v4i32)_sum2, output2_tm, 0);
-                __msa_st_w((v4i32)_sum3, output3_tm, 0);
-
-                output0_tm += 4;
-                output1_tm += 4;
-                output2_tm += 4;
-                output3_tm += 4;
-            }
-            for (; i < tiles; i++)
-            {
-                const float* r0 = bb2.row(i / 4 + i % 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-                float sum2 = 0.f;
-                float sum3 = 0.f;
-
-                int j = 0;
-                for (; j < nn; j++)
-                {
-                    sum0 += r0[0] * k0[0];
-                    sum1 += r0[0] * k0[1];
-                    sum2 += r0[0] * k0[2];
-                    sum3 += r0[0] * k0[3];
-
-                    r0 += 1;
-                    k0 += 4;
-                }
-
-                output0_tm[0] = sum0;
-                output1_tm[0] = sum1;
-                output2_tm[0] = sum2;
-                output3_tm[0] = sum3;
-
-                output0_tm++;
-                output1_tm++;
-                output2_tm++;
-                output3_tm++;
-            }
-        }
-    }
-
-    remain_outch_start += nn_outch << 2;
-#else
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        float* output0_tm = top_blob_tm.channel(p);
-        float* output1_tm = top_blob_tm.channel(p + 1);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 2);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 3 < tiles; i += 4)
-            {
-                const float* r0 = bb2.row(i / 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum00 = 0.f;
-                float sum01 = 0.f;
-                float sum02 = 0.f;
-                float sum03 = 0.f;
-                float sum10 = 0.f;
-                float sum11 = 0.f;
-                float sum12 = 0.f;
-                float sum13 = 0.f;
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 16);
-                    __builtin_prefetch(k0 + 8);
-                    float w0 = k0[0];
-                    float w1 = k0[1];
-                    sum00 += r0[0] * w0;
-                    sum01 += r0[1] * w0;
-                    sum02 += r0[2] * w0;
-                    sum03 += r0[3] * w0;
-                    sum10 += r0[0] * w1;
-                    sum11 += r0[1] * w1;
-                    sum12 += r0[2] * w1;
-                    sum13 += r0[3] * w1;
-
-                    r0 += 4;
-                    k0 += 2;
-                }
-
-                output0_tm[0] = sum00;
-                output0_tm[1] = sum01;
-                output0_tm[2] = sum02;
-                output0_tm[3] = sum03;
-                output1_tm[0] = sum10;
-                output1_tm[1] = sum11;
-                output1_tm[2] = sum12;
-                output1_tm[3] = sum13;
-
-                output0_tm += 4;
-                output1_tm += 4;
-            }
-            for (; i < tiles; i++)
-            {
-                const float* r0 = bb2.row(i / 4 + i % 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum00 = 0.f;
-                float sum10 = 0.f;
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 4);
-                    __builtin_prefetch(k0 + 8);
-                    float val0 = r0[0];
-                    sum00 += val0 * k0[0];
-                    sum10 += val0 * k0[1];
-
-                    r0 += 1;
-                    k0 += 2;
-                }
-
-                output0_tm[0] = sum00;
-                output1_tm[0] = sum10;
-                output0_tm++;
-                output1_tm++;
-            }
-        }
-    }
-#endif
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        float* output0_tm = top_blob_tm.channel(p);
-
-#if __mips_msa
-        const Mat kernel0_tm = kernel_tm.channel(p / 8 + (p % 8) / 4 + p % 4);
-#else
-        const Mat kernel0_tm = kernel_tm.channel(p / 2 + p % 2);
-#endif
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 3 < tiles; i += 4)
-            {
-                const float* r0 = bb2.row(i / 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                int j = 0;
-#if __mips_msa
-                v4f32 _sum0 = (v4f32)__msa_fill_w(0);
-
-                for (; j < nn; j++)
-                {
-                    _sum0 = __msa_fmadd_w(_sum0, __msa_fill_w_f32(k0[0]), (v4f32)__msa_ld_w(r0, 0));
-                    r0 += 4;
-                    k0++;
-                }
-
-                __msa_st_w((v4i32)_sum0, output0_tm, 0);
-                output0_tm += 4;
-#else  // __mips_msa
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-                float sum2 = 0.f;
-                float sum3 = 0.f;
-
-                for (; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 16);
-                    __builtin_prefetch(k0 + 4);
-                    float w0 = k0[0];
-                    sum0 += r0[0] * w0;
-                    sum1 += r0[1] * w0;
-                    sum2 += r0[2] * w0;
-                    sum3 += r0[3] * w0;
-
-                    r0 += 4;
-                    k0++;
-                }
-
-                output0_tm[0] = sum0;
-                output0_tm[1] = sum1;
-                output0_tm[2] = sum2;
-                output0_tm[3] = sum3;
-                output0_tm += 4;
-#endif // __mips_msa
-            }
-            for (; i < tiles; i++)
-            {
-                const float* r0 = bb2.row(i / 4 + i % 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum = 0.f;
-
-                for (int j = 0; j < nn; j++)
-                {
-                    float w0 = k0[0];
-                    float val0 = r0[0];
-                    sum += val0 * w0;
-
-                    r0 += 1;
-                    k0 += 1;
-                }
-
-                output0_tm[0] = sum;
-                output0_tm += 1;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_int8.h
deleted file mode 100644
index 1d5e6928..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_int8.h
+++ /dev/null
@@ -1,883 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if NCNN_RUNTIME_CPU && NCNN_MMI && !__mips_msa && !__mips_loongson_mmi
-void convolution_winograd_dot_int8_loongson_mmi(Mat& bottom_blob_tm, int outch, const Mat& kernel_tm, Mat& top_blob_tm, const Option& opt);
-#endif
-
-static void convolution_winograd_dot_int8_msa(Mat& bottom_blob_tm, int outch, const Mat& kernel_tm, Mat& top_blob_tm, const Option& opt)
-{
-#if NCNN_RUNTIME_CPU && NCNN_MMI && !__mips_msa && !__mips_loongson_mmi
-    if (ncnn::cpu_support_loongson_mmi())
-    {
-        convolution_winograd_dot_int8_loongson_mmi(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-        return;
-    }
-#endif
-
-    // Mat bottom_blob_tm(tiles, 16/36/64, inch, 2u, 1, opt.workspace_allocator);
-
-    const int tiles = bottom_blob_tm.w;
-    const int batch = bottom_blob_tm.h;
-    const int inch = bottom_blob_tm.c;
-
-    // permute
-    Mat bottom_blob_tm2;
-#if __mips_msa || __mips_loongson_mmi
-    if (inch >= 4)
-    {
-        if (tiles >= 2)
-            bottom_blob_tm2.create(inch / 4 + inch % 4, tiles / 2 + tiles % 2, batch, 16u, 8, opt.workspace_allocator);
-        else // if (tiles >= 1)
-            bottom_blob_tm2.create(inch / 4 + inch % 4, tiles, batch, 8u, 4, opt.workspace_allocator);
-    }
-    else
-#endif // __mips_msa || __mips_loongson_mmi
-    {
-        if (tiles >= 2)
-            bottom_blob_tm2.create(inch, tiles / 2 + tiles % 2, batch, 4u, 2, opt.workspace_allocator);
-        else // if (tiles >= 1)
-            bottom_blob_tm2.create(inch, tiles, batch, 2u, 1, opt.workspace_allocator);
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int r = 0; r < batch; r++)
-    {
-        Mat tm2 = bottom_blob_tm2.channel(r);
-
-        // tile
-        int i = 0;
-        for (; i + 1 < tiles; i += 2)
-        {
-            short* tmpptr = tm2.row<short>(i / 2);
-
-            const short* r0 = (const short*)bottom_blob_tm + r * tiles + i;
-
-            int q = 0;
-#if __mips_msa || __mips_loongson_mmi
-            const short* r1 = (const short*)bottom_blob_tm.channel(1) + r * tiles + i;
-            const short* r2 = (const short*)bottom_blob_tm.channel(2) + r * tiles + i;
-            const short* r3 = (const short*)bottom_blob_tm.channel(3) + r * tiles + i;
-            for (; q + 3 < inch; q += 4)
-            {
-#if __mips_loongson_mmi
-                tmpptr[0] = r0[0];
-                tmpptr[1] = r1[0];
-                tmpptr[2] = r0[1];
-                tmpptr[3] = r1[1];
-                tmpptr[4] = r2[0];
-                tmpptr[5] = r3[0];
-                tmpptr[6] = r2[1];
-                tmpptr[7] = r3[1];
-#else  // __mips_loongson_mmi
-                tmpptr[0] = r0[0];
-                tmpptr[1] = r1[0];
-                tmpptr[2] = r2[0];
-                tmpptr[3] = r3[0];
-                tmpptr[4] = r0[1];
-                tmpptr[5] = r1[1];
-                tmpptr[6] = r2[1];
-                tmpptr[7] = r3[1];
-#endif // __mips_loongson_mmi
-                r0 += bottom_blob_tm.cstep * 4;
-                r1 += bottom_blob_tm.cstep * 4;
-                r2 += bottom_blob_tm.cstep * 4;
-                r3 += bottom_blob_tm.cstep * 4;
-                tmpptr += 8;
-            }
-#endif // __mips_msa || __mips_loongson_mmi
-            for (; q < inch; q++)
-            {
-                tmpptr[0] = r0[0];
-                tmpptr[1] = r0[1];
-                r0 += bottom_blob_tm.cstep;
-                tmpptr += 2;
-            }
-        }
-        for (; i < tiles; i++)
-        {
-            short* tmpptr = tm2.row<short>(i / 2 + i % 2);
-
-            const short* r0 = (const short*)bottom_blob_tm + r * tiles + i;
-
-            int q = 0;
-#if __mips_msa || __mips_loongson_mmi
-            const short* r1 = (const short*)bottom_blob_tm.channel(1) + r * tiles + i;
-            const short* r2 = (const short*)bottom_blob_tm.channel(2) + r * tiles + i;
-            const short* r3 = (const short*)bottom_blob_tm.channel(3) + r * tiles + i;
-            for (; q + 3 < inch; q += 4)
-            {
-                tmpptr[0] = r0[0];
-                tmpptr[1] = r1[0];
-                tmpptr[2] = r2[0];
-                tmpptr[3] = r3[0];
-                r0 += bottom_blob_tm.cstep * 4;
-                r1 += bottom_blob_tm.cstep * 4;
-                r2 += bottom_blob_tm.cstep * 4;
-                r3 += bottom_blob_tm.cstep * 4;
-                tmpptr += 4;
-            }
-#endif // __mips_msa || __mips_loongson_mmi
-            for (; q < inch; q++)
-            {
-                tmpptr[0] = r0[0];
-                r0 += bottom_blob_tm.cstep;
-                tmpptr += 1;
-            }
-        }
-    }
-
-    bottom_blob_tm = Mat();
-    // permute end
-
-    top_blob_tm.create(tiles, batch, outch, 4u, 1, opt.workspace_allocator);
-
-#if __mips_msa
-    int nn_outch = outch >> 2;
-    int remain_outch_start = nn_outch << 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 4;
-
-        int* output0_tm = top_blob_tm.channel(p);
-        int* output1_tm = top_blob_tm.channel(p + 1);
-        int* output2_tm = top_blob_tm.channel(p + 2);
-        int* output3_tm = top_blob_tm.channel(p + 3);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 4);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 1 < tiles; i += 2)
-            {
-                const short* r0 = bb2.row<const short>(i / 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int nn4 = inch / 4;
-                int nn1 = inch % 4;
-
-                v4i32 _sum00 = __msa_fill_w(0);
-                v4i32 _sum10 = __msa_fill_w(0);
-
-                if (nn4 > 0)
-                {
-                    v4i32 _sum01 = __msa_fill_w(0);
-                    v4i32 _sum02 = __msa_fill_w(0);
-                    v4i32 _sum03 = __msa_fill_w(0);
-                    v4i32 _sum11 = __msa_fill_w(0);
-                    v4i32 _sum12 = __msa_fill_w(0);
-                    v4i32 _sum13 = __msa_fill_w(0);
-
-                    int j = 0;
-                    for (; j < nn4; j++)
-                    {
-                        v8i16 _val01 = __msa_ld_h(r0, 0);
-
-                        v8i16 _val0 = (v8i16)__msa_ilvr_d((v2i64)_val01, (v2i64)_val01);
-                        v8i16 _val1 = (v8i16)__msa_ilvl_d((v2i64)_val01, (v2i64)_val01);
-
-                        v8i16 _w0 = __msa_ld_h(k0, 0);
-                        v8i16 _w1 = __msa_ld_h(k0 + 8, 0);
-
-                        v8i16 _extval0 = __msa_clti_s_h(_val0, 0);
-                        v8i16 _extval1 = __msa_clti_s_h(_val1, 0);
-                        v8i16 _extw0 = __msa_clti_s_h(_w0, 0);
-                        v8i16 _extw1 = __msa_clti_s_h(_w1, 0);
-
-                        v4i32 _val0l = (v4i32)__msa_ilvr_h(_extval0, _val0);
-                        v4i32 _val0h = (v4i32)__msa_ilvl_h(_extval0, _val0);
-                        v4i32 _val1l = (v4i32)__msa_ilvr_h(_extval1, _val1);
-                        v4i32 _val1h = (v4i32)__msa_ilvl_h(_extval1, _val1);
-
-                        v4i32 _w0l = (v4i32)__msa_ilvr_h(_extw0, _w0);
-                        v4i32 _w0h = (v4i32)__msa_ilvl_h(_extw0, _w0);
-                        v4i32 _w1l = (v4i32)__msa_ilvr_h(_extw1, _w1);
-                        v4i32 _w1h = (v4i32)__msa_ilvl_h(_extw1, _w1);
-
-                        _sum00 = __msa_maddv_w(_sum00, _val0l, _w0l);
-                        _sum01 = __msa_maddv_w(_sum01, _val0h, _w0h);
-                        _sum02 = __msa_maddv_w(_sum02, _val0l, _w1l);
-                        _sum03 = __msa_maddv_w(_sum03, _val0h, _w1h);
-                        _sum10 = __msa_maddv_w(_sum10, _val1l, _w0l);
-                        _sum11 = __msa_maddv_w(_sum11, _val1h, _w0h);
-                        _sum12 = __msa_maddv_w(_sum12, _val1l, _w1l);
-                        _sum13 = __msa_maddv_w(_sum13, _val1h, _w1h);
-
-                        r0 += 8;
-                        k0 += 16;
-                    }
-
-                    // transpose 4x4
-                    {
-                        v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                        _tmp0 = __msa_ilvr_w(_sum01, _sum00);
-                        _tmp1 = __msa_ilvr_w(_sum03, _sum02);
-                        _tmp2 = __msa_ilvl_w(_sum01, _sum00);
-                        _tmp3 = __msa_ilvl_w(_sum03, _sum02);
-                        _sum00 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                        _sum01 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                        _sum02 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                        _sum03 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-                    }
-                    {
-                        v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                        _tmp0 = __msa_ilvr_w(_sum11, _sum10);
-                        _tmp1 = __msa_ilvr_w(_sum13, _sum12);
-                        _tmp2 = __msa_ilvl_w(_sum11, _sum10);
-                        _tmp3 = __msa_ilvl_w(_sum13, _sum12);
-                        _sum10 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                        _sum11 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                        _sum12 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                        _sum13 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-                    }
-
-                    _sum00 = __msa_addv_w(_sum00, _sum01);
-                    _sum02 = __msa_addv_w(_sum02, _sum03);
-                    _sum10 = __msa_addv_w(_sum10, _sum11);
-                    _sum12 = __msa_addv_w(_sum12, _sum13);
-
-                    _sum00 = __msa_addv_w(_sum00, _sum02);
-                    _sum10 = __msa_addv_w(_sum10, _sum12);
-                }
-
-                for (int j = 0; j < nn1; j++)
-                {
-                    v8i16 _val0 = __msa_fill_h(r0[0]);
-                    v8i16 _val1 = __msa_fill_h(r0[1]);
-                    v8i16 _val = (v8i16)__msa_ilvr_d((v2i64)_val1, (v2i64)_val0);
-
-                    v8i16 _w16 = __msa_ld_h(k0, 0);
-
-                    _w16 = (v8i16)__msa_ilvr_d((v2i64)_w16, (v2i64)_w16);
-
-                    v8i16 _extval = __msa_clti_s_h(_val, 0);
-                    v8i16 _extw16 = __msa_clti_s_h(_w16, 0);
-
-                    v4i32 _vall = (v4i32)__msa_ilvr_h(_extval, _val);
-                    v4i32 _valh = (v4i32)__msa_ilvl_h(_extval, _val);
-                    v4i32 _w0l = (v4i32)__msa_ilvr_h(_extw16, _w16);
-                    v4i32 _w0h = (v4i32)__msa_ilvl_h(_extw16, _w16);
-
-                    _sum00 = __msa_maddv_w(_sum00, _vall, _w0l);
-                    _sum10 = __msa_maddv_w(_sum10, _valh, _w0h);
-
-                    r0 += 2;
-                    k0 += 4;
-                }
-
-                int sum[8];
-                __msa_st_w(_sum00, sum, 0);
-                __msa_st_w(_sum10, sum + 4, 0);
-
-                output0_tm[0] = sum[0];
-                output1_tm[0] = sum[1];
-                output2_tm[0] = sum[2];
-                output3_tm[0] = sum[3];
-                output0_tm[1] = sum[4];
-                output1_tm[1] = sum[5];
-                output2_tm[1] = sum[6];
-                output3_tm[1] = sum[7];
-                output0_tm += 2;
-                output1_tm += 2;
-                output2_tm += 2;
-                output3_tm += 2;
-            }
-            for (; i < tiles; i++)
-            {
-                const short* r0 = bb2.row<const short>(i / 2 + i % 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int nn4 = inch / 4;
-                int nn1 = inch % 4;
-
-                v4i32 _sum0 = __msa_fill_w(0);
-
-                if (nn4 > 0)
-                {
-                    v4i32 _sum1 = __msa_fill_w(0);
-                    v4i32 _sum2 = __msa_fill_w(0);
-                    v4i32 _sum3 = __msa_fill_w(0);
-
-                    int j = 0;
-                    for (; j < nn4; j++)
-                    {
-                        v8i16 _val16 = __msa_ld_h(r0, 0);
-
-                        _val16 = (v8i16)__msa_ilvr_d((v2i64)_val16, (v2i64)_val16);
-
-                        v8i16 _w0 = __msa_ld_h(k0, 0);
-                        v8i16 _w1 = __msa_ld_h(k0 + 8, 0);
-
-                        v8i16 _extval16 = __msa_clti_s_h(_val16, 0);
-                        v8i16 _extw0 = __msa_clti_s_h(_w0, 0);
-                        v8i16 _extw1 = __msa_clti_s_h(_w1, 0);
-
-                        v4i32 _val0l = (v4i32)__msa_ilvr_h(_extval16, _val16);
-                        v4i32 _val0h = (v4i32)__msa_ilvl_h(_extval16, _val16);
-
-                        v4i32 _w0l = (v4i32)__msa_ilvr_h(_extw0, _w0);
-                        v4i32 _w0h = (v4i32)__msa_ilvl_h(_extw0, _w0);
-                        v4i32 _w1l = (v4i32)__msa_ilvr_h(_extw1, _w1);
-                        v4i32 _w1h = (v4i32)__msa_ilvl_h(_extw1, _w1);
-
-                        _sum0 = __msa_maddv_w(_sum0, _val0l, _w0l);
-                        _sum1 = __msa_maddv_w(_sum1, _val0h, _w0h);
-                        _sum2 = __msa_maddv_w(_sum2, _val0l, _w1l);
-                        _sum3 = __msa_maddv_w(_sum3, _val0h, _w1h);
-
-                        r0 += 4;
-                        k0 += 16;
-                    }
-
-                    // transpose 4x4
-                    {
-                        v4i32 _tmp0, _tmp1, _tmp2, _tmp3;
-                        _tmp0 = __msa_ilvr_w(_sum1, _sum0);
-                        _tmp1 = __msa_ilvr_w(_sum3, _sum2);
-                        _tmp2 = __msa_ilvl_w(_sum1, _sum0);
-                        _tmp3 = __msa_ilvl_w(_sum3, _sum2);
-                        _sum0 = (v4i32)__msa_ilvr_d((v2i64)_tmp1, (v2i64)_tmp0);
-                        _sum1 = (v4i32)__msa_ilvl_d((v2i64)_tmp1, (v2i64)_tmp0);
-                        _sum2 = (v4i32)__msa_ilvr_d((v2i64)_tmp3, (v2i64)_tmp2);
-                        _sum3 = (v4i32)__msa_ilvl_d((v2i64)_tmp3, (v2i64)_tmp2);
-                    }
-
-                    _sum0 = __msa_addv_w(_sum0, _sum1);
-                    _sum2 = __msa_addv_w(_sum2, _sum3);
-                    _sum0 = __msa_addv_w(_sum0, _sum2);
-                }
-
-                for (int j = 0; j < nn1; j++)
-                {
-                    v4i32 _val = __msa_fill_w(r0[0]);
-                    v8i16 _w16 = __msa_ld_h(k0, 0);
-
-                    v8i16 _extw16 = __msa_clti_s_h(_w16, 0);
-                    v4i32 _w0l = (v4i32)__msa_ilvr_h(_extw16, _w16);
-
-                    _sum0 = __msa_maddv_w(_sum0, _val, _w0l);
-
-                    r0 += 1;
-                    k0 += 4;
-                }
-
-                int sum[4];
-                __msa_st_w(_sum0, sum, 0);
-
-                output0_tm[0] = sum[0];
-                output1_tm[0] = sum[1];
-                output2_tm[0] = sum[2];
-                output3_tm[0] = sum[3];
-                output0_tm += 1;
-                output1_tm += 1;
-                output2_tm += 1;
-                output3_tm += 1;
-            }
-        }
-    }
-#else // __mips_msa
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        int* output0_tm = top_blob_tm.channel(p);
-        int* output1_tm = top_blob_tm.channel(p + 1);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 2);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 1 < tiles; i += 2)
-            {
-                const short* r0 = bb2.row<const short>(i / 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int sum00 = 0;
-                int sum01 = 0;
-                int sum10 = 0;
-                int sum11 = 0;
-
-#if __mips_loongson_mmi
-                int nn4 = inch / 4;
-                int nn1 = inch % 4;
-
-                if (nn4 > 0)
-                {
-                    int32x2_t _sum0 = __mmi_pzerow_s();
-                    int32x2_t _sum1 = __mmi_pzerow_s();
-
-                    double temp0 = 0;
-                    double temp1 = 0;
-                    double temp2 = 0;
-                    double temp3 = 0;
-                    double temp4 = 0;
-                    double temp5 = 0;
-
-                    uint64_t flag_0x44 = 0x44;
-                    uint64_t flag_0xee = 0xee;
-
-                    int j = 0;
-                    for (; j < nn4; j++)
-                    {
-                        asm volatile(
-                            "ld         $0, 64(%0)      \n" // __builtin_prefetch(r0 + 32);
-                            "ld         $0, 64(%1)      \n" // __builtin_prefetch(k0 + 32);
-
-                            "ldc1       %4, 0(%0)       \n" // int16x4_t _v0 = __mmi_pldh_s(r0);
-                            "ldc1       %5, 8(%0)       \n" // int16x4_t _v1 = __mmi_pldh_s(r0 + 4);
-                            "ldc1       %6, 0(%1)       \n" // int16x4_t _k0 = __mmi_pldh_s(k0);
-                            "ldc1       %7, 8(%1)       \n" // int16x4_t _k1 = __mmi_pldh_s(k0 + 4);
-
-                            "pshufh     %8, %6, %20     \n" // int16x4_t _k0202 = __mmi_pshufh_s(_k0, 0x44);
-                            "pshufh     %9, %6, %21     \n" // int16x4_t _k1313 = __mmi_pshufh_s(_k0, 0xee);
-                            "pshufh     %6, %7, %20     \n" // int16x4_t _k4646 = __mmi_pshufh_s(_k1, 0x44);
-                            "pshufh     %7, %7, %21     \n" // int16x4_t _k5757 = __mmi_pshufh_s(_k1, 0xee);
-
-                            "pmaddhw    %8, %4, %8      \n" // int32x2_t _s0x = __mmi_pmaddhw(_v0, _k0202);
-                            "pmaddhw    %9, %4, %9      \n" // int32x2_t _s1x = __mmi_pmaddhw(_v0, _k1313);
-                            "pmaddhw    %6, %5, %6      \n" // int32x2_t _s0y = __mmi_pmaddhw(_v1, _k4646);
-                            "pmaddhw    %7, %5, %7      \n" // int32x2_t _s1y = __mmi_pmaddhw(_v1, _k5757);
-
-                            "paddw      %2, %2, %8      \n" // _sum0 = __mmi_paddw_s(_sum0, _s0x);
-                            "paddw      %3, %3, %9      \n" // _sum1 = __mmi_paddw_s(_sum1, _s1x);
-                            "paddw      %2, %2, %6      \n" // _sum0 = __mmi_paddw_s(_sum0, _s0y);
-                            "paddw      %3, %3, %7      \n" // _sum1 = __mmi_paddw_s(_sum1, _s1y);
-
-                            : "=r"(r0),    // %0
-                            "=r"(k0),    // %1
-                            "=f"(_sum0), // %2
-                            "=f"(_sum1), // %3
-                            "=f"(temp0), // %4
-                            "=f"(temp1), // %5
-                            "=f"(temp2), // %6
-                            "=f"(temp3), // %7
-                            "=f"(temp4), // %8
-                            "=f"(temp5)  // %9
-                            : "0"(r0),
-                            "1"(k0),
-                            "2"(_sum0),
-                            "3"(_sum1),
-                            "4"(temp0),
-                            "5"(temp1),
-                            "6"(temp2),
-                            "7"(temp3),
-                            "8"(temp4),
-                            "9"(temp5),
-                            "f"(flag_0x44), // %20
-                            "f"(flag_0xee)  // %21
-                            : "memory");
-
-                        r0 += 8;
-                        k0 += 8;
-                    }
-
-                    int sum[4];
-                    __mmi_pstw_s(sum, _sum0);
-                    __mmi_pstw_s(sum + 2, _sum1);
-
-                    sum00 = sum[0];
-                    sum01 = sum[1];
-                    sum10 = sum[2];
-                    sum11 = sum[3];
-                }
-#else  // __mips_loongson_mmi
-                int nn1 = inch;
-#endif // __mips_loongson_mmi
-
-                for (int j = 0; j < nn1; j++)
-                {
-                    signed short val0 = r0[0];
-                    signed short val1 = r0[1];
-                    signed short w0 = k0[0];
-                    signed short w1 = k0[1];
-
-                    sum00 += val0 * w0;
-                    sum01 += val1 * w0;
-                    sum10 += val0 * w1;
-                    sum11 += val1 * w1;
-
-                    r0 += 2;
-                    k0 += 2;
-                }
-
-                output0_tm[0] = sum00;
-                output0_tm[1] = sum01;
-                output1_tm[0] = sum10;
-                output1_tm[1] = sum11;
-                output0_tm += 2;
-                output1_tm += 2;
-            }
-            for (; i < tiles; i++)
-            {
-                const short* r0 = bb2.row<const short>(i / 2 + i % 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int sum0 = 0;
-                int sum1 = 0;
-
-#if __mips_loongson_mmi
-                int nn4 = inch / 4;
-                int nn1 = inch % 4;
-
-                if (nn4 > 0)
-                {
-                    int32x2_t _sum01 = __mmi_pzerow_s();
-
-                    double temp0 = 0;
-                    double temp1 = 0;
-                    double temp2 = 0;
-                    double temp3 = 0;
-                    double temp4 = 0;
-
-                    uint64_t flag_0x44 = 0x44;
-                    uint64_t flag_0xee = 0xee;
-
-                    int j = 0;
-                    for (; j < nn4; j++)
-                    {
-                        asm volatile(
-                            "ld         $0, 32(%0)      \n" // __builtin_prefetch(r0 + 16);
-                            "ld         $0, 64(%1)      \n" // __builtin_prefetch(k0 + 32);
-
-                            "ldc1       %4, 0(%0)       \n" // int16x4_t _v0 = __mmi_pldh_s(r0);
-                            "ldc1       %5, 0(%1)       \n" // int16x4_t _k0 = __mmi_pldh_s(k0);
-                            "ldc1       %6, 8(%1)       \n" // int16x4_t _k1 = __mmi_pldh_s(k0 + 4);
-
-                            "pshufh     %7, %4, %16     \n" // int16x4_t _v0101 = __mmi_pshufh_s(_v0, 0x44);
-                            "pshufh     %4, %4, %17     \n" // int16x4_t _v2323 = __mmi_pshufh_s(_v0, 0xee);
-
-                            "pmaddhw    %5, %7, %5      \n" // int32x2_t _s01x = __mmi_pmaddhw(_v0101, _k0);
-                            "pmaddhw    %6, %4, %6      \n" // int32x2_t _s01y = __mmi_pmaddhw(_v2323, _k1);
-
-                            "paddw      %2, %2, %5      \n" // _sum01 = __mmi_paddw_s(_sum01, _s01x);
-                            "paddw      %2, %2, %6      \n" // _sum01 = __mmi_paddw_s(_sum01, _s01y);
-
-                            : "=r"(r0),     // %0
-                            "=r"(k0),     // %1
-                            "=f"(_sum01), // %2
-                            "=f"(temp0),  // %3
-                            "=f"(temp1),  // %4
-                            "=f"(temp2),  // %5
-                            "=f"(temp3),  // %6
-                            "=f"(temp4)   // %7
-                            : "0"(r0),
-                            "1"(k0),
-                            "2"(_sum01),
-                            "3"(temp0),
-                            "4"(temp1),
-                            "5"(temp2),
-                            "6"(temp3),
-                            "7"(temp4),
-                            "f"(flag_0x44), // %16
-                            "f"(flag_0xee)  // %17
-                            : "memory");
-
-                        r0 += 4;
-                        k0 += 8;
-                    }
-
-                    int sum[2];
-                    __mmi_pstw_s(sum, _sum01);
-
-                    sum0 = sum[0];
-                    sum1 = sum[1];
-                }
-#else  // __mips_loongson_mmi
-                int nn1 = inch;
-#endif // __mips_loongson_mmi
-
-                for (int j = 0; j < nn1; j++)
-                {
-                    signed short val0 = r0[0];
-                    signed short w0 = k0[0];
-                    signed short w1 = k0[1];
-
-                    sum0 += val0 * w0;
-                    sum1 += val0 * w1;
-
-                    r0 += 1;
-                    k0 += 2;
-                }
-
-                output0_tm[0] = sum0;
-                output1_tm[0] = sum1;
-                output0_tm += 1;
-                output1_tm += 1;
-            }
-        }
-    }
-#endif // __mips_msa
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        int* output0_tm = top_blob_tm.channel(p);
-
-#if __mips_msa
-        const Mat kernel0_tm = kernel_tm.channel(p / 4 + p % 4);
-#else
-        const Mat kernel0_tm = kernel_tm.channel(p / 2 + p % 2);
-#endif
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 1 < tiles; i += 2)
-            {
-                const short* r0 = bb2.row<const short>(i / 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int sum0 = 0;
-                int sum1 = 0;
-
-#if __mips_msa
-                int nn4 = inch / 4;
-                int nn1 = inch % 4;
-
-                if (nn4 > 0)
-                {
-                    v4i32 _sum0 = __msa_fill_w(0);
-                    v4i32 _sum1 = __msa_fill_w(0);
-
-                    int j = 0;
-                    for (; j < nn4; j++)
-                    {
-                        v8i16 _val16 = __msa_ld_h(r0, 0);
-
-                        v8i16 _w16 = __msa_ld_h(k0, 0);
-
-                        _w16 = (v8i16)__msa_ilvr_d((v2i64)_w16, (v2i64)_w16);
-
-                        v8i16 _extval16 = __msa_clti_s_h(_val16, 0);
-                        v8i16 _extw16 = __msa_clti_s_h(_w16, 0);
-
-                        v4i32 _val0l = (v4i32)__msa_ilvr_h(_extval16, _val16);
-                        v4i32 _val0h = (v4i32)__msa_ilvl_h(_extval16, _val16);
-
-                        v4i32 _w0l = (v4i32)__msa_ilvr_h(_extw16, _w16);
-                        v4i32 _w0h = (v4i32)__msa_ilvl_h(_extw16, _w16);
-
-                        _sum0 = __msa_maddv_w(_sum0, _val0l, _w0l);
-                        _sum1 = __msa_maddv_w(_sum1, _val0h, _w0h);
-
-                        r0 += 8;
-                        k0 += 4;
-                    }
-
-                    sum0 = _sum0[0] + _sum0[1] + _sum0[2] + _sum0[3];
-                    sum1 = _sum1[0] + _sum1[1] + _sum1[2] + _sum1[3];
-                }
-#elif __mips_loongson_mmi
-                int nn4 = inch / 4;
-                int nn1 = inch % 4;
-
-                if (nn4 > 0)
-                {
-                    int32x2_t _sum01 = __mmi_pzerow_s();
-
-                    double temp0 = 0;
-                    double temp1 = 0;
-                    double temp2 = 0;
-                    double temp3 = 0;
-                    double temp4 = 0;
-
-                    uint64_t flag_0x44 = 0x44;
-                    uint64_t flag_0xee = 0xee;
-
-                    int j = 0;
-                    for (; j < nn4; j++)
-                    {
-                        asm volatile(
-                            "ld         $0, 64(%0)      \n" // __builtin_prefetch(r0 + 32);
-                            "ld         $0, 32(%1)      \n" // __builtin_prefetch(k0 + 16);
-
-                            "ldc1       %4, 0(%0)       \n" // int16x4_t _v0 = __mmi_pldh_s(r0);
-                            "ldc1       %5, 8(%0)       \n" // int16x4_t _v1 = __mmi_pldh_s(r0 + 4);
-                            "ldc1       %6, 0(%1)       \n" // int16x4_t _k0 = __mmi_pldh_s(k0);
-
-                            "pshufh     %7, %6, %16     \n" // int16x4_t _k0202 = __mmi_pshufh_s(_k0, 0x44);
-                            "pshufh     %6, %6, %17     \n" // int16x4_t _k1313 = __mmi_pshufh_s(_k0, 0xee);
-
-                            "pmaddhw    %4, %4, %7      \n" // int32x2_t _s01x = __mmi_pmaddhw(_v0, _k0101);
-                            "pmaddhw    %5, %5, %6      \n" // int32x2_t _s01y = __mmi_pmaddhw(_v1, _k2323);
-
-                            "paddw      %2, %2, %4      \n" // _sum01 = __mmi_paddw_s(_sum01, _s01x);
-                            "paddw      %2, %2, %5      \n" // _sum01 = __mmi_paddw_s(_sum01, _s01y);
-
-                            : "=r"(r0),     // %0
-                            "=r"(k0),     // %1
-                            "=f"(_sum01), // %2
-                            "=f"(temp0),  // %3
-                            "=f"(temp1),  // %4
-                            "=f"(temp2),  // %5
-                            "=f"(temp3),  // %6
-                            "=f"(temp4)   // %7
-                            : "0"(r0),
-                            "1"(k0),
-                            "2"(_sum01),
-                            "3"(temp0),
-                            "4"(temp1),
-                            "5"(temp2),
-                            "6"(temp3),
-                            "7"(temp4),
-                            "f"(flag_0x44), // %16
-                            "f"(flag_0xee)  // %17
-                            : "memory");
-
-                        r0 += 8;
-                        k0 += 4;
-                    }
-
-                    int sum[2];
-                    __mmi_pstw_s(sum, _sum01);
-
-                    sum0 = sum[0];
-                    sum1 = sum[1];
-                }
-#else  // __mips_loongson_mmi
-                int nn1 = inch;
-#endif // __mips_msa || __mips_loongson_mmi
-
-                for (int q = 0; q < nn1; q++)
-                {
-                    signed short val0 = r0[0];
-                    signed short val1 = r0[1];
-                    signed short w = k0[0];
-
-                    sum0 += val0 * w;
-                    sum1 += val1 * w;
-
-                    k0 += 1;
-                    r0 += 2;
-                }
-
-                output0_tm[0] = sum0;
-                output0_tm[1] = sum1;
-                output0_tm += 2;
-            }
-            for (; i < tiles; i++)
-            {
-                const short* r0 = bb2.row<const short>(i / 2 + i % 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int sum = 0;
-
-#if __mips_msa
-                int nn4 = inch / 4;
-                int nn1 = inch % 4;
-
-                if (nn4 > 0)
-                {
-                    v4i32 _sum = __msa_fill_w(0);
-
-                    int j = 0;
-                    for (; j < nn4; j++)
-                    {
-                        v8i16 _val16 = __msa_ld_h(r0, 0);
-                        v8i16 _w16 = __msa_ld_h(k0, 0);
-
-                        v8i16 _extval16 = __msa_clti_s_h(_val16, 0);
-                        v8i16 _extw16 = __msa_clti_s_h(_w16, 0);
-
-                        v4i32 _val0l = (v4i32)__msa_ilvr_h(_extval16, _val16);
-                        v4i32 _w0l = (v4i32)__msa_ilvr_h(_extw16, _w16);
-
-                        _sum = __msa_maddv_w(_sum, _val0l, _w0l);
-
-                        r0 += 4;
-                        k0 += 4;
-                    }
-
-                    sum = _sum[0] + _sum[1] + _sum[2] + _sum[3];
-                }
-#elif __mips_loongson_mmi
-                int nn4 = inch / 4;
-                int nn1 = inch % 4;
-
-                if (nn4 > 0)
-                {
-                    int32x2_t _sum0 = __mmi_pzerow_s();
-
-                    double temp0 = 0;
-                    double temp1 = 0;
-
-                    int j = 0;
-                    for (; j < nn4; j++)
-                    {
-                        asm volatile(
-                            "ld         $0, 32(%0)      \n" // __builtin_prefetch(r0 + 16);
-                            "ld         $0, 32(%1)      \n" // __builtin_prefetch(k0 + 16);
-
-                            "ldc1       %3, 0(%0)       \n" // int16x4_t _v0 = __mmi_pldh_s(r0);
-                            "ldc1       %4, 0(%1)       \n" // int16x4_t _k0 = __mmi_pldh_s(k0);
-
-                            "pmaddhw    %3, %3, %4      \n" // int32x2_t _s0x = __mmi_pmaddhw(_v0, _k0);
-                            "paddw      %2, %2, %3      \n" // _sum0 = __mmi_paddw_s(_sum0, _s0x);
-
-                            : "=r"(r0),    // %0
-                            "=r"(k0),    // %1
-                            "=f"(_sum0), // %2
-                            "=f"(temp0), // %3
-                            "=f"(temp1)  // %4
-                            : "0"(r0),
-                            "1"(k0),
-                            "2"(_sum0),
-                            "3"(temp0),
-                            "4"(temp1)
-                            : "memory");
-
-                        r0 += 4;
-                        k0 += 4;
-                    }
-
-                    int tmp[2];
-                    __mmi_pstw_s(tmp, _sum0);
-
-                    sum = tmp[0] + tmp[1];
-                }
-#else  // __mips_loongson_mmi
-                int nn1 = inch;
-#endif // __mips_msa || __mips_loongson_mmi
-
-                for (int q = 0; q < nn1; q++)
-                {
-                    signed short val = r0[0];
-                    signed short w = k0[0];
-
-                    sum += val * w;
-
-                    k0 += 1;
-                    r0 += 1;
-                }
-
-                output0_tm[0] = sum;
-                output0_tm++;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_pack4.h
deleted file mode 100644
index 2d64ea7b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_pack4.h
+++ /dev/null
@@ -1,448 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_winograd_dot_pack4_msa(Mat& bottom_blob_tm, int outch, const Mat& kernel_tm, Mat& top_blob_tm, const Option& opt)
-{
-    // Mat bottom_blob_tm(tiles, 16/36/64, inch, 16u, 4, opt.workspace_allocator);
-
-    const int tiles = bottom_blob_tm.w;
-    const int batch = bottom_blob_tm.h;
-    const int inch = bottom_blob_tm.c;
-
-    // permute
-    Mat bottom_blob_tm2;
-    if (tiles >= 12)
-        bottom_blob_tm2.create(12 * inch, tiles / 12 + (tiles % 12) / 8 + (tiles % 12 % 8) / 4 + (tiles % 12 % 4) / 2 + tiles % 12 % 2, batch, 16u, 4, opt.workspace_allocator);
-    else if (tiles >= 8)
-        bottom_blob_tm2.create(8 * inch, tiles / 8 + (tiles % 8) / 4 + (tiles % 4) / 2 + tiles % 2, batch, 16u, 4, opt.workspace_allocator);
-    else if (tiles >= 4)
-        bottom_blob_tm2.create(4 * inch, tiles / 4 + (tiles % 4) / 2 + tiles % 2, batch, 16u, 4, opt.workspace_allocator);
-    else if (tiles >= 2)
-        bottom_blob_tm2.create(2 * inch, tiles / 2 + tiles % 2, batch, 16u, 4, opt.workspace_allocator);
-    else // if (tiles >= 1)
-        bottom_blob_tm2.create(1 * inch, tiles, batch, 16u, 4, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int r = 0; r < batch; r++)
-    {
-        Mat tm2 = bottom_blob_tm2.channel(r);
-
-        // tile
-        int i = 0;
-        for (; i + 11 < tiles; i += 12)
-        {
-            float* tmpptr = tm2.row(i / 12);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 4;
-
-            for (int q = 0; q < inch; q++)
-            {
-                // transpose 4x8
-                v4f32 _r0 = (v4f32)__msa_ld_w(r0, 0);
-                v4f32 _r1 = (v4f32)__msa_ld_w(r0 + 4, 0);
-                v4f32 _r2 = (v4f32)__msa_ld_w(r0 + 4 * 2, 0);
-                v4f32 _r3 = (v4f32)__msa_ld_w(r0 + 4 * 3, 0);
-                v4f32 _r4 = (v4f32)__msa_ld_w(r0 + 4 * 4, 0);
-                v4f32 _r5 = (v4f32)__msa_ld_w(r0 + 4 * 5, 0);
-                v4f32 _r6 = (v4f32)__msa_ld_w(r0 + 4 * 6, 0);
-                v4f32 _r7 = (v4f32)__msa_ld_w(r0 + 4 * 7, 0);
-                v4f32 _r8 = (v4f32)__msa_ld_w(r0 + 4 * 8, 0);
-                v4f32 _r9 = (v4f32)__msa_ld_w(r0 + 4 * 9, 0);
-                v4f32 _ra = (v4f32)__msa_ld_w(r0 + 4 * 10, 0);
-                v4f32 _rb = (v4f32)__msa_ld_w(r0 + 4 * 11, 0);
-
-                v4i32 _r01r = __msa_ilvr_w((v4i32)_r1, (v4i32)_r0);
-                v4i32 _r01l = __msa_ilvl_w((v4i32)_r1, (v4i32)_r0);
-                v4i32 _r23r = __msa_ilvr_w((v4i32)_r3, (v4i32)_r2);
-                v4i32 _r23l = __msa_ilvl_w((v4i32)_r3, (v4i32)_r2);
-                v4i32 _r45r = __msa_ilvr_w((v4i32)_r5, (v4i32)_r4);
-                v4i32 _r45l = __msa_ilvl_w((v4i32)_r5, (v4i32)_r4);
-                v4i32 _r67r = __msa_ilvr_w((v4i32)_r7, (v4i32)_r6);
-                v4i32 _r67l = __msa_ilvl_w((v4i32)_r7, (v4i32)_r6);
-                v4i32 _r89r = __msa_ilvr_w((v4i32)_r9, (v4i32)_r8);
-                v4i32 _r89l = __msa_ilvl_w((v4i32)_r9, (v4i32)_r8);
-                v4i32 _rabr = __msa_ilvr_w((v4i32)_rb, (v4i32)_ra);
-                v4i32 _rabl = __msa_ilvl_w((v4i32)_rb, (v4i32)_ra);
-                v2i64 _r0123_0 = __msa_ilvr_d((v2i64)_r23r, (v2i64)_r01r);
-                v2i64 _r0123_1 = __msa_ilvl_d((v2i64)_r23r, (v2i64)_r01r);
-                v2i64 _r0123_2 = __msa_ilvr_d((v2i64)_r23l, (v2i64)_r01l);
-                v2i64 _r0123_3 = __msa_ilvl_d((v2i64)_r23l, (v2i64)_r01l);
-                v2i64 _r4567_0 = __msa_ilvr_d((v2i64)_r67r, (v2i64)_r45r);
-                v2i64 _r4567_1 = __msa_ilvl_d((v2i64)_r67r, (v2i64)_r45r);
-                v2i64 _r4567_2 = __msa_ilvr_d((v2i64)_r67l, (v2i64)_r45l);
-                v2i64 _r4567_3 = __msa_ilvl_d((v2i64)_r67l, (v2i64)_r45l);
-                v2i64 _r89ab_0 = __msa_ilvr_d((v2i64)_rabr, (v2i64)_r89r);
-                v2i64 _r89ab_1 = __msa_ilvl_d((v2i64)_rabr, (v2i64)_r89r);
-                v2i64 _r89ab_2 = __msa_ilvr_d((v2i64)_rabl, (v2i64)_r89l);
-                v2i64 _r89ab_3 = __msa_ilvl_d((v2i64)_rabl, (v2i64)_r89l);
-
-                __msa_st_w((v4i32)_r0123_0, tmpptr, 0);
-                __msa_st_w((v4i32)_r4567_0, tmpptr + 4, 0);
-                __msa_st_w((v4i32)_r89ab_0, tmpptr + 4 * 2, 0);
-                __msa_st_w((v4i32)_r0123_1, tmpptr + 4 * 3, 0);
-                __msa_st_w((v4i32)_r4567_1, tmpptr + 4 * 4, 0);
-                __msa_st_w((v4i32)_r89ab_1, tmpptr + 4 * 5, 0);
-                __msa_st_w((v4i32)_r0123_2, tmpptr + 4 * 6, 0);
-                __msa_st_w((v4i32)_r4567_2, tmpptr + 4 * 7, 0);
-                __msa_st_w((v4i32)_r89ab_2, tmpptr + 4 * 8, 0);
-                __msa_st_w((v4i32)_r0123_3, tmpptr + 4 * 9, 0);
-                __msa_st_w((v4i32)_r4567_3, tmpptr + 4 * 10, 0);
-                __msa_st_w((v4i32)_r89ab_3, tmpptr + 4 * 11, 0);
-
-                r0 += bottom_blob_tm.cstep * 4;
-                tmpptr += 48;
-            }
-        }
-        for (; i + 7 < tiles; i += 8)
-        {
-            float* tmpptr = tm2.row(i / 12 + (i % 12) / 8);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 4;
-
-            for (int q = 0; q < inch; q++)
-            {
-                // transpose 4x8
-                v4f32 _r0 = (v4f32)__msa_ld_w(r0, 0);
-                v4f32 _r1 = (v4f32)__msa_ld_w(r0 + 4, 0);
-                v4f32 _r2 = (v4f32)__msa_ld_w(r0 + 4 * 2, 0);
-                v4f32 _r3 = (v4f32)__msa_ld_w(r0 + 4 * 3, 0);
-                v4f32 _r4 = (v4f32)__msa_ld_w(r0 + 4 * 4, 0);
-                v4f32 _r5 = (v4f32)__msa_ld_w(r0 + 4 * 5, 0);
-                v4f32 _r6 = (v4f32)__msa_ld_w(r0 + 4 * 6, 0);
-                v4f32 _r7 = (v4f32)__msa_ld_w(r0 + 4 * 7, 0);
-
-                v4i32 _r01r = __msa_ilvr_w((v4i32)_r1, (v4i32)_r0);
-                v4i32 _r01l = __msa_ilvl_w((v4i32)_r1, (v4i32)_r0);
-                v4i32 _r23r = __msa_ilvr_w((v4i32)_r3, (v4i32)_r2);
-                v4i32 _r23l = __msa_ilvl_w((v4i32)_r3, (v4i32)_r2);
-                v4i32 _r45r = __msa_ilvr_w((v4i32)_r5, (v4i32)_r4);
-                v4i32 _r45l = __msa_ilvl_w((v4i32)_r5, (v4i32)_r4);
-                v4i32 _r67r = __msa_ilvr_w((v4i32)_r7, (v4i32)_r6);
-                v4i32 _r67l = __msa_ilvl_w((v4i32)_r7, (v4i32)_r6);
-                v2i64 _r0123_0 = __msa_ilvr_d((v2i64)_r23r, (v2i64)_r01r);
-                v2i64 _r0123_1 = __msa_ilvl_d((v2i64)_r23r, (v2i64)_r01r);
-                v2i64 _r0123_2 = __msa_ilvr_d((v2i64)_r23l, (v2i64)_r01l);
-                v2i64 _r0123_3 = __msa_ilvl_d((v2i64)_r23l, (v2i64)_r01l);
-                v2i64 _r4567_0 = __msa_ilvr_d((v2i64)_r67r, (v2i64)_r45r);
-                v2i64 _r4567_1 = __msa_ilvl_d((v2i64)_r67r, (v2i64)_r45r);
-                v2i64 _r4567_2 = __msa_ilvr_d((v2i64)_r67l, (v2i64)_r45l);
-                v2i64 _r4567_3 = __msa_ilvl_d((v2i64)_r67l, (v2i64)_r45l);
-
-                __msa_st_w((v4i32)_r0123_0, tmpptr, 0);
-                __msa_st_w((v4i32)_r4567_0, tmpptr + 4, 0);
-                __msa_st_w((v4i32)_r0123_1, tmpptr + 4 * 2, 0);
-                __msa_st_w((v4i32)_r4567_1, tmpptr + 4 * 3, 0);
-                __msa_st_w((v4i32)_r0123_2, tmpptr + 4 * 4, 0);
-                __msa_st_w((v4i32)_r4567_2, tmpptr + 4 * 5, 0);
-                __msa_st_w((v4i32)_r0123_3, tmpptr + 4 * 6, 0);
-                __msa_st_w((v4i32)_r4567_3, tmpptr + 4 * 7, 0);
-
-                r0 += bottom_blob_tm.cstep * 4;
-                tmpptr += 32;
-            }
-        }
-        for (; i + 3 < tiles; i += 4)
-        {
-            float* tmpptr = tm2.row(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 4;
-
-            for (int q = 0; q < inch; q++)
-            {
-                // transpose 4x4
-                v4f32 _r0 = (v4f32)__msa_ld_w(r0, 0);
-                v4f32 _r1 = (v4f32)__msa_ld_w(r0 + 4, 0);
-                v4f32 _r2 = (v4f32)__msa_ld_w(r0 + 4 * 2, 0);
-                v4f32 _r3 = (v4f32)__msa_ld_w(r0 + 4 * 3, 0);
-
-                v4i32 _r01r = __msa_ilvr_w((v4i32)_r1, (v4i32)_r0);
-                v4i32 _r01l = __msa_ilvl_w((v4i32)_r1, (v4i32)_r0);
-                v4i32 _r23r = __msa_ilvr_w((v4i32)_r3, (v4i32)_r2);
-                v4i32 _r23l = __msa_ilvl_w((v4i32)_r3, (v4i32)_r2);
-                v2i64 _r0123_0 = __msa_ilvr_d((v2i64)_r23r, (v2i64)_r01r);
-                v2i64 _r0123_1 = __msa_ilvl_d((v2i64)_r23r, (v2i64)_r01r);
-                v2i64 _r0123_2 = __msa_ilvr_d((v2i64)_r23l, (v2i64)_r01l);
-                v2i64 _r0123_3 = __msa_ilvl_d((v2i64)_r23l, (v2i64)_r01l);
-
-                __msa_st_w((v4i32)_r0123_0, tmpptr, 0);
-                __msa_st_w((v4i32)_r0123_1, tmpptr + 4, 0);
-                __msa_st_w((v4i32)_r0123_2, tmpptr + 4 * 2, 0);
-                __msa_st_w((v4i32)_r0123_3, tmpptr + 4 * 3, 0);
-
-                r0 += bottom_blob_tm.cstep * 4;
-                tmpptr += 16;
-            }
-        }
-        for (; i + 1 < tiles; i += 2)
-        {
-            float* tmpptr = tm2.row(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 4;
-
-            for (int q = 0; q < inch; q++)
-            {
-                // transpose 4x2
-                v4f32 _r0 = (v4f32)__msa_ld_w(r0, 0);
-                v4f32 _r1 = (v4f32)__msa_ld_w(r0 + 4, 0);
-
-                v4i32 _r01_0 = __msa_ilvr_w((v4i32)_r1, (v4i32)_r0);
-                v4i32 _r01_1 = __msa_ilvl_w((v4i32)_r1, (v4i32)_r0);
-
-                __msa_st_w((v4i32)_r01_0, tmpptr, 0);
-                __msa_st_w((v4i32)_r01_1, tmpptr + 4, 0);
-
-                r0 += bottom_blob_tm.cstep * 4;
-                tmpptr += 8;
-            }
-        }
-        for (; i < tiles; i++)
-        {
-            float* tmpptr = tm2.row(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2 + i % 12 % 2);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 4;
-
-            for (int q = 0; q < inch; q++)
-            {
-                v4f32 _val = (v4f32)__msa_ld_w(r0, 0);
-                __msa_st_w((v4i32)_val, tmpptr, 0);
-
-                r0 += bottom_blob_tm.cstep * 4;
-                tmpptr += 4;
-            }
-        }
-    }
-
-    bottom_blob_tm = Mat();
-    // permute end
-
-    top_blob_tm.create(tiles, batch, outch, 16u, 4, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* output0_tm = top_blob_tm.channel(p);
-
-        const Mat kernel0_tm = kernel_tm.channel(p);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 11 < tiles; i += 12)
-            {
-                const float* r0 = bb2.row(i / 12);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch * 4; // inch always > 0
-
-                v4f32 _sum0 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum1 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum2 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum3 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum4 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum5 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum6 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum7 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum8 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum9 = (v4f32)__msa_fill_w(0);
-                v4f32 _suma = (v4f32)__msa_fill_w(0);
-                v4f32 _sumb = (v4f32)__msa_fill_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 48);
-                    __builtin_prefetch(k0 + 16);
-                    v4i32 _val0123 = __msa_ld_w(r0, 0);
-                    v4i32 _val4567 = __msa_ld_w(r0 + 4, 0);
-                    v4i32 _val89ab = __msa_ld_w(r0 + 8, 0);
-                    v4f32 _w0 = (v4f32)__msa_ld_w(k0, 0);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_val0123, 0), _w0);
-                    _sum1 = __msa_fmadd_w(_sum1, (v4f32)__msa_splati_w(_val0123, 1), _w0);
-                    _sum2 = __msa_fmadd_w(_sum2, (v4f32)__msa_splati_w(_val0123, 2), _w0);
-                    _sum3 = __msa_fmadd_w(_sum3, (v4f32)__msa_splati_w(_val0123, 3), _w0);
-                    _sum4 = __msa_fmadd_w(_sum4, (v4f32)__msa_splati_w(_val4567, 0), _w0);
-                    _sum5 = __msa_fmadd_w(_sum5, (v4f32)__msa_splati_w(_val4567, 1), _w0);
-                    _sum6 = __msa_fmadd_w(_sum6, (v4f32)__msa_splati_w(_val4567, 2), _w0);
-                    _sum7 = __msa_fmadd_w(_sum7, (v4f32)__msa_splati_w(_val4567, 3), _w0);
-                    _sum8 = __msa_fmadd_w(_sum8, (v4f32)__msa_splati_w(_val89ab, 0), _w0);
-                    _sum9 = __msa_fmadd_w(_sum9, (v4f32)__msa_splati_w(_val89ab, 1), _w0);
-                    _suma = __msa_fmadd_w(_suma, (v4f32)__msa_splati_w(_val89ab, 2), _w0);
-                    _sumb = __msa_fmadd_w(_sumb, (v4f32)__msa_splati_w(_val89ab, 3), _w0);
-
-                    r0 += 12;
-                    k0 += 4;
-                }
-
-                __msa_st_w((v4i32)_sum0, output0_tm, 0);
-                __msa_st_w((v4i32)_sum1, output0_tm + 4, 0);
-                __msa_st_w((v4i32)_sum2, output0_tm + 4 * 2, 0);
-                __msa_st_w((v4i32)_sum3, output0_tm + 4 * 3, 0);
-                __msa_st_w((v4i32)_sum4, output0_tm + 4 * 4, 0);
-                __msa_st_w((v4i32)_sum5, output0_tm + 4 * 5, 0);
-                __msa_st_w((v4i32)_sum6, output0_tm + 4 * 6, 0);
-                __msa_st_w((v4i32)_sum7, output0_tm + 4 * 7, 0);
-                __msa_st_w((v4i32)_sum8, output0_tm + 4 * 8, 0);
-                __msa_st_w((v4i32)_sum9, output0_tm + 4 * 9, 0);
-                __msa_st_w((v4i32)_suma, output0_tm + 4 * 10, 0);
-                __msa_st_w((v4i32)_sumb, output0_tm + 4 * 11, 0);
-
-                output0_tm += 4 * 12;
-            }
-            for (; i + 7 < tiles; i += 8)
-            {
-                const float* r0 = bb2.row(i / 12 + (i % 12) / 8);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch * 4; // inch always > 0
-
-                v4f32 _sum0 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum1 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum2 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum3 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum4 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum5 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum6 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum7 = (v4f32)__msa_fill_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 32);
-                    __builtin_prefetch(k0 + 16);
-                    v4i32 _val0123 = __msa_ld_w(r0, 0);
-                    v4i32 _val4567 = __msa_ld_w(r0 + 4, 0);
-                    v4f32 _w0 = (v4f32)__msa_ld_w(k0, 0);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_val0123, 0), _w0);
-                    _sum1 = __msa_fmadd_w(_sum1, (v4f32)__msa_splati_w(_val0123, 1), _w0);
-                    _sum2 = __msa_fmadd_w(_sum2, (v4f32)__msa_splati_w(_val0123, 2), _w0);
-                    _sum3 = __msa_fmadd_w(_sum3, (v4f32)__msa_splati_w(_val0123, 3), _w0);
-                    _sum4 = __msa_fmadd_w(_sum4, (v4f32)__msa_splati_w(_val4567, 0), _w0);
-                    _sum5 = __msa_fmadd_w(_sum5, (v4f32)__msa_splati_w(_val4567, 1), _w0);
-                    _sum6 = __msa_fmadd_w(_sum6, (v4f32)__msa_splati_w(_val4567, 2), _w0);
-                    _sum7 = __msa_fmadd_w(_sum7, (v4f32)__msa_splati_w(_val4567, 3), _w0);
-
-                    r0 += 8;
-                    k0 += 4;
-                }
-
-                __msa_st_w((v4i32)_sum0, output0_tm, 0);
-                __msa_st_w((v4i32)_sum1, output0_tm + 4, 0);
-                __msa_st_w((v4i32)_sum2, output0_tm + 4 * 2, 0);
-                __msa_st_w((v4i32)_sum3, output0_tm + 4 * 3, 0);
-                __msa_st_w((v4i32)_sum4, output0_tm + 4 * 4, 0);
-                __msa_st_w((v4i32)_sum5, output0_tm + 4 * 5, 0);
-                __msa_st_w((v4i32)_sum6, output0_tm + 4 * 6, 0);
-                __msa_st_w((v4i32)_sum7, output0_tm + 4 * 7, 0);
-
-                output0_tm += 4 * 8;
-            }
-            for (; i + 3 < tiles; i += 4)
-            {
-                const float* r0 = bb2.row(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch * 4; // inch always > 0
-
-                v4f32 _sum0 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum1 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum2 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum3 = (v4f32)__msa_fill_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 16);
-                    __builtin_prefetch(k0 + 16);
-                    v4i32 _val0123 = __msa_ld_w(r0, 0);
-                    v4f32 _w0 = (v4f32)__msa_ld_w(k0, 0);
-                    _sum0 = __msa_fmadd_w(_sum0, (v4f32)__msa_splati_w(_val0123, 0), _w0);
-                    _sum1 = __msa_fmadd_w(_sum1, (v4f32)__msa_splati_w(_val0123, 1), _w0);
-                    _sum2 = __msa_fmadd_w(_sum2, (v4f32)__msa_splati_w(_val0123, 2), _w0);
-                    _sum3 = __msa_fmadd_w(_sum3, (v4f32)__msa_splati_w(_val0123, 3), _w0);
-
-                    r0 += 4;
-                    k0 += 4;
-                }
-
-                __msa_st_w((v4i32)_sum0, output0_tm, 0);
-                __msa_st_w((v4i32)_sum1, output0_tm + 4, 0);
-                __msa_st_w((v4i32)_sum2, output0_tm + 4 * 2, 0);
-                __msa_st_w((v4i32)_sum3, output0_tm + 4 * 3, 0);
-
-                output0_tm += 4 * 4;
-            }
-            for (; i + 1 < tiles; i += 2)
-            {
-                const float* r0 = bb2.row(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch * 4; // inch always > 0
-
-                v4f32 _sum0 = (v4f32)__msa_fill_w(0);
-                v4f32 _sum1 = (v4f32)__msa_fill_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 8);
-                    __builtin_prefetch(k0 + 16);
-                    v4f32 _val0 = __msa_fill_w_f32(*r0++);
-                    v4f32 _val1 = __msa_fill_w_f32(*r0++);
-                    v4f32 _w0 = (v4f32)__msa_ld_w(k0, 0);
-                    _sum0 = __msa_fmadd_w(_sum0, _val0, _w0);
-                    _sum1 = __msa_fmadd_w(_sum1, _val1, _w0);
-
-                    k0 += 4;
-                }
-
-                __msa_st_w((v4i32)_sum0, output0_tm, 0);
-                __msa_st_w((v4i32)_sum1, output0_tm + 4, 0);
-
-                output0_tm += 4 * 2;
-            }
-            for (; i < tiles; i++)
-            {
-                const float* r0 = bb2.row(i / 12 + (i % 12) / 8 + (i % 12 % 8) / 4 + (i % 12 % 4) / 2 + i % 12 % 2);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch * 4; // inch always > 0
-
-                v4f32 _sum = (v4f32)__msa_fill_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 4);
-                    __builtin_prefetch(k0 + 16);
-                    v4f32 _val0 = __msa_fill_w_f32(*r0++);
-                    v4f32 _w0 = (v4f32)__msa_ld_w(k0, 0);
-                    _sum = __msa_fmadd_w(_sum, _val0, _w0);
-
-                    k0 += 4;
-                }
-
-                __msa_st_w((v4i32)_sum, output0_tm, 0);
-
-                output0_tm += 4;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_pack8to1_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_pack8to1_int8.h
deleted file mode 100644
index e293f392..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_pack8to1_int8.h
+++ /dev/null
@@ -1,363 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_winograd_dot_pack8to1_int8_msa(Mat& bottom_blob_tm, int outch, const Mat& kernel_tm, Mat& top_blob_tm, const Option& opt)
-{
-    // Mat bottom_blob_tm(tiles, 16/36/64, inch, 16u, 8, opt.workspace_allocator);
-
-    const int tiles = bottom_blob_tm.w;
-    const int batch = bottom_blob_tm.h;
-    const int inch = bottom_blob_tm.c;
-
-    // permute
-    Mat bottom_blob_tm2;
-    if (tiles >= 2)
-        bottom_blob_tm2.create(2 * inch, tiles / 2 + tiles % 2, batch, 16u, 8, opt.workspace_allocator);
-    else // if (tiles >= 1)
-        bottom_blob_tm2.create(1 * inch, tiles, batch, 16u, 8, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int r = 0; r < batch; r++)
-    {
-        Mat tm2 = bottom_blob_tm2.channel(r);
-
-        // tile
-        int i = 0;
-        for (; i + 1 < tiles; i += 2)
-        {
-            short* tmpptr = tm2.row<short>(i / 2);
-
-            const short* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 8;
-
-            for (int q = 0; q < inch; q++)
-            {
-                v8i16 _r0 = __msa_ld_h(r0, 0);
-                v8i16 _r1 = __msa_ld_h(r0 + 8, 0);
-                __msa_st_h(_r0, tmpptr, 0);
-                __msa_st_h(_r1, tmpptr + 8, 0);
-                r0 += bottom_blob_tm.cstep * 8;
-                tmpptr += 16;
-            }
-        }
-        for (; i < tiles; i++)
-        {
-            short* tmpptr = tm2.row<short>(i / 2 + i % 2);
-
-            const short* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 8;
-
-            for (int q = 0; q < inch; q++)
-            {
-                v8i16 _r0 = __msa_ld_h(r0, 0);
-                __msa_st_h(_r0, tmpptr, 0);
-                r0 += bottom_blob_tm.cstep * 8;
-                tmpptr += 8;
-            }
-        }
-    }
-
-    bottom_blob_tm = Mat();
-    // permute end
-
-    top_blob_tm.create(tiles, batch, outch, 4u, 1, opt.workspace_allocator);
-
-    int nn_outch = 0;
-    int remain_outch_start = 0;
-
-    nn_outch = outch >> 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 4;
-
-        int* output0_tm = top_blob_tm.channel(p);
-        int* output1_tm = top_blob_tm.channel(p + 1);
-        int* output2_tm = top_blob_tm.channel(p + 2);
-        int* output3_tm = top_blob_tm.channel(p + 3);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 4);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 1 < tiles; i += 2)
-            {
-                const short* r0 = bb2.row<const short>(i / 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int nn = inch; // inch always > 0
-
-                v4i32 _sum0 = __msa_fill_w(0);
-                v4i32 _sum1 = __msa_fill_w(0);
-                v4i32 _sum2 = __msa_fill_w(0);
-                v4i32 _sum3 = __msa_fill_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 64);
-                    __builtin_prefetch(k0 + 128);
-                    v8i16 _w0 = __msa_ld_h(k0, 0);
-                    v8i16 _w1 = __msa_ld_h(k0 + 8, 0);
-                    v8i16 _w2 = __msa_ld_h(k0 + 16, 0);
-                    v8i16 _w3 = __msa_ld_h(k0 + 24, 0);
-
-                    v8i16 _extw0 = __msa_clti_s_h(_w0, 0);
-                    v8i16 _extw1 = __msa_clti_s_h(_w1, 0);
-                    v8i16 _extw2 = __msa_clti_s_h(_w2, 0);
-                    v8i16 _extw3 = __msa_clti_s_h(_w3, 0);
-
-                    v4i32 _w0l = (v4i32)__msa_ilvr_h(_extw0, _w0);
-                    v4i32 _w0h = (v4i32)__msa_ilvl_h(_extw0, _w0);
-                    v4i32 _w1l = (v4i32)__msa_ilvr_h(_extw1, _w1);
-                    v4i32 _w1h = (v4i32)__msa_ilvl_h(_extw1, _w1);
-                    v4i32 _w2l = (v4i32)__msa_ilvr_h(_extw2, _w2);
-                    v4i32 _w2h = (v4i32)__msa_ilvl_h(_extw2, _w2);
-                    v4i32 _w3l = (v4i32)__msa_ilvr_h(_extw3, _w3);
-                    v4i32 _w3h = (v4i32)__msa_ilvl_h(_extw3, _w3);
-
-                    v4i32 _val0_0 = __msa_fill_w(r0[0]);
-                    v4i32 _val0_1 = __msa_fill_w(r0[1]);
-                    v4i32 _val0_2 = __msa_fill_w(r0[2]);
-                    v4i32 _val0_3 = __msa_fill_w(r0[3]);
-                    v4i32 _val0_4 = __msa_fill_w(r0[4]);
-                    v4i32 _val0_5 = __msa_fill_w(r0[5]);
-                    v4i32 _val0_6 = __msa_fill_w(r0[6]);
-                    v4i32 _val0_7 = __msa_fill_w(r0[7]);
-                    v4i32 _val1_0 = __msa_fill_w(r0[8]);
-                    v4i32 _val1_1 = __msa_fill_w(r0[9]);
-                    v4i32 _val1_2 = __msa_fill_w(r0[10]);
-                    v4i32 _val1_3 = __msa_fill_w(r0[11]);
-                    v4i32 _val1_4 = __msa_fill_w(r0[12]);
-                    v4i32 _val1_5 = __msa_fill_w(r0[13]);
-                    v4i32 _val1_6 = __msa_fill_w(r0[14]);
-                    v4i32 _val1_7 = __msa_fill_w(r0[15]);
-
-                    _sum0 = __msa_maddv_w(_sum0, _w0l, _val0_0);
-                    _sum1 = __msa_maddv_w(_sum1, _w0h, _val0_1);
-                    _sum2 = __msa_maddv_w(_sum2, _w0l, _val1_0);
-                    _sum3 = __msa_maddv_w(_sum3, _w0h, _val1_1);
-                    _sum0 = __msa_maddv_w(_sum0, _w1l, _val0_2);
-                    _sum1 = __msa_maddv_w(_sum1, _w1h, _val0_3);
-                    _sum2 = __msa_maddv_w(_sum2, _w1l, _val1_2);
-                    _sum3 = __msa_maddv_w(_sum3, _w1h, _val1_3);
-                    _sum0 = __msa_maddv_w(_sum0, _w2l, _val0_4);
-                    _sum1 = __msa_maddv_w(_sum1, _w2h, _val0_5);
-                    _sum2 = __msa_maddv_w(_sum2, _w2l, _val1_4);
-                    _sum3 = __msa_maddv_w(_sum3, _w2h, _val1_5);
-                    _sum0 = __msa_maddv_w(_sum0, _w3l, _val0_6);
-                    _sum1 = __msa_maddv_w(_sum1, _w3h, _val0_7);
-                    _sum2 = __msa_maddv_w(_sum2, _w3l, _val1_6);
-                    _sum3 = __msa_maddv_w(_sum3, _w3h, _val1_7);
-
-                    r0 += 16;
-                    k0 += 32;
-                }
-
-                _sum0 = __msa_addv_w(_sum0, _sum1);
-                _sum2 = __msa_addv_w(_sum2, _sum3);
-
-                int sum[8];
-                __msa_st_w(_sum0, sum, 0);
-                __msa_st_w(_sum2, sum + 4, 0);
-
-                output0_tm[0] = sum[0];
-                output1_tm[0] = sum[1];
-                output2_tm[0] = sum[2];
-                output3_tm[0] = sum[3];
-                output0_tm[1] = sum[4];
-                output1_tm[1] = sum[5];
-                output2_tm[1] = sum[6];
-                output3_tm[1] = sum[7];
-                output0_tm += 2;
-                output1_tm += 2;
-                output2_tm += 2;
-                output3_tm += 2;
-            }
-            for (; i < tiles; i++)
-            {
-                const short* r0 = bb2.row<const short>(i / 2 + i % 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int nn = inch; // inch always > 0
-
-                v4i32 _sum0 = __msa_fill_w(0);
-                v4i32 _sum1 = __msa_fill_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 32);
-                    __builtin_prefetch(k0 + 128);
-                    v8i16 _w0 = __msa_ld_h(k0, 0);
-                    v8i16 _w1 = __msa_ld_h(k0 + 8, 0);
-                    v8i16 _w2 = __msa_ld_h(k0 + 16, 0);
-                    v8i16 _w3 = __msa_ld_h(k0 + 24, 0);
-
-                    v8i16 _extw0 = __msa_clti_s_h(_w0, 0);
-                    v8i16 _extw1 = __msa_clti_s_h(_w1, 0);
-                    v8i16 _extw2 = __msa_clti_s_h(_w2, 0);
-                    v8i16 _extw3 = __msa_clti_s_h(_w3, 0);
-
-                    v4i32 _w0l = (v4i32)__msa_ilvr_h(_extw0, _w0);
-                    v4i32 _w0h = (v4i32)__msa_ilvl_h(_extw0, _w0);
-                    v4i32 _w1l = (v4i32)__msa_ilvr_h(_extw1, _w1);
-                    v4i32 _w1h = (v4i32)__msa_ilvl_h(_extw1, _w1);
-                    v4i32 _w2l = (v4i32)__msa_ilvr_h(_extw2, _w2);
-                    v4i32 _w2h = (v4i32)__msa_ilvl_h(_extw2, _w2);
-                    v4i32 _w3l = (v4i32)__msa_ilvr_h(_extw3, _w3);
-                    v4i32 _w3h = (v4i32)__msa_ilvl_h(_extw3, _w3);
-
-                    v4i32 _val0 = __msa_fill_w(r0[0]);
-                    v4i32 _val1 = __msa_fill_w(r0[1]);
-                    v4i32 _val2 = __msa_fill_w(r0[2]);
-                    v4i32 _val3 = __msa_fill_w(r0[3]);
-                    v4i32 _val4 = __msa_fill_w(r0[4]);
-                    v4i32 _val5 = __msa_fill_w(r0[5]);
-                    v4i32 _val6 = __msa_fill_w(r0[6]);
-                    v4i32 _val7 = __msa_fill_w(r0[7]);
-
-                    _sum0 = __msa_maddv_w(_sum0, _w0l, _val0);
-                    _sum1 = __msa_maddv_w(_sum1, _w0h, _val1);
-                    _sum0 = __msa_maddv_w(_sum0, _w1l, _val2);
-                    _sum1 = __msa_maddv_w(_sum1, _w1h, _val3);
-                    _sum0 = __msa_maddv_w(_sum0, _w2l, _val4);
-                    _sum1 = __msa_maddv_w(_sum1, _w2h, _val5);
-                    _sum0 = __msa_maddv_w(_sum0, _w3l, _val6);
-                    _sum1 = __msa_maddv_w(_sum1, _w3h, _val7);
-
-                    r0 += 8;
-                    k0 += 32;
-                }
-
-                _sum0 = __msa_addv_w(_sum0, _sum1);
-
-                int sum[4];
-                __msa_st_w(_sum0, sum, 0);
-
-                output0_tm[0] = sum[0];
-                output1_tm[0] = sum[1];
-                output2_tm[0] = sum[2];
-                output3_tm[0] = sum[3];
-                output0_tm += 1;
-                output1_tm += 1;
-                output2_tm += 1;
-                output3_tm += 1;
-            }
-        }
-    }
-
-    remain_outch_start += nn_outch << 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        int* output0_tm = top_blob_tm.channel(p);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 4 + p % 4);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 1 < tiles; i += 2)
-            {
-                const short* r0 = bb2.row<const short>(i / 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                v4i32 _sum0 = __msa_fill_w(0);
-                v4i32 _sum1 = __msa_fill_w(0);
-                v4i32 _sum2 = __msa_fill_w(0);
-                v4i32 _sum3 = __msa_fill_w(0);
-
-                for (int q = 0; q < inch; q++)
-                {
-                    __builtin_prefetch(r0 + 32);
-                    __builtin_prefetch(k0 + 64);
-                    v8i16 _val0 = __msa_ld_h(r0, 0);
-                    v8i16 _val1 = __msa_ld_h(r0 + 8, 0);
-
-                    v8i16 _extval0 = __msa_clti_s_h(_val0, 0);
-                    v8i16 _extval1 = __msa_clti_s_h(_val1, 0);
-                    v4i32 _val0l = (v4i32)__msa_ilvr_h(_extval0, _val0);
-                    v4i32 _val0h = (v4i32)__msa_ilvl_h(_extval0, _val0);
-                    v4i32 _val1l = (v4i32)__msa_ilvr_h(_extval1, _val1);
-                    v4i32 _val1h = (v4i32)__msa_ilvl_h(_extval1, _val1);
-
-                    v8i16 _w0 = __msa_ld_h(k0, 0);
-
-                    v8i16 _extw0 = __msa_clti_s_h(_w0, 0);
-                    v4i32 _w0l = (v4i32)__msa_ilvr_h(_extw0, _w0);
-                    v4i32 _w0h = (v4i32)__msa_ilvl_h(_extw0, _w0);
-
-                    _sum0 = __msa_maddv_w(_sum0, _w0l, _val0l);
-                    _sum1 = __msa_maddv_w(_sum1, _w0h, _val0h);
-                    _sum2 = __msa_maddv_w(_sum2, _w0l, _val1l);
-                    _sum3 = __msa_maddv_w(_sum3, _w0h, _val1h);
-
-                    k0 += 8;
-                    r0 += 16;
-                }
-
-                _sum0 = __msa_addv_w(_sum0, _sum1);
-                _sum2 = __msa_addv_w(_sum2, _sum3);
-
-                output0_tm[0] = __msa_reduce_add_w(_sum0);
-                output0_tm[1] = __msa_reduce_add_w(_sum2);
-                output0_tm += 2;
-            }
-            for (; i < tiles; i++)
-            {
-                const short* r0 = bb2.row<const short>(i / 2 + i % 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                v4i32 _sum0 = __msa_fill_w(0);
-                v4i32 _sum1 = __msa_fill_w(0);
-
-                for (int q = 0; q < inch; q++)
-                {
-                    __builtin_prefetch(r0 + 32);
-                    __builtin_prefetch(k0 + 32);
-                    v8i16 _val = __msa_ld_h(r0, 0);
-
-                    v8i16 _extval = __msa_clti_s_h(_val, 0);
-                    v4i32 _vall = (v4i32)__msa_ilvr_h(_extval, _val);
-                    v4i32 _valh = (v4i32)__msa_ilvl_h(_extval, _val);
-
-                    v8i16 _w0 = __msa_ld_h(k0, 0);
-
-                    v8i16 _extw0 = __msa_clti_s_h(_w0, 0);
-                    v4i32 _w0l = (v4i32)__msa_ilvr_h(_extw0, _w0);
-                    v4i32 _w0h = (v4i32)__msa_ilvl_h(_extw0, _w0);
-
-                    _sum0 = __msa_maddv_w(_sum0, _w0l, _vall);
-                    _sum1 = __msa_maddv_w(_sum1, _w0h, _valh);
-
-                    k0 += 8;
-                    r0 += 8;
-                }
-
-                _sum0 = __msa_addv_w(_sum0, _sum1);
-
-                output0_tm[0] = __msa_reduce_add_w(_sum0);
-                output0_tm++;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_pack8to4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_pack8to4_int8.h
deleted file mode 100644
index 5f58ca4d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_dot_pack8to4_int8.h
+++ /dev/null
@@ -1,233 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_winograd_dot_pack8to4_int8_msa(Mat& bottom_blob_tm, int outch, const Mat& kernel_tm, Mat& top_blob_tm, const Option& opt)
-{
-    // Mat bottom_blob_tm(tiles, 16/36/64, inch, 16u, 8, opt.workspace_allocator);
-
-    const int tiles = bottom_blob_tm.w;
-    const int batch = bottom_blob_tm.h;
-    const int inch = bottom_blob_tm.c;
-
-    // permute
-    Mat bottom_blob_tm2;
-    if (tiles >= 2)
-        bottom_blob_tm2.create(2 * inch, tiles / 2 + tiles % 2, batch, 16u, 8, opt.workspace_allocator);
-    else // if (tiles >= 1)
-        bottom_blob_tm2.create(1 * inch, tiles, batch, 16u, 8, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int r = 0; r < batch; r++)
-    {
-        Mat tm2 = bottom_blob_tm2.channel(r);
-
-        // tile
-        int i = 0;
-        for (; i + 1 < tiles; i += 2)
-        {
-            short* tmpptr = tm2.row<short>(i / 2);
-
-            const short* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 8;
-
-            for (int q = 0; q < inch; q++)
-            {
-                v8i16 _r0 = __msa_ld_h(r0, 0);
-                v8i16 _r1 = __msa_ld_h(r0 + 8, 0);
-                __msa_st_h(_r0, tmpptr, 0);
-                __msa_st_h(_r1, tmpptr + 8, 0);
-                r0 += bottom_blob_tm.cstep * 8;
-                tmpptr += 16;
-            }
-        }
-        for (; i < tiles; i++)
-        {
-            short* tmpptr = tm2.row<short>(i / 2 + i % 2);
-
-            const short* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * 8;
-
-            for (int q = 0; q < inch; q++)
-            {
-                v8i16 _r0 = __msa_ld_h(r0, 0);
-                __msa_st_h(_r0, tmpptr, 0);
-                r0 += bottom_blob_tm.cstep * 8;
-                tmpptr += 8;
-            }
-        }
-    }
-
-    bottom_blob_tm = Mat();
-    // permute end
-
-    top_blob_tm.create(tiles, batch, outch, 16u, 4, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        int* output0_tm = top_blob_tm.channel(p);
-
-        const Mat kernel0_tm = kernel_tm.channel(p);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 1 < tiles; i += 2)
-            {
-                const short* r0 = bb2.row<const short>(i / 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int nn = inch; // inch always > 0
-
-                v4i32 _sum0 = __msa_fill_w(0);
-                v4i32 _sum1 = __msa_fill_w(0);
-                v4i32 _sum2 = __msa_fill_w(0);
-                v4i32 _sum3 = __msa_fill_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 64);
-                    __builtin_prefetch(k0 + 128);
-                    v8i16 _w0 = __msa_ld_h(k0, 0);
-                    v8i16 _w1 = __msa_ld_h(k0 + 8, 0);
-                    v8i16 _w2 = __msa_ld_h(k0 + 16, 0);
-                    v8i16 _w3 = __msa_ld_h(k0 + 24, 0);
-
-                    v8i16 _extw0 = __msa_clti_s_h(_w0, 0);
-                    v8i16 _extw1 = __msa_clti_s_h(_w1, 0);
-                    v8i16 _extw2 = __msa_clti_s_h(_w2, 0);
-                    v8i16 _extw3 = __msa_clti_s_h(_w3, 0);
-
-                    v4i32 _w0l = (v4i32)__msa_ilvr_h(_extw0, _w0);
-                    v4i32 _w0h = (v4i32)__msa_ilvl_h(_extw0, _w0);
-                    v4i32 _w1l = (v4i32)__msa_ilvr_h(_extw1, _w1);
-                    v4i32 _w1h = (v4i32)__msa_ilvl_h(_extw1, _w1);
-                    v4i32 _w2l = (v4i32)__msa_ilvr_h(_extw2, _w2);
-                    v4i32 _w2h = (v4i32)__msa_ilvl_h(_extw2, _w2);
-                    v4i32 _w3l = (v4i32)__msa_ilvr_h(_extw3, _w3);
-                    v4i32 _w3h = (v4i32)__msa_ilvl_h(_extw3, _w3);
-
-                    v4i32 _val0_0 = __msa_fill_w(r0[0]);
-                    v4i32 _val0_1 = __msa_fill_w(r0[1]);
-                    v4i32 _val0_2 = __msa_fill_w(r0[2]);
-                    v4i32 _val0_3 = __msa_fill_w(r0[3]);
-                    v4i32 _val0_4 = __msa_fill_w(r0[4]);
-                    v4i32 _val0_5 = __msa_fill_w(r0[5]);
-                    v4i32 _val0_6 = __msa_fill_w(r0[6]);
-                    v4i32 _val0_7 = __msa_fill_w(r0[7]);
-                    v4i32 _val1_0 = __msa_fill_w(r0[8]);
-                    v4i32 _val1_1 = __msa_fill_w(r0[9]);
-                    v4i32 _val1_2 = __msa_fill_w(r0[10]);
-                    v4i32 _val1_3 = __msa_fill_w(r0[11]);
-                    v4i32 _val1_4 = __msa_fill_w(r0[12]);
-                    v4i32 _val1_5 = __msa_fill_w(r0[13]);
-                    v4i32 _val1_6 = __msa_fill_w(r0[14]);
-                    v4i32 _val1_7 = __msa_fill_w(r0[15]);
-
-                    _sum0 = __msa_maddv_w(_sum0, _w0l, _val0_0);
-                    _sum1 = __msa_maddv_w(_sum1, _w0h, _val0_1);
-                    _sum2 = __msa_maddv_w(_sum2, _w0l, _val1_0);
-                    _sum3 = __msa_maddv_w(_sum3, _w0h, _val1_1);
-                    _sum0 = __msa_maddv_w(_sum0, _w1l, _val0_2);
-                    _sum1 = __msa_maddv_w(_sum1, _w1h, _val0_3);
-                    _sum2 = __msa_maddv_w(_sum2, _w1l, _val1_2);
-                    _sum3 = __msa_maddv_w(_sum3, _w1h, _val1_3);
-                    _sum0 = __msa_maddv_w(_sum0, _w2l, _val0_4);
-                    _sum1 = __msa_maddv_w(_sum1, _w2h, _val0_5);
-                    _sum2 = __msa_maddv_w(_sum2, _w2l, _val1_4);
-                    _sum3 = __msa_maddv_w(_sum3, _w2h, _val1_5);
-                    _sum0 = __msa_maddv_w(_sum0, _w3l, _val0_6);
-                    _sum1 = __msa_maddv_w(_sum1, _w3h, _val0_7);
-                    _sum2 = __msa_maddv_w(_sum2, _w3l, _val1_6);
-                    _sum3 = __msa_maddv_w(_sum3, _w3h, _val1_7);
-
-                    r0 += 16;
-                    k0 += 32;
-                }
-
-                _sum0 = __msa_addv_w(_sum0, _sum1);
-                _sum2 = __msa_addv_w(_sum2, _sum3);
-
-                __msa_st_w(_sum0, output0_tm, 0);
-                __msa_st_w(_sum2, output0_tm + 4, 0);
-
-                output0_tm += 8;
-            }
-            for (; i < tiles; i++)
-            {
-                const short* r0 = bb2.row<const short>(i / 2 + i % 2);
-                const short* k0 = kernel0_tm.row<const short>(r);
-
-                int nn = inch; // inch always > 0
-
-                v4i32 _sum0 = __msa_fill_w(0);
-                v4i32 _sum1 = __msa_fill_w(0);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 32);
-                    __builtin_prefetch(k0 + 128);
-                    v8i16 _w0 = __msa_ld_h(k0, 0);
-                    v8i16 _w1 = __msa_ld_h(k0 + 8, 0);
-                    v8i16 _w2 = __msa_ld_h(k0 + 16, 0);
-                    v8i16 _w3 = __msa_ld_h(k0 + 24, 0);
-
-                    v8i16 _extw0 = __msa_clti_s_h(_w0, 0);
-                    v8i16 _extw1 = __msa_clti_s_h(_w1, 0);
-                    v8i16 _extw2 = __msa_clti_s_h(_w2, 0);
-                    v8i16 _extw3 = __msa_clti_s_h(_w3, 0);
-
-                    v4i32 _w0l = (v4i32)__msa_ilvr_h(_extw0, _w0);
-                    v4i32 _w0h = (v4i32)__msa_ilvl_h(_extw0, _w0);
-                    v4i32 _w1l = (v4i32)__msa_ilvr_h(_extw1, _w1);
-                    v4i32 _w1h = (v4i32)__msa_ilvl_h(_extw1, _w1);
-                    v4i32 _w2l = (v4i32)__msa_ilvr_h(_extw2, _w2);
-                    v4i32 _w2h = (v4i32)__msa_ilvl_h(_extw2, _w2);
-                    v4i32 _w3l = (v4i32)__msa_ilvr_h(_extw3, _w3);
-                    v4i32 _w3h = (v4i32)__msa_ilvl_h(_extw3, _w3);
-
-                    v4i32 _val0 = __msa_fill_w(r0[0]);
-                    v4i32 _val1 = __msa_fill_w(r0[1]);
-                    v4i32 _val2 = __msa_fill_w(r0[2]);
-                    v4i32 _val3 = __msa_fill_w(r0[3]);
-                    v4i32 _val4 = __msa_fill_w(r0[4]);
-                    v4i32 _val5 = __msa_fill_w(r0[5]);
-                    v4i32 _val6 = __msa_fill_w(r0[6]);
-                    v4i32 _val7 = __msa_fill_w(r0[7]);
-
-                    _sum0 = __msa_maddv_w(_sum0, _w0l, _val0);
-                    _sum1 = __msa_maddv_w(_sum1, _w0h, _val1);
-                    _sum0 = __msa_maddv_w(_sum0, _w1l, _val2);
-                    _sum1 = __msa_maddv_w(_sum1, _w1h, _val3);
-                    _sum0 = __msa_maddv_w(_sum0, _w2l, _val4);
-                    _sum1 = __msa_maddv_w(_sum1, _w2h, _val5);
-                    _sum0 = __msa_maddv_w(_sum0, _w3l, _val6);
-                    _sum1 = __msa_maddv_w(_sum1, _w3h, _val7);
-
-                    r0 += 8;
-                    k0 += 32;
-                }
-
-                _sum0 = __msa_addv_w(_sum0, _sum1);
-
-                __msa_st_w(_sum0, output0_tm, 0);
-                output0_tm += 4;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform.h
deleted file mode 100644
index 14578df5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform.h
+++ /dev/null
@@ -1,405 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd43_transform_input_msa(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 4;
-    const int h_tiles = (h - 2) / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[6][6] = {
-    //     {4.0f, 0.0f, -5.0f, 0.0f, 1.0f, 0.0f},
-    //     {0.0f,-4.0f, -4.0f, 1.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f, -4.0f,-1.0f, 1.0f, 0.0f},
-    //     {0.0f,-2.0f, -1.0f, 2.0f, 1.0f, 0.0f},
-    //     {0.0f, 2.0f, -1.0f,-2.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f,  0.0f,-5.0f, 0.0f, 1.0f}
-    // };
-
-    // 0 =  4 * r00 - 5 * r02 + r04
-    // 1 = -4 * (r01 + r02) + r04 + r03
-    // 2 =  4 * (r01 - r02) + r04 - r03
-    // 3 = -2 * (r01 - r03) + r04 - r02
-    // 4 =  2 * (r01 - r03) + r04 - r02
-    // 5 =  4 * r01 - 5 * r03 + r05
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        float tmp[6][6];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* r0 = img0.row(i * 4) + (j * 4);
-
-                for (int m = 0; m < 6; m++)
-                {
-                    float r00 = r0[0];
-                    float r01 = r0[1];
-                    float r02 = r0[2];
-                    float r03 = r0[3];
-                    float r04 = r0[4];
-                    float r05 = r0[5];
-
-                    float tmp0m = 4 * r00 - 5 * r02 + r04;
-                    float tmp1m = -4 * (r01 + r02) + r04 + r03;
-                    float tmp2m = 4 * (r01 - r02) + r04 - r03;
-                    float tmp3m = -2 * (r01 - r03) + r04 - r02;
-                    float tmp4m = 2 * (r01 - r03) + r04 - r02;
-                    float tmp5m = 4 * r01 - 5 * r03 + r05;
-
-                    tmp[0][m] = tmp0m;
-                    tmp[1][m] = tmp1m;
-                    tmp[2][m] = tmp2m;
-                    tmp[3][m] = tmp3m;
-                    tmp[4][m] = tmp4m;
-                    tmp[5][m] = tmp5m;
-
-                    r0 += w;
-                }
-
-                float* r0_tm_0 = (float*)img0_tm + (i * w_tiles + j);
-                float* r0_tm_1 = r0_tm_0 + tiles;
-                float* r0_tm_2 = r0_tm_0 + tiles * 2;
-                float* r0_tm_3 = r0_tm_0 + tiles * 3;
-                float* r0_tm_4 = r0_tm_0 + tiles * 4;
-                float* r0_tm_5 = r0_tm_0 + tiles * 5;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    float tmp00 = tmp[m][0];
-                    float tmp01 = tmp[m][1];
-                    float tmp02 = tmp[m][2];
-                    float tmp03 = tmp[m][3];
-                    float tmp04 = tmp[m][4];
-                    float tmp05 = tmp[m][5];
-
-                    float r0tm0 = 4 * tmp00 - 5 * tmp02 + tmp04;
-                    float r0tm1 = -4 * (tmp01 + tmp02) + tmp04 + tmp03;
-                    float r0tm2 = 4 * (tmp01 - tmp02) + tmp04 - tmp03;
-                    float r0tm3 = -2 * (tmp01 - tmp03) + tmp04 - tmp02;
-                    float r0tm4 = 2 * (tmp01 - tmp03) + tmp04 - tmp02;
-                    float r0tm5 = 4 * tmp01 - 5 * tmp03 + tmp05;
-
-                    r0_tm_0[0] = r0tm0;
-                    r0_tm_1[0] = r0tm1;
-                    r0_tm_2[0] = r0tm2;
-                    r0_tm_3[0] = r0tm3;
-                    r0_tm_4[0] = r0tm4;
-                    r0_tm_5[0] = r0tm5;
-
-                    r0_tm_0 += tiles * 6;
-                    r0_tm_1 += tiles * 6;
-                    r0_tm_2 += tiles * 6;
-                    r0_tm_3 += tiles * 6;
-                    r0_tm_4 += tiles * 6;
-                    r0_tm_5 += tiles * 6;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_output_msa(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 4;
-    const int h_tiles = outh / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    const float* biasptr = bias;
-
-    // const float otm[4][6] = {
-    //     {1.0f, 1.0f,  1.0f, 1.0f,  1.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 2.0f, -2.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f, 4.0f,  4.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 8.0f, -8.0f, 1.0f}
-    // };
-
-    // 0 = r00 + (r01 + r02) + (r03 + r04)
-    // 1 =       (r01 - r02) + (r03 - r04) * 2
-    // 2 =       (r01 + r02) + (r03 + r04) * 4
-    // 3 = r05 + (r01 - r02) + (r03 - r04) * 8
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        float bias0 = biasptr ? biasptr[p] : 0.f;
-
-        float tmp[4][6];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* output0_tm_0 = (const float*)out0_tm + (i * w_tiles + j);
-                const float* output0_tm_1 = output0_tm_0 + tiles;
-                const float* output0_tm_2 = output0_tm_0 + tiles * 2;
-                const float* output0_tm_3 = output0_tm_0 + tiles * 3;
-                const float* output0_tm_4 = output0_tm_0 + tiles * 4;
-                const float* output0_tm_5 = output0_tm_0 + tiles * 5;
-
-                float* output0 = out0.row(i * 4) + (j * 4);
-
-                for (int m = 0; m < 6; m++)
-                {
-                    float out0tm0 = output0_tm_0[0];
-                    float out0tm1 = output0_tm_1[0];
-                    float out0tm2 = output0_tm_2[0];
-                    float out0tm3 = output0_tm_3[0];
-                    float out0tm4 = output0_tm_4[0];
-                    float out0tm5 = output0_tm_5[0];
-
-                    float tmp02a = out0tm1 + out0tm2;
-                    float tmp13a = out0tm1 - out0tm2;
-
-                    float tmp02b = out0tm3 + out0tm4;
-                    float tmp13b = out0tm3 - out0tm4;
-
-                    float tmp0m = out0tm0 + tmp02a + tmp02b;
-                    float tmp1m = tmp13a + tmp13b * 2;
-                    float tmp2m = tmp02a + tmp02b * 4;
-                    float tmp3m = out0tm5 + tmp13a + tmp13b * 8;
-
-                    tmp[0][m] = tmp0m;
-                    tmp[1][m] = tmp1m;
-                    tmp[2][m] = tmp2m;
-                    tmp[3][m] = tmp3m;
-
-                    output0_tm_0 += tiles * 6;
-                    output0_tm_1 += tiles * 6;
-                    output0_tm_2 += tiles * 6;
-                    output0_tm_3 += tiles * 6;
-                    output0_tm_4 += tiles * 6;
-                    output0_tm_5 += tiles * 6;
-                }
-
-                for (int m = 0; m < 4; m++)
-                {
-                    float tmp00 = tmp[m][0];
-                    float tmp01 = tmp[m][1];
-                    float tmp02 = tmp[m][2];
-                    float tmp03 = tmp[m][3];
-                    float tmp04 = tmp[m][4];
-                    float tmp05 = tmp[m][5];
-
-                    float tmp02a = tmp01 + tmp02;
-                    float tmp13a = tmp01 - tmp02;
-
-                    float tmp02b = tmp03 + tmp04;
-                    float tmp13b = tmp03 - tmp04;
-
-                    float out00 = bias0 + tmp00 + tmp02a + tmp02b;
-                    float out01 = bias0 + tmp13a + tmp13b * 2;
-                    float out02 = bias0 + tmp02a + tmp02b * 4;
-                    float out03 = bias0 + tmp05 + tmp13a + tmp13b * 8;
-
-                    output0[0] = out00;
-                    output0[1] = out01;
-                    output0[2] = out02;
-                    output0[3] = out03;
-
-                    output0 += outw;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_input_msa(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 2;
-    const int h_tiles = (h - 2) / 2;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[4][4] = {
-    //     {1.0f,  0.0f, -1.0f,  0.0f},
-    //     {0.0f,  1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  0.00f, 1.0f}
-    // };
-
-    // 0 = r00 - r02
-    // 1 = r01 + r02
-    // 2 = r02 - r01
-    // 3 = r03 - r01
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        float tmp[4][4];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* r0 = img0.row(i * 2) + (j * 2);
-
-                for (int m = 0; m < 4; m++)
-                {
-                    float r00 = r0[0];
-                    float r01 = r0[1];
-                    float r02 = r0[2];
-                    float r03 = r0[3];
-
-                    float tmp0m = r00 - r02;
-                    float tmp1m = r01 + r02;
-                    float tmp2m = r02 - r01;
-                    float tmp3m = r03 - r01;
-
-                    tmp[0][m] = tmp0m;
-                    tmp[1][m] = tmp1m;
-                    tmp[2][m] = tmp2m;
-                    tmp[3][m] = tmp3m;
-
-                    r0 += w;
-                }
-
-                float* r0_tm_0 = (float*)img0_tm + (i * w_tiles + j);
-                float* r0_tm_1 = r0_tm_0 + tiles;
-                float* r0_tm_2 = r0_tm_0 + tiles * 2;
-                float* r0_tm_3 = r0_tm_0 + tiles * 3;
-
-                for (int m = 0; m < 4; m++)
-                {
-                    float tmp00 = tmp[m][0];
-                    float tmp01 = tmp[m][1];
-                    float tmp02 = tmp[m][2];
-                    float tmp03 = tmp[m][3];
-
-                    float r0tm0 = tmp00 - tmp02;
-                    float r0tm1 = tmp01 + tmp02;
-                    float r0tm2 = tmp02 - tmp01;
-                    float r0tm3 = tmp03 - tmp01;
-
-                    r0_tm_0[0] = r0tm0;
-                    r0_tm_1[0] = r0tm1;
-                    r0_tm_2[0] = r0tm2;
-                    r0_tm_3[0] = r0tm3;
-
-                    r0_tm_0 += tiles * 4;
-                    r0_tm_1 += tiles * 4;
-                    r0_tm_2 += tiles * 4;
-                    r0_tm_3 += tiles * 4;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_output_msa(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 2;
-    const int h_tiles = outh / 2;
-    const int tiles = w_tiles * h_tiles;
-
-    const float* biasptr = bias;
-
-    // const float otm[2][4] = {
-    //     {1.0f,  1.0f,  1.0f,  0.0f},
-    //     {0.0f,  1.0f, -1.0f,  1.0f}
-    // };
-
-    // 0 = r00 + r01 + r02
-    // 1 = r01 - r02 + r03
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        float bias0 = biasptr ? biasptr[p] : 0.f;
-
-        float tmp[2][4];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* output0_tm_0 = (const float*)out0_tm + (i * w_tiles + j);
-                const float* output0_tm_1 = output0_tm_0 + tiles;
-                const float* output0_tm_2 = output0_tm_0 + tiles * 2;
-                const float* output0_tm_3 = output0_tm_0 + tiles * 3;
-
-                float* output0 = out0.row(i * 2) + (j * 2);
-
-                for (int m = 0; m < 4; m++)
-                {
-                    float out0tm0 = output0_tm_0[0];
-                    float out0tm1 = output0_tm_1[0];
-                    float out0tm2 = output0_tm_2[0];
-                    float out0tm3 = output0_tm_3[0];
-
-                    float tmp0m = out0tm0 + out0tm1 + out0tm2;
-                    float tmp1m = out0tm1 - out0tm2 + out0tm3;
-
-                    tmp[0][m] = tmp0m;
-                    tmp[1][m] = tmp1m;
-
-                    output0_tm_0 += tiles * 4;
-                    output0_tm_1 += tiles * 4;
-                    output0_tm_2 += tiles * 4;
-                    output0_tm_3 += tiles * 4;
-                }
-
-                for (int m = 0; m < 2; m++)
-                {
-                    float tmp00 = tmp[m][0];
-                    float tmp01 = tmp[m][1];
-                    float tmp02 = tmp[m][2];
-                    float tmp03 = tmp[m][3];
-
-                    float out00 = bias0 + tmp00 + tmp01 + tmp02;
-                    float out01 = bias0 + tmp01 - tmp02 + tmp03;
-
-                    output0[0] = out00;
-                    output0[1] = out01;
-
-                    output0 += outw;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_int8.h
deleted file mode 100644
index 7bcb8b94..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_int8.h
+++ /dev/null
@@ -1,229 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd43_transform_input_int8_msa(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 4;
-    const int h_tiles = (h - 2) / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[6][6] = {
-    //     {4.0f, 0.0f, -5.0f, 0.0f, 1.0f, 0.0f},
-    //     {0.0f,-4.0f, -4.0f, 1.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f, -4.0f,-1.0f, 1.0f, 0.0f},
-    //     {0.0f,-2.0f, -1.0f, 2.0f, 1.0f, 0.0f},
-    //     {0.0f, 2.0f, -1.0f,-2.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f,  0.0f,-5.0f, 0.0f, 1.0f}
-    // };
-
-    // 0 =  4 * r00 - 5 * r02 + r04
-    // 1 = -4 * (r01 + r02) + r04 + r03
-    // 2 =  4 * (r01 - r02) + r04 - r03
-    // 3 = -2 * (r01 - r03) + r04 - r02
-    // 4 =  2 * (r01 - r03) + r04 - r02
-    // 5 =  4 * r01 - 5 * r03 + r05
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        short tmp[6][6];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const signed char* r0 = img0.row<const signed char>(i * 4) + (j * 4);
-
-                for (int m = 0; m < 6; m++)
-                {
-                    signed char r00 = r0[0];
-                    signed char r01 = r0[1];
-                    signed char r02 = r0[2];
-                    signed char r03 = r0[3];
-                    signed char r04 = r0[4];
-                    signed char r05 = r0[5];
-
-                    short tmp0m = 4 * r00 - 5 * r02 + r04;
-                    short tmp1m = -4 * (r01 + r02) + r04 + r03;
-                    short tmp2m = 4 * (r01 - r02) + r04 - r03;
-                    short tmp3m = -2 * (r01 - r03) + r04 - r02;
-                    short tmp4m = 2 * (r01 - r03) + r04 - r02;
-                    short tmp5m = 4 * r01 - 5 * r03 + r05;
-
-                    tmp[0][m] = tmp0m;
-                    tmp[1][m] = tmp1m;
-                    tmp[2][m] = tmp2m;
-                    tmp[3][m] = tmp3m;
-                    tmp[4][m] = tmp4m;
-                    tmp[5][m] = tmp5m;
-
-                    r0 += w;
-                }
-
-                short* r0_tm_0 = (short*)img0_tm + (i * w_tiles + j);
-                short* r0_tm_1 = r0_tm_0 + tiles;
-                short* r0_tm_2 = r0_tm_0 + tiles * 2;
-                short* r0_tm_3 = r0_tm_0 + tiles * 3;
-                short* r0_tm_4 = r0_tm_0 + tiles * 4;
-                short* r0_tm_5 = r0_tm_0 + tiles * 5;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    short tmp00 = tmp[m][0];
-                    short tmp01 = tmp[m][1];
-                    short tmp02 = tmp[m][2];
-                    short tmp03 = tmp[m][3];
-                    short tmp04 = tmp[m][4];
-                    short tmp05 = tmp[m][5];
-
-                    short r0tm0 = 4 * tmp00 - 5 * tmp02 + tmp04;
-                    short r0tm1 = -4 * (tmp01 + tmp02) + tmp04 + tmp03;
-                    short r0tm2 = 4 * (tmp01 - tmp02) + tmp04 - tmp03;
-                    short r0tm3 = -2 * (tmp01 - tmp03) + tmp04 - tmp02;
-                    short r0tm4 = 2 * (tmp01 - tmp03) + tmp04 - tmp02;
-                    short r0tm5 = 4 * tmp01 - 5 * tmp03 + tmp05;
-
-                    r0_tm_0[0] = r0tm0;
-                    r0_tm_1[0] = r0tm1;
-                    r0_tm_2[0] = r0tm2;
-                    r0_tm_3[0] = r0tm3;
-                    r0_tm_4[0] = r0tm4;
-                    r0_tm_5[0] = r0tm5;
-
-                    r0_tm_0 += tiles * 6;
-                    r0_tm_1 += tiles * 6;
-                    r0_tm_2 += tiles * 6;
-                    r0_tm_3 += tiles * 6;
-                    r0_tm_4 += tiles * 6;
-                    r0_tm_5 += tiles * 6;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_output_int8_msa(const Mat& top_blob_tm, Mat& top_blob, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 4;
-    const int h_tiles = outh / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float otm[4][6] = {
-    //     {1.0f, 1.0f,  1.0f, 1.0f,  1.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 2.0f, -2.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f, 4.0f,  4.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 8.0f, -8.0f, 1.0f}
-    // };
-
-    // 0 = r00 + (r01 + r02) + (r03 + r04)
-    // 1 =       (r01 - r02) + (r03 - r04) * 2
-    // 2 =       (r01 + r02) + (r03 + r04) * 4
-    // 3 = r05 + (r01 - r02) + (r03 - r04) * 8
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        int tmp[4][6];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const int* output0_tm_0 = (const int*)out0_tm + (i * w_tiles + j) * 1;
-                const int* output0_tm_1 = output0_tm_0 + tiles * 1;
-                const int* output0_tm_2 = output0_tm_0 + tiles * 2;
-                const int* output0_tm_3 = output0_tm_0 + tiles * 3;
-                const int* output0_tm_4 = output0_tm_0 + tiles * 4;
-                const int* output0_tm_5 = output0_tm_0 + tiles * 5;
-
-                int* output0 = out0.row<int>(i * 4) + j * 4;
-
-                for (int m = 0; m < 5; m++)
-                {
-                    int tmp02a = output0_tm_1[0] + output0_tm_2[0];
-                    int tmp13a = output0_tm_1[0] - output0_tm_2[0];
-
-                    int tmp02b = output0_tm_3[0] + output0_tm_4[0];
-                    int tmp13b = output0_tm_3[0] - output0_tm_4[0];
-
-                    tmp[0][m] = output0_tm_0[0] + tmp02a + tmp02b;
-                    tmp[1][m] = tmp13a + tmp13b * 2;
-                    tmp[2][m] = tmp02a + tmp02b * 4;
-                    tmp[3][m] = output0_tm_5[0] * 4 + tmp13a + tmp13b * 8;
-
-                    output0_tm_0 += tiles * 6;
-                    output0_tm_1 += tiles * 6;
-                    output0_tm_2 += tiles * 6;
-                    output0_tm_3 += tiles * 6;
-                    output0_tm_4 += tiles * 6;
-                    output0_tm_5 += tiles * 6;
-                }
-                for (int m = 5; m < 6; m++)
-                {
-                    int tmp02a = output0_tm_1[0] + output0_tm_2[0];
-                    int tmp13a = output0_tm_1[0] - output0_tm_2[0];
-
-                    int tmp02b = output0_tm_3[0] + output0_tm_4[0];
-                    int tmp13b = output0_tm_3[0] - output0_tm_4[0];
-
-                    tmp[0][m] = (output0_tm_0[0] + tmp02a + tmp02b) * 4;
-                    tmp[1][m] = (tmp13a + tmp13b * 2) * 4;
-                    tmp[2][m] = (tmp02a + tmp02b * 4) * 4;
-                    tmp[3][m] = (output0_tm_5[0] * 4 + tmp13a + tmp13b * 8) * 4;
-
-                    output0_tm_0 += tiles * 6;
-                    output0_tm_1 += tiles * 6;
-                    output0_tm_2 += tiles * 6;
-                    output0_tm_3 += tiles * 6;
-                    output0_tm_4 += tiles * 6;
-                    output0_tm_5 += tiles * 6;
-                }
-
-                for (int m = 0; m < 4; m++)
-                {
-                    const int* tmp0 = tmp[m];
-
-                    int tmp02a = tmp0[1] + tmp0[2];
-                    int tmp13a = tmp0[1] - tmp0[2];
-
-                    int tmp02b = tmp0[3] + tmp0[4];
-                    int tmp13b = tmp0[3] - tmp0[4];
-
-                    output0[0] = (tmp0[0] + tmp02a + tmp02b) / 576;
-                    output0[1] = (tmp13a + tmp13b * 2) / 576;
-                    output0[2] = (tmp02a + tmp02b * 4) / 576;
-                    output0[3] = (tmp0[5] + tmp13a + tmp13b * 8) / 576;
-
-                    output0 += outw;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_pack4.h
deleted file mode 100644
index 353dd930..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_pack4.h
+++ /dev/null
@@ -1,730 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd63_transform_input_pack4_msa(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 6;
-    const int h_tiles = (h - 2) / 6;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[8][8] = {
-    //     {1.0f,  0.0f, -5.25f,  0.00f,  5.25f,  0.00f, -1.0f, 0.0f},
-    //
-    //     {0.0f,  1.0f,  1.00f, -4.25f, -4.25f,  1.00f,  1.0f, 0.0f},
-    //     {0.0f, -1.0f,  1.00f,  4.25f, -4.25f, -1.00f,  1.0f, 0.0f},
-    //
-    //     {0.0f,  0.5f,  0.25f, -2.50f, -1.25f,  2.00f,  1.0f, 0.0f},
-    //     {0.0f, -0.5f,  0.25f,  2.50f, -1.25f, -2.00f,  1.0f, 0.0f},
-    //
-    //     {0.0f,  2.0f,  4.00f, -2.50f, -5.00f,  0.50f,  1.0f, 0.0f},
-    //     {0.0f, -2.0f,  4.00f,  2.50f, -5.00f, -0.50f,  1.0f, 0.0f},
-    //
-    //     {0.0f, -1.0f,  0.00f,  5.25f,  0.00f, -5.25f,  0.0f, 1.0f}
-    // };
-
-    // 0 = r00 - r06 + (r04 - r02) * 5.25
-    // 7 = r07 - r01 + (r03 - r05) * 5.25
-
-    // 1 = (r02 + r06 - r04 * 4.25) + (r01 - r03 * 4.25 + r05)
-    // 2 = (r02 + r06 - r04 * 4.25) - (r01 - r03 * 4.25 + r05)
-
-    // 3 = (r06 + r02 * 0.25 - r04 * 1.25) + (r01 * 0.5 - r03 * 2.5 + r05 * 2)
-    // 4 = (r06 + r02 * 0.25 - r04 * 1.25) - (r01 * 0.5 - r03 * 2.5 + r05 * 2)
-
-    // reuse r04 * 1.25
-    // reuse r03 * 2.5
-    // 5 = (r06 + (r02 - r04 * 1.25) * 4) + (r01 * 2 - r03 * 2.5 + r05 * 0.5)
-    // 6 = (r06 + (r02 - r04 * 1.25) * 4) - (r01 * 2 - r03 * 2.5 + r05 * 0.5)
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        float tmp[8][8][4];
-
-        v4f32 _v5_25 = __msa_fill_w_f32(5.25f);
-        v4f32 _vm4_25 = __msa_fill_w_f32(-4.25f);
-        v4f32 _vm1_25 = __msa_fill_w_f32(-1.25f);
-        v4f32 _v0_25 = __msa_fill_w_f32(0.25f);
-        v4f32 _vm2_5 = __msa_fill_w_f32(-2.5f);
-        v4f32 _v0_5 = __msa_fill_w_f32(0.5f);
-        v4f32 _v2 = __msa_fill_w_f32(2.f);
-        v4f32 _v4 = __msa_fill_w_f32(4.f);
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* r0 = img0.row(i * 6) + (j * 6) * 4;
-
-                for (int m = 0; m < 8; m++)
-                {
-                    v4f32 _r00 = (v4f32)__msa_ld_w(r0, 0);
-                    v4f32 _r01 = (v4f32)__msa_ld_w(r0 + 4, 0);
-                    v4f32 _r02 = (v4f32)__msa_ld_w(r0 + 4 * 2, 0);
-                    v4f32 _r03 = (v4f32)__msa_ld_w(r0 + 4 * 3, 0);
-                    v4f32 _r04 = (v4f32)__msa_ld_w(r0 + 4 * 4, 0);
-                    v4f32 _r05 = (v4f32)__msa_ld_w(r0 + 4 * 5, 0);
-                    v4f32 _r06 = (v4f32)__msa_ld_w(r0 + 4 * 6, 0);
-                    v4f32 _r07 = (v4f32)__msa_ld_w(r0 + 4 * 7, 0);
-
-                    v4f32 _tmp0m = __msa_fmadd_w(__msa_fsub_w(_r00, _r06), _v5_25, __msa_fsub_w(_r04, _r02));
-                    v4f32 _tmp7m = __msa_fmadd_w(__msa_fsub_w(_r07, _r01), _v5_25, __msa_fsub_w(_r03, _r05));
-                    __msa_st_w((v4i32)_tmp0m, tmp[0][m], 0);
-                    __msa_st_w((v4i32)_tmp7m, tmp[7][m], 0);
-
-                    v4f32 _tmp12a = __msa_fmadd_w(__msa_fadd_w(_r02, _r06), _vm4_25, _r04);
-                    v4f32 _tmp12b = __msa_fmadd_w(__msa_fadd_w(_r01, _r05), _vm4_25, _r03);
-
-                    v4f32 _tmp1m = __msa_fadd_w(_tmp12a, _tmp12b);
-                    v4f32 _tmp2m = __msa_fsub_w(_tmp12a, _tmp12b);
-                    __msa_st_w((v4i32)_tmp1m, tmp[1][m], 0);
-                    __msa_st_w((v4i32)_tmp2m, tmp[2][m], 0);
-
-                    v4f32 _tmp34a = __msa_fmadd_w(__msa_fmadd_w(_r06, _v0_25, _r02), _vm1_25, _r04);
-                    v4f32 _tmp34b = __msa_fmadd_w(__msa_fmadd_w(__msa_fmul_w(_r01, _v0_5), _vm2_5, _r03), _v2, _r05);
-
-                    v4f32 _tmp3m = __msa_fadd_w(_tmp34a, _tmp34b);
-                    v4f32 _tmp4m = __msa_fsub_w(_tmp34a, _tmp34b);
-                    __msa_st_w((v4i32)_tmp3m, tmp[3][m], 0);
-                    __msa_st_w((v4i32)_tmp4m, tmp[4][m], 0);
-
-                    v4f32 _tmp56a = __msa_fmadd_w(_r06, _v4, __msa_fmadd_w(_r02, _vm1_25, _r04));
-                    v4f32 _tmp56b = __msa_fmadd_w(__msa_fmadd_w(__msa_fmul_w(_r01, _v2), _vm2_5, _r03), _v0_5, _r05);
-
-                    v4f32 _tmp5m = __msa_fadd_w(_tmp56a, _tmp56b);
-                    v4f32 _tmp6m = __msa_fsub_w(_tmp56a, _tmp56b);
-                    __msa_st_w((v4i32)_tmp5m, tmp[5][m], 0);
-                    __msa_st_w((v4i32)_tmp6m, tmp[6][m], 0);
-
-                    r0 += w * 4;
-                }
-
-                float* r0_tm_0 = (float*)img0_tm + (i * w_tiles + j) * 4;
-                float* r0_tm_1 = r0_tm_0 + tiles * 4;
-                float* r0_tm_2 = r0_tm_0 + tiles * 4 * 2;
-                float* r0_tm_3 = r0_tm_0 + tiles * 4 * 3;
-                float* r0_tm_4 = r0_tm_0 + tiles * 4 * 4;
-                float* r0_tm_5 = r0_tm_0 + tiles * 4 * 5;
-                float* r0_tm_6 = r0_tm_0 + tiles * 4 * 6;
-                float* r0_tm_7 = r0_tm_0 + tiles * 4 * 7;
-
-                for (int m = 0; m < 8; m++)
-                {
-                    v4f32 _tmp00 = (v4f32)__msa_ld_w(tmp[m][0], 0);
-                    v4f32 _tmp01 = (v4f32)__msa_ld_w(tmp[m][1], 0);
-                    v4f32 _tmp02 = (v4f32)__msa_ld_w(tmp[m][2], 0);
-                    v4f32 _tmp03 = (v4f32)__msa_ld_w(tmp[m][3], 0);
-                    v4f32 _tmp04 = (v4f32)__msa_ld_w(tmp[m][4], 0);
-                    v4f32 _tmp05 = (v4f32)__msa_ld_w(tmp[m][5], 0);
-                    v4f32 _tmp06 = (v4f32)__msa_ld_w(tmp[m][6], 0);
-                    v4f32 _tmp07 = (v4f32)__msa_ld_w(tmp[m][7], 0);
-
-                    v4f32 _r0tm0 = __msa_fmadd_w(__msa_fsub_w(_tmp00, _tmp06), _v5_25, __msa_fsub_w(_tmp04, _tmp02));
-                    v4f32 _r0tm7 = __msa_fmadd_w(__msa_fsub_w(_tmp07, _tmp01), _v5_25, __msa_fsub_w(_tmp03, _tmp05));
-
-                    v4f32 _tmp12a = __msa_fmadd_w(__msa_fadd_w(_tmp02, _tmp06), _vm4_25, _tmp04);
-                    v4f32 _tmp12b = __msa_fmadd_w(__msa_fadd_w(_tmp01, _tmp05), _vm4_25, _tmp03);
-
-                    v4f32 _r0tm1 = __msa_fadd_w(_tmp12a, _tmp12b);
-                    v4f32 _r0tm2 = __msa_fsub_w(_tmp12a, _tmp12b);
-
-                    v4f32 _tmp34a = __msa_fmadd_w(__msa_fmadd_w(_tmp06, _v0_25, _tmp02), _vm1_25, _tmp04);
-                    v4f32 _tmp34b = __msa_fmadd_w(__msa_fmadd_w(__msa_fmul_w(_tmp01, _v0_5), _vm2_5, _tmp03), _v2, _tmp05);
-
-                    v4f32 _r0tm3 = __msa_fadd_w(_tmp34a, _tmp34b);
-                    v4f32 _r0tm4 = __msa_fsub_w(_tmp34a, _tmp34b);
-
-                    v4f32 _tmp56a = __msa_fmadd_w(_tmp06, _v4, __msa_fmadd_w(_tmp02, _vm1_25, _tmp04));
-                    v4f32 _tmp56b = __msa_fmadd_w(__msa_fmadd_w(__msa_fmul_w(_tmp01, _v2), _vm2_5, _tmp03), _v0_5, _tmp05);
-
-                    v4f32 _r0tm5 = __msa_fadd_w(_tmp56a, _tmp56b);
-                    v4f32 _r0tm6 = __msa_fsub_w(_tmp56a, _tmp56b);
-
-                    __msa_st_w((v4i32)_r0tm0, r0_tm_0, 0);
-                    __msa_st_w((v4i32)_r0tm1, r0_tm_1, 0);
-                    __msa_st_w((v4i32)_r0tm2, r0_tm_2, 0);
-                    __msa_st_w((v4i32)_r0tm3, r0_tm_3, 0);
-                    __msa_st_w((v4i32)_r0tm4, r0_tm_4, 0);
-                    __msa_st_w((v4i32)_r0tm5, r0_tm_5, 0);
-                    __msa_st_w((v4i32)_r0tm6, r0_tm_6, 0);
-                    __msa_st_w((v4i32)_r0tm7, r0_tm_7, 0);
-
-                    r0_tm_0 += tiles * 4 * 8;
-                    r0_tm_1 += tiles * 4 * 8;
-                    r0_tm_2 += tiles * 4 * 8;
-                    r0_tm_3 += tiles * 4 * 8;
-                    r0_tm_4 += tiles * 4 * 8;
-                    r0_tm_5 += tiles * 4 * 8;
-                    r0_tm_6 += tiles * 4 * 8;
-                    r0_tm_7 += tiles * 4 * 8;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd63_transform_output_pack4_msa(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 6;
-    const int h_tiles = outh / 6;
-    const int tiles = w_tiles * h_tiles;
-
-    const float* biasptr = bias;
-
-    // const float otm[6][8] = {
-    //     {1.0f,  1.0f,   1.0f,   1.0f,   1.0f,  32.0f, 32.0f, 0.0f},
-    //     {0.0f,  1.0f,  -1.0f,   2.0f,  -2.0f,  16.0f,-16.0f, 0.0f},
-    //     {0.0f,  1.0f,   1.0f,   4.0f,   4.0f,   8.0f,  8.0f, 0.0f},
-    //     {0.0f,  1.0f,  -1.0f,   8.0f,  -8.0f,   4.0f, -4.0f, 0.0f},
-    //     {0.0f,  1.0f,   1.0f,  16.0f,  16.0f,   2.0f,  2.0f, 0.0f},
-    //     {0.0f,  1.0f,  -1.0f,  32.0f, -32.0f,   1.0f, -1.0f, 1.0f}
-    // };
-
-    // 0 = r0 + (r1 + r2) + (r3 + r4)     + (r5 + r6) * 32
-    // 1 =      (r1 - r2) + (r3 - r4) * 2 + (r5 - r6) * 16
-    // 2 =      (r1 + r2) + (r3 + r4) * 4 + (r5 + r6) * 8
-    // 3 =      (r1 - r2) + (r3 - r4) * 8 + (r5 - r6) * 4
-    // 4 =      (r1 + r2) + (r3 + r4) * 16+ (r5 + r6) * 2
-    // 5 = r7 + (r1 - r2) + (r3 - r4) * 32+ (r5 - r6)
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        v4f32 _bias0 = biasptr ? (v4f32)__msa_ld_w(biasptr + p * 4, 0) : (v4f32)__msa_fill_w(0);
-
-        float tmp[6][8][4];
-
-        v4f32 _v32 = __msa_fill_w_f32(32.f);
-        v4f32 _v16 = __msa_fill_w_f32(16.f);
-        v4f32 _v8 = __msa_fill_w_f32(8.f);
-        v4f32 _v4 = __msa_fill_w_f32(4.f);
-        v4f32 _v2 = __msa_fill_w_f32(2.f);
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* output0_tm_0 = (const float*)out0_tm + (i * w_tiles + j) * 4;
-                const float* output0_tm_1 = output0_tm_0 + tiles * 4;
-                const float* output0_tm_2 = output0_tm_0 + tiles * 4 * 2;
-                const float* output0_tm_3 = output0_tm_0 + tiles * 4 * 3;
-                const float* output0_tm_4 = output0_tm_0 + tiles * 4 * 4;
-                const float* output0_tm_5 = output0_tm_0 + tiles * 4 * 5;
-                const float* output0_tm_6 = output0_tm_0 + tiles * 4 * 6;
-                const float* output0_tm_7 = output0_tm_0 + tiles * 4 * 7;
-
-                float* output0 = out0.row<float>(i * 6) + (j * 6) * 4;
-
-                for (int m = 0; m < 8; m++)
-                {
-                    v4f32 _out0tm0 = (v4f32)__msa_ld_w(output0_tm_0, 0);
-                    v4f32 _out0tm1 = (v4f32)__msa_ld_w(output0_tm_1, 0);
-                    v4f32 _out0tm2 = (v4f32)__msa_ld_w(output0_tm_2, 0);
-                    v4f32 _out0tm3 = (v4f32)__msa_ld_w(output0_tm_3, 0);
-                    v4f32 _out0tm4 = (v4f32)__msa_ld_w(output0_tm_4, 0);
-                    v4f32 _out0tm5 = (v4f32)__msa_ld_w(output0_tm_5, 0);
-                    v4f32 _out0tm6 = (v4f32)__msa_ld_w(output0_tm_6, 0);
-                    v4f32 _out0tm7 = (v4f32)__msa_ld_w(output0_tm_7, 0);
-
-                    v4f32 _tmp024a = __msa_fadd_w(_out0tm1, _out0tm2);
-                    v4f32 _tmp135a = __msa_fsub_w(_out0tm1, _out0tm2);
-
-                    v4f32 _tmp024b = __msa_fadd_w(_out0tm3, _out0tm4);
-                    v4f32 _tmp135b = __msa_fsub_w(_out0tm3, _out0tm4);
-
-                    v4f32 _tmp024c = __msa_fadd_w(_out0tm5, _out0tm6);
-                    v4f32 _tmp135c = __msa_fsub_w(_out0tm5, _out0tm6);
-
-                    v4f32 _tmp0m = __msa_fadd_w(__msa_fadd_w(_out0tm0, _tmp024a), __msa_fmadd_w(_tmp024b, _v32, _tmp024c));
-                    v4f32 _tmp2m = __msa_fmadd_w(__msa_fmadd_w(_tmp024a, _v4, _tmp024b), _v8, _tmp024c);
-                    v4f32 _tmp4m = __msa_fmadd_w(__msa_fmadd_w(_tmp024a, _v16, _tmp024b), _v2, _tmp024c);
-                    __msa_st_w((v4i32)_tmp0m, tmp[0][m], 0);
-                    __msa_st_w((v4i32)_tmp2m, tmp[2][m], 0);
-                    __msa_st_w((v4i32)_tmp4m, tmp[4][m], 0);
-
-                    v4f32 _tmp1m = __msa_fmadd_w(__msa_fmadd_w(_tmp135a, _v2, _tmp135b), _v16, _tmp135c);
-                    v4f32 _tmp3m = __msa_fmadd_w(__msa_fmadd_w(_tmp135a, _v8, _tmp135b), _v4, _tmp135c);
-                    v4f32 _tmp5m = __msa_fadd_w(__msa_fadd_w(_out0tm7, _tmp135a), __msa_fmadd_w(_tmp135c, _v32, _tmp135b));
-                    __msa_st_w((v4i32)_tmp1m, tmp[1][m], 0);
-                    __msa_st_w((v4i32)_tmp3m, tmp[3][m], 0);
-                    __msa_st_w((v4i32)_tmp5m, tmp[5][m], 0);
-
-                    output0_tm_0 += tiles * 4 * 8;
-                    output0_tm_1 += tiles * 4 * 8;
-                    output0_tm_2 += tiles * 4 * 8;
-                    output0_tm_3 += tiles * 4 * 8;
-                    output0_tm_4 += tiles * 4 * 8;
-                    output0_tm_5 += tiles * 4 * 8;
-                    output0_tm_6 += tiles * 4 * 8;
-                    output0_tm_7 += tiles * 4 * 8;
-                }
-
-                for (int m = 0; m < 6; m++)
-                {
-                    v4f32 _tmp00 = (v4f32)__msa_ld_w(tmp[m][0], 0);
-                    v4f32 _tmp01 = (v4f32)__msa_ld_w(tmp[m][1], 0);
-                    v4f32 _tmp02 = (v4f32)__msa_ld_w(tmp[m][2], 0);
-                    v4f32 _tmp03 = (v4f32)__msa_ld_w(tmp[m][3], 0);
-                    v4f32 _tmp04 = (v4f32)__msa_ld_w(tmp[m][4], 0);
-                    v4f32 _tmp05 = (v4f32)__msa_ld_w(tmp[m][5], 0);
-                    v4f32 _tmp06 = (v4f32)__msa_ld_w(tmp[m][6], 0);
-                    v4f32 _tmp07 = (v4f32)__msa_ld_w(tmp[m][7], 0);
-
-                    v4f32 _tmp024a = __msa_fadd_w(_tmp01, _tmp02);
-                    v4f32 _tmp135a = __msa_fsub_w(_tmp01, _tmp02);
-
-                    v4f32 _tmp024b = __msa_fadd_w(_tmp03, _tmp04);
-                    v4f32 _tmp135b = __msa_fsub_w(_tmp03, _tmp04);
-
-                    v4f32 _tmp024c = __msa_fadd_w(_tmp05, _tmp06);
-                    v4f32 _tmp135c = __msa_fsub_w(_tmp05, _tmp06);
-
-                    v4f32 _out00 = __msa_fadd_w(_bias0, __msa_fadd_w(__msa_fadd_w(_tmp00, _tmp024a), __msa_fmadd_w(_tmp024b, _v32, _tmp024c)));
-                    v4f32 _out02 = __msa_fadd_w(_bias0, __msa_fmadd_w(__msa_fmadd_w(_tmp024a, _v4, _tmp024b), _v8, _tmp024c));
-                    v4f32 _out04 = __msa_fadd_w(_bias0, __msa_fmadd_w(__msa_fmadd_w(_tmp024a, _v16, _tmp024b), _v2, _tmp024c));
-                    __msa_st_w((v4i32)_out00, output0, 0);
-                    __msa_st_w((v4i32)_out02, output0 + 4 * 2, 0);
-                    __msa_st_w((v4i32)_out04, output0 + 4 * 4, 0);
-
-                    v4f32 _out01 = __msa_fadd_w(_bias0, __msa_fmadd_w(__msa_fmadd_w(_tmp135a, _v2, _tmp135b), _v16, _tmp135c));
-                    v4f32 _out03 = __msa_fadd_w(_bias0, __msa_fmadd_w(__msa_fmadd_w(_tmp135a, _v8, _tmp135b), _v4, _tmp135c));
-                    v4f32 _out05 = __msa_fadd_w(_bias0, __msa_fadd_w(__msa_fadd_w(_tmp07, _tmp135a), __msa_fmadd_w(_tmp135c, _v32, _tmp135b)));
-                    __msa_st_w((v4i32)_out01, output0 + 4, 0);
-                    __msa_st_w((v4i32)_out03, output0 + 4 * 3, 0);
-                    __msa_st_w((v4i32)_out05, output0 + 4 * 5, 0);
-
-                    output0 += outw * 4;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_input_pack4_msa(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 4;
-    const int h_tiles = (h - 2) / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[6][6] = {
-    //     {4.0f, 0.0f, -5.0f, 0.0f, 1.0f, 0.0f},
-    //     {0.0f,-4.0f, -4.0f, 1.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f, -4.0f,-1.0f, 1.0f, 0.0f},
-    //     {0.0f,-2.0f, -1.0f, 2.0f, 1.0f, 0.0f},
-    //     {0.0f, 2.0f, -1.0f,-2.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f,  0.0f,-5.0f, 0.0f, 1.0f}
-    // };
-
-    // 0 =  4 * r00 - 5 * r02 + r04
-    // 1 = -4 * (r01 + r02) + r04 + r03
-    // 2 =  4 * (r01 - r02) + r04 - r03
-    // 3 = -2 * (r01 - r03) + r04 - r02
-    // 4 =  2 * (r01 - r03) + r04 - r02
-    // 5 =  4 * r01 - 5 * r03 + r05
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        float tmp[6][6][4];
-
-        v4f32 _vm5 = __msa_fill_w_f32(-5.f);
-        v4f32 _vm4 = __msa_fill_w_f32(-4.f);
-        v4f32 _v4 = __msa_fill_w_f32(4.f);
-        v4f32 _vm2 = __msa_fill_w_f32(-2.f);
-        v4f32 _v2 = __msa_fill_w_f32(2.f);
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* r0 = img0.row(i * 4) + (j * 4) * 4;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    v4f32 _r00 = (v4f32)__msa_ld_w(r0, 0);
-                    v4f32 _r01 = (v4f32)__msa_ld_w(r0 + 4, 0);
-                    v4f32 _r02 = (v4f32)__msa_ld_w(r0 + 4 * 2, 0);
-                    v4f32 _r03 = (v4f32)__msa_ld_w(r0 + 4 * 3, 0);
-                    v4f32 _r04 = (v4f32)__msa_ld_w(r0 + 4 * 4, 0);
-                    v4f32 _r05 = (v4f32)__msa_ld_w(r0 + 4 * 5, 0);
-
-                    v4f32 _tmp0m = __msa_fmadd_w(__msa_fmadd_w(_r04, _v4, _r00), _vm5, _r02);
-                    v4f32 _tmp1m = __msa_fmadd_w(__msa_fadd_w(_r04, _r03), _vm4, __msa_fadd_w(_r01, _r02));
-                    v4f32 _tmp2m = __msa_fmadd_w(__msa_fsub_w(_r04, _r03), _v4, __msa_fsub_w(_r01, _r02));
-                    v4f32 _tmp3m = __msa_fmadd_w(__msa_fsub_w(_r04, _r02), _vm2, __msa_fsub_w(_r01, _r03));
-                    v4f32 _tmp4m = __msa_fmadd_w(__msa_fsub_w(_r04, _r02), _v2, __msa_fsub_w(_r01, _r03));
-                    v4f32 _tmp5m = __msa_fmadd_w(__msa_fmadd_w(_r05, _v4, _r01), _vm5, _r03);
-
-                    __msa_st_w((v4i32)_tmp0m, tmp[0][m], 0);
-                    __msa_st_w((v4i32)_tmp1m, tmp[1][m], 0);
-                    __msa_st_w((v4i32)_tmp2m, tmp[2][m], 0);
-                    __msa_st_w((v4i32)_tmp3m, tmp[3][m], 0);
-                    __msa_st_w((v4i32)_tmp4m, tmp[4][m], 0);
-                    __msa_st_w((v4i32)_tmp5m, tmp[5][m], 0);
-
-                    r0 += w * 4;
-                }
-
-                float* r0_tm_0 = (float*)img0_tm + (i * w_tiles + j) * 4;
-                float* r0_tm_1 = r0_tm_0 + tiles * 4;
-                float* r0_tm_2 = r0_tm_0 + tiles * 4 * 2;
-                float* r0_tm_3 = r0_tm_0 + tiles * 4 * 3;
-                float* r0_tm_4 = r0_tm_0 + tiles * 4 * 4;
-                float* r0_tm_5 = r0_tm_0 + tiles * 4 * 5;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    v4f32 _tmp00 = (v4f32)__msa_ld_w(tmp[m][0], 0);
-                    v4f32 _tmp01 = (v4f32)__msa_ld_w(tmp[m][1], 0);
-                    v4f32 _tmp02 = (v4f32)__msa_ld_w(tmp[m][2], 0);
-                    v4f32 _tmp03 = (v4f32)__msa_ld_w(tmp[m][3], 0);
-                    v4f32 _tmp04 = (v4f32)__msa_ld_w(tmp[m][4], 0);
-                    v4f32 _tmp05 = (v4f32)__msa_ld_w(tmp[m][5], 0);
-
-                    v4f32 _r0tm0 = __msa_fmadd_w(__msa_fmadd_w(_tmp04, _v4, _tmp00), _vm5, _tmp02);
-                    v4f32 _r0tm1 = __msa_fmadd_w(__msa_fadd_w(_tmp04, _tmp03), _vm4, __msa_fadd_w(_tmp01, _tmp02));
-                    v4f32 _r0tm2 = __msa_fmadd_w(__msa_fsub_w(_tmp04, _tmp03), _v4, __msa_fsub_w(_tmp01, _tmp02));
-                    v4f32 _r0tm3 = __msa_fmadd_w(__msa_fsub_w(_tmp04, _tmp02), _vm2, __msa_fsub_w(_tmp01, _tmp03));
-                    v4f32 _r0tm4 = __msa_fmadd_w(__msa_fsub_w(_tmp04, _tmp02), _v2, __msa_fsub_w(_tmp01, _tmp03));
-                    v4f32 _r0tm5 = __msa_fmadd_w(__msa_fmadd_w(_tmp05, _v4, _tmp01), _vm5, _tmp03);
-
-                    __msa_st_w((v4i32)_r0tm0, r0_tm_0, 0);
-                    __msa_st_w((v4i32)_r0tm1, r0_tm_1, 0);
-                    __msa_st_w((v4i32)_r0tm2, r0_tm_2, 0);
-                    __msa_st_w((v4i32)_r0tm3, r0_tm_3, 0);
-                    __msa_st_w((v4i32)_r0tm4, r0_tm_4, 0);
-                    __msa_st_w((v4i32)_r0tm5, r0_tm_5, 0);
-
-                    r0_tm_0 += tiles * 4 * 6;
-                    r0_tm_1 += tiles * 4 * 6;
-                    r0_tm_2 += tiles * 4 * 6;
-                    r0_tm_3 += tiles * 4 * 6;
-                    r0_tm_4 += tiles * 4 * 6;
-                    r0_tm_5 += tiles * 4 * 6;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_output_pack4_msa(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 4;
-    const int h_tiles = outh / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    const float* biasptr = bias;
-
-    // const float otm[4][6] = {
-    //     {1.0f, 1.0f,  1.0f, 1.0f,  1.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 2.0f, -2.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f, 4.0f,  4.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 8.0f, -8.0f, 1.0f}
-    // };
-
-    // 0 = r00 + (r01 + r02) + (r03 + r04)
-    // 1 =       (r01 - r02) + (r03 - r04) * 2
-    // 2 =       (r01 + r02) + (r03 + r04) * 4
-    // 3 = r05 + (r01 - r02) + (r03 - r04) * 8
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        v4f32 _bias0 = biasptr ? (v4f32)__msa_ld_w(biasptr + p * 4, 0) : (v4f32)__msa_fill_w(0);
-
-        float tmp[4][6][4];
-
-        v4f32 _v2 = __msa_fill_w_f32(2.f);
-        v4f32 _v4 = __msa_fill_w_f32(4.f);
-        v4f32 _v8 = __msa_fill_w_f32(8.f);
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* output0_tm_0 = (const float*)out0_tm + (i * w_tiles + j) * 4;
-                const float* output0_tm_1 = output0_tm_0 + tiles * 4;
-                const float* output0_tm_2 = output0_tm_0 + tiles * 4 * 2;
-                const float* output0_tm_3 = output0_tm_0 + tiles * 4 * 3;
-                const float* output0_tm_4 = output0_tm_0 + tiles * 4 * 4;
-                const float* output0_tm_5 = output0_tm_0 + tiles * 4 * 5;
-
-                float* output0 = out0.row<float>(i * 4) + (j * 4) * 4;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    v4f32 _out0tm0 = (v4f32)__msa_ld_w(output0_tm_0, 0);
-                    v4f32 _out0tm1 = (v4f32)__msa_ld_w(output0_tm_1, 0);
-                    v4f32 _out0tm2 = (v4f32)__msa_ld_w(output0_tm_2, 0);
-                    v4f32 _out0tm3 = (v4f32)__msa_ld_w(output0_tm_3, 0);
-                    v4f32 _out0tm4 = (v4f32)__msa_ld_w(output0_tm_4, 0);
-                    v4f32 _out0tm5 = (v4f32)__msa_ld_w(output0_tm_5, 0);
-
-                    v4f32 _tmp02a = __msa_fadd_w(_out0tm1, _out0tm2);
-                    v4f32 _tmp13a = __msa_fsub_w(_out0tm1, _out0tm2);
-
-                    v4f32 _tmp02b = __msa_fadd_w(_out0tm3, _out0tm4);
-                    v4f32 _tmp13b = __msa_fsub_w(_out0tm3, _out0tm4);
-
-                    v4f32 _tmp0m = __msa_fadd_w(__msa_fadd_w(_out0tm0, _tmp02a), _tmp02b);
-                    v4f32 _tmp1m = __msa_fmadd_w(_tmp13a, _v2, _tmp13b);
-                    v4f32 _tmp2m = __msa_fmadd_w(_tmp02a, _v4, _tmp02b);
-                    v4f32 _tmp3m = __msa_fmadd_w(__msa_fadd_w(_out0tm5, _tmp13a), _v8, _tmp13b);
-
-                    __msa_st_w((v4i32)_tmp0m, tmp[0][m], 0);
-                    __msa_st_w((v4i32)_tmp1m, tmp[1][m], 0);
-                    __msa_st_w((v4i32)_tmp2m, tmp[2][m], 0);
-                    __msa_st_w((v4i32)_tmp3m, tmp[3][m], 0);
-
-                    output0_tm_0 += tiles * 4 * 6;
-                    output0_tm_1 += tiles * 4 * 6;
-                    output0_tm_2 += tiles * 4 * 6;
-                    output0_tm_3 += tiles * 4 * 6;
-                    output0_tm_4 += tiles * 4 * 6;
-                    output0_tm_5 += tiles * 4 * 6;
-                }
-
-                for (int m = 0; m < 4; m++)
-                {
-                    v4f32 _tmp00 = (v4f32)__msa_ld_w(tmp[m][0], 0);
-                    v4f32 _tmp01 = (v4f32)__msa_ld_w(tmp[m][1], 0);
-                    v4f32 _tmp02 = (v4f32)__msa_ld_w(tmp[m][2], 0);
-                    v4f32 _tmp03 = (v4f32)__msa_ld_w(tmp[m][3], 0);
-                    v4f32 _tmp04 = (v4f32)__msa_ld_w(tmp[m][4], 0);
-                    v4f32 _tmp05 = (v4f32)__msa_ld_w(tmp[m][5], 0);
-
-                    v4f32 _tmp02a = __msa_fadd_w(_tmp01, _tmp02);
-                    v4f32 _tmp13a = __msa_fsub_w(_tmp01, _tmp02);
-
-                    v4f32 _tmp02b = __msa_fadd_w(_tmp03, _tmp04);
-                    v4f32 _tmp13b = __msa_fsub_w(_tmp03, _tmp04);
-
-                    v4f32 _out00 = __msa_fadd_w(_bias0, __msa_fadd_w(__msa_fadd_w(_tmp00, _tmp02a), _tmp02b));
-                    v4f32 _out01 = __msa_fadd_w(_bias0, __msa_fmadd_w(_tmp13a, _v2, _tmp13b));
-                    v4f32 _out02 = __msa_fadd_w(_bias0, __msa_fmadd_w(_tmp02a, _v4, _tmp02b));
-                    v4f32 _out03 = __msa_fadd_w(_bias0, __msa_fmadd_w(__msa_fadd_w(_tmp05, _tmp13a), _v8, _tmp13b));
-
-                    __msa_st_w((v4i32)_out00, output0, 0);
-                    __msa_st_w((v4i32)_out01, output0 + 4, 0);
-                    __msa_st_w((v4i32)_out02, output0 + 4 * 2, 0);
-                    __msa_st_w((v4i32)_out03, output0 + 4 * 3, 0);
-
-                    output0 += outw * 4;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_input_pack4_msa(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 2;
-    const int h_tiles = (h - 2) / 2;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[4][4] = {
-    //     {1.0f,  0.0f, -1.0f,  0.0f},
-    //     {0.0f,  1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  0.00f, 1.0f}
-    // };
-
-    // 0 = r00 - r02
-    // 1 = r01 + r02
-    // 2 = r02 - r01
-    // 3 = r03 - r01
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        float tmp[4][4][4];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* r0 = img0.row(i * 2) + (j * 2) * 4;
-
-                for (int m = 0; m < 4; m++)
-                {
-                    v4f32 _r00 = (v4f32)__msa_ld_w(r0, 0);
-                    v4f32 _r01 = (v4f32)__msa_ld_w(r0 + 4, 0);
-                    v4f32 _r02 = (v4f32)__msa_ld_w(r0 + 4 * 2, 0);
-                    v4f32 _r03 = (v4f32)__msa_ld_w(r0 + 4 * 3, 0);
-
-                    v4f32 _tmp0m = __msa_fsub_w(_r00, _r02);
-                    v4f32 _tmp1m = __msa_fadd_w(_r01, _r02);
-                    v4f32 _tmp2m = __msa_fsub_w(_r02, _r01);
-                    v4f32 _tmp3m = __msa_fsub_w(_r03, _r01);
-
-                    __msa_st_w((v4i32)_tmp0m, tmp[0][m], 0);
-                    __msa_st_w((v4i32)_tmp1m, tmp[1][m], 0);
-                    __msa_st_w((v4i32)_tmp2m, tmp[2][m], 0);
-                    __msa_st_w((v4i32)_tmp3m, tmp[3][m], 0);
-
-                    r0 += w * 4;
-                }
-
-                float* r0_tm_0 = (float*)img0_tm + (i * w_tiles + j) * 4;
-                float* r0_tm_1 = r0_tm_0 + tiles * 4;
-                float* r0_tm_2 = r0_tm_0 + tiles * 4 * 2;
-                float* r0_tm_3 = r0_tm_0 + tiles * 4 * 3;
-
-                for (int m = 0; m < 4; m++)
-                {
-                    v4f32 _tmp00 = (v4f32)__msa_ld_w(tmp[m][0], 0);
-                    v4f32 _tmp01 = (v4f32)__msa_ld_w(tmp[m][1], 0);
-                    v4f32 _tmp02 = (v4f32)__msa_ld_w(tmp[m][2], 0);
-                    v4f32 _tmp03 = (v4f32)__msa_ld_w(tmp[m][3], 0);
-
-                    v4f32 _r0tm0 = __msa_fsub_w(_tmp00, _tmp02);
-                    v4f32 _r0tm1 = __msa_fadd_w(_tmp01, _tmp02);
-                    v4f32 _r0tm2 = __msa_fsub_w(_tmp02, _tmp01);
-                    v4f32 _r0tm3 = __msa_fsub_w(_tmp03, _tmp01);
-
-                    __msa_st_w((v4i32)_r0tm0, r0_tm_0, 0);
-                    __msa_st_w((v4i32)_r0tm1, r0_tm_1, 0);
-                    __msa_st_w((v4i32)_r0tm2, r0_tm_2, 0);
-                    __msa_st_w((v4i32)_r0tm3, r0_tm_3, 0);
-
-                    r0_tm_0 += tiles * 4 * 4;
-                    r0_tm_1 += tiles * 4 * 4;
-                    r0_tm_2 += tiles * 4 * 4;
-                    r0_tm_3 += tiles * 4 * 4;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_output_pack4_msa(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 2;
-    const int h_tiles = outh / 2;
-    const int tiles = w_tiles * h_tiles;
-
-    const float* biasptr = bias;
-
-    // const float otm[2][4] = {
-    //     {1.0f,  1.0f,  1.0f,  0.0f},
-    //     {0.0f,  1.0f, -1.0f,  1.0f}
-    // };
-
-    // 0 = r00 + r01 + r02
-    // 1 = r01 - r02 + r03
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        v4f32 _bias0 = biasptr ? (v4f32)__msa_ld_w(biasptr + p * 4, 0) : (v4f32)__msa_fill_w(0);
-
-        float tmp[2][4][4];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* output0_tm_0 = (const float*)out0_tm + (i * w_tiles + j) * 4;
-                const float* output0_tm_1 = output0_tm_0 + tiles * 4;
-                const float* output0_tm_2 = output0_tm_0 + tiles * 4 * 2;
-                const float* output0_tm_3 = output0_tm_0 + tiles * 4 * 3;
-
-                float* output0 = out0.row<float>(i * 2) + (j * 2) * 4;
-
-                for (int m = 0; m < 4; m++)
-                {
-                    v4f32 _out0tm0 = (v4f32)__msa_ld_w(output0_tm_0, 0);
-                    v4f32 _out0tm1 = (v4f32)__msa_ld_w(output0_tm_1, 0);
-                    v4f32 _out0tm2 = (v4f32)__msa_ld_w(output0_tm_2, 0);
-                    v4f32 _out0tm3 = (v4f32)__msa_ld_w(output0_tm_3, 0);
-
-                    v4f32 _tmp0m = __msa_fadd_w(__msa_fadd_w(_out0tm0, _out0tm1), _out0tm2);
-                    v4f32 _tmp1m = __msa_fadd_w(__msa_fsub_w(_out0tm1, _out0tm2), _out0tm3);
-
-                    __msa_st_w((v4i32)_tmp0m, tmp[0][m], 0);
-                    __msa_st_w((v4i32)_tmp1m, tmp[1][m], 0);
-
-                    output0_tm_0 += tiles * 4 * 4;
-                    output0_tm_1 += tiles * 4 * 4;
-                    output0_tm_2 += tiles * 4 * 4;
-                    output0_tm_3 += tiles * 4 * 4;
-                }
-
-                for (int m = 0; m < 2; m++)
-                {
-                    v4f32 _tmp00 = (v4f32)__msa_ld_w(tmp[m][0], 0);
-                    v4f32 _tmp01 = (v4f32)__msa_ld_w(tmp[m][1], 0);
-                    v4f32 _tmp02 = (v4f32)__msa_ld_w(tmp[m][2], 0);
-                    v4f32 _tmp03 = (v4f32)__msa_ld_w(tmp[m][3], 0);
-
-                    v4f32 _out00 = __msa_fadd_w(_bias0, __msa_fadd_w(__msa_fadd_w(_tmp00, _tmp01), _tmp02));
-                    v4f32 _out01 = __msa_fadd_w(_bias0, __msa_fadd_w(__msa_fsub_w(_tmp01, _tmp02), _tmp03));
-
-                    __msa_st_w((v4i32)_out00, output0, 0);
-                    __msa_st_w((v4i32)_out01, output0 + 4, 0);
-
-                    output0 += outw * 4;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_pack4_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_pack4_int8.h
deleted file mode 100644
index a8d7db65..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_pack4_int8.h
+++ /dev/null
@@ -1,166 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd43_transform_output_pack4_int8_msa(const Mat& top_blob_tm, Mat& top_blob, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 4;
-    const int h_tiles = outh / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float otm[4][6] = {
-    //     {1.0f, 1.0f,  1.0f, 1.0f,  1.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 2.0f, -2.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f, 4.0f,  4.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 8.0f, -8.0f, 1.0f}
-    // };
-
-    // 0 = r00 + (r01 + r02) + (r03 + r04)
-    // 1 =       (r01 - r02) + (r03 - r04) * 2
-    // 2 =       (r01 + r02) + (r03 + r04) * 4
-    // 3 = r05 + (r01 - r02) + (r03 - r04) * 8
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        int tmp[4][6][4];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const int* output0_tm_0 = (const int*)out0_tm + (i * w_tiles + j) * 4;
-                const int* output0_tm_1 = output0_tm_0 + tiles * 4;
-                const int* output0_tm_2 = output0_tm_0 + tiles * 8;
-                const int* output0_tm_3 = output0_tm_0 + tiles * 12;
-                const int* output0_tm_4 = output0_tm_0 + tiles * 16;
-                const int* output0_tm_5 = output0_tm_0 + tiles * 20;
-
-                int* output0 = out0.row<int>(i * 4) + (j * 4) * 4;
-
-                for (int m = 0; m < 5; m++)
-                {
-                    v4i32 _out0tm0 = __msa_ld_w(output0_tm_0, 0);
-                    v4i32 _out0tm1 = __msa_ld_w(output0_tm_1, 0);
-                    v4i32 _out0tm2 = __msa_ld_w(output0_tm_2, 0);
-                    v4i32 _out0tm3 = __msa_ld_w(output0_tm_3, 0);
-                    v4i32 _out0tm4 = __msa_ld_w(output0_tm_4, 0);
-                    v4i32 _out0tm5 = __msa_ld_w(output0_tm_5, 0);
-
-                    v4i32 _tmp02a = __msa_addv_w(_out0tm1, _out0tm2);
-                    v4i32 _tmp13a = __msa_subv_w(_out0tm1, _out0tm2);
-
-                    v4i32 _tmp02b = __msa_addv_w(_out0tm3, _out0tm4);
-                    v4i32 _tmp13b = __msa_subv_w(_out0tm3, _out0tm4);
-
-                    v4i32 _tmp0m = __msa_addv_w(__msa_addv_w(_out0tm0, _tmp02a), _tmp02b);
-                    v4i32 _tmp1m = __msa_addv_w(_tmp13a, __msa_slli_w(_tmp13b, 1));
-                    v4i32 _tmp2m = __msa_addv_w(_tmp02a, __msa_slli_w(_tmp02b, 2));
-                    v4i32 _tmp3m = __msa_addv_w(__msa_addv_w(_tmp13a, __msa_slli_w(_out0tm5, 2)), __msa_slli_w(_tmp13b, 3));
-
-                    __msa_st_w(_tmp0m, tmp[0][m], 0);
-                    __msa_st_w(_tmp1m, tmp[1][m], 0);
-                    __msa_st_w(_tmp2m, tmp[2][m], 0);
-                    __msa_st_w(_tmp3m, tmp[3][m], 0);
-
-                    output0_tm_0 += tiles * 24;
-                    output0_tm_1 += tiles * 24;
-                    output0_tm_2 += tiles * 24;
-                    output0_tm_3 += tiles * 24;
-                    output0_tm_4 += tiles * 24;
-                    output0_tm_5 += tiles * 24;
-                }
-                for (int m = 5; m < 6; m++)
-                {
-                    v4i32 _out0tm0 = __msa_ld_w(output0_tm_0, 0);
-                    v4i32 _out0tm1 = __msa_ld_w(output0_tm_1, 0);
-                    v4i32 _out0tm2 = __msa_ld_w(output0_tm_2, 0);
-                    v4i32 _out0tm3 = __msa_ld_w(output0_tm_3, 0);
-                    v4i32 _out0tm4 = __msa_ld_w(output0_tm_4, 0);
-                    v4i32 _out0tm5 = __msa_ld_w(output0_tm_5, 0);
-
-                    v4i32 _tmp02a = __msa_addv_w(_out0tm1, _out0tm2);
-                    v4i32 _tmp13a = __msa_subv_w(_out0tm1, _out0tm2);
-
-                    v4i32 _tmp02b = __msa_addv_w(_out0tm3, _out0tm4);
-                    v4i32 _tmp13b = __msa_subv_w(_out0tm3, _out0tm4);
-
-                    v4i32 _tmp0m = __msa_addv_w(__msa_addv_w(_out0tm0, _tmp02a), _tmp02b);
-                    v4i32 _tmp1m = __msa_addv_w(_tmp13a, __msa_slli_w(_tmp13b, 1));
-                    v4i32 _tmp2m = __msa_addv_w(_tmp02a, __msa_slli_w(_tmp02b, 2));
-                    v4i32 _tmp3m = __msa_addv_w(__msa_addv_w(_tmp13a, __msa_slli_w(_out0tm5, 2)), __msa_slli_w(_tmp13b, 3));
-
-                    _tmp0m = __msa_slli_w(_tmp0m, 2);
-                    _tmp1m = __msa_slli_w(_tmp1m, 2);
-                    _tmp2m = __msa_slli_w(_tmp2m, 2);
-                    _tmp3m = __msa_slli_w(_tmp3m, 2);
-
-                    __msa_st_w(_tmp0m, tmp[0][m], 0);
-                    __msa_st_w(_tmp1m, tmp[1][m], 0);
-                    __msa_st_w(_tmp2m, tmp[2][m], 0);
-                    __msa_st_w(_tmp3m, tmp[3][m], 0);
-
-                    output0_tm_0 += tiles * 24;
-                    output0_tm_1 += tiles * 24;
-                    output0_tm_2 += tiles * 24;
-                    output0_tm_3 += tiles * 24;
-                    output0_tm_4 += tiles * 24;
-                    output0_tm_5 += tiles * 24;
-                }
-
-                for (int m = 0; m < 4; m++)
-                {
-                    v4i32 _tmp00 = __msa_ld_w(tmp[m][0], 0);
-                    v4i32 _tmp01 = __msa_ld_w(tmp[m][1], 0);
-                    v4i32 _tmp02 = __msa_ld_w(tmp[m][2], 0);
-                    v4i32 _tmp03 = __msa_ld_w(tmp[m][3], 0);
-                    v4i32 _tmp04 = __msa_ld_w(tmp[m][4], 0);
-                    v4i32 _tmp05 = __msa_ld_w(tmp[m][5], 0);
-
-                    v4i32 _tmp02a = __msa_addv_w(_tmp01, _tmp02);
-                    v4i32 _tmp13a = __msa_subv_w(_tmp01, _tmp02);
-
-                    v4i32 _tmp02b = __msa_addv_w(_tmp03, _tmp04);
-                    v4i32 _tmp13b = __msa_subv_w(_tmp03, _tmp04);
-
-                    v4i32 _out00 = __msa_addv_w(__msa_addv_w(_tmp00, _tmp02a), _tmp02b);
-                    v4i32 _out01 = __msa_addv_w(_tmp13a, __msa_slli_w(_tmp13b, 1));
-                    v4i32 _out02 = __msa_addv_w(_tmp02a, __msa_slli_w(_tmp02b, 2));
-                    v4i32 _out03 = __msa_addv_w(__msa_addv_w(_tmp05, _tmp13a), __msa_slli_w(_tmp13b, 3));
-
-                    // TODO use integer trick for division by 576
-                    v4f32 _v576 = __msa_fill_w_f32(1.0 / 576);
-                    _out00 = __msa_ftint_s_w(__msa_fmul_w(__msa_ffint_s_w(_out00), _v576));
-                    _out01 = __msa_ftint_s_w(__msa_fmul_w(__msa_ffint_s_w(_out01), _v576));
-                    _out02 = __msa_ftint_s_w(__msa_fmul_w(__msa_ffint_s_w(_out02), _v576));
-                    _out03 = __msa_ftint_s_w(__msa_fmul_w(__msa_ffint_s_w(_out03), _v576));
-
-                    __msa_st_w(_out00, output0, 0);
-                    __msa_st_w(_out01, output0 + 4, 0);
-                    __msa_st_w(_out02, output0 + 8, 0);
-                    __msa_st_w(_out03, output0 + 12, 0);
-
-                    output0 += outw * 4;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_pack8_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_pack8_int8.h
deleted file mode 100644
index 8b188284..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolution_winograd_transform_pack8_int8.h
+++ /dev/null
@@ -1,132 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd43_transform_input_pack8_int8_msa(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 4;
-    const int h_tiles = (h - 2) / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[6][6] = {
-    //     {4.0f, 0.0f, -5.0f, 0.0f, 1.0f, 0.0f},
-    //     {0.0f,-4.0f, -4.0f, 1.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f, -4.0f,-1.0f, 1.0f, 0.0f},
-    //     {0.0f,-2.0f, -1.0f, 2.0f, 1.0f, 0.0f},
-    //     {0.0f, 2.0f, -1.0f,-2.0f, 1.0f, 0.0f},
-    //     {0.0f, 4.0f,  0.0f,-5.0f, 0.0f, 1.0f}
-    // };
-
-    // 0 =  4 * r00 - 5 * r02 + r04
-    // 1 = -4 * (r01 + r02) + r04 + r03
-    // 2 =  4 * (r01 - r02) + r04 - r03
-    // 3 = -2 * (r01 - r03) + r04 - r02
-    // 4 =  2 * (r01 - r03) + r04 - r02
-    // 5 =  4 * r01 - 5 * r03 + r05
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        short tmp[6][6][8];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const signed char* r0 = img0.row<const signed char>(i * 4) + (j * 4) * 8;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    v16i8 _r00_01 = __msa_ld_b(r0, 0);
-                    v16i8 _r02_03 = __msa_ld_b(r0 + 16, 0);
-                    v16i8 _r04_05 = __msa_ld_b(r0 + 32, 0);
-                    v16i8 _extr0001 = __msa_clti_s_b(_r00_01, 0);
-                    v16i8 _extr0203 = __msa_clti_s_b(_r02_03, 0);
-                    v16i8 _extr0405 = __msa_clti_s_b(_r04_05, 0);
-                    v8i16 _r00 = (v8i16)__msa_ilvr_b(_extr0001, _r00_01);
-                    v8i16 _r01 = (v8i16)__msa_ilvl_b(_extr0001, _r00_01);
-                    v8i16 _r02 = (v8i16)__msa_ilvr_b(_extr0203, _r02_03);
-                    v8i16 _r03 = (v8i16)__msa_ilvl_b(_extr0203, _r02_03);
-                    v8i16 _r04 = (v8i16)__msa_ilvr_b(_extr0405, _r04_05);
-                    v8i16 _r05 = (v8i16)__msa_ilvl_b(_extr0405, _r04_05);
-
-                    v8i16 _v5 = __msa_fill_h(5);
-
-                    v8i16 _tmp0m = __msa_subv_h(__msa_addv_h(__msa_slli_h(_r00, 2), _r04), __msa_mulv_h(_r02, _v5));
-                    v8i16 _tmp1m = __msa_subv_h(__msa_addv_h(_r04, _r03), __msa_slli_h(__msa_addv_h(_r01, _r02), 2));
-                    v8i16 _tmp2m = __msa_addv_h(__msa_subv_h(_r04, _r03), __msa_slli_h(__msa_subv_h(_r01, _r02), 2));
-                    v8i16 _tmp3m = __msa_subv_h(__msa_subv_h(_r04, _r02), __msa_slli_h(__msa_subv_h(_r01, _r03), 1));
-                    v8i16 _tmp4m = __msa_addv_h(__msa_subv_h(_r04, _r02), __msa_slli_h(__msa_subv_h(_r01, _r03), 1));
-                    v8i16 _tmp5m = __msa_subv_h(__msa_addv_h(__msa_slli_h(_r01, 2), _r05), __msa_mulv_h(_r03, _v5));
-
-                    __msa_st_h(_tmp0m, tmp[0][m], 0);
-                    __msa_st_h(_tmp1m, tmp[1][m], 0);
-                    __msa_st_h(_tmp2m, tmp[2][m], 0);
-                    __msa_st_h(_tmp3m, tmp[3][m], 0);
-                    __msa_st_h(_tmp4m, tmp[4][m], 0);
-                    __msa_st_h(_tmp5m, tmp[5][m], 0);
-
-                    r0 += w * 8;
-                }
-
-                short* r0_tm_0 = (short*)img0_tm + (i * w_tiles + j) * 8;
-                short* r0_tm_1 = r0_tm_0 + tiles * 8;
-                short* r0_tm_2 = r0_tm_0 + tiles * 16;
-                short* r0_tm_3 = r0_tm_0 + tiles * 24;
-                short* r0_tm_4 = r0_tm_0 + tiles * 32;
-                short* r0_tm_5 = r0_tm_0 + tiles * 40;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    v8i16 _tmp00 = __msa_ld_h(tmp[m][0], 0);
-                    v8i16 _tmp01 = __msa_ld_h(tmp[m][1], 0);
-                    v8i16 _tmp02 = __msa_ld_h(tmp[m][2], 0);
-                    v8i16 _tmp03 = __msa_ld_h(tmp[m][3], 0);
-                    v8i16 _tmp04 = __msa_ld_h(tmp[m][4], 0);
-                    v8i16 _tmp05 = __msa_ld_h(tmp[m][5], 0);
-
-                    v8i16 _v5 = __msa_fill_h(5);
-
-                    v8i16 _r0tm0 = __msa_subv_h(__msa_addv_h(__msa_slli_h(_tmp00, 2), _tmp04), __msa_mulv_h(_tmp02, _v5));
-                    v8i16 _r0tm1 = __msa_subv_h(__msa_addv_h(_tmp04, _tmp03), __msa_slli_h(__msa_addv_h(_tmp01, _tmp02), 2));
-                    v8i16 _r0tm2 = __msa_addv_h(__msa_subv_h(_tmp04, _tmp03), __msa_slli_h(__msa_subv_h(_tmp01, _tmp02), 2));
-                    v8i16 _r0tm3 = __msa_subv_h(__msa_subv_h(_tmp04, _tmp02), __msa_slli_h(__msa_subv_h(_tmp01, _tmp03), 1));
-                    v8i16 _r0tm4 = __msa_addv_h(__msa_subv_h(_tmp04, _tmp02), __msa_slli_h(__msa_subv_h(_tmp01, _tmp03), 1));
-                    v8i16 _r0tm5 = __msa_subv_h(__msa_addv_h(__msa_slli_h(_tmp01, 2), _tmp05), __msa_mulv_h(_tmp03, _v5));
-
-                    __msa_st_h(_r0tm0, r0_tm_0, 0);
-                    __msa_st_h(_r0tm1, r0_tm_1, 0);
-                    __msa_st_h(_r0tm2, r0_tm_2, 0);
-                    __msa_st_h(_r0tm3, r0_tm_3, 0);
-                    __msa_st_h(_r0tm4, r0_tm_4, 0);
-                    __msa_st_h(_r0tm5, r0_tm_5, 0);
-
-                    r0_tm_0 += tiles * 48;
-                    r0_tm_1 += tiles * 48;
-                    r0_tm_2 += tiles * 48;
-                    r0_tm_3 += tiles * 48;
-                    r0_tm_4 += tiles * 48;
-                    r0_tm_5 += tiles * 48;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_3x3.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_3x3.h
deleted file mode 100644
index e487b0b3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_3x3.h
+++ /dev/null
@@ -1,193 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const float bias0 = bias ? bias[g] : 0.f;
-
-        const float* kernel0 = kernel + g * 9;
-
-        float* outptr0 = out;
-        float* outptr1 = outptr0 + outw;
-
-        const float* img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0;
-        const float* r1 = img0 + w;
-        const float* r2 = img0 + w * 2;
-        const float* r3 = img0 + w * 3;
-
-        const float* k0 = kernel0;
-        const float* k1 = kernel0 + 3;
-        const float* k2 = kernel0 + 6;
-
-        int i = 0;
-
-        for (; i + 1 < outh; i += 2)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = bias0;
-                float sum2 = bias0;
-
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum2 += r1[0] * k0[0];
-                sum2 += r1[1] * k0[1];
-                sum2 += r1[2] * k0[2];
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum2 += r2[0] * k1[0];
-                sum2 += r2[1] * k1[1];
-                sum2 += r2[2] * k1[2];
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-                sum2 += r3[0] * k2[0];
-                sum2 += r3[1] * k2[1];
-                sum2 += r3[2] * k2[2];
-
-                *outptr0 = sum;
-                *outptr1 = sum2;
-
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                outptr0++;
-                outptr1++;
-            }
-
-            r0 += 2 + w;
-            r1 += 2 + w;
-            r2 += 2 + w;
-            r3 += 2 + w;
-
-            outptr0 += outw;
-            outptr1 += outw;
-        }
-
-        for (; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = bias0;
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-
-                *outptr0 = sum;
-
-                r0++;
-                r1++;
-                r2++;
-                outptr0++;
-            }
-
-            r0 += 2;
-            r1 += 2;
-            r2 += 2;
-        }
-    }
-}
-
-static void convdw3x3s2_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const float bias0 = bias ? bias[g] : 0.f;
-
-        const float* kernel0 = kernel + g * 9;
-
-        float* outptr = out;
-
-        const float* img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0;
-        const float* r1 = img0 + w;
-        const float* r2 = img0 + w * 2;
-
-        const float* k0 = kernel0;
-        const float* k1 = kernel0 + 3;
-        const float* k2 = kernel0 + 6;
-
-        int i = 0;
-
-        for (; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = bias0;
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-
-                *outptr = sum;
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-                outptr++;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_3x3_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_3x3_pack4.h
deleted file mode 100644
index f25a4453..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_3x3_pack4.h
+++ /dev/null
@@ -1,464 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        v4f32 _bias0 = bias ? (v4f32)__msa_ld_w(bias + g * 4, 0) : (v4f32)__msa_fill_w(0);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-        float* outptr1 = out.row(1);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-
-        v4f32 _k00 = (v4f32)__msa_ld_w(k0, 0);
-        v4f32 _k01 = (v4f32)__msa_ld_w(k0 + 4, 0);
-        v4f32 _k02 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-        v4f32 _k10 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-        v4f32 _k11 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-        v4f32 _k12 = (v4f32)__msa_ld_w(k0 + 4 * 5, 0);
-        v4f32 _k20 = (v4f32)__msa_ld_w(k0 + 4 * 6, 0);
-        v4f32 _k21 = (v4f32)__msa_ld_w(k0 + 4 * 7, 0);
-        v4f32 _k22 = (v4f32)__msa_ld_w(k0 + 4 * 8, 0);
-
-        int i = 0;
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                __builtin_prefetch(r0 + 32);
-                __builtin_prefetch(r1 + 32);
-                __builtin_prefetch(r2 + 32);
-                __builtin_prefetch(r3 + 32);
-
-                v4f32 _sum00 = _bias0;
-                v4f32 _sum01 = _bias0;
-                v4f32 _sum10 = _bias0;
-                v4f32 _sum11 = _bias0;
-
-                v4f32 _r00 = (v4f32)__msa_ld_w(r0, 0);
-                v4f32 _r01 = (v4f32)__msa_ld_w(r0 + 4, 0);
-                v4f32 _r02 = (v4f32)__msa_ld_w(r0 + 4 * 2, 0);
-                v4f32 _r03 = (v4f32)__msa_ld_w(r0 + 4 * 3, 0);
-
-                _sum00 = __msa_fmadd_w(_sum00, _k00, _r00);
-                _sum00 = __msa_fmadd_w(_sum00, _k01, _r01);
-                _sum00 = __msa_fmadd_w(_sum00, _k02, _r02);
-                _sum01 = __msa_fmadd_w(_sum01, _k00, _r01);
-                _sum01 = __msa_fmadd_w(_sum01, _k01, _r02);
-                _sum01 = __msa_fmadd_w(_sum01, _k02, _r03);
-
-                v4f32 _r10 = (v4f32)__msa_ld_w(r1, 0);
-                v4f32 _r11 = (v4f32)__msa_ld_w(r1 + 4, 0);
-                v4f32 _r12 = (v4f32)__msa_ld_w(r1 + 4 * 2, 0);
-                v4f32 _r13 = (v4f32)__msa_ld_w(r1 + 4 * 3, 0);
-
-                _sum00 = __msa_fmadd_w(_sum00, _k10, _r10);
-                _sum00 = __msa_fmadd_w(_sum00, _k11, _r11);
-                _sum00 = __msa_fmadd_w(_sum00, _k12, _r12);
-                _sum01 = __msa_fmadd_w(_sum01, _k10, _r11);
-                _sum01 = __msa_fmadd_w(_sum01, _k11, _r12);
-                _sum01 = __msa_fmadd_w(_sum01, _k12, _r13);
-                _sum10 = __msa_fmadd_w(_sum10, _k00, _r10);
-                _sum10 = __msa_fmadd_w(_sum10, _k01, _r11);
-                _sum10 = __msa_fmadd_w(_sum10, _k02, _r12);
-                _sum11 = __msa_fmadd_w(_sum11, _k00, _r11);
-                _sum11 = __msa_fmadd_w(_sum11, _k01, _r12);
-                _sum11 = __msa_fmadd_w(_sum11, _k02, _r13);
-
-                v4f32 _r20 = (v4f32)__msa_ld_w(r2, 0);
-                v4f32 _r21 = (v4f32)__msa_ld_w(r2 + 4, 0);
-                v4f32 _r22 = (v4f32)__msa_ld_w(r2 + 4 * 2, 0);
-                v4f32 _r23 = (v4f32)__msa_ld_w(r2 + 4 * 3, 0);
-
-                _sum00 = __msa_fmadd_w(_sum00, _k20, _r20);
-                _sum00 = __msa_fmadd_w(_sum00, _k21, _r21);
-                _sum00 = __msa_fmadd_w(_sum00, _k22, _r22);
-                _sum01 = __msa_fmadd_w(_sum01, _k20, _r21);
-                _sum01 = __msa_fmadd_w(_sum01, _k21, _r22);
-                _sum01 = __msa_fmadd_w(_sum01, _k22, _r23);
-                _sum10 = __msa_fmadd_w(_sum10, _k10, _r20);
-                _sum10 = __msa_fmadd_w(_sum10, _k11, _r21);
-                _sum10 = __msa_fmadd_w(_sum10, _k12, _r22);
-                _sum11 = __msa_fmadd_w(_sum11, _k10, _r21);
-                _sum11 = __msa_fmadd_w(_sum11, _k11, _r22);
-                _sum11 = __msa_fmadd_w(_sum11, _k12, _r23);
-
-                v4f32 _r30 = (v4f32)__msa_ld_w(r3, 0);
-                v4f32 _r31 = (v4f32)__msa_ld_w(r3 + 4, 0);
-                v4f32 _r32 = (v4f32)__msa_ld_w(r3 + 4 * 2, 0);
-                v4f32 _r33 = (v4f32)__msa_ld_w(r3 + 4 * 3, 0);
-
-                _sum10 = __msa_fmadd_w(_sum10, _k20, _r30);
-                _sum10 = __msa_fmadd_w(_sum10, _k21, _r31);
-                _sum10 = __msa_fmadd_w(_sum10, _k22, _r32);
-                _sum11 = __msa_fmadd_w(_sum11, _k20, _r31);
-                _sum11 = __msa_fmadd_w(_sum11, _k21, _r32);
-                _sum11 = __msa_fmadd_w(_sum11, _k22, _r33);
-
-                __msa_st_w((v4i32)_sum00, outptr0, 0);
-                __msa_st_w((v4i32)_sum01, outptr0 + 4, 0);
-                __msa_st_w((v4i32)_sum10, outptr1, 0);
-                __msa_st_w((v4i32)_sum11, outptr1 + 4, 0);
-
-                outptr0 += 4 * 2;
-                outptr1 += 4 * 2;
-
-                r0 += 4 * 2;
-                r1 += 4 * 2;
-                r2 += 4 * 2;
-                r3 += 4 * 2;
-            }
-            for (; j < outw; j++)
-            {
-                __builtin_prefetch(r0 + 16);
-                __builtin_prefetch(r1 + 16);
-                __builtin_prefetch(r2 + 16);
-                __builtin_prefetch(r3 + 16);
-
-                v4f32 _sum0 = _bias0;
-                v4f32 _sum1 = _bias0;
-
-                v4f32 _r00 = (v4f32)__msa_ld_w(r0, 0);
-                v4f32 _r01 = (v4f32)__msa_ld_w(r0 + 4, 0);
-                v4f32 _r02 = (v4f32)__msa_ld_w(r0 + 4 * 2, 0);
-
-                _sum0 = __msa_fmadd_w(_sum0, _k00, _r00);
-                _sum0 = __msa_fmadd_w(_sum0, _k01, _r01);
-                _sum0 = __msa_fmadd_w(_sum0, _k02, _r02);
-
-                v4f32 _r10 = (v4f32)__msa_ld_w(r1, 0);
-                v4f32 _r11 = (v4f32)__msa_ld_w(r1 + 4, 0);
-                v4f32 _r12 = (v4f32)__msa_ld_w(r1 + 4 * 2, 0);
-
-                _sum0 = __msa_fmadd_w(_sum0, _k10, _r10);
-                _sum0 = __msa_fmadd_w(_sum0, _k11, _r11);
-                _sum0 = __msa_fmadd_w(_sum0, _k12, _r12);
-                _sum1 = __msa_fmadd_w(_sum1, _k00, _r10);
-                _sum1 = __msa_fmadd_w(_sum1, _k01, _r11);
-                _sum1 = __msa_fmadd_w(_sum1, _k02, _r12);
-
-                v4f32 _r20 = (v4f32)__msa_ld_w(r2, 0);
-                v4f32 _r21 = (v4f32)__msa_ld_w(r2 + 4, 0);
-                v4f32 _r22 = (v4f32)__msa_ld_w(r2 + 4 * 2, 0);
-
-                _sum0 = __msa_fmadd_w(_sum0, _k20, _r20);
-                _sum0 = __msa_fmadd_w(_sum0, _k21, _r21);
-                _sum0 = __msa_fmadd_w(_sum0, _k22, _r22);
-                _sum1 = __msa_fmadd_w(_sum1, _k10, _r20);
-                _sum1 = __msa_fmadd_w(_sum1, _k11, _r21);
-                _sum1 = __msa_fmadd_w(_sum1, _k12, _r22);
-
-                v4f32 _r30 = (v4f32)__msa_ld_w(r3, 0);
-                v4f32 _r31 = (v4f32)__msa_ld_w(r3 + 4, 0);
-                v4f32 _r32 = (v4f32)__msa_ld_w(r3 + 4 * 2, 0);
-
-                _sum1 = __msa_fmadd_w(_sum1, _k20, _r30);
-                _sum1 = __msa_fmadd_w(_sum1, _k21, _r31);
-                _sum1 = __msa_fmadd_w(_sum1, _k22, _r32);
-
-                __msa_st_w((v4i32)_sum0, outptr0, 0);
-                __msa_st_w((v4i32)_sum1, outptr1, 0);
-
-                outptr0 += 4;
-                outptr1 += 4;
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                r3 += 4;
-            }
-
-            r0 += 2 * 4 + w * 4;
-            r1 += 2 * 4 + w * 4;
-            r2 += 2 * 4 + w * 4;
-            r3 += 2 * 4 + w * 4;
-
-            outptr0 += outw * 4;
-            outptr1 += outw * 4;
-        }
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                __builtin_prefetch(r0 + 32);
-                __builtin_prefetch(r1 + 32);
-                __builtin_prefetch(r2 + 32);
-
-                v4f32 _sum00 = _bias0;
-                v4f32 _sum01 = _bias0;
-
-                v4f32 _r00 = (v4f32)__msa_ld_w(r0, 0);
-                v4f32 _r01 = (v4f32)__msa_ld_w(r0 + 4, 0);
-                v4f32 _r02 = (v4f32)__msa_ld_w(r0 + 4 * 2, 0);
-                v4f32 _r03 = (v4f32)__msa_ld_w(r0 + 4 * 3, 0);
-
-                _sum00 = __msa_fmadd_w(_sum00, _k00, _r00);
-                _sum00 = __msa_fmadd_w(_sum00, _k01, _r01);
-                _sum00 = __msa_fmadd_w(_sum00, _k02, _r02);
-                _sum01 = __msa_fmadd_w(_sum01, _k00, _r01);
-                _sum01 = __msa_fmadd_w(_sum01, _k01, _r02);
-                _sum01 = __msa_fmadd_w(_sum01, _k02, _r03);
-
-                v4f32 _r10 = (v4f32)__msa_ld_w(r1, 0);
-                v4f32 _r11 = (v4f32)__msa_ld_w(r1 + 4, 0);
-                v4f32 _r12 = (v4f32)__msa_ld_w(r1 + 4 * 2, 0);
-                v4f32 _r13 = (v4f32)__msa_ld_w(r1 + 4 * 3, 0);
-
-                _sum00 = __msa_fmadd_w(_sum00, _k10, _r10);
-                _sum00 = __msa_fmadd_w(_sum00, _k11, _r11);
-                _sum00 = __msa_fmadd_w(_sum00, _k12, _r12);
-                _sum01 = __msa_fmadd_w(_sum01, _k10, _r11);
-                _sum01 = __msa_fmadd_w(_sum01, _k11, _r12);
-                _sum01 = __msa_fmadd_w(_sum01, _k12, _r13);
-
-                v4f32 _r20 = (v4f32)__msa_ld_w(r2, 0);
-                v4f32 _r21 = (v4f32)__msa_ld_w(r2 + 4, 0);
-                v4f32 _r22 = (v4f32)__msa_ld_w(r2 + 4 * 2, 0);
-                v4f32 _r23 = (v4f32)__msa_ld_w(r2 + 4 * 3, 0);
-
-                _sum00 = __msa_fmadd_w(_sum00, _k20, _r20);
-                _sum00 = __msa_fmadd_w(_sum00, _k21, _r21);
-                _sum00 = __msa_fmadd_w(_sum00, _k22, _r22);
-                _sum01 = __msa_fmadd_w(_sum01, _k20, _r21);
-                _sum01 = __msa_fmadd_w(_sum01, _k21, _r22);
-                _sum01 = __msa_fmadd_w(_sum01, _k22, _r23);
-
-                __msa_st_w((v4i32)_sum00, outptr0, 0);
-                __msa_st_w((v4i32)_sum01, outptr0 + 4, 0);
-
-                outptr0 += 4 * 2;
-
-                r0 += 4 * 2;
-                r1 += 4 * 2;
-                r2 += 4 * 2;
-            }
-            for (; j < outw; j++)
-            {
-                __builtin_prefetch(r0 + 16);
-                __builtin_prefetch(r1 + 16);
-                __builtin_prefetch(r2 + 16);
-
-                v4f32 _sum0 = _bias0;
-
-                v4f32 _r00 = (v4f32)__msa_ld_w(r0, 0);
-                v4f32 _r01 = (v4f32)__msa_ld_w(r0 + 4, 0);
-                v4f32 _r02 = (v4f32)__msa_ld_w(r0 + 4 * 2, 0);
-
-                _sum0 = __msa_fmadd_w(_sum0, _k00, _r00);
-                _sum0 = __msa_fmadd_w(_sum0, _k01, _r01);
-                _sum0 = __msa_fmadd_w(_sum0, _k02, _r02);
-
-                v4f32 _r10 = (v4f32)__msa_ld_w(r1, 0);
-                v4f32 _r11 = (v4f32)__msa_ld_w(r1 + 4, 0);
-                v4f32 _r12 = (v4f32)__msa_ld_w(r1 + 4 * 2, 0);
-
-                _sum0 = __msa_fmadd_w(_sum0, _k10, _r10);
-                _sum0 = __msa_fmadd_w(_sum0, _k11, _r11);
-                _sum0 = __msa_fmadd_w(_sum0, _k12, _r12);
-
-                v4f32 _r20 = (v4f32)__msa_ld_w(r2, 0);
-                v4f32 _r21 = (v4f32)__msa_ld_w(r2 + 4, 0);
-                v4f32 _r22 = (v4f32)__msa_ld_w(r2 + 4 * 2, 0);
-
-                _sum0 = __msa_fmadd_w(_sum0, _k20, _r20);
-                _sum0 = __msa_fmadd_w(_sum0, _k21, _r21);
-                _sum0 = __msa_fmadd_w(_sum0, _k22, _r22);
-
-                __msa_st_w((v4i32)_sum0, outptr0, 0);
-
-                outptr0 += 4;
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-            }
-
-            r0 += 2 * 4;
-            r1 += 2 * 4;
-            r2 += 2 * 4;
-        }
-    }
-}
-
-static void convdw3x3s2_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        v4f32 _bias0 = bias ? (v4f32)__msa_ld_w(bias + g * 4, 0) : (v4f32)__msa_fill_w(0);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out;
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-
-        v4f32 _k00 = (v4f32)__msa_ld_w(k0, 0);
-        v4f32 _k01 = (v4f32)__msa_ld_w(k0 + 4, 0);
-        v4f32 _k02 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-        v4f32 _k10 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-        v4f32 _k11 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-        v4f32 _k12 = (v4f32)__msa_ld_w(k0 + 4 * 5, 0);
-        v4f32 _k20 = (v4f32)__msa_ld_w(k0 + 4 * 6, 0);
-        v4f32 _k21 = (v4f32)__msa_ld_w(k0 + 4 * 7, 0);
-        v4f32 _k22 = (v4f32)__msa_ld_w(k0 + 4 * 8, 0);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                __builtin_prefetch(r0 + 64);
-                __builtin_prefetch(r1 + 64);
-                __builtin_prefetch(r2 + 64);
-
-                v4f32 _sum00 = _bias0;
-                v4f32 _sum01 = _bias0;
-
-                v4f32 _r00 = (v4f32)__msa_ld_w(r0, 0);
-                v4f32 _r01 = (v4f32)__msa_ld_w(r0 + 4, 0);
-                v4f32 _r02 = (v4f32)__msa_ld_w(r0 + 4 * 2, 0);
-                v4f32 _r03 = (v4f32)__msa_ld_w(r0 + 4 * 3, 0);
-                v4f32 _r04 = (v4f32)__msa_ld_w(r0 + 4 * 4, 0);
-
-                _sum00 = __msa_fmadd_w(_sum00, _k00, _r00);
-                _sum00 = __msa_fmadd_w(_sum00, _k01, _r01);
-                _sum00 = __msa_fmadd_w(_sum00, _k02, _r02);
-                _sum01 = __msa_fmadd_w(_sum01, _k00, _r02);
-                _sum01 = __msa_fmadd_w(_sum01, _k01, _r03);
-                _sum01 = __msa_fmadd_w(_sum01, _k02, _r04);
-
-                v4f32 _r10 = (v4f32)__msa_ld_w(r1, 0);
-                v4f32 _r11 = (v4f32)__msa_ld_w(r1 + 4, 0);
-                v4f32 _r12 = (v4f32)__msa_ld_w(r1 + 4 * 2, 0);
-                v4f32 _r13 = (v4f32)__msa_ld_w(r1 + 4 * 3, 0);
-                v4f32 _r14 = (v4f32)__msa_ld_w(r1 + 4 * 4, 0);
-
-                _sum00 = __msa_fmadd_w(_sum00, _k10, _r10);
-                _sum00 = __msa_fmadd_w(_sum00, _k11, _r11);
-                _sum00 = __msa_fmadd_w(_sum00, _k12, _r12);
-                _sum01 = __msa_fmadd_w(_sum01, _k10, _r12);
-                _sum01 = __msa_fmadd_w(_sum01, _k11, _r13);
-                _sum01 = __msa_fmadd_w(_sum01, _k12, _r14);
-
-                v4f32 _r20 = (v4f32)__msa_ld_w(r2, 0);
-                v4f32 _r21 = (v4f32)__msa_ld_w(r2 + 4, 0);
-                v4f32 _r22 = (v4f32)__msa_ld_w(r2 + 4 * 2, 0);
-                v4f32 _r23 = (v4f32)__msa_ld_w(r2 + 4 * 3, 0);
-                v4f32 _r24 = (v4f32)__msa_ld_w(r2 + 4 * 4, 0);
-
-                _sum00 = __msa_fmadd_w(_sum00, _k20, _r20);
-                _sum00 = __msa_fmadd_w(_sum00, _k21, _r21);
-                _sum00 = __msa_fmadd_w(_sum00, _k22, _r22);
-                _sum01 = __msa_fmadd_w(_sum01, _k20, _r22);
-                _sum01 = __msa_fmadd_w(_sum01, _k21, _r23);
-                _sum01 = __msa_fmadd_w(_sum01, _k22, _r24);
-
-                __msa_st_w((v4i32)_sum00, outptr0, 0);
-                __msa_st_w((v4i32)_sum01, outptr0 + 4, 0);
-
-                outptr0 += 4 * 2;
-
-                r0 += 4 * 4;
-                r1 += 4 * 4;
-                r2 += 4 * 4;
-            }
-            for (; j < outw; j++)
-            {
-                __builtin_prefetch(r0 + 32);
-                __builtin_prefetch(r1 + 32);
-                __builtin_prefetch(r2 + 32);
-
-                v4f32 _sum0 = _bias0;
-
-                v4f32 _r00 = (v4f32)__msa_ld_w(r0, 0);
-                v4f32 _r01 = (v4f32)__msa_ld_w(r0 + 4, 0);
-                v4f32 _r02 = (v4f32)__msa_ld_w(r0 + 4 * 2, 0);
-
-                _sum0 = __msa_fmadd_w(_sum0, _k00, _r00);
-                _sum0 = __msa_fmadd_w(_sum0, _k01, _r01);
-                _sum0 = __msa_fmadd_w(_sum0, _k02, _r02);
-
-                v4f32 _r10 = (v4f32)__msa_ld_w(r1, 0);
-                v4f32 _r11 = (v4f32)__msa_ld_w(r1 + 4, 0);
-                v4f32 _r12 = (v4f32)__msa_ld_w(r1 + 4 * 2, 0);
-
-                _sum0 = __msa_fmadd_w(_sum0, _k10, _r10);
-                _sum0 = __msa_fmadd_w(_sum0, _k11, _r11);
-                _sum0 = __msa_fmadd_w(_sum0, _k12, _r12);
-
-                v4f32 _r20 = (v4f32)__msa_ld_w(r2, 0);
-                v4f32 _r21 = (v4f32)__msa_ld_w(r2 + 4, 0);
-                v4f32 _r22 = (v4f32)__msa_ld_w(r2 + 4 * 2, 0);
-
-                _sum0 = __msa_fmadd_w(_sum0, _k20, _r20);
-                _sum0 = __msa_fmadd_w(_sum0, _k21, _r21);
-                _sum0 = __msa_fmadd_w(_sum0, _k22, _r22);
-
-                __msa_st_w((v4i32)_sum0, outptr0, 0);
-
-                outptr0 += 4;
-
-                r0 += 4 * 2;
-                r1 += 4 * 2;
-                r2 += 4 * 2;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_5x5_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_5x5_pack4.h
deleted file mode 100644
index 3b524424..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_5x5_pack4.h
+++ /dev/null
@@ -1,511 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw5x5s1_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        v4f32 _bias0 = bias ? (v4f32)__msa_ld_w(bias + g * 4, 0) : (v4f32)__msa_fill_w(0);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-        float* outptr1 = out.row(1);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-        const float* r4 = img0.row(4);
-        const float* r5 = img0.row(5);
-
-        int i = 0;
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                __builtin_prefetch(r0 + 16);
-                __builtin_prefetch(r1 + 16);
-                __builtin_prefetch(r2 + 16);
-                __builtin_prefetch(r3 + 16);
-                __builtin_prefetch(r4 + 16);
-                __builtin_prefetch(r5 + 16);
-
-                __builtin_prefetch(k0 + 400);
-
-                v4f32 _sum0 = _bias0;
-                v4f32 _sum1 = _bias0;
-
-                v4f32 _r00 = (v4f32)__msa_ld_w(r0, 0);
-                v4f32 _r01 = (v4f32)__msa_ld_w(r0 + 4, 0);
-                v4f32 _r02 = (v4f32)__msa_ld_w(r0 + 4 * 2, 0);
-                v4f32 _r03 = (v4f32)__msa_ld_w(r0 + 4 * 3, 0);
-                v4f32 _r04 = (v4f32)__msa_ld_w(r0 + 4 * 4, 0);
-
-                v4f32 _k00 = (v4f32)__msa_ld_w(k0, 0);
-                v4f32 _k01 = (v4f32)__msa_ld_w(k0 + 4, 0);
-                v4f32 _k02 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-                v4f32 _k03 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-                v4f32 _k04 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __msa_fmadd_w(_sum0, _k00, _r00);
-                _sum0 = __msa_fmadd_w(_sum0, _k01, _r01);
-                _sum0 = __msa_fmadd_w(_sum0, _k02, _r02);
-                _sum0 = __msa_fmadd_w(_sum0, _k03, _r03);
-                _sum0 = __msa_fmadd_w(_sum0, _k04, _r04);
-
-                v4f32 _r10 = (v4f32)__msa_ld_w(r1, 0);
-                v4f32 _r11 = (v4f32)__msa_ld_w(r1 + 4, 0);
-                v4f32 _r12 = (v4f32)__msa_ld_w(r1 + 4 * 2, 0);
-                v4f32 _r13 = (v4f32)__msa_ld_w(r1 + 4 * 3, 0);
-                v4f32 _r14 = (v4f32)__msa_ld_w(r1 + 4 * 4, 0);
-
-                _sum1 = __msa_fmadd_w(_sum1, _k00, _r10);
-                _sum1 = __msa_fmadd_w(_sum1, _k01, _r11);
-                _sum1 = __msa_fmadd_w(_sum1, _k02, _r12);
-                _sum1 = __msa_fmadd_w(_sum1, _k03, _r13);
-                _sum1 = __msa_fmadd_w(_sum1, _k04, _r14);
-
-                v4f32 _k10 = (v4f32)__msa_ld_w(k0, 0);
-                v4f32 _k11 = (v4f32)__msa_ld_w(k0 + 4, 0);
-                v4f32 _k12 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-                v4f32 _k13 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-                v4f32 _k14 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __msa_fmadd_w(_sum0, _k10, _r10);
-                _sum0 = __msa_fmadd_w(_sum0, _k11, _r11);
-                _sum0 = __msa_fmadd_w(_sum0, _k12, _r12);
-                _sum0 = __msa_fmadd_w(_sum0, _k13, _r13);
-                _sum0 = __msa_fmadd_w(_sum0, _k14, _r14);
-
-                v4f32 _r20 = (v4f32)__msa_ld_w(r2, 0);
-                v4f32 _r21 = (v4f32)__msa_ld_w(r2 + 4, 0);
-                v4f32 _r22 = (v4f32)__msa_ld_w(r2 + 4 * 2, 0);
-                v4f32 _r23 = (v4f32)__msa_ld_w(r2 + 4 * 3, 0);
-                v4f32 _r24 = (v4f32)__msa_ld_w(r2 + 4 * 4, 0);
-
-                _sum1 = __msa_fmadd_w(_sum1, _k10, _r20);
-                _sum1 = __msa_fmadd_w(_sum1, _k11, _r21);
-                _sum1 = __msa_fmadd_w(_sum1, _k12, _r22);
-                _sum1 = __msa_fmadd_w(_sum1, _k13, _r23);
-                _sum1 = __msa_fmadd_w(_sum1, _k14, _r24);
-
-                v4f32 _k20 = (v4f32)__msa_ld_w(k0, 0);
-                v4f32 _k21 = (v4f32)__msa_ld_w(k0 + 4, 0);
-                v4f32 _k22 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-                v4f32 _k23 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-                v4f32 _k24 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __msa_fmadd_w(_sum0, _k20, _r20);
-                _sum0 = __msa_fmadd_w(_sum0, _k21, _r21);
-                _sum0 = __msa_fmadd_w(_sum0, _k22, _r22);
-                _sum0 = __msa_fmadd_w(_sum0, _k23, _r23);
-                _sum0 = __msa_fmadd_w(_sum0, _k24, _r24);
-
-                v4f32 _r30 = (v4f32)__msa_ld_w(r3, 0);
-                v4f32 _r31 = (v4f32)__msa_ld_w(r3 + 4, 0);
-                v4f32 _r32 = (v4f32)__msa_ld_w(r3 + 4 * 2, 0);
-                v4f32 _r33 = (v4f32)__msa_ld_w(r3 + 4 * 3, 0);
-                v4f32 _r34 = (v4f32)__msa_ld_w(r3 + 4 * 4, 0);
-
-                _sum1 = __msa_fmadd_w(_sum1, _k20, _r30);
-                _sum1 = __msa_fmadd_w(_sum1, _k21, _r31);
-                _sum1 = __msa_fmadd_w(_sum1, _k22, _r32);
-                _sum1 = __msa_fmadd_w(_sum1, _k23, _r33);
-                _sum1 = __msa_fmadd_w(_sum1, _k24, _r34);
-
-                v4f32 _k30 = (v4f32)__msa_ld_w(k0, 0);
-                v4f32 _k31 = (v4f32)__msa_ld_w(k0 + 4, 0);
-                v4f32 _k32 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-                v4f32 _k33 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-                v4f32 _k34 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __msa_fmadd_w(_sum0, _k30, _r30);
-                _sum0 = __msa_fmadd_w(_sum0, _k31, _r31);
-                _sum0 = __msa_fmadd_w(_sum0, _k32, _r32);
-                _sum0 = __msa_fmadd_w(_sum0, _k33, _r33);
-                _sum0 = __msa_fmadd_w(_sum0, _k34, _r34);
-
-                v4f32 _r40 = (v4f32)__msa_ld_w(r4, 0);
-                v4f32 _r41 = (v4f32)__msa_ld_w(r4 + 4, 0);
-                v4f32 _r42 = (v4f32)__msa_ld_w(r4 + 4 * 2, 0);
-                v4f32 _r43 = (v4f32)__msa_ld_w(r4 + 4 * 3, 0);
-                v4f32 _r44 = (v4f32)__msa_ld_w(r4 + 4 * 4, 0);
-
-                _sum1 = __msa_fmadd_w(_sum1, _k30, _r40);
-                _sum1 = __msa_fmadd_w(_sum1, _k31, _r41);
-                _sum1 = __msa_fmadd_w(_sum1, _k32, _r42);
-                _sum1 = __msa_fmadd_w(_sum1, _k33, _r43);
-                _sum1 = __msa_fmadd_w(_sum1, _k34, _r44);
-
-                v4f32 _k40 = (v4f32)__msa_ld_w(k0, 0);
-                v4f32 _k41 = (v4f32)__msa_ld_w(k0 + 4, 0);
-                v4f32 _k42 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-                v4f32 _k43 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-                v4f32 _k44 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-                k0 -= 4 * 20;
-
-                _sum0 = __msa_fmadd_w(_sum0, _k40, _r40);
-                _sum0 = __msa_fmadd_w(_sum0, _k41, _r41);
-                _sum0 = __msa_fmadd_w(_sum0, _k42, _r42);
-                _sum0 = __msa_fmadd_w(_sum0, _k43, _r43);
-                _sum0 = __msa_fmadd_w(_sum0, _k44, _r44);
-
-                v4f32 _r50 = (v4f32)__msa_ld_w(r5, 0);
-                v4f32 _r51 = (v4f32)__msa_ld_w(r5 + 4, 0);
-                v4f32 _r52 = (v4f32)__msa_ld_w(r5 + 4 * 2, 0);
-                v4f32 _r53 = (v4f32)__msa_ld_w(r5 + 4 * 3, 0);
-                v4f32 _r54 = (v4f32)__msa_ld_w(r5 + 4 * 4, 0);
-
-                _sum1 = __msa_fmadd_w(_sum1, _k40, _r50);
-                _sum1 = __msa_fmadd_w(_sum1, _k41, _r51);
-                _sum1 = __msa_fmadd_w(_sum1, _k42, _r52);
-                _sum1 = __msa_fmadd_w(_sum1, _k43, _r53);
-                _sum1 = __msa_fmadd_w(_sum1, _k44, _r54);
-
-                __msa_st_w((v4i32)_sum0, outptr0, 0);
-                __msa_st_w((v4i32)_sum1, outptr1, 0);
-
-                outptr0 += 4;
-                outptr1 += 4;
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                r3 += 4;
-                r4 += 4;
-                r5 += 4;
-            }
-
-            r0 += 4 * 4 + w * 4;
-            r1 += 4 * 4 + w * 4;
-            r2 += 4 * 4 + w * 4;
-            r3 += 4 * 4 + w * 4;
-            r4 += 4 * 4 + w * 4;
-            r5 += 4 * 4 + w * 4;
-
-            outptr0 += outw * 4;
-            outptr1 += outw * 4;
-        }
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                __builtin_prefetch(r0 + 16);
-                __builtin_prefetch(r1 + 16);
-                __builtin_prefetch(r2 + 16);
-                __builtin_prefetch(r3 + 16);
-                __builtin_prefetch(r4 + 16);
-
-                __builtin_prefetch(k0 + 400);
-
-                v4f32 _sum0 = _bias0;
-
-                v4f32 _r00 = (v4f32)__msa_ld_w(r0, 0);
-                v4f32 _r01 = (v4f32)__msa_ld_w(r0 + 4, 0);
-                v4f32 _r02 = (v4f32)__msa_ld_w(r0 + 4 * 2, 0);
-                v4f32 _r03 = (v4f32)__msa_ld_w(r0 + 4 * 3, 0);
-                v4f32 _r04 = (v4f32)__msa_ld_w(r0 + 4 * 4, 0);
-
-                v4f32 _k00 = (v4f32)__msa_ld_w(k0, 0);
-                v4f32 _k01 = (v4f32)__msa_ld_w(k0 + 4, 0);
-                v4f32 _k02 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-                v4f32 _k03 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-                v4f32 _k04 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __msa_fmadd_w(_sum0, _k00, _r00);
-                _sum0 = __msa_fmadd_w(_sum0, _k01, _r01);
-                _sum0 = __msa_fmadd_w(_sum0, _k02, _r02);
-                _sum0 = __msa_fmadd_w(_sum0, _k03, _r03);
-                _sum0 = __msa_fmadd_w(_sum0, _k04, _r04);
-
-                v4f32 _r10 = (v4f32)__msa_ld_w(r1, 0);
-                v4f32 _r11 = (v4f32)__msa_ld_w(r1 + 4, 0);
-                v4f32 _r12 = (v4f32)__msa_ld_w(r1 + 4 * 2, 0);
-                v4f32 _r13 = (v4f32)__msa_ld_w(r1 + 4 * 3, 0);
-                v4f32 _r14 = (v4f32)__msa_ld_w(r1 + 4 * 4, 0);
-
-                v4f32 _k10 = (v4f32)__msa_ld_w(k0, 0);
-                v4f32 _k11 = (v4f32)__msa_ld_w(k0 + 4, 0);
-                v4f32 _k12 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-                v4f32 _k13 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-                v4f32 _k14 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __msa_fmadd_w(_sum0, _k10, _r10);
-                _sum0 = __msa_fmadd_w(_sum0, _k11, _r11);
-                _sum0 = __msa_fmadd_w(_sum0, _k12, _r12);
-                _sum0 = __msa_fmadd_w(_sum0, _k13, _r13);
-                _sum0 = __msa_fmadd_w(_sum0, _k14, _r14);
-
-                v4f32 _r20 = (v4f32)__msa_ld_w(r2, 0);
-                v4f32 _r21 = (v4f32)__msa_ld_w(r2 + 4, 0);
-                v4f32 _r22 = (v4f32)__msa_ld_w(r2 + 4 * 2, 0);
-                v4f32 _r23 = (v4f32)__msa_ld_w(r2 + 4 * 3, 0);
-                v4f32 _r24 = (v4f32)__msa_ld_w(r2 + 4 * 4, 0);
-
-                v4f32 _k20 = (v4f32)__msa_ld_w(k0, 0);
-                v4f32 _k21 = (v4f32)__msa_ld_w(k0 + 4, 0);
-                v4f32 _k22 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-                v4f32 _k23 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-                v4f32 _k24 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __msa_fmadd_w(_sum0, _k20, _r20);
-                _sum0 = __msa_fmadd_w(_sum0, _k21, _r21);
-                _sum0 = __msa_fmadd_w(_sum0, _k22, _r22);
-                _sum0 = __msa_fmadd_w(_sum0, _k23, _r23);
-                _sum0 = __msa_fmadd_w(_sum0, _k24, _r24);
-
-                v4f32 _r30 = (v4f32)__msa_ld_w(r3, 0);
-                v4f32 _r31 = (v4f32)__msa_ld_w(r3 + 4, 0);
-                v4f32 _r32 = (v4f32)__msa_ld_w(r3 + 4 * 2, 0);
-                v4f32 _r33 = (v4f32)__msa_ld_w(r3 + 4 * 3, 0);
-                v4f32 _r34 = (v4f32)__msa_ld_w(r3 + 4 * 4, 0);
-
-                v4f32 _k30 = (v4f32)__msa_ld_w(k0, 0);
-                v4f32 _k31 = (v4f32)__msa_ld_w(k0 + 4, 0);
-                v4f32 _k32 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-                v4f32 _k33 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-                v4f32 _k34 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __msa_fmadd_w(_sum0, _k30, _r30);
-                _sum0 = __msa_fmadd_w(_sum0, _k31, _r31);
-                _sum0 = __msa_fmadd_w(_sum0, _k32, _r32);
-                _sum0 = __msa_fmadd_w(_sum0, _k33, _r33);
-                _sum0 = __msa_fmadd_w(_sum0, _k34, _r34);
-
-                v4f32 _r40 = (v4f32)__msa_ld_w(r4, 0);
-                v4f32 _r41 = (v4f32)__msa_ld_w(r4 + 4, 0);
-                v4f32 _r42 = (v4f32)__msa_ld_w(r4 + 4 * 2, 0);
-                v4f32 _r43 = (v4f32)__msa_ld_w(r4 + 4 * 3, 0);
-                v4f32 _r44 = (v4f32)__msa_ld_w(r4 + 4 * 4, 0);
-
-                v4f32 _k40 = (v4f32)__msa_ld_w(k0, 0);
-                v4f32 _k41 = (v4f32)__msa_ld_w(k0 + 4, 0);
-                v4f32 _k42 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-                v4f32 _k43 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-                v4f32 _k44 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-                k0 -= 4 * 20;
-
-                _sum0 = __msa_fmadd_w(_sum0, _k40, _r40);
-                _sum0 = __msa_fmadd_w(_sum0, _k41, _r41);
-                _sum0 = __msa_fmadd_w(_sum0, _k42, _r42);
-                _sum0 = __msa_fmadd_w(_sum0, _k43, _r43);
-                _sum0 = __msa_fmadd_w(_sum0, _k44, _r44);
-
-                __msa_st_w((v4i32)_sum0, outptr0, 0);
-
-                outptr0 += 4;
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                r3 += 4;
-                r4 += 4;
-            }
-
-            r0 += 4 * 4;
-            r1 += 4 * 4;
-            r2 += 4 * 4;
-            r3 += 4 * 4;
-            r4 += 4 * 4;
-        }
-    }
-}
-
-static void convdw5x5s2_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        v4f32 _bias0 = bias ? (v4f32)__msa_ld_w(bias + g * 4, 0) : (v4f32)__msa_fill_w(0);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out;
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-        const float* r4 = img0.row(4);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                __builtin_prefetch(r0 + 32);
-                __builtin_prefetch(r1 + 32);
-                __builtin_prefetch(r2 + 32);
-                __builtin_prefetch(r3 + 32);
-                __builtin_prefetch(r4 + 32);
-
-                __builtin_prefetch(k0 + 400);
-
-                v4f32 _sum0 = _bias0;
-
-                v4f32 _r00 = (v4f32)__msa_ld_w(r0, 0);
-                v4f32 _r01 = (v4f32)__msa_ld_w(r0 + 4, 0);
-                v4f32 _r02 = (v4f32)__msa_ld_w(r0 + 4 * 2, 0);
-                v4f32 _r03 = (v4f32)__msa_ld_w(r0 + 4 * 3, 0);
-                v4f32 _r04 = (v4f32)__msa_ld_w(r0 + 4 * 4, 0);
-
-                v4f32 _k00 = (v4f32)__msa_ld_w(k0, 0);
-                v4f32 _k01 = (v4f32)__msa_ld_w(k0 + 4, 0);
-                v4f32 _k02 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-                v4f32 _k03 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-                v4f32 _k04 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __msa_fmadd_w(_sum0, _k00, _r00);
-                _sum0 = __msa_fmadd_w(_sum0, _k01, _r01);
-                _sum0 = __msa_fmadd_w(_sum0, _k02, _r02);
-                _sum0 = __msa_fmadd_w(_sum0, _k03, _r03);
-                _sum0 = __msa_fmadd_w(_sum0, _k04, _r04);
-
-                v4f32 _r10 = (v4f32)__msa_ld_w(r1, 0);
-                v4f32 _r11 = (v4f32)__msa_ld_w(r1 + 4, 0);
-                v4f32 _r12 = (v4f32)__msa_ld_w(r1 + 4 * 2, 0);
-                v4f32 _r13 = (v4f32)__msa_ld_w(r1 + 4 * 3, 0);
-                v4f32 _r14 = (v4f32)__msa_ld_w(r1 + 4 * 4, 0);
-
-                v4f32 _k10 = (v4f32)__msa_ld_w(k0, 0);
-                v4f32 _k11 = (v4f32)__msa_ld_w(k0 + 4, 0);
-                v4f32 _k12 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-                v4f32 _k13 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-                v4f32 _k14 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __msa_fmadd_w(_sum0, _k10, _r10);
-                _sum0 = __msa_fmadd_w(_sum0, _k11, _r11);
-                _sum0 = __msa_fmadd_w(_sum0, _k12, _r12);
-                _sum0 = __msa_fmadd_w(_sum0, _k13, _r13);
-                _sum0 = __msa_fmadd_w(_sum0, _k14, _r14);
-
-                v4f32 _r20 = (v4f32)__msa_ld_w(r2, 0);
-                v4f32 _r21 = (v4f32)__msa_ld_w(r2 + 4, 0);
-                v4f32 _r22 = (v4f32)__msa_ld_w(r2 + 4 * 2, 0);
-                v4f32 _r23 = (v4f32)__msa_ld_w(r2 + 4 * 3, 0);
-                v4f32 _r24 = (v4f32)__msa_ld_w(r2 + 4 * 4, 0);
-
-                v4f32 _k20 = (v4f32)__msa_ld_w(k0, 0);
-                v4f32 _k21 = (v4f32)__msa_ld_w(k0 + 4, 0);
-                v4f32 _k22 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-                v4f32 _k23 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-                v4f32 _k24 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __msa_fmadd_w(_sum0, _k20, _r20);
-                _sum0 = __msa_fmadd_w(_sum0, _k21, _r21);
-                _sum0 = __msa_fmadd_w(_sum0, _k22, _r22);
-                _sum0 = __msa_fmadd_w(_sum0, _k23, _r23);
-                _sum0 = __msa_fmadd_w(_sum0, _k24, _r24);
-
-                v4f32 _r30 = (v4f32)__msa_ld_w(r3, 0);
-                v4f32 _r31 = (v4f32)__msa_ld_w(r3 + 4, 0);
-                v4f32 _r32 = (v4f32)__msa_ld_w(r3 + 4 * 2, 0);
-                v4f32 _r33 = (v4f32)__msa_ld_w(r3 + 4 * 3, 0);
-                v4f32 _r34 = (v4f32)__msa_ld_w(r3 + 4 * 4, 0);
-
-                v4f32 _k30 = (v4f32)__msa_ld_w(k0, 0);
-                v4f32 _k31 = (v4f32)__msa_ld_w(k0 + 4, 0);
-                v4f32 _k32 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-                v4f32 _k33 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-                v4f32 _k34 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-                k0 += 4 * 5;
-
-                _sum0 = __msa_fmadd_w(_sum0, _k30, _r30);
-                _sum0 = __msa_fmadd_w(_sum0, _k31, _r31);
-                _sum0 = __msa_fmadd_w(_sum0, _k32, _r32);
-                _sum0 = __msa_fmadd_w(_sum0, _k33, _r33);
-                _sum0 = __msa_fmadd_w(_sum0, _k34, _r34);
-
-                v4f32 _r40 = (v4f32)__msa_ld_w(r4, 0);
-                v4f32 _r41 = (v4f32)__msa_ld_w(r4 + 4, 0);
-                v4f32 _r42 = (v4f32)__msa_ld_w(r4 + 4 * 2, 0);
-                v4f32 _r43 = (v4f32)__msa_ld_w(r4 + 4 * 3, 0);
-                v4f32 _r44 = (v4f32)__msa_ld_w(r4 + 4 * 4, 0);
-
-                v4f32 _k40 = (v4f32)__msa_ld_w(k0, 0);
-                v4f32 _k41 = (v4f32)__msa_ld_w(k0 + 4, 0);
-                v4f32 _k42 = (v4f32)__msa_ld_w(k0 + 4 * 2, 0);
-                v4f32 _k43 = (v4f32)__msa_ld_w(k0 + 4 * 3, 0);
-                v4f32 _k44 = (v4f32)__msa_ld_w(k0 + 4 * 4, 0);
-                k0 -= 4 * 20;
-
-                _sum0 = __msa_fmadd_w(_sum0, _k40, _r40);
-                _sum0 = __msa_fmadd_w(_sum0, _k41, _r41);
-                _sum0 = __msa_fmadd_w(_sum0, _k42, _r42);
-                _sum0 = __msa_fmadd_w(_sum0, _k43, _r43);
-                _sum0 = __msa_fmadd_w(_sum0, _k44, _r44);
-
-                __msa_st_w((v4i32)_sum0, outptr0, 0);
-
-                outptr0 += 4;
-
-                r0 += 4 * 2;
-                r1 += 4 * 2;
-                r2 += 4 * 2;
-                r3 += 4 * 2;
-                r4 += 4 * 2;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-            r3 += tailstep;
-            r4 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_mips.h
deleted file mode 100644
index 24d1650b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/convolutiondepthwise_mips.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTIONDEPTHWISE_MIPS_H
-#define LAYER_CONVOLUTIONDEPTHWISE_MIPS_H
-
-#include "convolutiondepthwise.h"
-
-namespace ncnn {
-
-class ConvolutionDepthWise_mips : public ConvolutionDepthWise
-{
-public:
-    ConvolutionDepthWise_mips();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    int create_group_ops(const Option& opt);
-#if NCNN_INT8
-    int create_pipeline_int8_mips(const Option& opt);
-    int forward_int8_mips(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Layer* activation;
-    std::vector<ncnn::Layer*> group_ops;
-
-    Mat weight_data_tm;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTIONDEPTHWISE_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/crop_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/crop_mips.h
deleted file mode 100644
index 77c077e7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/crop_mips.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CROP_MIPS_H
-#define LAYER_CROP_MIPS_H
-
-#include "crop.h"
-
-namespace ncnn {
-
-class Crop_mips : public Crop
-{
-public:
-    Crop_mips();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CROP_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_mips.h
deleted file mode 100644
index b7c0d2e7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_mips.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTION_MIPS_H
-#define LAYER_DECONVOLUTION_MIPS_H
-
-#include "deconvolution.h"
-
-namespace ncnn {
-
-class Deconvolution_mips : public Deconvolution
-{
-public:
-    Deconvolution_mips();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    Mat weight_data_tm;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTION_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_pack1to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_pack1to4.h
deleted file mode 100644
index b9ea4ef2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_pack1to4.h
+++ /dev/null
@@ -1,99 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack1to4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_pack1ton, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                v4f32 _sum = (v4f32)__msa_fill_w(0);
-
-                if (bias_data_ptr)
-                {
-                    _sum = (v4f32)__msa_ld_w((const float*)bias_data_ptr + p * 4, 0);
-                }
-
-                const float* kptr = (const float*)weight_data_pack1ton + maxk * channels * p * 4;
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        const float* sptr = m.row(sy);
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            float val = sptr[sx];
-
-                            int k = y * kernel_w + x;
-
-                            v4f32 _val = (v4f32)__msa_fill_w_f32(val);
-                            v4f32 _w = (v4f32)__msa_ld_w(kptr + k * 4, 0);
-                            _sum = __msa_fmadd_w(_sum, _val, _w);
-                        }
-                    }
-
-                    kptr += maxk * 4;
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params);
-
-                __msa_st_w((v4i32)_sum, outptr + j * 4, 0);
-            }
-
-            outptr += outw * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_pack4.h
deleted file mode 100644
index 5f07cc19..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_pack4.h
+++ /dev/null
@@ -1,106 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_pack4, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                v4f32 _sum = (v4f32)__msa_fill_w(0);
-
-                if (bias_data_ptr)
-                {
-                    _sum = (v4f32)__msa_ld_w((const float*)bias_data_ptr + p * 4, 0);
-                }
-
-                const float* kptr = (const float*)weight_data_pack4.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 4;
-
-                            int k = (y * kernel_w + x) * 16;
-
-                            v4f32 _val0 = (v4f32)__msa_fill_w_f32(*sptr++);
-                            v4f32 _val1 = (v4f32)__msa_fill_w_f32(*sptr++);
-                            v4f32 _val2 = (v4f32)__msa_fill_w_f32(*sptr++);
-                            v4f32 _val3 = (v4f32)__msa_fill_w_f32(*sptr++);
-                            v4f32 _w0 = (v4f32)__msa_ld_w(kptr + k, 0);
-                            v4f32 _w1 = (v4f32)__msa_ld_w(kptr + k + 4, 0);
-                            v4f32 _w2 = (v4f32)__msa_ld_w(kptr + k + 8, 0);
-                            v4f32 _w3 = (v4f32)__msa_ld_w(kptr + k + 12, 0);
-                            _sum = __msa_fmadd_w(_sum, _val0, _w0);
-                            _sum = __msa_fmadd_w(_sum, _val1, _w1);
-                            _sum = __msa_fmadd_w(_sum, _val2, _w2);
-                            _sum = __msa_fmadd_w(_sum, _val3, _w3);
-                        }
-                    }
-
-                    kptr += maxk * 16;
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params);
-
-                __msa_st_w((v4i32)_sum, outptr + j * 4, 0);
-            }
-
-            outptr += outw * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_pack4to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_pack4to1.h
deleted file mode 100644
index b825e146..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolution_pack4to1.h
+++ /dev/null
@@ -1,101 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack4to1_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_pack4to1, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                v4f32 _sum = (v4f32)__msa_fill_w(0);
-
-                const float* kptr = (const float*)weight_data_pack4to1 + maxk * channels * p * 4;
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 4;
-
-                            int k = y * kernel_w + x;
-
-                            v4f32 _val = (v4f32)__msa_ld_w(sptr, 0);
-                            v4f32 _w = (v4f32)__msa_ld_w(kptr + k * 4, 0);
-                            _sum = __msa_fmadd_w(_sum, _val, _w);
-                        }
-                    }
-
-                    kptr += maxk * 4;
-                }
-
-                sum += __msa_reduce_fadd_w(_sum);
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[j] = sum;
-            }
-
-            outptr += outw;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolutiondepthwise_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolutiondepthwise_mips.h
deleted file mode 100644
index 24e7a481..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/deconvolutiondepthwise_mips.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTIONDEPTHWISE_MIPS_H
-#define LAYER_DECONVOLUTIONDEPTHWISE_MIPS_H
-
-#include "deconvolutiondepthwise.h"
-
-namespace ncnn {
-
-class DeconvolutionDepthWise_mips : public DeconvolutionDepthWise
-{
-public:
-    DeconvolutionDepthWise_mips();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    int create_group_ops(const Option& opt);
-
-public:
-    std::vector<ncnn::Layer*> group_ops;
-
-    Mat weight_data_tm;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTIONDEPTHWISE_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/dequantize_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/dequantize_mips.h
deleted file mode 100644
index 8ae7e542..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/dequantize_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DEQUANTIZE_MIPS_H
-#define LAYER_DEQUANTIZE_MIPS_H
-
-#include "dequantize.h"
-
-namespace ncnn {
-
-class Dequantize_mips : public Dequantize
-{
-public:
-    Dequantize_mips();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DEQUANTIZE_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/dropout_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/dropout_mips.h
deleted file mode 100644
index 05fa3846..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/dropout_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DROPOUT_MIPS_H
-#define LAYER_DROPOUT_MIPS_H
-
-#include "dropout.h"
-
-namespace ncnn {
-
-class Dropout_mips : public Dropout
-{
-public:
-    Dropout_mips();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DROPOUT_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/eltwise_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/eltwise_mips.h
deleted file mode 100644
index 9b4ac773..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/eltwise_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ELTWISE_MIPS_H
-#define LAYER_ELTWISE_MIPS_H
-
-#include "eltwise.h"
-
-namespace ncnn {
-
-class Eltwise_mips : public Eltwise
-{
-public:
-    Eltwise_mips();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ELTWISE_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/flatten_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/flatten_mips.h
deleted file mode 100644
index c9f33225..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/flatten_mips.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_FLATTEN_MIPS_H
-#define LAYER_FLATTEN_MIPS_H
-
-#include "flatten.h"
-
-namespace ncnn {
-
-class Flatten_mips : public Flatten
-{
-public:
-    Flatten_mips();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_FLATTEN_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/hardsigmoid_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/hardsigmoid_mips.h
deleted file mode 100644
index 51cab826..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/hardsigmoid_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_HARDSIGMOID_MIPS_H
-#define LAYER_HARDSIGMOID_MIPS_H
-
-#include "hardsigmoid.h"
-
-namespace ncnn {
-
-class HardSigmoid_mips : public HardSigmoid
-{
-public:
-    HardSigmoid_mips();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_HARDSIGMOID_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/hardswish_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/hardswish_mips.h
deleted file mode 100644
index 8ace7fe7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/hardswish_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_HARDSWISH_MIPS_H
-#define LAYER_HARDSWISH_MIPS_H
-
-#include "hardswish.h"
-
-namespace ncnn {
-
-class HardSwish_mips : public HardSwish
-{
-public:
-    HardSwish_mips();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_HARDSWISH_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/innerproduct_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/innerproduct_mips.h
deleted file mode 100644
index c96db3f9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/innerproduct_mips.h
+++ /dev/null
@@ -1,54 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INNERPRODUCT_MIPS_H
-#define LAYER_INNERPRODUCT_MIPS_H
-
-#include "innerproduct.h"
-
-namespace ncnn {
-
-class InnerProduct_mips : public InnerProduct
-{
-public:
-    InnerProduct_mips();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-#if __mips_msa
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_INT8
-    int create_pipeline_int8_mips(const Option& opt);
-    int forward_int8_mips(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Layer* flatten;
-
-    Mat weight_data_tm;
-
-#if NCNN_INT8
-    Mat scale_in_data;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INNERPRODUCT_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bicubic.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bicubic.h
deleted file mode 100644
index efff2d4d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bicubic.h
+++ /dev/null
@@ -1,261 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static inline void interpolate_cubic(float fx, float* coeffs)
-{
-    const float A = -0.75f;
-
-    float fx0 = fx + 1;
-    float fx1 = fx;
-    float fx2 = 1 - fx;
-    // float fx3 = 2 - fx;
-
-    coeffs[0] = A * fx0 * fx0 * fx0 - 5 * A * fx0 * fx0 + 8 * A * fx0 - 4 * A;
-    coeffs[1] = (A + 2) * fx1 * fx1 * fx1 - (A + 3) * fx1 * fx1 + 1;
-    coeffs[2] = (A + 2) * fx2 * fx2 * fx2 - (A + 3) * fx2 * fx2 + 1;
-    coeffs[3] = 1.f - coeffs[0] - coeffs[1] - coeffs[2];
-}
-
-static void cubic_coeffs(int w, int outw, int* xofs, float* alpha, int align_corner)
-{
-    double scale = (double)w / outw;
-    if (align_corner)
-    {
-        scale = (double)(w - 1) / (outw - 1);
-    }
-
-    for (int dx = 0; dx < outw; dx++)
-    {
-        float fx = (float)((dx + 0.5) * scale - 0.5);
-        if (align_corner)
-        {
-            fx = (float)(dx * scale);
-        }
-
-        int sx = static_cast<int>(floor(fx));
-        fx -= sx;
-
-        interpolate_cubic(fx, alpha + dx * 4);
-
-        if (sx <= -1)
-        {
-            sx = 1;
-            alpha[dx * 4 + 0] = 1.f - alpha[dx * 4 + 3];
-            alpha[dx * 4 + 1] = alpha[dx * 4 + 3];
-            alpha[dx * 4 + 2] = 0.f;
-            alpha[dx * 4 + 3] = 0.f;
-        }
-        if (sx == 0)
-        {
-            sx = 1;
-            alpha[dx * 4 + 0] = alpha[dx * 4 + 0] + alpha[dx * 4 + 1];
-            alpha[dx * 4 + 1] = alpha[dx * 4 + 2];
-            alpha[dx * 4 + 2] = alpha[dx * 4 + 3];
-            alpha[dx * 4 + 3] = 0.f;
-        }
-        if (sx == w - 2)
-        {
-            sx = w - 3;
-            alpha[dx * 4 + 3] = alpha[dx * 4 + 2] + alpha[dx * 4 + 3];
-            alpha[dx * 4 + 2] = alpha[dx * 4 + 1];
-            alpha[dx * 4 + 1] = alpha[dx * 4 + 0];
-            alpha[dx * 4 + 0] = 0.f;
-        }
-        if (sx >= w - 1)
-        {
-            sx = w - 3;
-            alpha[dx * 4 + 3] = 1.f - alpha[dx * 4 + 0];
-            alpha[dx * 4 + 2] = alpha[dx * 4 + 0];
-            alpha[dx * 4 + 1] = 0.f;
-            alpha[dx * 4 + 0] = 0.f;
-        }
-
-        xofs[dx] = sx;
-    }
-}
-
-static void resize_bicubic_image(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    Mat rowsbuf2(w);
-    Mat rowsbuf3(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows1p[dx] = S1p[-1] * a0 + S1p[0] * a1 + S1p[1] * a2 + S1p[2] * a3;
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const float* S0 = src.row(sy - 1);
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows0p[dx] = S0p[-1] * a0 + S0p[0] * a1 + S0p[1] * a2 + S0p[2] * a3;
-                rows1p[dx] = S1p[-1] * a0 + S1p[0] * a1 + S1p[1] * a2 + S1p[2] * a3;
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-        float b2 = beta[2];
-        float b3 = beta[3];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        float* Dp = dst.row(dy);
-        for (int dx = 0; dx < w; dx++)
-        {
-            //             D[x] = rows0[x]*b0 + rows1[x]*b1 + rows2[x]*b2 + rows3[x]*b3;
-            *Dp++ = *rows0p++ * b0 + *rows1p++ * b1 + *rows2p++ * b2 + *rows3p++ * b3;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bicubic_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bicubic_pack4.h
deleted file mode 100644
index 22a07e16..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bicubic_pack4.h
+++ /dev/null
@@ -1,286 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bicubic_image_pack4(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf1(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf2(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf3(w, (size_t)4 * 4u, 4);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S3p = S3 + sx;
-
-                v4f32 _a0 = __msa_fill_w_f32(alphap[0]);
-                v4f32 _a1 = __msa_fill_w_f32(alphap[1]);
-                v4f32 _a2 = __msa_fill_w_f32(alphap[2]);
-                v4f32 _a3 = __msa_fill_w_f32(alphap[3]);
-
-                v4f32 _S30 = (v4f32)__msa_ld_w(S3p - 4, 0);
-                v4f32 _S31 = (v4f32)__msa_ld_w(S3p + 0, 0);
-                v4f32 _S32 = (v4f32)__msa_ld_w(S3p + 4, 0);
-                v4f32 _S33 = (v4f32)__msa_ld_w(S3p + 8, 0);
-                v4f32 _rows3 = __msa_fmul_w(_S30, _a0);
-                _rows3 = __msa_fmadd_w(_rows3, _S31, _a1);
-                _rows3 = __msa_fmadd_w(_rows3, _S32, _a2);
-                _rows3 = __msa_fmadd_w(_rows3, _S33, _a3);
-                __msa_st_w((v4i32)_rows3, rows3p + dx * 4, 0);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                v4f32 _a0 = __msa_fill_w_f32(alphap[0]);
-                v4f32 _a1 = __msa_fill_w_f32(alphap[1]);
-                v4f32 _a2 = __msa_fill_w_f32(alphap[2]);
-                v4f32 _a3 = __msa_fill_w_f32(alphap[3]);
-
-                v4f32 _S20 = (v4f32)__msa_ld_w(S2p - 4, 0);
-                v4f32 _S21 = (v4f32)__msa_ld_w(S2p + 0, 0);
-                v4f32 _S22 = (v4f32)__msa_ld_w(S2p + 4, 0);
-                v4f32 _S23 = (v4f32)__msa_ld_w(S2p + 8, 0);
-                v4f32 _S30 = (v4f32)__msa_ld_w(S3p - 4, 0);
-                v4f32 _S31 = (v4f32)__msa_ld_w(S3p + 0, 0);
-                v4f32 _S32 = (v4f32)__msa_ld_w(S3p + 4, 0);
-                v4f32 _S33 = (v4f32)__msa_ld_w(S3p + 8, 0);
-                v4f32 _rows2 = __msa_fmul_w(_S20, _a0);
-                v4f32 _rows3 = __msa_fmul_w(_S30, _a0);
-                _rows2 = __msa_fmadd_w(_rows2, _S21, _a1);
-                _rows3 = __msa_fmadd_w(_rows3, _S31, _a1);
-                _rows2 = __msa_fmadd_w(_rows2, _S22, _a2);
-                _rows3 = __msa_fmadd_w(_rows3, _S32, _a2);
-                _rows2 = __msa_fmadd_w(_rows2, _S23, _a3);
-                _rows3 = __msa_fmadd_w(_rows3, _S33, _a3);
-                __msa_st_w((v4i32)_rows2, rows2p + dx * 4, 0);
-                __msa_st_w((v4i32)_rows3, rows3p + dx * 4, 0);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                v4f32 _a0 = __msa_fill_w_f32(alphap[0]);
-                v4f32 _a1 = __msa_fill_w_f32(alphap[1]);
-                v4f32 _a2 = __msa_fill_w_f32(alphap[2]);
-                v4f32 _a3 = __msa_fill_w_f32(alphap[3]);
-
-                v4f32 _S10 = (v4f32)__msa_ld_w(S1p - 4, 0);
-                v4f32 _S11 = (v4f32)__msa_ld_w(S1p + 0, 0);
-                v4f32 _S12 = (v4f32)__msa_ld_w(S1p + 4, 0);
-                v4f32 _S13 = (v4f32)__msa_ld_w(S1p + 8, 0);
-                v4f32 _S20 = (v4f32)__msa_ld_w(S2p - 4, 0);
-                v4f32 _S21 = (v4f32)__msa_ld_w(S2p + 0, 0);
-                v4f32 _S22 = (v4f32)__msa_ld_w(S2p + 4, 0);
-                v4f32 _S23 = (v4f32)__msa_ld_w(S2p + 8, 0);
-                v4f32 _S30 = (v4f32)__msa_ld_w(S3p - 4, 0);
-                v4f32 _S31 = (v4f32)__msa_ld_w(S3p + 0, 0);
-                v4f32 _S32 = (v4f32)__msa_ld_w(S3p + 4, 0);
-                v4f32 _S33 = (v4f32)__msa_ld_w(S3p + 8, 0);
-                v4f32 _rows1 = __msa_fmul_w(_S10, _a0);
-                v4f32 _rows2 = __msa_fmul_w(_S20, _a0);
-                v4f32 _rows3 = __msa_fmul_w(_S30, _a0);
-                _rows1 = __msa_fmadd_w(_rows1, _S11, _a1);
-                _rows2 = __msa_fmadd_w(_rows2, _S21, _a1);
-                _rows3 = __msa_fmadd_w(_rows3, _S31, _a1);
-                _rows1 = __msa_fmadd_w(_rows1, _S12, _a2);
-                _rows2 = __msa_fmadd_w(_rows2, _S22, _a2);
-                _rows3 = __msa_fmadd_w(_rows3, _S32, _a2);
-                _rows1 = __msa_fmadd_w(_rows1, _S13, _a3);
-                _rows2 = __msa_fmadd_w(_rows2, _S23, _a3);
-                _rows3 = __msa_fmadd_w(_rows3, _S33, _a3);
-                __msa_st_w((v4i32)_rows1, rows1p + dx * 4, 0);
-                __msa_st_w((v4i32)_rows2, rows2p + dx * 4, 0);
-                __msa_st_w((v4i32)_rows3, rows3p + dx * 4, 0);
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const float* S0 = src.row(sy - 1);
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                v4f32 _a0 = __msa_fill_w_f32(alphap[0]);
-                v4f32 _a1 = __msa_fill_w_f32(alphap[1]);
-                v4f32 _a2 = __msa_fill_w_f32(alphap[2]);
-                v4f32 _a3 = __msa_fill_w_f32(alphap[3]);
-
-                v4f32 _S00 = (v4f32)__msa_ld_w(S0p - 4, 0);
-                v4f32 _S01 = (v4f32)__msa_ld_w(S0p + 0, 0);
-                v4f32 _S02 = (v4f32)__msa_ld_w(S0p + 4, 0);
-                v4f32 _S03 = (v4f32)__msa_ld_w(S0p + 8, 0);
-                v4f32 _S10 = (v4f32)__msa_ld_w(S1p - 4, 0);
-                v4f32 _S11 = (v4f32)__msa_ld_w(S1p + 0, 0);
-                v4f32 _S12 = (v4f32)__msa_ld_w(S1p + 4, 0);
-                v4f32 _S13 = (v4f32)__msa_ld_w(S1p + 8, 0);
-                v4f32 _S20 = (v4f32)__msa_ld_w(S2p - 4, 0);
-                v4f32 _S21 = (v4f32)__msa_ld_w(S2p + 0, 0);
-                v4f32 _S22 = (v4f32)__msa_ld_w(S2p + 4, 0);
-                v4f32 _S23 = (v4f32)__msa_ld_w(S2p + 8, 0);
-                v4f32 _S30 = (v4f32)__msa_ld_w(S3p - 4, 0);
-                v4f32 _S31 = (v4f32)__msa_ld_w(S3p + 0, 0);
-                v4f32 _S32 = (v4f32)__msa_ld_w(S3p + 4, 0);
-                v4f32 _S33 = (v4f32)__msa_ld_w(S3p + 8, 0);
-                v4f32 _rows0 = __msa_fmul_w(_S00, _a0);
-                v4f32 _rows1 = __msa_fmul_w(_S10, _a0);
-                v4f32 _rows2 = __msa_fmul_w(_S20, _a0);
-                v4f32 _rows3 = __msa_fmul_w(_S30, _a0);
-                _rows0 = __msa_fmadd_w(_rows0, _S01, _a1);
-                _rows1 = __msa_fmadd_w(_rows1, _S11, _a1);
-                _rows2 = __msa_fmadd_w(_rows2, _S21, _a1);
-                _rows3 = __msa_fmadd_w(_rows3, _S31, _a1);
-                _rows0 = __msa_fmadd_w(_rows0, _S02, _a2);
-                _rows1 = __msa_fmadd_w(_rows1, _S12, _a2);
-                _rows2 = __msa_fmadd_w(_rows2, _S22, _a2);
-                _rows3 = __msa_fmadd_w(_rows3, _S32, _a2);
-                _rows0 = __msa_fmadd_w(_rows0, _S03, _a3);
-                _rows1 = __msa_fmadd_w(_rows1, _S13, _a3);
-                _rows2 = __msa_fmadd_w(_rows2, _S23, _a3);
-                _rows3 = __msa_fmadd_w(_rows3, _S33, _a3);
-                __msa_st_w((v4i32)_rows0, rows0p + dx * 4, 0);
-                __msa_st_w((v4i32)_rows1, rows1p + dx * 4, 0);
-                __msa_st_w((v4i32)_rows2, rows2p + dx * 4, 0);
-                __msa_st_w((v4i32)_rows3, rows3p + dx * 4, 0);
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        v4f32 _b0 = __msa_fill_w_f32(beta[0]);
-        v4f32 _b1 = __msa_fill_w_f32(beta[1]);
-        v4f32 _b2 = __msa_fill_w_f32(beta[2]);
-        v4f32 _b3 = __msa_fill_w_f32(beta[3]);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        float* Dp = dst.row(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            v4f32 _rows0 = (v4f32)__msa_ld_w(rows0p, 0);
-            v4f32 _rows1 = (v4f32)__msa_ld_w(rows1p, 0);
-            v4f32 _rows2 = (v4f32)__msa_ld_w(rows2p, 0);
-            v4f32 _rows3 = (v4f32)__msa_ld_w(rows3p, 0);
-            v4f32 _Dp = __msa_fmul_w(_rows0, _b0);
-            _Dp = __msa_fmadd_w(_Dp, _rows1, _b1);
-            _Dp = __msa_fmadd_w(_Dp, _rows2, _b2);
-            _Dp = __msa_fmadd_w(_Dp, _rows3, _b3);
-            __msa_st_w((v4i32)_Dp, Dp, 0);
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-            rows2p += 4;
-            rows3p += 4;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bilinear.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bilinear.h
deleted file mode 100644
index eba05714..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bilinear.h
+++ /dev/null
@@ -1,172 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void linear_coeffs(int w, int outw, int* xofs, float* alpha, int align_corner)
-{
-    double scale = (double)w / outw;
-    if (align_corner)
-    {
-        scale = (double)(w - 1) / (outw - 1);
-    }
-
-    for (int dx = 0; dx < outw; dx++)
-    {
-        float fx = (float)((dx + 0.5) * scale - 0.5);
-        if (align_corner)
-        {
-            fx = (float)(dx * scale);
-        }
-
-        int sx = floor(fx);
-        fx -= sx;
-
-        if (sx < 0)
-        {
-            sx = 0;
-            fx = 0.f;
-        }
-        if (sx >= w - 1)
-        {
-            sx = w - 2;
-            fx = 1.f;
-        }
-
-        xofs[dx] = sx;
-
-        alpha[dx * 2] = 1.f - fx;
-        alpha[dx * 2 + 1] = fx;
-    }
-}
-
-static void resize_bilinear_image(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows1p[dx] = S1p[0] * a0 + S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const float* S0 = src.row(sy);
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows0p[dx] = S0p[0] * a0 + S0p[1] * a1;
-                rows1p[dx] = S1p[0] * a0 + S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* Dp = dst.row(dy);
-
-#if __mips_msa
-        int nn = w >> 3;
-#else
-        int nn = 0;
-#endif
-        int remain = w - (nn << 3);
-
-#if __mips_msa
-        v4f32 _b0 = __msa_fill_w_f32(b0);
-        v4f32 _b1 = __msa_fill_w_f32(b1);
-        for (; nn > 0; nn--)
-        {
-            v4f32 _rows0 = (v4f32)__msa_ld_w(rows0p, 0);
-            v4f32 _rows1 = (v4f32)__msa_ld_w(rows1p, 0);
-
-            v4f32 _Dp = __msa_fmul_w(_rows0, _b0);
-            _Dp = __msa_fmadd_w(_Dp, _rows1, _b1);
-
-            __msa_st_w((v4i32)_Dp, Dp, 0);
-
-            v4f32 _rows0n = (v4f32)__msa_ld_w(rows0p + 4, 0);
-            v4f32 _rows1n = (v4f32)__msa_ld_w(rows1p + 4, 0);
-
-            v4f32 _Dpn = __msa_fmul_w(_rows0n, _b0);
-            _Dpn = __msa_fmadd_w(_Dpn, _rows1n, _b1);
-
-            __msa_st_w((v4i32)_Dpn, Dp + 4, 0);
-
-            Dp += 8;
-            rows0p += 8;
-            rows1p += 8;
-        }
-#endif // __mips_msa
-        for (; remain; --remain)
-        {
-            //             D[x] = rows0[x]*b0 + rows1[x]*b1;
-            *Dp++ = *rows0p++ * b0 + *rows1p++ * b1;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bilinear_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bilinear_pack4.h
deleted file mode 100644
index 500e8648..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_bilinear_pack4.h
+++ /dev/null
@@ -1,123 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bilinear_image_pack4(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf1(w, (size_t)4 * 4u, 4);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S1p = S1 + sx;
-
-                v4f32 _a0 = __msa_fill_w_f32(alphap[0]);
-                v4f32 _a1 = __msa_fill_w_f32(alphap[1]);
-
-                v4f32 _S10 = (v4f32)__msa_ld_w(S1p, 0);
-                v4f32 _S11 = (v4f32)__msa_ld_w(S1p + 4, 0);
-                v4f32 _rows1 = __msa_fmul_w(_S10, _a0);
-                _rows1 = __msa_fmadd_w(_rows1, _S11, _a1);
-                __msa_st_w((v4i32)_rows1, rows1p + dx * 4, 0);
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const float* S0 = src.row(sy);
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-
-                v4f32 _a0 = __msa_fill_w_f32(alphap[0]);
-                v4f32 _a1 = __msa_fill_w_f32(alphap[1]);
-
-                v4f32 _S00 = (v4f32)__msa_ld_w(S0p, 0);
-                v4f32 _S01 = (v4f32)__msa_ld_w(S0p + 4, 0);
-                v4f32 _S10 = (v4f32)__msa_ld_w(S1p, 0);
-                v4f32 _S11 = (v4f32)__msa_ld_w(S1p + 4, 0);
-                v4f32 _rows0 = __msa_fmul_w(_S00, _a0);
-                v4f32 _rows1 = __msa_fmul_w(_S10, _a0);
-                _rows0 = __msa_fmadd_w(_rows0, _S01, _a1);
-                _rows1 = __msa_fmadd_w(_rows1, _S11, _a1);
-                __msa_st_w((v4i32)_rows0, rows0p + dx * 4, 0);
-                __msa_st_w((v4i32)_rows1, rows1p + dx * 4, 0);
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        v4f32 _b0 = __msa_fill_w_f32(beta[0]);
-        v4f32 _b1 = __msa_fill_w_f32(beta[1]);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* Dp = dst.row(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            v4f32 _rows0 = (v4f32)__msa_ld_w(rows0p, 0);
-            v4f32 _rows1 = (v4f32)__msa_ld_w(rows1p, 0);
-            v4f32 _Dp = __msa_fmul_w(_rows0, _b0);
-            _Dp = __msa_fmadd_w(_Dp, _rows1, _b1);
-            __msa_st_w((v4i32)_Dp, Dp, 0);
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_mips.h
deleted file mode 100644
index baff10b4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/interp_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INTERP_MIPS_H
-#define LAYER_INTERP_MIPS_H
-
-#include "interp.h"
-
-namespace ncnn {
-
-class Interp_mips : public Interp
-{
-public:
-    Interp_mips();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INTERP_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/loongson_mmi.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/loongson_mmi.h
deleted file mode 100644
index 9ea57785..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/loongson_mmi.h
+++ /dev/null
@@ -1,268 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright(C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License(the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef _MIPS_LOONGSON_MMI_H
-#define _MIPS_LOONGSON_MMI_H
-
-#if __mips_loongson_mmi
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#include <stdint.h>
-
-typedef uint8_t uint8x8_t __attribute__((vector_size(8)));
-typedef uint16_t uint16x4_t __attribute__((vector_size(8)));
-typedef uint32_t uint32x2_t __attribute__((vector_size(8)));
-
-typedef int8_t int8x8_t __attribute__((vector_size(8)));
-typedef int16_t int16x4_t __attribute__((vector_size(8)));
-typedef int32_t int32x2_t __attribute__((vector_size(8)));
-
-static inline __attribute__((__always_inline__)) uint32x2_t __mmi_pzerow_u()
-{
-    return (uint32x2_t) {
-        0, 0
-    };
-}
-static inline __attribute__((__always_inline__)) uint16x4_t __mmi_pzeroh_u()
-{
-    return (uint16x4_t) {
-        0, 0, 0, 0
-    };
-}
-static inline __attribute__((__always_inline__)) uint8x8_t __mmi_pzerob_u()
-{
-    return (uint8x8_t) {
-        0, 0, 0, 0, 0, 0, 0, 0
-    };
-}
-
-static inline __attribute__((__always_inline__)) int32x2_t __mmi_pzerow_s()
-{
-    return (int32x2_t) {
-        0, 0
-    };
-}
-static inline __attribute__((__always_inline__)) int16x4_t __mmi_pzeroh_s()
-{
-    return (int16x4_t) {
-        0, 0, 0, 0
-    };
-}
-static inline __attribute__((__always_inline__)) int8x8_t __mmi_pzerob_s()
-{
-    return (int8x8_t) {
-        0, 0, 0, 0, 0, 0, 0, 0
-    };
-}
-
-static inline __attribute__((__always_inline__)) uint32x2_t __mmi_pfillw_u(uint32_t v)
-{
-    return (uint32x2_t) {
-        v, v
-    };
-}
-static inline __attribute__((__always_inline__)) uint16x4_t __mmi_pfillh_u(uint16_t v)
-{
-    return (uint16x4_t) {
-        v, v, v, v
-    };
-}
-static inline __attribute__((__always_inline__)) uint8x8_t __mmi_pfillb_u(uint8_t v)
-{
-    return (uint8x8_t) {
-        v, v, v, v, v, v, v, v
-    };
-}
-
-static inline __attribute__((__always_inline__)) int32x2_t __mmi_pfillw_s(int32_t v)
-{
-    return (int32x2_t) {
-        v, v
-    };
-}
-static inline __attribute__((__always_inline__)) int16x4_t __mmi_pfillh_s(int16_t v)
-{
-    return (int16x4_t) {
-        v, v, v, v
-    };
-}
-static inline __attribute__((__always_inline__)) int8x8_t __mmi_pfillb_s(int8_t v)
-{
-    return (int8x8_t) {
-        v, v, v, v, v, v, v, v
-    };
-}
-
-static inline __attribute__((__always_inline__)) uint32x2_t __mmi_pldw_u(const void* p)
-{
-    return *(const uint32x2_t*)p;
-}
-static inline __attribute__((__always_inline__)) uint16x4_t __mmi_pldh_u(const void* p)
-{
-    return *(const uint16x4_t*)p;
-}
-static inline __attribute__((__always_inline__)) uint8x8_t __mmi_pldb_u(const void* p)
-{
-    return *(const uint8x8_t*)p;
-}
-
-static inline __attribute__((__always_inline__)) int32x2_t __mmi_pldw_s(const void* p)
-{
-    return *(const int32x2_t*)p;
-}
-static inline __attribute__((__always_inline__)) int16x4_t __mmi_pldh_s(const void* p)
-{
-    return *(const int16x4_t*)p;
-}
-static inline __attribute__((__always_inline__)) int8x8_t __mmi_pldb_s(const void* p)
-{
-    return *(const int8x8_t*)p;
-}
-
-static inline __attribute__((__always_inline__)) void __mmi_pstw_u(void* p, uint32x2_t v)
-{
-    *(uint32x2_t*)p = v;
-}
-static inline __attribute__((__always_inline__)) void __mmi_psth_u(void* p, uint16x4_t v)
-{
-    *(uint16x4_t*)p = v;
-}
-static inline __attribute__((__always_inline__)) void __mmi_pstb_u(void* p, uint8x8_t v)
-{
-    *(uint8x8_t*)p = v;
-}
-
-static inline __attribute__((__always_inline__)) void __mmi_pstw_s(void* p, int32x2_t v)
-{
-    *(int32x2_t*)p = v;
-}
-static inline __attribute__((__always_inline__)) void __mmi_psth_s(void* p, int16x4_t v)
-{
-    *(int16x4_t*)p = v;
-}
-static inline __attribute__((__always_inline__)) void __mmi_pstb_s(void* p, int8x8_t v)
-{
-    *(int8x8_t*)p = v;
-}
-
-#define __mmi_packsswh    __builtin_loongson_packsswh
-#define __mmi_packsshb    __builtin_loongson_packsshb
-#define __mmi_packushb    __builtin_loongson_packushb
-#define __mmi_paddw_u     __builtin_loongson_paddw_u
-#define __mmi_paddh_u     __builtin_loongson_paddh_u
-#define __mmi_paddb_u     __builtin_loongson_paddb_u
-#define __mmi_paddw_s     __builtin_loongson_paddw_s
-#define __mmi_paddh_s     __builtin_loongson_paddh_s
-#define __mmi_paddb_s     __builtin_loongson_paddb_s
-#define __mmi_paddd_u     __builtin_loongson_paddd_u
-#define __mmi_paddd_s     __builtin_loongson_paddd_s
-#define __mmi_paddsh      __builtin_loongson_paddsh
-#define __mmi_paddsb      __builtin_loongson_paddsb
-#define __mmi_paddush     __builtin_loongson_paddush
-#define __mmi_paddusb     __builtin_loongson_paddusb
-#define __mmi_pandn_ud    __builtin_loongson_pandn_ud
-#define __mmi_pandn_uw    __builtin_loongson_pandn_uw
-#define __mmi_pandn_uh    __builtin_loongson_pandn_uh
-#define __mmi_pandn_ub    __builtin_loongson_pandn_ub
-#define __mmi_pandn_sd    __builtin_loongson_pandn_sd
-#define __mmi_pandn_sw    __builtin_loongson_pandn_sw
-#define __mmi_pandn_sh    __builtin_loongson_pandn_sh
-#define __mmi_pandn_sb    __builtin_loongson_pandn_sb
-#define __mmi_pavgh       __builtin_loongson_pavgh
-#define __mmi_pavgb       __builtin_loongson_pavgb
-#define __mmi_pcmpeqw_u   __builtin_loongson_pcmpeqw_u
-#define __mmi_pcmpeqh_u   __builtin_loongson_pcmpeqh_u
-#define __mmi_pcmpeqb_u   __builtin_loongson_pcmpeqb_u
-#define __mmi_pcmpeqw_s   __builtin_loongson_pcmpeqw_s
-#define __mmi_pcmpeqh_s   __builtin_loongson_pcmpeqh_s
-#define __mmi_pcmpeqb_s   __builtin_loongson_pcmpeqb_s
-#define __mmi_pcmpgtw_u   __builtin_loongson_pcmpgtw_u
-#define __mmi_pcmpgth_u   __builtin_loongson_pcmpgth_u
-#define __mmi_pcmpgtb_u   __builtin_loongson_pcmpgtb_u
-#define __mmi_pcmpgtw_s   __builtin_loongson_pcmpgtw_s
-#define __mmi_pcmpgth_s   __builtin_loongson_pcmpgth_s
-#define __mmi_pcmpgtb_s   __builtin_loongson_pcmpgtb_s
-#define __mmi_pextrh_u    __builtin_loongson_pextrh_u
-#define __mmi_pextrh_s    __builtin_loongson_pextrh_s
-#define __mmi_pinsrh_0_u  __builtin_loongson_pinsrh_0_u
-#define __mmi_pinsrh_1_u  __builtin_loongson_pinsrh_1_u
-#define __mmi_pinsrh_2_u  __builtin_loongson_pinsrh_2_u
-#define __mmi_pinsrh_3_u  __builtin_loongson_pinsrh_3_u
-#define __mmi_pinsrh_0_s  __builtin_loongson_pinsrh_0_s
-#define __mmi_pinsrh_1_s  __builtin_loongson_pinsrh_1_s
-#define __mmi_pinsrh_2_s  __builtin_loongson_pinsrh_2_s
-#define __mmi_pinsrh_3_s  __builtin_loongson_pinsrh_3_s
-#define __mmi_pmaddhw     __builtin_loongson_pmaddhw
-#define __mmi_pmaxsh      __builtin_loongson_pmaxsh
-#define __mmi_pmaxub      __builtin_loongson_pmaxub
-#define __mmi_pminsh      __builtin_loongson_pminsh
-#define __mmi_pminub      __builtin_loongson_pminub
-#define __mmi_pmovmskb_u  __builtin_loongson_pmovmskb_u
-#define __mmi_pmovmskb_s  __builtin_loongson_pmovmskb_s
-#define __mmi_pmulhuh     __builtin_loongson_pmulhuh
-#define __mmi_pmulhh      __builtin_loongson_pmulhh
-#define __mmi_pmullh      __builtin_loongson_pmullh
-#define __mmi_pmuluw      __builtin_loongson_pmuluw
-#define __mmi_pasubub     __builtin_loongson_pasubub
-#define __mmi_biadd       __builtin_loongson_biadd
-#define __mmi_psadbh      __builtin_loongson_psadbh
-#define __mmi_pshufh_u    __builtin_loongson_pshufh_u
-#define __mmi_pshufh_s    __builtin_loongson_pshufh_s
-#define __mmi_psllh_u     __builtin_loongson_psllh_u
-#define __mmi_psllh_s     __builtin_loongson_psllh_s
-#define __mmi_psllw_u     __builtin_loongson_psllw_u
-#define __mmi_psllw_s     __builtin_loongson_psllw_s
-#define __mmi_psrlh_u     __builtin_loongson_psrlh_u
-#define __mmi_psrlh_s     __builtin_loongson_psrlh_s
-#define __mmi_psrlw_u     __builtin_loongson_psrlw_u
-#define __mmi_psrlw_s     __builtin_loongson_psrlw_s
-#define __mmi_psrah_u     __builtin_loongson_psrah_u
-#define __mmi_psrah_s     __builtin_loongson_psrah_s
-#define __mmi_psraw_u     __builtin_loongson_psraw_u
-#define __mmi_psraw_s     __builtin_loongson_psraw_s
-#define __mmi_psubw_u     __builtin_loongson_psubw_u
-#define __mmi_psubh_u     __builtin_loongson_psubh_u
-#define __mmi_psubb_u     __builtin_loongson_psubb_u
-#define __mmi_psubw_s     __builtin_loongson_psubw_s
-#define __mmi_psubh_s     __builtin_loongson_psubh_s
-#define __mmi_psubb_s     __builtin_loongson_psubb_s
-#define __mmi_psubd_u     __builtin_loongson_psubd_u
-#define __mmi_psubd_s     __builtin_loongson_psubd_s
-#define __mmi_psubsh      __builtin_loongson_psubsh
-#define __mmi_psubsb      __builtin_loongson_psubsb
-#define __mmi_psubush     __builtin_loongson_psubush
-#define __mmi_psubusb     __builtin_loongson_psubusb
-#define __mmi_punpckhwd_u __builtin_loongson_punpckhwd_u
-#define __mmi_punpckhhw_u __builtin_loongson_punpckhhw_u
-#define __mmi_punpckhbh_u __builtin_loongson_punpckhbh_u
-#define __mmi_punpckhwd_s __builtin_loongson_punpckhwd_s
-#define __mmi_punpckhhw_s __builtin_loongson_punpckhhw_s
-#define __mmi_punpckhbh_s __builtin_loongson_punpckhbh_s
-#define __mmi_punpcklwd_u __builtin_loongson_punpcklwd_u
-#define __mmi_punpcklhw_u __builtin_loongson_punpcklhw_u
-#define __mmi_punpcklbh_u __builtin_loongson_punpcklbh_u
-#define __mmi_punpcklwd_s __builtin_loongson_punpcklwd_s
-#define __mmi_punpcklhw_s __builtin_loongson_punpcklhw_s
-#define __mmi_punpcklbh_s __builtin_loongson_punpcklbh_s
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif // __mips_loongson_mmi
-
-#endif // _MIPS_LOONGSON_MMI_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/mips_activation.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/mips_activation.h
deleted file mode 100644
index 4f4ad235..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/mips_activation.h
+++ /dev/null
@@ -1,70 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef MIPS_ACTIVATION_H
-#define MIPS_ACTIVATION_H
-
-#include "fused_activation.h"
-
-#if __mips_msa
-#include <msa.h>
-#include "msa_mathfun.h"
-
-static inline v4f32 activation_ps(v4f32 _v, int activation_type, const ncnn::Mat& activation_params)
-{
-    if (activation_type == 1)
-    {
-        v4f32 _zero = (v4f32)__msa_fill_w(0);
-        _v = __msa_fmax_w(_v, _zero);
-    }
-    else if (activation_type == 2)
-    {
-        v4f32 _zero = (v4f32)__msa_fill_w(0);
-        v4f32 _slope = (v4f32)__msa_fill_w_f32(activation_params[0]);
-        v4i32_w _lemask = __msa_fcle_w(_v, _zero);
-        v4f32 _ps = __msa_fmul_w(_v, _slope);
-        _v = (v4f32)__msa_bsel_v((v16u8)_lemask, (v16u8)_v, (v16u8)_ps);
-    }
-    else if (activation_type == 3)
-    {
-        v4f32 _min = (v4f32)__msa_fill_w_f32(activation_params[0]);
-        v4f32 _max = (v4f32)__msa_fill_w_f32(activation_params[1]);
-        _v = __msa_fmax_w(_v, _min);
-        _v = __msa_fmin_w(_v, _max);
-    }
-    else if (activation_type == 4)
-    {
-        _v = sigmoid_ps(_v);
-    }
-    else if (activation_type == 5)
-    {
-        _v = __msa_fmul_w(_v, tanh_ps(log_ps(__msa_fadd_w(exp_ps(_v), (v4f32)__msa_fill_w_f32(1.f)))));
-    }
-    else if (activation_type == 6)
-    {
-        v4f32 _alpha = (v4f32)__msa_fill_w_f32(activation_params[0]);
-        v4f32 _beta = (v4f32)__msa_fill_w_f32(activation_params[1]);
-        v4f32 _zero = (v4f32)__msa_fill_w(0);
-        v4f32 _one = (v4f32)__msa_fill_w_f32(1.f);
-        v4f32 _outp = __msa_fmadd_w(_beta, _v, _alpha);
-        _outp = __msa_fmax_w(_outp, _zero);
-        _outp = __msa_fmin_w(_outp, _one);
-        _v = __msa_fmul_w(_outp, _v);
-    }
-
-    return _v;
-}
-#endif // __mips_msa
-
-#endif // MIPS_ACTIVATION_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/mips_usability.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/mips_usability.h
deleted file mode 100644
index 662320ee..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/mips_usability.h
+++ /dev/null
@@ -1,251 +0,0 @@
-// Leo is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 Leo <leo@nullptr.com.cn>. All rights reserved.
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef MIPS_USABILITY_H
-#define MIPS_USABILITY_H
-
-#if __mips_msa
-#include <msa.h>
-#endif // __mips_msa
-
-#include <stdint.h>
-
-namespace ncnn {
-
-typedef union
-{
-    int32_t i;
-    float f;
-} FloatInt;
-
-} // namespace ncnn
-
-#if __mips_msa
-/* declare some mips constants with union */
-#define _MIPS_FLOAT_CONST(Name, Val) \
-    static const ncnn::FloatInt Name = {.f = Val}
-
-/* float type data load instructions */
-static NCNN_FORCEINLINE v4f32 __msa_fill_w_f32(float val)
-{
-    ncnn::FloatInt fi_tmpval = {.f = val};
-    return (v4f32)__msa_fill_w(fi_tmpval.i);
-}
-
-static NCNN_FORCEINLINE float __msa_reduce_fadd_w(v4f32 _v)
-{
-    // TODO find a more efficient way
-    return _v[0] + _v[1] + _v[2] + _v[3];
-}
-
-static NCNN_FORCEINLINE int __msa_reduce_add_w(v4i32 _v)
-{
-    // TODO find a more efficient way
-    return _v[0] + _v[1] + _v[2] + _v[3];
-}
-
-static NCNN_FORCEINLINE int __msa_cfcmsa_msacsr()
-{
-    int v;
-    asm volatile("cfcmsa %0, $1 \n"
-                 : "=r"(v)
-                 :
-                 :);
-    return v;
-}
-
-static NCNN_FORCEINLINE void __msa_ctcmsa_msacsr(int v)
-{
-    asm volatile("ctcmsa $1, %0 \n"
-                 :
-                 : "r"(v)
-                 :);
-}
-#endif // __mips_msa
-
-static NCNN_FORCEINLINE signed char float2int8(float v)
-{
-    int int32 = round(v);
-    if (int32 > 127) return 127;
-    if (int32 < -127) return -127;
-    return (signed char)int32;
-}
-
-#if __mips_msa
-static NCNN_FORCEINLINE v16i8 float2int8(v4f32 _v)
-{
-    // simulate round to nearest via +/-0.5
-    v4f32 _p5 = (v4f32)__msa_fill_w_f32(0.5f);
-    v16u8 _signmask = (v16u8)__msa_fill_w(1 << 31);
-
-    v16u8 _sign = __msa_and_v((v16u8)_v, _signmask);
-    v4f32 _p5s = (v4f32)__msa_or_v((v16u8)_p5, _sign);
-    v4f32 _v5 = __msa_fadd_w(_v, _p5s);
-    v4i32 _v32 = __msa_ftrunc_s_w(_v5);
-
-    v8i16 _v32_16 = (v8i16)__msa_sat_s_w(_v32, 15);
-    v8i16 _v16 = __msa_pckev_h(_v32_16, _v32_16);
-    _v16 = __msa_max_s_h(_v16, __msa_fill_h(-127));
-    v16i8 _v16_8 = (v16i8)__msa_sat_s_h(_v16, 7);
-    v16i8 _v8 = __msa_pckev_b(_v16_8, _v16_8);
-
-    return _v8;
-}
-
-static NCNN_FORCEINLINE int64_t float2int8(v4f32 _vlow, v4f32 _vhigh)
-{
-    // simulate round to nearest via +/-0.5
-    v4f32 _p5 = (v4f32)__msa_fill_w_f32(0.5f);
-    v16u8 _signmask = (v16u8)__msa_fill_w(1 << 31);
-
-    v16u8 _signlow = __msa_and_v((v16u8)_vlow, _signmask);
-    v16u8 _signhigh = __msa_and_v((v16u8)_vhigh, _signmask);
-    v4f32 _p5low = (v4f32)__msa_or_v((v16u8)_p5, _signlow);
-    v4f32 _p5high = (v4f32)__msa_or_v((v16u8)_p5, _signhigh);
-    v4f32 _vlow5 = __msa_fadd_w(_vlow, _p5low);
-    v4f32 _vhigh5 = __msa_fadd_w(_vhigh, _p5high);
-    v4i32 _vlow32 = __msa_ftrunc_s_w(_vlow5);
-    v4i32 _vhigh32 = __msa_ftrunc_s_w(_vhigh5);
-
-    v8i16 _vlow32_16 = (v8i16)__msa_sat_s_w(_vlow32, 15);
-    v8i16 _vhigh32_16 = (v8i16)__msa_sat_s_w(_vhigh32, 15);
-    v8i16 _v16 = __msa_pckev_h(_vhigh32_16, _vlow32_16);
-    _v16 = __msa_max_s_h(_v16, __msa_fill_h(-127));
-    v16i8 _v16_8 = (v16i8)__msa_sat_s_h(_v16, 7);
-    v2i64 _v8 = (v2i64)__msa_pckev_b(_v16_8, _v16_8);
-
-    return _v8[0];
-}
-
-static NCNN_FORCEINLINE v16i8 float2int8relu(v4f32 _v)
-{
-    // simulate round to nearest via +/-0.5
-    v4f32 _p5 = (v4f32)__msa_fill_w_f32(0.5f);
-    v16u8 _signmask = (v16u8)__msa_fill_w(1 << 31);
-
-    v16u8 _sign = __msa_and_v((v16u8)_v, _signmask);
-    v4f32 _p5s = (v4f32)__msa_or_v((v16u8)_p5, _sign);
-    v4f32 _v5 = __msa_fadd_w(_v, _p5s);
-    v4i32 _v32 = __msa_ftrunc_s_w(_v5);
-
-    v8i16 _v32_16 = (v8i16)__msa_sat_s_w(_v32, 15);
-    v8i16 _v16 = __msa_pckev_h(_v32_16, _v32_16);
-    _v16 = __msa_maxi_s_h(_v16, 0);
-    v16i8 _v16_8 = (v16i8)__msa_sat_s_h(_v16, 7);
-    v16i8 _v8 = __msa_pckev_b(_v16_8, _v16_8);
-
-    return _v8;
-}
-
-static NCNN_FORCEINLINE int64_t float2int8relu(v4f32 _vlow, v4f32 _vhigh)
-{
-    // simulate round to nearest via +/-0.5
-    v4f32 _p5 = (v4f32)__msa_fill_w_f32(0.5f);
-    v16u8 _signmask = (v16u8)__msa_fill_w(1 << 31);
-
-    v16u8 _signlow = __msa_and_v((v16u8)_vlow, _signmask);
-    v16u8 _signhigh = __msa_and_v((v16u8)_vhigh, _signmask);
-    v4f32 _p5low = (v4f32)__msa_or_v((v16u8)_p5, _signlow);
-    v4f32 _p5high = (v4f32)__msa_or_v((v16u8)_p5, _signhigh);
-    v4f32 _vlow5 = __msa_fadd_w(_vlow, _p5low);
-    v4f32 _vhigh5 = __msa_fadd_w(_vhigh, _p5high);
-    v4i32 _vlow32 = __msa_ftrunc_s_w(_vlow5);
-    v4i32 _vhigh32 = __msa_ftrunc_s_w(_vhigh5);
-
-    v8i16 _vlow32_16 = (v8i16)__msa_sat_s_w(_vlow32, 15);
-    v8i16 _vhigh32_16 = (v8i16)__msa_sat_s_w(_vhigh32, 15);
-    v8i16 _v16 = __msa_pckev_h(_vhigh32_16, _vlow32_16);
-    _v16 = __msa_maxi_s_h(_v16, 0);
-    v16i8 _v16_8 = (v16i8)__msa_sat_s_h(_v16, 7);
-    v2i64 _v8 = (v2i64)__msa_pckev_b(_v16_8, _v16_8);
-
-    return _v8[0];
-}
-
-static NCNN_FORCEINLINE v16i8 float2int8leakyrelu(v4f32 _v, v4f32 _slope)
-{
-    v4f32 _v_leaky = __msa_fmul_w(_v, _slope);
-
-    // simulate round to nearest via +/-0.5
-    v4f32 _p5 = (v4f32)__msa_fill_w_f32(0.5f);
-    v16u8 _signmask = (v16u8)__msa_fill_w(1 << 31);
-
-    v16u8 _sign = __msa_and_v((v16u8)_v, _signmask);
-    v4f32 _p5s = (v4f32)__msa_or_v((v16u8)_p5, _sign);
-    v4f32 _v5 = __msa_fadd_w(_v, _p5s);
-    v4i32 _v32 = __msa_ftrunc_s_w(_v5);
-
-    v16u8 _sign_leaky = __msa_and_v((v16u8)_v_leaky, _signmask);
-    v4f32 _p5_leaky = (v4f32)__msa_or_v((v16u8)_p5, _sign_leaky);
-    v4f32 _v5_leaky = __msa_fadd_w(_v_leaky, _p5_leaky);
-    v4i32 _v32_leaky = __msa_ftrunc_s_w(_v5_leaky);
-
-    v8i16 _v32_16 = (v8i16)__msa_sat_s_w(_v32, 15);
-    v8i16 _v16 = __msa_pckev_h(_v32_16, _v32_16);
-
-    v8i16 _v32_16_leaky = (v8i16)__msa_sat_s_w(_v32_leaky, 15);
-    v8i16 _v16_leaky = __msa_pckev_h(_v32_16_leaky, _v32_16_leaky);
-
-    _v16 = __msa_max_s_h(_v16, _v16_leaky);
-    v16i8 _v16_8 = (v16i8)__msa_sat_s_h(_v16, 7);
-    v16i8 _v8 = __msa_pckev_b(_v16_8, _v16_8);
-
-    return _v8;
-}
-
-static NCNN_FORCEINLINE int64_t float2int8leakyrelu(v4f32 _vlow, v4f32 _vhigh, v4f32 _slope)
-{
-    v4f32 _vlow_leaky = __msa_fmul_w(_vlow, _slope);
-    v4f32 _vhigh_leaky = __msa_fmul_w(_vhigh, _slope);
-
-    // simulate round to nearest via +/-0.5
-    v4f32 _p5 = (v4f32)__msa_fill_w_f32(0.5f);
-    v16u8 _signmask = (v16u8)__msa_fill_w(1 << 31);
-
-    v16u8 _signlow = __msa_and_v((v16u8)_vlow, _signmask);
-    v16u8 _signhigh = __msa_and_v((v16u8)_vhigh, _signmask);
-    v4f32 _p5low = (v4f32)__msa_or_v((v16u8)_p5, _signlow);
-    v4f32 _p5high = (v4f32)__msa_or_v((v16u8)_p5, _signhigh);
-    v4f32 _vlow5 = __msa_fadd_w(_vlow, _p5low);
-    v4f32 _vhigh5 = __msa_fadd_w(_vhigh, _p5high);
-    v4i32 _vlow32 = __msa_ftrunc_s_w(_vlow5);
-    v4i32 _vhigh32 = __msa_ftrunc_s_w(_vhigh5);
-
-    v16u8 _signlow_leaky = __msa_and_v((v16u8)_vlow_leaky, _signmask);
-    v16u8 _signhigh_leaky = __msa_and_v((v16u8)_vhigh_leaky, _signmask);
-    v4f32 _p5low_leaky = (v4f32)__msa_or_v((v16u8)_p5, _signlow_leaky);
-    v4f32 _p5high_leaky = (v4f32)__msa_or_v((v16u8)_p5, _signhigh_leaky);
-    v4f32 _vlow5_leaky = __msa_fadd_w(_vlow_leaky, _p5low_leaky);
-    v4f32 _vhigh5_leaky = __msa_fadd_w(_vhigh_leaky, _p5high_leaky);
-    v4i32 _vlow32_leaky = __msa_ftrunc_s_w(_vlow5_leaky);
-    v4i32 _vhigh32_leaky = __msa_ftrunc_s_w(_vhigh5_leaky);
-
-    v8i16 _vlow32_16 = (v8i16)__msa_sat_s_w(_vlow32, 15);
-    v8i16 _vhigh32_16 = (v8i16)__msa_sat_s_w(_vhigh32, 15);
-    v8i16 _v16 = __msa_pckev_h(_vhigh32_16, _vlow32_16);
-
-    v8i16 _vlow32_16_leaky = (v8i16)__msa_sat_s_w(_vlow32_leaky, 15);
-    v8i16 _vhigh32_16_leaky = (v8i16)__msa_sat_s_w(_vhigh32_leaky, 15);
-    v8i16 _v16_leaky = __msa_pckev_h(_vhigh32_16_leaky, _vlow32_16_leaky);
-
-    _v16 = __msa_max_s_h(_v16, _v16_leaky);
-    v16i8 _v16_8 = (v16i8)__msa_sat_s_h(_v16, 7);
-    v2i64 _v8 = (v2i64)__msa_pckev_b(_v16_8, _v16_8);
-
-    return _v8[0];
-}
-#endif // __mips_msa
-
-#endif // MIPS_USABILITY_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/mish_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/mish_mips.h
deleted file mode 100644
index 33342a4f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/mish_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MISH_MIPS_H
-#define LAYER_MISH_MIPS_H
-
-#include "mish.h"
-
-namespace ncnn {
-
-class Mish_mips : public Mish
-{
-public:
-    Mish_mips();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MISH_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/msa_mathfun.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/msa_mathfun.h
deleted file mode 100644
index cab71acb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/msa_mathfun.h
+++ /dev/null
@@ -1,270 +0,0 @@
-/* MIPS implementation of exp
- *
- *   Inspired by Intel Approximate Math library, and based on the
- *   corresponding algorithms of the cephes math library
- */
-
-/* Copyright (C) 2020 Leo <leo@nullptr.com.cn>. All rights reserved.
- *
- *  This software is provided 'as-is', without any express or implied
- *  warranty.  In no event will the authors be held liable for any damages
- *  arising from the use of this software.
- *
- *  Permission is granted to anyone to use this software for any purpose,
- *  including commercial applications, and to alter it and redistribute it
- *  freely, subject to the following restrictions:
- *
- *  1. The origin of this software must not be misrepresented; you must not
- *     claim that you wrote the original software. If you use this software
- *     in a product, an acknowledgment in the product documentation would be
- *     appreciated but is not required.
- *  2. Altered source versions must be plainly marked as such, and must not be
- *     misrepresented as being the original software.
- *  3. This notice may not be removed or altered from any source distribution.
- *
- *  (this is the zlib license)
- */
-
-#ifndef MSA_MATHFUN_H
-#define MSA_MATHFUN_H
-
-#include "mips_usability.h"
-
-#include <msa.h>
-
-_MIPS_FLOAT_CONST(c_1, 1.0f);
-_MIPS_FLOAT_CONST(c_2, 2.0f);
-_MIPS_FLOAT_CONST(c_n1, -1.0f);
-_MIPS_FLOAT_CONST(c_0p5, 0.5f);
-
-#define c_inv_mant_mask ~0x7f800000u
-_MIPS_FLOAT_CONST(c_cephes_SQRTHF, 0.707106781186547524);
-_MIPS_FLOAT_CONST(c_cephes_log_p0, 7.0376836292E-2);
-_MIPS_FLOAT_CONST(c_cephes_log_p1, -1.1514610310E-1);
-_MIPS_FLOAT_CONST(c_cephes_log_p2, 1.1676998740E-1);
-_MIPS_FLOAT_CONST(c_cephes_log_p3, -1.2420140846E-1);
-_MIPS_FLOAT_CONST(c_cephes_log_p4, +1.4249322787E-1);
-_MIPS_FLOAT_CONST(c_cephes_log_p5, -1.6668057665E-1);
-_MIPS_FLOAT_CONST(c_cephes_log_p6, +2.0000714765E-1);
-_MIPS_FLOAT_CONST(c_cephes_log_p7, -2.4999993993E-1);
-_MIPS_FLOAT_CONST(c_cephes_log_p8, +3.3333331174E-1);
-_MIPS_FLOAT_CONST(c_cephes_log_q1, -2.12194440e-4);
-_MIPS_FLOAT_CONST(c_cephes_log_q2, 0.693359375);
-
-/* natural logarithm computed for 4 simultaneous float
- *   return NaN for x <= 0
- */
-static inline v4f32 log_ps(v4f32 x)
-{
-    v4f32 one = (v4f32)__msa_fill_w(c_1.i);
-
-    x = __msa_fmax_w(x, (v4f32)__msa_fill_w(0)); /* force flush to zero on denormal values */
-    v4i32 invalid_mask = __msa_fcle_w(x, (v4f32)__msa_fill_w(0));
-
-    v4i32 ux = (v4i32)(x);
-
-    v4i32 emm0 = __msa_srl_w(ux, (v4i32)__msa_fill_w(23));
-
-    /* keep only the fractional part */
-    ux = (v4i32)__msa_and_v((v16u8)ux, (v16u8)__msa_fill_w(c_inv_mant_mask));
-    ux = (v4i32)__msa_or_v((v16u8)ux, (v16u8)__msa_fill_w(c_0p5.i));
-    x = (v4f32)(ux);
-
-    emm0 = __msa_subv_w(emm0, (v4i32)__msa_fill_w(0x7f));
-    v4f32 e = __msa_ffint_s_w(emm0);
-
-    e = __msa_fadd_w(e, one);
-
-    /* part2:
-     *     if( x < SQRTHF ) {
-     *       e -= 1;
-     *       x = x + x - 1.0;
-     *     } else { x = x - 1.0; }
-     */
-    v4i32 mask = __msa_fclt_w(x, (v4f32)__msa_fill_w(c_cephes_SQRTHF.i));
-    v4f32 tmp = (v4f32)(__msa_and_v((v16u8)(x), (v16u8)mask));
-    x = __msa_fsub_w(x, one);
-    e = __msa_fsub_w(e, (v4f32)(__msa_and_v((v16u8)(one), (v16u8)mask)));
-    x = __msa_fadd_w(x, tmp);
-
-    v4f32 z = __msa_fmul_w(x, x);
-
-    v4f32 y = (v4f32)__msa_fill_w(c_cephes_log_p0.i);
-
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_cephes_log_p1.i), y, x);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_cephes_log_p2.i), y, x);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_cephes_log_p3.i), y, x);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_cephes_log_p4.i), y, x);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_cephes_log_p5.i), y, x);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_cephes_log_p6.i), y, x);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_cephes_log_p7.i), y, x);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_cephes_log_p8.i), y, x);
-    y = __msa_fmul_w(y, x);
-
-    y = __msa_fmul_w(y, z);
-
-    tmp = __msa_fmul_w(e, (v4f32)__msa_fill_w(c_cephes_log_q1.i));
-    y = __msa_fadd_w(y, tmp);
-
-    tmp = __msa_fmul_w(z, (v4f32)__msa_fill_w(c_0p5.i));
-    y = __msa_fsub_w(y, tmp);
-
-    tmp = __msa_fmul_w(e, (v4f32)__msa_fill_w(c_cephes_log_q2.i));
-    x = __msa_fadd_w(x, y);
-    x = __msa_fadd_w(x, tmp);
-    x = (v4f32)(__msa_or_v((v16u8)(x), (v16u8)invalid_mask)); // negative arg will be NAN
-    return x;
-}
-
-_MIPS_FLOAT_CONST(c_exp_hi, 88.3762626647949f);
-_MIPS_FLOAT_CONST(c_exp_lo, -88.3762626647949f);
-
-_MIPS_FLOAT_CONST(c_cephes_LOG2EF, 1.44269504088896341);
-_MIPS_FLOAT_CONST(c_cephes_exp_C1, 0.693359375);
-_MIPS_FLOAT_CONST(c_cephes_exp_C2, -2.12194440e-4);
-
-_MIPS_FLOAT_CONST(c_cephes_exp_p0, 1.9875691500E-4);
-_MIPS_FLOAT_CONST(c_cephes_exp_p1, 1.3981999507E-3);
-_MIPS_FLOAT_CONST(c_cephes_exp_p2, 8.3334519073E-3);
-_MIPS_FLOAT_CONST(c_cephes_exp_p3, 4.1665795894E-2);
-_MIPS_FLOAT_CONST(c_cephes_exp_p4, 1.6666665459E-1);
-_MIPS_FLOAT_CONST(c_cephes_exp_p5, 5.0000001201E-1);
-
-/* exp() computed for 4 float at once */
-static inline v4f32 exp_ps(v4f32 x)
-{
-    v4f32 tmp, fx;
-
-    v4f32 one = (v4f32)__msa_fill_w(c_1.i);
-    x = __msa_fmin_w(x, (v4f32)__msa_fill_w(c_exp_hi.i));
-    x = __msa_fmax_w(x, (v4f32)__msa_fill_w(c_exp_lo.i));
-
-    /* express exp(x) as exp(g + n*log(2)) */
-    fx = __msa_fmul_w(x, (v4f32)__msa_fill_w(c_cephes_LOG2EF.i));
-    fx = __msa_fadd_w(fx, (v4f32)__msa_fill_w(c_0p5.i));
-
-    /* perform a floorf */
-    tmp = __msa_ffint_s_w(__msa_ftint_s_w(fx));
-
-    /* if greater, substract 1 */
-    v4i32_w mask = __msa_fclt_w(fx, tmp);
-    mask = (v4i32_w)__msa_and_v((v16u8)mask, (v16u8)one);
-
-    fx = __msa_fsub_w(tmp, (v4f32)mask);
-
-    tmp = __msa_fmul_w(fx, (v4f32)__msa_fill_w(c_cephes_exp_C1.i));
-    v4f32 z = __msa_fmul_w(fx, (v4f32)__msa_fill_w(c_cephes_exp_C2.i));
-    x = __msa_fsub_w(x, tmp);
-    x = __msa_fsub_w(x, z);
-
-    v4f32 y = (v4f32)__msa_fill_w(c_cephes_exp_p0.i);
-
-    z = __msa_fmul_w(x, x);
-
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_cephes_exp_p1.i), y, x);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_cephes_exp_p2.i), y, x);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_cephes_exp_p3.i), y, x);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_cephes_exp_p4.i), y, x);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_cephes_exp_p5.i), y, x);
-
-    y = __msa_fmul_w(y, z);
-    y = __msa_fadd_w(y, x);
-    y = __msa_fadd_w(y, one);
-
-    /* build 2^n */
-    v4i32 mm;
-    mm = __msa_ftrunc_s_w(fx);
-    mm = __msa_addv_w(mm, __msa_fill_w(0x7f));
-    mm = __msa_sll_w(mm, __msa_fill_w(23));
-
-    y = __msa_fmul_w(y, (v4f32)mm);
-    return y;
-}
-
-_MIPS_FLOAT_CONST(c_tanh_tiny, 1e-4f);
-_MIPS_FLOAT_CONST(c_tanh_hi, 9.0f);
-// The monomial coefficients of the numerator polynomial (odd).
-_MIPS_FLOAT_CONST(c_tanh_alpha_1, 4.89352455891786e-3f);
-_MIPS_FLOAT_CONST(c_tanh_alpha_3, 6.37261928875436e-4f);
-_MIPS_FLOAT_CONST(c_tanh_alpha_5, 1.48572235717979e-5f);
-_MIPS_FLOAT_CONST(c_tanh_alpha_7, 5.12229709037114e-8f);
-_MIPS_FLOAT_CONST(c_tanh_alpha_9, -8.60467152213735e-11f);
-_MIPS_FLOAT_CONST(c_tanh_alpha_11, 2.00018790482477e-13f);
-_MIPS_FLOAT_CONST(c_tanh_alpha_13, -2.76076847742355e-16f);
-// The monomial coefficients of the denominator polynomial (even).
-_MIPS_FLOAT_CONST(c_tanh_beta_0, 4.89352518554385e-3f);
-_MIPS_FLOAT_CONST(c_tanh_beta_2, 2.26843463243900e-3f);
-_MIPS_FLOAT_CONST(c_tanh_beta_4, 1.18534705686654e-4f);
-_MIPS_FLOAT_CONST(c_tanh_beta_6, 1.19825839466702e-6f);
-
-/* tanh() computed for 4 float at once */
-static inline v4f32 tanh_ps(v4f32 x)
-{
-    v4f32 x2 = (v4f32)__msa_bclri_w((v4u32)x, 31);
-    v4i32 tiny_mask = __msa_fclt_w(x2, (v4f32)__msa_fill_w(c_tanh_tiny.i));
-
-    // clamp the inputs to the range [-9, 9] since anything outside
-    // this range is -/+1.0f in single-precision.
-    x2 = (v4f32)__msa_bsel_v((v16u8)__msa_fclt_w((v4f32)__msa_fill_w(c_tanh_hi.i), x2), (v16u8)x2, (v16u8)__msa_fill_w(c_tanh_hi.i));
-
-    // since the polynomials are odd/even, we need x**2.
-    v4f32 z = __msa_fmul_w(x2, x2);
-
-    // evaluate the numerator polynomial y.
-    v4f32 y = (v4f32)__msa_fill_w(c_tanh_alpha_13.i);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_tanh_alpha_11.i), y, z);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_tanh_alpha_9.i), y, z);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_tanh_alpha_7.i), y, z);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_tanh_alpha_5.i), y, z);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_tanh_alpha_3.i), y, z);
-    y = __msa_fmadd_w((v4f32)__msa_fill_w(c_tanh_alpha_1.i), y, z);
-    y = __msa_fmul_w(y, x2);
-
-    // evaluate the denominator polynomial w.
-    v4f32 w = (v4f32)__msa_fill_w(c_tanh_beta_6.i);
-    w = __msa_fmadd_w((v4f32)__msa_fill_w(c_tanh_beta_4.i), w, z);
-    w = __msa_fmadd_w((v4f32)__msa_fill_w(c_tanh_beta_2.i), w, z);
-    w = __msa_fmadd_w((v4f32)__msa_fill_w(c_tanh_beta_0.i), w, z);
-
-    // divide the numerator by the denominator.
-    y = __msa_fdiv_w(y, w);
-
-    // reinstate the sign.
-    y = (v4f32)__msa_binsli_w((v4u32)y, (v4u32)x, 0);
-
-    // when the argument is very small in magnitude it's more accurate to just return it.
-    y = (v4f32)__msa_bsel_v((v16u8)tiny_mask, (v16u8)y, (v16u8)x);
-
-    return y;
-}
-
-static inline v4f32 pow_ps(v4f32 a, v4f32 b)
-{
-    // pow(x, m) = exp(m * log(x))
-    return exp_ps(__msa_fmul_w(b, log_ps(a)));
-}
-
-static inline v4f32 sigmoid_ps(v4f32 _v)
-{
-    v4f32 _one = __msa_fill_w_f32(1.f);
-    _v = (v4f32)__msa_bnegi_w((v4u32)_v, 31);
-    _v = exp_ps(_v);
-    _v = __msa_fadd_w(_v, _one);
-    return __msa_fdiv_w(_one, _v);
-}
-
-static inline v4f32 atan2_ps(v4f32 a, v4f32 b)
-{
-    //TODO msa optimize
-    float tmpx[4];
-    float tmpy[4];
-    __msa_st_w((v4i32)a, tmpx, 0);
-    __msa_st_w((v4i32)b, tmpy, 0);
-    tmpx[0] = atan2(tmpx[0], tmpy[0]);
-    tmpx[1] = atan2(tmpx[1], tmpy[1]);
-    tmpx[2] = atan2(tmpx[2], tmpy[2]);
-    tmpx[3] = atan2(tmpx[3], tmpy[3]);
-    return (v4f32)__msa_ld_w(tmpx, 0);
-}
-
-#endif // MSA_MATHFUN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/packing_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/packing_mips.h
deleted file mode 100644
index ccc57f8a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/packing_mips.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PACKING_MIPS_H
-#define LAYER_PACKING_MIPS_H
-
-#include "packing.h"
-
-namespace ncnn {
-
-class Packing_mips : public Packing
-{
-public:
-    Packing_mips();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PACKING_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/padding_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/padding_mips.h
deleted file mode 100644
index 6d4ae8c2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/padding_mips.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PADDING_MIPS_H
-#define LAYER_PADDING_MIPS_H
-
-#include "padding.h"
-
-namespace ncnn {
-
-class Padding_mips : public Padding
-{
-public:
-    Padding_mips();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PADDING_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/padding_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/padding_pack4.h
deleted file mode 100644
index 9121a56b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/padding_pack4.h
+++ /dev/null
@@ -1,213 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void padding_constant_pack4_msa(const Mat& src, Mat& dst, int top, int bottom, int left, int right, v4f32 v)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-    int top_size = top * dst.w;
-    int bottom_size = bottom * dst.w;
-
-    // fill top
-    for (int y = 0; y < top_size; y++)
-    {
-        __msa_st_w((v4i32)v, outptr, 0);
-        outptr += 4;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            __msa_st_w((v4i32)v, outptr, 0);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            __builtin_prefetch(ptr + 32);
-            __msa_st_w(__msa_ld_w(ptr, 0), outptr, 0);
-            ptr += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __msa_st_w((v4i32)v, outptr, 0);
-            outptr += 4;
-        }
-    }
-    // fill top
-    for (int y = 0; y < bottom_size; y++)
-    {
-        __msa_st_w((v4i32)v, outptr, 0);
-        outptr += 4;
-    }
-}
-
-static void padding_replicate_pack4_msa(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-
-    // fill top
-    for (int y = 0; y < top; y++)
-    {
-        const float* ptr0 = ptr;
-        v4f32 _p = (v4f32)__msa_ld_w(ptr0, 0);
-        for (int x = 0; x < left; x++)
-        {
-            __msa_st_w((v4i32)_p, outptr, 0);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = (v4f32)__msa_ld_w(ptr0, 0);
-            __msa_st_w((v4i32)_p, outptr, 0);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __msa_st_w((v4i32)_p, outptr, 0);
-            outptr += 4;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        v4f32 _p = (v4f32)__msa_ld_w(ptr, 0);
-        for (int x = 0; x < left; x++)
-        {
-            __msa_st_w((v4i32)_p, outptr, 0);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = (v4f32)__msa_ld_w(ptr, 0);
-            __msa_st_w((v4i32)_p, outptr, 0);
-            ptr += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __msa_st_w((v4i32)_p, outptr, 0);
-            outptr += 4;
-        }
-    }
-    // fill bottom
-    ptr -= src.w * 4;
-    for (int y = 0; y < bottom; y++)
-    {
-        const float* ptr0 = ptr;
-        v4f32 _p = (v4f32)__msa_ld_w(ptr0, 0);
-        for (int x = 0; x < left; x++)
-        {
-            __msa_st_w((v4i32)_p, outptr, 0);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = (v4f32)__msa_ld_w(ptr0, 0);
-            __msa_st_w((v4i32)_p, outptr, 0);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __msa_st_w((v4i32)_p, outptr, 0);
-            outptr += 4;
-        }
-    }
-}
-
-static void padding_reflect_pack4_msa(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-
-    // fill top
-    ptr += top * src.w * 4;
-    for (int y = 0; y < top; y++)
-    {
-        const float* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            v4f32 _p = (v4f32)__msa_ld_w(ptr0 + (left - x) * 4, 0);
-            __msa_st_w((v4i32)_p, outptr, 0);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            v4f32 _p = (v4f32)__msa_ld_w(ptr0, 0);
-            __msa_st_w((v4i32)_p, outptr, 0);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            v4f32 _p = (v4f32)__msa_ld_w(ptr0 - 8 - x * 4, 0);
-            __msa_st_w((v4i32)_p, outptr, 0);
-            outptr += 4;
-        }
-        ptr -= src.w * 4;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            v4f32 _p = (v4f32)__msa_ld_w(ptr + (left - x) * 4, 0);
-            __msa_st_w((v4i32)_p, outptr, 0);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            v4f32 _p = (v4f32)__msa_ld_w(ptr, 0);
-            __msa_st_w((v4i32)_p, outptr, 0);
-            ptr += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            v4f32 _p = (v4f32)__msa_ld_w(ptr - 8 - x * 4, 0);
-            __msa_st_w((v4i32)_p, outptr, 0);
-            outptr += 4;
-        }
-    }
-    // fill bottom
-    ptr -= 2 * src.w * 4;
-    for (int y = 0; y < bottom; y++)
-    {
-        const float* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            v4f32 _p = (v4f32)__msa_ld_w(ptr0 + (left - x) * 4, 0);
-            __msa_st_w((v4i32)_p, outptr, 0);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            v4f32 _p = (v4f32)__msa_ld_w(ptr0, 0);
-            __msa_st_w((v4i32)_p, outptr, 0);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            v4f32 _p = (v4f32)__msa_ld_w(ptr0 - 8 - x * 4, 0);
-            __msa_st_w((v4i32)_p, outptr, 0);
-            outptr += 4;
-        }
-        ptr -= src.w * 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/padding_pack8_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/padding_pack8_int8.h
deleted file mode 100644
index 88fc52c8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/padding_pack8_int8.h
+++ /dev/null
@@ -1,171 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void padding_constant_pack8_int8_msa(const Mat& src, Mat& dst, int top, int bottom, int left, int right, int64_t _v)
-{
-    const int64_t* ptr = src;
-    int64_t* outptr = dst;
-
-    // fill top
-    for (int y = 0; y < top; y++)
-    {
-        for (int x = 0; x < dst.w; x++)
-        {
-            *outptr++ = _v;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = _v;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = _v;
-        }
-    }
-    // fill bottom
-    for (int y = 0; y < bottom; y++)
-    {
-        for (int x = 0; x < dst.w; x++)
-        {
-            *outptr++ = _v;
-        }
-    }
-}
-
-static void padding_replicate_pack8_int8_msa(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const int64_t* ptr = src;
-    int64_t* outptr = dst;
-
-    // fill top
-    for (int y = 0; y < top; y++)
-    {
-        const int64_t* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = *ptr0;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr0++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr0[-1];
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = *ptr;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr[-1];
-        }
-    }
-    // fill bottom
-    ptr -= src.w;
-    for (int y = 0; y < bottom; y++)
-    {
-        const int64_t* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = *ptr0;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr0++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr0[-1];
-        }
-    }
-}
-
-static void padding_reflect_pack8_int8_msa(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const int64_t* ptr = src;
-    int64_t* outptr = dst;
-
-    // fill top
-    ptr += top * src.w;
-    for (int y = 0; y < top; y++)
-    {
-        const int64_t* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = ptr0[left - x];
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr0++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr0[-2 - x];
-        }
-        ptr -= src.w;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = ptr[left - x];
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr[-2 - x];
-        }
-    }
-    // fill bottom
-    ptr -= 2 * src.w;
-    for (int y = 0; y < bottom; y++)
-    {
-        const int64_t* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = ptr0[left - x];
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr0++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr0[-2 - x];
-        }
-        ptr -= src.w;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/pooling_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/pooling_mips.h
deleted file mode 100644
index ec17a06a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/pooling_mips.h
+++ /dev/null
@@ -1,33 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_POOLING_MIPS_H
-#define LAYER_POOLING_MIPS_H
-
-#include "pooling.h"
-
-namespace ncnn {
-
-class Pooling_mips : public Pooling
-{
-public:
-    Pooling_mips();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_POOLING_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/prelu_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/prelu_mips.h
deleted file mode 100644
index 6174c257..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/prelu_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PRELU_MIPS_H
-#define LAYER_PRELU_MIPS_H
-
-#include "prelu.h"
-
-namespace ncnn {
-
-class PReLU_mips : public PReLU
-{
-public:
-    PReLU_mips();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PRELU_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/quantize_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/quantize_mips.h
deleted file mode 100644
index 220d73af..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/quantize_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_QUANTIZE_MIPS_H
-#define LAYER_QUANTIZE_MIPS_H
-
-#include "quantize.h"
-
-namespace ncnn {
-
-class Quantize_mips : public Quantize
-{
-public:
-    Quantize_mips();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_QUANTIZE_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/relu_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/relu_mips.h
deleted file mode 100644
index 74e55a6b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/relu_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_RELU_MIPS_H
-#define LAYER_RELU_MIPS_H
-
-#include "relu.h"
-
-namespace ncnn {
-
-class ReLU_mips : public ReLU
-{
-public:
-    ReLU_mips();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_RELU_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_leakyrelu_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_leakyrelu_pack4.h
deleted file mode 100644
index 89bc14bd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_leakyrelu_pack4.h
+++ /dev/null
@@ -1,267 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_leakyrelu_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, float slope, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-    int size = w * h;
-    int outc = top_blob.c;
-    int out_elempack = top_blob.elempack;
-
-    int scale_in_data_size = scale_in_data.w;
-    int scale_out_data_size = scale_out_data.w;
-    int bias_data_size = bias_data.w;
-
-    // int8(leakyrelu(v * scale_in, slope) * scale_out)
-    // int8_leakyrelu(v * (scale_in * scale_out), slope)
-
-    // int8(leakyrelu(v * scale_in + bias, slope) * scale_out)
-    // int8_leakyrelu(v * (scale_in * scale_out) + (bias * scale_out), slope)
-
-    if (out_elempack == 8)
-    {
-        if (bias_data_size == 0)
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < outc; q++)
-            {
-                const int* intptr0 = bottom_blob.channel(q * 2);
-                const int* intptr1 = bottom_blob.channel(q * 2 + 1);
-                signed char* ptr = top_blob.channel(q);
-
-                v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
-                v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
-                v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
-                v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
-
-                v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
-                v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
-                v4f32 _slope = (v4f32)__msa_fill_w_f32(slope);
-
-                int i = 0;
-                for (; i + 3 < size; i += 4)
-                {
-                    __builtin_prefetch(intptr0 + 64);
-                    __builtin_prefetch(intptr1 + 64);
-                    v4f32 _v00 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
-                    v4f32 _v01 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 4, 0));
-                    v4f32 _v02 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 8, 0));
-                    v4f32 _v03 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 12, 0));
-                    v4f32 _v10 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
-                    v4f32 _v11 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 4, 0));
-                    v4f32 _v12 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 8, 0));
-                    v4f32 _v13 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 12, 0));
-                    _v00 = __msa_fmul_w(_v00, _scale0);
-                    _v01 = __msa_fmul_w(_v01, _scale0);
-                    _v02 = __msa_fmul_w(_v02, _scale0);
-                    _v03 = __msa_fmul_w(_v03, _scale0);
-                    _v10 = __msa_fmul_w(_v10, _scale1);
-                    _v11 = __msa_fmul_w(_v11, _scale1);
-                    _v12 = __msa_fmul_w(_v12, _scale1);
-                    _v13 = __msa_fmul_w(_v13, _scale1);
-                    *((int64_t*)ptr) = float2int8leakyrelu(_v00, _v10, _slope);
-                    *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v01, _v11, _slope);
-                    *((int64_t*)(ptr + 16)) = float2int8leakyrelu(_v02, _v12, _slope);
-                    *((int64_t*)(ptr + 24)) = float2int8leakyrelu(_v03, _v13, _slope);
-
-                    intptr0 += 16;
-                    intptr1 += 16;
-                    ptr += 32;
-                }
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr0 + 16);
-                    __builtin_prefetch(intptr1 + 16);
-                    v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
-                    v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
-                    _v0 = __msa_fmul_w(_v0, _scale0);
-                    _v1 = __msa_fmul_w(_v1, _scale1);
-                    *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-
-                    intptr0 += 4;
-                    intptr1 += 4;
-                    ptr += 8;
-                }
-            }
-        }
-        else
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < outc; q++)
-            {
-                const int* intptr0 = bottom_blob.channel(q * 2);
-                const int* intptr1 = bottom_blob.channel(q * 2 + 1);
-                signed char* ptr = top_blob.channel(q);
-
-                v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
-                v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
-                v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
-                v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
-                v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8, 0);
-                v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8 + 4, 0);
-
-                v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
-                v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
-                _bias0 = __msa_fmul_w(_bias0, _scale_out0);
-                _bias1 = __msa_fmul_w(_bias1, _scale_out1);
-                v4f32 _slope = (v4f32)__msa_fill_w_f32(slope);
-
-                int i = 0;
-                for (; i + 3 < size; i += 4)
-                {
-                    __builtin_prefetch(intptr0 + 64);
-                    __builtin_prefetch(intptr1 + 64);
-                    v4f32 _v00 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
-                    v4f32 _v01 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 4, 0));
-                    v4f32 _v02 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 8, 0));
-                    v4f32 _v03 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 12, 0));
-                    v4f32 _v10 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
-                    v4f32 _v11 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 4, 0));
-                    v4f32 _v12 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 8, 0));
-                    v4f32 _v13 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 12, 0));
-                    _v00 = __msa_fmadd_w(_bias0, _v00, _scale0);
-                    _v01 = __msa_fmadd_w(_bias0, _v01, _scale0);
-                    _v02 = __msa_fmadd_w(_bias0, _v02, _scale0);
-                    _v03 = __msa_fmadd_w(_bias0, _v03, _scale0);
-                    _v10 = __msa_fmadd_w(_bias1, _v10, _scale1);
-                    _v11 = __msa_fmadd_w(_bias1, _v11, _scale1);
-                    _v12 = __msa_fmadd_w(_bias1, _v12, _scale1);
-                    _v13 = __msa_fmadd_w(_bias1, _v13, _scale1);
-                    *((int64_t*)ptr) = float2int8leakyrelu(_v00, _v10, _slope);
-                    *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v01, _v11, _slope);
-                    *((int64_t*)(ptr + 16)) = float2int8leakyrelu(_v02, _v12, _slope);
-                    *((int64_t*)(ptr + 24)) = float2int8leakyrelu(_v03, _v13, _slope);
-
-                    intptr0 += 16;
-                    intptr1 += 16;
-                    ptr += 32;
-                }
-                for (; i + 1 < size; i += 2)
-                {
-                    __builtin_prefetch(intptr0 + 32);
-                    __builtin_prefetch(intptr1 + 32);
-                    v4f32 _v00 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
-                    v4f32 _v01 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 4, 0));
-                    v4f32 _v10 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
-                    v4f32 _v11 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 4, 0));
-                    _v00 = __msa_fmadd_w(_bias0, _v00, _scale0);
-                    _v01 = __msa_fmadd_w(_bias0, _v01, _scale0);
-                    _v10 = __msa_fmadd_w(_bias1, _v10, _scale1);
-                    _v11 = __msa_fmadd_w(_bias1, _v11, _scale1);
-                    *((int64_t*)ptr) = float2int8leakyrelu(_v00, _v10, _slope);
-                    *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v01, _v11, _slope);
-
-                    intptr0 += 8;
-                    intptr1 += 8;
-                    ptr += 16;
-                }
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr0 + 16);
-                    __builtin_prefetch(intptr1 + 16);
-                    v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
-                    v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
-                    _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
-                    _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
-                    *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-
-                    intptr0 += 4;
-                    intptr1 += 4;
-                    ptr += 8;
-                }
-            }
-        }
-    }
-    if (out_elempack == 1)
-    {
-        if (bias_data_size == 0)
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < channels; q++)
-            {
-                const int* intptr = bottom_blob.channel(q);
-                signed char* ptr0 = top_blob.channel(q * 4);
-                signed char* ptr1 = top_blob.channel(q * 4 + 1);
-                signed char* ptr2 = top_blob.channel(q * 4 + 2);
-                signed char* ptr3 = top_blob.channel(q * 4 + 3);
-
-                v4f32 _scale_in = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 4, 0);
-                v4f32 _scale_out = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 4, 0);
-
-                v4f32 _scale = __msa_fmul_w(_scale_in, _scale_out);
-                v4f32 _slope = (v4f32)__msa_fill_w_f32(slope);
-
-                int i = 0;
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr + 16);
-                    v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                    _v = __msa_fmul_w(_v, _scale);
-                    v16i8 v = float2int8leakyrelu(_v, _slope);
-                    ptr0[0] = v[0];
-                    ptr1[0] = v[1];
-                    ptr2[0] = v[2];
-                    ptr3[0] = v[3];
-
-                    intptr += 4;
-                    ptr0 += 1;
-                    ptr1 += 1;
-                    ptr2 += 1;
-                    ptr3 += 1;
-                }
-            }
-        }
-        else
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < channels; q++)
-            {
-                const int* intptr = bottom_blob.channel(q);
-                signed char* ptr0 = top_blob.channel(q * 4);
-                signed char* ptr1 = top_blob.channel(q * 4 + 1);
-                signed char* ptr2 = top_blob.channel(q * 4 + 2);
-                signed char* ptr3 = top_blob.channel(q * 4 + 3);
-
-                v4f32 _scale_in = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 4, 0);
-                v4f32 _scale_out = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 4, 0);
-                v4f32 _bias = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 4, 0);
-
-                v4f32 _scale = __msa_fmul_w(_scale_in, _scale_out);
-                _bias = __msa_fmul_w(_bias, _scale_out);
-                v4f32 _slope = (v4f32)__msa_fill_w_f32(slope);
-
-                int i = 0;
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr + 16);
-                    v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                    _v = __msa_fmadd_w(_bias, _v, _scale);
-                    v16i8 v = float2int8leakyrelu(_v, _slope);
-                    ptr0[0] = v[0];
-                    ptr1[0] = v[1];
-                    ptr2[0] = v[2];
-                    ptr3[0] = v[3];
-
-                    intptr += 4;
-                    ptr0 += 1;
-                    ptr1 += 1;
-                    ptr2 += 1;
-                    ptr3 += 1;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_leakyrelu_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_leakyrelu_pack8.h
deleted file mode 100644
index f7968c9d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_leakyrelu_pack8.h
+++ /dev/null
@@ -1,188 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_leakyrelu_pack8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, float slope, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-    int size = w * h;
-
-    int scale_in_data_size = scale_in_data.w;
-    int scale_out_data_size = scale_out_data.w;
-    int bias_data_size = bias_data.w;
-
-    // int8(leakyrelu(v * scale_in, slope) * scale_out)
-    // int8_leakyrelu(v * (scale_in * scale_out), slope)
-
-    // int8(leakyrelu(v * scale_in + bias, slope) * scale_out)
-    // int8_leakyrelu(v * (scale_in * scale_out) + (bias * scale_out), slope)
-
-    if (bias_data_size == 0)
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int q = 0; q < channels; q++)
-        {
-            const int* intptr = bottom_blob.channel(q);
-            signed char* ptr = top_blob.channel(q);
-
-            v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
-            v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
-            v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
-            v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
-
-            v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
-            v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
-            v4f32 _slope = (v4f32)__msa_fill_w_f32(slope);
-
-            int i = 0;
-            for (; i + 3 < size; i += 4)
-            {
-                __builtin_prefetch(intptr + 128);
-                v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
-                v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
-                v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
-                v4f32 _v4 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 16, 0));
-                v4f32 _v5 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 20, 0));
-                v4f32 _v6 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 24, 0));
-                v4f32 _v7 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 28, 0));
-                _v0 = __msa_fmul_w(_v0, _scale0);
-                _v1 = __msa_fmul_w(_v1, _scale1);
-                _v2 = __msa_fmul_w(_v2, _scale0);
-                _v3 = __msa_fmul_w(_v3, _scale1);
-                _v4 = __msa_fmul_w(_v4, _scale0);
-                _v5 = __msa_fmul_w(_v5, _scale1);
-                _v6 = __msa_fmul_w(_v6, _scale0);
-                _v7 = __msa_fmul_w(_v7, _scale1);
-                *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-                *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v2, _v3, _slope);
-                *((int64_t*)(ptr + 16)) = float2int8leakyrelu(_v4, _v5, _slope);
-                *((int64_t*)(ptr + 24)) = float2int8leakyrelu(_v6, _v7, _slope);
-
-                intptr += 32;
-                ptr += 32;
-            }
-            for (; i + 1 < size; i += 2)
-            {
-                __builtin_prefetch(intptr + 64);
-                v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
-                v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
-                v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
-                _v0 = __msa_fmul_w(_v0, _scale0);
-                _v1 = __msa_fmul_w(_v1, _scale1);
-                _v2 = __msa_fmul_w(_v2, _scale0);
-                _v3 = __msa_fmul_w(_v3, _scale1);
-                *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-                *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v2, _v3, _slope);
-
-                intptr += 16;
-                ptr += 16;
-            }
-            for (; i < size; i++)
-            {
-                __builtin_prefetch(intptr + 32);
-                v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
-                _v0 = __msa_fmul_w(_v0, _scale0);
-                _v1 = __msa_fmul_w(_v1, _scale1);
-                *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-
-                intptr += 8;
-                ptr += 8;
-            }
-        }
-    }
-    else
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int q = 0; q < channels; q++)
-        {
-            const int* intptr = bottom_blob.channel(q);
-            signed char* ptr = top_blob.channel(q);
-
-            v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
-            v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
-            v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
-            v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
-            v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8, 0);
-            v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8 + 4, 0);
-
-            v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
-            v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
-            _bias0 = __msa_fmul_w(_bias0, _scale_out0);
-            _bias1 = __msa_fmul_w(_bias1, _scale_out1);
-            v4f32 _slope = (v4f32)__msa_fill_w_f32(slope);
-
-            int i = 0;
-            for (; i + 3 < size; i += 4)
-            {
-                __builtin_prefetch(intptr + 128);
-                v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
-                v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
-                v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
-                v4f32 _v4 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 16, 0));
-                v4f32 _v5 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 20, 0));
-                v4f32 _v6 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 24, 0));
-                v4f32 _v7 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 28, 0));
-                _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
-                _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
-                _v2 = __msa_fmadd_w(_bias0, _v2, _scale0);
-                _v3 = __msa_fmadd_w(_bias1, _v3, _scale1);
-                _v4 = __msa_fmadd_w(_bias0, _v4, _scale0);
-                _v5 = __msa_fmadd_w(_bias1, _v5, _scale1);
-                _v6 = __msa_fmadd_w(_bias0, _v6, _scale0);
-                _v7 = __msa_fmadd_w(_bias1, _v7, _scale1);
-                *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-                *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v2, _v3, _slope);
-                *((int64_t*)(ptr + 16)) = float2int8leakyrelu(_v4, _v5, _slope);
-                *((int64_t*)(ptr + 24)) = float2int8leakyrelu(_v6, _v7, _slope);
-
-                intptr += 32;
-                ptr += 32;
-            }
-            for (; i + 1 < size; i += 2)
-            {
-                __builtin_prefetch(intptr + 64);
-                v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
-                v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
-                v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
-                _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
-                _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
-                _v2 = __msa_fmadd_w(_bias0, _v2, _scale0);
-                _v3 = __msa_fmadd_w(_bias1, _v3, _scale1);
-                *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-                *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v2, _v3, _slope);
-
-                intptr += 16;
-                ptr += 16;
-            }
-            for (; i < size; i++)
-            {
-                __builtin_prefetch(intptr + 32);
-                v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
-                _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
-                _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
-                *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-
-                intptr += 8;
-                ptr += 8;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_mips.h
deleted file mode 100644
index 6ba74089..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_REQUANTIZE_MIPS_H
-#define LAYER_REQUANTIZE_MIPS_H
-
-#include "requantize.h"
-
-namespace ncnn {
-
-class Requantize_mips : public Requantize
-{
-public:
-    Requantize_mips();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_REQUANTIZE_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_relu_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_relu_pack4.h
deleted file mode 100644
index e43797bd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_relu_pack4.h
+++ /dev/null
@@ -1,263 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_relu_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-    int size = w * h;
-    int outc = top_blob.c;
-    int out_elempack = top_blob.elempack;
-
-    int scale_in_data_size = scale_in_data.w;
-    int scale_out_data_size = scale_out_data.w;
-    int bias_data_size = bias_data.w;
-
-    // int8(relu(v * scale_in) * scale_out)
-    // int8_relu(v * (scale_in * scale_out))
-
-    // int8(relu(v * scale_in + bias) * scale_out)
-    // int8_relu(v * (scale_in * scale_out) + (bias * scale_out))
-
-    if (out_elempack == 8)
-    {
-        if (bias_data_size == 0)
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < outc; q++)
-            {
-                const int* intptr0 = bottom_blob.channel(q * 2);
-                const int* intptr1 = bottom_blob.channel(q * 2 + 1);
-                signed char* ptr = top_blob.channel(q);
-
-                v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
-                v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
-                v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
-                v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
-
-                v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
-                v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
-
-                int i = 0;
-                for (; i + 3 < size; i += 4)
-                {
-                    __builtin_prefetch(intptr0 + 64);
-                    __builtin_prefetch(intptr1 + 64);
-                    v4f32 _v00 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
-                    v4f32 _v01 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 4, 0));
-                    v4f32 _v02 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 8, 0));
-                    v4f32 _v03 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 12, 0));
-                    v4f32 _v10 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
-                    v4f32 _v11 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 4, 0));
-                    v4f32 _v12 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 8, 0));
-                    v4f32 _v13 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 12, 0));
-                    _v00 = __msa_fmul_w(_v00, _scale0);
-                    _v01 = __msa_fmul_w(_v01, _scale0);
-                    _v02 = __msa_fmul_w(_v02, _scale0);
-                    _v03 = __msa_fmul_w(_v03, _scale0);
-                    _v10 = __msa_fmul_w(_v10, _scale1);
-                    _v11 = __msa_fmul_w(_v11, _scale1);
-                    _v12 = __msa_fmul_w(_v12, _scale1);
-                    _v13 = __msa_fmul_w(_v13, _scale1);
-                    *((int64_t*)ptr) = float2int8relu(_v00, _v10);
-                    *((int64_t*)(ptr + 8)) = float2int8relu(_v01, _v11);
-                    *((int64_t*)(ptr + 16)) = float2int8relu(_v02, _v12);
-                    *((int64_t*)(ptr + 24)) = float2int8relu(_v03, _v13);
-
-                    intptr0 += 16;
-                    intptr1 += 16;
-                    ptr += 32;
-                }
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr0 + 16);
-                    __builtin_prefetch(intptr1 + 16);
-                    v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
-                    v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
-                    _v0 = __msa_fmul_w(_v0, _scale0);
-                    _v1 = __msa_fmul_w(_v1, _scale1);
-                    *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-
-                    intptr0 += 4;
-                    intptr1 += 4;
-                    ptr += 8;
-                }
-            }
-        }
-        else
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < outc; q++)
-            {
-                const int* intptr0 = bottom_blob.channel(q * 2);
-                const int* intptr1 = bottom_blob.channel(q * 2 + 1);
-                signed char* ptr = top_blob.channel(q);
-
-                v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
-                v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
-                v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
-                v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
-                v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8, 0);
-                v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8 + 4, 0);
-
-                v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
-                v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
-                _bias0 = __msa_fmul_w(_bias0, _scale_out0);
-                _bias1 = __msa_fmul_w(_bias1, _scale_out1);
-
-                int i = 0;
-                for (; i + 3 < size; i += 4)
-                {
-                    __builtin_prefetch(intptr0 + 64);
-                    __builtin_prefetch(intptr1 + 64);
-                    v4f32 _v00 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
-                    v4f32 _v01 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 4, 0));
-                    v4f32 _v02 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 8, 0));
-                    v4f32 _v03 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 12, 0));
-                    v4f32 _v10 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
-                    v4f32 _v11 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 4, 0));
-                    v4f32 _v12 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 8, 0));
-                    v4f32 _v13 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 12, 0));
-                    _v00 = __msa_fmadd_w(_bias0, _v00, _scale0);
-                    _v01 = __msa_fmadd_w(_bias0, _v01, _scale0);
-                    _v02 = __msa_fmadd_w(_bias0, _v02, _scale0);
-                    _v03 = __msa_fmadd_w(_bias0, _v03, _scale0);
-                    _v10 = __msa_fmadd_w(_bias1, _v10, _scale1);
-                    _v11 = __msa_fmadd_w(_bias1, _v11, _scale1);
-                    _v12 = __msa_fmadd_w(_bias1, _v12, _scale1);
-                    _v13 = __msa_fmadd_w(_bias1, _v13, _scale1);
-                    *((int64_t*)ptr) = float2int8relu(_v00, _v10);
-                    *((int64_t*)(ptr + 8)) = float2int8relu(_v01, _v11);
-                    *((int64_t*)(ptr + 16)) = float2int8relu(_v02, _v12);
-                    *((int64_t*)(ptr + 24)) = float2int8relu(_v03, _v13);
-
-                    intptr0 += 16;
-                    intptr1 += 16;
-                    ptr += 32;
-                }
-                for (; i + 1 < size; i += 2)
-                {
-                    __builtin_prefetch(intptr0 + 32);
-                    __builtin_prefetch(intptr1 + 32);
-                    v4f32 _v00 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
-                    v4f32 _v01 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 4, 0));
-                    v4f32 _v10 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
-                    v4f32 _v11 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 4, 0));
-                    _v00 = __msa_fmadd_w(_bias0, _v00, _scale0);
-                    _v01 = __msa_fmadd_w(_bias0, _v01, _scale0);
-                    _v10 = __msa_fmadd_w(_bias1, _v10, _scale1);
-                    _v11 = __msa_fmadd_w(_bias1, _v11, _scale1);
-                    *((int64_t*)ptr) = float2int8relu(_v00, _v10);
-                    *((int64_t*)(ptr + 8)) = float2int8relu(_v01, _v11);
-
-                    intptr0 += 8;
-                    intptr1 += 8;
-                    ptr += 16;
-                }
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr0 + 16);
-                    __builtin_prefetch(intptr1 + 16);
-                    v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
-                    v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
-                    _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
-                    _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
-                    *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-
-                    intptr0 += 4;
-                    intptr1 += 4;
-                    ptr += 8;
-                }
-            }
-        }
-    }
-    if (out_elempack == 1)
-    {
-        if (bias_data_size == 0)
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < channels; q++)
-            {
-                const int* intptr = bottom_blob.channel(q);
-                signed char* ptr0 = top_blob.channel(q * 4);
-                signed char* ptr1 = top_blob.channel(q * 4 + 1);
-                signed char* ptr2 = top_blob.channel(q * 4 + 2);
-                signed char* ptr3 = top_blob.channel(q * 4 + 3);
-
-                v4f32 _scale_in = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 4, 0);
-                v4f32 _scale_out = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 4, 0);
-
-                v4f32 _scale = __msa_fmul_w(_scale_in, _scale_out);
-
-                int i = 0;
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr + 16);
-                    v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                    _v = __msa_fmul_w(_v, _scale);
-                    v16i8 v = float2int8relu(_v);
-                    ptr0[0] = v[0];
-                    ptr1[0] = v[1];
-                    ptr2[0] = v[2];
-                    ptr3[0] = v[3];
-
-                    intptr += 4;
-                    ptr0 += 1;
-                    ptr1 += 1;
-                    ptr2 += 1;
-                    ptr3 += 1;
-                }
-            }
-        }
-        else
-        {
-            #pragma omp parallel for num_threads(opt.num_threads)
-            for (int q = 0; q < channels; q++)
-            {
-                const int* intptr = bottom_blob.channel(q);
-                signed char* ptr0 = top_blob.channel(q * 4);
-                signed char* ptr1 = top_blob.channel(q * 4 + 1);
-                signed char* ptr2 = top_blob.channel(q * 4 + 2);
-                signed char* ptr3 = top_blob.channel(q * 4 + 3);
-
-                v4f32 _scale_in = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 4, 0);
-                v4f32 _scale_out = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 4, 0);
-                v4f32 _bias = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 4, 0);
-
-                v4f32 _scale = __msa_fmul_w(_scale_in, _scale_out);
-                _bias = __msa_fmul_w(_bias, _scale_out);
-
-                int i = 0;
-                for (; i < size; i++)
-                {
-                    __builtin_prefetch(intptr + 16);
-                    v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                    _v = __msa_fmadd_w(_bias, _v, _scale);
-                    v16i8 v = float2int8relu(_v);
-                    ptr0[0] = v[0];
-                    ptr1[0] = v[1];
-                    ptr2[0] = v[2];
-                    ptr3[0] = v[3];
-
-                    intptr += 4;
-                    ptr0 += 1;
-                    ptr1 += 1;
-                    ptr2 += 1;
-                    ptr3 += 1;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_relu_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_relu_pack8.h
deleted file mode 100644
index 824b668c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/requantize_relu_pack8.h
+++ /dev/null
@@ -1,186 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_relu_pack8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-    int size = w * h;
-
-    int scale_in_data_size = scale_in_data.w;
-    int scale_out_data_size = scale_out_data.w;
-    int bias_data_size = bias_data.w;
-
-    // int8(relu(v * scale_in) * scale_out)
-    // int8_relu(v * (scale_in * scale_out))
-
-    // int8(relu(v * scale_in + bias) * scale_out)
-    // int8_relu(v * (scale_in * scale_out) + (bias * scale_out))
-
-    if (bias_data_size == 0)
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int q = 0; q < channels; q++)
-        {
-            const int* intptr = bottom_blob.channel(q);
-            signed char* ptr = top_blob.channel(q);
-
-            v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
-            v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
-            v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
-            v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
-
-            v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
-            v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
-
-            int i = 0;
-            for (; i + 3 < size; i += 4)
-            {
-                __builtin_prefetch(intptr + 128);
-                v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
-                v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
-                v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
-                v4f32 _v4 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 16, 0));
-                v4f32 _v5 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 20, 0));
-                v4f32 _v6 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 24, 0));
-                v4f32 _v7 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 28, 0));
-                _v0 = __msa_fmul_w(_v0, _scale0);
-                _v1 = __msa_fmul_w(_v1, _scale1);
-                _v2 = __msa_fmul_w(_v2, _scale0);
-                _v3 = __msa_fmul_w(_v3, _scale1);
-                _v4 = __msa_fmul_w(_v4, _scale0);
-                _v5 = __msa_fmul_w(_v5, _scale1);
-                _v6 = __msa_fmul_w(_v6, _scale0);
-                _v7 = __msa_fmul_w(_v7, _scale1);
-                *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-                *((int64_t*)(ptr + 8)) = float2int8relu(_v2, _v3);
-                *((int64_t*)(ptr + 16)) = float2int8relu(_v4, _v5);
-                *((int64_t*)(ptr + 24)) = float2int8relu(_v6, _v7);
-
-                intptr += 32;
-                ptr += 32;
-            }
-            for (; i + 1 < size; i += 2)
-            {
-                __builtin_prefetch(intptr + 64);
-                v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
-                v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
-                v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
-                _v0 = __msa_fmul_w(_v0, _scale0);
-                _v1 = __msa_fmul_w(_v1, _scale1);
-                _v2 = __msa_fmul_w(_v2, _scale0);
-                _v3 = __msa_fmul_w(_v3, _scale1);
-                *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-                *((int64_t*)(ptr + 8)) = float2int8relu(_v2, _v3);
-
-                intptr += 16;
-                ptr += 16;
-            }
-            for (; i < size; i++)
-            {
-                __builtin_prefetch(intptr + 32);
-                v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
-                _v0 = __msa_fmul_w(_v0, _scale0);
-                _v1 = __msa_fmul_w(_v1, _scale1);
-                *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-
-                intptr += 8;
-                ptr += 8;
-            }
-        }
-    }
-    else
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int q = 0; q < channels; q++)
-        {
-            const int* intptr = bottom_blob.channel(q);
-            signed char* ptr = top_blob.channel(q);
-
-            v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
-            v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
-            v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
-            v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
-            v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8, 0);
-            v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8 + 4, 0);
-
-            v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
-            v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
-            _bias0 = __msa_fmul_w(_bias0, _scale_out0);
-            _bias1 = __msa_fmul_w(_bias1, _scale_out1);
-
-            int i = 0;
-            for (; i + 3 < size; i += 4)
-            {
-                __builtin_prefetch(intptr + 128);
-                v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
-                v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
-                v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
-                v4f32 _v4 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 16, 0));
-                v4f32 _v5 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 20, 0));
-                v4f32 _v6 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 24, 0));
-                v4f32 _v7 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 28, 0));
-                _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
-                _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
-                _v2 = __msa_fmadd_w(_bias0, _v2, _scale0);
-                _v3 = __msa_fmadd_w(_bias1, _v3, _scale1);
-                _v4 = __msa_fmadd_w(_bias0, _v4, _scale0);
-                _v5 = __msa_fmadd_w(_bias1, _v5, _scale1);
-                _v6 = __msa_fmadd_w(_bias0, _v6, _scale0);
-                _v7 = __msa_fmadd_w(_bias1, _v7, _scale1);
-                *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-                *((int64_t*)(ptr + 8)) = float2int8relu(_v2, _v3);
-                *((int64_t*)(ptr + 16)) = float2int8relu(_v4, _v5);
-                *((int64_t*)(ptr + 24)) = float2int8relu(_v6, _v7);
-
-                intptr += 32;
-                ptr += 32;
-            }
-            for (; i + 1 < size; i += 2)
-            {
-                __builtin_prefetch(intptr + 64);
-                v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
-                v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
-                v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
-                _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
-                _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
-                _v2 = __msa_fmadd_w(_bias0, _v2, _scale0);
-                _v3 = __msa_fmadd_w(_bias1, _v3, _scale1);
-                *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-                *((int64_t*)(ptr + 8)) = float2int8relu(_v2, _v3);
-
-                intptr += 16;
-                ptr += 16;
-            }
-            for (; i < size; i++)
-            {
-                __builtin_prefetch(intptr + 32);
-                v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
-                v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
-                _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
-                _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
-                *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-
-                intptr += 8;
-                ptr += 8;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/sigmoid_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/sigmoid_mips.h
deleted file mode 100644
index 2bf166e9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/sigmoid_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Leo is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 Leo <leo@nullptr.com.cn>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SIGMOID_MIPS_H
-#define LAYER_SIGMOID_MIPS_H
-
-#include "sigmoid.h"
-
-namespace ncnn {
-
-class Sigmoid_mips : public Sigmoid
-{
-public:
-    Sigmoid_mips();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SIGMOID_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/slice_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/slice_mips.h
deleted file mode 100644
index 73274d86..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/slice_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SLICE_MIPS_H
-#define LAYER_SLICE_MIPS_H
-
-#include "slice.h"
-
-namespace ncnn {
-
-class Slice_mips : public Slice
-{
-public:
-    Slice_mips();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SLICE_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/softmax_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/softmax_mips.h
deleted file mode 100644
index 91437c13..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/softmax_mips.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// Leo is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 Leo <leo@nullptr.com.cn>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SOFTMAX_MIPS_H
-#define LAYER_SOFTMAX_MIPS_H
-
-#include "softmax.h"
-
-namespace ncnn {
-
-class Softmax_mips : public Softmax
-{
-public:
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SOFTMAX_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/swish_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/swish_mips.h
deleted file mode 100644
index 1dc6753a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/swish_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SWISH_MIPS_H
-#define LAYER_SWISH_MIPS_H
-
-#include "swish.h"
-
-namespace ncnn {
-
-class Swish_mips : public Swish
-{
-public:
-    Swish_mips();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SWISH_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/tanh_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/tanh_mips.h
deleted file mode 100644
index 12e38d07..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/tanh_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Leo is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 Leo <leo@nullptr.com.cn>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_TANH_MIPS_H
-#define LAYER_TANH_MIPS_H
-
-#include "tanh.h"
-
-namespace ncnn {
-
-class TanH_mips : public TanH
-{
-public:
-    TanH_mips();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_TANH_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/unaryop_mips.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/unaryop_mips.h
deleted file mode 100644
index 800d028b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mips/unaryop_mips.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_UNARYOP_MIPS_H
-#define LAYER_UNARYOP_MIPS_H
-
-#include "unaryop.h"
-
-namespace ncnn {
-
-class UnaryOp_mips : public UnaryOp
-{
-public:
-    UnaryOp_mips();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_UNARYOP_MIPS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mish.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mish.h
deleted file mode 100644
index e35b8ceb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mish.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MISH_H
-#define LAYER_MISH_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Mish : public Layer
-{
-public:
-    Mish();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MISH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/multiheadattention.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/multiheadattention.h
deleted file mode 100644
index 50c8549a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/multiheadattention.h
+++ /dev/null
@@ -1,53 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MULTIHEADATTENTION_H
-#define LAYER_MULTIHEADATTENTION_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class MultiHeadAttention : public Layer
-{
-public:
-    MultiHeadAttention();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    int embed_dim;
-    int num_heads;
-    int weight_data_size;
-    int kdim;
-    int vdim;
-    int attn_mask;
-
-    Mat q_weight_data;
-    Mat q_bias_data;
-    Mat k_weight_data;
-    Mat k_bias_data;
-    Mat v_weight_data;
-    Mat v_bias_data;
-    Mat out_weight_data;
-    Mat out_bias_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MULTIHEADATTENTION_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mvn.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mvn.h
deleted file mode 100644
index 90fb2460..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/mvn.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MVN_H
-#define LAYER_MVN_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class MVN : public Layer
-{
-public:
-    MVN();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int normalize_variance;
-    int across_channels;
-    float eps;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MVN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/noop.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/noop.h
deleted file mode 100644
index 75bbdd1a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/noop.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_NOOP_H
-#define LAYER_NOOP_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Noop : public Layer
-{
-public:
-    Noop();
-
-    virtual int forward_inplace(std::vector<Mat>& bottom_top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_NOOP_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/normalize.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/normalize.h
deleted file mode 100644
index c4ff6964..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/normalize.h
+++ /dev/null
@@ -1,51 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_NORMALIZE_H
-#define LAYER_NORMALIZE_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Normalize : public Layer
-{
-public:
-    Normalize();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    // param
-    int across_spatial;
-    int across_channel;
-    int channel_shared;
-    float eps;
-    int scale_data_size;
-
-    // 0 = v / sqrt(v2 + eps) caffe/mxnet
-    // 1 = v / max(sqrt(v2), eps) pytorch
-    // 2 = v / sqrt(max(v2, eps)) tensorflow
-    int eps_mode;
-
-    Mat scale_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_NORMALIZE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/packing.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/packing.h
deleted file mode 100644
index ad54f907..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/packing.h
+++ /dev/null
@@ -1,52 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PACKING_H
-#define LAYER_PACKING_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Packing : public Layer
-{
-public:
-    Packing();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int out_elempack;
-    int use_padding;
-
-    // element type
-    // 0 = auto
-    // 1 = fp32
-    // 2 = fp16p
-    // 3 = fp16s
-    int cast_type_from;
-    int cast_type_to;
-
-    // storage type
-    // 0 = buffer
-    // 1 = image
-    int storage_type_from;
-    int storage_type_to;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PACKING_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/padding.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/padding.h
deleted file mode 100644
index 36c5be14..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/padding.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PADDING_H
-#define LAYER_PADDING_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Padding : public Layer
-{
-public:
-    Padding();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int top;
-    int bottom;
-    int left;
-    int right;
-    int type; // 0=CONSTANT 1=REPLICATE 2=REFLECT
-    float value;
-    int front;
-    int behind;
-
-    // per channel pad value
-    int per_channel_pad_data_size;
-    Mat per_channel_pad_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PADDING_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/permute.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/permute.h
deleted file mode 100644
index 963aa5f2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/permute.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PERMUTE_H
-#define LAYER_PERMUTE_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Permute : public Layer
-{
-public:
-    Permute();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int order_type;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PERMUTE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pixelshuffle.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pixelshuffle.h
deleted file mode 100644
index 728dc5f0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pixelshuffle.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PIXELSHUFFLE_H
-#define LAYER_PIXELSHUFFLE_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class PixelShuffle : public Layer
-{
-public:
-    PixelShuffle();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int upscale_factor;
-    int mode;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PIXELSHUFFLE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pooling.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pooling.h
deleted file mode 100644
index 919127d8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pooling.h
+++ /dev/null
@@ -1,61 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_POOLING_H
-#define LAYER_POOLING_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Pooling : public Layer
-{
-public:
-    Pooling();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    enum PoolMethod
-    {
-        PoolMethod_MAX = 0,
-        PoolMethod_AVE = 1
-    };
-
-protected:
-    void make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, const Option& opt) const;
-
-public:
-    // param
-    int pooling_type;
-    int kernel_w;
-    int kernel_h;
-    int stride_w;
-    int stride_h;
-    int pad_left;
-    int pad_right;
-    int pad_top;
-    int pad_bottom;
-    int global_pooling;
-    int pad_mode; // 0=full 1=valid 2=SAME_UPPER 3=SAME_LOWER
-    int avgpool_count_include_pad;
-    int adaptive_pooling;
-    int out_w;
-    int out_h;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_POOLING_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pooling1d.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pooling1d.h
deleted file mode 100644
index 38a2e659..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pooling1d.h
+++ /dev/null
@@ -1,56 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_POOLING1D_H
-#define LAYER_POOLING1D_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Pooling1D : public Layer
-{
-public:
-    Pooling1D();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    enum PoolMethod
-    {
-        PoolMethod_MAX = 0,
-        PoolMethod_AVE = 1
-    };
-
-protected:
-    void make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, const Option& opt) const;
-
-public:
-    // param
-    int pooling_type;
-    int kernel_w;
-    int stride_w;
-    int pad_left;
-    int pad_right;
-    int global_pooling;
-    int pad_mode; // 0=full 1=valid 2=SAME_UPPER 3=SAME_LOWER
-    int avgpool_count_include_pad;
-    int adaptive_pooling;
-    int out_w;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_POOLING1D_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pooling3d.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pooling3d.h
deleted file mode 100644
index 9e8bcf49..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/pooling3d.h
+++ /dev/null
@@ -1,66 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_POOLING3D_H
-#define LAYER_POOLING3D_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Pooling3D : public Layer
-{
-public:
-    Pooling3D();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    enum PoolMethod
-    {
-        PoolMethod_MAX = 0,
-        PoolMethod_AVE = 1
-    };
-
-protected:
-    void make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, const Option& opt) const;
-
-public:
-    // param
-    int pooling_type;
-    int kernel_w;
-    int kernel_h;
-    int kernel_d;
-    int stride_w;
-    int stride_h;
-    int stride_d;
-    int pad_left;
-    int pad_right;
-    int pad_top;
-    int pad_bottom;
-    int pad_front;
-    int pad_behind;
-    int global_pooling;
-    int pad_mode; // 0=full 1=valid 2=SAME_UPPER 3=SAME_LOWER
-    int avgpool_count_include_pad;
-    int adaptive_pooling;
-    int out_w;
-    int out_h;
-    int out_d;
-};
-
-} //namespace ncnn
-
-#endif // LAYER_POOLING3D_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/power.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/power.h
deleted file mode 100644
index 8a54cc63..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/power.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_POWER_H
-#define LAYER_POWER_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Power : public Layer
-{
-public:
-    Power();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    float power;
-    float scale;
-    float shift;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_POWER_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/prelu.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/prelu.h
deleted file mode 100644
index ca155d6e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/prelu.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PRELU_H
-#define LAYER_PRELU_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class PReLU : public Layer
-{
-public:
-    PReLU();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    int num_slope;
-    Mat slope_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PRELU_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/priorbox.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/priorbox.h
deleted file mode 100644
index 5e734e1b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/priorbox.h
+++ /dev/null
@@ -1,49 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PRIORBOX_H
-#define LAYER_PRIORBOX_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class PriorBox : public Layer
-{
-public:
-    PriorBox();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    Mat min_sizes;
-    Mat max_sizes;
-    Mat aspect_ratios;
-    float variances[4];
-    int flip;
-    int clip;
-    int image_width;
-    int image_height;
-    float step_width;
-    float step_height;
-    float offset;
-    bool step_mmdetection;
-    bool center_mmdetection;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PRIORBOX_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/proposal.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/proposal.h
deleted file mode 100644
index ac23ec09..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/proposal.h
+++ /dev/null
@@ -1,48 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PROPOSAL_H
-#define LAYER_PROPOSAL_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Proposal : public Layer
-{
-public:
-    Proposal();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    // param
-    int feat_stride;
-    int base_size;
-    int pre_nms_topN;
-    int after_nms_topN;
-    float nms_thresh;
-    int min_size;
-
-    Mat ratios;
-    Mat scales;
-
-    Mat anchors;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PROPOSAL_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/psroipooling.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/psroipooling.h
deleted file mode 100644
index 41ab8e7a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/psroipooling.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2018 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PSROIPOOLING_H
-#define LAYER_PSROIPOOLING_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class PSROIPooling : public Layer
-{
-public:
-    PSROIPooling();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    int pooled_width;
-    int pooled_height;
-    float spatial_scale;
-    int output_dim;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PSROIPOOLING_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/quantize.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/quantize.h
deleted file mode 100644
index bfafd628..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/quantize.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2018 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_QUANTIZE_H
-#define LAYER_QUANTIZE_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Quantize : public Layer
-{
-public:
-    Quantize();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int scale_data_size;
-    Mat scale_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_QUANTIZE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/reduction.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/reduction.h
deleted file mode 100644
index 209f6b7e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/reduction.h
+++ /dev/null
@@ -1,57 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_REDUCTION_H
-#define LAYER_REDUCTION_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Reduction : public Layer
-{
-public:
-    Reduction();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    enum ReductionOp
-    {
-        ReductionOp_SUM = 0,
-        ReductionOp_ASUM = 1,
-        ReductionOp_SUMSQ = 2,
-        ReductionOp_MEAN = 3,
-        ReductionOp_MAX = 4,
-        ReductionOp_MIN = 5,
-        ReductionOp_PROD = 6,
-        ReductionOp_L1 = 7,
-        ReductionOp_L2 = 8,
-        ReductionOp_LogSum = 9,
-        ReductionOp_LogSumExp = 10
-    };
-
-public:
-    // param
-    int operation;
-    int reduce_all;
-    float coeff;
-    Mat axes;
-    int keepdims;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_REDUCTION_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/relu.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/relu.h
deleted file mode 100644
index 9460bcdc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/relu.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_RELU_H
-#define LAYER_RELU_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class ReLU : public Layer
-{
-public:
-    ReLU();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    float slope;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_RELU_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/reorg.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/reorg.h
deleted file mode 100644
index deb30321..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/reorg.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2018 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_REORG_H
-#define LAYER_REORG_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Reorg : public Layer
-{
-public:
-    Reorg();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int stride;
-    int mode;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_REORG_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/requantize.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/requantize.h
deleted file mode 100644
index 4bf87bc9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/requantize.h
+++ /dev/null
@@ -1,57 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 BUG1989. All rights reserved.
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_REQUANTIZE_H
-#define LAYER_REQUANTIZE_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Requantize : public Layer
-{
-public:
-    Requantize();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int scale_in_data_size;
-    int scale_out_data_size;
-    int bias_data_size;
-
-    // 0=none 1=relu 2=leakyrelu 3=clip 4=sigmoid
-    int activation_type;
-    Mat activation_params;
-
-    Mat scale_in_data;
-    Mat scale_out_data;
-    Mat bias_data;
-
-    //     float scale_in;  // bottom_blob_scale * weight_scale
-    //     float scale_out; // top_blob_scale / (bottom_blob_scale * weight_scale)
-    //     int bias_term;
-    //     int bias_data_size;
-    //
-    //     Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_REQUANTIZE_H
\ No newline at end of file
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/reshape.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/reshape.h
deleted file mode 100644
index b1072a48..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/reshape.h
+++ /dev/null
@@ -1,49 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_RESHAPE_H
-#define LAYER_RESHAPE_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Reshape : public Layer
-{
-public:
-    Reshape();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    // reshape flag
-    // 0 = copy from bottom
-    // -1 = remaining
-    // -233 = drop this dim (default)
-    int w;
-    int h;
-    int d;
-    int c;
-
-    // flag permute chw->hwc or hw->wh before and after reshape
-    int permute;
-
-    int ndim;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_RESHAPE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/absval_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/absval_riscv.h
deleted file mode 100644
index 0d35c6b6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/absval_riscv.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Xavier Hsinyuan is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 Xavier Hsinyuan <thelastlinex@hotmail.com>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ABSVAL_RISCV_H
-#define LAYER_ABSVAL_RISCV_H
-
-#include "absval.h"
-
-namespace ncnn {
-
-class AbsVal_riscv : public AbsVal
-{
-public:
-    AbsVal_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ABSVAL_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/batchnorm_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/batchnorm_riscv.h
deleted file mode 100644
index 1ed4dc63..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/batchnorm_riscv.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Xavier Hsinyuan is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 Xavier Hsinyuan <me@lstlx.com>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BATCHNORM_RISCV_H
-#define LAYER_BATCHNORM_RISCV_H
-
-#include "batchnorm.h"
-
-namespace ncnn {
-class BatchNorm_riscv : public BatchNorm
-{
-public:
-    BatchNorm_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BATCHNORM_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/binaryop_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/binaryop_riscv.h
deleted file mode 100644
index afc728b6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/binaryop_riscv.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Xavier Hsinyuan is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 Xavier Hsinyuan <thelastlinex@hotmail.com>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this
-// file except in compliance with the License. You may obtain a copy of the
-// License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
-// WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
-// License for the specific language governing permissions and limitations under
-// the License.
-
-#ifndef LAYER_BINARYOP_RISCV_H
-#define LAYER_BINARYOP_RISCV_H
-
-#include "binaryop.h"
-
-namespace ncnn {
-
-class BinaryOp_riscv : public BinaryOp
-{
-public:
-    BinaryOp_riscv();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_fp16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BINARYOP_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/cast_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/cast_riscv.h
deleted file mode 100644
index 7c6fbb6d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/cast_riscv.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CAST_RISCV_H
-#define LAYER_CAST_RISCV_H
-
-#include "cast.h"
-
-namespace ncnn {
-
-class Cast_riscv : public Cast
-{
-public:
-    Cast_riscv();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CAST_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/clip_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/clip_riscv.h
deleted file mode 100644
index 051995e1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/clip_riscv.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CLIP_RISCV_H
-#define LAYER_CLIP_RISCV_H
-
-#include "clip.h"
-
-namespace ncnn {
-
-class Clip_riscv : public Clip
-{
-public:
-    Clip_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CLIP_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/concat_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/concat_riscv.h
deleted file mode 100644
index 23029340..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/concat_riscv.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONCAT_RISCV_H
-#define LAYER_CONCAT_RISCV_H
-
-#include "concat.h"
-
-namespace ncnn {
-
-class Concat_riscv : public Concat
-{
-public:
-    Concat_riscv();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    int forward_bf16s_fp16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONCAT_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution1d_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution1d_riscv.h
deleted file mode 100644
index f0e7f881..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution1d_riscv.h
+++ /dev/null
@@ -1,52 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTION1D_RISCV_H
-#define LAYER_CONVOLUTION1D_RISCV_H
-
-#include "convolution1d.h"
-
-namespace ncnn {
-
-class Convolution1D_riscv : public Convolution1D
-{
-public:
-    Convolution1D_riscv();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    // packn
-    Mat weight_data_packed;
-
-    // fp16
-    Mat weight_data_fp16;
-    Mat bias_data_fp16;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTION1D_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1.h
deleted file mode 100644
index 5566d543..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1.h
+++ /dev/null
@@ -1,26 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_fp16s.h
deleted file mode 100644
index 2b652eec..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_fp16s.h
+++ /dev/null
@@ -1,26 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_fp16sa_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_pack1ton.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_pack1ton.h
deleted file mode 100644
index e53083aa..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_pack1ton.h
+++ /dev/null
@@ -1,26 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_pack1ton_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_pack1ton_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_pack1ton_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_pack1ton_fp16s.h
deleted file mode 100644
index 87d51019..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_pack1ton_fp16s.h
+++ /dev/null
@@ -1,26 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_pack1ton_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_pack1ton_fp16sa_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packn.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packn.h
deleted file mode 100644
index 31bf72ba..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packn.h
+++ /dev/null
@@ -1,68 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_packn_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_packn_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
-
-static void conv1x1s2_sgemm_packn_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * packn;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const float* r0 = bottom_blob.channel(p);
-        float* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat32m1_t _val = vle32_v_f32m1(r0, vl);
-                vse32_v_f32m1(outptr, _val, vl);
-
-                r0 += packn * 2;
-                outptr += packn;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_packn_rvv(bottom_blob_shrinked, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packn_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packn_fp16s.h
deleted file mode 100644
index 5ac3f896..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packn_fp16s.h
+++ /dev/null
@@ -1,68 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_packn_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_packn_fp16sa_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
-
-static void conv1x1s2_sgemm_packn_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * packn;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const __fp16* r0 = bottom_blob.channel(p);
-        __fp16* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat16m1_t _val = vle16_v_f16m1(r0, vl);
-                vse16_v_f16m1(outptr, _val, vl);
-
-                r0 += packn * 2;
-                outptr += packn;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_packn_fp16sa_rvv(bottom_blob_shrinked, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packnto1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packnto1.h
deleted file mode 100644
index a3e1204a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packnto1.h
+++ /dev/null
@@ -1,68 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_packnto1_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_packnto1_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
-
-static void conv1x1s2_sgemm_packnto1_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * packn;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const float* r0 = bottom_blob.channel(p);
-        float* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat32m1_t _val = vle32_v_f32m1(r0, vl);
-                vse32_v_f32m1(outptr, _val, vl);
-
-                r0 += packn * 2;
-                outptr += packn;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_packnto1_rvv(bottom_blob_shrinked, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packnto1_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packnto1_fp16s.h
deleted file mode 100644
index 10591ab2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_1x1_packnto1_fp16s.h
+++ /dev/null
@@ -1,68 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sgemm_packnto1_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    const int size = w * h;
-
-    Mat bottom_im2col = bottom_blob;
-    bottom_im2col.w = size;
-    bottom_im2col.h = 1;
-
-    im2col_sgemm_packnto1_fp16sa_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
-
-static void conv1x1s2_sgemm_packnto1_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * packn;
-
-    Mat bottom_blob_shrinked;
-    bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < channels; p++)
-    {
-        const __fp16* r0 = bottom_blob.channel(p);
-        __fp16* outptr = bottom_blob_shrinked.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat16m1_t _val = vle16_v_f16m1(r0, vl);
-                vse16_v_f16m1(outptr, _val, vl);
-
-                r0 += packn * 2;
-                outptr += packn;
-            }
-
-            r0 += tailstep;
-        }
-    }
-
-    conv1x1s1_sgemm_packnto1_fp16sa_rvv(bottom_blob_shrinked, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3.h
deleted file mode 100644
index 115a299b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3.h
+++ /dev/null
@@ -1,413 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd23_transform_kernel_rvv(const Mat& kernel, Mat& kernel_tm2, int inch, int outch, const Option& opt)
-{
-    Mat kernel_tm(4 * 4, inch, outch);
-
-    // G
-    const float ktm[4][3] = {
-        {1.0f, 0.0f, 0.0f},
-        {1.0f / 2, 1.0f / 2, 1.0f / 2},
-        {1.0f / 2, -1.0f / 2, 1.0f / 2},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[4][3];
-            for (int i = 0; i < 4; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 4; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 4; i++)
-                {
-                    kernel_tm0[j * 4 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 16-inch-outch
-    // dst = inch-16-outch
-#if __riscv_vector
-    kernel_tm2.create(8 * inch, 16, outch / 8 + (outch % 8) / 4 + outch % 4);
-#else
-    kernel_tm2.create(2 * inch, 16, outch / 2 + outch % 2);
-#endif
-
-    int q = 0;
-#if __riscv_vector
-    for (; q + 7 < outch; q += 8)
-    {
-        Mat g0 = kernel_tm2.channel(q / 8);
-
-        for (int k = 0; k < 16; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 8; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-    for (; q + 3 < outch; q += 4)
-    {
-        Mat g0 = kernel_tm2.channel(q / 8 + (q % 8) / 4);
-
-        for (int k = 0; k < 16; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#else  // __riscv_vector
-    for (; q + 1 < outch; q += 2)
-    {
-        Mat g0 = kernel_tm2.channel(q / 2);
-
-        for (int k = 0; k < 16; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 2; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#endif // __riscv_vector
-    for (; q < outch; q++)
-    {
-#if __riscv_vector
-        Mat g0 = kernel_tm2.channel(q / 8 + (q % 8) / 4 + q % 4);
-#else
-        Mat g0 = kernel_tm2.channel(q / 2 + q % 2);
-#endif
-
-        for (int k = 0; k < 16; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                const float* k00 = kernel_tm.channel(q).row(p);
-                g00[0] = k00[k];
-                g00++;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 2n+2, winograd F(2,3)
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 1) / 2 * 2;
-    outh = (outh + 1) / 2 * 2;
-
-    w = outw + 2;
-    h = outh + 2;
-    Option opt_b = opt;
-    opt_b.blob_allocator = opt.workspace_allocator;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, 0, 0.f, opt_b);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 2;
-        int h_tiles = outh / 2;
-        int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 16, inch, 4u, opt.workspace_allocator);
-        conv3x3s1_winograd23_transform_input_rvv(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_rvv(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, 4u, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd23_transform_output_rvv(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
-
-static void conv3x3s1_winograd43_transform_kernel_rvv(const Mat& kernel, Mat& kernel_tm2, int inch, int outch, const Option& opt)
-{
-    Mat kernel_tm(6 * 6, inch, outch);
-
-    // G
-    const float sq2 = 1.41421356237f;
-    const float ktm[6][3] = {
-        {1.0f, 0.0f, 0.0f},
-        {-2.0f / 3, -sq2 / 3, -1.0f / 3},
-        {-2.0f / 3, sq2 / 3, -1.0f / 3},
-        {1.0f / 6, sq2 / 6, 1.0f / 3},
-        {1.0f / 6, -sq2 / 6, 1.0f / 3},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[6][3];
-            for (int i = 0; i < 6; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 6; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 6; i++)
-                {
-                    kernel_tm0[j * 6 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 36-inch-outch
-    // dst = inch-36-outch
-#if __riscv_vector
-    kernel_tm2.create(8 * inch, 36, outch / 8 + (outch % 8) / 4 + outch % 4);
-#else
-    kernel_tm2.create(2 * inch, 36, outch / 2 + outch % 2);
-#endif
-
-    int q = 0;
-#if __riscv_vector
-    for (; q + 7 < outch; q += 8)
-    {
-        Mat g0 = kernel_tm2.channel(q / 8);
-
-        for (int k = 0; k < 36; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 8; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-    for (; q + 3 < outch; q += 4)
-    {
-        Mat g0 = kernel_tm2.channel(q / 8 + (q % 8) / 4);
-
-        for (int k = 0; k < 36; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 4; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#else  // __riscv_vector
-    for (; q + 1 < outch; q += 2)
-    {
-        Mat g0 = kernel_tm2.channel(q / 2);
-
-        for (int k = 0; k < 36; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                for (int i = 0; i < 2; i++)
-                {
-                    const float* k00 = kernel_tm.channel(q + i).row(p);
-                    g00[0] = k00[k];
-                    g00++;
-                }
-            }
-        }
-    }
-#endif // __riscv_vector
-    for (; q < outch; q++)
-    {
-#if __riscv_vector
-        Mat g0 = kernel_tm2.channel(q / 8 + (q % 8) / 4 + q % 4);
-#else
-        Mat g0 = kernel_tm2.channel(q / 2 + q % 2);
-#endif
-
-        for (int k = 0; k < 36; k++)
-        {
-            float* g00 = g0.row(k);
-
-            for (int p = 0; p < inch; p++)
-            {
-                const float* k00 = kernel_tm.channel(q).row(p);
-                g00[0] = k00[k];
-                g00++;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 4n+2, winograd F(4,3)
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 3) / 4 * 4;
-    outh = (outh + 3) / 4 * 4;
-
-    w = outw + 2;
-    h = outh + 2;
-
-    Option opt_b = opt;
-    opt_b.blob_allocator = opt.workspace_allocator;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, 0, 0.f, opt_b);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 4;
-        int h_tiles = outh / 4;
-        int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 36, inch, 4u, opt.workspace_allocator);
-        conv3x3s1_winograd43_transform_input_rvv(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_rvv(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, 4u, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd43_transform_output_rvv(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_pack1ton.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_pack1ton.h
deleted file mode 100644
index 9adcfb1e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_pack1ton.h
+++ /dev/null
@@ -1,566 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack1ton_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        vfloat32m1_t _bias0 = bias ? vle32_v_f32m1(bias + p * packn, vl) : vfmv_v_f_f32m1(0.f, vl);
-        out0.fill(_bias0);
-
-        const float* k0 = kernel.channel(p);
-
-        int q = 0;
-        for (; q < inch; q++)
-        {
-            float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            vfloat32m1_t _k00 = vle32_v_f32m1(k0, vl);
-            vfloat32m1_t _k01 = vle32_v_f32m1(k0 + packn, vl);
-            vfloat32m1_t _k02 = vle32_v_f32m1(k0 + packn * 2, vl);
-            vfloat32m1_t _k10 = vle32_v_f32m1(k0 + packn * 3, vl);
-            vfloat32m1_t _k11 = vle32_v_f32m1(k0 + packn * 4, vl);
-            vfloat32m1_t _k12 = vle32_v_f32m1(k0 + packn * 5, vl);
-            vfloat32m1_t _k20 = vle32_v_f32m1(k0 + packn * 6, vl);
-            vfloat32m1_t _k21 = vle32_v_f32m1(k0 + packn * 7, vl);
-            vfloat32m1_t _k22 = vle32_v_f32m1(k0 + packn * 8, vl);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    vfloat32m1_t _sum0 = vle32_v_f32m1(outptr0, vl);
-                    vfloat32m1_t _sum1 = vle32_v_f32m1(outptr0 + packn, vl);
-                    vfloat32m1_t _sum2 = vle32_v_f32m1(outptr0 + packn * 2, vl);
-                    vfloat32m1_t _sum3 = vle32_v_f32m1(outptr0 + packn * 3, vl);
-                    vfloat32m1_t _sum4 = vle32_v_f32m1(outptr0 + packn * 4, vl);
-                    vfloat32m1_t _sum5 = vle32_v_f32m1(outptr0 + packn * 5, vl);
-                    vfloat32m1_t _sum6 = vle32_v_f32m1(outptr0 + packn * 6, vl);
-                    vfloat32m1_t _sum7 = vle32_v_f32m1(outptr0 + packn * 7, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[1], _k00, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[2], _k00, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[3], _k00, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r0[4], _k00, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r0[5], _k00, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r0[6], _k00, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r0[7], _k00, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[2], _k01, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[3], _k01, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[4], _k01, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r0[5], _k01, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r0[6], _k01, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r0[7], _k01, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r0[8], _k01, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[3], _k02, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[4], _k02, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[5], _k02, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r0[6], _k02, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r0[7], _k02, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r0[8], _k02, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r0[9], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[1], _k10, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[2], _k10, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[3], _k10, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r1[4], _k10, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r1[5], _k10, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r1[6], _k10, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r1[7], _k10, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[2], _k11, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[3], _k11, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[4], _k11, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r1[5], _k11, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r1[6], _k11, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r1[7], _k11, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r1[8], _k11, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[3], _k12, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[4], _k12, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[5], _k12, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r1[6], _k12, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r1[7], _k12, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r1[8], _k12, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r1[9], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[1], _k20, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[2], _k20, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[3], _k20, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r2[4], _k20, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r2[5], _k20, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r2[6], _k20, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r2[7], _k20, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[2], _k21, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[3], _k21, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[4], _k21, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r2[5], _k21, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r2[6], _k21, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r2[7], _k21, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r2[8], _k21, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[3], _k22, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[4], _k22, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[5], _k22, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r2[6], _k22, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r2[7], _k22, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r2[8], _k22, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r2[9], _k22, vl);
-
-                    vse32_v_f32m1(outptr0, _sum0, vl);
-                    vse32_v_f32m1(outptr0 + packn, _sum1, vl);
-                    vse32_v_f32m1(outptr0 + packn * 2, _sum2, vl);
-                    vse32_v_f32m1(outptr0 + packn * 3, _sum3, vl);
-                    vse32_v_f32m1(outptr0 + packn * 4, _sum4, vl);
-                    vse32_v_f32m1(outptr0 + packn * 5, _sum5, vl);
-                    vse32_v_f32m1(outptr0 + packn * 6, _sum6, vl);
-                    vse32_v_f32m1(outptr0 + packn * 7, _sum7, vl);
-
-                    outptr0 += packn * 8;
-
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                }
-                for (; j + 3 < outw; j += 4)
-                {
-                    vfloat32m1_t _sum0 = vle32_v_f32m1(outptr0, vl);
-                    vfloat32m1_t _sum1 = vle32_v_f32m1(outptr0 + packn, vl);
-                    vfloat32m1_t _sum2 = vle32_v_f32m1(outptr0 + packn * 2, vl);
-                    vfloat32m1_t _sum3 = vle32_v_f32m1(outptr0 + packn * 3, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[1], _k00, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[2], _k00, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[3], _k00, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[2], _k01, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[3], _k01, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[4], _k01, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[3], _k02, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[4], _k02, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[5], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[1], _k10, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[2], _k10, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[3], _k10, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[2], _k11, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[3], _k11, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[4], _k11, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[3], _k12, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[4], _k12, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[5], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[1], _k20, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[2], _k20, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[3], _k20, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[2], _k21, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[3], _k21, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[4], _k21, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[3], _k22, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[4], _k22, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[5], _k22, vl);
-
-                    vse32_v_f32m1(outptr0, _sum0, vl);
-                    vse32_v_f32m1(outptr0 + packn, _sum1, vl);
-                    vse32_v_f32m1(outptr0 + packn * 2, _sum2, vl);
-                    vse32_v_f32m1(outptr0 + packn * 3, _sum3, vl);
-
-                    outptr0 += packn * 4;
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    vfloat32m1_t _sum0 = vle32_v_f32m1(outptr0, vl);
-                    vfloat32m1_t _sum1 = vle32_v_f32m1(outptr0 + packn, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[1], _k00, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[2], _k01, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[3], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[1], _k10, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[2], _k11, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[3], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[1], _k20, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[2], _k21, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[3], _k22, vl);
-
-                    vse32_v_f32m1(outptr0, _sum0, vl);
-                    vse32_v_f32m1(outptr0 + packn, _sum1, vl);
-
-                    outptr0 += packn * 2;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                }
-                for (; j < outw; j++)
-                {
-                    vfloat32m1_t _sum0 = vle32_v_f32m1(outptr0, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[0], _k00, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[1], _k01, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[2], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[0], _k10, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[1], _k11, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[2], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[0], _k20, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[1], _k21, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[2], _k22, vl);
-
-                    vse32_v_f32m1(outptr0, _sum0, vl);
-
-                    outptr0 += packn;
-
-                    r0 += 1;
-                    r1 += 1;
-                    r2 += 1;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * packn;
-        }
-    }
-}
-
-static void conv3x3s2_pack1ton_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        vfloat32m1_t _bias0 = bias ? vle32_v_f32m1(bias + p * packn, vl) : vfmv_v_f_f32m1(0.f, vl);
-        out0.fill(_bias0);
-
-        const float* k0 = kernel.channel(p);
-
-        int q = 0;
-        for (; q < inch; q++)
-        {
-            float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            vfloat32m1_t _k00 = vle32_v_f32m1(k0, vl);
-            vfloat32m1_t _k01 = vle32_v_f32m1(k0 + packn, vl);
-            vfloat32m1_t _k02 = vle32_v_f32m1(k0 + packn * 2, vl);
-            vfloat32m1_t _k10 = vle32_v_f32m1(k0 + packn * 3, vl);
-            vfloat32m1_t _k11 = vle32_v_f32m1(k0 + packn * 4, vl);
-            vfloat32m1_t _k12 = vle32_v_f32m1(k0 + packn * 5, vl);
-            vfloat32m1_t _k20 = vle32_v_f32m1(k0 + packn * 6, vl);
-            vfloat32m1_t _k21 = vle32_v_f32m1(k0 + packn * 7, vl);
-            vfloat32m1_t _k22 = vle32_v_f32m1(k0 + packn * 8, vl);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    vfloat32m1_t _sum0 = vle32_v_f32m1(outptr0, vl);
-                    vfloat32m1_t _sum1 = vle32_v_f32m1(outptr0 + packn, vl);
-                    vfloat32m1_t _sum2 = vle32_v_f32m1(outptr0 + packn * 2, vl);
-                    vfloat32m1_t _sum3 = vle32_v_f32m1(outptr0 + packn * 3, vl);
-                    vfloat32m1_t _sum4 = vle32_v_f32m1(outptr0 + packn * 4, vl);
-                    vfloat32m1_t _sum5 = vle32_v_f32m1(outptr0 + packn * 5, vl);
-                    vfloat32m1_t _sum6 = vle32_v_f32m1(outptr0 + packn * 6, vl);
-                    vfloat32m1_t _sum7 = vle32_v_f32m1(outptr0 + packn * 7, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[2], _k00, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[4], _k00, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[6], _k00, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r0[8], _k00, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r0[10], _k00, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r0[12], _k00, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r0[14], _k00, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[3], _k01, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[5], _k01, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[7], _k01, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r0[9], _k01, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r0[11], _k01, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r0[13], _k01, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r0[15], _k01, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[4], _k02, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[6], _k02, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[8], _k02, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r0[10], _k02, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r0[12], _k02, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r0[14], _k02, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r0[16], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[2], _k10, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[4], _k10, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[6], _k10, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r1[8], _k10, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r1[10], _k10, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r1[12], _k10, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r1[14], _k10, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[3], _k11, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[5], _k11, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[7], _k11, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r1[9], _k11, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r1[11], _k11, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r1[13], _k11, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r1[15], _k11, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[4], _k12, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[6], _k12, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[8], _k12, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r1[10], _k12, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r1[12], _k12, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r1[14], _k12, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r1[16], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[2], _k20, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[4], _k20, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[6], _k20, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r2[8], _k20, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r2[10], _k20, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r2[12], _k20, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r2[14], _k20, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[3], _k21, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[5], _k21, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[7], _k21, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r2[9], _k21, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r2[11], _k21, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r2[13], _k21, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r2[15], _k21, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[4], _k22, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[6], _k22, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[8], _k22, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r2[10], _k22, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r2[12], _k22, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r2[14], _k22, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r2[16], _k22, vl);
-
-                    vse32_v_f32m1(outptr0, _sum0, vl);
-                    vse32_v_f32m1(outptr0 + packn, _sum1, vl);
-                    vse32_v_f32m1(outptr0 + packn * 2, _sum2, vl);
-                    vse32_v_f32m1(outptr0 + packn * 3, _sum3, vl);
-                    vse32_v_f32m1(outptr0 + packn * 4, _sum4, vl);
-                    vse32_v_f32m1(outptr0 + packn * 5, _sum5, vl);
-                    vse32_v_f32m1(outptr0 + packn * 6, _sum6, vl);
-                    vse32_v_f32m1(outptr0 + packn * 7, _sum7, vl);
-
-                    outptr0 += packn * 8;
-
-                    r0 += 16;
-                    r1 += 16;
-                    r2 += 16;
-                }
-                for (; j + 3 < outw; j += 4)
-                {
-                    vfloat32m1_t _sum0 = vle32_v_f32m1(outptr0, vl);
-                    vfloat32m1_t _sum1 = vle32_v_f32m1(outptr0 + packn, vl);
-                    vfloat32m1_t _sum2 = vle32_v_f32m1(outptr0 + packn * 2, vl);
-                    vfloat32m1_t _sum3 = vle32_v_f32m1(outptr0 + packn * 3, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[2], _k00, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[4], _k00, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[6], _k00, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[3], _k01, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[5], _k01, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[7], _k01, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[4], _k02, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[6], _k02, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[8], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[2], _k10, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[4], _k10, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[6], _k10, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[3], _k11, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[5], _k11, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[7], _k11, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[4], _k12, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[6], _k12, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[8], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[2], _k20, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[4], _k20, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[6], _k20, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[3], _k21, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[5], _k21, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[7], _k21, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[4], _k22, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[6], _k22, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[8], _k22, vl);
-
-                    vse32_v_f32m1(outptr0, _sum0, vl);
-                    vse32_v_f32m1(outptr0 + packn, _sum1, vl);
-                    vse32_v_f32m1(outptr0 + packn * 2, _sum2, vl);
-                    vse32_v_f32m1(outptr0 + packn * 3, _sum3, vl);
-
-                    outptr0 += packn * 4;
-
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    vfloat32m1_t _sum0 = vle32_v_f32m1(outptr0, vl);
-                    vfloat32m1_t _sum1 = vle32_v_f32m1(outptr0 + packn, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[2], _k00, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[3], _k01, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[4], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[2], _k10, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[3], _k11, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[4], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[2], _k20, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[3], _k21, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[4], _k22, vl);
-
-                    vse32_v_f32m1(outptr0, _sum0, vl);
-                    vse32_v_f32m1(outptr0 + packn, _sum1, vl);
-
-                    outptr0 += packn * 2;
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                }
-                for (; j < outw; j++)
-                {
-                    vfloat32m1_t _sum0 = vle32_v_f32m1(outptr0, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[0], _k00, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[1], _k01, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[2], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[0], _k10, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[1], _k11, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[2], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[0], _k20, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[1], _k21, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[2], _k22, vl);
-
-                    vse32_v_f32m1(outptr0, _sum0, vl);
-
-                    outptr0 += packn;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * packn;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_pack1ton_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_pack1ton_fp16s.h
deleted file mode 100644
index bff24a00..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_pack1ton_fp16s.h
+++ /dev/null
@@ -1,566 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack1ton_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        vfloat16m1_t _bias0 = bias ? vle16_v_f16m1(bias + p * packn, vl) : vfmv_v_f_f16m1(0.f, vl);
-        out0.fill(_bias0);
-
-        const __fp16* k0 = kernel.channel(p);
-
-        int q = 0;
-        for (; q < inch; q++)
-        {
-            __fp16* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const __fp16* r0 = img0.row<const __fp16>(0);
-            const __fp16* r1 = img0.row<const __fp16>(1);
-            const __fp16* r2 = img0.row<const __fp16>(2);
-
-            vfloat16m1_t _k00 = vle16_v_f16m1(k0, vl);
-            vfloat16m1_t _k01 = vle16_v_f16m1(k0 + packn, vl);
-            vfloat16m1_t _k02 = vle16_v_f16m1(k0 + packn * 2, vl);
-            vfloat16m1_t _k10 = vle16_v_f16m1(k0 + packn * 3, vl);
-            vfloat16m1_t _k11 = vle16_v_f16m1(k0 + packn * 4, vl);
-            vfloat16m1_t _k12 = vle16_v_f16m1(k0 + packn * 5, vl);
-            vfloat16m1_t _k20 = vle16_v_f16m1(k0 + packn * 6, vl);
-            vfloat16m1_t _k21 = vle16_v_f16m1(k0 + packn * 7, vl);
-            vfloat16m1_t _k22 = vle16_v_f16m1(k0 + packn * 8, vl);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    vfloat16m1_t _sum0 = vle16_v_f16m1(outptr0, vl);
-                    vfloat16m1_t _sum1 = vle16_v_f16m1(outptr0 + packn, vl);
-                    vfloat16m1_t _sum2 = vle16_v_f16m1(outptr0 + packn * 2, vl);
-                    vfloat16m1_t _sum3 = vle16_v_f16m1(outptr0 + packn * 3, vl);
-                    vfloat16m1_t _sum4 = vle16_v_f16m1(outptr0 + packn * 4, vl);
-                    vfloat16m1_t _sum5 = vle16_v_f16m1(outptr0 + packn * 5, vl);
-                    vfloat16m1_t _sum6 = vle16_v_f16m1(outptr0 + packn * 6, vl);
-                    vfloat16m1_t _sum7 = vle16_v_f16m1(outptr0 + packn * 7, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[1], _k00, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[2], _k00, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[3], _k00, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r0[4], _k00, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r0[5], _k00, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r0[6], _k00, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r0[7], _k00, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[2], _k01, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[3], _k01, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[4], _k01, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r0[5], _k01, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r0[6], _k01, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r0[7], _k01, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r0[8], _k01, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[3], _k02, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[4], _k02, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[5], _k02, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r0[6], _k02, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r0[7], _k02, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r0[8], _k02, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r0[9], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[1], _k10, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[2], _k10, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[3], _k10, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r1[4], _k10, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r1[5], _k10, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r1[6], _k10, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r1[7], _k10, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[2], _k11, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[3], _k11, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[4], _k11, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r1[5], _k11, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r1[6], _k11, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r1[7], _k11, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r1[8], _k11, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[3], _k12, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[4], _k12, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[5], _k12, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r1[6], _k12, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r1[7], _k12, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r1[8], _k12, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r1[9], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[1], _k20, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[2], _k20, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[3], _k20, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r2[4], _k20, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r2[5], _k20, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r2[6], _k20, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r2[7], _k20, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[2], _k21, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[3], _k21, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[4], _k21, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r2[5], _k21, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r2[6], _k21, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r2[7], _k21, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r2[8], _k21, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[3], _k22, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[4], _k22, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[5], _k22, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r2[6], _k22, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r2[7], _k22, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r2[8], _k22, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r2[9], _k22, vl);
-
-                    vse16_v_f16m1(outptr0, _sum0, vl);
-                    vse16_v_f16m1(outptr0 + packn, _sum1, vl);
-                    vse16_v_f16m1(outptr0 + packn * 2, _sum2, vl);
-                    vse16_v_f16m1(outptr0 + packn * 3, _sum3, vl);
-                    vse16_v_f16m1(outptr0 + packn * 4, _sum4, vl);
-                    vse16_v_f16m1(outptr0 + packn * 5, _sum5, vl);
-                    vse16_v_f16m1(outptr0 + packn * 6, _sum6, vl);
-                    vse16_v_f16m1(outptr0 + packn * 7, _sum7, vl);
-
-                    outptr0 += packn * 8;
-
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                }
-                for (; j + 3 < outw; j += 4)
-                {
-                    vfloat16m1_t _sum0 = vle16_v_f16m1(outptr0, vl);
-                    vfloat16m1_t _sum1 = vle16_v_f16m1(outptr0 + packn, vl);
-                    vfloat16m1_t _sum2 = vle16_v_f16m1(outptr0 + packn * 2, vl);
-                    vfloat16m1_t _sum3 = vle16_v_f16m1(outptr0 + packn * 3, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[1], _k00, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[2], _k00, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[3], _k00, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[2], _k01, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[3], _k01, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[4], _k01, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[3], _k02, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[4], _k02, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[5], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[1], _k10, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[2], _k10, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[3], _k10, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[2], _k11, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[3], _k11, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[4], _k11, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[3], _k12, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[4], _k12, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[5], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[1], _k20, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[2], _k20, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[3], _k20, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[2], _k21, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[3], _k21, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[4], _k21, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[3], _k22, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[4], _k22, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[5], _k22, vl);
-
-                    vse16_v_f16m1(outptr0, _sum0, vl);
-                    vse16_v_f16m1(outptr0 + packn, _sum1, vl);
-                    vse16_v_f16m1(outptr0 + packn * 2, _sum2, vl);
-                    vse16_v_f16m1(outptr0 + packn * 3, _sum3, vl);
-
-                    outptr0 += packn * 4;
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    vfloat16m1_t _sum0 = vle16_v_f16m1(outptr0, vl);
-                    vfloat16m1_t _sum1 = vle16_v_f16m1(outptr0 + packn, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[1], _k00, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[2], _k01, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[3], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[1], _k10, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[2], _k11, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[3], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[1], _k20, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[2], _k21, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[3], _k22, vl);
-
-                    vse16_v_f16m1(outptr0, _sum0, vl);
-                    vse16_v_f16m1(outptr0 + packn, _sum1, vl);
-
-                    outptr0 += packn * 2;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                }
-                for (; j < outw; j++)
-                {
-                    vfloat16m1_t _sum0 = vle16_v_f16m1(outptr0, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[0], _k00, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[1], _k01, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[2], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[0], _k10, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[1], _k11, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[2], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[0], _k20, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[1], _k21, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[2], _k22, vl);
-
-                    vse16_v_f16m1(outptr0, _sum0, vl);
-
-                    outptr0 += packn;
-
-                    r0 += 1;
-                    r1 += 1;
-                    r2 += 1;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * packn;
-        }
-    }
-}
-
-static void conv3x3s2_pack1ton_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        vfloat16m1_t _bias0 = bias ? vle16_v_f16m1(bias + p * packn, vl) : vfmv_v_f_f16m1(0.f, vl);
-        out0.fill(_bias0);
-
-        const __fp16* k0 = kernel.channel(p);
-
-        int q = 0;
-        for (; q < inch; q++)
-        {
-            __fp16* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const __fp16* r0 = img0.row<const __fp16>(0);
-            const __fp16* r1 = img0.row<const __fp16>(1);
-            const __fp16* r2 = img0.row<const __fp16>(2);
-
-            vfloat16m1_t _k00 = vle16_v_f16m1(k0, vl);
-            vfloat16m1_t _k01 = vle16_v_f16m1(k0 + packn, vl);
-            vfloat16m1_t _k02 = vle16_v_f16m1(k0 + packn * 2, vl);
-            vfloat16m1_t _k10 = vle16_v_f16m1(k0 + packn * 3, vl);
-            vfloat16m1_t _k11 = vle16_v_f16m1(k0 + packn * 4, vl);
-            vfloat16m1_t _k12 = vle16_v_f16m1(k0 + packn * 5, vl);
-            vfloat16m1_t _k20 = vle16_v_f16m1(k0 + packn * 6, vl);
-            vfloat16m1_t _k21 = vle16_v_f16m1(k0 + packn * 7, vl);
-            vfloat16m1_t _k22 = vle16_v_f16m1(k0 + packn * 8, vl);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    vfloat16m1_t _sum0 = vle16_v_f16m1(outptr0, vl);
-                    vfloat16m1_t _sum1 = vle16_v_f16m1(outptr0 + packn, vl);
-                    vfloat16m1_t _sum2 = vle16_v_f16m1(outptr0 + packn * 2, vl);
-                    vfloat16m1_t _sum3 = vle16_v_f16m1(outptr0 + packn * 3, vl);
-                    vfloat16m1_t _sum4 = vle16_v_f16m1(outptr0 + packn * 4, vl);
-                    vfloat16m1_t _sum5 = vle16_v_f16m1(outptr0 + packn * 5, vl);
-                    vfloat16m1_t _sum6 = vle16_v_f16m1(outptr0 + packn * 6, vl);
-                    vfloat16m1_t _sum7 = vle16_v_f16m1(outptr0 + packn * 7, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[2], _k00, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[4], _k00, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[6], _k00, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r0[8], _k00, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r0[10], _k00, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r0[12], _k00, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r0[14], _k00, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[3], _k01, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[5], _k01, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[7], _k01, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r0[9], _k01, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r0[11], _k01, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r0[13], _k01, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r0[15], _k01, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[4], _k02, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[6], _k02, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[8], _k02, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r0[10], _k02, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r0[12], _k02, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r0[14], _k02, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r0[16], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[2], _k10, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[4], _k10, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[6], _k10, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r1[8], _k10, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r1[10], _k10, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r1[12], _k10, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r1[14], _k10, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[3], _k11, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[5], _k11, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[7], _k11, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r1[9], _k11, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r1[11], _k11, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r1[13], _k11, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r1[15], _k11, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[4], _k12, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[6], _k12, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[8], _k12, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r1[10], _k12, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r1[12], _k12, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r1[14], _k12, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r1[16], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[2], _k20, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[4], _k20, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[6], _k20, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r2[8], _k20, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r2[10], _k20, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r2[12], _k20, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r2[14], _k20, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[3], _k21, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[5], _k21, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[7], _k21, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r2[9], _k21, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r2[11], _k21, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r2[13], _k21, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r2[15], _k21, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[4], _k22, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[6], _k22, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[8], _k22, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r2[10], _k22, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r2[12], _k22, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r2[14], _k22, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r2[16], _k22, vl);
-
-                    vse16_v_f16m1(outptr0, _sum0, vl);
-                    vse16_v_f16m1(outptr0 + packn, _sum1, vl);
-                    vse16_v_f16m1(outptr0 + packn * 2, _sum2, vl);
-                    vse16_v_f16m1(outptr0 + packn * 3, _sum3, vl);
-                    vse16_v_f16m1(outptr0 + packn * 4, _sum4, vl);
-                    vse16_v_f16m1(outptr0 + packn * 5, _sum5, vl);
-                    vse16_v_f16m1(outptr0 + packn * 6, _sum6, vl);
-                    vse16_v_f16m1(outptr0 + packn * 7, _sum7, vl);
-
-                    outptr0 += packn * 8;
-
-                    r0 += 16;
-                    r1 += 16;
-                    r2 += 16;
-                }
-                for (; j + 3 < outw; j += 4)
-                {
-                    vfloat16m1_t _sum0 = vle16_v_f16m1(outptr0, vl);
-                    vfloat16m1_t _sum1 = vle16_v_f16m1(outptr0 + packn, vl);
-                    vfloat16m1_t _sum2 = vle16_v_f16m1(outptr0 + packn * 2, vl);
-                    vfloat16m1_t _sum3 = vle16_v_f16m1(outptr0 + packn * 3, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[2], _k00, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[4], _k00, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[6], _k00, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[3], _k01, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[5], _k01, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[7], _k01, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[4], _k02, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[6], _k02, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[8], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[2], _k10, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[4], _k10, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[6], _k10, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[3], _k11, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[5], _k11, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[7], _k11, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[4], _k12, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[6], _k12, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[8], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[2], _k20, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[4], _k20, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[6], _k20, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[3], _k21, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[5], _k21, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[7], _k21, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[4], _k22, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[6], _k22, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[8], _k22, vl);
-
-                    vse16_v_f16m1(outptr0, _sum0, vl);
-                    vse16_v_f16m1(outptr0 + packn, _sum1, vl);
-                    vse16_v_f16m1(outptr0 + packn * 2, _sum2, vl);
-                    vse16_v_f16m1(outptr0 + packn * 3, _sum3, vl);
-
-                    outptr0 += packn * 4;
-
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    vfloat16m1_t _sum0 = vle16_v_f16m1(outptr0, vl);
-                    vfloat16m1_t _sum1 = vle16_v_f16m1(outptr0 + packn, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[2], _k00, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[3], _k01, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[4], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[2], _k10, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[3], _k11, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[4], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[2], _k20, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[3], _k21, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[4], _k22, vl);
-
-                    vse16_v_f16m1(outptr0, _sum0, vl);
-                    vse16_v_f16m1(outptr0 + packn, _sum1, vl);
-
-                    outptr0 += packn * 2;
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                }
-                for (; j < outw; j++)
-                {
-                    vfloat16m1_t _sum0 = vle16_v_f16m1(outptr0, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[0], _k00, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[1], _k01, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[2], _k02, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[0], _k10, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[1], _k11, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[2], _k12, vl);
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[0], _k20, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[1], _k21, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[2], _k22, vl);
-
-                    vse16_v_f16m1(outptr0, _sum0, vl);
-
-                    outptr0 += packn;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * packn;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_packn.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_packn.h
deleted file mode 100644
index 8734f016..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_packn.h
+++ /dev/null
@@ -1,432 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd63_transform_kernel_packn_rvv(const Mat& kernel, Mat& kernel_tm_packn, int inch, int outch, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-
-    // winograd63 transform kernel
-    Mat kernel_tm;
-    kernel_tm.create(8 * 8, inch, outch);
-
-    const float ktm[8][3] = {
-        {1.0f, 0.0f, 0.0f},
-        {-2.0f / 9, -2.0f / 9, -2.0f / 9},
-        {-2.0f / 9, 2.0f / 9, -2.0f / 9},
-        {1.0f / 90, 1.0f / 45, 2.0f / 45},
-        {1.0f / 90, -1.0f / 45, 2.0f / 45},
-        {1.0f / 45, 1.0f / 90, 1.0f / 180},
-        {1.0f / 45, -1.0f / 90, 1.0f / 180},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel, transposed
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[8][3];
-            for (int i = 0; i < 8; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // v
-            for (int j = 0; j < 8; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 8; i++)
-                {
-                    kernel_tm0[j * 8 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 64-inch-outch
-    // dst = pb-pa-inch/pa-64-outch/pb
-    kernel_tm_packn.create(inch / packn, 64, outch / packn, (size_t)4u * packn * packn, packn * packn);
-
-    for (int q = 0; q + (packn - 1) < outch; q += packn)
-    {
-        Mat g0 = kernel_tm_packn.channel(q / packn);
-
-        for (int k = 0; k < 64; k++)
-        {
-            float* g00 = g0.row<float>(k);
-
-            for (int p = 0; p + (packn - 1) < inch; p += packn)
-            {
-                for (int i = 0; i < packn; i++)
-                {
-                    for (int j = 0; j < packn; j++)
-                    {
-                        const float* k00 = kernel_tm.channel(q + j).row(p + i);
-                        g00[0] = (float)k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd63_packn_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 6n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 5) / 6 * 6;
-    outh = (outh + 5) / 6 * 6;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 6;
-        int h_tiles = outh / 6;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 64, inch, elemsize, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd63_transform_input_packn_rvv(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_packn_rvv(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, elemsize, elempack, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd63_transform_output_packn_rvv(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
-
-static void conv3x3s1_winograd43_transform_kernel_packn_rvv(const Mat& kernel, Mat& kernel_tm_packn, int inch, int outch, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-
-    // winograd43 transform kernel
-    Mat kernel_tm(6 * 6, inch, outch);
-
-    const float sq2 = 1.41421356237f;
-    const float ktm[6][3] = {
-        {1.0f, 0.0f, 0.0f},
-        {-2.0f / 3, -sq2 / 3, -1.0f / 3},
-        {-2.0f / 3, sq2 / 3, -1.0f / 3},
-        {1.0f / 6, sq2 / 6, 1.0f / 3},
-        {1.0f / 6, -sq2 / 6, 1.0f / 3},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[6][3];
-            for (int i = 0; i < 6; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 6; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 6; i++)
-                {
-                    kernel_tm0[j * 6 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 36-inch-outch
-    // dst = pb-pa-inch/pa-36-outch/pb
-    kernel_tm_packn.create(inch / packn, 36, outch / packn, (size_t)4u * packn * packn, packn * packn);
-
-    for (int q = 0; q + (packn - 1) < outch; q += packn)
-    {
-        Mat g0 = kernel_tm_packn.channel(q / packn);
-
-        for (int k = 0; k < 36; k++)
-        {
-            float* g00 = g0.row<float>(k);
-
-            for (int p = 0; p + (packn - 1) < inch; p += packn)
-            {
-                for (int i = 0; i < packn; i++)
-                {
-                    for (int j = 0; j < packn; j++)
-                    {
-                        const float* k00 = kernel_tm.channel(q + j).row(p + i);
-                        g00[0] = (float)k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_packn_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 4n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 3) / 4 * 4;
-    outh = (outh + 3) / 4 * 4;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 4;
-        int h_tiles = outh / 4;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 36, inch, elemsize, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd43_transform_input_packn_rvv(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_packn_rvv(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, elemsize, elempack, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd43_transform_output_packn_rvv(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
-
-static void conv3x3s1_winograd23_transform_kernel_packn_rvv(const Mat& kernel, Mat& kernel_tm_packn, int inch, int outch, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-
-    // winograd23 transform kernel
-    Mat kernel_tm(4 * 4, inch, outch);
-
-    const float ktm[4][3] = {
-        {1.0f, 0.0f, 0.0f},
-        {1.0f / 2, 1.0f / 2, 1.0f / 2},
-        {1.0f / 2, -1.0f / 2, 1.0f / 2},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[4][3];
-            for (int i = 0; i < 4; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 4; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 4; i++)
-                {
-                    kernel_tm0[j * 4 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 16-inch-outch
-    // dst = pb-pa-inch/pa-16-outch/pb
-    kernel_tm_packn.create(inch / packn, 16, outch / packn, (size_t)4u * packn * packn, packn * packn);
-
-    for (int q = 0; q + (packn - 1) < outch; q += packn)
-    {
-        Mat g0 = kernel_tm_packn.channel(q / packn);
-
-        for (int k = 0; k < 16; k++)
-        {
-            float* g00 = g0.row<float>(k);
-
-            for (int p = 0; p + (packn - 1) < inch; p += packn)
-            {
-                for (int i = 0; i < packn; i++)
-                {
-                    for (int j = 0; j < packn; j++)
-                    {
-                        const float* k00 = kernel_tm.channel(q + j).row(p + i);
-                        g00[0] = (float)k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_packn_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 2n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 1) / 2 * 2;
-    outh = (outh + 1) / 2 * 2;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 2;
-        int h_tiles = outh / 2;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 16, inch, elemsize, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd23_transform_input_packn_rvv(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_packn_rvv(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, elemsize, elempack, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd23_transform_output_packn_rvv(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_packn_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_packn_fp16s.h
deleted file mode 100644
index 022e00ff..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_3x3_packn_fp16s.h
+++ /dev/null
@@ -1,432 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd63_transform_kernel_packn_fp16sa_rvv(const Mat& kernel, Mat& kernel_tm_packn, int inch, int outch, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-
-    // winograd63 transform kernel
-    Mat kernel_tm;
-    kernel_tm.create(8 * 8, inch, outch);
-
-    const float ktm[8][3] = {
-        {1.0f, 0.0f, 0.0f},
-        {-2.0f / 9, -2.0f / 9, -2.0f / 9},
-        {-2.0f / 9, 2.0f / 9, -2.0f / 9},
-        {1.0f / 90, 1.0f / 45, 2.0f / 45},
-        {1.0f / 90, -1.0f / 45, 2.0f / 45},
-        {1.0f / 45, 1.0f / 90, 1.0f / 180},
-        {1.0f / 45, -1.0f / 90, 1.0f / 180},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel, transposed
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[8][3];
-            for (int i = 0; i < 8; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // v
-            for (int j = 0; j < 8; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 8; i++)
-                {
-                    kernel_tm0[j * 8 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 64-inch-outch
-    // dst = pb-pa-inch/pa-64-outch/pb
-    kernel_tm_packn.create(inch / packn, 64, outch / packn, (size_t)2u * packn * packn, packn * packn);
-
-    for (int q = 0; q + (packn - 1) < outch; q += packn)
-    {
-        Mat g0 = kernel_tm_packn.channel(q / packn);
-
-        for (int k = 0; k < 64; k++)
-        {
-            __fp16* g00 = g0.row<__fp16>(k);
-
-            for (int p = 0; p + (packn - 1) < inch; p += packn)
-            {
-                for (int i = 0; i < packn; i++)
-                {
-                    for (int j = 0; j < packn; j++)
-                    {
-                        const float* k00 = kernel_tm.channel(q + j).row(p + i);
-                        g00[0] = (__fp16)k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd63_packn_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 6n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 5) / 6 * 6;
-    outh = (outh + 5) / 6 * 6;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 6;
-        int h_tiles = outh / 6;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 64, inch, elemsize, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd63_transform_input_packn_fp16sa_rvv(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_packn_fp16sa_rvv(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, elemsize, elempack, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd63_transform_output_packn_fp16sa_rvv(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
-
-static void conv3x3s1_winograd43_transform_kernel_packn_fp16sa_rvv(const Mat& kernel, Mat& kernel_tm_packn, int inch, int outch, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-
-    // winograd43 transform kernel
-    Mat kernel_tm(6 * 6, inch, outch);
-
-    const float sq2 = 1.41421356237f;
-    const float ktm[6][3] = {
-        {1.0f, 0.0f, 0.0f},
-        {-2.0f / 3, -sq2 / 3, -1.0f / 3},
-        {-2.0f / 3, sq2 / 3, -1.0f / 3},
-        {1.0f / 6, sq2 / 6, 1.0f / 3},
-        {1.0f / 6, -sq2 / 6, 1.0f / 3},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[6][3];
-            for (int i = 0; i < 6; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 6; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 6; i++)
-                {
-                    kernel_tm0[j * 6 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 36-inch-outch
-    // dst = pb-pa-inch/pa-36-outch/pb
-    kernel_tm_packn.create(inch / packn, 36, outch / packn, (size_t)2u * packn * packn, packn * packn);
-
-    for (int q = 0; q + (packn - 1) < outch; q += packn)
-    {
-        Mat g0 = kernel_tm_packn.channel(q / packn);
-
-        for (int k = 0; k < 36; k++)
-        {
-            __fp16* g00 = g0.row<__fp16>(k);
-
-            for (int p = 0; p + (packn - 1) < inch; p += packn)
-            {
-                for (int i = 0; i < packn; i++)
-                {
-                    for (int j = 0; j < packn; j++)
-                    {
-                        const float* k00 = kernel_tm.channel(q + j).row(p + i);
-                        g00[0] = (__fp16)k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_packn_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 4n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 3) / 4 * 4;
-    outh = (outh + 3) / 4 * 4;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 4;
-        int h_tiles = outh / 4;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 36, inch, elemsize, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd43_transform_input_packn_fp16sa_rvv(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_packn_fp16sa_rvv(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, elemsize, elempack, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd43_transform_output_packn_fp16sa_rvv(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
-
-static void conv3x3s1_winograd23_transform_kernel_packn_fp16sa_rvv(const Mat& kernel, Mat& kernel_tm_packn, int inch, int outch, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-
-    // winograd23 transform kernel
-    Mat kernel_tm(4 * 4, inch, outch);
-
-    const float ktm[4][3] = {
-        {1.0f, 0.0f, 0.0f},
-        {1.0f / 2, 1.0f / 2, 1.0f / 2},
-        {1.0f / 2, -1.0f / 2, 1.0f / 2},
-        {0.0f, 0.0f, 1.0f}
-    };
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        for (int q = 0; q < inch; q++)
-        {
-            const float* kernel0 = (const float*)kernel + p * inch * 9 + q * 9;
-            float* kernel_tm0 = kernel_tm.channel(p).row(q);
-
-            // transform kernel
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            // h
-            float tmp[4][3];
-            for (int i = 0; i < 4; i++)
-            {
-                tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
-                tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
-                tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
-            }
-
-            // U
-            for (int j = 0; j < 4; j++)
-            {
-                float* tmpp = &tmp[j][0];
-
-                for (int i = 0; i < 4; i++)
-                {
-                    kernel_tm0[j * 4 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
-                }
-            }
-        }
-    }
-
-    // interleave
-    // src = 16-inch-outch
-    // dst = pb-pa-inch/pa-16-outch/pb
-    kernel_tm_packn.create(inch / packn, 16, outch / packn, (size_t)2u * packn * packn, packn * packn);
-
-    for (int q = 0; q + (packn - 1) < outch; q += packn)
-    {
-        Mat g0 = kernel_tm_packn.channel(q / packn);
-
-        for (int k = 0; k < 16; k++)
-        {
-            __fp16* g00 = g0.row<__fp16>(k);
-
-            for (int p = 0; p + (packn - 1) < inch; p += packn)
-            {
-                for (int i = 0; i < packn; i++)
-                {
-                    for (int j = 0; j < packn; j++)
-                    {
-                        const float* k00 = kernel_tm.channel(q + j).row(p + i);
-                        g00[0] = (__fp16)k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_packn_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel_tm, const Mat& bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    // pad to 2n+2
-    Mat bottom_blob_bordered = bottom_blob;
-
-    outw = (outw + 1) / 2 * 2;
-    outh = (outh + 1) / 2 * 2;
-
-    w = outw + 2;
-    h = outh + 2;
-    copy_make_border(bottom_blob, bottom_blob_bordered, 0, h - bottom_blob.h, 0, w - bottom_blob.w, BORDER_CONSTANT, 0.f, opt);
-
-    // BEGIN transform input
-    Mat bottom_blob_tm;
-    {
-        int w_tiles = outw / 2;
-        int h_tiles = outh / 2;
-        const int tiles = w_tiles * h_tiles;
-
-        bottom_blob_tm.create(tiles, 16, inch, elemsize, elempack, opt.workspace_allocator);
-        conv3x3s1_winograd23_transform_input_packn_fp16sa_rvv(bottom_blob_bordered, bottom_blob_tm, opt);
-    }
-    bottom_blob_bordered = Mat();
-    // END transform input
-
-    // BEGIN dot
-    Mat top_blob_tm;
-    convolution_winograd_dot_packn_fp16sa_rvv(bottom_blob_tm, outch, kernel_tm, top_blob_tm, opt);
-    // END dot
-
-    // BEGIN transform output
-    Mat top_blob_bordered;
-    if (outw == top_blob.w && outh == top_blob.h)
-    {
-        top_blob_bordered = top_blob;
-    }
-    else
-    {
-        top_blob_bordered.create(outw, outh, outch, elemsize, elempack, opt.workspace_allocator);
-    }
-    {
-        conv3x3s1_winograd23_transform_output_packn_fp16sa_rvv(top_blob_tm, top_blob_bordered, bias, opt);
-    }
-    // END transform output
-
-    // cut result pad
-    copy_cut_border(top_blob_bordered, top_blob, 0, top_blob_bordered.h - top_blob.h, 0, top_blob_bordered.w - top_blob.w, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_7x7_pack1ton.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_7x7_pack1ton.h
deleted file mode 100644
index 3605ed02..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_7x7_pack1ton.h
+++ /dev/null
@@ -1,1007 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv7x7s2_pack1ton_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        vfloat32m1_t _bias0 = bias ? vle32_v_f32m1(bias + p * packn, vl) : vfmv_v_f_f32m1(0.f, vl);
-        out0.fill(_bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-            const float* r3 = img0.row(3);
-            const float* r4 = img0.row(4);
-            const float* r5 = img0.row(5);
-            const float* r6 = img0.row(6);
-
-            const float* kptr = kernel.channel(p).row(q);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    vfloat32m1_t _sum0 = vle32_v_f32m1(outptr0, vl);
-                    vfloat32m1_t _sum1 = vle32_v_f32m1(outptr0 + packn, vl);
-                    vfloat32m1_t _sum2 = vle32_v_f32m1(outptr0 + packn * 2, vl);
-                    vfloat32m1_t _sum3 = vle32_v_f32m1(outptr0 + packn * 3, vl);
-                    vfloat32m1_t _sum4 = vle32_v_f32m1(outptr0 + packn * 4, vl);
-                    vfloat32m1_t _sum5 = vle32_v_f32m1(outptr0 + packn * 5, vl);
-                    vfloat32m1_t _sum6 = vle32_v_f32m1(outptr0 + packn * 6, vl);
-                    vfloat32m1_t _sum7 = vle32_v_f32m1(outptr0 + packn * 7, vl);
-
-                    vfloat32m1_t _k00 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k01 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k02 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k03 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k04 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k05 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k06 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[2], _k00, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[4], _k00, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[6], _k00, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r0[8], _k00, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r0[10], _k00, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r0[12], _k00, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r0[14], _k00, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[3], _k01, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[5], _k01, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[7], _k01, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r0[9], _k01, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r0[11], _k01, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r0[13], _k01, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r0[15], _k01, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[4], _k02, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[6], _k02, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[8], _k02, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r0[10], _k02, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r0[12], _k02, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r0[14], _k02, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r0[16], _k02, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[3], _k03, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[5], _k03, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[7], _k03, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[9], _k03, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r0[11], _k03, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r0[13], _k03, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r0[15], _k03, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r0[17], _k03, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[4], _k04, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[6], _k04, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[8], _k04, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[10], _k04, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r0[12], _k04, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r0[14], _k04, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r0[16], _k04, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r0[18], _k04, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[5], _k05, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[7], _k05, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[9], _k05, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[11], _k05, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r0[13], _k05, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r0[15], _k05, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r0[17], _k05, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r0[19], _k05, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[6], _k06, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[8], _k06, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[10], _k06, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[12], _k06, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r0[14], _k06, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r0[16], _k06, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r0[18], _k06, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r0[20], _k06, vl);
-
-                    vfloat32m1_t _k10 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k11 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k12 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k13 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k14 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k15 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k16 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[2], _k10, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[4], _k10, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[6], _k10, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r1[8], _k10, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r1[10], _k10, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r1[12], _k10, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r1[14], _k10, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[3], _k11, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[5], _k11, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[7], _k11, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r1[9], _k11, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r1[11], _k11, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r1[13], _k11, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r1[15], _k11, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[4], _k12, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[6], _k12, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[8], _k12, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r1[10], _k12, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r1[12], _k12, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r1[14], _k12, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r1[16], _k12, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[3], _k13, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[5], _k13, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[7], _k13, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[9], _k13, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r1[11], _k13, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r1[13], _k13, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r1[15], _k13, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r1[17], _k13, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[4], _k14, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[6], _k14, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[8], _k14, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[10], _k14, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r1[12], _k14, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r1[14], _k14, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r1[16], _k14, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r1[18], _k14, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[5], _k15, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[7], _k15, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[9], _k15, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[11], _k15, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r1[13], _k15, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r1[15], _k15, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r1[17], _k15, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r1[19], _k15, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[6], _k16, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[8], _k16, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[10], _k16, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[12], _k16, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r1[14], _k16, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r1[16], _k16, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r1[18], _k16, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r1[20], _k16, vl);
-
-                    vfloat32m1_t _k20 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k21 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k22 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k23 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k24 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k25 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k26 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[2], _k20, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[4], _k20, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[6], _k20, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r2[8], _k20, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r2[10], _k20, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r2[12], _k20, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r2[14], _k20, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[3], _k21, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[5], _k21, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[7], _k21, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r2[9], _k21, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r2[11], _k21, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r2[13], _k21, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r2[15], _k21, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[4], _k22, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[6], _k22, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[8], _k22, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r2[10], _k22, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r2[12], _k22, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r2[14], _k22, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r2[16], _k22, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[3], _k23, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[5], _k23, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[7], _k23, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[9], _k23, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r2[11], _k23, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r2[13], _k23, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r2[15], _k23, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r2[17], _k23, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[4], _k24, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[6], _k24, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[8], _k24, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[10], _k24, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r2[12], _k24, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r2[14], _k24, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r2[16], _k24, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r2[18], _k24, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[5], _k25, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[7], _k25, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[9], _k25, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[11], _k25, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r2[13], _k25, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r2[15], _k25, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r2[17], _k25, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r2[19], _k25, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[6], _k26, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[8], _k26, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[10], _k26, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[12], _k26, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r2[14], _k26, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r2[16], _k26, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r2[18], _k26, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r2[20], _k26, vl);
-
-                    vfloat32m1_t _k30 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k31 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k32 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k33 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k34 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k35 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k36 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[0], _k30, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r3[2], _k30, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r3[4], _k30, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r3[6], _k30, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r3[8], _k30, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r3[10], _k30, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r3[12], _k30, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r3[14], _k30, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[1], _k31, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r3[3], _k31, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r3[5], _k31, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r3[7], _k31, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r3[9], _k31, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r3[11], _k31, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r3[13], _k31, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r3[15], _k31, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[2], _k32, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r3[4], _k32, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r3[6], _k32, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r3[8], _k32, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r3[10], _k32, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r3[12], _k32, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r3[14], _k32, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r3[16], _k32, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[3], _k33, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r3[5], _k33, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r3[7], _k33, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r3[9], _k33, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r3[11], _k33, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r3[13], _k33, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r3[15], _k33, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r3[17], _k33, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[4], _k34, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r3[6], _k34, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r3[8], _k34, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r3[10], _k34, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r3[12], _k34, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r3[14], _k34, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r3[16], _k34, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r3[18], _k34, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[5], _k35, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r3[7], _k35, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r3[9], _k35, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r3[11], _k35, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r3[13], _k35, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r3[15], _k35, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r3[17], _k35, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r3[19], _k35, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[6], _k36, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r3[8], _k36, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r3[10], _k36, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r3[12], _k36, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r3[14], _k36, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r3[16], _k36, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r3[18], _k36, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r3[20], _k36, vl);
-
-                    vfloat32m1_t _k40 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k41 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k42 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k43 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k44 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k45 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k46 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[0], _k40, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r4[2], _k40, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r4[4], _k40, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r4[6], _k40, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r4[8], _k40, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r4[10], _k40, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r4[12], _k40, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r4[14], _k40, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[1], _k41, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r4[3], _k41, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r4[5], _k41, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r4[7], _k41, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r4[9], _k41, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r4[11], _k41, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r4[13], _k41, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r4[15], _k41, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[2], _k42, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r4[4], _k42, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r4[6], _k42, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r4[8], _k42, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r4[10], _k42, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r4[12], _k42, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r4[14], _k42, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r4[16], _k42, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[3], _k43, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r4[5], _k43, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r4[7], _k43, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r4[9], _k43, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r4[11], _k43, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r4[13], _k43, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r4[15], _k43, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r4[17], _k43, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[4], _k44, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r4[6], _k44, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r4[8], _k44, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r4[10], _k44, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r4[12], _k44, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r4[14], _k44, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r4[16], _k44, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r4[18], _k44, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[5], _k45, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r4[7], _k45, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r4[9], _k45, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r4[11], _k45, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r4[13], _k45, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r4[15], _k45, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r4[17], _k45, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r4[19], _k45, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[6], _k46, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r4[8], _k46, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r4[10], _k46, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r4[12], _k46, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r4[14], _k46, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r4[16], _k46, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r4[18], _k46, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r4[20], _k46, vl);
-
-                    vfloat32m1_t _k50 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k51 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k52 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k53 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k54 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k55 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k56 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[0], _k50, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r5[2], _k50, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r5[4], _k50, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r5[6], _k50, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r5[8], _k50, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r5[10], _k50, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r5[12], _k50, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r5[14], _k50, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[1], _k51, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r5[3], _k51, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r5[5], _k51, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r5[7], _k51, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r5[9], _k51, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r5[11], _k51, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r5[13], _k51, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r5[15], _k51, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[2], _k52, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r5[4], _k52, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r5[6], _k52, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r5[8], _k52, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r5[10], _k52, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r5[12], _k52, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r5[14], _k52, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r5[16], _k52, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[3], _k53, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r5[5], _k53, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r5[7], _k53, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r5[9], _k53, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r5[11], _k53, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r5[13], _k53, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r5[15], _k53, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r5[17], _k53, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[4], _k54, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r5[6], _k54, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r5[8], _k54, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r5[10], _k54, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r5[12], _k54, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r5[14], _k54, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r5[16], _k54, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r5[18], _k54, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[5], _k55, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r5[7], _k55, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r5[9], _k55, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r5[11], _k55, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r5[13], _k55, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r5[15], _k55, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r5[17], _k55, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r5[19], _k55, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[6], _k56, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r5[8], _k56, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r5[10], _k56, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r5[12], _k56, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r5[14], _k56, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r5[16], _k56, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r5[18], _k56, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r5[20], _k56, vl);
-
-                    vfloat32m1_t _k60 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k61 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k62 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k63 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k64 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k65 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k66 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr -= packn * 42;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[0], _k60, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r6[2], _k60, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r6[4], _k60, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r6[6], _k60, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r6[8], _k60, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r6[10], _k60, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r6[12], _k60, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r6[14], _k60, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[1], _k61, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r6[3], _k61, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r6[5], _k61, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r6[7], _k61, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r6[9], _k61, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r6[11], _k61, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r6[13], _k61, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r6[15], _k61, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[2], _k62, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r6[4], _k62, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r6[6], _k62, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r6[8], _k62, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r6[10], _k62, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r6[12], _k62, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r6[14], _k62, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r6[16], _k62, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[3], _k63, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r6[5], _k63, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r6[7], _k63, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r6[9], _k63, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r6[11], _k63, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r6[13], _k63, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r6[15], _k63, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r6[17], _k63, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[4], _k64, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r6[6], _k64, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r6[8], _k64, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r6[10], _k64, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r6[12], _k64, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r6[14], _k64, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r6[16], _k64, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r6[18], _k64, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[5], _k65, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r6[7], _k65, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r6[9], _k65, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r6[11], _k65, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r6[13], _k65, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r6[15], _k65, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r6[17], _k65, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r6[19], _k65, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[6], _k66, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r6[8], _k66, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r6[10], _k66, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r6[12], _k66, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, r6[14], _k66, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, r6[16], _k66, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, r6[18], _k66, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, r6[20], _k66, vl);
-
-                    vse32_v_f32m1(outptr0, _sum0, vl);
-                    vse32_v_f32m1(outptr0 + packn, _sum1, vl);
-                    vse32_v_f32m1(outptr0 + packn * 2, _sum2, vl);
-                    vse32_v_f32m1(outptr0 + packn * 3, _sum3, vl);
-                    vse32_v_f32m1(outptr0 + packn * 4, _sum4, vl);
-                    vse32_v_f32m1(outptr0 + packn * 5, _sum5, vl);
-                    vse32_v_f32m1(outptr0 + packn * 6, _sum6, vl);
-                    vse32_v_f32m1(outptr0 + packn * 7, _sum7, vl);
-
-                    outptr0 += packn * 8;
-
-                    r0 += 16;
-                    r1 += 16;
-                    r2 += 16;
-                    r3 += 16;
-                    r4 += 16;
-                    r5 += 16;
-                    r6 += 16;
-                }
-                for (; j + 3 < outw; j += 4)
-                {
-                    vfloat32m1_t _sum0 = vle32_v_f32m1(outptr0, vl);
-                    vfloat32m1_t _sum1 = vle32_v_f32m1(outptr0 + packn, vl);
-                    vfloat32m1_t _sum2 = vle32_v_f32m1(outptr0 + packn * 2, vl);
-                    vfloat32m1_t _sum3 = vle32_v_f32m1(outptr0 + packn * 3, vl);
-
-                    vfloat32m1_t _k00 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k01 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k02 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k03 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k04 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k05 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k06 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[2], _k00, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[4], _k00, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[6], _k00, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[3], _k01, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[5], _k01, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[7], _k01, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[4], _k02, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[6], _k02, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[8], _k02, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[3], _k03, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[5], _k03, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[7], _k03, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[9], _k03, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[4], _k04, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[6], _k04, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[8], _k04, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[10], _k04, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[5], _k05, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[7], _k05, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[9], _k05, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[11], _k05, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[6], _k06, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r0[8], _k06, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r0[10], _k06, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r0[12], _k06, vl);
-
-                    vfloat32m1_t _k10 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k11 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k12 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k13 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k14 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k15 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k16 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[2], _k10, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[4], _k10, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[6], _k10, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[3], _k11, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[5], _k11, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[7], _k11, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[4], _k12, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[6], _k12, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[8], _k12, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[3], _k13, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[5], _k13, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[7], _k13, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[9], _k13, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[4], _k14, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[6], _k14, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[8], _k14, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[10], _k14, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[5], _k15, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[7], _k15, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[9], _k15, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[11], _k15, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[6], _k16, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r1[8], _k16, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r1[10], _k16, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r1[12], _k16, vl);
-
-                    vfloat32m1_t _k20 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k21 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k22 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k23 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k24 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k25 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k26 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[2], _k20, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[4], _k20, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[6], _k20, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[3], _k21, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[5], _k21, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[7], _k21, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[4], _k22, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[6], _k22, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[8], _k22, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[3], _k23, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[5], _k23, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[7], _k23, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[9], _k23, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[4], _k24, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[6], _k24, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[8], _k24, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[10], _k24, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[5], _k25, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[7], _k25, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[9], _k25, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[11], _k25, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[6], _k26, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r2[8], _k26, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r2[10], _k26, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r2[12], _k26, vl);
-
-                    vfloat32m1_t _k30 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k31 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k32 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k33 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k34 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k35 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k36 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[0], _k30, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r3[2], _k30, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r3[4], _k30, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r3[6], _k30, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[1], _k31, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r3[3], _k31, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r3[5], _k31, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r3[7], _k31, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[2], _k32, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r3[4], _k32, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r3[6], _k32, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r3[8], _k32, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[3], _k33, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r3[5], _k33, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r3[7], _k33, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r3[9], _k33, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[4], _k34, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r3[6], _k34, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r3[8], _k34, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r3[10], _k34, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[5], _k35, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r3[7], _k35, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r3[9], _k35, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r3[11], _k35, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[6], _k36, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r3[8], _k36, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r3[10], _k36, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r3[12], _k36, vl);
-
-                    vfloat32m1_t _k40 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k41 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k42 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k43 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k44 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k45 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k46 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[0], _k40, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r4[2], _k40, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r4[4], _k40, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r4[6], _k40, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[1], _k41, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r4[3], _k41, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r4[5], _k41, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r4[7], _k41, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[2], _k42, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r4[4], _k42, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r4[6], _k42, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r4[8], _k42, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[3], _k43, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r4[5], _k43, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r4[7], _k43, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r4[9], _k43, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[4], _k44, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r4[6], _k44, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r4[8], _k44, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r4[10], _k44, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[5], _k45, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r4[7], _k45, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r4[9], _k45, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r4[11], _k45, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[6], _k46, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r4[8], _k46, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r4[10], _k46, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r4[12], _k46, vl);
-
-                    vfloat32m1_t _k50 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k51 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k52 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k53 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k54 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k55 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k56 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[0], _k50, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r5[2], _k50, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r5[4], _k50, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r5[6], _k50, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[1], _k51, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r5[3], _k51, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r5[5], _k51, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r5[7], _k51, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[2], _k52, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r5[4], _k52, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r5[6], _k52, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r5[8], _k52, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[3], _k53, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r5[5], _k53, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r5[7], _k53, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r5[9], _k53, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[4], _k54, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r5[6], _k54, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r5[8], _k54, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r5[10], _k54, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[5], _k55, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r5[7], _k55, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r5[9], _k55, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r5[11], _k55, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[6], _k56, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r5[8], _k56, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r5[10], _k56, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r5[12], _k56, vl);
-
-                    vfloat32m1_t _k60 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k61 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k62 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k63 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k64 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k65 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k66 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr -= packn * 42;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[0], _k60, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r6[2], _k60, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r6[4], _k60, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r6[6], _k60, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[1], _k61, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r6[3], _k61, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r6[5], _k61, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r6[7], _k61, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[2], _k62, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r6[4], _k62, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r6[6], _k62, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r6[8], _k62, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[3], _k63, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r6[5], _k63, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r6[7], _k63, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r6[9], _k63, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[4], _k64, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r6[6], _k64, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r6[8], _k64, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r6[10], _k64, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[5], _k65, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r6[7], _k65, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r6[9], _k65, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r6[11], _k65, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[6], _k66, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, r6[8], _k66, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, r6[10], _k66, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, r6[12], _k66, vl);
-
-                    vse32_v_f32m1(outptr0, _sum0, vl);
-                    vse32_v_f32m1(outptr0 + packn, _sum1, vl);
-                    vse32_v_f32m1(outptr0 + packn * 2, _sum2, vl);
-                    vse32_v_f32m1(outptr0 + packn * 3, _sum3, vl);
-
-                    outptr0 += packn * 4;
-
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                    r3 += 8;
-                    r4 += 8;
-                    r5 += 8;
-                    r6 += 8;
-                }
-                for (; j < outw; j++)
-                {
-                    vfloat32m1_t _sum0 = vle32_v_f32m1(outptr0, vl);
-
-                    vfloat32m1_t _k00 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k01 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k02 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k03 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k04 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k05 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k06 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[0], _k00, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[1], _k01, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[2], _k02, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[3], _k03, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[4], _k04, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[5], _k05, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r0[6], _k06, vl);
-
-                    vfloat32m1_t _k10 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k11 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k12 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k13 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k14 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k15 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k16 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[0], _k10, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[1], _k11, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[2], _k12, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[3], _k13, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[4], _k14, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[5], _k15, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r1[6], _k16, vl);
-
-                    vfloat32m1_t _k20 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k21 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k22 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k23 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k24 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k25 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k26 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[0], _k20, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[1], _k21, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[2], _k22, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[3], _k23, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[4], _k24, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[5], _k25, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r2[6], _k26, vl);
-
-                    vfloat32m1_t _k30 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k31 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k32 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k33 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k34 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k35 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k36 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[0], _k30, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[1], _k31, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[2], _k32, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[3], _k33, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[4], _k34, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[5], _k35, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r3[6], _k36, vl);
-
-                    vfloat32m1_t _k40 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k41 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k42 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k43 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k44 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k45 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k46 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[0], _k40, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[1], _k41, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[2], _k42, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[3], _k43, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[4], _k44, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[5], _k45, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r4[6], _k46, vl);
-
-                    vfloat32m1_t _k50 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k51 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k52 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k53 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k54 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k55 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k56 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[0], _k50, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[1], _k51, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[2], _k52, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[3], _k53, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[4], _k54, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[5], _k55, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r5[6], _k56, vl);
-
-                    vfloat32m1_t _k60 = vle32_v_f32m1(kptr, vl);
-                    vfloat32m1_t _k61 = vle32_v_f32m1(kptr + packn, vl);
-                    vfloat32m1_t _k62 = vle32_v_f32m1(kptr + packn * 2, vl);
-                    vfloat32m1_t _k63 = vle32_v_f32m1(kptr + packn * 3, vl);
-                    vfloat32m1_t _k64 = vle32_v_f32m1(kptr + packn * 4, vl);
-                    vfloat32m1_t _k65 = vle32_v_f32m1(kptr + packn * 5, vl);
-                    vfloat32m1_t _k66 = vle32_v_f32m1(kptr + packn * 6, vl);
-
-                    kptr -= packn * 42;
-
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[0], _k60, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[1], _k61, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[2], _k62, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[3], _k63, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[4], _k64, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[5], _k65, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, r6[6], _k66, vl);
-
-                    vse32_v_f32m1(outptr0, _sum0, vl);
-
-                    outptr0 += packn;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    r3 += 2;
-                    r4 += 2;
-                    r5 += 2;
-                    r6 += 2;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-                r4 += tailstep;
-                r5 += tailstep;
-                r6 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_7x7_pack1ton_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_7x7_pack1ton_fp16s.h
deleted file mode 100644
index 01804bf3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_7x7_pack1ton_fp16s.h
+++ /dev/null
@@ -1,1007 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv7x7s2_pack1ton_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        vfloat16m1_t _bias0 = bias ? vle16_v_f16m1(bias + p * packn, vl) : vfmv_v_f_f16m1(0.f, vl);
-        out0.fill(_bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            __fp16* outptr0 = out0;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const __fp16* r0 = img0.row<const __fp16>(0);
-            const __fp16* r1 = img0.row<const __fp16>(1);
-            const __fp16* r2 = img0.row<const __fp16>(2);
-            const __fp16* r3 = img0.row<const __fp16>(3);
-            const __fp16* r4 = img0.row<const __fp16>(4);
-            const __fp16* r5 = img0.row<const __fp16>(5);
-            const __fp16* r6 = img0.row<const __fp16>(6);
-
-            const __fp16* kptr = kernel.channel(p).row<const __fp16>(q);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    vfloat16m1_t _sum0 = vle16_v_f16m1(outptr0, vl);
-                    vfloat16m1_t _sum1 = vle16_v_f16m1(outptr0 + packn, vl);
-                    vfloat16m1_t _sum2 = vle16_v_f16m1(outptr0 + packn * 2, vl);
-                    vfloat16m1_t _sum3 = vle16_v_f16m1(outptr0 + packn * 3, vl);
-                    vfloat16m1_t _sum4 = vle16_v_f16m1(outptr0 + packn * 4, vl);
-                    vfloat16m1_t _sum5 = vle16_v_f16m1(outptr0 + packn * 5, vl);
-                    vfloat16m1_t _sum6 = vle16_v_f16m1(outptr0 + packn * 6, vl);
-                    vfloat16m1_t _sum7 = vle16_v_f16m1(outptr0 + packn * 7, vl);
-
-                    vfloat16m1_t _k00 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k01 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k02 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k03 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k04 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k05 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k06 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[2], _k00, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[4], _k00, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[6], _k00, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r0[8], _k00, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r0[10], _k00, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r0[12], _k00, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r0[14], _k00, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[3], _k01, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[5], _k01, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[7], _k01, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r0[9], _k01, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r0[11], _k01, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r0[13], _k01, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r0[15], _k01, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[4], _k02, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[6], _k02, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[8], _k02, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r0[10], _k02, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r0[12], _k02, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r0[14], _k02, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r0[16], _k02, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[3], _k03, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[5], _k03, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[7], _k03, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[9], _k03, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r0[11], _k03, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r0[13], _k03, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r0[15], _k03, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r0[17], _k03, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[4], _k04, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[6], _k04, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[8], _k04, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[10], _k04, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r0[12], _k04, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r0[14], _k04, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r0[16], _k04, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r0[18], _k04, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[5], _k05, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[7], _k05, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[9], _k05, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[11], _k05, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r0[13], _k05, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r0[15], _k05, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r0[17], _k05, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r0[19], _k05, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[6], _k06, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[8], _k06, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[10], _k06, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[12], _k06, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r0[14], _k06, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r0[16], _k06, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r0[18], _k06, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r0[20], _k06, vl);
-
-                    vfloat16m1_t _k10 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k11 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k12 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k13 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k14 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k15 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k16 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[2], _k10, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[4], _k10, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[6], _k10, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r1[8], _k10, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r1[10], _k10, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r1[12], _k10, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r1[14], _k10, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[3], _k11, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[5], _k11, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[7], _k11, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r1[9], _k11, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r1[11], _k11, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r1[13], _k11, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r1[15], _k11, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[4], _k12, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[6], _k12, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[8], _k12, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r1[10], _k12, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r1[12], _k12, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r1[14], _k12, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r1[16], _k12, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[3], _k13, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[5], _k13, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[7], _k13, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[9], _k13, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r1[11], _k13, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r1[13], _k13, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r1[15], _k13, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r1[17], _k13, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[4], _k14, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[6], _k14, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[8], _k14, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[10], _k14, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r1[12], _k14, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r1[14], _k14, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r1[16], _k14, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r1[18], _k14, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[5], _k15, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[7], _k15, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[9], _k15, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[11], _k15, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r1[13], _k15, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r1[15], _k15, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r1[17], _k15, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r1[19], _k15, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[6], _k16, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[8], _k16, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[10], _k16, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[12], _k16, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r1[14], _k16, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r1[16], _k16, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r1[18], _k16, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r1[20], _k16, vl);
-
-                    vfloat16m1_t _k20 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k21 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k22 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k23 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k24 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k25 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k26 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[2], _k20, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[4], _k20, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[6], _k20, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r2[8], _k20, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r2[10], _k20, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r2[12], _k20, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r2[14], _k20, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[3], _k21, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[5], _k21, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[7], _k21, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r2[9], _k21, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r2[11], _k21, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r2[13], _k21, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r2[15], _k21, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[4], _k22, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[6], _k22, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[8], _k22, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r2[10], _k22, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r2[12], _k22, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r2[14], _k22, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r2[16], _k22, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[3], _k23, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[5], _k23, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[7], _k23, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[9], _k23, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r2[11], _k23, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r2[13], _k23, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r2[15], _k23, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r2[17], _k23, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[4], _k24, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[6], _k24, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[8], _k24, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[10], _k24, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r2[12], _k24, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r2[14], _k24, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r2[16], _k24, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r2[18], _k24, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[5], _k25, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[7], _k25, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[9], _k25, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[11], _k25, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r2[13], _k25, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r2[15], _k25, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r2[17], _k25, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r2[19], _k25, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[6], _k26, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[8], _k26, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[10], _k26, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[12], _k26, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r2[14], _k26, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r2[16], _k26, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r2[18], _k26, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r2[20], _k26, vl);
-
-                    vfloat16m1_t _k30 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k31 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k32 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k33 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k34 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k35 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k36 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[0], _k30, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r3[2], _k30, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r3[4], _k30, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r3[6], _k30, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r3[8], _k30, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r3[10], _k30, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r3[12], _k30, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r3[14], _k30, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[1], _k31, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r3[3], _k31, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r3[5], _k31, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r3[7], _k31, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r3[9], _k31, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r3[11], _k31, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r3[13], _k31, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r3[15], _k31, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[2], _k32, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r3[4], _k32, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r3[6], _k32, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r3[8], _k32, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r3[10], _k32, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r3[12], _k32, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r3[14], _k32, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r3[16], _k32, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[3], _k33, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r3[5], _k33, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r3[7], _k33, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r3[9], _k33, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r3[11], _k33, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r3[13], _k33, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r3[15], _k33, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r3[17], _k33, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[4], _k34, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r3[6], _k34, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r3[8], _k34, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r3[10], _k34, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r3[12], _k34, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r3[14], _k34, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r3[16], _k34, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r3[18], _k34, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[5], _k35, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r3[7], _k35, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r3[9], _k35, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r3[11], _k35, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r3[13], _k35, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r3[15], _k35, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r3[17], _k35, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r3[19], _k35, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[6], _k36, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r3[8], _k36, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r3[10], _k36, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r3[12], _k36, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r3[14], _k36, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r3[16], _k36, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r3[18], _k36, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r3[20], _k36, vl);
-
-                    vfloat16m1_t _k40 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k41 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k42 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k43 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k44 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k45 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k46 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[0], _k40, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r4[2], _k40, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r4[4], _k40, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r4[6], _k40, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r4[8], _k40, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r4[10], _k40, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r4[12], _k40, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r4[14], _k40, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[1], _k41, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r4[3], _k41, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r4[5], _k41, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r4[7], _k41, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r4[9], _k41, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r4[11], _k41, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r4[13], _k41, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r4[15], _k41, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[2], _k42, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r4[4], _k42, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r4[6], _k42, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r4[8], _k42, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r4[10], _k42, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r4[12], _k42, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r4[14], _k42, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r4[16], _k42, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[3], _k43, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r4[5], _k43, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r4[7], _k43, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r4[9], _k43, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r4[11], _k43, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r4[13], _k43, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r4[15], _k43, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r4[17], _k43, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[4], _k44, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r4[6], _k44, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r4[8], _k44, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r4[10], _k44, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r4[12], _k44, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r4[14], _k44, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r4[16], _k44, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r4[18], _k44, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[5], _k45, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r4[7], _k45, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r4[9], _k45, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r4[11], _k45, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r4[13], _k45, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r4[15], _k45, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r4[17], _k45, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r4[19], _k45, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[6], _k46, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r4[8], _k46, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r4[10], _k46, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r4[12], _k46, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r4[14], _k46, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r4[16], _k46, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r4[18], _k46, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r4[20], _k46, vl);
-
-                    vfloat16m1_t _k50 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k51 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k52 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k53 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k54 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k55 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k56 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[0], _k50, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r5[2], _k50, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r5[4], _k50, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r5[6], _k50, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r5[8], _k50, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r5[10], _k50, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r5[12], _k50, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r5[14], _k50, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[1], _k51, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r5[3], _k51, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r5[5], _k51, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r5[7], _k51, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r5[9], _k51, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r5[11], _k51, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r5[13], _k51, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r5[15], _k51, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[2], _k52, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r5[4], _k52, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r5[6], _k52, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r5[8], _k52, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r5[10], _k52, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r5[12], _k52, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r5[14], _k52, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r5[16], _k52, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[3], _k53, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r5[5], _k53, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r5[7], _k53, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r5[9], _k53, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r5[11], _k53, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r5[13], _k53, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r5[15], _k53, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r5[17], _k53, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[4], _k54, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r5[6], _k54, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r5[8], _k54, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r5[10], _k54, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r5[12], _k54, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r5[14], _k54, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r5[16], _k54, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r5[18], _k54, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[5], _k55, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r5[7], _k55, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r5[9], _k55, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r5[11], _k55, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r5[13], _k55, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r5[15], _k55, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r5[17], _k55, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r5[19], _k55, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[6], _k56, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r5[8], _k56, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r5[10], _k56, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r5[12], _k56, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r5[14], _k56, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r5[16], _k56, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r5[18], _k56, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r5[20], _k56, vl);
-
-                    vfloat16m1_t _k60 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k61 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k62 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k63 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k64 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k65 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k66 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr -= packn * 42;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[0], _k60, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r6[2], _k60, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r6[4], _k60, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r6[6], _k60, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r6[8], _k60, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r6[10], _k60, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r6[12], _k60, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r6[14], _k60, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[1], _k61, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r6[3], _k61, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r6[5], _k61, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r6[7], _k61, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r6[9], _k61, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r6[11], _k61, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r6[13], _k61, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r6[15], _k61, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[2], _k62, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r6[4], _k62, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r6[6], _k62, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r6[8], _k62, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r6[10], _k62, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r6[12], _k62, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r6[14], _k62, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r6[16], _k62, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[3], _k63, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r6[5], _k63, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r6[7], _k63, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r6[9], _k63, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r6[11], _k63, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r6[13], _k63, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r6[15], _k63, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r6[17], _k63, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[4], _k64, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r6[6], _k64, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r6[8], _k64, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r6[10], _k64, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r6[12], _k64, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r6[14], _k64, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r6[16], _k64, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r6[18], _k64, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[5], _k65, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r6[7], _k65, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r6[9], _k65, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r6[11], _k65, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r6[13], _k65, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r6[15], _k65, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r6[17], _k65, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r6[19], _k65, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[6], _k66, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r6[8], _k66, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r6[10], _k66, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r6[12], _k66, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, r6[14], _k66, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, r6[16], _k66, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, r6[18], _k66, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, r6[20], _k66, vl);
-
-                    vse16_v_f16m1(outptr0, _sum0, vl);
-                    vse16_v_f16m1(outptr0 + packn, _sum1, vl);
-                    vse16_v_f16m1(outptr0 + packn * 2, _sum2, vl);
-                    vse16_v_f16m1(outptr0 + packn * 3, _sum3, vl);
-                    vse16_v_f16m1(outptr0 + packn * 4, _sum4, vl);
-                    vse16_v_f16m1(outptr0 + packn * 5, _sum5, vl);
-                    vse16_v_f16m1(outptr0 + packn * 6, _sum6, vl);
-                    vse16_v_f16m1(outptr0 + packn * 7, _sum7, vl);
-
-                    outptr0 += packn * 8;
-
-                    r0 += 16;
-                    r1 += 16;
-                    r2 += 16;
-                    r3 += 16;
-                    r4 += 16;
-                    r5 += 16;
-                    r6 += 16;
-                }
-                for (; j + 3 < outw; j += 4)
-                {
-                    vfloat16m1_t _sum0 = vle16_v_f16m1(outptr0, vl);
-                    vfloat16m1_t _sum1 = vle16_v_f16m1(outptr0 + packn, vl);
-                    vfloat16m1_t _sum2 = vle16_v_f16m1(outptr0 + packn * 2, vl);
-                    vfloat16m1_t _sum3 = vle16_v_f16m1(outptr0 + packn * 3, vl);
-
-                    vfloat16m1_t _k00 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k01 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k02 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k03 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k04 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k05 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k06 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[0], _k00, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[2], _k00, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[4], _k00, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[6], _k00, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[1], _k01, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[3], _k01, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[5], _k01, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[7], _k01, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[2], _k02, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[4], _k02, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[6], _k02, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[8], _k02, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[3], _k03, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[5], _k03, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[7], _k03, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[9], _k03, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[4], _k04, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[6], _k04, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[8], _k04, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[10], _k04, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[5], _k05, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[7], _k05, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[9], _k05, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[11], _k05, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[6], _k06, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r0[8], _k06, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r0[10], _k06, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r0[12], _k06, vl);
-
-                    vfloat16m1_t _k10 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k11 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k12 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k13 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k14 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k15 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k16 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[0], _k10, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[2], _k10, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[4], _k10, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[6], _k10, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[1], _k11, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[3], _k11, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[5], _k11, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[7], _k11, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[2], _k12, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[4], _k12, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[6], _k12, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[8], _k12, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[3], _k13, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[5], _k13, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[7], _k13, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[9], _k13, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[4], _k14, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[6], _k14, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[8], _k14, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[10], _k14, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[5], _k15, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[7], _k15, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[9], _k15, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[11], _k15, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[6], _k16, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r1[8], _k16, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r1[10], _k16, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r1[12], _k16, vl);
-
-                    vfloat16m1_t _k20 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k21 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k22 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k23 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k24 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k25 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k26 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[0], _k20, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[2], _k20, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[4], _k20, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[6], _k20, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[1], _k21, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[3], _k21, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[5], _k21, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[7], _k21, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[2], _k22, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[4], _k22, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[6], _k22, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[8], _k22, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[3], _k23, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[5], _k23, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[7], _k23, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[9], _k23, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[4], _k24, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[6], _k24, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[8], _k24, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[10], _k24, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[5], _k25, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[7], _k25, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[9], _k25, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[11], _k25, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[6], _k26, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r2[8], _k26, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r2[10], _k26, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r2[12], _k26, vl);
-
-                    vfloat16m1_t _k30 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k31 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k32 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k33 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k34 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k35 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k36 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[0], _k30, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r3[2], _k30, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r3[4], _k30, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r3[6], _k30, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[1], _k31, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r3[3], _k31, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r3[5], _k31, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r3[7], _k31, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[2], _k32, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r3[4], _k32, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r3[6], _k32, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r3[8], _k32, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[3], _k33, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r3[5], _k33, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r3[7], _k33, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r3[9], _k33, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[4], _k34, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r3[6], _k34, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r3[8], _k34, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r3[10], _k34, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[5], _k35, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r3[7], _k35, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r3[9], _k35, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r3[11], _k35, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[6], _k36, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r3[8], _k36, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r3[10], _k36, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r3[12], _k36, vl);
-
-                    vfloat16m1_t _k40 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k41 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k42 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k43 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k44 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k45 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k46 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[0], _k40, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r4[2], _k40, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r4[4], _k40, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r4[6], _k40, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[1], _k41, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r4[3], _k41, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r4[5], _k41, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r4[7], _k41, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[2], _k42, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r4[4], _k42, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r4[6], _k42, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r4[8], _k42, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[3], _k43, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r4[5], _k43, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r4[7], _k43, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r4[9], _k43, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[4], _k44, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r4[6], _k44, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r4[8], _k44, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r4[10], _k44, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[5], _k45, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r4[7], _k45, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r4[9], _k45, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r4[11], _k45, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[6], _k46, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r4[8], _k46, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r4[10], _k46, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r4[12], _k46, vl);
-
-                    vfloat16m1_t _k50 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k51 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k52 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k53 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k54 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k55 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k56 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[0], _k50, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r5[2], _k50, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r5[4], _k50, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r5[6], _k50, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[1], _k51, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r5[3], _k51, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r5[5], _k51, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r5[7], _k51, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[2], _k52, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r5[4], _k52, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r5[6], _k52, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r5[8], _k52, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[3], _k53, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r5[5], _k53, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r5[7], _k53, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r5[9], _k53, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[4], _k54, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r5[6], _k54, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r5[8], _k54, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r5[10], _k54, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[5], _k55, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r5[7], _k55, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r5[9], _k55, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r5[11], _k55, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[6], _k56, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r5[8], _k56, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r5[10], _k56, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r5[12], _k56, vl);
-
-                    vfloat16m1_t _k60 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k61 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k62 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k63 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k64 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k65 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k66 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr -= packn * 42;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[0], _k60, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r6[2], _k60, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r6[4], _k60, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r6[6], _k60, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[1], _k61, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r6[3], _k61, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r6[5], _k61, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r6[7], _k61, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[2], _k62, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r6[4], _k62, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r6[6], _k62, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r6[8], _k62, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[3], _k63, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r6[5], _k63, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r6[7], _k63, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r6[9], _k63, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[4], _k64, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r6[6], _k64, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r6[8], _k64, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r6[10], _k64, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[5], _k65, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r6[7], _k65, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r6[9], _k65, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r6[11], _k65, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[6], _k66, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, r6[8], _k66, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, r6[10], _k66, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, r6[12], _k66, vl);
-
-                    vse16_v_f16m1(outptr0, _sum0, vl);
-                    vse16_v_f16m1(outptr0 + packn, _sum1, vl);
-                    vse16_v_f16m1(outptr0 + packn * 2, _sum2, vl);
-                    vse16_v_f16m1(outptr0 + packn * 3, _sum3, vl);
-
-                    outptr0 += packn * 4;
-
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                    r3 += 8;
-                    r4 += 8;
-                    r5 += 8;
-                    r6 += 8;
-                }
-                for (; j < outw; j++)
-                {
-                    vfloat16m1_t _sum0 = vle16_v_f16m1(outptr0, vl);
-
-                    vfloat16m1_t _k00 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k01 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k02 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k03 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k04 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k05 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k06 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[0], _k00, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[1], _k01, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[2], _k02, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[3], _k03, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[4], _k04, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[5], _k05, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r0[6], _k06, vl);
-
-                    vfloat16m1_t _k10 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k11 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k12 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k13 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k14 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k15 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k16 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[0], _k10, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[1], _k11, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[2], _k12, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[3], _k13, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[4], _k14, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[5], _k15, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r1[6], _k16, vl);
-
-                    vfloat16m1_t _k20 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k21 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k22 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k23 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k24 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k25 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k26 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[0], _k20, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[1], _k21, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[2], _k22, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[3], _k23, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[4], _k24, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[5], _k25, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r2[6], _k26, vl);
-
-                    vfloat16m1_t _k30 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k31 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k32 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k33 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k34 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k35 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k36 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[0], _k30, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[1], _k31, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[2], _k32, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[3], _k33, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[4], _k34, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[5], _k35, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r3[6], _k36, vl);
-
-                    vfloat16m1_t _k40 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k41 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k42 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k43 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k44 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k45 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k46 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[0], _k40, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[1], _k41, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[2], _k42, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[3], _k43, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[4], _k44, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[5], _k45, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r4[6], _k46, vl);
-
-                    vfloat16m1_t _k50 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k51 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k52 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k53 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k54 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k55 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k56 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr += packn * 7;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[0], _k50, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[1], _k51, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[2], _k52, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[3], _k53, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[4], _k54, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[5], _k55, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r5[6], _k56, vl);
-
-                    vfloat16m1_t _k60 = vle16_v_f16m1(kptr, vl);
-                    vfloat16m1_t _k61 = vle16_v_f16m1(kptr + packn, vl);
-                    vfloat16m1_t _k62 = vle16_v_f16m1(kptr + packn * 2, vl);
-                    vfloat16m1_t _k63 = vle16_v_f16m1(kptr + packn * 3, vl);
-                    vfloat16m1_t _k64 = vle16_v_f16m1(kptr + packn * 4, vl);
-                    vfloat16m1_t _k65 = vle16_v_f16m1(kptr + packn * 5, vl);
-                    vfloat16m1_t _k66 = vle16_v_f16m1(kptr + packn * 6, vl);
-
-                    kptr -= packn * 42;
-
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[0], _k60, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[1], _k61, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[2], _k62, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[3], _k63, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[4], _k64, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[5], _k65, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, r6[6], _k66, vl);
-
-                    vse16_v_f16m1(outptr0, _sum0, vl);
-
-                    outptr0 += packn;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    r3 += 2;
-                    r4 += 2;
-                    r5 += 2;
-                    r6 += 2;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-                r4 += tailstep;
-                r5 += tailstep;
-                r6 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_fp16s.h
deleted file mode 100644
index 7b08a985..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_fp16s.h
+++ /dev/null
@@ -1,90 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_fp16s(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                const __fp16* kptr = weight_data_fp16.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const __fp16* sptr = m.row<__fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float val = (float)sptr[space_ofs[k]];
-                        float w = (float)kptr[k];
-                        sum += val * w;
-                    }
-
-                    kptr += maxk;
-                }
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[j] = (__fp16)sum;
-            }
-
-            outptr += outw;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_pack1ton.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_pack1ton.h
deleted file mode 100644
index 15eec7ba..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_pack1ton.h
+++ /dev/null
@@ -1,93 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_pack1ton_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_pack1ton, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
-
-                if (bias_data_ptr)
-                {
-                    _sum = vle32_v_f32m1(bias_data_ptr + p * packn, vl);
-                }
-
-                const float* kptr = (const float*)weight_data_pack1ton + maxk * channels * p * packn;
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const float* sptr = m.row(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++) // 29.23
-                    {
-                        float val = sptr[space_ofs[k]];
-                        vfloat32m1_t _w = vle32_v_f32m1(kptr, vl);
-                        _sum = vfmacc_vf_f32m1(_sum, val, _w, vl);
-
-                        kptr += packn;
-                    }
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params, vl);
-
-                vse32_v_f32m1(outptr + j * packn, _sum, vl);
-            }
-
-            outptr += outw * packn;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_pack1ton_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_pack1ton_fp16s.h
deleted file mode 100644
index 6f8c649e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_pack1ton_fp16s.h
+++ /dev/null
@@ -1,173 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_pack1ton_fp16s_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat32m2_t _sum = vfmv_v_f_f32m2(0.f, vl);
-
-                if (bias_data_ptr)
-                {
-                    _sum = vle32_v_f32m2(bias_data_ptr + p * packn, vl);
-                }
-
-                const __fp16* kptr = weight_data_fp16.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const __fp16* sptr = m.row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        float val = (float)sptr[space_ofs[k]];
-                        vfloat16m1_t _w = vle16_v_f16m1(kptr, vl);
-                        _sum = vfwmacc_vf_f32m2(_sum, val, _w, vl);
-
-                        kptr += packn;
-                    }
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params, vl);
-
-                vse16_v_f16m1(outptr + j * packn, vfncvt_f_f_w_f16m1(_sum, vl), vl);
-            }
-
-            outptr += outw * packn;
-        }
-    }
-}
-
-static void convolution_pack1ton_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data_fp16, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const __fp16* bias_data_ptr = bias_data_fp16;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat16m1_t _sum = vfmv_v_f_f16m1(0.f, vl);
-
-                if (bias_data_ptr)
-                {
-                    _sum = vle16_v_f16m1(bias_data_ptr + p * packn, vl);
-                }
-
-                const __fp16* kptr = weight_data_fp16.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const __fp16* sptr = m.row<const __fp16>(i * stride_h) + j * stride_w;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        __fp16 val = sptr[space_ofs[k]];
-                        vfloat16m1_t _w = vle16_v_f16m1(kptr, vl);
-                        _sum = vfmacc_vf_f16m1(_sum, val, _w, vl);
-
-                        kptr += packn;
-                    }
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params, vl);
-
-                vse16_v_f16m1(outptr + j * packn, _sum, vl);
-            }
-
-            outptr += outw * packn;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packn.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packn.h
deleted file mode 100644
index 9d18c1d8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packn.h
+++ /dev/null
@@ -1,97 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_packn_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packn, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
-
-                if (bias_data_ptr)
-                {
-                    _sum = vle32_v_f32m1(bias_data_ptr + p * packn, vl);
-                }
-
-                const float* kptr = (const float*)weight_data_packn.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const float* sptr = m.row(i * stride_h) + j * stride_w * packn;
-
-                    for (int k = 0; k < maxk; k++) // 29.23
-                    {
-                        const float* slptr = sptr + space_ofs[k] * packn;
-
-                        for (int l = 0; l < packn; l++)
-                        {
-                            float val = *slptr++;
-                            vfloat32m1_t _w0 = vle32_v_f32m1(kptr, vl);
-                            _sum = vfmacc_vf_f32m1(_sum, val, _w0, vl);
-
-                            kptr += packn;
-                        }
-                    }
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params, vl);
-
-                vse32_v_f32m1(outptr + j * packn, _sum, vl);
-            }
-
-            outptr += outw * packn;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packn_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packn_fp16s.h
deleted file mode 100644
index 0d36cf20..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packn_fp16s.h
+++ /dev/null
@@ -1,186 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_packn_fp16s_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat32m2_t _sum = vfmv_v_f_f32m2(0.f, vl);
-
-                if (bias_data_ptr)
-                {
-                    _sum = vle32_v_f32m2(bias_data_ptr + p * packn, vl);
-                }
-
-                const __fp16* kptr = weight_data_fp16.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const __fp16* sptr = m.row<const __fp16>(i * stride_h) + j * stride_w * packn;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const __fp16* slptr = sptr + space_ofs[k] * packn;
-
-                        for (int l = 0; l < packn; l++)
-                        {
-                            float val = (float)*slptr++;
-                            vfloat16m1_t _w0 = vle16_v_f16m1(kptr, vl);
-                            // _sum = vfwmacc_vf_f32m2(_sum, val, _w0, vl);
-
-                            vfloat32m2_t _qwq = vfwmul_vf_f32m2(_w0, val, vl);
-                            _sum = vfadd_vv_f32m2(_sum, _qwq, vl);
-
-                            kptr += packn;
-                        }
-                    }
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params, vl);
-
-                vse16_v_f16m1(outptr + j * packn, vfncvt_f_f_w_f16m1(_sum, vl), vl);
-            }
-
-            outptr += outw * packn;
-        }
-    }
-}
-
-static void convolution_packn_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data_fp16, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const __fp16* bias_data_ptr = bias_data_fp16;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat16m1_t _sum = vfmv_v_f_f16m1(0.f, vl);
-
-                if (bias_data_ptr)
-                {
-                    _sum = vle16_v_f16m1(bias_data_ptr + p * packn, vl);
-                }
-
-                const __fp16* kptr = weight_data_fp16.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const __fp16* sptr = m.row<const __fp16>(i * stride_h) + j * stride_w * packn;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const __fp16* slptr = sptr + space_ofs[k] * packn;
-
-                        for (int l = 0; l < packn; l++)
-                        {
-                            __fp16 val = *slptr++;
-                            vfloat16m1_t _w0 = vle16_v_f16m1(kptr, vl);
-                            _sum = vfmacc_vf_f16m1(_sum, val, _w0, vl);
-
-                            kptr += packn;
-                        }
-                    }
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params, vl);
-
-                vse16_v_f16m1(outptr + j * packn, _sum, vl);
-            }
-
-            outptr += outw * packn;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packnto1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packnto1.h
deleted file mode 100644
index 4c66116d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packnto1.h
+++ /dev/null
@@ -1,97 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_packnto1_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packnto1, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
-
-                const float* kptr = (const float*)weight_data_packnto1.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const float* sptr = m.row(i * stride_h) + j * stride_w * packn;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        vfloat32m1_t _val = vle32_v_f32m1(sptr + space_ofs[k] * packn, vl);
-                        vfloat32m1_t _w = vle32_v_f32m1(kptr, vl);
-                        _sum = vfmacc_vv_f32m1(_sum, _val, _w, vl);
-
-                        kptr += packn;
-                    }
-                }
-
-                sum = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum, vfmv_s_f_f32m1(vfloat32m1_t(), sum, vl), vl));
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[j] = sum;
-            }
-
-            outptr += outw;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packnto1_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packnto1_fp16s.h
deleted file mode 100644
index 83efd308..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_packnto1_fp16s.h
+++ /dev/null
@@ -1,191 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_packnto1_fp16s_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                vfloat32m2_t _sum = vfmv_v_f_f32m2(0.f, vl);
-
-                const __fp16* kptr = weight_data_fp16.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const __fp16* sptr = m.row<const __fp16>(i * stride_h) + j * stride_w * packn;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        vfloat16m1_t _val = vle16_v_f16m1(sptr + space_ofs[k] * packn, vl);
-                        vfloat16m1_t _w = vle16_v_f16m1(kptr, vl);
-                        _sum = vfwmacc_vv_f32m2(_sum, _val, _w, vl);
-
-                        kptr += packn;
-                    }
-                }
-
-#if C906
-                // TODO
-                std::vector<float> ss(packn);
-                vse32_v_f32m2((float*)ss.data(), _sum, vl);
-                for (int i = 0; i < packn; i++)
-                {
-                    sum += ss[i];
-                }
-#else
-                sum = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m2_f32m1(vfloat32m1_t(), _sum, vfmv_s_f_f32m1(vfloat32m1_t(), sum, vl), vl));
-#endif
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[j] = (__fp16)sum;
-            }
-
-            outptr += outw;
-        }
-    }
-}
-
-static void convolution_packnto1_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data_fp16, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const __fp16* bias_data_ptr = bias_data_fp16;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __fp16 sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                vfloat16m1_t _sum = vfmv_v_f_f16m1(0.f, vl);
-
-                const __fp16* kptr = weight_data_fp16.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-                    const __fp16* sptr = m.row<const __fp16>(i * stride_h) + j * stride_w * packn;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        vfloat16m1_t _val = vle16_v_f16m1(sptr + space_ofs[k] * packn, vl);
-                        vfloat16m1_t _w = vle16_v_f16m1(kptr, vl);
-                        _sum = vfmacc_vv_f16m1(_sum, _val, _w, vl);
-
-                        kptr += packn;
-                    }
-                }
-
-                sum = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum, vfmv_s_f_f16m1(vfloat16m1_t(), sum, vl), vl));
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[j] = sum;
-            }
-
-            outptr += outw;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_riscv.h
deleted file mode 100644
index a4e008c9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_riscv.h
+++ /dev/null
@@ -1,55 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTION_RISCV_H
-#define LAYER_CONVOLUTION_RISCV_H
-
-#include "convolution.h"
-
-namespace ncnn {
-
-class Convolution_riscv : public Convolution
-{
-public:
-    Convolution_riscv();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Layer* activation;
-
-    Mat weight_data_tm;
-    Mat weight_winograd23_data;
-    Mat weight_winograd43_data;
-    Mat weight_winograd63_data;
-
-    // fp16
-    Mat bias_data_fp16;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTION_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm.h
deleted file mode 100644
index 801b7cc4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm.h
+++ /dev/null
@@ -1,675 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_rvv(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-#if __riscv_vector
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-#endif
-
-    // Mat bottom_im2col(size, maxk, inch, 4u, 1, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    // permute
-    Mat tmp;
-#if __riscv_vector
-    if (size >= packn)
-        tmp.create(packn * maxk, inch, size / packn + size % packn, 4u, 1, opt.workspace_allocator);
-#else
-    if (size >= 4)
-        tmp.create(4 * maxk, inch, size / 4 + size % 4, 4u, 1, opt.workspace_allocator);
-#endif
-    else
-        tmp.create(maxk, inch, size, 4u, 1, opt.workspace_allocator);
-    {
-#if __riscv_vector
-        int nn_size = size / packn;
-        int remain_size_start = nn_size * packn;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = ii * packn;
-
-            float* tmpptr = tmp.channel(i / packn);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    vse32_v_f32m1(tmpptr, vle32_v_f32m1(img0, vl), vl);
-                    img0 += size;
-                    tmpptr += packn;
-                }
-            }
-        }
-#else // __riscv_vector
-        int nn_size = size / 4;
-        int remain_size_start = nn_size * 4;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = ii * 4;
-
-            float* tmpptr = tmp.channel(i / 4);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    tmpptr[1] = img0[1];
-                    tmpptr[2] = img0[2];
-                    tmpptr[3] = img0[3];
-                    img0 += size;
-                    tmpptr += 4;
-                }
-            }
-        }
-#endif // __riscv_vector
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-#if __riscv_vector
-            float* tmpptr = tmp.channel(i / packn + i % packn);
-#else
-            float* tmpptr = tmp.channel(i / 4 + i % 4);
-#endif
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    img0 += size;
-                    tmpptr += 1;
-                }
-            }
-        }
-    }
-
-#if __riscv_vector
-    int nn_outch = outch >> 3;
-    int remain_outch_start = nn_outch << 3;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 8;
-
-        float* outptr0 = top_blob.channel(p);
-        float* outptr1 = top_blob.channel(p + 1);
-        float* outptr2 = top_blob.channel(p + 2);
-        float* outptr3 = top_blob.channel(p + 3);
-        float* outptr4 = top_blob.channel(p + 4);
-        float* outptr5 = top_blob.channel(p + 5);
-        float* outptr6 = top_blob.channel(p + 6);
-        float* outptr7 = top_blob.channel(p + 7);
-
-        const float zeros[8] = {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
-        const float* biasptr = bias ? bias + p : zeros;
-
-        int i = 0;
-        for (; i + (packn - 1) < size; i += packn)
-        {
-            const float* tmpptr = tmp.channel(i / packn);
-            const float* kptr = kernel.channel(p / 8);
-
-            int nn = inch * maxk; // inch always > 0
-
-            vfloat32m1_t _sum0 = vfmv_v_f_f32m1(biasptr[0], vl);
-            vfloat32m1_t _sum1 = vfmv_v_f_f32m1(biasptr[1], vl);
-            vfloat32m1_t _sum2 = vfmv_v_f_f32m1(biasptr[2], vl);
-            vfloat32m1_t _sum3 = vfmv_v_f_f32m1(biasptr[3], vl);
-            vfloat32m1_t _sum4 = vfmv_v_f_f32m1(biasptr[4], vl);
-            vfloat32m1_t _sum5 = vfmv_v_f_f32m1(biasptr[5], vl);
-            vfloat32m1_t _sum6 = vfmv_v_f_f32m1(biasptr[6], vl);
-            vfloat32m1_t _sum7 = vfmv_v_f_f32m1(biasptr[7], vl);
-
-            for (int q = 0; q < nn; q++)
-            {
-                vfloat32m1_t _val = vle32_v_f32m1(tmpptr, vl);
-                _sum0 = vfmacc_vf_f32m1(_sum0, kptr[0], _val, vl);
-                _sum1 = vfmacc_vf_f32m1(_sum1, kptr[1], _val, vl);
-                _sum2 = vfmacc_vf_f32m1(_sum2, kptr[2], _val, vl);
-                _sum3 = vfmacc_vf_f32m1(_sum3, kptr[3], _val, vl);
-                _sum4 = vfmacc_vf_f32m1(_sum4, kptr[4], _val, vl);
-                _sum5 = vfmacc_vf_f32m1(_sum5, kptr[5], _val, vl);
-                _sum6 = vfmacc_vf_f32m1(_sum6, kptr[6], _val, vl);
-                _sum7 = vfmacc_vf_f32m1(_sum7, kptr[7], _val, vl);
-                tmpptr += packn;
-                kptr += 8;
-            }
-
-            vse32_v_f32m1(outptr0, _sum0, vl);
-            vse32_v_f32m1(outptr1, _sum1, vl);
-            vse32_v_f32m1(outptr2, _sum2, vl);
-            vse32_v_f32m1(outptr3, _sum3, vl);
-            vse32_v_f32m1(outptr4, _sum4, vl);
-            vse32_v_f32m1(outptr5, _sum5, vl);
-            vse32_v_f32m1(outptr6, _sum6, vl);
-            vse32_v_f32m1(outptr7, _sum7, vl);
-
-            outptr0 += packn;
-            outptr1 += packn;
-            outptr2 += packn;
-            outptr3 += packn;
-            outptr4 += packn;
-            outptr5 += packn;
-            outptr6 += packn;
-            outptr7 += packn;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / packn + i % packn);
-            const float* kptr = kernel.channel(p / 8);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = biasptr[0];
-            float sum1 = biasptr[1];
-            float sum2 = biasptr[2];
-            float sum3 = biasptr[3];
-            float sum4 = biasptr[4];
-            float sum5 = biasptr[5];
-            float sum6 = biasptr[6];
-            float sum7 = biasptr[7];
-
-            for (int q = 0; q < nn; q++)
-            {
-                sum0 += tmpptr[0] * kptr[0];
-                sum1 += tmpptr[0] * kptr[1];
-                sum2 += tmpptr[0] * kptr[2];
-                sum3 += tmpptr[0] * kptr[3];
-                sum4 += tmpptr[0] * kptr[4];
-                sum5 += tmpptr[0] * kptr[5];
-                sum6 += tmpptr[0] * kptr[6];
-                sum7 += tmpptr[0] * kptr[7];
-                tmpptr++;
-                kptr += 8;
-            }
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-            outptr2[0] = sum2;
-            outptr3[0] = sum3;
-            outptr4[0] = sum4;
-            outptr5[0] = sum5;
-            outptr6[0] = sum6;
-            outptr7[0] = sum7;
-
-            outptr0++;
-            outptr1++;
-            outptr2++;
-            outptr3++;
-            outptr4++;
-            outptr5++;
-            outptr6++;
-            outptr7++;
-        }
-    }
-
-    nn_outch = (outch - remain_outch_start) >> 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = remain_outch_start + pp * 4;
-
-        float* outptr0 = top_blob.channel(p);
-        float* outptr1 = top_blob.channel(p + 1);
-        float* outptr2 = top_blob.channel(p + 2);
-        float* outptr3 = top_blob.channel(p + 3);
-
-        const float zeros[4] = {0.f, 0.f, 0.f, 0.f};
-        const float* biasptr = bias ? bias + p : zeros;
-
-        int i = 0;
-        for (; i + (packn - 1) < size; i += packn)
-        {
-            const float* tmpptr = tmp.channel(i / packn);
-            const float* kptr = kernel.channel(p / 8 + (p % 8) / 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            vfloat32m1_t _sum0 = vfmv_v_f_f32m1(biasptr[0], vl);
-            vfloat32m1_t _sum1 = vfmv_v_f_f32m1(biasptr[1], vl);
-            vfloat32m1_t _sum2 = vfmv_v_f_f32m1(biasptr[2], vl);
-            vfloat32m1_t _sum3 = vfmv_v_f_f32m1(biasptr[3], vl);
-
-            for (int q = 0; q < nn; q++)
-            {
-                vfloat32m1_t _val = vle32_v_f32m1(tmpptr, vl);
-                _sum0 = vfmacc_vf_f32m1(_sum0, kptr[0], _val, vl);
-                _sum1 = vfmacc_vf_f32m1(_sum1, kptr[1], _val, vl);
-                _sum2 = vfmacc_vf_f32m1(_sum2, kptr[2], _val, vl);
-                _sum3 = vfmacc_vf_f32m1(_sum3, kptr[3], _val, vl);
-                tmpptr += packn;
-                kptr += 4;
-            }
-
-            vse32_v_f32m1(outptr0, _sum0, vl);
-            vse32_v_f32m1(outptr1, _sum1, vl);
-            vse32_v_f32m1(outptr2, _sum2, vl);
-            vse32_v_f32m1(outptr3, _sum3, vl);
-
-            outptr0 += packn;
-            outptr1 += packn;
-            outptr2 += packn;
-            outptr3 += packn;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / packn + i % packn);
-            const float* kptr = kernel.channel(p / 8 + (p % 8) / 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = biasptr[0];
-            float sum1 = biasptr[1];
-            float sum2 = biasptr[2];
-            float sum3 = biasptr[3];
-
-            for (int q = 0; q < nn; q++)
-            {
-                sum0 += tmpptr[0] * kptr[0];
-                sum1 += tmpptr[0] * kptr[1];
-                sum2 += tmpptr[0] * kptr[2];
-                sum3 += tmpptr[0] * kptr[3];
-                tmpptr++;
-                kptr += 4;
-            }
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-            outptr2[0] = sum2;
-            outptr3[0] = sum3;
-
-            outptr0++;
-            outptr1++;
-            outptr2++;
-            outptr3++;
-        }
-    }
-
-    remain_outch_start += nn_outch << 2;
-#else // __riscv_vector
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        float* outptr0 = top_blob.channel(p);
-        float* outptr1 = top_blob.channel(p + 1);
-
-        const float zeros[2] = {0.f, 0.f};
-        const float* biasptr = bias ? bias + p : zeros;
-
-        int i = 0;
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 4);
-            const float* kptr = kernel.channel(p / 2);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum00 = biasptr[0];
-            float sum01 = biasptr[0];
-            float sum02 = biasptr[0];
-            float sum03 = biasptr[0];
-            float sum10 = biasptr[1];
-            float sum11 = biasptr[1];
-            float sum12 = biasptr[1];
-            float sum13 = biasptr[1];
-
-            for (int q = 0; q < nn; q++)
-            {
-                float k0 = kptr[0];
-                float k1 = kptr[1];
-                sum00 += tmpptr[0] * k0;
-                sum01 += tmpptr[1] * k0;
-                sum02 += tmpptr[2] * k0;
-                sum03 += tmpptr[3] * k0;
-                sum10 += tmpptr[0] * k1;
-                sum11 += tmpptr[1] * k1;
-                sum12 += tmpptr[2] * k1;
-                sum13 += tmpptr[3] * k1;
-                tmpptr += 4;
-                kptr += 2;
-            }
-
-            outptr0[0] = sum00;
-            outptr0[1] = sum01;
-            outptr0[2] = sum02;
-            outptr0[3] = sum03;
-            outptr1[0] = sum10;
-            outptr1[1] = sum11;
-            outptr1[2] = sum12;
-            outptr1[3] = sum13;
-
-            outptr0 += 4;
-            outptr1 += 4;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 4 + i % 4);
-            const float* kptr = kernel.channel(p / 2);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = biasptr[0];
-            float sum1 = biasptr[1];
-
-            for (int q = 0; q < nn; q++)
-            {
-                sum0 += tmpptr[0] * kptr[0];
-                sum1 += tmpptr[0] * kptr[1];
-                tmpptr++;
-                kptr += 2;
-            }
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-
-            outptr0++;
-            outptr1++;
-        }
-    }
-#endif // __riscv_vector
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        float* outptr0 = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        int i = 0;
-#if __riscv_vector
-        for (; i + (packn - 1) < size; i += packn)
-        {
-            const float* tmpptr = tmp.channel(i / packn);
-            const float* kptr = kernel.channel(p / 8 + (p % 8) / 4 + p % 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            vfloat32m1_t _sum0 = vfmv_v_f_f32m1(bias0, vl);
-
-            for (int q = 0; q < nn; q++)
-            {
-                _sum0 = vfmacc_vf_f32m1(_sum0, kptr[0], vle32_v_f32m1(tmpptr, vl), vl);
-                tmpptr += packn;
-                kptr++;
-            }
-
-            vse32_v_f32m1(outptr0, _sum0, vl);
-
-            outptr0 += packn;
-        }
-#else  // __riscv_vector
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 4);
-            const float* kptr = kernel.channel(p / 2 + p % 2);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = bias0;
-            float sum1 = bias0;
-            float sum2 = bias0;
-            float sum3 = bias0;
-
-            for (int q = 0; q < nn; q++)
-            {
-                float k0 = kptr[0];
-                sum0 += tmpptr[0] * k0;
-                sum1 += tmpptr[1] * k0;
-                sum2 += tmpptr[2] * k0;
-                sum3 += tmpptr[3] * k0;
-                tmpptr += 4;
-                kptr++;
-            }
-
-            outptr0[0] = sum0;
-            outptr0[1] = sum1;
-            outptr0[2] = sum2;
-            outptr0[3] = sum3;
-
-            outptr0 += 4;
-        }
-#endif // __riscv_vector
-        for (; i < size; i++)
-        {
-#if __riscv_vector
-            const float* tmpptr = tmp.channel(i / packn + i % packn);
-            const float* kptr = kernel.channel(p / 8 + (p % 8) / 4 + p % 4);
-#else
-            const float* tmpptr = tmp.channel(i / 4 + i % 4);
-            const float* kptr = kernel.channel(p / 2 + p % 2);
-#endif
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = bias0;
-
-            for (int q = 0; q < nn; q++)
-            {
-                sum0 += tmpptr[0] * kptr[0];
-                tmpptr++;
-                kptr++;
-            }
-
-            outptr0[0] = sum0;
-
-            outptr0++;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_transform_kernel_rvv(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = 8b-maxk-inch-outch/8b
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-#if __riscv_vector
-    kernel_tm.create(8 * maxk, inch, outch / 8 + (outch % 8) / 4 + outch % 4);
-#else
-    kernel_tm.create(2 * maxk, inch, outch / 2 + outch % 2);
-#endif
-
-    int q = 0;
-#if __riscv_vector
-    for (; q + 7 < outch; q += 8)
-    {
-        const Mat k0 = kernel.channel(q);
-        const Mat k1 = kernel.channel(q + 1);
-        const Mat k2 = kernel.channel(q + 2);
-        const Mat k3 = kernel.channel(q + 3);
-        const Mat k4 = kernel.channel(q + 4);
-        const Mat k5 = kernel.channel(q + 5);
-        const Mat k6 = kernel.channel(q + 6);
-        const Mat k7 = kernel.channel(q + 7);
-
-        float* g00 = kernel_tm.channel(q / 8);
-
-        for (int p = 0; p < inch; p++)
-        {
-            const float* k00 = k0.row(p);
-            const float* k10 = k1.row(p);
-            const float* k20 = k2.row(p);
-            const float* k30 = k3.row(p);
-            const float* k40 = k4.row(p);
-            const float* k50 = k5.row(p);
-            const float* k60 = k6.row(p);
-            const float* k70 = k7.row(p);
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k00[k];
-                g00[1] = k10[k];
-                g00[2] = k20[k];
-                g00[3] = k30[k];
-                g00[4] = k40[k];
-                g00[5] = k50[k];
-                g00[6] = k60[k];
-                g00[7] = k70[k];
-
-                g00 += 8;
-            }
-        }
-    }
-    for (; q + 3 < outch; q += 4)
-    {
-        const Mat k0 = kernel.channel(q);
-        const Mat k1 = kernel.channel(q + 1);
-        const Mat k2 = kernel.channel(q + 2);
-        const Mat k3 = kernel.channel(q + 3);
-
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4);
-
-        for (int p = 0; p < inch; p++)
-        {
-            const float* k00 = k0.row(p);
-            const float* k10 = k1.row(p);
-            const float* k20 = k2.row(p);
-            const float* k30 = k3.row(p);
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k00[k];
-                g00[1] = k10[k];
-                g00[2] = k20[k];
-                g00[3] = k30[k];
-
-                g00 += 4;
-            }
-        }
-    }
-#else
-    for (; q + 1 < outch; q += 2)
-    {
-        const Mat k0 = kernel.channel(q);
-        const Mat k1 = kernel.channel(q + 1);
-
-        float* g00 = kernel_tm.channel(q / 2);
-
-        for (int p = 0; p < inch; p++)
-        {
-            const float* k00 = k0.row(p);
-            const float* k10 = k1.row(p);
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k00[k];
-                g00[1] = k10[k];
-
-                g00 += 2;
-            }
-        }
-    }
-#endif // __riscv_vector
-    for (; q < outch; q++)
-    {
-        const Mat k0 = kernel.channel(q);
-
-#if __riscv_vector
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + q % 4);
-#else
-        float* g00 = kernel_tm.channel(q / 2 + q % 2);
-#endif
-
-        for (int p = 0; p < inch; p++)
-        {
-            const float* k00 = k0.row(p);
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = k00[k];
-
-                g00 += 1;
-            }
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 4u, 1, opt.workspace_allocator);
-    {
-        const int gap = w * stride_h - outw * stride_w;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            float* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const float* sptr = img.row<const float>(dilation_h * u) + dilation_w * v;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            ptr[0] = sptr[0];
-
-                            sptr += stride_w;
-                            ptr += 1;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_fp16s.h
deleted file mode 100644
index 72a62164..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_fp16s.h
+++ /dev/null
@@ -1,540 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_fp16sa_rvv(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-#if __riscv_vector
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-#endif
-
-    // Mat bottom_im2col(size, maxk, inch, 2u, 1, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    const __fp16* bias = _bias;
-
-    // permute
-    Mat tmp;
-#if __riscv_vector
-    if (size >= packn)
-        tmp.create(packn * maxk, inch, size / packn + size % packn, 2u, 1, opt.workspace_allocator);
-    else
-        tmp.create(maxk, inch, size, 2u, 1, opt.workspace_allocator);
-    {
-        int nn_size = size / packn;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = ii * packn;
-
-            __fp16* tmpptr = tmp.channel(i / packn);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const __fp16* img0 = (const __fp16*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    vse16_v_f16m1(tmpptr, vle16_v_f16m1(img0, vl), vl);
-                    img0 += size;
-                    tmpptr += packn;
-                }
-            }
-        }
-
-        int remain_size_start = nn_size * packn;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            __fp16* tmpptr = tmp.channel(i / packn + i % packn);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const __fp16* img0 = (const __fp16*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    img0 += size;
-                    tmpptr += 1;
-                }
-            }
-        }
-    }
-#else // __riscv_vector
-    tmp.create(maxk, inch, size, 2u, 1, opt.workspace_allocator);
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = 0; i < size; i++)
-        {
-            __fp16* tmpptr = tmp.channel(i);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const __fp16* img0 = (const __fp16*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    img0 += size;
-                    tmpptr += 1;
-                }
-            }
-        }
-    }
-#endif // __riscv_vector
-
-#if __riscv_vector
-    int nn_outch = outch >> 3;
-    int remain_outch_start = nn_outch << 3;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 8;
-
-        __fp16* outptr0 = top_blob.channel(p);
-        __fp16* outptr1 = top_blob.channel(p + 1);
-        __fp16* outptr2 = top_blob.channel(p + 2);
-        __fp16* outptr3 = top_blob.channel(p + 3);
-        __fp16* outptr4 = top_blob.channel(p + 4);
-        __fp16* outptr5 = top_blob.channel(p + 5);
-        __fp16* outptr6 = top_blob.channel(p + 6);
-        __fp16* outptr7 = top_blob.channel(p + 7);
-
-        const __fp16 zeros[8] = {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
-        const __fp16* biasptr = bias ? bias + p : zeros;
-
-        int i = 0;
-        for (; i + (packn - 1) < size; i += packn)
-        {
-            const __fp16* tmpptr = tmp.channel(i / packn);
-            const __fp16* kptr = kernel.channel(p / 8);
-
-            int nn = inch * maxk; // inch always > 0
-
-            vfloat16m1_t _sum0 = vfmv_v_f_f16m1(biasptr[0], vl);
-            vfloat16m1_t _sum1 = vfmv_v_f_f16m1(biasptr[1], vl);
-            vfloat16m1_t _sum2 = vfmv_v_f_f16m1(biasptr[2], vl);
-            vfloat16m1_t _sum3 = vfmv_v_f_f16m1(biasptr[3], vl);
-            vfloat16m1_t _sum4 = vfmv_v_f_f16m1(biasptr[4], vl);
-            vfloat16m1_t _sum5 = vfmv_v_f_f16m1(biasptr[5], vl);
-            vfloat16m1_t _sum6 = vfmv_v_f_f16m1(biasptr[6], vl);
-            vfloat16m1_t _sum7 = vfmv_v_f_f16m1(biasptr[7], vl);
-
-            for (int q = 0; q < nn; q++)
-            {
-                vfloat16m1_t _val = vle16_v_f16m1(tmpptr, vl);
-                _sum0 = vfmacc_vf_f16m1(_sum0, kptr[0], _val, vl);
-                _sum1 = vfmacc_vf_f16m1(_sum1, kptr[1], _val, vl);
-                _sum2 = vfmacc_vf_f16m1(_sum2, kptr[2], _val, vl);
-                _sum3 = vfmacc_vf_f16m1(_sum3, kptr[3], _val, vl);
-                _sum4 = vfmacc_vf_f16m1(_sum4, kptr[4], _val, vl);
-                _sum5 = vfmacc_vf_f16m1(_sum5, kptr[5], _val, vl);
-                _sum6 = vfmacc_vf_f16m1(_sum6, kptr[6], _val, vl);
-                _sum7 = vfmacc_vf_f16m1(_sum7, kptr[7], _val, vl);
-                tmpptr += packn;
-                kptr += 8;
-            }
-
-            vse16_v_f16m1(outptr0, _sum0, vl);
-            vse16_v_f16m1(outptr1, _sum1, vl);
-            vse16_v_f16m1(outptr2, _sum2, vl);
-            vse16_v_f16m1(outptr3, _sum3, vl);
-            vse16_v_f16m1(outptr4, _sum4, vl);
-            vse16_v_f16m1(outptr5, _sum5, vl);
-            vse16_v_f16m1(outptr6, _sum6, vl);
-            vse16_v_f16m1(outptr7, _sum7, vl);
-
-            outptr0 += packn;
-            outptr1 += packn;
-            outptr2 += packn;
-            outptr3 += packn;
-            outptr4 += packn;
-            outptr5 += packn;
-            outptr6 += packn;
-            outptr7 += packn;
-        }
-        for (; i < size; i++)
-        {
-            const __fp16* tmpptr = tmp.channel(i / packn + i % packn);
-            const __fp16* kptr = kernel.channel(p / 8);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __fp16 sum0 = biasptr[0];
-            __fp16 sum1 = biasptr[1];
-            __fp16 sum2 = biasptr[2];
-            __fp16 sum3 = biasptr[3];
-            __fp16 sum4 = biasptr[4];
-            __fp16 sum5 = biasptr[5];
-            __fp16 sum6 = biasptr[6];
-            __fp16 sum7 = biasptr[7];
-
-            for (int q = 0; q < nn; q++)
-            {
-                sum0 += tmpptr[0] * kptr[0];
-                sum1 += tmpptr[0] * kptr[1];
-                sum2 += tmpptr[0] * kptr[2];
-                sum3 += tmpptr[0] * kptr[3];
-                sum4 += tmpptr[0] * kptr[4];
-                sum5 += tmpptr[0] * kptr[5];
-                sum6 += tmpptr[0] * kptr[6];
-                sum7 += tmpptr[0] * kptr[7];
-                tmpptr++;
-                kptr += 8;
-            }
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-            outptr2[0] = sum2;
-            outptr3[0] = sum3;
-            outptr4[0] = sum4;
-            outptr5[0] = sum5;
-            outptr6[0] = sum6;
-            outptr7[0] = sum7;
-
-            outptr0++;
-            outptr1++;
-            outptr2++;
-            outptr3++;
-            outptr4++;
-            outptr5++;
-            outptr6++;
-            outptr7++;
-        }
-    }
-
-    nn_outch = (outch - remain_outch_start) >> 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = remain_outch_start + pp * 4;
-
-        __fp16* outptr0 = top_blob.channel(p);
-        __fp16* outptr1 = top_blob.channel(p + 1);
-        __fp16* outptr2 = top_blob.channel(p + 2);
-        __fp16* outptr3 = top_blob.channel(p + 3);
-
-        const __fp16 zeros[4] = {0.f, 0.f, 0.f, 0.f};
-        const __fp16* biasptr = bias ? bias + p : zeros;
-
-        int i = 0;
-        for (; i + (packn - 1) < size; i += packn)
-        {
-            const __fp16* tmpptr = tmp.channel(i / packn);
-            const __fp16* kptr = kernel.channel(p / 8 + (p % 8) / 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            vfloat16m1_t _sum0 = vfmv_v_f_f16m1(biasptr[0], vl);
-            vfloat16m1_t _sum1 = vfmv_v_f_f16m1(biasptr[1], vl);
-            vfloat16m1_t _sum2 = vfmv_v_f_f16m1(biasptr[2], vl);
-            vfloat16m1_t _sum3 = vfmv_v_f_f16m1(biasptr[3], vl);
-
-            for (int q = 0; q < nn; q++)
-            {
-                vfloat16m1_t _val = vle16_v_f16m1(tmpptr, vl);
-                _sum0 = vfmacc_vf_f16m1(_sum0, kptr[0], _val, vl);
-                _sum1 = vfmacc_vf_f16m1(_sum1, kptr[1], _val, vl);
-                _sum2 = vfmacc_vf_f16m1(_sum2, kptr[2], _val, vl);
-                _sum3 = vfmacc_vf_f16m1(_sum3, kptr[3], _val, vl);
-                tmpptr += packn;
-                kptr += 4;
-            }
-
-            vse16_v_f16m1(outptr0, _sum0, vl);
-            vse16_v_f16m1(outptr1, _sum1, vl);
-            vse16_v_f16m1(outptr2, _sum2, vl);
-            vse16_v_f16m1(outptr3, _sum3, vl);
-
-            outptr0 += packn;
-            outptr1 += packn;
-            outptr2 += packn;
-            outptr3 += packn;
-        }
-        for (; i < size; i++)
-        {
-            const __fp16* tmpptr = tmp.channel(i / packn + i % packn);
-            const __fp16* kptr = kernel.channel(p / 8 + (p % 8) / 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __fp16 sum0 = biasptr[0];
-            __fp16 sum1 = biasptr[1];
-            __fp16 sum2 = biasptr[2];
-            __fp16 sum3 = biasptr[3];
-
-            for (int q = 0; q < nn; q++)
-            {
-                sum0 += tmpptr[0] * kptr[0];
-                sum1 += tmpptr[0] * kptr[1];
-                sum2 += tmpptr[0] * kptr[2];
-                sum3 += tmpptr[0] * kptr[3];
-                tmpptr++;
-                kptr += 4;
-            }
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-            outptr2[0] = sum2;
-            outptr3[0] = sum3;
-
-            outptr0++;
-            outptr1++;
-            outptr2++;
-            outptr3++;
-        }
-    }
-
-    remain_outch_start += nn_outch << 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        __fp16* outptr0 = top_blob.channel(p);
-
-        const __fp16 bias0 = bias ? bias[p] : 0.f;
-
-        int i = 0;
-        for (; i + (packn - 1) < size; i += packn)
-        {
-            const __fp16* tmpptr = tmp.channel(i / packn);
-            const __fp16* kptr = kernel.channel(p / 8 + (p % 8) / 4 + p % 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            vfloat16m1_t _sum0 = vfmv_v_f_f16m1(bias0, vl);
-
-            for (int q = 0; q < nn; q++)
-            {
-                _sum0 = vfmacc_vf_f16m1(_sum0, kptr[0], vle16_v_f16m1(tmpptr, vl), vl);
-                tmpptr += packn;
-                kptr++;
-            }
-
-            vse16_v_f16m1(outptr0, _sum0, vl);
-
-            outptr0 += packn;
-        }
-        for (; i < size; i++)
-        {
-            const __fp16* tmpptr = tmp.channel(i / packn + i % packn);
-            const __fp16* kptr = kernel.channel(p / 8 + (p % 8) / 4 + p % 4);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __fp16 sum0 = bias0;
-
-            for (int q = 0; q < nn; q++)
-            {
-                sum0 += tmpptr[0] * kptr[0];
-                tmpptr++;
-                kptr++;
-            }
-
-            outptr0[0] = sum0;
-
-            outptr0++;
-        }
-    }
-#else // __riscv_vector
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr0 = top_blob.channel(p);
-
-        const __fp16 bias0 = bias ? bias[p] : 0.f;
-
-        for (int i = 0; i < size; i++)
-        {
-            const __fp16* tmpptr = tmp.channel(i);
-            const __fp16* kptr = kernel.channel(p);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __fp16 sum0 = bias0;
-
-            for (int q = 0; q < nn; q++)
-            {
-                sum0 += tmpptr[0] * kptr[0];
-                tmpptr++;
-                kptr++;
-            }
-
-            outptr0[0] = sum0;
-
-            outptr0++;
-        }
-    }
-#endif // __riscv_vector
-}
-
-static void convolution_im2col_sgemm_transform_kernel_fp16sa_rvv(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = 8b-maxk-inch-outch/8b
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-#if __riscv_vector
-    kernel_tm.create(8 * maxk, inch, outch / 8 + (outch % 8) / 4 + outch % 4, (size_t)2u);
-
-    int q = 0;
-    for (; q + 7 < outch; q += 8)
-    {
-        const Mat k0 = kernel.channel(q);
-        const Mat k1 = kernel.channel(q + 1);
-        const Mat k2 = kernel.channel(q + 2);
-        const Mat k3 = kernel.channel(q + 3);
-        const Mat k4 = kernel.channel(q + 4);
-        const Mat k5 = kernel.channel(q + 5);
-        const Mat k6 = kernel.channel(q + 6);
-        const Mat k7 = kernel.channel(q + 7);
-
-        __fp16* g00 = kernel_tm.channel(q / 8);
-
-        for (int p = 0; p < inch; p++)
-        {
-            const float* k00 = k0.row(p);
-            const float* k10 = k1.row(p);
-            const float* k20 = k2.row(p);
-            const float* k30 = k3.row(p);
-            const float* k40 = k4.row(p);
-            const float* k50 = k5.row(p);
-            const float* k60 = k6.row(p);
-            const float* k70 = k7.row(p);
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = (__fp16)k00[k];
-                g00[1] = (__fp16)k10[k];
-                g00[2] = (__fp16)k20[k];
-                g00[3] = (__fp16)k30[k];
-                g00[4] = (__fp16)k40[k];
-                g00[5] = (__fp16)k50[k];
-                g00[6] = (__fp16)k60[k];
-                g00[7] = (__fp16)k70[k];
-
-                g00 += 8;
-            }
-        }
-    }
-    for (; q + 3 < outch; q += 4)
-    {
-        const Mat k0 = kernel.channel(q);
-        const Mat k1 = kernel.channel(q + 1);
-        const Mat k2 = kernel.channel(q + 2);
-        const Mat k3 = kernel.channel(q + 3);
-
-        __fp16* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4);
-
-        for (int p = 0; p < inch; p++)
-        {
-            const float* k00 = k0.row(p);
-            const float* k10 = k1.row(p);
-            const float* k20 = k2.row(p);
-            const float* k30 = k3.row(p);
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = (__fp16)k00[k];
-                g00[1] = (__fp16)k10[k];
-                g00[2] = (__fp16)k20[k];
-                g00[3] = (__fp16)k30[k];
-
-                g00 += 4;
-            }
-        }
-    }
-    for (; q < outch; q++)
-    {
-        const Mat k0 = kernel.channel(q);
-
-        __fp16* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + q % 4);
-
-        for (int p = 0; p < inch; p++)
-        {
-            const float* k00 = k0.row(p);
-
-            for (int k = 0; k < maxk; k++)
-            {
-                g00[0] = (__fp16)k00[k];
-
-                g00 += 1;
-            }
-        }
-    }
-#else
-    kernel_tm = kernel;
-#endif // __riscv_vector
-}
-
-static void convolution_im2col_sgemm_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 2u, 1, opt.workspace_allocator);
-    {
-        const int gap = w * stride_h - outw * stride_w;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            __fp16* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const __fp16* sptr = img.row<const __fp16>(dilation_h * u) + dilation_w * v;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            ptr[0] = sptr[0];
-
-                            sptr += stride_w;
-                            ptr += 1;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_fp16sa_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_pack1ton.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_pack1ton.h
deleted file mode 100644
index 8a3e6ffb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_pack1ton.h
+++ /dev/null
@@ -1,136 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_pack1ton_rvv(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    // Mat bottom_im2col(size, maxk, inch, 4u, 1, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    // permute
-    Mat tmp;
-    tmp.create(maxk, inch, size, 4u, 1, opt.workspace_allocator);
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = 0; i < size; i++)
-        {
-            float* tmpptr = tmp.channel(i);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    img0 += size;
-                    tmpptr += 1;
-                }
-            }
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr0 = top_blob.channel(p);
-
-        int i = 0;
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i);
-            const float* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk; // inch always > 0
-
-            vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
-
-            if (bias)
-            {
-                _sum = vle32_v_f32m1(bias + p * packn, vl);
-            }
-
-            for (int j = 0; j < nn; j++)
-            {
-                float val = *tmpptr++;
-                vfloat32m1_t _w0 = vle32_v_f32m1(kptr0, vl);
-                _sum = vfmacc_vf_f32m1(_sum, val, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-            vse32_v_f32m1(outptr0, _sum, vl);
-
-            outptr0 += packn;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_pack1ton_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 4u, 1, opt.workspace_allocator);
-    {
-        const int gap = w * stride_h - outw * stride_w;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            float* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const float* sptr = img.row(dilation_h * u) + dilation_w * v;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            ptr[0] = sptr[0];
-
-                            sptr += stride_w;
-                            ptr += 1;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_pack1ton_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_pack1ton_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_pack1ton_fp16s.h
deleted file mode 100644
index 0c0b2791..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_pack1ton_fp16s.h
+++ /dev/null
@@ -1,136 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_pack1ton_fp16sa_rvv(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    // Mat bottom_im2col(size, maxk, inch, 2u, 1, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    const __fp16* bias = _bias;
-
-    // permute
-    Mat tmp;
-    tmp.create(maxk, inch, size, 2u, 1, opt.workspace_allocator);
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = 0; i < size; i++)
-        {
-            __fp16* tmpptr = tmp.channel(i);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const __fp16* img0 = (const __fp16*)bottom_im2col.channel(q) + i;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    tmpptr[0] = img0[0];
-                    img0 += size;
-                    tmpptr += 1;
-                }
-            }
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr0 = top_blob.channel(p);
-
-        int i = 0;
-        for (; i < size; i++)
-        {
-            const __fp16* tmpptr = tmp.channel(i);
-            const __fp16* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk; // inch always > 0
-
-            vfloat16m1_t _sum = vfmv_v_f_f16m1(0.f, vl);
-
-            if (bias)
-            {
-                _sum = vle16_v_f16m1(bias + p * packn, vl);
-            }
-
-            for (int j = 0; j < nn; j++)
-            {
-                __fp16 val = *tmpptr++;
-                vfloat16m1_t _w0 = vle16_v_f16m1(kptr0, vl);
-                _sum = vfmacc_vf_f16m1(_sum, val, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-            vse16_v_f16m1(outptr0, _sum, vl);
-
-            outptr0 += packn;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_pack1ton_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 2u, 1, opt.workspace_allocator);
-    {
-        const int gap = w * stride_h - outw * stride_w;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            __fp16* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const __fp16* sptr = img.row<const __fp16>(dilation_h * u) + dilation_w * v;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            ptr[0] = sptr[0];
-
-                            sptr += stride_w;
-                            ptr += 1;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_pack1ton_fp16sa_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packn.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packn.h
deleted file mode 100644
index 9255c092..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packn.h
+++ /dev/null
@@ -1,414 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_packn_rvv(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    // Mat bottom_im2col(size, maxk, inch, 4u * packn, packn, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    // permute
-    Mat tmp;
-    if (size >= 8)
-        tmp.create(8 * maxk, inch, size / 8 + (size % 8) / 4 + (size % 4) / 2 + size % 2, 4u * packn, packn, opt.workspace_allocator);
-    else if (size >= 4)
-        tmp.create(4 * maxk, inch, size / 4 + (size % 4) / 2 + size % 2, 4u * packn, packn, opt.workspace_allocator);
-    else if (size >= 2)
-        tmp.create(2 * maxk, inch, size / 2 + size % 2, 4u * packn, packn, opt.workspace_allocator);
-    else
-        tmp.create(maxk, inch, size, 4u * packn, packn, opt.workspace_allocator);
-    {
-        int remain_size_start = 0;
-        int nn_size = size >> 3;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 8;
-
-            float* tmpptr = tmp.channel(i / 8);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-#if C906
-                    for (int l = 0; l < packn; l++)
-                    {
-                        tmpptr[0] = img0[l];
-                        tmpptr[1] = img0[l + packn];
-                        tmpptr[2] = img0[l + packn * 2];
-                        tmpptr[3] = img0[l + packn * 3];
-                        tmpptr[4] = img0[l + packn * 4];
-                        tmpptr[5] = img0[l + packn * 5];
-                        tmpptr[6] = img0[l + packn * 6];
-                        tmpptr[7] = img0[l + packn * 7];
-                        tmpptr += 8;
-                    }
-
-                    img0 += size * packn;
-#else
-                    vfloat32m1_t _val0 = vle32_v_f32m1(img0, vl);
-                    vfloat32m1_t _val1 = vle32_v_f32m1(img0 + packn, vl);
-                    vfloat32m1_t _val2 = vle32_v_f32m1(img0 + packn * 2, vl);
-                    vfloat32m1_t _val3 = vle32_v_f32m1(img0 + packn * 3, vl);
-                    vfloat32m1_t _val4 = vle32_v_f32m1(img0 + packn * 4, vl);
-                    vfloat32m1_t _val5 = vle32_v_f32m1(img0 + packn * 5, vl);
-                    vfloat32m1_t _val6 = vle32_v_f32m1(img0 + packn * 6, vl);
-                    vfloat32m1_t _val7 = vle32_v_f32m1(img0 + packn * 7, vl);
-                    vsseg8e32_v_f32m1(tmpptr, _val0, _val1, _val2, _val3, _val4, _val5, _val6, _val7, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn * 8;
-#endif
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 3;
-        nn_size = (size - remain_size_start) >> 2;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 4;
-
-            float* tmpptr = tmp.channel(i / 8 + (i % 8) / 4);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-#if C906
-                    for (int l = 0; l < packn; l++)
-                    {
-                        tmpptr[0] = img0[l];
-                        tmpptr[1] = img0[l + packn];
-                        tmpptr[2] = img0[l + packn * 2];
-                        tmpptr[3] = img0[l + packn * 3];
-                        tmpptr += 4;
-                    }
-
-                    img0 += size * packn;
-#else
-                    vfloat32m1_t _val0 = vle32_v_f32m1(img0, vl);
-                    vfloat32m1_t _val1 = vle32_v_f32m1(img0 + packn, vl);
-                    vfloat32m1_t _val2 = vle32_v_f32m1(img0 + packn * 2, vl);
-                    vfloat32m1_t _val3 = vle32_v_f32m1(img0 + packn * 3, vl);
-                    vsseg4e32_v_f32m1(tmpptr, _val0, _val1, _val2, _val3, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn * 4;
-#endif
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 2;
-        nn_size = (size - remain_size_start) >> 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 2;
-
-            float* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-#if C906
-                    for (int l = 0; l < packn; l++)
-                    {
-                        tmpptr[0] = img0[l];
-                        tmpptr[1] = img0[l + packn];
-                        tmpptr += 2;
-                    }
-
-                    img0 += size * packn;
-#else
-                    vfloat32m1_t _val0 = vle32_v_f32m1(img0, vl);
-                    vfloat32m1_t _val1 = vle32_v_f32m1(img0 + packn, vl);
-                    vsseg2e32_v_f32m1(tmpptr, _val0, _val1, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn * 2;
-#endif
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            float* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2 + i % 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    vfloat32m1_t _val = vle32_v_f32m1(img0, vl);
-                    vse32_v_f32m1(tmpptr, _val, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn;
-                }
-            }
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr0 = top_blob.channel(p);
-
-        int i = 0;
-        for (; i + 7 < size; i += 8)
-        {
-            const float* tmpptr = tmp.channel(i / 8);
-            const float* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat32m1_t _sum0 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum1 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum2 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum3 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum4 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum5 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum6 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum7 = vfmv_v_f_f32m1(0.f, vl);
-
-            if (bias)
-            {
-                _sum0 = vle32_v_f32m1(bias + p * packn, vl);
-                _sum1 = vle32_v_f32m1(bias + p * packn, vl);
-                _sum2 = vle32_v_f32m1(bias + p * packn, vl);
-                _sum3 = vle32_v_f32m1(bias + p * packn, vl);
-                _sum4 = vle32_v_f32m1(bias + p * packn, vl);
-                _sum5 = vle32_v_f32m1(bias + p * packn, vl);
-                _sum6 = vle32_v_f32m1(bias + p * packn, vl);
-                _sum7 = vle32_v_f32m1(bias + p * packn, vl);
-            }
-
-            for (int j = 0; j < nn; j++)
-            {
-                float val0 = *tmpptr++;
-                float val1 = *tmpptr++;
-                float val2 = *tmpptr++;
-                float val3 = *tmpptr++;
-                float val4 = *tmpptr++;
-                float val5 = *tmpptr++;
-                float val6 = *tmpptr++;
-                float val7 = *tmpptr++;
-                vfloat32m1_t _w0 = vle32_v_f32m1(kptr0, vl);
-                _sum0 = vfmacc_vf_f32m1(_sum0, val0, _w0, vl);
-                _sum1 = vfmacc_vf_f32m1(_sum1, val1, _w0, vl);
-                _sum2 = vfmacc_vf_f32m1(_sum2, val2, _w0, vl);
-                _sum3 = vfmacc_vf_f32m1(_sum3, val3, _w0, vl);
-                _sum4 = vfmacc_vf_f32m1(_sum4, val4, _w0, vl);
-                _sum5 = vfmacc_vf_f32m1(_sum5, val5, _w0, vl);
-                _sum6 = vfmacc_vf_f32m1(_sum6, val6, _w0, vl);
-                _sum7 = vfmacc_vf_f32m1(_sum7, val7, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-            vse32_v_f32m1(outptr0, _sum0, vl);
-            vse32_v_f32m1(outptr0 + packn, _sum1, vl);
-            vse32_v_f32m1(outptr0 + packn * 2, _sum2, vl);
-            vse32_v_f32m1(outptr0 + packn * 3, _sum3, vl);
-            vse32_v_f32m1(outptr0 + packn * 4, _sum4, vl);
-            vse32_v_f32m1(outptr0 + packn * 5, _sum5, vl);
-            vse32_v_f32m1(outptr0 + packn * 6, _sum6, vl);
-            vse32_v_f32m1(outptr0 + packn * 7, _sum7, vl);
-
-            outptr0 += packn * 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 8 + (i % 8) / 4);
-            const float* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat32m1_t _sum0 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum1 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum2 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum3 = vfmv_v_f_f32m1(0.f, vl);
-
-            if (bias)
-            {
-                _sum0 = vle32_v_f32m1(bias + p * packn, vl);
-                _sum1 = vle32_v_f32m1(bias + p * packn, vl);
-                _sum2 = vle32_v_f32m1(bias + p * packn, vl);
-                _sum3 = vle32_v_f32m1(bias + p * packn, vl);
-            }
-
-            for (int j = 0; j < nn; j++)
-            {
-                float val0 = *tmpptr++;
-                float val1 = *tmpptr++;
-                float val2 = *tmpptr++;
-                float val3 = *tmpptr++;
-                vfloat32m1_t _w0 = vle32_v_f32m1(kptr0, vl);
-                _sum0 = vfmacc_vf_f32m1(_sum0, val0, _w0, vl);
-                _sum1 = vfmacc_vf_f32m1(_sum1, val1, _w0, vl);
-                _sum2 = vfmacc_vf_f32m1(_sum2, val2, _w0, vl);
-                _sum3 = vfmacc_vf_f32m1(_sum3, val3, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-            vse32_v_f32m1(outptr0, _sum0, vl);
-            vse32_v_f32m1(outptr0 + packn, _sum1, vl);
-            vse32_v_f32m1(outptr0 + packn * 2, _sum2, vl);
-            vse32_v_f32m1(outptr0 + packn * 3, _sum3, vl);
-
-            outptr0 += packn * 4;
-        }
-        for (; i + 1 < size; i += 2)
-        {
-            const float* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2);
-            const float* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat32m1_t _sum0 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum1 = vfmv_v_f_f32m1(0.f, vl);
-
-            if (bias)
-            {
-                _sum0 = vle32_v_f32m1(bias + p * packn, vl);
-                _sum1 = vle32_v_f32m1(bias + p * packn, vl);
-            }
-
-            for (int j = 0; j < nn; j++)
-            {
-                float val0 = *tmpptr++;
-                float val1 = *tmpptr++;
-                vfloat32m1_t _w0 = vle32_v_f32m1(kptr0, vl);
-                _sum0 = vfmacc_vf_f32m1(_sum0, val0, _w0, vl);
-                _sum1 = vfmacc_vf_f32m1(_sum1, val1, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-            vse32_v_f32m1(outptr0, _sum0, vl);
-            vse32_v_f32m1(outptr0 + packn, _sum1, vl);
-
-            outptr0 += packn * 2;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2 + i % 2);
-            const float* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
-
-            if (bias)
-            {
-                _sum = vle32_v_f32m1(bias + p * packn, vl);
-            }
-
-            for (int j = 0; j < nn; j++)
-            {
-                float val = *tmpptr++;
-                vfloat32m1_t _w0 = vle32_v_f32m1(kptr0, vl);
-                _sum = vfmacc_vf_f32m1(_sum, val, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-            vse32_v_f32m1(outptr0, _sum, vl);
-
-            outptr0 += packn;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_packn_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 4u * packn, packn, opt.workspace_allocator);
-    {
-        const int gap = (w * stride_h - outw * stride_w) * packn;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            float* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const float* sptr = img.row<const float>(dilation_h * u) + dilation_w * v * packn;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            vfloat32m1_t _val = vle32_v_f32m1(sptr, vl);
-                            vse32_v_f32m1(ptr, _val, vl);
-
-                            sptr += stride_w * packn;
-                            ptr += packn;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_packn_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packn_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packn_fp16s.h
deleted file mode 100644
index cb3b6519..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packn_fp16s.h
+++ /dev/null
@@ -1,486 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_packn_fp16sa_rvv(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    // Mat bottom_im2col(size, maxk, inch, 2u * packn, packn, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    const __fp16* bias = _bias;
-
-    // permute
-    Mat tmp;
-    if (size >= 8)
-        tmp.create(8 * maxk, inch, size / 8 + (size % 8) / 4 + (size % 4) / 2 + size % 2, 2u * packn, packn, opt.workspace_allocator);
-    else if (size >= 4)
-        tmp.create(4 * maxk, inch, size / 4 + (size % 4) / 2 + size % 2, 2u * packn, packn, opt.workspace_allocator);
-    else if (size >= 2)
-        tmp.create(2 * maxk, inch, size / 2 + size % 2, 2u * packn, packn, opt.workspace_allocator);
-    else
-        tmp.create(maxk, inch, size, 2u * packn, packn, opt.workspace_allocator);
-    {
-        int remain_size_start = 0;
-        int nn_size = size >> 3;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 8;
-
-            __fp16* tmpptr = tmp.channel(i / 8);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const __fp16* img0 = (const __fp16*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-#if C906
-#ifdef RVV_SPEC_0_7
-                    asm volatile(
-                        "mv        t3,     %[LEN]  \n\t"
-                        "mv        t1,     %[SRC]  \n\t"
-                        "mv        t2,     %[TMP]  \n\t"
-                        "slli      t3,     t3,     1       \n\t"
-                        "vle.v     v0,     (t1)    \n\t"
-                        "add       t1,     t1,     t3      \n\t"
-                        "vle.v     v1,     (t1)    \n\t"
-                        "add       t1,     t1,     t3      \n\t"
-                        "vle.v     v2,     (t1)    \n\t"
-                        "add       t1,     t1,     t3      \n\t"
-                        "vle.v     v3,     (t1)    \n\t"
-                        "add       t1,     t1,     t3      \n\t"
-                        "vle.v     v4,     (t1)    \n\t"
-                        "add       t1,     t1,     t3      \n\t"
-                        "vle.v     v5,     (t1)    \n\t"
-                        "add       t1,     t1,     t3      \n\t"
-                        "vle.v     v6,     (t1)    \n\t"
-                        "add       t1,     t1,     t3      \n\t"
-                        "vle.v     v7,     (t1)    \n\t"
-                        "add       t1,     t1,     t3      \n\t"
-                        "vsseg8e.v v0,     (t2)    \n\t"
-                        :
-                        : [LEN] "r"(packn), [SRC] "r"(img0), [TMP] "r"(tmpptr)
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "t1", "t2", "t3");
-
-                    img0 += size * packn;
-                    tmpptr += packn * 8;
-#else
-                    for (int l = 0; l < packn; l++)
-                    {
-                        tmpptr[0] = img0[l];
-                        tmpptr[1] = img0[l + packn];
-                        tmpptr[2] = img0[l + packn * 2];
-                        tmpptr[3] = img0[l + packn * 3];
-                        tmpptr[4] = img0[l + packn * 4];
-                        tmpptr[5] = img0[l + packn * 5];
-                        tmpptr[6] = img0[l + packn * 6];
-                        tmpptr[7] = img0[l + packn * 7];
-                        tmpptr += 8;
-                    }
-
-                    img0 += size * packn;
-#endif
-#else
-                    vfloat16m1_t _val0 = vle16_v_f16m1(img0, vl);
-                    vfloat16m1_t _val1 = vle16_v_f16m1(img0 + packn, vl);
-                    vfloat16m1_t _val2 = vle16_v_f16m1(img0 + packn * 2, vl);
-                    vfloat16m1_t _val3 = vle16_v_f16m1(img0 + packn * 3, vl);
-                    vfloat16m1_t _val4 = vle16_v_f16m1(img0 + packn * 4, vl);
-                    vfloat16m1_t _val5 = vle16_v_f16m1(img0 + packn * 5, vl);
-                    vfloat16m1_t _val6 = vle16_v_f16m1(img0 + packn * 6, vl);
-                    vfloat16m1_t _val7 = vle16_v_f16m1(img0 + packn * 7, vl);
-                    vsseg8e16_v_f16m1(tmpptr, _val0, _val1, _val2, _val3, _val4, _val5, _val6, _val7, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn * 8;
-#endif
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 3;
-        nn_size = (size - remain_size_start) >> 2;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 4;
-
-            __fp16* tmpptr = tmp.channel(i / 8 + (i % 8) / 4);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const __fp16* img0 = (const __fp16*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-#if C906
-#ifdef RVV_SPEC_0_7
-                    asm volatile(
-                        "mv        t3,     %[LEN]  \n\t"
-                        "mv        t1,     %[SRC]  \n\t"
-                        "mv        t2,     %[TMP]  \n\t"
-                        "slli      t3,     t3,     1       \n\t"
-                        "vle.v     v0,     (t1)    \n\t"
-                        "add       t1,     t1,     t3      \n\t"
-                        "vle.v     v1,     (t1)    \n\t"
-                        "add       t1,     t1,     t3      \n\t"
-                        "vle.v     v2,     (t1)    \n\t"
-                        "add       t1,     t1,     t3      \n\t"
-                        "vle.v     v3,     (t1)    \n\t"
-                        "vsseg4e.v v0,     (t2)    \n\t"
-                        :
-                        : [LEN] "r"(packn), [SRC] "r"(img0), [TMP] "r"(tmpptr)
-                        : "cc", "memory", "v0", "v1", "v2", "v3", "t1", "t2", "t3");
-
-                    img0 += size * packn;
-                    tmpptr += packn * 4;
-#else
-                    for (int l = 0; l < packn; l++)
-                    {
-                        tmpptr[0] = img0[l];
-                        tmpptr[1] = img0[l + packn];
-                        tmpptr[2] = img0[l + packn * 2];
-                        tmpptr[3] = img0[l + packn * 3];
-                        tmpptr += 4;
-                    }
-
-                    img0 += size * packn;
-#endif
-#else
-                    vfloat16m1_t _val0 = vle16_v_f16m1(img0, vl);
-                    vfloat16m1_t _val1 = vle16_v_f16m1(img0 + packn, vl);
-                    vfloat16m1_t _val2 = vle16_v_f16m1(img0 + packn * 2, vl);
-                    vfloat16m1_t _val3 = vle16_v_f16m1(img0 + packn * 3, vl);
-                    vsseg4e16_v_f16m1(tmpptr, _val0, _val1, _val2, _val3, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn * 4;
-#endif
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 2;
-
-        nn_size = (size - remain_size_start) >> 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 2;
-
-            __fp16* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const __fp16* img0 = (const __fp16*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-#if C906
-#ifdef RVV_SPEC_0_7
-                    asm volatile(
-                        "mv        t3,     %[LEN]  \n\t"
-                        "mv        t1,     %[SRC]  \n\t"
-                        "mv        t2,     %[TMP]  \n\t"
-                        "slli      t3,     t3,     1       \n\t"
-                        "vle.v     v0,     (t1)    \n\t"
-                        "add       t1,     t1,     t3      \n\t"
-                        "vle.v     v1,     (t1)    \n\t"
-                        "add       t1,     t1,     t3      \n\t"
-                        :
-                        : [LEN] "r"(packn), [SRC] "r"(img0), [TMP] "r"(tmpptr)
-                        : "cc", "memory", "v0", "v1", "t1", "t2", "t3");
-
-                    img0 += size * packn;
-                    tmpptr += packn * 2;
-#else
-                    for (int l = 0; l < packn; l++)
-                    {
-                        tmpptr[0] = img0[l];
-                        tmpptr[1] = img0[l + packn];
-                        tmpptr += 2;
-                    }
-
-                    img0 += size * packn;
-#endif
-#else
-                    vfloat16m1_t _val0 = vle16_v_f16m1(img0, vl);
-                    vfloat16m1_t _val1 = vle16_v_f16m1(img0 + packn, vl);
-                    vsseg2e16_v_f16m1(tmpptr, _val0, _val1, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn * 2;
-#endif
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            __fp16* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2 + i % 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const __fp16* img0 = (const __fp16*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    vfloat16m1_t _val = vle16_v_f16m1(img0, vl);
-                    vse16_v_f16m1(tmpptr, _val, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn;
-                }
-            }
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr0 = top_blob.channel(p);
-
-        int i = 0;
-        for (; i + 7 < size; i += 8)
-        {
-            const __fp16* tmpptr = tmp.channel(i / 8);
-            const __fp16* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat16m1_t _sum0 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum1 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum2 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum3 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum4 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum5 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum6 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum7 = vfmv_v_f_f16m1(0.f, vl);
-
-            if (bias)
-            {
-                _sum0 = vle16_v_f16m1(bias + p * packn, vl);
-                _sum1 = vle16_v_f16m1(bias + p * packn, vl);
-                _sum2 = vle16_v_f16m1(bias + p * packn, vl);
-                _sum3 = vle16_v_f16m1(bias + p * packn, vl);
-                _sum4 = vle16_v_f16m1(bias + p * packn, vl);
-                _sum5 = vle16_v_f16m1(bias + p * packn, vl);
-                _sum6 = vle16_v_f16m1(bias + p * packn, vl);
-                _sum7 = vle16_v_f16m1(bias + p * packn, vl);
-            }
-
-            for (int j = 0; j < nn; j++)
-            {
-                __fp16 val0 = *tmpptr++;
-                __fp16 val1 = *tmpptr++;
-                __fp16 val2 = *tmpptr++;
-                __fp16 val3 = *tmpptr++;
-                __fp16 val4 = *tmpptr++;
-                __fp16 val5 = *tmpptr++;
-                __fp16 val6 = *tmpptr++;
-                __fp16 val7 = *tmpptr++;
-                vfloat16m1_t _w0 = vle16_v_f16m1(kptr0, vl);
-                _sum0 = vfmacc_vf_f16m1(_sum0, val0, _w0, vl);
-                _sum1 = vfmacc_vf_f16m1(_sum1, val1, _w0, vl);
-                _sum2 = vfmacc_vf_f16m1(_sum2, val2, _w0, vl);
-                _sum3 = vfmacc_vf_f16m1(_sum3, val3, _w0, vl);
-                _sum4 = vfmacc_vf_f16m1(_sum4, val4, _w0, vl);
-                _sum5 = vfmacc_vf_f16m1(_sum5, val5, _w0, vl);
-                _sum6 = vfmacc_vf_f16m1(_sum6, val6, _w0, vl);
-                _sum7 = vfmacc_vf_f16m1(_sum7, val7, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-            vse16_v_f16m1(outptr0, _sum0, vl);
-            vse16_v_f16m1(outptr0 + packn, _sum1, vl);
-            vse16_v_f16m1(outptr0 + packn * 2, _sum2, vl);
-            vse16_v_f16m1(outptr0 + packn * 3, _sum3, vl);
-            vse16_v_f16m1(outptr0 + packn * 4, _sum4, vl);
-            vse16_v_f16m1(outptr0 + packn * 5, _sum5, vl);
-            vse16_v_f16m1(outptr0 + packn * 6, _sum6, vl);
-            vse16_v_f16m1(outptr0 + packn * 7, _sum7, vl);
-
-            outptr0 += packn * 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-            const __fp16* tmpptr = tmp.channel(i / 8 + (i % 8) / 4);
-            const __fp16* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat16m1_t _sum0 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum1 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum2 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum3 = vfmv_v_f_f16m1(0.f, vl);
-
-            if (bias)
-            {
-                _sum0 = vle16_v_f16m1(bias + p * packn, vl);
-                _sum1 = vle16_v_f16m1(bias + p * packn, vl);
-                _sum2 = vle16_v_f16m1(bias + p * packn, vl);
-                _sum3 = vle16_v_f16m1(bias + p * packn, vl);
-            }
-
-            for (int j = 0; j < nn; j++)
-            {
-                __fp16 val0 = *tmpptr++;
-                __fp16 val1 = *tmpptr++;
-                __fp16 val2 = *tmpptr++;
-                __fp16 val3 = *tmpptr++;
-                vfloat16m1_t _w0 = vle16_v_f16m1(kptr0, vl);
-                _sum0 = vfmacc_vf_f16m1(_sum0, val0, _w0, vl);
-                _sum1 = vfmacc_vf_f16m1(_sum1, val1, _w0, vl);
-                _sum2 = vfmacc_vf_f16m1(_sum2, val2, _w0, vl);
-                _sum3 = vfmacc_vf_f16m1(_sum3, val3, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-            vse16_v_f16m1(outptr0, _sum0, vl);
-            vse16_v_f16m1(outptr0 + packn, _sum1, vl);
-            vse16_v_f16m1(outptr0 + packn * 2, _sum2, vl);
-            vse16_v_f16m1(outptr0 + packn * 3, _sum3, vl);
-
-            outptr0 += packn * 4;
-        }
-        for (; i + 1 < size; i += 2)
-        {
-            const __fp16* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2);
-            const __fp16* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat16m1_t _sum0 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum1 = vfmv_v_f_f16m1(0.f, vl);
-
-            if (bias)
-            {
-                _sum0 = vle16_v_f16m1(bias + p * packn, vl);
-                _sum1 = vle16_v_f16m1(bias + p * packn, vl);
-            }
-
-            for (int j = 0; j < nn; j++)
-            {
-                __fp16 val0 = *tmpptr++;
-                __fp16 val1 = *tmpptr++;
-                vfloat16m1_t _w0 = vle16_v_f16m1(kptr0, vl);
-                _sum0 = vfmacc_vf_f16m1(_sum0, val0, _w0, vl);
-                _sum1 = vfmacc_vf_f16m1(_sum1, val1, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-            vse16_v_f16m1(outptr0, _sum0, vl);
-            vse16_v_f16m1(outptr0 + packn, _sum1, vl);
-
-            outptr0 += packn * 2;
-        }
-        for (; i < size; i++)
-        {
-            const __fp16* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2 + i % 2);
-            const __fp16* kptr0 = kernel.channel(p);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat16m1_t _sum = vfmv_v_f_f16m1(0.f, vl);
-
-            if (bias)
-            {
-                _sum = vle16_v_f16m1(bias + p * packn, vl);
-            }
-
-            for (int j = 0; j < nn; j++)
-            {
-                __fp16 val = *tmpptr++;
-                vfloat16m1_t _w0 = vle16_v_f16m1(kptr0, vl);
-                _sum = vfmacc_vf_f16m1(_sum, val, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-            vse16_v_f16m1(outptr0, _sum, vl);
-
-            outptr0 += packn;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_packn_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 2u * packn, packn, opt.workspace_allocator);
-    {
-        const int gap = (w * stride_h - outw * stride_w) * packn;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            __fp16* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const __fp16* sptr = img.row<const __fp16>(dilation_h * u) + dilation_w * v * packn;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            vfloat16m1_t _val = vle16_v_f16m1(sptr, vl);
-                            vse16_v_f16m1(ptr, _val, vl);
-
-                            sptr += stride_w * packn;
-                            ptr += packn;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_packn_fp16sa_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packnto1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packnto1.h
deleted file mode 100644
index 2df2c7d7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packnto1.h
+++ /dev/null
@@ -1,728 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_packnto1_rvv(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    // Mat bottom_im2col(size, maxk, inch, 4u * packn, packn, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    Mat tmp;
-    if (size >= 8)
-        tmp.create(8 * maxk, inch, size / 8 + (size % 8) / 4 + (size % 4) / 2 + size % 2, 4u * packn, packn, opt.workspace_allocator);
-    else if (size >= 4)
-        tmp.create(4 * maxk, inch, size / 4 + (size % 4) / 2 + size % 2, 4u * packn, packn, opt.workspace_allocator);
-    else if (size >= 2)
-        tmp.create(2 * maxk, inch, size / 2 + size % 2, 4u * packn, packn, opt.workspace_allocator);
-    else
-        tmp.create(maxk, inch, size, 4u * packn, packn, opt.workspace_allocator);
-    {
-        int remain_size_start = 0;
-        int nn_size = size >> 3;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 8;
-
-            float* tmpptr = tmp.channel(i / 8);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-#if C906
-                    for (int l = 0; l < packn; l++)
-                    {
-                        tmpptr[0] = img0[l];
-                        tmpptr[1] = img0[l + packn];
-                        tmpptr[2] = img0[l + packn * 2];
-                        tmpptr[3] = img0[l + packn * 3];
-                        tmpptr[4] = img0[l + packn * 4];
-                        tmpptr[5] = img0[l + packn * 5];
-                        tmpptr[6] = img0[l + packn * 6];
-                        tmpptr[7] = img0[l + packn * 7];
-                        tmpptr += 8;
-                    }
-
-                    img0 += size * packn;
-#else
-                    vfloat32m1_t _val0 = vle32_v_f32m1(img0, vl);
-                    vfloat32m1_t _val1 = vle32_v_f32m1(img0 + packn, vl);
-                    vfloat32m1_t _val2 = vle32_v_f32m1(img0 + packn * 2, vl);
-                    vfloat32m1_t _val3 = vle32_v_f32m1(img0 + packn * 3, vl);
-                    vfloat32m1_t _val4 = vle32_v_f32m1(img0 + packn * 4, vl);
-                    vfloat32m1_t _val5 = vle32_v_f32m1(img0 + packn * 5, vl);
-                    vfloat32m1_t _val6 = vle32_v_f32m1(img0 + packn * 6, vl);
-                    vfloat32m1_t _val7 = vle32_v_f32m1(img0 + packn * 7, vl);
-                    vsseg8e32_v_f32m1(tmpptr, _val0, _val1, _val2, _val3, _val4, _val5, _val6, _val7, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn * 8;
-#endif
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 3;
-        nn_size = (size - remain_size_start) >> 2;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 4;
-
-            float* tmpptr = tmp.channel(i / 8 + (i % 8) / 4);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-#if C906
-                    for (int l = 0; l < packn; l++)
-                    {
-                        tmpptr[0] = img0[l];
-                        tmpptr[1] = img0[l + packn];
-                        tmpptr[2] = img0[l + packn * 2];
-                        tmpptr[3] = img0[l + packn * 3];
-                        tmpptr += 4;
-                    }
-
-                    img0 += size * packn;
-#else
-                    vfloat32m1_t _val0 = vle32_v_f32m1(img0, vl);
-                    vfloat32m1_t _val1 = vle32_v_f32m1(img0 + packn, vl);
-                    vfloat32m1_t _val2 = vle32_v_f32m1(img0 + packn * 2, vl);
-                    vfloat32m1_t _val3 = vle32_v_f32m1(img0 + packn * 3, vl);
-                    vsseg4e32_v_f32m1(tmpptr, _val0, _val1, _val2, _val3, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn * 4;
-#endif
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 2;
-        nn_size = (size - remain_size_start) >> 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 2;
-
-            float* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-#if C906
-                    for (int l = 0; l < packn; l++)
-                    {
-                        tmpptr[0] = img0[l];
-                        tmpptr[1] = img0[l + packn];
-                        tmpptr += 2;
-                    }
-
-                    img0 += size * packn;
-#else
-                    vfloat32m1_t _val0 = vle32_v_f32m1(img0, vl);
-                    vfloat32m1_t _val1 = vle32_v_f32m1(img0 + packn, vl);
-                    vsseg2e32_v_f32m1(tmpptr, _val0, _val1, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn * 2;
-#endif
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            float* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2 + i % 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const float* img0 = (const float*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    vfloat32m1_t _val = vle32_v_f32m1(img0, vl);
-                    vse32_v_f32m1(tmpptr, _val, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn;
-                }
-            }
-        }
-    }
-
-    int nn_outch = outch / packn;
-    int remain_outch_start = nn_outch * packn;
-
-#ifdef __clang__
-    // clang complains about VLA in the following loop
-    float* _zero_tmp = new float[packn]();
-    for (int _zero_clean_idx = 0; _zero_clean_idx < packn; _zero_clean_idx++)
-    {
-        _zero_tmp[_zero_clean_idx] = 0.f;
-    }
-#endif // __clang__
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * packn;
-
-        float* outptr0 = top_blob.channel(p);
-
-#ifdef __clang__
-        const float* zeros = _zero_tmp;
-#else
-        const float zeros[packn] = {0.f};
-#endif // __clang__
-        const float* biasptr = bias ? bias + p : zeros;
-
-        int i = 0;
-        for (; i + 7 < size; i += 8)
-        {
-            const float* tmpptr = tmp.channel(i / 8);
-            const float* kptr0 = kernel.channel(p / packn);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat32m1_t _sum0 = vle32_v_f32m1(biasptr, vl);
-            vfloat32m1_t _sum1 = vle32_v_f32m1(biasptr, vl);
-            vfloat32m1_t _sum2 = vle32_v_f32m1(biasptr, vl);
-            vfloat32m1_t _sum3 = vle32_v_f32m1(biasptr, vl);
-            vfloat32m1_t _sum4 = vle32_v_f32m1(biasptr, vl);
-            vfloat32m1_t _sum5 = vle32_v_f32m1(biasptr, vl);
-            vfloat32m1_t _sum6 = vle32_v_f32m1(biasptr, vl);
-            vfloat32m1_t _sum7 = vle32_v_f32m1(biasptr, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                float val0 = *tmpptr++;
-                float val1 = *tmpptr++;
-                float val2 = *tmpptr++;
-                float val3 = *tmpptr++;
-                float val4 = *tmpptr++;
-                float val5 = *tmpptr++;
-                float val6 = *tmpptr++;
-                float val7 = *tmpptr++;
-                vfloat32m1_t _w0 = vle32_v_f32m1(kptr0, vl);
-                _sum0 = vfmacc_vf_f32m1(_sum0, val0, _w0, vl);
-                _sum1 = vfmacc_vf_f32m1(_sum1, val1, _w0, vl);
-                _sum2 = vfmacc_vf_f32m1(_sum2, val2, _w0, vl);
-                _sum3 = vfmacc_vf_f32m1(_sum3, val3, _w0, vl);
-                _sum4 = vfmacc_vf_f32m1(_sum4, val4, _w0, vl);
-                _sum5 = vfmacc_vf_f32m1(_sum5, val5, _w0, vl);
-                _sum6 = vfmacc_vf_f32m1(_sum6, val6, _w0, vl);
-                _sum7 = vfmacc_vf_f32m1(_sum7, val7, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-#if C906
-            vsse32_v_f32m1(outptr0, top_blob.cstep * sizeof(float), _sum0, vl);
-            vsse32_v_f32m1(outptr0 + 1, top_blob.cstep * sizeof(float), _sum1, vl);
-            vsse32_v_f32m1(outptr0 + 2, top_blob.cstep * sizeof(float), _sum2, vl);
-            vsse32_v_f32m1(outptr0 + 3, top_blob.cstep * sizeof(float), _sum3, vl);
-            vsse32_v_f32m1(outptr0 + 4, top_blob.cstep * sizeof(float), _sum4, vl);
-            vsse32_v_f32m1(outptr0 + 5, top_blob.cstep * sizeof(float), _sum5, vl);
-            vsse32_v_f32m1(outptr0 + 6, top_blob.cstep * sizeof(float), _sum6, vl);
-            vsse32_v_f32m1(outptr0 + 7, top_blob.cstep * sizeof(float), _sum7, vl);
-#else
-            vssseg8e32_v_f32m1(outptr0, top_blob.cstep * sizeof(float), _sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7, vl);
-#endif
-            outptr0 += 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 8 + (i % 8) / 4);
-            const float* kptr0 = kernel.channel(p / packn);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat32m1_t _sum0 = vle32_v_f32m1(biasptr, vl);
-            vfloat32m1_t _sum1 = vle32_v_f32m1(biasptr, vl);
-            vfloat32m1_t _sum2 = vle32_v_f32m1(biasptr, vl);
-            vfloat32m1_t _sum3 = vle32_v_f32m1(biasptr, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                float val0 = *tmpptr++;
-                float val1 = *tmpptr++;
-                float val2 = *tmpptr++;
-                float val3 = *tmpptr++;
-                vfloat32m1_t _w0 = vle32_v_f32m1(kptr0, vl);
-                _sum0 = vfmacc_vf_f32m1(_sum0, val0, _w0, vl);
-                _sum1 = vfmacc_vf_f32m1(_sum1, val1, _w0, vl);
-                _sum2 = vfmacc_vf_f32m1(_sum2, val2, _w0, vl);
-                _sum3 = vfmacc_vf_f32m1(_sum3, val3, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-#if C906
-            vsse32_v_f32m1(outptr0, top_blob.cstep * sizeof(float), _sum0, vl);
-            vsse32_v_f32m1(outptr0 + 1, top_blob.cstep * sizeof(float), _sum1, vl);
-            vsse32_v_f32m1(outptr0 + 2, top_blob.cstep * sizeof(float), _sum2, vl);
-            vsse32_v_f32m1(outptr0 + 3, top_blob.cstep * sizeof(float), _sum3, vl);
-#else
-            vssseg4e32_v_f32m1(outptr0, top_blob.cstep * sizeof(float), _sum0, _sum1, _sum2, _sum3, vl);
-#endif
-            outptr0 += 4;
-        }
-        for (; i + 1 < size; i += 2)
-        {
-            const float* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2);
-            const float* kptr0 = kernel.channel(p / packn);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat32m1_t _sum0 = vle32_v_f32m1(biasptr, vl);
-            vfloat32m1_t _sum1 = vle32_v_f32m1(biasptr, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                float val0 = *tmpptr++;
-                float val1 = *tmpptr++;
-                vfloat32m1_t _w0 = vle32_v_f32m1(kptr0, vl);
-                _sum0 = vfmacc_vf_f32m1(_sum0, val0, _w0, vl);
-                _sum1 = vfmacc_vf_f32m1(_sum1, val1, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-#if C906
-            vsse32_v_f32m1(outptr0, top_blob.cstep * sizeof(float), _sum0, vl);
-            vsse32_v_f32m1(outptr0 + 1, top_blob.cstep * sizeof(float), _sum1, vl);
-#else
-            vssseg2e32_v_f32m1(outptr0, top_blob.cstep * sizeof(float), _sum0, _sum1, vl);
-#endif
-            outptr0 += 2;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2 + i % 2);
-            const float* kptr0 = kernel.channel(p / packn);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat32m1_t _sum = vle32_v_f32m1(biasptr, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                float val = *tmpptr++;
-                vfloat32m1_t _w0 = vle32_v_f32m1(kptr0, vl);
-                _sum = vfmacc_vf_f32m1(_sum, val, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-            vsse32_v_f32m1(outptr0, top_blob.cstep * sizeof(float), _sum, vl);
-
-            outptr0 += 1;
-        }
-    }
-#ifdef __clang__
-    delete[] _zero_tmp;
-#endif
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        float* outptr0 = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        int i = 0;
-        for (; i + 7 < size; i += 8)
-        {
-            const float* tmpptr = tmp.channel(i / 8);
-            const float* kptr0 = kernel.channel(p / packn + p % packn);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = bias0;
-            float sum1 = bias0;
-            float sum2 = bias0;
-            float sum3 = bias0;
-            float sum4 = bias0;
-            float sum5 = bias0;
-            float sum6 = bias0;
-            float sum7 = bias0;
-
-            vfloat32m1_t _sum0 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum1 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum2 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum3 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum4 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum5 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum6 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum7 = vfmv_v_f_f32m1(0.f, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                vfloat32m1_t _val0;
-                vfloat32m1_t _val1;
-                vfloat32m1_t _val2;
-                vfloat32m1_t _val3;
-                vfloat32m1_t _val4;
-                vfloat32m1_t _val5;
-                vfloat32m1_t _val6;
-                vfloat32m1_t _val7;
-                vlseg8e32_v_f32m1(&_val0, &_val1, &_val2, &_val3, &_val4, &_val5, &_val6, &_val7, tmpptr, vl);
-                vfloat32m1_t _w0 = vle32_v_f32m1(kptr0, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _val0, _w0, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _val1, _w0, vl);
-                _sum2 = vfmacc_vv_f32m1(_sum2, _val2, _w0, vl);
-                _sum3 = vfmacc_vv_f32m1(_sum3, _val3, _w0, vl);
-                _sum4 = vfmacc_vv_f32m1(_sum4, _val4, _w0, vl);
-                _sum5 = vfmacc_vv_f32m1(_sum5, _val5, _w0, vl);
-                _sum6 = vfmacc_vv_f32m1(_sum6, _val6, _w0, vl);
-                _sum7 = vfmacc_vv_f32m1(_sum7, _val7, _w0, vl);
-                tmpptr += packn * 8;
-                kptr0 += packn;
-            }
-
-#if C906
-            // TODO
-            std::vector<float> ss0(packn);
-            std::vector<float> ss1(packn);
-            std::vector<float> ss2(packn);
-            std::vector<float> ss3(packn);
-            std::vector<float> ss4(packn);
-            std::vector<float> ss5(packn);
-            std::vector<float> ss6(packn);
-            std::vector<float> ss7(packn);
-            vse32_v_f32m1((float*)ss0.data(), _sum0, vl);
-            vse32_v_f32m1((float*)ss1.data(), _sum1, vl);
-            vse32_v_f32m1((float*)ss2.data(), _sum2, vl);
-            vse32_v_f32m1((float*)ss3.data(), _sum3, vl);
-            vse32_v_f32m1((float*)ss4.data(), _sum4, vl);
-            vse32_v_f32m1((float*)ss5.data(), _sum5, vl);
-            vse32_v_f32m1((float*)ss6.data(), _sum6, vl);
-            vse32_v_f32m1((float*)ss7.data(), _sum7, vl);
-            for (int i = 0; i < packn; i++)
-            {
-                sum0 += ss0[i];
-                sum1 += ss1[i];
-                sum2 += ss2[i];
-                sum3 += ss3[i];
-                sum4 += ss4[i];
-                sum5 += ss5[i];
-                sum6 += ss6[i];
-                sum7 += ss7[i];
-            }
-#else
-            sum0 = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum0, vfmv_s_f_f32m1(vfloat32m1_t(), sum0, vl), vl));
-            sum1 = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum1, vfmv_s_f_f32m1(vfloat32m1_t(), sum1, vl), vl));
-            sum2 = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum2, vfmv_s_f_f32m1(vfloat32m1_t(), sum2, vl), vl));
-            sum3 = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum3, vfmv_s_f_f32m1(vfloat32m1_t(), sum3, vl), vl));
-            sum4 = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum4, vfmv_s_f_f32m1(vfloat32m1_t(), sum4, vl), vl));
-            sum5 = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum5, vfmv_s_f_f32m1(vfloat32m1_t(), sum5, vl), vl));
-            sum6 = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum6, vfmv_s_f_f32m1(vfloat32m1_t(), sum6, vl), vl));
-            sum7 = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum7, vfmv_s_f_f32m1(vfloat32m1_t(), sum7, vl), vl));
-#endif
-
-            outptr0[0] = sum0;
-            outptr0[1] = sum1;
-            outptr0[2] = sum2;
-            outptr0[3] = sum3;
-            outptr0[4] = sum4;
-            outptr0[5] = sum5;
-            outptr0[6] = sum6;
-            outptr0[7] = sum7;
-
-            outptr0 += 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-            const float* tmpptr = tmp.channel(i / 8 + (i % 8) / 4);
-            const float* kptr0 = kernel.channel(p / packn + p % packn);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = bias0;
-            float sum1 = bias0;
-            float sum2 = bias0;
-            float sum3 = bias0;
-
-            vfloat32m1_t _sum0 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum1 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum2 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum3 = vfmv_v_f_f32m1(0.f, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                vfloat32m1_t _val0;
-                vfloat32m1_t _val1;
-                vfloat32m1_t _val2;
-                vfloat32m1_t _val3;
-                vlseg4e32_v_f32m1(&_val0, &_val1, &_val2, &_val3, tmpptr, vl);
-                vfloat32m1_t _w0 = vle32_v_f32m1(kptr0, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _val0, _w0, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _val1, _w0, vl);
-                _sum2 = vfmacc_vv_f32m1(_sum2, _val2, _w0, vl);
-                _sum3 = vfmacc_vv_f32m1(_sum3, _val3, _w0, vl);
-                tmpptr += packn * 4;
-                kptr0 += packn;
-            }
-
-#if C906
-            // TODO
-            std::vector<float> ss0(packn);
-            std::vector<float> ss1(packn);
-            std::vector<float> ss2(packn);
-            std::vector<float> ss3(packn);
-            vse32_v_f32m1((float*)ss0.data(), _sum0, vl);
-            vse32_v_f32m1((float*)ss1.data(), _sum1, vl);
-            vse32_v_f32m1((float*)ss2.data(), _sum2, vl);
-            vse32_v_f32m1((float*)ss3.data(), _sum3, vl);
-            for (int i = 0; i < packn; i++)
-            {
-                sum0 += ss0[i];
-                sum1 += ss1[i];
-                sum2 += ss2[i];
-                sum3 += ss3[i];
-            }
-#else
-            sum0 = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum0, vfmv_s_f_f32m1(vfloat32m1_t(), sum0, vl), vl));
-            sum1 = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum1, vfmv_s_f_f32m1(vfloat32m1_t(), sum1, vl), vl));
-            sum2 = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum2, vfmv_s_f_f32m1(vfloat32m1_t(), sum2, vl), vl));
-            sum3 = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum3, vfmv_s_f_f32m1(vfloat32m1_t(), sum3, vl), vl));
-#endif
-
-            outptr0[0] = sum0;
-            outptr0[1] = sum1;
-            outptr0[2] = sum2;
-            outptr0[3] = sum3;
-
-            outptr0 += 4;
-        }
-        for (; i + 1 < size; i += 2)
-        {
-            const float* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2);
-            const float* kptr0 = kernel.channel(p / packn + p % packn);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = bias0;
-            float sum1 = bias0;
-
-            vfloat32m1_t _sum0 = vfmv_v_f_f32m1(0.f, vl);
-            vfloat32m1_t _sum1 = vfmv_v_f_f32m1(0.f, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                vfloat32m1_t _val0;
-                vfloat32m1_t _val1;
-                vlseg2e32_v_f32m1(&_val0, &_val1, tmpptr, vl);
-                vfloat32m1_t _w0 = vle32_v_f32m1(kptr0, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _val0, _w0, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _val1, _w0, vl);
-                tmpptr += packn * 2;
-                kptr0 += packn;
-            }
-
-#if C906
-            // TODO
-            std::vector<float> ss0(packn);
-            std::vector<float> ss1(packn);
-            vse32_v_f32m1((float*)ss0.data(), _sum0, vl);
-            vse32_v_f32m1((float*)ss1.data(), _sum1, vl);
-            for (int i = 0; i < packn; i++)
-            {
-                sum0 += ss0[i];
-                sum1 += ss1[i];
-            }
-#else
-            sum0 = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum0, vfmv_s_f_f32m1(vfloat32m1_t(), sum0, vl), vl));
-            sum1 = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum1, vfmv_s_f_f32m1(vfloat32m1_t(), sum1, vl), vl));
-#endif
-
-            outptr0[0] = sum0;
-            outptr0[1] = sum1;
-
-            outptr0 += 2;
-        }
-        for (; i < size; i++)
-        {
-            const float* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2 + i % 2);
-            const float* kptr0 = kernel.channel(p / packn + p % packn);
-
-            int nn = inch * maxk; // inch always > 0
-
-            float sum0 = bias0;
-
-            vfloat32m1_t _sum0 = vfmv_v_f_f32m1(0.f, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                vfloat32m1_t _val0 = vle32_v_f32m1(tmpptr, vl);
-                vfloat32m1_t _w0 = vle32_v_f32m1(kptr0, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _val0, _w0, vl);
-                tmpptr += packn;
-                kptr0 += packn;
-            }
-
-#if C906
-            // TODO
-            std::vector<float> ss0(packn);
-            vse32_v_f32m1((float*)ss0.data(), _sum0, vl);
-            for (int i = 0; i < packn; i++)
-            {
-                sum0 += ss0[i];
-            }
-#else
-            sum0 = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum0, vfmv_s_f_f32m1(vfloat32m1_t(), sum0, vl), vl));
-#endif
-
-            outptr0[0] = sum0;
-
-            outptr0 += 1;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_transform_kernel_packnto1_rvv(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int packn = csrr_vlenb() / 4;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = pb-pa-maxk-inch/pa-outch/pb
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-    kernel_tm.create(packn * packn * maxk, inch / packn, outch / packn + outch % packn);
-
-    int q = 0;
-    for (; q + (packn - 1) < outch; q += packn)
-    {
-        float* g00 = kernel_tm.channel(q / packn);
-
-        for (int p = 0; p + (packn - 1) < inch; p += packn)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < packn; i++)
-                {
-                    for (int j = 0; j < packn; j++)
-                    {
-                        const float* k00 = kernel.channel(q + j).row(p + i);
-
-                        g00[0] = k00[k];
-
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-    for (; q < outch; q++)
-    {
-        const Mat k0 = kernel.channel(q);
-
-        float* g00 = kernel_tm.channel(q / packn + q % packn);
-
-        for (int p = 0; p + (packn - 1) < inch; p += packn)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int j = 0; j < packn; j++)
-                {
-                    const float* k00 = k0.row(p + j);
-
-                    g00[0] = k00[k];
-
-                    g00++;
-                }
-            }
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_packnto1_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 4u * packn, packn, opt.workspace_allocator);
-    {
-        const int gap = (w * stride_h - outw * stride_w) * packn;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            float* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const float* sptr = img.row(dilation_h * u) + dilation_w * v * packn;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            vfloat32m1_t _val = vle32_v_f32m1(sptr, vl);
-                            vse32_v_f32m1(ptr, _val, vl);
-
-                            sptr += stride_w * packn;
-                            ptr += packn;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_packnto1_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packnto1_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packnto1_fp16s.h
deleted file mode 100644
index 925713d9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_sgemm_packnto1_fp16s.h
+++ /dev/null
@@ -1,729 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void im2col_sgemm_packnto1_fp16sa_rvv(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    // Mat bottom_im2col(size, maxk, inch, 2u * packn, packn, opt.workspace_allocator);
-
-    const int size = bottom_im2col.w;
-    const int maxk = bottom_im2col.h;
-    const int inch = bottom_im2col.c;
-
-    const int outch = top_blob.c;
-
-    const __fp16* bias = _bias;
-
-    Mat tmp;
-    if (size >= 8)
-        tmp.create(8 * maxk, inch, size / 8 + (size % 8) / 4 + (size % 4) / 2 + size % 2, 2u * packn, packn, opt.workspace_allocator);
-    else if (size >= 4)
-        tmp.create(4 * maxk, inch, size / 4 + (size % 4) / 2 + size % 2, 2u * packn, packn, opt.workspace_allocator);
-    else if (size >= 2)
-        tmp.create(2 * maxk, inch, size / 2 + size % 2, 2u * packn, packn, opt.workspace_allocator);
-    else
-        tmp.create(maxk, inch, size, 2u * packn, packn, opt.workspace_allocator);
-    {
-        int remain_size_start = 0;
-        int nn_size = size >> 3;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 8;
-
-            __fp16* tmpptr = tmp.channel(i / 8);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const __fp16* img0 = (const __fp16*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-#if C906
-                    for (int l = 0; l < packn; l++)
-                    {
-                        tmpptr[0] = img0[l];
-                        tmpptr[1] = img0[l + packn];
-                        tmpptr[2] = img0[l + packn * 2];
-                        tmpptr[3] = img0[l + packn * 3];
-                        tmpptr[4] = img0[l + packn * 4];
-                        tmpptr[5] = img0[l + packn * 5];
-                        tmpptr[6] = img0[l + packn * 6];
-                        tmpptr[7] = img0[l + packn * 7];
-                        tmpptr += 8;
-                    }
-
-                    img0 += size * packn;
-#else
-                    vfloat16m1_t _val0 = vle16_v_f16m1(img0, vl);
-                    vfloat16m1_t _val1 = vle16_v_f16m1(img0 + packn, vl);
-                    vfloat16m1_t _val2 = vle16_v_f16m1(img0 + packn * 2, vl);
-                    vfloat16m1_t _val3 = vle16_v_f16m1(img0 + packn * 3, vl);
-                    vfloat16m1_t _val4 = vle16_v_f16m1(img0 + packn * 4, vl);
-                    vfloat16m1_t _val5 = vle16_v_f16m1(img0 + packn * 5, vl);
-                    vfloat16m1_t _val6 = vle16_v_f16m1(img0 + packn * 6, vl);
-                    vfloat16m1_t _val7 = vle16_v_f16m1(img0 + packn * 7, vl);
-                    vsseg8e16_v_f16m1(tmpptr, _val0, _val1, _val2, _val3, _val4, _val5, _val6, _val7, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn * 8;
-#endif
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 3;
-        nn_size = (size - remain_size_start) >> 2;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 4;
-
-            __fp16* tmpptr = tmp.channel(i / 8 + (i % 8) / 4);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const __fp16* img0 = (const __fp16*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-#if C906
-                    for (int l = 0; l < packn; l++)
-                    {
-                        tmpptr[0] = img0[l];
-                        tmpptr[1] = img0[l + packn];
-                        tmpptr[2] = img0[l + packn * 2];
-                        tmpptr[3] = img0[l + packn * 3];
-                        tmpptr += 4;
-                    }
-
-                    img0 += size * packn;
-#else
-                    vfloat16m1_t _val0 = vle16_v_f16m1(img0, vl);
-                    vfloat16m1_t _val1 = vle16_v_f16m1(img0 + packn, vl);
-                    vfloat16m1_t _val2 = vle16_v_f16m1(img0 + packn * 2, vl);
-                    vfloat16m1_t _val3 = vle16_v_f16m1(img0 + packn * 3, vl);
-                    vsseg4e16_v_f16m1(tmpptr, _val0, _val1, _val2, _val3, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn * 4;
-#endif
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 2;
-        nn_size = (size - remain_size_start) >> 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int ii = 0; ii < nn_size; ii++)
-        {
-            int i = remain_size_start + ii * 2;
-
-            __fp16* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const __fp16* img0 = (const __fp16*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-#if C906
-                    for (int l = 0; l < packn; l++)
-                    {
-                        tmpptr[0] = img0[l];
-                        tmpptr[1] = img0[l + packn];
-                        tmpptr += 2;
-                    }
-
-                    img0 += size * packn;
-#else
-                    vfloat16m1_t _val0 = vle16_v_f16m1(img0, vl);
-                    vfloat16m1_t _val1 = vle16_v_f16m1(img0 + packn, vl);
-                    vsseg2e16_v_f16m1(tmpptr, _val0, _val1, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn * 2;
-#endif
-                }
-            }
-        }
-
-        remain_size_start += nn_size << 1;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int i = remain_size_start; i < size; i++)
-        {
-            __fp16* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2 + i % 2);
-
-            for (int q = 0; q < inch; q++)
-            {
-                const __fp16* img0 = (const __fp16*)bottom_im2col.channel(q) + i * packn;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    vfloat16m1_t _val = vle16_v_f16m1(img0, vl);
-                    vse16_v_f16m1(tmpptr, _val, vl);
-
-                    img0 += size * packn;
-                    tmpptr += packn;
-                }
-            }
-        }
-    }
-
-    int nn_outch = outch / packn;
-    int remain_outch_start = nn_outch * packn;
-
-    // make clang happy with the following loop
-#ifdef __clang__
-    __fp16* _zero_tmp = new __fp16[packn]();
-    for (int _zero_clean_idx = 0; _zero_clean_idx < packn; _zero_clean_idx++)
-    {
-        _zero_tmp[_zero_clean_idx] = 0.f;
-    }
-#endif // __clang__
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * packn;
-
-        __fp16* outptr0 = top_blob.channel(p);
-
-#ifdef __clang__
-        const __fp16* zeros = _zero_tmp;
-#else
-        const __fp16 zeros[packn] = {0.f};
-#endif // __clang__
-        const __fp16* biasptr = bias ? bias + p : zeros;
-
-        int i = 0;
-        for (; i + 7 < size; i += 8)
-        {
-            const __fp16* tmpptr = tmp.channel(i / 8);
-            const __fp16* kptr0 = kernel.channel(p / packn);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat16m1_t _sum0 = vle16_v_f16m1(biasptr, vl);
-            vfloat16m1_t _sum1 = vle16_v_f16m1(biasptr, vl);
-            vfloat16m1_t _sum2 = vle16_v_f16m1(biasptr, vl);
-            vfloat16m1_t _sum3 = vle16_v_f16m1(biasptr, vl);
-            vfloat16m1_t _sum4 = vle16_v_f16m1(biasptr, vl);
-            vfloat16m1_t _sum5 = vle16_v_f16m1(biasptr, vl);
-            vfloat16m1_t _sum6 = vle16_v_f16m1(biasptr, vl);
-            vfloat16m1_t _sum7 = vle16_v_f16m1(biasptr, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __fp16 val0 = *tmpptr++;
-                __fp16 val1 = *tmpptr++;
-                __fp16 val2 = *tmpptr++;
-                __fp16 val3 = *tmpptr++;
-                __fp16 val4 = *tmpptr++;
-                __fp16 val5 = *tmpptr++;
-                __fp16 val6 = *tmpptr++;
-                __fp16 val7 = *tmpptr++;
-                vfloat16m1_t _w0 = vle16_v_f16m1(kptr0, vl);
-                _sum0 = vfmacc_vf_f16m1(_sum0, val0, _w0, vl);
-                _sum1 = vfmacc_vf_f16m1(_sum1, val1, _w0, vl);
-                _sum2 = vfmacc_vf_f16m1(_sum2, val2, _w0, vl);
-                _sum3 = vfmacc_vf_f16m1(_sum3, val3, _w0, vl);
-                _sum4 = vfmacc_vf_f16m1(_sum4, val4, _w0, vl);
-                _sum5 = vfmacc_vf_f16m1(_sum5, val5, _w0, vl);
-                _sum6 = vfmacc_vf_f16m1(_sum6, val6, _w0, vl);
-                _sum7 = vfmacc_vf_f16m1(_sum7, val7, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-#if C906
-            vsse16_v_f16m1(outptr0, top_blob.cstep * sizeof(__fp16), _sum0, vl);
-            vsse16_v_f16m1(outptr0 + 1, top_blob.cstep * sizeof(__fp16), _sum1, vl);
-            vsse16_v_f16m1(outptr0 + 2, top_blob.cstep * sizeof(__fp16), _sum2, vl);
-            vsse16_v_f16m1(outptr0 + 3, top_blob.cstep * sizeof(__fp16), _sum3, vl);
-            vsse16_v_f16m1(outptr0 + 4, top_blob.cstep * sizeof(__fp16), _sum4, vl);
-            vsse16_v_f16m1(outptr0 + 5, top_blob.cstep * sizeof(__fp16), _sum5, vl);
-            vsse16_v_f16m1(outptr0 + 6, top_blob.cstep * sizeof(__fp16), _sum6, vl);
-            vsse16_v_f16m1(outptr0 + 7, top_blob.cstep * sizeof(__fp16), _sum7, vl);
-#else
-            vssseg8e16_v_f16m1(outptr0, top_blob.cstep * sizeof(__fp16), _sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7, vl);
-#endif
-            outptr0 += 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-            const __fp16* tmpptr = tmp.channel(i / 8 + (i % 8) / 4);
-            const __fp16* kptr0 = kernel.channel(p / packn);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat16m1_t _sum0 = vle16_v_f16m1(biasptr, vl);
-            vfloat16m1_t _sum1 = vle16_v_f16m1(biasptr, vl);
-            vfloat16m1_t _sum2 = vle16_v_f16m1(biasptr, vl);
-            vfloat16m1_t _sum3 = vle16_v_f16m1(biasptr, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __fp16 val0 = *tmpptr++;
-                __fp16 val1 = *tmpptr++;
-                __fp16 val2 = *tmpptr++;
-                __fp16 val3 = *tmpptr++;
-                vfloat16m1_t _w0 = vle16_v_f16m1(kptr0, vl);
-                _sum0 = vfmacc_vf_f16m1(_sum0, val0, _w0, vl);
-                _sum1 = vfmacc_vf_f16m1(_sum1, val1, _w0, vl);
-                _sum2 = vfmacc_vf_f16m1(_sum2, val2, _w0, vl);
-                _sum3 = vfmacc_vf_f16m1(_sum3, val3, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-#if C906
-            vsse16_v_f16m1(outptr0, top_blob.cstep * sizeof(__fp16), _sum0, vl);
-            vsse16_v_f16m1(outptr0 + 1, top_blob.cstep * sizeof(__fp16), _sum1, vl);
-            vsse16_v_f16m1(outptr0 + 2, top_blob.cstep * sizeof(__fp16), _sum2, vl);
-            vsse16_v_f16m1(outptr0 + 3, top_blob.cstep * sizeof(__fp16), _sum3, vl);
-#else
-            vssseg4e16_v_f16m1(outptr0, top_blob.cstep * sizeof(__fp16), _sum0, _sum1, _sum2, _sum3, vl);
-#endif
-            outptr0 += 4;
-        }
-        for (; i + 1 < size; i += 2)
-        {
-            const __fp16* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2);
-            const __fp16* kptr0 = kernel.channel(p / packn);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat16m1_t _sum0 = vle16_v_f16m1(biasptr, vl);
-            vfloat16m1_t _sum1 = vle16_v_f16m1(biasptr, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __fp16 val0 = *tmpptr++;
-                __fp16 val1 = *tmpptr++;
-                vfloat16m1_t _w0 = vle16_v_f16m1(kptr0, vl);
-                _sum0 = vfmacc_vf_f16m1(_sum0, val0, _w0, vl);
-                _sum1 = vfmacc_vf_f16m1(_sum1, val1, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-#if C906
-            vsse16_v_f16m1(outptr0, top_blob.cstep * sizeof(__fp16), _sum0, vl);
-            vsse16_v_f16m1(outptr0 + 1, top_blob.cstep * sizeof(__fp16), _sum1, vl);
-#else
-            vssseg2e16_v_f16m1(outptr0, top_blob.cstep * sizeof(__fp16), _sum0, _sum1, vl);
-#endif
-            outptr0 += 2;
-        }
-        for (; i < size; i++)
-        {
-            const __fp16* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2 + i % 2);
-            const __fp16* kptr0 = kernel.channel(p / packn);
-
-            int nn = inch * maxk * packn; // inch always > 0
-
-            vfloat16m1_t _sum = vle16_v_f16m1(biasptr, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                __fp16 val = *tmpptr++;
-                vfloat16m1_t _w0 = vle16_v_f16m1(kptr0, vl);
-                _sum = vfmacc_vf_f16m1(_sum, val, _w0, vl);
-
-                kptr0 += packn;
-            }
-
-            vsse16_v_f16m1(outptr0, top_blob.cstep * sizeof(__fp16), _sum, vl);
-
-            outptr0 += 1;
-        }
-    }
-#ifdef __clang__
-    delete[] _zero_tmp;
-#endif // __clang__
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        __fp16* outptr0 = top_blob.channel(p);
-
-        const __fp16 bias0 = bias ? bias[p] : 0.f;
-
-        int i = 0;
-        for (; i + 7 < size; i += 8)
-        {
-            const __fp16* tmpptr = tmp.channel(i / 8);
-            const __fp16* kptr0 = kernel.channel(p / packn + p % packn);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __fp16 sum0 = bias0;
-            __fp16 sum1 = bias0;
-            __fp16 sum2 = bias0;
-            __fp16 sum3 = bias0;
-            __fp16 sum4 = bias0;
-            __fp16 sum5 = bias0;
-            __fp16 sum6 = bias0;
-            __fp16 sum7 = bias0;
-
-            vfloat16m1_t _sum0 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum1 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum2 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum3 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum4 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum5 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum6 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum7 = vfmv_v_f_f16m1(0.f, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                vfloat16m1_t _val0;
-                vfloat16m1_t _val1;
-                vfloat16m1_t _val2;
-                vfloat16m1_t _val3;
-                vfloat16m1_t _val4;
-                vfloat16m1_t _val5;
-                vfloat16m1_t _val6;
-                vfloat16m1_t _val7;
-                vlseg8e16_v_f16m1(&_val0, &_val1, &_val2, &_val3, &_val4, &_val5, &_val6, &_val7, tmpptr, vl);
-                vfloat16m1_t _w0 = vle16_v_f16m1(kptr0, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _val0, _w0, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _val1, _w0, vl);
-                _sum2 = vfmacc_vv_f16m1(_sum2, _val2, _w0, vl);
-                _sum3 = vfmacc_vv_f16m1(_sum3, _val3, _w0, vl);
-                _sum4 = vfmacc_vv_f16m1(_sum4, _val4, _w0, vl);
-                _sum5 = vfmacc_vv_f16m1(_sum5, _val5, _w0, vl);
-                _sum6 = vfmacc_vv_f16m1(_sum6, _val6, _w0, vl);
-                _sum7 = vfmacc_vv_f16m1(_sum7, _val7, _w0, vl);
-                tmpptr += packn * 8;
-                kptr0 += packn;
-            }
-
-#if C906
-            // TODO
-            std::vector<__fp16> ss0(packn);
-            std::vector<__fp16> ss1(packn);
-            std::vector<__fp16> ss2(packn);
-            std::vector<__fp16> ss3(packn);
-            std::vector<__fp16> ss4(packn);
-            std::vector<__fp16> ss5(packn);
-            std::vector<__fp16> ss6(packn);
-            std::vector<__fp16> ss7(packn);
-            vse16_v_f16m1((__fp16*)ss0.data(), _sum0, vl);
-            vse16_v_f16m1((__fp16*)ss1.data(), _sum1, vl);
-            vse16_v_f16m1((__fp16*)ss2.data(), _sum2, vl);
-            vse16_v_f16m1((__fp16*)ss3.data(), _sum3, vl);
-            vse16_v_f16m1((__fp16*)ss4.data(), _sum4, vl);
-            vse16_v_f16m1((__fp16*)ss5.data(), _sum5, vl);
-            vse16_v_f16m1((__fp16*)ss6.data(), _sum6, vl);
-            vse16_v_f16m1((__fp16*)ss7.data(), _sum7, vl);
-            for (int i = 0; i < packn; i++)
-            {
-                sum0 += ss0[i];
-                sum1 += ss1[i];
-                sum2 += ss2[i];
-                sum3 += ss3[i];
-                sum4 += ss4[i];
-                sum5 += ss5[i];
-                sum6 += ss6[i];
-                sum7 += ss7[i];
-            }
-#else
-            sum0 = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum0, vfmv_s_f_f16m1(vfloat16m1_t(), sum0, vl), vl));
-            sum1 = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum1, vfmv_s_f_f16m1(vfloat16m1_t(), sum1, vl), vl));
-            sum2 = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum2, vfmv_s_f_f16m1(vfloat16m1_t(), sum2, vl), vl));
-            sum3 = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum3, vfmv_s_f_f16m1(vfloat16m1_t(), sum3, vl), vl));
-            sum4 = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum4, vfmv_s_f_f16m1(vfloat16m1_t(), sum4, vl), vl));
-            sum5 = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum5, vfmv_s_f_f16m1(vfloat16m1_t(), sum5, vl), vl));
-            sum6 = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum6, vfmv_s_f_f16m1(vfloat16m1_t(), sum6, vl), vl));
-            sum7 = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum7, vfmv_s_f_f16m1(vfloat16m1_t(), sum7, vl), vl));
-#endif
-
-            outptr0[0] = sum0;
-            outptr0[1] = sum1;
-            outptr0[2] = sum2;
-            outptr0[3] = sum3;
-            outptr0[4] = sum4;
-            outptr0[5] = sum5;
-            outptr0[6] = sum6;
-            outptr0[7] = sum7;
-
-            outptr0 += 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-            const __fp16* tmpptr = tmp.channel(i / 8 + (i % 8) / 4);
-            const __fp16* kptr0 = kernel.channel(p / packn + p % packn);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __fp16 sum0 = bias0;
-            __fp16 sum1 = bias0;
-            __fp16 sum2 = bias0;
-            __fp16 sum3 = bias0;
-
-            vfloat16m1_t _sum0 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum1 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum2 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum3 = vfmv_v_f_f16m1(0.f, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                vfloat16m1_t _val0;
-                vfloat16m1_t _val1;
-                vfloat16m1_t _val2;
-                vfloat16m1_t _val3;
-
-                vlseg4e16_v_f16m1(&_val0, &_val1, &_val2, &_val3, tmpptr, vl);
-                vfloat16m1_t _w0 = vle16_v_f16m1(kptr0, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _val0, _w0, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _val1, _w0, vl);
-                _sum2 = vfmacc_vv_f16m1(_sum2, _val2, _w0, vl);
-                _sum3 = vfmacc_vv_f16m1(_sum3, _val3, _w0, vl);
-                tmpptr += packn * 4;
-                kptr0 += packn;
-            }
-
-#if C906
-            // TODO
-            std::vector<__fp16> ss0(packn);
-            std::vector<__fp16> ss1(packn);
-            std::vector<__fp16> ss2(packn);
-            std::vector<__fp16> ss3(packn);
-            vse16_v_f16m1((__fp16*)ss0.data(), _sum0, vl);
-            vse16_v_f16m1((__fp16*)ss1.data(), _sum1, vl);
-            vse16_v_f16m1((__fp16*)ss2.data(), _sum2, vl);
-            vse16_v_f16m1((__fp16*)ss3.data(), _sum3, vl);
-            for (int i = 0; i < packn; i++)
-            {
-                sum0 += ss0[i];
-                sum1 += ss1[i];
-                sum2 += ss2[i];
-                sum3 += ss3[i];
-            }
-#else
-            sum0 = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum0, vfmv_s_f_f16m1(vfloat16m1_t(), sum0, vl), vl));
-            sum1 = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum1, vfmv_s_f_f16m1(vfloat16m1_t(), sum1, vl), vl));
-            sum2 = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum2, vfmv_s_f_f16m1(vfloat16m1_t(), sum2, vl), vl));
-            sum3 = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum3, vfmv_s_f_f16m1(vfloat16m1_t(), sum3, vl), vl));
-#endif
-
-            outptr0[0] = sum0;
-            outptr0[1] = sum1;
-            outptr0[2] = sum2;
-            outptr0[3] = sum3;
-
-            outptr0 += 4;
-        }
-        for (; i + 1 < size; i += 2)
-        {
-            const __fp16* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2);
-            const __fp16* kptr0 = kernel.channel(p / packn + p % packn);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __fp16 sum0 = bias0;
-            __fp16 sum1 = bias0;
-
-            vfloat16m1_t _sum0 = vfmv_v_f_f16m1(0.f, vl);
-            vfloat16m1_t _sum1 = vfmv_v_f_f16m1(0.f, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                vfloat16m1_t _val0;
-                vfloat16m1_t _val1;
-                vlseg2e16_v_f16m1(&_val0, &_val1, tmpptr, vl);
-                vfloat16m1_t _w0 = vle16_v_f16m1(kptr0, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _val0, _w0, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _val1, _w0, vl);
-                tmpptr += packn * 2;
-                kptr0 += packn;
-            }
-
-#if C906
-            // TODO
-            std::vector<__fp16> ss0(packn);
-            std::vector<__fp16> ss1(packn);
-            vse16_v_f16m1((__fp16*)ss0.data(), _sum0, vl);
-            vse16_v_f16m1((__fp16*)ss1.data(), _sum1, vl);
-            for (int i = 0; i < packn; i++)
-            {
-                sum0 += ss0[i];
-                sum1 += ss1[i];
-            }
-#else
-            sum0 = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum0, vfmv_s_f_f16m1(vfloat16m1_t(), sum0, vl), vl));
-            sum1 = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum1, vfmv_s_f_f16m1(vfloat16m1_t(), sum1, vl), vl));
-#endif
-
-            outptr0[0] = sum0;
-            outptr0[1] = sum1;
-
-            outptr0 += 2;
-        }
-        for (; i < size; i++)
-        {
-            const __fp16* tmpptr = tmp.channel(i / 8 + (i % 8) / 4 + (i % 4) / 2 + i % 2);
-            const __fp16* kptr0 = kernel.channel(p / packn + p % packn);
-
-            int nn = inch * maxk; // inch always > 0
-
-            __fp16 sum0 = bias0;
-
-            vfloat16m1_t _sum0 = vfmv_v_f_f16m1(0.f, vl);
-
-            for (int j = 0; j < nn; j++)
-            {
-                vfloat16m1_t _val0 = vle16_v_f16m1(tmpptr, vl);
-                vfloat16m1_t _w0 = vle16_v_f16m1(kptr0, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _val0, _w0, vl);
-                tmpptr += packn;
-                kptr0 += packn;
-            }
-
-#if C906
-            // TODO
-            std::vector<__fp16> ss0(packn);
-            vse16_v_f16m1((__fp16*)ss0.data(), _sum0, vl);
-            for (int i = 0; i < packn; i++)
-            {
-                sum0 += ss0[i];
-            }
-#else
-            sum0 = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum0, vfmv_s_f_f16m1(vfloat16m1_t(), sum0, vl), vl));
-#endif
-
-            outptr0[0] = sum0;
-
-            outptr0 += 1;
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_transform_kernel_packnto1_fp16sa_rvv(const Mat& _kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int packn = csrr_vlenb() / 2;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // interleave
-    // src = maxk-inch-outch
-    // dst = pb-pa-maxk-inch/pa-outch/pb
-    Mat kernel = _kernel.reshape(maxk, inch, outch);
-    kernel_tm.create(packn * packn * maxk, inch / packn, outch / packn + outch % packn, (size_t)2u);
-
-    int q = 0;
-    for (; q + (packn - 1) < outch; q += packn)
-    {
-        __fp16* g00 = kernel_tm.channel(q / packn);
-
-        for (int p = 0; p + (packn - 1) < inch; p += packn)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int i = 0; i < packn; i++)
-                {
-                    for (int j = 0; j < packn; j++)
-                    {
-                        const float* k00 = kernel.channel(q + j).row(p + i);
-
-                        g00[0] = (__fp16)k00[k];
-
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-    for (; q < outch; q++)
-    {
-        const Mat k0 = kernel.channel(q);
-
-        __fp16* g00 = kernel_tm.channel(q / packn + q % packn);
-
-        for (int p = 0; p + (packn - 1) < inch; p += packn)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                for (int j = 0; j < packn; j++)
-                {
-                    const float* k00 = k0.row(p + j);
-
-                    g00[0] = (__fp16)k00[k];
-
-                    g00++;
-                }
-            }
-        }
-    }
-}
-
-static void convolution_im2col_sgemm_packnto1_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    const int size = outw * outh;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // im2col
-    Mat bottom_im2col(size, maxk, inch, 2u * packn, packn, opt.workspace_allocator);
-    {
-        const int gap = (w * stride_h - outw * stride_w) * packn;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < inch; p++)
-        {
-            const Mat img = bottom_blob.channel(p);
-            __fp16* ptr = bottom_im2col.channel(p);
-
-            for (int u = 0; u < kernel_h; u++)
-            {
-                for (int v = 0; v < kernel_w; v++)
-                {
-                    const __fp16* sptr = img.row<const __fp16>(dilation_h * u) + dilation_w * v * packn;
-
-                    for (int i = 0; i < outh; i++)
-                    {
-                        int j = 0;
-                        for (; j < outw; j++)
-                        {
-                            vfloat16m1_t _val = vle16_v_f16m1(sptr, vl);
-                            vse16_v_f16m1(ptr, _val, vl);
-
-                            sptr += stride_w * packn;
-                            ptr += packn;
-                        }
-
-                        sptr += gap;
-                    }
-                }
-            }
-        }
-    }
-
-    im2col_sgemm_packnto1_fp16sa_rvv(bottom_im2col, top_blob, kernel, _bias, opt);
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_dot.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_dot.h
deleted file mode 100644
index c0a7b768..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_dot.h
+++ /dev/null
@@ -1,518 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_winograd_dot_rvv(Mat& bottom_blob_tm, int outch, const Mat& kernel_tm, Mat& top_blob_tm, const Option& opt)
-{
-#if __riscv_vector
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-#endif
-
-    // Mat bottom_blob_tm(tiles, 16/36/64, inch, 4u, opt.workspace_allocator);
-
-    const int tiles = bottom_blob_tm.w;
-    const int batch = bottom_blob_tm.h;
-    const int inch = bottom_blob_tm.c;
-
-    // permute
-    Mat bottom_blob_tm2;
-#if __riscv_vector
-    if (tiles >= packn)
-        bottom_blob_tm2.create(packn * inch, tiles / packn + tiles % packn, batch, 4u, opt.workspace_allocator);
-#else
-    if (tiles >= 4)
-        bottom_blob_tm2.create(4 * inch, tiles / 4 + tiles % 4, batch, 4u, opt.workspace_allocator);
-#endif
-    else
-        bottom_blob_tm2.create(1 * inch, tiles, batch, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int r = 0; r < batch; r++)
-    {
-        Mat tm2 = bottom_blob_tm2.channel(r);
-
-        // tile
-        int i = 0;
-#if __riscv_vector
-        for (; i + (packn - 1) < tiles; i += packn)
-        {
-            float* tmpptr = tm2.row(i / packn);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i);
-
-            for (int q = 0; q < inch; q++)
-            {
-                vse32_v_f32m1(tmpptr, vle32_v_f32m1(r0, vl), vl);
-                r0 += bottom_blob_tm.cstep;
-                tmpptr += packn;
-            }
-        }
-#else  // __riscv_vector
-        for (; i + 3 < tiles; i += 4)
-        {
-            float* tmpptr = tm2.row(i / 4);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i);
-
-            for (int q = 0; q < inch; q++)
-            {
-                tmpptr[0] = r0[0];
-                tmpptr[1] = r0[1];
-                tmpptr[2] = r0[2];
-                tmpptr[3] = r0[3];
-
-                r0 += bottom_blob_tm.cstep;
-                tmpptr += 4;
-            }
-        }
-#endif // __riscv_vector
-        for (; i < tiles; i++)
-        {
-#if __riscv_vector
-            float* tmpptr = tm2.row(i / packn + i % packn);
-#else
-            float* tmpptr = tm2.row(i / 4 + i % 4);
-#endif
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i);
-
-            for (int q = 0; q < inch; q++)
-            {
-                tmpptr[0] = r0[0];
-
-                r0 += bottom_blob_tm.cstep;
-                tmpptr += 1;
-            }
-        }
-    }
-
-    bottom_blob_tm = Mat();
-    // permute end
-
-    top_blob_tm.create(tiles, batch, outch, 4u, opt.workspace_allocator);
-
-#if __riscv_vector
-    int nn_outch = outch >> 3;
-    int remain_outch_start = nn_outch << 3;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 8;
-
-        float* output0_tm = top_blob_tm.channel(p);
-        float* output1_tm = top_blob_tm.channel(p + 1);
-        float* output2_tm = top_blob_tm.channel(p + 2);
-        float* output3_tm = top_blob_tm.channel(p + 3);
-        float* output4_tm = top_blob_tm.channel(p + 4);
-        float* output5_tm = top_blob_tm.channel(p + 5);
-        float* output6_tm = top_blob_tm.channel(p + 6);
-        float* output7_tm = top_blob_tm.channel(p + 7);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 8);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + (packn - 1) < tiles; i += packn)
-            {
-                const float* r0 = bb2.row(i / packn);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                vfloat32m1_t _sum0 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum1 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum2 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum3 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum4 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum5 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum6 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum7 = vfmv_v_f_f32m1(0.f, vl);
-
-                int j = 0;
-                for (; j < nn; j++)
-                {
-                    vfloat32m1_t _val = vle32_v_f32m1(r0, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, k0[0], _val, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, k0[1], _val, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, k0[2], _val, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, k0[3], _val, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, k0[4], _val, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, k0[5], _val, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, k0[6], _val, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, k0[7], _val, vl);
-                    r0 += packn;
-                    k0 += 8;
-                }
-
-                vse32_v_f32m1(output0_tm, _sum0, vl);
-                vse32_v_f32m1(output1_tm, _sum1, vl);
-                vse32_v_f32m1(output2_tm, _sum2, vl);
-                vse32_v_f32m1(output3_tm, _sum3, vl);
-                vse32_v_f32m1(output4_tm, _sum4, vl);
-                vse32_v_f32m1(output5_tm, _sum5, vl);
-                vse32_v_f32m1(output6_tm, _sum6, vl);
-                vse32_v_f32m1(output7_tm, _sum7, vl);
-
-                output0_tm += packn;
-                output1_tm += packn;
-                output2_tm += packn;
-                output3_tm += packn;
-                output4_tm += packn;
-                output5_tm += packn;
-                output6_tm += packn;
-                output7_tm += packn;
-            }
-            for (; i < tiles; i++)
-            {
-                const float* r0 = bb2.row(i / packn + i % packn);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-                float sum2 = 0.f;
-                float sum3 = 0.f;
-                float sum4 = 0.f;
-                float sum5 = 0.f;
-                float sum6 = 0.f;
-                float sum7 = 0.f;
-
-                int j = 0;
-                for (; j < nn; j++)
-                {
-                    sum0 += r0[0] * k0[0];
-                    sum1 += r0[0] * k0[1];
-                    sum2 += r0[0] * k0[2];
-                    sum3 += r0[0] * k0[3];
-                    sum4 += r0[0] * k0[4];
-                    sum5 += r0[0] * k0[5];
-                    sum6 += r0[0] * k0[6];
-                    sum7 += r0[0] * k0[7];
-
-                    r0 += 1;
-                    k0 += 8;
-                }
-
-                output0_tm[0] = sum0;
-                output1_tm[0] = sum1;
-                output2_tm[0] = sum2;
-                output3_tm[0] = sum3;
-                output4_tm[0] = sum4;
-                output5_tm[0] = sum5;
-                output6_tm[0] = sum6;
-                output7_tm[0] = sum7;
-
-                output0_tm++;
-                output1_tm++;
-                output2_tm++;
-                output3_tm++;
-                output4_tm++;
-                output5_tm++;
-                output6_tm++;
-                output7_tm++;
-            }
-        }
-    }
-
-    nn_outch = (outch - remain_outch_start) >> 2;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = remain_outch_start + pp * 4;
-
-        float* output0_tm = top_blob_tm.channel(p);
-        float* output1_tm = top_blob_tm.channel(p + 1);
-        float* output2_tm = top_blob_tm.channel(p + 2);
-        float* output3_tm = top_blob_tm.channel(p + 3);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 8 + (p % 8) / 4);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + (packn - 1) < tiles; i += packn)
-            {
-                const float* r0 = bb2.row(i / packn);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                vfloat32m1_t _sum0 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum1 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum2 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum3 = vfmv_v_f_f32m1(0.f, vl);
-
-                int j = 0;
-                for (; j < nn; j++)
-                {
-                    vfloat32m1_t _val = vle32_v_f32m1(r0, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, k0[0], _val, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, k0[1], _val, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, k0[2], _val, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, k0[3], _val, vl);
-                    r0 += packn;
-                    k0 += 4;
-                }
-
-                vse32_v_f32m1(output0_tm, _sum0, vl);
-                vse32_v_f32m1(output1_tm, _sum1, vl);
-                vse32_v_f32m1(output2_tm, _sum2, vl);
-                vse32_v_f32m1(output3_tm, _sum3, vl);
-
-                output0_tm += packn;
-                output1_tm += packn;
-                output2_tm += packn;
-                output3_tm += packn;
-            }
-            for (; i < tiles; i++)
-            {
-                const float* r0 = bb2.row(i / packn + i % packn);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-                float sum2 = 0.f;
-                float sum3 = 0.f;
-
-                int j = 0;
-                for (; j < nn; j++)
-                {
-                    sum0 += r0[0] * k0[0];
-                    sum1 += r0[0] * k0[1];
-                    sum2 += r0[0] * k0[2];
-                    sum3 += r0[0] * k0[3];
-
-                    r0 += 1;
-                    k0 += 4;
-                }
-
-                output0_tm[0] = sum0;
-                output1_tm[0] = sum1;
-                output2_tm[0] = sum2;
-                output3_tm[0] = sum3;
-
-                output0_tm++;
-                output1_tm++;
-                output2_tm++;
-                output3_tm++;
-            }
-        }
-    }
-
-    remain_outch_start += nn_outch << 2;
-#else
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        float* output0_tm = top_blob_tm.channel(p);
-        float* output1_tm = top_blob_tm.channel(p + 1);
-
-        const Mat kernel0_tm = kernel_tm.channel(p / 2);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 3 < tiles; i += 4)
-            {
-                const float* r0 = bb2.row(i / 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum00 = 0.f;
-                float sum01 = 0.f;
-                float sum02 = 0.f;
-                float sum03 = 0.f;
-                float sum10 = 0.f;
-                float sum11 = 0.f;
-                float sum12 = 0.f;
-                float sum13 = 0.f;
-
-                for (int j = 0; j < nn; j++)
-                {
-                    float w0 = k0[0];
-                    float w1 = k0[1];
-                    sum00 += r0[0] * w0;
-                    sum01 += r0[1] * w0;
-                    sum02 += r0[2] * w0;
-                    sum03 += r0[3] * w0;
-                    sum10 += r0[0] * w1;
-                    sum11 += r0[1] * w1;
-                    sum12 += r0[2] * w1;
-                    sum13 += r0[3] * w1;
-
-                    r0 += 4;
-                    k0 += 2;
-                }
-
-                output0_tm[0] = sum00;
-                output0_tm[1] = sum01;
-                output0_tm[2] = sum02;
-                output0_tm[3] = sum03;
-                output1_tm[0] = sum10;
-                output1_tm[1] = sum11;
-                output1_tm[2] = sum12;
-                output1_tm[3] = sum13;
-
-                output0_tm += 4;
-                output1_tm += 4;
-            }
-            for (; i < tiles; i++)
-            {
-                const float* r0 = bb2.row(i / 4 + i % 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum00 = 0.f;
-                float sum10 = 0.f;
-
-                for (int j = 0; j < nn; j++)
-                {
-                    float val0 = r0[0];
-                    sum00 += val0 * k0[0];
-                    sum10 += val0 * k0[1];
-
-                    r0 += 1;
-                    k0 += 2;
-                }
-
-                output0_tm[0] = sum00;
-                output1_tm[0] = sum10;
-                output0_tm++;
-                output1_tm++;
-            }
-        }
-    }
-#endif
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        float* output0_tm = top_blob_tm.channel(p);
-
-#if __riscv_vector
-        const Mat kernel0_tm = kernel_tm.channel(p / 8 + (p % 8) / 4 + p % 4);
-#else
-        const Mat kernel0_tm = kernel_tm.channel(p / 2 + p % 2);
-#endif
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-#if __riscv_vector
-            for (; i + (packn - 1) < tiles; i += packn)
-            {
-                const float* r0 = bb2.row(i / packn);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                vfloat32m1_t _sum0 = vfmv_v_f_f32m1(0.f, vl);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    _sum0 = vfmacc_vf_f32m1(_sum0, k0[0], vle32_v_f32m1(r0, vl), vl);
-                    r0 += packn;
-                    k0++;
-                }
-
-                vse32_v_f32m1(output0_tm, _sum0, vl);
-                output0_tm += packn;
-            }
-#else  // __riscv_vector
-            for (; i + 3 < tiles; i += 4)
-            {
-                const float* r0 = bb2.row(i / 4);
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-                float sum2 = 0.f;
-                float sum3 = 0.f;
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __builtin_prefetch(r0 + 16);
-                    __builtin_prefetch(k0 + 4);
-                    float w0 = k0[0];
-                    sum0 += r0[0] * w0;
-                    sum1 += r0[1] * w0;
-                    sum2 += r0[2] * w0;
-                    sum3 += r0[3] * w0;
-
-                    r0 += 4;
-                    k0++;
-                }
-
-                output0_tm[0] = sum0;
-                output0_tm[1] = sum1;
-                output0_tm[2] = sum2;
-                output0_tm[3] = sum3;
-                output0_tm += 4;
-            }
-#endif // __riscv_vector
-            for (; i < tiles; i++)
-            {
-#if __riscv_vector
-                const float* r0 = bb2.row(i / packn + i % packn);
-#else
-                const float* r0 = bb2.row(i / 4 + i % 4);
-#endif
-                const float* k0 = kernel0_tm.row(r);
-
-                int nn = inch; // inch always > 0
-
-                float sum = 0.f;
-
-                for (int j = 0; j < nn; j++)
-                {
-                    sum += r0[0] * k0[0];
-                    r0 += 1;
-                    k0 += 1;
-                }
-
-                output0_tm[0] = sum;
-                output0_tm += 1;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_dot_packn.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_dot_packn.h
deleted file mode 100644
index 1c505d5c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_dot_packn.h
+++ /dev/null
@@ -1,317 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_winograd_dot_packn_rvv(Mat& bottom_blob_tm, int outch, const Mat& kernel_tm, Mat& top_blob_tm, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    // Mat bottom_blob_tm(tiles, 16/36/64, inch, 4u * packn, packn, opt.workspace_allocator);
-
-    const int tiles = bottom_blob_tm.w;
-    const int batch = bottom_blob_tm.h;
-    const int inch = bottom_blob_tm.c;
-
-    // permute
-    Mat bottom_blob_tm2;
-    if (tiles >= 8)
-        bottom_blob_tm2.create(8 * inch, tiles / 8 + (tiles % 8) / 4 + (tiles % 4) / 2 + tiles % 2, batch, 4u * packn, packn, opt.workspace_allocator);
-    else if (tiles >= 4)
-        bottom_blob_tm2.create(4 * inch, tiles / 4 + (tiles % 4) / 2 + tiles % 2, batch, 4u * packn, packn, opt.workspace_allocator);
-    else if (tiles >= 2)
-        bottom_blob_tm2.create(2 * inch, tiles / 2 + tiles % 2, batch, 4u * packn, packn, opt.workspace_allocator);
-    else // if (tiles >= 1)
-        bottom_blob_tm2.create(1 * inch, tiles, batch, 4u * packn, packn, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int r = 0; r < batch; r++)
-    {
-        Mat tm2 = bottom_blob_tm2.channel(r);
-
-        // tile
-        int i = 0;
-        for (; i + 7 < tiles; i += 8)
-        {
-            float* tmpptr = tm2.row<float>(i / 8);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * packn;
-
-            for (int q = 0; q < inch; q++)
-            {
-#if C906
-                for (int l = 0; l < packn; l++)
-                {
-                    tmpptr[0] = r0[l];
-                    tmpptr[1] = r0[l + packn];
-                    tmpptr[2] = r0[l + packn * 2];
-                    tmpptr[3] = r0[l + packn * 3];
-                    tmpptr[4] = r0[l + packn * 4];
-                    tmpptr[5] = r0[l + packn * 5];
-                    tmpptr[6] = r0[l + packn * 6];
-                    tmpptr[7] = r0[l + packn * 7];
-                    tmpptr += 8;
-                }
-
-                r0 += bottom_blob_tm.cstep * packn;
-#else
-                vfloat32m1_t _val0 = vle32_v_f32m1(r0, vl);
-                vfloat32m1_t _val1 = vle32_v_f32m1(r0 + packn, vl);
-                vfloat32m1_t _val2 = vle32_v_f32m1(r0 + packn * 2, vl);
-                vfloat32m1_t _val3 = vle32_v_f32m1(r0 + packn * 3, vl);
-                vfloat32m1_t _val4 = vle32_v_f32m1(r0 + packn * 4, vl);
-                vfloat32m1_t _val5 = vle32_v_f32m1(r0 + packn * 5, vl);
-                vfloat32m1_t _val6 = vle32_v_f32m1(r0 + packn * 6, vl);
-                vfloat32m1_t _val7 = vle32_v_f32m1(r0 + packn * 7, vl);
-                vsseg8e32_v_f32m1(tmpptr, _val0, _val1, _val2, _val3, _val4, _val5, _val6, _val7, vl);
-
-                r0 += bottom_blob_tm.cstep * packn;
-                tmpptr += packn * 8;
-#endif
-            }
-        }
-        for (; i + 3 < tiles; i += 4)
-        {
-            float* tmpptr = tm2.row<float>(i / 8 + (i % 8) / 4);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * packn;
-
-            for (int q = 0; q < inch; q++)
-            {
-#if C906
-                for (int l = 0; l < packn; l++)
-                {
-                    tmpptr[0] = r0[l];
-                    tmpptr[1] = r0[l + packn];
-                    tmpptr[2] = r0[l + packn * 2];
-                    tmpptr[3] = r0[l + packn * 3];
-                    tmpptr += 4;
-                }
-
-                r0 += bottom_blob_tm.cstep * packn;
-#else
-                vfloat32m1_t _val0 = vle32_v_f32m1(r0, vl);
-                vfloat32m1_t _val1 = vle32_v_f32m1(r0 + packn, vl);
-                vfloat32m1_t _val2 = vle32_v_f32m1(r0 + packn * 2, vl);
-                vfloat32m1_t _val3 = vle32_v_f32m1(r0 + packn * 3, vl);
-                vsseg4e32_v_f32m1(tmpptr, _val0, _val1, _val2, _val3, vl);
-
-                r0 += bottom_blob_tm.cstep * packn;
-                tmpptr += packn * 4;
-#endif
-            }
-        }
-        for (; i + 1 < tiles; i += 2)
-        {
-            float* tmpptr = tm2.row<float>(i / 8 + (i % 8) / 4 + (i % 4) / 2);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * packn;
-
-            for (int q = 0; q < inch; q++)
-            {
-#if C906
-                for (int l = 0; l < packn; l++)
-                {
-                    tmpptr[0] = r0[l];
-                    tmpptr[1] = r0[l + packn];
-                    tmpptr += 2;
-                }
-
-                r0 += bottom_blob_tm.cstep * packn;
-#else
-                vfloat32m1_t _val0 = vle32_v_f32m1(r0, vl);
-                vfloat32m1_t _val1 = vle32_v_f32m1(r0 + packn, vl);
-                vsseg2e32_v_f32m1(tmpptr, _val0, _val1, vl);
-
-                r0 += bottom_blob_tm.cstep * packn;
-                tmpptr += packn * 2;
-#endif
-            }
-        }
-        for (; i < tiles; i++)
-        {
-            float* tmpptr = tm2.row<float>(i / 8 + (i % 8) / 4 + (i % 4) / 2 + i % 2);
-
-            const float* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * packn;
-
-            for (int q = 0; q < inch; q++)
-            {
-                vfloat32m1_t _val = vle32_v_f32m1(r0, vl);
-                vse32_v_f32m1(tmpptr, _val, vl);
-
-                r0 += bottom_blob_tm.cstep * packn;
-                tmpptr += packn;
-            }
-        }
-    }
-
-    bottom_blob_tm = Mat();
-    // permute end
-
-    top_blob_tm.create(tiles, batch, outch, 4u * packn, packn, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* output0_tm = top_blob_tm.channel(p);
-
-        const Mat kernel0_tm = kernel_tm.channel(p);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 7 < tiles; i += 8)
-            {
-                const float* r0 = bb2.row<const float>(i / 8);
-                const float* k0 = kernel0_tm.row<const float>(r);
-
-                int nn = inch * packn; // inch always > 0
-
-                vfloat32m1_t _sum0 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum1 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum2 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum3 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum4 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum5 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum6 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum7 = vfmv_v_f_f32m1(0.f, vl);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    float val0 = *r0++;
-                    float val1 = *r0++;
-                    float val2 = *r0++;
-                    float val3 = *r0++;
-                    float val4 = *r0++;
-                    float val5 = *r0++;
-                    float val6 = *r0++;
-                    float val7 = *r0++;
-                    vfloat32m1_t _w0 = vle32_v_f32m1(k0, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, val0, _w0, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, val1, _w0, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, val2, _w0, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, val3, _w0, vl);
-                    _sum4 = vfmacc_vf_f32m1(_sum4, val4, _w0, vl);
-                    _sum5 = vfmacc_vf_f32m1(_sum5, val5, _w0, vl);
-                    _sum6 = vfmacc_vf_f32m1(_sum6, val6, _w0, vl);
-                    _sum7 = vfmacc_vf_f32m1(_sum7, val7, _w0, vl);
-
-                    k0 += packn;
-                }
-
-                vse32_v_f32m1(output0_tm, _sum0, vl);
-                vse32_v_f32m1(output0_tm + packn, _sum1, vl);
-                vse32_v_f32m1(output0_tm + packn * 2, _sum2, vl);
-                vse32_v_f32m1(output0_tm + packn * 3, _sum3, vl);
-                vse32_v_f32m1(output0_tm + packn * 4, _sum4, vl);
-                vse32_v_f32m1(output0_tm + packn * 5, _sum5, vl);
-                vse32_v_f32m1(output0_tm + packn * 6, _sum6, vl);
-                vse32_v_f32m1(output0_tm + packn * 7, _sum7, vl);
-
-                output0_tm += packn * 8;
-            }
-            for (; i + 3 < tiles; i += 4)
-            {
-                const float* r0 = bb2.row<const float>(i / 8 + (i % 8) / 4);
-                const float* k0 = kernel0_tm.row<const float>(r);
-
-                int nn = inch * packn; // inch always > 0
-
-                vfloat32m1_t _sum0 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum1 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum2 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum3 = vfmv_v_f_f32m1(0.f, vl);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    float val0 = *r0++;
-                    float val1 = *r0++;
-                    float val2 = *r0++;
-                    float val3 = *r0++;
-                    vfloat32m1_t _w0 = vle32_v_f32m1(k0, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, val0, _w0, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, val1, _w0, vl);
-                    _sum2 = vfmacc_vf_f32m1(_sum2, val2, _w0, vl);
-                    _sum3 = vfmacc_vf_f32m1(_sum3, val3, _w0, vl);
-
-                    k0 += packn;
-                }
-
-                vse32_v_f32m1(output0_tm, _sum0, vl);
-                vse32_v_f32m1(output0_tm + packn, _sum1, vl);
-                vse32_v_f32m1(output0_tm + packn * 2, _sum2, vl);
-                vse32_v_f32m1(output0_tm + packn * 3, _sum3, vl);
-
-                output0_tm += packn * 4;
-            }
-            for (; i + 1 < tiles; i += 2)
-            {
-                const float* r0 = bb2.row<const float>(i / 8 + (i % 8) / 4 + (i % 4) / 2);
-                const float* k0 = kernel0_tm.row<const float>(r);
-
-                int nn = inch * packn; // inch always > 0
-
-                vfloat32m1_t _sum0 = vfmv_v_f_f32m1(0.f, vl);
-                vfloat32m1_t _sum1 = vfmv_v_f_f32m1(0.f, vl);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    float val0 = *r0++;
-                    float val1 = *r0++;
-                    vfloat32m1_t _w0 = vle32_v_f32m1(k0, vl);
-                    _sum0 = vfmacc_vf_f32m1(_sum0, val0, _w0, vl);
-                    _sum1 = vfmacc_vf_f32m1(_sum1, val1, _w0, vl);
-
-                    k0 += packn;
-                }
-
-                vse32_v_f32m1(output0_tm, _sum0, vl);
-                vse32_v_f32m1(output0_tm + packn, _sum1, vl);
-
-                output0_tm += packn * 2;
-            }
-            for (; i < tiles; i++)
-            {
-                const float* r0 = bb2.row<const float>(i / 8 + (i % 8) / 4 + (i % 4) / 2 + i % 2);
-                const float* k0 = kernel0_tm.row<const float>(r);
-
-                int nn = inch * packn; // inch always > 0
-
-                vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    float val = *r0++;
-                    vfloat32m1_t _w0 = vle32_v_f32m1(k0, vl);
-                    _sum = vfmacc_vf_f32m1(_sum, val, _w0, vl);
-
-                    k0 += packn;
-                }
-
-                vse32_v_f32m1(output0_tm, _sum, vl);
-
-                output0_tm += packn;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_dot_packn_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_dot_packn_fp16s.h
deleted file mode 100644
index ed35ad3e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_dot_packn_fp16s.h
+++ /dev/null
@@ -1,317 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_winograd_dot_packn_fp16sa_rvv(Mat& bottom_blob_tm, int outch, const Mat& kernel_tm, Mat& top_blob_tm, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    // Mat bottom_blob_tm(tiles, 16/36/64, inch, 2u * packn, packn, opt.workspace_allocator);
-
-    const int tiles = bottom_blob_tm.w;
-    const int batch = bottom_blob_tm.h;
-    const int inch = bottom_blob_tm.c;
-
-    // permute
-    Mat bottom_blob_tm2;
-    if (tiles >= 8)
-        bottom_blob_tm2.create(8 * inch, tiles / 8 + (tiles % 8) / 4 + (tiles % 4) / 2 + tiles % 2, batch, 2u * packn, packn, opt.workspace_allocator);
-    else if (tiles >= 4)
-        bottom_blob_tm2.create(4 * inch, tiles / 4 + (tiles % 4) / 2 + tiles % 2, batch, 2u * packn, packn, opt.workspace_allocator);
-    else if (tiles >= 2)
-        bottom_blob_tm2.create(2 * inch, tiles / 2 + tiles % 2, batch, 2u * packn, packn, opt.workspace_allocator);
-    else // if (tiles >= 1)
-        bottom_blob_tm2.create(1 * inch, tiles, batch, 2u * packn, packn, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int r = 0; r < batch; r++)
-    {
-        Mat tm2 = bottom_blob_tm2.channel(r);
-
-        // tile
-        int i = 0;
-        for (; i + 7 < tiles; i += 8)
-        {
-            __fp16* tmpptr = tm2.row<__fp16>(i / 8);
-
-            const __fp16* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * packn;
-
-            for (int q = 0; q < inch; q++)
-            {
-#if C906
-                for (int l = 0; l < packn; l++)
-                {
-                    tmpptr[0] = r0[l];
-                    tmpptr[1] = r0[l + packn];
-                    tmpptr[2] = r0[l + packn * 2];
-                    tmpptr[3] = r0[l + packn * 3];
-                    tmpptr[4] = r0[l + packn * 4];
-                    tmpptr[5] = r0[l + packn * 5];
-                    tmpptr[6] = r0[l + packn * 6];
-                    tmpptr[7] = r0[l + packn * 7];
-                    tmpptr += 8;
-                }
-
-                r0 += bottom_blob_tm.cstep * packn;
-#else
-                vfloat16m1_t _val0 = vle16_v_f16m1(r0, vl);
-                vfloat16m1_t _val1 = vle16_v_f16m1(r0 + packn, vl);
-                vfloat16m1_t _val2 = vle16_v_f16m1(r0 + packn * 2, vl);
-                vfloat16m1_t _val3 = vle16_v_f16m1(r0 + packn * 3, vl);
-                vfloat16m1_t _val4 = vle16_v_f16m1(r0 + packn * 4, vl);
-                vfloat16m1_t _val5 = vle16_v_f16m1(r0 + packn * 5, vl);
-                vfloat16m1_t _val6 = vle16_v_f16m1(r0 + packn * 6, vl);
-                vfloat16m1_t _val7 = vle16_v_f16m1(r0 + packn * 7, vl);
-                vsseg8e16_v_f16m1(tmpptr, _val0, _val1, _val2, _val3, _val4, _val5, _val6, _val7, vl);
-
-                r0 += bottom_blob_tm.cstep * packn;
-                tmpptr += packn * 8;
-#endif
-            }
-        }
-        for (; i + 3 < tiles; i += 4)
-        {
-            __fp16* tmpptr = tm2.row<__fp16>(i / 8 + (i % 8) / 4);
-
-            const __fp16* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * packn;
-
-            for (int q = 0; q < inch; q++)
-            {
-#if C906
-                for (int l = 0; l < packn; l++)
-                {
-                    tmpptr[0] = r0[l];
-                    tmpptr[1] = r0[l + packn];
-                    tmpptr[2] = r0[l + packn * 2];
-                    tmpptr[3] = r0[l + packn * 3];
-                    tmpptr += 4;
-                }
-
-                r0 += bottom_blob_tm.cstep * packn;
-#else
-                vfloat16m1_t _val0 = vle16_v_f16m1(r0, vl);
-                vfloat16m1_t _val1 = vle16_v_f16m1(r0 + packn, vl);
-                vfloat16m1_t _val2 = vle16_v_f16m1(r0 + packn * 2, vl);
-                vfloat16m1_t _val3 = vle16_v_f16m1(r0 + packn * 3, vl);
-                vsseg4e16_v_f16m1(tmpptr, _val0, _val1, _val2, _val3, vl);
-
-                r0 += bottom_blob_tm.cstep * packn;
-                tmpptr += packn * 4;
-#endif
-            }
-        }
-        for (; i + 1 < tiles; i += 2)
-        {
-            __fp16* tmpptr = tm2.row<__fp16>(i / 8 + (i % 8) / 4 + (i % 4) / 2);
-
-            const __fp16* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * packn;
-
-            for (int q = 0; q < inch; q++)
-            {
-#if C906
-                for (int l = 0; l < packn; l++)
-                {
-                    tmpptr[0] = r0[l];
-                    tmpptr[1] = r0[l + packn];
-                    tmpptr += 2;
-                }
-
-                r0 += bottom_blob_tm.cstep * packn;
-#else
-                vfloat16m1_t _val0 = vle16_v_f16m1(r0, vl);
-                vfloat16m1_t _val1 = vle16_v_f16m1(r0 + packn, vl);
-                vsseg2e16_v_f16m1(tmpptr, _val0, _val1, vl);
-
-                r0 += bottom_blob_tm.cstep * packn;
-                tmpptr += packn * 2;
-#endif
-            }
-        }
-        for (; i < tiles; i++)
-        {
-            __fp16* tmpptr = tm2.row<__fp16>(i / 8 + (i % 8) / 4 + (i % 4) / 2 + i % 2);
-
-            const __fp16* r0 = bottom_blob_tm;
-
-            r0 += (r * tiles + i) * packn;
-
-            for (int q = 0; q < inch; q++)
-            {
-                vfloat16m1_t _val = vle16_v_f16m1(r0, vl);
-                vse16_v_f16m1(tmpptr, _val, vl);
-
-                r0 += bottom_blob_tm.cstep * packn;
-                tmpptr += packn;
-            }
-        }
-    }
-
-    bottom_blob_tm = Mat();
-    // permute end
-
-    top_blob_tm.create(tiles, batch, outch, 2u * packn, packn, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* output0_tm = top_blob_tm.channel(p);
-
-        const Mat kernel0_tm = kernel_tm.channel(p);
-
-        for (int r = 0; r < batch; r++)
-        {
-            const Mat bb2 = bottom_blob_tm2.channel(r);
-
-            int i = 0;
-            for (; i + 7 < tiles; i += 8)
-            {
-                const __fp16* r0 = bb2.row<const __fp16>(i / 8);
-                const __fp16* k0 = kernel0_tm.row<const __fp16>(r);
-
-                int nn = inch * packn; // inch always > 0
-
-                vfloat16m1_t _sum0 = vfmv_v_f_f16m1(0.f, vl);
-                vfloat16m1_t _sum1 = vfmv_v_f_f16m1(0.f, vl);
-                vfloat16m1_t _sum2 = vfmv_v_f_f16m1(0.f, vl);
-                vfloat16m1_t _sum3 = vfmv_v_f_f16m1(0.f, vl);
-                vfloat16m1_t _sum4 = vfmv_v_f_f16m1(0.f, vl);
-                vfloat16m1_t _sum5 = vfmv_v_f_f16m1(0.f, vl);
-                vfloat16m1_t _sum6 = vfmv_v_f_f16m1(0.f, vl);
-                vfloat16m1_t _sum7 = vfmv_v_f_f16m1(0.f, vl);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __fp16 val0 = *r0++;
-                    __fp16 val1 = *r0++;
-                    __fp16 val2 = *r0++;
-                    __fp16 val3 = *r0++;
-                    __fp16 val4 = *r0++;
-                    __fp16 val5 = *r0++;
-                    __fp16 val6 = *r0++;
-                    __fp16 val7 = *r0++;
-                    vfloat16m1_t _w0 = vle16_v_f16m1(k0, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, val0, _w0, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, val1, _w0, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, val2, _w0, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, val3, _w0, vl);
-                    _sum4 = vfmacc_vf_f16m1(_sum4, val4, _w0, vl);
-                    _sum5 = vfmacc_vf_f16m1(_sum5, val5, _w0, vl);
-                    _sum6 = vfmacc_vf_f16m1(_sum6, val6, _w0, vl);
-                    _sum7 = vfmacc_vf_f16m1(_sum7, val7, _w0, vl);
-
-                    k0 += packn;
-                }
-
-                vse16_v_f16m1(output0_tm, _sum0, vl);
-                vse16_v_f16m1(output0_tm + packn, _sum1, vl);
-                vse16_v_f16m1(output0_tm + packn * 2, _sum2, vl);
-                vse16_v_f16m1(output0_tm + packn * 3, _sum3, vl);
-                vse16_v_f16m1(output0_tm + packn * 4, _sum4, vl);
-                vse16_v_f16m1(output0_tm + packn * 5, _sum5, vl);
-                vse16_v_f16m1(output0_tm + packn * 6, _sum6, vl);
-                vse16_v_f16m1(output0_tm + packn * 7, _sum7, vl);
-
-                output0_tm += packn * 8;
-            }
-            for (; i + 3 < tiles; i += 4)
-            {
-                const __fp16* r0 = bb2.row<const __fp16>(i / 8 + (i % 8) / 4);
-                const __fp16* k0 = kernel0_tm.row<const __fp16>(r);
-
-                int nn = inch * packn; // inch always > 0
-
-                vfloat16m1_t _sum0 = vfmv_v_f_f16m1(0.f, vl);
-                vfloat16m1_t _sum1 = vfmv_v_f_f16m1(0.f, vl);
-                vfloat16m1_t _sum2 = vfmv_v_f_f16m1(0.f, vl);
-                vfloat16m1_t _sum3 = vfmv_v_f_f16m1(0.f, vl);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __fp16 val0 = *r0++;
-                    __fp16 val1 = *r0++;
-                    __fp16 val2 = *r0++;
-                    __fp16 val3 = *r0++;
-                    vfloat16m1_t _w0 = vle16_v_f16m1(k0, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, val0, _w0, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, val1, _w0, vl);
-                    _sum2 = vfmacc_vf_f16m1(_sum2, val2, _w0, vl);
-                    _sum3 = vfmacc_vf_f16m1(_sum3, val3, _w0, vl);
-
-                    k0 += packn;
-                }
-
-                vse16_v_f16m1(output0_tm, _sum0, vl);
-                vse16_v_f16m1(output0_tm + packn, _sum1, vl);
-                vse16_v_f16m1(output0_tm + packn * 2, _sum2, vl);
-                vse16_v_f16m1(output0_tm + packn * 3, _sum3, vl);
-
-                output0_tm += packn * 4;
-            }
-            for (; i + 1 < tiles; i += 2)
-            {
-                const __fp16* r0 = bb2.row<const __fp16>(i / 8 + (i % 8) / 4 + (i % 4) / 2);
-                const __fp16* k0 = kernel0_tm.row<const __fp16>(r);
-
-                int nn = inch * packn; // inch always > 0
-
-                vfloat16m1_t _sum0 = vfmv_v_f_f16m1(0.f, vl);
-                vfloat16m1_t _sum1 = vfmv_v_f_f16m1(0.f, vl);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __fp16 val0 = *r0++;
-                    __fp16 val1 = *r0++;
-                    vfloat16m1_t _w0 = vle16_v_f16m1(k0, vl);
-                    _sum0 = vfmacc_vf_f16m1(_sum0, val0, _w0, vl);
-                    _sum1 = vfmacc_vf_f16m1(_sum1, val1, _w0, vl);
-
-                    k0 += packn;
-                }
-
-                vse16_v_f16m1(output0_tm, _sum0, vl);
-                vse16_v_f16m1(output0_tm + packn, _sum1, vl);
-
-                output0_tm += packn * 2;
-            }
-            for (; i < tiles; i++)
-            {
-                const __fp16* r0 = bb2.row<const __fp16>(i / 8 + (i % 8) / 4 + (i % 4) / 2 + i % 2);
-                const __fp16* k0 = kernel0_tm.row<const __fp16>(r);
-
-                int nn = inch * packn; // inch always > 0
-
-                vfloat16m1_t _sum = vfmv_v_f_f16m1(0.f, vl);
-
-                for (int j = 0; j < nn; j++)
-                {
-                    __fp16 val = *r0++;
-                    vfloat16m1_t _w0 = vle16_v_f16m1(k0, vl);
-                    _sum = vfmacc_vf_f16m1(_sum, val, _w0, vl);
-
-                    k0 += packn;
-                }
-
-                vse16_v_f16m1(output0_tm, _sum, vl);
-
-                output0_tm += packn;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_transform.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_transform.h
deleted file mode 100644
index e286df54..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_transform.h
+++ /dev/null
@@ -1,402 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd43_transform_input_rvv(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 4;
-    const int h_tiles = (h - 2) / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    const float sq2 = 1.41421356237;
-    const float sq2_d2 = 1.41421356237 / 2;
-
-    // const float itm[6][6] = {
-    //     {1.0f,  0.0f,  -2.5f,  0.0f,  1.0f, 0.0f},
-    //     {0.0f, -sq2,   -2.0f,  sq2/2, 1.0f, 0.0f},
-    //     {0.0f,  sq2,   -2.0f, -sq2/2, 1.0f, 0.0f},
-    //     {0.0f, -sq2/2, -0.5f,  sq2,   1.0f, 0.0f},
-    //     {0.0f,  sq2/2, -0.5f, -sq2,   1.0f, 0.0f},
-    //     {0.0f,  1.0f,   0.0f,  -2.5f, 0.0f, 1.0f}
-    // };
-
-    // 0 =  r00 - 2.5f * r02 + r04
-    // 1 = -(sq2 * r01 - sq2_d2 * r03) + (r04 - 2 * r02)
-    // 2 =  (sq2 * r01 - sq2_d2 * r03) + (r04 - 2 * r02)
-    // 3 = -(sq2_d2 * r01 - sq2 * r03) + (r04 - 0.5f * r02)
-    // 4 =  (sq2_d2 * r01 - sq2 * r03) + (r04 - 0.5f * r02)
-    // 5 =  r01 - 2.5f * r03 + r05
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        float tmp[6][6];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* r0 = img0.row(i * 4) + (j * 4);
-
-                for (int m = 0; m < 6; m++)
-                {
-                    float r00 = r0[0];
-                    float r01 = r0[1];
-                    float r02 = r0[2];
-                    float r03 = r0[3];
-                    float r04 = r0[4];
-                    float r05 = r0[5];
-
-                    float tmp01a = sq2 * r01 - sq2_d2 * r03;
-                    float tmp01b = r04 - 2 * r02;
-                    float tmp23a = sq2_d2 * r01 - sq2 * r03;
-                    float tmp23b = r04 - 0.5f * r02;
-
-                    tmp[0][m] = r00 - 2.5f * r02 + r04;
-                    tmp[1][m] = tmp01b - tmp01a;
-                    tmp[2][m] = tmp01b + tmp01a;
-                    tmp[3][m] = tmp23b - tmp23a;
-                    tmp[4][m] = tmp23b + tmp23a;
-                    tmp[5][m] = r01 - 2.5f * r03 + r05;
-
-                    r0 += w;
-                }
-
-                float* r0_tm_0 = (float*)img0_tm + (i * w_tiles + j);
-                float* r0_tm_1 = r0_tm_0 + tiles;
-                float* r0_tm_2 = r0_tm_0 + tiles * 2;
-                float* r0_tm_3 = r0_tm_0 + tiles * 3;
-                float* r0_tm_4 = r0_tm_0 + tiles * 4;
-                float* r0_tm_5 = r0_tm_0 + tiles * 5;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    float r00 = tmp[m][0];
-                    float r01 = tmp[m][1];
-                    float r02 = tmp[m][2];
-                    float r03 = tmp[m][3];
-                    float r04 = tmp[m][4];
-                    float r05 = tmp[m][5];
-
-                    float tmp01a = sq2 * r01 - sq2_d2 * r03;
-                    float tmp01b = r04 - 2 * r02;
-                    float tmp23a = sq2_d2 * r01 - sq2 * r03;
-                    float tmp23b = r04 - 0.5f * r02;
-
-                    r0_tm_0[0] = r00 - 2.5f * r02 + r04;
-                    r0_tm_1[0] = tmp01b - tmp01a;
-                    r0_tm_2[0] = tmp01b + tmp01a;
-                    r0_tm_3[0] = tmp23b - tmp23a;
-                    r0_tm_4[0] = tmp23b + tmp23a;
-                    r0_tm_5[0] = r01 - 2.5f * r03 + r05;
-
-                    r0_tm_0 += tiles * 6;
-                    r0_tm_1 += tiles * 6;
-                    r0_tm_2 += tiles * 6;
-                    r0_tm_3 += tiles * 6;
-                    r0_tm_4 += tiles * 6;
-                    r0_tm_5 += tiles * 6;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_output_rvv(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 4;
-    const int h_tiles = outh / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    const float* biasptr = bias;
-
-    const float sq2 = 1.41421356237;
-    const float sq2_m2 = 1.41421356237 * 2;
-    const float sq2_d2 = 1.41421356237 / 2;
-    const float sq2_d4 = 1.41421356237 / 4;
-
-    // const float otm[4][6] = {
-    //     {1.0f, 1.0f,   1.0f,  1.0f,  1.0f,   0.0f},
-    //     {0.0f, sq2/2, -sq2/2, sq2,   -sq2,   0.0f},
-    //     {0.0f, 0.5f,   0.5f,  2.0f,  2.0f,   0.0f},
-    //     {0.0f, sq2/4, -sq2/4, sq2*2, -sq2*2, 1.0f}
-    // };
-
-    // 0 = r00 + (r01 + r02) + (r03 + r04)
-    // 1 =       (r01 - r02) * sq2_d2 + (r03 - r04) * sq2
-    // 2 =       (r01 + r02) * 0.5f + (r03 + r04) * 2
-    // 3 = r05 + (r01 - r02) * sq2_d4 + (r03 - r04) * sq2_m2
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        float bias0 = biasptr ? biasptr[p] : 0.f;
-
-        float tmp[4][6];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* output0_tm_0 = (const float*)out0_tm + (i * w_tiles + j);
-                const float* output0_tm_1 = output0_tm_0 + tiles;
-                const float* output0_tm_2 = output0_tm_0 + tiles * 2;
-                const float* output0_tm_3 = output0_tm_0 + tiles * 3;
-                const float* output0_tm_4 = output0_tm_0 + tiles * 4;
-                const float* output0_tm_5 = output0_tm_0 + tiles * 5;
-
-                float* output0 = out0.row(i * 4) + (j * 4);
-
-                for (int m = 0; m < 6; m++)
-                {
-                    float r00 = output0_tm_0[0];
-                    float r01 = output0_tm_1[0];
-                    float r02 = output0_tm_2[0];
-                    float r03 = output0_tm_3[0];
-                    float r04 = output0_tm_4[0];
-                    float r05 = output0_tm_5[0];
-
-                    float tmp02a = r01 + r02;
-                    float tmp02b = r03 + r04;
-                    float tmp13a = r01 - r02;
-                    float tmp13b = r03 - r04;
-
-                    tmp[0][m] = r00 + tmp02a + tmp02b;
-                    tmp[1][m] = tmp13a * sq2_d2 + tmp13b * sq2;
-                    tmp[2][m] = tmp02a * 0.5f + tmp02b * 2;
-                    tmp[3][m] = r05 + tmp13a * sq2_d4 + tmp13b * sq2_m2;
-
-                    output0_tm_0 += tiles * 6;
-                    output0_tm_1 += tiles * 6;
-                    output0_tm_2 += tiles * 6;
-                    output0_tm_3 += tiles * 6;
-                    output0_tm_4 += tiles * 6;
-                    output0_tm_5 += tiles * 6;
-                }
-
-                for (int m = 0; m < 4; m++)
-                {
-                    float r00 = tmp[m][0];
-                    float r01 = tmp[m][1];
-                    float r02 = tmp[m][2];
-                    float r03 = tmp[m][3];
-                    float r04 = tmp[m][4];
-                    float r05 = tmp[m][5];
-
-                    float tmp02a = r01 + r02;
-                    float tmp02b = r03 + r04;
-                    float tmp13a = r01 - r02;
-                    float tmp13b = r03 - r04;
-
-                    float out00 = bias0 + r00 + tmp02a + tmp02b;
-                    float out01 = bias0 + tmp13a * sq2_d2 + tmp13b * sq2;
-                    float out02 = bias0 + tmp02a * 0.5f + tmp02b * 2;
-                    float out03 = bias0 + r05 + tmp13a * sq2_d4 + tmp13b * sq2_m2;
-
-                    output0[0] = out00;
-                    output0[1] = out01;
-                    output0[2] = out02;
-                    output0[3] = out03;
-
-                    output0 += outw;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_input_rvv(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 2;
-    const int h_tiles = (h - 2) / 2;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[4][4] = {
-    //     {1.0f,  0.0f, -1.0f,  0.0f},
-    //     {0.0f,  1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  0.00f, 1.0f}
-    // };
-
-    // 0 = r00 - r02
-    // 1 = r01 + r02
-    // 2 = r02 - r01
-    // 3 = r03 - r01
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        float tmp[4][4];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* r0 = img0.row(i * 2) + (j * 2);
-
-                for (int m = 0; m < 4; m++)
-                {
-                    float r00 = r0[0];
-                    float r01 = r0[1];
-                    float r02 = r0[2];
-                    float r03 = r0[3];
-
-                    float tmp0m = r00 - r02;
-                    float tmp1m = r01 + r02;
-                    float tmp2m = r02 - r01;
-                    float tmp3m = r03 - r01;
-
-                    tmp[0][m] = tmp0m;
-                    tmp[1][m] = tmp1m;
-                    tmp[2][m] = tmp2m;
-                    tmp[3][m] = tmp3m;
-
-                    r0 += w;
-                }
-
-                float* r0_tm_0 = (float*)img0_tm + (i * w_tiles + j);
-                float* r0_tm_1 = r0_tm_0 + tiles;
-                float* r0_tm_2 = r0_tm_0 + tiles * 2;
-                float* r0_tm_3 = r0_tm_0 + tiles * 3;
-
-                for (int m = 0; m < 4; m++)
-                {
-                    float tmp00 = tmp[m][0];
-                    float tmp01 = tmp[m][1];
-                    float tmp02 = tmp[m][2];
-                    float tmp03 = tmp[m][3];
-
-                    float r0tm0 = tmp00 - tmp02;
-                    float r0tm1 = tmp01 + tmp02;
-                    float r0tm2 = tmp02 - tmp01;
-                    float r0tm3 = tmp03 - tmp01;
-
-                    r0_tm_0[0] = r0tm0;
-                    r0_tm_1[0] = r0tm1;
-                    r0_tm_2[0] = r0tm2;
-                    r0_tm_3[0] = r0tm3;
-
-                    r0_tm_0 += tiles * 4;
-                    r0_tm_1 += tiles * 4;
-                    r0_tm_2 += tiles * 4;
-                    r0_tm_3 += tiles * 4;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_output_rvv(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 2;
-    const int h_tiles = outh / 2;
-    const int tiles = w_tiles * h_tiles;
-
-    const float* biasptr = bias;
-
-    // const float otm[2][4] = {
-    //     {1.0f,  1.0f,  1.0f,  0.0f},
-    //     {0.0f,  1.0f, -1.0f,  1.0f}
-    // };
-
-    // 0 = r00 + r01 + r02
-    // 1 = r01 - r02 + r03
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        float bias0 = biasptr ? biasptr[p] : 0.f;
-
-        float tmp[2][4];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* output0_tm_0 = (const float*)out0_tm + (i * w_tiles + j);
-                const float* output0_tm_1 = output0_tm_0 + tiles;
-                const float* output0_tm_2 = output0_tm_0 + tiles * 2;
-                const float* output0_tm_3 = output0_tm_0 + tiles * 3;
-
-                float* output0 = out0.row(i * 2) + (j * 2);
-
-                for (int m = 0; m < 4; m++)
-                {
-                    float out0tm0 = output0_tm_0[0];
-                    float out0tm1 = output0_tm_1[0];
-                    float out0tm2 = output0_tm_2[0];
-                    float out0tm3 = output0_tm_3[0];
-
-                    float tmp0m = out0tm0 + out0tm1 + out0tm2;
-                    float tmp1m = out0tm1 - out0tm2 + out0tm3;
-
-                    tmp[0][m] = tmp0m;
-                    tmp[1][m] = tmp1m;
-
-                    output0_tm_0 += tiles * 4;
-                    output0_tm_1 += tiles * 4;
-                    output0_tm_2 += tiles * 4;
-                    output0_tm_3 += tiles * 4;
-                }
-
-                for (int m = 0; m < 2; m++)
-                {
-                    float tmp00 = tmp[m][0];
-                    float tmp01 = tmp[m][1];
-                    float tmp02 = tmp[m][2];
-                    float tmp03 = tmp[m][3];
-
-                    float out00 = bias0 + tmp00 + tmp01 + tmp02;
-                    float out01 = bias0 + tmp01 - tmp02 + tmp03;
-
-                    output0[0] = out00;
-                    output0[1] = out01;
-
-                    output0 += outw;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_transform_packn.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_transform_packn.h
deleted file mode 100644
index 874796af..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_transform_packn.h
+++ /dev/null
@@ -1,745 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd63_transform_input_packn_rvv(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 6;
-    const int h_tiles = (h - 2) / 6;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[8][8] = {
-    //     {1.0f,  0.0f, -5.25f,  0.00f,  5.25f,  0.00f, -1.0f, 0.0f},
-    //
-    //     {0.0f,  1.0f,  1.00f, -4.25f, -4.25f,  1.00f,  1.0f, 0.0f},
-    //     {0.0f, -1.0f,  1.00f,  4.25f, -4.25f, -1.00f,  1.0f, 0.0f},
-    //
-    //     {0.0f,  0.5f,  0.25f, -2.50f, -1.25f,  2.00f,  1.0f, 0.0f},
-    //     {0.0f, -0.5f,  0.25f,  2.50f, -1.25f, -2.00f,  1.0f, 0.0f},
-    //
-    //     {0.0f,  2.0f,  4.00f, -2.50f, -5.00f,  0.50f,  1.0f, 0.0f},
-    //     {0.0f, -2.0f,  4.00f,  2.50f, -5.00f, -0.50f,  1.0f, 0.0f},
-    //
-    //     {0.0f, -1.0f,  0.00f,  5.25f,  0.00f, -5.25f,  0.0f, 1.0f}
-    // };
-
-    // 0 = r00 - r06 + (r04 - r02) * 5.25
-    // 7 = r07 - r01 + (r03 - r05) * 5.25
-
-    // 1 = (r02 + r06 - r04 * 4.25) + (r01 - r03 * 4.25 + r05)
-    // 2 = (r02 + r06 - r04 * 4.25) - (r01 - r03 * 4.25 + r05)
-
-    // 3 = (r06 + r02 * 0.25 - r04 * 1.25) + (r01 * 0.5 - r03 * 2.5 + r05 * 2)
-    // 4 = (r06 + r02 * 0.25 - r04 * 1.25) - (r01 * 0.5 - r03 * 2.5 + r05 * 2)
-
-    // reuse r04 * 1.25
-    // reuse r03 * 2.5
-    // 5 = (r06 + (r02 - r04 * 1.25) * 4) + (r01 * 2 - r03 * 2.5 + r05 * 0.5)
-    // 6 = (r06 + (r02 - r04 * 1.25) * 4) - (r01 * 2 - r03 * 2.5 + r05 * 0.5)
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        // NOTE c99 variable length array
-        float tmp[8][8][packn];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* r0 = img0.row(i * 6) + (j * 6) * packn;
-
-                for (int m = 0; m < 8; m++)
-                {
-                    vfloat32m1_t _r00 = vle32_v_f32m1(r0, vl);
-                    vfloat32m1_t _r01 = vle32_v_f32m1(r0 + packn, vl);
-                    vfloat32m1_t _r02 = vle32_v_f32m1(r0 + packn * 2, vl);
-                    vfloat32m1_t _r03 = vle32_v_f32m1(r0 + packn * 3, vl);
-                    vfloat32m1_t _r04 = vle32_v_f32m1(r0 + packn * 4, vl);
-                    vfloat32m1_t _r05 = vle32_v_f32m1(r0 + packn * 5, vl);
-                    vfloat32m1_t _r06 = vle32_v_f32m1(r0 + packn * 6, vl);
-                    vfloat32m1_t _r07 = vle32_v_f32m1(r0 + packn * 7, vl);
-
-                    vfloat32m1_t _tmp0m = vfmacc_vf_f32m1(vfsub_vv_f32m1(_r00, _r06, vl), 5.25f, vfsub_vv_f32m1(_r04, _r02, vl), vl);
-                    vfloat32m1_t _tmp7m = vfmacc_vf_f32m1(vfsub_vv_f32m1(_r07, _r01, vl), 5.25f, vfsub_vv_f32m1(_r03, _r05, vl), vl);
-                    vse32_v_f32m1(tmp[0][m], _tmp0m, vl);
-                    vse32_v_f32m1(tmp[7][m], _tmp7m, vl);
-
-                    vfloat32m1_t _tmp12a = vfmacc_vf_f32m1(vfadd_vv_f32m1(_r02, _r06, vl), -4.25f, _r04, vl);
-                    vfloat32m1_t _tmp12b = vfmacc_vf_f32m1(vfadd_vv_f32m1(_r01, _r05, vl), -4.25f, _r03, vl);
-
-                    vfloat32m1_t _tmp1m = vfadd_vv_f32m1(_tmp12a, _tmp12b, vl);
-                    vfloat32m1_t _tmp2m = vfsub_vv_f32m1(_tmp12a, _tmp12b, vl);
-                    vse32_v_f32m1(tmp[1][m], _tmp1m, vl);
-                    vse32_v_f32m1(tmp[2][m], _tmp2m, vl);
-
-                    vfloat32m1_t _tmp34a = vfmacc_vf_f32m1(vfmacc_vf_f32m1(_r06, 0.25f, _r02, vl), -1.25f, _r04, vl);
-                    vfloat32m1_t _tmp34b = vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmul_vf_f32m1(_r01, 0.5f, vl), -2.5f, _r03, vl), 2.f, _r05, vl);
-
-                    vfloat32m1_t _tmp3m = vfadd_vv_f32m1(_tmp34a, _tmp34b, vl);
-                    vfloat32m1_t _tmp4m = vfsub_vv_f32m1(_tmp34a, _tmp34b, vl);
-                    vse32_v_f32m1(tmp[3][m], _tmp3m, vl);
-                    vse32_v_f32m1(tmp[4][m], _tmp4m, vl);
-
-                    vfloat32m1_t _tmp56a = vfmacc_vf_f32m1(_r06, 4.f, vfmacc_vf_f32m1(_r02, -1.25f, _r04, vl), vl);
-                    vfloat32m1_t _tmp56b = vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmul_vf_f32m1(_r01, 2.f, vl), -2.5f, _r03, vl), 0.5f, _r05, vl);
-
-                    vfloat32m1_t _tmp5m = vfadd_vv_f32m1(_tmp56a, _tmp56b, vl);
-                    vfloat32m1_t _tmp6m = vfsub_vv_f32m1(_tmp56a, _tmp56b, vl);
-                    vse32_v_f32m1(tmp[5][m], _tmp5m, vl);
-                    vse32_v_f32m1(tmp[6][m], _tmp6m, vl);
-
-                    r0 += w * packn;
-                }
-
-                float* r0_tm_0 = (float*)img0_tm + (i * w_tiles + j) * packn;
-                float* r0_tm_1 = r0_tm_0 + tiles * packn;
-                float* r0_tm_2 = r0_tm_0 + tiles * packn * 2;
-                float* r0_tm_3 = r0_tm_0 + tiles * packn * 3;
-                float* r0_tm_4 = r0_tm_0 + tiles * packn * 4;
-                float* r0_tm_5 = r0_tm_0 + tiles * packn * 5;
-                float* r0_tm_6 = r0_tm_0 + tiles * packn * 6;
-                float* r0_tm_7 = r0_tm_0 + tiles * packn * 7;
-
-                for (int m = 0; m < 8; m++)
-                {
-                    vfloat32m1_t _tmp00 = vle32_v_f32m1(tmp[m][0], vl);
-                    vfloat32m1_t _tmp01 = vle32_v_f32m1(tmp[m][1], vl);
-                    vfloat32m1_t _tmp02 = vle32_v_f32m1(tmp[m][2], vl);
-                    vfloat32m1_t _tmp03 = vle32_v_f32m1(tmp[m][3], vl);
-                    vfloat32m1_t _tmp04 = vle32_v_f32m1(tmp[m][4], vl);
-                    vfloat32m1_t _tmp05 = vle32_v_f32m1(tmp[m][5], vl);
-                    vfloat32m1_t _tmp06 = vle32_v_f32m1(tmp[m][6], vl);
-                    vfloat32m1_t _tmp07 = vle32_v_f32m1(tmp[m][7], vl);
-
-                    vfloat32m1_t _r0tm0 = vfmacc_vf_f32m1(vfsub_vv_f32m1(_tmp00, _tmp06, vl), 5.25f, vfsub_vv_f32m1(_tmp04, _tmp02, vl), vl);
-                    vfloat32m1_t _r0tm7 = vfmacc_vf_f32m1(vfsub_vv_f32m1(_tmp07, _tmp01, vl), 5.25f, vfsub_vv_f32m1(_tmp03, _tmp05, vl), vl);
-
-                    vfloat32m1_t _tmp12a = vfmacc_vf_f32m1(vfadd_vv_f32m1(_tmp02, _tmp06, vl), -4.25f, _tmp04, vl);
-                    vfloat32m1_t _tmp12b = vfmacc_vf_f32m1(vfadd_vv_f32m1(_tmp01, _tmp05, vl), -4.25f, _tmp03, vl);
-
-                    vfloat32m1_t _r0tm1 = vfadd_vv_f32m1(_tmp12a, _tmp12b, vl);
-                    vfloat32m1_t _r0tm2 = vfsub_vv_f32m1(_tmp12a, _tmp12b, vl);
-
-                    vfloat32m1_t _tmp34a = vfmacc_vf_f32m1(vfmacc_vf_f32m1(_tmp06, 0.25f, _tmp02, vl), -1.25f, _tmp04, vl);
-                    vfloat32m1_t _tmp34b = vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmul_vf_f32m1(_tmp01, 0.5f, vl), -2.5f, _tmp03, vl), 2.f, _tmp05, vl);
-
-                    vfloat32m1_t _r0tm3 = vfadd_vv_f32m1(_tmp34a, _tmp34b, vl);
-                    vfloat32m1_t _r0tm4 = vfsub_vv_f32m1(_tmp34a, _tmp34b, vl);
-
-                    vfloat32m1_t _tmp56a = vfmacc_vf_f32m1(_tmp06, 4.f, vfmacc_vf_f32m1(_tmp02, -1.25f, _tmp04, vl), vl);
-                    vfloat32m1_t _tmp56b = vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmul_vf_f32m1(_tmp01, 2.f, vl), -2.5f, _tmp03, vl), 0.5f, _tmp05, vl);
-
-                    vfloat32m1_t _r0tm5 = vfadd_vv_f32m1(_tmp56a, _tmp56b, vl);
-                    vfloat32m1_t _r0tm6 = vfsub_vv_f32m1(_tmp56a, _tmp56b, vl);
-
-                    vse32_v_f32m1(r0_tm_0, _r0tm0, vl);
-                    vse32_v_f32m1(r0_tm_1, _r0tm1, vl);
-                    vse32_v_f32m1(r0_tm_2, _r0tm2, vl);
-                    vse32_v_f32m1(r0_tm_3, _r0tm3, vl);
-                    vse32_v_f32m1(r0_tm_4, _r0tm4, vl);
-                    vse32_v_f32m1(r0_tm_5, _r0tm5, vl);
-                    vse32_v_f32m1(r0_tm_6, _r0tm6, vl);
-                    vse32_v_f32m1(r0_tm_7, _r0tm7, vl);
-
-                    r0_tm_0 += tiles * packn * 8;
-                    r0_tm_1 += tiles * packn * 8;
-                    r0_tm_2 += tiles * packn * 8;
-                    r0_tm_3 += tiles * packn * 8;
-                    r0_tm_4 += tiles * packn * 8;
-                    r0_tm_5 += tiles * packn * 8;
-                    r0_tm_6 += tiles * packn * 8;
-                    r0_tm_7 += tiles * packn * 8;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd63_transform_output_packn_rvv(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 6;
-    const int h_tiles = outh / 6;
-    const int tiles = w_tiles * h_tiles;
-
-    const float* biasptr = bias;
-
-    // const float otm[6][8] = {
-    //     {1.0f,  1.0f,   1.0f,   1.0f,   1.0f,  32.0f, 32.0f, 0.0f},
-    //     {0.0f,  1.0f,  -1.0f,   2.0f,  -2.0f,  16.0f,-16.0f, 0.0f},
-    //     {0.0f,  1.0f,   1.0f,   4.0f,   4.0f,   8.0f,  8.0f, 0.0f},
-    //     {0.0f,  1.0f,  -1.0f,   8.0f,  -8.0f,   4.0f, -4.0f, 0.0f},
-    //     {0.0f,  1.0f,   1.0f,  16.0f,  16.0f,   2.0f,  2.0f, 0.0f},
-    //     {0.0f,  1.0f,  -1.0f,  32.0f, -32.0f,   1.0f, -1.0f, 1.0f}
-    // };
-
-    // 0 = r0 + (r1 + r2) + (r3 + r4)     + (r5 + r6) * 32
-    // 1 =      (r1 - r2) + (r3 - r4) * 2 + (r5 - r6) * 16
-    // 2 =      (r1 + r2) + (r3 + r4) * 4 + (r5 + r6) * 8
-    // 3 =      (r1 - r2) + (r3 - r4) * 8 + (r5 - r6) * 4
-    // 4 =      (r1 + r2) + (r3 + r4) * 16+ (r5 + r6) * 2
-    // 5 = r7 + (r1 - r2) + (r3 - r4) * 32+ (r5 - r6)
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        vfloat32m1_t _bias0 = biasptr ? vle32_v_f32m1(biasptr + p * packn, vl) : vfmv_v_f_f32m1(0.f, vl);
-
-        // NOTE c99 variable length array
-        float tmp[6][8][packn];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* output0_tm_0 = (const float*)out0_tm + (i * w_tiles + j) * packn;
-                const float* output0_tm_1 = output0_tm_0 + tiles * packn;
-                const float* output0_tm_2 = output0_tm_0 + tiles * packn * 2;
-                const float* output0_tm_3 = output0_tm_0 + tiles * packn * 3;
-                const float* output0_tm_4 = output0_tm_0 + tiles * packn * 4;
-                const float* output0_tm_5 = output0_tm_0 + tiles * packn * 5;
-                const float* output0_tm_6 = output0_tm_0 + tiles * packn * 6;
-                const float* output0_tm_7 = output0_tm_0 + tiles * packn * 7;
-
-                float* output0 = out0.row(i * 6) + (j * 6) * packn;
-
-                for (int m = 0; m < 8; m++)
-                {
-                    vfloat32m1_t _out0tm0 = vle32_v_f32m1(output0_tm_0, vl);
-                    vfloat32m1_t _out0tm1 = vle32_v_f32m1(output0_tm_1, vl);
-                    vfloat32m1_t _out0tm2 = vle32_v_f32m1(output0_tm_2, vl);
-                    vfloat32m1_t _out0tm3 = vle32_v_f32m1(output0_tm_3, vl);
-                    vfloat32m1_t _out0tm4 = vle32_v_f32m1(output0_tm_4, vl);
-                    vfloat32m1_t _out0tm5 = vle32_v_f32m1(output0_tm_5, vl);
-                    vfloat32m1_t _out0tm6 = vle32_v_f32m1(output0_tm_6, vl);
-                    vfloat32m1_t _out0tm7 = vle32_v_f32m1(output0_tm_7, vl);
-
-                    vfloat32m1_t _tmp024a = vfadd_vv_f32m1(_out0tm1, _out0tm2, vl);
-                    vfloat32m1_t _tmp135a = vfsub_vv_f32m1(_out0tm1, _out0tm2, vl);
-
-                    vfloat32m1_t _tmp024b = vfadd_vv_f32m1(_out0tm3, _out0tm4, vl);
-                    vfloat32m1_t _tmp135b = vfsub_vv_f32m1(_out0tm3, _out0tm4, vl);
-
-                    vfloat32m1_t _tmp024c = vfadd_vv_f32m1(_out0tm5, _out0tm6, vl);
-                    vfloat32m1_t _tmp135c = vfsub_vv_f32m1(_out0tm5, _out0tm6, vl);
-
-                    vfloat32m1_t _tmp0m = vfadd_vv_f32m1(vfadd_vv_f32m1(_out0tm0, _tmp024a, vl), vfmacc_vf_f32m1(_tmp024b, 32.f, _tmp024c, vl), vl);
-                    vfloat32m1_t _tmp2m = vfmacc_vf_f32m1(vfmacc_vf_f32m1(_tmp024a, 4.f, _tmp024b, vl), 8.f, _tmp024c, vl);
-                    vfloat32m1_t _tmp4m = vfmacc_vf_f32m1(vfmacc_vf_f32m1(_tmp024a, 16.f, _tmp024b, vl), 2.f, _tmp024c, vl);
-                    vse32_v_f32m1(tmp[0][m], _tmp0m, vl);
-                    vse32_v_f32m1(tmp[2][m], _tmp2m, vl);
-                    vse32_v_f32m1(tmp[4][m], _tmp4m, vl);
-
-                    vfloat32m1_t _tmp1m = vfmacc_vf_f32m1(vfmacc_vf_f32m1(_tmp135a, 2.f, _tmp135b, vl), 16.f, _tmp135c, vl);
-                    vfloat32m1_t _tmp3m = vfmacc_vf_f32m1(vfmacc_vf_f32m1(_tmp135a, 8.f, _tmp135b, vl), 4.f, _tmp135c, vl);
-                    vfloat32m1_t _tmp5m = vfadd_vv_f32m1(vfadd_vv_f32m1(_out0tm7, _tmp135a, vl), vfmacc_vf_f32m1(_tmp135c, 32.f, _tmp135b, vl), vl);
-                    vse32_v_f32m1(tmp[1][m], _tmp1m, vl);
-                    vse32_v_f32m1(tmp[3][m], _tmp3m, vl);
-                    vse32_v_f32m1(tmp[5][m], _tmp5m, vl);
-
-                    output0_tm_0 += tiles * packn * 8;
-                    output0_tm_1 += tiles * packn * 8;
-                    output0_tm_2 += tiles * packn * 8;
-                    output0_tm_3 += tiles * packn * 8;
-                    output0_tm_4 += tiles * packn * 8;
-                    output0_tm_5 += tiles * packn * 8;
-                    output0_tm_6 += tiles * packn * 8;
-                    output0_tm_7 += tiles * packn * 8;
-                }
-
-                for (int m = 0; m < 6; m++)
-                {
-                    vfloat32m1_t _tmp00 = vle32_v_f32m1(tmp[m][0], vl);
-                    vfloat32m1_t _tmp01 = vle32_v_f32m1(tmp[m][1], vl);
-                    vfloat32m1_t _tmp02 = vle32_v_f32m1(tmp[m][2], vl);
-                    vfloat32m1_t _tmp03 = vle32_v_f32m1(tmp[m][3], vl);
-                    vfloat32m1_t _tmp04 = vle32_v_f32m1(tmp[m][4], vl);
-                    vfloat32m1_t _tmp05 = vle32_v_f32m1(tmp[m][5], vl);
-                    vfloat32m1_t _tmp06 = vle32_v_f32m1(tmp[m][6], vl);
-                    vfloat32m1_t _tmp07 = vle32_v_f32m1(tmp[m][7], vl);
-
-                    vfloat32m1_t _tmp024a = vfadd_vv_f32m1(_tmp01, _tmp02, vl);
-                    vfloat32m1_t _tmp135a = vfsub_vv_f32m1(_tmp01, _tmp02, vl);
-
-                    vfloat32m1_t _tmp024b = vfadd_vv_f32m1(_tmp03, _tmp04, vl);
-                    vfloat32m1_t _tmp135b = vfsub_vv_f32m1(_tmp03, _tmp04, vl);
-
-                    vfloat32m1_t _tmp024c = vfadd_vv_f32m1(_tmp05, _tmp06, vl);
-                    vfloat32m1_t _tmp135c = vfsub_vv_f32m1(_tmp05, _tmp06, vl);
-
-                    vfloat32m1_t _out00 = vfadd_vv_f32m1(_bias0, vfadd_vv_f32m1(vfadd_vv_f32m1(_tmp00, _tmp024a, vl), vfmacc_vf_f32m1(_tmp024b, 32.f, _tmp024c, vl), vl), vl);
-                    vfloat32m1_t _out02 = vfadd_vv_f32m1(_bias0, vfmacc_vf_f32m1(vfmacc_vf_f32m1(_tmp024a, 4.f, _tmp024b, vl), 8.f, _tmp024c, vl), vl);
-                    vfloat32m1_t _out04 = vfadd_vv_f32m1(_bias0, vfmacc_vf_f32m1(vfmacc_vf_f32m1(_tmp024a, 16.f, _tmp024b, vl), 2.f, _tmp024c, vl), vl);
-                    vse32_v_f32m1(output0, _out00, vl);
-                    vse32_v_f32m1(output0 + packn * 2, _out02, vl);
-                    vse32_v_f32m1(output0 + packn * 4, _out04, vl);
-
-                    vfloat32m1_t _out01 = vfadd_vv_f32m1(_bias0, vfmacc_vf_f32m1(vfmacc_vf_f32m1(_tmp135a, 2.f, _tmp135b, vl), 16.f, _tmp135c, vl), vl);
-                    vfloat32m1_t _out03 = vfadd_vv_f32m1(_bias0, vfmacc_vf_f32m1(vfmacc_vf_f32m1(_tmp135a, 8.f, _tmp135b, vl), 4.f, _tmp135c, vl), vl);
-                    vfloat32m1_t _out05 = vfadd_vv_f32m1(_bias0, vfadd_vv_f32m1(vfadd_vv_f32m1(_tmp07, _tmp135a, vl), vfmacc_vf_f32m1(_tmp135c, 32.f, _tmp135b, vl), vl), vl);
-                    vse32_v_f32m1(output0 + packn, _out01, vl);
-                    vse32_v_f32m1(output0 + packn * 3, _out03, vl);
-                    vse32_v_f32m1(output0 + packn * 5, _out05, vl);
-
-                    output0 += outw * packn;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_input_packn_rvv(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 4;
-    const int h_tiles = (h - 2) / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    const float sq2 = 1.41421356237;
-    const float sq2_d2 = 1.41421356237 / 2;
-
-    // const float itm[6][6] = {
-    //     {1.0f,  0.0f,  -2.5f,  0.0f,  1.0f, 0.0f},
-    //     {0.0f, -sq2,   -2.0f,  sq2/2, 1.0f, 0.0f},
-    //     {0.0f,  sq2,   -2.0f, -sq2/2, 1.0f, 0.0f},
-    //     {0.0f, -sq2/2, -0.5f,  sq2,   1.0f, 0.0f},
-    //     {0.0f,  sq2/2, -0.5f, -sq2,   1.0f, 0.0f},
-    //     {0.0f,  1.0f,   0.0f,  -2.5f, 0.0f, 1.0f}
-    // };
-
-    // 0 =  r00 - 2.5f * r02 + r04
-    // 1 = -(sq2 * r01 - sq2_d2 * r03) + (r04 - 2 * r02)
-    // 2 =  (sq2 * r01 - sq2_d2 * r03) + (r04 - 2 * r02)
-    // 3 = -(sq2_d2 * r01 - sq2 * r03) + (r04 - 0.5f * r02)
-    // 4 =  (sq2_d2 * r01 - sq2 * r03) + (r04 - 0.5f * r02)
-    // 5 =  r01 - 2.5f * r03 + r05
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        // NOTE c99 variable length array
-        float tmp[6][6][packn];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* r0 = img0.row(i * 4) + (j * 4) * packn;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    vfloat32m1_t _r00 = vle32_v_f32m1(r0, vl);
-                    vfloat32m1_t _r01 = vle32_v_f32m1(r0 + packn, vl);
-                    vfloat32m1_t _r02 = vle32_v_f32m1(r0 + packn * 2, vl);
-                    vfloat32m1_t _r03 = vle32_v_f32m1(r0 + packn * 3, vl);
-                    vfloat32m1_t _r04 = vle32_v_f32m1(r0 + packn * 4, vl);
-                    vfloat32m1_t _r05 = vle32_v_f32m1(r0 + packn * 5, vl);
-
-                    vfloat32m1_t _tmp01a = vfmacc_vf_f32m1(vfmul_vf_f32m1(_r01, sq2, vl), -sq2_d2, _r03, vl);
-                    vfloat32m1_t _tmp01b = vfmacc_vf_f32m1(_r04, -2.f, _r02, vl);
-                    vfloat32m1_t _tmp23a = vfmacc_vf_f32m1(vfmul_vf_f32m1(_r01, sq2_d2, vl), -sq2, _r03, vl);
-                    vfloat32m1_t _tmp23b = vfmacc_vf_f32m1(_r04, -0.5f, _r02, vl);
-
-                    vfloat32m1_t _tmp0m = vfmacc_vf_f32m1(vfadd_vv_f32m1(_r00, _r04, vl), -2.5f, _r02, vl);
-                    vfloat32m1_t _tmp1m = vfsub_vv_f32m1(_tmp01b, _tmp01a, vl);
-                    vfloat32m1_t _tmp2m = vfadd_vv_f32m1(_tmp01b, _tmp01a, vl);
-                    vfloat32m1_t _tmp3m = vfsub_vv_f32m1(_tmp23b, _tmp23a, vl);
-                    vfloat32m1_t _tmp4m = vfadd_vv_f32m1(_tmp23b, _tmp23a, vl);
-                    vfloat32m1_t _tmp5m = vfmacc_vf_f32m1(vfadd_vv_f32m1(_r01, _r05, vl), -2.5f, _r03, vl);
-
-                    vse32_v_f32m1(tmp[0][m], _tmp0m, vl);
-                    vse32_v_f32m1(tmp[1][m], _tmp1m, vl);
-                    vse32_v_f32m1(tmp[2][m], _tmp2m, vl);
-                    vse32_v_f32m1(tmp[3][m], _tmp3m, vl);
-                    vse32_v_f32m1(tmp[4][m], _tmp4m, vl);
-                    vse32_v_f32m1(tmp[5][m], _tmp5m, vl);
-
-                    r0 += w * packn;
-                }
-
-                float* r0_tm_0 = (float*)img0_tm + (i * w_tiles + j) * packn;
-                float* r0_tm_1 = r0_tm_0 + tiles * packn;
-                float* r0_tm_2 = r0_tm_0 + tiles * packn * 2;
-                float* r0_tm_3 = r0_tm_0 + tiles * packn * 3;
-                float* r0_tm_4 = r0_tm_0 + tiles * packn * 4;
-                float* r0_tm_5 = r0_tm_0 + tiles * packn * 5;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    vfloat32m1_t _r00 = vle32_v_f32m1(tmp[m][0], vl);
-                    vfloat32m1_t _r01 = vle32_v_f32m1(tmp[m][1], vl);
-                    vfloat32m1_t _r02 = vle32_v_f32m1(tmp[m][2], vl);
-                    vfloat32m1_t _r03 = vle32_v_f32m1(tmp[m][3], vl);
-                    vfloat32m1_t _r04 = vle32_v_f32m1(tmp[m][4], vl);
-                    vfloat32m1_t _r05 = vle32_v_f32m1(tmp[m][5], vl);
-
-                    vfloat32m1_t _tmp01a = vfmacc_vf_f32m1(vfmul_vf_f32m1(_r01, sq2, vl), -sq2_d2, _r03, vl);
-                    vfloat32m1_t _tmp01b = vfmacc_vf_f32m1(_r04, -2.f, _r02, vl);
-                    vfloat32m1_t _tmp23a = vfmacc_vf_f32m1(vfmul_vf_f32m1(_r01, sq2_d2, vl), -sq2, _r03, vl);
-                    vfloat32m1_t _tmp23b = vfmacc_vf_f32m1(_r04, -0.5f, _r02, vl);
-
-                    vfloat32m1_t _tmp0m = vfmacc_vf_f32m1(vfadd_vv_f32m1(_r00, _r04, vl), -2.5f, _r02, vl);
-                    vfloat32m1_t _tmp1m = vfsub_vv_f32m1(_tmp01b, _tmp01a, vl);
-                    vfloat32m1_t _tmp2m = vfadd_vv_f32m1(_tmp01b, _tmp01a, vl);
-                    vfloat32m1_t _tmp3m = vfsub_vv_f32m1(_tmp23b, _tmp23a, vl);
-                    vfloat32m1_t _tmp4m = vfadd_vv_f32m1(_tmp23b, _tmp23a, vl);
-                    vfloat32m1_t _tmp5m = vfmacc_vf_f32m1(vfadd_vv_f32m1(_r01, _r05, vl), -2.5f, _r03, vl);
-
-                    vse32_v_f32m1(r0_tm_0, _tmp0m, vl);
-                    vse32_v_f32m1(r0_tm_1, _tmp1m, vl);
-                    vse32_v_f32m1(r0_tm_2, _tmp2m, vl);
-                    vse32_v_f32m1(r0_tm_3, _tmp3m, vl);
-                    vse32_v_f32m1(r0_tm_4, _tmp4m, vl);
-                    vse32_v_f32m1(r0_tm_5, _tmp5m, vl);
-
-                    r0_tm_0 += tiles * packn * 6;
-                    r0_tm_1 += tiles * packn * 6;
-                    r0_tm_2 += tiles * packn * 6;
-                    r0_tm_3 += tiles * packn * 6;
-                    r0_tm_4 += tiles * packn * 6;
-                    r0_tm_5 += tiles * packn * 6;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_output_packn_rvv(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 4;
-    const int h_tiles = outh / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    const float* biasptr = bias;
-
-    const float sq2 = 1.41421356237;
-    const float sq2_m2 = 1.41421356237 * 2;
-    const float sq2_d2 = 1.41421356237 / 2;
-    const float sq2_d4 = 1.41421356237 / 4;
-
-    // const float otm[4][6] = {
-    //     {1.0f, 1.0f,   1.0f,  1.0f,  1.0f,   0.0f},
-    //     {0.0f, sq2/2, -sq2/2, sq2,   -sq2,   0.0f},
-    //     {0.0f, 0.5f,   0.5f,  2.0f,  2.0f,   0.0f},
-    //     {0.0f, sq2/4, -sq2/4, sq2*2, -sq2*2, 1.0f}
-    // };
-
-    // 0 = r00 + (r01 + r02) + (r03 + r04)
-    // 1 =       (r01 - r02) * sq2_d2 + (r03 - r04) * sq2
-    // 2 =       (r01 + r02) * 0.5f + (r03 + r04) * 2
-    // 3 = r05 + (r01 - r02) * sq2_d4 + (r03 - r04) * sq2_m2
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        vfloat32m1_t _bias0 = biasptr ? vle32_v_f32m1(biasptr + p * packn, vl) : vfmv_v_f_f32m1(0.f, vl);
-
-        // NOTE variable length array
-        float tmp[4][6][packn];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* output0_tm_0 = (const float*)out0_tm + (i * w_tiles + j) * packn;
-                const float* output0_tm_1 = output0_tm_0 + tiles * packn;
-                const float* output0_tm_2 = output0_tm_0 + tiles * packn * 2;
-                const float* output0_tm_3 = output0_tm_0 + tiles * packn * 3;
-                const float* output0_tm_4 = output0_tm_0 + tiles * packn * 4;
-                const float* output0_tm_5 = output0_tm_0 + tiles * packn * 5;
-
-                float* output0 = out0.row(i * 4) + (j * 4) * packn;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    vfloat32m1_t _r00 = vle32_v_f32m1(output0_tm_0, vl);
-                    vfloat32m1_t _r01 = vle32_v_f32m1(output0_tm_1, vl);
-                    vfloat32m1_t _r02 = vle32_v_f32m1(output0_tm_2, vl);
-                    vfloat32m1_t _r03 = vle32_v_f32m1(output0_tm_3, vl);
-                    vfloat32m1_t _r04 = vle32_v_f32m1(output0_tm_4, vl);
-                    vfloat32m1_t _r05 = vle32_v_f32m1(output0_tm_5, vl);
-
-                    vfloat32m1_t _tmp02a = vfadd_vv_f32m1(_r01, _r02, vl);
-                    vfloat32m1_t _tmp02b = vfadd_vv_f32m1(_r03, _r04, vl);
-                    vfloat32m1_t _tmp13a = vfsub_vv_f32m1(_r01, _r02, vl);
-                    vfloat32m1_t _tmp13b = vfsub_vv_f32m1(_r03, _r04, vl);
-
-                    vfloat32m1_t _tmp0m = vfadd_vv_f32m1(vfadd_vv_f32m1(_r00, _tmp02a, vl), _tmp02b, vl);
-                    vfloat32m1_t _tmp1m = vfmacc_vf_f32m1(vfmul_vf_f32m1(_tmp13a, sq2_d2, vl), sq2, _tmp13b, vl);
-                    vfloat32m1_t _tmp2m = vfmacc_vf_f32m1(vfmul_vf_f32m1(_tmp02a, 0.5f, vl), 2.f, _tmp02b, vl);
-                    vfloat32m1_t _tmp3m = vfmacc_vf_f32m1(vfmacc_vf_f32m1(_r05, sq2_d4, _tmp13a, vl), sq2_m2, _tmp13b, vl);
-
-                    vse32_v_f32m1(tmp[0][m], _tmp0m, vl);
-                    vse32_v_f32m1(tmp[1][m], _tmp1m, vl);
-                    vse32_v_f32m1(tmp[2][m], _tmp2m, vl);
-                    vse32_v_f32m1(tmp[3][m], _tmp3m, vl);
-
-                    output0_tm_0 += tiles * packn * 6;
-                    output0_tm_1 += tiles * packn * 6;
-                    output0_tm_2 += tiles * packn * 6;
-                    output0_tm_3 += tiles * packn * 6;
-                    output0_tm_4 += tiles * packn * 6;
-                    output0_tm_5 += tiles * packn * 6;
-                }
-
-                for (int m = 0; m < 4; m++)
-                {
-                    vfloat32m1_t _r00 = vle32_v_f32m1(tmp[m][0], vl);
-                    vfloat32m1_t _r01 = vle32_v_f32m1(tmp[m][1], vl);
-                    vfloat32m1_t _r02 = vle32_v_f32m1(tmp[m][2], vl);
-                    vfloat32m1_t _r03 = vle32_v_f32m1(tmp[m][3], vl);
-                    vfloat32m1_t _r04 = vle32_v_f32m1(tmp[m][4], vl);
-                    vfloat32m1_t _r05 = vle32_v_f32m1(tmp[m][5], vl);
-
-                    vfloat32m1_t _tmp02a = vfadd_vv_f32m1(_r01, _r02, vl);
-                    vfloat32m1_t _tmp02b = vfadd_vv_f32m1(_r03, _r04, vl);
-                    vfloat32m1_t _tmp13a = vfsub_vv_f32m1(_r01, _r02, vl);
-                    vfloat32m1_t _tmp13b = vfsub_vv_f32m1(_r03, _r04, vl);
-
-                    vfloat32m1_t _out00 = vfadd_vv_f32m1(_bias0, vfadd_vv_f32m1(vfadd_vv_f32m1(_r00, _tmp02a, vl), _tmp02b, vl), vl);
-                    vfloat32m1_t _out01 = vfadd_vv_f32m1(_bias0, vfmacc_vf_f32m1(vfmul_vf_f32m1(_tmp13a, sq2_d2, vl), sq2, _tmp13b, vl), vl);
-                    vfloat32m1_t _out02 = vfadd_vv_f32m1(_bias0, vfmacc_vf_f32m1(vfmul_vf_f32m1(_tmp02a, 0.5f, vl), 2.f, _tmp02b, vl), vl);
-                    vfloat32m1_t _out03 = vfadd_vv_f32m1(_bias0, vfmacc_vf_f32m1(vfmacc_vf_f32m1(_r05, sq2_d4, _tmp13a, vl), sq2_m2, _tmp13b, vl), vl);
-
-                    vse32_v_f32m1(output0, _out00, vl);
-                    vse32_v_f32m1(output0 + packn, _out01, vl);
-                    vse32_v_f32m1(output0 + packn * 2, _out02, vl);
-                    vse32_v_f32m1(output0 + packn * 3, _out03, vl);
-
-                    output0 += outw * packn;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_input_packn_rvv(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 2;
-    const int h_tiles = (h - 2) / 2;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[4][4] = {
-    //     {1.0f,  0.0f, -1.0f,  0.0f},
-    //     {0.0f,  1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  0.00f, 1.0f}
-    // };
-
-    // 0 = r00 - r02
-    // 1 = r01 + r02
-    // 2 = r02 - r01
-    // 3 = r03 - r01
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        // NOTE c99 variable length array
-        float tmp[4][4][packn];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* r0 = img0.row(i * 2) + (j * 2) * packn;
-
-                for (int m = 0; m < 4; m++)
-                {
-                    vfloat32m1_t _r00 = vle32_v_f32m1(r0, vl);
-                    vfloat32m1_t _r01 = vle32_v_f32m1(r0 + packn, vl);
-                    vfloat32m1_t _r02 = vle32_v_f32m1(r0 + packn * 2, vl);
-                    vfloat32m1_t _r03 = vle32_v_f32m1(r0 + packn * 3, vl);
-
-                    vfloat32m1_t _tmp0m = vfsub_vv_f32m1(_r00, _r02, vl);
-                    vfloat32m1_t _tmp1m = vfadd_vv_f32m1(_r01, _r02, vl);
-                    vfloat32m1_t _tmp2m = vfsub_vv_f32m1(_r02, _r01, vl);
-                    vfloat32m1_t _tmp3m = vfsub_vv_f32m1(_r03, _r01, vl);
-
-                    vse32_v_f32m1(tmp[0][m], _tmp0m, vl);
-                    vse32_v_f32m1(tmp[1][m], _tmp1m, vl);
-                    vse32_v_f32m1(tmp[2][m], _tmp2m, vl);
-                    vse32_v_f32m1(tmp[3][m], _tmp3m, vl);
-
-                    r0 += w * packn;
-                }
-
-                float* r0_tm_0 = (float*)img0_tm + (i * w_tiles + j) * packn;
-                float* r0_tm_1 = r0_tm_0 + tiles * packn;
-                float* r0_tm_2 = r0_tm_0 + tiles * packn * 2;
-                float* r0_tm_3 = r0_tm_0 + tiles * packn * 3;
-
-                for (int m = 0; m < 4; m++)
-                {
-                    vfloat32m1_t _tmp00 = vle32_v_f32m1(tmp[m][0], vl);
-                    vfloat32m1_t _tmp01 = vle32_v_f32m1(tmp[m][1], vl);
-                    vfloat32m1_t _tmp02 = vle32_v_f32m1(tmp[m][2], vl);
-                    vfloat32m1_t _tmp03 = vle32_v_f32m1(tmp[m][3], vl);
-
-                    vfloat32m1_t _r0tm0 = vfsub_vv_f32m1(_tmp00, _tmp02, vl);
-                    vfloat32m1_t _r0tm1 = vfadd_vv_f32m1(_tmp01, _tmp02, vl);
-                    vfloat32m1_t _r0tm2 = vfsub_vv_f32m1(_tmp02, _tmp01, vl);
-                    vfloat32m1_t _r0tm3 = vfsub_vv_f32m1(_tmp03, _tmp01, vl);
-
-                    vse32_v_f32m1(r0_tm_0, _r0tm0, vl);
-                    vse32_v_f32m1(r0_tm_1, _r0tm1, vl);
-                    vse32_v_f32m1(r0_tm_2, _r0tm2, vl);
-                    vse32_v_f32m1(r0_tm_3, _r0tm3, vl);
-
-                    r0_tm_0 += tiles * packn * 4;
-                    r0_tm_1 += tiles * packn * 4;
-                    r0_tm_2 += tiles * packn * 4;
-                    r0_tm_3 += tiles * packn * 4;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_output_packn_rvv(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 2;
-    const int h_tiles = outh / 2;
-    const int tiles = w_tiles * h_tiles;
-
-    const float* biasptr = bias;
-
-    // const float otm[2][4] = {
-    //     {1.0f,  1.0f,  1.0f,  0.0f},
-    //     {0.0f,  1.0f, -1.0f,  1.0f}
-    // };
-
-    // 0 = r00 + r01 + r02
-    // 1 = r01 - r02 + r03
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        vfloat32m1_t _bias0 = biasptr ? vle32_v_f32m1(biasptr + p * packn, vl) : vfmv_v_f_f32m1(0.f, vl);
-
-        // NOTE variable length array
-        float tmp[2][4][packn];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const float* output0_tm_0 = (const float*)out0_tm + (i * w_tiles + j) * packn;
-                const float* output0_tm_1 = output0_tm_0 + tiles * packn;
-                const float* output0_tm_2 = output0_tm_0 + tiles * packn * 2;
-                const float* output0_tm_3 = output0_tm_0 + tiles * packn * 3;
-
-                float* output0 = out0.row(i * 2) + (j * 2) * packn;
-
-                for (int m = 0; m < 4; m++)
-                {
-                    vfloat32m1_t _out0tm0 = vle32_v_f32m1(output0_tm_0, vl);
-                    vfloat32m1_t _out0tm1 = vle32_v_f32m1(output0_tm_1, vl);
-                    vfloat32m1_t _out0tm2 = vle32_v_f32m1(output0_tm_2, vl);
-                    vfloat32m1_t _out0tm3 = vle32_v_f32m1(output0_tm_3, vl);
-
-                    vfloat32m1_t _tmp0m = vfadd_vv_f32m1(vfadd_vv_f32m1(_out0tm0, _out0tm1, vl), _out0tm2, vl);
-                    vfloat32m1_t _tmp1m = vfadd_vv_f32m1(vfsub_vv_f32m1(_out0tm1, _out0tm2, vl), _out0tm3, vl);
-
-                    vse32_v_f32m1(tmp[0][m], _tmp0m, vl);
-                    vse32_v_f32m1(tmp[1][m], _tmp1m, vl);
-
-                    output0_tm_0 += tiles * packn * 4;
-                    output0_tm_1 += tiles * packn * 4;
-                    output0_tm_2 += tiles * packn * 4;
-                    output0_tm_3 += tiles * packn * 4;
-                }
-
-                for (int m = 0; m < 2; m++)
-                {
-                    vfloat32m1_t _tmp00 = vle32_v_f32m1(tmp[m][0], vl);
-                    vfloat32m1_t _tmp01 = vle32_v_f32m1(tmp[m][1], vl);
-                    vfloat32m1_t _tmp02 = vle32_v_f32m1(tmp[m][2], vl);
-                    vfloat32m1_t _tmp03 = vle32_v_f32m1(tmp[m][3], vl);
-
-                    vfloat32m1_t _out00 = vfadd_vv_f32m1(_bias0, vfadd_vv_f32m1(vfadd_vv_f32m1(_tmp00, _tmp01, vl), _tmp02, vl), vl);
-                    vfloat32m1_t _out01 = vfadd_vv_f32m1(_bias0, vfadd_vv_f32m1(vfsub_vv_f32m1(_tmp01, _tmp02, vl), _tmp03, vl), vl);
-
-                    vse32_v_f32m1(output0, _out00, vl);
-                    vse32_v_f32m1(output0 + packn, _out01, vl);
-
-                    output0 += outw * packn;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_transform_packn_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_transform_packn_fp16s.h
deleted file mode 100644
index 145ca220..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolution_winograd_transform_packn_fp16s.h
+++ /dev/null
@@ -1,745 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_winograd63_transform_input_packn_fp16sa_rvv(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 6;
-    const int h_tiles = (h - 2) / 6;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[8][8] = {
-    //     {1.0f,  0.0f, -5.25f,  0.00f,  5.25f,  0.00f, -1.0f, 0.0f},
-    //
-    //     {0.0f,  1.0f,  1.00f, -4.25f, -4.25f,  1.00f,  1.0f, 0.0f},
-    //     {0.0f, -1.0f,  1.00f,  4.25f, -4.25f, -1.00f,  1.0f, 0.0f},
-    //
-    //     {0.0f,  0.5f,  0.25f, -2.50f, -1.25f,  2.00f,  1.0f, 0.0f},
-    //     {0.0f, -0.5f,  0.25f,  2.50f, -1.25f, -2.00f,  1.0f, 0.0f},
-    //
-    //     {0.0f,  2.0f,  4.00f, -2.50f, -5.00f,  0.50f,  1.0f, 0.0f},
-    //     {0.0f, -2.0f,  4.00f,  2.50f, -5.00f, -0.50f,  1.0f, 0.0f},
-    //
-    //     {0.0f, -1.0f,  0.00f,  5.25f,  0.00f, -5.25f,  0.0f, 1.0f}
-    // };
-
-    // 0 = r00 - r06 + (r04 - r02) * 5.25
-    // 7 = r07 - r01 + (r03 - r05) * 5.25
-
-    // 1 = (r02 + r06 - r04 * 4.25) + (r01 - r03 * 4.25 + r05)
-    // 2 = (r02 + r06 - r04 * 4.25) - (r01 - r03 * 4.25 + r05)
-
-    // 3 = (r06 + r02 * 0.25 - r04 * 1.25) + (r01 * 0.5 - r03 * 2.5 + r05 * 2)
-    // 4 = (r06 + r02 * 0.25 - r04 * 1.25) - (r01 * 0.5 - r03 * 2.5 + r05 * 2)
-
-    // reuse r04 * 1.25
-    // reuse r03 * 2.5
-    // 5 = (r06 + (r02 - r04 * 1.25) * 4) + (r01 * 2 - r03 * 2.5 + r05 * 0.5)
-    // 6 = (r06 + (r02 - r04 * 1.25) * 4) - (r01 * 2 - r03 * 2.5 + r05 * 0.5)
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        // NOTE c99 variable length array
-        __fp16 tmp[8][8][packn];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const __fp16* r0 = img0.row<const __fp16>(i * 6) + (j * 6) * packn;
-
-                for (int m = 0; m < 8; m++)
-                {
-                    vfloat16m1_t _r00 = vle16_v_f16m1(r0, vl);
-                    vfloat16m1_t _r01 = vle16_v_f16m1(r0 + packn, vl);
-                    vfloat16m1_t _r02 = vle16_v_f16m1(r0 + packn * 2, vl);
-                    vfloat16m1_t _r03 = vle16_v_f16m1(r0 + packn * 3, vl);
-                    vfloat16m1_t _r04 = vle16_v_f16m1(r0 + packn * 4, vl);
-                    vfloat16m1_t _r05 = vle16_v_f16m1(r0 + packn * 5, vl);
-                    vfloat16m1_t _r06 = vle16_v_f16m1(r0 + packn * 6, vl);
-                    vfloat16m1_t _r07 = vle16_v_f16m1(r0 + packn * 7, vl);
-
-                    vfloat16m1_t _tmp0m = vfmacc_vf_f16m1(vfsub_vv_f16m1(_r00, _r06, vl), 5.25f, vfsub_vv_f16m1(_r04, _r02, vl), vl);
-                    vfloat16m1_t _tmp7m = vfmacc_vf_f16m1(vfsub_vv_f16m1(_r07, _r01, vl), 5.25f, vfsub_vv_f16m1(_r03, _r05, vl), vl);
-                    vse16_v_f16m1(tmp[0][m], _tmp0m, vl);
-                    vse16_v_f16m1(tmp[7][m], _tmp7m, vl);
-
-                    vfloat16m1_t _tmp12a = vfmacc_vf_f16m1(vfadd_vv_f16m1(_r02, _r06, vl), -4.25f, _r04, vl);
-                    vfloat16m1_t _tmp12b = vfmacc_vf_f16m1(vfadd_vv_f16m1(_r01, _r05, vl), -4.25f, _r03, vl);
-
-                    vfloat16m1_t _tmp1m = vfadd_vv_f16m1(_tmp12a, _tmp12b, vl);
-                    vfloat16m1_t _tmp2m = vfsub_vv_f16m1(_tmp12a, _tmp12b, vl);
-                    vse16_v_f16m1(tmp[1][m], _tmp1m, vl);
-                    vse16_v_f16m1(tmp[2][m], _tmp2m, vl);
-
-                    vfloat16m1_t _tmp34a = vfmacc_vf_f16m1(vfmacc_vf_f16m1(_r06, 0.25f, _r02, vl), -1.25f, _r04, vl);
-                    vfloat16m1_t _tmp34b = vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmul_vf_f16m1(_r01, 0.5f, vl), -2.5f, _r03, vl), 2.f, _r05, vl);
-
-                    vfloat16m1_t _tmp3m = vfadd_vv_f16m1(_tmp34a, _tmp34b, vl);
-                    vfloat16m1_t _tmp4m = vfsub_vv_f16m1(_tmp34a, _tmp34b, vl);
-                    vse16_v_f16m1(tmp[3][m], _tmp3m, vl);
-                    vse16_v_f16m1(tmp[4][m], _tmp4m, vl);
-
-                    vfloat16m1_t _tmp56a = vfmacc_vf_f16m1(_r06, 4.f, vfmacc_vf_f16m1(_r02, -1.25f, _r04, vl), vl);
-                    vfloat16m1_t _tmp56b = vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmul_vf_f16m1(_r01, 2.f, vl), -2.5f, _r03, vl), 0.5f, _r05, vl);
-
-                    vfloat16m1_t _tmp5m = vfadd_vv_f16m1(_tmp56a, _tmp56b, vl);
-                    vfloat16m1_t _tmp6m = vfsub_vv_f16m1(_tmp56a, _tmp56b, vl);
-                    vse16_v_f16m1(tmp[5][m], _tmp5m, vl);
-                    vse16_v_f16m1(tmp[6][m], _tmp6m, vl);
-
-                    r0 += w * packn;
-                }
-
-                __fp16* r0_tm_0 = (__fp16*)img0_tm + (i * w_tiles + j) * packn;
-                __fp16* r0_tm_1 = r0_tm_0 + tiles * packn;
-                __fp16* r0_tm_2 = r0_tm_0 + tiles * packn * 2;
-                __fp16* r0_tm_3 = r0_tm_0 + tiles * packn * 3;
-                __fp16* r0_tm_4 = r0_tm_0 + tiles * packn * 4;
-                __fp16* r0_tm_5 = r0_tm_0 + tiles * packn * 5;
-                __fp16* r0_tm_6 = r0_tm_0 + tiles * packn * 6;
-                __fp16* r0_tm_7 = r0_tm_0 + tiles * packn * 7;
-
-                for (int m = 0; m < 8; m++)
-                {
-                    vfloat16m1_t _tmp00 = vle16_v_f16m1(tmp[m][0], vl);
-                    vfloat16m1_t _tmp01 = vle16_v_f16m1(tmp[m][1], vl);
-                    vfloat16m1_t _tmp02 = vle16_v_f16m1(tmp[m][2], vl);
-                    vfloat16m1_t _tmp03 = vle16_v_f16m1(tmp[m][3], vl);
-                    vfloat16m1_t _tmp04 = vle16_v_f16m1(tmp[m][4], vl);
-                    vfloat16m1_t _tmp05 = vle16_v_f16m1(tmp[m][5], vl);
-                    vfloat16m1_t _tmp06 = vle16_v_f16m1(tmp[m][6], vl);
-                    vfloat16m1_t _tmp07 = vle16_v_f16m1(tmp[m][7], vl);
-
-                    vfloat16m1_t _r0tm0 = vfmacc_vf_f16m1(vfsub_vv_f16m1(_tmp00, _tmp06, vl), 5.25f, vfsub_vv_f16m1(_tmp04, _tmp02, vl), vl);
-                    vfloat16m1_t _r0tm7 = vfmacc_vf_f16m1(vfsub_vv_f16m1(_tmp07, _tmp01, vl), 5.25f, vfsub_vv_f16m1(_tmp03, _tmp05, vl), vl);
-
-                    vfloat16m1_t _tmp12a = vfmacc_vf_f16m1(vfadd_vv_f16m1(_tmp02, _tmp06, vl), -4.25f, _tmp04, vl);
-                    vfloat16m1_t _tmp12b = vfmacc_vf_f16m1(vfadd_vv_f16m1(_tmp01, _tmp05, vl), -4.25f, _tmp03, vl);
-
-                    vfloat16m1_t _r0tm1 = vfadd_vv_f16m1(_tmp12a, _tmp12b, vl);
-                    vfloat16m1_t _r0tm2 = vfsub_vv_f16m1(_tmp12a, _tmp12b, vl);
-
-                    vfloat16m1_t _tmp34a = vfmacc_vf_f16m1(vfmacc_vf_f16m1(_tmp06, 0.25f, _tmp02, vl), -1.25f, _tmp04, vl);
-                    vfloat16m1_t _tmp34b = vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmul_vf_f16m1(_tmp01, 0.5f, vl), -2.5f, _tmp03, vl), 2.f, _tmp05, vl);
-
-                    vfloat16m1_t _r0tm3 = vfadd_vv_f16m1(_tmp34a, _tmp34b, vl);
-                    vfloat16m1_t _r0tm4 = vfsub_vv_f16m1(_tmp34a, _tmp34b, vl);
-
-                    vfloat16m1_t _tmp56a = vfmacc_vf_f16m1(_tmp06, 4.f, vfmacc_vf_f16m1(_tmp02, -1.25f, _tmp04, vl), vl);
-                    vfloat16m1_t _tmp56b = vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmul_vf_f16m1(_tmp01, 2.f, vl), -2.5f, _tmp03, vl), 0.5f, _tmp05, vl);
-
-                    vfloat16m1_t _r0tm5 = vfadd_vv_f16m1(_tmp56a, _tmp56b, vl);
-                    vfloat16m1_t _r0tm6 = vfsub_vv_f16m1(_tmp56a, _tmp56b, vl);
-
-                    vse16_v_f16m1(r0_tm_0, _r0tm0, vl);
-                    vse16_v_f16m1(r0_tm_1, _r0tm1, vl);
-                    vse16_v_f16m1(r0_tm_2, _r0tm2, vl);
-                    vse16_v_f16m1(r0_tm_3, _r0tm3, vl);
-                    vse16_v_f16m1(r0_tm_4, _r0tm4, vl);
-                    vse16_v_f16m1(r0_tm_5, _r0tm5, vl);
-                    vse16_v_f16m1(r0_tm_6, _r0tm6, vl);
-                    vse16_v_f16m1(r0_tm_7, _r0tm7, vl);
-
-                    r0_tm_0 += tiles * packn * 8;
-                    r0_tm_1 += tiles * packn * 8;
-                    r0_tm_2 += tiles * packn * 8;
-                    r0_tm_3 += tiles * packn * 8;
-                    r0_tm_4 += tiles * packn * 8;
-                    r0_tm_5 += tiles * packn * 8;
-                    r0_tm_6 += tiles * packn * 8;
-                    r0_tm_7 += tiles * packn * 8;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd63_transform_output_packn_fp16sa_rvv(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 6;
-    const int h_tiles = outh / 6;
-    const int tiles = w_tiles * h_tiles;
-
-    const __fp16* biasptr = bias;
-
-    // const float otm[6][8] = {
-    //     {1.0f,  1.0f,   1.0f,   1.0f,   1.0f,  32.0f, 32.0f, 0.0f},
-    //     {0.0f,  1.0f,  -1.0f,   2.0f,  -2.0f,  16.0f,-16.0f, 0.0f},
-    //     {0.0f,  1.0f,   1.0f,   4.0f,   4.0f,   8.0f,  8.0f, 0.0f},
-    //     {0.0f,  1.0f,  -1.0f,   8.0f,  -8.0f,   4.0f, -4.0f, 0.0f},
-    //     {0.0f,  1.0f,   1.0f,  16.0f,  16.0f,   2.0f,  2.0f, 0.0f},
-    //     {0.0f,  1.0f,  -1.0f,  32.0f, -32.0f,   1.0f, -1.0f, 1.0f}
-    // };
-
-    // 0 = r0 + (r1 + r2) + (r3 + r4)     + (r5 + r6) * 32
-    // 1 =      (r1 - r2) + (r3 - r4) * 2 + (r5 - r6) * 16
-    // 2 =      (r1 + r2) + (r3 + r4) * 4 + (r5 + r6) * 8
-    // 3 =      (r1 - r2) + (r3 - r4) * 8 + (r5 - r6) * 4
-    // 4 =      (r1 + r2) + (r3 + r4) * 16+ (r5 + r6) * 2
-    // 5 = r7 + (r1 - r2) + (r3 - r4) * 32+ (r5 - r6)
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        vfloat16m1_t _bias0 = biasptr ? vle16_v_f16m1(biasptr + p * packn, vl) : vfmv_v_f_f16m1(0.f, vl);
-
-        // NOTE c99 variable length array
-        __fp16 tmp[6][8][packn];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const __fp16* output0_tm_0 = (const __fp16*)out0_tm + (i * w_tiles + j) * packn;
-                const __fp16* output0_tm_1 = output0_tm_0 + tiles * packn;
-                const __fp16* output0_tm_2 = output0_tm_0 + tiles * packn * 2;
-                const __fp16* output0_tm_3 = output0_tm_0 + tiles * packn * 3;
-                const __fp16* output0_tm_4 = output0_tm_0 + tiles * packn * 4;
-                const __fp16* output0_tm_5 = output0_tm_0 + tiles * packn * 5;
-                const __fp16* output0_tm_6 = output0_tm_0 + tiles * packn * 6;
-                const __fp16* output0_tm_7 = output0_tm_0 + tiles * packn * 7;
-
-                __fp16* output0 = out0.row<__fp16>(i * 6) + (j * 6) * packn;
-
-                for (int m = 0; m < 8; m++)
-                {
-                    vfloat16m1_t _out0tm0 = vle16_v_f16m1(output0_tm_0, vl);
-                    vfloat16m1_t _out0tm1 = vle16_v_f16m1(output0_tm_1, vl);
-                    vfloat16m1_t _out0tm2 = vle16_v_f16m1(output0_tm_2, vl);
-                    vfloat16m1_t _out0tm3 = vle16_v_f16m1(output0_tm_3, vl);
-                    vfloat16m1_t _out0tm4 = vle16_v_f16m1(output0_tm_4, vl);
-                    vfloat16m1_t _out0tm5 = vle16_v_f16m1(output0_tm_5, vl);
-                    vfloat16m1_t _out0tm6 = vle16_v_f16m1(output0_tm_6, vl);
-                    vfloat16m1_t _out0tm7 = vle16_v_f16m1(output0_tm_7, vl);
-
-                    vfloat16m1_t _tmp024a = vfadd_vv_f16m1(_out0tm1, _out0tm2, vl);
-                    vfloat16m1_t _tmp135a = vfsub_vv_f16m1(_out0tm1, _out0tm2, vl);
-
-                    vfloat16m1_t _tmp024b = vfadd_vv_f16m1(_out0tm3, _out0tm4, vl);
-                    vfloat16m1_t _tmp135b = vfsub_vv_f16m1(_out0tm3, _out0tm4, vl);
-
-                    vfloat16m1_t _tmp024c = vfadd_vv_f16m1(_out0tm5, _out0tm6, vl);
-                    vfloat16m1_t _tmp135c = vfsub_vv_f16m1(_out0tm5, _out0tm6, vl);
-
-                    vfloat16m1_t _tmp0m = vfadd_vv_f16m1(vfadd_vv_f16m1(_out0tm0, _tmp024a, vl), vfmacc_vf_f16m1(_tmp024b, 32.f, _tmp024c, vl), vl);
-                    vfloat16m1_t _tmp2m = vfmacc_vf_f16m1(vfmacc_vf_f16m1(_tmp024a, 4.f, _tmp024b, vl), 8.f, _tmp024c, vl);
-                    vfloat16m1_t _tmp4m = vfmacc_vf_f16m1(vfmacc_vf_f16m1(_tmp024a, 16.f, _tmp024b, vl), 2.f, _tmp024c, vl);
-                    vse16_v_f16m1(tmp[0][m], _tmp0m, vl);
-                    vse16_v_f16m1(tmp[2][m], _tmp2m, vl);
-                    vse16_v_f16m1(tmp[4][m], _tmp4m, vl);
-
-                    vfloat16m1_t _tmp1m = vfmacc_vf_f16m1(vfmacc_vf_f16m1(_tmp135a, 2.f, _tmp135b, vl), 16.f, _tmp135c, vl);
-                    vfloat16m1_t _tmp3m = vfmacc_vf_f16m1(vfmacc_vf_f16m1(_tmp135a, 8.f, _tmp135b, vl), 4.f, _tmp135c, vl);
-                    vfloat16m1_t _tmp5m = vfadd_vv_f16m1(vfadd_vv_f16m1(_out0tm7, _tmp135a, vl), vfmacc_vf_f16m1(_tmp135c, 32.f, _tmp135b, vl), vl);
-                    vse16_v_f16m1(tmp[1][m], _tmp1m, vl);
-                    vse16_v_f16m1(tmp[3][m], _tmp3m, vl);
-                    vse16_v_f16m1(tmp[5][m], _tmp5m, vl);
-
-                    output0_tm_0 += tiles * packn * 8;
-                    output0_tm_1 += tiles * packn * 8;
-                    output0_tm_2 += tiles * packn * 8;
-                    output0_tm_3 += tiles * packn * 8;
-                    output0_tm_4 += tiles * packn * 8;
-                    output0_tm_5 += tiles * packn * 8;
-                    output0_tm_6 += tiles * packn * 8;
-                    output0_tm_7 += tiles * packn * 8;
-                }
-
-                for (int m = 0; m < 6; m++)
-                {
-                    vfloat16m1_t _tmp00 = vle16_v_f16m1(tmp[m][0], vl);
-                    vfloat16m1_t _tmp01 = vle16_v_f16m1(tmp[m][1], vl);
-                    vfloat16m1_t _tmp02 = vle16_v_f16m1(tmp[m][2], vl);
-                    vfloat16m1_t _tmp03 = vle16_v_f16m1(tmp[m][3], vl);
-                    vfloat16m1_t _tmp04 = vle16_v_f16m1(tmp[m][4], vl);
-                    vfloat16m1_t _tmp05 = vle16_v_f16m1(tmp[m][5], vl);
-                    vfloat16m1_t _tmp06 = vle16_v_f16m1(tmp[m][6], vl);
-                    vfloat16m1_t _tmp07 = vle16_v_f16m1(tmp[m][7], vl);
-
-                    vfloat16m1_t _tmp024a = vfadd_vv_f16m1(_tmp01, _tmp02, vl);
-                    vfloat16m1_t _tmp135a = vfsub_vv_f16m1(_tmp01, _tmp02, vl);
-
-                    vfloat16m1_t _tmp024b = vfadd_vv_f16m1(_tmp03, _tmp04, vl);
-                    vfloat16m1_t _tmp135b = vfsub_vv_f16m1(_tmp03, _tmp04, vl);
-
-                    vfloat16m1_t _tmp024c = vfadd_vv_f16m1(_tmp05, _tmp06, vl);
-                    vfloat16m1_t _tmp135c = vfsub_vv_f16m1(_tmp05, _tmp06, vl);
-
-                    vfloat16m1_t _out00 = vfadd_vv_f16m1(_bias0, vfadd_vv_f16m1(vfadd_vv_f16m1(_tmp00, _tmp024a, vl), vfmacc_vf_f16m1(_tmp024b, 32.f, _tmp024c, vl), vl), vl);
-                    vfloat16m1_t _out02 = vfadd_vv_f16m1(_bias0, vfmacc_vf_f16m1(vfmacc_vf_f16m1(_tmp024a, 4.f, _tmp024b, vl), 8.f, _tmp024c, vl), vl);
-                    vfloat16m1_t _out04 = vfadd_vv_f16m1(_bias0, vfmacc_vf_f16m1(vfmacc_vf_f16m1(_tmp024a, 16.f, _tmp024b, vl), 2.f, _tmp024c, vl), vl);
-                    vse16_v_f16m1(output0, _out00, vl);
-                    vse16_v_f16m1(output0 + packn * 2, _out02, vl);
-                    vse16_v_f16m1(output0 + packn * 4, _out04, vl);
-
-                    vfloat16m1_t _out01 = vfadd_vv_f16m1(_bias0, vfmacc_vf_f16m1(vfmacc_vf_f16m1(_tmp135a, 2.f, _tmp135b, vl), 16.f, _tmp135c, vl), vl);
-                    vfloat16m1_t _out03 = vfadd_vv_f16m1(_bias0, vfmacc_vf_f16m1(vfmacc_vf_f16m1(_tmp135a, 8.f, _tmp135b, vl), 4.f, _tmp135c, vl), vl);
-                    vfloat16m1_t _out05 = vfadd_vv_f16m1(_bias0, vfadd_vv_f16m1(vfadd_vv_f16m1(_tmp07, _tmp135a, vl), vfmacc_vf_f16m1(_tmp135c, 32.f, _tmp135b, vl), vl), vl);
-                    vse16_v_f16m1(output0 + packn, _out01, vl);
-                    vse16_v_f16m1(output0 + packn * 3, _out03, vl);
-                    vse16_v_f16m1(output0 + packn * 5, _out05, vl);
-
-                    output0 += outw * packn;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_input_packn_fp16sa_rvv(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 4;
-    const int h_tiles = (h - 2) / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    const float sq2 = 1.41421356237;
-    const float sq2_d2 = 1.41421356237 / 2;
-
-    // const float itm[6][6] = {
-    //     {1.0f,  0.0f,  -2.5f,  0.0f,  1.0f, 0.0f},
-    //     {0.0f, -sq2,   -2.0f,  sq2/2, 1.0f, 0.0f},
-    //     {0.0f,  sq2,   -2.0f, -sq2/2, 1.0f, 0.0f},
-    //     {0.0f, -sq2/2, -0.5f,  sq2,   1.0f, 0.0f},
-    //     {0.0f,  sq2/2, -0.5f, -sq2,   1.0f, 0.0f},
-    //     {0.0f,  1.0f,   0.0f,  -2.5f, 0.0f, 1.0f}
-    // };
-
-    // 0 =  r00 - 2.5f * r02 + r04
-    // 1 = -(sq2 * r01 - sq2_d2 * r03) + (r04 - 2 * r02)
-    // 2 =  (sq2 * r01 - sq2_d2 * r03) + (r04 - 2 * r02)
-    // 3 = -(sq2_d2 * r01 - sq2 * r03) + (r04 - 0.5f * r02)
-    // 4 =  (sq2_d2 * r01 - sq2 * r03) + (r04 - 0.5f * r02)
-    // 5 =  r01 - 2.5f * r03 + r05
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        // NOTE c99 variable length array
-        __fp16 tmp[6][6][packn];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const __fp16* r0 = img0.row<const __fp16>(i * 4) + (j * 4) * packn;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    vfloat16m1_t _r00 = vle16_v_f16m1(r0, vl);
-                    vfloat16m1_t _r01 = vle16_v_f16m1(r0 + packn, vl);
-                    vfloat16m1_t _r02 = vle16_v_f16m1(r0 + packn * 2, vl);
-                    vfloat16m1_t _r03 = vle16_v_f16m1(r0 + packn * 3, vl);
-                    vfloat16m1_t _r04 = vle16_v_f16m1(r0 + packn * 4, vl);
-                    vfloat16m1_t _r05 = vle16_v_f16m1(r0 + packn * 5, vl);
-
-                    vfloat16m1_t _tmp01a = vfmacc_vf_f16m1(vfmul_vf_f16m1(_r01, sq2, vl), -sq2_d2, _r03, vl);
-                    vfloat16m1_t _tmp01b = vfmacc_vf_f16m1(_r04, -2.f, _r02, vl);
-                    vfloat16m1_t _tmp23a = vfmacc_vf_f16m1(vfmul_vf_f16m1(_r01, sq2_d2, vl), -sq2, _r03, vl);
-                    vfloat16m1_t _tmp23b = vfmacc_vf_f16m1(_r04, -0.5f, _r02, vl);
-
-                    vfloat16m1_t _tmp0m = vfmacc_vf_f16m1(vfadd_vv_f16m1(_r00, _r04, vl), -2.5f, _r02, vl);
-                    vfloat16m1_t _tmp1m = vfsub_vv_f16m1(_tmp01b, _tmp01a, vl);
-                    vfloat16m1_t _tmp2m = vfadd_vv_f16m1(_tmp01b, _tmp01a, vl);
-                    vfloat16m1_t _tmp3m = vfsub_vv_f16m1(_tmp23b, _tmp23a, vl);
-                    vfloat16m1_t _tmp4m = vfadd_vv_f16m1(_tmp23b, _tmp23a, vl);
-                    vfloat16m1_t _tmp5m = vfmacc_vf_f16m1(vfadd_vv_f16m1(_r01, _r05, vl), -2.5f, _r03, vl);
-
-                    vse16_v_f16m1(tmp[0][m], _tmp0m, vl);
-                    vse16_v_f16m1(tmp[1][m], _tmp1m, vl);
-                    vse16_v_f16m1(tmp[2][m], _tmp2m, vl);
-                    vse16_v_f16m1(tmp[3][m], _tmp3m, vl);
-                    vse16_v_f16m1(tmp[4][m], _tmp4m, vl);
-                    vse16_v_f16m1(tmp[5][m], _tmp5m, vl);
-
-                    r0 += w * packn;
-                }
-
-                __fp16* r0_tm_0 = (__fp16*)img0_tm + (i * w_tiles + j) * packn;
-                __fp16* r0_tm_1 = r0_tm_0 + tiles * packn;
-                __fp16* r0_tm_2 = r0_tm_0 + tiles * packn * 2;
-                __fp16* r0_tm_3 = r0_tm_0 + tiles * packn * 3;
-                __fp16* r0_tm_4 = r0_tm_0 + tiles * packn * 4;
-                __fp16* r0_tm_5 = r0_tm_0 + tiles * packn * 5;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    vfloat16m1_t _r00 = vle16_v_f16m1(tmp[m][0], vl);
-                    vfloat16m1_t _r01 = vle16_v_f16m1(tmp[m][1], vl);
-                    vfloat16m1_t _r02 = vle16_v_f16m1(tmp[m][2], vl);
-                    vfloat16m1_t _r03 = vle16_v_f16m1(tmp[m][3], vl);
-                    vfloat16m1_t _r04 = vle16_v_f16m1(tmp[m][4], vl);
-                    vfloat16m1_t _r05 = vle16_v_f16m1(tmp[m][5], vl);
-
-                    vfloat16m1_t _tmp01a = vfmacc_vf_f16m1(vfmul_vf_f16m1(_r01, sq2, vl), -sq2_d2, _r03, vl);
-                    vfloat16m1_t _tmp01b = vfmacc_vf_f16m1(_r04, -2.f, _r02, vl);
-                    vfloat16m1_t _tmp23a = vfmacc_vf_f16m1(vfmul_vf_f16m1(_r01, sq2_d2, vl), -sq2, _r03, vl);
-                    vfloat16m1_t _tmp23b = vfmacc_vf_f16m1(_r04, -0.5f, _r02, vl);
-
-                    vfloat16m1_t _tmp0m = vfmacc_vf_f16m1(vfadd_vv_f16m1(_r00, _r04, vl), -2.5f, _r02, vl);
-                    vfloat16m1_t _tmp1m = vfsub_vv_f16m1(_tmp01b, _tmp01a, vl);
-                    vfloat16m1_t _tmp2m = vfadd_vv_f16m1(_tmp01b, _tmp01a, vl);
-                    vfloat16m1_t _tmp3m = vfsub_vv_f16m1(_tmp23b, _tmp23a, vl);
-                    vfloat16m1_t _tmp4m = vfadd_vv_f16m1(_tmp23b, _tmp23a, vl);
-                    vfloat16m1_t _tmp5m = vfmacc_vf_f16m1(vfadd_vv_f16m1(_r01, _r05, vl), -2.5f, _r03, vl);
-
-                    vse16_v_f16m1(r0_tm_0, _tmp0m, vl);
-                    vse16_v_f16m1(r0_tm_1, _tmp1m, vl);
-                    vse16_v_f16m1(r0_tm_2, _tmp2m, vl);
-                    vse16_v_f16m1(r0_tm_3, _tmp3m, vl);
-                    vse16_v_f16m1(r0_tm_4, _tmp4m, vl);
-                    vse16_v_f16m1(r0_tm_5, _tmp5m, vl);
-
-                    r0_tm_0 += tiles * packn * 6;
-                    r0_tm_1 += tiles * packn * 6;
-                    r0_tm_2 += tiles * packn * 6;
-                    r0_tm_3 += tiles * packn * 6;
-                    r0_tm_4 += tiles * packn * 6;
-                    r0_tm_5 += tiles * packn * 6;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_output_packn_fp16sa_rvv(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 4;
-    const int h_tiles = outh / 4;
-    const int tiles = w_tiles * h_tiles;
-
-    const __fp16* biasptr = bias;
-
-    const float sq2 = 1.41421356237;
-    const float sq2_m2 = 1.41421356237 * 2;
-    const float sq2_d2 = 1.41421356237 / 2;
-    const float sq2_d4 = 1.41421356237 / 4;
-
-    // const float otm[4][6] = {
-    //     {1.0f, 1.0f,   1.0f,  1.0f,  1.0f,   0.0f},
-    //     {0.0f, sq2/2, -sq2/2, sq2,   -sq2,   0.0f},
-    //     {0.0f, 0.5f,   0.5f,  2.0f,  2.0f,   0.0f},
-    //     {0.0f, sq2/4, -sq2/4, sq2*2, -sq2*2, 1.0f}
-    // };
-
-    // 0 = r00 + (r01 + r02) + (r03 + r04)
-    // 1 =       (r01 - r02) * sq2_d2 + (r03 - r04) * sq2
-    // 2 =       (r01 + r02) * 0.5f + (r03 + r04) * 2
-    // 3 = r05 + (r01 - r02) * sq2_d4 + (r03 - r04) * sq2_m2
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        vfloat16m1_t _bias0 = biasptr ? vle16_v_f16m1(biasptr + p * packn, vl) : vfmv_v_f_f16m1(0.f, vl);
-
-        // NOTE variable length array
-        __fp16 tmp[4][6][packn];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const __fp16* output0_tm_0 = (const __fp16*)out0_tm + (i * w_tiles + j) * packn;
-                const __fp16* output0_tm_1 = output0_tm_0 + tiles * packn;
-                const __fp16* output0_tm_2 = output0_tm_0 + tiles * packn * 2;
-                const __fp16* output0_tm_3 = output0_tm_0 + tiles * packn * 3;
-                const __fp16* output0_tm_4 = output0_tm_0 + tiles * packn * 4;
-                const __fp16* output0_tm_5 = output0_tm_0 + tiles * packn * 5;
-
-                __fp16* output0 = out0.row<__fp16>(i * 4) + (j * 4) * packn;
-
-                for (int m = 0; m < 6; m++)
-                {
-                    vfloat16m1_t _r00 = vle16_v_f16m1(output0_tm_0, vl);
-                    vfloat16m1_t _r01 = vle16_v_f16m1(output0_tm_1, vl);
-                    vfloat16m1_t _r02 = vle16_v_f16m1(output0_tm_2, vl);
-                    vfloat16m1_t _r03 = vle16_v_f16m1(output0_tm_3, vl);
-                    vfloat16m1_t _r04 = vle16_v_f16m1(output0_tm_4, vl);
-                    vfloat16m1_t _r05 = vle16_v_f16m1(output0_tm_5, vl);
-
-                    vfloat16m1_t _tmp02a = vfadd_vv_f16m1(_r01, _r02, vl);
-                    vfloat16m1_t _tmp02b = vfadd_vv_f16m1(_r03, _r04, vl);
-                    vfloat16m1_t _tmp13a = vfsub_vv_f16m1(_r01, _r02, vl);
-                    vfloat16m1_t _tmp13b = vfsub_vv_f16m1(_r03, _r04, vl);
-
-                    vfloat16m1_t _tmp0m = vfadd_vv_f16m1(vfadd_vv_f16m1(_r00, _tmp02a, vl), _tmp02b, vl);
-                    vfloat16m1_t _tmp1m = vfmacc_vf_f16m1(vfmul_vf_f16m1(_tmp13a, sq2_d2, vl), sq2, _tmp13b, vl);
-                    vfloat16m1_t _tmp2m = vfmacc_vf_f16m1(vfmul_vf_f16m1(_tmp02a, 0.5f, vl), 2.f, _tmp02b, vl);
-                    vfloat16m1_t _tmp3m = vfmacc_vf_f16m1(vfmacc_vf_f16m1(_r05, sq2_d4, _tmp13a, vl), sq2_m2, _tmp13b, vl);
-
-                    vse16_v_f16m1(tmp[0][m], _tmp0m, vl);
-                    vse16_v_f16m1(tmp[1][m], _tmp1m, vl);
-                    vse16_v_f16m1(tmp[2][m], _tmp2m, vl);
-                    vse16_v_f16m1(tmp[3][m], _tmp3m, vl);
-
-                    output0_tm_0 += tiles * packn * 6;
-                    output0_tm_1 += tiles * packn * 6;
-                    output0_tm_2 += tiles * packn * 6;
-                    output0_tm_3 += tiles * packn * 6;
-                    output0_tm_4 += tiles * packn * 6;
-                    output0_tm_5 += tiles * packn * 6;
-                }
-
-                for (int m = 0; m < 4; m++)
-                {
-                    vfloat16m1_t _r00 = vle16_v_f16m1(tmp[m][0], vl);
-                    vfloat16m1_t _r01 = vle16_v_f16m1(tmp[m][1], vl);
-                    vfloat16m1_t _r02 = vle16_v_f16m1(tmp[m][2], vl);
-                    vfloat16m1_t _r03 = vle16_v_f16m1(tmp[m][3], vl);
-                    vfloat16m1_t _r04 = vle16_v_f16m1(tmp[m][4], vl);
-                    vfloat16m1_t _r05 = vle16_v_f16m1(tmp[m][5], vl);
-
-                    vfloat16m1_t _tmp02a = vfadd_vv_f16m1(_r01, _r02, vl);
-                    vfloat16m1_t _tmp02b = vfadd_vv_f16m1(_r03, _r04, vl);
-                    vfloat16m1_t _tmp13a = vfsub_vv_f16m1(_r01, _r02, vl);
-                    vfloat16m1_t _tmp13b = vfsub_vv_f16m1(_r03, _r04, vl);
-
-                    vfloat16m1_t _out00 = vfadd_vv_f16m1(_bias0, vfadd_vv_f16m1(vfadd_vv_f16m1(_r00, _tmp02a, vl), _tmp02b, vl), vl);
-                    vfloat16m1_t _out01 = vfadd_vv_f16m1(_bias0, vfmacc_vf_f16m1(vfmul_vf_f16m1(_tmp13a, sq2_d2, vl), sq2, _tmp13b, vl), vl);
-                    vfloat16m1_t _out02 = vfadd_vv_f16m1(_bias0, vfmacc_vf_f16m1(vfmul_vf_f16m1(_tmp02a, 0.5f, vl), 2.f, _tmp02b, vl), vl);
-                    vfloat16m1_t _out03 = vfadd_vv_f16m1(_bias0, vfmacc_vf_f16m1(vfmacc_vf_f16m1(_r05, sq2_d4, _tmp13a, vl), sq2_m2, _tmp13b, vl), vl);
-
-                    vse16_v_f16m1(output0, _out00, vl);
-                    vse16_v_f16m1(output0 + packn, _out01, vl);
-                    vse16_v_f16m1(output0 + packn * 2, _out02, vl);
-                    vse16_v_f16m1(output0 + packn * 3, _out03, vl);
-
-                    output0 += outw * packn;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_input_packn_fp16sa_rvv(const Mat& bottom_blob, Mat& bottom_blob_tm, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int inch = bottom_blob.c;
-
-    const int w_tiles = (w - 2) / 2;
-    const int h_tiles = (h - 2) / 2;
-    const int tiles = w_tiles * h_tiles;
-
-    // const float itm[4][4] = {
-    //     {1.0f,  0.0f, -1.0f,  0.0f},
-    //     {0.0f,  1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  0.00f, 1.0f}
-    // };
-
-    // 0 = r00 - r02
-    // 1 = r01 + r02
-    // 2 = r02 - r01
-    // 3 = r03 - r01
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        Mat img0_tm = bottom_blob_tm.channel(q);
-
-        // NOTE c99 variable length array
-        __fp16 tmp[4][4][packn];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const __fp16* r0 = img0.row<const __fp16>(i * 2) + (j * 2) * packn;
-
-                for (int m = 0; m < 4; m++)
-                {
-                    vfloat16m1_t _r00 = vle16_v_f16m1(r0, vl);
-                    vfloat16m1_t _r01 = vle16_v_f16m1(r0 + packn, vl);
-                    vfloat16m1_t _r02 = vle16_v_f16m1(r0 + packn * 2, vl);
-                    vfloat16m1_t _r03 = vle16_v_f16m1(r0 + packn * 3, vl);
-
-                    vfloat16m1_t _tmp0m = vfsub_vv_f16m1(_r00, _r02, vl);
-                    vfloat16m1_t _tmp1m = vfadd_vv_f16m1(_r01, _r02, vl);
-                    vfloat16m1_t _tmp2m = vfsub_vv_f16m1(_r02, _r01, vl);
-                    vfloat16m1_t _tmp3m = vfsub_vv_f16m1(_r03, _r01, vl);
-
-                    vse16_v_f16m1(tmp[0][m], _tmp0m, vl);
-                    vse16_v_f16m1(tmp[1][m], _tmp1m, vl);
-                    vse16_v_f16m1(tmp[2][m], _tmp2m, vl);
-                    vse16_v_f16m1(tmp[3][m], _tmp3m, vl);
-
-                    r0 += w * packn;
-                }
-
-                __fp16* r0_tm_0 = (__fp16*)img0_tm + (i * w_tiles + j) * packn;
-                __fp16* r0_tm_1 = r0_tm_0 + tiles * packn;
-                __fp16* r0_tm_2 = r0_tm_0 + tiles * packn * 2;
-                __fp16* r0_tm_3 = r0_tm_0 + tiles * packn * 3;
-
-                for (int m = 0; m < 4; m++)
-                {
-                    vfloat16m1_t _tmp00 = vle16_v_f16m1(tmp[m][0], vl);
-                    vfloat16m1_t _tmp01 = vle16_v_f16m1(tmp[m][1], vl);
-                    vfloat16m1_t _tmp02 = vle16_v_f16m1(tmp[m][2], vl);
-                    vfloat16m1_t _tmp03 = vle16_v_f16m1(tmp[m][3], vl);
-
-                    vfloat16m1_t _r0tm0 = vfsub_vv_f16m1(_tmp00, _tmp02, vl);
-                    vfloat16m1_t _r0tm1 = vfadd_vv_f16m1(_tmp01, _tmp02, vl);
-                    vfloat16m1_t _r0tm2 = vfsub_vv_f16m1(_tmp02, _tmp01, vl);
-                    vfloat16m1_t _r0tm3 = vfsub_vv_f16m1(_tmp03, _tmp01, vl);
-
-                    vse16_v_f16m1(r0_tm_0, _r0tm0, vl);
-                    vse16_v_f16m1(r0_tm_1, _r0tm1, vl);
-                    vse16_v_f16m1(r0_tm_2, _r0tm2, vl);
-                    vse16_v_f16m1(r0_tm_3, _r0tm3, vl);
-
-                    r0_tm_0 += tiles * packn * 4;
-                    r0_tm_1 += tiles * packn * 4;
-                    r0_tm_2 += tiles * packn * 4;
-                    r0_tm_3 += tiles * packn * 4;
-                }
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_output_packn_fp16sa_rvv(const Mat& top_blob_tm, Mat& top_blob, const Mat& bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int outch = top_blob.c;
-
-    const int w_tiles = outw / 2;
-    const int h_tiles = outh / 2;
-    const int tiles = w_tiles * h_tiles;
-
-    const __fp16* biasptr = bias;
-
-    // const float otm[2][4] = {
-    //     {1.0f,  1.0f,  1.0f,  0.0f},
-    //     {0.0f,  1.0f, -1.0f,  1.0f}
-    // };
-
-    // 0 = r00 + r01 + r02
-    // 1 = r01 - r02 + r03
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        const Mat out0_tm = top_blob_tm.channel(p);
-        Mat out0 = top_blob.channel(p);
-
-        vfloat16m1_t _bias0 = biasptr ? vle16_v_f16m1(biasptr + p * packn, vl) : vfmv_v_f_f16m1(0.f, vl);
-
-        // NOTE variable length array
-        __fp16 tmp[2][4][packn];
-
-        // tile
-        for (int i = 0; i < h_tiles; i++)
-        {
-            for (int j = 0; j < w_tiles; j++)
-            {
-                const __fp16* output0_tm_0 = (const __fp16*)out0_tm + (i * w_tiles + j) * packn;
-                const __fp16* output0_tm_1 = output0_tm_0 + tiles * packn;
-                const __fp16* output0_tm_2 = output0_tm_0 + tiles * packn * 2;
-                const __fp16* output0_tm_3 = output0_tm_0 + tiles * packn * 3;
-
-                __fp16* output0 = out0.row<__fp16>(i * 2) + (j * 2) * packn;
-
-                for (int m = 0; m < 4; m++)
-                {
-                    vfloat16m1_t _out0tm0 = vle16_v_f16m1(output0_tm_0, vl);
-                    vfloat16m1_t _out0tm1 = vle16_v_f16m1(output0_tm_1, vl);
-                    vfloat16m1_t _out0tm2 = vle16_v_f16m1(output0_tm_2, vl);
-                    vfloat16m1_t _out0tm3 = vle16_v_f16m1(output0_tm_3, vl);
-
-                    vfloat16m1_t _tmp0m = vfadd_vv_f16m1(vfadd_vv_f16m1(_out0tm0, _out0tm1, vl), _out0tm2, vl);
-                    vfloat16m1_t _tmp1m = vfadd_vv_f16m1(vfsub_vv_f16m1(_out0tm1, _out0tm2, vl), _out0tm3, vl);
-
-                    vse16_v_f16m1(tmp[0][m], _tmp0m, vl);
-                    vse16_v_f16m1(tmp[1][m], _tmp1m, vl);
-
-                    output0_tm_0 += tiles * packn * 4;
-                    output0_tm_1 += tiles * packn * 4;
-                    output0_tm_2 += tiles * packn * 4;
-                    output0_tm_3 += tiles * packn * 4;
-                }
-
-                for (int m = 0; m < 2; m++)
-                {
-                    vfloat16m1_t _tmp00 = vle16_v_f16m1(tmp[m][0], vl);
-                    vfloat16m1_t _tmp01 = vle16_v_f16m1(tmp[m][1], vl);
-                    vfloat16m1_t _tmp02 = vle16_v_f16m1(tmp[m][2], vl);
-                    vfloat16m1_t _tmp03 = vle16_v_f16m1(tmp[m][3], vl);
-
-                    vfloat16m1_t _out00 = vfadd_vv_f16m1(_bias0, vfadd_vv_f16m1(vfadd_vv_f16m1(_tmp00, _tmp01, vl), _tmp02, vl), vl);
-                    vfloat16m1_t _out01 = vfadd_vv_f16m1(_bias0, vfadd_vv_f16m1(vfsub_vv_f16m1(_tmp01, _tmp02, vl), _tmp03, vl), vl);
-
-                    vse16_v_f16m1(output0, _out00, vl);
-                    vse16_v_f16m1(output0 + packn, _out01, vl);
-
-                    output0 += outw * packn;
-                }
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_3x3.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_3x3.h
deleted file mode 100644
index 1e8dad3e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_3x3.h
+++ /dev/null
@@ -1,193 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const float bias0 = bias ? bias[g] : 0.f;
-
-        const float* kernel0 = kernel + g * 9;
-
-        float* outptr0 = out;
-        float* outptr1 = outptr0 + outw;
-
-        const float* img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0;
-        const float* r1 = img0 + w;
-        const float* r2 = img0 + w * 2;
-        const float* r3 = img0 + w * 3;
-
-        const float* k0 = kernel0;
-        const float* k1 = kernel0 + 3;
-        const float* k2 = kernel0 + 6;
-
-        int i = 0;
-
-        for (; i + 1 < outh; i += 2)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = bias0;
-                float sum2 = bias0;
-
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum2 += r1[0] * k0[0];
-                sum2 += r1[1] * k0[1];
-                sum2 += r1[2] * k0[2];
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum2 += r2[0] * k1[0];
-                sum2 += r2[1] * k1[1];
-                sum2 += r2[2] * k1[2];
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-                sum2 += r3[0] * k2[0];
-                sum2 += r3[1] * k2[1];
-                sum2 += r3[2] * k2[2];
-
-                *outptr0 = sum;
-                *outptr1 = sum2;
-
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                outptr0++;
-                outptr1++;
-            }
-
-            r0 += 2 + w;
-            r1 += 2 + w;
-            r2 += 2 + w;
-            r3 += 2 + w;
-
-            outptr0 += outw;
-            outptr1 += outw;
-        }
-
-        for (; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = bias0;
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-
-                *outptr0 = sum;
-
-                r0++;
-                r1++;
-                r2++;
-                outptr0++;
-            }
-
-            r0 += 2;
-            r1 += 2;
-            r2 += 2;
-        }
-    }
-}
-
-static void convdw3x3s2_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const float bias0 = bias ? bias[g] : 0.f;
-
-        const float* kernel0 = kernel + g * 9;
-
-        float* outptr = out;
-
-        const float* img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0;
-        const float* r1 = img0 + w;
-        const float* r2 = img0 + w * 2;
-
-        const float* k0 = kernel0;
-        const float* k1 = kernel0 + 3;
-        const float* k2 = kernel0 + 6;
-
-        int i = 0;
-
-        for (; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = bias0;
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-
-                *outptr = sum;
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-                outptr++;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_3x3_packn.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_3x3_packn.h
deleted file mode 100644
index 0cab1af0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_3x3_packn.h
+++ /dev/null
@@ -1,444 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_packn_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        vfloat32m1_t _bias0 = bias ? vle32_v_f32m1(bias + g * packn, vl) : vfmv_v_f_f32m1(0.f, vl);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-        float* outptr1 = out.row(1);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-
-        vfloat32m1_t _k00 = vle32_v_f32m1(k0, vl);
-        vfloat32m1_t _k01 = vle32_v_f32m1(k0 + packn, vl);
-        vfloat32m1_t _k02 = vle32_v_f32m1(k0 + packn * 2, vl);
-        vfloat32m1_t _k10 = vle32_v_f32m1(k0 + packn * 3, vl);
-        vfloat32m1_t _k11 = vle32_v_f32m1(k0 + packn * 4, vl);
-        vfloat32m1_t _k12 = vle32_v_f32m1(k0 + packn * 5, vl);
-        vfloat32m1_t _k20 = vle32_v_f32m1(k0 + packn * 6, vl);
-        vfloat32m1_t _k21 = vle32_v_f32m1(k0 + packn * 7, vl);
-        vfloat32m1_t _k22 = vle32_v_f32m1(k0 + packn * 8, vl);
-
-        int i = 0;
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                vfloat32m1_t _sum00 = _bias0;
-                vfloat32m1_t _sum01 = _bias0;
-                vfloat32m1_t _sum10 = _bias0;
-                vfloat32m1_t _sum11 = _bias0;
-
-                vfloat32m1_t _r00 = vle32_v_f32m1(r0, vl);
-                vfloat32m1_t _r01 = vle32_v_f32m1(r0 + packn, vl);
-                vfloat32m1_t _r02 = vle32_v_f32m1(r0 + packn * 2, vl);
-                vfloat32m1_t _r03 = vle32_v_f32m1(r0 + packn * 3, vl);
-
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k00, _r00, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k01, _r01, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k02, _r02, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k00, _r01, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k01, _r02, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k02, _r03, vl);
-
-                vfloat32m1_t _r10 = vle32_v_f32m1(r1, vl);
-                vfloat32m1_t _r11 = vle32_v_f32m1(r1 + packn, vl);
-                vfloat32m1_t _r12 = vle32_v_f32m1(r1 + packn * 2, vl);
-                vfloat32m1_t _r13 = vle32_v_f32m1(r1 + packn * 3, vl);
-
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k10, _r10, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k11, _r11, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k12, _r12, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k10, _r11, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k11, _r12, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k12, _r13, vl);
-                _sum10 = vfmacc_vv_f32m1(_sum10, _k00, _r10, vl);
-                _sum10 = vfmacc_vv_f32m1(_sum10, _k01, _r11, vl);
-                _sum10 = vfmacc_vv_f32m1(_sum10, _k02, _r12, vl);
-                _sum11 = vfmacc_vv_f32m1(_sum11, _k00, _r11, vl);
-                _sum11 = vfmacc_vv_f32m1(_sum11, _k01, _r12, vl);
-                _sum11 = vfmacc_vv_f32m1(_sum11, _k02, _r13, vl);
-
-                vfloat32m1_t _r20 = vle32_v_f32m1(r2, vl);
-                vfloat32m1_t _r21 = vle32_v_f32m1(r2 + packn, vl);
-                vfloat32m1_t _r22 = vle32_v_f32m1(r2 + packn * 2, vl);
-                vfloat32m1_t _r23 = vle32_v_f32m1(r2 + packn * 3, vl);
-
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k20, _r20, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k21, _r21, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k22, _r22, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k20, _r21, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k21, _r22, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k22, _r23, vl);
-                _sum10 = vfmacc_vv_f32m1(_sum10, _k10, _r20, vl);
-                _sum10 = vfmacc_vv_f32m1(_sum10, _k11, _r21, vl);
-                _sum10 = vfmacc_vv_f32m1(_sum10, _k12, _r22, vl);
-                _sum11 = vfmacc_vv_f32m1(_sum11, _k10, _r21, vl);
-                _sum11 = vfmacc_vv_f32m1(_sum11, _k11, _r22, vl);
-                _sum11 = vfmacc_vv_f32m1(_sum11, _k12, _r23, vl);
-
-                vfloat32m1_t _r30 = vle32_v_f32m1(r3, vl);
-                vfloat32m1_t _r31 = vle32_v_f32m1(r3 + packn, vl);
-                vfloat32m1_t _r32 = vle32_v_f32m1(r3 + packn * 2, vl);
-                vfloat32m1_t _r33 = vle32_v_f32m1(r3 + packn * 3, vl);
-
-                _sum10 = vfmacc_vv_f32m1(_sum10, _k20, _r30, vl);
-                _sum10 = vfmacc_vv_f32m1(_sum10, _k21, _r31, vl);
-                _sum10 = vfmacc_vv_f32m1(_sum10, _k22, _r32, vl);
-                _sum11 = vfmacc_vv_f32m1(_sum11, _k20, _r31, vl);
-                _sum11 = vfmacc_vv_f32m1(_sum11, _k21, _r32, vl);
-                _sum11 = vfmacc_vv_f32m1(_sum11, _k22, _r33, vl);
-
-                vse32_v_f32m1(outptr0, _sum00, vl);
-                vse32_v_f32m1(outptr0 + packn, _sum01, vl);
-                vse32_v_f32m1(outptr1, _sum10, vl);
-                vse32_v_f32m1(outptr1 + packn, _sum11, vl);
-
-                outptr0 += packn * 2;
-                outptr1 += packn * 2;
-
-                r0 += packn * 2;
-                r1 += packn * 2;
-                r2 += packn * 2;
-                r3 += packn * 2;
-            }
-            for (; j < outw; j++)
-            {
-                vfloat32m1_t _sum0 = _bias0;
-                vfloat32m1_t _sum1 = _bias0;
-
-                vfloat32m1_t _r00 = vle32_v_f32m1(r0, vl);
-                vfloat32m1_t _r01 = vle32_v_f32m1(r0 + packn, vl);
-                vfloat32m1_t _r02 = vle32_v_f32m1(r0 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k00, _r00, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k01, _r01, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k02, _r02, vl);
-
-                vfloat32m1_t _r10 = vle32_v_f32m1(r1, vl);
-                vfloat32m1_t _r11 = vle32_v_f32m1(r1 + packn, vl);
-                vfloat32m1_t _r12 = vle32_v_f32m1(r1 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k10, _r10, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k11, _r11, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k12, _r12, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k00, _r10, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k01, _r11, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k02, _r12, vl);
-
-                vfloat32m1_t _r20 = vle32_v_f32m1(r2, vl);
-                vfloat32m1_t _r21 = vle32_v_f32m1(r2 + packn, vl);
-                vfloat32m1_t _r22 = vle32_v_f32m1(r2 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k20, _r20, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k21, _r21, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k22, _r22, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k10, _r20, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k11, _r21, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k12, _r22, vl);
-
-                vfloat32m1_t _r30 = vle32_v_f32m1(r3, vl);
-                vfloat32m1_t _r31 = vle32_v_f32m1(r3 + packn, vl);
-                vfloat32m1_t _r32 = vle32_v_f32m1(r3 + packn * 2, vl);
-
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k20, _r30, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k21, _r31, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k22, _r32, vl);
-
-                vse32_v_f32m1(outptr0, _sum0, vl);
-                vse32_v_f32m1(outptr1, _sum1, vl);
-
-                outptr0 += packn;
-                outptr1 += packn;
-
-                r0 += packn;
-                r1 += packn;
-                r2 += packn;
-                r3 += packn;
-            }
-
-            r0 += 2 * packn + w * packn;
-            r1 += 2 * packn + w * packn;
-            r2 += 2 * packn + w * packn;
-            r3 += 2 * packn + w * packn;
-
-            outptr0 += outw * packn;
-            outptr1 += outw * packn;
-        }
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                vfloat32m1_t _sum00 = _bias0;
-                vfloat32m1_t _sum01 = _bias0;
-
-                vfloat32m1_t _r00 = vle32_v_f32m1(r0, vl);
-                vfloat32m1_t _r01 = vle32_v_f32m1(r0 + packn, vl);
-                vfloat32m1_t _r02 = vle32_v_f32m1(r0 + packn * 2, vl);
-                vfloat32m1_t _r03 = vle32_v_f32m1(r0 + packn * 3, vl);
-
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k00, _r00, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k01, _r01, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k02, _r02, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k00, _r01, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k01, _r02, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k02, _r03, vl);
-
-                vfloat32m1_t _r10 = vle32_v_f32m1(r1, vl);
-                vfloat32m1_t _r11 = vle32_v_f32m1(r1 + packn, vl);
-                vfloat32m1_t _r12 = vle32_v_f32m1(r1 + packn * 2, vl);
-                vfloat32m1_t _r13 = vle32_v_f32m1(r1 + packn * 3, vl);
-
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k10, _r10, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k11, _r11, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k12, _r12, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k10, _r11, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k11, _r12, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k12, _r13, vl);
-
-                vfloat32m1_t _r20 = vle32_v_f32m1(r2, vl);
-                vfloat32m1_t _r21 = vle32_v_f32m1(r2 + packn, vl);
-                vfloat32m1_t _r22 = vle32_v_f32m1(r2 + packn * 2, vl);
-                vfloat32m1_t _r23 = vle32_v_f32m1(r2 + packn * 3, vl);
-
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k20, _r20, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k21, _r21, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k22, _r22, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k20, _r21, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k21, _r22, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k22, _r23, vl);
-
-                vse32_v_f32m1(outptr0, _sum00, vl);
-                vse32_v_f32m1(outptr0 + packn, _sum01, vl);
-
-                outptr0 += packn * 2;
-
-                r0 += packn * 2;
-                r1 += packn * 2;
-                r2 += packn * 2;
-            }
-            for (; j < outw; j++)
-            {
-                vfloat32m1_t _sum0 = _bias0;
-
-                vfloat32m1_t _r00 = vle32_v_f32m1(r0, vl);
-                vfloat32m1_t _r01 = vle32_v_f32m1(r0 + packn, vl);
-                vfloat32m1_t _r02 = vle32_v_f32m1(r0 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k00, _r00, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k01, _r01, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k02, _r02, vl);
-
-                vfloat32m1_t _r10 = vle32_v_f32m1(r1, vl);
-                vfloat32m1_t _r11 = vle32_v_f32m1(r1 + packn, vl);
-                vfloat32m1_t _r12 = vle32_v_f32m1(r1 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k10, _r10, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k11, _r11, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k12, _r12, vl);
-
-                vfloat32m1_t _r20 = vle32_v_f32m1(r2, vl);
-                vfloat32m1_t _r21 = vle32_v_f32m1(r2 + packn, vl);
-                vfloat32m1_t _r22 = vle32_v_f32m1(r2 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k20, _r20, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k21, _r21, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k22, _r22, vl);
-
-                vse32_v_f32m1(outptr0, _sum0, vl);
-
-                outptr0 += packn;
-
-                r0 += packn;
-                r1 += packn;
-                r2 += packn;
-            }
-
-            r0 += 2 * packn;
-            r1 += 2 * packn;
-            r2 += 2 * packn;
-        }
-    }
-}
-
-static void convdw3x3s2_packn_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * packn;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        vfloat32m1_t _bias0 = bias ? vle32_v_f32m1(bias + g * packn, vl) : vfmv_v_f_f32m1(0.f, vl);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out;
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-
-        vfloat32m1_t _k00 = vle32_v_f32m1(k0, vl);
-        vfloat32m1_t _k01 = vle32_v_f32m1(k0 + packn, vl);
-        vfloat32m1_t _k02 = vle32_v_f32m1(k0 + packn * 2, vl);
-        vfloat32m1_t _k10 = vle32_v_f32m1(k0 + packn * 3, vl);
-        vfloat32m1_t _k11 = vle32_v_f32m1(k0 + packn * 4, vl);
-        vfloat32m1_t _k12 = vle32_v_f32m1(k0 + packn * 5, vl);
-        vfloat32m1_t _k20 = vle32_v_f32m1(k0 + packn * 6, vl);
-        vfloat32m1_t _k21 = vle32_v_f32m1(k0 + packn * 7, vl);
-        vfloat32m1_t _k22 = vle32_v_f32m1(k0 + packn * 8, vl);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                vfloat32m1_t _sum00 = _bias0;
-                vfloat32m1_t _sum01 = _bias0;
-
-                vfloat32m1_t _r00 = vle32_v_f32m1(r0, vl);
-                vfloat32m1_t _r01 = vle32_v_f32m1(r0 + packn, vl);
-                vfloat32m1_t _r02 = vle32_v_f32m1(r0 + packn * 2, vl);
-                vfloat32m1_t _r03 = vle32_v_f32m1(r0 + packn * 3, vl);
-                vfloat32m1_t _r04 = vle32_v_f32m1(r0 + packn * 4, vl);
-
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k00, _r00, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k01, _r01, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k02, _r02, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k00, _r02, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k01, _r03, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k02, _r04, vl);
-
-                vfloat32m1_t _r10 = vle32_v_f32m1(r1, vl);
-                vfloat32m1_t _r11 = vle32_v_f32m1(r1 + packn, vl);
-                vfloat32m1_t _r12 = vle32_v_f32m1(r1 + packn * 2, vl);
-                vfloat32m1_t _r13 = vle32_v_f32m1(r1 + packn * 3, vl);
-                vfloat32m1_t _r14 = vle32_v_f32m1(r1 + packn * 4, vl);
-
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k10, _r10, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k11, _r11, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k12, _r12, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k10, _r12, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k11, _r13, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k12, _r14, vl);
-
-                vfloat32m1_t _r20 = vle32_v_f32m1(r2, vl);
-                vfloat32m1_t _r21 = vle32_v_f32m1(r2 + packn, vl);
-                vfloat32m1_t _r22 = vle32_v_f32m1(r2 + packn * 2, vl);
-                vfloat32m1_t _r23 = vle32_v_f32m1(r2 + packn * 3, vl);
-                vfloat32m1_t _r24 = vle32_v_f32m1(r2 + packn * 4, vl);
-
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k20, _r20, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k21, _r21, vl);
-                _sum00 = vfmacc_vv_f32m1(_sum00, _k22, _r22, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k20, _r22, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k21, _r23, vl);
-                _sum01 = vfmacc_vv_f32m1(_sum01, _k22, _r24, vl);
-
-                vse32_v_f32m1(outptr0, _sum00, vl);
-                vse32_v_f32m1(outptr0 + packn, _sum01, vl);
-
-                outptr0 += packn * 2;
-
-                r0 += packn * 4;
-                r1 += packn * 4;
-                r2 += packn * 4;
-            }
-            for (; j < outw; j++)
-            {
-                vfloat32m1_t _sum0 = _bias0;
-
-                vfloat32m1_t _r00 = vle32_v_f32m1(r0, vl);
-                vfloat32m1_t _r01 = vle32_v_f32m1(r0 + packn, vl);
-                vfloat32m1_t _r02 = vle32_v_f32m1(r0 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k00, _r00, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k01, _r01, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k02, _r02, vl);
-
-                vfloat32m1_t _r10 = vle32_v_f32m1(r1, vl);
-                vfloat32m1_t _r11 = vle32_v_f32m1(r1 + packn, vl);
-                vfloat32m1_t _r12 = vle32_v_f32m1(r1 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k10, _r10, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k11, _r11, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k12, _r12, vl);
-
-                vfloat32m1_t _r20 = vle32_v_f32m1(r2, vl);
-                vfloat32m1_t _r21 = vle32_v_f32m1(r2 + packn, vl);
-                vfloat32m1_t _r22 = vle32_v_f32m1(r2 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k20, _r20, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k21, _r21, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k22, _r22, vl);
-
-                vse32_v_f32m1(outptr0, _sum0, vl);
-
-                outptr0 += packn;
-
-                r0 += packn * 2;
-                r1 += packn * 2;
-                r2 += packn * 2;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_3x3_packn_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_3x3_packn_fp16s.h
deleted file mode 100644
index d479385f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_3x3_packn_fp16s.h
+++ /dev/null
@@ -1,444 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_packn_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        vfloat16m1_t _bias0 = bias ? vle16_v_f16m1(bias + g * packn, vl) : vfmv_v_f_f16m1((__fp16)0.f, vl);
-
-        const __fp16* k0 = kernel.row<const __fp16>(g);
-
-        __fp16* outptr0 = out.row<__fp16>(0);
-        __fp16* outptr1 = out.row<__fp16>(1);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const __fp16* r0 = img0.row<const __fp16>(0);
-        const __fp16* r1 = img0.row<const __fp16>(1);
-        const __fp16* r2 = img0.row<const __fp16>(2);
-        const __fp16* r3 = img0.row<const __fp16>(3);
-
-        vfloat16m1_t _k00 = vle16_v_f16m1(k0, vl);
-        vfloat16m1_t _k01 = vle16_v_f16m1(k0 + packn, vl);
-        vfloat16m1_t _k02 = vle16_v_f16m1(k0 + packn * 2, vl);
-        vfloat16m1_t _k10 = vle16_v_f16m1(k0 + packn * 3, vl);
-        vfloat16m1_t _k11 = vle16_v_f16m1(k0 + packn * 4, vl);
-        vfloat16m1_t _k12 = vle16_v_f16m1(k0 + packn * 5, vl);
-        vfloat16m1_t _k20 = vle16_v_f16m1(k0 + packn * 6, vl);
-        vfloat16m1_t _k21 = vle16_v_f16m1(k0 + packn * 7, vl);
-        vfloat16m1_t _k22 = vle16_v_f16m1(k0 + packn * 8, vl);
-
-        int i = 0;
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                vfloat16m1_t _sum00 = _bias0;
-                vfloat16m1_t _sum01 = _bias0;
-                vfloat16m1_t _sum10 = _bias0;
-                vfloat16m1_t _sum11 = _bias0;
-
-                vfloat16m1_t _r00 = vle16_v_f16m1(r0, vl);
-                vfloat16m1_t _r01 = vle16_v_f16m1(r0 + packn, vl);
-                vfloat16m1_t _r02 = vle16_v_f16m1(r0 + packn * 2, vl);
-                vfloat16m1_t _r03 = vle16_v_f16m1(r0 + packn * 3, vl);
-
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k00, _r00, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k01, _r01, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k02, _r02, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k00, _r01, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k01, _r02, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k02, _r03, vl);
-
-                vfloat16m1_t _r10 = vle16_v_f16m1(r1, vl);
-                vfloat16m1_t _r11 = vle16_v_f16m1(r1 + packn, vl);
-                vfloat16m1_t _r12 = vle16_v_f16m1(r1 + packn * 2, vl);
-                vfloat16m1_t _r13 = vle16_v_f16m1(r1 + packn * 3, vl);
-
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k10, _r10, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k11, _r11, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k12, _r12, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k10, _r11, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k11, _r12, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k12, _r13, vl);
-                _sum10 = vfmacc_vv_f16m1(_sum10, _k00, _r10, vl);
-                _sum10 = vfmacc_vv_f16m1(_sum10, _k01, _r11, vl);
-                _sum10 = vfmacc_vv_f16m1(_sum10, _k02, _r12, vl);
-                _sum11 = vfmacc_vv_f16m1(_sum11, _k00, _r11, vl);
-                _sum11 = vfmacc_vv_f16m1(_sum11, _k01, _r12, vl);
-                _sum11 = vfmacc_vv_f16m1(_sum11, _k02, _r13, vl);
-
-                vfloat16m1_t _r20 = vle16_v_f16m1(r2, vl);
-                vfloat16m1_t _r21 = vle16_v_f16m1(r2 + packn, vl);
-                vfloat16m1_t _r22 = vle16_v_f16m1(r2 + packn * 2, vl);
-                vfloat16m1_t _r23 = vle16_v_f16m1(r2 + packn * 3, vl);
-
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k20, _r20, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k21, _r21, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k22, _r22, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k20, _r21, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k21, _r22, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k22, _r23, vl);
-                _sum10 = vfmacc_vv_f16m1(_sum10, _k10, _r20, vl);
-                _sum10 = vfmacc_vv_f16m1(_sum10, _k11, _r21, vl);
-                _sum10 = vfmacc_vv_f16m1(_sum10, _k12, _r22, vl);
-                _sum11 = vfmacc_vv_f16m1(_sum11, _k10, _r21, vl);
-                _sum11 = vfmacc_vv_f16m1(_sum11, _k11, _r22, vl);
-                _sum11 = vfmacc_vv_f16m1(_sum11, _k12, _r23, vl);
-
-                vfloat16m1_t _r30 = vle16_v_f16m1(r3, vl);
-                vfloat16m1_t _r31 = vle16_v_f16m1(r3 + packn, vl);
-                vfloat16m1_t _r32 = vle16_v_f16m1(r3 + packn * 2, vl);
-                vfloat16m1_t _r33 = vle16_v_f16m1(r3 + packn * 3, vl);
-
-                _sum10 = vfmacc_vv_f16m1(_sum10, _k20, _r30, vl);
-                _sum10 = vfmacc_vv_f16m1(_sum10, _k21, _r31, vl);
-                _sum10 = vfmacc_vv_f16m1(_sum10, _k22, _r32, vl);
-                _sum11 = vfmacc_vv_f16m1(_sum11, _k20, _r31, vl);
-                _sum11 = vfmacc_vv_f16m1(_sum11, _k21, _r32, vl);
-                _sum11 = vfmacc_vv_f16m1(_sum11, _k22, _r33, vl);
-
-                vse16_v_f16m1(outptr0, _sum00, vl);
-                vse16_v_f16m1(outptr0 + packn, _sum01, vl);
-                vse16_v_f16m1(outptr1, _sum10, vl);
-                vse16_v_f16m1(outptr1 + packn, _sum11, vl);
-
-                outptr0 += packn * 2;
-                outptr1 += packn * 2;
-
-                r0 += packn * 2;
-                r1 += packn * 2;
-                r2 += packn * 2;
-                r3 += packn * 2;
-            }
-            for (; j < outw; j++)
-            {
-                vfloat16m1_t _sum0 = _bias0;
-                vfloat16m1_t _sum1 = _bias0;
-
-                vfloat16m1_t _r00 = vle16_v_f16m1(r0, vl);
-                vfloat16m1_t _r01 = vle16_v_f16m1(r0 + packn, vl);
-                vfloat16m1_t _r02 = vle16_v_f16m1(r0 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k00, _r00, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k01, _r01, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k02, _r02, vl);
-
-                vfloat16m1_t _r10 = vle16_v_f16m1(r1, vl);
-                vfloat16m1_t _r11 = vle16_v_f16m1(r1 + packn, vl);
-                vfloat16m1_t _r12 = vle16_v_f16m1(r1 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k10, _r10, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k11, _r11, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k12, _r12, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k00, _r10, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k01, _r11, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k02, _r12, vl);
-
-                vfloat16m1_t _r20 = vle16_v_f16m1(r2, vl);
-                vfloat16m1_t _r21 = vle16_v_f16m1(r2 + packn, vl);
-                vfloat16m1_t _r22 = vle16_v_f16m1(r2 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k20, _r20, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k21, _r21, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k22, _r22, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k10, _r20, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k11, _r21, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k12, _r22, vl);
-
-                vfloat16m1_t _r30 = vle16_v_f16m1(r3, vl);
-                vfloat16m1_t _r31 = vle16_v_f16m1(r3 + packn, vl);
-                vfloat16m1_t _r32 = vle16_v_f16m1(r3 + packn * 2, vl);
-
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k20, _r30, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k21, _r31, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k22, _r32, vl);
-
-                vse16_v_f16m1(outptr0, _sum0, vl);
-                vse16_v_f16m1(outptr1, _sum1, vl);
-
-                outptr0 += packn;
-                outptr1 += packn;
-
-                r0 += packn;
-                r1 += packn;
-                r2 += packn;
-                r3 += packn;
-            }
-
-            r0 += 2 * packn + w * packn;
-            r1 += 2 * packn + w * packn;
-            r2 += 2 * packn + w * packn;
-            r3 += 2 * packn + w * packn;
-
-            outptr0 += outw * packn;
-            outptr1 += outw * packn;
-        }
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                vfloat16m1_t _sum00 = _bias0;
-                vfloat16m1_t _sum01 = _bias0;
-
-                vfloat16m1_t _r00 = vle16_v_f16m1(r0, vl);
-                vfloat16m1_t _r01 = vle16_v_f16m1(r0 + packn, vl);
-                vfloat16m1_t _r02 = vle16_v_f16m1(r0 + packn * 2, vl);
-                vfloat16m1_t _r03 = vle16_v_f16m1(r0 + packn * 3, vl);
-
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k00, _r00, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k01, _r01, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k02, _r02, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k00, _r01, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k01, _r02, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k02, _r03, vl);
-
-                vfloat16m1_t _r10 = vle16_v_f16m1(r1, vl);
-                vfloat16m1_t _r11 = vle16_v_f16m1(r1 + packn, vl);
-                vfloat16m1_t _r12 = vle16_v_f16m1(r1 + packn * 2, vl);
-                vfloat16m1_t _r13 = vle16_v_f16m1(r1 + packn * 3, vl);
-
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k10, _r10, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k11, _r11, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k12, _r12, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k10, _r11, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k11, _r12, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k12, _r13, vl);
-
-                vfloat16m1_t _r20 = vle16_v_f16m1(r2, vl);
-                vfloat16m1_t _r21 = vle16_v_f16m1(r2 + packn, vl);
-                vfloat16m1_t _r22 = vle16_v_f16m1(r2 + packn * 2, vl);
-                vfloat16m1_t _r23 = vle16_v_f16m1(r2 + packn * 3, vl);
-
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k20, _r20, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k21, _r21, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k22, _r22, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k20, _r21, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k21, _r22, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k22, _r23, vl);
-
-                vse16_v_f16m1(outptr0, _sum00, vl);
-                vse16_v_f16m1(outptr0 + packn, _sum01, vl);
-
-                outptr0 += packn * 2;
-
-                r0 += packn * 2;
-                r1 += packn * 2;
-                r2 += packn * 2;
-            }
-            for (; j < outw; j++)
-            {
-                vfloat16m1_t _sum0 = _bias0;
-
-                vfloat16m1_t _r00 = vle16_v_f16m1(r0, vl);
-                vfloat16m1_t _r01 = vle16_v_f16m1(r0 + packn, vl);
-                vfloat16m1_t _r02 = vle16_v_f16m1(r0 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k00, _r00, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k01, _r01, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k02, _r02, vl);
-
-                vfloat16m1_t _r10 = vle16_v_f16m1(r1, vl);
-                vfloat16m1_t _r11 = vle16_v_f16m1(r1 + packn, vl);
-                vfloat16m1_t _r12 = vle16_v_f16m1(r1 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k10, _r10, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k11, _r11, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k12, _r12, vl);
-
-                vfloat16m1_t _r20 = vle16_v_f16m1(r2, vl);
-                vfloat16m1_t _r21 = vle16_v_f16m1(r2 + packn, vl);
-                vfloat16m1_t _r22 = vle16_v_f16m1(r2 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k20, _r20, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k21, _r21, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k22, _r22, vl);
-
-                vse16_v_f16m1(outptr0, _sum0, vl);
-
-                outptr0 += packn;
-
-                r0 += packn;
-                r1 += packn;
-                r2 += packn;
-            }
-
-            r0 += 2 * packn;
-            r1 += 2 * packn;
-            r2 += 2 * packn;
-        }
-    }
-}
-
-static void convdw3x3s2_packn_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * packn;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        vfloat16m1_t _bias0 = bias ? vle16_v_f16m1(bias + g * packn, vl) : vfmv_v_f_f16m1((__fp16)0.f, vl);
-
-        const __fp16* k0 = kernel.row<const __fp16>(g);
-
-        __fp16* outptr0 = out;
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const __fp16* r0 = img0.row<const __fp16>(0);
-        const __fp16* r1 = img0.row<const __fp16>(1);
-        const __fp16* r2 = img0.row<const __fp16>(2);
-
-        vfloat16m1_t _k00 = vle16_v_f16m1(k0, vl);
-        vfloat16m1_t _k01 = vle16_v_f16m1(k0 + packn, vl);
-        vfloat16m1_t _k02 = vle16_v_f16m1(k0 + packn * 2, vl);
-        vfloat16m1_t _k10 = vle16_v_f16m1(k0 + packn * 3, vl);
-        vfloat16m1_t _k11 = vle16_v_f16m1(k0 + packn * 4, vl);
-        vfloat16m1_t _k12 = vle16_v_f16m1(k0 + packn * 5, vl);
-        vfloat16m1_t _k20 = vle16_v_f16m1(k0 + packn * 6, vl);
-        vfloat16m1_t _k21 = vle16_v_f16m1(k0 + packn * 7, vl);
-        vfloat16m1_t _k22 = vle16_v_f16m1(k0 + packn * 8, vl);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                vfloat16m1_t _sum00 = _bias0;
-                vfloat16m1_t _sum01 = _bias0;
-
-                vfloat16m1_t _r00 = vle16_v_f16m1(r0, vl);
-                vfloat16m1_t _r01 = vle16_v_f16m1(r0 + packn, vl);
-                vfloat16m1_t _r02 = vle16_v_f16m1(r0 + packn * 2, vl);
-                vfloat16m1_t _r03 = vle16_v_f16m1(r0 + packn * 3, vl);
-                vfloat16m1_t _r04 = vle16_v_f16m1(r0 + packn * 4, vl);
-
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k00, _r00, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k01, _r01, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k02, _r02, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k00, _r02, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k01, _r03, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k02, _r04, vl);
-
-                vfloat16m1_t _r10 = vle16_v_f16m1(r1, vl);
-                vfloat16m1_t _r11 = vle16_v_f16m1(r1 + packn, vl);
-                vfloat16m1_t _r12 = vle16_v_f16m1(r1 + packn * 2, vl);
-                vfloat16m1_t _r13 = vle16_v_f16m1(r1 + packn * 3, vl);
-                vfloat16m1_t _r14 = vle16_v_f16m1(r1 + packn * 4, vl);
-
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k10, _r10, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k11, _r11, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k12, _r12, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k10, _r12, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k11, _r13, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k12, _r14, vl);
-
-                vfloat16m1_t _r20 = vle16_v_f16m1(r2, vl);
-                vfloat16m1_t _r21 = vle16_v_f16m1(r2 + packn, vl);
-                vfloat16m1_t _r22 = vle16_v_f16m1(r2 + packn * 2, vl);
-                vfloat16m1_t _r23 = vle16_v_f16m1(r2 + packn * 3, vl);
-                vfloat16m1_t _r24 = vle16_v_f16m1(r2 + packn * 4, vl);
-
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k20, _r20, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k21, _r21, vl);
-                _sum00 = vfmacc_vv_f16m1(_sum00, _k22, _r22, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k20, _r22, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k21, _r23, vl);
-                _sum01 = vfmacc_vv_f16m1(_sum01, _k22, _r24, vl);
-
-                vse16_v_f16m1(outptr0, _sum00, vl);
-                vse16_v_f16m1(outptr0 + packn, _sum01, vl);
-
-                outptr0 += packn * 2;
-
-                r0 += packn * 4;
-                r1 += packn * 4;
-                r2 += packn * 4;
-            }
-            for (; j < outw; j++)
-            {
-                vfloat16m1_t _sum0 = _bias0;
-
-                vfloat16m1_t _r00 = vle16_v_f16m1(r0, vl);
-                vfloat16m1_t _r01 = vle16_v_f16m1(r0 + packn, vl);
-                vfloat16m1_t _r02 = vle16_v_f16m1(r0 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k00, _r00, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k01, _r01, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k02, _r02, vl);
-
-                vfloat16m1_t _r10 = vle16_v_f16m1(r1, vl);
-                vfloat16m1_t _r11 = vle16_v_f16m1(r1 + packn, vl);
-                vfloat16m1_t _r12 = vle16_v_f16m1(r1 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k10, _r10, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k11, _r11, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k12, _r12, vl);
-
-                vfloat16m1_t _r20 = vle16_v_f16m1(r2, vl);
-                vfloat16m1_t _r21 = vle16_v_f16m1(r2 + packn, vl);
-                vfloat16m1_t _r22 = vle16_v_f16m1(r2 + packn * 2, vl);
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k20, _r20, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k21, _r21, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k22, _r22, vl);
-
-                vse16_v_f16m1(outptr0, _sum0, vl);
-
-                outptr0 += packn;
-
-                r0 += packn * 2;
-                r1 += packn * 2;
-                r2 += packn * 2;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_5x5_packn.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_5x5_packn.h
deleted file mode 100644
index 2ef2fea7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_5x5_packn.h
+++ /dev/null
@@ -1,492 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw5x5s1_packn_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        vfloat32m1_t _bias0 = bias ? vle32_v_f32m1(bias + g * packn, vl) : vfmv_v_f_f32m1(0.f, vl);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-        float* outptr1 = out.row(1);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-        const float* r4 = img0.row(4);
-        const float* r5 = img0.row(5);
-
-        int i = 0;
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                vfloat32m1_t _sum0 = _bias0;
-                vfloat32m1_t _sum1 = _bias0;
-
-                vfloat32m1_t _r00 = vle32_v_f32m1(r0, vl);
-                vfloat32m1_t _r01 = vle32_v_f32m1(r0 + packn, vl);
-                vfloat32m1_t _r02 = vle32_v_f32m1(r0 + packn * 2, vl);
-                vfloat32m1_t _r03 = vle32_v_f32m1(r0 + packn * 3, vl);
-                vfloat32m1_t _r04 = vle32_v_f32m1(r0 + packn * 4, vl);
-
-                vfloat32m1_t _k00 = vle32_v_f32m1(k0, vl);
-                vfloat32m1_t _k01 = vle32_v_f32m1(k0 + packn, vl);
-                vfloat32m1_t _k02 = vle32_v_f32m1(k0 + packn * 2, vl);
-                vfloat32m1_t _k03 = vle32_v_f32m1(k0 + packn * 3, vl);
-                vfloat32m1_t _k04 = vle32_v_f32m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k00, _r00, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k01, _r01, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k02, _r02, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k03, _r03, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k04, _r04, vl);
-
-                vfloat32m1_t _r10 = vle32_v_f32m1(r1, vl);
-                vfloat32m1_t _r11 = vle32_v_f32m1(r1 + packn, vl);
-                vfloat32m1_t _r12 = vle32_v_f32m1(r1 + packn * 2, vl);
-                vfloat32m1_t _r13 = vle32_v_f32m1(r1 + packn * 3, vl);
-                vfloat32m1_t _r14 = vle32_v_f32m1(r1 + packn * 4, vl);
-
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k00, _r10, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k01, _r11, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k02, _r12, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k03, _r13, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k04, _r14, vl);
-
-                vfloat32m1_t _k10 = vle32_v_f32m1(k0, vl);
-                vfloat32m1_t _k11 = vle32_v_f32m1(k0 + packn, vl);
-                vfloat32m1_t _k12 = vle32_v_f32m1(k0 + packn * 2, vl);
-                vfloat32m1_t _k13 = vle32_v_f32m1(k0 + packn * 3, vl);
-                vfloat32m1_t _k14 = vle32_v_f32m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k10, _r10, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k11, _r11, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k12, _r12, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k13, _r13, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k14, _r14, vl);
-
-                vfloat32m1_t _r20 = vle32_v_f32m1(r2, vl);
-                vfloat32m1_t _r21 = vle32_v_f32m1(r2 + packn, vl);
-                vfloat32m1_t _r22 = vle32_v_f32m1(r2 + packn * 2, vl);
-                vfloat32m1_t _r23 = vle32_v_f32m1(r2 + packn * 3, vl);
-                vfloat32m1_t _r24 = vle32_v_f32m1(r2 + packn * 4, vl);
-
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k10, _r20, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k11, _r21, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k12, _r22, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k13, _r23, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k14, _r24, vl);
-
-                vfloat32m1_t _k20 = vle32_v_f32m1(k0, vl);
-                vfloat32m1_t _k21 = vle32_v_f32m1(k0 + packn, vl);
-                vfloat32m1_t _k22 = vle32_v_f32m1(k0 + packn * 2, vl);
-                vfloat32m1_t _k23 = vle32_v_f32m1(k0 + packn * 3, vl);
-                vfloat32m1_t _k24 = vle32_v_f32m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k20, _r20, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k21, _r21, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k22, _r22, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k23, _r23, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k24, _r24, vl);
-
-                vfloat32m1_t _r30 = vle32_v_f32m1(r3, vl);
-                vfloat32m1_t _r31 = vle32_v_f32m1(r3 + packn, vl);
-                vfloat32m1_t _r32 = vle32_v_f32m1(r3 + packn * 2, vl);
-                vfloat32m1_t _r33 = vle32_v_f32m1(r3 + packn * 3, vl);
-                vfloat32m1_t _r34 = vle32_v_f32m1(r3 + packn * 4, vl);
-
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k20, _r30, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k21, _r31, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k22, _r32, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k23, _r33, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k24, _r34, vl);
-
-                vfloat32m1_t _k30 = vle32_v_f32m1(k0, vl);
-                vfloat32m1_t _k31 = vle32_v_f32m1(k0 + packn, vl);
-                vfloat32m1_t _k32 = vle32_v_f32m1(k0 + packn * 2, vl);
-                vfloat32m1_t _k33 = vle32_v_f32m1(k0 + packn * 3, vl);
-                vfloat32m1_t _k34 = vle32_v_f32m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k30, _r30, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k31, _r31, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k32, _r32, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k33, _r33, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k34, _r34, vl);
-
-                vfloat32m1_t _r40 = vle32_v_f32m1(r4, vl);
-                vfloat32m1_t _r41 = vle32_v_f32m1(r4 + packn, vl);
-                vfloat32m1_t _r42 = vle32_v_f32m1(r4 + packn * 2, vl);
-                vfloat32m1_t _r43 = vle32_v_f32m1(r4 + packn * 3, vl);
-                vfloat32m1_t _r44 = vle32_v_f32m1(r4 + packn * 4, vl);
-
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k30, _r40, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k31, _r41, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k32, _r42, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k33, _r43, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k34, _r44, vl);
-
-                vfloat32m1_t _k40 = vle32_v_f32m1(k0, vl);
-                vfloat32m1_t _k41 = vle32_v_f32m1(k0 + packn, vl);
-                vfloat32m1_t _k42 = vle32_v_f32m1(k0 + packn * 2, vl);
-                vfloat32m1_t _k43 = vle32_v_f32m1(k0 + packn * 3, vl);
-                vfloat32m1_t _k44 = vle32_v_f32m1(k0 + packn * 4, vl);
-                k0 -= packn * 20;
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k40, _r40, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k41, _r41, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k42, _r42, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k43, _r43, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k44, _r44, vl);
-
-                vfloat32m1_t _r50 = vle32_v_f32m1(r5, vl);
-                vfloat32m1_t _r51 = vle32_v_f32m1(r5 + packn, vl);
-                vfloat32m1_t _r52 = vle32_v_f32m1(r5 + packn * 2, vl);
-                vfloat32m1_t _r53 = vle32_v_f32m1(r5 + packn * 3, vl);
-                vfloat32m1_t _r54 = vle32_v_f32m1(r5 + packn * 4, vl);
-
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k40, _r50, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k41, _r51, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k42, _r52, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k43, _r53, vl);
-                _sum1 = vfmacc_vv_f32m1(_sum1, _k44, _r54, vl);
-
-                vse32_v_f32m1(outptr0, _sum0, vl);
-                vse32_v_f32m1(outptr1, _sum1, vl);
-
-                outptr0 += packn;
-                outptr1 += packn;
-
-                r0 += packn;
-                r1 += packn;
-                r2 += packn;
-                r3 += packn;
-                r4 += packn;
-                r5 += packn;
-            }
-
-            r0 += 4 * packn + w * packn;
-            r1 += 4 * packn + w * packn;
-            r2 += 4 * packn + w * packn;
-            r3 += 4 * packn + w * packn;
-            r4 += 4 * packn + w * packn;
-            r5 += 4 * packn + w * packn;
-
-            outptr0 += outw * packn;
-            outptr1 += outw * packn;
-        }
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                vfloat32m1_t _sum0 = _bias0;
-
-                vfloat32m1_t _r00 = vle32_v_f32m1(r0, vl);
-                vfloat32m1_t _r01 = vle32_v_f32m1(r0 + packn, vl);
-                vfloat32m1_t _r02 = vle32_v_f32m1(r0 + packn * 2, vl);
-                vfloat32m1_t _r03 = vle32_v_f32m1(r0 + packn * 3, vl);
-                vfloat32m1_t _r04 = vle32_v_f32m1(r0 + packn * 4, vl);
-
-                vfloat32m1_t _k00 = vle32_v_f32m1(k0, vl);
-                vfloat32m1_t _k01 = vle32_v_f32m1(k0 + packn, vl);
-                vfloat32m1_t _k02 = vle32_v_f32m1(k0 + packn * 2, vl);
-                vfloat32m1_t _k03 = vle32_v_f32m1(k0 + packn * 3, vl);
-                vfloat32m1_t _k04 = vle32_v_f32m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k00, _r00, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k01, _r01, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k02, _r02, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k03, _r03, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k04, _r04, vl);
-
-                vfloat32m1_t _r10 = vle32_v_f32m1(r1, vl);
-                vfloat32m1_t _r11 = vle32_v_f32m1(r1 + packn, vl);
-                vfloat32m1_t _r12 = vle32_v_f32m1(r1 + packn * 2, vl);
-                vfloat32m1_t _r13 = vle32_v_f32m1(r1 + packn * 3, vl);
-                vfloat32m1_t _r14 = vle32_v_f32m1(r1 + packn * 4, vl);
-
-                vfloat32m1_t _k10 = vle32_v_f32m1(k0, vl);
-                vfloat32m1_t _k11 = vle32_v_f32m1(k0 + packn, vl);
-                vfloat32m1_t _k12 = vle32_v_f32m1(k0 + packn * 2, vl);
-                vfloat32m1_t _k13 = vle32_v_f32m1(k0 + packn * 3, vl);
-                vfloat32m1_t _k14 = vle32_v_f32m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k10, _r10, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k11, _r11, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k12, _r12, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k13, _r13, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k14, _r14, vl);
-
-                vfloat32m1_t _r20 = vle32_v_f32m1(r2, vl);
-                vfloat32m1_t _r21 = vle32_v_f32m1(r2 + packn, vl);
-                vfloat32m1_t _r22 = vle32_v_f32m1(r2 + packn * 2, vl);
-                vfloat32m1_t _r23 = vle32_v_f32m1(r2 + packn * 3, vl);
-                vfloat32m1_t _r24 = vle32_v_f32m1(r2 + packn * 4, vl);
-
-                vfloat32m1_t _k20 = vle32_v_f32m1(k0, vl);
-                vfloat32m1_t _k21 = vle32_v_f32m1(k0 + packn, vl);
-                vfloat32m1_t _k22 = vle32_v_f32m1(k0 + packn * 2, vl);
-                vfloat32m1_t _k23 = vle32_v_f32m1(k0 + packn * 3, vl);
-                vfloat32m1_t _k24 = vle32_v_f32m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k20, _r20, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k21, _r21, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k22, _r22, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k23, _r23, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k24, _r24, vl);
-
-                vfloat32m1_t _r30 = vle32_v_f32m1(r3, vl);
-                vfloat32m1_t _r31 = vle32_v_f32m1(r3 + packn, vl);
-                vfloat32m1_t _r32 = vle32_v_f32m1(r3 + packn * 2, vl);
-                vfloat32m1_t _r33 = vle32_v_f32m1(r3 + packn * 3, vl);
-                vfloat32m1_t _r34 = vle32_v_f32m1(r3 + packn * 4, vl);
-
-                vfloat32m1_t _k30 = vle32_v_f32m1(k0, vl);
-                vfloat32m1_t _k31 = vle32_v_f32m1(k0 + packn, vl);
-                vfloat32m1_t _k32 = vle32_v_f32m1(k0 + packn * 2, vl);
-                vfloat32m1_t _k33 = vle32_v_f32m1(k0 + packn * 3, vl);
-                vfloat32m1_t _k34 = vle32_v_f32m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k30, _r30, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k31, _r31, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k32, _r32, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k33, _r33, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k34, _r34, vl);
-
-                vfloat32m1_t _r40 = vle32_v_f32m1(r4, vl);
-                vfloat32m1_t _r41 = vle32_v_f32m1(r4 + packn, vl);
-                vfloat32m1_t _r42 = vle32_v_f32m1(r4 + packn * 2, vl);
-                vfloat32m1_t _r43 = vle32_v_f32m1(r4 + packn * 3, vl);
-                vfloat32m1_t _r44 = vle32_v_f32m1(r4 + packn * 4, vl);
-
-                vfloat32m1_t _k40 = vle32_v_f32m1(k0, vl);
-                vfloat32m1_t _k41 = vle32_v_f32m1(k0 + packn, vl);
-                vfloat32m1_t _k42 = vle32_v_f32m1(k0 + packn * 2, vl);
-                vfloat32m1_t _k43 = vle32_v_f32m1(k0 + packn * 3, vl);
-                vfloat32m1_t _k44 = vle32_v_f32m1(k0 + packn * 4, vl);
-                k0 -= packn * 20;
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k40, _r40, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k41, _r41, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k42, _r42, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k43, _r43, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k44, _r44, vl);
-
-                vse32_v_f32m1(outptr0, _sum0, vl);
-
-                outptr0 += packn;
-
-                r0 += packn;
-                r1 += packn;
-                r2 += packn;
-                r3 += packn;
-                r4 += packn;
-            }
-
-            r0 += 4 * packn;
-            r1 += 4 * packn;
-            r2 += 4 * packn;
-            r3 += 4 * packn;
-            r4 += 4 * packn;
-        }
-    }
-}
-
-static void convdw5x5s2_packn_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * packn;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        vfloat32m1_t _bias0 = bias ? vle32_v_f32m1(bias + g * packn, vl) : vfmv_v_f_f32m1(0.f, vl);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out;
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-        const float* r4 = img0.row(4);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                vfloat32m1_t _sum0 = _bias0;
-
-                vfloat32m1_t _r00 = vle32_v_f32m1(r0, vl);
-                vfloat32m1_t _r01 = vle32_v_f32m1(r0 + packn, vl);
-                vfloat32m1_t _r02 = vle32_v_f32m1(r0 + packn * 2, vl);
-                vfloat32m1_t _r03 = vle32_v_f32m1(r0 + packn * 3, vl);
-                vfloat32m1_t _r04 = vle32_v_f32m1(r0 + packn * 4, vl);
-
-                vfloat32m1_t _k00 = vle32_v_f32m1(k0, vl);
-                vfloat32m1_t _k01 = vle32_v_f32m1(k0 + packn, vl);
-                vfloat32m1_t _k02 = vle32_v_f32m1(k0 + packn * 2, vl);
-                vfloat32m1_t _k03 = vle32_v_f32m1(k0 + packn * 3, vl);
-                vfloat32m1_t _k04 = vle32_v_f32m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k00, _r00, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k01, _r01, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k02, _r02, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k03, _r03, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k04, _r04, vl);
-
-                vfloat32m1_t _r10 = vle32_v_f32m1(r1, vl);
-                vfloat32m1_t _r11 = vle32_v_f32m1(r1 + packn, vl);
-                vfloat32m1_t _r12 = vle32_v_f32m1(r1 + packn * 2, vl);
-                vfloat32m1_t _r13 = vle32_v_f32m1(r1 + packn * 3, vl);
-                vfloat32m1_t _r14 = vle32_v_f32m1(r1 + packn * 4, vl);
-
-                vfloat32m1_t _k10 = vle32_v_f32m1(k0, vl);
-                vfloat32m1_t _k11 = vle32_v_f32m1(k0 + packn, vl);
-                vfloat32m1_t _k12 = vle32_v_f32m1(k0 + packn * 2, vl);
-                vfloat32m1_t _k13 = vle32_v_f32m1(k0 + packn * 3, vl);
-                vfloat32m1_t _k14 = vle32_v_f32m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k10, _r10, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k11, _r11, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k12, _r12, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k13, _r13, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k14, _r14, vl);
-
-                vfloat32m1_t _r20 = vle32_v_f32m1(r2, vl);
-                vfloat32m1_t _r21 = vle32_v_f32m1(r2 + packn, vl);
-                vfloat32m1_t _r22 = vle32_v_f32m1(r2 + packn * 2, vl);
-                vfloat32m1_t _r23 = vle32_v_f32m1(r2 + packn * 3, vl);
-                vfloat32m1_t _r24 = vle32_v_f32m1(r2 + packn * 4, vl);
-
-                vfloat32m1_t _k20 = vle32_v_f32m1(k0, vl);
-                vfloat32m1_t _k21 = vle32_v_f32m1(k0 + packn, vl);
-                vfloat32m1_t _k22 = vle32_v_f32m1(k0 + packn * 2, vl);
-                vfloat32m1_t _k23 = vle32_v_f32m1(k0 + packn * 3, vl);
-                vfloat32m1_t _k24 = vle32_v_f32m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k20, _r20, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k21, _r21, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k22, _r22, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k23, _r23, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k24, _r24, vl);
-
-                vfloat32m1_t _r30 = vle32_v_f32m1(r3, vl);
-                vfloat32m1_t _r31 = vle32_v_f32m1(r3 + packn, vl);
-                vfloat32m1_t _r32 = vle32_v_f32m1(r3 + packn * 2, vl);
-                vfloat32m1_t _r33 = vle32_v_f32m1(r3 + packn * 3, vl);
-                vfloat32m1_t _r34 = vle32_v_f32m1(r3 + packn * 4, vl);
-
-                vfloat32m1_t _k30 = vle32_v_f32m1(k0, vl);
-                vfloat32m1_t _k31 = vle32_v_f32m1(k0 + packn, vl);
-                vfloat32m1_t _k32 = vle32_v_f32m1(k0 + packn * 2, vl);
-                vfloat32m1_t _k33 = vle32_v_f32m1(k0 + packn * 3, vl);
-                vfloat32m1_t _k34 = vle32_v_f32m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k30, _r30, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k31, _r31, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k32, _r32, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k33, _r33, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k34, _r34, vl);
-
-                vfloat32m1_t _r40 = vle32_v_f32m1(r4, vl);
-                vfloat32m1_t _r41 = vle32_v_f32m1(r4 + packn, vl);
-                vfloat32m1_t _r42 = vle32_v_f32m1(r4 + packn * 2, vl);
-                vfloat32m1_t _r43 = vle32_v_f32m1(r4 + packn * 3, vl);
-                vfloat32m1_t _r44 = vle32_v_f32m1(r4 + packn * 4, vl);
-
-                vfloat32m1_t _k40 = vle32_v_f32m1(k0, vl);
-                vfloat32m1_t _k41 = vle32_v_f32m1(k0 + packn, vl);
-                vfloat32m1_t _k42 = vle32_v_f32m1(k0 + packn * 2, vl);
-                vfloat32m1_t _k43 = vle32_v_f32m1(k0 + packn * 3, vl);
-                vfloat32m1_t _k44 = vle32_v_f32m1(k0 + packn * 4, vl);
-                k0 -= packn * 20;
-
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k40, _r40, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k41, _r41, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k42, _r42, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k43, _r43, vl);
-                _sum0 = vfmacc_vv_f32m1(_sum0, _k44, _r44, vl);
-
-                vse32_v_f32m1(outptr0, _sum0, vl);
-
-                outptr0 += packn;
-
-                r0 += packn * 2;
-                r1 += packn * 2;
-                r2 += packn * 2;
-                r3 += packn * 2;
-                r4 += packn * 2;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-            r3 += tailstep;
-            r4 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_5x5_packn_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_5x5_packn_fp16s.h
deleted file mode 100644
index 08270e30..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_5x5_packn_fp16s.h
+++ /dev/null
@@ -1,492 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw5x5s1_packn_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        vfloat16m1_t _bias0 = bias ? vle16_v_f16m1(bias + g * packn, vl) : vfmv_v_f_f16m1((__fp16)0.f, vl);
-
-        const __fp16* k0 = kernel.row<const __fp16>(g);
-
-        __fp16* outptr0 = out.row<__fp16>(0);
-        __fp16* outptr1 = out.row<__fp16>(1);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const __fp16* r0 = img0.row<const __fp16>(0);
-        const __fp16* r1 = img0.row<const __fp16>(1);
-        const __fp16* r2 = img0.row<const __fp16>(2);
-        const __fp16* r3 = img0.row<const __fp16>(3);
-        const __fp16* r4 = img0.row<const __fp16>(4);
-        const __fp16* r5 = img0.row<const __fp16>(5);
-
-        int i = 0;
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                vfloat16m1_t _sum0 = _bias0;
-                vfloat16m1_t _sum1 = _bias0;
-
-                vfloat16m1_t _r00 = vle16_v_f16m1(r0, vl);
-                vfloat16m1_t _r01 = vle16_v_f16m1(r0 + packn, vl);
-                vfloat16m1_t _r02 = vle16_v_f16m1(r0 + packn * 2, vl);
-                vfloat16m1_t _r03 = vle16_v_f16m1(r0 + packn * 3, vl);
-                vfloat16m1_t _r04 = vle16_v_f16m1(r0 + packn * 4, vl);
-
-                vfloat16m1_t _k00 = vle16_v_f16m1(k0, vl);
-                vfloat16m1_t _k01 = vle16_v_f16m1(k0 + packn, vl);
-                vfloat16m1_t _k02 = vle16_v_f16m1(k0 + packn * 2, vl);
-                vfloat16m1_t _k03 = vle16_v_f16m1(k0 + packn * 3, vl);
-                vfloat16m1_t _k04 = vle16_v_f16m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k00, _r00, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k01, _r01, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k02, _r02, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k03, _r03, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k04, _r04, vl);
-
-                vfloat16m1_t _r10 = vle16_v_f16m1(r1, vl);
-                vfloat16m1_t _r11 = vle16_v_f16m1(r1 + packn, vl);
-                vfloat16m1_t _r12 = vle16_v_f16m1(r1 + packn * 2, vl);
-                vfloat16m1_t _r13 = vle16_v_f16m1(r1 + packn * 3, vl);
-                vfloat16m1_t _r14 = vle16_v_f16m1(r1 + packn * 4, vl);
-
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k00, _r10, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k01, _r11, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k02, _r12, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k03, _r13, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k04, _r14, vl);
-
-                vfloat16m1_t _k10 = vle16_v_f16m1(k0, vl);
-                vfloat16m1_t _k11 = vle16_v_f16m1(k0 + packn, vl);
-                vfloat16m1_t _k12 = vle16_v_f16m1(k0 + packn * 2, vl);
-                vfloat16m1_t _k13 = vle16_v_f16m1(k0 + packn * 3, vl);
-                vfloat16m1_t _k14 = vle16_v_f16m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k10, _r10, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k11, _r11, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k12, _r12, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k13, _r13, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k14, _r14, vl);
-
-                vfloat16m1_t _r20 = vle16_v_f16m1(r2, vl);
-                vfloat16m1_t _r21 = vle16_v_f16m1(r2 + packn, vl);
-                vfloat16m1_t _r22 = vle16_v_f16m1(r2 + packn * 2, vl);
-                vfloat16m1_t _r23 = vle16_v_f16m1(r2 + packn * 3, vl);
-                vfloat16m1_t _r24 = vle16_v_f16m1(r2 + packn * 4, vl);
-
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k10, _r20, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k11, _r21, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k12, _r22, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k13, _r23, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k14, _r24, vl);
-
-                vfloat16m1_t _k20 = vle16_v_f16m1(k0, vl);
-                vfloat16m1_t _k21 = vle16_v_f16m1(k0 + packn, vl);
-                vfloat16m1_t _k22 = vle16_v_f16m1(k0 + packn * 2, vl);
-                vfloat16m1_t _k23 = vle16_v_f16m1(k0 + packn * 3, vl);
-                vfloat16m1_t _k24 = vle16_v_f16m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k20, _r20, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k21, _r21, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k22, _r22, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k23, _r23, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k24, _r24, vl);
-
-                vfloat16m1_t _r30 = vle16_v_f16m1(r3, vl);
-                vfloat16m1_t _r31 = vle16_v_f16m1(r3 + packn, vl);
-                vfloat16m1_t _r32 = vle16_v_f16m1(r3 + packn * 2, vl);
-                vfloat16m1_t _r33 = vle16_v_f16m1(r3 + packn * 3, vl);
-                vfloat16m1_t _r34 = vle16_v_f16m1(r3 + packn * 4, vl);
-
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k20, _r30, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k21, _r31, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k22, _r32, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k23, _r33, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k24, _r34, vl);
-
-                vfloat16m1_t _k30 = vle16_v_f16m1(k0, vl);
-                vfloat16m1_t _k31 = vle16_v_f16m1(k0 + packn, vl);
-                vfloat16m1_t _k32 = vle16_v_f16m1(k0 + packn * 2, vl);
-                vfloat16m1_t _k33 = vle16_v_f16m1(k0 + packn * 3, vl);
-                vfloat16m1_t _k34 = vle16_v_f16m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k30, _r30, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k31, _r31, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k32, _r32, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k33, _r33, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k34, _r34, vl);
-
-                vfloat16m1_t _r40 = vle16_v_f16m1(r4, vl);
-                vfloat16m1_t _r41 = vle16_v_f16m1(r4 + packn, vl);
-                vfloat16m1_t _r42 = vle16_v_f16m1(r4 + packn * 2, vl);
-                vfloat16m1_t _r43 = vle16_v_f16m1(r4 + packn * 3, vl);
-                vfloat16m1_t _r44 = vle16_v_f16m1(r4 + packn * 4, vl);
-
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k30, _r40, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k31, _r41, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k32, _r42, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k33, _r43, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k34, _r44, vl);
-
-                vfloat16m1_t _k40 = vle16_v_f16m1(k0, vl);
-                vfloat16m1_t _k41 = vle16_v_f16m1(k0 + packn, vl);
-                vfloat16m1_t _k42 = vle16_v_f16m1(k0 + packn * 2, vl);
-                vfloat16m1_t _k43 = vle16_v_f16m1(k0 + packn * 3, vl);
-                vfloat16m1_t _k44 = vle16_v_f16m1(k0 + packn * 4, vl);
-                k0 -= packn * 20;
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k40, _r40, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k41, _r41, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k42, _r42, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k43, _r43, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k44, _r44, vl);
-
-                vfloat16m1_t _r50 = vle16_v_f16m1(r5, vl);
-                vfloat16m1_t _r51 = vle16_v_f16m1(r5 + packn, vl);
-                vfloat16m1_t _r52 = vle16_v_f16m1(r5 + packn * 2, vl);
-                vfloat16m1_t _r53 = vle16_v_f16m1(r5 + packn * 3, vl);
-                vfloat16m1_t _r54 = vle16_v_f16m1(r5 + packn * 4, vl);
-
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k40, _r50, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k41, _r51, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k42, _r52, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k43, _r53, vl);
-                _sum1 = vfmacc_vv_f16m1(_sum1, _k44, _r54, vl);
-
-                vse16_v_f16m1(outptr0, _sum0, vl);
-                vse16_v_f16m1(outptr1, _sum1, vl);
-
-                outptr0 += packn;
-                outptr1 += packn;
-
-                r0 += packn;
-                r1 += packn;
-                r2 += packn;
-                r3 += packn;
-                r4 += packn;
-                r5 += packn;
-            }
-
-            r0 += 4 * packn + w * packn;
-            r1 += 4 * packn + w * packn;
-            r2 += 4 * packn + w * packn;
-            r3 += 4 * packn + w * packn;
-            r4 += 4 * packn + w * packn;
-            r5 += 4 * packn + w * packn;
-
-            outptr0 += outw * packn;
-            outptr1 += outw * packn;
-        }
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                vfloat16m1_t _sum0 = _bias0;
-
-                vfloat16m1_t _r00 = vle16_v_f16m1(r0, vl);
-                vfloat16m1_t _r01 = vle16_v_f16m1(r0 + packn, vl);
-                vfloat16m1_t _r02 = vle16_v_f16m1(r0 + packn * 2, vl);
-                vfloat16m1_t _r03 = vle16_v_f16m1(r0 + packn * 3, vl);
-                vfloat16m1_t _r04 = vle16_v_f16m1(r0 + packn * 4, vl);
-
-                vfloat16m1_t _k00 = vle16_v_f16m1(k0, vl);
-                vfloat16m1_t _k01 = vle16_v_f16m1(k0 + packn, vl);
-                vfloat16m1_t _k02 = vle16_v_f16m1(k0 + packn * 2, vl);
-                vfloat16m1_t _k03 = vle16_v_f16m1(k0 + packn * 3, vl);
-                vfloat16m1_t _k04 = vle16_v_f16m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k00, _r00, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k01, _r01, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k02, _r02, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k03, _r03, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k04, _r04, vl);
-
-                vfloat16m1_t _r10 = vle16_v_f16m1(r1, vl);
-                vfloat16m1_t _r11 = vle16_v_f16m1(r1 + packn, vl);
-                vfloat16m1_t _r12 = vle16_v_f16m1(r1 + packn * 2, vl);
-                vfloat16m1_t _r13 = vle16_v_f16m1(r1 + packn * 3, vl);
-                vfloat16m1_t _r14 = vle16_v_f16m1(r1 + packn * 4, vl);
-
-                vfloat16m1_t _k10 = vle16_v_f16m1(k0, vl);
-                vfloat16m1_t _k11 = vle16_v_f16m1(k0 + packn, vl);
-                vfloat16m1_t _k12 = vle16_v_f16m1(k0 + packn * 2, vl);
-                vfloat16m1_t _k13 = vle16_v_f16m1(k0 + packn * 3, vl);
-                vfloat16m1_t _k14 = vle16_v_f16m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k10, _r10, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k11, _r11, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k12, _r12, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k13, _r13, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k14, _r14, vl);
-
-                vfloat16m1_t _r20 = vle16_v_f16m1(r2, vl);
-                vfloat16m1_t _r21 = vle16_v_f16m1(r2 + packn, vl);
-                vfloat16m1_t _r22 = vle16_v_f16m1(r2 + packn * 2, vl);
-                vfloat16m1_t _r23 = vle16_v_f16m1(r2 + packn * 3, vl);
-                vfloat16m1_t _r24 = vle16_v_f16m1(r2 + packn * 4, vl);
-
-                vfloat16m1_t _k20 = vle16_v_f16m1(k0, vl);
-                vfloat16m1_t _k21 = vle16_v_f16m1(k0 + packn, vl);
-                vfloat16m1_t _k22 = vle16_v_f16m1(k0 + packn * 2, vl);
-                vfloat16m1_t _k23 = vle16_v_f16m1(k0 + packn * 3, vl);
-                vfloat16m1_t _k24 = vle16_v_f16m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k20, _r20, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k21, _r21, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k22, _r22, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k23, _r23, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k24, _r24, vl);
-
-                vfloat16m1_t _r30 = vle16_v_f16m1(r3, vl);
-                vfloat16m1_t _r31 = vle16_v_f16m1(r3 + packn, vl);
-                vfloat16m1_t _r32 = vle16_v_f16m1(r3 + packn * 2, vl);
-                vfloat16m1_t _r33 = vle16_v_f16m1(r3 + packn * 3, vl);
-                vfloat16m1_t _r34 = vle16_v_f16m1(r3 + packn * 4, vl);
-
-                vfloat16m1_t _k30 = vle16_v_f16m1(k0, vl);
-                vfloat16m1_t _k31 = vle16_v_f16m1(k0 + packn, vl);
-                vfloat16m1_t _k32 = vle16_v_f16m1(k0 + packn * 2, vl);
-                vfloat16m1_t _k33 = vle16_v_f16m1(k0 + packn * 3, vl);
-                vfloat16m1_t _k34 = vle16_v_f16m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k30, _r30, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k31, _r31, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k32, _r32, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k33, _r33, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k34, _r34, vl);
-
-                vfloat16m1_t _r40 = vle16_v_f16m1(r4, vl);
-                vfloat16m1_t _r41 = vle16_v_f16m1(r4 + packn, vl);
-                vfloat16m1_t _r42 = vle16_v_f16m1(r4 + packn * 2, vl);
-                vfloat16m1_t _r43 = vle16_v_f16m1(r4 + packn * 3, vl);
-                vfloat16m1_t _r44 = vle16_v_f16m1(r4 + packn * 4, vl);
-
-                vfloat16m1_t _k40 = vle16_v_f16m1(k0, vl);
-                vfloat16m1_t _k41 = vle16_v_f16m1(k0 + packn, vl);
-                vfloat16m1_t _k42 = vle16_v_f16m1(k0 + packn * 2, vl);
-                vfloat16m1_t _k43 = vle16_v_f16m1(k0 + packn * 3, vl);
-                vfloat16m1_t _k44 = vle16_v_f16m1(k0 + packn * 4, vl);
-                k0 -= packn * 20;
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k40, _r40, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k41, _r41, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k42, _r42, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k43, _r43, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k44, _r44, vl);
-
-                vse16_v_f16m1(outptr0, _sum0, vl);
-
-                outptr0 += packn;
-
-                r0 += packn;
-                r1 += packn;
-                r2 += packn;
-                r3 += packn;
-                r4 += packn;
-            }
-
-            r0 += 4 * packn;
-            r1 += 4 * packn;
-            r2 += 4 * packn;
-            r3 += 4 * packn;
-            r4 += 4 * packn;
-        }
-    }
-}
-
-static void convdw5x5s2_packn_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * packn;
-
-    const __fp16* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        vfloat16m1_t _bias0 = bias ? vle16_v_f16m1(bias + g * packn, vl) : vfmv_v_f_f16m1((__fp16)0.f, vl);
-
-        const __fp16* k0 = kernel.row<const __fp16>(g);
-
-        __fp16* outptr0 = out.row<__fp16>(0);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const __fp16* r0 = img0.row<const __fp16>(0);
-        const __fp16* r1 = img0.row<const __fp16>(1);
-        const __fp16* r2 = img0.row<const __fp16>(2);
-        const __fp16* r3 = img0.row<const __fp16>(3);
-        const __fp16* r4 = img0.row<const __fp16>(4);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                vfloat16m1_t _sum0 = _bias0;
-
-                vfloat16m1_t _r00 = vle16_v_f16m1(r0, vl);
-                vfloat16m1_t _r01 = vle16_v_f16m1(r0 + packn, vl);
-                vfloat16m1_t _r02 = vle16_v_f16m1(r0 + packn * 2, vl);
-                vfloat16m1_t _r03 = vle16_v_f16m1(r0 + packn * 3, vl);
-                vfloat16m1_t _r04 = vle16_v_f16m1(r0 + packn * 4, vl);
-
-                vfloat16m1_t _k00 = vle16_v_f16m1(k0, vl);
-                vfloat16m1_t _k01 = vle16_v_f16m1(k0 + packn, vl);
-                vfloat16m1_t _k02 = vle16_v_f16m1(k0 + packn * 2, vl);
-                vfloat16m1_t _k03 = vle16_v_f16m1(k0 + packn * 3, vl);
-                vfloat16m1_t _k04 = vle16_v_f16m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k00, _r00, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k01, _r01, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k02, _r02, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k03, _r03, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k04, _r04, vl);
-
-                vfloat16m1_t _r10 = vle16_v_f16m1(r1, vl);
-                vfloat16m1_t _r11 = vle16_v_f16m1(r1 + packn, vl);
-                vfloat16m1_t _r12 = vle16_v_f16m1(r1 + packn * 2, vl);
-                vfloat16m1_t _r13 = vle16_v_f16m1(r1 + packn * 3, vl);
-                vfloat16m1_t _r14 = vle16_v_f16m1(r1 + packn * 4, vl);
-
-                vfloat16m1_t _k10 = vle16_v_f16m1(k0, vl);
-                vfloat16m1_t _k11 = vle16_v_f16m1(k0 + packn, vl);
-                vfloat16m1_t _k12 = vle16_v_f16m1(k0 + packn * 2, vl);
-                vfloat16m1_t _k13 = vle16_v_f16m1(k0 + packn * 3, vl);
-                vfloat16m1_t _k14 = vle16_v_f16m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k10, _r10, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k11, _r11, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k12, _r12, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k13, _r13, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k14, _r14, vl);
-
-                vfloat16m1_t _r20 = vle16_v_f16m1(r2, vl);
-                vfloat16m1_t _r21 = vle16_v_f16m1(r2 + packn, vl);
-                vfloat16m1_t _r22 = vle16_v_f16m1(r2 + packn * 2, vl);
-                vfloat16m1_t _r23 = vle16_v_f16m1(r2 + packn * 3, vl);
-                vfloat16m1_t _r24 = vle16_v_f16m1(r2 + packn * 4, vl);
-
-                vfloat16m1_t _k20 = vle16_v_f16m1(k0, vl);
-                vfloat16m1_t _k21 = vle16_v_f16m1(k0 + packn, vl);
-                vfloat16m1_t _k22 = vle16_v_f16m1(k0 + packn * 2, vl);
-                vfloat16m1_t _k23 = vle16_v_f16m1(k0 + packn * 3, vl);
-                vfloat16m1_t _k24 = vle16_v_f16m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k20, _r20, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k21, _r21, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k22, _r22, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k23, _r23, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k24, _r24, vl);
-
-                vfloat16m1_t _r30 = vle16_v_f16m1(r3, vl);
-                vfloat16m1_t _r31 = vle16_v_f16m1(r3 + packn, vl);
-                vfloat16m1_t _r32 = vle16_v_f16m1(r3 + packn * 2, vl);
-                vfloat16m1_t _r33 = vle16_v_f16m1(r3 + packn * 3, vl);
-                vfloat16m1_t _r34 = vle16_v_f16m1(r3 + packn * 4, vl);
-
-                vfloat16m1_t _k30 = vle16_v_f16m1(k0, vl);
-                vfloat16m1_t _k31 = vle16_v_f16m1(k0 + packn, vl);
-                vfloat16m1_t _k32 = vle16_v_f16m1(k0 + packn * 2, vl);
-                vfloat16m1_t _k33 = vle16_v_f16m1(k0 + packn * 3, vl);
-                vfloat16m1_t _k34 = vle16_v_f16m1(k0 + packn * 4, vl);
-                k0 += packn * 5;
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k30, _r30, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k31, _r31, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k32, _r32, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k33, _r33, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k34, _r34, vl);
-
-                vfloat16m1_t _r40 = vle16_v_f16m1(r4, vl);
-                vfloat16m1_t _r41 = vle16_v_f16m1(r4 + packn, vl);
-                vfloat16m1_t _r42 = vle16_v_f16m1(r4 + packn * 2, vl);
-                vfloat16m1_t _r43 = vle16_v_f16m1(r4 + packn * 3, vl);
-                vfloat16m1_t _r44 = vle16_v_f16m1(r4 + packn * 4, vl);
-
-                vfloat16m1_t _k40 = vle16_v_f16m1(k0, vl);
-                vfloat16m1_t _k41 = vle16_v_f16m1(k0 + packn, vl);
-                vfloat16m1_t _k42 = vle16_v_f16m1(k0 + packn * 2, vl);
-                vfloat16m1_t _k43 = vle16_v_f16m1(k0 + packn * 3, vl);
-                vfloat16m1_t _k44 = vle16_v_f16m1(k0 + packn * 4, vl);
-                k0 -= packn * 20;
-
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k40, _r40, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k41, _r41, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k42, _r42, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k43, _r43, vl);
-                _sum0 = vfmacc_vv_f16m1(_sum0, _k44, _r44, vl);
-
-                vse16_v_f16m1(outptr0, _sum0, vl);
-
-                outptr0 += packn;
-
-                r0 += packn * 2;
-                r1 += packn * 2;
-                r2 += packn * 2;
-                r3 += packn * 2;
-                r4 += packn * 2;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-            r3 += tailstep;
-            r4 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_riscv.h
deleted file mode 100644
index f9503975..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/convolutiondepthwise_riscv.h
+++ /dev/null
@@ -1,54 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTIONDEPTHWISE_RISCV_H
-#define LAYER_CONVOLUTIONDEPTHWISE_RISCV_H
-
-#include "convolutiondepthwise.h"
-
-namespace ncnn {
-
-class ConvolutionDepthWise_riscv : public ConvolutionDepthWise
-{
-public:
-    ConvolutionDepthWise_riscv();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    int create_group_ops(const Option& opt);
-#if __riscv_vector && __riscv_zfh
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Layer* activation;
-    std::vector<ncnn::Layer*> group_ops;
-
-    Mat weight_data_tm;
-
-    // fp16
-    Mat bias_data_fp16;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTIONDEPTHWISE_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/crop_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/crop_riscv.h
deleted file mode 100644
index 404022fa..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/crop_riscv.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CROP_RISCV_H
-#define LAYER_CROP_RISCV_H
-
-#include "crop.h"
-
-namespace ncnn {
-
-class Crop_riscv : public Crop
-{
-public:
-    Crop_riscv();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CROP_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_fp16s.h
deleted file mode 100644
index bf4e89ae..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_fp16s.h
+++ /dev/null
@@ -1,99 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_fp16s(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                const __fp16* kptr = weight_data_fp16.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        const __fp16* sptr = m.row<const __fp16>(sy);
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            float val = (float)sptr[sx];
-
-                            int k = y * kernel_w + x;
-
-                            float w = (float)kptr[k];
-
-                            sum += val * w;
-                        }
-                    }
-
-                    kptr += maxk;
-                }
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[j] = (__fp16)sum;
-            }
-
-            outptr += outw;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_pack1ton.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_pack1ton.h
deleted file mode 100644
index ec18f62c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_pack1ton.h
+++ /dev/null
@@ -1,101 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack1ton_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_pack1ton, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
-
-                if (bias_data_ptr)
-                {
-                    _sum = vle32_v_f32m1(bias_data_ptr + p * packn, vl);
-                }
-
-                const float* kptr = (const float*)weight_data_pack1ton + maxk * channels * p * packn;
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        const float* sptr = m.row(sy);
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            float val = sptr[sx];
-
-                            int k = y * kernel_w + x;
-
-                            vfloat32m1_t _w = vle32_v_f32m1(kptr + k * packn, vl);
-                            _sum = vfmacc_vf_f32m1(_sum, val, _w, vl);
-                        }
-                    }
-
-                    kptr += maxk * packn;
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params, vl);
-
-                vse32_v_f32m1(outptr + j * packn, _sum, vl);
-            }
-
-            outptr += outw * packn;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_pack1ton_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_pack1ton_fp16s.h
deleted file mode 100644
index 168c7092..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_pack1ton_fp16s.h
+++ /dev/null
@@ -1,189 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack1ton_fp16s_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat32m2_t _sum = vfmv_v_f_f32m2(0.f, vl);
-
-                if (bias_data_ptr)
-                {
-                    _sum = vle32_v_f32m2(bias_data_ptr + p * packn, vl);
-                }
-
-                const __fp16* kptr = (const __fp16*)weight_data_fp16 + maxk * channels * p * packn;
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        const __fp16* sptr = m.row<const __fp16>(sy);
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            __fp16 val = sptr[sx];
-
-                            int k = y * kernel_w + x;
-
-                            vfloat16m1_t _w = vle16_v_f16m1(kptr + k * packn, vl);
-                            _sum = vfwmacc_vf_f32m2(_sum, val, _w, vl);
-                        }
-                    }
-
-                    kptr += maxk * packn;
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params, vl);
-
-                vse16_v_f16m1(outptr + j * packn, vfncvt_f_f_w_f16m1(_sum, vl), vl);
-            }
-
-            outptr += outw * packn;
-        }
-    }
-}
-
-static void deconvolution_pack1ton_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data_fp16, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const __fp16* bias_data_ptr = bias_data_fp16;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat16m1_t _sum = vfmv_v_f_f16m1(0.f, vl);
-
-                if (bias_data_ptr)
-                {
-                    _sum = vle16_v_f16m1(bias_data_ptr + p * packn, vl);
-                }
-
-                const __fp16* kptr = (const __fp16*)weight_data_fp16 + maxk * channels * p * packn;
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        const __fp16* sptr = m.row<const __fp16>(sy);
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            __fp16 val = sptr[sx];
-
-                            int k = y * kernel_w + x;
-
-                            vfloat16m1_t _w = vle16_v_f16m1(kptr + k * packn, vl);
-                            _sum = vfmacc_vf_f16m1(_sum, val, _w, vl);
-                        }
-                    }
-
-                    kptr += maxk * packn;
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params, vl);
-
-                vse16_v_f16m1(outptr + j * packn, _sum, vl);
-            }
-
-            outptr += outw * packn;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packn.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packn.h
deleted file mode 100644
index 8cab6c3b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packn.h
+++ /dev/null
@@ -1,103 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_packn_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packn, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
-
-                if (bias_data_ptr)
-                {
-                    _sum = vle32_v_f32m1(bias_data_ptr + p * packn, vl);
-                }
-
-                const float* kptr = (const float*)weight_data_packn.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * packn;
-
-                            int k = y * kernel_w + x;
-
-                            for (int l = 0; l < packn; l++)
-                            {
-                                float val = *sptr++;
-                                vfloat32m1_t _w0 = vle32_v_f32m1(kptr + k * packn * packn + packn * l, vl);
-                                _sum = vfmacc_vf_f32m1(_sum, val, _w0, vl);
-                            }
-                        }
-                    }
-
-                    kptr += maxk * packn * packn;
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params, vl);
-
-                vse32_v_f32m1(outptr + j * packn, _sum, vl);
-            }
-
-            outptr += outw * packn;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packn_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packn_fp16s.h
deleted file mode 100644
index 62fbd2eb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packn_fp16s.h
+++ /dev/null
@@ -1,193 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_packn_fp16s_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat32m2_t _sum = vfmv_v_f_f32m2(0.f, vl);
-
-                if (bias_data_ptr)
-                {
-                    _sum = vle32_v_f32m2(bias_data_ptr + p * packn, vl);
-                }
-
-                const __fp16* kptr = (const __fp16*)weight_data_fp16.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const __fp16* sptr = m.row<const __fp16>(sy) + sx * packn;
-
-                            int k = y * kernel_w + x;
-
-                            for (int l = 0; l < packn; l++)
-                            {
-                                __fp16 val = *sptr++;
-                                vfloat16m1_t _w0 = vle16_v_f16m1(kptr + k * packn * packn + packn * l, vl);
-                                _sum = vfwmacc_vf_f32m2(_sum, val, _w0, vl);
-                            }
-                        }
-                    }
-
-                    kptr += maxk * packn * packn;
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params, vl);
-
-                vse16_v_f16m1(outptr + j * packn, vfncvt_f_f_w_f16m1(_sum, vl), vl);
-            }
-
-            outptr += outw * packn;
-        }
-    }
-}
-
-static void deconvolution_packn_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data_fp16, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const __fp16* bias_data_ptr = bias_data_fp16;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                vfloat16m1_t _sum = vfmv_v_f_f16m1(0.f, vl);
-
-                if (bias_data_ptr)
-                {
-                    _sum = vle16_v_f16m1(bias_data_ptr + p * packn, vl);
-                }
-
-                const __fp16* kptr = (const __fp16*)weight_data_fp16.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const __fp16* sptr = m.row<const __fp16>(sy) + sx * packn;
-
-                            int k = y * kernel_w + x;
-
-                            for (int l = 0; l < packn; l++)
-                            {
-                                __fp16 val = *sptr++;
-                                vfloat16m1_t _w0 = vle16_v_f16m1(kptr + k * packn * packn + packn * l, vl);
-                                _sum = vfmacc_vf_f16m1(_sum, val, _w0, vl);
-                            }
-                        }
-                    }
-
-                    kptr += maxk * packn * packn;
-                }
-
-                _sum = activation_ps(_sum, activation_type, activation_params, vl);
-
-                vse16_v_f16m1(outptr + j * packn, _sum, vl);
-            }
-
-            outptr += outw * packn;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packnto1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packnto1.h
deleted file mode 100644
index 2efa9b15..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packnto1.h
+++ /dev/null
@@ -1,104 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_packnto1_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packnto1, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                vfloat32m1_t _sum = vfmv_v_f_f32m1(0.f, vl);
-
-                const float* kptr = (const float*)weight_data_packnto1 + maxk * channels * p * packn;
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * packn;
-
-                            int k = y * kernel_w + x;
-
-                            vfloat32m1_t _val = vle32_v_f32m1(sptr, vl);
-                            vfloat32m1_t _w = vle32_v_f32m1(kptr + k * packn, vl);
-                            _sum = vfmacc_vv_f32m1(_sum, _val, _w, vl);
-                        }
-                    }
-
-                    kptr += maxk * packn;
-                }
-
-                sum = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m1_f32m1(vfloat32m1_t(), _sum, vfmv_s_f_f32m1(vfloat32m1_t(), sum, vl), vl));
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[j] = sum;
-            }
-
-            outptr += outw;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packnto1_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packnto1_fp16s.h
deleted file mode 100644
index ab70100f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_packnto1_fp16s.h
+++ /dev/null
@@ -1,205 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_packnto1_fp16s_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                vfloat32m2_t _sum = vfmv_v_f_f32m2(0.f, vl);
-
-                const __fp16* kptr = (const __fp16*)weight_data_fp16 + maxk * channels * p * packn;
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const __fp16* sptr = m.row<const __fp16>(sy) + sx * packn;
-
-                            int k = y * kernel_w + x;
-
-                            vfloat16m1_t _val = vle16_v_f16m1(sptr, vl);
-                            vfloat16m1_t _w = vle16_v_f16m1(kptr + k * packn, vl);
-                            _sum = vfwmacc_vv_f32m2(_sum, _val, _w, vl);
-                        }
-                    }
-
-                    kptr += maxk * packn;
-                }
-
-#if C906
-                // TODO
-                std::vector<float> ss(packn);
-                vse32_v_f32m2((float*)ss.data(), _sum, vl);
-                for (int i = 0; i < packn; i++)
-                {
-                    sum += ss[i];
-                }
-#else
-                sum = vfmv_f_s_f32m1_f32(vfredusum_vs_f32m2_f32m1(vfloat32m1_t(), _sum, vfmv_s_f_f32m1(vfloat32m1_t(), sum, vl), vl));
-#endif
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[j] = sum;
-            }
-
-            outptr += outw;
-        }
-    }
-}
-
-static void deconvolution_packnto1_fp16sa_rvv(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_fp16, const Mat& bias_data_fp16, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const int maxk = kernel_w * kernel_h;
-
-    const __fp16* bias_data_ptr = bias_data_fp16;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        __fp16* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __fp16 sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                vfloat16m1_t _sum = vfmv_v_f_f16m1(0.f, vl);
-
-                const __fp16* kptr = (const __fp16*)weight_data_fp16 + maxk * channels * p * packn;
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const __fp16* sptr = m.row<const __fp16>(sy) + sx * packn;
-
-                            int k = y * kernel_w + x;
-
-                            vfloat16m1_t _val = vle16_v_f16m1(sptr, vl);
-                            vfloat16m1_t _w = vle16_v_f16m1(kptr + k * packn, vl);
-                            _sum = vfmacc_vv_f16m1(_sum, _val, _w, vl);
-                        }
-                    }
-
-                    kptr += maxk * packn;
-                }
-
-                sum = vfmv_f_s_f16m1_f16(vfredusum_vs_f16m1_f16m1(vfloat16m1_t(), _sum, vfmv_s_f_f16m1(vfloat16m1_t(), sum, vl), vl));
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[j] = sum;
-            }
-
-            outptr += outw;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_riscv.h
deleted file mode 100644
index 57d30349..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolution_riscv.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTION_RISCV_H
-#define LAYER_DECONVOLUTION_RISCV_H
-
-#include "deconvolution.h"
-
-namespace ncnn {
-
-class Deconvolution_riscv : public Deconvolution
-{
-public:
-    Deconvolution_riscv();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Mat weight_data_tm;
-
-    // fp16
-    Mat bias_data_fp16;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTION_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolutiondepthwise_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolutiondepthwise_riscv.h
deleted file mode 100644
index b0c8f7b0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/deconvolutiondepthwise_riscv.h
+++ /dev/null
@@ -1,53 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTIONDEPTHWISE_RISCV_H
-#define LAYER_DECONVOLUTIONDEPTHWISE_RISCV_H
-
-#include "deconvolutiondepthwise.h"
-
-namespace ncnn {
-
-class DeconvolutionDepthWise_riscv : public DeconvolutionDepthWise
-{
-public:
-    DeconvolutionDepthWise_riscv();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    int create_group_ops(const Option& opt);
-#if __riscv_vector && __riscv_zfh
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    std::vector<ncnn::Layer*> group_ops;
-
-    Mat weight_data_tm;
-
-    // fp16
-    Mat bias_data_fp16;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTIONDEPTHWISE_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/dropout_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/dropout_riscv.h
deleted file mode 100644
index 9c28d867..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/dropout_riscv.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// Xavier Hsinyuan is pleased to support the open source community by making
-// ncnn available.
-//
-// Copyright (C) 2021 Xavier Hsinyuan <me@lstlx.com>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this
-// file except in compliance with the License. You may obtain a copy of the
-// License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
-// WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
-// License for the specific language governing permissions and limitations under
-// the License.
-
-#ifndef LAYER_DROPOUT_RISCV_H
-#define LAYER_DROPOUT_RISCV_H
-
-#include "dropout.h"
-
-namespace ncnn {
-
-class Dropout_riscv : public Dropout
-{
-public:
-    Dropout_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DROPOUT_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/flatten_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/flatten_riscv.h
deleted file mode 100644
index 31860340..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/flatten_riscv.h
+++ /dev/null
@@ -1,36 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_FLATTEN_RISCV_H
-#define LAYER_FLATTEN_RISCV_H
-
-#include "flatten.h"
-
-namespace ncnn {
-
-class Flatten_riscv : public Flatten
-{
-public:
-    Flatten_riscv();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_bf16s_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_FLATTEN_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/gelu_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/gelu_riscv.h
deleted file mode 100644
index 8a2e9492..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/gelu_riscv.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Xavier Hsinyuan is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 Xavier Hsinyuan <me@lstlx.com>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GELU_RISCV_H
-#define LAYER_GELU_RISCV_H
-
-#include "gelu.h"
-
-namespace ncnn {
-
-class GELU_riscv : public GELU
-{
-public:
-    GELU_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GELU_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/gemm_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/gemm_riscv.h
deleted file mode 100644
index 6bca092f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/gemm_riscv.h
+++ /dev/null
@@ -1,43 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GEMM_RISCV_H
-#define LAYER_GEMM_RISCV_H
-
-#include "gemm.h"
-
-namespace ncnn {
-
-class Gemm_riscv : public Gemm
-{
-public:
-    Gemm_riscv();
-
-    virtual int create_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-    // public:
-    int nT;
-    size_t vl;
-    Mat AT_data;
-    Mat BT_data;
-    Mat CT_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GEMM_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/gru_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/gru_riscv.h
deleted file mode 100644
index 46bb6245..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/gru_riscv.h
+++ /dev/null
@@ -1,48 +0,0 @@
-// Xavier Hsinyuan is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 Xavier Hsinyuan <me@lstlx.com>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GRU_RISCV_H
-#define LAYER_GRU_RISCV_H
-
-#include "gru.h"
-
-namespace ncnn {
-
-class GRU_riscv : public GRU
-{
-public:
-    GRU_riscv();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-    virtual int create_pipeline(const Option& opt);
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-    int create_pipeline_fp16sa(const Option& opt);
-#endif
-
-public:
-    Mat weight_xc_data_fp16sa;
-    Mat bias_c_data_fp16sa;
-    Mat weight_hc_data_fp16sa;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GRU_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/hardsigmoid_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/hardsigmoid_riscv.h
deleted file mode 100644
index 3c264b31..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/hardsigmoid_riscv.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Xavier Hsinyuan is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 Xavier Hsinyuan <me@lstlx.com>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_HARDSIGMOID_RISCV_H
-#define LAYER_HARDSIGMOID_RISCV_H
-
-#include "hardsigmoid.h"
-
-namespace ncnn {
-
-class HardSigmoid_riscv : public HardSigmoid
-{
-public:
-    HardSigmoid_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_HARDSIGMOID_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/hardswish_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/hardswish_riscv.h
deleted file mode 100644
index cfec7916..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/hardswish_riscv.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// Xavier Hsinyuan is pleased to support the open source community by making
-// ncnn available.
-//
-// Copyright (C) 2021 Xavier Hsinyuan <me@lstlx.com>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this
-// file except in compliance with the License. You may obtain a copy of the
-// License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
-// WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
-// License for the specific language governing permissions and limitations under
-// the License.
-
-#ifndef LAYER_HARDSWISH_RISCV_H
-#define LAYER_HARDSWISH_RISCV_H
-
-#include "hardswish.h"
-
-namespace ncnn {
-
-class HardSwish_riscv : public HardSwish
-{
-public:
-    HardSwish_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_HARDSWISH_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/innerproduct_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/innerproduct_riscv.h
deleted file mode 100644
index d3056d58..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/innerproduct_riscv.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INNERPRODUCT_RISCV_H
-#define LAYER_INNERPRODUCT_RISCV_H
-
-#include "innerproduct.h"
-
-namespace ncnn {
-
-class InnerProduct_riscv : public InnerProduct
-{
-public:
-    InnerProduct_riscv();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Layer* flatten;
-
-    Mat weight_data_tm;
-
-    // fp16
-    Mat bias_data_fp16;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INNERPRODUCT_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/instancenorm_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/instancenorm_riscv.h
deleted file mode 100644
index b0d2e900..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/instancenorm_riscv.h
+++ /dev/null
@@ -1,36 +0,0 @@
-// Xavier Hsinyuan is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 Xavier Hsinyuan <me@lstlx.com>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INSTANCENORM_RISCV_H
-#define LAYER_INSTANCENORM_RISCV_H
-
-#include "instancenorm.h"
-
-namespace ncnn {
-class InstanceNorm_riscv : public InstanceNorm
-{
-public:
-    InstanceNorm_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-} // namespace ncnn
-
-#endif // LAYER_INSTANCENORM_RISCV_H
\ No newline at end of file
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic.h
deleted file mode 100644
index efff2d4d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic.h
+++ /dev/null
@@ -1,261 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static inline void interpolate_cubic(float fx, float* coeffs)
-{
-    const float A = -0.75f;
-
-    float fx0 = fx + 1;
-    float fx1 = fx;
-    float fx2 = 1 - fx;
-    // float fx3 = 2 - fx;
-
-    coeffs[0] = A * fx0 * fx0 * fx0 - 5 * A * fx0 * fx0 + 8 * A * fx0 - 4 * A;
-    coeffs[1] = (A + 2) * fx1 * fx1 * fx1 - (A + 3) * fx1 * fx1 + 1;
-    coeffs[2] = (A + 2) * fx2 * fx2 * fx2 - (A + 3) * fx2 * fx2 + 1;
-    coeffs[3] = 1.f - coeffs[0] - coeffs[1] - coeffs[2];
-}
-
-static void cubic_coeffs(int w, int outw, int* xofs, float* alpha, int align_corner)
-{
-    double scale = (double)w / outw;
-    if (align_corner)
-    {
-        scale = (double)(w - 1) / (outw - 1);
-    }
-
-    for (int dx = 0; dx < outw; dx++)
-    {
-        float fx = (float)((dx + 0.5) * scale - 0.5);
-        if (align_corner)
-        {
-            fx = (float)(dx * scale);
-        }
-
-        int sx = static_cast<int>(floor(fx));
-        fx -= sx;
-
-        interpolate_cubic(fx, alpha + dx * 4);
-
-        if (sx <= -1)
-        {
-            sx = 1;
-            alpha[dx * 4 + 0] = 1.f - alpha[dx * 4 + 3];
-            alpha[dx * 4 + 1] = alpha[dx * 4 + 3];
-            alpha[dx * 4 + 2] = 0.f;
-            alpha[dx * 4 + 3] = 0.f;
-        }
-        if (sx == 0)
-        {
-            sx = 1;
-            alpha[dx * 4 + 0] = alpha[dx * 4 + 0] + alpha[dx * 4 + 1];
-            alpha[dx * 4 + 1] = alpha[dx * 4 + 2];
-            alpha[dx * 4 + 2] = alpha[dx * 4 + 3];
-            alpha[dx * 4 + 3] = 0.f;
-        }
-        if (sx == w - 2)
-        {
-            sx = w - 3;
-            alpha[dx * 4 + 3] = alpha[dx * 4 + 2] + alpha[dx * 4 + 3];
-            alpha[dx * 4 + 2] = alpha[dx * 4 + 1];
-            alpha[dx * 4 + 1] = alpha[dx * 4 + 0];
-            alpha[dx * 4 + 0] = 0.f;
-        }
-        if (sx >= w - 1)
-        {
-            sx = w - 3;
-            alpha[dx * 4 + 3] = 1.f - alpha[dx * 4 + 0];
-            alpha[dx * 4 + 2] = alpha[dx * 4 + 0];
-            alpha[dx * 4 + 1] = 0.f;
-            alpha[dx * 4 + 0] = 0.f;
-        }
-
-        xofs[dx] = sx;
-    }
-}
-
-static void resize_bicubic_image(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    Mat rowsbuf2(w);
-    Mat rowsbuf3(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows1p[dx] = S1p[-1] * a0 + S1p[0] * a1 + S1p[1] * a2 + S1p[2] * a3;
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const float* S0 = src.row(sy - 1);
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows0p[dx] = S0p[-1] * a0 + S0p[0] * a1 + S0p[1] * a2 + S0p[2] * a3;
-                rows1p[dx] = S1p[-1] * a0 + S1p[0] * a1 + S1p[1] * a2 + S1p[2] * a3;
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-        float b2 = beta[2];
-        float b3 = beta[3];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        float* Dp = dst.row(dy);
-        for (int dx = 0; dx < w; dx++)
-        {
-            //             D[x] = rows0[x]*b0 + rows1[x]*b1 + rows2[x]*b2 + rows3[x]*b3;
-            *Dp++ = *rows0p++ * b0 + *rows1p++ * b1 + *rows2p++ * b2 + *rows3p++ * b3;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic_fp16s.h
deleted file mode 100644
index be3e8159..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic_fp16s.h
+++ /dev/null
@@ -1,436 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static inline void interpolate_cubic_fp16sa(float fx, __fp16* coeffs)
-{
-    const float A = -0.75f;
-
-    float fx0 = fx + 1;
-    float fx1 = fx;
-    float fx2 = 1 - fx;
-    // float fx3 = 2 - fx;
-
-    coeffs[0] = (__fp16)(A * fx0 * fx0 * fx0 - 5 * A * fx0 * fx0 + 8 * A * fx0 - 4 * A);
-    coeffs[1] = (__fp16)((A + 2) * fx1 * fx1 * fx1 - (A + 3) * fx1 * fx1 + 1);
-    coeffs[2] = (__fp16)((A + 2) * fx2 * fx2 * fx2 - (A + 3) * fx2 * fx2 + 1);
-    coeffs[3] = (__fp16)(1.f - coeffs[0] - coeffs[1] - coeffs[2]);
-}
-
-static void cubic_coeffs_fp16sa(int w, int outw, int* xofs, __fp16* alpha, int align_corner)
-{
-    double scale = (double)w / outw;
-    if (align_corner)
-    {
-        scale = (double)(w - 1) / (outw - 1);
-    }
-
-    for (int dx = 0; dx < outw; dx++)
-    {
-        float fx = (float)((dx + 0.5) * scale - 0.5);
-        if (align_corner)
-        {
-            fx = static_cast<float>(dx * scale);
-        }
-
-        int sx = static_cast<int>(floor(fx));
-        fx -= sx;
-
-        interpolate_cubic_fp16sa(fx, alpha + dx * 4);
-
-        if (sx <= -1)
-        {
-            sx = 1;
-            alpha[dx * 4 + 0] = (__fp16)(1.f - alpha[dx * 4 + 3]);
-            alpha[dx * 4 + 1] = (__fp16)alpha[dx * 4 + 3];
-            alpha[dx * 4 + 2] = (__fp16)0.f;
-            alpha[dx * 4 + 3] = (__fp16)0.f;
-        }
-        if (sx == 0)
-        {
-            sx = 1;
-            alpha[dx * 4 + 0] = (__fp16)(alpha[dx * 4 + 0] + alpha[dx * 4 + 1]);
-            alpha[dx * 4 + 1] = (__fp16)alpha[dx * 4 + 2];
-            alpha[dx * 4 + 2] = (__fp16)alpha[dx * 4 + 3];
-            alpha[dx * 4 + 3] = (__fp16)0.f;
-        }
-        if (sx == w - 2)
-        {
-            sx = w - 3;
-            alpha[dx * 4 + 3] = (__fp16)(alpha[dx * 4 + 2] + alpha[dx * 4 + 3]);
-            alpha[dx * 4 + 2] = (__fp16)alpha[dx * 4 + 1];
-            alpha[dx * 4 + 1] = (__fp16)alpha[dx * 4 + 0];
-            alpha[dx * 4 + 0] = (__fp16)0.f;
-        }
-        if (sx >= w - 1)
-        {
-            sx = w - 3;
-            alpha[dx * 4 + 3] = (__fp16)(1.f - alpha[dx * 4 + 0]);
-            alpha[dx * 4 + 2] = (__fp16)(alpha[dx * 4 + 0]);
-            alpha[dx * 4 + 1] = (__fp16)0.f;
-            alpha[dx * 4 + 0] = (__fp16)0.f;
-        }
-
-        xofs[dx] = sx;
-    }
-}
-
-static void resize_bicubic_image_fp16s(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    Mat rowsbuf2(w);
-    Mat rowsbuf3(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows3p[dx] = (float)S3p[-1] * a0 + (float)S3p[0] * a1 + (float)S3p[1] * a2 + (float)S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows2p[dx] = (float)S2p[-1] * a0 + (float)S2p[0] * a1 + (float)S2p[1] * a2 + (float)S2p[2] * a3;
-                rows3p[dx] = (float)S3p[-1] * a0 + (float)S3p[0] * a1 + (float)S3p[1] * a2 + (float)S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows1p[dx] = (float)S1p[-1] * a0 + (float)S1p[0] * a1 + (float)S1p[1] * a2 + (float)S1p[2] * a3;
-                rows2p[dx] = (float)S2p[-1] * a0 + (float)S2p[0] * a1 + (float)S2p[1] * a2 + (float)S2p[2] * a3;
-                rows3p[dx] = (float)S3p[-1] * a0 + (float)S3p[0] * a1 + (float)S3p[1] * a2 + (float)S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const __fp16* S0 = src.row<const __fp16>(sy - 1);
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows0p[dx] = (float)S0p[-1] * a0 + (float)S0p[0] * a1 + (float)S0p[1] * a2 + (float)S0p[2] * a3;
-                rows1p[dx] = (float)S1p[-1] * a0 + (float)S1p[0] * a1 + (float)S1p[1] * a2 + (float)S1p[2] * a3;
-                rows2p[dx] = (float)S2p[-1] * a0 + (float)S2p[0] * a1 + (float)S2p[1] * a2 + (float)S2p[2] * a3;
-                rows3p[dx] = (float)S3p[-1] * a0 + (float)S3p[0] * a1 + (float)S3p[1] * a2 + (float)S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-        float b2 = beta[2];
-        float b3 = beta[3];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        __fp16* Dp = dst.row<__fp16>(dy);
-        for (int dx = 0; dx < w; dx++)
-        {
-            // D[x] = rows0[x]*b0 + rows1[x]*b1 + rows2[x]*b2 + rows3[x]*b3;
-            *Dp++ = (__fp16)(*rows0p++ * b0 + *rows1p++ * b1 + *rows2p++ * b2 + *rows3p++ * b3);
-        }
-
-        beta += 4;
-    }
-}
-
-static void resize_bicubic_image_fp16sa(const Mat& src, Mat& dst, __fp16* alpha, int* xofs, __fp16* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)2u);
-    Mat rowsbuf1(w, (size_t)2u);
-    Mat rowsbuf2(w, (size_t)2u);
-    Mat rowsbuf3(w, (size_t)2u);
-    __fp16* rows0 = rowsbuf0;
-    __fp16* rows1 = rowsbuf1;
-    __fp16* rows2 = rowsbuf2;
-    __fp16* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            __fp16* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S3p = S3 + sx;
-
-                __fp16 a0 = alphap[0];
-                __fp16 a1 = alphap[1];
-                __fp16 a2 = alphap[2];
-                __fp16 a3 = alphap[3];
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            __fp16* rows0_old = rows0;
-            __fp16* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows2p = rows2;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                __fp16 a0 = alphap[0];
-                __fp16 a1 = alphap[1];
-                __fp16 a2 = alphap[2];
-                __fp16 a3 = alphap[3];
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            __fp16* rows0_old = rows0;
-            __fp16* rows1_old = rows1;
-            __fp16* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows1p = rows1;
-            __fp16* rows2p = rows2;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                __fp16 a0 = alphap[0];
-                __fp16 a1 = alphap[1];
-                __fp16 a2 = alphap[2];
-                __fp16 a3 = alphap[3];
-                rows1p[dx] = S1p[-1] * a0 + S1p[0] * a1 + S1p[1] * a2 + S1p[2] * a3;
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const __fp16* S0 = src.row<const __fp16>(sy - 1);
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows0p = rows0;
-            __fp16* rows1p = rows1;
-            __fp16* rows2p = rows2;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                __fp16 a0 = alphap[0];
-                __fp16 a1 = alphap[1];
-                __fp16 a2 = alphap[2];
-                __fp16 a3 = alphap[3];
-                rows0p[dx] = S0p[-1] * a0 + S0p[0] * a1 + S0p[1] * a2 + S0p[2] * a3;
-                rows1p[dx] = S1p[-1] * a0 + S1p[0] * a1 + S1p[1] * a2 + S1p[2] * a3;
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        __fp16 b0 = beta[0];
-        __fp16 b1 = beta[1];
-        __fp16 b2 = beta[2];
-        __fp16 b3 = beta[3];
-
-        __fp16* rows0p = rows0;
-        __fp16* rows1p = rows1;
-        __fp16* rows2p = rows2;
-        __fp16* rows3p = rows3;
-        __fp16* Dp = dst.row<__fp16>(dy);
-        for (int dx = 0; dx < w; dx++)
-        {
-            // D[x] = rows0[x]*b0 + rows1[x]*b1 + rows2[x]*b2 + rows3[x]*b3;
-            *Dp++ = (*rows0p++ * b0 + *rows1p++ * b1 + *rows2p++ * b2 + *rows3p++ * b3);
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic_packn.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic_packn.h
deleted file mode 100644
index b19af95f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic_packn.h
+++ /dev/null
@@ -1,242 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bicubic_image_packn(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)packn * 4u, packn);
-    Mat rowsbuf1(w, (size_t)packn * 4u, packn);
-    Mat rowsbuf2(w, (size_t)packn * 4u, packn);
-    Mat rowsbuf3(w, (size_t)packn * 4u, packn);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const float* S3p = S3 + sx;
-
-                vfloat32m1_t _S30 = vle32_v_f32m1(S3p - packn, vl);
-                vfloat32m1_t _S31 = vle32_v_f32m1(S3p, vl);
-                vfloat32m1_t _S32 = vle32_v_f32m1(S3p + packn, vl);
-                vfloat32m1_t _S33 = vle32_v_f32m1(S3p + packn * 2, vl);
-                vfloat32m1_t _rows3 = vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmul_vf_f32m1(_S30, alphap[0], vl), alphap[1], _S31, vl), alphap[2], _S32, vl), alphap[3], _S33, vl);
-
-                vse32_v_f32m1(rows3p + dx * packn, _rows3, vl);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                vfloat32m1_t _S20 = vle32_v_f32m1(S2p - packn, vl);
-                vfloat32m1_t _S21 = vle32_v_f32m1(S2p, vl);
-                vfloat32m1_t _S22 = vle32_v_f32m1(S2p + packn, vl);
-                vfloat32m1_t _S23 = vle32_v_f32m1(S2p + packn * 2, vl);
-                vfloat32m1_t _S30 = vle32_v_f32m1(S3p - packn, vl);
-                vfloat32m1_t _S31 = vle32_v_f32m1(S3p, vl);
-                vfloat32m1_t _S32 = vle32_v_f32m1(S3p + packn, vl);
-                vfloat32m1_t _S33 = vle32_v_f32m1(S3p + packn * 2, vl);
-                vfloat32m1_t _rows2 = vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmul_vf_f32m1(_S20, alphap[0], vl), alphap[1], _S21, vl), alphap[2], _S22, vl), alphap[3], _S23, vl);
-                vfloat32m1_t _rows3 = vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmul_vf_f32m1(_S30, alphap[0], vl), alphap[1], _S31, vl), alphap[2], _S32, vl), alphap[3], _S33, vl);
-
-                vse32_v_f32m1(rows2p + dx * packn, _rows2, vl);
-                vse32_v_f32m1(rows3p + dx * packn, _rows3, vl);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                vfloat32m1_t _S10 = vle32_v_f32m1(S1p - packn, vl);
-                vfloat32m1_t _S11 = vle32_v_f32m1(S1p, vl);
-                vfloat32m1_t _S12 = vle32_v_f32m1(S1p + packn, vl);
-                vfloat32m1_t _S13 = vle32_v_f32m1(S1p + packn * 2, vl);
-                vfloat32m1_t _S20 = vle32_v_f32m1(S2p - packn, vl);
-                vfloat32m1_t _S21 = vle32_v_f32m1(S2p, vl);
-                vfloat32m1_t _S22 = vle32_v_f32m1(S2p + packn, vl);
-                vfloat32m1_t _S23 = vle32_v_f32m1(S2p + packn * 2, vl);
-                vfloat32m1_t _S30 = vle32_v_f32m1(S3p - packn, vl);
-                vfloat32m1_t _S31 = vle32_v_f32m1(S3p, vl);
-                vfloat32m1_t _S32 = vle32_v_f32m1(S3p + packn, vl);
-                vfloat32m1_t _S33 = vle32_v_f32m1(S3p + packn * 2, vl);
-                vfloat32m1_t _rows1 = vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmul_vf_f32m1(_S10, alphap[0], vl), alphap[1], _S11, vl), alphap[2], _S12, vl), alphap[3], _S13, vl);
-                vfloat32m1_t _rows2 = vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmul_vf_f32m1(_S20, alphap[0], vl), alphap[1], _S21, vl), alphap[2], _S22, vl), alphap[3], _S23, vl);
-                vfloat32m1_t _rows3 = vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmul_vf_f32m1(_S30, alphap[0], vl), alphap[1], _S31, vl), alphap[2], _S32, vl), alphap[3], _S33, vl);
-
-                vse32_v_f32m1(rows1p + dx * packn, _rows1, vl);
-                vse32_v_f32m1(rows2p + dx * packn, _rows2, vl);
-                vse32_v_f32m1(rows3p + dx * packn, _rows3, vl);
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const float* S0 = src.row(sy - 1);
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                vfloat32m1_t _S00 = vle32_v_f32m1(S0p - packn, vl);
-                vfloat32m1_t _S01 = vle32_v_f32m1(S0p, vl);
-                vfloat32m1_t _S02 = vle32_v_f32m1(S0p + packn, vl);
-                vfloat32m1_t _S03 = vle32_v_f32m1(S0p + packn * 2, vl);
-                vfloat32m1_t _S10 = vle32_v_f32m1(S1p - packn, vl);
-                vfloat32m1_t _S11 = vle32_v_f32m1(S1p, vl);
-                vfloat32m1_t _S12 = vle32_v_f32m1(S1p + packn, vl);
-                vfloat32m1_t _S13 = vle32_v_f32m1(S1p + packn * 2, vl);
-                vfloat32m1_t _S20 = vle32_v_f32m1(S2p - packn, vl);
-                vfloat32m1_t _S21 = vle32_v_f32m1(S2p, vl);
-                vfloat32m1_t _S22 = vle32_v_f32m1(S2p + packn, vl);
-                vfloat32m1_t _S23 = vle32_v_f32m1(S2p + packn * 2, vl);
-                vfloat32m1_t _S30 = vle32_v_f32m1(S3p - packn, vl);
-                vfloat32m1_t _S31 = vle32_v_f32m1(S3p, vl);
-                vfloat32m1_t _S32 = vle32_v_f32m1(S3p + packn, vl);
-                vfloat32m1_t _S33 = vle32_v_f32m1(S3p + packn * 2, vl);
-                vfloat32m1_t _rows0 = vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmul_vf_f32m1(_S00, alphap[0], vl), alphap[1], _S01, vl), alphap[2], _S02, vl), alphap[3], _S03, vl);
-                vfloat32m1_t _rows1 = vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmul_vf_f32m1(_S10, alphap[0], vl), alphap[1], _S11, vl), alphap[2], _S12, vl), alphap[3], _S13, vl);
-                vfloat32m1_t _rows2 = vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmul_vf_f32m1(_S20, alphap[0], vl), alphap[1], _S21, vl), alphap[2], _S22, vl), alphap[3], _S23, vl);
-                vfloat32m1_t _rows3 = vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmul_vf_f32m1(_S30, alphap[0], vl), alphap[1], _S31, vl), alphap[2], _S32, vl), alphap[3], _S33, vl);
-
-                vse32_v_f32m1(rows0p + dx * packn, _rows0, vl);
-                vse32_v_f32m1(rows1p + dx * packn, _rows1, vl);
-                vse32_v_f32m1(rows2p + dx * packn, _rows2, vl);
-                vse32_v_f32m1(rows3p + dx * packn, _rows3, vl);
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-        float b2 = beta[2];
-        float b3 = beta[3];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        float* Dp = dst.row(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            vfloat32m1_t _rows0 = vle32_v_f32m1(rows0p, vl);
-            vfloat32m1_t _rows1 = vle32_v_f32m1(rows1p, vl);
-            vfloat32m1_t _rows2 = vle32_v_f32m1(rows2p, vl);
-            vfloat32m1_t _rows3 = vle32_v_f32m1(rows3p, vl);
-
-            vfloat32m1_t _Dp = vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmacc_vf_f32m1(vfmul_vf_f32m1(_rows0, b0, vl), b1, _rows1, vl), b2, _rows2, vl), b3, _rows3, vl);
-
-            vse32_v_f32m1(Dp, _Dp, vl);
-
-            Dp += packn;
-            rows0p += packn;
-            rows1p += packn;
-            rows2p += packn;
-            rows3p += packn;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic_packn_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic_packn_fp16s.h
deleted file mode 100644
index f87d5bb5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bicubic_packn_fp16s.h
+++ /dev/null
@@ -1,471 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bicubic_image_packn_fp16s(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)packn * 4u, packn);
-    Mat rowsbuf1(w, (size_t)packn * 4u, packn);
-    Mat rowsbuf2(w, (size_t)packn * 4u, packn);
-    Mat rowsbuf3(w, (size_t)packn * 4u, packn);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const __fp16* S3p = S3 + sx;
-
-                vfloat16m1_t _S30 = vle16_v_f16m1(S3p - packn, vl);
-                vfloat16m1_t _S31 = vle16_v_f16m1(S3p, vl);
-                vfloat16m1_t _S32 = vle16_v_f16m1(S3p + packn, vl);
-                vfloat16m1_t _S33 = vle16_v_f16m1(S3p + packn * 2, vl);
-                vfloat32m2_t _rows3 = vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmul_vf_f32m2(_S30, alphap[0], vl), alphap[1], _S31, vl), alphap[2], _S32, vl), alphap[3], _S33, vl);
-
-                vse32_v_f32m2(rows3p + dx * packn, _rows3, vl);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                vfloat16m1_t _S20 = vle16_v_f16m1(S2p - packn, vl);
-                vfloat16m1_t _S21 = vle16_v_f16m1(S2p, vl);
-                vfloat16m1_t _S22 = vle16_v_f16m1(S2p + packn, vl);
-                vfloat16m1_t _S23 = vle16_v_f16m1(S2p + packn * 2, vl);
-                vfloat16m1_t _S30 = vle16_v_f16m1(S3p - packn, vl);
-                vfloat16m1_t _S31 = vle16_v_f16m1(S3p, vl);
-                vfloat16m1_t _S32 = vle16_v_f16m1(S3p + packn, vl);
-                vfloat16m1_t _S33 = vle16_v_f16m1(S3p + packn * 2, vl);
-                vfloat32m2_t _rows2 = vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmul_vf_f32m2(_S20, alphap[0], vl), alphap[1], _S21, vl), alphap[2], _S22, vl), alphap[3], _S23, vl);
-                vfloat32m2_t _rows3 = vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmul_vf_f32m2(_S30, alphap[0], vl), alphap[1], _S31, vl), alphap[2], _S32, vl), alphap[3], _S33, vl);
-
-                vse32_v_f32m2(rows2p + dx * packn, _rows2, vl);
-                vse32_v_f32m2(rows3p + dx * packn, _rows3, vl);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                vfloat16m1_t _S10 = vle16_v_f16m1(S1p - packn, vl);
-                vfloat16m1_t _S11 = vle16_v_f16m1(S1p, vl);
-                vfloat16m1_t _S12 = vle16_v_f16m1(S1p + packn, vl);
-                vfloat16m1_t _S13 = vle16_v_f16m1(S1p + packn * 2, vl);
-                vfloat16m1_t _S20 = vle16_v_f16m1(S2p - packn, vl);
-                vfloat16m1_t _S21 = vle16_v_f16m1(S2p, vl);
-                vfloat16m1_t _S22 = vle16_v_f16m1(S2p + packn, vl);
-                vfloat16m1_t _S23 = vle16_v_f16m1(S2p + packn * 2, vl);
-                vfloat16m1_t _S30 = vle16_v_f16m1(S3p - packn, vl);
-                vfloat16m1_t _S31 = vle16_v_f16m1(S3p, vl);
-                vfloat16m1_t _S32 = vle16_v_f16m1(S3p + packn, vl);
-                vfloat16m1_t _S33 = vle16_v_f16m1(S3p + packn * 2, vl);
-                vfloat32m2_t _rows1 = vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmul_vf_f32m2(_S10, alphap[0], vl), alphap[1], _S11, vl), alphap[2], _S12, vl), alphap[3], _S13, vl);
-                vfloat32m2_t _rows2 = vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmul_vf_f32m2(_S20, alphap[0], vl), alphap[1], _S21, vl), alphap[2], _S22, vl), alphap[3], _S23, vl);
-                vfloat32m2_t _rows3 = vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmul_vf_f32m2(_S30, alphap[0], vl), alphap[1], _S31, vl), alphap[2], _S32, vl), alphap[3], _S33, vl);
-
-                vse32_v_f32m2(rows1p + dx * packn, _rows1, vl);
-                vse32_v_f32m2(rows2p + dx * packn, _rows2, vl);
-                vse32_v_f32m2(rows3p + dx * packn, _rows3, vl);
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const __fp16* S0 = src.row<const __fp16>(sy - 1);
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                vfloat16m1_t _S00 = vle16_v_f16m1(S0p - packn, vl);
-                vfloat16m1_t _S01 = vle16_v_f16m1(S0p, vl);
-                vfloat16m1_t _S02 = vle16_v_f16m1(S0p + packn, vl);
-                vfloat16m1_t _S03 = vle16_v_f16m1(S0p + packn * 2, vl);
-                vfloat16m1_t _S10 = vle16_v_f16m1(S1p - packn, vl);
-                vfloat16m1_t _S11 = vle16_v_f16m1(S1p, vl);
-                vfloat16m1_t _S12 = vle16_v_f16m1(S1p + packn, vl);
-                vfloat16m1_t _S13 = vle16_v_f16m1(S1p + packn * 2, vl);
-                vfloat16m1_t _S20 = vle16_v_f16m1(S2p - packn, vl);
-                vfloat16m1_t _S21 = vle16_v_f16m1(S2p, vl);
-                vfloat16m1_t _S22 = vle16_v_f16m1(S2p + packn, vl);
-                vfloat16m1_t _S23 = vle16_v_f16m1(S2p + packn * 2, vl);
-                vfloat16m1_t _S30 = vle16_v_f16m1(S3p - packn, vl);
-                vfloat16m1_t _S31 = vle16_v_f16m1(S3p, vl);
-                vfloat16m1_t _S32 = vle16_v_f16m1(S3p + packn, vl);
-                vfloat16m1_t _S33 = vle16_v_f16m1(S3p + packn * 2, vl);
-                vfloat32m2_t _rows0 = vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmul_vf_f32m2(_S00, alphap[0], vl), alphap[1], _S01, vl), alphap[2], _S02, vl), alphap[3], _S03, vl);
-                vfloat32m2_t _rows1 = vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmul_vf_f32m2(_S10, alphap[0], vl), alphap[1], _S11, vl), alphap[2], _S12, vl), alphap[3], _S13, vl);
-                vfloat32m2_t _rows2 = vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmul_vf_f32m2(_S20, alphap[0], vl), alphap[1], _S21, vl), alphap[2], _S22, vl), alphap[3], _S23, vl);
-                vfloat32m2_t _rows3 = vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmacc_vf_f32m2(vfwmul_vf_f32m2(_S30, alphap[0], vl), alphap[1], _S31, vl), alphap[2], _S32, vl), alphap[3], _S33, vl);
-
-                vse32_v_f32m2(rows0p + dx * packn, _rows0, vl);
-                vse32_v_f32m2(rows1p + dx * packn, _rows1, vl);
-                vse32_v_f32m2(rows2p + dx * packn, _rows2, vl);
-                vse32_v_f32m2(rows3p + dx * packn, _rows3, vl);
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-        float b2 = beta[2];
-        float b3 = beta[3];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        __fp16* Dp = dst.row<__fp16>(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            vfloat32m2_t _rows0 = vle32_v_f32m2(rows0p, vl);
-            vfloat32m2_t _rows1 = vle32_v_f32m2(rows1p, vl);
-            vfloat32m2_t _rows2 = vle32_v_f32m2(rows2p, vl);
-            vfloat32m2_t _rows3 = vle32_v_f32m2(rows3p, vl);
-
-            vfloat32m2_t _Dp = vfmacc_vf_f32m2(vfmacc_vf_f32m2(vfmacc_vf_f32m2(vfmul_vf_f32m2(_rows0, b0, vl), b1, _rows1, vl), b2, _rows2, vl), b3, _rows3, vl);
-
-            vse16_v_f16m1(Dp, vfncvt_f_f_w_f16m1(_Dp, vl), vl);
-
-            Dp += packn;
-            rows0p += packn;
-            rows1p += packn;
-            rows2p += packn;
-            rows3p += packn;
-        }
-
-        beta += 4;
-    }
-}
-
-static void resize_bicubic_image_packn_fp16sa(const Mat& src, Mat& dst, __fp16* alpha, int* xofs, __fp16* beta, int* yofs)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)packn * 2u, packn);
-    Mat rowsbuf1(w, (size_t)packn * 2u, packn);
-    Mat rowsbuf2(w, (size_t)packn * 2u, packn);
-    Mat rowsbuf3(w, (size_t)packn * 2u, packn);
-    __fp16* rows0 = rowsbuf0;
-    __fp16* rows1 = rowsbuf1;
-    __fp16* rows2 = rowsbuf2;
-    __fp16* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            __fp16* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const __fp16* S3p = S3 + sx;
-
-                vfloat16m1_t _S30 = vle16_v_f16m1(S3p - packn, vl);
-                vfloat16m1_t _S31 = vle16_v_f16m1(S3p, vl);
-                vfloat16m1_t _S32 = vle16_v_f16m1(S3p + packn, vl);
-                vfloat16m1_t _S33 = vle16_v_f16m1(S3p + packn * 2, vl);
-                vfloat16m1_t _rows3 = vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmul_vf_f16m1(_S30, alphap[0], vl), alphap[1], _S31, vl), alphap[2], _S32, vl), alphap[3], _S33, vl);
-
-                vse16_v_f16m1(rows3p + dx * packn, _rows3, vl);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            __fp16* rows0_old = rows0;
-            __fp16* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows2p = rows2;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                vfloat16m1_t _S20 = vle16_v_f16m1(S2p - packn, vl);
-                vfloat16m1_t _S21 = vle16_v_f16m1(S2p, vl);
-                vfloat16m1_t _S22 = vle16_v_f16m1(S2p + packn, vl);
-                vfloat16m1_t _S23 = vle16_v_f16m1(S2p + packn * 2, vl);
-                vfloat16m1_t _S30 = vle16_v_f16m1(S3p - packn, vl);
-                vfloat16m1_t _S31 = vle16_v_f16m1(S3p, vl);
-                vfloat16m1_t _S32 = vle16_v_f16m1(S3p + packn, vl);
-                vfloat16m1_t _S33 = vle16_v_f16m1(S3p + packn * 2, vl);
-                vfloat16m1_t _rows2 = vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmul_vf_f16m1(_S20, alphap[0], vl), alphap[1], _S21, vl), alphap[2], _S22, vl), alphap[3], _S23, vl);
-                vfloat16m1_t _rows3 = vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmul_vf_f16m1(_S30, alphap[0], vl), alphap[1], _S31, vl), alphap[2], _S32, vl), alphap[3], _S33, vl);
-
-                vse16_v_f16m1(rows2p + dx * packn, _rows2, vl);
-                vse16_v_f16m1(rows3p + dx * packn, _rows3, vl);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            __fp16* rows0_old = rows0;
-            __fp16* rows1_old = rows1;
-            __fp16* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows1p = rows1;
-            __fp16* rows2p = rows2;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                vfloat16m1_t _S10 = vle16_v_f16m1(S1p - packn, vl);
-                vfloat16m1_t _S11 = vle16_v_f16m1(S1p, vl);
-                vfloat16m1_t _S12 = vle16_v_f16m1(S1p + packn, vl);
-                vfloat16m1_t _S13 = vle16_v_f16m1(S1p + packn * 2, vl);
-                vfloat16m1_t _S20 = vle16_v_f16m1(S2p - packn, vl);
-                vfloat16m1_t _S21 = vle16_v_f16m1(S2p, vl);
-                vfloat16m1_t _S22 = vle16_v_f16m1(S2p + packn, vl);
-                vfloat16m1_t _S23 = vle16_v_f16m1(S2p + packn * 2, vl);
-                vfloat16m1_t _S30 = vle16_v_f16m1(S3p - packn, vl);
-                vfloat16m1_t _S31 = vle16_v_f16m1(S3p, vl);
-                vfloat16m1_t _S32 = vle16_v_f16m1(S3p + packn, vl);
-                vfloat16m1_t _S33 = vle16_v_f16m1(S3p + packn * 2, vl);
-                vfloat16m1_t _rows1 = vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmul_vf_f16m1(_S10, alphap[0], vl), alphap[1], _S11, vl), alphap[2], _S12, vl), alphap[3], _S13, vl);
-                vfloat16m1_t _rows2 = vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmul_vf_f16m1(_S20, alphap[0], vl), alphap[1], _S21, vl), alphap[2], _S22, vl), alphap[3], _S23, vl);
-                vfloat16m1_t _rows3 = vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmul_vf_f16m1(_S30, alphap[0], vl), alphap[1], _S31, vl), alphap[2], _S32, vl), alphap[3], _S33, vl);
-
-                vse16_v_f16m1(rows1p + dx * packn, _rows1, vl);
-                vse16_v_f16m1(rows2p + dx * packn, _rows2, vl);
-                vse16_v_f16m1(rows3p + dx * packn, _rows3, vl);
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const __fp16* S0 = src.row<const __fp16>(sy - 1);
-            const __fp16* S1 = src.row<const __fp16>(sy);
-            const __fp16* S2 = src.row<const __fp16>(sy + 1);
-            const __fp16* S3 = src.row<const __fp16>(sy + 2);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows0p = rows0;
-            __fp16* rows1p = rows1;
-            __fp16* rows2p = rows2;
-            __fp16* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-                const __fp16* S2p = S2 + sx;
-                const __fp16* S3p = S3 + sx;
-
-                vfloat16m1_t _S00 = vle16_v_f16m1(S0p - packn, vl);
-                vfloat16m1_t _S01 = vle16_v_f16m1(S0p, vl);
-                vfloat16m1_t _S02 = vle16_v_f16m1(S0p + packn, vl);
-                vfloat16m1_t _S03 = vle16_v_f16m1(S0p + packn * 2, vl);
-                vfloat16m1_t _S10 = vle16_v_f16m1(S1p - packn, vl);
-                vfloat16m1_t _S11 = vle16_v_f16m1(S1p, vl);
-                vfloat16m1_t _S12 = vle16_v_f16m1(S1p + packn, vl);
-                vfloat16m1_t _S13 = vle16_v_f16m1(S1p + packn * 2, vl);
-                vfloat16m1_t _S20 = vle16_v_f16m1(S2p - packn, vl);
-                vfloat16m1_t _S21 = vle16_v_f16m1(S2p, vl);
-                vfloat16m1_t _S22 = vle16_v_f16m1(S2p + packn, vl);
-                vfloat16m1_t _S23 = vle16_v_f16m1(S2p + packn * 2, vl);
-                vfloat16m1_t _S30 = vle16_v_f16m1(S3p - packn, vl);
-                vfloat16m1_t _S31 = vle16_v_f16m1(S3p, vl);
-                vfloat16m1_t _S32 = vle16_v_f16m1(S3p + packn, vl);
-                vfloat16m1_t _S33 = vle16_v_f16m1(S3p + packn * 2, vl);
-                vfloat16m1_t _rows0 = vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmul_vf_f16m1(_S00, alphap[0], vl), alphap[1], _S01, vl), alphap[2], _S02, vl), alphap[3], _S03, vl);
-                vfloat16m1_t _rows1 = vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmul_vf_f16m1(_S10, alphap[0], vl), alphap[1], _S11, vl), alphap[2], _S12, vl), alphap[3], _S13, vl);
-                vfloat16m1_t _rows2 = vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmul_vf_f16m1(_S20, alphap[0], vl), alphap[1], _S21, vl), alphap[2], _S22, vl), alphap[3], _S23, vl);
-                vfloat16m1_t _rows3 = vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmul_vf_f16m1(_S30, alphap[0], vl), alphap[1], _S31, vl), alphap[2], _S32, vl), alphap[3], _S33, vl);
-
-                vse16_v_f16m1(rows0p + dx * packn, _rows0, vl);
-                vse16_v_f16m1(rows1p + dx * packn, _rows1, vl);
-                vse16_v_f16m1(rows2p + dx * packn, _rows2, vl);
-                vse16_v_f16m1(rows3p + dx * packn, _rows3, vl);
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        __fp16 b0 = beta[0];
-        __fp16 b1 = beta[1];
-        __fp16 b2 = beta[2];
-        __fp16 b3 = beta[3];
-
-        __fp16* rows0p = rows0;
-        __fp16* rows1p = rows1;
-        __fp16* rows2p = rows2;
-        __fp16* rows3p = rows3;
-        __fp16* Dp = dst.row<__fp16>(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            vfloat16m1_t _rows0 = vle16_v_f16m1(rows0p, vl);
-            vfloat16m1_t _rows1 = vle16_v_f16m1(rows1p, vl);
-            vfloat16m1_t _rows2 = vle16_v_f16m1(rows2p, vl);
-            vfloat16m1_t _rows3 = vle16_v_f16m1(rows3p, vl);
-
-            vfloat16m1_t _Dp = vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmacc_vf_f16m1(vfmul_vf_f16m1(_rows0, b0, vl), b1, _rows1, vl), b2, _rows2, vl), b3, _rows3, vl);
-
-            vse16_v_f16m1(Dp, _Dp, vl);
-
-            Dp += packn;
-            rows0p += packn;
-            rows1p += packn;
-            rows2p += packn;
-            rows3p += packn;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear.h
deleted file mode 100644
index ffd613a6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear.h
+++ /dev/null
@@ -1,222 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void linear_coeffs(int w, int outw, int* xofs, float* alpha, int align_corner)
-{
-    double scale = (double)w / outw;
-    if (align_corner)
-    {
-        scale = (double)(w - 1) / (outw - 1);
-    }
-
-    for (int dx = 0; dx < outw; dx++)
-    {
-        float fx = (float)((dx + 0.5) * scale - 0.5);
-        if (align_corner)
-        {
-            fx = (float)(dx * scale);
-        }
-
-        int sx = floor(fx);
-        fx -= sx;
-
-        if (sx < 0)
-        {
-            sx = 0;
-            fx = 0.f;
-        }
-        if (sx >= w - 1)
-        {
-            sx = w - 2;
-            fx = 1.f;
-        }
-
-        xofs[dx] = sx;
-
-        alpha[dx * 2] = 1.f - fx;
-        alpha[dx * 2 + 1] = fx;
-    }
-}
-
-static void resize_bilinear_image(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-
-#if __riscv_vector
-            const unsigned int* pxofs = (const unsigned int*)xofs;
-            int n = w;
-            while (n > 0)
-            {
-                size_t vl = vsetvl_e32m4(n);
-
-                vuint32m4_t _sx = vmul_vx_u32m4(vle32_v_u32m4(pxofs, vl), sizeof(float), vl);
-
-                vfloat32m4_t _S1p0;
-                vfloat32m4_t _S1p1;
-                vloxseg2ei32_v_f32m4(&_S1p0, &_S1p1, S1, _sx, vl);
-
-                vfloat32m4_t _a0;
-                vfloat32m4_t _a1;
-                vlseg2e32_v_f32m4(&_a0, &_a1, alphap, vl);
-
-                vfloat32m4_t _rows1 = vfmacc_vv_f32m4(vfmul_vv_f32m4(_S1p0, _a0, vl), _S1p1, _a1, vl);
-
-                vse32_v_f32m4(rows1p, _rows1, vl);
-
-                pxofs += vl;
-                alphap += vl * 2;
-                rows1p += vl;
-                n -= vl;
-            }
-#else  // __riscv_vector
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows1p[dx] = S1p[0] * a0 + S1p[1] * a1;
-
-                alphap += 2;
-            }
-#endif // __riscv_vector
-        }
-        else
-        {
-            // hresize two rows
-            const float* S0 = src.row(sy);
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-
-#if __riscv_vector
-            const unsigned int* pxofs = (const unsigned int*)xofs;
-            int n = w;
-            while (n > 0)
-            {
-                size_t vl = vsetvl_e32m4(n);
-
-                vuint32m4_t _sx = vmul_vx_u32m4(vle32_v_u32m4(pxofs, vl), sizeof(float), vl);
-
-                vfloat32m4_t _S0p0;
-                vfloat32m4_t _S0p1;
-                vfloat32m4_t _S1p0;
-                vfloat32m4_t _S1p1;
-
-                vloxseg2ei32_v_f32m4(&_S0p0, &_S0p1, S0, _sx, vl);
-                vloxseg2ei32_v_f32m4(&_S1p0, &_S1p1, S1, _sx, vl);
-
-                vfloat32m4_t _a0;
-                vfloat32m4_t _a1;
-                vlseg2e32_v_f32m4(&_a0, &_a1, alphap, vl);
-
-                vfloat32m4_t _rows0 = vfmacc_vv_f32m4(vfmul_vv_f32m4(_S0p0, _a0, vl), _S0p1, _a1, vl);
-                vfloat32m4_t _rows1 = vfmacc_vv_f32m4(vfmul_vv_f32m4(_S1p0, _a0, vl), _S1p1, _a1, vl);
-
-                vse32_v_f32m4(rows0p, _rows0, vl);
-                vse32_v_f32m4(rows1p, _rows1, vl);
-
-                pxofs += vl;
-                alphap += vl * 2;
-                rows0p += vl;
-                rows1p += vl;
-                n -= vl;
-            }
-#else  // __riscv_vector
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows0p[dx] = S0p[0] * a0 + S0p[1] * a1;
-                rows1p[dx] = S1p[0] * a0 + S1p[1] * a1;
-
-                alphap += 2;
-            }
-#endif // __riscv_vector
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* Dp = dst.row(dy);
-
-#if __riscv_vector
-        int n = w;
-        while (n > 0)
-        {
-            size_t vl = vsetvl_e32m8(n);
-
-            vfloat32m8_t _rows0 = vle32_v_f32m8(rows0p, vl);
-            vfloat32m8_t _rows1 = vle32_v_f32m8(rows1p, vl);
-
-            vfloat32m8_t _Dp = vfmacc_vf_f32m8(vfmul_vf_f32m8(_rows0, b0, vl), b1, _rows1, vl);
-
-            vse32_v_f32m8(Dp, _Dp, vl);
-
-            Dp += vl;
-            rows0p += vl;
-            rows1p += vl;
-            n -= vl;
-        }
-#else  // __riscv_vector
-        for (int i = 0; i < w; i++)
-        {
-            //             D[x] = rows0[x]*b0 + rows1[x]*b1;
-            *Dp++ = *rows0p++ * b0 + *rows1p++ * b1;
-        }
-#endif // __riscv_vector
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear_fp16s.h
deleted file mode 100644
index 318f36f8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear_fp16s.h
+++ /dev/null
@@ -1,252 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void linear_coeffs_fp16sa(int w, int outw, int* xofs, __fp16* alpha, int align_corner)
-{
-    double scale = (double)w / outw;
-    if (align_corner)
-    {
-        scale = (double)(w - 1) / (outw - 1);
-    }
-
-    for (int dx = 0; dx < outw; dx++)
-    {
-        float fx = (float)((dx + 0.5) * scale - 0.5);
-        if (align_corner)
-        {
-            fx = static_cast<float>(dx * scale);
-        }
-
-        int sx = floor(fx);
-        fx -= sx;
-
-        if (sx < 0)
-        {
-            sx = 0;
-            fx = 0.f;
-        }
-        if (sx >= w - 1)
-        {
-            sx = w - 2;
-            fx = 1.f;
-        }
-
-        xofs[dx] = sx;
-
-        alpha[dx * 2] = (__fp16)(1.f - fx);
-        alpha[dx * 2 + 1] = (__fp16)fx;
-    }
-}
-
-static void resize_bilinear_image_fp16s(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows1p[dx] = (float)S1p[0] * a0 + (float)S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const __fp16* S0 = src.row<const __fp16>(sy);
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows0p[dx] = (float)S0p[0] * a0 + (float)S0p[1] * a1;
-                rows1p[dx] = (float)S1p[0] * a0 + (float)S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        __fp16* Dp = dst.row<__fp16>(dy);
-
-        int n = w;
-        while (n > 0)
-        {
-            size_t vl = vsetvl_e16m4(n);
-
-            vfloat32m8_t _rows0 = vle32_v_f32m8(rows0p, vl);
-            vfloat32m8_t _rows1 = vle32_v_f32m8(rows1p, vl);
-
-            vfloat32m8_t _Dp = vfmacc_vf_f32m8(vfmul_vf_f32m8(_rows0, b0, vl), b1, _rows1, vl);
-
-            vse16_v_f16m4(Dp, vfncvt_f_f_w_f16m4(_Dp, vl), vl);
-
-            Dp += vl;
-            rows0p += vl;
-            rows1p += vl;
-            n -= vl;
-        }
-
-        beta += 2;
-    }
-}
-
-static void resize_bilinear_image_fp16sa(const Mat& src, Mat& dst, __fp16* alpha, int* xofs, __fp16* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)2u);
-    Mat rowsbuf1(w, (size_t)2u);
-    __fp16* rows0 = rowsbuf0;
-    __fp16* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            __fp16* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S1p = S1 + sx;
-
-                __fp16 a0 = alphap[0];
-                __fp16 a1 = alphap[1];
-                rows1p[dx] = S1p[0] * a0 + S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const __fp16* S0 = src.row<const __fp16>(sy);
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows0p = rows0;
-            __fp16* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-
-                __fp16 a0 = alphap[0];
-                __fp16 a1 = alphap[1];
-                rows0p[dx] = S0p[0] * a0 + S0p[1] * a1;
-                rows1p[dx] = S1p[0] * a0 + S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        __fp16 b0 = beta[0];
-        __fp16 b1 = beta[1];
-
-        __fp16* rows0p = rows0;
-        __fp16* rows1p = rows1;
-        __fp16* Dp = dst.row<__fp16>(dy);
-
-        int n = w;
-        while (n > 0)
-        {
-            size_t vl = vsetvl_e16m8(n);
-
-            vfloat16m8_t _rows0 = vle16_v_f16m8(rows0p, vl);
-            vfloat16m8_t _rows1 = vle16_v_f16m8(rows1p, vl);
-
-            vfloat16m8_t _Dp = vfmacc_vf_f16m8(vfmul_vf_f16m8(_rows0, b0, vl), b1, _rows1, vl);
-
-            vse16_v_f16m8(Dp, _Dp, vl);
-
-            Dp += vl;
-            rows0p += vl;
-            rows1p += vl;
-            n -= vl;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear_packn.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear_packn.h
deleted file mode 100644
index 725651dd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear_packn.h
+++ /dev/null
@@ -1,120 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bilinear_image_packn(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)packn * 4u, packn);
-    Mat rowsbuf1(w, (size_t)packn * 4u, packn);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const float* S1p = S1 + sx;
-
-                vfloat32m1_t _S10 = vle32_v_f32m1(S1p, vl);
-                vfloat32m1_t _S11 = vle32_v_f32m1(S1p + packn, vl);
-                vfloat32m1_t _rows1 = vfmacc_vf_f32m1(vfmul_vf_f32m1(_S10, alphap[0], vl), alphap[1], _S11, vl);
-
-                vse32_v_f32m1(rows1p + dx * packn, _rows1, vl);
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const float* S0 = src.row(sy);
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-
-                vfloat32m1_t _S00 = vle32_v_f32m1(S0p, vl);
-                vfloat32m1_t _S01 = vle32_v_f32m1(S0p + packn, vl);
-                vfloat32m1_t _S10 = vle32_v_f32m1(S1p, vl);
-                vfloat32m1_t _S11 = vle32_v_f32m1(S1p + packn, vl);
-                vfloat32m1_t _rows0 = vfmacc_vf_f32m1(vfmul_vf_f32m1(_S00, alphap[0], vl), alphap[1], _S01, vl);
-                vfloat32m1_t _rows1 = vfmacc_vf_f32m1(vfmul_vf_f32m1(_S10, alphap[0], vl), alphap[1], _S11, vl);
-
-                vse32_v_f32m1(rows0p + dx * packn, _rows0, vl);
-                vse32_v_f32m1(rows1p + dx * packn, _rows1, vl);
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* Dp = dst.row(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            vfloat32m1_t _rows0 = vle32_v_f32m1(rows0p, vl);
-            vfloat32m1_t _rows1 = vle32_v_f32m1(rows1p, vl);
-
-            vfloat32m1_t _Dp = vfmacc_vf_f32m1(vfmul_vf_f32m1(_rows0, b0, vl), b1, _rows1, vl);
-
-            vse32_v_f32m1(Dp, _Dp, vl);
-
-            Dp += packn;
-            rows0p += packn;
-            rows1p += packn;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear_packn_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear_packn_fp16s.h
deleted file mode 100644
index bfa23943..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_bilinear_packn_fp16s.h
+++ /dev/null
@@ -1,227 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bilinear_image_packn_fp16s(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)packn * 4u, packn);
-    Mat rowsbuf1(w, (size_t)packn * 4u, packn);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const __fp16* S1p = S1 + sx;
-
-                vfloat16m1_t _S10 = vle16_v_f16m1(S1p, vl);
-                vfloat16m1_t _S11 = vle16_v_f16m1(S1p + packn, vl);
-                vfloat32m2_t _rows1 = vfwmacc_vf_f32m2(vfwmul_vf_f32m2(_S10, alphap[0], vl), alphap[1], _S11, vl);
-
-                vse32_v_f32m2(rows1p + dx * packn, _rows1, vl);
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const __fp16* S0 = src.row<const __fp16>(sy);
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-
-                vfloat16m1_t _S00 = vle16_v_f16m1(S0p, vl);
-                vfloat16m1_t _S01 = vle16_v_f16m1(S0p + packn, vl);
-                vfloat16m1_t _S10 = vle16_v_f16m1(S1p, vl);
-                vfloat16m1_t _S11 = vle16_v_f16m1(S1p + packn, vl);
-                vfloat32m2_t _rows0 = vfwmacc_vf_f32m2(vfwmul_vf_f32m2(_S00, alphap[0], vl), alphap[1], _S01, vl);
-                vfloat32m2_t _rows1 = vfwmacc_vf_f32m2(vfwmul_vf_f32m2(_S10, alphap[0], vl), alphap[1], _S11, vl);
-
-                vse32_v_f32m2(rows0p + dx * packn, _rows0, vl);
-                vse32_v_f32m2(rows1p + dx * packn, _rows1, vl);
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        __fp16* Dp = dst.row<__fp16>(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            vfloat32m2_t _rows0 = vle32_v_f32m2(rows0p, vl);
-            vfloat32m2_t _rows1 = vle32_v_f32m2(rows1p, vl);
-
-            vfloat32m2_t _Dp = vfmacc_vf_f32m2(vfmul_vf_f32m2(_rows0, b0, vl), b1, _rows1, vl);
-
-            vse16_v_f16m1(Dp, vfncvt_f_f_w_f16m1(_Dp, vl), vl);
-
-            Dp += packn;
-            rows0p += packn;
-            rows1p += packn;
-        }
-
-        beta += 2;
-    }
-}
-
-static void resize_bilinear_image_packn_fp16sa(const Mat& src, Mat& dst, __fp16* alpha, int* xofs, __fp16* beta, int* yofs)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)packn * 2u, packn);
-    Mat rowsbuf1(w, (size_t)packn * 2u, packn);
-    __fp16* rows0 = rowsbuf0;
-    __fp16* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            __fp16* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const __fp16* S1p = S1 + sx;
-
-                vfloat16m1_t _S10 = vle16_v_f16m1(S1p, vl);
-                vfloat16m1_t _S11 = vle16_v_f16m1(S1p + packn, vl);
-                vfloat16m1_t _rows1 = vfmacc_vf_f16m1(vfmul_vf_f16m1(_S10, alphap[0], vl), alphap[1], _S11, vl);
-
-                vse16_v_f16m1(rows1p + dx * packn, _rows1, vl);
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const __fp16* S0 = src.row<const __fp16>(sy);
-            const __fp16* S1 = src.row<const __fp16>(sy + 1);
-
-            const __fp16* alphap = alpha;
-            __fp16* rows0p = rows0;
-            __fp16* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * packn;
-                const __fp16* S0p = S0 + sx;
-                const __fp16* S1p = S1 + sx;
-
-                vfloat16m1_t _S00 = vle16_v_f16m1(S0p, vl);
-                vfloat16m1_t _S01 = vle16_v_f16m1(S0p + packn, vl);
-                vfloat16m1_t _S10 = vle16_v_f16m1(S1p, vl);
-                vfloat16m1_t _S11 = vle16_v_f16m1(S1p + packn, vl);
-                vfloat16m1_t _rows0 = vfmacc_vf_f16m1(vfmul_vf_f16m1(_S00, alphap[0], vl), alphap[1], _S01, vl);
-                vfloat16m1_t _rows1 = vfmacc_vf_f16m1(vfmul_vf_f16m1(_S10, alphap[0], vl), alphap[1], _S11, vl);
-
-                vse16_v_f16m1(rows0p + dx * packn, _rows0, vl);
-                vse16_v_f16m1(rows1p + dx * packn, _rows1, vl);
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        __fp16 b0 = beta[0];
-        __fp16 b1 = beta[1];
-
-        __fp16* rows0p = rows0;
-        __fp16* rows1p = rows1;
-        __fp16* Dp = dst.row<__fp16>(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            vfloat16m1_t _rows0 = vle16_v_f16m1(rows0p, vl);
-            vfloat16m1_t _rows1 = vle16_v_f16m1(rows1p, vl);
-
-            vfloat16m1_t _Dp = vfmacc_vf_f16m1(vfmul_vf_f16m1(_rows0, b0, vl), b1, _rows1, vl);
-
-            vse16_v_f16m1(Dp, _Dp, vl);
-
-            Dp += packn;
-            rows0p += packn;
-            rows1p += packn;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_riscv.h
deleted file mode 100644
index f4792235..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/interp_riscv.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INTERP_RISCV_H
-#define LAYER_INTERP_RISCV_H
-
-#include "interp.h"
-
-namespace ncnn {
-
-class Interp_riscv : public Interp
-{
-public:
-    Interp_riscv();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_fp16s(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-    int forward_fp16sa(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INTERP_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/mish_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/mish_riscv.h
deleted file mode 100644
index 2e2be1a2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/mish_riscv.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MISH_RISCV_H
-#define LAYER_MISH_RISCV_H
-
-#include "mish.h"
-
-namespace ncnn {
-
-class Mish_riscv : public Mish
-{
-public:
-    Mish_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MISH_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/packing_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/packing_riscv.h
deleted file mode 100644
index 097d7749..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/packing_riscv.h
+++ /dev/null
@@ -1,36 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PACKING_RISCV_H
-#define LAYER_PACKING_RISCV_H
-
-#include "packing.h"
-
-namespace ncnn {
-
-class Packing_riscv : public Packing
-{
-public:
-    Packing_riscv();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_bf16s_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PACKING_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/padding_packn.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/padding_packn.h
deleted file mode 100644
index 50f5efe1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/padding_packn.h
+++ /dev/null
@@ -1,231 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#define _PADDING_PACKN_RVV(SEW, TSEW, LMUL, T, VT)                                                                                          \
-    static void padding_constant_packn_##VT##_rvv(const Mat& src, Mat& dst, int top, int bottom, int left, int right, v##VT##m##LMUL##_t v) \
-    {                                                                                                                                       \
-        const int packn = csrr_vlenb() / sizeof(T);                                                                                         \
-        const size_t vl = vsetvl_e##SEW##m##LMUL(packn);                                                                                    \
-                                                                                                                                            \
-        const T* ptr = src;                                                                                                                 \
-        T* outptr = dst;                                                                                                                    \
-                                                                                                                                            \
-        /* fill top */                                                                                                                      \
-        for (int y = 0; y < top; y++)                                                                                                       \
-        {                                                                                                                                   \
-            for (int x = 0; x < dst.w; x++)                                                                                                 \
-            {                                                                                                                               \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, v, vl);                                                                                \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-        }                                                                                                                                   \
-        /* fill center */                                                                                                                   \
-        for (int y = 0; y < src.h; y++)                                                                                                     \
-        {                                                                                                                                   \
-            for (int x = 0; x < left; x++)                                                                                                  \
-            {                                                                                                                               \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, v, vl);                                                                                \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            for (int x = 0; x < src.w; x++)                                                                                                 \
-            {                                                                                                                               \
-                v##VT##m##LMUL##_t _p = vle##SEW##_v_##TSEW##m##LMUL(ptr, vl);                                                              \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                ptr += packn;                                                                                                               \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            for (int x = 0; x < right; x++)                                                                                                 \
-            {                                                                                                                               \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, v, vl);                                                                                \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-        }                                                                                                                                   \
-        /* fill bottom */                                                                                                                   \
-        for (int y = 0; y < bottom; y++)                                                                                                    \
-        {                                                                                                                                   \
-            for (int x = 0; x < dst.w; x++)                                                                                                 \
-            {                                                                                                                               \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, v, vl);                                                                                \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-        }                                                                                                                                   \
-    }                                                                                                                                       \
-                                                                                                                                            \
-    static void padding_replicate_packn_##VT##_rvv(const Mat& src, Mat& dst, int top, int bottom, int left, int right)                      \
-    {                                                                                                                                       \
-        const int packn = csrr_vlenb() / sizeof(T);                                                                                         \
-        const size_t vl = vsetvl_e##SEW##m##LMUL(packn);                                                                                    \
-                                                                                                                                            \
-        const T* ptr = src;                                                                                                                 \
-        T* outptr = dst;                                                                                                                    \
-                                                                                                                                            \
-        /* fill top */                                                                                                                      \
-        for (int y = 0; y < top; y++)                                                                                                       \
-        {                                                                                                                                   \
-            const T* ptr0 = ptr;                                                                                                            \
-            v##VT##m##LMUL##_t _p = vle##SEW##_v_##TSEW##m##LMUL(ptr0, vl);                                                                 \
-            for (int x = 0; x < left; x++)                                                                                                  \
-            {                                                                                                                               \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            for (int x = 0; x < src.w; x++)                                                                                                 \
-            {                                                                                                                               \
-                _p = vle##SEW##_v_##TSEW##m##LMUL(ptr0, vl);                                                                                \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                ptr0 += packn;                                                                                                              \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            for (int x = 0; x < right; x++)                                                                                                 \
-            {                                                                                                                               \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-        }                                                                                                                                   \
-        /* fill center */                                                                                                                   \
-        for (int y = 0; y < src.h; y++)                                                                                                     \
-        {                                                                                                                                   \
-            v##VT##m##LMUL##_t _p = vle##SEW##_v_##TSEW##m##LMUL(ptr, vl);                                                                  \
-            for (int x = 0; x < left; x++)                                                                                                  \
-            {                                                                                                                               \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            for (int x = 0; x < src.w; x++)                                                                                                 \
-            {                                                                                                                               \
-                _p = vle##SEW##_v_##TSEW##m##LMUL(ptr, vl);                                                                                 \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                ptr += packn;                                                                                                               \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            for (int x = 0; x < right; x++)                                                                                                 \
-            {                                                                                                                               \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-        }                                                                                                                                   \
-        /* fill bottom */                                                                                                                   \
-        ptr -= src.w * packn;                                                                                                               \
-        for (int y = 0; y < bottom; y++)                                                                                                    \
-        {                                                                                                                                   \
-            const T* ptr0 = ptr;                                                                                                            \
-            v##VT##m##LMUL##_t _p = vle##SEW##_v_##TSEW##m##LMUL(ptr0, vl);                                                                 \
-            for (int x = 0; x < left; x++)                                                                                                  \
-            {                                                                                                                               \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            for (int x = 0; x < src.w; x++)                                                                                                 \
-            {                                                                                                                               \
-                _p = vle##SEW##_v_##TSEW##m##LMUL(ptr0, vl);                                                                                \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                ptr0 += packn;                                                                                                              \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            for (int x = 0; x < right; x++)                                                                                                 \
-            {                                                                                                                               \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-        }                                                                                                                                   \
-    }                                                                                                                                       \
-                                                                                                                                            \
-    static void padding_reflect_packn_##VT##_rvv(const Mat& src, Mat& dst, int top, int bottom, int left, int right)                        \
-    {                                                                                                                                       \
-        const int packn = csrr_vlenb() / sizeof(T);                                                                                         \
-        const size_t vl = vsetvl_e##SEW##m##LMUL(packn);                                                                                    \
-                                                                                                                                            \
-        const T* ptr = src;                                                                                                                 \
-        T* outptr = dst;                                                                                                                    \
-                                                                                                                                            \
-        /* fill top */                                                                                                                      \
-        ptr += top * src.w * packn;                                                                                                         \
-        for (int y = 0; y < top; y++)                                                                                                       \
-        {                                                                                                                                   \
-            const T* ptr0 = ptr;                                                                                                            \
-            for (int x = 0; x < left; x++)                                                                                                  \
-            {                                                                                                                               \
-                v##VT##m##LMUL##_t _p = vle##SEW##_v_##TSEW##m##LMUL(ptr0 + (left - x) * packn, vl);                                        \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            for (int x = 0; x < src.w; x++)                                                                                                 \
-            {                                                                                                                               \
-                v##VT##m##LMUL##_t _p = vle##SEW##_v_##TSEW##m##LMUL(ptr0, vl);                                                             \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                ptr0 += packn;                                                                                                              \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            for (int x = 0; x < right; x++)                                                                                                 \
-            {                                                                                                                               \
-                v##VT##m##LMUL##_t _p = vle##SEW##_v_##TSEW##m##LMUL(ptr0 - packn * 2 - x * packn, vl);                                     \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            ptr -= src.w * packn;                                                                                                           \
-        }                                                                                                                                   \
-        /* fill center */                                                                                                                   \
-        for (int y = 0; y < src.h; y++)                                                                                                     \
-        {                                                                                                                                   \
-            for (int x = 0; x < left; x++)                                                                                                  \
-            {                                                                                                                               \
-                v##VT##m##LMUL##_t _p = vle##SEW##_v_##TSEW##m##LMUL(ptr + (left - x) * packn, vl);                                         \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            for (int x = 0; x < src.w; x++)                                                                                                 \
-            {                                                                                                                               \
-                v##VT##m##LMUL##_t _p = vle##SEW##_v_##TSEW##m##LMUL(ptr, vl);                                                              \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                ptr += packn;                                                                                                               \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            for (int x = 0; x < right; x++)                                                                                                 \
-            {                                                                                                                               \
-                v##VT##m##LMUL##_t _p = vle##SEW##_v_##TSEW##m##LMUL(ptr - packn * 2 - x * packn, vl);                                      \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-        }                                                                                                                                   \
-        /* fill bottom */                                                                                                                   \
-        ptr -= 2 * src.w * packn;                                                                                                           \
-        for (int y = 0; y < bottom; y++)                                                                                                    \
-        {                                                                                                                                   \
-            const T* ptr0 = ptr;                                                                                                            \
-            for (int x = 0; x < left; x++)                                                                                                  \
-            {                                                                                                                               \
-                v##VT##m##LMUL##_t _p = vle##SEW##_v_##TSEW##m##LMUL(ptr0 + (left - x) * packn, vl);                                        \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            for (int x = 0; x < src.w; x++)                                                                                                 \
-            {                                                                                                                               \
-                v##VT##m##LMUL##_t _p = vle##SEW##_v_##TSEW##m##LMUL(ptr0, vl);                                                             \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                ptr0 += packn;                                                                                                              \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            for (int x = 0; x < right; x++)                                                                                                 \
-            {                                                                                                                               \
-                v##VT##m##LMUL##_t _p = vle##SEW##_v_##TSEW##m##LMUL(ptr0 - packn * 2 - x * packn, vl);                                     \
-                vse##SEW##_v_##TSEW##m##LMUL(outptr, _p, vl);                                                                               \
-                outptr += packn;                                                                                                            \
-            }                                                                                                                               \
-            ptr -= src.w * packn;                                                                                                           \
-        }                                                                                                                                   \
-    }
-
-_PADDING_PACKN_RVV(32, f32, 1, float, float32)
-_PADDING_PACKN_RVV(16, u16, 1, unsigned short, uint16)
-_PADDING_PACKN_RVV(8, i8, 1, signed char, int8)
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/padding_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/padding_riscv.h
deleted file mode 100644
index 7642dcca..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/padding_riscv.h
+++ /dev/null
@@ -1,49 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PADDING_RISCV_H
-#define LAYER_PADDING_RISCV_H
-
-#include "padding.h"
-
-namespace ncnn {
-
-class Padding_riscv : public Padding
-{
-public:
-    Padding_riscv();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_bf16s_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-#if NCNN_BF16
-    // bf16
-    unsigned short value_bf16;
-    Mat per_channel_pad_data_bf16;
-#endif
-
-    // fp16
-    Mat per_channel_pad_data_fp16;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PADDING_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/pooling_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/pooling_riscv.h
deleted file mode 100644
index e285b58e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/pooling_riscv.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_POOLING_RISCV_H
-#define LAYER_POOLING_RISCV_H
-
-#include "pooling.h"
-
-namespace ncnn {
-
-class Pooling_riscv : public Pooling
-{
-public:
-    Pooling_riscv();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-    int forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_POOLING_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/prelu_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/prelu_riscv.h
deleted file mode 100644
index 70acbc5d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/prelu_riscv.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Xavier Hsinyuan is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 Xavier Hsinyuan <me@lstlx.com>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PRELU_RISCV_H
-#define LAYER_PRELU_RISCV_H
-
-#include "prelu.h"
-
-namespace ncnn {
-
-class PReLU_riscv : public PReLU
-{
-public:
-    PReLU_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-#if __riscv_vector && __riscv_zfh
-protected:
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PRELU_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/relu_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/relu_riscv.h
deleted file mode 100644
index 58181b53..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/relu_riscv.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_RELU_RISCV_H
-#define LAYER_RELU_RISCV_H
-
-#include "relu.h"
-
-namespace ncnn {
-
-class ReLU_riscv : public ReLU
-{
-public:
-    ReLU_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_RELU_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/riscv_activation.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/riscv_activation.h
deleted file mode 100644
index d5f114f3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/riscv_activation.h
+++ /dev/null
@@ -1,81 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef RISCV_ACTIVATION_H
-#define RISCV_ACTIVATION_H
-
-#include "fused_activation.h"
-
-#if __riscv_vector
-#include <riscv_vector.h>
-#include "rvv_mathfun.h"
-#include "rvv_mathfun_fp16s.h"
-
-#define _RVV_FLOAT_ACTIVATION_PS(SEW, LMUL, MLEN)                                                                                                         \
-    static inline vfloat##SEW##m##LMUL##_t activation_ps(vfloat##SEW##m##LMUL##_t _v, int activation_type, const ncnn::Mat& activation_params, size_t vl) \
-    {                                                                                                                                                     \
-        if (activation_type == 1)                                                                                                                         \
-        {                                                                                                                                                 \
-            _v = vfmax_vf_f##SEW##m##LMUL(_v, 0.f, vl);                                                                                                   \
-        }                                                                                                                                                 \
-        else if (activation_type == 2)                                                                                                                    \
-        {                                                                                                                                                 \
-            vbool##MLEN##_t _lemask = vmfle_vf_f##SEW##m##LMUL##_b##MLEN(_v, 0.f, vl);                                                                    \
-            _v = vfmul_vf_f##SEW##m##LMUL##_m(_lemask, _v, _v, activation_params[0], vl);                                                                 \
-        }                                                                                                                                                 \
-        else if (activation_type == 3)                                                                                                                    \
-        {                                                                                                                                                 \
-            _v = vfmax_vf_f##SEW##m##LMUL(_v, activation_params[0], vl);                                                                                  \
-            _v = vfmin_vf_f##SEW##m##LMUL(_v, activation_params[1], vl);                                                                                  \
-        }                                                                                                                                                 \
-        else if (activation_type == 4)                                                                                                                    \
-        {                                                                                                                                                 \
-            _v = sigmoid_ps(_v, vl);                                                                                                                      \
-        }                                                                                                                                                 \
-        else if (activation_type == 5)                                                                                                                    \
-        {                                                                                                                                                 \
-            _v = vfmul_vv_f##SEW##m##LMUL(_v, tanh_ps(log_ps(vfadd_vf_f##SEW##m##LMUL(exp_ps(_v, vl), 1.f, vl), vl), vl), vl);                            \
-        }                                                                                                                                                 \
-        else if (activation_type == 6)                                                                                                                    \
-        {                                                                                                                                                 \
-            const float alpha = activation_params[0];                                                                                                     \
-            const float beta = activation_params[1];                                                                                                      \
-            const float lower = -beta / alpha;                                                                                                            \
-            const float upper = (1.f / alpha) + lower;                                                                                                    \
-            vbool##MLEN##_t _lower = vmflt_vf_f##SEW##m##LMUL##_b##MLEN(_v, lower, vl);                                                                   \
-            vbool##MLEN##_t _higher = vmfgt_vf_f##SEW##m##LMUL##_b##MLEN(_v, upper, vl);                                                                  \
-            vbool##MLEN##_t _apply = vmnor_mm_b##MLEN(_lower, _higher, vl);                                                                               \
-            _v = vfmerge_vfm_f##SEW##m##LMUL(_lower, _v, .0f, vl);                                                                                        \
-                                                                                                                                                          \
-            vfloat##SEW##m##LMUL##_t _p0 = vfadd_vf_f##SEW##m##LMUL##_m(                                                                                  \
-                _apply, _v, /*op1*/ vfmul_vf_f##SEW##m##LMUL##_m(_apply, _v, _v, alpha, vl), beta,                                                        \
-                vl);                                                                                                                                      \
-            _v = vfmul_vv_f##SEW##m##LMUL##_m(_apply, _v, /*op1*/ _v, _p0, vl);                                                                           \
-        }                                                                                                                                                 \
-                                                                                                                                                          \
-        return _v;                                                                                                                                        \
-    }
-
-_RVV_FLOAT_ACTIVATION_PS(16, 1, 16)
-_RVV_FLOAT_ACTIVATION_PS(16, 2, 8)
-_RVV_FLOAT_ACTIVATION_PS(16, 4, 4)
-_RVV_FLOAT_ACTIVATION_PS(16, 8, 2)
-_RVV_FLOAT_ACTIVATION_PS(32, 1, 32)
-_RVV_FLOAT_ACTIVATION_PS(32, 2, 16)
-_RVV_FLOAT_ACTIVATION_PS(32, 4, 8)
-_RVV_FLOAT_ACTIVATION_PS(32, 8, 4)
-
-#endif // __riscv_vector
-
-#endif // RISCV_ACTIVATION_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/riscv_usability.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/riscv_usability.h
deleted file mode 100644
index e2824646..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/riscv_usability.h
+++ /dev/null
@@ -1,619 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef RISCV_USABILITY_H
-#define RISCV_USABILITY_H
-
-#if __riscv_vector
-#include <riscv_vector.h>
-#endif // __riscv_vector
-
-#if __riscv_vector
-static inline int csrr_vl()
-{
-    int a = 0;
-    asm volatile("csrr %0, vl"
-                 : "=r"(a)
-                 :
-                 : "memory");
-    return a;
-}
-
-static inline int csrr_vtype()
-{
-    int a = 0;
-    asm volatile("csrr %0, vtype"
-                 : "=r"(a)
-                 :
-                 : "memory");
-    return a;
-}
-
-static inline int csrr_vlenb()
-{
-    int a = 0;
-    asm volatile("csrr %0, vlenb"
-                 : "=r"(a)
-                 :
-                 : "memory");
-    return a;
-}
-
-static inline vfloat32m8_t vle32_v_f32m8_f32m1(const float* ptr)
-{
-    const int packn = csrr_vlenb() / 4;
-    const size_t vl = vsetvl_e32m8(packn * 8);
-
-    // NOTE vloxei8_v_f32m8 gets illegal instruction on d1  --- nihui
-
-    // 128bit
-    static const uint32_t index_128bit[32] = {
-        0, 4, 8, 12,
-        0, 4, 8, 12,
-        0, 4, 8, 12,
-        0, 4, 8, 12,
-        0, 4, 8, 12,
-        0, 4, 8, 12,
-        0, 4, 8, 12,
-        0, 4, 8, 12
-    };
-
-    // 256bit
-    static const uint32_t index_256bit[64] = {
-        0, 4, 8, 12, 16, 20, 24, 28,
-        0, 4, 8, 12, 16, 20, 24, 28,
-        0, 4, 8, 12, 16, 20, 24, 28,
-        0, 4, 8, 12, 16, 20, 24, 28,
-        0, 4, 8, 12, 16, 20, 24, 28,
-        0, 4, 8, 12, 16, 20, 24, 28,
-        0, 4, 8, 12, 16, 20, 24, 28,
-        0, 4, 8, 12, 16, 20, 24, 28
-    };
-
-    const uint32_t* index = packn == 4 ? index_128bit : index_256bit;
-    vuint32m8_t bindex = vle32_v_u32m8(index, vl);
-    return vloxei32_v_f32m8(ptr, bindex, vl);
-}
-
-#if __riscv_zfh
-static inline vfloat16m8_t vle16_v_f16m8_f16m1(const __fp16* ptr)
-{
-    const int packn = csrr_vlenb() / 2;
-    const size_t vl = vsetvl_e16m8(packn * 8);
-
-    // NOTE vloxei8_v_f16m8 gets illegal instruction on d1  --- nihui
-
-    // 128bit
-    static const uint16_t index_128bit[64] = {
-        0, 2, 4, 6, 8, 10, 12, 14,
-        0, 2, 4, 6, 8, 10, 12, 14,
-        0, 2, 4, 6, 8, 10, 12, 14,
-        0, 2, 4, 6, 8, 10, 12, 14,
-        0, 2, 4, 6, 8, 10, 12, 14,
-        0, 2, 4, 6, 8, 10, 12, 14,
-        0, 2, 4, 6, 8, 10, 12, 14,
-        0, 2, 4, 6, 8, 10, 12, 14
-    };
-
-    // 256bit
-    static const uint16_t index_256bit[128] = {
-        0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30,
-        0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30,
-        0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30,
-        0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30,
-        0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30,
-        0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30,
-        0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30,
-        0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
-    };
-
-    const uint16_t* index = packn == 8 ? index_128bit : index_256bit;
-    vuint16m8_t bindex = vle16_v_u16m8(index, vl);
-    return vloxei16_v_f16m8(ptr, bindex, vl);
-}
-#endif // __riscv_zfh
-#endif // __riscv_vector
-
-#if __riscv_vector && __rvv_tuple
-
-// f32m1, vsseg.v
-static inline void vsseg8e32_v_f32m1(float32_t* base, vfloat32m1_t v0, vfloat32m1_t v1, vfloat32m1_t v2, vfloat32m1_t v3, vfloat32m1_t v4, vfloat32m1_t v5, vfloat32m1_t v6, vfloat32m1_t v7, size_t vl)
-{
-    vfloat32m1x8_t _tmp = vcreate_f32m1x8(v0, v1, v2, v3, v4, v5, v6, v7);
-    vsseg8e32_v_f32m1x8(base, _tmp, vl);
-}
-
-static inline void vsseg4e32_v_f32m1(float32_t* base, vfloat32m1_t v0, vfloat32m1_t v1, vfloat32m1_t v2, vfloat32m1_t v3, size_t vl)
-{
-    vfloat32m1x4_t _tmp = vcreate_f32m1x4(v0, v1, v2, v3);
-    vsseg4e32_v_f32m1x4(base, _tmp, vl);
-}
-
-static inline void vsseg2e32_v_f32m1(float32_t* base, vfloat32m1_t v0, vfloat32m1_t v1, size_t vl)
-{
-    vfloat32m1x2_t _tmp = vcreate_f32m1x2(v0, v1);
-    vsseg2e32_v_f32m1x2(base, _tmp, vl);
-}
-
-// f32m1, vssseg.v, 8/4/2
-static inline void vssseg8e32_v_f32m1(float32_t* base, ptrdiff_t bstride, vfloat32m1_t v0, vfloat32m1_t v1, vfloat32m1_t v2, vfloat32m1_t v3, vfloat32m1_t v4, vfloat32m1_t v5, vfloat32m1_t v6, vfloat32m1_t v7, size_t vl)
-{
-    vfloat32m1x8_t _tmp = vcreate_f32m1x8(v0, v1, v2, v3, v4, v5, v6, v7);
-    vssseg8e32_v_f32m1x8(base, bstride, _tmp, vl);
-}
-
-static inline void vssseg4e32_v_f32m1(float32_t* base, ptrdiff_t bstride, vfloat32m1_t v0, vfloat32m1_t v1, vfloat32m1_t v2, vfloat32m1_t v3, size_t vl)
-{
-    vfloat32m1x4_t _tmp = vcreate_f32m1x4(v0, v1, v2, v3);
-    vssseg4e32_v_f32m1x4(base, bstride, _tmp, vl);
-}
-
-static inline void vssseg2e32_v_f32m1(float32_t* base, ptrdiff_t bstride, vfloat32m1_t v0, vfloat32m1_t v1, size_t vl)
-{
-    vfloat32m1x2_t _tmp = vcreate_f32m1x2(v0, v1);
-    vssseg2e32_v_f32m1x2(base, bstride, _tmp, vl);
-}
-
-// f32m2, vsseg.v, 4/2
-static inline void vsseg4e32_v_f32m2(float32_t* base, vfloat32m2_t v0, vfloat32m2_t v1, vfloat32m2_t v2, vfloat32m2_t v3, size_t vl)
-{
-    vfloat32m2x4_t _tmp = vcreate_f32m2x4(v0, v1, v2, v3);
-    vsseg4e32_v_f32m2x4(base, _tmp, vl);
-}
-
-static inline void vsseg2e32_v_f32m2(float32_t* base, vfloat32m2_t v0, vfloat32m2_t v1, size_t vl)
-{
-    vfloat32m2x2_t _tmp = vcreate_f32m2x2(v0, v1);
-    vsseg2e32_v_f32m2x2(base, _tmp, vl);
-}
-
-// u16m1, vsseg.v, 8/4
-static inline void vsseg8e16_v_u16m1(uint16_t* base, vuint16m1_t v0, vuint16m1_t v1, vuint16m1_t v2, vuint16m1_t v3, vuint16m1_t v4, vuint16m1_t v5, vuint16m1_t v6, vuint16m1_t v7, size_t vl)
-{
-    vuint16m1x8_t _tmp = vcreate_u16m1x8(v0, v1, v2, v3, v4, v5, v6, v7);
-    vsseg8e16_v_u16m1x8(base, _tmp, vl);
-}
-
-static inline void vsseg4e16_v_u16m1(uint16_t* base, vuint16m1_t v0, vuint16m1_t v1, vuint16m1_t v2, vuint16m1_t v3, size_t vl)
-{
-    vuint16m1x4_t _tmp = vcreate_u16m1x4(v0, v1, v2, v3);
-    vsseg4e16_v_u16m1x4(base, _tmp, vl);
-}
-
-// u16m2, vsseg.v, 4/2
-static inline void vsseg4e16_v_u16m2(uint16_t* base, vuint16m2_t v0, vuint16m2_t v1, vuint16m2_t v2, vuint16m2_t v3, size_t vl)
-{
-    vuint16m2x4_t _tmp = vcreate_u16m2x4(v0, v1, v2, v3);
-    vsseg4e16_v_u16m2x4(base, _tmp, vl);
-}
-
-static inline void vsseg2e16_v_u16m2(uint16_t* base, vuint16m2_t v0, vuint16m2_t v1, size_t vl)
-{
-    vuint16m2x2_t _tmp = vcreate_u16m2x2(v0, v1);
-    vsseg2e16_v_u16m2x2(base, _tmp, vl);
-}
-
-// f32m1, vlseg.v 8/4/2
-static inline void vlseg8e32_v_f32m1(vfloat32m1_t* v0, vfloat32m1_t* v1, vfloat32m1_t* v2, vfloat32m1_t* v3, vfloat32m1_t* v4, vfloat32m1_t* v5, vfloat32m1_t* v6, vfloat32m1_t* v7, const float32_t* base, size_t vl)
-{
-    vfloat32m1x8_t _tmp = vlseg8e32_v_f32m1x8(base, vl);
-    *v0 = vget_f32m1x8_f32m1(_tmp, 0);
-    *v1 = vget_f32m1x8_f32m1(_tmp, 1);
-    *v2 = vget_f32m1x8_f32m1(_tmp, 2);
-    *v3 = vget_f32m1x8_f32m1(_tmp, 3);
-    *v4 = vget_f32m1x8_f32m1(_tmp, 4);
-    *v5 = vget_f32m1x8_f32m1(_tmp, 5);
-    *v6 = vget_f32m1x8_f32m1(_tmp, 6);
-    *v7 = vget_f32m1x8_f32m1(_tmp, 7);
-}
-
-static inline void vlseg4e32_v_f32m1(vfloat32m1_t* v0, vfloat32m1_t* v1, vfloat32m1_t* v2, vfloat32m1_t* v3, const float32_t* base, size_t vl)
-{
-    vfloat32m1x4_t _tmp = vlseg4e32_v_f32m1x4(base, vl);
-    *v0 = vget_f32m1x4_f32m1(_tmp, 0);
-    *v1 = vget_f32m1x4_f32m1(_tmp, 1);
-    *v2 = vget_f32m1x4_f32m1(_tmp, 2);
-    *v3 = vget_f32m1x4_f32m1(_tmp, 3);
-}
-
-static inline void vlseg2e32_v_f32m1(vfloat32m1_t* v0, vfloat32m1_t* v1, const float32_t* base, size_t vl)
-{
-    vfloat32m1x2_t _tmp = vlseg2e32_v_f32m1x2(base, vl);
-    *v0 = vget_f32m1x2_f32m1(_tmp, 0);
-    *v1 = vget_f32m1x2_f32m1(_tmp, 1);
-}
-
-// f32m2, vlseg.v, 4
-static inline void vlseg4e32_v_f32m2(vfloat32m2_t* v0, vfloat32m2_t* v1, vfloat32m2_t* v2, vfloat32m2_t* v3, const float32_t* base, size_t vl)
-{
-    vfloat32m2x4_t _tmp = vlseg4e32_v_f32m2x4(base, vl);
-    *v0 = vget_f32m2x4_f32m2(_tmp, 0);
-    *v1 = vget_f32m2x4_f32m2(_tmp, 1);
-    *v2 = vget_f32m2x4_f32m2(_tmp, 2);
-    *v3 = vget_f32m2x4_f32m2(_tmp, 3);
-}
-
-// f32m4, vlseg.v, 2
-static inline void vlseg2e32_v_f32m4(vfloat32m4_t* v0, vfloat32m4_t* v1, const float32_t* base, size_t vl)
-{
-    vfloat32m4x2_t _tmp = vlseg2e32_v_f32m4x2(base, vl);
-    *v0 = vget_f32m4x2_f32m4(_tmp, 0);
-    *v1 = vget_f32m4x2_f32m4(_tmp, 1);
-}
-
-// f32m4, vloxseg.v
-static inline void vloxseg2ei32_v_f32m4(vfloat32m4_t* v0, vfloat32m4_t* v1, const float32_t* base, vuint32m4_t bindex, size_t vl)
-{
-    vfloat32m4x2_t _tmp = vloxseg2ei32_v_f32m4x2(base, bindex, vl);
-    *v0 = vget_f32m4x2_f32m4(_tmp, 0);
-    *v1 = vget_f32m4x2_f32m4(_tmp, 1);
-}
-
-// u16m1, vlseg.v 8/4/2
-static inline void vlseg8e16_v_u16m1(vuint16m1_t* v0, vuint16m1_t* v1, vuint16m1_t* v2, vuint16m1_t* v3, vuint16m1_t* v4, vuint16m1_t* v5, vuint16m1_t* v6, vuint16m1_t* v7, const uint16_t* base, size_t vl)
-{
-    vuint16m1x8_t _tmp = vlseg8e16_v_u16m1x8(base, vl);
-    *v0 = vget_u16m1x8_u16m1(_tmp, 0);
-    *v1 = vget_u16m1x8_u16m1(_tmp, 1);
-    *v2 = vget_u16m1x8_u16m1(_tmp, 2);
-    *v3 = vget_u16m1x8_u16m1(_tmp, 3);
-    *v4 = vget_u16m1x8_u16m1(_tmp, 4);
-    *v5 = vget_u16m1x8_u16m1(_tmp, 5);
-    *v6 = vget_u16m1x8_u16m1(_tmp, 6);
-    *v7 = vget_u16m1x8_u16m1(_tmp, 7);
-}
-
-static inline void vlseg4e16_v_u16m1(vuint16m1_t* v0, vuint16m1_t* v1, vuint16m1_t* v2, vuint16m1_t* v3, const uint16_t* base, size_t vl)
-{
-    vuint16m1x4_t _tmp = vlseg4e16_v_u16m1x4(base, vl);
-    *v0 = vget_u16m1x4_u16m1(_tmp, 0);
-    *v1 = vget_u16m1x4_u16m1(_tmp, 1);
-    *v2 = vget_u16m1x4_u16m1(_tmp, 2);
-    *v3 = vget_u16m1x4_u16m1(_tmp, 3);
-}
-
-static inline void vlseg2e16_v_u16m1(vuint16m1_t* v0, vuint16m1_t* v1, const uint16_t* base, size_t vl)
-{
-    vuint16m1x2_t _tmp = vlseg2e16_v_u16m1x2(base, vl);
-    *v0 = vget_u16m1x2_u16m1(_tmp, 0);
-    *v1 = vget_u16m1x2_u16m1(_tmp, 1);
-}
-
-// u16m2, vlseg.v, 4
-static inline void vlseg4e16_v_u16m2(vuint16m2_t* v0, vuint16m2_t* v1, vuint16m2_t* v2, vuint16m2_t* v3, const uint16_t* base, size_t vl)
-{
-    vuint16m2x4_t _tmp = vlseg4e16_v_u16m2x4(base, vl);
-    *v0 = vget_u16m2x4_u16m2(_tmp, 0);
-    *v1 = vget_u16m2x4_u16m2(_tmp, 1);
-    *v2 = vget_u16m2x4_u16m2(_tmp, 2);
-    *v3 = vget_u16m2x4_u16m2(_tmp, 3);
-}
-
-// u16m4, vlseg.v, 2
-static inline void vlseg2e16_v_u16m4(vuint16m4_t* v0, vuint16m4_t* v1, const uint16_t* base, size_t vl)
-{
-    vuint16m4x2_t _tmp = vlseg2e16_v_u16m4x2(base, vl);
-    *v0 = vget_u16m4x2_u16m4(_tmp, 0);
-    *v1 = vget_u16m4x2_u16m4(_tmp, 1);
-}
-
-#if __riscv_zfh
-
-// f16m1, vsseg.v, 8/4/2
-static inline void vsseg8e16_v_f16m1(float16_t* base, vfloat16m1_t v0, vfloat16m1_t v1, vfloat16m1_t v2, vfloat16m1_t v3, vfloat16m1_t v4, vfloat16m1_t v5, vfloat16m1_t v6, vfloat16m1_t v7, size_t vl)
-{
-    vfloat16m1x8_t _tmp = vcreate_f16m1x8(v0, v1, v2, v3, v4, v5, v6, v7);
-    vsseg8e16_v_f16m1x8(base, _tmp, vl);
-}
-
-static inline void vsseg4e16_v_f16m1(float16_t* base, vfloat16m1_t v0, vfloat16m1_t v1, vfloat16m1_t v2, vfloat16m1_t v3, size_t vl)
-{
-    vfloat16m1x4_t _tmp = vcreate_f16m1x4(v0, v1, v2, v3);
-    vsseg4e16_v_f16m1x4(base, _tmp, vl);
-}
-
-static inline void vsseg2e16_v_f16m1(float16_t* base, vfloat16m1_t v0, vfloat16m1_t v1, size_t vl)
-{
-    vfloat16m1x2_t _tmp = vcreate_f16m1x2(v0, v1);
-    vsseg2e16_v_f16m1x2(base, _tmp, vl);
-}
-
-// f16m1, vssseg.v, 8/4/2
-static inline void vssseg8e16_v_f16m1(float16_t* base, ptrdiff_t bstride, vfloat16m1_t v0, vfloat16m1_t v1, vfloat16m1_t v2, vfloat16m1_t v3, vfloat16m1_t v4, vfloat16m1_t v5, vfloat16m1_t v6, vfloat16m1_t v7, size_t vl)
-{
-    vfloat16m1x8_t _tmp = vcreate_f16m1x8(v0, v1, v2, v3, v4, v5, v6, v7);
-    vssseg8e16_v_f16m1x8(base, bstride, _tmp, vl);
-}
-
-static inline void vssseg4e16_v_f16m1(float16_t* base, ptrdiff_t bstride, vfloat16m1_t v0, vfloat16m1_t v1, vfloat16m1_t v2, vfloat16m1_t v3, size_t vl)
-{
-    vfloat16m1x4_t _tmp = vcreate_f16m1x4(v0, v1, v2, v3);
-    vssseg4e16_v_f16m1x4(base, bstride, _tmp, vl);
-}
-
-static inline void vssseg2e16_v_f16m1(float16_t* base, ptrdiff_t bstride, vfloat16m1_t v0, vfloat16m1_t v1, size_t vl)
-{
-    vfloat16m1x2_t _tmp = vcreate_f16m1x2(v0, v1);
-    vssseg2e16_v_f16m1x2(base, bstride, _tmp, vl);
-}
-
-// f16m1, vlseg.v 8/4/2
-static inline void vlseg8e16_v_f16m1(vfloat16m1_t* v0, vfloat16m1_t* v1, vfloat16m1_t* v2, vfloat16m1_t* v3, vfloat16m1_t* v4, vfloat16m1_t* v5, vfloat16m1_t* v6, vfloat16m1_t* v7, const float16_t* base, size_t vl)
-{
-    vfloat16m1x8_t _tmp = vlseg8e16_v_f16m1x8(base, vl);
-    *v0 = vget_f16m1x8_f16m1(_tmp, 0);
-    *v1 = vget_f16m1x8_f16m1(_tmp, 1);
-    *v2 = vget_f16m1x8_f16m1(_tmp, 2);
-    *v3 = vget_f16m1x8_f16m1(_tmp, 3);
-    *v4 = vget_f16m1x8_f16m1(_tmp, 4);
-    *v5 = vget_f16m1x8_f16m1(_tmp, 5);
-    *v6 = vget_f16m1x8_f16m1(_tmp, 6);
-    *v7 = vget_f16m1x8_f16m1(_tmp, 7);
-}
-
-static inline void vlseg4e16_v_f16m1(vfloat16m1_t* v0, vfloat16m1_t* v1, vfloat16m1_t* v2, vfloat16m1_t* v3, const float16_t* base, size_t vl)
-{
-    vfloat16m1x4_t _tmp = vlseg4e16_v_f16m1x4(base, vl);
-    *v0 = vget_f16m1x4_f16m1(_tmp, 0);
-    *v1 = vget_f16m1x4_f16m1(_tmp, 1);
-    *v2 = vget_f16m1x4_f16m1(_tmp, 2);
-    *v3 = vget_f16m1x4_f16m1(_tmp, 3);
-}
-
-static inline void vlseg2e16_v_f16m1(vfloat16m1_t* v0, vfloat16m1_t* v1, const float16_t* base, size_t vl)
-{
-    vfloat16m1x2_t _tmp = vlseg2e16_v_f16m1x2(base, vl);
-    *v0 = vget_f16m1x2_f16m1(_tmp, 0);
-    *v1 = vget_f16m1x2_f16m1(_tmp, 1);
-}
-
-// f16m2, vlseg.v, 4
-static inline void vlseg4e16_v_f16m2(vfloat16m2_t* v0, vfloat16m2_t* v1, vfloat16m2_t* v2, vfloat16m2_t* v3, const float16_t* base, size_t vl)
-{
-    vfloat16m2x4_t _tmp = vlseg4e16_v_f16m2x4(base, vl);
-    *v0 = vget_f16m2x4_f16m2(_tmp, 0);
-    *v1 = vget_f16m2x4_f16m2(_tmp, 1);
-    *v2 = vget_f16m2x4_f16m2(_tmp, 2);
-    *v3 = vget_f16m2x4_f16m2(_tmp, 3);
-}
-
-// f16m4, vlseg.v, 2
-static inline void vlseg2e16_v_f16m4(vfloat16m4_t* v0, vfloat16m4_t* v1, const float16_t* base, size_t vl)
-{
-    vfloat16m4x2_t _tmp = vlseg2e16_v_f16m4x2(base, vl);
-    *v0 = vget_f16m4x2_f16m4(_tmp, 0);
-    *v1 = vget_f16m4x2_f16m4(_tmp, 1);
-}
-
-#endif // __riscv_zfh
-#endif // __riscv_vector
-
-#ifdef __riscv_vector
-
-static inline void transpose8x8_ps(vfloat32m1_t& _r0l, vfloat32m1_t& _r0h,
-                                   vfloat32m1_t& _r1l, vfloat32m1_t& _r1h,
-                                   vfloat32m1_t& _r2l, vfloat32m1_t& _r2h,
-                                   vfloat32m1_t& _r3l, vfloat32m1_t& _r3h,
-                                   vfloat32m1_t& _r4l, vfloat32m1_t& _r4h,
-                                   vfloat32m1_t& _r5l, vfloat32m1_t& _r5h,
-                                   vfloat32m1_t& _r6l, vfloat32m1_t& _r6h,
-                                   vfloat32m1_t& _r7l, vfloat32m1_t& _r7h, size_t vl)
-{
-    float tmp[64];
-    vsseg8e32_v_f32m1(&tmp[0], _r0l, _r1l, _r2l, _r3l, _r4l, _r5l, _r6l, _r7l, vl);
-    vsseg8e32_v_f32m1(&tmp[32], _r0h, _r1h, _r2h, _r3h, _r4h, _r5h, _r6h, _r7h, vl);
-    float* ptr = (float*)tmp;
-    _r0l = vle32_v_f32m1(ptr + 0 * 4, vl);
-    _r0h = vle32_v_f32m1(ptr + 1 * 4, vl);
-    _r1l = vle32_v_f32m1(ptr + 2 * 4, vl);
-    _r1h = vle32_v_f32m1(ptr + 3 * 4, vl);
-    _r2l = vle32_v_f32m1(ptr + 4 * 4, vl);
-    _r2h = vle32_v_f32m1(ptr + 5 * 4, vl);
-    _r3l = vle32_v_f32m1(ptr + 6 * 4, vl);
-    _r3h = vle32_v_f32m1(ptr + 7 * 4, vl);
-    _r4l = vle32_v_f32m1(ptr + 8 * 4, vl);
-    _r4h = vle32_v_f32m1(ptr + 9 * 4, vl);
-    _r5l = vle32_v_f32m1(ptr + 10 * 4, vl);
-    _r5h = vle32_v_f32m1(ptr + 11 * 4, vl);
-    _r6l = vle32_v_f32m1(ptr + 12 * 4, vl);
-    _r6h = vle32_v_f32m1(ptr + 13 * 4, vl);
-    _r7l = vle32_v_f32m1(ptr + 14 * 4, vl);
-    _r7h = vle32_v_f32m1(ptr + 15 * 4, vl);
-}
-
-static inline void transpose4x4_ps(vfloat32m1_t& _r0, vfloat32m1_t& _r1, vfloat32m1_t& _r2, vfloat32m1_t& _r3, size_t vl)
-{
-    float tmp[16];
-    vsseg4e32_v_f32m1(&tmp[0], _r0, _r1, _r2, _r3, vl);
-    float* ptr = (float*)tmp;
-    _r0 = vle32_v_f32m1(ptr + 0 * 4, vl);
-    _r1 = vle32_v_f32m1(ptr + 1 * 4, vl);
-    _r2 = vle32_v_f32m1(ptr + 2 * 4, vl);
-    _r3 = vle32_v_f32m1(ptr + 3 * 4, vl);
-}
-
-static inline void transpose8x12_ps(vfloat32m1_t& _r0l, vfloat32m1_t& _r0h,
-                                    vfloat32m1_t& _r1l, vfloat32m1_t& _r1h,
-                                    vfloat32m1_t& _r2l, vfloat32m1_t& _r2h,
-                                    vfloat32m1_t& _r3l, vfloat32m1_t& _r3h,
-                                    vfloat32m1_t& _r4l, vfloat32m1_t& _r4h,
-                                    vfloat32m1_t& _r5l, vfloat32m1_t& _r5h,
-                                    vfloat32m1_t& _r6l, vfloat32m1_t& _r6h,
-                                    vfloat32m1_t& _r7l, vfloat32m1_t& _r7h,
-                                    vfloat32m1_t& _r8l, vfloat32m1_t& _r8h,
-                                    vfloat32m1_t& _r9l, vfloat32m1_t& _r9h,
-                                    vfloat32m1_t& _ral, vfloat32m1_t& _rah,
-                                    vfloat32m1_t& _rbl, vfloat32m1_t& _rbh, size_t vl)
-{
-    float tmp[8][12];
-
-    vsse32_v_f32m1(&tmp[0][0], sizeof(float) * 12, _r0l, vl);
-    vsse32_v_f32m1(&tmp[4][0], sizeof(float) * 12, _r0h, vl);
-    vsse32_v_f32m1(&tmp[0][1], sizeof(float) * 12, _r1l, vl);
-    vsse32_v_f32m1(&tmp[4][1], sizeof(float) * 12, _r1h, vl);
-    vsse32_v_f32m1(&tmp[0][2], sizeof(float) * 12, _r2l, vl);
-    vsse32_v_f32m1(&tmp[4][2], sizeof(float) * 12, _r2h, vl);
-    vsse32_v_f32m1(&tmp[0][3], sizeof(float) * 12, _r3l, vl);
-    vsse32_v_f32m1(&tmp[4][3], sizeof(float) * 12, _r3h, vl);
-    vsse32_v_f32m1(&tmp[0][4], sizeof(float) * 12, _r4l, vl);
-    vsse32_v_f32m1(&tmp[4][4], sizeof(float) * 12, _r4h, vl);
-    vsse32_v_f32m1(&tmp[0][5], sizeof(float) * 12, _r5l, vl);
-    vsse32_v_f32m1(&tmp[4][5], sizeof(float) * 12, _r5h, vl);
-    vsse32_v_f32m1(&tmp[0][6], sizeof(float) * 12, _r6l, vl);
-    vsse32_v_f32m1(&tmp[4][6], sizeof(float) * 12, _r6h, vl);
-    vsse32_v_f32m1(&tmp[0][7], sizeof(float) * 12, _r7l, vl);
-    vsse32_v_f32m1(&tmp[4][7], sizeof(float) * 12, _r7h, vl);
-    vsse32_v_f32m1(&tmp[0][8], sizeof(float) * 12, _r8l, vl);
-    vsse32_v_f32m1(&tmp[4][8], sizeof(float) * 12, _r8h, vl);
-    vsse32_v_f32m1(&tmp[0][9], sizeof(float) * 12, _r9l, vl);
-    vsse32_v_f32m1(&tmp[4][9], sizeof(float) * 12, _r9h, vl);
-    vsse32_v_f32m1(&tmp[0][10], sizeof(float) * 12, _ral, vl);
-    vsse32_v_f32m1(&tmp[4][10], sizeof(float) * 12, _rah, vl);
-    vsse32_v_f32m1(&tmp[0][11], sizeof(float) * 12, _rbl, vl);
-    vsse32_v_f32m1(&tmp[4][11], sizeof(float) * 12, _rbh, vl);
-    float* ptr = (float*)tmp;
-    _r0l = vle32_v_f32m1(ptr + 0 * 4, vl);
-    _r0h = vle32_v_f32m1(ptr + 1 * 4, vl);
-    _r1l = vle32_v_f32m1(ptr + 2 * 4, vl);
-    _r1h = vle32_v_f32m1(ptr + 3 * 4, vl);
-    _r2l = vle32_v_f32m1(ptr + 4 * 4, vl);
-    _r2h = vle32_v_f32m1(ptr + 5 * 4, vl);
-    _r3l = vle32_v_f32m1(ptr + 6 * 4, vl);
-    _r3h = vle32_v_f32m1(ptr + 7 * 4, vl);
-    _r4l = vle32_v_f32m1(ptr + 8 * 4, vl);
-    _r4h = vle32_v_f32m1(ptr + 9 * 4, vl);
-    _r5l = vle32_v_f32m1(ptr + 10 * 4, vl);
-    _r5h = vle32_v_f32m1(ptr + 11 * 4, vl);
-    _r6l = vle32_v_f32m1(ptr + 12 * 4, vl);
-    _r6h = vle32_v_f32m1(ptr + 13 * 4, vl);
-    _r7l = vle32_v_f32m1(ptr + 14 * 4, vl);
-    _r7h = vle32_v_f32m1(ptr + 15 * 4, vl);
-    _r8l = vle32_v_f32m1(ptr + 16 * 4, vl);
-    _r8h = vle32_v_f32m1(ptr + 17 * 4, vl);
-    _r9l = vle32_v_f32m1(ptr + 18 * 4, vl);
-    _r9h = vle32_v_f32m1(ptr + 19 * 4, vl);
-    _ral = vle32_v_f32m1(ptr + 20 * 4, vl);
-    _rah = vle32_v_f32m1(ptr + 21 * 4, vl);
-    _rbl = vle32_v_f32m1(ptr + 22 * 4, vl);
-    _rbh = vle32_v_f32m1(ptr + 23 * 4, vl);
-}
-
-static inline void transpose12x8_ps(vfloat32m1_t& _r0l, vfloat32m1_t& _r0m, vfloat32m1_t& _r0h,
-                                    vfloat32m1_t& _r1l, vfloat32m1_t& _r1m, vfloat32m1_t& _r1h,
-                                    vfloat32m1_t& _r2l, vfloat32m1_t& _r2m, vfloat32m1_t& _r2h,
-                                    vfloat32m1_t& _r3l, vfloat32m1_t& _r3m, vfloat32m1_t& _r3h,
-                                    vfloat32m1_t& _r4l, vfloat32m1_t& _r4m, vfloat32m1_t& _r4h,
-                                    vfloat32m1_t& _r5l, vfloat32m1_t& _r5m, vfloat32m1_t& _r5h,
-                                    vfloat32m1_t& _r6l, vfloat32m1_t& _r6m, vfloat32m1_t& _r6h,
-                                    vfloat32m1_t& _r7l, vfloat32m1_t& _r7m, vfloat32m1_t& _r7h, size_t vl)
-{
-    float tmp[96];
-    vsseg8e32_v_f32m1(&tmp[0], _r0l, _r1l, _r2l, _r3l, _r4l, _r5l, _r6l, _r7l, vl);
-    vsseg8e32_v_f32m1(&tmp[32], _r0m, _r1m, _r2m, _r3m, _r4m, _r5m, _r6m, _r7m, vl);
-    vsseg8e32_v_f32m1(&tmp[64], _r0h, _r1h, _r2h, _r3h, _r4h, _r5h, _r6h, _r7h, vl);
-
-    float* ptr = (float*)tmp;
-    _r0l = vle32_v_f32m1(ptr + 0 * 4, vl);
-    _r0m = vle32_v_f32m1(ptr + 1 * 4, vl);
-    _r0h = vle32_v_f32m1(ptr + 2 * 4, vl);
-    _r1l = vle32_v_f32m1(ptr + 3 * 4, vl);
-    _r1m = vle32_v_f32m1(ptr + 4 * 4, vl);
-    _r1h = vle32_v_f32m1(ptr + 5 * 4, vl);
-    _r2l = vle32_v_f32m1(ptr + 6 * 4, vl);
-    _r2m = vle32_v_f32m1(ptr + 7 * 4, vl);
-    _r2h = vle32_v_f32m1(ptr + 8 * 4, vl);
-    _r3l = vle32_v_f32m1(ptr + 9 * 4, vl);
-    _r3m = vle32_v_f32m1(ptr + 10 * 4, vl);
-    _r3h = vle32_v_f32m1(ptr + 11 * 4, vl);
-    _r4l = vle32_v_f32m1(ptr + 12 * 4, vl);
-    _r4m = vle32_v_f32m1(ptr + 13 * 4, vl);
-    _r4h = vle32_v_f32m1(ptr + 14 * 4, vl);
-    _r5l = vle32_v_f32m1(ptr + 15 * 4, vl);
-    _r5m = vle32_v_f32m1(ptr + 16 * 4, vl);
-    _r5h = vle32_v_f32m1(ptr + 17 * 4, vl);
-    _r6l = vle32_v_f32m1(ptr + 18 * 4, vl);
-    _r6m = vle32_v_f32m1(ptr + 19 * 4, vl);
-    _r6h = vle32_v_f32m1(ptr + 20 * 4, vl);
-    _r7l = vle32_v_f32m1(ptr + 21 * 4, vl);
-    _r7m = vle32_v_f32m1(ptr + 22 * 4, vl);
-    _r7h = vle32_v_f32m1(ptr + 23 * 4, vl);
-}
-
-static inline void transpose4x8_ps(vfloat32m1_t& _r0, vfloat32m1_t& _r1, vfloat32m1_t& _r2, vfloat32m1_t& _r3, vfloat32m1_t& _r4, vfloat32m1_t& _r5, vfloat32m1_t& _r6, vfloat32m1_t& _r7, size_t vl)
-{
-    float tmp[32];
-    vsseg8e32_v_f32m1(&tmp[0], _r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7, vl);
-
-    float* ptr = (float*)tmp;
-    _r0 = vle32_v_f32m1(ptr + 0 * 4, vl);
-    _r1 = vle32_v_f32m1(ptr + 1 * 4, vl);
-    _r2 = vle32_v_f32m1(ptr + 2 * 4, vl);
-    _r3 = vle32_v_f32m1(ptr + 3 * 4, vl);
-    _r4 = vle32_v_f32m1(ptr + 4 * 4, vl);
-    _r5 = vle32_v_f32m1(ptr + 5 * 4, vl);
-    _r6 = vle32_v_f32m1(ptr + 6 * 4, vl);
-    _r7 = vle32_v_f32m1(ptr + 7 * 4, vl);
-}
-
-static inline void transpose4x12_ps(vfloat32m1_t& _r0, vfloat32m1_t& _r1, vfloat32m1_t& _r2, vfloat32m1_t& _r3, vfloat32m1_t& _r4, vfloat32m1_t& _r5, vfloat32m1_t& _r6, vfloat32m1_t& _r7, vfloat32m1_t& _r8, vfloat32m1_t& _r9, vfloat32m1_t& _ra, vfloat32m1_t& _rb, size_t vl)
-{
-    float tmp[4][12];
-    vsse32_v_f32m1(&tmp[0][0], sizeof(float) * 12, _r0, vl);
-    vsse32_v_f32m1(&tmp[0][1], sizeof(float) * 12, _r1, vl);
-    vsse32_v_f32m1(&tmp[0][2], sizeof(float) * 12, _r2, vl);
-    vsse32_v_f32m1(&tmp[0][3], sizeof(float) * 12, _r3, vl);
-    vsse32_v_f32m1(&tmp[0][4], sizeof(float) * 12, _r4, vl);
-    vsse32_v_f32m1(&tmp[0][5], sizeof(float) * 12, _r5, vl);
-    vsse32_v_f32m1(&tmp[0][6], sizeof(float) * 12, _r6, vl);
-    vsse32_v_f32m1(&tmp[0][7], sizeof(float) * 12, _r7, vl);
-    vsse32_v_f32m1(&tmp[0][8], sizeof(float) * 12, _r8, vl);
-    vsse32_v_f32m1(&tmp[0][9], sizeof(float) * 12, _r9, vl);
-    vsse32_v_f32m1(&tmp[0][10], sizeof(float) * 12, _ra, vl);
-    vsse32_v_f32m1(&tmp[0][11], sizeof(float) * 12, _rb, vl);
-    float* ptr = (float*)tmp;
-    _r0 = vle32_v_f32m1(ptr + 0 * 4, vl);
-    _r1 = vle32_v_f32m1(ptr + 1 * 4, vl);
-    _r2 = vle32_v_f32m1(ptr + 2 * 4, vl);
-    _r3 = vle32_v_f32m1(ptr + 3 * 4, vl);
-    _r4 = vle32_v_f32m1(ptr + 4 * 4, vl);
-    _r5 = vle32_v_f32m1(ptr + 5 * 4, vl);
-    _r6 = vle32_v_f32m1(ptr + 6 * 4, vl);
-    _r7 = vle32_v_f32m1(ptr + 7 * 4, vl);
-    _r8 = vle32_v_f32m1(ptr + 8 * 4, vl);
-    _r9 = vle32_v_f32m1(ptr + 9 * 4, vl);
-    _ra = vle32_v_f32m1(ptr + 10 * 4, vl);
-    _rb = vle32_v_f32m1(ptr + 11 * 4, vl);
-}
-
-static inline void transpose8x4_ps(vfloat32m1_t& _r0l, vfloat32m1_t& _r0h,
-                                   vfloat32m1_t& _r1l, vfloat32m1_t& _r1h,
-                                   vfloat32m1_t& _r2l, vfloat32m1_t& _r2h,
-                                   vfloat32m1_t& _r3l, vfloat32m1_t& _r3h, size_t vl)
-{
-    float tmp[32];
-    vsseg4e32_v_f32m1(&tmp[0], _r0l, _r1l, _r2l, _r3l, vl);
-    vsseg4e32_v_f32m1(&tmp[16], _r0h, _r1h, _r2h, _r3h, vl);
-    float* ptr = (float*)tmp;
-    _r0l = vle32_v_f32m1(ptr + 0 * 4, vl);
-    _r0h = vle32_v_f32m1(ptr + 1 * 4, vl);
-    _r1l = vle32_v_f32m1(ptr + 2 * 4, vl);
-    _r1h = vle32_v_f32m1(ptr + 3 * 4, vl);
-    _r2l = vle32_v_f32m1(ptr + 4 * 4, vl);
-    _r2h = vle32_v_f32m1(ptr + 5 * 4, vl);
-    _r3l = vle32_v_f32m1(ptr + 6 * 4, vl);
-    _r3h = vle32_v_f32m1(ptr + 7 * 4, vl);
-}
-#endif
-
-#endif // RISCV_USABILITY_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/rvv_mathfun.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/rvv_mathfun.h
deleted file mode 100644
index 34f07278..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/rvv_mathfun.h
+++ /dev/null
@@ -1,583 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef RVV_MATHFUN_H
-#define RVV_MATHFUN_H
-
-#include <riscv_vector.h>
-
-#define c_inv_mant_mask ~0x7f800000u
-#define c_cephes_SQRTHF 0.707106781186547524
-#define c_cephes_log_p0 7.0376836292E-2
-#define c_cephes_log_p1 -1.1514610310E-1
-#define c_cephes_log_p2 1.1676998740E-1
-#define c_cephes_log_p3 -1.2420140846E-1
-#define c_cephes_log_p4 +1.4249322787E-1
-#define c_cephes_log_p5 -1.6668057665E-1
-#define c_cephes_log_p6 +2.0000714765E-1
-#define c_cephes_log_p7 -2.4999993993E-1
-#define c_cephes_log_p8 +3.3333331174E-1
-#define c_cephes_log_q1 -2.12194440e-4
-#define c_cephes_log_q2 0.693359375
-
-#define _RVV_FLOAT32_LOG_OP(LMUL, MLEN)                                                                              \
-    static inline vfloat32m##LMUL##_t log_ps(vfloat32m##LMUL##_t x, size_t vl)                                       \
-    {                                                                                                                \
-        x = vfmax_vf_f32m##LMUL(x, 0.f, vl); /* force flush to zero on denormal values */                            \
-        vbool##MLEN##_t invalid_mask = vmfle_vf_f32m##LMUL##_b##MLEN(x, 0.f, vl);                                    \
-                                                                                                                     \
-        vint32m##LMUL##_t ux = vreinterpret_v_f32m##LMUL##_i32m##LMUL(x);                                            \
-                                                                                                                     \
-        vint32m##LMUL##_t emm0 = vsra_vx_i32m##LMUL(ux, 23, vl);                                                     \
-                                                                                                                     \
-        /* keep only the fractional part */                                                                          \
-        ux = vand_vx_i32m##LMUL(ux, c_inv_mant_mask, vl);                                                            \
-        ux = vor_vx_i32m##LMUL(ux, 1056964608 /* reinterpret_cast<int>(0.5) */, vl);                                 \
-        x = vreinterpret_v_i32m##LMUL##_f32m##LMUL(ux);                                                              \
-                                                                                                                     \
-        emm0 = vsub_vx_i32m##LMUL(emm0, 0x7f, vl);                                                                   \
-        vfloat32m##LMUL##_t e = vfcvt_f_x_v_f32m##LMUL(emm0, vl);                                                    \
-                                                                                                                     \
-        e = vfadd_vf_f32m##LMUL(e, 1.f, vl);                                                                         \
-                                                                                                                     \
-        /* part2:                      */                                                                            \
-        /*     if( x < SQRTHF ) {      */                                                                            \
-        /*       e -= 1;               */                                                                            \
-        /*       x = x + x - 1.0;      */                                                                            \
-        /*     } else { x = x - 1.0; } */                                                                            \
-        vbool##MLEN##_t mask = vmflt_vf_f32m##LMUL##_b##MLEN(x, c_cephes_SQRTHF, vl);                                \
-        x = vfadd_vv_f32m##LMUL##_m(mask, x, x, x, vl);                                                              \
-        x = vfsub_vf_f32m##LMUL(x, 1.f, vl);                                                                         \
-        e = vfsub_vf_f32m##LMUL##_m(mask, e, e, 1.f, vl);                                                            \
-                                                                                                                     \
-        vfloat32m##LMUL##_t z = vfmul_vv_f32m##LMUL(x, x, vl);                                                       \
-                                                                                                                     \
-        vfloat32m##LMUL##_t y = vfmul_vf_f32m##LMUL(x, c_cephes_log_p0, vl);                                         \
-        y = vfadd_vf_f32m##LMUL(y, c_cephes_log_p1, vl);                                                             \
-        y = vfmul_vv_f32m##LMUL(y, x, vl);                                                                           \
-        y = vfadd_vf_f32m##LMUL(y, c_cephes_log_p2, vl);                                                             \
-        y = vfmul_vv_f32m##LMUL(y, x, vl);                                                                           \
-        y = vfadd_vf_f32m##LMUL(y, c_cephes_log_p3, vl);                                                             \
-        y = vfmul_vv_f32m##LMUL(y, x, vl);                                                                           \
-        y = vfadd_vf_f32m##LMUL(y, c_cephes_log_p4, vl);                                                             \
-        y = vfmul_vv_f32m##LMUL(y, x, vl);                                                                           \
-        y = vfadd_vf_f32m##LMUL(y, c_cephes_log_p5, vl);                                                             \
-        y = vfmul_vv_f32m##LMUL(y, x, vl);                                                                           \
-        y = vfadd_vf_f32m##LMUL(y, c_cephes_log_p6, vl);                                                             \
-        y = vfmul_vv_f32m##LMUL(y, x, vl);                                                                           \
-        y = vfadd_vf_f32m##LMUL(y, c_cephes_log_p7, vl);                                                             \
-        y = vfmul_vv_f32m##LMUL(y, x, vl);                                                                           \
-        y = vfadd_vf_f32m##LMUL(y, c_cephes_log_p8, vl);                                                             \
-        y = vfmul_vv_f32m##LMUL(y, x, vl);                                                                           \
-                                                                                                                     \
-        y = vfmul_vv_f32m##LMUL(y, z, vl);                                                                           \
-                                                                                                                     \
-        vfloat32m##LMUL##_t tmp = vfmul_vf_f32m##LMUL(e, c_cephes_log_q1, vl);                                       \
-        y = vfadd_vv_f32m##LMUL(y, tmp, vl);                                                                         \
-                                                                                                                     \
-        tmp = vfmul_vf_f32m##LMUL(z, 0.5f, vl);                                                                      \
-        y = vfsub_vv_f32m##LMUL(y, tmp, vl);                                                                         \
-                                                                                                                     \
-        tmp = vfmul_vf_f32m##LMUL(e, c_cephes_log_q2, vl);                                                           \
-        x = vfadd_vv_f32m##LMUL(x, y, vl);                                                                           \
-        x = vfadd_vv_f32m##LMUL(x, tmp, vl);                                                                         \
-        /* negative arg will be NAN */                                                                               \
-        vuint32m##LMUL##_t xtmp = vreinterpret_v_f32m##LMUL##_u32m##LMUL(x);                                         \
-        x = vreinterpret_v_u32m##LMUL##_f32m##LMUL(vor_vx_u32m##LMUL##_m(invalid_mask, xtmp, xtmp, 0xffffffff, vl)); \
-        return x;                                                                                                    \
-    }
-
-_RVV_FLOAT32_LOG_OP(1, 32)
-_RVV_FLOAT32_LOG_OP(2, 16)
-_RVV_FLOAT32_LOG_OP(4, 8)
-_RVV_FLOAT32_LOG_OP(8, 4)
-
-#define c_exp_hi 88.3762626647949f
-#define c_exp_lo -88.3762626647949f
-
-#define c_cephes_LOG2EF 1.44269504088896341
-#define c_cephes_exp_C1 0.693359375
-#define c_cephes_exp_C2 -2.12194440e-4
-
-#define c_cephes_exp_p0 1.9875691500E-4
-#define c_cephes_exp_p1 1.3981999507E-3
-#define c_cephes_exp_p2 8.3334519073E-3
-#define c_cephes_exp_p3 4.1665795894E-2
-#define c_cephes_exp_p4 1.6666665459E-1
-#define c_cephes_exp_p5 5.0000001201E-1
-
-#define _RVV_FLOAT32_EXP_OP(LMUL, MLEN)                                                   \
-    static inline vfloat32m##LMUL##_t exp_ps(vfloat32m##LMUL##_t x, size_t vl)            \
-    {                                                                                     \
-        vfloat32m##LMUL##_t tmp, fx;                                                      \
-                                                                                          \
-        x = vfmin_vf_f32m##LMUL(x, c_exp_hi, vl);                                         \
-        x = vfmax_vf_f32m##LMUL(x, c_exp_lo, vl);                                         \
-                                                                                          \
-        /* express exp(x) as exp(g + n*log(2)) */                                         \
-        fx = vfmacc_vf_f32m##LMUL(vfmv_v_f_f32m##LMUL(0.5f, vl), c_cephes_LOG2EF, x, vl); \
-                                                                                          \
-        /* perform a floorf */                                                            \
-        tmp = vfcvt_f_x_v_f32m##LMUL(vfcvt_x_f_v_i32m##LMUL(fx, vl), vl);                 \
-                                                                                          \
-        /* if greater, substract 1 */                                                     \
-        vbool##MLEN##_t mask = vmfgt_vv_f32m##LMUL##_b##MLEN(tmp, fx, vl);                \
-        fx = vfsub_vf_f32m##LMUL##_m(mask, tmp, tmp, 1.f, vl);                            \
-                                                                                          \
-        tmp = vfmul_vf_f32m##LMUL(fx, c_cephes_exp_C1, vl);                               \
-        vfloat32m##LMUL##_t z = vfmul_vf_f32m##LMUL(fx, c_cephes_exp_C2, vl);             \
-        x = vfsub_vv_f32m##LMUL(x, tmp, vl);                                              \
-        x = vfsub_vv_f32m##LMUL(x, z, vl);                                                \
-                                                                                          \
-        vfloat32m##LMUL##_t y = vfmul_vf_f32m##LMUL(x, c_cephes_exp_p0, vl);              \
-        z = vfmul_vv_f32m##LMUL(x, x, vl);                                                \
-                                                                                          \
-        y = vfadd_vf_f32m##LMUL(y, c_cephes_exp_p1, vl);                                  \
-        y = vfmul_vv_f32m##LMUL(y, x, vl);                                                \
-        y = vfadd_vf_f32m##LMUL(y, c_cephes_exp_p2, vl);                                  \
-        y = vfmul_vv_f32m##LMUL(y, x, vl);                                                \
-        y = vfadd_vf_f32m##LMUL(y, c_cephes_exp_p3, vl);                                  \
-        y = vfmul_vv_f32m##LMUL(y, x, vl);                                                \
-        y = vfadd_vf_f32m##LMUL(y, c_cephes_exp_p4, vl);                                  \
-        y = vfmul_vv_f32m##LMUL(y, x, vl);                                                \
-        y = vfadd_vf_f32m##LMUL(y, c_cephes_exp_p5, vl);                                  \
-                                                                                          \
-        y = vfmul_vv_f32m##LMUL(y, z, vl);                                                \
-        y = vfadd_vv_f32m##LMUL(y, x, vl);                                                \
-        y = vfadd_vf_f32m##LMUL(y, 1.f, vl);                                              \
-                                                                                          \
-        /* build 2^n */                                                                   \
-        vint32m##LMUL##_t mm = vfcvt_x_f_v_i32m##LMUL(fx, vl);                            \
-        mm = vadd_vx_i32m##LMUL(mm, 0x7f, vl);                                            \
-        mm = vsll_vx_i32m##LMUL(mm, 23, vl);                                              \
-        vfloat32m##LMUL##_t pow2n = vreinterpret_v_i32m##LMUL##_f32m##LMUL(mm);           \
-                                                                                          \
-        y = vfmul_vv_f32m##LMUL(y, pow2n, vl);                                            \
-        return y;                                                                         \
-    }
-
-_RVV_FLOAT32_EXP_OP(1, 32)
-_RVV_FLOAT32_EXP_OP(2, 16)
-_RVV_FLOAT32_EXP_OP(4, 8)
-_RVV_FLOAT32_EXP_OP(8, 4)
-
-#define c_minus_cephes_DP1 -0.78515625
-#define c_minus_cephes_DP2 -2.4187564849853515625e-4
-#define c_minus_cephes_DP3 -3.77489497744594108e-8
-#define c_sincof_p0        -1.9515295891E-4
-#define c_sincof_p1        8.3321608736E-3
-#define c_sincof_p2        -1.6666654611E-1
-#define c_coscof_p0        2.443315711809948E-005
-#define c_coscof_p1        -1.388731625493765E-003
-#define c_coscof_p2        4.166664568298827E-002
-#define c_cephes_FOPI      1.27323954473516 // 4 / M_PI
-
-#define _RVV_FLOAT32_SINCOS_OP(LMUL, MLEN)                                                                                          \
-    static inline void sincos_ps(vfloat32m##LMUL##_t x, vfloat32m##LMUL##_t* ysin, vfloat32m##LMUL##_t* ycos, size_t vl)            \
-    {                                                                                                                               \
-        /* any x */                                                                                                                 \
-        vfloat32m##LMUL##_t xmm1, xmm2, xmm3, y;                                                                                    \
-                                                                                                                                    \
-        vuint32m##LMUL##_t emm2;                                                                                                    \
-                                                                                                                                    \
-        vbool##MLEN##_t sign_mask_sin, sign_mask_cos;                                                                               \
-        sign_mask_sin = vmflt_vf_f32m##LMUL##_b##MLEN(x, 0.f, vl);                                                                  \
-        x = vfsgnj_vf_f32m##LMUL(x, 1.f, vl);                                                                                       \
-                                                                                                                                    \
-        /* scale by 4/Pi */                                                                                                         \
-        y = vfmul_vf_f32m##LMUL(x, c_cephes_FOPI, vl);                                                                              \
-                                                                                                                                    \
-        /* store the integer part of y in mm0 */                                                                                    \
-        emm2 = vfcvt_xu_f_v_u32m##LMUL(y, vl);                                                                                      \
-        /* j=(j+1) & (~1) (see the cephes sources) */                                                                               \
-        emm2 = vadd_vx_u32m##LMUL(emm2, 1, vl);                                                                                     \
-        emm2 = vand_vx_u32m##LMUL(emm2, ~1, vl);                                                                                    \
-        y = vfcvt_f_xu_v_f32m##LMUL(emm2, vl);                                                                                      \
-                                                                                                                                    \
-        /* get the polynom selection mask              */                                                                           \
-        /*     there is one polynom for 0 <= x <= Pi/4 */                                                                           \
-        /*     and another one for Pi/4<x<=Pi/2        */                                                                           \
-        /*                                             */                                                                           \
-        /*     Both branches will be computed.         */                                                                           \
-        vbool##MLEN##_t poly_mask = vmsne_vx_u32m##LMUL##_b##MLEN(vand_vx_u32m##LMUL(emm2, 2, vl), 0, vl);                          \
-                                                                                                                                    \
-        /* The magic pass: "Extended precision modular arithmetic" */                                                               \
-        /*     x = ((x - y * DP1) - y * DP2) - y * DP3;            */                                                               \
-        xmm1 = vfmul_vf_f32m##LMUL(y, c_minus_cephes_DP1, vl);                                                                      \
-        xmm2 = vfmul_vf_f32m##LMUL(y, c_minus_cephes_DP2, vl);                                                                      \
-        xmm3 = vfmul_vf_f32m##LMUL(y, c_minus_cephes_DP3, vl);                                                                      \
-        x = vfadd_vv_f32m##LMUL(x, xmm1, vl);                                                                                       \
-        x = vfadd_vv_f32m##LMUL(x, xmm2, vl);                                                                                       \
-        x = vfadd_vv_f32m##LMUL(x, xmm3, vl);                                                                                       \
-                                                                                                                                    \
-        sign_mask_sin = vmxor_mm_b##MLEN(sign_mask_sin, vmsne_vx_u32m##LMUL##_b##MLEN(vand_vx_u32m##LMUL(emm2, 4, vl), 0, vl), vl); \
-        sign_mask_cos = vmsne_vx_u32m##LMUL##_b##MLEN(vand_vx_u32m##LMUL(vsub_vx_u32m##LMUL(emm2, 2, vl), 4, vl), 0, vl);           \
-                                                                                                                                    \
-        /* Evaluate the first polynom  (0 <= x <= Pi/4) in y1, */                                                                   \
-        /*     and the second polynom  (Pi/4 <= x <= 0) in y2  */                                                                   \
-        vfloat32m##LMUL##_t z = vfmul_vv_f32m##LMUL(x, x, vl);                                                                      \
-        vfloat32m##LMUL##_t y1, y2;                                                                                                 \
-                                                                                                                                    \
-        y1 = vfmul_vf_f32m##LMUL(z, c_coscof_p0, vl);                                                                               \
-        y2 = vfmul_vf_f32m##LMUL(z, c_sincof_p0, vl);                                                                               \
-        y1 = vfadd_vf_f32m##LMUL(y1, c_coscof_p1, vl);                                                                              \
-        y2 = vfadd_vf_f32m##LMUL(y2, c_sincof_p1, vl);                                                                              \
-        y1 = vfmul_vv_f32m##LMUL(y1, z, vl);                                                                                        \
-        y2 = vfmul_vv_f32m##LMUL(y2, z, vl);                                                                                        \
-        y1 = vfadd_vf_f32m##LMUL(y1, c_coscof_p2, vl);                                                                              \
-        y2 = vfadd_vf_f32m##LMUL(y2, c_sincof_p2, vl);                                                                              \
-        y1 = vfmul_vv_f32m##LMUL(y1, z, vl);                                                                                        \
-        y2 = vfmul_vv_f32m##LMUL(y2, z, vl);                                                                                        \
-        y1 = vfmul_vv_f32m##LMUL(y1, z, vl);                                                                                        \
-        y2 = vfmul_vv_f32m##LMUL(y2, x, vl);                                                                                        \
-        y1 = vfsub_vv_f32m##LMUL(y1, vfmul_vf_f32m##LMUL(z, 0.5f, vl), vl);                                                         \
-        y2 = vfadd_vv_f32m##LMUL(y2, x, vl);                                                                                        \
-        y1 = vfadd_vf_f32m##LMUL(y1, 1.f, vl);                                                                                      \
-                                                                                                                                    \
-        /* select the correct result from the two polynoms */                                                                       \
-        vfloat32m##LMUL##_t ys = vmerge_vvm_f32m##LMUL(poly_mask, y2, y1, vl);                                                      \
-        vfloat32m##LMUL##_t yc = vmerge_vvm_f32m##LMUL(poly_mask, y1, y2, vl);                                                      \
-        *ysin = vmerge_vvm_f32m##LMUL(sign_mask_sin, ys, vfneg_v_f32m##LMUL(ys, vl), vl);                                           \
-        *ycos = vmerge_vvm_f32m##LMUL(sign_mask_cos, vfneg_v_f32m##LMUL(yc, vl), yc, vl);                                           \
-    }
-
-_RVV_FLOAT32_SINCOS_OP(1, 32)
-_RVV_FLOAT32_SINCOS_OP(2, 16)
-_RVV_FLOAT32_SINCOS_OP(4, 8)
-_RVV_FLOAT32_SINCOS_OP(8, 4)
-
-#define _RVV_FLOAT32_SIN_OP(LMUL, MLEN)                                        \
-    static inline vfloat32m##LMUL##_t sin_ps(vfloat32m##LMUL##_t x, size_t vl) \
-    {                                                                          \
-        vfloat32m##LMUL##_t ysin, ycos;                                        \
-        sincos_ps(x, &ysin, &ycos, vl);                                        \
-        return ysin;                                                           \
-    }
-
-_RVV_FLOAT32_SIN_OP(1, 32)
-_RVV_FLOAT32_SIN_OP(2, 16)
-_RVV_FLOAT32_SIN_OP(4, 8)
-_RVV_FLOAT32_SIN_OP(8, 4)
-
-#define _RVV_FLOAT32_COS_OP(LMUL, MLEN)                                        \
-    static inline vfloat32m##LMUL##_t cos_ps(vfloat32m##LMUL##_t x, size_t vl) \
-    {                                                                          \
-        vfloat32m##LMUL##_t ysin, ycos;                                        \
-        sincos_ps(x, &ysin, &ycos, vl);                                        \
-        return ycos;                                                           \
-    }
-
-_RVV_FLOAT32_COS_OP(1, 32)
-_RVV_FLOAT32_COS_OP(2, 16)
-_RVV_FLOAT32_COS_OP(4, 8)
-_RVV_FLOAT32_COS_OP(8, 4)
-
-#define c_tanh_tiny 1e-4f
-#define c_tanh_hi   9.0f
-// The monomial coefficients of the numerator polynomial (odd).
-#define c_tanh_alpha_1  4.89352455891786e-3f
-#define c_tanh_alpha_3  6.37261928875436e-4f
-#define c_tanh_alpha_5  1.48572235717979e-5f
-#define c_tanh_alpha_7  5.12229709037114e-8f
-#define c_tanh_alpha_9  -8.60467152213735e-11f
-#define c_tanh_alpha_11 2.00018790482477e-13f
-#define c_tanh_alpha_13 -2.76076847742355e-16f
-// The monomial coefficients of the denominator polynomial (even).
-#define c_tanh_beta_0 4.89352518554385e-3f
-#define c_tanh_beta_2 2.26843463243900e-3f
-#define c_tanh_beta_4 1.18534705686654e-4f
-#define c_tanh_beta_6 1.19825839466702e-6f
-
-#define _RVV_FLOAT32_TANH_OP(LMUL, MLEN)                                                         \
-    static inline vfloat32m##LMUL##_t tanh_ps(vfloat32m##LMUL##_t x, size_t vl)                  \
-    {                                                                                            \
-        vfloat32m##LMUL##_t x2 = vfsgnj_vf_f32m##LMUL(x, 1.f, vl);                               \
-                                                                                                 \
-        vbool##MLEN##_t tiny_mask = vmfge_vf_f32m##LMUL##_b##MLEN(x2, c_tanh_tiny, vl);          \
-                                                                                                 \
-        /* clamp the inputs to the range [-9, 9] since anything outside */                       \
-        /* this range is -/+1.0f in single-precision.                   */                       \
-        x2 = vfmin_vf_f32m##LMUL(x, c_tanh_hi, vl);                                              \
-                                                                                                 \
-        /* since the polynomials are odd/even, we need x**2. */                                  \
-        vfloat32m##LMUL##_t z = vfmul_vv_f32m##LMUL(x2, x2, vl);                                 \
-                                                                                                 \
-        /* evaluate the numerator polynomial y. */                                               \
-        vfloat32m##LMUL##_t y = vfmul_vf_f32m##LMUL(z, c_tanh_alpha_13, vl);                     \
-        y = vfadd_vf_f32m##LMUL(y, c_tanh_alpha_11, vl);                                         \
-        y = vfmul_vv_f32m##LMUL(y, z, vl);                                                       \
-        y = vfadd_vf_f32m##LMUL(y, c_tanh_alpha_9, vl);                                          \
-        y = vfmul_vv_f32m##LMUL(y, z, vl);                                                       \
-        y = vfadd_vf_f32m##LMUL(y, c_tanh_alpha_7, vl);                                          \
-        y = vfmul_vv_f32m##LMUL(y, z, vl);                                                       \
-        y = vfadd_vf_f32m##LMUL(y, c_tanh_alpha_5, vl);                                          \
-        y = vfmul_vv_f32m##LMUL(y, z, vl);                                                       \
-        y = vfadd_vf_f32m##LMUL(y, c_tanh_alpha_3, vl);                                          \
-        y = vfmul_vv_f32m##LMUL(y, z, vl);                                                       \
-        y = vfadd_vf_f32m##LMUL(y, c_tanh_alpha_1, vl);                                          \
-        y = vfmul_vv_f32m##LMUL(y, x2, vl);                                                      \
-                                                                                                 \
-        /* evaluate the denominator polynomial w. */                                             \
-        vfloat32m##LMUL##_t w = vfmul_vf_f32m##LMUL(z, c_tanh_beta_6, vl);                       \
-        w = vfadd_vf_f32m##LMUL(w, c_tanh_beta_4, vl);                                           \
-        w = vfmul_vv_f32m##LMUL(w, z, vl);                                                       \
-        w = vfadd_vf_f32m##LMUL(w, c_tanh_beta_2, vl);                                           \
-        w = vfmul_vv_f32m##LMUL(w, z, vl);                                                       \
-        w = vfadd_vf_f32m##LMUL(w, c_tanh_beta_0, vl);                                           \
-                                                                                                 \
-        /* divide the numerator by the denominator. */                                           \
-        y = vfdiv_vv_f32m##LMUL(y, w, vl);                                                       \
-                                                                                                 \
-        /* reinstate the sign.  */                                                               \
-        y = vfsgnj_vv_f32m##LMUL(y, x, vl);                                                      \
-                                                                                                 \
-        /* when the argument is very small in magnitude it's more accurate to just return it. */ \
-        y = vmerge_vvm_f32m##LMUL(tiny_mask, x, y, vl);                                          \
-                                                                                                 \
-        return y;                                                                                \
-    }
-
-_RVV_FLOAT32_TANH_OP(1, 32)
-_RVV_FLOAT32_TANH_OP(2, 16)
-_RVV_FLOAT32_TANH_OP(4, 8)
-_RVV_FLOAT32_TANH_OP(8, 4)
-
-#define _RVV_FLOAT32_POW_OP(LMUL, MLEN)                                                               \
-    static inline vfloat32m##LMUL##_t pow_ps(vfloat32m##LMUL##_t a, vfloat32m##LMUL##_t b, size_t vl) \
-    {                                                                                                 \
-        /* pow(x, m) = exp(m * log(x)) */                                                             \
-        return exp_ps(vfmul_vv_f32m##LMUL(b, log_ps(a, vl), vl), vl);                                 \
-    }
-
-_RVV_FLOAT32_POW_OP(1, 32)
-_RVV_FLOAT32_POW_OP(2, 16)
-_RVV_FLOAT32_POW_OP(4, 8)
-_RVV_FLOAT32_POW_OP(8, 4)
-
-#if C906
-#define _RVV_FLOAT32_SIGMOID_OP(LMUL, MLEN)                                                                                                \
-    static inline vfloat32m##LMUL##_t sigmoid_ps(vfloat32m##LMUL##_t _v, size_t vl)                                                        \
-    {                                                                                                                                      \
-        _v = vfneg_v_f32m##LMUL(_v, vl);                                                                                                   \
-        _v = exp_ps(_v, vl);                                                                                                               \
-        _v = vfadd_vf_f32m##LMUL(_v, 1.f, vl);                                                                                             \
-        vfloat32m##LMUL##_t _reciprocal = vfrdiv_vf_f32m##LMUL(_v, 1.f, vl);                                                               \
-        _reciprocal = vfmul_vv_f32m##LMUL(vfrsub_vf_f32m##LMUL(vfmul_vv_f32m##LMUL(_v, _reciprocal, vl), 2.f, vl), _reciprocal, vl);       \
-        /* _reciprocal = vfmul_vv_f32m##LMUL(vfrsub_vf_f32m##LMUL(vfmul_vv_f32m##LMUL(_v, _reciprocal, vl), 2.f, vl), _reciprocal, vl); */ \
-        return _reciprocal;                                                                                                                \
-    }
-#else // C906
-#define _RVV_FLOAT32_SIGMOID_OP(LMUL, MLEN)                                                                                                \
-    static inline vfloat32m##LMUL##_t sigmoid_ps(vfloat32m##LMUL##_t _v, size_t vl)                                                        \
-    {                                                                                                                                      \
-        _v = vfneg_v_f32m##LMUL(_v, vl);                                                                                                   \
-        _v = exp_ps(_v, vl);                                                                                                               \
-        _v = vfadd_vf_f32m##LMUL(_v, 1.f, vl);                                                                                             \
-        vfloat32m##LMUL##_t _reciprocal = vfrec7_v_f32m##LMUL(_v, vl);                                                                     \
-        _reciprocal = vfmul_vv_f32m##LMUL(vfrsub_vf_f32m##LMUL(vfmul_vv_f32m##LMUL(_v, _reciprocal, vl), 2.f, vl), _reciprocal, vl);       \
-        /* _reciprocal = vfmul_vv_f32m##LMUL(vfrsub_vf_f32m##LMUL(vfmul_vv_f32m##LMUL(_v, _reciprocal, vl), 2.f, vl), _reciprocal, vl); */ \
-        return _reciprocal;                                                                                                                \
-    }
-#endif // C906
-
-_RVV_FLOAT32_SIGMOID_OP(1, 32)
-_RVV_FLOAT32_SIGMOID_OP(2, 16)
-_RVV_FLOAT32_SIGMOID_OP(4, 8)
-_RVV_FLOAT32_SIGMOID_OP(8, 4)
-
-/*
- * ====================================================
- * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
- *
- * Developed at SunPro, a Sun Microsystems, Inc. business.
- * Permission to use, copy, modify, and distribute this
- * software is freely granted, provided that this notice
- * is preserved.
- * ====================================================
- */
-#define c_erfc_erx_f 8.4506291151e-01f /* 0x3f58560b */
-// Coefficients for approximation to  erf on [00.84375]
-#define c_erfc_efx  1.2837916613e-01f /* 0x3e0375d4 */
-#define c_erfc_efx8 1.0270333290e+00f /* 0x3f8375d4 */
-
-#define c_erfc_pp0 1.2837916613e-01f  /* 0x3e0375d4 */
-#define c_erfc_pp1 -3.2504209876e-01f /* 0xbea66beb */
-#define c_erfc_pp2 -2.8481749818e-02f /* 0xbce9528f */
-#define c_erfc_pp3 -5.7702702470e-03f /* 0xbbbd1489 */
-#define c_erfc_pp4 -2.3763017452e-05f /* 0xb7c756b1 */
-#define c_erfc_qq1 3.9791721106e-01f  /* 0x3ecbbbce */
-#define c_erfc_qq2 6.5022252500e-02f  /* 0x3d852a63 */
-#define c_erfc_qq3 5.0813062117e-03f  /* 0x3ba68116 */
-#define c_erfc_qq4 1.3249473704e-04f  /* 0x390aee49 */
-#define c_erfc_qq5 -3.9602282413e-06f /* 0xb684e21a */
-
-// Coefficients for approximation to  erf  in [0.843751.25]
-#define c_erfc_pa0 -2.3621185683e-03f /* 0xbb1acdc6 */
-#define c_erfc_pa1 4.1485610604e-01f  /* 0x3ed46805 */
-#define c_erfc_pa2 -3.7220788002e-01f /* 0xbebe9208 */
-#define c_erfc_pa3 3.1834661961e-01f  /* 0x3ea2fe54 */
-#define c_erfc_pa4 -1.1089469492e-01f /* 0xbde31cc2 */
-#define c_erfc_pa5 3.5478305072e-02f  /* 0x3d1151b3 */
-#define c_erfc_pa6 -2.1663755178e-03f /* 0xbb0df9c0 */
-#define c_erfc_qa1 1.0642088205e-01f  /* 0x3dd9f331 */
-#define c_erfc_qa2 5.4039794207e-01f  /* 0x3f0a5785 */
-#define c_erfc_qa3 7.1828655899e-02f  /* 0x3d931ae7 */
-#define c_erfc_qa4 1.2617121637e-01f  /* 0x3e013307 */
-#define c_erfc_qa5 1.3637083583e-02f  /* 0x3c5f6e13 */
-#define c_erfc_qa6 1.1984500103e-02f  /* 0x3c445aa3 */
-
-// Coefficients for approximation to  erfc in [1.251/0.35]
-#define c_erfc_ra0 -9.8649440333e-03f /* 0xbc21a093 */
-#define c_erfc_ra1 -6.9385856390e-01f /* 0xbf31a0b7 */
-#define c_erfc_ra2 -1.0558626175e+01f /* 0xc128f022 */
-#define c_erfc_ra3 -6.2375331879e+01f /* 0xc2798057 */
-#define c_erfc_ra4 -1.6239666748e+02f /* 0xc322658c */
-#define c_erfc_ra5 -1.8460508728e+02f /* 0xc3389ae7 */
-#define c_erfc_ra6 -8.1287437439e+01f /* 0xc2a2932b */
-#define c_erfc_ra7 -9.8143291473e+00f /* 0xc11d077e */
-#define c_erfc_sa1 1.9651271820e+01f  /* 0x419d35ce */
-#define c_erfc_sa2 1.3765776062e+02f  /* 0x4309a863 */
-#define c_erfc_sa3 4.3456588745e+02f  /* 0x43d9486f */
-#define c_erfc_sa4 6.4538726807e+02f  /* 0x442158c9 */
-#define c_erfc_sa5 4.2900814819e+02f  /* 0x43d6810b */
-#define c_erfc_sa6 1.0863500214e+02f  /* 0x42d9451f */
-#define c_erfc_sa7 6.5702495575e+00f  /* 0x40d23f7c */
-#define c_erfc_sa8 -6.0424413532e-02f /* 0xbd777f97 */
-
-// Coefficients for approximation to  erfc in [1/.3528]
-
-#define c_erfc_rb0 -9.8649431020e-03f /* 0xbc21a092 */
-#define c_erfc_rb1 -7.9928326607e-01f /* 0xbf4c9dd4 */
-#define c_erfc_rb2 -1.7757955551e+01f /* 0xc18e104b */
-#define c_erfc_rb3 -1.6063638306e+02f /* 0xc320a2ea */
-#define c_erfc_rb4 -6.3756646729e+02f /* 0xc41f6441 */
-#define c_erfc_rb5 -1.0250950928e+03f /* 0xc480230b */
-#define c_erfc_rb6 -4.8351919556e+02f /* 0xc3f1c275 */
-#define c_erfc_sb1 3.0338060379e+01f  /* 0x41f2b459 */
-#define c_erfc_sb2 3.2579251099e+02f  /* 0x43a2e571 */
-#define c_erfc_sb3 1.5367296143e+03f  /* 0x44c01759 */
-#define c_erfc_sb4 3.1998581543e+03f  /* 0x4547fdbb */
-#define c_erfc_sb5 2.5530502930e+03f  /* 0x451f90ce */
-#define c_erfc_sb6 4.7452853394e+02f  /* 0x43ed43a7 */
-#define c_erfc_sb7 -2.2440952301e+01f /* 0xc1b38712 */
-
-#define _RVV_FLOAT32_FMA_HELPER(LMUL)                                                                     \
-    static inline vfloat32m##LMUL##_t vfmadd_vff_f32m##LMUL(vfloat32m##LMUL##_t a, float b,               \
-                                                            float c, size_t vl)                           \
-    {                                                                                                     \
-        vfloat32m##LMUL##_t ret = vfmul_vf_f32m##LMUL(a, b, vl);                                          \
-        ret = vfadd_vf_f32m##LMUL(ret, c, vl);                                                            \
-        return ret;                                                                                       \
-    }                                                                                                     \
-                                                                                                          \
-    static inline vfloat32m##LMUL##_t vfmadd_vvf_f32m##LMUL(vfloat32m##LMUL##_t a, vfloat32m##LMUL##_t b, \
-                                                            float c, size_t vl)                           \
-    {                                                                                                     \
-        vfloat32m##LMUL##_t ret = vfmul_vv_f32m##LMUL(a, b, vl);                                          \
-        ret = vfadd_vf_f32m##LMUL(ret, c, vl);                                                            \
-        return ret;                                                                                       \
-    }
-
-_RVV_FLOAT32_FMA_HELPER(8)
-_RVV_FLOAT32_FMA_HELPER(4)
-_RVV_FLOAT32_FMA_HELPER(2)
-_RVV_FLOAT32_FMA_HELPER(1)
-
-#define _RVV_FLOAT32_ERFC_OP(LMUL, MLEN)                                                                                                                                                                                                                                                                                                           \
-    static inline vfloat32m##LMUL##_t erfc_ps(vfloat32m##LMUL##_t x, size_t vl)                                                                                                                                                                                                                                                                    \
-    {                                                                                                                                                                                                                                                                                                                                              \
-        /* Argument for polys */                                                                                                                                                                                                                                                                                                                   \
-        vfloat32m##LMUL##_t absx = vfsgnjx_vv_f32m##LMUL(x, x, vl);                                                                                                                                                                                                                                                                                \
-        vfloat32m##LMUL##_t x2 = vfmul_vv_f32m##LMUL(x, x, vl);                                                                                                                                                                                                                                                                                    \
-        vfloat32m##LMUL##_t t = vfrdiv_vf_f32m##LMUL(x2, 1.0f, vl);                                                                                                                                                                                                                                                                                \
-        vfloat32m##LMUL##_t tt = vfsub_vf_f32m##LMUL(absx, 1.0f, vl);                                                                                                                                                                                                                                                                              \
-        /* absx < 1.25f ? tt:t */                                                                                                                                                                                                                                                                                                                  \
-        t = vmerge_vvm_f32m##LMUL(vmfge_vf_f32m##LMUL##_b##MLEN(absx, 1.25f, vl), tt, t, vl);                                                                                                                                                                                                                                                      \
-        /* absx < 0.84375f ? x2:t*/                                                                                                                                                                                                                                                                                                                \
-        t = vmerge_vvm_f32m##LMUL(vmfge_vf_f32m##LMUL##_b##MLEN(absx, 0.84375f, vl), x2, t, vl);                                                                                                                                                                                                                                                   \
-                                                                                                                                                                                                                                                                                                                                                   \
-        vfloat32m##LMUL##_t u = vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vff_f32m##LMUL(t, c_erfc_rb6, c_erfc_rb5, vl), c_erfc_rb4, vl), c_erfc_rb3, vl), c_erfc_rb2, vl), c_erfc_rb1, vl), c_erfc_rb0, vl);                                            \
-        vfloat32m##LMUL##_t v = vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vff_f32m##LMUL(t, c_erfc_sb7, c_erfc_sb6, vl), c_erfc_sb5, vl), c_erfc_sb4, vl), c_erfc_sb3, vl), c_erfc_sb2, vl), c_erfc_sb1, vl);                                            \
-                                                                                                                                                                                                                                                                                                                                                   \
-        vfloat32m##LMUL##_t tu = vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vff_f32m##LMUL(t, c_erfc_ra7, c_erfc_ra6, vl), c_erfc_ra5, vl), c_erfc_ra4, vl), c_erfc_ra3, vl), c_erfc_ra2, vl), c_erfc_ra1, vl), c_erfc_ra0, vl); \
-        vfloat32m##LMUL##_t tv = vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vff_f32m##LMUL(t, c_erfc_sa8, c_erfc_sa7, vl), c_erfc_sa6, vl), c_erfc_sa5, vl), c_erfc_sa4, vl), c_erfc_sa3, vl), c_erfc_sa2, vl), c_erfc_sa1, vl); \
-        u = vmerge_vvm_f32m##LMUL(vmfge_vf_f32m##LMUL##_b##MLEN(absx, 0x1.6db6dap+1f, vl), tu, u, vl); /* u = absx < 0x1.6db6dap+1f ? tu : u;*/                                                                                                                                                                                                    \
-        v = vmerge_vvm_f32m##LMUL(vmfge_vf_f32m##LMUL##_b##MLEN(absx, 0x1.6db6dap+1f, vl), tv, v, vl); /* v = absx < 0x1.6db6dap+1f ? tv : v;*/                                                                                                                                                                                                    \
-                                                                                                                                                                                                                                                                                                                                                   \
-        tu = vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vff_f32m##LMUL(t, c_erfc_pa6, c_erfc_pa5, vl), c_erfc_pa4, vl), c_erfc_pa3, vl), c_erfc_pa2, vl), c_erfc_pa1, vl), c_erfc_pa0, vl);                                                               \
-        tv = vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vff_f32m##LMUL(t, c_erfc_qa6, c_erfc_qa5, vl), c_erfc_qa4, vl), c_erfc_qa3, vl), c_erfc_qa2, vl), c_erfc_qa1, vl);                                                                                                         \
-        u = vmerge_vvm_f32m##LMUL(vmfge_vf_f32m##LMUL##_b##MLEN(absx, 1.25f, vl), tu, u,                                                                                                                                                                                                                                                           \
-                                  vl); /* absx < 1.25f ? tu : u */                                                                                                                                                                                                                                                                                 \
-        v = vmerge_vvm_f32m##LMUL(vmfge_vf_f32m##LMUL##_b##MLEN(absx, 1.25f, vl), tv, v,                                                                                                                                                                                                                                                           \
-                                  vl); /* absx < 1.25f ? tv : v  */                                                                                                                                                                                                                                                                                \
-                                                                                                                                                                                                                                                                                                                                                   \
-        tu = vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vff_f32m##LMUL(t, c_erfc_pp4, c_erfc_pp3, vl), c_erfc_pp2, vl), c_erfc_pp1, vl), c_erfc_pp0, vl);                                                                                                                                                   \
-        tv = vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vvf_f32m##LMUL(t, vfmadd_vff_f32m##LMUL(t, c_erfc_qq5, c_erfc_qq4, vl), c_erfc_qq3, vl), c_erfc_qq2, vl), c_erfc_qq1, vl);                                                                                                                                                   \
-        u = vmerge_vvm_f32m##LMUL(vmfge_vf_f32m##LMUL##_b##MLEN(absx, 0.84375f, vl), tu, u, vl); /* absx < 0.84375f ? tu : u */                                                                                                                                                                                                                    \
-        v = vmerge_vvm_f32m##LMUL(vmfge_vf_f32m##LMUL##_b##MLEN(absx, 0.84375f, vl), tv, v, vl); /* absx <  0.84375f ? tv : v */                                                                                                                                                                                                                   \
-                                                                                                                                                                                                                                                                                                                                                   \
-        v = vfmadd_vvf_f32m##LMUL(t, v, 1.f, vl);                                                                                                                                                                                                                                                                                                  \
-                                                                                                                                                                                                                                                                                                                                                   \
-        vfloat32m##LMUL##_t q = vfdiv_vv_f32m##LMUL(u, v, vl);                                                                                                                                                                                                                                                                                     \
-        vfloat32m##LMUL##_t ret = vfmv_v_f_f32m##LMUL(0.f, vl);                                                                                                                                                                                                                                                                                    \
-                                                                                                                                                                                                                                                                                                                                                   \
-    vfloat32m##LMUL##_t z = vreinterpret_v_u32m##LMUL##_f32m##LMUL( vand_vx_u32m##LMUL(vreinterpret_v_f32m##LMUL##_u32m##LMUL(absx), 0xffff'f000, vl));     \
-       \
-    vfloat32m##LMUL##_t r = vfmul_vv_f32m##LMUL( exp_ps(vfmadd_vvf_f32m##LMUL(vfneg_v_f32m##LMUL(z, vl), z, -0.5625f, vl), vl), exp_ps(vfmadd_vv_f32m##LMUL(vfsub_vv_f32m##LMUL(z, absx, vl), vfadd_vv_f32m##LMUL(z, absx, vl), q, vl), vl), vl);     \
-    r = vfdiv_vv_f32m##LMUL(r, absx, vl);     \
-    t = vfrsub_vf_f32m##LMUL(r, 2.f, vl);     \
-    r = vmerge_vvm_f32m##LMUL(vmfge_vf_f32m##LMUL##_b##MLEN(x, 0.f, vl), t, r, vl); /* x < 0.f ? t:r  */   \
-    ret = vmerge_vvm_f32m##LMUL(vmfge_vf_f32m##LMUL##_b##MLEN(absx, 28.f, vl), r, ret, vl); /*  abs < 28.f ? r : ret  */   \
-       \
-    r = vfrsub_vf_f32m##LMUL(q, 1.f - c_erfc_erx_f, vl);     \
-    t = vfadd_vf_f32m##LMUL(q, 1.f + c_erfc_erx_f, vl);     \
-    r = vmerge_vvm_f32m##LMUL(vmfge_vf_f32m##LMUL##_b##MLEN(x, 0.f, vl), t, r, vl); /*  x < 0.f ? t:r*/     \
-    ret = vmerge_vvm_f32m##LMUL(vmfge_vf_f32m##LMUL##_b##MLEN(absx, 1.25f, vl), r, ret, vl); /*  absx < 1.25f ? r : ret*/     \
-       \
-    r = vfrsub_vf_f32m##LMUL(vfmadd_vv_f32m##LMUL(x, q, vfsub_vf_f32m##LMUL(x, 0.5f, vl), vl), .5, vl);     \
-    ret = vmerge_vvm_f32m##LMUL(vmfge_vf_f32m##LMUL##_b##MLEN(absx, 0.84375f, vl), r, ret, vl); /*  absx < 0.84375f ? r : ret*/     \
-       \
-    ret = vfmerge_vfm_f32m##LMUL(vmflt_vf_f32m##LMUL##_b##MLEN(x, -6.0f, vl), ret, 2.f, vl); /*  x< -6.0f ? 2.0f: ret*/     \
-       \
-    ret = vmerge_vvm_f32m##LMUL(vmfeq_vv_f32m##LMUL##_b##MLEN(x, x, vl), x, ret, vl); /*  erfc(NaN) = NaN*/     \
-       \
-    return ret;     \
-}
-
-_RVV_FLOAT32_ERFC_OP(1, 32)
-_RVV_FLOAT32_ERFC_OP(2, 16)
-_RVV_FLOAT32_ERFC_OP(4, 8)
-_RVV_FLOAT32_ERFC_OP(8, 4)
-
-//TODO rvv optimize
-#define _RVV_FLOAT32_ATAN2_OP(LMUL, MLEN)                                                               \
-    static inline vfloat32m##LMUL##_t atan2_ps(vfloat32m##LMUL##_t a, vfloat32m##LMUL##_t b, size_t vl) \
-    {                                                                                                   \
-        std::vector<float> tmpx(vl);                                                                    \
-        std::vector<float> tmpy(vl);                                                                    \
-        vse32_v_f32m##LMUL(tmpx.data(), a, vl);                                                         \
-        vse32_v_f32m##LMUL(tmpy.data(), b, vl);                                                         \
-        for (size_t i = 0; i < vl; i++)                                                                 \
-        {                                                                                               \
-            tmpx[i] = atan2(tmpx[i], tmpy[i]);                                                          \
-        }                                                                                               \
-        return vle32_v_f32m##LMUL(tmpx.data(), vl);                                                     \
-    }
-
-_RVV_FLOAT32_ATAN2_OP(1, 32)
-_RVV_FLOAT32_ATAN2_OP(2, 16)
-_RVV_FLOAT32_ATAN2_OP(4, 8)
-_RVV_FLOAT32_ATAN2_OP(8, 4)
-
-#endif // RVV_MATHFUN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/rvv_mathfun_fp16s.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/rvv_mathfun_fp16s.h
deleted file mode 100644
index ee5ffe4a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/rvv_mathfun_fp16s.h
+++ /dev/null
@@ -1,419 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef RVV_MATHFUN_FP16S_H
-#define RVV_MATHFUN_FP16S_H
-
-#include <riscv_vector.h>
-
-#define c_inv_mant_mask_f16 -31745 // ~0x7c00u
-#define c_cephes_SQRTHF     0.707106781186547524
-#define c_cephes_log_p0     7.0376836292E-2
-#define c_cephes_log_p1     -1.1514610310E-1
-#define c_cephes_log_p2     1.1676998740E-1
-#define c_cephes_log_p3     -1.2420140846E-1
-#define c_cephes_log_p4     +1.4249322787E-1
-#define c_cephes_log_p5     -1.6668057665E-1
-#define c_cephes_log_p6     +2.0000714765E-1
-#define c_cephes_log_p7     -2.4999993993E-1
-#define c_cephes_log_p8     +3.3333331174E-1
-#define c_cephes_log_q1     -2.12194440e-4
-#define c_cephes_log_q2     0.693359375
-
-#define _RVV_FLOAT16_LOG_OP(LMUL, MLEN)                                                                          \
-    static inline vfloat16m##LMUL##_t log_ps(vfloat16m##LMUL##_t x, size_t vl)                                   \
-    {                                                                                                            \
-        x = vfmax_vf_f16m##LMUL(x, 0.f, vl); /* force flush to zero on denormal values */                        \
-        vbool##MLEN##_t invalid_mask = vmfle_vf_f16m##LMUL##_b##MLEN(x, 0.f, vl);                                \
-                                                                                                                 \
-        vint16m##LMUL##_t ux = vreinterpret_v_f16m##LMUL##_i16m##LMUL(x);                                        \
-                                                                                                                 \
-        vint16m##LMUL##_t emm0 = vsra_vx_i16m##LMUL(ux, 10, vl);                                                 \
-                                                                                                                 \
-        /* keep only the fractional part */                                                                      \
-        ux = vand_vx_i16m##LMUL(ux, c_inv_mant_mask_f16, vl);                                                    \
-        ux = vor_vx_i16m##LMUL(ux, 14336 /* reinterpret_cast<short>((__fp16)0.5) */, vl);                        \
-        x = vreinterpret_v_i16m##LMUL##_f16m##LMUL(ux);                                                          \
-                                                                                                                 \
-        emm0 = vsub_vx_i16m##LMUL(emm0, 0xf, vl);                                                                \
-        vfloat16m##LMUL##_t e = vfcvt_f_x_v_f16m##LMUL(emm0, vl);                                                \
-                                                                                                                 \
-        e = vfadd_vf_f16m##LMUL(e, 1.f, vl);                                                                     \
-                                                                                                                 \
-        /* part2:                      */                                                                        \
-        /*     if( x < SQRTHF ) {      */                                                                        \
-        /*       e -= 1;               */                                                                        \
-        /*       x = x + x - 1.0;      */                                                                        \
-        /*     } else { x = x - 1.0; } */                                                                        \
-        vbool##MLEN##_t mask = vmflt_vf_f16m##LMUL##_b##MLEN(x, c_cephes_SQRTHF, vl);                            \
-        x = vfadd_vv_f16m##LMUL##_m(mask, x, x, x, vl);                                                          \
-        x = vfsub_vf_f16m##LMUL(x, 1.f, vl);                                                                     \
-        e = vfsub_vf_f16m##LMUL##_m(mask, e, e, 1.f, vl);                                                        \
-                                                                                                                 \
-        vfloat16m##LMUL##_t z = vfmul_vv_f16m##LMUL(x, x, vl);                                                   \
-                                                                                                                 \
-        vfloat16m##LMUL##_t y = vfmul_vf_f16m##LMUL(x, c_cephes_log_p0, vl);                                     \
-        y = vfadd_vf_f16m##LMUL(y, c_cephes_log_p1, vl);                                                         \
-        y = vfmul_vv_f16m##LMUL(y, x, vl);                                                                       \
-        y = vfadd_vf_f16m##LMUL(y, c_cephes_log_p2, vl);                                                         \
-        y = vfmul_vv_f16m##LMUL(y, x, vl);                                                                       \
-        y = vfadd_vf_f16m##LMUL(y, c_cephes_log_p3, vl);                                                         \
-        y = vfmul_vv_f16m##LMUL(y, x, vl);                                                                       \
-        y = vfadd_vf_f16m##LMUL(y, c_cephes_log_p4, vl);                                                         \
-        y = vfmul_vv_f16m##LMUL(y, x, vl);                                                                       \
-        y = vfadd_vf_f16m##LMUL(y, c_cephes_log_p5, vl);                                                         \
-        y = vfmul_vv_f16m##LMUL(y, x, vl);                                                                       \
-        y = vfadd_vf_f16m##LMUL(y, c_cephes_log_p6, vl);                                                         \
-        y = vfmul_vv_f16m##LMUL(y, x, vl);                                                                       \
-        y = vfadd_vf_f16m##LMUL(y, c_cephes_log_p7, vl);                                                         \
-        y = vfmul_vv_f16m##LMUL(y, x, vl);                                                                       \
-        y = vfadd_vf_f16m##LMUL(y, c_cephes_log_p8, vl);                                                         \
-        y = vfmul_vv_f16m##LMUL(y, x, vl);                                                                       \
-                                                                                                                 \
-        y = vfmul_vv_f16m##LMUL(y, z, vl);                                                                       \
-                                                                                                                 \
-        vfloat16m##LMUL##_t tmp = vfmul_vf_f16m##LMUL(e, c_cephes_log_q1, vl);                                   \
-        y = vfadd_vv_f16m##LMUL(y, tmp, vl);                                                                     \
-                                                                                                                 \
-        tmp = vfmul_vf_f16m##LMUL(z, 0.5f, vl);                                                                  \
-        y = vfsub_vv_f16m##LMUL(y, tmp, vl);                                                                     \
-                                                                                                                 \
-        tmp = vfmul_vf_f16m##LMUL(e, c_cephes_log_q2, vl);                                                       \
-        x = vfadd_vv_f16m##LMUL(x, y, vl);                                                                       \
-        x = vfadd_vv_f16m##LMUL(x, tmp, vl);                                                                     \
-        /* negative arg will be NAN */                                                                           \
-        vuint16m##LMUL##_t xtmp = vreinterpret_v_f16m##LMUL##_u16m##LMUL(x);                                     \
-        x = vreinterpret_v_u16m##LMUL##_f16m##LMUL(vor_vx_u16m##LMUL##_m(invalid_mask, xtmp, xtmp, 0xffff, vl)); \
-        return x;                                                                                                \
-    }
-
-_RVV_FLOAT16_LOG_OP(1, 16)
-_RVV_FLOAT16_LOG_OP(2, 8)
-_RVV_FLOAT16_LOG_OP(4, 4)
-_RVV_FLOAT16_LOG_OP(8, 2)
-
-#define c_exp_hi_f16 10.7421875f
-#define c_exp_lo_f16 -10.7421875f
-
-#define c_cephes_LOG2EF 1.44269504088896341
-#define c_cephes_exp_C1 0.693359375
-#define c_cephes_exp_C2 -2.12194440e-4
-
-#define c_cephes_exp_p0 1.9875691500E-4
-#define c_cephes_exp_p1 1.3981999507E-3
-#define c_cephes_exp_p2 8.3334519073E-3
-#define c_cephes_exp_p3 4.1665795894E-2
-#define c_cephes_exp_p4 1.6666665459E-1
-#define c_cephes_exp_p5 5.0000001201E-1
-
-#define _RVV_FLOAT16_EXP_OP(LMUL, MLEN)                                                   \
-    static inline vfloat16m##LMUL##_t exp_ps(vfloat16m##LMUL##_t x, size_t vl)            \
-    {                                                                                     \
-        vfloat16m##LMUL##_t tmp, fx;                                                      \
-                                                                                          \
-        x = vfmin_vf_f16m##LMUL(x, c_exp_hi_f16, vl);                                     \
-        x = vfmax_vf_f16m##LMUL(x, c_exp_lo_f16, vl);                                     \
-                                                                                          \
-        /* express exp(x) as exp(g + n*log(2)) */                                         \
-        fx = vfmacc_vf_f16m##LMUL(vfmv_v_f_f16m##LMUL(0.5f, vl), c_cephes_LOG2EF, x, vl); \
-                                                                                          \
-        /* perform a floorf */                                                            \
-        tmp = vfcvt_f_x_v_f16m##LMUL(vfcvt_x_f_v_i16m##LMUL(fx, vl), vl);                 \
-                                                                                          \
-        /* if greater, substract 1 */                                                     \
-        vbool##MLEN##_t mask = vmfgt_vv_f16m##LMUL##_b##MLEN(tmp, fx, vl);                \
-        fx = vfsub_vf_f16m##LMUL##_m(mask, tmp, tmp, 1.f, vl);                            \
-                                                                                          \
-        tmp = vfmul_vf_f16m##LMUL(fx, c_cephes_exp_C1, vl);                               \
-        vfloat16m##LMUL##_t z = vfmul_vf_f16m##LMUL(fx, c_cephes_exp_C2, vl);             \
-        x = vfsub_vv_f16m##LMUL(x, tmp, vl);                                              \
-        x = vfsub_vv_f16m##LMUL(x, z, vl);                                                \
-                                                                                          \
-        vfloat16m##LMUL##_t y = vfmul_vf_f16m##LMUL(x, c_cephes_exp_p0, vl);              \
-        z = vfmul_vv_f16m##LMUL(x, x, vl);                                                \
-                                                                                          \
-        y = vfadd_vf_f16m##LMUL(y, c_cephes_exp_p1, vl);                                  \
-        y = vfmul_vv_f16m##LMUL(y, x, vl);                                                \
-        y = vfadd_vf_f16m##LMUL(y, c_cephes_exp_p2, vl);                                  \
-        y = vfmul_vv_f16m##LMUL(y, x, vl);                                                \
-        y = vfadd_vf_f16m##LMUL(y, c_cephes_exp_p3, vl);                                  \
-        y = vfmul_vv_f16m##LMUL(y, x, vl);                                                \
-        y = vfadd_vf_f16m##LMUL(y, c_cephes_exp_p4, vl);                                  \
-        y = vfmul_vv_f16m##LMUL(y, x, vl);                                                \
-        y = vfadd_vf_f16m##LMUL(y, c_cephes_exp_p5, vl);                                  \
-                                                                                          \
-        y = vfmul_vv_f16m##LMUL(y, z, vl);                                                \
-        y = vfadd_vv_f16m##LMUL(y, x, vl);                                                \
-        y = vfadd_vf_f16m##LMUL(y, 1.f, vl);                                              \
-                                                                                          \
-        /* build 2^n */                                                                   \
-        vint16m##LMUL##_t mm = vfcvt_x_f_v_i16m##LMUL(fx, vl);                            \
-        mm = vadd_vx_i16m##LMUL(mm, 0xf, vl);                                             \
-        mm = vsll_vx_i16m##LMUL(mm, 10, vl);                                              \
-        vfloat16m##LMUL##_t pow2n = vreinterpret_v_i16m##LMUL##_f16m##LMUL(mm);           \
-                                                                                          \
-        y = vfmul_vv_f16m##LMUL(y, pow2n, vl);                                            \
-        return y;                                                                         \
-    }
-
-_RVV_FLOAT16_EXP_OP(1, 16)
-_RVV_FLOAT16_EXP_OP(2, 8)
-_RVV_FLOAT16_EXP_OP(4, 4)
-_RVV_FLOAT16_EXP_OP(8, 2)
-
-#define c_minus_cephes_DP1 -0.78515625
-#define c_minus_cephes_DP2 -2.4187564849853515625e-4
-#define c_minus_cephes_DP3 -3.77489497744594108e-8
-#define c_sincof_p0        -1.9515295891E-4
-#define c_sincof_p1        8.3321608736E-3
-#define c_sincof_p2        -1.6666654611E-1
-#define c_coscof_p0        2.443315711809948E-005
-#define c_coscof_p1        -1.388731625493765E-003
-#define c_coscof_p2        4.166664568298827E-002
-#define c_cephes_FOPI      1.27323954473516 // 4 / M_PI
-
-#define _RVV_FLOAT16_SINCOS_OP(LMUL, MLEN)                                                                                          \
-    static inline void sincos_ps(vfloat16m##LMUL##_t x, vfloat16m##LMUL##_t* ysin, vfloat16m##LMUL##_t* ycos, size_t vl)            \
-    {                                                                                                                               \
-        /* any x */                                                                                                                 \
-        vfloat16m##LMUL##_t xmm1, xmm2, xmm3, y;                                                                                    \
-                                                                                                                                    \
-        vuint16m##LMUL##_t emm2;                                                                                                    \
-                                                                                                                                    \
-        vbool##MLEN##_t sign_mask_sin, sign_mask_cos;                                                                               \
-        sign_mask_sin = vmflt_vf_f16m##LMUL##_b##MLEN(x, 0.f, vl);                                                                  \
-        x = vfsgnj_vf_f16m##LMUL(x, 1.f, vl);                                                                                       \
-                                                                                                                                    \
-        /* scale by 4/Pi */                                                                                                         \
-        y = vfmul_vf_f16m##LMUL(x, c_cephes_FOPI, vl);                                                                              \
-                                                                                                                                    \
-        /* store the integer part of y in mm0 */                                                                                    \
-        emm2 = vfcvt_xu_f_v_u16m##LMUL(y, vl);                                                                                      \
-        /* j=(j+1) & (~1) (see the cephes sources) */                                                                               \
-        emm2 = vadd_vx_u16m##LMUL(emm2, 1, vl);                                                                                     \
-        emm2 = vand_vx_u16m##LMUL(emm2, ~1, vl);                                                                                    \
-        y = vfcvt_f_xu_v_f16m##LMUL(emm2, vl);                                                                                      \
-                                                                                                                                    \
-        /* get the polynom selection mask              */                                                                           \
-        /*     there is one polynom for 0 <= x <= Pi/4 */                                                                           \
-        /*     and another one for Pi/4<x<=Pi/2        */                                                                           \
-        /*                                             */                                                                           \
-        /*     Both branches will be computed.         */                                                                           \
-        vbool##MLEN##_t poly_mask = vmsne_vx_u16m##LMUL##_b##MLEN(vand_vx_u16m##LMUL(emm2, 2, vl), 0, vl);                          \
-                                                                                                                                    \
-        /* The magic pass: "Extended precision modular arithmetic" */                                                               \
-        /*     x = ((x - y * DP1) - y * DP2) - y * DP3;            */                                                               \
-        xmm1 = vfmul_vf_f16m##LMUL(y, c_minus_cephes_DP1, vl);                                                                      \
-        xmm2 = vfmul_vf_f16m##LMUL(y, c_minus_cephes_DP2, vl);                                                                      \
-        xmm3 = vfmul_vf_f16m##LMUL(y, c_minus_cephes_DP3, vl);                                                                      \
-        x = vfadd_vv_f16m##LMUL(x, xmm1, vl);                                                                                       \
-        x = vfadd_vv_f16m##LMUL(x, xmm2, vl);                                                                                       \
-        x = vfadd_vv_f16m##LMUL(x, xmm3, vl);                                                                                       \
-                                                                                                                                    \
-        sign_mask_sin = vmxor_mm_b##MLEN(sign_mask_sin, vmsne_vx_u16m##LMUL##_b##MLEN(vand_vx_u16m##LMUL(emm2, 4, vl), 0, vl), vl); \
-        sign_mask_cos = vmsne_vx_u16m##LMUL##_b##MLEN(vand_vx_u16m##LMUL(vsub_vx_u16m##LMUL(emm2, 2, vl), 4, vl), 0, vl);           \
-                                                                                                                                    \
-        /* Evaluate the first polynom  (0 <= x <= Pi/4) in y1, */                                                                   \
-        /*     and the second polynom  (Pi/4 <= x <= 0) in y2  */                                                                   \
-        vfloat16m##LMUL##_t z = vfmul_vv_f16m##LMUL(x, x, vl);                                                                      \
-        vfloat16m##LMUL##_t y1, y2;                                                                                                 \
-                                                                                                                                    \
-        y1 = vfmul_vf_f16m##LMUL(z, c_coscof_p0, vl);                                                                               \
-        y2 = vfmul_vf_f16m##LMUL(z, c_sincof_p0, vl);                                                                               \
-        y1 = vfadd_vf_f16m##LMUL(y1, c_coscof_p1, vl);                                                                              \
-        y2 = vfadd_vf_f16m##LMUL(y2, c_sincof_p1, vl);                                                                              \
-        y1 = vfmul_vv_f16m##LMUL(y1, z, vl);                                                                                        \
-        y2 = vfmul_vv_f16m##LMUL(y2, z, vl);                                                                                        \
-        y1 = vfadd_vf_f16m##LMUL(y1, c_coscof_p2, vl);                                                                              \
-        y2 = vfadd_vf_f16m##LMUL(y2, c_sincof_p2, vl);                                                                              \
-        y1 = vfmul_vv_f16m##LMUL(y1, z, vl);                                                                                        \
-        y2 = vfmul_vv_f16m##LMUL(y2, z, vl);                                                                                        \
-        y1 = vfmul_vv_f16m##LMUL(y1, z, vl);                                                                                        \
-        y2 = vfmul_vv_f16m##LMUL(y2, x, vl);                                                                                        \
-        y1 = vfsub_vv_f16m##LMUL(y1, vfmul_vf_f16m##LMUL(z, 0.5f, vl), vl);                                                         \
-        y2 = vfadd_vv_f16m##LMUL(y2, x, vl);                                                                                        \
-        y1 = vfadd_vf_f16m##LMUL(y1, 1.f, vl);                                                                                      \
-                                                                                                                                    \
-        /* select the correct result from the two polynoms */                                                                       \
-        vfloat16m##LMUL##_t ys = vmerge_vvm_f16m##LMUL(poly_mask, y2, y1, vl);                                                      \
-        vfloat16m##LMUL##_t yc = vmerge_vvm_f16m##LMUL(poly_mask, y1, y2, vl);                                                      \
-        *ysin = vmerge_vvm_f16m##LMUL(sign_mask_sin, ys, vfneg_v_f16m##LMUL(ys, vl), vl);                                           \
-        *ycos = vmerge_vvm_f16m##LMUL(sign_mask_cos, vfneg_v_f16m##LMUL(yc, vl), yc, vl);                                           \
-    }
-
-_RVV_FLOAT16_SINCOS_OP(1, 16)
-_RVV_FLOAT16_SINCOS_OP(2, 8)
-_RVV_FLOAT16_SINCOS_OP(4, 4)
-_RVV_FLOAT16_SINCOS_OP(8, 2)
-
-#define _RVV_FLOAT16_SIN_OP(LMUL, MLEN)                                        \
-    static inline vfloat16m##LMUL##_t sin_ps(vfloat16m##LMUL##_t x, size_t vl) \
-    {                                                                          \
-        vfloat16m##LMUL##_t ysin, ycos;                                        \
-        sincos_ps(x, &ysin, &ycos, vl);                                        \
-        return ysin;                                                           \
-    }
-
-_RVV_FLOAT16_SIN_OP(1, 16)
-_RVV_FLOAT16_SIN_OP(2, 8)
-_RVV_FLOAT16_SIN_OP(4, 4)
-_RVV_FLOAT16_SIN_OP(8, 2)
-
-#define _RVV_FLOAT16_COS_OP(LMUL, MLEN)                                        \
-    static inline vfloat16m##LMUL##_t cos_ps(vfloat16m##LMUL##_t x, size_t vl) \
-    {                                                                          \
-        vfloat16m##LMUL##_t ysin, ycos;                                        \
-        sincos_ps(x, &ysin, &ycos, vl);                                        \
-        return ycos;                                                           \
-    }
-
-_RVV_FLOAT16_COS_OP(1, 16)
-_RVV_FLOAT16_COS_OP(2, 8)
-_RVV_FLOAT16_COS_OP(4, 4)
-_RVV_FLOAT16_COS_OP(8, 2)
-
-#define c_tanh_tiny 1e-4f
-#define c_tanh_hi   9.0f
-// The monomial coefficients of the numerator polynomial (odd).
-#define c_tanh_alpha_1  4.89352455891786e-3f
-#define c_tanh_alpha_3  6.37261928875436e-4f
-#define c_tanh_alpha_5  1.48572235717979e-5f
-#define c_tanh_alpha_7  5.12229709037114e-8f
-#define c_tanh_alpha_9  -8.60467152213735e-11f
-#define c_tanh_alpha_11 2.00018790482477e-13f
-#define c_tanh_alpha_13 -2.76076847742355e-16f
-// The monomial coefficients of the denominator polynomial (even).
-#define c_tanh_beta_0 4.89352518554385e-3f
-#define c_tanh_beta_2 2.26843463243900e-3f
-#define c_tanh_beta_4 1.18534705686654e-4f
-#define c_tanh_beta_6 1.19825839466702e-6f
-
-#define _RVV_FLOAT16_TANH_OP(LMUL, MLEN)                                                         \
-    static inline vfloat16m##LMUL##_t tanh_ps(vfloat16m##LMUL##_t x, size_t vl)                  \
-    {                                                                                            \
-        vfloat16m##LMUL##_t x2 = vfsgnj_vf_f16m##LMUL(x, 1.f, vl);                               \
-                                                                                                 \
-        vbool##MLEN##_t tiny_mask = vmfge_vf_f16m##LMUL##_b##MLEN(x2, c_tanh_tiny, vl);          \
-                                                                                                 \
-        /* clamp the inputs to the range [-9, 9] since anything outside */                       \
-        /* this range is -/+1.0f in single-precision.                   */                       \
-        x2 = vfmin_vf_f16m##LMUL(x, c_tanh_hi, vl);                                              \
-                                                                                                 \
-        /* since the polynomials are odd/even, we need x**2. */                                  \
-        vfloat16m##LMUL##_t z = vfmul_vv_f16m##LMUL(x2, x2, vl);                                 \
-                                                                                                 \
-        /* evaluate the numerator polynomial y. */                                               \
-        vfloat16m##LMUL##_t y = vfmul_vf_f16m##LMUL(z, c_tanh_alpha_13, vl);                     \
-        y = vfadd_vf_f16m##LMUL(y, c_tanh_alpha_11, vl);                                         \
-        y = vfmul_vv_f16m##LMUL(y, z, vl);                                                       \
-        y = vfadd_vf_f16m##LMUL(y, c_tanh_alpha_9, vl);                                          \
-        y = vfmul_vv_f16m##LMUL(y, z, vl);                                                       \
-        y = vfadd_vf_f16m##LMUL(y, c_tanh_alpha_7, vl);                                          \
-        y = vfmul_vv_f16m##LMUL(y, z, vl);                                                       \
-        y = vfadd_vf_f16m##LMUL(y, c_tanh_alpha_5, vl);                                          \
-        y = vfmul_vv_f16m##LMUL(y, z, vl);                                                       \
-        y = vfadd_vf_f16m##LMUL(y, c_tanh_alpha_3, vl);                                          \
-        y = vfmul_vv_f16m##LMUL(y, z, vl);                                                       \
-        y = vfadd_vf_f16m##LMUL(y, c_tanh_alpha_1, vl);                                          \
-        y = vfmul_vv_f16m##LMUL(y, x2, vl);                                                      \
-                                                                                                 \
-        /* evaluate the denominator polynomial w. */                                             \
-        vfloat16m##LMUL##_t w = vfmul_vf_f16m##LMUL(z, c_tanh_beta_6, vl);                       \
-        w = vfadd_vf_f16m##LMUL(w, c_tanh_beta_4, vl);                                           \
-        w = vfmul_vv_f16m##LMUL(w, z, vl);                                                       \
-        w = vfadd_vf_f16m##LMUL(w, c_tanh_beta_2, vl);                                           \
-        w = vfmul_vv_f16m##LMUL(w, z, vl);                                                       \
-        w = vfadd_vf_f16m##LMUL(w, c_tanh_beta_0, vl);                                           \
-                                                                                                 \
-        /* divide the numerator by the denominator. */                                           \
-        y = vfdiv_vv_f16m##LMUL(y, w, vl);                                                       \
-                                                                                                 \
-        /* reinstate the sign.  */                                                               \
-        y = vfsgnj_vv_f16m##LMUL(y, x, vl);                                                      \
-                                                                                                 \
-        /* when the argument is very small in magnitude it's more accurate to just return it. */ \
-        y = vmerge_vvm_f16m##LMUL(tiny_mask, x, y, vl);                                          \
-                                                                                                 \
-        return y;                                                                                \
-    }
-
-_RVV_FLOAT16_TANH_OP(1, 16)
-_RVV_FLOAT16_TANH_OP(2, 8)
-_RVV_FLOAT16_TANH_OP(4, 4)
-_RVV_FLOAT16_TANH_OP(8, 2)
-
-#define _RVV_FLOAT16_POW_OP(LMUL, MLEN)                                                               \
-    static inline vfloat16m##LMUL##_t pow_ps(vfloat16m##LMUL##_t a, vfloat16m##LMUL##_t b, size_t vl) \
-    {                                                                                                 \
-        /* pow(x, m) = exp(m * log(x)) */                                                             \
-        return exp_ps(vfmul_vv_f16m##LMUL(b, log_ps(a, vl), vl), vl);                                 \
-    }
-
-_RVV_FLOAT16_POW_OP(1, 16)
-_RVV_FLOAT16_POW_OP(2, 8)
-_RVV_FLOAT16_POW_OP(4, 4)
-_RVV_FLOAT16_POW_OP(8, 2)
-
-#if C906
-#define _RVV_FLOAT16_SIGMOID_OP(LMUL, MLEN)                                                                                                \
-    static inline vfloat16m##LMUL##_t sigmoid_ps(vfloat16m##LMUL##_t _v, size_t vl)                                                        \
-    {                                                                                                                                      \
-        _v = vfneg_v_f16m##LMUL(_v, vl);                                                                                                   \
-        _v = exp_ps(_v, vl);                                                                                                               \
-        _v = vfadd_vf_f16m##LMUL(_v, 1.f, vl);                                                                                             \
-        vfloat16m##LMUL##_t _reciprocal = vfrdiv_vf_f16m##LMUL(_v, 1.f, vl);                                                               \
-        _reciprocal = vfmul_vv_f16m##LMUL(vfrsub_vf_f16m##LMUL(vfmul_vv_f16m##LMUL(_v, _reciprocal, vl), 2.f, vl), _reciprocal, vl);       \
-        /* _reciprocal = vfmul_vv_f16m##LMUL(vfrsub_vf_f16m##LMUL(vfmul_vv_f16m##LMUL(_v, _reciprocal, vl), 2.f, vl), _reciprocal, vl); */ \
-        return _reciprocal;                                                                                                                \
-    }
-#else // C906
-#define _RVV_FLOAT16_SIGMOID_OP(LMUL, MLEN)                                                                                                \
-    static inline vfloat16m##LMUL##_t sigmoid_ps(vfloat16m##LMUL##_t _v, size_t vl)                                                        \
-    {                                                                                                                                      \
-        _v = vfneg_v_f16m##LMUL(_v, vl);                                                                                                   \
-        _v = exp_ps(_v, vl);                                                                                                               \
-        _v = vfadd_vf_f16m##LMUL(_v, 1.f, vl);                                                                                             \
-        vfloat16m##LMUL##_t _reciprocal = vfrec7_v_f16m##LMUL(_v, vl);                                                                     \
-        _reciprocal = vfmul_vv_f16m##LMUL(vfrsub_vf_f16m##LMUL(vfmul_vv_f16m##LMUL(_v, _reciprocal, vl), 2.f, vl), _reciprocal, vl);       \
-        /* _reciprocal = vfmul_vv_f16m##LMUL(vfrsub_vf_f16m##LMUL(vfmul_vv_f16m##LMUL(_v, _reciprocal, vl), 2.f, vl), _reciprocal, vl); */ \
-        return _reciprocal;                                                                                                                \
-    }
-#endif // C906
-
-_RVV_FLOAT16_SIGMOID_OP(1, 16)
-_RVV_FLOAT16_SIGMOID_OP(2, 8)
-_RVV_FLOAT16_SIGMOID_OP(4, 4)
-_RVV_FLOAT16_SIGMOID_OP(8, 2)
-
-//TODO rvv optimize
-#define _RVV_FLOAT16_ATAN2_OP(LMUL, MLEN)                                                               \
-    static inline vfloat16m##LMUL##_t atan2_ps(vfloat16m##LMUL##_t a, vfloat16m##LMUL##_t b, size_t vl) \
-    {                                                                                                   \
-        std::vector<__fp16> tmpx(vl);                                                                   \
-        std::vector<__fp16> tmpy(vl);                                                                   \
-        vse16_v_f16m##LMUL(tmpx.data(), a, vl);                                                         \
-        vse16_v_f16m##LMUL(tmpy.data(), b, vl);                                                         \
-        for (size_t i = 0; i < vl; i++)                                                                 \
-        {                                                                                               \
-            tmpx[i] = (__fp16)atan2((float)tmpx[i], (float)tmpy[i]);                                    \
-        }                                                                                               \
-        return vle16_v_f16m##LMUL(tmpx.data(), vl);                                                     \
-    }
-
-_RVV_FLOAT16_ATAN2_OP(1, 32)
-_RVV_FLOAT16_ATAN2_OP(2, 16)
-_RVV_FLOAT16_ATAN2_OP(4, 8)
-_RVV_FLOAT16_ATAN2_OP(8, 4)
-
-#endif // RVV_MATHFUN_FP16S_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/selu_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/selu_riscv.h
deleted file mode 100644
index 185b7f5b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/selu_riscv.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// Xavier Hsinyuan is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 Xavier Hsinyuan <me@lstlx.com>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SELU_RISCV_H
-#define LAYER_SELU_RISCV_H
-
-#include "selu.h"
-
-namespace ncnn {
-
-class SELU_riscv : public SELU
-{
-public:
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SELU_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/sigmoid_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/sigmoid_riscv.h
deleted file mode 100644
index 8f014e6c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/sigmoid_riscv.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SIGMOID_RISCV_H
-#define LAYER_SIGMOID_RISCV_H
-
-#include "sigmoid.h"
-
-namespace ncnn {
-
-class Sigmoid_riscv : public Sigmoid
-{
-public:
-    Sigmoid_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SIGMOID_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/softmax_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/softmax_riscv.h
deleted file mode 100644
index f93dc302..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/softmax_riscv.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Xavier Hsinyuan is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 Xavier Hsinyuan <me@lstlx.com>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SOFTMAX_RISCV_H
-#define LAYER_SOFTMAX_RISCV_H
-
-#include "softmax.h"
-
-namespace ncnn {
-
-class Softmax_riscv : public Softmax
-{
-public:
-    Softmax_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SOFTMAX_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/swish_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/swish_riscv.h
deleted file mode 100644
index 05d5cbe1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/swish_riscv.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SWISH_RISCV_H
-#define LAYER_SWISH_RISCV_H
-
-#include "swish.h"
-
-namespace ncnn {
-
-class Swish_riscv : public Swish
-{
-public:
-    Swish_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SWISH_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/tanh_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/tanh_riscv.h
deleted file mode 100644
index 6fb22ce9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/tanh_riscv.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_TANH_RISCV_H
-#define LAYER_TANH_RISCV_H
-
-#include "tanh.h"
-
-namespace ncnn {
-
-class TanH_riscv : public TanH
-{
-public:
-    TanH_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-    int forward_inplace_fp16sa(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_TANH_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/unaryop_riscv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/unaryop_riscv.h
deleted file mode 100644
index 215ad342..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/riscv/unaryop_riscv.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_UNARYOP_RISCV_H
-#define LAYER_UNARYOP_RISCV_H
-
-#include "unaryop.h"
-
-namespace ncnn {
-
-class UnaryOp_riscv : public UnaryOp
-{
-public:
-    UnaryOp_riscv();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-#if __riscv_vector && __riscv_zfh
-    int forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_UNARYOP_RISCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/rnn.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/rnn.h
deleted file mode 100644
index c52a9208..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/rnn.h
+++ /dev/null
@@ -1,47 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_RNN_H
-#define LAYER_RNN_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class RNN : public Layer
-{
-public:
-    RNN();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    int num_output;
-    int weight_data_size;
-    int direction; // 0=forward 1=reverse 2=bidirectional
-
-    Mat weight_hc_data;
-    Mat weight_xc_data;
-    Mat bias_c_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_RNN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/roialign.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/roialign.h
deleted file mode 100644
index 410ad02f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/roialign.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ROIALIGN_H
-#define LAYER_ROIALIGN_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class ROIAlign : public Layer
-{
-public:
-    ROIAlign();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    int pooled_width;
-    int pooled_height;
-    float spatial_scale;
-    int sampling_ratio;
-    bool aligned;
-    int version;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ROIALIGN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/roipooling.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/roipooling.h
deleted file mode 100644
index 3f73bf31..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/roipooling.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ROIPOOLING_H
-#define LAYER_ROIPOOLING_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class ROIPooling : public Layer
-{
-public:
-    ROIPooling();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    int pooled_width;
-    int pooled_height;
-    float spatial_scale;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ROIPOOLING_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/scale.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/scale.h
deleted file mode 100644
index 3ca87950..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/scale.h
+++ /dev/null
@@ -1,46 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SCALE_H
-#define LAYER_SCALE_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Scale : public Layer
-{
-public:
-    Scale();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int load_model(const ModelBin& mb);
-
-    virtual int forward_inplace(std::vector<Mat>& bottom_top_blobs, const Option& opt) const;
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    // param
-    int scale_data_size;
-    int bias_term;
-
-    // model
-    Mat scale_data;
-    Mat bias_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SCALE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/selu.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/selu.h
deleted file mode 100644
index dc4de694..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/selu.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SELU_H
-#define LAYER_SELU_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class SELU : public Layer
-{
-public:
-    SELU();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    float alpha;
-    float lambda;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SELU_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/shrink.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/shrink.h
deleted file mode 100644
index 26ef6883..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/shrink.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SHRINK_H
-#define LAYER_SHRINK_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Shrink : public Layer
-{
-public:
-    Shrink();
-
-    virtual int load_param(const ParamDict& pd);
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    float bias;
-    float lambd;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SHRINK_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/shufflechannel.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/shufflechannel.h
deleted file mode 100644
index 09631b4e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/shufflechannel.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SHUFFLECHANNEL_H
-#define LAYER_SHUFFLECHANNEL_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class ShuffleChannel : public Layer
-{
-public:
-    ShuffleChannel();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int group;
-    int reverse;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SHUFFLECHANNEL_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/sigmoid.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/sigmoid.h
deleted file mode 100644
index 4403441e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/sigmoid.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SIGMOID_H
-#define LAYER_SIGMOID_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Sigmoid : public Layer
-{
-public:
-    Sigmoid();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SIGMOID_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/slice.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/slice.h
deleted file mode 100644
index c3bcacb9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/slice.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SLICE_H
-#define LAYER_SLICE_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Slice : public Layer
-{
-public:
-    Slice();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    Mat slices;
-    Mat indices;
-    int axis;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SLICE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/softmax.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/softmax.h
deleted file mode 100644
index a40237cb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/softmax.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SOFTMAX_H
-#define LAYER_SOFTMAX_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Softmax : public Layer
-{
-public:
-    Softmax();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    int axis;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SOFTMAX_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/softplus.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/softplus.h
deleted file mode 100644
index 415ad502..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/softplus.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SOFTPLUS_H
-#define LAYER_SOFTPLUS_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Softplus : public Layer
-{
-public:
-    Softplus();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SOFTPLUS_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/split.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/split.h
deleted file mode 100644
index 53686f82..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/split.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SPLIT_H
-#define LAYER_SPLIT_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Split : public Layer
-{
-public:
-    Split();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SPLIT_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/spp.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/spp.h
deleted file mode 100644
index 4161abf4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/spp.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SPP_H
-#define LAYER_SPP_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class SPP : public Layer
-{
-public:
-    SPP();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    enum PoolMethod
-    {
-        PoolMethod_MAX = 0,
-        PoolMethod_AVE = 1
-    };
-
-public:
-    // param
-    int pooling_type;
-    int pyramid_height;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SPP_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/squeeze.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/squeeze.h
deleted file mode 100644
index 536a3b97..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/squeeze.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SQUEEZE_H
-#define LAYER_SQUEEZE_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Squeeze : public Layer
-{
-public:
-    Squeeze();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int squeeze_w;
-    int squeeze_h;
-    int squeeze_d;
-    int squeeze_c;
-    Mat axes;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SQUEEZE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/statisticspooling.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/statisticspooling.h
deleted file mode 100644
index c7000a55..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/statisticspooling.h
+++ /dev/null
@@ -1,36 +0,0 @@
-// Copyright 2016 SoundAI Technology Co., Ltd. (author: Charles Wang)
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_STATISTICSPOOLING_H
-#define LAYER_STATISTICSPOOLING_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class StatisticsPooling : public Layer
-{
-public:
-    StatisticsPooling();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    // param
-    int include_stddev;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_STATISTICSPOOLING_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/swish.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/swish.h
deleted file mode 100644
index 40b29af7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/swish.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SWISH_H
-#define LAYER_SWISH_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Swish : public Layer
-{
-public:
-    Swish();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SWISH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/tanh.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/tanh.h
deleted file mode 100644
index 479cf7e3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/tanh.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_TANH_H
-#define LAYER_TANH_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class TanH : public Layer
-{
-public:
-    TanH();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_TANH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/threshold.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/threshold.h
deleted file mode 100644
index f49f3ec2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/threshold.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_THRESHOLD_H
-#define LAYER_THRESHOLD_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Threshold : public Layer
-{
-public:
-    Threshold();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-    float threshold;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_THRESHOLD_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/tile.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/tile.h
deleted file mode 100644
index c3e65085..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/tile.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_TILE_H
-#define LAYER_TILE_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Tile : public Layer
-{
-public:
-    Tile();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    int axis;
-    int tiles;
-    Mat repeats;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_TILE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/unaryop.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/unaryop.h
deleted file mode 100644
index 3a6926ce..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/unaryop.h
+++ /dev/null
@@ -1,62 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_UNARYOP_H
-#define LAYER_UNARYOP_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class UnaryOp : public Layer
-{
-public:
-    UnaryOp();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-    enum OperationType
-    {
-        Operation_ABS = 0,
-        Operation_NEG = 1,
-        Operation_FLOOR = 2,
-        Operation_CEIL = 3,
-        Operation_SQUARE = 4,
-        Operation_SQRT = 5,
-        Operation_RSQRT = 6,
-        Operation_EXP = 7,
-        Operation_LOG = 8,
-        Operation_SIN = 9,
-        Operation_COS = 10,
-        Operation_TAN = 11,
-        Operation_ASIN = 12,
-        Operation_ACOS = 13,
-        Operation_ATAN = 14,
-        Operation_RECIPROCAL = 15,
-        Operation_TANH = 16,
-        Operation_LOG10 = 17,
-        Operation_ROUND = 18,
-        Operation_TRUNC = 19
-    };
-
-public:
-    // param
-    int op_type;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_UNARYOP_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/unfold.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/unfold.h
deleted file mode 100644
index ff7860b7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/unfold.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_UNFOLD_H
-#define LAYER_UNFOLD_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Unfold : public Layer
-{
-public:
-    Unfold();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    void make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, const Option& opt) const;
-
-public:
-    int kernel_w;
-    int kernel_h;
-    int dilation_w;
-    int dilation_h;
-    int stride_w;
-    int stride_h;
-    int pad_left; // -233=SAME_UPPER -234=SAME_LOWER
-    int pad_right;
-    int pad_top;
-    int pad_bottom;
-    float pad_value;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_UNFOLD_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/absval_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/absval_vulkan.h
deleted file mode 100644
index 9652aac9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/absval_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ABSVAL_VULKAN_H
-#define LAYER_ABSVAL_VULKAN_H
-
-#include "absval.h"
-
-namespace ncnn {
-
-class AbsVal_vulkan : public AbsVal
-{
-public:
-    AbsVal_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using AbsVal::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_absval;
-    Pipeline* pipeline_absval_pack4;
-    Pipeline* pipeline_absval_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ABSVAL_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/batchnorm_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/batchnorm_vulkan.h
deleted file mode 100644
index eedf0491..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/batchnorm_vulkan.h
+++ /dev/null
@@ -1,49 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BATCHNORM_VULKAN_H
-#define LAYER_BATCHNORM_VULKAN_H
-
-#include "batchnorm.h"
-
-namespace ncnn {
-
-class BatchNorm_vulkan : public BatchNorm
-{
-public:
-    BatchNorm_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using BatchNorm::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    VkMat a_data_gpu;
-    VkMat b_data_gpu;
-    VkImageMat a_data_gpu_image;
-    VkImageMat b_data_gpu_image;
-
-    Pipeline* pipeline_batchnorm;
-    Pipeline* pipeline_batchnorm_pack4;
-    Pipeline* pipeline_batchnorm_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BATCHNORM_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/binaryop_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/binaryop_vulkan.h
deleted file mode 100644
index 1c66186a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/binaryop_vulkan.h
+++ /dev/null
@@ -1,55 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BINARYOP_VULKAN_H
-#define LAYER_BINARYOP_VULKAN_H
-
-#include "binaryop.h"
-
-namespace ncnn {
-
-class BinaryOp_vulkan : public BinaryOp
-{
-public:
-    BinaryOp_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using BinaryOp::forward;
-    using BinaryOp::forward_inplace;
-    virtual int forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-    virtual int forward(const std::vector<VkImageMat>& bottom_blobs, std::vector<VkImageMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_binaryop;
-    Pipeline* pipeline_binaryop_pack4;
-    Pipeline* pipeline_binaryop_pack8;
-
-    // broadcast
-    Pipeline* pipeline_binaryop_broadcast[2];
-    Pipeline* pipeline_binaryop_broadcast_pack4[2];
-    Pipeline* pipeline_binaryop_broadcast_pack1to4[2];
-    Pipeline* pipeline_binaryop_broadcast_pack8[2];
-    Pipeline* pipeline_binaryop_broadcast_pack1to8[2];
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BINARYOP_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/cast_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/cast_vulkan.h
deleted file mode 100644
index 47ce3b27..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/cast_vulkan.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CAST_VULKAN_H
-#define LAYER_CAST_VULKAN_H
-
-#include "cast.h"
-
-namespace ncnn {
-
-class Cast_vulkan : public Cast
-{
-public:
-    Cast_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Cast::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_cast_fp32_to_fp16;
-    Pipeline* pipeline_cast_fp32_to_fp16_pack4;
-    Pipeline* pipeline_cast_fp32_to_fp16_pack8;
-    Pipeline* pipeline_cast_fp16_to_fp32;
-    Pipeline* pipeline_cast_fp16_to_fp32_pack4;
-    Pipeline* pipeline_cast_fp16_to_fp32_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CAST_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/celu_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/celu_vulkan.h
deleted file mode 100644
index 2c03a4b9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/celu_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CELU_VULKAN_H
-#define LAYER_CELU_VULKAN_H
-
-#include "celu.h"
-
-namespace ncnn {
-
-class CELU_vulkan : public CELU
-{
-public:
-    CELU_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using CELU::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_celu;
-    Pipeline* pipeline_celu_pack4;
-    Pipeline* pipeline_celu_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CELU_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/clip_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/clip_vulkan.h
deleted file mode 100644
index 79e7745f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/clip_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CLIP_VULKAN_H
-#define LAYER_CLIP_VULKAN_H
-
-#include "clip.h"
-
-namespace ncnn {
-
-class Clip_vulkan : public Clip
-{
-public:
-    Clip_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Clip::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_clip;
-    Pipeline* pipeline_clip_pack4;
-    Pipeline* pipeline_clip_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CLIP_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/concat_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/concat_vulkan.h
deleted file mode 100644
index 109750f3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/concat_vulkan.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONCAT_VULKAN_H
-#define LAYER_CONCAT_VULKAN_H
-
-#include "concat.h"
-
-namespace ncnn {
-
-class Concat_vulkan : public Concat
-{
-public:
-    Concat_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Concat::forward;
-    virtual int forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const std::vector<VkImageMat>& bottom_blobs, std::vector<VkImageMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_concat[2];
-    Pipeline* pipeline_concat_pack4[2];
-    Pipeline* pipeline_concat_pack4to1[2];
-    Pipeline* pipeline_concat_pack8[2];
-    Pipeline* pipeline_concat_pack8to4[2];
-    Pipeline* pipeline_concat_pack8to1[2];
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONCAT_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/convolution1d_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/convolution1d_vulkan.h
deleted file mode 100644
index 28d692ae..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/convolution1d_vulkan.h
+++ /dev/null
@@ -1,55 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTION1D_VULKAN_H
-#define LAYER_CONVOLUTION1D_VULKAN_H
-
-#include "convolution1d.h"
-
-namespace ncnn {
-
-class Convolution1D_vulkan : public Convolution1D
-{
-public:
-    Convolution1D_vulkan();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using Convolution1D::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    ncnn::Layer* padding;
-
-    Mat weight_data_packed;
-    Mat bias_data_packed;
-
-    VkMat weight_data_gpu;
-    VkMat bias_data_gpu;
-
-    VkImageMat weight_data_gpu_image;
-    VkImageMat bias_data_gpu_image;
-
-    Pipeline* pipeline_convolution1d;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTION1D_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/convolution_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/convolution_vulkan.h
deleted file mode 100644
index fa4bdbc5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/convolution_vulkan.h
+++ /dev/null
@@ -1,77 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTION_VULKAN_H
-#define LAYER_CONVOLUTION_VULKAN_H
-
-#include "convolution.h"
-
-namespace ncnn {
-
-class Convolution_vulkan : public Convolution
-{
-public:
-    Convolution_vulkan();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using Convolution::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    ncnn::Layer* padding;
-
-    Mat weight_data_packed;
-    Mat weight_winograd23_data_packed;
-    Mat weight_winograd43_data_packed;
-    Mat bias_data_packed;
-
-    VkMat weight_data_gpu;
-    VkMat bias_data_gpu;
-
-    VkImageMat weight_data_gpu_image;
-    VkImageMat bias_data_gpu_image;
-
-    Pipeline* pipeline_convolution;
-    Pipeline* pipeline_convolution_1x1s1d1;
-
-    Pipeline* pipeline_convolution_gemm;
-
-    // winograd23 and winograd43
-    VkMat weight_data_gpu_tm_winograd23;
-    VkImageMat weight_data_gpu_tm_winograd23_image;
-    Pipeline* pipeline_convolution_3x3s1d1_winograd23_transform_input;
-    Pipeline* pipeline_convolution_3x3s1d1_winograd23_gemm;
-    Pipeline* pipeline_convolution_3x3s1d1_winograd23_transform_output;
-
-    VkMat weight_data_gpu_tm_winograd43;
-    VkImageMat weight_data_gpu_tm_winograd43_image;
-    Pipeline* pipeline_convolution_3x3s1d1_winograd43_transform_input;
-    Pipeline* pipeline_convolution_3x3s1d1_winograd43_gemm;
-    Pipeline* pipeline_convolution_3x3s1d1_winograd43_transform_output;
-
-    // convolution as fc
-    ncnn::Layer* reshape_1x1xw;
-    ncnn::Layer* reshape_w;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTION_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/convolutiondepthwise_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/convolutiondepthwise_vulkan.h
deleted file mode 100644
index 7a6cfe1f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/convolutiondepthwise_vulkan.h
+++ /dev/null
@@ -1,68 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTIONDEPTHWISE_VULKAN_H
-#define LAYER_CONVOLUTIONDEPTHWISE_VULKAN_H
-
-#include "convolutiondepthwise.h"
-
-namespace ncnn {
-
-class ConvolutionDepthWise_vulkan : public ConvolutionDepthWise
-{
-public:
-    ConvolutionDepthWise_vulkan();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using ConvolutionDepthWise::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Mat weight_data_packed;
-    Mat weight_data_packed_groups;
-    Mat bias_data_packed;
-
-    VkMat weight_data_gpu;
-    VkMat bias_data_gpu;
-
-    VkImageMat weight_data_gpu_image;
-    VkImageMat bias_data_gpu_image;
-
-    ncnn::Layer* padding;
-
-    Pipeline* pipeline_convolutiondepthwise;
-    Pipeline* pipeline_convolutiondepthwise_pack4;
-    Pipeline* pipeline_convolutiondepthwise_pack8;
-
-    Pipeline* pipeline_convolutiondepthwise_group;
-    Pipeline* pipeline_convolutiondepthwise_group_pack4;
-    Pipeline* pipeline_convolutiondepthwise_group_pack1to4;
-    Pipeline* pipeline_convolutiondepthwise_group_pack4to1;
-    Pipeline* pipeline_convolutiondepthwise_group_pack8;
-    Pipeline* pipeline_convolutiondepthwise_group_pack1to8;
-    Pipeline* pipeline_convolutiondepthwise_group_pack4to8;
-    Pipeline* pipeline_convolutiondepthwise_group_pack8to4;
-    Pipeline* pipeline_convolutiondepthwise_group_pack8to1;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTIONDEPTHWISE_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/crop_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/crop_vulkan.h
deleted file mode 100644
index 44802688..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/crop_vulkan.h
+++ /dev/null
@@ -1,53 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CROP_VULKAN_H
-#define LAYER_CROP_VULKAN_H
-
-#include "crop.h"
-
-namespace ncnn {
-
-class Crop_vulkan : public Crop
-{
-public:
-    Crop_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Crop::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-    virtual int forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-    virtual int forward(const std::vector<VkImageMat>& bottom_blobs, std::vector<VkImageMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_crop;
-    Pipeline* pipeline_crop_pack4;
-    Pipeline* pipeline_crop_pack1to4;
-    Pipeline* pipeline_crop_pack4to1;
-    Pipeline* pipeline_crop_pack8;
-    Pipeline* pipeline_crop_pack1to8;
-    Pipeline* pipeline_crop_pack4to8;
-    Pipeline* pipeline_crop_pack8to4;
-    Pipeline* pipeline_crop_pack8to1;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CROP_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/deconvolution_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/deconvolution_vulkan.h
deleted file mode 100644
index 6e18c38d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/deconvolution_vulkan.h
+++ /dev/null
@@ -1,59 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTION_VULKAN_H
-#define LAYER_DECONVOLUTION_VULKAN_H
-
-#include "deconvolution.h"
-
-namespace ncnn {
-
-class Deconvolution_vulkan : public Deconvolution
-{
-public:
-    Deconvolution_vulkan();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using Deconvolution::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Mat weight_data_packed;
-    Mat bias_data_packed;
-
-    VkMat weight_data_gpu;
-    VkMat bias_data_gpu;
-
-    VkImageMat weight_data_gpu_image;
-    VkImageMat bias_data_gpu_image;
-
-    ncnn::Layer* crop;
-    ncnn::Layer* output_crop;
-
-    Pipeline* pipeline_deconvolution;
-
-    Pipeline* pipeline_deconvolution_gemm;
-    Pipeline* pipeline_deconvolution_col2im;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTION_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/deconvolutiondepthwise_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/deconvolutiondepthwise_vulkan.h
deleted file mode 100644
index 5346de8e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/deconvolutiondepthwise_vulkan.h
+++ /dev/null
@@ -1,68 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTIONDEPTHWISE_VULKAN_H
-#define LAYER_DECONVOLUTIONDEPTHWISE_VULKAN_H
-
-#include "deconvolutiondepthwise.h"
-
-namespace ncnn {
-
-class DeconvolutionDepthWise_vulkan : public DeconvolutionDepthWise
-{
-public:
-    DeconvolutionDepthWise_vulkan();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using DeconvolutionDepthWise::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Mat weight_data_packed;
-    Mat bias_data_packed;
-
-    VkMat weight_data_gpu;
-    VkMat bias_data_gpu;
-
-    VkImageMat weight_data_gpu_image;
-    VkImageMat bias_data_gpu_image;
-
-    ncnn::Layer* crop;
-    ncnn::Layer* output_crop;
-
-    Pipeline* pipeline_deconvolutiondepthwise;
-    Pipeline* pipeline_deconvolutiondepthwise_pack4;
-    Pipeline* pipeline_deconvolutiondepthwise_pack8;
-
-    Pipeline* pipeline_deconvolutiondepthwise_group;
-    Pipeline* pipeline_deconvolutiondepthwise_group_pack4;
-    Pipeline* pipeline_deconvolutiondepthwise_group_pack1to4;
-    Pipeline* pipeline_deconvolutiondepthwise_group_pack4to1;
-    Pipeline* pipeline_deconvolutiondepthwise_group_pack8;
-    Pipeline* pipeline_deconvolutiondepthwise_group_pack1to8;
-    Pipeline* pipeline_deconvolutiondepthwise_group_pack4to8;
-    Pipeline* pipeline_deconvolutiondepthwise_group_pack8to4;
-    Pipeline* pipeline_deconvolutiondepthwise_group_pack8to1;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTIONDEPTHWISE_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/deepcopy_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/deepcopy_vulkan.h
deleted file mode 100644
index 867ff1af..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/deepcopy_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DEEPCOPY_VULKAN_H
-#define LAYER_DEEPCOPY_VULKAN_H
-
-#include "deepcopy.h"
-
-namespace ncnn {
-
-class DeepCopy_vulkan : public DeepCopy
-{
-public:
-    DeepCopy_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using DeepCopy::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_deepcopy;
-    Pipeline* pipeline_deepcopy_pack4;
-    Pipeline* pipeline_deepcopy_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DEEPCOPY_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/dropout_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/dropout_vulkan.h
deleted file mode 100644
index e45159b7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/dropout_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DROPOUT_VULKAN_H
-#define LAYER_DROPOUT_VULKAN_H
-
-#include "dropout.h"
-
-namespace ncnn {
-
-class Dropout_vulkan : public Dropout
-{
-public:
-    Dropout_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Dropout::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_dropout;
-    Pipeline* pipeline_dropout_pack4;
-    Pipeline* pipeline_dropout_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DROPOUT_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/eltwise_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/eltwise_vulkan.h
deleted file mode 100644
index 09418657..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/eltwise_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ELTWISE_VULKAN_H
-#define LAYER_ELTWISE_VULKAN_H
-
-#include "eltwise.h"
-
-namespace ncnn {
-
-class Eltwise_vulkan : public Eltwise
-{
-public:
-    Eltwise_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Eltwise::forward;
-    virtual int forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const std::vector<VkImageMat>& bottom_blobs, std::vector<VkImageMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_eltwise[2];
-    Pipeline* pipeline_eltwise_pack4[2];
-    Pipeline* pipeline_eltwise_pack8[2];
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ELTWISE_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/elu_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/elu_vulkan.h
deleted file mode 100644
index c616c3be..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/elu_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ELU_VULKAN_H
-#define LAYER_ELU_VULKAN_H
-
-#include "elu.h"
-
-namespace ncnn {
-
-class ELU_vulkan : public ELU
-{
-public:
-    ELU_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using ELU::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_elu;
-    Pipeline* pipeline_elu_pack4;
-    Pipeline* pipeline_elu_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ELU_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/erf_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/erf_vulkan.h
deleted file mode 100644
index 3f2ae5ac..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/erf_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ERF_VULKAN_H
-#define LAYER_ERF_VULKAN_H
-
-#include "erf.h"
-
-namespace ncnn {
-
-class Erf_vulkan : public Erf
-{
-public:
-    Erf_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Erf::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_erf;
-    Pipeline* pipeline_erf_pack4;
-    Pipeline* pipeline_erf_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ERF_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/flatten_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/flatten_vulkan.h
deleted file mode 100644
index 1068ce54..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/flatten_vulkan.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_FLATTEN_VULKAN_H
-#define LAYER_FLATTEN_VULKAN_H
-
-#include "flatten.h"
-
-namespace ncnn {
-
-class Flatten_vulkan : public Flatten
-{
-public:
-    Flatten_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Flatten::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_flatten;
-    Pipeline* pipeline_flatten_pack4;
-    Pipeline* pipeline_flatten_pack1to4;
-    Pipeline* pipeline_flatten_pack8;
-    Pipeline* pipeline_flatten_pack1to8;
-    Pipeline* pipeline_flatten_pack4to8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_FLATTEN_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/gelu_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/gelu_vulkan.h
deleted file mode 100644
index ced6f07a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/gelu_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GELU_VULKAN_H
-#define LAYER_GELU_VULKAN_H
-
-#include "gelu.h"
-
-namespace ncnn {
-
-class GELU_vulkan : public GELU
-{
-public:
-    GELU_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using GELU::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_gelu;
-    Pipeline* pipeline_gelu_pack4;
-    Pipeline* pipeline_gelu_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GELU_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/gemm_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/gemm_vulkan.h
deleted file mode 100644
index d9fa9201..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/gemm_vulkan.h
+++ /dev/null
@@ -1,56 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GEMM_VULKAN_H
-#define LAYER_GEMM_VULKAN_H
-
-#include "gemm.h"
-
-namespace ncnn {
-
-class Gemm_vulkan : public Gemm
-{
-public:
-    Gemm_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using Gemm::forward;
-    virtual int forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const std::vector<VkImageMat>& bottom_blobs, std::vector<VkImageMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Mat A_data_packed;
-    Mat B_data_packed;
-    Mat C_data_packed;
-
-    VkMat A_data_gpu;
-    VkMat B_data_gpu;
-    VkMat C_data_gpu;
-
-    VkImageMat A_data_gpu_image;
-    VkImageMat B_data_gpu_image;
-    VkImageMat C_data_gpu_image;
-
-    Pipeline* pipeline_gemm;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GEMM_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/hardsigmoid_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/hardsigmoid_vulkan.h
deleted file mode 100644
index b0902948..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/hardsigmoid_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_HARDSIGMOID_VULKAN_H
-#define LAYER_HARDSIGMOID_VULKAN_H
-
-#include "hardsigmoid.h"
-
-namespace ncnn {
-
-class HardSigmoid_vulkan : public HardSigmoid
-{
-public:
-    HardSigmoid_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using HardSigmoid::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_hardsigmoid;
-    Pipeline* pipeline_hardsigmoid_pack4;
-    Pipeline* pipeline_hardsigmoid_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_HARDSIGMOID_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/hardswish_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/hardswish_vulkan.h
deleted file mode 100644
index ab472687..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/hardswish_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_HARDSWISH_VULKAN_H
-#define LAYER_HARDSWISH_VULKAN_H
-
-#include "hardswish.h"
-
-namespace ncnn {
-
-class HardSwish_vulkan : public HardSwish
-{
-public:
-    HardSwish_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using HardSwish::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_hardswish;
-    Pipeline* pipeline_hardswish_pack4;
-    Pipeline* pipeline_hardswish_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_HARDSWISH_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/innerproduct_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/innerproduct_vulkan.h
deleted file mode 100644
index 9002c581..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/innerproduct_vulkan.h
+++ /dev/null
@@ -1,58 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INNERPRODUCT_VULKAN_H
-#define LAYER_INNERPRODUCT_VULKAN_H
-
-#include "innerproduct.h"
-
-namespace ncnn {
-
-class InnerProduct_vulkan : public InnerProduct
-{
-public:
-    InnerProduct_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using InnerProduct::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    ncnn::Layer* flatten;
-
-    Mat weight_data_packed;
-    Mat bias_data_packed;
-
-    VkMat weight_data_gpu;
-    VkMat bias_data_gpu;
-
-    VkImageMat weight_data_gpu_image;
-    VkImageMat bias_data_gpu_image;
-
-    Pipeline* pipeline_innerproduct;
-
-    Pipeline* pipeline_innerproduct_sum8;
-    Pipeline* pipeline_innerproduct_reduce_sum8;
-
-    Pipeline* pipeline_innerproduct_gemm;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INNERPRODUCT_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/instancenorm_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/instancenorm_vulkan.h
deleted file mode 100644
index 943fff65..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/instancenorm_vulkan.h
+++ /dev/null
@@ -1,66 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INSTANCENORM_VULKAN_H
-#define LAYER_INSTANCENORM_VULKAN_H
-
-#include "instancenorm.h"
-
-namespace ncnn {
-
-class InstanceNorm_vulkan : public InstanceNorm
-{
-public:
-    InstanceNorm_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using InstanceNorm::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    VkMat gamma_data_gpu;
-    VkMat beta_data_gpu;
-    VkImageMat gamma_data_gpu_image;
-    VkImageMat beta_data_gpu_image;
-
-    Pipeline* pipeline_instancenorm_reduce_sum4_fp16_to_fp32;
-    Pipeline* pipeline_instancenorm_reduce_sum4_fp32[2];
-    Pipeline* pipeline_instancenorm_reduce_mean;
-    Pipeline* pipeline_instancenorm_sub_mean_square;
-    Pipeline* pipeline_instancenorm_coeffs;
-    Pipeline* pipeline_instancenorm_norm;
-
-    Pipeline* pipeline_instancenorm_reduce_sum4_fp16_to_fp32_pack4;
-    Pipeline* pipeline_instancenorm_reduce_sum4_fp32_pack4[2];
-    Pipeline* pipeline_instancenorm_reduce_mean_pack4;
-    Pipeline* pipeline_instancenorm_sub_mean_square_pack4;
-    Pipeline* pipeline_instancenorm_coeffs_pack4;
-    Pipeline* pipeline_instancenorm_norm_pack4;
-
-    Pipeline* pipeline_instancenorm_reduce_sum4_fp16_to_fp32_pack8;
-    Pipeline* pipeline_instancenorm_reduce_sum4_fp32_pack8[2];
-    Pipeline* pipeline_instancenorm_reduce_mean_pack8;
-    Pipeline* pipeline_instancenorm_sub_mean_square_pack8;
-    Pipeline* pipeline_instancenorm_coeffs_pack8;
-    Pipeline* pipeline_instancenorm_norm_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INSTANCENORM_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/interp_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/interp_vulkan.h
deleted file mode 100644
index 5f175234..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/interp_vulkan.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INTERP_VULKAN_H
-#define LAYER_INTERP_VULKAN_H
-
-#include "interp.h"
-
-namespace ncnn {
-
-class Interp_vulkan : public Interp
-{
-public:
-    Interp_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Interp::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const std::vector<VkImageMat>& bottom_blobs, std::vector<VkImageMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_interp;
-    Pipeline* pipeline_interp_pack4;
-    Pipeline* pipeline_interp_pack8;
-
-    Pipeline* pipeline_interp_bicubic_coeffs_x;
-    Pipeline* pipeline_interp_bicubic_coeffs_y;
-    Pipeline* pipeline_interp_bicubic;
-    Pipeline* pipeline_interp_bicubic_pack4;
-    Pipeline* pipeline_interp_bicubic_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INTERP_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/lrn_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/lrn_vulkan.h
deleted file mode 100644
index ad8cc993..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/lrn_vulkan.h
+++ /dev/null
@@ -1,49 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_LRN_VULKAN_H
-#define LAYER_LRN_VULKAN_H
-
-#include "lrn.h"
-
-namespace ncnn {
-
-class LRN_vulkan : public LRN
-{
-public:
-    LRN_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using LRN::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_lrn_square_pad;
-    Pipeline* pipeline_lrn_norm;
-    Pipeline* pipeline_lrn_square_pad_across_channel_pack4;
-    Pipeline* pipeline_lrn_norm_across_channel_pack4;
-    Pipeline* pipeline_lrn_square_pad_within_channel_pack4;
-    Pipeline* pipeline_lrn_norm_within_channel_pack4;
-    Pipeline* pipeline_lrn_square_pad_across_channel_pack8;
-    Pipeline* pipeline_lrn_norm_across_channel_pack8;
-    Pipeline* pipeline_lrn_square_pad_within_channel_pack8;
-    Pipeline* pipeline_lrn_norm_within_channel_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_LRN_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/memorydata_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/memorydata_vulkan.h
deleted file mode 100644
index 32655abd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/memorydata_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MEMORYDATA_VULKAN_H
-#define LAYER_MEMORYDATA_VULKAN_H
-
-#include "memorydata.h"
-
-namespace ncnn {
-
-class MemoryData_vulkan : public MemoryData
-{
-public:
-    MemoryData_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using MemoryData::forward;
-    virtual int forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const std::vector<VkImageMat>& bottom_blobs, std::vector<VkImageMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-
-public:
-    VkMat data_gpu;
-    VkImageMat data_gpu_image;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MEMORYDATA_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/mish_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/mish_vulkan.h
deleted file mode 100644
index 86488438..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/mish_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MISH_VULKAN_H
-#define LAYER_MISH_VULKAN_H
-
-#include "mish.h"
-
-namespace ncnn {
-
-class Mish_vulkan : public Mish
-{
-public:
-    Mish_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Mish::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_mish;
-    Pipeline* pipeline_mish_pack4;
-    Pipeline* pipeline_mish_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MISH_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/multiheadattention_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/multiheadattention_vulkan.h
deleted file mode 100644
index 3b77d96d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/multiheadattention_vulkan.h
+++ /dev/null
@@ -1,57 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MULTIHEADATTENTION_VULKAN_H
-#define LAYER_MULTIHEADATTENTION_VULKAN_H
-
-#include "multiheadattention.h"
-
-namespace ncnn {
-
-class MultiHeadAttention_vulkan : public MultiHeadAttention
-{
-public:
-    MultiHeadAttention_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using MultiHeadAttention::forward;
-    virtual int forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const std::vector<VkImageMat>& bottom_blobs, std::vector<VkImageMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Layer* q_gemm;
-    Layer* k_gemm;
-    Layer* v_gemm;
-    Layer* o_gemm;
-
-    Layer* qk_softmax;
-
-    Pipeline* pipeline_multiheadattention_qk_cross;
-    Pipeline* pipeline_multiheadattention_qk_cross_pack4;
-    Pipeline* pipeline_multiheadattention_qk_cross_pack1to4;
-    Pipeline* pipeline_multiheadattention_qk_cross_pack4to1;
-
-    Pipeline* pipeline_multiheadattention_qkv_cross;
-    Pipeline* pipeline_multiheadattention_qkv_cross_pack4;
-    Pipeline* pipeline_multiheadattention_qkv_cross_pack1to4;
-    Pipeline* pipeline_multiheadattention_qkv_cross_pack4to1;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MULTIHEADATTENTION_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/noop_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/noop_vulkan.h
deleted file mode 100644
index 84d05d07..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/noop_vulkan.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_NOOP_VULKAN_H
-#define LAYER_NOOP_VULKAN_H
-
-#include "noop.h"
-
-namespace ncnn {
-
-class Noop_vulkan : public Noop
-{
-public:
-    Noop_vulkan();
-
-    using Noop::forward;
-    virtual int forward_inplace(std::vector<VkMat>& bottom_top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(std::vector<VkImageMat>& bottom_top_blobs, VkCompute& cmd, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_NOOP_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/normalize_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/normalize_vulkan.h
deleted file mode 100644
index 4ad20cc4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/normalize_vulkan.h
+++ /dev/null
@@ -1,58 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_NORMALIZE_VULKAN_H
-#define LAYER_NORMALIZE_VULKAN_H
-
-#include "normalize.h"
-
-namespace ncnn {
-
-class Normalize_vulkan : public Normalize
-{
-public:
-    Normalize_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using Normalize::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    VkMat scale_data_gpu;
-    VkImageMat scale_data_gpu_image;
-
-    Pipeline* pipeline_normalize_reduce_sum4_fp16_to_fp32;
-    Pipeline* pipeline_normalize_reduce_sum4_fp32[2];
-    Pipeline* pipeline_normalize_coeffs;
-    Pipeline* pipeline_normalize_norm;
-
-    Pipeline* pipeline_normalize_reduce_sum4_fp16_to_fp32_pack4;
-    Pipeline* pipeline_normalize_reduce_sum4_fp32_pack4[2];
-    Pipeline* pipeline_normalize_coeffs_pack4;
-    Pipeline* pipeline_normalize_norm_pack4;
-
-    Pipeline* pipeline_normalize_reduce_sum4_fp16_to_fp32_pack8;
-    Pipeline* pipeline_normalize_reduce_sum4_fp32_pack8[2];
-    Pipeline* pipeline_normalize_coeffs_pack8;
-    Pipeline* pipeline_normalize_norm_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_NORMALIZE_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/packing_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/packing_vulkan.h
deleted file mode 100644
index fb9d1cd1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/packing_vulkan.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PACKING_VULKAN_H
-#define LAYER_PACKING_VULKAN_H
-
-#include "packing.h"
-
-namespace ncnn {
-
-class Packing_vulkan : public Packing
-{
-public:
-    Packing_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Packing::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_packing;
-    Pipeline* pipeline_packing_pack4;
-    Pipeline* pipeline_packing_pack8;
-    Pipeline* pipeline_packing_pack1to4;
-    Pipeline* pipeline_packing_pack4to1;
-    Pipeline* pipeline_packing_pack1to8;
-    Pipeline* pipeline_packing_pack4to8;
-    Pipeline* pipeline_packing_pack8to4;
-    Pipeline* pipeline_packing_pack8to1;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PACKING_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/padding_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/padding_vulkan.h
deleted file mode 100644
index bc6a235e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/padding_vulkan.h
+++ /dev/null
@@ -1,61 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PADDING_VULKAN_H
-#define LAYER_PADDING_VULKAN_H
-
-#include "padding.h"
-
-namespace ncnn {
-
-class Padding_vulkan : public Padding
-{
-public:
-    Padding_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using Padding::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-    virtual int forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-    virtual int forward(const std::vector<VkImageMat>& bottom_blobs, std::vector<VkImageMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-
-public:
-    VkMat per_channel_pad_data_gpu;
-    VkImageMat per_channel_pad_data_gpu_image;
-    Pipeline* pipeline_padding;
-    Pipeline* pipeline_padding_pack4;
-    Pipeline* pipeline_padding_pack1to4;
-    Pipeline* pipeline_padding_pack4to1;
-    Pipeline* pipeline_padding_pack8;
-    Pipeline* pipeline_padding_pack1to8;
-    Pipeline* pipeline_padding_pack4to8;
-    Pipeline* pipeline_padding_pack8to4;
-    Pipeline* pipeline_padding_pack8to1;
-
-    Pipeline* pipeline_padding_3d;
-    Pipeline* pipeline_padding_3d_pack4;
-    Pipeline* pipeline_padding_3d_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PADDING_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/permute_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/permute_vulkan.h
deleted file mode 100644
index fd073bec..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/permute_vulkan.h
+++ /dev/null
@@ -1,48 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PERMUTE_VULKAN_H
-#define LAYER_PERMUTE_VULKAN_H
-
-#include "permute.h"
-
-namespace ncnn {
-
-class Permute_vulkan : public Permute
-{
-public:
-    Permute_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Permute::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_permute;
-    Pipeline* pipeline_permute_pack4;
-    Pipeline* pipeline_permute_pack1to4;
-    Pipeline* pipeline_permute_pack4to1;
-    Pipeline* pipeline_permute_pack8;
-    Pipeline* pipeline_permute_pack1to8;
-    Pipeline* pipeline_permute_pack4to8;
-    Pipeline* pipeline_permute_pack8to4;
-    Pipeline* pipeline_permute_pack8to1;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PERMUTE_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/pixelshuffle_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/pixelshuffle_vulkan.h
deleted file mode 100644
index d0b812f2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/pixelshuffle_vulkan.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PIXELSHUFFLE_VULKAN_H
-#define LAYER_PIXELSHUFFLE_VULKAN_H
-
-#include "pixelshuffle.h"
-
-namespace ncnn {
-
-class PixelShuffle_vulkan : public PixelShuffle
-{
-public:
-    PixelShuffle_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using PixelShuffle::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_pixelshuffle;
-    Pipeline* pipeline_pixelshuffle_pack4;
-    Pipeline* pipeline_pixelshuffle_pack4to1;
-    Pipeline* pipeline_pixelshuffle_pack8;
-    Pipeline* pipeline_pixelshuffle_pack8to1;
-    Pipeline* pipeline_pixelshuffle_pack8to4;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PIXELSHUFFLE_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/pooling_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/pooling_vulkan.h
deleted file mode 100644
index a336908d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/pooling_vulkan.h
+++ /dev/null
@@ -1,60 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_POOLING_VULKAN_H
-#define LAYER_POOLING_VULKAN_H
-
-#include "pooling.h"
-
-namespace ncnn {
-
-class Pooling_vulkan : public Pooling
-{
-public:
-    Pooling_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using Pooling::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    ncnn::Layer* padding;
-
-    Pipeline* pipeline_pooling;
-    Pipeline* pipeline_pooling_pack4;
-    Pipeline* pipeline_pooling_pack8;
-
-    Pipeline* pipeline_pooling_adaptive;
-    Pipeline* pipeline_pooling_adaptive_pack4;
-    Pipeline* pipeline_pooling_adaptive_pack8;
-
-    Pipeline* pipeline_pooling_global_reduce_first;
-    Pipeline* pipeline_pooling_global_reduce_first_pack4;
-    Pipeline* pipeline_pooling_global_reduce_first_pack8;
-    Pipeline* pipeline_pooling_global_reduce;
-    Pipeline* pipeline_pooling_global_reduce_pack4;
-    Pipeline* pipeline_pooling_global_reduce_pack8;
-    Pipeline* pipeline_pooling_global_reduce_last;
-    Pipeline* pipeline_pooling_global_reduce_last_pack4;
-    Pipeline* pipeline_pooling_global_reduce_last_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_POOLING_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/prelu_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/prelu_vulkan.h
deleted file mode 100644
index d2bae5ea..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/prelu_vulkan.h
+++ /dev/null
@@ -1,47 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PRELU_VULKAN_H
-#define LAYER_PRELU_VULKAN_H
-
-#include "prelu.h"
-
-namespace ncnn {
-
-class PReLU_vulkan : public PReLU
-{
-public:
-    PReLU_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using PReLU::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    VkMat slope_data_gpu;
-    VkImageMat slope_data_gpu_image;
-
-    Pipeline* pipeline_prelu;
-    Pipeline* pipeline_prelu_pack4;
-    Pipeline* pipeline_prelu_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PRELU_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/priorbox_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/priorbox_vulkan.h
deleted file mode 100644
index 394b12d0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/priorbox_vulkan.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PRIORBOX_VULKAN_H
-#define LAYER_PRIORBOX_VULKAN_H
-
-#include "priorbox.h"
-
-namespace ncnn {
-
-class PriorBox_vulkan : public PriorBox
-{
-public:
-    PriorBox_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using PriorBox::forward;
-    virtual int forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-
-public:
-    VkMat min_sizes_gpu;
-    VkMat max_sizes_gpu;
-    VkMat aspect_ratios_gpu;
-    Pipeline* pipeline_priorbox;
-    Pipeline* pipeline_priorbox_mxnet;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PRIORBOX_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/relu_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/relu_vulkan.h
deleted file mode 100644
index 287781fd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/relu_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_RELU_VULKAN_H
-#define LAYER_RELU_VULKAN_H
-
-#include "relu.h"
-
-namespace ncnn {
-
-class ReLU_vulkan : public ReLU
-{
-public:
-    ReLU_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using ReLU::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_relu;
-    Pipeline* pipeline_relu_pack4;
-    Pipeline* pipeline_relu_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_RELU_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/reorg_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/reorg_vulkan.h
deleted file mode 100644
index f1565486..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/reorg_vulkan.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_REORG_VULKAN_H
-#define LAYER_REORG_VULKAN_H
-
-#include "reorg.h"
-
-namespace ncnn {
-
-class Reorg_vulkan : public Reorg
-{
-public:
-    Reorg_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Reorg::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_reorg;
-    Pipeline* pipeline_reorg_pack4;
-    Pipeline* pipeline_reorg_pack1to4;
-    Pipeline* pipeline_reorg_pack8;
-    Pipeline* pipeline_reorg_pack1to8;
-    Pipeline* pipeline_reorg_pack4to8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_REORG_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/reshape_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/reshape_vulkan.h
deleted file mode 100644
index 6b408f79..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/reshape_vulkan.h
+++ /dev/null
@@ -1,55 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_RESHAPE_VULKAN_H
-#define LAYER_RESHAPE_VULKAN_H
-
-#include "reshape.h"
-
-namespace ncnn {
-
-class Reshape_vulkan : public Reshape
-{
-public:
-    Reshape_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Reshape::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    ncnn::Layer* permute_wh;
-    ncnn::Layer* permute_hw;
-    ncnn::Layer* permute_hwc;
-    ncnn::Layer* permute_chw;
-    ncnn::Layer* permute_dhwc;
-    ncnn::Layer* permute_cdhw;
-
-    Pipeline* pipeline_reshape;
-    Pipeline* pipeline_reshape_pack4;
-    Pipeline* pipeline_reshape_pack1to4;
-    Pipeline* pipeline_reshape_pack4to1;
-    Pipeline* pipeline_reshape_pack8;
-    Pipeline* pipeline_reshape_pack1to8;
-    Pipeline* pipeline_reshape_pack4to8;
-    Pipeline* pipeline_reshape_pack8to4;
-    Pipeline* pipeline_reshape_pack8to1;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_RESHAPE_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/scale_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/scale_vulkan.h
deleted file mode 100644
index 72851030..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/scale_vulkan.h
+++ /dev/null
@@ -1,51 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SCALE_VULKAN_H
-#define LAYER_SCALE_VULKAN_H
-
-#include "scale.h"
-
-namespace ncnn {
-
-class Scale_vulkan : public Scale
-{
-public:
-    Scale_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int upload_model(VkTransfer& cmd, const Option& opt);
-
-    using Scale::forward_inplace;
-    virtual int forward_inplace(std::vector<VkMat>& bottom_top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(std::vector<VkImageMat>& bottom_top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    VkMat scale_data_gpu;
-    VkMat bias_data_gpu;
-    VkImageMat scale_data_gpu_image;
-    VkImageMat bias_data_gpu_image;
-
-    Pipeline* pipeline_scale;
-    Pipeline* pipeline_scale_pack4;
-    Pipeline* pipeline_scale_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SCALE_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/shufflechannel_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/shufflechannel_vulkan.h
deleted file mode 100644
index 1cbc706b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/shufflechannel_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SHUFFLECHANNEL_VULKAN_H
-#define LAYER_SHUFFLECHANNEL_VULKAN_H
-
-#include "shufflechannel.h"
-
-namespace ncnn {
-
-class ShuffleChannel_vulkan : public ShuffleChannel
-{
-public:
-    ShuffleChannel_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using ShuffleChannel::forward;
-    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_shufflechannel;
-    Pipeline* pipeline_shufflechannel_pack4;
-    Pipeline* pipeline_shufflechannel_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SHUFFLECHANNEL_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/sigmoid_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/sigmoid_vulkan.h
deleted file mode 100644
index 1350f6a4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/sigmoid_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SIGMOID_VULKAN_H
-#define LAYER_SIGMOID_VULKAN_H
-
-#include "sigmoid.h"
-
-namespace ncnn {
-
-class Sigmoid_vulkan : public Sigmoid
-{
-public:
-    Sigmoid_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Sigmoid::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_sigmoid;
-    Pipeline* pipeline_sigmoid_pack4;
-    Pipeline* pipeline_sigmoid_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SIGMOID_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/slice_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/slice_vulkan.h
deleted file mode 100644
index 92f9ad15..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/slice_vulkan.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SLICE_VULKAN_H
-#define LAYER_SLICE_VULKAN_H
-
-#include "slice.h"
-
-namespace ncnn {
-
-class Slice_vulkan : public Slice
-{
-public:
-    Slice_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Slice::forward;
-    virtual int forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const std::vector<VkImageMat>& bottom_blobs, std::vector<VkImageMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_slice[2];
-    Pipeline* pipeline_slice_pack4[2];
-    Pipeline* pipeline_slice_pack1to4[2];
-    Pipeline* pipeline_slice_pack8[2];
-    Pipeline* pipeline_slice_pack1to8[2];
-    Pipeline* pipeline_slice_pack4to8[2];
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SLICE_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/softmax_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/softmax_vulkan.h
deleted file mode 100644
index aeff8d40..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/softmax_vulkan.h
+++ /dev/null
@@ -1,53 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SOFTMAX_VULKAN_H
-#define LAYER_SOFTMAX_VULKAN_H
-
-#include "softmax.h"
-
-namespace ncnn {
-
-class Softmax_vulkan : public Softmax
-{
-public:
-    Softmax_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Softmax::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_softmax_reduce_max;
-    Pipeline* pipeline_softmax_exp_sub_max;
-    Pipeline* pipeline_softmax_reduce_sum;
-    Pipeline* pipeline_softmax_div_sum;
-
-    Pipeline* pipeline_softmax_reduce_max_pack4;
-    Pipeline* pipeline_softmax_exp_sub_max_pack4;
-    Pipeline* pipeline_softmax_reduce_sum_pack4;
-    Pipeline* pipeline_softmax_div_sum_pack4;
-
-    Pipeline* pipeline_softmax_reduce_max_pack8;
-    Pipeline* pipeline_softmax_exp_sub_max_pack8;
-    Pipeline* pipeline_softmax_reduce_sum_pack8;
-    Pipeline* pipeline_softmax_div_sum_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SOFTMAX_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/split_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/split_vulkan.h
deleted file mode 100644
index 8e1998a3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/split_vulkan.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SPLIT_VULKAN_H
-#define LAYER_SPLIT_VULKAN_H
-
-#include "split.h"
-
-namespace ncnn {
-
-class Split_vulkan : public Split
-{
-public:
-    Split_vulkan();
-
-    using Split::forward;
-    virtual int forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-    virtual int forward(const std::vector<VkImageMat>& bottom_blobs, std::vector<VkImageMat>& top_blobs, VkCompute& cmd, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SPLIT_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/swish_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/swish_vulkan.h
deleted file mode 100644
index a562767c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/swish_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SWISH_VULKAN_H
-#define LAYER_SWISH_VULKAN_H
-
-#include "swish.h"
-
-namespace ncnn {
-
-class Swish_vulkan : public Swish
-{
-public:
-    Swish_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using Swish::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_swish;
-    Pipeline* pipeline_swish_pack4;
-    Pipeline* pipeline_swish_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SWISH_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/tanh_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/tanh_vulkan.h
deleted file mode 100644
index 1926363a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/tanh_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_TANH_VULKAN_H
-#define LAYER_TANH_VULKAN_H
-
-#include "tanh.h"
-
-namespace ncnn {
-
-class TanH_vulkan : public TanH
-{
-public:
-    TanH_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using TanH::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_tanh;
-    Pipeline* pipeline_tanh_pack4;
-    Pipeline* pipeline_tanh_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_TANH_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/unaryop_vulkan.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/unaryop_vulkan.h
deleted file mode 100644
index bad5377f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/vulkan/unaryop_vulkan.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_UNARYOP_VULKAN_H
-#define LAYER_UNARYOP_VULKAN_H
-
-#include "unaryop.h"
-
-namespace ncnn {
-
-class UnaryOp_vulkan : public UnaryOp
-{
-public:
-    UnaryOp_vulkan();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    using UnaryOp::forward_inplace;
-    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-    virtual int forward_inplace(VkImageMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
-
-public:
-    Pipeline* pipeline_unaryop;
-    Pipeline* pipeline_unaryop_pack4;
-    Pipeline* pipeline_unaryop_pack8;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_UNARYOP_VULKAN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/avx512_mathfun.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/avx512_mathfun.h
deleted file mode 100644
index b5e47bdb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/avx512_mathfun.h
+++ /dev/null
@@ -1,859 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef AVX512_MATHFUN_H
-#define AVX512_MATHFUN_H
-
-#include <immintrin.h>
-#include <emmintrin.h>
-
-/* yes I know, the top of this file is quite ugly */
-
-#ifdef _MSC_VER /* visual c++ */
-#define ALIGN64_BEG __declspec(align(64))
-#define ALIGN64_END
-#else /* gcc or icc */
-#define ALIGN64_BEG
-#define ALIGN64_END __attribute__((aligned(64)))
-#endif
-
-/* declare some AVX constants -- why can't I figure a better way to do that? */
-#define _PS512_CONST(Name, Val) \
-    static const ALIGN64_BEG float _ps512_##Name[16] ALIGN64_END = {Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val}
-#define _PI32_CONST512(Name, Val) \
-    static const ALIGN64_BEG int _pi32_512_##Name[16] ALIGN64_END = {Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val}
-#define _PS512_CONST_TYPE(Name, Type, Val) \
-    static const ALIGN64_BEG Type _ps512_##Name[16] ALIGN64_END = {Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val, Val}
-
-_PS512_CONST(1, 1.0f);
-_PS512_CONST(0p5, 0.5f);
-/* the smallest non denormalized float number */
-_PS512_CONST_TYPE(min_norm_pos, int, 0x00800000);
-_PS512_CONST_TYPE(mant_mask, int, 0x7f800000);
-_PS512_CONST_TYPE(inv_mant_mask, int, ~0x7f800000);
-
-_PS512_CONST_TYPE(sign_mask, int, (int)0x80000000);
-
-_PI32_CONST512(0, 0);
-_PI32_CONST512(1, 1);
-_PI32_CONST512(inv1, ~1);
-_PI32_CONST512(2, 2);
-_PI32_CONST512(4, 4);
-_PI32_CONST512(0x7f, 0x7f);
-
-_PS512_CONST(cephes_SQRTHF, 0.707106781186547524f);
-_PS512_CONST(cephes_log_p0, 7.0376836292E-2f);
-_PS512_CONST(cephes_log_p1, -1.1514610310E-1f);
-_PS512_CONST(cephes_log_p2, 1.1676998740E-1f);
-_PS512_CONST(cephes_log_p3, -1.2420140846E-1f);
-_PS512_CONST(cephes_log_p4, +1.4249322787E-1f);
-_PS512_CONST(cephes_log_p5, -1.6668057665E-1f);
-_PS512_CONST(cephes_log_p6, +2.0000714765E-1f);
-_PS512_CONST(cephes_log_p7, -2.4999993993E-1f);
-_PS512_CONST(cephes_log_p8, +3.3333331174E-1f);
-_PS512_CONST(cephes_log_q1, -2.12194440e-4f);
-_PS512_CONST(cephes_log_q2, 0.693359375f);
-
-/* natural logarithm computed for 8 simultaneous float
-   return NaN for x <= 0
-*/
-static NCNN_FORCEINLINE __m512 log512_ps(__m512 x)
-{
-    __m512i imm0;
-    __m512 one = *(__m512*)_ps512_1;
-
-    __mmask16 invalid_mask = _mm512_cmp_ps_mask(x, _mm512_setzero_ps(), _CMP_LE_OQ);
-
-    x = _mm512_max_ps(x, *(__m512*)_ps512_min_norm_pos); /* cut off denormalized stuff */
-
-    imm0 = _mm512_srli_epi32(_mm512_castps_si512(x), 23);
-
-    /* keep only the fractional part */
-    x = _mm512_and_ps(x, *(__m512*)_ps512_inv_mant_mask);
-    x = _mm512_or_ps(x, *(__m512*)_ps512_0p5);
-
-    imm0 = _mm512_sub_epi32(imm0, *(__m512i*)_pi32_512_0x7f);
-    __m512 e = _mm512_cvtepi32_ps(imm0);
-
-    e = _mm512_add_ps(e, one);
-
-    /* part2:
-       if( x < SQRTHF ) {
-         e -= 1;
-         x = x + x - 1.0;
-       } else { x = x - 1.0; }
-    */
-    __mmask16 mask = _mm512_cmp_ps_mask(x, *(__m512*)_ps512_cephes_SQRTHF, _CMP_LT_OQ);
-    __m512 tmp = _mm512_sub_ps(x, one);
-    e = _mm512_mask_sub_ps(e, mask, e, one);
-    x = _mm512_mask_add_ps(tmp, mask, tmp, x);
-
-    __m512 z = _mm512_mul_ps(x, x);
-
-    __m512 y = *(__m512*)_ps512_cephes_log_p0;
-    y = _mm512_fmadd_ps(y, x, *(__m512*)_ps512_cephes_log_p1);
-    y = _mm512_fmadd_ps(y, x, *(__m512*)_ps512_cephes_log_p2);
-    y = _mm512_fmadd_ps(y, x, *(__m512*)_ps512_cephes_log_p3);
-    y = _mm512_fmadd_ps(y, x, *(__m512*)_ps512_cephes_log_p4);
-    y = _mm512_fmadd_ps(y, x, *(__m512*)_ps512_cephes_log_p5);
-    y = _mm512_fmadd_ps(y, x, *(__m512*)_ps512_cephes_log_p6);
-    y = _mm512_fmadd_ps(y, x, *(__m512*)_ps512_cephes_log_p7);
-    y = _mm512_fmadd_ps(y, x, *(__m512*)_ps512_cephes_log_p8);
-    y = _mm512_mul_ps(y, x);
-
-    y = _mm512_mul_ps(y, z);
-
-    y = _mm512_fmadd_ps(e, *(__m512*)_ps512_cephes_log_q1, y);
-
-    //y = -z * 0.5 + y
-    y = _mm512_fnmadd_ps(z, *(__m512*)_ps512_0p5, y);
-
-    x = _mm512_add_ps(x, y);
-    x = _mm512_fmadd_ps(e, *(__m512*)_ps512_cephes_log_q2, x);
-    y = _mm512_or_ps(x, _mm512_castsi512_ps(_mm512_movm_epi32(invalid_mask))); // negative arg will be NAN
-    return y;
-}
-
-_PS512_CONST(exp_hi, 88.3762626647949f);
-_PS512_CONST(exp_lo, -88.3762626647949f);
-
-_PS512_CONST(cephes_LOG2EF, 1.44269504088896341f);
-_PS512_CONST(cephes_exp_C1, 0.693359375f);
-_PS512_CONST(cephes_exp_C2, -2.12194440e-4f);
-
-_PS512_CONST(cephes_exp_p0, 1.9875691500E-4f);
-_PS512_CONST(cephes_exp_p1, 1.3981999507E-3f);
-_PS512_CONST(cephes_exp_p2, 8.3334519073E-3f);
-_PS512_CONST(cephes_exp_p3, 4.1665795894E-2f);
-_PS512_CONST(cephes_exp_p4, 1.6666665459E-1f);
-_PS512_CONST(cephes_exp_p5, 5.0000001201E-1f);
-
-static NCNN_FORCEINLINE __m512 exp512_ps(__m512 x)
-{
-    __m512 tmp = _mm512_setzero_ps(), fx;
-    __m512i imm0;
-    __m512 one = *(__m512*)_ps512_1;
-
-    x = _mm512_min_ps(x, *(__m512*)_ps512_exp_hi);
-    x = _mm512_max_ps(x, *(__m512*)_ps512_exp_lo);
-
-    /* express exp(x) as exp(g + n*log(2)) */
-    fx = _mm512_fmadd_ps(x, *(__m512*)_ps512_cephes_LOG2EF, *(__m512*)_ps512_0p5);
-
-    tmp = _mm512_roundscale_ps(fx, _MM_FROUND_TO_NEG_INF);
-
-    /* if greater, subtract 1 */
-    __mmask16 mask = _mm512_cmp_ps_mask(tmp, fx, _CMP_GT_OQ);
-    fx = _mm512_mask_sub_ps(tmp, mask, tmp, one);
-
-    // x = x - fx * exp_C1
-    x = _mm512_fnmadd_ps(fx, *(__m512*)_ps512_cephes_exp_C1, x);
-    // x = x - fx * exp_C2
-    x = _mm512_fnmadd_ps(fx, *(__m512*)_ps512_cephes_exp_C2, x);
-
-    tmp = _mm512_mul_ps(x, x);
-
-    __m512 y = *(__m512*)_ps512_cephes_exp_p0;
-    y = _mm512_fmadd_ps(y, x, *(__m512*)_ps512_cephes_exp_p1);
-    y = _mm512_fmadd_ps(y, x, *(__m512*)_ps512_cephes_exp_p2);
-    y = _mm512_fmadd_ps(y, x, *(__m512*)_ps512_cephes_exp_p3);
-    y = _mm512_fmadd_ps(y, x, *(__m512*)_ps512_cephes_exp_p4);
-    y = _mm512_fmadd_ps(y, x, *(__m512*)_ps512_cephes_exp_p5);
-    y = _mm512_fmadd_ps(y, tmp, x);
-    y = _mm512_add_ps(y, one);
-
-    /* build 2^n */
-    imm0 = _mm512_cvttps_epi32(fx);
-    imm0 = _mm512_add_epi32(imm0, *(__m512i*)_pi32_512_0x7f);
-    imm0 = _mm512_slli_epi32(imm0, 23);
-    __m512 pow2n = _mm512_castsi512_ps(imm0);
-    y = _mm512_mul_ps(y, pow2n);
-    return y;
-}
-
-_PS512_CONST(tanh_hi, 9.0f);
-_PS512_CONST(tanh_lo, -9.0f);
-
-_PS512_CONST(cephes_tanh_p0, -2.76076847742355E-16f);
-_PS512_CONST(cephes_tanh_p1, 2.00018790482477E-13f);
-_PS512_CONST(cephes_tanh_p2, -8.60467152213735E-11f);
-_PS512_CONST(cephes_tanh_p3, 5.12229709037114E-08f);
-_PS512_CONST(cephes_tanh_p4, 1.48572235717979E-05f);
-_PS512_CONST(cephes_tanh_p5, 6.37261928875436E-04f);
-_PS512_CONST(cephes_tanh_p6, 4.89352455891786E-03f);
-
-_PS512_CONST(cephes_tanh_p7, 1.19825839466702e-06f);
-_PS512_CONST(cephes_tanh_p8, 1.18534705686654e-04f);
-_PS512_CONST(cephes_tanh_p9, 2.26843463243900e-03f);
-
-// an approximation of tanh
-static inline __m512 tanh512_ps(const __m512 x)
-{
-    __m512 value = x;
-    value = _mm512_max_ps(*(__m512*)_ps512_tanh_lo, value);
-    value = _mm512_min_ps(*(__m512*)_ps512_tanh_hi, value);
-
-    __m512 value_squared = _mm512_mul_ps(value, value);
-
-    __m512 p;
-    p = _mm512_fmadd_ps(value_squared, *(__m512*)_ps512_cephes_tanh_p0, *(__m512*)_ps512_cephes_tanh_p1);
-    p = _mm512_fmadd_ps(p, value_squared, *(__m512*)_ps512_cephes_tanh_p2);
-    p = _mm512_fmadd_ps(p, value_squared, *(__m512*)_ps512_cephes_tanh_p3);
-    p = _mm512_fmadd_ps(p, value_squared, *(__m512*)_ps512_cephes_tanh_p4);
-    p = _mm512_fmadd_ps(p, value_squared, *(__m512*)_ps512_cephes_tanh_p5);
-    p = _mm512_fmadd_ps(p, value_squared, *(__m512*)_ps512_cephes_tanh_p6);
-    p = _mm512_mul_ps(p, value);
-
-    __m512 q;
-    q = _mm512_fmadd_ps(value_squared, *(__m512*)_ps512_cephes_tanh_p7, *(__m512*)_ps512_cephes_tanh_p8);
-    q = _mm512_fmadd_ps(q, value_squared, *(__m512*)_ps512_cephes_tanh_p9);
-    q = _mm512_fmadd_ps(q, value_squared, *(__m512*)_ps512_cephes_tanh_p6);
-
-    __m512 dst = _mm512_div_ps(p, q);
-    return dst;
-}
-
-_PS512_CONST(minus_cephes_DP1, -0.78515625f);
-_PS512_CONST(minus_cephes_DP2, -2.4187564849853515625e-4f);
-_PS512_CONST(minus_cephes_DP3, -3.77489497744594108e-8f);
-_PS512_CONST(sincof_p0, -1.9515295891E-4f);
-_PS512_CONST(sincof_p1, 8.3321608736E-3f);
-_PS512_CONST(sincof_p2, -1.6666654611E-1f);
-_PS512_CONST(coscof_p0, 2.443315711809948E-005f);
-_PS512_CONST(coscof_p1, -1.388731625493765E-003f);
-_PS512_CONST(coscof_p2, 4.166664568298827E-002f);
-_PS512_CONST(cephes_FOPI, 1.27323954473516f); // 4 / M_PI
-
-/* evaluation of 8 sines at onces using AVX intrisics
-
-   The code is the exact rewriting of the cephes sinf function.
-   Precision is excellent as long as x < 8192 (I did not bother to
-   take into account the special handling they have for greater values
-   -- it does not return garbage for arguments over 8192, though, but
-   the extra precision is missing).
-
-   Note that it is such that sinf((float)M_PI) = 8.74e-8, which is the
-   surprising but correct result.
-
-*/
-static NCNN_FORCEINLINE __m512 sin512_ps(__m512 x)
-{   // any x
-    __m512 xmm1, xmm2, xmm3, sign_bit, y;
-    __m512i imm0, imm2;
-
-    sign_bit = x;
-    /* take the absolute value */
-    x = _mm512_castsi512_ps(_mm512_and_epi32(_mm512_castps_si512(x), _mm512_set1_epi32(0x7fffffff)));
-    /* extract the sign bit (upper one) */
-    sign_bit = _mm512_and_ps(sign_bit, *(__m512*)_ps512_sign_mask);
-
-    /* scale by 4/Pi */
-    y = _mm512_mul_ps(x, *(__m512*)_ps512_cephes_FOPI);
-
-    /* store the integer part of y in mm0 */
-    imm2 = _mm512_cvttps_epi32(y);
-    /* j=(j+1) & (~1) (see the cephes sources) */
-    imm2 = _mm512_add_epi32(imm2, *(__m512i*)_pi32_512_1);
-    imm2 = _mm512_and_si512(imm2, *(__m512i*)_pi32_512_inv1);
-    y = _mm512_cvtepi32_ps(imm2);
-
-    /* get the swap sign flag */
-    imm0 = _mm512_and_si512(imm2, *(__m512i*)_pi32_512_4);
-    imm0 = _mm512_slli_epi32(imm0, 29);
-    /* get the polynom selection mask
-       there is one polynom for 0 <= x <= Pi/4
-       and another one for Pi/4<x<=Pi/2
-
-       Both branches will be computed.
-    */
-    imm2 = _mm512_and_si512(imm2, *(__m512i*)_pi32_512_2);
-    imm2 = _mm512_movm_epi32(_mm512_cmpeq_epi32_mask(imm2, *(__m512i*)_pi32_512_0));
-
-    __m512 swap_sign_bit = _mm512_castsi512_ps(imm0);
-    __m512 poly_mask = _mm512_castsi512_ps(imm2);
-    sign_bit = _mm512_xor_ps(sign_bit, swap_sign_bit);
-
-    /* The magic pass: "Extended precision modular arithmetic"
-       x = ((x - y * DP1) - y * DP2) - y * DP3; */
-    xmm1 = *(__m512*)_ps512_minus_cephes_DP1;
-    xmm2 = *(__m512*)_ps512_minus_cephes_DP2;
-    xmm3 = *(__m512*)_ps512_minus_cephes_DP3;
-    x = _mm512_fmadd_ps(y, xmm1, x);
-    x = _mm512_fmadd_ps(y, xmm2, x);
-    x = _mm512_fmadd_ps(y, xmm3, x);
-
-    /* Evaluate the first polynom  (0 <= x <= Pi/4) */
-    y = *(__m512*)_ps512_coscof_p0;
-    __m512 z = _mm512_mul_ps(x, x);
-
-    y = _mm512_fmadd_ps(y, z, *(__m512*)_ps512_coscof_p1);
-    y = _mm512_fmadd_ps(y, z, *(__m512*)_ps512_coscof_p2);
-    y = _mm512_mul_ps(y, z);
-    y = _mm512_mul_ps(y, z);
-    // y = y - z * 0.5
-    y = _mm512_fnmadd_ps(z, *(__m512*)_ps512_0p5, y);
-    y = _mm512_add_ps(y, *(__m512*)_ps512_1);
-
-    /* Evaluate the second polynom  (Pi/4 <= x <= 0) */
-
-    __m512 y2 = *(__m512*)_ps512_sincof_p0;
-    y2 = _mm512_fmadd_ps(y2, z, *(__m512*)_ps512_sincof_p1);
-    y2 = _mm512_fmadd_ps(y2, z, *(__m512*)_ps512_sincof_p2);
-    y2 = _mm512_mul_ps(y2, z);
-    y2 = _mm512_fmadd_ps(y2, x, x);
-
-    /* select the correct result from the two polynoms */
-    xmm3 = poly_mask;
-    y2 = _mm512_and_ps(xmm3, y2); //, xmm3);
-    y = _mm512_andnot_ps(xmm3, y);
-    y = _mm512_add_ps(y, y2);
-    /* update the sign */
-    y = _mm512_xor_ps(y, sign_bit);
-
-    return y;
-}
-
-/* almost the same as sin_ps */
-static NCNN_FORCEINLINE __m512 cos512_ps(__m512 x)
-{   // any x
-    __m512 xmm1, xmm2, xmm3, y;
-    __m512i imm0, imm2;
-
-    /* take the absolute value */
-    x = _mm512_castsi512_ps(_mm512_and_epi32(_mm512_castps_si512(x), _mm512_set1_epi32(0x7fffffff)));
-
-    /* scale by 4/Pi */
-    y = _mm512_mul_ps(x, *(__m512*)_ps512_cephes_FOPI);
-
-    /* store the integer part of y in mm0 */
-    imm2 = _mm512_cvttps_epi32(y);
-    /* j=(j+1) & (~1) (see the cephes sources) */
-    imm2 = _mm512_add_epi32(imm2, *(__m512i*)_pi32_512_1);
-    imm2 = _mm512_and_si512(imm2, *(__m512i*)_pi32_512_inv1);
-    y = _mm512_cvtepi32_ps(imm2);
-    imm2 = _mm512_sub_epi32(imm2, *(__m512i*)_pi32_512_2);
-
-    /* get the swap sign flag */
-    imm0 = _mm512_andnot_si512(imm2, *(__m512i*)_pi32_512_4);
-    imm0 = _mm512_slli_epi32(imm0, 29);
-    /* get the polynom selection mask */
-    imm2 = _mm512_and_si512(imm2, *(__m512i*)_pi32_512_2);
-    imm2 = _mm512_movm_epi32(_mm512_cmpeq_epi32_mask(imm2, *(__m512i*)_pi32_512_0));
-
-    __m512 sign_bit = _mm512_castsi512_ps(imm0);
-    __m512 poly_mask = _mm512_castsi512_ps(imm2);
-
-    /* The magic pass: "Extended precision modular arithmetic"
-       x = ((x - y * DP1) - y * DP2) - y * DP3; */
-    xmm1 = *(__m512*)_ps512_minus_cephes_DP1;
-    xmm2 = *(__m512*)_ps512_minus_cephes_DP2;
-    xmm3 = *(__m512*)_ps512_minus_cephes_DP3;
-    x = _mm512_fmadd_ps(y, xmm1, x);
-    x = _mm512_fmadd_ps(y, xmm2, x);
-    x = _mm512_fmadd_ps(y, xmm3, x);
-
-    /* Evaluate the first polynom  (0 <= x <= Pi/4) */
-    y = *(__m512*)_ps512_coscof_p0;
-    __m512 z = _mm512_mul_ps(x, x);
-
-    y = _mm512_fmadd_ps(y, z, *(__m512*)_ps512_coscof_p1);
-    y = _mm512_fmadd_ps(y, z, *(__m512*)_ps512_coscof_p2);
-    y = _mm512_mul_ps(y, z);
-    y = _mm512_mul_ps(y, z);
-    // y = y - z * 0.5
-    y = _mm512_fnmadd_ps(z, *(__m512*)_ps512_0p5, y);
-    y = _mm512_add_ps(y, *(__m512*)_ps512_1);
-
-    /* Evaluate the second polynom  (Pi/4 <= x <= 0) */
-
-    __m512 y2 = *(__m512*)_ps512_sincof_p0;
-    y2 = _mm512_mul_ps(y2, z);
-    y2 = _mm512_add_ps(y2, *(__m512*)_ps512_sincof_p1);
-    y2 = _mm512_mul_ps(y2, z);
-    y2 = _mm512_add_ps(y2, *(__m512*)_ps512_sincof_p2);
-    y2 = _mm512_mul_ps(y2, z);
-    y2 = _mm512_mul_ps(y2, x);
-    y2 = _mm512_add_ps(y2, x);
-
-    /* select the correct result from the two polynoms */
-    xmm3 = poly_mask;
-    y2 = _mm512_and_ps(xmm3, y2); //, xmm3);
-    y = _mm512_andnot_ps(xmm3, y);
-    y = _mm512_add_ps(y, y2);
-    /* update the sign */
-    y = _mm512_xor_ps(y, sign_bit);
-
-    return y;
-}
-
-/* since sin512_ps and cos512_ps are almost identical, sincos512_ps could replace both of them..
-   it is almost as fast, and gives you a free cosine with your sine */
-static NCNN_FORCEINLINE void sincos512_ps(__m512 x, __m512* s, __m512* c)
-{
-    __m512 xmm1, xmm2, xmm3, sign_bit_sin, y;
-    __m512i imm0, imm2, imm4;
-
-    sign_bit_sin = x;
-    /* take the absolute value */
-    x = _mm512_castsi512_ps(_mm512_and_epi32(_mm512_castps_si512(x), _mm512_set1_epi32(0x7fffffff)));
-    /* extract the sign bit (upper one) */
-    sign_bit_sin = _mm512_and_ps(sign_bit_sin, *(__m512*)_ps512_sign_mask);
-
-    /* scale by 4/Pi */
-    y = _mm512_mul_ps(x, *(__m512*)_ps512_cephes_FOPI);
-
-    /* store the integer part of y in imm2 */
-    imm2 = _mm512_cvttps_epi32(y);
-
-    /* j=(j+1) & (~1) (see the cephes sources) */
-    imm2 = _mm512_add_epi32(imm2, *(__m512i*)_pi32_512_1);
-    imm2 = _mm512_and_si512(imm2, *(__m512i*)_pi32_512_inv1);
-
-    y = _mm512_cvtepi32_ps(imm2);
-    imm4 = imm2;
-
-    /* get the swap sign flag for the sine */
-    imm0 = _mm512_and_si512(imm2, *(__m512i*)_pi32_512_4);
-    imm0 = _mm512_slli_epi32(imm0, 29);
-    //__m512 swap_sign_bit_sin = _mm512_castsi512_ps(imm0);
-
-    /* get the polynom selection mask for the sine*/
-    imm2 = _mm512_and_si512(imm2, *(__m512i*)_pi32_512_2);
-    imm2 = _mm512_movm_epi32(_mm512_cmpeq_epi32_mask(imm2, *(__m512i*)_pi32_512_0));
-    //__m512 poly_mask = _mm512_castsi512_ps(imm2);
-
-    __m512 swap_sign_bit_sin = _mm512_castsi512_ps(imm0);
-    __m512 poly_mask = _mm512_castsi512_ps(imm2);
-
-    /* The magic pass: "Extended precision modular arithmetic"
-       x = ((x - y * DP1) - y * DP2) - y * DP3; */
-    xmm1 = *(__m512*)_ps512_minus_cephes_DP1;
-    xmm2 = *(__m512*)_ps512_minus_cephes_DP2;
-    xmm3 = *(__m512*)_ps512_minus_cephes_DP3;
-    x = _mm512_fmadd_ps(y, xmm1, x);
-    x = _mm512_fmadd_ps(y, xmm2, x);
-    x = _mm512_fmadd_ps(y, xmm3, x);
-
-    imm4 = _mm512_sub_epi32(imm4, *(__m512i*)_pi32_512_2);
-    imm4 = _mm512_andnot_si512(imm4, *(__m512i*)_pi32_512_4);
-    imm4 = _mm512_slli_epi32(imm4, 29);
-
-    __m512 sign_bit_cos = _mm512_castsi512_ps(imm4);
-
-    sign_bit_sin = _mm512_xor_ps(sign_bit_sin, swap_sign_bit_sin);
-
-    /* Evaluate the first polynom  (0 <= x <= Pi/4) */
-    __m512 z = _mm512_mul_ps(x, x);
-    y = *(__m512*)_ps512_coscof_p0;
-
-    y = _mm512_fmadd_ps(y, z, *(__m512*)_ps512_coscof_p1);
-    y = _mm512_fmadd_ps(y, z, *(__m512*)_ps512_coscof_p2);
-    y = _mm512_mul_ps(y, z);
-    y = _mm512_mul_ps(y, z);
-    // y = y - z * 0.5
-    y = _mm512_fnmadd_ps(z, *(__m512*)_ps512_0p5, y);
-    y = _mm512_add_ps(y, *(__m512*)_ps512_1);
-
-    /* Evaluate the second polynom  (Pi/4 <= x <= 0) */
-
-    __m512 y2 = *(__m512*)_ps512_sincof_p0;
-    y2 = _mm512_fmadd_ps(y2, z, *(__m512*)_ps512_sincof_p1);
-    y2 = _mm512_fmadd_ps(y2, z, *(__m512*)_ps512_sincof_p2);
-    y2 = _mm512_mul_ps(y2, z);
-    y2 = _mm512_fmadd_ps(y2, x, x);
-
-    /* select the correct result from the two polynoms */
-    xmm3 = poly_mask;
-    __m512 ysin2 = _mm512_and_ps(xmm3, y2);
-    __m512 ysin1 = _mm512_andnot_ps(xmm3, y);
-    y2 = _mm512_sub_ps(y2, ysin2);
-    y = _mm512_sub_ps(y, ysin1);
-
-    xmm1 = _mm512_add_ps(ysin1, ysin2);
-    xmm2 = _mm512_add_ps(y, y2);
-
-    /* update the sign */
-    *s = _mm512_xor_ps(xmm1, sign_bit_sin);
-    *c = _mm512_xor_ps(xmm2, sign_bit_cos);
-}
-
-static NCNN_FORCEINLINE __m512 tan512_ps(__m512 x)
-{
-    __m512 ysin, ycos;
-    __m512 eps = _mm512_set1_ps(1E-8f);
-    sincos512_ps(x, &ysin, &ycos);
-    __mmask16 mask = _mm512_cmp_ps_mask(ycos, _mm512_setzero_ps(), _CMP_EQ_OS);
-    ycos = _mm512_mask_add_ps(ycos, mask, ycos, eps);
-    __m512 ytan = _mm512_div_ps(ysin, ycos);
-    return ytan;
-}
-
-static NCNN_FORCEINLINE __m512 pow512_ps(__m512 a, __m512 b)
-{
-    // pow(x, m) = exp(m * log(x))
-    return exp512_ps(_mm512_mul_ps(b, log512_ps(a)));
-}
-
-static NCNN_FORCEINLINE __m512 asin512_ps(__m512 x)
-{
-    const __m512 magic_negative_zero = _mm512_set1_ps(-0.0f);
-    const __m512 magic_half_one = _mm512_set1_ps(0.5f);
-    const __m512 magic_one = _mm512_set1_ps(1.0f);
-    const __m512 magic_a4 = _mm512_set1_ps(0.023994016f);
-    const __m512 magic_a5 = _mm512_set1_ps(0.042417344f);
-    const __m512 magic_a2 = _mm512_set1_ps(0.07494697f);
-    const __m512 magic_a3 = _mm512_set1_ps(0.045520633f);
-    const __m512 magic_a0 = _mm512_set1_ps(1.0f);
-    const __m512 magic_a1 = _mm512_set1_ps(0.166667819f);
-    const __m512 magic_half_pi = _mm512_set1_ps(1.5707964f);
-    const __m512 magic_three = _mm512_set1_ps(3.0f);
-
-    // negative_mask = magic_negative_zero && x;
-    __m512 negative_mask = _mm512_and_ps(magic_negative_zero, x);
-
-    // absolute = abs(x);
-    __m512 absolute = _mm512_andnot_ps(magic_negative_zero, x);
-
-    // Reference: https://en.wikipedia.org/wiki/Small-angle_approximation
-
-    // is_small_input = (absolute <= 0.5f);
-    __mmask16 is_small_input = _mm512_cmp_ps_mask(
-                                   absolute,
-                                   magic_half_one,
-                                   _CMP_LE_OQ);
-
-    // is_big_input = (is_small_input ? 0.0f : 1.0f);
-    __m512 is_big_input = _mm512_maskz_mov_ps(~is_small_input, magic_one);
-
-    // big_input_approx = sqrt(0.5f * (1 - absolute));
-    __m512 big_input_approx = _mm512_sqrt_ps(_mm512_mul_ps(
-                                  magic_half_one,
-                                  _mm512_sub_ps(magic_one, absolute)));
-
-    // input_approx = (is_small_input ? absolute : big_input_approx);
-    __m512 input_approx = _mm512_mask_mov_ps(
-                              big_input_approx,
-                              is_small_input,
-                              absolute);
-
-    // square_of_input_approx = input_approx * input_approx;
-    __m512 square_of_input_approx = _mm512_mul_ps(input_approx, input_approx);
-
-    // fourth_power_of_input_approx =
-    //     square_of_input_approx * square_of_input_approx;
-    __m512 fourth_power_of_input_approx = _mm512_mul_ps(
-            square_of_input_approx, square_of_input_approx);
-
-    // TODO: Need more explanations.
-    // x1 = ((fourth_power_of_input_approx * magic_a4) + magic_a2);
-    // x2 = ((fourth_power_of_input_approx * magic_a5) + magic_a3);
-    // x3 = ((fourth_power_of_input_approx * x1) + magic_a0);
-    // x4 = ((fourth_power_of_input_approx * x2) + magic_a1);
-    // output_approx = ((square_of_input_approx * x4) + x3);
-    __m512 output_approx = _mm512_fmadd_ps(
-                               square_of_input_approx,
-                               _mm512_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm512_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       magic_a5,
-                                       magic_a3),
-                                   magic_a1),
-                               _mm512_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm512_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       magic_a4,
-                                       magic_a2),
-                                   magic_a0));
-
-    // TODO: Need more explanations.
-    // x1 = ((0.5 * PI) * is_big_input);
-    // x2 = (output_approx * input_approx);
-    // x3 = (-(3.0f * is_big_input) + 1.0f);
-    // final_approx = ((x2 * x3) + x1);
-    __m512 final_approx = _mm512_fmadd_ps(
-                              _mm512_mul_ps(output_approx, input_approx),
-                              _mm512_fnmadd_ps(magic_three, is_big_input, magic_one),
-                              _mm512_mul_ps(magic_half_pi, is_big_input));
-
-    // return (final_approx || negative_mask);
-    return _mm512_or_ps(final_approx, negative_mask);
-}
-
-static NCNN_FORCEINLINE __m512 acos512_ps(__m512 x)
-{
-    const __m512 magic_negative_zero = _mm512_set1_ps(-0.0f);
-    const __m512 magic_zero = _mm512_set1_ps(0.0f);
-    const __m512 magic_half_one = _mm512_set1_ps(0.5f);
-    const __m512 magic_one = _mm512_set1_ps(1.0f);
-    const __m512 magic_a4 = _mm512_set1_ps(0.023994016f);
-    const __m512 magic_a5 = _mm512_set1_ps(0.042417344f);
-    const __m512 magic_a2 = _mm512_set1_ps(0.07494697f);
-    const __m512 magic_a3 = _mm512_set1_ps(0.045520633f);
-    const __m512 magic_a0 = _mm512_set1_ps(1.0f);
-    const __m512 magic_a1 = _mm512_set1_ps(0.166667819f);
-    const __m512 magic_half_pi = _mm512_set1_ps(1.5707964f);
-    const __m512 magic_pi = _mm512_set1_ps(3.1415927f);
-
-    // negative_mask = magic_negative_zero && x;
-    __m512 negative_mask = _mm512_and_ps(magic_negative_zero, x);
-
-    // absolute = abs(x);
-    __m512 absolute = _mm512_andnot_ps(magic_negative_zero, x);
-
-    // Reference: https://en.wikipedia.org/wiki/Small-angle_approximation
-
-    // is_small_input = (absolute <= 0.5f);
-    __mmask16 is_small_input = _mm512_cmp_ps_mask(
-                                   absolute,
-                                   magic_half_one,
-                                   _CMP_LE_OQ);
-
-    // big_input_approx = sqrt(0.5f * (1 - absolute));
-    __m512 big_input_approx = _mm512_sqrt_ps(_mm512_mul_ps(
-                                  magic_half_one,
-                                  _mm512_sub_ps(magic_one, absolute)));
-
-    // input_approx = (is_small_input ? absolute : big_input_approx);
-    __m512 input_approx = _mm512_mask_mov_ps(
-                              big_input_approx,
-                              is_small_input,
-                              absolute);
-
-    // square_of_input_approx = input_approx * input_approx;
-    __m512 square_of_input_approx = _mm512_mul_ps(input_approx, input_approx);
-
-    // fourth_power_of_input_approx =
-    //     square_of_input_approx * square_of_input_approx;
-    __m512 fourth_power_of_input_approx = _mm512_mul_ps(
-            square_of_input_approx, square_of_input_approx);
-
-    // TODO: Need more explanations.
-    // x1 = ((fourth_power_of_input_approx * magic_a4) + magic_a2);
-    // x2 = ((fourth_power_of_input_approx * magic_a5) + magic_a3);
-    // x3 = ((fourth_power_of_input_approx * x1) + magic_a0);
-    // x4 = ((fourth_power_of_input_approx * x2) + magic_a1);
-    // output_approx = ((square_of_input_approx * x4) + x3);
-    __m512 output_approx = _mm512_fmadd_ps(
-                               square_of_input_approx,
-                               _mm512_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm512_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       magic_a5,
-                                       magic_a3),
-                                   magic_a1),
-                               _mm512_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm512_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       magic_a4,
-                                       magic_a2),
-                                   magic_a0));
-
-    // TODO: Need more explanations.
-    // x1 = (output_approx * input_approx);
-    __m512 x1 = _mm512_mul_ps(output_approx, input_approx);
-
-    // TODO: Need more explanations.
-    // small_final_approx = ((0.5 * PI) - (x1 | negative_mask));
-    __m512 small_final_approx = _mm512_sub_ps(
-                                    magic_half_pi,
-                                    _mm512_or_ps(x1, negative_mask));
-
-    // TODO: Need more explanations.
-    // big_final_approx = (((x < 0.0f) & PI) + ((x1 * 2) | negative_mask));
-    __m512 big_final_approx = _mm512_add_ps(
-                                  _mm512_maskz_mov_ps(
-                                      _mm512_cmp_ps_mask(x, magic_zero, _CMP_LT_OQ),
-                                      magic_pi),
-                                  _mm512_or_ps(_mm512_add_ps(x1, x1), negative_mask));
-
-    // return (is_small_input ? small_final_approx : big_final_approx);
-    return _mm512_mask_mov_ps(
-               big_final_approx,
-               is_small_input,
-               small_final_approx);
-}
-
-static NCNN_FORCEINLINE __m512 atan512_ps(__m512 x)
-{
-    const __m512 magic_negative_zero = _mm512_set1_ps(-0.0f);
-    const __m512 magic_one = _mm512_set1_ps(1.0f);
-    const __m512 magic_negative_one = _mm512_set1_ps(-1.0f);
-    const __m512 magic_half_pi = _mm512_set1_ps(1.5707964f);
-    const __m512 magic_a0 = _mm512_set1_ps(1.0f);
-    const __m512 magic_a1 = _mm512_set1_ps(-0.33333072f);
-    const __m512 magic_a2 = _mm512_set1_ps(0.1999262f);
-    const __m512 magic_a3 = _mm512_set1_ps(-0.14203644f);
-    const __m512 magic_a4 = _mm512_set1_ps(0.10640934f);
-    const __m512 magic_a5 = _mm512_set1_ps(-0.07504295f);
-    const __m512 magic_a6 = _mm512_set1_ps(0.04269152f);
-    const __m512 magic_a7 = _mm512_set1_ps(-0.01606863f);
-    const __m512 magic_a8 = _mm512_set1_ps(0.0028498897f);
-
-    // negative_mask = magic_negative_zero && x;
-    __m512 negative_mask = _mm512_and_ps(magic_negative_zero, x);
-
-    // absolute = abs(x);
-    __m512 absolute = _mm512_andnot_ps(magic_negative_zero, x);
-
-    // Reference: https://en.wikipedia.org/wiki/Small-angle_approximation
-
-    // is_small_input = (1.0f < absolute);
-    __mmask16 is_small_input = _mm512_cmp_ps_mask(
-                                   magic_one,
-                                   absolute,
-                                   _CMP_LT_OQ);
-
-    // x1 = (is_small_input ? -1.0f : absolute);
-    // x2 = (is_small_input ? absolute : 1.0f)
-    // input_approx = x1 / x2;
-    __m512 input_approx = _mm512_div_ps(
-                              _mm512_mask_mov_ps(absolute, is_small_input, magic_negative_one),
-                              _mm512_mask_mov_ps(magic_one, is_small_input, absolute));
-
-    // square_of_input_approx = input_approx * input_approx;
-    __m512 square_of_input_approx = _mm512_mul_ps(input_approx, input_approx);
-
-    // fourth_power_of_input_approx =
-    //     square_of_input_approx * square_of_input_approx;
-    __m512 fourth_power_of_input_approx = _mm512_mul_ps(
-            square_of_input_approx, square_of_input_approx);
-
-    // TODO: Need more explanations.
-    // x1 = ((fourth_power_of_input_approx * magic_a7) + magic_a5);
-    // x2 = ((fourth_power_of_input_approx * magic_a8) + magic_a6);
-    // x3 = ((fourth_power_of_input_approx * x1) + magic_a3);
-    // x4 = ((fourth_power_of_input_approx * x2) + magic_a4);
-    // x5 = ((fourth_power_of_input_approx * x3) + magic_a1);
-    // x6 = ((fourth_power_of_input_approx * x4) + magic_a2);
-    // x7 = ((fourth_power_of_input_approx * x6) + magic_a0);
-    // output_approx = ((square_of_input_approx * x5) + x7);
-    __m512 output_approx = _mm512_fmadd_ps(
-                               square_of_input_approx,
-                               _mm512_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm512_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       _mm512_fmadd_ps(
-                                           fourth_power_of_input_approx,
-                                           magic_a7,
-                                           magic_a5),
-                                       magic_a3),
-                                   magic_a1),
-                               _mm512_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm512_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       _mm512_fmadd_ps(
-                                           fourth_power_of_input_approx,
-                                           _mm512_fmadd_ps(
-                                                   fourth_power_of_input_approx,
-                                                   magic_a8,
-                                                   magic_a6),
-                                           magic_a4),
-                                       magic_a2),
-                                   magic_a0));
-
-    // TODO: Need more explanations.
-    // x1 = (output_approx * input_approx);
-    // if (is_small_input) x1 += (0.5 * PI);
-    // return (negative_mask ? -x1 : x1);
-    return _mm512_or_ps(
-               _mm512_add_ps(
-                   _mm512_mul_ps(output_approx, input_approx),
-                   _mm512_maskz_mov_ps(is_small_input, magic_half_pi)),
-               negative_mask);
-}
-
-// MSVC 2017 x86 CI will be broken if use NCNN_FORCEINLINE for atan2512_ps.
-// This function still be inlined compiled by MSVC 2017 even without that.
-#if _MSC_VER < 1920
-static __m512 atan2512_ps(__m512 y, __m512 x)
-#else
-static NCNN_FORCEINLINE __m512 atan2512_ps(__m512 y, __m512 x)
-#endif
-{
-    // Reference: https://mazzo.li/posts/vectorized-atan2.html
-
-    const __m512 magic_zero = _mm512_set1_ps(0.0f);
-    const __m512 magic_negative_zero = _mm512_set1_ps(-0.0f);
-    const __m512 magic_pi = _mm512_set1_ps(3.1415927f);
-    const __m512 magic_half_pi = _mm512_set1_ps(1.5707964f);
-
-    // not_equal_zero_x = (x != 0.0f);
-    __mmask16 not_equal_zero_x = _mm512_cmp_ps_mask(
-                                     x,
-                                     magic_zero,
-                                     _CMP_NEQ_OQ);
-
-    // not_equal_zero_y = (y != 0.0f);
-    __mmask16 not_equal_zero_y = _mm512_cmp_ps_mask(
-                                     y,
-                                     magic_zero,
-                                     _CMP_NEQ_OQ);
-
-    // normal_mode = ((x != 0.0f) & (y != 0.0f));
-    __mmask16 normal_mode = (not_equal_zero_x & not_equal_zero_y);
-
-    // negative_mask_x = magic_negative_zero && x;
-    __m512 negative_mask_x = _mm512_and_ps(magic_negative_zero, x);
-
-    // negative_mask_y = magic_negative_zero && y;
-    __m512 negative_mask_y = _mm512_and_ps(magic_negative_zero, y);
-
-    // pi_additions = ((x < 0.0f) ? ((y < 0.0f) ? -PI : PI) : 0.0f);
-    __m512 pi_additions = _mm512_mask_mov_ps(
-                              magic_zero,
-                              _mm512_cmp_ps_mask(x, magic_zero, _CMP_LT_OQ),
-                              _mm512_mask_mov_ps(
-                                  magic_pi,
-                                  _mm512_cmp_ps_mask(y, magic_zero, _CMP_LT_OQ),
-                                  _mm512_or_ps(magic_negative_zero, magic_pi)));
-
-    // normal_result = (atan(y / x) + pi_additions);
-    __m512 normal_result = _mm512_add_ps(
-                               atan512_ps(_mm512_div_ps(y, x)),
-                               pi_additions);
-
-    // negative_mask_full_x = ((negative_mask_x | PI) < 0.0f);
-    __mmask16 negative_mask_full_x = _mm512_cmp_ps_mask(
-                                         _mm512_or_ps(negative_mask_x, magic_pi),
-                                         magic_zero,
-                                         _CMP_LT_OQ);
-
-    // x1 = (negative_mask_y ? -(0.5 * PI) : (0.5 * PI));
-    // x2 = (negative_mask_full_x ? PI : 0.0f);
-    // special_result = ((y != 0.0f) ? x1 : x2);
-    __m512 special_result = _mm512_mask_mov_ps(
-                                _mm512_mask_mov_ps(magic_zero, negative_mask_full_x, magic_pi),
-                                not_equal_zero_y,
-                                _mm512_or_ps(negative_mask_y, magic_half_pi));
-
-    // return (normal_mode ? normal_result : special_result);
-    return _mm512_mask_mov_ps(special_result, normal_mode, normal_result);
-}
-
-static NCNN_FORCEINLINE __m512 abs512_ps(__m512 x)
-{
-    // Use negative zero as the sign bit mask.
-    const __m512 magic_negative_zero = _mm512_set1_ps(-0.0f);
-
-    // return (!magic_negative_zero && x);
-    return _mm512_andnot_ps(magic_negative_zero, x);
-}
-
-#endif // AVX512_MATHFUN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/avx_mathfun.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/avx_mathfun.h
deleted file mode 100644
index 65c34efc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/avx_mathfun.h
+++ /dev/null
@@ -1,1090 +0,0 @@
-/*
-   AVX implementation of sin, cos, sincos, exp and log
-
-   Based on "sse_mathfun.h", by Julien Pommier
-   http://gruntthepeon.free.fr/ssemath/
-
-   Copyright (C) 2012 Giovanni Garberoglio
-   Interdisciplinary Laboratory for Computational Science (LISC)
-   Fondazione Bruno Kessler and University of Trento
-   via Sommarive, 18
-   I-38123 Trento (Italy)
-
-  This software is provided 'as-is', without any express or implied
-  warranty.  In no event will the authors be held liable for any damages
-  arising from the use of this software.
-
-  Permission is granted to anyone to use this software for any purpose,
-  including commercial applications, and to alter it and redistribute it
-  freely, subject to the following restrictions:
-
-  1. The origin of this software must not be misrepresented; you must not
-     claim that you wrote the original software. If you use this software
-     in a product, an acknowledgment in the product documentation would be
-     appreciated but is not required.
-  2. Altered source versions must be plainly marked as such, and must not be
-     misrepresented as being the original software.
-  3. This notice may not be removed or altered from any source distribution.
-
-  (this is the zlib license)
-*/
-
-#ifndef AVX_MATHFUN_H
-#define AVX_MATHFUN_H
-
-#include <immintrin.h>
-#include <emmintrin.h>
-#include <x86_usability.h>
-
-/* yes I know, the top of this file is quite ugly */
-
-#ifdef _MSC_VER /* visual c++ */
-#define ALIGN32_BEG __declspec(align(32))
-#define ALIGN32_END
-#else /* gcc or icc */
-#define ALIGN32_BEG
-#define ALIGN32_END __attribute__((aligned(32)))
-#endif
-
-#define _PI32AVX_CONST(Name, Val) \
-    static const ALIGN32_BEG int _pi32avx_##Name[4] ALIGN32_END = {Val, Val, Val, Val}
-
-_PI32AVX_CONST(1, 1);
-_PI32AVX_CONST(inv1, ~1);
-_PI32AVX_CONST(2, 2);
-_PI32AVX_CONST(4, 4);
-
-/* declare some AVX constants -- why can't I figure a better way to do that? */
-#define _PS256_CONST(Name, Val) \
-    static const ALIGN32_BEG float _ps256_##Name[8] ALIGN32_END = {Val, Val, Val, Val, Val, Val, Val, Val}
-#define _PI32_CONST256(Name, Val) \
-    static const ALIGN32_BEG int _pi32_256_##Name[8] ALIGN32_END = {Val, Val, Val, Val, Val, Val, Val, Val}
-#define _PS256_CONST_TYPE(Name, Type, Val) \
-    static const ALIGN32_BEG Type _ps256_##Name[8] ALIGN32_END = {Val, Val, Val, Val, Val, Val, Val, Val}
-
-_PS256_CONST(1, 1.0f);
-_PS256_CONST(0p5, 0.5f);
-/* the smallest non denormalized float number */
-_PS256_CONST_TYPE(min_norm_pos, int, 0x00800000);
-_PS256_CONST_TYPE(mant_mask, int, 0x7f800000);
-_PS256_CONST_TYPE(inv_mant_mask, int, ~0x7f800000);
-
-_PS256_CONST_TYPE(sign_mask, int, (int)0x80000000);
-_PS256_CONST_TYPE(inv_sign_mask, int, ~0x80000000);
-
-_PI32_CONST256(0, 0);
-_PI32_CONST256(1, 1);
-_PI32_CONST256(inv1, ~1);
-_PI32_CONST256(2, 2);
-_PI32_CONST256(4, 4);
-_PI32_CONST256(0x7f, 0x7f);
-
-_PS256_CONST(cephes_SQRTHF, 0.707106781186547524f);
-_PS256_CONST(cephes_log_p0, 7.0376836292E-2f);
-_PS256_CONST(cephes_log_p1, -1.1514610310E-1f);
-_PS256_CONST(cephes_log_p2, 1.1676998740E-1f);
-_PS256_CONST(cephes_log_p3, -1.2420140846E-1f);
-_PS256_CONST(cephes_log_p4, +1.4249322787E-1f);
-_PS256_CONST(cephes_log_p5, -1.6668057665E-1f);
-_PS256_CONST(cephes_log_p6, +2.0000714765E-1f);
-_PS256_CONST(cephes_log_p7, -2.4999993993E-1f);
-_PS256_CONST(cephes_log_p8, +3.3333331174E-1f);
-_PS256_CONST(cephes_log_q1, -2.12194440e-4f);
-_PS256_CONST(cephes_log_q2, 0.693359375f);
-
-#ifndef __AVX2__
-typedef union imm_xmm_union
-{
-    __m256i imm;
-    __m128i xmm[2];
-} imm_xmm_union;
-
-#define COPY_IMM_TO_XMM(imm_, xmm0_, xmm1_)      \
-    {                                            \
-        ALIGN32_BEG imm_xmm_union u ALIGN32_END; \
-        u.imm = imm_;                            \
-        xmm0_ = u.xmm[0];                        \
-        xmm1_ = u.xmm[1];                        \
-    }
-
-#define COPY_XMM_TO_IMM(xmm0_, xmm1_, imm_)      \
-    {                                            \
-        ALIGN32_BEG imm_xmm_union u ALIGN32_END; \
-        u.xmm[0] = xmm0_;                        \
-        u.xmm[1] = xmm1_;                        \
-        imm_ = u.imm;                            \
-    }
-
-#define AVX2_BITOP_USING_SSE2(fn)                                      \
-    static NCNN_FORCEINLINE __m256i _mm256_comp_##fn(__m256i x, int a) \
-    {                                                                  \
-        /* use SSE2 instruction to perform the bitop AVX2 */           \
-        __m128i x1, x2;                                                \
-        __m256i ret;                                                   \
-        COPY_IMM_TO_XMM(x, x1, x2);                                    \
-        x1 = _mm_##fn(x1, a);                                          \
-        x2 = _mm_##fn(x2, a);                                          \
-        COPY_XMM_TO_IMM(x1, x2, ret);                                  \
-        return (ret);                                                  \
-    }
-#define AVX2_INTOP_USING_SSE2(fn)                                          \
-    static NCNN_FORCEINLINE __m256i _mm256_comp_##fn(__m256i x, __m256i y) \
-    {                                                                      \
-        /* use SSE2 instructions to perform the AVX2 integer operation */  \
-        __m128i x1, x2;                                                    \
-        __m128i y1, y2;                                                    \
-        __m256i ret;                                                       \
-        COPY_IMM_TO_XMM(x, x1, x2);                                        \
-        COPY_IMM_TO_XMM(y, y1, y2);                                        \
-        x1 = _mm_##fn(x1, y1);                                             \
-        x2 = _mm_##fn(x2, y2);                                             \
-        COPY_XMM_TO_IMM(x1, x2, ret);                                      \
-        return (ret);                                                      \
-    }
-#else
-#define AVX2_BITOP_USING_SSE2(fn)                                      \
-    static NCNN_FORCEINLINE __m256i _mm256_comp_##fn(__m256i x, int a) \
-    {                                                                  \
-        return _mm256_##fn(x, a);                                      \
-    }
-#define AVX2_INTOP_USING_SSE2(fn)                                          \
-    static NCNN_FORCEINLINE __m256i _mm256_comp_##fn(__m256i x, __m256i y) \
-    {                                                                      \
-        return _mm256_##fn(x, y);                                          \
-    }
-#endif
-
-AVX2_BITOP_USING_SSE2(slli_epi32)
-AVX2_BITOP_USING_SSE2(srli_epi32)
-AVX2_INTOP_USING_SSE2(cmpeq_epi32)
-AVX2_INTOP_USING_SSE2(sub_epi32)
-AVX2_INTOP_USING_SSE2(add_epi32)
-
-// Replace 256 bit operations with 128 bit ones when AVX2 is disabled
-#ifndef __AVX2__
-AVX2_INTOP_USING_SSE2(and_si128)
-AVX2_INTOP_USING_SSE2(andnot_si128)
-#endif
-
-/* natural logarithm computed for 8 simultaneous float
-   return NaN for x <= 0
-*/
-static NCNN_FORCEINLINE __m256 log256_ps(__m256 x)
-{
-    __m256i imm0;
-    __m256 one = *(__m256*)_ps256_1;
-
-    //__m256 invalid_mask = _mm256_cmple_ps(x, _mm256_setzero_ps());
-    __m256 invalid_mask = _mm256_cmp_ps(x, _mm256_setzero_ps(), _CMP_LE_OS);
-
-    x = _mm256_max_ps(x, *(__m256*)_ps256_min_norm_pos); /* cut off denormalized stuff */
-
-    // can be done with AVX2
-    imm0 = _mm256_comp_srli_epi32(_mm256_castps_si256(x), 23);
-
-    /* keep only the fractional part */
-    x = _mm256_and_ps(x, *(__m256*)_ps256_inv_mant_mask);
-    x = _mm256_or_ps(x, *(__m256*)_ps256_0p5);
-
-    // this is again another AVX2 instruction
-    imm0 = _mm256_comp_sub_epi32(imm0, *(__m256i*)_pi32_256_0x7f);
-    __m256 e = _mm256_cvtepi32_ps(imm0);
-
-    e = _mm256_add_ps(e, one);
-
-    /* part2:
-       if( x < SQRTHF ) {
-         e -= 1;
-         x = x + x - 1.0;
-       } else { x = x - 1.0; }
-    */
-    //__m256 mask = _mm256_cmplt_ps(x, *(__m256*)_ps256_cephes_SQRTHF);
-    __m256 mask = _mm256_cmp_ps(x, *(__m256*)_ps256_cephes_SQRTHF, _CMP_LT_OS);
-    __m256 tmp = _mm256_and_ps(x, mask);
-    x = _mm256_sub_ps(x, one);
-    e = _mm256_sub_ps(e, _mm256_and_ps(one, mask));
-    x = _mm256_add_ps(x, tmp);
-
-    __m256 z = _mm256_mul_ps(x, x);
-
-    __m256 y = *(__m256*)_ps256_cephes_log_p0;
-    y = _mm256_comp_fmadd_ps(y, x, *(__m256*)_ps256_cephes_log_p1);
-    y = _mm256_comp_fmadd_ps(y, x, *(__m256*)_ps256_cephes_log_p2);
-    y = _mm256_comp_fmadd_ps(y, x, *(__m256*)_ps256_cephes_log_p3);
-    y = _mm256_comp_fmadd_ps(y, x, *(__m256*)_ps256_cephes_log_p4);
-    y = _mm256_comp_fmadd_ps(y, x, *(__m256*)_ps256_cephes_log_p5);
-    y = _mm256_comp_fmadd_ps(y, x, *(__m256*)_ps256_cephes_log_p6);
-    y = _mm256_comp_fmadd_ps(y, x, *(__m256*)_ps256_cephes_log_p7);
-    y = _mm256_comp_fmadd_ps(y, x, *(__m256*)_ps256_cephes_log_p8);
-    y = _mm256_mul_ps(y, x);
-
-    y = _mm256_mul_ps(y, z);
-
-    y = _mm256_comp_fmadd_ps(e, *(__m256*)_ps256_cephes_log_q1, y);
-
-    //y = -z * 0.5 + y
-    y = _mm256_comp_fnmadd_ps(z, *(__m256*)_ps256_0p5, y);
-
-    x = _mm256_add_ps(x, y);
-    x = _mm256_comp_fmadd_ps(e, *(__m256*)_ps256_cephes_log_q2, x);
-    y = _mm256_or_ps(x, invalid_mask); // negative arg will be NAN
-    return y;
-}
-
-_PS256_CONST(exp_hi, 88.3762626647949f);
-_PS256_CONST(exp_lo, -88.3762626647949f);
-
-_PS256_CONST(cephes_LOG2EF, 1.44269504088896341f);
-_PS256_CONST(cephes_exp_C1, 0.693359375f);
-_PS256_CONST(cephes_exp_C2, -2.12194440e-4f);
-
-_PS256_CONST(cephes_exp_p0, 1.9875691500E-4f);
-_PS256_CONST(cephes_exp_p1, 1.3981999507E-3f);
-_PS256_CONST(cephes_exp_p2, 8.3334519073E-3f);
-_PS256_CONST(cephes_exp_p3, 4.1665795894E-2f);
-_PS256_CONST(cephes_exp_p4, 1.6666665459E-1f);
-_PS256_CONST(cephes_exp_p5, 5.0000001201E-1f);
-
-static NCNN_FORCEINLINE __m256 exp256_ps(__m256 x)
-{
-    __m256 tmp = _mm256_setzero_ps(), fx;
-    __m256i imm0;
-    __m256 one = *(__m256*)_ps256_1;
-
-    x = _mm256_min_ps(x, *(__m256*)_ps256_exp_hi);
-    x = _mm256_max_ps(x, *(__m256*)_ps256_exp_lo);
-
-    /* express exp(x) as exp(g + n*log(2)) */
-    fx = _mm256_comp_fmadd_ps(x, *(__m256*)_ps256_cephes_LOG2EF, *(__m256*)_ps256_0p5);
-
-    /* how to perform a floorf with SSE: just below */
-    //imm0 = _mm256_cvttps_epi32(fx);
-    //tmp  = _mm256_cvtepi32_ps(imm0);
-
-    tmp = _mm256_floor_ps(fx);
-
-    /* if greater, subtract 1 */
-    //__m256 mask = _mm256_cmpgt_ps(tmp, fx);
-    __m256 mask = _mm256_cmp_ps(tmp, fx, _CMP_GT_OS);
-    mask = _mm256_and_ps(mask, one);
-    fx = _mm256_sub_ps(tmp, mask);
-
-    // x = x - fx * exp_C1
-    x = _mm256_comp_fnmadd_ps(fx, *(__m256*)_ps256_cephes_exp_C1, x);
-    // x = x - fx * exp_C2
-    x = _mm256_comp_fnmadd_ps(fx, *(__m256*)_ps256_cephes_exp_C2, x);
-
-    tmp = _mm256_mul_ps(x, x);
-
-    __m256 y = *(__m256*)_ps256_cephes_exp_p0;
-    y = _mm256_comp_fmadd_ps(y, x, *(__m256*)_ps256_cephes_exp_p1);
-    y = _mm256_comp_fmadd_ps(y, x, *(__m256*)_ps256_cephes_exp_p2);
-    y = _mm256_comp_fmadd_ps(y, x, *(__m256*)_ps256_cephes_exp_p3);
-    y = _mm256_comp_fmadd_ps(y, x, *(__m256*)_ps256_cephes_exp_p4);
-    y = _mm256_comp_fmadd_ps(y, x, *(__m256*)_ps256_cephes_exp_p5);
-    y = _mm256_comp_fmadd_ps(y, tmp, x);
-    y = _mm256_add_ps(y, one);
-
-    /* build 2^n */
-    imm0 = _mm256_cvttps_epi32(fx);
-    // another two AVX2 instructions
-    imm0 = _mm256_comp_add_epi32(imm0, *(__m256i*)_pi32_256_0x7f);
-    imm0 = _mm256_comp_slli_epi32(imm0, 23);
-    __m256 pow2n = _mm256_castsi256_ps(imm0);
-    y = _mm256_mul_ps(y, pow2n);
-    return y;
-}
-
-_PS256_CONST(tanh_hi, 9.0f);
-_PS256_CONST(tanh_lo, -9.0f);
-
-_PS256_CONST(cephes_tanh_p0, -2.76076847742355E-16f);
-_PS256_CONST(cephes_tanh_p1, 2.00018790482477E-13f);
-_PS256_CONST(cephes_tanh_p2, -8.60467152213735E-11f);
-_PS256_CONST(cephes_tanh_p3, 5.12229709037114E-08f);
-_PS256_CONST(cephes_tanh_p4, 1.48572235717979E-05f);
-_PS256_CONST(cephes_tanh_p5, 6.37261928875436E-04f);
-_PS256_CONST(cephes_tanh_p6, 4.89352455891786E-03f);
-
-_PS256_CONST(cephes_tanh_p7, 1.19825839466702e-06f);
-_PS256_CONST(cephes_tanh_p8, 1.18534705686654e-04f);
-_PS256_CONST(cephes_tanh_p9, 2.26843463243900e-03f);
-
-// an approximation of tanh
-static inline __m256 tanh256_ps(const __m256 x)
-{
-    __m256 value = x;
-    value = _mm256_max_ps(*(__m256*)_ps256_tanh_lo, value);
-    value = _mm256_min_ps(*(__m256*)_ps256_tanh_hi, value);
-
-    __m256 value_squared = _mm256_mul_ps(value, value);
-
-    __m256 p;
-    p = _mm256_comp_fmadd_ps(value_squared, *(__m256*)_ps256_cephes_tanh_p0, *(__m256*)_ps256_cephes_tanh_p1);
-    p = _mm256_comp_fmadd_ps(p, value_squared, *(__m256*)_ps256_cephes_tanh_p2);
-    p = _mm256_comp_fmadd_ps(p, value_squared, *(__m256*)_ps256_cephes_tanh_p3);
-    p = _mm256_comp_fmadd_ps(p, value_squared, *(__m256*)_ps256_cephes_tanh_p4);
-    p = _mm256_comp_fmadd_ps(p, value_squared, *(__m256*)_ps256_cephes_tanh_p5);
-    p = _mm256_comp_fmadd_ps(p, value_squared, *(__m256*)_ps256_cephes_tanh_p6);
-    p = _mm256_mul_ps(p, value);
-
-    __m256 q;
-    q = _mm256_comp_fmadd_ps(value_squared, *(__m256*)_ps256_cephes_tanh_p7, *(__m256*)_ps256_cephes_tanh_p8);
-    q = _mm256_comp_fmadd_ps(q, value_squared, *(__m256*)_ps256_cephes_tanh_p9);
-    q = _mm256_comp_fmadd_ps(q, value_squared, *(__m256*)_ps256_cephes_tanh_p6);
-
-    __m256 dst = _mm256_div_ps(p, q);
-    return dst;
-}
-
-_PS256_CONST(minus_cephes_DP1, -0.78515625f);
-_PS256_CONST(minus_cephes_DP2, -2.4187564849853515625e-4f);
-_PS256_CONST(minus_cephes_DP3, -3.77489497744594108e-8f);
-_PS256_CONST(sincof_p0, -1.9515295891E-4f);
-_PS256_CONST(sincof_p1, 8.3321608736E-3f);
-_PS256_CONST(sincof_p2, -1.6666654611E-1f);
-_PS256_CONST(coscof_p0, 2.443315711809948E-005f);
-_PS256_CONST(coscof_p1, -1.388731625493765E-003f);
-_PS256_CONST(coscof_p2, 4.166664568298827E-002f);
-_PS256_CONST(cephes_FOPI, 1.27323954473516f); // 4 / M_PI
-
-/* evaluation of 8 sines at onces using AVX intrisics
-
-   The code is the exact rewriting of the cephes sinf function.
-   Precision is excellent as long as x < 8192 (I did not bother to
-   take into account the special handling they have for greater values
-   -- it does not return garbage for arguments over 8192, though, but
-   the extra precision is missing).
-
-   Note that it is such that sinf((float)M_PI) = 8.74e-8, which is the
-   surprising but correct result.
-
-*/
-static NCNN_FORCEINLINE __m256 sin256_ps(__m256 x)
-{   // any x
-    __m256 xmm1, xmm2 = _mm256_setzero_ps(), xmm3, sign_bit, y;
-    __m256i imm0, imm2;
-
-#ifndef __AVX2__
-    __m128i imm0_1, imm0_2;
-    __m128i imm2_1, imm2_2;
-#endif
-
-    sign_bit = x;
-    /* take the absolute value */
-    x = _mm256_and_ps(x, *(__m256*)_ps256_inv_sign_mask);
-    /* extract the sign bit (upper one) */
-    sign_bit = _mm256_and_ps(sign_bit, *(__m256*)_ps256_sign_mask);
-
-    /* scale by 4/Pi */
-    y = _mm256_mul_ps(x, *(__m256*)_ps256_cephes_FOPI);
-
-    /*
-      Here we start a series of integer operations, which are in the
-      realm of AVX2.
-      If we don't have AVX, let's perform them using SSE2 directives
-    */
-
-#ifdef __AVX2__
-    /* store the integer part of y in mm0 */
-    imm2 = _mm256_cvttps_epi32(y);
-    /* j=(j+1) & (~1) (see the cephes sources) */
-    // another two AVX2 instruction
-    imm2 = _mm256_comp_add_epi32(imm2, *(__m256i*)_pi32_256_1);
-    imm2 = _mm256_and_si256(imm2, *(__m256i*)_pi32_256_inv1);
-    y = _mm256_cvtepi32_ps(imm2);
-
-    /* get the swap sign flag */
-    imm0 = _mm256_and_si256(imm2, *(__m256i*)_pi32_256_4);
-    imm0 = _mm256_comp_slli_epi32(imm0, 29);
-    /* get the polynom selection mask
-       there is one polynom for 0 <= x <= Pi/4
-       and another one for Pi/4<x<=Pi/2
-
-       Both branches will be computed.
-    */
-    imm2 = _mm256_and_si256(imm2, *(__m256i*)_pi32_256_2);
-    imm2 = _mm256_cmpeq_epi32(imm2, *(__m256i*)_pi32_256_0);
-#else
-    /* we use SSE2 routines to perform the integer ops */
-    COPY_IMM_TO_XMM(_mm256_cvttps_epi32(y), imm2_1, imm2_2);
-
-    imm2_1 = _mm_add_epi32(imm2_1, *(__m128i*)_pi32avx_1);
-    imm2_2 = _mm_add_epi32(imm2_2, *(__m128i*)_pi32avx_1);
-
-    imm2_1 = _mm_and_si128(imm2_1, *(__m128i*)_pi32avx_inv1);
-    imm2_2 = _mm_and_si128(imm2_2, *(__m128i*)_pi32avx_inv1);
-
-    COPY_XMM_TO_IMM(imm2_1, imm2_2, imm2);
-    y = _mm256_cvtepi32_ps(imm2);
-
-    imm0_1 = _mm_and_si128(imm2_1, *(__m128i*)_pi32avx_4);
-    imm0_2 = _mm_and_si128(imm2_2, *(__m128i*)_pi32avx_4);
-
-    imm0_1 = _mm_slli_epi32(imm0_1, 29);
-    imm0_2 = _mm_slli_epi32(imm0_2, 29);
-
-    COPY_XMM_TO_IMM(imm0_1, imm0_2, imm0);
-
-    imm2_1 = _mm_and_si128(imm2_1, *(__m128i*)_pi32avx_2);
-    imm2_2 = _mm_and_si128(imm2_2, *(__m128i*)_pi32avx_2);
-
-    imm2_1 = _mm_cmpeq_epi32(imm2_1, _mm_setzero_si128());
-    imm2_2 = _mm_cmpeq_epi32(imm2_2, _mm_setzero_si128());
-
-    COPY_XMM_TO_IMM(imm2_1, imm2_2, imm2);
-#endif
-
-    __m256 swap_sign_bit = _mm256_castsi256_ps(imm0);
-    __m256 poly_mask = _mm256_castsi256_ps(imm2);
-    sign_bit = _mm256_xor_ps(sign_bit, swap_sign_bit);
-
-    /* The magic pass: "Extended precision modular arithmetic"
-       x = ((x - y * DP1) - y * DP2) - y * DP3; */
-    xmm1 = *(__m256*)_ps256_minus_cephes_DP1;
-    xmm2 = *(__m256*)_ps256_minus_cephes_DP2;
-    xmm3 = *(__m256*)_ps256_minus_cephes_DP3;
-    x = _mm256_comp_fmadd_ps(y, xmm1, x);
-    x = _mm256_comp_fmadd_ps(y, xmm2, x);
-    x = _mm256_comp_fmadd_ps(y, xmm3, x);
-
-    /* Evaluate the first polynom  (0 <= x <= Pi/4) */
-    y = *(__m256*)_ps256_coscof_p0;
-    __m256 z = _mm256_mul_ps(x, x);
-
-    y = _mm256_comp_fmadd_ps(y, z, *(__m256*)_ps256_coscof_p1);
-    y = _mm256_comp_fmadd_ps(y, z, *(__m256*)_ps256_coscof_p2);
-    y = _mm256_mul_ps(y, z);
-    y = _mm256_mul_ps(y, z);
-    // y = y - z * 0.5
-    y = _mm256_comp_fnmadd_ps(z, *(__m256*)_ps256_0p5, y);
-    y = _mm256_add_ps(y, *(__m256*)_ps256_1);
-
-    /* Evaluate the second polynom  (Pi/4 <= x <= 0) */
-
-    __m256 y2 = *(__m256*)_ps256_sincof_p0;
-    y2 = _mm256_comp_fmadd_ps(y2, z, *(__m256*)_ps256_sincof_p1);
-    y2 = _mm256_comp_fmadd_ps(y2, z, *(__m256*)_ps256_sincof_p2);
-    y2 = _mm256_mul_ps(y2, z);
-    y2 = _mm256_comp_fmadd_ps(y2, x, x);
-
-    /* select the correct result from the two polynoms */
-    xmm3 = poly_mask;
-    y2 = _mm256_and_ps(xmm3, y2); //, xmm3);
-    y = _mm256_andnot_ps(xmm3, y);
-    y = _mm256_add_ps(y, y2);
-    /* update the sign */
-    y = _mm256_xor_ps(y, sign_bit);
-
-    return y;
-}
-
-/* almost the same as sin_ps */
-static NCNN_FORCEINLINE __m256 cos256_ps(__m256 x)
-{   // any x
-    __m256 xmm1, xmm2 = _mm256_setzero_ps(), xmm3, y;
-    __m256i imm0, imm2;
-
-#ifndef __AVX2__
-    __m128i imm0_1, imm0_2;
-    __m128i imm2_1, imm2_2;
-#endif
-
-    /* take the absolute value */
-    x = _mm256_and_ps(x, *(__m256*)_ps256_inv_sign_mask);
-
-    /* scale by 4/Pi */
-    y = _mm256_mul_ps(x, *(__m256*)_ps256_cephes_FOPI);
-
-#ifdef __AVX2__
-    /* store the integer part of y in mm0 */
-    imm2 = _mm256_cvttps_epi32(y);
-    /* j=(j+1) & (~1) (see the cephes sources) */
-    imm2 = _mm256_comp_add_epi32(imm2, *(__m256i*)_pi32_256_1);
-    imm2 = _mm256_and_si256(imm2, *(__m256i*)_pi32_256_inv1);
-    y = _mm256_cvtepi32_ps(imm2);
-    imm2 = _mm256_comp_sub_epi32(imm2, *(__m256i*)_pi32_256_2);
-
-    /* get the swap sign flag */
-    imm0 = _mm256_andnot_si256(imm2, *(__m256i*)_pi32_256_4);
-    imm0 = _mm256_comp_slli_epi32(imm0, 29);
-    /* get the polynom selection mask */
-    imm2 = _mm256_and_si256(imm2, *(__m256i*)_pi32_256_2);
-    imm2 = _mm256_cmpeq_epi32(imm2, *(__m256i*)_pi32_256_0);
-#else
-
-    /* we use SSE2 routines to perform the integer ops */
-    COPY_IMM_TO_XMM(_mm256_cvttps_epi32(y), imm2_1, imm2_2);
-
-    imm2_1 = _mm_add_epi32(imm2_1, *(__m128i*)_pi32avx_1);
-    imm2_2 = _mm_add_epi32(imm2_2, *(__m128i*)_pi32avx_1);
-
-    imm2_1 = _mm_and_si128(imm2_1, *(__m128i*)_pi32avx_inv1);
-    imm2_2 = _mm_and_si128(imm2_2, *(__m128i*)_pi32avx_inv1);
-
-    COPY_XMM_TO_IMM(imm2_1, imm2_2, imm2);
-    y = _mm256_cvtepi32_ps(imm2);
-
-    imm2_1 = _mm_sub_epi32(imm2_1, *(__m128i*)_pi32avx_2);
-    imm2_2 = _mm_sub_epi32(imm2_2, *(__m128i*)_pi32avx_2);
-
-    imm0_1 = _mm_andnot_si128(imm2_1, *(__m128i*)_pi32avx_4);
-    imm0_2 = _mm_andnot_si128(imm2_2, *(__m128i*)_pi32avx_4);
-
-    imm0_1 = _mm_slli_epi32(imm0_1, 29);
-    imm0_2 = _mm_slli_epi32(imm0_2, 29);
-
-    COPY_XMM_TO_IMM(imm0_1, imm0_2, imm0);
-
-    imm2_1 = _mm_and_si128(imm2_1, *(__m128i*)_pi32avx_2);
-    imm2_2 = _mm_and_si128(imm2_2, *(__m128i*)_pi32avx_2);
-
-    imm2_1 = _mm_cmpeq_epi32(imm2_1, _mm_setzero_si128());
-    imm2_2 = _mm_cmpeq_epi32(imm2_2, _mm_setzero_si128());
-
-    COPY_XMM_TO_IMM(imm2_1, imm2_2, imm2);
-#endif
-
-    __m256 sign_bit = _mm256_castsi256_ps(imm0);
-    __m256 poly_mask = _mm256_castsi256_ps(imm2);
-
-    /* The magic pass: "Extended precision modular arithmetic"
-       x = ((x - y * DP1) - y * DP2) - y * DP3; */
-    xmm1 = *(__m256*)_ps256_minus_cephes_DP1;
-    xmm2 = *(__m256*)_ps256_minus_cephes_DP2;
-    xmm3 = *(__m256*)_ps256_minus_cephes_DP3;
-    x = _mm256_comp_fmadd_ps(y, xmm1, x);
-    x = _mm256_comp_fmadd_ps(y, xmm2, x);
-    x = _mm256_comp_fmadd_ps(y, xmm3, x);
-
-    /* Evaluate the first polynom  (0 <= x <= Pi/4) */
-    y = *(__m256*)_ps256_coscof_p0;
-    __m256 z = _mm256_mul_ps(x, x);
-
-    y = _mm256_comp_fmadd_ps(y, z, *(__m256*)_ps256_coscof_p1);
-    y = _mm256_comp_fmadd_ps(y, z, *(__m256*)_ps256_coscof_p2);
-    y = _mm256_mul_ps(y, z);
-    y = _mm256_mul_ps(y, z);
-    // y = y - z * 0.5
-    y = _mm256_comp_fnmadd_ps(z, *(__m256*)_ps256_0p5, y);
-    y = _mm256_add_ps(y, *(__m256*)_ps256_1);
-
-    /* Evaluate the second polynom  (Pi/4 <= x <= 0) */
-
-    __m256 y2 = *(__m256*)_ps256_sincof_p0;
-    y2 = _mm256_mul_ps(y2, z);
-    y2 = _mm256_add_ps(y2, *(__m256*)_ps256_sincof_p1);
-    y2 = _mm256_mul_ps(y2, z);
-    y2 = _mm256_add_ps(y2, *(__m256*)_ps256_sincof_p2);
-    y2 = _mm256_mul_ps(y2, z);
-    y2 = _mm256_mul_ps(y2, x);
-    y2 = _mm256_add_ps(y2, x);
-
-    /* select the correct result from the two polynoms */
-    xmm3 = poly_mask;
-    y2 = _mm256_and_ps(xmm3, y2); //, xmm3);
-    y = _mm256_andnot_ps(xmm3, y);
-    y = _mm256_add_ps(y, y2);
-    /* update the sign */
-    y = _mm256_xor_ps(y, sign_bit);
-
-    return y;
-}
-
-/* since sin256_ps and cos256_ps are almost identical, sincos256_ps could replace both of them..
-   it is almost as fast, and gives you a free cosine with your sine */
-static NCNN_FORCEINLINE void sincos256_ps(__m256 x, __m256* s, __m256* c)
-{
-    __m256 xmm1, xmm2, xmm3 = _mm256_setzero_ps(), sign_bit_sin, y;
-    __m256i imm0, imm2, imm4;
-
-#ifndef __AVX2__
-    __m128i imm0_1, imm0_2;
-    __m128i imm2_1, imm2_2;
-    __m128i imm4_1, imm4_2;
-#endif
-
-    sign_bit_sin = x;
-    /* take the absolute value */
-    x = _mm256_and_ps(x, *(__m256*)_ps256_inv_sign_mask);
-    /* extract the sign bit (upper one) */
-    sign_bit_sin = _mm256_and_ps(sign_bit_sin, *(__m256*)_ps256_sign_mask);
-
-    /* scale by 4/Pi */
-    y = _mm256_mul_ps(x, *(__m256*)_ps256_cephes_FOPI);
-
-#ifdef __AVX2__
-    /* store the integer part of y in imm2 */
-    imm2 = _mm256_cvttps_epi32(y);
-
-    /* j=(j+1) & (~1) (see the cephes sources) */
-    imm2 = _mm256_comp_add_epi32(imm2, *(__m256i*)_pi32_256_1);
-    imm2 = _mm256_and_si256(imm2, *(__m256i*)_pi32_256_inv1);
-
-    y = _mm256_cvtepi32_ps(imm2);
-    imm4 = imm2;
-
-    /* get the swap sign flag for the sine */
-    imm0 = _mm256_and_si256(imm2, *(__m256i*)_pi32_256_4);
-    imm0 = _mm256_comp_slli_epi32(imm0, 29);
-    //__m256 swap_sign_bit_sin = _mm256_castsi256_ps(imm0);
-
-    /* get the polynom selection mask for the sine*/
-    imm2 = _mm256_and_si256(imm2, *(__m256i*)_pi32_256_2);
-    imm2 = _mm256_cmpeq_epi32(imm2, *(__m256i*)_pi32_256_0);
-    //__m256 poly_mask = _mm256_castsi256_ps(imm2);
-#else
-    /* we use SSE2 routines to perform the integer ops */
-    COPY_IMM_TO_XMM(_mm256_cvttps_epi32(y), imm2_1, imm2_2);
-
-    imm2_1 = _mm_add_epi32(imm2_1, *(__m128i*)_pi32avx_1);
-    imm2_2 = _mm_add_epi32(imm2_2, *(__m128i*)_pi32avx_1);
-
-    imm2_1 = _mm_and_si128(imm2_1, *(__m128i*)_pi32avx_inv1);
-    imm2_2 = _mm_and_si128(imm2_2, *(__m128i*)_pi32avx_inv1);
-
-    COPY_XMM_TO_IMM(imm2_1, imm2_2, imm2);
-    y = _mm256_cvtepi32_ps(imm2);
-
-    imm4_1 = imm2_1;
-    imm4_2 = imm2_2;
-
-    imm0_1 = _mm_and_si128(imm2_1, *(__m128i*)_pi32avx_4);
-    imm0_2 = _mm_and_si128(imm2_2, *(__m128i*)_pi32avx_4);
-
-    imm0_1 = _mm_slli_epi32(imm0_1, 29);
-    imm0_2 = _mm_slli_epi32(imm0_2, 29);
-
-    COPY_XMM_TO_IMM(imm0_1, imm0_2, imm0);
-
-    imm2_1 = _mm_and_si128(imm2_1, *(__m128i*)_pi32avx_2);
-    imm2_2 = _mm_and_si128(imm2_2, *(__m128i*)_pi32avx_2);
-
-    imm2_1 = _mm_cmpeq_epi32(imm2_1, _mm_setzero_si128());
-    imm2_2 = _mm_cmpeq_epi32(imm2_2, _mm_setzero_si128());
-
-    COPY_XMM_TO_IMM(imm2_1, imm2_2, imm2);
-#endif
-    __m256 swap_sign_bit_sin = _mm256_castsi256_ps(imm0);
-    __m256 poly_mask = _mm256_castsi256_ps(imm2);
-
-    /* The magic pass: "Extended precision modular arithmetic"
-       x = ((x - y * DP1) - y * DP2) - y * DP3; */
-    xmm1 = *(__m256*)_ps256_minus_cephes_DP1;
-    xmm2 = *(__m256*)_ps256_minus_cephes_DP2;
-    xmm3 = *(__m256*)_ps256_minus_cephes_DP3;
-    x = _mm256_comp_fmadd_ps(y, xmm1, x);
-    x = _mm256_comp_fmadd_ps(y, xmm2, x);
-    x = _mm256_comp_fmadd_ps(y, xmm3, x);
-
-#ifdef __AVX2__
-    imm4 = _mm256_comp_sub_epi32(imm4, *(__m256i*)_pi32_256_2);
-    imm4 = _mm256_andnot_si256(imm4, *(__m256i*)_pi32_256_4);
-    imm4 = _mm256_comp_slli_epi32(imm4, 29);
-#else
-    imm4_1 = _mm_sub_epi32(imm4_1, *(__m128i*)_pi32avx_2);
-    imm4_2 = _mm_sub_epi32(imm4_2, *(__m128i*)_pi32avx_2);
-
-    imm4_1 = _mm_andnot_si128(imm4_1, *(__m128i*)_pi32avx_4);
-    imm4_2 = _mm_andnot_si128(imm4_2, *(__m128i*)_pi32avx_4);
-
-    imm4_1 = _mm_slli_epi32(imm4_1, 29);
-    imm4_2 = _mm_slli_epi32(imm4_2, 29);
-
-    COPY_XMM_TO_IMM(imm4_1, imm4_2, imm4);
-#endif
-
-    __m256 sign_bit_cos = _mm256_castsi256_ps(imm4);
-
-    sign_bit_sin = _mm256_xor_ps(sign_bit_sin, swap_sign_bit_sin);
-
-    /* Evaluate the first polynom  (0 <= x <= Pi/4) */
-    __m256 z = _mm256_mul_ps(x, x);
-    y = *(__m256*)_ps256_coscof_p0;
-
-    y = _mm256_comp_fmadd_ps(y, z, *(__m256*)_ps256_coscof_p1);
-    y = _mm256_comp_fmadd_ps(y, z, *(__m256*)_ps256_coscof_p2);
-    y = _mm256_mul_ps(y, z);
-    y = _mm256_mul_ps(y, z);
-    // y = y - z * 0.5
-    y = _mm256_comp_fnmadd_ps(z, *(__m256*)_ps256_0p5, y);
-    y = _mm256_add_ps(y, *(__m256*)_ps256_1);
-
-    /* Evaluate the second polynom  (Pi/4 <= x <= 0) */
-
-    __m256 y2 = *(__m256*)_ps256_sincof_p0;
-    y2 = _mm256_comp_fmadd_ps(y2, z, *(__m256*)_ps256_sincof_p1);
-    y2 = _mm256_comp_fmadd_ps(y2, z, *(__m256*)_ps256_sincof_p2);
-    y2 = _mm256_mul_ps(y2, z);
-    y2 = _mm256_comp_fmadd_ps(y2, x, x);
-
-    /* select the correct result from the two polynoms */
-    xmm3 = poly_mask;
-    __m256 ysin2 = _mm256_and_ps(xmm3, y2);
-    __m256 ysin1 = _mm256_andnot_ps(xmm3, y);
-    y2 = _mm256_sub_ps(y2, ysin2);
-    y = _mm256_sub_ps(y, ysin1);
-
-    xmm1 = _mm256_add_ps(ysin1, ysin2);
-    xmm2 = _mm256_add_ps(y, y2);
-
-    /* update the sign */
-    *s = _mm256_xor_ps(xmm1, sign_bit_sin);
-    *c = _mm256_xor_ps(xmm2, sign_bit_cos);
-}
-
-static NCNN_FORCEINLINE __m256 tan256_ps(__m256 x)
-{
-    __m256 ysin, ycos;
-    __m256 eps = _mm256_set1_ps(1E-8f);
-    sincos256_ps(x, &ysin, &ycos);
-    __m256 mask = _mm256_cmp_ps(ycos, _mm256_setzero_ps(), _CMP_EQ_OS);
-    __m256 _tmp = _mm256_and_ps(eps, mask);
-    ycos = _mm256_add_ps(ycos, _tmp);
-    __m256 ytan = _mm256_div_ps(ysin, ycos);
-    return ytan;
-}
-
-static NCNN_FORCEINLINE __m256 pow256_ps(__m256 a, __m256 b)
-{
-    // pow(x, m) = exp(m * log(x))
-    return exp256_ps(_mm256_mul_ps(b, log256_ps(a)));
-}
-
-static NCNN_FORCEINLINE __m256 asin256_ps(__m256 x)
-{
-    const __m256 magic_negative_zero = _mm256_set1_ps(-0.0f);
-    const __m256 magic_half_one = _mm256_set1_ps(0.5f);
-    const __m256 magic_one = _mm256_set1_ps(1.0f);
-    const __m256 magic_a4 = _mm256_set1_ps(0.023994016f);
-    const __m256 magic_a5 = _mm256_set1_ps(0.042417344f);
-    const __m256 magic_a2 = _mm256_set1_ps(0.07494697f);
-    const __m256 magic_a3 = _mm256_set1_ps(0.045520633f);
-    const __m256 magic_a0 = _mm256_set1_ps(1.0f);
-    const __m256 magic_a1 = _mm256_set1_ps(0.166667819f);
-    const __m256 magic_half_pi = _mm256_set1_ps(1.5707964f);
-    const __m256 magic_three = _mm256_set1_ps(3.0f);
-
-    // negative_mask = magic_negative_zero && x;
-    __m256 negative_mask = _mm256_and_ps(magic_negative_zero, x);
-
-    // absolute = abs(x);
-    __m256 absolute = _mm256_andnot_ps(magic_negative_zero, x);
-
-    // Reference: https://en.wikipedia.org/wiki/Small-angle_approximation
-
-    // is_small_input = (absolute <= 0.5f);
-    __m256 is_small_input = _mm256_cmp_ps(absolute, magic_half_one, _CMP_LE_OQ);
-
-    // is_big_input = (is_small_input ? 0.0f : 1.0f);
-    __m256 is_big_input = _mm256_andnot_ps(is_small_input, magic_one);
-
-    // big_input_approx = sqrt(0.5f * (1 - absolute));
-    __m256 big_input_approx = _mm256_sqrt_ps(_mm256_mul_ps(
-                                  magic_half_one,
-                                  _mm256_sub_ps(magic_one, absolute)));
-
-    // input_approx = (is_small_input ? absolute : big_input_approx);
-    __m256 input_approx = _mm256_or_ps(
-                              _mm256_and_ps(is_small_input, absolute),
-                              _mm256_andnot_ps(is_small_input, big_input_approx));
-
-    // square_of_input_approx = input_approx * input_approx;
-    __m256 square_of_input_approx = _mm256_mul_ps(input_approx, input_approx);
-
-    // fourth_power_of_input_approx =
-    //     square_of_input_approx * square_of_input_approx;
-    __m256 fourth_power_of_input_approx = _mm256_mul_ps(
-            square_of_input_approx, square_of_input_approx);
-
-    // TODO: Need more explanations.
-    // x1 = ((fourth_power_of_input_approx * magic_a4) + magic_a2);
-    // x2 = ((fourth_power_of_input_approx * magic_a5) + magic_a3);
-    // x3 = ((fourth_power_of_input_approx * x1) + magic_a0);
-    // x4 = ((fourth_power_of_input_approx * x2) + magic_a1);
-    // output_approx = ((square_of_input_approx * x4) + x3);
-    __m256 output_approx = _mm256_comp_fmadd_ps(
-                               square_of_input_approx,
-                               _mm256_comp_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm256_comp_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       magic_a5,
-                                       magic_a3),
-                                   magic_a1),
-                               _mm256_comp_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm256_comp_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       magic_a4,
-                                       magic_a2),
-                                   magic_a0));
-
-    // TODO: Need more explanations.
-    // x1 = ((0.5 * PI) * is_big_input);
-    // x2 = (output_approx * input_approx);
-    // x3 = (-(3.0f * is_big_input) + 1.0f);
-    // final_approx = ((x2 * x3) + x1);
-    __m256 final_approx = _mm256_comp_fmadd_ps(
-                              _mm256_mul_ps(output_approx, input_approx),
-                              _mm256_comp_fnmadd_ps(magic_three, is_big_input, magic_one),
-                              _mm256_mul_ps(magic_half_pi, is_big_input));
-
-    // return (final_approx || negative_mask);
-    return _mm256_or_ps(final_approx, negative_mask);
-}
-
-static NCNN_FORCEINLINE __m256 acos256_ps(__m256 x)
-{
-    const __m256 magic_negative_zero = _mm256_set1_ps(-0.0f);
-    const __m256 magic_zero = _mm256_set1_ps(0.0f);
-    const __m256 magic_half_one = _mm256_set1_ps(0.5f);
-    const __m256 magic_one = _mm256_set1_ps(1.0f);
-    const __m256 magic_a4 = _mm256_set1_ps(0.023994016f);
-    const __m256 magic_a5 = _mm256_set1_ps(0.042417344f);
-    const __m256 magic_a2 = _mm256_set1_ps(0.07494697f);
-    const __m256 magic_a3 = _mm256_set1_ps(0.045520633f);
-    const __m256 magic_a0 = _mm256_set1_ps(1.0f);
-    const __m256 magic_a1 = _mm256_set1_ps(0.166667819f);
-    const __m256 magic_half_pi = _mm256_set1_ps(1.5707964f);
-    const __m256 magic_pi = _mm256_set1_ps(3.1415927f);
-
-    // negative_mask = magic_negative_zero && x;
-    __m256 negative_mask = _mm256_and_ps(magic_negative_zero, x);
-
-    // absolute = abs(x);
-    __m256 absolute = _mm256_andnot_ps(magic_negative_zero, x);
-
-    // Reference: https://en.wikipedia.org/wiki/Small-angle_approximation
-
-    // is_small_input = (absolute <= 0.5f);
-    __m256 is_small_input = _mm256_cmp_ps(absolute, magic_half_one, _CMP_LE_OQ);
-
-    // big_input_approx = sqrt(0.5f * (1 - absolute));
-    __m256 big_input_approx = _mm256_sqrt_ps(_mm256_mul_ps(
-                                  magic_half_one,
-                                  _mm256_sub_ps(magic_one, absolute)));
-
-    // input_approx = (is_small_input ? absolute : big_input_approx);
-    __m256 input_approx = _mm256_or_ps(
-                              _mm256_and_ps(is_small_input, absolute),
-                              _mm256_andnot_ps(is_small_input, big_input_approx));
-
-    // square_of_input_approx = input_approx * input_approx;
-    __m256 square_of_input_approx = _mm256_mul_ps(input_approx, input_approx);
-
-    // fourth_power_of_input_approx =
-    //     square_of_input_approx * square_of_input_approx;
-    __m256 fourth_power_of_input_approx = _mm256_mul_ps(
-            square_of_input_approx, square_of_input_approx);
-
-    // TODO: Need more explanations.
-    // x1 = ((fourth_power_of_input_approx * magic_a4) + magic_a2);
-    // x2 = ((fourth_power_of_input_approx * magic_a5) + magic_a3);
-    // x3 = ((fourth_power_of_input_approx * x1) + magic_a0);
-    // x4 = ((fourth_power_of_input_approx * x2) + magic_a1);
-    // output_approx = ((square_of_input_approx * x4) + x3);
-    __m256 output_approx = _mm256_comp_fmadd_ps(
-                               square_of_input_approx,
-                               _mm256_comp_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm256_comp_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       magic_a5,
-                                       magic_a3),
-                                   magic_a1),
-                               _mm256_comp_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm256_comp_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       magic_a4,
-                                       magic_a2),
-                                   magic_a0));
-
-    // TODO: Need more explanations.
-    // x1 = (output_approx * input_approx);
-    __m256 x1 = _mm256_mul_ps(output_approx, input_approx);
-
-    // TODO: Need more explanations.
-    // small_final_approx = ((0.5 * PI) - (x1 | negative_mask));
-    __m256 small_final_approx = _mm256_sub_ps(
-                                    magic_half_pi,
-                                    _mm256_or_ps(x1, negative_mask));
-
-    // TODO: Need more explanations.
-    // big_final_approx = (((x < 0.0f) & PI) + ((x1 * 2) | negative_mask));
-    __m256 big_final_approx = _mm256_add_ps(
-                                  _mm256_and_ps(_mm256_cmp_ps(x, magic_zero, _CMP_LT_OQ), magic_pi),
-                                  _mm256_or_ps(_mm256_add_ps(x1, x1), negative_mask));
-
-    // return (is_small_input ? small_final_approx : big_final_approx);
-    return _mm256_or_ps(
-               _mm256_and_ps(is_small_input, small_final_approx),
-               _mm256_andnot_ps(is_small_input, big_final_approx));
-}
-
-static NCNN_FORCEINLINE __m256 atan256_ps(__m256 x)
-{
-    const __m256 magic_negative_zero = _mm256_set1_ps(-0.0f);
-    const __m256 magic_one = _mm256_set1_ps(1.0f);
-    const __m256 magic_negative_one = _mm256_set1_ps(-1.0f);
-    const __m256 magic_half_pi = _mm256_set1_ps(1.5707964f);
-    const __m256 magic_a0 = _mm256_set1_ps(1.0f);
-    const __m256 magic_a1 = _mm256_set1_ps(-0.33333072f);
-    const __m256 magic_a2 = _mm256_set1_ps(0.1999262f);
-    const __m256 magic_a3 = _mm256_set1_ps(-0.14203644f);
-    const __m256 magic_a4 = _mm256_set1_ps(0.10640934f);
-    const __m256 magic_a5 = _mm256_set1_ps(-0.07504295f);
-    const __m256 magic_a6 = _mm256_set1_ps(0.04269152f);
-    const __m256 magic_a7 = _mm256_set1_ps(-0.01606863f);
-    const __m256 magic_a8 = _mm256_set1_ps(0.0028498897f);
-
-    // negative_mask = magic_negative_zero && x;
-    __m256 negative_mask = _mm256_and_ps(magic_negative_zero, x);
-
-    // absolute = abs(x);
-    __m256 absolute = _mm256_andnot_ps(magic_negative_zero, x);
-
-    // Reference: https://en.wikipedia.org/wiki/Small-angle_approximation
-
-    // is_small_input = (1.0f < absolute);
-    __m256 is_small_input = _mm256_cmp_ps(magic_one, absolute, _CMP_LT_OQ);
-
-    // x1 = (is_small_input ? -1.0f : absolute);
-    // x2 = (is_small_input ? absolute : 1.0f)
-    // input_approx = x1 / x2;
-    __m256 input_approx = _mm256_div_ps(
-                              _mm256_or_ps(
-                                  _mm256_and_ps(is_small_input, magic_negative_one),
-                                  _mm256_andnot_ps(is_small_input, absolute)),
-                              _mm256_or_ps(
-                                  _mm256_and_ps(is_small_input, absolute),
-                                  _mm256_andnot_ps(is_small_input, magic_one)));
-
-    // square_of_input_approx = input_approx * input_approx;
-    __m256 square_of_input_approx = _mm256_mul_ps(input_approx, input_approx);
-
-    // fourth_power_of_input_approx =
-    //     square_of_input_approx * square_of_input_approx;
-    __m256 fourth_power_of_input_approx = _mm256_mul_ps(
-            square_of_input_approx, square_of_input_approx);
-
-    // TODO: Need more explanations.
-    // x1 = ((fourth_power_of_input_approx * magic_a7) + magic_a5);
-    // x2 = ((fourth_power_of_input_approx * magic_a8) + magic_a6);
-    // x3 = ((fourth_power_of_input_approx * x1) + magic_a3);
-    // x4 = ((fourth_power_of_input_approx * x2) + magic_a4);
-    // x5 = ((fourth_power_of_input_approx * x3) + magic_a1);
-    // x6 = ((fourth_power_of_input_approx * x4) + magic_a2);
-    // x7 = ((fourth_power_of_input_approx * x6) + magic_a0);
-    // output_approx = ((square_of_input_approx * x5) + x7);
-    __m256 output_approx = _mm256_comp_fmadd_ps(
-                               square_of_input_approx,
-                               _mm256_comp_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm256_comp_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       _mm256_comp_fmadd_ps(
-                                           fourth_power_of_input_approx,
-                                           magic_a7,
-                                           magic_a5),
-                                       magic_a3),
-                                   magic_a1),
-                               _mm256_comp_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm256_comp_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       _mm256_comp_fmadd_ps(
-                                           fourth_power_of_input_approx,
-                                           _mm256_comp_fmadd_ps(
-                                                   fourth_power_of_input_approx,
-                                                   magic_a8,
-                                                   magic_a6),
-                                           magic_a4),
-                                       magic_a2),
-                                   magic_a0));
-
-    // TODO: Need more explanations.
-    // x1 = (output_approx * input_approx);
-    // if (is_small_input) x1 += (0.5 * PI);
-    // return (negative_mask ? -x1 : x1);
-    return _mm256_or_ps(
-               _mm256_add_ps(
-                   _mm256_mul_ps(output_approx, input_approx),
-                   _mm256_and_ps(is_small_input, magic_half_pi)),
-               negative_mask);
-}
-
-static NCNN_FORCEINLINE __m256 atan2256_ps(__m256 y, __m256 x)
-{
-    // Reference: https://mazzo.li/posts/vectorized-atan2.html
-
-    const __m256 magic_zero = _mm256_set1_ps(0.0f);
-    const __m256 magic_negative_zero = _mm256_set1_ps(-0.0f);
-    const __m256 magic_pi = _mm256_set1_ps(3.1415927f);
-    const __m256 magic_half_pi = _mm256_set1_ps(1.5707964f);
-
-    // not_equal_zero_x = (x != 0.0f);
-    __m256 not_equal_zero_x = _mm256_cmp_ps(x, magic_zero, _CMP_NEQ_OQ);
-
-    // not_equal_zero_y = (y != 0.0f);
-    __m256 not_equal_zero_y = _mm256_cmp_ps(y, magic_zero, _CMP_NEQ_OQ);
-
-    // normal_mode = ((x != 0.0f) & (y != 0.0f));
-    __m256 normal_mode = _mm256_and_ps(not_equal_zero_x, not_equal_zero_y);
-
-    // negative_mask_x = magic_negative_zero && x;
-    __m256 negative_mask_x = _mm256_and_ps(magic_negative_zero, x);
-
-    // negative_mask_y = magic_negative_zero && y;
-    __m256 negative_mask_y = _mm256_and_ps(magic_negative_zero, y);
-
-    // pi_additions = ((x < 0.0f) ? ((y < 0.0f) ? -PI : PI) : 0.0f);
-    __m256 pi_additions = _mm256_and_ps(
-                              _mm256_cmp_ps(x, magic_zero, _CMP_LT_OQ),
-                              _mm256_or_ps(
-                                  _mm256_and_ps(
-                                      _mm256_cmp_ps(y, magic_zero, _CMP_LT_OQ),
-                                      magic_negative_zero),
-                                  magic_pi));
-
-    // normal_result = (atan(y / x) + pi_additions);
-    __m256 normal_result = _mm256_add_ps(
-                               atan256_ps(_mm256_div_ps(y, x)),
-                               pi_additions);
-
-    // negative_mask_full_x = ((negative_mask_x | PI) < 0.0f);
-    __m256 negative_mask_full_x = _mm256_cmp_ps(
-                                      _mm256_or_ps(negative_mask_x, magic_pi),
-                                      magic_zero,
-                                      _CMP_LT_OQ);
-
-    // x1 = (negative_mask_y ? -(0.5 * PI) : (0.5 * PI));
-    // x2 = (negative_mask_full_x ? PI : 0.0f);
-    // special_result = ((y != 0.0f) ? x1 : x2);
-    __m256 special_result = _mm256_or_ps(
-                                _mm256_and_ps(
-                                    not_equal_zero_y,
-                                    _mm256_or_ps(negative_mask_y, magic_half_pi)),
-                                _mm256_andnot_ps(
-                                    not_equal_zero_y,
-                                    _mm256_or_ps(
-                                        _mm256_and_ps(negative_mask_full_x, magic_pi),
-                                        _mm256_andnot_ps(negative_mask_full_x, magic_zero))));
-
-    // return (normal_mode ? normal_result : special_result);
-    return _mm256_or_ps(
-               _mm256_and_ps(normal_mode, normal_result),
-               _mm256_andnot_ps(normal_mode, special_result));
-}
-
-static NCNN_FORCEINLINE __m256 abs256_ps(__m256 x)
-{
-    // Use negative zero as the sign bit mask.
-    const __m256 magic_negative_zero = _mm256_set1_ps(-0.0f);
-
-    // return (!magic_negative_zero && x);
-    return _mm256_andnot_ps(magic_negative_zero, x);
-}
-
-#endif // AVX_MATHFUN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/batchnorm_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/batchnorm_x86.h
deleted file mode 100644
index 7168332a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/batchnorm_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BATCHNORM_X86_H
-#define LAYER_BATCHNORM_X86_H
-
-#include "batchnorm.h"
-
-namespace ncnn {
-
-class BatchNorm_x86 : public BatchNorm
-{
-public:
-    BatchNorm_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BATCHNORM_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/bias_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/bias_x86.h
deleted file mode 100644
index ab8e30de..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/bias_x86.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BIAS_X86_H
-#define LAYER_BIAS_X86_H
-
-#include "bias.h"
-
-namespace ncnn {
-
-class Bias_x86 : public Bias
-{
-public:
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BIAS_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/binaryop_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/binaryop_x86.h
deleted file mode 100644
index cd3ff12a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/binaryop_x86.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BINARYOP_X86_H
-#define LAYER_BINARYOP_X86_H
-
-#include "binaryop.h"
-
-namespace ncnn {
-
-class BinaryOp_x86 : public BinaryOp
-{
-public:
-    BinaryOp_x86();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BINARYOP_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/bnll_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/bnll_x86.h
deleted file mode 100644
index b3fad45c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/bnll_x86.h
+++ /dev/null
@@ -1,33 +0,0 @@
-// JasonZhang892 is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 JasonZhang892 <zqhy_0929@163.com>. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_BNLL_X86_H
-#define LAYER_BNLL_X86_H
-
-#include "bnll.h"
-
-namespace ncnn {
-
-class BNLL_x86 : public BNLL
-{
-public:
-    BNLL_x86();
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-public:
-};
-
-} // namespace ncnn
-
-#endif // LAYER_BNLL_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/cast_bf16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/cast_bf16.h
deleted file mode 100644
index 15939b92..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/cast_bf16.h
+++ /dev/null
@@ -1,159 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if NCNN_RUNTIME_CPU && NCNN_AVX512BF16 && __AVX512F__ && !__AVX512BF16__
-void cast_fp32_to_bf16_sse_avx512bf16(const Mat& bottom_blob, Mat& top_blob, const Option& opt);
-void cast_bf16_to_fp32_sse_avx512bf16(const Mat& bottom_blob, Mat& top_blob, const Option& opt);
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-void cast_fp32_to_bf16_sse_avx2(const Mat& bottom_blob, Mat& top_blob, const Option& opt);
-void cast_bf16_to_fp32_sse_avx2(const Mat& bottom_blob, Mat& top_blob, const Option& opt);
-#endif
-
-static void cast_fp32_to_bf16_sse(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-#if NCNN_RUNTIME_CPU && NCNN_AVX512BF16 && __AVX512F__ && !__AVX512BF16__
-    if (ncnn::cpu_support_x86_avx512_bf16())
-    {
-        cast_fp32_to_bf16_sse_avx512bf16(bottom_blob, top_blob, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-    if (ncnn::cpu_support_x86_avx2())
-    {
-        cast_fp32_to_bf16_sse_avx2(bottom_blob, top_blob, opt);
-        return;
-    }
-#endif
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int d = bottom_blob.d;
-    const int channels = bottom_blob.c;
-    const int elempack = bottom_blob.elempack;
-
-    const int size = w * h * d * elempack;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* ptr = bottom_blob.channel(q);
-        unsigned short* outptr = top_blob.channel(q);
-
-        int i = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-        for (; i + 31 < size; i += 32)
-        {
-            _mm512_storeu_si512((__m512i*)outptr, float2bfloat_avx512(_mm512_loadu_ps(ptr), _mm512_loadu_ps(ptr + 16)));
-            ptr += 32;
-            outptr += 32;
-        }
-#endif // __AVX512F__
-        for (; i + 15 < size; i += 16)
-        {
-#if __AVX512F__
-            _mm256_storeu_si256((__m256i*)outptr, float2bfloat_avx512(_mm512_loadu_ps(ptr)));
-#else
-            _mm256_storeu_si256((__m256i*)outptr, float2bfloat_avx(_mm256_loadu_ps(ptr), _mm256_loadu_ps(ptr + 8)));
-#endif
-            ptr += 16;
-            outptr += 16;
-        }
-#endif // __AVX__
-        for (; i + 7 < size; i += 8)
-        {
-#if __AVX__
-            _mm_store_si128((__m128i*)outptr, float2bfloat_avx(_mm256_loadu_ps(ptr)));
-#else
-            _mm_store_si128((__m128i*)outptr, float2bfloat_sse(_mm_loadu_ps(ptr), _mm_loadu_ps(ptr + 4)));
-#endif
-            ptr += 8;
-            outptr += 8;
-        }
-#endif // __SSE2__
-        for (; i < size; i++)
-        {
-            *outptr++ = float32_to_bfloat16(*ptr++);
-        }
-    }
-}
-
-static void cast_bf16_to_fp32_sse(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-#if NCNN_AVX512BF16 && __AVX512F__ && !__AVX512BF16__
-    if (ncnn::cpu_support_x86_avx512_bf16())
-    {
-        cast_bf16_to_fp32_sse_avx512bf16(bottom_blob, top_blob, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-    if (ncnn::cpu_support_x86_avx2())
-    {
-        cast_bf16_to_fp32_sse_avx2(bottom_blob, top_blob, opt);
-        return;
-    }
-#endif
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int d = bottom_blob.d;
-    const int channels = bottom_blob.c;
-    const int elempack = bottom_blob.elempack;
-
-    const int size = w * h * d * elempack;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const unsigned short* ptr = bottom_blob.channel(q);
-        float* outptr = top_blob.channel(q);
-
-        int i = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-        for (; i + 15 < size; i += 16)
-        {
-            _mm512_storeu_ps(outptr, bfloat2float_avx512(_mm256_loadu_si256((const __m256i*)ptr)));
-            ptr += 16;
-            outptr += 16;
-        }
-#endif // __AVX512F__
-        for (; i + 7 < size; i += 8)
-        {
-            _mm256_storeu_ps(outptr, bfloat2float_avx(_mm_loadu_si128((const __m128i*)ptr)));
-            ptr += 8;
-            outptr += 8;
-        }
-#endif // __AVX__
-        for (; i + 3 < size; i += 4)
-        {
-            _mm_storeu_ps(outptr, bfloat2float_sse(_mm_loadl_epi64((const __m128i*)ptr)));
-            ptr += 4;
-            outptr += 4;
-        }
-#endif // __SSE2__
-        for (; i < size; i++)
-        {
-            *outptr++ = bfloat16_to_float32(*ptr++);
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/cast_fp16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/cast_fp16.h
deleted file mode 100644
index fc190265..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/cast_fp16.h
+++ /dev/null
@@ -1,138 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if NCNN_RUNTIME_CPU && NCNN_F16C && __AVX__ && !__F16C__
-void cast_fp32_to_fp16_sse_f16c(const Mat& bottom_blob, Mat& top_blob, const Option& opt);
-void cast_fp16_to_fp32_sse_f16c(const Mat& bottom_blob, Mat& top_blob, const Option& opt);
-#endif
-
-static void cast_fp32_to_fp16_sse(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-#if NCNN_RUNTIME_CPU && NCNN_F16C && __AVX__ && !__F16C__
-    if (ncnn::cpu_support_x86_f16c())
-    {
-        cast_fp32_to_fp16_sse_f16c(bottom_blob, top_blob, opt);
-        return;
-    }
-#endif
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int d = bottom_blob.d;
-    const int channels = bottom_blob.c;
-    const int elempack = bottom_blob.elempack;
-
-    const int size = w * h * d * elempack;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* ptr = bottom_blob.channel(q);
-        unsigned short* outptr = top_blob.channel(q);
-
-        int i = 0;
-#if __F16C__
-#if __AVX512F__
-        for (; i + 15 < size; i += 16)
-        {
-            __m512 _v_fp32 = _mm512_loadu_ps(ptr);
-            __m256i _v_fp16 = _mm512_cvtps_ph(_v_fp32, _MM_ROUND_NEAREST | _MM_FROUND_NO_EXC);
-            _mm256_storeu_si256((__m256i*)outptr, _v_fp16);
-            ptr += 16;
-            outptr += 16;
-        }
-#endif // __AVX512F__
-        for (; i + 7 < size; i += 8)
-        {
-            __m256 _v_fp32 = _mm256_loadu_ps(ptr);
-            __m128i _v_fp16 = _mm256_cvtps_ph(_v_fp32, _MM_ROUND_NEAREST | _MM_FROUND_NO_EXC);
-            _mm_storeu_si128((__m128i*)outptr, _v_fp16);
-            ptr += 8;
-            outptr += 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-            __m128 _v_fp32 = _mm_loadu_ps(ptr);
-            __m128i _v_fp16 = _mm_cvtps_ph(_v_fp32, _MM_ROUND_NEAREST | _MM_FROUND_NO_EXC);
-            _mm_storel_epi64((__m128i*)outptr, _v_fp16);
-            ptr += 4;
-            outptr += 4;
-        }
-#endif // __F16C__
-        for (; i < size; i++)
-        {
-            *outptr++ = float32_to_float16(*ptr++);
-        }
-    }
-}
-
-static void cast_fp16_to_fp32_sse(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-#if NCNN_F16C && __AVX__ && !__F16C__
-    if (ncnn::cpu_support_x86_f16c())
-    {
-        cast_fp16_to_fp32_sse_f16c(bottom_blob, top_blob, opt);
-        return;
-    }
-#endif
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int d = bottom_blob.d;
-    const int channels = bottom_blob.c;
-    const int elempack = bottom_blob.elempack;
-
-    const int size = w * h * d * elempack;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const unsigned short* ptr = bottom_blob.channel(q);
-        float* outptr = top_blob.channel(q);
-
-        int i = 0;
-#if __F16C__
-#if __AVX512F__
-        for (; i + 15 < size; i += 16)
-        {
-            __m256i _v_fp16 = _mm256_loadu_si256((const __m256i*)ptr);
-            __m512 _v_fp32 = _mm512_cvtph_ps(_v_fp16);
-            _mm512_storeu_ps(outptr, _v_fp32);
-            ptr += 16;
-            outptr += 16;
-        }
-#endif // __AVX512F__
-        for (; i + 7 < size; i += 8)
-        {
-            __m128i _v_fp16 = _mm_loadu_si128((const __m128i*)ptr);
-            __m256 _v_fp32 = _mm256_cvtph_ps(_v_fp16);
-            _mm256_storeu_ps(outptr, _v_fp32);
-            ptr += 8;
-            outptr += 8;
-        }
-        for (; i + 3 < size; i += 4)
-        {
-            __m128i _v_fp16 = _mm_loadl_epi64((const __m128i*)ptr);
-            __m128 _v_fp32 = _mm_cvtph_ps(_v_fp16);
-            _mm_storeu_ps(outptr, _v_fp32);
-            ptr += 4;
-            outptr += 4;
-        }
-#endif // __F16C__
-        for (; i < size; i++)
-        {
-            *outptr++ = float16_to_float32(*ptr++);
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/cast_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/cast_x86.h
deleted file mode 100644
index 45b27a8c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/cast_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CAST_X86_H
-#define LAYER_CAST_X86_H
-
-#include "cast.h"
-
-namespace ncnn {
-
-class Cast_x86 : public Cast
-{
-public:
-    Cast_x86();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CAST_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/clip_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/clip_x86.h
deleted file mode 100644
index 45a4058e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/clip_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CLIP_X86_H
-#define LAYER_CLIP_X86_H
-
-#include "clip.h"
-
-namespace ncnn {
-
-class Clip_x86 : public Clip
-{
-public:
-    Clip_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CLIP_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/concat_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/concat_x86.h
deleted file mode 100644
index 28ff162d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/concat_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONCAT_X86_H
-#define LAYER_CONCAT_X86_H
-
-#include "concat.h"
-
-namespace ncnn {
-
-class Concat_x86 : public Concat
-{
-public:
-    Concat_x86();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONCAT_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution1d_packed.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution1d_packed.h
deleted file mode 100644
index 5d20b83f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution1d_packed.h
+++ /dev/null
@@ -1,2819 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution1d_transform_kernel_packed(const Mat& kernel, Mat& kernel_tm, int inh, int outh, int kernel_w)
-{
-    // src = kw-inh-outh
-    // dst = pb-pa-kw-inh/pa-outh/pb
-
-    // clang-format off
-    // *INDENT-OFF*
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    if (outh >= 16)
-    {
-        if (inh >= 16)
-            kernel_tm.create(16 * 16 * kernel_w, inh / 16 + (inh % 16) / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 16 + (outh % 16) / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2);
-        else if (inh >= 8)
-            kernel_tm.create(16 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 16 + (outh % 16) / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2);
-        else if (inh >= 4)
-            kernel_tm.create(16 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 16 + (outh % 16) / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2);
-        else if (inh >= 2)
-            kernel_tm.create(16 * 2 * kernel_w, inh / 2 + inh % 2, outh / 16 + (outh % 16) / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2);
-        else
-            kernel_tm.create(16 * kernel_w, inh, outh / 16 + (outh % 16) / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2);
-    }
-    else
-#endif // __AVX512F__
-    if (outh >= 8)
-    {
-#if __AVX512F__
-        if (inh >= 16)
-            kernel_tm.create(8 * 16 * kernel_w, inh / 16 + (inh % 16) / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2);
-        else
-#endif // __AVX512F__
-        if (inh >= 8)
-            kernel_tm.create(8 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2);
-        else if (inh >= 4)
-            kernel_tm.create(8 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2);
-        else if (inh >= 2)
-            kernel_tm.create(8 * 2 * kernel_w, inh / 2 + inh % 2, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2);
-        else
-            kernel_tm.create(8 * kernel_w, inh, outh / 8 + (outh % 8) / 4 + (outh % 4) / 2 + outh % 2);
-    }
-    else
-#endif // __AVX__
-    if (outh >= 4)
-    {
-#if __AVX__
-#if __AVX512F__
-        if (inh >= 16)
-            kernel_tm.create(4 * 16 * kernel_w, inh / 16 + (inh % 16) / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2);
-        else
-#endif // __AVX512F__
-        if (inh >= 8)
-            kernel_tm.create(4 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2);
-        else
-#endif // __AVX__
-        if (inh >= 4)
-            kernel_tm.create(4 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2);
-        else if (inh >= 2)
-            kernel_tm.create(4 * 2 * kernel_w, inh / 2 + inh % 2, outh / 4 + (outh % 4) / 2 + outh % 2);
-        else
-            kernel_tm.create(4 * kernel_w, inh, outh / 4 + (outh % 4) / 2 + outh % 2);
-    }
-    else
-#endif // __SSE2__
-    if (outh >= 2)
-    {
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-        if (inh >= 16)
-            kernel_tm.create(2 * 16 * kernel_w, inh / 16 + (inh % 16) / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 2 + outh % 2);
-        else
-#endif // __AVX512F__
-        if (inh >= 8)
-            kernel_tm.create(2 * 8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh / 2 + outh % 2);
-        else
-#endif // __AVX__
-        if (inh >= 4)
-            kernel_tm.create(2 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 2 + outh % 2);
-        else
-#endif // __SSE2__
-        if (inh >= 2)
-            kernel_tm.create(2 * 2 * kernel_w, inh / 2 + inh % 2, outh / 2 + outh % 2);
-        else
-            kernel_tm.create(2 * kernel_w, inh, outh / 2 + outh % 2);
-    }
-    else
-    {
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-        if (inh >= 16)
-            kernel_tm.create(16 * kernel_w, inh / 16 + (inh % 16) / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh);
-        else
-#endif // __AVX512F__
-        if (inh >= 8)
-            kernel_tm.create(8 * kernel_w, inh / 8 + (inh % 8) / 4 + (inh % 4) / 2 + inh % 2, outh);
-        else
-#endif // __AVX__
-        if (inh >= 4)
-            kernel_tm.create(4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh);
-        else
-#endif // __SSE2__
-        if (inh >= 2)
-            kernel_tm.create(2 * kernel_w, inh / 2 + inh % 2, outh);
-        else
-            kernel_tm.create(kernel_w, inh, outh);
-    }
-    // *INDENT-ON*
-    // clang-format on
-
-    int q = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    for (; q + 15 < outh; q += 16)
-    {
-        const float* kptr0 = (const float*)kernel + q * inh * kernel_w;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inh * kernel_w;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inh * kernel_w;
-        const float* kptr4 = (const float*)kernel + (q + 4) * inh * kernel_w;
-        const float* kptr5 = (const float*)kernel + (q + 5) * inh * kernel_w;
-        const float* kptr6 = (const float*)kernel + (q + 6) * inh * kernel_w;
-        const float* kptr7 = (const float*)kernel + (q + 7) * inh * kernel_w;
-        const float* kptr8 = (const float*)kernel + (q + 8) * inh * kernel_w;
-        const float* kptr9 = (const float*)kernel + (q + 9) * inh * kernel_w;
-        const float* kptra = (const float*)kernel + (q + 10) * inh * kernel_w;
-        const float* kptrb = (const float*)kernel + (q + 11) * inh * kernel_w;
-        const float* kptrc = (const float*)kernel + (q + 12) * inh * kernel_w;
-        const float* kptrd = (const float*)kernel + (q + 13) * inh * kernel_w;
-        const float* kptre = (const float*)kernel + (q + 14) * inh * kernel_w;
-        const float* kptrf = (const float*)kernel + (q + 15) * inh * kernel_w;
-
-        float* g00 = kernel_tm.channel(q / 16);
-
-        __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-        _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(kernel_w));
-
-        int p = 0;
-        for (; p + 15 < inh; p += 16)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-                const float* k4 = kptr4 + k;
-                const float* k5 = kptr5 + k;
-                const float* k6 = kptr6 + k;
-                const float* k7 = kptr7 + k;
-                const float* k8 = kptr8 + k;
-                const float* k9 = kptr9 + k;
-                const float* ka = kptra + k;
-                const float* kb = kptrb + k;
-                const float* kc = kptrc + k;
-                const float* kd = kptrd + k;
-                const float* ke = kptre + k;
-                const float* kf = kptrf + k;
-
-                __m512 _k0 = _mm512_i32gather_ps(_vindex, k0, sizeof(float));
-                __m512 _k1 = _mm512_i32gather_ps(_vindex, k1, sizeof(float));
-                __m512 _k2 = _mm512_i32gather_ps(_vindex, k2, sizeof(float));
-                __m512 _k3 = _mm512_i32gather_ps(_vindex, k3, sizeof(float));
-                __m512 _k4 = _mm512_i32gather_ps(_vindex, k4, sizeof(float));
-                __m512 _k5 = _mm512_i32gather_ps(_vindex, k5, sizeof(float));
-                __m512 _k6 = _mm512_i32gather_ps(_vindex, k6, sizeof(float));
-                __m512 _k7 = _mm512_i32gather_ps(_vindex, k7, sizeof(float));
-                __m512 _k8 = _mm512_i32gather_ps(_vindex, k8, sizeof(float));
-                __m512 _k9 = _mm512_i32gather_ps(_vindex, k9, sizeof(float));
-                __m512 _ka = _mm512_i32gather_ps(_vindex, ka, sizeof(float));
-                __m512 _kb = _mm512_i32gather_ps(_vindex, kb, sizeof(float));
-                __m512 _kc = _mm512_i32gather_ps(_vindex, kc, sizeof(float));
-                __m512 _kd = _mm512_i32gather_ps(_vindex, kd, sizeof(float));
-                __m512 _ke = _mm512_i32gather_ps(_vindex, ke, sizeof(float));
-                __m512 _kf = _mm512_i32gather_ps(_vindex, kf, sizeof(float));
-
-                transpose16x16_ps(_k0, _k1, _k2, _k3, _k4, _k5, _k6, _k7, _k8, _k9, _ka, _kb, _kc, _kd, _ke, _kf);
-
-                _mm512_store_ps(g00, _k0);
-                _mm512_store_ps(g00 + 16, _k1);
-                _mm512_store_ps(g00 + 16 * 2, _k2);
-                _mm512_store_ps(g00 + 16 * 3, _k3);
-                _mm512_store_ps(g00 + 16 * 4, _k4);
-                _mm512_store_ps(g00 + 16 * 5, _k5);
-                _mm512_store_ps(g00 + 16 * 6, _k6);
-                _mm512_store_ps(g00 + 16 * 7, _k7);
-                _mm512_store_ps(g00 + 16 * 8, _k8);
-                _mm512_store_ps(g00 + 16 * 9, _k9);
-                _mm512_store_ps(g00 + 16 * 10, _ka);
-                _mm512_store_ps(g00 + 16 * 11, _kb);
-                _mm512_store_ps(g00 + 16 * 12, _kc);
-                _mm512_store_ps(g00 + 16 * 13, _kd);
-                _mm512_store_ps(g00 + 16 * 14, _ke);
-                _mm512_store_ps(g00 + 16 * 15, _kf);
-
-                g00 += 256;
-            }
-
-            kptr0 += kernel_w * 16;
-            kptr1 += kernel_w * 16;
-            kptr2 += kernel_w * 16;
-            kptr3 += kernel_w * 16;
-            kptr4 += kernel_w * 16;
-            kptr5 += kernel_w * 16;
-            kptr6 += kernel_w * 16;
-            kptr7 += kernel_w * 16;
-            kptr8 += kernel_w * 16;
-            kptr9 += kernel_w * 16;
-            kptra += kernel_w * 16;
-            kptrb += kernel_w * 16;
-            kptrc += kernel_w * 16;
-            kptrd += kernel_w * 16;
-            kptre += kernel_w * 16;
-            kptrf += kernel_w * 16;
-        }
-
-        _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(inh));
-
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    __m512 _k0 = _mm512_i32gather_ps(_vindex, k0, sizeof(float));
-                    _mm512_store_ps(g00, _k0);
-                    k0 += kernel_w;
-                    g00 += 16;
-                }
-            }
-
-            kptr0 += kernel_w * 8;
-        }
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    __m512 _k0 = _mm512_i32gather_ps(_vindex, k0, sizeof(float));
-                    _mm512_store_ps(g00, _k0);
-                    k0 += kernel_w;
-                    g00 += 16;
-                }
-            }
-
-            kptr0 += kernel_w * 4;
-        }
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    __m512 _k0 = _mm512_i32gather_ps(_vindex, k0, sizeof(float));
-                    _mm512_store_ps(g00, _k0);
-                    k0 += kernel_w;
-                    g00 += 16;
-                }
-            }
-
-            kptr0 += kernel_w * 2;
-        }
-        for (; p < inh; p++)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-
-                __m512 _k0 = _mm512_i32gather_ps(_vindex, k0, sizeof(float));
-                _mm512_store_ps(g00, _k0);
-                g00 += 16;
-            }
-        }
-    }
-#endif // __AVX512F__
-    for (; q + 7 < outh; q += 8)
-    {
-        const float* kptr0 = (const float*)kernel + q * inh * kernel_w;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inh * kernel_w;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inh * kernel_w;
-        const float* kptr4 = (const float*)kernel + (q + 4) * inh * kernel_w;
-        const float* kptr5 = (const float*)kernel + (q + 5) * inh * kernel_w;
-        const float* kptr6 = (const float*)kernel + (q + 6) * inh * kernel_w;
-        const float* kptr7 = (const float*)kernel + (q + 7) * inh * kernel_w;
-
-#if __AVX512F__
-        float* g00 = kernel_tm.channel(q / 16 + (q % 16) / 8);
-#else
-        float* g00 = kernel_tm.channel(q / 8);
-#endif
-
-#if __AVX2__
-        __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-        _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(kernel_w));
-#if __AVX512F__
-        __m512i _vindex_512 = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-        _vindex_512 = _mm512_mullo_epi32(_vindex_512, _mm512_set1_epi32(kernel_w));
-#endif // __AVX512F__
-#endif // __AVX2__
-
-        int p = 0;
-#if __AVX512F__
-        for (; p + 15 < inh; p += 16)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-                const float* k4 = kptr4 + k;
-                const float* k5 = kptr5 + k;
-                const float* k6 = kptr6 + k;
-                const float* k7 = kptr7 + k;
-
-                __m512 _k0 = _mm512_i32gather_ps(_vindex_512, k0, sizeof(float));
-                __m512 _k1 = _mm512_i32gather_ps(_vindex_512, k1, sizeof(float));
-                __m512 _k2 = _mm512_i32gather_ps(_vindex_512, k2, sizeof(float));
-                __m512 _k3 = _mm512_i32gather_ps(_vindex_512, k3, sizeof(float));
-                __m512 _k4 = _mm512_i32gather_ps(_vindex_512, k4, sizeof(float));
-                __m512 _k5 = _mm512_i32gather_ps(_vindex_512, k5, sizeof(float));
-                __m512 _k6 = _mm512_i32gather_ps(_vindex_512, k6, sizeof(float));
-                __m512 _k7 = _mm512_i32gather_ps(_vindex_512, k7, sizeof(float));
-
-                transpose16x8_ps(_k0, _k1, _k2, _k3, _k4, _k5, _k6, _k7);
-
-                _mm512_storeu_ps(g00, _k0);
-                _mm512_storeu_ps(g00 + 16, _k1);
-                _mm512_storeu_ps(g00 + 16 * 2, _k2);
-                _mm512_storeu_ps(g00 + 16 * 3, _k3);
-                _mm512_storeu_ps(g00 + 16 * 4, _k4);
-                _mm512_storeu_ps(g00 + 16 * 5, _k5);
-                _mm512_storeu_ps(g00 + 16 * 6, _k6);
-                _mm512_storeu_ps(g00 + 16 * 7, _k7);
-
-                g00 += 128;
-            }
-
-            kptr0 += kernel_w * 16;
-            kptr1 += kernel_w * 16;
-            kptr2 += kernel_w * 16;
-            kptr3 += kernel_w * 16;
-            kptr4 += kernel_w * 16;
-            kptr5 += kernel_w * 16;
-            kptr6 += kernel_w * 16;
-            kptr7 += kernel_w * 16;
-        }
-#endif // __AVX512F__
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-                const float* k4 = kptr4 + k;
-                const float* k5 = kptr5 + k;
-                const float* k6 = kptr6 + k;
-                const float* k7 = kptr7 + k;
-
-#if __AVX2__
-                __m256 _k0 = _mm256_i32gather_ps(k0, _vindex, sizeof(float));
-                __m256 _k1 = _mm256_i32gather_ps(k1, _vindex, sizeof(float));
-                __m256 _k2 = _mm256_i32gather_ps(k2, _vindex, sizeof(float));
-                __m256 _k3 = _mm256_i32gather_ps(k3, _vindex, sizeof(float));
-                __m256 _k4 = _mm256_i32gather_ps(k4, _vindex, sizeof(float));
-                __m256 _k5 = _mm256_i32gather_ps(k5, _vindex, sizeof(float));
-                __m256 _k6 = _mm256_i32gather_ps(k6, _vindex, sizeof(float));
-                __m256 _k7 = _mm256_i32gather_ps(k7, _vindex, sizeof(float));
-
-                transpose8x8_ps(_k0, _k1, _k2, _k3, _k4, _k5, _k6, _k7);
-
-                _mm256_store_ps(g00, _k0);
-                _mm256_store_ps(g00 + 8, _k1);
-                _mm256_store_ps(g00 + 8 * 2, _k2);
-                _mm256_store_ps(g00 + 8 * 3, _k3);
-                _mm256_store_ps(g00 + 8 * 4, _k4);
-                _mm256_store_ps(g00 + 8 * 5, _k5);
-                _mm256_store_ps(g00 + 8 * 6, _k6);
-                _mm256_store_ps(g00 + 8 * 7, _k7);
-
-                g00 += 64;
-#else  // __AVX2__
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[0];
-                    g00[1] = k1[0];
-                    g00[2] = k2[0];
-                    g00[3] = k3[0];
-                    g00[4] = k4[0];
-                    g00[5] = k5[0];
-                    g00[6] = k6[0];
-                    g00[7] = k7[0];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    k4 += kernel_w;
-                    k5 += kernel_w;
-                    k6 += kernel_w;
-                    k7 += kernel_w;
-                    g00 += 8;
-                }
-#endif // __AVX2__
-            }
-
-            kptr0 += kernel_w * 8;
-            kptr1 += kernel_w * 8;
-            kptr2 += kernel_w * 8;
-            kptr3 += kernel_w * 8;
-            kptr4 += kernel_w * 8;
-            kptr5 += kernel_w * 8;
-            kptr6 += kernel_w * 8;
-            kptr7 += kernel_w * 8;
-        }
-
-#if __AVX2__
-        _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(inh));
-#endif // __AVX2__
-
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-#if !__AVX2__
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-                const float* k4 = kptr4 + k;
-                const float* k5 = kptr5 + k;
-                const float* k6 = kptr6 + k;
-                const float* k7 = kptr7 + k;
-#endif // !__AVX2__
-
-                for (int i = 0; i < 4; i++)
-                {
-#if __AVX2__
-                    __m256 _k0 = _mm256_i32gather_ps(k0, _vindex, sizeof(float));
-                    _mm256_store_ps(g00, _k0);
-                    k0 += kernel_w;
-                    g00 += 8;
-#else  // __AVX2__
-                    g00[0] = k0[0];
-                    g00[1] = k1[0];
-                    g00[2] = k2[0];
-                    g00[3] = k3[0];
-                    g00[4] = k4[0];
-                    g00[5] = k5[0];
-                    g00[6] = k6[0];
-                    g00[7] = k7[0];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    k4 += kernel_w;
-                    k5 += kernel_w;
-                    k6 += kernel_w;
-                    k7 += kernel_w;
-                    g00 += 8;
-#endif // __AVX2__
-                }
-            }
-
-            kptr0 += kernel_w * 4;
-#if !__AVX2__
-            kptr1 += kernel_w * 4;
-            kptr2 += kernel_w * 4;
-            kptr3 += kernel_w * 4;
-            kptr4 += kernel_w * 4;
-            kptr5 += kernel_w * 4;
-            kptr6 += kernel_w * 4;
-            kptr7 += kernel_w * 4;
-#endif // !__AVX2__
-        }
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-#if !__AVX2__
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-                const float* k4 = kptr4 + k;
-                const float* k5 = kptr5 + k;
-                const float* k6 = kptr6 + k;
-                const float* k7 = kptr7 + k;
-#endif // !__AVX2__
-
-                for (int i = 0; i < 2; i++)
-                {
-#if __AVX2__
-                    __m256 _k0 = _mm256_i32gather_ps(k0, _vindex, sizeof(float));
-                    _mm256_store_ps(g00, _k0);
-                    k0 += kernel_w;
-                    g00 += 8;
-#else  // __AVX2__
-                    g00[0] = k0[0];
-                    g00[1] = k1[0];
-                    g00[2] = k2[0];
-                    g00[3] = k3[0];
-                    g00[4] = k4[0];
-                    g00[5] = k5[0];
-                    g00[6] = k6[0];
-                    g00[7] = k7[0];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    k4 += kernel_w;
-                    k5 += kernel_w;
-                    k6 += kernel_w;
-                    k7 += kernel_w;
-                    g00 += 8;
-#endif // __AVX2__
-                }
-            }
-
-            kptr0 += kernel_w * 2;
-#if !__AVX2__
-            kptr1 += kernel_w * 2;
-            kptr2 += kernel_w * 2;
-            kptr3 += kernel_w * 2;
-            kptr4 += kernel_w * 2;
-            kptr5 += kernel_w * 2;
-            kptr6 += kernel_w * 2;
-            kptr7 += kernel_w * 2;
-#endif // !__AVX2__
-        }
-        for (; p < inh; p++)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-#if __AVX2__
-                __m256 _k0 = _mm256_i32gather_ps(k0, _vindex, sizeof(float));
-                _mm256_store_ps(g00, _k0);
-                g00 += 8;
-#else  // __AVX2__
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-                const float* k4 = kptr4 + k;
-                const float* k5 = kptr5 + k;
-                const float* k6 = kptr6 + k;
-                const float* k7 = kptr7 + k;
-
-                g00[0] = k0[0];
-                g00[1] = k1[0];
-                g00[2] = k2[0];
-                g00[3] = k3[0];
-                g00[4] = k4[0];
-                g00[5] = k5[0];
-                g00[6] = k6[0];
-                g00[7] = k7[0];
-                g00 += 8;
-#endif // __AVX2__
-            }
-        }
-    }
-#endif // __AVX__
-    for (; q + 3 < outh; q += 4)
-    {
-        const float* kptr0 = (const float*)kernel + q * inh * kernel_w;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inh * kernel_w;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inh * kernel_w;
-
-#if __AVX512F__
-        float* g00 = kernel_tm.channel(q / 16 + (q % 16) / 8 + (q % 8) / 4);
-#elif __AVX__
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4);
-#else
-        float* g00 = kernel_tm.channel(q / 4);
-#endif
-
-#if __AVX2__
-        __m128i _vindex = _mm_setr_epi32(0, 1, 2, 3);
-        _vindex = _mm_mullo_epi32(_vindex, _mm_set1_epi32(kernel_w));
-        __m256i _vindex_256 = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-        _vindex_256 = _mm256_mullo_epi32(_vindex_256, _mm256_set1_epi32(kernel_w));
-#if __AVX512F__
-        __m512i _vindex_512 = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-        _vindex_512 = _mm512_mullo_epi32(_vindex_512, _mm512_set1_epi32(kernel_w));
-#endif // __AVX512F__
-#endif // __AVX2__
-
-        int p = 0;
-#if __AVX__
-#if __AVX512F__
-        for (; p + 15 < inh; p += 16)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-
-                __m512 _k0 = _mm512_i32gather_ps(_vindex_512, k0, sizeof(float));
-                __m512 _k1 = _mm512_i32gather_ps(_vindex_512, k1, sizeof(float));
-                __m512 _k2 = _mm512_i32gather_ps(_vindex_512, k2, sizeof(float));
-                __m512 _k3 = _mm512_i32gather_ps(_vindex_512, k3, sizeof(float));
-
-                transpose16x4_ps(_k0, _k1, _k2, _k3);
-
-                _mm512_storeu_ps(g00, _k0);
-                _mm512_storeu_ps(g00 + 16, _k1);
-                _mm512_storeu_ps(g00 + 16 * 2, _k2);
-                _mm512_storeu_ps(g00 + 16 * 3, _k3);
-
-                g00 += 64;
-            }
-
-            kptr0 += kernel_w * 16;
-            kptr1 += kernel_w * 16;
-            kptr2 += kernel_w * 16;
-            kptr3 += kernel_w * 16;
-        }
-#endif // __AVX512F__
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-
-#if __AVX2__
-                __m256 _k0 = _mm256_i32gather_ps(k0, _vindex_256, sizeof(float));
-                __m256 _k1 = _mm256_i32gather_ps(k1, _vindex_256, sizeof(float));
-                __m256 _k2 = _mm256_i32gather_ps(k2, _vindex_256, sizeof(float));
-                __m256 _k3 = _mm256_i32gather_ps(k3, _vindex_256, sizeof(float));
-
-                transpose8x4_ps(_k0, _k1, _k2, _k3);
-
-                _mm256_storeu_ps(g00, _k0);
-                _mm256_storeu_ps(g00 + 8, _k1);
-                _mm256_storeu_ps(g00 + 8 * 2, _k2);
-                _mm256_storeu_ps(g00 + 8 * 3, _k3);
-
-                g00 += 32;
-#else  // __AVX2__
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[0];
-                    g00[1] = k1[0];
-                    g00[2] = k2[0];
-                    g00[3] = k3[0];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    g00 += 4;
-                }
-#endif // __AVX2__
-            }
-
-            kptr0 += kernel_w * 8;
-            kptr1 += kernel_w * 8;
-            kptr2 += kernel_w * 8;
-            kptr3 += kernel_w * 8;
-        }
-#endif // __AVX__
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-
-#if __AVX2__
-                __m128 _k0 = _mm_i32gather_ps(k0, _vindex, sizeof(float));
-                __m128 _k1 = _mm_i32gather_ps(k1, _vindex, sizeof(float));
-                __m128 _k2 = _mm_i32gather_ps(k2, _vindex, sizeof(float));
-                __m128 _k3 = _mm_i32gather_ps(k3, _vindex, sizeof(float));
-
-                _MM_TRANSPOSE4_PS(_k0, _k1, _k2, _k3);
-
-                _mm_store_ps(g00, _k0);
-                _mm_store_ps(g00 + 4, _k1);
-                _mm_store_ps(g00 + 4 * 2, _k2);
-                _mm_store_ps(g00 + 4 * 3, _k3);
-
-                g00 += 16;
-#else  // __AVX2__
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k0[0];
-                    g00[1] = k1[0];
-                    g00[2] = k2[0];
-                    g00[3] = k3[0];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    g00 += 4;
-                }
-#endif // __AVX2__
-            }
-
-            kptr0 += kernel_w * 4;
-            kptr1 += kernel_w * 4;
-            kptr2 += kernel_w * 4;
-            kptr3 += kernel_w * 4;
-        }
-
-#if __AVX2__
-        _vindex = _mm_mullo_epi32(_vindex, _mm_set1_epi32(inh));
-#endif // __AVX2__
-
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-#if !__AVX2__
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-#endif // !__AVX2__
-
-                for (int i = 0; i < 2; i++)
-                {
-#if __AVX2__
-                    __m128 _k0 = _mm_i32gather_ps(k0, _vindex, sizeof(float));
-                    _mm_store_ps(g00, _k0);
-                    k0 += kernel_w;
-                    g00 += 4;
-#else  // __AVX2__
-                    g00[0] = k0[0];
-                    g00[1] = k1[0];
-                    g00[2] = k2[0];
-                    g00[3] = k3[0];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    k2 += kernel_w;
-                    k3 += kernel_w;
-                    g00 += 4;
-#endif // __AVX2__
-                }
-            }
-
-            kptr0 += kernel_w * 2;
-#if !__AVX2__
-            kptr1 += kernel_w * 2;
-            kptr2 += kernel_w * 2;
-            kptr3 += kernel_w * 2;
-#endif // !__AVX2__
-        }
-        for (; p < inh; p++)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-#if __AVX2__
-                __m128 _k0 = _mm_i32gather_ps(k0, _vindex, sizeof(float));
-                _mm_store_ps(g00, _k0);
-                g00 += 4;
-#else  // __AVX2__
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-
-                g00[0] = k0[0];
-                g00[1] = k1[0];
-                g00[2] = k2[0];
-                g00[3] = k3[0];
-                g00 += 4;
-#endif // __AVX2__
-            }
-        }
-    }
-#endif // __SSE2__
-    for (; q + 1 < outh; q += 2)
-    {
-        const float* kptr0 = (const float*)kernel + q * inh * kernel_w;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inh * kernel_w;
-
-#if __AVX512F__
-        float* g00 = kernel_tm.channel(q / 16 + (q % 16) / 8 + (q % 8) / 4 + (q % 4) / 2);
-#elif __AVX__
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2);
-#elif __SSE2__
-        float* g00 = kernel_tm.channel(q / 4 + (q % 4) / 2);
-#else
-        float* g00 = kernel_tm.channel(q / 2);
-#endif
-
-#if __AVX2__
-        __m128i _vindex = _mm_setr_epi32(0, 1, 2, 3);
-        _vindex = _mm_mullo_epi32(_vindex, _mm_set1_epi32(kernel_w));
-        __m256i _vindex_256 = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-        _vindex_256 = _mm256_mullo_epi32(_vindex_256, _mm256_set1_epi32(kernel_w));
-#if __AVX512F__
-        __m512i _vindex_512 = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-        _vindex_512 = _mm512_mullo_epi32(_vindex_512, _mm512_set1_epi32(kernel_w));
-#endif // __AVX512F__
-#endif // __AVX2__
-
-        int p = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-        for (; p + 15 < inh; p += 16)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-
-                __m512 _k0 = _mm512_i32gather_ps(_vindex_512, k0, sizeof(float));
-                __m512 _k1 = _mm512_i32gather_ps(_vindex_512, k1, sizeof(float));
-                _mm512_storeu_ps(g00, _k0);
-                _mm512_storeu_ps(g00 + 16, _k1);
-                g00 += 32;
-            }
-
-            kptr0 += kernel_w * 16;
-            kptr1 += kernel_w * 16;
-        }
-#endif // __AVX512F__
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-
-#if __AVX2__
-                __m256 _k0 = _mm256_i32gather_ps(k0, _vindex_256, sizeof(float));
-                __m256 _k1 = _mm256_i32gather_ps(k1, _vindex_256, sizeof(float));
-                _mm256_storeu_ps(g00, _k0);
-                _mm256_storeu_ps(g00 + 8, _k1);
-                g00 += 16;
-#else  // __AVX2__
-                g00[0] = k0[0];
-                g00[1] = k0[kernel_w];
-                g00[2] = k0[kernel_w * 2];
-                g00[3] = k0[kernel_w * 3];
-                g00[4] = k0[kernel_w * 4];
-                g00[5] = k0[kernel_w * 5];
-                g00[6] = k0[kernel_w * 6];
-                g00[7] = k0[kernel_w * 7];
-                g00[8] = k1[0];
-                g00[9] = k1[kernel_w];
-                g00[10] = k1[kernel_w * 2];
-                g00[11] = k1[kernel_w * 3];
-                g00[12] = k1[kernel_w * 4];
-                g00[13] = k1[kernel_w * 5];
-                g00[14] = k1[kernel_w * 6];
-                g00[15] = k1[kernel_w * 7];
-                g00 += 16;
-#endif // __AVX2__
-            }
-
-            kptr0 += kernel_w * 8;
-            kptr1 += kernel_w * 8;
-        }
-#endif // __AVX__
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-
-#if __AVX2__
-                __m128 _k0 = _mm_i32gather_ps(k0, _vindex, sizeof(float));
-                __m128 _k1 = _mm_i32gather_ps(k1, _vindex, sizeof(float));
-                _mm_storeu_ps(g00, _k0);
-                _mm_storeu_ps(g00 + 4, _k1);
-                g00 += 8;
-#else  // __AVX2__
-                g00[0] = k0[0];
-                g00[1] = k0[kernel_w];
-                g00[2] = k0[kernel_w * 2];
-                g00[3] = k0[kernel_w * 3];
-                g00[4] = k1[0];
-                g00[5] = k1[kernel_w];
-                g00[6] = k1[kernel_w * 2];
-                g00[7] = k1[kernel_w * 3];
-                g00 += 8;
-#endif // __AVX2__
-            }
-
-            kptr0 += kernel_w * 4;
-            kptr1 += kernel_w * 4;
-        }
-#endif // __SSE2__
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = k0[0];
-                    g00[1] = k1[0];
-                    k0 += kernel_w;
-                    k1 += kernel_w;
-                    g00 += 2;
-                }
-            }
-
-            kptr0 += kernel_w * 2;
-            kptr1 += kernel_w * 2;
-        }
-        for (; p < inh; p++)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-
-                g00[0] = k0[0];
-                g00[1] = k1[0];
-                g00 += 2;
-            }
-        }
-    }
-    for (; q < outh; q++)
-    {
-        const float* kptr = (const float*)kernel + q * inh * kernel_w;
-
-#if __AVX512F__
-        float* g00 = kernel_tm.channel(q / 16 + (q % 16) / 8 + (q % 8) / 4 + (q % 4) / 2 + q % 2);
-#elif __AVX__
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2 + q % 2);
-#elif __SSE2__
-        float* g00 = kernel_tm.channel(q / 4 + (q % 4) / 2 + q % 2);
-#else
-        float* g00 = kernel_tm.channel(q / 2 + q % 2);
-#endif
-
-#if __AVX2__
-        __m128i _vindex = _mm_setr_epi32(0, 1, 2, 3);
-        _vindex = _mm_mullo_epi32(_vindex, _mm_set1_epi32(kernel_w));
-        __m256i _vindex_256 = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-        _vindex_256 = _mm256_mullo_epi32(_vindex_256, _mm256_set1_epi32(kernel_w));
-#if __AVX512F__
-        __m512i _vindex_512 = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-        _vindex_512 = _mm512_mullo_epi32(_vindex_512, _mm512_set1_epi32(kernel_w));
-#endif // __AVX512F__
-#endif // __AVX2__
-
-        int p = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-        for (; p + 15 < inh; p += 16)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr + k;
-
-                __m512 _k0 = _mm512_i32gather_ps(_vindex_512, k0, sizeof(float));
-                _mm512_storeu_ps(g00, _k0);
-                g00 += 16;
-            }
-
-            kptr += kernel_w * 16;
-        }
-#endif // __AVX512F__
-        for (; p + 7 < inh; p += 8)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr + k;
-
-#if __AVX2__
-                __m256 _k0 = _mm256_i32gather_ps(k0, _vindex_256, sizeof(float));
-                _mm256_storeu_ps(g00, _k0);
-                g00 += 8;
-#else  // __AVX2__
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[0];
-                    k0 += kernel_w;
-                    g00 += 1;
-                }
-#endif // __AVX2__
-            }
-
-            kptr += kernel_w * 8;
-        }
-#endif // __AVX__
-        for (; p + 3 < inh; p += 4)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr + k;
-
-#if __AVX2__
-                __m128 _k0 = _mm_i32gather_ps(k0, _vindex, sizeof(float));
-                _mm_storeu_ps(g00, _k0);
-                g00 += 4;
-#else  // __AVX2__
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k0[0];
-                    k0 += kernel_w;
-                    g00 += 1;
-                }
-#endif // __AVX2__
-            }
-
-            kptr += kernel_w * 4;
-        }
-#endif // __SSE2__
-        for (; p + 1 < inh; p += 2)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr + k;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = k0[0];
-                    k0 += kernel_w;
-                    g00 += 1;
-                }
-            }
-
-            kptr += kernel_w * 2;
-        }
-        for (; p < inh; p++)
-        {
-            for (int k = 0; k < kernel_w; k++)
-            {
-                const float* k0 = kptr + k;
-                g00[0] = k0[0];
-                g00++;
-            }
-        }
-    }
-}
-
-static void convolution1d_packed(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int kernel_w, int dilation_w, int stride_w, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int elempack = bottom_blob.elempack;
-    const int inh = bottom_blob.h * elempack;
-
-    const int N = bottom_blob.w * elempack;
-
-    const int outw = top_blob.w;
-    const int out_elempack = top_blob.elempack;
-    const int outh = top_blob.h * out_elempack;
-
-    const int M = top_blob.w * out_elempack;
-
-    const float* bias_data_ptr = bias_data;
-
-    int nn_outh = 0;
-    int remain_outh_start = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    nn_outh = outh / 16;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = pp * 16;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-        const int out_elempack = top_blob.elempack;
-
-        float* outptr = top_blob.row(p / out_elempack);
-
-        for (int j = 0; j < outw; j++)
-        {
-            __m512 _sum0 = _mm512_setzero_ps();
-            __m512 _sum1 = _mm512_setzero_ps();
-            __m512 _sum2 = _mm512_setzero_ps();
-            __m512 _sum3 = _mm512_setzero_ps();
-
-            if (bias_data_ptr)
-            {
-                _sum0 = _mm512_loadu_ps(bias_data_ptr + p);
-            }
-
-            const float* kptr = weight_data_tm.channel(p / 16);
-
-            int q = 0;
-            for (; q + 15 < inh; q += 16)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                if (elempack == 16)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _w0 = _mm512_load_ps(kptr + 16 * 0);
-                        __m512 _w1 = _mm512_load_ps(kptr + 16 * 1);
-                        __m512 _w2 = _mm512_load_ps(kptr + 16 * 2);
-                        __m512 _w3 = _mm512_load_ps(kptr + 16 * 3);
-                        __m512 _w4 = _mm512_load_ps(kptr + 16 * 4);
-                        __m512 _w5 = _mm512_load_ps(kptr + 16 * 5);
-                        __m512 _w6 = _mm512_load_ps(kptr + 16 * 6);
-                        __m512 _w7 = _mm512_load_ps(kptr + 16 * 7);
-                        __m512 _w8 = _mm512_load_ps(kptr + 16 * 8);
-                        __m512 _w9 = _mm512_load_ps(kptr + 16 * 9);
-                        __m512 _wa = _mm512_load_ps(kptr + 16 * 10);
-                        __m512 _wb = _mm512_load_ps(kptr + 16 * 11);
-                        __m512 _wc = _mm512_load_ps(kptr + 16 * 12);
-                        __m512 _wd = _mm512_load_ps(kptr + 16 * 13);
-                        __m512 _we = _mm512_load_ps(kptr + 16 * 14);
-                        __m512 _wf = _mm512_load_ps(kptr + 16 * 15);
-
-                        _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0[3]), _sum3);
-                        _sum0 = _mm512_fmadd_ps(_w4, _mm512_set1_ps(r0[4]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w5, _mm512_set1_ps(r0[5]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w6, _mm512_set1_ps(r0[6]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w7, _mm512_set1_ps(r0[7]), _sum3);
-                        _sum0 = _mm512_fmadd_ps(_w8, _mm512_set1_ps(r0[8]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w9, _mm512_set1_ps(r0[9]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_wa, _mm512_set1_ps(r0[10]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_wb, _mm512_set1_ps(r0[11]), _sum3);
-                        _sum0 = _mm512_fmadd_ps(_wc, _mm512_set1_ps(r0[12]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_wd, _mm512_set1_ps(r0[13]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_we, _mm512_set1_ps(r0[14]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_wf, _mm512_set1_ps(r0[15]), _sum3);
-
-                        r0 += dilation_w * 16;
-                        kptr += 256;
-                    }
-                }
-                if (elempack == 8)
-                {
-                    const float* r1 = r0 + N;
-
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _w0 = _mm512_load_ps(kptr + 16 * 0);
-                        __m512 _w1 = _mm512_load_ps(kptr + 16 * 1);
-                        __m512 _w2 = _mm512_load_ps(kptr + 16 * 2);
-                        __m512 _w3 = _mm512_load_ps(kptr + 16 * 3);
-                        __m512 _w4 = _mm512_load_ps(kptr + 16 * 4);
-                        __m512 _w5 = _mm512_load_ps(kptr + 16 * 5);
-                        __m512 _w6 = _mm512_load_ps(kptr + 16 * 6);
-                        __m512 _w7 = _mm512_load_ps(kptr + 16 * 7);
-                        __m512 _w8 = _mm512_load_ps(kptr + 16 * 8);
-                        __m512 _w9 = _mm512_load_ps(kptr + 16 * 9);
-                        __m512 _wa = _mm512_load_ps(kptr + 16 * 10);
-                        __m512 _wb = _mm512_load_ps(kptr + 16 * 11);
-                        __m512 _wc = _mm512_load_ps(kptr + 16 * 12);
-                        __m512 _wd = _mm512_load_ps(kptr + 16 * 13);
-                        __m512 _we = _mm512_load_ps(kptr + 16 * 14);
-                        __m512 _wf = _mm512_load_ps(kptr + 16 * 15);
-
-                        _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0[3]), _sum3);
-                        _sum0 = _mm512_fmadd_ps(_w4, _mm512_set1_ps(r0[4]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w5, _mm512_set1_ps(r0[5]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w6, _mm512_set1_ps(r0[6]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w7, _mm512_set1_ps(r0[7]), _sum3);
-                        _sum0 = _mm512_fmadd_ps(_w8, _mm512_set1_ps(r1[0]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w9, _mm512_set1_ps(r1[1]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_wa, _mm512_set1_ps(r1[2]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_wb, _mm512_set1_ps(r1[3]), _sum3);
-                        _sum0 = _mm512_fmadd_ps(_wc, _mm512_set1_ps(r1[4]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_wd, _mm512_set1_ps(r1[5]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_we, _mm512_set1_ps(r1[6]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_wf, _mm512_set1_ps(r1[7]), _sum3);
-
-                        r0 += dilation_w * 8;
-                        r1 += dilation_w * 8;
-                        kptr += 256;
-                    }
-                }
-                if (elempack == 4)
-                {
-                    const float* r1 = r0 + N;
-                    const float* r2 = r0 + N * 2;
-                    const float* r3 = r0 + N * 3;
-
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _w0 = _mm512_load_ps(kptr + 16 * 0);
-                        __m512 _w1 = _mm512_load_ps(kptr + 16 * 1);
-                        __m512 _w2 = _mm512_load_ps(kptr + 16 * 2);
-                        __m512 _w3 = _mm512_load_ps(kptr + 16 * 3);
-                        __m512 _w4 = _mm512_load_ps(kptr + 16 * 4);
-                        __m512 _w5 = _mm512_load_ps(kptr + 16 * 5);
-                        __m512 _w6 = _mm512_load_ps(kptr + 16 * 6);
-                        __m512 _w7 = _mm512_load_ps(kptr + 16 * 7);
-                        __m512 _w8 = _mm512_load_ps(kptr + 16 * 8);
-                        __m512 _w9 = _mm512_load_ps(kptr + 16 * 9);
-                        __m512 _wa = _mm512_load_ps(kptr + 16 * 10);
-                        __m512 _wb = _mm512_load_ps(kptr + 16 * 11);
-                        __m512 _wc = _mm512_load_ps(kptr + 16 * 12);
-                        __m512 _wd = _mm512_load_ps(kptr + 16 * 13);
-                        __m512 _we = _mm512_load_ps(kptr + 16 * 14);
-                        __m512 _wf = _mm512_load_ps(kptr + 16 * 15);
-
-                        _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0[3]), _sum3);
-                        _sum0 = _mm512_fmadd_ps(_w4, _mm512_set1_ps(r1[0]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w5, _mm512_set1_ps(r1[1]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w6, _mm512_set1_ps(r1[2]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w7, _mm512_set1_ps(r1[3]), _sum3);
-                        _sum0 = _mm512_fmadd_ps(_w8, _mm512_set1_ps(r2[0]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w9, _mm512_set1_ps(r2[1]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_wa, _mm512_set1_ps(r2[2]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_wb, _mm512_set1_ps(r2[3]), _sum3);
-                        _sum0 = _mm512_fmadd_ps(_wc, _mm512_set1_ps(r3[0]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_wd, _mm512_set1_ps(r3[1]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_we, _mm512_set1_ps(r3[2]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_wf, _mm512_set1_ps(r3[3]), _sum3);
-
-                        r0 += dilation_w * 4;
-                        r1 += dilation_w * 4;
-                        r2 += dilation_w * 4;
-                        r3 += dilation_w * 4;
-                        kptr += 256;
-                    }
-                }
-                if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _w0 = _mm512_load_ps(kptr + 16 * 0);
-                        __m512 _w1 = _mm512_load_ps(kptr + 16 * 1);
-                        __m512 _w2 = _mm512_load_ps(kptr + 16 * 2);
-                        __m512 _w3 = _mm512_load_ps(kptr + 16 * 3);
-                        __m512 _w4 = _mm512_load_ps(kptr + 16 * 4);
-                        __m512 _w5 = _mm512_load_ps(kptr + 16 * 5);
-                        __m512 _w6 = _mm512_load_ps(kptr + 16 * 6);
-                        __m512 _w7 = _mm512_load_ps(kptr + 16 * 7);
-                        __m512 _w8 = _mm512_load_ps(kptr + 16 * 8);
-                        __m512 _w9 = _mm512_load_ps(kptr + 16 * 9);
-                        __m512 _wa = _mm512_load_ps(kptr + 16 * 10);
-                        __m512 _wb = _mm512_load_ps(kptr + 16 * 11);
-                        __m512 _wc = _mm512_load_ps(kptr + 16 * 12);
-                        __m512 _wd = _mm512_load_ps(kptr + 16 * 13);
-                        __m512 _we = _mm512_load_ps(kptr + 16 * 14);
-                        __m512 _wf = _mm512_load_ps(kptr + 16 * 15);
-
-                        _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0[N]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0[N * 2]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0[N * 3]), _sum3);
-                        _sum0 = _mm512_fmadd_ps(_w4, _mm512_set1_ps(r0[N * 4]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w5, _mm512_set1_ps(r0[N * 5]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w6, _mm512_set1_ps(r0[N * 6]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w7, _mm512_set1_ps(r0[N * 7]), _sum3);
-                        _sum0 = _mm512_fmadd_ps(_w8, _mm512_set1_ps(r0[N * 8]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w9, _mm512_set1_ps(r0[N * 9]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_wa, _mm512_set1_ps(r0[N * 10]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_wb, _mm512_set1_ps(r0[N * 11]), _sum3);
-                        _sum0 = _mm512_fmadd_ps(_wc, _mm512_set1_ps(r0[N * 12]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_wd, _mm512_set1_ps(r0[N * 13]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_we, _mm512_set1_ps(r0[N * 14]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_wf, _mm512_set1_ps(r0[N * 15]), _sum3);
-
-                        r0 += dilation_w;
-                        kptr += 256;
-                    }
-                }
-            }
-            for (; q + 7 < inh; q += 8)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                if (elempack == 8)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _w0 = _mm512_load_ps(kptr + 16 * 0);
-                        __m512 _w1 = _mm512_load_ps(kptr + 16 * 1);
-                        __m512 _w2 = _mm512_load_ps(kptr + 16 * 2);
-                        __m512 _w3 = _mm512_load_ps(kptr + 16 * 3);
-                        __m512 _w4 = _mm512_load_ps(kptr + 16 * 4);
-                        __m512 _w5 = _mm512_load_ps(kptr + 16 * 5);
-                        __m512 _w6 = _mm512_load_ps(kptr + 16 * 6);
-                        __m512 _w7 = _mm512_load_ps(kptr + 16 * 7);
-
-                        _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0[3]), _sum3);
-                        _sum0 = _mm512_fmadd_ps(_w4, _mm512_set1_ps(r0[4]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w5, _mm512_set1_ps(r0[5]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w6, _mm512_set1_ps(r0[6]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w7, _mm512_set1_ps(r0[7]), _sum3);
-
-                        r0 += dilation_w * 8;
-                        kptr += 128;
-                    }
-                }
-                if (elempack == 4)
-                {
-                    const float* r1 = r0 + N;
-
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _w0 = _mm512_load_ps(kptr + 16 * 0);
-                        __m512 _w1 = _mm512_load_ps(kptr + 16 * 1);
-                        __m512 _w2 = _mm512_load_ps(kptr + 16 * 2);
-                        __m512 _w3 = _mm512_load_ps(kptr + 16 * 3);
-                        __m512 _w4 = _mm512_load_ps(kptr + 16 * 4);
-                        __m512 _w5 = _mm512_load_ps(kptr + 16 * 5);
-                        __m512 _w6 = _mm512_load_ps(kptr + 16 * 6);
-                        __m512 _w7 = _mm512_load_ps(kptr + 16 * 7);
-
-                        _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0[3]), _sum3);
-                        _sum0 = _mm512_fmadd_ps(_w4, _mm512_set1_ps(r1[0]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w5, _mm512_set1_ps(r1[1]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w6, _mm512_set1_ps(r1[2]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w7, _mm512_set1_ps(r1[3]), _sum3);
-
-                        r0 += dilation_w * 4;
-                        r1 += dilation_w * 4;
-                        kptr += 128;
-                    }
-                }
-                if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _w0 = _mm512_load_ps(kptr + 16 * 0);
-                        __m512 _w1 = _mm512_load_ps(kptr + 16 * 1);
-                        __m512 _w2 = _mm512_load_ps(kptr + 16 * 2);
-                        __m512 _w3 = _mm512_load_ps(kptr + 16 * 3);
-                        __m512 _w4 = _mm512_load_ps(kptr + 16 * 4);
-                        __m512 _w5 = _mm512_load_ps(kptr + 16 * 5);
-                        __m512 _w6 = _mm512_load_ps(kptr + 16 * 6);
-                        __m512 _w7 = _mm512_load_ps(kptr + 16 * 7);
-
-                        _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0[N]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0[N * 2]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0[N * 3]), _sum3);
-                        _sum0 = _mm512_fmadd_ps(_w4, _mm512_set1_ps(r0[N * 4]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w5, _mm512_set1_ps(r0[N * 5]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w6, _mm512_set1_ps(r0[N * 6]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w7, _mm512_set1_ps(r0[N * 7]), _sum3);
-
-                        r0 += dilation_w;
-                        kptr += 128;
-                    }
-                }
-            }
-            for (; q + 3 < inh; q += 4)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                if (elempack == 4)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _w0 = _mm512_load_ps(kptr);
-                        __m512 _w1 = _mm512_load_ps(kptr + 16);
-                        __m512 _w2 = _mm512_load_ps(kptr + 32);
-                        __m512 _w3 = _mm512_load_ps(kptr + 48);
-
-                        _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0[3]), _sum3);
-
-                        r0 += dilation_w * 4;
-                        kptr += 64;
-                    }
-                }
-                if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _w0 = _mm512_load_ps(kptr);
-                        __m512 _w1 = _mm512_load_ps(kptr + 16);
-                        __m512 _w2 = _mm512_load_ps(kptr + 32);
-                        __m512 _w3 = _mm512_load_ps(kptr + 48);
-
-                        _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0[N]), _sum1);
-                        _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0[N * 2]), _sum2);
-                        _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0[N * 3]), _sum3);
-
-                        r0 += dilation_w;
-                        kptr += 64;
-                    }
-                }
-            }
-            for (; q + 1 < inh; q += 2)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                // if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _w0 = _mm512_load_ps(kptr);
-                        __m512 _w1 = _mm512_load_ps(kptr + 16);
-
-                        _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0[N]), _sum1);
-
-                        r0 += dilation_w;
-                        kptr += 32;
-                    }
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                // if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _val = _mm512_set1_ps(r0[0]);
-                        __m512 _w = _mm512_load_ps(kptr);
-                        _sum0 = _mm512_fmadd_ps(_val, _w, _sum0);
-
-                        r0 += dilation_w;
-                        kptr += 16;
-                    }
-                }
-            }
-
-            _sum0 = _mm512_add_ps(_sum0, _sum1);
-            _sum2 = _mm512_add_ps(_sum2, _sum3);
-            _sum0 = _mm512_add_ps(_sum0, _sum2);
-
-            _sum0 = activation_avx512(_sum0, activation_type, activation_params);
-
-            if (out_elempack == 16)
-            {
-                _mm512_store_ps(outptr, _sum0);
-                outptr += 16;
-            }
-            if (out_elempack == 8)
-            {
-                _mm256_store_ps(outptr, _mm512_extractf32x8_ps(_sum0, 0));
-                _mm256_store_ps(outptr + M, _mm512_extractf32x8_ps(_sum0, 1));
-                outptr += 8;
-            }
-            if (out_elempack == 4)
-            {
-                _mm_store_ps(outptr, _mm512_extractf32x4_ps(_sum0, 0));
-                _mm_store_ps(outptr + M, _mm512_extractf32x4_ps(_sum0, 1));
-                _mm_store_ps(outptr + M * 2, _mm512_extractf32x4_ps(_sum0, 2));
-                _mm_store_ps(outptr + M * 3, _mm512_extractf32x4_ps(_sum0, 3));
-                outptr += 4;
-            }
-            if (out_elempack == 1)
-            {
-                float sum[16];
-                _mm512_storeu_ps(sum, _sum0);
-
-                outptr[0] = sum[0];
-                outptr[M] = sum[1];
-                outptr[M * 2] = sum[2];
-                outptr[M * 3] = sum[3];
-                outptr[M * 4] = sum[4];
-                outptr[M * 5] = sum[5];
-                outptr[M * 6] = sum[6];
-                outptr[M * 7] = sum[7];
-                outptr[M * 8] = sum[8];
-                outptr[M * 9] = sum[9];
-                outptr[M * 10] = sum[10];
-                outptr[M * 11] = sum[11];
-                outptr[M * 12] = sum[12];
-                outptr[M * 13] = sum[13];
-                outptr[M * 14] = sum[14];
-                outptr[M * 15] = sum[15];
-                outptr += 1;
-            }
-        }
-    }
-    remain_outh_start += nn_outh * 16;
-    nn_outh = (outh - remain_outh_start) / 8;
-#else // __AVX512F__
-    nn_outh = (outh - remain_outh_start) / 8;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __AVX512F__
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = remain_outh_start + pp * 8;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-        const int out_elempack = top_blob.elempack;
-
-        float* outptr = top_blob.row(p / out_elempack);
-
-        for (int j = 0; j < outw; j++)
-        {
-            __m256 _sum0 = _mm256_setzero_ps();
-            __m256 _sum1 = _mm256_setzero_ps();
-            __m256 _sum2 = _mm256_setzero_ps();
-            __m256 _sum3 = _mm256_setzero_ps();
-
-            if (bias_data_ptr)
-            {
-                _sum0 = _mm256_loadu_ps(bias_data_ptr + p);
-            }
-
-#if __AVX512F__
-            const float* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8);
-#else
-            const float* kptr = weight_data_tm.channel(p / 8);
-#endif
-
-            int q = 0;
-#if __AVX512F__
-            for (; q + 15 < inh; q += 16)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                if (elempack == 16)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _w0 = _mm256_load_ps(kptr + 8 * 0);
-                        __m256 _w1 = _mm256_load_ps(kptr + 8 * 1);
-                        __m256 _w2 = _mm256_load_ps(kptr + 8 * 2);
-                        __m256 _w3 = _mm256_load_ps(kptr + 8 * 3);
-                        __m256 _w4 = _mm256_load_ps(kptr + 8 * 4);
-                        __m256 _w5 = _mm256_load_ps(kptr + 8 * 5);
-                        __m256 _w6 = _mm256_load_ps(kptr + 8 * 6);
-                        __m256 _w7 = _mm256_load_ps(kptr + 8 * 7);
-                        __m256 _w8 = _mm256_load_ps(kptr + 8 * 8);
-                        __m256 _w9 = _mm256_load_ps(kptr + 8 * 9);
-                        __m256 _wa = _mm256_load_ps(kptr + 8 * 10);
-                        __m256 _wb = _mm256_load_ps(kptr + 8 * 11);
-                        __m256 _wc = _mm256_load_ps(kptr + 8 * 12);
-                        __m256 _wd = _mm256_load_ps(kptr + 8 * 13);
-                        __m256 _we = _mm256_load_ps(kptr + 8 * 14);
-                        __m256 _wf = _mm256_load_ps(kptr + 8 * 15);
-
-                        _sum0 = _mm256_fmadd_ps(_w0, _mm256_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_w1, _mm256_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_w2, _mm256_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_w3, _mm256_set1_ps(r0[3]), _sum3);
-                        _sum0 = _mm256_fmadd_ps(_w4, _mm256_set1_ps(r0[4]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_w5, _mm256_set1_ps(r0[5]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_w6, _mm256_set1_ps(r0[6]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_w7, _mm256_set1_ps(r0[7]), _sum3);
-                        _sum0 = _mm256_fmadd_ps(_w8, _mm256_set1_ps(r0[8]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_w9, _mm256_set1_ps(r0[9]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_wa, _mm256_set1_ps(r0[10]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_wb, _mm256_set1_ps(r0[11]), _sum3);
-                        _sum0 = _mm256_fmadd_ps(_wc, _mm256_set1_ps(r0[12]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_wd, _mm256_set1_ps(r0[13]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_we, _mm256_set1_ps(r0[14]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_wf, _mm256_set1_ps(r0[15]), _sum3);
-
-                        r0 += dilation_w * 16;
-                        kptr += 128;
-                    }
-                }
-                if (elempack == 8)
-                {
-                    const float* r1 = r0 + N;
-
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _w0 = _mm256_load_ps(kptr + 8 * 0);
-                        __m256 _w1 = _mm256_load_ps(kptr + 8 * 1);
-                        __m256 _w2 = _mm256_load_ps(kptr + 8 * 2);
-                        __m256 _w3 = _mm256_load_ps(kptr + 8 * 3);
-                        __m256 _w4 = _mm256_load_ps(kptr + 8 * 4);
-                        __m256 _w5 = _mm256_load_ps(kptr + 8 * 5);
-                        __m256 _w6 = _mm256_load_ps(kptr + 8 * 6);
-                        __m256 _w7 = _mm256_load_ps(kptr + 8 * 7);
-                        __m256 _w8 = _mm256_load_ps(kptr + 8 * 8);
-                        __m256 _w9 = _mm256_load_ps(kptr + 8 * 9);
-                        __m256 _wa = _mm256_load_ps(kptr + 8 * 10);
-                        __m256 _wb = _mm256_load_ps(kptr + 8 * 11);
-                        __m256 _wc = _mm256_load_ps(kptr + 8 * 12);
-                        __m256 _wd = _mm256_load_ps(kptr + 8 * 13);
-                        __m256 _we = _mm256_load_ps(kptr + 8 * 14);
-                        __m256 _wf = _mm256_load_ps(kptr + 8 * 15);
-
-                        _sum0 = _mm256_fmadd_ps(_w0, _mm256_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_w1, _mm256_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_w2, _mm256_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_w3, _mm256_set1_ps(r0[3]), _sum3);
-                        _sum0 = _mm256_fmadd_ps(_w4, _mm256_set1_ps(r0[4]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_w5, _mm256_set1_ps(r0[5]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_w6, _mm256_set1_ps(r0[6]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_w7, _mm256_set1_ps(r0[7]), _sum3);
-                        _sum0 = _mm256_fmadd_ps(_w8, _mm256_set1_ps(r1[0]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_w9, _mm256_set1_ps(r1[1]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_wa, _mm256_set1_ps(r1[2]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_wb, _mm256_set1_ps(r1[3]), _sum3);
-                        _sum0 = _mm256_fmadd_ps(_wc, _mm256_set1_ps(r1[4]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_wd, _mm256_set1_ps(r1[5]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_we, _mm256_set1_ps(r1[6]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_wf, _mm256_set1_ps(r1[7]), _sum3);
-
-                        r0 += dilation_w * 8;
-                        r1 += dilation_w * 8;
-                        kptr += 128;
-                    }
-                }
-                if (elempack == 4)
-                {
-                    const float* r1 = r0 + N;
-                    const float* r2 = r0 + N * 2;
-                    const float* r3 = r0 + N * 3;
-
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _w0 = _mm256_load_ps(kptr + 8 * 0);
-                        __m256 _w1 = _mm256_load_ps(kptr + 8 * 1);
-                        __m256 _w2 = _mm256_load_ps(kptr + 8 * 2);
-                        __m256 _w3 = _mm256_load_ps(kptr + 8 * 3);
-                        __m256 _w4 = _mm256_load_ps(kptr + 8 * 4);
-                        __m256 _w5 = _mm256_load_ps(kptr + 8 * 5);
-                        __m256 _w6 = _mm256_load_ps(kptr + 8 * 6);
-                        __m256 _w7 = _mm256_load_ps(kptr + 8 * 7);
-                        __m256 _w8 = _mm256_load_ps(kptr + 8 * 8);
-                        __m256 _w9 = _mm256_load_ps(kptr + 8 * 9);
-                        __m256 _wa = _mm256_load_ps(kptr + 8 * 10);
-                        __m256 _wb = _mm256_load_ps(kptr + 8 * 11);
-                        __m256 _wc = _mm256_load_ps(kptr + 8 * 12);
-                        __m256 _wd = _mm256_load_ps(kptr + 8 * 13);
-                        __m256 _we = _mm256_load_ps(kptr + 8 * 14);
-                        __m256 _wf = _mm256_load_ps(kptr + 8 * 15);
-
-                        _sum0 = _mm256_fmadd_ps(_w0, _mm256_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_w1, _mm256_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_w2, _mm256_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_w3, _mm256_set1_ps(r0[3]), _sum3);
-                        _sum0 = _mm256_fmadd_ps(_w4, _mm256_set1_ps(r1[0]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_w5, _mm256_set1_ps(r1[1]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_w6, _mm256_set1_ps(r1[2]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_w7, _mm256_set1_ps(r1[3]), _sum3);
-                        _sum0 = _mm256_fmadd_ps(_w8, _mm256_set1_ps(r2[0]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_w9, _mm256_set1_ps(r2[1]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_wa, _mm256_set1_ps(r2[2]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_wb, _mm256_set1_ps(r2[3]), _sum3);
-                        _sum0 = _mm256_fmadd_ps(_wc, _mm256_set1_ps(r3[0]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_wd, _mm256_set1_ps(r3[1]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_we, _mm256_set1_ps(r3[2]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_wf, _mm256_set1_ps(r3[3]), _sum3);
-
-                        r0 += dilation_w * 4;
-                        r1 += dilation_w * 4;
-                        r2 += dilation_w * 4;
-                        r3 += dilation_w * 4;
-                        kptr += 128;
-                    }
-                }
-                if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _w0 = _mm256_load_ps(kptr + 8 * 0);
-                        __m256 _w1 = _mm256_load_ps(kptr + 8 * 1);
-                        __m256 _w2 = _mm256_load_ps(kptr + 8 * 2);
-                        __m256 _w3 = _mm256_load_ps(kptr + 8 * 3);
-                        __m256 _w4 = _mm256_load_ps(kptr + 8 * 4);
-                        __m256 _w5 = _mm256_load_ps(kptr + 8 * 5);
-                        __m256 _w6 = _mm256_load_ps(kptr + 8 * 6);
-                        __m256 _w7 = _mm256_load_ps(kptr + 8 * 7);
-                        __m256 _w8 = _mm256_load_ps(kptr + 8 * 8);
-                        __m256 _w9 = _mm256_load_ps(kptr + 8 * 9);
-                        __m256 _wa = _mm256_load_ps(kptr + 8 * 10);
-                        __m256 _wb = _mm256_load_ps(kptr + 8 * 11);
-                        __m256 _wc = _mm256_load_ps(kptr + 8 * 12);
-                        __m256 _wd = _mm256_load_ps(kptr + 8 * 13);
-                        __m256 _we = _mm256_load_ps(kptr + 8 * 14);
-                        __m256 _wf = _mm256_load_ps(kptr + 8 * 15);
-
-                        _sum0 = _mm256_fmadd_ps(_w0, _mm256_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_w1, _mm256_set1_ps(r0[N]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_w2, _mm256_set1_ps(r0[N * 2]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_w3, _mm256_set1_ps(r0[N * 3]), _sum3);
-                        _sum0 = _mm256_fmadd_ps(_w4, _mm256_set1_ps(r0[N * 4]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_w5, _mm256_set1_ps(r0[N * 5]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_w6, _mm256_set1_ps(r0[N * 6]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_w7, _mm256_set1_ps(r0[N * 7]), _sum3);
-                        _sum0 = _mm256_fmadd_ps(_w8, _mm256_set1_ps(r0[N * 8]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_w9, _mm256_set1_ps(r0[N * 9]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_wa, _mm256_set1_ps(r0[N * 10]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_wb, _mm256_set1_ps(r0[N * 11]), _sum3);
-                        _sum0 = _mm256_fmadd_ps(_wc, _mm256_set1_ps(r0[N * 12]), _sum0);
-                        _sum1 = _mm256_fmadd_ps(_wd, _mm256_set1_ps(r0[N * 13]), _sum1);
-                        _sum2 = _mm256_fmadd_ps(_we, _mm256_set1_ps(r0[N * 14]), _sum2);
-                        _sum3 = _mm256_fmadd_ps(_wf, _mm256_set1_ps(r0[N * 15]), _sum3);
-
-                        r0 += dilation_w;
-                        kptr += 128;
-                    }
-                }
-            }
-#endif // __AVX512F__
-            for (; q + 7 < inh; q += 8)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                if (elempack == 8)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _w0 = _mm256_load_ps(kptr);
-                        __m256 _w1 = _mm256_load_ps(kptr + 8);
-                        __m256 _w2 = _mm256_load_ps(kptr + 16);
-                        __m256 _w3 = _mm256_load_ps(kptr + 24);
-                        __m256 _w4 = _mm256_load_ps(kptr + 32);
-                        __m256 _w5 = _mm256_load_ps(kptr + 40);
-                        __m256 _w6 = _mm256_load_ps(kptr + 48);
-                        __m256 _w7 = _mm256_load_ps(kptr + 56);
-
-                        _sum0 = _mm256_comp_fmadd_ps(_w0, _mm256_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm256_comp_fmadd_ps(_w1, _mm256_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm256_comp_fmadd_ps(_w2, _mm256_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm256_comp_fmadd_ps(_w3, _mm256_set1_ps(r0[3]), _sum3);
-                        _sum0 = _mm256_comp_fmadd_ps(_w4, _mm256_set1_ps(r0[4]), _sum0);
-                        _sum1 = _mm256_comp_fmadd_ps(_w5, _mm256_set1_ps(r0[5]), _sum1);
-                        _sum2 = _mm256_comp_fmadd_ps(_w6, _mm256_set1_ps(r0[6]), _sum2);
-                        _sum3 = _mm256_comp_fmadd_ps(_w7, _mm256_set1_ps(r0[7]), _sum3);
-
-                        r0 += dilation_w * 8;
-                        kptr += 64;
-                    }
-                }
-                if (elempack == 4)
-                {
-                    const float* r1 = r0 + N;
-
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _w0 = _mm256_load_ps(kptr);
-                        __m256 _w1 = _mm256_load_ps(kptr + 8);
-                        __m256 _w2 = _mm256_load_ps(kptr + 16);
-                        __m256 _w3 = _mm256_load_ps(kptr + 24);
-                        __m256 _w4 = _mm256_load_ps(kptr + 32);
-                        __m256 _w5 = _mm256_load_ps(kptr + 40);
-                        __m256 _w6 = _mm256_load_ps(kptr + 48);
-                        __m256 _w7 = _mm256_load_ps(kptr + 56);
-
-                        _sum0 = _mm256_comp_fmadd_ps(_w0, _mm256_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm256_comp_fmadd_ps(_w1, _mm256_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm256_comp_fmadd_ps(_w2, _mm256_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm256_comp_fmadd_ps(_w3, _mm256_set1_ps(r0[3]), _sum3);
-                        _sum0 = _mm256_comp_fmadd_ps(_w4, _mm256_set1_ps(r1[0]), _sum0);
-                        _sum1 = _mm256_comp_fmadd_ps(_w5, _mm256_set1_ps(r1[1]), _sum1);
-                        _sum2 = _mm256_comp_fmadd_ps(_w6, _mm256_set1_ps(r1[2]), _sum2);
-                        _sum3 = _mm256_comp_fmadd_ps(_w7, _mm256_set1_ps(r1[3]), _sum3);
-
-                        r0 += dilation_w * 4;
-                        r1 += dilation_w * 4;
-                        kptr += 64;
-                    }
-                }
-                if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _w0 = _mm256_load_ps(kptr);
-                        __m256 _w1 = _mm256_load_ps(kptr + 8);
-                        __m256 _w2 = _mm256_load_ps(kptr + 16);
-                        __m256 _w3 = _mm256_load_ps(kptr + 24);
-                        __m256 _w4 = _mm256_load_ps(kptr + 32);
-                        __m256 _w5 = _mm256_load_ps(kptr + 40);
-                        __m256 _w6 = _mm256_load_ps(kptr + 48);
-                        __m256 _w7 = _mm256_load_ps(kptr + 56);
-
-                        _sum0 = _mm256_comp_fmadd_ps(_w0, _mm256_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm256_comp_fmadd_ps(_w1, _mm256_set1_ps(r0[N]), _sum1);
-                        _sum2 = _mm256_comp_fmadd_ps(_w2, _mm256_set1_ps(r0[N * 2]), _sum2);
-                        _sum3 = _mm256_comp_fmadd_ps(_w3, _mm256_set1_ps(r0[N * 3]), _sum3);
-                        _sum0 = _mm256_comp_fmadd_ps(_w4, _mm256_set1_ps(r0[N * 4]), _sum0);
-                        _sum1 = _mm256_comp_fmadd_ps(_w5, _mm256_set1_ps(r0[N * 5]), _sum1);
-                        _sum2 = _mm256_comp_fmadd_ps(_w6, _mm256_set1_ps(r0[N * 6]), _sum2);
-                        _sum3 = _mm256_comp_fmadd_ps(_w7, _mm256_set1_ps(r0[N * 7]), _sum3);
-
-                        r0 += dilation_w;
-                        kptr += 64;
-                    }
-                }
-            }
-            for (; q + 3 < inh; q += 4)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                if (elempack == 4)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _w0 = _mm256_load_ps(kptr);
-                        __m256 _w1 = _mm256_load_ps(kptr + 8);
-                        __m256 _w2 = _mm256_load_ps(kptr + 16);
-                        __m256 _w3 = _mm256_load_ps(kptr + 24);
-
-                        _sum0 = _mm256_comp_fmadd_ps(_w0, _mm256_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm256_comp_fmadd_ps(_w1, _mm256_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm256_comp_fmadd_ps(_w2, _mm256_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm256_comp_fmadd_ps(_w3, _mm256_set1_ps(r0[3]), _sum3);
-
-                        r0 += dilation_w * 4;
-                        kptr += 32;
-                    }
-                }
-                if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _w0 = _mm256_load_ps(kptr);
-                        __m256 _w1 = _mm256_load_ps(kptr + 8);
-                        __m256 _w2 = _mm256_load_ps(kptr + 16);
-                        __m256 _w3 = _mm256_load_ps(kptr + 24);
-
-                        _sum0 = _mm256_comp_fmadd_ps(_w0, _mm256_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm256_comp_fmadd_ps(_w1, _mm256_set1_ps(r0[N]), _sum1);
-                        _sum2 = _mm256_comp_fmadd_ps(_w2, _mm256_set1_ps(r0[N * 2]), _sum2);
-                        _sum3 = _mm256_comp_fmadd_ps(_w3, _mm256_set1_ps(r0[N * 3]), _sum3);
-
-                        r0 += dilation_w;
-                        kptr += 32;
-                    }
-                }
-            }
-            for (; q + 1 < inh; q += 2)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                // if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _w0 = _mm256_load_ps(kptr);
-                        __m256 _w1 = _mm256_load_ps(kptr + 8);
-
-                        _sum0 = _mm256_comp_fmadd_ps(_w0, _mm256_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm256_comp_fmadd_ps(_w1, _mm256_set1_ps(r0[N]), _sum1);
-
-                        r0 += dilation_w;
-                        kptr += 16;
-                    }
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                // if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _val = _mm256_set1_ps(r0[0]);
-                        __m256 _w = _mm256_load_ps(kptr);
-                        _sum0 = _mm256_comp_fmadd_ps(_val, _w, _sum0);
-
-                        r0 += dilation_w;
-                        kptr += 8;
-                    }
-                }
-            }
-
-            _sum0 = _mm256_add_ps(_sum0, _sum1);
-            _sum2 = _mm256_add_ps(_sum2, _sum3);
-            _sum0 = _mm256_add_ps(_sum0, _sum2);
-
-            _sum0 = activation_avx(_sum0, activation_type, activation_params);
-
-            if (out_elempack == 8)
-            {
-                _mm256_store_ps(outptr, _sum0);
-                outptr += 8;
-            }
-            if (out_elempack == 4)
-            {
-                _mm_store_ps(outptr, _mm256_extractf128_ps(_sum0, 0));
-                _mm_store_ps(outptr + M, _mm256_extractf128_ps(_sum0, 1));
-                outptr += 4;
-            }
-            if (out_elempack == 1)
-            {
-                float sum[8];
-                _mm256_storeu_ps(sum, _sum0);
-
-                outptr[0] = sum[0];
-                outptr[M] = sum[1];
-                outptr[M * 2] = sum[2];
-                outptr[M * 3] = sum[3];
-                outptr[M * 4] = sum[4];
-                outptr[M * 5] = sum[5];
-                outptr[M * 6] = sum[6];
-                outptr[M * 7] = sum[7];
-                outptr += 1;
-            }
-        }
-    }
-    remain_outh_start += nn_outh * 8;
-    nn_outh = (outh - remain_outh_start) / 4;
-#else // __AVX__
-    nn_outh = (outh - remain_outh_start) / 4;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __AVX__
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = remain_outh_start + pp * 4;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-        const int out_elempack = top_blob.elempack;
-
-        float* outptr = top_blob.row(p / out_elempack);
-
-        for (int j = 0; j < outw; j++)
-        {
-            __m128 _sum0 = _mm_setzero_ps();
-            __m128 _sum1 = _mm_setzero_ps();
-            __m128 _sum2 = _mm_setzero_ps();
-            __m128 _sum3 = _mm_setzero_ps();
-
-            if (bias_data_ptr)
-            {
-                _sum0 = _mm_loadu_ps(bias_data_ptr + p);
-            }
-
-#if __AVX512F__
-            const float* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8 + (p % 8) / 4);
-#elif __AVX__
-            const float* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4);
-#else
-            const float* kptr = weight_data_tm.channel(p / 4);
-#endif
-
-            int q = 0;
-#if __AVX__
-#if __AVX512F__
-            for (; q + 15 < inh; q += 16)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                if (elempack == 16)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m128 _w0 = _mm_load_ps(kptr + 4 * 0);
-                        __m128 _w1 = _mm_load_ps(kptr + 4 * 1);
-                        __m128 _w2 = _mm_load_ps(kptr + 4 * 2);
-                        __m128 _w3 = _mm_load_ps(kptr + 4 * 3);
-                        __m128 _w4 = _mm_load_ps(kptr + 4 * 4);
-                        __m128 _w5 = _mm_load_ps(kptr + 4 * 5);
-                        __m128 _w6 = _mm_load_ps(kptr + 4 * 6);
-                        __m128 _w7 = _mm_load_ps(kptr + 4 * 7);
-                        __m128 _w8 = _mm_load_ps(kptr + 4 * 8);
-                        __m128 _w9 = _mm_load_ps(kptr + 4 * 9);
-                        __m128 _wa = _mm_load_ps(kptr + 4 * 10);
-                        __m128 _wb = _mm_load_ps(kptr + 4 * 11);
-                        __m128 _wc = _mm_load_ps(kptr + 4 * 12);
-                        __m128 _wd = _mm_load_ps(kptr + 4 * 13);
-                        __m128 _we = _mm_load_ps(kptr + 4 * 14);
-                        __m128 _wf = _mm_load_ps(kptr + 4 * 15);
-
-                        _sum0 = _mm_fmadd_ps(_w0, _mm_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_w1, _mm_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_w2, _mm_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_w3, _mm_set1_ps(r0[3]), _sum3);
-                        _sum0 = _mm_fmadd_ps(_w4, _mm_set1_ps(r0[4]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_w5, _mm_set1_ps(r0[5]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_w6, _mm_set1_ps(r0[6]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_w7, _mm_set1_ps(r0[7]), _sum3);
-                        _sum0 = _mm_fmadd_ps(_w8, _mm_set1_ps(r0[8]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_w9, _mm_set1_ps(r0[9]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_wa, _mm_set1_ps(r0[10]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_wb, _mm_set1_ps(r0[11]), _sum3);
-                        _sum0 = _mm_fmadd_ps(_wc, _mm_set1_ps(r0[12]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_wd, _mm_set1_ps(r0[13]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_we, _mm_set1_ps(r0[14]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_wf, _mm_set1_ps(r0[15]), _sum3);
-
-                        r0 += dilation_w * 16;
-                        kptr += 64;
-                    }
-                }
-                if (elempack == 8)
-                {
-                    const float* r1 = r0 + N;
-
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m128 _w0 = _mm_load_ps(kptr + 4 * 0);
-                        __m128 _w1 = _mm_load_ps(kptr + 4 * 1);
-                        __m128 _w2 = _mm_load_ps(kptr + 4 * 2);
-                        __m128 _w3 = _mm_load_ps(kptr + 4 * 3);
-                        __m128 _w4 = _mm_load_ps(kptr + 4 * 4);
-                        __m128 _w5 = _mm_load_ps(kptr + 4 * 5);
-                        __m128 _w6 = _mm_load_ps(kptr + 4 * 6);
-                        __m128 _w7 = _mm_load_ps(kptr + 4 * 7);
-                        __m128 _w8 = _mm_load_ps(kptr + 4 * 8);
-                        __m128 _w9 = _mm_load_ps(kptr + 4 * 9);
-                        __m128 _wa = _mm_load_ps(kptr + 4 * 10);
-                        __m128 _wb = _mm_load_ps(kptr + 4 * 11);
-                        __m128 _wc = _mm_load_ps(kptr + 4 * 12);
-                        __m128 _wd = _mm_load_ps(kptr + 4 * 13);
-                        __m128 _we = _mm_load_ps(kptr + 4 * 14);
-                        __m128 _wf = _mm_load_ps(kptr + 4 * 15);
-
-                        _sum0 = _mm_fmadd_ps(_w0, _mm_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_w1, _mm_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_w2, _mm_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_w3, _mm_set1_ps(r0[3]), _sum3);
-                        _sum0 = _mm_fmadd_ps(_w4, _mm_set1_ps(r0[4]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_w5, _mm_set1_ps(r0[5]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_w6, _mm_set1_ps(r0[6]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_w7, _mm_set1_ps(r0[7]), _sum3);
-                        _sum0 = _mm_fmadd_ps(_w8, _mm_set1_ps(r1[0]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_w9, _mm_set1_ps(r1[1]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_wa, _mm_set1_ps(r1[2]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_wb, _mm_set1_ps(r1[3]), _sum3);
-                        _sum0 = _mm_fmadd_ps(_wc, _mm_set1_ps(r1[4]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_wd, _mm_set1_ps(r1[5]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_we, _mm_set1_ps(r1[6]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_wf, _mm_set1_ps(r1[7]), _sum3);
-
-                        r0 += dilation_w * 8;
-                        r1 += dilation_w * 8;
-                        kptr += 64;
-                    }
-                }
-                if (elempack == 4)
-                {
-                    const float* r1 = r0 + N;
-                    const float* r2 = r0 + N * 2;
-                    const float* r3 = r0 + N * 3;
-
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m128 _w0 = _mm_load_ps(kptr + 4 * 0);
-                        __m128 _w1 = _mm_load_ps(kptr + 4 * 1);
-                        __m128 _w2 = _mm_load_ps(kptr + 4 * 2);
-                        __m128 _w3 = _mm_load_ps(kptr + 4 * 3);
-                        __m128 _w4 = _mm_load_ps(kptr + 4 * 4);
-                        __m128 _w5 = _mm_load_ps(kptr + 4 * 5);
-                        __m128 _w6 = _mm_load_ps(kptr + 4 * 6);
-                        __m128 _w7 = _mm_load_ps(kptr + 4 * 7);
-                        __m128 _w8 = _mm_load_ps(kptr + 4 * 8);
-                        __m128 _w9 = _mm_load_ps(kptr + 4 * 9);
-                        __m128 _wa = _mm_load_ps(kptr + 4 * 10);
-                        __m128 _wb = _mm_load_ps(kptr + 4 * 11);
-                        __m128 _wc = _mm_load_ps(kptr + 4 * 12);
-                        __m128 _wd = _mm_load_ps(kptr + 4 * 13);
-                        __m128 _we = _mm_load_ps(kptr + 4 * 14);
-                        __m128 _wf = _mm_load_ps(kptr + 4 * 15);
-
-                        _sum0 = _mm_fmadd_ps(_w0, _mm_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_w1, _mm_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_w2, _mm_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_w3, _mm_set1_ps(r0[3]), _sum3);
-                        _sum0 = _mm_fmadd_ps(_w4, _mm_set1_ps(r1[0]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_w5, _mm_set1_ps(r1[1]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_w6, _mm_set1_ps(r1[2]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_w7, _mm_set1_ps(r1[3]), _sum3);
-                        _sum0 = _mm_fmadd_ps(_w8, _mm_set1_ps(r2[0]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_w9, _mm_set1_ps(r2[1]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_wa, _mm_set1_ps(r2[2]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_wb, _mm_set1_ps(r2[3]), _sum3);
-                        _sum0 = _mm_fmadd_ps(_wc, _mm_set1_ps(r3[0]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_wd, _mm_set1_ps(r3[1]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_we, _mm_set1_ps(r3[2]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_wf, _mm_set1_ps(r3[3]), _sum3);
-
-                        r0 += dilation_w * 4;
-                        r1 += dilation_w * 4;
-                        r2 += dilation_w * 4;
-                        r3 += dilation_w * 4;
-                        kptr += 64;
-                    }
-                }
-                if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m128 _w0 = _mm_load_ps(kptr + 4 * 0);
-                        __m128 _w1 = _mm_load_ps(kptr + 4 * 1);
-                        __m128 _w2 = _mm_load_ps(kptr + 4 * 2);
-                        __m128 _w3 = _mm_load_ps(kptr + 4 * 3);
-                        __m128 _w4 = _mm_load_ps(kptr + 4 * 4);
-                        __m128 _w5 = _mm_load_ps(kptr + 4 * 5);
-                        __m128 _w6 = _mm_load_ps(kptr + 4 * 6);
-                        __m128 _w7 = _mm_load_ps(kptr + 4 * 7);
-                        __m128 _w8 = _mm_load_ps(kptr + 4 * 8);
-                        __m128 _w9 = _mm_load_ps(kptr + 4 * 9);
-                        __m128 _wa = _mm_load_ps(kptr + 4 * 10);
-                        __m128 _wb = _mm_load_ps(kptr + 4 * 11);
-                        __m128 _wc = _mm_load_ps(kptr + 4 * 12);
-                        __m128 _wd = _mm_load_ps(kptr + 4 * 13);
-                        __m128 _we = _mm_load_ps(kptr + 4 * 14);
-                        __m128 _wf = _mm_load_ps(kptr + 4 * 15);
-
-                        _sum0 = _mm_fmadd_ps(_w0, _mm_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_w1, _mm_set1_ps(r0[N]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_w2, _mm_set1_ps(r0[N * 2]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_w3, _mm_set1_ps(r0[N * 3]), _sum3);
-                        _sum0 = _mm_fmadd_ps(_w4, _mm_set1_ps(r0[N * 4]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_w5, _mm_set1_ps(r0[N * 5]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_w6, _mm_set1_ps(r0[N * 6]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_w7, _mm_set1_ps(r0[N * 7]), _sum3);
-                        _sum0 = _mm_fmadd_ps(_w8, _mm_set1_ps(r0[N * 8]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_w9, _mm_set1_ps(r0[N * 9]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_wa, _mm_set1_ps(r0[N * 10]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_wb, _mm_set1_ps(r0[N * 11]), _sum3);
-                        _sum0 = _mm_fmadd_ps(_wc, _mm_set1_ps(r0[N * 12]), _sum0);
-                        _sum1 = _mm_fmadd_ps(_wd, _mm_set1_ps(r0[N * 13]), _sum1);
-                        _sum2 = _mm_fmadd_ps(_we, _mm_set1_ps(r0[N * 14]), _sum2);
-                        _sum3 = _mm_fmadd_ps(_wf, _mm_set1_ps(r0[N * 15]), _sum3);
-
-                        r0 += dilation_w;
-                        kptr += 64;
-                    }
-                }
-            }
-#endif // __AVX512F__
-            for (; q + 7 < inh; q += 8)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                if (elempack == 8)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m128 _w0 = _mm_load_ps(kptr);
-                        __m128 _w1 = _mm_load_ps(kptr + 4);
-                        __m128 _w2 = _mm_load_ps(kptr + 8);
-                        __m128 _w3 = _mm_load_ps(kptr + 12);
-                        __m128 _w4 = _mm_load_ps(kptr + 16);
-                        __m128 _w5 = _mm_load_ps(kptr + 20);
-                        __m128 _w6 = _mm_load_ps(kptr + 24);
-                        __m128 _w7 = _mm_load_ps(kptr + 28);
-
-                        _sum0 = _mm_comp_fmadd_ps(_w0, _mm_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm_comp_fmadd_ps(_w1, _mm_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm_comp_fmadd_ps(_w2, _mm_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm_comp_fmadd_ps(_w3, _mm_set1_ps(r0[3]), _sum3);
-                        _sum0 = _mm_comp_fmadd_ps(_w4, _mm_set1_ps(r0[4]), _sum0);
-                        _sum1 = _mm_comp_fmadd_ps(_w5, _mm_set1_ps(r0[5]), _sum1);
-                        _sum2 = _mm_comp_fmadd_ps(_w6, _mm_set1_ps(r0[6]), _sum2);
-                        _sum3 = _mm_comp_fmadd_ps(_w7, _mm_set1_ps(r0[7]), _sum3);
-
-                        r0 += dilation_w * 8;
-                        kptr += 32;
-                    }
-                }
-                if (elempack == 4)
-                {
-                    const float* r1 = r0 + N;
-
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m128 _w0 = _mm_load_ps(kptr);
-                        __m128 _w1 = _mm_load_ps(kptr + 4);
-                        __m128 _w2 = _mm_load_ps(kptr + 8);
-                        __m128 _w3 = _mm_load_ps(kptr + 12);
-                        __m128 _w4 = _mm_load_ps(kptr + 16);
-                        __m128 _w5 = _mm_load_ps(kptr + 20);
-                        __m128 _w6 = _mm_load_ps(kptr + 24);
-                        __m128 _w7 = _mm_load_ps(kptr + 28);
-
-                        _sum0 = _mm_comp_fmadd_ps(_w0, _mm_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm_comp_fmadd_ps(_w1, _mm_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm_comp_fmadd_ps(_w2, _mm_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm_comp_fmadd_ps(_w3, _mm_set1_ps(r0[3]), _sum3);
-                        _sum0 = _mm_comp_fmadd_ps(_w4, _mm_set1_ps(r1[0]), _sum0);
-                        _sum1 = _mm_comp_fmadd_ps(_w5, _mm_set1_ps(r1[1]), _sum1);
-                        _sum2 = _mm_comp_fmadd_ps(_w6, _mm_set1_ps(r1[2]), _sum2);
-                        _sum3 = _mm_comp_fmadd_ps(_w7, _mm_set1_ps(r1[3]), _sum3);
-
-                        r0 += dilation_w * 4;
-                        r1 += dilation_w * 4;
-                        kptr += 32;
-                    }
-                }
-                if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m128 _w0 = _mm_load_ps(kptr);
-                        __m128 _w1 = _mm_load_ps(kptr + 4);
-                        __m128 _w2 = _mm_load_ps(kptr + 8);
-                        __m128 _w3 = _mm_load_ps(kptr + 12);
-                        __m128 _w4 = _mm_load_ps(kptr + 16);
-                        __m128 _w5 = _mm_load_ps(kptr + 20);
-                        __m128 _w6 = _mm_load_ps(kptr + 24);
-                        __m128 _w7 = _mm_load_ps(kptr + 28);
-
-                        _sum0 = _mm_comp_fmadd_ps(_w0, _mm_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm_comp_fmadd_ps(_w1, _mm_set1_ps(r0[N]), _sum1);
-                        _sum2 = _mm_comp_fmadd_ps(_w2, _mm_set1_ps(r0[N * 2]), _sum2);
-                        _sum3 = _mm_comp_fmadd_ps(_w3, _mm_set1_ps(r0[N * 3]), _sum3);
-                        _sum0 = _mm_comp_fmadd_ps(_w4, _mm_set1_ps(r0[N * 4]), _sum0);
-                        _sum1 = _mm_comp_fmadd_ps(_w5, _mm_set1_ps(r0[N * 5]), _sum1);
-                        _sum2 = _mm_comp_fmadd_ps(_w6, _mm_set1_ps(r0[N * 6]), _sum2);
-                        _sum3 = _mm_comp_fmadd_ps(_w7, _mm_set1_ps(r0[N * 7]), _sum3);
-
-                        r0 += dilation_w;
-                        kptr += 32;
-                    }
-                }
-            }
-#endif // __AVX__
-            for (; q + 3 < inh; q += 4)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                if (elempack == 4)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m128 _w0 = _mm_load_ps(kptr);
-                        __m128 _w1 = _mm_load_ps(kptr + 4);
-                        __m128 _w2 = _mm_load_ps(kptr + 8);
-                        __m128 _w3 = _mm_load_ps(kptr + 12);
-
-                        _sum0 = _mm_comp_fmadd_ps(_w0, _mm_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm_comp_fmadd_ps(_w1, _mm_set1_ps(r0[1]), _sum1);
-                        _sum2 = _mm_comp_fmadd_ps(_w2, _mm_set1_ps(r0[2]), _sum2);
-                        _sum3 = _mm_comp_fmadd_ps(_w3, _mm_set1_ps(r0[3]), _sum3);
-
-                        r0 += dilation_w * 4;
-                        kptr += 16;
-                    }
-                }
-                if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m128 _w0 = _mm_load_ps(kptr);
-                        __m128 _w1 = _mm_load_ps(kptr + 4);
-                        __m128 _w2 = _mm_load_ps(kptr + 8);
-                        __m128 _w3 = _mm_load_ps(kptr + 12);
-
-                        _sum0 = _mm_comp_fmadd_ps(_w0, _mm_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm_comp_fmadd_ps(_w1, _mm_set1_ps(r0[N]), _sum1);
-                        _sum2 = _mm_comp_fmadd_ps(_w2, _mm_set1_ps(r0[N * 2]), _sum2);
-                        _sum3 = _mm_comp_fmadd_ps(_w3, _mm_set1_ps(r0[N * 3]), _sum3);
-
-                        r0 += dilation_w;
-                        kptr += 16;
-                    }
-                }
-            }
-            for (; q + 1 < inh; q += 2)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                // if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m128 _w0 = _mm_load_ps(kptr);
-                        __m128 _w1 = _mm_load_ps(kptr + 4);
-
-                        _sum0 = _mm_comp_fmadd_ps(_w0, _mm_set1_ps(r0[0]), _sum0);
-                        _sum1 = _mm_comp_fmadd_ps(_w1, _mm_set1_ps(r0[N]), _sum1);
-
-                        r0 += dilation_w;
-                        kptr += 8;
-                    }
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                // if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m128 _val = _mm_set1_ps(r0[0]);
-                        __m128 _w = _mm_load_ps(kptr);
-                        _sum0 = _mm_comp_fmadd_ps(_val, _w, _sum0);
-
-                        r0 += dilation_w;
-                        kptr += 4;
-                    }
-                }
-            }
-
-            _sum0 = _mm_add_ps(_sum0, _sum1);
-            _sum2 = _mm_add_ps(_sum2, _sum3);
-            _sum0 = _mm_add_ps(_sum0, _sum2);
-
-            _sum0 = activation_sse(_sum0, activation_type, activation_params);
-
-            if (out_elempack == 4)
-            {
-                _mm_storeu_ps(outptr, _sum0);
-                outptr += 4;
-            }
-            if (out_elempack == 1)
-            {
-                float sum[4];
-                _mm_storeu_ps(sum, _sum0);
-
-                outptr[0] = sum[0];
-                outptr[M] = sum[1];
-                outptr[M * 2] = sum[2];
-                outptr[M * 3] = sum[3];
-                outptr += 1;
-            }
-        }
-    }
-    remain_outh_start += nn_outh * 4;
-    nn_outh = (outh - remain_outh_start) / 2;
-#else // __SSE2__
-    nn_outh = (outh - remain_outh_start) / 2;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __SSE2__
-    for (int pp = 0; pp < nn_outh; pp++)
-    {
-        const int p = remain_outh_start + pp * 2;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inh = bottom_blob.h * elempack;
-        const int outw = top_blob.w;
-
-        float* outptr0 = top_blob.row(p);
-        float* outptr1 = top_blob.row(p + 1);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float sum0 = 0.f;
-            float sum1 = 0.f;
-
-            if (bias_data_ptr)
-            {
-                sum0 = bias_data_ptr[p];
-                sum1 = bias_data_ptr[p + 1];
-            }
-
-#if __AVX512F__
-            const float* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8 + (p % 8) / 4 + (p % 4) / 2);
-#elif __AVX__
-            const float* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2);
-#elif __SSE2__
-            const float* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2);
-#else
-            const float* kptr = weight_data_tm.channel(p / 2);
-#endif
-
-            int q = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-            __m512 _sum0_avx512 = _mm512_setzero_ps();
-            __m512 _sum1_avx512 = _mm512_setzero_ps();
-            for (; q + 15 < inh; q += 16)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                if (elempack == 16)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _r0 = _mm512_load_ps(r0);
-                        __m512 _w0 = _mm512_load_ps(kptr);
-                        __m512 _w1 = _mm512_load_ps(kptr + 16);
-                        _sum0_avx512 = _mm512_fmadd_ps(_r0, _w0, _sum0_avx512);
-                        _sum1_avx512 = _mm512_fmadd_ps(_r0, _w1, _sum1_avx512);
-
-                        r0 += dilation_w * 16;
-                        kptr += 32;
-                    }
-                }
-                if (elempack == 8)
-                {
-                    const float* r1 = r0 + N;
-
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0)), _mm256_load_ps(r1), 1);
-                        __m512 _w0 = _mm512_load_ps(kptr);
-                        __m512 _w1 = _mm512_load_ps(kptr + 16);
-                        _sum0_avx512 = _mm512_fmadd_ps(_r0, _w0, _sum0_avx512);
-                        _sum1_avx512 = _mm512_fmadd_ps(_r0, _w1, _sum1_avx512);
-
-                        r0 += dilation_w * 8;
-                        r1 += dilation_w * 8;
-                        kptr += 32;
-                    }
-                }
-                if (elempack == 4)
-                {
-                    const float* r1 = r0 + N;
-                    const float* r2 = r0 + N * 2;
-                    const float* r3 = r0 + N * 3;
-
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0)), _mm_load_ps(r1), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2)), _mm_load_ps(r3), 1), 1);
-                        __m512 _w0 = _mm512_load_ps(kptr);
-                        __m512 _w1 = _mm512_load_ps(kptr + 16);
-                        _sum0_avx512 = _mm512_fmadd_ps(_r0, _w0, _sum0_avx512);
-                        _sum1_avx512 = _mm512_fmadd_ps(_r0, _w1, _sum1_avx512);
-
-                        r0 += dilation_w * 4;
-                        r1 += dilation_w * 4;
-                        r2 += dilation_w * 4;
-                        r3 += dilation_w * 4;
-                        kptr += 32;
-                    }
-                }
-                if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _r0 = _mm512_set_ps(r0[N * 15], r0[N * 14], r0[N * 13], r0[N * 12], r0[N * 11], r0[N * 10], r0[N * 9], r0[N * 8], r0[N * 7], r0[N * 6], r0[N * 5], r0[N * 4], r0[N * 3], r0[N * 2], r0[N], r0[0]);
-                        __m512 _w0 = _mm512_load_ps(kptr);
-                        __m512 _w1 = _mm512_load_ps(kptr + 16);
-                        _sum0_avx512 = _mm512_fmadd_ps(_r0, _w0, _sum0_avx512);
-                        _sum1_avx512 = _mm512_fmadd_ps(_r0, _w1, _sum1_avx512);
-
-                        r0 += dilation_w;
-                        kptr += 32;
-                    }
-                }
-            }
-            sum0 += _mm512_comp_reduce_add_ps(_sum0_avx512);
-            sum1 += _mm512_comp_reduce_add_ps(_sum1_avx512);
-#endif // __AVX512F__
-            __m256 _sum0_avx = _mm256_setzero_ps();
-            __m256 _sum1_avx = _mm256_setzero_ps();
-            for (; q + 7 < inh; q += 8)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                if (elempack == 8)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _r0 = _mm256_load_ps(r0);
-                        __m256 _w0 = _mm256_load_ps(kptr);
-                        __m256 _w1 = _mm256_load_ps(kptr + 8);
-                        _sum0_avx = _mm256_comp_fmadd_ps(_r0, _w0, _sum0_avx);
-                        _sum1_avx = _mm256_comp_fmadd_ps(_r0, _w1, _sum1_avx);
-
-                        r0 += dilation_w * 8;
-                        kptr += 16;
-                    }
-                }
-                if (elempack == 4)
-                {
-                    const float* r1 = r0 + N;
-
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _r0 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0)), _mm_load_ps(r1), 1);
-                        __m256 _w0 = _mm256_load_ps(kptr);
-                        __m256 _w1 = _mm256_load_ps(kptr + 8);
-                        _sum0_avx = _mm256_comp_fmadd_ps(_r0, _w0, _sum0_avx);
-                        _sum1_avx = _mm256_comp_fmadd_ps(_r0, _w1, _sum1_avx);
-
-                        r0 += dilation_w * 4;
-                        r1 += dilation_w * 4;
-                        kptr += 16;
-                    }
-                }
-                if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _r0 = _mm256_set_ps(r0[N * 7], r0[N * 6], r0[N * 5], r0[N * 4], r0[N * 3], r0[N * 2], r0[N], r0[0]);
-                        __m256 _w0 = _mm256_load_ps(kptr);
-                        __m256 _w1 = _mm256_load_ps(kptr + 8);
-                        _sum0_avx = _mm256_comp_fmadd_ps(_r0, _w0, _sum0_avx);
-                        _sum1_avx = _mm256_comp_fmadd_ps(_r0, _w1, _sum1_avx);
-
-                        r0 += dilation_w;
-                        kptr += 16;
-                    }
-                }
-            }
-            sum0 += _mm256_reduce_add_ps(_sum0_avx);
-            sum1 += _mm256_reduce_add_ps(_sum1_avx);
-#endif // __AVX__
-            __m128 _sum0 = _mm_setzero_ps();
-            __m128 _sum1 = _mm_setzero_ps();
-            for (; q + 3 < inh; q += 4)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                if (elempack == 4)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m128 _r0 = _mm_load_ps(r0);
-                        __m128 _w0 = _mm_load_ps(kptr);
-                        __m128 _w1 = _mm_load_ps(kptr + 4);
-                        _sum0 = _mm_comp_fmadd_ps(_r0, _w0, _sum0);
-                        _sum1 = _mm_comp_fmadd_ps(_r0, _w1, _sum1);
-
-                        r0 += dilation_w * 4;
-                        kptr += 8;
-                    }
-                }
-                if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m128 _r0 = _mm_set_ps(r0[N * 3], r0[N * 2], r0[N], r0[0]);
-                        __m128 _w0 = _mm_load_ps(kptr);
-                        __m128 _w1 = _mm_load_ps(kptr + 4);
-                        _sum0 = _mm_comp_fmadd_ps(_r0, _w0, _sum0);
-                        _sum1 = _mm_comp_fmadd_ps(_r0, _w1, _sum1);
-
-                        r0 += dilation_w;
-                        kptr += 8;
-                    }
-                }
-            }
-            sum0 += _mm_reduce_add_ps(_sum0);
-            sum1 += _mm_reduce_add_ps(_sum1);
-#endif // __SSE2__
-            for (; q + 1 < inh; q += 2)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                // if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        sum0 += r0[0] * kptr[0];
-                        sum1 += r0[0] * kptr[1];
-                        sum0 += r0[N] * kptr[2];
-                        sum1 += r0[N] * kptr[3];
-
-                        r0 += dilation_w;
-                        kptr += 4;
-                    }
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                // if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        float val = r0[0];
-                        sum0 += val * kptr[0];
-                        sum1 += val * kptr[1];
-
-                        r0 += dilation_w;
-                        kptr += 2;
-                    }
-                }
-            }
-
-            sum0 = activation_ss(sum0, activation_type, activation_params);
-            sum1 = activation_ss(sum1, activation_type, activation_params);
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-            outptr0 += 1;
-            outptr1 += 1;
-        }
-    }
-    remain_outh_start += nn_outh * 2;
-    for (int p = remain_outh_start; p < outh; p++)
-    {
-        float* outptr = top_blob.row(p);
-
-        for (int j = 0; j < outw; j++)
-        {
-            float sum = 0.f;
-
-            if (bias_data_ptr)
-            {
-                sum = bias_data_ptr[p];
-            }
-
-#if __AVX512F__
-            const float* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#elif __AVX__
-            const float* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#elif __SSE2__
-            const float* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2 + p % 2);
-#else
-            const float* kptr = weight_data_tm.channel(p / 2 + p % 2);
-#endif
-
-            int q = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-            __m512 _sum_avx512 = _mm512_setzero_ps();
-            for (; q + 15 < inh; q += 16)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                if (elempack == 16)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _r0 = _mm512_load_ps(r0);
-                        __m512 _w = _mm512_load_ps(kptr);
-                        _sum_avx512 = _mm512_fmadd_ps(_r0, _w, _sum_avx512);
-
-                        r0 += dilation_w * 16;
-                        kptr += 16;
-                    }
-                }
-                if (elempack == 8)
-                {
-                    const float* r1 = r0 + N;
-
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0)), _mm256_load_ps(r1), 1);
-                        __m512 _w = _mm512_load_ps(kptr);
-                        _sum_avx512 = _mm512_fmadd_ps(_r0, _w, _sum_avx512);
-
-                        r0 += dilation_w * 8;
-                        r1 += dilation_w * 8;
-                        kptr += 16;
-                    }
-                }
-                if (elempack == 4)
-                {
-                    const float* r1 = r0 + N;
-                    const float* r2 = r0 + N * 2;
-                    const float* r3 = r0 + N * 3;
-
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0)), _mm_load_ps(r1), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2)), _mm_load_ps(r3), 1), 1);
-                        __m512 _w = _mm512_load_ps(kptr);
-                        _sum_avx512 = _mm512_fmadd_ps(_r0, _w, _sum_avx512);
-
-                        r0 += dilation_w * 4;
-                        r1 += dilation_w * 4;
-                        r2 += dilation_w * 4;
-                        r3 += dilation_w * 4;
-                        kptr += 16;
-                    }
-                }
-                if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m512 _r0 = _mm512_set_ps(r0[N * 15], r0[N * 14], r0[N * 13], r0[N * 12], r0[N * 11], r0[N * 10], r0[N * 9], r0[N * 8], r0[N * 7], r0[N * 6], r0[N * 5], r0[N * 4], r0[N * 3], r0[N * 2], r0[N], r0[0]);
-                        __m512 _w = _mm512_load_ps(kptr);
-                        _sum_avx512 = _mm512_fmadd_ps(_r0, _w, _sum_avx512);
-
-                        r0 += dilation_w;
-                        kptr += 16;
-                    }
-                }
-            }
-            sum += _mm512_comp_reduce_add_ps(_sum_avx512);
-#endif // __AVX512F__
-            __m256 _sum_avx = _mm256_setzero_ps();
-            for (; q + 7 < inh; q += 8)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                if (elempack == 8)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _r0 = _mm256_load_ps(r0);
-                        __m256 _w = _mm256_load_ps(kptr);
-                        _sum_avx = _mm256_comp_fmadd_ps(_r0, _w, _sum_avx);
-
-                        r0 += dilation_w * 8;
-                        kptr += 8;
-                    }
-                }
-                if (elempack == 4)
-                {
-                    const float* r1 = r0 + N;
-
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _r0 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0)), _mm_load_ps(r1), 1);
-                        __m256 _w = _mm256_load_ps(kptr);
-                        _sum_avx = _mm256_comp_fmadd_ps(_r0, _w, _sum_avx);
-
-                        r0 += dilation_w * 4;
-                        r1 += dilation_w * 4;
-                        kptr += 8;
-                    }
-                }
-                if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m256 _r0 = _mm256_set_ps(r0[N * 7], r0[N * 6], r0[N * 5], r0[N * 4], r0[N * 3], r0[N * 2], r0[N], r0[0]);
-                        __m256 _w = _mm256_load_ps(kptr);
-                        _sum_avx = _mm256_comp_fmadd_ps(_r0, _w, _sum_avx);
-
-                        r0 += dilation_w;
-                        kptr += 8;
-                    }
-                }
-            }
-            sum += _mm256_reduce_add_ps(_sum_avx);
-#endif // __AVX__
-            __m128 _sum = _mm_setzero_ps();
-            for (; q + 3 < inh; q += 4)
-            {
-                const float* r0 = bottom_blob.row(q / elempack) + j * stride_w * elempack;
-
-                if (elempack == 4)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m128 _r0 = _mm_load_ps(r0);
-                        __m128 _w = _mm_load_ps(kptr);
-                        _sum = _mm_comp_fmadd_ps(_r0, _w, _sum);
-
-                        r0 += dilation_w * 4;
-                        kptr += 4;
-                    }
-                }
-                if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        __m128 _r0 = _mm_set_ps(r0[N * 3], r0[N * 2], r0[N], r0[0]);
-                        __m128 _w = _mm_load_ps(kptr);
-                        _sum = _mm_comp_fmadd_ps(_r0, _w, _sum);
-
-                        r0 += dilation_w;
-                        kptr += 4;
-                    }
-                }
-            }
-            sum += _mm_reduce_add_ps(_sum);
-#endif // __SSE2__
-            for (; q + 1 < inh; q += 2)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                // if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        sum += r0[0] * kptr[0];
-                        sum += r0[N] * kptr[1];
-
-                        r0 += dilation_w;
-                        kptr += 2;
-                    }
-                }
-            }
-            for (; q < inh; q++)
-            {
-                const float* r0 = bottom_blob.row(q) + j * stride_w;
-
-                // if (elempack == 1)
-                {
-                    for (int k = 0; k < kernel_w; k++)
-                    {
-                        float val = r0[0];
-                        sum += val * kptr[0];
-
-                        r0 += dilation_w;
-                        kptr += 1;
-                    }
-                }
-            }
-
-            sum = activation_ss(sum, activation_type, activation_params);
-
-            outptr[0] = sum;
-            outptr += 1;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution1d_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution1d_x86.h
deleted file mode 100644
index 497b34e5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution1d_x86.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTION1D_X86_H
-#define LAYER_CONVOLUTION1D_X86_H
-
-#include "convolution1d.h"
-
-namespace ncnn {
-
-class Convolution1D_x86 : public Convolution1D
-{
-public:
-    Convolution1D_x86();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    Mat weight_data_tm;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTION1D_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_1x1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_1x1.h
deleted file mode 100644
index acf0f8c0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_1x1.h
+++ /dev/null
@@ -1,207 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv1x1s1_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        int q = 0;
-
-        for (; q + 3 < inch; q += 4)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-            const float* img1 = bottom_blob.channel(q + 1);
-            const float* img2 = bottom_blob.channel(q + 2);
-            const float* img3 = bottom_blob.channel(q + 3);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float k0 = kernel0[0];
-            const float k1 = kernel0[1];
-            const float k2 = kernel0[2];
-            const float k3 = kernel0[3];
-
-            const float* r0 = img0;
-            const float* r1 = img1;
-            const float* r2 = img2;
-            const float* r3 = img3;
-
-            int size = outw * outh;
-
-            int remain = size;
-
-            for (; remain > 0; remain--)
-            {
-                float sum = *r0 * k0;
-                float sum1 = *r1 * k1;
-                float sum2 = *r2 * k2;
-                float sum3 = *r3 * k3;
-
-                *outptr += sum + sum1 + sum2 + sum3;
-
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                outptr++;
-            }
-        }
-
-        for (; q < inch; q++)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float k0 = kernel0[0];
-
-            const float* r0 = img0;
-
-            int size = outw * outh;
-
-            int remain = size;
-
-            for (; remain > 0; remain--)
-            {
-                float sum = *r0 * k0;
-
-                *outptr += sum;
-
-                r0++;
-                outptr++;
-            }
-        }
-    }
-}
-
-static void conv1x1s2_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        int q = 0;
-
-        for (; q + 3 < inch; q += 4)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-            const float* img1 = bottom_blob.channel(q + 1);
-            const float* img2 = bottom_blob.channel(q + 2);
-            const float* img3 = bottom_blob.channel(q + 3);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float k0 = kernel0[0];
-            const float k1 = kernel0[1];
-            const float k2 = kernel0[2];
-            const float k3 = kernel0[3];
-
-            const float* r0 = img0;
-            const float* r1 = img1;
-            const float* r2 = img2;
-            const float* r3 = img3;
-
-            for (int i = 0; i < outh; i++)
-            {
-                int remain = outw;
-
-                for (; remain > 0; remain--)
-                {
-                    float sum = *r0 * k0;
-                    float sum1 = *r1 * k1;
-                    float sum2 = *r2 * k2;
-                    float sum3 = *r3 * k3;
-
-                    *outptr += sum + sum1 + sum2 + sum3;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    r3 += 2;
-                    outptr++;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-                r3 += tailstep;
-            }
-        }
-
-        for (; q < inch; q++)
-        {
-            float* outptr = out;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch + q;
-            const float k0 = kernel0[0];
-
-            const float* r0 = img0;
-
-            for (int i = 0; i < outh; i++)
-            {
-                int remain = outw;
-
-                for (; remain > 0; remain--)
-                {
-                    float sum = *r0 * k0;
-
-                    *outptr += sum;
-
-                    r0 += 2;
-                    outptr++;
-                }
-
-                r0 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_2x2_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_2x2_pack8.h
deleted file mode 100644
index 24812b25..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_2x2_pack8.h
+++ /dev/null
@@ -1,384 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv2x2s1_pack8_avx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        __m256 _bias0 = bias ? _mm256_loadu_ps((const float*)bias + p * 8) : _mm256_set1_ps(0.f);
-        out0.fill(_bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-
-            const float* kptr = (const float*)kernel.channel(p).row(q);
-            // const float* kptr = (const float*)kernel + 4 * inch * p * 64;
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-
-                for (; j + 1 < outw; j += 2)
-                {
-                    __m256 _sum0 = _mm256_loadu_ps(outptr0);
-                    __m256 _sum1 = _mm256_loadu_ps(outptr0 + 8);
-
-                    __m256 _r00 = _mm256_broadcast_ss(r0);
-                    __m256 _r01 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 3);
-                    __m256 _r04 = _mm256_broadcast_ss(r0 + 4);
-                    __m256 _r05 = _mm256_broadcast_ss(r0 + 5);
-                    __m256 _r06 = _mm256_broadcast_ss(r0 + 6);
-                    __m256 _r07 = _mm256_broadcast_ss(r0 + 7);
-                    r0 += 8;
-
-                    __m256 _k00 = _mm256_loadu_ps(kptr);
-                    __m256 _k01 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k02 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k03 = _mm256_loadu_ps(kptr + 24);
-                    kptr += 32;
-
-                    _sum0 = _mm256_comp_fmadd_ps(_k00, _r00, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k01, _r01, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k02, _r02, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k03, _r03, _sum0);
-
-                    __m256 _k04 = _mm256_loadu_ps(kptr);
-                    __m256 _k05 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k06 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k07 = _mm256_loadu_ps(kptr + 24);
-                    kptr += 32;
-
-                    _sum0 = _mm256_comp_fmadd_ps(_k04, _r04, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k05, _r05, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k06, _r06, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k07, _r07, _sum0);
-
-                    //========================================
-
-                    _r00 = _mm256_broadcast_ss(r0);
-                    _r01 = _mm256_broadcast_ss(r0 + 1);
-                    _r02 = _mm256_broadcast_ss(r0 + 2);
-                    _r03 = _mm256_broadcast_ss(r0 + 3);
-                    _r04 = _mm256_broadcast_ss(r0 + 4);
-                    _r05 = _mm256_broadcast_ss(r0 + 5);
-                    _r06 = _mm256_broadcast_ss(r0 + 6);
-                    _r07 = _mm256_broadcast_ss(r0 + 7);
-                    r0 += 8;
-
-                    _sum1 = _mm256_comp_fmadd_ps(_k00, _r00, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k01, _r01, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k02, _r02, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k03, _r03, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k04, _r04, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k05, _r05, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k06, _r06, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k07, _r07, _sum1);
-
-                    _k00 = _mm256_loadu_ps(kptr);
-                    _k01 = _mm256_loadu_ps(kptr + 8);
-                    _k02 = _mm256_loadu_ps(kptr + 16);
-                    _k03 = _mm256_loadu_ps(kptr + 24);
-                    kptr += 32;
-
-                    _sum0 = _mm256_comp_fmadd_ps(_k00, _r00, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k01, _r01, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k02, _r02, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k03, _r03, _sum0);
-
-                    _k04 = _mm256_loadu_ps(kptr);
-                    _k05 = _mm256_loadu_ps(kptr + 8);
-                    _k06 = _mm256_loadu_ps(kptr + 16);
-                    _k07 = _mm256_loadu_ps(kptr + 24);
-                    kptr += 32;
-
-                    _sum0 = _mm256_comp_fmadd_ps(_k04, _r04, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k05, _r05, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k06, _r06, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k07, _r07, _sum0);
-
-                    _r00 = _mm256_broadcast_ss(r0);
-                    _r01 = _mm256_broadcast_ss(r0 + 1);
-                    _r02 = _mm256_broadcast_ss(r0 + 2);
-                    _r03 = _mm256_broadcast_ss(r0 + 3);
-                    _r04 = _mm256_broadcast_ss(r0 + 4);
-                    _r05 = _mm256_broadcast_ss(r0 + 5);
-                    _r06 = _mm256_broadcast_ss(r0 + 6);
-                    _r07 = _mm256_broadcast_ss(r0 + 7);
-
-                    _sum1 = _mm256_comp_fmadd_ps(_k00, _r00, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k01, _r01, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k02, _r02, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k03, _r03, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k04, _r04, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k05, _r05, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k06, _r06, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k07, _r07, _sum1);
-                    //===============
-
-                    __m256 _r10 = _mm256_broadcast_ss(r1);
-                    __m256 _r11 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 3);
-                    __m256 _r14 = _mm256_broadcast_ss(r1 + 4);
-                    __m256 _r15 = _mm256_broadcast_ss(r1 + 5);
-                    __m256 _r16 = _mm256_broadcast_ss(r1 + 6);
-                    __m256 _r17 = _mm256_broadcast_ss(r1 + 7);
-
-                    __m256 _k10 = _mm256_loadu_ps(kptr);
-                    __m256 _k11 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k12 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k13 = _mm256_loadu_ps(kptr + 24);
-                    kptr += 32;
-
-                    _sum0 = _mm256_comp_fmadd_ps(_k10, _r10, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k11, _r11, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k12, _r12, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k13, _r13, _sum0);
-
-                    __m256 _k14 = _mm256_loadu_ps(kptr);
-                    __m256 _k15 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k16 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k17 = _mm256_loadu_ps(kptr + 24);
-                    kptr += 32;
-
-                    _sum0 = _mm256_comp_fmadd_ps(_k14, _r14, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k15, _r15, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k16, _r16, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k17, _r17, _sum0);
-
-                    //=======================================
-                    r1 += 8;
-                    _r10 = _mm256_broadcast_ss(r1);
-                    _r11 = _mm256_broadcast_ss(r1 + 1);
-                    _r12 = _mm256_broadcast_ss(r1 + 2);
-                    _r13 = _mm256_broadcast_ss(r1 + 3);
-                    _r14 = _mm256_broadcast_ss(r1 + 4);
-                    _r15 = _mm256_broadcast_ss(r1 + 5);
-                    _r16 = _mm256_broadcast_ss(r1 + 6);
-                    _r17 = _mm256_broadcast_ss(r1 + 7);
-
-                    _sum1 = _mm256_comp_fmadd_ps(_k10, _r10, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k11, _r11, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k12, _r12, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k13, _r13, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k14, _r14, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k15, _r15, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k16, _r16, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k17, _r17, _sum1);
-
-                    _k10 = _mm256_loadu_ps(kptr);
-                    _k11 = _mm256_loadu_ps(kptr + 8);
-                    _k12 = _mm256_loadu_ps(kptr + 16);
-                    _k13 = _mm256_loadu_ps(kptr + 24);
-                    kptr += 32;
-
-                    _sum0 = _mm256_comp_fmadd_ps(_k10, _r10, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k11, _r11, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k12, _r12, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k13, _r13, _sum0);
-
-                    _k14 = _mm256_loadu_ps(kptr);
-                    _k15 = _mm256_loadu_ps(kptr + 8);
-                    _k16 = _mm256_loadu_ps(kptr + 16);
-                    _k17 = _mm256_loadu_ps(kptr + 24);
-                    _sum0 = _mm256_comp_fmadd_ps(_k14, _r14, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k15, _r15, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k16, _r16, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_k17, _r17, _sum0);
-
-                    r1 += 8;
-                    _r10 = _mm256_broadcast_ss(r1);
-                    _r11 = _mm256_broadcast_ss(r1 + 1);
-                    _r12 = _mm256_broadcast_ss(r1 + 2);
-                    _r13 = _mm256_broadcast_ss(r1 + 3);
-                    _r14 = _mm256_broadcast_ss(r1 + 4);
-                    _r15 = _mm256_broadcast_ss(r1 + 5);
-                    _r16 = _mm256_broadcast_ss(r1 + 6);
-                    _r17 = _mm256_broadcast_ss(r1 + 7);
-
-                    _sum1 = _mm256_comp_fmadd_ps(_k10, _r10, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k11, _r11, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k12, _r12, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k13, _r13, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k14, _r14, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k15, _r15, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k16, _r16, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_k17, _r17, _sum1);
-
-                    kptr -= 224;
-                    _mm256_storeu_ps(outptr0, _sum0);
-                    _mm256_storeu_ps(outptr0 + 8, _sum1);
-                    outptr0 += 16;
-                }
-
-                for (; j < outw; j++)
-                {
-                    __m256 _sum = _mm256_loadu_ps(outptr0);
-
-                    __m256 _r00 = _mm256_broadcast_ss(r0);
-                    __m256 _r01 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 3);
-                    __m256 _r04 = _mm256_broadcast_ss(r0 + 4);
-                    __m256 _r05 = _mm256_broadcast_ss(r0 + 5);
-                    __m256 _r06 = _mm256_broadcast_ss(r0 + 6);
-                    __m256 _r07 = _mm256_broadcast_ss(r0 + 7);
-
-                    __m256 _k00 = _mm256_loadu_ps(kptr);
-                    __m256 _k01 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k02 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k03 = _mm256_loadu_ps(kptr + 24);
-                    kptr += 32;
-
-                    _sum = _mm256_comp_fmadd_ps(_k00, _r00, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k01, _r01, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k02, _r02, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k03, _r03, _sum);
-
-                    __m256 _k04 = _mm256_loadu_ps(kptr);
-                    __m256 _k05 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k06 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k07 = _mm256_loadu_ps(kptr + 24);
-                    kptr += 32;
-
-                    _sum = _mm256_comp_fmadd_ps(_k04, _r04, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k05, _r05, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k06, _r06, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k07, _r07, _sum);
-
-                    //========================================
-                    r0 += 8;
-                    _r00 = _mm256_broadcast_ss(r0);
-                    _r01 = _mm256_broadcast_ss(r0 + 1);
-                    _r02 = _mm256_broadcast_ss(r0 + 2);
-                    _r03 = _mm256_broadcast_ss(r0 + 3);
-                    _r04 = _mm256_broadcast_ss(r0 + 4);
-                    _r05 = _mm256_broadcast_ss(r0 + 5);
-                    _r06 = _mm256_broadcast_ss(r0 + 6);
-                    _r07 = _mm256_broadcast_ss(r0 + 7);
-
-                    _k00 = _mm256_loadu_ps(kptr);
-                    _k01 = _mm256_loadu_ps(kptr + 8);
-                    _k02 = _mm256_loadu_ps(kptr + 16);
-                    _k03 = _mm256_loadu_ps(kptr + 24);
-                    kptr += 32;
-
-                    _sum = _mm256_comp_fmadd_ps(_k00, _r00, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k01, _r01, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k02, _r02, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k03, _r03, _sum);
-
-                    _k04 = _mm256_loadu_ps(kptr);
-                    _k05 = _mm256_loadu_ps(kptr + 8);
-                    _k06 = _mm256_loadu_ps(kptr + 16);
-                    _k07 = _mm256_loadu_ps(kptr + 24);
-                    kptr += 32;
-
-                    _sum = _mm256_comp_fmadd_ps(_k04, _r04, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k05, _r05, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k06, _r06, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k07, _r07, _sum);
-                    //===============
-
-                    __m256 _r10 = _mm256_broadcast_ss(r1);
-                    __m256 _r11 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 3);
-                    __m256 _r14 = _mm256_broadcast_ss(r1 + 4);
-                    __m256 _r15 = _mm256_broadcast_ss(r1 + 5);
-                    __m256 _r16 = _mm256_broadcast_ss(r1 + 6);
-                    __m256 _r17 = _mm256_broadcast_ss(r1 + 7);
-
-                    __m256 _k10 = _mm256_loadu_ps(kptr);
-                    __m256 _k11 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k12 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k13 = _mm256_loadu_ps(kptr + 24);
-                    kptr += 32;
-
-                    _sum = _mm256_comp_fmadd_ps(_k10, _r10, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k11, _r11, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k12, _r12, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k13, _r13, _sum);
-
-                    __m256 _k14 = _mm256_loadu_ps(kptr);
-                    __m256 _k15 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k16 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k17 = _mm256_loadu_ps(kptr + 24);
-                    kptr += 32;
-
-                    _sum = _mm256_comp_fmadd_ps(_k14, _r14, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k15, _r15, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k16, _r16, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k17, _r17, _sum);
-
-                    //=======================================
-                    r1 += 8;
-                    _r10 = _mm256_broadcast_ss(r1);
-                    _r11 = _mm256_broadcast_ss(r1 + 1);
-                    _r12 = _mm256_broadcast_ss(r1 + 2);
-                    _r13 = _mm256_broadcast_ss(r1 + 3);
-                    _r14 = _mm256_broadcast_ss(r1 + 4);
-                    _r15 = _mm256_broadcast_ss(r1 + 5);
-                    _r16 = _mm256_broadcast_ss(r1 + 6);
-                    _r17 = _mm256_broadcast_ss(r1 + 7);
-
-                    _k10 = _mm256_loadu_ps(kptr);
-                    _k11 = _mm256_loadu_ps(kptr + 8);
-                    _k12 = _mm256_loadu_ps(kptr + 16);
-                    _k13 = _mm256_loadu_ps(kptr + 24);
-                    kptr += 32;
-
-                    _sum = _mm256_comp_fmadd_ps(_k10, _r10, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k11, _r11, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k12, _r12, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k13, _r13, _sum);
-
-                    _k14 = _mm256_loadu_ps(kptr);
-                    _k15 = _mm256_loadu_ps(kptr + 8);
-                    _k16 = _mm256_loadu_ps(kptr + 16);
-                    _k17 = _mm256_loadu_ps(kptr + 24);
-                    _sum = _mm256_comp_fmadd_ps(_k14, _r14, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k15, _r15, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k16, _r16, _sum);
-                    _sum = _mm256_comp_fmadd_ps(_k17, _r17, _sum);
-
-                    kptr -= 224;
-                    _mm256_storeu_ps(outptr0, _sum);
-                    outptr0 += 8;
-                }
-
-                r0 += 8;
-                r1 += 8;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3.h
deleted file mode 100644
index caa60fd5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3.h
+++ /dev/null
@@ -1,209 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr = out;
-            float* outptr2 = outptr + outw;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch * 9 + q * 9;
-
-            const float* r0 = img0;
-            const float* r1 = img0 + w;
-            const float* r2 = img0 + w * 2;
-            const float* r3 = img0 + w * 3;
-
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            int i = 0;
-
-            for (; i + 1 < outh; i += 2)
-            {
-                int remain = outw;
-
-                for (; remain > 0; remain--)
-                {
-                    float sum = 0;
-                    float sum2 = 0;
-
-                    sum += r0[0] * k0[0];
-                    sum += r0[1] * k0[1];
-                    sum += r0[2] * k0[2];
-                    sum += r1[0] * k1[0];
-                    sum += r1[1] * k1[1];
-                    sum += r1[2] * k1[2];
-                    sum += r2[0] * k2[0];
-                    sum += r2[1] * k2[1];
-                    sum += r2[2] * k2[2];
-
-                    sum2 += r1[0] * k0[0];
-                    sum2 += r1[1] * k0[1];
-                    sum2 += r1[2] * k0[2];
-                    sum2 += r2[0] * k1[0];
-                    sum2 += r2[1] * k1[1];
-                    sum2 += r2[2] * k1[2];
-                    sum2 += r3[0] * k2[0];
-                    sum2 += r3[1] * k2[1];
-                    sum2 += r3[2] * k2[2];
-
-                    *outptr += sum;
-                    *outptr2 += sum2;
-
-                    r0++;
-                    r1++;
-                    r2++;
-                    r3++;
-                    outptr++;
-                    outptr2++;
-                }
-
-                r0 += 2 + w;
-                r1 += 2 + w;
-                r2 += 2 + w;
-                r3 += 2 + w;
-
-                outptr += outw;
-                outptr2 += outw;
-            }
-
-            for (; i < outh; i++)
-            {
-                int remain = outw;
-
-                for (; remain > 0; remain--)
-                {
-                    float sum = 0;
-
-                    sum += r0[0] * k0[0];
-                    sum += r0[1] * k0[1];
-                    sum += r0[2] * k0[2];
-                    sum += r1[0] * k1[0];
-                    sum += r1[1] * k1[1];
-                    sum += r1[2] * k1[2];
-                    sum += r2[0] * k2[0];
-                    sum += r2[1] * k2[1];
-                    sum += r2[2] * k2[2];
-
-                    *outptr += sum;
-
-                    r0++;
-                    r1++;
-                    r2++;
-                    outptr++;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-        }
-    }
-}
-
-static void conv3x3s2_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr = out;
-
-            const float* img = bottom_blob.channel(q);
-            const float* kernel0 = kernel + p * inch * 9 + q * 9;
-
-            const float* r0 = img;
-            const float* r1 = img + w;
-            const float* r2 = img + w * 2;
-
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 3;
-            const float* k2 = kernel0 + 6;
-
-            for (int i = 0; i < outh; i++)
-            {
-                int remain = outw;
-
-                for (; remain > 0; remain--)
-                {
-                    float sum = 0;
-
-                    sum += r0[0] * k0[0];
-                    sum += r0[1] * k0[1];
-                    sum += r0[2] * k0[2];
-                    sum += r1[0] * k1[0];
-                    sum += r1[1] * k1[1];
-                    sum += r1[2] * k1[2];
-                    sum += r2[0] * k2[0];
-                    sum += r2[1] * k2[1];
-                    sum += r2[2] * k2[2];
-
-                    *outptr += sum;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr++;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_int8.h
deleted file mode 100644
index ceaf75b9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_int8.h
+++ /dev/null
@@ -1,147 +0,0 @@
-// BUG1989 is pleased to support the open source community by supporting ncnn available.
-//
-// Copyright (C) 2019 BUG1989. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_int8_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const signed char* kernel = _kernel;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        out0.fill(0);
-
-        const signed char* kernel0 = (const signed char*)kernel + p * inch * 9;
-
-        for (int q = 0; q < inch; q++)
-        {
-            int* outptr0 = out0;
-
-            const signed char* img0 = bottom_blob.channel(q);
-
-            const signed char* r0 = img0;
-            const signed char* r1 = img0 + w;
-            const signed char* r2 = img0 + w * 2;
-
-            for (int i = 0; i < outh; i++)
-            {
-                int remain = outw;
-
-                for (; remain > 0; remain--)
-                {
-                    int sum0 = 0;
-
-                    sum0 += (int)r0[0] * kernel0[0];
-                    sum0 += (int)r0[1] * kernel0[1];
-                    sum0 += (int)r0[2] * kernel0[2];
-                    sum0 += (int)r1[0] * kernel0[3];
-                    sum0 += (int)r1[1] * kernel0[4];
-                    sum0 += (int)r1[2] * kernel0[5];
-                    sum0 += (int)r2[0] * kernel0[6];
-                    sum0 += (int)r2[1] * kernel0[7];
-                    sum0 += (int)r2[2] * kernel0[8];
-
-                    *outptr0 += sum0;
-
-                    r0++;
-                    r1++;
-                    r2++;
-                    outptr0++;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            kernel0 += 9;
-        }
-    }
-}
-
-static void conv3x3s2_int8_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const signed char* kernel = _kernel;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        out0.fill(0);
-
-        const signed char* kernel0 = (const signed char*)kernel + p * inch * 9;
-
-        for (int q = 0; q < inch; q++)
-        {
-            int* outptr0 = out0;
-
-            const signed char* img0 = bottom_blob.channel(q);
-
-            const signed char* r0 = img0;
-            const signed char* r1 = img0 + w;
-            const signed char* r2 = img0 + w * 2;
-
-            for (int i = 0; i < outh; i++)
-            {
-                int remain = outw;
-
-                for (; remain > 0; remain--)
-                {
-                    int sum0 = 0;
-
-                    sum0 += (int)r0[0] * kernel0[0];
-                    sum0 += (int)r0[1] * kernel0[1];
-                    sum0 += (int)r0[2] * kernel0[2];
-                    sum0 += (int)r1[0] * kernel0[3];
-                    sum0 += (int)r1[1] * kernel0[4];
-                    sum0 += (int)r1[2] * kernel0[5];
-                    sum0 += (int)r2[0] * kernel0[6];
-                    sum0 += (int)r2[1] * kernel0[7];
-                    sum0 += (int)r2[2] * kernel0[8];
-
-                    *outptr0 += sum0;
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0++;
-                }
-
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            kernel0 += 9;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack16to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack16to1.h
deleted file mode 100644
index fd6eb2dc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack16to1.h
+++ /dev/null
@@ -1,100 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack16to1_avx512(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    int remain_outch_start = 0;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        out0.fill(bias0);
-
-        const float* k0 = kernel.channel(p);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            __m512 _k00 = _mm512_loadu_ps(k0);
-            __m512 _k01 = _mm512_loadu_ps(k0 + 16);
-            __m512 _k02 = _mm512_loadu_ps(k0 + 16 * 2);
-            __m512 _k10 = _mm512_loadu_ps(k0 + 16 * 3);
-            __m512 _k11 = _mm512_loadu_ps(k0 + 16 * 4);
-            __m512 _k12 = _mm512_loadu_ps(k0 + 16 * 5);
-            __m512 _k20 = _mm512_loadu_ps(k0 + 16 * 6);
-            __m512 _k21 = _mm512_loadu_ps(k0 + 16 * 7);
-            __m512 _k22 = _mm512_loadu_ps(k0 + 16 * 8);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                const float* r0 = img0.row(i);
-                const float* r1 = img0.row(i + 1);
-                const float* r2 = img0.row(i + 2);
-
-                int j = 0;
-                for (; j < outw; j++)
-                {
-                    __m512 _r00 = _mm512_loadu_ps(r0);
-                    __m512 _r01 = _mm512_loadu_ps(r0 + 16);
-                    __m512 _r02 = _mm512_loadu_ps(r0 + 32);
-
-                    __m512 _sum0 = _mm512_mul_ps(_k00, _r00);
-                    __m512 _sum1 = _mm512_mul_ps(_k01, _r01);
-                    __m512 _sum2 = _mm512_mul_ps(_k02, _r02);
-
-                    __m512 _r10 = _mm512_loadu_ps(r1);
-                    __m512 _r11 = _mm512_loadu_ps(r1 + 16);
-                    __m512 _r12 = _mm512_loadu_ps(r1 + 32);
-
-                    _sum0 = _mm512_fmadd_ps(_k10, _r10, _sum0);
-                    _sum1 = _mm512_fmadd_ps(_k11, _r11, _sum1);
-                    _sum2 = _mm512_fmadd_ps(_k12, _r12, _sum2);
-
-                    __m512 _r20 = _mm512_loadu_ps(r2);
-                    __m512 _r21 = _mm512_loadu_ps(r2 + 16);
-                    __m512 _r22 = _mm512_loadu_ps(r2 + 32);
-
-                    _sum0 = _mm512_fmadd_ps(_k20, _r20, _sum0);
-                    _sum1 = _mm512_fmadd_ps(_k21, _r21, _sum1);
-                    _sum2 = _mm512_fmadd_ps(_k22, _r22, _sum2);
-
-                    __m512 _sum = _mm512_add_ps(_sum0, _mm512_add_ps(_sum1, _sum2));
-
-                    *outptr0 += _mm512_comp_reduce_add_ps(_sum);
-                    outptr0++;
-                    r0 += 16;
-                    r1 += 16;
-                    r2 += 16;
-                }
-            }
-
-            k0 += 9 * 16;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack1to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack1to4.h
deleted file mode 100644
index 31707546..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack1to4.h
+++ /dev/null
@@ -1,1283 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack1to4_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const float* bias = _bias;
-
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        Mat out0 = top_blob.channel(p);
-        Mat out1 = top_blob.channel(p + 1);
-
-        __m128 _bias0 = bias ? _mm_loadu_ps((const float*)bias + p * 4) : _mm_set1_ps(0.f);
-        __m128 _bias1 = bias ? _mm_loadu_ps((const float*)bias + (p + 1) * 4) : _mm_set1_ps(0.f);
-        out0.fill(_bias0);
-        out1.fill(_bias1);
-
-        const float* k0 = kernel.channel(p);
-        const float* k1 = kernel.channel(p + 1);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            __m128 _k00_0 = _mm_loadu_ps(k0);
-            __m128 _k01_0 = _mm_loadu_ps(k0 + 4);
-            __m128 _k02_0 = _mm_loadu_ps(k0 + 8);
-            __m128 _k10_0 = _mm_loadu_ps(k0 + 12);
-            __m128 _k11_0 = _mm_loadu_ps(k0 + 16);
-            __m128 _k12_0 = _mm_loadu_ps(k0 + 20);
-            __m128 _k20_0 = _mm_loadu_ps(k0 + 24);
-            __m128 _k21_0 = _mm_loadu_ps(k0 + 28);
-            __m128 _k22_0 = _mm_loadu_ps(k0 + 32);
-
-            __m128 _k00_1 = _mm_loadu_ps(k1);
-            __m128 _k01_1 = _mm_loadu_ps(k1 + 4);
-            __m128 _k02_1 = _mm_loadu_ps(k1 + 8);
-            __m128 _k10_1 = _mm_loadu_ps(k1 + 12);
-            __m128 _k11_1 = _mm_loadu_ps(k1 + 16);
-            __m128 _k12_1 = _mm_loadu_ps(k1 + 20);
-            __m128 _k20_1 = _mm_loadu_ps(k1 + 24);
-            __m128 _k21_1 = _mm_loadu_ps(k1 + 28);
-            __m128 _k22_1 = _mm_loadu_ps(k1 + 32);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-                    __m128 _sum00 = _mm_loadu_ps(outptr0);
-                    __m128 _sum10 = _mm_loadu_ps(outptr1);
-
-                    __m128 _r01 = _mm_set1_ps(*(r0));
-                    __m128 _r02 = _mm_set1_ps(*(r0 + 1));
-                    __m128 _r03 = _mm_set1_ps(*(r0 + 2));
-                    __m128 _r11 = _mm_set1_ps(*(r1));
-                    __m128 _r12 = _mm_set1_ps(*(r1 + 1));
-                    __m128 _r13 = _mm_set1_ps(*(r1 + 2));
-                    __m128 _r21 = _mm_set1_ps(*(r2));
-                    __m128 _r22 = _mm_set1_ps(*(r2 + 1));
-                    __m128 _r23 = _mm_set1_ps(*(r2 + 2));
-
-                    _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _sum10 = _mm_comp_fmadd_ps(_r01, _k00_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r02, _k01_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r03, _k02_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r11, _k10_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r12, _k11_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r13, _k12_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r21, _k20_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r22, _k21_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r23, _k22_1, _sum10);
-
-                    _mm_storeu_ps(outptr0, _sum00);
-                    _mm_storeu_ps(outptr1, _sum10);
-
-                    __m128 _sum01 = _mm_loadu_ps(outptr0 + 4);
-                    __m128 _sum11 = _mm_loadu_ps(outptr1 + 4);
-
-                    __m128 _r04 = _mm_set1_ps(*(r0 + 3));
-                    __m128 _r14 = _mm_set1_ps(*(r1 + 3));
-                    __m128 _r24 = _mm_set1_ps(*(r2 + 3));
-
-                    _sum01 = _mm_comp_fmadd_ps(_r02, _k00_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r03, _k01_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r04, _k02_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r12, _k10_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r13, _k11_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r14, _k12_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r22, _k20_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r23, _k21_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r24, _k22_0, _sum01);
-
-                    _sum11 = _mm_comp_fmadd_ps(_r02, _k00_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r03, _k01_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r04, _k02_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r12, _k10_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r13, _k11_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r14, _k12_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r22, _k20_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r23, _k21_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r24, _k22_1, _sum11);
-
-                    _mm_storeu_ps(outptr0 + 4, _sum01);
-                    _mm_storeu_ps(outptr1 + 4, _sum11);
-
-                    __m128 _sum02 = _mm_loadu_ps(outptr0 + 8);
-                    __m128 _sum12 = _mm_loadu_ps(outptr1 + 8);
-
-                    __m128 _r05 = _mm_set1_ps(*(r0 + 4));
-                    __m128 _r15 = _mm_set1_ps(*(r1 + 4));
-                    __m128 _r25 = _mm_set1_ps(*(r2 + 4));
-
-                    _sum02 = _mm_comp_fmadd_ps(_r03, _k00_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r04, _k01_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r05, _k02_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r13, _k10_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r14, _k11_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r15, _k12_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r23, _k20_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r24, _k21_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r25, _k22_0, _sum02);
-
-                    _sum12 = _mm_comp_fmadd_ps(_r03, _k00_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r04, _k01_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r05, _k02_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r13, _k10_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r14, _k11_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r15, _k12_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r23, _k20_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r24, _k21_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r25, _k22_1, _sum12);
-
-                    _mm_storeu_ps(outptr0 + 8, _sum02);
-                    _mm_storeu_ps(outptr1 + 8, _sum12);
-
-                    __m128 _r06 = _mm_set1_ps(*(r0 + 5));
-                    __m128 _r16 = _mm_set1_ps(*(r1 + 5));
-                    __m128 _r26 = _mm_set1_ps(*(r2 + 5));
-
-                    __m128 _sum03 = _mm_loadu_ps(outptr0 + 12);
-                    __m128 _sum13 = _mm_loadu_ps(outptr1 + 12);
-
-                    _sum03 = _mm_comp_fmadd_ps(_r04, _k00_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r05, _k01_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r06, _k02_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r14, _k10_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r15, _k11_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r16, _k12_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r24, _k20_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r25, _k21_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r26, _k22_0, _sum03);
-
-                    _sum13 = _mm_comp_fmadd_ps(_r04, _k00_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r05, _k01_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r06, _k02_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r14, _k10_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r15, _k11_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r16, _k12_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r24, _k20_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r25, _k21_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r26, _k22_1, _sum13);
-
-                    _mm_storeu_ps(outptr0 + 12, _sum03);
-                    _mm_storeu_ps(outptr1 + 12, _sum13);
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                    outptr0 += 16;
-                    outptr1 += 16;
-                }
-
-                for (; j + 1 < outw; j += 2)
-                {
-                    __m128 _sum00 = _mm_loadu_ps(outptr0);
-                    __m128 _sum10 = _mm_loadu_ps(outptr1);
-
-                    __m128 _r01 = _mm_set1_ps(*(r0));
-                    __m128 _r02 = _mm_set1_ps(*(r0 + 1));
-                    __m128 _r03 = _mm_set1_ps(*(r0 + 2));
-                    __m128 _r11 = _mm_set1_ps(*(r1));
-                    __m128 _r12 = _mm_set1_ps(*(r1 + 1));
-                    __m128 _r13 = _mm_set1_ps(*(r1 + 2));
-                    __m128 _r21 = _mm_set1_ps(*(r2));
-                    __m128 _r22 = _mm_set1_ps(*(r2 + 1));
-                    __m128 _r23 = _mm_set1_ps(*(r2 + 2));
-
-                    _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _sum10 = _mm_comp_fmadd_ps(_r01, _k00_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r02, _k01_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r03, _k02_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r11, _k10_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r12, _k11_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r13, _k12_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r21, _k20_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r22, _k21_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r23, _k22_1, _sum10);
-
-                    _mm_storeu_ps(outptr0, _sum00);
-                    _mm_storeu_ps(outptr1, _sum10);
-
-                    __m128 _sum01 = _mm_loadu_ps(outptr0 + 4);
-                    __m128 _sum11 = _mm_loadu_ps(outptr1 + 4);
-
-                    __m128 _r04 = _mm_set1_ps(*(r0 + 3));
-                    __m128 _r14 = _mm_set1_ps(*(r1 + 3));
-                    __m128 _r24 = _mm_set1_ps(*(r2 + 3));
-
-                    _sum01 = _mm_comp_fmadd_ps(_r02, _k00_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r03, _k01_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r04, _k02_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r12, _k10_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r13, _k11_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r14, _k12_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r22, _k20_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r23, _k21_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r24, _k22_0, _sum01);
-
-                    _sum11 = _mm_comp_fmadd_ps(_r02, _k00_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r03, _k01_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r04, _k02_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r12, _k10_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r13, _k11_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r14, _k12_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r22, _k20_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r23, _k21_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r24, _k22_1, _sum11);
-
-                    _mm_storeu_ps(outptr0 + 4, _sum01);
-                    _mm_storeu_ps(outptr1 + 4, _sum11);
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0 += 8;
-                    outptr1 += 8;
-                }
-
-                for (; j < outw; j++)
-                {
-                    __m128 _sum00 = _mm_loadu_ps(outptr0);
-                    __m128 _sum10 = _mm_loadu_ps(outptr1);
-
-                    __m128 _r01 = _mm_set1_ps(*(r0));
-                    __m128 _r02 = _mm_set1_ps(*(r0 + 1));
-                    __m128 _r03 = _mm_set1_ps(*(r0 + 2));
-                    __m128 _r11 = _mm_set1_ps(*(r1));
-                    __m128 _r12 = _mm_set1_ps(*(r1 + 1));
-                    __m128 _r13 = _mm_set1_ps(*(r1 + 2));
-                    __m128 _r21 = _mm_set1_ps(*(r2));
-                    __m128 _r22 = _mm_set1_ps(*(r2 + 1));
-                    __m128 _r23 = _mm_set1_ps(*(r2 + 2));
-
-                    _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _sum10 = _mm_comp_fmadd_ps(_r01, _k00_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r02, _k01_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r03, _k02_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r11, _k10_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r12, _k11_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r13, _k12_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r21, _k20_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r22, _k21_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r23, _k22_1, _sum10);
-
-                    _mm_storeu_ps(outptr0, _sum00);
-                    _mm_storeu_ps(outptr1, _sum10);
-
-                    r0 += 1;
-                    r1 += 1;
-                    r2 += 1;
-                    outptr0 += 4;
-                    outptr1 += 4;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * 4;
-            k1 += 9 * 4;
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        __m128 _bias0 = bias ? _mm_loadu_ps((const float*)bias + p * 4) : _mm_set1_ps(0.f);
-        out0.fill(_bias0);
-
-        const float* k0 = kernel.channel(p);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            __m128 _k00 = _mm_loadu_ps(k0);
-            __m128 _k01 = _mm_loadu_ps(k0 + 4);
-            __m128 _k02 = _mm_loadu_ps(k0 + 8);
-            __m128 _k10 = _mm_loadu_ps(k0 + 12);
-            __m128 _k11 = _mm_loadu_ps(k0 + 16);
-            __m128 _k12 = _mm_loadu_ps(k0 + 20);
-            __m128 _k20 = _mm_loadu_ps(k0 + 24);
-            __m128 _k21 = _mm_loadu_ps(k0 + 28);
-            __m128 _k22 = _mm_loadu_ps(k0 + 32);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-                    __m128 _sum0 = _mm_loadu_ps(outptr0);
-
-                    __m128 _r01 = _mm_set1_ps(*(r0));
-                    __m128 _r02 = _mm_set1_ps(*(r0 + 1));
-                    __m128 _r03 = _mm_set1_ps(*(r0 + 2));
-                    __m128 _r11 = _mm_set1_ps(*(r1));
-                    __m128 _r12 = _mm_set1_ps(*(r1 + 1));
-                    __m128 _r13 = _mm_set1_ps(*(r1 + 2));
-                    __m128 _r21 = _mm_set1_ps(*(r2));
-                    __m128 _r22 = _mm_set1_ps(*(r2 + 1));
-                    __m128 _r23 = _mm_set1_ps(*(r2 + 2));
-
-                    _sum0 = _mm_comp_fmadd_ps(_r01, _k00, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r02, _k01, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r03, _k02, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r11, _k10, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r12, _k11, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r13, _k12, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r21, _k20, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r22, _k21, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r23, _k22, _sum0);
-
-                    __m128 _sum1 = _mm_loadu_ps(outptr0 + 4);
-                    __m128 _r04 = _mm_set1_ps(*(r0 + 3));
-                    __m128 _r14 = _mm_set1_ps(*(r1 + 3));
-                    __m128 _r24 = _mm_set1_ps(*(r2 + 3));
-                    _mm_storeu_ps(outptr0, _sum0);
-
-                    _sum1 = _mm_comp_fmadd_ps(_r02, _k00, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r03, _k01, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r04, _k02, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r12, _k10, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r13, _k11, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r14, _k12, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r22, _k20, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r23, _k21, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r24, _k22, _sum1);
-
-                    __m128 _sum2 = _mm_loadu_ps(outptr0 + 8);
-                    __m128 _r05 = _mm_set1_ps(*(r0 + 4));
-                    __m128 _r15 = _mm_set1_ps(*(r1 + 4));
-                    __m128 _r25 = _mm_set1_ps(*(r2 + 4));
-                    _mm_storeu_ps(outptr0 + 4, _sum1);
-
-                    _sum2 = _mm_comp_fmadd_ps(_r03, _k00, _sum2);
-                    _sum2 = _mm_comp_fmadd_ps(_r04, _k01, _sum2);
-                    _sum2 = _mm_comp_fmadd_ps(_r05, _k02, _sum2);
-                    _sum2 = _mm_comp_fmadd_ps(_r13, _k10, _sum2);
-                    _sum2 = _mm_comp_fmadd_ps(_r14, _k11, _sum2);
-                    _sum2 = _mm_comp_fmadd_ps(_r15, _k12, _sum2);
-                    _sum2 = _mm_comp_fmadd_ps(_r23, _k20, _sum2);
-                    _sum2 = _mm_comp_fmadd_ps(_r24, _k21, _sum2);
-                    _sum2 = _mm_comp_fmadd_ps(_r25, _k22, _sum2);
-
-                    __m128 _sum3 = _mm_loadu_ps(outptr0 + 12);
-                    __m128 _r06 = _mm_set1_ps(*(r0 + 5));
-                    __m128 _r16 = _mm_set1_ps(*(r1 + 5));
-                    __m128 _r26 = _mm_set1_ps(*(r2 + 5));
-                    _mm_storeu_ps(outptr0 + 8, _sum2);
-
-                    _sum3 = _mm_comp_fmadd_ps(_r04, _k00, _sum3);
-                    _sum3 = _mm_comp_fmadd_ps(_r05, _k01, _sum3);
-                    _sum3 = _mm_comp_fmadd_ps(_r06, _k02, _sum3);
-                    _sum3 = _mm_comp_fmadd_ps(_r14, _k10, _sum3);
-                    _sum3 = _mm_comp_fmadd_ps(_r15, _k11, _sum3);
-                    _sum3 = _mm_comp_fmadd_ps(_r16, _k12, _sum3);
-                    _sum3 = _mm_comp_fmadd_ps(_r24, _k20, _sum3);
-                    _sum3 = _mm_comp_fmadd_ps(_r25, _k21, _sum3);
-                    _sum3 = _mm_comp_fmadd_ps(_r26, _k22, _sum3);
-
-                    _mm_storeu_ps(outptr0 + 12, _sum3);
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                    outptr0 += 16;
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    __m128 _sum0 = _mm_loadu_ps(outptr0);
-
-                    __m128 _r01 = _mm_set1_ps(*(r0));
-                    __m128 _r02 = _mm_set1_ps(*(r0 + 1));
-                    __m128 _r03 = _mm_set1_ps(*(r0 + 2));
-                    __m128 _r11 = _mm_set1_ps(*(r1));
-                    __m128 _r12 = _mm_set1_ps(*(r1 + 1));
-                    __m128 _r13 = _mm_set1_ps(*(r1 + 2));
-                    __m128 _r21 = _mm_set1_ps(*(r2));
-                    __m128 _r22 = _mm_set1_ps(*(r2 + 1));
-                    __m128 _r23 = _mm_set1_ps(*(r2 + 2));
-
-                    _sum0 = _mm_comp_fmadd_ps(_r01, _k00, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r02, _k01, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r03, _k02, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r11, _k10, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r12, _k11, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r13, _k12, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r21, _k20, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r22, _k21, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r23, _k22, _sum0);
-
-                    __m128 _sum1 = _mm_loadu_ps(outptr0 + 4);
-                    __m128 _r04 = _mm_set1_ps(*(r0 + 3));
-                    __m128 _r14 = _mm_set1_ps(*(r1 + 3));
-                    __m128 _r24 = _mm_set1_ps(*(r2 + 3));
-                    _mm_storeu_ps(outptr0, _sum0);
-
-                    _sum1 = _mm_comp_fmadd_ps(_r02, _k00, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r03, _k01, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r04, _k02, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r12, _k10, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r13, _k11, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r14, _k12, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r22, _k20, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r23, _k21, _sum1);
-                    _sum1 = _mm_comp_fmadd_ps(_r24, _k22, _sum1);
-
-                    _mm_storeu_ps(outptr0 + 4, _sum1);
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0 += 8;
-                }
-                for (; j < outw; j++)
-                {
-                    __m128 _sum0 = _mm_loadu_ps(outptr0);
-
-                    __m128 _r01 = _mm_set1_ps(*(r0));
-                    __m128 _r02 = _mm_set1_ps(*(r0 + 1));
-                    __m128 _r03 = _mm_set1_ps(*(r0 + 2));
-                    __m128 _r11 = _mm_set1_ps(*(r1));
-                    __m128 _r12 = _mm_set1_ps(*(r1 + 1));
-                    __m128 _r13 = _mm_set1_ps(*(r1 + 2));
-                    __m128 _r21 = _mm_set1_ps(*(r2));
-                    __m128 _r22 = _mm_set1_ps(*(r2 + 1));
-                    __m128 _r23 = _mm_set1_ps(*(r2 + 2));
-
-                    _sum0 = _mm_comp_fmadd_ps(_r01, _k00, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r02, _k01, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r03, _k02, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r11, _k10, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r12, _k11, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r13, _k12, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r21, _k20, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r22, _k21, _sum0);
-                    _sum0 = _mm_comp_fmadd_ps(_r23, _k22, _sum0);
-
-                    _mm_storeu_ps(outptr0, _sum0);
-                    r0 += 1;
-                    r1 += 1;
-                    r2 += 1;
-                    outptr0 += 4;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * 4;
-        }
-    }
-}
-
-static void conv3x3s2_pack1to4_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* bias = _bias;
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        Mat out0 = top_blob.channel(p);
-        Mat out1 = top_blob.channel(p + 1);
-
-        __m128 _bias0 = bias ? _mm_loadu_ps((const float*)bias + p * 4) : _mm_set1_ps(0.f);
-        __m128 _bias1 = bias ? _mm_loadu_ps((const float*)bias + (p + 1) * 4) : _mm_set1_ps(0.f);
-        out0.fill(_bias0);
-        out1.fill(_bias1);
-
-        const float* k0 = kernel.channel(p);
-        const float* k1 = kernel.channel(p + 1);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            __m128 _k00_0 = _mm_loadu_ps(k0);
-            __m128 _k01_0 = _mm_loadu_ps(k0 + 4);
-            __m128 _k02_0 = _mm_loadu_ps(k0 + 8);
-            __m128 _k10_0 = _mm_loadu_ps(k0 + 12);
-            __m128 _k11_0 = _mm_loadu_ps(k0 + 16);
-            __m128 _k12_0 = _mm_loadu_ps(k0 + 20);
-            __m128 _k20_0 = _mm_loadu_ps(k0 + 24);
-            __m128 _k21_0 = _mm_loadu_ps(k0 + 28);
-            __m128 _k22_0 = _mm_loadu_ps(k0 + 32);
-
-            __m128 _k00_1 = _mm_loadu_ps(k1);
-            __m128 _k01_1 = _mm_loadu_ps(k1 + 4);
-            __m128 _k02_1 = _mm_loadu_ps(k1 + 8);
-            __m128 _k10_1 = _mm_loadu_ps(k1 + 12);
-            __m128 _k11_1 = _mm_loadu_ps(k1 + 16);
-            __m128 _k12_1 = _mm_loadu_ps(k1 + 20);
-            __m128 _k20_1 = _mm_loadu_ps(k1 + 24);
-            __m128 _k21_1 = _mm_loadu_ps(k1 + 28);
-            __m128 _k22_1 = _mm_loadu_ps(k1 + 32);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-
-                for (; j + 3 < outw; j += 4)
-                {
-                    __m128 _sum00 = _mm_loadu_ps(outptr0);
-                    __m128 _sum10 = _mm_loadu_ps(outptr1);
-
-                    __m128 _r01 = _mm_set1_ps(*(r0));
-                    __m128 _r02 = _mm_set1_ps(*(r0 + 1));
-                    __m128 _r03 = _mm_set1_ps(*(r0 + 2));
-                    __m128 _r11 = _mm_set1_ps(*(r1));
-                    __m128 _r12 = _mm_set1_ps(*(r1 + 1));
-                    __m128 _r13 = _mm_set1_ps(*(r1 + 2));
-                    __m128 _r21 = _mm_set1_ps(*(r2));
-                    __m128 _r22 = _mm_set1_ps(*(r2 + 1));
-                    __m128 _r23 = _mm_set1_ps(*(r2 + 2));
-
-                    _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _sum10 = _mm_comp_fmadd_ps(_r01, _k00_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r02, _k01_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r03, _k02_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r11, _k10_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r12, _k11_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r13, _k12_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r21, _k20_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r22, _k21_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r23, _k22_1, _sum10);
-
-                    _mm_storeu_ps(outptr0, _sum00);
-                    _mm_storeu_ps(outptr1, _sum10);
-
-                    __m128 _sum01 = _mm_loadu_ps(outptr0 + 4);
-                    __m128 _sum11 = _mm_loadu_ps(outptr1 + 4);
-
-                    __m128 _r04 = _mm_set1_ps(*(r0 + 3));
-                    __m128 _r14 = _mm_set1_ps(*(r1 + 3));
-                    __m128 _r24 = _mm_set1_ps(*(r2 + 3));
-                    __m128 _r05 = _mm_set1_ps(*(r0 + 4));
-                    __m128 _r15 = _mm_set1_ps(*(r1 + 4));
-                    __m128 _r25 = _mm_set1_ps(*(r2 + 4));
-
-                    _sum01 = _mm_comp_fmadd_ps(_r03, _k00_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r04, _k01_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r05, _k02_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r13, _k10_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r14, _k11_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r15, _k12_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r23, _k20_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r24, _k21_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r25, _k22_0, _sum01);
-
-                    _sum11 = _mm_comp_fmadd_ps(_r03, _k00_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r04, _k01_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r05, _k02_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r13, _k10_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r14, _k11_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r15, _k12_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r23, _k20_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r24, _k21_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r25, _k22_1, _sum11);
-
-                    _mm_storeu_ps(outptr0 + 4, _sum01);
-                    _mm_storeu_ps(outptr1 + 4, _sum11);
-
-                    __m128 _sum02 = _mm_loadu_ps(outptr0 + 8);
-                    __m128 _sum12 = _mm_loadu_ps(outptr1 + 8);
-
-                    __m128 _r06 = _mm_set1_ps(*(r0 + 5));
-                    __m128 _r16 = _mm_set1_ps(*(r1 + 5));
-                    __m128 _r26 = _mm_set1_ps(*(r2 + 5));
-                    __m128 _r07 = _mm_set1_ps(*(r0 + 6));
-                    __m128 _r17 = _mm_set1_ps(*(r1 + 6));
-                    __m128 _r27 = _mm_set1_ps(*(r2 + 6));
-
-                    _sum02 = _mm_comp_fmadd_ps(_r05, _k00_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r06, _k01_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r07, _k02_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r15, _k10_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r16, _k11_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r17, _k12_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r25, _k20_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r26, _k21_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r27, _k22_0, _sum02);
-
-                    _sum12 = _mm_comp_fmadd_ps(_r05, _k00_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r06, _k01_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r07, _k02_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r15, _k10_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r16, _k11_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r17, _k12_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r25, _k20_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r26, _k21_1, _sum12);
-                    _sum12 = _mm_comp_fmadd_ps(_r27, _k22_1, _sum12);
-
-                    _mm_storeu_ps(outptr0 + 8, _sum02);
-                    _mm_storeu_ps(outptr1 + 8, _sum12);
-
-                    __m128 _r08 = _mm_set1_ps(*(r0 + 7));
-                    __m128 _r18 = _mm_set1_ps(*(r1 + 7));
-                    __m128 _r28 = _mm_set1_ps(*(r2 + 7));
-                    __m128 _r09 = _mm_set1_ps(*(r0 + 8));
-                    __m128 _r19 = _mm_set1_ps(*(r1 + 8));
-                    __m128 _r29 = _mm_set1_ps(*(r2 + 8));
-
-                    __m128 _sum03 = _mm_loadu_ps(outptr0 + 12);
-                    __m128 _sum13 = _mm_loadu_ps(outptr1 + 12);
-
-                    _sum03 = _mm_comp_fmadd_ps(_r07, _k00_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r08, _k01_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r09, _k02_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r17, _k10_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r18, _k11_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r19, _k12_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r27, _k20_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r28, _k21_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r29, _k22_0, _sum03);
-
-                    _sum13 = _mm_comp_fmadd_ps(_r07, _k00_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r08, _k01_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r09, _k02_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r17, _k10_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r18, _k11_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r19, _k12_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r27, _k20_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r28, _k21_1, _sum13);
-                    _sum13 = _mm_comp_fmadd_ps(_r29, _k22_1, _sum13);
-
-                    _mm_storeu_ps(outptr0 + 12, _sum03);
-                    _mm_storeu_ps(outptr1 + 12, _sum13);
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                    outptr0 += 16;
-                    outptr1 += 16;
-                }
-
-                for (; j + 1 < outw; j += 2)
-                {
-                    __m128 _sum00 = _mm_loadu_ps(outptr0);
-                    __m128 _sum10 = _mm_loadu_ps(outptr1);
-
-                    __m128 _r01 = _mm_set1_ps(*(r0));
-                    __m128 _r02 = _mm_set1_ps(*(r0 + 1));
-                    __m128 _r03 = _mm_set1_ps(*(r0 + 2));
-                    __m128 _r11 = _mm_set1_ps(*(r1));
-                    __m128 _r12 = _mm_set1_ps(*(r1 + 1));
-                    __m128 _r13 = _mm_set1_ps(*(r1 + 2));
-                    __m128 _r21 = _mm_set1_ps(*(r2));
-                    __m128 _r22 = _mm_set1_ps(*(r2 + 1));
-                    __m128 _r23 = _mm_set1_ps(*(r2 + 2));
-
-                    _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _sum10 = _mm_comp_fmadd_ps(_r01, _k00_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r02, _k01_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r03, _k02_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r11, _k10_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r12, _k11_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r13, _k12_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r21, _k20_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r22, _k21_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r23, _k22_1, _sum10);
-
-                    _mm_storeu_ps(outptr0, _sum00);
-                    _mm_storeu_ps(outptr1, _sum10);
-
-                    __m128 _sum01 = _mm_loadu_ps(outptr0 + 4);
-                    __m128 _sum11 = _mm_loadu_ps(outptr1 + 4);
-
-                    __m128 _r04 = _mm_set1_ps(*(r0 + 3));
-                    __m128 _r14 = _mm_set1_ps(*(r1 + 3));
-                    __m128 _r24 = _mm_set1_ps(*(r2 + 3));
-                    __m128 _r05 = _mm_set1_ps(*(r0 + 4));
-                    __m128 _r15 = _mm_set1_ps(*(r1 + 4));
-                    __m128 _r25 = _mm_set1_ps(*(r2 + 4));
-
-                    _sum01 = _mm_comp_fmadd_ps(_r03, _k00_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r04, _k01_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r05, _k02_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r13, _k10_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r14, _k11_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r15, _k12_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r23, _k20_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r24, _k21_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r25, _k22_0, _sum01);
-
-                    _sum11 = _mm_comp_fmadd_ps(_r03, _k00_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r04, _k01_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r05, _k02_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r13, _k10_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r14, _k11_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r15, _k12_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r23, _k20_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r24, _k21_1, _sum11);
-                    _sum11 = _mm_comp_fmadd_ps(_r25, _k22_1, _sum11);
-
-                    _mm_storeu_ps(outptr0 + 4, _sum01);
-                    _mm_storeu_ps(outptr1 + 4, _sum11);
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                    outptr0 += 8;
-                    outptr1 += 8;
-                }
-                for (; j < outw; j++)
-                {
-                    __m128 _sum00 = _mm_loadu_ps(outptr0);
-                    __m128 _sum10 = _mm_loadu_ps(outptr1);
-
-                    __m128 _r01 = _mm_set1_ps(*(r0));
-                    __m128 _r02 = _mm_set1_ps(*(r0 + 1));
-                    __m128 _r03 = _mm_set1_ps(*(r0 + 2));
-                    __m128 _r11 = _mm_set1_ps(*(r1));
-                    __m128 _r12 = _mm_set1_ps(*(r1 + 1));
-                    __m128 _r13 = _mm_set1_ps(*(r1 + 2));
-                    __m128 _r21 = _mm_set1_ps(*(r2));
-                    __m128 _r22 = _mm_set1_ps(*(r2 + 1));
-                    __m128 _r23 = _mm_set1_ps(*(r2 + 2));
-
-                    _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _sum10 = _mm_comp_fmadd_ps(_r01, _k00_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r02, _k01_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r03, _k02_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r11, _k10_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r12, _k11_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r13, _k12_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r21, _k20_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r22, _k21_1, _sum10);
-                    _sum10 = _mm_comp_fmadd_ps(_r23, _k22_1, _sum10);
-
-                    _mm_storeu_ps(outptr0, _sum00);
-                    _mm_storeu_ps(outptr1, _sum10);
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0 += 4;
-                    outptr1 += 4;
-                }
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * 4;
-            k1 += 9 * 4;
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        __m128 _bias0 = bias ? _mm_loadu_ps((const float*)bias + p * 4) : _mm_set1_ps(0.f);
-        out0.fill(_bias0);
-
-        const float* k0 = kernel.channel(p);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            __m128 _k00_0 = _mm_loadu_ps(k0);
-            __m128 _k01_0 = _mm_loadu_ps(k0 + 4);
-            __m128 _k02_0 = _mm_loadu_ps(k0 + 8);
-            __m128 _k10_0 = _mm_loadu_ps(k0 + 12);
-            __m128 _k11_0 = _mm_loadu_ps(k0 + 16);
-            __m128 _k12_0 = _mm_loadu_ps(k0 + 20);
-            __m128 _k20_0 = _mm_loadu_ps(k0 + 24);
-            __m128 _k21_0 = _mm_loadu_ps(k0 + 28);
-            __m128 _k22_0 = _mm_loadu_ps(k0 + 32);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    __m128 _sum00 = _mm_loadu_ps(outptr0);
-
-                    __m128 _r01 = _mm_set1_ps(*(r0));
-                    __m128 _r02 = _mm_set1_ps(*(r0 + 1));
-                    __m128 _r03 = _mm_set1_ps(*(r0 + 2));
-                    __m128 _r11 = _mm_set1_ps(*(r1));
-                    __m128 _r12 = _mm_set1_ps(*(r1 + 1));
-                    __m128 _r13 = _mm_set1_ps(*(r1 + 2));
-                    __m128 _r21 = _mm_set1_ps(*(r2));
-                    __m128 _r22 = _mm_set1_ps(*(r2 + 1));
-                    __m128 _r23 = _mm_set1_ps(*(r2 + 2));
-
-                    _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _mm_storeu_ps(outptr0, _sum00);
-
-                    __m128 _sum01 = _mm_loadu_ps(outptr0 + 4);
-
-                    __m128 _r04 = _mm_set1_ps(*(r0 + 3));
-                    __m128 _r14 = _mm_set1_ps(*(r1 + 3));
-                    __m128 _r24 = _mm_set1_ps(*(r2 + 3));
-                    __m128 _r05 = _mm_set1_ps(*(r0 + 4));
-                    __m128 _r15 = _mm_set1_ps(*(r1 + 4));
-                    __m128 _r25 = _mm_set1_ps(*(r2 + 4));
-
-                    _sum01 = _mm_comp_fmadd_ps(_r03, _k00_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r04, _k01_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r05, _k02_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r13, _k10_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r14, _k11_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r15, _k12_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r23, _k20_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r24, _k21_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r25, _k22_0, _sum01);
-
-                    _mm_storeu_ps(outptr0 + 4, _sum01);
-
-                    __m128 _sum02 = _mm_loadu_ps(outptr0 + 8);
-
-                    __m128 _r06 = _mm_set1_ps(*(r0 + 5));
-                    __m128 _r16 = _mm_set1_ps(*(r1 + 5));
-                    __m128 _r26 = _mm_set1_ps(*(r2 + 5));
-                    __m128 _r07 = _mm_set1_ps(*(r0 + 6));
-                    __m128 _r17 = _mm_set1_ps(*(r1 + 6));
-                    __m128 _r27 = _mm_set1_ps(*(r2 + 6));
-
-                    _sum02 = _mm_comp_fmadd_ps(_r05, _k00_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r06, _k01_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r07, _k02_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r15, _k10_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r16, _k11_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r17, _k12_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r25, _k20_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r26, _k21_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r27, _k22_0, _sum02);
-
-                    _mm_storeu_ps(outptr0 + 8, _sum02);
-
-                    __m128 _r08 = _mm_set1_ps(*(r0 + 7));
-                    __m128 _r18 = _mm_set1_ps(*(r1 + 7));
-                    __m128 _r28 = _mm_set1_ps(*(r2 + 7));
-                    __m128 _r09 = _mm_set1_ps(*(r0 + 8));
-                    __m128 _r19 = _mm_set1_ps(*(r1 + 8));
-                    __m128 _r29 = _mm_set1_ps(*(r2 + 8));
-
-                    __m128 _sum03 = _mm_loadu_ps(outptr0 + 12);
-
-                    _sum03 = _mm_comp_fmadd_ps(_r07, _k00_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r08, _k01_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r09, _k02_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r17, _k10_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r18, _k11_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r19, _k12_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r27, _k20_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r28, _k21_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r29, _k22_0, _sum03);
-
-                    _mm_storeu_ps(outptr0 + 12, _sum03);
-
-                    __m128 _r010 = _mm_set1_ps(*(r0 + 9));
-                    __m128 _r110 = _mm_set1_ps(*(r1 + 9));
-                    __m128 _r210 = _mm_set1_ps(*(r2 + 9));
-                    __m128 _r011 = _mm_set1_ps(*(r0 + 10));
-                    __m128 _r111 = _mm_set1_ps(*(r1 + 10));
-                    __m128 _r211 = _mm_set1_ps(*(r2 + 10));
-
-                    __m128 _sum04 = _mm_loadu_ps(outptr0 + 16);
-
-                    _sum04 = _mm_comp_fmadd_ps(_r09, _k00_0, _sum04);
-                    _sum04 = _mm_comp_fmadd_ps(_r010, _k01_0, _sum04);
-                    _sum04 = _mm_comp_fmadd_ps(_r011, _k02_0, _sum04);
-                    _sum04 = _mm_comp_fmadd_ps(_r19, _k10_0, _sum04);
-                    _sum04 = _mm_comp_fmadd_ps(_r110, _k11_0, _sum04);
-                    _sum04 = _mm_comp_fmadd_ps(_r111, _k12_0, _sum04);
-                    _sum04 = _mm_comp_fmadd_ps(_r29, _k20_0, _sum04);
-                    _sum04 = _mm_comp_fmadd_ps(_r210, _k21_0, _sum04);
-                    _sum04 = _mm_comp_fmadd_ps(_r211, _k22_0, _sum04);
-
-                    _mm_storeu_ps(outptr0 + 16, _sum04);
-
-                    __m128 _r012 = _mm_set1_ps(*(r0 + 11));
-                    __m128 _r112 = _mm_set1_ps(*(r1 + 11));
-                    __m128 _r212 = _mm_set1_ps(*(r2 + 11));
-                    __m128 _r013 = _mm_set1_ps(*(r0 + 12));
-                    __m128 _r113 = _mm_set1_ps(*(r1 + 12));
-                    __m128 _r213 = _mm_set1_ps(*(r2 + 12));
-
-                    __m128 _sum05 = _mm_loadu_ps(outptr0 + 20);
-
-                    _sum05 = _mm_comp_fmadd_ps(_r011, _k00_0, _sum05);
-                    _sum05 = _mm_comp_fmadd_ps(_r012, _k01_0, _sum05);
-                    _sum05 = _mm_comp_fmadd_ps(_r013, _k02_0, _sum05);
-                    _sum05 = _mm_comp_fmadd_ps(_r111, _k10_0, _sum05);
-                    _sum05 = _mm_comp_fmadd_ps(_r112, _k11_0, _sum05);
-                    _sum05 = _mm_comp_fmadd_ps(_r113, _k12_0, _sum05);
-                    _sum05 = _mm_comp_fmadd_ps(_r211, _k20_0, _sum05);
-                    _sum05 = _mm_comp_fmadd_ps(_r212, _k21_0, _sum05);
-                    _sum05 = _mm_comp_fmadd_ps(_r213, _k22_0, _sum05);
-
-                    _mm_storeu_ps(outptr0 + 20, _sum05);
-
-                    __m128 _r014 = _mm_set1_ps(*(r0 + 13));
-                    __m128 _r114 = _mm_set1_ps(*(r1 + 13));
-                    __m128 _r214 = _mm_set1_ps(*(r2 + 13));
-                    __m128 _r015 = _mm_set1_ps(*(r0 + 14));
-                    __m128 _r115 = _mm_set1_ps(*(r1 + 14));
-                    __m128 _r215 = _mm_set1_ps(*(r2 + 14));
-
-                    __m128 _sum06 = _mm_loadu_ps(outptr0 + 24);
-
-                    _sum06 = _mm_comp_fmadd_ps(_r013, _k00_0, _sum06);
-                    _sum06 = _mm_comp_fmadd_ps(_r014, _k01_0, _sum06);
-                    _sum06 = _mm_comp_fmadd_ps(_r015, _k02_0, _sum06);
-                    _sum06 = _mm_comp_fmadd_ps(_r113, _k10_0, _sum06);
-                    _sum06 = _mm_comp_fmadd_ps(_r114, _k11_0, _sum06);
-                    _sum06 = _mm_comp_fmadd_ps(_r115, _k12_0, _sum06);
-                    _sum06 = _mm_comp_fmadd_ps(_r213, _k20_0, _sum06);
-                    _sum06 = _mm_comp_fmadd_ps(_r214, _k21_0, _sum06);
-                    _sum06 = _mm_comp_fmadd_ps(_r215, _k22_0, _sum06);
-
-                    _mm_storeu_ps(outptr0 + 24, _sum06);
-
-                    __m128 _r016 = _mm_set1_ps(*(r0 + 15));
-                    __m128 _r116 = _mm_set1_ps(*(r1 + 15));
-                    __m128 _r216 = _mm_set1_ps(*(r2 + 15));
-                    __m128 _r017 = _mm_set1_ps(*(r0 + 16));
-                    __m128 _r117 = _mm_set1_ps(*(r1 + 16));
-                    __m128 _r217 = _mm_set1_ps(*(r2 + 16));
-
-                    __m128 _sum07 = _mm_loadu_ps(outptr0 + 28);
-
-                    _sum07 = _mm_comp_fmadd_ps(_r015, _k00_0, _sum07);
-                    _sum07 = _mm_comp_fmadd_ps(_r016, _k01_0, _sum07);
-                    _sum07 = _mm_comp_fmadd_ps(_r017, _k02_0, _sum07);
-                    _sum07 = _mm_comp_fmadd_ps(_r115, _k10_0, _sum07);
-                    _sum07 = _mm_comp_fmadd_ps(_r116, _k11_0, _sum07);
-                    _sum07 = _mm_comp_fmadd_ps(_r117, _k12_0, _sum07);
-                    _sum07 = _mm_comp_fmadd_ps(_r215, _k20_0, _sum07);
-                    _sum07 = _mm_comp_fmadd_ps(_r216, _k21_0, _sum07);
-                    _sum07 = _mm_comp_fmadd_ps(_r217, _k22_0, _sum07);
-
-                    _mm_storeu_ps(outptr0 + 28, _sum07);
-
-                    r0 += 16;
-                    r1 += 16;
-                    r2 += 16;
-                    outptr0 += 32;
-                }
-
-                for (; j + 3 < outw; j += 4)
-                {
-                    __m128 _sum00 = _mm_loadu_ps(outptr0);
-
-                    __m128 _r01 = _mm_set1_ps(*(r0));
-                    __m128 _r02 = _mm_set1_ps(*(r0 + 1));
-                    __m128 _r03 = _mm_set1_ps(*(r0 + 2));
-                    __m128 _r11 = _mm_set1_ps(*(r1));
-                    __m128 _r12 = _mm_set1_ps(*(r1 + 1));
-                    __m128 _r13 = _mm_set1_ps(*(r1 + 2));
-                    __m128 _r21 = _mm_set1_ps(*(r2));
-                    __m128 _r22 = _mm_set1_ps(*(r2 + 1));
-                    __m128 _r23 = _mm_set1_ps(*(r2 + 2));
-
-                    _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _mm_storeu_ps(outptr0, _sum00);
-
-                    __m128 _sum01 = _mm_loadu_ps(outptr0 + 4);
-
-                    __m128 _r04 = _mm_set1_ps(*(r0 + 3));
-                    __m128 _r14 = _mm_set1_ps(*(r1 + 3));
-                    __m128 _r24 = _mm_set1_ps(*(r2 + 3));
-                    __m128 _r05 = _mm_set1_ps(*(r0 + 4));
-                    __m128 _r15 = _mm_set1_ps(*(r1 + 4));
-                    __m128 _r25 = _mm_set1_ps(*(r2 + 4));
-
-                    _sum01 = _mm_comp_fmadd_ps(_r03, _k00_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r04, _k01_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r05, _k02_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r13, _k10_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r14, _k11_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r15, _k12_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r23, _k20_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r24, _k21_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r25, _k22_0, _sum01);
-
-                    _mm_storeu_ps(outptr0 + 4, _sum01);
-
-                    __m128 _sum02 = _mm_loadu_ps(outptr0 + 8);
-
-                    __m128 _r06 = _mm_set1_ps(*(r0 + 5));
-                    __m128 _r16 = _mm_set1_ps(*(r1 + 5));
-                    __m128 _r26 = _mm_set1_ps(*(r2 + 5));
-                    __m128 _r07 = _mm_set1_ps(*(r0 + 6));
-                    __m128 _r17 = _mm_set1_ps(*(r1 + 6));
-                    __m128 _r27 = _mm_set1_ps(*(r2 + 6));
-
-                    _sum02 = _mm_comp_fmadd_ps(_r05, _k00_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r06, _k01_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r07, _k02_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r15, _k10_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r16, _k11_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r17, _k12_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r25, _k20_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r26, _k21_0, _sum02);
-                    _sum02 = _mm_comp_fmadd_ps(_r27, _k22_0, _sum02);
-
-                    _mm_storeu_ps(outptr0 + 8, _sum02);
-
-                    __m128 _r08 = _mm_set1_ps(*(r0 + 7));
-                    __m128 _r18 = _mm_set1_ps(*(r1 + 7));
-                    __m128 _r28 = _mm_set1_ps(*(r2 + 7));
-                    __m128 _r09 = _mm_set1_ps(*(r0 + 8));
-                    __m128 _r19 = _mm_set1_ps(*(r1 + 8));
-                    __m128 _r29 = _mm_set1_ps(*(r2 + 8));
-
-                    __m128 _sum03 = _mm_loadu_ps(outptr0 + 12);
-
-                    _sum03 = _mm_comp_fmadd_ps(_r07, _k00_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r08, _k01_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r09, _k02_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r17, _k10_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r18, _k11_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r19, _k12_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r27, _k20_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r28, _k21_0, _sum03);
-                    _sum03 = _mm_comp_fmadd_ps(_r29, _k22_0, _sum03);
-
-                    _mm_storeu_ps(outptr0 + 12, _sum03);
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                    outptr0 += 16;
-                }
-
-                for (; j + 1 < outw; j += 2)
-                {
-                    __m128 _sum00 = _mm_loadu_ps(outptr0);
-
-                    __m128 _r01 = _mm_set1_ps(*(r0));
-                    __m128 _r02 = _mm_set1_ps(*(r0 + 1));
-                    __m128 _r03 = _mm_set1_ps(*(r0 + 2));
-                    __m128 _r11 = _mm_set1_ps(*(r1));
-                    __m128 _r12 = _mm_set1_ps(*(r1 + 1));
-                    __m128 _r13 = _mm_set1_ps(*(r1 + 2));
-                    __m128 _r21 = _mm_set1_ps(*(r2));
-                    __m128 _r22 = _mm_set1_ps(*(r2 + 1));
-                    __m128 _r23 = _mm_set1_ps(*(r2 + 2));
-
-                    _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _mm_storeu_ps(outptr0, _sum00);
-
-                    __m128 _sum01 = _mm_loadu_ps(outptr0 + 4);
-
-                    __m128 _r04 = _mm_set1_ps(*(r0 + 3));
-                    __m128 _r14 = _mm_set1_ps(*(r1 + 3));
-                    __m128 _r24 = _mm_set1_ps(*(r2 + 3));
-                    __m128 _r05 = _mm_set1_ps(*(r0 + 4));
-                    __m128 _r15 = _mm_set1_ps(*(r1 + 4));
-                    __m128 _r25 = _mm_set1_ps(*(r2 + 4));
-
-                    _sum01 = _mm_comp_fmadd_ps(_r03, _k00_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r04, _k01_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r05, _k02_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r13, _k10_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r14, _k11_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r15, _k12_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r23, _k20_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r24, _k21_0, _sum01);
-                    _sum01 = _mm_comp_fmadd_ps(_r25, _k22_0, _sum01);
-
-                    _mm_storeu_ps(outptr0 + 4, _sum01);
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                    outptr0 += 8;
-                }
-                for (; j < outw; j++)
-                {
-                    __m128 _sum00 = _mm_loadu_ps(outptr0);
-                    __m128 _r01 = _mm_set1_ps(*(r0));
-                    __m128 _r02 = _mm_set1_ps(*(r0 + 1));
-                    __m128 _r03 = _mm_set1_ps(*(r0 + 2));
-                    __m128 _r11 = _mm_set1_ps(*(r1));
-                    __m128 _r12 = _mm_set1_ps(*(r1 + 1));
-                    __m128 _r13 = _mm_set1_ps(*(r1 + 2));
-                    __m128 _r21 = _mm_set1_ps(*(r2));
-                    __m128 _r22 = _mm_set1_ps(*(r2 + 1));
-                    __m128 _r23 = _mm_set1_ps(*(r2 + 2));
-
-                    _sum00 = _mm_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm_comp_fmadd_ps(_r23, _k22_0, _sum00);
-                    _mm_storeu_ps(outptr0, _sum00);
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0 += 4;
-                }
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * 4;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack1to8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack1to8.h
deleted file mode 100644
index 3c4dccce..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack1to8.h
+++ /dev/null
@@ -1,1555 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack1to8_avx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const float* bias = _bias;
-
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        Mat out0 = top_blob.channel(p);
-        Mat out1 = top_blob.channel(p + 1);
-
-        __m256 _bias0 = bias ? _mm256_loadu_ps((const float*)bias + p * 8) : _mm256_set1_ps(0.f);
-        __m256 _bias1 = bias ? _mm256_loadu_ps((const float*)bias + (p + 1) * 8) : _mm256_set1_ps(0.f);
-        out0.fill(_bias0);
-        out1.fill(_bias1);
-
-        const float* k0 = kernel.channel(p);
-        const float* k1 = kernel.channel(p + 1);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            __m256 _k00_0 = _mm256_loadu_ps(k0);
-            __m256 _k01_0 = _mm256_loadu_ps(k0 + 8);
-            __m256 _k02_0 = _mm256_loadu_ps(k0 + 16);
-            __m256 _k10_0 = _mm256_loadu_ps(k0 + 24);
-            __m256 _k11_0 = _mm256_loadu_ps(k0 + 32);
-            __m256 _k12_0 = _mm256_loadu_ps(k0 + 40);
-            __m256 _k20_0 = _mm256_loadu_ps(k0 + 48);
-            __m256 _k21_0 = _mm256_loadu_ps(k0 + 56);
-            __m256 _k22_0 = _mm256_loadu_ps(k0 + 64);
-
-            __m256 _k00_1 = _mm256_loadu_ps(k1);
-            __m256 _k01_1 = _mm256_loadu_ps(k1 + 8);
-            __m256 _k02_1 = _mm256_loadu_ps(k1 + 16);
-            __m256 _k10_1 = _mm256_loadu_ps(k1 + 24);
-            __m256 _k11_1 = _mm256_loadu_ps(k1 + 32);
-            __m256 _k12_1 = _mm256_loadu_ps(k1 + 40);
-            __m256 _k20_1 = _mm256_loadu_ps(k1 + 48);
-            __m256 _k21_1 = _mm256_loadu_ps(k1 + 56);
-            __m256 _k22_1 = _mm256_loadu_ps(k1 + 64);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-                    __m256 _sum00 = _mm256_loadu_ps(outptr0);
-                    __m256 _sum10 = _mm256_loadu_ps(outptr1);
-
-                    __m256 _r01 = _mm256_broadcast_ss(r0);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r11 = _mm256_broadcast_ss(r1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r21 = _mm256_broadcast_ss(r2);
-                    __m256 _r22 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r23 = _mm256_broadcast_ss(r2 + 2);
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r01, _k00_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r02, _k01_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r03, _k02_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r11, _k10_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r12, _k11_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r13, _k12_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r21, _k20_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r22, _k21_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r23, _k22_1, _sum10);
-
-                    _mm256_storeu_ps(outptr0, _sum00);
-                    _mm256_storeu_ps(outptr1, _sum10);
-
-                    __m256 _sum01 = _mm256_loadu_ps(outptr0 + 8);
-                    __m256 _sum11 = _mm256_loadu_ps(outptr1 + 8);
-
-                    __m256 _r04 = _mm256_broadcast_ss(r0 + 3);
-                    __m256 _r14 = _mm256_broadcast_ss(r1 + 3);
-                    __m256 _r24 = _mm256_broadcast_ss(r2 + 3);
-
-                    _sum01 = _mm256_comp_fmadd_ps(_r02, _k00_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r03, _k01_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r04, _k02_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r12, _k10_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r13, _k11_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r14, _k12_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r22, _k20_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r23, _k21_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r24, _k22_0, _sum01);
-
-                    _sum11 = _mm256_comp_fmadd_ps(_r02, _k00_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r03, _k01_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r04, _k02_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r12, _k10_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r13, _k11_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r14, _k12_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r22, _k20_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r23, _k21_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r24, _k22_1, _sum11);
-
-                    _mm256_storeu_ps(outptr0 + 8, _sum01);
-                    _mm256_storeu_ps(outptr1 + 8, _sum11);
-
-                    __m256 _sum02 = _mm256_loadu_ps(outptr0 + 16);
-                    __m256 _sum12 = _mm256_loadu_ps(outptr1 + 16);
-
-                    __m256 _r05 = _mm256_broadcast_ss(r0 + 4);
-                    __m256 _r15 = _mm256_broadcast_ss(r1 + 4);
-                    __m256 _r25 = _mm256_broadcast_ss(r2 + 4);
-
-                    _sum02 = _mm256_comp_fmadd_ps(_r03, _k00_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r04, _k01_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r05, _k02_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r13, _k10_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r14, _k11_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r15, _k12_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r23, _k20_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r24, _k21_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r25, _k22_0, _sum02);
-
-                    _sum12 = _mm256_comp_fmadd_ps(_r03, _k00_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r04, _k01_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r05, _k02_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r13, _k10_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r14, _k11_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r15, _k12_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r23, _k20_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r24, _k21_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r25, _k22_1, _sum12);
-
-                    _mm256_storeu_ps(outptr0 + 16, _sum02);
-                    _mm256_storeu_ps(outptr1 + 16, _sum12);
-
-                    __m256 _r06 = _mm256_broadcast_ss(r0 + 5);
-                    __m256 _r16 = _mm256_broadcast_ss(r1 + 5);
-                    __m256 _r26 = _mm256_broadcast_ss(r2 + 5);
-
-                    __m256 _sum03 = _mm256_loadu_ps(outptr0 + 24);
-                    __m256 _sum13 = _mm256_loadu_ps(outptr1 + 24);
-
-                    _sum03 = _mm256_comp_fmadd_ps(_r04, _k00_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r05, _k01_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r06, _k02_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r14, _k10_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r15, _k11_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r16, _k12_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r24, _k20_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r25, _k21_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r26, _k22_0, _sum03);
-
-                    _sum13 = _mm256_comp_fmadd_ps(_r04, _k00_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r05, _k01_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r06, _k02_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r14, _k10_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r15, _k11_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r16, _k12_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r24, _k20_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r25, _k21_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r26, _k22_1, _sum13);
-
-                    _mm256_storeu_ps(outptr0 + 24, _sum03);
-                    _mm256_storeu_ps(outptr1 + 24, _sum13);
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                    outptr0 += 32;
-                    outptr1 += 32;
-                }
-
-                for (; j + 1 < outw; j += 2)
-                {
-                    __m256 _sum00 = _mm256_loadu_ps(outptr0);
-                    __m256 _sum10 = _mm256_loadu_ps(outptr1);
-
-                    __m256 _r01 = _mm256_broadcast_ss(r0);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r11 = _mm256_broadcast_ss(r1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r21 = _mm256_broadcast_ss(r2);
-                    __m256 _r22 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r23 = _mm256_broadcast_ss(r2 + 2);
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r01, _k00_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r02, _k01_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r03, _k02_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r11, _k10_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r12, _k11_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r13, _k12_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r21, _k20_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r22, _k21_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r23, _k22_1, _sum10);
-
-                    _mm256_storeu_ps(outptr0, _sum00);
-                    _mm256_storeu_ps(outptr1, _sum10);
-
-                    __m256 _sum01 = _mm256_loadu_ps(outptr0 + 8);
-                    __m256 _sum11 = _mm256_loadu_ps(outptr1 + 8);
-
-                    __m256 _r04 = _mm256_broadcast_ss(r0 + 3);
-                    __m256 _r14 = _mm256_broadcast_ss(r1 + 3);
-                    __m256 _r24 = _mm256_broadcast_ss(r2 + 3);
-
-                    _sum01 = _mm256_comp_fmadd_ps(_r02, _k00_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r03, _k01_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r04, _k02_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r12, _k10_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r13, _k11_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r14, _k12_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r22, _k20_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r23, _k21_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r24, _k22_0, _sum01);
-
-                    _sum11 = _mm256_comp_fmadd_ps(_r02, _k00_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r03, _k01_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r04, _k02_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r12, _k10_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r13, _k11_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r14, _k12_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r22, _k20_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r23, _k21_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r24, _k22_1, _sum11);
-
-                    _mm256_storeu_ps(outptr0 + 8, _sum01);
-                    _mm256_storeu_ps(outptr1 + 8, _sum11);
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0 += 16;
-                    outptr1 += 16;
-                }
-
-                for (; j < outw; j++)
-                {
-                    __m256 _sum00 = _mm256_loadu_ps(outptr0);
-                    __m256 _sum10 = _mm256_loadu_ps(outptr1);
-
-                    __m256 _r01 = _mm256_broadcast_ss(r0);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r11 = _mm256_broadcast_ss(r1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r21 = _mm256_broadcast_ss(r2);
-                    __m256 _r22 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r23 = _mm256_broadcast_ss(r2 + 2);
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r01, _k00_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r02, _k01_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r03, _k02_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r11, _k10_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r12, _k11_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r13, _k12_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r21, _k20_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r22, _k21_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r23, _k22_1, _sum10);
-
-                    _mm256_storeu_ps(outptr0, _sum00);
-                    _mm256_storeu_ps(outptr1, _sum10);
-
-                    r0 += 1;
-                    r1 += 1;
-                    r2 += 1;
-                    outptr0 += 8;
-                    outptr1 += 8;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * 8;
-            k1 += 9 * 8;
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        __m256 _bias0 = bias ? _mm256_loadu_ps((const float*)bias + p * 8) : _mm256_set1_ps(0.f);
-        out0.fill(_bias0);
-
-        const float* k0 = kernel.channel(p);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            __m256 _k00 = _mm256_loadu_ps(k0);
-            __m256 _k01 = _mm256_loadu_ps(k0 + 8);
-            __m256 _k02 = _mm256_loadu_ps(k0 + 16);
-            __m256 _k10 = _mm256_loadu_ps(k0 + 24);
-            __m256 _k11 = _mm256_loadu_ps(k0 + 32);
-            __m256 _k12 = _mm256_loadu_ps(k0 + 40);
-            __m256 _k20 = _mm256_loadu_ps(k0 + 48);
-            __m256 _k21 = _mm256_loadu_ps(k0 + 56);
-            __m256 _k22 = _mm256_loadu_ps(k0 + 64);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 3 < outw; j += 4)
-                {
-                    __m256 _sum0 = _mm256_loadu_ps(outptr0);
-
-                    __m256 _r01 = _mm256_broadcast_ss(r0);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r11 = _mm256_broadcast_ss(r1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r21 = _mm256_broadcast_ss(r2);
-                    __m256 _r22 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r23 = _mm256_broadcast_ss(r2 + 2);
-
-                    _sum0 = _mm256_comp_fmadd_ps(_r01, _k00, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r02, _k01, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r03, _k02, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r11, _k10, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r12, _k11, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r13, _k12, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r21, _k20, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r22, _k21, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r23, _k22, _sum0);
-
-                    __m256 _sum1 = _mm256_loadu_ps(outptr0 + 8);
-                    __m256 _r04 = _mm256_broadcast_ss(r0 + 3);
-                    __m256 _r14 = _mm256_broadcast_ss(r1 + 3);
-                    __m256 _r24 = _mm256_broadcast_ss(r2 + 3);
-                    _mm256_storeu_ps(outptr0, _sum0);
-
-                    _sum1 = _mm256_comp_fmadd_ps(_r02, _k00, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r03, _k01, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r04, _k02, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r12, _k10, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r13, _k11, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r14, _k12, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r22, _k20, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r23, _k21, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r24, _k22, _sum1);
-
-                    __m256 _sum2 = _mm256_loadu_ps(outptr0 + 16);
-                    __m256 _r05 = _mm256_broadcast_ss(r0 + 4);
-                    __m256 _r15 = _mm256_broadcast_ss(r1 + 4);
-                    __m256 _r25 = _mm256_broadcast_ss(r2 + 4);
-                    _mm256_storeu_ps(outptr0 + 8, _sum1);
-
-                    _sum2 = _mm256_comp_fmadd_ps(_r03, _k00, _sum2);
-                    _sum2 = _mm256_comp_fmadd_ps(_r04, _k01, _sum2);
-                    _sum2 = _mm256_comp_fmadd_ps(_r05, _k02, _sum2);
-                    _sum2 = _mm256_comp_fmadd_ps(_r13, _k10, _sum2);
-                    _sum2 = _mm256_comp_fmadd_ps(_r14, _k11, _sum2);
-                    _sum2 = _mm256_comp_fmadd_ps(_r15, _k12, _sum2);
-                    _sum2 = _mm256_comp_fmadd_ps(_r23, _k20, _sum2);
-                    _sum2 = _mm256_comp_fmadd_ps(_r24, _k21, _sum2);
-                    _sum2 = _mm256_comp_fmadd_ps(_r25, _k22, _sum2);
-
-                    __m256 _sum3 = _mm256_loadu_ps(outptr0 + 24);
-                    __m256 _r06 = _mm256_broadcast_ss(r0 + 5);
-                    __m256 _r16 = _mm256_broadcast_ss(r1 + 5);
-                    __m256 _r26 = _mm256_broadcast_ss(r2 + 5);
-                    _mm256_storeu_ps(outptr0 + 16, _sum2);
-
-                    _sum3 = _mm256_comp_fmadd_ps(_r04, _k00, _sum3);
-                    _sum3 = _mm256_comp_fmadd_ps(_r05, _k01, _sum3);
-                    _sum3 = _mm256_comp_fmadd_ps(_r06, _k02, _sum3);
-                    _sum3 = _mm256_comp_fmadd_ps(_r14, _k10, _sum3);
-                    _sum3 = _mm256_comp_fmadd_ps(_r15, _k11, _sum3);
-                    _sum3 = _mm256_comp_fmadd_ps(_r16, _k12, _sum3);
-                    _sum3 = _mm256_comp_fmadd_ps(_r24, _k20, _sum3);
-                    _sum3 = _mm256_comp_fmadd_ps(_r25, _k21, _sum3);
-                    _sum3 = _mm256_comp_fmadd_ps(_r26, _k22, _sum3);
-
-                    _mm256_storeu_ps(outptr0 + 24, _sum3);
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                    outptr0 += 32;
-                }
-                for (; j + 1 < outw; j += 2)
-                {
-                    __m256 _sum0 = _mm256_loadu_ps(outptr0);
-
-                    __m256 _r01 = _mm256_broadcast_ss(r0);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r11 = _mm256_broadcast_ss(r1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r21 = _mm256_broadcast_ss(r2);
-                    __m256 _r22 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r23 = _mm256_broadcast_ss(r2 + 2);
-
-                    _sum0 = _mm256_comp_fmadd_ps(_r01, _k00, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r02, _k01, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r03, _k02, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r11, _k10, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r12, _k11, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r13, _k12, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r21, _k20, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r22, _k21, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r23, _k22, _sum0);
-
-                    __m256 _sum1 = _mm256_loadu_ps(outptr0 + 8);
-                    __m256 _r04 = _mm256_broadcast_ss(r0 + 3);
-                    __m256 _r14 = _mm256_broadcast_ss(r1 + 3);
-                    __m256 _r24 = _mm256_broadcast_ss(r2 + 3);
-                    _mm256_storeu_ps(outptr0, _sum0);
-
-                    _sum1 = _mm256_comp_fmadd_ps(_r02, _k00, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r03, _k01, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r04, _k02, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r12, _k10, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r13, _k11, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r14, _k12, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r22, _k20, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r23, _k21, _sum1);
-                    _sum1 = _mm256_comp_fmadd_ps(_r24, _k22, _sum1);
-
-                    _mm256_storeu_ps(outptr0 + 8, _sum1);
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0 += 16;
-                }
-                for (; j < outw; j++)
-                {
-                    __m256 _sum0 = _mm256_loadu_ps(outptr0);
-
-                    __m256 _r01 = _mm256_broadcast_ss(r0);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r11 = _mm256_broadcast_ss(r1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r21 = _mm256_broadcast_ss(r2);
-                    __m256 _r22 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r23 = _mm256_broadcast_ss(r2 + 2);
-
-                    _sum0 = _mm256_comp_fmadd_ps(_r01, _k00, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r02, _k01, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r03, _k02, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r11, _k10, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r12, _k11, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r13, _k12, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r21, _k20, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r22, _k21, _sum0);
-                    _sum0 = _mm256_comp_fmadd_ps(_r23, _k22, _sum0);
-
-                    _mm256_storeu_ps(outptr0, _sum0);
-                    r0 += 1;
-                    r1 += 1;
-                    r2 += 1;
-                    outptr0 += 8;
-                }
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-            }
-
-            k0 += 9 * 8;
-        }
-    }
-}
-
-static void conv3x3s2_pack1to8_avx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* bias = _bias;
-
-    int nn_outch = outch >> 1;
-    int remain_outch_start = nn_outch << 1;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        int p = pp * 2;
-
-        Mat out0 = top_blob.channel(p);
-        Mat out1 = top_blob.channel(p + 1);
-
-        __m256 _bias0 = bias ? _mm256_loadu_ps((const float*)bias + p * 8) : _mm256_set1_ps(0.f);
-        __m256 _bias1 = bias ? _mm256_loadu_ps((const float*)bias + (p + 1) * 8) : _mm256_set1_ps(0.f);
-        out0.fill(_bias0);
-        out1.fill(_bias1);
-
-        const float* k0 = kernel.channel(p);
-        const float* k1 = kernel.channel(p + 1);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0;
-            float* outptr1 = out1;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            __m256 _k00_0 = _mm256_loadu_ps(k0);
-            __m256 _k01_0 = _mm256_loadu_ps(k0 + 8);
-            __m256 _k02_0 = _mm256_loadu_ps(k0 + 16);
-            __m256 _k10_0 = _mm256_loadu_ps(k0 + 24);
-            __m256 _k11_0 = _mm256_loadu_ps(k0 + 32);
-            __m256 _k12_0 = _mm256_loadu_ps(k0 + 40);
-            __m256 _k20_0 = _mm256_loadu_ps(k0 + 48);
-            __m256 _k21_0 = _mm256_loadu_ps(k0 + 56);
-            __m256 _k22_0 = _mm256_loadu_ps(k0 + 64);
-
-            __m256 _k00_1 = _mm256_loadu_ps(k1);
-            __m256 _k01_1 = _mm256_loadu_ps(k1 + 8);
-            __m256 _k02_1 = _mm256_loadu_ps(k1 + 16);
-            __m256 _k10_1 = _mm256_loadu_ps(k1 + 24);
-            __m256 _k11_1 = _mm256_loadu_ps(k1 + 32);
-            __m256 _k12_1 = _mm256_loadu_ps(k1 + 40);
-            __m256 _k20_1 = _mm256_loadu_ps(k1 + 48);
-            __m256 _k21_1 = _mm256_loadu_ps(k1 + 56);
-            __m256 _k22_1 = _mm256_loadu_ps(k1 + 64);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    __m256 _sum00 = _mm256_loadu_ps(outptr0);
-                    __m256 _sum10 = _mm256_loadu_ps(outptr1);
-
-                    __m256 _r01 = _mm256_broadcast_ss(r0);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r11 = _mm256_broadcast_ss(r1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r21 = _mm256_broadcast_ss(r2);
-                    __m256 _r22 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r23 = _mm256_broadcast_ss(r2 + 2);
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r01, _k00_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r02, _k01_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r03, _k02_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r11, _k10_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r12, _k11_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r13, _k12_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r21, _k20_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r22, _k21_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r23, _k22_1, _sum10);
-
-                    _mm256_storeu_ps(outptr0, _sum00);
-                    _mm256_storeu_ps(outptr1, _sum10);
-
-                    __m256 _sum01 = _mm256_loadu_ps(outptr0 + 8);
-                    __m256 _sum11 = _mm256_loadu_ps(outptr1 + 8);
-
-                    __m256 _r04 = _mm256_broadcast_ss(r0 + 3);
-                    __m256 _r14 = _mm256_broadcast_ss(r1 + 3);
-                    __m256 _r24 = _mm256_broadcast_ss(r2 + 3);
-                    __m256 _r05 = _mm256_broadcast_ss(r0 + 4);
-                    __m256 _r15 = _mm256_broadcast_ss(r1 + 4);
-                    __m256 _r25 = _mm256_broadcast_ss(r2 + 4);
-
-                    _sum01 = _mm256_comp_fmadd_ps(_r03, _k00_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r04, _k01_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r05, _k02_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r13, _k10_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r14, _k11_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r15, _k12_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r23, _k20_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r24, _k21_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r25, _k22_0, _sum01);
-
-                    _sum11 = _mm256_comp_fmadd_ps(_r03, _k00_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r04, _k01_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r05, _k02_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r13, _k10_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r14, _k11_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r15, _k12_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r23, _k20_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r24, _k21_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r25, _k22_1, _sum11);
-
-                    _mm256_storeu_ps(outptr0 + 8, _sum01);
-                    _mm256_storeu_ps(outptr1 + 8, _sum11);
-
-                    __m256 _sum02 = _mm256_loadu_ps(outptr0 + 16);
-                    __m256 _sum12 = _mm256_loadu_ps(outptr1 + 16);
-
-                    __m256 _r06 = _mm256_broadcast_ss(r0 + 5);
-                    __m256 _r16 = _mm256_broadcast_ss(r1 + 5);
-                    __m256 _r26 = _mm256_broadcast_ss(r2 + 5);
-                    __m256 _r07 = _mm256_broadcast_ss(r0 + 6);
-                    __m256 _r17 = _mm256_broadcast_ss(r1 + 6);
-                    __m256 _r27 = _mm256_broadcast_ss(r2 + 6);
-
-                    _sum02 = _mm256_comp_fmadd_ps(_r05, _k00_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r06, _k01_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r07, _k02_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r15, _k10_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r16, _k11_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r17, _k12_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r25, _k20_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r26, _k21_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r27, _k22_0, _sum02);
-
-                    _sum12 = _mm256_comp_fmadd_ps(_r05, _k00_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r06, _k01_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r07, _k02_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r15, _k10_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r16, _k11_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r17, _k12_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r25, _k20_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r26, _k21_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r27, _k22_1, _sum12);
-
-                    _mm256_storeu_ps(outptr0 + 16, _sum02);
-                    _mm256_storeu_ps(outptr1 + 16, _sum12);
-
-                    __m256 _r08 = _mm256_broadcast_ss(r0 + 7);
-                    __m256 _r18 = _mm256_broadcast_ss(r1 + 7);
-                    __m256 _r28 = _mm256_broadcast_ss(r2 + 7);
-                    __m256 _r09 = _mm256_broadcast_ss(r0 + 8);
-                    __m256 _r19 = _mm256_broadcast_ss(r1 + 8);
-                    __m256 _r29 = _mm256_broadcast_ss(r2 + 8);
-
-                    __m256 _sum03 = _mm256_loadu_ps(outptr0 + 24);
-                    __m256 _sum13 = _mm256_loadu_ps(outptr1 + 24);
-
-                    _sum03 = _mm256_comp_fmadd_ps(_r07, _k00_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r08, _k01_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r09, _k02_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r17, _k10_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r18, _k11_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r19, _k12_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r27, _k20_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r28, _k21_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r29, _k22_0, _sum03);
-
-                    _sum13 = _mm256_comp_fmadd_ps(_r07, _k00_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r08, _k01_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r09, _k02_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r17, _k10_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r18, _k11_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r19, _k12_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r27, _k20_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r28, _k21_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r29, _k22_1, _sum13);
-
-                    _mm256_storeu_ps(outptr0 + 24, _sum03);
-                    _mm256_storeu_ps(outptr1 + 24, _sum13);
-
-                    __m256 _r010 = _mm256_broadcast_ss(r0 + 9);
-                    __m256 _r110 = _mm256_broadcast_ss(r1 + 9);
-                    __m256 _r210 = _mm256_broadcast_ss(r2 + 9);
-                    __m256 _r011 = _mm256_broadcast_ss(r0 + 10);
-                    __m256 _r111 = _mm256_broadcast_ss(r1 + 10);
-                    __m256 _r211 = _mm256_broadcast_ss(r2 + 10);
-
-                    __m256 _sum04 = _mm256_loadu_ps(outptr0 + 32);
-                    __m256 _sum14 = _mm256_loadu_ps(outptr1 + 32);
-
-                    _sum04 = _mm256_comp_fmadd_ps(_r09, _k00_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r010, _k01_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r011, _k02_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r19, _k10_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r110, _k11_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r111, _k12_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r29, _k20_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r210, _k21_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r211, _k22_0, _sum04);
-
-                    _sum14 = _mm256_comp_fmadd_ps(_r09, _k00_1, _sum14);
-                    _sum14 = _mm256_comp_fmadd_ps(_r010, _k01_1, _sum14);
-                    _sum14 = _mm256_comp_fmadd_ps(_r011, _k02_1, _sum14);
-                    _sum14 = _mm256_comp_fmadd_ps(_r19, _k10_1, _sum14);
-                    _sum14 = _mm256_comp_fmadd_ps(_r110, _k11_1, _sum14);
-                    _sum14 = _mm256_comp_fmadd_ps(_r111, _k12_1, _sum14);
-                    _sum14 = _mm256_comp_fmadd_ps(_r29, _k20_1, _sum14);
-                    _sum14 = _mm256_comp_fmadd_ps(_r210, _k21_1, _sum14);
-                    _sum14 = _mm256_comp_fmadd_ps(_r211, _k22_1, _sum14);
-
-                    _mm256_storeu_ps(outptr0 + 32, _sum04);
-                    _mm256_storeu_ps(outptr1 + 32, _sum14);
-
-                    __m256 _r012 = _mm256_broadcast_ss(r0 + 11);
-                    __m256 _r112 = _mm256_broadcast_ss(r1 + 11);
-                    __m256 _r212 = _mm256_broadcast_ss(r2 + 11);
-                    __m256 _r013 = _mm256_broadcast_ss(r0 + 12);
-                    __m256 _r113 = _mm256_broadcast_ss(r1 + 12);
-                    __m256 _r213 = _mm256_broadcast_ss(r2 + 12);
-
-                    __m256 _sum05 = _mm256_loadu_ps(outptr0 + 40);
-                    __m256 _sum15 = _mm256_loadu_ps(outptr1 + 40);
-
-                    _sum05 = _mm256_comp_fmadd_ps(_r011, _k00_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r012, _k01_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r013, _k02_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r111, _k10_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r112, _k11_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r113, _k12_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r211, _k20_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r212, _k21_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r213, _k22_0, _sum05);
-                    _sum15 = _mm256_comp_fmadd_ps(_r011, _k00_1, _sum15);
-                    _sum15 = _mm256_comp_fmadd_ps(_r012, _k01_1, _sum15);
-                    _sum15 = _mm256_comp_fmadd_ps(_r013, _k02_1, _sum15);
-                    _sum15 = _mm256_comp_fmadd_ps(_r111, _k10_1, _sum15);
-                    _sum15 = _mm256_comp_fmadd_ps(_r112, _k11_1, _sum15);
-                    _sum15 = _mm256_comp_fmadd_ps(_r113, _k12_1, _sum15);
-                    _sum15 = _mm256_comp_fmadd_ps(_r211, _k20_1, _sum15);
-                    _sum15 = _mm256_comp_fmadd_ps(_r212, _k21_1, _sum15);
-                    _sum15 = _mm256_comp_fmadd_ps(_r213, _k22_1, _sum15);
-
-                    _mm256_storeu_ps(outptr0 + 40, _sum05);
-                    _mm256_storeu_ps(outptr1 + 40, _sum15);
-
-                    __m256 _r014 = _mm256_broadcast_ss(r0 + 13);
-                    __m256 _r114 = _mm256_broadcast_ss(r1 + 13);
-                    __m256 _r214 = _mm256_broadcast_ss(r2 + 13);
-                    __m256 _r015 = _mm256_broadcast_ss(r0 + 14);
-                    __m256 _r115 = _mm256_broadcast_ss(r1 + 14);
-                    __m256 _r215 = _mm256_broadcast_ss(r2 + 14);
-
-                    __m256 _sum06 = _mm256_loadu_ps(outptr0 + 48);
-                    __m256 _sum16 = _mm256_loadu_ps(outptr1 + 48);
-
-                    _sum06 = _mm256_comp_fmadd_ps(_r013, _k00_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r014, _k01_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r015, _k02_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r113, _k10_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r114, _k11_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r115, _k12_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r213, _k20_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r214, _k21_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r215, _k22_0, _sum06);
-                    _sum16 = _mm256_comp_fmadd_ps(_r013, _k00_1, _sum16);
-                    _sum16 = _mm256_comp_fmadd_ps(_r014, _k01_1, _sum16);
-                    _sum16 = _mm256_comp_fmadd_ps(_r015, _k02_1, _sum16);
-                    _sum16 = _mm256_comp_fmadd_ps(_r113, _k10_1, _sum16);
-                    _sum16 = _mm256_comp_fmadd_ps(_r114, _k11_1, _sum16);
-                    _sum16 = _mm256_comp_fmadd_ps(_r115, _k12_1, _sum16);
-                    _sum16 = _mm256_comp_fmadd_ps(_r213, _k20_1, _sum16);
-                    _sum16 = _mm256_comp_fmadd_ps(_r214, _k21_1, _sum16);
-                    _sum16 = _mm256_comp_fmadd_ps(_r215, _k22_1, _sum16);
-
-                    _mm256_storeu_ps(outptr0 + 48, _sum06);
-                    _mm256_storeu_ps(outptr1 + 48, _sum16);
-
-                    __m256 _r016 = _mm256_broadcast_ss(r0 + 15);
-                    __m256 _r116 = _mm256_broadcast_ss(r1 + 15);
-                    __m256 _r216 = _mm256_broadcast_ss(r2 + 15);
-                    __m256 _r017 = _mm256_broadcast_ss(r0 + 16);
-                    __m256 _r117 = _mm256_broadcast_ss(r1 + 16);
-                    __m256 _r217 = _mm256_broadcast_ss(r2 + 16);
-
-                    __m256 _sum07 = _mm256_loadu_ps(outptr0 + 56);
-                    __m256 _sum17 = _mm256_loadu_ps(outptr1 + 56);
-
-                    _sum07 = _mm256_comp_fmadd_ps(_r015, _k00_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r016, _k01_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r017, _k02_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r115, _k10_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r116, _k11_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r117, _k12_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r215, _k20_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r216, _k21_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r217, _k22_0, _sum07);
-                    _sum17 = _mm256_comp_fmadd_ps(_r015, _k00_1, _sum17);
-                    _sum17 = _mm256_comp_fmadd_ps(_r016, _k01_1, _sum17);
-                    _sum17 = _mm256_comp_fmadd_ps(_r017, _k02_1, _sum17);
-                    _sum17 = _mm256_comp_fmadd_ps(_r115, _k10_1, _sum17);
-                    _sum17 = _mm256_comp_fmadd_ps(_r116, _k11_1, _sum17);
-                    _sum17 = _mm256_comp_fmadd_ps(_r117, _k12_1, _sum17);
-                    _sum17 = _mm256_comp_fmadd_ps(_r215, _k20_1, _sum17);
-                    _sum17 = _mm256_comp_fmadd_ps(_r216, _k21_1, _sum17);
-                    _sum17 = _mm256_comp_fmadd_ps(_r217, _k22_1, _sum17);
-
-                    _mm256_storeu_ps(outptr0 + 56, _sum07);
-                    _mm256_storeu_ps(outptr1 + 56, _sum17);
-
-                    r0 += 16;
-                    r1 += 16;
-                    r2 += 16;
-                    outptr0 += 64;
-                    outptr1 += 64;
-                }
-
-                for (; j + 3 < outw; j += 4)
-                {
-                    __m256 _sum00 = _mm256_loadu_ps(outptr0);
-                    __m256 _sum10 = _mm256_loadu_ps(outptr1);
-
-                    __m256 _r01 = _mm256_broadcast_ss(r0);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r11 = _mm256_broadcast_ss(r1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r21 = _mm256_broadcast_ss(r2);
-                    __m256 _r22 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r23 = _mm256_broadcast_ss(r2 + 2);
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r01, _k00_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r02, _k01_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r03, _k02_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r11, _k10_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r12, _k11_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r13, _k12_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r21, _k20_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r22, _k21_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r23, _k22_1, _sum10);
-
-                    _mm256_storeu_ps(outptr0, _sum00);
-                    _mm256_storeu_ps(outptr1, _sum10);
-
-                    __m256 _sum01 = _mm256_loadu_ps(outptr0 + 8);
-                    __m256 _sum11 = _mm256_loadu_ps(outptr1 + 8);
-
-                    __m256 _r04 = _mm256_broadcast_ss(r0 + 3);
-                    __m256 _r14 = _mm256_broadcast_ss(r1 + 3);
-                    __m256 _r24 = _mm256_broadcast_ss(r2 + 3);
-                    __m256 _r05 = _mm256_broadcast_ss(r0 + 4);
-                    __m256 _r15 = _mm256_broadcast_ss(r1 + 4);
-                    __m256 _r25 = _mm256_broadcast_ss(r2 + 4);
-
-                    _sum01 = _mm256_comp_fmadd_ps(_r03, _k00_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r04, _k01_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r05, _k02_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r13, _k10_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r14, _k11_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r15, _k12_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r23, _k20_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r24, _k21_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r25, _k22_0, _sum01);
-
-                    _sum11 = _mm256_comp_fmadd_ps(_r03, _k00_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r04, _k01_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r05, _k02_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r13, _k10_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r14, _k11_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r15, _k12_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r23, _k20_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r24, _k21_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r25, _k22_1, _sum11);
-
-                    _mm256_storeu_ps(outptr0 + 8, _sum01);
-                    _mm256_storeu_ps(outptr1 + 8, _sum11);
-
-                    __m256 _sum02 = _mm256_loadu_ps(outptr0 + 16);
-                    __m256 _sum12 = _mm256_loadu_ps(outptr1 + 16);
-
-                    __m256 _r06 = _mm256_broadcast_ss(r0 + 5);
-                    __m256 _r16 = _mm256_broadcast_ss(r1 + 5);
-                    __m256 _r26 = _mm256_broadcast_ss(r2 + 5);
-                    __m256 _r07 = _mm256_broadcast_ss(r0 + 6);
-                    __m256 _r17 = _mm256_broadcast_ss(r1 + 6);
-                    __m256 _r27 = _mm256_broadcast_ss(r2 + 6);
-
-                    _sum02 = _mm256_comp_fmadd_ps(_r05, _k00_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r06, _k01_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r07, _k02_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r15, _k10_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r16, _k11_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r17, _k12_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r25, _k20_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r26, _k21_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r27, _k22_0, _sum02);
-
-                    _sum12 = _mm256_comp_fmadd_ps(_r05, _k00_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r06, _k01_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r07, _k02_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r15, _k10_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r16, _k11_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r17, _k12_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r25, _k20_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r26, _k21_1, _sum12);
-                    _sum12 = _mm256_comp_fmadd_ps(_r27, _k22_1, _sum12);
-
-                    _mm256_storeu_ps(outptr0 + 16, _sum02);
-                    _mm256_storeu_ps(outptr1 + 16, _sum12);
-
-                    __m256 _r08 = _mm256_broadcast_ss(r0 + 7);
-                    __m256 _r18 = _mm256_broadcast_ss(r1 + 7);
-                    __m256 _r28 = _mm256_broadcast_ss(r2 + 7);
-                    __m256 _r09 = _mm256_broadcast_ss(r0 + 8);
-                    __m256 _r19 = _mm256_broadcast_ss(r1 + 8);
-                    __m256 _r29 = _mm256_broadcast_ss(r2 + 8);
-
-                    __m256 _sum03 = _mm256_loadu_ps(outptr0 + 24);
-                    __m256 _sum13 = _mm256_loadu_ps(outptr1 + 24);
-
-                    _sum03 = _mm256_comp_fmadd_ps(_r07, _k00_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r08, _k01_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r09, _k02_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r17, _k10_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r18, _k11_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r19, _k12_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r27, _k20_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r28, _k21_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r29, _k22_0, _sum03);
-
-                    _sum13 = _mm256_comp_fmadd_ps(_r07, _k00_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r08, _k01_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r09, _k02_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r17, _k10_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r18, _k11_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r19, _k12_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r27, _k20_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r28, _k21_1, _sum13);
-                    _sum13 = _mm256_comp_fmadd_ps(_r29, _k22_1, _sum13);
-
-                    _mm256_storeu_ps(outptr0 + 24, _sum03);
-                    _mm256_storeu_ps(outptr1 + 24, _sum13);
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                    outptr0 += 32;
-                    outptr1 += 32;
-                }
-
-                for (; j + 1 < outw; j += 2)
-                {
-                    __m256 _sum00 = _mm256_loadu_ps(outptr0);
-                    __m256 _sum10 = _mm256_loadu_ps(outptr1);
-
-                    __m256 _r01 = _mm256_broadcast_ss(r0);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r11 = _mm256_broadcast_ss(r1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r21 = _mm256_broadcast_ss(r2);
-                    __m256 _r22 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r23 = _mm256_broadcast_ss(r2 + 2);
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r01, _k00_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r02, _k01_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r03, _k02_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r11, _k10_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r12, _k11_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r13, _k12_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r21, _k20_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r22, _k21_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r23, _k22_1, _sum10);
-
-                    _mm256_storeu_ps(outptr0, _sum00);
-                    _mm256_storeu_ps(outptr1, _sum10);
-
-                    __m256 _sum01 = _mm256_loadu_ps(outptr0 + 8);
-                    __m256 _sum11 = _mm256_loadu_ps(outptr1 + 8);
-
-                    __m256 _r04 = _mm256_broadcast_ss(r0 + 3);
-                    __m256 _r14 = _mm256_broadcast_ss(r1 + 3);
-                    __m256 _r24 = _mm256_broadcast_ss(r2 + 3);
-                    __m256 _r05 = _mm256_broadcast_ss(r0 + 4);
-                    __m256 _r15 = _mm256_broadcast_ss(r1 + 4);
-                    __m256 _r25 = _mm256_broadcast_ss(r2 + 4);
-
-                    _sum01 = _mm256_comp_fmadd_ps(_r03, _k00_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r04, _k01_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r05, _k02_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r13, _k10_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r14, _k11_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r15, _k12_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r23, _k20_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r24, _k21_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r25, _k22_0, _sum01);
-
-                    _sum11 = _mm256_comp_fmadd_ps(_r03, _k00_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r04, _k01_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r05, _k02_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r13, _k10_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r14, _k11_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r15, _k12_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r23, _k20_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r24, _k21_1, _sum11);
-                    _sum11 = _mm256_comp_fmadd_ps(_r25, _k22_1, _sum11);
-
-                    _mm256_storeu_ps(outptr0 + 8, _sum01);
-                    _mm256_storeu_ps(outptr1 + 8, _sum11);
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                    outptr0 += 16;
-                    outptr1 += 16;
-                }
-                for (; j < outw; j++)
-                {
-                    __m256 _sum00 = _mm256_loadu_ps(outptr0);
-                    __m256 _sum10 = _mm256_loadu_ps(outptr1);
-
-                    __m256 _r01 = _mm256_broadcast_ss(r0);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r11 = _mm256_broadcast_ss(r1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r21 = _mm256_broadcast_ss(r2);
-                    __m256 _r22 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r23 = _mm256_broadcast_ss(r2 + 2);
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r01, _k00_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r02, _k01_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r03, _k02_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r11, _k10_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r12, _k11_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r13, _k12_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r21, _k20_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r22, _k21_1, _sum10);
-                    _sum10 = _mm256_comp_fmadd_ps(_r23, _k22_1, _sum10);
-
-                    _mm256_storeu_ps(outptr0, _sum00);
-                    _mm256_storeu_ps(outptr1, _sum10);
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0 += 8;
-                    outptr1 += 8;
-                }
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * 8;
-            k1 += 9 * 8;
-        }
-    }
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        __m256 _bias0 = bias ? _mm256_loadu_ps((const float*)bias + p * 8) : _mm256_set1_ps(0.f);
-        out0.fill(_bias0);
-
-        const float* k0 = kernel.channel(p);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            __m256 _k00_0 = _mm256_loadu_ps(k0);
-            __m256 _k01_0 = _mm256_loadu_ps(k0 + 8);
-            __m256 _k02_0 = _mm256_loadu_ps(k0 + 16);
-            __m256 _k10_0 = _mm256_loadu_ps(k0 + 24);
-            __m256 _k11_0 = _mm256_loadu_ps(k0 + 32);
-            __m256 _k12_0 = _mm256_loadu_ps(k0 + 40);
-            __m256 _k20_0 = _mm256_loadu_ps(k0 + 48);
-            __m256 _k21_0 = _mm256_loadu_ps(k0 + 56);
-            __m256 _k22_0 = _mm256_loadu_ps(k0 + 64);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 7 < outw; j += 8)
-                {
-                    __m256 _sum00 = _mm256_loadu_ps(outptr0);
-
-                    __m256 _r01 = _mm256_broadcast_ss(r0);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r11 = _mm256_broadcast_ss(r1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r21 = _mm256_broadcast_ss(r2);
-                    __m256 _r22 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r23 = _mm256_broadcast_ss(r2 + 2);
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _mm256_storeu_ps(outptr0, _sum00);
-
-                    __m256 _sum01 = _mm256_loadu_ps(outptr0 + 8);
-
-                    __m256 _r04 = _mm256_broadcast_ss(r0 + 3);
-                    __m256 _r14 = _mm256_broadcast_ss(r1 + 3);
-                    __m256 _r24 = _mm256_broadcast_ss(r2 + 3);
-                    __m256 _r05 = _mm256_broadcast_ss(r0 + 4);
-                    __m256 _r15 = _mm256_broadcast_ss(r1 + 4);
-                    __m256 _r25 = _mm256_broadcast_ss(r2 + 4);
-
-                    _sum01 = _mm256_comp_fmadd_ps(_r03, _k00_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r04, _k01_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r05, _k02_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r13, _k10_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r14, _k11_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r15, _k12_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r23, _k20_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r24, _k21_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r25, _k22_0, _sum01);
-
-                    _mm256_storeu_ps(outptr0 + 8, _sum01);
-
-                    __m256 _sum02 = _mm256_loadu_ps(outptr0 + 16);
-
-                    __m256 _r06 = _mm256_broadcast_ss(r0 + 5);
-                    __m256 _r16 = _mm256_broadcast_ss(r1 + 5);
-                    __m256 _r26 = _mm256_broadcast_ss(r2 + 5);
-                    __m256 _r07 = _mm256_broadcast_ss(r0 + 6);
-                    __m256 _r17 = _mm256_broadcast_ss(r1 + 6);
-                    __m256 _r27 = _mm256_broadcast_ss(r2 + 6);
-
-                    _sum02 = _mm256_comp_fmadd_ps(_r05, _k00_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r06, _k01_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r07, _k02_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r15, _k10_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r16, _k11_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r17, _k12_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r25, _k20_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r26, _k21_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r27, _k22_0, _sum02);
-
-                    _mm256_storeu_ps(outptr0 + 16, _sum02);
-
-                    __m256 _r08 = _mm256_broadcast_ss(r0 + 7);
-                    __m256 _r18 = _mm256_broadcast_ss(r1 + 7);
-                    __m256 _r28 = _mm256_broadcast_ss(r2 + 7);
-                    __m256 _r09 = _mm256_broadcast_ss(r0 + 8);
-                    __m256 _r19 = _mm256_broadcast_ss(r1 + 8);
-                    __m256 _r29 = _mm256_broadcast_ss(r2 + 8);
-
-                    __m256 _sum03 = _mm256_loadu_ps(outptr0 + 24);
-
-                    _sum03 = _mm256_comp_fmadd_ps(_r07, _k00_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r08, _k01_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r09, _k02_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r17, _k10_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r18, _k11_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r19, _k12_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r27, _k20_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r28, _k21_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r29, _k22_0, _sum03);
-
-                    _mm256_storeu_ps(outptr0 + 24, _sum03);
-
-                    __m256 _r010 = _mm256_broadcast_ss(r0 + 9);
-                    __m256 _r110 = _mm256_broadcast_ss(r1 + 9);
-                    __m256 _r210 = _mm256_broadcast_ss(r2 + 9);
-                    __m256 _r011 = _mm256_broadcast_ss(r0 + 10);
-                    __m256 _r111 = _mm256_broadcast_ss(r1 + 10);
-                    __m256 _r211 = _mm256_broadcast_ss(r2 + 10);
-
-                    __m256 _sum04 = _mm256_loadu_ps(outptr0 + 32);
-
-                    _sum04 = _mm256_comp_fmadd_ps(_r09, _k00_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r010, _k01_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r011, _k02_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r19, _k10_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r110, _k11_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r111, _k12_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r29, _k20_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r210, _k21_0, _sum04);
-                    _sum04 = _mm256_comp_fmadd_ps(_r211, _k22_0, _sum04);
-
-                    _mm256_storeu_ps(outptr0 + 32, _sum04);
-
-                    __m256 _r012 = _mm256_broadcast_ss(r0 + 11);
-                    __m256 _r112 = _mm256_broadcast_ss(r1 + 11);
-                    __m256 _r212 = _mm256_broadcast_ss(r2 + 11);
-                    __m256 _r013 = _mm256_broadcast_ss(r0 + 12);
-                    __m256 _r113 = _mm256_broadcast_ss(r1 + 12);
-                    __m256 _r213 = _mm256_broadcast_ss(r2 + 12);
-
-                    __m256 _sum05 = _mm256_loadu_ps(outptr0 + 40);
-
-                    _sum05 = _mm256_comp_fmadd_ps(_r011, _k00_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r012, _k01_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r013, _k02_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r111, _k10_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r112, _k11_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r113, _k12_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r211, _k20_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r212, _k21_0, _sum05);
-                    _sum05 = _mm256_comp_fmadd_ps(_r213, _k22_0, _sum05);
-
-                    _mm256_storeu_ps(outptr0 + 40, _sum05);
-
-                    __m256 _r014 = _mm256_broadcast_ss(r0 + 13);
-                    __m256 _r114 = _mm256_broadcast_ss(r1 + 13);
-                    __m256 _r214 = _mm256_broadcast_ss(r2 + 13);
-                    __m256 _r015 = _mm256_broadcast_ss(r0 + 14);
-                    __m256 _r115 = _mm256_broadcast_ss(r1 + 14);
-                    __m256 _r215 = _mm256_broadcast_ss(r2 + 14);
-
-                    __m256 _sum06 = _mm256_loadu_ps(outptr0 + 48);
-
-                    _sum06 = _mm256_comp_fmadd_ps(_r013, _k00_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r014, _k01_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r015, _k02_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r113, _k10_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r114, _k11_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r115, _k12_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r213, _k20_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r214, _k21_0, _sum06);
-                    _sum06 = _mm256_comp_fmadd_ps(_r215, _k22_0, _sum06);
-
-                    _mm256_storeu_ps(outptr0 + 48, _sum06);
-
-                    __m256 _r016 = _mm256_broadcast_ss(r0 + 15);
-                    __m256 _r116 = _mm256_broadcast_ss(r1 + 15);
-                    __m256 _r216 = _mm256_broadcast_ss(r2 + 15);
-                    __m256 _r017 = _mm256_broadcast_ss(r0 + 16);
-                    __m256 _r117 = _mm256_broadcast_ss(r1 + 16);
-                    __m256 _r217 = _mm256_broadcast_ss(r2 + 16);
-
-                    __m256 _sum07 = _mm256_loadu_ps(outptr0 + 56);
-
-                    _sum07 = _mm256_comp_fmadd_ps(_r015, _k00_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r016, _k01_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r017, _k02_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r115, _k10_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r116, _k11_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r117, _k12_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r215, _k20_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r216, _k21_0, _sum07);
-                    _sum07 = _mm256_comp_fmadd_ps(_r217, _k22_0, _sum07);
-
-                    _mm256_storeu_ps(outptr0 + 56, _sum07);
-
-                    r0 += 16;
-                    r1 += 16;
-                    r2 += 16;
-                    outptr0 += 64;
-                }
-
-                for (; j + 3 < outw; j += 4)
-                {
-                    __m256 _sum00 = _mm256_loadu_ps(outptr0);
-                    __m256 _r01 = _mm256_broadcast_ss(r0);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r11 = _mm256_broadcast_ss(r1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r21 = _mm256_broadcast_ss(r2);
-                    __m256 _r22 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r23 = _mm256_broadcast_ss(r2 + 2);
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _mm256_storeu_ps(outptr0, _sum00);
-
-                    __m256 _sum01 = _mm256_loadu_ps(outptr0 + 8);
-
-                    __m256 _r04 = _mm256_broadcast_ss(r0 + 3);
-                    __m256 _r14 = _mm256_broadcast_ss(r1 + 3);
-                    __m256 _r24 = _mm256_broadcast_ss(r2 + 3);
-                    __m256 _r05 = _mm256_broadcast_ss(r0 + 4);
-                    __m256 _r15 = _mm256_broadcast_ss(r1 + 4);
-                    __m256 _r25 = _mm256_broadcast_ss(r2 + 4);
-
-                    _sum01 = _mm256_comp_fmadd_ps(_r03, _k00_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r04, _k01_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r05, _k02_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r13, _k10_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r14, _k11_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r15, _k12_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r23, _k20_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r24, _k21_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r25, _k22_0, _sum01);
-
-                    _mm256_storeu_ps(outptr0 + 8, _sum01);
-
-                    __m256 _sum02 = _mm256_loadu_ps(outptr0 + 16);
-
-                    __m256 _r06 = _mm256_broadcast_ss(r0 + 5);
-                    __m256 _r16 = _mm256_broadcast_ss(r1 + 5);
-                    __m256 _r26 = _mm256_broadcast_ss(r2 + 5);
-                    __m256 _r07 = _mm256_broadcast_ss(r0 + 6);
-                    __m256 _r17 = _mm256_broadcast_ss(r1 + 6);
-                    __m256 _r27 = _mm256_broadcast_ss(r2 + 6);
-
-                    _sum02 = _mm256_comp_fmadd_ps(_r05, _k00_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r06, _k01_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r07, _k02_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r15, _k10_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r16, _k11_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r17, _k12_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r25, _k20_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r26, _k21_0, _sum02);
-                    _sum02 = _mm256_comp_fmadd_ps(_r27, _k22_0, _sum02);
-
-                    _mm256_storeu_ps(outptr0 + 16, _sum02);
-
-                    __m256 _r08 = _mm256_broadcast_ss(r0 + 7);
-                    __m256 _r18 = _mm256_broadcast_ss(r1 + 7);
-                    __m256 _r28 = _mm256_broadcast_ss(r2 + 7);
-                    __m256 _r09 = _mm256_broadcast_ss(r0 + 8);
-                    __m256 _r19 = _mm256_broadcast_ss(r1 + 8);
-                    __m256 _r29 = _mm256_broadcast_ss(r2 + 8);
-
-                    __m256 _sum03 = _mm256_loadu_ps(outptr0 + 24);
-
-                    _sum03 = _mm256_comp_fmadd_ps(_r07, _k00_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r08, _k01_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r09, _k02_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r17, _k10_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r18, _k11_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r19, _k12_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r27, _k20_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r28, _k21_0, _sum03);
-                    _sum03 = _mm256_comp_fmadd_ps(_r29, _k22_0, _sum03);
-
-                    _mm256_storeu_ps(outptr0 + 24, _sum03);
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                    outptr0 += 32;
-                }
-
-                for (; j + 1 < outw; j += 2)
-                {
-                    __m256 _sum00 = _mm256_loadu_ps(outptr0);
-
-                    __m256 _r01 = _mm256_broadcast_ss(r0);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r11 = _mm256_broadcast_ss(r1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r21 = _mm256_broadcast_ss(r2);
-                    __m256 _r22 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r23 = _mm256_broadcast_ss(r2 + 2);
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r23, _k22_0, _sum00);
-
-                    _mm256_storeu_ps(outptr0, _sum00);
-
-                    __m256 _sum01 = _mm256_loadu_ps(outptr0 + 8);
-
-                    __m256 _r04 = _mm256_broadcast_ss(r0 + 3);
-                    __m256 _r14 = _mm256_broadcast_ss(r1 + 3);
-                    __m256 _r24 = _mm256_broadcast_ss(r2 + 3);
-                    __m256 _r05 = _mm256_broadcast_ss(r0 + 4);
-                    __m256 _r15 = _mm256_broadcast_ss(r1 + 4);
-                    __m256 _r25 = _mm256_broadcast_ss(r2 + 4);
-
-                    _sum01 = _mm256_comp_fmadd_ps(_r03, _k00_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r04, _k01_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r05, _k02_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r13, _k10_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r14, _k11_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r15, _k12_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r23, _k20_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r24, _k21_0, _sum01);
-                    _sum01 = _mm256_comp_fmadd_ps(_r25, _k22_0, _sum01);
-
-                    _mm256_storeu_ps(outptr0 + 8, _sum01);
-
-                    r0 += 4;
-                    r1 += 4;
-                    r2 += 4;
-                    outptr0 += 16;
-                }
-                for (; j < outw; j++)
-                {
-                    __m256 _sum00 = _mm256_loadu_ps(outptr0);
-                    __m256 _r01 = _mm256_broadcast_ss(r0);
-                    __m256 _r02 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r03 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r11 = _mm256_broadcast_ss(r1);
-                    __m256 _r12 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r13 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r21 = _mm256_broadcast_ss(r2);
-                    __m256 _r22 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r23 = _mm256_broadcast_ss(r2 + 2);
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r01, _k00_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r02, _k01_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r03, _k02_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r11, _k10_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r12, _k11_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r13, _k12_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r21, _k20_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r22, _k21_0, _sum00);
-                    _sum00 = _mm256_comp_fmadd_ps(_r23, _k22_0, _sum00);
-                    _mm256_storeu_ps(outptr0, _sum00);
-
-                    r0 += 2;
-                    r1 += 2;
-                    r2 += 2;
-                    outptr0 += 8;
-                }
-                r0 += tailstep;
-                r1 += tailstep;
-                r2 += tailstep;
-            }
-
-            k0 += 9 * 8;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack8.h
deleted file mode 100644
index fea49bea..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack8.h
+++ /dev/null
@@ -1,718 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack8_avx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        __m256 _bias0 = bias ? _mm256_loadu_ps(bias + p * 8) : _mm256_setzero_ps();
-        out.fill(_bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr = out;
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            const float* r0 = img0.row(0);
-            const float* r1 = img0.row(1);
-            const float* r2 = img0.row(2);
-
-            const float* kptr = kernel.channel(p).row(q);
-
-            int i = 0;
-            for (; i < outh; i++)
-            {
-                int j = 0;
-                for (; j + 1 < outw; j += 2)
-                {
-                    __m256 _sum00 = _mm256_loadu_ps(outptr);
-                    __m256 _sum01 = _mm256_setzero_ps();
-                    __m256 _sum10 = _mm256_loadu_ps(outptr + 8);
-                    __m256 _sum11 = _mm256_setzero_ps();
-
-                    __m256 _r000 = _mm256_broadcast_ss(r0 + 0);
-                    __m256 _r001 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r002 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r003 = _mm256_broadcast_ss(r0 + 3);
-                    __m256 _r004 = _mm256_broadcast_ss(r0 + 4);
-                    __m256 _r005 = _mm256_broadcast_ss(r0 + 5);
-                    __m256 _r006 = _mm256_broadcast_ss(r0 + 6);
-                    __m256 _r007 = _mm256_broadcast_ss(r0 + 7);
-
-                    __m256 _k00 = _mm256_loadu_ps(kptr);
-                    __m256 _k01 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k02 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k03 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k04 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k05 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k06 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k07 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r000, _k00, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r001, _k01, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r002, _k02, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r003, _k03, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r004, _k04, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r005, _k05, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r006, _k06, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r007, _k07, _sum01);
-
-                    __m256 _r010 = _mm256_broadcast_ss(r0 + 8);
-                    __m256 _r011 = _mm256_broadcast_ss(r0 + 9);
-                    __m256 _r012 = _mm256_broadcast_ss(r0 + 10);
-                    __m256 _r013 = _mm256_broadcast_ss(r0 + 11);
-                    __m256 _r014 = _mm256_broadcast_ss(r0 + 12);
-                    __m256 _r015 = _mm256_broadcast_ss(r0 + 13);
-                    __m256 _r016 = _mm256_broadcast_ss(r0 + 14);
-                    __m256 _r017 = _mm256_broadcast_ss(r0 + 15);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r010, _k00, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r011, _k01, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r012, _k02, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r013, _k03, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r014, _k04, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r015, _k05, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r016, _k06, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r017, _k07, _sum11);
-
-                    __m256 _k10 = _mm256_loadu_ps(kptr);
-                    __m256 _k11 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k12 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k13 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k14 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k15 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k16 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k17 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r010, _k10, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r011, _k11, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r012, _k12, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r013, _k13, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r014, _k14, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r015, _k15, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r016, _k16, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r017, _k17, _sum01);
-
-                    __m256 _r020 = _mm256_broadcast_ss(r0 + 16);
-                    __m256 _r021 = _mm256_broadcast_ss(r0 + 17);
-                    __m256 _r022 = _mm256_broadcast_ss(r0 + 18);
-                    __m256 _r023 = _mm256_broadcast_ss(r0 + 19);
-                    __m256 _r024 = _mm256_broadcast_ss(r0 + 20);
-                    __m256 _r025 = _mm256_broadcast_ss(r0 + 21);
-                    __m256 _r026 = _mm256_broadcast_ss(r0 + 22);
-                    __m256 _r027 = _mm256_broadcast_ss(r0 + 23);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r020, _k10, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r021, _k11, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r022, _k12, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r023, _k13, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r024, _k14, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r025, _k15, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r026, _k16, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r027, _k17, _sum11);
-
-                    __m256 _k20 = _mm256_loadu_ps(kptr);
-                    __m256 _k21 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k22 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k23 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k24 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k25 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k26 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k27 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r020, _k20, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r021, _k21, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r022, _k22, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r023, _k23, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r024, _k24, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r025, _k25, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r026, _k26, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r027, _k27, _sum01);
-
-                    __m256 _r030 = _mm256_broadcast_ss(r0 + 24);
-                    __m256 _r031 = _mm256_broadcast_ss(r0 + 25);
-                    __m256 _r032 = _mm256_broadcast_ss(r0 + 26);
-                    __m256 _r033 = _mm256_broadcast_ss(r0 + 27);
-                    __m256 _r034 = _mm256_broadcast_ss(r0 + 28);
-                    __m256 _r035 = _mm256_broadcast_ss(r0 + 29);
-                    __m256 _r036 = _mm256_broadcast_ss(r0 + 30);
-                    __m256 _r037 = _mm256_broadcast_ss(r0 + 31);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r030, _k20, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r031, _k21, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r032, _k22, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r033, _k23, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r034, _k24, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r035, _k25, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r036, _k26, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r037, _k27, _sum11);
-
-                    __m256 _r100 = _mm256_broadcast_ss(r1 + 0);
-                    __m256 _r101 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r102 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r103 = _mm256_broadcast_ss(r1 + 3);
-                    __m256 _r104 = _mm256_broadcast_ss(r1 + 4);
-                    __m256 _r105 = _mm256_broadcast_ss(r1 + 5);
-                    __m256 _r106 = _mm256_broadcast_ss(r1 + 6);
-                    __m256 _r107 = _mm256_broadcast_ss(r1 + 7);
-
-                    __m256 _k30 = _mm256_loadu_ps(kptr);
-                    __m256 _k31 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k32 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k33 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k34 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k35 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k36 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k37 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r100, _k30, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r101, _k31, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r102, _k32, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r103, _k33, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r104, _k34, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r105, _k35, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r106, _k36, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r107, _k37, _sum01);
-
-                    __m256 _r110 = _mm256_broadcast_ss(r1 + 8);
-                    __m256 _r111 = _mm256_broadcast_ss(r1 + 9);
-                    __m256 _r112 = _mm256_broadcast_ss(r1 + 10);
-                    __m256 _r113 = _mm256_broadcast_ss(r1 + 11);
-                    __m256 _r114 = _mm256_broadcast_ss(r1 + 12);
-                    __m256 _r115 = _mm256_broadcast_ss(r1 + 13);
-                    __m256 _r116 = _mm256_broadcast_ss(r1 + 14);
-                    __m256 _r117 = _mm256_broadcast_ss(r1 + 15);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r110, _k30, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r111, _k31, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r112, _k32, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r113, _k33, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r114, _k34, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r115, _k35, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r116, _k36, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r117, _k37, _sum11);
-
-                    __m256 _k40 = _mm256_loadu_ps(kptr);
-                    __m256 _k41 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k42 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k43 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k44 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k45 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k46 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k47 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r110, _k40, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r111, _k41, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r112, _k42, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r113, _k43, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r114, _k44, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r115, _k45, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r116, _k46, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r117, _k47, _sum01);
-
-                    __m256 _r120 = _mm256_broadcast_ss(r1 + 16);
-                    __m256 _r121 = _mm256_broadcast_ss(r1 + 17);
-                    __m256 _r122 = _mm256_broadcast_ss(r1 + 18);
-                    __m256 _r123 = _mm256_broadcast_ss(r1 + 19);
-                    __m256 _r124 = _mm256_broadcast_ss(r1 + 20);
-                    __m256 _r125 = _mm256_broadcast_ss(r1 + 21);
-                    __m256 _r126 = _mm256_broadcast_ss(r1 + 22);
-                    __m256 _r127 = _mm256_broadcast_ss(r1 + 23);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r120, _k40, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r121, _k41, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r122, _k42, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r123, _k43, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r124, _k44, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r125, _k45, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r126, _k46, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r127, _k47, _sum11);
-
-                    __m256 _k50 = _mm256_loadu_ps(kptr);
-                    __m256 _k51 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k52 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k53 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k54 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k55 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k56 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k57 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r120, _k50, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r121, _k51, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r122, _k52, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r123, _k53, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r124, _k54, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r125, _k55, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r126, _k56, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r127, _k57, _sum01);
-
-                    __m256 _r130 = _mm256_broadcast_ss(r1 + 24);
-                    __m256 _r131 = _mm256_broadcast_ss(r1 + 25);
-                    __m256 _r132 = _mm256_broadcast_ss(r1 + 26);
-                    __m256 _r133 = _mm256_broadcast_ss(r1 + 27);
-                    __m256 _r134 = _mm256_broadcast_ss(r1 + 28);
-                    __m256 _r135 = _mm256_broadcast_ss(r1 + 29);
-                    __m256 _r136 = _mm256_broadcast_ss(r1 + 30);
-                    __m256 _r137 = _mm256_broadcast_ss(r1 + 31);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r130, _k50, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r131, _k51, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r132, _k52, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r133, _k53, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r134, _k54, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r135, _k55, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r136, _k56, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r137, _k57, _sum11);
-
-                    __m256 _r200 = _mm256_broadcast_ss(r2 + 0);
-                    __m256 _r201 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r202 = _mm256_broadcast_ss(r2 + 2);
-                    __m256 _r203 = _mm256_broadcast_ss(r2 + 3);
-                    __m256 _r204 = _mm256_broadcast_ss(r2 + 4);
-                    __m256 _r205 = _mm256_broadcast_ss(r2 + 5);
-                    __m256 _r206 = _mm256_broadcast_ss(r2 + 6);
-                    __m256 _r207 = _mm256_broadcast_ss(r2 + 7);
-
-                    __m256 _k60 = _mm256_loadu_ps(kptr);
-                    __m256 _k61 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k62 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k63 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k64 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k65 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k66 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k67 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r200, _k60, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r201, _k61, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r202, _k62, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r203, _k63, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r204, _k64, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r205, _k65, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r206, _k66, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r207, _k67, _sum01);
-
-                    __m256 _r210 = _mm256_broadcast_ss(r2 + 8);
-                    __m256 _r211 = _mm256_broadcast_ss(r2 + 9);
-                    __m256 _r212 = _mm256_broadcast_ss(r2 + 10);
-                    __m256 _r213 = _mm256_broadcast_ss(r2 + 11);
-                    __m256 _r214 = _mm256_broadcast_ss(r2 + 12);
-                    __m256 _r215 = _mm256_broadcast_ss(r2 + 13);
-                    __m256 _r216 = _mm256_broadcast_ss(r2 + 14);
-                    __m256 _r217 = _mm256_broadcast_ss(r2 + 15);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r210, _k60, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r211, _k61, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r212, _k62, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r213, _k63, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r214, _k64, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r215, _k65, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r216, _k66, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r217, _k67, _sum11);
-
-                    __m256 _k70 = _mm256_loadu_ps(kptr);
-                    __m256 _k71 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k72 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k73 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k74 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k75 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k76 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k77 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r210, _k70, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r211, _k71, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r212, _k72, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r213, _k73, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r214, _k74, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r215, _k75, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r216, _k76, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r217, _k77, _sum01);
-
-                    __m256 _r220 = _mm256_broadcast_ss(r2 + 16);
-                    __m256 _r221 = _mm256_broadcast_ss(r2 + 17);
-                    __m256 _r222 = _mm256_broadcast_ss(r2 + 18);
-                    __m256 _r223 = _mm256_broadcast_ss(r2 + 19);
-                    __m256 _r224 = _mm256_broadcast_ss(r2 + 20);
-                    __m256 _r225 = _mm256_broadcast_ss(r2 + 21);
-                    __m256 _r226 = _mm256_broadcast_ss(r2 + 22);
-                    __m256 _r227 = _mm256_broadcast_ss(r2 + 23);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r220, _k70, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r221, _k71, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r222, _k72, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r223, _k73, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r224, _k74, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r225, _k75, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r226, _k76, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r227, _k77, _sum11);
-
-                    __m256 _k80 = _mm256_loadu_ps(kptr);
-                    __m256 _k81 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k82 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k83 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k84 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k85 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k86 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k87 = _mm256_loadu_ps(kptr + 56);
-
-                    _sum00 = _mm256_comp_fmadd_ps(_r220, _k80, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r221, _k81, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r222, _k82, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r223, _k83, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r224, _k84, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r225, _k85, _sum01);
-                    _sum00 = _mm256_comp_fmadd_ps(_r226, _k86, _sum00);
-                    _sum01 = _mm256_comp_fmadd_ps(_r227, _k87, _sum01);
-
-                    __m256 _r230 = _mm256_broadcast_ss(r2 + 24);
-                    __m256 _r231 = _mm256_broadcast_ss(r2 + 25);
-                    __m256 _r232 = _mm256_broadcast_ss(r2 + 26);
-                    __m256 _r233 = _mm256_broadcast_ss(r2 + 27);
-                    __m256 _r234 = _mm256_broadcast_ss(r2 + 28);
-                    __m256 _r235 = _mm256_broadcast_ss(r2 + 29);
-                    __m256 _r236 = _mm256_broadcast_ss(r2 + 30);
-                    __m256 _r237 = _mm256_broadcast_ss(r2 + 31);
-
-                    _sum10 = _mm256_comp_fmadd_ps(_r230, _k80, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r231, _k81, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r232, _k82, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r233, _k83, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r234, _k84, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r235, _k85, _sum11);
-                    _sum10 = _mm256_comp_fmadd_ps(_r236, _k86, _sum10);
-                    _sum11 = _mm256_comp_fmadd_ps(_r237, _k87, _sum11);
-
-                    kptr -= 64 * 8;
-
-                    _sum00 = _mm256_add_ps(_sum00, _sum01);
-                    _sum10 = _mm256_add_ps(_sum10, _sum11);
-
-                    _mm256_storeu_ps(outptr, _sum00);
-                    _mm256_storeu_ps(outptr + 8, _sum10);
-
-                    r0 += 16;
-                    r1 += 16;
-                    r2 += 16;
-                    outptr += 16;
-                }
-                for (; j < outw; j++)
-                {
-                    __m256 _sum0 = _mm256_loadu_ps(outptr);
-                    __m256 _sum1 = _mm256_setzero_ps();
-
-                    __m256 _r000 = _mm256_broadcast_ss(r0 + 0);
-                    __m256 _r001 = _mm256_broadcast_ss(r0 + 1);
-                    __m256 _r002 = _mm256_broadcast_ss(r0 + 2);
-                    __m256 _r003 = _mm256_broadcast_ss(r0 + 3);
-                    __m256 _r004 = _mm256_broadcast_ss(r0 + 4);
-                    __m256 _r005 = _mm256_broadcast_ss(r0 + 5);
-                    __m256 _r006 = _mm256_broadcast_ss(r0 + 6);
-                    __m256 _r007 = _mm256_broadcast_ss(r0 + 7);
-
-                    __m256 _k00 = _mm256_loadu_ps(kptr);
-                    __m256 _k01 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k02 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k03 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k04 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k05 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k06 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k07 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum0 = _mm256_comp_fmadd_ps(_r000, _k00, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r001, _k01, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r002, _k02, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r003, _k03, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r004, _k04, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r005, _k05, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r006, _k06, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r007, _k07, _sum1);
-
-                    __m256 _r010 = _mm256_broadcast_ss(r0 + 8);
-                    __m256 _r011 = _mm256_broadcast_ss(r0 + 9);
-                    __m256 _r012 = _mm256_broadcast_ss(r0 + 10);
-                    __m256 _r013 = _mm256_broadcast_ss(r0 + 11);
-                    __m256 _r014 = _mm256_broadcast_ss(r0 + 12);
-                    __m256 _r015 = _mm256_broadcast_ss(r0 + 13);
-                    __m256 _r016 = _mm256_broadcast_ss(r0 + 14);
-                    __m256 _r017 = _mm256_broadcast_ss(r0 + 15);
-
-                    __m256 _k10 = _mm256_loadu_ps(kptr);
-                    __m256 _k11 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k12 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k13 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k14 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k15 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k16 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k17 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum0 = _mm256_comp_fmadd_ps(_r010, _k10, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r011, _k11, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r012, _k12, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r013, _k13, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r014, _k14, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r015, _k15, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r016, _k16, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r017, _k17, _sum1);
-
-                    __m256 _r020 = _mm256_broadcast_ss(r0 + 16);
-                    __m256 _r021 = _mm256_broadcast_ss(r0 + 17);
-                    __m256 _r022 = _mm256_broadcast_ss(r0 + 18);
-                    __m256 _r023 = _mm256_broadcast_ss(r0 + 19);
-                    __m256 _r024 = _mm256_broadcast_ss(r0 + 20);
-                    __m256 _r025 = _mm256_broadcast_ss(r0 + 21);
-                    __m256 _r026 = _mm256_broadcast_ss(r0 + 22);
-                    __m256 _r027 = _mm256_broadcast_ss(r0 + 23);
-
-                    __m256 _k20 = _mm256_loadu_ps(kptr);
-                    __m256 _k21 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k22 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k23 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k24 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k25 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k26 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k27 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum0 = _mm256_comp_fmadd_ps(_r020, _k20, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r021, _k21, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r022, _k22, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r023, _k23, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r024, _k24, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r025, _k25, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r026, _k26, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r027, _k27, _sum1);
-
-                    __m256 _r100 = _mm256_broadcast_ss(r1 + 0);
-                    __m256 _r101 = _mm256_broadcast_ss(r1 + 1);
-                    __m256 _r102 = _mm256_broadcast_ss(r1 + 2);
-                    __m256 _r103 = _mm256_broadcast_ss(r1 + 3);
-                    __m256 _r104 = _mm256_broadcast_ss(r1 + 4);
-                    __m256 _r105 = _mm256_broadcast_ss(r1 + 5);
-                    __m256 _r106 = _mm256_broadcast_ss(r1 + 6);
-                    __m256 _r107 = _mm256_broadcast_ss(r1 + 7);
-
-                    __m256 _k30 = _mm256_loadu_ps(kptr);
-                    __m256 _k31 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k32 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k33 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k34 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k35 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k36 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k37 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum0 = _mm256_comp_fmadd_ps(_r100, _k30, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r101, _k31, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r102, _k32, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r103, _k33, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r104, _k34, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r105, _k35, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r106, _k36, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r107, _k37, _sum1);
-
-                    __m256 _r110 = _mm256_broadcast_ss(r1 + 8);
-                    __m256 _r111 = _mm256_broadcast_ss(r1 + 9);
-                    __m256 _r112 = _mm256_broadcast_ss(r1 + 10);
-                    __m256 _r113 = _mm256_broadcast_ss(r1 + 11);
-                    __m256 _r114 = _mm256_broadcast_ss(r1 + 12);
-                    __m256 _r115 = _mm256_broadcast_ss(r1 + 13);
-                    __m256 _r116 = _mm256_broadcast_ss(r1 + 14);
-                    __m256 _r117 = _mm256_broadcast_ss(r1 + 15);
-
-                    __m256 _k40 = _mm256_loadu_ps(kptr);
-                    __m256 _k41 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k42 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k43 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k44 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k45 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k46 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k47 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum0 = _mm256_comp_fmadd_ps(_r110, _k40, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r111, _k41, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r112, _k42, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r113, _k43, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r114, _k44, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r115, _k45, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r116, _k46, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r117, _k47, _sum1);
-
-                    __m256 _r120 = _mm256_broadcast_ss(r1 + 16);
-                    __m256 _r121 = _mm256_broadcast_ss(r1 + 17);
-                    __m256 _r122 = _mm256_broadcast_ss(r1 + 18);
-                    __m256 _r123 = _mm256_broadcast_ss(r1 + 19);
-                    __m256 _r124 = _mm256_broadcast_ss(r1 + 20);
-                    __m256 _r125 = _mm256_broadcast_ss(r1 + 21);
-                    __m256 _r126 = _mm256_broadcast_ss(r1 + 22);
-                    __m256 _r127 = _mm256_broadcast_ss(r1 + 23);
-
-                    __m256 _k50 = _mm256_loadu_ps(kptr);
-                    __m256 _k51 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k52 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k53 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k54 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k55 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k56 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k57 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum0 = _mm256_comp_fmadd_ps(_r120, _k50, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r121, _k51, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r122, _k52, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r123, _k53, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r124, _k54, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r125, _k55, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r126, _k56, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r127, _k57, _sum1);
-
-                    __m256 _r200 = _mm256_broadcast_ss(r2 + 0);
-                    __m256 _r201 = _mm256_broadcast_ss(r2 + 1);
-                    __m256 _r202 = _mm256_broadcast_ss(r2 + 2);
-                    __m256 _r203 = _mm256_broadcast_ss(r2 + 3);
-                    __m256 _r204 = _mm256_broadcast_ss(r2 + 4);
-                    __m256 _r205 = _mm256_broadcast_ss(r2 + 5);
-                    __m256 _r206 = _mm256_broadcast_ss(r2 + 6);
-                    __m256 _r207 = _mm256_broadcast_ss(r2 + 7);
-
-                    __m256 _k60 = _mm256_loadu_ps(kptr);
-                    __m256 _k61 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k62 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k63 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k64 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k65 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k66 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k67 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum0 = _mm256_comp_fmadd_ps(_r200, _k60, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r201, _k61, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r202, _k62, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r203, _k63, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r204, _k64, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r205, _k65, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r206, _k66, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r207, _k67, _sum1);
-
-                    __m256 _r210 = _mm256_broadcast_ss(r2 + 8);
-                    __m256 _r211 = _mm256_broadcast_ss(r2 + 9);
-                    __m256 _r212 = _mm256_broadcast_ss(r2 + 10);
-                    __m256 _r213 = _mm256_broadcast_ss(r2 + 11);
-                    __m256 _r214 = _mm256_broadcast_ss(r2 + 12);
-                    __m256 _r215 = _mm256_broadcast_ss(r2 + 13);
-                    __m256 _r216 = _mm256_broadcast_ss(r2 + 14);
-                    __m256 _r217 = _mm256_broadcast_ss(r2 + 15);
-
-                    __m256 _k70 = _mm256_loadu_ps(kptr);
-                    __m256 _k71 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k72 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k73 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k74 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k75 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k76 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k77 = _mm256_loadu_ps(kptr + 56);
-
-                    kptr += 64;
-
-                    _sum0 = _mm256_comp_fmadd_ps(_r210, _k70, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r211, _k71, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r212, _k72, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r213, _k73, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r214, _k74, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r215, _k75, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r216, _k76, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r217, _k77, _sum1);
-
-                    __m256 _r220 = _mm256_broadcast_ss(r2 + 16);
-                    __m256 _r221 = _mm256_broadcast_ss(r2 + 17);
-                    __m256 _r222 = _mm256_broadcast_ss(r2 + 18);
-                    __m256 _r223 = _mm256_broadcast_ss(r2 + 19);
-                    __m256 _r224 = _mm256_broadcast_ss(r2 + 20);
-                    __m256 _r225 = _mm256_broadcast_ss(r2 + 21);
-                    __m256 _r226 = _mm256_broadcast_ss(r2 + 22);
-                    __m256 _r227 = _mm256_broadcast_ss(r2 + 23);
-
-                    __m256 _k80 = _mm256_loadu_ps(kptr);
-                    __m256 _k81 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _k82 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _k83 = _mm256_loadu_ps(kptr + 24);
-                    __m256 _k84 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _k85 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _k86 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _k87 = _mm256_loadu_ps(kptr + 56);
-
-                    _sum0 = _mm256_comp_fmadd_ps(_r220, _k80, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r221, _k81, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r222, _k82, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r223, _k83, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r224, _k84, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r225, _k85, _sum1);
-                    _sum0 = _mm256_comp_fmadd_ps(_r226, _k86, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_r227, _k87, _sum1);
-
-                    kptr -= 64 * 8;
-
-                    _sum0 = _mm256_add_ps(_sum0, _sum1);
-
-                    _mm256_storeu_ps(outptr, _sum0);
-
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                    outptr += 8;
-                }
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack8to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack8to1.h
deleted file mode 100644
index f7b5301f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_pack8to1.h
+++ /dev/null
@@ -1,98 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv3x3s1_pack8to1_avx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int inch = bottom_blob.c;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* bias = _bias;
-
-    int remain_outch_start = 0;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        Mat out0 = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        out0.fill(bias0);
-
-        const float* k0 = kernel.channel(p);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr0 = out0.row(0);
-
-            const Mat img0 = bottom_blob.channel(q);
-
-            __m256 _k00 = _mm256_loadu_ps(k0);
-            __m256 _k01 = _mm256_loadu_ps(k0 + 8);
-            __m256 _k02 = _mm256_loadu_ps(k0 + 16);
-            __m256 _k10 = _mm256_loadu_ps(k0 + 24);
-            __m256 _k11 = _mm256_loadu_ps(k0 + 32);
-            __m256 _k12 = _mm256_loadu_ps(k0 + 40);
-            __m256 _k20 = _mm256_loadu_ps(k0 + 48);
-            __m256 _k21 = _mm256_loadu_ps(k0 + 56);
-            __m256 _k22 = _mm256_loadu_ps(k0 + 64);
-
-            int i = 0;
-
-            for (; i < outh; i++)
-            {
-                const float* r0 = img0.row(i);
-                const float* r1 = img0.row(i + 1);
-                const float* r2 = img0.row(i + 2);
-                int j = 0;
-                for (; j < outw; j++)
-                {
-                    __m256 _r00 = _mm256_loadu_ps(r0);
-                    __m256 _r01 = _mm256_loadu_ps(r0 + 8);
-                    __m256 _r02 = _mm256_loadu_ps(r0 + 16);
-
-                    __m256 _sum0 = _mm256_mul_ps(_k00, _r00);
-                    __m256 _sum1 = _mm256_mul_ps(_k01, _r01);
-                    __m256 _sum2 = _mm256_mul_ps(_k02, _r02);
-
-                    __m256 _r10 = _mm256_loadu_ps(r1);
-                    __m256 _r11 = _mm256_loadu_ps(r1 + 8);
-                    __m256 _r12 = _mm256_loadu_ps(r1 + 16);
-
-                    _sum0 = _mm256_comp_fmadd_ps(_k10, _r10, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_k11, _r11, _sum1);
-                    _sum2 = _mm256_comp_fmadd_ps(_k12, _r12, _sum2);
-
-                    __m256 _r20 = _mm256_loadu_ps(r2);
-                    __m256 _r21 = _mm256_loadu_ps(r2 + 8);
-                    __m256 _r22 = _mm256_loadu_ps(r2 + 16);
-
-                    _sum0 = _mm256_comp_fmadd_ps(_k20, _r20, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_k21, _r21, _sum1);
-                    _sum2 = _mm256_comp_fmadd_ps(_k22, _r22, _sum2);
-                    __m128 _sum = HorizontalSums(_sum0, _sum1, _sum2);
-
-                    *outptr0 += _mm_reduce_add_ps(_sum); // dot
-                    outptr0++;
-                    r0 += 8;
-                    r1 += 8;
-                    r2 += 8;
-                }
-            }
-
-            k0 += 9 * 8;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_winograd.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_winograd.h
deleted file mode 100644
index ca7751c2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_winograd.h
+++ /dev/null
@@ -1,7380 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pack_A_tile(const Mat& A, Mat& AT, int batch, int max_ii, int max_kk)
-{
-    const int N = max_kk * batch;
-
-    for (int b = 0; b < batch; b++)
-    {
-        float* pp = AT.row(b);
-
-        int ii = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-        for (; ii + 15 < max_ii; ii += 16)
-        {
-            const float* p0 = (const float*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                pp[2] = p0[2 * N];
-                pp[3] = p0[3 * N];
-                pp[4] = p0[4 * N];
-                pp[5] = p0[5 * N];
-                pp[6] = p0[6 * N];
-                pp[7] = p0[7 * N];
-                pp[8] = p0[8 * N];
-                pp[9] = p0[9 * N];
-                pp[10] = p0[10 * N];
-                pp[11] = p0[11 * N];
-                pp[12] = p0[12 * N];
-                pp[13] = p0[13 * N];
-                pp[14] = p0[14 * N];
-                pp[15] = p0[15 * N];
-                p0 += batch;
-                pp += 16;
-            }
-        }
-#endif // __AVX512F__
-        for (; ii + 7 < max_ii; ii += 8)
-        {
-            const float* p0 = (const float*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                pp[2] = p0[2 * N];
-                pp[3] = p0[3 * N];
-                pp[4] = p0[4 * N];
-                pp[5] = p0[5 * N];
-                pp[6] = p0[6 * N];
-                pp[7] = p0[7 * N];
-                p0 += batch;
-                pp += 8;
-            }
-        }
-#endif // __AVX__
-        for (; ii + 3 < max_ii; ii += 4)
-        {
-            const float* p0 = (const float*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                pp[2] = p0[2 * N];
-                pp[3] = p0[3 * N];
-                p0 += batch;
-                pp += 4;
-            }
-        }
-#endif // __SSE2__
-        for (; ii + 1 < max_ii; ii += 2)
-        {
-            const float* p0 = (const float*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                p0 += batch;
-                pp += 2;
-            }
-        }
-        for (; ii < max_ii; ii++)
-        {
-            const float* p0 = (const float*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                p0 += batch;
-                pp += 1;
-            }
-        }
-    }
-}
-
-static void transpose_pack_B_tile(const Mat& B, Mat& BT, int batch, int max_jj, int max_kk, int nT)
-{
-    #pragma omp parallel for num_threads(nT)
-    for (int b = 0; b < batch; b++)
-    {
-        float* pp = BT.row(b);
-
-        int jj = 0;
-#if __SSE2__
-        for (; jj + 11 < max_jj; jj += 12)
-        {
-            const float* p0 = B;
-
-            int kk = 0;
-#if __AVX__
-#if __AVX512F__
-            p0 += (b * max_jj + jj) * 16;
-            for (; kk + 15 < max_kk; kk += 16)
-            {
-                __m512 _r0 = _mm512_load_ps(p0);
-                __m512 _r1 = _mm512_load_ps(p0 + 16);
-                __m512 _r2 = _mm512_load_ps(p0 + 16 * 2);
-                __m512 _r3 = _mm512_load_ps(p0 + 16 * 3);
-                __m512 _r4 = _mm512_load_ps(p0 + 16 * 4);
-                __m512 _r5 = _mm512_load_ps(p0 + 16 * 5);
-                __m512 _r6 = _mm512_load_ps(p0 + 16 * 6);
-                __m512 _r7 = _mm512_load_ps(p0 + 16 * 7);
-                __m512 _r8 = _mm512_load_ps(p0 + 16 * 8);
-                __m512 _r9 = _mm512_load_ps(p0 + 16 * 9);
-                __m512 _ra = _mm512_load_ps(p0 + 16 * 10);
-                __m512 _rb = _mm512_load_ps(p0 + 16 * 11);
-                transpose16x12_ps(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7, _r8, _r9, _ra, _rb);
-                _mm512_storeu_ps(pp, _r0);
-                _mm512_storeu_ps(pp + 16, _r1);
-                _mm512_storeu_ps(pp + 16 * 2, _r2);
-                _mm512_storeu_ps(pp + 16 * 3, _r3);
-                _mm512_storeu_ps(pp + 16 * 4, _r4);
-                _mm512_storeu_ps(pp + 16 * 5, _r5);
-                _mm512_storeu_ps(pp + 16 * 6, _r6);
-                _mm512_storeu_ps(pp + 16 * 7, _r7);
-                _mm512_storeu_ps(pp + 16 * 8, _r8);
-                _mm512_storeu_ps(pp + 16 * 9, _r9);
-                _mm512_storeu_ps(pp + 16 * 10, _ra);
-                _mm512_storeu_ps(pp + 16 * 11, _rb);
-                p0 += max_jj * batch * 16;
-                pp += 192;
-            }
-            p0 -= (b * max_jj + jj) * 16;
-#endif // __AVX512F__
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                __m256 _r0 = _mm256_load_ps(p0);
-                __m256 _r1 = _mm256_load_ps(p0 + 8);
-                __m256 _r2 = _mm256_load_ps(p0 + 8 * 2);
-                __m256 _r3 = _mm256_load_ps(p0 + 8 * 3);
-                __m256 _r4 = _mm256_load_ps(p0 + 8 * 4);
-                __m256 _r5 = _mm256_load_ps(p0 + 8 * 5);
-                __m256 _r6 = _mm256_load_ps(p0 + 8 * 6);
-                __m256 _r7 = _mm256_load_ps(p0 + 8 * 7);
-                __m256 _r8 = _mm256_load_ps(p0 + 8 * 8);
-                __m256 _r9 = _mm256_load_ps(p0 + 8 * 9);
-                __m256 _ra = _mm256_load_ps(p0 + 8 * 10);
-                __m256 _rb = _mm256_load_ps(p0 + 8 * 11);
-                transpose8x12_ps(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7, _r8, _r9, _ra, _rb);
-                _mm256_storeu_ps(pp, _r0);
-                _mm256_storeu_ps(pp + 8, _r1);
-                _mm256_storeu_ps(pp + 8 * 2, _r2);
-                _mm256_storeu_ps(pp + 8 * 3, _r3);
-                _mm256_storeu_ps(pp + 8 * 4, _r4);
-                _mm256_storeu_ps(pp + 8 * 5, _r5);
-                _mm256_storeu_ps(pp + 8 * 6, _r6);
-                _mm256_storeu_ps(pp + 8 * 7, _r7);
-                _mm256_storeu_ps(pp + 8 * 8, _r8);
-                _mm256_storeu_ps(pp + 8 * 9, _r9);
-                _mm256_storeu_ps(pp + 8 * 10, _ra);
-                _mm256_storeu_ps(pp + 8 * 11, _rb);
-                p0 += max_jj * batch * 8;
-                pp += 96;
-            }
-            p0 -= (b * max_jj + jj) * 8;
-#endif // __AVX__
-            p0 += (b * max_jj + jj) * 4;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                __m128 _r0 = _mm_load_ps(p0);
-                __m128 _r1 = _mm_load_ps(p0 + 4);
-                __m128 _r2 = _mm_load_ps(p0 + 4 * 2);
-                __m128 _r3 = _mm_load_ps(p0 + 4 * 3);
-                __m128 _r4 = _mm_load_ps(p0 + 4 * 4);
-                __m128 _r5 = _mm_load_ps(p0 + 4 * 5);
-                __m128 _r6 = _mm_load_ps(p0 + 4 * 6);
-                __m128 _r7 = _mm_load_ps(p0 + 4 * 7);
-                __m128 _r8 = _mm_load_ps(p0 + 4 * 8);
-                __m128 _r9 = _mm_load_ps(p0 + 4 * 9);
-                __m128 _ra = _mm_load_ps(p0 + 4 * 10);
-                __m128 _rb = _mm_load_ps(p0 + 4 * 11);
-                _MM_TRANSPOSE4_PS(_r0, _r1, _r2, _r3);
-                _MM_TRANSPOSE4_PS(_r4, _r5, _r6, _r7);
-                _MM_TRANSPOSE4_PS(_r8, _r9, _ra, _rb);
-                _mm_store_ps(pp, _r0);
-                _mm_store_ps(pp + 4, _r4);
-                _mm_store_ps(pp + 4 * 2, _r8);
-                _mm_store_ps(pp + 4 * 3, _r1);
-                _mm_store_ps(pp + 4 * 4, _r5);
-                _mm_store_ps(pp + 4 * 5, _r9);
-                _mm_store_ps(pp + 4 * 6, _r2);
-                _mm_store_ps(pp + 4 * 7, _r6);
-                _mm_store_ps(pp + 4 * 8, _ra);
-                _mm_store_ps(pp + 4 * 9, _r3);
-                _mm_store_ps(pp + 4 * 10, _r7);
-                _mm_store_ps(pp + 4 * 11, _rb);
-                p0 += max_jj * batch * 4;
-                pp += 48;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[2];
-                pp[2] = p0[2 * 2];
-                pp[3] = p0[3 * 2];
-                pp[4] = p0[4 * 2];
-                pp[5] = p0[5 * 2];
-                pp[6] = p0[6 * 2];
-                pp[7] = p0[7 * 2];
-                pp[8] = p0[8 * 2];
-                pp[9] = p0[9 * 2];
-                pp[10] = p0[10 * 2];
-                pp[11] = p0[11 * 2];
-                pp[12] = p0[1];
-                pp[13] = p0[2 + 1];
-                pp[14] = p0[2 * 2 + 1];
-                pp[15] = p0[3 * 2 + 1];
-                pp[16] = p0[4 * 2 + 1];
-                pp[17] = p0[5 * 2 + 1];
-                pp[18] = p0[6 * 2 + 1];
-                pp[19] = p0[7 * 2 + 1];
-                pp[20] = p0[8 * 2 + 1];
-                pp[21] = p0[9 * 2 + 1];
-                pp[22] = p0[10 * 2 + 1];
-                pp[23] = p0[11 * 2 + 1];
-                p0 += max_jj * batch * 2;
-                pp += 24;
-            }
-            p0 -= (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                pp[2] = p0[2];
-                pp[3] = p0[3];
-                pp[4] = p0[4];
-                pp[5] = p0[5];
-                pp[6] = p0[6];
-                pp[7] = p0[7];
-                pp[8] = p0[8];
-                pp[9] = p0[9];
-                pp[10] = p0[10];
-                pp[11] = p0[11];
-                p0 += max_jj * batch;
-                pp += 12;
-            }
-        }
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            const float* p0 = B;
-
-            int kk = 0;
-#if __AVX__
-#if __AVX512F__
-            p0 += (b * max_jj + jj) * 16;
-            for (; kk + 15 < max_kk; kk += 16)
-            {
-                __m512 _r0 = _mm512_load_ps(p0);
-                __m512 _r1 = _mm512_load_ps(p0 + 16);
-                __m512 _r2 = _mm512_load_ps(p0 + 16 * 2);
-                __m512 _r3 = _mm512_load_ps(p0 + 16 * 3);
-                __m512 _r4 = _mm512_load_ps(p0 + 16 * 4);
-                __m512 _r5 = _mm512_load_ps(p0 + 16 * 5);
-                __m512 _r6 = _mm512_load_ps(p0 + 16 * 6);
-                __m512 _r7 = _mm512_load_ps(p0 + 16 * 7);
-                transpose16x8_ps(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-                _mm512_storeu_ps(pp, _r0);
-                _mm512_storeu_ps(pp + 16, _r1);
-                _mm512_storeu_ps(pp + 16 * 2, _r2);
-                _mm512_storeu_ps(pp + 16 * 3, _r3);
-                _mm512_storeu_ps(pp + 16 * 4, _r4);
-                _mm512_storeu_ps(pp + 16 * 5, _r5);
-                _mm512_storeu_ps(pp + 16 * 6, _r6);
-                _mm512_storeu_ps(pp + 16 * 7, _r7);
-                p0 += max_jj * batch * 16;
-                pp += 128;
-            }
-            p0 -= (b * max_jj + jj) * 16;
-#endif // __AVX512F__
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                __m256 _r0 = _mm256_load_ps(p0);
-                __m256 _r1 = _mm256_load_ps(p0 + 8);
-                __m256 _r2 = _mm256_load_ps(p0 + 8 * 2);
-                __m256 _r3 = _mm256_load_ps(p0 + 8 * 3);
-                __m256 _r4 = _mm256_load_ps(p0 + 8 * 4);
-                __m256 _r5 = _mm256_load_ps(p0 + 8 * 5);
-                __m256 _r6 = _mm256_load_ps(p0 + 8 * 6);
-                __m256 _r7 = _mm256_load_ps(p0 + 8 * 7);
-                transpose8x8_ps(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-                _mm256_storeu_ps(pp, _r0);
-                _mm256_storeu_ps(pp + 8, _r1);
-                _mm256_storeu_ps(pp + 8 * 2, _r2);
-                _mm256_storeu_ps(pp + 8 * 3, _r3);
-                _mm256_storeu_ps(pp + 8 * 4, _r4);
-                _mm256_storeu_ps(pp + 8 * 5, _r5);
-                _mm256_storeu_ps(pp + 8 * 6, _r6);
-                _mm256_storeu_ps(pp + 8 * 7, _r7);
-                p0 += max_jj * batch * 8;
-                pp += 64;
-            }
-            p0 -= (b * max_jj + jj) * 8;
-#endif // __AVX__
-            p0 += (b * max_jj + jj) * 4;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                __m128 _r0 = _mm_load_ps(p0);
-                __m128 _r1 = _mm_load_ps(p0 + 4);
-                __m128 _r2 = _mm_load_ps(p0 + 4 * 2);
-                __m128 _r3 = _mm_load_ps(p0 + 4 * 3);
-                __m128 _r4 = _mm_load_ps(p0 + 4 * 4);
-                __m128 _r5 = _mm_load_ps(p0 + 4 * 5);
-                __m128 _r6 = _mm_load_ps(p0 + 4 * 6);
-                __m128 _r7 = _mm_load_ps(p0 + 4 * 7);
-                _MM_TRANSPOSE4_PS(_r0, _r1, _r2, _r3);
-                _MM_TRANSPOSE4_PS(_r4, _r5, _r6, _r7);
-                _mm_store_ps(pp, _r0);
-                _mm_store_ps(pp + 4, _r4);
-                _mm_store_ps(pp + 8, _r1);
-                _mm_store_ps(pp + 12, _r5);
-                _mm_store_ps(pp + 16, _r2);
-                _mm_store_ps(pp + 20, _r6);
-                _mm_store_ps(pp + 24, _r3);
-                _mm_store_ps(pp + 28, _r7);
-                p0 += max_jj * batch * 4;
-                pp += 32;
-            }
-            p0 -= (b * max_jj + jj) * 4;
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[2];
-                pp[2] = p0[4];
-                pp[3] = p0[6];
-                pp[4] = p0[8];
-                pp[5] = p0[10];
-                pp[6] = p0[12];
-                pp[7] = p0[14];
-                pp[8] = p0[1];
-                pp[9] = p0[3];
-                pp[10] = p0[5];
-                pp[11] = p0[7];
-                pp[12] = p0[9];
-                pp[13] = p0[11];
-                pp[14] = p0[13];
-                pp[15] = p0[15];
-                p0 += max_jj * batch * 2;
-                pp += 16;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                pp[2] = p0[2];
-                pp[3] = p0[3];
-                pp[4] = p0[4];
-                pp[5] = p0[5];
-                pp[6] = p0[6];
-                pp[7] = p0[7];
-                p0 += max_jj * batch;
-                pp += 8;
-            }
-        }
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            const float* p0 = B;
-
-            int kk = 0;
-#if __AVX__
-#if __AVX512F__
-            p0 += (b * max_jj + jj) * 16;
-            for (; kk + 15 < max_kk; kk += 16)
-            {
-                __m512 _r0 = _mm512_load_ps(p0);
-                __m512 _r1 = _mm512_load_ps(p0 + 16);
-                __m512 _r2 = _mm512_load_ps(p0 + 16 * 2);
-                __m512 _r3 = _mm512_load_ps(p0 + 16 * 3);
-                transpose16x4_ps(_r0, _r1, _r2, _r3);
-                _mm512_storeu_ps(pp, _r0);
-                _mm512_storeu_ps(pp + 16, _r1);
-                _mm512_storeu_ps(pp + 32, _r2);
-                _mm512_storeu_ps(pp + 48, _r3);
-                p0 += max_jj * batch * 16;
-                pp += 64;
-            }
-            p0 -= (b * max_jj + jj) * 16;
-#endif // __AVX512F__
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                __m256 _r0 = _mm256_load_ps(p0);
-                __m256 _r1 = _mm256_load_ps(p0 + 8);
-                __m256 _r2 = _mm256_load_ps(p0 + 8 * 2);
-                __m256 _r3 = _mm256_load_ps(p0 + 8 * 3);
-                transpose8x4_ps(_r0, _r1, _r2, _r3);
-                _mm256_storeu_ps(pp, _r0);
-                _mm256_storeu_ps(pp + 8, _r1);
-                _mm256_storeu_ps(pp + 8 * 2, _r2);
-                _mm256_storeu_ps(pp + 8 * 3, _r3);
-                p0 += max_jj * batch * 8;
-                pp += 32;
-            }
-            p0 -= (b * max_jj + jj) * 8;
-#endif // __AVX__
-            p0 += (b * max_jj + jj) * 4;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                __m128 _r0 = _mm_load_ps(p0);
-                __m128 _r1 = _mm_load_ps(p0 + 4);
-                __m128 _r2 = _mm_load_ps(p0 + 4 * 2);
-                __m128 _r3 = _mm_load_ps(p0 + 4 * 3);
-                _MM_TRANSPOSE4_PS(_r0, _r1, _r2, _r3);
-                _mm_store_ps(pp, _r0);
-                _mm_store_ps(pp + 4, _r1);
-                _mm_store_ps(pp + 8, _r2);
-                _mm_store_ps(pp + 12, _r3);
-                p0 += max_jj * batch * 4;
-                pp += 16;
-            }
-            p0 -= (b * max_jj + jj) * 4;
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[2];
-                pp[2] = p0[4];
-                pp[3] = p0[6];
-                pp[4] = p0[1];
-                pp[5] = p0[3];
-                pp[6] = p0[5];
-                pp[7] = p0[7];
-                p0 += max_jj * batch * 2;
-                pp += 8;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                pp[2] = p0[2];
-                pp[3] = p0[3];
-                p0 += max_jj * batch;
-                pp += 4;
-            }
-        }
-#endif // __SSE2__
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            const float* p0 = B;
-
-            int kk = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-            p0 += (b * max_jj + jj) * 16;
-            for (; kk + 15 < max_kk; kk += 16)
-            {
-                __m512 _r0 = _mm512_load_ps(p0);
-                __m512 _r1 = _mm512_load_ps(p0 + 16);
-                transpose16x2_ps(_r0, _r1);
-                _mm512_storeu_ps(pp, _r0);
-                _mm512_storeu_ps(pp + 16, _r1);
-                p0 += max_jj * batch * 16;
-                pp += 32;
-            }
-            p0 -= (b * max_jj + jj) * 16;
-#endif // __AVX512F__
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                __m256 _r0 = _mm256_load_ps(p0);
-                __m256 _r1 = _mm256_load_ps(p0 + 8);
-                transpose8x2_ps(_r0, _r1);
-                _mm256_storeu_ps(pp, _r0);
-                _mm256_storeu_ps(pp + 8, _r1);
-                p0 += max_jj * batch * 8;
-                pp += 16;
-            }
-            p0 -= (b * max_jj + jj) * 8;
-#endif // __AVX__
-            p0 += (b * max_jj + jj) * 4;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                __m128 _r0 = _mm_load_ps(p0);
-                __m128 _r1 = _mm_load_ps(p0 + 4);
-                __m128 _tmp0 = _mm_unpacklo_ps(_r0, _r1);
-                __m128 _tmp1 = _mm_unpackhi_ps(_r0, _r1);
-                _mm_storeu_ps(pp, _tmp0);
-                _mm_storeu_ps(pp + 4, _tmp1);
-                p0 += max_jj * batch * 4;
-                pp += 8;
-            }
-            p0 -= (b * max_jj + jj) * 4;
-#endif // __SSE2__
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[2];
-                pp[2] = p0[1];
-                pp[3] = p0[3];
-                p0 += max_jj * batch * 2;
-                pp += 4;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                p0 += max_jj * batch;
-                pp += 2;
-            }
-        }
-        for (; jj < max_jj; jj++)
-        {
-            const float* p0 = B;
-
-            int kk = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-            p0 += (b * max_jj + jj) * 16;
-            for (; kk + 15 < max_kk; kk += 16)
-            {
-                __m512 _r0 = _mm512_load_ps(p0);
-                _mm512_storeu_ps(pp, _r0);
-                p0 += max_jj * batch * 16;
-                pp += 16;
-            }
-            p0 -= (b * max_jj + jj) * 16;
-#endif // __AVX512F__
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                __m256 _r0 = _mm256_load_ps(p0);
-                _mm256_storeu_ps(pp, _r0);
-                p0 += max_jj * batch * 8;
-                pp += 8;
-            }
-            p0 -= (b * max_jj + jj) * 8;
-#endif // __AVX__
-            p0 += (b * max_jj + jj) * 4;
-            for (; kk + 3 < max_kk; kk += 4)
-            {
-                __m128 _r0 = _mm_load_ps(p0);
-                _mm_storeu_ps(pp, _r0);
-                p0 += max_jj * batch * 4;
-                pp += 4;
-            }
-            p0 -= (b * max_jj + jj) * 4;
-#endif // __SSE2__
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                p0 += max_jj * batch * 2;
-                pp += 2;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                p0 += max_jj * batch;
-                pp += 1;
-            }
-        }
-    }
-}
-
-static void gemm_transB_packed_tile(const Mat& AT_tile, const Mat& BT_tile, Mat& top_blob, int batch, int max_ii, int max_jj, int k, int max_kk)
-{
-    float* outptr = top_blob;
-
-    int ii = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    for (; ii + 15 < max_ii; ii += 16)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const float* pAT = AT_tile.row(b) + max_kk * ii;
-            const float* pB = BT_tile.row(b);
-
-            int jj = 0;
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const float* pA = pAT;
-
-                __m512 _sum0;
-                __m512 _sum1;
-                __m512 _sum2;
-                __m512 _sum3;
-                __m512 _sum4;
-                __m512 _sum5;
-                __m512 _sum6;
-                __m512 _sum7;
-                __m512 _sum8;
-                __m512 _sum9;
-                __m512 _suma;
-                __m512 _sumb;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm512_setzero_ps();
-                    _sum1 = _mm512_setzero_ps();
-                    _sum2 = _mm512_setzero_ps();
-                    _sum3 = _mm512_setzero_ps();
-                    _sum4 = _mm512_setzero_ps();
-                    _sum5 = _mm512_setzero_ps();
-                    _sum6 = _mm512_setzero_ps();
-                    _sum7 = _mm512_setzero_ps();
-                    _sum8 = _mm512_setzero_ps();
-                    _sum9 = _mm512_setzero_ps();
-                    _suma = _mm512_setzero_ps();
-                    _sumb = _mm512_setzero_ps();
-                }
-                else
-                {
-                    _sum0 = _mm512_load_ps(outptr);
-                    _sum1 = _mm512_load_ps(outptr + 16);
-                    _sum2 = _mm512_load_ps(outptr + 16 * 2);
-                    _sum3 = _mm512_load_ps(outptr + 16 * 3);
-                    _sum4 = _mm512_load_ps(outptr + 16 * 4);
-                    _sum5 = _mm512_load_ps(outptr + 16 * 5);
-                    _sum6 = _mm512_load_ps(outptr + 16 * 6);
-                    _sum7 = _mm512_load_ps(outptr + 16 * 7);
-                    _sum8 = _mm512_load_ps(outptr + 16 * 8);
-                    _sum9 = _mm512_load_ps(outptr + 16 * 9);
-                    _suma = _mm512_load_ps(outptr + 16 * 10);
-                    _sumb = _mm512_load_ps(outptr + 16 * 11);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m512 _pA = _mm512_load_ps(pA);
-                    _sum0 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[0]), _sum0);
-                    _sum1 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[1]), _sum1);
-                    _sum2 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[2]), _sum2);
-                    _sum3 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[3]), _sum3);
-                    _sum4 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[4]), _sum4);
-                    _sum5 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[5]), _sum5);
-                    _sum6 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[6]), _sum6);
-                    _sum7 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[7]), _sum7);
-                    _sum8 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[8]), _sum8);
-                    _sum9 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[9]), _sum9);
-                    _suma = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[10]), _suma);
-                    _sumb = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[11]), _sumb);
-
-                    pA += 16;
-                    pB += 12;
-                }
-
-                _mm512_store_ps(outptr, _sum0);
-                _mm512_store_ps(outptr + 16, _sum1);
-                _mm512_store_ps(outptr + 16 * 2, _sum2);
-                _mm512_store_ps(outptr + 16 * 3, _sum3);
-                _mm512_store_ps(outptr + 16 * 4, _sum4);
-                _mm512_store_ps(outptr + 16 * 5, _sum5);
-                _mm512_store_ps(outptr + 16 * 6, _sum6);
-                _mm512_store_ps(outptr + 16 * 7, _sum7);
-                _mm512_store_ps(outptr + 16 * 8, _sum8);
-                _mm512_store_ps(outptr + 16 * 9, _sum9);
-                _mm512_store_ps(outptr + 16 * 10, _suma);
-                _mm512_store_ps(outptr + 16 * 11, _sumb);
-                outptr += 16 * 12;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const float* pA = pAT;
-
-                __m512 _sum0;
-                __m512 _sum1;
-                __m512 _sum2;
-                __m512 _sum3;
-                __m512 _sum4;
-                __m512 _sum5;
-                __m512 _sum6;
-                __m512 _sum7;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm512_setzero_ps();
-                    _sum1 = _mm512_setzero_ps();
-                    _sum2 = _mm512_setzero_ps();
-                    _sum3 = _mm512_setzero_ps();
-                    _sum4 = _mm512_setzero_ps();
-                    _sum5 = _mm512_setzero_ps();
-                    _sum6 = _mm512_setzero_ps();
-                    _sum7 = _mm512_setzero_ps();
-                }
-                else
-                {
-                    _sum0 = _mm512_load_ps(outptr);
-                    _sum1 = _mm512_load_ps(outptr + 16);
-                    _sum2 = _mm512_load_ps(outptr + 16 * 2);
-                    _sum3 = _mm512_load_ps(outptr + 16 * 3);
-                    _sum4 = _mm512_load_ps(outptr + 16 * 4);
-                    _sum5 = _mm512_load_ps(outptr + 16 * 5);
-                    _sum6 = _mm512_load_ps(outptr + 16 * 6);
-                    _sum7 = _mm512_load_ps(outptr + 16 * 7);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m512 _pA = _mm512_load_ps(pA);
-                    _sum0 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[0]), _sum0);
-                    _sum1 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[1]), _sum1);
-                    _sum2 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[2]), _sum2);
-                    _sum3 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[3]), _sum3);
-                    _sum4 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[4]), _sum4);
-                    _sum5 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[5]), _sum5);
-                    _sum6 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[6]), _sum6);
-                    _sum7 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[7]), _sum7);
-
-                    pA += 16;
-                    pB += 8;
-                }
-
-                _mm512_store_ps(outptr, _sum0);
-                _mm512_store_ps(outptr + 16, _sum1);
-                _mm512_store_ps(outptr + 16 * 2, _sum2);
-                _mm512_store_ps(outptr + 16 * 3, _sum3);
-                _mm512_store_ps(outptr + 16 * 4, _sum4);
-                _mm512_store_ps(outptr + 16 * 5, _sum5);
-                _mm512_store_ps(outptr + 16 * 6, _sum6);
-                _mm512_store_ps(outptr + 16 * 7, _sum7);
-                outptr += 16 * 8;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const float* pA = pAT;
-
-                __m512 _sum0;
-                __m512 _sum1;
-                __m512 _sum2;
-                __m512 _sum3;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm512_setzero_ps();
-                    _sum1 = _mm512_setzero_ps();
-                    _sum2 = _mm512_setzero_ps();
-                    _sum3 = _mm512_setzero_ps();
-                }
-                else
-                {
-                    _sum0 = _mm512_load_ps(outptr);
-                    _sum1 = _mm512_load_ps(outptr + 16);
-                    _sum2 = _mm512_load_ps(outptr + 16 * 2);
-                    _sum3 = _mm512_load_ps(outptr + 16 * 3);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m512 _pA = _mm512_load_ps(pA);
-                    _sum0 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[0]), _sum0);
-                    _sum1 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[1]), _sum1);
-                    _sum2 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[2]), _sum2);
-                    _sum3 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[3]), _sum3);
-
-                    pA += 16;
-                    pB += 4;
-                }
-
-                _mm512_store_ps(outptr, _sum0);
-                _mm512_store_ps(outptr + 16, _sum1);
-                _mm512_store_ps(outptr + 16 * 2, _sum2);
-                _mm512_store_ps(outptr + 16 * 3, _sum3);
-                outptr += 16 * 4;
-            }
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const float* pA = pAT;
-
-                __m512 _sum0;
-                __m512 _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm512_setzero_ps();
-                    _sum1 = _mm512_setzero_ps();
-                }
-                else
-                {
-                    _sum0 = _mm512_load_ps(outptr);
-                    _sum1 = _mm512_load_ps(outptr + 16);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m512 _pA = _mm512_load_ps(pA);
-                    _sum0 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[0]), _sum0);
-                    _sum1 = _mm512_fmadd_ps(_pA, _mm512_set1_ps(pB[1]), _sum1);
-
-                    pA += 16;
-                    pB += 2;
-                }
-
-                _mm512_store_ps(outptr, _sum0);
-                _mm512_store_ps(outptr + 16, _sum1);
-                outptr += 16 * 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const float* pA = pAT;
-
-                __m512 _sum;
-
-                if (k == 0)
-                {
-                    _sum = _mm512_setzero_ps();
-                }
-                else
-                {
-                    _sum = _mm512_load_ps(outptr);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m512 _pA = _mm512_load_ps(pA);
-                    __m512 _pB = _mm512_set1_ps(pB[0]);
-                    _sum = _mm512_fmadd_ps(_pA, _pB, _sum);
-
-                    pA += 16;
-                    pB += 1;
-                }
-
-                _mm512_store_ps(outptr, _sum);
-                outptr += 16;
-            }
-        }
-    }
-#endif // __AVX512F__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const float* pAT = AT_tile.row(b) + max_kk * ii;
-            const float* pB = BT_tile.row(b);
-
-            int jj = 0;
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const float* pA = pAT;
-
-                __m256 _sum0;
-                __m256 _sum1;
-                __m256 _sum2;
-                __m256 _sum3;
-                __m256 _sum4;
-                __m256 _sum5;
-                __m256 _sum6;
-                __m256 _sum7;
-                __m256 _sum8;
-                __m256 _sum9;
-                __m256 _suma;
-                __m256 _sumb;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm256_setzero_ps();
-                    _sum1 = _mm256_setzero_ps();
-                    _sum2 = _mm256_setzero_ps();
-                    _sum3 = _mm256_setzero_ps();
-                    _sum4 = _mm256_setzero_ps();
-                    _sum5 = _mm256_setzero_ps();
-                    _sum6 = _mm256_setzero_ps();
-                    _sum7 = _mm256_setzero_ps();
-                    _sum8 = _mm256_setzero_ps();
-                    _sum9 = _mm256_setzero_ps();
-                    _suma = _mm256_setzero_ps();
-                    _sumb = _mm256_setzero_ps();
-                }
-                else
-                {
-                    _sum0 = _mm256_load_ps(outptr);
-                    _sum1 = _mm256_load_ps(outptr + 8);
-                    _sum2 = _mm256_load_ps(outptr + 16);
-                    _sum3 = _mm256_load_ps(outptr + 24);
-                    _sum4 = _mm256_load_ps(outptr + 32);
-                    _sum5 = _mm256_load_ps(outptr + 40);
-                    _sum6 = _mm256_load_ps(outptr + 48);
-                    _sum7 = _mm256_load_ps(outptr + 56);
-                    _sum8 = _mm256_load_ps(outptr + 64);
-                    _sum9 = _mm256_load_ps(outptr + 72);
-                    _suma = _mm256_load_ps(outptr + 80);
-                    _sumb = _mm256_load_ps(outptr + 88);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m256 _pA = _mm256_load_ps(pA);
-                    _sum0 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[0]), _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[1]), _sum1);
-                    _sum2 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[2]), _sum2);
-                    _sum3 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[3]), _sum3);
-                    _sum4 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[4]), _sum4);
-                    _sum5 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[5]), _sum5);
-                    _sum6 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[6]), _sum6);
-                    _sum7 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[7]), _sum7);
-                    _sum8 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[8]), _sum8);
-                    _sum9 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[9]), _sum9);
-                    _suma = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[10]), _suma);
-                    _sumb = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[11]), _sumb);
-
-                    pA += 8;
-                    pB += 12;
-                }
-
-                _mm256_store_ps(outptr, _sum0);
-                _mm256_store_ps(outptr + 8, _sum1);
-                _mm256_store_ps(outptr + 8 * 2, _sum2);
-                _mm256_store_ps(outptr + 8 * 3, _sum3);
-                _mm256_store_ps(outptr + 8 * 4, _sum4);
-                _mm256_store_ps(outptr + 8 * 5, _sum5);
-                _mm256_store_ps(outptr + 8 * 6, _sum6);
-                _mm256_store_ps(outptr + 8 * 7, _sum7);
-                _mm256_store_ps(outptr + 8 * 8, _sum8);
-                _mm256_store_ps(outptr + 8 * 9, _sum9);
-                _mm256_store_ps(outptr + 8 * 10, _suma);
-                _mm256_store_ps(outptr + 8 * 11, _sumb);
-                outptr += 8 * 12;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const float* pA = pAT;
-
-                __m256 _sum0;
-                __m256 _sum1;
-                __m256 _sum2;
-                __m256 _sum3;
-                __m256 _sum4;
-                __m256 _sum5;
-                __m256 _sum6;
-                __m256 _sum7;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm256_setzero_ps();
-                    _sum1 = _mm256_setzero_ps();
-                    _sum2 = _mm256_setzero_ps();
-                    _sum3 = _mm256_setzero_ps();
-                    _sum4 = _mm256_setzero_ps();
-                    _sum5 = _mm256_setzero_ps();
-                    _sum6 = _mm256_setzero_ps();
-                    _sum7 = _mm256_setzero_ps();
-                }
-                else
-                {
-                    _sum0 = _mm256_load_ps(outptr);
-                    _sum1 = _mm256_load_ps(outptr + 8);
-                    _sum2 = _mm256_load_ps(outptr + 16);
-                    _sum3 = _mm256_load_ps(outptr + 24);
-                    _sum4 = _mm256_load_ps(outptr + 32);
-                    _sum5 = _mm256_load_ps(outptr + 40);
-                    _sum6 = _mm256_load_ps(outptr + 48);
-                    _sum7 = _mm256_load_ps(outptr + 56);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m256 _pA = _mm256_load_ps(pA);
-                    _sum0 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[0]), _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[1]), _sum1);
-                    _sum2 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[2]), _sum2);
-                    _sum3 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[3]), _sum3);
-                    _sum4 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[4]), _sum4);
-                    _sum5 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[5]), _sum5);
-                    _sum6 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[6]), _sum6);
-                    _sum7 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[7]), _sum7);
-
-                    pA += 8;
-                    pB += 8;
-                }
-
-                _mm256_store_ps(outptr, _sum0);
-                _mm256_store_ps(outptr + 8, _sum1);
-                _mm256_store_ps(outptr + 8 * 2, _sum2);
-                _mm256_store_ps(outptr + 8 * 3, _sum3);
-                _mm256_store_ps(outptr + 8 * 4, _sum4);
-                _mm256_store_ps(outptr + 8 * 5, _sum5);
-                _mm256_store_ps(outptr + 8 * 6, _sum6);
-                _mm256_store_ps(outptr + 8 * 7, _sum7);
-                outptr += 8 * 8;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const float* pA = pAT;
-
-                __m256 _sum0;
-                __m256 _sum1;
-                __m256 _sum2;
-                __m256 _sum3;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm256_setzero_ps();
-                    _sum1 = _mm256_setzero_ps();
-                    _sum2 = _mm256_setzero_ps();
-                    _sum3 = _mm256_setzero_ps();
-                }
-                else
-                {
-                    _sum0 = _mm256_load_ps(outptr);
-                    _sum1 = _mm256_load_ps(outptr + 8);
-                    _sum2 = _mm256_load_ps(outptr + 16);
-                    _sum3 = _mm256_load_ps(outptr + 24);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m256 _pA = _mm256_load_ps(pA);
-                    _sum0 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[0]), _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[1]), _sum1);
-                    _sum2 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[2]), _sum2);
-                    _sum3 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[3]), _sum3);
-
-                    pA += 8;
-                    pB += 4;
-                }
-
-                _mm256_store_ps(outptr, _sum0);
-                _mm256_store_ps(outptr + 8, _sum1);
-                _mm256_store_ps(outptr + 8 * 2, _sum2);
-                _mm256_store_ps(outptr + 8 * 3, _sum3);
-                outptr += 8 * 4;
-            }
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const float* pA = pAT;
-
-                __m256 _sum0;
-                __m256 _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm256_setzero_ps();
-                    _sum1 = _mm256_setzero_ps();
-                }
-                else
-                {
-                    _sum0 = _mm256_load_ps(outptr);
-                    _sum1 = _mm256_load_ps(outptr + 8);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m256 _pA = _mm256_load_ps(pA);
-                    _sum0 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[0]), _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_pA, _mm256_set1_ps(pB[1]), _sum1);
-
-                    pA += 8;
-                    pB += 2;
-                }
-
-                _mm256_store_ps(outptr, _sum0);
-                _mm256_store_ps(outptr + 8, _sum1);
-                outptr += 8 * 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const float* pA = pAT;
-
-                __m256 _sum;
-
-                if (k == 0)
-                {
-                    _sum = _mm256_setzero_ps();
-                }
-                else
-                {
-                    _sum = _mm256_load_ps(outptr);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m256 _pA = _mm256_load_ps(pA);
-                    __m256 _pB = _mm256_set1_ps(pB[0]);
-                    _sum = _mm256_comp_fmadd_ps(_pA, _pB, _sum);
-
-                    pA += 8;
-                    pB += 1;
-                }
-
-                _mm256_store_ps(outptr, _sum);
-                outptr += 8;
-            }
-        }
-    }
-#endif // __AVX__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const float* pAT = AT_tile.row(b) + max_kk * ii;
-            const float* pB = BT_tile.row(b);
-
-            int jj = 0;
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const float* pA = pAT;
-
-                __m128 _sum0;
-                __m128 _sum1;
-                __m128 _sum2;
-                __m128 _sum3;
-                __m128 _sum4;
-                __m128 _sum5;
-                __m128 _sum6;
-                __m128 _sum7;
-                __m128 _sum8;
-                __m128 _sum9;
-                __m128 _suma;
-                __m128 _sumb;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm_setzero_ps();
-                    _sum1 = _mm_setzero_ps();
-                    _sum2 = _mm_setzero_ps();
-                    _sum3 = _mm_setzero_ps();
-                    _sum4 = _mm_setzero_ps();
-                    _sum5 = _mm_setzero_ps();
-                    _sum6 = _mm_setzero_ps();
-                    _sum7 = _mm_setzero_ps();
-                    _sum8 = _mm_setzero_ps();
-                    _sum9 = _mm_setzero_ps();
-                    _suma = _mm_setzero_ps();
-                    _sumb = _mm_setzero_ps();
-                }
-                else
-                {
-                    _sum0 = _mm_load_ps(outptr);
-                    _sum1 = _mm_load_ps(outptr + 4);
-                    _sum2 = _mm_load_ps(outptr + 8);
-                    _sum3 = _mm_load_ps(outptr + 12);
-                    _sum4 = _mm_load_ps(outptr + 16);
-                    _sum5 = _mm_load_ps(outptr + 20);
-                    _sum6 = _mm_load_ps(outptr + 24);
-                    _sum7 = _mm_load_ps(outptr + 28);
-                    _sum8 = _mm_load_ps(outptr + 32);
-                    _sum9 = _mm_load_ps(outptr + 36);
-                    _suma = _mm_load_ps(outptr + 40);
-                    _sumb = _mm_load_ps(outptr + 44);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m128 _pA = _mm_load_ps(pA);
-                    _sum0 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[0]), _sum0);
-                    _sum1 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[1]), _sum1);
-                    _sum2 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[2]), _sum2);
-                    _sum3 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[3]), _sum3);
-                    _sum4 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[4]), _sum4);
-                    _sum5 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[5]), _sum5);
-                    _sum6 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[6]), _sum6);
-                    _sum7 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[7]), _sum7);
-                    _sum8 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[8]), _sum8);
-                    _sum9 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[9]), _sum9);
-                    _suma = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[10]), _suma);
-                    _sumb = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[11]), _sumb);
-
-                    pA += 4;
-                    pB += 12;
-                }
-
-                _mm_store_ps(outptr, _sum0);
-                _mm_store_ps(outptr + 4, _sum1);
-                _mm_store_ps(outptr + 4 * 2, _sum2);
-                _mm_store_ps(outptr + 4 * 3, _sum3);
-                _mm_store_ps(outptr + 4 * 4, _sum4);
-                _mm_store_ps(outptr + 4 * 5, _sum5);
-                _mm_store_ps(outptr + 4 * 6, _sum6);
-                _mm_store_ps(outptr + 4 * 7, _sum7);
-                _mm_store_ps(outptr + 4 * 8, _sum8);
-                _mm_store_ps(outptr + 4 * 9, _sum9);
-                _mm_store_ps(outptr + 4 * 10, _suma);
-                _mm_store_ps(outptr + 4 * 11, _sumb);
-                outptr += 4 * 12;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const float* pA = pAT;
-
-                __m128 _sum0;
-                __m128 _sum1;
-                __m128 _sum2;
-                __m128 _sum3;
-                __m128 _sum4;
-                __m128 _sum5;
-                __m128 _sum6;
-                __m128 _sum7;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm_setzero_ps();
-                    _sum1 = _mm_setzero_ps();
-                    _sum2 = _mm_setzero_ps();
-                    _sum3 = _mm_setzero_ps();
-                    _sum4 = _mm_setzero_ps();
-                    _sum5 = _mm_setzero_ps();
-                    _sum6 = _mm_setzero_ps();
-                    _sum7 = _mm_setzero_ps();
-                }
-                else
-                {
-                    _sum0 = _mm_load_ps(outptr);
-                    _sum1 = _mm_load_ps(outptr + 4);
-                    _sum2 = _mm_load_ps(outptr + 8);
-                    _sum3 = _mm_load_ps(outptr + 12);
-                    _sum4 = _mm_load_ps(outptr + 16);
-                    _sum5 = _mm_load_ps(outptr + 20);
-                    _sum6 = _mm_load_ps(outptr + 24);
-                    _sum7 = _mm_load_ps(outptr + 28);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m128 _pA = _mm_load_ps(pA);
-                    _sum0 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[0]), _sum0);
-                    _sum1 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[1]), _sum1);
-                    _sum2 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[2]), _sum2);
-                    _sum3 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[3]), _sum3);
-                    _sum4 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[4]), _sum4);
-                    _sum5 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[5]), _sum5);
-                    _sum6 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[6]), _sum6);
-                    _sum7 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[7]), _sum7);
-
-                    pA += 4;
-                    pB += 8;
-                }
-
-                _mm_store_ps(outptr, _sum0);
-                _mm_store_ps(outptr + 4, _sum1);
-                _mm_store_ps(outptr + 4 * 2, _sum2);
-                _mm_store_ps(outptr + 4 * 3, _sum3);
-                _mm_store_ps(outptr + 4 * 4, _sum4);
-                _mm_store_ps(outptr + 4 * 5, _sum5);
-                _mm_store_ps(outptr + 4 * 6, _sum6);
-                _mm_store_ps(outptr + 4 * 7, _sum7);
-                outptr += 4 * 8;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const float* pA = pAT;
-
-                __m128 _sum0;
-                __m128 _sum1;
-                __m128 _sum2;
-                __m128 _sum3;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm_setzero_ps();
-                    _sum1 = _mm_setzero_ps();
-                    _sum2 = _mm_setzero_ps();
-                    _sum3 = _mm_setzero_ps();
-                }
-                else
-                {
-                    _sum0 = _mm_load_ps(outptr);
-                    _sum1 = _mm_load_ps(outptr + 4);
-                    _sum2 = _mm_load_ps(outptr + 8);
-                    _sum3 = _mm_load_ps(outptr + 12);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m128 _pA = _mm_load_ps(pA);
-                    _sum0 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[0]), _sum0);
-                    _sum1 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[1]), _sum1);
-                    _sum2 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[2]), _sum2);
-                    _sum3 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[3]), _sum3);
-
-                    pA += 4;
-                    pB += 4;
-                }
-
-                _mm_store_ps(outptr, _sum0);
-                _mm_store_ps(outptr + 4, _sum1);
-                _mm_store_ps(outptr + 4 * 2, _sum2);
-                _mm_store_ps(outptr + 4 * 3, _sum3);
-                outptr += 4 * 4;
-            }
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const float* pA = pAT;
-
-                __m128 _sum0;
-                __m128 _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm_setzero_ps();
-                    _sum1 = _mm_setzero_ps();
-                }
-                else
-                {
-                    _sum0 = _mm_load_ps(outptr);
-                    _sum1 = _mm_load_ps(outptr + 4);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m128 _pA = _mm_load_ps(pA);
-                    _sum0 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[0]), _sum0);
-                    _sum1 = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[1]), _sum1);
-
-                    pA += 4;
-                    pB += 2;
-                }
-
-                _mm_store_ps(outptr, _sum0);
-                _mm_store_ps(outptr + 4, _sum1);
-                outptr += 4 * 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const float* pA = pAT;
-
-                __m128 _sum;
-
-                if (k == 0)
-                {
-                    _sum = _mm_setzero_ps();
-                }
-                else
-                {
-                    _sum = _mm_load_ps(outptr);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m128 _pA = _mm_load_ps(pA);
-                    _sum = _mm_comp_fmadd_ps(_pA, _mm_set1_ps(pB[0]), _sum);
-
-                    pA += 4;
-                    pB += 1;
-                }
-
-                _mm_store_ps(outptr, _sum);
-                outptr += 4;
-            }
-        }
-    }
-#endif // __SSE2__
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const float* pAT = AT_tile.row(b) + max_kk * ii;
-            const float* pB = BT_tile.row(b);
-
-            int jj = 0;
-#if __SSE2__
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const float* pA = pAT;
-
-                __m128 _sum0;
-                __m128 _sum1;
-                __m128 _sum2;
-                __m128 _sum3;
-                __m128 _sum4;
-                __m128 _sum5;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm_setzero_ps();
-                    _sum1 = _mm_setzero_ps();
-                    _sum2 = _mm_setzero_ps();
-                    _sum3 = _mm_setzero_ps();
-                    _sum4 = _mm_setzero_ps();
-                    _sum5 = _mm_setzero_ps();
-                }
-                else
-                {
-                    __m128 _tmp0 = _mm_loadu_ps(outptr);
-                    __m128 _tmp1 = _mm_loadu_ps(outptr + 4);
-                    __m128 _tmp2 = _mm_loadu_ps(outptr + 8);
-                    __m128 _tmp3 = _mm_loadu_ps(outptr + 12);
-                    __m128 _tmp4 = _mm_loadu_ps(outptr + 16);
-                    __m128 _tmp5 = _mm_loadu_ps(outptr + 20);
-                    _sum0 = _mm_shuffle_ps(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum1 = _mm_shuffle_ps(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum2 = _mm_shuffle_ps(_tmp4, _tmp5, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum3 = _mm_shuffle_ps(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum4 = _mm_shuffle_ps(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum5 = _mm_shuffle_ps(_tmp4, _tmp5, _MM_SHUFFLE(3, 1, 3, 1));
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m128 _pA0 = _mm_set1_ps(pA[0]);
-                    __m128 _pA1 = _mm_set1_ps(pA[1]);
-                    __m128 _pB0 = _mm_load_ps(pB);
-                    __m128 _pB1 = _mm_load_ps(pB + 4);
-                    __m128 _pB2 = _mm_load_ps(pB + 8);
-                    _sum0 = _mm_comp_fmadd_ps(_pA0, _pB0, _sum0);
-                    _sum1 = _mm_comp_fmadd_ps(_pA0, _pB1, _sum1);
-                    _sum2 = _mm_comp_fmadd_ps(_pA0, _pB2, _sum2);
-                    _sum3 = _mm_comp_fmadd_ps(_pA1, _pB0, _sum3);
-                    _sum4 = _mm_comp_fmadd_ps(_pA1, _pB1, _sum4);
-                    _sum5 = _mm_comp_fmadd_ps(_pA1, _pB2, _sum5);
-                    pA += 2;
-                    pB += 12;
-                }
-
-                __m128 _tmp0 = _mm_unpacklo_ps(_sum0, _sum3);
-                __m128 _tmp1 = _mm_unpackhi_ps(_sum0, _sum3);
-                __m128 _tmp2 = _mm_unpacklo_ps(_sum1, _sum4);
-                __m128 _tmp3 = _mm_unpackhi_ps(_sum1, _sum4);
-                __m128 _tmp4 = _mm_unpacklo_ps(_sum2, _sum5);
-                __m128 _tmp5 = _mm_unpackhi_ps(_sum2, _sum5);
-                _mm_storeu_ps(outptr, _tmp0);
-                _mm_storeu_ps(outptr + 4, _tmp1);
-                _mm_storeu_ps(outptr + 8, _tmp2);
-                _mm_storeu_ps(outptr + 12, _tmp3);
-                _mm_storeu_ps(outptr + 16, _tmp4);
-                _mm_storeu_ps(outptr + 20, _tmp5);
-                outptr += 2 * 12;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const float* pA = pAT;
-
-                __m128 _sum0;
-                __m128 _sum1;
-                __m128 _sum2;
-                __m128 _sum3;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm_setzero_ps();
-                    _sum1 = _mm_setzero_ps();
-                    _sum2 = _mm_setzero_ps();
-                    _sum3 = _mm_setzero_ps();
-                }
-                else
-                {
-                    __m128 _tmp0 = _mm_loadu_ps(outptr);
-                    __m128 _tmp1 = _mm_loadu_ps(outptr + 4);
-                    __m128 _tmp2 = _mm_loadu_ps(outptr + 8);
-                    __m128 _tmp3 = _mm_loadu_ps(outptr + 12);
-                    _sum0 = _mm_shuffle_ps(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum1 = _mm_shuffle_ps(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum2 = _mm_shuffle_ps(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum3 = _mm_shuffle_ps(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m128 _pA0 = _mm_set1_ps(pA[0]);
-                    __m128 _pA1 = _mm_set1_ps(pA[1]);
-                    __m128 _pB0 = _mm_load_ps(pB);
-                    __m128 _pB1 = _mm_load_ps(pB + 4);
-                    _sum0 = _mm_comp_fmadd_ps(_pA0, _pB0, _sum0);
-                    _sum1 = _mm_comp_fmadd_ps(_pA0, _pB1, _sum1);
-                    _sum2 = _mm_comp_fmadd_ps(_pA1, _pB0, _sum2);
-                    _sum3 = _mm_comp_fmadd_ps(_pA1, _pB1, _sum3);
-                    pA += 2;
-                    pB += 8;
-                }
-
-                __m128 _tmp0 = _mm_unpacklo_ps(_sum0, _sum2);
-                __m128 _tmp1 = _mm_unpackhi_ps(_sum0, _sum2);
-                __m128 _tmp2 = _mm_unpacklo_ps(_sum1, _sum3);
-                __m128 _tmp3 = _mm_unpackhi_ps(_sum1, _sum3);
-                _mm_storeu_ps(outptr, _tmp0);
-                _mm_storeu_ps(outptr + 4, _tmp1);
-                _mm_storeu_ps(outptr + 8, _tmp2);
-                _mm_storeu_ps(outptr + 12, _tmp3);
-                outptr += 2 * 8;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const float* pA = pAT;
-
-                __m128 _sum0;
-                __m128 _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm_setzero_ps();
-                    _sum1 = _mm_setzero_ps();
-                }
-                else
-                {
-                    __m128 _tmp0 = _mm_loadu_ps(outptr);
-                    __m128 _tmp1 = _mm_loadu_ps(outptr + 4);
-                    _sum0 = _mm_shuffle_ps(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum1 = _mm_shuffle_ps(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m128 _pB = _mm_load_ps(pB);
-                    _sum0 = _mm_comp_fmadd_ps(_mm_set1_ps(pA[0]), _pB, _sum0);
-                    _sum1 = _mm_comp_fmadd_ps(_mm_set1_ps(pA[1]), _pB, _sum1);
-                    pA += 2;
-                    pB += 4;
-                }
-
-                __m128 _tmp0 = _mm_unpacklo_ps(_sum0, _sum1);
-                __m128 _tmp1 = _mm_unpackhi_ps(_sum0, _sum1);
-                _mm_storeu_ps(outptr, _tmp0);
-                _mm_storeu_ps(outptr + 4, _tmp1);
-                outptr += 2 * 4;
-            }
-#endif // __SSE2__
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const float* pA = pAT;
-
-                float sum00 = 0.f;
-                float sum01 = 0.f;
-                float sum10 = 0.f;
-                float sum11 = 0.f;
-
-                if (k == 0)
-                {
-                    sum00 = 0.f;
-                    sum01 = 0.f;
-                    sum10 = 0.f;
-                    sum11 = 0.f;
-                }
-                else
-                {
-                    sum00 = outptr[0];
-                    sum01 = outptr[1];
-                    sum10 = outptr[2];
-                    sum11 = outptr[3];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    sum00 += pA[0] * pB[0];
-                    sum01 += pA[1] * pB[0];
-                    sum10 += pA[0] * pB[1];
-                    sum11 += pA[1] * pB[1];
-                    pA += 2;
-                    pB += 2;
-                }
-
-                outptr[0] = sum00;
-                outptr[1] = sum01;
-                outptr[2] = sum10;
-                outptr[3] = sum11;
-                outptr += 2 * 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const float* pA = pAT;
-
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-
-                if (k == 0)
-                {
-                    sum0 = 0.f;
-                    sum1 = 0.f;
-                }
-                else
-                {
-                    sum0 = outptr[0];
-                    sum1 = outptr[1];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    sum0 += pA[0] * pB[0];
-                    sum1 += pA[1] * pB[0];
-                    pA += 2;
-                    pB += 1;
-                }
-
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-                outptr += 2;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const float* pAT = AT_tile.row(b) + max_kk * ii;
-            const float* pB = BT_tile.row(b);
-
-            int jj = 0;
-#if __SSE2__
-            for (; jj + 11 < max_jj; jj += 12)
-            {
-                const float* pA = pAT;
-
-                __m128 _sum0;
-                __m128 _sum1;
-                __m128 _sum2;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm_setzero_ps();
-                    _sum1 = _mm_setzero_ps();
-                    _sum2 = _mm_setzero_ps();
-                }
-                else
-                {
-                    _sum0 = _mm_loadu_ps(outptr);
-                    _sum1 = _mm_loadu_ps(outptr + 4);
-                    _sum2 = _mm_loadu_ps(outptr + 8);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m128 _pA = _mm_set1_ps(pA[0]);
-                    __m128 _pB0 = _mm_load_ps(pB);
-                    __m128 _pB1 = _mm_load_ps(pB + 4);
-                    __m128 _pB2 = _mm_load_ps(pB + 8);
-                    _sum0 = _mm_comp_fmadd_ps(_pA, _pB0, _sum0);
-                    _sum1 = _mm_comp_fmadd_ps(_pA, _pB1, _sum1);
-                    _sum2 = _mm_comp_fmadd_ps(_pA, _pB2, _sum2);
-                    pA += 1;
-                    pB += 12;
-                }
-
-                _mm_storeu_ps(outptr, _sum0);
-                _mm_storeu_ps(outptr + 4, _sum1);
-                _mm_storeu_ps(outptr + 8, _sum2);
-                outptr += 12;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const float* pA = pAT;
-
-                __m128 _sum0;
-                __m128 _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm_setzero_ps();
-                    _sum1 = _mm_setzero_ps();
-                }
-                else
-                {
-                    _sum0 = _mm_loadu_ps(outptr);
-                    _sum1 = _mm_loadu_ps(outptr + 4);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m128 _pA = _mm_set1_ps(pA[0]);
-                    __m128 _pB0 = _mm_load_ps(pB);
-                    __m128 _pB1 = _mm_load_ps(pB + 4);
-                    _sum0 = _mm_comp_fmadd_ps(_pA, _pB0, _sum0);
-                    _sum1 = _mm_comp_fmadd_ps(_pA, _pB1, _sum1);
-                    pA += 1;
-                    pB += 8;
-                }
-
-                _mm_storeu_ps(outptr, _sum0);
-                _mm_storeu_ps(outptr + 4, _sum1);
-                outptr += 8;
-            }
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const float* pA = pAT;
-
-                __m128 _sum;
-
-                if (k == 0)
-                {
-                    _sum = _mm_setzero_ps();
-                }
-                else
-                {
-                    _sum = _mm_loadu_ps(outptr);
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    __m128 _pA = _mm_set1_ps(pA[0]);
-                    __m128 _pB = _mm_load_ps(pB);
-                    _sum = _mm_comp_fmadd_ps(_pA, _pB, _sum);
-                    pA += 1;
-                    pB += 4;
-                }
-
-                _mm_storeu_ps(outptr, _sum);
-                outptr += 4;
-            }
-#endif // __SSE2__
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const float* pA = pAT;
-
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-
-                if (k == 0)
-                {
-                    sum0 = 0.f;
-                    sum1 = 0.f;
-                }
-                else
-                {
-                    sum0 = outptr[0];
-                    sum1 = outptr[1];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    sum0 += pA[0] * pB[0];
-                    sum1 += pA[0] * pB[1];
-                    pA += 1;
-                    pB += 2;
-                }
-
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-                outptr += 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const float* pA = pAT;
-
-                float sum = 0.f;
-
-                if (k == 0)
-                {
-                    sum = 0.f;
-                }
-                else
-                {
-                    sum = outptr[0];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    sum += pA[0] * pB[0];
-                    pA += 1;
-                    pB += 1;
-                }
-
-                outptr[0] = sum;
-                outptr += 1;
-            }
-        }
-    }
-}
-
-static void get_optimal_tile_mnk(int M, int N, int K, int& TILE_M, int& TILE_N, int& TILE_K, int nT)
-{
-    // resolve optimal tile size from cache size
-    const size_t l2_cache_size = get_cpu_level2_cache_size();
-
-    if (nT == 0)
-        nT = get_physical_big_cpu_count();
-
-    // solve M
-    {
-        int tile_size = (int)sqrt((float)l2_cache_size / sizeof(float) / 3);
-
-#if __AVX512F__
-        TILE_M = std::max(16, tile_size / 16 * 16);
-#elif __AVX__
-        TILE_M = std::max(8, tile_size / 8 * 8);
-#elif __SSE2__
-        TILE_M = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_M = std::max(2, tile_size / 2 * 2);
-#endif
-
-        TILE_M *= std::min(nT, get_physical_cpu_count());
-
-        int nn_M = (M + TILE_M - 1) / TILE_M;
-#if __AVX512F__
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 15) / 16 * 16);
-#elif __AVX__
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-#elif __SSE2__
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 3) / 4 * 4);
-#else
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 1) / 2 * 2);
-#endif
-
-        if (nT > 1)
-        {
-#if __AVX512F__
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 15) / 16 * 16);
-#elif __AVX__
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 7) / 8 * 8);
-#elif __SSE2__
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 3) / 4 * 4);
-#else
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 1) / 2 * 2);
-#endif
-        }
-    }
-
-    // solve K
-    {
-        int tile_size = (int)(sqrt((float)l2_cache_size / sizeof(float)) - TILE_M);
-
-#if __AVX512F__
-        TILE_K = std::max(16, tile_size / 16 * 16);
-#elif __AVX__
-        TILE_K = std::max(8, tile_size / 8 * 8);
-#elif __SSE2__
-        TILE_K = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_K = std::max(2, tile_size / 2 * 2);
-#endif
-
-        int nn_K = (K + TILE_K - 1) / TILE_K;
-#if __AVX512F__
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 15) / 16 * 16);
-#elif __AVX__
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 7) / 8 * 8);
-#elif __SSE2__
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 3) / 4 * 4);
-#else
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 1) / 2 * 2);
-#endif
-    }
-
-    if (N > 0)
-    {
-        int tile_size = (int)(((float)l2_cache_size / sizeof(float) - TILE_M * TILE_K) / (TILE_M + TILE_K));
-
-#if __AVX512F__
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#elif __AVX__
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#elif __SSE2__
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_N = std::max(1, tile_size);
-#endif
-
-        int nn_N = (N + TILE_N - 1) / TILE_N;
-#if __AVX512F__
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#elif __AVX__
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#elif __SSE2__
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#else
-        TILE_N = std::min(TILE_N, (N + nn_N - 1) / nn_N);
-#endif
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_kernel_tile(const Mat& kernel, Mat& A, int inch, int i, int max_ii, int k, int max_kk)
-{
-    // const float ktm[4][3] = {
-    //     {1.0f, 0.0f, 0.0f},
-    //     {1.0f / 2, 1.0f / 2, 1.0f / 2},
-    //     {1.0f / 2, -1.0f / 2, 1.0f / 2},
-    //     {0.0f, 0.0f, 1.0f}
-    // };
-
-    float* ptmp = A;
-
-    int ii = 0;
-    for (; ii < max_ii; ii++)
-    {
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            float tmp[4][3];
-
-            const float* k0 = (const float*)kernel + (i + ii) * inch * 9 + (k + kk) * 9;
-
-            for (int m = 0; m < 3; m++)
-            {
-                float r0 = k0[0];
-                float r1 = k0[1];
-                float r2 = k0[2];
-
-                tmp[0][m] = r0;
-                tmp[1][m] = r0 * 0.5f + r1 * 0.5f + r2 * 0.5f;
-                tmp[2][m] = r0 * 0.5f - r1 * 0.5f + r2 * 0.5f;
-                tmp[3][m] = r2;
-
-                k0 += 3;
-            }
-
-            for (int m = 0; m < 4; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-
-                float z0 = r0;
-                float z1 = r0 * 0.5f + r1 * 0.5f + r2 * 0.5f;
-                float z2 = r0 * 0.5f - r1 * 0.5f + r2 * 0.5f;
-                float z3 = r2;
-
-                ptmp[0] = z0;
-                ptmp[1] = z1;
-                ptmp[2] = z2;
-                ptmp[3] = z3;
-                ptmp += 4;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_kernel(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt)
-{
-    const int M = outch;
-    const int K = inch;
-    const int B = 16;
-
-    int TILE_M, TILE_N, TILE_K;
-    get_optimal_tile_mnk(M, 0, K, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    Mat A_tileX(B * TILE_M * TILE_K, 1, opt.num_threads, 4u, (Allocator*)0);
-
-    AT.create(TILE_K * TILE_M, B, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M, 4u, (Allocator*)0);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat A_tile = A_tileX.channel(get_omp_thread_num());
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_ii = std::min((M - i), TILE_M);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            conv3x3s1_winograd23_transform_kernel_tile(kernel, A_tile, inch, i, max_ii, k, max_kk);
-
-            Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-            pack_A_tile(A_tile, AT_tile, B, max_ii, max_kk);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_input_tile(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    // const float itm[4][4] = {
-    //     {1.0f,  0.0f, -1.0f,  0.0f},
-    //     {0.0f,  1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  1.00f, 0.0f},
-    //     {0.0f, -1.0f,  0.00f, 1.0f}
-    // };
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w - 1) / 2;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    nn_max_kk = max_kk / 16;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = ppkk * 16;
-
-#ifdef _MSC_VER
-        __declspec(align(64))
-#else
-        __attribute__((aligned(64)))
-#endif
-        float tmp[4][4][16];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel((k + kk) / elempack).row(ti * 2) + (tj * 2) * elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                __m512 _r0 = _mm512_setzero_ps();
-                __m512 _r1 = _mm512_setzero_ps();
-                __m512 _r2 = _mm512_setzero_ps();
-                __m512 _r3 = _mm512_setzero_ps();
-
-                if (ti * 2 + m < h)
-                {
-                    if (elempack == 16)
-                    {
-                        _r0 = _mm512_load_ps(r0);
-                        if (tj * 2 + 1 < w) _r1 = _mm512_load_ps(r0 + 16);
-                        if (tj * 2 + 2 < w) _r2 = _mm512_load_ps(r0 + 32);
-                        if (tj * 2 + 3 < w) _r3 = _mm512_load_ps(r0 + 48);
-                    }
-                    if (elempack == 8)
-                    {
-                        const float* r1 = r0 + N;
-
-                        _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0)), _mm256_load_ps(r1), 1);
-                        if (tj * 2 + 1 < w) _r1 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + 8)), _mm256_load_ps(r1 + 8), 1);
-                        if (tj * 2 + 2 < w) _r2 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + 16)), _mm256_load_ps(r1 + 16), 1);
-                        if (tj * 2 + 3 < w) _r3 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + 24)), _mm256_load_ps(r1 + 24), 1);
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-
-                        _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0)), _mm_load_ps(r1), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2)), _mm_load_ps(r3), 1), 1);
-                        if (tj * 2 + 1 < w) _r1 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 4)), _mm_load_ps(r1 + 4), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + 4)), _mm_load_ps(r3 + 4), 1), 1);
-                        if (tj * 2 + 2 < w) _r2 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 8)), _mm_load_ps(r1 + 8), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + 8)), _mm_load_ps(r3 + 8), 1), 1);
-                        if (tj * 2 + 3 < w) _r3 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 12)), _mm_load_ps(r1 + 12), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + 12)), _mm_load_ps(r3 + 12), 1), 1);
-                    }
-                    if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-                        const float* r4 = r0 + N * 4;
-                        const float* r5 = r0 + N * 5;
-                        const float* r6 = r0 + N * 6;
-                        const float* r7 = r0 + N * 7;
-                        const float* r8 = r0 + N * 8;
-                        const float* r9 = r0 + N * 9;
-                        const float* ra = r0 + N * 10;
-                        const float* rb = r0 + N * 11;
-                        const float* rc = r0 + N * 12;
-                        const float* rd = r0 + N * 13;
-                        const float* re = r0 + N * 14;
-                        const float* rf = r0 + N * 15;
-
-                        __m128 _t0 = _mm_loadu_ps(r0);
-                        __m128 _t1 = _mm_loadu_ps(r1);
-                        __m128 _t2 = _mm_loadu_ps(r2);
-                        __m128 _t3 = _mm_loadu_ps(r3);
-                        __m128 _t4 = _mm_loadu_ps(r4);
-                        __m128 _t5 = _mm_loadu_ps(r5);
-                        __m128 _t6 = _mm_loadu_ps(r6);
-                        __m128 _t7 = _mm_loadu_ps(r7);
-                        __m128 _t8 = _mm_loadu_ps(r8);
-                        __m128 _t9 = _mm_loadu_ps(r9);
-                        __m128 _ta = _mm_loadu_ps(ra);
-                        __m128 _tb = _mm_loadu_ps(rb);
-                        __m128 _tc = _mm_loadu_ps(rc);
-                        __m128 _td = _mm_loadu_ps(rd);
-                        __m128 _te = _mm_loadu_ps(re);
-                        __m128 _tf = _mm_loadu_ps(rf);
-
-                        _MM_TRANSPOSE4_PS(_t0, _t1, _t2, _t3);
-                        _MM_TRANSPOSE4_PS(_t4, _t5, _t6, _t7);
-                        _MM_TRANSPOSE4_PS(_t8, _t9, _ta, _tb);
-                        _MM_TRANSPOSE4_PS(_tc, _td, _te, _tf);
-
-                        _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t0), _t4, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_t8), _tc, 1), 1);
-                        if (tj * 2 + 1 < w) _r1 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t1), _t5, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_t9), _td, 1), 1);
-                        if (tj * 2 + 2 < w) _r2 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t2), _t6, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_ta), _te, 1), 1);
-                        if (tj * 2 + 3 < w) _r3 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t3), _t7, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_tb), _tf, 1), 1);
-                    }
-                }
-
-                __m512 _tmp0 = _mm512_sub_ps(_r0, _r2);
-                __m512 _tmp1 = _mm512_add_ps(_r1, _r2);
-                __m512 _tmp2 = _mm512_sub_ps(_r2, _r1);
-                __m512 _tmp3 = _mm512_sub_ps(_r3, _r1);
-
-                _mm512_store_ps(tmp[0][m], _tmp0);
-                _mm512_store_ps(tmp[1][m], _tmp1);
-                _mm512_store_ps(tmp[2][m], _tmp2);
-                _mm512_store_ps(tmp[3][m], _tmp3);
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 16 + jj * 16;
-            float* p1 = p0 + max_jj * 16;
-            float* p2 = p0 + max_jj * 16 * 2;
-            float* p3 = p0 + max_jj * 16 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                __m512 _r0 = _mm512_load_ps(tmp[m][0]);
-                __m512 _r1 = _mm512_load_ps(tmp[m][1]);
-                __m512 _r2 = _mm512_load_ps(tmp[m][2]);
-                __m512 _r3 = _mm512_load_ps(tmp[m][3]);
-
-                __m512 _tmp0 = _mm512_sub_ps(_r0, _r2);
-                __m512 _tmp1 = _mm512_add_ps(_r1, _r2);
-                __m512 _tmp2 = _mm512_sub_ps(_r2, _r1);
-                __m512 _tmp3 = _mm512_sub_ps(_r3, _r1);
-
-                _mm512_store_ps(p0, _tmp0);
-                _mm512_store_ps(p1, _tmp1);
-                _mm512_store_ps(p2, _tmp2);
-                _mm512_store_ps(p3, _tmp3);
-
-                p0 += max_jj * 4 * 16;
-                p1 += max_jj * 4 * 16;
-                p2 += max_jj * 4 * 16;
-                p3 += max_jj * 4 * 16;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 16;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-#else // __AVX512F__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-    #pragma omp parallel for num_threads(nT)
-#endif // __AVX512F__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-#ifdef _MSC_VER
-        __declspec(align(32))
-#else
-        __attribute__((aligned(32)))
-#endif
-        float tmp[4][4][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel((k + kk) / elempack).row(ti * 2) + (tj * 2) * elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                __m256 _r0 = _mm256_setzero_ps();
-                __m256 _r1 = _mm256_setzero_ps();
-                __m256 _r2 = _mm256_setzero_ps();
-                __m256 _r3 = _mm256_setzero_ps();
-
-                if (ti * 2 + m < h)
-                {
-                    if (elempack == 8)
-                    {
-                        _r0 = _mm256_load_ps(r0);
-                        if (tj * 2 + 1 < w) _r1 = _mm256_load_ps(r0 + 8);
-                        if (tj * 2 + 2 < w) _r2 = _mm256_load_ps(r0 + 16);
-                        if (tj * 2 + 3 < w) _r3 = _mm256_load_ps(r0 + 24);
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-
-                        _r0 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0)), _mm_load_ps(r1), 1);
-                        if (tj * 2 + 1 < w) _r1 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 4)), _mm_load_ps(r1 + 4), 1);
-                        if (tj * 2 + 2 < w) _r2 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 8)), _mm_load_ps(r1 + 8), 1);
-                        if (tj * 2 + 3 < w) _r3 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 12)), _mm_load_ps(r1 + 12), 1);
-                    }
-                    if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-                        const float* r4 = r0 + N * 4;
-                        const float* r5 = r0 + N * 5;
-                        const float* r6 = r0 + N * 6;
-                        const float* r7 = r0 + N * 7;
-
-                        __m128 _t0 = _mm_loadu_ps(r0);
-                        __m128 _t1 = _mm_loadu_ps(r1);
-                        __m128 _t2 = _mm_loadu_ps(r2);
-                        __m128 _t3 = _mm_loadu_ps(r3);
-                        __m128 _t4 = _mm_loadu_ps(r4);
-                        __m128 _t5 = _mm_loadu_ps(r5);
-                        __m128 _t6 = _mm_loadu_ps(r6);
-                        __m128 _t7 = _mm_loadu_ps(r7);
-
-                        _MM_TRANSPOSE4_PS(_t0, _t1, _t2, _t3);
-                        _MM_TRANSPOSE4_PS(_t4, _t5, _t6, _t7);
-
-                        _r0 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t0), _t4, 1);
-                        if (tj * 2 + 1 < w) _r1 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t1), _t5, 1);
-                        if (tj * 2 + 2 < w) _r2 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t2), _t6, 1);
-                        if (tj * 2 + 3 < w) _r3 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t3), _t7, 1);
-                    }
-                }
-
-                __m256 _tmp0 = _mm256_sub_ps(_r0, _r2);
-                __m256 _tmp1 = _mm256_add_ps(_r1, _r2);
-                __m256 _tmp2 = _mm256_sub_ps(_r2, _r1);
-                __m256 _tmp3 = _mm256_sub_ps(_r3, _r1);
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                // old gcc breaks stack variable alignement
-                // ref https://gcc.gnu.org/bugzilla/show_bug.cgi?id=16660
-                _mm256_storeu_ps(tmp[0][m], _tmp0);
-                _mm256_storeu_ps(tmp[1][m], _tmp1);
-                _mm256_storeu_ps(tmp[2][m], _tmp2);
-                _mm256_storeu_ps(tmp[3][m], _tmp3);
-#else
-                _mm256_store_ps(tmp[0][m], _tmp0);
-                _mm256_store_ps(tmp[1][m], _tmp1);
-                _mm256_store_ps(tmp[2][m], _tmp2);
-                _mm256_store_ps(tmp[3][m], _tmp3);
-#endif
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 16 + jj * 8;
-            float* p1 = p0 + max_jj * 8;
-            float* p2 = p0 + max_jj * 8 * 2;
-            float* p3 = p0 + max_jj * 8 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m256 _r0 = _mm256_loadu_ps(tmp[m][0]);
-                __m256 _r1 = _mm256_loadu_ps(tmp[m][1]);
-                __m256 _r2 = _mm256_loadu_ps(tmp[m][2]);
-                __m256 _r3 = _mm256_loadu_ps(tmp[m][3]);
-#else
-                __m256 _r0 = _mm256_load_ps(tmp[m][0]);
-                __m256 _r1 = _mm256_load_ps(tmp[m][1]);
-                __m256 _r2 = _mm256_load_ps(tmp[m][2]);
-                __m256 _r3 = _mm256_load_ps(tmp[m][3]);
-#endif
-
-                __m256 _tmp0 = _mm256_sub_ps(_r0, _r2);
-                __m256 _tmp1 = _mm256_add_ps(_r1, _r2);
-                __m256 _tmp2 = _mm256_sub_ps(_r2, _r1);
-                __m256 _tmp3 = _mm256_sub_ps(_r3, _r1);
-
-                _mm256_store_ps(p0, _tmp0);
-                _mm256_store_ps(p1, _tmp1);
-                _mm256_store_ps(p2, _tmp2);
-                _mm256_store_ps(p3, _tmp3);
-
-                p0 += max_jj * 4 * 8;
-                p1 += max_jj * 4 * 8;
-                p2 += max_jj * 4 * 8;
-                p3 += max_jj * 4 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-#else // __AVX__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-    #pragma omp parallel for num_threads(nT)
-#endif // __AVX__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 4;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[4][4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel((k + kk) / elempack).row(ti * 2) + (tj * 2) * elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                __m128 _r0 = _mm_setzero_ps();
-                __m128 _r1 = _mm_setzero_ps();
-                __m128 _r2 = _mm_setzero_ps();
-                __m128 _r3 = _mm_setzero_ps();
-
-                if (ti * 2 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        _r0 = _mm_load_ps(r0);
-                        if (tj * 2 + 1 < w) _r1 = _mm_load_ps(r0 + 4);
-                        if (tj * 2 + 2 < w) _r2 = _mm_load_ps(r0 + 8);
-                        if (tj * 2 + 3 < w) _r3 = _mm_load_ps(r0 + 12);
-                    }
-                    if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-
-                        __m128 _t0 = _mm_loadu_ps(r0);
-                        __m128 _t1 = _mm_loadu_ps(r1);
-                        __m128 _t2 = _mm_loadu_ps(r2);
-                        __m128 _t3 = _mm_loadu_ps(r3);
-
-                        _MM_TRANSPOSE4_PS(_t0, _t1, _t2, _t3);
-
-                        _r0 = _t0;
-                        if (tj * 2 + 1 < w) _r1 = _t1;
-                        if (tj * 2 + 2 < w) _r2 = _t2;
-                        if (tj * 2 + 3 < w) _r3 = _t3;
-                    }
-                }
-
-                __m128 _tmp0 = _mm_sub_ps(_r0, _r2);
-                __m128 _tmp1 = _mm_add_ps(_r1, _r2);
-                __m128 _tmp2 = _mm_sub_ps(_r2, _r1);
-                __m128 _tmp3 = _mm_sub_ps(_r3, _r1);
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm_storeu_ps(tmp[0][m], _tmp0);
-                _mm_storeu_ps(tmp[1][m], _tmp1);
-                _mm_storeu_ps(tmp[2][m], _tmp2);
-                _mm_storeu_ps(tmp[3][m], _tmp3);
-#else
-                _mm_store_ps(tmp[0][m], _tmp0);
-                _mm_store_ps(tmp[1][m], _tmp1);
-                _mm_store_ps(tmp[2][m], _tmp2);
-                _mm_store_ps(tmp[3][m], _tmp3);
-#endif
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 16 + jj * 4;
-            float* p1 = p0 + max_jj * 4;
-            float* p2 = p0 + max_jj * 4 * 2;
-            float* p3 = p0 + max_jj * 4 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m128 _r0 = _mm_loadu_ps(tmp[m][0]);
-                __m128 _r1 = _mm_loadu_ps(tmp[m][1]);
-                __m128 _r2 = _mm_loadu_ps(tmp[m][2]);
-                __m128 _r3 = _mm_loadu_ps(tmp[m][3]);
-#else
-                __m128 _r0 = _mm_load_ps(tmp[m][0]);
-                __m128 _r1 = _mm_load_ps(tmp[m][1]);
-                __m128 _r2 = _mm_load_ps(tmp[m][2]);
-                __m128 _r3 = _mm_load_ps(tmp[m][3]);
-#endif
-
-                __m128 _tmp0 = _mm_sub_ps(_r0, _r2);
-                __m128 _tmp1 = _mm_add_ps(_r1, _r2);
-                __m128 _tmp2 = _mm_sub_ps(_r2, _r1);
-                __m128 _tmp3 = _mm_sub_ps(_r3, _r1);
-
-                _mm_store_ps(p0, _tmp0);
-                _mm_store_ps(p1, _tmp1);
-                _mm_store_ps(p2, _tmp2);
-                _mm_store_ps(p3, _tmp3);
-
-                p0 += max_jj * 4 * 4;
-                p1 += max_jj * 4 * 4;
-                p2 += max_jj * 4 * 4;
-                p3 += max_jj * 4 * 4;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 4;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-#else // __SSE2__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    #pragma omp parallel for num_threads(nT)
-#endif // __SSE2__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-        float tmp[4][4][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel(k + kk).row(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 4; m++)
-            {
-                float r00 = 0.f;
-                float r01 = 0.f;
-                float r10 = 0.f;
-                float r11 = 0.f;
-                float r20 = 0.f;
-                float r21 = 0.f;
-                float r30 = 0.f;
-                float r31 = 0.f;
-
-                if (ti * 2 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-
-                        r00 = r0[0];
-                        r01 = r1[0];
-                        if (tj * 2 + 1 < w)
-                        {
-                            r10 = r0[1];
-                            r11 = r1[1];
-                        }
-                        if (tj * 2 + 2 < w)
-                        {
-                            r20 = r0[2];
-                            r21 = r1[2];
-                        }
-                        if (tj * 2 + 3 < w)
-                        {
-                            r30 = r0[3];
-                            r31 = r1[3];
-                        }
-                    }
-                }
-
-                tmp[0][m][0] = r00 - r20;
-                tmp[0][m][1] = r01 - r21;
-                tmp[1][m][0] = r10 + r20;
-                tmp[1][m][1] = r11 + r21;
-                tmp[2][m][0] = r20 - r10;
-                tmp[2][m][1] = r21 - r11;
-                tmp[3][m][0] = r30 - r10;
-                tmp[3][m][1] = r31 - r11;
-
-                r0 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 16 + jj * 2;
-            float* p1 = p0 + max_jj * 2;
-            float* p2 = p0 + max_jj * 2 * 2;
-            float* p3 = p0 + max_jj * 2 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-
-                p0[0] = r00 - r20;
-                p0[1] = r01 - r21;
-                p1[0] = r10 + r20;
-                p1[1] = r11 + r21;
-                p2[0] = r20 - r10;
-                p2[1] = r21 - r11;
-                p3[0] = r30 - r10;
-                p3[1] = r31 - r11;
-
-                p0 += max_jj * 4 * 2;
-                p1 += max_jj * 4 * 2;
-                p2 += max_jj * 4 * 2;
-                p3 += max_jj * 4 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        float tmp[4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0123 = bottom_blob.channel(k + kk).row(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 4; m++)
-            {
-                float r0 = 0.f;
-                float r1 = 0.f;
-                float r2 = 0.f;
-                float r3 = 0.f;
-
-                if (ti * 2 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = r0123[0];
-                        if (tj * 2 + 1 < w) r1 = r0123[1];
-                        if (tj * 2 + 2 < w) r2 = r0123[2];
-                        if (tj * 2 + 3 < w) r3 = r0123[3];
-                    }
-                }
-
-                tmp[0][m] = r0 - r2;
-                tmp[1][m] = r1 + r2;
-                tmp[2][m] = r2 - r1;
-                tmp[3][m] = r3 - r1;
-
-                r0123 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 16 + jj;
-            float* p1 = p0 + max_jj;
-            float* p2 = p0 + max_jj * 2;
-            float* p3 = p0 + max_jj * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-
-                p0[0] = r0 - r2;
-                p1[0] = r1 + r2;
-                p2[0] = r2 - r1;
-                p3[0] = r3 - r1;
-
-                p0 += max_jj * 4;
-                p1 += max_jj * 4;
-                p2 += max_jj * 4;
-                p3 += max_jj * 4;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_output_tile(const Mat& top_tile, Mat& top_blob, const Mat& bias, int i, int max_ii, int j, int max_jj)
-{
-    // const float otm[2][4] = {
-    //     {1.0f,  1.0f,  1.0f,  0.0f},
-    //     {0.0f,  1.0f, -1.0f,  1.0f}
-    // };
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 1) / 2;
-
-    const float* biasptr = bias;
-
-    int ii = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    for (; ii + 15 < max_ii; ii += 16)
-    {
-        __m512 _bias0 = biasptr ? _mm512_loadu_ps(biasptr + i + ii) : _mm512_setzero_ps();
-
-#ifdef _MSC_VER
-        __declspec(align(64))
-#else
-        __attribute__((aligned(64)))
-#endif
-        float tmp[2][4][16];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 16 + jj * 16;
-            const float* r1 = r0 + max_jj * 16;
-            const float* r2 = r0 + max_jj * 16 * 2;
-            const float* r3 = r0 + max_jj * 16 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                __m512 _r0 = _mm512_load_ps(r0);
-                __m512 _r1 = _mm512_load_ps(r1);
-                __m512 _r2 = _mm512_load_ps(r2);
-                __m512 _r3 = _mm512_load_ps(r3);
-
-                __m512 _tmp0 = _mm512_add_ps(_mm512_add_ps(_r0, _r1), _r2);
-                __m512 _tmp1 = _mm512_add_ps(_mm512_sub_ps(_r1, _r2), _r3);
-
-                _mm512_store_ps(tmp[0][m], _tmp0);
-                _mm512_store_ps(tmp[1][m], _tmp1);
-
-                r0 += max_jj * 4 * 16;
-                r1 += max_jj * 4 * 16;
-                r2 += max_jj * 4 * 16;
-                r3 += max_jj * 4 * 16;
-            }
-
-            float* outptr0 = top_blob.channel((i + ii) / out_elempack).row(ti * 2) + (tj * 2) * out_elempack;
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                __m512 _r0 = _mm512_load_ps(tmp[m][0]);
-                __m512 _r1 = _mm512_load_ps(tmp[m][1]);
-                __m512 _r2 = _mm512_load_ps(tmp[m][2]);
-                __m512 _r3 = _mm512_load_ps(tmp[m][3]);
-
-                __m512 _tmp0 = _mm512_add_ps(_bias0, _mm512_add_ps(_mm512_add_ps(_r0, _r1), _r2));
-                __m512 _tmp1 = _mm512_add_ps(_bias0, _mm512_add_ps(_mm512_sub_ps(_r1, _r2), _r3));
-
-                if (out_elempack == 16)
-                {
-                    _mm512_store_ps(outptr0, _tmp0);
-                    if (tj * 2 + 1 < outw)
-                    {
-                        _mm512_store_ps(outptr0 + 16, _tmp1);
-                    }
-                }
-                if (out_elempack == 8)
-                {
-                    float* outptr1 = outptr0 + N;
-
-                    _mm256_store_ps(outptr0, _mm512_extractf32x8_ps(_tmp0, 0));
-                    _mm256_store_ps(outptr1, _mm512_extractf32x8_ps(_tmp0, 1));
-                    if (tj * 2 + 1 < outw)
-                    {
-                        _mm256_store_ps(outptr0 + 8, _mm512_extractf32x8_ps(_tmp1, 0));
-                        _mm256_store_ps(outptr1 + 8, _mm512_extractf32x8_ps(_tmp1, 1));
-                    }
-                }
-                if (out_elempack == 4)
-                {
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-
-                    _mm_store_ps(outptr0, _mm512_extractf32x4_ps(_tmp0, 0));
-                    _mm_store_ps(outptr1, _mm512_extractf32x4_ps(_tmp0, 1));
-                    _mm_store_ps(outptr2, _mm512_extractf32x4_ps(_tmp0, 2));
-                    _mm_store_ps(outptr3, _mm512_extractf32x4_ps(_tmp0, 3));
-                    if (tj * 2 + 1 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 4, _mm512_extractf32x4_ps(_tmp1, 0));
-                        _mm_store_ps(outptr1 + 4, _mm512_extractf32x4_ps(_tmp1, 1));
-                        _mm_store_ps(outptr2 + 4, _mm512_extractf32x4_ps(_tmp1, 2));
-                        _mm_store_ps(outptr3 + 4, _mm512_extractf32x4_ps(_tmp1, 3));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    float tmp0[16];
-                    float tmp1[16];
-                    _mm512_storeu_ps(tmp0, _tmp0);
-                    _mm512_storeu_ps(tmp1, _tmp1);
-
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-                    float* outptr4 = outptr0 + N * 4;
-                    float* outptr5 = outptr0 + N * 5;
-                    float* outptr6 = outptr0 + N * 6;
-                    float* outptr7 = outptr0 + N * 7;
-                    float* outptr8 = outptr0 + N * 8;
-                    float* outptr9 = outptr0 + N * 9;
-                    float* outptra = outptr0 + N * 10;
-                    float* outptrb = outptr0 + N * 11;
-                    float* outptrc = outptr0 + N * 12;
-                    float* outptrd = outptr0 + N * 13;
-                    float* outptre = outptr0 + N * 14;
-                    float* outptrf = outptr0 + N * 15;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-                    outptr8[0] = tmp0[8];
-                    outptr9[0] = tmp0[9];
-                    outptra[0] = tmp0[10];
-                    outptrb[0] = tmp0[11];
-                    outptrc[0] = tmp0[12];
-                    outptrd[0] = tmp0[13];
-                    outptre[0] = tmp0[14];
-                    outptrf[0] = tmp0[15];
-
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                        outptr8[1] = tmp1[8];
-                        outptr9[1] = tmp1[9];
-                        outptra[1] = tmp1[10];
-                        outptrb[1] = tmp1[11];
-                        outptrc[1] = tmp1[12];
-                        outptrd[1] = tmp1[13];
-                        outptre[1] = tmp1[14];
-                        outptrf[1] = tmp1[15];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __AVX512F__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        __m256 _bias0 = biasptr ? _mm256_loadu_ps(biasptr + i + ii) : _mm256_setzero_ps();
-
-#ifdef _MSC_VER
-        __declspec(align(32))
-#else
-        __attribute__((aligned(32)))
-#endif
-        float tmp[2][4][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 16 + jj * 8;
-            const float* r1 = r0 + max_jj * 8;
-            const float* r2 = r0 + max_jj * 8 * 2;
-            const float* r3 = r0 + max_jj * 8 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                __m256 _r0 = _mm256_load_ps(r0);
-                __m256 _r1 = _mm256_load_ps(r1);
-                __m256 _r2 = _mm256_load_ps(r2);
-                __m256 _r3 = _mm256_load_ps(r3);
-
-                __m256 _tmp0 = _mm256_add_ps(_mm256_add_ps(_r0, _r1), _r2);
-                __m256 _tmp1 = _mm256_add_ps(_mm256_sub_ps(_r1, _r2), _r3);
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm256_storeu_ps(tmp[0][m], _tmp0);
-                _mm256_storeu_ps(tmp[1][m], _tmp1);
-#else
-                _mm256_store_ps(tmp[0][m], _tmp0);
-                _mm256_store_ps(tmp[1][m], _tmp1);
-#endif
-
-                r0 += max_jj * 4 * 8;
-                r1 += max_jj * 4 * 8;
-                r2 += max_jj * 4 * 8;
-                r3 += max_jj * 4 * 8;
-            }
-
-            float* outptr0 = top_blob.channel((i + ii) / out_elempack).row(ti * 2) + (tj * 2) * out_elempack;
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m256 _r0 = _mm256_loadu_ps(tmp[m][0]);
-                __m256 _r1 = _mm256_loadu_ps(tmp[m][1]);
-                __m256 _r2 = _mm256_loadu_ps(tmp[m][2]);
-                __m256 _r3 = _mm256_loadu_ps(tmp[m][3]);
-#else
-                __m256 _r0 = _mm256_load_ps(tmp[m][0]);
-                __m256 _r1 = _mm256_load_ps(tmp[m][1]);
-                __m256 _r2 = _mm256_load_ps(tmp[m][2]);
-                __m256 _r3 = _mm256_load_ps(tmp[m][3]);
-#endif
-
-                __m256 _tmp0 = _mm256_add_ps(_bias0, _mm256_add_ps(_mm256_add_ps(_r0, _r1), _r2));
-                __m256 _tmp1 = _mm256_add_ps(_bias0, _mm256_add_ps(_mm256_sub_ps(_r1, _r2), _r3));
-
-                if (out_elempack == 8)
-                {
-                    _mm256_store_ps(outptr0, _tmp0);
-                    if (tj * 2 + 1 < outw)
-                    {
-                        _mm256_store_ps(outptr0 + 8, _tmp1);
-                    }
-                }
-                if (out_elempack == 4)
-                {
-                    float* outptr1 = outptr0 + N;
-
-                    _mm_store_ps(outptr0, _mm256_extractf128_ps(_tmp0, 0));
-                    _mm_store_ps(outptr1, _mm256_extractf128_ps(_tmp0, 1));
-                    if (tj * 2 + 1 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 4, _mm256_extractf128_ps(_tmp1, 0));
-                        _mm_store_ps(outptr1 + 4, _mm256_extractf128_ps(_tmp1, 1));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    float tmp0[8];
-                    float tmp1[8];
-                    _mm256_storeu_ps(tmp0, _tmp0);
-                    _mm256_storeu_ps(tmp1, _tmp1);
-
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-                    float* outptr4 = outptr0 + N * 4;
-                    float* outptr5 = outptr0 + N * 5;
-                    float* outptr6 = outptr0 + N * 6;
-                    float* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __AVX__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        __m128 _bias0 = biasptr ? _mm_loadu_ps(biasptr + i + ii) : _mm_setzero_ps();
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[2][4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 16 + jj * 4;
-            const float* r1 = r0 + max_jj * 4;
-            const float* r2 = r0 + max_jj * 4 * 2;
-            const float* r3 = r0 + max_jj * 4 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                __m128 _r0 = _mm_load_ps(r0);
-                __m128 _r1 = _mm_load_ps(r1);
-                __m128 _r2 = _mm_load_ps(r2);
-                __m128 _r3 = _mm_load_ps(r3);
-
-                __m128 _tmp0 = _mm_add_ps(_mm_add_ps(_r0, _r1), _r2);
-                __m128 _tmp1 = _mm_add_ps(_mm_sub_ps(_r1, _r2), _r3);
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm_storeu_ps(tmp[0][m], _tmp0);
-                _mm_storeu_ps(tmp[1][m], _tmp1);
-#else
-                _mm_store_ps(tmp[0][m], _tmp0);
-                _mm_store_ps(tmp[1][m], _tmp1);
-#endif
-
-                r0 += max_jj * 4 * 4;
-                r1 += max_jj * 4 * 4;
-                r2 += max_jj * 4 * 4;
-                r3 += max_jj * 4 * 4;
-            }
-
-            float* outptr0 = top_blob.channel((i + ii) / out_elempack).row(ti * 2) + (tj * 2) * out_elempack;
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m128 _r0 = _mm_loadu_ps(tmp[m][0]);
-                __m128 _r1 = _mm_loadu_ps(tmp[m][1]);
-                __m128 _r2 = _mm_loadu_ps(tmp[m][2]);
-                __m128 _r3 = _mm_loadu_ps(tmp[m][3]);
-#else
-                __m128 _r0 = _mm_load_ps(tmp[m][0]);
-                __m128 _r1 = _mm_load_ps(tmp[m][1]);
-                __m128 _r2 = _mm_load_ps(tmp[m][2]);
-                __m128 _r3 = _mm_load_ps(tmp[m][3]);
-#endif
-
-                __m128 _tmp0 = _mm_add_ps(_bias0, _mm_add_ps(_mm_add_ps(_r0, _r1), _r2));
-                __m128 _tmp1 = _mm_add_ps(_bias0, _mm_add_ps(_mm_sub_ps(_r1, _r2), _r3));
-
-                if (out_elempack == 4)
-                {
-                    _mm_store_ps(outptr0, _tmp0);
-                    if (tj * 2 + 1 < outw) _mm_store_ps(outptr0 + 4, _tmp1);
-                }
-                if (out_elempack == 1)
-                {
-                    float tmp0[4];
-                    float tmp1[4];
-                    _mm_storeu_ps(tmp0, _tmp0);
-                    _mm_storeu_ps(tmp1, _tmp1);
-
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __SSE2__
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-        float bias1 = biasptr ? biasptr[i + ii + 1] : 0.f;
-
-        float tmp[2][4][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 16 + jj * 2;
-            const float* r1 = r0 + max_jj * 2;
-            const float* r2 = r0 + max_jj * 2 * 2;
-            const float* r3 = r0 + max_jj * 2 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                tmp[0][m][0] = r0[0] + r1[0] + r2[0];
-                tmp[0][m][1] = r0[1] + r1[1] + r2[1];
-                tmp[1][m][0] = r1[0] - r2[0] + r3[0];
-                tmp[1][m][1] = r1[1] - r2[1] + r3[1];
-
-                r0 += max_jj * 4 * 2;
-                r1 += max_jj * 4 * 2;
-                r2 += max_jj * 4 * 2;
-                r3 += max_jj * 4 * 2;
-            }
-
-            float* outptr0 = top_blob.channel(i + ii).row(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-
-                float tmp00 = bias0 + r00 + r10 + r20;
-                float tmp01 = bias1 + r01 + r11 + r21;
-                float tmp10 = bias0 + r10 - r20 + r30;
-                float tmp11 = bias1 + r11 - r21 + r31;
-
-                // if (out_elempack == 1)
-                {
-                    float* outptr1 = outptr0 + N;
-
-                    outptr0[0] = tmp00;
-                    outptr1[0] = tmp01;
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp10;
-                        outptr1[1] = tmp11;
-                    }
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-
-        float tmp[2][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 16 + jj;
-            const float* r1 = r0 + max_jj;
-            const float* r2 = r0 + max_jj * 2;
-            const float* r3 = r0 + max_jj * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                tmp[0][m] = r0[0] + r1[0] + r2[0];
-                tmp[1][m] = r1[0] - r2[0] + r3[0];
-
-                r0 += max_jj * 4;
-                r1 += max_jj * 4;
-                r2 += max_jj * 4;
-                r3 += max_jj * 4;
-            }
-
-            float* outptr0 = top_blob.channel(i + ii).row(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-
-                float tmp0 = bias0 + r0 + r1 + r2;
-                float tmp1 = bias0 + r1 - r2 + r3;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = tmp0;
-                    if (tj * 2 + 1 < outw) outptr0[1] = tmp1;
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, const Mat& bias, int nT, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 2n+2, winograd F(2,3)
-    int w_tiles = (outw + 1) / 2;
-    int h_tiles = (outh + 1) / 2;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 16;
-
-    // NCNN_LOGE("conv3x3s1_winograd23 %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    get_optimal_tile_mnk(M, N, K, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 4u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 4u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd23_transform_input_tile(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 4u, opt.workspace_allocator);
-
-        #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd23_transform_input_tile(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                gemm_transB_packed_tile(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk);
-            }
-
-            // transform output
-            conv3x3s1_winograd23_transform_output_tile(top_tile, top_blob, bias, i, max_ii, j, max_jj);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_kernel_tile(const Mat& kernel, Mat& A, int inch, int i, int max_ii, int k, int max_kk)
-{
-    float* ptmp = A;
-
-    int ii = 0;
-    for (; ii < max_ii; ii++)
-    {
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            const float sq2 = 1.41421356237f;
-            // const float ktm[6][3] = {
-            //     {1.0f, 0.0f, 0.0f},
-            //     {-2.0f / 3, -sq2 / 3, -1.0f / 3},
-            //     {-2.0f / 3, sq2 / 3, -1.0f / 3},
-            //     {1.0f / 6, sq2 / 6, 1.0f / 3},
-            //     {1.0f / 6, -sq2 / 6, 1.0f / 3},
-            //     {0.0f, 0.0f, 1.0f}
-            // };
-            const float ktm0 = 2.0f / 3;
-            const float ktm1 = sq2 / 3;
-            const float ktm2 = 1.0f / 3;
-            const float ktm3 = 1.0f / 6;
-            const float ktm4 = sq2 / 6;
-
-            float tmp[6][3];
-
-            const float* k0 = (const float*)kernel + (i + ii) * inch * 9 + (k + kk) * 9;
-
-            for (int m = 0; m < 3; m++)
-            {
-                float r0 = k0[0];
-                float r1 = k0[1];
-                float r2 = k0[2];
-
-                tmp[0][m] = r0;
-                tmp[1][m] = -r0 * ktm0 - r1 * ktm1 - r2 * ktm2;
-                tmp[2][m] = -r0 * ktm0 + r1 * ktm1 - r2 * ktm2;
-                tmp[3][m] = r0 * ktm3 + r1 * ktm4 + r2 * ktm2;
-                tmp[4][m] = r0 * ktm3 - r1 * ktm4 + r2 * ktm2;
-                tmp[5][m] = r2;
-
-                k0 += 3;
-            }
-
-            for (int m = 0; m < 6; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-
-                float z0 = r0;
-                float z1 = -r0 * ktm0 - r1 * ktm1 - r2 * ktm2;
-                float z2 = -r0 * ktm0 + r1 * ktm1 - r2 * ktm2;
-                float z3 = r0 * ktm3 + r1 * ktm4 + r2 * ktm2;
-                float z4 = r0 * ktm3 - r1 * ktm4 + r2 * ktm2;
-                float z5 = r2;
-
-                ptmp[0] = z0;
-                ptmp[1] = z1;
-                ptmp[2] = z2;
-                ptmp[3] = z3;
-                ptmp[4] = z4;
-                ptmp[5] = z5;
-                ptmp += 6;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_kernel(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt)
-{
-    const int M = outch;
-    const int K = inch;
-    const int B = 36;
-
-    int TILE_M, TILE_N, TILE_K;
-    get_optimal_tile_mnk(M, 0, K, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    Mat A_tileX(B * TILE_M * TILE_K, 1, opt.num_threads, 4u, (Allocator*)0);
-
-    AT.create(TILE_K * TILE_M, B, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M, 4u, (Allocator*)0);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat A_tile = A_tileX.channel(get_omp_thread_num());
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_ii = std::min((M - i), TILE_M);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            conv3x3s1_winograd43_transform_kernel_tile(kernel, A_tile, inch, i, max_ii, k, max_kk);
-
-            Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-            pack_A_tile(A_tile, AT_tile, B, max_ii, max_kk);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_input_tile(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    const float sq2 = 1.41421356237;
-    const float sq2_d2 = 1.41421356237 / 2;
-
-    // const float itm[6][6] = {
-    //     {1.0f,  0.0f,  -2.5f,  0.0f,  1.0f, 0.0f},
-    //     {0.0f, -sq2,   -2.0f,  sq2/2, 1.0f, 0.0f},
-    //     {0.0f,  sq2,   -2.0f, -sq2/2, 1.0f, 0.0f},
-    //     {0.0f, -sq2/2, -0.5f,  sq2,   1.0f, 0.0f},
-    //     {0.0f,  sq2/2, -0.5f, -sq2,   1.0f, 0.0f},
-    //     {0.0f,  1.0f,   0.0f,  -2.5f, 0.0f, 1.0f}
-    // };
-
-    // 0 =  r00 + r04 - 2.5f * r02
-    // 1 = -(sq2 * r01 - sq2_d2 * r03) + (r04 - 2 * r02)
-    // 2 =  (sq2 * r01 - sq2_d2 * r03) + (r04 - 2 * r02)
-    // 3 =  (sq2 * r03 - sq2_d2 * r01) + (r04 - 0.5f * r02)
-    // 4 = -(sq2 * r03 - sq2_d2 * r01) + (r04 - 0.5f * r02)
-    // 5 =  r01 + r05 - 2.5f * r03
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w + 1) / 4;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    nn_max_kk = max_kk / 16;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = ppkk * 16;
-
-#ifdef _MSC_VER
-        __declspec(align(64))
-#else
-        __attribute__((aligned(64)))
-#endif
-        float tmp[6][6][16];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel((k + kk) / elempack).row(ti * 4) + (tj * 4) * elempack;
-
-            __m512 _vm2_5 = _mm512_set1_ps(-2.5f);
-            __m512 _vsq2 = _mm512_set1_ps(sq2);
-            __m512 _vmsq2_d2 = _mm512_set1_ps(-sq2_d2);
-            __m512 _vm2 = _mm512_set1_ps(-2.f);
-            __m512 _vm0_5 = _mm512_set1_ps(-0.5f);
-
-            for (int m = 0; m < 6; m++)
-            {
-                __m512 _r0 = _mm512_setzero_ps();
-                __m512 _r1 = _mm512_setzero_ps();
-                __m512 _r2 = _mm512_setzero_ps();
-                __m512 _r3 = _mm512_setzero_ps();
-                __m512 _r4 = _mm512_setzero_ps();
-                __m512 _r5 = _mm512_setzero_ps();
-
-                if (ti * 4 + m < h)
-                {
-                    if (elempack == 16)
-                    {
-                        _r0 = _mm512_load_ps(r0);
-                        if (tj * 4 + 1 < w) _r1 = _mm512_load_ps(r0 + 16);
-                        if (tj * 4 + 2 < w) _r2 = _mm512_load_ps(r0 + 32);
-                        if (tj * 4 + 3 < w) _r3 = _mm512_load_ps(r0 + 48);
-                        if (tj * 4 + 4 < w) _r4 = _mm512_load_ps(r0 + 64);
-                        if (tj * 4 + 5 < w) _r5 = _mm512_load_ps(r0 + 80);
-                    }
-                    if (elempack == 8)
-                    {
-                        const float* r1 = r0 + N;
-
-                        _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0)), _mm256_load_ps(r1), 1);
-                        if (tj * 4 + 1 < w) _r1 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + 8)), _mm256_load_ps(r1 + 8), 1);
-                        if (tj * 4 + 2 < w) _r2 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + 16)), _mm256_load_ps(r1 + 16), 1);
-                        if (tj * 4 + 3 < w) _r3 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + 24)), _mm256_load_ps(r1 + 24), 1);
-                        if (tj * 4 + 4 < w) _r4 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + 32)), _mm256_load_ps(r1 + 32), 1);
-                        if (tj * 4 + 5 < w) _r5 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + 40)), _mm256_load_ps(r1 + 40), 1);
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-
-                        _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0)), _mm_load_ps(r1), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2)), _mm_load_ps(r3), 1), 1);
-                        if (tj * 4 + 1 < w) _r1 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 4)), _mm_load_ps(r1 + 4), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + 4)), _mm_load_ps(r3 + 4), 1), 1);
-                        if (tj * 4 + 2 < w) _r2 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 8)), _mm_load_ps(r1 + 8), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + 8)), _mm_load_ps(r3 + 8), 1), 1);
-                        if (tj * 4 + 3 < w) _r3 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 12)), _mm_load_ps(r1 + 12), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + 12)), _mm_load_ps(r3 + 12), 1), 1);
-                        if (tj * 4 + 4 < w) _r4 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 16)), _mm_load_ps(r1 + 16), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + 16)), _mm_load_ps(r3 + 16), 1), 1);
-                        if (tj * 4 + 5 < w) _r5 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 20)), _mm_load_ps(r1 + 20), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + 20)), _mm_load_ps(r3 + 20), 1), 1);
-                    }
-                    if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-                        const float* r4 = r0 + N * 4;
-                        const float* r5 = r0 + N * 5;
-                        const float* r6 = r0 + N * 6;
-                        const float* r7 = r0 + N * 7;
-                        const float* r8 = r0 + N * 8;
-                        const float* r9 = r0 + N * 9;
-                        const float* ra = r0 + N * 10;
-                        const float* rb = r0 + N * 11;
-                        const float* rc = r0 + N * 12;
-                        const float* rd = r0 + N * 13;
-                        const float* re = r0 + N * 14;
-                        const float* rf = r0 + N * 15;
-
-                        __m128 _t0 = _mm_loadu_ps(r0);
-                        __m128 _t1 = _mm_loadu_ps(r1);
-                        __m128 _t2 = _mm_loadu_ps(r2);
-                        __m128 _t3 = _mm_loadu_ps(r3);
-                        __m128 _t4 = _mm_loadu_ps(r4);
-                        __m128 _t5 = _mm_loadu_ps(r5);
-                        __m128 _t6 = _mm_loadu_ps(r6);
-                        __m128 _t7 = _mm_loadu_ps(r7);
-                        __m128 _t8 = _mm_loadu_ps(r8);
-                        __m128 _t9 = _mm_loadu_ps(r9);
-                        __m128 _ta = _mm_loadu_ps(ra);
-                        __m128 _tb = _mm_loadu_ps(rb);
-                        __m128 _tc = _mm_loadu_ps(rc);
-                        __m128 _td = _mm_loadu_ps(rd);
-                        __m128 _te = _mm_loadu_ps(re);
-                        __m128 _tf = _mm_loadu_ps(rf);
-
-                        _MM_TRANSPOSE4_PS(_t0, _t1, _t2, _t3);
-                        _MM_TRANSPOSE4_PS(_t4, _t5, _t6, _t7);
-                        _MM_TRANSPOSE4_PS(_t8, _t9, _ta, _tb);
-                        _MM_TRANSPOSE4_PS(_tc, _td, _te, _tf);
-
-                        _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t0), _t4, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_t8), _tc, 1), 1);
-                        if (tj * 4 + 1 < w) _r1 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t1), _t5, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_t9), _td, 1), 1);
-                        if (tj * 4 + 2 < w) _r2 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t2), _t6, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_ta), _te, 1), 1);
-                        if (tj * 4 + 3 < w) _r3 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t3), _t7, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_tb), _tf, 1), 1);
-                        if (tj * 4 + 4 < w) _r4 = _mm512_set_ps(rf[4], re[4], rd[4], rc[4], rb[4], ra[4], r9[4], r8[4], r7[4], r6[4], r5[4], r4[4], r3[4], r2[4], r1[4], r0[4]);
-                        if (tj * 4 + 5 < w) _r5 = _mm512_set_ps(rf[5], re[5], rd[5], rc[5], rb[5], ra[5], r9[5], r8[5], r7[5], r6[5], r5[5], r4[5], r3[5], r2[5], r1[5], r0[5]);
-                    }
-                }
-
-                __m512 _tmp12a = _mm512_fmadd_ps(_vmsq2_d2, _r3, _mm512_mul_ps(_r1, _vsq2));
-                __m512 _tmp12b = _mm512_fmadd_ps(_vm2, _r2, _r4);
-                __m512 _tmp34a = _mm512_fmadd_ps(_vmsq2_d2, _r1, _mm512_mul_ps(_r3, _vsq2));
-                __m512 _tmp34b = _mm512_fmadd_ps(_vm0_5, _r2, _r4);
-
-                __m512 _tmp0 = _mm512_fmadd_ps(_vm2_5, _r2, _mm512_add_ps(_r0, _r4));
-                __m512 _tmp1 = _mm512_sub_ps(_tmp12b, _tmp12a);
-                __m512 _tmp2 = _mm512_add_ps(_tmp12b, _tmp12a);
-                __m512 _tmp3 = _mm512_add_ps(_tmp34b, _tmp34a);
-                __m512 _tmp4 = _mm512_sub_ps(_tmp34b, _tmp34a);
-                __m512 _tmp5 = _mm512_fmadd_ps(_vm2_5, _r3, _mm512_add_ps(_r1, _r5));
-
-                _mm512_store_ps(tmp[0][m], _tmp0);
-                _mm512_store_ps(tmp[1][m], _tmp1);
-                _mm512_store_ps(tmp[2][m], _tmp2);
-                _mm512_store_ps(tmp[3][m], _tmp3);
-                _mm512_store_ps(tmp[4][m], _tmp4);
-                _mm512_store_ps(tmp[5][m], _tmp5);
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 36 + jj * 16;
-            float* p1 = p0 + max_jj * 16;
-            float* p2 = p0 + max_jj * 16 * 2;
-            float* p3 = p0 + max_jj * 16 * 3;
-            float* p4 = p0 + max_jj * 16 * 4;
-            float* p5 = p0 + max_jj * 16 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                __m512 _r0 = _mm512_load_ps(tmp[m][0]);
-                __m512 _r1 = _mm512_load_ps(tmp[m][1]);
-                __m512 _r2 = _mm512_load_ps(tmp[m][2]);
-                __m512 _r3 = _mm512_load_ps(tmp[m][3]);
-                __m512 _r4 = _mm512_load_ps(tmp[m][4]);
-                __m512 _r5 = _mm512_load_ps(tmp[m][5]);
-
-                __m512 _tmp12a = _mm512_fmadd_ps(_vmsq2_d2, _r3, _mm512_mul_ps(_r1, _vsq2));
-                __m512 _tmp12b = _mm512_fmadd_ps(_vm2, _r2, _r4);
-                __m512 _tmp34a = _mm512_fmadd_ps(_vmsq2_d2, _r1, _mm512_mul_ps(_r3, _vsq2));
-                __m512 _tmp34b = _mm512_fmadd_ps(_vm0_5, _r2, _r4);
-
-                __m512 _tmp0 = _mm512_fmadd_ps(_vm2_5, _r2, _mm512_add_ps(_r0, _r4));
-                __m512 _tmp1 = _mm512_sub_ps(_tmp12b, _tmp12a);
-                __m512 _tmp2 = _mm512_add_ps(_tmp12b, _tmp12a);
-                __m512 _tmp3 = _mm512_add_ps(_tmp34b, _tmp34a);
-                __m512 _tmp4 = _mm512_sub_ps(_tmp34b, _tmp34a);
-                __m512 _tmp5 = _mm512_fmadd_ps(_vm2_5, _r3, _mm512_add_ps(_r1, _r5));
-
-                _mm512_store_ps(p0, _tmp0);
-                _mm512_store_ps(p1, _tmp1);
-                _mm512_store_ps(p2, _tmp2);
-                _mm512_store_ps(p3, _tmp3);
-                _mm512_store_ps(p4, _tmp4);
-                _mm512_store_ps(p5, _tmp5);
-
-                p0 += max_jj * 6 * 16;
-                p1 += max_jj * 6 * 16;
-                p2 += max_jj * 6 * 16;
-                p3 += max_jj * 6 * 16;
-                p4 += max_jj * 6 * 16;
-                p5 += max_jj * 6 * 16;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 16;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-#else // __AVX512F__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-    #pragma omp parallel for num_threads(nT)
-#endif // __AVX512F__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-#ifdef _MSC_VER
-        __declspec(align(32))
-#else
-        __attribute__((aligned(32)))
-#endif
-        float tmp[6][6][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel((k + kk) / elempack).row(ti * 4) + (tj * 4) * elempack;
-
-            __m256 _vm2_5 = _mm256_set1_ps(-2.5f);
-            __m256 _vsq2 = _mm256_set1_ps(sq2);
-            __m256 _vmsq2_d2 = _mm256_set1_ps(-sq2_d2);
-            __m256 _vm2 = _mm256_set1_ps(-2.f);
-            __m256 _vm0_5 = _mm256_set1_ps(-0.5f);
-
-            for (int m = 0; m < 6; m++)
-            {
-                __m256 _r0 = _mm256_setzero_ps();
-                __m256 _r1 = _mm256_setzero_ps();
-                __m256 _r2 = _mm256_setzero_ps();
-                __m256 _r3 = _mm256_setzero_ps();
-                __m256 _r4 = _mm256_setzero_ps();
-                __m256 _r5 = _mm256_setzero_ps();
-
-                if (ti * 4 + m < h)
-                {
-                    if (elempack == 8)
-                    {
-                        _r0 = _mm256_load_ps(r0);
-                        if (tj * 4 + 1 < w) _r1 = _mm256_load_ps(r0 + 8);
-                        if (tj * 4 + 2 < w) _r2 = _mm256_load_ps(r0 + 16);
-                        if (tj * 4 + 3 < w) _r3 = _mm256_load_ps(r0 + 24);
-                        if (tj * 4 + 4 < w) _r4 = _mm256_load_ps(r0 + 32);
-                        if (tj * 4 + 5 < w) _r5 = _mm256_load_ps(r0 + 40);
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-
-                        _r0 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0)), _mm_load_ps(r1), 1);
-                        if (tj * 4 + 1 < w) _r1 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 4)), _mm_load_ps(r1 + 4), 1);
-                        if (tj * 4 + 2 < w) _r2 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 8)), _mm_load_ps(r1 + 8), 1);
-                        if (tj * 4 + 3 < w) _r3 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 12)), _mm_load_ps(r1 + 12), 1);
-                        if (tj * 4 + 4 < w) _r4 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 16)), _mm_load_ps(r1 + 16), 1);
-                        if (tj * 4 + 5 < w) _r5 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 20)), _mm_load_ps(r1 + 20), 1);
-                    }
-                    if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-                        const float* r4 = r0 + N * 4;
-                        const float* r5 = r0 + N * 5;
-                        const float* r6 = r0 + N * 6;
-                        const float* r7 = r0 + N * 7;
-
-                        __m128 _t0 = _mm_loadu_ps(r0);
-                        __m128 _t1 = _mm_loadu_ps(r1);
-                        __m128 _t2 = _mm_loadu_ps(r2);
-                        __m128 _t3 = _mm_loadu_ps(r3);
-                        __m128 _t4 = _mm_loadu_ps(r4);
-                        __m128 _t5 = _mm_loadu_ps(r5);
-                        __m128 _t6 = _mm_loadu_ps(r6);
-                        __m128 _t7 = _mm_loadu_ps(r7);
-
-                        _MM_TRANSPOSE4_PS(_t0, _t1, _t2, _t3);
-                        _MM_TRANSPOSE4_PS(_t4, _t5, _t6, _t7);
-
-                        _r0 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t0), _t4, 1);
-                        if (tj * 4 + 1 < w) _r1 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t1), _t5, 1);
-                        if (tj * 4 + 2 < w) _r2 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t2), _t6, 1);
-                        if (tj * 4 + 3 < w) _r3 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t3), _t7, 1);
-                        if (tj * 4 + 4 < w) _r4 = _mm256_set_ps(r7[4], r6[4], r5[4], r4[4], r3[4], r2[4], r1[4], r0[4]);
-                        if (tj * 4 + 5 < w) _r5 = _mm256_set_ps(r7[5], r6[5], r5[5], r4[5], r3[5], r2[5], r1[5], r0[5]);
-                    }
-                }
-
-                __m256 _tmp12a = _mm256_comp_fmadd_ps(_vmsq2_d2, _r3, _mm256_mul_ps(_r1, _vsq2));
-                __m256 _tmp12b = _mm256_comp_fmadd_ps(_vm2, _r2, _r4);
-                __m256 _tmp34a = _mm256_comp_fmadd_ps(_vmsq2_d2, _r1, _mm256_mul_ps(_r3, _vsq2));
-                __m256 _tmp34b = _mm256_comp_fmadd_ps(_vm0_5, _r2, _r4);
-
-                __m256 _tmp0 = _mm256_comp_fmadd_ps(_vm2_5, _r2, _mm256_add_ps(_r0, _r4));
-                __m256 _tmp1 = _mm256_sub_ps(_tmp12b, _tmp12a);
-                __m256 _tmp2 = _mm256_add_ps(_tmp12b, _tmp12a);
-                __m256 _tmp3 = _mm256_add_ps(_tmp34b, _tmp34a);
-                __m256 _tmp4 = _mm256_sub_ps(_tmp34b, _tmp34a);
-                __m256 _tmp5 = _mm256_comp_fmadd_ps(_vm2_5, _r3, _mm256_add_ps(_r1, _r5));
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm256_storeu_ps(tmp[0][m], _tmp0);
-                _mm256_storeu_ps(tmp[1][m], _tmp1);
-                _mm256_storeu_ps(tmp[2][m], _tmp2);
-                _mm256_storeu_ps(tmp[3][m], _tmp3);
-                _mm256_storeu_ps(tmp[4][m], _tmp4);
-                _mm256_storeu_ps(tmp[5][m], _tmp5);
-#else
-                _mm256_store_ps(tmp[0][m], _tmp0);
-                _mm256_store_ps(tmp[1][m], _tmp1);
-                _mm256_store_ps(tmp[2][m], _tmp2);
-                _mm256_store_ps(tmp[3][m], _tmp3);
-                _mm256_store_ps(tmp[4][m], _tmp4);
-                _mm256_store_ps(tmp[5][m], _tmp5);
-#endif
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 36 + jj * 8;
-            float* p1 = p0 + max_jj * 8;
-            float* p2 = p0 + max_jj * 8 * 2;
-            float* p3 = p0 + max_jj * 8 * 3;
-            float* p4 = p0 + max_jj * 8 * 4;
-            float* p5 = p0 + max_jj * 8 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m256 _r0 = _mm256_loadu_ps(tmp[m][0]);
-                __m256 _r1 = _mm256_loadu_ps(tmp[m][1]);
-                __m256 _r2 = _mm256_loadu_ps(tmp[m][2]);
-                __m256 _r3 = _mm256_loadu_ps(tmp[m][3]);
-                __m256 _r4 = _mm256_loadu_ps(tmp[m][4]);
-                __m256 _r5 = _mm256_loadu_ps(tmp[m][5]);
-#else
-                __m256 _r0 = _mm256_load_ps(tmp[m][0]);
-                __m256 _r1 = _mm256_load_ps(tmp[m][1]);
-                __m256 _r2 = _mm256_load_ps(tmp[m][2]);
-                __m256 _r3 = _mm256_load_ps(tmp[m][3]);
-                __m256 _r4 = _mm256_load_ps(tmp[m][4]);
-                __m256 _r5 = _mm256_load_ps(tmp[m][5]);
-#endif
-
-                __m256 _tmp12a = _mm256_comp_fmadd_ps(_vmsq2_d2, _r3, _mm256_mul_ps(_r1, _vsq2));
-                __m256 _tmp12b = _mm256_comp_fmadd_ps(_vm2, _r2, _r4);
-                __m256 _tmp34a = _mm256_comp_fmadd_ps(_vmsq2_d2, _r1, _mm256_mul_ps(_r3, _vsq2));
-                __m256 _tmp34b = _mm256_comp_fmadd_ps(_vm0_5, _r2, _r4);
-
-                __m256 _tmp0 = _mm256_comp_fmadd_ps(_vm2_5, _r2, _mm256_add_ps(_r0, _r4));
-                __m256 _tmp1 = _mm256_sub_ps(_tmp12b, _tmp12a);
-                __m256 _tmp2 = _mm256_add_ps(_tmp12b, _tmp12a);
-                __m256 _tmp3 = _mm256_add_ps(_tmp34b, _tmp34a);
-                __m256 _tmp4 = _mm256_sub_ps(_tmp34b, _tmp34a);
-                __m256 _tmp5 = _mm256_comp_fmadd_ps(_vm2_5, _r3, _mm256_add_ps(_r1, _r5));
-
-                _mm256_store_ps(p0, _tmp0);
-                _mm256_store_ps(p1, _tmp1);
-                _mm256_store_ps(p2, _tmp2);
-                _mm256_store_ps(p3, _tmp3);
-                _mm256_store_ps(p4, _tmp4);
-                _mm256_store_ps(p5, _tmp5);
-
-                p0 += max_jj * 6 * 8;
-                p1 += max_jj * 6 * 8;
-                p2 += max_jj * 6 * 8;
-                p3 += max_jj * 6 * 8;
-                p4 += max_jj * 6 * 8;
-                p5 += max_jj * 6 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-#else // __AVX__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-    #pragma omp parallel for num_threads(nT)
-#endif // __AVX__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 4;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[6][6][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel((k + kk) / elempack).row(ti * 4) + (tj * 4) * elempack;
-
-            __m128 _vm2_5 = _mm_set1_ps(-2.5f);
-            __m128 _vsq2 = _mm_set1_ps(sq2);
-            __m128 _vmsq2_d2 = _mm_set1_ps(-sq2_d2);
-            __m128 _vm2 = _mm_set1_ps(-2.f);
-            __m128 _vm0_5 = _mm_set1_ps(-0.5f);
-
-            for (int m = 0; m < 6; m++)
-            {
-                __m128 _r0 = _mm_setzero_ps();
-                __m128 _r1 = _mm_setzero_ps();
-                __m128 _r2 = _mm_setzero_ps();
-                __m128 _r3 = _mm_setzero_ps();
-                __m128 _r4 = _mm_setzero_ps();
-                __m128 _r5 = _mm_setzero_ps();
-
-                if (ti * 4 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        _r0 = _mm_load_ps(r0);
-                        if (tj * 4 + 1 < w) _r1 = _mm_load_ps(r0 + 4);
-                        if (tj * 4 + 2 < w) _r2 = _mm_load_ps(r0 + 8);
-                        if (tj * 4 + 3 < w) _r3 = _mm_load_ps(r0 + 12);
-                        if (tj * 4 + 4 < w) _r4 = _mm_load_ps(r0 + 16);
-                        if (tj * 4 + 5 < w) _r5 = _mm_load_ps(r0 + 20);
-                    }
-                    if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-
-                        __m128 _t0 = _mm_loadu_ps(r0);
-                        __m128 _t1 = _mm_loadu_ps(r1);
-                        __m128 _t2 = _mm_loadu_ps(r2);
-                        __m128 _t3 = _mm_loadu_ps(r3);
-
-                        _MM_TRANSPOSE4_PS(_t0, _t1, _t2, _t3);
-
-                        _r0 = _t0;
-                        if (tj * 4 + 1 < w) _r1 = _t1;
-                        if (tj * 4 + 2 < w) _r2 = _t2;
-                        if (tj * 4 + 3 < w) _r3 = _t3;
-                        if (tj * 4 + 4 < w) _r4 = _mm_set_ps(r3[4], r2[4], r1[4], r0[4]);
-                        if (tj * 4 + 5 < w) _r5 = _mm_set_ps(r3[5], r2[5], r1[5], r0[5]);
-                    }
-                }
-
-                __m128 _tmp12a = _mm_comp_fmadd_ps(_vmsq2_d2, _r3, _mm_mul_ps(_r1, _vsq2));
-                __m128 _tmp12b = _mm_comp_fmadd_ps(_vm2, _r2, _r4);
-                __m128 _tmp34a = _mm_comp_fmadd_ps(_vmsq2_d2, _r1, _mm_mul_ps(_r3, _vsq2));
-                __m128 _tmp34b = _mm_comp_fmadd_ps(_vm0_5, _r2, _r4);
-
-                __m128 _tmp0 = _mm_comp_fmadd_ps(_vm2_5, _r2, _mm_add_ps(_r0, _r4));
-                __m128 _tmp1 = _mm_sub_ps(_tmp12b, _tmp12a);
-                __m128 _tmp2 = _mm_add_ps(_tmp12b, _tmp12a);
-                __m128 _tmp3 = _mm_add_ps(_tmp34b, _tmp34a);
-                __m128 _tmp4 = _mm_sub_ps(_tmp34b, _tmp34a);
-                __m128 _tmp5 = _mm_comp_fmadd_ps(_vm2_5, _r3, _mm_add_ps(_r1, _r5));
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm_storeu_ps(tmp[0][m], _tmp0);
-                _mm_storeu_ps(tmp[1][m], _tmp1);
-                _mm_storeu_ps(tmp[2][m], _tmp2);
-                _mm_storeu_ps(tmp[3][m], _tmp3);
-                _mm_storeu_ps(tmp[4][m], _tmp4);
-                _mm_storeu_ps(tmp[5][m], _tmp5);
-#else
-                _mm_store_ps(tmp[0][m], _tmp0);
-                _mm_store_ps(tmp[1][m], _tmp1);
-                _mm_store_ps(tmp[2][m], _tmp2);
-                _mm_store_ps(tmp[3][m], _tmp3);
-                _mm_store_ps(tmp[4][m], _tmp4);
-                _mm_store_ps(tmp[5][m], _tmp5);
-#endif
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 36 + jj * 4;
-            float* p1 = p0 + max_jj * 4;
-            float* p2 = p0 + max_jj * 4 * 2;
-            float* p3 = p0 + max_jj * 4 * 3;
-            float* p4 = p0 + max_jj * 4 * 4;
-            float* p5 = p0 + max_jj * 4 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m128 _r0 = _mm_loadu_ps(tmp[m][0]);
-                __m128 _r1 = _mm_loadu_ps(tmp[m][1]);
-                __m128 _r2 = _mm_loadu_ps(tmp[m][2]);
-                __m128 _r3 = _mm_loadu_ps(tmp[m][3]);
-                __m128 _r4 = _mm_loadu_ps(tmp[m][4]);
-                __m128 _r5 = _mm_loadu_ps(tmp[m][5]);
-#else
-                __m128 _r0 = _mm_load_ps(tmp[m][0]);
-                __m128 _r1 = _mm_load_ps(tmp[m][1]);
-                __m128 _r2 = _mm_load_ps(tmp[m][2]);
-                __m128 _r3 = _mm_load_ps(tmp[m][3]);
-                __m128 _r4 = _mm_load_ps(tmp[m][4]);
-                __m128 _r5 = _mm_load_ps(tmp[m][5]);
-#endif
-
-                __m128 _tmp12a = _mm_comp_fmadd_ps(_vmsq2_d2, _r3, _mm_mul_ps(_r1, _vsq2));
-                __m128 _tmp12b = _mm_comp_fmadd_ps(_vm2, _r2, _r4);
-                __m128 _tmp34a = _mm_comp_fmadd_ps(_vmsq2_d2, _r1, _mm_mul_ps(_r3, _vsq2));
-                __m128 _tmp34b = _mm_comp_fmadd_ps(_vm0_5, _r2, _r4);
-
-                __m128 _tmp0 = _mm_comp_fmadd_ps(_vm2_5, _r2, _mm_add_ps(_r0, _r4));
-                __m128 _tmp1 = _mm_sub_ps(_tmp12b, _tmp12a);
-                __m128 _tmp2 = _mm_add_ps(_tmp12b, _tmp12a);
-                __m128 _tmp3 = _mm_add_ps(_tmp34b, _tmp34a);
-                __m128 _tmp4 = _mm_sub_ps(_tmp34b, _tmp34a);
-                __m128 _tmp5 = _mm_comp_fmadd_ps(_vm2_5, _r3, _mm_add_ps(_r1, _r5));
-
-                _mm_store_ps(p0, _tmp0);
-                _mm_store_ps(p1, _tmp1);
-                _mm_store_ps(p2, _tmp2);
-                _mm_store_ps(p3, _tmp3);
-                _mm_store_ps(p4, _tmp4);
-                _mm_store_ps(p5, _tmp5);
-
-                p0 += max_jj * 6 * 4;
-                p1 += max_jj * 6 * 4;
-                p2 += max_jj * 6 * 4;
-                p3 += max_jj * 6 * 4;
-                p4 += max_jj * 6 * 4;
-                p5 += max_jj * 6 * 4;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 4;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-#else // __SSE2__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    #pragma omp parallel for num_threads(nT)
-#endif // __SSE2__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-        float tmp[6][6][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel(k + kk).row(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 6; m++)
-            {
-                float r00 = 0.f;
-                float r01 = 0.f;
-                float r10 = 0.f;
-                float r11 = 0.f;
-                float r20 = 0.f;
-                float r21 = 0.f;
-                float r30 = 0.f;
-                float r31 = 0.f;
-                float r40 = 0.f;
-                float r41 = 0.f;
-                float r50 = 0.f;
-                float r51 = 0.f;
-
-                if (ti * 4 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-
-                        r00 = r0[0];
-                        r01 = r1[0];
-                        if (tj * 4 + 1 < w)
-                        {
-                            r10 = r0[1];
-                            r11 = r1[1];
-                        }
-                        if (tj * 4 + 2 < w)
-                        {
-                            r20 = r0[2];
-                            r21 = r1[2];
-                        }
-                        if (tj * 4 + 3 < w)
-                        {
-                            r30 = r0[3];
-                            r31 = r1[3];
-                        }
-                        if (tj * 4 + 4 < w)
-                        {
-                            r40 = r0[4];
-                            r41 = r1[4];
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            r50 = r0[5];
-                            r51 = r1[5];
-                        }
-                    }
-                }
-
-                float tmp12a0 = sq2 * r10 - sq2_d2 * r30;
-                float tmp12a1 = sq2 * r11 - sq2_d2 * r31;
-                float tmp12b0 = r40 - 2 * r20;
-                float tmp12b1 = r41 - 2 * r21;
-                float tmp34a0 = sq2 * r30 - sq2_d2 * r10;
-                float tmp34a1 = sq2 * r31 - sq2_d2 * r11;
-                float tmp34b0 = r40 - 0.5f * r20;
-                float tmp34b1 = r41 - 0.5f * r21;
-
-                tmp[0][m][0] = r00 + r40 - 2.5f * r20;
-                tmp[0][m][1] = r01 + r41 - 2.5f * r21;
-                tmp[1][m][0] = tmp12b0 - tmp12a0;
-                tmp[1][m][1] = tmp12b1 - tmp12a1;
-                tmp[2][m][0] = tmp12b0 + tmp12a0;
-                tmp[2][m][1] = tmp12b1 + tmp12a1;
-                tmp[3][m][0] = tmp34b0 + tmp34a0;
-                tmp[3][m][1] = tmp34b1 + tmp34a1;
-                tmp[4][m][0] = tmp34b0 - tmp34a0;
-                tmp[4][m][1] = tmp34b1 - tmp34a1;
-                tmp[5][m][0] = r10 + r50 - 2.5f * r30;
-                tmp[5][m][1] = r11 + r51 - 2.5f * r31;
-
-                r0 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 36 + jj * 2;
-            float* p1 = p0 + max_jj * 2;
-            float* p2 = p0 + max_jj * 2 * 2;
-            float* p3 = p0 + max_jj * 2 * 3;
-            float* p4 = p0 + max_jj * 2 * 4;
-            float* p5 = p0 + max_jj * 2 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-                float r40 = tmp[m][4][0];
-                float r41 = tmp[m][4][1];
-                float r50 = tmp[m][5][0];
-                float r51 = tmp[m][5][1];
-
-                float tmp12a0 = sq2 * r10 - sq2_d2 * r30;
-                float tmp12a1 = sq2 * r11 - sq2_d2 * r31;
-                float tmp12b0 = r40 - 2 * r20;
-                float tmp12b1 = r41 - 2 * r21;
-                float tmp34a0 = sq2 * r30 - sq2_d2 * r10;
-                float tmp34a1 = sq2 * r31 - sq2_d2 * r11;
-                float tmp34b0 = r40 - 0.5f * r20;
-                float tmp34b1 = r41 - 0.5f * r21;
-
-                p0[0] = r00 + r40 - 2.5f * r20;
-                p0[1] = r01 + r41 - 2.5f * r21;
-                p1[0] = tmp12b0 - tmp12a0;
-                p1[1] = tmp12b1 - tmp12a1;
-                p2[0] = tmp12b0 + tmp12a0;
-                p2[1] = tmp12b1 + tmp12a1;
-                p3[0] = tmp34b0 + tmp34a0;
-                p3[1] = tmp34b1 + tmp34a1;
-                p4[0] = tmp34b0 - tmp34a0;
-                p4[1] = tmp34b1 - tmp34a1;
-                p5[0] = r10 + r50 - 2.5f * r30;
-                p5[1] = r11 + r51 - 2.5f * r31;
-
-                p0 += max_jj * 6 * 2;
-                p1 += max_jj * 6 * 2;
-                p2 += max_jj * 6 * 2;
-                p3 += max_jj * 6 * 2;
-                p4 += max_jj * 6 * 2;
-                p5 += max_jj * 6 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        float tmp[6][6];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0123 = bottom_blob.channel(k + kk).row(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 6; m++)
-            {
-                float r0 = 0.f;
-                float r1 = 0.f;
-                float r2 = 0.f;
-                float r3 = 0.f;
-                float r4 = 0.f;
-                float r5 = 0.f;
-
-                if (ti * 4 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = r0123[0];
-                        if (tj * 4 + 1 < w) r1 = r0123[1];
-                        if (tj * 4 + 2 < w) r2 = r0123[2];
-                        if (tj * 4 + 3 < w) r3 = r0123[3];
-                        if (tj * 4 + 4 < w) r4 = r0123[4];
-                        if (tj * 4 + 5 < w) r5 = r0123[5];
-                    }
-                }
-
-                float tmp12a = sq2 * r1 - sq2_d2 * r3;
-                float tmp12b = r4 - 2 * r2;
-                float tmp34a = sq2 * r3 - sq2_d2 * r1;
-                float tmp34b = r4 - 0.5f * r2;
-
-                tmp[0][m] = r0 + r4 - 2.5f * r2;
-                tmp[1][m] = tmp12b - tmp12a;
-                tmp[2][m] = tmp12b + tmp12a;
-                tmp[3][m] = tmp34b + tmp34a;
-                tmp[4][m] = tmp34b - tmp34a;
-                tmp[5][m] = r1 + r5 - 2.5f * r3;
-
-                r0123 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 36 + jj;
-            float* p1 = p0 + max_jj;
-            float* p2 = p0 + max_jj * 2;
-            float* p3 = p0 + max_jj * 3;
-            float* p4 = p0 + max_jj * 4;
-            float* p5 = p0 + max_jj * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-                float r4 = tmp[m][4];
-                float r5 = tmp[m][5];
-
-                float tmp12a = sq2 * r1 - sq2_d2 * r3;
-                float tmp12b = r4 - 2 * r2;
-                float tmp34a = sq2 * r3 - sq2_d2 * r1;
-                float tmp34b = r4 - 0.5f * r2;
-
-                p0[0] = r0 + r4 - 2.5f * r2;
-                p1[0] = tmp12b - tmp12a;
-                p2[0] = tmp12b + tmp12a;
-                p3[0] = tmp34b + tmp34a;
-                p4[0] = tmp34b - tmp34a;
-                p5[0] = r1 + r5 - 2.5f * r3;
-
-                p0 += max_jj * 6;
-                p1 += max_jj * 6;
-                p2 += max_jj * 6;
-                p3 += max_jj * 6;
-                p4 += max_jj * 6;
-                p5 += max_jj * 6;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_output_tile(const Mat& top_tile, Mat& top_blob, const Mat& bias, int i, int max_ii, int j, int max_jj)
-{
-    const float sq2 = 1.41421356237;
-    const float sq2_m2 = 1.41421356237 * 2;
-    const float sq2_d2 = 1.41421356237 / 2;
-    const float sq2_d4 = 1.41421356237 / 4;
-
-    // const float otm[4][6] = {
-    //     {1.0f, 1.0f,   1.0f,  1.0f,  1.0f,   0.0f},
-    //     {0.0f, sq2/2, -sq2/2, sq2,   -sq2,   0.0f},
-    //     {0.0f, 0.5f,   0.5f,  2.0f,  2.0f,   0.0f},
-    //     {0.0f, sq2/4, -sq2/4, sq2*2, -sq2*2, 1.0f}
-    // };
-
-    // 0 = r00 + (r01 + r02) + (r03 + r04)
-    // 1 =       (r01 - r02) * sq2_d2 + (r03 - r04) * sq2
-    // 2 =       (r01 + r02) * 0.5f + (r03 + r04) * 2
-    // 3 = r05 + (r01 - r02) * sq2_d4 + (r03 - r04) * sq2_m2
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 3) / 4;
-
-    const float* biasptr = bias;
-
-    int ii = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    for (; ii + 15 < max_ii; ii += 16)
-    {
-        __m512 _bias0 = biasptr ? _mm512_loadu_ps(biasptr + i + ii) : _mm512_setzero_ps();
-
-#ifdef _MSC_VER
-        __declspec(align(64))
-#else
-        __attribute__((aligned(64)))
-#endif
-        float tmp[4][6][16];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 36 + jj * 16;
-            const float* r1 = r0 + max_jj * 16;
-            const float* r2 = r0 + max_jj * 16 * 2;
-            const float* r3 = r0 + max_jj * 16 * 3;
-            const float* r4 = r0 + max_jj * 16 * 4;
-            const float* r5 = r0 + max_jj * 16 * 5;
-
-            __m512 _vsq2 = _mm512_set1_ps(sq2);
-            __m512 _vsq2_d2 = _mm512_set1_ps(sq2_d2);
-            __m512 _vsq2_d4 = _mm512_set1_ps(sq2_d4);
-            __m512 _vsq2_m2 = _mm512_set1_ps(sq2_m2);
-            __m512 _v0_5 = _mm512_set1_ps(0.5f);
-            __m512 _v2 = _mm512_set1_ps(2.f);
-
-            for (int m = 0; m < 6; m++)
-            {
-                __m512 _r0 = _mm512_load_ps(r0);
-                __m512 _r1 = _mm512_load_ps(r1);
-                __m512 _r2 = _mm512_load_ps(r2);
-                __m512 _r3 = _mm512_load_ps(r3);
-                __m512 _r4 = _mm512_load_ps(r4);
-                __m512 _r5 = _mm512_load_ps(r5);
-
-                __m512 _tmp02a = _mm512_add_ps(_r1, _r2);
-                __m512 _tmp02b = _mm512_add_ps(_r3, _r4);
-                __m512 _tmp13a = _mm512_sub_ps(_r1, _r2);
-                __m512 _tmp13b = _mm512_sub_ps(_r3, _r4);
-
-                __m512 _tmp0 = _mm512_add_ps(_mm512_add_ps(_r0, _tmp02a), _tmp02b);
-                __m512 _tmp1 = _mm512_fmadd_ps(_tmp13b, _vsq2, _mm512_mul_ps(_tmp13a, _vsq2_d2));
-                __m512 _tmp2 = _mm512_fmadd_ps(_tmp02b, _v2, _mm512_mul_ps(_tmp02a, _v0_5));
-                __m512 _tmp3 = _mm512_fmadd_ps(_tmp13b, _vsq2_m2, _mm512_fmadd_ps(_tmp13a, _vsq2_d4, _r5));
-
-                _mm512_store_ps(tmp[0][m], _tmp0);
-                _mm512_store_ps(tmp[1][m], _tmp1);
-                _mm512_store_ps(tmp[2][m], _tmp2);
-                _mm512_store_ps(tmp[3][m], _tmp3);
-
-                r0 += max_jj * 6 * 16;
-                r1 += max_jj * 6 * 16;
-                r2 += max_jj * 6 * 16;
-                r3 += max_jj * 6 * 16;
-                r4 += max_jj * 6 * 16;
-                r5 += max_jj * 6 * 16;
-            }
-
-            float* outptr0 = top_blob.channel((i + ii) / out_elempack).row(ti * 4) + (tj * 4) * out_elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                __m512 _r0 = _mm512_load_ps(tmp[m][0]);
-                __m512 _r1 = _mm512_load_ps(tmp[m][1]);
-                __m512 _r2 = _mm512_load_ps(tmp[m][2]);
-                __m512 _r3 = _mm512_load_ps(tmp[m][3]);
-                __m512 _r4 = _mm512_load_ps(tmp[m][4]);
-                __m512 _r5 = _mm512_load_ps(tmp[m][5]);
-
-                __m512 _tmp02a = _mm512_add_ps(_r1, _r2);
-                __m512 _tmp02b = _mm512_add_ps(_r3, _r4);
-                __m512 _tmp13a = _mm512_sub_ps(_r1, _r2);
-                __m512 _tmp13b = _mm512_sub_ps(_r3, _r4);
-
-                __m512 _tmp0 = _mm512_add_ps(_mm512_add_ps(_r0, _tmp02a), _mm512_add_ps(_tmp02b, _bias0));
-                __m512 _tmp1 = _mm512_fmadd_ps(_tmp13b, _vsq2, _mm512_fmadd_ps(_tmp13a, _vsq2_d2, _bias0));
-                __m512 _tmp2 = _mm512_fmadd_ps(_tmp02b, _v2, _mm512_fmadd_ps(_tmp02a, _v0_5, _bias0));
-                __m512 _tmp3 = _mm512_fmadd_ps(_tmp13b, _vsq2_m2, _mm512_fmadd_ps(_tmp13a, _vsq2_d4, _mm512_add_ps(_r5, _bias0)));
-
-                if (out_elempack == 16)
-                {
-                    _mm512_store_ps(outptr0, _tmp0);
-                    if (tj * 4 + 1 < outw) _mm512_store_ps(outptr0 + 16, _tmp1);
-                    if (tj * 4 + 2 < outw) _mm512_store_ps(outptr0 + 32, _tmp2);
-                    if (tj * 4 + 3 < outw) _mm512_store_ps(outptr0 + 48, _tmp3);
-                }
-                if (out_elempack == 8)
-                {
-                    float* outptr1 = outptr0 + N;
-
-                    _mm256_store_ps(outptr0, _mm512_extractf32x8_ps(_tmp0, 0));
-                    _mm256_store_ps(outptr1, _mm512_extractf32x8_ps(_tmp0, 1));
-                    if (tj * 4 + 1 < outw)
-                    {
-                        _mm256_store_ps(outptr0 + 8, _mm512_extractf32x8_ps(_tmp1, 0));
-                        _mm256_store_ps(outptr1 + 8, _mm512_extractf32x8_ps(_tmp1, 1));
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        _mm256_store_ps(outptr0 + 16, _mm512_extractf32x8_ps(_tmp2, 0));
-                        _mm256_store_ps(outptr1 + 16, _mm512_extractf32x8_ps(_tmp2, 1));
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        _mm256_store_ps(outptr0 + 24, _mm512_extractf32x8_ps(_tmp3, 0));
-                        _mm256_store_ps(outptr1 + 24, _mm512_extractf32x8_ps(_tmp3, 1));
-                    }
-                }
-                if (out_elempack == 4)
-                {
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-
-                    _mm_store_ps(outptr0, _mm512_extractf32x4_ps(_tmp0, 0));
-                    _mm_store_ps(outptr1, _mm512_extractf32x4_ps(_tmp0, 1));
-                    _mm_store_ps(outptr2, _mm512_extractf32x4_ps(_tmp0, 2));
-                    _mm_store_ps(outptr3, _mm512_extractf32x4_ps(_tmp0, 3));
-                    if (tj * 4 + 1 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 4, _mm512_extractf32x4_ps(_tmp1, 0));
-                        _mm_store_ps(outptr1 + 4, _mm512_extractf32x4_ps(_tmp1, 1));
-                        _mm_store_ps(outptr2 + 4, _mm512_extractf32x4_ps(_tmp1, 2));
-                        _mm_store_ps(outptr3 + 4, _mm512_extractf32x4_ps(_tmp1, 3));
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 8, _mm512_extractf32x4_ps(_tmp2, 0));
-                        _mm_store_ps(outptr1 + 8, _mm512_extractf32x4_ps(_tmp2, 1));
-                        _mm_store_ps(outptr2 + 8, _mm512_extractf32x4_ps(_tmp2, 2));
-                        _mm_store_ps(outptr3 + 8, _mm512_extractf32x4_ps(_tmp2, 3));
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 12, _mm512_extractf32x4_ps(_tmp3, 0));
-                        _mm_store_ps(outptr1 + 12, _mm512_extractf32x4_ps(_tmp3, 1));
-                        _mm_store_ps(outptr2 + 12, _mm512_extractf32x4_ps(_tmp3, 2));
-                        _mm_store_ps(outptr3 + 12, _mm512_extractf32x4_ps(_tmp3, 3));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    float tmp0[16];
-                    float tmp1[16];
-                    float tmp2[16];
-                    float tmp3[16];
-                    _mm512_storeu_ps(tmp0, _tmp0);
-                    _mm512_storeu_ps(tmp1, _tmp1);
-                    _mm512_storeu_ps(tmp2, _tmp2);
-                    _mm512_storeu_ps(tmp3, _tmp3);
-
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-                    float* outptr4 = outptr0 + N * 4;
-                    float* outptr5 = outptr0 + N * 5;
-                    float* outptr6 = outptr0 + N * 6;
-                    float* outptr7 = outptr0 + N * 7;
-                    float* outptr8 = outptr0 + N * 8;
-                    float* outptr9 = outptr0 + N * 9;
-                    float* outptra = outptr0 + N * 10;
-                    float* outptrb = outptr0 + N * 11;
-                    float* outptrc = outptr0 + N * 12;
-                    float* outptrd = outptr0 + N * 13;
-                    float* outptre = outptr0 + N * 14;
-                    float* outptrf = outptr0 + N * 15;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-                    outptr8[0] = tmp0[8];
-                    outptr9[0] = tmp0[9];
-                    outptra[0] = tmp0[10];
-                    outptrb[0] = tmp0[11];
-                    outptrc[0] = tmp0[12];
-                    outptrd[0] = tmp0[13];
-                    outptre[0] = tmp0[14];
-                    outptrf[0] = tmp0[15];
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                        outptr8[1] = tmp1[8];
-                        outptr9[1] = tmp1[9];
-                        outptra[1] = tmp1[10];
-                        outptrb[1] = tmp1[11];
-                        outptrc[1] = tmp1[12];
-                        outptrd[1] = tmp1[13];
-                        outptre[1] = tmp1[14];
-                        outptrf[1] = tmp1[15];
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                        outptr4[2] = tmp2[4];
-                        outptr5[2] = tmp2[5];
-                        outptr6[2] = tmp2[6];
-                        outptr7[2] = tmp2[7];
-                        outptr8[2] = tmp2[8];
-                        outptr9[2] = tmp2[9];
-                        outptra[2] = tmp2[10];
-                        outptrb[2] = tmp2[11];
-                        outptrc[2] = tmp2[12];
-                        outptrd[2] = tmp2[13];
-                        outptre[2] = tmp2[14];
-                        outptrf[2] = tmp2[15];
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                        outptr4[3] = tmp3[4];
-                        outptr5[3] = tmp3[5];
-                        outptr6[3] = tmp3[6];
-                        outptr7[3] = tmp3[7];
-                        outptr8[3] = tmp3[8];
-                        outptr9[3] = tmp3[9];
-                        outptra[3] = tmp3[10];
-                        outptrb[3] = tmp3[11];
-                        outptrc[3] = tmp3[12];
-                        outptrd[3] = tmp3[13];
-                        outptre[3] = tmp3[14];
-                        outptrf[3] = tmp3[15];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __AVX512F__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        __m256 _bias0 = biasptr ? _mm256_loadu_ps(biasptr + i + ii) : _mm256_setzero_ps();
-
-#ifdef _MSC_VER
-        __declspec(align(32))
-#else
-        __attribute__((aligned(32)))
-#endif
-        float tmp[4][6][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 36 + jj * 8;
-            const float* r1 = r0 + max_jj * 8;
-            const float* r2 = r0 + max_jj * 8 * 2;
-            const float* r3 = r0 + max_jj * 8 * 3;
-            const float* r4 = r0 + max_jj * 8 * 4;
-            const float* r5 = r0 + max_jj * 8 * 5;
-
-            __m256 _vsq2 = _mm256_set1_ps(sq2);
-            __m256 _vsq2_d2 = _mm256_set1_ps(sq2_d2);
-            __m256 _vsq2_d4 = _mm256_set1_ps(sq2_d4);
-            __m256 _vsq2_m2 = _mm256_set1_ps(sq2_m2);
-            __m256 _v0_5 = _mm256_set1_ps(0.5f);
-            __m256 _v2 = _mm256_set1_ps(2.f);
-
-            for (int m = 0; m < 6; m++)
-            {
-                __m256 _r0 = _mm256_load_ps(r0);
-                __m256 _r1 = _mm256_load_ps(r1);
-                __m256 _r2 = _mm256_load_ps(r2);
-                __m256 _r3 = _mm256_load_ps(r3);
-                __m256 _r4 = _mm256_load_ps(r4);
-                __m256 _r5 = _mm256_load_ps(r5);
-
-                __m256 _tmp02a = _mm256_add_ps(_r1, _r2);
-                __m256 _tmp02b = _mm256_add_ps(_r3, _r4);
-                __m256 _tmp13a = _mm256_sub_ps(_r1, _r2);
-                __m256 _tmp13b = _mm256_sub_ps(_r3, _r4);
-
-                __m256 _tmp0 = _mm256_add_ps(_mm256_add_ps(_r0, _tmp02a), _tmp02b);
-                __m256 _tmp1 = _mm256_comp_fmadd_ps(_tmp13b, _vsq2, _mm256_mul_ps(_tmp13a, _vsq2_d2));
-                __m256 _tmp2 = _mm256_comp_fmadd_ps(_tmp02b, _v2, _mm256_mul_ps(_tmp02a, _v0_5));
-                __m256 _tmp3 = _mm256_comp_fmadd_ps(_tmp13b, _vsq2_m2, _mm256_comp_fmadd_ps(_tmp13a, _vsq2_d4, _r5));
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm256_storeu_ps(tmp[0][m], _tmp0);
-                _mm256_storeu_ps(tmp[1][m], _tmp1);
-                _mm256_storeu_ps(tmp[2][m], _tmp2);
-                _mm256_storeu_ps(tmp[3][m], _tmp3);
-#else
-                _mm256_store_ps(tmp[0][m], _tmp0);
-                _mm256_store_ps(tmp[1][m], _tmp1);
-                _mm256_store_ps(tmp[2][m], _tmp2);
-                _mm256_store_ps(tmp[3][m], _tmp3);
-#endif
-
-                r0 += max_jj * 6 * 8;
-                r1 += max_jj * 6 * 8;
-                r2 += max_jj * 6 * 8;
-                r3 += max_jj * 6 * 8;
-                r4 += max_jj * 6 * 8;
-                r5 += max_jj * 6 * 8;
-            }
-
-            float* outptr0 = top_blob.channel((i + ii) / out_elempack).row(ti * 4) + (tj * 4) * out_elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m256 _r0 = _mm256_loadu_ps(tmp[m][0]);
-                __m256 _r1 = _mm256_loadu_ps(tmp[m][1]);
-                __m256 _r2 = _mm256_loadu_ps(tmp[m][2]);
-                __m256 _r3 = _mm256_loadu_ps(tmp[m][3]);
-                __m256 _r4 = _mm256_loadu_ps(tmp[m][4]);
-                __m256 _r5 = _mm256_loadu_ps(tmp[m][5]);
-#else
-                __m256 _r0 = _mm256_load_ps(tmp[m][0]);
-                __m256 _r1 = _mm256_load_ps(tmp[m][1]);
-                __m256 _r2 = _mm256_load_ps(tmp[m][2]);
-                __m256 _r3 = _mm256_load_ps(tmp[m][3]);
-                __m256 _r4 = _mm256_load_ps(tmp[m][4]);
-                __m256 _r5 = _mm256_load_ps(tmp[m][5]);
-#endif
-
-                __m256 _tmp02a = _mm256_add_ps(_r1, _r2);
-                __m256 _tmp02b = _mm256_add_ps(_r3, _r4);
-                __m256 _tmp13a = _mm256_sub_ps(_r1, _r2);
-                __m256 _tmp13b = _mm256_sub_ps(_r3, _r4);
-
-                __m256 _tmp0 = _mm256_add_ps(_mm256_add_ps(_r0, _tmp02a), _mm256_add_ps(_tmp02b, _bias0));
-                __m256 _tmp1 = _mm256_comp_fmadd_ps(_tmp13b, _vsq2, _mm256_comp_fmadd_ps(_tmp13a, _vsq2_d2, _bias0));
-                __m256 _tmp2 = _mm256_comp_fmadd_ps(_tmp02b, _v2, _mm256_comp_fmadd_ps(_tmp02a, _v0_5, _bias0));
-                __m256 _tmp3 = _mm256_comp_fmadd_ps(_tmp13b, _vsq2_m2, _mm256_comp_fmadd_ps(_tmp13a, _vsq2_d4, _mm256_add_ps(_r5, _bias0)));
-
-                if (out_elempack == 8)
-                {
-                    _mm256_store_ps(outptr0, _tmp0);
-                    if (tj * 4 + 1 < outw) _mm256_store_ps(outptr0 + 8, _tmp1);
-                    if (tj * 4 + 2 < outw) _mm256_store_ps(outptr0 + 16, _tmp2);
-                    if (tj * 4 + 3 < outw) _mm256_store_ps(outptr0 + 24, _tmp3);
-                }
-                if (out_elempack == 4)
-                {
-                    float* outptr1 = outptr0 + N;
-
-                    _mm_store_ps(outptr0, _mm256_extractf128_ps(_tmp0, 0));
-                    _mm_store_ps(outptr1, _mm256_extractf128_ps(_tmp0, 1));
-                    if (tj * 4 + 1 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 4, _mm256_extractf128_ps(_tmp1, 0));
-                        _mm_store_ps(outptr1 + 4, _mm256_extractf128_ps(_tmp1, 1));
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 8, _mm256_extractf128_ps(_tmp2, 0));
-                        _mm_store_ps(outptr1 + 8, _mm256_extractf128_ps(_tmp2, 1));
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 12, _mm256_extractf128_ps(_tmp3, 0));
-                        _mm_store_ps(outptr1 + 12, _mm256_extractf128_ps(_tmp3, 1));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    float tmp0[8];
-                    float tmp1[8];
-                    float tmp2[8];
-                    float tmp3[8];
-                    _mm256_storeu_ps(tmp0, _tmp0);
-                    _mm256_storeu_ps(tmp1, _tmp1);
-                    _mm256_storeu_ps(tmp2, _tmp2);
-                    _mm256_storeu_ps(tmp3, _tmp3);
-
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-                    float* outptr4 = outptr0 + N * 4;
-                    float* outptr5 = outptr0 + N * 5;
-                    float* outptr6 = outptr0 + N * 6;
-                    float* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                        outptr4[2] = tmp2[4];
-                        outptr5[2] = tmp2[5];
-                        outptr6[2] = tmp2[6];
-                        outptr7[2] = tmp2[7];
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                        outptr4[3] = tmp3[4];
-                        outptr5[3] = tmp3[5];
-                        outptr6[3] = tmp3[6];
-                        outptr7[3] = tmp3[7];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __AVX__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        __m128 _bias0 = biasptr ? _mm_loadu_ps(biasptr + i + ii) : _mm_setzero_ps();
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[4][6][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 36 + jj * 4;
-            const float* r1 = r0 + max_jj * 4;
-            const float* r2 = r0 + max_jj * 4 * 2;
-            const float* r3 = r0 + max_jj * 4 * 3;
-            const float* r4 = r0 + max_jj * 4 * 4;
-            const float* r5 = r0 + max_jj * 4 * 5;
-
-            __m128 _vsq2 = _mm_set1_ps(sq2);
-            __m128 _vsq2_d2 = _mm_set1_ps(sq2_d2);
-            __m128 _vsq2_d4 = _mm_set1_ps(sq2_d4);
-            __m128 _vsq2_m2 = _mm_set1_ps(sq2_m2);
-            __m128 _v0_5 = _mm_set1_ps(0.5f);
-            __m128 _v2 = _mm_set1_ps(2.f);
-
-            for (int m = 0; m < 6; m++)
-            {
-                __m128 _r0 = _mm_load_ps(r0);
-                __m128 _r1 = _mm_load_ps(r1);
-                __m128 _r2 = _mm_load_ps(r2);
-                __m128 _r3 = _mm_load_ps(r3);
-                __m128 _r4 = _mm_load_ps(r4);
-                __m128 _r5 = _mm_load_ps(r5);
-
-                __m128 _tmp02a = _mm_add_ps(_r1, _r2);
-                __m128 _tmp02b = _mm_add_ps(_r3, _r4);
-                __m128 _tmp13a = _mm_sub_ps(_r1, _r2);
-                __m128 _tmp13b = _mm_sub_ps(_r3, _r4);
-
-                __m128 _tmp0 = _mm_add_ps(_mm_add_ps(_r0, _tmp02a), _tmp02b);
-                __m128 _tmp1 = _mm_comp_fmadd_ps(_tmp13b, _vsq2, _mm_mul_ps(_tmp13a, _vsq2_d2));
-                __m128 _tmp2 = _mm_comp_fmadd_ps(_tmp02b, _v2, _mm_mul_ps(_tmp02a, _v0_5));
-                __m128 _tmp3 = _mm_comp_fmadd_ps(_tmp13b, _vsq2_m2, _mm_comp_fmadd_ps(_tmp13a, _vsq2_d4, _r5));
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm_storeu_ps(tmp[0][m], _tmp0);
-                _mm_storeu_ps(tmp[1][m], _tmp1);
-                _mm_storeu_ps(tmp[2][m], _tmp2);
-                _mm_storeu_ps(tmp[3][m], _tmp3);
-#else
-                _mm_store_ps(tmp[0][m], _tmp0);
-                _mm_store_ps(tmp[1][m], _tmp1);
-                _mm_store_ps(tmp[2][m], _tmp2);
-                _mm_store_ps(tmp[3][m], _tmp3);
-#endif
-
-                r0 += max_jj * 6 * 4;
-                r1 += max_jj * 6 * 4;
-                r2 += max_jj * 6 * 4;
-                r3 += max_jj * 6 * 4;
-                r4 += max_jj * 6 * 4;
-                r5 += max_jj * 6 * 4;
-            }
-
-            float* outptr0 = top_blob.channel((i + ii) / out_elempack).row(ti * 4) + (tj * 4) * out_elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m128 _r0 = _mm_loadu_ps(tmp[m][0]);
-                __m128 _r1 = _mm_loadu_ps(tmp[m][1]);
-                __m128 _r2 = _mm_loadu_ps(tmp[m][2]);
-                __m128 _r3 = _mm_loadu_ps(tmp[m][3]);
-                __m128 _r4 = _mm_loadu_ps(tmp[m][4]);
-                __m128 _r5 = _mm_loadu_ps(tmp[m][5]);
-#else
-                __m128 _r0 = _mm_load_ps(tmp[m][0]);
-                __m128 _r1 = _mm_load_ps(tmp[m][1]);
-                __m128 _r2 = _mm_load_ps(tmp[m][2]);
-                __m128 _r3 = _mm_load_ps(tmp[m][3]);
-                __m128 _r4 = _mm_load_ps(tmp[m][4]);
-                __m128 _r5 = _mm_load_ps(tmp[m][5]);
-#endif
-
-                __m128 _tmp02a = _mm_add_ps(_r1, _r2);
-                __m128 _tmp02b = _mm_add_ps(_r3, _r4);
-                __m128 _tmp13a = _mm_sub_ps(_r1, _r2);
-                __m128 _tmp13b = _mm_sub_ps(_r3, _r4);
-
-                __m128 _tmp0 = _mm_add_ps(_mm_add_ps(_r0, _tmp02a), _mm_add_ps(_tmp02b, _bias0));
-                __m128 _tmp1 = _mm_comp_fmadd_ps(_tmp13b, _vsq2, _mm_comp_fmadd_ps(_tmp13a, _vsq2_d2, _bias0));
-                __m128 _tmp2 = _mm_comp_fmadd_ps(_tmp02b, _v2, _mm_comp_fmadd_ps(_tmp02a, _v0_5, _bias0));
-                __m128 _tmp3 = _mm_comp_fmadd_ps(_tmp13b, _vsq2_m2, _mm_comp_fmadd_ps(_tmp13a, _vsq2_d4, _mm_add_ps(_r5, _bias0)));
-
-                if (out_elempack == 4)
-                {
-                    _mm_store_ps(outptr0, _tmp0);
-                    if (tj * 4 + 1 < outw) _mm_store_ps(outptr0 + 4, _tmp1);
-                    if (tj * 4 + 2 < outw) _mm_store_ps(outptr0 + 8, _tmp2);
-                    if (tj * 4 + 3 < outw) _mm_store_ps(outptr0 + 12, _tmp3);
-                }
-                if (out_elempack == 1)
-                {
-                    float tmp0[4];
-                    float tmp1[4];
-                    float tmp2[4];
-                    float tmp3[4];
-                    _mm_storeu_ps(tmp0, _tmp0);
-                    _mm_storeu_ps(tmp1, _tmp1);
-                    _mm_storeu_ps(tmp2, _tmp2);
-                    _mm_storeu_ps(tmp3, _tmp3);
-
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __SSE2__
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-        float bias1 = biasptr ? biasptr[i + ii + 1] : 0.f;
-
-        float tmp[4][6][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 36 + jj * 2;
-            const float* r1 = r0 + max_jj * 2;
-            const float* r2 = r0 + max_jj * 2 * 2;
-            const float* r3 = r0 + max_jj * 2 * 3;
-            const float* r4 = r0 + max_jj * 2 * 4;
-            const float* r5 = r0 + max_jj * 2 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float tmp02a0 = r1[0] + r2[0];
-                float tmp02a1 = r1[1] + r2[1];
-                float tmp02b0 = r3[0] + r4[0];
-                float tmp02b1 = r3[1] + r4[1];
-                float tmp13a0 = r1[0] - r2[0];
-                float tmp13a1 = r1[1] - r2[1];
-                float tmp13b0 = r3[0] - r4[0];
-                float tmp13b1 = r3[1] - r4[1];
-
-                tmp[0][m][0] = r0[0] + tmp02a0 + tmp02b0;
-                tmp[0][m][1] = r0[1] + tmp02a1 + tmp02b1;
-                tmp[1][m][0] = tmp13a0 * sq2_d2 + tmp13b0 * sq2;
-                tmp[1][m][1] = tmp13a1 * sq2_d2 + tmp13b1 * sq2;
-                tmp[2][m][0] = tmp02a0 * 0.5f + tmp02b0 * 2;
-                tmp[2][m][1] = tmp02a1 * 0.5f + tmp02b1 * 2;
-                tmp[3][m][0] = r5[0] + tmp13a0 * sq2_d4 + tmp13b0 * sq2_m2;
-                tmp[3][m][1] = r5[1] + tmp13a1 * sq2_d4 + tmp13b1 * sq2_m2;
-
-                r0 += max_jj * 6 * 2;
-                r1 += max_jj * 6 * 2;
-                r2 += max_jj * 6 * 2;
-                r3 += max_jj * 6 * 2;
-                r4 += max_jj * 6 * 2;
-                r5 += max_jj * 6 * 2;
-            }
-
-            float* outptr0 = top_blob.channel(i + ii).row(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-                float r40 = tmp[m][4][0];
-                float r41 = tmp[m][4][1];
-                float r50 = tmp[m][5][0];
-                float r51 = tmp[m][5][1];
-
-                float tmp02a0 = r10 + r20;
-                float tmp02a1 = r11 + r21;
-                float tmp02b0 = r30 + r40;
-                float tmp02b1 = r31 + r41;
-                float tmp13a0 = r10 - r20;
-                float tmp13a1 = r11 - r21;
-                float tmp13b0 = r30 - r40;
-                float tmp13b1 = r31 - r41;
-
-                float tmp00 = bias0 + r00 + tmp02a0 + tmp02b0;
-                float tmp01 = bias1 + r01 + tmp02a1 + tmp02b1;
-                float tmp10 = bias0 + tmp13a0 * sq2_d2 + tmp13b0 * sq2;
-                float tmp11 = bias1 + tmp13a1 * sq2_d2 + tmp13b1 * sq2;
-                float tmp20 = bias0 + tmp02a0 * 0.5f + tmp02b0 * 2;
-                float tmp21 = bias1 + tmp02a1 * 0.5f + tmp02b1 * 2;
-                float tmp30 = bias0 + r50 + tmp13a0 * sq2_d4 + tmp13b0 * sq2_m2;
-                float tmp31 = bias1 + r51 + tmp13a1 * sq2_d4 + tmp13b1 * sq2_m2;
-
-                // if (out_elempack == 1)
-                {
-                    float* outptr1 = outptr0 + N;
-
-                    outptr0[0] = tmp00;
-                    outptr1[0] = tmp01;
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp10;
-                        outptr1[1] = tmp11;
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp20;
-                        outptr1[2] = tmp21;
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp30;
-                        outptr1[3] = tmp31;
-                    }
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-
-        float tmp[4][6];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 36 + jj;
-            const float* r1 = r0 + max_jj;
-            const float* r2 = r0 + max_jj * 2;
-            const float* r3 = r0 + max_jj * 3;
-            const float* r4 = r0 + max_jj * 4;
-            const float* r5 = r0 + max_jj * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                float tmp02a = r1[0] + r2[0];
-                float tmp02b = r3[0] + r4[0];
-                float tmp13a = r1[0] - r2[0];
-                float tmp13b = r3[0] - r4[0];
-
-                tmp[0][m] = r0[0] + tmp02a + tmp02b;
-                tmp[1][m] = tmp13a * sq2_d2 + tmp13b * sq2;
-                tmp[2][m] = tmp02a * 0.5f + tmp02b * 2;
-                tmp[3][m] = r5[0] + tmp13a * sq2_d4 + tmp13b * sq2_m2;
-
-                r0 += max_jj * 6;
-                r1 += max_jj * 6;
-                r2 += max_jj * 6;
-                r3 += max_jj * 6;
-                r4 += max_jj * 6;
-                r5 += max_jj * 6;
-            }
-
-            float* outptr0 = top_blob.channel(i + ii).row(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-                float r4 = tmp[m][4];
-                float r5 = tmp[m][5];
-
-                float tmp02a = r1 + r2;
-                float tmp02b = r3 + r4;
-                float tmp13a = r1 - r2;
-                float tmp13b = r3 - r4;
-
-                float tmp0 = bias0 + r0 + tmp02a + tmp02b;
-                float tmp1 = bias0 + tmp13a * sq2_d2 + tmp13b * sq2;
-                float tmp2 = bias0 + tmp02a * 0.5f + tmp02b * 2;
-                float tmp3 = bias0 + r5 + tmp13a * sq2_d4 + tmp13b * sq2_m2;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = tmp0;
-                    if (tj * 4 + 1 < outw) outptr0[1] = tmp1;
-                    if (tj * 4 + 2 < outw) outptr0[2] = tmp2;
-                    if (tj * 4 + 3 < outw) outptr0[3] = tmp3;
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, const Mat& bias, int nT, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 4n+2, winograd F(4,3)
-    int w_tiles = (outw + 3) / 4;
-    int h_tiles = (outh + 3) / 4;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 36;
-
-    // NCNN_LOGE("conv3x3s1_winograd43 %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    get_optimal_tile_mnk(M, N, K, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 4u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 4u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd43_transform_input_tile(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 4u, opt.workspace_allocator);
-
-        #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd43_transform_input_tile(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                gemm_transB_packed_tile(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk);
-            }
-
-            // transform output
-            conv3x3s1_winograd43_transform_output_tile(top_tile, top_blob, bias, i, max_ii, j, max_jj);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd63_transform_kernel_tile(const Mat& kernel, Mat& A, int inch, int i, int max_ii, int k, int max_kk)
-{
-    float* ptmp = A;
-
-    int ii = 0;
-    for (; ii < max_ii; ii++)
-    {
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            // const float ktm[8][3] = {
-            //     {1.0f, 0.0f, 0.0f},
-            //     {-2.0f / 9, -2.0f / 9, -2.0f / 9},
-            //     {-2.0f / 9, 2.0f / 9, -2.0f / 9},
-            //     {1.0f / 90, 1.0f / 45, 2.0f / 45},
-            //     {1.0f / 90, -1.0f / 45, 2.0f / 45},
-            //     {1.0f / 45, 1.0f / 90, 1.0f / 180},
-            //     {1.0f / 45, -1.0f / 90, 1.0f / 180},
-            //     {0.0f, 0.0f, 1.0f}
-            // };
-            const float ktm0 = 2.0f / 9;
-            const float ktm1 = 1.0f / 45;
-            const float ktm2 = 2.0f / 45;
-            const float ktm3 = 1.0f / 90;
-            const float ktm4 = 1.0f / 180;
-
-            float tmp[8][3];
-
-            const float* k0 = (const float*)kernel + (i + ii) * inch * 9 + (k + kk) * 9;
-
-            for (int m = 0; m < 3; m++)
-            {
-                float r0 = k0[0];
-                float r1 = k0[1];
-                float r2 = k0[2];
-
-                tmp[0][m] = r0;
-                tmp[1][m] = -r0 * ktm0 - r1 * ktm0 - r2 * ktm0;
-                tmp[2][m] = -r0 * ktm0 + r1 * ktm0 - r2 * ktm0;
-                tmp[3][m] = r0 * ktm3 + r1 * ktm1 + r2 * ktm2;
-                tmp[4][m] = r0 * ktm3 - r1 * ktm1 + r2 * ktm2;
-                tmp[5][m] = r0 * ktm1 + r1 * ktm3 + r2 * ktm4;
-                tmp[6][m] = r0 * ktm1 - r1 * ktm3 + r2 * ktm4;
-                tmp[7][m] = r2;
-
-                k0 += 3;
-            }
-
-            for (int m = 0; m < 8; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-
-                float z0 = r0;
-                float z1 = -r0 * ktm0 - r1 * ktm0 - r2 * ktm0;
-                float z2 = -r0 * ktm0 + r1 * ktm0 - r2 * ktm0;
-                float z3 = r0 * ktm3 + r1 * ktm1 + r2 * ktm2;
-                float z4 = r0 * ktm3 - r1 * ktm1 + r2 * ktm2;
-                float z5 = r0 * ktm1 + r1 * ktm3 + r2 * ktm4;
-                float z6 = r0 * ktm1 - r1 * ktm3 + r2 * ktm4;
-                float z7 = r2;
-
-                ptmp[0] = z0;
-                ptmp[1] = z1;
-                ptmp[2] = z2;
-                ptmp[3] = z3;
-                ptmp[4] = z4;
-                ptmp[5] = z5;
-                ptmp[6] = z6;
-                ptmp[7] = z7;
-                ptmp += 8;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd63_transform_kernel(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt)
-{
-    const int M = outch;
-    const int K = inch;
-    const int B = 64;
-
-    int TILE_M, TILE_N, TILE_K;
-    get_optimal_tile_mnk(M, 0, K, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    Mat A_tileX(B * TILE_M * TILE_K, 1, opt.num_threads, 4u, (Allocator*)0);
-
-    AT.create(TILE_K * TILE_M, B, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M, 4u, (Allocator*)0);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat A_tile = A_tileX.channel(get_omp_thread_num());
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_ii = std::min((M - i), TILE_M);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            conv3x3s1_winograd63_transform_kernel_tile(kernel, A_tile, inch, i, max_ii, k, max_kk);
-
-            Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-            pack_A_tile(A_tile, AT_tile, B, max_ii, max_kk);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd63_transform_input_tile(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    // const float itm[8][8] = {
-    //     {1.0f, 0.0f,-5.25f, 0.00f, 5.25f, 0.00f,-1.0f, 0.0f},
-    //     {0.0f, 1.0f, 1.00f,-4.25f,-4.25f, 1.00f, 1.0f, 0.0f},
-    //     {0.0f,-1.0f, 1.00f, 4.25f,-4.25f,-1.00f, 1.0f, 0.0f},
-    //     {0.0f, 0.5f, 0.25f,-2.50f,-1.25f, 2.00f, 1.0f, 0.0f},
-    //     {0.0f,-0.5f, 0.25f, 2.50f,-1.25f,-2.00f, 1.0f, 0.0f},
-    //     {0.0f, 2.0f, 4.00f,-2.50f,-5.00f, 0.50f, 1.0f, 0.0f},
-    //     {0.0f,-2.0f, 4.00f, 2.50f,-5.00f,-0.50f, 1.0f, 0.0f},
-    //     {0.0f,-1.0f, 0.00f, 5.25f, 0.00f,-5.25f, 0.0f, 1.0f}
-    // };
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w + 3) / 6;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    nn_max_kk = max_kk / 16;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = ppkk * 16;
-
-#ifdef _MSC_VER
-        __declspec(align(64))
-#else
-        __attribute__((aligned(64)))
-#endif
-        float tmp[8][8][16];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel((k + kk) / elempack).row(ti * 6) + (tj * 6) * elempack;
-
-            for (int m = 0; m < 8; m++)
-            {
-                __m512 _r0 = _mm512_setzero_ps();
-                __m512 _r1 = _mm512_setzero_ps();
-                __m512 _r2 = _mm512_setzero_ps();
-                __m512 _r3 = _mm512_setzero_ps();
-                __m512 _r4 = _mm512_setzero_ps();
-                __m512 _r5 = _mm512_setzero_ps();
-                __m512 _r6 = _mm512_setzero_ps();
-                __m512 _r7 = _mm512_setzero_ps();
-
-                if (ti * 6 + m < h)
-                {
-                    if (elempack == 16)
-                    {
-                        _r0 = _mm512_load_ps(r0);
-                        if (tj * 6 + 1 < w) _r1 = _mm512_load_ps(r0 + 16);
-                        if (tj * 6 + 2 < w) _r2 = _mm512_load_ps(r0 + 32);
-                        if (tj * 6 + 3 < w) _r3 = _mm512_load_ps(r0 + 48);
-                        if (tj * 6 + 4 < w) _r4 = _mm512_load_ps(r0 + 64);
-                        if (tj * 6 + 5 < w) _r5 = _mm512_load_ps(r0 + 80);
-                        if (tj * 6 + 6 < w) _r6 = _mm512_load_ps(r0 + 96);
-                        if (tj * 6 + 7 < w) _r7 = _mm512_load_ps(r0 + 112);
-                    }
-                    if (elempack == 8)
-                    {
-                        const float* r1 = r0 + N;
-
-                        _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0)), _mm256_load_ps(r1), 1);
-                        if (tj * 6 + 1 < w) _r1 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + 8)), _mm256_load_ps(r1 + 8), 1);
-                        if (tj * 6 + 2 < w) _r2 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + 16)), _mm256_load_ps(r1 + 16), 1);
-                        if (tj * 6 + 3 < w) _r3 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + 24)), _mm256_load_ps(r1 + 24), 1);
-                        if (tj * 6 + 4 < w) _r4 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + 32)), _mm256_load_ps(r1 + 32), 1);
-                        if (tj * 6 + 5 < w) _r5 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + 40)), _mm256_load_ps(r1 + 40), 1);
-                        if (tj * 6 + 6 < w) _r6 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + 48)), _mm256_load_ps(r1 + 48), 1);
-                        if (tj * 6 + 7 < w) _r7 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + 56)), _mm256_load_ps(r1 + 56), 1);
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-
-                        _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0)), _mm_load_ps(r1), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2)), _mm_load_ps(r3), 1), 1);
-                        if (tj * 6 + 1 < w) _r1 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 4)), _mm_load_ps(r1 + 4), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + 4)), _mm_load_ps(r3 + 4), 1), 1);
-                        if (tj * 6 + 2 < w) _r2 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 8)), _mm_load_ps(r1 + 8), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + 8)), _mm_load_ps(r3 + 8), 1), 1);
-                        if (tj * 6 + 3 < w) _r3 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 12)), _mm_load_ps(r1 + 12), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + 12)), _mm_load_ps(r3 + 12), 1), 1);
-                        if (tj * 6 + 4 < w) _r4 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 16)), _mm_load_ps(r1 + 16), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + 16)), _mm_load_ps(r3 + 16), 1), 1);
-                        if (tj * 6 + 5 < w) _r5 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 20)), _mm_load_ps(r1 + 20), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + 20)), _mm_load_ps(r3 + 20), 1), 1);
-                        if (tj * 6 + 6 < w) _r6 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 24)), _mm_load_ps(r1 + 24), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + 24)), _mm_load_ps(r3 + 24), 1), 1);
-                        if (tj * 6 + 7 < w) _r7 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 28)), _mm_load_ps(r1 + 28), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + 28)), _mm_load_ps(r3 + 28), 1), 1);
-                    }
-                    if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-                        const float* r4 = r0 + N * 4;
-                        const float* r5 = r0 + N * 5;
-                        const float* r6 = r0 + N * 6;
-                        const float* r7 = r0 + N * 7;
-                        const float* r8 = r0 + N * 8;
-                        const float* r9 = r0 + N * 9;
-                        const float* ra = r0 + N * 10;
-                        const float* rb = r0 + N * 11;
-                        const float* rc = r0 + N * 12;
-                        const float* rd = r0 + N * 13;
-                        const float* re = r0 + N * 14;
-                        const float* rf = r0 + N * 15;
-
-                        __m128 _t0 = _mm_loadu_ps(r0);
-                        __m128 _t1 = _mm_loadu_ps(r1);
-                        __m128 _t2 = _mm_loadu_ps(r2);
-                        __m128 _t3 = _mm_loadu_ps(r3);
-                        __m128 _t4 = _mm_loadu_ps(r4);
-                        __m128 _t5 = _mm_loadu_ps(r5);
-                        __m128 _t6 = _mm_loadu_ps(r6);
-                        __m128 _t7 = _mm_loadu_ps(r7);
-                        __m128 _t8 = _mm_loadu_ps(r8);
-                        __m128 _t9 = _mm_loadu_ps(r9);
-                        __m128 _ta = _mm_loadu_ps(ra);
-                        __m128 _tb = _mm_loadu_ps(rb);
-                        __m128 _tc = _mm_loadu_ps(rc);
-                        __m128 _td = _mm_loadu_ps(rd);
-                        __m128 _te = _mm_loadu_ps(re);
-                        __m128 _tf = _mm_loadu_ps(rf);
-
-                        _MM_TRANSPOSE4_PS(_t0, _t1, _t2, _t3);
-                        _MM_TRANSPOSE4_PS(_t4, _t5, _t6, _t7);
-                        _MM_TRANSPOSE4_PS(_t8, _t9, _ta, _tb);
-                        _MM_TRANSPOSE4_PS(_tc, _td, _te, _tf);
-
-                        _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t0), _t4, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_t8), _tc, 1), 1);
-                        if (tj * 6 + 1 < w) _r1 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t1), _t5, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_t9), _td, 1), 1);
-                        if (tj * 6 + 2 < w) _r2 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t2), _t6, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_ta), _te, 1), 1);
-                        if (tj * 6 + 3 < w) _r3 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t3), _t7, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_tb), _tf, 1), 1);
-                        if (tj * 6 + 4 < w)
-                        {
-                            _t0 = _mm_loadu_ps(r0 + 4);
-                            _t1 = _mm_loadu_ps(r1 + 4);
-                            _t2 = _mm_loadu_ps(r2 + 4);
-                            _t3 = _mm_loadu_ps(r3 + 4);
-                            _t4 = _mm_loadu_ps(r4 + 4);
-                            _t5 = _mm_loadu_ps(r5 + 4);
-                            _t6 = _mm_loadu_ps(r6 + 4);
-                            _t7 = _mm_loadu_ps(r7 + 4);
-                            _t8 = _mm_loadu_ps(r8 + 4);
-                            _t9 = _mm_loadu_ps(r9 + 4);
-                            _ta = _mm_loadu_ps(ra + 4);
-                            _tb = _mm_loadu_ps(rb + 4);
-                            _tc = _mm_loadu_ps(rc + 4);
-                            _td = _mm_loadu_ps(rd + 4);
-                            _te = _mm_loadu_ps(re + 4);
-                            _tf = _mm_loadu_ps(rf + 4);
-
-                            _MM_TRANSPOSE4_PS(_t0, _t1, _t2, _t3);
-                            _MM_TRANSPOSE4_PS(_t4, _t5, _t6, _t7);
-                            _MM_TRANSPOSE4_PS(_t8, _t9, _ta, _tb);
-                            _MM_TRANSPOSE4_PS(_tc, _td, _te, _tf);
-
-                            _r4 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t0), _t4, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_t8), _tc, 1), 1);
-                            if (tj * 6 + 5 < w) _r5 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t1), _t5, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_t9), _td, 1), 1);
-                            if (tj * 6 + 6 < w) _r6 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t2), _t6, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_ta), _te, 1), 1);
-                            if (tj * 6 + 7 < w) _r7 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_t3), _t7, 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_tb), _tf, 1), 1);
-                        }
-                    }
-                }
-
-                __m512 _v5_25 = _mm512_set1_ps(5.25f);
-                __m512 _vm4_25 = _mm512_set1_ps(-4.25f);
-                __m512 _vm1_25 = _mm512_set1_ps(-1.25f);
-                __m512 _v0_25 = _mm512_set1_ps(0.25f);
-                __m512 _vm2_5 = _mm512_set1_ps(-2.5f);
-                __m512 _v0_5 = _mm512_set1_ps(0.5f);
-                __m512 _v2 = _mm512_set1_ps(2.f);
-                __m512 _v4 = _mm512_set1_ps(4.f);
-
-                __m512 _tmp12a = _mm512_fmadd_ps(_vm4_25, _r4, _mm512_add_ps(_r2, _r6));
-                __m512 _tmp12b = _mm512_fmadd_ps(_vm4_25, _r3, _mm512_add_ps(_r1, _r5));
-                __m512 _tmp34a = _mm512_fmadd_ps(_vm1_25, _r4, _mm512_fmadd_ps(_v0_25, _r2, _r6));
-                __m512 _tmp34b = _mm512_fmadd_ps(_v2, _r5, _mm512_fmadd_ps(_vm2_5, _r3, _mm512_mul_ps(_r1, _v0_5)));
-                __m512 _tmp56a = _mm512_fmadd_ps(_v4, _mm512_fmadd_ps(_vm1_25, _r4, _r2), _r6);
-                __m512 _tmp56b = _mm512_fmadd_ps(_v0_5, _r5, _mm512_fmadd_ps(_vm2_5, _r3, _mm512_mul_ps(_r1, _v2)));
-
-                __m512 _tmp0 = _mm512_fmadd_ps(_v5_25, _mm512_sub_ps(_r4, _r2), _mm512_sub_ps(_r0, _r6));
-                __m512 _tmp1 = _mm512_add_ps(_tmp12a, _tmp12b);
-                __m512 _tmp2 = _mm512_sub_ps(_tmp12a, _tmp12b);
-                __m512 _tmp3 = _mm512_add_ps(_tmp34a, _tmp34b);
-                __m512 _tmp4 = _mm512_sub_ps(_tmp34a, _tmp34b);
-                __m512 _tmp5 = _mm512_add_ps(_tmp56a, _tmp56b);
-                __m512 _tmp6 = _mm512_sub_ps(_tmp56a, _tmp56b);
-                __m512 _tmp7 = _mm512_fmadd_ps(_v5_25, _mm512_sub_ps(_r3, _r5), _mm512_sub_ps(_r7, _r1));
-
-                _mm512_store_ps(tmp[0][m], _tmp0);
-                _mm512_store_ps(tmp[1][m], _tmp1);
-                _mm512_store_ps(tmp[2][m], _tmp2);
-                _mm512_store_ps(tmp[3][m], _tmp3);
-                _mm512_store_ps(tmp[4][m], _tmp4);
-                _mm512_store_ps(tmp[5][m], _tmp5);
-                _mm512_store_ps(tmp[6][m], _tmp6);
-                _mm512_store_ps(tmp[7][m], _tmp7);
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 64 + jj * 16;
-            float* p1 = p0 + max_jj * 16;
-            float* p2 = p0 + max_jj * 16 * 2;
-            float* p3 = p0 + max_jj * 16 * 3;
-            float* p4 = p0 + max_jj * 16 * 4;
-            float* p5 = p0 + max_jj * 16 * 5;
-            float* p6 = p0 + max_jj * 16 * 6;
-            float* p7 = p0 + max_jj * 16 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                __m512 _r0 = _mm512_load_ps(tmp[m][0]);
-                __m512 _r1 = _mm512_load_ps(tmp[m][1]);
-                __m512 _r2 = _mm512_load_ps(tmp[m][2]);
-                __m512 _r3 = _mm512_load_ps(tmp[m][3]);
-                __m512 _r4 = _mm512_load_ps(tmp[m][4]);
-                __m512 _r5 = _mm512_load_ps(tmp[m][5]);
-                __m512 _r6 = _mm512_load_ps(tmp[m][6]);
-                __m512 _r7 = _mm512_load_ps(tmp[m][7]);
-
-                __m512 _v5_25 = _mm512_set1_ps(5.25f);
-                __m512 _vm4_25 = _mm512_set1_ps(-4.25f);
-                __m512 _vm1_25 = _mm512_set1_ps(-1.25f);
-                __m512 _v0_25 = _mm512_set1_ps(0.25f);
-                __m512 _vm2_5 = _mm512_set1_ps(-2.5f);
-                __m512 _v0_5 = _mm512_set1_ps(0.5f);
-                __m512 _v2 = _mm512_set1_ps(2.f);
-                __m512 _v4 = _mm512_set1_ps(4.f);
-
-                __m512 _tmp12a = _mm512_fmadd_ps(_vm4_25, _r4, _mm512_add_ps(_r2, _r6));
-                __m512 _tmp12b = _mm512_fmadd_ps(_vm4_25, _r3, _mm512_add_ps(_r1, _r5));
-                __m512 _tmp34a = _mm512_fmadd_ps(_vm1_25, _r4, _mm512_fmadd_ps(_v0_25, _r2, _r6));
-                __m512 _tmp34b = _mm512_fmadd_ps(_v2, _r5, _mm512_fmadd_ps(_vm2_5, _r3, _mm512_mul_ps(_r1, _v0_5)));
-                __m512 _tmp56a = _mm512_fmadd_ps(_v4, _mm512_fmadd_ps(_vm1_25, _r4, _r2), _r6);
-                __m512 _tmp56b = _mm512_fmadd_ps(_v0_5, _r5, _mm512_fmadd_ps(_vm2_5, _r3, _mm512_mul_ps(_r1, _v2)));
-
-                __m512 _tmp0 = _mm512_fmadd_ps(_v5_25, _mm512_sub_ps(_r4, _r2), _mm512_sub_ps(_r0, _r6));
-                __m512 _tmp1 = _mm512_add_ps(_tmp12a, _tmp12b);
-                __m512 _tmp2 = _mm512_sub_ps(_tmp12a, _tmp12b);
-                __m512 _tmp3 = _mm512_add_ps(_tmp34a, _tmp34b);
-                __m512 _tmp4 = _mm512_sub_ps(_tmp34a, _tmp34b);
-                __m512 _tmp5 = _mm512_add_ps(_tmp56a, _tmp56b);
-                __m512 _tmp6 = _mm512_sub_ps(_tmp56a, _tmp56b);
-                __m512 _tmp7 = _mm512_fmadd_ps(_v5_25, _mm512_sub_ps(_r3, _r5), _mm512_sub_ps(_r7, _r1));
-
-                _mm512_store_ps(p0, _tmp0);
-                _mm512_store_ps(p1, _tmp1);
-                _mm512_store_ps(p2, _tmp2);
-                _mm512_store_ps(p3, _tmp3);
-                _mm512_store_ps(p4, _tmp4);
-                _mm512_store_ps(p5, _tmp5);
-                _mm512_store_ps(p6, _tmp6);
-                _mm512_store_ps(p7, _tmp7);
-
-                p0 += max_jj * 8 * 16;
-                p1 += max_jj * 8 * 16;
-                p2 += max_jj * 8 * 16;
-                p3 += max_jj * 8 * 16;
-                p4 += max_jj * 8 * 16;
-                p5 += max_jj * 8 * 16;
-                p6 += max_jj * 8 * 16;
-                p7 += max_jj * 8 * 16;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 16;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-#else // __AVX512F__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-    #pragma omp parallel for num_threads(nT)
-#endif // __AVX512F__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-#ifdef _MSC_VER
-        __declspec(align(32))
-#else
-        __attribute__((aligned(32)))
-#endif
-        float tmp[8][8][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel((k + kk) / elempack).row(ti * 6) + (tj * 6) * elempack;
-
-            for (int m = 0; m < 8; m++)
-            {
-                __m256 _r0 = _mm256_setzero_ps();
-                __m256 _r1 = _mm256_setzero_ps();
-                __m256 _r2 = _mm256_setzero_ps();
-                __m256 _r3 = _mm256_setzero_ps();
-                __m256 _r4 = _mm256_setzero_ps();
-                __m256 _r5 = _mm256_setzero_ps();
-                __m256 _r6 = _mm256_setzero_ps();
-                __m256 _r7 = _mm256_setzero_ps();
-
-                if (ti * 6 + m < h)
-                {
-                    if (elempack == 8)
-                    {
-                        _r0 = _mm256_load_ps(r0);
-                        if (tj * 6 + 1 < w) _r1 = _mm256_load_ps(r0 + 8);
-                        if (tj * 6 + 2 < w) _r2 = _mm256_load_ps(r0 + 16);
-                        if (tj * 6 + 3 < w) _r3 = _mm256_load_ps(r0 + 24);
-                        if (tj * 6 + 4 < w) _r4 = _mm256_load_ps(r0 + 32);
-                        if (tj * 6 + 5 < w) _r5 = _mm256_load_ps(r0 + 40);
-                        if (tj * 6 + 6 < w) _r6 = _mm256_load_ps(r0 + 48);
-                        if (tj * 6 + 7 < w) _r7 = _mm256_load_ps(r0 + 56);
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-
-                        _r0 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0)), _mm_load_ps(r1), 1);
-                        if (tj * 6 + 1 < w) _r1 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 4)), _mm_load_ps(r1 + 4), 1);
-                        if (tj * 6 + 2 < w) _r2 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 8)), _mm_load_ps(r1 + 8), 1);
-                        if (tj * 6 + 3 < w) _r3 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 12)), _mm_load_ps(r1 + 12), 1);
-                        if (tj * 6 + 4 < w) _r4 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 16)), _mm_load_ps(r1 + 16), 1);
-                        if (tj * 6 + 5 < w) _r5 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 20)), _mm_load_ps(r1 + 20), 1);
-                        if (tj * 6 + 6 < w) _r6 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 24)), _mm_load_ps(r1 + 24), 1);
-                        if (tj * 6 + 7 < w) _r7 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + 28)), _mm_load_ps(r1 + 28), 1);
-                    }
-                    if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-                        const float* r4 = r0 + N * 4;
-                        const float* r5 = r0 + N * 5;
-                        const float* r6 = r0 + N * 6;
-                        const float* r7 = r0 + N * 7;
-
-                        __m128 _t0 = _mm_loadu_ps(r0);
-                        __m128 _t1 = _mm_loadu_ps(r1);
-                        __m128 _t2 = _mm_loadu_ps(r2);
-                        __m128 _t3 = _mm_loadu_ps(r3);
-                        __m128 _t4 = _mm_loadu_ps(r4);
-                        __m128 _t5 = _mm_loadu_ps(r5);
-                        __m128 _t6 = _mm_loadu_ps(r6);
-                        __m128 _t7 = _mm_loadu_ps(r7);
-
-                        _MM_TRANSPOSE4_PS(_t0, _t1, _t2, _t3);
-                        _MM_TRANSPOSE4_PS(_t4, _t5, _t6, _t7);
-
-                        _r0 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t0), _t4, 1);
-                        if (tj * 6 + 1 < w) _r1 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t1), _t5, 1);
-                        if (tj * 6 + 2 < w) _r2 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t2), _t6, 1);
-                        if (tj * 6 + 3 < w) _r3 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t3), _t7, 1);
-                        if (tj * 6 + 4 < w)
-                        {
-                            _t0 = _mm_loadu_ps(r0 + 4);
-                            _t1 = _mm_loadu_ps(r1 + 4);
-                            _t2 = _mm_loadu_ps(r2 + 4);
-                            _t3 = _mm_loadu_ps(r3 + 4);
-                            _t4 = _mm_loadu_ps(r4 + 4);
-                            _t5 = _mm_loadu_ps(r5 + 4);
-                            _t6 = _mm_loadu_ps(r6 + 4);
-                            _t7 = _mm_loadu_ps(r7 + 4);
-
-                            _MM_TRANSPOSE4_PS(_t0, _t1, _t2, _t3);
-                            _MM_TRANSPOSE4_PS(_t4, _t5, _t6, _t7);
-
-                            _r4 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t0), _t4, 1);
-                            if (tj * 6 + 5 < w) _r5 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t1), _t5, 1);
-                            if (tj * 6 + 6 < w) _r6 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t2), _t6, 1);
-                            if (tj * 6 + 7 < w) _r7 = _mm256_insertf128_ps(_mm256_castps128_ps256(_t3), _t7, 1);
-                        }
-                    }
-                }
-
-                __m256 _v5_25 = _mm256_set1_ps(5.25f);
-                __m256 _vm4_25 = _mm256_set1_ps(-4.25f);
-                __m256 _vm1_25 = _mm256_set1_ps(-1.25f);
-                __m256 _v0_25 = _mm256_set1_ps(0.25f);
-                __m256 _vm2_5 = _mm256_set1_ps(-2.5f);
-                __m256 _v0_5 = _mm256_set1_ps(0.5f);
-                __m256 _v2 = _mm256_set1_ps(2.f);
-                __m256 _v4 = _mm256_set1_ps(4.f);
-
-                __m256 _tmp12a = _mm256_comp_fmadd_ps(_vm4_25, _r4, _mm256_add_ps(_r2, _r6));
-                __m256 _tmp12b = _mm256_comp_fmadd_ps(_vm4_25, _r3, _mm256_add_ps(_r1, _r5));
-                __m256 _tmp34a = _mm256_comp_fmadd_ps(_vm1_25, _r4, _mm256_comp_fmadd_ps(_v0_25, _r2, _r6));
-                __m256 _tmp34b = _mm256_comp_fmadd_ps(_v2, _r5, _mm256_comp_fmadd_ps(_vm2_5, _r3, _mm256_mul_ps(_r1, _v0_5)));
-                __m256 _tmp56a = _mm256_comp_fmadd_ps(_v4, _mm256_comp_fmadd_ps(_vm1_25, _r4, _r2), _r6);
-                __m256 _tmp56b = _mm256_comp_fmadd_ps(_v0_5, _r5, _mm256_comp_fmadd_ps(_vm2_5, _r3, _mm256_mul_ps(_r1, _v2)));
-
-                __m256 _tmp0 = _mm256_comp_fmadd_ps(_v5_25, _mm256_sub_ps(_r4, _r2), _mm256_sub_ps(_r0, _r6));
-                __m256 _tmp1 = _mm256_add_ps(_tmp12a, _tmp12b);
-                __m256 _tmp2 = _mm256_sub_ps(_tmp12a, _tmp12b);
-                __m256 _tmp3 = _mm256_add_ps(_tmp34a, _tmp34b);
-                __m256 _tmp4 = _mm256_sub_ps(_tmp34a, _tmp34b);
-                __m256 _tmp5 = _mm256_add_ps(_tmp56a, _tmp56b);
-                __m256 _tmp6 = _mm256_sub_ps(_tmp56a, _tmp56b);
-                __m256 _tmp7 = _mm256_comp_fmadd_ps(_v5_25, _mm256_sub_ps(_r3, _r5), _mm256_sub_ps(_r7, _r1));
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm256_storeu_ps(tmp[0][m], _tmp0);
-                _mm256_storeu_ps(tmp[1][m], _tmp1);
-                _mm256_storeu_ps(tmp[2][m], _tmp2);
-                _mm256_storeu_ps(tmp[3][m], _tmp3);
-                _mm256_storeu_ps(tmp[4][m], _tmp4);
-                _mm256_storeu_ps(tmp[5][m], _tmp5);
-                _mm256_storeu_ps(tmp[6][m], _tmp6);
-                _mm256_storeu_ps(tmp[7][m], _tmp7);
-#else
-                _mm256_store_ps(tmp[0][m], _tmp0);
-                _mm256_store_ps(tmp[1][m], _tmp1);
-                _mm256_store_ps(tmp[2][m], _tmp2);
-                _mm256_store_ps(tmp[3][m], _tmp3);
-                _mm256_store_ps(tmp[4][m], _tmp4);
-                _mm256_store_ps(tmp[5][m], _tmp5);
-                _mm256_store_ps(tmp[6][m], _tmp6);
-                _mm256_store_ps(tmp[7][m], _tmp7);
-#endif
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 64 + jj * 8;
-            float* p1 = p0 + max_jj * 8;
-            float* p2 = p0 + max_jj * 8 * 2;
-            float* p3 = p0 + max_jj * 8 * 3;
-            float* p4 = p0 + max_jj * 8 * 4;
-            float* p5 = p0 + max_jj * 8 * 5;
-            float* p6 = p0 + max_jj * 8 * 6;
-            float* p7 = p0 + max_jj * 8 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m256 _r0 = _mm256_loadu_ps(tmp[m][0]);
-                __m256 _r1 = _mm256_loadu_ps(tmp[m][1]);
-                __m256 _r2 = _mm256_loadu_ps(tmp[m][2]);
-                __m256 _r3 = _mm256_loadu_ps(tmp[m][3]);
-                __m256 _r4 = _mm256_loadu_ps(tmp[m][4]);
-                __m256 _r5 = _mm256_loadu_ps(tmp[m][5]);
-                __m256 _r6 = _mm256_loadu_ps(tmp[m][6]);
-                __m256 _r7 = _mm256_loadu_ps(tmp[m][7]);
-#else
-                __m256 _r0 = _mm256_load_ps(tmp[m][0]);
-                __m256 _r1 = _mm256_load_ps(tmp[m][1]);
-                __m256 _r2 = _mm256_load_ps(tmp[m][2]);
-                __m256 _r3 = _mm256_load_ps(tmp[m][3]);
-                __m256 _r4 = _mm256_load_ps(tmp[m][4]);
-                __m256 _r5 = _mm256_load_ps(tmp[m][5]);
-                __m256 _r6 = _mm256_load_ps(tmp[m][6]);
-                __m256 _r7 = _mm256_load_ps(tmp[m][7]);
-#endif
-
-                __m256 _v5_25 = _mm256_set1_ps(5.25f);
-                __m256 _vm4_25 = _mm256_set1_ps(-4.25f);
-                __m256 _vm1_25 = _mm256_set1_ps(-1.25f);
-                __m256 _v0_25 = _mm256_set1_ps(0.25f);
-                __m256 _vm2_5 = _mm256_set1_ps(-2.5f);
-                __m256 _v0_5 = _mm256_set1_ps(0.5f);
-                __m256 _v2 = _mm256_set1_ps(2.f);
-                __m256 _v4 = _mm256_set1_ps(4.f);
-
-                __m256 _tmp12a = _mm256_comp_fmadd_ps(_vm4_25, _r4, _mm256_add_ps(_r2, _r6));
-                __m256 _tmp12b = _mm256_comp_fmadd_ps(_vm4_25, _r3, _mm256_add_ps(_r1, _r5));
-                __m256 _tmp34a = _mm256_comp_fmadd_ps(_vm1_25, _r4, _mm256_comp_fmadd_ps(_v0_25, _r2, _r6));
-                __m256 _tmp34b = _mm256_comp_fmadd_ps(_v2, _r5, _mm256_comp_fmadd_ps(_vm2_5, _r3, _mm256_mul_ps(_r1, _v0_5)));
-                __m256 _tmp56a = _mm256_comp_fmadd_ps(_v4, _mm256_comp_fmadd_ps(_vm1_25, _r4, _r2), _r6);
-                __m256 _tmp56b = _mm256_comp_fmadd_ps(_v0_5, _r5, _mm256_comp_fmadd_ps(_vm2_5, _r3, _mm256_mul_ps(_r1, _v2)));
-
-                __m256 _tmp0 = _mm256_comp_fmadd_ps(_v5_25, _mm256_sub_ps(_r4, _r2), _mm256_sub_ps(_r0, _r6));
-                __m256 _tmp1 = _mm256_add_ps(_tmp12a, _tmp12b);
-                __m256 _tmp2 = _mm256_sub_ps(_tmp12a, _tmp12b);
-                __m256 _tmp3 = _mm256_add_ps(_tmp34a, _tmp34b);
-                __m256 _tmp4 = _mm256_sub_ps(_tmp34a, _tmp34b);
-                __m256 _tmp5 = _mm256_add_ps(_tmp56a, _tmp56b);
-                __m256 _tmp6 = _mm256_sub_ps(_tmp56a, _tmp56b);
-                __m256 _tmp7 = _mm256_comp_fmadd_ps(_v5_25, _mm256_sub_ps(_r3, _r5), _mm256_sub_ps(_r7, _r1));
-
-                _mm256_store_ps(p0, _tmp0);
-                _mm256_store_ps(p1, _tmp1);
-                _mm256_store_ps(p2, _tmp2);
-                _mm256_store_ps(p3, _tmp3);
-                _mm256_store_ps(p4, _tmp4);
-                _mm256_store_ps(p5, _tmp5);
-                _mm256_store_ps(p6, _tmp6);
-                _mm256_store_ps(p7, _tmp7);
-
-                p0 += max_jj * 8 * 8;
-                p1 += max_jj * 8 * 8;
-                p2 += max_jj * 8 * 8;
-                p3 += max_jj * 8 * 8;
-                p4 += max_jj * 8 * 8;
-                p5 += max_jj * 8 * 8;
-                p6 += max_jj * 8 * 8;
-                p7 += max_jj * 8 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-#else // __AVX__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 4;
-    #pragma omp parallel for num_threads(nT)
-#endif // __AVX__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 4;
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[8][8][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel((k + kk) / elempack).row(ti * 6) + (tj * 6) * elempack;
-
-            for (int m = 0; m < 8; m++)
-            {
-                __m128 _r0 = _mm_setzero_ps();
-                __m128 _r1 = _mm_setzero_ps();
-                __m128 _r2 = _mm_setzero_ps();
-                __m128 _r3 = _mm_setzero_ps();
-                __m128 _r4 = _mm_setzero_ps();
-                __m128 _r5 = _mm_setzero_ps();
-                __m128 _r6 = _mm_setzero_ps();
-                __m128 _r7 = _mm_setzero_ps();
-
-                if (ti * 6 + m < h)
-                {
-                    if (elempack == 4)
-                    {
-                        _r0 = _mm_load_ps(r0);
-                        if (tj * 6 + 1 < w) _r1 = _mm_load_ps(r0 + 4);
-                        if (tj * 6 + 2 < w) _r2 = _mm_load_ps(r0 + 8);
-                        if (tj * 6 + 3 < w) _r3 = _mm_load_ps(r0 + 12);
-                        if (tj * 6 + 4 < w) _r4 = _mm_load_ps(r0 + 16);
-                        if (tj * 6 + 5 < w) _r5 = _mm_load_ps(r0 + 20);
-                        if (tj * 6 + 6 < w) _r6 = _mm_load_ps(r0 + 24);
-                        if (tj * 6 + 7 < w) _r7 = _mm_load_ps(r0 + 28);
-                    }
-                    if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-
-                        __m128 _t0 = _mm_loadu_ps(r0);
-                        __m128 _t1 = _mm_loadu_ps(r1);
-                        __m128 _t2 = _mm_loadu_ps(r2);
-                        __m128 _t3 = _mm_loadu_ps(r3);
-
-                        _MM_TRANSPOSE4_PS(_t0, _t1, _t2, _t3);
-
-                        _r0 = _t0;
-                        if (tj * 6 + 1 < w) _r1 = _t1;
-                        if (tj * 6 + 2 < w) _r2 = _t2;
-                        if (tj * 6 + 3 < w) _r3 = _t3;
-                        if (tj * 6 + 4 < w)
-                        {
-                            _t0 = _mm_loadu_ps(r0 + 4);
-                            _t1 = _mm_loadu_ps(r1 + 4);
-                            _t2 = _mm_loadu_ps(r2 + 4);
-                            _t3 = _mm_loadu_ps(r3 + 4);
-
-                            _MM_TRANSPOSE4_PS(_t0, _t1, _t2, _t3);
-
-                            _r4 = _t0;
-                            if (tj * 6 + 5 < w) _r5 = _t1;
-                            if (tj * 6 + 6 < w) _r6 = _t2;
-                            if (tj * 6 + 7 < w) _r7 = _t3;
-                        }
-                    }
-                }
-
-                __m128 _v5_25 = _mm_set1_ps(5.25f);
-                __m128 _vm4_25 = _mm_set1_ps(-4.25f);
-                __m128 _vm1_25 = _mm_set1_ps(-1.25f);
-                __m128 _v0_25 = _mm_set1_ps(0.25f);
-                __m128 _vm2_5 = _mm_set1_ps(-2.5f);
-                __m128 _v0_5 = _mm_set1_ps(0.5f);
-                __m128 _v2 = _mm_set1_ps(2.f);
-                __m128 _v4 = _mm_set1_ps(4.f);
-
-                __m128 _tmp12a = _mm_comp_fmadd_ps(_vm4_25, _r4, _mm_add_ps(_r2, _r6));
-                __m128 _tmp12b = _mm_comp_fmadd_ps(_vm4_25, _r3, _mm_add_ps(_r1, _r5));
-                __m128 _tmp34a = _mm_comp_fmadd_ps(_vm1_25, _r4, _mm_comp_fmadd_ps(_v0_25, _r2, _r6));
-                __m128 _tmp34b = _mm_comp_fmadd_ps(_v2, _r5, _mm_comp_fmadd_ps(_vm2_5, _r3, _mm_mul_ps(_r1, _v0_5)));
-                __m128 _tmp56a = _mm_comp_fmadd_ps(_v4, _mm_comp_fmadd_ps(_vm1_25, _r4, _r2), _r6);
-                __m128 _tmp56b = _mm_comp_fmadd_ps(_v0_5, _r5, _mm_comp_fmadd_ps(_vm2_5, _r3, _mm_mul_ps(_r1, _v2)));
-
-                __m128 _tmp0 = _mm_comp_fmadd_ps(_v5_25, _mm_sub_ps(_r4, _r2), _mm_sub_ps(_r0, _r6));
-                __m128 _tmp1 = _mm_add_ps(_tmp12a, _tmp12b);
-                __m128 _tmp2 = _mm_sub_ps(_tmp12a, _tmp12b);
-                __m128 _tmp3 = _mm_add_ps(_tmp34a, _tmp34b);
-                __m128 _tmp4 = _mm_sub_ps(_tmp34a, _tmp34b);
-                __m128 _tmp5 = _mm_add_ps(_tmp56a, _tmp56b);
-                __m128 _tmp6 = _mm_sub_ps(_tmp56a, _tmp56b);
-                __m128 _tmp7 = _mm_comp_fmadd_ps(_v5_25, _mm_sub_ps(_r3, _r5), _mm_sub_ps(_r7, _r1));
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm_storeu_ps(tmp[0][m], _tmp0);
-                _mm_storeu_ps(tmp[1][m], _tmp1);
-                _mm_storeu_ps(tmp[2][m], _tmp2);
-                _mm_storeu_ps(tmp[3][m], _tmp3);
-                _mm_storeu_ps(tmp[4][m], _tmp4);
-                _mm_storeu_ps(tmp[5][m], _tmp5);
-                _mm_storeu_ps(tmp[6][m], _tmp6);
-                _mm_storeu_ps(tmp[7][m], _tmp7);
-#else
-                _mm_store_ps(tmp[0][m], _tmp0);
-                _mm_store_ps(tmp[1][m], _tmp1);
-                _mm_store_ps(tmp[2][m], _tmp2);
-                _mm_store_ps(tmp[3][m], _tmp3);
-                _mm_store_ps(tmp[4][m], _tmp4);
-                _mm_store_ps(tmp[5][m], _tmp5);
-                _mm_store_ps(tmp[6][m], _tmp6);
-                _mm_store_ps(tmp[7][m], _tmp7);
-#endif
-
-                r0 += w * elempack;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 64 + jj * 4;
-            float* p1 = p0 + max_jj * 4;
-            float* p2 = p0 + max_jj * 4 * 2;
-            float* p3 = p0 + max_jj * 4 * 3;
-            float* p4 = p0 + max_jj * 4 * 4;
-            float* p5 = p0 + max_jj * 4 * 5;
-            float* p6 = p0 + max_jj * 4 * 6;
-            float* p7 = p0 + max_jj * 4 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m128 _r0 = _mm_loadu_ps(tmp[m][0]);
-                __m128 _r1 = _mm_loadu_ps(tmp[m][1]);
-                __m128 _r2 = _mm_loadu_ps(tmp[m][2]);
-                __m128 _r3 = _mm_loadu_ps(tmp[m][3]);
-                __m128 _r4 = _mm_loadu_ps(tmp[m][4]);
-                __m128 _r5 = _mm_loadu_ps(tmp[m][5]);
-                __m128 _r6 = _mm_loadu_ps(tmp[m][6]);
-                __m128 _r7 = _mm_loadu_ps(tmp[m][7]);
-#else
-                __m128 _r0 = _mm_load_ps(tmp[m][0]);
-                __m128 _r1 = _mm_load_ps(tmp[m][1]);
-                __m128 _r2 = _mm_load_ps(tmp[m][2]);
-                __m128 _r3 = _mm_load_ps(tmp[m][3]);
-                __m128 _r4 = _mm_load_ps(tmp[m][4]);
-                __m128 _r5 = _mm_load_ps(tmp[m][5]);
-                __m128 _r6 = _mm_load_ps(tmp[m][6]);
-                __m128 _r7 = _mm_load_ps(tmp[m][7]);
-#endif
-
-                __m128 _v5_25 = _mm_set1_ps(5.25f);
-                __m128 _vm4_25 = _mm_set1_ps(-4.25f);
-                __m128 _vm1_25 = _mm_set1_ps(-1.25f);
-                __m128 _v0_25 = _mm_set1_ps(0.25f);
-                __m128 _vm2_5 = _mm_set1_ps(-2.5f);
-                __m128 _v0_5 = _mm_set1_ps(0.5f);
-                __m128 _v2 = _mm_set1_ps(2.f);
-                __m128 _v4 = _mm_set1_ps(4.f);
-
-                __m128 _tmp12a = _mm_comp_fmadd_ps(_vm4_25, _r4, _mm_add_ps(_r2, _r6));
-                __m128 _tmp12b = _mm_comp_fmadd_ps(_vm4_25, _r3, _mm_add_ps(_r1, _r5));
-                __m128 _tmp34a = _mm_comp_fmadd_ps(_vm1_25, _r4, _mm_comp_fmadd_ps(_v0_25, _r2, _r6));
-                __m128 _tmp34b = _mm_comp_fmadd_ps(_v2, _r5, _mm_comp_fmadd_ps(_vm2_5, _r3, _mm_mul_ps(_r1, _v0_5)));
-                __m128 _tmp56a = _mm_comp_fmadd_ps(_v4, _mm_comp_fmadd_ps(_vm1_25, _r4, _r2), _r6);
-                __m128 _tmp56b = _mm_comp_fmadd_ps(_v0_5, _r5, _mm_comp_fmadd_ps(_vm2_5, _r3, _mm_mul_ps(_r1, _v2)));
-
-                __m128 _tmp0 = _mm_comp_fmadd_ps(_v5_25, _mm_sub_ps(_r4, _r2), _mm_sub_ps(_r0, _r6));
-                __m128 _tmp1 = _mm_add_ps(_tmp12a, _tmp12b);
-                __m128 _tmp2 = _mm_sub_ps(_tmp12a, _tmp12b);
-                __m128 _tmp3 = _mm_add_ps(_tmp34a, _tmp34b);
-                __m128 _tmp4 = _mm_sub_ps(_tmp34a, _tmp34b);
-                __m128 _tmp5 = _mm_add_ps(_tmp56a, _tmp56b);
-                __m128 _tmp6 = _mm_sub_ps(_tmp56a, _tmp56b);
-                __m128 _tmp7 = _mm_comp_fmadd_ps(_v5_25, _mm_sub_ps(_r3, _r5), _mm_sub_ps(_r7, _r1));
-
-                _mm_store_ps(p0, _tmp0);
-                _mm_store_ps(p1, _tmp1);
-                _mm_store_ps(p2, _tmp2);
-                _mm_store_ps(p3, _tmp3);
-                _mm_store_ps(p4, _tmp4);
-                _mm_store_ps(p5, _tmp5);
-                _mm_store_ps(p6, _tmp6);
-                _mm_store_ps(p7, _tmp7);
-
-                p0 += max_jj * 8 * 4;
-                p1 += max_jj * 8 * 4;
-                p2 += max_jj * 8 * 4;
-                p3 += max_jj * 8 * 4;
-                p4 += max_jj * 8 * 4;
-                p5 += max_jj * 8 * 4;
-                p6 += max_jj * 8 * 4;
-                p7 += max_jj * 8 * 4;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 4;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-#else // __SSE2__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    #pragma omp parallel for num_threads(nT)
-#endif // __SSE2__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-        float tmp[8][8][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = bottom_blob.channel(k + kk).row(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 8; m++)
-            {
-                float r00 = 0.f;
-                float r01 = 0.f;
-                float r10 = 0.f;
-                float r11 = 0.f;
-                float r20 = 0.f;
-                float r21 = 0.f;
-                float r30 = 0.f;
-                float r31 = 0.f;
-                float r40 = 0.f;
-                float r41 = 0.f;
-                float r50 = 0.f;
-                float r51 = 0.f;
-                float r60 = 0.f;
-                float r61 = 0.f;
-                float r70 = 0.f;
-                float r71 = 0.f;
-
-                if (ti * 6 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const float* r1 = r0 + N;
-
-                        r00 = r0[0];
-                        r01 = r1[0];
-                        if (tj * 6 + 1 < w)
-                        {
-                            r10 = r0[1];
-                            r11 = r1[1];
-                        }
-                        if (tj * 6 + 2 < w)
-                        {
-                            r20 = r0[2];
-                            r21 = r1[2];
-                        }
-                        if (tj * 6 + 3 < w)
-                        {
-                            r30 = r0[3];
-                            r31 = r1[3];
-                        }
-                        if (tj * 6 + 4 < w)
-                        {
-                            r40 = r0[4];
-                            r41 = r1[4];
-                        }
-                        if (tj * 6 + 5 < w)
-                        {
-                            r50 = r0[5];
-                            r51 = r1[5];
-                        }
-                        if (tj * 6 + 6 < w)
-                        {
-                            r60 = r0[6];
-                            r61 = r1[6];
-                        }
-                        if (tj * 6 + 7 < w)
-                        {
-                            r70 = r0[7];
-                            r71 = r1[7];
-                        }
-                    }
-                }
-
-                float tmp12a0 = r20 + r60 - r40 * 4.25f;
-                float tmp12a1 = r21 + r61 - r41 * 4.25f;
-                float tmp12b0 = r10 + r50 - r30 * 4.25f;
-                float tmp12b1 = r11 + r51 - r31 * 4.25f;
-                float tmp34a0 = r60 + r20 * 0.25f - r40 * 1.25f;
-                float tmp34a1 = r61 + r21 * 0.25f - r41 * 1.25f;
-                float tmp34b0 = r10 * 0.5f - r30 * 2.5f + r50 * 2.f;
-                float tmp34b1 = r11 * 0.5f - r31 * 2.5f + r51 * 2.f;
-                float tmp56a0 = r20 * 4.f - r40 * 5.f + r60;
-                float tmp56a1 = r21 * 4.f - r41 * 5.f + r61;
-                float tmp56b0 = r10 * 2.f - r30 * 2.5f + r50 * 0.5f;
-                float tmp56b1 = r11 * 2.f - r31 * 2.5f + r51 * 0.5f;
-
-                tmp[0][m][0] = r00 - r60 + (r40 - r20) * 5.25f;
-                tmp[0][m][1] = r01 - r61 + (r41 - r21) * 5.25f;
-                tmp[1][m][0] = tmp12a0 + tmp12b0;
-                tmp[1][m][1] = tmp12a1 + tmp12b1;
-                tmp[2][m][0] = tmp12a0 - tmp12b0;
-                tmp[2][m][1] = tmp12a1 - tmp12b1;
-                tmp[3][m][0] = tmp34a0 + tmp34b0;
-                tmp[3][m][1] = tmp34a1 + tmp34b1;
-                tmp[4][m][0] = tmp34a0 - tmp34b0;
-                tmp[4][m][1] = tmp34a1 - tmp34b1;
-                tmp[5][m][0] = tmp56a0 + tmp56b0;
-                tmp[5][m][1] = tmp56a1 + tmp56b1;
-                tmp[6][m][0] = tmp56a0 - tmp56b0;
-                tmp[6][m][1] = tmp56a1 - tmp56b1;
-                tmp[7][m][0] = r70 - r10 + (r30 - r50) * 5.25f;
-                tmp[7][m][1] = r71 - r11 + (r31 - r51) * 5.25f;
-
-                r0 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 64 + jj * 2;
-            float* p1 = p0 + max_jj * 2;
-            float* p2 = p0 + max_jj * 2 * 2;
-            float* p3 = p0 + max_jj * 2 * 3;
-            float* p4 = p0 + max_jj * 2 * 4;
-            float* p5 = p0 + max_jj * 2 * 5;
-            float* p6 = p0 + max_jj * 2 * 6;
-            float* p7 = p0 + max_jj * 2 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-                float r40 = tmp[m][4][0];
-                float r41 = tmp[m][4][1];
-                float r50 = tmp[m][5][0];
-                float r51 = tmp[m][5][1];
-                float r60 = tmp[m][6][0];
-                float r61 = tmp[m][6][1];
-                float r70 = tmp[m][7][0];
-                float r71 = tmp[m][7][1];
-
-                float tmp12a0 = r20 + r60 - r40 * 4.25f;
-                float tmp12a1 = r21 + r61 - r41 * 4.25f;
-                float tmp12b0 = r10 + r50 - r30 * 4.25f;
-                float tmp12b1 = r11 + r51 - r31 * 4.25f;
-                float tmp34a0 = r60 + r20 * 0.25f - r40 * 1.25f;
-                float tmp34a1 = r61 + r21 * 0.25f - r41 * 1.25f;
-                float tmp34b0 = r10 * 0.5f - r30 * 2.5f + r50 * 2.f;
-                float tmp34b1 = r11 * 0.5f - r31 * 2.5f + r51 * 2.f;
-                float tmp56a0 = r20 * 4.f - r40 * 5.f + r60;
-                float tmp56a1 = r21 * 4.f - r41 * 5.f + r61;
-                float tmp56b0 = r10 * 2.f - r30 * 2.5f + r50 * 0.5f;
-                float tmp56b1 = r11 * 2.f - r31 * 2.5f + r51 * 0.5f;
-
-                p0[0] = r00 - r60 + (r40 - r20) * 5.25f;
-                p0[1] = r01 - r61 + (r41 - r21) * 5.25f;
-                p1[0] = tmp12a0 + tmp12b0;
-                p1[1] = tmp12a1 + tmp12b1;
-                p2[0] = tmp12a0 - tmp12b0;
-                p2[1] = tmp12a1 - tmp12b1;
-                p3[0] = tmp34a0 + tmp34b0;
-                p3[1] = tmp34a1 + tmp34b1;
-                p4[0] = tmp34a0 - tmp34b0;
-                p4[1] = tmp34a1 - tmp34b1;
-                p5[0] = tmp56a0 + tmp56b0;
-                p5[1] = tmp56a1 + tmp56b1;
-                p6[0] = tmp56a0 - tmp56b0;
-                p6[1] = tmp56a1 - tmp56b1;
-                p7[0] = r70 - r10 + (r30 - r50) * 5.25f;
-                p7[1] = r71 - r11 + (r31 - r51) * 5.25f;
-
-                p0 += max_jj * 8 * 2;
-                p1 += max_jj * 8 * 2;
-                p2 += max_jj * 8 * 2;
-                p3 += max_jj * 8 * 2;
-                p4 += max_jj * 8 * 2;
-                p5 += max_jj * 8 * 2;
-                p6 += max_jj * 8 * 2;
-                p7 += max_jj * 8 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        float tmp[8][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0123 = bottom_blob.channel(k + kk).row(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 8; m++)
-            {
-                float r0 = 0.f;
-                float r1 = 0.f;
-                float r2 = 0.f;
-                float r3 = 0.f;
-                float r4 = 0.f;
-                float r5 = 0.f;
-                float r6 = 0.f;
-                float r7 = 0.f;
-
-                if (ti * 6 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = r0123[0];
-                        if (tj * 6 + 1 < w) r1 = r0123[1];
-                        if (tj * 6 + 2 < w) r2 = r0123[2];
-                        if (tj * 6 + 3 < w) r3 = r0123[3];
-                        if (tj * 6 + 4 < w) r4 = r0123[4];
-                        if (tj * 6 + 5 < w) r5 = r0123[5];
-                        if (tj * 6 + 6 < w) r6 = r0123[6];
-                        if (tj * 6 + 7 < w) r7 = r0123[7];
-                    }
-                }
-
-                float tmp12a = r2 + r6 - r4 * 4.25f;
-                float tmp12b = r1 + r5 - r3 * 4.25f;
-                float tmp34a = r6 + r2 * 0.25f - r4 * 1.25f;
-                float tmp34b = r1 * 0.5f - r3 * 2.5f + r5 * 2.f;
-                float tmp56a = r2 * 4.f - r4 * 5.f + r6;
-                float tmp56b = r1 * 2.f - r3 * 2.5f + r5 * 0.5f;
-
-                tmp[0][m] = r0 - r6 + (r4 - r2) * 5.25f;
-                tmp[1][m] = tmp12a + tmp12b;
-                tmp[2][m] = tmp12a - tmp12b;
-                tmp[3][m] = tmp34a + tmp34b;
-                tmp[4][m] = tmp34a - tmp34b;
-                tmp[5][m] = tmp56a + tmp56b;
-                tmp[6][m] = tmp56a - tmp56b;
-                tmp[7][m] = r7 - r1 + (r3 - r5) * 5.25f;
-
-                r0123 += w;
-            }
-
-            float* p0 = (float*)B + kk * max_jj * 64 + jj;
-            float* p1 = p0 + max_jj;
-            float* p2 = p0 + max_jj * 2;
-            float* p3 = p0 + max_jj * 3;
-            float* p4 = p0 + max_jj * 4;
-            float* p5 = p0 + max_jj * 5;
-            float* p6 = p0 + max_jj * 6;
-            float* p7 = p0 + max_jj * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-                float r4 = tmp[m][4];
-                float r5 = tmp[m][5];
-                float r6 = tmp[m][6];
-                float r7 = tmp[m][7];
-
-                float tmp12a = r2 + r6 - r4 * 4.25f;
-                float tmp12b = r1 + r5 - r3 * 4.25f;
-                float tmp34a = r6 + r2 * 0.25f - r4 * 1.25f;
-                float tmp34b = r1 * 0.5f - r3 * 2.5f + r5 * 2.f;
-                float tmp56a = r2 * 4.f - r4 * 5.f + r6;
-                float tmp56b = r1 * 2.f - r3 * 2.5f + r5 * 0.5f;
-
-                p0[0] = r0 - r6 + (r4 - r2) * 5.25f;
-                p1[0] = tmp12a + tmp12b;
-                p2[0] = tmp12a - tmp12b;
-                p3[0] = tmp34a + tmp34b;
-                p4[0] = tmp34a - tmp34b;
-                p5[0] = tmp56a + tmp56b;
-                p6[0] = tmp56a - tmp56b;
-                p7[0] = r7 - r1 + (r3 - r5) * 5.25f;
-
-                p0 += max_jj * 8;
-                p1 += max_jj * 8;
-                p2 += max_jj * 8;
-                p3 += max_jj * 8;
-                p4 += max_jj * 8;
-                p5 += max_jj * 8;
-                p6 += max_jj * 8;
-                p7 += max_jj * 8;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd63_transform_output_tile(const Mat& top_tile, Mat& top_blob, const Mat& bias, int i, int max_ii, int j, int max_jj)
-{
-    // const float otm[6][8] = {
-    //     {1.0f, 1.0f,  1.0f,  1.0f,  1.0f, 32.0f, 32.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f,  2.0f, -2.0f, 16.0f,-16.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f,  4.0f,  4.0f,  8.0f,  8.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f,  8.0f, -8.0f,  4.0f, -4.0f, 0.0f},
-    //     {0.0f, 1.0f,  1.0f, 16.0f, 16.0f,  2.0f,  2.0f, 0.0f},
-    //     {0.0f, 1.0f, -1.0f, 32.0f,-32.0f,  1.0f, -1.0f, 1.0f}
-    // };
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 5) / 6;
-
-    const float* biasptr = bias;
-
-    int ii = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    for (; ii + 15 < max_ii; ii += 16)
-    {
-        __m512 _bias0 = biasptr ? _mm512_loadu_ps(biasptr + i + ii) : _mm512_setzero_ps();
-
-#ifdef _MSC_VER
-        __declspec(align(64))
-#else
-        __attribute__((aligned(64)))
-#endif
-        float tmp[6][8][16];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 64 + jj * 16;
-            const float* r1 = r0 + max_jj * 16;
-            const float* r2 = r0 + max_jj * 16 * 2;
-            const float* r3 = r0 + max_jj * 16 * 3;
-            const float* r4 = r0 + max_jj * 16 * 4;
-            const float* r5 = r0 + max_jj * 16 * 5;
-            const float* r6 = r0 + max_jj * 16 * 6;
-            const float* r7 = r0 + max_jj * 16 * 7;
-
-            __m512 _v32 = _mm512_set1_ps(32.f);
-            __m512 _v16 = _mm512_set1_ps(16.f);
-            __m512 _v8 = _mm512_set1_ps(8.f);
-            __m512 _v4 = _mm512_set1_ps(4.f);
-            __m512 _v2 = _mm512_set1_ps(2.f);
-
-            for (int m = 0; m < 8; m++)
-            {
-                __m512 _r0 = _mm512_load_ps(r0);
-                __m512 _r1 = _mm512_load_ps(r1);
-                __m512 _r2 = _mm512_load_ps(r2);
-                __m512 _r3 = _mm512_load_ps(r3);
-                __m512 _r4 = _mm512_load_ps(r4);
-                __m512 _r5 = _mm512_load_ps(r5);
-                __m512 _r6 = _mm512_load_ps(r6);
-                __m512 _r7 = _mm512_load_ps(r7);
-
-                __m512 _tmp024a = _mm512_add_ps(_r1, _r2);
-                __m512 _tmp135a = _mm512_sub_ps(_r1, _r2);
-                __m512 _tmp024b = _mm512_add_ps(_r3, _r4);
-                __m512 _tmp135b = _mm512_sub_ps(_r3, _r4);
-                __m512 _tmp024c = _mm512_add_ps(_r5, _r6);
-                __m512 _tmp135c = _mm512_sub_ps(_r5, _r6);
-                __m512 _tmp0 = _mm512_add_ps(_mm512_add_ps(_r0, _tmp024a), _mm512_fmadd_ps(_v32, _tmp024c, _tmp024b));
-                __m512 _tmp1 = _mm512_fmadd_ps(_v16, _tmp135c, _mm512_fmadd_ps(_v2, _tmp135b, _tmp135a));
-                __m512 _tmp2 = _mm512_fmadd_ps(_v8, _tmp024c, _mm512_fmadd_ps(_v4, _tmp024b, _tmp024a));
-                __m512 _tmp3 = _mm512_fmadd_ps(_v4, _tmp135c, _mm512_fmadd_ps(_v8, _tmp135b, _tmp135a));
-                __m512 _tmp4 = _mm512_fmadd_ps(_v2, _tmp024c, _mm512_fmadd_ps(_v16, _tmp024b, _tmp024a));
-                __m512 _tmp5 = _mm512_add_ps(_mm512_add_ps(_r7, _tmp135a), _mm512_fmadd_ps(_v32, _tmp135b, _tmp135c));
-
-                _mm512_store_ps(tmp[0][m], _tmp0);
-                _mm512_store_ps(tmp[1][m], _tmp1);
-                _mm512_store_ps(tmp[2][m], _tmp2);
-                _mm512_store_ps(tmp[3][m], _tmp3);
-                _mm512_store_ps(tmp[4][m], _tmp4);
-                _mm512_store_ps(tmp[5][m], _tmp5);
-
-                r0 += max_jj * 8 * 16;
-                r1 += max_jj * 8 * 16;
-                r2 += max_jj * 8 * 16;
-                r3 += max_jj * 8 * 16;
-                r4 += max_jj * 8 * 16;
-                r5 += max_jj * 8 * 16;
-                r6 += max_jj * 8 * 16;
-                r7 += max_jj * 8 * 16;
-            }
-
-            float* outptr0 = top_blob.channel((i + ii) / out_elempack).row(ti * 6) + (tj * 6) * out_elempack;
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-                __m512 _r0 = _mm512_load_ps(tmp[m][0]);
-                __m512 _r1 = _mm512_load_ps(tmp[m][1]);
-                __m512 _r2 = _mm512_load_ps(tmp[m][2]);
-                __m512 _r3 = _mm512_load_ps(tmp[m][3]);
-                __m512 _r4 = _mm512_load_ps(tmp[m][4]);
-                __m512 _r5 = _mm512_load_ps(tmp[m][5]);
-                __m512 _r6 = _mm512_load_ps(tmp[m][6]);
-                __m512 _r7 = _mm512_load_ps(tmp[m][7]);
-
-                __m512 _tmp024a = _mm512_add_ps(_r1, _r2);
-                __m512 _tmp135a = _mm512_sub_ps(_r1, _r2);
-                __m512 _tmp024b = _mm512_add_ps(_r3, _r4);
-                __m512 _tmp135b = _mm512_sub_ps(_r3, _r4);
-                __m512 _tmp024c = _mm512_add_ps(_r5, _r6);
-                __m512 _tmp135c = _mm512_sub_ps(_r5, _r6);
-                __m512 _tmp0 = _mm512_add_ps(_bias0, _mm512_add_ps(_mm512_add_ps(_r0, _tmp024a), _mm512_fmadd_ps(_v32, _tmp024c, _tmp024b)));
-                __m512 _tmp1 = _mm512_add_ps(_bias0, _mm512_fmadd_ps(_v16, _tmp135c, _mm512_fmadd_ps(_v2, _tmp135b, _tmp135a)));
-                __m512 _tmp2 = _mm512_add_ps(_bias0, _mm512_fmadd_ps(_v8, _tmp024c, _mm512_fmadd_ps(_v4, _tmp024b, _tmp024a)));
-                __m512 _tmp3 = _mm512_add_ps(_bias0, _mm512_fmadd_ps(_v4, _tmp135c, _mm512_fmadd_ps(_v8, _tmp135b, _tmp135a)));
-                __m512 _tmp4 = _mm512_add_ps(_bias0, _mm512_fmadd_ps(_v2, _tmp024c, _mm512_fmadd_ps(_v16, _tmp024b, _tmp024a)));
-                __m512 _tmp5 = _mm512_add_ps(_bias0, _mm512_add_ps(_mm512_add_ps(_r7, _tmp135a), _mm512_fmadd_ps(_v32, _tmp135b, _tmp135c)));
-
-                if (out_elempack == 16)
-                {
-                    _mm512_store_ps(outptr0, _tmp0);
-                    if (tj * 6 + 1 < outw) _mm512_store_ps(outptr0 + 16, _tmp1);
-                    if (tj * 6 + 2 < outw) _mm512_store_ps(outptr0 + 32, _tmp2);
-                    if (tj * 6 + 3 < outw) _mm512_store_ps(outptr0 + 48, _tmp3);
-                    if (tj * 6 + 4 < outw) _mm512_store_ps(outptr0 + 64, _tmp4);
-                    if (tj * 6 + 5 < outw) _mm512_store_ps(outptr0 + 80, _tmp5);
-                }
-                if (out_elempack == 8)
-                {
-                    float* outptr1 = outptr0 + N;
-
-                    _mm256_store_ps(outptr0, _mm512_extractf32x8_ps(_tmp0, 0));
-                    _mm256_store_ps(outptr1, _mm512_extractf32x8_ps(_tmp0, 1));
-                    if (tj * 6 + 1 < outw)
-                    {
-                        _mm256_store_ps(outptr0 + 8, _mm512_extractf32x8_ps(_tmp1, 0));
-                        _mm256_store_ps(outptr1 + 8, _mm512_extractf32x8_ps(_tmp1, 1));
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        _mm256_store_ps(outptr0 + 16, _mm512_extractf32x8_ps(_tmp2, 0));
-                        _mm256_store_ps(outptr1 + 16, _mm512_extractf32x8_ps(_tmp2, 1));
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        _mm256_store_ps(outptr0 + 24, _mm512_extractf32x8_ps(_tmp3, 0));
-                        _mm256_store_ps(outptr1 + 24, _mm512_extractf32x8_ps(_tmp3, 1));
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        _mm256_store_ps(outptr0 + 32, _mm512_extractf32x8_ps(_tmp4, 0));
-                        _mm256_store_ps(outptr1 + 32, _mm512_extractf32x8_ps(_tmp4, 1));
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        _mm256_store_ps(outptr0 + 40, _mm512_extractf32x8_ps(_tmp5, 0));
-                        _mm256_store_ps(outptr1 + 40, _mm512_extractf32x8_ps(_tmp5, 1));
-                    }
-                }
-                if (out_elempack == 4)
-                {
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-
-                    _mm_store_ps(outptr0, _mm512_extractf32x4_ps(_tmp0, 0));
-                    _mm_store_ps(outptr1, _mm512_extractf32x4_ps(_tmp0, 1));
-                    _mm_store_ps(outptr2, _mm512_extractf32x4_ps(_tmp0, 2));
-                    _mm_store_ps(outptr3, _mm512_extractf32x4_ps(_tmp0, 3));
-                    if (tj * 6 + 1 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 4, _mm512_extractf32x4_ps(_tmp1, 0));
-                        _mm_store_ps(outptr1 + 4, _mm512_extractf32x4_ps(_tmp1, 1));
-                        _mm_store_ps(outptr2 + 4, _mm512_extractf32x4_ps(_tmp1, 2));
-                        _mm_store_ps(outptr3 + 4, _mm512_extractf32x4_ps(_tmp1, 3));
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 8, _mm512_extractf32x4_ps(_tmp2, 0));
-                        _mm_store_ps(outptr1 + 8, _mm512_extractf32x4_ps(_tmp2, 1));
-                        _mm_store_ps(outptr2 + 8, _mm512_extractf32x4_ps(_tmp2, 2));
-                        _mm_store_ps(outptr3 + 8, _mm512_extractf32x4_ps(_tmp2, 3));
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 12, _mm512_extractf32x4_ps(_tmp3, 0));
-                        _mm_store_ps(outptr1 + 12, _mm512_extractf32x4_ps(_tmp3, 1));
-                        _mm_store_ps(outptr2 + 12, _mm512_extractf32x4_ps(_tmp3, 2));
-                        _mm_store_ps(outptr3 + 12, _mm512_extractf32x4_ps(_tmp3, 3));
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 16, _mm512_extractf32x4_ps(_tmp4, 0));
-                        _mm_store_ps(outptr1 + 16, _mm512_extractf32x4_ps(_tmp4, 1));
-                        _mm_store_ps(outptr2 + 16, _mm512_extractf32x4_ps(_tmp4, 2));
-                        _mm_store_ps(outptr3 + 16, _mm512_extractf32x4_ps(_tmp4, 3));
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 20, _mm512_extractf32x4_ps(_tmp5, 0));
-                        _mm_store_ps(outptr1 + 20, _mm512_extractf32x4_ps(_tmp5, 1));
-                        _mm_store_ps(outptr2 + 20, _mm512_extractf32x4_ps(_tmp5, 2));
-                        _mm_store_ps(outptr3 + 20, _mm512_extractf32x4_ps(_tmp5, 3));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    float tmp0[16];
-                    float tmp1[16];
-                    float tmp2[16];
-                    float tmp3[16];
-                    float tmp4[16];
-                    float tmp5[16];
-                    _mm512_storeu_ps(tmp0, _tmp0);
-                    _mm512_storeu_ps(tmp1, _tmp1);
-                    _mm512_storeu_ps(tmp2, _tmp2);
-                    _mm512_storeu_ps(tmp3, _tmp3);
-                    _mm512_storeu_ps(tmp4, _tmp4);
-                    _mm512_storeu_ps(tmp5, _tmp5);
-
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-                    float* outptr4 = outptr0 + N * 4;
-                    float* outptr5 = outptr0 + N * 5;
-                    float* outptr6 = outptr0 + N * 6;
-                    float* outptr7 = outptr0 + N * 7;
-                    float* outptr8 = outptr0 + N * 8;
-                    float* outptr9 = outptr0 + N * 9;
-                    float* outptra = outptr0 + N * 10;
-                    float* outptrb = outptr0 + N * 11;
-                    float* outptrc = outptr0 + N * 12;
-                    float* outptrd = outptr0 + N * 13;
-                    float* outptre = outptr0 + N * 14;
-                    float* outptrf = outptr0 + N * 15;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-                    outptr8[0] = tmp0[8];
-                    outptr9[0] = tmp0[9];
-                    outptra[0] = tmp0[10];
-                    outptrb[0] = tmp0[11];
-                    outptrc[0] = tmp0[12];
-                    outptrd[0] = tmp0[13];
-                    outptre[0] = tmp0[14];
-                    outptrf[0] = tmp0[15];
-                    if (tj * 6 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                        outptr8[1] = tmp1[8];
-                        outptr9[1] = tmp1[9];
-                        outptra[1] = tmp1[10];
-                        outptrb[1] = tmp1[11];
-                        outptrc[1] = tmp1[12];
-                        outptrd[1] = tmp1[13];
-                        outptre[1] = tmp1[14];
-                        outptrf[1] = tmp1[15];
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                        outptr4[2] = tmp2[4];
-                        outptr5[2] = tmp2[5];
-                        outptr6[2] = tmp2[6];
-                        outptr7[2] = tmp2[7];
-                        outptr8[2] = tmp2[8];
-                        outptr9[2] = tmp2[9];
-                        outptra[2] = tmp2[10];
-                        outptrb[2] = tmp2[11];
-                        outptrc[2] = tmp2[12];
-                        outptrd[2] = tmp2[13];
-                        outptre[2] = tmp2[14];
-                        outptrf[2] = tmp2[15];
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                        outptr4[3] = tmp3[4];
-                        outptr5[3] = tmp3[5];
-                        outptr6[3] = tmp3[6];
-                        outptr7[3] = tmp3[7];
-                        outptr8[3] = tmp3[8];
-                        outptr9[3] = tmp3[9];
-                        outptra[3] = tmp3[10];
-                        outptrb[3] = tmp3[11];
-                        outptrc[3] = tmp3[12];
-                        outptrd[3] = tmp3[13];
-                        outptre[3] = tmp3[14];
-                        outptrf[3] = tmp3[15];
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        outptr0[4] = tmp4[0];
-                        outptr1[4] = tmp4[1];
-                        outptr2[4] = tmp4[2];
-                        outptr3[4] = tmp4[3];
-                        outptr4[4] = tmp4[4];
-                        outptr5[4] = tmp4[5];
-                        outptr6[4] = tmp4[6];
-                        outptr7[4] = tmp4[7];
-                        outptr8[4] = tmp4[8];
-                        outptr9[4] = tmp4[9];
-                        outptra[4] = tmp4[10];
-                        outptrb[4] = tmp4[11];
-                        outptrc[4] = tmp4[12];
-                        outptrd[4] = tmp4[13];
-                        outptre[4] = tmp4[14];
-                        outptrf[4] = tmp4[15];
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        outptr0[5] = tmp5[0];
-                        outptr1[5] = tmp5[1];
-                        outptr2[5] = tmp5[2];
-                        outptr3[5] = tmp5[3];
-                        outptr4[5] = tmp5[4];
-                        outptr5[5] = tmp5[5];
-                        outptr6[5] = tmp5[6];
-                        outptr7[5] = tmp5[7];
-                        outptr8[5] = tmp5[8];
-                        outptr9[5] = tmp5[9];
-                        outptra[5] = tmp5[10];
-                        outptrb[5] = tmp5[11];
-                        outptrc[5] = tmp5[12];
-                        outptrd[5] = tmp5[13];
-                        outptre[5] = tmp5[14];
-                        outptrf[5] = tmp5[15];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __AVX512F__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        __m256 _bias0 = biasptr ? _mm256_loadu_ps(biasptr + i + ii) : _mm256_setzero_ps();
-
-#ifdef _MSC_VER
-        __declspec(align(32))
-#else
-        __attribute__((aligned(32)))
-#endif
-        float tmp[6][8][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 64 + jj * 8;
-            const float* r1 = r0 + max_jj * 8;
-            const float* r2 = r0 + max_jj * 8 * 2;
-            const float* r3 = r0 + max_jj * 8 * 3;
-            const float* r4 = r0 + max_jj * 8 * 4;
-            const float* r5 = r0 + max_jj * 8 * 5;
-            const float* r6 = r0 + max_jj * 8 * 6;
-            const float* r7 = r0 + max_jj * 8 * 7;
-
-            __m256 _v32 = _mm256_set1_ps(32.f);
-            __m256 _v16 = _mm256_set1_ps(16.f);
-            __m256 _v8 = _mm256_set1_ps(8.f);
-            __m256 _v4 = _mm256_set1_ps(4.f);
-            __m256 _v2 = _mm256_set1_ps(2.f);
-
-            for (int m = 0; m < 8; m++)
-            {
-                __m256 _r0 = _mm256_load_ps(r0);
-                __m256 _r1 = _mm256_load_ps(r1);
-                __m256 _r2 = _mm256_load_ps(r2);
-                __m256 _r3 = _mm256_load_ps(r3);
-                __m256 _r4 = _mm256_load_ps(r4);
-                __m256 _r5 = _mm256_load_ps(r5);
-                __m256 _r6 = _mm256_load_ps(r6);
-                __m256 _r7 = _mm256_load_ps(r7);
-
-                __m256 _tmp024a = _mm256_add_ps(_r1, _r2);
-                __m256 _tmp135a = _mm256_sub_ps(_r1, _r2);
-                __m256 _tmp024b = _mm256_add_ps(_r3, _r4);
-                __m256 _tmp135b = _mm256_sub_ps(_r3, _r4);
-                __m256 _tmp024c = _mm256_add_ps(_r5, _r6);
-                __m256 _tmp135c = _mm256_sub_ps(_r5, _r6);
-                __m256 _tmp0 = _mm256_add_ps(_mm256_add_ps(_r0, _tmp024a), _mm256_comp_fmadd_ps(_v32, _tmp024c, _tmp024b));
-                __m256 _tmp1 = _mm256_comp_fmadd_ps(_v16, _tmp135c, _mm256_comp_fmadd_ps(_v2, _tmp135b, _tmp135a));
-                __m256 _tmp2 = _mm256_comp_fmadd_ps(_v8, _tmp024c, _mm256_comp_fmadd_ps(_v4, _tmp024b, _tmp024a));
-                __m256 _tmp3 = _mm256_comp_fmadd_ps(_v4, _tmp135c, _mm256_comp_fmadd_ps(_v8, _tmp135b, _tmp135a));
-                __m256 _tmp4 = _mm256_comp_fmadd_ps(_v2, _tmp024c, _mm256_comp_fmadd_ps(_v16, _tmp024b, _tmp024a));
-                __m256 _tmp5 = _mm256_add_ps(_mm256_add_ps(_r7, _tmp135a), _mm256_comp_fmadd_ps(_v32, _tmp135b, _tmp135c));
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm256_storeu_ps(tmp[0][m], _tmp0);
-                _mm256_storeu_ps(tmp[1][m], _tmp1);
-                _mm256_storeu_ps(tmp[2][m], _tmp2);
-                _mm256_storeu_ps(tmp[3][m], _tmp3);
-                _mm256_storeu_ps(tmp[4][m], _tmp4);
-                _mm256_storeu_ps(tmp[5][m], _tmp5);
-#else
-                _mm256_store_ps(tmp[0][m], _tmp0);
-                _mm256_store_ps(tmp[1][m], _tmp1);
-                _mm256_store_ps(tmp[2][m], _tmp2);
-                _mm256_store_ps(tmp[3][m], _tmp3);
-                _mm256_store_ps(tmp[4][m], _tmp4);
-                _mm256_store_ps(tmp[5][m], _tmp5);
-#endif
-
-                r0 += max_jj * 8 * 8;
-                r1 += max_jj * 8 * 8;
-                r2 += max_jj * 8 * 8;
-                r3 += max_jj * 8 * 8;
-                r4 += max_jj * 8 * 8;
-                r5 += max_jj * 8 * 8;
-                r6 += max_jj * 8 * 8;
-                r7 += max_jj * 8 * 8;
-            }
-
-            float* outptr0 = top_blob.channel((i + ii) / out_elempack).row(ti * 6) + (tj * 6) * out_elempack;
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m256 _r0 = _mm256_loadu_ps(tmp[m][0]);
-                __m256 _r1 = _mm256_loadu_ps(tmp[m][1]);
-                __m256 _r2 = _mm256_loadu_ps(tmp[m][2]);
-                __m256 _r3 = _mm256_loadu_ps(tmp[m][3]);
-                __m256 _r4 = _mm256_loadu_ps(tmp[m][4]);
-                __m256 _r5 = _mm256_loadu_ps(tmp[m][5]);
-                __m256 _r6 = _mm256_loadu_ps(tmp[m][6]);
-                __m256 _r7 = _mm256_loadu_ps(tmp[m][7]);
-#else
-                __m256 _r0 = _mm256_load_ps(tmp[m][0]);
-                __m256 _r1 = _mm256_load_ps(tmp[m][1]);
-                __m256 _r2 = _mm256_load_ps(tmp[m][2]);
-                __m256 _r3 = _mm256_load_ps(tmp[m][3]);
-                __m256 _r4 = _mm256_load_ps(tmp[m][4]);
-                __m256 _r5 = _mm256_load_ps(tmp[m][5]);
-                __m256 _r6 = _mm256_load_ps(tmp[m][6]);
-                __m256 _r7 = _mm256_load_ps(tmp[m][7]);
-#endif
-
-                __m256 _tmp024a = _mm256_add_ps(_r1, _r2);
-                __m256 _tmp135a = _mm256_sub_ps(_r1, _r2);
-                __m256 _tmp024b = _mm256_add_ps(_r3, _r4);
-                __m256 _tmp135b = _mm256_sub_ps(_r3, _r4);
-                __m256 _tmp024c = _mm256_add_ps(_r5, _r6);
-                __m256 _tmp135c = _mm256_sub_ps(_r5, _r6);
-                __m256 _tmp0 = _mm256_add_ps(_bias0, _mm256_add_ps(_mm256_add_ps(_r0, _tmp024a), _mm256_comp_fmadd_ps(_v32, _tmp024c, _tmp024b)));
-                __m256 _tmp1 = _mm256_add_ps(_bias0, _mm256_comp_fmadd_ps(_v16, _tmp135c, _mm256_comp_fmadd_ps(_v2, _tmp135b, _tmp135a)));
-                __m256 _tmp2 = _mm256_add_ps(_bias0, _mm256_comp_fmadd_ps(_v8, _tmp024c, _mm256_comp_fmadd_ps(_v4, _tmp024b, _tmp024a)));
-                __m256 _tmp3 = _mm256_add_ps(_bias0, _mm256_comp_fmadd_ps(_v4, _tmp135c, _mm256_comp_fmadd_ps(_v8, _tmp135b, _tmp135a)));
-                __m256 _tmp4 = _mm256_add_ps(_bias0, _mm256_comp_fmadd_ps(_v2, _tmp024c, _mm256_comp_fmadd_ps(_v16, _tmp024b, _tmp024a)));
-                __m256 _tmp5 = _mm256_add_ps(_bias0, _mm256_add_ps(_mm256_add_ps(_r7, _tmp135a), _mm256_comp_fmadd_ps(_v32, _tmp135b, _tmp135c)));
-
-                if (out_elempack == 8)
-                {
-                    _mm256_store_ps(outptr0, _tmp0);
-                    if (tj * 6 + 1 < outw) _mm256_store_ps(outptr0 + 8, _tmp1);
-                    if (tj * 6 + 2 < outw) _mm256_store_ps(outptr0 + 16, _tmp2);
-                    if (tj * 6 + 3 < outw) _mm256_store_ps(outptr0 + 24, _tmp3);
-                    if (tj * 6 + 4 < outw) _mm256_store_ps(outptr0 + 32, _tmp4);
-                    if (tj * 6 + 5 < outw) _mm256_store_ps(outptr0 + 40, _tmp5);
-                }
-                if (out_elempack == 4)
-                {
-                    float* outptr1 = outptr0 + N;
-
-                    _mm_store_ps(outptr0, _mm256_extractf128_ps(_tmp0, 0));
-                    _mm_store_ps(outptr1, _mm256_extractf128_ps(_tmp0, 1));
-                    if (tj * 6 + 1 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 4, _mm256_extractf128_ps(_tmp1, 0));
-                        _mm_store_ps(outptr1 + 4, _mm256_extractf128_ps(_tmp1, 1));
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 8, _mm256_extractf128_ps(_tmp2, 0));
-                        _mm_store_ps(outptr1 + 8, _mm256_extractf128_ps(_tmp2, 1));
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 12, _mm256_extractf128_ps(_tmp3, 0));
-                        _mm_store_ps(outptr1 + 12, _mm256_extractf128_ps(_tmp3, 1));
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 16, _mm256_extractf128_ps(_tmp4, 0));
-                        _mm_store_ps(outptr1 + 16, _mm256_extractf128_ps(_tmp4, 1));
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        _mm_store_ps(outptr0 + 20, _mm256_extractf128_ps(_tmp5, 0));
-                        _mm_store_ps(outptr1 + 20, _mm256_extractf128_ps(_tmp5, 1));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    float tmp0[8];
-                    float tmp1[8];
-                    float tmp2[8];
-                    float tmp3[8];
-                    float tmp4[8];
-                    float tmp5[8];
-                    _mm256_storeu_ps(tmp0, _tmp0);
-                    _mm256_storeu_ps(tmp1, _tmp1);
-                    _mm256_storeu_ps(tmp2, _tmp2);
-                    _mm256_storeu_ps(tmp3, _tmp3);
-                    _mm256_storeu_ps(tmp4, _tmp4);
-                    _mm256_storeu_ps(tmp5, _tmp5);
-
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-                    float* outptr4 = outptr0 + N * 4;
-                    float* outptr5 = outptr0 + N * 5;
-                    float* outptr6 = outptr0 + N * 6;
-                    float* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-                    if (tj * 6 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                        outptr4[2] = tmp2[4];
-                        outptr5[2] = tmp2[5];
-                        outptr6[2] = tmp2[6];
-                        outptr7[2] = tmp2[7];
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                        outptr4[3] = tmp3[4];
-                        outptr5[3] = tmp3[5];
-                        outptr6[3] = tmp3[6];
-                        outptr7[3] = tmp3[7];
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        outptr0[4] = tmp4[0];
-                        outptr1[4] = tmp4[1];
-                        outptr2[4] = tmp4[2];
-                        outptr3[4] = tmp4[3];
-                        outptr4[4] = tmp4[4];
-                        outptr5[4] = tmp4[5];
-                        outptr6[4] = tmp4[6];
-                        outptr7[4] = tmp4[7];
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        outptr0[5] = tmp5[0];
-                        outptr1[5] = tmp5[1];
-                        outptr2[5] = tmp5[2];
-                        outptr3[5] = tmp5[3];
-                        outptr4[5] = tmp5[4];
-                        outptr5[5] = tmp5[5];
-                        outptr6[5] = tmp5[6];
-                        outptr7[5] = tmp5[7];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __AVX__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        __m128 _bias0 = biasptr ? _mm_loadu_ps(biasptr + i + ii) : _mm_setzero_ps();
-
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        float tmp[6][8][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 64 + jj * 4;
-            const float* r1 = r0 + max_jj * 4;
-            const float* r2 = r0 + max_jj * 4 * 2;
-            const float* r3 = r0 + max_jj * 4 * 3;
-            const float* r4 = r0 + max_jj * 4 * 4;
-            const float* r5 = r0 + max_jj * 4 * 5;
-            const float* r6 = r0 + max_jj * 4 * 6;
-            const float* r7 = r0 + max_jj * 4 * 7;
-
-            __m128 _v32 = _mm_set1_ps(32.f);
-            __m128 _v16 = _mm_set1_ps(16.f);
-            __m128 _v8 = _mm_set1_ps(8.f);
-            __m128 _v4 = _mm_set1_ps(4.f);
-            __m128 _v2 = _mm_set1_ps(2.f);
-
-            for (int m = 0; m < 8; m++)
-            {
-                __m128 _r0 = _mm_load_ps(r0);
-                __m128 _r1 = _mm_load_ps(r1);
-                __m128 _r2 = _mm_load_ps(r2);
-                __m128 _r3 = _mm_load_ps(r3);
-                __m128 _r4 = _mm_load_ps(r4);
-                __m128 _r5 = _mm_load_ps(r5);
-                __m128 _r6 = _mm_load_ps(r6);
-                __m128 _r7 = _mm_load_ps(r7);
-
-                __m128 _tmp024a = _mm_add_ps(_r1, _r2);
-                __m128 _tmp135a = _mm_sub_ps(_r1, _r2);
-                __m128 _tmp024b = _mm_add_ps(_r3, _r4);
-                __m128 _tmp135b = _mm_sub_ps(_r3, _r4);
-                __m128 _tmp024c = _mm_add_ps(_r5, _r6);
-                __m128 _tmp135c = _mm_sub_ps(_r5, _r6);
-                __m128 _tmp0 = _mm_add_ps(_mm_add_ps(_r0, _tmp024a), _mm_comp_fmadd_ps(_v32, _tmp024c, _tmp024b));
-                __m128 _tmp1 = _mm_comp_fmadd_ps(_v16, _tmp135c, _mm_comp_fmadd_ps(_v2, _tmp135b, _tmp135a));
-                __m128 _tmp2 = _mm_comp_fmadd_ps(_v8, _tmp024c, _mm_comp_fmadd_ps(_v4, _tmp024b, _tmp024a));
-                __m128 _tmp3 = _mm_comp_fmadd_ps(_v4, _tmp135c, _mm_comp_fmadd_ps(_v8, _tmp135b, _tmp135a));
-                __m128 _tmp4 = _mm_comp_fmadd_ps(_v2, _tmp024c, _mm_comp_fmadd_ps(_v16, _tmp024b, _tmp024a));
-                __m128 _tmp5 = _mm_add_ps(_mm_add_ps(_r7, _tmp135a), _mm_comp_fmadd_ps(_v32, _tmp135b, _tmp135c));
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm_storeu_ps(tmp[0][m], _tmp0);
-                _mm_storeu_ps(tmp[1][m], _tmp1);
-                _mm_storeu_ps(tmp[2][m], _tmp2);
-                _mm_storeu_ps(tmp[3][m], _tmp3);
-                _mm_storeu_ps(tmp[4][m], _tmp4);
-                _mm_storeu_ps(tmp[5][m], _tmp5);
-#else
-                _mm_store_ps(tmp[0][m], _tmp0);
-                _mm_store_ps(tmp[1][m], _tmp1);
-                _mm_store_ps(tmp[2][m], _tmp2);
-                _mm_store_ps(tmp[3][m], _tmp3);
-                _mm_store_ps(tmp[4][m], _tmp4);
-                _mm_store_ps(tmp[5][m], _tmp5);
-#endif
-
-                r0 += max_jj * 8 * 4;
-                r1 += max_jj * 8 * 4;
-                r2 += max_jj * 8 * 4;
-                r3 += max_jj * 8 * 4;
-                r4 += max_jj * 8 * 4;
-                r5 += max_jj * 8 * 4;
-                r6 += max_jj * 8 * 4;
-                r7 += max_jj * 8 * 4;
-            }
-
-            float* outptr0 = top_blob.channel((i + ii) / out_elempack).row(ti * 6) + (tj * 6) * out_elempack;
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m128 _r0 = _mm_loadu_ps(tmp[m][0]);
-                __m128 _r1 = _mm_loadu_ps(tmp[m][1]);
-                __m128 _r2 = _mm_loadu_ps(tmp[m][2]);
-                __m128 _r3 = _mm_loadu_ps(tmp[m][3]);
-                __m128 _r4 = _mm_loadu_ps(tmp[m][4]);
-                __m128 _r5 = _mm_loadu_ps(tmp[m][5]);
-                __m128 _r6 = _mm_loadu_ps(tmp[m][6]);
-                __m128 _r7 = _mm_loadu_ps(tmp[m][7]);
-#else
-                __m128 _r0 = _mm_load_ps(tmp[m][0]);
-                __m128 _r1 = _mm_load_ps(tmp[m][1]);
-                __m128 _r2 = _mm_load_ps(tmp[m][2]);
-                __m128 _r3 = _mm_load_ps(tmp[m][3]);
-                __m128 _r4 = _mm_load_ps(tmp[m][4]);
-                __m128 _r5 = _mm_load_ps(tmp[m][5]);
-                __m128 _r6 = _mm_load_ps(tmp[m][6]);
-                __m128 _r7 = _mm_load_ps(tmp[m][7]);
-#endif
-
-                __m128 _tmp024a = _mm_add_ps(_r1, _r2);
-                __m128 _tmp135a = _mm_sub_ps(_r1, _r2);
-                __m128 _tmp024b = _mm_add_ps(_r3, _r4);
-                __m128 _tmp135b = _mm_sub_ps(_r3, _r4);
-                __m128 _tmp024c = _mm_add_ps(_r5, _r6);
-                __m128 _tmp135c = _mm_sub_ps(_r5, _r6);
-                __m128 _tmp0 = _mm_add_ps(_bias0, _mm_add_ps(_mm_add_ps(_r0, _tmp024a), _mm_comp_fmadd_ps(_v32, _tmp024c, _tmp024b)));
-                __m128 _tmp1 = _mm_add_ps(_bias0, _mm_comp_fmadd_ps(_v16, _tmp135c, _mm_comp_fmadd_ps(_v2, _tmp135b, _tmp135a)));
-                __m128 _tmp2 = _mm_add_ps(_bias0, _mm_comp_fmadd_ps(_v8, _tmp024c, _mm_comp_fmadd_ps(_v4, _tmp024b, _tmp024a)));
-                __m128 _tmp3 = _mm_add_ps(_bias0, _mm_comp_fmadd_ps(_v4, _tmp135c, _mm_comp_fmadd_ps(_v8, _tmp135b, _tmp135a)));
-                __m128 _tmp4 = _mm_add_ps(_bias0, _mm_comp_fmadd_ps(_v2, _tmp024c, _mm_comp_fmadd_ps(_v16, _tmp024b, _tmp024a)));
-                __m128 _tmp5 = _mm_add_ps(_bias0, _mm_add_ps(_mm_add_ps(_r7, _tmp135a), _mm_comp_fmadd_ps(_v32, _tmp135b, _tmp135c)));
-
-                if (out_elempack == 4)
-                {
-                    _mm_store_ps(outptr0, _tmp0);
-                    if (tj * 6 + 1 < outw) _mm_store_ps(outptr0 + 4, _tmp1);
-                    if (tj * 6 + 2 < outw) _mm_store_ps(outptr0 + 8, _tmp2);
-                    if (tj * 6 + 3 < outw) _mm_store_ps(outptr0 + 12, _tmp3);
-                    if (tj * 6 + 4 < outw) _mm_store_ps(outptr0 + 16, _tmp4);
-                    if (tj * 6 + 5 < outw) _mm_store_ps(outptr0 + 20, _tmp5);
-                }
-                if (out_elempack == 1)
-                {
-                    float tmp0[4];
-                    float tmp1[4];
-                    float tmp2[4];
-                    float tmp3[4];
-                    float tmp4[4];
-                    float tmp5[4];
-                    _mm_storeu_ps(tmp0, _tmp0);
-                    _mm_storeu_ps(tmp1, _tmp1);
-                    _mm_storeu_ps(tmp2, _tmp2);
-                    _mm_storeu_ps(tmp3, _tmp3);
-                    _mm_storeu_ps(tmp4, _tmp4);
-                    _mm_storeu_ps(tmp5, _tmp5);
-
-                    float* outptr1 = outptr0 + N;
-                    float* outptr2 = outptr0 + N * 2;
-                    float* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    if (tj * 6 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        outptr0[4] = tmp4[0];
-                        outptr1[4] = tmp4[1];
-                        outptr2[4] = tmp4[2];
-                        outptr3[4] = tmp4[3];
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        outptr0[5] = tmp5[0];
-                        outptr1[5] = tmp5[1];
-                        outptr2[5] = tmp5[2];
-                        outptr3[5] = tmp5[3];
-                    }
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __SSE2__
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-        float bias1 = biasptr ? biasptr[i + ii + 1] : 0.f;
-
-        float tmp[6][8][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 64 + jj * 2;
-            const float* r1 = r0 + max_jj * 2;
-            const float* r2 = r0 + max_jj * 2 * 2;
-            const float* r3 = r0 + max_jj * 2 * 3;
-            const float* r4 = r0 + max_jj * 2 * 4;
-            const float* r5 = r0 + max_jj * 2 * 5;
-            const float* r6 = r0 + max_jj * 2 * 6;
-            const float* r7 = r0 + max_jj * 2 * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float tmp024a0 = r1[0] + r2[0];
-                float tmp024a1 = r1[1] + r2[1];
-                float tmp135a0 = r1[0] - r2[0];
-                float tmp135a1 = r1[1] - r2[1];
-                float tmp024b0 = r3[0] + r4[0];
-                float tmp024b1 = r3[1] + r4[1];
-                float tmp135b0 = r3[0] - r4[0];
-                float tmp135b1 = r3[1] - r4[1];
-                float tmp024c0 = r5[0] + r6[0];
-                float tmp024c1 = r5[1] + r6[1];
-                float tmp135c0 = r5[0] - r6[0];
-                float tmp135c1 = r5[1] - r6[1];
-
-                tmp[0][m][0] = r0[0] + tmp024a0 + tmp024b0 + tmp024c0 * 32;
-                tmp[0][m][1] = r0[1] + tmp024a1 + tmp024b1 + tmp024c1 * 32;
-                tmp[1][m][0] = tmp135a0 + tmp135b0 + tmp135b0 + tmp135c0 * 16;
-                tmp[1][m][1] = tmp135a1 + tmp135b1 + tmp135b1 + tmp135c1 * 16;
-                tmp[2][m][0] = tmp024a0 + tmp024b0 * 4 + tmp024c0 * 8;
-                tmp[2][m][1] = tmp024a1 + tmp024b1 * 4 + tmp024c1 * 8;
-                tmp[3][m][0] = tmp135a0 + tmp135b0 * 8 + tmp135c0 * 4;
-                tmp[3][m][1] = tmp135a1 + tmp135b1 * 8 + tmp135c1 * 4;
-                tmp[4][m][0] = tmp024a0 + tmp024b0 * 16 + tmp024c0 + tmp024c0;
-                tmp[4][m][1] = tmp024a1 + tmp024b1 * 16 + tmp024c1 + tmp024c1;
-                tmp[5][m][0] = r7[0] + tmp135a0 + tmp135b0 * 32 + tmp135c0;
-                tmp[5][m][1] = r7[1] + tmp135a1 + tmp135b1 * 32 + tmp135c1;
-
-                r0 += max_jj * 8 * 2;
-                r1 += max_jj * 8 * 2;
-                r2 += max_jj * 8 * 2;
-                r3 += max_jj * 8 * 2;
-                r4 += max_jj * 8 * 2;
-                r5 += max_jj * 8 * 2;
-                r6 += max_jj * 8 * 2;
-                r7 += max_jj * 8 * 2;
-            }
-
-            float* outptr0 = top_blob.channel(i + ii).row(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-                float r00 = tmp[m][0][0];
-                float r01 = tmp[m][0][1];
-                float r10 = tmp[m][1][0];
-                float r11 = tmp[m][1][1];
-                float r20 = tmp[m][2][0];
-                float r21 = tmp[m][2][1];
-                float r30 = tmp[m][3][0];
-                float r31 = tmp[m][3][1];
-                float r40 = tmp[m][4][0];
-                float r41 = tmp[m][4][1];
-                float r50 = tmp[m][5][0];
-                float r51 = tmp[m][5][1];
-                float r60 = tmp[m][6][0];
-                float r61 = tmp[m][6][1];
-                float r70 = tmp[m][7][0];
-                float r71 = tmp[m][7][1];
-
-                float tmp024a0 = r10 + r20;
-                float tmp024a1 = r11 + r21;
-                float tmp135a0 = r10 - r20;
-                float tmp135a1 = r11 - r21;
-                float tmp024b0 = r30 + r40;
-                float tmp024b1 = r31 + r41;
-                float tmp135b0 = r30 - r40;
-                float tmp135b1 = r31 - r41;
-                float tmp024c0 = r50 + r60;
-                float tmp024c1 = r51 + r61;
-                float tmp135c0 = r50 - r60;
-                float tmp135c1 = r51 - r61;
-
-                float tmp00 = bias0 + r00 + tmp024a0 + tmp024b0 + tmp024c0 * 32;
-                float tmp01 = bias1 + r01 + tmp024a1 + tmp024b1 + tmp024c1 * 32;
-                float tmp10 = bias0 + tmp135a0 + tmp135b0 + tmp135b0 + tmp135c0 * 16;
-                float tmp11 = bias1 + tmp135a1 + tmp135b1 + tmp135b1 + tmp135c1 * 16;
-                float tmp20 = bias0 + tmp024a0 + tmp024b0 * 4 + tmp024c0 * 8;
-                float tmp21 = bias1 + tmp024a1 + tmp024b1 * 4 + tmp024c1 * 8;
-                float tmp30 = bias0 + tmp135a0 + tmp135b0 * 8 + tmp135c0 * 4;
-                float tmp31 = bias1 + tmp135a1 + tmp135b1 * 8 + tmp135c1 * 4;
-                float tmp40 = bias0 + tmp024a0 + tmp024b0 * 16 + tmp024c0 + tmp024c0;
-                float tmp41 = bias1 + tmp024a1 + tmp024b1 * 16 + tmp024c1 + tmp024c1;
-                float tmp50 = bias0 + r70 + tmp135a0 + tmp135b0 * 32 + tmp135c0;
-                float tmp51 = bias1 + r71 + tmp135a1 + tmp135b1 * 32 + tmp135c1;
-
-                // if (out_elempack == 1)
-                {
-                    float* outptr1 = outptr0 + N;
-
-                    outptr0[0] = tmp00;
-                    outptr1[0] = tmp01;
-                    if (tj * 6 + 1 < outw)
-                    {
-                        outptr0[1] = tmp10;
-                        outptr1[1] = tmp11;
-                    }
-                    if (tj * 6 + 2 < outw)
-                    {
-                        outptr0[2] = tmp20;
-                        outptr1[2] = tmp21;
-                    }
-                    if (tj * 6 + 3 < outw)
-                    {
-                        outptr0[3] = tmp30;
-                        outptr1[3] = tmp31;
-                    }
-                    if (tj * 6 + 4 < outw)
-                    {
-                        outptr0[4] = tmp40;
-                        outptr1[4] = tmp41;
-                    }
-                    if (tj * 6 + 5 < outw)
-                    {
-                        outptr0[5] = tmp50;
-                        outptr1[5] = tmp51;
-                    }
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        float bias0 = biasptr ? biasptr[i + ii] : 0.f;
-
-        float tmp[6][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const float* r0 = (const float*)top_tile + ii * max_jj * 64 + jj;
-            const float* r1 = r0 + max_jj;
-            const float* r2 = r0 + max_jj * 2;
-            const float* r3 = r0 + max_jj * 3;
-            const float* r4 = r0 + max_jj * 4;
-            const float* r5 = r0 + max_jj * 5;
-            const float* r6 = r0 + max_jj * 6;
-            const float* r7 = r0 + max_jj * 7;
-
-            for (int m = 0; m < 8; m++)
-            {
-                float tmp024a = r1[0] + r2[0];
-                float tmp135a = r1[0] - r2[0];
-                float tmp024b = r3[0] + r4[0];
-                float tmp135b = r3[0] - r4[0];
-                float tmp024c = r5[0] + r6[0];
-                float tmp135c = r5[0] - r6[0];
-
-                tmp[0][m] = r0[0] + tmp024a + tmp024b + tmp024c * 32;
-                tmp[1][m] = tmp135a + tmp135b + tmp135b + tmp135c * 16;
-                tmp[2][m] = tmp024a + tmp024b * 4 + tmp024c * 8;
-                tmp[3][m] = tmp135a + tmp135b * 8 + tmp135c * 4;
-                tmp[4][m] = tmp024a + tmp024b * 16 + tmp024c + tmp024c;
-                tmp[5][m] = r7[0] + tmp135a + tmp135b * 32 + tmp135c;
-
-                r0 += max_jj * 8;
-                r1 += max_jj * 8;
-                r2 += max_jj * 8;
-                r3 += max_jj * 8;
-                r4 += max_jj * 8;
-                r5 += max_jj * 8;
-                r6 += max_jj * 8;
-                r7 += max_jj * 8;
-            }
-
-            float* outptr0 = top_blob.channel(i + ii).row(ti * 6) + (tj * 6);
-
-            for (int m = 0; m < 6; m++)
-            {
-                if (ti * 6 + m >= outh)
-                    continue;
-
-                float r0 = tmp[m][0];
-                float r1 = tmp[m][1];
-                float r2 = tmp[m][2];
-                float r3 = tmp[m][3];
-                float r4 = tmp[m][4];
-                float r5 = tmp[m][5];
-                float r6 = tmp[m][6];
-                float r7 = tmp[m][7];
-
-                float tmp024a = r1 + r2;
-                float tmp135a = r1 - r2;
-                float tmp024b = r3 + r4;
-                float tmp135b = r3 - r4;
-                float tmp024c = r5 + r6;
-                float tmp135c = r5 - r6;
-
-                float tmp0 = bias0 + r0 + tmp024a + tmp024b + tmp024c * 32;
-                float tmp1 = bias0 + tmp135a + tmp135b + tmp135b + tmp135c * 16;
-                float tmp2 = bias0 + tmp024a + tmp024b * 4 + tmp024c * 8;
-                float tmp3 = bias0 + tmp135a + tmp135b * 8 + tmp135c * 4;
-                float tmp4 = bias0 + tmp024a + tmp024b * 16 + tmp024c + tmp024c;
-                float tmp5 = bias0 + r7 + tmp135a + tmp135b * 32 + tmp135c;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = tmp0;
-                    if (tj * 6 + 1 < outw) outptr0[1] = tmp1;
-                    if (tj * 6 + 2 < outw) outptr0[2] = tmp2;
-                    if (tj * 6 + 3 < outw) outptr0[3] = tmp3;
-                    if (tj * 6 + 4 < outw) outptr0[4] = tmp4;
-                    if (tj * 6 + 5 < outw) outptr0[5] = tmp5;
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd63(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, const Mat& bias, int nT, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 6n+2, winograd F(6,3)
-    int w_tiles = (outw + 5) / 6;
-    int h_tiles = (outh + 5) / 6;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 64;
-
-    // NCNN_LOGE("conv3x3s1_winograd63 %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    get_optimal_tile_mnk(M, N, K, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 4u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 4u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd63_transform_input_tile(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 4u, opt.workspace_allocator);
-
-        #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd63_transform_input_tile(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            transpose_pack_B_tile(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                gemm_transB_packed_tile(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk);
-            }
-
-            // transform output
-            conv3x3s1_winograd63_transform_output_tile(top_tile, top_blob, bias, i, max_ii, j, max_jj);
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_winograd_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_winograd_int8.h
deleted file mode 100644
index 94ea79d4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_3x3_winograd_int8.h
+++ /dev/null
@@ -1,6407 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
-#if NCNN_RUNTIME_CPU && NCNN_AVX512VNNI && __AVX512F__ && !__AVX512VNNI__
-void conv3x3s1_winograd23_int8_avx512vnni(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
-void conv3x3s1_winograd43_int8_avx512vnni(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__
-void conv3x3s1_winograd23_int8_avxvnni(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
-void conv3x3s1_winograd43_int8_avxvnni(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-void conv3x3s1_winograd23_transform_kernel_int8_avx2(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt);
-void conv3x3s1_winograd23_int8_avx2(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
-void conv3x3s1_winograd43_transform_kernel_int8_avx2(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt);
-void conv3x3s1_winograd43_int8_avx2(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__
-void conv3x3s1_winograd23_int8_xop(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
-void conv3x3s1_winograd43_int8_xop(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt);
-#endif
-#endif
-
-static void pack_A_tile_int8(const Mat& A, Mat& AT, int batch, int max_ii, int max_kk)
-{
-    const int N = max_kk * batch;
-
-    for (int b = 0; b < batch; b++)
-    {
-        short* pp = AT.row<short>(b);
-
-        int ii = 0;
-#if __SSE2__
-#if __AVX2__
-#if __AVX512F__
-        for (; ii + 15 < max_ii; ii += 16)
-        {
-            const short* p0 = (const short*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[batch];
-                pp[2] = p0[N];
-                pp[3] = p0[N + batch];
-                pp[4] = p0[N * 2];
-                pp[5] = p0[N * 2 + batch];
-                pp[6] = p0[N * 3];
-                pp[7] = p0[N * 3 + batch];
-                pp[8] = p0[N * 4];
-                pp[9] = p0[N * 4 + batch];
-                pp[10] = p0[N * 5];
-                pp[11] = p0[N * 5 + batch];
-                pp[12] = p0[N * 6];
-                pp[13] = p0[N * 6 + batch];
-                pp[14] = p0[N * 7];
-                pp[15] = p0[N * 7 + batch];
-                pp[16] = p0[N * 8];
-                pp[17] = p0[N * 8 + batch];
-                pp[18] = p0[N * 9];
-                pp[19] = p0[N * 9 + batch];
-                pp[20] = p0[N * 10];
-                pp[21] = p0[N * 10 + batch];
-                pp[22] = p0[N * 11];
-                pp[23] = p0[N * 11 + batch];
-                pp[24] = p0[N * 12];
-                pp[25] = p0[N * 12 + batch];
-                pp[26] = p0[N * 13];
-                pp[27] = p0[N * 13 + batch];
-                pp[28] = p0[N * 14];
-                pp[29] = p0[N * 14 + batch];
-                pp[30] = p0[N * 15];
-                pp[31] = p0[N * 15 + batch];
-                p0 += batch * 2;
-                pp += 32;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                pp[2] = p0[N * 2];
-                pp[3] = p0[N * 3];
-                pp[4] = p0[N * 4];
-                pp[5] = p0[N * 5];
-                pp[6] = p0[N * 6];
-                pp[7] = p0[N * 7];
-                pp[8] = p0[N * 8];
-                pp[9] = p0[N * 9];
-                pp[10] = p0[N * 10];
-                pp[11] = p0[N * 11];
-                pp[12] = p0[N * 12];
-                pp[13] = p0[N * 13];
-                pp[14] = p0[N * 14];
-                pp[15] = p0[N * 15];
-                p0 += batch;
-                pp += 16;
-            }
-        }
-#endif // __AVX512F__
-        for (; ii + 7 < max_ii; ii += 8)
-        {
-            const short* p0 = (const short*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[batch];
-                pp[2] = p0[N];
-                pp[3] = p0[N + batch];
-                pp[4] = p0[N * 2];
-                pp[5] = p0[N * 2 + batch];
-                pp[6] = p0[N * 3];
-                pp[7] = p0[N * 3 + batch];
-                pp[8] = p0[N * 4];
-                pp[9] = p0[N * 4 + batch];
-                pp[10] = p0[N * 5];
-                pp[11] = p0[N * 5 + batch];
-                pp[12] = p0[N * 6];
-                pp[13] = p0[N * 6 + batch];
-                pp[14] = p0[N * 7];
-                pp[15] = p0[N * 7 + batch];
-                p0 += batch * 2;
-                pp += 16;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                pp[2] = p0[N * 2];
-                pp[3] = p0[N * 3];
-                pp[4] = p0[N * 4];
-                pp[5] = p0[N * 5];
-                pp[6] = p0[N * 6];
-                pp[7] = p0[N * 7];
-                p0 += batch;
-                pp += 8;
-            }
-        }
-#endif // __AVX2__
-        for (; ii + 3 < max_ii; ii += 4)
-        {
-            const short* p0 = (const short*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[batch];
-                pp[2] = p0[N];
-                pp[3] = p0[N + batch];
-                pp[4] = p0[N * 2];
-                pp[5] = p0[N * 2 + batch];
-                pp[6] = p0[N * 3];
-                pp[7] = p0[N * 3 + batch];
-                p0 += batch * 2;
-                pp += 8;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                pp[2] = p0[N * 2];
-                pp[3] = p0[N * 3];
-                p0 += batch;
-                pp += 4;
-            }
-        }
-#endif // __SSE2__
-        for (; ii + 1 < max_ii; ii += 2)
-        {
-            const short* p0 = (const short*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[batch];
-                pp[2] = p0[N];
-                pp[3] = p0[N + batch];
-                p0 += batch * 2;
-                pp += 4;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[N];
-                p0 += batch;
-                pp += 2;
-            }
-        }
-        for (; ii < max_ii; ii++)
-        {
-            const short* p0 = (const short*)A + ii * N + b;
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                p0 += batch;
-                pp += 1;
-            }
-        }
-    }
-}
-
-static void transpose_pack_B_tile_int8(const Mat& B, Mat& BT, int batch, int max_jj, int max_kk, int nT)
-{
-    #pragma omp parallel for num_threads(nT)
-    for (int b = 0; b < batch; b++)
-    {
-        short* pp = BT.row<short>(b);
-
-        int jj = 0;
-#if __SSE2__
-#if defined(__x86_64__) || defined(_M_X64)
-#if __AVX512F__
-        for (; jj + 15 < max_jj; jj += 16)
-        {
-            const short* p0 = B;
-
-            int kk = 0;
-            p0 += (b * max_jj + jj) * 16;
-            for (; kk + 15 < max_kk; kk += 16)
-            {
-                __m512i _r0 = _mm512_loadu_si512((const __m512i*)p0);
-                __m512i _r1 = _mm512_loadu_si512((const __m512i*)(p0 + 32));
-                __m512i _r2 = _mm512_loadu_si512((const __m512i*)(p0 + 64));
-                __m512i _r3 = _mm512_loadu_si512((const __m512i*)(p0 + 96));
-                __m512i _r4 = _mm512_loadu_si512((const __m512i*)(p0 + 128));
-                __m512i _r5 = _mm512_loadu_si512((const __m512i*)(p0 + 160));
-                __m512i _r6 = _mm512_loadu_si512((const __m512i*)(p0 + 192));
-                __m512i _r7 = _mm512_loadu_si512((const __m512i*)(p0 + 224));
-                __m512i _tmp0 = _mm512_shuffle_i32x4(_r0, _r2, _MM_SHUFFLE(1, 0, 1, 0));
-                __m512i _tmp1 = _mm512_shuffle_i32x4(_r0, _r2, _MM_SHUFFLE(3, 2, 3, 2));
-                __m512i _tmp2 = _mm512_shuffle_i32x4(_r1, _r3, _MM_SHUFFLE(1, 0, 1, 0));
-                __m512i _tmp3 = _mm512_shuffle_i32x4(_r1, _r3, _MM_SHUFFLE(3, 2, 3, 2));
-                __m512i _tmp4 = _mm512_shuffle_i32x4(_r4, _r6, _MM_SHUFFLE(1, 0, 1, 0));
-                __m512i _tmp5 = _mm512_shuffle_i32x4(_r4, _r6, _MM_SHUFFLE(3, 2, 3, 2));
-                __m512i _tmp6 = _mm512_shuffle_i32x4(_r5, _r7, _MM_SHUFFLE(1, 0, 1, 0));
-                __m512i _tmp7 = _mm512_shuffle_i32x4(_r5, _r7, _MM_SHUFFLE(3, 2, 3, 2));
-                _r0 = _mm512_unpacklo_epi32(_tmp0, _tmp1);
-                _r1 = _mm512_unpackhi_epi32(_tmp0, _tmp1);
-                _r2 = _mm512_unpacklo_epi32(_tmp2, _tmp3);
-                _r3 = _mm512_unpackhi_epi32(_tmp2, _tmp3);
-                _r4 = _mm512_unpacklo_epi32(_tmp4, _tmp5);
-                _r5 = _mm512_unpackhi_epi32(_tmp4, _tmp5);
-                _r6 = _mm512_unpacklo_epi32(_tmp6, _tmp7);
-                _r7 = _mm512_unpackhi_epi32(_tmp6, _tmp7);
-                _tmp0 = _mm512_unpacklo_epi64(_r0, _r2);
-                _tmp1 = _mm512_unpackhi_epi64(_r0, _r2);
-                _tmp2 = _mm512_unpacklo_epi64(_r1, _r3);
-                _tmp3 = _mm512_unpackhi_epi64(_r1, _r3);
-                _tmp4 = _mm512_unpacklo_epi64(_r4, _r6);
-                _tmp5 = _mm512_unpackhi_epi64(_r4, _r6);
-                _tmp6 = _mm512_unpacklo_epi64(_r5, _r7);
-                _tmp7 = _mm512_unpackhi_epi64(_r5, _r7);
-                _r0 = _mm512_shuffle_i32x4(_tmp0, _tmp4, _MM_SHUFFLE(2, 0, 2, 0));
-                _r1 = _mm512_shuffle_i32x4(_tmp1, _tmp5, _MM_SHUFFLE(2, 0, 2, 0));
-                _r2 = _mm512_shuffle_i32x4(_tmp2, _tmp6, _MM_SHUFFLE(2, 0, 2, 0));
-                _r3 = _mm512_shuffle_i32x4(_tmp3, _tmp7, _MM_SHUFFLE(2, 0, 2, 0));
-                _r4 = _mm512_shuffle_i32x4(_tmp0, _tmp4, _MM_SHUFFLE(3, 1, 3, 1));
-                _r5 = _mm512_shuffle_i32x4(_tmp1, _tmp5, _MM_SHUFFLE(3, 1, 3, 1));
-                _r6 = _mm512_shuffle_i32x4(_tmp2, _tmp6, _MM_SHUFFLE(3, 1, 3, 1));
-                _r7 = _mm512_shuffle_i32x4(_tmp3, _tmp7, _MM_SHUFFLE(3, 1, 3, 1));
-                _mm512_storeu_si512((__m512i*)pp, _r0);
-                _mm512_storeu_si512((__m512i*)(pp + 32), _r1);
-                _mm512_storeu_si512((__m512i*)(pp + 64), _r2);
-                _mm512_storeu_si512((__m512i*)(pp + 96), _r3);
-                _mm512_storeu_si512((__m512i*)(pp + 128), _r4);
-                _mm512_storeu_si512((__m512i*)(pp + 160), _r5);
-                _mm512_storeu_si512((__m512i*)(pp + 192), _r6);
-                _mm512_storeu_si512((__m512i*)(pp + 224), _r7);
-                p0 += max_jj * batch * 16;
-                pp += 256;
-            }
-            p0 -= (b * max_jj + jj) * 16;
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                __m512i _r0 = _mm512_loadu_si512((const __m512i*)p0);
-                __m512i _r1 = _mm512_loadu_si512((const __m512i*)(p0 + 32));
-                __m512i _r2 = _mm512_loadu_si512((const __m512i*)(p0 + 64));
-                __m512i _r3 = _mm512_loadu_si512((const __m512i*)(p0 + 96));
-                __m512i _tmp0 = _mm512_shuffle_i32x4(_r0, _r1, _MM_SHUFFLE(2, 0, 2, 0));
-                __m512i _tmp1 = _mm512_shuffle_i32x4(_r0, _r1, _MM_SHUFFLE(3, 1, 3, 1));
-                __m512i _tmp2 = _mm512_shuffle_i32x4(_r2, _r3, _MM_SHUFFLE(2, 0, 2, 0));
-                __m512i _tmp3 = _mm512_shuffle_i32x4(_r2, _r3, _MM_SHUFFLE(3, 1, 3, 1));
-                _r0 = _mm512_unpacklo_epi32(_tmp0, _tmp1);
-                _r1 = _mm512_unpackhi_epi32(_tmp0, _tmp1);
-                _r2 = _mm512_unpacklo_epi32(_tmp2, _tmp3);
-                _r3 = _mm512_unpackhi_epi32(_tmp2, _tmp3);
-                _tmp0 = _mm512_permutex_epi64(_r0, _MM_SHUFFLE(3, 1, 2, 0));
-                _tmp1 = _mm512_permutex_epi64(_r1, _MM_SHUFFLE(3, 1, 2, 0));
-                _tmp2 = _mm512_permutex_epi64(_r2, _MM_SHUFFLE(3, 1, 2, 0));
-                _tmp3 = _mm512_permutex_epi64(_r3, _MM_SHUFFLE(3, 1, 2, 0));
-                _r0 = _mm512_shuffle_i32x4(_tmp0, _tmp2, _MM_SHUFFLE(2, 0, 2, 0));
-                _r1 = _mm512_shuffle_i32x4(_tmp0, _tmp2, _MM_SHUFFLE(3, 1, 3, 1));
-                _r2 = _mm512_shuffle_i32x4(_tmp1, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-                _r3 = _mm512_shuffle_i32x4(_tmp1, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-                _mm512_storeu_si512((__m512i*)pp, _r0);
-                _mm512_storeu_si512((__m512i*)(pp + 32), _r1);
-                _mm512_storeu_si512((__m512i*)(pp + 64), _r2);
-                _mm512_storeu_si512((__m512i*)(pp + 96), _r3);
-                p0 += max_jj * batch * 8;
-                pp += 128;
-            }
-            p0 -= (b * max_jj + jj) * 8;
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m512i _r0 = _mm512_loadu_si512((const __m512i*)p0);
-                _mm512_storeu_si512((__m512i*)pp, _r0);
-                p0 += max_jj * batch * 2;
-                pp += 32;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                __m256i _r0 = _mm256_loadu_si256((const __m256i*)p0);
-                _mm256_store_si256((__m256i*)pp, _r0);
-                p0 += max_jj * batch;
-                pp += 16;
-            }
-        }
-#endif // __AVX512F__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            const short* p0 = B;
-
-            int kk = 0;
-#if __AVX512F__
-            p0 += (b * max_jj + jj) * 16;
-            for (; kk + 15 < max_kk; kk += 16)
-            {
-                __m512i _r0 = _mm512_loadu_si512((const __m512i*)p0);
-                __m512i _r1 = _mm512_loadu_si512((const __m512i*)(p0 + 32));
-                __m512i _r2 = _mm512_loadu_si512((const __m512i*)(p0 + 64));
-                __m512i _r3 = _mm512_loadu_si512((const __m512i*)(p0 + 96));
-                __m512i _tmp0 = _mm512_shuffle_i32x4(_r0, _r2, _MM_SHUFFLE(1, 0, 1, 0));
-                __m512i _tmp1 = _mm512_shuffle_i32x4(_r0, _r2, _MM_SHUFFLE(3, 2, 3, 2));
-                __m512i _tmp2 = _mm512_shuffle_i32x4(_r1, _r3, _MM_SHUFFLE(1, 0, 1, 0));
-                __m512i _tmp3 = _mm512_shuffle_i32x4(_r1, _r3, _MM_SHUFFLE(3, 2, 3, 2));
-                _r0 = _mm512_unpacklo_epi32(_tmp0, _tmp1);
-                _r1 = _mm512_unpackhi_epi32(_tmp0, _tmp1);
-                _r2 = _mm512_unpacklo_epi32(_tmp2, _tmp3);
-                _r3 = _mm512_unpackhi_epi32(_tmp2, _tmp3);
-                _tmp0 = _mm512_unpacklo_epi64(_r0, _r2);
-                _tmp1 = _mm512_unpackhi_epi64(_r0, _r2);
-                _tmp2 = _mm512_unpacklo_epi64(_r1, _r3);
-                _tmp3 = _mm512_unpackhi_epi64(_r1, _r3);
-                _r0 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-                _r1 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-                _r2 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-                _r3 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-                _mm512_storeu_si512((__m512i*)pp, _r0);
-                _mm512_storeu_si512((__m512i*)(pp + 32), _r1);
-                _mm512_storeu_si512((__m512i*)(pp + 64), _r2);
-                _mm512_storeu_si512((__m512i*)(pp + 96), _r3);
-                p0 += max_jj * batch * 16;
-                pp += 128;
-            }
-            p0 -= (b * max_jj + jj) * 16;
-#endif // __AVX512F__
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-#if __AVX__
-                __m256 _r0 = _mm256_loadu_ps((const float*)p0);
-                __m256 _r1 = _mm256_loadu_ps((const float*)(p0 + 16));
-                __m256 _r2 = _mm256_loadu_ps((const float*)(p0 + 32));
-                __m256 _r3 = _mm256_loadu_ps((const float*)(p0 + 48));
-                __m256 _tmp0 = _mm256_permute2f128_ps(_r0, _r2, _MM_SHUFFLE(0, 2, 0, 0));
-                __m256 _tmp1 = _mm256_permute2f128_ps(_r0, _r2, _MM_SHUFFLE(0, 3, 0, 1));
-                __m256 _tmp2 = _mm256_permute2f128_ps(_r1, _r3, _MM_SHUFFLE(0, 2, 0, 0));
-                __m256 _tmp3 = _mm256_permute2f128_ps(_r1, _r3, _MM_SHUFFLE(0, 3, 0, 1));
-                _r0 = _mm256_unpacklo_ps(_tmp0, _tmp1);
-                _r1 = _mm256_unpackhi_ps(_tmp0, _tmp1);
-                _r2 = _mm256_unpacklo_ps(_tmp2, _tmp3);
-                _r3 = _mm256_unpackhi_ps(_tmp2, _tmp3);
-                _tmp0 = _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(_r0), _mm256_castps_pd(_r2)));
-                _tmp1 = _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(_r0), _mm256_castps_pd(_r2)));
-                _tmp2 = _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(_r1), _mm256_castps_pd(_r3)));
-                _tmp3 = _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(_r1), _mm256_castps_pd(_r3)));
-                _mm256_storeu_ps((float*)pp, _tmp0);
-                _mm256_storeu_ps((float*)(pp + 16), _tmp1);
-                _mm256_storeu_ps((float*)(pp + 32), _tmp2);
-                _mm256_storeu_ps((float*)(pp + 48), _tmp3);
-#else
-                __m128i _r0 = _mm_load_si128((const __m128i*)p0);
-                __m128i _r1 = _mm_load_si128((const __m128i*)(p0 + 8));
-                __m128i _r2 = _mm_load_si128((const __m128i*)(p0 + 8 * 2));
-                __m128i _r3 = _mm_load_si128((const __m128i*)(p0 + 8 * 3));
-                __m128i _r4 = _mm_load_si128((const __m128i*)(p0 + 8 * 4));
-                __m128i _r5 = _mm_load_si128((const __m128i*)(p0 + 8 * 5));
-                __m128i _r6 = _mm_load_si128((const __m128i*)(p0 + 8 * 6));
-                __m128i _r7 = _mm_load_si128((const __m128i*)(p0 + 8 * 7));
-                transpose4x8_epi32(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-                _mm_store_si128((__m128i*)pp, _r0);
-                _mm_store_si128((__m128i*)(pp + 8), _r1);
-                _mm_store_si128((__m128i*)(pp + 8 * 2), _r2);
-                _mm_store_si128((__m128i*)(pp + 8 * 3), _r3);
-                _mm_store_si128((__m128i*)(pp + 8 * 4), _r4);
-                _mm_store_si128((__m128i*)(pp + 8 * 5), _r5);
-                _mm_store_si128((__m128i*)(pp + 8 * 6), _r6);
-                _mm_store_si128((__m128i*)(pp + 8 * 7), _r7);
-#endif // __AVX__
-                p0 += max_jj * batch * 8;
-                pp += 64;
-            }
-            p0 -= (b * max_jj + jj) * 8;
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-#if __AVX__
-                __m256 _r0 = _mm256_loadu_ps((const float*)p0);
-                _mm256_storeu_ps((float*)pp, _r0);
-#else
-                __m128i _r0 = _mm_loadu_si128((const __m128i*)p0);
-                __m128i _r1 = _mm_loadu_si128((const __m128i*)(p0 + 8));
-                _mm_store_si128((__m128i*)pp, _r0);
-                _mm_store_si128((__m128i*)(pp + 8), _r1);
-#endif // __AVX__
-                p0 += max_jj * batch * 2;
-                pp += 16;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                __m128i _r0 = _mm_loadu_si128((const __m128i*)p0);
-                _mm_store_si128((__m128i*)pp, _r0);
-                p0 += max_jj * batch;
-                pp += 8;
-            }
-        }
-#endif // defined(__x86_64__) || defined(_M_X64)
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            const short* p0 = B;
-
-            int kk = 0;
-#if __AVX512F__
-            p0 += (b * max_jj + jj) * 16;
-            for (; kk + 15 < max_kk; kk += 16)
-            {
-                __m512i _r0 = _mm512_loadu_si512((const __m512i*)p0);
-                __m512i _r1 = _mm512_loadu_si512((const __m512i*)(p0 + 32));
-                __m512i _tmp0 = _mm512_shuffle_i32x4(_r0, _r1, _MM_SHUFFLE(1, 0, 1, 0));
-                __m512i _tmp1 = _mm512_shuffle_i32x4(_r0, _r1, _MM_SHUFFLE(3, 2, 3, 2));
-                _r0 = _mm512_unpacklo_epi32(_tmp0, _tmp1);
-                _r1 = _mm512_unpackhi_epi32(_tmp0, _tmp1);
-                _r0 = _mm512_permutex_epi64(_r0, _MM_SHUFFLE(3, 1, 2, 0));
-                _r1 = _mm512_permutex_epi64(_r1, _MM_SHUFFLE(3, 1, 2, 0));
-                _tmp0 = _mm512_shuffle_i32x4(_r0, _r1, _MM_SHUFFLE(1, 0, 1, 0));
-                _tmp1 = _mm512_shuffle_i32x4(_r0, _r1, _MM_SHUFFLE(3, 2, 3, 2));
-                _r0 = _mm512_unpacklo_epi64(_tmp0, _tmp1);
-                _r1 = _mm512_unpackhi_epi64(_tmp0, _tmp1);
-                _mm512_storeu_si512((__m512i*)pp, _r0);
-                _mm512_storeu_si512((__m512i*)(pp + 32), _r1);
-                p0 += max_jj * batch * 16;
-                pp += 64;
-            }
-            p0 -= (b * max_jj + jj) * 16;
-#endif // __AVX512F__
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                __m128i _r0 = _mm_load_si128((const __m128i*)p0);
-                __m128i _r1 = _mm_load_si128((const __m128i*)(p0 + 8));
-                __m128i _r2 = _mm_load_si128((const __m128i*)(p0 + 8 * 2));
-                __m128i _r3 = _mm_load_si128((const __m128i*)(p0 + 8 * 3));
-                transpose4x4_epi32(_r0, _r1, _r2, _r3);
-                _mm_storeu_si128((__m128i*)pp, _r0);
-                _mm_storeu_si128((__m128i*)(pp + 8), _r1);
-                _mm_storeu_si128((__m128i*)(pp + 8 * 2), _r2);
-                _mm_storeu_si128((__m128i*)(pp + 8 * 3), _r3);
-                p0 += max_jj * batch * 8;
-                pp += 32;
-            }
-            p0 -= (b * max_jj + jj) * 8;
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _r0 = _mm_loadu_si128((const __m128i*)p0);
-                _mm_storeu_si128((__m128i*)pp, _r0);
-                p0 += max_jj * batch * 2;
-                pp += 8;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                pp[2] = p0[2];
-                pp[3] = p0[3];
-                p0 += max_jj * batch;
-                pp += 4;
-            }
-        }
-#endif // __SSE2__
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            const short* p0 = B;
-
-            int kk = 0;
-#if __SSE2__
-#if __AVX512F__
-            p0 += (b * max_jj + jj) * 16;
-            for (; kk + 15 < max_kk; kk += 16)
-            {
-                __m256i _r0 = _mm256_load_si256((const __m256i*)p0);
-                __m256i _r1 = _mm256_load_si256((const __m256i*)(p0 + 16));
-                transpose8x2_epi32(_r0, _r1);
-                _mm256_storeu_si256((__m256i*)pp, _r0);
-                _mm256_storeu_si256((__m256i*)(pp + 16), _r1);
-                p0 += max_jj * batch * 16;
-                pp += 32;
-            }
-            p0 -= (b * max_jj + jj) * 16;
-#endif // __AVX512F__
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                __m128i _r0 = _mm_load_si128((const __m128i*)p0);
-                __m128i _r1 = _mm_load_si128((const __m128i*)(p0 + 8));
-                __m128i _tmp0 = _mm_unpacklo_epi32(_r0, _r1);
-                __m128i _tmp1 = _mm_unpackhi_epi32(_r0, _r1);
-                _mm_storeu_si128((__m128i*)pp, _tmp0);
-                _mm_storeu_si128((__m128i*)(pp + 8), _tmp1);
-                p0 += max_jj * batch * 8;
-                pp += 16;
-            }
-            p0 -= (b * max_jj + jj) * 8;
-#endif // __SSE2__
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                pp[2] = p0[2];
-                pp[3] = p0[3];
-                p0 += max_jj * batch * 2;
-                pp += 4;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                p0 += max_jj * batch;
-                pp += 2;
-            }
-        }
-        for (; jj < max_jj; jj++)
-        {
-            const short* p0 = B;
-
-            int kk = 0;
-#if __SSE2__
-#if __AVX512F__
-            p0 += (b * max_jj + jj) * 16;
-            for (; kk + 15 < max_kk; kk += 16)
-            {
-                __m256i _r0 = _mm256_load_si256((const __m256i*)p0);
-                _mm256_storeu_si256((__m256i*)pp, _r0);
-                p0 += max_jj * batch * 16;
-                pp += 16;
-            }
-            p0 -= (b * max_jj + jj) * 16;
-#endif // __AVX512F__
-            p0 += (b * max_jj + jj) * 8;
-            for (; kk + 7 < max_kk; kk += 8)
-            {
-                __m128i _r0 = _mm_load_si128((const __m128i*)p0);
-                _mm_storeu_si128((__m128i*)pp, _r0);
-                p0 += max_jj * batch * 8;
-                pp += 8;
-            }
-            p0 -= (b * max_jj + jj) * 8;
-#endif // __SSE2__
-            p0 += (b * max_jj + jj) * 2;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                p0 += max_jj * batch * 2;
-                pp += 2;
-            }
-            p0 -= (b * max_jj + jj) * 2;
-            p0 += (b * max_jj + jj);
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                p0 += max_jj * batch;
-                pp += 1;
-            }
-        }
-    }
-}
-
-static void gemm_transB_packed_tile_int8(const Mat& AT_tile, const Mat& BT_tile, Mat& top_blob, int batch, int max_ii, int max_jj, int k, int max_kk, bool k_end)
-{
-    int* outptr = top_blob;
-
-    int ii = 0;
-#if __SSE2__
-#if __AVX2__
-#if __AVX512F__
-    for (; ii + 15 < max_ii; ii += 16)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const short* pAT = AT_tile.row<const short>(b) + max_kk * ii;
-            const short* pB = BT_tile.row<const short>(b);
-
-            int jj = 0;
-#if defined(__x86_64__) || defined(_M_X64)
-            for (; jj + 15 < max_jj; jj += 16)
-            {
-                const short* pA = pAT;
-
-                __m512i _sum0;
-                __m512i _sum1;
-                __m512i _sum2;
-                __m512i _sum3;
-                __m512i _sum4;
-                __m512i _sum5;
-                __m512i _sum6;
-                __m512i _sum7;
-                __m512i _sum8;
-                __m512i _sum9;
-                __m512i _suma;
-                __m512i _sumb;
-                __m512i _sumc;
-                __m512i _sumd;
-                __m512i _sume;
-                __m512i _sumf;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm512_setzero_si512();
-                    _sum1 = _mm512_setzero_si512();
-                    _sum2 = _mm512_setzero_si512();
-                    _sum3 = _mm512_setzero_si512();
-                    _sum4 = _mm512_setzero_si512();
-                    _sum5 = _mm512_setzero_si512();
-                    _sum6 = _mm512_setzero_si512();
-                    _sum7 = _mm512_setzero_si512();
-                    _sum8 = _mm512_setzero_si512();
-                    _sum9 = _mm512_setzero_si512();
-                    _suma = _mm512_setzero_si512();
-                    _sumb = _mm512_setzero_si512();
-                    _sumc = _mm512_setzero_si512();
-                    _sumd = _mm512_setzero_si512();
-                    _sume = _mm512_setzero_si512();
-                    _sumf = _mm512_setzero_si512();
-                }
-                else
-                {
-                    _sum0 = _mm512_load_si512((const __m512i*)outptr);
-                    _sum1 = _mm512_load_si512((const __m512i*)(outptr + 16));
-                    _sum2 = _mm512_load_si512((const __m512i*)(outptr + 32));
-                    _sum3 = _mm512_load_si512((const __m512i*)(outptr + 48));
-                    _sum4 = _mm512_load_si512((const __m512i*)(outptr + 64));
-                    _sum5 = _mm512_load_si512((const __m512i*)(outptr + 80));
-                    _sum6 = _mm512_load_si512((const __m512i*)(outptr + 96));
-                    _sum7 = _mm512_load_si512((const __m512i*)(outptr + 112));
-                    _sum8 = _mm512_load_si512((const __m512i*)(outptr + 128));
-                    _sum9 = _mm512_load_si512((const __m512i*)(outptr + 128 + 16));
-                    _suma = _mm512_load_si512((const __m512i*)(outptr + 128 + 32));
-                    _sumb = _mm512_load_si512((const __m512i*)(outptr + 128 + 48));
-                    _sumc = _mm512_load_si512((const __m512i*)(outptr + 128 + 64));
-                    _sumd = _mm512_load_si512((const __m512i*)(outptr + 128 + 80));
-                    _sume = _mm512_load_si512((const __m512i*)(outptr + 128 + 96));
-                    _sumf = _mm512_load_si512((const __m512i*)(outptr + 128 + 112));
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m512i _pA0 = _mm512_loadu_si512((const __m512i*)pA);
-                    __m512i _pB0 = _mm512_loadu_si512((const __m512i*)pB);
-                    __m512i _pA1 = _mm512_shuffle_i32x4(_pA0, _pA0, _MM_SHUFFLE(2, 1, 0, 3));
-                    __m512i _pA2 = _mm512_shuffle_i32x4(_pA0, _pA0, _MM_SHUFFLE(1, 0, 3, 2));
-                    __m512i _pA3 = _mm512_shuffle_i32x4(_pA0, _pA0, _MM_SHUFFLE(0, 3, 2, 1));
-                    __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-                    __m512i _pB2 = _mm512_shuffle_epi32(_pB0, _MM_PERM_BADC);
-                    __m512i _pB3 = _mm512_shuffle_epi32(_pB0, _MM_PERM_CBAD);
-
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _pA0, _pB0);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _pA0, _pB1);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _pA0, _pB2);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _pA0, _pB3);
-                    _sum4 = _mm512_dpwssd_epi32(_sum4, _pA1, _pB0);
-                    _sum5 = _mm512_dpwssd_epi32(_sum5, _pA1, _pB1);
-                    _sum6 = _mm512_dpwssd_epi32(_sum6, _pA1, _pB2);
-                    _sum7 = _mm512_dpwssd_epi32(_sum7, _pA1, _pB3);
-                    _sum8 = _mm512_dpwssd_epi32(_sum8, _pA2, _pB0);
-                    _sum9 = _mm512_dpwssd_epi32(_sum9, _pA2, _pB1);
-                    _suma = _mm512_dpwssd_epi32(_suma, _pA2, _pB2);
-                    _sumb = _mm512_dpwssd_epi32(_sumb, _pA2, _pB3);
-                    _sumc = _mm512_dpwssd_epi32(_sumc, _pA3, _pB0);
-                    _sumd = _mm512_dpwssd_epi32(_sumd, _pA3, _pB1);
-                    _sume = _mm512_dpwssd_epi32(_sume, _pA3, _pB2);
-                    _sumf = _mm512_dpwssd_epi32(_sumf, _pA3, _pB3);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA0, _pB0));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA0, _pB1));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_pA0, _pB2));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_pA0, _pB3));
-                    _sum4 = _mm512_add_epi32(_sum4, _mm512_madd_epi16(_pA1, _pB0));
-                    _sum5 = _mm512_add_epi32(_sum5, _mm512_madd_epi16(_pA1, _pB1));
-                    _sum6 = _mm512_add_epi32(_sum6, _mm512_madd_epi16(_pA1, _pB2));
-                    _sum7 = _mm512_add_epi32(_sum7, _mm512_madd_epi16(_pA1, _pB3));
-                    _sum8 = _mm512_add_epi32(_sum8, _mm512_madd_epi16(_pA2, _pB0));
-                    _sum9 = _mm512_add_epi32(_sum9, _mm512_madd_epi16(_pA2, _pB1));
-                    _suma = _mm512_add_epi32(_suma, _mm512_madd_epi16(_pA2, _pB2));
-                    _sumb = _mm512_add_epi32(_sumb, _mm512_madd_epi16(_pA2, _pB3));
-                    _sumc = _mm512_add_epi32(_sumc, _mm512_madd_epi16(_pA3, _pB0));
-                    _sumd = _mm512_add_epi32(_sumd, _mm512_madd_epi16(_pA3, _pB1));
-                    _sume = _mm512_add_epi32(_sume, _mm512_madd_epi16(_pA3, _pB2));
-                    _sumf = _mm512_add_epi32(_sumf, _mm512_madd_epi16(_pA3, _pB3));
-#endif
-
-                    pA += 32;
-                    pB += 32;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m512i _pA0 = _mm512_cvtepi16_epi32(_mm256_load_si256((const __m256i*)pA));
-                    __m512i _pB0 = _mm512_cvtepi16_epi32(_mm256_load_si256((const __m256i*)pB));
-
-                    __m512i _pA1 = _mm512_shuffle_i32x4(_pA0, _pA0, _MM_SHUFFLE(2, 1, 0, 3));
-                    __m512i _pA2 = _mm512_shuffle_i32x4(_pA0, _pA0, _MM_SHUFFLE(1, 0, 3, 2));
-                    __m512i _pA3 = _mm512_shuffle_i32x4(_pA0, _pA0, _MM_SHUFFLE(0, 3, 2, 1));
-
-                    __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-                    __m512i _pB2 = _mm512_permutex_epi64(_pB0, _MM_SHUFFLE(2, 3, 0, 1));
-                    __m512i _pB3 = _mm512_shuffle_epi32(_pB0, _MM_PERM_CBAD);
-
-                    __m512i _s0 = _mm512_mullo_epi32(_pA0, _pB0);
-                    __m512i _s1 = _mm512_mullo_epi32(_pA0, _pB1);
-                    __m512i _s2 = _mm512_mullo_epi32(_pA0, _pB2);
-                    __m512i _s3 = _mm512_mullo_epi32(_pA0, _pB3);
-                    __m512i _s4 = _mm512_mullo_epi32(_pA1, _pB0);
-                    __m512i _s5 = _mm512_mullo_epi32(_pA1, _pB1);
-                    __m512i _s6 = _mm512_mullo_epi32(_pA1, _pB2);
-                    __m512i _s7 = _mm512_mullo_epi32(_pA1, _pB3);
-                    __m512i _s8 = _mm512_mullo_epi32(_pA2, _pB0);
-                    __m512i _s9 = _mm512_mullo_epi32(_pA2, _pB1);
-                    __m512i _sa = _mm512_mullo_epi32(_pA2, _pB2);
-                    __m512i _sb = _mm512_mullo_epi32(_pA2, _pB3);
-                    __m512i _sc = _mm512_mullo_epi32(_pA3, _pB0);
-                    __m512i _sd = _mm512_mullo_epi32(_pA3, _pB1);
-                    __m512i _se = _mm512_mullo_epi32(_pA3, _pB2);
-                    __m512i _sf = _mm512_mullo_epi32(_pA3, _pB3);
-                    _sum0 = _mm512_add_epi32(_sum0, _s0);
-                    _sum1 = _mm512_add_epi32(_sum1, _s1);
-                    _sum2 = _mm512_add_epi32(_sum2, _s2);
-                    _sum3 = _mm512_add_epi32(_sum3, _s3);
-                    _sum4 = _mm512_add_epi32(_sum4, _s4);
-                    _sum5 = _mm512_add_epi32(_sum5, _s5);
-                    _sum6 = _mm512_add_epi32(_sum6, _s6);
-                    _sum7 = _mm512_add_epi32(_sum7, _s7);
-                    _sum8 = _mm512_add_epi32(_sum8, _s8);
-                    _sum9 = _mm512_add_epi32(_sum9, _s9);
-                    _suma = _mm512_add_epi32(_suma, _sa);
-                    _sumb = _mm512_add_epi32(_sumb, _sb);
-                    _sumc = _mm512_add_epi32(_sumc, _sc);
-                    _sumd = _mm512_add_epi32(_sumd, _sd);
-                    _sume = _mm512_add_epi32(_sume, _se);
-                    _sumf = _mm512_add_epi32(_sumf, _sf);
-
-                    pA += 16;
-                    pB += 16;
-                }
-
-                if (k_end)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77 88 99 aa bb cc dd ee ff
-                    //      01 12 23 30 45 56 67 74 89 9a ab b8 cd de ef fc
-                    //      02 13 20 31 46 57 64 75 8a 9b a8 b9 ce df ec fd
-                    //      03 10 21 32 47 54 65 76 8b 98 a9 ba cf dc ed fe
-                    //      c0 d1 e2 f3 04 15 26 37 48 59 6a 7b 8c 9d ae bf
-                    //      c1 d2 e3 f0 05 16 27 34 49 5a 6b 78 8d 9e af bc
-                    //      c2 d3 e0 f1 06 17 24 35 4a 5b 68 79 8e 9f ac bd
-                    //      c3 d0 e1 f2 07 14 25 36 4b 58 69 7a 8f 9c ad be
-                    //      80 91 a2 b3 c4 d5 e6 f7 08 19 2a 3b 4c 5d 6e 7f
-                    //      81 92 a3 b0 c5 d6 e7 f4 09 1a 2b 38 4d 5e 6f 7c
-                    //      82 93 a0 b1 c6 d7 e4 f5 0a 1b 28 39 4e 5f 6c 7d
-                    //      83 90 a1 b2 c7 d4 e5 f6 0b 18 29 3a 4f 5c 6d 7e
-                    //      40 51 62 73 84 95 a6 b7 c8 d9 ea fb 0c 1d 2e 3f
-                    //      41 52 63 70 85 96 a7 b4 c9 da eb f8 0d 1e 2f 3c
-                    //      42 53 60 71 86 97 a4 b5 ca db e8 f9 0e 1f 2c 3d
-                    //      43 50 61 72 87 94 a5 b6 cb d8 e9 fa 0f 1c 2d 3e
-                    // to
-                    //      00 10 20 30 44 54 64 74 88 98 a8 b8 cc dc ec fc
-                    //      01 11 21 31 45 55 65 75 89 99 a9 b9 cd dd ed fd
-                    //      02 12 22 32 46 56 66 76 8a 9a aa ba ce de ee fe
-                    //      03 13 23 33 47 57 67 77 8b 9b ab bb cf df ef ff
-                    //      c0 d0 e0 f0 04 14 24 34 48 58 68 78 8c 9c ac bc
-                    //      c1 d1 e1 f1 05 15 25 35 49 59 69 79 8d 9d ad bd
-                    //      c2 d2 e2 f2 06 16 26 36 4a 5a 6a 7a 8e 9e ae be
-                    //      c3 d3 e3 f3 07 17 27 37 4b 5b 6b 7b 8f 9f af bf
-                    //      80 90 a0 b0 c4 d4 e4 f4 08 18 28 38 4c 5c 6c 7c
-                    //      81 91 a1 b1 c5 d5 e5 f5 09 19 29 39 4d 5d 6d 7d
-                    //      82 92 a2 b2 c6 d6 e6 f6 0a 1a 2a 3a 4e 5e 6e 7e
-                    //      83 93 a3 b3 c7 d7 e7 f7 0b 1b 2b 3b 4f 5f 6f 7f
-                    //      40 50 60 70 84 94 a4 b4 c8 d8 e8 f8 0c 1c 2c 3c
-                    //      41 51 61 71 85 95 a5 b5 c9 d9 e9 f9 0d 1d 2d 3d
-                    //      42 52 62 72 86 96 a6 b6 ca da ea fa 0e 1e 2e 3e
-                    //      43 53 63 73 87 97 a7 b7 cb db eb fb 0f 1f 2f 3f
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum2 = _mm512_shuffle_epi32(_sum2, _MM_PERM_BADC);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_ADCB);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum6 = _mm512_shuffle_epi32(_sum6, _MM_PERM_BADC);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_ADCB);
-                        _sum9 = _mm512_shuffle_epi32(_sum9, _MM_PERM_CBAD);
-                        _suma = _mm512_shuffle_epi32(_suma, _MM_PERM_BADC);
-                        _sumb = _mm512_shuffle_epi32(_sumb, _MM_PERM_ADCB);
-                        _sumd = _mm512_shuffle_epi32(_sumd, _MM_PERM_CBAD);
-                        _sume = _mm512_shuffle_epi32(_sume, _MM_PERM_BADC);
-                        _sumf = _mm512_shuffle_epi32(_sumf, _MM_PERM_ADCB);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        __m512i _tmp4 = _mm512_unpacklo_epi32(_sum4, _sum7);
-                        __m512i _tmp5 = _mm512_unpackhi_epi32(_sum4, _sum7);
-                        __m512i _tmp6 = _mm512_unpacklo_epi32(_sum6, _sum5);
-                        __m512i _tmp7 = _mm512_unpackhi_epi32(_sum6, _sum5);
-                        __m512i _tmp8 = _mm512_unpacklo_epi32(_sum8, _sumb);
-                        __m512i _tmp9 = _mm512_unpackhi_epi32(_sum8, _sumb);
-                        __m512i _tmpa = _mm512_unpacklo_epi32(_suma, _sum9);
-                        __m512i _tmpb = _mm512_unpackhi_epi32(_suma, _sum9);
-                        __m512i _tmpc = _mm512_unpacklo_epi32(_sumc, _sumf);
-                        __m512i _tmpd = _mm512_unpackhi_epi32(_sumc, _sumf);
-                        __m512i _tmpe = _mm512_unpacklo_epi32(_sume, _sumd);
-                        __m512i _tmpf = _mm512_unpackhi_epi32(_sume, _sumd);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm512_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm512_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm512_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm512_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum8 = _mm512_unpacklo_epi64(_tmp8, _tmpa);
-                        _sum9 = _mm512_unpackhi_epi64(_tmp8, _tmpa);
-                        _suma = _mm512_unpacklo_epi64(_tmpb, _tmp9);
-                        _sumb = _mm512_unpackhi_epi64(_tmpb, _tmp9);
-                        _sumc = _mm512_unpacklo_epi64(_tmpc, _tmpe);
-                        _sumd = _mm512_unpackhi_epi64(_tmpc, _tmpe);
-                        _sume = _mm512_unpacklo_epi64(_tmpf, _tmpd);
-                        _sumf = _mm512_unpackhi_epi64(_tmpf, _tmpd);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_CBAD);
-                        _sum9 = _mm512_shuffle_epi32(_sum9, _MM_PERM_CBAD);
-                        _sumb = _mm512_shuffle_epi32(_sumb, _MM_PERM_CBAD);
-                        _sumd = _mm512_shuffle_epi32(_sumd, _MM_PERM_CBAD);
-                        _sumf = _mm512_shuffle_epi32(_sumf, _MM_PERM_CBAD);
-                    }
-
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sumc, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum4, _sum0, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum8, _sum4, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sumc, _sum8, _MM_SHUFFLE(1, 3, 1, 3));
-                    __m512i _tmp4 = _mm512_shuffle_i32x4(_sum1, _sumd, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp5 = _mm512_shuffle_i32x4(_sum5, _sum1, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp6 = _mm512_shuffle_i32x4(_sum9, _sum5, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmp7 = _mm512_shuffle_i32x4(_sumd, _sum9, _MM_SHUFFLE(1, 3, 1, 3));
-                    __m512i _tmp8 = _mm512_shuffle_i32x4(_sum2, _sume, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp9 = _mm512_shuffle_i32x4(_sum6, _sum2, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmpa = _mm512_shuffle_i32x4(_suma, _sum6, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmpb = _mm512_shuffle_i32x4(_sume, _suma, _MM_SHUFFLE(1, 3, 1, 3));
-                    __m512i _tmpc = _mm512_shuffle_i32x4(_sum3, _sumf, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmpd = _mm512_shuffle_i32x4(_sum7, _sum3, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmpe = _mm512_shuffle_i32x4(_sumb, _sum7, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmpf = _mm512_shuffle_i32x4(_sumf, _sumb, _MM_SHUFFLE(1, 3, 1, 3));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp2, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp4, _tmp6, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp8, _tmpa, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum3 = _mm512_shuffle_i32x4(_tmpc, _tmpe, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum4 = _mm512_shuffle_i32x4(_tmp1, _tmp3, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum5 = _mm512_shuffle_i32x4(_tmp5, _tmp7, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum6 = _mm512_shuffle_i32x4(_tmp9, _tmpb, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum7 = _mm512_shuffle_i32x4(_tmpd, _tmpf, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum8 = _mm512_shuffle_i32x4(_tmp2, _tmp0, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum9 = _mm512_shuffle_i32x4(_tmp6, _tmp4, _MM_SHUFFLE(3, 1, 2, 0));
-                    _suma = _mm512_shuffle_i32x4(_tmpa, _tmp8, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sumb = _mm512_shuffle_i32x4(_tmpe, _tmpc, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sumc = _mm512_shuffle_i32x4(_tmp3, _tmp1, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sumd = _mm512_shuffle_i32x4(_tmp7, _tmp5, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sume = _mm512_shuffle_i32x4(_tmpb, _tmp9, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sumf = _mm512_shuffle_i32x4(_tmpf, _tmpd, _MM_SHUFFLE(3, 1, 2, 0));
-                }
-
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                _mm512_store_si512((__m512i*)(outptr + 16), _sum1);
-                _mm512_store_si512((__m512i*)(outptr + 32), _sum2);
-                _mm512_store_si512((__m512i*)(outptr + 48), _sum3);
-                _mm512_store_si512((__m512i*)(outptr + 64), _sum4);
-                _mm512_store_si512((__m512i*)(outptr + 80), _sum5);
-                _mm512_store_si512((__m512i*)(outptr + 96), _sum6);
-                _mm512_store_si512((__m512i*)(outptr + 112), _sum7);
-                _mm512_store_si512((__m512i*)(outptr + 128), _sum8);
-                _mm512_store_si512((__m512i*)(outptr + 128 + 16), _sum9);
-                _mm512_store_si512((__m512i*)(outptr + 128 + 32), _suma);
-                _mm512_store_si512((__m512i*)(outptr + 128 + 48), _sumb);
-                _mm512_store_si512((__m512i*)(outptr + 128 + 64), _sumc);
-                _mm512_store_si512((__m512i*)(outptr + 128 + 80), _sumd);
-                _mm512_store_si512((__m512i*)(outptr + 128 + 96), _sume);
-                _mm512_store_si512((__m512i*)(outptr + 128 + 112), _sumf);
-                outptr += 256;
-            }
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const short* pA = pAT;
-
-                __m512i _sum0;
-                __m512i _sum1;
-                __m512i _sum2;
-                __m512i _sum3;
-                __m512i _sum4;
-                __m512i _sum5;
-                __m512i _sum6;
-                __m512i _sum7;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm512_setzero_si512();
-                    _sum1 = _mm512_setzero_si512();
-                    _sum2 = _mm512_setzero_si512();
-                    _sum3 = _mm512_setzero_si512();
-                    _sum4 = _mm512_setzero_si512();
-                    _sum5 = _mm512_setzero_si512();
-                    _sum6 = _mm512_setzero_si512();
-                    _sum7 = _mm512_setzero_si512();
-                }
-                else
-                {
-                    _sum0 = _mm512_load_si512((const __m512i*)outptr);
-                    _sum1 = _mm512_load_si512((const __m512i*)(outptr + 16));
-                    _sum2 = _mm512_load_si512((const __m512i*)(outptr + 16 * 2));
-                    _sum3 = _mm512_load_si512((const __m512i*)(outptr + 16 * 3));
-                    _sum4 = _mm512_load_si512((const __m512i*)(outptr + 16 * 4));
-                    _sum5 = _mm512_load_si512((const __m512i*)(outptr + 16 * 5));
-                    _sum6 = _mm512_load_si512((const __m512i*)(outptr + 16 * 6));
-                    _sum7 = _mm512_load_si512((const __m512i*)(outptr + 16 * 7));
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m512i _pA0 = _mm512_loadu_si512((const __m512i*)pA);
-                    __m256i _pB = _mm256_loadu_si256((const __m256i*)pB);
-                    __m512i _pA1 = _mm512_permutex_epi64(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-                    __m512i _pB0 = _mm512_inserti32x8(_mm512_castsi256_si512(_pB), _pB, 1);
-                    __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-                    __m512i _pB2 = _mm512_shuffle_epi32(_pB0, _MM_PERM_BADC);
-                    __m512i _pB3 = _mm512_shuffle_epi32(_pB0, _MM_PERM_CBAD);
-
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _pA0, _pB0);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _pA0, _pB1);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _pA0, _pB2);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _pA0, _pB3);
-                    _sum4 = _mm512_dpwssd_epi32(_sum4, _pA1, _pB0);
-                    _sum5 = _mm512_dpwssd_epi32(_sum5, _pA1, _pB1);
-                    _sum6 = _mm512_dpwssd_epi32(_sum6, _pA1, _pB2);
-                    _sum7 = _mm512_dpwssd_epi32(_sum7, _pA1, _pB3);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA0, _pB0));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA0, _pB1));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_pA0, _pB2));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_pA0, _pB3));
-                    _sum4 = _mm512_add_epi32(_sum4, _mm512_madd_epi16(_pA1, _pB0));
-                    _sum5 = _mm512_add_epi32(_sum5, _mm512_madd_epi16(_pA1, _pB1));
-                    _sum6 = _mm512_add_epi32(_sum6, _mm512_madd_epi16(_pA1, _pB2));
-                    _sum7 = _mm512_add_epi32(_sum7, _mm512_madd_epi16(_pA1, _pB3));
-#endif
-
-                    pA += 32;
-                    pB += 16;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m512i _pA0 = _mm512_cvtepi16_epi32(_mm256_load_si256((const __m256i*)pA));
-                    __m256i _pB = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i*)pB));
-                    __m512i _pA1 = _mm512_permutex_epi64(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-                    __m512i _pB0 = _mm512_inserti32x8(_mm512_castsi256_si512(_pB), _pB, 1);
-                    __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-                    __m512i _pB2 = _mm512_permutex_epi64(_pB0, _MM_SHUFFLE(2, 3, 0, 1));
-                    __m512i _pB3 = _mm512_shuffle_epi32(_pB0, _MM_PERM_CBAD);
-
-                    __m512i _s0 = _mm512_mullo_epi32(_pA0, _pB0);
-                    __m512i _s1 = _mm512_mullo_epi32(_pA0, _pB1);
-                    __m512i _s2 = _mm512_mullo_epi32(_pA0, _pB2);
-                    __m512i _s3 = _mm512_mullo_epi32(_pA0, _pB3);
-                    __m512i _s4 = _mm512_mullo_epi32(_pA1, _pB0);
-                    __m512i _s5 = _mm512_mullo_epi32(_pA1, _pB1);
-                    __m512i _s6 = _mm512_mullo_epi32(_pA1, _pB2);
-                    __m512i _s7 = _mm512_mullo_epi32(_pA1, _pB3);
-                    _sum0 = _mm512_add_epi32(_sum0, _s0);
-                    _sum1 = _mm512_add_epi32(_sum1, _s1);
-                    _sum2 = _mm512_add_epi32(_sum2, _s2);
-                    _sum3 = _mm512_add_epi32(_sum3, _s3);
-                    _sum4 = _mm512_add_epi32(_sum4, _s4);
-                    _sum5 = _mm512_add_epi32(_sum5, _s5);
-                    _sum6 = _mm512_add_epi32(_sum6, _s6);
-                    _sum7 = _mm512_add_epi32(_sum7, _s7);
-
-                    pA += 16;
-                    pB += 8;
-                }
-
-                if (k_end)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77 80 91 a2 b3 c4 d5 e6 f7
-                    //      01 12 23 30 45 56 67 74 81 92 a3 b0 c5 d6 e7 f4
-                    //      02 13 20 31 46 57 64 75 82 93 a0 b1 c6 d7 e4 f5
-                    //      03 10 21 32 47 54 65 76 83 90 a1 b2 c7 d4 e5 f6
-                    //      40 51 62 73 04 15 26 37 c0 d1 e2 f3 84 95 a6 b7
-                    //      41 52 63 70 05 16 27 34 c1 d2 e3 f0 85 96 a7 b4
-                    //      42 53 60 71 06 17 24 35 c2 d3 e0 f1 86 97 a4 b5
-                    //      43 50 61 72 07 14 25 36 c3 d0 e1 f2 87 94 a5 b6
-                    // to
-                    //      00 10 20 30 44 54 64 74 80 90 a0 b0 c4 d4 e4 f4
-                    //      01 11 21 31 45 55 65 75 81 91 a1 b1 c5 d5 e5 f5
-                    //      02 12 22 32 46 56 66 76 82 92 a2 b2 c6 d6 e6 f6
-                    //      03 13 23 33 47 57 67 77 83 93 a3 b3 c7 d7 e7 f7
-                    //      40 50 60 70 04 14 24 34 c0 d0 e0 f0 84 94 a4 b4
-                    //      41 51 61 71 05 15 25 35 c1 d1 e1 f1 85 95 a5 b5
-                    //      42 52 62 72 06 16 26 36 c2 d2 e2 f2 86 96 a6 b6
-                    //      43 53 63 73 07 17 27 37 c3 d3 e3 f3 87 97 a7 b7
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum2 = _mm512_shuffle_epi32(_sum2, _MM_PERM_BADC);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_ADCB);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum6 = _mm512_shuffle_epi32(_sum6, _MM_PERM_BADC);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_ADCB);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        __m512i _tmp4 = _mm512_unpacklo_epi32(_sum4, _sum7);
-                        __m512i _tmp5 = _mm512_unpackhi_epi32(_sum4, _sum7);
-                        __m512i _tmp6 = _mm512_unpacklo_epi32(_sum6, _sum5);
-                        __m512i _tmp7 = _mm512_unpackhi_epi32(_sum6, _sum5);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm512_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm512_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm512_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm512_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_CBAD);
-                    }
-
-                    // TODO
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum4, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum1, _sum5, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum2, _sum6, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum3, _sum7, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp4 = _mm512_shuffle_i32x4(_sum4, _sum0, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp5 = _mm512_shuffle_i32x4(_sum5, _sum1, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp6 = _mm512_shuffle_i32x4(_sum6, _sum2, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp7 = _mm512_shuffle_i32x4(_sum7, _sum3, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp0, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp1, _tmp1, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp2, _tmp2, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp3, _tmp3, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum4 = _mm512_shuffle_i32x4(_tmp4, _tmp4, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum5 = _mm512_shuffle_i32x4(_tmp5, _tmp5, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum6 = _mm512_shuffle_i32x4(_tmp6, _tmp6, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum7 = _mm512_shuffle_i32x4(_tmp7, _tmp7, _MM_SHUFFLE(3, 1, 2, 0));
-                }
-
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                _mm512_store_si512((__m512i*)(outptr + 16), _sum1);
-                _mm512_store_si512((__m512i*)(outptr + 16 * 2), _sum2);
-                _mm512_store_si512((__m512i*)(outptr + 16 * 3), _sum3);
-                _mm512_store_si512((__m512i*)(outptr + 16 * 4), _sum4);
-                _mm512_store_si512((__m512i*)(outptr + 16 * 5), _sum5);
-                _mm512_store_si512((__m512i*)(outptr + 16 * 6), _sum6);
-                _mm512_store_si512((__m512i*)(outptr + 16 * 7), _sum7);
-                outptr += 16 * 8;
-            }
-#endif // defined(__x86_64__) || defined(_M_X64)
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const short* pA = pAT;
-
-                __m512i _sum0;
-                __m512i _sum1;
-                __m512i _sum2;
-                __m512i _sum3;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm512_setzero_si512();
-                    _sum1 = _mm512_setzero_si512();
-                    _sum2 = _mm512_setzero_si512();
-                    _sum3 = _mm512_setzero_si512();
-                }
-                else
-                {
-                    _sum0 = _mm512_load_si512((const __m512i*)outptr);
-                    _sum1 = _mm512_load_si512((const __m512i*)(outptr + 16));
-                    _sum2 = _mm512_load_si512((const __m512i*)(outptr + 16 * 2));
-                    _sum3 = _mm512_load_si512((const __m512i*)(outptr + 16 * 3));
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m512i _pA0 = _mm512_loadu_si512((const __m512i*)pA);
-                    __m512i _pB = _mm512_castsi128_si512(_mm_loadu_si128((const __m128i*)pB));
-                    __m512i _pB0 = _mm512_shuffle_i32x4(_pB, _pB, _MM_SHUFFLE(0, 0, 0, 0));
-                    __m512i _pA1 = _mm512_shuffle_epi32(_pA0, _MM_PERM_BADC);
-                    __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _pA0, _pB0);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _pA0, _pB1);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _pA1, _pB0);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _pA1, _pB1);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA0, _pB0));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA0, _pB1));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_pA1, _pB0));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_pA1, _pB1));
-#endif
-
-                    pA += 32;
-                    pB += 8;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m512i _pA0 = _mm512_cvtepi16_epi32(_mm256_load_si256((const __m256i*)pA));
-                    __m512i _pB0 = _mm512_cvtepi16_epi32(_mm256_castpd_si256(_mm256_broadcast_sd((const double*)pB)));
-                    __m512i _pA1 = _mm512_permutex_epi64(_pA0, _MM_SHUFFLE(2, 3, 0, 1));
-                    __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-
-                    __m512i _s0 = _mm512_mullo_epi32(_pA0, _pB0);
-                    __m512i _s1 = _mm512_mullo_epi32(_pA0, _pB1);
-                    __m512i _s2 = _mm512_mullo_epi32(_pA1, _pB0);
-                    __m512i _s3 = _mm512_mullo_epi32(_pA1, _pB1);
-                    _sum0 = _mm512_add_epi32(_sum0, _s0);
-                    _sum1 = _mm512_add_epi32(_sum1, _s1);
-                    _sum2 = _mm512_add_epi32(_sum2, _s2);
-                    _sum3 = _mm512_add_epi32(_sum3, _s3);
-
-                    pA += 16;
-                    pB += 4;
-                }
-
-                if (k_end)
-                {
-                    // from
-                    //      00 11 22 33 40 51 62 73 80 91 a2 b3 c0 d1 e2 f3
-                    //      01 12 23 30 41 52 63 70 81 92 a3 b0 c1 d2 e3 f0
-                    //      20 31 02 13 60 71 42 53 a0 b1 82 93 e0 f1 c2 d3
-                    //      21 32 03 10 61 72 43 50 a1 b2 83 90 e1 f2 c3 d0
-                    // to
-                    //      00 10 20 30 40 50 60 70 80 90 a0 b0 c0 d0 e0 f0
-                    //      01 11 21 31 41 51 61 71 81 91 a1 b1 c1 d1 e1 f1
-                    //      02 12 22 32 42 52 62 72 82 92 a2 b2 c2 d2 e2 f2
-                    //      03 13 23 33 43 53 63 73 83 93 a3 b3 c3 d3 e3 f3
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                    }
-                }
-
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                _mm512_store_si512((__m512i*)(outptr + 16), _sum1);
-                _mm512_store_si512((__m512i*)(outptr + 16 * 2), _sum2);
-                _mm512_store_si512((__m512i*)(outptr + 16 * 3), _sum3);
-                outptr += 16 * 4;
-            }
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const short* pA = pAT;
-
-                __m512i _sum0;
-                __m512i _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm512_setzero_si512();
-                    _sum1 = _mm512_setzero_si512();
-                }
-                else
-                {
-                    _sum0 = _mm512_load_si512((const __m512i*)outptr);
-                    _sum1 = _mm512_load_si512((const __m512i*)(outptr + 16));
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m512i _pA = _mm512_loadu_si512((const __m512i*)pA);
-                    __m512i _pB0 = _mm512_castpd_si512(_mm512_set1_pd(((const double*)pB)[0]));
-                    __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_CDAB);
-
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _pA, _pB0);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _pA, _pB1);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA, _pB0));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA, _pB1));
-#endif
-
-                    pA += 32;
-                    pB += 4;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m512i _pA = _mm512_cvtepi16_epi32(_mm256_load_si256((const __m256i*)pA));
-                    __m512i _pB0 = _mm512_cvtepi16_epi32(_mm256_set1_epi32(((const int*)pB)[0]));
-                    __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ABAB);
-
-                    __m512i _s0 = _mm512_mullo_epi32(_pA, _pB0);
-                    __m512i _s1 = _mm512_mullo_epi32(_pA, _pB1);
-                    _sum0 = _mm512_add_epi32(_sum0, _s0);
-                    _sum1 = _mm512_add_epi32(_sum1, _s1);
-
-                    pA += 16;
-                    pB += 2;
-                }
-
-                if (k_end)
-                {
-                    // from
-                    //      00 11 20 31 40 51 60 71 80 91 a0 b1 c0 d1 e0 f1
-                    //      01 10 21 30 41 50 61 70 81 90 a1 b0 c1 d0 e1 f0
-                    // to
-                    //      00 10 20 30 40 50 60 70 80 90 a0 b0 c0 d0 e0 f0
-                    //      01 11 21 31 41 51 61 71 81 91 a1 b1 c1 d1 e1 f1
-                    {
-                        __m512i _tmp0 = _mm512_shuffle_epi32(_sum0, _MM_PERM_DBCA);
-                        __m512i _tmp1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_ACDB);
-                        _sum0 = _mm512_unpacklo_epi32(_tmp0, _tmp1);
-                        _sum1 = _mm512_unpackhi_epi32(_tmp0, _tmp1);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                    }
-                }
-
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                _mm512_store_si512((__m512i*)(outptr + 16), _sum1);
-                outptr += 16 * 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const short* pA = pAT;
-
-                __m512i _sum0;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm512_setzero_si512();
-                }
-                else
-                {
-                    _sum0 = _mm512_load_si512((const __m512i*)outptr);
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m512i _pA = _mm512_loadu_si512((const __m512i*)pA);
-                    __m512i _pB = _mm512_set1_epi32(((const int*)pB)[0]);
-
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _pA, _pB);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA, _pB));
-#endif
-
-                    pA += 32;
-                    pB += 2;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m512i _pA = _mm512_cvtepi16_epi32(_mm256_load_si256((const __m256i*)pA));
-                    __m512i _pB = _mm512_set1_epi32(pB[0]);
-                    __m512i _s0 = _mm512_mullo_epi32(_pA, _pB);
-                    _sum0 = _mm512_add_epi32(_sum0, _s0);
-
-                    pA += 16;
-                    pB += 1;
-                }
-
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                outptr += 16;
-            }
-        }
-    }
-#endif // __AVX512F__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const short* pAT = AT_tile.row<const short>(b) + max_kk * ii;
-            const short* pB = BT_tile.row<const short>(b);
-
-            int jj = 0;
-#if defined(__x86_64__) || defined(_M_X64)
-#if __AVX512F__
-            for (; jj + 15 < max_jj; jj += 16)
-            {
-                const short* pA = pAT;
-
-                __m512i _sum0;
-                __m512i _sum1;
-                __m512i _sum2;
-                __m512i _sum3;
-                __m512i _sum4;
-                __m512i _sum5;
-                __m512i _sum6;
-                __m512i _sum7;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm512_setzero_si512();
-                    _sum1 = _mm512_setzero_si512();
-                    _sum2 = _mm512_setzero_si512();
-                    _sum3 = _mm512_setzero_si512();
-                    _sum4 = _mm512_setzero_si512();
-                    _sum5 = _mm512_setzero_si512();
-                    _sum6 = _mm512_setzero_si512();
-                    _sum7 = _mm512_setzero_si512();
-                }
-                else
-                {
-                    _sum0 = _mm512_loadu_si512((const __m512i*)outptr);
-                    _sum1 = _mm512_loadu_si512((const __m512i*)(outptr + 16));
-                    _sum2 = _mm512_loadu_si512((const __m512i*)(outptr + 32));
-                    _sum3 = _mm512_loadu_si512((const __m512i*)(outptr + 48));
-                    _sum4 = _mm512_loadu_si512((const __m512i*)(outptr + 64));
-                    _sum5 = _mm512_loadu_si512((const __m512i*)(outptr + 80));
-                    _sum6 = _mm512_loadu_si512((const __m512i*)(outptr + 96));
-                    _sum7 = _mm512_loadu_si512((const __m512i*)(outptr + 112));
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m256i _pA0 = _mm256_loadu_si256((const __m256i*)pA);
-                    __m512i _pB0 = _mm512_loadu_si512((const __m512i*)pB);
-
-                    __m512i _pA00 = _mm512_inserti32x8(_mm512_castsi256_si512(_pA0), _pA0, 1);
-                    __m512i _pA11 = _mm512_permutex_epi64(_pA00, _MM_SHUFFLE(1, 0, 3, 2));
-                    __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-                    __m512i _pB2 = _mm512_shuffle_epi32(_pB0, _MM_PERM_BADC);
-                    __m512i _pB3 = _mm512_shuffle_epi32(_pB0, _MM_PERM_CBAD);
-
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _pA00, _pB0);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _pA00, _pB1);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _pA00, _pB2);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _pA00, _pB3);
-                    _sum4 = _mm512_dpwssd_epi32(_sum4, _pA11, _pB0);
-                    _sum5 = _mm512_dpwssd_epi32(_sum5, _pA11, _pB1);
-                    _sum6 = _mm512_dpwssd_epi32(_sum6, _pA11, _pB2);
-                    _sum7 = _mm512_dpwssd_epi32(_sum7, _pA11, _pB3);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA00, _pB0));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA00, _pB1));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_pA00, _pB2));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_pA00, _pB3));
-                    _sum4 = _mm512_add_epi32(_sum4, _mm512_madd_epi16(_pA11, _pB0));
-                    _sum5 = _mm512_add_epi32(_sum5, _mm512_madd_epi16(_pA11, _pB1));
-                    _sum6 = _mm512_add_epi32(_sum6, _mm512_madd_epi16(_pA11, _pB2));
-                    _sum7 = _mm512_add_epi32(_sum7, _mm512_madd_epi16(_pA11, _pB3));
-#endif // __AVX512VNNI__
-
-                    pA += 16;
-                    pB += 32;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m128i _pA = _mm_load_si128((const __m128i*)pA);
-                    __m256i _pB = _mm256_loadu_si256((const __m256i*)pB);
-
-                    __m256i _pA0 = _mm256_cvtepi16_epi32(_pA);
-                    __m512i _pB0 = _mm512_cvtepi16_epi32(_pB);
-
-                    __m512i _pA00 = _mm512_inserti32x8(_mm512_castsi256_si512(_pA0), _pA0, 1);
-                    __m512i _pA11 = _mm512_permutex_epi64(_pA00, _MM_SHUFFLE(1, 0, 3, 2));
-
-                    __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-                    __m512i _pB2 = _mm512_permutex_epi64(_pB0, _MM_SHUFFLE(2, 3, 0, 1));
-                    __m512i _pB3 = _mm512_shuffle_epi32(_pB0, _MM_PERM_CBAD);
-
-                    __m512i _s0 = _mm512_mullo_epi32(_pA00, _pB0);
-                    __m512i _s1 = _mm512_mullo_epi32(_pA00, _pB1);
-                    __m512i _s2 = _mm512_mullo_epi32(_pA00, _pB2);
-                    __m512i _s3 = _mm512_mullo_epi32(_pA00, _pB3);
-                    __m512i _s4 = _mm512_mullo_epi32(_pA11, _pB0);
-                    __m512i _s5 = _mm512_mullo_epi32(_pA11, _pB1);
-                    __m512i _s6 = _mm512_mullo_epi32(_pA11, _pB2);
-                    __m512i _s7 = _mm512_mullo_epi32(_pA11, _pB3);
-                    _sum0 = _mm512_add_epi32(_sum0, _s0);
-                    _sum1 = _mm512_add_epi32(_sum1, _s1);
-                    _sum2 = _mm512_add_epi32(_sum2, _s2);
-                    _sum3 = _mm512_add_epi32(_sum3, _s3);
-                    _sum4 = _mm512_add_epi32(_sum4, _s4);
-                    _sum5 = _mm512_add_epi32(_sum5, _s5);
-                    _sum6 = _mm512_add_epi32(_sum6, _s6);
-                    _sum7 = _mm512_add_epi32(_sum7, _s7);
-
-                    pA += 8;
-                    pB += 16;
-                }
-
-                if (k_end)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77 08 19 2a 3b 4c 5d 6e 7f
-                    //      01 12 23 30 45 56 67 74 09 1a 2b 38 4d 5e 6f 7c
-                    //      02 13 20 31 46 57 64 75 0a 1b 28 39 4e 5f 6c 7d
-                    //      03 10 21 32 47 54 65 76 0b 18 29 3a 4f 5c 6d 7e
-                    //      40 51 62 73 04 15 26 37 48 59 6a 7b 0c 1d 2e 3f
-                    //      41 52 63 70 05 16 27 34 49 5a 6b 78 0d 1e 2f 3c
-                    //      42 53 60 71 06 17 24 35 4a 5b 68 79 0e 1f 2c 3d
-                    //      43 50 61 72 07 14 25 36 4b 58 69 7a 0f 1c 2d 3e
-                    // to
-                    //      00 10 20 30 44 54 64 74 08 18 28 38 4c 5c 6c 7c
-                    //      01 11 21 31 45 55 65 75 09 19 29 39 4d 5d 6d 7d
-                    //      02 12 22 32 46 56 66 76 0a 1a 2a 3a 4e 5e 6e 7e
-                    //      03 13 23 33 47 57 67 77 0b 1b 2b 3b 4f 5f 6f 7f
-                    //      40 50 60 70 04 14 24 34 48 58 68 78 0c 1c 2c 3c
-                    //      41 51 61 71 05 15 25 35 49 59 69 79 0d 1d 2d 3d
-                    //      42 52 62 72 06 16 26 36 4a 5a 6a 7a 0e 1e 2e 3e
-                    //      43 53 63 73 07 17 27 37 4b 5b 6b 7b 0f 1f 2f 3f
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum2 = _mm512_shuffle_epi32(_sum2, _MM_PERM_BADC);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_ADCB);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum6 = _mm512_shuffle_epi32(_sum6, _MM_PERM_BADC);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_ADCB);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        __m512i _tmp4 = _mm512_unpacklo_epi32(_sum4, _sum7);
-                        __m512i _tmp5 = _mm512_unpackhi_epi32(_sum4, _sum7);
-                        __m512i _tmp6 = _mm512_unpacklo_epi32(_sum6, _sum5);
-                        __m512i _tmp7 = _mm512_unpackhi_epi32(_sum6, _sum5);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm512_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm512_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm512_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm512_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_CBAD);
-                    }
-
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum4, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum0, _sum4, _MM_SHUFFLE(3, 2, 3, 2));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum1, _sum5, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum1, _sum5, _MM_SHUFFLE(3, 2, 3, 2));
-                    __m512i _tmp4 = _mm512_shuffle_i32x4(_sum2, _sum6, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp5 = _mm512_shuffle_i32x4(_sum2, _sum6, _MM_SHUFFLE(3, 2, 3, 2));
-                    __m512i _tmp6 = _mm512_shuffle_i32x4(_sum3, _sum7, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp7 = _mm512_shuffle_i32x4(_sum3, _sum7, _MM_SHUFFLE(3, 2, 3, 2));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp2, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp4, _tmp6, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp0, _tmp2, _MM_SHUFFLE(1, 3, 1, 3));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp4, _tmp6, _MM_SHUFFLE(1, 3, 1, 3));
-                    _sum4 = _mm512_shuffle_i32x4(_tmp1, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum5 = _mm512_shuffle_i32x4(_tmp5, _tmp7, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum6 = _mm512_shuffle_i32x4(_tmp1, _tmp3, _MM_SHUFFLE(1, 3, 1, 3));
-                    _sum7 = _mm512_shuffle_i32x4(_tmp5, _tmp7, _MM_SHUFFLE(1, 3, 1, 3));
-                }
-
-                _mm512_storeu_si512((__m512i*)outptr, _sum0);
-                _mm512_storeu_si512((__m512i*)(outptr + 16), _sum1);
-                _mm512_storeu_si512((__m512i*)(outptr + 32), _sum2);
-                _mm512_storeu_si512((__m512i*)(outptr + 48), _sum3);
-                _mm512_storeu_si512((__m512i*)(outptr + 64), _sum4);
-                _mm512_storeu_si512((__m512i*)(outptr + 80), _sum5);
-                _mm512_storeu_si512((__m512i*)(outptr + 96), _sum6);
-                _mm512_storeu_si512((__m512i*)(outptr + 112), _sum7);
-                outptr += 128;
-            }
-#endif // __AVX512F__
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const short* pA = pAT;
-
-#if __AVX512F__
-                __m512i _sum0;
-                __m512i _sum1;
-                __m512i _sum2;
-                __m512i _sum3;
-#else
-                __m256i _sum0;
-                __m256i _sum1;
-                __m256i _sum2;
-                __m256i _sum3;
-                __m256i _sum4;
-                __m256i _sum5;
-                __m256i _sum6;
-                __m256i _sum7;
-#endif // __AVX512F__
-
-                if (k == 0)
-                {
-#if __AVX512F__
-                    _sum0 = _mm512_setzero_si512();
-                    _sum1 = _mm512_setzero_si512();
-                    _sum2 = _mm512_setzero_si512();
-                    _sum3 = _mm512_setzero_si512();
-#else
-                    _sum0 = _mm256_setzero_si256();
-                    _sum1 = _mm256_setzero_si256();
-                    _sum2 = _mm256_setzero_si256();
-                    _sum3 = _mm256_setzero_si256();
-                    _sum4 = _mm256_setzero_si256();
-                    _sum5 = _mm256_setzero_si256();
-                    _sum6 = _mm256_setzero_si256();
-                    _sum7 = _mm256_setzero_si256();
-#endif // __AVX512F__
-                }
-                else
-                {
-#if __AVX512F__
-                    _sum0 = _mm512_loadu_si512((const __m512i*)outptr);
-                    _sum1 = _mm512_loadu_si512((const __m512i*)(outptr + 16));
-                    _sum2 = _mm512_loadu_si512((const __m512i*)(outptr + 32));
-                    _sum3 = _mm512_loadu_si512((const __m512i*)(outptr + 48));
-#else
-                    _sum0 = _mm256_load_si256((const __m256i*)outptr);
-                    _sum1 = _mm256_load_si256((const __m256i*)(outptr + 8));
-                    _sum2 = _mm256_load_si256((const __m256i*)(outptr + 16));
-                    _sum3 = _mm256_load_si256((const __m256i*)(outptr + 24));
-                    _sum4 = _mm256_load_si256((const __m256i*)(outptr + 32));
-                    _sum5 = _mm256_load_si256((const __m256i*)(outptr + 40));
-                    _sum6 = _mm256_load_si256((const __m256i*)(outptr + 48));
-                    _sum7 = _mm256_load_si256((const __m256i*)(outptr + 56));
-#endif // __AVX512F__
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m256i _pA0 = _mm256_loadu_si256((const __m256i*)pA);
-                    __m256i _pB0 = _mm256_loadu_si256((const __m256i*)pB);
-#if __AVX512F__
-                    __m512i _pA00 = _mm512_inserti32x8(_mm512_castsi256_si512(_pA0), _pA0, 1);
-                    __m512i _pA11 = _mm512_permutex_epi64(_pA00, _MM_SHUFFLE(1, 0, 3, 2));
-                    __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-                    __m512i _pB01 = _mm512_inserti32x8(_mm512_castsi256_si512(_pB0), _pB1, 1);
-                    __m512i _pB23 = _mm512_shuffle_epi32(_pB01, _MM_PERM_BADC);
-
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _pA00, _pB01);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _pA00, _pB23);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _pA11, _pB01);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _pA11, _pB23);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA00, _pB01));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA00, _pB23));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_pA11, _pB01));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_pA11, _pB23));
-#endif // __AVX512VNNI__
-#else  // __AVX512F__
-                    __m256i _pA1 = _mm256_permute4x64_epi64(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-                    __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-                    __m256i _pB2 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(1, 0, 3, 2));
-                    __m256i _pB3 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(2, 1, 0, 3));
-
-#if __AVXVNNI__
-                    _sum0 = _mm256_dpwssd_epi32(_sum0, _pA0, _pB0);
-                    _sum1 = _mm256_dpwssd_epi32(_sum1, _pA0, _pB1);
-                    _sum2 = _mm256_dpwssd_epi32(_sum2, _pA0, _pB2);
-                    _sum3 = _mm256_dpwssd_epi32(_sum3, _pA0, _pB3);
-                    _sum4 = _mm256_dpwssd_epi32(_sum4, _pA1, _pB0);
-                    _sum5 = _mm256_dpwssd_epi32(_sum5, _pA1, _pB1);
-                    _sum6 = _mm256_dpwssd_epi32(_sum6, _pA1, _pB2);
-                    _sum7 = _mm256_dpwssd_epi32(_sum7, _pA1, _pB3);
-#else
-                    _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_pA0, _pB0));
-                    _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_pA0, _pB1));
-                    _sum2 = _mm256_add_epi32(_sum2, _mm256_madd_epi16(_pA0, _pB2));
-                    _sum3 = _mm256_add_epi32(_sum3, _mm256_madd_epi16(_pA0, _pB3));
-                    _sum4 = _mm256_add_epi32(_sum4, _mm256_madd_epi16(_pA1, _pB0));
-                    _sum5 = _mm256_add_epi32(_sum5, _mm256_madd_epi16(_pA1, _pB1));
-                    _sum6 = _mm256_add_epi32(_sum6, _mm256_madd_epi16(_pA1, _pB2));
-                    _sum7 = _mm256_add_epi32(_sum7, _mm256_madd_epi16(_pA1, _pB3));
-#endif // __AVXVNNI__
-#endif // __AVX512F__
-
-                    pA += 16;
-                    pB += 16;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m128i _pA = _mm_load_si128((const __m128i*)pA);
-                    __m128i _pB = _mm_load_si128((const __m128i*)pB);
-
-                    __m256i _pA0 = _mm256_cvtepi16_epi32(_pA);
-                    __m256i _pB0 = _mm256_cvtepi16_epi32(_pB);
-#if __AVX512F__
-                    __m512i _pA00 = _mm512_inserti32x8(_mm512_castsi256_si512(_pA0), _pA0, 1);
-                    __m512i _pA11 = _mm512_shuffle_i32x4(_pA00, _pA00, _MM_SHUFFLE(2, 3, 0, 1));
-                    __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-                    __m512i _pB01 = _mm512_inserti32x8(_mm512_castsi256_si512(_pB0), _pB1, 1);
-                    __m512i _pB23 = _mm512_permutex_epi64(_pB01, _MM_SHUFFLE(2, 3, 0, 1));
-
-                    __m512i _s01 = _mm512_mullo_epi32(_pA00, _pB01);
-                    __m512i _s23 = _mm512_mullo_epi32(_pA00, _pB23);
-                    __m512i _s45 = _mm512_mullo_epi32(_pA11, _pB01);
-                    __m512i _s67 = _mm512_mullo_epi32(_pA11, _pB23);
-                    _sum0 = _mm512_add_epi32(_sum0, _s01);
-                    _sum1 = _mm512_add_epi32(_sum1, _s23);
-                    _sum2 = _mm512_add_epi32(_sum2, _s45);
-                    _sum3 = _mm512_add_epi32(_sum3, _s67);
-#else
-                    __m256i _pA1 = _mm256_permute4x64_epi64(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-                    __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-                    __m256i _pB2 = _mm256_permute4x64_epi64(_pB0, _MM_SHUFFLE(2, 3, 0, 1));
-                    __m256i _pB3 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(2, 1, 0, 3));
-
-                    __m256i _s0 = _mm256_mullo_epi32(_pA0, _pB0);
-                    __m256i _s1 = _mm256_mullo_epi32(_pA0, _pB1);
-                    __m256i _s2 = _mm256_mullo_epi32(_pA0, _pB2);
-                    __m256i _s3 = _mm256_mullo_epi32(_pA0, _pB3);
-                    __m256i _s4 = _mm256_mullo_epi32(_pA1, _pB0);
-                    __m256i _s5 = _mm256_mullo_epi32(_pA1, _pB1);
-                    __m256i _s6 = _mm256_mullo_epi32(_pA1, _pB2);
-                    __m256i _s7 = _mm256_mullo_epi32(_pA1, _pB3);
-                    _sum0 = _mm256_add_epi32(_sum0, _s0);
-                    _sum1 = _mm256_add_epi32(_sum1, _s1);
-                    _sum2 = _mm256_add_epi32(_sum2, _s2);
-                    _sum3 = _mm256_add_epi32(_sum3, _s3);
-                    _sum4 = _mm256_add_epi32(_sum4, _s4);
-                    _sum5 = _mm256_add_epi32(_sum5, _s5);
-                    _sum6 = _mm256_add_epi32(_sum6, _s6);
-                    _sum7 = _mm256_add_epi32(_sum7, _s7);
-#endif // __AVX512F__
-
-                    pA += 8;
-                    pB += 8;
-                }
-
-#if __AVX512F__
-                if (k_end)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77 01 12 23 30 45 56 67 74
-                    //      02 13 20 31 46 57 64 75 03 10 21 32 47 54 65 76
-                    //      40 51 62 73 04 15 26 37 41 52 63 70 05 16 27 34
-                    //      42 53 60 71 06 17 24 35 43 50 61 72 07 14 25 36
-                    // to
-                    //      00 10 20 30 44 54 64 74 04 14 24 34 40 50 60 70
-                    //      01 11 21 31 45 55 65 75 05 15 25 35 41 51 61 71
-                    //      02 12 22 32 46 56 66 76 06 16 26 36 42 52 62 72
-                    //      03 13 23 33 47 57 67 77 07 17 27 37 43 53 63 73
-                    {
-                        __m512i _s0 = _mm512_shuffle_i32x4(_sum0, _sum2, _MM_SHUFFLE(0, 1, 1, 0));
-                        __m512i _s1 = _mm512_shuffle_i32x4(_sum1, _sum3, _MM_SHUFFLE(2, 3, 3, 2));
-                        __m512i _s2 = _mm512_shuffle_i32x4(_sum1, _sum3, _MM_SHUFFLE(0, 1, 1, 0));
-                        __m512i _s3 = _mm512_shuffle_i32x4(_sum0, _sum2, _MM_SHUFFLE(2, 3, 3, 2));
-                        _s1 = _mm512_shuffle_epi32(_s1, _MM_PERM_ADCB);
-                        _s2 = _mm512_shuffle_epi32(_s2, _MM_PERM_BADC);
-                        _s3 = _mm512_shuffle_epi32(_s3, _MM_PERM_CBAD);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_s0, _s1);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_s0, _s1);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_s2, _s3);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_s2, _s3);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                    }
-
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(3, 0, 3, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(3, 0, 3, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(1, 2, 1, 2));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(1, 2, 1, 2));
-                    _sum0 = _tmp0;
-                    _sum1 = _tmp1;
-                    _sum2 = _tmp2;
-                    _sum3 = _tmp3;
-                }
-
-                _mm512_storeu_si512((__m512i*)outptr, _sum0);
-                _mm512_storeu_si512((__m512i*)(outptr + 16), _sum1);
-                _mm512_storeu_si512((__m512i*)(outptr + 32), _sum2);
-                _mm512_storeu_si512((__m512i*)(outptr + 48), _sum3);
-                outptr += 64;
-#else
-                if (k_end)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77
-                    //      01 12 23 30 45 56 67 74
-                    //      02 13 20 31 46 57 64 75
-                    //      03 10 21 32 47 54 65 76
-                    //      40 51 62 73 04 15 26 37
-                    //      41 52 63 70 05 16 27 34
-                    //      42 53 60 71 06 17 24 35
-                    //      43 50 61 72 07 14 25 36
-                    // to
-                    //      00 10 20 30 44 54 64 74
-                    //      01 11 21 31 45 55 65 75
-                    //      02 12 22 32 46 56 66 76
-                    //      03 13 23 33 47 57 67 77
-                    //      40 50 60 70 04 14 24 34
-                    //      41 51 61 71 05 15 25 35
-                    //      42 52 62 72 06 16 26 36
-                    //      43 53 63 73 07 17 27 37
-                    {
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum2 = _mm256_shuffle_epi32(_sum2, _MM_SHUFFLE(1, 0, 3, 2));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(0, 3, 2, 1));
-                        _sum5 = _mm256_shuffle_epi32(_sum5, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum6 = _mm256_shuffle_epi32(_sum6, _MM_SHUFFLE(1, 0, 3, 2));
-                        _sum7 = _mm256_shuffle_epi32(_sum7, _MM_SHUFFLE(0, 3, 2, 1));
-                        __m256i _tmp0 = _mm256_unpacklo_epi32(_sum0, _sum3);
-                        __m256i _tmp1 = _mm256_unpackhi_epi32(_sum0, _sum3);
-                        __m256i _tmp2 = _mm256_unpacklo_epi32(_sum2, _sum1);
-                        __m256i _tmp3 = _mm256_unpackhi_epi32(_sum2, _sum1);
-                        __m256i _tmp4 = _mm256_unpacklo_epi32(_sum4, _sum7);
-                        __m256i _tmp5 = _mm256_unpackhi_epi32(_sum4, _sum7);
-                        __m256i _tmp6 = _mm256_unpacklo_epi32(_sum6, _sum5);
-                        __m256i _tmp7 = _mm256_unpackhi_epi32(_sum6, _sum5);
-                        _sum0 = _mm256_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm256_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm256_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm256_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm256_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm256_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm256_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm256_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum5 = _mm256_shuffle_epi32(_sum5, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum7 = _mm256_shuffle_epi32(_sum7, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-
-                    __m256i _tmp0 = _mm256_permute2x128_si256(_sum0, _sum4, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp1 = _mm256_permute2x128_si256(_sum1, _sum5, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp2 = _mm256_permute2x128_si256(_sum2, _sum6, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp3 = _mm256_permute2x128_si256(_sum3, _sum7, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp4 = _mm256_permute2x128_si256(_sum4, _sum0, _MM_SHUFFLE(0, 3, 0, 1));
-                    __m256i _tmp5 = _mm256_permute2x128_si256(_sum5, _sum1, _MM_SHUFFLE(0, 3, 0, 1));
-                    __m256i _tmp6 = _mm256_permute2x128_si256(_sum6, _sum2, _MM_SHUFFLE(0, 3, 0, 1));
-                    __m256i _tmp7 = _mm256_permute2x128_si256(_sum7, _sum3, _MM_SHUFFLE(0, 3, 0, 1));
-                    _sum0 = _tmp0;
-                    _sum1 = _tmp1;
-                    _sum2 = _tmp2;
-                    _sum3 = _tmp3;
-                    _sum4 = _tmp4;
-                    _sum5 = _tmp5;
-                    _sum6 = _tmp6;
-                    _sum7 = _tmp7;
-                }
-
-                _mm256_store_si256((__m256i*)outptr, _sum0);
-                _mm256_store_si256((__m256i*)(outptr + 8), _sum1);
-                _mm256_store_si256((__m256i*)(outptr + 8 * 2), _sum2);
-                _mm256_store_si256((__m256i*)(outptr + 8 * 3), _sum3);
-                _mm256_store_si256((__m256i*)(outptr + 8 * 4), _sum4);
-                _mm256_store_si256((__m256i*)(outptr + 8 * 5), _sum5);
-                _mm256_store_si256((__m256i*)(outptr + 8 * 6), _sum6);
-                _mm256_store_si256((__m256i*)(outptr + 8 * 7), _sum7);
-                outptr += 8 * 8;
-#endif // __AVX512F__
-            }
-#endif // defined(__x86_64__) || defined(_M_X64)
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const short* pA = pAT;
-
-                __m256i _sum0;
-                __m256i _sum1;
-                __m256i _sum2;
-                __m256i _sum3;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm256_setzero_si256();
-                    _sum1 = _mm256_setzero_si256();
-                    _sum2 = _mm256_setzero_si256();
-                    _sum3 = _mm256_setzero_si256();
-                }
-                else
-                {
-                    _sum0 = _mm256_load_si256((const __m256i*)outptr);
-                    _sum1 = _mm256_load_si256((const __m256i*)(outptr + 8));
-                    _sum2 = _mm256_load_si256((const __m256i*)(outptr + 16));
-                    _sum3 = _mm256_load_si256((const __m256i*)(outptr + 24));
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m256i _pA0 = _mm256_loadu_si256((const __m256i*)pA);
-                    __m128i _pB = _mm_loadu_si128((const __m128i*)pB);
-                    __m256i _pA1 = _mm256_shuffle_epi32(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-                    __m256i _pB0 = _mm256_inserti128_si256(_mm256_castsi128_si256(_pB), _pB, 1);
-                    __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                    _sum0 = _mm256_dpwssd_epi32(_sum0, _pA0, _pB0);
-                    _sum1 = _mm256_dpwssd_epi32(_sum1, _pA0, _pB1);
-                    _sum2 = _mm256_dpwssd_epi32(_sum2, _pA1, _pB0);
-                    _sum3 = _mm256_dpwssd_epi32(_sum3, _pA1, _pB1);
-#else
-                    _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_pA0, _pB0));
-                    _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_pA0, _pB1));
-                    _sum2 = _mm256_add_epi32(_sum2, _mm256_madd_epi16(_pA1, _pB0));
-                    _sum3 = _mm256_add_epi32(_sum3, _mm256_madd_epi16(_pA1, _pB1));
-#endif
-
-                    pA += 16;
-                    pB += 8;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m256i _pA0 = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i*)pA));
-                    __m256i _pB0 = _mm256_cvtepi16_epi32(_mm_castpd_si128(_mm_load1_pd((const double*)pB)));
-                    __m256i _pA1 = _mm256_permute4x64_epi64(_pA0, _MM_SHUFFLE(2, 3, 0, 1));
-                    __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-
-                    __m256i _s0 = _mm256_mullo_epi32(_pA0, _pB0);
-                    __m256i _s1 = _mm256_mullo_epi32(_pA0, _pB1);
-                    __m256i _s2 = _mm256_mullo_epi32(_pA1, _pB0);
-                    __m256i _s3 = _mm256_mullo_epi32(_pA1, _pB1);
-                    _sum0 = _mm256_add_epi32(_sum0, _s0);
-                    _sum1 = _mm256_add_epi32(_sum1, _s1);
-                    _sum2 = _mm256_add_epi32(_sum2, _s2);
-                    _sum3 = _mm256_add_epi32(_sum3, _s3);
-
-                    pA += 8;
-                    pB += 4;
-                }
-
-                if (k_end)
-                {
-                    // from
-                    //      00 11 22 33 40 51 62 73
-                    //      01 12 23 30 41 52 63 70
-                    //      20 31 02 13 60 71 42 53
-                    //      21 32 03 10 61 72 43 50
-                    // to
-                    //      00 10 20 30 40 50 60 70
-                    //      01 11 21 31 41 51 61 71
-                    //      02 12 22 32 42 52 62 72
-                    //      03 13 23 33 43 53 63 73
-                    {
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        __m256i _tmp0 = _mm256_unpacklo_epi32(_sum0, _sum3);
-                        __m256i _tmp1 = _mm256_unpackhi_epi32(_sum0, _sum3);
-                        __m256i _tmp2 = _mm256_unpacklo_epi32(_sum2, _sum1);
-                        __m256i _tmp3 = _mm256_unpackhi_epi32(_sum2, _sum1);
-                        _sum0 = _mm256_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm256_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm256_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm256_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-                }
-
-                _mm256_store_si256((__m256i*)outptr, _sum0);
-                _mm256_store_si256((__m256i*)(outptr + 8), _sum1);
-                _mm256_store_si256((__m256i*)(outptr + 16), _sum2);
-                _mm256_store_si256((__m256i*)(outptr + 24), _sum3);
-                outptr += 32;
-            }
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const short* pA = pAT;
-
-                __m256i _sum0;
-                __m256i _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm256_setzero_si256();
-                    _sum1 = _mm256_setzero_si256();
-                }
-                else
-                {
-                    _sum0 = _mm256_load_si256((const __m256i*)outptr);
-                    _sum1 = _mm256_load_si256((const __m256i*)(outptr + 8));
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m256i _pA = _mm256_loadu_si256((const __m256i*)pA);
-                    __m256i _pB0 = _mm256_castpd_si256(_mm256_broadcast_sd((const double*)pB));
-                    __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 1, 0, 1));
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                    _sum0 = _mm256_dpwssd_epi32(_sum0, _pA, _pB0);
-                    _sum1 = _mm256_dpwssd_epi32(_sum1, _pA, _pB1);
-#else
-                    _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_pA, _pB0));
-                    _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_pA, _pB1));
-#endif
-
-                    pA += 16;
-                    pB += 4;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m256i _pA = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i*)pA));
-                    __m256i _pB0 = _mm256_cvtepi16_epi32(_mm_castps_si128(_mm_load1_ps((const float*)pB)));
-                    __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 1, 0, 1));
-
-                    __m256i _s0 = _mm256_mullo_epi32(_pA, _pB0);
-                    __m256i _s1 = _mm256_mullo_epi32(_pA, _pB1);
-                    _sum0 = _mm256_add_epi32(_sum0, _s0);
-                    _sum1 = _mm256_add_epi32(_sum1, _s1);
-
-                    pA += 8;
-                    pB += 2;
-                }
-
-                if (k_end)
-                {
-                    // from
-                    //      00 11 20 31 40 51 60 71
-                    //      01 10 21 30 41 50 61 70
-                    // to
-                    //      00 10 20 30 40 50 60 70
-                    //      01 11 21 31 41 51 61 71
-                    {
-                        __m256i _tmp0 = _mm256_shuffle_epi32(_sum0, _MM_SHUFFLE(3, 1, 2, 0));
-                        __m256i _tmp1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(0, 2, 3, 1));
-                        _sum0 = _mm256_unpacklo_epi32(_tmp0, _tmp1);
-                        _sum1 = _mm256_unpackhi_epi32(_tmp0, _tmp1);
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-                }
-
-                _mm256_store_si256((__m256i*)outptr, _sum0);
-                _mm256_store_si256((__m256i*)(outptr + 8), _sum1);
-                outptr += 16;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const short* pA = pAT;
-
-                __m256i _sum0;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm256_setzero_si256();
-                }
-                else
-                {
-                    _sum0 = _mm256_load_si256((const __m256i*)outptr);
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m256i _pA = _mm256_loadu_si256((const __m256i*)pA);
-                    __m256i _pB = _mm256_castps_si256(_mm256_broadcast_ss((const float*)pB));
-                    _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_pA, _pB));
-
-                    pA += 16;
-                    pB += 2;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m256i _pA = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i*)pA));
-                    __m256i _pB = _mm256_set1_epi32(pB[0]);
-                    __m256i _s0 = _mm256_mullo_epi32(_pA, _pB);
-                    _sum0 = _mm256_add_epi32(_sum0, _s0);
-
-                    pA += 8;
-                    pB += 1;
-                }
-
-                _mm256_store_si256((__m256i*)outptr, _sum0);
-                outptr += 8;
-            }
-        }
-    }
-#endif // __AVX2__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const short* pAT = AT_tile.row<const short>(b) + max_kk * ii;
-            const short* pB = BT_tile.row<const short>(b);
-
-            int jj = 0;
-#if defined(__x86_64__) || defined(_M_X64)
-#if __AVX512F__
-            for (; jj + 15 < max_jj; jj += 16)
-            {
-                const short* pA = pAT;
-
-                __m512i _sum0;
-                __m512i _sum1;
-                __m512i _sum2;
-                __m512i _sum3;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm512_setzero_si512();
-                    _sum1 = _mm512_setzero_si512();
-                    _sum2 = _mm512_setzero_si512();
-                    _sum3 = _mm512_setzero_si512();
-                }
-                else
-                {
-                    _sum0 = _mm512_loadu_si512((const __m512i*)outptr);
-                    _sum1 = _mm512_loadu_si512((const __m512i*)(outptr + 16));
-                    _sum2 = _mm512_loadu_si512((const __m512i*)(outptr + 32));
-                    _sum3 = _mm512_loadu_si512((const __m512i*)(outptr + 48));
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m128i _pA = _mm_loadu_si128((const __m128i*)pA);
-                    __m512i _pB0 = _mm512_loadu_si512((const __m512i*)pB);
-                    __m256i _pAA = _mm256_inserti128_si256(_mm256_castsi128_si256(_pA), _pA, 1);
-                    __m512i _pA0 = _mm512_inserti32x8(_mm512_castsi256_si512(_pAA), _pAA, 1);
-                    __m512i _pA1 = _mm512_shuffle_epi32(_pA0, _MM_PERM_BADC);
-                    __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _pA0, _pB0);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _pA0, _pB1);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _pA1, _pB0);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _pA1, _pB1);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA0, _pB0));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA0, _pB1));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_pA1, _pB0));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_pA1, _pB1));
-#endif
-
-                    pA += 8;
-                    pB += 32;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m256i _pA = _mm256_castpd_si256(_mm256_broadcast_sd((const double*)pA));
-                    __m256i _pB = _mm256_loadu_si256((const __m256i*)pB);
-
-                    __m512i _pA0 = _mm512_cvtepi16_epi32(_pA);
-                    __m512i _pB0 = _mm512_cvtepi16_epi32(_pB);
-                    __m512i _pA1 = _mm512_permutex_epi64(_pA0, _MM_SHUFFLE(2, 3, 0, 1));
-                    __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-
-                    __m512i _s0 = _mm512_mullo_epi32(_pA0, _pB0);
-                    __m512i _s1 = _mm512_mullo_epi32(_pA0, _pB1);
-                    __m512i _s2 = _mm512_mullo_epi32(_pA1, _pB0);
-                    __m512i _s3 = _mm512_mullo_epi32(_pA1, _pB1);
-
-                    _sum0 = _mm512_add_epi32(_sum0, _s0);
-                    _sum1 = _mm512_add_epi32(_sum1, _s1);
-                    _sum2 = _mm512_add_epi32(_sum2, _s2);
-                    _sum3 = _mm512_add_epi32(_sum3, _s3);
-
-                    pA += 4;
-                    pB += 16;
-                }
-
-                if (k_end)
-                {
-                    // from
-                    //      00 11 22 33 04 15 26 37 08 19 2a 3b 0c 1d 2e 3f
-                    //      01 12 23 30 05 16 27 34 09 1a 2b 38 0d 1e 2f 3c
-                    //      20 31 02 13 24 35 06 17 28 3a 0a 1b 2c 3d 0e 1f
-                    //      21 32 03 10 25 36 07 14 29 3a 0b 18 2d 3e 0f 1c
-                    // to
-                    //      00 10 20 30 04 14 24 34 08 18 28 38 0c 1c 2c 3c
-                    //      01 11 21 31 05 15 25 35 09 19 29 39 0d 1d 2d 3d
-                    //      02 12 22 32 06 16 26 36 0a 1a 2a 3a 0e 1e 2e 3e
-                    //      03 13 23 33 07 17 27 37 0b 1b 2b 3b 0f 1f 2f 3f
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                    }
-
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(3, 2, 3, 2));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(3, 2, 3, 2));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-                }
-
-                _mm512_storeu_si512((__m512i*)outptr, _sum0);
-                _mm512_storeu_si512((__m512i*)(outptr + 16), _sum1);
-                _mm512_storeu_si512((__m512i*)(outptr + 32), _sum2);
-                _mm512_storeu_si512((__m512i*)(outptr + 48), _sum3);
-                outptr += 64;
-            }
-#endif // __AVX512F__
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const short* pA = pAT;
-
-#if __AVX2__
-                __m256i _sum0;
-                __m256i _sum1;
-                __m256i _sum2;
-                __m256i _sum3;
-#else
-                __m128i _sum0;
-                __m128i _sum1;
-                __m128i _sum2;
-                __m128i _sum3;
-                __m128i _sum4;
-                __m128i _sum5;
-                __m128i _sum6;
-                __m128i _sum7;
-#endif
-
-                if (k == 0)
-                {
-#if __AVX2__
-                    _sum0 = _mm256_setzero_si256();
-                    _sum1 = _mm256_setzero_si256();
-                    _sum2 = _mm256_setzero_si256();
-                    _sum3 = _mm256_setzero_si256();
-#else
-                    _sum0 = _mm_setzero_si128();
-                    _sum1 = _mm_setzero_si128();
-                    _sum2 = _mm_setzero_si128();
-                    _sum3 = _mm_setzero_si128();
-                    _sum4 = _mm_setzero_si128();
-                    _sum5 = _mm_setzero_si128();
-                    _sum6 = _mm_setzero_si128();
-                    _sum7 = _mm_setzero_si128();
-#endif
-                }
-                else
-                {
-#if __AVX2__
-                    _sum0 = _mm256_loadu_si256((const __m256i*)outptr);
-                    _sum1 = _mm256_loadu_si256((const __m256i*)(outptr + 8));
-                    _sum2 = _mm256_loadu_si256((const __m256i*)(outptr + 16));
-                    _sum3 = _mm256_loadu_si256((const __m256i*)(outptr + 24));
-#else
-                    _sum0 = _mm_load_si128((const __m128i*)outptr);
-                    _sum1 = _mm_load_si128((const __m128i*)(outptr + 4));
-                    _sum2 = _mm_load_si128((const __m128i*)(outptr + 8));
-                    _sum3 = _mm_load_si128((const __m128i*)(outptr + 12));
-                    _sum4 = _mm_load_si128((const __m128i*)(outptr + 16));
-                    _sum5 = _mm_load_si128((const __m128i*)(outptr + 20));
-                    _sum6 = _mm_load_si128((const __m128i*)(outptr + 24));
-                    _sum7 = _mm_load_si128((const __m128i*)(outptr + 28));
-#endif
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-#if __AVX2__
-                    __m128i _pA = _mm_loadu_si128((const __m128i*)pA);
-                    __m256i _pB0 = _mm256_loadu_si256((const __m256i*)pB);
-                    __m256i _pA0 = _mm256_inserti128_si256(_mm256_castsi128_si256(_pA), _pA, 1);
-                    __m256i _pA1 = _mm256_shuffle_epi32(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-                    __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                    _sum0 = _mm256_dpwssd_epi32(_sum0, _pA0, _pB0);
-                    _sum1 = _mm256_dpwssd_epi32(_sum1, _pA0, _pB1);
-                    _sum2 = _mm256_dpwssd_epi32(_sum2, _pA1, _pB0);
-                    _sum3 = _mm256_dpwssd_epi32(_sum3, _pA1, _pB1);
-#else
-                    _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_pA0, _pB0));
-                    _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_pA0, _pB1));
-                    _sum2 = _mm256_add_epi32(_sum2, _mm256_madd_epi16(_pA1, _pB0));
-                    _sum3 = _mm256_add_epi32(_sum3, _mm256_madd_epi16(_pA1, _pB1));
-#endif
-#else // __AVX2__
-                    __m128i _pA0 = _mm_loadu_si128((const __m128i*)pA);
-                    __m128i _pB0 = _mm_loadu_si128((const __m128i*)pB);
-                    __m128i _pB1 = _mm_loadu_si128((const __m128i*)(pB + 8));
-                    __m128i _pA1 = _mm_shuffle_epi32(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-                    __m128i _pB2 = _mm_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-                    __m128i _pB3 = _mm_shuffle_epi32(_pB1, _MM_SHUFFLE(0, 3, 2, 1));
-
-#if __XOP__
-                    _sum0 = _mm_maddd_epi16(_pA0, _pB0, _sum0);
-                    _sum1 = _mm_maddd_epi16(_pA0, _pB1, _sum1);
-                    _sum2 = _mm_maddd_epi16(_pA0, _pB2, _sum2);
-                    _sum3 = _mm_maddd_epi16(_pA0, _pB3, _sum3);
-                    _sum4 = _mm_maddd_epi16(_pA1, _pB0, _sum4);
-                    _sum5 = _mm_maddd_epi16(_pA1, _pB1, _sum5);
-                    _sum6 = _mm_maddd_epi16(_pA1, _pB2, _sum6);
-                    _sum7 = _mm_maddd_epi16(_pA1, _pB3, _sum7);
-#else
-                    _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA0, _pB0));
-                    _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_pA0, _pB1));
-                    _sum2 = _mm_add_epi32(_sum2, _mm_madd_epi16(_pA0, _pB2));
-                    _sum3 = _mm_add_epi32(_sum3, _mm_madd_epi16(_pA0, _pB3));
-                    _sum4 = _mm_add_epi32(_sum4, _mm_madd_epi16(_pA1, _pB0));
-                    _sum5 = _mm_add_epi32(_sum5, _mm_madd_epi16(_pA1, _pB1));
-                    _sum6 = _mm_add_epi32(_sum6, _mm_madd_epi16(_pA1, _pB2));
-                    _sum7 = _mm_add_epi32(_sum7, _mm_madd_epi16(_pA1, _pB3));
-#endif
-#endif // __AVX2__
-
-                    pA += 8;
-                    pB += 16;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m128i _pA = _mm_castpd_si128(_mm_load1_pd((const double*)pA));
-                    __m128i _pB = _mm_loadu_si128((const __m128i*)pB);
-
-#if __AVX2__
-                    __m256i _pA0 = _mm256_cvtepi16_epi32(_pA);
-                    __m256i _pB0 = _mm256_cvtepi16_epi32(_pB);
-                    __m256i _pA1 = _mm256_permute4x64_epi64(_pA0, _MM_SHUFFLE(2, 3, 0, 1));
-                    __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-
-                    __m256i _s0 = _mm256_mullo_epi32(_pA0, _pB0);
-                    __m256i _s1 = _mm256_mullo_epi32(_pA0, _pB1);
-                    __m256i _s2 = _mm256_mullo_epi32(_pA1, _pB0);
-                    __m256i _s3 = _mm256_mullo_epi32(_pA1, _pB1);
-
-                    _sum0 = _mm256_add_epi32(_sum0, _s0);
-                    _sum1 = _mm256_add_epi32(_sum1, _s1);
-                    _sum2 = _mm256_add_epi32(_sum2, _s2);
-                    _sum3 = _mm256_add_epi32(_sum3, _s3);
-#else // __AVX2__
-#if __XOP__
-                    __m128i _pA0 = _mm_unpacklo_epi16(_pA, _pA);
-                    __m128i _pA1 = _mm_shuffle_epi32(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-                    __m128i _pB0 = _mm_unpacklo_epi16(_pB, _pB);
-                    __m128i _pB1 = _mm_unpackhi_epi16(_pB, _pB);
-                    __m128i _pB2 = _mm_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-                    __m128i _pB3 = _mm_shuffle_epi32(_pB1, _MM_SHUFFLE(0, 3, 2, 1));
-                    _sum0 = _mm_maccd_epi16(_pA0, _pB0, _sum0);
-                    _sum1 = _mm_maccd_epi16(_pA0, _pB1, _sum1);
-                    _sum2 = _mm_maccd_epi16(_pA0, _pB2, _sum2);
-                    _sum3 = _mm_maccd_epi16(_pA0, _pB3, _sum3);
-                    _sum4 = _mm_maccd_epi16(_pA1, _pB0, _sum4);
-                    _sum5 = _mm_maccd_epi16(_pA1, _pB1, _sum5);
-                    _sum6 = _mm_maccd_epi16(_pA1, _pB2, _sum6);
-                    _sum7 = _mm_maccd_epi16(_pA1, _pB3, _sum7);
-#else
-                    __m128i _pA0 = _pA;
-                    __m128i _pA1 = _mm_shuffle_epi32(_pA, _MM_SHUFFLE(2, 3, 0, 1));
-                    __m128i _pB01 = _pB;
-                    __m128i _pB23 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_pB, _MM_SHUFFLE(0, 3, 2, 1)), _MM_SHUFFLE(0, 3, 2, 1));
-                    __m128i _sl0 = _mm_mullo_epi16(_pA0, _pB01);
-                    __m128i _sh0 = _mm_mulhi_epi16(_pA0, _pB01);
-                    __m128i _sl1 = _mm_mullo_epi16(_pA0, _pB23);
-                    __m128i _sh1 = _mm_mulhi_epi16(_pA0, _pB23);
-                    __m128i _sl2 = _mm_mullo_epi16(_pA1, _pB01);
-                    __m128i _sh2 = _mm_mulhi_epi16(_pA1, _pB01);
-                    __m128i _sl3 = _mm_mullo_epi16(_pA1, _pB23);
-                    __m128i _sh3 = _mm_mulhi_epi16(_pA1, _pB23);
-                    __m128i _s0 = _mm_unpacklo_epi16(_sl0, _sh0);
-                    __m128i _s1 = _mm_unpackhi_epi16(_sl0, _sh0);
-                    __m128i _s2 = _mm_unpacklo_epi16(_sl1, _sh1);
-                    __m128i _s3 = _mm_unpackhi_epi16(_sl1, _sh1);
-                    __m128i _s4 = _mm_unpacklo_epi16(_sl2, _sh2);
-                    __m128i _s5 = _mm_unpackhi_epi16(_sl2, _sh2);
-                    __m128i _s6 = _mm_unpacklo_epi16(_sl3, _sh3);
-                    __m128i _s7 = _mm_unpackhi_epi16(_sl3, _sh3);
-                    _sum0 = _mm_add_epi32(_sum0, _s0);
-                    _sum1 = _mm_add_epi32(_sum1, _s1);
-                    _sum2 = _mm_add_epi32(_sum2, _s2);
-                    _sum3 = _mm_add_epi32(_sum3, _s3);
-                    _sum4 = _mm_add_epi32(_sum4, _s4);
-                    _sum5 = _mm_add_epi32(_sum5, _s5);
-                    _sum6 = _mm_add_epi32(_sum6, _s6);
-                    _sum7 = _mm_add_epi32(_sum7, _s7);
-#endif
-#endif // __AVX2__
-
-                    pA += 4;
-                    pB += 8;
-                }
-
-#if __AVX2__
-                if (k_end)
-                {
-                    // from
-                    //      00 11 22 33 04 15 26 37
-                    //      01 12 23 30 05 16 27 34
-                    //      20 31 02 13 24 35 06 17
-                    //      21 32 03 10 25 36 07 14
-                    // to
-                    //      00 10 20 30 04 14 24 34
-                    //      01 11 21 31 05 15 25 35
-                    //      02 12 22 32 06 16 26 36
-                    //      03 13 23 33 07 17 27 37
-                    {
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        __m256i _tmp0 = _mm256_unpacklo_epi32(_sum0, _sum3);
-                        __m256i _tmp1 = _mm256_unpackhi_epi32(_sum0, _sum3);
-                        __m256i _tmp2 = _mm256_unpacklo_epi32(_sum2, _sum1);
-                        __m256i _tmp3 = _mm256_unpackhi_epi32(_sum2, _sum1);
-                        _sum0 = _mm256_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm256_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm256_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm256_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        _tmp0 = _mm256_permute2x128_si256(_sum0, _sum1, _MM_SHUFFLE(0, 2, 0, 0));
-                        _tmp1 = _mm256_permute2x128_si256(_sum2, _sum3, _MM_SHUFFLE(0, 2, 0, 0));
-                        _tmp2 = _mm256_permute2x128_si256(_sum0, _sum1, _MM_SHUFFLE(0, 3, 0, 1));
-                        _tmp3 = _mm256_permute2x128_si256(_sum2, _sum3, _MM_SHUFFLE(0, 3, 0, 1));
-                        _sum0 = _tmp0;
-                        _sum1 = _tmp1;
-                        _sum2 = _tmp2;
-                        _sum3 = _tmp3;
-                    }
-                }
-
-                _mm256_storeu_si256((__m256i*)outptr, _sum0);
-                _mm256_storeu_si256((__m256i*)(outptr + 8), _sum1);
-                _mm256_storeu_si256((__m256i*)(outptr + 16), _sum2);
-                _mm256_storeu_si256((__m256i*)(outptr + 24), _sum3);
-                outptr += 32;
-#else  // __AVX2__
-                if (k_end)
-                {
-                    // from
-                    //      00 11 22 33  04 15 26 37
-                    //      01 12 23 30  05 16 27 34
-                    //      20 31 02 13  24 35 06 17
-                    //      21 32 03 10  25 36 07 14
-                    // to
-                    //      00 10 20 30  04 14 24 34
-                    //      01 11 21 31  05 15 25 35
-                    //      02 12 22 32  06 16 26 36
-                    //      03 13 23 33  07 17 27 37
-                    {
-                        _sum2 = _mm_shuffle_epi32(_sum2, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum6 = _mm_shuffle_epi32(_sum6, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum7 = _mm_shuffle_epi32(_sum7, _MM_SHUFFLE(2, 1, 0, 3));
-                        __m128i _tmp0 = _mm_unpacklo_epi32(_sum0, _sum6);
-                        __m128i _tmp1 = _mm_unpackhi_epi32(_sum0, _sum6);
-                        __m128i _tmp2 = _mm_unpacklo_epi32(_sum1, _sum7);
-                        __m128i _tmp3 = _mm_unpackhi_epi32(_sum1, _sum7);
-                        __m128i _tmp4 = _mm_unpacklo_epi32(_sum4, _sum2);
-                        __m128i _tmp5 = _mm_unpackhi_epi32(_sum4, _sum2);
-                        __m128i _tmp6 = _mm_unpacklo_epi32(_sum5, _sum3);
-                        __m128i _tmp7 = _mm_unpackhi_epi32(_sum5, _sum3);
-                        _sum0 = _mm_unpacklo_epi64(_tmp0, _tmp4);
-                        _sum1 = _mm_unpackhi_epi64(_tmp0, _tmp4);
-                        _sum2 = _mm_unpacklo_epi64(_tmp5, _tmp1);
-                        _sum3 = _mm_unpackhi_epi64(_tmp5, _tmp1);
-                        _sum4 = _mm_unpacklo_epi64(_tmp2, _tmp6);
-                        _sum5 = _mm_unpackhi_epi64(_tmp2, _tmp6);
-                        _sum6 = _mm_unpacklo_epi64(_tmp7, _tmp3);
-                        _sum7 = _mm_unpackhi_epi64(_tmp7, _tmp3);
-                        _sum1 = _mm_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum5 = _mm_shuffle_epi32(_sum5, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum7 = _mm_shuffle_epi32(_sum7, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-                }
-
-                _mm_store_si128((__m128i*)outptr, _sum0);
-                _mm_store_si128((__m128i*)(outptr + 4), _sum1);
-                _mm_store_si128((__m128i*)(outptr + 8), _sum2);
-                _mm_store_si128((__m128i*)(outptr + 12), _sum3);
-                _mm_store_si128((__m128i*)(outptr + 16), _sum4);
-                _mm_store_si128((__m128i*)(outptr + 20), _sum5);
-                _mm_store_si128((__m128i*)(outptr + 24), _sum6);
-                _mm_store_si128((__m128i*)(outptr + 28), _sum7);
-                outptr += 32;
-#endif // __AVX2__
-            }
-#endif // defined(__x86_64__) || defined(_M_X64)
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const short* pA = pAT;
-
-                __m128i _sum0;
-                __m128i _sum1;
-                __m128i _sum2;
-                __m128i _sum3;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm_setzero_si128();
-                    _sum1 = _mm_setzero_si128();
-                    _sum2 = _mm_setzero_si128();
-                    _sum3 = _mm_setzero_si128();
-                }
-                else
-                {
-                    _sum0 = _mm_load_si128((const __m128i*)outptr);
-                    _sum1 = _mm_load_si128((const __m128i*)(outptr + 4));
-                    _sum2 = _mm_load_si128((const __m128i*)(outptr + 8));
-                    _sum3 = _mm_load_si128((const __m128i*)(outptr + 12));
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m128i _pA0 = _mm_loadu_si128((const __m128i*)pA);
-                    __m128i _pB0 = _mm_loadu_si128((const __m128i*)pB);
-                    __m128i _pA1 = _mm_shuffle_epi32(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-                    __m128i _pB1 = _mm_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-
-#if __XOP__
-                    _sum0 = _mm_maddd_epi16(_pA0, _pB0, _sum0);
-                    _sum1 = _mm_maddd_epi16(_pA0, _pB1, _sum1);
-                    _sum2 = _mm_maddd_epi16(_pA1, _pB0, _sum2);
-                    _sum3 = _mm_maddd_epi16(_pA1, _pB1, _sum3);
-#else
-                    _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA0, _pB0));
-                    _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_pA0, _pB1));
-                    _sum2 = _mm_add_epi32(_sum2, _mm_madd_epi16(_pA1, _pB0));
-                    _sum3 = _mm_add_epi32(_sum3, _mm_madd_epi16(_pA1, _pB1));
-#endif
-
-                    pA += 8;
-                    pB += 8;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m128i _pA = _mm_castpd_si128(_mm_load1_pd((const double*)pA));
-                    __m128i _pB = _mm_castpd_si128(_mm_load1_pd((const double*)pB));
-#if __XOP__
-                    __m128i _pA0 = _mm_unpacklo_epi16(_pA, _pA);
-                    __m128i _pA1 = _mm_shuffle_epi32(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-                    __m128i _pB0 = _mm_unpacklo_epi16(_pB, _pB);
-                    __m128i _pB1 = _mm_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-                    _sum0 = _mm_maccd_epi16(_pA0, _pB0, _sum0);
-                    _sum1 = _mm_maccd_epi16(_pA0, _pB1, _sum1);
-                    _sum2 = _mm_maccd_epi16(_pA1, _pB0, _sum2);
-                    _sum3 = _mm_maccd_epi16(_pA1, _pB1, _sum3);
-#else
-                    __m128i _pA0 = _pA;
-                    __m128i _pA1 = _mm_shuffle_epi32(_pA, _MM_SHUFFLE(2, 3, 0, 1));
-                    __m128i _pB01 = _mm_shufflehi_epi16(_pB, _MM_SHUFFLE(0, 3, 2, 1));
-                    __m128i _sl0 = _mm_mullo_epi16(_pA0, _pB01);
-                    __m128i _sh0 = _mm_mulhi_epi16(_pA0, _pB01);
-                    __m128i _sl1 = _mm_mullo_epi16(_pA1, _pB01);
-                    __m128i _sh1 = _mm_mulhi_epi16(_pA1, _pB01);
-                    __m128i _s0 = _mm_unpacklo_epi16(_sl0, _sh0);
-                    __m128i _s1 = _mm_unpackhi_epi16(_sl0, _sh0);
-                    __m128i _s2 = _mm_unpacklo_epi16(_sl1, _sh1);
-                    __m128i _s3 = _mm_unpackhi_epi16(_sl1, _sh1);
-                    _sum0 = _mm_add_epi32(_sum0, _s0);
-                    _sum1 = _mm_add_epi32(_sum1, _s1);
-                    _sum2 = _mm_add_epi32(_sum2, _s2);
-                    _sum3 = _mm_add_epi32(_sum3, _s3);
-#endif
-
-                    pA += 4;
-                    pB += 4;
-                }
-
-                if (k_end)
-                {
-                    // from
-                    //      00 11 22 33
-                    //      01 12 23 30
-                    //      20 31 02 13
-                    //      21 32 03 10
-                    // to
-                    //      00 10 20 30
-                    //      01 11 21 31
-                    //      02 12 22 32
-                    //      03 13 23 33
-                    {
-                        _sum1 = _mm_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        __m128i _tmp0 = _mm_unpacklo_epi32(_sum0, _sum3);
-                        __m128i _tmp1 = _mm_unpackhi_epi32(_sum0, _sum3);
-                        __m128i _tmp2 = _mm_unpacklo_epi32(_sum2, _sum1);
-                        __m128i _tmp3 = _mm_unpackhi_epi32(_sum2, _sum1);
-                        _sum0 = _mm_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-                }
-
-                _mm_store_si128((__m128i*)outptr, _sum0);
-                _mm_store_si128((__m128i*)(outptr + 4), _sum1);
-                _mm_store_si128((__m128i*)(outptr + 8), _sum2);
-                _mm_store_si128((__m128i*)(outptr + 12), _sum3);
-                outptr += 16;
-            }
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const short* pA = pAT;
-
-                __m128i _sum0;
-                __m128i _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm_setzero_si128();
-                    _sum1 = _mm_setzero_si128();
-                }
-                else
-                {
-                    _sum0 = _mm_load_si128((const __m128i*)outptr);
-                    _sum1 = _mm_load_si128((const __m128i*)(outptr + 4));
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m128i _pA = _mm_loadu_si128((const __m128i*)pA);
-                    __m128i _pB0 = _mm_castpd_si128(_mm_load1_pd((const double*)pB));
-                    __m128i _pB1 = _mm_shuffle_epi32(_pB0, _MM_SHUFFLE(2, 3, 0, 1));
-
-#if __XOP__
-                    _sum0 = _mm_maddd_epi16(_pA, _pB0, _sum0);
-                    _sum1 = _mm_maddd_epi16(_pA, _pB1, _sum1);
-#else
-                    _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA, _pB0));
-                    _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_pA, _pB1));
-#endif
-
-                    pA += 8;
-                    pB += 4;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m128i _pA = _mm_castpd_si128(_mm_load1_pd((const double*)pA));
-                    __m128i _pB = _mm_castps_si128(_mm_load1_ps((const float*)pB));
-
-#if __XOP__
-                    _pA = _mm_unpacklo_epi16(_pA, _pA);
-                    __m128i _pB0 = _mm_unpacklo_epi16(_pB, _pB);
-                    __m128i _pB1 = _mm_shuffle_epi32(_pB0, _MM_SHUFFLE(2, 3, 0, 1));
-                    _sum0 = _mm_maccd_epi16(_pA, _pB0, _sum0);
-                    _sum1 = _mm_maccd_epi16(_pA, _pB1, _sum1);
-#else
-                    __m128i _pB01 = _mm_shufflehi_epi16(_pB, _MM_SHUFFLE(0, 1, 0, 1));
-                    __m128i _sl = _mm_mullo_epi16(_pA, _pB01);
-                    __m128i _sh = _mm_mulhi_epi16(_pA, _pB01);
-                    __m128i _s0 = _mm_unpacklo_epi16(_sl, _sh);
-                    __m128i _s1 = _mm_unpackhi_epi16(_sl, _sh);
-                    _sum0 = _mm_add_epi32(_sum0, _s0);
-                    _sum1 = _mm_add_epi32(_sum1, _s1);
-#endif
-
-                    pA += 4;
-                    pB += 2;
-                }
-
-                if (k_end)
-                {
-                    // from
-                    //      00 11 20 31
-                    //      01 10 21 30
-                    // to
-                    //      00 10 20 30
-                    //      01 11 21 31
-                    {
-                        __m128i _tmp0 = _mm_shuffle_epi32(_sum0, _MM_SHUFFLE(3, 1, 2, 0));
-                        __m128i _tmp1 = _mm_shuffle_epi32(_sum1, _MM_SHUFFLE(0, 2, 3, 1));
-                        _sum0 = _mm_unpacklo_epi32(_tmp0, _tmp1);
-                        _sum1 = _mm_unpackhi_epi32(_tmp0, _tmp1);
-                        _sum1 = _mm_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-                }
-
-                _mm_store_si128((__m128i*)outptr, _sum0);
-                _mm_store_si128((__m128i*)(outptr + 4), _sum1);
-                outptr += 8;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const short* pA = pAT;
-
-                __m128i _sum0;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm_setzero_si128();
-                }
-                else
-                {
-                    _sum0 = _mm_load_si128((const __m128i*)outptr);
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m128i _pA = _mm_loadu_si128((const __m128i*)pA);
-                    __m128i _pB = _mm_castps_si128(_mm_load1_ps((const float*)pB));
-
-#if __XOP__
-                    _sum0 = _mm_maddd_epi16(_pA, _pB, _sum0);
-#else
-                    _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA, _pB));
-#endif
-
-                    pA += 8;
-                    pB += 2;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m128i _pA = _mm_loadl_epi64((const __m128i*)pA);
-                    __m128i _pB = _mm_set1_epi16(pB[0]);
-
-#if __XOP__
-                    _pA = _mm_unpacklo_epi16(_pA, _pA);
-                    _sum0 = _mm_maccd_epi16(_pA, _pB, _sum0);
-#else
-                    __m128i _sl = _mm_mullo_epi16(_pA, _pB);
-                    __m128i _sh = _mm_mulhi_epi16(_pA, _pB);
-                    __m128i _s0 = _mm_unpacklo_epi16(_sl, _sh);
-                    _sum0 = _mm_add_epi32(_sum0, _s0);
-#endif
-
-                    pA += 4;
-                    pB += 1;
-                }
-
-                _mm_store_si128((__m128i*)outptr, _sum0);
-                outptr += 4;
-            }
-        }
-    }
-#endif // __SSE2__
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const short* pAT = AT_tile.row<const short>(b) + max_kk * ii;
-            const short* pB = BT_tile.row<const short>(b);
-
-            int jj = 0;
-#if __SSE2__
-#if defined(__x86_64__) || defined(_M_X64)
-#if __AVX512F__
-            for (; jj + 15 < max_jj; jj += 16)
-            {
-                const short* pA = pAT;
-
-                __m512i _sum0;
-                __m512i _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm512_setzero_si512();
-                    _sum1 = _mm512_setzero_si512();
-                }
-                else
-                {
-                    _sum0 = _mm512_loadu_si512((const __m512i*)outptr);
-                    _sum1 = _mm512_loadu_si512((const __m512i*)(outptr + 16));
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m512i _pA0 = _mm512_set1_epi32(((const int*)pA)[0]);
-                    __m512i _pA1 = _mm512_set1_epi32(((const int*)pA)[1]);
-                    __m512i _pB0 = _mm512_loadu_si512((const __m512i*)pB);
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _pA0, _pB0);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _pA1, _pB0);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA0, _pB0));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA1, _pB0));
-#endif // __AVX512VNNI__
-
-                    pA += 4;
-                    pB += 32;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m512i _pA0 = _mm512_set1_epi32(pA[0]);
-                    __m512i _pA1 = _mm512_set1_epi32(pA[1]);
-                    __m256i _pB = _mm256_loadu_si256((const __m256i*)pB);
-                    __m512i _pB0 = _mm512_cvtepi16_epi32(_pB);
-
-                    __m512i _s0 = _mm512_mullo_epi32(_pA0, _pB0);
-                    __m512i _s1 = _mm512_mullo_epi32(_pA1, _pB0);
-                    _sum0 = _mm512_add_epi32(_sum0, _s0);
-                    _sum1 = _mm512_add_epi32(_sum1, _s1);
-
-                    pA += 2;
-                    pB += 16;
-                }
-
-                if (k_end)
-                {
-                    __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum1);
-                    __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum1);
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(1, 0, 1, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(3, 2, 3, 2));
-                    _sum0 = _mm512_shuffle_i32x4(_sum0, _sum0, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_sum1, _sum1, _MM_SHUFFLE(3, 1, 2, 0));
-                }
-
-                _mm512_storeu_si512((__m512i*)outptr, _sum0);
-                _mm512_storeu_si512((__m512i*)(outptr + 16), _sum1);
-                outptr += 32;
-            }
-#endif // __AVX512F__
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const short* pA = pAT;
-
-#if __AVX2__
-                __m256i _sum0;
-                __m256i _sum1;
-#else
-                __m128i _sum0;
-                __m128i _sum1;
-                __m128i _sum2;
-                __m128i _sum3;
-#endif
-
-                if (k == 0)
-                {
-#if __AVX2__
-                    _sum0 = _mm256_setzero_si256();
-                    _sum1 = _mm256_setzero_si256();
-#else
-                    _sum0 = _mm_setzero_si128();
-                    _sum1 = _mm_setzero_si128();
-                    _sum2 = _mm_setzero_si128();
-                    _sum3 = _mm_setzero_si128();
-#endif
-                }
-                else
-                {
-#if __AVX2__
-                    _sum0 = _mm256_loadu_si256((const __m256i*)outptr);
-                    _sum1 = _mm256_loadu_si256((const __m256i*)(outptr + 8));
-#else
-                    _sum0 = _mm_loadu_si128((const __m128i*)outptr);
-                    _sum1 = _mm_loadu_si128((const __m128i*)(outptr + 4));
-                    _sum2 = _mm_loadu_si128((const __m128i*)(outptr + 8));
-                    _sum3 = _mm_loadu_si128((const __m128i*)(outptr + 12));
-#endif
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-#if __AVX2__
-                    __m256i _pA0 = _mm256_castps_si256(_mm256_broadcast_ss((const float*)pA));
-                    __m256i _pA1 = _mm256_castps_si256(_mm256_broadcast_ss((const float*)(pA + 2)));
-                    __m256i _pB0 = _mm256_loadu_si256((const __m256i*)pB);
-
-                    // vs2019 internal compiler error with avx512 vnni intrinsics here
-                    // fallback to avx2 madd anyway as a workaround   --- nihui
-                    _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_pA0, _pB0));
-                    _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_pA1, _pB0));
-#else  // __AVX2__
-                    __m128i _pA0 = _mm_castps_si128(_mm_load1_ps((const float*)pA));
-                    __m128i _pA1 = _mm_castps_si128(_mm_load1_ps((const float*)(pA + 2)));
-                    __m128i _pB0 = _mm_loadu_si128((const __m128i*)pB);
-                    __m128i _pB1 = _mm_loadu_si128((const __m128i*)(pB + 8));
-                    _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA0, _pB0));
-                    _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_pA0, _pB1));
-                    _sum2 = _mm_add_epi32(_sum2, _mm_madd_epi16(_pA1, _pB0));
-                    _sum3 = _mm_add_epi32(_sum3, _mm_madd_epi16(_pA1, _pB1));
-#endif // __AVX2__
-
-                    pA += 4;
-                    pB += 16;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m128i _pB = _mm_load_si128((const __m128i*)pB);
-#if __AVX2__
-                    __m256i _pA0 = _mm256_set1_epi32(pA[0]);
-                    __m256i _pA1 = _mm256_set1_epi32(pA[1]);
-                    __m256i _pB0 = _mm256_cvtepi16_epi32(_pB);
-
-                    __m256i _s0 = _mm256_mullo_epi32(_pA0, _pB0);
-                    __m256i _s1 = _mm256_mullo_epi32(_pA1, _pB0);
-                    _sum0 = _mm256_add_epi32(_sum0, _s0);
-                    _sum1 = _mm256_add_epi32(_sum1, _s1);
-#else  // __AVX2__
-                    __m128i _pA0 = _mm_set1_epi16(pA[0]);
-                    __m128i _pA1 = _mm_set1_epi16(pA[1]);
-
-                    __m128i _sl0 = _mm_mullo_epi16(_pA0, _pB);
-                    __m128i _sh0 = _mm_mulhi_epi16(_pA0, _pB);
-                    __m128i _sl1 = _mm_mullo_epi16(_pA1, _pB);
-                    __m128i _sh1 = _mm_mulhi_epi16(_pA1, _pB);
-                    __m128i _s0 = _mm_unpacklo_epi16(_sl0, _sh0);
-                    __m128i _s1 = _mm_unpackhi_epi16(_sl0, _sh0);
-                    __m128i _s2 = _mm_unpacklo_epi16(_sl1, _sh1);
-                    __m128i _s3 = _mm_unpackhi_epi16(_sl1, _sh1);
-                    _sum0 = _mm_add_epi32(_sum0, _s0);
-                    _sum1 = _mm_add_epi32(_sum1, _s1);
-                    _sum2 = _mm_add_epi32(_sum2, _s2);
-                    _sum3 = _mm_add_epi32(_sum3, _s3);
-#endif // __AVX2__
-                    pA += 2;
-                    pB += 8;
-                }
-
-#if __AVX2__
-                if (k_end)
-                {
-                    __m256i _tmp0 = _mm256_unpacklo_epi32(_sum0, _sum1);
-                    __m256i _tmp1 = _mm256_unpackhi_epi32(_sum0, _sum1);
-                    _sum0 = _mm256_permute2x128_si256(_tmp0, _tmp1, _MM_SHUFFLE(0, 2, 0, 0));
-                    _sum1 = _mm256_permute2x128_si256(_tmp0, _tmp1, _MM_SHUFFLE(0, 3, 0, 1));
-                }
-
-                _mm256_storeu_si256((__m256i*)outptr, _sum0);
-                _mm256_storeu_si256((__m256i*)(outptr + 8), _sum1);
-                outptr += 16;
-#else  // __AVX2__
-                if (k_end)
-                {
-                    __m128i _tmp0 = _mm_unpacklo_epi32(_sum0, _sum2);
-                    __m128i _tmp1 = _mm_unpackhi_epi32(_sum0, _sum2);
-                    __m128i _tmp2 = _mm_unpacklo_epi32(_sum1, _sum3);
-                    __m128i _tmp3 = _mm_unpackhi_epi32(_sum1, _sum3);
-                    _sum0 = _tmp0;
-                    _sum1 = _tmp1;
-                    _sum2 = _tmp2;
-                    _sum3 = _tmp3;
-                }
-
-                _mm_storeu_si128((__m128i*)outptr, _sum0);
-                _mm_storeu_si128((__m128i*)(outptr + 4), _sum1);
-                _mm_storeu_si128((__m128i*)(outptr + 8), _sum2);
-                _mm_storeu_si128((__m128i*)(outptr + 12), _sum3);
-                outptr += 16;
-#endif // __AVX2__
-            }
-#endif // defined(__x86_64__) || defined(_M_X64)
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const short* pA = pAT;
-
-                __m128i _sum0;
-                __m128i _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm_setzero_si128();
-                    _sum1 = _mm_setzero_si128();
-                }
-                else
-                {
-                    _sum0 = _mm_loadu_si128((const __m128i*)outptr);
-                    _sum1 = _mm_loadu_si128((const __m128i*)(outptr + 4));
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m128i _pA0 = _mm_castps_si128(_mm_load1_ps((const float*)pA));
-                    __m128i _pA1 = _mm_castps_si128(_mm_load1_ps((const float*)(pA + 2)));
-                    __m128i _pB = _mm_loadu_si128((const __m128i*)pB);
-                    _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA0, _pB));
-                    _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_pA1, _pB));
-
-                    pA += 4;
-                    pB += 8;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m128i _pA0 = _mm_set1_epi16(pA[0]);
-                    __m128i _pA1 = _mm_set1_epi16(pA[1]);
-                    __m128i _pB = _mm_loadl_epi64((const __m128i*)pB);
-                    __m128i _sl0 = _mm_mullo_epi16(_pA0, _pB);
-                    __m128i _sh0 = _mm_mulhi_epi16(_pA0, _pB);
-                    __m128i _sl1 = _mm_mullo_epi16(_pA1, _pB);
-                    __m128i _sh1 = _mm_mulhi_epi16(_pA1, _pB);
-                    __m128i _s0 = _mm_unpacklo_epi16(_sl0, _sh0);
-                    __m128i _s1 = _mm_unpacklo_epi16(_sl1, _sh1);
-                    _sum0 = _mm_add_epi32(_sum0, _s0);
-                    _sum1 = _mm_add_epi32(_sum1, _s1);
-                    pA += 2;
-                    pB += 4;
-                }
-
-                if (k_end)
-                {
-                    __m128i _tmp0 = _mm_unpacklo_epi32(_sum0, _sum1);
-                    __m128i _tmp1 = _mm_unpackhi_epi32(_sum0, _sum1);
-                    _sum0 = _tmp0;
-                    _sum1 = _tmp1;
-                }
-
-                _mm_storeu_si128((__m128i*)outptr, _sum0);
-                _mm_storeu_si128((__m128i*)(outptr + 4), _sum1);
-                outptr += 2 * 4;
-            }
-#endif // __SSE2__
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const short* pA = pAT;
-
-                int sum00 = 0;
-                int sum01 = 0;
-                int sum10 = 0;
-                int sum11 = 0;
-
-                if (k == 0)
-                {
-                    sum00 = 0;
-                    sum01 = 0;
-                    sum10 = 0;
-                    sum11 = 0;
-                }
-                else
-                {
-                    sum00 = outptr[0];
-                    sum01 = outptr[1];
-                    sum10 = outptr[2];
-                    sum11 = outptr[3];
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    sum00 += pA[0] * pB[0];
-                    sum00 += pA[1] * pB[1];
-                    sum01 += pA[2] * pB[0];
-                    sum01 += pA[3] * pB[1];
-                    sum10 += pA[0] * pB[2];
-                    sum10 += pA[1] * pB[3];
-                    sum11 += pA[2] * pB[2];
-                    sum11 += pA[3] * pB[3];
-
-                    pA += 4;
-                    pB += 4;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    sum00 += pA[0] * pB[0];
-                    sum01 += pA[1] * pB[0];
-                    sum10 += pA[0] * pB[1];
-                    sum11 += pA[1] * pB[1];
-                    pA += 2;
-                    pB += 2;
-                }
-
-                outptr[0] = sum00;
-                outptr[1] = sum01;
-                outptr[2] = sum10;
-                outptr[3] = sum11;
-                outptr += 2 * 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const short* pA = pAT;
-
-                int sum0 = 0;
-                int sum1 = 0;
-
-                if (k == 0)
-                {
-                    sum0 = 0;
-                    sum1 = 0;
-                }
-                else
-                {
-                    sum0 = outptr[0];
-                    sum1 = outptr[1];
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    sum0 += pA[0] * pB[0];
-                    sum0 += pA[1] * pB[1];
-                    sum1 += pA[2] * pB[0];
-                    sum1 += pA[3] * pB[1];
-                    pA += 4;
-                    pB += 2;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    sum0 += pA[0] * pB[0];
-                    sum1 += pA[1] * pB[0];
-                    pA += 2;
-                    pB += 1;
-                }
-
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-                outptr += 2;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        for (int b = 0; b < batch; b++)
-        {
-            const short* pAT = AT_tile.row<const short>(b) + max_kk * ii;
-            const short* pB = BT_tile.row<const short>(b);
-
-            int jj = 0;
-#if __SSE2__
-#if defined(__x86_64__) || defined(_M_X64)
-#if __AVX512F__
-            for (; jj + 15 < max_jj; jj += 16)
-            {
-                const short* pA = pAT;
-
-                __m512i _sum0;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm512_setzero_si512();
-                }
-                else
-                {
-                    _sum0 = _mm512_loadu_si512((const __m512i*)outptr);
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m512i _pA0 = _mm512_set1_epi32(((const int*)pA)[0]);
-                    __m512i _pB0 = _mm512_loadu_si512((const __m512i*)pB);
-
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _pA0, _pB0);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA0, _pB0));
-#endif
-
-                    pA += 2;
-                    pB += 32;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m512i _pA = _mm512_set1_epi32(pA[0]);
-                    __m512i _pB0 = _mm512_cvtepi16_epi32(_mm256_loadu_si256((const __m256i*)pB));
-
-                    __m512i _s0 = _mm512_mullo_epi32(_pA, _pB0);
-                    _sum0 = _mm512_add_epi32(_sum0, _s0);
-
-                    pA += 1;
-                    pB += 16;
-                }
-
-                _mm512_storeu_si512((__m512i*)outptr, _sum0);
-                outptr += 16;
-            }
-#endif // __AVX512F__
-            for (; jj + 7 < max_jj; jj += 8)
-            {
-                const short* pA = pAT;
-
-                __m128i _sum0;
-                __m128i _sum1;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm_setzero_si128();
-                    _sum1 = _mm_setzero_si128();
-                }
-                else
-                {
-                    _sum0 = _mm_loadu_si128((const __m128i*)outptr);
-                    _sum1 = _mm_loadu_si128((const __m128i*)(outptr + 4));
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m128i _pA = _mm_castps_si128(_mm_load1_ps((const float*)pA));
-                    __m128i _pB0 = _mm_loadu_si128((const __m128i*)pB);
-                    __m128i _pB1 = _mm_loadu_si128((const __m128i*)(pB + 8));
-                    _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA, _pB0));
-                    _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_pA, _pB1));
-
-                    pA += 2;
-                    pB += 16;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m128i _pA = _mm_set1_epi16(pA[0]);
-                    __m128i _pB = _mm_load_si128((const __m128i*)pB);
-                    __m128i _sl = _mm_mullo_epi16(_pA, _pB);
-                    __m128i _sh = _mm_mulhi_epi16(_pA, _pB);
-                    __m128i _s0 = _mm_unpacklo_epi16(_sl, _sh);
-                    __m128i _s1 = _mm_unpackhi_epi16(_sl, _sh);
-                    _sum0 = _mm_add_epi32(_sum0, _s0);
-                    _sum1 = _mm_add_epi32(_sum1, _s1);
-                    pA += 1;
-                    pB += 8;
-                }
-
-                _mm_storeu_si128((__m128i*)outptr, _sum0);
-                _mm_storeu_si128((__m128i*)(outptr + 4), _sum1);
-                outptr += 8;
-            }
-#endif // defined(__x86_64__) || defined(_M_X64)
-            for (; jj + 3 < max_jj; jj += 4)
-            {
-                const short* pA = pAT;
-
-                __m128i _sum0;
-
-                if (k == 0)
-                {
-                    _sum0 = _mm_setzero_si128();
-                }
-                else
-                {
-                    _sum0 = _mm_loadu_si128((const __m128i*)outptr);
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    __m128i _pA = _mm_castps_si128(_mm_load1_ps((const float*)pA));
-                    __m128i _pB = _mm_loadu_si128((const __m128i*)pB);
-                    _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA, _pB));
-
-                    pA += 2;
-                    pB += 8;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    __m128i _pA = _mm_set1_epi16(pA[0]);
-                    __m128i _pB = _mm_loadl_epi64((const __m128i*)pB);
-                    __m128i _sl = _mm_mullo_epi16(_pA, _pB);
-                    __m128i _sh = _mm_mulhi_epi16(_pA, _pB);
-                    __m128i _s0 = _mm_unpacklo_epi16(_sl, _sh);
-                    _sum0 = _mm_add_epi32(_sum0, _s0);
-                    pA += 1;
-                    pB += 4;
-                }
-
-                _mm_storeu_si128((__m128i*)outptr, _sum0);
-                outptr += 4;
-            }
-#endif // __SSE2__
-            for (; jj + 1 < max_jj; jj += 2)
-            {
-                const short* pA = pAT;
-
-                int sum0 = 0;
-                int sum1 = 0;
-
-                if (k == 0)
-                {
-                    sum0 = 0;
-                    sum1 = 0;
-                }
-                else
-                {
-                    sum0 = outptr[0];
-                    sum1 = outptr[1];
-                }
-
-                int kk = 0;
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    sum0 += pA[0] * pB[0];
-                    sum0 += pA[1] * pB[1];
-                    sum1 += pA[0] * pB[2];
-                    sum1 += pA[1] * pB[3];
-                    pA += 2;
-                    pB += 4;
-                }
-                for (; kk < max_kk; kk++)
-                {
-                    sum0 += pA[0] * pB[0];
-                    sum1 += pA[0] * pB[1];
-                    pA += 1;
-                    pB += 2;
-                }
-
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-                outptr += 2;
-            }
-            for (; jj < max_jj; jj++)
-            {
-                const short* pA = pAT;
-
-                int sum = 0;
-
-                if (k == 0)
-                {
-                    sum = 0;
-                }
-                else
-                {
-                    sum = outptr[0];
-                }
-
-                int kk = 0;
-                for (; kk < max_kk; kk++)
-                {
-                    sum += pA[0] * pB[0];
-                    pA += 1;
-                    pB += 1;
-                }
-
-                outptr[0] = sum;
-                outptr += 1;
-            }
-        }
-    }
-}
-
-static void get_optimal_tile_mnk_int8(int M, int N, int K, int& TILE_M, int& TILE_N, int& TILE_K, int nT)
-{
-    // resolve optimal tile size from cache size
-    const size_t l2_cache_size_int8 = (int)(get_cpu_level2_cache_size() / sizeof(short));
-
-    if (nT == 0)
-        nT = get_physical_big_cpu_count();
-
-    // solve M
-    {
-        int tile_size = (int)sqrt((float)l2_cache_size_int8 / 3);
-
-#if __AVX512F__
-        TILE_M = std::max(16, tile_size / 16 * 16);
-#elif __AVX2__
-        TILE_M = std::max(8, tile_size / 8 * 8);
-#elif __SSE2__
-        TILE_M = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_M = std::max(2, tile_size / 2 * 2);
-#endif
-
-        TILE_M *= std::min(nT, get_physical_cpu_count());
-
-        int nn_M = (M + TILE_M - 1) / TILE_M;
-#if __AVX512F__
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 15) / 16 * 16);
-#elif __AVX2__
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-#elif __SSE2__
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 3) / 4 * 4);
-#else
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 1) / 2 * 2);
-#endif
-
-        if (nT > 1)
-        {
-#if __AVX512F__
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 15) / 16 * 16);
-#elif __AVX2__
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 7) / 8 * 8);
-#elif __SSE2__
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 3) / 4 * 4);
-#else
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 1) / 2 * 2);
-#endif
-        }
-    }
-
-    // solve K
-    {
-        int tile_size = (int)(sqrt((float)l2_cache_size_int8) - TILE_M);
-
-#if __AVX512F__
-        TILE_K = std::max(16, tile_size / 16 * 16);
-#elif __AVX2__
-        TILE_K = std::max(8, tile_size / 8 * 8);
-#elif __SSE2__
-        TILE_K = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_K = std::max(2, tile_size / 2 * 2);
-#endif
-
-        int nn_K = (K + TILE_K - 1) / TILE_K;
-#if __AVX512F__
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 15) / 16 * 16);
-#elif __AVX2__
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 7) / 8 * 8);
-#elif __SSE2__
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 3) / 4 * 4);
-#else
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 1) / 2 * 2);
-#endif
-    }
-
-    if (N > 0)
-    {
-        int tile_size = (int)((l2_cache_size_int8 - TILE_M * TILE_K) / (TILE_M * 2 + TILE_K));
-
-#if __SSE2__
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_N = std::max(1, tile_size);
-#endif
-
-        int nn_N = (N + TILE_N - 1) / TILE_N;
-#if __SSE2__
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#else
-        TILE_N = std::min(TILE_N, (N + nn_N - 1) / nn_N);
-#endif
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_kernel_tile_int8(const Mat& kernel, Mat& A, int inch, int i, int max_ii, int k, int max_kk)
-{
-    // const signed char ktm[4][3] = {
-    //     {2, 0, 0},
-    //     {1, 1, 1},
-    //     {1, -1, 1},
-    //     {0, 0, 2}
-    // };
-
-    short* ptmp = A;
-
-    int ii = 0;
-    for (; ii < max_ii; ii++)
-    {
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            short tmp[4][3];
-
-            const signed char* k0 = (const signed char*)kernel + (i + ii) * inch * 9 + (k + kk) * 9;
-
-            for (int m = 0; m < 3; m++)
-            {
-                signed char r0 = k0[0];
-                signed char r1 = k0[1];
-                signed char r2 = k0[2];
-
-                tmp[0][m] = r0 * 2;
-                tmp[1][m] = r0 + r1 + r2;
-                tmp[2][m] = r0 - r1 + r2;
-                tmp[3][m] = r2 * 2;
-
-                k0 += 3;
-            }
-
-            for (int m = 0; m < 4; m++)
-            {
-                short r0 = tmp[m][0];
-                short r1 = tmp[m][1];
-                short r2 = tmp[m][2];
-
-                short z0 = r0 * 2;
-                short z1 = r0 + r1 + r2;
-                short z2 = r0 - r1 + r2;
-                short z3 = r2 * 2;
-
-                ptmp[0] = z0;
-                ptmp[1] = z1;
-                ptmp[2] = z2;
-                ptmp[3] = z3;
-                ptmp += 4;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_transform_kernel_int8(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt)
-{
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-    if (ncnn::cpu_support_x86_avx2())
-    {
-        conv3x3s1_winograd23_transform_kernel_int8_avx2(kernel, AT, inch, outch, opt);
-        return;
-    }
-#endif
-#endif
-
-    const int M = outch;
-    const int K = inch;
-    const int B = 16;
-
-    int TILE_M, TILE_N, TILE_K;
-    get_optimal_tile_mnk_int8(M, 0, K, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    Mat A_tileX(B * TILE_M * TILE_K, 1, opt.num_threads, 2u, (Allocator*)0);
-
-    AT.create(TILE_K * TILE_M, B, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M, 2u, (Allocator*)0);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat A_tile = A_tileX.channel(get_omp_thread_num());
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_ii = std::min((M - i), TILE_M);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            conv3x3s1_winograd23_transform_kernel_tile_int8(kernel, A_tile, inch, i, max_ii, k, max_kk);
-
-            Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-            pack_A_tile_int8(A_tile, AT_tile, B, max_ii, max_kk);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_input_tile_int8(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    // const signed char itm[4][4] = {
-    //     {1,  0, -1,  0},
-    //     {0,  1,  1,  0},
-    //     {0, -1,  1,  0},
-    //     {0, -1,  0,  1}
-    // };
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w - 1) / 2;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-#if __SSE2__
-#if __AVX512F__
-    nn_max_kk = max_kk / 16;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = ppkk * 16;
-
-#ifdef _MSC_VER
-        __declspec(align(64))
-#else
-        __attribute__((aligned(64)))
-#endif
-        short tmp[4][4][16];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const signed char* r0 = bottom_blob.channel((k + kk) / elempack).row<const signed char>(ti * 2) + (tj * 2) * elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                __m256i _r0 = _mm256_setzero_si256();
-                __m256i _r1 = _mm256_setzero_si256();
-                __m256i _r2 = _mm256_setzero_si256();
-                __m256i _r3 = _mm256_setzero_si256();
-
-                if (ti * 2 + m < h)
-                {
-                    if (elempack == 16)
-                    {
-                        _r0 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)r0));
-                        if (tj * 2 + 1 < w) _r1 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)(r0 + 16)));
-                        if (tj * 2 + 2 < w) _r2 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)(r0 + 32)));
-                        if (tj * 2 + 3 < w) _r3 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)(r0 + 48)));
-                    }
-                    if (elempack == 8)
-                    {
-                        const signed char* r1 = r0 + N;
-
-                        _r0 = _mm256_cvtepi8_epi16(_mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i*)r0), _mm_loadl_epi64((const __m128i*)r1)));
-                        if (tj * 2 + 1 < w) _r1 = _mm256_cvtepi8_epi16(_mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i*)(r0 + 8)), _mm_loadl_epi64((const __m128i*)(r1 + 8))));
-                        if (tj * 2 + 2 < w) _r2 = _mm256_cvtepi8_epi16(_mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i*)(r0 + 16)), _mm_loadl_epi64((const __m128i*)(r1 + 16))));
-                        if (tj * 2 + 3 < w) _r3 = _mm256_cvtepi8_epi16(_mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i*)(r0 + 24)), _mm_loadl_epi64((const __m128i*)(r1 + 24))));
-                    }
-                    if (elempack == 1)
-                    {
-                        __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                        _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                        _r0 = _mm256_cvtepi8_epi16(_mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)r0, 1)));
-                        if (tj * 2 + 1 < w) _r1 = _mm256_cvtepi8_epi16(_mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0 + 1), 1)));
-                        if (tj * 2 + 2 < w) _r2 = _mm256_cvtepi8_epi16(_mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0 + 2), 1)));
-                        if (tj * 2 + 3 < w) _r3 = _mm256_cvtepi8_epi16(_mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0 + 3), 1)));
-                    }
-                }
-
-                __m256i _tmp0 = _mm256_sub_epi16(_r0, _r2);
-                __m256i _tmp1 = _mm256_add_epi16(_r1, _r2);
-                __m256i _tmp2 = _mm256_sub_epi16(_r2, _r1);
-                __m256i _tmp3 = _mm256_sub_epi16(_r3, _r1);
-
-                _mm256_store_si256((__m256i*)tmp[0][m], _tmp0);
-                _mm256_store_si256((__m256i*)tmp[1][m], _tmp1);
-                _mm256_store_si256((__m256i*)tmp[2][m], _tmp2);
-                _mm256_store_si256((__m256i*)tmp[3][m], _tmp3);
-
-                r0 += w * elempack;
-            }
-
-            short* p0 = (short*)B + kk * max_jj * 16 + jj * 16;
-            short* p1 = p0 + max_jj * 16;
-            short* p2 = p0 + max_jj * 16 * 2;
-            short* p3 = p0 + max_jj * 16 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                __m256i _r0 = _mm256_load_si256((const __m256i*)tmp[m][0]);
-                __m256i _r1 = _mm256_load_si256((const __m256i*)tmp[m][1]);
-                __m256i _r2 = _mm256_load_si256((const __m256i*)tmp[m][2]);
-                __m256i _r3 = _mm256_load_si256((const __m256i*)tmp[m][3]);
-
-                __m256i _tmp0 = _mm256_sub_epi16(_r0, _r2);
-                __m256i _tmp1 = _mm256_add_epi16(_r1, _r2);
-                __m256i _tmp2 = _mm256_sub_epi16(_r2, _r1);
-                __m256i _tmp3 = _mm256_sub_epi16(_r3, _r1);
-
-                _mm256_store_si256((__m256i*)p0, _tmp0);
-                _mm256_store_si256((__m256i*)p1, _tmp1);
-                _mm256_store_si256((__m256i*)p2, _tmp2);
-                _mm256_store_si256((__m256i*)p3, _tmp3);
-
-                p0 += max_jj * 4 * 16;
-                p1 += max_jj * 4 * 16;
-                p2 += max_jj * 4 * 16;
-                p3 += max_jj * 4 * 16;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 16;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-#else // __AVX512F__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-    #pragma omp parallel for num_threads(nT)
-#endif // __AVX512F__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-#ifdef _MSC_VER
-        __declspec(align(32))
-#else
-        __attribute__((aligned(32)))
-#endif
-        short tmp[4][4][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const signed char* r0 = bottom_blob.channel((k + kk) / elempack).row<const signed char>(ti * 2) + (tj * 2) * elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                __m128i _r0 = _mm_setzero_si128();
-                __m128i _r1 = _mm_setzero_si128();
-                __m128i _r2 = _mm_setzero_si128();
-                __m128i _r3 = _mm_setzero_si128();
-
-                if (ti * 2 + m < h)
-                {
-                    if (elempack == 8)
-                    {
-                        _r0 = _mm_loadl_epi64((const __m128i*)r0);
-                        _r0 = _mm_unpacklo_epi8(_r0, _mm_cmpgt_epi8(_mm_setzero_si128(), _r0));
-                        if (tj * 2 + 1 < w)
-                        {
-                            _r1 = _mm_loadl_epi64((const __m128i*)(r0 + 8));
-                            _r1 = _mm_unpacklo_epi8(_r1, _mm_cmpgt_epi8(_mm_setzero_si128(), _r1));
-                        }
-                        if (tj * 2 + 2 < w)
-                        {
-                            _r2 = _mm_loadl_epi64((const __m128i*)(r0 + 16));
-                            _r2 = _mm_unpacklo_epi8(_r2, _mm_cmpgt_epi8(_mm_setzero_si128(), _r2));
-                        }
-                        if (tj * 2 + 3 < w)
-                        {
-                            _r3 = _mm_loadl_epi64((const __m128i*)(r0 + 24));
-                            _r3 = _mm_unpacklo_epi8(_r3, _mm_cmpgt_epi8(_mm_setzero_si128(), _r3));
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-#if __AVX2__
-                        __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                        _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-#if __AVX512F__
-                        _r0 = _mm_cvtepi8_epi16(_mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0, _vindex, 1)));
-                        if (tj * 2 + 1 < w) _r1 = _mm_cvtepi8_epi16(_mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(r0 + 1), _vindex, 1)));
-                        if (tj * 2 + 2 < w) _r2 = _mm_cvtepi8_epi16(_mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(r0 + 2), _vindex, 1)));
-                        if (tj * 2 + 3 < w) _r3 = _mm_cvtepi8_epi16(_mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(r0 + 3), _vindex, 1)));
-#else
-                        __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-                        __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-                        __m256i _val0_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r0, _vindex, 1), _sindex88);
-                        _r0 = _mm_cvtepi8_epi16(_mm_unpacklo_epi32(_mm256_extracti128_si256(_val0_32, 0), _mm256_extracti128_si256(_val0_32, 1)));
-                        if (tj * 2 + 1 < w)
-                        {
-                            __m256i _val1_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)(r0 + 1), _vindex, 1), _sindex88);
-                            _r1 = _mm_cvtepi8_epi16(_mm_unpacklo_epi32(_mm256_extracti128_si256(_val1_32, 0), _mm256_extracti128_si256(_val1_32, 1)));
-                        }
-                        if (tj * 2 + 2 < w)
-                        {
-                            __m256i _val2_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)(r0 + 2), _vindex, 1), _sindex88);
-                            _r2 = _mm_cvtepi8_epi16(_mm_unpacklo_epi32(_mm256_extracti128_si256(_val2_32, 0), _mm256_extracti128_si256(_val2_32, 1)));
-                        }
-                        if (tj * 2 + 3 < w)
-                        {
-                            __m256i _val3_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)(r0 + 3), _vindex, 1), _sindex88);
-                            _r3 = _mm_cvtepi8_epi16(_mm_unpacklo_epi32(_mm256_extracti128_si256(_val3_32, 0), _mm256_extracti128_si256(_val3_32, 1)));
-                        }
-#endif // __AVX512F__
-#else  // __AVX2__
-                        const signed char* r1 = r0 + N;
-                        const signed char* r2 = r0 + N * 2;
-                        const signed char* r3 = r0 + N * 3;
-                        const signed char* r4 = r0 + N * 4;
-                        const signed char* r5 = r0 + N * 5;
-                        const signed char* r6 = r0 + N * 6;
-                        const signed char* r7 = r0 + N * 7;
-
-                        __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0);
-                        __m128i _t1 = _mm_loadl_epi64((const __m128i*)r1);
-                        __m128i _t2 = _mm_loadl_epi64((const __m128i*)r2);
-                        __m128i _t3 = _mm_loadl_epi64((const __m128i*)r3);
-                        __m128i _t4 = _mm_loadl_epi64((const __m128i*)r4);
-                        __m128i _t5 = _mm_loadl_epi64((const __m128i*)r5);
-                        __m128i _t6 = _mm_loadl_epi64((const __m128i*)r6);
-                        __m128i _t7 = _mm_loadl_epi64((const __m128i*)r7);
-
-                        __m128i _t01 = _mm_unpacklo_epi8(_t0, _t1);
-                        __m128i _t23 = _mm_unpacklo_epi8(_t2, _t3);
-                        __m128i _t45 = _mm_unpacklo_epi8(_t4, _t5);
-                        __m128i _t67 = _mm_unpacklo_epi8(_t6, _t7);
-                        _t0 = _mm_unpacklo_epi16(_t01, _t23);
-                        _t1 = _mm_unpacklo_epi16(_t45, _t67);
-                        _t2 = _mm_unpacklo_epi32(_t0, _t1);
-                        _t3 = _mm_unpackhi_epi32(_t0, _t1);
-
-                        __m128i _extt2 = _mm_cmpgt_epi8(_mm_setzero_si128(), _t2);
-                        __m128i _extt3 = _mm_cmpgt_epi8(_mm_setzero_si128(), _t3);
-
-                        _r0 = _mm_unpacklo_epi8(_t2, _extt2);
-                        if (tj * 2 + 1 < w) _r1 = _mm_unpackhi_epi8(_t2, _extt2);
-                        if (tj * 2 + 2 < w) _r2 = _mm_unpacklo_epi8(_t3, _extt3);
-                        if (tj * 2 + 3 < w) _r3 = _mm_unpackhi_epi8(_t3, _extt3);
-#endif // __AVX2__
-                    }
-                }
-
-                __m128i _tmp0 = _mm_sub_epi16(_r0, _r2);
-                __m128i _tmp1 = _mm_add_epi16(_r1, _r2);
-                __m128i _tmp2 = _mm_sub_epi16(_r2, _r1);
-                __m128i _tmp3 = _mm_sub_epi16(_r3, _r1);
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                // old gcc breaks stack variable alignement
-                // ref https://gcc.gnu.org/bugzilla/show_bug.cgi?id=16660
-                _mm_storeu_si128((__m128i*)tmp[0][m], _tmp0);
-                _mm_storeu_si128((__m128i*)tmp[1][m], _tmp1);
-                _mm_storeu_si128((__m128i*)tmp[2][m], _tmp2);
-                _mm_storeu_si128((__m128i*)tmp[3][m], _tmp3);
-#else
-                _mm_store_si128((__m128i*)tmp[0][m], _tmp0);
-                _mm_store_si128((__m128i*)tmp[1][m], _tmp1);
-                _mm_store_si128((__m128i*)tmp[2][m], _tmp2);
-                _mm_store_si128((__m128i*)tmp[3][m], _tmp3);
-#endif
-
-                r0 += w * elempack;
-            }
-
-            short* p0 = (short*)B + kk * max_jj * 16 + jj * 8;
-            short* p1 = p0 + max_jj * 8;
-            short* p2 = p0 + max_jj * 8 * 2;
-            short* p3 = p0 + max_jj * 8 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m128i _r0 = _mm_loadu_si128((const __m128i*)tmp[m][0]);
-                __m128i _r1 = _mm_loadu_si128((const __m128i*)tmp[m][1]);
-                __m128i _r2 = _mm_loadu_si128((const __m128i*)tmp[m][2]);
-                __m128i _r3 = _mm_loadu_si128((const __m128i*)tmp[m][3]);
-#else
-                __m128i _r0 = _mm_load_si128((const __m128i*)tmp[m][0]);
-                __m128i _r1 = _mm_load_si128((const __m128i*)tmp[m][1]);
-                __m128i _r2 = _mm_load_si128((const __m128i*)tmp[m][2]);
-                __m128i _r3 = _mm_load_si128((const __m128i*)tmp[m][3]);
-#endif
-
-                __m128i _tmp0 = _mm_sub_epi16(_r0, _r2);
-                __m128i _tmp1 = _mm_add_epi16(_r1, _r2);
-                __m128i _tmp2 = _mm_sub_epi16(_r2, _r1);
-                __m128i _tmp3 = _mm_sub_epi16(_r3, _r1);
-
-                _mm_store_si128((__m128i*)p0, _tmp0);
-                _mm_store_si128((__m128i*)p1, _tmp1);
-                _mm_store_si128((__m128i*)p2, _tmp2);
-                _mm_store_si128((__m128i*)p3, _tmp3);
-
-                p0 += max_jj * 4 * 8;
-                p1 += max_jj * 4 * 8;
-                p2 += max_jj * 4 * 8;
-                p3 += max_jj * 4 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-#else // __SSE2__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    #pragma omp parallel for num_threads(nT)
-#endif // __SSE2__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-        short tmp[4][4][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const signed char* r0 = bottom_blob.channel(k + kk).row<const signed char>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 4; m++)
-            {
-                signed char r00 = 0;
-                signed char r01 = 0;
-                signed char r10 = 0;
-                signed char r11 = 0;
-                signed char r20 = 0;
-                signed char r21 = 0;
-                signed char r30 = 0;
-                signed char r31 = 0;
-
-                if (ti * 2 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const signed char* r1 = r0 + N;
-
-                        r00 = r0[0];
-                        r01 = r1[0];
-                        if (tj * 2 + 1 < w)
-                        {
-                            r10 = r0[1];
-                            r11 = r1[1];
-                        }
-                        if (tj * 2 + 2 < w)
-                        {
-                            r20 = r0[2];
-                            r21 = r1[2];
-                        }
-                        if (tj * 2 + 3 < w)
-                        {
-                            r30 = r0[3];
-                            r31 = r1[3];
-                        }
-                    }
-                }
-
-                tmp[0][m][0] = r00 - r20;
-                tmp[0][m][1] = r01 - r21;
-                tmp[1][m][0] = r10 + r20;
-                tmp[1][m][1] = r11 + r21;
-                tmp[2][m][0] = r20 - r10;
-                tmp[2][m][1] = r21 - r11;
-                tmp[3][m][0] = r30 - r10;
-                tmp[3][m][1] = r31 - r11;
-
-                r0 += w;
-            }
-
-            short* p0 = (short*)B + kk * max_jj * 16 + jj * 2;
-            short* p1 = p0 + max_jj * 2;
-            short* p2 = p0 + max_jj * 2 * 2;
-            short* p3 = p0 + max_jj * 2 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                short r00 = tmp[m][0][0];
-                short r01 = tmp[m][0][1];
-                short r10 = tmp[m][1][0];
-                short r11 = tmp[m][1][1];
-                short r20 = tmp[m][2][0];
-                short r21 = tmp[m][2][1];
-                short r30 = tmp[m][3][0];
-                short r31 = tmp[m][3][1];
-
-                p0[0] = r00 - r20;
-                p0[1] = r01 - r21;
-                p1[0] = r10 + r20;
-                p1[1] = r11 + r21;
-                p2[0] = r20 - r10;
-                p2[1] = r21 - r11;
-                p3[0] = r30 - r10;
-                p3[1] = r31 - r11;
-
-                p0 += max_jj * 4 * 2;
-                p1 += max_jj * 4 * 2;
-                p2 += max_jj * 4 * 2;
-                p3 += max_jj * 4 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        short tmp[4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const signed char* r0123 = bottom_blob.channel(k + kk).row<const signed char>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 4; m++)
-            {
-                signed char r0 = 0;
-                signed char r1 = 0;
-                signed char r2 = 0;
-                signed char r3 = 0;
-
-                if (ti * 2 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = r0123[0];
-                        if (tj * 2 + 1 < w) r1 = r0123[1];
-                        if (tj * 2 + 2 < w) r2 = r0123[2];
-                        if (tj * 2 + 3 < w) r3 = r0123[3];
-                    }
-                }
-
-                tmp[0][m] = r0 - r2;
-                tmp[1][m] = r1 + r2;
-                tmp[2][m] = r2 - r1;
-                tmp[3][m] = r3 - r1;
-
-                r0123 += w;
-            }
-
-            short* p0 = (short*)B + kk * max_jj * 16 + jj;
-            short* p1 = p0 + max_jj;
-            short* p2 = p0 + max_jj * 2;
-            short* p3 = p0 + max_jj * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                short r0 = tmp[m][0];
-                short r1 = tmp[m][1];
-                short r2 = tmp[m][2];
-                short r3 = tmp[m][3];
-
-                p0[0] = r0 - r2;
-                p1[0] = r1 + r2;
-                p2[0] = r2 - r1;
-                p3[0] = r3 - r1;
-
-                p0 += max_jj * 4;
-                p1 += max_jj * 4;
-                p2 += max_jj * 4;
-                p3 += max_jj * 4;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd23_transform_output_tile_int8(const Mat& top_tile, Mat& top_blob, int i, int max_ii, int j, int max_jj)
-{
-    // const int otm[2][4] = {
-    //     {1,  1,  1,  0},
-    //     {0,  1, -1,  1}
-    // };
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 1) / 2;
-
-    int ii = 0;
-#if __SSE2__
-#if __AVX2__
-#if __AVX512F__
-    for (; ii + 15 < max_ii; ii += 16)
-    {
-#ifdef _MSC_VER
-        __declspec(align(64))
-#else
-        __attribute__((aligned(64)))
-#endif
-        int tmp[2][4][16];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 16 + jj * 16;
-            const int* r1 = r0 + max_jj * 16;
-            const int* r2 = r0 + max_jj * 16 * 2;
-            const int* r3 = r0 + max_jj * 16 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                __m512i _r0 = _mm512_load_si512((const __m512i*)r0);
-                __m512i _r1 = _mm512_load_si512((const __m512i*)r1);
-                __m512i _r2 = _mm512_load_si512((const __m512i*)r2);
-                __m512i _r3 = _mm512_load_si512((const __m512i*)r3);
-
-                __m512i _tmp0 = _mm512_add_epi32(_mm512_add_epi32(_r0, _r1), _r2);
-                __m512i _tmp1 = _mm512_add_epi32(_mm512_sub_epi32(_r1, _r2), _r3);
-
-                _mm512_store_si512((__m512i*)tmp[0][m], _tmp0);
-                _mm512_store_si512((__m512i*)tmp[1][m], _tmp1);
-
-                r0 += max_jj * 4 * 16;
-                r1 += max_jj * 4 * 16;
-                r2 += max_jj * 4 * 16;
-                r3 += max_jj * 4 * 16;
-            }
-
-            int* outptr0 = top_blob.channel((i + ii) / out_elempack).row<int>(ti * 2) + (tj * 2) * out_elempack;
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                __m512i _r0 = _mm512_load_si512((const __m512i*)tmp[m][0]);
-                __m512i _r1 = _mm512_load_si512((const __m512i*)tmp[m][1]);
-                __m512i _r2 = _mm512_load_si512((const __m512i*)tmp[m][2]);
-                __m512i _r3 = _mm512_load_si512((const __m512i*)tmp[m][3]);
-
-                __m512i _tmp0 = _mm512_add_epi32(_mm512_add_epi32(_r0, _r1), _r2);
-                __m512i _tmp1 = _mm512_add_epi32(_mm512_sub_epi32(_r1, _r2), _r3);
-
-                _tmp0 = _mm512_srai_epi32(_tmp0, 2);
-                _tmp1 = _mm512_srai_epi32(_tmp1, 2);
-
-                if (out_elempack == 16)
-                {
-                    _mm512_store_si512((__m512i*)outptr0, _tmp0);
-                    if (tj * 2 + 1 < outw)
-                    {
-                        _mm512_store_si512((__m512i*)(outptr0 + 16), _tmp1);
-                    }
-                }
-                if (out_elempack == 8)
-                {
-                    int* outptr1 = outptr0 + N;
-
-                    _mm256_store_si256((__m256i*)outptr0, _mm512_extracti32x8_epi32(_tmp0, 0));
-                    _mm256_store_si256((__m256i*)outptr1, _mm512_extracti32x8_epi32(_tmp0, 1));
-                    if (tj * 2 + 1 < outw)
-                    {
-                        _mm256_store_si256((__m256i*)(outptr0 + 8), _mm512_extracti32x8_epi32(_tmp1, 0));
-                        _mm256_store_si256((__m256i*)(outptr1 + 8), _mm512_extracti32x8_epi32(_tmp1, 1));
-                    }
-                }
-                if (out_elempack == 4)
-                {
-                    int* outptr1 = outptr0 + N;
-                    int* outptr2 = outptr0 + N * 2;
-                    int* outptr3 = outptr0 + N * 3;
-
-                    _mm_store_si128((__m128i*)outptr0, _mm512_extracti32x4_epi32(_tmp0, 0));
-                    _mm_store_si128((__m128i*)outptr1, _mm512_extracti32x4_epi32(_tmp0, 1));
-                    _mm_store_si128((__m128i*)outptr2, _mm512_extracti32x4_epi32(_tmp0, 2));
-                    _mm_store_si128((__m128i*)outptr3, _mm512_extracti32x4_epi32(_tmp0, 3));
-                    if (tj * 2 + 1 < outw)
-                    {
-                        _mm_store_si128((__m128i*)(outptr0 + 4), _mm512_extracti32x4_epi32(_tmp1, 0));
-                        _mm_store_si128((__m128i*)(outptr1 + 4), _mm512_extracti32x4_epi32(_tmp1, 1));
-                        _mm_store_si128((__m128i*)(outptr2 + 4), _mm512_extracti32x4_epi32(_tmp1, 2));
-                        _mm_store_si128((__m128i*)(outptr3 + 4), _mm512_extracti32x4_epi32(_tmp1, 3));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                    _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                    _mm512_i32scatter_epi32(outptr0, _vindex, _tmp0, sizeof(int));
-                    if (tj * 2 + 1 < outw) _mm512_i32scatter_epi32(outptr0 + 1, _vindex, _tmp1, sizeof(int));
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __AVX512F__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-#ifdef _MSC_VER
-        __declspec(align(32))
-#else
-        __attribute__((aligned(32)))
-#endif
-        int tmp[2][4][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 16 + jj * 8;
-            const int* r1 = r0 + max_jj * 8;
-            const int* r2 = r0 + max_jj * 8 * 2;
-            const int* r3 = r0 + max_jj * 8 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                __m256i _r0 = _mm256_load_si256((const __m256i*)r0);
-                __m256i _r1 = _mm256_load_si256((const __m256i*)r1);
-                __m256i _r2 = _mm256_load_si256((const __m256i*)r2);
-                __m256i _r3 = _mm256_load_si256((const __m256i*)r3);
-
-                __m256i _tmp0 = _mm256_add_epi32(_mm256_add_epi32(_r0, _r1), _r2);
-                __m256i _tmp1 = _mm256_add_epi32(_mm256_sub_epi32(_r1, _r2), _r3);
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm256_storeu_si256((__m256i*)tmp[0][m], _tmp0);
-                _mm256_storeu_si256((__m256i*)tmp[1][m], _tmp1);
-#else
-                _mm256_store_si256((__m256i*)tmp[0][m], _tmp0);
-                _mm256_store_si256((__m256i*)tmp[1][m], _tmp1);
-#endif
-
-                r0 += max_jj * 4 * 8;
-                r1 += max_jj * 4 * 8;
-                r2 += max_jj * 4 * 8;
-                r3 += max_jj * 4 * 8;
-            }
-
-            int* outptr0 = top_blob.channel((i + ii) / out_elempack).row<int>(ti * 2) + (tj * 2) * out_elempack;
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m256i _r0 = _mm256_loadu_si256((const __m256i*)tmp[m][0]);
-                __m256i _r1 = _mm256_loadu_si256((const __m256i*)tmp[m][1]);
-                __m256i _r2 = _mm256_loadu_si256((const __m256i*)tmp[m][2]);
-                __m256i _r3 = _mm256_loadu_si256((const __m256i*)tmp[m][3]);
-#else
-                __m256i _r0 = _mm256_load_si256((const __m256i*)tmp[m][0]);
-                __m256i _r1 = _mm256_load_si256((const __m256i*)tmp[m][1]);
-                __m256i _r2 = _mm256_load_si256((const __m256i*)tmp[m][2]);
-                __m256i _r3 = _mm256_load_si256((const __m256i*)tmp[m][3]);
-#endif
-
-                __m256i _tmp0 = _mm256_add_epi32(_mm256_add_epi32(_r0, _r1), _r2);
-                __m256i _tmp1 = _mm256_add_epi32(_mm256_sub_epi32(_r1, _r2), _r3);
-
-                _tmp0 = _mm256_srai_epi32(_tmp0, 2);
-                _tmp1 = _mm256_srai_epi32(_tmp1, 2);
-
-                if (out_elempack == 8)
-                {
-                    _mm256_store_si256((__m256i*)outptr0, _tmp0);
-                    if (tj * 2 + 1 < outw)
-                    {
-                        _mm256_store_si256((__m256i*)(outptr0 + 8), _tmp1);
-                    }
-                }
-                if (out_elempack == 4)
-                {
-                    int* outptr1 = outptr0 + N;
-
-                    _mm_store_si128((__m128i*)outptr0, _mm256_extracti128_si256(_tmp0, 0));
-                    _mm_store_si128((__m128i*)outptr1, _mm256_extracti128_si256(_tmp0, 1));
-                    if (tj * 2 + 1 < outw)
-                    {
-                        _mm_store_si128((__m128i*)(outptr0 + 4), _mm256_extracti128_si256(_tmp1, 0));
-                        _mm_store_si128((__m128i*)(outptr1 + 4), _mm256_extracti128_si256(_tmp1, 1));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-#if __AVX512F__
-                    __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                    _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-                    _mm256_i32scatter_epi32(outptr0, _vindex, _tmp0, sizeof(int));
-                    if (tj * 2 + 1 < outw) _mm256_i32scatter_epi32(outptr0 + 1, _vindex, _tmp1, sizeof(int));
-#else
-                    int tmp0[8];
-                    int tmp1[8];
-                    _mm256_storeu_si256((__m256i*)tmp0, _tmp0);
-                    _mm256_storeu_si256((__m256i*)tmp1, _tmp1);
-
-                    int* outptr1 = outptr0 + N;
-                    int* outptr2 = outptr0 + N * 2;
-                    int* outptr3 = outptr0 + N * 3;
-                    int* outptr4 = outptr0 + N * 4;
-                    int* outptr5 = outptr0 + N * 5;
-                    int* outptr6 = outptr0 + N * 6;
-                    int* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                    }
-#endif // __AVX512F__
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __AVX2__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        int tmp[2][4][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 16 + jj * 4;
-            const int* r1 = r0 + max_jj * 4;
-            const int* r2 = r0 + max_jj * 4 * 2;
-            const int* r3 = r0 + max_jj * 4 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                __m128i _r0 = _mm_load_si128((const __m128i*)r0);
-                __m128i _r1 = _mm_load_si128((const __m128i*)r1);
-                __m128i _r2 = _mm_load_si128((const __m128i*)r2);
-                __m128i _r3 = _mm_load_si128((const __m128i*)r3);
-
-                __m128i _tmp0 = _mm_add_epi32(_mm_add_epi32(_r0, _r1), _r2);
-                __m128i _tmp1 = _mm_add_epi32(_mm_sub_epi32(_r1, _r2), _r3);
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm_storeu_si128((__m128i*)tmp[0][m], _tmp0);
-                _mm_storeu_si128((__m128i*)tmp[1][m], _tmp1);
-#else
-                _mm_store_si128((__m128i*)tmp[0][m], _tmp0);
-                _mm_store_si128((__m128i*)tmp[1][m], _tmp1);
-#endif
-
-                r0 += max_jj * 4 * 4;
-                r1 += max_jj * 4 * 4;
-                r2 += max_jj * 4 * 4;
-                r3 += max_jj * 4 * 4;
-            }
-
-            int* outptr0 = top_blob.channel((i + ii) / out_elempack).row<int>(ti * 2) + (tj * 2) * out_elempack;
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m128i _r0 = _mm_loadu_si128((const __m128i*)tmp[m][0]);
-                __m128i _r1 = _mm_loadu_si128((const __m128i*)tmp[m][1]);
-                __m128i _r2 = _mm_loadu_si128((const __m128i*)tmp[m][2]);
-                __m128i _r3 = _mm_loadu_si128((const __m128i*)tmp[m][3]);
-#else
-                __m128i _r0 = _mm_load_si128((const __m128i*)tmp[m][0]);
-                __m128i _r1 = _mm_load_si128((const __m128i*)tmp[m][1]);
-                __m128i _r2 = _mm_load_si128((const __m128i*)tmp[m][2]);
-                __m128i _r3 = _mm_load_si128((const __m128i*)tmp[m][3]);
-#endif
-
-                __m128i _tmp0 = _mm_add_epi32(_mm_add_epi32(_r0, _r1), _r2);
-                __m128i _tmp1 = _mm_add_epi32(_mm_sub_epi32(_r1, _r2), _r3);
-
-                _tmp0 = _mm_srai_epi32(_tmp0, 2);
-                _tmp1 = _mm_srai_epi32(_tmp1, 2);
-
-                if (out_elempack == 4)
-                {
-                    _mm_store_si128((__m128i*)outptr0, _tmp0);
-                    if (tj * 2 + 1 < outw) _mm_store_si128((__m128i*)(outptr0 + 4), _tmp1);
-                }
-                if (out_elempack == 1)
-                {
-#if __AVX512F__
-                    __m128i _vindex = _mm_setr_epi32(0, 1, 2, 3);
-                    _vindex = _mm_mullo_epi32(_vindex, _mm_set1_epi32(N));
-                    _mm_i32scatter_epi32(outptr0, _vindex, _tmp0, sizeof(int));
-                    if (tj * 2 + 1 < outw) _mm_i32scatter_epi32(outptr0 + 1, _vindex, _tmp1, sizeof(int));
-#else
-                    int tmp0[4];
-                    int tmp1[4];
-                    _mm_storeu_si128((__m128i*)tmp0, _tmp0);
-                    _mm_storeu_si128((__m128i*)tmp1, _tmp1);
-
-                    int* outptr1 = outptr0 + N;
-                    int* outptr2 = outptr0 + N * 2;
-                    int* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                    }
-#endif // __AVX512F__
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __SSE2__
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        int tmp[2][4][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 16 + jj * 2;
-            const int* r1 = r0 + max_jj * 2;
-            const int* r2 = r0 + max_jj * 2 * 2;
-            const int* r3 = r0 + max_jj * 2 * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                tmp[0][m][0] = r0[0] + r1[0] + r2[0];
-                tmp[0][m][1] = r0[1] + r1[1] + r2[1];
-                tmp[1][m][0] = r1[0] - r2[0] + r3[0];
-                tmp[1][m][1] = r1[1] - r2[1] + r3[1];
-
-                r0 += max_jj * 4 * 2;
-                r1 += max_jj * 4 * 2;
-                r2 += max_jj * 4 * 2;
-                r3 += max_jj * 4 * 2;
-            }
-
-            int* outptr0 = top_blob.channel(i + ii).row<int>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                int tmp00 = tmp[m][0][0] + tmp[m][1][0] + tmp[m][2][0];
-                int tmp01 = tmp[m][0][1] + tmp[m][1][1] + tmp[m][2][1];
-                int tmp10 = tmp[m][1][0] - tmp[m][2][0] + tmp[m][3][0];
-                int tmp11 = tmp[m][1][1] - tmp[m][2][1] + tmp[m][3][1];
-
-                tmp00 = tmp00 >> 2;
-                tmp01 = tmp01 >> 2;
-                tmp10 = tmp10 >> 2;
-                tmp11 = tmp11 >> 2;
-
-                // if (out_elempack == 1)
-                {
-                    int* outptr1 = outptr0 + N;
-
-                    outptr0[0] = tmp00;
-                    outptr1[0] = tmp01;
-                    if (tj * 2 + 1 < outw)
-                    {
-                        outptr0[1] = tmp10;
-                        outptr1[1] = tmp11;
-                    }
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        int tmp[2][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 16 + jj;
-            const int* r1 = r0 + max_jj;
-            const int* r2 = r0 + max_jj * 2;
-            const int* r3 = r0 + max_jj * 3;
-
-            for (int m = 0; m < 4; m++)
-            {
-                tmp[0][m] = r0[0] + r1[0] + r2[0];
-                tmp[1][m] = r1[0] - r2[0] + r3[0];
-
-                r0 += max_jj * 4;
-                r1 += max_jj * 4;
-                r2 += max_jj * 4;
-                r3 += max_jj * 4;
-            }
-
-            int* outptr0 = top_blob.channel(i + ii).row<int>(ti * 2) + (tj * 2);
-
-            for (int m = 0; m < 2; m++)
-            {
-                if (ti * 2 + m >= outh)
-                    continue;
-
-                int tmp0 = tmp[m][0] + tmp[m][1] + tmp[m][2];
-                int tmp1 = tmp[m][1] - tmp[m][2] + tmp[m][3];
-
-                tmp0 = tmp0 >> 2;
-                tmp1 = tmp1 >> 2;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = tmp0;
-                    if (tj * 2 + 1 < outw) outptr0[1] = tmp1;
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd23_int8(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt)
-{
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
-#if NCNN_RUNTIME_CPU && NCNN_AVX512VNNI && __AVX512F__ && !__AVX512VNNI__
-    if (ncnn::cpu_support_x86_avx512_vnni())
-    {
-        conv3x3s1_winograd23_int8_avx512vnni(bottom_blob, top_blob, AT, nT, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__
-    if (ncnn::cpu_support_x86_avx_vnni())
-    {
-        conv3x3s1_winograd23_int8_avxvnni(bottom_blob, top_blob, AT, nT, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-    if (ncnn::cpu_support_x86_avx2())
-    {
-        conv3x3s1_winograd23_int8_avx2(bottom_blob, top_blob, AT, nT, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__
-    if (ncnn::cpu_support_x86_xop())
-    {
-        conv3x3s1_winograd23_int8_xop(bottom_blob, top_blob, AT, nT, opt);
-        return;
-    }
-#endif
-#endif
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 2n+2, winograd F(2,3)
-    int w_tiles = (outw + 1) / 2;
-    int h_tiles = (outh + 1) / 2;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 16;
-
-    // NCNN_LOGE("conv3x3s1_winograd23_int8 %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    get_optimal_tile_mnk_int8(M, N, K, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 2u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 2u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd23_transform_input_tile_int8(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            transpose_pack_B_tile_int8(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 2u, opt.workspace_allocator);
-
-        #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd23_transform_input_tile_int8(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            transpose_pack_B_tile_int8(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                bool k_end = k + TILE_K >= K;
-
-                gemm_transB_packed_tile_int8(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk, k_end);
-            }
-
-            // transform output
-            conv3x3s1_winograd23_transform_output_tile_int8(top_tile, top_blob, i, max_ii, j, max_jj);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_kernel_tile_int8(const Mat& kernel, Mat& A, int inch, int i, int max_ii, int k, int max_kk)
-{
-    // const short ktm[6][3] = {
-    //     {6, 0, 0},
-    //     {-4, -4, -4},
-    //     {-4, 4, -4},
-    //     {1, 2, 4},
-    //     {1, -2, 4},
-    //     {0, 0, 6}
-    // };
-
-    short* ptmp = A;
-
-    int ii = 0;
-    for (; ii < max_ii; ii++)
-    {
-        int kk = 0;
-        for (; kk < max_kk; kk++)
-        {
-            short tmp[6][3];
-
-            const signed char* k0 = (const signed char*)kernel + (i + ii) * inch * 9 + (k + kk) * 9;
-
-            for (int m = 0; m < 3; m++)
-            {
-                signed char r0 = k0[0];
-                signed char r1 = k0[1];
-                signed char r2 = k0[2];
-
-                tmp[0][m] = r0 * 6;
-                tmp[1][m] = -r0 * 4 - r1 * 4 - r2 * 4;
-                tmp[2][m] = -r0 * 4 + r1 * 4 - r2 * 4;
-                tmp[3][m] = r0 + r1 * 2 + r2 * 4;
-                tmp[4][m] = r0 - r1 * 2 + r2 * 4;
-                tmp[5][m] = r2 * 6;
-
-                k0 += 3;
-            }
-
-            for (int m = 0; m < 6; m++)
-            {
-                short r0 = tmp[m][0];
-                short r1 = tmp[m][1];
-                short r2 = tmp[m][2];
-
-                short z0 = r0 * 6;
-                short z1 = -r0 * 4 - r1 * 4 - r2 * 4;
-                short z2 = -r0 * 4 + r1 * 4 - r2 * 4;
-                short z3 = r0 + r1 * 2 + r2 * 4;
-                short z4 = r0 - r1 * 2 + r2 * 4;
-                short z5 = r2 * 6;
-
-                ptmp[0] = z0;
-                ptmp[1] = z1;
-                ptmp[2] = z2;
-                ptmp[3] = z3;
-                ptmp[4] = z4;
-                ptmp[5] = z5;
-                ptmp += 6;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_transform_kernel_int8(const Mat& kernel, Mat& AT, int inch, int outch, const Option& opt)
-{
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-    if (ncnn::cpu_support_x86_avx2())
-    {
-        conv3x3s1_winograd43_transform_kernel_int8_avx2(kernel, AT, inch, outch, opt);
-        return;
-    }
-#endif
-#endif
-
-    const int M = outch;
-    const int K = inch;
-    const int B = 36;
-
-    int TILE_M, TILE_N, TILE_K;
-    get_optimal_tile_mnk_int8(M, 0, K, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    Mat A_tileX(B * TILE_M * TILE_K, 1, opt.num_threads, 4u, (Allocator*)0);
-
-    AT.create(TILE_K * TILE_M, B, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M, 4u, (Allocator*)0);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat A_tile = A_tileX.channel(get_omp_thread_num());
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_ii = std::min((M - i), TILE_M);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            conv3x3s1_winograd43_transform_kernel_tile_int8(kernel, A_tile, inch, i, max_ii, k, max_kk);
-
-            Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-            pack_A_tile_int8(A_tile, AT_tile, B, max_ii, max_kk);
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_input_tile_int8(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int nT)
-{
-    // const float itm[4][4] = {
-    //     {4,  0, -5,  0, 1, 0},
-    //     {0, -4, -4,  1, 1, 0},
-    //     {0,  4, -4, -1, 1, 0},
-    //     {0, -2, -1,  2, 1, 0},
-    //     {0,  2, -1, -2, 1, 0},
-    //     {0,  4,  0, -5, 0, 1}
-    // };
-
-    const int w = bottom_blob.w;
-    const int h = bottom_blob.h;
-    const int elempack = bottom_blob.elempack;
-    const int N = bottom_blob.cstep * elempack;
-
-    const int w_tiles = (w + 1) / 4;
-
-    int nn_max_kk = 0;
-    int remain_max_kk_start = 0;
-#if __SSE2__
-#if __AVX512F__
-    nn_max_kk = max_kk / 16;
-    #pragma omp parallel for num_threads(nT)
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = ppkk * 16;
-
-#ifdef _MSC_VER
-        __declspec(align(64))
-#else
-        __attribute__((aligned(64)))
-#endif
-        short tmp[6][6][16];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const signed char* r0 = bottom_blob.channel((k + kk) / elempack).row<const signed char>(ti * 4) + (tj * 4) * elempack;
-
-            __m256i _v2 = _mm256_set1_epi16(2);
-            __m256i _v4 = _mm256_set1_epi16(4);
-            __m256i _v5 = _mm256_set1_epi16(5);
-
-            for (int m = 0; m < 6; m++)
-            {
-                __m256i _r0 = _mm256_setzero_si256();
-                __m256i _r1 = _mm256_setzero_si256();
-                __m256i _r2 = _mm256_setzero_si256();
-                __m256i _r3 = _mm256_setzero_si256();
-                __m256i _r4 = _mm256_setzero_si256();
-                __m256i _r5 = _mm256_setzero_si256();
-
-                if (ti * 4 + m < h)
-                {
-                    if (elempack == 16)
-                    {
-                        _r0 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)r0));
-                        if (tj * 4 + 1 < w) _r1 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)(r0 + 16)));
-                        if (tj * 4 + 2 < w) _r2 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)(r0 + 32)));
-                        if (tj * 4 + 3 < w) _r3 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)(r0 + 48)));
-                        if (tj * 4 + 4 < w) _r4 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)(r0 + 64)));
-                        if (tj * 4 + 5 < w) _r5 = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)(r0 + 80)));
-                    }
-                    if (elempack == 8)
-                    {
-                        const signed char* r1 = r0 + N;
-
-                        _r0 = _mm256_cvtepi8_epi16(_mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i*)r0), _mm_loadl_epi64((const __m128i*)r1)));
-                        if (tj * 4 + 1 < w) _r1 = _mm256_cvtepi8_epi16(_mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i*)(r0 + 8)), _mm_loadl_epi64((const __m128i*)(r1 + 8))));
-                        if (tj * 4 + 2 < w) _r2 = _mm256_cvtepi8_epi16(_mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i*)(r0 + 16)), _mm_loadl_epi64((const __m128i*)(r1 + 16))));
-                        if (tj * 4 + 3 < w) _r3 = _mm256_cvtepi8_epi16(_mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i*)(r0 + 24)), _mm_loadl_epi64((const __m128i*)(r1 + 24))));
-                        if (tj * 4 + 4 < w) _r4 = _mm256_cvtepi8_epi16(_mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i*)(r0 + 32)), _mm_loadl_epi64((const __m128i*)(r1 + 32))));
-                        if (tj * 4 + 5 < w) _r5 = _mm256_cvtepi8_epi16(_mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i*)(r0 + 40)), _mm_loadl_epi64((const __m128i*)(r1 + 40))));
-                    }
-                    if (elempack == 1)
-                    {
-                        __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                        _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                        _r0 = _mm256_cvtepi8_epi16(_mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)r0, 1)));
-                        if (tj * 4 + 1 < w) _r1 = _mm256_cvtepi8_epi16(_mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0 + 1), 1)));
-                        if (tj * 4 + 2 < w) _r2 = _mm256_cvtepi8_epi16(_mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0 + 2), 1)));
-                        if (tj * 4 + 3 < w) _r3 = _mm256_cvtepi8_epi16(_mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0 + 3), 1)));
-                        if (tj * 4 + 4 < w) _r4 = _mm256_cvtepi8_epi16(_mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0 + 4), 1)));
-                        if (tj * 4 + 5 < w) _r5 = _mm256_cvtepi8_epi16(_mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0 + 5), 1)));
-                    }
-                }
-
-                __m256i _tmp12a = _mm256_sub_epi16(_r3, _mm256_mullo_epi16(_r1, _v4));
-                __m256i _tmp12b = _mm256_sub_epi16(_r4, _mm256_mullo_epi16(_r2, _v4));
-                __m256i _tmp34a = _mm256_mullo_epi16(_mm256_sub_epi16(_r3, _r1), _v2);
-                __m256i _tmp34b = _mm256_sub_epi16(_r4, _r2);
-
-                __m256i _tmp0 = _mm256_add_epi16(_r4, _mm256_sub_epi16(_mm256_mullo_epi16(_r0, _v4), _mm256_mullo_epi16(_r2, _v5)));
-                __m256i _tmp1 = _mm256_add_epi16(_tmp12b, _tmp12a);
-                __m256i _tmp2 = _mm256_sub_epi16(_tmp12b, _tmp12a);
-                __m256i _tmp3 = _mm256_add_epi16(_tmp34b, _tmp34a);
-                __m256i _tmp4 = _mm256_sub_epi16(_tmp34b, _tmp34a);
-                __m256i _tmp5 = _mm256_add_epi16(_r5, _mm256_sub_epi16(_mm256_mullo_epi16(_r1, _v4), _mm256_mullo_epi16(_r3, _v5)));
-
-                _mm256_store_si256((__m256i*)tmp[0][m], _tmp0);
-                _mm256_store_si256((__m256i*)tmp[1][m], _tmp1);
-                _mm256_store_si256((__m256i*)tmp[2][m], _tmp2);
-                _mm256_store_si256((__m256i*)tmp[3][m], _tmp3);
-                _mm256_store_si256((__m256i*)tmp[4][m], _tmp4);
-                _mm256_store_si256((__m256i*)tmp[5][m], _tmp5);
-
-                r0 += w * elempack;
-            }
-
-            short* p0 = (short*)B + kk * max_jj * 36 + jj * 16;
-            short* p1 = p0 + max_jj * 16;
-            short* p2 = p0 + max_jj * 16 * 2;
-            short* p3 = p0 + max_jj * 16 * 3;
-            short* p4 = p0 + max_jj * 16 * 4;
-            short* p5 = p0 + max_jj * 16 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                __m256i _r0 = _mm256_load_si256((const __m256i*)tmp[m][0]);
-                __m256i _r1 = _mm256_load_si256((const __m256i*)tmp[m][1]);
-                __m256i _r2 = _mm256_load_si256((const __m256i*)tmp[m][2]);
-                __m256i _r3 = _mm256_load_si256((const __m256i*)tmp[m][3]);
-                __m256i _r4 = _mm256_load_si256((const __m256i*)tmp[m][4]);
-                __m256i _r5 = _mm256_load_si256((const __m256i*)tmp[m][5]);
-
-                __m256i _tmp12a = _mm256_sub_epi16(_r3, _mm256_mullo_epi16(_r1, _v4));
-                __m256i _tmp12b = _mm256_sub_epi16(_r4, _mm256_mullo_epi16(_r2, _v4));
-                __m256i _tmp34a = _mm256_mullo_epi16(_mm256_sub_epi16(_r3, _r1), _v2);
-                __m256i _tmp34b = _mm256_sub_epi16(_r4, _r2);
-
-                __m256i _tmp0 = _mm256_add_epi16(_r4, _mm256_sub_epi16(_mm256_mullo_epi16(_r0, _v4), _mm256_mullo_epi16(_r2, _v5)));
-                __m256i _tmp1 = _mm256_add_epi16(_tmp12b, _tmp12a);
-                __m256i _tmp2 = _mm256_sub_epi16(_tmp12b, _tmp12a);
-                __m256i _tmp3 = _mm256_add_epi16(_tmp34b, _tmp34a);
-                __m256i _tmp4 = _mm256_sub_epi16(_tmp34b, _tmp34a);
-                __m256i _tmp5 = _mm256_add_epi16(_r5, _mm256_sub_epi16(_mm256_mullo_epi16(_r1, _v4), _mm256_mullo_epi16(_r3, _v5)));
-
-                _mm256_store_si256((__m256i*)p0, _tmp0);
-                _mm256_store_si256((__m256i*)p1, _tmp1);
-                _mm256_store_si256((__m256i*)p2, _tmp2);
-                _mm256_store_si256((__m256i*)p3, _tmp3);
-                _mm256_store_si256((__m256i*)p4, _tmp4);
-                _mm256_store_si256((__m256i*)p5, _tmp5);
-
-                p0 += max_jj * 6 * 16;
-                p1 += max_jj * 6 * 16;
-                p2 += max_jj * 6 * 16;
-                p3 += max_jj * 6 * 16;
-                p4 += max_jj * 6 * 16;
-                p5 += max_jj * 6 * 16;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 16;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-#else // __AVX512F__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 8;
-    #pragma omp parallel for num_threads(nT)
-#endif // __AVX512F__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 8;
-
-#ifdef _MSC_VER
-        __declspec(align(32))
-#else
-        __attribute__((aligned(32)))
-#endif
-        short tmp[6][6][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const signed char* r0 = bottom_blob.channel((k + kk) / elempack).row<const signed char>(ti * 4) + (tj * 4) * elempack;
-
-            __m128i _v2 = _mm_set1_epi16(2);
-            __m128i _v4 = _mm_set1_epi16(4);
-            __m128i _v5 = _mm_set1_epi16(5);
-
-            for (int m = 0; m < 6; m++)
-            {
-                __m128i _r0 = _mm_setzero_si128();
-                __m128i _r1 = _mm_setzero_si128();
-                __m128i _r2 = _mm_setzero_si128();
-                __m128i _r3 = _mm_setzero_si128();
-                __m128i _r4 = _mm_setzero_si128();
-                __m128i _r5 = _mm_setzero_si128();
-
-                if (ti * 4 + m < h)
-                {
-                    if (elempack == 8)
-                    {
-                        _r0 = _mm_loadl_epi64((const __m128i*)r0);
-                        _r0 = _mm_unpacklo_epi8(_r0, _mm_cmpgt_epi8(_mm_setzero_si128(), _r0));
-                        if (tj * 4 + 1 < w)
-                        {
-                            _r1 = _mm_loadl_epi64((const __m128i*)(r0 + 8));
-                            _r1 = _mm_unpacklo_epi8(_r1, _mm_cmpgt_epi8(_mm_setzero_si128(), _r1));
-                        }
-                        if (tj * 4 + 2 < w)
-                        {
-                            _r2 = _mm_loadl_epi64((const __m128i*)(r0 + 16));
-                            _r2 = _mm_unpacklo_epi8(_r2, _mm_cmpgt_epi8(_mm_setzero_si128(), _r2));
-                        }
-                        if (tj * 4 + 3 < w)
-                        {
-                            _r3 = _mm_loadl_epi64((const __m128i*)(r0 + 24));
-                            _r3 = _mm_unpacklo_epi8(_r3, _mm_cmpgt_epi8(_mm_setzero_si128(), _r3));
-                        }
-                        if (tj * 4 + 4 < w)
-                        {
-                            _r4 = _mm_loadl_epi64((const __m128i*)(r0 + 32));
-                            _r4 = _mm_unpacklo_epi8(_r4, _mm_cmpgt_epi8(_mm_setzero_si128(), _r4));
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            _r5 = _mm_loadl_epi64((const __m128i*)(r0 + 40));
-                            _r5 = _mm_unpacklo_epi8(_r5, _mm_cmpgt_epi8(_mm_setzero_si128(), _r5));
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-#if __AVX2__
-                        __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                        _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-#if __AVX512F__
-                        _r0 = _mm_cvtepi8_epi16(_mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0, _vindex, 1)));
-                        if (tj * 4 + 1 < w) _r1 = _mm_cvtepi8_epi16(_mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(r0 + 1), _vindex, 1)));
-                        if (tj * 4 + 2 < w) _r2 = _mm_cvtepi8_epi16(_mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(r0 + 2), _vindex, 1)));
-                        if (tj * 4 + 3 < w) _r3 = _mm_cvtepi8_epi16(_mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(r0 + 3), _vindex, 1)));
-                        if (tj * 4 + 4 < w) _r4 = _mm_cvtepi8_epi16(_mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(r0 + 4), _vindex, 1)));
-                        if (tj * 4 + 5 < w) _r5 = _mm_cvtepi8_epi16(_mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(r0 + 5), _vindex, 1)));
-#else
-                        __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-                        __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-                        __m256i _val0_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r0, _vindex, 1), _sindex88);
-                        _r0 = _mm_cvtepi8_epi16(_mm_unpacklo_epi32(_mm256_extracti128_si256(_val0_32, 0), _mm256_extracti128_si256(_val0_32, 1)));
-                        if (tj * 4 + 1 < w)
-                        {
-                            __m256i _val1_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)(r0 + 1), _vindex, 1), _sindex88);
-                            _r1 = _mm_cvtepi8_epi16(_mm_unpacklo_epi32(_mm256_extracti128_si256(_val1_32, 0), _mm256_extracti128_si256(_val1_32, 1)));
-                        }
-                        if (tj * 4 + 2 < w)
-                        {
-                            __m256i _val2_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)(r0 + 2), _vindex, 1), _sindex88);
-                            _r2 = _mm_cvtepi8_epi16(_mm_unpacklo_epi32(_mm256_extracti128_si256(_val2_32, 0), _mm256_extracti128_si256(_val2_32, 1)));
-                        }
-                        if (tj * 4 + 3 < w)
-                        {
-                            __m256i _val3_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)(r0 + 3), _vindex, 1), _sindex88);
-                            _r3 = _mm_cvtepi8_epi16(_mm_unpacklo_epi32(_mm256_extracti128_si256(_val3_32, 0), _mm256_extracti128_si256(_val3_32, 1)));
-                        }
-                        if (tj * 4 + 4 < w)
-                        {
-                            __m256i _val4_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)(r0 + 4), _vindex, 1), _sindex88);
-                            _r4 = _mm_cvtepi8_epi16(_mm_unpacklo_epi32(_mm256_extracti128_si256(_val4_32, 0), _mm256_extracti128_si256(_val4_32, 1)));
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            __m256i _val5_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)(r0 + 5), _vindex, 1), _sindex88);
-                            _r5 = _mm_cvtepi8_epi16(_mm_unpacklo_epi32(_mm256_extracti128_si256(_val5_32, 0), _mm256_extracti128_si256(_val5_32, 1)));
-                        }
-#endif // __AVX512F__
-#else  // __AVX2__
-                        const signed char* r1 = r0 + N;
-                        const signed char* r2 = r0 + N * 2;
-                        const signed char* r3 = r0 + N * 3;
-                        const signed char* r4 = r0 + N * 4;
-                        const signed char* r5 = r0 + N * 5;
-                        const signed char* r6 = r0 + N * 6;
-                        const signed char* r7 = r0 + N * 7;
-
-                        __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0);
-                        __m128i _t1 = _mm_loadl_epi64((const __m128i*)r1);
-                        __m128i _t2 = _mm_loadl_epi64((const __m128i*)r2);
-                        __m128i _t3 = _mm_loadl_epi64((const __m128i*)r3);
-                        __m128i _t4 = _mm_loadl_epi64((const __m128i*)r4);
-                        __m128i _t5 = _mm_loadl_epi64((const __m128i*)r5);
-                        __m128i _t6 = _mm_loadl_epi64((const __m128i*)r6);
-                        __m128i _t7 = _mm_loadl_epi64((const __m128i*)r7);
-
-                        __m128i _t01 = _mm_unpacklo_epi8(_t0, _t1);
-                        __m128i _t23 = _mm_unpacklo_epi8(_t2, _t3);
-                        __m128i _t45 = _mm_unpacklo_epi8(_t4, _t5);
-                        __m128i _t67 = _mm_unpacklo_epi8(_t6, _t7);
-                        _t0 = _mm_unpacklo_epi16(_t01, _t23);
-                        _t1 = _mm_unpacklo_epi16(_t45, _t67);
-                        _t2 = _mm_unpacklo_epi32(_t0, _t1);
-                        _t3 = _mm_unpackhi_epi32(_t0, _t1);
-
-                        __m128i _extt2 = _mm_cmpgt_epi8(_mm_setzero_si128(), _t2);
-                        __m128i _extt3 = _mm_cmpgt_epi8(_mm_setzero_si128(), _t3);
-
-                        _r0 = _mm_unpacklo_epi8(_t2, _extt2);
-                        if (tj * 4 + 1 < w) _r1 = _mm_unpackhi_epi8(_t2, _extt2);
-                        if (tj * 4 + 2 < w) _r2 = _mm_unpacklo_epi8(_t3, _extt3);
-                        if (tj * 4 + 3 < w) _r3 = _mm_unpackhi_epi8(_t3, _extt3);
-                        if (tj * 4 + 4 < w) _r4 = _mm_setr_epi16(r0[4], r1[4], r2[4], r3[4], r4[4], r5[4], r6[4], r7[4]);
-                        if (tj * 4 + 5 < w) _r5 = _mm_setr_epi16(r0[5], r1[5], r2[5], r3[5], r4[5], r5[5], r6[5], r7[5]);
-#endif // __AVX2__
-                    }
-                }
-
-                __m128i _tmp12a = _mm_sub_epi16(_r3, _mm_mullo_epi16(_r1, _v4));
-                __m128i _tmp12b = _mm_sub_epi16(_r4, _mm_mullo_epi16(_r2, _v4));
-                __m128i _tmp34a = _mm_mullo_epi16(_mm_sub_epi16(_r3, _r1), _v2);
-                __m128i _tmp34b = _mm_sub_epi16(_r4, _r2);
-
-                __m128i _tmp0 = _mm_add_epi16(_r4, _mm_sub_epi16(_mm_mullo_epi16(_r0, _v4), _mm_mullo_epi16(_r2, _v5)));
-                __m128i _tmp1 = _mm_add_epi16(_tmp12b, _tmp12a);
-                __m128i _tmp2 = _mm_sub_epi16(_tmp12b, _tmp12a);
-                __m128i _tmp3 = _mm_add_epi16(_tmp34b, _tmp34a);
-                __m128i _tmp4 = _mm_sub_epi16(_tmp34b, _tmp34a);
-                __m128i _tmp5 = _mm_add_epi16(_r5, _mm_sub_epi16(_mm_mullo_epi16(_r1, _v4), _mm_mullo_epi16(_r3, _v5)));
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm_storeu_si128((__m128i*)tmp[0][m], _tmp0);
-                _mm_storeu_si128((__m128i*)tmp[1][m], _tmp1);
-                _mm_storeu_si128((__m128i*)tmp[2][m], _tmp2);
-                _mm_storeu_si128((__m128i*)tmp[3][m], _tmp3);
-                _mm_storeu_si128((__m128i*)tmp[4][m], _tmp4);
-                _mm_storeu_si128((__m128i*)tmp[5][m], _tmp5);
-#else
-                _mm_store_si128((__m128i*)tmp[0][m], _tmp0);
-                _mm_store_si128((__m128i*)tmp[1][m], _tmp1);
-                _mm_store_si128((__m128i*)tmp[2][m], _tmp2);
-                _mm_store_si128((__m128i*)tmp[3][m], _tmp3);
-                _mm_store_si128((__m128i*)tmp[4][m], _tmp4);
-                _mm_store_si128((__m128i*)tmp[5][m], _tmp5);
-#endif
-
-                r0 += w * elempack;
-            }
-
-            short* p0 = (short*)B + kk * max_jj * 36 + jj * 8;
-            short* p1 = p0 + max_jj * 8;
-            short* p2 = p0 + max_jj * 8 * 2;
-            short* p3 = p0 + max_jj * 8 * 3;
-            short* p4 = p0 + max_jj * 8 * 4;
-            short* p5 = p0 + max_jj * 8 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m128i _r0 = _mm_loadu_si128((const __m128i*)tmp[m][0]);
-                __m128i _r1 = _mm_loadu_si128((const __m128i*)tmp[m][1]);
-                __m128i _r2 = _mm_loadu_si128((const __m128i*)tmp[m][2]);
-                __m128i _r3 = _mm_loadu_si128((const __m128i*)tmp[m][3]);
-                __m128i _r4 = _mm_loadu_si128((const __m128i*)tmp[m][4]);
-                __m128i _r5 = _mm_loadu_si128((const __m128i*)tmp[m][5]);
-#else
-                __m128i _r0 = _mm_load_si128((const __m128i*)tmp[m][0]);
-                __m128i _r1 = _mm_load_si128((const __m128i*)tmp[m][1]);
-                __m128i _r2 = _mm_load_si128((const __m128i*)tmp[m][2]);
-                __m128i _r3 = _mm_load_si128((const __m128i*)tmp[m][3]);
-                __m128i _r4 = _mm_load_si128((const __m128i*)tmp[m][4]);
-                __m128i _r5 = _mm_load_si128((const __m128i*)tmp[m][5]);
-#endif
-
-                __m128i _tmp12a = _mm_sub_epi16(_r3, _mm_mullo_epi16(_r1, _v4));
-                __m128i _tmp12b = _mm_sub_epi16(_r4, _mm_mullo_epi16(_r2, _v4));
-                __m128i _tmp34a = _mm_mullo_epi16(_mm_sub_epi16(_r3, _r1), _v2);
-                __m128i _tmp34b = _mm_sub_epi16(_r4, _r2);
-
-                __m128i _tmp0 = _mm_add_epi16(_r4, _mm_sub_epi16(_mm_mullo_epi16(_r0, _v4), _mm_mullo_epi16(_r2, _v5)));
-                __m128i _tmp1 = _mm_add_epi16(_tmp12b, _tmp12a);
-                __m128i _tmp2 = _mm_sub_epi16(_tmp12b, _tmp12a);
-                __m128i _tmp3 = _mm_add_epi16(_tmp34b, _tmp34a);
-                __m128i _tmp4 = _mm_sub_epi16(_tmp34b, _tmp34a);
-                __m128i _tmp5 = _mm_add_epi16(_r5, _mm_sub_epi16(_mm_mullo_epi16(_r1, _v4), _mm_mullo_epi16(_r3, _v5)));
-
-                _mm_store_si128((__m128i*)p0, _tmp0);
-                _mm_store_si128((__m128i*)p1, _tmp1);
-                _mm_store_si128((__m128i*)p2, _tmp2);
-                _mm_store_si128((__m128i*)p3, _tmp3);
-                _mm_store_si128((__m128i*)p4, _tmp4);
-                _mm_store_si128((__m128i*)p5, _tmp5);
-
-                p0 += max_jj * 6 * 8;
-                p1 += max_jj * 6 * 8;
-                p2 += max_jj * 6 * 8;
-                p3 += max_jj * 6 * 8;
-                p4 += max_jj * 6 * 8;
-                p5 += max_jj * 6 * 8;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 8;
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-#else // __SSE2__
-    nn_max_kk = (max_kk - remain_max_kk_start) / 2;
-    #pragma omp parallel for num_threads(nT)
-#endif // __SSE2__
-    for (int ppkk = 0; ppkk < nn_max_kk; ppkk++)
-    {
-        const int kk = remain_max_kk_start + ppkk * 2;
-
-        short tmp[6][6][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const signed char* r0 = bottom_blob.channel(k + kk).row<const signed char>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 6; m++)
-            {
-                signed char r00 = 0;
-                signed char r01 = 0;
-                signed char r10 = 0;
-                signed char r11 = 0;
-                signed char r20 = 0;
-                signed char r21 = 0;
-                signed char r30 = 0;
-                signed char r31 = 0;
-                signed char r40 = 0;
-                signed char r41 = 0;
-                signed char r50 = 0;
-                signed char r51 = 0;
-
-                if (ti * 4 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        const signed char* r1 = r0 + N;
-
-                        r00 = r0[0];
-                        r01 = r1[0];
-                        if (tj * 4 + 1 < w)
-                        {
-                            r10 = r0[1];
-                            r11 = r1[1];
-                        }
-                        if (tj * 4 + 2 < w)
-                        {
-                            r20 = r0[2];
-                            r21 = r1[2];
-                        }
-                        if (tj * 4 + 3 < w)
-                        {
-                            r30 = r0[3];
-                            r31 = r1[3];
-                        }
-                        if (tj * 4 + 4 < w)
-                        {
-                            r40 = r0[4];
-                            r41 = r1[4];
-                        }
-                        if (tj * 4 + 5 < w)
-                        {
-                            r50 = r0[5];
-                            r51 = r1[5];
-                        }
-                    }
-                }
-
-                short tmp120a = r30 - r10 * 4;
-                short tmp121a = r31 - r11 * 4;
-                short tmp120b = r40 - r20 * 4;
-                short tmp121b = r41 - r21 * 4;
-                short tmp340a = (r30 - r10) * 2;
-                short tmp341a = (r31 - r11) * 2;
-                short tmp340b = r40 - r20;
-                short tmp341b = r41 - r21;
-
-                tmp[0][m][0] = r40 + r00 * 4 - r20 * 5;
-                tmp[0][m][1] = r41 + r01 * 4 - r21 * 5;
-                tmp[1][m][0] = tmp120b + tmp120a;
-                tmp[1][m][1] = tmp121b + tmp121a;
-                tmp[2][m][0] = tmp120b - tmp120a;
-                tmp[2][m][1] = tmp121b - tmp121a;
-                tmp[3][m][0] = tmp340b + tmp340a;
-                tmp[3][m][1] = tmp341b + tmp341a;
-                tmp[4][m][0] = tmp340b - tmp340a;
-                tmp[4][m][1] = tmp341b - tmp341a;
-                tmp[5][m][0] = r50 + r10 * 4 - r30 * 5;
-                tmp[5][m][1] = r51 + r11 * 4 - r31 * 5;
-
-                r0 += w;
-            }
-
-            short* p0 = (short*)B + kk * max_jj * 36 + jj * 2;
-            short* p1 = p0 + max_jj * 2;
-            short* p2 = p0 + max_jj * 2 * 2;
-            short* p3 = p0 + max_jj * 2 * 3;
-            short* p4 = p0 + max_jj * 2 * 4;
-            short* p5 = p0 + max_jj * 2 * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                short r00 = tmp[m][0][0];
-                short r01 = tmp[m][0][1];
-                short r10 = tmp[m][1][0];
-                short r11 = tmp[m][1][1];
-                short r20 = tmp[m][2][0];
-                short r21 = tmp[m][2][1];
-                short r30 = tmp[m][3][0];
-                short r31 = tmp[m][3][1];
-                short r40 = tmp[m][4][0];
-                short r41 = tmp[m][4][1];
-                short r50 = tmp[m][5][0];
-                short r51 = tmp[m][5][1];
-
-                short tmp120a = r30 - r10 * 4;
-                short tmp121a = r31 - r11 * 4;
-                short tmp120b = r40 - r20 * 4;
-                short tmp121b = r41 - r21 * 4;
-                short tmp340a = (r30 - r10) * 2;
-                short tmp341a = (r31 - r11) * 2;
-                short tmp340b = r40 - r20;
-                short tmp341b = r41 - r21;
-
-                p0[0] = r40 + r00 * 4 - r20 * 5;
-                p0[1] = r41 + r01 * 4 - r21 * 5;
-                p1[0] = tmp120b + tmp120a;
-                p1[1] = tmp121b + tmp121a;
-                p2[0] = tmp120b - tmp120a;
-                p2[1] = tmp121b - tmp121a;
-                p3[0] = tmp340b + tmp340a;
-                p3[1] = tmp341b + tmp341a;
-                p4[0] = tmp340b - tmp340a;
-                p4[1] = tmp341b - tmp341a;
-                p5[0] = r50 + r10 * 4 - r30 * 5;
-                p5[1] = r51 + r11 * 4 - r31 * 5;
-
-                p0 += max_jj * 6 * 2;
-                p1 += max_jj * 6 * 2;
-                p2 += max_jj * 6 * 2;
-                p3 += max_jj * 6 * 2;
-                p4 += max_jj * 6 * 2;
-                p5 += max_jj * 6 * 2;
-            }
-        }
-    }
-    remain_max_kk_start += nn_max_kk * 2;
-    for (int kk = remain_max_kk_start; kk < max_kk; kk++)
-    {
-        short tmp[6][6];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const signed char* r0123 = bottom_blob.channel(k + kk).row<const signed char>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 6; m++)
-            {
-                signed char r0 = 0;
-                signed char r1 = 0;
-                signed char r2 = 0;
-                signed char r3 = 0;
-                signed char r4 = 0;
-                signed char r5 = 0;
-
-                if (ti * 4 + m < h)
-                {
-                    // if (elempack == 1)
-                    {
-                        r0 = r0123[0];
-                        if (tj * 4 + 1 < w) r1 = r0123[1];
-                        if (tj * 4 + 2 < w) r2 = r0123[2];
-                        if (tj * 4 + 3 < w) r3 = r0123[3];
-                        if (tj * 4 + 4 < w) r4 = r0123[4];
-                        if (tj * 4 + 5 < w) r5 = r0123[5];
-                    }
-                }
-
-                short tmp12a = r3 - r1 * 4;
-                short tmp12b = r4 - r2 * 4;
-                short tmp34a = (r3 - r1) * 2;
-                short tmp34b = r4 - r2;
-
-                tmp[0][m] = r4 + r0 * 4 - r2 * 5;
-                tmp[1][m] = tmp12b + tmp12a;
-                tmp[2][m] = tmp12b - tmp12a;
-                tmp[3][m] = tmp34b + tmp34a;
-                tmp[4][m] = tmp34b - tmp34a;
-                tmp[5][m] = r5 + r1 * 4 - r3 * 5;
-
-                r0123 += w;
-            }
-
-            short* p0 = (short*)B + kk * max_jj * 36 + jj;
-            short* p1 = p0 + max_jj;
-            short* p2 = p0 + max_jj * 2;
-            short* p3 = p0 + max_jj * 3;
-            short* p4 = p0 + max_jj * 4;
-            short* p5 = p0 + max_jj * 5;
-
-            for (int m = 0; m < 6; m++)
-            {
-                short r0 = tmp[m][0];
-                short r1 = tmp[m][1];
-                short r2 = tmp[m][2];
-                short r3 = tmp[m][3];
-                short r4 = tmp[m][4];
-                short r5 = tmp[m][5];
-
-                short tmp12a = r3 - r1 * 4;
-                short tmp12b = r4 - r2 * 4;
-                short tmp34a = (r3 - r1) * 2;
-                short tmp34b = r4 - r2;
-
-                p0[0] = r4 + r0 * 4 - r2 * 5;
-                p1[0] = tmp12b + tmp12a;
-                p2[0] = tmp12b - tmp12a;
-                p3[0] = tmp34b + tmp34a;
-                p4[0] = tmp34b - tmp34a;
-                p5[0] = r5 + r1 * 4 - r3 * 5;
-
-                p0 += max_jj * 6;
-                p1 += max_jj * 6;
-                p2 += max_jj * 6;
-                p3 += max_jj * 6;
-                p4 += max_jj * 6;
-                p5 += max_jj * 6;
-            }
-        }
-    }
-}
-
-static inline void conv3x3s1_winograd43_transform_output_tile_int8(const Mat& top_tile, Mat& top_blob, int i, int max_ii, int j, int max_jj)
-{
-    // const int otm[4][6] = {
-    //     {1, 1,  1, 1,  1, 0},
-    //     {0, 1, -1, 2, -2, 0},
-    //     {0, 1,  1, 4,  4, 0},
-    //     {0, 1, -1, 8, -8, 1}
-    // };
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int N = top_blob.cstep * out_elempack;
-
-    const int w_tiles = (outw + 3) / 4;
-
-    int ii = 0;
-#if __SSE2__
-#if __AVX2__
-#if __AVX512F__
-    for (; ii + 15 < max_ii; ii += 16)
-    {
-#ifdef _MSC_VER
-        __declspec(align(64))
-#else
-        __attribute__((aligned(64)))
-#endif
-        int tmp[4][6][16];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 36 + jj * 16;
-            const int* r1 = r0 + max_jj * 16;
-            const int* r2 = r0 + max_jj * 16 * 2;
-            const int* r3 = r0 + max_jj * 16 * 3;
-            const int* r4 = r0 + max_jj * 16 * 4;
-            const int* r5 = r0 + max_jj * 16 * 5;
-
-            for (int m = 0; m < 5; m++)
-            {
-                __m512i _r0 = _mm512_load_si512((const __m512i*)r0);
-                __m512i _r1 = _mm512_load_si512((const __m512i*)r1);
-                __m512i _r2 = _mm512_load_si512((const __m512i*)r2);
-                __m512i _r3 = _mm512_load_si512((const __m512i*)r3);
-                __m512i _r4 = _mm512_load_si512((const __m512i*)r4);
-                __m512i _r5 = _mm512_load_si512((const __m512i*)r5);
-
-                __m512i _tmp02a = _mm512_add_epi32(_r1, _r2);
-                __m512i _tmp02b = _mm512_add_epi32(_r3, _r4);
-                __m512i _tmp13a = _mm512_sub_epi32(_r1, _r2);
-                __m512i _tmp13b = _mm512_sub_epi32(_r3, _r4);
-
-                __m512i _tmp0 = _mm512_add_epi32(_mm512_add_epi32(_tmp02a, _tmp02b), _r0);
-                __m512i _tmp1 = _mm512_add_epi32(_tmp13a, _mm512_slli_epi32(_tmp13b, 1));
-                __m512i _tmp2 = _mm512_add_epi32(_tmp02a, _mm512_slli_epi32(_tmp02b, 2));
-                __m512i _tmp3 = _mm512_add_epi32(_mm512_add_epi32(_tmp13a, _mm512_slli_epi32(_tmp13b, 3)), _mm512_slli_epi32(_r5, 2));
-
-                _mm512_store_si512((__m512i*)tmp[0][m], _tmp0);
-                _mm512_store_si512((__m512i*)tmp[1][m], _tmp1);
-                _mm512_store_si512((__m512i*)tmp[2][m], _tmp2);
-                _mm512_store_si512((__m512i*)tmp[3][m], _tmp3);
-
-                r0 += max_jj * 6 * 16;
-                r1 += max_jj * 6 * 16;
-                r2 += max_jj * 6 * 16;
-                r3 += max_jj * 6 * 16;
-                r4 += max_jj * 6 * 16;
-                r5 += max_jj * 6 * 16;
-            }
-            for (int m = 5; m < 6; m++)
-            {
-                __m512i _r0 = _mm512_load_si512((const __m512i*)r0);
-                __m512i _r1 = _mm512_load_si512((const __m512i*)r1);
-                __m512i _r2 = _mm512_load_si512((const __m512i*)r2);
-                __m512i _r3 = _mm512_load_si512((const __m512i*)r3);
-                __m512i _r4 = _mm512_load_si512((const __m512i*)r4);
-                __m512i _r5 = _mm512_load_si512((const __m512i*)r5);
-
-                __m512i _tmp02a = _mm512_add_epi32(_r1, _r2);
-                __m512i _tmp02b = _mm512_add_epi32(_r3, _r4);
-                __m512i _tmp13a = _mm512_sub_epi32(_r1, _r2);
-                __m512i _tmp13b = _mm512_sub_epi32(_r3, _r4);
-
-                __m512i _tmp0 = _mm512_add_epi32(_mm512_add_epi32(_tmp02a, _tmp02b), _r0);
-                __m512i _tmp1 = _mm512_add_epi32(_tmp13a, _mm512_slli_epi32(_tmp13b, 1));
-                __m512i _tmp2 = _mm512_add_epi32(_tmp02a, _mm512_slli_epi32(_tmp02b, 2));
-                __m512i _tmp3 = _mm512_add_epi32(_mm512_add_epi32(_tmp13a, _mm512_slli_epi32(_tmp13b, 3)), _mm512_slli_epi32(_r5, 2));
-
-                _tmp0 = _mm512_slli_epi32(_tmp0, 2);
-                _tmp1 = _mm512_slli_epi32(_tmp1, 2);
-                _tmp2 = _mm512_slli_epi32(_tmp2, 2);
-                _tmp3 = _mm512_slli_epi32(_tmp3, 2);
-
-                _mm512_store_si512((__m512i*)tmp[0][m], _tmp0);
-                _mm512_store_si512((__m512i*)tmp[1][m], _tmp1);
-                _mm512_store_si512((__m512i*)tmp[2][m], _tmp2);
-                _mm512_store_si512((__m512i*)tmp[3][m], _tmp3);
-
-                r0 += max_jj * 6 * 16;
-                r1 += max_jj * 6 * 16;
-                r2 += max_jj * 6 * 16;
-                r3 += max_jj * 6 * 16;
-                r4 += max_jj * 6 * 16;
-                r5 += max_jj * 6 * 16;
-            }
-
-            int* outptr0 = top_blob.channel((i + ii) / out_elempack).row<int>(ti * 4) + (tj * 4) * out_elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                __m512i _r0 = _mm512_load_si512((const __m512i*)tmp[m][0]);
-                __m512i _r1 = _mm512_load_si512((const __m512i*)tmp[m][1]);
-                __m512i _r2 = _mm512_load_si512((const __m512i*)tmp[m][2]);
-                __m512i _r3 = _mm512_load_si512((const __m512i*)tmp[m][3]);
-                __m512i _r4 = _mm512_load_si512((const __m512i*)tmp[m][4]);
-                __m512i _r5 = _mm512_load_si512((const __m512i*)tmp[m][5]);
-
-                __m512i _tmp02a = _mm512_add_epi32(_r1, _r2);
-                __m512i _tmp02b = _mm512_add_epi32(_r3, _r4);
-                __m512i _tmp13a = _mm512_sub_epi32(_r1, _r2);
-                __m512i _tmp13b = _mm512_sub_epi32(_r3, _r4);
-
-                __m512i _tmp0 = _mm512_add_epi32(_mm512_add_epi32(_tmp02a, _tmp02b), _r0);
-                __m512i _tmp1 = _mm512_add_epi32(_tmp13a, _mm512_slli_epi32(_tmp13b, 1));
-                __m512i _tmp2 = _mm512_add_epi32(_tmp02a, _mm512_slli_epi32(_tmp02b, 2));
-                __m512i _tmp3 = _mm512_add_epi32(_mm512_add_epi32(_tmp13a, _mm512_slli_epi32(_tmp13b, 3)), _r5);
-
-                // TODO use integer trick for division by 576
-                __m512 _v576 = _mm512_set1_ps(1.0 / 576);
-                _tmp0 = _mm512_cvttps_epi32(_mm512_mul_ps(_mm512_cvtepi32_ps(_tmp0), _v576));
-                _tmp1 = _mm512_cvttps_epi32(_mm512_mul_ps(_mm512_cvtepi32_ps(_tmp1), _v576));
-                _tmp2 = _mm512_cvttps_epi32(_mm512_mul_ps(_mm512_cvtepi32_ps(_tmp2), _v576));
-                _tmp3 = _mm512_cvttps_epi32(_mm512_mul_ps(_mm512_cvtepi32_ps(_tmp3), _v576));
-
-                if (out_elempack == 16)
-                {
-                    _mm512_store_si512((__m512i*)outptr0, _tmp0);
-                    if (tj * 4 + 1 < outw) _mm512_store_si512((__m512i*)(outptr0 + 16), _tmp1);
-                    if (tj * 4 + 2 < outw) _mm512_store_si512((__m512i*)(outptr0 + 32), _tmp2);
-                    if (tj * 4 + 3 < outw) _mm512_store_si512((__m512i*)(outptr0 + 48), _tmp3);
-                }
-                if (out_elempack == 8)
-                {
-                    int* outptr1 = outptr0 + N;
-
-                    _mm256_store_si256((__m256i*)outptr0, _mm512_extracti32x8_epi32(_tmp0, 0));
-                    _mm256_store_si256((__m256i*)outptr1, _mm512_extracti32x8_epi32(_tmp0, 1));
-                    if (tj * 4 + 1 < outw)
-                    {
-                        _mm256_store_si256((__m256i*)(outptr0 + 8), _mm512_extracti32x8_epi32(_tmp1, 0));
-                        _mm256_store_si256((__m256i*)(outptr1 + 8), _mm512_extracti32x8_epi32(_tmp1, 1));
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        _mm256_store_si256((__m256i*)(outptr0 + 16), _mm512_extracti32x8_epi32(_tmp2, 0));
-                        _mm256_store_si256((__m256i*)(outptr1 + 16), _mm512_extracti32x8_epi32(_tmp2, 1));
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        _mm256_store_si256((__m256i*)(outptr0 + 24), _mm512_extracti32x8_epi32(_tmp3, 0));
-                        _mm256_store_si256((__m256i*)(outptr1 + 24), _mm512_extracti32x8_epi32(_tmp3, 1));
-                    }
-                }
-                if (out_elempack == 4)
-                {
-                    int* outptr1 = outptr0 + N;
-                    int* outptr2 = outptr0 + N * 2;
-                    int* outptr3 = outptr0 + N * 3;
-
-                    _mm_store_si128((__m128i*)outptr0, _mm512_extracti32x4_epi32(_tmp0, 0));
-                    _mm_store_si128((__m128i*)outptr1, _mm512_extracti32x4_epi32(_tmp0, 1));
-                    _mm_store_si128((__m128i*)outptr2, _mm512_extracti32x4_epi32(_tmp0, 2));
-                    _mm_store_si128((__m128i*)outptr3, _mm512_extracti32x4_epi32(_tmp0, 3));
-                    if (tj * 4 + 1 < outw)
-                    {
-                        _mm_store_si128((__m128i*)(outptr0 + 4), _mm512_extracti32x4_epi32(_tmp1, 0));
-                        _mm_store_si128((__m128i*)(outptr1 + 4), _mm512_extracti32x4_epi32(_tmp1, 1));
-                        _mm_store_si128((__m128i*)(outptr2 + 4), _mm512_extracti32x4_epi32(_tmp1, 2));
-                        _mm_store_si128((__m128i*)(outptr3 + 4), _mm512_extracti32x4_epi32(_tmp1, 3));
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        _mm_store_si128((__m128i*)(outptr0 + 8), _mm512_extracti32x4_epi32(_tmp2, 0));
-                        _mm_store_si128((__m128i*)(outptr1 + 8), _mm512_extracti32x4_epi32(_tmp2, 1));
-                        _mm_store_si128((__m128i*)(outptr2 + 8), _mm512_extracti32x4_epi32(_tmp2, 2));
-                        _mm_store_si128((__m128i*)(outptr3 + 8), _mm512_extracti32x4_epi32(_tmp2, 3));
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        _mm_store_si128((__m128i*)(outptr0 + 12), _mm512_extracti32x4_epi32(_tmp3, 0));
-                        _mm_store_si128((__m128i*)(outptr1 + 12), _mm512_extracti32x4_epi32(_tmp3, 1));
-                        _mm_store_si128((__m128i*)(outptr2 + 12), _mm512_extracti32x4_epi32(_tmp3, 2));
-                        _mm_store_si128((__m128i*)(outptr3 + 12), _mm512_extracti32x4_epi32(_tmp3, 3));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-                    __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                    _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                    _mm512_i32scatter_epi32(outptr0, _vindex, _tmp0, sizeof(int));
-                    if (tj * 4 + 1 < outw) _mm512_i32scatter_epi32(outptr0 + 1, _vindex, _tmp1, sizeof(int));
-                    if (tj * 4 + 2 < outw) _mm512_i32scatter_epi32(outptr0 + 2, _vindex, _tmp2, sizeof(int));
-                    if (tj * 4 + 3 < outw) _mm512_i32scatter_epi32(outptr0 + 3, _vindex, _tmp3, sizeof(int));
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __AVX512F__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-#ifdef _MSC_VER
-        __declspec(align(32))
-#else
-        __attribute__((aligned(32)))
-#endif
-        int tmp[4][6][8];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 36 + jj * 8;
-            const int* r1 = r0 + max_jj * 8;
-            const int* r2 = r0 + max_jj * 8 * 2;
-            const int* r3 = r0 + max_jj * 8 * 3;
-            const int* r4 = r0 + max_jj * 8 * 4;
-            const int* r5 = r0 + max_jj * 8 * 5;
-
-            for (int m = 0; m < 5; m++)
-            {
-                __m256i _r0 = _mm256_load_si256((const __m256i*)r0);
-                __m256i _r1 = _mm256_load_si256((const __m256i*)r1);
-                __m256i _r2 = _mm256_load_si256((const __m256i*)r2);
-                __m256i _r3 = _mm256_load_si256((const __m256i*)r3);
-                __m256i _r4 = _mm256_load_si256((const __m256i*)r4);
-                __m256i _r5 = _mm256_load_si256((const __m256i*)r5);
-
-                __m256i _tmp02a = _mm256_add_epi32(_r1, _r2);
-                __m256i _tmp02b = _mm256_add_epi32(_r3, _r4);
-                __m256i _tmp13a = _mm256_sub_epi32(_r1, _r2);
-                __m256i _tmp13b = _mm256_sub_epi32(_r3, _r4);
-
-                __m256i _tmp0 = _mm256_add_epi32(_mm256_add_epi32(_tmp02a, _tmp02b), _r0);
-                __m256i _tmp1 = _mm256_add_epi32(_tmp13a, _mm256_slli_epi32(_tmp13b, 1));
-                __m256i _tmp2 = _mm256_add_epi32(_tmp02a, _mm256_slli_epi32(_tmp02b, 2));
-                __m256i _tmp3 = _mm256_add_epi32(_mm256_add_epi32(_tmp13a, _mm256_slli_epi32(_tmp13b, 3)), _mm256_slli_epi32(_r5, 2));
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm256_storeu_si256((__m256i*)tmp[0][m], _tmp0);
-                _mm256_storeu_si256((__m256i*)tmp[1][m], _tmp1);
-                _mm256_storeu_si256((__m256i*)tmp[2][m], _tmp2);
-                _mm256_storeu_si256((__m256i*)tmp[3][m], _tmp3);
-#else
-                _mm256_store_si256((__m256i*)tmp[0][m], _tmp0);
-                _mm256_store_si256((__m256i*)tmp[1][m], _tmp1);
-                _mm256_store_si256((__m256i*)tmp[2][m], _tmp2);
-                _mm256_store_si256((__m256i*)tmp[3][m], _tmp3);
-#endif
-
-                r0 += max_jj * 6 * 8;
-                r1 += max_jj * 6 * 8;
-                r2 += max_jj * 6 * 8;
-                r3 += max_jj * 6 * 8;
-                r4 += max_jj * 6 * 8;
-                r5 += max_jj * 6 * 8;
-            }
-            for (int m = 5; m < 6; m++)
-            {
-                __m256i _r0 = _mm256_load_si256((const __m256i*)r0);
-                __m256i _r1 = _mm256_load_si256((const __m256i*)r1);
-                __m256i _r2 = _mm256_load_si256((const __m256i*)r2);
-                __m256i _r3 = _mm256_load_si256((const __m256i*)r3);
-                __m256i _r4 = _mm256_load_si256((const __m256i*)r4);
-                __m256i _r5 = _mm256_load_si256((const __m256i*)r5);
-
-                __m256i _tmp02a = _mm256_add_epi32(_r1, _r2);
-                __m256i _tmp02b = _mm256_add_epi32(_r3, _r4);
-                __m256i _tmp13a = _mm256_sub_epi32(_r1, _r2);
-                __m256i _tmp13b = _mm256_sub_epi32(_r3, _r4);
-
-                __m256i _tmp0 = _mm256_add_epi32(_mm256_add_epi32(_tmp02a, _tmp02b), _r0);
-                __m256i _tmp1 = _mm256_add_epi32(_tmp13a, _mm256_slli_epi32(_tmp13b, 1));
-                __m256i _tmp2 = _mm256_add_epi32(_tmp02a, _mm256_slli_epi32(_tmp02b, 2));
-                __m256i _tmp3 = _mm256_add_epi32(_mm256_add_epi32(_tmp13a, _mm256_slli_epi32(_tmp13b, 3)), _mm256_slli_epi32(_r5, 2));
-
-                _tmp0 = _mm256_slli_epi32(_tmp0, 2);
-                _tmp1 = _mm256_slli_epi32(_tmp1, 2);
-                _tmp2 = _mm256_slli_epi32(_tmp2, 2);
-                _tmp3 = _mm256_slli_epi32(_tmp3, 2);
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm256_storeu_si256((__m256i*)tmp[0][m], _tmp0);
-                _mm256_storeu_si256((__m256i*)tmp[1][m], _tmp1);
-                _mm256_storeu_si256((__m256i*)tmp[2][m], _tmp2);
-                _mm256_storeu_si256((__m256i*)tmp[3][m], _tmp3);
-#else
-                _mm256_store_si256((__m256i*)tmp[0][m], _tmp0);
-                _mm256_store_si256((__m256i*)tmp[1][m], _tmp1);
-                _mm256_store_si256((__m256i*)tmp[2][m], _tmp2);
-                _mm256_store_si256((__m256i*)tmp[3][m], _tmp3);
-#endif
-
-                r0 += max_jj * 6 * 8;
-                r1 += max_jj * 6 * 8;
-                r2 += max_jj * 6 * 8;
-                r3 += max_jj * 6 * 8;
-                r4 += max_jj * 6 * 8;
-                r5 += max_jj * 6 * 8;
-            }
-
-            int* outptr0 = top_blob.channel((i + ii) / out_elempack).row<int>(ti * 4) + (tj * 4) * out_elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m256i _r0 = _mm256_loadu_si256((const __m256i*)tmp[m][0]);
-                __m256i _r1 = _mm256_loadu_si256((const __m256i*)tmp[m][1]);
-                __m256i _r2 = _mm256_loadu_si256((const __m256i*)tmp[m][2]);
-                __m256i _r3 = _mm256_loadu_si256((const __m256i*)tmp[m][3]);
-                __m256i _r4 = _mm256_loadu_si256((const __m256i*)tmp[m][4]);
-                __m256i _r5 = _mm256_loadu_si256((const __m256i*)tmp[m][5]);
-#else
-                __m256i _r0 = _mm256_load_si256((const __m256i*)tmp[m][0]);
-                __m256i _r1 = _mm256_load_si256((const __m256i*)tmp[m][1]);
-                __m256i _r2 = _mm256_load_si256((const __m256i*)tmp[m][2]);
-                __m256i _r3 = _mm256_load_si256((const __m256i*)tmp[m][3]);
-                __m256i _r4 = _mm256_load_si256((const __m256i*)tmp[m][4]);
-                __m256i _r5 = _mm256_load_si256((const __m256i*)tmp[m][5]);
-#endif
-
-                __m256i _tmp02a = _mm256_add_epi32(_r1, _r2);
-                __m256i _tmp02b = _mm256_add_epi32(_r3, _r4);
-                __m256i _tmp13a = _mm256_sub_epi32(_r1, _r2);
-                __m256i _tmp13b = _mm256_sub_epi32(_r3, _r4);
-
-                __m256i _tmp0 = _mm256_add_epi32(_mm256_add_epi32(_tmp02a, _tmp02b), _r0);
-                __m256i _tmp1 = _mm256_add_epi32(_tmp13a, _mm256_slli_epi32(_tmp13b, 1));
-                __m256i _tmp2 = _mm256_add_epi32(_tmp02a, _mm256_slli_epi32(_tmp02b, 2));
-                __m256i _tmp3 = _mm256_add_epi32(_mm256_add_epi32(_tmp13a, _mm256_slli_epi32(_tmp13b, 3)), _r5);
-
-                // TODO use integer trick for division by 576
-                __m256 _v576 = _mm256_set1_ps(1.0 / 576);
-                _tmp0 = _mm256_cvttps_epi32(_mm256_mul_ps(_mm256_cvtepi32_ps(_tmp0), _v576));
-                _tmp1 = _mm256_cvttps_epi32(_mm256_mul_ps(_mm256_cvtepi32_ps(_tmp1), _v576));
-                _tmp2 = _mm256_cvttps_epi32(_mm256_mul_ps(_mm256_cvtepi32_ps(_tmp2), _v576));
-                _tmp3 = _mm256_cvttps_epi32(_mm256_mul_ps(_mm256_cvtepi32_ps(_tmp3), _v576));
-
-                if (out_elempack == 8)
-                {
-                    _mm256_store_si256((__m256i*)outptr0, _tmp0);
-                    if (tj * 4 + 1 < outw) _mm256_store_si256((__m256i*)(outptr0 + 8), _tmp1);
-                    if (tj * 4 + 2 < outw) _mm256_store_si256((__m256i*)(outptr0 + 16), _tmp2);
-                    if (tj * 4 + 3 < outw) _mm256_store_si256((__m256i*)(outptr0 + 24), _tmp3);
-                }
-                if (out_elempack == 4)
-                {
-                    int* outptr1 = outptr0 + N;
-
-                    _mm_store_si128((__m128i*)(outptr0), _mm256_extracti128_si256(_tmp0, 0));
-                    _mm_store_si128((__m128i*)(outptr1), _mm256_extracti128_si256(_tmp0, 1));
-                    if (tj * 4 + 1 < outw)
-                    {
-                        _mm_store_si128((__m128i*)(outptr0 + 4), _mm256_extracti128_si256(_tmp1, 0));
-                        _mm_store_si128((__m128i*)(outptr1 + 4), _mm256_extracti128_si256(_tmp1, 1));
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        _mm_store_si128((__m128i*)(outptr0 + 8), _mm256_extracti128_si256(_tmp2, 0));
-                        _mm_store_si128((__m128i*)(outptr1 + 8), _mm256_extracti128_si256(_tmp2, 1));
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        _mm_store_si128((__m128i*)(outptr0 + 12), _mm256_extracti128_si256(_tmp3, 0));
-                        _mm_store_si128((__m128i*)(outptr1 + 12), _mm256_extracti128_si256(_tmp3, 1));
-                    }
-                }
-                if (out_elempack == 1)
-                {
-#if __AVX512F__
-                    __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                    _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-                    _mm256_i32scatter_epi32(outptr0, _vindex, _tmp0, sizeof(int));
-                    if (tj * 4 + 1 < outw) _mm256_i32scatter_epi32(outptr0 + 1, _vindex, _tmp1, sizeof(int));
-                    if (tj * 4 + 2 < outw) _mm256_i32scatter_epi32(outptr0 + 2, _vindex, _tmp2, sizeof(int));
-                    if (tj * 4 + 3 < outw) _mm256_i32scatter_epi32(outptr0 + 3, _vindex, _tmp3, sizeof(int));
-#else
-                    int tmp0[8];
-                    int tmp1[8];
-                    int tmp2[8];
-                    int tmp3[8];
-                    _mm256_storeu_si256((__m256i*)tmp0, _tmp0);
-                    _mm256_storeu_si256((__m256i*)tmp1, _tmp1);
-                    _mm256_storeu_si256((__m256i*)tmp2, _tmp2);
-                    _mm256_storeu_si256((__m256i*)tmp3, _tmp3);
-
-                    int* outptr1 = outptr0 + N;
-                    int* outptr2 = outptr0 + N * 2;
-                    int* outptr3 = outptr0 + N * 3;
-                    int* outptr4 = outptr0 + N * 4;
-                    int* outptr5 = outptr0 + N * 5;
-                    int* outptr6 = outptr0 + N * 6;
-                    int* outptr7 = outptr0 + N * 7;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    outptr4[0] = tmp0[4];
-                    outptr5[0] = tmp0[5];
-                    outptr6[0] = tmp0[6];
-                    outptr7[0] = tmp0[7];
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                        outptr4[1] = tmp1[4];
-                        outptr5[1] = tmp1[5];
-                        outptr6[1] = tmp1[6];
-                        outptr7[1] = tmp1[7];
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                        outptr4[2] = tmp2[4];
-                        outptr5[2] = tmp2[5];
-                        outptr6[2] = tmp2[6];
-                        outptr7[2] = tmp2[7];
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                        outptr4[3] = tmp3[4];
-                        outptr5[3] = tmp3[5];
-                        outptr6[3] = tmp3[6];
-                        outptr7[3] = tmp3[7];
-                    }
-#endif // __AVX512F__
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __AVX2__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-#ifdef _MSC_VER
-        __declspec(align(16))
-#else
-        __attribute__((aligned(16)))
-#endif
-        int tmp[4][6][4];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 36 + jj * 4;
-            const int* r1 = r0 + max_jj * 4;
-            const int* r2 = r0 + max_jj * 4 * 2;
-            const int* r3 = r0 + max_jj * 4 * 3;
-            const int* r4 = r0 + max_jj * 4 * 4;
-            const int* r5 = r0 + max_jj * 4 * 5;
-
-            for (int m = 0; m < 5; m++)
-            {
-                __m128i _r0 = _mm_load_si128((const __m128i*)r0);
-                __m128i _r1 = _mm_load_si128((const __m128i*)r1);
-                __m128i _r2 = _mm_load_si128((const __m128i*)r2);
-                __m128i _r3 = _mm_load_si128((const __m128i*)r3);
-                __m128i _r4 = _mm_load_si128((const __m128i*)r4);
-                __m128i _r5 = _mm_load_si128((const __m128i*)r5);
-
-                __m128i _tmp02a = _mm_add_epi32(_r1, _r2);
-                __m128i _tmp02b = _mm_add_epi32(_r3, _r4);
-                __m128i _tmp13a = _mm_sub_epi32(_r1, _r2);
-                __m128i _tmp13b = _mm_sub_epi32(_r3, _r4);
-
-                __m128i _tmp0 = _mm_add_epi32(_mm_add_epi32(_tmp02a, _tmp02b), _r0);
-                __m128i _tmp1 = _mm_add_epi32(_tmp13a, _mm_slli_epi32(_tmp13b, 1));
-                __m128i _tmp2 = _mm_add_epi32(_tmp02a, _mm_slli_epi32(_tmp02b, 2));
-                __m128i _tmp3 = _mm_add_epi32(_mm_add_epi32(_tmp13a, _mm_slli_epi32(_tmp13b, 3)), _mm_slli_epi32(_r5, 2));
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm_storeu_si128((__m128i*)tmp[0][m], _tmp0);
-                _mm_storeu_si128((__m128i*)tmp[1][m], _tmp1);
-                _mm_storeu_si128((__m128i*)tmp[2][m], _tmp2);
-                _mm_storeu_si128((__m128i*)tmp[3][m], _tmp3);
-#else
-                _mm_store_si128((__m128i*)tmp[0][m], _tmp0);
-                _mm_store_si128((__m128i*)tmp[1][m], _tmp1);
-                _mm_store_si128((__m128i*)tmp[2][m], _tmp2);
-                _mm_store_si128((__m128i*)tmp[3][m], _tmp3);
-#endif
-
-                r0 += max_jj * 6 * 4;
-                r1 += max_jj * 6 * 4;
-                r2 += max_jj * 6 * 4;
-                r3 += max_jj * 6 * 4;
-                r4 += max_jj * 6 * 4;
-                r5 += max_jj * 6 * 4;
-            }
-            for (int m = 5; m < 6; m++)
-            {
-                __m128i _r0 = _mm_load_si128((const __m128i*)r0);
-                __m128i _r1 = _mm_load_si128((const __m128i*)r1);
-                __m128i _r2 = _mm_load_si128((const __m128i*)r2);
-                __m128i _r3 = _mm_load_si128((const __m128i*)r3);
-                __m128i _r4 = _mm_load_si128((const __m128i*)r4);
-                __m128i _r5 = _mm_load_si128((const __m128i*)r5);
-
-                __m128i _tmp02a = _mm_add_epi32(_r1, _r2);
-                __m128i _tmp02b = _mm_add_epi32(_r3, _r4);
-                __m128i _tmp13a = _mm_sub_epi32(_r1, _r2);
-                __m128i _tmp13b = _mm_sub_epi32(_r3, _r4);
-
-                __m128i _tmp0 = _mm_add_epi32(_mm_add_epi32(_tmp02a, _tmp02b), _r0);
-                __m128i _tmp1 = _mm_add_epi32(_tmp13a, _mm_slli_epi32(_tmp13b, 1));
-                __m128i _tmp2 = _mm_add_epi32(_tmp02a, _mm_slli_epi32(_tmp02b, 2));
-                __m128i _tmp3 = _mm_add_epi32(_mm_add_epi32(_tmp13a, _mm_slli_epi32(_tmp13b, 3)), _mm_slli_epi32(_r5, 2));
-
-                _tmp0 = _mm_slli_epi32(_tmp0, 2);
-                _tmp1 = _mm_slli_epi32(_tmp1, 2);
-                _tmp2 = _mm_slli_epi32(_tmp2, 2);
-                _tmp3 = _mm_slli_epi32(_tmp3, 2);
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                _mm_storeu_si128((__m128i*)tmp[0][m], _tmp0);
-                _mm_storeu_si128((__m128i*)tmp[1][m], _tmp1);
-                _mm_storeu_si128((__m128i*)tmp[2][m], _tmp2);
-                _mm_storeu_si128((__m128i*)tmp[3][m], _tmp3);
-#else
-                _mm_store_si128((__m128i*)tmp[0][m], _tmp0);
-                _mm_store_si128((__m128i*)tmp[1][m], _tmp1);
-                _mm_store_si128((__m128i*)tmp[2][m], _tmp2);
-                _mm_store_si128((__m128i*)tmp[3][m], _tmp3);
-#endif
-
-                r0 += max_jj * 6 * 4;
-                r1 += max_jj * 6 * 4;
-                r2 += max_jj * 6 * 4;
-                r3 += max_jj * 6 * 4;
-                r4 += max_jj * 6 * 4;
-                r5 += max_jj * 6 * 4;
-            }
-
-            int* outptr0 = top_blob.channel((i + ii) / out_elempack).row<int>(ti * 4) + (tj * 4) * out_elempack;
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-#if defined(__GNUC__) && (__GNUC__ <= 4) && (__GNUC_MINOR__ < 6)
-                __m128i _r0 = _mm_loadu_si128((const __m128i*)tmp[m][0]);
-                __m128i _r1 = _mm_loadu_si128((const __m128i*)tmp[m][1]);
-                __m128i _r2 = _mm_loadu_si128((const __m128i*)tmp[m][2]);
-                __m128i _r3 = _mm_loadu_si128((const __m128i*)tmp[m][3]);
-                __m128i _r4 = _mm_loadu_si128((const __m128i*)tmp[m][4]);
-                __m128i _r5 = _mm_loadu_si128((const __m128i*)tmp[m][5]);
-#else
-                __m128i _r0 = _mm_load_si128((const __m128i*)tmp[m][0]);
-                __m128i _r1 = _mm_load_si128((const __m128i*)tmp[m][1]);
-                __m128i _r2 = _mm_load_si128((const __m128i*)tmp[m][2]);
-                __m128i _r3 = _mm_load_si128((const __m128i*)tmp[m][3]);
-                __m128i _r4 = _mm_load_si128((const __m128i*)tmp[m][4]);
-                __m128i _r5 = _mm_load_si128((const __m128i*)tmp[m][5]);
-#endif
-
-                __m128i _tmp02a = _mm_add_epi32(_r1, _r2);
-                __m128i _tmp02b = _mm_add_epi32(_r3, _r4);
-                __m128i _tmp13a = _mm_sub_epi32(_r1, _r2);
-                __m128i _tmp13b = _mm_sub_epi32(_r3, _r4);
-
-                __m128i _tmp0 = _mm_add_epi32(_mm_add_epi32(_tmp02a, _tmp02b), _r0);
-                __m128i _tmp1 = _mm_add_epi32(_tmp13a, _mm_slli_epi32(_tmp13b, 1));
-                __m128i _tmp2 = _mm_add_epi32(_tmp02a, _mm_slli_epi32(_tmp02b, 2));
-                __m128i _tmp3 = _mm_add_epi32(_mm_add_epi32(_tmp13a, _mm_slli_epi32(_tmp13b, 3)), _r5);
-
-                // TODO use integer trick for division by 576
-                __m128 _v576 = _mm_set1_ps(1.0 / 576);
-                _tmp0 = _mm_cvttps_epi32(_mm_mul_ps(_mm_cvtepi32_ps(_tmp0), _v576));
-                _tmp1 = _mm_cvttps_epi32(_mm_mul_ps(_mm_cvtepi32_ps(_tmp1), _v576));
-                _tmp2 = _mm_cvttps_epi32(_mm_mul_ps(_mm_cvtepi32_ps(_tmp2), _v576));
-                _tmp3 = _mm_cvttps_epi32(_mm_mul_ps(_mm_cvtepi32_ps(_tmp3), _v576));
-
-                if (out_elempack == 4)
-                {
-                    _mm_store_si128((__m128i*)outptr0, _tmp0);
-                    if (tj * 4 + 1 < outw) _mm_store_si128((__m128i*)(outptr0 + 4), _tmp1);
-                    if (tj * 4 + 2 < outw) _mm_store_si128((__m128i*)(outptr0 + 8), _tmp2);
-                    if (tj * 4 + 3 < outw) _mm_store_si128((__m128i*)(outptr0 + 12), _tmp3);
-                }
-                if (out_elempack == 1)
-                {
-#if __AVX512F__
-                    __m128i _vindex = _mm_setr_epi32(0, 1, 2, 3);
-                    _vindex = _mm_mullo_epi32(_vindex, _mm_set1_epi32(N));
-                    _mm_i32scatter_epi32(outptr0, _vindex, _tmp0, sizeof(int));
-                    if (tj * 4 + 1 < outw) _mm_i32scatter_epi32(outptr0 + 1, _vindex, _tmp1, sizeof(int));
-                    if (tj * 4 + 2 < outw) _mm_i32scatter_epi32(outptr0 + 2, _vindex, _tmp2, sizeof(int));
-                    if (tj * 4 + 3 < outw) _mm_i32scatter_epi32(outptr0 + 3, _vindex, _tmp3, sizeof(int));
-#else
-                    int tmp0[4];
-                    int tmp1[4];
-                    int tmp2[4];
-                    int tmp3[4];
-                    _mm_storeu_si128((__m128i*)tmp0, _tmp0);
-                    _mm_storeu_si128((__m128i*)tmp1, _tmp1);
-                    _mm_storeu_si128((__m128i*)tmp2, _tmp2);
-                    _mm_storeu_si128((__m128i*)tmp3, _tmp3);
-
-                    int* outptr1 = outptr0 + N;
-                    int* outptr2 = outptr0 + N * 2;
-                    int* outptr3 = outptr0 + N * 3;
-
-                    outptr0[0] = tmp0[0];
-                    outptr1[0] = tmp0[1];
-                    outptr2[0] = tmp0[2];
-                    outptr3[0] = tmp0[3];
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp1[0];
-                        outptr1[1] = tmp1[1];
-                        outptr2[1] = tmp1[2];
-                        outptr3[1] = tmp1[3];
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp2[0];
-                        outptr1[2] = tmp2[1];
-                        outptr2[2] = tmp2[2];
-                        outptr3[2] = tmp2[3];
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp3[0];
-                        outptr1[3] = tmp3[1];
-                        outptr2[3] = tmp3[2];
-                        outptr3[3] = tmp3[3];
-                    }
-#endif // __AVX512F__
-                }
-
-                outptr0 += outw * out_elempack;
-            }
-        }
-    }
-#endif // __SSE2__
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        int tmp[4][6][2];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 36 + jj * 2;
-            const int* r1 = r0 + max_jj * 2;
-            const int* r2 = r0 + max_jj * 2 * 2;
-            const int* r3 = r0 + max_jj * 2 * 3;
-            const int* r4 = r0 + max_jj * 2 * 4;
-            const int* r5 = r0 + max_jj * 2 * 5;
-
-            for (int m = 0; m < 5; m++)
-            {
-                int tmp02a0 = r1[0] + r2[0];
-                int tmp02a1 = r1[1] + r2[1];
-                int tmp02b0 = r3[0] + r4[0];
-                int tmp02b1 = r3[1] + r4[1];
-                int tmp13a0 = r1[0] - r2[0];
-                int tmp13a1 = r1[1] - r2[1];
-                int tmp13b0 = r3[0] - r4[0];
-                int tmp13b1 = r3[1] - r4[1];
-
-                int tmp00 = tmp02a0 + tmp02b0 + r0[0];
-                int tmp01 = tmp02a1 + tmp02b1 + r0[1];
-                int tmp10 = tmp13a0 + tmp13b0 * 2;
-                int tmp11 = tmp13a1 + tmp13b1 * 2;
-                int tmp20 = tmp02a0 + tmp02b0 * 4;
-                int tmp21 = tmp02a1 + tmp02b1 * 4;
-                int tmp30 = tmp13a0 + tmp13b0 * 8 + r5[0] * 4;
-                int tmp31 = tmp13a1 + tmp13b1 * 8 + r5[1] * 4;
-
-                tmp[0][m][0] = tmp00;
-                tmp[0][m][1] = tmp01;
-                tmp[1][m][0] = tmp10;
-                tmp[1][m][1] = tmp11;
-                tmp[2][m][0] = tmp20;
-                tmp[2][m][1] = tmp21;
-                tmp[3][m][0] = tmp30;
-                tmp[3][m][1] = tmp31;
-
-                r0 += max_jj * 6 * 2;
-                r1 += max_jj * 6 * 2;
-                r2 += max_jj * 6 * 2;
-                r3 += max_jj * 6 * 2;
-                r4 += max_jj * 6 * 2;
-                r5 += max_jj * 6 * 2;
-            }
-            for (int m = 5; m < 6; m++)
-            {
-                int tmp02a0 = r1[0] + r2[0];
-                int tmp02a1 = r1[1] + r2[1];
-                int tmp02b0 = r3[0] + r4[0];
-                int tmp02b1 = r3[1] + r4[1];
-                int tmp13a0 = r1[0] - r2[0];
-                int tmp13a1 = r1[1] - r2[1];
-                int tmp13b0 = r3[0] - r4[0];
-                int tmp13b1 = r3[1] - r4[1];
-
-                int tmp00 = tmp02a0 + tmp02b0 + r0[0];
-                int tmp01 = tmp02a1 + tmp02b1 + r0[1];
-                int tmp10 = tmp13a0 + tmp13b0 * 2;
-                int tmp11 = tmp13a1 + tmp13b1 * 2;
-                int tmp20 = tmp02a0 + tmp02b0 * 4;
-                int tmp21 = tmp02a1 + tmp02b1 * 4;
-                int tmp30 = tmp13a0 + tmp13b0 * 8 + r5[0] * 4;
-                int tmp31 = tmp13a1 + tmp13b1 * 8 + r5[1] * 4;
-
-                tmp00 = tmp00 * 4;
-                tmp01 = tmp01 * 4;
-                tmp10 = tmp10 * 4;
-                tmp11 = tmp11 * 4;
-                tmp20 = tmp20 * 4;
-                tmp21 = tmp21 * 4;
-                tmp30 = tmp30 * 4;
-                tmp31 = tmp31 * 4;
-
-                tmp[0][m][0] = tmp00;
-                tmp[0][m][1] = tmp01;
-                tmp[1][m][0] = tmp10;
-                tmp[1][m][1] = tmp11;
-                tmp[2][m][0] = tmp20;
-                tmp[2][m][1] = tmp21;
-                tmp[3][m][0] = tmp30;
-                tmp[3][m][1] = tmp31;
-
-                r0 += max_jj * 6 * 2;
-                r1 += max_jj * 6 * 2;
-                r2 += max_jj * 6 * 2;
-                r3 += max_jj * 6 * 2;
-                r4 += max_jj * 6 * 2;
-                r5 += max_jj * 6 * 2;
-            }
-
-            int* outptr0 = top_blob.channel(i + ii).row<int>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                int tmp02a0 = tmp[m][1][0] + tmp[m][2][0];
-                int tmp02a1 = tmp[m][1][1] + tmp[m][2][1];
-                int tmp02b0 = tmp[m][3][0] + tmp[m][4][0];
-                int tmp02b1 = tmp[m][3][1] + tmp[m][4][1];
-                int tmp13a0 = tmp[m][1][0] - tmp[m][2][0];
-                int tmp13a1 = tmp[m][1][1] - tmp[m][2][1];
-                int tmp13b0 = tmp[m][3][0] - tmp[m][4][0];
-                int tmp13b1 = tmp[m][3][1] - tmp[m][4][1];
-
-                int tmp00 = tmp02a0 + tmp02b0 + tmp[m][0][0];
-                int tmp01 = tmp02a1 + tmp02b1 + tmp[m][0][1];
-                int tmp10 = tmp13a0 + tmp13b0 * 2;
-                int tmp11 = tmp13a1 + tmp13b1 * 2;
-                int tmp20 = tmp02a0 + tmp02b0 * 4;
-                int tmp21 = tmp02a1 + tmp02b1 * 4;
-                int tmp30 = tmp13a0 + tmp13b0 * 8 + tmp[m][5][0];
-                int tmp31 = tmp13a1 + tmp13b1 * 8 + tmp[m][5][1];
-
-                tmp00 = tmp00 / 576;
-                tmp01 = tmp01 / 576;
-                tmp10 = tmp10 / 576;
-                tmp11 = tmp11 / 576;
-                tmp20 = tmp20 / 576;
-                tmp21 = tmp21 / 576;
-                tmp30 = tmp30 / 576;
-                tmp31 = tmp31 / 576;
-
-                // if (out_elempack == 1)
-                {
-                    int* outptr1 = outptr0 + N;
-
-                    outptr0[0] = tmp00;
-                    outptr1[0] = tmp01;
-                    if (tj * 4 + 1 < outw)
-                    {
-                        outptr0[1] = tmp10;
-                        outptr1[1] = tmp11;
-                    }
-                    if (tj * 4 + 2 < outw)
-                    {
-                        outptr0[2] = tmp20;
-                        outptr1[2] = tmp21;
-                    }
-                    if (tj * 4 + 3 < outw)
-                    {
-                        outptr0[3] = tmp30;
-                        outptr1[3] = tmp31;
-                    }
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-    for (; ii < max_ii; ii++)
-    {
-        int tmp[4][6];
-
-        int jj = 0;
-        for (; jj < max_jj; jj++)
-        {
-            int ti = (j + jj) / w_tiles;
-            int tj = (j + jj) % w_tiles;
-
-            const int* r0 = (const int*)top_tile + ii * max_jj * 36 + jj;
-            const int* r1 = r0 + max_jj;
-            const int* r2 = r0 + max_jj * 2;
-            const int* r3 = r0 + max_jj * 3;
-            const int* r4 = r0 + max_jj * 4;
-            const int* r5 = r0 + max_jj * 5;
-
-            for (int m = 0; m < 5; m++)
-            {
-                int tmp02a = r1[0] + r2[0];
-                int tmp02b = r3[0] + r4[0];
-                int tmp13a = r1[0] - r2[0];
-                int tmp13b = r3[0] - r4[0];
-
-                int tmp0 = tmp02a + tmp02b + r0[0];
-                int tmp1 = tmp13a + tmp13b * 2;
-                int tmp2 = tmp02a + tmp02b * 4;
-                int tmp3 = tmp13a + tmp13b * 8 + r5[0] * 4;
-
-                tmp[0][m] = tmp0;
-                tmp[1][m] = tmp1;
-                tmp[2][m] = tmp2;
-                tmp[3][m] = tmp3;
-
-                r0 += max_jj * 6;
-                r1 += max_jj * 6;
-                r2 += max_jj * 6;
-                r3 += max_jj * 6;
-                r4 += max_jj * 6;
-                r5 += max_jj * 6;
-            }
-            for (int m = 5; m < 6; m++)
-            {
-                int tmp02a = r1[0] + r2[0];
-                int tmp02b = r3[0] + r4[0];
-                int tmp13a = r1[0] - r2[0];
-                int tmp13b = r3[0] - r4[0];
-
-                int tmp0 = tmp02a + tmp02b + r0[0];
-                int tmp1 = tmp13a + tmp13b * 2;
-                int tmp2 = tmp02a + tmp02b * 4;
-                int tmp3 = tmp13a + tmp13b * 8 + r5[0] * 4;
-
-                tmp0 = tmp0 * 4;
-                tmp1 = tmp1 * 4;
-                tmp2 = tmp2 * 4;
-                tmp3 = tmp3 * 4;
-
-                tmp[0][m] = tmp0;
-                tmp[1][m] = tmp1;
-                tmp[2][m] = tmp2;
-                tmp[3][m] = tmp3;
-
-                r0 += max_jj * 6;
-                r1 += max_jj * 6;
-                r2 += max_jj * 6;
-                r3 += max_jj * 6;
-                r4 += max_jj * 6;
-                r5 += max_jj * 6;
-            }
-
-            int* outptr0 = top_blob.channel(i + ii).row<int>(ti * 4) + (tj * 4);
-
-            for (int m = 0; m < 4; m++)
-            {
-                if (ti * 4 + m >= outh)
-                    continue;
-
-                int tmp02a = tmp[m][1] + tmp[m][2];
-                int tmp02b = tmp[m][3] + tmp[m][4];
-                int tmp13a = tmp[m][1] - tmp[m][2];
-                int tmp13b = tmp[m][3] - tmp[m][4];
-
-                int tmp0 = tmp02a + tmp02b + tmp[m][0];
-                int tmp1 = tmp13a + tmp13b * 2;
-                int tmp2 = tmp02a + tmp02b * 4;
-                int tmp3 = tmp13a + tmp13b * 8 + tmp[m][5];
-
-                tmp0 = tmp0 / 576;
-                tmp1 = tmp1 / 576;
-                tmp2 = tmp2 / 576;
-                tmp3 = tmp3 / 576;
-
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = tmp0;
-                    if (tj * 4 + 1 < outw) outptr0[1] = tmp1;
-                    if (tj * 4 + 2 < outw) outptr0[2] = tmp2;
-                    if (tj * 4 + 3 < outw) outptr0[3] = tmp3;
-                }
-
-                outptr0 += outw;
-            }
-        }
-    }
-}
-
-static void conv3x3s1_winograd43_int8(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int nT, const Option& opt)
-{
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
-#if NCNN_RUNTIME_CPU && NCNN_AVX512VNNI && __AVX512F__ && !__AVX512VNNI__
-    if (ncnn::cpu_support_x86_avx512_vnni())
-    {
-        conv3x3s1_winograd43_int8_avx512vnni(bottom_blob, top_blob, AT, nT, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__
-    if (ncnn::cpu_support_x86_avx_vnni())
-    {
-        conv3x3s1_winograd43_int8_avxvnni(bottom_blob, top_blob, AT, nT, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-    if (ncnn::cpu_support_x86_avx2())
-    {
-        conv3x3s1_winograd43_int8_avx2(bottom_blob, top_blob, AT, nT, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__
-    if (ncnn::cpu_support_x86_xop())
-    {
-        conv3x3s1_winograd43_int8_xop(bottom_blob, top_blob, AT, nT, opt);
-        return;
-    }
-#endif
-#endif
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    // pad to 4n+2, winograd F(4,3)
-    int w_tiles = (outw + 3) / 4;
-    int h_tiles = (outh + 3) / 4;
-    int tiles = w_tiles * h_tiles;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = tiles;
-    const int K = bottom_blob.c * bottom_blob.elempack;
-    const int B = 36;
-
-    // NCNN_LOGE("conv3x3s1_winograd43_int8 %d %d %d", M, N, K);
-
-    int TILE_M, TILE_N, TILE_K;
-    get_optimal_tile_mnk_int8(M, N, K, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, B, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 4u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    if (nT > 1 && nn_NK < nT)
-    {
-        Mat B_tile(TILE_N * B * TILE_K, 4u, opt.workspace_allocator);
-
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            // transform input
-            conv3x3s1_winograd43_transform_input_tile_int8(bottom_blob, B_tile, j, max_jj, k, max_kk, nT);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            transpose_pack_B_tile_int8(B_tile, BT_tile, B, max_jj, max_kk, nT);
-        }
-    }
-    else
-    {
-        Mat B_tileX(TILE_N * B * TILE_K, 1, nT, 4u, opt.workspace_allocator);
-
-        #pragma omp parallel for num_threads(nT)
-        for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-        {
-            const int ppj = ppjk / nn_K;
-            const int ppk = ppjk % nn_K;
-
-            const int j = ppj * TILE_N;
-            const int k = ppk * TILE_K;
-
-            const int max_jj = std::min((N - j), TILE_N);
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat B_tile = B_tileX.channel(get_omp_thread_num());
-
-            // transform input
-            conv3x3s1_winograd43_transform_input_tile_int8(bottom_blob, B_tile, j, max_jj, k, max_kk, 1);
-
-            Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-            transpose_pack_B_tile_int8(B_tile, BT_tile, B, max_jj, max_kk, 1);
-        }
-    }
-
-    Mat top_tileX(TILE_N * B * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat top_tile = top_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).depth(k / TILE_K);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).depth(k / TILE_K);
-
-                bool k_end = k + TILE_K >= K;
-
-                gemm_transB_packed_tile_int8(AT_tile, BT_tile, top_tile, B, max_ii, max_jj, k, max_kk, k_end);
-            }
-
-            // transform output
-            conv3x3s1_winograd43_transform_output_tile_int8(top_tile, top_blob, i, max_ii, j, max_jj);
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_5x5.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_5x5.h
deleted file mode 100644
index bdbc2e10..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_5x5.h
+++ /dev/null
@@ -1,209 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void conv5x5s1_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        const float bias0 = bias ? bias[p] : 0.f;
-
-        out.fill(bias0);
-
-        for (int q = 0; q < inch; q++)
-        {
-            float* outptr = out;
-            float* outptr2 = outptr + outw;
-
-            const float* img0 = bottom_blob.channel(q);
-
-            const float* kernel0 = kernel + p * inch * 25 + q * 25;
-
-            const float* r0 = img0;
-            const float* r1 = img0 + w;
-            const float* r2 = img0 + w * 2;
-            const float* r3 = img0 + w * 3;
-            const float* r4 = img0 + w * 4;
-            const float* r5 = img0 + w * 5;
-
-            const float* k0 = kernel0;
-            const float* k1 = kernel0 + 5;
-            const float* k2 = kernel0 + 10;
-            const float* k3 = kernel0 + 15;
-            const float* k4 = kernel0 + 20;
-
-            int i = 0;
-
-            for (; i + 1 < outh; i += 2)
-            {
-                int remain = outw;
-
-                for (; remain > 0; remain--)
-                {
-                    float sum = 0;
-                    float sum2 = 0;
-
-                    sum += r0[0] * k0[0];
-                    sum += r0[1] * k0[1];
-                    sum += r0[2] * k0[2];
-                    sum += r0[3] * k0[3];
-                    sum += r0[4] * k0[4];
-
-                    sum += r1[0] * k1[0];
-                    sum += r1[1] * k1[1];
-                    sum += r1[2] * k1[2];
-                    sum += r1[3] * k1[3];
-                    sum += r1[4] * k1[4];
-
-                    sum += r2[0] * k2[0];
-                    sum += r2[1] * k2[1];
-                    sum += r2[2] * k2[2];
-                    sum += r2[3] * k2[3];
-                    sum += r2[4] * k2[4];
-
-                    sum += r3[0] * k3[0];
-                    sum += r3[1] * k3[1];
-                    sum += r3[2] * k3[2];
-                    sum += r3[3] * k3[3];
-                    sum += r3[4] * k3[4];
-
-                    sum += r4[0] * k4[0];
-                    sum += r4[1] * k4[1];
-                    sum += r4[2] * k4[2];
-                    sum += r4[3] * k4[3];
-                    sum += r4[4] * k4[4];
-
-                    sum2 += r1[0] * k0[0];
-                    sum2 += r1[1] * k0[1];
-                    sum2 += r1[2] * k0[2];
-                    sum2 += r1[3] * k0[3];
-                    sum2 += r1[4] * k0[4];
-
-                    sum2 += r2[0] * k1[0];
-                    sum2 += r2[1] * k1[1];
-                    sum2 += r2[2] * k1[2];
-                    sum2 += r2[3] * k1[3];
-                    sum2 += r2[4] * k1[4];
-
-                    sum2 += r3[0] * k2[0];
-                    sum2 += r3[1] * k2[1];
-                    sum2 += r3[2] * k2[2];
-                    sum2 += r3[3] * k2[3];
-                    sum2 += r3[4] * k2[4];
-
-                    sum2 += r4[0] * k3[0];
-                    sum2 += r4[1] * k3[1];
-                    sum2 += r4[2] * k3[2];
-                    sum2 += r4[3] * k3[3];
-                    sum2 += r4[4] * k3[4];
-
-                    sum2 += r5[0] * k4[0];
-                    sum2 += r5[1] * k4[1];
-                    sum2 += r5[2] * k4[2];
-                    sum2 += r5[3] * k4[3];
-                    sum2 += r5[4] * k4[4];
-
-                    *outptr += sum;
-                    *outptr2 += sum2;
-
-                    r0++;
-                    r1++;
-                    r2++;
-                    r3++;
-                    r4++;
-                    r5++;
-                    outptr++;
-                    outptr2++;
-                }
-
-                r0 += 4 + w;
-                r1 += 4 + w;
-                r2 += 4 + w;
-                r3 += 4 + w;
-                r4 += 4 + w;
-                r5 += 4 + w;
-
-                outptr += outw;
-                outptr2 += outw;
-            }
-
-            for (; i < outh; i++)
-            {
-                int remain = outw;
-
-                for (; remain > 0; remain--)
-                {
-                    float sum = 0;
-
-                    sum += r0[0] * k0[0];
-                    sum += r0[1] * k0[1];
-                    sum += r0[2] * k0[2];
-                    sum += r0[3] * k0[3];
-                    sum += r0[4] * k0[4];
-
-                    sum += r1[0] * k1[0];
-                    sum += r1[1] * k1[1];
-                    sum += r1[2] * k1[2];
-                    sum += r1[3] * k1[3];
-                    sum += r1[4] * k1[4];
-
-                    sum += r2[0] * k2[0];
-                    sum += r2[1] * k2[1];
-                    sum += r2[2] * k2[2];
-                    sum += r2[3] * k2[3];
-                    sum += r2[4] * k2[4];
-
-                    sum += r3[0] * k3[0];
-                    sum += r3[1] * k3[1];
-                    sum += r3[2] * k3[2];
-                    sum += r3[3] * k3[3];
-                    sum += r3[4] * k3[4];
-
-                    sum += r4[0] * k4[0];
-                    sum += r4[1] * k4[1];
-                    sum += r4[2] * k4[2];
-                    sum += r4[3] * k4[3];
-                    sum += r4[4] * k4[4];
-
-                    *outptr += sum;
-
-                    r0++;
-                    r1++;
-                    r2++;
-                    r3++;
-                    r4++;
-                    outptr++;
-                }
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                r3 += 4;
-                r4 += 4;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_im2col_gemm_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_im2col_gemm_int8.h
deleted file mode 100644
index ef8fe0ab..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_im2col_gemm_int8.h
+++ /dev/null
@@ -1,8895 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
-#if NCNN_RUNTIME_CPU && NCNN_AVX512VNNI && __AVX512F__ && !__AVX512VNNI__
-void convolution_im2col_gemm_int8_avx512vnni(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt);
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__
-void convolution_im2col_gemm_int8_avxvnni(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt);
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-void convolution_im2col_gemm_transform_kernel_int8_avx2(const Mat& kernel, Mat& AT, int inch, int outch, int kernel_w, int kernel_h, const Option& opt);
-void convolution_im2col_gemm_int8_avx2(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt);
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__
-void convolution_im2col_gemm_int8_xop(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt);
-#endif
-#endif
-
-static void convolution_im2col_pack_A_tile_int8(const Mat& A, Mat& AT, int i, int max_ii, int k, int max_kk)
-{
-    // A = (pa, maxk, inch/pa), outch
-    const int A_hstep = A.w;
-
-    signed char* pp = AT;
-
-    int ii = 0;
-#if __SSE2__
-#if __AVX2__
-#if __AVX512F__
-    for (; ii + 15 < max_ii; ii += 16)
-    {
-        const signed char* p0 = (const signed char*)A + (i + ii) * A_hstep + k;
-        const signed char* p1 = (const signed char*)A + (i + ii + 1) * A_hstep + k;
-        const signed char* p2 = (const signed char*)A + (i + ii + 2) * A_hstep + k;
-        const signed char* p3 = (const signed char*)A + (i + ii + 3) * A_hstep + k;
-        const signed char* p4 = (const signed char*)A + (i + ii + 4) * A_hstep + k;
-        const signed char* p5 = (const signed char*)A + (i + ii + 5) * A_hstep + k;
-        const signed char* p6 = (const signed char*)A + (i + ii + 6) * A_hstep + k;
-        const signed char* p7 = (const signed char*)A + (i + ii + 7) * A_hstep + k;
-        const signed char* p8 = (const signed char*)A + (i + ii + 8) * A_hstep + k;
-        const signed char* p9 = (const signed char*)A + (i + ii + 9) * A_hstep + k;
-        const signed char* pa = (const signed char*)A + (i + ii + 10) * A_hstep + k;
-        const signed char* pb = (const signed char*)A + (i + ii + 11) * A_hstep + k;
-        const signed char* pc = (const signed char*)A + (i + ii + 12) * A_hstep + k;
-        const signed char* pd = (const signed char*)A + (i + ii + 13) * A_hstep + k;
-        const signed char* pe = (const signed char*)A + (i + ii + 14) * A_hstep + k;
-        const signed char* pf = (const signed char*)A + (i + ii + 15) * A_hstep + k;
-
-        int kk = 0;
-#if defined(__x86_64__) || defined(_M_X64)
-        for (; kk + 15 < max_kk; kk += 16)
-        {
-            __m128i _r0 = _mm_loadu_si128((const __m128i*)p0);
-            __m128i _r1 = _mm_loadu_si128((const __m128i*)p1);
-            __m128i _r2 = _mm_loadu_si128((const __m128i*)p2);
-            __m128i _r3 = _mm_loadu_si128((const __m128i*)p3);
-            __m128i _r4 = _mm_loadu_si128((const __m128i*)p4);
-            __m128i _r5 = _mm_loadu_si128((const __m128i*)p5);
-            __m128i _r6 = _mm_loadu_si128((const __m128i*)p6);
-            __m128i _r7 = _mm_loadu_si128((const __m128i*)p7);
-            __m128i _r8 = _mm_loadu_si128((const __m128i*)p8);
-            __m128i _r9 = _mm_loadu_si128((const __m128i*)p9);
-            __m128i _ra = _mm_loadu_si128((const __m128i*)pa);
-            __m128i _rb = _mm_loadu_si128((const __m128i*)pb);
-            __m128i _rc = _mm_loadu_si128((const __m128i*)pc);
-            __m128i _rd = _mm_loadu_si128((const __m128i*)pd);
-            __m128i _re = _mm_loadu_si128((const __m128i*)pe);
-            __m128i _rf = _mm_loadu_si128((const __m128i*)pf);
-            __m128i _t0 = _mm_unpacklo_epi16(_r0, _r1);
-            __m128i _t1 = _mm_unpackhi_epi16(_r0, _r1);
-            __m128i _t2 = _mm_unpacklo_epi16(_r2, _r3);
-            __m128i _t3 = _mm_unpackhi_epi16(_r2, _r3);
-            __m128i _t4 = _mm_unpacklo_epi16(_r4, _r5);
-            __m128i _t5 = _mm_unpackhi_epi16(_r4, _r5);
-            __m128i _t6 = _mm_unpacklo_epi16(_r6, _r7);
-            __m128i _t7 = _mm_unpackhi_epi16(_r6, _r7);
-            __m128i _t8 = _mm_unpacklo_epi16(_r8, _r9);
-            __m128i _t9 = _mm_unpackhi_epi16(_r8, _r9);
-            __m128i _ta = _mm_unpacklo_epi16(_ra, _rb);
-            __m128i _tb = _mm_unpackhi_epi16(_ra, _rb);
-            __m128i _tc = _mm_unpacklo_epi16(_rc, _rd);
-            __m128i _td = _mm_unpackhi_epi16(_rc, _rd);
-            __m128i _te = _mm_unpacklo_epi16(_re, _rf);
-            __m128i _tf = _mm_unpackhi_epi16(_re, _rf);
-            _r0 = _mm_unpacklo_epi32(_t0, _t2);
-            _r1 = _mm_unpackhi_epi32(_t0, _t2);
-            _r2 = _mm_unpacklo_epi32(_t4, _t6);
-            _r3 = _mm_unpackhi_epi32(_t4, _t6);
-            _r4 = _mm_unpacklo_epi32(_t8, _ta);
-            _r5 = _mm_unpackhi_epi32(_t8, _ta);
-            _r6 = _mm_unpacklo_epi32(_tc, _te);
-            _r7 = _mm_unpackhi_epi32(_tc, _te);
-            _r8 = _mm_unpacklo_epi32(_t1, _t3);
-            _r9 = _mm_unpackhi_epi32(_t1, _t3);
-            _ra = _mm_unpacklo_epi32(_t5, _t7);
-            _rb = _mm_unpackhi_epi32(_t5, _t7);
-            _rc = _mm_unpacklo_epi32(_t9, _tb);
-            _rd = _mm_unpackhi_epi32(_t9, _tb);
-            _re = _mm_unpacklo_epi32(_td, _tf);
-            _rf = _mm_unpackhi_epi32(_td, _tf);
-            _t0 = _mm_unpacklo_epi64(_r0, _r2);
-            _t1 = _mm_unpackhi_epi64(_r0, _r2);
-            _t2 = _mm_unpacklo_epi64(_r1, _r3);
-            _t3 = _mm_unpackhi_epi64(_r1, _r3);
-            _t4 = _mm_unpacklo_epi64(_r4, _r6);
-            _t5 = _mm_unpackhi_epi64(_r4, _r6);
-            _t6 = _mm_unpacklo_epi64(_r5, _r7);
-            _t7 = _mm_unpackhi_epi64(_r5, _r7);
-            _t8 = _mm_unpacklo_epi64(_r8, _ra);
-            _t9 = _mm_unpackhi_epi64(_r8, _ra);
-            _ta = _mm_unpacklo_epi64(_r9, _rb);
-            _tb = _mm_unpackhi_epi64(_r9, _rb);
-            _tc = _mm_unpacklo_epi64(_rc, _re);
-            _td = _mm_unpackhi_epi64(_rc, _re);
-            _te = _mm_unpacklo_epi64(_rd, _rf);
-            _tf = _mm_unpackhi_epi64(_rd, _rf);
-            _mm_storeu_si128((__m128i*)pp, _t0);
-            _mm_storeu_si128((__m128i*)(pp + 16), _t4);
-            _mm_storeu_si128((__m128i*)(pp + 32), _t1);
-            _mm_storeu_si128((__m128i*)(pp + 48), _t5);
-            _mm_storeu_si128((__m128i*)(pp + 64), _t2);
-            _mm_storeu_si128((__m128i*)(pp + 80), _t6);
-            _mm_storeu_si128((__m128i*)(pp + 96), _t3);
-            _mm_storeu_si128((__m128i*)(pp + 112), _t7);
-            _mm_storeu_si128((__m128i*)(pp + 128), _t8);
-            _mm_storeu_si128((__m128i*)(pp + 144), _tc);
-            _mm_storeu_si128((__m128i*)(pp + 160), _t9);
-            _mm_storeu_si128((__m128i*)(pp + 176), _td);
-            _mm_storeu_si128((__m128i*)(pp + 192), _ta);
-            _mm_storeu_si128((__m128i*)(pp + 208), _te);
-            _mm_storeu_si128((__m128i*)(pp + 224), _tb);
-            _mm_storeu_si128((__m128i*)(pp + 240), _tf);
-            pp += 256;
-            p0 += 16;
-            p1 += 16;
-            p2 += 16;
-            p3 += 16;
-            p4 += 16;
-            p5 += 16;
-            p6 += 16;
-            p7 += 16;
-            p8 += 16;
-            p9 += 16;
-            pa += 16;
-            pb += 16;
-            pc += 16;
-            pd += 16;
-            pe += 16;
-            pf += 16;
-        }
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            __m128i _r0 = _mm_loadl_epi64((const __m128i*)p0);
-            __m128i _r1 = _mm_loadl_epi64((const __m128i*)p1);
-            __m128i _r2 = _mm_loadl_epi64((const __m128i*)p2);
-            __m128i _r3 = _mm_loadl_epi64((const __m128i*)p3);
-            __m128i _r4 = _mm_loadl_epi64((const __m128i*)p4);
-            __m128i _r5 = _mm_loadl_epi64((const __m128i*)p5);
-            __m128i _r6 = _mm_loadl_epi64((const __m128i*)p6);
-            __m128i _r7 = _mm_loadl_epi64((const __m128i*)p7);
-            __m128i _r8 = _mm_loadl_epi64((const __m128i*)p8);
-            __m128i _r9 = _mm_loadl_epi64((const __m128i*)p9);
-            __m128i _ra = _mm_loadl_epi64((const __m128i*)pa);
-            __m128i _rb = _mm_loadl_epi64((const __m128i*)pb);
-            __m128i _rc = _mm_loadl_epi64((const __m128i*)pc);
-            __m128i _rd = _mm_loadl_epi64((const __m128i*)pd);
-            __m128i _re = _mm_loadl_epi64((const __m128i*)pe);
-            __m128i _rf = _mm_loadl_epi64((const __m128i*)pf);
-            __m128i _t0 = _mm_unpacklo_epi16(_r0, _r1);
-            __m128i _t1 = _mm_unpacklo_epi16(_r2, _r3);
-            __m128i _t2 = _mm_unpacklo_epi16(_r4, _r5);
-            __m128i _t3 = _mm_unpacklo_epi16(_r6, _r7);
-            __m128i _t4 = _mm_unpacklo_epi16(_r8, _r9);
-            __m128i _t5 = _mm_unpacklo_epi16(_ra, _rb);
-            __m128i _t6 = _mm_unpacklo_epi16(_rc, _rd);
-            __m128i _t7 = _mm_unpacklo_epi16(_re, _rf);
-            _r0 = _mm_unpacklo_epi32(_t0, _t1);
-            _r1 = _mm_unpackhi_epi32(_t0, _t1);
-            _r2 = _mm_unpacklo_epi32(_t2, _t3);
-            _r3 = _mm_unpackhi_epi32(_t2, _t3);
-            _r4 = _mm_unpacklo_epi32(_t4, _t5);
-            _r5 = _mm_unpackhi_epi32(_t4, _t5);
-            _r6 = _mm_unpacklo_epi32(_t6, _t7);
-            _r7 = _mm_unpackhi_epi32(_t6, _t7);
-            _t0 = _mm_unpacklo_epi64(_r0, _r2);
-            _t1 = _mm_unpackhi_epi64(_r0, _r2);
-            _t2 = _mm_unpacklo_epi64(_r1, _r3);
-            _t3 = _mm_unpackhi_epi64(_r1, _r3);
-            _t4 = _mm_unpacklo_epi64(_r4, _r6);
-            _t5 = _mm_unpackhi_epi64(_r4, _r6);
-            _t6 = _mm_unpacklo_epi64(_r5, _r7);
-            _t7 = _mm_unpackhi_epi64(_r5, _r7);
-            _mm_storeu_si128((__m128i*)pp, _t0);
-            _mm_storeu_si128((__m128i*)(pp + 16), _t4);
-            _mm_storeu_si128((__m128i*)(pp + 32), _t1);
-            _mm_storeu_si128((__m128i*)(pp + 48), _t5);
-            _mm_storeu_si128((__m128i*)(pp + 64), _t2);
-            _mm_storeu_si128((__m128i*)(pp + 80), _t6);
-            _mm_storeu_si128((__m128i*)(pp + 96), _t3);
-            _mm_storeu_si128((__m128i*)(pp + 112), _t7);
-            pp += 128;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-            p4 += 8;
-            p5 += 8;
-            p6 += 8;
-            p7 += 8;
-            p8 += 8;
-            p9 += 8;
-            pa += 8;
-            pb += 8;
-            pc += 8;
-            pd += 8;
-            pe += 8;
-            pf += 8;
-        }
-#endif // defined(__x86_64__) || defined(_M_X64)
-        for (; kk + 1 < max_kk; kk += 2)
-        {
-            pp[0] = p0[0];
-            pp[1] = p0[1];
-            pp[2] = p1[0];
-            pp[3] = p1[1];
-            pp[4] = p2[0];
-            pp[5] = p2[1];
-            pp[6] = p3[0];
-            pp[7] = p3[1];
-            pp[8] = p4[0];
-            pp[9] = p4[1];
-            pp[10] = p5[0];
-            pp[11] = p5[1];
-            pp[12] = p6[0];
-            pp[13] = p6[1];
-            pp[14] = p7[0];
-            pp[15] = p7[1];
-            pp[16] = p8[0];
-            pp[17] = p8[1];
-            pp[18] = p9[0];
-            pp[19] = p9[1];
-            pp[20] = pa[0];
-            pp[21] = pa[1];
-            pp[22] = pb[0];
-            pp[23] = pb[1];
-            pp[24] = pc[0];
-            pp[25] = pc[1];
-            pp[26] = pd[0];
-            pp[27] = pd[1];
-            pp[28] = pe[0];
-            pp[29] = pe[1];
-            pp[30] = pf[0];
-            pp[31] = pf[1];
-            pp += 32;
-            p0 += 2;
-            p1 += 2;
-            p2 += 2;
-            p3 += 2;
-            p4 += 2;
-            p5 += 2;
-            p6 += 2;
-            p7 += 2;
-            p8 += 2;
-            p9 += 2;
-            pa += 2;
-            pb += 2;
-            pc += 2;
-            pd += 2;
-            pe += 2;
-            pf += 2;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp[1] = p1[0];
-            pp[2] = p2[0];
-            pp[3] = p3[0];
-            pp[4] = p4[0];
-            pp[5] = p5[0];
-            pp[6] = p6[0];
-            pp[7] = p7[0];
-            pp[8] = p8[0];
-            pp[9] = p9[0];
-            pp[10] = pa[0];
-            pp[11] = pb[0];
-            pp[12] = pc[0];
-            pp[13] = pd[0];
-            pp[14] = pe[0];
-            pp[15] = pf[0];
-            pp += 16;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-            p4++;
-            p5++;
-            p6++;
-            p7++;
-            p8++;
-            p9++;
-            pa++;
-            pb++;
-            pc++;
-            pd++;
-            pe++;
-            pf++;
-        }
-    }
-#endif // __AVX512F__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        const signed char* p0 = (const signed char*)A + (i + ii) * A_hstep + k;
-        const signed char* p1 = (const signed char*)A + (i + ii + 1) * A_hstep + k;
-        const signed char* p2 = (const signed char*)A + (i + ii + 2) * A_hstep + k;
-        const signed char* p3 = (const signed char*)A + (i + ii + 3) * A_hstep + k;
-        const signed char* p4 = (const signed char*)A + (i + ii + 4) * A_hstep + k;
-        const signed char* p5 = (const signed char*)A + (i + ii + 5) * A_hstep + k;
-        const signed char* p6 = (const signed char*)A + (i + ii + 6) * A_hstep + k;
-        const signed char* p7 = (const signed char*)A + (i + ii + 7) * A_hstep + k;
-
-        int kk = 0;
-#if defined(__x86_64__) || defined(_M_X64)
-        for (; kk + 15 < max_kk; kk += 16)
-        {
-            __m128i _r0 = _mm_loadu_si128((const __m128i*)p0);
-            __m128i _r1 = _mm_loadu_si128((const __m128i*)p1);
-            __m128i _r2 = _mm_loadu_si128((const __m128i*)p2);
-            __m128i _r3 = _mm_loadu_si128((const __m128i*)p3);
-            __m128i _r4 = _mm_loadu_si128((const __m128i*)p4);
-            __m128i _r5 = _mm_loadu_si128((const __m128i*)p5);
-            __m128i _r6 = _mm_loadu_si128((const __m128i*)p6);
-            __m128i _r7 = _mm_loadu_si128((const __m128i*)p7);
-            __m128i _t0 = _mm_unpacklo_epi16(_r0, _r1);
-            __m128i _t1 = _mm_unpackhi_epi16(_r0, _r1);
-            __m128i _t2 = _mm_unpacklo_epi16(_r2, _r3);
-            __m128i _t3 = _mm_unpackhi_epi16(_r2, _r3);
-            __m128i _t4 = _mm_unpacklo_epi16(_r4, _r5);
-            __m128i _t5 = _mm_unpackhi_epi16(_r4, _r5);
-            __m128i _t6 = _mm_unpacklo_epi16(_r6, _r7);
-            __m128i _t7 = _mm_unpackhi_epi16(_r6, _r7);
-            _r0 = _mm_unpacklo_epi32(_t0, _t2);
-            _r1 = _mm_unpackhi_epi32(_t0, _t2);
-            _r2 = _mm_unpacklo_epi32(_t4, _t6);
-            _r3 = _mm_unpackhi_epi32(_t4, _t6);
-            _r4 = _mm_unpacklo_epi32(_t1, _t3);
-            _r5 = _mm_unpackhi_epi32(_t1, _t3);
-            _r6 = _mm_unpacklo_epi32(_t5, _t7);
-            _r7 = _mm_unpackhi_epi32(_t5, _t7);
-            _t0 = _mm_unpacklo_epi64(_r0, _r2);
-            _t1 = _mm_unpackhi_epi64(_r0, _r2);
-            _t2 = _mm_unpacklo_epi64(_r1, _r3);
-            _t3 = _mm_unpackhi_epi64(_r1, _r3);
-            _t4 = _mm_unpacklo_epi64(_r4, _r6);
-            _t5 = _mm_unpackhi_epi64(_r4, _r6);
-            _t6 = _mm_unpacklo_epi64(_r5, _r7);
-            _t7 = _mm_unpackhi_epi64(_r5, _r7);
-            _mm_storeu_si128((__m128i*)pp, _t0);
-            _mm_storeu_si128((__m128i*)(pp + 16), _t1);
-            _mm_storeu_si128((__m128i*)(pp + 32), _t2);
-            _mm_storeu_si128((__m128i*)(pp + 48), _t3);
-            _mm_storeu_si128((__m128i*)(pp + 64), _t4);
-            _mm_storeu_si128((__m128i*)(pp + 80), _t5);
-            _mm_storeu_si128((__m128i*)(pp + 96), _t6);
-            _mm_storeu_si128((__m128i*)(pp + 112), _t7);
-            pp += 128;
-            p0 += 16;
-            p1 += 16;
-            p2 += 16;
-            p3 += 16;
-            p4 += 16;
-            p5 += 16;
-            p6 += 16;
-            p7 += 16;
-        }
-#endif // defined(__x86_64__) || defined(_M_X64)
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            __m128i _r0 = _mm_loadl_epi64((const __m128i*)p0);
-            __m128i _r1 = _mm_loadl_epi64((const __m128i*)p1);
-            __m128i _r2 = _mm_loadl_epi64((const __m128i*)p2);
-            __m128i _r3 = _mm_loadl_epi64((const __m128i*)p3);
-            __m128i _r4 = _mm_loadl_epi64((const __m128i*)p4);
-            __m128i _r5 = _mm_loadl_epi64((const __m128i*)p5);
-            __m128i _r6 = _mm_loadl_epi64((const __m128i*)p6);
-            __m128i _r7 = _mm_loadl_epi64((const __m128i*)p7);
-            __m128i _t0 = _mm_unpacklo_epi16(_r0, _r1);
-            __m128i _t1 = _mm_unpacklo_epi16(_r2, _r3);
-            __m128i _t2 = _mm_unpacklo_epi16(_r4, _r5);
-            __m128i _t3 = _mm_unpacklo_epi16(_r6, _r7);
-            _r0 = _mm_unpacklo_epi32(_t0, _t1);
-            _r1 = _mm_unpackhi_epi32(_t0, _t1);
-            _r2 = _mm_unpacklo_epi32(_t2, _t3);
-            _r3 = _mm_unpackhi_epi32(_t2, _t3);
-            _r4 = _mm_unpacklo_epi64(_r0, _r2);
-            _r5 = _mm_unpackhi_epi64(_r0, _r2);
-            _r6 = _mm_unpacklo_epi64(_r1, _r3);
-            _r7 = _mm_unpackhi_epi64(_r1, _r3);
-            _mm_storeu_si128((__m128i*)pp, _r4);
-            _mm_storeu_si128((__m128i*)(pp + 16), _r5);
-            _mm_storeu_si128((__m128i*)(pp + 32), _r6);
-            _mm_storeu_si128((__m128i*)(pp + 48), _r7);
-            pp += 64;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-            p4 += 8;
-            p5 += 8;
-            p6 += 8;
-            p7 += 8;
-        }
-        for (; kk + 1 < max_kk; kk += 2)
-        {
-            pp[0] = p0[0];
-            pp[1] = p0[1];
-            pp[2] = p1[0];
-            pp[3] = p1[1];
-            pp[4] = p2[0];
-            pp[5] = p2[1];
-            pp[6] = p3[0];
-            pp[7] = p3[1];
-            pp[8] = p4[0];
-            pp[9] = p4[1];
-            pp[10] = p5[0];
-            pp[11] = p5[1];
-            pp[12] = p6[0];
-            pp[13] = p6[1];
-            pp[14] = p7[0];
-            pp[15] = p7[1];
-            pp += 16;
-            p0 += 2;
-            p1 += 2;
-            p2 += 2;
-            p3 += 2;
-            p4 += 2;
-            p5 += 2;
-            p6 += 2;
-            p7 += 2;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp[1] = p1[0];
-            pp[2] = p2[0];
-            pp[3] = p3[0];
-            pp[4] = p4[0];
-            pp[5] = p5[0];
-            pp[6] = p6[0];
-            pp[7] = p7[0];
-            pp += 8;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-            p4++;
-            p5++;
-            p6++;
-            p7++;
-        }
-    }
-#endif // __AVX2__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        const signed char* p0 = (const signed char*)A + (i + ii) * A_hstep + k;
-        const signed char* p1 = (const signed char*)A + (i + ii + 1) * A_hstep + k;
-        const signed char* p2 = (const signed char*)A + (i + ii + 2) * A_hstep + k;
-        const signed char* p3 = (const signed char*)A + (i + ii + 3) * A_hstep + k;
-
-        int kk = 0;
-        for (; kk + 15 < max_kk; kk += 16)
-        {
-            __m128i _r0 = _mm_loadu_si128((const __m128i*)p0);
-            __m128i _r1 = _mm_loadu_si128((const __m128i*)p1);
-            __m128i _r2 = _mm_loadu_si128((const __m128i*)p2);
-            __m128i _r3 = _mm_loadu_si128((const __m128i*)p3);
-            __m128i _t0 = _mm_unpacklo_epi16(_r0, _r1);
-            __m128i _t1 = _mm_unpackhi_epi16(_r0, _r1);
-            __m128i _t2 = _mm_unpacklo_epi16(_r2, _r3);
-            __m128i _t3 = _mm_unpackhi_epi16(_r2, _r3);
-            _r0 = _mm_unpacklo_epi32(_t0, _t2);
-            _r1 = _mm_unpackhi_epi32(_t0, _t2);
-            _r2 = _mm_unpacklo_epi32(_t1, _t3);
-            _r3 = _mm_unpackhi_epi32(_t1, _t3);
-            _mm_storeu_si128((__m128i*)pp, _r0);
-            _mm_storeu_si128((__m128i*)(pp + 16), _r1);
-            _mm_storeu_si128((__m128i*)(pp + 32), _r2);
-            _mm_storeu_si128((__m128i*)(pp + 48), _r3);
-            pp += 64;
-            p0 += 16;
-            p1 += 16;
-            p2 += 16;
-            p3 += 16;
-        }
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            __m128i _r0 = _mm_loadl_epi64((const __m128i*)p0);
-            __m128i _r1 = _mm_loadl_epi64((const __m128i*)p1);
-            __m128i _r2 = _mm_loadl_epi64((const __m128i*)p2);
-            __m128i _r3 = _mm_loadl_epi64((const __m128i*)p3);
-            __m128i _t0 = _mm_unpacklo_epi16(_r0, _r1);
-            __m128i _t1 = _mm_unpacklo_epi16(_r2, _r3);
-            _r0 = _mm_unpacklo_epi32(_t0, _t1);
-            _r1 = _mm_unpackhi_epi32(_t0, _t1);
-            _mm_storeu_si128((__m128i*)pp, _r0);
-            _mm_storeu_si128((__m128i*)(pp + 16), _r1);
-            pp += 32;
-            p0 += 8;
-            p1 += 8;
-            p2 += 8;
-            p3 += 8;
-        }
-        for (; kk + 1 < max_kk; kk += 2)
-        {
-            pp[0] = p0[0];
-            pp[1] = p0[1];
-            pp[2] = p1[0];
-            pp[3] = p1[1];
-            pp[4] = p2[0];
-            pp[5] = p2[1];
-            pp[6] = p3[0];
-            pp[7] = p3[1];
-            pp += 8;
-            p0 += 2;
-            p1 += 2;
-            p2 += 2;
-            p3 += 2;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp[1] = p1[0];
-            pp[2] = p2[0];
-            pp[3] = p3[0];
-            pp += 4;
-            p0++;
-            p1++;
-            p2++;
-            p3++;
-        }
-    }
-#endif // __SSE2__
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        const signed char* p0 = (const signed char*)A + (i + ii) * A_hstep + k;
-        const signed char* p1 = (const signed char*)A + (i + ii + 1) * A_hstep + k;
-
-        int kk = 0;
-#if __SSE2__
-        for (; kk + 15 < max_kk; kk += 16)
-        {
-            __m128i _r0 = _mm_loadu_si128((const __m128i*)p0);
-            __m128i _r1 = _mm_loadu_si128((const __m128i*)p1);
-            __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-            __m128i _r23 = _mm_unpackhi_epi16(_r0, _r1);
-            _mm_storeu_si128((__m128i*)pp, _r01);
-            _mm_storeu_si128((__m128i*)(pp + 16), _r23);
-            pp += 32;
-            p0 += 16;
-            p1 += 16;
-        }
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            __m128i _r0 = _mm_loadl_epi64((const __m128i*)p0);
-            __m128i _r1 = _mm_loadl_epi64((const __m128i*)p1);
-            __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-            _mm_storeu_si128((__m128i*)pp, _r01);
-            pp += 16;
-            p0 += 8;
-            p1 += 8;
-        }
-#endif // __SSE2__
-        for (; kk + 1 < max_kk; kk += 2)
-        {
-            pp[0] = p0[0];
-            pp[1] = p0[1];
-            pp[2] = p1[0];
-            pp[3] = p1[1];
-            pp += 4;
-            p0 += 2;
-            p1 += 2;
-        }
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp[1] = p1[0];
-            pp += 2;
-            p0++;
-            p1++;
-        }
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        const signed char* p0 = (const signed char*)A + (i + ii) * A_hstep + k;
-
-        int kk = 0;
-#if __SSE2__
-        for (; kk + 15 < max_kk; kk += 15)
-        {
-            _mm_storeu_si128((__m128i*)pp, _mm_loadu_si128((const __m128i*)p0));
-            pp += 16;
-            p0 += 16;
-        }
-        for (; kk + 7 < max_kk; kk += 8)
-        {
-            _mm_storel_epi64((__m128i*)pp, _mm_loadl_epi64((const __m128i*)p0));
-            pp += 8;
-            p0 += 8;
-        }
-#endif // __SSE2__
-        for (; kk < max_kk; kk++)
-        {
-            pp[0] = p0[0];
-            pp += 1;
-            p0++;
-        }
-    }
-}
-
-static void convolution_gemm_transB_packed_tile_int8(const Mat& AT_tile, const Mat& BT_tile, Mat& topT_tile, Mat& top_blob, int i, int max_ii, int j, int max_jj, int k, int max_kk, bool k_end)
-{
-    // NCNN_LOGE("convolution_gemm_transB_packed_tile_int8 %d %d %d %d %d %d", i, max_ii, j, max_jj, k, max_kk);
-
-    const int out_elempack = top_blob.elempack;
-    const int out_hstep = (int)top_blob.cstep;
-
-    const signed char* pAT = AT_tile;
-    const signed char* pBT = BT_tile;
-
-    int* outptr = topT_tile;
-
-    int ii = 0;
-#if __SSE2__
-#if __AVX2__
-#if __AVX512F__
-    for (; ii + 15 < max_ii; ii += 16)
-    {
-        int* outptr0 = (int*)top_blob + (i + ii) * out_hstep + j * out_elempack;
-
-        const signed char* pB = pBT;
-
-        int jj = 0;
-#if defined(__x86_64__) || defined(_M_X64)
-        for (; jj + 15 < max_jj; jj += 16)
-        {
-            const signed char* pA = pAT;
-
-            __m512i _sum0;
-            __m512i _sum1;
-            __m512i _sum2;
-            __m512i _sum3;
-            __m512i _sum4;
-            __m512i _sum5;
-            __m512i _sum6;
-            __m512i _sum7;
-            __m512i _sum8;
-            __m512i _sum9;
-            __m512i _suma;
-            __m512i _sumb;
-            __m512i _sumc;
-            __m512i _sumd;
-            __m512i _sume;
-            __m512i _sumf;
-
-            if (k == 0)
-            {
-                _sum0 = _mm512_setzero_si512();
-                _sum1 = _mm512_setzero_si512();
-                _sum2 = _mm512_setzero_si512();
-                _sum3 = _mm512_setzero_si512();
-                _sum4 = _mm512_setzero_si512();
-                _sum5 = _mm512_setzero_si512();
-                _sum6 = _mm512_setzero_si512();
-                _sum7 = _mm512_setzero_si512();
-                _sum8 = _mm512_setzero_si512();
-                _sum9 = _mm512_setzero_si512();
-                _suma = _mm512_setzero_si512();
-                _sumb = _mm512_setzero_si512();
-                _sumc = _mm512_setzero_si512();
-                _sumd = _mm512_setzero_si512();
-                _sume = _mm512_setzero_si512();
-                _sumf = _mm512_setzero_si512();
-            }
-            else
-            {
-                _sum0 = _mm512_load_si512((const __m512i*)outptr);
-                _sum1 = _mm512_load_si512((const __m512i*)(outptr + 16));
-                _sum2 = _mm512_load_si512((const __m512i*)(outptr + 32));
-                _sum3 = _mm512_load_si512((const __m512i*)(outptr + 48));
-                _sum4 = _mm512_load_si512((const __m512i*)(outptr + 64));
-                _sum5 = _mm512_load_si512((const __m512i*)(outptr + 80));
-                _sum6 = _mm512_load_si512((const __m512i*)(outptr + 96));
-                _sum7 = _mm512_load_si512((const __m512i*)(outptr + 112));
-                _sum8 = _mm512_load_si512((const __m512i*)(outptr + 128));
-                _sum9 = _mm512_load_si512((const __m512i*)(outptr + 128 + 16));
-                _suma = _mm512_load_si512((const __m512i*)(outptr + 128 + 32));
-                _sumb = _mm512_load_si512((const __m512i*)(outptr + 128 + 48));
-                _sumc = _mm512_load_si512((const __m512i*)(outptr + 128 + 64));
-                _sumd = _mm512_load_si512((const __m512i*)(outptr + 128 + 80));
-                _sume = _mm512_load_si512((const __m512i*)(outptr + 128 + 96));
-                _sumf = _mm512_load_si512((const __m512i*)(outptr + 128 + 112));
-            }
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m256i _pA = _mm256_loadu_si256((const __m256i*)pA);
-                __m256i _pB = _mm256_loadu_si256((const __m256i*)pB);
-
-                __m512i _pA0 = _mm512_cvtepi8_epi16(_pA);
-                __m512i _pB0 = _mm512_cvtepi8_epi16(_pB);
-
-                // 0123 4567 89ab cdef
-                // cdef 0123 4567 89ab
-                // 89ab cdef 0123 4567
-                // 4567 89ab cdef 0123
-                __m512i _pA1 = _mm512_shuffle_i32x4(_pA0, _pA0, _MM_SHUFFLE(2, 1, 0, 3));
-                __m512i _pA2 = _mm512_shuffle_i32x4(_pA0, _pA0, _MM_SHUFFLE(1, 0, 3, 2));
-                __m512i _pA3 = _mm512_shuffle_i32x4(_pA0, _pA0, _MM_SHUFFLE(0, 3, 2, 1));
-
-                // 0123 4567 89ab cdef
-                // 1230 5674 9ab8 defc
-                // 2301 6745 ab89 efcd
-                // 3012 7456 b89a fcde
-                __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-                __m512i _pB2 = _mm512_shuffle_epi32(_pB0, _MM_PERM_BADC);
-                __m512i _pB3 = _mm512_shuffle_epi32(_pB0, _MM_PERM_CBAD);
-
-#if __AVX512VNNI__
-                _sum0 = _mm512_dpwssd_epi32(_sum0, _pA0, _pB0);
-                _sum1 = _mm512_dpwssd_epi32(_sum1, _pA0, _pB1);
-                _sum2 = _mm512_dpwssd_epi32(_sum2, _pA0, _pB2);
-                _sum3 = _mm512_dpwssd_epi32(_sum3, _pA0, _pB3);
-                _sum4 = _mm512_dpwssd_epi32(_sum4, _pA1, _pB0);
-                _sum5 = _mm512_dpwssd_epi32(_sum5, _pA1, _pB1);
-                _sum6 = _mm512_dpwssd_epi32(_sum6, _pA1, _pB2);
-                _sum7 = _mm512_dpwssd_epi32(_sum7, _pA1, _pB3);
-                _sum8 = _mm512_dpwssd_epi32(_sum8, _pA2, _pB0);
-                _sum9 = _mm512_dpwssd_epi32(_sum9, _pA2, _pB1);
-                _suma = _mm512_dpwssd_epi32(_suma, _pA2, _pB2);
-                _sumb = _mm512_dpwssd_epi32(_sumb, _pA2, _pB3);
-                _sumc = _mm512_dpwssd_epi32(_sumc, _pA3, _pB0);
-                _sumd = _mm512_dpwssd_epi32(_sumd, _pA3, _pB1);
-                _sume = _mm512_dpwssd_epi32(_sume, _pA3, _pB2);
-                _sumf = _mm512_dpwssd_epi32(_sumf, _pA3, _pB3);
-#else
-                _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA0, _pB0));
-                _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA0, _pB1));
-                _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_pA0, _pB2));
-                _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_pA0, _pB3));
-                _sum4 = _mm512_add_epi32(_sum4, _mm512_madd_epi16(_pA1, _pB0));
-                _sum5 = _mm512_add_epi32(_sum5, _mm512_madd_epi16(_pA1, _pB1));
-                _sum6 = _mm512_add_epi32(_sum6, _mm512_madd_epi16(_pA1, _pB2));
-                _sum7 = _mm512_add_epi32(_sum7, _mm512_madd_epi16(_pA1, _pB3));
-                _sum8 = _mm512_add_epi32(_sum8, _mm512_madd_epi16(_pA2, _pB0));
-                _sum9 = _mm512_add_epi32(_sum9, _mm512_madd_epi16(_pA2, _pB1));
-                _suma = _mm512_add_epi32(_suma, _mm512_madd_epi16(_pA2, _pB2));
-                _sumb = _mm512_add_epi32(_sumb, _mm512_madd_epi16(_pA2, _pB3));
-                _sumc = _mm512_add_epi32(_sumc, _mm512_madd_epi16(_pA3, _pB0));
-                _sumd = _mm512_add_epi32(_sumd, _mm512_madd_epi16(_pA3, _pB1));
-                _sume = _mm512_add_epi32(_sume, _mm512_madd_epi16(_pA3, _pB2));
-                _sumf = _mm512_add_epi32(_sumf, _mm512_madd_epi16(_pA3, _pB3));
-#endif
-
-                pA += 32;
-                pB += 32;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_load_si128((const __m128i*)pA);
-                __m128i _pB = _mm_load_si128((const __m128i*)pB);
-
-                __m256i _pA0 = _mm256_cvtepi8_epi16(_pA);
-                __m256i _pB0 = _mm256_cvtepi8_epi16(_pB);
-
-                // 01234567 89abcdef
-                // cdef0123 456789ab
-                // 89abcdef 01234567
-                // 456789ab cdef0123
-                __m256i _pA1 = _mm256_permute4x64_epi64(_pA0, _MM_SHUFFLE(2, 1, 0, 3));
-                __m256i _pA2 = _mm256_permute4x64_epi64(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-                __m256i _pA3 = _mm256_permute4x64_epi64(_pA0, _MM_SHUFFLE(0, 3, 2, 1));
-
-                // 01234567 89abcdef
-                // 12305674 9ab8defc
-                // 23016745 ab89efcd
-                // 30127456 b89afcde
-                __m256i _pB1 = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(_pB0, _MM_SHUFFLE(0, 3, 2, 1)), _MM_SHUFFLE(0, 3, 2, 1));
-                __m256i _pB2 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(2, 3, 0, 1));
-                __m256i _pB3 = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(_pB0, _MM_SHUFFLE(2, 1, 0, 3)), _MM_SHUFFLE(2, 1, 0, 3));
-
-                __m512i _s0 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB0));
-                __m512i _s1 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB1));
-                __m512i _s2 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB2));
-                __m512i _s3 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB3));
-                __m512i _s4 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA1, _pB0));
-                __m512i _s5 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA1, _pB1));
-                __m512i _s6 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA1, _pB2));
-                __m512i _s7 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA1, _pB3));
-                __m512i _s8 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA2, _pB0));
-                __m512i _s9 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA2, _pB1));
-                __m512i _sa = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA2, _pB2));
-                __m512i _sb = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA2, _pB3));
-                __m512i _sc = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA3, _pB0));
-                __m512i _sd = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA3, _pB1));
-                __m512i _se = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA3, _pB2));
-                __m512i _sf = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA3, _pB3));
-
-                _sum0 = _mm512_add_epi32(_sum0, _s0);
-                _sum1 = _mm512_add_epi32(_sum1, _s1);
-                _sum2 = _mm512_add_epi32(_sum2, _s2);
-                _sum3 = _mm512_add_epi32(_sum3, _s3);
-                _sum4 = _mm512_add_epi32(_sum4, _s4);
-                _sum5 = _mm512_add_epi32(_sum5, _s5);
-                _sum6 = _mm512_add_epi32(_sum6, _s6);
-                _sum7 = _mm512_add_epi32(_sum7, _s7);
-                _sum8 = _mm512_add_epi32(_sum8, _s8);
-                _sum9 = _mm512_add_epi32(_sum9, _s9);
-                _suma = _mm512_add_epi32(_suma, _sa);
-                _sumb = _mm512_add_epi32(_sumb, _sb);
-                _sumc = _mm512_add_epi32(_sumc, _sc);
-                _sumd = _mm512_add_epi32(_sumd, _sd);
-                _sume = _mm512_add_epi32(_sume, _se);
-                _sumf = _mm512_add_epi32(_sumf, _sf);
-
-                pA += 16;
-                pB += 16;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 16)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77 88 99 aa bb cc dd ee ff
-                    //      01 12 23 30 45 56 67 74 89 9a ab b8 cd de ef fc
-                    //      02 13 20 31 46 57 64 75 8a 9b a8 b9 ce df ec fd
-                    //      03 10 21 32 47 54 65 76 8b 98 a9 ba cf dc ed fe
-                    //      c0 d1 e2 f3 04 15 26 37 48 59 6a 7b 8c 9d ae bf
-                    //      c1 d2 e3 f0 05 16 27 34 49 5a 6b 78 8d 9e af bc
-                    //      c2 d3 e0 f1 06 17 24 35 4a 5b 68 79 8e 9f ac bd
-                    //      c3 d0 e1 f2 07 14 25 36 4b 58 69 7a 8f 9c ad be
-                    //      80 91 a2 b3 c4 d5 e6 f7 08 19 2a 3b 4c 5d 6e 7f
-                    //      81 92 a3 b0 c5 d6 e7 f4 09 1a 2b 38 4d 5e 6f 7c
-                    //      82 93 a0 b1 c6 d7 e4 f5 0a 1b 28 39 4e 5f 6c 7d
-                    //      83 90 a1 b2 c7 d4 e5 f6 0b 18 29 3a 4f 5c 6d 7e
-                    //      40 51 62 73 84 95 a6 b7 c8 d9 ea fb 0c 1d 2e 3f
-                    //      41 52 63 70 85 96 a7 b4 c9 da eb f8 0d 1e 2f 3c
-                    //      42 53 60 71 86 97 a4 b5 ca db e8 f9 0e 1f 2c 3d
-                    //      43 50 61 72 87 94 a5 b6 cb d8 e9 fa 0f 1c 2d 3e
-                    // to
-                    //      00 10 20 30 44 54 64 74 88 98 a8 b8 cc dc ec fc
-                    //      01 11 21 31 45 55 65 75 89 99 a9 b9 cd dd ed fd
-                    //      02 12 22 32 46 56 66 76 8a 9a aa ba ce de ee fe
-                    //      03 13 23 33 47 57 67 77 8b 9b ab bb cf df ef ff
-                    //      c0 d0 e0 f0 04 14 24 34 48 58 68 78 8c 9c ac bc
-                    //      c1 d1 e1 f1 05 15 25 35 49 59 69 79 8d 9d ad bd
-                    //      c2 d2 e2 f2 06 16 26 36 4a 5a 6a 7a 8e 9e ae be
-                    //      c3 d3 e3 f3 07 17 27 37 4b 5b 6b 7b 8f 9f af bf
-                    //      80 90 a0 b0 c4 d4 e4 f4 08 18 28 38 4c 5c 6c 7c
-                    //      81 91 a1 b1 c5 d5 e5 f5 09 19 29 39 4d 5d 6d 7d
-                    //      82 92 a2 b2 c6 d6 e6 f6 0a 1a 2a 3a 4e 5e 6e 7e
-                    //      83 93 a3 b3 c7 d7 e7 f7 0b 1b 2b 3b 4f 5f 6f 7f
-                    //      40 50 60 70 84 94 a4 b4 c8 d8 e8 f8 0c 1c 2c 3c
-                    //      41 51 61 71 85 95 a5 b5 c9 d9 e9 f9 0d 1d 2d 3d
-                    //      42 52 62 72 86 96 a6 b6 ca da ea fa 0e 1e 2e 3e
-                    //      43 53 63 73 87 97 a7 b7 cb db eb fb 0f 1f 2f 3f
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum2 = _mm512_shuffle_epi32(_sum2, _MM_PERM_BADC);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_ADCB);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum6 = _mm512_shuffle_epi32(_sum6, _MM_PERM_BADC);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_ADCB);
-                        _sum9 = _mm512_shuffle_epi32(_sum9, _MM_PERM_CBAD);
-                        _suma = _mm512_shuffle_epi32(_suma, _MM_PERM_BADC);
-                        _sumb = _mm512_shuffle_epi32(_sumb, _MM_PERM_ADCB);
-                        _sumd = _mm512_shuffle_epi32(_sumd, _MM_PERM_CBAD);
-                        _sume = _mm512_shuffle_epi32(_sume, _MM_PERM_BADC);
-                        _sumf = _mm512_shuffle_epi32(_sumf, _MM_PERM_ADCB);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        __m512i _tmp4 = _mm512_unpacklo_epi32(_sum4, _sum7);
-                        __m512i _tmp5 = _mm512_unpackhi_epi32(_sum4, _sum7);
-                        __m512i _tmp6 = _mm512_unpacklo_epi32(_sum6, _sum5);
-                        __m512i _tmp7 = _mm512_unpackhi_epi32(_sum6, _sum5);
-                        __m512i _tmp8 = _mm512_unpacklo_epi32(_sum8, _sumb);
-                        __m512i _tmp9 = _mm512_unpackhi_epi32(_sum8, _sumb);
-                        __m512i _tmpa = _mm512_unpacklo_epi32(_suma, _sum9);
-                        __m512i _tmpb = _mm512_unpackhi_epi32(_suma, _sum9);
-                        __m512i _tmpc = _mm512_unpacklo_epi32(_sumc, _sumf);
-                        __m512i _tmpd = _mm512_unpackhi_epi32(_sumc, _sumf);
-                        __m512i _tmpe = _mm512_unpacklo_epi32(_sume, _sumd);
-                        __m512i _tmpf = _mm512_unpackhi_epi32(_sume, _sumd);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm512_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm512_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm512_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm512_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum8 = _mm512_unpacklo_epi64(_tmp8, _tmpa);
-                        _sum9 = _mm512_unpackhi_epi64(_tmp8, _tmpa);
-                        _suma = _mm512_unpacklo_epi64(_tmpb, _tmp9);
-                        _sumb = _mm512_unpackhi_epi64(_tmpb, _tmp9);
-                        _sumc = _mm512_unpacklo_epi64(_tmpc, _tmpe);
-                        _sumd = _mm512_unpackhi_epi64(_tmpc, _tmpe);
-                        _sume = _mm512_unpacklo_epi64(_tmpf, _tmpd);
-                        _sumf = _mm512_unpackhi_epi64(_tmpf, _tmpd);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_CBAD);
-                        _sum9 = _mm512_shuffle_epi32(_sum9, _MM_PERM_CBAD);
-                        _sumb = _mm512_shuffle_epi32(_sumb, _MM_PERM_CBAD);
-                        _sumd = _mm512_shuffle_epi32(_sumd, _MM_PERM_CBAD);
-                        _sumf = _mm512_shuffle_epi32(_sumf, _MM_PERM_CBAD);
-                    }
-
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sumc, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum4, _sum0, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum8, _sum4, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sumc, _sum8, _MM_SHUFFLE(1, 3, 1, 3));
-                    __m512i _tmp4 = _mm512_shuffle_i32x4(_sum1, _sumd, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp5 = _mm512_shuffle_i32x4(_sum5, _sum1, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp6 = _mm512_shuffle_i32x4(_sum9, _sum5, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmp7 = _mm512_shuffle_i32x4(_sumd, _sum9, _MM_SHUFFLE(1, 3, 1, 3));
-                    __m512i _tmp8 = _mm512_shuffle_i32x4(_sum2, _sume, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp9 = _mm512_shuffle_i32x4(_sum6, _sum2, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmpa = _mm512_shuffle_i32x4(_suma, _sum6, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmpb = _mm512_shuffle_i32x4(_sume, _suma, _MM_SHUFFLE(1, 3, 1, 3));
-                    __m512i _tmpc = _mm512_shuffle_i32x4(_sum3, _sumf, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmpd = _mm512_shuffle_i32x4(_sum7, _sum3, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmpe = _mm512_shuffle_i32x4(_sumb, _sum7, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmpf = _mm512_shuffle_i32x4(_sumf, _sumb, _MM_SHUFFLE(1, 3, 1, 3));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp2, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp4, _tmp6, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp8, _tmpa, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum3 = _mm512_shuffle_i32x4(_tmpc, _tmpe, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum4 = _mm512_shuffle_i32x4(_tmp1, _tmp3, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum5 = _mm512_shuffle_i32x4(_tmp5, _tmp7, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum6 = _mm512_shuffle_i32x4(_tmp9, _tmpb, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum7 = _mm512_shuffle_i32x4(_tmpd, _tmpf, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum8 = _mm512_shuffle_i32x4(_tmp2, _tmp0, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum9 = _mm512_shuffle_i32x4(_tmp6, _tmp4, _MM_SHUFFLE(3, 1, 2, 0));
-                    _suma = _mm512_shuffle_i32x4(_tmpa, _tmp8, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sumb = _mm512_shuffle_i32x4(_tmpe, _tmpc, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sumc = _mm512_shuffle_i32x4(_tmp3, _tmp1, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sumd = _mm512_shuffle_i32x4(_tmp7, _tmp5, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sume = _mm512_shuffle_i32x4(_tmpb, _tmp9, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sumf = _mm512_shuffle_i32x4(_tmpf, _tmpd, _MM_SHUFFLE(3, 1, 2, 0));
-
-                    _mm512_store_si512((__m512i*)outptr0, _sum0);
-                    _mm512_store_si512((__m512i*)(outptr0 + 16), _sum1);
-                    _mm512_store_si512((__m512i*)(outptr0 + 32), _sum2);
-                    _mm512_store_si512((__m512i*)(outptr0 + 48), _sum3);
-                    _mm512_store_si512((__m512i*)(outptr0 + 64), _sum4);
-                    _mm512_store_si512((__m512i*)(outptr0 + 80), _sum5);
-                    _mm512_store_si512((__m512i*)(outptr0 + 96), _sum6);
-                    _mm512_store_si512((__m512i*)(outptr0 + 112), _sum7);
-                    _mm512_store_si512((__m512i*)(outptr0 + 128), _sum8);
-                    _mm512_store_si512((__m512i*)(outptr0 + 128 + 16), _sum9);
-                    _mm512_store_si512((__m512i*)(outptr0 + 128 + 32), _suma);
-                    _mm512_store_si512((__m512i*)(outptr0 + 128 + 48), _sumb);
-                    _mm512_store_si512((__m512i*)(outptr0 + 128 + 64), _sumc);
-                    _mm512_store_si512((__m512i*)(outptr0 + 128 + 80), _sumd);
-                    _mm512_store_si512((__m512i*)(outptr0 + 128 + 96), _sume);
-                    _mm512_store_si512((__m512i*)(outptr0 + 128 + 112), _sumf);
-                    outptr0 += 256;
-                }
-                if (out_elempack == 8)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77 88 99 aa bb cc dd ee ff
-                    //      01 12 23 30 45 56 67 74 89 9a ab b8 cd de ef fc
-                    //      02 13 20 31 46 57 64 75 8a 9b a8 b9 ce df ec fd
-                    //      03 10 21 32 47 54 65 76 8b 98 a9 ba cf dc ed fe
-                    //      c0 d1 e2 f3 04 15 26 37 48 59 6a 7b 8c 9d ae bf
-                    //      c1 d2 e3 f0 05 16 27 34 49 5a 6b 78 8d 9e af bc
-                    //      c2 d3 e0 f1 06 17 24 35 4a 5b 68 79 8e 9f ac bd
-                    //      c3 d0 e1 f2 07 14 25 36 4b 58 69 7a 8f 9c ad be
-                    //      80 91 a2 b3 c4 d5 e6 f7 08 19 2a 3b 4c 5d 6e 7f
-                    //      81 92 a3 b0 c5 d6 e7 f4 09 1a 2b 38 4d 5e 6f 7c
-                    //      82 93 a0 b1 c6 d7 e4 f5 0a 1b 28 39 4e 5f 6c 7d
-                    //      83 90 a1 b2 c7 d4 e5 f6 0b 18 29 3a 4f 5c 6d 7e
-                    //      40 51 62 73 84 95 a6 b7 c8 d9 ea fb 0c 1d 2e 3f
-                    //      41 52 63 70 85 96 a7 b4 c9 da eb f8 0d 1e 2f 3c
-                    //      42 53 60 71 86 97 a4 b5 ca db e8 f9 0e 1f 2c 3d
-                    //      43 50 61 72 87 94 a5 b6 cb d8 e9 fa 0f 1c 2d 3e
-                    // to
-                    //      00 10 20 30 44 54 64 74 88 98 a8 b8 cc dc ec fc
-                    //      01 11 21 31 45 55 65 75 89 99 a9 b9 cd dd ed fd
-                    //      02 12 22 32 46 56 66 76 8a 9a aa ba ce de ee fe
-                    //      03 13 23 33 47 57 67 77 8b 9b ab bb cf df ef ff
-                    //      c0 d0 e0 f0 04 14 24 34 48 58 68 78 8c 9c ac bc
-                    //      c1 d1 e1 f1 05 15 25 35 49 59 69 79 8d 9d ad bd
-                    //      c2 d2 e2 f2 06 16 26 36 4a 5a 6a 7a 8e 9e ae be
-                    //      c3 d3 e3 f3 07 17 27 37 4b 5b 6b 7b 8f 9f af bf
-                    //      80 90 a0 b0 c4 d4 e4 f4 08 18 28 38 4c 5c 6c 7c
-                    //      81 91 a1 b1 c5 d5 e5 f5 09 19 29 39 4d 5d 6d 7d
-                    //      82 92 a2 b2 c6 d6 e6 f6 0a 1a 2a 3a 4e 5e 6e 7e
-                    //      83 93 a3 b3 c7 d7 e7 f7 0b 1b 2b 3b 4f 5f 6f 7f
-                    //      40 50 60 70 84 94 a4 b4 c8 d8 e8 f8 0c 1c 2c 3c
-                    //      41 51 61 71 85 95 a5 b5 c9 d9 e9 f9 0d 1d 2d 3d
-                    //      42 52 62 72 86 96 a6 b6 ca da ea fa 0e 1e 2e 3e
-                    //      43 53 63 73 87 97 a7 b7 cb db eb fb 0f 1f 2f 3f
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum2 = _mm512_shuffle_epi32(_sum2, _MM_PERM_BADC);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_ADCB);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum6 = _mm512_shuffle_epi32(_sum6, _MM_PERM_BADC);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_ADCB);
-                        _sum9 = _mm512_shuffle_epi32(_sum9, _MM_PERM_CBAD);
-                        _suma = _mm512_shuffle_epi32(_suma, _MM_PERM_BADC);
-                        _sumb = _mm512_shuffle_epi32(_sumb, _MM_PERM_ADCB);
-                        _sumd = _mm512_shuffle_epi32(_sumd, _MM_PERM_CBAD);
-                        _sume = _mm512_shuffle_epi32(_sume, _MM_PERM_BADC);
-                        _sumf = _mm512_shuffle_epi32(_sumf, _MM_PERM_ADCB);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        __m512i _tmp4 = _mm512_unpacklo_epi32(_sum4, _sum7);
-                        __m512i _tmp5 = _mm512_unpackhi_epi32(_sum4, _sum7);
-                        __m512i _tmp6 = _mm512_unpacklo_epi32(_sum6, _sum5);
-                        __m512i _tmp7 = _mm512_unpackhi_epi32(_sum6, _sum5);
-                        __m512i _tmp8 = _mm512_unpacklo_epi32(_sum8, _sumb);
-                        __m512i _tmp9 = _mm512_unpackhi_epi32(_sum8, _sumb);
-                        __m512i _tmpa = _mm512_unpacklo_epi32(_suma, _sum9);
-                        __m512i _tmpb = _mm512_unpackhi_epi32(_suma, _sum9);
-                        __m512i _tmpc = _mm512_unpacklo_epi32(_sumc, _sumf);
-                        __m512i _tmpd = _mm512_unpackhi_epi32(_sumc, _sumf);
-                        __m512i _tmpe = _mm512_unpacklo_epi32(_sume, _sumd);
-                        __m512i _tmpf = _mm512_unpackhi_epi32(_sume, _sumd);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm512_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm512_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm512_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm512_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum8 = _mm512_unpacklo_epi64(_tmp8, _tmpa);
-                        _sum9 = _mm512_unpackhi_epi64(_tmp8, _tmpa);
-                        _suma = _mm512_unpacklo_epi64(_tmpb, _tmp9);
-                        _sumb = _mm512_unpackhi_epi64(_tmpb, _tmp9);
-                        _sumc = _mm512_unpacklo_epi64(_tmpc, _tmpe);
-                        _sumd = _mm512_unpackhi_epi64(_tmpc, _tmpe);
-                        _sume = _mm512_unpacklo_epi64(_tmpf, _tmpd);
-                        _sumf = _mm512_unpackhi_epi64(_tmpf, _tmpd);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_CBAD);
-                        _sum9 = _mm512_shuffle_epi32(_sum9, _MM_PERM_CBAD);
-                        _sumb = _mm512_shuffle_epi32(_sumb, _MM_PERM_CBAD);
-                        _sumd = _mm512_shuffle_epi32(_sumd, _MM_PERM_CBAD);
-                        _sumf = _mm512_shuffle_epi32(_sumf, _MM_PERM_CBAD);
-                    }
-
-                    // TODO
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sumc, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum4, _sum0, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum8, _sum4, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sumc, _sum8, _MM_SHUFFLE(1, 3, 1, 3));
-                    __m512i _tmp4 = _mm512_shuffle_i32x4(_sum1, _sumd, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp5 = _mm512_shuffle_i32x4(_sum5, _sum1, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp6 = _mm512_shuffle_i32x4(_sum9, _sum5, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmp7 = _mm512_shuffle_i32x4(_sumd, _sum9, _MM_SHUFFLE(1, 3, 1, 3));
-                    __m512i _tmp8 = _mm512_shuffle_i32x4(_sum2, _sume, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp9 = _mm512_shuffle_i32x4(_sum6, _sum2, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmpa = _mm512_shuffle_i32x4(_suma, _sum6, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmpb = _mm512_shuffle_i32x4(_sume, _suma, _MM_SHUFFLE(1, 3, 1, 3));
-                    __m512i _tmpc = _mm512_shuffle_i32x4(_sum3, _sumf, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmpd = _mm512_shuffle_i32x4(_sum7, _sum3, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmpe = _mm512_shuffle_i32x4(_sumb, _sum7, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmpf = _mm512_shuffle_i32x4(_sumf, _sumb, _MM_SHUFFLE(1, 3, 1, 3));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp4, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp8, _tmpc, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp1, _tmp5, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp9, _tmpd, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum4 = _mm512_shuffle_i32x4(_tmp2, _tmp6, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum5 = _mm512_shuffle_i32x4(_tmpa, _tmpe, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum6 = _mm512_shuffle_i32x4(_tmp3, _tmp7, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum7 = _mm512_shuffle_i32x4(_tmpb, _tmpf, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum8 = _mm512_shuffle_i32x4(_tmp2, _tmp6, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum9 = _mm512_shuffle_i32x4(_tmpa, _tmpe, _MM_SHUFFLE(3, 1, 3, 1));
-                    _suma = _mm512_shuffle_i32x4(_tmp3, _tmp7, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sumb = _mm512_shuffle_i32x4(_tmpb, _tmpf, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sumc = _mm512_shuffle_i32x4(_tmp0, _tmp4, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sumd = _mm512_shuffle_i32x4(_tmp8, _tmpc, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sume = _mm512_shuffle_i32x4(_tmp1, _tmp5, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sumf = _mm512_shuffle_i32x4(_tmp9, _tmpc, _MM_SHUFFLE(3, 1, 3, 1));
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _sum0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16), _sum1);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16 * 2), _sum2);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16 * 3), _sum3);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16 * 4), _sum4);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16 * 5), _sum5);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16 * 6), _sum6);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16 * 7), _sum7);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8), _sum8);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8 + 16), _sum9);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8 + 16 * 2), _suma);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8 + 16 * 3), _sumb);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8 + 16 * 4), _sumc);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8 + 16 * 5), _sumd);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8 + 16 * 6), _sume);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8 + 16 * 7), _sumf);
-                    outptr0 += 128;
-                }
-                if (out_elempack == 4)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77 88 99 aa bb cc dd ee ff
-                    //      01 12 23 30 45 56 67 74 89 9a ab b8 cd de ef fc
-                    //      02 13 20 31 46 57 64 75 8a 9b a8 b9 ce df ec fd
-                    //      03 10 21 32 47 54 65 76 8b 98 a9 ba cf dc ed fe
-                    //      c0 d1 e2 f3 04 15 26 37 48 59 6a 7b 8c 9d ae bf
-                    //      c1 d2 e3 f0 05 16 27 34 49 5a 6b 78 8d 9e af bc
-                    //      c2 d3 e0 f1 06 17 24 35 4a 5b 68 79 8e 9f ac bd
-                    //      c3 d0 e1 f2 07 14 25 36 4b 58 69 7a 8f 9c ad be
-                    //      80 91 a2 b3 c4 d5 e6 f7 08 19 2a 3b 4c 5d 6e 7f
-                    //      81 92 a3 b0 c5 d6 e7 f4 09 1a 2b 38 4d 5e 6f 7c
-                    //      82 93 a0 b1 c6 d7 e4 f5 0a 1b 28 39 4e 5f 6c 7d
-                    //      83 90 a1 b2 c7 d4 e5 f6 0b 18 29 3a 4f 5c 6d 7e
-                    //      40 51 62 73 84 95 a6 b7 c8 d9 ea fb 0c 1d 2e 3f
-                    //      41 52 63 70 85 96 a7 b4 c9 da eb f8 0d 1e 2f 3c
-                    //      42 53 60 71 86 97 a4 b5 ca db e8 f9 0e 1f 2c 3d
-                    //      43 50 61 72 87 94 a5 b6 cb d8 e9 fa 0f 1c 2d 3e
-                    // to
-                    //      00 10 20 30 44 54 64 74 88 98 a8 b8 cc dc ec fc
-                    //      01 11 21 31 45 55 65 75 89 99 a9 b9 cd dd ed fd
-                    //      02 12 22 32 46 56 66 76 8a 9a aa ba ce de ee fe
-                    //      03 13 23 33 47 57 67 77 8b 9b ab bb cf df ef ff
-                    //      c0 d0 e0 f0 04 14 24 34 48 58 68 78 8c 9c ac bc
-                    //      c1 d1 e1 f1 05 15 25 35 49 59 69 79 8d 9d ad bd
-                    //      c2 d2 e2 f2 06 16 26 36 4a 5a 6a 7a 8e 9e ae be
-                    //      c3 d3 e3 f3 07 17 27 37 4b 5b 6b 7b 8f 9f af bf
-                    //      80 90 a0 b0 c4 d4 e4 f4 08 18 28 38 4c 5c 6c 7c
-                    //      81 91 a1 b1 c5 d5 e5 f5 09 19 29 39 4d 5d 6d 7d
-                    //      82 92 a2 b2 c6 d6 e6 f6 0a 1a 2a 3a 4e 5e 6e 7e
-                    //      83 93 a3 b3 c7 d7 e7 f7 0b 1b 2b 3b 4f 5f 6f 7f
-                    //      40 50 60 70 84 94 a4 b4 c8 d8 e8 f8 0c 1c 2c 3c
-                    //      41 51 61 71 85 95 a5 b5 c9 d9 e9 f9 0d 1d 2d 3d
-                    //      42 52 62 72 86 96 a6 b6 ca da ea fa 0e 1e 2e 3e
-                    //      43 53 63 73 87 97 a7 b7 cb db eb fb 0f 1f 2f 3f
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum2 = _mm512_shuffle_epi32(_sum2, _MM_PERM_BADC);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_ADCB);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum6 = _mm512_shuffle_epi32(_sum6, _MM_PERM_BADC);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_ADCB);
-                        _sum9 = _mm512_shuffle_epi32(_sum9, _MM_PERM_CBAD);
-                        _suma = _mm512_shuffle_epi32(_suma, _MM_PERM_BADC);
-                        _sumb = _mm512_shuffle_epi32(_sumb, _MM_PERM_ADCB);
-                        _sumd = _mm512_shuffle_epi32(_sumd, _MM_PERM_CBAD);
-                        _sume = _mm512_shuffle_epi32(_sume, _MM_PERM_BADC);
-                        _sumf = _mm512_shuffle_epi32(_sumf, _MM_PERM_ADCB);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        __m512i _tmp4 = _mm512_unpacklo_epi32(_sum4, _sum7);
-                        __m512i _tmp5 = _mm512_unpackhi_epi32(_sum4, _sum7);
-                        __m512i _tmp6 = _mm512_unpacklo_epi32(_sum6, _sum5);
-                        __m512i _tmp7 = _mm512_unpackhi_epi32(_sum6, _sum5);
-                        __m512i _tmp8 = _mm512_unpacklo_epi32(_sum8, _sumb);
-                        __m512i _tmp9 = _mm512_unpackhi_epi32(_sum8, _sumb);
-                        __m512i _tmpa = _mm512_unpacklo_epi32(_suma, _sum9);
-                        __m512i _tmpb = _mm512_unpackhi_epi32(_suma, _sum9);
-                        __m512i _tmpc = _mm512_unpacklo_epi32(_sumc, _sumf);
-                        __m512i _tmpd = _mm512_unpackhi_epi32(_sumc, _sumf);
-                        __m512i _tmpe = _mm512_unpacklo_epi32(_sume, _sumd);
-                        __m512i _tmpf = _mm512_unpackhi_epi32(_sume, _sumd);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm512_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm512_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm512_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm512_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum8 = _mm512_unpacklo_epi64(_tmp8, _tmpa);
-                        _sum9 = _mm512_unpackhi_epi64(_tmp8, _tmpa);
-                        _suma = _mm512_unpacklo_epi64(_tmpb, _tmp9);
-                        _sumb = _mm512_unpackhi_epi64(_tmpb, _tmp9);
-                        _sumc = _mm512_unpacklo_epi64(_tmpc, _tmpe);
-                        _sumd = _mm512_unpackhi_epi64(_tmpc, _tmpe);
-                        _sume = _mm512_unpacklo_epi64(_tmpf, _tmpd);
-                        _sumf = _mm512_unpackhi_epi64(_tmpf, _tmpd);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_CBAD);
-                        _sum9 = _mm512_shuffle_epi32(_sum9, _MM_PERM_CBAD);
-                        _sumb = _mm512_shuffle_epi32(_sumb, _MM_PERM_CBAD);
-                        _sumd = _mm512_shuffle_epi32(_sumd, _MM_PERM_CBAD);
-                        _sumf = _mm512_shuffle_epi32(_sumf, _MM_PERM_CBAD);
-                    }
-
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum4, _sum5, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum6, _sum7, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp4 = _mm512_shuffle_i32x4(_sum8, _sum9, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmp5 = _mm512_shuffle_i32x4(_suma, _sumb, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmp6 = _mm512_shuffle_i32x4(_sumc, _sumd, _MM_SHUFFLE(1, 3, 1, 3));
-                    __m512i _tmp7 = _mm512_shuffle_i32x4(_sume, _sumf, _MM_SHUFFLE(1, 3, 1, 3));
-                    __m512i _tmp8 = _mm512_shuffle_i32x4(_sumc, _sumd, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp9 = _mm512_shuffle_i32x4(_sume, _sumf, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmpa = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmpb = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmpc = _mm512_shuffle_i32x4(_sum4, _sum5, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmpd = _mm512_shuffle_i32x4(_sum6, _sum7, _MM_SHUFFLE(0, 2, 0, 2));
-                    __m512i _tmpe = _mm512_shuffle_i32x4(_sum8, _sum9, _MM_SHUFFLE(1, 3, 1, 3));
-                    __m512i _tmpf = _mm512_shuffle_i32x4(_suma, _sumb, _MM_SHUFFLE(1, 3, 1, 3));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp4, _tmp5, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp6, _tmp7, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum4 = _mm512_shuffle_i32x4(_tmp8, _tmp9, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum5 = _mm512_shuffle_i32x4(_tmpa, _tmpb, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum6 = _mm512_shuffle_i32x4(_tmpc, _tmpd, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum7 = _mm512_shuffle_i32x4(_tmpe, _tmpf, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum8 = _mm512_shuffle_i32x4(_tmp4, _tmp5, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum9 = _mm512_shuffle_i32x4(_tmp6, _tmp7, _MM_SHUFFLE(3, 1, 3, 1));
-                    _suma = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sumb = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sumc = _mm512_shuffle_i32x4(_tmpc, _tmpd, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sumd = _mm512_shuffle_i32x4(_tmpe, _tmpf, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sume = _mm512_shuffle_i32x4(_tmp8, _tmp9, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sumf = _mm512_shuffle_i32x4(_tmpa, _tmpb, _MM_SHUFFLE(3, 1, 3, 1));
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _sum0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16), _sum1);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 32), _sum2);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 48), _sum3);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 4), _sum4);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 4 + 16), _sum5);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 4 + 32), _sum6);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 4 + 48), _sum7);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8), _sum8);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8 + 16), _sum9);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8 + 32), _suma);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8 + 48), _sumb);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 12), _sumc);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 12 + 16), _sumd);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 12 + 32), _sume);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 12 + 48), _sumf);
-                    outptr0 += 64;
-                }
-                if (out_elempack == 1)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77 88 99 aa bb cc dd ee ff
-                    //      01 12 23 30 45 56 67 74 89 9a ab b8 cd de ef fc
-                    //      02 13 20 31 46 57 64 75 8a 9b a8 b9 ce df ec fd
-                    //      03 10 21 32 47 54 65 76 8b 98 a9 ba cf dc ed fe
-                    //      c0 d1 e2 f3 04 15 26 37 48 59 6a 7b 8c 9d ae bf
-                    //      c1 d2 e3 f0 05 16 27 34 49 5a 6b 78 8d 9e af bc
-                    //      c2 d3 e0 f1 06 17 24 35 4a 5b 68 79 8e 9f ac bd
-                    //      c3 d0 e1 f2 07 14 25 36 4b 58 69 7a 8f 9c ad be
-                    //      80 91 a2 b3 c4 d5 e6 f7 08 19 2a 3b 4c 5d 6e 7f
-                    //      81 92 a3 b0 c5 d6 e7 f4 09 1a 2b 38 4d 5e 6f 7c
-                    //      82 93 a0 b1 c6 d7 e4 f5 0a 1b 28 39 4e 5f 6c 7d
-                    //      83 90 a1 b2 c7 d4 e5 f6 0b 18 29 3a 4f 5c 6d 7e
-                    //      40 51 62 73 84 95 a6 b7 c8 d9 ea fb 0c 1d 2e 3f
-                    //      41 52 63 70 85 96 a7 b4 c9 da eb f8 0d 1e 2f 3c
-                    //      42 53 60 71 86 97 a4 b5 ca db e8 f9 0e 1f 2c 3d
-                    //      43 50 61 72 87 94 a5 b6 cb d8 e9 fa 0f 1c 2d 3e
-                    // to
-                    //      00 01 02 03 44 45 46 47 88 80 8a 8b cc cd ce cf
-                    //      10 11 12 13 54 55 56 57 98 90 9a 9b dc dd de df
-                    //      20 21 22 23 64 65 66 67 a8 a0 aa ab ec ed ee ef
-                    //      30 31 32 33 74 75 76 77 b8 b0 ba bb fc fd fe ff
-                    //      c0 c1 c2 c3 04 05 06 07 48 40 4a 4b 8c 8d 8e 8f
-                    //      d0 d1 d2 d3 14 15 16 17 58 50 5a 5b 9c 9d 9e 9f
-                    //      e0 e1 e2 e3 24 25 26 27 68 60 6a 6b ac ad ae af
-                    //      f0 f1 f2 f3 34 35 36 37 78 70 7a 7b bc bd be bf
-                    //      80 81 82 83 c4 c5 c6 c7 08 00 0a 0b 4c 4d 4e 4f
-                    //      90 91 92 93 d4 d5 d6 d7 18 10 1a 1b 5c 5d 5e 5f
-                    //      a0 a1 a2 a3 e4 e5 e6 e7 28 20 2a 2b 6c 6d 6e 6f
-                    //      b0 b1 b2 b3 f4 f5 f6 f7 38 30 3a 3b 7c 7d 7e 7f
-                    //      40 41 42 43 84 85 86 87 c8 c0 ca cb 0c 0d 0e 0f
-                    //      50 51 52 53 94 95 96 97 d8 d0 da db 1c 1d 1e 1f
-                    //      60 61 62 63 a4 a5 a6 a7 e8 e0 ea eb 2c 2d 2e 2f
-                    //      70 71 72 73 b4 b5 b6 b7 f8 f0 fa fb 3c 3d 3e 3f
-                    {
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum1);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum1);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum3);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum3);
-                        __m512i _tmp4 = _mm512_unpacklo_epi32(_sum4, _sum5);
-                        __m512i _tmp5 = _mm512_unpackhi_epi32(_sum4, _sum5);
-                        __m512i _tmp6 = _mm512_unpacklo_epi32(_sum6, _sum7);
-                        __m512i _tmp7 = _mm512_unpackhi_epi32(_sum6, _sum7);
-                        __m512i _tmp8 = _mm512_unpacklo_epi32(_sum8, _sum9);
-                        __m512i _tmp9 = _mm512_unpackhi_epi32(_sum8, _sum9);
-                        __m512i _tmpa = _mm512_unpacklo_epi32(_suma, _sumb);
-                        __m512i _tmpb = _mm512_unpackhi_epi32(_suma, _sumb);
-                        __m512i _tmpc = _mm512_unpacklo_epi32(_sumc, _sumd);
-                        __m512i _tmpd = _mm512_unpackhi_epi32(_sumc, _sumd);
-                        __m512i _tmpe = _mm512_unpacklo_epi32(_sume, _sumf);
-                        __m512i _tmpf = _mm512_unpackhi_epi32(_sume, _sumf);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm512_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm512_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm512_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm512_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum8 = _mm512_unpacklo_epi64(_tmp8, _tmpa);
-                        _sum9 = _mm512_unpackhi_epi64(_tmp8, _tmpa);
-                        _suma = _mm512_unpacklo_epi64(_tmpb, _tmp9);
-                        _sumb = _mm512_unpackhi_epi64(_tmpb, _tmp9);
-                        _sumc = _mm512_unpacklo_epi64(_tmpc, _tmpe);
-                        _sumd = _mm512_unpackhi_epi64(_tmpc, _tmpe);
-                        _sume = _mm512_unpacklo_epi64(_tmpf, _tmpd);
-                        _sumf = _mm512_unpackhi_epi64(_tmpf, _tmpd);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_CBAD);
-                        _sum9 = _mm512_shuffle_epi32(_sum9, _MM_PERM_CBAD);
-                        _sumb = _mm512_shuffle_epi32(_sumb, _MM_PERM_CBAD);
-                        _sumd = _mm512_shuffle_epi32(_sumd, _MM_PERM_CBAD);
-                        _sumf = _mm512_shuffle_epi32(_sumf, _MM_PERM_CBAD);
-                    }
-
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum4, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum1, _sum5, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum2, _sum6, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum3, _sum7, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp4 = _mm512_shuffle_i32x4(_sumc, _sum0, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp5 = _mm512_shuffle_i32x4(_sumd, _sum1, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp6 = _mm512_shuffle_i32x4(_sume, _sum2, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp7 = _mm512_shuffle_i32x4(_sumf, _sum3, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp8 = _mm512_shuffle_i32x4(_sum8, _sumc, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp9 = _mm512_shuffle_i32x4(_sum9, _sumd, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmpa = _mm512_shuffle_i32x4(_suma, _sume, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmpb = _mm512_shuffle_i32x4(_sumb, _sumf, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmpc = _mm512_shuffle_i32x4(_sum4, _sum8, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmpd = _mm512_shuffle_i32x4(_sum5, _sum9, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmpe = _mm512_shuffle_i32x4(_sum6, _suma, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmpf = _mm512_shuffle_i32x4(_sum7, _sumb, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp8, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp1, _tmp9, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp2, _tmpa, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp3, _tmpb, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum4 = _mm512_shuffle_i32x4(_tmp4, _tmpc, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum5 = _mm512_shuffle_i32x4(_tmp5, _tmpd, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum6 = _mm512_shuffle_i32x4(_tmp6, _tmpe, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum7 = _mm512_shuffle_i32x4(_tmp7, _tmpf, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum8 = _mm512_shuffle_i32x4(_tmp8, _tmp0, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum9 = _mm512_shuffle_i32x4(_tmp9, _tmp1, _MM_SHUFFLE(3, 1, 2, 0));
-                    _suma = _mm512_shuffle_i32x4(_tmpa, _tmp2, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sumb = _mm512_shuffle_i32x4(_tmpb, _tmp3, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sumc = _mm512_shuffle_i32x4(_tmpc, _tmp4, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sumd = _mm512_shuffle_i32x4(_tmpd, _tmp5, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sume = _mm512_shuffle_i32x4(_tmpe, _tmp6, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sumf = _mm512_shuffle_i32x4(_tmpf, _tmp7, _MM_SHUFFLE(3, 1, 2, 0));
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _sum0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep), _sum1);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 2), _sum2);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 3), _sum3);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 4), _sum4);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 5), _sum5);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 6), _sum6);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 7), _sum7);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8), _sum8);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 9), _sum9);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 10), _suma);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 11), _sumb);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 12), _sumc);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 13), _sumd);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 14), _sume);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 15), _sumf);
-                    outptr0 += 16;
-                }
-            }
-            else
-            {
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                _mm512_store_si512((__m512i*)(outptr + 16), _sum1);
-                _mm512_store_si512((__m512i*)(outptr + 32), _sum2);
-                _mm512_store_si512((__m512i*)(outptr + 48), _sum3);
-                _mm512_store_si512((__m512i*)(outptr + 64), _sum4);
-                _mm512_store_si512((__m512i*)(outptr + 80), _sum5);
-                _mm512_store_si512((__m512i*)(outptr + 96), _sum6);
-                _mm512_store_si512((__m512i*)(outptr + 112), _sum7);
-                _mm512_store_si512((__m512i*)(outptr + 128), _sum8);
-                _mm512_store_si512((__m512i*)(outptr + 128 + 16), _sum9);
-                _mm512_store_si512((__m512i*)(outptr + 128 + 32), _suma);
-                _mm512_store_si512((__m512i*)(outptr + 128 + 48), _sumb);
-                _mm512_store_si512((__m512i*)(outptr + 128 + 64), _sumc);
-                _mm512_store_si512((__m512i*)(outptr + 128 + 80), _sumd);
-                _mm512_store_si512((__m512i*)(outptr + 128 + 96), _sume);
-                _mm512_store_si512((__m512i*)(outptr + 128 + 112), _sumf);
-            }
-
-            outptr += 256;
-        }
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            const signed char* pA = pAT;
-
-            __m512i _sum0;
-            __m512i _sum1;
-            __m512i _sum2;
-            __m512i _sum3;
-            __m512i _sum4;
-            __m512i _sum5;
-            __m512i _sum6;
-            __m512i _sum7;
-
-            if (k == 0)
-            {
-                _sum0 = _mm512_setzero_si512();
-                _sum1 = _mm512_setzero_si512();
-                _sum2 = _mm512_setzero_si512();
-                _sum3 = _mm512_setzero_si512();
-                _sum4 = _mm512_setzero_si512();
-                _sum5 = _mm512_setzero_si512();
-                _sum6 = _mm512_setzero_si512();
-                _sum7 = _mm512_setzero_si512();
-            }
-            else
-            {
-                _sum0 = _mm512_load_si512((const __m512i*)outptr);
-                _sum1 = _mm512_load_si512((const __m512i*)(outptr + 16));
-                _sum2 = _mm512_load_si512((const __m512i*)(outptr + 32));
-                _sum3 = _mm512_load_si512((const __m512i*)(outptr + 48));
-                _sum4 = _mm512_load_si512((const __m512i*)(outptr + 64));
-                _sum5 = _mm512_load_si512((const __m512i*)(outptr + 80));
-                _sum6 = _mm512_load_si512((const __m512i*)(outptr + 96));
-                _sum7 = _mm512_load_si512((const __m512i*)(outptr + 112));
-            }
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m256i _pA = _mm256_loadu_si256((const __m256i*)pA);
-                __m128i _pB = _mm_loadu_si128((const __m128i*)pB);
-
-                __m512i _pA0 = _mm512_cvtepi8_epi16(_pA);
-                __m256i _pBB = _mm256_cvtepi8_epi16(_pB);
-
-                // 0123 4567 89ab cdef
-                // 4567 0123 cdef 89ab
-                __m512i _pA1 = _mm512_permutex_epi64(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-
-                // 0123 4567
-                // 0123 4567 0123 4567
-                // 1230 5674 1230 5674
-                // 2301 6745 2301 6745
-                // 3012 7456 3012 7456
-                __m512i _pB0 = _mm512_inserti32x8(_mm512_castsi256_si512(_pBB), _pBB, 1);
-                __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-                __m512i _pB2 = _mm512_shuffle_epi32(_pB0, _MM_PERM_BADC);
-                __m512i _pB3 = _mm512_shuffle_epi32(_pB0, _MM_PERM_CBAD);
-
-#if __AVX512VNNI__
-                _sum0 = _mm512_dpwssd_epi32(_sum0, _pA0, _pB0);
-                _sum1 = _mm512_dpwssd_epi32(_sum1, _pA0, _pB1);
-                _sum2 = _mm512_dpwssd_epi32(_sum2, _pA0, _pB2);
-                _sum3 = _mm512_dpwssd_epi32(_sum3, _pA0, _pB3);
-                _sum4 = _mm512_dpwssd_epi32(_sum4, _pA1, _pB0);
-                _sum5 = _mm512_dpwssd_epi32(_sum5, _pA1, _pB1);
-                _sum6 = _mm512_dpwssd_epi32(_sum6, _pA1, _pB2);
-                _sum7 = _mm512_dpwssd_epi32(_sum7, _pA1, _pB3);
-#else
-                _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA0, _pB0));
-                _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA0, _pB1));
-                _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_pA0, _pB2));
-                _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_pA0, _pB3));
-                _sum4 = _mm512_add_epi32(_sum4, _mm512_madd_epi16(_pA1, _pB0));
-                _sum5 = _mm512_add_epi32(_sum5, _mm512_madd_epi16(_pA1, _pB1));
-                _sum6 = _mm512_add_epi32(_sum6, _mm512_madd_epi16(_pA1, _pB2));
-                _sum7 = _mm512_add_epi32(_sum7, _mm512_madd_epi16(_pA1, _pB3));
-#endif
-
-                pA += 32;
-                pB += 16;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_load_si128((const __m128i*)pA);
-                __m128i _pB = _mm_loadl_epi64((const __m128i*)pB);
-
-                __m256i _pA0 = _mm256_cvtepi8_epi16(_pA);
-                _pB = _mm_cvtepi8_epi16(_pB);
-
-                // 01234567 89abcdef
-                // 45670123 cdef89ab
-                __m256i _pA1 = _mm256_shuffle_epi32(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-
-                // 01234567
-                // 01234567 01234567
-                // 12305674 12305674
-                // 23016745 23016745
-                // 30127456 30127456
-                __m256i _pB0 = _mm256_inserti128_si256(_mm256_castsi128_si256(_pB), _pB, 1);
-                __m256i _pB1 = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(_pB0, _MM_SHUFFLE(0, 3, 2, 1)), _MM_SHUFFLE(0, 3, 2, 1));
-                __m256i _pB2 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(2, 3, 0, 1));
-                __m256i _pB3 = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(_pB0, _MM_SHUFFLE(2, 1, 0, 3)), _MM_SHUFFLE(2, 1, 0, 3));
-
-                __m512i _s0 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB0));
-                __m512i _s1 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB1));
-                __m512i _s2 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB2));
-                __m512i _s3 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB3));
-                __m512i _s4 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA1, _pB0));
-                __m512i _s5 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA1, _pB1));
-                __m512i _s6 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA1, _pB2));
-                __m512i _s7 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA1, _pB3));
-
-                _sum0 = _mm512_add_epi32(_sum0, _s0);
-                _sum1 = _mm512_add_epi32(_sum1, _s1);
-                _sum2 = _mm512_add_epi32(_sum2, _s2);
-                _sum3 = _mm512_add_epi32(_sum3, _s3);
-                _sum4 = _mm512_add_epi32(_sum4, _s4);
-                _sum5 = _mm512_add_epi32(_sum5, _s5);
-                _sum6 = _mm512_add_epi32(_sum6, _s6);
-                _sum7 = _mm512_add_epi32(_sum7, _s7);
-
-                pA += 16;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 16)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77 80 91 a2 b3 c4 d5 e6 f7
-                    //      01 12 23 30 45 56 67 74 81 92 a3 b0 c5 d6 e7 f4
-                    //      02 13 20 31 46 57 64 75 82 93 a0 b1 c6 d7 e4 f5
-                    //      03 10 21 32 47 54 65 76 83 90 a1 b2 c7 d4 e5 f6
-                    //      40 51 62 73 04 15 26 37 c0 d1 e2 f3 84 95 a6 b7
-                    //      41 52 63 70 05 16 27 34 c1 d2 e3 f0 85 96 a7 b4
-                    //      42 53 60 71 06 17 24 35 c2 d3 e0 f1 86 97 a4 b5
-                    //      43 50 61 72 07 14 25 36 c3 d0 e1 f2 87 94 a5 b6
-                    // to
-                    //      00 10 20 30 44 54 64 74 80 90 a0 b0 c4 d4 e4 f4
-                    //      01 11 21 31 45 55 65 75 81 91 a1 b1 c5 d5 e5 f5
-                    //      02 12 22 32 46 56 66 76 82 92 a2 b2 c6 d6 e6 f6
-                    //      03 13 23 33 47 57 67 77 83 93 a3 b3 c7 d7 e7 f7
-                    //      40 50 60 70 04 14 24 34 c0 d0 e0 f0 84 94 a4 b4
-                    //      41 51 61 71 05 15 25 35 c1 d1 e1 f1 85 95 a5 b5
-                    //      42 52 62 72 06 16 26 36 c2 d2 e2 f2 86 96 a6 b6
-                    //      43 53 63 73 07 17 27 37 c3 d3 e3 f3 87 97 a7 b7
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum2 = _mm512_shuffle_epi32(_sum2, _MM_PERM_BADC);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_ADCB);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum6 = _mm512_shuffle_epi32(_sum6, _MM_PERM_BADC);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_ADCB);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        __m512i _tmp4 = _mm512_unpacklo_epi32(_sum4, _sum7);
-                        __m512i _tmp5 = _mm512_unpackhi_epi32(_sum4, _sum7);
-                        __m512i _tmp6 = _mm512_unpacklo_epi32(_sum6, _sum5);
-                        __m512i _tmp7 = _mm512_unpackhi_epi32(_sum6, _sum5);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm512_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm512_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm512_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm512_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_CBAD);
-                    }
-
-                    // TODO
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum4, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum1, _sum5, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum2, _sum6, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum3, _sum7, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp4 = _mm512_shuffle_i32x4(_sum4, _sum0, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp5 = _mm512_shuffle_i32x4(_sum5, _sum1, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp6 = _mm512_shuffle_i32x4(_sum6, _sum2, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp7 = _mm512_shuffle_i32x4(_sum7, _sum3, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp0, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp1, _tmp1, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp2, _tmp2, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp3, _tmp3, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum4 = _mm512_shuffle_i32x4(_tmp4, _tmp4, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum5 = _mm512_shuffle_i32x4(_tmp5, _tmp5, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum6 = _mm512_shuffle_i32x4(_tmp6, _tmp6, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum7 = _mm512_shuffle_i32x4(_tmp7, _tmp7, _MM_SHUFFLE(3, 1, 2, 0));
-
-                    _mm512_store_si512((__m512i*)outptr0, _sum0);
-                    _mm512_store_si512((__m512i*)(outptr0 + 16), _sum1);
-                    _mm512_store_si512((__m512i*)(outptr0 + 32), _sum2);
-                    _mm512_store_si512((__m512i*)(outptr0 + 48), _sum3);
-                    _mm512_store_si512((__m512i*)(outptr0 + 64), _sum4);
-                    _mm512_store_si512((__m512i*)(outptr0 + 80), _sum5);
-                    _mm512_store_si512((__m512i*)(outptr0 + 96), _sum6);
-                    _mm512_store_si512((__m512i*)(outptr0 + 112), _sum7);
-                    outptr0 += 128;
-                }
-                if (out_elempack == 8)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77 80 91 a2 b3 c4 d5 e6 f7
-                    //      01 12 23 30 45 56 67 74 81 92 a3 b0 c5 d6 e7 f4
-                    //      02 13 20 31 46 57 64 75 82 93 a0 b1 c6 d7 e4 f5
-                    //      03 10 21 32 47 54 65 76 83 90 a1 b2 c7 d4 e5 f6
-                    //      40 51 62 73 04 15 26 37 c0 d1 e2 f3 84 95 a6 b7
-                    //      41 52 63 70 05 16 27 34 c1 d2 e3 f0 85 96 a7 b4
-                    //      42 53 60 71 06 17 24 35 c2 d3 e0 f1 86 97 a4 b5
-                    //      43 50 61 72 07 14 25 36 c3 d0 e1 f2 87 94 a5 b6
-                    // to
-                    //      00 10 20 30 44 54 64 74 80 90 a0 b0 c4 d4 e4 f4
-                    //      01 11 21 31 45 55 65 75 81 91 a1 b1 c5 d5 e5 f5
-                    //      02 12 22 32 46 56 66 76 82 92 a2 b2 c6 d6 e6 f6
-                    //      03 13 23 33 47 57 67 77 83 93 a3 b3 c7 d7 e7 f7
-                    //      40 50 60 70 04 14 24 34 c0 d0 e0 f0 84 94 a4 b4
-                    //      41 51 61 71 05 15 25 35 c1 d1 e1 f1 85 95 a5 b5
-                    //      42 52 62 72 06 16 26 36 c2 d2 e2 f2 86 96 a6 b6
-                    //      43 53 63 73 07 17 27 37 c3 d3 e3 f3 87 97 a7 b7
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum2 = _mm512_shuffle_epi32(_sum2, _MM_PERM_BADC);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_ADCB);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum6 = _mm512_shuffle_epi32(_sum6, _MM_PERM_BADC);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_ADCB);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        __m512i _tmp4 = _mm512_unpacklo_epi32(_sum4, _sum7);
-                        __m512i _tmp5 = _mm512_unpackhi_epi32(_sum4, _sum7);
-                        __m512i _tmp6 = _mm512_unpacklo_epi32(_sum6, _sum5);
-                        __m512i _tmp7 = _mm512_unpackhi_epi32(_sum6, _sum5);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm512_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm512_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm512_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm512_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_CBAD);
-                    }
-
-                    // TODO
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum4, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum1, _sum5, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum2, _sum6, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum3, _sum7, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp4 = _mm512_shuffle_i32x4(_sum4, _sum0, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp5 = _mm512_shuffle_i32x4(_sum5, _sum1, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp6 = _mm512_shuffle_i32x4(_sum6, _sum2, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp7 = _mm512_shuffle_i32x4(_sum7, _sum3, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp4, _tmp5, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp6, _tmp7, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum4 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum5 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum6 = _mm512_shuffle_i32x4(_tmp4, _tmp5, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum7 = _mm512_shuffle_i32x4(_tmp6, _tmp7, _MM_SHUFFLE(3, 1, 3, 1));
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _sum0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16), _sum1);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16 * 2), _sum2);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16 * 3), _sum3);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8), _sum4);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8 + 16), _sum5);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8 + 16 * 2), _sum6);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8 + 16 * 3), _sum7);
-                    outptr0 += 64;
-                }
-                if (out_elempack == 4)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77 80 91 a2 b3 c4 d5 e6 f7
-                    //      01 12 23 30 45 56 67 74 81 92 a3 b0 c5 d6 e7 f4
-                    //      02 13 20 31 46 57 64 75 82 93 a0 b1 c6 d7 e4 f5
-                    //      03 10 21 32 47 54 65 76 83 90 a1 b2 c7 d4 e5 f6
-                    //      40 51 62 73 04 15 26 37 c0 d1 e2 f3 84 95 a6 b7
-                    //      41 52 63 70 05 16 27 34 c1 d2 e3 f0 85 96 a7 b4
-                    //      42 53 60 71 06 17 24 35 c2 d3 e0 f1 86 97 a4 b5
-                    //      43 50 61 72 07 14 25 36 c3 d0 e1 f2 87 94 a5 b6
-                    // to
-                    //      00 10 20 30 44 54 64 74 80 90 a0 b0 c4 d4 e4 f4
-                    //      01 11 21 31 45 55 65 75 81 91 a1 b1 c5 d5 e5 f5
-                    //      02 12 22 32 46 56 66 76 82 92 a2 b2 c6 d6 e6 f6
-                    //      03 13 23 33 47 57 67 77 83 93 a3 b3 c7 d7 e7 f7
-                    //      40 50 60 70 04 14 24 34 c0 d0 e0 f0 84 94 a4 b4
-                    //      41 51 61 71 05 15 25 35 c1 d1 e1 f1 85 95 a5 b5
-                    //      42 52 62 72 06 16 26 36 c2 d2 e2 f2 86 96 a6 b6
-                    //      43 53 63 73 07 17 27 37 c3 d3 e3 f3 87 97 a7 b7
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum2 = _mm512_shuffle_epi32(_sum2, _MM_PERM_BADC);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_ADCB);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum6 = _mm512_shuffle_epi32(_sum6, _MM_PERM_BADC);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_ADCB);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        __m512i _tmp4 = _mm512_unpacklo_epi32(_sum4, _sum7);
-                        __m512i _tmp5 = _mm512_unpackhi_epi32(_sum4, _sum7);
-                        __m512i _tmp6 = _mm512_unpacklo_epi32(_sum6, _sum5);
-                        __m512i _tmp7 = _mm512_unpackhi_epi32(_sum6, _sum5);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm512_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm512_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm512_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm512_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_CBAD);
-                    }
-
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum4, _sum5, _MM_SHUFFLE(0, 1, 0, 1));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum6, _sum7, _MM_SHUFFLE(0, 1, 0, 1));
-                    __m512i _tmp4 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(3, 2, 3, 2));
-                    __m512i _tmp5 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(3, 2, 3, 2));
-                    __m512i _tmp6 = _mm512_shuffle_i32x4(_sum4, _sum5, _MM_SHUFFLE(2, 3, 2, 3));
-                    __m512i _tmp7 = _mm512_shuffle_i32x4(_sum6, _sum7, _MM_SHUFFLE(2, 3, 2, 3));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum4 = _mm512_shuffle_i32x4(_tmp4, _tmp5, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum5 = _mm512_shuffle_i32x4(_tmp6, _tmp7, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum6 = _mm512_shuffle_i32x4(_tmp6, _tmp7, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum7 = _mm512_shuffle_i32x4(_tmp4, _tmp5, _MM_SHUFFLE(3, 1, 3, 1));
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _sum0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16), _sum1);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 4), _sum2);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 4 + 16), _sum3);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8), _sum4);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8 + 16), _sum5);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 12), _sum6);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 12 + 16), _sum7);
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77 80 91 a2 b3 c4 d5 e6 f7
-                    //      01 12 23 30 45 56 67 74 81 92 a3 b0 c5 d6 e7 f4
-                    //      02 13 20 31 46 57 64 75 82 93 a0 b1 c6 d7 e4 f5
-                    //      03 10 21 32 47 54 65 76 83 90 a1 b2 c7 d4 e5 f6
-                    //      40 51 62 73 04 15 26 37 c0 d1 e2 f3 84 95 a6 b7
-                    //      41 52 63 70 05 16 27 34 c1 d2 e3 f0 85 96 a7 b4
-                    //      42 53 60 71 06 17 24 35 c2 d3 e0 f1 86 97 a4 b5
-                    //      43 50 61 72 07 14 25 36 c3 d0 e1 f2 87 94 a5 b6
-                    // to
-                    //      00 01 02 03 44 45 46 47 80 81 82 83 c4 c5 c6 c7
-                    //      10 11 12 13 54 55 56 57 90 91 92 93 d4 d5 d6 d7
-                    //      20 21 22 23 64 65 66 67 a0 a1 a2 a3 e4 e5 e6 e7
-                    //      30 31 32 33 74 75 76 77 b0 b1 b2 b3 f4 f5 f6 f7
-                    //      40 41 42 43 04 05 06 07 c0 c1 c2 c3 84 85 86 87
-                    //      50 51 52 53 14 15 16 17 d0 d1 d2 d3 94 95 96 97
-                    //      60 61 62 63 24 25 26 27 e0 e1 e2 e3 a4 a5 a6 a7
-                    //      70 71 72 73 34 35 36 37 f0 f1 f2 f3 b4 b5 b6 b7
-                    {
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum1);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum1);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum3);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum3);
-                        __m512i _tmp4 = _mm512_unpacklo_epi32(_sum4, _sum5);
-                        __m512i _tmp5 = _mm512_unpackhi_epi32(_sum4, _sum5);
-                        __m512i _tmp6 = _mm512_unpacklo_epi32(_sum6, _sum7);
-                        __m512i _tmp7 = _mm512_unpackhi_epi32(_sum6, _sum7);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm512_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm512_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm512_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm512_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_CBAD);
-                    }
-
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum4, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum1, _sum5, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum2, _sum6, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum3, _sum7, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp4 = _mm512_shuffle_i32x4(_sum4, _sum0, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp5 = _mm512_shuffle_i32x4(_sum5, _sum1, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp6 = _mm512_shuffle_i32x4(_sum6, _sum2, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp7 = _mm512_shuffle_i32x4(_sum7, _sum3, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp0, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp1, _tmp1, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp2, _tmp2, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp3, _tmp3, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum4 = _mm512_shuffle_i32x4(_tmp4, _tmp4, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum5 = _mm512_shuffle_i32x4(_tmp5, _tmp5, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum6 = _mm512_shuffle_i32x4(_tmp6, _tmp6, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum7 = _mm512_shuffle_i32x4(_tmp7, _tmp7, _MM_SHUFFLE(3, 1, 2, 0));
-
-                    _mm256_storeu_si256((__m256i*)outptr0, _mm512_extracti32x8_epi32(_sum0, 0));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep), _mm512_extracti32x8_epi32(_sum1, 0));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 2), _mm512_extracti32x8_epi32(_sum2, 0));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 3), _mm512_extracti32x8_epi32(_sum3, 0));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 4), _mm512_extracti32x8_epi32(_sum4, 0));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 5), _mm512_extracti32x8_epi32(_sum5, 0));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 6), _mm512_extracti32x8_epi32(_sum6, 0));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 7), _mm512_extracti32x8_epi32(_sum7, 0));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 8), _mm512_extracti32x8_epi32(_sum0, 1));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 9), _mm512_extracti32x8_epi32(_sum1, 1));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 10), _mm512_extracti32x8_epi32(_sum2, 1));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 11), _mm512_extracti32x8_epi32(_sum3, 1));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 12), _mm512_extracti32x8_epi32(_sum4, 1));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 13), _mm512_extracti32x8_epi32(_sum5, 1));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 14), _mm512_extracti32x8_epi32(_sum6, 1));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 15), _mm512_extracti32x8_epi32(_sum7, 1));
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                _mm512_store_si512((__m512i*)(outptr + 16), _sum1);
-                _mm512_store_si512((__m512i*)(outptr + 32), _sum2);
-                _mm512_store_si512((__m512i*)(outptr + 48), _sum3);
-                _mm512_store_si512((__m512i*)(outptr + 64), _sum4);
-                _mm512_store_si512((__m512i*)(outptr + 80), _sum5);
-                _mm512_store_si512((__m512i*)(outptr + 96), _sum6);
-                _mm512_store_si512((__m512i*)(outptr + 112), _sum7);
-            }
-
-            outptr += 128;
-        }
-#endif // defined(__x86_64__) || defined(_M_X64)
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            const signed char* pA = pAT;
-
-            __m512i _sum0;
-            __m512i _sum1;
-            __m512i _sum2;
-            __m512i _sum3;
-
-            if (k == 0)
-            {
-                _sum0 = _mm512_setzero_si512();
-                _sum1 = _mm512_setzero_si512();
-                _sum2 = _mm512_setzero_si512();
-                _sum3 = _mm512_setzero_si512();
-            }
-            else
-            {
-                _sum0 = _mm512_load_si512((const __m512i*)outptr);
-                _sum1 = _mm512_load_si512((const __m512i*)(outptr + 16));
-                _sum2 = _mm512_load_si512((const __m512i*)(outptr + 32));
-                _sum3 = _mm512_load_si512((const __m512i*)(outptr + 48));
-            }
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m256i _pA = _mm256_loadu_si256((const __m256i*)pA);
-                __m256i _pB = _mm256_castpd_si256(_mm256_broadcast_sd((const double*)pB));
-
-                __m512i _pA0 = _mm512_cvtepi8_epi16(_pA);
-                __m512i _pB0 = _mm512_cvtepi8_epi16(_pB);
-
-                // 0123 4567 89ab cdef
-                // 2301 6745 ab89 efcd
-                __m512i _pA1 = _mm512_shuffle_epi32(_pA0, _MM_PERM_BADC);
-
-                // 0123 0123 0123 0123
-                // 1230 1230 1230 1230
-                __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-
-#if __AVX512VNNI__
-                _sum0 = _mm512_dpwssd_epi32(_sum0, _pA0, _pB0);
-                _sum1 = _mm512_dpwssd_epi32(_sum1, _pA0, _pB1);
-                _sum2 = _mm512_dpwssd_epi32(_sum2, _pA1, _pB0);
-                _sum3 = _mm512_dpwssd_epi32(_sum3, _pA1, _pB1);
-#else
-                _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA0, _pB0));
-                _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA0, _pB1));
-                _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_pA1, _pB0));
-                _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_pA1, _pB1));
-#endif
-
-                pA += 32;
-                pB += 8;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_load_si128((const __m128i*)pA);
-                __m128i _pB = _mm_castps_si128(_mm_load1_ps((const float*)pB));
-
-                __m256i _pA0 = _mm256_cvtepi8_epi16(_pA);
-                __m256i _pB0 = _mm256_cvtepi8_epi16(_pB);
-
-                // 01234567 89abcdef
-                // 23016745 ab89efcd
-                __m256i _pA1 = _mm256_shuffle_epi32(_pA0, _MM_SHUFFLE(2, 3, 0, 1));
-
-                // 01230123 01230123
-                // 12301230 12301230
-                __m256i _pB1 = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(_pB0, _MM_SHUFFLE(0, 3, 2, 1)), _MM_SHUFFLE(0, 3, 2, 1));
-
-                __m512i _s0 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB0));
-                __m512i _s1 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB1));
-                __m512i _s2 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA1, _pB0));
-                __m512i _s3 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA1, _pB1));
-
-                _sum0 = _mm512_add_epi32(_sum0, _s0);
-                _sum1 = _mm512_add_epi32(_sum1, _s1);
-                _sum2 = _mm512_add_epi32(_sum2, _s2);
-                _sum3 = _mm512_add_epi32(_sum3, _s3);
-
-                pA += 16;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 16)
-                {
-                    // from
-                    //      00 11 22 33 40 51 62 73 80 91 a2 b3 c0 d1 e2 f3
-                    //      01 12 23 30 41 52 63 70 81 92 a3 b0 c1 d2 e3 f0
-                    //      20 31 02 13 60 71 42 53 a0 b1 82 93 e0 f1 c2 d3
-                    //      21 32 03 10 61 72 43 50 a1 b2 83 90 e1 f2 c3 d0
-                    // to
-                    //      00 10 20 30 40 50 60 70 80 90 a0 b0 c0 d0 e0 f0
-                    //      01 11 21 31 41 51 61 71 81 91 a1 b1 c1 d1 e1 f1
-                    //      02 12 22 32 42 52 62 72 82 92 a2 b2 c2 d2 e2 f2
-                    //      03 13 23 33 43 53 63 73 83 93 a3 b3 c3 d3 e3 f3
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                    }
-
-                    _mm512_store_si512((__m512i*)outptr0, _sum0);
-                    _mm512_store_si512((__m512i*)(outptr0 + 16), _sum1);
-                    _mm512_store_si512((__m512i*)(outptr0 + 32), _sum2);
-                    _mm512_store_si512((__m512i*)(outptr0 + 48), _sum3);
-                    outptr0 += 64;
-                }
-                if (out_elempack == 8)
-                {
-                    // from
-                    //      00 11 22 33 40 51 62 73 80 91 a2 b3 c0 d1 e2 f3
-                    //      01 12 23 30 41 52 63 70 81 92 a3 b0 c1 d2 e3 f0
-                    //      20 31 02 13 60 71 42 53 a0 b1 82 93 e0 f1 c2 d3
-                    //      21 32 03 10 61 72 43 50 a1 b2 83 90 e1 f2 c3 d0
-                    // to
-                    //      00 10 20 30 40 50 60 70 80 90 a0 b0 c0 d0 e0 f0
-                    //      01 11 21 31 41 51 61 71 81 91 a1 b1 c1 d1 e1 f1
-                    //      02 12 22 32 42 52 62 72 82 92 a2 b2 c2 d2 e2 f2
-                    //      03 13 23 33 43 53 63 73 83 93 a3 b3 c3 d3 e3 f3
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                    }
-
-                    // TODO
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(3, 2, 3, 2));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(3, 2, 3, 2));
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _tmp0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16), _tmp1);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8), _tmp2);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8 + 16), _tmp3);
-                    outptr0 += 32;
-                }
-                if (out_elempack == 4)
-                {
-                    // from
-                    //      00 11 22 33 40 51 62 73 80 91 a2 b3 c0 d1 e2 f3
-                    //      01 12 23 30 41 52 63 70 81 92 a3 b0 c1 d2 e3 f0
-                    //      20 31 02 13 60 71 42 53 a0 b1 82 93 e0 f1 c2 d3
-                    //      21 32 03 10 61 72 43 50 a1 b2 83 90 e1 f2 c3 d0
-                    // to
-                    //      00 10 20 30 40 50 60 70 80 90 a0 b0 c0 d0 e0 f0
-                    //      01 11 21 31 41 51 61 71 81 91 a1 b1 c1 d1 e1 f1
-                    //      02 12 22 32 42 52 62 72 82 92 a2 b2 c2 d2 e2 f2
-                    //      03 13 23 33 43 53 63 73 83 93 a3 b3 c3 d3 e3 f3
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                    }
-
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(3, 2, 3, 2));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(3, 2, 3, 2));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _sum0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 4), _sum1);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8), _sum2);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 12), _sum3);
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    // from
-                    //      00 11 22 33 40 51 62 73 80 91 a2 b3 c0 d1 e2 f3
-                    //      01 12 23 30 41 52 63 70 81 92 a3 b0 c1 d2 e3 f0
-                    //      20 31 02 13 60 71 42 53 a0 b1 82 93 e0 f1 c2 d3
-                    //      21 32 03 10 61 72 43 50 a1 b2 83 90 e1 f2 c3 d0
-                    // to
-                    //      00 01 02 03 40 41 42 43 80 81 82 83 c0 c1 c2 c3
-                    //      10 11 12 13 50 51 52 53 90 91 92 93 d0 d1 d2 d3
-                    //      20 21 22 23 60 61 62 63 a0 a1 a2 a3 e0 e1 e2 e3
-                    //      30 31 32 33 70 71 72 73 b0 b1 b2 b3 f0 f1 f2 f2
-                    {
-                        _sum2 = _mm512_shuffle_epi32(_sum2, _MM_PERM_BADC);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_BADC);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum1);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum1);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum3);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum3);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                    }
-
-                    _mm_storeu_si128((__m128i*)outptr0, _mm512_extracti32x4_epi32(_sum0, 0));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep), _mm512_extracti32x4_epi32(_sum1, 0));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 2), _mm512_extracti32x4_epi32(_sum2, 0));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 3), _mm512_extracti32x4_epi32(_sum3, 0));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 4), _mm512_extracti32x4_epi32(_sum0, 1));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 5), _mm512_extracti32x4_epi32(_sum1, 1));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 6), _mm512_extracti32x4_epi32(_sum2, 1));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 7), _mm512_extracti32x4_epi32(_sum3, 1));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 8), _mm512_extracti32x4_epi32(_sum0, 2));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 9), _mm512_extracti32x4_epi32(_sum1, 2));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 10), _mm512_extracti32x4_epi32(_sum2, 2));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 11), _mm512_extracti32x4_epi32(_sum3, 2));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 12), _mm512_extracti32x4_epi32(_sum0, 3));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 13), _mm512_extracti32x4_epi32(_sum1, 3));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 14), _mm512_extracti32x4_epi32(_sum2, 3));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 15), _mm512_extracti32x4_epi32(_sum3, 3));
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                _mm512_store_si512((__m512i*)(outptr + 16), _sum1);
-                _mm512_store_si512((__m512i*)(outptr + 32), _sum2);
-                _mm512_store_si512((__m512i*)(outptr + 48), _sum3);
-            }
-
-            outptr += 64;
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            const signed char* pA = pAT;
-
-            __m512i _sum0;
-            __m512i _sum1;
-
-            if (k == 0)
-            {
-                _sum0 = _mm512_setzero_si512();
-                _sum1 = _mm512_setzero_si512();
-            }
-            else
-            {
-                _sum0 = _mm512_load_si512((const __m512i*)outptr);
-                _sum1 = _mm512_load_si512((const __m512i*)(outptr + 16));
-            }
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m256i _pA = _mm256_loadu_si256((const __m256i*)pA);
-                __m256i _pB = _mm256_castps_si256(_mm256_broadcast_ss((const float*)pB));
-
-                __m512i _pA0 = _mm512_cvtepi8_epi16(_pA);
-                __m512i _pB0 = _mm512_cvtepi8_epi16(_pB);
-
-                // 0123 4567 89ab cdef
-
-                // 0101 0101 0101 0101
-                // 1010 1010 1010 1010
-                __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_CDAB);
-
-#if __AVX512VNNI__
-                _sum0 = _mm512_dpwssd_epi32(_sum0, _pA0, _pB0);
-                _sum1 = _mm512_dpwssd_epi32(_sum1, _pA0, _pB1);
-#else
-                _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA0, _pB0));
-                _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA0, _pB1));
-#endif
-
-                pA += 32;
-                pB += 4;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_load_si128((const __m128i*)pA);
-                __m128i _pB = _mm_set1_epi16(((const short*)pB)[0]);
-
-                __m256i _pA0 = _mm256_cvtepi8_epi16(_pA);
-                __m256i _pB0 = _mm256_cvtepi8_epi16(_pB);
-
-                // 01234567 89abcdef
-
-                // 01010101 01010101
-                // 10101010 10101010
-                __m256i _pB1 = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(_pB0, _MM_SHUFFLE(0, 1, 0, 1)), _MM_SHUFFLE(0, 1, 0, 1));
-
-                __m512i _s0 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB0));
-                __m512i _s1 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB1));
-
-                _sum0 = _mm512_add_epi32(_sum0, _s0);
-                _sum1 = _mm512_add_epi32(_sum1, _s1);
-
-                pA += 16;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                // from
-                //      00 11 20 31 40 51 60 71 80 91 a0 b1 c0 d1 e0 f1
-                //      01 10 21 30 41 50 61 70 81 90 a1 b0 c1 d0 e1 f0
-                // to
-                //      00 10 20 30 40 50 60 70 80 90 a0 b0 c0 d0 e0 f0
-                //      01 11 21 31 41 51 61 71 81 91 a1 b1 c1 d1 e1 f1
-                {
-                    __m512i _tmp0 = _mm512_shuffle_epi32(_sum0, _MM_PERM_DBCA);
-                    __m512i _tmp1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_ACDB);
-                    _sum0 = _mm512_unpacklo_epi32(_tmp0, _tmp1);
-                    _sum1 = _mm512_unpackhi_epi32(_tmp0, _tmp1);
-                    _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                }
-
-                if (out_elempack == 16)
-                {
-                    _mm512_store_si512((__m512i*)outptr0, _sum0);
-                    _mm512_store_si512((__m512i*)(outptr0 + 16), _sum1);
-                    outptr0 += 32;
-                }
-                if (out_elempack == 8)
-                {
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(3, 2, 3, 2));
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _tmp0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 8), _tmp1);
-                    outptr0 += 16;
-                }
-                if (out_elempack == 4)
-                {
-                    _mm_storeu_si128((__m128i*)outptr0, _mm512_extracti32x4_epi32(_sum0, 0));
-                    _mm_storeu_si128((__m128i*)(outptr0 + 4), _mm512_extracti32x4_epi32(_sum1, 0));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 4), _mm512_extracti32x4_epi32(_sum0, 1));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 4 + 4), _mm512_extracti32x4_epi32(_sum1, 1));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 8), _mm512_extracti32x4_epi32(_sum0, 2));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 8 + 4), _mm512_extracti32x4_epi32(_sum1, 2));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 12), _mm512_extracti32x4_epi32(_sum0, 3));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 12 + 4), _mm512_extracti32x4_epi32(_sum1, 3));
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-                    __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                    _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(out_hstep));
-                    _mm512_i32scatter_epi32(outptr0, _vindex, _sum0, sizeof(float));
-                    _mm512_i32scatter_epi32(outptr0 + 1, _vindex, _sum1, sizeof(float));
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                _mm512_store_si512((__m512i*)(outptr + 16), _sum1);
-            }
-
-            outptr += 32;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            const signed char* pA = pAT;
-
-            __m512i _sum0;
-
-            if (k == 0)
-            {
-                _sum0 = _mm512_setzero_si512();
-            }
-            else
-            {
-                _sum0 = _mm512_load_si512((const __m512i*)outptr);
-            }
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m256i _pA = _mm256_loadu_si256((const __m256i*)pA);
-                __m256i _pB = _mm256_castps_si256(_mm256_broadcast_ss((const float*)pB));
-
-                __m512i _pA0 = _mm512_cvtepi8_epi16(_pA);
-                __m512i _pBBBB = _mm512_cvtepi8_epi16(_pB);
-
-                // 0xxx0xxx0xxx0xxx -> 00000000...
-                __m512i _pB0 = _mm512_shuffle_epi32(_pBBBB, _MM_PERM_AAAA);
-
-#if __AVX512VNNI__
-                _sum0 = _mm512_dpwssd_epi32(_sum0, _pA0, _pB0);
-#else
-                _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA0, _pB0));
-#endif
-
-                pA += 32;
-                pB += 2;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_load_si128((const __m128i*)pA);
-                __m256i _pB = _mm256_set1_epi16(pB[0]);
-
-                __m256i _pA0 = _mm256_cvtepi8_epi16(_pA);
-
-                __m512i _s0 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB));
-
-                _sum0 = _mm512_add_epi32(_sum0, _s0);
-
-                pA += 16;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 16)
-                {
-                    _mm512_store_si512((__m512i*)outptr0, _sum0);
-                    outptr0 += 16;
-                }
-                if (out_elempack == 8)
-                {
-                    _mm256_storeu_si256((__m256i*)outptr0, _mm512_extracti32x8_epi32(_sum0, 0));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 8), _mm512_extracti32x8_epi32(_sum0, 1));
-                    outptr0 += 8;
-                }
-                if (out_elempack == 4)
-                {
-                    _mm_storeu_si128((__m128i*)outptr0, _mm512_extracti32x4_epi32(_sum0, 0));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 4), _mm512_extracti32x4_epi32(_sum0, 1));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 8), _mm512_extracti32x4_epi32(_sum0, 2));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 12), _mm512_extracti32x4_epi32(_sum0, 3));
-                    outptr0 += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                    _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(out_hstep));
-                    _mm512_i32scatter_epi32(outptr0, _vindex, _sum0, sizeof(float));
-                    outptr0++;
-                }
-            }
-            else
-            {
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-            }
-
-            outptr += 16;
-        }
-
-        pAT += max_kk * 16;
-    }
-#endif // __AVX512F__
-    for (; ii + 7 < max_ii; ii += 8)
-    {
-        int* outptr0 = (int*)top_blob + (i + ii) * out_hstep + j * out_elempack;
-
-        const signed char* pB = pBT;
-
-        int jj = 0;
-#if defined(__x86_64__) || defined(_M_X64)
-#if __AVX512F__
-        for (; jj + 15 < max_jj; jj += 16)
-        {
-            const signed char* pA = pAT;
-
-            __m512i _sum0;
-            __m512i _sum1;
-            __m512i _sum2;
-            __m512i _sum3;
-            __m512i _sum4;
-            __m512i _sum5;
-            __m512i _sum6;
-            __m512i _sum7;
-
-            if (k == 0)
-            {
-                _sum0 = _mm512_setzero_si512();
-                _sum1 = _mm512_setzero_si512();
-                _sum2 = _mm512_setzero_si512();
-                _sum3 = _mm512_setzero_si512();
-                _sum4 = _mm512_setzero_si512();
-                _sum5 = _mm512_setzero_si512();
-                _sum6 = _mm512_setzero_si512();
-                _sum7 = _mm512_setzero_si512();
-            }
-            else
-            {
-                _sum0 = _mm512_load_si512((const __m512i*)outptr);
-                _sum1 = _mm512_load_si512((const __m512i*)(outptr + 16));
-                _sum2 = _mm512_load_si512((const __m512i*)(outptr + 32));
-                _sum3 = _mm512_load_si512((const __m512i*)(outptr + 48));
-                _sum4 = _mm512_load_si512((const __m512i*)(outptr + 64));
-                _sum5 = _mm512_load_si512((const __m512i*)(outptr + 80));
-                _sum6 = _mm512_load_si512((const __m512i*)(outptr + 96));
-                _sum7 = _mm512_load_si512((const __m512i*)(outptr + 112));
-            }
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _pA = _mm_loadu_si128((const __m128i*)pA);
-                __m256i _pB = _mm256_loadu_si256((const __m256i*)pB);
-
-                __m256i _pA0 = _mm256_cvtepi8_epi16(_pA);
-                __m512i _pB0 = _mm512_cvtepi8_epi16(_pB);
-
-                // 0123 4567 0123 4567
-                // 4567 0123 4567 0123
-                __m512i _pA00 = _mm512_inserti32x8(_mm512_castsi256_si512(_pA0), _pA0, 1);
-                __m512i _pA11 = _mm512_permutex_epi64(_pA00, _MM_SHUFFLE(1, 0, 3, 2));
-
-                // 0123 4567 89ab cdef
-                // 1230 5674 9ab8 defc
-                // 2301 6745 ab89 efcd
-                // 3012 7456 b89a fcde
-                __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-                __m512i _pB2 = _mm512_shuffle_epi32(_pB0, _MM_PERM_BADC);
-                __m512i _pB3 = _mm512_shuffle_epi32(_pB0, _MM_PERM_CBAD);
-
-#if __AVX512VNNI__
-                _sum0 = _mm512_dpwssd_epi32(_sum0, _pA00, _pB0);
-                _sum1 = _mm512_dpwssd_epi32(_sum1, _pA00, _pB1);
-                _sum2 = _mm512_dpwssd_epi32(_sum2, _pA00, _pB2);
-                _sum3 = _mm512_dpwssd_epi32(_sum3, _pA00, _pB3);
-                _sum4 = _mm512_dpwssd_epi32(_sum4, _pA11, _pB0);
-                _sum5 = _mm512_dpwssd_epi32(_sum5, _pA11, _pB1);
-                _sum6 = _mm512_dpwssd_epi32(_sum6, _pA11, _pB2);
-                _sum7 = _mm512_dpwssd_epi32(_sum7, _pA11, _pB3);
-#else
-                _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA00, _pB0));
-                _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA00, _pB1));
-                _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_pA00, _pB2));
-                _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_pA00, _pB3));
-                _sum4 = _mm512_add_epi32(_sum4, _mm512_madd_epi16(_pA11, _pB0));
-                _sum5 = _mm512_add_epi32(_sum5, _mm512_madd_epi16(_pA11, _pB1));
-                _sum6 = _mm512_add_epi32(_sum6, _mm512_madd_epi16(_pA11, _pB2));
-                _sum7 = _mm512_add_epi32(_sum7, _mm512_madd_epi16(_pA11, _pB3));
-#endif // __AVX512VNNI__
-
-                pA += 16;
-                pB += 32;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_loadl_epi64((const __m128i*)pA);
-                __m128i _pB = _mm_load_si128((const __m128i*)pB);
-
-                _pA = _mm_cvtepi8_epi16(_pA);
-                __m256i _pB0 = _mm256_cvtepi8_epi16(_pB);
-
-                // 01234567 01234567
-                // 45670123 45670123
-                __m256i _pA00 = _mm256_inserti128_si256(_mm256_castsi128_si256(_pA), _pA, 1);
-                __m256i _pA11 = _mm256_permute4x64_epi64(_pA00, _MM_SHUFFLE(2, 3, 0, 1));
-
-                // 01234567 89abcdef
-                // 12305674 9ab8defc
-                // 23016745 ab89efcd
-                // 30127456 b89afcde
-                __m256i _pB1 = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(_pB0, _MM_SHUFFLE(0, 3, 2, 1)), _MM_SHUFFLE(0, 3, 2, 1));
-                __m256i _pB2 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(2, 3, 0, 1));
-                __m256i _pB3 = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(_pB0, _MM_SHUFFLE(2, 1, 0, 3)), _MM_SHUFFLE(2, 1, 0, 3));
-
-                __m512i _s0 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA00, _pB0));
-                __m512i _s1 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA00, _pB1));
-                __m512i _s2 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA00, _pB2));
-                __m512i _s3 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA00, _pB3));
-                __m512i _s4 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA11, _pB0));
-                __m512i _s5 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA11, _pB1));
-                __m512i _s6 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA11, _pB2));
-                __m512i _s7 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA11, _pB3));
-
-                _sum0 = _mm512_add_epi32(_sum0, _s0);
-                _sum1 = _mm512_add_epi32(_sum1, _s1);
-                _sum2 = _mm512_add_epi32(_sum2, _s2);
-                _sum3 = _mm512_add_epi32(_sum3, _s3);
-                _sum4 = _mm512_add_epi32(_sum4, _s4);
-                _sum5 = _mm512_add_epi32(_sum5, _s5);
-                _sum6 = _mm512_add_epi32(_sum6, _s6);
-                _sum7 = _mm512_add_epi32(_sum7, _s7);
-
-                pA += 8;
-                pB += 16;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 8)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77 08 19 2a 3b 4c 5d 6e 7f
-                    //      01 12 23 30 45 56 67 74 09 1a 2b 38 4d 5e 6f 7c
-                    //      02 13 20 31 46 57 64 75 0a 1b 28 39 4e 5f 6c 7d
-                    //      03 10 21 32 47 54 65 76 0b 18 29 3a 4f 5c 6d 7e
-                    //      40 51 62 73 04 15 26 37 48 59 6a 7b 0c 1d 2e 3f
-                    //      41 52 63 70 05 16 27 34 49 5a 6b 78 0d 1e 2f 3c
-                    //      42 53 60 71 06 17 24 35 4a 5b 68 79 0e 1f 2c 3d
-                    //      43 50 61 72 07 14 25 36 4b 58 69 7a 0f 1c 2d 3e
-                    // to
-                    //      00 10 20 30 44 54 64 74 08 18 28 38 4c 5c 6c 7c
-                    //      01 11 21 31 45 55 65 75 09 19 29 39 4d 5d 6d 7d
-                    //      02 12 22 32 46 56 66 76 0a 1a 2a 3a 4e 5e 6e 7e
-                    //      03 13 23 33 47 57 67 77 0b 1b 2b 3b 4f 5f 6f 7f
-                    //      40 50 60 70 04 14 24 34 48 58 68 78 0c 1c 2c 3c
-                    //      41 51 61 71 05 15 25 35 49 59 69 79 0d 1d 2d 3d
-                    //      42 52 62 72 06 16 26 36 4a 5a 6a 7a 0e 1e 2e 3e
-                    //      43 53 63 73 07 17 27 37 4b 5b 6b 7b 0f 1f 2f 3f
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum2 = _mm512_shuffle_epi32(_sum2, _MM_PERM_BADC);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_ADCB);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum6 = _mm512_shuffle_epi32(_sum6, _MM_PERM_BADC);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_ADCB);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        __m512i _tmp4 = _mm512_unpacklo_epi32(_sum4, _sum7);
-                        __m512i _tmp5 = _mm512_unpackhi_epi32(_sum4, _sum7);
-                        __m512i _tmp6 = _mm512_unpacklo_epi32(_sum6, _sum5);
-                        __m512i _tmp7 = _mm512_unpackhi_epi32(_sum6, _sum5);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm512_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm512_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm512_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm512_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_CBAD);
-                    }
-
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum4, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum0, _sum4, _MM_SHUFFLE(3, 2, 3, 2));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum1, _sum5, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum1, _sum5, _MM_SHUFFLE(3, 2, 3, 2));
-                    __m512i _tmp4 = _mm512_shuffle_i32x4(_sum2, _sum6, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp5 = _mm512_shuffle_i32x4(_sum2, _sum6, _MM_SHUFFLE(3, 2, 3, 2));
-                    __m512i _tmp6 = _mm512_shuffle_i32x4(_sum3, _sum7, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp7 = _mm512_shuffle_i32x4(_sum3, _sum7, _MM_SHUFFLE(3, 2, 3, 2));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp2, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp4, _tmp6, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp0, _tmp2, _MM_SHUFFLE(1, 3, 1, 3));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp4, _tmp6, _MM_SHUFFLE(1, 3, 1, 3));
-                    _sum4 = _mm512_shuffle_i32x4(_tmp1, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum5 = _mm512_shuffle_i32x4(_tmp5, _tmp7, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum6 = _mm512_shuffle_i32x4(_tmp1, _tmp3, _MM_SHUFFLE(1, 3, 1, 3));
-                    _sum7 = _mm512_shuffle_i32x4(_tmp5, _tmp7, _MM_SHUFFLE(1, 3, 1, 3));
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _sum0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16), _sum1);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 32), _sum2);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 48), _sum3);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 64), _sum4);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 80), _sum5);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 96), _sum6);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 112), _sum7);
-                    outptr0 += 128;
-                }
-                if (out_elempack == 4)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77 08 19 2a 3b 4c 5d 6e 7f
-                    //      01 12 23 30 45 56 67 74 09 1a 2b 38 4d 5e 6f 7c
-                    //      02 13 20 31 46 57 64 75 0a 1b 28 39 4e 5f 6c 7d
-                    //      03 10 21 32 47 54 65 76 0b 18 29 3a 4f 5c 6d 7e
-                    //      40 51 62 73 04 15 26 37 48 59 6a 7b 0c 1d 2e 3f
-                    //      41 52 63 70 05 16 27 34 49 5a 6b 78 0d 1e 2f 3c
-                    //      42 53 60 71 06 17 24 35 4a 5b 68 79 0e 1f 2c 3d
-                    //      43 50 61 72 07 14 25 36 4b 58 69 7a 0f 1c 2d 3e
-                    // to
-                    //      00 10 20 30 44 54 64 74 08 18 28 38 4c 5c 6c 7c
-                    //      01 11 21 31 45 55 65 75 09 19 29 39 4d 5d 6d 7d
-                    //      02 12 22 32 46 56 66 76 0a 1a 2a 3a 4e 5e 6e 7e
-                    //      03 13 23 33 47 57 67 77 0b 1b 2b 3b 4f 5f 6f 7f
-                    //      40 50 60 70 04 14 24 34 48 58 68 78 0c 1c 2c 3c
-                    //      41 51 61 71 05 15 25 35 49 59 69 79 0d 1d 2d 3d
-                    //      42 52 62 72 06 16 26 36 4a 5a 6a 7a 0e 1e 2e 3e
-                    //      43 53 63 73 07 17 27 37 4b 5b 6b 7b 0f 1f 2f 3f
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum2 = _mm512_shuffle_epi32(_sum2, _MM_PERM_BADC);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_ADCB);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum6 = _mm512_shuffle_epi32(_sum6, _MM_PERM_BADC);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_ADCB);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        __m512i _tmp4 = _mm512_unpacklo_epi32(_sum4, _sum7);
-                        __m512i _tmp5 = _mm512_unpackhi_epi32(_sum4, _sum7);
-                        __m512i _tmp6 = _mm512_unpacklo_epi32(_sum6, _sum5);
-                        __m512i _tmp7 = _mm512_unpackhi_epi32(_sum6, _sum5);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm512_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm512_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm512_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm512_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_CBAD);
-                    }
-
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum4, _sum5, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum6, _sum7, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp4 = _mm512_shuffle_i32x4(_sum4, _sum5, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp5 = _mm512_shuffle_i32x4(_sum6, _sum7, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp6 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(3, 1, 3, 1));
-                    __m512i _tmp7 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum4 = _mm512_shuffle_i32x4(_tmp4, _tmp5, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum5 = _mm512_shuffle_i32x4(_tmp6, _tmp7, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum6 = _mm512_shuffle_i32x4(_tmp4, _tmp5, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum7 = _mm512_shuffle_i32x4(_tmp6, _tmp7, _MM_SHUFFLE(3, 1, 3, 1));
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _sum0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16), _sum1);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 32), _sum2);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 48), _sum3);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 4), _sum4);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 4 + 16), _sum5);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 4 + 32), _sum6);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 4 + 48), _sum7);
-                    outptr0 += 64;
-                }
-                if (out_elempack == 1)
-                {
-                    // from
-                    //      00 11 22 33 44 55 66 77 08 19 2a 3b 4c 5d 6e 7f
-                    //      01 12 23 30 45 56 67 74 09 1a 2b 38 4d 5e 6f 7c
-                    //      02 13 20 31 46 57 64 75 0a 1b 28 39 4e 5f 6c 7d
-                    //      03 10 21 32 47 54 65 76 0b 18 29 3a 4f 5c 6d 7e
-                    //      40 51 62 73 04 15 26 37 48 59 6a 7b 0c 1d 2e 3f
-                    //      41 52 63 70 05 16 27 34 49 5a 6b 78 0d 1e 2f 3c
-                    //      42 53 60 71 06 17 24 35 4a 5b 68 79 0e 1f 2c 3d
-                    //      43 50 61 72 07 14 25 36 4b 58 69 7a 0f 1c 2d 3e
-                    // to
-                    //      00 01 02 03 44 45 46 47 08 09 0a 0b 4c 4d 4e 4f
-                    //      10 11 12 13 54 55 56 57 18 19 1a 1b 5c 5d 5e 5f
-                    //      20 21 22 23 64 65 66 67 28 29 2a 2b 6c 6d 6e 6f
-                    //      30 31 32 33 74 75 76 77 38 39 3a 3b 7c 7d 7e 7f
-                    //      40 41 42 43 04 05 06 07 48 49 4a 4b 0c 0d 0e 0f
-                    //      50 51 52 53 14 15 16 17 58 59 5a 5b 1c 1d 1e 1f
-                    //      60 61 62 63 24 25 26 27 68 69 6a 6b 2c 2d 2e 2f
-                    //      70 71 72 73 34 35 36 37 78 79 7a 7b 3c 3d 3e 3f
-                    {
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum1);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum1);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum3);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum3);
-                        __m512i _tmp4 = _mm512_unpacklo_epi32(_sum4, _sum5);
-                        __m512i _tmp5 = _mm512_unpackhi_epi32(_sum4, _sum5);
-                        __m512i _tmp6 = _mm512_unpacklo_epi32(_sum6, _sum7);
-                        __m512i _tmp7 = _mm512_unpackhi_epi32(_sum6, _sum7);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm512_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm512_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm512_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm512_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        _sum5 = _mm512_shuffle_epi32(_sum5, _MM_PERM_CBAD);
-                        _sum7 = _mm512_shuffle_epi32(_sum7, _MM_PERM_CBAD);
-                    }
-
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum4, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum1, _sum5, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum2, _sum6, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum3, _sum7, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp4 = _mm512_shuffle_i32x4(_sum4, _sum0, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp5 = _mm512_shuffle_i32x4(_sum5, _sum1, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp6 = _mm512_shuffle_i32x4(_sum6, _sum2, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m512i _tmp7 = _mm512_shuffle_i32x4(_sum7, _sum3, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp0, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp1, _tmp1, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp2, _tmp2, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp3, _tmp3, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum4 = _mm512_shuffle_i32x4(_tmp4, _tmp4, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum5 = _mm512_shuffle_i32x4(_tmp5, _tmp5, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum6 = _mm512_shuffle_i32x4(_tmp6, _tmp6, _MM_SHUFFLE(3, 1, 2, 0));
-                    _sum7 = _mm512_shuffle_i32x4(_tmp7, _tmp7, _MM_SHUFFLE(3, 1, 2, 0));
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _sum0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep), _sum1);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 2), _sum2);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 3), _sum3);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 4), _sum4);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 5), _sum5);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 6), _sum6);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 7), _sum7);
-
-                    outptr0 += 16;
-                }
-            }
-            else
-            {
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                _mm512_store_si512((__m512i*)(outptr + 16), _sum1);
-                _mm512_store_si512((__m512i*)(outptr + 32), _sum2);
-                _mm512_store_si512((__m512i*)(outptr + 48), _sum3);
-                _mm512_store_si512((__m512i*)(outptr + 64), _sum4);
-                _mm512_store_si512((__m512i*)(outptr + 80), _sum5);
-                _mm512_store_si512((__m512i*)(outptr + 96), _sum6);
-                _mm512_store_si512((__m512i*)(outptr + 112), _sum7);
-            }
-
-            outptr += 128;
-        }
-#endif // __AVX512F__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            const signed char* pA = pAT;
-
-#if __AVX512F__
-            __m512i _sum0;
-            __m512i _sum1;
-            __m512i _sum2;
-            __m512i _sum3;
-#else
-            __m256i _sum0;
-            __m256i _sum1;
-            __m256i _sum2;
-            __m256i _sum3;
-            __m256i _sum4;
-            __m256i _sum5;
-            __m256i _sum6;
-            __m256i _sum7;
-#endif // __AVX512F__
-
-            if (k == 0)
-            {
-#if __AVX512F__
-                _sum0 = _mm512_setzero_si512();
-                _sum1 = _mm512_setzero_si512();
-                _sum2 = _mm512_setzero_si512();
-                _sum3 = _mm512_setzero_si512();
-#else
-                _sum0 = _mm256_setzero_si256();
-                _sum1 = _mm256_setzero_si256();
-                _sum2 = _mm256_setzero_si256();
-                _sum3 = _mm256_setzero_si256();
-                _sum4 = _mm256_setzero_si256();
-                _sum5 = _mm256_setzero_si256();
-                _sum6 = _mm256_setzero_si256();
-                _sum7 = _mm256_setzero_si256();
-#endif // __AVX512F__
-            }
-            else
-            {
-#if __AVX512F__
-                _sum0 = _mm512_load_si512((const __m512i*)outptr);
-                _sum1 = _mm512_load_si512((const __m512i*)(outptr + 16));
-                _sum2 = _mm512_load_si512((const __m512i*)(outptr + 32));
-                _sum3 = _mm512_load_si512((const __m512i*)(outptr + 48));
-#else
-                _sum0 = _mm256_load_si256((const __m256i*)outptr);
-                _sum1 = _mm256_load_si256((const __m256i*)(outptr + 8));
-                _sum2 = _mm256_load_si256((const __m256i*)(outptr + 16));
-                _sum3 = _mm256_load_si256((const __m256i*)(outptr + 24));
-                _sum4 = _mm256_load_si256((const __m256i*)(outptr + 32));
-                _sum5 = _mm256_load_si256((const __m256i*)(outptr + 40));
-                _sum6 = _mm256_load_si256((const __m256i*)(outptr + 48));
-                _sum7 = _mm256_load_si256((const __m256i*)(outptr + 56));
-#endif // __AVX512F__
-            }
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _pA = _mm_loadu_si128((const __m128i*)pA);
-                __m128i _pB = _mm_loadu_si128((const __m128i*)pB);
-
-                __m256i _pA0 = _mm256_cvtepi8_epi16(_pA);
-                __m256i _pB0 = _mm256_cvtepi8_epi16(_pB);
-
-#if __AVX512F__
-                // 0123 4567 0123 4567
-                // 4567 0123 4567 0123
-                __m512i _pA00 = _mm512_inserti32x8(_mm512_castsi256_si512(_pA0), _pA0, 1);
-                __m512i _pA11 = _mm512_permutex_epi64(_pA00, _MM_SHUFFLE(1, 0, 3, 2));
-
-                // 0123 4567 1230 5674
-                // 2301 6745 3012 7456
-                __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-                __m512i _pB01 = _mm512_inserti32x8(_mm512_castsi256_si512(_pB0), _pB1, 1);
-                __m512i _pB23 = _mm512_shuffle_epi32(_pB01, _MM_PERM_BADC);
-
-#if __AVX512VNNI__
-                _sum0 = _mm512_dpwssd_epi32(_sum0, _pA00, _pB01);
-                _sum1 = _mm512_dpwssd_epi32(_sum1, _pA00, _pB23);
-                _sum2 = _mm512_dpwssd_epi32(_sum2, _pA11, _pB01);
-                _sum3 = _mm512_dpwssd_epi32(_sum3, _pA11, _pB23);
-#else
-                _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA00, _pB01));
-                _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA00, _pB23));
-                _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_pA11, _pB01));
-                _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_pA11, _pB23));
-#endif // __AVX512VNNI__
-#else  // __AVX512F__
-
-                // 0123 4567
-                // 4567 0123
-                __m256i _pA1 = _mm256_permute4x64_epi64(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-
-                // 0123 4567
-                // 1230 5674
-                // 2301 6745
-                // 3012 7456
-                __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-                __m256i _pB2 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(1, 0, 3, 2));
-                __m256i _pB3 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(2, 1, 0, 3));
-
-#if __AVXVNNI__
-                _sum0 = _mm256_dpwssd_epi32(_sum0, _pA0, _pB0);
-                _sum1 = _mm256_dpwssd_epi32(_sum1, _pA0, _pB1);
-                _sum2 = _mm256_dpwssd_epi32(_sum2, _pA0, _pB2);
-                _sum3 = _mm256_dpwssd_epi32(_sum3, _pA0, _pB3);
-                _sum4 = _mm256_dpwssd_epi32(_sum4, _pA1, _pB0);
-                _sum5 = _mm256_dpwssd_epi32(_sum5, _pA1, _pB1);
-                _sum6 = _mm256_dpwssd_epi32(_sum6, _pA1, _pB2);
-                _sum7 = _mm256_dpwssd_epi32(_sum7, _pA1, _pB3);
-#else
-                _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_pA0, _pB0));
-                _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_pA0, _pB1));
-                _sum2 = _mm256_add_epi32(_sum2, _mm256_madd_epi16(_pA0, _pB2));
-                _sum3 = _mm256_add_epi32(_sum3, _mm256_madd_epi16(_pA0, _pB3));
-                _sum4 = _mm256_add_epi32(_sum4, _mm256_madd_epi16(_pA1, _pB0));
-                _sum5 = _mm256_add_epi32(_sum5, _mm256_madd_epi16(_pA1, _pB1));
-                _sum6 = _mm256_add_epi32(_sum6, _mm256_madd_epi16(_pA1, _pB2));
-                _sum7 = _mm256_add_epi32(_sum7, _mm256_madd_epi16(_pA1, _pB3));
-#endif // __AVXVNNI__
-#endif // __AVX512F__
-
-                pA += 16;
-                pB += 16;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_loadl_epi64((const __m128i*)pA);
-                __m128i _pB = _mm_loadl_epi64((const __m128i*)pB);
-
-                _pA = _mm_cvtepi8_epi16(_pA);
-                _pB = _mm_cvtepi8_epi16(_pB);
-
-#if __AVX512F__
-                // 01234567 01234567
-                // 45670123 45670123
-                __m256i _pA00 = _mm256_inserti128_si256(_mm256_castsi128_si256(_pA), _pA, 1);
-                __m256i _pA11 = _mm256_permute4x64_epi64(_pA00, _MM_SHUFFLE(2, 3, 0, 1));
-
-                // 01234567 12305674
-                // 23016745 30127456
-                __m128i _pB1 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_pB, _MM_SHUFFLE(0, 3, 2, 1)), _MM_SHUFFLE(0, 3, 2, 1));
-                __m256i _pB01 = _mm256_inserti128_si256(_mm256_castsi128_si256(_pB), _pB1, 1);
-                __m256i _pB23 = _mm256_shuffle_epi32(_pB01, _MM_SHUFFLE(2, 3, 0, 1));
-
-                __m512i _s01 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA00, _pB01));
-                __m512i _s23 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA00, _pB23));
-                __m512i _s45 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA11, _pB01));
-                __m512i _s67 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA11, _pB23));
-
-                _sum0 = _mm512_add_epi32(_sum0, _s01);
-                _sum1 = _mm512_add_epi32(_sum1, _s23);
-                _sum2 = _mm512_add_epi32(_sum2, _s45);
-                _sum3 = _mm512_add_epi32(_sum3, _s67);
-#else
-                // 0123 4567
-                // 4567 0123
-                __m128i _pA0 = _pA;
-                __m128i _pA1 = _mm_shuffle_epi32(_pA, _MM_SHUFFLE(1, 0, 3, 2));
-
-                // 0123 4567
-                // 1230 5674
-                // 2301 6745
-                // 3012 7456
-                __m128i _pB0 = _pB;
-                __m128i _pB1 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_pB, _MM_SHUFFLE(0, 3, 2, 1)), _MM_SHUFFLE(0, 3, 2, 1));
-                __m128i _pB2 = _mm_shuffle_epi32(_pB, _MM_SHUFFLE(2, 3, 0, 1));
-                __m128i _pB3 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_pB, _MM_SHUFFLE(2, 1, 0, 3)), _MM_SHUFFLE(2, 1, 0, 3));
-
-                __m256i _s0 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA0, _pB0));
-                __m256i _s1 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA0, _pB1));
-                __m256i _s2 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA0, _pB2));
-                __m256i _s3 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA0, _pB3));
-                __m256i _s4 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA1, _pB0));
-                __m256i _s5 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA1, _pB1));
-                __m256i _s6 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA1, _pB2));
-                __m256i _s7 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA1, _pB3));
-
-                _sum0 = _mm256_add_epi32(_sum0, _s0);
-                _sum1 = _mm256_add_epi32(_sum1, _s1);
-                _sum2 = _mm256_add_epi32(_sum2, _s2);
-                _sum3 = _mm256_add_epi32(_sum3, _s3);
-                _sum4 = _mm256_add_epi32(_sum4, _s4);
-                _sum5 = _mm256_add_epi32(_sum5, _s5);
-                _sum6 = _mm256_add_epi32(_sum6, _s6);
-                _sum7 = _mm256_add_epi32(_sum7, _s7);
-#endif // __AVX512F__
-
-                pA += 8;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 8)
-                {
-#if __AVX512F__
-                    // from
-                    //      00 11 22 33 44 55 66 77 01 12 23 30 45 56 67 74
-                    //      02 13 20 31 46 57 64 75 03 10 21 32 47 54 65 76
-                    //      40 51 62 73 04 15 26 37 41 52 63 70 05 16 27 34
-                    //      42 53 60 71 06 17 24 35 43 50 61 72 07 14 25 36
-                    // to
-                    //      00 10 20 30 44 54 64 74 04 14 24 34 40 50 60 70
-                    //      01 11 21 31 45 55 65 75 05 15 25 35 41 51 61 71
-                    //      02 12 22 32 46 56 66 76 06 16 26 36 42 52 62 72
-                    //      03 13 23 33 47 57 67 77 07 17 27 37 43 53 63 73
-                    {
-                        __m512i _s0 = _mm512_shuffle_i32x4(_sum0, _sum2, _MM_SHUFFLE(0, 1, 1, 0));
-                        __m512i _s1 = _mm512_shuffle_i32x4(_sum1, _sum3, _MM_SHUFFLE(2, 3, 3, 2));
-                        __m512i _s2 = _mm512_shuffle_i32x4(_sum1, _sum3, _MM_SHUFFLE(0, 1, 1, 0));
-                        __m512i _s3 = _mm512_shuffle_i32x4(_sum0, _sum2, _MM_SHUFFLE(2, 3, 3, 2));
-                        _s1 = _mm512_shuffle_epi32(_s1, _MM_PERM_ADCB);
-                        _s2 = _mm512_shuffle_epi32(_s2, _MM_PERM_BADC);
-                        _s3 = _mm512_shuffle_epi32(_s3, _MM_PERM_CBAD);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_s0, _s1);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_s0, _s1);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_s2, _s3);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_s2, _s3);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                    }
-
-                    // TODO
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(3, 0, 3, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(3, 0, 3, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(1, 2, 1, 2));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(1, 2, 1, 2));
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _tmp0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16), _tmp1);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 32), _tmp2);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 48), _tmp3);
-#else
-                    // from
-                    //      00 11 22 33 44 55 66 77
-                    //      01 12 23 30 45 56 67 74
-                    //      02 13 20 31 46 57 64 75
-                    //      03 10 21 32 47 54 65 76
-                    //      40 51 62 73 04 15 26 37
-                    //      41 52 63 70 05 16 27 34
-                    //      42 53 60 71 06 17 24 35
-                    //      43 50 61 72 07 14 25 36
-                    // to
-                    //      00 10 20 30 44 54 64 74
-                    //      01 11 21 31 45 55 65 75
-                    //      02 12 22 32 46 56 66 76
-                    //      03 13 23 33 47 57 67 77
-                    //      40 50 60 70 04 14 24 34
-                    //      41 51 61 71 05 15 25 35
-                    //      42 52 62 72 06 16 26 36
-                    //      43 53 63 73 07 17 27 37
-                    {
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum2 = _mm256_shuffle_epi32(_sum2, _MM_SHUFFLE(1, 0, 3, 2));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(0, 3, 2, 1));
-                        _sum5 = _mm256_shuffle_epi32(_sum5, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum6 = _mm256_shuffle_epi32(_sum6, _MM_SHUFFLE(1, 0, 3, 2));
-                        _sum7 = _mm256_shuffle_epi32(_sum7, _MM_SHUFFLE(0, 3, 2, 1));
-                        __m256i _tmp0 = _mm256_unpacklo_epi32(_sum0, _sum3);
-                        __m256i _tmp1 = _mm256_unpackhi_epi32(_sum0, _sum3);
-                        __m256i _tmp2 = _mm256_unpacklo_epi32(_sum2, _sum1);
-                        __m256i _tmp3 = _mm256_unpackhi_epi32(_sum2, _sum1);
-                        __m256i _tmp4 = _mm256_unpacklo_epi32(_sum4, _sum7);
-                        __m256i _tmp5 = _mm256_unpackhi_epi32(_sum4, _sum7);
-                        __m256i _tmp6 = _mm256_unpacklo_epi32(_sum6, _sum5);
-                        __m256i _tmp7 = _mm256_unpackhi_epi32(_sum6, _sum5);
-                        _sum0 = _mm256_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm256_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm256_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm256_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm256_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm256_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm256_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm256_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum5 = _mm256_shuffle_epi32(_sum5, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum7 = _mm256_shuffle_epi32(_sum7, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-
-                    // TODO
-                    __m256i _tmp0 = _mm256_permute2x128_si256(_sum0, _sum4, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp1 = _mm256_permute2x128_si256(_sum1, _sum5, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp2 = _mm256_permute2x128_si256(_sum2, _sum6, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp3 = _mm256_permute2x128_si256(_sum3, _sum7, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp4 = _mm256_permute2x128_si256(_sum4, _sum0, _MM_SHUFFLE(0, 3, 0, 1));
-                    __m256i _tmp5 = _mm256_permute2x128_si256(_sum5, _sum1, _MM_SHUFFLE(0, 3, 0, 1));
-                    __m256i _tmp6 = _mm256_permute2x128_si256(_sum6, _sum2, _MM_SHUFFLE(0, 3, 0, 1));
-                    __m256i _tmp7 = _mm256_permute2x128_si256(_sum7, _sum3, _MM_SHUFFLE(0, 3, 0, 1));
-
-                    _mm256_store_si256((__m256i*)outptr0, _tmp0);
-                    _mm256_store_si256((__m256i*)(outptr0 + 8), _tmp1);
-                    _mm256_store_si256((__m256i*)(outptr0 + 16), _tmp2);
-                    _mm256_store_si256((__m256i*)(outptr0 + 24), _tmp3);
-                    _mm256_store_si256((__m256i*)(outptr0 + 32), _tmp4);
-                    _mm256_store_si256((__m256i*)(outptr0 + 40), _tmp5);
-                    _mm256_store_si256((__m256i*)(outptr0 + 48), _tmp6);
-                    _mm256_store_si256((__m256i*)(outptr0 + 56), _tmp7);
-#endif //__AVX512F__
-                    outptr0 += 64;
-                }
-                if (out_elempack == 4)
-                {
-#if __AVX512F__
-                    // from
-                    //      00 11 22 33 44 55 66 77 01 12 23 30 45 56 67 74
-                    //      02 13 20 31 46 57 64 75 03 10 21 32 47 54 65 76
-                    //      40 51 62 73 04 15 26 37 41 52 63 70 05 16 27 34
-                    //      42 53 60 71 06 17 24 35 43 50 61 72 07 14 25 36
-                    // to
-                    //      00 10 20 30 44 54 64 74 04 14 24 34 40 50 60 70
-                    //      01 11 21 31 45 55 65 75 05 15 25 35 41 51 61 71
-                    //      02 12 22 32 46 56 66 76 06 16 26 36 42 52 62 72
-                    //      03 13 23 33 47 57 67 77 07 17 27 37 43 53 63 73
-                    {
-                        __m512i _s0 = _mm512_shuffle_i32x4(_sum0, _sum2, _MM_SHUFFLE(0, 1, 1, 0));
-                        __m512i _s1 = _mm512_shuffle_i32x4(_sum1, _sum3, _MM_SHUFFLE(2, 3, 3, 2));
-                        __m512i _s2 = _mm512_shuffle_i32x4(_sum1, _sum3, _MM_SHUFFLE(0, 1, 1, 0));
-                        __m512i _s3 = _mm512_shuffle_i32x4(_sum0, _sum2, _MM_SHUFFLE(2, 3, 3, 2));
-                        _s1 = _mm512_shuffle_epi32(_s1, _MM_PERM_ADCB);
-                        _s2 = _mm512_shuffle_epi32(_s2, _MM_PERM_BADC);
-                        _s3 = _mm512_shuffle_epi32(_s3, _MM_PERM_CBAD);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_s0, _s1);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_s0, _s1);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_s2, _s3);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_s2, _s3);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                    }
-
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(2, 0, 2, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(1, 3, 1, 3));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(1, 3, 1, 3));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _sum0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16), _sum1);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 4), _sum2);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 4 + 16), _sum3);
-#else
-                    // from
-                    //      00 11 22 33 44 55 66 77
-                    //      01 12 23 30 45 56 67 74
-                    //      02 13 20 31 46 57 64 75
-                    //      03 10 21 32 47 54 65 76
-                    //      40 51 62 73 04 15 26 37
-                    //      41 52 63 70 05 16 27 34
-                    //      42 53 60 71 06 17 24 35
-                    //      43 50 61 72 07 14 25 36
-                    // to
-                    //      00 10 20 30 44 54 64 74
-                    //      01 11 21 31 45 55 65 75
-                    //      02 12 22 32 46 56 66 76
-                    //      03 13 23 33 47 57 67 77
-                    //      40 50 60 70 04 14 24 34
-                    //      41 51 61 71 05 15 25 35
-                    //      42 52 62 72 06 16 26 36
-                    //      43 53 63 73 07 17 27 37
-                    {
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum2 = _mm256_shuffle_epi32(_sum2, _MM_SHUFFLE(1, 0, 3, 2));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(0, 3, 2, 1));
-                        _sum5 = _mm256_shuffle_epi32(_sum5, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum6 = _mm256_shuffle_epi32(_sum6, _MM_SHUFFLE(1, 0, 3, 2));
-                        _sum7 = _mm256_shuffle_epi32(_sum7, _MM_SHUFFLE(0, 3, 2, 1));
-                        __m256i _tmp0 = _mm256_unpacklo_epi32(_sum0, _sum3);
-                        __m256i _tmp1 = _mm256_unpackhi_epi32(_sum0, _sum3);
-                        __m256i _tmp2 = _mm256_unpacklo_epi32(_sum2, _sum1);
-                        __m256i _tmp3 = _mm256_unpackhi_epi32(_sum2, _sum1);
-                        __m256i _tmp4 = _mm256_unpacklo_epi32(_sum4, _sum7);
-                        __m256i _tmp5 = _mm256_unpackhi_epi32(_sum4, _sum7);
-                        __m256i _tmp6 = _mm256_unpacklo_epi32(_sum6, _sum5);
-                        __m256i _tmp7 = _mm256_unpackhi_epi32(_sum6, _sum5);
-                        _sum0 = _mm256_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm256_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm256_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm256_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm256_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm256_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm256_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm256_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum5 = _mm256_shuffle_epi32(_sum5, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum7 = _mm256_shuffle_epi32(_sum7, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-
-                    __m256i _tmp0 = _mm256_permute2x128_si256(_sum0, _sum1, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp1 = _mm256_permute2x128_si256(_sum2, _sum3, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp2 = _mm256_permute2x128_si256(_sum4, _sum5, _MM_SHUFFLE(0, 3, 0, 1));
-                    __m256i _tmp3 = _mm256_permute2x128_si256(_sum6, _sum7, _MM_SHUFFLE(0, 3, 0, 1));
-                    __m256i _tmp4 = _mm256_permute2x128_si256(_sum4, _sum5, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp5 = _mm256_permute2x128_si256(_sum6, _sum7, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp6 = _mm256_permute2x128_si256(_sum0, _sum1, _MM_SHUFFLE(0, 3, 0, 1));
-                    __m256i _tmp7 = _mm256_permute2x128_si256(_sum2, _sum3, _MM_SHUFFLE(0, 3, 0, 1));
-
-                    _mm256_storeu_si256((__m256i*)outptr0, _tmp0);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + 8), _tmp1);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + 8 * 2), _tmp2);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + 8 * 3), _tmp3);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 4), _tmp4);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 4 + 8), _tmp5);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 4 + 8 * 2), _tmp6);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 4 + 8 * 3), _tmp7);
-#endif // __AVX512F__
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-#if __AVX512F__
-                    // from
-                    //      00 11 22 33 44 55 66 77 01 12 23 30 45 56 67 74
-                    //      02 13 20 31 46 57 64 75 03 10 21 32 47 54 65 76
-                    //      40 51 62 73 04 15 26 37 41 52 63 70 05 16 27 34
-                    //      42 53 60 71 06 17 24 35 43 50 61 72 07 14 25 36
-                    // to
-                    //      00 01 02 03 44 45 46 47 04 05 06 07 40 41 42 43
-                    //      10 11 12 13 54 55 56 57 14 15 16 17 50 51 52 53
-                    //      20 21 22 23 64 65 66 67 24 25 26 27 60 61 62 63
-                    //      30 31 32 33 74 75 76 77 34 35 36 37 70 71 72 73
-                    {
-                        __m512i _s0 = _mm512_shuffle_i32x4(_sum0, _sum2, _MM_SHUFFLE(0, 1, 1, 0));
-                        __m512i _s1 = _mm512_shuffle_i32x4(_sum0, _sum2, _MM_SHUFFLE(2, 3, 3, 2));
-                        __m512i _s2 = _mm512_shuffle_i32x4(_sum1, _sum3, _MM_SHUFFLE(0, 1, 1, 0));
-                        __m512i _s3 = _mm512_shuffle_i32x4(_sum1, _sum3, _MM_SHUFFLE(2, 3, 3, 2));
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_s0, _s1);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_s0, _s1);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_s2, _s3);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_s2, _s3);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                    }
-
-                    _sum0 = _mm512_shuffle_i32x4(_sum0, _sum0, _MM_SHUFFLE(1, 3, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_sum1, _sum1, _MM_SHUFFLE(1, 3, 2, 0));
-                    _sum2 = _mm512_shuffle_i32x4(_sum2, _sum2, _MM_SHUFFLE(1, 3, 2, 0));
-                    _sum3 = _mm512_shuffle_i32x4(_sum3, _sum3, _MM_SHUFFLE(1, 3, 2, 0));
-
-                    _mm256_storeu_si256((__m256i*)outptr0, _mm512_extracti32x8_epi32(_sum0, 0));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep), _mm512_extracti32x8_epi32(_sum1, 0));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 2), _mm512_extracti32x8_epi32(_sum2, 0));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 3), _mm512_extracti32x8_epi32(_sum3, 0));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 4), _mm512_extracti32x8_epi32(_sum0, 1));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 5), _mm512_extracti32x8_epi32(_sum1, 1));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 6), _mm512_extracti32x8_epi32(_sum2, 1));
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 7), _mm512_extracti32x8_epi32(_sum3, 1));
-#else
-                    // from
-                    //      00 11 22 33 44 55 66 77
-                    //      01 12 23 30 45 56 67 74
-                    //      02 13 20 31 46 57 64 75
-                    //      03 10 21 32 47 54 65 76
-                    //      40 51 62 73 04 15 26 37
-                    //      41 52 63 70 05 16 27 34
-                    //      42 53 60 71 06 17 24 35
-                    //      43 50 61 72 07 14 25 36
-                    // to
-                    //      00 01 02 03 44 45 46 47
-                    //      10 11 12 13 54 55 56 57
-                    //      20 21 22 23 64 65 66 67
-                    //      30 31 32 33 74 75 76 77
-                    //      40 41 42 43 04 05 06 07
-                    //      50 51 52 53 14 15 16 17
-                    //      60 61 62 63 24 25 26 27
-                    //      70 71 72 73 34 35 36 37
-                    {
-                        __m256i _tmp0 = _mm256_unpacklo_epi32(_sum0, _sum1);
-                        __m256i _tmp1 = _mm256_unpackhi_epi32(_sum0, _sum1);
-                        __m256i _tmp2 = _mm256_unpacklo_epi32(_sum2, _sum3);
-                        __m256i _tmp3 = _mm256_unpackhi_epi32(_sum2, _sum3);
-                        __m256i _tmp4 = _mm256_unpacklo_epi32(_sum4, _sum5);
-                        __m256i _tmp5 = _mm256_unpackhi_epi32(_sum4, _sum5);
-                        __m256i _tmp6 = _mm256_unpacklo_epi32(_sum6, _sum7);
-                        __m256i _tmp7 = _mm256_unpackhi_epi32(_sum6, _sum7);
-                        _sum0 = _mm256_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm256_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm256_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm256_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum4 = _mm256_unpacklo_epi64(_tmp4, _tmp6);
-                        _sum5 = _mm256_unpackhi_epi64(_tmp4, _tmp6);
-                        _sum6 = _mm256_unpacklo_epi64(_tmp7, _tmp5);
-                        _sum7 = _mm256_unpackhi_epi64(_tmp7, _tmp5);
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum5 = _mm256_shuffle_epi32(_sum5, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum7 = _mm256_shuffle_epi32(_sum7, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-
-                    __m256i _tmp0 = _mm256_permute2x128_si256(_sum0, _sum4, _MM_SHUFFLE(0, 3, 0, 0));
-                    __m256i _tmp1 = _mm256_permute2x128_si256(_sum1, _sum5, _MM_SHUFFLE(0, 3, 0, 0));
-                    __m256i _tmp2 = _mm256_permute2x128_si256(_sum2, _sum6, _MM_SHUFFLE(0, 3, 0, 0));
-                    __m256i _tmp3 = _mm256_permute2x128_si256(_sum3, _sum7, _MM_SHUFFLE(0, 3, 0, 0));
-                    __m256i _tmp4 = _mm256_permute2x128_si256(_sum4, _sum0, _MM_SHUFFLE(0, 3, 0, 0));
-                    __m256i _tmp5 = _mm256_permute2x128_si256(_sum5, _sum1, _MM_SHUFFLE(0, 3, 0, 0));
-                    __m256i _tmp6 = _mm256_permute2x128_si256(_sum6, _sum2, _MM_SHUFFLE(0, 3, 0, 0));
-                    __m256i _tmp7 = _mm256_permute2x128_si256(_sum7, _sum3, _MM_SHUFFLE(0, 3, 0, 0));
-
-                    _mm256_storeu_si256((__m256i*)outptr0, _tmp0);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep), _tmp1);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 2), _tmp2);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 3), _tmp3);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 4), _tmp4);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 5), _tmp5);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 6), _tmp6);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 7), _tmp7);
-#endif // __AVX512F__
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-#if __AVX512F__
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                _mm512_store_si512((__m512i*)(outptr + 16), _sum1);
-                _mm512_store_si512((__m512i*)(outptr + 32), _sum2);
-                _mm512_store_si512((__m512i*)(outptr + 48), _sum3);
-#else
-                _mm256_store_si256((__m256i*)outptr, _sum0);
-                _mm256_store_si256((__m256i*)(outptr + 8), _sum1);
-                _mm256_store_si256((__m256i*)(outptr + 16), _sum2);
-                _mm256_store_si256((__m256i*)(outptr + 24), _sum3);
-                _mm256_store_si256((__m256i*)(outptr + 32), _sum4);
-                _mm256_store_si256((__m256i*)(outptr + 40), _sum5);
-                _mm256_store_si256((__m256i*)(outptr + 48), _sum6);
-                _mm256_store_si256((__m256i*)(outptr + 56), _sum7);
-#endif // __AVX512F__
-            }
-
-            outptr += 64;
-        }
-#endif // defined(__x86_64__) || defined(_M_X64)
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            const signed char* pA = pAT;
-
-            __m256i _sum0;
-            __m256i _sum1;
-            __m256i _sum2;
-            __m256i _sum3;
-
-            if (k == 0)
-            {
-                _sum0 = _mm256_setzero_si256();
-                _sum1 = _mm256_setzero_si256();
-                _sum2 = _mm256_setzero_si256();
-                _sum3 = _mm256_setzero_si256();
-            }
-            else
-            {
-                _sum0 = _mm256_load_si256((const __m256i*)outptr);
-                _sum1 = _mm256_load_si256((const __m256i*)(outptr + 8));
-                _sum2 = _mm256_load_si256((const __m256i*)(outptr + 16));
-                _sum3 = _mm256_load_si256((const __m256i*)(outptr + 24));
-            }
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _pA = _mm_loadu_si128((const __m128i*)pA);
-                __m128i _pB = _mm_castpd_si128(_mm_load1_pd((const double*)pB));
-
-                __m256i _pA0 = _mm256_cvtepi8_epi16(_pA);
-                __m256i _pB0 = _mm256_cvtepi8_epi16(_pB);
-
-                // 0123 4567
-                // 2301 6745
-                __m256i _pA1 = _mm256_shuffle_epi32(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-
-                // 0123 0123
-                // 1230 1230
-                __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                _sum0 = _mm256_dpwssd_epi32(_sum0, _pA0, _pB0);
-                _sum1 = _mm256_dpwssd_epi32(_sum1, _pA0, _pB1);
-                _sum2 = _mm256_dpwssd_epi32(_sum2, _pA1, _pB0);
-                _sum3 = _mm256_dpwssd_epi32(_sum3, _pA1, _pB1);
-#else
-                _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_pA0, _pB0));
-                _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_pA0, _pB1));
-                _sum2 = _mm256_add_epi32(_sum2, _mm256_madd_epi16(_pA1, _pB0));
-                _sum3 = _mm256_add_epi32(_sum3, _mm256_madd_epi16(_pA1, _pB1));
-#endif
-
-                pA += 16;
-                pB += 8;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_loadl_epi64((const __m128i*)pA);
-                __m128i _pB = _mm_castps_si128(_mm_load1_ps((const float*)pB));
-
-                _pA = _mm_cvtepi8_epi16(_pA);
-                _pB = _mm_cvtepi8_epi16(_pB);
-
-                // 01234567
-                // 23016745
-                __m128i _pA0 = _pA;
-                __m128i _pA1 = _mm_shuffle_epi32(_pA, _MM_SHUFFLE(2, 3, 0, 1));
-
-                // 01230123
-                // 12301230
-                __m128i _pB0 = _pB;
-                __m128i _pB1 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_pB, _MM_SHUFFLE(0, 3, 2, 1)), _MM_SHUFFLE(0, 3, 2, 1));
-
-                __m256i _s0 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA0, _pB0));
-                __m256i _s1 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA0, _pB1));
-                __m256i _s2 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA1, _pB0));
-                __m256i _s3 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA1, _pB1));
-
-                _sum0 = _mm256_add_epi32(_sum0, _s0);
-                _sum1 = _mm256_add_epi32(_sum1, _s1);
-                _sum2 = _mm256_add_epi32(_sum2, _s2);
-                _sum3 = _mm256_add_epi32(_sum3, _s3);
-
-                pA += 8;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 8)
-                {
-                    // from
-                    //      00 11 22 33 40 51 62 73
-                    //      01 12 23 30 41 52 63 70
-                    //      20 31 02 13 60 71 42 53
-                    //      21 32 03 10 61 72 43 50
-                    // to
-                    //      00 10 20 30 40 50 60 70
-                    //      01 11 21 31 41 51 61 71
-                    //      02 12 22 32 42 52 62 72
-                    //      03 13 23 33 43 53 63 73
-                    {
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        __m256i _tmp0 = _mm256_unpacklo_epi32(_sum0, _sum3);
-                        __m256i _tmp1 = _mm256_unpackhi_epi32(_sum0, _sum3);
-                        __m256i _tmp2 = _mm256_unpacklo_epi32(_sum2, _sum1);
-                        __m256i _tmp3 = _mm256_unpackhi_epi32(_sum2, _sum1);
-                        _sum0 = _mm256_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm256_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm256_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm256_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-
-                    _mm256_store_si256((__m256i*)outptr0, _sum0);
-                    _mm256_store_si256((__m256i*)(outptr0 + 8), _sum1);
-                    _mm256_store_si256((__m256i*)(outptr0 + 16), _sum2);
-                    _mm256_store_si256((__m256i*)(outptr0 + 24), _sum3);
-                    outptr0 += 32;
-                }
-                if (out_elempack == 4)
-                {
-                    // from
-                    //      00 11 22 33 40 51 62 73
-                    //      01 12 23 30 41 52 63 70
-                    //      20 31 02 13 60 71 42 53
-                    //      21 32 03 10 61 72 43 50
-                    // to
-                    //      00 10 20 30 40 50 60 70
-                    //      01 11 21 31 41 51 61 71
-                    //      02 12 22 32 42 52 62 72
-                    //      03 13 23 33 43 53 63 73
-                    {
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        __m256i _tmp0 = _mm256_unpacklo_epi32(_sum0, _sum3);
-                        __m256i _tmp1 = _mm256_unpackhi_epi32(_sum0, _sum3);
-                        __m256i _tmp2 = _mm256_unpacklo_epi32(_sum2, _sum1);
-                        __m256i _tmp3 = _mm256_unpackhi_epi32(_sum2, _sum1);
-                        _sum0 = _mm256_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm256_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm256_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm256_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-
-                    __m256i _tmp0 = _mm256_permute2x128_si256(_sum0, _sum1, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp1 = _mm256_permute2x128_si256(_sum2, _sum3, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp2 = _mm256_permute2x128_si256(_sum0, _sum1, _MM_SHUFFLE(0, 3, 0, 1));
-                    __m256i _tmp3 = _mm256_permute2x128_si256(_sum2, _sum3, _MM_SHUFFLE(0, 3, 0, 1));
-
-                    _mm256_storeu_si256((__m256i*)outptr0, _tmp0);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + 8), _tmp1);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 4), _tmp2);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 4 + 8), _tmp3);
-
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    // from
-                    //      00 11 22 33 40 51 62 73
-                    //      01 12 23 30 41 52 63 70
-                    //      20 31 02 13 60 71 42 53
-                    //      21 32 03 10 61 72 43 50
-                    // to
-                    //      00 01 02 03 40 41 42 43
-                    //      10 11 12 13 50 51 52 53
-                    //      20 21 22 23 60 61 62 63
-                    //      30 31 32 33 70 71 72 73
-                    {
-                        _sum2 = _mm256_shuffle_epi32(_sum2, _MM_SHUFFLE(1, 0, 3, 2));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(1, 0, 3, 2));
-                        __m256i _tmp0 = _mm256_unpacklo_epi32(_sum0, _sum1);
-                        __m256i _tmp1 = _mm256_unpackhi_epi32(_sum0, _sum1);
-                        __m256i _tmp2 = _mm256_unpacklo_epi32(_sum2, _sum3);
-                        __m256i _tmp3 = _mm256_unpackhi_epi32(_sum2, _sum3);
-                        _sum0 = _mm256_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm256_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm256_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm256_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-
-                    _mm_storeu_si128((__m128i*)outptr0, _mm256_extracti128_si256(_sum0, 0));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep), _mm256_extracti128_si256(_sum1, 0));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 2), _mm256_extracti128_si256(_sum2, 0));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 3), _mm256_extracti128_si256(_sum3, 0));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 4), _mm256_extracti128_si256(_sum0, 1));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 5), _mm256_extracti128_si256(_sum1, 1));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 6), _mm256_extracti128_si256(_sum2, 1));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 7), _mm256_extracti128_si256(_sum3, 1));
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                _mm256_store_si256((__m256i*)outptr, _sum0);
-                _mm256_store_si256((__m256i*)(outptr + 8), _sum1);
-                _mm256_store_si256((__m256i*)(outptr + 16), _sum2);
-                _mm256_store_si256((__m256i*)(outptr + 24), _sum3);
-            }
-
-            outptr += 32;
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            const signed char* pA = pAT;
-
-            __m256i _sum0;
-            __m256i _sum1;
-
-            if (k == 0)
-            {
-                _sum0 = _mm256_setzero_si256();
-                _sum1 = _mm256_setzero_si256();
-            }
-            else
-            {
-                _sum0 = _mm256_load_si256((const __m256i*)outptr);
-                _sum1 = _mm256_load_si256((const __m256i*)(outptr + 8));
-            }
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _pA = _mm_loadu_si128((const __m128i*)pA);
-                __m128i _pB = _mm_castps_si128(_mm_load1_ps((const float*)pB));
-
-                __m256i _pA0 = _mm256_cvtepi8_epi16(_pA);
-                __m256i _pB0 = _mm256_cvtepi8_epi16(_pB);
-
-                // 0123 4567
-
-                // 0101 0101
-                // 1010 1010
-                __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 1, 0, 1));
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                _sum0 = _mm256_dpwssd_epi32(_sum0, _pA0, _pB0);
-                _sum1 = _mm256_dpwssd_epi32(_sum1, _pA0, _pB1);
-#else
-                _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_pA0, _pB0));
-                _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_pA0, _pB1));
-#endif
-
-                pA += 16;
-                pB += 4;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_loadl_epi64((const __m128i*)pA);
-                __m128i _pB = _mm_set1_epi16(((const short*)pB)[0]);
-
-                _pA = _mm_cvtepi8_epi16(_pA);
-                _pB = _mm_cvtepi8_epi16(_pB);
-
-                // 01234567
-
-                // 01010101
-                // 10101010
-                __m128i _pB0 = _pB;
-                __m128i _pB1 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_pB, _MM_SHUFFLE(0, 1, 0, 1)), _MM_SHUFFLE(0, 1, 0, 1));
-
-                __m256i _s0 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA, _pB0));
-                __m256i _s1 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA, _pB1));
-
-                _sum0 = _mm256_add_epi32(_sum0, _s0);
-                _sum1 = _mm256_add_epi32(_sum1, _s1);
-
-                pA += 8;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                // from
-                //      00 11 20 31 40 51 60 71
-                //      01 10 21 30 41 50 61 70
-                // to
-                //      00 10 20 30 40 50 60 70
-                //      01 11 21 31 41 51 61 71
-                {
-                    __m256i _tmp0 = _mm256_shuffle_epi32(_sum0, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m256i _tmp1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(0, 2, 3, 1));
-                    _sum0 = _mm256_unpacklo_epi32(_tmp0, _tmp1);
-                    _sum1 = _mm256_unpackhi_epi32(_tmp0, _tmp1);
-                    _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                }
-
-                if (out_elempack == 8)
-                {
-                    _mm256_store_si256((__m256i*)outptr0, _sum0);
-                    _mm256_store_si256((__m256i*)(outptr0 + 8), _sum1);
-                    outptr0 += 16;
-                }
-                if (out_elempack == 4)
-                {
-                    __m256i _tmp0 = _mm256_permute2x128_si256(_sum0, _sum1, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp1 = _mm256_permute2x128_si256(_sum0, _sum1, _MM_SHUFFLE(0, 3, 0, 1));
-
-                    _mm256_storeu_si256((__m256i*)outptr0, _tmp0);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 4), _tmp1);
-
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-#if __AVX512F__
-                    __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                    _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(out_hstep));
-                    _mm256_i32scatter_epi32(outptr0, _vindex, _sum0, sizeof(float));
-                    _mm256_i32scatter_epi32(outptr0 + 1, _vindex, _sum1, sizeof(float));
-#else
-                    int sum0[8];
-                    int sum1[8];
-                    _mm256_storeu_si256((__m256i*)sum0, _sum0);
-                    _mm256_storeu_si256((__m256i*)sum1, _sum1);
-
-                    outptr0[0] = sum0[0];
-                    outptr0[1] = sum1[0];
-                    outptr0[out_hstep * 1] = sum0[1];
-                    outptr0[out_hstep * 1 + 1] = sum1[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 2 + 1] = sum1[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[out_hstep * 3 + 1] = sum1[3];
-                    outptr0[out_hstep * 4] = sum0[4];
-                    outptr0[out_hstep * 4 + 1] = sum1[4];
-                    outptr0[out_hstep * 5] = sum0[5];
-                    outptr0[out_hstep * 5 + 1] = sum1[5];
-                    outptr0[out_hstep * 6] = sum0[6];
-                    outptr0[out_hstep * 6 + 1] = sum1[6];
-                    outptr0[out_hstep * 7] = sum0[7];
-                    outptr0[out_hstep * 7 + 1] = sum1[7];
-#endif // __AVX512F__
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                _mm256_store_si256((__m256i*)outptr, _sum0);
-                _mm256_store_si256((__m256i*)(outptr + 8), _sum1);
-            }
-
-            outptr += 16;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            const signed char* pA = pAT;
-
-            __m256i _sum0;
-
-            if (k == 0)
-            {
-                _sum0 = _mm256_setzero_si256();
-            }
-            else
-            {
-                _sum0 = _mm256_load_si256((const __m256i*)outptr);
-            }
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _pA = _mm_loadu_si128((const __m128i*)pA);
-                __m128i _pB = _mm_castps_si128(_mm_load1_ps((const float*)pB));
-
-                __m256i _pA0 = _mm256_cvtepi8_epi16(_pA);
-                __m256i _pBB = _mm256_cvtepi8_epi16(_pB);
-
-                // 0xxx0xxx -> 00000000 11111111
-                __m256i _pB0 = _mm256_shuffle_epi32(_pBB, _MM_SHUFFLE(0, 0, 0, 0));
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                _sum0 = _mm256_dpwssd_epi32(_sum0, _pA0, _pB0);
-#else
-                _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_pA0, _pB0));
-#endif
-
-                pA += 16;
-                pB += 2;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_loadl_epi64((const __m128i*)pA);
-                __m128i _pB = _mm_set1_epi16(pB[0]);
-
-                _pA = _mm_cvtepi8_epi16(_pA);
-
-                __m256i _s0 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA, _pB));
-
-                _sum0 = _mm256_add_epi32(_sum0, _s0);
-
-                pA += 8;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 8)
-                {
-                    _mm256_store_si256((__m256i*)outptr0, _sum0);
-                    outptr0 += 8;
-                }
-                if (out_elempack == 4)
-                {
-                    _mm_storeu_si128((__m128i*)outptr0, _mm256_extracti128_si256(_sum0, 0));
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 4), _mm256_extracti128_si256(_sum0, 1));
-                    outptr0 += 4;
-                }
-                if (out_elempack == 1)
-                {
-#if __AVX512F__
-                    __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                    _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(out_hstep));
-                    _mm256_i32scatter_epi32(outptr0, _vindex, _sum0, sizeof(float));
-#else
-                    int sum0[8];
-                    _mm256_storeu_si256((__m256i*)sum0, _sum0);
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep * 1] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[out_hstep * 4] = sum0[4];
-                    outptr0[out_hstep * 5] = sum0[5];
-                    outptr0[out_hstep * 6] = sum0[6];
-                    outptr0[out_hstep * 7] = sum0[7];
-#endif // __AVX512F__
-                    outptr0++;
-                }
-            }
-            else
-            {
-                _mm256_store_si256((__m256i*)outptr, _sum0);
-            }
-
-            outptr += 8;
-        }
-
-        pAT += max_kk * 8;
-    }
-#endif // __AVX2__
-    for (; ii + 3 < max_ii; ii += 4)
-    {
-        int* outptr0 = (int*)top_blob + (i + ii) * out_hstep + j * out_elempack;
-
-        const signed char* pB = pBT;
-
-        int jj = 0;
-#if defined(__x86_64__) || defined(_M_X64)
-#if __AVX512F__
-        for (; jj + 15 < max_jj; jj += 16)
-        {
-            const signed char* pA = pAT;
-
-            __m512i _sum0;
-            __m512i _sum1;
-            __m512i _sum2;
-            __m512i _sum3;
-
-            if (k == 0)
-            {
-                _sum0 = _mm512_setzero_si512();
-                _sum1 = _mm512_setzero_si512();
-                _sum2 = _mm512_setzero_si512();
-                _sum3 = _mm512_setzero_si512();
-            }
-            else
-            {
-                _sum0 = _mm512_load_si512((const __m512i*)outptr);
-                _sum1 = _mm512_load_si512((const __m512i*)(outptr + 16));
-                _sum2 = _mm512_load_si512((const __m512i*)(outptr + 32));
-                _sum3 = _mm512_load_si512((const __m512i*)(outptr + 48));
-            }
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m256i _pA = _mm256_castpd_si256(_mm256_broadcast_sd((const double*)pA));
-                __m256i _pB = _mm256_loadu_si256((const __m256i*)pB);
-
-                __m512i _pA0 = _mm512_cvtepi8_epi16(_pA);
-                __m512i _pB0 = _mm512_cvtepi8_epi16(_pB);
-
-                // 0123 0123 0123 0123
-                // 2301 2301 2301 2301
-                __m512i _pA1 = _mm512_shuffle_epi32(_pA0, _MM_PERM_BADC);
-
-                // 0123 4567 89ab cdef
-                // 1230 5674 9ab8 defc
-                __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-
-#if __AVX512VNNI__
-                _sum0 = _mm512_dpwssd_epi32(_sum0, _pA0, _pB0);
-                _sum1 = _mm512_dpwssd_epi32(_sum1, _pA0, _pB1);
-                _sum2 = _mm512_dpwssd_epi32(_sum2, _pA1, _pB0);
-                _sum3 = _mm512_dpwssd_epi32(_sum3, _pA1, _pB1);
-#else
-                _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA0, _pB0));
-                _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA0, _pB1));
-                _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_pA1, _pB0));
-                _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_pA1, _pB1));
-#endif
-
-                pA += 8;
-                pB += 32;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_castps_si128(_mm_load1_ps((const float*)pA));
-                __m128i _pB = _mm_load_si128((const __m128i*)pB);
-
-                __m256i _pA0 = _mm256_cvtepi8_epi16(_pA);
-                __m256i _pB0 = _mm256_cvtepi8_epi16(_pB);
-
-                // 01230123 01230123
-                // 23012301 23012301
-                __m256i _pA1 = _mm256_shuffle_epi32(_pA0, _MM_SHUFFLE(2, 3, 0, 1));
-
-                // 01234567 89abcdef
-                // 12305674 9ab8defc
-                __m256i _pB1 = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(_pB0, _MM_SHUFFLE(0, 3, 2, 1)), _MM_SHUFFLE(0, 3, 2, 1));
-
-                __m512i _s0 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB0));
-                __m512i _s1 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB1));
-                __m512i _s2 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA1, _pB0));
-                __m512i _s3 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA1, _pB1));
-
-                _sum0 = _mm512_add_epi32(_sum0, _s0);
-                _sum1 = _mm512_add_epi32(_sum1, _s1);
-                _sum2 = _mm512_add_epi32(_sum2, _s2);
-                _sum3 = _mm512_add_epi32(_sum3, _s3);
-
-                pA += 4;
-                pB += 16;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    // from
-                    //      00 11 22 33 04 15 26 37 08 19 2a 3b 0c 1d 2e 3f
-                    //      01 12 23 30 05 16 27 34 09 1a 2b 38 0d 1e 2f 3c
-                    //      20 31 02 13 24 35 06 17 28 3a 0a 1b 2c 3d 0e 1f
-                    //      21 32 03 10 25 36 07 14 29 3a 0b 18 2d 3e 0f 1c
-                    // to
-                    //      00 10 20 30 04 14 24 34 08 18 28 38 0c 1c 2c 3c
-                    //      01 11 21 31 05 15 25 35 09 19 29 39 0d 1d 2d 3d
-                    //      02 12 22 32 06 16 26 36 0a 1a 2a 3a 0e 1e 2e 3e
-                    //      03 13 23 33 07 17 27 37 0b 1b 2b 3b 0f 1f 2f 3f
-                    {
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum3);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum3);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum1);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum1);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                    }
-
-                    __m512i _tmp0 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp1 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(1, 0, 1, 0));
-                    __m512i _tmp2 = _mm512_shuffle_i32x4(_sum0, _sum1, _MM_SHUFFLE(3, 2, 3, 2));
-                    __m512i _tmp3 = _mm512_shuffle_i32x4(_sum2, _sum3, _MM_SHUFFLE(3, 2, 3, 2));
-                    _sum0 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum1 = _mm512_shuffle_i32x4(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-                    _sum2 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-                    _sum3 = _mm512_shuffle_i32x4(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _sum0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 16), _sum1);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 32), _sum2);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + 48), _sum3);
-                    outptr0 += 64;
-                }
-                if (out_elempack == 1)
-                {
-                    // from
-                    //      00 11 22 33 04 15 26 37 08 19 2a 3b 0c 1d 2e 3f
-                    //      01 12 23 30 05 16 27 34 09 1a 2b 38 0d 1e 2f 3c
-                    //      20 31 02 13 24 35 06 17 28 3a 0a 1b 2c 3d 0e 1f
-                    //      21 32 03 10 25 36 07 14 29 3a 0b 18 2d 3e 0f 1c
-                    // to
-                    //      00 01 02 03 04 05 06 07 08 09 0a 0b 0c 0d 0e 0f
-                    //      10 11 12 13 14 15 16 17 18 19 1a 1b 1c 1d 1e 1f
-                    //      20 21 22 23 24 25 26 27 28 29 2a 2b 2c 2d 2e 2f
-                    //      30 31 32 33 34 35 36 37 38 39 3a 3b 3c 3d 3e 3f
-                    {
-                        _sum2 = _mm512_shuffle_epi32(_sum2, _MM_PERM_BADC);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_BADC);
-                        __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum1);
-                        __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum1);
-                        __m512i _tmp2 = _mm512_unpacklo_epi32(_sum2, _sum3);
-                        __m512i _tmp3 = _mm512_unpackhi_epi32(_sum2, _sum3);
-                        _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm512_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm512_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-                        _sum3 = _mm512_shuffle_epi32(_sum3, _MM_PERM_CBAD);
-                    }
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _sum0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep), _sum1);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 2), _sum2);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep * 3), _sum3);
-
-                    outptr0 += 16;
-                }
-            }
-            else
-            {
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                _mm512_store_si512((__m512i*)(outptr + 16), _sum1);
-                _mm512_store_si512((__m512i*)(outptr + 32), _sum2);
-                _mm512_store_si512((__m512i*)(outptr + 48), _sum3);
-            }
-
-            outptr += 64;
-        }
-#endif // __AVX512F__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-            const signed char* pA = pAT;
-
-#if __AVX2__
-            __m256i _sum0;
-            __m256i _sum1;
-            __m256i _sum2;
-            __m256i _sum3;
-#else
-            __m128i _sum0;
-            __m128i _sum1;
-            __m128i _sum2;
-            __m128i _sum3;
-            __m128i _sum4;
-            __m128i _sum5;
-            __m128i _sum6;
-            __m128i _sum7;
-#endif
-
-            if (k == 0)
-            {
-#if __AVX2__
-                _sum0 = _mm256_setzero_si256();
-                _sum1 = _mm256_setzero_si256();
-                _sum2 = _mm256_setzero_si256();
-                _sum3 = _mm256_setzero_si256();
-#else
-                _sum0 = _mm_setzero_si128();
-                _sum1 = _mm_setzero_si128();
-                _sum2 = _mm_setzero_si128();
-                _sum3 = _mm_setzero_si128();
-                _sum4 = _mm_setzero_si128();
-                _sum5 = _mm_setzero_si128();
-                _sum6 = _mm_setzero_si128();
-                _sum7 = _mm_setzero_si128();
-#endif
-            }
-            else
-            {
-#if __AVX2__
-                _sum0 = _mm256_load_si256((const __m256i*)outptr);
-                _sum1 = _mm256_load_si256((const __m256i*)(outptr + 8));
-                _sum2 = _mm256_load_si256((const __m256i*)(outptr + 16));
-                _sum3 = _mm256_load_si256((const __m256i*)(outptr + 24));
-#else
-                _sum0 = _mm_load_si128((const __m128i*)outptr);
-                _sum1 = _mm_load_si128((const __m128i*)(outptr + 4));
-                _sum2 = _mm_load_si128((const __m128i*)(outptr + 8));
-                _sum3 = _mm_load_si128((const __m128i*)(outptr + 12));
-                _sum4 = _mm_load_si128((const __m128i*)(outptr + 16));
-                _sum5 = _mm_load_si128((const __m128i*)(outptr + 20));
-                _sum6 = _mm_load_si128((const __m128i*)(outptr + 24));
-                _sum7 = _mm_load_si128((const __m128i*)(outptr + 28));
-#endif
-            }
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _pA = _mm_castpd_si128(_mm_load1_pd((const double*)pA));
-                __m128i _pB = _mm_loadu_si128((const __m128i*)pB);
-
-#if __AVX2__
-                __m256i _pA0 = _mm256_cvtepi8_epi16(_pA);
-                __m256i _pB0 = _mm256_cvtepi8_epi16(_pB);
-
-                // 0123 0123
-                // 2301 2301
-                __m256i _pA1 = _mm256_shuffle_epi32(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-
-                // 0123 4567
-                // 1230 5674
-                __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                _sum0 = _mm256_dpwssd_epi32(_sum0, _pA0, _pB0);
-                _sum1 = _mm256_dpwssd_epi32(_sum1, _pA0, _pB1);
-                _sum2 = _mm256_dpwssd_epi32(_sum2, _pA1, _pB0);
-                _sum3 = _mm256_dpwssd_epi32(_sum3, _pA1, _pB1);
-#else
-                _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_pA0, _pB0));
-                _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_pA0, _pB1));
-                _sum2 = _mm256_add_epi32(_sum2, _mm256_madd_epi16(_pA1, _pB0));
-                _sum3 = _mm256_add_epi32(_sum3, _mm256_madd_epi16(_pA1, _pB1));
-#endif
-#else // __AVX2__
-#if __SSE4_1__
-                _pA = _mm_cvtepi8_epi16(_pA);
-#else
-                _pA = _mm_unpacklo_epi8(_pA, _mm_cmpgt_epi8(_mm_setzero_si128(), _pA));
-#endif
-                __m128i _extpB = _mm_cmpgt_epi8(_mm_setzero_si128(), _pB);
-                __m128i _pBl = _mm_unpacklo_epi8(_pB, _extpB);
-                __m128i _pBh = _mm_unpackhi_epi8(_pB, _extpB);
-
-                // 0123
-                // 2301
-                __m128i _pA0 = _pA;
-                __m128i _pA1 = _mm_shuffle_epi32(_pA, _MM_SHUFFLE(1, 0, 3, 2));
-
-                // 0123
-                // 4567
-                // 1230
-                // 5674
-                __m128i _pB0 = _pBl;
-                __m128i _pB1 = _pBh;
-                __m128i _pB2 = _mm_shuffle_epi32(_pBl, _MM_SHUFFLE(0, 3, 2, 1));
-                __m128i _pB3 = _mm_shuffle_epi32(_pBh, _MM_SHUFFLE(0, 3, 2, 1));
-
-#if __XOP__
-                _sum0 = _mm_maddd_epi16(_pA0, _pB0, _sum0);
-                _sum1 = _mm_maddd_epi16(_pA0, _pB1, _sum1);
-                _sum2 = _mm_maddd_epi16(_pA0, _pB2, _sum2);
-                _sum3 = _mm_maddd_epi16(_pA0, _pB3, _sum3);
-                _sum4 = _mm_maddd_epi16(_pA1, _pB0, _sum4);
-                _sum5 = _mm_maddd_epi16(_pA1, _pB1, _sum5);
-                _sum6 = _mm_maddd_epi16(_pA1, _pB2, _sum6);
-                _sum7 = _mm_maddd_epi16(_pA1, _pB3, _sum7);
-#else
-                _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA0, _pB0));
-                _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_pA0, _pB1));
-                _sum2 = _mm_add_epi32(_sum2, _mm_madd_epi16(_pA0, _pB2));
-                _sum3 = _mm_add_epi32(_sum3, _mm_madd_epi16(_pA0, _pB3));
-                _sum4 = _mm_add_epi32(_sum4, _mm_madd_epi16(_pA1, _pB0));
-                _sum5 = _mm_add_epi32(_sum5, _mm_madd_epi16(_pA1, _pB1));
-                _sum6 = _mm_add_epi32(_sum6, _mm_madd_epi16(_pA1, _pB2));
-                _sum7 = _mm_add_epi32(_sum7, _mm_madd_epi16(_pA1, _pB3));
-#endif
-#endif // __AVX2__
-
-                pA += 8;
-                pB += 16;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_castps_si128(_mm_load1_ps((const float*)pA));
-                __m128i _pB = _mm_loadl_epi64((const __m128i*)pB);
-
-#if __SSE4_1__
-                _pA = _mm_cvtepi8_epi16(_pA);
-                _pB = _mm_cvtepi8_epi16(_pB);
-#else
-                _pA = _mm_unpacklo_epi8(_pA, _mm_cmpgt_epi8(_mm_setzero_si128(), _pA));
-                _pB = _mm_unpacklo_epi8(_pB, _mm_cmpgt_epi8(_mm_setzero_si128(), _pB));
-#endif
-
-#if __AVX2__
-                // 01230123
-                // 23012301
-                __m128i _pA0 = _pA;
-                __m128i _pA1 = _mm_shuffle_epi32(_pA, _MM_SHUFFLE(2, 3, 0, 1));
-
-                // 01234567
-                // 12305674
-                __m128i _pB0 = _pB;
-                __m128i _pB1 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_pB, _MM_SHUFFLE(0, 3, 2, 1)), _MM_SHUFFLE(0, 3, 2, 1));
-
-                __m256i _s0 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA0, _pB0));
-                __m256i _s1 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA0, _pB1));
-                __m256i _s2 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA1, _pB0));
-                __m256i _s3 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA1, _pB1));
-
-                _sum0 = _mm256_add_epi32(_sum0, _s0);
-                _sum1 = _mm256_add_epi32(_sum1, _s1);
-                _sum2 = _mm256_add_epi32(_sum2, _s2);
-                _sum3 = _mm256_add_epi32(_sum3, _s3);
-#else // __AVX2__
-#if __XOP__
-                // 00112233
-                // 22330011
-                __m128i _pA0 = _mm_unpacklo_epi16(_pA, _pA);
-                __m128i _pA1 = _mm_shuffle_epi32(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-
-                // 00112233
-                // 44556677
-                // 1.2.3.0.
-                __m128i _pB0 = _mm_unpacklo_epi16(_pB, _pB);
-                __m128i _pB1 = _mm_unpackhi_epi16(_pB, _pB);
-                __m128i _pB2 = _mm_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-                __m128i _pB3 = _mm_shuffle_epi32(_pB1, _MM_SHUFFLE(0, 3, 2, 1));
-
-                _sum0 = _mm_maccd_epi16(_pA0, _pB0, _sum0);
-                _sum1 = _mm_maccd_epi16(_pA0, _pB1, _sum1);
-                _sum2 = _mm_maccd_epi16(_pA0, _pB2, _sum2);
-                _sum3 = _mm_maccd_epi16(_pA0, _pB3, _sum3);
-                _sum4 = _mm_maccd_epi16(_pA1, _pB0, _sum4);
-                _sum5 = _mm_maccd_epi16(_pA1, _pB1, _sum5);
-                _sum6 = _mm_maccd_epi16(_pA1, _pB2, _sum6);
-                _sum7 = _mm_maccd_epi16(_pA1, _pB3, _sum7);
-#else
-                // 01230123
-                // 23012301
-                __m128i _pA0 = _pA;
-                __m128i _pA1 = _mm_shuffle_epi32(_pA, _MM_SHUFFLE(2, 3, 0, 1));
-
-                // 01234567
-                // 12305674
-                __m128i _pB01 = _pB;
-                __m128i _pB23 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_pB, _MM_SHUFFLE(0, 3, 2, 1)), _MM_SHUFFLE(0, 3, 2, 1));
-
-                __m128i _sl0 = _mm_mullo_epi16(_pA0, _pB01);
-                __m128i _sh0 = _mm_mulhi_epi16(_pA0, _pB01);
-                __m128i _sl1 = _mm_mullo_epi16(_pA0, _pB23);
-                __m128i _sh1 = _mm_mulhi_epi16(_pA0, _pB23);
-                __m128i _sl2 = _mm_mullo_epi16(_pA1, _pB01);
-                __m128i _sh2 = _mm_mulhi_epi16(_pA1, _pB01);
-                __m128i _sl3 = _mm_mullo_epi16(_pA1, _pB23);
-                __m128i _sh3 = _mm_mulhi_epi16(_pA1, _pB23);
-                __m128i _s0 = _mm_unpacklo_epi16(_sl0, _sh0);
-                __m128i _s1 = _mm_unpackhi_epi16(_sl0, _sh0);
-                __m128i _s2 = _mm_unpacklo_epi16(_sl1, _sh1);
-                __m128i _s3 = _mm_unpackhi_epi16(_sl1, _sh1);
-                __m128i _s4 = _mm_unpacklo_epi16(_sl2, _sh2);
-                __m128i _s5 = _mm_unpackhi_epi16(_sl2, _sh2);
-                __m128i _s6 = _mm_unpacklo_epi16(_sl3, _sh3);
-                __m128i _s7 = _mm_unpackhi_epi16(_sl3, _sh3);
-
-                _sum0 = _mm_add_epi32(_sum0, _s0);
-                _sum1 = _mm_add_epi32(_sum1, _s1);
-                _sum2 = _mm_add_epi32(_sum2, _s2);
-                _sum3 = _mm_add_epi32(_sum3, _s3);
-                _sum4 = _mm_add_epi32(_sum4, _s4);
-                _sum5 = _mm_add_epi32(_sum5, _s5);
-                _sum6 = _mm_add_epi32(_sum6, _s6);
-                _sum7 = _mm_add_epi32(_sum7, _s7);
-#endif
-#endif // __AVX2__
-
-                pA += 4;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-#if __AVX2__
-                    // from
-                    //      00 11 22 33 04 15 26 37
-                    //      01 12 23 30 05 16 27 34
-                    //      20 31 02 13 24 35 06 17
-                    //      21 32 03 10 25 36 07 14
-                    // to
-                    //      00 10 20 30 04 14 24 34
-                    //      01 11 21 31 05 15 25 35
-                    //      02 12 22 32 06 16 26 36
-                    //      03 13 23 33 07 17 27 37
-                    {
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        __m256i _tmp0 = _mm256_unpacklo_epi32(_sum0, _sum3);
-                        __m256i _tmp1 = _mm256_unpackhi_epi32(_sum0, _sum3);
-                        __m256i _tmp2 = _mm256_unpacklo_epi32(_sum2, _sum1);
-                        __m256i _tmp3 = _mm256_unpackhi_epi32(_sum2, _sum1);
-                        _sum0 = _mm256_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm256_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm256_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm256_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-
-                    __m256i _tmp0 = _mm256_permute2x128_si256(_sum0, _sum1, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp1 = _mm256_permute2x128_si256(_sum2, _sum3, _MM_SHUFFLE(0, 2, 0, 0));
-                    __m256i _tmp2 = _mm256_permute2x128_si256(_sum0, _sum1, _MM_SHUFFLE(0, 3, 0, 1));
-                    __m256i _tmp3 = _mm256_permute2x128_si256(_sum2, _sum3, _MM_SHUFFLE(0, 3, 0, 1));
-
-                    _mm256_storeu_si256((__m256i*)outptr0, _tmp0);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + 8), _tmp1);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + 16), _tmp2);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + 24), _tmp3);
-#else
-                    // from
-                    //      00 11 22 33  04 15 26 37
-                    //      01 12 23 30  05 16 27 34
-                    //      20 31 02 13  24 35 06 17
-                    //      21 32 03 10  25 36 07 14
-                    // to
-                    //      00 10 20 30  04 14 24 34
-                    //      01 11 21 31  05 15 25 35
-                    //      02 12 22 32  06 16 26 36
-                    //      03 13 23 33  07 17 27 37
-                    {
-                        _sum2 = _mm_shuffle_epi32(_sum2, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum6 = _mm_shuffle_epi32(_sum6, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum7 = _mm_shuffle_epi32(_sum7, _MM_SHUFFLE(2, 1, 0, 3));
-                        __m128i _tmp0 = _mm_unpacklo_epi32(_sum0, _sum6);
-                        __m128i _tmp1 = _mm_unpackhi_epi32(_sum0, _sum6);
-                        __m128i _tmp2 = _mm_unpacklo_epi32(_sum1, _sum7);
-                        __m128i _tmp3 = _mm_unpackhi_epi32(_sum1, _sum7);
-                        __m128i _tmp4 = _mm_unpacklo_epi32(_sum4, _sum2);
-                        __m128i _tmp5 = _mm_unpackhi_epi32(_sum4, _sum2);
-                        __m128i _tmp6 = _mm_unpacklo_epi32(_sum5, _sum3);
-                        __m128i _tmp7 = _mm_unpackhi_epi32(_sum5, _sum3);
-                        _sum0 = _mm_unpacklo_epi64(_tmp0, _tmp4);
-                        _sum1 = _mm_unpacklo_epi64(_tmp2, _tmp6);
-                        _sum2 = _mm_unpackhi_epi64(_tmp0, _tmp4);
-                        _sum3 = _mm_unpackhi_epi64(_tmp2, _tmp6);
-                        _sum4 = _mm_unpacklo_epi64(_tmp5, _tmp1);
-                        _sum5 = _mm_unpacklo_epi64(_tmp7, _tmp3);
-                        _sum6 = _mm_unpackhi_epi64(_tmp5, _tmp1);
-                        _sum7 = _mm_unpackhi_epi64(_tmp7, _tmp3);
-                        _sum2 = _mm_shuffle_epi32(_sum2, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum6 = _mm_shuffle_epi32(_sum6, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum7 = _mm_shuffle_epi32(_sum7, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-
-                    _mm_store_si128((__m128i*)outptr0, _sum0);
-                    _mm_store_si128((__m128i*)(outptr0 + 4), _sum2);
-                    _mm_store_si128((__m128i*)(outptr0 + 8), _sum4);
-                    _mm_store_si128((__m128i*)(outptr0 + 12), _sum6);
-                    _mm_store_si128((__m128i*)(outptr0 + 16), _sum1);
-                    _mm_store_si128((__m128i*)(outptr0 + 20), _sum3);
-                    _mm_store_si128((__m128i*)(outptr0 + 24), _sum5);
-                    _mm_store_si128((__m128i*)(outptr0 + 28), _sum7);
-#endif // __AVX2__
-                    outptr0 += 32;
-                }
-                if (out_elempack == 1)
-                {
-#if __AVX2__
-                    // from
-                    //      00 11 22 33 04 15 26 37
-                    //      01 12 23 30 05 16 27 34
-                    //      20 31 02 13 24 35 06 17
-                    //      21 32 03 10 25 36 07 14
-                    // to
-                    //      00 01 02 03 04 05 06 07
-                    //      10 11 12 13 14 15 16 17
-                    //      20 21 22 23 24 25 26 27
-                    //      30 31 32 33 34 35 36 37
-                    {
-                        _sum2 = _mm256_shuffle_epi32(_sum2, _MM_SHUFFLE(1, 0, 3, 2));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(1, 0, 3, 2));
-                        __m256i _tmp0 = _mm256_unpacklo_epi32(_sum0, _sum1);
-                        __m256i _tmp1 = _mm256_unpackhi_epi32(_sum0, _sum1);
-                        __m256i _tmp2 = _mm256_unpacklo_epi32(_sum2, _sum3);
-                        __m256i _tmp3 = _mm256_unpackhi_epi32(_sum2, _sum3);
-                        _sum0 = _mm256_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm256_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm256_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm256_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm256_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-
-                    _mm256_storeu_si256((__m256i*)outptr0, _sum0);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep), _sum1);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 2), _sum2);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep * 3), _sum3);
-#else
-                    // from
-                    //      00 11 22 33  04 15 26 37
-                    //      01 12 23 30  05 16 27 34
-                    //      20 31 02 13  24 35 06 17
-                    //      21 32 03 10  25 36 07 14
-                    // to
-                    //      00 01 02 03  04 05 06 07
-                    //      10 11 12 13  14 15 16 17
-                    //      20 21 22 23  24 25 26 27
-                    //      30 31 32 33  34 35 36 37
-                    {
-                        _sum4 = _mm_shuffle_epi32(_sum4, _MM_SHUFFLE(1, 0, 3, 2));
-                        _sum5 = _mm_shuffle_epi32(_sum5, _MM_SHUFFLE(1, 0, 3, 2));
-                        _sum6 = _mm_shuffle_epi32(_sum6, _MM_SHUFFLE(1, 0, 3, 2));
-                        _sum7 = _mm_shuffle_epi32(_sum7, _MM_SHUFFLE(1, 0, 3, 2));
-                        __m128i _tmp0 = _mm_unpacklo_epi32(_sum0, _sum2);
-                        __m128i _tmp1 = _mm_unpackhi_epi32(_sum0, _sum2);
-                        __m128i _tmp2 = _mm_unpacklo_epi32(_sum1, _sum3);
-                        __m128i _tmp3 = _mm_unpackhi_epi32(_sum1, _sum3);
-                        __m128i _tmp4 = _mm_unpacklo_epi32(_sum4, _sum6);
-                        __m128i _tmp5 = _mm_unpackhi_epi32(_sum4, _sum6);
-                        __m128i _tmp6 = _mm_unpacklo_epi32(_sum5, _sum7);
-                        __m128i _tmp7 = _mm_unpackhi_epi32(_sum5, _sum7);
-                        _sum0 = _mm_unpacklo_epi64(_tmp0, _tmp4);
-                        _sum1 = _mm_unpacklo_epi64(_tmp2, _tmp6);
-                        _sum2 = _mm_unpackhi_epi64(_tmp0, _tmp4);
-                        _sum3 = _mm_unpackhi_epi64(_tmp2, _tmp6);
-                        _sum4 = _mm_unpacklo_epi64(_tmp5, _tmp1);
-                        _sum5 = _mm_unpacklo_epi64(_tmp7, _tmp3);
-                        _sum6 = _mm_unpackhi_epi64(_tmp5, _tmp1);
-                        _sum7 = _mm_unpackhi_epi64(_tmp7, _tmp3);
-                        _sum2 = _mm_shuffle_epi32(_sum2, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum6 = _mm_shuffle_epi32(_sum6, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum7 = _mm_shuffle_epi32(_sum7, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-
-                    _mm_storeu_si128((__m128i*)outptr0, _sum0);
-                    _mm_storeu_si128((__m128i*)(outptr0 + 4), _sum1);
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep), _sum2);
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep + 4), _sum3);
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 2), _sum4);
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 2 + 4), _sum5);
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 3), _sum6);
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 3 + 4), _sum7);
-#endif // __AVX2__
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-#if __AVX2__
-                _mm256_store_si256((__m256i*)outptr, _sum0);
-                _mm256_store_si256((__m256i*)(outptr + 8), _sum1);
-                _mm256_store_si256((__m256i*)(outptr + 16), _sum2);
-                _mm256_store_si256((__m256i*)(outptr + 24), _sum3);
-#else
-                _mm_store_si128((__m128i*)outptr, _sum0);
-                _mm_store_si128((__m128i*)(outptr + 4), _sum1);
-                _mm_store_si128((__m128i*)(outptr + 8), _sum2);
-                _mm_store_si128((__m128i*)(outptr + 12), _sum3);
-                _mm_store_si128((__m128i*)(outptr + 16), _sum4);
-                _mm_store_si128((__m128i*)(outptr + 20), _sum5);
-                _mm_store_si128((__m128i*)(outptr + 24), _sum6);
-                _mm_store_si128((__m128i*)(outptr + 28), _sum7);
-#endif
-            }
-
-            outptr += 32;
-        }
-#endif // defined(__x86_64__) || defined(_M_X64)
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            const signed char* pA = pAT;
-
-            __m128i _sum0;
-            __m128i _sum1;
-            __m128i _sum2;
-            __m128i _sum3;
-
-            if (k == 0)
-            {
-                _sum0 = _mm_setzero_si128();
-                _sum1 = _mm_setzero_si128();
-                _sum2 = _mm_setzero_si128();
-                _sum3 = _mm_setzero_si128();
-            }
-            else
-            {
-                _sum0 = _mm_load_si128((const __m128i*)outptr);
-                _sum1 = _mm_load_si128((const __m128i*)(outptr + 4));
-                _sum2 = _mm_load_si128((const __m128i*)(outptr + 8));
-                _sum3 = _mm_load_si128((const __m128i*)(outptr + 12));
-            }
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _pA = _mm_loadl_epi64((const __m128i*)pA);
-                __m128i _pB = _mm_loadl_epi64((const __m128i*)pB);
-
-#if __SSE4_1__
-                _pA = _mm_cvtepi8_epi16(_pA);
-                _pB = _mm_cvtepi8_epi16(_pB);
-#else
-                _pA = _mm_unpacklo_epi8(_pA, _mm_cmpgt_epi8(_mm_setzero_si128(), _pA));
-                _pB = _mm_unpacklo_epi8(_pB, _mm_cmpgt_epi8(_mm_setzero_si128(), _pB));
-#endif
-
-                // 0123
-                // 2301
-                __m128i _pA0 = _pA;
-                __m128i _pA1 = _mm_shuffle_epi32(_pA, _MM_SHUFFLE(1, 0, 3, 2));
-
-                // 0123
-                // 1230
-                __m128i _pB0 = _pB;
-                __m128i _pB1 = _mm_shuffle_epi32(_pB, _MM_SHUFFLE(0, 3, 2, 1));
-
-#if __XOP__
-                _sum0 = _mm_maddd_epi16(_pA0, _pB0, _sum0);
-                _sum1 = _mm_maddd_epi16(_pA0, _pB1, _sum1);
-                _sum2 = _mm_maddd_epi16(_pA1, _pB0, _sum2);
-                _sum3 = _mm_maddd_epi16(_pA1, _pB1, _sum3);
-#else
-                _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA0, _pB0));
-                _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_pA0, _pB1));
-                _sum2 = _mm_add_epi32(_sum2, _mm_madd_epi16(_pA1, _pB0));
-                _sum3 = _mm_add_epi32(_sum3, _mm_madd_epi16(_pA1, _pB1));
-#endif
-
-                pA += 8;
-                pB += 8;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_castps_si128(_mm_load1_ps((const float*)pA));
-                __m128i _pB = _mm_castps_si128(_mm_load1_ps((const float*)pB));
-
-#if __SSE4_1__
-                _pA = _mm_cvtepi8_epi16(_pA);
-                _pB = _mm_cvtepi8_epi16(_pB);
-#else
-                _pA = _mm_unpacklo_epi8(_pA, _mm_cmpgt_epi8(_mm_setzero_si128(), _pA));
-                _pB = _mm_unpacklo_epi8(_pB, _mm_cmpgt_epi8(_mm_setzero_si128(), _pB));
-#endif
-
-#if __XOP__
-                // 00112233
-                // 22330011
-                __m128i _pA0 = _mm_unpacklo_epi16(_pA, _pA);
-                __m128i _pA1 = _mm_shuffle_epi32(_pA0, _MM_SHUFFLE(1, 0, 3, 2));
-
-                // 00112233
-                // 1.2.3.0.
-                __m128i _pB0 = _mm_unpacklo_epi16(_pB, _pB);
-                __m128i _pB1 = _mm_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-
-                _sum0 = _mm_maccd_epi16(_pA0, _pB0, _sum0);
-                _sum1 = _mm_maccd_epi16(_pA0, _pB1, _sum1);
-                _sum2 = _mm_maccd_epi16(_pA1, _pB0, _sum2);
-                _sum3 = _mm_maccd_epi16(_pA1, _pB1, _sum3);
-#else
-                // 0123 0123
-                // 2301 2301
-                __m128i _pA0 = _pA;
-                __m128i _pA1 = _mm_shuffle_epi32(_pA, _MM_SHUFFLE(2, 3, 0, 1));
-
-                // 0123 1230
-                __m128i _pB01 = _mm_shufflehi_epi16(_pB, _MM_SHUFFLE(0, 3, 2, 1));
-
-                __m128i _sl0 = _mm_mullo_epi16(_pA0, _pB01);
-                __m128i _sh0 = _mm_mulhi_epi16(_pA0, _pB01);
-                __m128i _sl1 = _mm_mullo_epi16(_pA1, _pB01);
-                __m128i _sh1 = _mm_mulhi_epi16(_pA1, _pB01);
-                __m128i _s0 = _mm_unpacklo_epi16(_sl0, _sh0);
-                __m128i _s1 = _mm_unpackhi_epi16(_sl0, _sh0);
-                __m128i _s2 = _mm_unpacklo_epi16(_sl1, _sh1);
-                __m128i _s3 = _mm_unpackhi_epi16(_sl1, _sh1);
-
-                _sum0 = _mm_add_epi32(_sum0, _s0);
-                _sum1 = _mm_add_epi32(_sum1, _s1);
-                _sum2 = _mm_add_epi32(_sum2, _s2);
-                _sum3 = _mm_add_epi32(_sum3, _s3);
-#endif
-
-                pA += 4;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    // from
-                    //      00 11 22 33
-                    //      01 12 23 30
-                    //      20 31 02 13
-                    //      21 32 03 10
-                    // to
-                    //      00 10 20 30
-                    //      01 11 21 31
-                    //      02 12 22 32
-                    //      03 13 23 33
-                    {
-                        _sum1 = _mm_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                        __m128i _tmp0 = _mm_unpacklo_epi32(_sum0, _sum3);
-                        __m128i _tmp1 = _mm_unpackhi_epi32(_sum0, _sum3);
-                        __m128i _tmp2 = _mm_unpacklo_epi32(_sum2, _sum1);
-                        __m128i _tmp3 = _mm_unpackhi_epi32(_sum2, _sum1);
-                        _sum0 = _mm_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-
-                    _mm_store_si128((__m128i*)outptr0, _sum0);
-                    _mm_store_si128((__m128i*)(outptr0 + 4), _sum1);
-                    _mm_store_si128((__m128i*)(outptr0 + 8), _sum2);
-                    _mm_store_si128((__m128i*)(outptr0 + 12), _sum3);
-                    outptr0 += 16;
-                }
-                if (out_elempack == 1)
-                {
-                    // from
-                    //      00 11 22 33
-                    //      01 12 23 30
-                    //      20 31 02 13
-                    //      21 32 03 10
-                    // to
-                    //      00 01 02 03
-                    //      10 11 12 13
-                    //      20 21 22 23
-                    //      30 31 32 33
-                    {
-                        _sum2 = _mm_shuffle_epi32(_sum2, _MM_SHUFFLE(1, 0, 3, 2));
-                        _sum3 = _mm_shuffle_epi32(_sum3, _MM_SHUFFLE(1, 0, 3, 2));
-                        __m128i _tmp0 = _mm_unpacklo_epi32(_sum0, _sum1);
-                        __m128i _tmp1 = _mm_unpackhi_epi32(_sum0, _sum1);
-                        __m128i _tmp2 = _mm_unpacklo_epi32(_sum2, _sum3);
-                        __m128i _tmp3 = _mm_unpackhi_epi32(_sum2, _sum3);
-                        _sum0 = _mm_unpacklo_epi64(_tmp0, _tmp2);
-                        _sum1 = _mm_unpackhi_epi64(_tmp0, _tmp2);
-                        _sum2 = _mm_unpacklo_epi64(_tmp3, _tmp1);
-                        _sum3 = _mm_unpackhi_epi64(_tmp3, _tmp1);
-                        _sum1 = _mm_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                        _sum3 = _mm_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 1, 0, 3));
-                    }
-
-                    _mm_storeu_si128((__m128i*)outptr0, _sum0);
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep), _sum1);
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 2), _sum2);
-                    _mm_storeu_si128((__m128i*)(outptr0 + out_hstep * 3), _sum3);
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                _mm_store_si128((__m128i*)outptr, _sum0);
-                _mm_store_si128((__m128i*)(outptr + 4), _sum1);
-                _mm_store_si128((__m128i*)(outptr + 8), _sum2);
-                _mm_store_si128((__m128i*)(outptr + 12), _sum3);
-            }
-
-            outptr += 16;
-        }
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            const signed char* pA = pAT;
-
-            __m128i _sum0;
-            __m128i _sum1;
-
-            if (k == 0)
-            {
-                _sum0 = _mm_setzero_si128();
-                _sum1 = _mm_setzero_si128();
-            }
-            else
-            {
-                _sum0 = _mm_load_si128((const __m128i*)outptr);
-                _sum1 = _mm_load_si128((const __m128i*)(outptr + 4));
-            }
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _pA = _mm_loadl_epi64((const __m128i*)pA);
-                __m128i _pB = _mm_castps_si128(_mm_load1_ps((const float*)pB));
-
-#if __SSE4_1__
-                _pA = _mm_cvtepi8_epi16(_pA);
-                _pB = _mm_cvtepi8_epi16(_pB);
-#else
-                _pA = _mm_unpacklo_epi8(_pA, _mm_cmpgt_epi8(_mm_setzero_si128(), _pA));
-                _pB = _mm_unpacklo_epi8(_pB, _mm_cmpgt_epi8(_mm_setzero_si128(), _pB));
-#endif
-
-                // 0123
-
-                // 0101
-                // 1010
-                __m128i _pB0 = _pB;
-                __m128i _pB1 = _mm_shuffle_epi32(_pB, _MM_SHUFFLE(2, 3, 0, 1));
-
-#if __XOP__
-                _sum0 = _mm_maddd_epi16(_pA, _pB0, _sum0);
-                _sum1 = _mm_maddd_epi16(_pA, _pB1, _sum1);
-#else
-                _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA, _pB0));
-                _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_pA, _pB1));
-#endif
-
-                pA += 8;
-                pB += 4;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_castps_si128(_mm_load1_ps((const float*)pA));
-                __m128i _pB = _mm_set1_epi16(((const short*)pB)[0]);
-
-#if __SSE4_1__
-                _pA = _mm_cvtepi8_epi16(_pA);
-                _pB = _mm_cvtepi8_epi16(_pB);
-#else
-                _pA = _mm_unpacklo_epi8(_pA, _mm_cmpgt_epi8(_mm_setzero_si128(), _pA));
-                _pB = _mm_unpacklo_epi8(_pB, _mm_cmpgt_epi8(_mm_setzero_si128(), _pB));
-#endif
-
-#if __XOP__
-                // 00112233
-                _pA = _mm_unpacklo_epi16(_pA, _pA);
-
-                // 00110011
-                // 1.0.1.0.
-                __m128i _pB0 = _mm_unpacklo_epi16(_pB, _pB);
-                __m128i _pB1 = _mm_shuffle_epi32(_pB0, _MM_SHUFFLE(2, 3, 0, 1));
-
-                _sum0 = _mm_maccd_epi16(_pA, _pB0, _sum0);
-                _sum1 = _mm_maccd_epi16(_pA, _pB1, _sum1);
-#else
-                // 01230123
-                // 01011010
-                __m128i _pB01 = _mm_shufflehi_epi16(_pB, _MM_SHUFFLE(0, 1, 0, 1));
-
-                __m128i _sl = _mm_mullo_epi16(_pA, _pB01);
-                __m128i _sh = _mm_mulhi_epi16(_pA, _pB01);
-                __m128i _s0 = _mm_unpacklo_epi16(_sl, _sh);
-                __m128i _s1 = _mm_unpackhi_epi16(_sl, _sh);
-
-                _sum0 = _mm_add_epi32(_sum0, _s0);
-                _sum1 = _mm_add_epi32(_sum1, _s1);
-#endif
-
-                pA += 4;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                // from
-                //      00 11 20 31
-                //      01 10 21 30
-                // to
-                //      00 10 20 30
-                //      01 11 21 31
-                {
-                    __m128i _tmp0 = _mm_shuffle_epi32(_sum0, _MM_SHUFFLE(3, 1, 2, 0));
-                    __m128i _tmp1 = _mm_shuffle_epi32(_sum1, _MM_SHUFFLE(0, 2, 3, 1));
-                    _sum0 = _mm_unpacklo_epi32(_tmp0, _tmp1);
-                    _sum1 = _mm_unpackhi_epi32(_tmp0, _tmp1);
-                    _sum1 = _mm_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-                }
-
-                if (out_elempack == 4)
-                {
-                    _mm_store_si128((__m128i*)outptr0, _sum0);
-                    _mm_store_si128((__m128i*)(outptr0 + 4), _sum1);
-                    outptr0 += 8;
-                }
-                if (out_elempack == 1)
-                {
-#if __AVX512F__
-                    __m128i _vindex = _mm_setr_epi32(0, 1, 2, 3);
-                    _vindex = _mm_mullo_epi32(_vindex, _mm_set1_epi32(out_hstep));
-                    _mm_i32scatter_epi32(outptr0, _vindex, _sum0, sizeof(float));
-                    _mm_i32scatter_epi32(outptr0 + 1, _vindex, _sum1, sizeof(float));
-#else
-                    int sum0[4];
-                    int sum1[4];
-                    _mm_storeu_si128((__m128i*)sum0, _sum0);
-                    _mm_storeu_si128((__m128i*)sum1, _sum1);
-
-                    outptr0[0] = sum0[0];
-                    outptr0[1] = sum1[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep + 1] = sum1[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 2 + 1] = sum1[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-                    outptr0[out_hstep * 3 + 1] = sum1[3];
-#endif // __AVX512F__
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                _mm_store_si128((__m128i*)outptr, _sum0);
-                _mm_store_si128((__m128i*)(outptr + 4), _sum1);
-            }
-
-            outptr += 8;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            const signed char* pA = pAT;
-
-            __m128i _sum0;
-
-            if (k == 0)
-            {
-                _sum0 = _mm_setzero_si128();
-            }
-            else
-            {
-                _sum0 = _mm_load_si128((const __m128i*)outptr);
-            }
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _pA = _mm_loadl_epi64((const __m128i*)pA);
-                __m128i _pB = _mm_set1_epi16(((const short*)pB)[0]);
-
-#if __SSE4_1__
-                _pA = _mm_cvtepi8_epi16(_pA);
-                _pB = _mm_cvtepi8_epi16(_pB);
-#else
-                _pA = _mm_unpacklo_epi8(_pA, _mm_cmpgt_epi8(_mm_setzero_si128(), _pA));
-                _pB = _mm_unpacklo_epi8(_pB, _mm_cmpgt_epi8(_mm_setzero_si128(), _pB));
-#endif
-
-#if __XOP__
-                _sum0 = _mm_maddd_epi16(_pA, _pB, _sum0);
-#else
-                _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA, _pB));
-#endif
-
-                pA += 8;
-                pB += 2;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_loadl_epi64((const __m128i*)pA);
-                __m128i _pB = _mm_set1_epi16(pB[0]);
-
-#if __SSE4_1__
-                _pA = _mm_cvtepi8_epi16(_pA);
-#else
-                _pA = _mm_unpacklo_epi8(_pA, _mm_cmpgt_epi8(_mm_setzero_si128(), _pA));
-#endif
-
-#if __XOP__
-                _pA = _mm_unpacklo_epi16(_pA, _pA);
-
-                _sum0 = _mm_maccd_epi16(_pA, _pB, _sum0);
-#else
-                __m128i _sl = _mm_mullo_epi16(_pA, _pB);
-                __m128i _sh = _mm_mulhi_epi16(_pA, _pB);
-                __m128i _s0 = _mm_unpacklo_epi16(_sl, _sh);
-
-                _sum0 = _mm_add_epi32(_sum0, _s0);
-#endif
-
-                pA += 4;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                if (out_elempack == 4)
-                {
-                    _mm_store_si128((__m128i*)outptr0, _sum0);
-                    outptr0 += 4;
-                }
-                if (out_elempack == 1)
-                {
-#if __AVX512F__
-                    __m128i _vindex = _mm_setr_epi32(0, 1, 2, 3);
-                    _vindex = _mm_mullo_epi32(_vindex, _mm_set1_epi32(out_hstep));
-                    _mm_i32scatter_epi32(outptr0, _vindex, _sum0, sizeof(float));
-#else
-                    int sum0[4];
-                    _mm_storeu_si128((__m128i*)sum0, _sum0);
-
-                    outptr0[0] = sum0[0];
-                    outptr0[out_hstep] = sum0[1];
-                    outptr0[out_hstep * 2] = sum0[2];
-                    outptr0[out_hstep * 3] = sum0[3];
-#endif // __AVX512F__
-                    outptr0++;
-                }
-            }
-            else
-            {
-                _mm_store_si128((__m128i*)outptr, _sum0);
-            }
-
-            outptr += 4;
-        }
-
-        pAT += max_kk * 4;
-    }
-#endif // __SSE2__
-    for (; ii + 1 < max_ii; ii += 2)
-    {
-        int* outptr0 = (int*)top_blob + (i + ii) * out_hstep + j;
-
-        const signed char* pB = pBT;
-
-        int jj = 0;
-#if __SSE2__
-#if defined(__x86_64__) || defined(_M_X64)
-#if __AVX512F__
-        for (; jj + 15 < max_jj; jj += 16)
-        {
-            __m512i _sum0;
-            __m512i _sum1;
-
-            if (k == 0)
-            {
-                _sum0 = _mm512_setzero_si512();
-                _sum1 = _mm512_setzero_si512();
-            }
-            else
-            {
-                _sum0 = _mm512_load_si512((const __m512i*)outptr);
-                _sum1 = _mm512_load_si512((const __m512i*)(outptr + 16));
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m256i _pA = _mm256_castps_si256(_mm256_broadcast_ss((const float*)pA));
-                __m256i _pB = _mm256_loadu_si256((const __m256i*)pB);
-
-                __m512i _pA0 = _mm512_cvtepi8_epi16(_pA);
-                __m512i _pB0 = _mm512_cvtepi8_epi16(_pB);
-
-                // 0101 0101 0101 0101
-
-                // 0123 4567 89ab cdef
-                // 1230 5674 9ab8 defc
-                __m512i _pB1 = _mm512_shuffle_epi32(_pB0, _MM_PERM_ADCB);
-
-#if __AVX512VNNI__
-                _sum0 = _mm512_dpwssd_epi32(_sum0, _pA0, _pB0);
-                _sum1 = _mm512_dpwssd_epi32(_sum1, _pA0, _pB1);
-#else
-                _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA0, _pB0));
-                _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_pA0, _pB1));
-#endif
-
-                pA += 4;
-                pB += 32;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_set1_epi16(((const short*)pA)[0]);
-                __m128i _pB = _mm_load_si128((const __m128i*)pB);
-
-                __m256i _pA0 = _mm256_cvtepi8_epi16(_pA);
-                __m256i _pB0 = _mm256_cvtepi8_epi16(_pB);
-
-                // 01010101 01010101
-
-                // 01234567 89abcdef
-                // 12305674 9ab8defc
-                __m256i _pB1 = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(_pB0, _MM_SHUFFLE(0, 3, 2, 1)), _MM_SHUFFLE(0, 3, 2, 1));
-
-                __m512i _s0 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB0));
-                __m512i _s1 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA0, _pB1));
-
-                _sum0 = _mm512_add_epi32(_sum0, _s0);
-                _sum1 = _mm512_add_epi32(_sum1, _s1);
-
-                pA += 2;
-                pB += 16;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    // 00 11 02 13  04 15 06 17  08 19 0a 1b  0c 1d 0e 1f
-                    // 01 12 03 10  05 16 07 14  09 1a 0b 18  0d 1e 0f 1c
-
-                    __m512i _tmp0 = _mm512_unpacklo_epi32(_sum0, _sum1);
-                    __m512i _tmp1 = _mm512_unpackhi_epi32(_sum0, _sum1);
-
-                    _sum0 = _mm512_unpacklo_epi64(_tmp0, _tmp1);
-                    _sum1 = _mm512_unpackhi_epi64(_tmp0, _tmp1);
-
-                    _sum0 = _sum0;
-                    _sum1 = _mm512_shuffle_epi32(_sum1, _MM_PERM_CBAD);
-
-                    // 0123 4567 89ab cdef  x 0
-                    // 0123 4567 89ab cdef  x 1
-
-                    _mm512_storeu_si512((__m512i*)outptr0, _sum0);
-                    _mm512_storeu_si512((__m512i*)(outptr0 + out_hstep), _sum1);
-                    outptr0 += 16;
-                }
-            }
-            else
-            {
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                _mm512_store_si512((__m512i*)(outptr + 16), _sum1);
-            }
-
-            outptr += 32;
-        }
-#endif // __AVX512F__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-#if __AVX2__
-            __m256i _sum0;
-            __m256i _sum1;
-#else
-            __m128i _sum0;
-            __m128i _sum1;
-            __m128i _sum2;
-            __m128i _sum3;
-#endif
-
-            if (k == 0)
-            {
-#if __AVX2__
-                _sum0 = _mm256_setzero_si256();
-                _sum1 = _mm256_setzero_si256();
-#else
-                _sum0 = _mm_setzero_si128();
-                _sum1 = _mm_setzero_si128();
-                _sum2 = _mm_setzero_si128();
-                _sum3 = _mm_setzero_si128();
-#endif
-            }
-            else
-            {
-#if __AVX2__
-                _sum0 = _mm256_load_si256((const __m256i*)outptr);
-                _sum1 = _mm256_load_si256((const __m256i*)(outptr + 8));
-#else
-                _sum0 = _mm_load_si128((const __m128i*)outptr);
-                _sum1 = _mm_load_si128((const __m128i*)(outptr + 4));
-                _sum2 = _mm_load_si128((const __m128i*)(outptr + 8));
-                _sum3 = _mm_load_si128((const __m128i*)(outptr + 12));
-#endif
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _pA = _mm_castps_si128(_mm_load1_ps((const float*)pA));
-                __m128i _pB = _mm_loadu_si128((const __m128i*)pB);
-
-#if __AVX2__
-                __m256i _pA0 = _mm256_cvtepi8_epi16(_pA);
-                __m256i _pB0 = _mm256_cvtepi8_epi16(_pB);
-
-                // 0101 0101
-
-                // 0123 4567
-                // 1230 5674
-                __m256i _pB1 = _mm256_shuffle_epi32(_pB0, _MM_SHUFFLE(0, 3, 2, 1));
-
-#if __AVX512VNNI__ || __AVXVNNI__
-                _sum0 = _mm256_dpwssd_epi32(_sum0, _pA0, _pB0);
-                _sum1 = _mm256_dpwssd_epi32(_sum1, _pA0, _pB1);
-#else
-                _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_pA0, _pB0));
-                _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_pA0, _pB1));
-#endif // __AVX512VNNI__ || __AVXVNNI__
-#else  // __AVX2__
-#if __SSE4_1__
-                _pA = _mm_cvtepi8_epi16(_pA);
-#else
-                _pA = _mm_unpacklo_epi8(_pA, _mm_cmpgt_epi8(_mm_setzero_si128(), _pA));
-#endif
-
-                __m128i _extpB = _mm_cmpgt_epi8(_mm_setzero_si128(), _pB);
-                __m128i _pB0 = _mm_unpacklo_epi8(_pB, _extpB);
-                __m128i _pB1 = _mm_unpackhi_epi8(_pB, _extpB);
-
-                // 0101
-                // 1010
-                __m128i _pA0 = _pA;
-                __m128i _pA1 = _mm_shuffle_epi32(_pA, _MM_SHUFFLE(2, 3, 0, 1));
-
-                // 0123
-                // 4567
-
-#if __XOP__
-                _sum0 = _mm_maddd_epi16(_pA0, _pB0, _sum0);
-                _sum1 = _mm_maddd_epi16(_pA0, _pB1, _sum1);
-                _sum2 = _mm_maddd_epi16(_pA1, _pB0, _sum2);
-                _sum3 = _mm_maddd_epi16(_pA1, _pB1, _sum3);
-#else
-                _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA0, _pB0));
-                _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_pA0, _pB1));
-                _sum2 = _mm_add_epi32(_sum2, _mm_madd_epi16(_pA1, _pB0));
-                _sum3 = _mm_add_epi32(_sum3, _mm_madd_epi16(_pA1, _pB1));
-#endif
-#endif // __AVX2__
-
-                pA += 4;
-                pB += 16;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_set1_epi16(((const short*)pA)[0]);
-                __m128i _pB = _mm_loadl_epi64((const __m128i*)pB);
-
-#if __AVX2__
-                _pA = _mm_cvtepi8_epi16(_pA);
-                _pB = _mm_cvtepi8_epi16(_pB);
-
-                // 01010101
-
-                // 01234567
-                // 12305674
-                __m128i _pB0 = _pB;
-                __m128i _pB1 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_pB, _MM_SHUFFLE(0, 3, 2, 1)), _MM_SHUFFLE(0, 3, 2, 1));
-
-                __m256i _s0 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA, _pB0));
-                __m256i _s1 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA, _pB1));
-
-                _sum0 = _mm256_add_epi32(_sum0, _s0);
-                _sum1 = _mm256_add_epi32(_sum1, _s1);
-#else // __AVX2__
-#if __SSE4_1__
-                _pA = _mm_cvtepi8_epi16(_pA);
-                _pB = _mm_cvtepi8_epi16(_pB);
-#else
-                _pA = _mm_unpacklo_epi8(_pA, _mm_cmpgt_epi8(_mm_setzero_si128(), _pA));
-                _pB = _mm_unpacklo_epi8(_pB, _mm_cmpgt_epi8(_mm_setzero_si128(), _pB));
-#endif
-
-                // 01010101
-                // 10101010
-                __m128i _pA0 = _pA;
-                __m128i _pA1 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_pA, _MM_SHUFFLE(2, 3, 0, 1)), _MM_SHUFFLE(2, 3, 0, 1));
-
-                // 01234567
-
-                __m128i _sl0 = _mm_mullo_epi16(_pA0, _pB);
-                __m128i _sh0 = _mm_mulhi_epi16(_pA0, _pB);
-                __m128i _sl1 = _mm_mullo_epi16(_pA1, _pB);
-                __m128i _sh1 = _mm_mulhi_epi16(_pA1, _pB);
-                __m128i _s0 = _mm_unpacklo_epi16(_sl0, _sh0);
-                __m128i _s1 = _mm_unpackhi_epi16(_sl0, _sh0);
-                __m128i _s2 = _mm_unpacklo_epi16(_sl1, _sh1);
-                __m128i _s3 = _mm_unpackhi_epi16(_sl1, _sh1);
-
-                _sum0 = _mm_add_epi32(_sum0, _s0);
-                _sum1 = _mm_add_epi32(_sum1, _s1);
-                _sum2 = _mm_add_epi32(_sum2, _s2);
-                _sum3 = _mm_add_epi32(_sum3, _s3);
-#endif // __AVX2__
-
-                pA += 2;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-#if __AVX2__
-                    // 00 11 02 13  04 15 06 17
-                    // 01 12 03 10  05 16 07 14
-                    __m256i _tmp0 = _mm256_unpacklo_epi32(_sum0, _sum1);
-                    __m256i _tmp1 = _mm256_unpackhi_epi32(_sum0, _sum1);
-
-                    // 00 01 11 12  04 05 15 16
-                    // 02 03 13 10  06 07 17 14
-                    _sum0 = _mm256_unpacklo_epi64(_tmp0, _tmp1);
-                    _sum1 = _mm256_unpackhi_epi64(_tmp0, _tmp1);
-
-                    // 00 01 02 03  04 05 06 07
-                    // 11 12 13 10  15 16 17 14
-                    _sum1 = _mm256_shuffle_epi32(_sum1, _MM_SHUFFLE(2, 1, 0, 3));
-
-                    _mm256_storeu_si256((__m256i*)outptr0, _sum0);
-                    _mm256_storeu_si256((__m256i*)(outptr0 + out_hstep), _sum1);
-#else
-                    // 00 11 02 13
-                    // 04 15 06 17
-                    // 10 01 12 03
-                    // 14 05 16 07
-                    _sum2 = _mm_shuffle_epi32(_sum2, _MM_SHUFFLE(2, 3, 0, 1));
-                    _sum3 = _mm_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 3, 0, 1));
-
-                    // 00 11 02 13
-                    // 04 15 06 17
-                    // 01 10 03 12
-                    // 05 14 07 16
-
-                    __m128i _tmp0 = _mm_unpacklo_epi32(_sum0, _sum2);
-                    __m128i _tmp1 = _mm_unpackhi_epi32(_sum0, _sum2);
-                    __m128i _tmp2 = _mm_unpacklo_epi32(_sum1, _sum3);
-                    __m128i _tmp3 = _mm_unpackhi_epi32(_sum1, _sum3);
-
-                    // 00 01 11 10
-                    // 02 03 13 12
-                    // 04 05 15 14
-                    // 06 07 17 16
-
-                    _sum0 = _mm_unpacklo_epi64(_tmp0, _tmp1);
-                    _sum1 = _mm_unpacklo_epi64(_tmp2, _tmp3);
-                    _sum2 = _mm_unpackhi_epi64(_tmp0, _tmp1);
-                    _sum3 = _mm_unpackhi_epi64(_tmp2, _tmp3);
-
-                    // 00 01 02 03
-                    // 04 05 06 07
-                    // 11 10 13 12
-                    // 15 14 17 16
-                    _sum2 = _mm_shuffle_epi32(_sum2, _MM_SHUFFLE(2, 3, 0, 1));
-                    _sum3 = _mm_shuffle_epi32(_sum3, _MM_SHUFFLE(2, 3, 0, 1));
-
-                    _mm_store_si128((__m128i*)outptr0, _sum0);
-                    _mm_store_si128((__m128i*)(outptr0 + 4), _sum1);
-                    _mm_store_si128((__m128i*)(outptr0 + out_hstep), _sum2);
-                    _mm_store_si128((__m128i*)(outptr0 + out_hstep + 4), _sum3);
-#endif // __AVX2__
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-#if __AVX2__
-                _mm256_store_si256((__m256i*)outptr, _sum0);
-                _mm256_store_si256((__m256i*)(outptr + 8), _sum1);
-#else
-                _mm_store_si128((__m128i*)outptr, _sum0);
-                _mm_store_si128((__m128i*)(outptr + 4), _sum1);
-                _mm_store_si128((__m128i*)(outptr + 8), _sum2);
-                _mm_store_si128((__m128i*)(outptr + 12), _sum3);
-#endif
-            }
-
-            outptr += 16;
-        }
-#endif // defined(__x86_64__) || defined(_M_X64)
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            __m128i _sum0;
-            __m128i _sum1;
-
-            if (k == 0)
-            {
-                _sum0 = _mm_setzero_si128();
-                _sum1 = _mm_setzero_si128();
-            }
-            else
-            {
-                _sum0 = _mm_load_si128((const __m128i*)outptr);
-                _sum1 = _mm_load_si128((const __m128i*)(outptr + 4));
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _pA = _mm_castps_si128(_mm_load1_ps((const float*)pA));
-                __m128i _pB = _mm_loadl_epi64((const __m128i*)pB);
-
-#if __SSE4_1__
-                _pA = _mm_cvtepi8_epi16(_pA);
-                _pB = _mm_cvtepi8_epi16(_pB);
-#else
-                _pA = _mm_unpacklo_epi8(_pA, _mm_cmpgt_epi8(_mm_setzero_si128(), _pA));
-                _pB = _mm_unpacklo_epi8(_pB, _mm_cmpgt_epi8(_mm_setzero_si128(), _pB));
-#endif
-
-                // 0101
-
-                // 0123
-                // 1230
-                __m128i _pB0 = _pB;
-                __m128i _pB1 = _mm_shuffle_epi32(_pB, _MM_SHUFFLE(0, 3, 2, 1));
-
-#if __XOP__
-                _sum0 = _mm_maddd_epi16(_pA, _pB0, _sum0);
-                _sum1 = _mm_maddd_epi16(_pA, _pB1, _sum1);
-#else
-                _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA, _pB0));
-                _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_pA, _pB1));
-#endif
-
-                pA += 4;
-                pB += 8;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_set1_epi16(((const short*)pA)[0]);
-                __m128i _pB = _mm_castps_si128(_mm_load1_ps((const float*)pB));
-
-#if __SSE4_1__
-                _pA = _mm_cvtepi8_epi16(_pA);
-                _pB = _mm_cvtepi8_epi16(_pB);
-#else
-                _pA = _mm_unpacklo_epi8(_pA, _mm_cmpgt_epi8(_mm_setzero_si128(), _pA));
-                _pB = _mm_unpacklo_epi8(_pB, _mm_cmpgt_epi8(_mm_setzero_si128(), _pB));
-#endif
-
-                // 01010101
-
-                // 01231230
-                __m128i _pB01 = _mm_shufflehi_epi16(_pB, _MM_SHUFFLE(0, 3, 2, 1));
-
-                __m128i _sl = _mm_mullo_epi16(_pA, _pB01);
-                __m128i _sh = _mm_mulhi_epi16(_pA, _pB01);
-                __m128i _s0 = _mm_unpacklo_epi16(_sl, _sh);
-                __m128i _s1 = _mm_unpackhi_epi16(_sl, _sh);
-
-                _sum0 = _mm_add_epi32(_sum0, _s0);
-                _sum1 = _mm_add_epi32(_sum1, _s1);
-
-                pA += 2;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    // 00 11 02 13
-                    // 01 12 03 10
-                    __m128i _tmp0 = _mm_unpacklo_epi32(_sum0, _sum1);
-                    __m128i _tmp1 = _mm_unpackhi_epi32(_sum0, _sum1);
-
-                    // 00 01 11 12
-                    // 02 03 13 10
-                    _sum0 = _mm_unpacklo_epi64(_tmp0, _tmp1);
-                    _sum1 = _mm_unpackhi_epi64(_tmp1, _tmp0);
-
-                    // 00 01 02 03
-                    // 13 10 11 12
-                    _sum1 = _mm_shuffle_epi32(_sum1, _MM_SHUFFLE(0, 3, 2, 1));
-
-                    _mm_store_si128((__m128i*)outptr0, _sum0);
-                    _mm_store_si128((__m128i*)(outptr0 + out_hstep), _sum1);
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                _mm_store_si128((__m128i*)outptr, _sum0);
-                _mm_store_si128((__m128i*)(outptr + 4), _sum1);
-            }
-
-            outptr += 8;
-        }
-#endif // __SSE2__
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            int sum00;
-            int sum10;
-            int sum01;
-            int sum11;
-
-            if (k == 0)
-            {
-                sum00 = 0;
-                sum10 = 0;
-                sum01 = 0;
-                sum11 = 0;
-            }
-            else
-            {
-                sum00 = outptr[0];
-                sum10 = outptr[1];
-                sum01 = outptr[2];
-                sum11 = outptr[3];
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                sum00 += pA[0] * pB[0];
-                sum00 += pA[1] * pB[1];
-                sum10 += pA[2] * pB[0];
-                sum10 += pA[3] * pB[1];
-                sum01 += pA[0] * pB[2];
-                sum01 += pA[1] * pB[3];
-                sum11 += pA[2] * pB[2];
-                sum11 += pA[3] * pB[3];
-                pA += 4;
-                pB += 4;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                sum00 += pA[0] * pB[0];
-                sum10 += pA[1] * pB[0];
-                sum01 += pA[0] * pB[1];
-                sum11 += pA[1] * pB[1];
-                pA += 2;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum00;
-                    outptr0[1] = sum01;
-                    outptr0[out_hstep] = sum10;
-                    outptr0[out_hstep + 1] = sum11;
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                outptr[0] = sum00;
-                outptr[1] = sum10;
-                outptr[2] = sum01;
-                outptr[3] = sum11;
-            }
-
-            outptr += 4;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            int sum0;
-            int sum1;
-
-            if (k == 0)
-            {
-                sum0 = 0;
-                sum1 = 0;
-            }
-            else
-            {
-                sum0 = outptr[0];
-                sum1 = outptr[1];
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                sum0 += pA[0] * pB[0];
-                sum0 += pA[1] * pB[1];
-                sum1 += pA[2] * pB[0];
-                sum1 += pA[3] * pB[1];
-                pA += 4;
-                pB += 2;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                sum0 += pA[0] * pB[0];
-                sum1 += pA[1] * pB[0];
-                pA += 2;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum0;
-                    outptr0[out_hstep] = sum1;
-                    outptr0++;
-                }
-            }
-            else
-            {
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-            }
-
-            outptr += 2;
-        }
-
-        pAT += max_kk * 2;
-    }
-    for (; ii < max_ii; ii += 1)
-    {
-        int* outptr0 = (int*)top_blob + (i + ii) * out_hstep + j;
-
-        const signed char* pB = pBT;
-
-        int jj = 0;
-#if __SSE2__
-#if defined(__x86_64__) || defined(_M_X64)
-#if __AVX512F__
-        for (; jj + 15 < max_jj; jj += 16)
-        {
-            __m512i _sum0;
-
-            if (k == 0)
-            {
-                _sum0 = _mm512_setzero_si512();
-            }
-            else
-            {
-                _sum0 = _mm512_load_si512((const __m512i*)outptr);
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m256i _pA = _mm256_set1_epi16(((const short*)pA)[0]);
-                __m256i _pB = _mm256_loadu_si256((const __m256i*)pB);
-
-                __m512i _pA0 = _mm512_cvtepi8_epi16(_pA);
-                __m512i _pB0 = _mm512_cvtepi8_epi16(_pB);
-
-#if __AVX512VNNI__
-                _sum0 = _mm512_dpwssd_epi32(_sum0, _pA0, _pB0);
-#else
-                _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_pA0, _pB0));
-#endif
-
-                pA += 2;
-                pB += 32;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m256i _pA = _mm256_set1_epi16(pA[0]);
-                __m128i _pB = _mm_load_si128((const __m128i*)pB);
-
-                __m256i _pB0 = _mm256_cvtepi8_epi16(_pB);
-
-                __m512i _s0 = _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_pA, _pB0));
-
-                _sum0 = _mm512_add_epi32(_sum0, _s0);
-
-                pA += 1;
-                pB += 16;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    _mm512_storeu_si512((__m512i*)outptr0, _sum0);
-                    outptr0 += 16;
-                }
-            }
-            else
-            {
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-            }
-
-            outptr += 16;
-        }
-#endif // __AVX512F__
-        for (; jj + 7 < max_jj; jj += 8)
-        {
-#if __AVX2__
-            __m256i _sum0;
-#else
-            __m128i _sum0;
-            __m128i _sum1;
-#endif
-
-            if (k == 0)
-            {
-#if __AVX2__
-                _sum0 = _mm256_setzero_si256();
-#else
-                _sum0 = _mm_setzero_si128();
-                _sum1 = _mm_setzero_si128();
-#endif
-            }
-            else
-            {
-#if __AVX2__
-                _sum0 = _mm256_load_si256((const __m256i*)outptr);
-#else
-                _sum0 = _mm_load_si128((const __m128i*)outptr);
-                _sum1 = _mm_load_si128((const __m128i*)(outptr + 4));
-#endif
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _pA = _mm_set1_epi16(((const short*)pA)[0]);
-                __m128i _pB = _mm_loadu_si128((const __m128i*)pB);
-
-#if __AVX2__
-                __m256i _pA0 = _mm256_cvtepi8_epi16(_pA);
-                __m256i _pB0 = _mm256_cvtepi8_epi16(_pB);
-
-#if __AVX512VNNI__ || __AVXVNNI__
-                _sum0 = _mm256_dpwssd_epi32(_sum0, _pA0, _pB0);
-#else
-                _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_pA0, _pB0));
-#endif // __AVX512VNNI__ || __AVXVNNI__
-#else  // __AVX2__
-#if __SSE4_1__
-                _pA = _mm_cvtepi8_epi16(_pA);
-#else
-                _pA = _mm_unpacklo_epi8(_pA, _mm_cmpgt_epi8(_mm_setzero_si128(), _pA));
-#endif
-
-                __m128i _extpB = _mm_cmpgt_epi8(_mm_setzero_si128(), _pB);
-                __m128i _pB0 = _mm_unpacklo_epi8(_pB, _extpB);
-                __m128i _pB1 = _mm_unpackhi_epi8(_pB, _extpB);
-
-#if __XOP__
-                _sum0 = _mm_maddd_epi16(_pA, _pB0, _sum0);
-                _sum1 = _mm_maddd_epi16(_pA, _pB1, _sum1);
-#else
-                _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA, _pB0));
-                _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_pA, _pB1));
-#endif
-#endif // __AVX2__
-
-                pA += 2;
-                pB += 16;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_set1_epi16(pA[0]);
-                __m128i _pB = _mm_loadl_epi64((const __m128i*)pB);
-
-#if __SSE4_1__
-                _pB = _mm_cvtepi8_epi16(_pB);
-#else
-                _pB = _mm_unpacklo_epi8(_pB, _mm_cmpgt_epi8(_mm_setzero_si128(), _pB));
-#endif
-
-#if __AVX2__
-                __m256i _s0 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_pA, _pB));
-
-                _sum0 = _mm256_add_epi32(_sum0, _s0);
-#else  // __AVX2__
-                __m128i _sl = _mm_mullo_epi16(_pA, _pB);
-                __m128i _sh = _mm_mulhi_epi16(_pA, _pB);
-                __m128i _s0 = _mm_unpacklo_epi16(_sl, _sh);
-                __m128i _s1 = _mm_unpackhi_epi16(_sl, _sh);
-
-                _sum0 = _mm_add_epi32(_sum0, _s0);
-                _sum1 = _mm_add_epi32(_sum1, _s1);
-#endif // __AVX2__
-
-                pA += 1;
-                pB += 8;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-#if __AVX2__
-                    _mm256_storeu_si256((__m256i*)outptr0, _sum0);
-#else
-                    _mm_store_si128((__m128i*)outptr0, _sum0);
-                    _mm_store_si128((__m128i*)(outptr0 + 4), _sum1);
-#endif
-                    outptr0 += 8;
-                }
-            }
-            else
-            {
-#if __AVX2__
-                _mm256_store_si256((__m256i*)outptr, _sum0);
-#else
-                _mm_store_si128((__m128i*)outptr, _sum0);
-                _mm_store_si128((__m128i*)(outptr + 4), _sum1);
-#endif
-            }
-
-            outptr += 8;
-        }
-#endif // defined(__x86_64__) || defined(_M_X64)
-        for (; jj + 3 < max_jj; jj += 4)
-        {
-            __m128i _sum0;
-
-            if (k == 0)
-            {
-                _sum0 = _mm_setzero_si128();
-            }
-            else
-            {
-                _sum0 = _mm_load_si128((const __m128i*)outptr);
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _pA = _mm_castps_si128(_mm_load1_ps((const float*)pA));
-                __m128i _pB = _mm_loadl_epi64((const __m128i*)pB);
-
-#if __SSE4_1__
-                _pA = _mm_cvtepi8_epi16(_pA);
-                _pB = _mm_cvtepi8_epi16(_pB);
-#else
-                _pA = _mm_unpacklo_epi8(_pA, _mm_cmpgt_epi8(_mm_setzero_si128(), _pA));
-                _pB = _mm_unpacklo_epi8(_pB, _mm_cmpgt_epi8(_mm_setzero_si128(), _pB));
-#endif
-
-                // 0xxx -> 0000
-                __m128i _pA0 = _mm_shuffle_epi32(_pA, _MM_SHUFFLE(0, 0, 0, 0));
-
-#if __XOP__
-                _sum0 = _mm_maddd_epi16(_pA0, _pB, _sum0);
-#else
-                _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_pA0, _pB));
-#endif
-
-                pA += 2;
-                pB += 8;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                __m128i _pA = _mm_set1_epi16(pA[0]);
-                __m128i _pB = _mm_loadl_epi64((const __m128i*)pB);
-
-#if __SSE4_1__
-                _pB = _mm_cvtepi8_epi16(_pB);
-#else
-                _pB = _mm_unpacklo_epi8(_pB, _mm_cmpgt_epi8(_mm_setzero_si128(), _pB));
-#endif
-
-                __m128i _sl = _mm_mullo_epi16(_pA, _pB);
-                __m128i _sh = _mm_mulhi_epi16(_pA, _pB);
-                __m128i _s0 = _mm_unpacklo_epi16(_sl, _sh);
-
-                _sum0 = _mm_add_epi32(_sum0, _s0);
-
-                pA += 1;
-                pB += 4;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    _mm_store_si128((__m128i*)outptr0, _sum0);
-                    outptr0 += 4;
-                }
-            }
-            else
-            {
-                _mm_store_si128((__m128i*)outptr, _sum0);
-            }
-
-            outptr += 4;
-        }
-#endif // __SSE2__
-        for (; jj + 1 < max_jj; jj += 2)
-        {
-            int sum0;
-            int sum1;
-
-            if (k == 0)
-            {
-                sum0 = 0;
-                sum1 = 0;
-            }
-            else
-            {
-                sum0 = outptr[0];
-                sum1 = outptr[1];
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                sum0 += pA[0] * pB[0];
-                sum0 += pA[1] * pB[1];
-                sum1 += pA[0] * pB[2];
-                sum1 += pA[1] * pB[3];
-                pA += 2;
-                pB += 4;
-            }
-            for (; kk < max_kk; kk += 1)
-            {
-                sum0 += pA[0] * pB[0];
-                sum1 += pA[0] * pB[1];
-                pA += 1;
-                pB += 2;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum0;
-                    outptr0[1] = sum1;
-                    outptr0 += 2;
-                }
-            }
-            else
-            {
-                outptr[0] = sum0;
-                outptr[1] = sum1;
-            }
-
-            outptr += 2;
-        }
-        for (; jj < max_jj; jj += 1)
-        {
-            int sum;
-
-            if (k == 0)
-            {
-                sum = 0;
-            }
-            else
-            {
-                sum = outptr[0];
-            }
-
-            const signed char* pA = pAT;
-            int kk = 0;
-            for (; kk < max_kk; kk += 1)
-            {
-                sum += pA[0] * pB[0];
-                pA += 1;
-                pB += 1;
-            }
-
-            if (k_end)
-            {
-                // if (out_elempack == 1)
-                {
-                    outptr0[0] = sum;
-                    outptr0++;
-                }
-            }
-            else
-            {
-                outptr[0] = sum;
-            }
-
-            outptr += 1;
-        }
-
-        pAT += max_kk;
-    }
-}
-
-static void convolution_im2col_gemm_get_optimal_tile_mnk_int8(int M, int N, int K, int& TILE_M, int& TILE_N, int& TILE_K, int nT)
-{
-    // resolve optimal tile size from cache size
-    const size_t l2_cache_size_int8 = (int)(get_cpu_level2_cache_size() / sizeof(signed char));
-
-    if (nT == 0)
-        nT = get_physical_big_cpu_count();
-
-    // solve K
-    {
-        // try not to split K
-#if __AVX512F__
-        int tile_size = (l2_cache_size_int8 - 64) / 16;
-#elif __AVX2__
-        int tile_size = (l2_cache_size_int8 - 32) / 8;
-#elif __SSE2__
-        int tile_size = (l2_cache_size_int8 - 16) / 8;
-#else
-        int tile_size = (l2_cache_size_int8 - 2) / 3;
-#endif
-
-#if __AVX512F__
-        TILE_K = std::max(16, tile_size / 16 * 16);
-#elif __AVX2__
-        TILE_K = std::max(8, tile_size / 8 * 8);
-#elif __SSE2__
-        TILE_K = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_K = std::max(2, tile_size / 2 * 2);
-#endif
-
-        int nn_K = (K + TILE_K - 1) / TILE_K;
-#if __AVX512F__
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 15) / 16 * 16);
-#elif __AVX2__
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 7) / 8 * 8);
-#elif __SSE2__
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 3) / 4 * 4);
-#else
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 1) / 2 * 2);
-#endif
-    }
-
-    // solve M
-    {
-#if __AVX512F__
-        int nn_M = (M + 63) / 64;
-#elif __AVX2__
-        int nn_M = (M + 31) / 32;
-#elif __SSE2__
-        int nn_M = (M + 15) / 16;
-#else
-        int nn_M = (M + 7) / 8;
-#endif
-
-#if __AVX512F__
-        TILE_M = std::max(16, ((M + nn_M - 1) / nn_M + 15) / 16 * 16);
-#elif __AVX2__
-        TILE_M = std::max(8, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-#elif __SSE2__
-        TILE_M = std::max(4, ((M + nn_M - 1) / nn_M + 3) / 4 * 4);
-#else
-        TILE_M = std::max(2, ((M + nn_M - 1) / nn_M + 1) / 2 * 2);
-#endif
-    }
-
-    {
-        TILE_M *= std::min(nT, get_physical_cpu_count());
-
-        int nn_M = (M + TILE_M - 1) / TILE_M;
-#if __AVX512F__
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 15) / 16 * 16);
-#elif __AVX2__
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 7) / 8 * 8);
-#elif __SSE2__
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 3) / 4 * 4);
-#else
-        TILE_M = std::min(TILE_M, ((M + nn_M - 1) / nn_M + 1) / 2 * 2);
-#endif
-
-        if (nT > 1)
-        {
-#if __AVX512F__
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 15) / 16 * 16);
-#elif __AVX2__
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 7) / 8 * 8);
-#elif __SSE2__
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 3) / 4 * 4);
-#else
-            TILE_M = std::min(TILE_M, (std::max(1, TILE_M / nT) + 1) / 2 * 2);
-#endif
-        }
-    }
-
-    if (N > 0)
-    {
-        int tile_size;
-        if (TILE_K >= K)
-        {
-            tile_size = (l2_cache_size_int8 - TILE_M * TILE_K) / TILE_K;
-        }
-        else
-        {
-            tile_size = (l2_cache_size_int8 - TILE_M * TILE_K) / (TILE_M * 4 + TILE_K);
-        }
-
-#if __AVX512F__
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#elif __AVX2__
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#elif __SSE2__
-        TILE_N = std::max(4, tile_size / 4 * 4);
-#else
-        TILE_N = std::max(1, tile_size);
-#endif
-
-        int nn_N = (N + TILE_N - 1) / TILE_N;
-#if __AVX512F__
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#elif __AVX2__
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#elif __SSE2__
-        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
-#else
-        TILE_N = std::min(TILE_N, (N + nn_N - 1) / nn_N);
-#endif
-    }
-}
-
-static void convolution_im2col_input_tile_conv1x1s1d1_int8(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk)
-{
-    const int elempack = bottom_blob.elempack;
-
-    signed char* pp = B;
-
-    int jj = 0;
-#if __SSE2__
-#if defined(__x86_64__) || defined(_M_X64)
-#if __AVX512F__
-    for (; jj + 15 < max_jj; jj += 16)
-    {
-        if (elempack == 8)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k / 8) + (j + jj) * 8;
-
-            int kk = 0;
-            for (; kk < max_kk / 8; kk++)
-            {
-                __m128i _r01 = _mm_load_si128((const __m128i*)p0);
-                __m128i _r23 = _mm_load_si128((const __m128i*)(p0 + 16));
-                __m128i _r45 = _mm_load_si128((const __m128i*)(p0 + 32));
-                __m128i _r67 = _mm_load_si128((const __m128i*)(p0 + 48));
-                __m128i _r89 = _mm_load_si128((const __m128i*)(p0 + 64));
-                __m128i _rab = _mm_load_si128((const __m128i*)(p0 + 80));
-                __m128i _rcd = _mm_load_si128((const __m128i*)(p0 + 96));
-                __m128i _ref = _mm_load_si128((const __m128i*)(p0 + 112));
-
-                __m128i _t0 = _mm_unpacklo_epi16(_r01, _r23);
-                __m128i _t1 = _mm_unpackhi_epi16(_r01, _r23);
-                __m128i _t2 = _mm_unpacklo_epi16(_r45, _r67);
-                __m128i _t3 = _mm_unpackhi_epi16(_r45, _r67);
-                __m128i _t4 = _mm_unpacklo_epi16(_r89, _rab);
-                __m128i _t5 = _mm_unpackhi_epi16(_r89, _rab);
-                __m128i _t6 = _mm_unpacklo_epi16(_rcd, _ref);
-                __m128i _t7 = _mm_unpackhi_epi16(_rcd, _ref);
-
-                _r01 = _mm_unpacklo_epi16(_t0, _t1);
-                _r23 = _mm_unpackhi_epi16(_t0, _t1);
-                _r45 = _mm_unpacklo_epi16(_t2, _t3);
-                _r67 = _mm_unpackhi_epi16(_t2, _t3);
-                _r89 = _mm_unpacklo_epi16(_t4, _t5);
-                _rab = _mm_unpackhi_epi16(_t4, _t5);
-                _rcd = _mm_unpacklo_epi16(_t6, _t7);
-                _ref = _mm_unpackhi_epi16(_t6, _t7);
-
-                __m128i _r0 = _mm_unpacklo_epi64(_r01, _r45);
-                __m128i _r1 = _mm_unpacklo_epi64(_r89, _rcd);
-                __m128i _r2 = _mm_unpackhi_epi64(_r01, _r45);
-                __m128i _r3 = _mm_unpackhi_epi64(_r89, _rcd);
-                __m128i _r4 = _mm_unpacklo_epi64(_r23, _r67);
-                __m128i _r5 = _mm_unpacklo_epi64(_rab, _ref);
-                __m128i _r6 = _mm_unpackhi_epi64(_r23, _r67);
-                __m128i _r7 = _mm_unpackhi_epi64(_rab, _ref);
-
-                _mm_store_si128((__m128i*)pp, _r0);
-                _mm_store_si128((__m128i*)(pp + 16), _r1);
-                _mm_store_si128((__m128i*)(pp + 32), _r2);
-                _mm_store_si128((__m128i*)(pp + 48), _r3);
-                _mm_store_si128((__m128i*)(pp + 64), _r4);
-                _mm_store_si128((__m128i*)(pp + 80), _r5);
-                _mm_store_si128((__m128i*)(pp + 96), _r6);
-                _mm_store_si128((__m128i*)(pp + 112), _r7);
-
-                pp += 128;
-                p0 += bottom_blob.cstep * 8;
-            }
-        }
-
-        if (elempack == 1)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k) + (j + jj);
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _r0 = _mm_loadu_si128((const __m128i*)p0);
-                __m128i _r1 = _mm_loadu_si128((const __m128i*)(p0 + bottom_blob.cstep));
-                __m128i _r01 = _mm_unpacklo_epi8(_r0, _r1);
-                __m128i _r23 = _mm_unpackhi_epi8(_r0, _r1);
-                _mm_storeu_si128((__m128i*)pp, _r01);
-                _mm_storeu_si128((__m128i*)(pp + 16), _r23);
-                pp += 32;
-                p0 += bottom_blob.cstep * 2;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                _mm_storeu_si128((__m128i*)pp, _mm_loadu_si128((const __m128i*)p0));
-                pp += 16;
-                p0 += bottom_blob.cstep;
-            }
-        }
-    }
-#endif // __AVX512F__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-        if (elempack == 8)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k / 8) + (j + jj) * 8;
-
-            int kk = 0;
-            for (; kk < max_kk / 8; kk++)
-            {
-                __m128i _r01 = _mm_load_si128((const __m128i*)p0);
-                __m128i _r23 = _mm_load_si128((const __m128i*)(p0 + 16));
-                __m128i _r45 = _mm_load_si128((const __m128i*)(p0 + 32));
-                __m128i _r67 = _mm_load_si128((const __m128i*)(p0 + 48));
-
-                __m128i _t0 = _mm_unpacklo_epi16(_r01, _r23);
-                __m128i _t1 = _mm_unpackhi_epi16(_r01, _r23);
-                __m128i _t2 = _mm_unpacklo_epi16(_r45, _r67);
-                __m128i _t3 = _mm_unpackhi_epi16(_r45, _r67);
-
-                _r01 = _mm_unpacklo_epi16(_t0, _t1);
-                _r23 = _mm_unpackhi_epi16(_t0, _t1);
-                _r45 = _mm_unpacklo_epi16(_t2, _t3);
-                _r67 = _mm_unpackhi_epi16(_t2, _t3);
-
-                __m128i _r0 = _mm_unpacklo_epi64(_r01, _r45);
-                __m128i _r1 = _mm_unpackhi_epi64(_r01, _r45);
-                __m128i _r2 = _mm_unpacklo_epi64(_r23, _r67);
-                __m128i _r3 = _mm_unpackhi_epi64(_r23, _r67);
-
-                _mm_store_si128((__m128i*)pp, _r0);
-                _mm_store_si128((__m128i*)(pp + 16), _r1);
-                _mm_store_si128((__m128i*)(pp + 32), _r2);
-                _mm_store_si128((__m128i*)(pp + 48), _r3);
-
-                pp += 64;
-                p0 += bottom_blob.cstep * 8;
-            }
-        }
-
-        if (elempack == 1)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k) + (j + jj);
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                __m128i _r0 = _mm_loadl_epi64((const __m128i*)p0);
-                __m128i _r1 = _mm_loadl_epi64((const __m128i*)(p0 + bottom_blob.cstep));
-                __m128i _r01 = _mm_unpacklo_epi8(_r0, _r1);
-                _mm_storeu_si128((__m128i*)pp, _r01);
-                pp += 16;
-                p0 += bottom_blob.cstep * 2;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                _mm_storel_epi64((__m128i*)pp, _mm_loadl_epi64((const __m128i*)p0));
-                pp += 8;
-                p0 += bottom_blob.cstep;
-            }
-        }
-    }
-#endif // defined(__x86_64__) || defined(_M_X64)
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-        if (elempack == 8)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k / 8) + (j + jj) * 8;
-
-            int kk = 0;
-            for (; kk < max_kk / 8; kk++)
-            {
-                __m128i _r01 = _mm_load_si128((const __m128i*)p0);
-                __m128i _r23 = _mm_load_si128((const __m128i*)(p0 + 16));
-
-                __m128i _t0 = _mm_unpacklo_epi16(_r01, _r23);
-                __m128i _t1 = _mm_unpackhi_epi16(_r01, _r23);
-
-                __m128i _r0 = _mm_unpacklo_epi16(_t0, _t1);
-                __m128i _r1 = _mm_unpackhi_epi16(_t0, _t1);
-
-                _mm_store_si128((__m128i*)pp, _r0);
-                _mm_store_si128((__m128i*)(pp + 16), _r1);
-
-                pp += 32;
-                p0 += bottom_blob.cstep * 8;
-            }
-        }
-
-        if (elempack == 1)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k) + (j + jj);
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[bottom_blob.cstep + 0];
-                pp[2] = p0[1];
-                pp[3] = p0[bottom_blob.cstep + 1];
-                pp[4] = p0[2];
-                pp[5] = p0[bottom_blob.cstep + 2];
-                pp[6] = p0[3];
-                pp[7] = p0[bottom_blob.cstep + 3];
-                pp += 8;
-                p0 += bottom_blob.cstep * 2;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                pp[2] = p0[2];
-                pp[3] = p0[3];
-                pp += 4;
-                p0 += bottom_blob.cstep;
-            }
-        }
-    }
-#endif // __SSE2__
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-#if __SSE2__
-        if (elempack == 8)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k / 8) + (j + jj) * 8;
-
-            int kk = 0;
-            for (; kk < max_kk / 8; kk++)
-            {
-                __m128i _r0 = _mm_loadl_epi64((const __m128i*)p0);
-                __m128i _r1 = _mm_loadl_epi64((const __m128i*)(p0 + 8));
-                __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                _mm_storeu_si128((__m128i*)pp, _r01);
-                pp += 16;
-                p0 += bottom_blob.cstep * 8;
-            }
-        }
-#endif // __SSE2__
-
-        if (elempack == 1)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k) + (j + jj);
-
-            int kk = 0;
-            for (; kk + 1 < max_kk; kk += 2)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[bottom_blob.cstep];
-                pp[2] = p0[1];
-                pp[3] = p0[bottom_blob.cstep + 1];
-                pp += 4;
-                p0 += bottom_blob.cstep * 2;
-            }
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp[1] = p0[1];
-                pp += 2;
-                p0 += bottom_blob.cstep;
-            }
-        }
-    }
-    for (; jj < max_jj; jj++)
-    {
-#if __SSE2__
-        if (elempack == 8)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k / 8) + (j + jj) * 8;
-
-            int kk = 0;
-            for (; kk < max_kk / 8; kk++)
-            {
-                _mm_storel_epi64((__m128i*)pp, _mm_loadl_epi64((const __m128i*)p0));
-                pp += 8;
-                p0 += bottom_blob.cstep * 8;
-            }
-        }
-#endif // __SSE2__
-
-        if (elempack == 1)
-        {
-            const signed char* p0 = (const signed char*)bottom_blob.channel(k) + (j + jj);
-
-            int kk = 0;
-            for (; kk < max_kk; kk++)
-            {
-                pp[0] = p0[0];
-                pp += 1;
-                p0 += bottom_blob.cstep;
-            }
-        }
-    }
-}
-
-template<int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h>
-#if __AVX512F__
-void convolution_im2col_input_tile_int8_avx512(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk)
-#else  // __AVX512F__
-void convolution_im2col_input_tile_int8(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk)
-#endif // __AVX512F__
-{
-    const int w = bottom_blob.w;
-    // const int channels = bottom_blob.c;
-    const int elempack = bottom_blob.elempack;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int outw = (w - kernel_extent_w) / stride_w + 1;
-
-    // j max_jj     outw*outh    split w and h
-
-    // k max_kk     pa*maxk*(inch/pa)    split inch
-
-    // k/max_kk shall be multiple of maxk
-
-    const int maxk = kernel_w * kernel_h;
-
-    signed char* pp = B;
-
-    int jj = 0;
-#if __SSE2__
-#if defined(__x86_64__) || defined(_M_X64)
-#if __AVX512F__
-    for (; jj + 15 < max_jj; jj += 16)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dy2 = (j + jj + 2) / outw;
-        int dy3 = (j + jj + 3) / outw;
-        int dy4 = (j + jj + 4) / outw;
-        int dy5 = (j + jj + 5) / outw;
-        int dy6 = (j + jj + 6) / outw;
-        int dy7 = (j + jj + 7) / outw;
-        int dy8 = (j + jj + 8) / outw;
-        int dy9 = (j + jj + 9) / outw;
-        int dya = (j + jj + 10) / outw;
-        int dyb = (j + jj + 11) / outw;
-        int dyc = (j + jj + 12) / outw;
-        int dyd = (j + jj + 13) / outw;
-        int dye = (j + jj + 14) / outw;
-        int dyf = (j + jj + 15) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-        int dx2 = (j + jj + 2) % outw;
-        int dx3 = (j + jj + 3) % outw;
-        int dx4 = (j + jj + 4) % outw;
-        int dx5 = (j + jj + 5) % outw;
-        int dx6 = (j + jj + 6) % outw;
-        int dx7 = (j + jj + 7) % outw;
-        int dx8 = (j + jj + 8) % outw;
-        int dx9 = (j + jj + 9) % outw;
-        int dxa = (j + jj + 10) % outw;
-        int dxb = (j + jj + 11) % outw;
-        int dxc = (j + jj + 12) % outw;
-        int dxd = (j + jj + 13) % outw;
-        int dxe = (j + jj + 14) % outw;
-        int dxf = (j + jj + 15) % outw;
-
-        if (dy0 == dyf)
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-
-                    if (stride_w == 1)
-                    {
-                        __m128i _r0 = _mm_loadu_si128((const __m128i*)sptr0);
-                        __m128i _r1 = _mm_loadu_si128((const __m128i*)sptr1);
-                        __m128i _tmp0 = _mm_unpacklo_epi8(_r0, _r1);
-                        __m128i _tmp1 = _mm_unpackhi_epi8(_r0, _r1);
-                        _mm_store_si128((__m128i*)pp, _tmp0);
-                        _mm_store_si128((__m128i*)(pp + 16), _tmp1);
-                        pp += 32;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        __m256i _r0 = _mm256_loadu_si256((const __m256i*)sptr0);
-                        __m256i _r1 = _mm256_loadu_si256((const __m256i*)sptr1);
-                        __m256i _tmp0 = _mm256_unpacklo_epi8(_r0, _r1);
-                        __m256i _tmp1 = _mm256_unpackhi_epi8(_r0, _r1);
-                        _tmp0 = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(_tmp0, _MM_SHUFFLE(3, 1, 2, 0)), _MM_SHUFFLE(3, 1, 2, 0));
-                        _tmp1 = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(_tmp1, _MM_SHUFFLE(3, 1, 2, 0)), _MM_SHUFFLE(3, 1, 2, 0));
-                        _tmp0 = _mm256_shuffle_epi32(_tmp0, _MM_SHUFFLE(3, 1, 2, 0));
-                        _tmp1 = _mm256_shuffle_epi32(_tmp1, _MM_SHUFFLE(3, 1, 2, 0));
-                        __m256i _r01 = _mm256_unpacklo_epi64(_tmp0, _tmp1);
-                        _mm256_storeu_si256((__m256i*)pp, _r01);
-                        pp += 32;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr0[stride_w];
-                        pp[3] = sptr1[stride_w];
-                        pp[4] = sptr0[stride_w * 2];
-                        pp[5] = sptr1[stride_w * 2];
-                        pp[6] = sptr0[stride_w * 3];
-                        pp[7] = sptr1[stride_w * 3];
-                        pp[8] = sptr0[stride_w * 4];
-                        pp[9] = sptr1[stride_w * 4];
-                        pp[10] = sptr0[stride_w * 5];
-                        pp[11] = sptr1[stride_w * 5];
-                        pp[12] = sptr0[stride_w * 6];
-                        pp[13] = sptr1[stride_w * 6];
-                        pp[14] = sptr0[stride_w * 7];
-                        pp[15] = sptr1[stride_w * 7];
-                        pp[16 + 0] = sptr0[stride_w * 8];
-                        pp[16 + 1] = sptr1[stride_w * 8];
-                        pp[16 + 2] = sptr0[stride_w * 9];
-                        pp[16 + 3] = sptr1[stride_w * 9];
-                        pp[16 + 4] = sptr0[stride_w * 10];
-                        pp[16 + 5] = sptr1[stride_w * 10];
-                        pp[16 + 6] = sptr0[stride_w * 11];
-                        pp[16 + 7] = sptr1[stride_w * 11];
-                        pp[16 + 8] = sptr0[stride_w * 12];
-                        pp[16 + 9] = sptr1[stride_w * 12];
-                        pp[16 + 10] = sptr0[stride_w * 13];
-                        pp[16 + 11] = sptr1[stride_w * 13];
-                        pp[16 + 12] = sptr0[stride_w * 14];
-                        pp[16 + 13] = sptr1[stride_w * 14];
-                        pp[16 + 14] = sptr0[stride_w * 15];
-                        pp[16 + 15] = sptr1[stride_w * 15];
-                        pp += 32;
-                    }
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-
-                const signed char* sptr = img.row<const signed char>(y0) + x0 * elempack;
-
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 8));
-                    __m128i _r2 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 16));
-                    __m128i _r3 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 24));
-                    __m128i _r4 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 32));
-                    __m128i _r5 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 40));
-                    __m128i _r6 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 48));
-                    __m128i _r7 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 56));
-                    __m128i _r8 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 64));
-                    __m128i _r9 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 72));
-                    __m128i _ra = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 80));
-                    __m128i _rb = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 88));
-                    __m128i _rc = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 96));
-                    __m128i _rd = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 104));
-                    __m128i _re = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 112));
-                    __m128i _rf = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 120));
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    __m128i _r23 = _mm_unpacklo_epi16(_r2, _r3);
-                    __m128i _r45 = _mm_unpacklo_epi16(_r4, _r5);
-                    __m128i _r67 = _mm_unpacklo_epi16(_r6, _r7);
-                    __m128i _r89 = _mm_unpacklo_epi16(_r8, _r9);
-                    __m128i _rab = _mm_unpacklo_epi16(_ra, _rb);
-                    __m128i _rcd = _mm_unpacklo_epi16(_rc, _rd);
-                    __m128i _ref = _mm_unpacklo_epi16(_re, _rf);
-                    _r0 = _mm_unpacklo_epi32(_r01, _r23);
-                    _r1 = _mm_unpackhi_epi32(_r01, _r23);
-                    _r2 = _mm_unpacklo_epi32(_r45, _r67);
-                    _r3 = _mm_unpackhi_epi32(_r45, _r67);
-                    _r4 = _mm_unpacklo_epi32(_r89, _rab);
-                    _r5 = _mm_unpackhi_epi32(_r89, _rab);
-                    _r6 = _mm_unpacklo_epi32(_rcd, _ref);
-                    _r7 = _mm_unpackhi_epi32(_rcd, _ref);
-                    _r8 = _mm_unpacklo_epi64(_r0, _r2);
-                    _r9 = _mm_unpacklo_epi64(_r4, _r6);
-                    _ra = _mm_unpackhi_epi64(_r0, _r2);
-                    _rb = _mm_unpackhi_epi64(_r4, _r6);
-                    _rc = _mm_unpacklo_epi64(_r1, _r3);
-                    _rd = _mm_unpacklo_epi64(_r5, _r7);
-                    _re = _mm_unpackhi_epi64(_r1, _r3);
-                    _rf = _mm_unpackhi_epi64(_r5, _r7);
-                    _mm_store_si128((__m128i*)pp, _r8);
-                    _mm_store_si128((__m128i*)(pp + 16), _r9);
-                    _mm_store_si128((__m128i*)(pp + 32), _ra);
-                    _mm_store_si128((__m128i*)(pp + 48), _rb);
-                    _mm_store_si128((__m128i*)(pp + 64), _rc);
-                    _mm_store_si128((__m128i*)(pp + 80), _rd);
-                    _mm_store_si128((__m128i*)(pp + 96), _re);
-                    _mm_store_si128((__m128i*)(pp + 112), _rf);
-                    pp += 128;
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr[0];
-                    pp[1] = sptr[stride_w];
-                    pp[2] = sptr[stride_w * 2];
-                    pp[3] = sptr[stride_w * 3];
-                    pp[4] = sptr[stride_w * 4];
-                    pp[5] = sptr[stride_w * 5];
-                    pp[6] = sptr[stride_w * 6];
-                    pp[7] = sptr[stride_w * 7];
-                    pp[8] = sptr[stride_w * 8];
-                    pp[9] = sptr[stride_w * 9];
-                    pp[10] = sptr[stride_w * 10];
-                    pp[11] = sptr[stride_w * 11];
-                    pp[12] = sptr[stride_w * 12];
-                    pp[13] = sptr[stride_w * 13];
-                    pp[14] = sptr[stride_w * 14];
-                    pp[15] = sptr[stride_w * 15];
-                    pp += 16;
-                }
-            }
-        }
-        else
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int x01 = stride_w * dx1 + dilation_w * v0;
-                    int x02 = stride_w * dx2 + dilation_w * v0;
-                    int x03 = stride_w * dx3 + dilation_w * v0;
-                    int x04 = stride_w * dx4 + dilation_w * v0;
-                    int x05 = stride_w * dx5 + dilation_w * v0;
-                    int x06 = stride_w * dx6 + dilation_w * v0;
-                    int x07 = stride_w * dx7 + dilation_w * v0;
-                    int x08 = stride_w * dx8 + dilation_w * v0;
-                    int x09 = stride_w * dx9 + dilation_w * v0;
-                    int x0a = stride_w * dxa + dilation_w * v0;
-                    int x0b = stride_w * dxb + dilation_w * v0;
-                    int x0c = stride_w * dxc + dilation_w * v0;
-                    int x0d = stride_w * dxd + dilation_w * v0;
-                    int x0e = stride_w * dxe + dilation_w * v0;
-                    int x0f = stride_w * dxf + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int y01 = stride_h * dy1 + dilation_h * u0;
-                    int y02 = stride_h * dy2 + dilation_h * u0;
-                    int y03 = stride_h * dy3 + dilation_h * u0;
-                    int y04 = stride_h * dy4 + dilation_h * u0;
-                    int y05 = stride_h * dy5 + dilation_h * u0;
-                    int y06 = stride_h * dy6 + dilation_h * u0;
-                    int y07 = stride_h * dy7 + dilation_h * u0;
-                    int y08 = stride_h * dy8 + dilation_h * u0;
-                    int y09 = stride_h * dy9 + dilation_h * u0;
-                    int y0a = stride_h * dya + dilation_h * u0;
-                    int y0b = stride_h * dyb + dilation_h * u0;
-                    int y0c = stride_h * dyc + dilation_h * u0;
-                    int y0d = stride_h * dyd + dilation_h * u0;
-                    int y0e = stride_h * dye + dilation_h * u0;
-                    int y0f = stride_h * dyf + dilation_h * u0;
-
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int x11 = stride_w * dx1 + dilation_w * v1;
-                    int x12 = stride_w * dx2 + dilation_w * v1;
-                    int x13 = stride_w * dx3 + dilation_w * v1;
-                    int x14 = stride_w * dx4 + dilation_w * v1;
-                    int x15 = stride_w * dx5 + dilation_w * v1;
-                    int x16 = stride_w * dx6 + dilation_w * v1;
-                    int x17 = stride_w * dx7 + dilation_w * v1;
-                    int x18 = stride_w * dx8 + dilation_w * v1;
-                    int x19 = stride_w * dx9 + dilation_w * v1;
-                    int x1a = stride_w * dxa + dilation_w * v1;
-                    int x1b = stride_w * dxb + dilation_w * v1;
-                    int x1c = stride_w * dxc + dilation_w * v1;
-                    int x1d = stride_w * dxd + dilation_w * v1;
-                    int x1e = stride_w * dxe + dilation_w * v1;
-                    int x1f = stride_w * dxf + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-                    int y11 = stride_h * dy1 + dilation_h * u1;
-                    int y12 = stride_h * dy2 + dilation_h * u1;
-                    int y13 = stride_h * dy3 + dilation_h * u1;
-                    int y14 = stride_h * dy4 + dilation_h * u1;
-                    int y15 = stride_h * dy5 + dilation_h * u1;
-                    int y16 = stride_h * dy6 + dilation_h * u1;
-                    int y17 = stride_h * dy7 + dilation_h * u1;
-                    int y18 = stride_h * dy8 + dilation_h * u1;
-                    int y19 = stride_h * dy9 + dilation_h * u1;
-                    int y1a = stride_h * dya + dilation_h * u1;
-                    int y1b = stride_h * dyb + dilation_h * u1;
-                    int y1c = stride_h * dyc + dilation_h * u1;
-                    int y1d = stride_h * dyd + dilation_h * u1;
-                    int y1e = stride_h * dye + dilation_h * u1;
-                    int y1f = stride_h * dyf + dilation_h * u1;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-                    const signed char* sptr04 = img0.row<const signed char>(y04) + x04;
-                    const signed char* sptr05 = img0.row<const signed char>(y05) + x05;
-                    const signed char* sptr06 = img0.row<const signed char>(y06) + x06;
-                    const signed char* sptr07 = img0.row<const signed char>(y07) + x07;
-                    const signed char* sptr08 = img0.row<const signed char>(y08) + x08;
-                    const signed char* sptr09 = img0.row<const signed char>(y09) + x09;
-                    const signed char* sptr0a = img0.row<const signed char>(y0a) + x0a;
-                    const signed char* sptr0b = img0.row<const signed char>(y0b) + x0b;
-                    const signed char* sptr0c = img0.row<const signed char>(y0c) + x0c;
-                    const signed char* sptr0d = img0.row<const signed char>(y0d) + x0d;
-                    const signed char* sptr0e = img0.row<const signed char>(y0e) + x0e;
-                    const signed char* sptr0f = img0.row<const signed char>(y0f) + x0f;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-                    const signed char* sptr14 = img1.row<const signed char>(y14) + x14;
-                    const signed char* sptr15 = img1.row<const signed char>(y15) + x15;
-                    const signed char* sptr16 = img1.row<const signed char>(y16) + x16;
-                    const signed char* sptr17 = img1.row<const signed char>(y17) + x17;
-                    const signed char* sptr18 = img1.row<const signed char>(y18) + x18;
-                    const signed char* sptr19 = img1.row<const signed char>(y19) + x19;
-                    const signed char* sptr1a = img1.row<const signed char>(y1a) + x1a;
-                    const signed char* sptr1b = img1.row<const signed char>(y1b) + x1b;
-                    const signed char* sptr1c = img1.row<const signed char>(y1c) + x1c;
-                    const signed char* sptr1d = img1.row<const signed char>(y1d) + x1d;
-                    const signed char* sptr1e = img1.row<const signed char>(y1e) + x1e;
-                    const signed char* sptr1f = img1.row<const signed char>(y1f) + x1f;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr01[0];
-                    pp[3] = sptr11[0];
-                    pp[4] = sptr02[0];
-                    pp[5] = sptr12[0];
-                    pp[6] = sptr03[0];
-                    pp[7] = sptr13[0];
-                    pp[8] = sptr04[0];
-                    pp[9] = sptr14[0];
-                    pp[10] = sptr05[0];
-                    pp[11] = sptr15[0];
-                    pp[12] = sptr06[0];
-                    pp[13] = sptr16[0];
-                    pp[14] = sptr07[0];
-                    pp[15] = sptr17[0];
-                    pp[16 + 0] = sptr08[0];
-                    pp[16 + 1] = sptr18[0];
-                    pp[16 + 2] = sptr09[0];
-                    pp[16 + 3] = sptr19[0];
-                    pp[16 + 4] = sptr0a[0];
-                    pp[16 + 5] = sptr1a[0];
-                    pp[16 + 6] = sptr0b[0];
-                    pp[16 + 7] = sptr1b[0];
-                    pp[16 + 8] = sptr0c[0];
-                    pp[16 + 9] = sptr1c[0];
-                    pp[16 + 10] = sptr0d[0];
-                    pp[16 + 11] = sptr1d[0];
-                    pp[16 + 12] = sptr0e[0];
-                    pp[16 + 13] = sptr1e[0];
-                    pp[16 + 14] = sptr0f[0];
-                    pp[16 + 15] = sptr1f[0];
-                    pp += 32;
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int x1 = stride_w * dx1 + dilation_w * v;
-                int x2 = stride_w * dx2 + dilation_w * v;
-                int x3 = stride_w * dx3 + dilation_w * v;
-                int x4 = stride_w * dx4 + dilation_w * v;
-                int x5 = stride_w * dx5 + dilation_w * v;
-                int x6 = stride_w * dx6 + dilation_w * v;
-                int x7 = stride_w * dx7 + dilation_w * v;
-                int x8 = stride_w * dx8 + dilation_w * v;
-                int x9 = stride_w * dx9 + dilation_w * v;
-                int xa = stride_w * dxa + dilation_w * v;
-                int xb = stride_w * dxb + dilation_w * v;
-                int xc = stride_w * dxc + dilation_w * v;
-                int xd = stride_w * dxd + dilation_w * v;
-                int xe = stride_w * dxe + dilation_w * v;
-                int xf = stride_w * dxf + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-                int y1 = stride_h * dy1 + dilation_h * u;
-                int y2 = stride_h * dy2 + dilation_h * u;
-                int y3 = stride_h * dy3 + dilation_h * u;
-                int y4 = stride_h * dy4 + dilation_h * u;
-                int y5 = stride_h * dy5 + dilation_h * u;
-                int y6 = stride_h * dy6 + dilation_h * u;
-                int y7 = stride_h * dy7 + dilation_h * u;
-                int y8 = stride_h * dy8 + dilation_h * u;
-                int y9 = stride_h * dy9 + dilation_h * u;
-                int ya = stride_h * dya + dilation_h * u;
-                int yb = stride_h * dyb + dilation_h * u;
-                int yc = stride_h * dyc + dilation_h * u;
-                int yd = stride_h * dyd + dilation_h * u;
-                int ye = stride_h * dye + dilation_h * u;
-                int yf = stride_h * dyf + dilation_h * u;
-
-                const signed char* sptr0 = img.row<const signed char>(y0) + x0 * elempack;
-                const signed char* sptr1 = img.row<const signed char>(y1) + x1 * elempack;
-                const signed char* sptr2 = img.row<const signed char>(y2) + x2 * elempack;
-                const signed char* sptr3 = img.row<const signed char>(y3) + x3 * elempack;
-                const signed char* sptr4 = img.row<const signed char>(y4) + x4 * elempack;
-                const signed char* sptr5 = img.row<const signed char>(y5) + x5 * elempack;
-                const signed char* sptr6 = img.row<const signed char>(y6) + x6 * elempack;
-                const signed char* sptr7 = img.row<const signed char>(y7) + x7 * elempack;
-                const signed char* sptr8 = img.row<const signed char>(y8) + x8 * elempack;
-                const signed char* sptr9 = img.row<const signed char>(y9) + x9 * elempack;
-                const signed char* sptra = img.row<const signed char>(ya) + xa * elempack;
-                const signed char* sptrb = img.row<const signed char>(yb) + xb * elempack;
-                const signed char* sptrc = img.row<const signed char>(yc) + xc * elempack;
-                const signed char* sptrd = img.row<const signed char>(yd) + xd * elempack;
-                const signed char* sptre = img.row<const signed char>(ye) + xe * elempack;
-                const signed char* sptrf = img.row<const signed char>(yf) + xf * elempack;
-
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr0);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)sptr1);
-                    __m128i _r2 = _mm_loadl_epi64((const __m128i*)sptr2);
-                    __m128i _r3 = _mm_loadl_epi64((const __m128i*)sptr3);
-                    __m128i _r4 = _mm_loadl_epi64((const __m128i*)sptr4);
-                    __m128i _r5 = _mm_loadl_epi64((const __m128i*)sptr5);
-                    __m128i _r6 = _mm_loadl_epi64((const __m128i*)sptr6);
-                    __m128i _r7 = _mm_loadl_epi64((const __m128i*)sptr7);
-                    __m128i _r8 = _mm_loadl_epi64((const __m128i*)sptr8);
-                    __m128i _r9 = _mm_loadl_epi64((const __m128i*)sptr9);
-                    __m128i _ra = _mm_loadl_epi64((const __m128i*)sptra);
-                    __m128i _rb = _mm_loadl_epi64((const __m128i*)sptrb);
-                    __m128i _rc = _mm_loadl_epi64((const __m128i*)sptrc);
-                    __m128i _rd = _mm_loadl_epi64((const __m128i*)sptrd);
-                    __m128i _re = _mm_loadl_epi64((const __m128i*)sptre);
-                    __m128i _rf = _mm_loadl_epi64((const __m128i*)sptrf);
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    __m128i _r23 = _mm_unpacklo_epi16(_r2, _r3);
-                    __m128i _r45 = _mm_unpacklo_epi16(_r4, _r5);
-                    __m128i _r67 = _mm_unpacklo_epi16(_r6, _r7);
-                    __m128i _r89 = _mm_unpacklo_epi16(_r8, _r9);
-                    __m128i _rab = _mm_unpacklo_epi16(_ra, _rb);
-                    __m128i _rcd = _mm_unpacklo_epi16(_rc, _rd);
-                    __m128i _ref = _mm_unpacklo_epi16(_re, _rf);
-                    _r0 = _mm_unpacklo_epi32(_r01, _r23);
-                    _r1 = _mm_unpackhi_epi32(_r01, _r23);
-                    _r2 = _mm_unpacklo_epi32(_r45, _r67);
-                    _r3 = _mm_unpackhi_epi32(_r45, _r67);
-                    _r4 = _mm_unpacklo_epi32(_r89, _rab);
-                    _r5 = _mm_unpackhi_epi32(_r89, _rab);
-                    _r6 = _mm_unpacklo_epi32(_rcd, _ref);
-                    _r7 = _mm_unpackhi_epi32(_rcd, _ref);
-                    _r8 = _mm_unpacklo_epi64(_r0, _r2);
-                    _r9 = _mm_unpacklo_epi64(_r4, _r6);
-                    _ra = _mm_unpackhi_epi64(_r0, _r2);
-                    _rb = _mm_unpackhi_epi64(_r4, _r6);
-                    _rc = _mm_unpacklo_epi64(_r1, _r3);
-                    _rd = _mm_unpacklo_epi64(_r5, _r7);
-                    _re = _mm_unpackhi_epi64(_r1, _r3);
-                    _rf = _mm_unpackhi_epi64(_r5, _r7);
-                    _mm_store_si128((__m128i*)pp, _r8);
-                    _mm_store_si128((__m128i*)(pp + 16), _r9);
-                    _mm_store_si128((__m128i*)(pp + 32), _ra);
-                    _mm_store_si128((__m128i*)(pp + 48), _rb);
-                    _mm_store_si128((__m128i*)(pp + 64), _rc);
-                    _mm_store_si128((__m128i*)(pp + 80), _rd);
-                    _mm_store_si128((__m128i*)(pp + 96), _re);
-                    _mm_store_si128((__m128i*)(pp + 112), _rf);
-                    pp += 128;
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr2[0];
-                    pp[3] = sptr3[0];
-                    pp[4] = sptr4[0];
-                    pp[5] = sptr5[0];
-                    pp[6] = sptr6[0];
-                    pp[7] = sptr7[0];
-                    pp[8] = sptr8[0];
-                    pp[9] = sptr9[0];
-                    pp[10] = sptra[0];
-                    pp[11] = sptrb[0];
-                    pp[12] = sptrc[0];
-                    pp[13] = sptrd[0];
-                    pp[14] = sptre[0];
-                    pp[15] = sptrf[0];
-                    pp += 16;
-                }
-            }
-        }
-    }
-#endif // __AVX512F__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dy2 = (j + jj + 2) / outw;
-        int dy3 = (j + jj + 3) / outw;
-        int dy4 = (j + jj + 4) / outw;
-        int dy5 = (j + jj + 5) / outw;
-        int dy6 = (j + jj + 6) / outw;
-        int dy7 = (j + jj + 7) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-        int dx2 = (j + jj + 2) % outw;
-        int dx3 = (j + jj + 3) % outw;
-        int dx4 = (j + jj + 4) % outw;
-        int dx5 = (j + jj + 5) % outw;
-        int dx6 = (j + jj + 6) % outw;
-        int dx7 = (j + jj + 7) % outw;
-
-        if (dy0 == dy7)
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-
-                    if (stride_w == 1)
-                    {
-                        __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr0);
-                        __m128i _r1 = _mm_loadl_epi64((const __m128i*)sptr1);
-                        __m128i _r01 = _mm_unpacklo_epi8(_r0, _r1);
-                        _mm_storeu_si128((__m128i*)pp, _r01);
-                        pp += 16;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        __m128i _r0 = _mm_loadu_si128((const __m128i*)sptr0);
-                        __m128i _r1 = _mm_loadu_si128((const __m128i*)sptr1);
-                        __m128i _tmp0 = _mm_unpacklo_epi8(_r0, _r1);
-                        __m128i _tmp1 = _mm_unpackhi_epi8(_r0, _r1);
-                        _tmp0 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_tmp0, _MM_SHUFFLE(3, 1, 2, 0)), _MM_SHUFFLE(3, 1, 2, 0));
-                        _tmp1 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_tmp1, _MM_SHUFFLE(3, 1, 2, 0)), _MM_SHUFFLE(3, 1, 2, 0));
-                        _tmp0 = _mm_shuffle_epi32(_tmp0, _MM_SHUFFLE(3, 1, 2, 0));
-                        _tmp1 = _mm_shuffle_epi32(_tmp1, _MM_SHUFFLE(3, 1, 2, 0));
-                        __m128i _r01 = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(_tmp0), _mm_castsi128_ps(_tmp1), _MM_SHUFFLE(1, 0, 1, 0)));
-                        _mm_storeu_si128((__m128i*)pp, _r01);
-                        pp += 16;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr0[stride_w];
-                        pp[3] = sptr1[stride_w];
-                        pp[4] = sptr0[stride_w * 2];
-                        pp[5] = sptr1[stride_w * 2];
-                        pp[6] = sptr0[stride_w * 3];
-                        pp[7] = sptr1[stride_w * 3];
-                        pp[8] = sptr0[stride_w * 4];
-                        pp[9] = sptr1[stride_w * 4];
-                        pp[10] = sptr0[stride_w * 5];
-                        pp[11] = sptr1[stride_w * 5];
-                        pp[12] = sptr0[stride_w * 6];
-                        pp[13] = sptr1[stride_w * 6];
-                        pp[14] = sptr0[stride_w * 7];
-                        pp[15] = sptr1[stride_w * 7];
-                        pp += 16;
-                    }
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-
-                const signed char* sptr = img.row<const signed char>(y0) + x0 * elempack;
-
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 8));
-                    __m128i _r2 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 16));
-                    __m128i _r3 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 24));
-                    __m128i _r4 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 32));
-                    __m128i _r5 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 40));
-                    __m128i _r6 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 48));
-                    __m128i _r7 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 56));
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    __m128i _r23 = _mm_unpacklo_epi16(_r2, _r3);
-                    __m128i _r45 = _mm_unpacklo_epi16(_r4, _r5);
-                    __m128i _r67 = _mm_unpacklo_epi16(_r6, _r7);
-                    _r0 = _mm_unpacklo_epi32(_r01, _r23);
-                    _r1 = _mm_unpackhi_epi32(_r01, _r23);
-                    _r2 = _mm_unpacklo_epi32(_r45, _r67);
-                    _r3 = _mm_unpackhi_epi32(_r45, _r67);
-                    _r4 = _mm_unpacklo_epi64(_r0, _r2);
-                    _r5 = _mm_unpackhi_epi64(_r0, _r2);
-                    _r6 = _mm_unpacklo_epi64(_r1, _r3);
-                    _r7 = _mm_unpackhi_epi64(_r1, _r3);
-                    _mm_storeu_si128((__m128i*)pp, _r4);
-                    _mm_storeu_si128((__m128i*)(pp + 16), _r5);
-                    _mm_storeu_si128((__m128i*)(pp + 32), _r6);
-                    _mm_storeu_si128((__m128i*)(pp + 48), _r7);
-                    pp += 64;
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr[0];
-                    pp[1] = sptr[stride_w];
-                    pp[2] = sptr[stride_w * 2];
-                    pp[3] = sptr[stride_w * 3];
-                    pp[4] = sptr[stride_w * 4];
-                    pp[5] = sptr[stride_w * 5];
-                    pp[6] = sptr[stride_w * 6];
-                    pp[7] = sptr[stride_w * 7];
-                    pp += 8;
-                }
-            }
-        }
-        else
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int x01 = stride_w * dx1 + dilation_w * v0;
-                    int x02 = stride_w * dx2 + dilation_w * v0;
-                    int x03 = stride_w * dx3 + dilation_w * v0;
-                    int x04 = stride_w * dx4 + dilation_w * v0;
-                    int x05 = stride_w * dx5 + dilation_w * v0;
-                    int x06 = stride_w * dx6 + dilation_w * v0;
-                    int x07 = stride_w * dx7 + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int y01 = stride_h * dy1 + dilation_h * u0;
-                    int y02 = stride_h * dy2 + dilation_h * u0;
-                    int y03 = stride_h * dy3 + dilation_h * u0;
-                    int y04 = stride_h * dy4 + dilation_h * u0;
-                    int y05 = stride_h * dy5 + dilation_h * u0;
-                    int y06 = stride_h * dy6 + dilation_h * u0;
-                    int y07 = stride_h * dy7 + dilation_h * u0;
-
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int x11 = stride_w * dx1 + dilation_w * v1;
-                    int x12 = stride_w * dx2 + dilation_w * v1;
-                    int x13 = stride_w * dx3 + dilation_w * v1;
-                    int x14 = stride_w * dx4 + dilation_w * v1;
-                    int x15 = stride_w * dx5 + dilation_w * v1;
-                    int x16 = stride_w * dx6 + dilation_w * v1;
-                    int x17 = stride_w * dx7 + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-                    int y11 = stride_h * dy1 + dilation_h * u1;
-                    int y12 = stride_h * dy2 + dilation_h * u1;
-                    int y13 = stride_h * dy3 + dilation_h * u1;
-                    int y14 = stride_h * dy4 + dilation_h * u1;
-                    int y15 = stride_h * dy5 + dilation_h * u1;
-                    int y16 = stride_h * dy6 + dilation_h * u1;
-                    int y17 = stride_h * dy7 + dilation_h * u1;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-                    const signed char* sptr04 = img0.row<const signed char>(y04) + x04;
-                    const signed char* sptr05 = img0.row<const signed char>(y05) + x05;
-                    const signed char* sptr06 = img0.row<const signed char>(y06) + x06;
-                    const signed char* sptr07 = img0.row<const signed char>(y07) + x07;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-                    const signed char* sptr14 = img1.row<const signed char>(y14) + x14;
-                    const signed char* sptr15 = img1.row<const signed char>(y15) + x15;
-                    const signed char* sptr16 = img1.row<const signed char>(y16) + x16;
-                    const signed char* sptr17 = img1.row<const signed char>(y17) + x17;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr01[0];
-                    pp[3] = sptr11[0];
-                    pp[4] = sptr02[0];
-                    pp[5] = sptr12[0];
-                    pp[6] = sptr03[0];
-                    pp[7] = sptr13[0];
-                    pp[8] = sptr04[0];
-                    pp[9] = sptr14[0];
-                    pp[10] = sptr05[0];
-                    pp[11] = sptr15[0];
-                    pp[12] = sptr06[0];
-                    pp[13] = sptr16[0];
-                    pp[14] = sptr07[0];
-                    pp[15] = sptr17[0];
-                    pp += 16;
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int x1 = stride_w * dx1 + dilation_w * v;
-                int x2 = stride_w * dx2 + dilation_w * v;
-                int x3 = stride_w * dx3 + dilation_w * v;
-                int x4 = stride_w * dx4 + dilation_w * v;
-                int x5 = stride_w * dx5 + dilation_w * v;
-                int x6 = stride_w * dx6 + dilation_w * v;
-                int x7 = stride_w * dx7 + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-                int y1 = stride_h * dy1 + dilation_h * u;
-                int y2 = stride_h * dy2 + dilation_h * u;
-                int y3 = stride_h * dy3 + dilation_h * u;
-                int y4 = stride_h * dy4 + dilation_h * u;
-                int y5 = stride_h * dy5 + dilation_h * u;
-                int y6 = stride_h * dy6 + dilation_h * u;
-                int y7 = stride_h * dy7 + dilation_h * u;
-
-                const signed char* sptr0 = img.row<const signed char>(y0) + x0 * elempack;
-                const signed char* sptr1 = img.row<const signed char>(y1) + x1 * elempack;
-                const signed char* sptr2 = img.row<const signed char>(y2) + x2 * elempack;
-                const signed char* sptr3 = img.row<const signed char>(y3) + x3 * elempack;
-                const signed char* sptr4 = img.row<const signed char>(y4) + x4 * elempack;
-                const signed char* sptr5 = img.row<const signed char>(y5) + x5 * elempack;
-                const signed char* sptr6 = img.row<const signed char>(y6) + x6 * elempack;
-                const signed char* sptr7 = img.row<const signed char>(y7) + x7 * elempack;
-
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr0);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)sptr1);
-                    __m128i _r2 = _mm_loadl_epi64((const __m128i*)sptr2);
-                    __m128i _r3 = _mm_loadl_epi64((const __m128i*)sptr3);
-                    __m128i _r4 = _mm_loadl_epi64((const __m128i*)sptr4);
-                    __m128i _r5 = _mm_loadl_epi64((const __m128i*)sptr5);
-                    __m128i _r6 = _mm_loadl_epi64((const __m128i*)sptr6);
-                    __m128i _r7 = _mm_loadl_epi64((const __m128i*)sptr7);
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    __m128i _r23 = _mm_unpacklo_epi16(_r2, _r3);
-                    __m128i _r45 = _mm_unpacklo_epi16(_r4, _r5);
-                    __m128i _r67 = _mm_unpacklo_epi16(_r6, _r7);
-                    _r0 = _mm_unpacklo_epi32(_r01, _r23);
-                    _r1 = _mm_unpackhi_epi32(_r01, _r23);
-                    _r2 = _mm_unpacklo_epi32(_r45, _r67);
-                    _r3 = _mm_unpackhi_epi32(_r45, _r67);
-                    _r4 = _mm_unpacklo_epi64(_r0, _r2);
-                    _r5 = _mm_unpackhi_epi64(_r0, _r2);
-                    _r6 = _mm_unpacklo_epi64(_r1, _r3);
-                    _r7 = _mm_unpackhi_epi64(_r1, _r3);
-                    _mm_storeu_si128((__m128i*)pp, _r4);
-                    _mm_storeu_si128((__m128i*)(pp + 16), _r5);
-                    _mm_storeu_si128((__m128i*)(pp + 32), _r6);
-                    _mm_storeu_si128((__m128i*)(pp + 48), _r7);
-                    pp += 64;
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr2[0];
-                    pp[3] = sptr3[0];
-                    pp[4] = sptr4[0];
-                    pp[5] = sptr5[0];
-                    pp[6] = sptr6[0];
-                    pp[7] = sptr7[0];
-                    pp += 8;
-                }
-            }
-        }
-    }
-#endif // defined(__x86_64__) || defined(_M_X64)
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dy2 = (j + jj + 2) / outw;
-        int dy3 = (j + jj + 3) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-        int dx2 = (j + jj + 2) % outw;
-        int dx3 = (j + jj + 3) % outw;
-
-        if (dy0 == dy3)
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-
-                    if (stride_w == 1)
-                    {
-                        __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr0);
-                        __m128i _r1 = _mm_loadl_epi64((const __m128i*)sptr1);
-                        __m128i _r01 = _mm_unpacklo_epi8(_r0, _r1);
-                        _mm_storel_epi64((__m128i*)pp, _r01);
-                        pp += 8;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr0);
-                        __m128i _r1 = _mm_loadl_epi64((const __m128i*)sptr1);
-                        __m128i _r01 = _mm_unpacklo_epi8(_r0, _r1);
-                        _r01 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_r01, _MM_SHUFFLE(3, 1, 2, 0)), _MM_SHUFFLE(3, 1, 2, 0));
-                        _r01 = _mm_shuffle_epi32(_r01, _MM_SHUFFLE(3, 1, 2, 0));
-                        _mm_storel_epi64((__m128i*)pp, _r01);
-                        pp += 8;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr0[stride_w];
-                        pp[3] = sptr1[stride_w];
-                        pp[4] = sptr0[stride_w * 2];
-                        pp[5] = sptr1[stride_w * 2];
-                        pp[6] = sptr0[stride_w * 3];
-                        pp[7] = sptr1[stride_w * 3];
-                        pp += 8;
-                    }
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-
-                const signed char* sptr = img.row<const signed char>(y0) + x0 * elempack;
-
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 8));
-                    __m128i _r2 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 16));
-                    __m128i _r3 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 24));
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    __m128i _r23 = _mm_unpacklo_epi16(_r2, _r3);
-                    _r0 = _mm_unpacklo_epi32(_r01, _r23);
-                    _r1 = _mm_unpackhi_epi32(_r01, _r23);
-                    _mm_storeu_si128((__m128i*)pp, _r0);
-                    _mm_storeu_si128((__m128i*)(pp + 16), _r1);
-                    pp += 32;
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr[0];
-                    pp[1] = sptr[stride_w];
-                    pp[2] = sptr[stride_w * 2];
-                    pp[3] = sptr[stride_w * 3];
-                    pp += 4;
-                }
-            }
-        }
-        else
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int x01 = stride_w * dx1 + dilation_w * v0;
-                    int x02 = stride_w * dx2 + dilation_w * v0;
-                    int x03 = stride_w * dx3 + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int y01 = stride_h * dy1 + dilation_h * u0;
-                    int y02 = stride_h * dy2 + dilation_h * u0;
-                    int y03 = stride_h * dy3 + dilation_h * u0;
-
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int x11 = stride_w * dx1 + dilation_w * v1;
-                    int x12 = stride_w * dx2 + dilation_w * v1;
-                    int x13 = stride_w * dx3 + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-                    int y11 = stride_h * dy1 + dilation_h * u1;
-                    int y12 = stride_h * dy2 + dilation_h * u1;
-                    int y13 = stride_h * dy3 + dilation_h * u1;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr01[0];
-                    pp[3] = sptr11[0];
-                    pp[4] = sptr02[0];
-                    pp[5] = sptr12[0];
-                    pp[6] = sptr03[0];
-                    pp[7] = sptr13[0];
-                    pp += 8;
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int x1 = stride_w * dx1 + dilation_w * v;
-                int x2 = stride_w * dx2 + dilation_w * v;
-                int x3 = stride_w * dx3 + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-                int y1 = stride_h * dy1 + dilation_h * u;
-                int y2 = stride_h * dy2 + dilation_h * u;
-                int y3 = stride_h * dy3 + dilation_h * u;
-
-                const signed char* sptr0 = img.row<const signed char>(y0) + x0 * elempack;
-                const signed char* sptr1 = img.row<const signed char>(y1) + x1 * elempack;
-                const signed char* sptr2 = img.row<const signed char>(y2) + x2 * elempack;
-                const signed char* sptr3 = img.row<const signed char>(y3) + x3 * elempack;
-
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr0);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)sptr1);
-                    __m128i _r2 = _mm_loadl_epi64((const __m128i*)sptr2);
-                    __m128i _r3 = _mm_loadl_epi64((const __m128i*)sptr3);
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    __m128i _r23 = _mm_unpacklo_epi16(_r2, _r3);
-                    _r0 = _mm_unpacklo_epi32(_r01, _r23);
-                    _r1 = _mm_unpackhi_epi32(_r01, _r23);
-                    _mm_storeu_si128((__m128i*)pp, _r0);
-                    _mm_storeu_si128((__m128i*)(pp + 16), _r1);
-                    pp += 32;
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr2[0];
-                    pp[3] = sptr3[0];
-                    pp += 4;
-                }
-            }
-        }
-    }
-#endif // __SSE2__
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-
-        if (dy0 == dy1)
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr0[stride_w];
-                    pp[3] = sptr1[stride_w];
-                    pp += 4;
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-
-                const signed char* sptr = img.row<const signed char>(y0) + x0 * elempack;
-
-#if __SSE2__
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 8));
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    _mm_storeu_si128((__m128i*)pp, _r01);
-                    pp += 16;
-                }
-#endif // __SSE2__
-                if (elempack == 1)
-                {
-                    pp[0] = sptr[0];
-                    pp[1] = sptr[stride_w];
-                    pp += 2;
-                }
-            }
-        }
-        else
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int x01 = stride_w * dx1 + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int y01 = stride_h * dy1 + dilation_h * u0;
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int x11 = stride_w * dx1 + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-                    int y11 = stride_h * dy1 + dilation_h * u1;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr01[0];
-                    pp[3] = sptr11[0];
-                    pp += 4;
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int x1 = stride_w * dx1 + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-                int y1 = stride_h * dy1 + dilation_h * u;
-
-                const signed char* sptr0 = img.row<const signed char>(y0) + x0 * elempack;
-                const signed char* sptr1 = img.row<const signed char>(y1) + x1 * elempack;
-
-#if __SSE2__
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr0);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)sptr1);
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    _mm_storeu_si128((__m128i*)pp, _r01);
-                    pp += 16;
-                }
-#endif // __SSE2__
-                if (elempack == 1)
-                {
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp += 2;
-                }
-            }
-        }
-    }
-    for (; jj < max_jj; jj++)
-    {
-        int dy = (j + jj) / outw;
-        int dx = (j + jj) % outw;
-
-        int kk = 0;
-        for (; kk < max_kk / elempack; kk++)
-        {
-            int p = (k / elempack + kk) / maxk;
-            int uv = (k / elempack + kk) % maxk;
-            int u = uv / kernel_w;
-            int v = uv % kernel_w;
-
-            const Mat img = bottom_blob.channel(p);
-
-            int x = stride_w * dx + dilation_w * v;
-            int y = stride_h * dy + dilation_h * u;
-
-            const signed char* sptr = img.row<const signed char>(y) + x * elempack;
-
-#if __SSE2__
-            if (elempack == 8)
-            {
-                _mm_storel_epi64((__m128i*)pp, _mm_loadl_epi64((const __m128i*)sptr));
-                pp += 8;
-            }
-#endif // __SSE2__
-            if (elempack == 1)
-            {
-                pp[0] = sptr[0];
-                pp += 1;
-            }
-        }
-    }
-}
-
-#if __AVX512F__
-template void convolution_im2col_input_tile_int8_avx512<1, 1, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8_avx512<3, 3, 1, 1, 1, 1>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8_avx512<3, 3, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8_avx512<5, 5, 1, 1, 1, 1>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8_avx512<5, 5, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8_avx512<7, 7, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-#else  // __AVX512F__
-template void convolution_im2col_input_tile_int8<1, 1, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8<3, 3, 1, 1, 1, 1>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8<3, 3, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8<5, 5, 1, 1, 1, 1>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8<5, 5, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-template void convolution_im2col_input_tile_int8<7, 7, 1, 1, 2, 2>(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk);
-#endif // __AVX512F__
-
-static void convolution_im2col_input_tile_int8(const Mat& bottom_blob, Mat& B, int j, int max_jj, int k, int max_kk, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h)
-{
-    if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
-    {
-        convolution_im2col_input_tile_conv1x1s1d1_int8(bottom_blob, B, j, max_jj, k, max_kk);
-        return;
-    }
-
-    if (kernel_w == 1 && kernel_h == 1 && stride_w == 2 && stride_h == 2)
-    {
-#if __AVX512F__
-        convolution_im2col_input_tile_int8_avx512<1, 1, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#else  // __AVX512F__
-        convolution_im2col_input_tile_int8<1, 1, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#endif // __AVX512F__
-        return;
-    }
-
-    if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
-    {
-#if __AVX512F__
-        convolution_im2col_input_tile_int8_avx512<3, 3, 1, 1, 1, 1>(bottom_blob, B, j, max_jj, k, max_kk);
-#else  // __AVX512F__
-        convolution_im2col_input_tile_int8<3, 3, 1, 1, 1, 1>(bottom_blob, B, j, max_jj, k, max_kk);
-#endif // __AVX512F__
-        return;
-    }
-
-    if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
-    {
-#if __AVX512F__
-        convolution_im2col_input_tile_int8_avx512<3, 3, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#else  // __AVX512F__
-        convolution_im2col_input_tile_int8<3, 3, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#endif // __AVX512F__
-        return;
-    }
-
-    if (kernel_w == 5 && kernel_h == 5 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
-    {
-#if __AVX512F__
-        convolution_im2col_input_tile_int8_avx512<5, 5, 1, 1, 1, 1>(bottom_blob, B, j, max_jj, k, max_kk);
-#else  // __AVX512F__
-        convolution_im2col_input_tile_int8<5, 5, 1, 1, 1, 1>(bottom_blob, B, j, max_jj, k, max_kk);
-#endif // __AVX512F__
-        return;
-    }
-
-    if (kernel_w == 5 && kernel_h == 5 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
-    {
-#if __AVX512F__
-        convolution_im2col_input_tile_int8_avx512<5, 5, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#else  // __AVX512F__
-        convolution_im2col_input_tile_int8<5, 5, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#endif // __AVX512F__
-        return;
-    }
-
-    if (kernel_w == 7 && kernel_h == 7 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
-    {
-#if __AVX512F__
-        convolution_im2col_input_tile_int8_avx512<7, 7, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#else  // __AVX512F__
-        convolution_im2col_input_tile_int8<7, 7, 1, 1, 2, 2>(bottom_blob, B, j, max_jj, k, max_kk);
-#endif // __AVX512F__
-        return;
-    }
-
-    const int w = bottom_blob.w;
-    // const int channels = bottom_blob.c;
-    const int elempack = bottom_blob.elempack;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int outw = (w - kernel_extent_w) / stride_w + 1;
-
-    // j max_jj     outw*outh    split w and h
-
-    // k max_kk     pa*maxk*(inch/pa)    split inch
-
-    // k/max_kk shall be multiple of maxk
-
-    const int maxk = kernel_w * kernel_h;
-
-    signed char* pp = B;
-
-    int jj = 0;
-#if __SSE2__
-#if defined(__x86_64__) || defined(_M_X64)
-#if __AVX512F__
-    for (; jj + 15 < max_jj; jj += 16)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dy2 = (j + jj + 2) / outw;
-        int dy3 = (j + jj + 3) / outw;
-        int dy4 = (j + jj + 4) / outw;
-        int dy5 = (j + jj + 5) / outw;
-        int dy6 = (j + jj + 6) / outw;
-        int dy7 = (j + jj + 7) / outw;
-        int dy8 = (j + jj + 8) / outw;
-        int dy9 = (j + jj + 9) / outw;
-        int dya = (j + jj + 10) / outw;
-        int dyb = (j + jj + 11) / outw;
-        int dyc = (j + jj + 12) / outw;
-        int dyd = (j + jj + 13) / outw;
-        int dye = (j + jj + 14) / outw;
-        int dyf = (j + jj + 15) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-        int dx2 = (j + jj + 2) % outw;
-        int dx3 = (j + jj + 3) % outw;
-        int dx4 = (j + jj + 4) % outw;
-        int dx5 = (j + jj + 5) % outw;
-        int dx6 = (j + jj + 6) % outw;
-        int dx7 = (j + jj + 7) % outw;
-        int dx8 = (j + jj + 8) % outw;
-        int dx9 = (j + jj + 9) % outw;
-        int dxa = (j + jj + 10) % outw;
-        int dxb = (j + jj + 11) % outw;
-        int dxc = (j + jj + 12) % outw;
-        int dxd = (j + jj + 13) % outw;
-        int dxe = (j + jj + 14) % outw;
-        int dxf = (j + jj + 15) % outw;
-
-        if (dy0 == dyf)
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-
-                    if (stride_w == 1)
-                    {
-                        __m128i _r0 = _mm_loadu_si128((const __m128i*)sptr0);
-                        __m128i _r1 = _mm_loadu_si128((const __m128i*)sptr1);
-                        __m128i _tmp0 = _mm_unpacklo_epi8(_r0, _r1);
-                        __m128i _tmp1 = _mm_unpackhi_epi8(_r0, _r1);
-                        _mm_store_si128((__m128i*)pp, _tmp0);
-                        _mm_store_si128((__m128i*)(pp + 16), _tmp1);
-                        pp += 32;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        __m256i _r0 = _mm256_loadu_si256((const __m256i*)sptr0);
-                        __m256i _r1 = _mm256_loadu_si256((const __m256i*)sptr1);
-                        __m256i _tmp0 = _mm256_unpacklo_epi8(_r0, _r1);
-                        __m256i _tmp1 = _mm256_unpackhi_epi8(_r0, _r1);
-                        _tmp0 = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(_tmp0, _MM_SHUFFLE(3, 1, 2, 0)), _MM_SHUFFLE(3, 1, 2, 0));
-                        _tmp1 = _mm256_shufflehi_epi16(_mm256_shufflelo_epi16(_tmp1, _MM_SHUFFLE(3, 1, 2, 0)), _MM_SHUFFLE(3, 1, 2, 0));
-                        _tmp0 = _mm256_shuffle_epi32(_tmp0, _MM_SHUFFLE(3, 1, 2, 0));
-                        _tmp1 = _mm256_shuffle_epi32(_tmp1, _MM_SHUFFLE(3, 1, 2, 0));
-                        __m256i _r01 = _mm256_unpacklo_epi64(_tmp0, _tmp1);
-                        _mm256_storeu_si256((__m256i*)pp, _r01);
-                        pp += 32;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr0[stride_w];
-                        pp[3] = sptr1[stride_w];
-                        pp[4] = sptr0[stride_w * 2];
-                        pp[5] = sptr1[stride_w * 2];
-                        pp[6] = sptr0[stride_w * 3];
-                        pp[7] = sptr1[stride_w * 3];
-                        pp[8] = sptr0[stride_w * 4];
-                        pp[9] = sptr1[stride_w * 4];
-                        pp[10] = sptr0[stride_w * 5];
-                        pp[11] = sptr1[stride_w * 5];
-                        pp[12] = sptr0[stride_w * 6];
-                        pp[13] = sptr1[stride_w * 6];
-                        pp[14] = sptr0[stride_w * 7];
-                        pp[15] = sptr1[stride_w * 7];
-                        pp[16 + 0] = sptr0[stride_w * 8];
-                        pp[16 + 1] = sptr1[stride_w * 8];
-                        pp[16 + 2] = sptr0[stride_w * 9];
-                        pp[16 + 3] = sptr1[stride_w * 9];
-                        pp[16 + 4] = sptr0[stride_w * 10];
-                        pp[16 + 5] = sptr1[stride_w * 10];
-                        pp[16 + 6] = sptr0[stride_w * 11];
-                        pp[16 + 7] = sptr1[stride_w * 11];
-                        pp[16 + 8] = sptr0[stride_w * 12];
-                        pp[16 + 9] = sptr1[stride_w * 12];
-                        pp[16 + 10] = sptr0[stride_w * 13];
-                        pp[16 + 11] = sptr1[stride_w * 13];
-                        pp[16 + 12] = sptr0[stride_w * 14];
-                        pp[16 + 13] = sptr1[stride_w * 14];
-                        pp[16 + 14] = sptr0[stride_w * 15];
-                        pp[16 + 15] = sptr1[stride_w * 15];
-                        pp += 32;
-                    }
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-
-                const signed char* sptr = img.row<const signed char>(y0) + x0 * elempack;
-
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 8));
-                    __m128i _r2 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 16));
-                    __m128i _r3 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 24));
-                    __m128i _r4 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 32));
-                    __m128i _r5 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 40));
-                    __m128i _r6 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 48));
-                    __m128i _r7 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 56));
-                    __m128i _r8 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 64));
-                    __m128i _r9 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 72));
-                    __m128i _ra = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 80));
-                    __m128i _rb = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 88));
-                    __m128i _rc = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 96));
-                    __m128i _rd = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 104));
-                    __m128i _re = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 112));
-                    __m128i _rf = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 120));
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    __m128i _r23 = _mm_unpacklo_epi16(_r2, _r3);
-                    __m128i _r45 = _mm_unpacklo_epi16(_r4, _r5);
-                    __m128i _r67 = _mm_unpacklo_epi16(_r6, _r7);
-                    __m128i _r89 = _mm_unpacklo_epi16(_r8, _r9);
-                    __m128i _rab = _mm_unpacklo_epi16(_ra, _rb);
-                    __m128i _rcd = _mm_unpacklo_epi16(_rc, _rd);
-                    __m128i _ref = _mm_unpacklo_epi16(_re, _rf);
-                    _r0 = _mm_unpacklo_epi32(_r01, _r23);
-                    _r1 = _mm_unpackhi_epi32(_r01, _r23);
-                    _r2 = _mm_unpacklo_epi32(_r45, _r67);
-                    _r3 = _mm_unpackhi_epi32(_r45, _r67);
-                    _r4 = _mm_unpacklo_epi32(_r89, _rab);
-                    _r5 = _mm_unpackhi_epi32(_r89, _rab);
-                    _r6 = _mm_unpacklo_epi32(_rcd, _ref);
-                    _r7 = _mm_unpackhi_epi32(_rcd, _ref);
-                    _r8 = _mm_unpacklo_epi64(_r0, _r2);
-                    _r9 = _mm_unpacklo_epi64(_r4, _r6);
-                    _ra = _mm_unpackhi_epi64(_r0, _r2);
-                    _rb = _mm_unpackhi_epi64(_r4, _r6);
-                    _rc = _mm_unpacklo_epi64(_r1, _r3);
-                    _rd = _mm_unpacklo_epi64(_r5, _r7);
-                    _re = _mm_unpackhi_epi64(_r1, _r3);
-                    _rf = _mm_unpackhi_epi64(_r5, _r7);
-                    _mm_store_si128((__m128i*)pp, _r8);
-                    _mm_store_si128((__m128i*)(pp + 16), _r9);
-                    _mm_store_si128((__m128i*)(pp + 32), _ra);
-                    _mm_store_si128((__m128i*)(pp + 48), _rb);
-                    _mm_store_si128((__m128i*)(pp + 64), _rc);
-                    _mm_store_si128((__m128i*)(pp + 80), _rd);
-                    _mm_store_si128((__m128i*)(pp + 96), _re);
-                    _mm_store_si128((__m128i*)(pp + 112), _rf);
-                    pp += 128;
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr[0];
-                    pp[1] = sptr[stride_w];
-                    pp[2] = sptr[stride_w * 2];
-                    pp[3] = sptr[stride_w * 3];
-                    pp[4] = sptr[stride_w * 4];
-                    pp[5] = sptr[stride_w * 5];
-                    pp[6] = sptr[stride_w * 6];
-                    pp[7] = sptr[stride_w * 7];
-                    pp[8] = sptr[stride_w * 8];
-                    pp[9] = sptr[stride_w * 9];
-                    pp[10] = sptr[stride_w * 10];
-                    pp[11] = sptr[stride_w * 11];
-                    pp[12] = sptr[stride_w * 12];
-                    pp[13] = sptr[stride_w * 13];
-                    pp[14] = sptr[stride_w * 14];
-                    pp[15] = sptr[stride_w * 15];
-                    pp += 16;
-                }
-            }
-        }
-        else
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int x01 = stride_w * dx1 + dilation_w * v0;
-                    int x02 = stride_w * dx2 + dilation_w * v0;
-                    int x03 = stride_w * dx3 + dilation_w * v0;
-                    int x04 = stride_w * dx4 + dilation_w * v0;
-                    int x05 = stride_w * dx5 + dilation_w * v0;
-                    int x06 = stride_w * dx6 + dilation_w * v0;
-                    int x07 = stride_w * dx7 + dilation_w * v0;
-                    int x08 = stride_w * dx8 + dilation_w * v0;
-                    int x09 = stride_w * dx9 + dilation_w * v0;
-                    int x0a = stride_w * dxa + dilation_w * v0;
-                    int x0b = stride_w * dxb + dilation_w * v0;
-                    int x0c = stride_w * dxc + dilation_w * v0;
-                    int x0d = stride_w * dxd + dilation_w * v0;
-                    int x0e = stride_w * dxe + dilation_w * v0;
-                    int x0f = stride_w * dxf + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int y01 = stride_h * dy1 + dilation_h * u0;
-                    int y02 = stride_h * dy2 + dilation_h * u0;
-                    int y03 = stride_h * dy3 + dilation_h * u0;
-                    int y04 = stride_h * dy4 + dilation_h * u0;
-                    int y05 = stride_h * dy5 + dilation_h * u0;
-                    int y06 = stride_h * dy6 + dilation_h * u0;
-                    int y07 = stride_h * dy7 + dilation_h * u0;
-                    int y08 = stride_h * dy8 + dilation_h * u0;
-                    int y09 = stride_h * dy9 + dilation_h * u0;
-                    int y0a = stride_h * dya + dilation_h * u0;
-                    int y0b = stride_h * dyb + dilation_h * u0;
-                    int y0c = stride_h * dyc + dilation_h * u0;
-                    int y0d = stride_h * dyd + dilation_h * u0;
-                    int y0e = stride_h * dye + dilation_h * u0;
-                    int y0f = stride_h * dyf + dilation_h * u0;
-
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int x11 = stride_w * dx1 + dilation_w * v1;
-                    int x12 = stride_w * dx2 + dilation_w * v1;
-                    int x13 = stride_w * dx3 + dilation_w * v1;
-                    int x14 = stride_w * dx4 + dilation_w * v1;
-                    int x15 = stride_w * dx5 + dilation_w * v1;
-                    int x16 = stride_w * dx6 + dilation_w * v1;
-                    int x17 = stride_w * dx7 + dilation_w * v1;
-                    int x18 = stride_w * dx8 + dilation_w * v1;
-                    int x19 = stride_w * dx9 + dilation_w * v1;
-                    int x1a = stride_w * dxa + dilation_w * v1;
-                    int x1b = stride_w * dxb + dilation_w * v1;
-                    int x1c = stride_w * dxc + dilation_w * v1;
-                    int x1d = stride_w * dxd + dilation_w * v1;
-                    int x1e = stride_w * dxe + dilation_w * v1;
-                    int x1f = stride_w * dxf + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-                    int y11 = stride_h * dy1 + dilation_h * u1;
-                    int y12 = stride_h * dy2 + dilation_h * u1;
-                    int y13 = stride_h * dy3 + dilation_h * u1;
-                    int y14 = stride_h * dy4 + dilation_h * u1;
-                    int y15 = stride_h * dy5 + dilation_h * u1;
-                    int y16 = stride_h * dy6 + dilation_h * u1;
-                    int y17 = stride_h * dy7 + dilation_h * u1;
-                    int y18 = stride_h * dy8 + dilation_h * u1;
-                    int y19 = stride_h * dy9 + dilation_h * u1;
-                    int y1a = stride_h * dya + dilation_h * u1;
-                    int y1b = stride_h * dyb + dilation_h * u1;
-                    int y1c = stride_h * dyc + dilation_h * u1;
-                    int y1d = stride_h * dyd + dilation_h * u1;
-                    int y1e = stride_h * dye + dilation_h * u1;
-                    int y1f = stride_h * dyf + dilation_h * u1;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-                    const signed char* sptr04 = img0.row<const signed char>(y04) + x04;
-                    const signed char* sptr05 = img0.row<const signed char>(y05) + x05;
-                    const signed char* sptr06 = img0.row<const signed char>(y06) + x06;
-                    const signed char* sptr07 = img0.row<const signed char>(y07) + x07;
-                    const signed char* sptr08 = img0.row<const signed char>(y08) + x08;
-                    const signed char* sptr09 = img0.row<const signed char>(y09) + x09;
-                    const signed char* sptr0a = img0.row<const signed char>(y0a) + x0a;
-                    const signed char* sptr0b = img0.row<const signed char>(y0b) + x0b;
-                    const signed char* sptr0c = img0.row<const signed char>(y0c) + x0c;
-                    const signed char* sptr0d = img0.row<const signed char>(y0d) + x0d;
-                    const signed char* sptr0e = img0.row<const signed char>(y0e) + x0e;
-                    const signed char* sptr0f = img0.row<const signed char>(y0f) + x0f;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-                    const signed char* sptr14 = img1.row<const signed char>(y14) + x14;
-                    const signed char* sptr15 = img1.row<const signed char>(y15) + x15;
-                    const signed char* sptr16 = img1.row<const signed char>(y16) + x16;
-                    const signed char* sptr17 = img1.row<const signed char>(y17) + x17;
-                    const signed char* sptr18 = img1.row<const signed char>(y18) + x18;
-                    const signed char* sptr19 = img1.row<const signed char>(y19) + x19;
-                    const signed char* sptr1a = img1.row<const signed char>(y1a) + x1a;
-                    const signed char* sptr1b = img1.row<const signed char>(y1b) + x1b;
-                    const signed char* sptr1c = img1.row<const signed char>(y1c) + x1c;
-                    const signed char* sptr1d = img1.row<const signed char>(y1d) + x1d;
-                    const signed char* sptr1e = img1.row<const signed char>(y1e) + x1e;
-                    const signed char* sptr1f = img1.row<const signed char>(y1f) + x1f;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr01[0];
-                    pp[3] = sptr11[0];
-                    pp[4] = sptr02[0];
-                    pp[5] = sptr12[0];
-                    pp[6] = sptr03[0];
-                    pp[7] = sptr13[0];
-                    pp[8] = sptr04[0];
-                    pp[9] = sptr14[0];
-                    pp[10] = sptr05[0];
-                    pp[11] = sptr15[0];
-                    pp[12] = sptr06[0];
-                    pp[13] = sptr16[0];
-                    pp[14] = sptr07[0];
-                    pp[15] = sptr17[0];
-                    pp[16 + 0] = sptr08[0];
-                    pp[16 + 1] = sptr18[0];
-                    pp[16 + 2] = sptr09[0];
-                    pp[16 + 3] = sptr19[0];
-                    pp[16 + 4] = sptr0a[0];
-                    pp[16 + 5] = sptr1a[0];
-                    pp[16 + 6] = sptr0b[0];
-                    pp[16 + 7] = sptr1b[0];
-                    pp[16 + 8] = sptr0c[0];
-                    pp[16 + 9] = sptr1c[0];
-                    pp[16 + 10] = sptr0d[0];
-                    pp[16 + 11] = sptr1d[0];
-                    pp[16 + 12] = sptr0e[0];
-                    pp[16 + 13] = sptr1e[0];
-                    pp[16 + 14] = sptr0f[0];
-                    pp[16 + 15] = sptr1f[0];
-                    pp += 32;
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int x1 = stride_w * dx1 + dilation_w * v;
-                int x2 = stride_w * dx2 + dilation_w * v;
-                int x3 = stride_w * dx3 + dilation_w * v;
-                int x4 = stride_w * dx4 + dilation_w * v;
-                int x5 = stride_w * dx5 + dilation_w * v;
-                int x6 = stride_w * dx6 + dilation_w * v;
-                int x7 = stride_w * dx7 + dilation_w * v;
-                int x8 = stride_w * dx8 + dilation_w * v;
-                int x9 = stride_w * dx9 + dilation_w * v;
-                int xa = stride_w * dxa + dilation_w * v;
-                int xb = stride_w * dxb + dilation_w * v;
-                int xc = stride_w * dxc + dilation_w * v;
-                int xd = stride_w * dxd + dilation_w * v;
-                int xe = stride_w * dxe + dilation_w * v;
-                int xf = stride_w * dxf + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-                int y1 = stride_h * dy1 + dilation_h * u;
-                int y2 = stride_h * dy2 + dilation_h * u;
-                int y3 = stride_h * dy3 + dilation_h * u;
-                int y4 = stride_h * dy4 + dilation_h * u;
-                int y5 = stride_h * dy5 + dilation_h * u;
-                int y6 = stride_h * dy6 + dilation_h * u;
-                int y7 = stride_h * dy7 + dilation_h * u;
-                int y8 = stride_h * dy8 + dilation_h * u;
-                int y9 = stride_h * dy9 + dilation_h * u;
-                int ya = stride_h * dya + dilation_h * u;
-                int yb = stride_h * dyb + dilation_h * u;
-                int yc = stride_h * dyc + dilation_h * u;
-                int yd = stride_h * dyd + dilation_h * u;
-                int ye = stride_h * dye + dilation_h * u;
-                int yf = stride_h * dyf + dilation_h * u;
-
-                const signed char* sptr0 = img.row<const signed char>(y0) + x0 * elempack;
-                const signed char* sptr1 = img.row<const signed char>(y1) + x1 * elempack;
-                const signed char* sptr2 = img.row<const signed char>(y2) + x2 * elempack;
-                const signed char* sptr3 = img.row<const signed char>(y3) + x3 * elempack;
-                const signed char* sptr4 = img.row<const signed char>(y4) + x4 * elempack;
-                const signed char* sptr5 = img.row<const signed char>(y5) + x5 * elempack;
-                const signed char* sptr6 = img.row<const signed char>(y6) + x6 * elempack;
-                const signed char* sptr7 = img.row<const signed char>(y7) + x7 * elempack;
-                const signed char* sptr8 = img.row<const signed char>(y8) + x8 * elempack;
-                const signed char* sptr9 = img.row<const signed char>(y9) + x9 * elempack;
-                const signed char* sptra = img.row<const signed char>(ya) + xa * elempack;
-                const signed char* sptrb = img.row<const signed char>(yb) + xb * elempack;
-                const signed char* sptrc = img.row<const signed char>(yc) + xc * elempack;
-                const signed char* sptrd = img.row<const signed char>(yd) + xd * elempack;
-                const signed char* sptre = img.row<const signed char>(ye) + xe * elempack;
-                const signed char* sptrf = img.row<const signed char>(yf) + xf * elempack;
-
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr0);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)sptr1);
-                    __m128i _r2 = _mm_loadl_epi64((const __m128i*)sptr2);
-                    __m128i _r3 = _mm_loadl_epi64((const __m128i*)sptr3);
-                    __m128i _r4 = _mm_loadl_epi64((const __m128i*)sptr4);
-                    __m128i _r5 = _mm_loadl_epi64((const __m128i*)sptr5);
-                    __m128i _r6 = _mm_loadl_epi64((const __m128i*)sptr6);
-                    __m128i _r7 = _mm_loadl_epi64((const __m128i*)sptr7);
-                    __m128i _r8 = _mm_loadl_epi64((const __m128i*)sptr8);
-                    __m128i _r9 = _mm_loadl_epi64((const __m128i*)sptr9);
-                    __m128i _ra = _mm_loadl_epi64((const __m128i*)sptra);
-                    __m128i _rb = _mm_loadl_epi64((const __m128i*)sptrb);
-                    __m128i _rc = _mm_loadl_epi64((const __m128i*)sptrc);
-                    __m128i _rd = _mm_loadl_epi64((const __m128i*)sptrd);
-                    __m128i _re = _mm_loadl_epi64((const __m128i*)sptre);
-                    __m128i _rf = _mm_loadl_epi64((const __m128i*)sptrf);
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    __m128i _r23 = _mm_unpacklo_epi16(_r2, _r3);
-                    __m128i _r45 = _mm_unpacklo_epi16(_r4, _r5);
-                    __m128i _r67 = _mm_unpacklo_epi16(_r6, _r7);
-                    __m128i _r89 = _mm_unpacklo_epi16(_r8, _r9);
-                    __m128i _rab = _mm_unpacklo_epi16(_ra, _rb);
-                    __m128i _rcd = _mm_unpacklo_epi16(_rc, _rd);
-                    __m128i _ref = _mm_unpacklo_epi16(_re, _rf);
-                    _r0 = _mm_unpacklo_epi32(_r01, _r23);
-                    _r1 = _mm_unpackhi_epi32(_r01, _r23);
-                    _r2 = _mm_unpacklo_epi32(_r45, _r67);
-                    _r3 = _mm_unpackhi_epi32(_r45, _r67);
-                    _r4 = _mm_unpacklo_epi32(_r89, _rab);
-                    _r5 = _mm_unpackhi_epi32(_r89, _rab);
-                    _r6 = _mm_unpacklo_epi32(_rcd, _ref);
-                    _r7 = _mm_unpackhi_epi32(_rcd, _ref);
-                    _r8 = _mm_unpacklo_epi64(_r0, _r2);
-                    _r9 = _mm_unpacklo_epi64(_r4, _r6);
-                    _ra = _mm_unpackhi_epi64(_r0, _r2);
-                    _rb = _mm_unpackhi_epi64(_r4, _r6);
-                    _rc = _mm_unpacklo_epi64(_r1, _r3);
-                    _rd = _mm_unpacklo_epi64(_r5, _r7);
-                    _re = _mm_unpackhi_epi64(_r1, _r3);
-                    _rf = _mm_unpackhi_epi64(_r5, _r7);
-                    _mm_store_si128((__m128i*)pp, _r8);
-                    _mm_store_si128((__m128i*)(pp + 16), _r9);
-                    _mm_store_si128((__m128i*)(pp + 32), _ra);
-                    _mm_store_si128((__m128i*)(pp + 48), _rb);
-                    _mm_store_si128((__m128i*)(pp + 64), _rc);
-                    _mm_store_si128((__m128i*)(pp + 80), _rd);
-                    _mm_store_si128((__m128i*)(pp + 96), _re);
-                    _mm_store_si128((__m128i*)(pp + 112), _rf);
-                    pp += 128;
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr2[0];
-                    pp[3] = sptr3[0];
-                    pp[4] = sptr4[0];
-                    pp[5] = sptr5[0];
-                    pp[6] = sptr6[0];
-                    pp[7] = sptr7[0];
-                    pp[8] = sptr8[0];
-                    pp[9] = sptr9[0];
-                    pp[10] = sptra[0];
-                    pp[11] = sptrb[0];
-                    pp[12] = sptrc[0];
-                    pp[13] = sptrd[0];
-                    pp[14] = sptre[0];
-                    pp[15] = sptrf[0];
-                    pp += 16;
-                }
-            }
-        }
-    }
-#endif // __AVX512F__
-    for (; jj + 7 < max_jj; jj += 8)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dy2 = (j + jj + 2) / outw;
-        int dy3 = (j + jj + 3) / outw;
-        int dy4 = (j + jj + 4) / outw;
-        int dy5 = (j + jj + 5) / outw;
-        int dy6 = (j + jj + 6) / outw;
-        int dy7 = (j + jj + 7) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-        int dx2 = (j + jj + 2) % outw;
-        int dx3 = (j + jj + 3) % outw;
-        int dx4 = (j + jj + 4) % outw;
-        int dx5 = (j + jj + 5) % outw;
-        int dx6 = (j + jj + 6) % outw;
-        int dx7 = (j + jj + 7) % outw;
-
-        if (dy0 == dy7)
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-
-                    if (stride_w == 1)
-                    {
-                        __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr0);
-                        __m128i _r1 = _mm_loadl_epi64((const __m128i*)sptr1);
-                        __m128i _r01 = _mm_unpacklo_epi8(_r0, _r1);
-                        _mm_storeu_si128((__m128i*)pp, _r01);
-                        pp += 16;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        __m128i _r0 = _mm_loadu_si128((const __m128i*)sptr0);
-                        __m128i _r1 = _mm_loadu_si128((const __m128i*)sptr1);
-                        __m128i _tmp0 = _mm_unpacklo_epi8(_r0, _r1);
-                        __m128i _tmp1 = _mm_unpackhi_epi8(_r0, _r1);
-                        _tmp0 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_tmp0, _MM_SHUFFLE(3, 1, 2, 0)), _MM_SHUFFLE(3, 1, 2, 0));
-                        _tmp1 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_tmp1, _MM_SHUFFLE(3, 1, 2, 0)), _MM_SHUFFLE(3, 1, 2, 0));
-                        _tmp0 = _mm_shuffle_epi32(_tmp0, _MM_SHUFFLE(3, 1, 2, 0));
-                        _tmp1 = _mm_shuffle_epi32(_tmp1, _MM_SHUFFLE(3, 1, 2, 0));
-                        __m128i _r01 = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(_tmp0), _mm_castsi128_ps(_tmp1), _MM_SHUFFLE(1, 0, 1, 0)));
-                        _mm_storeu_si128((__m128i*)pp, _r01);
-                        pp += 16;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr0[stride_w];
-                        pp[3] = sptr1[stride_w];
-                        pp[4] = sptr0[stride_w * 2];
-                        pp[5] = sptr1[stride_w * 2];
-                        pp[6] = sptr0[stride_w * 3];
-                        pp[7] = sptr1[stride_w * 3];
-                        pp[8] = sptr0[stride_w * 4];
-                        pp[9] = sptr1[stride_w * 4];
-                        pp[10] = sptr0[stride_w * 5];
-                        pp[11] = sptr1[stride_w * 5];
-                        pp[12] = sptr0[stride_w * 6];
-                        pp[13] = sptr1[stride_w * 6];
-                        pp[14] = sptr0[stride_w * 7];
-                        pp[15] = sptr1[stride_w * 7];
-                        pp += 16;
-                    }
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-
-                const signed char* sptr = img.row<const signed char>(y0) + x0 * elempack;
-
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 8));
-                    __m128i _r2 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 16));
-                    __m128i _r3 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 24));
-                    __m128i _r4 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 32));
-                    __m128i _r5 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 40));
-                    __m128i _r6 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 48));
-                    __m128i _r7 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 56));
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    __m128i _r23 = _mm_unpacklo_epi16(_r2, _r3);
-                    __m128i _r45 = _mm_unpacklo_epi16(_r4, _r5);
-                    __m128i _r67 = _mm_unpacklo_epi16(_r6, _r7);
-                    _r0 = _mm_unpacklo_epi32(_r01, _r23);
-                    _r1 = _mm_unpackhi_epi32(_r01, _r23);
-                    _r2 = _mm_unpacklo_epi32(_r45, _r67);
-                    _r3 = _mm_unpackhi_epi32(_r45, _r67);
-                    _r4 = _mm_unpacklo_epi64(_r0, _r2);
-                    _r5 = _mm_unpackhi_epi64(_r0, _r2);
-                    _r6 = _mm_unpacklo_epi64(_r1, _r3);
-                    _r7 = _mm_unpackhi_epi64(_r1, _r3);
-                    _mm_storeu_si128((__m128i*)pp, _r4);
-                    _mm_storeu_si128((__m128i*)(pp + 16), _r5);
-                    _mm_storeu_si128((__m128i*)(pp + 32), _r6);
-                    _mm_storeu_si128((__m128i*)(pp + 48), _r7);
-                    pp += 64;
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr[0];
-                    pp[1] = sptr[stride_w];
-                    pp[2] = sptr[stride_w * 2];
-                    pp[3] = sptr[stride_w * 3];
-                    pp[4] = sptr[stride_w * 4];
-                    pp[5] = sptr[stride_w * 5];
-                    pp[6] = sptr[stride_w * 6];
-                    pp[7] = sptr[stride_w * 7];
-                    pp += 8;
-                }
-            }
-        }
-        else
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int x01 = stride_w * dx1 + dilation_w * v0;
-                    int x02 = stride_w * dx2 + dilation_w * v0;
-                    int x03 = stride_w * dx3 + dilation_w * v0;
-                    int x04 = stride_w * dx4 + dilation_w * v0;
-                    int x05 = stride_w * dx5 + dilation_w * v0;
-                    int x06 = stride_w * dx6 + dilation_w * v0;
-                    int x07 = stride_w * dx7 + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int y01 = stride_h * dy1 + dilation_h * u0;
-                    int y02 = stride_h * dy2 + dilation_h * u0;
-                    int y03 = stride_h * dy3 + dilation_h * u0;
-                    int y04 = stride_h * dy4 + dilation_h * u0;
-                    int y05 = stride_h * dy5 + dilation_h * u0;
-                    int y06 = stride_h * dy6 + dilation_h * u0;
-                    int y07 = stride_h * dy7 + dilation_h * u0;
-
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int x11 = stride_w * dx1 + dilation_w * v1;
-                    int x12 = stride_w * dx2 + dilation_w * v1;
-                    int x13 = stride_w * dx3 + dilation_w * v1;
-                    int x14 = stride_w * dx4 + dilation_w * v1;
-                    int x15 = stride_w * dx5 + dilation_w * v1;
-                    int x16 = stride_w * dx6 + dilation_w * v1;
-                    int x17 = stride_w * dx7 + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-                    int y11 = stride_h * dy1 + dilation_h * u1;
-                    int y12 = stride_h * dy2 + dilation_h * u1;
-                    int y13 = stride_h * dy3 + dilation_h * u1;
-                    int y14 = stride_h * dy4 + dilation_h * u1;
-                    int y15 = stride_h * dy5 + dilation_h * u1;
-                    int y16 = stride_h * dy6 + dilation_h * u1;
-                    int y17 = stride_h * dy7 + dilation_h * u1;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-                    const signed char* sptr04 = img0.row<const signed char>(y04) + x04;
-                    const signed char* sptr05 = img0.row<const signed char>(y05) + x05;
-                    const signed char* sptr06 = img0.row<const signed char>(y06) + x06;
-                    const signed char* sptr07 = img0.row<const signed char>(y07) + x07;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-                    const signed char* sptr14 = img1.row<const signed char>(y14) + x14;
-                    const signed char* sptr15 = img1.row<const signed char>(y15) + x15;
-                    const signed char* sptr16 = img1.row<const signed char>(y16) + x16;
-                    const signed char* sptr17 = img1.row<const signed char>(y17) + x17;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr01[0];
-                    pp[3] = sptr11[0];
-                    pp[4] = sptr02[0];
-                    pp[5] = sptr12[0];
-                    pp[6] = sptr03[0];
-                    pp[7] = sptr13[0];
-                    pp[8] = sptr04[0];
-                    pp[9] = sptr14[0];
-                    pp[10] = sptr05[0];
-                    pp[11] = sptr15[0];
-                    pp[12] = sptr06[0];
-                    pp[13] = sptr16[0];
-                    pp[14] = sptr07[0];
-                    pp[15] = sptr17[0];
-                    pp += 16;
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int x1 = stride_w * dx1 + dilation_w * v;
-                int x2 = stride_w * dx2 + dilation_w * v;
-                int x3 = stride_w * dx3 + dilation_w * v;
-                int x4 = stride_w * dx4 + dilation_w * v;
-                int x5 = stride_w * dx5 + dilation_w * v;
-                int x6 = stride_w * dx6 + dilation_w * v;
-                int x7 = stride_w * dx7 + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-                int y1 = stride_h * dy1 + dilation_h * u;
-                int y2 = stride_h * dy2 + dilation_h * u;
-                int y3 = stride_h * dy3 + dilation_h * u;
-                int y4 = stride_h * dy4 + dilation_h * u;
-                int y5 = stride_h * dy5 + dilation_h * u;
-                int y6 = stride_h * dy6 + dilation_h * u;
-                int y7 = stride_h * dy7 + dilation_h * u;
-
-                const signed char* sptr0 = img.row<const signed char>(y0) + x0 * elempack;
-                const signed char* sptr1 = img.row<const signed char>(y1) + x1 * elempack;
-                const signed char* sptr2 = img.row<const signed char>(y2) + x2 * elempack;
-                const signed char* sptr3 = img.row<const signed char>(y3) + x3 * elempack;
-                const signed char* sptr4 = img.row<const signed char>(y4) + x4 * elempack;
-                const signed char* sptr5 = img.row<const signed char>(y5) + x5 * elempack;
-                const signed char* sptr6 = img.row<const signed char>(y6) + x6 * elempack;
-                const signed char* sptr7 = img.row<const signed char>(y7) + x7 * elempack;
-
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr0);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)sptr1);
-                    __m128i _r2 = _mm_loadl_epi64((const __m128i*)sptr2);
-                    __m128i _r3 = _mm_loadl_epi64((const __m128i*)sptr3);
-                    __m128i _r4 = _mm_loadl_epi64((const __m128i*)sptr4);
-                    __m128i _r5 = _mm_loadl_epi64((const __m128i*)sptr5);
-                    __m128i _r6 = _mm_loadl_epi64((const __m128i*)sptr6);
-                    __m128i _r7 = _mm_loadl_epi64((const __m128i*)sptr7);
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    __m128i _r23 = _mm_unpacklo_epi16(_r2, _r3);
-                    __m128i _r45 = _mm_unpacklo_epi16(_r4, _r5);
-                    __m128i _r67 = _mm_unpacklo_epi16(_r6, _r7);
-                    _r0 = _mm_unpacklo_epi32(_r01, _r23);
-                    _r1 = _mm_unpackhi_epi32(_r01, _r23);
-                    _r2 = _mm_unpacklo_epi32(_r45, _r67);
-                    _r3 = _mm_unpackhi_epi32(_r45, _r67);
-                    _r4 = _mm_unpacklo_epi64(_r0, _r2);
-                    _r5 = _mm_unpackhi_epi64(_r0, _r2);
-                    _r6 = _mm_unpacklo_epi64(_r1, _r3);
-                    _r7 = _mm_unpackhi_epi64(_r1, _r3);
-                    _mm_storeu_si128((__m128i*)pp, _r4);
-                    _mm_storeu_si128((__m128i*)(pp + 16), _r5);
-                    _mm_storeu_si128((__m128i*)(pp + 32), _r6);
-                    _mm_storeu_si128((__m128i*)(pp + 48), _r7);
-                    pp += 64;
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr2[0];
-                    pp[3] = sptr3[0];
-                    pp[4] = sptr4[0];
-                    pp[5] = sptr5[0];
-                    pp[6] = sptr6[0];
-                    pp[7] = sptr7[0];
-                    pp += 8;
-                }
-            }
-        }
-    }
-#endif // defined(__x86_64__) || defined(_M_X64)
-    for (; jj + 3 < max_jj; jj += 4)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dy2 = (j + jj + 2) / outw;
-        int dy3 = (j + jj + 3) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-        int dx2 = (j + jj + 2) % outw;
-        int dx3 = (j + jj + 3) % outw;
-
-        if (dy0 == dy3)
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-
-                    if (stride_w == 1)
-                    {
-                        __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr0);
-                        __m128i _r1 = _mm_loadl_epi64((const __m128i*)sptr1);
-                        __m128i _r01 = _mm_unpacklo_epi8(_r0, _r1);
-                        _mm_storel_epi64((__m128i*)pp, _r01);
-                        pp += 8;
-                    }
-                    else if (stride_w == 2)
-                    {
-                        __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr0);
-                        __m128i _r1 = _mm_loadl_epi64((const __m128i*)sptr1);
-                        __m128i _r01 = _mm_unpacklo_epi8(_r0, _r1);
-                        _r01 = _mm_shufflehi_epi16(_mm_shufflelo_epi16(_r01, _MM_SHUFFLE(3, 1, 2, 0)), _MM_SHUFFLE(3, 1, 2, 0));
-                        _r01 = _mm_shuffle_epi32(_r01, _MM_SHUFFLE(3, 1, 2, 0));
-                        _mm_storel_epi64((__m128i*)pp, _r01);
-                        pp += 8;
-                    }
-                    else
-                    {
-                        pp[0] = sptr0[0];
-                        pp[1] = sptr1[0];
-                        pp[2] = sptr0[stride_w];
-                        pp[3] = sptr1[stride_w];
-                        pp[4] = sptr0[stride_w * 2];
-                        pp[5] = sptr1[stride_w * 2];
-                        pp[6] = sptr0[stride_w * 3];
-                        pp[7] = sptr1[stride_w * 3];
-                        pp += 8;
-                    }
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-
-                const signed char* sptr = img.row<const signed char>(y0) + x0 * elempack;
-
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 8));
-                    __m128i _r2 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 16));
-                    __m128i _r3 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 24));
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    __m128i _r23 = _mm_unpacklo_epi16(_r2, _r3);
-                    _r0 = _mm_unpacklo_epi32(_r01, _r23);
-                    _r1 = _mm_unpackhi_epi32(_r01, _r23);
-                    _mm_storeu_si128((__m128i*)pp, _r0);
-                    _mm_storeu_si128((__m128i*)(pp + 16), _r1);
-                    pp += 32;
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr[0];
-                    pp[1] = sptr[stride_w];
-                    pp[2] = sptr[stride_w * 2];
-                    pp[3] = sptr[stride_w * 3];
-                    pp += 4;
-                }
-            }
-        }
-        else
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int x01 = stride_w * dx1 + dilation_w * v0;
-                    int x02 = stride_w * dx2 + dilation_w * v0;
-                    int x03 = stride_w * dx3 + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int y01 = stride_h * dy1 + dilation_h * u0;
-                    int y02 = stride_h * dy2 + dilation_h * u0;
-                    int y03 = stride_h * dy3 + dilation_h * u0;
-
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int x11 = stride_w * dx1 + dilation_w * v1;
-                    int x12 = stride_w * dx2 + dilation_w * v1;
-                    int x13 = stride_w * dx3 + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-                    int y11 = stride_h * dy1 + dilation_h * u1;
-                    int y12 = stride_h * dy2 + dilation_h * u1;
-                    int y13 = stride_h * dy3 + dilation_h * u1;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr02 = img0.row<const signed char>(y02) + x02;
-                    const signed char* sptr03 = img0.row<const signed char>(y03) + x03;
-
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-                    const signed char* sptr12 = img1.row<const signed char>(y12) + x12;
-                    const signed char* sptr13 = img1.row<const signed char>(y13) + x13;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr01[0];
-                    pp[3] = sptr11[0];
-                    pp[4] = sptr02[0];
-                    pp[5] = sptr12[0];
-                    pp[6] = sptr03[0];
-                    pp[7] = sptr13[0];
-                    pp += 8;
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int x1 = stride_w * dx1 + dilation_w * v;
-                int x2 = stride_w * dx2 + dilation_w * v;
-                int x3 = stride_w * dx3 + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-                int y1 = stride_h * dy1 + dilation_h * u;
-                int y2 = stride_h * dy2 + dilation_h * u;
-                int y3 = stride_h * dy3 + dilation_h * u;
-
-                const signed char* sptr0 = img.row<const signed char>(y0) + x0 * elempack;
-                const signed char* sptr1 = img.row<const signed char>(y1) + x1 * elempack;
-                const signed char* sptr2 = img.row<const signed char>(y2) + x2 * elempack;
-                const signed char* sptr3 = img.row<const signed char>(y3) + x3 * elempack;
-
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr0);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)sptr1);
-                    __m128i _r2 = _mm_loadl_epi64((const __m128i*)sptr2);
-                    __m128i _r3 = _mm_loadl_epi64((const __m128i*)sptr3);
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    __m128i _r23 = _mm_unpacklo_epi16(_r2, _r3);
-                    _r0 = _mm_unpacklo_epi32(_r01, _r23);
-                    _r1 = _mm_unpackhi_epi32(_r01, _r23);
-                    _mm_storeu_si128((__m128i*)pp, _r0);
-                    _mm_storeu_si128((__m128i*)(pp + 16), _r1);
-                    pp += 32;
-                }
-                if (elempack == 1)
-                {
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr2[0];
-                    pp[3] = sptr3[0];
-                    pp += 4;
-                }
-            }
-        }
-    }
-#endif // __SSE2__
-    for (; jj + 1 < max_jj; jj += 2)
-    {
-        int dy0 = (j + jj) / outw;
-        int dy1 = (j + jj + 1) / outw;
-        int dx0 = (j + jj) % outw;
-        int dx1 = (j + jj + 1) % outw;
-
-        if (dy0 == dy1)
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-
-                    const signed char* sptr0 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr1 = img1.row<const signed char>(y10) + x10;
-
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp[2] = sptr0[stride_w];
-                    pp[3] = sptr1[stride_w];
-                    pp += 4;
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-
-                const signed char* sptr = img.row<const signed char>(y0) + x0 * elempack;
-
-#if __SSE2__
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)(sptr + stride_w * 8));
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    _mm_storeu_si128((__m128i*)pp, _r01);
-                    pp += 16;
-                }
-#endif // __SSE2__
-                if (elempack == 1)
-                {
-                    pp[0] = sptr[0];
-                    pp[1] = sptr[stride_w];
-                    pp += 2;
-                }
-            }
-        }
-        else
-        {
-            int kk = 0;
-            if (elempack == 1)
-            {
-                for (; kk + 1 < max_kk; kk += 2)
-                {
-                    int p0 = (k + kk) / maxk;
-                    int p1 = (k + kk + 1) / maxk;
-                    int uv0 = (k + kk) % maxk;
-                    int uv1 = (k + kk + 1) % maxk;
-                    int u0 = uv0 / kernel_w;
-                    int u1 = uv1 / kernel_w;
-                    int v0 = uv0 % kernel_w;
-                    int v1 = uv1 % kernel_w;
-
-                    const Mat img0 = bottom_blob.channel(p0);
-                    const Mat img1 = bottom_blob.channel(p1);
-
-                    int x00 = stride_w * dx0 + dilation_w * v0;
-                    int x01 = stride_w * dx1 + dilation_w * v0;
-                    int y00 = stride_h * dy0 + dilation_h * u0;
-                    int y01 = stride_h * dy1 + dilation_h * u0;
-                    int x10 = stride_w * dx0 + dilation_w * v1;
-                    int x11 = stride_w * dx1 + dilation_w * v1;
-                    int y10 = stride_h * dy0 + dilation_h * u1;
-                    int y11 = stride_h * dy1 + dilation_h * u1;
-
-                    const signed char* sptr00 = img0.row<const signed char>(y00) + x00;
-                    const signed char* sptr01 = img0.row<const signed char>(y01) + x01;
-                    const signed char* sptr10 = img1.row<const signed char>(y10) + x10;
-                    const signed char* sptr11 = img1.row<const signed char>(y11) + x11;
-
-                    pp[0] = sptr00[0];
-                    pp[1] = sptr10[0];
-                    pp[2] = sptr01[0];
-                    pp[3] = sptr11[0];
-                    pp += 4;
-                }
-            }
-            for (; kk < max_kk / elempack; kk++)
-            {
-                int p = (k / elempack + kk) / maxk;
-                int uv = (k / elempack + kk) % maxk;
-                int u = uv / kernel_w;
-                int v = uv % kernel_w;
-
-                const Mat img = bottom_blob.channel(p);
-
-                int x0 = stride_w * dx0 + dilation_w * v;
-                int x1 = stride_w * dx1 + dilation_w * v;
-                int y0 = stride_h * dy0 + dilation_h * u;
-                int y1 = stride_h * dy1 + dilation_h * u;
-
-                const signed char* sptr0 = img.row<const signed char>(y0) + x0 * elempack;
-                const signed char* sptr1 = img.row<const signed char>(y1) + x1 * elempack;
-
-#if __SSE2__
-                if (elempack == 8)
-                {
-                    __m128i _r0 = _mm_loadl_epi64((const __m128i*)sptr0);
-                    __m128i _r1 = _mm_loadl_epi64((const __m128i*)sptr1);
-                    __m128i _r01 = _mm_unpacklo_epi16(_r0, _r1);
-                    _mm_storeu_si128((__m128i*)pp, _r01);
-                    pp += 16;
-                }
-#endif // __SSE2__
-                if (elempack == 1)
-                {
-                    pp[0] = sptr0[0];
-                    pp[1] = sptr1[0];
-                    pp += 2;
-                }
-            }
-        }
-    }
-    for (; jj < max_jj; jj++)
-    {
-        int dy = (j + jj) / outw;
-        int dx = (j + jj) % outw;
-
-        int kk = 0;
-        for (; kk < max_kk / elempack; kk++)
-        {
-            int p = (k / elempack + kk) / maxk;
-            int uv = (k / elempack + kk) % maxk;
-            int u = uv / kernel_w;
-            int v = uv % kernel_w;
-
-            const Mat img = bottom_blob.channel(p);
-
-            int x = stride_w * dx + dilation_w * v;
-            int y = stride_h * dy + dilation_h * u;
-
-            const signed char* sptr = img.row<const signed char>(y) + x * elempack;
-
-#if __SSE2__
-            if (elempack == 8)
-            {
-                _mm_storel_epi64((__m128i*)pp, _mm_loadl_epi64((const __m128i*)sptr));
-                pp += 8;
-            }
-#endif // __SSE2__
-            if (elempack == 1)
-            {
-                pp[0] = sptr[0];
-                pp += 1;
-            }
-        }
-    }
-}
-
-static void convolution_im2col_gemm_transform_kernel_int8(const Mat& kernel, Mat& AT, int inch, int outch, int kernel_w, int kernel_h, const Option& opt)
-{
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-    if (ncnn::cpu_support_x86_avx2())
-    {
-        convolution_im2col_gemm_transform_kernel_int8_avx2(kernel, AT, inch, outch, kernel_w, kernel_h, opt);
-        return;
-    }
-#endif
-#endif
-
-    // NCNN_LOGE("convolution_im2col_gemm_transform_kernel");
-    const int maxk = kernel_w * kernel_h;
-
-    const int M = outch;
-    const int K = inch * maxk;
-
-    int TILE_M, TILE_N, TILE_K;
-    convolution_im2col_gemm_get_optimal_tile_mnk_int8(M, 0, K, TILE_M, TILE_N, TILE_K, opt.num_threads);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-
-    int elempack = 1;
-#if __SSE2__
-    if (opt.use_packing_layout)
-    {
-        elempack = inch % 8 == 0 ? 8 : 1;
-    }
-#endif // __SSE2__
-
-    // maxk-inch-outch to pa-maxk-inch/pa-outch
-    Mat A_data;
-    if (maxk == 1)
-    {
-        A_data = kernel.reshape(maxk * inch, outch);
-    }
-    else
-    {
-        Mat weight_data_r2 = kernel.reshape(maxk, inch, outch);
-
-        A_data.create(maxk * inch, outch, (size_t)1u, 1);
-
-        for (int q = 0; q < outch; q += 1)
-        {
-            signed char* g00 = A_data.row<signed char>(q);
-
-            for (int p = 0; p + (elempack - 1) < inch; p += elempack)
-            {
-                for (int k = 0; k < maxk; k++)
-                {
-                    for (int i = 0; i < elempack; i++)
-                    {
-                        const signed char* k00 = weight_data_r2.channel(q).row<const signed char>(p + i);
-                        g00[0] = k00[k];
-                        g00++;
-                    }
-                }
-            }
-        }
-    }
-
-    AT.create(TILE_K * TILE_M, (K + TILE_K - 1) / TILE_K, (M + TILE_M - 1) / TILE_M, (size_t)1u, 1);
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int k = 0; k < K; k += TILE_K)
-        {
-            const int max_kk = std::min((K - k), TILE_K);
-
-            Mat AT_tile = AT.channel(i / TILE_M).row_range(k / TILE_K, 1);
-
-            convolution_im2col_pack_A_tile_int8(A_data, AT_tile, i, max_ii, k, max_kk);
-        }
-    }
-}
-
-static void convolution_im2col_gemm_int8(const Mat& bottom_blob, Mat& top_blob, const Mat& AT, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int nT, const Option& opt)
-{
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
-#if NCNN_RUNTIME_CPU && NCNN_AVX512VNNI && __AVX512F__ && !__AVX512VNNI__
-    if (ncnn::cpu_support_x86_avx512vnni())
-    {
-        convolution_im2col_gemm_int8_avx512vnni(bottom_blob, top_blob, AT, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, nT, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__
-    if (ncnn::cpu_support_x86_avxvnni())
-    {
-        convolution_im2col_gemm_int8_avxvnni(bottom_blob, top_blob, AT, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, nT, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-    if (ncnn::cpu_support_x86_avx2())
-    {
-        convolution_im2col_gemm_int8_avx2(bottom_blob, top_blob, AT, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, nT, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__
-    if (ncnn::cpu_support_x86_xop())
-    {
-        convolution_im2col_gemm_int8_xop(bottom_blob, top_blob, AT, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, nT, opt);
-        return;
-    }
-#endif
-#endif
-
-    const int maxk = kernel_w * kernel_h;
-
-    const int M = top_blob.c * top_blob.elempack;
-    const int N = top_blob.w * top_blob.h;
-    const int K = bottom_blob.c * bottom_blob.elempack * maxk;
-
-    int TILE_M, TILE_N, TILE_K;
-    convolution_im2col_gemm_get_optimal_tile_mnk_int8(M, N, K, TILE_M, TILE_N, TILE_K, nT);
-
-    const int nn_M = (M + TILE_M - 1) / TILE_M;
-    const int nn_N = (N + TILE_N - 1) / TILE_N;
-    const int nn_K = (K + TILE_K - 1) / TILE_K;
-
-    // NCNN_LOGE("TILE M/N/K = %d %d %d -> %d %d %d", M, N, K, TILE_M, TILE_N, TILE_K);
-
-    Mat BT(TILE_K * TILE_N, (K + TILE_K - 1) / TILE_K, (N + TILE_N - 1) / TILE_N, 1u, opt.workspace_allocator);
-
-    const int nn_NK = nn_N * nn_K;
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppjk = 0; ppjk < nn_NK; ppjk++)
-    {
-        const int ppj = ppjk / nn_K;
-        const int ppk = ppjk % nn_K;
-
-        const int j = ppj * TILE_N;
-        const int k = ppk * TILE_K;
-
-        const int max_jj = std::min((N - j), TILE_N);
-        const int max_kk = std::min((K - k), TILE_K);
-
-        Mat BT_tile = BT.channel(j / TILE_N).row_range(k / TILE_K, 1);
-
-        // im2col
-        convolution_im2col_input_tile_int8(bottom_blob, BT_tile, j, max_jj, k, max_kk, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h);
-    }
-
-    Mat topT_tileX;
-    if (K > TILE_K)
-        topT_tileX.create(TILE_N * TILE_M, 1, nT, 4u, opt.workspace_allocator);
-
-    #pragma omp parallel for num_threads(nT)
-    for (int ppj = 0; ppj < nn_M; ppj++)
-    {
-        const int i = ppj * TILE_M;
-
-        Mat topT_tile;
-        if (K > TILE_K)
-            topT_tile = topT_tileX.channel(get_omp_thread_num());
-
-        const int max_ii = std::min((M - i), TILE_M);
-
-        for (int j = 0; j < N; j += TILE_N)
-        {
-            const int max_jj = std::min((N - j), TILE_N);
-
-            for (int k = 0; k < K; k += TILE_K)
-            {
-                const int max_kk = std::min((K - k), TILE_K);
-
-                const Mat AT_tile = AT.channel(i / TILE_M).row_range(k / TILE_K, 1);
-
-                const Mat BT_tile = BT.channel(j / TILE_N).row_range(k / TILE_K, 1);
-
-                bool k_end = k + TILE_K >= K;
-
-                convolution_gemm_transB_packed_tile_int8(AT_tile, BT_tile, topT_tile, top_blob, i, max_ii, j, max_jj, k, max_kk, k_end);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_packed.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_packed.h
deleted file mode 100644
index 97ba6583..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_packed.h
+++ /dev/null
@@ -1,2881 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convolution_transform_kernel_packed(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-    const int maxk = kernel_w * kernel_h;
-
-    // src = kw-kh-inch-outch
-    // dst = pb-pa-kw-kh-inch/pa-outch/pb
-
-    // clang-format off
-    // *INDENT-OFF*
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    if (outch >= 16)
-    {
-        if (inch >= 16)
-            kernel_tm.create(16 * 16 * maxk, inch / 16 + (inch % 16) / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 16 + (outch % 16) / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2);
-        else if (inch >= 8)
-            kernel_tm.create(16 * 8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 16 + (outch % 16) / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2);
-        else if (inch >= 4)
-            kernel_tm.create(16 * 4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch / 16 + (outch % 16) / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2);
-        else if (inch >= 2)
-            kernel_tm.create(16 * 2 * maxk, inch / 2 + inch % 2, outch / 16 + (outch % 16) / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2);
-        else
-            kernel_tm.create(16 * maxk, inch, outch / 16 + (outch % 16) / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2);
-    }
-    else
-#endif // __AVX512F__
-    if (outch >= 8)
-    {
-#if __AVX512F__
-        if (inch >= 16)
-            kernel_tm.create(8 * 16 * maxk, inch / 16 + (inch % 16) / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2);
-        else
-#endif // __AVX512F__
-        if (inch >= 8)
-            kernel_tm.create(8 * 8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2);
-        else if (inch >= 4)
-            kernel_tm.create(8 * 4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2);
-        else if (inch >= 2)
-            kernel_tm.create(8 * 2 * maxk, inch / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2);
-        else
-            kernel_tm.create(8 * maxk, inch, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2);
-    }
-    else
-#endif // __AVX__
-    if (outch >= 4)
-    {
-#if __AVX__
-#if __AVX512F__
-        if (inch >= 16)
-            kernel_tm.create(4 * 16 * maxk, inch / 16 + (inch % 16) / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2);
-        else
-#endif // __AVX512F__
-        if (inch >= 8)
-            kernel_tm.create(4 * 8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2);
-        else
-#endif // __AVX__
-        if (inch >= 4)
-            kernel_tm.create(4 * 4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2);
-        else if (inch >= 2)
-            kernel_tm.create(4 * 2 * maxk, inch / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2);
-        else
-            kernel_tm.create(4 * maxk, inch, outch / 4 + (outch % 4) / 2 + outch % 2);
-    }
-    else
-#endif // __SSE2__
-    if (outch >= 2)
-    {
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-        if (inch >= 16)
-            kernel_tm.create(2 * 16 * maxk, inch / 16 + (inch % 16) / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 2 + outch % 2);
-        else
-#endif // __AVX512F__
-        if (inch >= 8)
-            kernel_tm.create(2 * 8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch / 2 + outch % 2);
-        else
-#endif // __AVX__
-        if (inch >= 4)
-            kernel_tm.create(2 * 4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch / 2 + outch % 2);
-        else
-#endif // __SSE2__
-        if (inch >= 2)
-            kernel_tm.create(2 * 2 * maxk, inch / 2 + inch % 2, outch / 2 + outch % 2);
-        else
-            kernel_tm.create(2 * maxk, inch, outch / 2 + outch % 2);
-    }
-    else
-    {
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-        if (inch >= 16)
-            kernel_tm.create(16 * maxk, inch / 16 + (inch % 16) / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch);
-        else
-#endif // __AVX512F__
-        if (inch >= 8)
-            kernel_tm.create(8 * maxk, inch / 8 + (inch % 8) / 4 + (inch % 4) / 2 + inch % 2, outch);
-        else
-#endif // __AVX__
-        if (inch >= 4)
-            kernel_tm.create(4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch);
-        else
-#endif // __SSE2__
-        if (inch >= 2)
-            kernel_tm.create(2 * maxk, inch / 2 + inch % 2, outch);
-        else
-            kernel_tm.create(maxk, inch, outch);
-    }
-    // *INDENT-ON*
-    // clang-format on
-
-    int q = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    for (; q + 15 < outch; q += 16)
-    {
-        const float* kptr0 = (const float*)kernel + q * inch * maxk;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inch * maxk;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inch * maxk;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inch * maxk;
-        const float* kptr4 = (const float*)kernel + (q + 4) * inch * maxk;
-        const float* kptr5 = (const float*)kernel + (q + 5) * inch * maxk;
-        const float* kptr6 = (const float*)kernel + (q + 6) * inch * maxk;
-        const float* kptr7 = (const float*)kernel + (q + 7) * inch * maxk;
-        const float* kptr8 = (const float*)kernel + (q + 8) * inch * maxk;
-        const float* kptr9 = (const float*)kernel + (q + 9) * inch * maxk;
-        const float* kptra = (const float*)kernel + (q + 10) * inch * maxk;
-        const float* kptrb = (const float*)kernel + (q + 11) * inch * maxk;
-        const float* kptrc = (const float*)kernel + (q + 12) * inch * maxk;
-        const float* kptrd = (const float*)kernel + (q + 13) * inch * maxk;
-        const float* kptre = (const float*)kernel + (q + 14) * inch * maxk;
-        const float* kptrf = (const float*)kernel + (q + 15) * inch * maxk;
-
-        float* g00 = kernel_tm.channel(q / 16);
-
-        __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-        _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(maxk));
-
-        int p = 0;
-        for (; p + 15 < inch; p += 16)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-                const float* k4 = kptr4 + k;
-                const float* k5 = kptr5 + k;
-                const float* k6 = kptr6 + k;
-                const float* k7 = kptr7 + k;
-                const float* k8 = kptr8 + k;
-                const float* k9 = kptr9 + k;
-                const float* ka = kptra + k;
-                const float* kb = kptrb + k;
-                const float* kc = kptrc + k;
-                const float* kd = kptrd + k;
-                const float* ke = kptre + k;
-                const float* kf = kptrf + k;
-
-                __m512 _k0 = _mm512_i32gather_ps(_vindex, k0, sizeof(float));
-                __m512 _k1 = _mm512_i32gather_ps(_vindex, k1, sizeof(float));
-                __m512 _k2 = _mm512_i32gather_ps(_vindex, k2, sizeof(float));
-                __m512 _k3 = _mm512_i32gather_ps(_vindex, k3, sizeof(float));
-                __m512 _k4 = _mm512_i32gather_ps(_vindex, k4, sizeof(float));
-                __m512 _k5 = _mm512_i32gather_ps(_vindex, k5, sizeof(float));
-                __m512 _k6 = _mm512_i32gather_ps(_vindex, k6, sizeof(float));
-                __m512 _k7 = _mm512_i32gather_ps(_vindex, k7, sizeof(float));
-                __m512 _k8 = _mm512_i32gather_ps(_vindex, k8, sizeof(float));
-                __m512 _k9 = _mm512_i32gather_ps(_vindex, k9, sizeof(float));
-                __m512 _ka = _mm512_i32gather_ps(_vindex, ka, sizeof(float));
-                __m512 _kb = _mm512_i32gather_ps(_vindex, kb, sizeof(float));
-                __m512 _kc = _mm512_i32gather_ps(_vindex, kc, sizeof(float));
-                __m512 _kd = _mm512_i32gather_ps(_vindex, kd, sizeof(float));
-                __m512 _ke = _mm512_i32gather_ps(_vindex, ke, sizeof(float));
-                __m512 _kf = _mm512_i32gather_ps(_vindex, kf, sizeof(float));
-
-                transpose16x16_ps(_k0, _k1, _k2, _k3, _k4, _k5, _k6, _k7, _k8, _k9, _ka, _kb, _kc, _kd, _ke, _kf);
-
-                _mm512_store_ps(g00, _k0);
-                _mm512_store_ps(g00 + 16, _k1);
-                _mm512_store_ps(g00 + 16 * 2, _k2);
-                _mm512_store_ps(g00 + 16 * 3, _k3);
-                _mm512_store_ps(g00 + 16 * 4, _k4);
-                _mm512_store_ps(g00 + 16 * 5, _k5);
-                _mm512_store_ps(g00 + 16 * 6, _k6);
-                _mm512_store_ps(g00 + 16 * 7, _k7);
-                _mm512_store_ps(g00 + 16 * 8, _k8);
-                _mm512_store_ps(g00 + 16 * 9, _k9);
-                _mm512_store_ps(g00 + 16 * 10, _ka);
-                _mm512_store_ps(g00 + 16 * 11, _kb);
-                _mm512_store_ps(g00 + 16 * 12, _kc);
-                _mm512_store_ps(g00 + 16 * 13, _kd);
-                _mm512_store_ps(g00 + 16 * 14, _ke);
-                _mm512_store_ps(g00 + 16 * 15, _kf);
-
-                g00 += 256;
-            }
-
-            kptr0 += maxk * 16;
-            kptr1 += maxk * 16;
-            kptr2 += maxk * 16;
-            kptr3 += maxk * 16;
-            kptr4 += maxk * 16;
-            kptr5 += maxk * 16;
-            kptr6 += maxk * 16;
-            kptr7 += maxk * 16;
-            kptr8 += maxk * 16;
-            kptr9 += maxk * 16;
-            kptra += maxk * 16;
-            kptrb += maxk * 16;
-            kptrc += maxk * 16;
-            kptrd += maxk * 16;
-            kptre += maxk * 16;
-            kptrf += maxk * 16;
-        }
-
-        _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(inch));
-
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    __m512 _k0 = _mm512_i32gather_ps(_vindex, k0, sizeof(float));
-                    _mm512_store_ps(g00, _k0);
-                    k0 += maxk;
-                    g00 += 16;
-                }
-            }
-
-            kptr0 += maxk * 8;
-        }
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    __m512 _k0 = _mm512_i32gather_ps(_vindex, k0, sizeof(float));
-                    _mm512_store_ps(g00, _k0);
-                    k0 += maxk;
-                    g00 += 16;
-                }
-            }
-
-            kptr0 += maxk * 4;
-        }
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    __m512 _k0 = _mm512_i32gather_ps(_vindex, k0, sizeof(float));
-                    _mm512_store_ps(g00, _k0);
-                    k0 += maxk;
-                    g00 += 16;
-                }
-            }
-
-            kptr0 += maxk * 2;
-        }
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-
-                __m512 _k0 = _mm512_i32gather_ps(_vindex, k0, sizeof(float));
-                _mm512_store_ps(g00, _k0);
-                g00 += 16;
-            }
-        }
-    }
-#endif // __AVX512F__
-    for (; q + 7 < outch; q += 8)
-    {
-        const float* kptr0 = (const float*)kernel + q * inch * maxk;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inch * maxk;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inch * maxk;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inch * maxk;
-        const float* kptr4 = (const float*)kernel + (q + 4) * inch * maxk;
-        const float* kptr5 = (const float*)kernel + (q + 5) * inch * maxk;
-        const float* kptr6 = (const float*)kernel + (q + 6) * inch * maxk;
-        const float* kptr7 = (const float*)kernel + (q + 7) * inch * maxk;
-
-#if __AVX512F__
-        float* g00 = kernel_tm.channel(q / 16 + (q % 16) / 8);
-#else
-        float* g00 = kernel_tm.channel(q / 8);
-#endif
-
-#if __AVX2__
-        __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-        _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(maxk));
-#if __AVX512F__
-        __m512i _vindex_512 = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-        _vindex_512 = _mm512_mullo_epi32(_vindex_512, _mm512_set1_epi32(maxk));
-#endif // __AVX512F__
-#endif // __AVX2__
-
-        int p = 0;
-#if __AVX512F__
-        for (; p + 15 < inch; p += 16)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-                const float* k4 = kptr4 + k;
-                const float* k5 = kptr5 + k;
-                const float* k6 = kptr6 + k;
-                const float* k7 = kptr7 + k;
-
-                __m512 _k0 = _mm512_i32gather_ps(_vindex_512, k0, sizeof(float));
-                __m512 _k1 = _mm512_i32gather_ps(_vindex_512, k1, sizeof(float));
-                __m512 _k2 = _mm512_i32gather_ps(_vindex_512, k2, sizeof(float));
-                __m512 _k3 = _mm512_i32gather_ps(_vindex_512, k3, sizeof(float));
-                __m512 _k4 = _mm512_i32gather_ps(_vindex_512, k4, sizeof(float));
-                __m512 _k5 = _mm512_i32gather_ps(_vindex_512, k5, sizeof(float));
-                __m512 _k6 = _mm512_i32gather_ps(_vindex_512, k6, sizeof(float));
-                __m512 _k7 = _mm512_i32gather_ps(_vindex_512, k7, sizeof(float));
-
-                transpose16x8_ps(_k0, _k1, _k2, _k3, _k4, _k5, _k6, _k7);
-
-                _mm512_storeu_ps(g00, _k0);
-                _mm512_storeu_ps(g00 + 16, _k1);
-                _mm512_storeu_ps(g00 + 16 * 2, _k2);
-                _mm512_storeu_ps(g00 + 16 * 3, _k3);
-                _mm512_storeu_ps(g00 + 16 * 4, _k4);
-                _mm512_storeu_ps(g00 + 16 * 5, _k5);
-                _mm512_storeu_ps(g00 + 16 * 6, _k6);
-                _mm512_storeu_ps(g00 + 16 * 7, _k7);
-
-                g00 += 128;
-            }
-
-            kptr0 += maxk * 16;
-            kptr1 += maxk * 16;
-            kptr2 += maxk * 16;
-            kptr3 += maxk * 16;
-            kptr4 += maxk * 16;
-            kptr5 += maxk * 16;
-            kptr6 += maxk * 16;
-            kptr7 += maxk * 16;
-        }
-#endif // __AVX512F__
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-                const float* k4 = kptr4 + k;
-                const float* k5 = kptr5 + k;
-                const float* k6 = kptr6 + k;
-                const float* k7 = kptr7 + k;
-
-#if __AVX2__
-                __m256 _k0 = _mm256_i32gather_ps(k0, _vindex, sizeof(float));
-                __m256 _k1 = _mm256_i32gather_ps(k1, _vindex, sizeof(float));
-                __m256 _k2 = _mm256_i32gather_ps(k2, _vindex, sizeof(float));
-                __m256 _k3 = _mm256_i32gather_ps(k3, _vindex, sizeof(float));
-                __m256 _k4 = _mm256_i32gather_ps(k4, _vindex, sizeof(float));
-                __m256 _k5 = _mm256_i32gather_ps(k5, _vindex, sizeof(float));
-                __m256 _k6 = _mm256_i32gather_ps(k6, _vindex, sizeof(float));
-                __m256 _k7 = _mm256_i32gather_ps(k7, _vindex, sizeof(float));
-
-                transpose8x8_ps(_k0, _k1, _k2, _k3, _k4, _k5, _k6, _k7);
-
-                _mm256_store_ps(g00, _k0);
-                _mm256_store_ps(g00 + 8, _k1);
-                _mm256_store_ps(g00 + 8 * 2, _k2);
-                _mm256_store_ps(g00 + 8 * 3, _k3);
-                _mm256_store_ps(g00 + 8 * 4, _k4);
-                _mm256_store_ps(g00 + 8 * 5, _k5);
-                _mm256_store_ps(g00 + 8 * 6, _k6);
-                _mm256_store_ps(g00 + 8 * 7, _k7);
-
-                g00 += 64;
-#else  // __AVX2__
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[0];
-                    g00[1] = k1[0];
-                    g00[2] = k2[0];
-                    g00[3] = k3[0];
-                    g00[4] = k4[0];
-                    g00[5] = k5[0];
-                    g00[6] = k6[0];
-                    g00[7] = k7[0];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    k4 += maxk;
-                    k5 += maxk;
-                    k6 += maxk;
-                    k7 += maxk;
-                    g00 += 8;
-                }
-#endif // __AVX2__
-            }
-
-            kptr0 += maxk * 8;
-            kptr1 += maxk * 8;
-            kptr2 += maxk * 8;
-            kptr3 += maxk * 8;
-            kptr4 += maxk * 8;
-            kptr5 += maxk * 8;
-            kptr6 += maxk * 8;
-            kptr7 += maxk * 8;
-        }
-
-#if __AVX2__
-        _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(inch));
-#endif // __AVX2__
-
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-#if !__AVX2__
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-                const float* k4 = kptr4 + k;
-                const float* k5 = kptr5 + k;
-                const float* k6 = kptr6 + k;
-                const float* k7 = kptr7 + k;
-#endif // !__AVX2__
-
-                for (int i = 0; i < 4; i++)
-                {
-#if __AVX2__
-                    __m256 _k0 = _mm256_i32gather_ps(k0, _vindex, sizeof(float));
-                    _mm256_store_ps(g00, _k0);
-                    k0 += maxk;
-                    g00 += 8;
-#else  // __AVX2__
-                    g00[0] = k0[0];
-                    g00[1] = k1[0];
-                    g00[2] = k2[0];
-                    g00[3] = k3[0];
-                    g00[4] = k4[0];
-                    g00[5] = k5[0];
-                    g00[6] = k6[0];
-                    g00[7] = k7[0];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    k4 += maxk;
-                    k5 += maxk;
-                    k6 += maxk;
-                    k7 += maxk;
-                    g00 += 8;
-#endif // __AVX2__
-                }
-            }
-
-            kptr0 += maxk * 4;
-#if !__AVX2__
-            kptr1 += maxk * 4;
-            kptr2 += maxk * 4;
-            kptr3 += maxk * 4;
-            kptr4 += maxk * 4;
-            kptr5 += maxk * 4;
-            kptr6 += maxk * 4;
-            kptr7 += maxk * 4;
-#endif // !__AVX2__
-        }
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-#if !__AVX2__
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-                const float* k4 = kptr4 + k;
-                const float* k5 = kptr5 + k;
-                const float* k6 = kptr6 + k;
-                const float* k7 = kptr7 + k;
-#endif // !__AVX2__
-
-                for (int i = 0; i < 2; i++)
-                {
-#if __AVX2__
-                    __m256 _k0 = _mm256_i32gather_ps(k0, _vindex, sizeof(float));
-                    _mm256_store_ps(g00, _k0);
-                    k0 += maxk;
-                    g00 += 8;
-#else  // __AVX2__
-                    g00[0] = k0[0];
-                    g00[1] = k1[0];
-                    g00[2] = k2[0];
-                    g00[3] = k3[0];
-                    g00[4] = k4[0];
-                    g00[5] = k5[0];
-                    g00[6] = k6[0];
-                    g00[7] = k7[0];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    k4 += maxk;
-                    k5 += maxk;
-                    k6 += maxk;
-                    k7 += maxk;
-                    g00 += 8;
-#endif // __AVX2__
-                }
-            }
-
-            kptr0 += maxk * 2;
-#if !__AVX2__
-            kptr1 += maxk * 2;
-            kptr2 += maxk * 2;
-            kptr3 += maxk * 2;
-            kptr4 += maxk * 2;
-            kptr5 += maxk * 2;
-            kptr6 += maxk * 2;
-            kptr7 += maxk * 2;
-#endif // !__AVX2__
-        }
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-#if __AVX2__
-                __m256 _k0 = _mm256_i32gather_ps(k0, _vindex, sizeof(float));
-                _mm256_store_ps(g00, _k0);
-                g00 += 8;
-#else  // __AVX2__
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-                const float* k4 = kptr4 + k;
-                const float* k5 = kptr5 + k;
-                const float* k6 = kptr6 + k;
-                const float* k7 = kptr7 + k;
-
-                g00[0] = k0[0];
-                g00[1] = k1[0];
-                g00[2] = k2[0];
-                g00[3] = k3[0];
-                g00[4] = k4[0];
-                g00[5] = k5[0];
-                g00[6] = k6[0];
-                g00[7] = k7[0];
-                g00 += 8;
-#endif // __AVX2__
-            }
-        }
-    }
-#endif // __AVX__
-    for (; q + 3 < outch; q += 4)
-    {
-        const float* kptr0 = (const float*)kernel + q * inch * maxk;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inch * maxk;
-        const float* kptr2 = (const float*)kernel + (q + 2) * inch * maxk;
-        const float* kptr3 = (const float*)kernel + (q + 3) * inch * maxk;
-
-#if __AVX512F__
-        float* g00 = kernel_tm.channel(q / 16 + (q % 16) / 8 + (q % 8) / 4);
-#elif __AVX__
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4);
-#else
-        float* g00 = kernel_tm.channel(q / 4);
-#endif
-
-#if __AVX2__
-        __m128i _vindex = _mm_setr_epi32(0, 1, 2, 3);
-        _vindex = _mm_mullo_epi32(_vindex, _mm_set1_epi32(maxk));
-        __m256i _vindex_256 = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-        _vindex_256 = _mm256_mullo_epi32(_vindex_256, _mm256_set1_epi32(maxk));
-#if __AVX512F__
-        __m512i _vindex_512 = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-        _vindex_512 = _mm512_mullo_epi32(_vindex_512, _mm512_set1_epi32(maxk));
-#endif // __AVX512F__
-#endif // __AVX2__
-
-        int p = 0;
-#if __AVX__
-#if __AVX512F__
-        for (; p + 15 < inch; p += 16)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-
-                __m512 _k0 = _mm512_i32gather_ps(_vindex_512, k0, sizeof(float));
-                __m512 _k1 = _mm512_i32gather_ps(_vindex_512, k1, sizeof(float));
-                __m512 _k2 = _mm512_i32gather_ps(_vindex_512, k2, sizeof(float));
-                __m512 _k3 = _mm512_i32gather_ps(_vindex_512, k3, sizeof(float));
-
-                transpose16x4_ps(_k0, _k1, _k2, _k3);
-
-                _mm512_storeu_ps(g00, _k0);
-                _mm512_storeu_ps(g00 + 16, _k1);
-                _mm512_storeu_ps(g00 + 16 * 2, _k2);
-                _mm512_storeu_ps(g00 + 16 * 3, _k3);
-
-                g00 += 64;
-            }
-
-            kptr0 += maxk * 16;
-            kptr1 += maxk * 16;
-            kptr2 += maxk * 16;
-            kptr3 += maxk * 16;
-        }
-#endif // __AVX512F__
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-
-#if __AVX2__
-                __m256 _k0 = _mm256_i32gather_ps(k0, _vindex_256, sizeof(float));
-                __m256 _k1 = _mm256_i32gather_ps(k1, _vindex_256, sizeof(float));
-                __m256 _k2 = _mm256_i32gather_ps(k2, _vindex_256, sizeof(float));
-                __m256 _k3 = _mm256_i32gather_ps(k3, _vindex_256, sizeof(float));
-
-                transpose8x4_ps(_k0, _k1, _k2, _k3);
-
-                _mm256_storeu_ps(g00, _k0);
-                _mm256_storeu_ps(g00 + 8, _k1);
-                _mm256_storeu_ps(g00 + 8 * 2, _k2);
-                _mm256_storeu_ps(g00 + 8 * 3, _k3);
-
-                g00 += 32;
-#else  // __AVX2__
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[0];
-                    g00[1] = k1[0];
-                    g00[2] = k2[0];
-                    g00[3] = k3[0];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    g00 += 4;
-                }
-#endif // __AVX2__
-            }
-
-            kptr0 += maxk * 8;
-            kptr1 += maxk * 8;
-            kptr2 += maxk * 8;
-            kptr3 += maxk * 8;
-        }
-#endif // __AVX__
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-
-#if __AVX2__
-                __m128 _k0 = _mm_i32gather_ps(k0, _vindex, sizeof(float));
-                __m128 _k1 = _mm_i32gather_ps(k1, _vindex, sizeof(float));
-                __m128 _k2 = _mm_i32gather_ps(k2, _vindex, sizeof(float));
-                __m128 _k3 = _mm_i32gather_ps(k3, _vindex, sizeof(float));
-
-                _MM_TRANSPOSE4_PS(_k0, _k1, _k2, _k3);
-
-                _mm_store_ps(g00, _k0);
-                _mm_store_ps(g00 + 4, _k1);
-                _mm_store_ps(g00 + 4 * 2, _k2);
-                _mm_store_ps(g00 + 4 * 3, _k3);
-
-                g00 += 16;
-#else  // __AVX2__
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k0[0];
-                    g00[1] = k1[0];
-                    g00[2] = k2[0];
-                    g00[3] = k3[0];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    g00 += 4;
-                }
-#endif // __AVX2__
-            }
-
-            kptr0 += maxk * 4;
-            kptr1 += maxk * 4;
-            kptr2 += maxk * 4;
-            kptr3 += maxk * 4;
-        }
-
-#if __AVX2__
-        _vindex = _mm_mullo_epi32(_vindex, _mm_set1_epi32(inch));
-#endif // __AVX2__
-
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-#if !__AVX2__
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-#endif // !__AVX2__
-
-                for (int i = 0; i < 2; i++)
-                {
-#if __AVX2__
-                    __m128 _k0 = _mm_i32gather_ps(k0, _vindex, sizeof(float));
-                    _mm_store_ps(g00, _k0);
-                    k0 += maxk;
-                    g00 += 4;
-#else  // __AVX2__
-                    g00[0] = k0[0];
-                    g00[1] = k1[0];
-                    g00[2] = k2[0];
-                    g00[3] = k3[0];
-                    k0 += maxk;
-                    k1 += maxk;
-                    k2 += maxk;
-                    k3 += maxk;
-                    g00 += 4;
-#endif // __AVX2__
-                }
-            }
-
-            kptr0 += maxk * 2;
-#if !__AVX2__
-            kptr1 += maxk * 2;
-            kptr2 += maxk * 2;
-            kptr3 += maxk * 2;
-#endif // !__AVX2__
-        }
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-#if __AVX2__
-                __m128 _k0 = _mm_i32gather_ps(k0, _vindex, sizeof(float));
-                _mm_store_ps(g00, _k0);
-                g00 += 4;
-#else  // __AVX2__
-                const float* k1 = kptr1 + k;
-                const float* k2 = kptr2 + k;
-                const float* k3 = kptr3 + k;
-
-                g00[0] = k0[0];
-                g00[1] = k1[0];
-                g00[2] = k2[0];
-                g00[3] = k3[0];
-                g00 += 4;
-#endif // __AVX2__
-            }
-        }
-    }
-#endif // __SSE2__
-    for (; q + 1 < outch; q += 2)
-    {
-        const float* kptr0 = (const float*)kernel + q * inch * maxk;
-        const float* kptr1 = (const float*)kernel + (q + 1) * inch * maxk;
-
-#if __AVX512F__
-        float* g00 = kernel_tm.channel(q / 16 + (q % 16) / 8 + (q % 8) / 4 + (q % 4) / 2);
-#elif __AVX__
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2);
-#elif __SSE2__
-        float* g00 = kernel_tm.channel(q / 4 + (q % 4) / 2);
-#else
-        float* g00 = kernel_tm.channel(q / 2);
-#endif
-
-#if __AVX2__
-        __m128i _vindex = _mm_setr_epi32(0, 1, 2, 3);
-        _vindex = _mm_mullo_epi32(_vindex, _mm_set1_epi32(maxk));
-        __m256i _vindex_256 = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-        _vindex_256 = _mm256_mullo_epi32(_vindex_256, _mm256_set1_epi32(maxk));
-#if __AVX512F__
-        __m512i _vindex_512 = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-        _vindex_512 = _mm512_mullo_epi32(_vindex_512, _mm512_set1_epi32(maxk));
-#endif // __AVX512F__
-#endif // __AVX2__
-
-        int p = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-        for (; p + 15 < inch; p += 16)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-
-                __m512 _k0 = _mm512_i32gather_ps(_vindex_512, k0, sizeof(float));
-                __m512 _k1 = _mm512_i32gather_ps(_vindex_512, k1, sizeof(float));
-                _mm512_storeu_ps(g00, _k0);
-                _mm512_storeu_ps(g00 + 16, _k1);
-                g00 += 32;
-            }
-
-            kptr0 += maxk * 16;
-            kptr1 += maxk * 16;
-        }
-#endif // __AVX512F__
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-
-#if __AVX2__
-                __m256 _k0 = _mm256_i32gather_ps(k0, _vindex_256, sizeof(float));
-                __m256 _k1 = _mm256_i32gather_ps(k1, _vindex_256, sizeof(float));
-                _mm256_storeu_ps(g00, _k0);
-                _mm256_storeu_ps(g00 + 8, _k1);
-                g00 += 16;
-#else  // __AVX2__
-                g00[0] = k0[0];
-                g00[1] = k0[maxk];
-                g00[2] = k0[maxk * 2];
-                g00[3] = k0[maxk * 3];
-                g00[4] = k0[maxk * 4];
-                g00[5] = k0[maxk * 5];
-                g00[6] = k0[maxk * 6];
-                g00[7] = k0[maxk * 7];
-                g00[8] = k1[0];
-                g00[9] = k1[maxk];
-                g00[10] = k1[maxk * 2];
-                g00[11] = k1[maxk * 3];
-                g00[12] = k1[maxk * 4];
-                g00[13] = k1[maxk * 5];
-                g00[14] = k1[maxk * 6];
-                g00[15] = k1[maxk * 7];
-                g00 += 16;
-#endif // __AVX2__
-            }
-
-            kptr0 += maxk * 8;
-            kptr1 += maxk * 8;
-        }
-#endif // __AVX__
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-
-#if __AVX2__
-                __m128 _k0 = _mm_i32gather_ps(k0, _vindex, sizeof(float));
-                __m128 _k1 = _mm_i32gather_ps(k1, _vindex, sizeof(float));
-                _mm_storeu_ps(g00, _k0);
-                _mm_storeu_ps(g00 + 4, _k1);
-                g00 += 8;
-#else  // __AVX2__
-                g00[0] = k0[0];
-                g00[1] = k0[maxk];
-                g00[2] = k0[maxk * 2];
-                g00[3] = k0[maxk * 3];
-                g00[4] = k1[0];
-                g00[5] = k1[maxk];
-                g00[6] = k1[maxk * 2];
-                g00[7] = k1[maxk * 3];
-                g00 += 8;
-#endif // __AVX2__
-            }
-
-            kptr0 += maxk * 4;
-            kptr1 += maxk * 4;
-        }
-#endif // __SSE2__
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = k0[0];
-                    g00[1] = k1[0];
-                    k0 += maxk;
-                    k1 += maxk;
-                    g00 += 2;
-                }
-            }
-
-            kptr0 += maxk * 2;
-            kptr1 += maxk * 2;
-        }
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr0 + k;
-                const float* k1 = kptr1 + k;
-
-                g00[0] = k0[0];
-                g00[1] = k1[0];
-                g00 += 2;
-            }
-        }
-    }
-    for (; q < outch; q++)
-    {
-        const float* kptr = (const float*)kernel + q * inch * maxk;
-
-#if __AVX512F__
-        float* g00 = kernel_tm.channel(q / 16 + (q % 16) / 8 + (q % 8) / 4 + (q % 4) / 2 + q % 2);
-#elif __AVX__
-        float* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2 + q % 2);
-#elif __SSE2__
-        float* g00 = kernel_tm.channel(q / 4 + (q % 4) / 2 + q % 2);
-#else
-        float* g00 = kernel_tm.channel(q / 2 + q % 2);
-#endif
-
-#if __AVX2__
-        __m128i _vindex = _mm_setr_epi32(0, 1, 2, 3);
-        _vindex = _mm_mullo_epi32(_vindex, _mm_set1_epi32(maxk));
-        __m256i _vindex_256 = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-        _vindex_256 = _mm256_mullo_epi32(_vindex_256, _mm256_set1_epi32(maxk));
-#if __AVX512F__
-        __m512i _vindex_512 = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-        _vindex_512 = _mm512_mullo_epi32(_vindex_512, _mm512_set1_epi32(maxk));
-#endif // __AVX512F__
-#endif // __AVX2__
-
-        int p = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-        for (; p + 15 < inch; p += 16)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr + k;
-
-                __m512 _k0 = _mm512_i32gather_ps(_vindex_512, k0, sizeof(float));
-                _mm512_storeu_ps(g00, _k0);
-                g00 += 16;
-            }
-
-            kptr += maxk * 16;
-        }
-#endif // __AVX512F__
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr + k;
-
-#if __AVX2__
-                __m256 _k0 = _mm256_i32gather_ps(k0, _vindex_256, sizeof(float));
-                _mm256_storeu_ps(g00, _k0);
-                g00 += 8;
-#else  // __AVX2__
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[0];
-                    k0 += maxk;
-                    g00 += 1;
-                }
-#endif // __AVX2__
-            }
-
-            kptr += maxk * 8;
-        }
-#endif // __AVX__
-        for (; p + 3 < inch; p += 4)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr + k;
-
-#if __AVX2__
-                __m128 _k0 = _mm_i32gather_ps(k0, _vindex, sizeof(float));
-                _mm_storeu_ps(g00, _k0);
-                g00 += 4;
-#else  // __AVX2__
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k0[0];
-                    k0 += maxk;
-                    g00 += 1;
-                }
-#endif // __AVX2__
-            }
-
-            kptr += maxk * 4;
-        }
-#endif // __SSE2__
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr + k;
-
-                for (int i = 0; i < 2; i++)
-                {
-                    g00[0] = k0[0];
-                    k0 += maxk;
-                    g00 += 1;
-                }
-            }
-
-            kptr += maxk * 2;
-        }
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const float* k0 = kptr + k;
-
-                g00[0] = k0[0];
-                g00++;
-            }
-        }
-    }
-}
-
-static void convolution_packed(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const int w = bottom_blob.w;
-    const int elempack = bottom_blob.elempack;
-    const int inch = bottom_blob.c * elempack;
-
-    const int N = bottom_blob.cstep * elempack;
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int outch = top_blob.c * out_elempack;
-
-    const int M = top_blob.cstep * out_elempack;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2 * elempack;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    const float* bias_data_ptr = bias_data;
-
-    int nn_outch = 0;
-    int remain_outch_start = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    nn_outch = outch / 16;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = pp * 16;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int out_elempack = top_blob.elempack;
-
-        float* outptr = top_blob.channel(p / out_elempack);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m512 _sum0 = _mm512_setzero_ps();
-                __m512 _sum1 = _mm512_setzero_ps();
-                __m512 _sum2 = _mm512_setzero_ps();
-                __m512 _sum3 = _mm512_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm512_loadu_ps(bias_data_ptr + p);
-                }
-
-                const float* kptr = weight_data_tm.channel(p / 16);
-
-                int q = 0;
-                for (; q + 15 < inch; q += 16)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    if (elempack == 16)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-
-                            __m512 _w0 = _mm512_load_ps(kptr + 16 * 0);
-                            __m512 _w1 = _mm512_load_ps(kptr + 16 * 1);
-                            __m512 _w2 = _mm512_load_ps(kptr + 16 * 2);
-                            __m512 _w3 = _mm512_load_ps(kptr + 16 * 3);
-                            __m512 _w4 = _mm512_load_ps(kptr + 16 * 4);
-                            __m512 _w5 = _mm512_load_ps(kptr + 16 * 5);
-                            __m512 _w6 = _mm512_load_ps(kptr + 16 * 6);
-                            __m512 _w7 = _mm512_load_ps(kptr + 16 * 7);
-                            __m512 _w8 = _mm512_load_ps(kptr + 16 * 8);
-                            __m512 _w9 = _mm512_load_ps(kptr + 16 * 9);
-                            __m512 _wa = _mm512_load_ps(kptr + 16 * 10);
-                            __m512 _wb = _mm512_load_ps(kptr + 16 * 11);
-                            __m512 _wc = _mm512_load_ps(kptr + 16 * 12);
-                            __m512 _wd = _mm512_load_ps(kptr + 16 * 13);
-                            __m512 _we = _mm512_load_ps(kptr + 16 * 14);
-                            __m512 _wf = _mm512_load_ps(kptr + 16 * 15);
-
-                            _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0s[3]), _sum3);
-                            _sum0 = _mm512_fmadd_ps(_w4, _mm512_set1_ps(r0s[4]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w5, _mm512_set1_ps(r0s[5]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w6, _mm512_set1_ps(r0s[6]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w7, _mm512_set1_ps(r0s[7]), _sum3);
-                            _sum0 = _mm512_fmadd_ps(_w8, _mm512_set1_ps(r0s[8]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w9, _mm512_set1_ps(r0s[9]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_wa, _mm512_set1_ps(r0s[10]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_wb, _mm512_set1_ps(r0s[11]), _sum3);
-                            _sum0 = _mm512_fmadd_ps(_wc, _mm512_set1_ps(r0s[12]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_wd, _mm512_set1_ps(r0s[13]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_we, _mm512_set1_ps(r0s[14]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_wf, _mm512_set1_ps(r0s[15]), _sum3);
-
-                            kptr += 256;
-                        }
-                    }
-                    if (elempack == 8)
-                    {
-                        const float* r1 = r0 + N;
-
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-                            const float* r1s = r1 + space_ofs[k];
-
-                            __m512 _w0 = _mm512_load_ps(kptr + 16 * 0);
-                            __m512 _w1 = _mm512_load_ps(kptr + 16 * 1);
-                            __m512 _w2 = _mm512_load_ps(kptr + 16 * 2);
-                            __m512 _w3 = _mm512_load_ps(kptr + 16 * 3);
-                            __m512 _w4 = _mm512_load_ps(kptr + 16 * 4);
-                            __m512 _w5 = _mm512_load_ps(kptr + 16 * 5);
-                            __m512 _w6 = _mm512_load_ps(kptr + 16 * 6);
-                            __m512 _w7 = _mm512_load_ps(kptr + 16 * 7);
-                            __m512 _w8 = _mm512_load_ps(kptr + 16 * 8);
-                            __m512 _w9 = _mm512_load_ps(kptr + 16 * 9);
-                            __m512 _wa = _mm512_load_ps(kptr + 16 * 10);
-                            __m512 _wb = _mm512_load_ps(kptr + 16 * 11);
-                            __m512 _wc = _mm512_load_ps(kptr + 16 * 12);
-                            __m512 _wd = _mm512_load_ps(kptr + 16 * 13);
-                            __m512 _we = _mm512_load_ps(kptr + 16 * 14);
-                            __m512 _wf = _mm512_load_ps(kptr + 16 * 15);
-
-                            _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0s[3]), _sum3);
-                            _sum0 = _mm512_fmadd_ps(_w4, _mm512_set1_ps(r0s[4]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w5, _mm512_set1_ps(r0s[5]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w6, _mm512_set1_ps(r0s[6]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w7, _mm512_set1_ps(r0s[7]), _sum3);
-                            _sum0 = _mm512_fmadd_ps(_w8, _mm512_set1_ps(r1s[0]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w9, _mm512_set1_ps(r1s[1]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_wa, _mm512_set1_ps(r1s[2]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_wb, _mm512_set1_ps(r1s[3]), _sum3);
-                            _sum0 = _mm512_fmadd_ps(_wc, _mm512_set1_ps(r1s[4]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_wd, _mm512_set1_ps(r1s[5]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_we, _mm512_set1_ps(r1s[6]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_wf, _mm512_set1_ps(r1s[7]), _sum3);
-
-                            kptr += 256;
-                        }
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-                            const float* r1s = r1 + space_ofs[k];
-                            const float* r2s = r2 + space_ofs[k];
-                            const float* r3s = r3 + space_ofs[k];
-
-                            __m512 _w0 = _mm512_load_ps(kptr + 16 * 0);
-                            __m512 _w1 = _mm512_load_ps(kptr + 16 * 1);
-                            __m512 _w2 = _mm512_load_ps(kptr + 16 * 2);
-                            __m512 _w3 = _mm512_load_ps(kptr + 16 * 3);
-                            __m512 _w4 = _mm512_load_ps(kptr + 16 * 4);
-                            __m512 _w5 = _mm512_load_ps(kptr + 16 * 5);
-                            __m512 _w6 = _mm512_load_ps(kptr + 16 * 6);
-                            __m512 _w7 = _mm512_load_ps(kptr + 16 * 7);
-                            __m512 _w8 = _mm512_load_ps(kptr + 16 * 8);
-                            __m512 _w9 = _mm512_load_ps(kptr + 16 * 9);
-                            __m512 _wa = _mm512_load_ps(kptr + 16 * 10);
-                            __m512 _wb = _mm512_load_ps(kptr + 16 * 11);
-                            __m512 _wc = _mm512_load_ps(kptr + 16 * 12);
-                            __m512 _wd = _mm512_load_ps(kptr + 16 * 13);
-                            __m512 _we = _mm512_load_ps(kptr + 16 * 14);
-                            __m512 _wf = _mm512_load_ps(kptr + 16 * 15);
-
-                            _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0s[3]), _sum3);
-                            _sum0 = _mm512_fmadd_ps(_w4, _mm512_set1_ps(r1s[0]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w5, _mm512_set1_ps(r1s[1]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w6, _mm512_set1_ps(r1s[2]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w7, _mm512_set1_ps(r1s[3]), _sum3);
-                            _sum0 = _mm512_fmadd_ps(_w8, _mm512_set1_ps(r2s[0]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w9, _mm512_set1_ps(r2s[1]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_wa, _mm512_set1_ps(r2s[2]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_wb, _mm512_set1_ps(r2s[3]), _sum3);
-                            _sum0 = _mm512_fmadd_ps(_wc, _mm512_set1_ps(r3s[0]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_wd, _mm512_set1_ps(r3s[1]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_we, _mm512_set1_ps(r3s[2]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_wf, _mm512_set1_ps(r3s[3]), _sum3);
-
-                            kptr += 256;
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-
-                            __m512 _w0 = _mm512_load_ps(kptr + 16 * 0);
-                            __m512 _w1 = _mm512_load_ps(kptr + 16 * 1);
-                            __m512 _w2 = _mm512_load_ps(kptr + 16 * 2);
-                            __m512 _w3 = _mm512_load_ps(kptr + 16 * 3);
-                            __m512 _w4 = _mm512_load_ps(kptr + 16 * 4);
-                            __m512 _w5 = _mm512_load_ps(kptr + 16 * 5);
-                            __m512 _w6 = _mm512_load_ps(kptr + 16 * 6);
-                            __m512 _w7 = _mm512_load_ps(kptr + 16 * 7);
-                            __m512 _w8 = _mm512_load_ps(kptr + 16 * 8);
-                            __m512 _w9 = _mm512_load_ps(kptr + 16 * 9);
-                            __m512 _wa = _mm512_load_ps(kptr + 16 * 10);
-                            __m512 _wb = _mm512_load_ps(kptr + 16 * 11);
-                            __m512 _wc = _mm512_load_ps(kptr + 16 * 12);
-                            __m512 _wd = _mm512_load_ps(kptr + 16 * 13);
-                            __m512 _we = _mm512_load_ps(kptr + 16 * 14);
-                            __m512 _wf = _mm512_load_ps(kptr + 16 * 15);
-
-                            _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0[sok]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0[sok + N]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0[sok + N * 2]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0[sok + N * 3]), _sum3);
-                            _sum0 = _mm512_fmadd_ps(_w4, _mm512_set1_ps(r0[sok + N * 4]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w5, _mm512_set1_ps(r0[sok + N * 5]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w6, _mm512_set1_ps(r0[sok + N * 6]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w7, _mm512_set1_ps(r0[sok + N * 7]), _sum3);
-                            _sum0 = _mm512_fmadd_ps(_w8, _mm512_set1_ps(r0[sok + N * 8]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w9, _mm512_set1_ps(r0[sok + N * 9]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_wa, _mm512_set1_ps(r0[sok + N * 10]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_wb, _mm512_set1_ps(r0[sok + N * 11]), _sum3);
-                            _sum0 = _mm512_fmadd_ps(_wc, _mm512_set1_ps(r0[sok + N * 12]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_wd, _mm512_set1_ps(r0[sok + N * 13]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_we, _mm512_set1_ps(r0[sok + N * 14]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_wf, _mm512_set1_ps(r0[sok + N * 15]), _sum3);
-
-                            kptr += 256;
-                        }
-                    }
-                }
-                for (; q + 7 < inch; q += 8)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    if (elempack == 8)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-
-                            __m512 _w0 = _mm512_load_ps(kptr + 16 * 0);
-                            __m512 _w1 = _mm512_load_ps(kptr + 16 * 1);
-                            __m512 _w2 = _mm512_load_ps(kptr + 16 * 2);
-                            __m512 _w3 = _mm512_load_ps(kptr + 16 * 3);
-                            __m512 _w4 = _mm512_load_ps(kptr + 16 * 4);
-                            __m512 _w5 = _mm512_load_ps(kptr + 16 * 5);
-                            __m512 _w6 = _mm512_load_ps(kptr + 16 * 6);
-                            __m512 _w7 = _mm512_load_ps(kptr + 16 * 7);
-
-                            _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0s[3]), _sum3);
-                            _sum0 = _mm512_fmadd_ps(_w4, _mm512_set1_ps(r0s[4]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w5, _mm512_set1_ps(r0s[5]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w6, _mm512_set1_ps(r0s[6]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w7, _mm512_set1_ps(r0s[7]), _sum3);
-
-                            kptr += 128;
-                        }
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-                            const float* r1s = r1 + space_ofs[k];
-
-                            __m512 _w0 = _mm512_load_ps(kptr + 16 * 0);
-                            __m512 _w1 = _mm512_load_ps(kptr + 16 * 1);
-                            __m512 _w2 = _mm512_load_ps(kptr + 16 * 2);
-                            __m512 _w3 = _mm512_load_ps(kptr + 16 * 3);
-                            __m512 _w4 = _mm512_load_ps(kptr + 16 * 4);
-                            __m512 _w5 = _mm512_load_ps(kptr + 16 * 5);
-                            __m512 _w6 = _mm512_load_ps(kptr + 16 * 6);
-                            __m512 _w7 = _mm512_load_ps(kptr + 16 * 7);
-
-                            _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0s[3]), _sum3);
-                            _sum0 = _mm512_fmadd_ps(_w4, _mm512_set1_ps(r1s[0]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w5, _mm512_set1_ps(r1s[1]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w6, _mm512_set1_ps(r1s[2]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w7, _mm512_set1_ps(r1s[3]), _sum3);
-
-                            kptr += 128;
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-
-                            __m512 _w0 = _mm512_load_ps(kptr + 16 * 0);
-                            __m512 _w1 = _mm512_load_ps(kptr + 16 * 1);
-                            __m512 _w2 = _mm512_load_ps(kptr + 16 * 2);
-                            __m512 _w3 = _mm512_load_ps(kptr + 16 * 3);
-                            __m512 _w4 = _mm512_load_ps(kptr + 16 * 4);
-                            __m512 _w5 = _mm512_load_ps(kptr + 16 * 5);
-                            __m512 _w6 = _mm512_load_ps(kptr + 16 * 6);
-                            __m512 _w7 = _mm512_load_ps(kptr + 16 * 7);
-
-                            _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0[sok]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0[sok + N]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0[sok + N * 2]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0[sok + N * 3]), _sum3);
-                            _sum0 = _mm512_fmadd_ps(_w4, _mm512_set1_ps(r0[sok + N * 4]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w5, _mm512_set1_ps(r0[sok + N * 5]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w6, _mm512_set1_ps(r0[sok + N * 6]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w7, _mm512_set1_ps(r0[sok + N * 7]), _sum3);
-
-                            kptr += 128;
-                        }
-                    }
-                }
-                for (; q + 3 < inch; q += 4)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    if (elempack == 4)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-
-                            __m512 _w0 = _mm512_load_ps(kptr);
-                            __m512 _w1 = _mm512_load_ps(kptr + 16);
-                            __m512 _w2 = _mm512_load_ps(kptr + 32);
-                            __m512 _w3 = _mm512_load_ps(kptr + 48);
-
-                            _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0s[3]), _sum3);
-
-                            kptr += 64;
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-
-                            __m512 _w0 = _mm512_load_ps(kptr);
-                            __m512 _w1 = _mm512_load_ps(kptr + 16);
-                            __m512 _w2 = _mm512_load_ps(kptr + 32);
-                            __m512 _w3 = _mm512_load_ps(kptr + 48);
-
-                            _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0[sok]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0[sok + N]), _sum1);
-                            _sum2 = _mm512_fmadd_ps(_w2, _mm512_set1_ps(r0[sok + N * 2]), _sum2);
-                            _sum3 = _mm512_fmadd_ps(_w3, _mm512_set1_ps(r0[sok + N * 3]), _sum3);
-
-                            kptr += 64;
-                        }
-                    }
-                }
-                for (; q + 1 < inch; q += 2)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    // if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-
-                            __m512 _w0 = _mm512_load_ps(kptr);
-                            __m512 _w1 = _mm512_load_ps(kptr + 16);
-
-                            _sum0 = _mm512_fmadd_ps(_w0, _mm512_set1_ps(r0[sok]), _sum0);
-                            _sum1 = _mm512_fmadd_ps(_w1, _mm512_set1_ps(r0[sok + N]), _sum1);
-
-                            kptr += 32;
-                        }
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    // if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            __m512 _val = _mm512_set1_ps(r0[space_ofs[k]]);
-                            __m512 _w = _mm512_load_ps(kptr);
-                            _sum0 = _mm512_fmadd_ps(_val, _w, _sum0);
-
-                            kptr += 16;
-                        }
-                    }
-                }
-
-                _sum0 = _mm512_add_ps(_sum0, _sum1);
-                _sum2 = _mm512_add_ps(_sum2, _sum3);
-                _sum0 = _mm512_add_ps(_sum0, _sum2);
-
-                _sum0 = activation_avx512(_sum0, activation_type, activation_params);
-
-                if (out_elempack == 16)
-                {
-                    _mm512_store_ps(outptr, _sum0);
-                    outptr += 16;
-                }
-                if (out_elempack == 8)
-                {
-                    _mm256_store_ps(outptr, _mm512_extractf32x8_ps(_sum0, 0));
-                    _mm256_store_ps(outptr + M, _mm512_extractf32x8_ps(_sum0, 1));
-                    outptr += 8;
-                }
-                if (out_elempack == 4)
-                {
-                    _mm_store_ps(outptr, _mm512_extractf32x4_ps(_sum0, 0));
-                    _mm_store_ps(outptr + M, _mm512_extractf32x4_ps(_sum0, 1));
-                    _mm_store_ps(outptr + M * 2, _mm512_extractf32x4_ps(_sum0, 2));
-                    _mm_store_ps(outptr + M * 3, _mm512_extractf32x4_ps(_sum0, 3));
-                    outptr += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    float sum[16];
-                    _mm512_storeu_ps(sum, _sum0);
-
-                    outptr[0] = sum[0];
-                    outptr[M] = sum[1];
-                    outptr[M * 2] = sum[2];
-                    outptr[M * 3] = sum[3];
-                    outptr[M * 4] = sum[4];
-                    outptr[M * 5] = sum[5];
-                    outptr[M * 6] = sum[6];
-                    outptr[M * 7] = sum[7];
-                    outptr[M * 8] = sum[8];
-                    outptr[M * 9] = sum[9];
-                    outptr[M * 10] = sum[10];
-                    outptr[M * 11] = sum[11];
-                    outptr[M * 12] = sum[12];
-                    outptr[M * 13] = sum[13];
-                    outptr[M * 14] = sum[14];
-                    outptr[M * 15] = sum[15];
-                    outptr += 1;
-                }
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 16;
-    nn_outch = (outch - remain_outch_start) / 8;
-#else // __AVX512F__
-    nn_outch = (outch - remain_outch_start) / 8;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __AVX512F__
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 8;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int out_elempack = top_blob.elempack;
-
-        float* outptr = top_blob.channel(p / out_elempack);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m256 _sum0 = _mm256_setzero_ps();
-                __m256 _sum1 = _mm256_setzero_ps();
-                __m256 _sum2 = _mm256_setzero_ps();
-                __m256 _sum3 = _mm256_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm256_loadu_ps(bias_data_ptr + p);
-                }
-
-#if __AVX512F__
-                const float* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8);
-#else
-                const float* kptr = weight_data_tm.channel(p / 8);
-#endif
-
-                int q = 0;
-#if __AVX512F__
-                for (; q + 15 < inch; q += 16)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    if (elempack == 16)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-
-                            __m256 _w0 = _mm256_load_ps(kptr + 8 * 0);
-                            __m256 _w1 = _mm256_load_ps(kptr + 8 * 1);
-                            __m256 _w2 = _mm256_load_ps(kptr + 8 * 2);
-                            __m256 _w3 = _mm256_load_ps(kptr + 8 * 3);
-                            __m256 _w4 = _mm256_load_ps(kptr + 8 * 4);
-                            __m256 _w5 = _mm256_load_ps(kptr + 8 * 5);
-                            __m256 _w6 = _mm256_load_ps(kptr + 8 * 6);
-                            __m256 _w7 = _mm256_load_ps(kptr + 8 * 7);
-                            __m256 _w8 = _mm256_load_ps(kptr + 8 * 8);
-                            __m256 _w9 = _mm256_load_ps(kptr + 8 * 9);
-                            __m256 _wa = _mm256_load_ps(kptr + 8 * 10);
-                            __m256 _wb = _mm256_load_ps(kptr + 8 * 11);
-                            __m256 _wc = _mm256_load_ps(kptr + 8 * 12);
-                            __m256 _wd = _mm256_load_ps(kptr + 8 * 13);
-                            __m256 _we = _mm256_load_ps(kptr + 8 * 14);
-                            __m256 _wf = _mm256_load_ps(kptr + 8 * 15);
-
-                            _sum0 = _mm256_fmadd_ps(_w0, _mm256_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_w1, _mm256_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_w2, _mm256_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_w3, _mm256_set1_ps(r0s[3]), _sum3);
-                            _sum0 = _mm256_fmadd_ps(_w4, _mm256_set1_ps(r0s[4]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_w5, _mm256_set1_ps(r0s[5]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_w6, _mm256_set1_ps(r0s[6]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_w7, _mm256_set1_ps(r0s[7]), _sum3);
-                            _sum0 = _mm256_fmadd_ps(_w8, _mm256_set1_ps(r0s[8]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_w9, _mm256_set1_ps(r0s[9]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_wa, _mm256_set1_ps(r0s[10]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_wb, _mm256_set1_ps(r0s[11]), _sum3);
-                            _sum0 = _mm256_fmadd_ps(_wc, _mm256_set1_ps(r0s[12]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_wd, _mm256_set1_ps(r0s[13]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_we, _mm256_set1_ps(r0s[14]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_wf, _mm256_set1_ps(r0s[15]), _sum3);
-
-                            kptr += 128;
-                        }
-                    }
-                    if (elempack == 8)
-                    {
-                        const float* r1 = r0 + N;
-
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-                            const float* r1s = r1 + space_ofs[k];
-
-                            __m256 _w0 = _mm256_load_ps(kptr + 8 * 0);
-                            __m256 _w1 = _mm256_load_ps(kptr + 8 * 1);
-                            __m256 _w2 = _mm256_load_ps(kptr + 8 * 2);
-                            __m256 _w3 = _mm256_load_ps(kptr + 8 * 3);
-                            __m256 _w4 = _mm256_load_ps(kptr + 8 * 4);
-                            __m256 _w5 = _mm256_load_ps(kptr + 8 * 5);
-                            __m256 _w6 = _mm256_load_ps(kptr + 8 * 6);
-                            __m256 _w7 = _mm256_load_ps(kptr + 8 * 7);
-                            __m256 _w8 = _mm256_load_ps(kptr + 8 * 8);
-                            __m256 _w9 = _mm256_load_ps(kptr + 8 * 9);
-                            __m256 _wa = _mm256_load_ps(kptr + 8 * 10);
-                            __m256 _wb = _mm256_load_ps(kptr + 8 * 11);
-                            __m256 _wc = _mm256_load_ps(kptr + 8 * 12);
-                            __m256 _wd = _mm256_load_ps(kptr + 8 * 13);
-                            __m256 _we = _mm256_load_ps(kptr + 8 * 14);
-                            __m256 _wf = _mm256_load_ps(kptr + 8 * 15);
-
-                            _sum0 = _mm256_fmadd_ps(_w0, _mm256_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_w1, _mm256_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_w2, _mm256_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_w3, _mm256_set1_ps(r0s[3]), _sum3);
-                            _sum0 = _mm256_fmadd_ps(_w4, _mm256_set1_ps(r0s[4]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_w5, _mm256_set1_ps(r0s[5]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_w6, _mm256_set1_ps(r0s[6]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_w7, _mm256_set1_ps(r0s[7]), _sum3);
-                            _sum0 = _mm256_fmadd_ps(_w8, _mm256_set1_ps(r1s[0]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_w9, _mm256_set1_ps(r1s[1]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_wa, _mm256_set1_ps(r1s[2]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_wb, _mm256_set1_ps(r1s[3]), _sum3);
-                            _sum0 = _mm256_fmadd_ps(_wc, _mm256_set1_ps(r1s[4]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_wd, _mm256_set1_ps(r1s[5]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_we, _mm256_set1_ps(r1s[6]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_wf, _mm256_set1_ps(r1s[7]), _sum3);
-
-                            kptr += 128;
-                        }
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-                            const float* r1s = r1 + space_ofs[k];
-                            const float* r2s = r2 + space_ofs[k];
-                            const float* r3s = r3 + space_ofs[k];
-
-                            __m256 _w0 = _mm256_load_ps(kptr + 8 * 0);
-                            __m256 _w1 = _mm256_load_ps(kptr + 8 * 1);
-                            __m256 _w2 = _mm256_load_ps(kptr + 8 * 2);
-                            __m256 _w3 = _mm256_load_ps(kptr + 8 * 3);
-                            __m256 _w4 = _mm256_load_ps(kptr + 8 * 4);
-                            __m256 _w5 = _mm256_load_ps(kptr + 8 * 5);
-                            __m256 _w6 = _mm256_load_ps(kptr + 8 * 6);
-                            __m256 _w7 = _mm256_load_ps(kptr + 8 * 7);
-                            __m256 _w8 = _mm256_load_ps(kptr + 8 * 8);
-                            __m256 _w9 = _mm256_load_ps(kptr + 8 * 9);
-                            __m256 _wa = _mm256_load_ps(kptr + 8 * 10);
-                            __m256 _wb = _mm256_load_ps(kptr + 8 * 11);
-                            __m256 _wc = _mm256_load_ps(kptr + 8 * 12);
-                            __m256 _wd = _mm256_load_ps(kptr + 8 * 13);
-                            __m256 _we = _mm256_load_ps(kptr + 8 * 14);
-                            __m256 _wf = _mm256_load_ps(kptr + 8 * 15);
-
-                            _sum0 = _mm256_fmadd_ps(_w0, _mm256_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_w1, _mm256_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_w2, _mm256_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_w3, _mm256_set1_ps(r0s[3]), _sum3);
-                            _sum0 = _mm256_fmadd_ps(_w4, _mm256_set1_ps(r1s[0]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_w5, _mm256_set1_ps(r1s[1]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_w6, _mm256_set1_ps(r1s[2]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_w7, _mm256_set1_ps(r1s[3]), _sum3);
-                            _sum0 = _mm256_fmadd_ps(_w8, _mm256_set1_ps(r2s[0]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_w9, _mm256_set1_ps(r2s[1]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_wa, _mm256_set1_ps(r2s[2]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_wb, _mm256_set1_ps(r2s[3]), _sum3);
-                            _sum0 = _mm256_fmadd_ps(_wc, _mm256_set1_ps(r3s[0]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_wd, _mm256_set1_ps(r3s[1]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_we, _mm256_set1_ps(r3s[2]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_wf, _mm256_set1_ps(r3s[3]), _sum3);
-
-                            kptr += 128;
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-
-                            __m256 _w0 = _mm256_load_ps(kptr + 8 * 0);
-                            __m256 _w1 = _mm256_load_ps(kptr + 8 * 1);
-                            __m256 _w2 = _mm256_load_ps(kptr + 8 * 2);
-                            __m256 _w3 = _mm256_load_ps(kptr + 8 * 3);
-                            __m256 _w4 = _mm256_load_ps(kptr + 8 * 4);
-                            __m256 _w5 = _mm256_load_ps(kptr + 8 * 5);
-                            __m256 _w6 = _mm256_load_ps(kptr + 8 * 6);
-                            __m256 _w7 = _mm256_load_ps(kptr + 8 * 7);
-                            __m256 _w8 = _mm256_load_ps(kptr + 8 * 8);
-                            __m256 _w9 = _mm256_load_ps(kptr + 8 * 9);
-                            __m256 _wa = _mm256_load_ps(kptr + 8 * 10);
-                            __m256 _wb = _mm256_load_ps(kptr + 8 * 11);
-                            __m256 _wc = _mm256_load_ps(kptr + 8 * 12);
-                            __m256 _wd = _mm256_load_ps(kptr + 8 * 13);
-                            __m256 _we = _mm256_load_ps(kptr + 8 * 14);
-                            __m256 _wf = _mm256_load_ps(kptr + 8 * 15);
-
-                            _sum0 = _mm256_fmadd_ps(_w0, _mm256_set1_ps(r0[sok]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_w1, _mm256_set1_ps(r0[sok + N]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_w2, _mm256_set1_ps(r0[sok + N * 2]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_w3, _mm256_set1_ps(r0[sok + N * 3]), _sum3);
-                            _sum0 = _mm256_fmadd_ps(_w4, _mm256_set1_ps(r0[sok + N * 4]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_w5, _mm256_set1_ps(r0[sok + N * 5]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_w6, _mm256_set1_ps(r0[sok + N * 6]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_w7, _mm256_set1_ps(r0[sok + N * 7]), _sum3);
-                            _sum0 = _mm256_fmadd_ps(_w8, _mm256_set1_ps(r0[sok + N * 8]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_w9, _mm256_set1_ps(r0[sok + N * 9]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_wa, _mm256_set1_ps(r0[sok + N * 10]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_wb, _mm256_set1_ps(r0[sok + N * 11]), _sum3);
-                            _sum0 = _mm256_fmadd_ps(_wc, _mm256_set1_ps(r0[sok + N * 12]), _sum0);
-                            _sum1 = _mm256_fmadd_ps(_wd, _mm256_set1_ps(r0[sok + N * 13]), _sum1);
-                            _sum2 = _mm256_fmadd_ps(_we, _mm256_set1_ps(r0[sok + N * 14]), _sum2);
-                            _sum3 = _mm256_fmadd_ps(_wf, _mm256_set1_ps(r0[sok + N * 15]), _sum3);
-
-                            kptr += 128;
-                        }
-                    }
-                }
-#endif // __AVX512F__
-                for (; q + 7 < inch; q += 8)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    if (elempack == 8)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-
-                            __m256 _w0 = _mm256_load_ps(kptr);
-                            __m256 _w1 = _mm256_load_ps(kptr + 8);
-                            __m256 _w2 = _mm256_load_ps(kptr + 16);
-                            __m256 _w3 = _mm256_load_ps(kptr + 24);
-                            __m256 _w4 = _mm256_load_ps(kptr + 32);
-                            __m256 _w5 = _mm256_load_ps(kptr + 40);
-                            __m256 _w6 = _mm256_load_ps(kptr + 48);
-                            __m256 _w7 = _mm256_load_ps(kptr + 56);
-
-                            _sum0 = _mm256_comp_fmadd_ps(_w0, _mm256_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm256_comp_fmadd_ps(_w1, _mm256_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm256_comp_fmadd_ps(_w2, _mm256_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm256_comp_fmadd_ps(_w3, _mm256_set1_ps(r0s[3]), _sum3);
-                            _sum0 = _mm256_comp_fmadd_ps(_w4, _mm256_set1_ps(r0s[4]), _sum0);
-                            _sum1 = _mm256_comp_fmadd_ps(_w5, _mm256_set1_ps(r0s[5]), _sum1);
-                            _sum2 = _mm256_comp_fmadd_ps(_w6, _mm256_set1_ps(r0s[6]), _sum2);
-                            _sum3 = _mm256_comp_fmadd_ps(_w7, _mm256_set1_ps(r0s[7]), _sum3);
-
-                            kptr += 64;
-                        }
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-                            const float* r1s = r1 + space_ofs[k];
-
-                            __m256 _w0 = _mm256_load_ps(kptr);
-                            __m256 _w1 = _mm256_load_ps(kptr + 8);
-                            __m256 _w2 = _mm256_load_ps(kptr + 16);
-                            __m256 _w3 = _mm256_load_ps(kptr + 24);
-                            __m256 _w4 = _mm256_load_ps(kptr + 32);
-                            __m256 _w5 = _mm256_load_ps(kptr + 40);
-                            __m256 _w6 = _mm256_load_ps(kptr + 48);
-                            __m256 _w7 = _mm256_load_ps(kptr + 56);
-
-                            _sum0 = _mm256_comp_fmadd_ps(_w0, _mm256_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm256_comp_fmadd_ps(_w1, _mm256_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm256_comp_fmadd_ps(_w2, _mm256_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm256_comp_fmadd_ps(_w3, _mm256_set1_ps(r0s[3]), _sum3);
-                            _sum0 = _mm256_comp_fmadd_ps(_w4, _mm256_set1_ps(r1s[0]), _sum0);
-                            _sum1 = _mm256_comp_fmadd_ps(_w5, _mm256_set1_ps(r1s[1]), _sum1);
-                            _sum2 = _mm256_comp_fmadd_ps(_w6, _mm256_set1_ps(r1s[2]), _sum2);
-                            _sum3 = _mm256_comp_fmadd_ps(_w7, _mm256_set1_ps(r1s[3]), _sum3);
-
-                            kptr += 64;
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-
-                            __m256 _w0 = _mm256_load_ps(kptr);
-                            __m256 _w1 = _mm256_load_ps(kptr + 8);
-                            __m256 _w2 = _mm256_load_ps(kptr + 16);
-                            __m256 _w3 = _mm256_load_ps(kptr + 24);
-                            __m256 _w4 = _mm256_load_ps(kptr + 32);
-                            __m256 _w5 = _mm256_load_ps(kptr + 40);
-                            __m256 _w6 = _mm256_load_ps(kptr + 48);
-                            __m256 _w7 = _mm256_load_ps(kptr + 56);
-
-                            _sum0 = _mm256_comp_fmadd_ps(_w0, _mm256_set1_ps(r0[sok]), _sum0);
-                            _sum1 = _mm256_comp_fmadd_ps(_w1, _mm256_set1_ps(r0[sok + N]), _sum1);
-                            _sum2 = _mm256_comp_fmadd_ps(_w2, _mm256_set1_ps(r0[sok + N * 2]), _sum2);
-                            _sum3 = _mm256_comp_fmadd_ps(_w3, _mm256_set1_ps(r0[sok + N * 3]), _sum3);
-                            _sum0 = _mm256_comp_fmadd_ps(_w4, _mm256_set1_ps(r0[sok + N * 4]), _sum0);
-                            _sum1 = _mm256_comp_fmadd_ps(_w5, _mm256_set1_ps(r0[sok + N * 5]), _sum1);
-                            _sum2 = _mm256_comp_fmadd_ps(_w6, _mm256_set1_ps(r0[sok + N * 6]), _sum2);
-                            _sum3 = _mm256_comp_fmadd_ps(_w7, _mm256_set1_ps(r0[sok + N * 7]), _sum3);
-
-                            kptr += 64;
-                        }
-                    }
-                }
-                for (; q + 3 < inch; q += 4)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    if (elempack == 4)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-
-                            __m256 _w0 = _mm256_load_ps(kptr);
-                            __m256 _w1 = _mm256_load_ps(kptr + 8);
-                            __m256 _w2 = _mm256_load_ps(kptr + 16);
-                            __m256 _w3 = _mm256_load_ps(kptr + 24);
-
-                            _sum0 = _mm256_comp_fmadd_ps(_w0, _mm256_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm256_comp_fmadd_ps(_w1, _mm256_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm256_comp_fmadd_ps(_w2, _mm256_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm256_comp_fmadd_ps(_w3, _mm256_set1_ps(r0s[3]), _sum3);
-
-                            kptr += 32;
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-
-                            __m256 _w0 = _mm256_load_ps(kptr);
-                            __m256 _w1 = _mm256_load_ps(kptr + 8);
-                            __m256 _w2 = _mm256_load_ps(kptr + 16);
-                            __m256 _w3 = _mm256_load_ps(kptr + 24);
-
-                            _sum0 = _mm256_comp_fmadd_ps(_w0, _mm256_set1_ps(r0[sok]), _sum0);
-                            _sum1 = _mm256_comp_fmadd_ps(_w1, _mm256_set1_ps(r0[sok + N]), _sum1);
-                            _sum2 = _mm256_comp_fmadd_ps(_w2, _mm256_set1_ps(r0[sok + N * 2]), _sum2);
-                            _sum3 = _mm256_comp_fmadd_ps(_w3, _mm256_set1_ps(r0[sok + N * 3]), _sum3);
-
-                            kptr += 32;
-                        }
-                    }
-                }
-                for (; q + 1 < inch; q += 2)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    // if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-
-                            __m256 _w0 = _mm256_load_ps(kptr);
-                            __m256 _w1 = _mm256_load_ps(kptr + 8);
-
-                            _sum0 = _mm256_comp_fmadd_ps(_w0, _mm256_set1_ps(r0[sok]), _sum0);
-                            _sum1 = _mm256_comp_fmadd_ps(_w1, _mm256_set1_ps(r0[sok + N]), _sum1);
-
-                            kptr += 16;
-                        }
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    // if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            __m256 _val = _mm256_set1_ps(r0[space_ofs[k]]);
-                            __m256 _w = _mm256_load_ps(kptr);
-                            _sum0 = _mm256_comp_fmadd_ps(_val, _w, _sum0);
-
-                            kptr += 8;
-                        }
-                    }
-                }
-
-                _sum0 = _mm256_add_ps(_sum0, _sum1);
-                _sum2 = _mm256_add_ps(_sum2, _sum3);
-                _sum0 = _mm256_add_ps(_sum0, _sum2);
-
-                _sum0 = activation_avx(_sum0, activation_type, activation_params);
-
-                if (out_elempack == 8)
-                {
-                    _mm256_store_ps(outptr, _sum0);
-                    outptr += 8;
-                }
-                if (out_elempack == 4)
-                {
-                    _mm_store_ps(outptr, _mm256_extractf128_ps(_sum0, 0));
-                    _mm_store_ps(outptr + M, _mm256_extractf128_ps(_sum0, 1));
-                    outptr += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    float sum[8];
-                    _mm256_storeu_ps(sum, _sum0);
-
-                    outptr[0] = sum[0];
-                    outptr[M] = sum[1];
-                    outptr[M * 2] = sum[2];
-                    outptr[M * 3] = sum[3];
-                    outptr[M * 4] = sum[4];
-                    outptr[M * 5] = sum[5];
-                    outptr[M * 6] = sum[6];
-                    outptr[M * 7] = sum[7];
-                    outptr += 1;
-                }
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 8;
-    nn_outch = (outch - remain_outch_start) / 4;
-#else // __AVX__
-    nn_outch = (outch - remain_outch_start) / 4;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __AVX__
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 4;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int out_elempack = top_blob.elempack;
-
-        float* outptr = top_blob.channel(p / out_elempack);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m128 _sum0 = _mm_setzero_ps();
-                __m128 _sum1 = _mm_setzero_ps();
-                __m128 _sum2 = _mm_setzero_ps();
-                __m128 _sum3 = _mm_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm_loadu_ps(bias_data_ptr + p);
-                }
-
-#if __AVX512F__
-                const float* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8 + (p % 8) / 4);
-#elif __AVX__
-                const float* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4);
-#else
-                const float* kptr = weight_data_tm.channel(p / 4);
-#endif
-
-                int q = 0;
-#if __AVX__
-#if __AVX512F__
-                for (; q + 15 < inch; q += 16)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    if (elempack == 16)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-
-                            __m128 _w0 = _mm_load_ps(kptr + 4 * 0);
-                            __m128 _w1 = _mm_load_ps(kptr + 4 * 1);
-                            __m128 _w2 = _mm_load_ps(kptr + 4 * 2);
-                            __m128 _w3 = _mm_load_ps(kptr + 4 * 3);
-                            __m128 _w4 = _mm_load_ps(kptr + 4 * 4);
-                            __m128 _w5 = _mm_load_ps(kptr + 4 * 5);
-                            __m128 _w6 = _mm_load_ps(kptr + 4 * 6);
-                            __m128 _w7 = _mm_load_ps(kptr + 4 * 7);
-                            __m128 _w8 = _mm_load_ps(kptr + 4 * 8);
-                            __m128 _w9 = _mm_load_ps(kptr + 4 * 9);
-                            __m128 _wa = _mm_load_ps(kptr + 4 * 10);
-                            __m128 _wb = _mm_load_ps(kptr + 4 * 11);
-                            __m128 _wc = _mm_load_ps(kptr + 4 * 12);
-                            __m128 _wd = _mm_load_ps(kptr + 4 * 13);
-                            __m128 _we = _mm_load_ps(kptr + 4 * 14);
-                            __m128 _wf = _mm_load_ps(kptr + 4 * 15);
-
-                            _sum0 = _mm_fmadd_ps(_w0, _mm_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_w1, _mm_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_w2, _mm_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_w3, _mm_set1_ps(r0s[3]), _sum3);
-                            _sum0 = _mm_fmadd_ps(_w4, _mm_set1_ps(r0s[4]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_w5, _mm_set1_ps(r0s[5]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_w6, _mm_set1_ps(r0s[6]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_w7, _mm_set1_ps(r0s[7]), _sum3);
-                            _sum0 = _mm_fmadd_ps(_w8, _mm_set1_ps(r0s[8]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_w9, _mm_set1_ps(r0s[9]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_wa, _mm_set1_ps(r0s[10]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_wb, _mm_set1_ps(r0s[11]), _sum3);
-                            _sum0 = _mm_fmadd_ps(_wc, _mm_set1_ps(r0s[12]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_wd, _mm_set1_ps(r0s[13]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_we, _mm_set1_ps(r0s[14]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_wf, _mm_set1_ps(r0s[15]), _sum3);
-
-                            kptr += 64;
-                        }
-                    }
-                    if (elempack == 8)
-                    {
-                        const float* r1 = r0 + N;
-
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-                            const float* r1s = r1 + space_ofs[k];
-
-                            __m128 _w0 = _mm_load_ps(kptr + 4 * 0);
-                            __m128 _w1 = _mm_load_ps(kptr + 4 * 1);
-                            __m128 _w2 = _mm_load_ps(kptr + 4 * 2);
-                            __m128 _w3 = _mm_load_ps(kptr + 4 * 3);
-                            __m128 _w4 = _mm_load_ps(kptr + 4 * 4);
-                            __m128 _w5 = _mm_load_ps(kptr + 4 * 5);
-                            __m128 _w6 = _mm_load_ps(kptr + 4 * 6);
-                            __m128 _w7 = _mm_load_ps(kptr + 4 * 7);
-                            __m128 _w8 = _mm_load_ps(kptr + 4 * 8);
-                            __m128 _w9 = _mm_load_ps(kptr + 4 * 9);
-                            __m128 _wa = _mm_load_ps(kptr + 4 * 10);
-                            __m128 _wb = _mm_load_ps(kptr + 4 * 11);
-                            __m128 _wc = _mm_load_ps(kptr + 4 * 12);
-                            __m128 _wd = _mm_load_ps(kptr + 4 * 13);
-                            __m128 _we = _mm_load_ps(kptr + 4 * 14);
-                            __m128 _wf = _mm_load_ps(kptr + 4 * 15);
-
-                            _sum0 = _mm_fmadd_ps(_w0, _mm_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_w1, _mm_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_w2, _mm_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_w3, _mm_set1_ps(r0s[3]), _sum3);
-                            _sum0 = _mm_fmadd_ps(_w4, _mm_set1_ps(r0s[4]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_w5, _mm_set1_ps(r0s[5]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_w6, _mm_set1_ps(r0s[6]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_w7, _mm_set1_ps(r0s[7]), _sum3);
-                            _sum0 = _mm_fmadd_ps(_w8, _mm_set1_ps(r1s[0]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_w9, _mm_set1_ps(r1s[1]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_wa, _mm_set1_ps(r1s[2]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_wb, _mm_set1_ps(r1s[3]), _sum3);
-                            _sum0 = _mm_fmadd_ps(_wc, _mm_set1_ps(r1s[4]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_wd, _mm_set1_ps(r1s[5]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_we, _mm_set1_ps(r1s[6]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_wf, _mm_set1_ps(r1s[7]), _sum3);
-
-                            kptr += 64;
-                        }
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-                            const float* r1s = r1 + space_ofs[k];
-                            const float* r2s = r2 + space_ofs[k];
-                            const float* r3s = r3 + space_ofs[k];
-
-                            __m128 _w0 = _mm_load_ps(kptr + 4 * 0);
-                            __m128 _w1 = _mm_load_ps(kptr + 4 * 1);
-                            __m128 _w2 = _mm_load_ps(kptr + 4 * 2);
-                            __m128 _w3 = _mm_load_ps(kptr + 4 * 3);
-                            __m128 _w4 = _mm_load_ps(kptr + 4 * 4);
-                            __m128 _w5 = _mm_load_ps(kptr + 4 * 5);
-                            __m128 _w6 = _mm_load_ps(kptr + 4 * 6);
-                            __m128 _w7 = _mm_load_ps(kptr + 4 * 7);
-                            __m128 _w8 = _mm_load_ps(kptr + 4 * 8);
-                            __m128 _w9 = _mm_load_ps(kptr + 4 * 9);
-                            __m128 _wa = _mm_load_ps(kptr + 4 * 10);
-                            __m128 _wb = _mm_load_ps(kptr + 4 * 11);
-                            __m128 _wc = _mm_load_ps(kptr + 4 * 12);
-                            __m128 _wd = _mm_load_ps(kptr + 4 * 13);
-                            __m128 _we = _mm_load_ps(kptr + 4 * 14);
-                            __m128 _wf = _mm_load_ps(kptr + 4 * 15);
-
-                            _sum0 = _mm_fmadd_ps(_w0, _mm_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_w1, _mm_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_w2, _mm_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_w3, _mm_set1_ps(r0s[3]), _sum3);
-                            _sum0 = _mm_fmadd_ps(_w4, _mm_set1_ps(r1s[0]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_w5, _mm_set1_ps(r1s[1]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_w6, _mm_set1_ps(r1s[2]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_w7, _mm_set1_ps(r1s[3]), _sum3);
-                            _sum0 = _mm_fmadd_ps(_w8, _mm_set1_ps(r2s[0]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_w9, _mm_set1_ps(r2s[1]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_wa, _mm_set1_ps(r2s[2]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_wb, _mm_set1_ps(r2s[3]), _sum3);
-                            _sum0 = _mm_fmadd_ps(_wc, _mm_set1_ps(r3s[0]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_wd, _mm_set1_ps(r3s[1]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_we, _mm_set1_ps(r3s[2]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_wf, _mm_set1_ps(r3s[3]), _sum3);
-
-                            kptr += 64;
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-
-                            __m128 _w0 = _mm_load_ps(kptr + 4 * 0);
-                            __m128 _w1 = _mm_load_ps(kptr + 4 * 1);
-                            __m128 _w2 = _mm_load_ps(kptr + 4 * 2);
-                            __m128 _w3 = _mm_load_ps(kptr + 4 * 3);
-                            __m128 _w4 = _mm_load_ps(kptr + 4 * 4);
-                            __m128 _w5 = _mm_load_ps(kptr + 4 * 5);
-                            __m128 _w6 = _mm_load_ps(kptr + 4 * 6);
-                            __m128 _w7 = _mm_load_ps(kptr + 4 * 7);
-                            __m128 _w8 = _mm_load_ps(kptr + 4 * 8);
-                            __m128 _w9 = _mm_load_ps(kptr + 4 * 9);
-                            __m128 _wa = _mm_load_ps(kptr + 4 * 10);
-                            __m128 _wb = _mm_load_ps(kptr + 4 * 11);
-                            __m128 _wc = _mm_load_ps(kptr + 4 * 12);
-                            __m128 _wd = _mm_load_ps(kptr + 4 * 13);
-                            __m128 _we = _mm_load_ps(kptr + 4 * 14);
-                            __m128 _wf = _mm_load_ps(kptr + 4 * 15);
-
-                            _sum0 = _mm_fmadd_ps(_w0, _mm_set1_ps(r0[sok]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_w1, _mm_set1_ps(r0[sok + N]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_w2, _mm_set1_ps(r0[sok + N * 2]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_w3, _mm_set1_ps(r0[sok + N * 3]), _sum3);
-                            _sum0 = _mm_fmadd_ps(_w4, _mm_set1_ps(r0[sok + N * 4]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_w5, _mm_set1_ps(r0[sok + N * 5]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_w6, _mm_set1_ps(r0[sok + N * 6]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_w7, _mm_set1_ps(r0[sok + N * 7]), _sum3);
-                            _sum0 = _mm_fmadd_ps(_w8, _mm_set1_ps(r0[sok + N * 8]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_w9, _mm_set1_ps(r0[sok + N * 9]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_wa, _mm_set1_ps(r0[sok + N * 10]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_wb, _mm_set1_ps(r0[sok + N * 11]), _sum3);
-                            _sum0 = _mm_fmadd_ps(_wc, _mm_set1_ps(r0[sok + N * 12]), _sum0);
-                            _sum1 = _mm_fmadd_ps(_wd, _mm_set1_ps(r0[sok + N * 13]), _sum1);
-                            _sum2 = _mm_fmadd_ps(_we, _mm_set1_ps(r0[sok + N * 14]), _sum2);
-                            _sum3 = _mm_fmadd_ps(_wf, _mm_set1_ps(r0[sok + N * 15]), _sum3);
-
-                            kptr += 64;
-                        }
-                    }
-                }
-#endif // __AVX512F__
-                for (; q + 7 < inch; q += 8)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    if (elempack == 8)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-
-                            __m128 _w0 = _mm_load_ps(kptr);
-                            __m128 _w1 = _mm_load_ps(kptr + 4);
-                            __m128 _w2 = _mm_load_ps(kptr + 8);
-                            __m128 _w3 = _mm_load_ps(kptr + 12);
-                            __m128 _w4 = _mm_load_ps(kptr + 16);
-                            __m128 _w5 = _mm_load_ps(kptr + 20);
-                            __m128 _w6 = _mm_load_ps(kptr + 24);
-                            __m128 _w7 = _mm_load_ps(kptr + 28);
-
-                            _sum0 = _mm_comp_fmadd_ps(_w0, _mm_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm_comp_fmadd_ps(_w1, _mm_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm_comp_fmadd_ps(_w2, _mm_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm_comp_fmadd_ps(_w3, _mm_set1_ps(r0s[3]), _sum3);
-                            _sum0 = _mm_comp_fmadd_ps(_w4, _mm_set1_ps(r0s[4]), _sum0);
-                            _sum1 = _mm_comp_fmadd_ps(_w5, _mm_set1_ps(r0s[5]), _sum1);
-                            _sum2 = _mm_comp_fmadd_ps(_w6, _mm_set1_ps(r0s[6]), _sum2);
-                            _sum3 = _mm_comp_fmadd_ps(_w7, _mm_set1_ps(r0s[7]), _sum3);
-
-                            kptr += 32;
-                        }
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-                            const float* r1s = r1 + space_ofs[k];
-
-                            __m128 _w0 = _mm_load_ps(kptr);
-                            __m128 _w1 = _mm_load_ps(kptr + 4);
-                            __m128 _w2 = _mm_load_ps(kptr + 8);
-                            __m128 _w3 = _mm_load_ps(kptr + 12);
-                            __m128 _w4 = _mm_load_ps(kptr + 16);
-                            __m128 _w5 = _mm_load_ps(kptr + 20);
-                            __m128 _w6 = _mm_load_ps(kptr + 24);
-                            __m128 _w7 = _mm_load_ps(kptr + 28);
-
-                            _sum0 = _mm_comp_fmadd_ps(_w0, _mm_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm_comp_fmadd_ps(_w1, _mm_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm_comp_fmadd_ps(_w2, _mm_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm_comp_fmadd_ps(_w3, _mm_set1_ps(r0s[3]), _sum3);
-                            _sum0 = _mm_comp_fmadd_ps(_w4, _mm_set1_ps(r1s[0]), _sum0);
-                            _sum1 = _mm_comp_fmadd_ps(_w5, _mm_set1_ps(r1s[1]), _sum1);
-                            _sum2 = _mm_comp_fmadd_ps(_w6, _mm_set1_ps(r1s[2]), _sum2);
-                            _sum3 = _mm_comp_fmadd_ps(_w7, _mm_set1_ps(r1s[3]), _sum3);
-
-                            kptr += 32;
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-
-                            __m128 _w0 = _mm_load_ps(kptr);
-                            __m128 _w1 = _mm_load_ps(kptr + 4);
-                            __m128 _w2 = _mm_load_ps(kptr + 8);
-                            __m128 _w3 = _mm_load_ps(kptr + 12);
-                            __m128 _w4 = _mm_load_ps(kptr + 16);
-                            __m128 _w5 = _mm_load_ps(kptr + 20);
-                            __m128 _w6 = _mm_load_ps(kptr + 24);
-                            __m128 _w7 = _mm_load_ps(kptr + 28);
-
-                            _sum0 = _mm_comp_fmadd_ps(_w0, _mm_set1_ps(r0[sok]), _sum0);
-                            _sum1 = _mm_comp_fmadd_ps(_w1, _mm_set1_ps(r0[sok + N]), _sum1);
-                            _sum2 = _mm_comp_fmadd_ps(_w2, _mm_set1_ps(r0[sok + N * 2]), _sum2);
-                            _sum3 = _mm_comp_fmadd_ps(_w3, _mm_set1_ps(r0[sok + N * 3]), _sum3);
-                            _sum0 = _mm_comp_fmadd_ps(_w4, _mm_set1_ps(r0[sok + N * 4]), _sum0);
-                            _sum1 = _mm_comp_fmadd_ps(_w5, _mm_set1_ps(r0[sok + N * 5]), _sum1);
-                            _sum2 = _mm_comp_fmadd_ps(_w6, _mm_set1_ps(r0[sok + N * 6]), _sum2);
-                            _sum3 = _mm_comp_fmadd_ps(_w7, _mm_set1_ps(r0[sok + N * 7]), _sum3);
-
-                            kptr += 32;
-                        }
-                    }
-                }
-#endif // __AVX__
-                for (; q + 3 < inch; q += 4)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    if (elempack == 4)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const float* r0s = r0 + space_ofs[k];
-
-                            __m128 _w0 = _mm_load_ps(kptr);
-                            __m128 _w1 = _mm_load_ps(kptr + 4);
-                            __m128 _w2 = _mm_load_ps(kptr + 8);
-                            __m128 _w3 = _mm_load_ps(kptr + 12);
-
-                            _sum0 = _mm_comp_fmadd_ps(_w0, _mm_set1_ps(r0s[0]), _sum0);
-                            _sum1 = _mm_comp_fmadd_ps(_w1, _mm_set1_ps(r0s[1]), _sum1);
-                            _sum2 = _mm_comp_fmadd_ps(_w2, _mm_set1_ps(r0s[2]), _sum2);
-                            _sum3 = _mm_comp_fmadd_ps(_w3, _mm_set1_ps(r0s[3]), _sum3);
-
-                            kptr += 16;
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-
-                            __m128 _w0 = _mm_load_ps(kptr);
-                            __m128 _w1 = _mm_load_ps(kptr + 4);
-                            __m128 _w2 = _mm_load_ps(kptr + 8);
-                            __m128 _w3 = _mm_load_ps(kptr + 12);
-
-                            _sum0 = _mm_comp_fmadd_ps(_w0, _mm_set1_ps(r0[sok]), _sum0);
-                            _sum1 = _mm_comp_fmadd_ps(_w1, _mm_set1_ps(r0[sok + N]), _sum1);
-                            _sum2 = _mm_comp_fmadd_ps(_w2, _mm_set1_ps(r0[sok + N * 2]), _sum2);
-                            _sum3 = _mm_comp_fmadd_ps(_w3, _mm_set1_ps(r0[sok + N * 3]), _sum3);
-
-                            kptr += 16;
-                        }
-                    }
-                }
-                for (; q + 1 < inch; q += 2)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    // if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-
-                            __m128 _w0 = _mm_load_ps(kptr);
-                            __m128 _w1 = _mm_load_ps(kptr + 4);
-
-                            _sum0 = _mm_comp_fmadd_ps(_w0, _mm_set1_ps(r0[sok]), _sum0);
-                            _sum1 = _mm_comp_fmadd_ps(_w1, _mm_set1_ps(r0[sok + N]), _sum1);
-
-                            kptr += 8;
-                        }
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    // if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            __m128 _val = _mm_set1_ps(r0[space_ofs[k]]);
-                            __m128 _w = _mm_load_ps(kptr);
-                            _sum0 = _mm_comp_fmadd_ps(_val, _w, _sum0);
-
-                            kptr += 4;
-                        }
-                    }
-                }
-
-                _sum0 = _mm_add_ps(_sum0, _sum1);
-                _sum2 = _mm_add_ps(_sum2, _sum3);
-                _sum0 = _mm_add_ps(_sum0, _sum2);
-
-                _sum0 = activation_sse(_sum0, activation_type, activation_params);
-
-                if (out_elempack == 4)
-                {
-                    _mm_storeu_ps(outptr, _sum0);
-                    outptr += 4;
-                }
-                if (out_elempack == 1)
-                {
-                    float sum[4];
-                    _mm_storeu_ps(sum, _sum0);
-
-                    outptr[0] = sum[0];
-                    outptr[M] = sum[1];
-                    outptr[M * 2] = sum[2];
-                    outptr[M * 3] = sum[3];
-                    outptr += 1;
-                }
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 4;
-    nn_outch = (outch - remain_outch_start) / 2;
-#else // __SSE2__
-    nn_outch = (outch - remain_outch_start) / 2;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __SSE2__
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 2;
-
-        // shadowed variable for less openmp task args
-        const int elempack = bottom_blob.elempack;
-        const int inch = bottom_blob.c * elempack;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        float* outptr0 = top_blob.channel(p);
-        float* outptr1 = top_blob.channel(p + 1);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum0 = 0.f;
-                float sum1 = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum0 = bias_data_ptr[p];
-                    sum1 = bias_data_ptr[p + 1];
-                }
-
-#if __AVX512F__
-                const float* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8 + (p % 8) / 4 + (p % 4) / 2);
-#elif __AVX__
-                const float* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2);
-#elif __SSE2__
-                const float* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2);
-#else
-                const float* kptr = weight_data_tm.channel(p / 2);
-#endif
-
-                int q = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-                __m512 _sum0_avx512 = _mm512_setzero_ps();
-                __m512 _sum1_avx512 = _mm512_setzero_ps();
-                for (; q + 15 < inch; q += 16)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    if (elempack == 16)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m512 _r0 = _mm512_load_ps(r0 + sok);
-                            __m512 _w0 = _mm512_load_ps(kptr);
-                            __m512 _w1 = _mm512_load_ps(kptr + 16);
-                            _sum0_avx512 = _mm512_fmadd_ps(_r0, _w0, _sum0_avx512);
-                            _sum1_avx512 = _mm512_fmadd_ps(_r0, _w1, _sum1_avx512);
-
-                            kptr += 32;
-                        }
-                    }
-                    if (elempack == 8)
-                    {
-                        const float* r1 = r0 + N;
-
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m512 _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + sok)), _mm256_load_ps(r1 + sok), 1);
-                            __m512 _w0 = _mm512_load_ps(kptr);
-                            __m512 _w1 = _mm512_load_ps(kptr + 16);
-                            _sum0_avx512 = _mm512_fmadd_ps(_r0, _w0, _sum0_avx512);
-                            _sum1_avx512 = _mm512_fmadd_ps(_r0, _w1, _sum1_avx512);
-
-                            kptr += 32;
-                        }
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m512 _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + sok)), _mm_load_ps(r1 + sok), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + sok)), _mm_load_ps(r3 + sok), 1), 1);
-                            __m512 _w0 = _mm512_load_ps(kptr);
-                            __m512 _w1 = _mm512_load_ps(kptr + 16);
-                            _sum0_avx512 = _mm512_fmadd_ps(_r0, _w0, _sum0_avx512);
-                            _sum1_avx512 = _mm512_fmadd_ps(_r0, _w1, _sum1_avx512);
-
-                            kptr += 32;
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m512 _r0 = _mm512_set_ps(r0[sok + N * 15], r0[sok + N * 14], r0[sok + N * 13], r0[sok + N * 12], r0[sok + N * 11], r0[sok + N * 10], r0[sok + N * 9], r0[sok + N * 8], r0[sok + N * 7], r0[sok + N * 6], r0[sok + N * 5], r0[sok + N * 4], r0[sok + N * 3], r0[sok + N * 2], r0[sok + N], r0[sok]);
-                            __m512 _w0 = _mm512_load_ps(kptr);
-                            __m512 _w1 = _mm512_load_ps(kptr + 16);
-                            _sum0_avx512 = _mm512_fmadd_ps(_r0, _w0, _sum0_avx512);
-                            _sum1_avx512 = _mm512_fmadd_ps(_r0, _w1, _sum1_avx512);
-
-                            kptr += 32;
-                        }
-                    }
-                }
-                sum0 += _mm512_comp_reduce_add_ps(_sum0_avx512);
-                sum1 += _mm512_comp_reduce_add_ps(_sum1_avx512);
-#endif // __AVX512F__
-                __m256 _sum0_avx = _mm256_setzero_ps();
-                __m256 _sum1_avx = _mm256_setzero_ps();
-                for (; q + 7 < inch; q += 8)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    if (elempack == 8)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m256 _r0 = _mm256_load_ps(r0 + sok);
-                            __m256 _w0 = _mm256_load_ps(kptr);
-                            __m256 _w1 = _mm256_load_ps(kptr + 8);
-                            _sum0_avx = _mm256_comp_fmadd_ps(_r0, _w0, _sum0_avx);
-                            _sum1_avx = _mm256_comp_fmadd_ps(_r0, _w1, _sum1_avx);
-
-                            kptr += 16;
-                        }
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m256 _r0 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + sok)), _mm_load_ps(r1 + sok), 1);
-                            __m256 _w0 = _mm256_load_ps(kptr);
-                            __m256 _w1 = _mm256_load_ps(kptr + 8);
-                            _sum0_avx = _mm256_comp_fmadd_ps(_r0, _w0, _sum0_avx);
-                            _sum1_avx = _mm256_comp_fmadd_ps(_r0, _w1, _sum1_avx);
-
-                            kptr += 16;
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m256 _r0 = _mm256_set_ps(r0[sok + N * 7], r0[sok + N * 6], r0[sok + N * 5], r0[sok + N * 4], r0[sok + N * 3], r0[sok + N * 2], r0[sok + N], r0[sok]);
-                            __m256 _w0 = _mm256_load_ps(kptr);
-                            __m256 _w1 = _mm256_load_ps(kptr + 8);
-                            _sum0_avx = _mm256_comp_fmadd_ps(_r0, _w0, _sum0_avx);
-                            _sum1_avx = _mm256_comp_fmadd_ps(_r0, _w1, _sum1_avx);
-
-                            kptr += 16;
-                        }
-                    }
-                }
-                sum0 += _mm256_reduce_add_ps(_sum0_avx);
-                sum1 += _mm256_reduce_add_ps(_sum1_avx);
-#endif // __AVX__
-                __m128 _sum0 = _mm_setzero_ps();
-                __m128 _sum1 = _mm_setzero_ps();
-                for (; q + 3 < inch; q += 4)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    if (elempack == 4)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m128 _r0 = _mm_load_ps(r0 + sok);
-                            __m128 _w0 = _mm_load_ps(kptr);
-                            __m128 _w1 = _mm_load_ps(kptr + 4);
-                            _sum0 = _mm_comp_fmadd_ps(_r0, _w0, _sum0);
-                            _sum1 = _mm_comp_fmadd_ps(_r0, _w1, _sum1);
-
-                            kptr += 8;
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m128 _r0 = _mm_set_ps(r0[sok + N * 3], r0[sok + N * 2], r0[sok + N], r0[sok]);
-                            __m128 _w0 = _mm_load_ps(kptr);
-                            __m128 _w1 = _mm_load_ps(kptr + 4);
-                            _sum0 = _mm_comp_fmadd_ps(_r0, _w0, _sum0);
-                            _sum1 = _mm_comp_fmadd_ps(_r0, _w1, _sum1);
-
-                            kptr += 8;
-                        }
-                    }
-                }
-                sum0 += _mm_reduce_add_ps(_sum0);
-                sum1 += _mm_reduce_add_ps(_sum1);
-#endif // __SSE2__
-                for (; q + 1 < inch; q += 2)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    // if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-
-                            sum0 += r0[sok] * kptr[0];
-                            sum1 += r0[sok] * kptr[1];
-                            sum0 += r0[sok + N] * kptr[2];
-                            sum1 += r0[sok + N] * kptr[3];
-
-                            kptr += 4;
-                        }
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    // if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            float val = r0[space_ofs[k]];
-                            sum0 += val * kptr[0];
-                            sum1 += val * kptr[1];
-
-                            kptr += 2;
-                        }
-                    }
-                }
-
-                sum0 = activation_ss(sum0, activation_type, activation_params);
-                sum1 = activation_ss(sum1, activation_type, activation_params);
-
-                outptr0[0] = sum0;
-                outptr1[0] = sum1;
-                outptr0 += 1;
-                outptr1 += 1;
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 2;
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-#if __AVX512F__
-                const float* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#elif __AVX__
-                const float* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#elif __SSE2__
-                const float* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2 + p % 2);
-#else
-                const float* kptr = weight_data_tm.channel(p / 2 + p % 2);
-#endif
-
-                int q = 0;
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-                __m512 _sum_avx512 = _mm512_setzero_ps();
-                for (; q + 15 < inch; q += 16)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    if (elempack == 16)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m512 _r0 = _mm512_load_ps(r0 + sok);
-                            __m512 _w = _mm512_load_ps(kptr);
-                            _sum_avx512 = _mm512_fmadd_ps(_r0, _w, _sum_avx512);
-
-                            kptr += 16;
-                        }
-                    }
-                    if (elempack == 8)
-                    {
-                        const float* r1 = r0 + N;
-
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m512 _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_load_ps(r0 + sok)), _mm256_load_ps(r1 + sok), 1);
-                            __m512 _w = _mm512_load_ps(kptr);
-                            _sum_avx512 = _mm512_fmadd_ps(_r0, _w, _sum_avx512);
-
-                            kptr += 16;
-                        }
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-                        const float* r2 = r0 + N * 2;
-                        const float* r3 = r0 + N * 3;
-
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m512 _r0 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + sok)), _mm_load_ps(r1 + sok), 1)), _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r2 + sok)), _mm_load_ps(r3 + sok), 1), 1);
-                            __m512 _w = _mm512_load_ps(kptr);
-                            _sum_avx512 = _mm512_fmadd_ps(_r0, _w, _sum_avx512);
-
-                            kptr += 16;
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m512 _r0 = _mm512_set_ps(r0[sok + N * 15], r0[sok + N * 14], r0[sok + N * 13], r0[sok + N * 12], r0[sok + N * 11], r0[sok + N * 10], r0[sok + N * 9], r0[sok + N * 8], r0[sok + N * 7], r0[sok + N * 6], r0[sok + N * 5], r0[sok + N * 4], r0[sok + N * 3], r0[sok + N * 2], r0[sok + N], r0[sok]);
-                            __m512 _w = _mm512_load_ps(kptr);
-                            _sum_avx512 = _mm512_fmadd_ps(_r0, _w, _sum_avx512);
-
-                            kptr += 16;
-                        }
-                    }
-                }
-                sum += _mm512_comp_reduce_add_ps(_sum_avx512);
-#endif // __AVX512F__
-                __m256 _sum_avx = _mm256_setzero_ps();
-                for (; q + 7 < inch; q += 8)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    if (elempack == 8)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m256 _r0 = _mm256_load_ps(r0 + sok);
-                            __m256 _w = _mm256_load_ps(kptr);
-                            _sum_avx = _mm256_comp_fmadd_ps(_r0, _w, _sum_avx);
-
-                            kptr += 8;
-                        }
-                    }
-                    if (elempack == 4)
-                    {
-                        const float* r1 = r0 + N;
-
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m256 _r0 = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_load_ps(r0 + sok)), _mm_load_ps(r1 + sok), 1);
-                            __m256 _w = _mm256_load_ps(kptr);
-                            _sum_avx = _mm256_comp_fmadd_ps(_r0, _w, _sum_avx);
-
-                            kptr += 8;
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m256 _r0 = _mm256_set_ps(r0[sok + N * 7], r0[sok + N * 6], r0[sok + N * 5], r0[sok + N * 4], r0[sok + N * 3], r0[sok + N * 2], r0[sok + N], r0[sok]);
-                            __m256 _w = _mm256_load_ps(kptr);
-                            _sum_avx = _mm256_comp_fmadd_ps(_r0, _w, _sum_avx);
-
-                            kptr += 8;
-                        }
-                    }
-                }
-                sum += _mm256_reduce_add_ps(_sum_avx);
-#endif // __AVX__
-                __m128 _sum = _mm_setzero_ps();
-                for (; q + 3 < inch; q += 4)
-                {
-                    const float* r0 = bottom_blob.channel(q / elempack).row(i * stride_h) + j * stride_w * elempack;
-
-                    if (elempack == 4)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m128 _r0 = _mm_load_ps(r0 + sok);
-                            __m128 _w = _mm_load_ps(kptr);
-                            _sum = _mm_comp_fmadd_ps(_r0, _w, _sum);
-
-                            kptr += 4;
-                        }
-                    }
-                    if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-                            __m128 _r0 = _mm_set_ps(r0[sok + N * 3], r0[sok + N * 2], r0[sok + N], r0[sok]);
-                            __m128 _w = _mm_load_ps(kptr);
-                            _sum = _mm_comp_fmadd_ps(_r0, _w, _sum);
-
-                            kptr += 4;
-                        }
-                    }
-                }
-                sum += _mm_reduce_add_ps(_sum);
-#endif // __SSE2__
-                for (; q + 1 < inch; q += 2)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    // if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            const int sok = space_ofs[k];
-
-                            sum += r0[sok] * kptr[0];
-                            sum += r0[sok + N] * kptr[1];
-
-                            kptr += 2;
-                        }
-                    }
-                }
-                for (; q < inch; q++)
-                {
-                    const float* r0 = bottom_blob.channel(q).row(i * stride_h) + j * stride_w;
-
-                    // if (elempack == 1)
-                    {
-                        for (int k = 0; k < maxk; k++)
-                        {
-                            float val = r0[space_ofs[k]];
-                            sum += val * kptr[0];
-
-                            kptr += 1;
-                        }
-                    }
-                }
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[0] = sum;
-                outptr += 1;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_packed_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_packed_int8.h
deleted file mode 100644
index 46c03f0c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_packed_int8.h
+++ /dev/null
@@ -1,5386 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-void convolution_transform_kernel_packed_int8_avx2(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h);
-#endif
-
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
-#if NCNN_RUNTIME_CPU && NCNN_AVX512VNNI && __AVX512F__ && !__AVX512VNNI__
-void convolution_packed_int8_avx512vnni(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt);
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__
-void convolution_packed_int8_avxvnni(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt);
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-void convolution_packed_int8_avx2(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt);
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__
-void convolution_packed_int8_xop(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt);
-#endif
-#endif
-
-static void convolution_transform_kernel_packed_int8(const Mat& kernel, Mat& kernel_tm, int inch, int outch, int kernel_w, int kernel_h)
-{
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-    if (ncnn::cpu_support_x86_avx2())
-    {
-        convolution_transform_kernel_packed_int8_avx2(kernel, kernel_tm, inch, outch, kernel_w, kernel_h);
-        return;
-    }
-#endif
-
-    const int maxk = kernel_w * kernel_h;
-
-    // src = kw-kh-inch-outch
-    // dst = pb-pa-kw-kh-inch/pa-outch/pb
-
-    // clang-format off
-    // *INDENT-OFF*
-#if __SSE2__
-#if __AVX2__
-#if __AVX512F__
-    if (outch >= 16)
-    {
-        if (inch >= 16)
-            kernel_tm.create(maxk, inch / 16 + (inch % 16) / 8 + (inch % 8) / 2 + inch % 2, outch / 16 + (outch % 16) / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)256u, 256);
-        else if (inch >= 8)
-            kernel_tm.create(maxk, inch / 8 + (inch % 8) / 2 + inch % 2, outch / 16 + (outch % 16) / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)128u, 128);
-        else if (inch >= 2)
-            kernel_tm.create(maxk, inch / 2 + inch % 2, outch / 16 + (outch % 16) / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)32u, 32);
-        else
-            kernel_tm.create(maxk, inch, outch / 16 + (outch % 16) / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)16u, 16);
-    }
-    else
-#endif // __AVX512F__
-    if (outch >= 8)
-    {
-#if __AVX512F__
-        if (inch >= 16)
-            kernel_tm.create(maxk, inch / 16 + (inch % 16) / 8 + (inch % 8) / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)128u, 128);
-        else
-#endif // __AVX512F__
-        if (inch >= 8)
-            kernel_tm.create(maxk, inch / 8 + (inch % 8) / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)64u, 64);
-        else if (inch >= 2)
-            kernel_tm.create(maxk, inch / 2 + inch % 2, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)16u, 16);
-        else
-            kernel_tm.create(maxk, inch, outch / 8 + (outch % 8) / 4 + (outch % 4) / 2 + outch % 2, (size_t)8u, 8);
-    }
-    else
-#endif // __AVX2__
-    if (outch >= 4)
-    {
-#if __AVX512F__
-        if (inch >= 16)
-            kernel_tm.create(maxk, inch / 16 + (inch % 16) / 8 + (inch % 8) / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2, (size_t)64u, 64);
-        else
-#endif // __AVX512F__
-        if (inch >= 8)
-            kernel_tm.create(maxk, inch / 8 + (inch % 8) / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2, (size_t)32u, 32);
-        else if (inch >= 2)
-            kernel_tm.create(maxk, inch / 2 + inch % 2, outch / 4 + (outch % 4) / 2 + outch % 2, (size_t)8u, 8);
-        else
-            kernel_tm.create(maxk, inch, outch / 4 + (outch % 4) / 2 + outch % 2, (size_t)4u, 4);
-    }
-    else
-#endif // __SSE2__
-    if (outch >= 2)
-    {
-#if __SSE2__
-#if __AVX512F__
-        if (inch >= 16)
-            kernel_tm.create(maxk, inch / 16 + (inch % 16) / 8 + (inch % 8) / 2 + inch % 2, outch / 2 + outch % 2, (size_t)32u, 32);
-        else
-#endif // __AVX512F__
-        if (inch >= 8)
-            kernel_tm.create(maxk, inch / 8 + (inch % 8) / 2 + inch % 2, outch / 2 + outch % 2, (size_t)16u, 16);
-        else
-#endif // __SSE2__
-        if (inch >= 2)
-            kernel_tm.create(maxk, inch / 2 + inch % 2, outch / 2 + outch % 2, (size_t)4u, 4);
-        else
-            kernel_tm.create(maxk, inch, outch / 2 + outch % 2, (size_t)2u, 2);
-    }
-    else
-    {
-#if __SSE2__
-#if __AVX512F__
-        if (inch >= 16)
-            kernel_tm.create(maxk, inch / 16 + (inch % 16) / 8 + (inch % 8) / 2 + inch % 2, outch, (size_t)16u, 16);
-        else
-#endif // __AVX512F__
-        if (inch >= 8)
-            kernel_tm.create(maxk, inch / 8 + (inch % 8) / 2 + inch % 2, outch, (size_t)8u, 8);
-        else
-#endif // __SSE2__
-        if (inch >= 2)
-            kernel_tm.create(maxk, inch / 2 + inch % 2, outch, (size_t)2u, 2);
-        else
-            kernel_tm.create(maxk, inch, outch, (size_t)1u, 1);
-    }
-    // *INDENT-ON*
-    // clang-format on
-
-    int q = 0;
-#if __SSE2__
-#if __AVX2__
-#if __AVX512F__
-    for (; q + 15 < outch; q += 16)
-    {
-        const signed char* kptr0 = (const signed char*)kernel + q * inch * maxk;
-        const signed char* kptr1 = (const signed char*)kernel + (q + 1) * inch * maxk;
-        const signed char* kptr2 = (const signed char*)kernel + (q + 2) * inch * maxk;
-        const signed char* kptr3 = (const signed char*)kernel + (q + 3) * inch * maxk;
-        const signed char* kptr4 = (const signed char*)kernel + (q + 4) * inch * maxk;
-        const signed char* kptr5 = (const signed char*)kernel + (q + 5) * inch * maxk;
-        const signed char* kptr6 = (const signed char*)kernel + (q + 6) * inch * maxk;
-        const signed char* kptr7 = (const signed char*)kernel + (q + 7) * inch * maxk;
-        const signed char* kptr8 = (const signed char*)kernel + (q + 8) * inch * maxk;
-        const signed char* kptr9 = (const signed char*)kernel + (q + 9) * inch * maxk;
-        const signed char* kptra = (const signed char*)kernel + (q + 10) * inch * maxk;
-        const signed char* kptrb = (const signed char*)kernel + (q + 11) * inch * maxk;
-        const signed char* kptrc = (const signed char*)kernel + (q + 12) * inch * maxk;
-        const signed char* kptrd = (const signed char*)kernel + (q + 13) * inch * maxk;
-        const signed char* kptre = (const signed char*)kernel + (q + 14) * inch * maxk;
-        const signed char* kptrf = (const signed char*)kernel + (q + 15) * inch * maxk;
-
-        signed char* g00 = kernel_tm.channel(q / 16);
-
-        int p = 0;
-        for (; p + 15 < inch; p += 16)
-        {
-            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(maxk));
-
-            for (int k = 0; k < maxk; k++)
-            {
-                __m128i _w0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr0 + k), 1));
-                __m128i _w1 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr1 + k), 1));
-                __m128i _w2 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr2 + k), 1));
-                __m128i _w3 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr3 + k), 1));
-                __m128i _w4 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr4 + k), 1));
-                __m128i _w5 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr5 + k), 1));
-                __m128i _w6 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr6 + k), 1));
-                __m128i _w7 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr7 + k), 1));
-                __m128i _w8 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr8 + k), 1));
-                __m128i _w9 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr9 + k), 1));
-                __m128i _wa = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptra + k), 1));
-                __m128i _wb = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptrb + k), 1));
-                __m128i _wc = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptrc + k), 1));
-                __m128i _wd = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptrd + k), 1));
-                __m128i _we = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptre + k), 1));
-                __m128i _wf = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptrf + k), 1));
-
-                transpose8x16_epi16(_w0, _w1, _w2, _w3, _w4, _w5, _w6, _w7, _w8, _w9, _wa, _wb, _wc, _wd, _we, _wf);
-
-                _mm_storeu_si128((__m128i*)g00, _w0);
-                _mm_storeu_si128((__m128i*)(g00 + 16), _w1);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 2), _w2);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 3), _w3);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 4), _w4);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 5), _w5);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 6), _w6);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 7), _w7);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 8), _w8);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 9), _w9);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 10), _wa);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 11), _wb);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 12), _wc);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 13), _wd);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 14), _we);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 15), _wf);
-                g00 += 256;
-            }
-
-            kptr0 += maxk * 16;
-            kptr1 += maxk * 16;
-            kptr2 += maxk * 16;
-            kptr3 += maxk * 16;
-            kptr4 += maxk * 16;
-            kptr5 += maxk * 16;
-            kptr6 += maxk * 16;
-            kptr7 += maxk * 16;
-            kptr8 += maxk * 16;
-            kptr9 += maxk * 16;
-            kptra += maxk * 16;
-            kptrb += maxk * 16;
-            kptrc += maxk * 16;
-            kptrd += maxk * 16;
-            kptre += maxk * 16;
-            kptrf += maxk * 16;
-        }
-        for (; p + 7 < inch; p += 8)
-        {
-            __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-            _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(maxk));
-
-            for (int k = 0; k < maxk; k++)
-            {
-                __m128i _w0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptr0 + k), _vindex, 1));
-                __m128i _w1 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptr1 + k), _vindex, 1));
-                __m128i _w2 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptr2 + k), _vindex, 1));
-                __m128i _w3 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptr3 + k), _vindex, 1));
-                __m128i _w4 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptr4 + k), _vindex, 1));
-                __m128i _w5 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptr5 + k), _vindex, 1));
-                __m128i _w6 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptr6 + k), _vindex, 1));
-                __m128i _w7 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptr7 + k), _vindex, 1));
-                __m128i _w8 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptr8 + k), _vindex, 1));
-                __m128i _w9 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptr9 + k), _vindex, 1));
-                __m128i _wa = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptra + k), _vindex, 1));
-                __m128i _wb = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptrb + k), _vindex, 1));
-                __m128i _wc = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptrc + k), _vindex, 1));
-                __m128i _wd = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptrd + k), _vindex, 1));
-                __m128i _we = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptre + k), _vindex, 1));
-                __m128i _wf = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptrf + k), _vindex, 1));
-
-                __m128i _w08 = _mm_unpacklo_epi64(_w0, _w8);
-                __m128i _w19 = _mm_unpacklo_epi64(_w1, _w9);
-                __m128i _w2a = _mm_unpacklo_epi64(_w2, _wa);
-                __m128i _w3b = _mm_unpacklo_epi64(_w3, _wb);
-                __m128i _w4c = _mm_unpacklo_epi64(_w4, _wc);
-                __m128i _w5d = _mm_unpacklo_epi64(_w5, _wd);
-                __m128i _w6e = _mm_unpacklo_epi64(_w6, _we);
-                __m128i _w7f = _mm_unpacklo_epi64(_w7, _wf);
-
-                transpose8x8_epi16(_w08, _w19, _w2a, _w3b, _w4c, _w5d, _w6e, _w7f);
-
-                _mm_storeu_si128((__m128i*)g00, _w08);
-                _mm_storeu_si128((__m128i*)(g00 + 16), _w4c);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 2), _w19);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 3), _w5d);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 4), _w2a);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 5), _w6e);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 6), _w3b);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 7), _w7f);
-                g00 += 128;
-            }
-
-            kptr0 += maxk * 8;
-            kptr1 += maxk * 8;
-            kptr2 += maxk * 8;
-            kptr3 += maxk * 8;
-            kptr4 += maxk * 8;
-            kptr5 += maxk * 8;
-            kptr6 += maxk * 8;
-            kptr7 += maxk * 8;
-            kptr8 += maxk * 8;
-            kptr9 += maxk * 8;
-            kptra += maxk * 8;
-            kptrb += maxk * 8;
-            kptrc += maxk * 8;
-            kptrd += maxk * 8;
-            kptre += maxk * 8;
-            kptrf += maxk * 8;
-        }
-        for (; p + 1 < inch; p += 2)
-        {
-            __m128i _vindex0 = _mm_setr_epi32(0, maxk, inch * maxk, inch * maxk + maxk);
-            __m128i _vindex1 = _mm_add_epi32(_vindex0, _mm_set1_epi32(inch * maxk * 2));
-            __m256i _vindex01 = _mm256_inserti128_si256(_mm256_castsi128_si256(_vindex0), _vindex1, 1);
-            __m256i _vindex23 = _mm256_add_epi32(_vindex01, _mm256_set1_epi32(inch * maxk * 4));
-            __m512i _vindex = _mm512_inserti64x4(_mm512_castsi256_si512(_vindex01), _vindex23, 1);
-            for (int k = 0; k < maxk; k++)
-            {
-                __m128i _w0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr0 + k), 1));
-                __m128i _w1 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr8 + k), 1));
-
-                _mm_storeu_si128((__m128i*)g00, _w0);
-                _mm_storeu_si128((__m128i*)(g00 + 16), _w1);
-                g00 += 32;
-            }
-
-            kptr0 += maxk * 2;
-            kptr8 += maxk * 2;
-        }
-        for (; p < inch; p++)
-        {
-            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(inch * maxk));
-            for (int k = 0; k < maxk; k++)
-            {
-                __m128i _w0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr0 + k), 1));
-                _mm_storeu_si128((__m128i*)g00, _w0);
-                g00 += 16;
-            }
-        }
-    }
-#endif // __AVX512F__
-    for (; q + 7 < outch; q += 8)
-    {
-        const signed char* kptr0 = (const signed char*)kernel + q * inch * maxk;
-        const signed char* kptr1 = (const signed char*)kernel + (q + 1) * inch * maxk;
-        const signed char* kptr2 = (const signed char*)kernel + (q + 2) * inch * maxk;
-        const signed char* kptr3 = (const signed char*)kernel + (q + 3) * inch * maxk;
-        const signed char* kptr4 = (const signed char*)kernel + (q + 4) * inch * maxk;
-        const signed char* kptr5 = (const signed char*)kernel + (q + 5) * inch * maxk;
-        const signed char* kptr6 = (const signed char*)kernel + (q + 6) * inch * maxk;
-        const signed char* kptr7 = (const signed char*)kernel + (q + 7) * inch * maxk;
-
-#if __AVX512F__
-        signed char* g00 = kernel_tm.channel(q / 16 + (q % 16) / 8);
-#else
-        signed char* g00 = kernel_tm.channel(q / 8);
-#endif
-
-        int p = 0;
-#if __AVX512F__
-        for (; p + 15 < inch; p += 16)
-        {
-            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(maxk));
-
-            for (int k = 0; k < maxk; k++)
-            {
-                __m128i _w0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr0 + k), 1));
-                __m128i _w1 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr1 + k), 1));
-                __m128i _w2 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr2 + k), 1));
-                __m128i _w3 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr3 + k), 1));
-                __m128i _w4 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr4 + k), 1));
-                __m128i _w5 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr5 + k), 1));
-                __m128i _w6 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr6 + k), 1));
-                __m128i _w7 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr7 + k), 1));
-
-                transpose8x8_epi16(_w0, _w1, _w2, _w3, _w4, _w5, _w6, _w7);
-
-                _mm_storeu_si128((__m128i*)g00, _w0);
-                _mm_storeu_si128((__m128i*)(g00 + 16), _w1);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 2), _w2);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 3), _w3);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 4), _w4);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 5), _w5);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 6), _w6);
-                _mm_storeu_si128((__m128i*)(g00 + 16 * 7), _w7);
-                g00 += 128;
-            }
-
-            kptr0 += maxk * 16;
-            kptr1 += maxk * 16;
-            kptr2 += maxk * 16;
-            kptr3 += maxk * 16;
-            kptr4 += maxk * 16;
-            kptr5 += maxk * 16;
-            kptr6 += maxk * 16;
-            kptr7 += maxk * 16;
-        }
-#endif // __AVX512F__
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k0 = kptr0 + k;
-                const signed char* k1 = kptr1 + k;
-                const signed char* k2 = kptr2 + k;
-                const signed char* k3 = kptr3 + k;
-                const signed char* k4 = kptr4 + k;
-                const signed char* k5 = kptr5 + k;
-                const signed char* k6 = kptr6 + k;
-                const signed char* k7 = kptr7 + k;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k0[0];
-                    g00[1] = k0[maxk];
-                    g00[2] = k1[0];
-                    g00[3] = k1[maxk];
-                    g00[4] = k2[0];
-                    g00[5] = k2[maxk];
-                    g00[6] = k3[0];
-                    g00[7] = k3[maxk];
-                    g00[8] = k4[0];
-                    g00[9] = k4[maxk];
-                    g00[10] = k5[0];
-                    g00[11] = k5[maxk];
-                    g00[12] = k6[0];
-                    g00[13] = k6[maxk];
-                    g00[14] = k7[0];
-                    g00[15] = k7[maxk];
-
-                    g00 += 16;
-                    k0 += maxk * 2;
-                    k1 += maxk * 2;
-                    k2 += maxk * 2;
-                    k3 += maxk * 2;
-                    k4 += maxk * 2;
-                    k5 += maxk * 2;
-                    k6 += maxk * 2;
-                    k7 += maxk * 2;
-                }
-            }
-
-            kptr0 += maxk * 8;
-            kptr1 += maxk * 8;
-            kptr2 += maxk * 8;
-            kptr3 += maxk * 8;
-            kptr4 += maxk * 8;
-            kptr5 += maxk * 8;
-            kptr6 += maxk * 8;
-            kptr7 += maxk * 8;
-        }
-        for (; p + 1 < inch; p += 2)
-        {
-            __m128i _vindex0 = _mm_setr_epi32(0, maxk, inch * maxk, inch * maxk + maxk);
-            __m128i _vindex1 = _mm_add_epi32(_vindex0, _mm_set1_epi32(inch * maxk * 2));
-            __m256i _vindex01 = _mm256_inserti128_si256(_mm256_castsi128_si256(_vindex0), _vindex1, 1);
-#if __AVX512F__
-            __m256i _vindex23 = _mm256_add_epi32(_vindex01, _mm256_set1_epi32(inch * maxk * 4));
-            __m512i _vindex = _mm512_inserti64x4(_mm512_castsi256_si512(_vindex01), _vindex23, 1);
-#else
-            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-            __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-#endif
-            for (int k = 0; k < maxk; k++)
-            {
-#if __AVX512F__
-                __m128i _w0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr0 + k), 1));
-#else
-                __m256i _w01 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)(kptr0 + k), _vindex01, 1), _sindex88);
-                __m256i _w23 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)(kptr4 + k), _vindex01, 1), _sindex88);
-                __m128i _w01xx = _mm_unpacklo_epi32(_mm256_extracti128_si256(_w01, 0), _mm256_extracti128_si256(_w01, 1));
-                __m128i _w23xx = _mm_unpacklo_epi32(_mm256_extracti128_si256(_w23, 0), _mm256_extracti128_si256(_w23, 1));
-                __m128i _w0 = _mm_unpacklo_epi64(_w01xx, _w23xx);
-#endif
-                _mm_storeu_si128((__m128i*)g00, _w0);
-                g00 += 16;
-            }
-
-            kptr0 += maxk * 2;
-            kptr4 += maxk * 2;
-        }
-        for (; p < inch; p++)
-        {
-            __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-            _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(inch * maxk));
-#if !__AVX512F__
-            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-            __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-#endif
-            for (int k = 0; k < maxk; k++)
-            {
-                __m256i _w32 = _mm256_i32gather_epi32((const int*)(kptr0 + k), _vindex, 1);
-#if __AVX512F__
-                __m128i _w0 = _mm256_cvtepi32_epi8(_w32);
-#else
-                __m256i _w01 = _mm256_shuffle_epi8(_w32, _sindex88);
-                __m128i _w0 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_w01, 0), _mm256_extracti128_si256(_w01, 1));
-#endif
-                _mm_storel_epi64((__m128i*)g00, _w0);
-                g00 += 8;
-            }
-        }
-    }
-#endif // __AVX2__
-    for (; q + 3 < outch; q += 4)
-    {
-        const signed char* kptr0 = (const signed char*)kernel + q * inch * maxk;
-        const signed char* kptr1 = (const signed char*)kernel + (q + 1) * inch * maxk;
-        const signed char* kptr2 = (const signed char*)kernel + (q + 2) * inch * maxk;
-        const signed char* kptr3 = (const signed char*)kernel + (q + 3) * inch * maxk;
-
-#if __AVX512F__
-        signed char* g00 = kernel_tm.channel(q / 16 + (q % 16) / 8 + (q % 8) / 4);
-#elif __AVX2__
-        signed char* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4);
-#else
-        signed char* g00 = kernel_tm.channel(q / 4);
-#endif
-
-        int p = 0;
-#if __AVX512F__
-        for (; p + 15 < inch; p += 16)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k0 = kptr0 + k;
-                const signed char* k1 = kptr1 + k;
-                const signed char* k2 = kptr2 + k;
-                const signed char* k3 = kptr3 + k;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[0];
-                    g00[1] = k0[maxk];
-                    g00[2] = k1[0];
-                    g00[3] = k1[maxk];
-                    g00[4] = k2[0];
-                    g00[5] = k2[maxk];
-                    g00[6] = k3[0];
-                    g00[7] = k3[maxk];
-
-                    g00 += 8;
-                    k0 += maxk * 2;
-                    k1 += maxk * 2;
-                    k2 += maxk * 2;
-                    k3 += maxk * 2;
-                }
-            }
-
-            kptr0 += maxk * 16;
-            kptr1 += maxk * 16;
-            kptr2 += maxk * 16;
-            kptr3 += maxk * 16;
-        }
-#endif // __AVX512F__
-        for (; p + 7 < inch; p += 8)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k0 = kptr0 + k;
-                const signed char* k1 = kptr1 + k;
-                const signed char* k2 = kptr2 + k;
-                const signed char* k3 = kptr3 + k;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    g00[0] = k0[0];
-                    g00[1] = k0[maxk];
-                    g00[2] = k1[0];
-                    g00[3] = k1[maxk];
-                    g00[4] = k2[0];
-                    g00[5] = k2[maxk];
-                    g00[6] = k3[0];
-                    g00[7] = k3[maxk];
-
-                    g00 += 8;
-                    k0 += maxk * 2;
-                    k1 += maxk * 2;
-                    k2 += maxk * 2;
-                    k3 += maxk * 2;
-                }
-            }
-
-            kptr0 += maxk * 8;
-            kptr1 += maxk * 8;
-            kptr2 += maxk * 8;
-            kptr3 += maxk * 8;
-        }
-        for (; p + 1 < inch; p += 2)
-        {
-#if __AVX2__
-            __m128i _vindex0 = _mm_setr_epi32(0, maxk, inch * maxk, inch * maxk + maxk);
-            __m128i _vindex1 = _mm_add_epi32(_vindex0, _mm_set1_epi32(inch * maxk * 2));
-            __m256i _vindex01 = _mm256_inserti128_si256(_mm256_castsi128_si256(_vindex0), _vindex1, 1);
-#if !__AVX512F__
-            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-            __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-#endif
-#endif
-
-            for (int k = 0; k < maxk; k++)
-            {
-#if __AVX512F__
-                __m128i _w0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptr0 + k), _vindex01, 1));
-                _mm_storel_epi64((__m128i*)g00, _w0);
-#elif __AVX2__
-                __m256i _w01 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)(kptr0 + k), _vindex01, 1), _sindex88);
-                __m128i _w0 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_w01, 0), _mm256_extracti128_si256(_w01, 1));
-                _mm_storel_epi64((__m128i*)g00, _w0);
-#else
-                const signed char* k0 = kptr0 + k;
-                const signed char* k1 = kptr1 + k;
-                const signed char* k2 = kptr2 + k;
-                const signed char* k3 = kptr3 + k;
-
-                g00[0] = k0[0];
-                g00[1] = k0[maxk];
-                g00[2] = k1[0];
-                g00[3] = k1[maxk];
-                g00[4] = k2[0];
-                g00[5] = k2[maxk];
-                g00[6] = k3[0];
-                g00[7] = k3[maxk];
-#endif
-                g00 += 8;
-            }
-
-            kptr0 += maxk * 2;
-            kptr1 += maxk * 2;
-            kptr2 += maxk * 2;
-            kptr3 += maxk * 2;
-        }
-        for (; p < inch; p++)
-        {
-#if __AVX2__
-            __m128i _vindex = _mm_mullo_epi32(_mm_setr_epi32(0, 1, 2, 3), _mm_set1_epi32(inch * maxk));
-#if !__AVX512F__
-            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-#endif
-#endif
-
-            for (int k = 0; k < maxk; k++)
-            {
-#if __AVX512F__
-                __m128i _w0 = _mm_cvtepi32_epi8(_mm_i32gather_epi32((const int*)(kptr0 + k), _vindex, 1));
-                _mm_store_ss((float*)g00, _mm_castsi128_ps(_w0));
-#elif __AVX2__
-                __m128i _w0 = _mm_shuffle_epi8(_mm_i32gather_epi32((const int*)(kptr0 + k), _vindex, 1), _sindex8);
-                _mm_store_ss((float*)g00, _mm_castsi128_ps(_w0));
-#else
-                const signed char* k0 = kptr0 + k;
-                const signed char* k1 = kptr1 + k;
-                const signed char* k2 = kptr2 + k;
-                const signed char* k3 = kptr3 + k;
-
-                g00[0] = k0[0];
-                g00[1] = k1[0];
-                g00[2] = k2[0];
-                g00[3] = k3[0];
-#endif
-                g00 += 4;
-            }
-        }
-    }
-#endif // __SSE2__
-    for (; q + 1 < outch; q += 2)
-    {
-        const signed char* kptr0 = (const signed char*)kernel + q * inch * maxk;
-        const signed char* kptr1 = (const signed char*)kernel + (q + 1) * inch * maxk;
-
-#if __AVX512F__
-        signed char* g00 = kernel_tm.channel(q / 16 + (q % 16) / 8 + (q % 8) / 4 + (q % 4) / 2);
-#elif __AVX2__
-        signed char* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2);
-#elif __SSE2__
-        signed char* g00 = kernel_tm.channel(q / 4 + (q % 4) / 2);
-#else
-        signed char* g00 = kernel_tm.channel(q / 2);
-#endif
-
-        int p = 0;
-#if __SSE2__
-#if __AVX512F__
-        for (; p + 15 < inch; p += 16)
-        {
-            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(maxk));
-
-            for (int k = 0; k < maxk; k++)
-            {
-                __m128i _w0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr0 + k), 1));
-                __m128i _w1 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr1 + k), 1));
-
-                _mm_storeu_si128((__m128i*)g00, _w0);
-                _mm_storeu_si128((__m128i*)(g00 + 16), _w1);
-                g00 += 32;
-            }
-
-            kptr0 += maxk * 16;
-            kptr1 += maxk * 16;
-        }
-#endif // __AVX512F__
-        for (; p + 7 < inch; p += 8)
-        {
-#if __AVX2__
-            __m256i _vindex0 = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-            _vindex0 = _mm256_mullo_epi32(_vindex0, _mm256_set1_epi32(maxk));
-#if __AVX512F__
-            __m256i _vindex1 = _mm256_add_epi32(_vindex0, _mm256_set1_epi32(inch * maxk));
-            __m512i _vindex = _mm512_inserti64x4(_mm512_castsi256_si512(_vindex0), _vindex1, 1);
-#else
-            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-            __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-#endif
-#endif
-
-            for (int k = 0; k < maxk; k++)
-            {
-#if __AVX512F__
-                __m128i _w0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr0 + k), 1));
-                _mm_storeu_si128((__m128i*)g00, _w0);
-#elif __AVX2__
-                __m256i _w00 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)(kptr0 + k), _vindex0, 1), _sindex88);
-                __m256i _w11 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)(kptr1 + k), _vindex0, 1), _sindex88);
-                __m128i _w0x = _mm_unpacklo_epi32(_mm256_extracti128_si256(_w00, 0), _mm256_extracti128_si256(_w00, 1));
-                __m128i _w1x = _mm_unpacklo_epi32(_mm256_extracti128_si256(_w11, 0), _mm256_extracti128_si256(_w11, 1));
-                __m128i _w0 = _mm_unpacklo_epi64(_w0x, _w1x);
-                _mm_storeu_si128((__m128i*)g00, _w0);
-#else
-                const signed char* k0 = kptr0 + k;
-                const signed char* k1 = kptr1 + k;
-
-                g00[0] = k0[0];
-                g00[1] = k0[maxk];
-                g00[2] = k0[maxk * 2];
-                g00[3] = k0[maxk * 3];
-                g00[4] = k0[maxk * 4];
-                g00[5] = k0[maxk * 5];
-                g00[6] = k0[maxk * 6];
-                g00[7] = k0[maxk * 7];
-                g00[8] = k1[0];
-                g00[9] = k1[maxk];
-                g00[10] = k1[maxk * 2];
-                g00[11] = k1[maxk * 3];
-                g00[12] = k1[maxk * 4];
-                g00[13] = k1[maxk * 5];
-                g00[14] = k1[maxk * 6];
-                g00[15] = k1[maxk * 7];
-#endif
-                g00 += 16;
-            }
-
-            kptr0 += maxk * 8;
-            kptr1 += maxk * 8;
-        }
-#endif // __SSE2__
-        for (; p + 1 < inch; p += 2)
-        {
-#if __AVX2__
-            __m128i _vindex = _mm_setr_epi32(0, inch * maxk, maxk, inch * maxk + maxk);
-#if !__AVX512F__
-            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-#endif
-#endif
-
-            for (int k = 0; k < maxk; k++)
-            {
-#if __AVX512F__
-                __m128i _w0 = _mm_cvtepi32_epi8(_mm_i32gather_epi32((const int*)(kptr0 + k), _vindex, 1));
-                _mm_store_ss((float*)g00, _mm_castsi128_ps(_w0));
-#elif __AVX2__
-                __m128i _w0 = _mm_shuffle_epi8(_mm_i32gather_epi32((const int*)(kptr0 + k), _vindex, 1), _sindex8);
-                _mm_store_ss((float*)g00, _mm_castsi128_ps(_w0));
-#else
-                const signed char* k0 = kptr0 + k;
-                const signed char* k1 = kptr1 + k;
-
-                g00[0] = k0[0];
-                g00[1] = k1[0];
-                g00[2] = k0[maxk];
-                g00[3] = k1[maxk];
-#endif
-                g00 += 4;
-            }
-
-            kptr0 += maxk * 2;
-            kptr1 += maxk * 2;
-        }
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k0 = kptr0 + k;
-                const signed char* k1 = kptr1 + k;
-
-                g00[0] = k0[0];
-                g00[1] = k1[0];
-                g00 += 2;
-            }
-        }
-    }
-    for (; q < outch; q++)
-    {
-        const signed char* kptr = (const signed char*)kernel + q * inch * maxk;
-
-#if __AVX512F__
-        signed char* g00 = kernel_tm.channel(q / 16 + (q % 16) / 8 + (q % 8) / 4 + (q % 4) / 2 + q % 2);
-#elif __AVX2__
-        signed char* g00 = kernel_tm.channel(q / 8 + (q % 8) / 4 + (q % 4) / 2 + q % 2);
-#elif __SSE2__
-        signed char* g00 = kernel_tm.channel(q / 4 + (q % 4) / 2 + q % 2);
-#else
-        signed char* g00 = kernel_tm.channel(q / 2 + q % 2);
-#endif
-
-        int p = 0;
-#if __SSE2__
-#if __AVX512F__
-        for (; p + 15 < inch; p += 16)
-        {
-            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(maxk));
-
-            for (int k = 0; k < maxk; k++)
-            {
-                __m128i _w0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(kptr + k), 1));
-
-                _mm_storeu_si128((__m128i*)g00, _w0);
-                g00 += 16;
-            }
-
-            kptr += maxk * 16;
-        }
-#endif // __AVX512F__
-        for (; p + 7 < inch; p += 8)
-        {
-#if __AVX2__
-            __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-            _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(maxk));
-#if !__AVX512F__
-            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-            __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-#endif
-#endif
-            for (int k = 0; k < maxk; k++)
-            {
-#if __AVX512F__
-                __m128i _w0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)(kptr + k), _vindex, 1));
-
-                _mm_storel_epi64((__m128i*)g00, _w0);
-                g00 += 8;
-#elif __AVX2__
-                __m256i _w00 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)(kptr + k), _vindex, 1), _sindex88);
-                __m128i _w0 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_w00, 0), _mm256_extracti128_si256(_w00, 1));
-
-                _mm_storel_epi64((__m128i*)g00, _w0);
-                g00 += 8;
-#else
-                const signed char* k0 = kptr + k;
-
-                for (int i = 0; i < 8; i++)
-                {
-                    g00[0] = k0[0];
-                    k0 += maxk;
-                    g00 += 1;
-                }
-#endif
-            }
-
-            kptr += maxk * 8;
-        }
-#endif // __SSE2__
-        for (; p + 1 < inch; p += 2)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k0 = kptr + k;
-
-                g00[0] = k0[0];
-                g00[1] = k0[maxk];
-                g00 += 2;
-            }
-
-            kptr += maxk * 2;
-        }
-        for (; p < inch; p++)
-        {
-            for (int k = 0; k < maxk; k++)
-            {
-                const signed char* k0 = kptr + k;
-
-                g00[0] = k0[0];
-                g00++;
-            }
-        }
-    }
-}
-
-static void convolution_packed_int8(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
-{
-#if !(__AVX512VNNI__ || __AVXVNNI__ || __AVX2__ || __XOP__)
-#if NCNN_RUNTIME_CPU && NCNN_AVX512VNNI && __AVX512F__ && !__AVX512VNNI__
-    if (ncnn::cpu_support_x86_avx512_vnni())
-    {
-        convolution_packed_int8_avx512vnni(bottom_blob, top_blob, weight_data_tm, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVXVNNI && __AVX2__ && !__AVXVNNI__
-    if (ncnn::cpu_support_x86_avx_vnni())
-    {
-        convolution_packed_int8_avxvnni(bottom_blob, top_blob, weight_data_tm, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_AVX2 && __AVX__ && !__AVX2__
-    if (ncnn::cpu_support_x86_avx2())
-    {
-        convolution_packed_int8_avx2(bottom_blob, top_blob, weight_data_tm, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
-        return;
-    }
-#endif
-
-#if NCNN_RUNTIME_CPU && NCNN_XOP && __SSE2__ && !__XOP__
-    if (ncnn::cpu_support_x86_xop())
-    {
-        convolution_packed_int8_xop(bottom_blob, top_blob, weight_data_tm, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
-        return;
-    }
-#endif
-#endif
-
-    const int w = bottom_blob.w;
-    const int elempack = bottom_blob.elempack;
-    const int inch = bottom_blob.c * elempack;
-
-    const int N = bottom_blob.cstep * elempack;
-
-    const int outw = top_blob.w;
-    const int outh = top_blob.h;
-    const int out_elempack = top_blob.elempack;
-    const int outch = top_blob.c * out_elempack;
-
-    const int maxk = kernel_w * kernel_h;
-
-    // kernel offsets
-    std::vector<int> _space_ofs(maxk);
-    int* space_ofs = &_space_ofs[0];
-    {
-        int p1 = 0;
-        int p2 = 0;
-        int gap = w * dilation_h - kernel_w * dilation_w;
-        for (int i = 0; i < kernel_h; i++)
-        {
-            for (int j = 0; j < kernel_w; j++)
-            {
-                space_ofs[p1] = p2 * elempack;
-                p1++;
-                p2 += dilation_w;
-            }
-            p2 += gap;
-        }
-    }
-
-    int nn_outch = 0;
-    int remain_outch_start = 0;
-#if __SSE2__
-#if __AVX2__
-#if __AVX512F__
-    nn_outch = outch / 16;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = pp * 16;
-
-        // shadowed variable for less openmp task args
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int N = bottom_blob.cstep * elempack;
-        const int M = top_blob.cstep * out_elempack;
-
-        int* outptr = top_blob.channel(p / out_elempack);
-
-        int ij = 0;
-        for (; ij + 3 < outw * outh; ij += 4)
-        {
-            const int i0 = ij / outw;
-            const int i1 = (ij + 1) / outw;
-            const int i2 = (ij + 2) / outw;
-            const int i3 = (ij + 3) / outw;
-            const int j0 = ij % outw;
-            const int j1 = (ij + 1) % outw;
-            const int j2 = (ij + 2) % outw;
-            const int j3 = (ij + 3) % outw;
-
-            __m512i _sum0 = _mm512_setzero_si512();
-            __m512i _sum1 = _mm512_setzero_si512();
-            __m512i _sum2 = _mm512_setzero_si512();
-            __m512i _sum3 = _mm512_setzero_si512();
-
-            const signed char* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8);
-
-            int q = 0;
-            for (; q + 15 < inch; q += 16)
-            {
-                const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-                const signed char* r2 = bottom_blob.channel(q / elempack).row<const signed char>(i2 * stride_h) + j2 * stride_w * elempack;
-                const signed char* r3 = bottom_blob.channel(q / elempack).row<const signed char>(i3 * stride_h) + j3 * stride_w * elempack;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-                    const signed char* r2s = r2 + space_ofs[k];
-                    const signed char* r3s = r3 + space_ofs[k];
-
-                    __m128i _r0;
-                    __m128i _r1;
-                    __m128i _r2;
-                    __m128i _r3;
-                    if (elempack == 16)
-                    {
-                        _r0 = _mm_load_si128((const __m128i*)r0s);
-                        _r1 = _mm_load_si128((const __m128i*)r1s);
-                        _r2 = _mm_load_si128((const __m128i*)r2s);
-                        _r3 = _mm_load_si128((const __m128i*)r3s);
-                    }
-                    else if (elempack == 8)
-                    {
-                        __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0s);
-                        __m128i _t1 = _mm_loadl_epi64((const __m128i*)(r0s + N));
-                        __m128i _t2 = _mm_loadl_epi64((const __m128i*)r1s);
-                        __m128i _t3 = _mm_loadl_epi64((const __m128i*)(r1s + N));
-                        __m128i _t4 = _mm_loadl_epi64((const __m128i*)r2s);
-                        __m128i _t5 = _mm_loadl_epi64((const __m128i*)(r2s + N));
-                        __m128i _t6 = _mm_loadl_epi64((const __m128i*)r3s);
-                        __m128i _t7 = _mm_loadl_epi64((const __m128i*)(r3s + N));
-                        _r0 = _mm_unpacklo_epi64(_t0, _t1);
-                        _r1 = _mm_unpacklo_epi64(_t2, _t3);
-                        _r2 = _mm_unpacklo_epi64(_t4, _t5);
-                        _r3 = _mm_unpacklo_epi64(_t6, _t7);
-                    }
-                    else // if (elempack == 1)
-                    {
-                        __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                        _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                        _r0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0s), 1));
-                        _r1 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r1s), 1));
-                        _r2 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r2s), 1));
-                        _r3 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r3s), 1));
-                    }
-
-                    __m256i _rr0 = _mm256_cvtepi8_epi16(_r0);
-                    __m256i _rr1 = _mm256_cvtepi8_epi16(_r1);
-                    __m256i _rr2 = _mm256_cvtepi8_epi16(_r2);
-                    __m256i _rr3 = _mm256_cvtepi8_epi16(_r3);
-
-                    __m512i _w01 = _mm512_load_si512((const __m512i*)kptr);
-                    __m512i _w23 = _mm512_load_si512((const __m512i*)(kptr + 64));
-                    __m512i _w45 = _mm512_load_si512((const __m512i*)(kptr + 128));
-                    __m512i _w67 = _mm512_load_si512((const __m512i*)(kptr + 192));
-                    __m512i _w0 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 0));
-                    __m512i _w1 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 1));
-                    __m512i _w2 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 0));
-                    __m512i _w3 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 1));
-                    __m512i _w4 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w45, 0));
-                    __m512i _w5 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w45, 1));
-                    __m512i _w6 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w67, 0));
-                    __m512i _w7 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w67, 1));
-
-                    // 01234567 89abcdef -> 01010101 01010101 01010101 01010101 01010101 01010101 01010101 01010101
-                    __m512i _rrrr00 = _mm512_broadcast_i32x4(_mm256_extracti128_si256(_rr0, 0));
-                    __m512i _rrrr01 = _mm512_broadcast_i32x4(_mm256_extracti128_si256(_rr0, 1));
-                    __m512i _rrrr10 = _mm512_broadcast_i32x4(_mm256_extracti128_si256(_rr1, 0));
-                    __m512i _rrrr11 = _mm512_broadcast_i32x4(_mm256_extracti128_si256(_rr1, 1));
-                    __m512i _rrrr20 = _mm512_broadcast_i32x4(_mm256_extracti128_si256(_rr2, 0));
-                    __m512i _rrrr21 = _mm512_broadcast_i32x4(_mm256_extracti128_si256(_rr2, 1));
-                    __m512i _rrrr30 = _mm512_broadcast_i32x4(_mm256_extracti128_si256(_rr3, 0));
-                    __m512i _rrrr31 = _mm512_broadcast_i32x4(_mm256_extracti128_si256(_rr3, 1));
-
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr00, _MM_PERM_AAAA), _w0);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr10, _MM_PERM_AAAA), _w0);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr20, _MM_PERM_AAAA), _w0);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr30, _MM_PERM_AAAA), _w0);
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr00, _MM_PERM_BBBB), _w1);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr10, _MM_PERM_BBBB), _w1);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr20, _MM_PERM_BBBB), _w1);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr30, _MM_PERM_BBBB), _w1);
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr00, _MM_PERM_CCCC), _w2);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr10, _MM_PERM_CCCC), _w2);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr20, _MM_PERM_CCCC), _w2);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr30, _MM_PERM_CCCC), _w2);
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr00, _MM_PERM_DDDD), _w3);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr10, _MM_PERM_DDDD), _w3);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr20, _MM_PERM_DDDD), _w3);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr30, _MM_PERM_DDDD), _w3);
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr01, _MM_PERM_AAAA), _w4);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr11, _MM_PERM_AAAA), _w4);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr21, _MM_PERM_AAAA), _w4);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr31, _MM_PERM_AAAA), _w4);
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr01, _MM_PERM_BBBB), _w5);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr11, _MM_PERM_BBBB), _w5);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr21, _MM_PERM_BBBB), _w5);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr31, _MM_PERM_BBBB), _w5);
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr01, _MM_PERM_CCCC), _w6);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr11, _MM_PERM_CCCC), _w6);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr21, _MM_PERM_CCCC), _w6);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr31, _MM_PERM_CCCC), _w6);
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr01, _MM_PERM_DDDD), _w7);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr11, _MM_PERM_DDDD), _w7);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr21, _MM_PERM_DDDD), _w7);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr31, _MM_PERM_DDDD), _w7);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr00, _MM_PERM_AAAA), _w0));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr10, _MM_PERM_AAAA), _w0));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr20, _MM_PERM_AAAA), _w0));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr30, _MM_PERM_AAAA), _w0));
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr00, _MM_PERM_BBBB), _w1));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr10, _MM_PERM_BBBB), _w1));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr20, _MM_PERM_BBBB), _w1));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr30, _MM_PERM_BBBB), _w1));
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr00, _MM_PERM_CCCC), _w2));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr10, _MM_PERM_CCCC), _w2));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr20, _MM_PERM_CCCC), _w2));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr30, _MM_PERM_CCCC), _w2));
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr00, _MM_PERM_DDDD), _w3));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr10, _MM_PERM_DDDD), _w3));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr20, _MM_PERM_DDDD), _w3));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr30, _MM_PERM_DDDD), _w3));
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr01, _MM_PERM_AAAA), _w4));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr11, _MM_PERM_AAAA), _w4));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr21, _MM_PERM_AAAA), _w4));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr31, _MM_PERM_AAAA), _w4));
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr01, _MM_PERM_BBBB), _w5));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr11, _MM_PERM_BBBB), _w5));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr21, _MM_PERM_BBBB), _w5));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr31, _MM_PERM_BBBB), _w5));
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr01, _MM_PERM_CCCC), _w6));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr11, _MM_PERM_CCCC), _w6));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr21, _MM_PERM_CCCC), _w6));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr31, _MM_PERM_CCCC), _w6));
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr01, _MM_PERM_DDDD), _w7));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr11, _MM_PERM_DDDD), _w7));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr21, _MM_PERM_DDDD), _w7));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr31, _MM_PERM_DDDD), _w7));
-#endif // __AVX512VNNI__
-
-                    kptr += 256;
-                }
-            }
-            for (; q + 7 < inch; q += 8)
-            {
-                const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-                const signed char* r2 = bottom_blob.channel(q / elempack).row<const signed char>(i2 * stride_h) + j2 * stride_w * elempack;
-                const signed char* r3 = bottom_blob.channel(q / elempack).row<const signed char>(i3 * stride_h) + j3 * stride_w * elempack;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-                    const signed char* r2s = r2 + space_ofs[k];
-                    const signed char* r3s = r3 + space_ofs[k];
-
-                    __m128i _r0;
-                    __m128i _r1;
-                    __m128i _r2;
-                    __m128i _r3;
-                    if (elempack == 8)
-                    {
-                        _r0 = _mm_loadl_epi64((const __m128i*)r0s);
-                        _r1 = _mm_loadl_epi64((const __m128i*)r1s);
-                        _r2 = _mm_loadl_epi64((const __m128i*)r2s);
-                        _r3 = _mm_loadl_epi64((const __m128i*)r3s);
-                    }
-                    else // if (elempack == 1)
-                    {
-                        __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                        _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-                        _r0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1));
-                        _r1 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1));
-                        _r2 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r2s, _vindex, 1));
-                        _r3 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r3s, _vindex, 1));
-                    }
-
-                    _r0 = _mm_cvtepi8_epi16(_r0);
-                    _r1 = _mm_cvtepi8_epi16(_r1);
-                    _r2 = _mm_cvtepi8_epi16(_r2);
-                    _r3 = _mm_cvtepi8_epi16(_r3);
-
-                    __m512i _w01 = _mm512_load_si512((const __m512i*)kptr);
-                    __m512i _w23 = _mm512_load_si512((const __m512i*)(kptr + 64));
-                    __m512i _w0 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 0));
-                    __m512i _w1 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 1));
-                    __m512i _w2 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 0));
-                    __m512i _w3 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 1));
-
-                    // 01234567 -> 01010101 01010101 01010101 01010101
-                    __m512i _rrrr0 = _mm512_broadcast_i32x4(_r0);
-                    __m512i _rrrr1 = _mm512_broadcast_i32x4(_r1);
-                    __m512i _rrrr2 = _mm512_broadcast_i32x4(_r2);
-                    __m512i _rrrr3 = _mm512_broadcast_i32x4(_r3);
-
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr0, _MM_PERM_AAAA), _w0);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr1, _MM_PERM_AAAA), _w0);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr2, _MM_PERM_AAAA), _w0);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr3, _MM_PERM_AAAA), _w0);
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr0, _MM_PERM_BBBB), _w1);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr1, _MM_PERM_BBBB), _w1);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr2, _MM_PERM_BBBB), _w1);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr3, _MM_PERM_BBBB), _w1);
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr0, _MM_PERM_CCCC), _w2);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr1, _MM_PERM_CCCC), _w2);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr2, _MM_PERM_CCCC), _w2);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr3, _MM_PERM_CCCC), _w2);
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr0, _MM_PERM_DDDD), _w3);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr1, _MM_PERM_DDDD), _w3);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr2, _MM_PERM_DDDD), _w3);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr3, _MM_PERM_DDDD), _w3);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr0, _MM_PERM_AAAA), _w0));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr1, _MM_PERM_AAAA), _w0));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr2, _MM_PERM_AAAA), _w0));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr3, _MM_PERM_AAAA), _w0));
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr0, _MM_PERM_BBBB), _w1));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr1, _MM_PERM_BBBB), _w1));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr2, _MM_PERM_BBBB), _w1));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr3, _MM_PERM_BBBB), _w1));
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr0, _MM_PERM_CCCC), _w2));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr1, _MM_PERM_CCCC), _w2));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr2, _MM_PERM_CCCC), _w2));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr3, _MM_PERM_CCCC), _w2));
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr0, _MM_PERM_DDDD), _w3));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr1, _MM_PERM_DDDD), _w3));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr2, _MM_PERM_DDDD), _w3));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr3, _MM_PERM_DDDD), _w3));
-#endif // __AVX512VNNI__
-
-                    kptr += 128;
-                }
-            }
-            for (; q + 1 < inch; q += 2)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-                const signed char* r2 = bottom_blob.channel(q).row<const signed char>(i2 * stride_h) + j2 * stride_w;
-                const signed char* r3 = bottom_blob.channel(q).row<const signed char>(i3 * stride_h) + j3 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-                    const signed char* r2s = r2 + space_ofs[k];
-                    const signed char* r3s = r3 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m512i _r0 = _mm512_broadcastd_epi32(_mm_setr_epi16(r0s[0], r0s[N], 0, 0, 0, 0, 0, 0));
-                        __m512i _r1 = _mm512_broadcastd_epi32(_mm_setr_epi16(r1s[0], r1s[N], 0, 0, 0, 0, 0, 0));
-                        __m512i _r2 = _mm512_broadcastd_epi32(_mm_setr_epi16(r2s[0], r2s[N], 0, 0, 0, 0, 0, 0));
-                        __m512i _r3 = _mm512_broadcastd_epi32(_mm_setr_epi16(r3s[0], r3s[N], 0, 0, 0, 0, 0, 0));
-
-                        __m512i _w = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*)kptr));
-
-#if __AVX512VNNI__
-                        _sum0 = _mm512_dpwssd_epi32(_sum0, _r0, _w);
-                        _sum1 = _mm512_dpwssd_epi32(_sum1, _r1, _w);
-                        _sum2 = _mm512_dpwssd_epi32(_sum2, _r2, _w);
-                        _sum3 = _mm512_dpwssd_epi32(_sum3, _r3, _w);
-#else
-                        _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_r0, _w));
-                        _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_r1, _w));
-                        _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_r2, _w));
-                        _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_r3, _w));
-#endif // __AVX512VNNI__
-
-                        kptr += 32;
-                    }
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-                const signed char* r2 = bottom_blob.channel(q).row<const signed char>(i2 * stride_h) + j2 * stride_w;
-                const signed char* r3 = bottom_blob.channel(q).row<const signed char>(i3 * stride_h) + j3 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-                    const signed char* r2s = r2 + space_ofs[k];
-                    const signed char* r3s = r3 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m256i _r0 = _mm256_set1_epi16(r0s[0]);
-                        __m256i _r1 = _mm256_set1_epi16(r1s[0]);
-                        __m256i _r2 = _mm256_set1_epi16(r2s[0]);
-                        __m256i _r3 = _mm256_set1_epi16(r3s[0]);
-
-                        __m256i _w = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)kptr));
-
-                        _sum0 = _mm512_add_epi32(_sum0, _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_r0, _w)));
-                        _sum1 = _mm512_add_epi32(_sum1, _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_r1, _w)));
-                        _sum2 = _mm512_add_epi32(_sum2, _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_r2, _w)));
-                        _sum3 = _mm512_add_epi32(_sum3, _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_r3, _w)));
-
-                        kptr += 16;
-                    }
-                }
-            }
-
-            if (out_elempack == 16)
-            {
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                _mm512_store_si512((__m512i*)(outptr + 16), _sum1);
-                _mm512_store_si512((__m512i*)(outptr + 32), _sum2);
-                _mm512_store_si512((__m512i*)(outptr + 48), _sum3);
-                outptr += 64;
-            }
-            if (out_elempack == 8)
-            {
-                _mm256_store_si256((__m256i*)outptr, _mm512_extracti32x8_epi32(_sum0, 0));
-                _mm256_store_si256((__m256i*)(outptr + M), _mm512_extracti32x8_epi32(_sum0, 1));
-                _mm256_store_si256((__m256i*)(outptr + 8), _mm512_extracti32x8_epi32(_sum1, 0));
-                _mm256_store_si256((__m256i*)(outptr + M + 8), _mm512_extracti32x8_epi32(_sum1, 1));
-                _mm256_store_si256((__m256i*)(outptr + 16), _mm512_extracti32x8_epi32(_sum2, 0));
-                _mm256_store_si256((__m256i*)(outptr + M + 16), _mm512_extracti32x8_epi32(_sum2, 1));
-                _mm256_store_si256((__m256i*)(outptr + 24), _mm512_extracti32x8_epi32(_sum3, 0));
-                _mm256_store_si256((__m256i*)(outptr + M + 24), _mm512_extracti32x8_epi32(_sum3, 1));
-                outptr += 32;
-            }
-            if (out_elempack == 4)
-            {
-                _mm_store_si128((__m128i*)outptr, _mm512_extracti32x4_epi32(_sum0, 0));
-                _mm_store_si128((__m128i*)(outptr + M), _mm512_extracti32x4_epi32(_sum0, 1));
-                _mm_store_si128((__m128i*)(outptr + M * 2), _mm512_extracti32x4_epi32(_sum0, 2));
-                _mm_store_si128((__m128i*)(outptr + M * 3), _mm512_extracti32x4_epi32(_sum0, 3));
-                _mm_store_si128((__m128i*)(outptr + 4), _mm512_extracti32x4_epi32(_sum1, 0));
-                _mm_store_si128((__m128i*)(outptr + M + 4), _mm512_extracti32x4_epi32(_sum1, 1));
-                _mm_store_si128((__m128i*)(outptr + M * 2 + 4), _mm512_extracti32x4_epi32(_sum1, 2));
-                _mm_store_si128((__m128i*)(outptr + M * 3 + 4), _mm512_extracti32x4_epi32(_sum1, 3));
-                _mm_store_si128((__m128i*)(outptr + 8), _mm512_extracti32x4_epi32(_sum2, 0));
-                _mm_store_si128((__m128i*)(outptr + M + 8), _mm512_extracti32x4_epi32(_sum2, 1));
-                _mm_store_si128((__m128i*)(outptr + M * 2 + 8), _mm512_extracti32x4_epi32(_sum2, 2));
-                _mm_store_si128((__m128i*)(outptr + M * 3 + 8), _mm512_extracti32x4_epi32(_sum2, 3));
-                _mm_store_si128((__m128i*)(outptr + 12), _mm512_extracti32x4_epi32(_sum3, 0));
-                _mm_store_si128((__m128i*)(outptr + M + 12), _mm512_extracti32x4_epi32(_sum3, 1));
-                _mm_store_si128((__m128i*)(outptr + M * 2 + 12), _mm512_extracti32x4_epi32(_sum3, 2));
-                _mm_store_si128((__m128i*)(outptr + M * 3 + 12), _mm512_extracti32x4_epi32(_sum3, 3));
-                outptr += 16;
-            }
-            if (out_elempack == 1)
-            {
-                __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(M));
-                _mm512_i32scatter_epi32(outptr, _vindex, _sum0, sizeof(int));
-                _mm512_i32scatter_epi32(outptr + 1, _vindex, _sum1, sizeof(int));
-                _mm512_i32scatter_epi32(outptr + 2, _vindex, _sum2, sizeof(int));
-                _mm512_i32scatter_epi32(outptr + 3, _vindex, _sum3, sizeof(int));
-                outptr += 4;
-            }
-        }
-        for (; ij + 1 < outw * outh; ij += 2)
-        {
-            const int i0 = ij / outw;
-            const int i1 = (ij + 1) / outw;
-            const int j0 = ij % outw;
-            const int j1 = (ij + 1) % outw;
-
-            __m512i _sum0 = _mm512_setzero_si512();
-            __m512i _sum1 = _mm512_setzero_si512();
-            __m512i _sum2 = _mm512_setzero_si512();
-            __m512i _sum3 = _mm512_setzero_si512();
-
-            const signed char* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8);
-
-            int q = 0;
-            for (; q + 15 < inch; q += 16)
-            {
-                const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    __m128i _r0;
-                    __m128i _r1;
-                    if (elempack == 16)
-                    {
-                        _r0 = _mm_load_si128((const __m128i*)r0s);
-                        _r1 = _mm_load_si128((const __m128i*)r1s);
-                    }
-                    else if (elempack == 8)
-                    {
-                        __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0s);
-                        __m128i _t1 = _mm_loadl_epi64((const __m128i*)(r0s + N));
-                        __m128i _t2 = _mm_loadl_epi64((const __m128i*)r1s);
-                        __m128i _t3 = _mm_loadl_epi64((const __m128i*)(r1s + N));
-                        _r0 = _mm_unpacklo_epi64(_t0, _t1);
-                        _r1 = _mm_unpacklo_epi64(_t2, _t3);
-                    }
-                    else // if (elempack == 1)
-                    {
-                        __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                        _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                        _r0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0s), 1));
-                        _r1 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r1s), 1));
-                    }
-
-                    __m256i _rr0 = _mm256_cvtepi8_epi16(_r0);
-                    __m256i _rr1 = _mm256_cvtepi8_epi16(_r1);
-
-                    __m512i _w01 = _mm512_load_si512((const __m512i*)kptr);
-                    __m512i _w23 = _mm512_load_si512((const __m512i*)(kptr + 64));
-                    __m512i _w45 = _mm512_load_si512((const __m512i*)(kptr + 128));
-                    __m512i _w67 = _mm512_load_si512((const __m512i*)(kptr + 192));
-                    __m512i _w0 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 0));
-                    __m512i _w1 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 1));
-                    __m512i _w2 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 0));
-                    __m512i _w3 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 1));
-                    __m512i _w4 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w45, 0));
-                    __m512i _w5 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w45, 1));
-                    __m512i _w6 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w67, 0));
-                    __m512i _w7 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w67, 1));
-
-                    // 01234567 89abcdef -> 01010101 01010101 01010101 01010101 01010101 01010101 01010101 01010101
-                    __m512i _rrrr00 = _mm512_broadcast_i32x4(_mm256_extracti128_si256(_rr0, 0));
-                    __m512i _rrrr01 = _mm512_broadcast_i32x4(_mm256_extracti128_si256(_rr0, 1));
-                    __m512i _rrrr10 = _mm512_broadcast_i32x4(_mm256_extracti128_si256(_rr1, 0));
-                    __m512i _rrrr11 = _mm512_broadcast_i32x4(_mm256_extracti128_si256(_rr1, 1));
-
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr00, _MM_PERM_AAAA), _w0);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr10, _MM_PERM_AAAA), _w0);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr00, _MM_PERM_BBBB), _w1);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr10, _MM_PERM_BBBB), _w1);
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr00, _MM_PERM_CCCC), _w2);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr10, _MM_PERM_CCCC), _w2);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr00, _MM_PERM_DDDD), _w3);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr10, _MM_PERM_DDDD), _w3);
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr01, _MM_PERM_AAAA), _w4);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr11, _MM_PERM_AAAA), _w4);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr01, _MM_PERM_BBBB), _w5);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr11, _MM_PERM_BBBB), _w5);
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr01, _MM_PERM_CCCC), _w6);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr11, _MM_PERM_CCCC), _w6);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr01, _MM_PERM_DDDD), _w7);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr11, _MM_PERM_DDDD), _w7);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr00, _MM_PERM_AAAA), _w0));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr10, _MM_PERM_AAAA), _w0));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr00, _MM_PERM_BBBB), _w1));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr10, _MM_PERM_BBBB), _w1));
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr00, _MM_PERM_CCCC), _w2));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr10, _MM_PERM_CCCC), _w2));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr00, _MM_PERM_DDDD), _w3));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr10, _MM_PERM_DDDD), _w3));
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr01, _MM_PERM_AAAA), _w4));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr11, _MM_PERM_AAAA), _w4));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr01, _MM_PERM_BBBB), _w5));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr11, _MM_PERM_BBBB), _w5));
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr01, _MM_PERM_CCCC), _w6));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr11, _MM_PERM_CCCC), _w6));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr01, _MM_PERM_DDDD), _w7));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr11, _MM_PERM_DDDD), _w7));
-#endif // __AVX512VNNI__
-
-                    kptr += 256;
-                }
-            }
-            for (; q + 7 < inch; q += 8)
-            {
-                const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    __m128i _r0;
-                    __m128i _r1;
-                    if (elempack == 8)
-                    {
-                        _r0 = _mm_loadl_epi64((const __m128i*)r0s);
-                        _r1 = _mm_loadl_epi64((const __m128i*)r1s);
-                    }
-                    else // if (elempack == 1)
-                    {
-                        __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                        _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-                        _r0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1));
-                        _r1 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1));
-                    }
-
-                    _r0 = _mm_cvtepi8_epi16(_r0);
-                    _r1 = _mm_cvtepi8_epi16(_r1);
-
-                    __m512i _w01 = _mm512_load_si512((const __m512i*)kptr);
-                    __m512i _w23 = _mm512_load_si512((const __m512i*)(kptr + 64));
-                    __m512i _w0 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 0));
-                    __m512i _w1 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 1));
-                    __m512i _w2 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 0));
-                    __m512i _w3 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 1));
-
-                    // 01234567 -> 01010101 01010101 01010101 01010101
-                    __m512i _rrrr0 = _mm512_broadcast_i32x4(_r0);
-                    __m512i _rrrr1 = _mm512_broadcast_i32x4(_r1);
-
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr0, _MM_PERM_AAAA), _w0);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr1, _MM_PERM_AAAA), _w0);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr0, _MM_PERM_BBBB), _w1);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr1, _MM_PERM_BBBB), _w1);
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr0, _MM_PERM_CCCC), _w2);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr1, _MM_PERM_CCCC), _w2);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr0, _MM_PERM_DDDD), _w3);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr1, _MM_PERM_DDDD), _w3);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr0, _MM_PERM_AAAA), _w0));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr1, _MM_PERM_AAAA), _w0));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr0, _MM_PERM_BBBB), _w1));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr1, _MM_PERM_BBBB), _w1));
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr0, _MM_PERM_CCCC), _w2));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr1, _MM_PERM_CCCC), _w2));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr0, _MM_PERM_DDDD), _w3));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr1, _MM_PERM_DDDD), _w3));
-#endif // __AVX512VNNI__
-
-                    kptr += 128;
-                }
-            }
-            for (; q + 1 < inch; q += 2)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m512i _r0 = _mm512_broadcastd_epi32(_mm_setr_epi16(r0s[0], r0s[N], 0, 0, 0, 0, 0, 0));
-                        __m512i _r1 = _mm512_broadcastd_epi32(_mm_setr_epi16(r1s[0], r1s[N], 0, 0, 0, 0, 0, 0));
-
-                        __m512i _w = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*)kptr));
-
-#if __AVX512VNNI__
-                        _sum0 = _mm512_dpwssd_epi32(_sum0, _r0, _w);
-                        _sum1 = _mm512_dpwssd_epi32(_sum1, _r1, _w);
-#else
-                        _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_r0, _w));
-                        _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_r1, _w));
-#endif // __AVX512VNNI__
-
-                        kptr += 32;
-                    }
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m256i _r0 = _mm256_set1_epi16(r0s[0]);
-                        __m256i _r1 = _mm256_set1_epi16(r1s[0]);
-
-                        __m256i _w = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)kptr));
-
-                        _sum0 = _mm512_add_epi32(_sum0, _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_r0, _w)));
-                        _sum1 = _mm512_add_epi32(_sum1, _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_r1, _w)));
-
-                        kptr += 16;
-                    }
-                }
-            }
-
-            _sum0 = _mm512_add_epi32(_sum0, _sum2);
-            _sum1 = _mm512_add_epi32(_sum1, _sum3);
-
-            if (out_elempack == 16)
-            {
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                _mm512_store_si512((__m512i*)(outptr + 16), _sum1);
-                outptr += 32;
-            }
-            if (out_elempack == 8)
-            {
-                _mm256_store_si256((__m256i*)outptr, _mm512_extracti32x8_epi32(_sum0, 0));
-                _mm256_store_si256((__m256i*)(outptr + M), _mm512_extracti32x8_epi32(_sum0, 1));
-                _mm256_store_si256((__m256i*)(outptr + 8), _mm512_extracti32x8_epi32(_sum1, 0));
-                _mm256_store_si256((__m256i*)(outptr + M + 8), _mm512_extracti32x8_epi32(_sum1, 1));
-                outptr += 16;
-            }
-            if (out_elempack == 4)
-            {
-                _mm_store_si128((__m128i*)outptr, _mm512_extracti32x4_epi32(_sum0, 0));
-                _mm_store_si128((__m128i*)(outptr + M), _mm512_extracti32x4_epi32(_sum0, 1));
-                _mm_store_si128((__m128i*)(outptr + M * 2), _mm512_extracti32x4_epi32(_sum0, 2));
-                _mm_store_si128((__m128i*)(outptr + M * 3), _mm512_extracti32x4_epi32(_sum0, 3));
-                _mm_store_si128((__m128i*)(outptr + 4), _mm512_extracti32x4_epi32(_sum1, 0));
-                _mm_store_si128((__m128i*)(outptr + M + 4), _mm512_extracti32x4_epi32(_sum1, 1));
-                _mm_store_si128((__m128i*)(outptr + M * 2 + 4), _mm512_extracti32x4_epi32(_sum1, 2));
-                _mm_store_si128((__m128i*)(outptr + M * 3 + 4), _mm512_extracti32x4_epi32(_sum1, 3));
-                outptr += 8;
-            }
-            if (out_elempack == 1)
-            {
-                __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(M));
-                _mm512_i32scatter_epi32(outptr, _vindex, _sum0, sizeof(int));
-                _mm512_i32scatter_epi32(outptr + 1, _vindex, _sum1, sizeof(int));
-                outptr += 2;
-            }
-        }
-        for (; ij < outw * outh; ij++)
-        {
-            const int i = ij / outw;
-            const int j = ij % outw;
-
-            __m512i _sum0 = _mm512_setzero_si512();
-            __m512i _sum1 = _mm512_setzero_si512();
-            __m512i _sum2 = _mm512_setzero_si512();
-            __m512i _sum3 = _mm512_setzero_si512();
-
-            const signed char* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8);
-
-            int q = 0;
-            for (; q + 15 < inch; q += 16)
-            {
-                const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i * stride_h) + j * stride_w * elempack;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    __m128i _r0;
-                    if (elempack == 16)
-                    {
-                        _r0 = _mm_load_si128((const __m128i*)r0s);
-                    }
-                    else if (elempack == 8)
-                    {
-                        __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0s);
-                        __m128i _t1 = _mm_loadl_epi64((const __m128i*)(r0s + N));
-                        _r0 = _mm_unpacklo_epi64(_t0, _t1);
-                    }
-                    else // if (elempack == 1)
-                    {
-                        __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                        _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                        _r0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0s), 1));
-                    }
-
-                    __m256i _rr0 = _mm256_cvtepi8_epi16(_r0);
-
-                    __m512i _w01 = _mm512_load_si512((const __m512i*)kptr);
-                    __m512i _w23 = _mm512_load_si512((const __m512i*)(kptr + 64));
-                    __m512i _w45 = _mm512_load_si512((const __m512i*)(kptr + 128));
-                    __m512i _w67 = _mm512_load_si512((const __m512i*)(kptr + 192));
-                    __m512i _w0 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 0));
-                    __m512i _w1 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 1));
-                    __m512i _w2 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 0));
-                    __m512i _w3 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 1));
-                    __m512i _w4 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w45, 0));
-                    __m512i _w5 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w45, 1));
-                    __m512i _w6 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w67, 0));
-                    __m512i _w7 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w67, 1));
-
-                    // 01234567 89abcdef -> 01010101 01010101 01010101 01010101 01010101 01010101 01010101 01010101
-                    __m512i _rrrr00 = _mm512_broadcast_i32x4(_mm256_extracti128_si256(_rr0, 0));
-                    __m512i _rrrr01 = _mm512_broadcast_i32x4(_mm256_extracti128_si256(_rr0, 1));
-
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr00, _MM_PERM_AAAA), _w0);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr00, _MM_PERM_BBBB), _w1);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr00, _MM_PERM_CCCC), _w2);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr00, _MM_PERM_DDDD), _w3);
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr01, _MM_PERM_AAAA), _w4);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr01, _MM_PERM_BBBB), _w5);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr01, _MM_PERM_CCCC), _w6);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr01, _MM_PERM_DDDD), _w7);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr00, _MM_PERM_AAAA), _w0));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr00, _MM_PERM_BBBB), _w1));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr00, _MM_PERM_CCCC), _w2));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr00, _MM_PERM_DDDD), _w3));
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr01, _MM_PERM_AAAA), _w4));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr01, _MM_PERM_BBBB), _w5));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr01, _MM_PERM_CCCC), _w6));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr01, _MM_PERM_DDDD), _w7));
-#endif // __AVX512VNNI__
-
-                    kptr += 256;
-                }
-            }
-            for (; q + 7 < inch; q += 8)
-            {
-                const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i * stride_h) + j * stride_w * elempack;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    __m128i _r0;
-                    if (elempack == 8)
-                    {
-                        _r0 = _mm_loadl_epi64((const __m128i*)r0s);
-                    }
-                    else // if (elempack == 1)
-                    {
-                        __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                        _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-                        _r0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1));
-                    }
-
-                    _r0 = _mm_cvtepi8_epi16(_r0);
-
-                    __m512i _w01 = _mm512_load_si512((const __m512i*)kptr);
-                    __m512i _w23 = _mm512_load_si512((const __m512i*)(kptr + 64));
-                    __m512i _w0 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 0));
-                    __m512i _w1 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 1));
-                    __m512i _w2 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 0));
-                    __m512i _w3 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 1));
-
-                    // 01234567 -> 01010101 01010101 01010101 01010101
-                    __m512i _rrrr0 = _mm512_broadcast_i32x4(_r0);
-
-#if __AVX512VNNI__
-                    _sum0 = _mm512_dpwssd_epi32(_sum0, _mm512_shuffle_epi32(_rrrr0, _MM_PERM_AAAA), _w0);
-                    _sum1 = _mm512_dpwssd_epi32(_sum1, _mm512_shuffle_epi32(_rrrr0, _MM_PERM_BBBB), _w1);
-                    _sum2 = _mm512_dpwssd_epi32(_sum2, _mm512_shuffle_epi32(_rrrr0, _MM_PERM_CCCC), _w2);
-                    _sum3 = _mm512_dpwssd_epi32(_sum3, _mm512_shuffle_epi32(_rrrr0, _MM_PERM_DDDD), _w3);
-#else
-                    _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr0, _MM_PERM_AAAA), _w0));
-                    _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr0, _MM_PERM_BBBB), _w1));
-                    _sum2 = _mm512_add_epi32(_sum2, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr0, _MM_PERM_CCCC), _w2));
-                    _sum3 = _mm512_add_epi32(_sum3, _mm512_madd_epi16(_mm512_shuffle_epi32(_rrrr0, _MM_PERM_DDDD), _w3));
-#endif // __AVX512VNNI__
-
-                    kptr += 128;
-                }
-            }
-            for (; q + 1 < inch; q += 2)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i * stride_h) + j * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m512i _val = _mm512_broadcastd_epi32(_mm_setr_epi16(r0s[0], r0s[N], 0, 0, 0, 0, 0, 0));
-
-                        __m512i _w = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*)kptr));
-
-#if __AVX512VNNI__
-                        _sum0 = _mm512_dpwssd_epi32(_sum0, _val, _w);
-#else
-                        _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_val, _w));
-#endif
-
-                        kptr += 32;
-                    }
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i * stride_h) + j * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m256i _val = _mm256_set1_epi16(r0s[0]);
-
-                        __m256i _w = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)kptr));
-
-                        _sum0 = _mm512_add_epi32(_sum0, _mm512_cvtepi16_epi32(_mm256_mullo_epi16(_val, _w)));
-
-                        kptr += 16;
-                    }
-                }
-            }
-
-            _sum0 = _mm512_add_epi32(_sum0, _sum1);
-            _sum2 = _mm512_add_epi32(_sum2, _sum3);
-            _sum0 = _mm512_add_epi32(_sum0, _sum2);
-
-            if (out_elempack == 16)
-            {
-                _mm512_store_si512((__m512i*)outptr, _sum0);
-                outptr += 16;
-            }
-            if (out_elempack == 8)
-            {
-                _mm256_store_si256((__m256i*)outptr, _mm512_extracti32x8_epi32(_sum0, 0));
-                _mm256_store_si256((__m256i*)(outptr + M), _mm512_extracti32x8_epi32(_sum0, 1));
-                outptr += 8;
-            }
-            if (out_elempack == 4)
-            {
-                _mm_store_si128((__m128i*)outptr, _mm512_extracti32x4_epi32(_sum0, 0));
-                _mm_store_si128((__m128i*)(outptr + M), _mm512_extracti32x4_epi32(_sum0, 1));
-                _mm_store_si128((__m128i*)(outptr + M * 2), _mm512_extracti32x4_epi32(_sum0, 2));
-                _mm_store_si128((__m128i*)(outptr + M * 3), _mm512_extracti32x4_epi32(_sum0, 3));
-                outptr += 4;
-            }
-            if (out_elempack == 1)
-            {
-                __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(M));
-                _mm512_i32scatter_epi32(outptr, _vindex, _sum0, sizeof(int));
-                outptr += 1;
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 16;
-    nn_outch = (outch - remain_outch_start) / 8;
-#else // __AVX512F__
-    nn_outch = (outch - remain_outch_start) / 8;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __AVX512F__
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 8;
-
-        // shadowed variable for less openmp task args
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int N = bottom_blob.cstep * elempack;
-        const int M = top_blob.cstep * out_elempack;
-
-        int* outptr = top_blob.channel(p / out_elempack);
-
-        int ij = 0;
-        for (; ij + 3 < outw * outh; ij += 4)
-        {
-            const int i0 = ij / outw;
-            const int i1 = (ij + 1) / outw;
-            const int i2 = (ij + 2) / outw;
-            const int i3 = (ij + 3) / outw;
-            const int j0 = ij % outw;
-            const int j1 = (ij + 1) % outw;
-            const int j2 = (ij + 2) % outw;
-            const int j3 = (ij + 3) % outw;
-
-            __m256i _sum0 = _mm256_setzero_si256();
-            __m256i _sum1 = _mm256_setzero_si256();
-            __m256i _sum2 = _mm256_setzero_si256();
-            __m256i _sum3 = _mm256_setzero_si256();
-
-#if __AVX512F__
-            const signed char* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 8);
-#endif
-
-            int q = 0;
-#if __AVX512F__
-            {
-                __m512i _sum00 = _mm512_setzero_si512();
-                __m512i _sum11 = _mm512_setzero_si512();
-                __m512i _sum22 = _mm512_setzero_si512();
-                __m512i _sum33 = _mm512_setzero_si512();
-                for (; q + 15 < inch; q += 16)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-                    const signed char* r2 = bottom_blob.channel(q / elempack).row<const signed char>(i2 * stride_h) + j2 * stride_w * elempack;
-                    const signed char* r3 = bottom_blob.channel(q / elempack).row<const signed char>(i3 * stride_h) + j3 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-                        const signed char* r2s = r2 + space_ofs[k];
-                        const signed char* r3s = r3 + space_ofs[k];
-
-                        __m128i _r0;
-                        __m128i _r1;
-                        __m128i _r2;
-                        __m128i _r3;
-                        if (elempack == 16)
-                        {
-                            _r0 = _mm_load_si128((const __m128i*)r0s);
-                            _r1 = _mm_load_si128((const __m128i*)r1s);
-                            _r2 = _mm_load_si128((const __m128i*)r2s);
-                            _r3 = _mm_load_si128((const __m128i*)r3s);
-                        }
-                        else if (elempack == 8)
-                        {
-                            __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            __m128i _t1 = _mm_loadl_epi64((const __m128i*)(r0s + N));
-                            __m128i _t2 = _mm_loadl_epi64((const __m128i*)r1s);
-                            __m128i _t3 = _mm_loadl_epi64((const __m128i*)(r1s + N));
-                            __m128i _t4 = _mm_loadl_epi64((const __m128i*)r2s);
-                            __m128i _t5 = _mm_loadl_epi64((const __m128i*)(r2s + N));
-                            __m128i _t6 = _mm_loadl_epi64((const __m128i*)r3s);
-                            __m128i _t7 = _mm_loadl_epi64((const __m128i*)(r3s + N));
-                            _r0 = _mm_unpacklo_epi64(_t0, _t1);
-                            _r1 = _mm_unpacklo_epi64(_t2, _t3);
-                            _r2 = _mm_unpacklo_epi64(_t4, _t5);
-                            _r3 = _mm_unpacklo_epi64(_t6, _t7);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                            _r0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0s), 1));
-                            _r1 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r1s), 1));
-                            _r2 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r2s), 1));
-                            _r3 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r3s), 1));
-                        }
-
-                        __m256i _rr0 = _mm256_cvtepi8_epi16(_r0);
-                        __m256i _rr1 = _mm256_cvtepi8_epi16(_r1);
-                        __m256i _rr2 = _mm256_cvtepi8_epi16(_r2);
-                        __m256i _rr3 = _mm256_cvtepi8_epi16(_r3);
-
-                        __m512i _w01 = _mm512_load_si512((const __m512i*)kptr);
-                        __m512i _w23 = _mm512_load_si512((const __m512i*)(kptr + 64));
-                        __m512i _w0 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 0));
-                        __m512i _w1 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 1));
-                        __m512i _w2 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 0));
-                        __m512i _w3 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 1));
-
-                        // 01234567 89abcdef -> 01010101 01010101 23232323 23232323
-                        __m512i _rrr0 = _mm512_inserti64x4(_mm512_castsi256_si512(_rr0), _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(1, 0, 3, 2)), 1);
-                        __m512i _rrr1 = _mm512_inserti64x4(_mm512_castsi256_si512(_rr1), _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(1, 0, 3, 2)), 1);
-                        __m512i _rrr2 = _mm512_inserti64x4(_mm512_castsi256_si512(_rr2), _mm256_shuffle_epi32(_rr2, _MM_SHUFFLE(1, 0, 3, 2)), 1);
-                        __m512i _rrr3 = _mm512_inserti64x4(_mm512_castsi256_si512(_rr3), _mm256_shuffle_epi32(_rr3, _MM_SHUFFLE(1, 0, 3, 2)), 1);
-
-                        _rrr0 = _mm512_unpacklo_epi32(_rrr0, _rrr0);
-                        _rrr1 = _mm512_unpacklo_epi32(_rrr1, _rrr1);
-                        _rrr2 = _mm512_unpacklo_epi32(_rrr2, _rrr2);
-                        _rrr3 = _mm512_unpacklo_epi32(_rrr3, _rrr3);
-
-                        __m512i _rrr0l = _mm512_unpacklo_epi64(_rrr0, _rrr0);
-                        __m512i _rrr0h = _mm512_unpackhi_epi64(_rrr0, _rrr0);
-                        __m512i _rrr1l = _mm512_unpacklo_epi64(_rrr1, _rrr1);
-                        __m512i _rrr1h = _mm512_unpackhi_epi64(_rrr1, _rrr1);
-                        __m512i _rrr2l = _mm512_unpacklo_epi64(_rrr2, _rrr2);
-                        __m512i _rrr2h = _mm512_unpackhi_epi64(_rrr2, _rrr2);
-                        __m512i _rrr3l = _mm512_unpacklo_epi64(_rrr3, _rrr3);
-                        __m512i _rrr3h = _mm512_unpackhi_epi64(_rrr3, _rrr3);
-
-#if __AVX512VNNI__
-                        _sum00 = _mm512_dpwssd_epi32(_sum00, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                        _sum11 = _mm512_dpwssd_epi32(_sum11, _mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                        _sum22 = _mm512_dpwssd_epi32(_sum22, _mm512_shuffle_i32x4(_rrr2l, _rrr2h, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                        _sum33 = _mm512_dpwssd_epi32(_sum33, _mm512_shuffle_i32x4(_rrr3l, _rrr3h, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                        _sum00 = _mm512_dpwssd_epi32(_sum00, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(2, 2, 2, 2)), _w1);
-                        _sum11 = _mm512_dpwssd_epi32(_sum11, _mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(2, 2, 2, 2)), _w1);
-                        _sum22 = _mm512_dpwssd_epi32(_sum22, _mm512_shuffle_i32x4(_rrr2l, _rrr2h, _MM_SHUFFLE(2, 2, 2, 2)), _w1);
-                        _sum33 = _mm512_dpwssd_epi32(_sum33, _mm512_shuffle_i32x4(_rrr3l, _rrr3h, _MM_SHUFFLE(2, 2, 2, 2)), _w1);
-                        _sum00 = _mm512_dpwssd_epi32(_sum00, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(1, 1, 1, 1)), _w2);
-                        _sum11 = _mm512_dpwssd_epi32(_sum11, _mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(1, 1, 1, 1)), _w2);
-                        _sum22 = _mm512_dpwssd_epi32(_sum22, _mm512_shuffle_i32x4(_rrr2l, _rrr2h, _MM_SHUFFLE(1, 1, 1, 1)), _w2);
-                        _sum33 = _mm512_dpwssd_epi32(_sum33, _mm512_shuffle_i32x4(_rrr3l, _rrr3h, _MM_SHUFFLE(1, 1, 1, 1)), _w2);
-                        _sum00 = _mm512_dpwssd_epi32(_sum00, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(3, 3, 3, 3)), _w3);
-                        _sum11 = _mm512_dpwssd_epi32(_sum11, _mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(3, 3, 3, 3)), _w3);
-                        _sum22 = _mm512_dpwssd_epi32(_sum22, _mm512_shuffle_i32x4(_rrr2l, _rrr2h, _MM_SHUFFLE(3, 3, 3, 3)), _w3);
-                        _sum33 = _mm512_dpwssd_epi32(_sum33, _mm512_shuffle_i32x4(_rrr3l, _rrr3h, _MM_SHUFFLE(3, 3, 3, 3)), _w3);
-#else
-                        _sum00 = _mm512_add_epi32(_sum00, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum11 = _mm512_add_epi32(_sum11, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum22 = _mm512_add_epi32(_sum22, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr2l, _rrr2h, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum33 = _mm512_add_epi32(_sum33, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr3l, _rrr3h, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum00 = _mm512_add_epi32(_sum00, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(2, 2, 2, 2)), _w1));
-                        _sum11 = _mm512_add_epi32(_sum11, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(2, 2, 2, 2)), _w1));
-                        _sum22 = _mm512_add_epi32(_sum22, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr2l, _rrr2h, _MM_SHUFFLE(2, 2, 2, 2)), _w1));
-                        _sum33 = _mm512_add_epi32(_sum33, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr3l, _rrr3h, _MM_SHUFFLE(2, 2, 2, 2)), _w1));
-                        _sum00 = _mm512_add_epi32(_sum00, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(1, 1, 1, 1)), _w2));
-                        _sum11 = _mm512_add_epi32(_sum11, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(1, 1, 1, 1)), _w2));
-                        _sum22 = _mm512_add_epi32(_sum22, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr2l, _rrr2h, _MM_SHUFFLE(1, 1, 1, 1)), _w2));
-                        _sum33 = _mm512_add_epi32(_sum33, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr3l, _rrr3h, _MM_SHUFFLE(1, 1, 1, 1)), _w2));
-                        _sum00 = _mm512_add_epi32(_sum00, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-                        _sum11 = _mm512_add_epi32(_sum11, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-                        _sum22 = _mm512_add_epi32(_sum22, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr2l, _rrr2h, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-                        _sum33 = _mm512_add_epi32(_sum33, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr3l, _rrr3h, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-#endif // __AVX512VNNI__
-
-                        kptr += 128;
-                    }
-                }
-                _sum0 = _mm256_add_epi32(_sum0, _mm512_extracti64x4_epi64(_sum00, 0));
-                _sum0 = _mm256_add_epi32(_sum0, _mm512_extracti64x4_epi64(_sum00, 1));
-                _sum1 = _mm256_add_epi32(_sum1, _mm512_extracti64x4_epi64(_sum11, 0));
-                _sum1 = _mm256_add_epi32(_sum1, _mm512_extracti64x4_epi64(_sum11, 1));
-                _sum2 = _mm256_add_epi32(_sum2, _mm512_extracti64x4_epi64(_sum22, 0));
-                _sum2 = _mm256_add_epi32(_sum2, _mm512_extracti64x4_epi64(_sum22, 1));
-                _sum3 = _mm256_add_epi32(_sum3, _mm512_extracti64x4_epi64(_sum33, 0));
-                _sum3 = _mm256_add_epi32(_sum3, _mm512_extracti64x4_epi64(_sum33, 1));
-            }
-#endif // __AVX512F__
-            for (; q + 7 < inch; q += 8)
-            {
-                const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-                const signed char* r2 = bottom_blob.channel(q / elempack).row<const signed char>(i2 * stride_h) + j2 * stride_w * elempack;
-                const signed char* r3 = bottom_blob.channel(q / elempack).row<const signed char>(i3 * stride_h) + j3 * stride_w * elempack;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-                    const signed char* r2s = r2 + space_ofs[k];
-                    const signed char* r3s = r3 + space_ofs[k];
-
-                    __m128i _r0;
-                    __m128i _r1;
-                    __m128i _r2;
-                    __m128i _r3;
-                    if (elempack == 8)
-                    {
-                        _r0 = _mm_loadl_epi64((const __m128i*)r0s);
-                        _r1 = _mm_loadl_epi64((const __m128i*)r1s);
-                        _r2 = _mm_loadl_epi64((const __m128i*)r2s);
-                        _r3 = _mm_loadl_epi64((const __m128i*)r3s);
-                    }
-                    else // if (elempack == 1)
-                    {
-#if __AVX2__
-                        __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                        _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-#if __AVX512F__
-                        _r0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1));
-                        _r1 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1));
-                        _r2 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r2s, _vindex, 1));
-                        _r3 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r3s, _vindex, 1));
-#else
-                        __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-                        __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-                        __m256i _val0_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1), _sindex88);
-                        __m256i _val1_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1), _sindex88);
-                        __m256i _val2_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r2s, _vindex, 1), _sindex88);
-                        __m256i _val3_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r3s, _vindex, 1), _sindex88);
-                        _r0 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val0_32, 0), _mm256_extracti128_si256(_val0_32, 1));
-                        _r1 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val1_32, 0), _mm256_extracti128_si256(_val1_32, 1));
-                        _r2 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val2_32, 0), _mm256_extracti128_si256(_val2_32, 1));
-                        _r3 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val3_32, 0), _mm256_extracti128_si256(_val3_32, 1));
-#endif // __AVX512F__
-#else
-                        _r0 = _mm_setr_epi8(r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                        _r1 = _mm_setr_epi8(r1s[0], r1s[N], r1s[N * 2], r1s[N * 3], r1s[N * 4], r1s[N * 5], r1s[N * 6], r1s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                        _r2 = _mm_setr_epi8(r2s[0], r2s[N], r2s[N * 2], r2s[N * 3], r2s[N * 4], r2s[N * 5], r2s[N * 6], r2s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                        _r3 = _mm_setr_epi8(r3s[0], r3s[N], r3s[N * 2], r3s[N * 3], r3s[N * 4], r3s[N * 5], r3s[N * 6], r3s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-#endif // __AVX2__
-                    }
-
-                    _r0 = _mm_cvtepi8_epi16(_r0);
-                    _r1 = _mm_cvtepi8_epi16(_r1);
-                    _r2 = _mm_cvtepi8_epi16(_r2);
-                    _r3 = _mm_cvtepi8_epi16(_r3);
-
-                    __m256i _w01 = _mm256_load_si256((const __m256i*)kptr);
-                    __m256i _w23 = _mm256_load_si256((const __m256i*)(kptr + 32));
-                    __m256i _w0 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w01, 0));
-                    __m256i _w1 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w01, 1));
-                    __m256i _w2 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w23, 0));
-                    __m256i _w3 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w23, 1));
-
-                    // 01234567 -> 01010101 01010101
-                    __m256i _rr0 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r0), _r0, 1);
-                    __m256i _rr1 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r1), _r1, 1);
-                    __m256i _rr2 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r2), _r2, 1);
-                    __m256i _rr3 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r3), _r3, 1);
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                    _sum0 = _mm256_dpwssd_epi32(_sum0, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                    _sum1 = _mm256_dpwssd_epi32(_sum1, _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                    _sum2 = _mm256_dpwssd_epi32(_sum2, _mm256_shuffle_epi32(_rr2, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                    _sum3 = _mm256_dpwssd_epi32(_sum3, _mm256_shuffle_epi32(_rr3, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                    _sum0 = _mm256_dpwssd_epi32(_sum0, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(1, 1, 1, 1)), _w1);
-                    _sum1 = _mm256_dpwssd_epi32(_sum1, _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(1, 1, 1, 1)), _w1);
-                    _sum2 = _mm256_dpwssd_epi32(_sum2, _mm256_shuffle_epi32(_rr2, _MM_SHUFFLE(1, 1, 1, 1)), _w1);
-                    _sum3 = _mm256_dpwssd_epi32(_sum3, _mm256_shuffle_epi32(_rr3, _MM_SHUFFLE(1, 1, 1, 1)), _w1);
-                    _sum0 = _mm256_dpwssd_epi32(_sum0, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(2, 2, 2, 2)), _w2);
-                    _sum1 = _mm256_dpwssd_epi32(_sum1, _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(2, 2, 2, 2)), _w2);
-                    _sum2 = _mm256_dpwssd_epi32(_sum2, _mm256_shuffle_epi32(_rr2, _MM_SHUFFLE(2, 2, 2, 2)), _w2);
-                    _sum3 = _mm256_dpwssd_epi32(_sum3, _mm256_shuffle_epi32(_rr3, _MM_SHUFFLE(2, 2, 2, 2)), _w2);
-                    _sum0 = _mm256_dpwssd_epi32(_sum0, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(3, 3, 3, 3)), _w3);
-                    _sum1 = _mm256_dpwssd_epi32(_sum1, _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(3, 3, 3, 3)), _w3);
-                    _sum2 = _mm256_dpwssd_epi32(_sum2, _mm256_shuffle_epi32(_rr2, _MM_SHUFFLE(3, 3, 3, 3)), _w3);
-                    _sum3 = _mm256_dpwssd_epi32(_sum3, _mm256_shuffle_epi32(_rr3, _MM_SHUFFLE(3, 3, 3, 3)), _w3);
-#else
-                    _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                    _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                    _sum2 = _mm256_add_epi32(_sum2, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr2, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                    _sum3 = _mm256_add_epi32(_sum3, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr3, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                    _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(1, 1, 1, 1)), _w1));
-                    _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(1, 1, 1, 1)), _w1));
-                    _sum2 = _mm256_add_epi32(_sum2, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr2, _MM_SHUFFLE(1, 1, 1, 1)), _w1));
-                    _sum3 = _mm256_add_epi32(_sum3, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr3, _MM_SHUFFLE(1, 1, 1, 1)), _w1));
-                    _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(2, 2, 2, 2)), _w2));
-                    _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(2, 2, 2, 2)), _w2));
-                    _sum2 = _mm256_add_epi32(_sum2, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr2, _MM_SHUFFLE(2, 2, 2, 2)), _w2));
-                    _sum3 = _mm256_add_epi32(_sum3, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr3, _MM_SHUFFLE(2, 2, 2, 2)), _w2));
-                    _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-                    _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-                    _sum2 = _mm256_add_epi32(_sum2, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr2, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-                    _sum3 = _mm256_add_epi32(_sum3, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr3, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-#endif
-
-                    kptr += 64;
-                }
-            }
-            for (; q + 1 < inch; q += 2)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-                const signed char* r2 = bottom_blob.channel(q).row<const signed char>(i2 * stride_h) + j2 * stride_w;
-                const signed char* r3 = bottom_blob.channel(q).row<const signed char>(i3 * stride_h) + j3 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-                    const signed char* r2s = r2 + space_ofs[k];
-                    const signed char* r3s = r3 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _r0 = _mm_setr_epi16(r0s[0], r0s[N], 0, 0, 0, 0, 0, 0);
-                        __m128i _r1 = _mm_setr_epi16(r1s[0], r1s[N], 0, 0, 0, 0, 0, 0);
-                        __m128i _r2 = _mm_setr_epi16(r2s[0], r2s[N], 0, 0, 0, 0, 0, 0);
-                        __m128i _r3 = _mm_setr_epi16(r3s[0], r3s[N], 0, 0, 0, 0, 0, 0);
-                        __m256i _rr0 = _mm256_broadcastd_epi32(_r0);
-                        __m256i _rr1 = _mm256_broadcastd_epi32(_r1);
-                        __m256i _rr2 = _mm256_broadcastd_epi32(_r2);
-                        __m256i _rr3 = _mm256_broadcastd_epi32(_r3);
-
-                        __m256i _w = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)kptr));
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum0 = _mm256_dpwssd_epi32(_sum0, _rr0, _w);
-                        _sum1 = _mm256_dpwssd_epi32(_sum1, _rr1, _w);
-                        _sum2 = _mm256_dpwssd_epi32(_sum2, _rr2, _w);
-                        _sum3 = _mm256_dpwssd_epi32(_sum3, _rr3, _w);
-#else
-                        _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_rr0, _w));
-                        _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_rr1, _w));
-                        _sum2 = _mm256_add_epi32(_sum2, _mm256_madd_epi16(_rr2, _w));
-                        _sum3 = _mm256_add_epi32(_sum3, _mm256_madd_epi16(_rr3, _w));
-#endif
-
-                        kptr += 16;
-                    }
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-                const signed char* r2 = bottom_blob.channel(q).row<const signed char>(i2 * stride_h) + j2 * stride_w;
-                const signed char* r3 = bottom_blob.channel(q).row<const signed char>(i3 * stride_h) + j3 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-                    const signed char* r2s = r2 + space_ofs[k];
-                    const signed char* r3s = r3 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _r0 = _mm_set1_epi16(r0s[0]);
-                        __m128i _r1 = _mm_set1_epi16(r1s[0]);
-                        __m128i _r2 = _mm_set1_epi16(r2s[0]);
-                        __m128i _r3 = _mm_set1_epi16(r3s[0]);
-                        __m128i _w = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*)kptr));
-                        __m256i _s0 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_r0, _w));
-                        __m256i _s1 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_r1, _w));
-                        __m256i _s2 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_r2, _w));
-                        __m256i _s3 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_r3, _w));
-                        _sum0 = _mm256_add_epi32(_sum0, _s0);
-                        _sum1 = _mm256_add_epi32(_sum1, _s1);
-                        _sum2 = _mm256_add_epi32(_sum2, _s2);
-                        _sum3 = _mm256_add_epi32(_sum3, _s3);
-
-                        kptr += 8;
-                    }
-                }
-            }
-
-            if (out_elempack == 8)
-            {
-                _mm256_store_si256((__m256i*)outptr, _sum0);
-                _mm256_store_si256((__m256i*)(outptr + 8), _sum1);
-                _mm256_store_si256((__m256i*)(outptr + 16), _sum2);
-                _mm256_store_si256((__m256i*)(outptr + 24), _sum3);
-                outptr += 32;
-            }
-            if (out_elempack == 4)
-            {
-                _mm_store_si128((__m128i*)outptr, _mm256_extracti128_si256(_sum0, 0));
-                _mm_store_si128((__m128i*)(outptr + M), _mm256_extracti128_si256(_sum0, 1));
-                _mm_store_si128((__m128i*)(outptr + 4), _mm256_extracti128_si256(_sum1, 0));
-                _mm_store_si128((__m128i*)(outptr + M + 4), _mm256_extracti128_si256(_sum1, 1));
-                _mm_store_si128((__m128i*)(outptr + 8), _mm256_extracti128_si256(_sum2, 0));
-                _mm_store_si128((__m128i*)(outptr + M + 8), _mm256_extracti128_si256(_sum2, 1));
-                _mm_store_si128((__m128i*)(outptr + 12), _mm256_extracti128_si256(_sum3, 0));
-                _mm_store_si128((__m128i*)(outptr + M + 12), _mm256_extracti128_si256(_sum3, 1));
-                outptr += 16;
-            }
-            if (out_elempack == 1)
-            {
-#if __AVX512F__
-                __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(M));
-                _mm256_i32scatter_epi32(outptr, _vindex, _sum0, sizeof(int));
-                _mm256_i32scatter_epi32(outptr + 1, _vindex, _sum1, sizeof(int));
-                _mm256_i32scatter_epi32(outptr + 2, _vindex, _sum2, sizeof(int));
-                _mm256_i32scatter_epi32(outptr + 3, _vindex, _sum3, sizeof(int));
-#else
-                int sum0[8];
-                int sum1[8];
-                int sum2[8];
-                int sum3[8];
-                _mm256_storeu_si256((__m256i*)sum0, _sum0);
-                _mm256_storeu_si256((__m256i*)sum1, _sum1);
-                _mm256_storeu_si256((__m256i*)sum2, _sum2);
-                _mm256_storeu_si256((__m256i*)sum3, _sum3);
-
-                outptr[0] = sum0[0];
-                outptr[1] = sum1[0];
-                outptr[2] = sum2[0];
-                outptr[3] = sum3[0];
-                outptr[M] = sum0[1];
-                outptr[M + 1] = sum1[1];
-                outptr[M + 2] = sum2[1];
-                outptr[M + 3] = sum3[1];
-                outptr[M * 2] = sum0[2];
-                outptr[M * 2 + 1] = sum1[2];
-                outptr[M * 2 + 2] = sum2[2];
-                outptr[M * 2 + 3] = sum3[2];
-                outptr[M * 3] = sum0[3];
-                outptr[M * 3 + 1] = sum1[3];
-                outptr[M * 3 + 2] = sum2[3];
-                outptr[M * 3 + 3] = sum3[3];
-                outptr[M * 4] = sum0[4];
-                outptr[M * 4 + 1] = sum1[4];
-                outptr[M * 4 + 2] = sum2[4];
-                outptr[M * 4 + 3] = sum3[4];
-                outptr[M * 5] = sum0[5];
-                outptr[M * 5 + 1] = sum1[5];
-                outptr[M * 5 + 2] = sum2[5];
-                outptr[M * 5 + 3] = sum3[5];
-                outptr[M * 6] = sum0[6];
-                outptr[M * 6 + 1] = sum1[6];
-                outptr[M * 6 + 2] = sum2[6];
-                outptr[M * 6 + 3] = sum3[6];
-                outptr[M * 7] = sum0[7];
-                outptr[M * 7 + 1] = sum1[7];
-                outptr[M * 7 + 2] = sum2[7];
-                outptr[M * 7 + 3] = sum3[7];
-#endif // __AVX512F__
-                outptr += 4;
-            }
-        }
-        for (; ij + 1 < outw * outh; ij += 2)
-        {
-            const int i0 = ij / outw;
-            const int i1 = (ij + 1) / outw;
-            const int j0 = ij % outw;
-            const int j1 = (ij + 1) % outw;
-
-            __m256i _sum0 = _mm256_setzero_si256();
-            __m256i _sum1 = _mm256_setzero_si256();
-            __m256i _sum2 = _mm256_setzero_si256();
-            __m256i _sum3 = _mm256_setzero_si256();
-
-#if __AVX512F__
-            const signed char* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 8);
-#endif
-
-            int q = 0;
-#if __AVX512F__
-            {
-                __m512i _sum00 = _mm512_setzero_si512();
-                __m512i _sum11 = _mm512_setzero_si512();
-                __m512i _sum22 = _mm512_setzero_si512();
-                __m512i _sum33 = _mm512_setzero_si512();
-                for (; q + 15 < inch; q += 16)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-
-                        __m128i _r0;
-                        __m128i _r1;
-                        if (elempack == 16)
-                        {
-                            _r0 = _mm_load_si128((const __m128i*)r0s);
-                            _r1 = _mm_load_si128((const __m128i*)r1s);
-                        }
-                        else if (elempack == 8)
-                        {
-                            __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            __m128i _t1 = _mm_loadl_epi64((const __m128i*)(r0s + N));
-                            __m128i _t2 = _mm_loadl_epi64((const __m128i*)r1s);
-                            __m128i _t3 = _mm_loadl_epi64((const __m128i*)(r1s + N));
-                            _r0 = _mm_unpacklo_epi64(_t0, _t1);
-                            _r1 = _mm_unpacklo_epi64(_t2, _t3);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                            _r0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0s), 1));
-                            _r1 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r1s), 1));
-                        }
-
-                        __m256i _rr0 = _mm256_cvtepi8_epi16(_r0);
-                        __m256i _rr1 = _mm256_cvtepi8_epi16(_r1);
-
-                        __m512i _w01 = _mm512_load_si512((const __m512i*)kptr);
-                        __m512i _w23 = _mm512_load_si512((const __m512i*)(kptr + 64));
-                        __m512i _w0 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 0));
-                        __m512i _w1 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 1));
-                        __m512i _w2 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 0));
-                        __m512i _w3 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 1));
-
-                        // 01234567 89abcdef -> 01010101 01010101 23232323 23232323
-                        __m512i _rrr0 = _mm512_inserti64x4(_mm512_castsi256_si512(_rr0), _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(1, 0, 3, 2)), 1);
-                        __m512i _rrr1 = _mm512_inserti64x4(_mm512_castsi256_si512(_rr1), _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(1, 0, 3, 2)), 1);
-
-                        _rrr0 = _mm512_unpacklo_epi32(_rrr0, _rrr0);
-                        _rrr1 = _mm512_unpacklo_epi32(_rrr1, _rrr1);
-
-                        __m512i _rrr0l = _mm512_unpacklo_epi64(_rrr0, _rrr0);
-                        __m512i _rrr0h = _mm512_unpackhi_epi64(_rrr0, _rrr0);
-                        __m512i _rrr1l = _mm512_unpacklo_epi64(_rrr1, _rrr1);
-                        __m512i _rrr1h = _mm512_unpackhi_epi64(_rrr1, _rrr1);
-
-#if __AVX512VNNI__
-                        _sum00 = _mm512_dpwssd_epi32(_sum00, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                        _sum11 = _mm512_dpwssd_epi32(_sum11, _mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                        _sum22 = _mm512_dpwssd_epi32(_sum22, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(2, 2, 2, 2)), _w1);
-                        _sum33 = _mm512_dpwssd_epi32(_sum33, _mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(2, 2, 2, 2)), _w1);
-                        _sum00 = _mm512_dpwssd_epi32(_sum00, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(1, 1, 1, 1)), _w2);
-                        _sum11 = _mm512_dpwssd_epi32(_sum11, _mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(1, 1, 1, 1)), _w2);
-                        _sum22 = _mm512_dpwssd_epi32(_sum22, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(3, 3, 3, 3)), _w3);
-                        _sum33 = _mm512_dpwssd_epi32(_sum33, _mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(3, 3, 3, 3)), _w3);
-#else
-                        _sum00 = _mm512_add_epi32(_sum00, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum11 = _mm512_add_epi32(_sum11, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum22 = _mm512_add_epi32(_sum22, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(2, 2, 2, 2)), _w1));
-                        _sum33 = _mm512_add_epi32(_sum33, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(2, 2, 2, 2)), _w1));
-                        _sum00 = _mm512_add_epi32(_sum00, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(1, 1, 1, 1)), _w2));
-                        _sum11 = _mm512_add_epi32(_sum11, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(1, 1, 1, 1)), _w2));
-                        _sum22 = _mm512_add_epi32(_sum22, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-                        _sum33 = _mm512_add_epi32(_sum33, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-#endif // __AVX512VNNI__
-
-                        kptr += 128;
-                    }
-                }
-                _sum0 = _mm256_add_epi32(_sum0, _mm512_extracti64x4_epi64(_sum00, 0));
-                _sum0 = _mm256_add_epi32(_sum0, _mm512_extracti64x4_epi64(_sum00, 1));
-                _sum1 = _mm256_add_epi32(_sum1, _mm512_extracti64x4_epi64(_sum11, 0));
-                _sum1 = _mm256_add_epi32(_sum1, _mm512_extracti64x4_epi64(_sum11, 1));
-                _sum2 = _mm256_add_epi32(_sum2, _mm512_extracti64x4_epi64(_sum22, 0));
-                _sum2 = _mm256_add_epi32(_sum2, _mm512_extracti64x4_epi64(_sum22, 1));
-                _sum3 = _mm256_add_epi32(_sum3, _mm512_extracti64x4_epi64(_sum33, 0));
-                _sum3 = _mm256_add_epi32(_sum3, _mm512_extracti64x4_epi64(_sum33, 1));
-            }
-#endif // __AVX512F__
-            for (; q + 7 < inch; q += 8)
-            {
-                const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    __m128i _r0;
-                    __m128i _r1;
-                    if (elempack == 8)
-                    {
-                        _r0 = _mm_loadl_epi64((const __m128i*)r0s);
-                        _r1 = _mm_loadl_epi64((const __m128i*)r1s);
-                    }
-                    else // if (elempack == 1)
-                    {
-#if __AVX2__
-                        __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                        _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-#if __AVX512F__
-                        _r0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1));
-                        _r1 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1));
-#else
-                        __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-                        __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-                        __m256i _val0_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1), _sindex88);
-                        __m256i _val1_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1), _sindex88);
-                        _r0 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val0_32, 0), _mm256_extracti128_si256(_val0_32, 1));
-                        _r1 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val1_32, 0), _mm256_extracti128_si256(_val1_32, 1));
-#endif // __AVX512F__
-#else
-                        _r0 = _mm_setr_epi8(r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                        _r1 = _mm_setr_epi8(r1s[0], r1s[N], r1s[N * 2], r1s[N * 3], r1s[N * 4], r1s[N * 5], r1s[N * 6], r1s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-#endif // __AVX2__
-                    }
-
-                    _r0 = _mm_cvtepi8_epi16(_r0);
-                    _r1 = _mm_cvtepi8_epi16(_r1);
-
-                    __m256i _w01 = _mm256_load_si256((const __m256i*)kptr);
-                    __m256i _w23 = _mm256_load_si256((const __m256i*)(kptr + 32));
-                    __m256i _w0 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w01, 0));
-                    __m256i _w1 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w01, 1));
-                    __m256i _w2 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w23, 0));
-                    __m256i _w3 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w23, 1));
-
-                    // 01234567 -> 01010101 01010101
-                    __m256i _rr0 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r0), _r0, 1);
-                    __m256i _rr1 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r1), _r1, 1);
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                    _sum0 = _mm256_dpwssd_epi32(_sum0, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                    _sum1 = _mm256_dpwssd_epi32(_sum1, _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                    _sum2 = _mm256_dpwssd_epi32(_sum2, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(1, 1, 1, 1)), _w1);
-                    _sum3 = _mm256_dpwssd_epi32(_sum3, _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(1, 1, 1, 1)), _w1);
-                    _sum0 = _mm256_dpwssd_epi32(_sum0, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(2, 2, 2, 2)), _w2);
-                    _sum1 = _mm256_dpwssd_epi32(_sum1, _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(2, 2, 2, 2)), _w2);
-                    _sum2 = _mm256_dpwssd_epi32(_sum2, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(3, 3, 3, 3)), _w3);
-                    _sum3 = _mm256_dpwssd_epi32(_sum3, _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(3, 3, 3, 3)), _w3);
-#else
-                    _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                    _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                    _sum2 = _mm256_add_epi32(_sum2, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(1, 1, 1, 1)), _w1));
-                    _sum3 = _mm256_add_epi32(_sum3, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(1, 1, 1, 1)), _w1));
-                    _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(2, 2, 2, 2)), _w2));
-                    _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(2, 2, 2, 2)), _w2));
-                    _sum2 = _mm256_add_epi32(_sum2, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-                    _sum3 = _mm256_add_epi32(_sum3, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-#endif
-
-                    kptr += 64;
-                }
-            }
-            for (; q + 1 < inch; q += 2)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _r0 = _mm_setr_epi16(r0s[0], r0s[N], 0, 0, 0, 0, 0, 0);
-                        __m128i _r1 = _mm_setr_epi16(r1s[0], r1s[N], 0, 0, 0, 0, 0, 0);
-                        __m256i _rr0 = _mm256_broadcastd_epi32(_r0);
-                        __m256i _rr1 = _mm256_broadcastd_epi32(_r1);
-
-                        __m256i _w = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)kptr));
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum0 = _mm256_dpwssd_epi32(_sum0, _rr0, _w);
-                        _sum1 = _mm256_dpwssd_epi32(_sum1, _rr1, _w);
-#else
-                        _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_rr0, _w));
-                        _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_rr1, _w));
-#endif
-
-                        kptr += 16;
-                    }
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _r0 = _mm_set1_epi16(r0s[0]);
-                        __m128i _r1 = _mm_set1_epi16(r1s[0]);
-                        __m128i _w = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*)kptr));
-                        __m256i _s0 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_r0, _w));
-                        __m256i _s1 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_r1, _w));
-                        _sum0 = _mm256_add_epi32(_sum0, _s0);
-                        _sum1 = _mm256_add_epi32(_sum1, _s1);
-
-                        kptr += 8;
-                    }
-                }
-            }
-
-            _sum0 = _mm256_add_epi32(_sum0, _sum2);
-            _sum1 = _mm256_add_epi32(_sum1, _sum3);
-
-            if (out_elempack == 8)
-            {
-                _mm256_store_si256((__m256i*)outptr, _sum0);
-                _mm256_store_si256((__m256i*)(outptr + 8), _sum1);
-                outptr += 16;
-            }
-            if (out_elempack == 4)
-            {
-                _mm_store_si128((__m128i*)outptr, _mm256_extracti128_si256(_sum0, 0));
-                _mm_store_si128((__m128i*)(outptr + M), _mm256_extracti128_si256(_sum0, 1));
-                _mm_store_si128((__m128i*)(outptr + 4), _mm256_extracti128_si256(_sum1, 0));
-                _mm_store_si128((__m128i*)(outptr + M + 4), _mm256_extracti128_si256(_sum1, 1));
-                outptr += 8;
-            }
-            if (out_elempack == 1)
-            {
-#if __AVX512F__
-                __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(M));
-                _mm256_i32scatter_epi32(outptr, _vindex, _sum0, sizeof(int));
-                _mm256_i32scatter_epi32(outptr + 1, _vindex, _sum1, sizeof(int));
-#else
-                int sum0[8];
-                int sum1[8];
-                _mm256_storeu_si256((__m256i*)sum0, _sum0);
-                _mm256_storeu_si256((__m256i*)sum1, _sum1);
-
-                outptr[0] = sum0[0];
-                outptr[1] = sum1[0];
-                outptr[M] = sum0[1];
-                outptr[M + 1] = sum1[1];
-                outptr[M * 2] = sum0[2];
-                outptr[M * 2 + 1] = sum1[2];
-                outptr[M * 3] = sum0[3];
-                outptr[M * 3 + 1] = sum1[3];
-                outptr[M * 4] = sum0[4];
-                outptr[M * 4 + 1] = sum1[4];
-                outptr[M * 5] = sum0[5];
-                outptr[M * 5 + 1] = sum1[5];
-                outptr[M * 6] = sum0[6];
-                outptr[M * 6 + 1] = sum1[6];
-                outptr[M * 7] = sum0[7];
-                outptr[M * 7 + 1] = sum1[7];
-#endif // __AVX512F__
-                outptr += 2;
-            }
-        }
-        for (; ij < outw * outh; ij++)
-        {
-            const int i = ij / outw;
-            const int j = ij % outw;
-
-            __m256i _sum0 = _mm256_setzero_si256();
-
-#if __AVX512F__
-            const signed char* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 8);
-#endif
-
-            int q = 0;
-#if __AVX512F__
-            {
-                __m512i _sum00 = _mm512_setzero_si512();
-                __m512i _sum11 = _mm512_setzero_si512();
-                __m512i _sum22 = _mm512_setzero_si512();
-                __m512i _sum33 = _mm512_setzero_si512();
-                for (; q + 15 < inch; q += 16)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-
-                        __m128i _r0;
-                        if (elempack == 16)
-                        {
-                            _r0 = _mm_load_si128((const __m128i*)r0s);
-                        }
-                        else if (elempack == 8)
-                        {
-                            __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            __m128i _t1 = _mm_loadl_epi64((const __m128i*)(r0s + N));
-                            _r0 = _mm_unpacklo_epi64(_t0, _t1);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                            _r0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0s), 1));
-                        }
-
-                        __m256i _rr0 = _mm256_cvtepi8_epi16(_r0);
-
-                        __m512i _w01 = _mm512_load_si512((const __m512i*)kptr);
-                        __m512i _w23 = _mm512_load_si512((const __m512i*)(kptr + 64));
-                        __m512i _w0 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 0));
-                        __m512i _w1 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w01, 1));
-                        __m512i _w2 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 0));
-                        __m512i _w3 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w23, 1));
-
-                        // 01234567 89abcdef -> 01010101 01010101 23232323 23232323
-                        __m512i _rrr0 = _mm512_inserti64x4(_mm512_castsi256_si512(_rr0), _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(1, 0, 3, 2)), 1);
-
-                        _rrr0 = _mm512_unpacklo_epi32(_rrr0, _rrr0);
-
-                        __m512i _rrr0l = _mm512_unpacklo_epi64(_rrr0, _rrr0);
-                        __m512i _rrr0h = _mm512_unpackhi_epi64(_rrr0, _rrr0);
-
-#if __AVX512VNNI__
-                        _sum00 = _mm512_dpwssd_epi32(_sum00, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                        _sum11 = _mm512_dpwssd_epi32(_sum11, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(2, 2, 2, 2)), _w1);
-                        _sum22 = _mm512_dpwssd_epi32(_sum22, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(1, 1, 1, 1)), _w2);
-                        _sum33 = _mm512_dpwssd_epi32(_sum33, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(3, 3, 3, 3)), _w3);
-#else
-                        _sum00 = _mm512_add_epi32(_sum00, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum11 = _mm512_add_epi32(_sum11, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(2, 2, 2, 2)), _w1));
-                        _sum22 = _mm512_add_epi32(_sum22, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(1, 1, 1, 1)), _w2));
-                        _sum33 = _mm512_add_epi32(_sum33, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-#endif // __AVX512VNNI__
-
-                        kptr += 128;
-                    }
-                }
-                _sum00 = _mm512_add_epi32(_sum00, _sum11);
-                _sum22 = _mm512_add_epi32(_sum22, _sum33);
-                _sum00 = _mm512_add_epi32(_sum00, _sum22);
-                _sum0 = _mm256_add_epi32(_sum0, _mm512_extracti64x4_epi64(_sum00, 0));
-                _sum0 = _mm256_add_epi32(_sum0, _mm512_extracti64x4_epi64(_sum00, 1));
-            }
-#endif // __AVX512F__
-            {
-                __m256i _sum1 = _mm256_setzero_si256();
-                __m256i _sum2 = _mm256_setzero_si256();
-                __m256i _sum3 = _mm256_setzero_si256();
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-
-                        __m128i _r0;
-                        if (elempack == 8)
-                        {
-                            _r0 = _mm_loadl_epi64((const __m128i*)r0s);
-                        }
-                        else // if (elempack == 1)
-                        {
-#if __AVX2__
-                            __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                            _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-#if __AVX512F__
-                            _r0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1));
-#else
-                            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-                            __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-                            __m256i _val32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1), _sindex88);
-                            _r0 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val32, 0), _mm256_extracti128_si256(_val32, 1));
-#endif // __AVX512F__
-#else
-                            _r0 = _mm_setr_epi8(r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-#endif // __AVX2__
-                        }
-
-                        _r0 = _mm_cvtepi8_epi16(_r0);
-
-                        __m256i _w01 = _mm256_load_si256((const __m256i*)kptr);
-                        __m256i _w23 = _mm256_load_si256((const __m256i*)(kptr + 32));
-                        __m256i _w0 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w01, 0));
-                        __m256i _w1 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w01, 1));
-                        __m256i _w2 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w23, 0));
-                        __m256i _w3 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w23, 1));
-
-                        // 01234567 -> 01010101 01010101
-                        __m256i _rr0 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r0), _r0, 1);
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum0 = _mm256_dpwssd_epi32(_sum0, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                        _sum1 = _mm256_dpwssd_epi32(_sum1, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(1, 1, 1, 1)), _w1);
-                        _sum2 = _mm256_dpwssd_epi32(_sum2, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(2, 2, 2, 2)), _w2);
-                        _sum3 = _mm256_dpwssd_epi32(_sum3, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(3, 3, 3, 3)), _w3);
-#else
-                        _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(1, 1, 1, 1)), _w1));
-                        _sum2 = _mm256_add_epi32(_sum2, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(2, 2, 2, 2)), _w2));
-                        _sum3 = _mm256_add_epi32(_sum3, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-#endif
-
-                        kptr += 64;
-                    }
-                }
-                _sum0 = _mm256_add_epi32(_sum0, _sum1);
-                _sum2 = _mm256_add_epi32(_sum2, _sum3);
-                _sum0 = _mm256_add_epi32(_sum0, _sum2);
-            }
-            for (; q + 1 < inch; q += 2)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i * stride_h) + j * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _val0 = _mm_setr_epi16(r0s[0], r0s[N], 0, 0, 0, 0, 0, 0);
-                        __m256i _val = _mm256_broadcastd_epi32(_val0);
-
-                        __m256i _w = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)kptr));
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum0 = _mm256_dpwssd_epi32(_sum0, _val, _w);
-#else
-                        _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_val, _w));
-#endif
-
-                        kptr += 16;
-                    }
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i * stride_h) + j * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _val = _mm_set1_epi16(r0s[0]);
-                        __m128i _w = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*)kptr));
-                        __m256i _s0 = _mm256_cvtepi16_epi32(_mm_mullo_epi16(_val, _w));
-                        _sum0 = _mm256_add_epi32(_sum0, _s0);
-
-                        kptr += 8;
-                    }
-                }
-            }
-
-            if (out_elempack == 8)
-            {
-                _mm256_store_si256((__m256i*)outptr, _sum0);
-                outptr += 8;
-            }
-            if (out_elempack == 4)
-            {
-                _mm_store_si128((__m128i*)outptr, _mm256_extracti128_si256(_sum0, 0));
-                _mm_store_si128((__m128i*)(outptr + M), _mm256_extracti128_si256(_sum0, 1));
-                outptr += 4;
-            }
-            if (out_elempack == 1)
-            {
-#if __AVX512F__
-                __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(M));
-                _mm256_i32scatter_epi32(outptr, _vindex, _sum0, sizeof(int));
-#else
-                int sum[8];
-                _mm256_storeu_si256((__m256i*)sum, _sum0);
-
-                outptr[0] = sum[0];
-                outptr[M] = sum[1];
-                outptr[M * 2] = sum[2];
-                outptr[M * 3] = sum[3];
-                outptr[M * 4] = sum[4];
-                outptr[M * 5] = sum[5];
-                outptr[M * 6] = sum[6];
-                outptr[M * 7] = sum[7];
-#endif // __AVX512F__
-                outptr += 1;
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 8;
-    nn_outch = (outch - remain_outch_start) / 4;
-#else // __AVX2__
-    nn_outch = (outch - remain_outch_start) / 4;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __AVX2__
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 4;
-
-        // shadowed variable for less openmp task args
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int N = bottom_blob.cstep * elempack;
-        const int M = top_blob.cstep * out_elempack;
-
-        int* outptr = top_blob.channel(p / out_elempack);
-
-        int ij = 0;
-        for (; ij + 3 < outw * outh; ij += 4)
-        {
-            const int i0 = ij / outw;
-            const int i1 = (ij + 1) / outw;
-            const int i2 = (ij + 2) / outw;
-            const int i3 = (ij + 3) / outw;
-            const int j0 = ij % outw;
-            const int j1 = (ij + 1) % outw;
-            const int j2 = (ij + 2) % outw;
-            const int j3 = (ij + 3) % outw;
-
-            __m128i _sum0 = _mm_setzero_si128();
-            __m128i _sum1 = _mm_setzero_si128();
-            __m128i _sum2 = _mm_setzero_si128();
-            __m128i _sum3 = _mm_setzero_si128();
-
-#if __AVX512F__
-            const signed char* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8 + (p % 8) / 4);
-#elif __AVX2__
-            const signed char* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 4);
-#endif
-
-            int q = 0;
-#if __AVX512F__
-            {
-                __m512i _sum00 = _mm512_setzero_si512();
-                __m512i _sum11 = _mm512_setzero_si512();
-                __m512i _sum22 = _mm512_setzero_si512();
-                __m512i _sum33 = _mm512_setzero_si512();
-                for (; q + 15 < inch; q += 16)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-                    const signed char* r2 = bottom_blob.channel(q / elempack).row<const signed char>(i2 * stride_h) + j2 * stride_w * elempack;
-                    const signed char* r3 = bottom_blob.channel(q / elempack).row<const signed char>(i3 * stride_h) + j3 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-                        const signed char* r2s = r2 + space_ofs[k];
-                        const signed char* r3s = r3 + space_ofs[k];
-
-                        __m128i _r0;
-                        __m128i _r1;
-                        __m128i _r2;
-                        __m128i _r3;
-                        if (elempack == 16)
-                        {
-                            _r0 = _mm_load_si128((const __m128i*)r0s);
-                            _r1 = _mm_load_si128((const __m128i*)r1s);
-                            _r2 = _mm_load_si128((const __m128i*)r2s);
-                            _r3 = _mm_load_si128((const __m128i*)r3s);
-                        }
-                        else if (elempack == 8)
-                        {
-                            __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            __m128i _t1 = _mm_loadl_epi64((const __m128i*)(r0s + N));
-                            __m128i _t2 = _mm_loadl_epi64((const __m128i*)r1s);
-                            __m128i _t3 = _mm_loadl_epi64((const __m128i*)(r1s + N));
-                            __m128i _t4 = _mm_loadl_epi64((const __m128i*)r2s);
-                            __m128i _t5 = _mm_loadl_epi64((const __m128i*)(r2s + N));
-                            __m128i _t6 = _mm_loadl_epi64((const __m128i*)r3s);
-                            __m128i _t7 = _mm_loadl_epi64((const __m128i*)(r3s + N));
-                            _r0 = _mm_unpacklo_epi64(_t0, _t1);
-                            _r1 = _mm_unpacklo_epi64(_t2, _t3);
-                            _r2 = _mm_unpacklo_epi64(_t4, _t5);
-                            _r3 = _mm_unpacklo_epi64(_t6, _t7);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                            _r0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0s), 1));
-                            _r1 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r1s), 1));
-                            _r2 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r2s), 1));
-                            _r3 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r3s), 1));
-                        }
-
-                        __m256i _rr0 = _mm256_cvtepi8_epi16(_r0);
-                        __m256i _rr1 = _mm256_cvtepi8_epi16(_r1);
-                        __m256i _rr2 = _mm256_cvtepi8_epi16(_r2);
-                        __m256i _rr3 = _mm256_cvtepi8_epi16(_r3);
-
-                        __m512i _w = _mm512_load_si512((const __m512i*)kptr);
-                        __m512i _w0 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w, 0));
-                        __m512i _w1 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w, 1));
-
-                        // 01234567 89abcdef -> 01010101 23232323 45454545 67676767
-                        _rr0 = _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(3, 1, 2, 0));
-                        _rr1 = _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(3, 1, 2, 0));
-                        _rr2 = _mm256_shuffle_epi32(_rr2, _MM_SHUFFLE(3, 1, 2, 0));
-                        _rr3 = _mm256_shuffle_epi32(_rr3, _MM_SHUFFLE(3, 1, 2, 0));
-
-                        __m512i _rrr0 = _mm512_inserti64x4(_mm512_castsi256_si512(_rr0), _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(1, 0, 3, 2)), 1);
-                        __m512i _rrr1 = _mm512_inserti64x4(_mm512_castsi256_si512(_rr1), _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(1, 0, 3, 2)), 1);
-                        __m512i _rrr2 = _mm512_inserti64x4(_mm512_castsi256_si512(_rr2), _mm256_shuffle_epi32(_rr2, _MM_SHUFFLE(1, 0, 3, 2)), 1);
-                        __m512i _rrr3 = _mm512_inserti64x4(_mm512_castsi256_si512(_rr3), _mm256_shuffle_epi32(_rr3, _MM_SHUFFLE(1, 0, 3, 2)), 1);
-
-                        _rrr0 = _mm512_unpacklo_epi32(_rrr0, _rrr0);
-                        _rrr1 = _mm512_unpacklo_epi32(_rrr1, _rrr1);
-                        _rrr2 = _mm512_unpacklo_epi32(_rrr2, _rrr2);
-                        _rrr3 = _mm512_unpacklo_epi32(_rrr3, _rrr3);
-
-                        __m512i _rrr0l = _mm512_unpacklo_epi64(_rrr0, _rrr0);
-                        __m512i _rrr0h = _mm512_unpackhi_epi64(_rrr0, _rrr0);
-                        __m512i _rrr1l = _mm512_unpacklo_epi64(_rrr1, _rrr1);
-                        __m512i _rrr1h = _mm512_unpackhi_epi64(_rrr1, _rrr1);
-                        __m512i _rrr2l = _mm512_unpacklo_epi64(_rrr2, _rrr2);
-                        __m512i _rrr2h = _mm512_unpackhi_epi64(_rrr2, _rrr2);
-                        __m512i _rrr3l = _mm512_unpacklo_epi64(_rrr3, _rrr3);
-                        __m512i _rrr3h = _mm512_unpackhi_epi64(_rrr3, _rrr3);
-
-#if __AVX512VNNI__
-                        _sum00 = _mm512_dpwssd_epi32(_sum00, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(2, 0, 2, 0)), _w0);
-                        _sum11 = _mm512_dpwssd_epi32(_sum11, _mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(2, 0, 2, 0)), _w0);
-                        _sum22 = _mm512_dpwssd_epi32(_sum22, _mm512_shuffle_i32x4(_rrr2l, _rrr2h, _MM_SHUFFLE(2, 0, 2, 0)), _w0);
-                        _sum33 = _mm512_dpwssd_epi32(_sum33, _mm512_shuffle_i32x4(_rrr3l, _rrr3h, _MM_SHUFFLE(2, 0, 2, 0)), _w0);
-                        _sum00 = _mm512_dpwssd_epi32(_sum00, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(3, 1, 3, 1)), _w1);
-                        _sum11 = _mm512_dpwssd_epi32(_sum11, _mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(3, 1, 3, 1)), _w1);
-                        _sum22 = _mm512_dpwssd_epi32(_sum22, _mm512_shuffle_i32x4(_rrr2l, _rrr2h, _MM_SHUFFLE(3, 1, 3, 1)), _w1);
-                        _sum33 = _mm512_dpwssd_epi32(_sum33, _mm512_shuffle_i32x4(_rrr3l, _rrr3h, _MM_SHUFFLE(3, 1, 3, 1)), _w1);
-#else
-                        _sum00 = _mm512_add_epi32(_sum00, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(2, 0, 2, 0)), _w0));
-                        _sum11 = _mm512_add_epi32(_sum11, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(2, 0, 2, 0)), _w0));
-                        _sum22 = _mm512_add_epi32(_sum22, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr2l, _rrr2h, _MM_SHUFFLE(2, 0, 2, 0)), _w0));
-                        _sum33 = _mm512_add_epi32(_sum33, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr3l, _rrr3h, _MM_SHUFFLE(2, 0, 2, 0)), _w0));
-                        _sum00 = _mm512_add_epi32(_sum00, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(3, 1, 3, 1)), _w1));
-                        _sum11 = _mm512_add_epi32(_sum11, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(3, 1, 3, 1)), _w1));
-                        _sum22 = _mm512_add_epi32(_sum22, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr2l, _rrr2h, _MM_SHUFFLE(3, 1, 3, 1)), _w1));
-                        _sum33 = _mm512_add_epi32(_sum33, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr3l, _rrr3h, _MM_SHUFFLE(3, 1, 3, 1)), _w1));
-#endif // __AVX512VNNI__
-
-                        kptr += 64;
-                    }
-                }
-                __m256i _ss0 = _mm256_add_epi32(_mm512_extracti64x4_epi64(_sum00, 0), _mm512_extracti64x4_epi64(_sum00, 1));
-                __m256i _ss1 = _mm256_add_epi32(_mm512_extracti64x4_epi64(_sum11, 0), _mm512_extracti64x4_epi64(_sum11, 1));
-                __m256i _ss2 = _mm256_add_epi32(_mm512_extracti64x4_epi64(_sum22, 0), _mm512_extracti64x4_epi64(_sum22, 1));
-                __m256i _ss3 = _mm256_add_epi32(_mm512_extracti64x4_epi64(_sum33, 0), _mm512_extracti64x4_epi64(_sum33, 1));
-                _sum0 = _mm_add_epi32(_sum0, _mm256_extracti128_si256(_ss0, 0));
-                _sum0 = _mm_add_epi32(_sum0, _mm256_extracti128_si256(_ss0, 1));
-                _sum1 = _mm_add_epi32(_sum1, _mm256_extracti128_si256(_ss1, 0));
-                _sum1 = _mm_add_epi32(_sum1, _mm256_extracti128_si256(_ss1, 1));
-                _sum2 = _mm_add_epi32(_sum2, _mm256_extracti128_si256(_ss2, 0));
-                _sum2 = _mm_add_epi32(_sum2, _mm256_extracti128_si256(_ss2, 1));
-                _sum3 = _mm_add_epi32(_sum3, _mm256_extracti128_si256(_ss3, 0));
-                _sum3 = _mm_add_epi32(_sum3, _mm256_extracti128_si256(_ss3, 1));
-            }
-#endif // __AVX512F__
-            {
-#if __AVX2__
-                __m256i _sum00 = _mm256_setzero_si256();
-                __m256i _sum11 = _mm256_setzero_si256();
-                __m256i _sum22 = _mm256_setzero_si256();
-                __m256i _sum33 = _mm256_setzero_si256();
-#endif
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-                    const signed char* r2 = bottom_blob.channel(q / elempack).row<const signed char>(i2 * stride_h) + j2 * stride_w * elempack;
-                    const signed char* r3 = bottom_blob.channel(q / elempack).row<const signed char>(i3 * stride_h) + j3 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-                        const signed char* r2s = r2 + space_ofs[k];
-                        const signed char* r3s = r3 + space_ofs[k];
-
-                        __m128i _r0;
-                        __m128i _r1;
-                        __m128i _r2;
-                        __m128i _r3;
-                        if (elempack == 8)
-                        {
-                            _r0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            _r1 = _mm_loadl_epi64((const __m128i*)r1s);
-                            _r2 = _mm_loadl_epi64((const __m128i*)r2s);
-                            _r3 = _mm_loadl_epi64((const __m128i*)r3s);
-                        }
-                        else // if (elempack == 1)
-                        {
-#if __AVX2__
-                            __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                            _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-#if __AVX512F__
-                            _r0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1));
-                            _r1 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1));
-                            _r2 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r2s, _vindex, 1));
-                            _r3 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r3s, _vindex, 1));
-#else
-                            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-                            __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-                            __m256i _val0_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1), _sindex88);
-                            __m256i _val1_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1), _sindex88);
-                            __m256i _val2_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r2s, _vindex, 1), _sindex88);
-                            __m256i _val3_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r3s, _vindex, 1), _sindex88);
-                            _r0 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val0_32, 0), _mm256_extracti128_si256(_val0_32, 1));
-                            _r1 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val1_32, 0), _mm256_extracti128_si256(_val1_32, 1));
-                            _r2 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val2_32, 0), _mm256_extracti128_si256(_val2_32, 1));
-                            _r3 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val3_32, 0), _mm256_extracti128_si256(_val3_32, 1));
-#endif // __AVX512F__
-#else
-                            _r0 = _mm_setr_epi8(r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                            _r1 = _mm_setr_epi8(r1s[0], r1s[N], r1s[N * 2], r1s[N * 3], r1s[N * 4], r1s[N * 5], r1s[N * 6], r1s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                            _r2 = _mm_setr_epi8(r2s[0], r2s[N], r2s[N * 2], r2s[N * 3], r2s[N * 4], r2s[N * 5], r2s[N * 6], r2s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                            _r3 = _mm_setr_epi8(r3s[0], r3s[N], r3s[N * 2], r3s[N * 3], r3s[N * 4], r3s[N * 5], r3s[N * 6], r3s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-#endif // __AVX2__
-                        }
-
-#if __SSE4_1__
-                        _r0 = _mm_cvtepi8_epi16(_r0);
-                        _r1 = _mm_cvtepi8_epi16(_r1);
-                        _r2 = _mm_cvtepi8_epi16(_r2);
-                        _r3 = _mm_cvtepi8_epi16(_r3);
-#else
-                        _r0 = _mm_unpacklo_epi8(_r0, _mm_cmpgt_epi8(_mm_setzero_si128(), _r0));
-                        _r1 = _mm_unpacklo_epi8(_r1, _mm_cmpgt_epi8(_mm_setzero_si128(), _r1));
-                        _r2 = _mm_unpacklo_epi8(_r2, _mm_cmpgt_epi8(_mm_setzero_si128(), _r2));
-                        _r3 = _mm_unpacklo_epi8(_r3, _mm_cmpgt_epi8(_mm_setzero_si128(), _r3));
-#endif
-
-#if __AVX2__
-                        __m256i _w = _mm256_load_si256((const __m256i*)kptr);
-                        __m256i _w0 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w, 0));
-                        __m256i _w1 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w, 1));
-
-                        // 01234567 -> 01010101 23232323
-                        __m256i _rr0 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r0), _mm_shuffle_epi32(_r0, _MM_SHUFFLE(2, 3, 0, 1)), 1);
-                        __m256i _rr1 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r1), _mm_shuffle_epi32(_r1, _MM_SHUFFLE(2, 3, 0, 1)), 1);
-                        __m256i _rr2 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r2), _mm_shuffle_epi32(_r2, _MM_SHUFFLE(2, 3, 0, 1)), 1);
-                        __m256i _rr3 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r3), _mm_shuffle_epi32(_r3, _MM_SHUFFLE(2, 3, 0, 1)), 1);
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum00 = _mm256_dpwssd_epi32(_sum00, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                        _sum11 = _mm256_dpwssd_epi32(_sum11, _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                        _sum22 = _mm256_dpwssd_epi32(_sum22, _mm256_shuffle_epi32(_rr2, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                        _sum33 = _mm256_dpwssd_epi32(_sum33, _mm256_shuffle_epi32(_rr3, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                        _sum00 = _mm256_dpwssd_epi32(_sum00, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(2, 2, 2, 2)), _w1);
-                        _sum11 = _mm256_dpwssd_epi32(_sum11, _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(2, 2, 2, 2)), _w1);
-                        _sum22 = _mm256_dpwssd_epi32(_sum22, _mm256_shuffle_epi32(_rr2, _MM_SHUFFLE(2, 2, 2, 2)), _w1);
-                        _sum33 = _mm256_dpwssd_epi32(_sum33, _mm256_shuffle_epi32(_rr3, _MM_SHUFFLE(2, 2, 2, 2)), _w1);
-#else
-                        _sum00 = _mm256_add_epi32(_sum00, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum11 = _mm256_add_epi32(_sum11, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum22 = _mm256_add_epi32(_sum22, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr2, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum33 = _mm256_add_epi32(_sum33, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr3, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum00 = _mm256_add_epi32(_sum00, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(2, 2, 2, 2)), _w1));
-                        _sum11 = _mm256_add_epi32(_sum11, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(2, 2, 2, 2)), _w1));
-                        _sum22 = _mm256_add_epi32(_sum22, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr2, _MM_SHUFFLE(2, 2, 2, 2)), _w1));
-                        _sum33 = _mm256_add_epi32(_sum33, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr3, _MM_SHUFFLE(2, 2, 2, 2)), _w1));
-#endif
-#else // __AVX2__
-                        __m128i _w01 = _mm_load_si128((const __m128i*)kptr);
-                        __m128i _w23 = _mm_load_si128((const __m128i*)(kptr + 16));
-                        __m128i _extw01 = _mm_cmpgt_epi8(_mm_setzero_si128(), _w01);
-                        __m128i _extw23 = _mm_cmpgt_epi8(_mm_setzero_si128(), _w23);
-                        __m128i _w0 = _mm_unpacklo_epi8(_w01, _extw01);
-                        __m128i _w1 = _mm_unpackhi_epi8(_w01, _extw01);
-                        __m128i _w2 = _mm_unpacklo_epi8(_w23, _extw23);
-                        __m128i _w3 = _mm_unpackhi_epi8(_w23, _extw23);
-
-                        // 01234567 -> 01010101
-#if __XOP__
-                        _sum0 = _mm_maddd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(0, 0, 0, 0)), _w0, _sum0);
-                        _sum1 = _mm_maddd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(0, 0, 0, 0)), _w0, _sum1);
-                        _sum2 = _mm_maddd_epi16(_mm_shuffle_epi32(_r2, _MM_SHUFFLE(0, 0, 0, 0)), _w0, _sum2);
-                        _sum3 = _mm_maddd_epi16(_mm_shuffle_epi32(_r3, _MM_SHUFFLE(0, 0, 0, 0)), _w0, _sum3);
-                        _sum0 = _mm_maddd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(1, 1, 1, 1)), _w1, _sum0);
-                        _sum1 = _mm_maddd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(1, 1, 1, 1)), _w1, _sum1);
-                        _sum2 = _mm_maddd_epi16(_mm_shuffle_epi32(_r2, _MM_SHUFFLE(1, 1, 1, 1)), _w1, _sum2);
-                        _sum3 = _mm_maddd_epi16(_mm_shuffle_epi32(_r3, _MM_SHUFFLE(1, 1, 1, 1)), _w1, _sum3);
-                        _sum0 = _mm_maddd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(2, 2, 2, 2)), _w2, _sum0);
-                        _sum1 = _mm_maddd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(2, 2, 2, 2)), _w2, _sum1);
-                        _sum2 = _mm_maddd_epi16(_mm_shuffle_epi32(_r2, _MM_SHUFFLE(2, 2, 2, 2)), _w2, _sum2);
-                        _sum3 = _mm_maddd_epi16(_mm_shuffle_epi32(_r3, _MM_SHUFFLE(2, 2, 2, 2)), _w2, _sum3);
-                        _sum0 = _mm_maddd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(3, 3, 3, 3)), _w3, _sum0);
-                        _sum1 = _mm_maddd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(3, 3, 3, 3)), _w3, _sum1);
-                        _sum2 = _mm_maddd_epi16(_mm_shuffle_epi32(_r2, _MM_SHUFFLE(3, 3, 3, 3)), _w3, _sum2);
-                        _sum3 = _mm_maddd_epi16(_mm_shuffle_epi32(_r3, _MM_SHUFFLE(3, 3, 3, 3)), _w3, _sum3);
-#else
-                        _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum2 = _mm_add_epi32(_sum2, _mm_madd_epi16(_mm_shuffle_epi32(_r2, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum3 = _mm_add_epi32(_sum3, _mm_madd_epi16(_mm_shuffle_epi32(_r3, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(1, 1, 1, 1)), _w1));
-                        _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(1, 1, 1, 1)), _w1));
-                        _sum2 = _mm_add_epi32(_sum2, _mm_madd_epi16(_mm_shuffle_epi32(_r2, _MM_SHUFFLE(1, 1, 1, 1)), _w1));
-                        _sum3 = _mm_add_epi32(_sum3, _mm_madd_epi16(_mm_shuffle_epi32(_r3, _MM_SHUFFLE(1, 1, 1, 1)), _w1));
-                        _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(2, 2, 2, 2)), _w2));
-                        _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(2, 2, 2, 2)), _w2));
-                        _sum2 = _mm_add_epi32(_sum2, _mm_madd_epi16(_mm_shuffle_epi32(_r2, _MM_SHUFFLE(2, 2, 2, 2)), _w2));
-                        _sum3 = _mm_add_epi32(_sum3, _mm_madd_epi16(_mm_shuffle_epi32(_r3, _MM_SHUFFLE(2, 2, 2, 2)), _w2));
-                        _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-                        _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-                        _sum2 = _mm_add_epi32(_sum2, _mm_madd_epi16(_mm_shuffle_epi32(_r2, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-                        _sum3 = _mm_add_epi32(_sum3, _mm_madd_epi16(_mm_shuffle_epi32(_r3, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-#endif // __XOP__
-#endif // __AVX2__
-
-                        kptr += 32;
-                    }
-                }
-#if __AVX2__
-                __m128i _ss0 = _mm_add_epi32(_mm256_extracti128_si256(_sum00, 0), _mm256_extracti128_si256(_sum00, 1));
-                __m128i _ss1 = _mm_add_epi32(_mm256_extracti128_si256(_sum11, 0), _mm256_extracti128_si256(_sum11, 1));
-                __m128i _ss2 = _mm_add_epi32(_mm256_extracti128_si256(_sum22, 0), _mm256_extracti128_si256(_sum22, 1));
-                __m128i _ss3 = _mm_add_epi32(_mm256_extracti128_si256(_sum33, 0), _mm256_extracti128_si256(_sum33, 1));
-                _sum0 = _mm_add_epi32(_sum0, _ss0);
-                _sum1 = _mm_add_epi32(_sum1, _ss1);
-                _sum2 = _mm_add_epi32(_sum2, _ss2);
-                _sum3 = _mm_add_epi32(_sum3, _ss3);
-#endif // __AVX2__
-            }
-            for (; q + 1 < inch; q += 2)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-                const signed char* r2 = bottom_blob.channel(q).row<const signed char>(i2 * stride_h) + j2 * stride_w;
-                const signed char* r3 = bottom_blob.channel(q).row<const signed char>(i3 * stride_h) + j3 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-                    const signed char* r2s = r2 + space_ofs[k];
-                    const signed char* r3s = r3 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _r0 = _mm_setr_epi16(r0s[0], r0s[N], r0s[0], r0s[N], r0s[0], r0s[N], r0s[0], r0s[N]);
-                        __m128i _r1 = _mm_setr_epi16(r1s[0], r1s[N], r1s[0], r1s[N], r1s[0], r1s[N], r1s[0], r1s[N]);
-                        __m128i _r2 = _mm_setr_epi16(r2s[0], r2s[N], r2s[0], r2s[N], r2s[0], r2s[N], r2s[0], r2s[N]);
-                        __m128i _r3 = _mm_setr_epi16(r3s[0], r3s[N], r3s[0], r3s[N], r3s[0], r3s[N], r3s[0], r3s[N]);
-
-                        __m128i _w = _mm_loadl_epi64((const __m128i*)kptr);
-#if __SSE4_1__
-                        _w = _mm_cvtepi8_epi16(_w);
-#else
-                        _w = _mm_unpacklo_epi8(_w, _mm_cmpgt_epi8(_mm_setzero_si128(), _w));
-#endif
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum0 = _mm_dpwssd_epi32(_sum0, _r0, _w);
-                        _sum1 = _mm_dpwssd_epi32(_sum1, _r1, _w);
-                        _sum2 = _mm_dpwssd_epi32(_sum2, _r2, _w);
-                        _sum3 = _mm_dpwssd_epi32(_sum3, _r3, _w);
-#elif __XOP__
-                        _sum0 = _mm_maddd_epi16(_r0, _w, _sum0);
-                        _sum1 = _mm_maddd_epi16(_r1, _w, _sum1);
-                        _sum2 = _mm_maddd_epi16(_r2, _w, _sum2);
-                        _sum3 = _mm_maddd_epi16(_r3, _w, _sum3);
-#else
-                        _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_r0, _w));
-                        _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_r1, _w));
-                        _sum2 = _mm_add_epi32(_sum2, _mm_madd_epi16(_r2, _w));
-                        _sum3 = _mm_add_epi32(_sum3, _mm_madd_epi16(_r3, _w));
-#endif
-
-                        kptr += 8;
-                    }
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-                const signed char* r2 = bottom_blob.channel(q).row<const signed char>(i2 * stride_h) + j2 * stride_w;
-                const signed char* r3 = bottom_blob.channel(q).row<const signed char>(i3 * stride_h) + j3 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-                    const signed char* r2s = r2 + space_ofs[k];
-                    const signed char* r3s = r3 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _r0 = _mm_set1_epi16(r0s[0]);
-                        __m128i _r1 = _mm_set1_epi16(r1s[0]);
-                        __m128i _r2 = _mm_set1_epi16(r2s[0]);
-                        __m128i _r3 = _mm_set1_epi16(r3s[0]);
-
-                        __m128i _w = _mm_loadl_epi64((const __m128i*)kptr);
-#if __SSE4_1__
-                        _w = _mm_cvtepi8_epi16(_w);
-#else
-                        _w = _mm_unpacklo_epi8(_w, _mm_cmpgt_epi8(_mm_setzero_si128(), _w));
-#endif
-
-#if __XOP__
-                        _w = _mm_unpacklo_epi16(_w, _mm_setzero_si128());
-                        _sum0 = _mm_maccd_epi16(_r0, _w, _sum0);
-                        _sum1 = _mm_maccd_epi16(_r1, _w, _sum1);
-                        _sum2 = _mm_maccd_epi16(_r2, _w, _sum2);
-                        _sum3 = _mm_maccd_epi16(_r3, _w, _sum3);
-#else
-                        __m128i _sl0 = _mm_mullo_epi16(_r0, _w);
-                        __m128i _sh0 = _mm_mulhi_epi16(_r0, _w);
-                        __m128i _sl1 = _mm_mullo_epi16(_r1, _w);
-                        __m128i _sh1 = _mm_mulhi_epi16(_r1, _w);
-                        __m128i _sl2 = _mm_mullo_epi16(_r2, _w);
-                        __m128i _sh2 = _mm_mulhi_epi16(_r2, _w);
-                        __m128i _sl3 = _mm_mullo_epi16(_r3, _w);
-                        __m128i _sh3 = _mm_mulhi_epi16(_r3, _w);
-                        __m128i _s0 = _mm_unpacklo_epi16(_sl0, _sh0);
-                        __m128i _s1 = _mm_unpacklo_epi16(_sl1, _sh1);
-                        __m128i _s2 = _mm_unpacklo_epi16(_sl2, _sh2);
-                        __m128i _s3 = _mm_unpacklo_epi16(_sl3, _sh3);
-
-                        _sum0 = _mm_add_epi32(_sum0, _s0);
-                        _sum1 = _mm_add_epi32(_sum1, _s1);
-                        _sum2 = _mm_add_epi32(_sum2, _s2);
-                        _sum3 = _mm_add_epi32(_sum3, _s3);
-#endif // __XOP__
-
-                        kptr += 4;
-                    }
-                }
-            }
-
-            if (out_elempack == 4)
-            {
-                _mm_store_si128((__m128i*)outptr, _sum0);
-                _mm_store_si128((__m128i*)(outptr + 4), _sum1);
-                _mm_store_si128((__m128i*)(outptr + 8), _sum2);
-                _mm_store_si128((__m128i*)(outptr + 12), _sum3);
-                outptr += 16;
-            }
-            if (out_elempack == 1)
-            {
-#if __AVX512F__
-                __m128i _vindex = _mm_setr_epi32(0, 1, 2, 3);
-                _vindex = _mm_mullo_epi32(_vindex, _mm_set1_epi32(M));
-                _mm_i32scatter_epi32(outptr, _vindex, _sum0, sizeof(int));
-                _mm_i32scatter_epi32(outptr + 1, _vindex, _sum1, sizeof(int));
-                _mm_i32scatter_epi32(outptr + 2, _vindex, _sum2, sizeof(int));
-                _mm_i32scatter_epi32(outptr + 3, _vindex, _sum3, sizeof(int));
-#else
-                int sum0[4];
-                int sum1[4];
-                int sum2[4];
-                int sum3[4];
-                _mm_storeu_si128((__m128i*)sum0, _sum0);
-                _mm_storeu_si128((__m128i*)sum1, _sum1);
-                _mm_storeu_si128((__m128i*)sum2, _sum2);
-                _mm_storeu_si128((__m128i*)sum3, _sum3);
-
-                outptr[0] = sum0[0];
-                outptr[1] = sum1[0];
-                outptr[2] = sum2[0];
-                outptr[3] = sum3[0];
-                outptr[M] = sum0[1];
-                outptr[M + 1] = sum1[1];
-                outptr[M + 2] = sum2[1];
-                outptr[M + 3] = sum3[1];
-                outptr[M * 2] = sum0[2];
-                outptr[M * 2 + 1] = sum1[2];
-                outptr[M * 2 + 2] = sum2[2];
-                outptr[M * 2 + 3] = sum3[2];
-                outptr[M * 3] = sum0[3];
-                outptr[M * 3 + 1] = sum1[3];
-                outptr[M * 3 + 2] = sum2[3];
-                outptr[M * 3 + 3] = sum3[3];
-#endif // __AVX512F__
-                outptr += 4;
-            }
-        }
-        for (; ij + 1 < outw * outh; ij += 2)
-        {
-            const int i0 = ij / outw;
-            const int i1 = (ij + 1) / outw;
-            const int j0 = ij % outw;
-            const int j1 = (ij + 1) % outw;
-
-            __m128i _sum0 = _mm_setzero_si128();
-            __m128i _sum1 = _mm_setzero_si128();
-            __m128i _sum2 = _mm_setzero_si128();
-            __m128i _sum3 = _mm_setzero_si128();
-
-#if __AVX512F__
-            const signed char* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8 + (p % 8) / 4);
-#elif __AVX2__
-            const signed char* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 4);
-#endif
-
-            int q = 0;
-#if __AVX512F__
-            {
-                __m512i _sum00 = _mm512_setzero_si512();
-                __m512i _sum11 = _mm512_setzero_si512();
-                __m512i _sum22 = _mm512_setzero_si512();
-                __m512i _sum33 = _mm512_setzero_si512();
-                for (; q + 15 < inch; q += 16)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-
-                        __m128i _r0;
-                        __m128i _r1;
-                        if (elempack == 16)
-                        {
-                            _r0 = _mm_load_si128((const __m128i*)r0s);
-                            _r1 = _mm_load_si128((const __m128i*)r1s);
-                        }
-                        else if (elempack == 8)
-                        {
-                            __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            __m128i _t1 = _mm_loadl_epi64((const __m128i*)(r0s + N));
-                            __m128i _t2 = _mm_loadl_epi64((const __m128i*)r1s);
-                            __m128i _t3 = _mm_loadl_epi64((const __m128i*)(r1s + N));
-                            _r0 = _mm_unpacklo_epi64(_t0, _t1);
-                            _r1 = _mm_unpacklo_epi64(_t2, _t3);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                            _r0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0s), 1));
-                            _r1 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r1s), 1));
-                        }
-
-                        __m256i _rr0 = _mm256_cvtepi8_epi16(_r0);
-                        __m256i _rr1 = _mm256_cvtepi8_epi16(_r1);
-
-                        __m512i _w = _mm512_load_si512((const __m512i*)kptr);
-                        __m512i _w0 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w, 0));
-                        __m512i _w1 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w, 1));
-
-                        // 01234567 89abcdef -> 01010101 23232323 45454545 67676767
-                        _rr0 = _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(3, 1, 2, 0));
-                        _rr1 = _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(3, 1, 2, 0));
-
-                        __m512i _rrr0 = _mm512_inserti64x4(_mm512_castsi256_si512(_rr0), _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(1, 0, 3, 2)), 1);
-                        __m512i _rrr1 = _mm512_inserti64x4(_mm512_castsi256_si512(_rr1), _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(1, 0, 3, 2)), 1);
-
-                        _rrr0 = _mm512_unpacklo_epi32(_rrr0, _rrr0);
-                        _rrr1 = _mm512_unpacklo_epi32(_rrr1, _rrr1);
-
-                        __m512i _rrr0l = _mm512_unpacklo_epi64(_rrr0, _rrr0);
-                        __m512i _rrr0h = _mm512_unpackhi_epi64(_rrr0, _rrr0);
-                        __m512i _rrr1l = _mm512_unpacklo_epi64(_rrr1, _rrr1);
-                        __m512i _rrr1h = _mm512_unpackhi_epi64(_rrr1, _rrr1);
-
-#if __AVX512VNNI__
-                        _sum00 = _mm512_dpwssd_epi32(_sum00, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(2, 0, 2, 0)), _w0);
-                        _sum11 = _mm512_dpwssd_epi32(_sum11, _mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(2, 0, 2, 0)), _w0);
-                        _sum22 = _mm512_dpwssd_epi32(_sum22, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(3, 1, 3, 1)), _w1);
-                        _sum33 = _mm512_dpwssd_epi32(_sum33, _mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(3, 1, 3, 1)), _w1);
-#else
-                        _sum00 = _mm512_add_epi32(_sum00, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(2, 0, 2, 0)), _w0));
-                        _sum11 = _mm512_add_epi32(_sum11, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(2, 0, 2, 0)), _w0));
-                        _sum22 = _mm512_add_epi32(_sum22, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(3, 1, 3, 1)), _w1));
-                        _sum33 = _mm512_add_epi32(_sum33, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr1l, _rrr1h, _MM_SHUFFLE(3, 1, 3, 1)), _w1));
-#endif // __AVX512VNNI__
-
-                        kptr += 64;
-                    }
-                }
-                __m256i _ss0 = _mm256_add_epi32(_mm512_extracti64x4_epi64(_sum00, 0), _mm512_extracti64x4_epi64(_sum00, 1));
-                __m256i _ss1 = _mm256_add_epi32(_mm512_extracti64x4_epi64(_sum11, 0), _mm512_extracti64x4_epi64(_sum11, 1));
-                __m256i _ss2 = _mm256_add_epi32(_mm512_extracti64x4_epi64(_sum22, 0), _mm512_extracti64x4_epi64(_sum22, 1));
-                __m256i _ss3 = _mm256_add_epi32(_mm512_extracti64x4_epi64(_sum33, 0), _mm512_extracti64x4_epi64(_sum33, 1));
-                _sum0 = _mm_add_epi32(_sum0, _mm256_extracti128_si256(_ss0, 0));
-                _sum0 = _mm_add_epi32(_sum0, _mm256_extracti128_si256(_ss0, 1));
-                _sum1 = _mm_add_epi32(_sum1, _mm256_extracti128_si256(_ss1, 0));
-                _sum1 = _mm_add_epi32(_sum1, _mm256_extracti128_si256(_ss1, 1));
-                _sum2 = _mm_add_epi32(_sum2, _mm256_extracti128_si256(_ss2, 0));
-                _sum2 = _mm_add_epi32(_sum2, _mm256_extracti128_si256(_ss2, 1));
-                _sum3 = _mm_add_epi32(_sum3, _mm256_extracti128_si256(_ss3, 0));
-                _sum3 = _mm_add_epi32(_sum3, _mm256_extracti128_si256(_ss3, 1));
-            }
-#endif // __AVX512F__
-            {
-#if __AVX2__
-                __m256i _sum00 = _mm256_setzero_si256();
-                __m256i _sum11 = _mm256_setzero_si256();
-                __m256i _sum22 = _mm256_setzero_si256();
-                __m256i _sum33 = _mm256_setzero_si256();
-#endif
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-
-                        __m128i _r0;
-                        __m128i _r1;
-                        if (elempack == 8)
-                        {
-                            _r0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            _r1 = _mm_loadl_epi64((const __m128i*)r1s);
-                        }
-                        else // if (elempack == 1)
-                        {
-#if __AVX2__
-                            __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                            _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-#if __AVX512F__
-                            _r0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1));
-                            _r1 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1));
-#else
-                            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-                            __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-                            __m256i _val0_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1), _sindex88);
-                            __m256i _val1_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1), _sindex88);
-                            _r0 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val0_32, 0), _mm256_extracti128_si256(_val0_32, 1));
-                            _r1 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val1_32, 0), _mm256_extracti128_si256(_val1_32, 1));
-#endif // __AVX512F__
-#else
-                            _r0 = _mm_setr_epi8(r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                            _r1 = _mm_setr_epi8(r1s[0], r1s[N], r1s[N * 2], r1s[N * 3], r1s[N * 4], r1s[N * 5], r1s[N * 6], r1s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-#endif // __AVX2__
-                        }
-
-#if __SSE4_1__
-                        _r0 = _mm_cvtepi8_epi16(_r0);
-                        _r1 = _mm_cvtepi8_epi16(_r1);
-#else
-                        _r0 = _mm_unpacklo_epi8(_r0, _mm_cmpgt_epi8(_mm_setzero_si128(), _r0));
-                        _r1 = _mm_unpacklo_epi8(_r1, _mm_cmpgt_epi8(_mm_setzero_si128(), _r1));
-#endif
-
-#if __AVX2__
-                        __m256i _w = _mm256_load_si256((const __m256i*)kptr);
-                        __m256i _w0 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w, 0));
-                        __m256i _w1 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w, 1));
-
-                        // 01234567 -> 01010101 23232323
-                        __m256i _rr0 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r0), _mm_shuffle_epi32(_r0, _MM_SHUFFLE(2, 3, 0, 1)), 1);
-                        __m256i _rr1 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r1), _mm_shuffle_epi32(_r1, _MM_SHUFFLE(2, 3, 0, 1)), 1);
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum00 = _mm256_dpwssd_epi32(_sum00, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                        _sum11 = _mm256_dpwssd_epi32(_sum11, _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                        _sum22 = _mm256_dpwssd_epi32(_sum22, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(2, 2, 2, 2)), _w1);
-                        _sum33 = _mm256_dpwssd_epi32(_sum33, _mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(2, 2, 2, 2)), _w1);
-#else
-                        _sum00 = _mm256_add_epi32(_sum00, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum11 = _mm256_add_epi32(_sum11, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum22 = _mm256_add_epi32(_sum22, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(2, 2, 2, 2)), _w1));
-                        _sum33 = _mm256_add_epi32(_sum33, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr1, _MM_SHUFFLE(2, 2, 2, 2)), _w1));
-#endif
-#else // __AVX2__
-                        __m128i _w01 = _mm_load_si128((const __m128i*)kptr);
-                        __m128i _w23 = _mm_load_si128((const __m128i*)(kptr + 16));
-                        __m128i _extw01 = _mm_cmpgt_epi8(_mm_setzero_si128(), _w01);
-                        __m128i _extw23 = _mm_cmpgt_epi8(_mm_setzero_si128(), _w23);
-                        __m128i _w0 = _mm_unpacklo_epi8(_w01, _extw01);
-                        __m128i _w1 = _mm_unpackhi_epi8(_w01, _extw01);
-                        __m128i _w2 = _mm_unpacklo_epi8(_w23, _extw23);
-                        __m128i _w3 = _mm_unpackhi_epi8(_w23, _extw23);
-
-                        // 01234567 -> 01010101
-#if __XOP__
-                        _sum0 = _mm_maddd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(0, 0, 0, 0)), _w0, _sum0);
-                        _sum1 = _mm_maddd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(0, 0, 0, 0)), _w0, _sum1);
-                        _sum2 = _mm_maddd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(1, 1, 1, 1)), _w1, _sum2);
-                        _sum3 = _mm_maddd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(1, 1, 1, 1)), _w1, _sum3);
-                        _sum0 = _mm_maddd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(2, 2, 2, 2)), _w2, _sum0);
-                        _sum1 = _mm_maddd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(2, 2, 2, 2)), _w2, _sum1);
-                        _sum2 = _mm_maddd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(3, 3, 3, 3)), _w3, _sum2);
-                        _sum3 = _mm_maddd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(3, 3, 3, 3)), _w3, _sum3);
-#else
-                        _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum2 = _mm_add_epi32(_sum2, _mm_madd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(1, 1, 1, 1)), _w1));
-                        _sum3 = _mm_add_epi32(_sum3, _mm_madd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(1, 1, 1, 1)), _w1));
-                        _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(2, 2, 2, 2)), _w2));
-                        _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(2, 2, 2, 2)), _w2));
-                        _sum2 = _mm_add_epi32(_sum2, _mm_madd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-                        _sum3 = _mm_add_epi32(_sum3, _mm_madd_epi16(_mm_shuffle_epi32(_r1, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-#endif // __XOP__
-#endif // __AVX2__
-
-                        kptr += 32;
-                    }
-                }
-#if __AVX2__
-                __m128i _ss0 = _mm_add_epi32(_mm256_extracti128_si256(_sum00, 0), _mm256_extracti128_si256(_sum00, 1));
-                __m128i _ss1 = _mm_add_epi32(_mm256_extracti128_si256(_sum11, 0), _mm256_extracti128_si256(_sum11, 1));
-                __m128i _ss2 = _mm_add_epi32(_mm256_extracti128_si256(_sum22, 0), _mm256_extracti128_si256(_sum22, 1));
-                __m128i _ss3 = _mm_add_epi32(_mm256_extracti128_si256(_sum33, 0), _mm256_extracti128_si256(_sum33, 1));
-                _sum0 = _mm_add_epi32(_sum0, _ss0);
-                _sum1 = _mm_add_epi32(_sum1, _ss1);
-                _sum2 = _mm_add_epi32(_sum2, _ss2);
-                _sum3 = _mm_add_epi32(_sum3, _ss3);
-#endif // __AVX2__
-            }
-            for (; q + 1 < inch; q += 2)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _r0 = _mm_setr_epi16(r0s[0], r0s[N], r0s[0], r0s[N], r0s[0], r0s[N], r0s[0], r0s[N]);
-                        __m128i _r1 = _mm_setr_epi16(r1s[0], r1s[N], r1s[0], r1s[N], r1s[0], r1s[N], r1s[0], r1s[N]);
-
-                        __m128i _w = _mm_loadl_epi64((const __m128i*)kptr);
-#if __SSE4_1__
-                        _w = _mm_cvtepi8_epi16(_w);
-#else
-                        _w = _mm_unpacklo_epi8(_w, _mm_cmpgt_epi8(_mm_setzero_si128(), _w));
-#endif
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum0 = _mm_dpwssd_epi32(_sum0, _r0, _w);
-                        _sum1 = _mm_dpwssd_epi32(_sum1, _r1, _w);
-#elif __XOP__
-                        _sum0 = _mm_maddd_epi16(_r0, _w, _sum0);
-                        _sum1 = _mm_maddd_epi16(_r1, _w, _sum1);
-#else
-                        _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_r0, _w));
-                        _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_r1, _w));
-#endif
-
-                        kptr += 8;
-                    }
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _r0 = _mm_set1_epi16(r0s[0]);
-                        __m128i _r1 = _mm_set1_epi16(r1s[0]);
-
-                        __m128i _w = _mm_loadl_epi64((const __m128i*)kptr);
-#if __SSE4_1__
-                        _w = _mm_cvtepi8_epi16(_w);
-#else
-                        _w = _mm_unpacklo_epi8(_w, _mm_cmpgt_epi8(_mm_setzero_si128(), _w));
-#endif
-
-#if __XOP__
-                        _w = _mm_unpacklo_epi16(_w, _mm_setzero_si128());
-                        _sum0 = _mm_maccd_epi16(_r0, _w, _sum0);
-                        _sum1 = _mm_maccd_epi16(_r1, _w, _sum1);
-#else
-                        __m128i _sl0 = _mm_mullo_epi16(_r0, _w);
-                        __m128i _sh0 = _mm_mulhi_epi16(_r0, _w);
-                        __m128i _sl1 = _mm_mullo_epi16(_r1, _w);
-                        __m128i _sh1 = _mm_mulhi_epi16(_r1, _w);
-                        __m128i _s0 = _mm_unpacklo_epi16(_sl0, _sh0);
-                        __m128i _s1 = _mm_unpacklo_epi16(_sl1, _sh1);
-
-                        _sum0 = _mm_add_epi32(_sum0, _s0);
-                        _sum1 = _mm_add_epi32(_sum1, _s1);
-#endif // __XOP__
-
-                        kptr += 4;
-                    }
-                }
-            }
-
-            _sum0 = _mm_add_epi32(_sum0, _sum2);
-            _sum1 = _mm_add_epi32(_sum1, _sum3);
-
-            if (out_elempack == 4)
-            {
-                _mm_store_si128((__m128i*)outptr, _sum0);
-                _mm_store_si128((__m128i*)(outptr + 4), _sum1);
-                outptr += 8;
-            }
-            if (out_elempack == 1)
-            {
-#if __AVX512F__
-                __m128i _vindex = _mm_setr_epi32(0, 1, 2, 3);
-                _vindex = _mm_mullo_epi32(_vindex, _mm_set1_epi32(M));
-                _mm_i32scatter_epi32(outptr, _vindex, _sum0, sizeof(int));
-                _mm_i32scatter_epi32(outptr + 1, _vindex, _sum1, sizeof(int));
-#else
-                int sum0[4];
-                int sum1[4];
-                _mm_storeu_si128((__m128i*)sum0, _sum0);
-                _mm_storeu_si128((__m128i*)sum1, _sum1);
-
-                outptr[0] = sum0[0];
-                outptr[1] = sum1[0];
-                outptr[M] = sum0[1];
-                outptr[M + 1] = sum1[1];
-                outptr[M * 2] = sum0[2];
-                outptr[M * 2 + 1] = sum1[2];
-                outptr[M * 3] = sum0[3];
-                outptr[M * 3 + 1] = sum1[3];
-#endif // __AVX512F__
-                outptr += 2;
-            }
-        }
-        for (; ij < outw * outh; ij++)
-        {
-            const int i = ij / outw;
-            const int j = ij % outw;
-
-            __m128i _sum0 = _mm_setzero_si128();
-
-#if __AVX512F__
-            const signed char* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8 + (p % 8) / 4);
-#elif __AVX2__
-            const signed char* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 4);
-#endif
-
-            int q = 0;
-#if __AVX512F__
-            {
-                __m512i _sum00 = _mm512_setzero_si512();
-                __m512i _sum11 = _mm512_setzero_si512();
-                for (; q + 15 < inch; q += 16)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-
-                        __m128i _r0;
-                        if (elempack == 16)
-                        {
-                            _r0 = _mm_load_si128((const __m128i*)r0s);
-                        }
-                        else if (elempack == 8)
-                        {
-                            __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            __m128i _t1 = _mm_loadl_epi64((const __m128i*)(r0s + N));
-                            _r0 = _mm_unpacklo_epi64(_t0, _t1);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                            _r0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0s), 1));
-                        }
-
-                        __m256i _rr0 = _mm256_cvtepi8_epi16(_r0);
-
-                        __m512i _w = _mm512_load_si512((const __m512i*)kptr);
-                        __m512i _w0 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w, 0));
-                        __m512i _w1 = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(_w, 1));
-
-                        // 01234567 89abcdef -> 01010101 23232323 45454545 67676767
-                        _rr0 = _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(3, 1, 2, 0));
-
-                        __m512i _rrr0 = _mm512_inserti64x4(_mm512_castsi256_si512(_rr0), _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(1, 0, 3, 2)), 1);
-
-                        _rrr0 = _mm512_unpacklo_epi32(_rrr0, _rrr0);
-
-                        __m512i _rrr0l = _mm512_unpacklo_epi64(_rrr0, _rrr0);
-                        __m512i _rrr0h = _mm512_unpackhi_epi64(_rrr0, _rrr0);
-
-#if __AVX512VNNI__
-                        _sum00 = _mm512_dpwssd_epi32(_sum00, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(2, 0, 2, 0)), _w0);
-                        _sum11 = _mm512_dpwssd_epi32(_sum11, _mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(3, 1, 3, 1)), _w1);
-#else
-                        _sum00 = _mm512_add_epi32(_sum00, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(2, 0, 2, 0)), _w0));
-                        _sum11 = _mm512_add_epi32(_sum11, _mm512_madd_epi16(_mm512_shuffle_i32x4(_rrr0l, _rrr0h, _MM_SHUFFLE(3, 1, 3, 1)), _w1));
-#endif // __AVX512VNNI__
-
-                        kptr += 64;
-                    }
-                }
-                _sum00 = _mm512_add_epi32(_sum00, _sum11);
-                __m256i _ss0 = _mm256_add_epi32(_mm512_extracti64x4_epi64(_sum00, 0), _mm512_extracti64x4_epi64(_sum00, 1));
-                _sum0 = _mm_add_epi32(_sum0, _mm256_extracti128_si256(_ss0, 0));
-                _sum0 = _mm_add_epi32(_sum0, _mm256_extracti128_si256(_ss0, 1));
-            }
-#endif // __AVX512F__
-            {
-#if __AVX2__
-                __m256i _sum00 = _mm256_setzero_si256();
-                __m256i _sum11 = _mm256_setzero_si256();
-#else
-                __m128i _sum1 = _mm_setzero_si128();
-                __m128i _sum2 = _mm_setzero_si128();
-                __m128i _sum3 = _mm_setzero_si128();
-#endif
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-
-                        __m128i _r0;
-                        if (elempack == 8)
-                        {
-                            _r0 = _mm_loadl_epi64((const __m128i*)r0s);
-                        }
-                        else // if (elempack == 1)
-                        {
-#if __AVX2__
-                            __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                            _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-#if __AVX512F__
-                            _r0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1));
-#else
-                            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-                            __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-                            __m256i _val32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1), _sindex88);
-                            _r0 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val32, 0), _mm256_extracti128_si256(_val32, 1));
-#endif // __AVX512F__
-#else
-                            _r0 = _mm_setr_epi8(r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-#endif // __AVX2__
-                        }
-
-#if __SSE4_1__
-                        _r0 = _mm_cvtepi8_epi16(_r0);
-#else
-                        _r0 = _mm_unpacklo_epi8(_r0, _mm_cmpgt_epi8(_mm_setzero_si128(), _r0));
-#endif
-
-#if __AVX2__
-                        __m256i _w = _mm256_load_si256((const __m256i*)kptr);
-                        __m256i _w0 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w, 0));
-                        __m256i _w1 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(_w, 1));
-
-                        // 01234567 -> 01010101 23232323
-                        __m256i _rr0 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r0), _mm_shuffle_epi32(_r0, _MM_SHUFFLE(2, 3, 0, 1)), 1);
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum00 = _mm256_dpwssd_epi32(_sum00, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(0, 0, 0, 0)), _w0);
-                        _sum11 = _mm256_dpwssd_epi32(_sum11, _mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(2, 2, 2, 2)), _w1);
-#else
-                        _sum00 = _mm256_add_epi32(_sum00, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum11 = _mm256_add_epi32(_sum11, _mm256_madd_epi16(_mm256_shuffle_epi32(_rr0, _MM_SHUFFLE(2, 2, 2, 2)), _w1));
-#endif
-#else // __AVX2__
-                        __m128i _w01 = _mm_load_si128((const __m128i*)kptr);
-                        __m128i _w23 = _mm_load_si128((const __m128i*)(kptr + 16));
-                        __m128i _extw01 = _mm_cmpgt_epi8(_mm_setzero_si128(), _w01);
-                        __m128i _extw23 = _mm_cmpgt_epi8(_mm_setzero_si128(), _w23);
-                        __m128i _w0 = _mm_unpacklo_epi8(_w01, _extw01);
-                        __m128i _w1 = _mm_unpackhi_epi8(_w01, _extw01);
-                        __m128i _w2 = _mm_unpacklo_epi8(_w23, _extw23);
-                        __m128i _w3 = _mm_unpackhi_epi8(_w23, _extw23);
-
-                        // 01234567 -> 01010101
-#if __XOP__
-                        _sum0 = _mm_maddd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(0, 0, 0, 0)), _w0, _sum0);
-                        _sum1 = _mm_maddd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(1, 1, 1, 1)), _w1, _sum1);
-                        _sum2 = _mm_maddd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(2, 2, 2, 2)), _w2, _sum2);
-                        _sum3 = _mm_maddd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(3, 3, 3, 3)), _w3, _sum3);
-#else
-                        _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(0, 0, 0, 0)), _w0));
-                        _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(1, 1, 1, 1)), _w1));
-                        _sum2 = _mm_add_epi32(_sum2, _mm_madd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(2, 2, 2, 2)), _w2));
-                        _sum3 = _mm_add_epi32(_sum3, _mm_madd_epi16(_mm_shuffle_epi32(_r0, _MM_SHUFFLE(3, 3, 3, 3)), _w3));
-#endif // __XOP__
-#endif // __AVX2__
-
-                        kptr += 32;
-                    }
-                }
-#if __AVX2__
-                _sum00 = _mm256_add_epi32(_sum00, _sum11);
-                __m128i _ss = _mm_add_epi32(_mm256_extracti128_si256(_sum00, 0), _mm256_extracti128_si256(_sum00, 1));
-                _sum0 = _mm_add_epi32(_sum0, _ss);
-#else
-                _sum0 = _mm_add_epi32(_sum0, _sum1);
-                _sum2 = _mm_add_epi32(_sum2, _sum3);
-                _sum0 = _mm_add_epi32(_sum0, _sum2);
-#endif // __AVX2__
-            }
-            for (; q + 1 < inch; q += 2)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i * stride_h) + j * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _r0 = _mm_setr_epi16(r0s[0], r0s[N], r0s[0], r0s[N], r0s[0], r0s[N], r0s[0], r0s[N]);
-
-                        __m128i _w = _mm_loadl_epi64((const __m128i*)kptr);
-#if __SSE4_1__
-                        _w = _mm_cvtepi8_epi16(_w);
-#else
-                        _w = _mm_unpacklo_epi8(_w, _mm_cmpgt_epi8(_mm_setzero_si128(), _w));
-#endif
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum0 = _mm_dpwssd_epi32(_sum0, _r0, _w);
-#elif __XOP__
-                        _sum0 = _mm_maddd_epi16(_r0, _w, _sum0);
-#else
-                        _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_r0, _w));
-#endif
-
-                        kptr += 8;
-                    }
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i * stride_h) + j * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _r0 = _mm_set1_epi16(r0s[0]);
-
-                        __m128i _w = _mm_loadl_epi64((const __m128i*)kptr);
-#if __SSE4_1__
-                        _w = _mm_cvtepi8_epi16(_w);
-#else
-                        _w = _mm_unpacklo_epi8(_w, _mm_cmpgt_epi8(_mm_setzero_si128(), _w));
-#endif
-
-#if __XOP__
-                        _w = _mm_unpacklo_epi16(_w, _mm_setzero_si128());
-                        _sum0 = _mm_maccd_epi16(_r0, _w, _sum0);
-#else
-                        __m128i _sl = _mm_mullo_epi16(_r0, _w);
-                        __m128i _sh = _mm_mulhi_epi16(_r0, _w);
-                        __m128i _s0 = _mm_unpacklo_epi16(_sl, _sh);
-
-                        _sum0 = _mm_add_epi32(_sum0, _s0);
-#endif // __XOP__
-
-                        kptr += 4;
-                    }
-                }
-            }
-
-            if (out_elempack == 4)
-            {
-                _mm_store_si128((__m128i*)outptr, _sum0);
-                outptr += 4;
-            }
-            if (out_elempack == 1)
-            {
-#if __AVX512F__
-                __m128i _vindex = _mm_setr_epi32(0, 1, 2, 3);
-                _vindex = _mm_mullo_epi32(_vindex, _mm_set1_epi32(M));
-                _mm_i32scatter_epi32(outptr, _vindex, _sum0, sizeof(int));
-#else
-                int sum[4];
-                _mm_storeu_si128((__m128i*)sum, _sum0);
-
-                outptr[0] = sum[0];
-                outptr[M] = sum[1];
-                outptr[M * 2] = sum[2];
-                outptr[M * 3] = sum[3];
-#endif // __AVX512F__
-                outptr += 1;
-            }
-        }
-    }
-    remain_outch_start += nn_outch * 4;
-    nn_outch = (outch - remain_outch_start) / 2;
-#else // __SSE2__
-    nn_outch = (outch - remain_outch_start) / 2;
-    #pragma omp parallel for num_threads(opt.num_threads)
-#endif // __SSE2__
-    for (int pp = 0; pp < nn_outch; pp++)
-    {
-        const int p = remain_outch_start + pp * 2;
-
-        // shadowed variable for less openmp task args
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-        const int N = bottom_blob.cstep * elempack;
-
-        int* outptr0 = top_blob.channel(p);
-        int* outptr1 = top_blob.channel(p + 1);
-
-        int ij = 0;
-#if __SSE2__
-        for (; ij + 3 < outw * outh; ij += 4)
-        {
-            const int i0 = ij / outw;
-            const int i1 = (ij + 1) / outw;
-            const int i2 = (ij + 2) / outw;
-            const int i3 = (ij + 3) / outw;
-            const int j0 = ij % outw;
-            const int j1 = (ij + 1) % outw;
-            const int j2 = (ij + 2) % outw;
-            const int j3 = (ij + 3) % outw;
-
-            __m128i _sum0 = _mm_setzero_si128();
-            __m128i _sum1 = _mm_setzero_si128();
-
-#if __AVX512F__
-            const signed char* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8 + (p % 8) / 4 + (p % 4) / 2);
-#elif __AVX2__
-            const signed char* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2);
-#endif
-
-            int q = 0;
-#if __AVX512F__
-            {
-                __m512i _sum00 = _mm512_setzero_si512();
-                __m512i _sum11 = _mm512_setzero_si512();
-                __m512i _sum22 = _mm512_setzero_si512();
-                __m512i _sum33 = _mm512_setzero_si512();
-                for (; q + 15 < inch; q += 16)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-                    const signed char* r2 = bottom_blob.channel(q / elempack).row<const signed char>(i2 * stride_h) + j2 * stride_w * elempack;
-                    const signed char* r3 = bottom_blob.channel(q / elempack).row<const signed char>(i3 * stride_h) + j3 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-                        const signed char* r2s = r2 + space_ofs[k];
-                        const signed char* r3s = r3 + space_ofs[k];
-
-                        __m128i _r0;
-                        __m128i _r1;
-                        __m128i _r2;
-                        __m128i _r3;
-                        if (elempack == 16)
-                        {
-                            _r0 = _mm_load_si128((const __m128i*)r0s);
-                            _r1 = _mm_load_si128((const __m128i*)r1s);
-                            _r2 = _mm_load_si128((const __m128i*)r2s);
-                            _r3 = _mm_load_si128((const __m128i*)r3s);
-                        }
-                        else if (elempack == 8)
-                        {
-                            __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            __m128i _t1 = _mm_loadl_epi64((const __m128i*)(r0s + N));
-                            __m128i _t2 = _mm_loadl_epi64((const __m128i*)r1s);
-                            __m128i _t3 = _mm_loadl_epi64((const __m128i*)(r1s + N));
-                            __m128i _t4 = _mm_loadl_epi64((const __m128i*)r2s);
-                            __m128i _t5 = _mm_loadl_epi64((const __m128i*)(r2s + N));
-                            __m128i _t6 = _mm_loadl_epi64((const __m128i*)r3s);
-                            __m128i _t7 = _mm_loadl_epi64((const __m128i*)(r3s + N));
-                            _r0 = _mm_unpacklo_epi64(_t0, _t1);
-                            _r1 = _mm_unpacklo_epi64(_t2, _t3);
-                            _r2 = _mm_unpacklo_epi64(_t4, _t5);
-                            _r3 = _mm_unpacklo_epi64(_t6, _t7);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                            _r0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0s), 1));
-                            _r1 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r1s), 1));
-                            _r2 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r2s), 1));
-                            _r3 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r3s), 1));
-                        }
-
-                        __m256i _val0 = _mm256_cvtepi8_epi16(_r0);
-                        __m256i _val1 = _mm256_cvtepi8_epi16(_r1);
-                        __m256i _val2 = _mm256_cvtepi8_epi16(_r2);
-                        __m256i _val3 = _mm256_cvtepi8_epi16(_r3);
-                        __m512i _valval0 = _mm512_inserti64x4(_mm512_castsi256_si512(_val0), _val0, 1);
-                        __m512i _valval1 = _mm512_inserti64x4(_mm512_castsi256_si512(_val1), _val1, 1);
-                        __m512i _valval2 = _mm512_inserti64x4(_mm512_castsi256_si512(_val2), _val2, 1);
-                        __m512i _valval3 = _mm512_inserti64x4(_mm512_castsi256_si512(_val3), _val3, 1);
-
-                        __m512i _w = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*)kptr));
-
-#if __AVX512VNNI__
-                        _sum00 = _mm512_dpwssd_epi32(_sum00, _valval0, _w);
-                        _sum11 = _mm512_dpwssd_epi32(_sum11, _valval1, _w);
-                        _sum22 = _mm512_dpwssd_epi32(_sum22, _valval2, _w);
-                        _sum33 = _mm512_dpwssd_epi32(_sum33, _valval3, _w);
-#else
-                        _sum00 = _mm512_add_epi32(_sum00, _mm512_madd_epi16(_valval0, _w));
-                        _sum11 = _mm512_add_epi32(_sum11, _mm512_madd_epi16(_valval1, _w));
-                        _sum22 = _mm512_add_epi32(_sum22, _mm512_madd_epi16(_valval2, _w));
-                        _sum33 = _mm512_add_epi32(_sum33, _mm512_madd_epi16(_valval3, _w));
-#endif // __AVX512VNNI__
-
-                        kptr += 32;
-                    }
-                }
-                __m256i _sum010 = _mm256_hadd_epi32(_mm512_extracti64x4_epi64(_sum00, 0), _mm512_extracti64x4_epi64(_sum11, 0));
-                __m256i _sum230 = _mm256_hadd_epi32(_mm512_extracti64x4_epi64(_sum22, 0), _mm512_extracti64x4_epi64(_sum33, 0));
-                __m256i _sum011 = _mm256_hadd_epi32(_mm512_extracti64x4_epi64(_sum00, 1), _mm512_extracti64x4_epi64(_sum11, 1));
-                __m256i _sum231 = _mm256_hadd_epi32(_mm512_extracti64x4_epi64(_sum22, 1), _mm512_extracti64x4_epi64(_sum33, 1));
-                __m256i _ss0 = _mm256_hadd_epi32(_sum010, _sum230);
-                __m256i _ss1 = _mm256_hadd_epi32(_sum011, _sum231);
-                _sum0 = _mm_add_epi32(_sum0, _mm256_extracti128_si256(_ss0, 0));
-                _sum0 = _mm_add_epi32(_sum0, _mm256_extracti128_si256(_ss0, 1));
-                _sum1 = _mm_add_epi32(_sum1, _mm256_extracti128_si256(_ss1, 0));
-                _sum1 = _mm_add_epi32(_sum1, _mm256_extracti128_si256(_ss1, 1));
-            }
-#endif // __AVX512F__
-            {
-#if __AVX2__
-                __m256i _sum00 = _mm256_setzero_si256();
-                __m256i _sum11 = _mm256_setzero_si256();
-                __m256i _sum22 = _mm256_setzero_si256();
-                __m256i _sum33 = _mm256_setzero_si256();
-#else
-                __m128i _sum00 = _mm_setzero_si128();
-                __m128i _sum10 = _mm_setzero_si128();
-                __m128i _sum01 = _mm_setzero_si128();
-                __m128i _sum11 = _mm_setzero_si128();
-                __m128i _sum02 = _mm_setzero_si128();
-                __m128i _sum12 = _mm_setzero_si128();
-                __m128i _sum03 = _mm_setzero_si128();
-                __m128i _sum13 = _mm_setzero_si128();
-#endif
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-                    const signed char* r2 = bottom_blob.channel(q / elempack).row<const signed char>(i2 * stride_h) + j2 * stride_w * elempack;
-                    const signed char* r3 = bottom_blob.channel(q / elempack).row<const signed char>(i3 * stride_h) + j3 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-                        const signed char* r2s = r2 + space_ofs[k];
-                        const signed char* r3s = r3 + space_ofs[k];
-
-                        __m128i _r0;
-                        __m128i _r1;
-                        __m128i _r2;
-                        __m128i _r3;
-                        if (elempack == 8)
-                        {
-                            _r0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            _r1 = _mm_loadl_epi64((const __m128i*)r1s);
-                            _r2 = _mm_loadl_epi64((const __m128i*)r2s);
-                            _r3 = _mm_loadl_epi64((const __m128i*)r3s);
-                        }
-                        else // if (elempack == 1)
-                        {
-#if __AVX2__
-                            __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                            _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-#if __AVX512F__
-                            _r0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1));
-                            _r1 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1));
-                            _r2 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r2s, _vindex, 1));
-                            _r3 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r3s, _vindex, 1));
-#else
-                            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-                            __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-                            __m256i _val0_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1), _sindex88);
-                            __m256i _val1_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1), _sindex88);
-                            __m256i _val2_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r2s, _vindex, 1), _sindex88);
-                            __m256i _val3_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r3s, _vindex, 1), _sindex88);
-                            _r0 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val0_32, 0), _mm256_extracti128_si256(_val0_32, 1));
-                            _r1 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val1_32, 0), _mm256_extracti128_si256(_val1_32, 1));
-                            _r2 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val2_32, 0), _mm256_extracti128_si256(_val2_32, 1));
-                            _r3 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val3_32, 0), _mm256_extracti128_si256(_val3_32, 1));
-#endif // __AVX512F__
-#else
-                            _r0 = _mm_setr_epi8(r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                            _r1 = _mm_setr_epi8(r1s[0], r1s[N], r1s[N * 2], r1s[N * 3], r1s[N * 4], r1s[N * 5], r1s[N * 6], r1s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                            _r2 = _mm_setr_epi8(r2s[0], r2s[N], r2s[N * 2], r2s[N * 3], r2s[N * 4], r2s[N * 5], r2s[N * 6], r2s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                            _r3 = _mm_setr_epi8(r3s[0], r3s[N], r3s[N * 2], r3s[N * 3], r3s[N * 4], r3s[N * 5], r3s[N * 6], r3s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-#endif // __AVX2__
-                        }
-
-#if __SSE4_1__
-                        _r0 = _mm_cvtepi8_epi16(_r0);
-                        _r1 = _mm_cvtepi8_epi16(_r1);
-                        _r2 = _mm_cvtepi8_epi16(_r2);
-                        _r3 = _mm_cvtepi8_epi16(_r3);
-#else
-                        _r0 = _mm_unpacklo_epi8(_r0, _mm_cmpgt_epi8(_mm_setzero_si128(), _r0));
-                        _r1 = _mm_unpacklo_epi8(_r1, _mm_cmpgt_epi8(_mm_setzero_si128(), _r1));
-                        _r2 = _mm_unpacklo_epi8(_r2, _mm_cmpgt_epi8(_mm_setzero_si128(), _r2));
-                        _r3 = _mm_unpacklo_epi8(_r3, _mm_cmpgt_epi8(_mm_setzero_si128(), _r3));
-#endif
-
-                        __m128i _w01 = _mm_load_si128((const __m128i*)kptr);
-#if __AVX2__
-                        __m256i _valval0 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r0), _r0, 1);
-                        __m256i _valval1 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r1), _r1, 1);
-                        __m256i _valval2 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r2), _r2, 1);
-                        __m256i _valval3 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r3), _r3, 1);
-
-                        __m256i _w = _mm256_cvtepi8_epi16(_w01);
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum00 = _mm256_dpwssd_epi32(_sum00, _valval0, _w);
-                        _sum11 = _mm256_dpwssd_epi32(_sum11, _valval1, _w);
-                        _sum22 = _mm256_dpwssd_epi32(_sum22, _valval2, _w);
-                        _sum33 = _mm256_dpwssd_epi32(_sum33, _valval3, _w);
-#else
-                        _sum00 = _mm256_add_epi32(_sum00, _mm256_madd_epi16(_valval0, _w));
-                        _sum11 = _mm256_add_epi32(_sum11, _mm256_madd_epi16(_valval1, _w));
-                        _sum22 = _mm256_add_epi32(_sum22, _mm256_madd_epi16(_valval2, _w));
-                        _sum33 = _mm256_add_epi32(_sum33, _mm256_madd_epi16(_valval3, _w));
-#endif
-#else // __AVX2__
-                        __m128i _extw01 = _mm_cmpgt_epi8(_mm_setzero_si128(), _w01);
-                        __m128i _w0 = _mm_unpacklo_epi8(_w01, _extw01);
-                        __m128i _w1 = _mm_unpackhi_epi8(_w01, _extw01);
-
-#if __XOP__
-                        _sum00 = _mm_maddd_epi16(_r0, _w0, _sum00);
-                        _sum10 = _mm_maddd_epi16(_r0, _w1, _sum10);
-                        _sum01 = _mm_maddd_epi16(_r1, _w0, _sum01);
-                        _sum11 = _mm_maddd_epi16(_r1, _w1, _sum11);
-                        _sum02 = _mm_maddd_epi16(_r2, _w0, _sum02);
-                        _sum12 = _mm_maddd_epi16(_r2, _w1, _sum12);
-                        _sum03 = _mm_maddd_epi16(_r3, _w0, _sum03);
-                        _sum13 = _mm_maddd_epi16(_r3, _w1, _sum13);
-#else
-                        _sum00 = _mm_add_epi32(_sum00, _mm_madd_epi16(_r0, _w0));
-                        _sum10 = _mm_add_epi32(_sum10, _mm_madd_epi16(_r0, _w1));
-                        _sum01 = _mm_add_epi32(_sum01, _mm_madd_epi16(_r1, _w0));
-                        _sum11 = _mm_add_epi32(_sum11, _mm_madd_epi16(_r1, _w1));
-                        _sum02 = _mm_add_epi32(_sum02, _mm_madd_epi16(_r2, _w0));
-                        _sum12 = _mm_add_epi32(_sum12, _mm_madd_epi16(_r2, _w1));
-                        _sum03 = _mm_add_epi32(_sum03, _mm_madd_epi16(_r3, _w0));
-                        _sum13 = _mm_add_epi32(_sum13, _mm_madd_epi16(_r3, _w1));
-#endif // __XOP__
-#endif // __AVX2__
-
-                        kptr += 16;
-                    }
-                }
-#if __AVX2__
-                __m256i _sum01 = _mm256_hadd_epi32(_sum00, _sum11);
-                __m256i _sum23 = _mm256_hadd_epi32(_sum22, _sum33);
-                __m256i _ss = _mm256_hadd_epi32(_sum01, _sum23);
-                _sum0 = _mm_add_epi32(_sum0, _mm256_extracti128_si256(_ss, 0));
-                _sum1 = _mm_add_epi32(_sum1, _mm256_extracti128_si256(_ss, 1));
-#else
-                // transpose 4x4
-                __m128i _tmp00 = _mm_unpacklo_epi32(_sum00, _sum01);
-                __m128i _tmp01 = _mm_unpacklo_epi32(_sum02, _sum03);
-                __m128i _tmp02 = _mm_unpackhi_epi32(_sum00, _sum01);
-                __m128i _tmp03 = _mm_unpackhi_epi32(_sum02, _sum03);
-                __m128i _tmp10 = _mm_unpacklo_epi32(_sum10, _sum11);
-                __m128i _tmp11 = _mm_unpacklo_epi32(_sum12, _sum13);
-                __m128i _tmp12 = _mm_unpackhi_epi32(_sum10, _sum11);
-                __m128i _tmp13 = _mm_unpackhi_epi32(_sum12, _sum13);
-                _sum00 = _mm_unpacklo_epi64(_tmp00, _tmp01);
-                _sum01 = _mm_unpackhi_epi64(_tmp00, _tmp01);
-                _sum02 = _mm_unpacklo_epi64(_tmp02, _tmp03);
-                _sum03 = _mm_unpackhi_epi64(_tmp02, _tmp03);
-                _sum10 = _mm_unpacklo_epi64(_tmp10, _tmp11);
-                _sum11 = _mm_unpackhi_epi64(_tmp10, _tmp11);
-                _sum12 = _mm_unpacklo_epi64(_tmp12, _tmp13);
-                _sum13 = _mm_unpackhi_epi64(_tmp12, _tmp13);
-                _sum00 = _mm_add_epi32(_sum00, _sum01);
-                _sum02 = _mm_add_epi32(_sum02, _sum03);
-                _sum10 = _mm_add_epi32(_sum10, _sum11);
-                _sum12 = _mm_add_epi32(_sum12, _sum13);
-                _sum0 = _mm_add_epi32(_sum0, _sum00);
-                _sum0 = _mm_add_epi32(_sum0, _sum02);
-                _sum1 = _mm_add_epi32(_sum1, _sum10);
-                _sum1 = _mm_add_epi32(_sum1, _sum12);
-#endif // __AVX2__
-            }
-            for (; q + 1 < inch; q += 2)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-                const signed char* r2 = bottom_blob.channel(q).row<const signed char>(i2 * stride_h) + j2 * stride_w;
-                const signed char* r3 = bottom_blob.channel(q).row<const signed char>(i3 * stride_h) + j3 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-                    const signed char* r2s = r2 + space_ofs[k];
-                    const signed char* r3s = r3 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _r = _mm_setr_epi16(r0s[0], r0s[N], r1s[0], r1s[N], r2s[0], r2s[N], r3s[0], r3s[N]);
-                        __m128i _w0 = _mm_setr_epi16(kptr[0], kptr[2], kptr[0], kptr[2], kptr[0], kptr[2], kptr[0], kptr[2]);
-                        __m128i _w1 = _mm_setr_epi16(kptr[1], kptr[3], kptr[1], kptr[3], kptr[1], kptr[3], kptr[1], kptr[3]);
-
-#if __XOP__
-                        _sum0 = _mm_maddd_epi16(_r, _w0, _sum0);
-                        _sum1 = _mm_maddd_epi16(_r, _w1, _sum1);
-#else
-                        _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_r, _w0));
-                        _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_r, _w1));
-#endif // __XOP__
-
-                        kptr += 4;
-                    }
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-                const signed char* r2 = bottom_blob.channel(q).row<const signed char>(i2 * stride_h) + j2 * stride_w;
-                const signed char* r3 = bottom_blob.channel(q).row<const signed char>(i3 * stride_h) + j3 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-                    const signed char* r2s = r2 + space_ofs[k];
-                    const signed char* r3s = r3 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _r = _mm_setr_epi16(r0s[0], r1s[0], r2s[0], r3s[0], r0s[0], r1s[0], r2s[0], r3s[0]);
-                        __m128i _w = _mm_setr_epi16(kptr[0], kptr[0], kptr[0], kptr[0], kptr[1], kptr[1], kptr[1], kptr[1]);
-
-                        __m128i _sl = _mm_mullo_epi16(_r, _w);
-                        __m128i _sh = _mm_mulhi_epi16(_r, _w);
-                        __m128i _s0 = _mm_unpacklo_epi16(_sl, _sh);
-                        __m128i _s1 = _mm_unpackhi_epi16(_sl, _sh);
-
-                        _sum0 = _mm_add_epi32(_sum0, _s0);
-                        _sum1 = _mm_add_epi32(_sum1, _s1);
-
-                        kptr += 2;
-                    }
-                }
-            }
-
-            _mm_store_si128((__m128i*)outptr0, _sum0);
-            _mm_store_si128((__m128i*)outptr1, _sum1);
-            outptr0 += 4;
-            outptr1 += 4;
-        }
-#endif // __SSE2__
-        for (; ij + 1 < outw * outh; ij += 2)
-        {
-            const int i0 = ij / outw;
-            const int i1 = (ij + 1) / outw;
-            const int j0 = ij % outw;
-            const int j1 = (ij + 1) % outw;
-
-            int sum00 = 0;
-            int sum01 = 0;
-            int sum10 = 0;
-            int sum11 = 0;
-
-#if __AVX512F__
-            const signed char* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8 + (p % 8) / 4 + (p % 4) / 2);
-#elif __AVX2__
-            const signed char* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2);
-#elif __SSE2__
-            const signed char* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 2);
-#endif
-
-            int q = 0;
-#if __SSE2__
-#if __AVX512F__
-            {
-                __m512i _sum0 = _mm512_setzero_si512();
-                __m512i _sum1 = _mm512_setzero_si512();
-                for (; q + 15 < inch; q += 16)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-
-                        __m128i _r0;
-                        __m128i _r1;
-                        if (elempack == 16)
-                        {
-                            _r0 = _mm_load_si128((const __m128i*)r0s);
-                            _r1 = _mm_load_si128((const __m128i*)r1s);
-                        }
-                        else if (elempack == 8)
-                        {
-                            __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            __m128i _t1 = _mm_loadl_epi64((const __m128i*)(r0s + N));
-                            __m128i _t2 = _mm_loadl_epi64((const __m128i*)r1s);
-                            __m128i _t3 = _mm_loadl_epi64((const __m128i*)(r1s + N));
-                            _r0 = _mm_unpacklo_epi64(_t0, _t1);
-                            _r1 = _mm_unpacklo_epi64(_t2, _t3);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                            _r0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0s), 1));
-                            _r1 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r1s), 1));
-                        }
-
-                        __m256i _val0 = _mm256_cvtepi8_epi16(_r0);
-                        __m256i _val1 = _mm256_cvtepi8_epi16(_r1);
-                        __m512i _valval0 = _mm512_inserti64x4(_mm512_castsi256_si512(_val0), _val0, 1);
-                        __m512i _valval1 = _mm512_inserti64x4(_mm512_castsi256_si512(_val1), _val1, 1);
-
-                        __m512i _w = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*)kptr));
-
-#if __AVX512VNNI__
-                        _sum0 = _mm512_dpwssd_epi32(_sum0, _valval0, _w);
-                        _sum1 = _mm512_dpwssd_epi32(_sum1, _valval1, _w);
-#else
-                        _sum0 = _mm512_add_epi32(_sum0, _mm512_madd_epi16(_valval0, _w));
-                        _sum1 = _mm512_add_epi32(_sum1, _mm512_madd_epi16(_valval1, _w));
-#endif // __AVX512VNNI__
-
-                        kptr += 32;
-                    }
-                }
-                __m256i _sum010 = _mm256_hadd_epi32(_mm512_extracti64x4_epi64(_sum0, 0), _mm512_extracti64x4_epi64(_sum0, 1));
-                __m256i _sum011 = _mm256_hadd_epi32(_mm512_extracti64x4_epi64(_sum1, 0), _mm512_extracti64x4_epi64(_sum1, 1));
-                __m128i _ss0 = _mm_add_epi32(_mm256_extracti128_si256(_sum010, 0), _mm256_extracti128_si256(_sum010, 1));
-                __m128i _ss1 = _mm_add_epi32(_mm256_extracti128_si256(_sum011, 0), _mm256_extracti128_si256(_sum011, 1));
-
-                sum00 += _mm_extract_epi32(_ss0, 0) + _mm_extract_epi32(_ss0, 1);
-                sum10 += _mm_extract_epi32(_ss0, 2) + _mm_extract_epi32(_ss0, 3);
-                sum01 += _mm_extract_epi32(_ss1, 0) + _mm_extract_epi32(_ss1, 1);
-                sum11 += _mm_extract_epi32(_ss1, 2) + _mm_extract_epi32(_ss1, 3);
-            }
-#endif // __AVX512F__
-            {
-#if __AVX2__
-                __m256i _sum0 = _mm256_setzero_si256();
-                __m256i _sum1 = _mm256_setzero_si256();
-#else
-                __m128i _sum00 = _mm_setzero_si128();
-                __m128i _sum10 = _mm_setzero_si128();
-                __m128i _sum01 = _mm_setzero_si128();
-                __m128i _sum11 = _mm_setzero_si128();
-#endif
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-
-                        __m128i _r0;
-                        __m128i _r1;
-                        if (elempack == 8)
-                        {
-                            _r0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            _r1 = _mm_loadl_epi64((const __m128i*)r1s);
-                        }
-                        else // if (elempack == 1)
-                        {
-#if __AVX2__
-                            __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                            _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-#if __AVX512F__
-                            _r0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1));
-                            _r1 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1));
-#else
-                            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-                            __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-                            __m256i _val0_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1), _sindex88);
-                            __m256i _val1_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1), _sindex88);
-                            _r0 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val0_32, 0), _mm256_extracti128_si256(_val0_32, 1));
-                            _r1 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val1_32, 0), _mm256_extracti128_si256(_val1_32, 1));
-#endif // __AVX512F__
-#else
-                            _r0 = _mm_setr_epi8(r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                            _r1 = _mm_setr_epi8(r1s[0], r1s[N], r1s[N * 2], r1s[N * 3], r1s[N * 4], r1s[N * 5], r1s[N * 6], r1s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-#endif // __AVX2__
-                        }
-
-#if __SSE4_1__
-                        _r0 = _mm_cvtepi8_epi16(_r0);
-                        _r1 = _mm_cvtepi8_epi16(_r1);
-#else
-                        _r0 = _mm_unpacklo_epi8(_r0, _mm_cmpgt_epi8(_mm_setzero_si128(), _r0));
-                        _r1 = _mm_unpacklo_epi8(_r1, _mm_cmpgt_epi8(_mm_setzero_si128(), _r1));
-#endif
-
-                        __m128i _w01 = _mm_load_si128((const __m128i*)kptr);
-#if __AVX2__
-                        __m256i _w = _mm256_cvtepi8_epi16(_w01);
-
-                        __m256i _valval0 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r0), _r0, 1);
-                        __m256i _valval1 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r1), _r1, 1);
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum0 = _mm256_dpwssd_epi32(_sum0, _valval0, _w);
-                        _sum1 = _mm256_dpwssd_epi32(_sum1, _valval1, _w);
-#else
-                        _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_valval0, _w));
-                        _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_valval1, _w));
-#endif
-#else // __AVX2__
-                        __m128i _extw01 = _mm_cmpgt_epi8(_mm_setzero_si128(), _w01);
-                        __m128i _w0 = _mm_unpacklo_epi8(_w01, _extw01);
-                        __m128i _w1 = _mm_unpackhi_epi8(_w01, _extw01);
-
-#if __XOP__
-                        _sum00 = _mm_maddd_epi16(_r0, _w0, _sum00);
-                        _sum10 = _mm_maddd_epi16(_r0, _w1, _sum10);
-                        _sum01 = _mm_maddd_epi16(_r1, _w0, _sum01);
-                        _sum11 = _mm_maddd_epi16(_r1, _w1, _sum11);
-#else
-                        _sum00 = _mm_add_epi32(_sum00, _mm_madd_epi16(_r0, _w0));
-                        _sum10 = _mm_add_epi32(_sum10, _mm_madd_epi16(_r0, _w1));
-                        _sum01 = _mm_add_epi32(_sum01, _mm_madd_epi16(_r1, _w0));
-                        _sum11 = _mm_add_epi32(_sum11, _mm_madd_epi16(_r1, _w1));
-#endif // __XOP__
-#endif // __AVX2__
-
-                        kptr += 16;
-                    }
-                }
-#if __AVX2__
-                sum00 += _mm_reduce_add_epi32(_mm256_extracti128_si256(_sum0, 0));
-                sum10 += _mm_reduce_add_epi32(_mm256_extracti128_si256(_sum0, 1));
-                sum01 += _mm_reduce_add_epi32(_mm256_extracti128_si256(_sum1, 0));
-                sum11 += _mm_reduce_add_epi32(_mm256_extracti128_si256(_sum1, 1));
-#else
-                sum00 += _mm_reduce_add_epi32(_sum00);
-                sum10 += _mm_reduce_add_epi32(_sum10);
-                sum01 += _mm_reduce_add_epi32(_sum01);
-                sum11 += _mm_reduce_add_epi32(_sum11);
-#endif // __AVX2__
-            }
-#endif // __SSE2__
-            for (; q + 1 < inch; q += 2)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        sum00 += r0s[0] * kptr[0];
-                        sum10 += r0s[0] * kptr[1];
-                        sum00 += r0s[N] * kptr[2];
-                        sum10 += r0s[N] * kptr[3];
-                        sum01 += r1s[0] * kptr[0];
-                        sum11 += r1s[0] * kptr[1];
-                        sum01 += r1s[N] * kptr[2];
-                        sum11 += r1s[N] * kptr[3];
-
-                        kptr += 4;
-                    }
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        sum00 += r0s[0] * kptr[0];
-                        sum10 += r0s[0] * kptr[1];
-                        sum01 += r1s[0] * kptr[0];
-                        sum11 += r1s[0] * kptr[1];
-
-                        kptr += 2;
-                    }
-                }
-            }
-
-            outptr0[0] = sum00;
-            outptr0[1] = sum01;
-            outptr1[0] = sum10;
-            outptr1[1] = sum11;
-            outptr0 += 2;
-            outptr1 += 2;
-        }
-        for (; ij < outw * outh; ij++)
-        {
-            const int i = ij / outw;
-            const int j = ij % outw;
-
-            int sum0 = 0;
-            int sum1 = 0;
-
-#if __AVX512F__
-            const signed char* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8 + (p % 8) / 4 + (p % 4) / 2);
-#elif __AVX2__
-            const signed char* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2);
-#elif __SSE2__
-            const signed char* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 2);
-#endif
-
-            int q = 0;
-#if __SSE2__
-#if __AVX512F__
-            {
-                __m512i _sum01 = _mm512_setzero_si512();
-                for (; q + 15 < inch; q += 16)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-
-                        __m128i _r0;
-                        if (elempack == 16)
-                        {
-                            _r0 = _mm_load_si128((const __m128i*)r0s);
-                        }
-                        else if (elempack == 8)
-                        {
-                            __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            __m128i _t1 = _mm_loadl_epi64((const __m128i*)(r0s + N));
-                            _r0 = _mm_unpacklo_epi64(_t0, _t1);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                            _r0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0s), 1));
-                        }
-
-                        __m256i _val = _mm256_cvtepi8_epi16(_r0);
-                        __m512i _valval = _mm512_inserti64x4(_mm512_castsi256_si512(_val), _val, 1);
-
-                        __m512i _w = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*)kptr));
-
-#if __AVX512VNNI__
-                        _sum01 = _mm512_dpwssd_epi32(_sum01, _valval, _w);
-#else
-                        _sum01 = _mm512_add_epi32(_sum01, _mm512_madd_epi16(_valval, _w));
-#endif
-
-                        kptr += 32;
-                    }
-                }
-                __m256i _sum0101 = _mm256_hadd_epi32(_mm512_extracti64x4_epi64(_sum01, 0), _mm512_extracti64x4_epi64(_sum01, 1));
-                __m128i _ss = _mm_add_epi32(_mm256_extracti128_si256(_sum0101, 0), _mm256_extracti128_si256(_sum0101, 1));
-
-                sum0 += _mm_extract_epi32(_ss, 0) + _mm_extract_epi32(_ss, 1);
-                sum1 += _mm_extract_epi32(_ss, 2) + _mm_extract_epi32(_ss, 3);
-            }
-#endif // __AVX512F__
-            {
-#if __AVX2__
-                __m256i _sum = _mm256_setzero_si256();
-#else
-                __m128i _sum0 = _mm_setzero_si128();
-                __m128i _sum1 = _mm_setzero_si128();
-#endif
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-
-                        __m128i _r0;
-                        if (elempack == 8)
-                        {
-                            _r0 = _mm_loadl_epi64((const __m128i*)r0s);
-                        }
-                        else // if (elempack == 1)
-                        {
-#if __AVX2__
-                            __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                            _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-#if __AVX512F__
-                            _r0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1));
-#else
-                            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-                            __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-                            __m256i _val32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1), _sindex88);
-                            _r0 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val32, 0), _mm256_extracti128_si256(_val32, 1));
-#endif // __AVX512F__
-#else
-                            _r0 = _mm_setr_epi8(r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-#endif // __AVX2__
-                        }
-
-#if __SSE4_1__
-                        _r0 = _mm_cvtepi8_epi16(_r0);
-#else
-                        _r0 = _mm_unpacklo_epi8(_r0, _mm_cmpgt_epi8(_mm_setzero_si128(), _r0));
-#endif
-
-                        __m128i _w01 = _mm_load_si128((const __m128i*)kptr);
-#if __AVX2__
-                        __m256i _w = _mm256_cvtepi8_epi16(_w01);
-
-                        __m256i _rr0 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r0), _r0, 1);
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum = _mm256_dpwssd_epi32(_sum, _rr0, _w);
-#else
-                        _sum = _mm256_add_epi32(_sum, _mm256_madd_epi16(_rr0, _w));
-#endif
-#else // __AVX2__
-                        __m128i _extw01 = _mm_cmpgt_epi8(_mm_setzero_si128(), _w01);
-                        __m128i _w0 = _mm_unpacklo_epi8(_w01, _extw01);
-                        __m128i _w1 = _mm_unpackhi_epi8(_w01, _extw01);
-
-#if __XOP__
-                        _sum0 = _mm_maddd_epi16(_r0, _w0, _sum0);
-                        _sum1 = _mm_maddd_epi16(_r0, _w1, _sum1);
-#else
-                        _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_r0, _w0));
-                        _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_r0, _w1));
-#endif // __XOP__
-#endif // __AVX2__
-
-                        kptr += 16;
-                    }
-                }
-#if __AVX2__
-                sum0 += _mm_reduce_add_epi32(_mm256_extracti128_si256(_sum, 0));
-                sum1 += _mm_reduce_add_epi32(_mm256_extracti128_si256(_sum, 1));
-#else
-                sum0 += _mm_reduce_add_epi32(_sum0);
-                sum1 += _mm_reduce_add_epi32(_sum1);
-#endif // __AVX2__
-            }
-#endif // __SSE2__
-            for (; q + 1 < inch; q += 2)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i * stride_h) + j * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        sum0 += r0s[0] * kptr[0];
-                        sum1 += r0s[0] * kptr[1];
-                        sum0 += r0s[N] * kptr[2];
-                        sum1 += r0s[N] * kptr[3];
-
-                        kptr += 4;
-                    }
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i * stride_h) + j * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        sum0 += r0s[0] * kptr[0];
-                        sum1 += r0s[0] * kptr[1];
-
-                        kptr += 2;
-                    }
-                }
-            }
-
-            outptr0[0] = sum0;
-            outptr1[0] = sum1;
-            outptr0 += 1;
-            outptr1 += 1;
-        }
-    }
-    remain_outch_start += nn_outch * 2;
-    for (int p = remain_outch_start; p < outch; p++)
-    {
-        int* outptr = top_blob.channel(p);
-
-        int ij = 0;
-#if __SSE2__
-        for (; ij + 3 < outw * outh; ij += 4)
-        {
-            const int i0 = ij / outw;
-            const int i1 = (ij + 1) / outw;
-            const int i2 = (ij + 2) / outw;
-            const int i3 = (ij + 3) / outw;
-            const int j0 = ij % outw;
-            const int j1 = (ij + 1) % outw;
-            const int j2 = (ij + 2) % outw;
-            const int j3 = (ij + 3) % outw;
-
-            __m128i _sum = _mm_setzero_si128();
-
-#if __AVX512F__
-            const signed char* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#elif __AVX2__
-            const signed char* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2 + p % 2);
-#endif
-
-            int q = 0;
-#if __AVX512F__
-            {
-                __m256i _sum0 = _mm256_setzero_si256();
-                __m256i _sum1 = _mm256_setzero_si256();
-                __m256i _sum2 = _mm256_setzero_si256();
-                __m256i _sum3 = _mm256_setzero_si256();
-                for (; q + 15 < inch; q += 16)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-                    const signed char* r2 = bottom_blob.channel(q / elempack).row<const signed char>(i2 * stride_h) + j2 * stride_w * elempack;
-                    const signed char* r3 = bottom_blob.channel(q / elempack).row<const signed char>(i3 * stride_h) + j3 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-                        const signed char* r2s = r2 + space_ofs[k];
-                        const signed char* r3s = r3 + space_ofs[k];
-
-                        __m128i _r0;
-                        __m128i _r1;
-                        __m128i _r2;
-                        __m128i _r3;
-                        if (elempack == 16)
-                        {
-                            _r0 = _mm_load_si128((const __m128i*)r0s);
-                            _r1 = _mm_load_si128((const __m128i*)r1s);
-                            _r2 = _mm_load_si128((const __m128i*)r2s);
-                            _r3 = _mm_load_si128((const __m128i*)r3s);
-                        }
-                        else if (elempack == 8)
-                        {
-                            __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            __m128i _t1 = _mm_loadl_epi64((const __m128i*)(r0s + N));
-                            __m128i _t2 = _mm_loadl_epi64((const __m128i*)r1s);
-                            __m128i _t3 = _mm_loadl_epi64((const __m128i*)(r1s + N));
-                            __m128i _t4 = _mm_loadl_epi64((const __m128i*)r2s);
-                            __m128i _t5 = _mm_loadl_epi64((const __m128i*)(r2s + N));
-                            __m128i _t6 = _mm_loadl_epi64((const __m128i*)r3s);
-                            __m128i _t7 = _mm_loadl_epi64((const __m128i*)(r3s + N));
-                            _r0 = _mm_unpacklo_epi64(_t0, _t1);
-                            _r1 = _mm_unpacklo_epi64(_t2, _t3);
-                            _r2 = _mm_unpacklo_epi64(_t4, _t5);
-                            _r3 = _mm_unpacklo_epi64(_t6, _t7);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-
-                            _r0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0s), 1));
-                            _r1 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r1s), 1));
-                            _r2 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r2s), 1));
-                            _r3 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r3s), 1));
-                        }
-
-                        __m256i _val0 = _mm256_cvtepi8_epi16(_r0);
-                        __m256i _val1 = _mm256_cvtepi8_epi16(_r1);
-                        __m256i _val2 = _mm256_cvtepi8_epi16(_r2);
-                        __m256i _val3 = _mm256_cvtepi8_epi16(_r3);
-
-                        __m256i _w = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)kptr));
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum0 = _mm256_dpwssd_epi32(_sum0, _val0, _w);
-                        _sum1 = _mm256_dpwssd_epi32(_sum1, _val1, _w);
-                        _sum2 = _mm256_dpwssd_epi32(_sum2, _val2, _w);
-                        _sum3 = _mm256_dpwssd_epi32(_sum3, _val3, _w);
-#else
-                        _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_val0, _w));
-                        _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_val1, _w));
-                        _sum2 = _mm256_add_epi32(_sum2, _mm256_madd_epi16(_val2, _w));
-                        _sum3 = _mm256_add_epi32(_sum3, _mm256_madd_epi16(_val3, _w));
-#endif
-
-                        kptr += 16;
-                    }
-                }
-                _sum0 = _mm256_hadd_epi32(_sum0, _sum1);
-                _sum2 = _mm256_hadd_epi32(_sum2, _sum3);
-                _sum0 = _mm256_hadd_epi32(_sum0, _sum2);
-                _sum = _mm_add_epi32(_sum, _mm256_extracti128_si256(_sum0, 0));
-                _sum = _mm_add_epi32(_sum, _mm256_extracti128_si256(_sum0, 1));
-            }
-#endif // __AVX512F__
-            {
-                __m128i _sum0 = _mm_setzero_si128();
-                __m128i _sum1 = _mm_setzero_si128();
-                __m128i _sum2 = _mm_setzero_si128();
-                __m128i _sum3 = _mm_setzero_si128();
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-                    const signed char* r2 = bottom_blob.channel(q / elempack).row<const signed char>(i2 * stride_h) + j2 * stride_w * elempack;
-                    const signed char* r3 = bottom_blob.channel(q / elempack).row<const signed char>(i3 * stride_h) + j3 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-                        const signed char* r2s = r2 + space_ofs[k];
-                        const signed char* r3s = r3 + space_ofs[k];
-
-                        __m128i _r0;
-                        __m128i _r1;
-                        __m128i _r2;
-                        __m128i _r3;
-                        if (elempack == 8)
-                        {
-                            _r0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            _r1 = _mm_loadl_epi64((const __m128i*)r1s);
-                            _r2 = _mm_loadl_epi64((const __m128i*)r2s);
-                            _r3 = _mm_loadl_epi64((const __m128i*)r3s);
-                        }
-                        else // if (elempack == 1)
-                        {
-#if __AVX2__
-                            __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                            _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-#if __AVX512F__
-                            _r0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1));
-                            _r1 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1));
-                            _r2 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r2s, _vindex, 1));
-                            _r3 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r3s, _vindex, 1));
-#else
-                            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-                            __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-
-                            __m256i _val0_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1), _sindex88);
-                            __m256i _val1_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1), _sindex88);
-                            __m256i _val2_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r2s, _vindex, 1), _sindex88);
-                            __m256i _val3_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r3s, _vindex, 1), _sindex88);
-                            _r0 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val0_32, 0), _mm256_extracti128_si256(_val0_32, 1));
-                            _r1 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val1_32, 0), _mm256_extracti128_si256(_val1_32, 1));
-                            _r2 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val2_32, 0), _mm256_extracti128_si256(_val2_32, 1));
-                            _r3 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val3_32, 0), _mm256_extracti128_si256(_val3_32, 1));
-#endif
-#else
-                            _r0 = _mm_setr_epi8(r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                            _r1 = _mm_setr_epi8(r1s[0], r1s[N], r1s[N * 2], r1s[N * 3], r1s[N * 4], r1s[N * 5], r1s[N * 6], r1s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                            _r2 = _mm_setr_epi8(r2s[0], r2s[N], r2s[N * 2], r2s[N * 3], r2s[N * 4], r2s[N * 5], r2s[N * 6], r2s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                            _r3 = _mm_setr_epi8(r3s[0], r3s[N], r3s[N * 2], r3s[N * 3], r3s[N * 4], r3s[N * 5], r3s[N * 6], r3s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-#endif
-                        }
-
-#if __SSE4_1__
-                        _r0 = _mm_cvtepi8_epi16(_r0);
-                        _r1 = _mm_cvtepi8_epi16(_r1);
-                        _r2 = _mm_cvtepi8_epi16(_r2);
-                        _r3 = _mm_cvtepi8_epi16(_r3);
-#else
-                        _r0 = _mm_unpacklo_epi8(_r0, _mm_cmpgt_epi8(_mm_setzero_si128(), _r0));
-                        _r1 = _mm_unpacklo_epi8(_r1, _mm_cmpgt_epi8(_mm_setzero_si128(), _r1));
-                        _r2 = _mm_unpacklo_epi8(_r2, _mm_cmpgt_epi8(_mm_setzero_si128(), _r2));
-                        _r3 = _mm_unpacklo_epi8(_r3, _mm_cmpgt_epi8(_mm_setzero_si128(), _r3));
-#endif
-
-                        __m128i _w = _mm_loadl_epi64((const __m128i*)kptr);
-#if __SSE4_1__
-                        _w = _mm_cvtepi8_epi16(_w);
-#else
-                        _w = _mm_unpacklo_epi8(_w, _mm_cmpgt_epi8(_mm_setzero_si128(), _w));
-#endif
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum0 = _mm_dpwssd_epi32(_sum0, _r0, _w);
-                        _sum1 = _mm_dpwssd_epi32(_sum1, _r1, _w);
-                        _sum2 = _mm_dpwssd_epi32(_sum2, _r2, _w);
-                        _sum3 = _mm_dpwssd_epi32(_sum3, _r3, _w);
-#elif __XOP__
-                        _sum0 = _mm_maddd_epi16(_r0, _w, _sum0);
-                        _sum1 = _mm_maddd_epi16(_r1, _w, _sum1);
-                        _sum2 = _mm_maddd_epi16(_r2, _w, _sum2);
-                        _sum3 = _mm_maddd_epi16(_r3, _w, _sum3);
-#else
-                        _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_r0, _w));
-                        _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_r1, _w));
-                        _sum2 = _mm_add_epi32(_sum2, _mm_madd_epi16(_r2, _w));
-                        _sum3 = _mm_add_epi32(_sum3, _mm_madd_epi16(_r3, _w));
-#endif
-
-                        kptr += 8;
-                    }
-                }
-#if __SSSE3__
-                __m128i _ss = _mm_hadd_epi32(_mm_hadd_epi32(_sum0, _sum1), _mm_hadd_epi32(_sum2, _sum3));
-                _sum = _mm_add_epi32(_sum, _ss);
-#else
-                // transpose 4x4
-                __m128i _tmp0 = _mm_unpacklo_epi32(_sum0, _sum1);
-                __m128i _tmp1 = _mm_unpacklo_epi32(_sum2, _sum3);
-                __m128i _tmp2 = _mm_unpackhi_epi32(_sum0, _sum1);
-                __m128i _tmp3 = _mm_unpackhi_epi32(_sum2, _sum3);
-                _sum0 = _mm_unpacklo_epi64(_tmp0, _tmp1);
-                _sum1 = _mm_unpackhi_epi64(_tmp0, _tmp1);
-                _sum2 = _mm_unpacklo_epi64(_tmp2, _tmp3);
-                _sum3 = _mm_unpackhi_epi64(_tmp2, _tmp3);
-                _sum0 = _mm_add_epi32(_sum0, _sum1);
-                _sum2 = _mm_add_epi32(_sum2, _sum3);
-                _sum = _mm_add_epi32(_sum, _sum0);
-                _sum = _mm_add_epi32(_sum, _sum2);
-#endif // __SSSE3__
-            }
-            for (; q + 1 < inch; q += 2)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-                const signed char* r2 = bottom_blob.channel(q).row<const signed char>(i2 * stride_h) + j2 * stride_w;
-                const signed char* r3 = bottom_blob.channel(q).row<const signed char>(i3 * stride_h) + j3 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-                    const signed char* r2s = r2 + space_ofs[k];
-                    const signed char* r3s = r3 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _r = _mm_setr_epi16(r0s[0], r0s[N], r1s[0], r1s[N], r2s[0], r2s[N], r3s[0], r3s[N]);
-                        __m128i _w = _mm_setr_epi16(kptr[0], kptr[1], kptr[0], kptr[1], kptr[0], kptr[1], kptr[0], kptr[1]);
-
-#if __XOP__
-                        _sum = _mm_maddd_epi16(_r, _w, _sum);
-#else
-                        _sum = _mm_add_epi32(_sum, _mm_madd_epi16(_r, _w));
-#endif // __XOP__
-
-                        kptr += 2;
-                    }
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-                const signed char* r2 = bottom_blob.channel(q).row<const signed char>(i2 * stride_h) + j2 * stride_w;
-                const signed char* r3 = bottom_blob.channel(q).row<const signed char>(i3 * stride_h) + j3 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-                    const signed char* r2s = r2 + space_ofs[k];
-                    const signed char* r3s = r3 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        __m128i _w = _mm_set1_epi16(kptr[0]);
-
-#if __XOP__
-                        __m128i _r = _mm_setr_epi16(r0s[0], 0, r1s[0], 0, r2s[0], 0, r3s[0], 0);
-
-                        _sum = _mm_maccd_epi16(_r, _w, _sum);
-#else
-                        __m128i _r = _mm_setr_epi16(r0s[0], r1s[0], r2s[0], r3s[0], 0, 0, 0, 0);
-
-                        __m128i _sl = _mm_mullo_epi16(_r, _w);
-                        __m128i _sh = _mm_mulhi_epi16(_r, _w);
-                        __m128i _s0 = _mm_unpacklo_epi16(_sl, _sh);
-
-                        _sum = _mm_add_epi32(_sum, _s0);
-#endif // __XOP__
-
-                        kptr += 1;
-                    }
-                }
-            }
-
-            _mm_store_si128((__m128i*)outptr, _sum);
-            outptr += 4;
-        }
-#endif // __SSE2__
-        for (; ij + 1 < outw * outh; ij += 2)
-        {
-            const int i0 = ij / outw;
-            const int i1 = (ij + 1) / outw;
-            const int j0 = ij % outw;
-            const int j1 = (ij + 1) % outw;
-
-            int sum0 = 0;
-            int sum1 = 0;
-
-#if __AVX512F__
-            const signed char* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#elif __AVX2__
-            const signed char* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#elif __SSE2__
-            const signed char* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2 + p % 2);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 2 + p % 2);
-#endif
-
-            int q = 0;
-#if __SSE2__
-#if __AVX512F__
-            {
-                __m256i _sum0 = _mm256_setzero_si256();
-                __m256i _sum1 = _mm256_setzero_si256();
-                for (; q + 15 < inch; q += 16)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-
-                        __m128i _r0;
-                        __m128i _r1;
-                        if (elempack == 16)
-                        {
-                            _r0 = _mm_load_si128((const __m128i*)r0s);
-                            _r1 = _mm_load_si128((const __m128i*)r1s);
-                        }
-                        else if (elempack == 8)
-                        {
-                            __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            __m128i _t1 = _mm_loadl_epi64((const __m128i*)(r0s + N));
-                            __m128i _t2 = _mm_loadl_epi64((const __m128i*)r1s);
-                            __m128i _t3 = _mm_loadl_epi64((const __m128i*)(r1s + N));
-                            _r0 = _mm_unpacklo_epi64(_t0, _t1);
-                            _r1 = _mm_unpacklo_epi64(_t2, _t3);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-
-                            _r0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0s), 1));
-                            _r1 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r1s), 1));
-                        }
-
-                        __m256i _val0 = _mm256_cvtepi8_epi16(_r0);
-                        __m256i _val1 = _mm256_cvtepi8_epi16(_r1);
-
-                        __m256i _w = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)kptr));
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum0 = _mm256_dpwssd_epi32(_sum0, _val0, _w);
-                        _sum1 = _mm256_dpwssd_epi32(_sum1, _val1, _w);
-#else
-                        _sum0 = _mm256_add_epi32(_sum0, _mm256_madd_epi16(_val0, _w));
-                        _sum1 = _mm256_add_epi32(_sum1, _mm256_madd_epi16(_val1, _w));
-#endif
-
-                        kptr += 16;
-                    }
-                }
-
-                __m128i _ss0 = _mm_add_epi32(_mm256_extracti128_si256(_sum0, 0), _mm256_extracti128_si256(_sum0, 1));
-                __m128i _ss1 = _mm_add_epi32(_mm256_extracti128_si256(_sum1, 0), _mm256_extracti128_si256(_sum1, 1));
-                sum0 += _mm_reduce_add_epi32(_ss0);
-                sum1 += _mm_reduce_add_epi32(_ss1);
-            }
-#endif // __AVX512F__
-            {
-                __m128i _sum0 = _mm_setzero_si128();
-                __m128i _sum1 = _mm_setzero_si128();
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i0 * stride_h) + j0 * stride_w * elempack;
-                    const signed char* r1 = bottom_blob.channel(q / elempack).row<const signed char>(i1 * stride_h) + j1 * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-                        const signed char* r1s = r1 + space_ofs[k];
-
-                        __m128i _r0;
-                        __m128i _r1;
-                        if (elempack == 8)
-                        {
-                            _r0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            _r1 = _mm_loadl_epi64((const __m128i*)r1s);
-                        }
-                        else // if (elempack == 1)
-                        {
-#if __AVX2__
-                            __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                            _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-#if __AVX512F__
-                            _r0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1));
-                            _r1 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1));
-#else
-                            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-                            __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-
-                            __m256i _val0_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1), _sindex88);
-                            __m256i _val1_32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r1s, _vindex, 1), _sindex88);
-                            _r0 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val0_32, 0), _mm256_extracti128_si256(_val0_32, 1));
-                            _r1 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val1_32, 0), _mm256_extracti128_si256(_val1_32, 1));
-#endif
-#else
-                            _r0 = _mm_setr_epi8(r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-                            _r1 = _mm_setr_epi8(r1s[0], r1s[N], r1s[N * 2], r1s[N * 3], r1s[N * 4], r1s[N * 5], r1s[N * 6], r1s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-#endif
-                        }
-
-#if __SSE4_1__
-                        _r0 = _mm_cvtepi8_epi16(_r0);
-                        _r1 = _mm_cvtepi8_epi16(_r1);
-#else
-                        _r0 = _mm_unpacklo_epi8(_r0, _mm_cmpgt_epi8(_mm_setzero_si128(), _r0));
-                        _r1 = _mm_unpacklo_epi8(_r1, _mm_cmpgt_epi8(_mm_setzero_si128(), _r1));
-#endif
-
-                        __m128i _w = _mm_loadl_epi64((const __m128i*)kptr);
-#if __SSE4_1__
-                        _w = _mm_cvtepi8_epi16(_w);
-#else
-                        _w = _mm_unpacklo_epi8(_w, _mm_cmpgt_epi8(_mm_setzero_si128(), _w));
-#endif
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum0 = _mm_dpwssd_epi32(_sum0, _r0, _w);
-                        _sum1 = _mm_dpwssd_epi32(_sum1, _r1, _w);
-#elif __XOP__
-                        _sum0 = _mm_maddd_epi16(_r0, _w, _sum0);
-                        _sum1 = _mm_maddd_epi16(_r1, _w, _sum1);
-#else
-                        _sum0 = _mm_add_epi32(_sum0, _mm_madd_epi16(_r0, _w));
-                        _sum1 = _mm_add_epi32(_sum1, _mm_madd_epi16(_r1, _w));
-#endif
-
-                        kptr += 8;
-                    }
-                }
-                sum0 += _mm_reduce_add_epi32(_sum0);
-                sum1 += _mm_reduce_add_epi32(_sum1);
-            }
-#endif // __SSE2__
-            for (; q + 1 < inch; q += 2)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        sum0 += r0s[0] * kptr[0];
-                        sum0 += r0s[N] * kptr[1];
-                        sum1 += r1s[0] * kptr[0];
-                        sum1 += r1s[N] * kptr[1];
-
-                        kptr += 2;
-                    }
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i0 * stride_h) + j0 * stride_w;
-                const signed char* r1 = bottom_blob.channel(q).row<const signed char>(i1 * stride_h) + j1 * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-                    const signed char* r1s = r1 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        sum0 += r0s[0] * kptr[0];
-                        sum1 += r1s[0] * kptr[0];
-
-                        kptr += 1;
-                    }
-                }
-            }
-
-            outptr[0] = sum0;
-            outptr[1] = sum1;
-            outptr += 2;
-        }
-        for (; ij < outw * outh; ij++)
-        {
-            const int i = ij / outw;
-            const int j = ij % outw;
-
-            int sum = 0;
-
-#if __AVX512F__
-            const signed char* kptr = weight_data_tm.channel(p / 16 + (p % 16) / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#elif __AVX2__
-            const signed char* kptr = weight_data_tm.channel(p / 8 + (p % 8) / 4 + (p % 4) / 2 + p % 2);
-#elif __SSE2__
-            const signed char* kptr = weight_data_tm.channel(p / 4 + (p % 4) / 2 + p % 2);
-#else
-            const signed char* kptr = weight_data_tm.channel(p / 2 + p % 2);
-#endif
-
-            int q = 0;
-#if __SSE2__
-#if __AVX512F__
-            {
-                __m256i _sum = _mm256_setzero_si256();
-                for (; q + 15 < inch; q += 16)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-
-                        __m128i _r0;
-                        if (elempack == 16)
-                        {
-                            _r0 = _mm_loadu_si128((const __m128i*)r0s);
-                        }
-                        else if (elempack == 8)
-                        {
-                            __m128i _t0 = _mm_loadl_epi64((const __m128i*)r0s);
-                            __m128i _t1 = _mm_loadl_epi64((const __m128i*)(r0s + N));
-                            _r0 = _mm_unpacklo_epi64(_t0, _t1);
-                        }
-                        else // if (elempack == 1)
-                        {
-                            __m512i _vindex = _mm512_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-                            _vindex = _mm512_mullo_epi32(_vindex, _mm512_set1_epi32(N));
-                            _r0 = _mm512_cvtepi32_epi8(_mm512_i32gather_epi32(_vindex, (const int*)(r0s), 1));
-                        }
-
-                        __m256i _val = _mm256_cvtepi8_epi16(_r0);
-
-                        __m256i _w = _mm256_cvtepi8_epi16(_mm_load_si128((const __m128i*)kptr));
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum = _mm256_dpwssd_epi32(_sum, _val, _w);
-#else
-                        _sum = _mm256_add_epi32(_sum, _mm256_madd_epi16(_val, _w));
-#endif
-
-                        kptr += 16;
-                    }
-                }
-
-                __m128i _ss = _mm_add_epi32(_mm256_extracti128_si256(_sum, 0), _mm256_extracti128_si256(_sum, 1));
-                sum += _mm_reduce_add_epi32(_ss);
-            }
-#endif // __AVX512F__
-            {
-                __m128i _sum = _mm_setzero_si128();
-                for (; q + 7 < inch; q += 8)
-                {
-                    const signed char* r0 = bottom_blob.channel(q / elempack).row<const signed char>(i * stride_h) + j * stride_w * elempack;
-
-                    for (int k = 0; k < maxk; k++)
-                    {
-                        const signed char* r0s = r0 + space_ofs[k];
-
-                        __m128i _r0;
-                        if (elempack == 8)
-                        {
-                            _r0 = _mm_loadl_epi64((const __m128i*)r0s);
-                        }
-                        else // if (elempack == 1)
-                        {
-#if __AVX2__
-                            __m256i _vindex = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
-                            _vindex = _mm256_mullo_epi32(_vindex, _mm256_set1_epi32(N));
-#if __AVX512F__
-                            _r0 = _mm256_cvtepi32_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1));
-#else
-                            __m128i _sindex8 = _mm_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1);
-                            __m256i _sindex88 = _mm256_inserti128_si256(_mm256_castsi128_si256(_sindex8), _sindex8, 1);
-                            __m256i _val32 = _mm256_shuffle_epi8(_mm256_i32gather_epi32((const int*)r0s, _vindex, 1), _sindex88);
-                            _r0 = _mm_unpacklo_epi32(_mm256_extracti128_si256(_val32, 0), _mm256_extracti128_si256(_val32, 1));
-#endif // __AVX512F__
-#else
-                            _r0 = _mm_setr_epi8(r0s[0], r0s[N], r0s[N * 2], r0s[N * 3], r0s[N * 4], r0s[N * 5], r0s[N * 6], r0s[N * 7], 0, 0, 0, 0, 0, 0, 0, 0);
-#endif // __AVX2__
-                        }
-
-#if __SSE4_1__
-                        _r0 = _mm_cvtepi8_epi16(_r0);
-#else
-                        _r0 = _mm_unpacklo_epi8(_r0, _mm_cmpgt_epi8(_mm_setzero_si128(), _r0));
-#endif
-
-                        __m128i _w = _mm_loadl_epi64((const __m128i*)kptr);
-#if __SSE4_1__
-                        _w = _mm_cvtepi8_epi16(_w);
-#else
-                        _w = _mm_unpacklo_epi8(_w, _mm_cmpgt_epi8(_mm_setzero_si128(), _w));
-#endif
-
-#if __AVXVNNI__ || __AVX512VNNI__
-                        _sum = _mm_dpwssd_epi32(_sum, _r0, _w);
-#elif __XOP__
-                        _sum = _mm_maddd_epi16(_r0, _w, _sum);
-#else
-                        _sum = _mm_add_epi32(_sum, _mm_madd_epi16(_r0, _w));
-#endif
-
-                        kptr += 8;
-                    }
-                }
-                sum += _mm_reduce_add_epi32(_sum);
-            }
-#endif // __SSE2__
-            for (; q + 1 < inch; q += 2)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i * stride_h) + j * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        sum += r0s[0] * kptr[0];
-                        sum += r0s[N] * kptr[1];
-
-                        kptr += 2;
-                    }
-                }
-            }
-            for (; q < inch; q++)
-            {
-                const signed char* r0 = bottom_blob.channel(q).row<const signed char>(i * stride_h) + j * stride_w;
-
-                for (int k = 0; k < maxk; k++)
-                {
-                    const signed char* r0s = r0 + space_ofs[k];
-
-                    // if (elempack == 1)
-                    {
-                        sum += r0s[0] * kptr[0];
-
-                        kptr += 1;
-                    }
-                }
-            }
-
-            outptr[0] = sum;
-            outptr += 1;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_x86.h
deleted file mode 100644
index fdfa88f7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolution_x86.h
+++ /dev/null
@@ -1,63 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTION_X86_H
-#define LAYER_CONVOLUTION_X86_H
-
-#include "convolution.h"
-
-namespace ncnn {
-
-class Convolution_x86 : public Convolution
-{
-public:
-    Convolution_x86();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-#if NCNN_INT8
-    int create_pipeline_int8_x86(const Option& opt);
-    int forward_int8_x86(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-    int forwardDilation_x86(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-public:
-    Layer* activation;
-
-    int nT;
-    Mat weight_data_tm;
-    Mat weight_sgemm_data;
-    Mat weight_winograd23_data;
-    Mat weight_winograd43_data;
-    Mat weight_winograd63_data;
-
-    // forwardDilation
-    Layer* convolution_dilation1;
-
-    Layer* gemm;
-
-#if NCNN_INT8
-    Mat scale_in_data;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTION_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3.h
deleted file mode 100644
index fd99f105..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3.h
+++ /dev/null
@@ -1,199 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const float bias0 = bias ? bias[g] : 0.f;
-
-        const float* kernel0 = kernel + g * 9;
-
-        float* outptr = out;
-        float* outptr2 = outptr + outw;
-
-        const float* img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0;
-        const float* r1 = img0 + w;
-        const float* r2 = img0 + w * 2;
-        const float* r3 = img0 + w * 3;
-
-        const float* k0 = kernel0;
-        const float* k1 = kernel0 + 3;
-        const float* k2 = kernel0 + 6;
-
-        int i = 0;
-
-        for (; i + 1 < outh; i += 2)
-        {
-            int remain = outw;
-
-            for (; remain > 0; remain--)
-            {
-                float sum = bias0;
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-
-                float sum2 = bias0;
-                sum2 += r1[0] * k0[0];
-                sum2 += r1[1] * k0[1];
-                sum2 += r1[2] * k0[2];
-                sum2 += r2[0] * k1[0];
-                sum2 += r2[1] * k1[1];
-                sum2 += r2[2] * k1[2];
-                sum2 += r3[0] * k2[0];
-                sum2 += r3[1] * k2[1];
-                sum2 += r3[2] * k2[2];
-
-                *outptr = sum;
-                *outptr2 = sum2;
-
-                r0++;
-                r1++;
-                r2++;
-                r3++;
-                outptr++;
-                outptr2++;
-            }
-
-            r0 += 2 + w;
-            r1 += 2 + w;
-            r2 += 2 + w;
-            r3 += 2 + w;
-
-            outptr += outw;
-            outptr2 += outw;
-        }
-
-        for (; i < outh; i++)
-        {
-            int remain = outw;
-
-            for (; remain > 0; remain--)
-            {
-                float sum = bias0;
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-
-                *outptr = sum;
-
-                r0++;
-                r1++;
-                r2++;
-                outptr++;
-            }
-
-            r0 += 2;
-            r1 += 2;
-            r2 += 2;
-        }
-    }
-}
-
-static void convdw3x3s2_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const float* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        const float bias0 = bias ? bias[g] : 0.f;
-
-        const float* kernel0 = kernel + g * 9;
-
-        float* outptr = out;
-
-        const float* img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0;
-        const float* r1 = img0 + w;
-        const float* r2 = img0 + w * 2;
-
-        const float* k0 = kernel0;
-        const float* k1 = kernel0 + 3;
-        const float* k2 = kernel0 + 6;
-
-        int i = 0;
-
-        for (; i < outh; i++)
-        {
-            int remain = outw;
-
-            for (; remain > 0; remain--)
-            {
-                float sum = bias0;
-                sum += r0[0] * k0[0];
-                sum += r0[1] * k0[1];
-                sum += r0[2] * k0[2];
-                sum += r1[0] * k1[0];
-                sum += r1[1] * k1[1];
-                sum += r1[2] * k1[2];
-                sum += r2[0] * k2[0];
-                sum += r2[1] * k2[1];
-                sum += r2[2] * k2[2];
-
-                *outptr = sum;
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-                outptr++;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_int8.h
deleted file mode 100644
index 6c424995..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_int8.h
+++ /dev/null
@@ -1,405 +0,0 @@
-// BUG1989 is pleased to support the open source community by supporting ncnn available.
-//
-// Copyright (C) 2019 BUG1989. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_int8_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    //int h = bottom_blob.h;
-    //int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const signed char* kernel = _kernel;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-
-        out.fill(0);
-
-        const signed char* kernel0 = (const signed char*)kernel + p * 9;
-
-        int* outptr = out;
-
-        const signed char* img0 = bottom_blob.channel(p);
-
-        const signed char* r0 = img0;
-        const signed char* r1 = img0 + w;
-        const signed char* r2 = img0 + w * 2;
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int remain = outw;
-
-            for (; remain > 0; remain--)
-            {
-                int sum = 0;
-
-                sum += (int)r0[0] * (int)kernel0[0];
-                sum += (int)r0[1] * (int)kernel0[1];
-                sum += (int)r0[2] * (int)kernel0[2];
-                sum += (int)r1[0] * (int)kernel0[3];
-                sum += (int)r1[1] * (int)kernel0[4];
-                sum += (int)r1[2] * (int)kernel0[5];
-                sum += (int)r2[0] * (int)kernel0[6];
-                sum += (int)r2[1] * (int)kernel0[7];
-                sum += (int)r2[2] * (int)kernel0[8];
-
-                *outptr += sum;
-
-                r0++;
-                r1++;
-                r2++;
-                outptr++;
-            }
-
-            r0 += 2;
-            r1 += 2;
-            r2 += 2;
-        }
-    }
-}
-
-static void convdw3x3s2_int8_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Option& opt)
-{
-    int w = bottom_blob.w;
-    //int h = bottom_blob.h;
-    //int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const signed char* kernel = _kernel;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-        out.fill(0);
-
-        const signed char* kernel0 = (const signed char*)kernel + p * 9;
-
-        int* outptr = out;
-
-        const signed char* img0 = bottom_blob.channel(p);
-
-        const signed char* r0 = img0;
-        const signed char* r1 = img0 + w;
-        const signed char* r2 = img0 + w * 2;
-
-        int i = 0;
-
-        for (; i < outh; i++)
-        {
-            int remain = outw;
-
-            for (; remain > 0; remain--)
-            {
-                int sum = 0;
-
-                sum += (int)r0[0] * (int)kernel0[0];
-                sum += (int)r0[1] * (int)kernel0[1];
-                sum += (int)r0[2] * (int)kernel0[2];
-                sum += (int)r1[0] * (int)kernel0[3];
-                sum += (int)r1[1] * (int)kernel0[4];
-                sum += (int)r1[2] * (int)kernel0[5];
-                sum += (int)r2[0] * (int)kernel0[6];
-                sum += (int)r2[1] * (int)kernel0[7];
-                sum += (int)r2[2] * (int)kernel0[8];
-
-                *outptr += sum;
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-                outptr++;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
-
-static void convdw3x3s1_int8_dequant_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, std::vector<float> scales_dequant, const Option& opt)
-{
-    int w = bottom_blob.w;
-    //int h = bottom_blob.h;
-    //int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const signed char* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-        float* outptr = out;
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        const float scale_dequant = scales_dequant[p];
-
-        out.fill(bias0);
-
-        const signed char* kernel0 = (const signed char*)kernel + p * 9;
-
-        const signed char* img0 = bottom_blob.channel(p);
-        const signed char* r0 = img0;
-        const signed char* r1 = img0 + w;
-        const signed char* r2 = img0 + w * 2;
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int remain = outw;
-
-            for (; remain > 0; remain--)
-            {
-                int sum = 0;
-
-                sum += (int)r0[0] * (int)kernel0[0];
-                sum += (int)r0[1] * (int)kernel0[1];
-                sum += (int)r0[2] * (int)kernel0[2];
-                sum += (int)r1[0] * (int)kernel0[3];
-                sum += (int)r1[1] * (int)kernel0[4];
-                sum += (int)r1[2] * (int)kernel0[5];
-                sum += (int)r2[0] * (int)kernel0[6];
-                sum += (int)r2[1] * (int)kernel0[7];
-                sum += (int)r2[2] * (int)kernel0[8];
-
-                *outptr += (float)sum * scale_dequant;
-
-                r0++;
-                r1++;
-                r2++;
-                outptr++;
-            }
-
-            r0 += 2;
-            r1 += 2;
-            r2 += 2;
-        }
-    }
-}
-
-static void convdw3x3s2_int8_dequant_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, std::vector<float> scales_dequant, const Option& opt)
-{
-    int w = bottom_blob.w;
-    //int h = bottom_blob.h;
-    //int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const signed char* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-        float* outptr = out;
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        const float scale_dequant = scales_dequant[p];
-
-        out.fill(bias0);
-
-        const signed char* kernel0 = (const signed char*)kernel + p * 9;
-
-        const signed char* img0 = bottom_blob.channel(p);
-        const signed char* r0 = img0;
-        const signed char* r1 = img0 + w;
-        const signed char* r2 = img0 + w * 2;
-
-        int i = 0;
-
-        for (; i < outh; i++)
-        {
-            int remain = outw;
-
-            for (; remain > 0; remain--)
-            {
-                int sum = 0;
-
-                sum += (int)r0[0] * (int)kernel0[0];
-                sum += (int)r0[1] * (int)kernel0[1];
-                sum += (int)r0[2] * (int)kernel0[2];
-                sum += (int)r1[0] * (int)kernel0[3];
-                sum += (int)r1[1] * (int)kernel0[4];
-                sum += (int)r1[2] * (int)kernel0[5];
-                sum += (int)r2[0] * (int)kernel0[6];
-                sum += (int)r2[1] * (int)kernel0[7];
-                sum += (int)r2[2] * (int)kernel0[8];
-
-                *outptr += (float)sum * scale_dequant;
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-                outptr++;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
-
-static void convdw3x3s1_int8_requant_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, std::vector<float> scales_requant, const Option& opt)
-{
-    int w = bottom_blob.w;
-    //int h = bottom_blob.h;
-    //int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const signed char* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-        signed char* outptr = out;
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        const float scale_requant_in = scales_requant[2 * p];
-        const float scale_requant_out = scales_requant[2 * p + 1];
-
-        const signed char* kernel0 = (const signed char*)kernel + p * 9;
-
-        const signed char* img0 = bottom_blob.channel(p);
-        const signed char* r0 = img0;
-        const signed char* r1 = img0 + w;
-        const signed char* r2 = img0 + w * 2;
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int remain = outw;
-
-            for (; remain > 0; remain--)
-            {
-                int sum = 0;
-
-                sum += (int)r0[0] * (int)kernel0[0];
-                sum += (int)r0[1] * (int)kernel0[1];
-                sum += (int)r0[2] * (int)kernel0[2];
-                sum += (int)r1[0] * (int)kernel0[3];
-                sum += (int)r1[1] * (int)kernel0[4];
-                sum += (int)r1[2] * (int)kernel0[5];
-                sum += (int)r2[0] * (int)kernel0[6];
-                sum += (int)r2[1] * (int)kernel0[7];
-                sum += (int)r2[2] * (int)kernel0[8];
-
-                *outptr = float2int8(((float)sum * scale_requant_in + bias0) * scale_requant_out);
-
-                r0++;
-                r1++;
-                r2++;
-                outptr++;
-            }
-
-            r0 += 2;
-            r1 += 2;
-            r2 += 2;
-        }
-    }
-}
-
-static void convdw3x3s2_int8_requant_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& _kernel, const Mat& _bias, std::vector<float> scales_requant, const Option& opt)
-{
-    int w = bottom_blob.w;
-    //int h = bottom_blob.h;
-    //int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-
-    const int tailstep = w - 2 * outw + w;
-
-    const signed char* kernel = _kernel;
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        Mat out = top_blob.channel(p);
-        signed char* outptr = out;
-
-        const float bias0 = bias ? bias[p] : 0.f;
-        const float scale_requant_in = scales_requant[2 * p];
-        const float scale_requant_out = scales_requant[2 * p + 1];
-
-        const signed char* kernel0 = (const signed char*)kernel + p * 9;
-
-        const signed char* img0 = bottom_blob.channel(p);
-        const signed char* r0 = img0;
-        const signed char* r1 = img0 + w;
-        const signed char* r2 = img0 + w * 2;
-
-        int i = 0;
-
-        for (; i < outh; i++)
-        {
-            int remain = outw;
-
-            for (; remain > 0; remain--)
-            {
-                int sum = 0;
-
-                sum += (int)r0[0] * (int)kernel0[0];
-                sum += (int)r0[1] * (int)kernel0[1];
-                sum += (int)r0[2] * (int)kernel0[2];
-                sum += (int)r1[0] * (int)kernel0[3];
-                sum += (int)r1[1] * (int)kernel0[4];
-                sum += (int)r1[2] * (int)kernel0[5];
-                sum += (int)r2[0] * (int)kernel0[6];
-                sum += (int)r2[1] * (int)kernel0[7];
-                sum += (int)r2[2] * (int)kernel0[8];
-
-                *outptr = float2int8(((float)sum * scale_requant_in + bias0) * scale_requant_out);
-
-                r0 += 2;
-                r1 += 2;
-                r2 += 2;
-                outptr++;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
\ No newline at end of file
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_pack16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_pack16.h
deleted file mode 100644
index c0bf77d3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_pack16.h
+++ /dev/null
@@ -1,733 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_pack16_avx512(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m512 _bias0 = bias ? _mm512_loadu_ps((const float*)bias + g * 16) : _mm512_setzero_ps();
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-        float* outptr1 = out.row(1);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-
-        __m512 _k00 = _mm512_load_ps(k0);
-        __m512 _k01 = _mm512_load_ps(k0 + 16);
-        __m512 _k02 = _mm512_load_ps(k0 + 32);
-        __m512 _k10 = _mm512_load_ps(k0 + 48);
-        __m512 _k11 = _mm512_load_ps(k0 + 64);
-        __m512 _k12 = _mm512_load_ps(k0 + 80);
-        __m512 _k20 = _mm512_load_ps(k0 + 96);
-        __m512 _k21 = _mm512_load_ps(k0 + 112);
-        __m512 _k22 = _mm512_load_ps(k0 + 128);
-
-        int i = 0;
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                __m512 _sum00 = _bias0;
-                __m512 _sum01 = _bias0;
-                __m512 _sum02 = _bias0;
-                __m512 _sum03 = _bias0;
-                __m512 _sum10 = _bias0;
-                __m512 _sum11 = _bias0;
-                __m512 _sum12 = _bias0;
-                __m512 _sum13 = _bias0;
-
-                __m512 _r00 = _mm512_load_ps(r0);
-                __m512 _r01 = _mm512_load_ps(r0 + 16);
-                __m512 _r02 = _mm512_load_ps(r0 + 32);
-                __m512 _r03 = _mm512_load_ps(r0 + 48);
-                __m512 _r04 = _mm512_load_ps(r0 + 64);
-                __m512 _r05 = _mm512_load_ps(r0 + 80);
-
-                _sum00 = _mm512_fmadd_ps(_k00, _r00, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k00, _r01, _sum01);
-                _sum02 = _mm512_fmadd_ps(_k00, _r02, _sum02);
-                _sum03 = _mm512_fmadd_ps(_k00, _r03, _sum03);
-                _sum00 = _mm512_fmadd_ps(_k01, _r01, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k01, _r02, _sum01);
-                _sum02 = _mm512_fmadd_ps(_k01, _r03, _sum02);
-                _sum03 = _mm512_fmadd_ps(_k01, _r04, _sum03);
-                _sum00 = _mm512_fmadd_ps(_k02, _r02, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k02, _r03, _sum01);
-                _sum02 = _mm512_fmadd_ps(_k02, _r04, _sum02);
-                _sum03 = _mm512_fmadd_ps(_k02, _r05, _sum03);
-
-                __m512 _r10 = _mm512_load_ps(r1);
-                __m512 _r11 = _mm512_load_ps(r1 + 16);
-                __m512 _r12 = _mm512_load_ps(r1 + 32);
-                __m512 _r13 = _mm512_load_ps(r1 + 48);
-                __m512 _r14 = _mm512_load_ps(r1 + 64);
-                __m512 _r15 = _mm512_load_ps(r1 + 80);
-
-                _sum10 = _mm512_fmadd_ps(_k00, _r10, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k00, _r11, _sum11);
-                _sum12 = _mm512_fmadd_ps(_k00, _r12, _sum12);
-                _sum13 = _mm512_fmadd_ps(_k00, _r13, _sum13);
-                _sum00 = _mm512_fmadd_ps(_k10, _r10, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k10, _r11, _sum01);
-                _sum02 = _mm512_fmadd_ps(_k10, _r12, _sum02);
-                _sum03 = _mm512_fmadd_ps(_k10, _r13, _sum03);
-
-                _sum10 = _mm512_fmadd_ps(_k01, _r11, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k01, _r12, _sum11);
-                _sum12 = _mm512_fmadd_ps(_k01, _r13, _sum12);
-                _sum13 = _mm512_fmadd_ps(_k01, _r14, _sum13);
-                _sum00 = _mm512_fmadd_ps(_k11, _r11, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k11, _r12, _sum01);
-                _sum02 = _mm512_fmadd_ps(_k11, _r13, _sum02);
-                _sum03 = _mm512_fmadd_ps(_k11, _r14, _sum03);
-
-                _sum10 = _mm512_fmadd_ps(_k02, _r12, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k02, _r13, _sum11);
-                _sum12 = _mm512_fmadd_ps(_k02, _r14, _sum12);
-                _sum13 = _mm512_fmadd_ps(_k02, _r15, _sum13);
-                _sum00 = _mm512_fmadd_ps(_k12, _r12, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k12, _r13, _sum01);
-                _sum02 = _mm512_fmadd_ps(_k12, _r14, _sum02);
-                _sum03 = _mm512_fmadd_ps(_k12, _r15, _sum03);
-
-                __m512 _r20 = _mm512_load_ps(r2);
-                __m512 _r21 = _mm512_load_ps(r2 + 16);
-                __m512 _r22 = _mm512_load_ps(r2 + 32);
-                __m512 _r23 = _mm512_load_ps(r2 + 48);
-                __m512 _r24 = _mm512_load_ps(r2 + 64);
-                __m512 _r25 = _mm512_load_ps(r2 + 80);
-
-                _sum10 = _mm512_fmadd_ps(_k10, _r20, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k10, _r21, _sum11);
-                _sum12 = _mm512_fmadd_ps(_k10, _r22, _sum12);
-                _sum13 = _mm512_fmadd_ps(_k10, _r23, _sum13);
-                _sum00 = _mm512_fmadd_ps(_k20, _r20, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k20, _r21, _sum01);
-                _sum02 = _mm512_fmadd_ps(_k20, _r22, _sum02);
-                _sum03 = _mm512_fmadd_ps(_k20, _r23, _sum03);
-
-                _sum10 = _mm512_fmadd_ps(_k11, _r21, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k11, _r22, _sum11);
-                _sum12 = _mm512_fmadd_ps(_k11, _r23, _sum12);
-                _sum13 = _mm512_fmadd_ps(_k11, _r24, _sum13);
-                _sum00 = _mm512_fmadd_ps(_k21, _r21, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k21, _r22, _sum01);
-                _sum02 = _mm512_fmadd_ps(_k21, _r23, _sum02);
-                _sum03 = _mm512_fmadd_ps(_k21, _r24, _sum03);
-
-                _sum10 = _mm512_fmadd_ps(_k12, _r22, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k12, _r23, _sum11);
-                _sum12 = _mm512_fmadd_ps(_k12, _r24, _sum12);
-                _sum13 = _mm512_fmadd_ps(_k12, _r25, _sum13);
-                _sum00 = _mm512_fmadd_ps(_k22, _r22, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k22, _r23, _sum01);
-                _sum02 = _mm512_fmadd_ps(_k22, _r24, _sum02);
-                _sum03 = _mm512_fmadd_ps(_k22, _r25, _sum03);
-
-                __m512 _r30 = _mm512_load_ps(r3);
-                __m512 _r31 = _mm512_load_ps(r3 + 16);
-                __m512 _r32 = _mm512_load_ps(r3 + 32);
-                __m512 _r33 = _mm512_load_ps(r3 + 48);
-                __m512 _r34 = _mm512_load_ps(r3 + 64);
-                __m512 _r35 = _mm512_load_ps(r3 + 80);
-
-                _sum10 = _mm512_fmadd_ps(_k20, _r30, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k20, _r31, _sum11);
-                _sum12 = _mm512_fmadd_ps(_k20, _r32, _sum12);
-                _sum13 = _mm512_fmadd_ps(_k20, _r33, _sum13);
-                _sum10 = _mm512_fmadd_ps(_k21, _r31, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k21, _r32, _sum11);
-                _sum12 = _mm512_fmadd_ps(_k21, _r33, _sum12);
-                _sum13 = _mm512_fmadd_ps(_k21, _r34, _sum13);
-                _sum10 = _mm512_fmadd_ps(_k22, _r32, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k22, _r33, _sum11);
-                _sum12 = _mm512_fmadd_ps(_k22, _r34, _sum12);
-                _sum13 = _mm512_fmadd_ps(_k22, _r35, _sum13);
-
-                _mm512_store_ps(outptr0, _sum00);
-                _mm512_store_ps(outptr0 + 16, _sum01);
-                _mm512_store_ps(outptr0 + 32, _sum02);
-                _mm512_store_ps(outptr0 + 48, _sum03);
-                _mm512_store_ps(outptr1, _sum10);
-                _mm512_store_ps(outptr1 + 16, _sum11);
-                _mm512_store_ps(outptr1 + 32, _sum12);
-                _mm512_store_ps(outptr1 + 48, _sum13);
-
-                r0 += 64;
-                r1 += 64;
-                r2 += 64;
-                r3 += 64;
-                outptr0 += 64;
-                outptr1 += 64;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                __m512 _sum00 = _bias0;
-                __m512 _sum01 = _bias0;
-                __m512 _sum10 = _bias0;
-                __m512 _sum11 = _bias0;
-
-                __m512 _r00 = _mm512_load_ps(r0);
-                __m512 _r01 = _mm512_load_ps(r0 + 16);
-                __m512 _r02 = _mm512_load_ps(r0 + 32);
-                __m512 _r03 = _mm512_load_ps(r0 + 48);
-
-                _sum00 = _mm512_fmadd_ps(_k00, _r00, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k00, _r01, _sum01);
-                _sum00 = _mm512_fmadd_ps(_k01, _r01, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k01, _r02, _sum01);
-                _sum00 = _mm512_fmadd_ps(_k02, _r02, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k02, _r03, _sum01);
-
-                __m512 _r10 = _mm512_load_ps(r1);
-                __m512 _r11 = _mm512_load_ps(r1 + 16);
-                __m512 _r12 = _mm512_load_ps(r1 + 32);
-                __m512 _r13 = _mm512_load_ps(r1 + 48);
-
-                _sum00 = _mm512_fmadd_ps(_k10, _r10, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k10, _r11, _sum01);
-                _sum10 = _mm512_fmadd_ps(_k00, _r10, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k00, _r11, _sum11);
-
-                _sum00 = _mm512_fmadd_ps(_k11, _r11, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k11, _r12, _sum01);
-                _sum10 = _mm512_fmadd_ps(_k01, _r11, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k01, _r12, _sum11);
-
-                _sum00 = _mm512_fmadd_ps(_k12, _r12, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k12, _r13, _sum01);
-                _sum10 = _mm512_fmadd_ps(_k02, _r12, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k02, _r13, _sum11);
-
-                __m512 _r20 = _mm512_load_ps(r2);
-                __m512 _r21 = _mm512_load_ps(r2 + 16);
-                __m512 _r22 = _mm512_load_ps(r2 + 32);
-                __m512 _r23 = _mm512_load_ps(r2 + 48);
-
-                _sum00 = _mm512_fmadd_ps(_k20, _r20, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k20, _r21, _sum01);
-                _sum10 = _mm512_fmadd_ps(_k10, _r20, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k10, _r21, _sum11);
-
-                _sum00 = _mm512_fmadd_ps(_k21, _r21, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k21, _r22, _sum01);
-                _sum10 = _mm512_fmadd_ps(_k11, _r21, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k11, _r22, _sum11);
-
-                _sum00 = _mm512_fmadd_ps(_k22, _r22, _sum00);
-                _sum01 = _mm512_fmadd_ps(_k22, _r23, _sum01);
-                _sum10 = _mm512_fmadd_ps(_k12, _r22, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k12, _r23, _sum11);
-
-                __m512 _r30 = _mm512_load_ps(r3);
-                __m512 _r31 = _mm512_load_ps(r3 + 16);
-                __m512 _r32 = _mm512_load_ps(r3 + 32);
-                __m512 _r33 = _mm512_load_ps(r3 + 48);
-
-                _sum10 = _mm512_fmadd_ps(_k20, _r30, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k20, _r31, _sum11);
-                _sum10 = _mm512_fmadd_ps(_k21, _r31, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k21, _r32, _sum11);
-                _sum10 = _mm512_fmadd_ps(_k22, _r32, _sum10);
-                _sum11 = _mm512_fmadd_ps(_k22, _r33, _sum11);
-
-                _mm512_store_ps(outptr0, _sum00);
-                _mm512_store_ps(outptr0 + 16, _sum01);
-                _mm512_store_ps(outptr1, _sum10);
-                _mm512_store_ps(outptr1 + 16, _sum11);
-
-                r0 += 32;
-                r1 += 32;
-                r2 += 32;
-                r3 += 32;
-                outptr0 += 32;
-                outptr1 += 32;
-            }
-            for (; j < outw; j++)
-            {
-                __m512 _sum0 = _bias0;
-                __m512 _sum1 = _bias0;
-
-                __m512 _r00 = _mm512_load_ps(r0);
-                __m512 _r01 = _mm512_load_ps(r0 + 16);
-                __m512 _r02 = _mm512_load_ps(r0 + 32);
-
-                _sum0 = _mm512_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k02, _r02, _sum0);
-
-                __m512 _r10 = _mm512_load_ps(r1);
-                __m512 _r11 = _mm512_load_ps(r1 + 16);
-                __m512 _r12 = _mm512_load_ps(r1 + 32);
-
-                _sum0 = _mm512_fmadd_ps(_k10, _r10, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k00, _r10, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k11, _r11, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k01, _r11, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k12, _r12, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k02, _r12, _sum1);
-
-                __m512 _r20 = _mm512_load_ps(r2);
-                __m512 _r21 = _mm512_load_ps(r2 + 16);
-                __m512 _r22 = _mm512_load_ps(r2 + 32);
-
-                _sum0 = _mm512_fmadd_ps(_k20, _r20, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k10, _r20, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k21, _r21, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k11, _r21, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k22, _r22, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k12, _r22, _sum1);
-
-                __m512 _r30 = _mm512_load_ps(r3);
-                __m512 _r31 = _mm512_load_ps(r3 + 16);
-                __m512 _r32 = _mm512_load_ps(r3 + 32);
-
-                _sum1 = _mm512_fmadd_ps(_k20, _r30, _sum1);
-                _sum1 = _mm512_fmadd_ps(_k21, _r31, _sum1);
-                _sum1 = _mm512_fmadd_ps(_k22, _r32, _sum1);
-
-                _mm512_store_ps(outptr0, _sum0);
-                _mm512_store_ps(outptr1, _sum1);
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                r3 += 16;
-                outptr0 += 16;
-                outptr1 += 16;
-            }
-
-            r0 += 2 * 16 + w * 16;
-            r1 += 2 * 16 + w * 16;
-            r2 += 2 * 16 + w * 16;
-            r3 += 2 * 16 + w * 16;
-
-            outptr0 += outw * 16;
-            outptr1 += outw * 16;
-        }
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                __m512 _sum0 = _bias0;
-                __m512 _sum1 = _bias0;
-                __m512 _sum2 = _bias0;
-                __m512 _sum3 = _bias0;
-
-                __m512 _r00 = _mm512_load_ps(r0);
-                __m512 _r01 = _mm512_load_ps(r0 + 16);
-                __m512 _r02 = _mm512_load_ps(r0 + 32);
-                __m512 _r03 = _mm512_load_ps(r0 + 48);
-                __m512 _r04 = _mm512_load_ps(r0 + 64);
-                __m512 _r05 = _mm512_load_ps(r0 + 80);
-
-                _sum0 = _mm512_fmadd_ps(_k00, _r00, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k00, _r01, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k00, _r02, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k00, _r03, _sum3);
-                _sum0 = _mm512_fmadd_ps(_k01, _r01, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k01, _r02, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k01, _r03, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k01, _r04, _sum3);
-                _sum0 = _mm512_fmadd_ps(_k02, _r02, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k02, _r03, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k02, _r04, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k02, _r05, _sum3);
-
-                __m512 _r10 = _mm512_load_ps(r1);
-                __m512 _r11 = _mm512_load_ps(r1 + 16);
-                __m512 _r12 = _mm512_load_ps(r1 + 32);
-                __m512 _r13 = _mm512_load_ps(r1 + 48);
-                __m512 _r14 = _mm512_load_ps(r1 + 64);
-                __m512 _r15 = _mm512_load_ps(r1 + 80);
-
-                _sum0 = _mm512_fmadd_ps(_k10, _r10, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k10, _r11, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k10, _r12, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k10, _r13, _sum3);
-                _sum0 = _mm512_fmadd_ps(_k11, _r11, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k11, _r12, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k11, _r13, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k11, _r14, _sum3);
-                _sum0 = _mm512_fmadd_ps(_k12, _r12, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k12, _r13, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k12, _r14, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k12, _r15, _sum3);
-
-                __m512 _r20 = _mm512_load_ps(r2);
-                __m512 _r21 = _mm512_load_ps(r2 + 16);
-                __m512 _r22 = _mm512_load_ps(r2 + 32);
-                __m512 _r23 = _mm512_load_ps(r2 + 48);
-                __m512 _r24 = _mm512_load_ps(r2 + 64);
-                __m512 _r25 = _mm512_load_ps(r2 + 80);
-
-                _sum0 = _mm512_fmadd_ps(_k20, _r20, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k20, _r21, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k20, _r22, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k20, _r23, _sum3);
-                _sum0 = _mm512_fmadd_ps(_k21, _r21, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k21, _r22, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k21, _r23, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k21, _r24, _sum3);
-                _sum0 = _mm512_fmadd_ps(_k22, _r22, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k22, _r23, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k22, _r24, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k22, _r25, _sum3);
-
-                _mm512_store_ps(outptr0, _sum0);
-                _mm512_store_ps(outptr0 + 16, _sum1);
-                _mm512_store_ps(outptr0 + 32, _sum2);
-                _mm512_store_ps(outptr0 + 48, _sum3);
-
-                r0 += 64;
-                r1 += 64;
-                r2 += 64;
-                outptr0 += 64;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                __m512 _sum0 = _bias0;
-                __m512 _sum1 = _bias0;
-
-                __m512 _r00 = _mm512_load_ps(r0);
-                __m512 _r01 = _mm512_load_ps(r0 + 16);
-                __m512 _r02 = _mm512_load_ps(r0 + 32);
-                __m512 _r03 = _mm512_load_ps(r0 + 48);
-
-                _sum0 = _mm512_fmadd_ps(_k00, _r00, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k00, _r01, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k01, _r01, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k01, _r02, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k02, _r02, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k02, _r03, _sum1);
-
-                __m512 _r10 = _mm512_load_ps(r1);
-                __m512 _r11 = _mm512_load_ps(r1 + 16);
-                __m512 _r12 = _mm512_load_ps(r1 + 32);
-                __m512 _r13 = _mm512_load_ps(r1 + 48);
-
-                _sum0 = _mm512_fmadd_ps(_k10, _r10, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k10, _r11, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k11, _r11, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k11, _r12, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k12, _r12, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k12, _r13, _sum1);
-
-                __m512 _r20 = _mm512_load_ps(r2);
-                __m512 _r21 = _mm512_load_ps(r2 + 16);
-                __m512 _r22 = _mm512_load_ps(r2 + 32);
-                __m512 _r23 = _mm512_load_ps(r2 + 48);
-
-                _sum0 = _mm512_fmadd_ps(_k20, _r20, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k20, _r21, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k21, _r21, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k21, _r22, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k22, _r22, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k22, _r23, _sum1);
-
-                _mm512_store_ps(outptr0, _sum0);
-                _mm512_store_ps(outptr0 + 16, _sum1);
-
-                r0 += 32;
-                r1 += 32;
-                r2 += 32;
-                outptr0 += 32;
-            }
-            for (; j < outw; j++)
-            {
-                __m512 _sum0 = _bias0;
-
-                __m512 _r00 = _mm512_load_ps(r0);
-                __m512 _r01 = _mm512_load_ps(r0 + 16);
-                __m512 _r02 = _mm512_load_ps(r0 + 32);
-
-                _sum0 = _mm512_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k02, _r02, _sum0);
-
-                __m512 _r10 = _mm512_load_ps(r1);
-                __m512 _r11 = _mm512_load_ps(r1 + 16);
-                __m512 _r12 = _mm512_load_ps(r1 + 32);
-
-                _sum0 = _mm512_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k12, _r12, _sum0);
-
-                __m512 _r20 = _mm512_load_ps(r2);
-                __m512 _r21 = _mm512_load_ps(r2 + 16);
-                __m512 _r22 = _mm512_load_ps(r2 + 32);
-
-                _sum0 = _mm512_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k22, _r22, _sum0);
-
-                _mm512_store_ps(outptr0, _sum0);
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                outptr0 += 16;
-            }
-
-            r0 += 2 * 16;
-            r1 += 2 * 16;
-            r2 += 2 * 16;
-        }
-    }
-}
-
-static void convdw3x3s2_pack16_avx512(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 16;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m512 _bias0 = bias ? _mm512_loadu_ps((const float*)bias + g * 16) : _mm512_setzero_ps();
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-
-        __m512 _k00 = _mm512_load_ps(k0);
-        __m512 _k01 = _mm512_load_ps(k0 + 16);
-        __m512 _k02 = _mm512_load_ps(k0 + 32);
-        __m512 _k10 = _mm512_load_ps(k0 + 48);
-        __m512 _k11 = _mm512_load_ps(k0 + 64);
-        __m512 _k12 = _mm512_load_ps(k0 + 80);
-        __m512 _k20 = _mm512_load_ps(k0 + 96);
-        __m512 _k21 = _mm512_load_ps(k0 + 112);
-        __m512 _k22 = _mm512_load_ps(k0 + 128);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                __m512 _sum0 = _bias0;
-                __m512 _sum1 = _bias0;
-                __m512 _sum2 = _bias0;
-                __m512 _sum3 = _bias0;
-
-                __m512 _r00 = _mm512_load_ps(r0);
-                __m512 _r01 = _mm512_load_ps(r0 + 16);
-                __m512 _r02 = _mm512_load_ps(r0 + 32);
-                __m512 _r03 = _mm512_load_ps(r0 + 48);
-                __m512 _r04 = _mm512_load_ps(r0 + 64);
-                __m512 _r05 = _mm512_load_ps(r0 + 80);
-                __m512 _r06 = _mm512_load_ps(r0 + 96);
-                __m512 _r07 = _mm512_load_ps(r0 + 112);
-                __m512 _r08 = _mm512_load_ps(r0 + 128);
-
-                _sum0 = _mm512_fmadd_ps(_k00, _r00, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k00, _r02, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k00, _r04, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k00, _r06, _sum3);
-                _sum0 = _mm512_fmadd_ps(_k01, _r01, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k01, _r03, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k01, _r05, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k01, _r07, _sum3);
-                _sum0 = _mm512_fmadd_ps(_k02, _r02, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k02, _r04, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k02, _r06, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k02, _r08, _sum3);
-
-                __m512 _r10 = _mm512_load_ps(r1);
-                __m512 _r11 = _mm512_load_ps(r1 + 16);
-                __m512 _r12 = _mm512_load_ps(r1 + 32);
-                __m512 _r13 = _mm512_load_ps(r1 + 48);
-                __m512 _r14 = _mm512_load_ps(r1 + 64);
-                __m512 _r15 = _mm512_load_ps(r1 + 80);
-                __m512 _r16 = _mm512_load_ps(r1 + 96);
-                __m512 _r17 = _mm512_load_ps(r1 + 112);
-                __m512 _r18 = _mm512_load_ps(r1 + 128);
-
-                _sum0 = _mm512_fmadd_ps(_k10, _r10, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k10, _r12, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k10, _r14, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k10, _r16, _sum3);
-                _sum0 = _mm512_fmadd_ps(_k11, _r11, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k11, _r13, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k11, _r15, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k11, _r17, _sum3);
-                _sum0 = _mm512_fmadd_ps(_k12, _r12, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k12, _r14, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k12, _r16, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k12, _r18, _sum3);
-
-                __m512 _r20 = _mm512_load_ps(r2);
-                __m512 _r21 = _mm512_load_ps(r2 + 16);
-                __m512 _r22 = _mm512_load_ps(r2 + 32);
-                __m512 _r23 = _mm512_load_ps(r2 + 48);
-                __m512 _r24 = _mm512_load_ps(r2 + 64);
-                __m512 _r25 = _mm512_load_ps(r2 + 80);
-                __m512 _r26 = _mm512_load_ps(r2 + 96);
-                __m512 _r27 = _mm512_load_ps(r2 + 112);
-                __m512 _r28 = _mm512_load_ps(r2 + 128);
-
-                _sum0 = _mm512_fmadd_ps(_k20, _r20, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k20, _r22, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k20, _r24, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k20, _r26, _sum3);
-                _sum0 = _mm512_fmadd_ps(_k21, _r21, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k21, _r23, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k21, _r25, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k21, _r27, _sum3);
-                _sum0 = _mm512_fmadd_ps(_k22, _r22, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k22, _r24, _sum1);
-                _sum2 = _mm512_fmadd_ps(_k22, _r26, _sum2);
-                _sum3 = _mm512_fmadd_ps(_k22, _r28, _sum3);
-
-                _mm512_store_ps(outptr0, _sum0);
-                _mm512_store_ps(outptr0 + 16, _sum1);
-                _mm512_store_ps(outptr0 + 32, _sum2);
-                _mm512_store_ps(outptr0 + 48, _sum3);
-
-                r0 += 2 * 64;
-                r1 += 2 * 64;
-                r2 += 2 * 64;
-                outptr0 += 64;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                __m512 _sum0 = _bias0;
-                __m512 _sum1 = _bias0;
-
-                __m512 _r00 = _mm512_load_ps(r0);
-                __m512 _r01 = _mm512_load_ps(r0 + 16);
-                __m512 _r02 = _mm512_load_ps(r0 + 32);
-                __m512 _r03 = _mm512_load_ps(r0 + 48);
-                __m512 _r04 = _mm512_load_ps(r0 + 64);
-
-                _sum0 = _mm512_fmadd_ps(_k00, _r00, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k00, _r02, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k01, _r01, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k01, _r03, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k02, _r02, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k02, _r04, _sum1);
-
-                __m512 _r10 = _mm512_load_ps(r1);
-                __m512 _r11 = _mm512_load_ps(r1 + 16);
-                __m512 _r12 = _mm512_load_ps(r1 + 32);
-                __m512 _r13 = _mm512_load_ps(r1 + 48);
-                __m512 _r14 = _mm512_load_ps(r1 + 64);
-
-                _sum0 = _mm512_fmadd_ps(_k10, _r10, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k10, _r12, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k11, _r11, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k11, _r13, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k12, _r12, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k12, _r14, _sum1);
-
-                __m512 _r20 = _mm512_load_ps(r2);
-                __m512 _r21 = _mm512_load_ps(r2 + 16);
-                __m512 _r22 = _mm512_load_ps(r2 + 32);
-                __m512 _r23 = _mm512_load_ps(r2 + 48);
-                __m512 _r24 = _mm512_load_ps(r2 + 64);
-
-                _sum0 = _mm512_fmadd_ps(_k20, _r20, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k20, _r22, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k21, _r21, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k21, _r23, _sum1);
-                _sum0 = _mm512_fmadd_ps(_k22, _r22, _sum0);
-                _sum1 = _mm512_fmadd_ps(_k22, _r24, _sum1);
-
-                _mm512_store_ps(outptr0, _sum0);
-                _mm512_store_ps(outptr0 + 16, _sum1);
-
-                r0 += 2 * 32;
-                r1 += 2 * 32;
-                r2 += 2 * 32;
-                outptr0 += 32;
-            }
-            for (; j < outw; j++)
-            {
-                __m512 _sum0 = _bias0;
-
-                __m512 _r00 = _mm512_load_ps(r0);
-                __m512 _r01 = _mm512_load_ps(r0 + 16);
-                __m512 _r02 = _mm512_load_ps(r0 + 32);
-
-                _sum0 = _mm512_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k02, _r02, _sum0);
-
-                __m512 _r10 = _mm512_load_ps(r1);
-                __m512 _r11 = _mm512_load_ps(r1 + 16);
-                __m512 _r12 = _mm512_load_ps(r1 + 32);
-
-                _sum0 = _mm512_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k12, _r12, _sum0);
-
-                __m512 _r20 = _mm512_load_ps(r2);
-                __m512 _r21 = _mm512_load_ps(r2 + 16);
-                __m512 _r22 = _mm512_load_ps(r2 + 32);
-
-                _sum0 = _mm512_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k22, _r22, _sum0);
-
-                _mm512_store_ps(outptr0, _sum0);
-
-                r0 += 2 * 16;
-                r1 += 2 * 16;
-                r2 += 2 * 16;
-                outptr0 += 16;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_pack4.h
deleted file mode 100644
index 97b8c0d9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_pack4.h
+++ /dev/null
@@ -1,581 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_pack4_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m128 _bias0 = bias ? _mm_loadu_ps((const float*)bias + g * 4) : _mm_set1_ps(0.f);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-
-        __m128 _k00 = _mm_load_ps(k0);
-        __m128 _k01 = _mm_load_ps(k0 + 4);
-        __m128 _k02 = _mm_load_ps(k0 + 8);
-        __m128 _k10 = _mm_load_ps(k0 + 12);
-        __m128 _k11 = _mm_load_ps(k0 + 16);
-        __m128 _k12 = _mm_load_ps(k0 + 20);
-        __m128 _k20 = _mm_load_ps(k0 + 24);
-        __m128 _k21 = _mm_load_ps(k0 + 28);
-        __m128 _k22 = _mm_load_ps(k0 + 32);
-
-        int i = 0;
-
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 7 < outw; j += 8)
-            {
-                __m128 _sum0 = _bias0;
-
-                __m128 _r00 = _mm_load_ps(r0);
-                __m128 _r01 = _mm_load_ps(r0 + 4);
-                __m128 _r02 = _mm_load_ps(r0 + 8);
-                __m128 _r10 = _mm_load_ps(r1);
-                __m128 _r11 = _mm_load_ps(r1 + 4);
-                __m128 _r12 = _mm_load_ps(r1 + 8);
-                __m128 _r20 = _mm_load_ps(r2);
-                __m128 _r21 = _mm_load_ps(r2 + 4);
-                __m128 _r22 = _mm_load_ps(r2 + 8);
-
-                _sum0 = _mm_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k22, _r22, _sum0);
-
-                __m128 _sum1 = _bias0;
-                __m128 _r03 = _mm_load_ps(r0 + 12);
-                __m128 _r13 = _mm_load_ps(r1 + 12);
-                __m128 _r23 = _mm_load_ps(r2 + 12);
-                _mm_store_ps(outptr0, _sum0);
-
-                _sum1 = _mm_comp_fmadd_ps(_k00, _r01, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k01, _r02, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k02, _r03, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k10, _r11, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k11, _r12, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k12, _r13, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k20, _r21, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k21, _r22, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k22, _r23, _sum1);
-
-                __m128 _sum2 = _bias0;
-                __m128 _r04 = _mm_load_ps(r0 + 16);
-                __m128 _r14 = _mm_load_ps(r1 + 16);
-                __m128 _r24 = _mm_load_ps(r2 + 16);
-                _mm_store_ps(outptr0 + 4, _sum1);
-
-                _sum2 = _mm_comp_fmadd_ps(_k00, _r02, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k01, _r03, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k02, _r04, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k10, _r12, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k11, _r13, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k12, _r14, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k20, _r22, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k21, _r23, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k22, _r24, _sum2);
-
-                __m128 _sum3 = _bias0;
-                __m128 _r05 = _mm_load_ps(r0 + 20);
-                __m128 _r15 = _mm_load_ps(r1 + 20);
-                __m128 _r25 = _mm_load_ps(r2 + 20);
-                _mm_store_ps(outptr0 + 8, _sum2);
-
-                _sum3 = _mm_comp_fmadd_ps(_k00, _r03, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k01, _r04, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k02, _r05, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k10, _r13, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k11, _r14, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k12, _r15, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k20, _r23, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k21, _r24, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k22, _r25, _sum3);
-
-                __m128 _sum4 = _bias0;
-                __m128 _r06 = _mm_load_ps(r0 + 24);
-                __m128 _r16 = _mm_load_ps(r1 + 24);
-                __m128 _r26 = _mm_load_ps(r2 + 24);
-                _mm_store_ps(outptr0 + 12, _sum3);
-
-                _sum4 = _mm_comp_fmadd_ps(_k00, _r04, _sum4);
-                _sum4 = _mm_comp_fmadd_ps(_k01, _r05, _sum4);
-                _sum4 = _mm_comp_fmadd_ps(_k02, _r06, _sum4);
-                _sum4 = _mm_comp_fmadd_ps(_k10, _r14, _sum4);
-                _sum4 = _mm_comp_fmadd_ps(_k11, _r15, _sum4);
-                _sum4 = _mm_comp_fmadd_ps(_k12, _r16, _sum4);
-                _sum4 = _mm_comp_fmadd_ps(_k20, _r24, _sum4);
-                _sum4 = _mm_comp_fmadd_ps(_k21, _r25, _sum4);
-                _sum4 = _mm_comp_fmadd_ps(_k22, _r26, _sum4);
-
-                __m128 _sum5 = _bias0;
-                __m128 _r07 = _mm_load_ps(r0 + 28);
-                __m128 _r17 = _mm_load_ps(r1 + 28);
-                __m128 _r27 = _mm_load_ps(r2 + 28);
-                _mm_store_ps(outptr0 + 16, _sum4);
-
-                _sum5 = _mm_comp_fmadd_ps(_k00, _r05, _sum5);
-                _sum5 = _mm_comp_fmadd_ps(_k01, _r06, _sum5);
-                _sum5 = _mm_comp_fmadd_ps(_k02, _r07, _sum5);
-                _sum5 = _mm_comp_fmadd_ps(_k10, _r15, _sum5);
-                _sum5 = _mm_comp_fmadd_ps(_k11, _r16, _sum5);
-                _sum5 = _mm_comp_fmadd_ps(_k12, _r17, _sum5);
-                _sum5 = _mm_comp_fmadd_ps(_k20, _r25, _sum5);
-                _sum5 = _mm_comp_fmadd_ps(_k21, _r26, _sum5);
-                _sum5 = _mm_comp_fmadd_ps(_k22, _r27, _sum5);
-
-                __m128 _sum6 = _bias0;
-                __m128 _r08 = _mm_load_ps(r0 + 32);
-                __m128 _r18 = _mm_load_ps(r1 + 32);
-                __m128 _r28 = _mm_load_ps(r2 + 32);
-                _mm_store_ps(outptr0 + 20, _sum5);
-
-                _sum6 = _mm_comp_fmadd_ps(_k00, _r06, _sum6);
-                _sum6 = _mm_comp_fmadd_ps(_k01, _r07, _sum6);
-                _sum6 = _mm_comp_fmadd_ps(_k02, _r08, _sum6);
-                _sum6 = _mm_comp_fmadd_ps(_k10, _r16, _sum6);
-                _sum6 = _mm_comp_fmadd_ps(_k11, _r17, _sum6);
-                _sum6 = _mm_comp_fmadd_ps(_k12, _r18, _sum6);
-                _sum6 = _mm_comp_fmadd_ps(_k20, _r26, _sum6);
-                _sum6 = _mm_comp_fmadd_ps(_k21, _r27, _sum6);
-                _sum6 = _mm_comp_fmadd_ps(_k22, _r28, _sum6);
-
-                __m128 _sum7 = _bias0;
-                __m128 _r09 = _mm_load_ps(r0 + 36);
-                __m128 _r19 = _mm_load_ps(r1 + 36);
-                __m128 _r29 = _mm_load_ps(r2 + 36);
-                _mm_store_ps(outptr0 + 24, _sum6);
-
-                _sum7 = _mm_comp_fmadd_ps(_k00, _r07, _sum7);
-                _sum7 = _mm_comp_fmadd_ps(_k01, _r08, _sum7);
-                _sum7 = _mm_comp_fmadd_ps(_k02, _r09, _sum7);
-                _sum7 = _mm_comp_fmadd_ps(_k10, _r17, _sum7);
-                _sum7 = _mm_comp_fmadd_ps(_k11, _r18, _sum7);
-                _sum7 = _mm_comp_fmadd_ps(_k12, _r19, _sum7);
-                _sum7 = _mm_comp_fmadd_ps(_k20, _r27, _sum7);
-                _sum7 = _mm_comp_fmadd_ps(_k21, _r28, _sum7);
-                _sum7 = _mm_comp_fmadd_ps(_k22, _r29, _sum7);
-                _mm_store_ps(outptr0 + 28, _sum7);
-
-                r0 += 32;
-                r1 += 32;
-                r2 += 32;
-                outptr0 += 32;
-            }
-            for (; j + 3 < outw; j += 4)
-            {
-                __m128 _sum0 = _bias0;
-
-                __m128 _r00 = _mm_load_ps(r0);
-                __m128 _r01 = _mm_load_ps(r0 + 4);
-                __m128 _r02 = _mm_load_ps(r0 + 8);
-                __m128 _r10 = _mm_load_ps(r1);
-                __m128 _r11 = _mm_load_ps(r1 + 4);
-                __m128 _r12 = _mm_load_ps(r1 + 8);
-                __m128 _r20 = _mm_load_ps(r2);
-                __m128 _r21 = _mm_load_ps(r2 + 4);
-                __m128 _r22 = _mm_load_ps(r2 + 8);
-
-                _sum0 = _mm_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k22, _r22, _sum0);
-
-                __m128 _sum1 = _bias0;
-                __m128 _r03 = _mm_load_ps(r0 + 12);
-                __m128 _r13 = _mm_load_ps(r1 + 12);
-                __m128 _r23 = _mm_load_ps(r2 + 12);
-                _mm_store_ps(outptr0, _sum0);
-
-                _sum1 = _mm_comp_fmadd_ps(_k00, _r01, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k01, _r02, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k02, _r03, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k10, _r11, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k11, _r12, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k12, _r13, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k20, _r21, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k21, _r22, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k22, _r23, _sum1);
-
-                __m128 _sum2 = _bias0;
-                __m128 _r04 = _mm_load_ps(r0 + 16);
-                __m128 _r14 = _mm_load_ps(r1 + 16);
-                __m128 _r24 = _mm_load_ps(r2 + 16);
-                _mm_store_ps(outptr0 + 4, _sum1);
-
-                _sum2 = _mm_comp_fmadd_ps(_k00, _r02, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k01, _r03, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k02, _r04, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k10, _r12, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k11, _r13, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k12, _r14, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k20, _r22, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k21, _r23, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k22, _r24, _sum2);
-
-                __m128 _sum3 = _bias0;
-                __m128 _r05 = _mm_load_ps(r0 + 20);
-                __m128 _r15 = _mm_load_ps(r1 + 20);
-                __m128 _r25 = _mm_load_ps(r2 + 20);
-                _mm_store_ps(outptr0 + 8, _sum2);
-
-                _sum3 = _mm_comp_fmadd_ps(_k00, _r03, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k01, _r04, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k02, _r05, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k10, _r13, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k11, _r14, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k12, _r15, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k20, _r23, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k21, _r24, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k22, _r25, _sum3);
-
-                _mm_store_ps(outptr0 + 12, _sum3);
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                outptr0 += 16;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                __m128 _sum0 = _bias0;
-
-                __m128 _r00 = _mm_load_ps(r0);
-                __m128 _r01 = _mm_load_ps(r0 + 4);
-                __m128 _r02 = _mm_load_ps(r0 + 8);
-                __m128 _r10 = _mm_load_ps(r1);
-                __m128 _r11 = _mm_load_ps(r1 + 4);
-                __m128 _r12 = _mm_load_ps(r1 + 8);
-                __m128 _r20 = _mm_load_ps(r2);
-                __m128 _r21 = _mm_load_ps(r2 + 4);
-                __m128 _r22 = _mm_load_ps(r2 + 8);
-
-                _sum0 = _mm_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k22, _r22, _sum0);
-
-                __m128 _sum1 = _bias0;
-                __m128 _r03 = _mm_load_ps(r0 + 12);
-                __m128 _r13 = _mm_load_ps(r1 + 12);
-                __m128 _r23 = _mm_load_ps(r2 + 12);
-                _mm_store_ps(outptr0, _sum0);
-
-                _sum1 = _mm_comp_fmadd_ps(_k00, _r01, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k01, _r02, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k02, _r03, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k10, _r11, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k11, _r12, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k12, _r13, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k20, _r21, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k21, _r22, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k22, _r23, _sum1);
-
-                _mm_store_ps(outptr0 + 4, _sum1);
-
-                r0 += 8;
-                r1 += 8;
-                r2 += 8;
-                outptr0 += 8;
-            }
-            for (; j < outw; j++)
-            {
-                __m128 _sum0 = _bias0;
-
-                __m128 _r00 = _mm_load_ps(r0);
-                __m128 _r01 = _mm_load_ps(r0 + 4);
-                __m128 _r02 = _mm_load_ps(r0 + 8);
-                __m128 _r10 = _mm_load_ps(r1);
-                __m128 _r11 = _mm_load_ps(r1 + 4);
-                __m128 _r12 = _mm_load_ps(r1 + 8);
-                __m128 _r20 = _mm_load_ps(r2);
-                __m128 _r21 = _mm_load_ps(r2 + 4);
-                __m128 _r22 = _mm_load_ps(r2 + 8);
-
-                _sum0 = _mm_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k22, _r22, _sum0);
-
-                _mm_store_ps(outptr0, _sum0);
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                outptr0 += 4;
-            }
-
-            r0 += 2 * 4;
-            r1 += 2 * 4;
-            r2 += 2 * 4;
-        }
-    }
-}
-
-static void convdw3x3s2_pack4_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m128 _bias0 = bias ? _mm_loadu_ps((const float*)bias + g * 4) : _mm_set1_ps(0.f);
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-
-        __m128 _k00 = _mm_load_ps(k0);
-        __m128 _k01 = _mm_load_ps(k0 + 4);
-        __m128 _k02 = _mm_load_ps(k0 + 8);
-        __m128 _k10 = _mm_load_ps(k0 + 12);
-        __m128 _k11 = _mm_load_ps(k0 + 16);
-        __m128 _k12 = _mm_load_ps(k0 + 20);
-        __m128 _k20 = _mm_load_ps(k0 + 24);
-        __m128 _k21 = _mm_load_ps(k0 + 28);
-        __m128 _k22 = _mm_load_ps(k0 + 32);
-
-        int i = 0;
-
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                __m128 _sum0 = _bias0;
-
-                __m128 _r00 = _mm_load_ps(r0);
-                __m128 _r01 = _mm_load_ps(r0 + 4);
-                __m128 _r02 = _mm_load_ps(r0 + 8);
-                __m128 _r10 = _mm_load_ps(r1);
-                __m128 _r11 = _mm_load_ps(r1 + 4);
-                __m128 _r12 = _mm_load_ps(r1 + 8);
-                __m128 _r20 = _mm_load_ps(r2);
-                __m128 _r21 = _mm_load_ps(r2 + 4);
-                __m128 _r22 = _mm_load_ps(r2 + 8);
-
-                _sum0 = _mm_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k22, _r22, _sum0);
-
-                __m128 _sum1 = _bias0;
-                __m128 _r03 = _mm_load_ps(r0 + 12);
-                __m128 _r13 = _mm_load_ps(r1 + 12);
-                __m128 _r23 = _mm_load_ps(r2 + 12);
-                __m128 _r04 = _mm_load_ps(r0 + 16);
-                __m128 _r14 = _mm_load_ps(r1 + 16);
-                __m128 _r24 = _mm_load_ps(r2 + 16);
-                _mm_store_ps(outptr0, _sum0);
-
-                _sum1 = _mm_comp_fmadd_ps(_k00, _r02, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k01, _r03, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k02, _r04, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k10, _r12, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k11, _r13, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k12, _r14, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k20, _r22, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k21, _r23, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k22, _r24, _sum1);
-
-                __m128 _sum2 = _bias0;
-                __m128 _r05 = _mm_load_ps(r0 + 20);
-                __m128 _r15 = _mm_load_ps(r1 + 20);
-                __m128 _r25 = _mm_load_ps(r2 + 20);
-                __m128 _r06 = _mm_load_ps(r0 + 24);
-                __m128 _r16 = _mm_load_ps(r1 + 24);
-                __m128 _r26 = _mm_load_ps(r2 + 24);
-                _mm_store_ps(outptr0 + 4, _sum1);
-
-                _sum2 = _mm_comp_fmadd_ps(_k00, _r04, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k01, _r05, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k02, _r06, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k10, _r14, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k11, _r15, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k12, _r16, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k20, _r24, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k21, _r25, _sum2);
-                _sum2 = _mm_comp_fmadd_ps(_k22, _r26, _sum2);
-
-                __m128 _sum3 = _bias0;
-                __m128 _r07 = _mm_load_ps(r0 + 28);
-                __m128 _r17 = _mm_load_ps(r1 + 28);
-                __m128 _r27 = _mm_load_ps(r2 + 28);
-                __m128 _r08 = _mm_load_ps(r0 + 32);
-                __m128 _r18 = _mm_load_ps(r1 + 32);
-                __m128 _r28 = _mm_load_ps(r2 + 32);
-                _mm_store_ps(outptr0 + 8, _sum2);
-
-                _sum3 = _mm_comp_fmadd_ps(_k00, _r06, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k01, _r07, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k02, _r08, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k10, _r16, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k11, _r17, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k12, _r18, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k20, _r26, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k21, _r27, _sum3);
-                _sum3 = _mm_comp_fmadd_ps(_k22, _r28, _sum3);
-
-                _mm_store_ps(outptr0 + 12, _sum3);
-
-                r0 += 2 * 16;
-                r1 += 2 * 16;
-                r2 += 2 * 16;
-                outptr0 += 16;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                __m128 _sum0 = _bias0;
-
-                __m128 _r00 = _mm_load_ps(r0);
-                __m128 _r01 = _mm_load_ps(r0 + 4);
-                __m128 _r02 = _mm_load_ps(r0 + 8);
-                __m128 _r10 = _mm_load_ps(r1);
-                __m128 _r11 = _mm_load_ps(r1 + 4);
-                __m128 _r12 = _mm_load_ps(r1 + 8);
-                __m128 _r20 = _mm_load_ps(r2);
-                __m128 _r21 = _mm_load_ps(r2 + 4);
-                __m128 _r22 = _mm_load_ps(r2 + 8);
-
-                _sum0 = _mm_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k22, _r22, _sum0);
-
-                __m128 _sum1 = _bias0;
-                __m128 _r03 = _mm_load_ps(r0 + 12);
-                __m128 _r13 = _mm_load_ps(r1 + 12);
-                __m128 _r23 = _mm_load_ps(r2 + 12);
-                __m128 _r04 = _mm_load_ps(r0 + 16);
-                __m128 _r14 = _mm_load_ps(r1 + 16);
-                __m128 _r24 = _mm_load_ps(r2 + 16);
-                _mm_store_ps(outptr0, _sum0);
-
-                _sum1 = _mm_comp_fmadd_ps(_k00, _r02, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k01, _r03, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k02, _r04, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k10, _r12, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k11, _r13, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k12, _r14, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k20, _r22, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k21, _r23, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k22, _r24, _sum1);
-
-                _mm_store_ps(outptr0 + 4, _sum1);
-
-                r0 += 2 * 8;
-                r1 += 2 * 8;
-                r2 += 2 * 8;
-                outptr0 += 8;
-            }
-            for (; j < outw; j++)
-            {
-                __m128 _sum0 = _bias0;
-
-                __m128 _r00 = _mm_load_ps(r0);
-                __m128 _r01 = _mm_load_ps(r0 + 4);
-                __m128 _r02 = _mm_load_ps(r0 + 8);
-                __m128 _r10 = _mm_load_ps(r1);
-                __m128 _r11 = _mm_load_ps(r1 + 4);
-                __m128 _r12 = _mm_load_ps(r1 + 8);
-                __m128 _r20 = _mm_load_ps(r2);
-                __m128 _r21 = _mm_load_ps(r2 + 4);
-                __m128 _r22 = _mm_load_ps(r2 + 8);
-
-                _sum0 = _mm_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k22, _r22, _sum0);
-                _mm_store_ps(outptr0, _sum0);
-                r0 += 2 * 4;
-                r1 += 2 * 4;
-                r2 += 2 * 4;
-                outptr0 += 4;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_pack8.h
deleted file mode 100644
index aa30143f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_3x3_pack8.h
+++ /dev/null
@@ -1,795 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw3x3s1_pack8_avx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m256 _bias0 = bias ? _mm256_loadu_ps((const float*)bias + g * 8) : _mm256_setzero_ps();
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-        float* outptr1 = out.row(1);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-
-        int i = 0;
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                __m256 _sum00 = _bias0;
-                __m256 _sum01 = _bias0;
-                __m256 _sum02 = _bias0;
-                __m256 _sum03 = _bias0;
-                __m256 _sum10 = _bias0;
-                __m256 _sum11 = _bias0;
-                __m256 _sum12 = _bias0;
-                __m256 _sum13 = _bias0;
-
-                __m256 _k00 = _mm256_load_ps(k0);
-                __m256 _k01 = _mm256_load_ps(k0 + 8);
-                __m256 _k02 = _mm256_load_ps(k0 + 16);
-
-                __m256 _r00 = _mm256_load_ps(r0);
-                __m256 _r01 = _mm256_load_ps(r0 + 8);
-                __m256 _r02 = _mm256_load_ps(r0 + 16);
-                __m256 _r03 = _mm256_load_ps(r0 + 24);
-                __m256 _r04 = _mm256_load_ps(r0 + 32);
-                __m256 _r05 = _mm256_load_ps(r0 + 40);
-
-                _sum00 = _mm256_comp_fmadd_ps(_k00, _r00, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k00, _r01, _sum01);
-                _sum02 = _mm256_comp_fmadd_ps(_k00, _r02, _sum02);
-                _sum03 = _mm256_comp_fmadd_ps(_k00, _r03, _sum03);
-                _sum00 = _mm256_comp_fmadd_ps(_k01, _r01, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k01, _r02, _sum01);
-                _sum02 = _mm256_comp_fmadd_ps(_k01, _r03, _sum02);
-                _sum03 = _mm256_comp_fmadd_ps(_k01, _r04, _sum03);
-                _sum00 = _mm256_comp_fmadd_ps(_k02, _r02, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k02, _r03, _sum01);
-                _sum02 = _mm256_comp_fmadd_ps(_k02, _r04, _sum02);
-                _sum03 = _mm256_comp_fmadd_ps(_k02, _r05, _sum03);
-
-                __m256 _k10 = _mm256_load_ps(k0 + 24);
-                __m256 _k11 = _mm256_load_ps(k0 + 32);
-                __m256 _k12 = _mm256_load_ps(k0 + 40);
-
-                __m256 _r10 = _mm256_load_ps(r1);
-                __m256 _r11 = _mm256_load_ps(r1 + 8);
-                __m256 _r12 = _mm256_load_ps(r1 + 16);
-                __m256 _r13 = _mm256_load_ps(r1 + 24);
-                __m256 _r14 = _mm256_load_ps(r1 + 32);
-                __m256 _r15 = _mm256_load_ps(r1 + 40);
-
-                _sum10 = _mm256_comp_fmadd_ps(_k00, _r10, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k00, _r11, _sum11);
-                _sum12 = _mm256_comp_fmadd_ps(_k00, _r12, _sum12);
-                _sum13 = _mm256_comp_fmadd_ps(_k00, _r13, _sum13);
-                _sum00 = _mm256_comp_fmadd_ps(_k10, _r10, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k10, _r11, _sum01);
-                _sum02 = _mm256_comp_fmadd_ps(_k10, _r12, _sum02);
-                _sum03 = _mm256_comp_fmadd_ps(_k10, _r13, _sum03);
-
-                _sum10 = _mm256_comp_fmadd_ps(_k01, _r11, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k01, _r12, _sum11);
-                _sum12 = _mm256_comp_fmadd_ps(_k01, _r13, _sum12);
-                _sum13 = _mm256_comp_fmadd_ps(_k01, _r14, _sum13);
-                _sum00 = _mm256_comp_fmadd_ps(_k11, _r11, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k11, _r12, _sum01);
-                _sum02 = _mm256_comp_fmadd_ps(_k11, _r13, _sum02);
-                _sum03 = _mm256_comp_fmadd_ps(_k11, _r14, _sum03);
-
-                _sum10 = _mm256_comp_fmadd_ps(_k02, _r12, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k02, _r13, _sum11);
-                _sum12 = _mm256_comp_fmadd_ps(_k02, _r14, _sum12);
-                _sum13 = _mm256_comp_fmadd_ps(_k02, _r15, _sum13);
-                _sum00 = _mm256_comp_fmadd_ps(_k12, _r12, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k12, _r13, _sum01);
-                _sum02 = _mm256_comp_fmadd_ps(_k12, _r14, _sum02);
-                _sum03 = _mm256_comp_fmadd_ps(_k12, _r15, _sum03);
-
-                __m256 _k20 = _mm256_load_ps(k0 + 48);
-                __m256 _k21 = _mm256_load_ps(k0 + 56);
-                __m256 _k22 = _mm256_load_ps(k0 + 64);
-
-                __m256 _r20 = _mm256_load_ps(r2);
-                __m256 _r21 = _mm256_load_ps(r2 + 8);
-                __m256 _r22 = _mm256_load_ps(r2 + 16);
-                __m256 _r23 = _mm256_load_ps(r2 + 24);
-                __m256 _r24 = _mm256_load_ps(r2 + 32);
-                __m256 _r25 = _mm256_load_ps(r2 + 40);
-
-                _sum10 = _mm256_comp_fmadd_ps(_k10, _r20, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k10, _r21, _sum11);
-                _sum12 = _mm256_comp_fmadd_ps(_k10, _r22, _sum12);
-                _sum13 = _mm256_comp_fmadd_ps(_k10, _r23, _sum13);
-                _sum00 = _mm256_comp_fmadd_ps(_k20, _r20, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k20, _r21, _sum01);
-                _sum02 = _mm256_comp_fmadd_ps(_k20, _r22, _sum02);
-                _sum03 = _mm256_comp_fmadd_ps(_k20, _r23, _sum03);
-
-                _sum10 = _mm256_comp_fmadd_ps(_k11, _r21, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k11, _r22, _sum11);
-                _sum12 = _mm256_comp_fmadd_ps(_k11, _r23, _sum12);
-                _sum13 = _mm256_comp_fmadd_ps(_k11, _r24, _sum13);
-                _sum00 = _mm256_comp_fmadd_ps(_k21, _r21, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k21, _r22, _sum01);
-                _sum02 = _mm256_comp_fmadd_ps(_k21, _r23, _sum02);
-                _sum03 = _mm256_comp_fmadd_ps(_k21, _r24, _sum03);
-
-                _sum10 = _mm256_comp_fmadd_ps(_k12, _r22, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k12, _r23, _sum11);
-                _sum12 = _mm256_comp_fmadd_ps(_k12, _r24, _sum12);
-                _sum13 = _mm256_comp_fmadd_ps(_k12, _r25, _sum13);
-                _sum00 = _mm256_comp_fmadd_ps(_k22, _r22, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k22, _r23, _sum01);
-                _sum02 = _mm256_comp_fmadd_ps(_k22, _r24, _sum02);
-                _sum03 = _mm256_comp_fmadd_ps(_k22, _r25, _sum03);
-
-                __m256 _r30 = _mm256_load_ps(r3);
-                __m256 _r31 = _mm256_load_ps(r3 + 8);
-                __m256 _r32 = _mm256_load_ps(r3 + 16);
-                __m256 _r33 = _mm256_load_ps(r3 + 24);
-                __m256 _r34 = _mm256_load_ps(r3 + 32);
-                __m256 _r35 = _mm256_load_ps(r3 + 40);
-
-                _sum10 = _mm256_comp_fmadd_ps(_k20, _r30, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k20, _r31, _sum11);
-                _sum12 = _mm256_comp_fmadd_ps(_k20, _r32, _sum12);
-                _sum13 = _mm256_comp_fmadd_ps(_k20, _r33, _sum13);
-                _sum10 = _mm256_comp_fmadd_ps(_k21, _r31, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k21, _r32, _sum11);
-                _sum12 = _mm256_comp_fmadd_ps(_k21, _r33, _sum12);
-                _sum13 = _mm256_comp_fmadd_ps(_k21, _r34, _sum13);
-                _sum10 = _mm256_comp_fmadd_ps(_k22, _r32, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k22, _r33, _sum11);
-                _sum12 = _mm256_comp_fmadd_ps(_k22, _r34, _sum12);
-                _sum13 = _mm256_comp_fmadd_ps(_k22, _r35, _sum13);
-
-                _mm256_store_ps(outptr0, _sum00);
-                _mm256_store_ps(outptr0 + 8, _sum01);
-                _mm256_store_ps(outptr0 + 16, _sum02);
-                _mm256_store_ps(outptr0 + 24, _sum03);
-                _mm256_store_ps(outptr1, _sum10);
-                _mm256_store_ps(outptr1 + 8, _sum11);
-                _mm256_store_ps(outptr1 + 16, _sum12);
-                _mm256_store_ps(outptr1 + 24, _sum13);
-
-                r0 += 32;
-                r1 += 32;
-                r2 += 32;
-                r3 += 32;
-                outptr0 += 32;
-                outptr1 += 32;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                __m256 _sum00 = _bias0;
-                __m256 _sum01 = _bias0;
-                __m256 _sum10 = _bias0;
-                __m256 _sum11 = _bias0;
-
-                __m256 _k00 = _mm256_load_ps(k0);
-                __m256 _k01 = _mm256_load_ps(k0 + 8);
-                __m256 _k02 = _mm256_load_ps(k0 + 16);
-
-                __m256 _r00 = _mm256_load_ps(r0);
-                __m256 _r01 = _mm256_load_ps(r0 + 8);
-                __m256 _r02 = _mm256_load_ps(r0 + 16);
-                __m256 _r03 = _mm256_load_ps(r0 + 24);
-
-                _sum00 = _mm256_comp_fmadd_ps(_k00, _r00, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k00, _r01, _sum01);
-                _sum00 = _mm256_comp_fmadd_ps(_k01, _r01, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k01, _r02, _sum01);
-                _sum00 = _mm256_comp_fmadd_ps(_k02, _r02, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k02, _r03, _sum01);
-
-                __m256 _k10 = _mm256_load_ps(k0 + 24);
-                __m256 _k11 = _mm256_load_ps(k0 + 32);
-                __m256 _k12 = _mm256_load_ps(k0 + 40);
-
-                __m256 _r10 = _mm256_load_ps(r1);
-                __m256 _r11 = _mm256_load_ps(r1 + 8);
-                __m256 _r12 = _mm256_load_ps(r1 + 16);
-                __m256 _r13 = _mm256_load_ps(r1 + 24);
-
-                _sum00 = _mm256_comp_fmadd_ps(_k10, _r10, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k10, _r11, _sum01);
-                _sum10 = _mm256_comp_fmadd_ps(_k00, _r10, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k00, _r11, _sum11);
-
-                _sum00 = _mm256_comp_fmadd_ps(_k11, _r11, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k11, _r12, _sum01);
-                _sum10 = _mm256_comp_fmadd_ps(_k01, _r11, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k01, _r12, _sum11);
-
-                _sum00 = _mm256_comp_fmadd_ps(_k12, _r12, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k12, _r13, _sum01);
-                _sum10 = _mm256_comp_fmadd_ps(_k02, _r12, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k02, _r13, _sum11);
-
-                __m256 _k20 = _mm256_load_ps(k0 + 48);
-                __m256 _k21 = _mm256_load_ps(k0 + 56);
-                __m256 _k22 = _mm256_load_ps(k0 + 64);
-
-                __m256 _r20 = _mm256_load_ps(r2);
-                __m256 _r21 = _mm256_load_ps(r2 + 8);
-                __m256 _r22 = _mm256_load_ps(r2 + 16);
-                __m256 _r23 = _mm256_load_ps(r2 + 24);
-
-                _sum00 = _mm256_comp_fmadd_ps(_k20, _r20, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k20, _r21, _sum01);
-                _sum10 = _mm256_comp_fmadd_ps(_k10, _r20, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k10, _r21, _sum11);
-
-                _sum00 = _mm256_comp_fmadd_ps(_k21, _r21, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k21, _r22, _sum01);
-                _sum10 = _mm256_comp_fmadd_ps(_k11, _r21, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k11, _r22, _sum11);
-
-                _sum00 = _mm256_comp_fmadd_ps(_k22, _r22, _sum00);
-                _sum01 = _mm256_comp_fmadd_ps(_k22, _r23, _sum01);
-                _sum10 = _mm256_comp_fmadd_ps(_k12, _r22, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k12, _r23, _sum11);
-
-                __m256 _r30 = _mm256_load_ps(r3);
-                __m256 _r31 = _mm256_load_ps(r3 + 8);
-                __m256 _r32 = _mm256_load_ps(r3 + 16);
-                __m256 _r33 = _mm256_load_ps(r3 + 24);
-
-                _sum10 = _mm256_comp_fmadd_ps(_k20, _r30, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k20, _r31, _sum11);
-                _sum10 = _mm256_comp_fmadd_ps(_k21, _r31, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k21, _r32, _sum11);
-                _sum10 = _mm256_comp_fmadd_ps(_k22, _r32, _sum10);
-                _sum11 = _mm256_comp_fmadd_ps(_k22, _r33, _sum11);
-
-                _mm256_store_ps(outptr0, _sum00);
-                _mm256_store_ps(outptr0 + 8, _sum01);
-                _mm256_store_ps(outptr1, _sum10);
-                _mm256_store_ps(outptr1 + 8, _sum11);
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                r3 += 16;
-                outptr0 += 16;
-                outptr1 += 16;
-            }
-            for (; j < outw; j++)
-            {
-                __m256 _sum0 = _bias0;
-                __m256 _sum1 = _bias0;
-
-                __m256 _k00 = _mm256_load_ps(k0);
-                __m256 _k01 = _mm256_load_ps(k0 + 8);
-                __m256 _k02 = _mm256_load_ps(k0 + 16);
-
-                __m256 _r00 = _mm256_load_ps(r0);
-                __m256 _r01 = _mm256_load_ps(r0 + 8);
-                __m256 _r02 = _mm256_load_ps(r0 + 16);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k02, _r02, _sum0);
-
-                __m256 _k10 = _mm256_load_ps(k0 + 24);
-                __m256 _k11 = _mm256_load_ps(k0 + 32);
-                __m256 _k12 = _mm256_load_ps(k0 + 40);
-
-                __m256 _r10 = _mm256_load_ps(r1);
-                __m256 _r11 = _mm256_load_ps(r1 + 8);
-                __m256 _r12 = _mm256_load_ps(r1 + 16);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k00, _r10, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k01, _r11, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k02, _r12, _sum1);
-
-                __m256 _k20 = _mm256_load_ps(k0 + 48);
-                __m256 _k21 = _mm256_load_ps(k0 + 56);
-                __m256 _k22 = _mm256_load_ps(k0 + 64);
-
-                __m256 _r20 = _mm256_load_ps(r2);
-                __m256 _r21 = _mm256_load_ps(r2 + 8);
-                __m256 _r22 = _mm256_load_ps(r2 + 16);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k10, _r20, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k11, _r21, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k22, _r22, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k12, _r22, _sum1);
-
-                __m256 _r30 = _mm256_load_ps(r3);
-                __m256 _r31 = _mm256_load_ps(r3 + 8);
-                __m256 _r32 = _mm256_load_ps(r3 + 16);
-
-                _sum1 = _mm256_comp_fmadd_ps(_k20, _r30, _sum1);
-                _sum1 = _mm256_comp_fmadd_ps(_k21, _r31, _sum1);
-                _sum1 = _mm256_comp_fmadd_ps(_k22, _r32, _sum1);
-
-                _mm256_store_ps(outptr0, _sum0);
-                _mm256_store_ps(outptr1, _sum1);
-
-                r0 += 8;
-                r1 += 8;
-                r2 += 8;
-                r3 += 8;
-                outptr0 += 8;
-                outptr1 += 8;
-            }
-
-            r0 += 2 * 8 + w * 8;
-            r1 += 2 * 8 + w * 8;
-            r2 += 2 * 8 + w * 8;
-            r3 += 2 * 8 + w * 8;
-
-            outptr0 += outw * 8;
-            outptr1 += outw * 8;
-        }
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                __m256 _sum0 = _bias0;
-                __m256 _sum1 = _bias0;
-                __m256 _sum2 = _bias0;
-                __m256 _sum3 = _bias0;
-
-                __m256 _k00 = _mm256_load_ps(k0);
-                __m256 _k01 = _mm256_load_ps(k0 + 8);
-                __m256 _k02 = _mm256_load_ps(k0 + 16);
-
-                __m256 _r00 = _mm256_load_ps(r0);
-                __m256 _r01 = _mm256_load_ps(r0 + 8);
-                __m256 _r02 = _mm256_load_ps(r0 + 16);
-                __m256 _r03 = _mm256_load_ps(r0 + 24);
-                __m256 _r04 = _mm256_load_ps(r0 + 32);
-                __m256 _r05 = _mm256_load_ps(r0 + 40);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k00, _r01, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k00, _r02, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k00, _r03, _sum3);
-                _sum0 = _mm256_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k01, _r02, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k01, _r03, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k01, _r04, _sum3);
-                _sum0 = _mm256_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k02, _r03, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k02, _r04, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k02, _r05, _sum3);
-
-                __m256 _k10 = _mm256_load_ps(k0 + 24);
-                __m256 _k11 = _mm256_load_ps(k0 + 32);
-                __m256 _k12 = _mm256_load_ps(k0 + 40);
-
-                __m256 _r10 = _mm256_load_ps(r1);
-                __m256 _r11 = _mm256_load_ps(r1 + 8);
-                __m256 _r12 = _mm256_load_ps(r1 + 16);
-                __m256 _r13 = _mm256_load_ps(r1 + 24);
-                __m256 _r14 = _mm256_load_ps(r1 + 32);
-                __m256 _r15 = _mm256_load_ps(r1 + 40);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k10, _r11, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k10, _r12, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k10, _r13, _sum3);
-                _sum0 = _mm256_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k11, _r12, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k11, _r13, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k11, _r14, _sum3);
-                _sum0 = _mm256_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k12, _r13, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k12, _r14, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k12, _r15, _sum3);
-
-                __m256 _k20 = _mm256_load_ps(k0 + 48);
-                __m256 _k21 = _mm256_load_ps(k0 + 56);
-                __m256 _k22 = _mm256_load_ps(k0 + 64);
-
-                __m256 _r20 = _mm256_load_ps(r2);
-                __m256 _r21 = _mm256_load_ps(r2 + 8);
-                __m256 _r22 = _mm256_load_ps(r2 + 16);
-                __m256 _r23 = _mm256_load_ps(r2 + 24);
-                __m256 _r24 = _mm256_load_ps(r2 + 32);
-                __m256 _r25 = _mm256_load_ps(r2 + 40);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k20, _r21, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k20, _r22, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k20, _r23, _sum3);
-                _sum0 = _mm256_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k21, _r22, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k21, _r23, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k21, _r24, _sum3);
-                _sum0 = _mm256_comp_fmadd_ps(_k22, _r22, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k22, _r23, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k22, _r24, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k22, _r25, _sum3);
-
-                _mm256_store_ps(outptr0, _sum0);
-                _mm256_store_ps(outptr0 + 8, _sum1);
-                _mm256_store_ps(outptr0 + 16, _sum2);
-                _mm256_store_ps(outptr0 + 24, _sum3);
-
-                r0 += 32;
-                r1 += 32;
-                r2 += 32;
-                outptr0 += 32;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                __m256 _sum0 = _bias0;
-                __m256 _sum1 = _bias0;
-
-                __m256 _k00 = _mm256_load_ps(k0);
-                __m256 _k01 = _mm256_load_ps(k0 + 8);
-                __m256 _k02 = _mm256_load_ps(k0 + 16);
-
-                __m256 _r00 = _mm256_load_ps(r0);
-                __m256 _r01 = _mm256_load_ps(r0 + 8);
-                __m256 _r02 = _mm256_load_ps(r0 + 16);
-                __m256 _r03 = _mm256_load_ps(r0 + 24);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k00, _r01, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k01, _r02, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k02, _r03, _sum1);
-
-                __m256 _k10 = _mm256_load_ps(k0 + 24);
-                __m256 _k11 = _mm256_load_ps(k0 + 32);
-                __m256 _k12 = _mm256_load_ps(k0 + 40);
-
-                __m256 _r10 = _mm256_load_ps(r1);
-                __m256 _r11 = _mm256_load_ps(r1 + 8);
-                __m256 _r12 = _mm256_load_ps(r1 + 16);
-                __m256 _r13 = _mm256_load_ps(r1 + 24);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k10, _r11, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k11, _r12, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k12, _r13, _sum1);
-
-                __m256 _k20 = _mm256_load_ps(k0 + 48);
-                __m256 _k21 = _mm256_load_ps(k0 + 56);
-                __m256 _k22 = _mm256_load_ps(k0 + 64);
-
-                __m256 _r20 = _mm256_load_ps(r2);
-                __m256 _r21 = _mm256_load_ps(r2 + 8);
-                __m256 _r22 = _mm256_load_ps(r2 + 16);
-                __m256 _r23 = _mm256_load_ps(r2 + 24);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k20, _r21, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k21, _r22, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k22, _r22, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k22, _r23, _sum1);
-
-                _mm256_store_ps(outptr0, _sum0);
-                _mm256_store_ps(outptr0 + 8, _sum1);
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                outptr0 += 16;
-            }
-            for (; j < outw; j++)
-            {
-                __m256 _sum0 = _bias0;
-
-                __m256 _k00 = _mm256_load_ps(k0);
-                __m256 _k01 = _mm256_load_ps(k0 + 8);
-                __m256 _k02 = _mm256_load_ps(k0 + 16);
-
-                __m256 _r00 = _mm256_load_ps(r0);
-                __m256 _r01 = _mm256_load_ps(r0 + 8);
-                __m256 _r02 = _mm256_load_ps(r0 + 16);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k02, _r02, _sum0);
-
-                __m256 _k10 = _mm256_load_ps(k0 + 24);
-                __m256 _k11 = _mm256_load_ps(k0 + 32);
-                __m256 _k12 = _mm256_load_ps(k0 + 40);
-
-                __m256 _r10 = _mm256_load_ps(r1);
-                __m256 _r11 = _mm256_load_ps(r1 + 8);
-                __m256 _r12 = _mm256_load_ps(r1 + 16);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k12, _r12, _sum0);
-
-                __m256 _k20 = _mm256_load_ps(k0 + 48);
-                __m256 _k21 = _mm256_load_ps(k0 + 56);
-                __m256 _k22 = _mm256_load_ps(k0 + 64);
-
-                __m256 _r20 = _mm256_load_ps(r2);
-                __m256 _r21 = _mm256_load_ps(r2 + 8);
-                __m256 _r22 = _mm256_load_ps(r2 + 16);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k22, _r22, _sum0);
-
-                _mm256_store_ps(outptr0, _sum0);
-
-                r0 += 8;
-                r1 += 8;
-                r2 += 8;
-                outptr0 += 8;
-            }
-
-            r0 += 2 * 8;
-            r1 += 2 * 8;
-            r2 += 2 * 8;
-        }
-    }
-}
-
-static void convdw3x3s2_pack8_avx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 8;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m256 _bias0 = bias ? _mm256_loadu_ps((const float*)bias + g * 8) : _mm256_setzero_ps();
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-
-        __m256 _k00 = _mm256_load_ps(k0);
-        __m256 _k01 = _mm256_load_ps(k0 + 8);
-        __m256 _k02 = _mm256_load_ps(k0 + 16);
-        __m256 _k10 = _mm256_load_ps(k0 + 24);
-        __m256 _k11 = _mm256_load_ps(k0 + 32);
-        __m256 _k12 = _mm256_load_ps(k0 + 40);
-        __m256 _k20 = _mm256_load_ps(k0 + 48);
-        __m256 _k21 = _mm256_load_ps(k0 + 56);
-        __m256 _k22 = _mm256_load_ps(k0 + 64);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                __m256 _sum0 = _bias0;
-                __m256 _sum1 = _bias0;
-                __m256 _sum2 = _bias0;
-                __m256 _sum3 = _bias0;
-
-                __m256 _r00 = _mm256_load_ps(r0);
-                __m256 _r01 = _mm256_load_ps(r0 + 8);
-                __m256 _r02 = _mm256_load_ps(r0 + 16);
-                __m256 _r03 = _mm256_load_ps(r0 + 24);
-                __m256 _r04 = _mm256_load_ps(r0 + 32);
-                __m256 _r05 = _mm256_load_ps(r0 + 40);
-                __m256 _r06 = _mm256_load_ps(r0 + 48);
-                __m256 _r07 = _mm256_load_ps(r0 + 56);
-                __m256 _r08 = _mm256_load_ps(r0 + 64);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k00, _r02, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k00, _r04, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k00, _r06, _sum3);
-                _sum0 = _mm256_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k01, _r03, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k01, _r05, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k01, _r07, _sum3);
-                _sum0 = _mm256_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k02, _r04, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k02, _r06, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k02, _r08, _sum3);
-
-                __m256 _r10 = _mm256_load_ps(r1);
-                __m256 _r11 = _mm256_load_ps(r1 + 8);
-                __m256 _r12 = _mm256_load_ps(r1 + 16);
-                __m256 _r13 = _mm256_load_ps(r1 + 24);
-                __m256 _r14 = _mm256_load_ps(r1 + 32);
-                __m256 _r15 = _mm256_load_ps(r1 + 40);
-                __m256 _r16 = _mm256_load_ps(r1 + 48);
-                __m256 _r17 = _mm256_load_ps(r1 + 56);
-                __m256 _r18 = _mm256_load_ps(r1 + 64);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k10, _r12, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k10, _r14, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k10, _r16, _sum3);
-                _sum0 = _mm256_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k11, _r13, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k11, _r15, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k11, _r17, _sum3);
-                _sum0 = _mm256_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k12, _r14, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k12, _r16, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k12, _r18, _sum3);
-
-                __m256 _r20 = _mm256_load_ps(r2);
-                __m256 _r21 = _mm256_load_ps(r2 + 8);
-                __m256 _r22 = _mm256_load_ps(r2 + 16);
-                __m256 _r23 = _mm256_load_ps(r2 + 24);
-                __m256 _r24 = _mm256_load_ps(r2 + 32);
-                __m256 _r25 = _mm256_load_ps(r2 + 40);
-                __m256 _r26 = _mm256_load_ps(r2 + 48);
-                __m256 _r27 = _mm256_load_ps(r2 + 56);
-                __m256 _r28 = _mm256_load_ps(r2 + 64);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k20, _r22, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k20, _r24, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k20, _r26, _sum3);
-                _sum0 = _mm256_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k21, _r23, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k21, _r25, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k21, _r27, _sum3);
-                _sum0 = _mm256_comp_fmadd_ps(_k22, _r22, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k22, _r24, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_k22, _r26, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_k22, _r28, _sum3);
-
-                _mm256_store_ps(outptr0, _sum0);
-                _mm256_store_ps(outptr0 + 8, _sum1);
-                _mm256_store_ps(outptr0 + 16, _sum2);
-                _mm256_store_ps(outptr0 + 24, _sum3);
-
-                r0 += 2 * 32;
-                r1 += 2 * 32;
-                r2 += 2 * 32;
-                outptr0 += 32;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                __m256 _sum0 = _bias0;
-                __m256 _sum1 = _bias0;
-
-                __m256 _r00 = _mm256_load_ps(r0);
-                __m256 _r01 = _mm256_load_ps(r0 + 8);
-                __m256 _r02 = _mm256_load_ps(r0 + 16);
-                __m256 _r03 = _mm256_load_ps(r0 + 24);
-                __m256 _r04 = _mm256_load_ps(r0 + 32);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k00, _r02, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k01, _r03, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k02, _r04, _sum1);
-
-                __m256 _r10 = _mm256_load_ps(r1);
-                __m256 _r11 = _mm256_load_ps(r1 + 8);
-                __m256 _r12 = _mm256_load_ps(r1 + 16);
-                __m256 _r13 = _mm256_load_ps(r1 + 24);
-                __m256 _r14 = _mm256_load_ps(r1 + 32);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k10, _r12, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k11, _r13, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k12, _r14, _sum1);
-
-                __m256 _r20 = _mm256_load_ps(r2);
-                __m256 _r21 = _mm256_load_ps(r2 + 8);
-                __m256 _r22 = _mm256_load_ps(r2 + 16);
-                __m256 _r23 = _mm256_load_ps(r2 + 24);
-                __m256 _r24 = _mm256_load_ps(r2 + 32);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k20, _r22, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k21, _r23, _sum1);
-                _sum0 = _mm256_comp_fmadd_ps(_k22, _r22, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_k22, _r24, _sum1);
-
-                _mm256_store_ps(outptr0, _sum0);
-                _mm256_store_ps(outptr0 + 8, _sum1);
-
-                r0 += 2 * 16;
-                r1 += 2 * 16;
-                r2 += 2 * 16;
-                outptr0 += 16;
-            }
-            for (; j < outw; j++)
-            {
-                __m256 _sum0 = _bias0;
-
-                __m256 _r00 = _mm256_load_ps(r0);
-                __m256 _r01 = _mm256_load_ps(r0 + 8);
-                __m256 _r02 = _mm256_load_ps(r0 + 16);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k02, _r02, _sum0);
-
-                __m256 _r10 = _mm256_load_ps(r1);
-                __m256 _r11 = _mm256_load_ps(r1 + 8);
-                __m256 _r12 = _mm256_load_ps(r1 + 16);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k12, _r12, _sum0);
-
-                __m256 _r20 = _mm256_load_ps(r2);
-                __m256 _r21 = _mm256_load_ps(r2 + 8);
-                __m256 _r22 = _mm256_load_ps(r2 + 16);
-
-                _sum0 = _mm256_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k22, _r22, _sum0);
-
-                _mm256_store_ps(outptr0, _sum0);
-
-                r0 += 2 * 8;
-                r1 += 2 * 8;
-                r2 += 2 * 8;
-                outptr0 += 8;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_5x5_pack16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_5x5_pack16.h
deleted file mode 100644
index 4485d982..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_5x5_pack16.h
+++ /dev/null
@@ -1,317 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw5x5s1_pack16_avx512(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m512 _bias0 = bias ? _mm512_loadu_ps((const float*)bias + g * 16) : _mm512_setzero_ps();
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-        const float* r4 = img0.row(4);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j < outw; j++)
-            {
-                __m512 _sum0 = _bias0;
-
-                __m512 _r00 = _mm512_load_ps(r0);
-                __m512 _r01 = _mm512_load_ps(r0 + 16);
-                __m512 _r02 = _mm512_load_ps(r0 + 32);
-                __m512 _r03 = _mm512_load_ps(r0 + 48);
-                __m512 _r04 = _mm512_load_ps(r0 + 64);
-
-                __m512 _k00 = _mm512_load_ps(k0);
-                __m512 _k01 = _mm512_load_ps(k0 + 16);
-                __m512 _k02 = _mm512_load_ps(k0 + 32);
-                __m512 _k03 = _mm512_load_ps(k0 + 48);
-                __m512 _k04 = _mm512_load_ps(k0 + 64);
-                k0 += 80;
-
-                _sum0 = _mm512_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k02, _r02, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k03, _r03, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k04, _r04, _sum0);
-
-                __m512 _r10 = _mm512_load_ps(r1);
-                __m512 _r11 = _mm512_load_ps(r1 + 16);
-                __m512 _r12 = _mm512_load_ps(r1 + 32);
-                __m512 _r13 = _mm512_load_ps(r1 + 48);
-                __m512 _r14 = _mm512_load_ps(r1 + 64);
-
-                __m512 _k10 = _mm512_load_ps(k0);
-                __m512 _k11 = _mm512_load_ps(k0 + 16);
-                __m512 _k12 = _mm512_load_ps(k0 + 32);
-                __m512 _k13 = _mm512_load_ps(k0 + 48);
-                __m512 _k14 = _mm512_load_ps(k0 + 64);
-                k0 += 80;
-
-                _sum0 = _mm512_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k12, _r12, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k13, _r13, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k14, _r14, _sum0);
-
-                __m512 _r20 = _mm512_load_ps(r2);
-                __m512 _r21 = _mm512_load_ps(r2 + 16);
-                __m512 _r22 = _mm512_load_ps(r2 + 32);
-                __m512 _r23 = _mm512_load_ps(r2 + 48);
-                __m512 _r24 = _mm512_load_ps(r2 + 64);
-
-                __m512 _k20 = _mm512_load_ps(k0);
-                __m512 _k21 = _mm512_load_ps(k0 + 16);
-                __m512 _k22 = _mm512_load_ps(k0 + 32);
-                __m512 _k23 = _mm512_load_ps(k0 + 48);
-                __m512 _k24 = _mm512_load_ps(k0 + 64);
-                k0 += 80;
-
-                _sum0 = _mm512_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k22, _r22, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k23, _r23, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k24, _r24, _sum0);
-
-                __m512 _r30 = _mm512_load_ps(r3);
-                __m512 _r31 = _mm512_load_ps(r3 + 16);
-                __m512 _r32 = _mm512_load_ps(r3 + 32);
-                __m512 _r33 = _mm512_load_ps(r3 + 48);
-                __m512 _r34 = _mm512_load_ps(r3 + 64);
-
-                __m512 _k30 = _mm512_load_ps(k0);
-                __m512 _k31 = _mm512_load_ps(k0 + 16);
-                __m512 _k32 = _mm512_load_ps(k0 + 32);
-                __m512 _k33 = _mm512_load_ps(k0 + 48);
-                __m512 _k34 = _mm512_load_ps(k0 + 64);
-                k0 += 80;
-
-                _sum0 = _mm512_fmadd_ps(_k30, _r30, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k31, _r31, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k32, _r32, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k33, _r33, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k34, _r34, _sum0);
-
-                __m512 _r40 = _mm512_load_ps(r4);
-                __m512 _r41 = _mm512_load_ps(r4 + 16);
-                __m512 _r42 = _mm512_load_ps(r4 + 32);
-                __m512 _r43 = _mm512_load_ps(r4 + 48);
-                __m512 _r44 = _mm512_load_ps(r4 + 64);
-
-                __m512 _k40 = _mm512_load_ps(k0);
-                __m512 _k41 = _mm512_load_ps(k0 + 16);
-                __m512 _k42 = _mm512_load_ps(k0 + 32);
-                __m512 _k43 = _mm512_load_ps(k0 + 48);
-                __m512 _k44 = _mm512_load_ps(k0 + 64);
-                k0 -= 320;
-
-                _sum0 = _mm512_fmadd_ps(_k40, _r40, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k41, _r41, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k42, _r42, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k43, _r43, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k44, _r44, _sum0);
-
-                _mm512_store_ps(outptr0, _sum0);
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                r3 += 16;
-                r4 += 16;
-                outptr0 += 16;
-            }
-
-            r0 += 4 * 16;
-            r1 += 4 * 16;
-            r2 += 4 * 16;
-            r3 += 4 * 16;
-            r4 += 4 * 16;
-        }
-    }
-}
-
-static void convdw5x5s2_pack16_avx512(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 16;
-
-    const float* bias = _bias;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m512 _bias0 = bias ? _mm512_loadu_ps((const float*)bias + g * 16) : _mm512_setzero_ps();
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-        const float* r4 = img0.row(4);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j < outw; j++)
-            {
-                __m512 _sum0 = _bias0;
-
-                __m512 _r00 = _mm512_load_ps(r0);
-                __m512 _r01 = _mm512_load_ps(r0 + 16);
-                __m512 _r02 = _mm512_load_ps(r0 + 32);
-                __m512 _r03 = _mm512_load_ps(r0 + 48);
-                __m512 _r04 = _mm512_load_ps(r0 + 64);
-
-                __m512 _k00 = _mm512_load_ps(k0);
-                __m512 _k01 = _mm512_load_ps(k0 + 16);
-                __m512 _k02 = _mm512_load_ps(k0 + 32);
-                __m512 _k03 = _mm512_load_ps(k0 + 48);
-                __m512 _k04 = _mm512_load_ps(k0 + 64);
-                k0 += 80;
-
-                _sum0 = _mm512_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k02, _r02, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k03, _r03, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k04, _r04, _sum0);
-
-                __m512 _r10 = _mm512_load_ps(r1);
-                __m512 _r11 = _mm512_load_ps(r1 + 16);
-                __m512 _r12 = _mm512_load_ps(r1 + 32);
-                __m512 _r13 = _mm512_load_ps(r1 + 48);
-                __m512 _r14 = _mm512_load_ps(r1 + 64);
-
-                __m512 _k10 = _mm512_load_ps(k0);
-                __m512 _k11 = _mm512_load_ps(k0 + 16);
-                __m512 _k12 = _mm512_load_ps(k0 + 32);
-                __m512 _k13 = _mm512_load_ps(k0 + 48);
-                __m512 _k14 = _mm512_load_ps(k0 + 64);
-                k0 += 80;
-
-                _sum0 = _mm512_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k12, _r12, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k13, _r13, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k14, _r14, _sum0);
-
-                __m512 _r20 = _mm512_load_ps(r2);
-                __m512 _r21 = _mm512_load_ps(r2 + 16);
-                __m512 _r22 = _mm512_load_ps(r2 + 32);
-                __m512 _r23 = _mm512_load_ps(r2 + 48);
-                __m512 _r24 = _mm512_load_ps(r2 + 64);
-
-                __m512 _k20 = _mm512_load_ps(k0);
-                __m512 _k21 = _mm512_load_ps(k0 + 16);
-                __m512 _k22 = _mm512_load_ps(k0 + 32);
-                __m512 _k23 = _mm512_load_ps(k0 + 48);
-                __m512 _k24 = _mm512_load_ps(k0 + 64);
-                k0 += 80;
-
-                _sum0 = _mm512_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k22, _r22, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k23, _r23, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k24, _r24, _sum0);
-
-                __m512 _r30 = _mm512_load_ps(r3);
-                __m512 _r31 = _mm512_load_ps(r3 + 16);
-                __m512 _r32 = _mm512_load_ps(r3 + 32);
-                __m512 _r33 = _mm512_load_ps(r3 + 48);
-                __m512 _r34 = _mm512_load_ps(r3 + 64);
-
-                __m512 _k30 = _mm512_load_ps(k0);
-                __m512 _k31 = _mm512_load_ps(k0 + 16);
-                __m512 _k32 = _mm512_load_ps(k0 + 32);
-                __m512 _k33 = _mm512_load_ps(k0 + 48);
-                __m512 _k34 = _mm512_load_ps(k0 + 64);
-                k0 += 80;
-
-                _sum0 = _mm512_fmadd_ps(_k30, _r30, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k31, _r31, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k32, _r32, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k33, _r33, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k34, _r34, _sum0);
-
-                __m512 _r40 = _mm512_load_ps(r4);
-                __m512 _r41 = _mm512_load_ps(r4 + 16);
-                __m512 _r42 = _mm512_load_ps(r4 + 32);
-                __m512 _r43 = _mm512_load_ps(r4 + 48);
-                __m512 _r44 = _mm512_load_ps(r4 + 64);
-
-                __m512 _k40 = _mm512_load_ps(k0);
-                __m512 _k41 = _mm512_load_ps(k0 + 16);
-                __m512 _k42 = _mm512_load_ps(k0 + 32);
-                __m512 _k43 = _mm512_load_ps(k0 + 48);
-                __m512 _k44 = _mm512_load_ps(k0 + 64);
-                k0 -= 320;
-
-                _sum0 = _mm512_fmadd_ps(_k40, _r40, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k41, _r41, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k42, _r42, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k43, _r43, _sum0);
-                _sum0 = _mm512_fmadd_ps(_k44, _r44, _sum0);
-
-                _mm512_store_ps(outptr0, _sum0);
-
-                r0 += 32;
-                r1 += 32;
-                r2 += 32;
-                r3 += 32;
-                r4 += 32;
-                outptr0 += 16;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-            r3 += tailstep;
-            r4 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_5x5_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_5x5_pack4.h
deleted file mode 100644
index 889df7ad..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_5x5_pack4.h
+++ /dev/null
@@ -1,486 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw5x5s1_pack4_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m128 _bias0 = bias ? _mm_loadu_ps(bias + g * 4) : _mm_setzero_ps();
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-        float* outptr1 = out.row(1);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-        const float* r4 = img0.row(4);
-        const float* r5 = img0.row(5);
-
-        int i = 0;
-        for (; i + 1 < outh; i += 2)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                __m128 _sum0 = _bias0;
-                __m128 _sum1 = _bias0;
-
-                __m128 _r00 = _mm_load_ps(r0);
-                __m128 _r01 = _mm_load_ps(r0 + 4);
-                __m128 _r02 = _mm_load_ps(r0 + 4 * 2);
-                __m128 _r03 = _mm_load_ps(r0 + 4 * 3);
-                __m128 _r04 = _mm_load_ps(r0 + 4 * 4);
-
-                __m128 _k00 = _mm_load_ps(k0);
-                __m128 _k01 = _mm_load_ps(k0 + 4);
-                __m128 _k02 = _mm_load_ps(k0 + 4 * 2);
-                __m128 _k03 = _mm_load_ps(k0 + 4 * 3);
-                __m128 _k04 = _mm_load_ps(k0 + 4 * 4);
-                k0 += 4 * 5;
-
-                _sum0 = _mm_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k03, _r03, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k04, _r04, _sum0);
-
-                __m128 _r10 = _mm_load_ps(r1);
-                __m128 _r11 = _mm_load_ps(r1 + 4);
-                __m128 _r12 = _mm_load_ps(r1 + 4 * 2);
-                __m128 _r13 = _mm_load_ps(r1 + 4 * 3);
-                __m128 _r14 = _mm_load_ps(r1 + 4 * 4);
-
-                _sum1 = _mm_comp_fmadd_ps(_k00, _r10, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k01, _r11, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k02, _r12, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k03, _r13, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k04, _r14, _sum1);
-
-                __m128 _k10 = _mm_load_ps(k0);
-                __m128 _k11 = _mm_load_ps(k0 + 4);
-                __m128 _k12 = _mm_load_ps(k0 + 4 * 2);
-                __m128 _k13 = _mm_load_ps(k0 + 4 * 3);
-                __m128 _k14 = _mm_load_ps(k0 + 4 * 4);
-                k0 += 4 * 5;
-
-                _sum0 = _mm_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k13, _r13, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k14, _r14, _sum0);
-
-                __m128 _r20 = _mm_load_ps(r2);
-                __m128 _r21 = _mm_load_ps(r2 + 4);
-                __m128 _r22 = _mm_load_ps(r2 + 4 * 2);
-                __m128 _r23 = _mm_load_ps(r2 + 4 * 3);
-                __m128 _r24 = _mm_load_ps(r2 + 4 * 4);
-
-                _sum1 = _mm_comp_fmadd_ps(_k10, _r20, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k11, _r21, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k12, _r22, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k13, _r23, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k14, _r24, _sum1);
-
-                __m128 _k20 = _mm_load_ps(k0);
-                __m128 _k21 = _mm_load_ps(k0 + 4);
-                __m128 _k22 = _mm_load_ps(k0 + 4 * 2);
-                __m128 _k23 = _mm_load_ps(k0 + 4 * 3);
-                __m128 _k24 = _mm_load_ps(k0 + 4 * 4);
-                k0 += 4 * 5;
-
-                _sum0 = _mm_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k22, _r22, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k23, _r23, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k24, _r24, _sum0);
-
-                __m128 _r30 = _mm_load_ps(r3);
-                __m128 _r31 = _mm_load_ps(r3 + 4);
-                __m128 _r32 = _mm_load_ps(r3 + 4 * 2);
-                __m128 _r33 = _mm_load_ps(r3 + 4 * 3);
-                __m128 _r34 = _mm_load_ps(r3 + 4 * 4);
-
-                _sum1 = _mm_comp_fmadd_ps(_k20, _r30, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k21, _r31, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k22, _r32, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k23, _r33, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k24, _r34, _sum1);
-
-                __m128 _k30 = _mm_load_ps(k0);
-                __m128 _k31 = _mm_load_ps(k0 + 4);
-                __m128 _k32 = _mm_load_ps(k0 + 4 * 2);
-                __m128 _k33 = _mm_load_ps(k0 + 4 * 3);
-                __m128 _k34 = _mm_load_ps(k0 + 4 * 4);
-                k0 += 4 * 5;
-
-                _sum0 = _mm_comp_fmadd_ps(_k30, _r30, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k31, _r31, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k32, _r32, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k33, _r33, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k34, _r34, _sum0);
-
-                __m128 _r40 = _mm_load_ps(r4);
-                __m128 _r41 = _mm_load_ps(r4 + 4);
-                __m128 _r42 = _mm_load_ps(r4 + 4 * 2);
-                __m128 _r43 = _mm_load_ps(r4 + 4 * 3);
-                __m128 _r44 = _mm_load_ps(r4 + 4 * 4);
-
-                _sum1 = _mm_comp_fmadd_ps(_k30, _r40, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k31, _r41, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k32, _r42, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k33, _r43, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k34, _r44, _sum1);
-
-                __m128 _k40 = _mm_load_ps(k0);
-                __m128 _k41 = _mm_load_ps(k0 + 4);
-                __m128 _k42 = _mm_load_ps(k0 + 4 * 2);
-                __m128 _k43 = _mm_load_ps(k0 + 4 * 3);
-                __m128 _k44 = _mm_load_ps(k0 + 4 * 4);
-                k0 -= 4 * 20;
-
-                _sum0 = _mm_comp_fmadd_ps(_k40, _r40, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k41, _r41, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k42, _r42, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k43, _r43, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k44, _r44, _sum0);
-
-                __m128 _r50 = _mm_load_ps(r5);
-                __m128 _r51 = _mm_load_ps(r5 + 4);
-                __m128 _r52 = _mm_load_ps(r5 + 4 * 2);
-                __m128 _r53 = _mm_load_ps(r5 + 4 * 3);
-                __m128 _r54 = _mm_load_ps(r5 + 4 * 4);
-
-                _sum1 = _mm_comp_fmadd_ps(_k40, _r50, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k41, _r51, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k42, _r52, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k43, _r53, _sum1);
-                _sum1 = _mm_comp_fmadd_ps(_k44, _r54, _sum1);
-
-                _mm_store_ps(outptr0, _sum0);
-                _mm_store_ps(outptr1, _sum1);
-
-                outptr0 += 4;
-                outptr1 += 4;
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                r3 += 4;
-                r4 += 4;
-                r5 += 4;
-            }
-
-            r0 += 4 * 4 + w * 4;
-            r1 += 4 * 4 + w * 4;
-            r2 += 4 * 4 + w * 4;
-            r3 += 4 * 4 + w * 4;
-            r4 += 4 * 4 + w * 4;
-            r5 += 4 * 4 + w * 4;
-
-            outptr0 += outw * 4;
-            outptr1 += outw * 4;
-        }
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                __m128 _sum0 = _bias0;
-
-                __m128 _r00 = _mm_load_ps(r0);
-                __m128 _r01 = _mm_load_ps(r0 + 4);
-                __m128 _r02 = _mm_load_ps(r0 + 4 * 2);
-                __m128 _r03 = _mm_load_ps(r0 + 4 * 3);
-                __m128 _r04 = _mm_load_ps(r0 + 4 * 4);
-
-                __m128 _k00 = _mm_load_ps(k0);
-                __m128 _k01 = _mm_load_ps(k0 + 4);
-                __m128 _k02 = _mm_load_ps(k0 + 4 * 2);
-                __m128 _k03 = _mm_load_ps(k0 + 4 * 3);
-                __m128 _k04 = _mm_load_ps(k0 + 4 * 4);
-                k0 += 4 * 5;
-
-                _sum0 = _mm_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k03, _r03, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k04, _r04, _sum0);
-
-                __m128 _r10 = _mm_load_ps(r1);
-                __m128 _r11 = _mm_load_ps(r1 + 4);
-                __m128 _r12 = _mm_load_ps(r1 + 4 * 2);
-                __m128 _r13 = _mm_load_ps(r1 + 4 * 3);
-                __m128 _r14 = _mm_load_ps(r1 + 4 * 4);
-
-                __m128 _k10 = _mm_load_ps(k0);
-                __m128 _k11 = _mm_load_ps(k0 + 4);
-                __m128 _k12 = _mm_load_ps(k0 + 4 * 2);
-                __m128 _k13 = _mm_load_ps(k0 + 4 * 3);
-                __m128 _k14 = _mm_load_ps(k0 + 4 * 4);
-                k0 += 4 * 5;
-
-                _sum0 = _mm_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k13, _r13, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k14, _r14, _sum0);
-
-                __m128 _r20 = _mm_load_ps(r2);
-                __m128 _r21 = _mm_load_ps(r2 + 4);
-                __m128 _r22 = _mm_load_ps(r2 + 4 * 2);
-                __m128 _r23 = _mm_load_ps(r2 + 4 * 3);
-                __m128 _r24 = _mm_load_ps(r2 + 4 * 4);
-
-                __m128 _k20 = _mm_load_ps(k0);
-                __m128 _k21 = _mm_load_ps(k0 + 4);
-                __m128 _k22 = _mm_load_ps(k0 + 4 * 2);
-                __m128 _k23 = _mm_load_ps(k0 + 4 * 3);
-                __m128 _k24 = _mm_load_ps(k0 + 4 * 4);
-                k0 += 4 * 5;
-
-                _sum0 = _mm_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k22, _r22, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k23, _r23, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k24, _r24, _sum0);
-
-                __m128 _r30 = _mm_load_ps(r3);
-                __m128 _r31 = _mm_load_ps(r3 + 4);
-                __m128 _r32 = _mm_load_ps(r3 + 4 * 2);
-                __m128 _r33 = _mm_load_ps(r3 + 4 * 3);
-                __m128 _r34 = _mm_load_ps(r3 + 4 * 4);
-
-                __m128 _k30 = _mm_load_ps(k0);
-                __m128 _k31 = _mm_load_ps(k0 + 4);
-                __m128 _k32 = _mm_load_ps(k0 + 4 * 2);
-                __m128 _k33 = _mm_load_ps(k0 + 4 * 3);
-                __m128 _k34 = _mm_load_ps(k0 + 4 * 4);
-                k0 += 4 * 5;
-
-                _sum0 = _mm_comp_fmadd_ps(_k30, _r30, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k31, _r31, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k32, _r32, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k33, _r33, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k34, _r34, _sum0);
-
-                __m128 _r40 = _mm_load_ps(r4);
-                __m128 _r41 = _mm_load_ps(r4 + 4);
-                __m128 _r42 = _mm_load_ps(r4 + 4 * 2);
-                __m128 _r43 = _mm_load_ps(r4 + 4 * 3);
-                __m128 _r44 = _mm_load_ps(r4 + 4 * 4);
-
-                __m128 _k40 = _mm_load_ps(k0);
-                __m128 _k41 = _mm_load_ps(k0 + 4);
-                __m128 _k42 = _mm_load_ps(k0 + 4 * 2);
-                __m128 _k43 = _mm_load_ps(k0 + 4 * 3);
-                __m128 _k44 = _mm_load_ps(k0 + 4 * 4);
-                k0 -= 4 * 20;
-
-                _sum0 = _mm_comp_fmadd_ps(_k40, _r40, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k41, _r41, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k42, _r42, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k43, _r43, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k44, _r44, _sum0);
-
-                _mm_store_ps(outptr0, _sum0);
-
-                outptr0 += 4;
-
-                r0 += 4;
-                r1 += 4;
-                r2 += 4;
-                r3 += 4;
-                r4 += 4;
-            }
-
-            r0 += 4 * 4;
-            r1 += 4 * 4;
-            r2 += 4 * 4;
-            r3 += 4 * 4;
-            r4 += 4 * 4;
-        }
-    }
-}
-
-static void convdw5x5s2_pack4_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    const float* bias = _bias;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m128 _bias0 = bias ? _mm_loadu_ps(bias + g * 4) : _mm_setzero_ps();
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out;
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-        const float* r4 = img0.row(4);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-            for (; j < outw; j++)
-            {
-                __m128 _sum0 = _bias0;
-
-                __m128 _r00 = _mm_load_ps(r0);
-                __m128 _r01 = _mm_load_ps(r0 + 4);
-                __m128 _r02 = _mm_load_ps(r0 + 4 * 2);
-                __m128 _r03 = _mm_load_ps(r0 + 4 * 3);
-                __m128 _r04 = _mm_load_ps(r0 + 4 * 4);
-
-                __m128 _k00 = _mm_load_ps(k0);
-                __m128 _k01 = _mm_load_ps(k0 + 4);
-                __m128 _k02 = _mm_load_ps(k0 + 4 * 2);
-                __m128 _k03 = _mm_load_ps(k0 + 4 * 3);
-                __m128 _k04 = _mm_load_ps(k0 + 4 * 4);
-                k0 += 4 * 5;
-
-                _sum0 = _mm_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k03, _r03, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k04, _r04, _sum0);
-
-                __m128 _r10 = _mm_load_ps(r1);
-                __m128 _r11 = _mm_load_ps(r1 + 4);
-                __m128 _r12 = _mm_load_ps(r1 + 4 * 2);
-                __m128 _r13 = _mm_load_ps(r1 + 4 * 3);
-                __m128 _r14 = _mm_load_ps(r1 + 4 * 4);
-
-                __m128 _k10 = _mm_load_ps(k0);
-                __m128 _k11 = _mm_load_ps(k0 + 4);
-                __m128 _k12 = _mm_load_ps(k0 + 4 * 2);
-                __m128 _k13 = _mm_load_ps(k0 + 4 * 3);
-                __m128 _k14 = _mm_load_ps(k0 + 4 * 4);
-                k0 += 4 * 5;
-
-                _sum0 = _mm_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k13, _r13, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k14, _r14, _sum0);
-
-                __m128 _r20 = _mm_load_ps(r2);
-                __m128 _r21 = _mm_load_ps(r2 + 4);
-                __m128 _r22 = _mm_load_ps(r2 + 4 * 2);
-                __m128 _r23 = _mm_load_ps(r2 + 4 * 3);
-                __m128 _r24 = _mm_load_ps(r2 + 4 * 4);
-
-                __m128 _k20 = _mm_load_ps(k0);
-                __m128 _k21 = _mm_load_ps(k0 + 4);
-                __m128 _k22 = _mm_load_ps(k0 + 4 * 2);
-                __m128 _k23 = _mm_load_ps(k0 + 4 * 3);
-                __m128 _k24 = _mm_load_ps(k0 + 4 * 4);
-                k0 += 4 * 5;
-
-                _sum0 = _mm_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k22, _r22, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k23, _r23, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k24, _r24, _sum0);
-
-                __m128 _r30 = _mm_load_ps(r3);
-                __m128 _r31 = _mm_load_ps(r3 + 4);
-                __m128 _r32 = _mm_load_ps(r3 + 4 * 2);
-                __m128 _r33 = _mm_load_ps(r3 + 4 * 3);
-                __m128 _r34 = _mm_load_ps(r3 + 4 * 4);
-
-                __m128 _k30 = _mm_load_ps(k0);
-                __m128 _k31 = _mm_load_ps(k0 + 4);
-                __m128 _k32 = _mm_load_ps(k0 + 4 * 2);
-                __m128 _k33 = _mm_load_ps(k0 + 4 * 3);
-                __m128 _k34 = _mm_load_ps(k0 + 4 * 4);
-                k0 += 4 * 5;
-
-                _sum0 = _mm_comp_fmadd_ps(_k30, _r30, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k31, _r31, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k32, _r32, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k33, _r33, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k34, _r34, _sum0);
-
-                __m128 _r40 = _mm_load_ps(r4);
-                __m128 _r41 = _mm_load_ps(r4 + 4);
-                __m128 _r42 = _mm_load_ps(r4 + 4 * 2);
-                __m128 _r43 = _mm_load_ps(r4 + 4 * 3);
-                __m128 _r44 = _mm_load_ps(r4 + 4 * 4);
-
-                __m128 _k40 = _mm_load_ps(k0);
-                __m128 _k41 = _mm_load_ps(k0 + 4);
-                __m128 _k42 = _mm_load_ps(k0 + 4 * 2);
-                __m128 _k43 = _mm_load_ps(k0 + 4 * 3);
-                __m128 _k44 = _mm_load_ps(k0 + 4 * 4);
-                k0 -= 4 * 20;
-
-                _sum0 = _mm_comp_fmadd_ps(_k40, _r40, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k41, _r41, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k42, _r42, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k43, _r43, _sum0);
-                _sum0 = _mm_comp_fmadd_ps(_k44, _r44, _sum0);
-
-                _mm_store_ps(outptr0, _sum0);
-
-                outptr0 += 4;
-
-                r0 += 4 * 2;
-                r1 += 4 * 2;
-                r2 += 4 * 2;
-                r3 += 4 * 2;
-                r4 += 4 * 2;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-            r3 += tailstep;
-            r4 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_5x5_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_5x5_pack8.h
deleted file mode 100644
index ec8189f0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_5x5_pack8.h
+++ /dev/null
@@ -1,317 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void convdw5x5s1_pack8_avx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const float* bias = _bias;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m256 _bias0 = bias ? _mm256_loadu_ps((const float*)bias + g * 8) : _mm256_setzero_ps();
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-        const float* r4 = img0.row(4);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j < outw; j++)
-            {
-                __m256 _sum0 = _bias0;
-
-                __m256 _r00 = _mm256_load_ps(r0);
-                __m256 _r01 = _mm256_load_ps(r0 + 8);
-                __m256 _r02 = _mm256_load_ps(r0 + 16);
-                __m256 _r03 = _mm256_load_ps(r0 + 24);
-                __m256 _r04 = _mm256_load_ps(r0 + 32);
-
-                __m256 _k00 = _mm256_load_ps(k0);
-                __m256 _k01 = _mm256_load_ps(k0 + 8);
-                __m256 _k02 = _mm256_load_ps(k0 + 16);
-                __m256 _k03 = _mm256_load_ps(k0 + 24);
-                __m256 _k04 = _mm256_load_ps(k0 + 32);
-                k0 += 40;
-
-                _sum0 = _mm256_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k03, _r03, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k04, _r04, _sum0);
-
-                __m256 _r10 = _mm256_load_ps(r1);
-                __m256 _r11 = _mm256_load_ps(r1 + 8);
-                __m256 _r12 = _mm256_load_ps(r1 + 16);
-                __m256 _r13 = _mm256_load_ps(r1 + 24);
-                __m256 _r14 = _mm256_load_ps(r1 + 32);
-
-                __m256 _k10 = _mm256_load_ps(k0);
-                __m256 _k11 = _mm256_load_ps(k0 + 8);
-                __m256 _k12 = _mm256_load_ps(k0 + 16);
-                __m256 _k13 = _mm256_load_ps(k0 + 24);
-                __m256 _k14 = _mm256_load_ps(k0 + 32);
-                k0 += 40;
-
-                _sum0 = _mm256_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k13, _r13, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k14, _r14, _sum0);
-
-                __m256 _r20 = _mm256_load_ps(r2);
-                __m256 _r21 = _mm256_load_ps(r2 + 8);
-                __m256 _r22 = _mm256_load_ps(r2 + 16);
-                __m256 _r23 = _mm256_load_ps(r2 + 24);
-                __m256 _r24 = _mm256_load_ps(r2 + 32);
-
-                __m256 _k20 = _mm256_load_ps(k0);
-                __m256 _k21 = _mm256_load_ps(k0 + 8);
-                __m256 _k22 = _mm256_load_ps(k0 + 16);
-                __m256 _k23 = _mm256_load_ps(k0 + 24);
-                __m256 _k24 = _mm256_load_ps(k0 + 32);
-                k0 += 40;
-
-                _sum0 = _mm256_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k22, _r22, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k23, _r23, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k24, _r24, _sum0);
-
-                __m256 _r30 = _mm256_load_ps(r3);
-                __m256 _r31 = _mm256_load_ps(r3 + 8);
-                __m256 _r32 = _mm256_load_ps(r3 + 16);
-                __m256 _r33 = _mm256_load_ps(r3 + 24);
-                __m256 _r34 = _mm256_load_ps(r3 + 32);
-
-                __m256 _k30 = _mm256_load_ps(k0);
-                __m256 _k31 = _mm256_load_ps(k0 + 8);
-                __m256 _k32 = _mm256_load_ps(k0 + 16);
-                __m256 _k33 = _mm256_load_ps(k0 + 24);
-                __m256 _k34 = _mm256_load_ps(k0 + 32);
-                k0 += 40;
-
-                _sum0 = _mm256_comp_fmadd_ps(_k30, _r30, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k31, _r31, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k32, _r32, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k33, _r33, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k34, _r34, _sum0);
-
-                __m256 _r40 = _mm256_load_ps(r4);
-                __m256 _r41 = _mm256_load_ps(r4 + 8);
-                __m256 _r42 = _mm256_load_ps(r4 + 16);
-                __m256 _r43 = _mm256_load_ps(r4 + 24);
-                __m256 _r44 = _mm256_load_ps(r4 + 32);
-
-                __m256 _k40 = _mm256_load_ps(k0);
-                __m256 _k41 = _mm256_load_ps(k0 + 8);
-                __m256 _k42 = _mm256_load_ps(k0 + 16);
-                __m256 _k43 = _mm256_load_ps(k0 + 24);
-                __m256 _k44 = _mm256_load_ps(k0 + 32);
-                k0 -= 160;
-
-                _sum0 = _mm256_comp_fmadd_ps(_k40, _r40, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k41, _r41, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k42, _r42, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k43, _r43, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k44, _r44, _sum0);
-
-                _mm256_store_ps(outptr0, _sum0);
-
-                r0 += 8;
-                r1 += 8;
-                r2 += 8;
-                r3 += 8;
-                r4 += 8;
-                outptr0 += 8;
-            }
-
-            r0 += 4 * 8;
-            r1 += 4 * 8;
-            r2 += 4 * 8;
-            r3 += 4 * 8;
-            r4 += 4 * 8;
-        }
-    }
-}
-
-static void convdw5x5s2_pack8_avx(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Mat& _bias, const Option& opt)
-{
-    int w = bottom_blob.w;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int group = bottom_blob.c;
-
-    const int tailstep = (w - 2 * outw + w) * 8;
-
-    const float* bias = _bias;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int g = 0; g < group; g++)
-    {
-        Mat out = top_blob.channel(g);
-
-        __m256 _bias0 = bias ? _mm256_loadu_ps((const float*)bias + g * 8) : _mm256_setzero_ps();
-
-        const float* k0 = kernel.row(g);
-
-        float* outptr0 = out.row(0);
-
-        const Mat img0 = bottom_blob.channel(g);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        const float* r3 = img0.row(3);
-        const float* r4 = img0.row(4);
-
-        int i = 0;
-        for (; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j < outw; j++)
-            {
-                __m256 _sum0 = _bias0;
-
-                __m256 _r00 = _mm256_load_ps(r0);
-                __m256 _r01 = _mm256_load_ps(r0 + 8);
-                __m256 _r02 = _mm256_load_ps(r0 + 16);
-                __m256 _r03 = _mm256_load_ps(r0 + 24);
-                __m256 _r04 = _mm256_load_ps(r0 + 32);
-
-                __m256 _k00 = _mm256_load_ps(k0);
-                __m256 _k01 = _mm256_load_ps(k0 + 8);
-                __m256 _k02 = _mm256_load_ps(k0 + 16);
-                __m256 _k03 = _mm256_load_ps(k0 + 24);
-                __m256 _k04 = _mm256_load_ps(k0 + 32);
-                k0 += 40;
-
-                _sum0 = _mm256_comp_fmadd_ps(_k00, _r00, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k01, _r01, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k02, _r02, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k03, _r03, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k04, _r04, _sum0);
-
-                __m256 _r10 = _mm256_load_ps(r1);
-                __m256 _r11 = _mm256_load_ps(r1 + 8);
-                __m256 _r12 = _mm256_load_ps(r1 + 16);
-                __m256 _r13 = _mm256_load_ps(r1 + 24);
-                __m256 _r14 = _mm256_load_ps(r1 + 32);
-
-                __m256 _k10 = _mm256_load_ps(k0);
-                __m256 _k11 = _mm256_load_ps(k0 + 8);
-                __m256 _k12 = _mm256_load_ps(k0 + 16);
-                __m256 _k13 = _mm256_load_ps(k0 + 24);
-                __m256 _k14 = _mm256_load_ps(k0 + 32);
-                k0 += 40;
-
-                _sum0 = _mm256_comp_fmadd_ps(_k10, _r10, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k11, _r11, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k12, _r12, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k13, _r13, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k14, _r14, _sum0);
-
-                __m256 _r20 = _mm256_load_ps(r2);
-                __m256 _r21 = _mm256_load_ps(r2 + 8);
-                __m256 _r22 = _mm256_load_ps(r2 + 16);
-                __m256 _r23 = _mm256_load_ps(r2 + 24);
-                __m256 _r24 = _mm256_load_ps(r2 + 32);
-
-                __m256 _k20 = _mm256_load_ps(k0);
-                __m256 _k21 = _mm256_load_ps(k0 + 8);
-                __m256 _k22 = _mm256_load_ps(k0 + 16);
-                __m256 _k23 = _mm256_load_ps(k0 + 24);
-                __m256 _k24 = _mm256_load_ps(k0 + 32);
-                k0 += 40;
-
-                _sum0 = _mm256_comp_fmadd_ps(_k20, _r20, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k21, _r21, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k22, _r22, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k23, _r23, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k24, _r24, _sum0);
-
-                __m256 _r30 = _mm256_load_ps(r3);
-                __m256 _r31 = _mm256_load_ps(r3 + 8);
-                __m256 _r32 = _mm256_load_ps(r3 + 16);
-                __m256 _r33 = _mm256_load_ps(r3 + 24);
-                __m256 _r34 = _mm256_load_ps(r3 + 32);
-
-                __m256 _k30 = _mm256_load_ps(k0);
-                __m256 _k31 = _mm256_load_ps(k0 + 8);
-                __m256 _k32 = _mm256_load_ps(k0 + 16);
-                __m256 _k33 = _mm256_load_ps(k0 + 24);
-                __m256 _k34 = _mm256_load_ps(k0 + 32);
-                k0 += 40;
-
-                _sum0 = _mm256_comp_fmadd_ps(_k30, _r30, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k31, _r31, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k32, _r32, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k33, _r33, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k34, _r34, _sum0);
-
-                __m256 _r40 = _mm256_load_ps(r4);
-                __m256 _r41 = _mm256_load_ps(r4 + 8);
-                __m256 _r42 = _mm256_load_ps(r4 + 16);
-                __m256 _r43 = _mm256_load_ps(r4 + 24);
-                __m256 _r44 = _mm256_load_ps(r4 + 32);
-
-                __m256 _k40 = _mm256_load_ps(k0);
-                __m256 _k41 = _mm256_load_ps(k0 + 8);
-                __m256 _k42 = _mm256_load_ps(k0 + 16);
-                __m256 _k43 = _mm256_load_ps(k0 + 24);
-                __m256 _k44 = _mm256_load_ps(k0 + 32);
-                k0 -= 160;
-
-                _sum0 = _mm256_comp_fmadd_ps(_k40, _r40, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k41, _r41, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k42, _r42, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k43, _r43, _sum0);
-                _sum0 = _mm256_comp_fmadd_ps(_k44, _r44, _sum0);
-
-                _mm256_store_ps(outptr0, _sum0);
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                r3 += 16;
-                r4 += 16;
-                outptr0 += 8;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-            r3 += tailstep;
-            r4 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_x86.h
deleted file mode 100644
index 1fedb119..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/convolutiondepthwise_x86.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CONVOLUTIONDEPTHWISE_X86_H
-#define LAYER_CONVOLUTIONDEPTHWISE_X86_H
-
-#include "convolutiondepthwise.h"
-
-namespace ncnn {
-
-class ConvolutionDepthWise_x86 : public ConvolutionDepthWise
-{
-public:
-    ConvolutionDepthWise_x86();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    int create_group_ops(const Option& opt);
-#if NCNN_INT8
-    int create_pipeline_int8_x86(const Option& opt);
-    int forward_int8_x86(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Layer* activation;
-    std::vector<ncnn::Layer*> group_ops;
-
-    Mat weight_data_tm;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CONVOLUTIONDEPTHWISE_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/crop_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/crop_x86.h
deleted file mode 100644
index ba0fc1b6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/crop_x86.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_CROP_X86_H
-#define LAYER_CROP_X86_H
-
-#include "crop.h"
-
-namespace ncnn {
-
-class Crop_x86 : public Crop
-{
-public:
-    Crop_x86();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_CROP_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16.h
deleted file mode 100644
index 8f21a393..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16.h
+++ /dev/null
@@ -1,142 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack16_avx512(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        const int maxk = kernel_w * kernel_h;
-
-        // shadowed variable for less openmp task args
-        const int w = bottom_blob.w;
-        const int h = bottom_blob.h;
-        const int channels = bottom_blob.c;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m512 _sum = _mm512_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum = _mm512_loadu_ps(bias_data_ptr + p * 16);
-                }
-
-                const float* kptr = weight_data_packed.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 16;
-
-                            int k = (y * kernel_w + x) * 256;
-
-                            __m512 _val0 = _mm512_set1_ps(sptr[0]);
-                            __m512 _val1 = _mm512_set1_ps(sptr[1]);
-                            __m512 _val2 = _mm512_set1_ps(sptr[2]);
-                            __m512 _val3 = _mm512_set1_ps(sptr[3]);
-                            __m512 _val4 = _mm512_set1_ps(sptr[4]);
-                            __m512 _val5 = _mm512_set1_ps(sptr[5]);
-                            __m512 _val6 = _mm512_set1_ps(sptr[6]);
-                            __m512 _val7 = _mm512_set1_ps(sptr[7]);
-                            __m512 _val8 = _mm512_set1_ps(sptr[8]);
-                            __m512 _val9 = _mm512_set1_ps(sptr[9]);
-                            __m512 _vala = _mm512_set1_ps(sptr[10]);
-                            __m512 _valb = _mm512_set1_ps(sptr[11]);
-                            __m512 _valc = _mm512_set1_ps(sptr[12]);
-                            __m512 _vald = _mm512_set1_ps(sptr[13]);
-                            __m512 _vale = _mm512_set1_ps(sptr[14]);
-                            __m512 _valf = _mm512_set1_ps(sptr[15]);
-                            __m512 _w0 = _mm512_load_ps(kptr + k);
-                            __m512 _w1 = _mm512_load_ps(kptr + k + 16);
-                            __m512 _w2 = _mm512_load_ps(kptr + k + 16 * 2);
-                            __m512 _w3 = _mm512_load_ps(kptr + k + 16 * 3);
-                            __m512 _w4 = _mm512_load_ps(kptr + k + 16 * 4);
-                            __m512 _w5 = _mm512_load_ps(kptr + k + 16 * 5);
-                            __m512 _w6 = _mm512_load_ps(kptr + k + 16 * 6);
-                            __m512 _w7 = _mm512_load_ps(kptr + k + 16 * 7);
-                            __m512 _w8 = _mm512_load_ps(kptr + k + 16 * 8);
-                            __m512 _w9 = _mm512_load_ps(kptr + k + 16 * 9);
-                            __m512 _wa = _mm512_load_ps(kptr + k + 16 * 10);
-                            __m512 _wb = _mm512_load_ps(kptr + k + 16 * 11);
-                            __m512 _wc = _mm512_load_ps(kptr + k + 16 * 12);
-                            __m512 _wd = _mm512_load_ps(kptr + k + 16 * 13);
-                            __m512 _we = _mm512_load_ps(kptr + k + 16 * 14);
-                            __m512 _wf = _mm512_load_ps(kptr + k + 16 * 15);
-                            _sum = _mm512_fmadd_ps(_val0, _w0, _sum);
-                            _sum = _mm512_fmadd_ps(_val1, _w1, _sum);
-                            _sum = _mm512_fmadd_ps(_val2, _w2, _sum);
-                            _sum = _mm512_fmadd_ps(_val3, _w3, _sum);
-                            _sum = _mm512_fmadd_ps(_val4, _w4, _sum);
-                            _sum = _mm512_fmadd_ps(_val5, _w5, _sum);
-                            _sum = _mm512_fmadd_ps(_val6, _w6, _sum);
-                            _sum = _mm512_fmadd_ps(_val7, _w7, _sum);
-                            _sum = _mm512_fmadd_ps(_val8, _w8, _sum);
-                            _sum = _mm512_fmadd_ps(_val9, _w9, _sum);
-                            _sum = _mm512_fmadd_ps(_vala, _wa, _sum);
-                            _sum = _mm512_fmadd_ps(_valb, _wb, _sum);
-                            _sum = _mm512_fmadd_ps(_valc, _wc, _sum);
-                            _sum = _mm512_fmadd_ps(_vald, _wd, _sum);
-                            _sum = _mm512_fmadd_ps(_vale, _we, _sum);
-                            _sum = _mm512_fmadd_ps(_valf, _wf, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 256;
-                }
-
-                _sum = activation_avx512(_sum, activation_type, activation_params);
-
-                _mm512_storeu_ps(outptr, _sum);
-                outptr += 16;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16to1.h
deleted file mode 100644
index 65a1cfae..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16to1.h
+++ /dev/null
@@ -1,101 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack16to1_avx512(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        const int maxk = kernel_w * kernel_h;
-
-        // shadowed variable for less openmp task args
-        const int w = bottom_blob.w;
-        const int h = bottom_blob.h;
-        const int channels = bottom_blob.c;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                __m512 _sum = _mm512_setzero_ps();
-
-                const float* kptr = weight_data_packed.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 16;
-
-                            int k = y * kernel_w + x;
-
-                            __m512 _val = _mm512_load_ps(sptr);
-                            __m512 _w = _mm512_load_ps(kptr + k * 16);
-                            _sum = _mm512_fmadd_ps(_val, _w, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 16;
-                }
-
-                sum += _mm512_comp_reduce_add_ps(_sum);
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[0] = sum;
-                outptr++;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16to4.h
deleted file mode 100644
index ce4f7fd3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16to4.h
+++ /dev/null
@@ -1,142 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack16to4_avx512(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        const int maxk = kernel_w * kernel_h;
-
-        // shadowed variable for less openmp task args
-        const int w = bottom_blob.w;
-        const int h = bottom_blob.h;
-        const int channels = bottom_blob.c;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m128 _sum = _mm_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum = _mm_loadu_ps(bias_data_ptr + p * 4);
-                }
-
-                const float* kptr = weight_data_packed.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 16;
-
-                            int k = (y * kernel_w + x) * 64;
-
-                            __m128 _val0 = _mm_broadcast_ss(sptr);
-                            __m128 _val1 = _mm_broadcast_ss(sptr + 1);
-                            __m128 _val2 = _mm_broadcast_ss(sptr + 2);
-                            __m128 _val3 = _mm_broadcast_ss(sptr + 3);
-                            __m128 _val4 = _mm_broadcast_ss(sptr + 4);
-                            __m128 _val5 = _mm_broadcast_ss(sptr + 5);
-                            __m128 _val6 = _mm_broadcast_ss(sptr + 6);
-                            __m128 _val7 = _mm_broadcast_ss(sptr + 7);
-                            __m128 _val8 = _mm_broadcast_ss(sptr + 8);
-                            __m128 _val9 = _mm_broadcast_ss(sptr + 9);
-                            __m128 _vala = _mm_broadcast_ss(sptr + 10);
-                            __m128 _valb = _mm_broadcast_ss(sptr + 11);
-                            __m128 _valc = _mm_broadcast_ss(sptr + 12);
-                            __m128 _vald = _mm_broadcast_ss(sptr + 13);
-                            __m128 _vale = _mm_broadcast_ss(sptr + 14);
-                            __m128 _valf = _mm_broadcast_ss(sptr + 15);
-                            __m128 _w0 = _mm_load_ps(kptr + k);
-                            __m128 _w1 = _mm_load_ps(kptr + k + 4);
-                            __m128 _w2 = _mm_load_ps(kptr + k + 8);
-                            __m128 _w3 = _mm_load_ps(kptr + k + 12);
-                            __m128 _w4 = _mm_load_ps(kptr + k + 16);
-                            __m128 _w5 = _mm_load_ps(kptr + k + 20);
-                            __m128 _w6 = _mm_load_ps(kptr + k + 24);
-                            __m128 _w7 = _mm_load_ps(kptr + k + 28);
-                            __m128 _w8 = _mm_load_ps(kptr + k + 32);
-                            __m128 _w9 = _mm_load_ps(kptr + k + 36);
-                            __m128 _wa = _mm_load_ps(kptr + k + 40);
-                            __m128 _wb = _mm_load_ps(kptr + k + 44);
-                            __m128 _wc = _mm_load_ps(kptr + k + 48);
-                            __m128 _wd = _mm_load_ps(kptr + k + 52);
-                            __m128 _we = _mm_load_ps(kptr + k + 56);
-                            __m128 _wf = _mm_load_ps(kptr + k + 60);
-                            _sum = _mm_fmadd_ps(_val0, _w0, _sum);
-                            _sum = _mm_fmadd_ps(_val1, _w1, _sum);
-                            _sum = _mm_fmadd_ps(_val2, _w2, _sum);
-                            _sum = _mm_fmadd_ps(_val3, _w3, _sum);
-                            _sum = _mm_fmadd_ps(_val4, _w4, _sum);
-                            _sum = _mm_fmadd_ps(_val5, _w5, _sum);
-                            _sum = _mm_fmadd_ps(_val6, _w6, _sum);
-                            _sum = _mm_fmadd_ps(_val7, _w7, _sum);
-                            _sum = _mm_fmadd_ps(_val8, _w8, _sum);
-                            _sum = _mm_fmadd_ps(_val9, _w9, _sum);
-                            _sum = _mm_fmadd_ps(_vala, _wa, _sum);
-                            _sum = _mm_fmadd_ps(_valb, _wb, _sum);
-                            _sum = _mm_fmadd_ps(_valc, _wc, _sum);
-                            _sum = _mm_fmadd_ps(_vald, _wd, _sum);
-                            _sum = _mm_fmadd_ps(_vale, _we, _sum);
-                            _sum = _mm_fmadd_ps(_valf, _wf, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 64;
-                }
-
-                _sum = activation_sse(_sum, activation_type, activation_params);
-
-                _mm_storeu_ps(outptr, _sum);
-                outptr += 4;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16to8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16to8.h
deleted file mode 100644
index 0001f7c3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack16to8.h
+++ /dev/null
@@ -1,142 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack16to8_avx512(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        const int maxk = kernel_w * kernel_h;
-
-        // shadowed variable for less openmp task args
-        const int w = bottom_blob.w;
-        const int h = bottom_blob.h;
-        const int channels = bottom_blob.c;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m256 _sum = _mm256_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum = _mm256_loadu_ps(bias_data_ptr + p * 8);
-                }
-
-                const float* kptr = weight_data_packed.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 16;
-
-                            int k = (y * kernel_w + x) * 128;
-
-                            __m256 _val0 = _mm256_broadcast_ss(sptr);
-                            __m256 _val1 = _mm256_broadcast_ss(sptr + 1);
-                            __m256 _val2 = _mm256_broadcast_ss(sptr + 2);
-                            __m256 _val3 = _mm256_broadcast_ss(sptr + 3);
-                            __m256 _val4 = _mm256_broadcast_ss(sptr + 4);
-                            __m256 _val5 = _mm256_broadcast_ss(sptr + 5);
-                            __m256 _val6 = _mm256_broadcast_ss(sptr + 6);
-                            __m256 _val7 = _mm256_broadcast_ss(sptr + 7);
-                            __m256 _val8 = _mm256_broadcast_ss(sptr + 8);
-                            __m256 _val9 = _mm256_broadcast_ss(sptr + 9);
-                            __m256 _vala = _mm256_broadcast_ss(sptr + 10);
-                            __m256 _valb = _mm256_broadcast_ss(sptr + 11);
-                            __m256 _valc = _mm256_broadcast_ss(sptr + 12);
-                            __m256 _vald = _mm256_broadcast_ss(sptr + 13);
-                            __m256 _vale = _mm256_broadcast_ss(sptr + 14);
-                            __m256 _valf = _mm256_broadcast_ss(sptr + 15);
-                            __m256 _w0 = _mm256_load_ps(kptr + k);
-                            __m256 _w1 = _mm256_load_ps(kptr + k + 8);
-                            __m256 _w2 = _mm256_load_ps(kptr + k + 8 * 2);
-                            __m256 _w3 = _mm256_load_ps(kptr + k + 8 * 3);
-                            __m256 _w4 = _mm256_load_ps(kptr + k + 8 * 4);
-                            __m256 _w5 = _mm256_load_ps(kptr + k + 8 * 5);
-                            __m256 _w6 = _mm256_load_ps(kptr + k + 8 * 6);
-                            __m256 _w7 = _mm256_load_ps(kptr + k + 8 * 7);
-                            __m256 _w8 = _mm256_load_ps(kptr + k + 8 * 8);
-                            __m256 _w9 = _mm256_load_ps(kptr + k + 8 * 9);
-                            __m256 _wa = _mm256_load_ps(kptr + k + 8 * 10);
-                            __m256 _wb = _mm256_load_ps(kptr + k + 8 * 11);
-                            __m256 _wc = _mm256_load_ps(kptr + k + 8 * 12);
-                            __m256 _wd = _mm256_load_ps(kptr + k + 8 * 13);
-                            __m256 _we = _mm256_load_ps(kptr + k + 8 * 14);
-                            __m256 _wf = _mm256_load_ps(kptr + k + 8 * 15);
-                            _sum = _mm256_fmadd_ps(_val0, _w0, _sum);
-                            _sum = _mm256_fmadd_ps(_val1, _w1, _sum);
-                            _sum = _mm256_fmadd_ps(_val2, _w2, _sum);
-                            _sum = _mm256_fmadd_ps(_val3, _w3, _sum);
-                            _sum = _mm256_fmadd_ps(_val4, _w4, _sum);
-                            _sum = _mm256_fmadd_ps(_val5, _w5, _sum);
-                            _sum = _mm256_fmadd_ps(_val6, _w6, _sum);
-                            _sum = _mm256_fmadd_ps(_val7, _w7, _sum);
-                            _sum = _mm256_fmadd_ps(_val8, _w8, _sum);
-                            _sum = _mm256_fmadd_ps(_val9, _w9, _sum);
-                            _sum = _mm256_fmadd_ps(_vala, _wa, _sum);
-                            _sum = _mm256_fmadd_ps(_valb, _wb, _sum);
-                            _sum = _mm256_fmadd_ps(_valc, _wc, _sum);
-                            _sum = _mm256_fmadd_ps(_vald, _wd, _sum);
-                            _sum = _mm256_fmadd_ps(_vale, _we, _sum);
-                            _sum = _mm256_fmadd_ps(_valf, _wf, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 128;
-                }
-
-                _sum = activation_avx(_sum, activation_type, activation_params);
-
-                _mm256_storeu_ps(outptr, _sum);
-                outptr += 8;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack1to16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack1to16.h
deleted file mode 100644
index c357e1f0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack1to16.h
+++ /dev/null
@@ -1,99 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack1to16_avx512(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        const int maxk = kernel_w * kernel_h;
-
-        // shadowed variable for less openmp task args
-        const int w = bottom_blob.w;
-        const int h = bottom_blob.h;
-        const int channels = bottom_blob.c;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m512 _sum = _mm512_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum = _mm512_loadu_ps(bias_data_ptr + p * 16);
-                }
-
-                const float* kptr = weight_data_packed.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        const float* sptr = m.row(sy);
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            float val = sptr[sx];
-
-                            int k = y * kernel_w + x;
-
-                            __m512 _val = _mm512_set1_ps(val);
-                            __m512 _w = _mm512_load_ps(kptr + k * 16);
-                            _sum = _mm512_fmadd_ps(_val, _w, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 16;
-                }
-
-                _sum = activation_avx512(_sum, activation_type, activation_params);
-
-                _mm512_storeu_ps(outptr, _sum);
-                outptr += 16;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack1to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack1to4.h
deleted file mode 100644
index e749f1f9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack1to4.h
+++ /dev/null
@@ -1,99 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack1to4_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        const int maxk = kernel_w * kernel_h;
-
-        // shadowed variable for less openmp task args
-        const int w = bottom_blob.w;
-        const int h = bottom_blob.h;
-        const int channels = bottom_blob.c;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m128 _sum = _mm_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum = _mm_loadu_ps(bias_data_ptr + p * 4);
-                }
-
-                const float* kptr = weight_data_packed.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        const float* sptr = m.row(sy);
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            float val = sptr[sx];
-
-                            int k = y * kernel_w + x;
-
-                            __m128 _val = _mm_set1_ps(val);
-                            __m128 _w = _mm_load_ps(kptr + k * 4);
-                            _sum = _mm_comp_fmadd_ps(_val, _w, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 4;
-                }
-
-                _sum = activation_sse(_sum, activation_type, activation_params);
-
-                _mm_storeu_ps(outptr, _sum);
-                outptr += 4;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack1to8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack1to8.h
deleted file mode 100644
index 31b7a672..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack1to8.h
+++ /dev/null
@@ -1,99 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack1to8_avx(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        const int maxk = kernel_w * kernel_h;
-
-        // shadowed variable for less openmp task args
-        const int w = bottom_blob.w;
-        const int h = bottom_blob.h;
-        const int channels = bottom_blob.c;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m256 _sum = _mm256_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum = _mm256_loadu_ps(bias_data_ptr + p * 8);
-                }
-
-                const float* kptr = weight_data_packed.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        const float* sptr = m.row(sy);
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            float val = sptr[sx];
-
-                            int k = y * kernel_w + x;
-
-                            __m256 _val = _mm256_set1_ps(val);
-                            __m256 _w = _mm256_load_ps(kptr + k * 8);
-                            _sum = _mm256_comp_fmadd_ps(_val, _w, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 8;
-                }
-
-                _sum = activation_avx(_sum, activation_type, activation_params);
-
-                _mm256_storeu_ps(outptr, _sum);
-                outptr += 8;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4.h
deleted file mode 100644
index 96955e20..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4.h
+++ /dev/null
@@ -1,106 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack4_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        const int maxk = kernel_w * kernel_h;
-
-        // shadowed variable for less openmp task args
-        const int w = bottom_blob.w;
-        const int h = bottom_blob.h;
-        const int channels = bottom_blob.c;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m128 _sum = _mm_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum = _mm_loadu_ps(bias_data_ptr + p * 4);
-                }
-
-                const float* kptr = weight_data_packed.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 4;
-
-                            int k = (y * kernel_w + x) * 16;
-
-                            __m128 _val0 = _mm_load1_ps(sptr);
-                            __m128 _val1 = _mm_load1_ps(sptr + 1);
-                            __m128 _val2 = _mm_load1_ps(sptr + 2);
-                            __m128 _val3 = _mm_load1_ps(sptr + 3);
-                            __m128 _w0 = _mm_load_ps(kptr + k);
-                            __m128 _w1 = _mm_load_ps(kptr + k + 4);
-                            __m128 _w2 = _mm_load_ps(kptr + k + 8);
-                            __m128 _w3 = _mm_load_ps(kptr + k + 12);
-                            _sum = _mm_comp_fmadd_ps(_val0, _w0, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val1, _w1, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val2, _w2, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val3, _w3, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 16;
-                }
-
-                _sum = activation_sse(_sum, activation_type, activation_params);
-
-                _mm_storeu_ps(outptr, _sum);
-                outptr += 4;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4to1.h
deleted file mode 100644
index 27f89043..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4to1.h
+++ /dev/null
@@ -1,101 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack4to1_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        const int maxk = kernel_w * kernel_h;
-
-        // shadowed variable for less openmp task args
-        const int w = bottom_blob.w;
-        const int h = bottom_blob.h;
-        const int channels = bottom_blob.c;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                __m128 _sum = _mm_setzero_ps();
-
-                const float* kptr = weight_data_packed.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 4;
-
-                            int k = y * kernel_w + x;
-
-                            __m128 _val = _mm_load_ps(sptr);
-                            __m128 _w = _mm_load_ps(kptr + k * 4);
-                            _sum = _mm_comp_fmadd_ps(_val, _w, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 4;
-                }
-
-                sum += _mm_reduce_add_ps(_sum);
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[0] = sum;
-                outptr++;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4to16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4to16.h
deleted file mode 100644
index 4510c28a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4to16.h
+++ /dev/null
@@ -1,106 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack4to16_avx512(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        const int maxk = kernel_w * kernel_h;
-
-        // shadowed variable for less openmp task args
-        const int w = bottom_blob.w;
-        const int h = bottom_blob.h;
-        const int channels = bottom_blob.c;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m512 _sum = _mm512_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum = _mm512_loadu_ps(bias_data_ptr + p * 16);
-                }
-
-                const float* kptr = weight_data_packed.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 4;
-
-                            int k = (y * kernel_w + x) * 64;
-
-                            __m512 _val0 = _mm512_set1_ps(sptr[0]);
-                            __m512 _val1 = _mm512_set1_ps(sptr[1]);
-                            __m512 _val2 = _mm512_set1_ps(sptr[2]);
-                            __m512 _val3 = _mm512_set1_ps(sptr[3]);
-                            __m512 _w0 = _mm512_load_ps(kptr + k);
-                            __m512 _w1 = _mm512_load_ps(kptr + k + 16);
-                            __m512 _w2 = _mm512_load_ps(kptr + k + 32);
-                            __m512 _w3 = _mm512_load_ps(kptr + k + 48);
-                            _sum = _mm512_fmadd_ps(_val0, _w0, _sum);
-                            _sum = _mm512_fmadd_ps(_val1, _w1, _sum);
-                            _sum = _mm512_fmadd_ps(_val2, _w2, _sum);
-                            _sum = _mm512_fmadd_ps(_val3, _w3, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 64;
-                }
-
-                _sum = activation_avx512(_sum, activation_type, activation_params);
-
-                _mm512_storeu_ps(outptr, _sum);
-                outptr += 16;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4to8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4to8.h
deleted file mode 100644
index 1ca09527..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack4to8.h
+++ /dev/null
@@ -1,106 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack4to8_avx(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        const int maxk = kernel_w * kernel_h;
-
-        // shadowed variable for less openmp task args
-        const int w = bottom_blob.w;
-        const int h = bottom_blob.h;
-        const int channels = bottom_blob.c;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m256 _sum = _mm256_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum = _mm256_loadu_ps(bias_data_ptr + p * 8);
-                }
-
-                const float* kptr = weight_data_packed.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 4;
-
-                            int k = (y * kernel_w + x) * 32;
-
-                            __m256 _val0 = _mm256_broadcast_ss(sptr);
-                            __m256 _val1 = _mm256_broadcast_ss(sptr + 1);
-                            __m256 _val2 = _mm256_broadcast_ss(sptr + 2);
-                            __m256 _val3 = _mm256_broadcast_ss(sptr + 3);
-                            __m256 _w0 = _mm256_load_ps(kptr + k);
-                            __m256 _w1 = _mm256_load_ps(kptr + k + 8);
-                            __m256 _w2 = _mm256_load_ps(kptr + k + 16);
-                            __m256 _w3 = _mm256_load_ps(kptr + k + 24);
-                            _sum = _mm256_comp_fmadd_ps(_val0, _w0, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val1, _w1, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val2, _w2, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val3, _w3, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 32;
-                }
-
-                _sum = activation_avx(_sum, activation_type, activation_params);
-
-                _mm256_storeu_ps(outptr, _sum);
-                outptr += 8;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8.h
deleted file mode 100644
index c0ec798e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8.h
+++ /dev/null
@@ -1,118 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack8_avx(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        const int maxk = kernel_w * kernel_h;
-
-        // shadowed variable for less openmp task args
-        const int w = bottom_blob.w;
-        const int h = bottom_blob.h;
-        const int channels = bottom_blob.c;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m256 _sum = _mm256_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum = _mm256_loadu_ps(bias_data_ptr + p * 8);
-                }
-
-                const float* kptr = weight_data_packed.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 8;
-
-                            int k = (y * kernel_w + x) * 64;
-
-                            __m256 _val0 = _mm256_broadcast_ss(sptr);
-                            __m256 _val1 = _mm256_broadcast_ss(sptr + 1);
-                            __m256 _val2 = _mm256_broadcast_ss(sptr + 2);
-                            __m256 _val3 = _mm256_broadcast_ss(sptr + 3);
-                            __m256 _val4 = _mm256_broadcast_ss(sptr + 4);
-                            __m256 _val5 = _mm256_broadcast_ss(sptr + 5);
-                            __m256 _val6 = _mm256_broadcast_ss(sptr + 6);
-                            __m256 _val7 = _mm256_broadcast_ss(sptr + 7);
-                            __m256 _w0 = _mm256_load_ps(kptr + k);
-                            __m256 _w1 = _mm256_load_ps(kptr + k + 8);
-                            __m256 _w2 = _mm256_load_ps(kptr + k + 16);
-                            __m256 _w3 = _mm256_load_ps(kptr + k + 24);
-                            __m256 _w4 = _mm256_load_ps(kptr + k + 32);
-                            __m256 _w5 = _mm256_load_ps(kptr + k + 40);
-                            __m256 _w6 = _mm256_load_ps(kptr + k + 48);
-                            __m256 _w7 = _mm256_load_ps(kptr + k + 56);
-                            _sum = _mm256_comp_fmadd_ps(_val0, _w0, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val1, _w1, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val2, _w2, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val3, _w3, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val4, _w4, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val5, _w5, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val6, _w6, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val7, _w7, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 64;
-                }
-
-                _sum = activation_avx(_sum, activation_type, activation_params);
-
-                _mm256_storeu_ps(outptr, _sum);
-                outptr += 8;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8to1.h
deleted file mode 100644
index 4d40bb42..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8to1.h
+++ /dev/null
@@ -1,101 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack8to1_avx(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        const int maxk = kernel_w * kernel_h;
-
-        // shadowed variable for less openmp task args
-        const int w = bottom_blob.w;
-        const int h = bottom_blob.h;
-        const int channels = bottom_blob.c;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                __m256 _sum = _mm256_setzero_ps();
-
-                const float* kptr = weight_data_packed.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 8;
-
-                            int k = y * kernel_w + x;
-
-                            __m256 _val = _mm256_load_ps(sptr);
-                            __m256 _w = _mm256_load_ps(kptr + k * 8);
-                            _sum = _mm256_comp_fmadd_ps(_val, _w, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 8;
-                }
-
-                sum += _mm256_reduce_add_ps(_sum);
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[0] = sum;
-                outptr++;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8to16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8to16.h
deleted file mode 100644
index 86f0d726..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8to16.h
+++ /dev/null
@@ -1,118 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack8to16_avx512(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        const int maxk = kernel_w * kernel_h;
-
-        // shadowed variable for less openmp task args
-        const int w = bottom_blob.w;
-        const int h = bottom_blob.h;
-        const int channels = bottom_blob.c;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m512 _sum = _mm512_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum = _mm512_loadu_ps(bias_data_ptr + p * 16);
-                }
-
-                const float* kptr = weight_data_packed.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 8;
-
-                            int k = (y * kernel_w + x) * 128;
-
-                            __m512 _val0 = _mm512_set1_ps(sptr[0]);
-                            __m512 _val1 = _mm512_set1_ps(sptr[1]);
-                            __m512 _val2 = _mm512_set1_ps(sptr[2]);
-                            __m512 _val3 = _mm512_set1_ps(sptr[3]);
-                            __m512 _val4 = _mm512_set1_ps(sptr[4]);
-                            __m512 _val5 = _mm512_set1_ps(sptr[5]);
-                            __m512 _val6 = _mm512_set1_ps(sptr[6]);
-                            __m512 _val7 = _mm512_set1_ps(sptr[7]);
-                            __m512 _w0 = _mm512_load_ps(kptr + k);
-                            __m512 _w1 = _mm512_load_ps(kptr + k + 16);
-                            __m512 _w2 = _mm512_load_ps(kptr + k + 16 * 2);
-                            __m512 _w3 = _mm512_load_ps(kptr + k + 16 * 3);
-                            __m512 _w4 = _mm512_load_ps(kptr + k + 16 * 4);
-                            __m512 _w5 = _mm512_load_ps(kptr + k + 16 * 5);
-                            __m512 _w6 = _mm512_load_ps(kptr + k + 16 * 6);
-                            __m512 _w7 = _mm512_load_ps(kptr + k + 16 * 7);
-                            _sum = _mm512_fmadd_ps(_val0, _w0, _sum);
-                            _sum = _mm512_fmadd_ps(_val1, _w1, _sum);
-                            _sum = _mm512_fmadd_ps(_val2, _w2, _sum);
-                            _sum = _mm512_fmadd_ps(_val3, _w3, _sum);
-                            _sum = _mm512_fmadd_ps(_val4, _w4, _sum);
-                            _sum = _mm512_fmadd_ps(_val5, _w5, _sum);
-                            _sum = _mm512_fmadd_ps(_val6, _w6, _sum);
-                            _sum = _mm512_fmadd_ps(_val7, _w7, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 128;
-                }
-
-                _sum = activation_avx512(_sum, activation_type, activation_params);
-
-                _mm512_storeu_ps(outptr, _sum);
-                outptr += 16;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8to4.h
deleted file mode 100644
index fea7a87b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_pack8to4.h
+++ /dev/null
@@ -1,118 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deconvolution_pack8to4_avx(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    int outch = top_blob.c;
-
-    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
-    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
-
-    const float* bias_data_ptr = bias_data;
-
-    // num_output
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int p = 0; p < outch; p++)
-    {
-        float* outptr = top_blob.channel(p);
-
-        const int maxk = kernel_w * kernel_h;
-
-        // shadowed variable for less openmp task args
-        const int w = bottom_blob.w;
-        const int h = bottom_blob.h;
-        const int channels = bottom_blob.c;
-        const int outw = top_blob.w;
-        const int outh = top_blob.h;
-
-        for (int i = 0; i < outh; i++)
-        {
-            for (int j = 0; j < outw; j++)
-            {
-                __m128 _sum = _mm_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum = _mm_loadu_ps(bias_data_ptr + p * 4);
-                }
-
-                const float* kptr = weight_data_packed.channel(p);
-
-                // channels
-                for (int q = 0; q < channels; q++)
-                {
-                    const Mat m = bottom_blob.channel(q);
-
-                    for (int y = 0; y < kernel_h; y++)
-                    {
-                        int sys = (i + y * dilation_h - (kernel_extent_h - 1));
-                        if (sys < 0 || sys % stride_h != 0)
-                            continue;
-
-                        int sy = sys / stride_h;
-                        if (sy >= h)
-                            continue;
-
-                        for (int x = 0; x < kernel_w; x++)
-                        {
-                            int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
-                            if (sxs < 0 || sxs % stride_w != 0)
-                                continue;
-
-                            int sx = sxs / stride_w;
-                            if (sx >= w)
-                                continue;
-
-                            const float* sptr = m.row(sy) + sx * 8;
-
-                            int k = (y * kernel_w + x) * 32;
-
-                            __m128 _val0 = _mm_broadcast_ss(sptr);
-                            __m128 _val1 = _mm_broadcast_ss(sptr + 1);
-                            __m128 _val2 = _mm_broadcast_ss(sptr + 2);
-                            __m128 _val3 = _mm_broadcast_ss(sptr + 3);
-                            __m128 _val4 = _mm_broadcast_ss(sptr + 4);
-                            __m128 _val5 = _mm_broadcast_ss(sptr + 5);
-                            __m128 _val6 = _mm_broadcast_ss(sptr + 6);
-                            __m128 _val7 = _mm_broadcast_ss(sptr + 7);
-                            __m128 _w0 = _mm_load_ps(kptr + k);
-                            __m128 _w1 = _mm_load_ps(kptr + k + 4);
-                            __m128 _w2 = _mm_load_ps(kptr + k + 8);
-                            __m128 _w3 = _mm_load_ps(kptr + k + 12);
-                            __m128 _w4 = _mm_load_ps(kptr + k + 16);
-                            __m128 _w5 = _mm_load_ps(kptr + k + 20);
-                            __m128 _w6 = _mm_load_ps(kptr + k + 24);
-                            __m128 _w7 = _mm_load_ps(kptr + k + 28);
-                            _sum = _mm_comp_fmadd_ps(_val0, _w0, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val1, _w1, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val2, _w2, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val3, _w3, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val4, _w4, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val5, _w5, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val6, _w6, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val7, _w7, _sum);
-                        }
-                    }
-
-                    kptr += maxk * 32;
-                }
-
-                _sum = activation_sse(_sum, activation_type, activation_params);
-
-                _mm_storeu_ps(outptr, _sum);
-                outptr += 4;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_x86.h
deleted file mode 100644
index 66c23eef..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolution_x86.h
+++ /dev/null
@@ -1,43 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTION_X86_H
-#define LAYER_DECONVOLUTION_X86_H
-
-#include "deconvolution.h"
-
-namespace ncnn {
-
-class Deconvolution_x86 : public Deconvolution
-{
-public:
-    Deconvolution_x86();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    Layer* activation;
-    Layer* gemm;
-
-    Mat weight_data_tm;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTION_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolutiondepthwise_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolutiondepthwise_x86.h
deleted file mode 100644
index 9c9e54cc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deconvolutiondepthwise_x86.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DECONVOLUTIONDEPTHWISE_X86_H
-#define LAYER_DECONVOLUTIONDEPTHWISE_X86_H
-
-#include "deconvolutiondepthwise.h"
-
-namespace ncnn {
-
-class DeconvolutionDepthWise_x86 : public DeconvolutionDepthWise
-{
-public:
-    DeconvolutionDepthWise_x86();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-protected:
-    int create_group_ops(const Option& opt);
-
-public:
-    std::vector<ncnn::Layer*> group_ops;
-
-    Mat weight_data_tm;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DECONVOLUTIONDEPTHWISE_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16.h
deleted file mode 100644
index f15d40f3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16.h
+++ /dev/null
@@ -1,435 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deformableconv2d_pack16_avx512(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const Mat& bottom_blob = bottom_blobs[0];
-    const Mat& offset = bottom_blobs[1];
-    const bool has_mask = (bottom_blobs.size() == 3);
-    const bool offset_not_pack = offset.elempack == 1;
-    const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const int size = outw * outh;
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-    const int elempack = 16;
-    const int out_elempack = 16;
-    const int wstep = out_elempack * elempack;
-    const float zeros[out_elempack] = {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
-    const float* zeros_ptr = zeros;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int h_col = 0; h_col < outh; h_col++)
-    {
-        for (int w_col = 0; w_col < outw; w_col++)
-        {
-            int h_in = h_col * stride_h - pad_top;
-            int w_in = w_col * stride_w - pad_left;
-            for (int oc = 0; oc < outch; oc++)
-            {
-                const float* kptr = weight_data_packed.channel(oc);
-                float* outptr = top_blob.channel(oc);
-                __m512 _sum = _mm512_setzero_ps();
-                if (bias_data_ptr)
-                    _sum = _mm512_loadu_ps(bias_data_ptr + oc * out_elempack);
-                for (int i = 0; i < kernel_h; i++)
-                {
-                    for (int j = 0; j < kernel_w; j++)
-                    {
-                        float offset_h = 0.f;
-                        float offset_w = 0.f;
-                        float mask_ = 1.f;
-                        if (offset_not_pack)
-                        {
-                            offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
-                            offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
-                        }
-                        else
-                        {
-                            const int y_c = (i * kernel_w + j) * 2;
-                            const int x_c = (i * kernel_w + j) * 2 + 1;
-                            offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
-                            offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
-                        }
-                        if (has_mask)
-                        {
-                            const Mat& mask = bottom_blobs[2];
-                            if (mask_not_pack)
-                            {
-                                mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
-                            }
-                            else
-                            {
-                                const int m_c = i * kernel_w + j;
-                                mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
-                            }
-                        }
-                        const float h_im = h_in + i * dilation_h + offset_h;
-                        const float w_im = w_in + j * dilation_w + offset_w;
-
-                        // Bilinear
-                        const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
-                        float w1 = 0.f;
-                        float w2 = 0.f;
-                        float w3 = 0.f;
-                        float w4 = 0.f;
-                        bool v1_cond = false;
-                        bool v2_cond = false;
-                        bool v3_cond = false;
-                        bool v4_cond = false;
-                        int v1_pos = 0;
-                        int v2_pos = 0;
-                        int v3_pos = 0;
-                        int v4_pos = 0;
-                        if (cond)
-                        {
-                            int h_low = (int)floorf(h_im);
-                            int w_low = (int)floorf(w_im);
-                            int h_high = h_low + 1;
-                            int w_high = w_low + 1;
-
-                            float lh = h_im - h_low;
-                            float lw = w_im - w_low;
-                            float hh = 1 - lh;
-                            float hw = 1 - lw;
-
-                            v1_cond = (h_low >= 0 && w_low >= 0);
-                            v2_cond = (h_low >= 0 && w_high <= w - 1);
-                            v3_cond = (h_high <= h - 1 && w_low >= 0);
-                            v4_cond = (h_high <= h - 1 && w_high <= w - 1);
-                            if (v1_cond)
-                                v1_pos = h_low * w + w_low;
-                            if (v2_cond)
-                                v2_pos = h_low * w + w_high;
-                            if (v3_cond)
-                                v3_pos = h_high * w + w_low;
-                            if (v4_cond)
-                                v4_pos = h_high * w + w_high;
-
-                            w1 = hh * hw;
-                            w2 = hh * lw;
-                            w3 = lh * hw;
-                            w4 = lh * lw;
-                        }
-                        const float w1s[out_elempack] = {w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1};
-                        const float* w1_ptr = w1s;
-                        const float w2s[out_elempack] = {w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2};
-                        const float* w2_ptr = w2s;
-                        const float w3s[out_elempack] = {w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3};
-                        const float* w3_ptr = w3s;
-                        const float w4s[out_elempack] = {w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4};
-                        const float* w4_ptr = w4s;
-                        const float masks[out_elempack] = {mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_};
-                        const float* mask_ptr = masks;
-
-                        for (int ic = 0; ic < inch; ic++)
-                        {
-                            const float* data_im_ptr = bottom_blob.channel(ic);
-                            __m512 _val_channel0 = _mm512_loadu_ps(zeros_ptr);
-                            __m512 _val_channel1 = _val_channel0;
-                            __m512 _val_channel2 = _val_channel0;
-                            __m512 _val_channel3 = _val_channel0;
-                            __m512 _val_channel4 = _val_channel0;
-                            __m512 _val_channel5 = _val_channel0;
-                            __m512 _val_channel6 = _val_channel0;
-                            __m512 _val_channel7 = _val_channel0;
-                            __m512 _val_channel8 = _val_channel0;
-                            __m512 _val_channel9 = _val_channel0;
-                            __m512 _val_channela = _val_channel0;
-                            __m512 _val_channelb = _val_channel0;
-                            __m512 _val_channelc = _val_channel0;
-                            __m512 _val_channeld = _val_channel0;
-                            __m512 _val_channele = _val_channel0;
-                            __m512 _val_channelf = _val_channel0;
-                            if (cond)
-                            {
-                                __m512 _v1_channel0 = _val_channel0;
-                                __m512 _v1_channel1 = _val_channel0;
-                                __m512 _v1_channel2 = _val_channel0;
-                                __m512 _v1_channel3 = _val_channel0;
-                                __m512 _v1_channel4 = _val_channel0;
-                                __m512 _v1_channel5 = _val_channel0;
-                                __m512 _v1_channel6 = _val_channel0;
-                                __m512 _v1_channel7 = _val_channel0;
-                                __m512 _v1_channel8 = _val_channel0;
-                                __m512 _v1_channel9 = _val_channel0;
-                                __m512 _v1_channela = _val_channel0;
-                                __m512 _v1_channelb = _val_channel0;
-                                __m512 _v1_channelc = _val_channel0;
-                                __m512 _v1_channeld = _val_channel0;
-                                __m512 _v1_channele = _val_channel0;
-                                __m512 _v1_channelf = _val_channel0;
-                                __m512 _v2_channel0 = _val_channel0;
-                                __m512 _v2_channel1 = _val_channel0;
-                                __m512 _v2_channel2 = _val_channel0;
-                                __m512 _v2_channel3 = _val_channel0;
-                                __m512 _v2_channel4 = _val_channel0;
-                                __m512 _v2_channel5 = _val_channel0;
-                                __m512 _v2_channel6 = _val_channel0;
-                                __m512 _v2_channel7 = _val_channel0;
-                                __m512 _v2_channel8 = _val_channel0;
-                                __m512 _v2_channel9 = _val_channel0;
-                                __m512 _v2_channela = _val_channel0;
-                                __m512 _v2_channelb = _val_channel0;
-                                __m512 _v2_channelc = _val_channel0;
-                                __m512 _v2_channeld = _val_channel0;
-                                __m512 _v2_channele = _val_channel0;
-                                __m512 _v2_channelf = _val_channel0;
-                                __m512 _v3_channel0 = _val_channel0;
-                                __m512 _v3_channel1 = _val_channel0;
-                                __m512 _v3_channel2 = _val_channel0;
-                                __m512 _v3_channel3 = _val_channel0;
-                                __m512 _v3_channel4 = _val_channel0;
-                                __m512 _v3_channel5 = _val_channel0;
-                                __m512 _v3_channel6 = _val_channel0;
-                                __m512 _v3_channel7 = _val_channel0;
-                                __m512 _v3_channel8 = _val_channel0;
-                                __m512 _v3_channel9 = _val_channel0;
-                                __m512 _v3_channela = _val_channel0;
-                                __m512 _v3_channelb = _val_channel0;
-                                __m512 _v3_channelc = _val_channel0;
-                                __m512 _v3_channeld = _val_channel0;
-                                __m512 _v3_channele = _val_channel0;
-                                __m512 _v3_channelf = _val_channel0;
-                                __m512 _v4_channel0 = _val_channel0;
-                                __m512 _v4_channel1 = _val_channel0;
-                                __m512 _v4_channel2 = _val_channel0;
-                                __m512 _v4_channel3 = _val_channel0;
-                                __m512 _v4_channel4 = _val_channel0;
-                                __m512 _v4_channel5 = _val_channel0;
-                                __m512 _v4_channel6 = _val_channel0;
-                                __m512 _v4_channel7 = _val_channel0;
-                                __m512 _v4_channel8 = _val_channel0;
-                                __m512 _v4_channel9 = _val_channel0;
-                                __m512 _v4_channela = _val_channel0;
-                                __m512 _v4_channelb = _val_channel0;
-                                __m512 _v4_channelc = _val_channel0;
-                                __m512 _v4_channeld = _val_channel0;
-                                __m512 _v4_channele = _val_channel0;
-                                __m512 _v4_channelf = _val_channel0;
-                                if (v1_cond)
-                                {
-                                    _v1_channel0 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack]);
-                                    _v1_channel1 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 1]);
-                                    _v1_channel2 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 2]);
-                                    _v1_channel3 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 3]);
-                                    _v1_channel4 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 4]);
-                                    _v1_channel5 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 5]);
-                                    _v1_channel6 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 6]);
-                                    _v1_channel7 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 7]);
-                                    _v1_channel8 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 8]);
-                                    _v1_channel9 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 9]);
-                                    _v1_channela = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 10]);
-                                    _v1_channelb = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 11]);
-                                    _v1_channelc = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 12]);
-                                    _v1_channeld = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 13]);
-                                    _v1_channele = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 14]);
-                                    _v1_channelf = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 15]);
-                                }
-                                if (v2_cond)
-                                {
-                                    _v2_channel0 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack]);
-                                    _v2_channel1 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 1]);
-                                    _v2_channel2 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 2]);
-                                    _v2_channel3 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 3]);
-                                    _v2_channel4 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 4]);
-                                    _v2_channel5 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 5]);
-                                    _v2_channel6 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 6]);
-                                    _v2_channel7 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 7]);
-                                    _v2_channel8 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 8]);
-                                    _v2_channel9 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 9]);
-                                    _v2_channela = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 10]);
-                                    _v2_channelb = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 11]);
-                                    _v2_channelc = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 12]);
-                                    _v2_channeld = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 13]);
-                                    _v2_channele = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 14]);
-                                    _v2_channelf = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 15]);
-                                }
-                                if (v3_cond)
-                                {
-                                    _v3_channel0 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack]);
-                                    _v3_channel1 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 1]);
-                                    _v3_channel2 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 2]);
-                                    _v3_channel3 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 3]);
-                                    _v3_channel4 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 4]);
-                                    _v3_channel5 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 5]);
-                                    _v3_channel6 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 6]);
-                                    _v3_channel7 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 7]);
-                                    _v3_channel8 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 8]);
-                                    _v3_channel9 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 9]);
-                                    _v3_channela = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 10]);
-                                    _v3_channelb = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 11]);
-                                    _v3_channelc = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 12]);
-                                    _v3_channeld = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 13]);
-                                    _v3_channele = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 14]);
-                                    _v3_channelf = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 15]);
-                                }
-                                if (v4_cond)
-                                {
-                                    _v4_channel0 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack]);
-                                    _v4_channel1 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 1]);
-                                    _v4_channel2 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 2]);
-                                    _v4_channel3 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 3]);
-                                    _v4_channel4 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 4]);
-                                    _v4_channel5 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 5]);
-                                    _v4_channel6 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 6]);
-                                    _v4_channel7 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 7]);
-                                    _v4_channel8 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 8]);
-                                    _v4_channel9 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 9]);
-                                    _v4_channela = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 10]);
-                                    _v4_channelb = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 11]);
-                                    _v4_channelc = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 12]);
-                                    _v4_channeld = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 13]);
-                                    _v4_channele = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 14]);
-                                    _v4_channelf = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 15]);
-                                }
-                                __m512 _w1 = _mm512_loadu_ps(w1_ptr);
-                                __m512 _w2 = _mm512_loadu_ps(w2_ptr);
-                                __m512 _w3 = _mm512_loadu_ps(w3_ptr);
-                                __m512 _w4 = _mm512_loadu_ps(w4_ptr);
-                                _val_channel0 = _mm512_fmadd_ps(_v1_channel0, _w1, _val_channel0);
-                                _val_channel0 = _mm512_fmadd_ps(_v2_channel0, _w2, _val_channel0);
-                                _val_channel0 = _mm512_fmadd_ps(_v3_channel0, _w3, _val_channel0);
-                                _val_channel0 = _mm512_fmadd_ps(_v4_channel0, _w4, _val_channel0);
-                                _val_channel1 = _mm512_fmadd_ps(_v1_channel1, _w1, _val_channel1);
-                                _val_channel1 = _mm512_fmadd_ps(_v2_channel1, _w2, _val_channel1);
-                                _val_channel1 = _mm512_fmadd_ps(_v3_channel1, _w3, _val_channel1);
-                                _val_channel1 = _mm512_fmadd_ps(_v4_channel1, _w4, _val_channel1);
-                                _val_channel2 = _mm512_fmadd_ps(_v1_channel2, _w1, _val_channel2);
-                                _val_channel2 = _mm512_fmadd_ps(_v2_channel2, _w2, _val_channel2);
-                                _val_channel2 = _mm512_fmadd_ps(_v3_channel2, _w3, _val_channel2);
-                                _val_channel2 = _mm512_fmadd_ps(_v4_channel2, _w4, _val_channel2);
-                                _val_channel3 = _mm512_fmadd_ps(_v1_channel3, _w1, _val_channel3);
-                                _val_channel3 = _mm512_fmadd_ps(_v2_channel3, _w2, _val_channel3);
-                                _val_channel3 = _mm512_fmadd_ps(_v3_channel3, _w3, _val_channel3);
-                                _val_channel3 = _mm512_fmadd_ps(_v4_channel3, _w4, _val_channel3);
-                                _val_channel4 = _mm512_fmadd_ps(_v1_channel4, _w1, _val_channel4);
-                                _val_channel4 = _mm512_fmadd_ps(_v2_channel4, _w2, _val_channel4);
-                                _val_channel4 = _mm512_fmadd_ps(_v3_channel4, _w3, _val_channel4);
-                                _val_channel4 = _mm512_fmadd_ps(_v4_channel4, _w4, _val_channel4);
-                                _val_channel5 = _mm512_fmadd_ps(_v1_channel5, _w1, _val_channel5);
-                                _val_channel5 = _mm512_fmadd_ps(_v2_channel5, _w2, _val_channel5);
-                                _val_channel5 = _mm512_fmadd_ps(_v3_channel5, _w3, _val_channel5);
-                                _val_channel5 = _mm512_fmadd_ps(_v4_channel5, _w4, _val_channel5);
-                                _val_channel6 = _mm512_fmadd_ps(_v1_channel6, _w1, _val_channel6);
-                                _val_channel6 = _mm512_fmadd_ps(_v2_channel6, _w2, _val_channel6);
-                                _val_channel6 = _mm512_fmadd_ps(_v3_channel6, _w3, _val_channel6);
-                                _val_channel6 = _mm512_fmadd_ps(_v4_channel6, _w4, _val_channel6);
-                                _val_channel7 = _mm512_fmadd_ps(_v1_channel7, _w1, _val_channel7);
-                                _val_channel7 = _mm512_fmadd_ps(_v2_channel7, _w2, _val_channel7);
-                                _val_channel7 = _mm512_fmadd_ps(_v3_channel7, _w3, _val_channel7);
-                                _val_channel7 = _mm512_fmadd_ps(_v4_channel7, _w4, _val_channel7);
-                                _val_channel8 = _mm512_fmadd_ps(_v1_channel8, _w1, _val_channel8);
-                                _val_channel8 = _mm512_fmadd_ps(_v2_channel8, _w2, _val_channel8);
-                                _val_channel8 = _mm512_fmadd_ps(_v3_channel8, _w3, _val_channel8);
-                                _val_channel8 = _mm512_fmadd_ps(_v4_channel8, _w4, _val_channel8);
-                                _val_channel9 = _mm512_fmadd_ps(_v1_channel9, _w1, _val_channel9);
-                                _val_channel9 = _mm512_fmadd_ps(_v2_channel9, _w2, _val_channel9);
-                                _val_channel9 = _mm512_fmadd_ps(_v3_channel9, _w3, _val_channel9);
-                                _val_channel9 = _mm512_fmadd_ps(_v4_channel9, _w4, _val_channel9);
-                                _val_channela = _mm512_fmadd_ps(_v1_channela, _w1, _val_channela);
-                                _val_channela = _mm512_fmadd_ps(_v2_channela, _w2, _val_channela);
-                                _val_channela = _mm512_fmadd_ps(_v3_channela, _w3, _val_channela);
-                                _val_channela = _mm512_fmadd_ps(_v4_channela, _w4, _val_channela);
-                                _val_channelb = _mm512_fmadd_ps(_v1_channelb, _w1, _val_channelb);
-                                _val_channelb = _mm512_fmadd_ps(_v2_channelb, _w2, _val_channelb);
-                                _val_channelb = _mm512_fmadd_ps(_v3_channelb, _w3, _val_channelb);
-                                _val_channelb = _mm512_fmadd_ps(_v4_channelb, _w4, _val_channelb);
-                                _val_channelc = _mm512_fmadd_ps(_v1_channelc, _w1, _val_channelc);
-                                _val_channelc = _mm512_fmadd_ps(_v2_channelc, _w2, _val_channelc);
-                                _val_channelc = _mm512_fmadd_ps(_v3_channelc, _w3, _val_channelc);
-                                _val_channelc = _mm512_fmadd_ps(_v4_channelc, _w4, _val_channelc);
-                                _val_channeld = _mm512_fmadd_ps(_v1_channeld, _w1, _val_channeld);
-                                _val_channeld = _mm512_fmadd_ps(_v2_channeld, _w2, _val_channeld);
-                                _val_channeld = _mm512_fmadd_ps(_v3_channeld, _w3, _val_channeld);
-                                _val_channeld = _mm512_fmadd_ps(_v4_channeld, _w4, _val_channeld);
-                                _val_channele = _mm512_fmadd_ps(_v1_channele, _w1, _val_channele);
-                                _val_channele = _mm512_fmadd_ps(_v2_channele, _w2, _val_channele);
-                                _val_channele = _mm512_fmadd_ps(_v3_channele, _w3, _val_channele);
-                                _val_channele = _mm512_fmadd_ps(_v4_channele, _w4, _val_channele);
-                                _val_channelf = _mm512_fmadd_ps(_v1_channelf, _w1, _val_channelf);
-                                _val_channelf = _mm512_fmadd_ps(_v2_channelf, _w2, _val_channelf);
-                                _val_channelf = _mm512_fmadd_ps(_v3_channelf, _w3, _val_channelf);
-                                _val_channelf = _mm512_fmadd_ps(_v4_channelf, _w4, _val_channelf);
-                            }
-                            if (has_mask)
-                            {
-                                __m512 _mask = _mm512_loadu_ps(mask_ptr);
-                                _val_channel0 = _mm512_mul_ps(_val_channel0, _mask);
-                                _val_channel1 = _mm512_mul_ps(_val_channel1, _mask);
-                                _val_channel2 = _mm512_mul_ps(_val_channel2, _mask);
-                                _val_channel3 = _mm512_mul_ps(_val_channel3, _mask);
-                                _val_channel4 = _mm512_mul_ps(_val_channel4, _mask);
-                                _val_channel5 = _mm512_mul_ps(_val_channel5, _mask);
-                                _val_channel6 = _mm512_mul_ps(_val_channel6, _mask);
-                                _val_channel7 = _mm512_mul_ps(_val_channel7, _mask);
-                                _val_channel8 = _mm512_mul_ps(_val_channel8, _mask);
-                                _val_channel9 = _mm512_mul_ps(_val_channel9, _mask);
-                                _val_channela = _mm512_mul_ps(_val_channela, _mask);
-                                _val_channelb = _mm512_mul_ps(_val_channelb, _mask);
-                                _val_channelc = _mm512_mul_ps(_val_channelc, _mask);
-                                _val_channeld = _mm512_mul_ps(_val_channeld, _mask);
-                                _val_channele = _mm512_mul_ps(_val_channele, _mask);
-                                _val_channelf = _mm512_mul_ps(_val_channelf, _mask);
-                            }
-                            __m512 _conv_w0 = _mm512_load_ps(kptr);
-                            __m512 _conv_w1 = _mm512_load_ps(kptr + out_elempack); // 1 * out_elempack
-                            _sum = _mm512_fmadd_ps(_val_channel0, _conv_w0, _sum);
-                            _sum = _mm512_fmadd_ps(_val_channel1, _conv_w1, _sum);
-                            __m512 _conv_w2 = _mm512_load_ps(kptr + 32); // 2 * out_elempack
-                            __m512 _conv_w3 = _mm512_load_ps(kptr + 48); // 3 * out_elempack
-                            _sum = _mm512_fmadd_ps(_val_channel2, _conv_w2, _sum);
-                            _sum = _mm512_fmadd_ps(_val_channel3, _conv_w3, _sum);
-                            __m512 _conv_w4 = _mm512_load_ps(kptr + 64); // 4 * out_elempack
-                            __m512 _conv_w5 = _mm512_load_ps(kptr + 80); // 5 * out_elempack
-                            _sum = _mm512_fmadd_ps(_val_channel4, _conv_w4, _sum);
-                            _sum = _mm512_fmadd_ps(_val_channel5, _conv_w5, _sum);
-                            __m512 _conv_w6 = _mm512_load_ps(kptr + 96);  // 6 * out_elempack
-                            __m512 _conv_w7 = _mm512_load_ps(kptr + 112); // 7 * out_elempack
-                            _sum = _mm512_fmadd_ps(_val_channel6, _conv_w6, _sum);
-                            _sum = _mm512_fmadd_ps(_val_channel7, _conv_w7, _sum);
-                            __m512 _conv_w8 = _mm512_load_ps(kptr + 128); // 8 * out_elempack
-                            __m512 _conv_w9 = _mm512_load_ps(kptr + 144); // 9 * out_elempack
-                            _sum = _mm512_fmadd_ps(_val_channel8, _conv_w8, _sum);
-                            _sum = _mm512_fmadd_ps(_val_channel9, _conv_w9, _sum);
-                            __m512 _conv_wa = _mm512_load_ps(kptr + 160); // 10 * out_elempack
-                            __m512 _conv_wb = _mm512_load_ps(kptr + 176); // 11 * out_elempack
-                            _sum = _mm512_fmadd_ps(_val_channela, _conv_wa, _sum);
-                            _sum = _mm512_fmadd_ps(_val_channelb, _conv_wb, _sum);
-                            __m512 _conv_wc = _mm512_load_ps(kptr + 192); // 12 * out_elempack
-                            __m512 _conv_wd = _mm512_load_ps(kptr + 208); // 13 * out_elempack
-                            _sum = _mm512_fmadd_ps(_val_channelc, _conv_wc, _sum);
-                            _sum = _mm512_fmadd_ps(_val_channeld, _conv_wd, _sum);
-                            __m512 _conv_we = _mm512_load_ps(kptr + 224); // 14 * out_elempack
-                            __m512 _conv_wf = _mm512_load_ps(kptr + 240); // 15 * out_elempack
-                            _sum = _mm512_fmadd_ps(_val_channele, _conv_we, _sum);
-                            _sum = _mm512_fmadd_ps(_val_channelf, _conv_wf, _sum);
-                            kptr += wstep;
-                        }
-                    }
-                }
-                _sum = activation_avx512(_sum, activation_type, activation_params);
-                _mm512_store_ps(outptr + (h_col * outw + w_col) * out_elempack, _sum);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16to1.h
deleted file mode 100644
index 1238537b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16to1.h
+++ /dev/null
@@ -1,370 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deformableconv2d_pack16to1_avx512(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const Mat& bottom_blob = bottom_blobs[0];
-    const Mat& offset = bottom_blobs[1];
-    const bool has_mask = (bottom_blobs.size() == 3);
-    const bool offset_not_pack = offset.elempack == 1;
-    const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const int size = outw * outh;
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-    const int elempack = 16;
-    const int out_elempack = 1;
-    const int wstep = out_elempack * elempack;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int h_col = 0; h_col < outh; h_col++)
-    {
-        for (int w_col = 0; w_col < outw; w_col++)
-        {
-            int h_in = h_col * stride_h - pad_top;
-            int w_in = w_col * stride_w - pad_left;
-            for (int oc = 0; oc < outch; oc++)
-            {
-                const float* kptr = weight_data_packed.channel(oc);
-                float* outptr = top_blob.channel(oc);
-                float _sum = 0.f;
-                if (bias_data_ptr)
-                    _sum = *(bias_data_ptr + oc);
-                for (int i = 0; i < kernel_h; i++)
-                {
-                    for (int j = 0; j < kernel_w; j++)
-                    {
-                        float offset_h = 0.f;
-                        float offset_w = 0.f;
-                        float mask_ = 1.f;
-                        if (offset_not_pack)
-                        {
-                            offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
-                            offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
-                        }
-                        else
-                        {
-                            const int y_c = (i * kernel_w + j) * 2;
-                            const int x_c = (i * kernel_w + j) * 2 + 1;
-                            offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
-                            offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
-                        }
-                        if (has_mask)
-                        {
-                            const Mat& mask = bottom_blobs[2];
-                            if (mask_not_pack)
-                            {
-                                mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
-                            }
-                            else
-                            {
-                                const int m_c = i * kernel_w + j;
-                                mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
-                            }
-                        }
-                        const float h_im = h_in + i * dilation_h + offset_h;
-                        const float w_im = w_in + j * dilation_w + offset_w;
-
-                        // Bilinear
-                        const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
-                        float w1 = 0.f;
-                        float w2 = 0.f;
-                        float w3 = 0.f;
-                        float w4 = 0.f;
-                        bool v1_cond = false;
-                        bool v2_cond = false;
-                        bool v3_cond = false;
-                        bool v4_cond = false;
-                        int v1_pos = 0;
-                        int v2_pos = 0;
-                        int v3_pos = 0;
-                        int v4_pos = 0;
-                        if (cond)
-                        {
-                            int h_low = (int)floorf(h_im);
-                            int w_low = (int)floorf(w_im);
-                            int h_high = h_low + 1;
-                            int w_high = w_low + 1;
-
-                            float lh = h_im - h_low;
-                            float lw = w_im - w_low;
-                            float hh = 1 - lh;
-                            float hw = 1 - lw;
-
-                            v1_cond = (h_low >= 0 && w_low >= 0);
-                            v2_cond = (h_low >= 0 && w_high <= w - 1);
-                            v3_cond = (h_high <= h - 1 && w_low >= 0);
-                            v4_cond = (h_high <= h - 1 && w_high <= w - 1);
-                            if (v1_cond)
-                                v1_pos = h_low * w + w_low;
-                            if (v2_cond)
-                                v2_pos = h_low * w + w_high;
-                            if (v3_cond)
-                                v3_pos = h_high * w + w_low;
-                            if (v4_cond)
-                                v4_pos = h_high * w + w_high;
-
-                            w1 = hh * hw;
-                            w2 = hh * lw;
-                            w3 = lh * hw;
-                            w4 = lh * lw;
-                        }
-
-                        for (int ic = 0; ic < inch; ic++)
-                        {
-                            const float* data_im_ptr = bottom_blob.channel(ic);
-                            float _val_channel0 = 0.f;
-                            float _val_channel1 = _val_channel0;
-                            float _val_channel2 = _val_channel0;
-                            float _val_channel3 = _val_channel0;
-                            float _val_channel4 = _val_channel0;
-                            float _val_channel5 = _val_channel0;
-                            float _val_channel6 = _val_channel0;
-                            float _val_channel7 = _val_channel0;
-                            float _val_channel8 = _val_channel0;
-                            float _val_channel9 = _val_channel0;
-                            float _val_channela = _val_channel0;
-                            float _val_channelb = _val_channel0;
-                            float _val_channelc = _val_channel0;
-                            float _val_channeld = _val_channel0;
-                            float _val_channele = _val_channel0;
-                            float _val_channelf = _val_channel0;
-                            if (cond)
-                            {
-                                float _v1_channel0 = _val_channel0;
-                                float _v1_channel1 = _val_channel0;
-                                float _v1_channel2 = _val_channel0;
-                                float _v1_channel3 = _val_channel0;
-                                float _v1_channel4 = _val_channel0;
-                                float _v1_channel5 = _val_channel0;
-                                float _v1_channel6 = _val_channel0;
-                                float _v1_channel7 = _val_channel0;
-                                float _v1_channel8 = _val_channel0;
-                                float _v1_channel9 = _val_channel0;
-                                float _v1_channela = _val_channel0;
-                                float _v1_channelb = _val_channel0;
-                                float _v1_channelc = _val_channel0;
-                                float _v1_channeld = _val_channel0;
-                                float _v1_channele = _val_channel0;
-                                float _v1_channelf = _val_channel0;
-                                float _v2_channel0 = _val_channel0;
-                                float _v2_channel1 = _val_channel0;
-                                float _v2_channel2 = _val_channel0;
-                                float _v2_channel3 = _val_channel0;
-                                float _v2_channel4 = _val_channel0;
-                                float _v2_channel5 = _val_channel0;
-                                float _v2_channel6 = _val_channel0;
-                                float _v2_channel7 = _val_channel0;
-                                float _v2_channel8 = _val_channel0;
-                                float _v2_channel9 = _val_channel0;
-                                float _v2_channela = _val_channel0;
-                                float _v2_channelb = _val_channel0;
-                                float _v2_channelc = _val_channel0;
-                                float _v2_channeld = _val_channel0;
-                                float _v2_channele = _val_channel0;
-                                float _v2_channelf = _val_channel0;
-                                float _v3_channel0 = _val_channel0;
-                                float _v3_channel1 = _val_channel0;
-                                float _v3_channel2 = _val_channel0;
-                                float _v3_channel3 = _val_channel0;
-                                float _v3_channel4 = _val_channel0;
-                                float _v3_channel5 = _val_channel0;
-                                float _v3_channel6 = _val_channel0;
-                                float _v3_channel7 = _val_channel0;
-                                float _v3_channel8 = _val_channel0;
-                                float _v3_channel9 = _val_channel0;
-                                float _v3_channela = _val_channel0;
-                                float _v3_channelb = _val_channel0;
-                                float _v3_channelc = _val_channel0;
-                                float _v3_channeld = _val_channel0;
-                                float _v3_channele = _val_channel0;
-                                float _v3_channelf = _val_channel0;
-                                float _v4_channel0 = _val_channel0;
-                                float _v4_channel1 = _val_channel0;
-                                float _v4_channel2 = _val_channel0;
-                                float _v4_channel3 = _val_channel0;
-                                float _v4_channel4 = _val_channel0;
-                                float _v4_channel5 = _val_channel0;
-                                float _v4_channel6 = _val_channel0;
-                                float _v4_channel7 = _val_channel0;
-                                float _v4_channel8 = _val_channel0;
-                                float _v4_channel9 = _val_channel0;
-                                float _v4_channela = _val_channel0;
-                                float _v4_channelb = _val_channel0;
-                                float _v4_channelc = _val_channel0;
-                                float _v4_channeld = _val_channel0;
-                                float _v4_channele = _val_channel0;
-                                float _v4_channelf = _val_channel0;
-                                if (v1_cond)
-                                {
-                                    _v1_channel0 = *(data_im_ptr + v1_pos * elempack);
-                                    _v1_channel1 = *(data_im_ptr + v1_pos * elempack + 1);
-                                    _v1_channel2 = *(data_im_ptr + v1_pos * elempack + 2);
-                                    _v1_channel3 = *(data_im_ptr + v1_pos * elempack + 3);
-                                    _v1_channel4 = *(data_im_ptr + v1_pos * elempack + 4);
-                                    _v1_channel5 = *(data_im_ptr + v1_pos * elempack + 5);
-                                    _v1_channel6 = *(data_im_ptr + v1_pos * elempack + 6);
-                                    _v1_channel7 = *(data_im_ptr + v1_pos * elempack + 7);
-                                    _v1_channel8 = *(data_im_ptr + v1_pos * elempack + 8);
-                                    _v1_channel9 = *(data_im_ptr + v1_pos * elempack + 9);
-                                    _v1_channela = *(data_im_ptr + v1_pos * elempack + 10);
-                                    _v1_channelb = *(data_im_ptr + v1_pos * elempack + 11);
-                                    _v1_channelc = *(data_im_ptr + v1_pos * elempack + 12);
-                                    _v1_channeld = *(data_im_ptr + v1_pos * elempack + 13);
-                                    _v1_channele = *(data_im_ptr + v1_pos * elempack + 14);
-                                    _v1_channelf = *(data_im_ptr + v1_pos * elempack + 15);
-                                }
-                                if (v2_cond)
-                                {
-                                    _v2_channel0 = *(data_im_ptr + v2_pos * elempack);
-                                    _v2_channel1 = *(data_im_ptr + v2_pos * elempack + 1);
-                                    _v2_channel2 = *(data_im_ptr + v2_pos * elempack + 2);
-                                    _v2_channel3 = *(data_im_ptr + v2_pos * elempack + 3);
-                                    _v2_channel4 = *(data_im_ptr + v2_pos * elempack + 4);
-                                    _v2_channel5 = *(data_im_ptr + v2_pos * elempack + 5);
-                                    _v2_channel6 = *(data_im_ptr + v2_pos * elempack + 6);
-                                    _v2_channel7 = *(data_im_ptr + v2_pos * elempack + 7);
-                                    _v2_channel8 = *(data_im_ptr + v2_pos * elempack + 8);
-                                    _v2_channel9 = *(data_im_ptr + v2_pos * elempack + 9);
-                                    _v2_channela = *(data_im_ptr + v2_pos * elempack + 10);
-                                    _v2_channelb = *(data_im_ptr + v2_pos * elempack + 11);
-                                    _v2_channelc = *(data_im_ptr + v2_pos * elempack + 12);
-                                    _v2_channeld = *(data_im_ptr + v2_pos * elempack + 13);
-                                    _v2_channele = *(data_im_ptr + v2_pos * elempack + 14);
-                                    _v2_channelf = *(data_im_ptr + v2_pos * elempack + 15);
-                                }
-                                if (v3_cond)
-                                {
-                                    _v3_channel0 = *(data_im_ptr + v3_pos * elempack);
-                                    _v3_channel1 = *(data_im_ptr + v3_pos * elempack + 1);
-                                    _v3_channel2 = *(data_im_ptr + v3_pos * elempack + 2);
-                                    _v3_channel3 = *(data_im_ptr + v3_pos * elempack + 3);
-                                    _v3_channel4 = *(data_im_ptr + v3_pos * elempack + 4);
-                                    _v3_channel5 = *(data_im_ptr + v3_pos * elempack + 5);
-                                    _v3_channel6 = *(data_im_ptr + v3_pos * elempack + 6);
-                                    _v3_channel7 = *(data_im_ptr + v3_pos * elempack + 7);
-                                    _v3_channel8 = *(data_im_ptr + v3_pos * elempack + 8);
-                                    _v3_channel9 = *(data_im_ptr + v3_pos * elempack + 9);
-                                    _v3_channela = *(data_im_ptr + v3_pos * elempack + 10);
-                                    _v3_channelb = *(data_im_ptr + v3_pos * elempack + 11);
-                                    _v3_channelc = *(data_im_ptr + v3_pos * elempack + 12);
-                                    _v3_channeld = *(data_im_ptr + v3_pos * elempack + 13);
-                                    _v3_channele = *(data_im_ptr + v3_pos * elempack + 14);
-                                    _v3_channelf = *(data_im_ptr + v3_pos * elempack + 15);
-                                }
-                                if (v4_cond)
-                                {
-                                    _v4_channel0 = *(data_im_ptr + v4_pos * elempack);
-                                    _v4_channel1 = *(data_im_ptr + v4_pos * elempack + 1);
-                                    _v4_channel2 = *(data_im_ptr + v4_pos * elempack + 2);
-                                    _v4_channel3 = *(data_im_ptr + v4_pos * elempack + 3);
-                                    _v4_channel4 = *(data_im_ptr + v4_pos * elempack + 4);
-                                    _v4_channel5 = *(data_im_ptr + v4_pos * elempack + 5);
-                                    _v4_channel6 = *(data_im_ptr + v4_pos * elempack + 6);
-                                    _v4_channel7 = *(data_im_ptr + v4_pos * elempack + 7);
-                                    _v4_channel8 = *(data_im_ptr + v4_pos * elempack + 8);
-                                    _v4_channel9 = *(data_im_ptr + v4_pos * elempack + 9);
-                                    _v4_channela = *(data_im_ptr + v4_pos * elempack + 10);
-                                    _v4_channelb = *(data_im_ptr + v4_pos * elempack + 11);
-                                    _v4_channelc = *(data_im_ptr + v4_pos * elempack + 12);
-                                    _v4_channeld = *(data_im_ptr + v4_pos * elempack + 13);
-                                    _v4_channele = *(data_im_ptr + v4_pos * elempack + 14);
-                                    _v4_channelf = *(data_im_ptr + v4_pos * elempack + 15);
-                                }
-                                _val_channel0 = w1 * _v1_channel0 + w2 * _v2_channel0 + w3 * _v3_channel0 + w4 * _v4_channel0;
-                                _val_channel1 = w1 * _v1_channel1 + w2 * _v2_channel1 + w3 * _v3_channel1 + w4 * _v4_channel1;
-                                _val_channel2 = w1 * _v1_channel2 + w2 * _v2_channel2 + w3 * _v3_channel2 + w4 * _v4_channel2;
-                                _val_channel3 = w1 * _v1_channel3 + w2 * _v2_channel3 + w3 * _v3_channel3 + w4 * _v4_channel3;
-                                _val_channel4 = w1 * _v1_channel4 + w2 * _v2_channel4 + w3 * _v3_channel4 + w4 * _v4_channel4;
-                                _val_channel5 = w1 * _v1_channel5 + w2 * _v2_channel5 + w3 * _v3_channel5 + w4 * _v4_channel5;
-                                _val_channel6 = w1 * _v1_channel6 + w2 * _v2_channel6 + w3 * _v3_channel6 + w4 * _v4_channel6;
-                                _val_channel7 = w1 * _v1_channel7 + w2 * _v2_channel7 + w3 * _v3_channel7 + w4 * _v4_channel7;
-                                _val_channel8 = w1 * _v1_channel8 + w2 * _v2_channel8 + w3 * _v3_channel8 + w4 * _v4_channel8;
-                                _val_channel9 = w1 * _v1_channel9 + w2 * _v2_channel9 + w3 * _v3_channel9 + w4 * _v4_channel9;
-                                _val_channela = w1 * _v1_channela + w2 * _v2_channela + w3 * _v3_channela + w4 * _v4_channela;
-                                _val_channelb = w1 * _v1_channelb + w2 * _v2_channelb + w3 * _v3_channelb + w4 * _v4_channelb;
-                                _val_channelc = w1 * _v1_channelc + w2 * _v2_channelc + w3 * _v3_channelc + w4 * _v4_channelc;
-                                _val_channeld = w1 * _v1_channeld + w2 * _v2_channeld + w3 * _v3_channeld + w4 * _v4_channeld;
-                                _val_channele = w1 * _v1_channele + w2 * _v2_channele + w3 * _v3_channele + w4 * _v4_channele;
-                                _val_channelf = w1 * _v1_channelf + w2 * _v2_channelf + w3 * _v3_channelf + w4 * _v4_channelf;
-                            }
-                            if (has_mask)
-                            {
-                                _val_channel0 *= mask_;
-                                _val_channel1 *= mask_;
-                                _val_channel2 *= mask_;
-                                _val_channel3 *= mask_;
-                                _val_channel4 *= mask_;
-                                _val_channel5 *= mask_;
-                                _val_channel6 *= mask_;
-                                _val_channel7 *= mask_;
-                                _val_channel8 *= mask_;
-                                _val_channel9 *= mask_;
-                                _val_channela *= mask_;
-                                _val_channelb *= mask_;
-                                _val_channelc *= mask_;
-                                _val_channeld *= mask_;
-                                _val_channele *= mask_;
-                                _val_channelf *= mask_;
-                            }
-                            float _conv_w0 = *(kptr);
-                            float _conv_w1 = *(kptr + out_elempack); // 1 * out_elempack
-                            _sum += (_val_channel0 * _conv_w0);
-                            _sum += (_val_channel1 * _conv_w1);
-                            float _conv_w2 = *(kptr + 2); // 2 * out_elempack
-                            float _conv_w3 = *(kptr + 3); // 3 * out_elempack
-                            _sum += (_val_channel2 * _conv_w2);
-                            _sum += (_val_channel3 * _conv_w3);
-                            float _conv_w4 = *(kptr + 4); // 4 * out_elempack
-                            float _conv_w5 = *(kptr + 5); // 5 * out_elempack
-                            _sum += (_val_channel4 * _conv_w4);
-                            _sum += (_val_channel5 * _conv_w5);
-                            float _conv_w6 = *(kptr + 6); // 6 * out_elempack
-                            float _conv_w7 = *(kptr + 7); // 7 * out_elempack
-                            _sum += (_val_channel6 * _conv_w6);
-                            _sum += (_val_channel7 * _conv_w7);
-                            float _conv_w8 = *(kptr + 8); // 8 * out_elempack
-                            float _conv_w9 = *(kptr + 9); // 9 * out_elempack
-                            _sum += (_val_channel8 * _conv_w8);
-                            _sum += (_val_channel9 * _conv_w9);
-                            float _conv_wa = *(kptr + 10); // 10 * out_elempack
-                            float _conv_wb = *(kptr + 11); // 11 * out_elempack
-                            _sum += (_val_channela * _conv_wa);
-                            _sum += (_val_channelb * _conv_wb);
-                            float _conv_wc = *(kptr + 12); // 12 * out_elempack
-                            float _conv_wd = *(kptr + 13); // 13 * out_elempack
-                            _sum += (_val_channelc * _conv_wc);
-                            _sum += (_val_channeld * _conv_wd);
-                            float _conv_we = *(kptr + 14); // 14 * out_elempack
-                            float _conv_wf = *(kptr + 15); // 15 * out_elempack
-                            _sum += (_val_channele * _conv_we);
-                            _sum += (_val_channelf * _conv_wf);
-                            kptr += wstep;
-                        }
-                    }
-                }
-                _sum = activation_ss(_sum, activation_type, activation_params);
-                *(outptr + h_col * outw + w_col) = _sum;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16to4.h
deleted file mode 100644
index 058f2bed..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16to4.h
+++ /dev/null
@@ -1,435 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deformableconv2d_pack16to4_avx512(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const Mat& bottom_blob = bottom_blobs[0];
-    const Mat& offset = bottom_blobs[1];
-    const bool has_mask = (bottom_blobs.size() == 3);
-    const bool offset_not_pack = offset.elempack == 1;
-    const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const int size = outw * outh;
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-    const int elempack = 16;
-    const int out_elempack = 4;
-    const int wstep = out_elempack * elempack;
-    const float zeros[out_elempack] = {0.f, 0.f, 0.f, 0.f};
-    const float* zeros_ptr = zeros;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int h_col = 0; h_col < outh; h_col++)
-    {
-        for (int w_col = 0; w_col < outw; w_col++)
-        {
-            int h_in = h_col * stride_h - pad_top;
-            int w_in = w_col * stride_w - pad_left;
-            for (int oc = 0; oc < outch; oc++)
-            {
-                const float* kptr = weight_data_packed.channel(oc);
-                float* outptr = top_blob.channel(oc);
-                __m128 _sum = _mm_setzero_ps();
-                if (bias_data_ptr)
-                    _sum = _mm_loadu_ps(bias_data_ptr + oc * out_elempack);
-                for (int i = 0; i < kernel_h; i++)
-                {
-                    for (int j = 0; j < kernel_w; j++)
-                    {
-                        float offset_h = 0.f;
-                        float offset_w = 0.f;
-                        float mask_ = 1.f;
-                        if (offset_not_pack)
-                        {
-                            offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
-                            offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
-                        }
-                        else
-                        {
-                            const int y_c = (i * kernel_w + j) * 2;
-                            const int x_c = (i * kernel_w + j) * 2 + 1;
-                            offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
-                            offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
-                        }
-                        if (has_mask)
-                        {
-                            const Mat& mask = bottom_blobs[2];
-                            if (mask_not_pack)
-                            {
-                                mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
-                            }
-                            else
-                            {
-                                const int m_c = i * kernel_w + j;
-                                mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
-                            }
-                        }
-                        const float h_im = h_in + i * dilation_h + offset_h;
-                        const float w_im = w_in + j * dilation_w + offset_w;
-
-                        // Bilinear
-                        const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
-                        float w1 = 0.f;
-                        float w2 = 0.f;
-                        float w3 = 0.f;
-                        float w4 = 0.f;
-                        bool v1_cond = false;
-                        bool v2_cond = false;
-                        bool v3_cond = false;
-                        bool v4_cond = false;
-                        int v1_pos = 0;
-                        int v2_pos = 0;
-                        int v3_pos = 0;
-                        int v4_pos = 0;
-                        if (cond)
-                        {
-                            int h_low = (int)floorf(h_im);
-                            int w_low = (int)floorf(w_im);
-                            int h_high = h_low + 1;
-                            int w_high = w_low + 1;
-
-                            float lh = h_im - h_low;
-                            float lw = w_im - w_low;
-                            float hh = 1 - lh;
-                            float hw = 1 - lw;
-
-                            v1_cond = (h_low >= 0 && w_low >= 0);
-                            v2_cond = (h_low >= 0 && w_high <= w - 1);
-                            v3_cond = (h_high <= h - 1 && w_low >= 0);
-                            v4_cond = (h_high <= h - 1 && w_high <= w - 1);
-                            if (v1_cond)
-                                v1_pos = h_low * w + w_low;
-                            if (v2_cond)
-                                v2_pos = h_low * w + w_high;
-                            if (v3_cond)
-                                v3_pos = h_high * w + w_low;
-                            if (v4_cond)
-                                v4_pos = h_high * w + w_high;
-
-                            w1 = hh * hw;
-                            w2 = hh * lw;
-                            w3 = lh * hw;
-                            w4 = lh * lw;
-                        }
-                        const float w1s[out_elempack] = {w1, w1, w1, w1};
-                        const float* w1_ptr = w1s;
-                        const float w2s[out_elempack] = {w2, w2, w2, w2};
-                        const float* w2_ptr = w2s;
-                        const float w3s[out_elempack] = {w3, w3, w3, w3};
-                        const float* w3_ptr = w3s;
-                        const float w4s[out_elempack] = {w4, w4, w4, w4};
-                        const float* w4_ptr = w4s;
-                        const float masks[out_elempack] = {mask_, mask_, mask_, mask_};
-                        const float* mask_ptr = masks;
-
-                        for (int ic = 0; ic < inch; ic++)
-                        {
-                            const float* data_im_ptr = bottom_blob.channel(ic);
-                            __m128 _val_channel0 = _mm_loadu_ps(zeros_ptr);
-                            __m128 _val_channel1 = _val_channel0;
-                            __m128 _val_channel2 = _val_channel0;
-                            __m128 _val_channel3 = _val_channel0;
-                            __m128 _val_channel4 = _val_channel0;
-                            __m128 _val_channel5 = _val_channel0;
-                            __m128 _val_channel6 = _val_channel0;
-                            __m128 _val_channel7 = _val_channel0;
-                            __m128 _val_channel8 = _val_channel0;
-                            __m128 _val_channel9 = _val_channel0;
-                            __m128 _val_channela = _val_channel0;
-                            __m128 _val_channelb = _val_channel0;
-                            __m128 _val_channelc = _val_channel0;
-                            __m128 _val_channeld = _val_channel0;
-                            __m128 _val_channele = _val_channel0;
-                            __m128 _val_channelf = _val_channel0;
-                            if (cond)
-                            {
-                                __m128 _v1_channel0 = _val_channel0;
-                                __m128 _v1_channel1 = _val_channel0;
-                                __m128 _v1_channel2 = _val_channel0;
-                                __m128 _v1_channel3 = _val_channel0;
-                                __m128 _v1_channel4 = _val_channel0;
-                                __m128 _v1_channel5 = _val_channel0;
-                                __m128 _v1_channel6 = _val_channel0;
-                                __m128 _v1_channel7 = _val_channel0;
-                                __m128 _v1_channel8 = _val_channel0;
-                                __m128 _v1_channel9 = _val_channel0;
-                                __m128 _v1_channela = _val_channel0;
-                                __m128 _v1_channelb = _val_channel0;
-                                __m128 _v1_channelc = _val_channel0;
-                                __m128 _v1_channeld = _val_channel0;
-                                __m128 _v1_channele = _val_channel0;
-                                __m128 _v1_channelf = _val_channel0;
-                                __m128 _v2_channel0 = _val_channel0;
-                                __m128 _v2_channel1 = _val_channel0;
-                                __m128 _v2_channel2 = _val_channel0;
-                                __m128 _v2_channel3 = _val_channel0;
-                                __m128 _v2_channel4 = _val_channel0;
-                                __m128 _v2_channel5 = _val_channel0;
-                                __m128 _v2_channel6 = _val_channel0;
-                                __m128 _v2_channel7 = _val_channel0;
-                                __m128 _v2_channel8 = _val_channel0;
-                                __m128 _v2_channel9 = _val_channel0;
-                                __m128 _v2_channela = _val_channel0;
-                                __m128 _v2_channelb = _val_channel0;
-                                __m128 _v2_channelc = _val_channel0;
-                                __m128 _v2_channeld = _val_channel0;
-                                __m128 _v2_channele = _val_channel0;
-                                __m128 _v2_channelf = _val_channel0;
-                                __m128 _v3_channel0 = _val_channel0;
-                                __m128 _v3_channel1 = _val_channel0;
-                                __m128 _v3_channel2 = _val_channel0;
-                                __m128 _v3_channel3 = _val_channel0;
-                                __m128 _v3_channel4 = _val_channel0;
-                                __m128 _v3_channel5 = _val_channel0;
-                                __m128 _v3_channel6 = _val_channel0;
-                                __m128 _v3_channel7 = _val_channel0;
-                                __m128 _v3_channel8 = _val_channel0;
-                                __m128 _v3_channel9 = _val_channel0;
-                                __m128 _v3_channela = _val_channel0;
-                                __m128 _v3_channelb = _val_channel0;
-                                __m128 _v3_channelc = _val_channel0;
-                                __m128 _v3_channeld = _val_channel0;
-                                __m128 _v3_channele = _val_channel0;
-                                __m128 _v3_channelf = _val_channel0;
-                                __m128 _v4_channel0 = _val_channel0;
-                                __m128 _v4_channel1 = _val_channel0;
-                                __m128 _v4_channel2 = _val_channel0;
-                                __m128 _v4_channel3 = _val_channel0;
-                                __m128 _v4_channel4 = _val_channel0;
-                                __m128 _v4_channel5 = _val_channel0;
-                                __m128 _v4_channel6 = _val_channel0;
-                                __m128 _v4_channel7 = _val_channel0;
-                                __m128 _v4_channel8 = _val_channel0;
-                                __m128 _v4_channel9 = _val_channel0;
-                                __m128 _v4_channela = _val_channel0;
-                                __m128 _v4_channelb = _val_channel0;
-                                __m128 _v4_channelc = _val_channel0;
-                                __m128 _v4_channeld = _val_channel0;
-                                __m128 _v4_channele = _val_channel0;
-                                __m128 _v4_channelf = _val_channel0;
-                                if (v1_cond)
-                                {
-                                    _v1_channel0 = _mm_load1_ps(data_im_ptr + v1_pos * elempack);
-                                    _v1_channel1 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 1);
-                                    _v1_channel2 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 2);
-                                    _v1_channel3 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 3);
-                                    _v1_channel4 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 4);
-                                    _v1_channel5 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 5);
-                                    _v1_channel6 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 6);
-                                    _v1_channel7 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 7);
-                                    _v1_channel8 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 8);
-                                    _v1_channel9 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 9);
-                                    _v1_channela = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 10);
-                                    _v1_channelb = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 11);
-                                    _v1_channelc = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 12);
-                                    _v1_channeld = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 13);
-                                    _v1_channele = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 14);
-                                    _v1_channelf = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 15);
-                                }
-                                if (v2_cond)
-                                {
-                                    _v2_channel0 = _mm_load1_ps(data_im_ptr + v2_pos * elempack);
-                                    _v2_channel1 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 1);
-                                    _v2_channel2 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 2);
-                                    _v2_channel3 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 3);
-                                    _v2_channel4 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 4);
-                                    _v2_channel5 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 5);
-                                    _v2_channel6 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 6);
-                                    _v2_channel7 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 7);
-                                    _v2_channel8 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 8);
-                                    _v2_channel9 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 9);
-                                    _v2_channela = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 10);
-                                    _v2_channelb = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 11);
-                                    _v2_channelc = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 12);
-                                    _v2_channeld = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 13);
-                                    _v2_channele = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 14);
-                                    _v2_channelf = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 15);
-                                }
-                                if (v3_cond)
-                                {
-                                    _v3_channel0 = _mm_load1_ps(data_im_ptr + v3_pos * elempack);
-                                    _v3_channel1 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 1);
-                                    _v3_channel2 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 2);
-                                    _v3_channel3 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 3);
-                                    _v3_channel4 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 4);
-                                    _v3_channel5 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 5);
-                                    _v3_channel6 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 6);
-                                    _v3_channel7 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 7);
-                                    _v3_channel8 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 8);
-                                    _v3_channel9 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 9);
-                                    _v3_channela = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 10);
-                                    _v3_channelb = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 11);
-                                    _v3_channelc = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 12);
-                                    _v3_channeld = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 13);
-                                    _v3_channele = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 14);
-                                    _v3_channelf = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 15);
-                                }
-                                if (v4_cond)
-                                {
-                                    _v4_channel0 = _mm_load1_ps(data_im_ptr + v4_pos * elempack);
-                                    _v4_channel1 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 1);
-                                    _v4_channel2 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 2);
-                                    _v4_channel3 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 3);
-                                    _v4_channel4 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 4);
-                                    _v4_channel5 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 5);
-                                    _v4_channel6 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 6);
-                                    _v4_channel7 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 7);
-                                    _v4_channel8 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 8);
-                                    _v4_channel9 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 9);
-                                    _v4_channela = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 10);
-                                    _v4_channelb = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 11);
-                                    _v4_channelc = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 12);
-                                    _v4_channeld = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 13);
-                                    _v4_channele = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 14);
-                                    _v4_channelf = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 15);
-                                }
-                                __m128 _w1 = _mm_loadu_ps(w1_ptr);
-                                __m128 _w2 = _mm_loadu_ps(w2_ptr);
-                                __m128 _w3 = _mm_loadu_ps(w3_ptr);
-                                __m128 _w4 = _mm_loadu_ps(w4_ptr);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v1_channel0, _w1, _val_channel0);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v2_channel0, _w2, _val_channel0);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v3_channel0, _w3, _val_channel0);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v4_channel0, _w4, _val_channel0);
-                                _val_channel1 = _mm_comp_fmadd_ps(_v1_channel1, _w1, _val_channel1);
-                                _val_channel1 = _mm_comp_fmadd_ps(_v2_channel1, _w2, _val_channel1);
-                                _val_channel1 = _mm_comp_fmadd_ps(_v3_channel1, _w3, _val_channel1);
-                                _val_channel1 = _mm_comp_fmadd_ps(_v4_channel1, _w4, _val_channel1);
-                                _val_channel2 = _mm_comp_fmadd_ps(_v1_channel2, _w1, _val_channel2);
-                                _val_channel2 = _mm_comp_fmadd_ps(_v2_channel2, _w2, _val_channel2);
-                                _val_channel2 = _mm_comp_fmadd_ps(_v3_channel2, _w3, _val_channel2);
-                                _val_channel2 = _mm_comp_fmadd_ps(_v4_channel2, _w4, _val_channel2);
-                                _val_channel3 = _mm_comp_fmadd_ps(_v1_channel3, _w1, _val_channel3);
-                                _val_channel3 = _mm_comp_fmadd_ps(_v2_channel3, _w2, _val_channel3);
-                                _val_channel3 = _mm_comp_fmadd_ps(_v3_channel3, _w3, _val_channel3);
-                                _val_channel3 = _mm_comp_fmadd_ps(_v4_channel3, _w4, _val_channel3);
-                                _val_channel4 = _mm_comp_fmadd_ps(_v1_channel4, _w1, _val_channel4);
-                                _val_channel4 = _mm_comp_fmadd_ps(_v2_channel4, _w2, _val_channel4);
-                                _val_channel4 = _mm_comp_fmadd_ps(_v3_channel4, _w3, _val_channel4);
-                                _val_channel4 = _mm_comp_fmadd_ps(_v4_channel4, _w4, _val_channel4);
-                                _val_channel5 = _mm_comp_fmadd_ps(_v1_channel5, _w1, _val_channel5);
-                                _val_channel5 = _mm_comp_fmadd_ps(_v2_channel5, _w2, _val_channel5);
-                                _val_channel5 = _mm_comp_fmadd_ps(_v3_channel5, _w3, _val_channel5);
-                                _val_channel5 = _mm_comp_fmadd_ps(_v4_channel5, _w4, _val_channel5);
-                                _val_channel6 = _mm_comp_fmadd_ps(_v1_channel6, _w1, _val_channel6);
-                                _val_channel6 = _mm_comp_fmadd_ps(_v2_channel6, _w2, _val_channel6);
-                                _val_channel6 = _mm_comp_fmadd_ps(_v3_channel6, _w3, _val_channel6);
-                                _val_channel6 = _mm_comp_fmadd_ps(_v4_channel6, _w4, _val_channel6);
-                                _val_channel7 = _mm_comp_fmadd_ps(_v1_channel7, _w1, _val_channel7);
-                                _val_channel7 = _mm_comp_fmadd_ps(_v2_channel7, _w2, _val_channel7);
-                                _val_channel7 = _mm_comp_fmadd_ps(_v3_channel7, _w3, _val_channel7);
-                                _val_channel7 = _mm_comp_fmadd_ps(_v4_channel7, _w4, _val_channel7);
-                                _val_channel8 = _mm_comp_fmadd_ps(_v1_channel8, _w1, _val_channel8);
-                                _val_channel8 = _mm_comp_fmadd_ps(_v2_channel8, _w2, _val_channel8);
-                                _val_channel8 = _mm_comp_fmadd_ps(_v3_channel8, _w3, _val_channel8);
-                                _val_channel8 = _mm_comp_fmadd_ps(_v4_channel8, _w4, _val_channel8);
-                                _val_channel9 = _mm_comp_fmadd_ps(_v1_channel9, _w1, _val_channel9);
-                                _val_channel9 = _mm_comp_fmadd_ps(_v2_channel9, _w2, _val_channel9);
-                                _val_channel9 = _mm_comp_fmadd_ps(_v3_channel9, _w3, _val_channel9);
-                                _val_channel9 = _mm_comp_fmadd_ps(_v4_channel9, _w4, _val_channel9);
-                                _val_channela = _mm_comp_fmadd_ps(_v1_channela, _w1, _val_channela);
-                                _val_channela = _mm_comp_fmadd_ps(_v2_channela, _w2, _val_channela);
-                                _val_channela = _mm_comp_fmadd_ps(_v3_channela, _w3, _val_channela);
-                                _val_channela = _mm_comp_fmadd_ps(_v4_channela, _w4, _val_channela);
-                                _val_channelb = _mm_comp_fmadd_ps(_v1_channelb, _w1, _val_channelb);
-                                _val_channelb = _mm_comp_fmadd_ps(_v2_channelb, _w2, _val_channelb);
-                                _val_channelb = _mm_comp_fmadd_ps(_v3_channelb, _w3, _val_channelb);
-                                _val_channelb = _mm_comp_fmadd_ps(_v4_channelb, _w4, _val_channelb);
-                                _val_channelc = _mm_comp_fmadd_ps(_v1_channelc, _w1, _val_channelc);
-                                _val_channelc = _mm_comp_fmadd_ps(_v2_channelc, _w2, _val_channelc);
-                                _val_channelc = _mm_comp_fmadd_ps(_v3_channelc, _w3, _val_channelc);
-                                _val_channelc = _mm_comp_fmadd_ps(_v4_channelc, _w4, _val_channelc);
-                                _val_channeld = _mm_comp_fmadd_ps(_v1_channeld, _w1, _val_channeld);
-                                _val_channeld = _mm_comp_fmadd_ps(_v2_channeld, _w2, _val_channeld);
-                                _val_channeld = _mm_comp_fmadd_ps(_v3_channeld, _w3, _val_channeld);
-                                _val_channeld = _mm_comp_fmadd_ps(_v4_channeld, _w4, _val_channeld);
-                                _val_channele = _mm_comp_fmadd_ps(_v1_channele, _w1, _val_channele);
-                                _val_channele = _mm_comp_fmadd_ps(_v2_channele, _w2, _val_channele);
-                                _val_channele = _mm_comp_fmadd_ps(_v3_channele, _w3, _val_channele);
-                                _val_channele = _mm_comp_fmadd_ps(_v4_channele, _w4, _val_channele);
-                                _val_channelf = _mm_comp_fmadd_ps(_v1_channelf, _w1, _val_channelf);
-                                _val_channelf = _mm_comp_fmadd_ps(_v2_channelf, _w2, _val_channelf);
-                                _val_channelf = _mm_comp_fmadd_ps(_v3_channelf, _w3, _val_channelf);
-                                _val_channelf = _mm_comp_fmadd_ps(_v4_channelf, _w4, _val_channelf);
-                            }
-                            if (has_mask)
-                            {
-                                __m128 _mask = _mm_loadu_ps(mask_ptr);
-                                _val_channel0 = _mm_mul_ps(_val_channel0, _mask);
-                                _val_channel1 = _mm_mul_ps(_val_channel1, _mask);
-                                _val_channel2 = _mm_mul_ps(_val_channel2, _mask);
-                                _val_channel3 = _mm_mul_ps(_val_channel3, _mask);
-                                _val_channel4 = _mm_mul_ps(_val_channel4, _mask);
-                                _val_channel5 = _mm_mul_ps(_val_channel5, _mask);
-                                _val_channel6 = _mm_mul_ps(_val_channel6, _mask);
-                                _val_channel7 = _mm_mul_ps(_val_channel7, _mask);
-                                _val_channel8 = _mm_mul_ps(_val_channel8, _mask);
-                                _val_channel9 = _mm_mul_ps(_val_channel9, _mask);
-                                _val_channela = _mm_mul_ps(_val_channela, _mask);
-                                _val_channelb = _mm_mul_ps(_val_channelb, _mask);
-                                _val_channelc = _mm_mul_ps(_val_channelc, _mask);
-                                _val_channeld = _mm_mul_ps(_val_channeld, _mask);
-                                _val_channele = _mm_mul_ps(_val_channele, _mask);
-                                _val_channelf = _mm_mul_ps(_val_channelf, _mask);
-                            }
-                            __m128 _conv_w0 = _mm_load_ps(kptr);
-                            __m128 _conv_w1 = _mm_load_ps(kptr + out_elempack); // 1 * out_elempack
-                            _sum = _mm_comp_fmadd_ps(_val_channel0, _conv_w0, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val_channel1, _conv_w1, _sum);
-                            __m128 _conv_w2 = _mm_load_ps(kptr + 8);  // 2 * out_elempack
-                            __m128 _conv_w3 = _mm_load_ps(kptr + 12); // 3 * out_elempack
-                            _sum = _mm_comp_fmadd_ps(_val_channel2, _conv_w2, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val_channel3, _conv_w3, _sum);
-                            __m128 _conv_w4 = _mm_load_ps(kptr + 16); // 4 * out_elempack
-                            __m128 _conv_w5 = _mm_load_ps(kptr + 20); // 5 * out_elempack
-                            _sum = _mm_comp_fmadd_ps(_val_channel4, _conv_w4, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val_channel5, _conv_w5, _sum);
-                            __m128 _conv_w6 = _mm_load_ps(kptr + 24); // 6 * out_elempack
-                            __m128 _conv_w7 = _mm_load_ps(kptr + 28); // 7 * out_elempack
-                            _sum = _mm_comp_fmadd_ps(_val_channel6, _conv_w6, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val_channel7, _conv_w7, _sum);
-                            __m128 _conv_w8 = _mm_load_ps(kptr + 32); // 8 * out_elempack
-                            __m128 _conv_w9 = _mm_load_ps(kptr + 36); // 9 * out_elempack
-                            _sum = _mm_comp_fmadd_ps(_val_channel8, _conv_w8, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val_channel9, _conv_w9, _sum);
-                            __m128 _conv_wa = _mm_load_ps(kptr + 40); // 10 * out_elempack
-                            __m128 _conv_wb = _mm_load_ps(kptr + 44); // 11 * out_elempack
-                            _sum = _mm_comp_fmadd_ps(_val_channela, _conv_wa, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val_channelb, _conv_wb, _sum);
-                            __m128 _conv_wc = _mm_load_ps(kptr + 48); // 12 * out_elempack
-                            __m128 _conv_wd = _mm_load_ps(kptr + 52); // 13 * out_elempack
-                            _sum = _mm_comp_fmadd_ps(_val_channelc, _conv_wc, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val_channeld, _conv_wd, _sum);
-                            __m128 _conv_we = _mm_load_ps(kptr + 56); // 14 * out_elempack
-                            __m128 _conv_wf = _mm_load_ps(kptr + 60); // 15 * out_elempack
-                            _sum = _mm_comp_fmadd_ps(_val_channele, _conv_we, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val_channelf, _conv_wf, _sum);
-                            kptr += wstep;
-                        }
-                    }
-                }
-                _sum = activation_sse(_sum, activation_type, activation_params);
-                _mm_storeu_ps(outptr + (h_col * outw + w_col) * out_elempack, _sum);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16to8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16to8.h
deleted file mode 100644
index c64f5e46..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack16to8.h
+++ /dev/null
@@ -1,435 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deformableconv2d_pack16to8_avx512(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const Mat& bottom_blob = bottom_blobs[0];
-    const Mat& offset = bottom_blobs[1];
-    const bool has_mask = (bottom_blobs.size() == 3);
-    const bool offset_not_pack = offset.elempack == 1;
-    const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const int size = outw * outh;
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-    const int elempack = 16;
-    const int out_elempack = 8;
-    const int wstep = out_elempack * elempack;
-    const float zeros[out_elempack] = {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
-    const float* zeros_ptr = zeros;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int h_col = 0; h_col < outh; h_col++)
-    {
-        for (int w_col = 0; w_col < outw; w_col++)
-        {
-            int h_in = h_col * stride_h - pad_top;
-            int w_in = w_col * stride_w - pad_left;
-            for (int oc = 0; oc < outch; oc++)
-            {
-                const float* kptr = weight_data_packed.channel(oc);
-                float* outptr = top_blob.channel(oc);
-                __m256 _sum = _mm256_setzero_ps();
-                if (bias_data_ptr)
-                    _sum = _mm256_loadu_ps(bias_data_ptr + oc * out_elempack);
-                for (int i = 0; i < kernel_h; i++)
-                {
-                    for (int j = 0; j < kernel_w; j++)
-                    {
-                        float offset_h = 0.f;
-                        float offset_w = 0.f;
-                        float mask_ = 1.f;
-                        if (offset_not_pack)
-                        {
-                            offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
-                            offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
-                        }
-                        else
-                        {
-                            const int y_c = (i * kernel_w + j) * 2;
-                            const int x_c = (i * kernel_w + j) * 2 + 1;
-                            offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
-                            offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
-                        }
-                        if (has_mask)
-                        {
-                            const Mat& mask = bottom_blobs[2];
-                            if (mask_not_pack)
-                            {
-                                mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
-                            }
-                            else
-                            {
-                                const int m_c = i * kernel_w + j;
-                                mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
-                            }
-                        }
-                        const float h_im = h_in + i * dilation_h + offset_h;
-                        const float w_im = w_in + j * dilation_w + offset_w;
-
-                        // Bilinear
-                        const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
-                        float w1 = 0.f;
-                        float w2 = 0.f;
-                        float w3 = 0.f;
-                        float w4 = 0.f;
-                        bool v1_cond = false;
-                        bool v2_cond = false;
-                        bool v3_cond = false;
-                        bool v4_cond = false;
-                        int v1_pos = 0;
-                        int v2_pos = 0;
-                        int v3_pos = 0;
-                        int v4_pos = 0;
-                        if (cond)
-                        {
-                            int h_low = (int)floorf(h_im);
-                            int w_low = (int)floorf(w_im);
-                            int h_high = h_low + 1;
-                            int w_high = w_low + 1;
-
-                            float lh = h_im - h_low;
-                            float lw = w_im - w_low;
-                            float hh = 1 - lh;
-                            float hw = 1 - lw;
-
-                            v1_cond = (h_low >= 0 && w_low >= 0);
-                            v2_cond = (h_low >= 0 && w_high <= w - 1);
-                            v3_cond = (h_high <= h - 1 && w_low >= 0);
-                            v4_cond = (h_high <= h - 1 && w_high <= w - 1);
-                            if (v1_cond)
-                                v1_pos = h_low * w + w_low;
-                            if (v2_cond)
-                                v2_pos = h_low * w + w_high;
-                            if (v3_cond)
-                                v3_pos = h_high * w + w_low;
-                            if (v4_cond)
-                                v4_pos = h_high * w + w_high;
-
-                            w1 = hh * hw;
-                            w2 = hh * lw;
-                            w3 = lh * hw;
-                            w4 = lh * lw;
-                        }
-                        const float w1s[out_elempack] = {w1, w1, w1, w1, w1, w1, w1, w1};
-                        const float* w1_ptr = w1s;
-                        const float w2s[out_elempack] = {w2, w2, w2, w2, w2, w2, w2, w2};
-                        const float* w2_ptr = w2s;
-                        const float w3s[out_elempack] = {w3, w3, w3, w3, w3, w3, w3, w3};
-                        const float* w3_ptr = w3s;
-                        const float w4s[out_elempack] = {w4, w4, w4, w4, w4, w4, w4, w4};
-                        const float* w4_ptr = w4s;
-                        const float masks[out_elempack] = {mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_};
-                        const float* mask_ptr = masks;
-
-                        for (int ic = 0; ic < inch; ic++)
-                        {
-                            const float* data_im_ptr = bottom_blob.channel(ic);
-                            __m256 _val_channel0 = _mm256_loadu_ps(zeros_ptr);
-                            __m256 _val_channel1 = _val_channel0;
-                            __m256 _val_channel2 = _val_channel0;
-                            __m256 _val_channel3 = _val_channel0;
-                            __m256 _val_channel4 = _val_channel0;
-                            __m256 _val_channel5 = _val_channel0;
-                            __m256 _val_channel6 = _val_channel0;
-                            __m256 _val_channel7 = _val_channel0;
-                            __m256 _val_channel8 = _val_channel0;
-                            __m256 _val_channel9 = _val_channel0;
-                            __m256 _val_channela = _val_channel0;
-                            __m256 _val_channelb = _val_channel0;
-                            __m256 _val_channelc = _val_channel0;
-                            __m256 _val_channeld = _val_channel0;
-                            __m256 _val_channele = _val_channel0;
-                            __m256 _val_channelf = _val_channel0;
-                            if (cond)
-                            {
-                                __m256 _v1_channel0 = _val_channel0;
-                                __m256 _v1_channel1 = _val_channel0;
-                                __m256 _v1_channel2 = _val_channel0;
-                                __m256 _v1_channel3 = _val_channel0;
-                                __m256 _v1_channel4 = _val_channel0;
-                                __m256 _v1_channel5 = _val_channel0;
-                                __m256 _v1_channel6 = _val_channel0;
-                                __m256 _v1_channel7 = _val_channel0;
-                                __m256 _v1_channel8 = _val_channel0;
-                                __m256 _v1_channel9 = _val_channel0;
-                                __m256 _v1_channela = _val_channel0;
-                                __m256 _v1_channelb = _val_channel0;
-                                __m256 _v1_channelc = _val_channel0;
-                                __m256 _v1_channeld = _val_channel0;
-                                __m256 _v1_channele = _val_channel0;
-                                __m256 _v1_channelf = _val_channel0;
-                                __m256 _v2_channel0 = _val_channel0;
-                                __m256 _v2_channel1 = _val_channel0;
-                                __m256 _v2_channel2 = _val_channel0;
-                                __m256 _v2_channel3 = _val_channel0;
-                                __m256 _v2_channel4 = _val_channel0;
-                                __m256 _v2_channel5 = _val_channel0;
-                                __m256 _v2_channel6 = _val_channel0;
-                                __m256 _v2_channel7 = _val_channel0;
-                                __m256 _v2_channel8 = _val_channel0;
-                                __m256 _v2_channel9 = _val_channel0;
-                                __m256 _v2_channela = _val_channel0;
-                                __m256 _v2_channelb = _val_channel0;
-                                __m256 _v2_channelc = _val_channel0;
-                                __m256 _v2_channeld = _val_channel0;
-                                __m256 _v2_channele = _val_channel0;
-                                __m256 _v2_channelf = _val_channel0;
-                                __m256 _v3_channel0 = _val_channel0;
-                                __m256 _v3_channel1 = _val_channel0;
-                                __m256 _v3_channel2 = _val_channel0;
-                                __m256 _v3_channel3 = _val_channel0;
-                                __m256 _v3_channel4 = _val_channel0;
-                                __m256 _v3_channel5 = _val_channel0;
-                                __m256 _v3_channel6 = _val_channel0;
-                                __m256 _v3_channel7 = _val_channel0;
-                                __m256 _v3_channel8 = _val_channel0;
-                                __m256 _v3_channel9 = _val_channel0;
-                                __m256 _v3_channela = _val_channel0;
-                                __m256 _v3_channelb = _val_channel0;
-                                __m256 _v3_channelc = _val_channel0;
-                                __m256 _v3_channeld = _val_channel0;
-                                __m256 _v3_channele = _val_channel0;
-                                __m256 _v3_channelf = _val_channel0;
-                                __m256 _v4_channel0 = _val_channel0;
-                                __m256 _v4_channel1 = _val_channel0;
-                                __m256 _v4_channel2 = _val_channel0;
-                                __m256 _v4_channel3 = _val_channel0;
-                                __m256 _v4_channel4 = _val_channel0;
-                                __m256 _v4_channel5 = _val_channel0;
-                                __m256 _v4_channel6 = _val_channel0;
-                                __m256 _v4_channel7 = _val_channel0;
-                                __m256 _v4_channel8 = _val_channel0;
-                                __m256 _v4_channel9 = _val_channel0;
-                                __m256 _v4_channela = _val_channel0;
-                                __m256 _v4_channelb = _val_channel0;
-                                __m256 _v4_channelc = _val_channel0;
-                                __m256 _v4_channeld = _val_channel0;
-                                __m256 _v4_channele = _val_channel0;
-                                __m256 _v4_channelf = _val_channel0;
-                                if (v1_cond)
-                                {
-                                    _v1_channel0 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack);
-                                    _v1_channel1 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 1);
-                                    _v1_channel2 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 2);
-                                    _v1_channel3 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 3);
-                                    _v1_channel4 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 4);
-                                    _v1_channel5 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 5);
-                                    _v1_channel6 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 6);
-                                    _v1_channel7 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 7);
-                                    _v1_channel8 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 8);
-                                    _v1_channel9 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 9);
-                                    _v1_channela = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 10);
-                                    _v1_channelb = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 11);
-                                    _v1_channelc = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 12);
-                                    _v1_channeld = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 13);
-                                    _v1_channele = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 14);
-                                    _v1_channelf = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 15);
-                                }
-                                if (v2_cond)
-                                {
-                                    _v2_channel0 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack);
-                                    _v2_channel1 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 1);
-                                    _v2_channel2 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 2);
-                                    _v2_channel3 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 3);
-                                    _v2_channel4 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 4);
-                                    _v2_channel5 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 5);
-                                    _v2_channel6 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 6);
-                                    _v2_channel7 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 7);
-                                    _v2_channel8 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 8);
-                                    _v2_channel9 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 9);
-                                    _v2_channela = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 10);
-                                    _v2_channelb = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 11);
-                                    _v2_channelc = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 12);
-                                    _v2_channeld = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 13);
-                                    _v2_channele = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 14);
-                                    _v2_channelf = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 15);
-                                }
-                                if (v3_cond)
-                                {
-                                    _v3_channel0 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack);
-                                    _v3_channel1 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 1);
-                                    _v3_channel2 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 2);
-                                    _v3_channel3 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 3);
-                                    _v3_channel4 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 4);
-                                    _v3_channel5 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 5);
-                                    _v3_channel6 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 6);
-                                    _v3_channel7 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 7);
-                                    _v3_channel8 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 8);
-                                    _v3_channel9 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 9);
-                                    _v3_channela = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 10);
-                                    _v3_channelb = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 11);
-                                    _v3_channelc = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 12);
-                                    _v3_channeld = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 13);
-                                    _v3_channele = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 14);
-                                    _v3_channelf = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 15);
-                                }
-                                if (v4_cond)
-                                {
-                                    _v4_channel0 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack);
-                                    _v4_channel1 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 1);
-                                    _v4_channel2 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 2);
-                                    _v4_channel3 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 3);
-                                    _v4_channel4 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 4);
-                                    _v4_channel5 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 5);
-                                    _v4_channel6 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 6);
-                                    _v4_channel7 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 7);
-                                    _v4_channel8 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 8);
-                                    _v4_channel9 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 9);
-                                    _v4_channela = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 10);
-                                    _v4_channelb = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 11);
-                                    _v4_channelc = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 12);
-                                    _v4_channeld = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 13);
-                                    _v4_channele = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 14);
-                                    _v4_channelf = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 15);
-                                }
-                                __m256 _w1 = _mm256_loadu_ps(w1_ptr);
-                                __m256 _w2 = _mm256_loadu_ps(w2_ptr);
-                                __m256 _w3 = _mm256_loadu_ps(w3_ptr);
-                                __m256 _w4 = _mm256_loadu_ps(w4_ptr);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v1_channel0, _w1, _val_channel0);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v2_channel0, _w2, _val_channel0);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v3_channel0, _w3, _val_channel0);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v4_channel0, _w4, _val_channel0);
-                                _val_channel1 = _mm256_comp_fmadd_ps(_v1_channel1, _w1, _val_channel1);
-                                _val_channel1 = _mm256_comp_fmadd_ps(_v2_channel1, _w2, _val_channel1);
-                                _val_channel1 = _mm256_comp_fmadd_ps(_v3_channel1, _w3, _val_channel1);
-                                _val_channel1 = _mm256_comp_fmadd_ps(_v4_channel1, _w4, _val_channel1);
-                                _val_channel2 = _mm256_comp_fmadd_ps(_v1_channel2, _w1, _val_channel2);
-                                _val_channel2 = _mm256_comp_fmadd_ps(_v2_channel2, _w2, _val_channel2);
-                                _val_channel2 = _mm256_comp_fmadd_ps(_v3_channel2, _w3, _val_channel2);
-                                _val_channel2 = _mm256_comp_fmadd_ps(_v4_channel2, _w4, _val_channel2);
-                                _val_channel3 = _mm256_comp_fmadd_ps(_v1_channel3, _w1, _val_channel3);
-                                _val_channel3 = _mm256_comp_fmadd_ps(_v2_channel3, _w2, _val_channel3);
-                                _val_channel3 = _mm256_comp_fmadd_ps(_v3_channel3, _w3, _val_channel3);
-                                _val_channel3 = _mm256_comp_fmadd_ps(_v4_channel3, _w4, _val_channel3);
-                                _val_channel4 = _mm256_comp_fmadd_ps(_v1_channel4, _w1, _val_channel4);
-                                _val_channel4 = _mm256_comp_fmadd_ps(_v2_channel4, _w2, _val_channel4);
-                                _val_channel4 = _mm256_comp_fmadd_ps(_v3_channel4, _w3, _val_channel4);
-                                _val_channel4 = _mm256_comp_fmadd_ps(_v4_channel4, _w4, _val_channel4);
-                                _val_channel5 = _mm256_comp_fmadd_ps(_v1_channel5, _w1, _val_channel5);
-                                _val_channel5 = _mm256_comp_fmadd_ps(_v2_channel5, _w2, _val_channel5);
-                                _val_channel5 = _mm256_comp_fmadd_ps(_v3_channel5, _w3, _val_channel5);
-                                _val_channel5 = _mm256_comp_fmadd_ps(_v4_channel5, _w4, _val_channel5);
-                                _val_channel6 = _mm256_comp_fmadd_ps(_v1_channel6, _w1, _val_channel6);
-                                _val_channel6 = _mm256_comp_fmadd_ps(_v2_channel6, _w2, _val_channel6);
-                                _val_channel6 = _mm256_comp_fmadd_ps(_v3_channel6, _w3, _val_channel6);
-                                _val_channel6 = _mm256_comp_fmadd_ps(_v4_channel6, _w4, _val_channel6);
-                                _val_channel7 = _mm256_comp_fmadd_ps(_v1_channel7, _w1, _val_channel7);
-                                _val_channel7 = _mm256_comp_fmadd_ps(_v2_channel7, _w2, _val_channel7);
-                                _val_channel7 = _mm256_comp_fmadd_ps(_v3_channel7, _w3, _val_channel7);
-                                _val_channel7 = _mm256_comp_fmadd_ps(_v4_channel7, _w4, _val_channel7);
-                                _val_channel8 = _mm256_comp_fmadd_ps(_v1_channel8, _w1, _val_channel8);
-                                _val_channel8 = _mm256_comp_fmadd_ps(_v2_channel8, _w2, _val_channel8);
-                                _val_channel8 = _mm256_comp_fmadd_ps(_v3_channel8, _w3, _val_channel8);
-                                _val_channel8 = _mm256_comp_fmadd_ps(_v4_channel8, _w4, _val_channel8);
-                                _val_channel9 = _mm256_comp_fmadd_ps(_v1_channel9, _w1, _val_channel9);
-                                _val_channel9 = _mm256_comp_fmadd_ps(_v2_channel9, _w2, _val_channel9);
-                                _val_channel9 = _mm256_comp_fmadd_ps(_v3_channel9, _w3, _val_channel9);
-                                _val_channel9 = _mm256_comp_fmadd_ps(_v4_channel9, _w4, _val_channel9);
-                                _val_channela = _mm256_comp_fmadd_ps(_v1_channela, _w1, _val_channela);
-                                _val_channela = _mm256_comp_fmadd_ps(_v2_channela, _w2, _val_channela);
-                                _val_channela = _mm256_comp_fmadd_ps(_v3_channela, _w3, _val_channela);
-                                _val_channela = _mm256_comp_fmadd_ps(_v4_channela, _w4, _val_channela);
-                                _val_channelb = _mm256_comp_fmadd_ps(_v1_channelb, _w1, _val_channelb);
-                                _val_channelb = _mm256_comp_fmadd_ps(_v2_channelb, _w2, _val_channelb);
-                                _val_channelb = _mm256_comp_fmadd_ps(_v3_channelb, _w3, _val_channelb);
-                                _val_channelb = _mm256_comp_fmadd_ps(_v4_channelb, _w4, _val_channelb);
-                                _val_channelc = _mm256_comp_fmadd_ps(_v1_channelc, _w1, _val_channelc);
-                                _val_channelc = _mm256_comp_fmadd_ps(_v2_channelc, _w2, _val_channelc);
-                                _val_channelc = _mm256_comp_fmadd_ps(_v3_channelc, _w3, _val_channelc);
-                                _val_channelc = _mm256_comp_fmadd_ps(_v4_channelc, _w4, _val_channelc);
-                                _val_channeld = _mm256_comp_fmadd_ps(_v1_channeld, _w1, _val_channeld);
-                                _val_channeld = _mm256_comp_fmadd_ps(_v2_channeld, _w2, _val_channeld);
-                                _val_channeld = _mm256_comp_fmadd_ps(_v3_channeld, _w3, _val_channeld);
-                                _val_channeld = _mm256_comp_fmadd_ps(_v4_channeld, _w4, _val_channeld);
-                                _val_channele = _mm256_comp_fmadd_ps(_v1_channele, _w1, _val_channele);
-                                _val_channele = _mm256_comp_fmadd_ps(_v2_channele, _w2, _val_channele);
-                                _val_channele = _mm256_comp_fmadd_ps(_v3_channele, _w3, _val_channele);
-                                _val_channele = _mm256_comp_fmadd_ps(_v4_channele, _w4, _val_channele);
-                                _val_channelf = _mm256_comp_fmadd_ps(_v1_channelf, _w1, _val_channelf);
-                                _val_channelf = _mm256_comp_fmadd_ps(_v2_channelf, _w2, _val_channelf);
-                                _val_channelf = _mm256_comp_fmadd_ps(_v3_channelf, _w3, _val_channelf);
-                                _val_channelf = _mm256_comp_fmadd_ps(_v4_channelf, _w4, _val_channelf);
-                            }
-                            if (has_mask)
-                            {
-                                __m256 _mask = _mm256_loadu_ps(mask_ptr);
-                                _val_channel0 = _mm256_mul_ps(_val_channel0, _mask);
-                                _val_channel1 = _mm256_mul_ps(_val_channel1, _mask);
-                                _val_channel2 = _mm256_mul_ps(_val_channel2, _mask);
-                                _val_channel3 = _mm256_mul_ps(_val_channel3, _mask);
-                                _val_channel4 = _mm256_mul_ps(_val_channel4, _mask);
-                                _val_channel5 = _mm256_mul_ps(_val_channel5, _mask);
-                                _val_channel6 = _mm256_mul_ps(_val_channel6, _mask);
-                                _val_channel7 = _mm256_mul_ps(_val_channel7, _mask);
-                                _val_channel8 = _mm256_mul_ps(_val_channel8, _mask);
-                                _val_channel9 = _mm256_mul_ps(_val_channel9, _mask);
-                                _val_channela = _mm256_mul_ps(_val_channela, _mask);
-                                _val_channelb = _mm256_mul_ps(_val_channelb, _mask);
-                                _val_channelc = _mm256_mul_ps(_val_channelc, _mask);
-                                _val_channeld = _mm256_mul_ps(_val_channeld, _mask);
-                                _val_channele = _mm256_mul_ps(_val_channele, _mask);
-                                _val_channelf = _mm256_mul_ps(_val_channelf, _mask);
-                            }
-                            __m256 _conv_w0 = _mm256_load_ps(kptr);
-                            __m256 _conv_w1 = _mm256_load_ps(kptr + out_elempack); // 1 * out_elempack
-                            _sum = _mm256_comp_fmadd_ps(_val_channel0, _conv_w0, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val_channel1, _conv_w1, _sum);
-                            __m256 _conv_w2 = _mm256_load_ps(kptr + 16); // 2 * out_elempack
-                            __m256 _conv_w3 = _mm256_load_ps(kptr + 24); // 3 * out_elempack
-                            _sum = _mm256_comp_fmadd_ps(_val_channel2, _conv_w2, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val_channel3, _conv_w3, _sum);
-                            __m256 _conv_w4 = _mm256_load_ps(kptr + 32); // 4 * out_elempack
-                            __m256 _conv_w5 = _mm256_load_ps(kptr + 40); // 5 * out_elempack
-                            _sum = _mm256_comp_fmadd_ps(_val_channel4, _conv_w4, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val_channel5, _conv_w5, _sum);
-                            __m256 _conv_w6 = _mm256_load_ps(kptr + 48); // 6 * out_elempack
-                            __m256 _conv_w7 = _mm256_load_ps(kptr + 56); // 7 * out_elempack
-                            _sum = _mm256_comp_fmadd_ps(_val_channel6, _conv_w6, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val_channel7, _conv_w7, _sum);
-                            __m256 _conv_w8 = _mm256_load_ps(kptr + 64); // 8 * out_elempack
-                            __m256 _conv_w9 = _mm256_load_ps(kptr + 72); // 9 * out_elempack
-                            _sum = _mm256_comp_fmadd_ps(_val_channel8, _conv_w8, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val_channel9, _conv_w9, _sum);
-                            __m256 _conv_wa = _mm256_load_ps(kptr + 80); // 10 * out_elempack
-                            __m256 _conv_wb = _mm256_load_ps(kptr + 88); // 11 * out_elempack
-                            _sum = _mm256_comp_fmadd_ps(_val_channela, _conv_wa, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val_channelb, _conv_wb, _sum);
-                            __m256 _conv_wc = _mm256_load_ps(kptr + 96);  // 12 * out_elempack
-                            __m256 _conv_wd = _mm256_load_ps(kptr + 104); // 13 * out_elempack
-                            _sum = _mm256_comp_fmadd_ps(_val_channelc, _conv_wc, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val_channeld, _conv_wd, _sum);
-                            __m256 _conv_we = _mm256_load_ps(kptr + 112); // 14 * out_elempack
-                            __m256 _conv_wf = _mm256_load_ps(kptr + 120); // 15 * out_elempack
-                            _sum = _mm256_comp_fmadd_ps(_val_channele, _conv_we, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val_channelf, _conv_wf, _sum);
-                            kptr += wstep;
-                        }
-                    }
-                }
-                _sum = activation_avx(_sum, activation_type, activation_params);
-                _mm256_store_ps(outptr + (h_col * outw + w_col) * out_elempack, _sum);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack1to16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack1to16.h
deleted file mode 100644
index d7e4a64f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack1to16.h
+++ /dev/null
@@ -1,195 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deformableconv2d_pack1to16_avx512(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const Mat& bottom_blob = bottom_blobs[0];
-    const Mat& offset = bottom_blobs[1];
-    const bool has_mask = (bottom_blobs.size() == 3);
-    const bool offset_not_pack = offset.elempack == 1;
-    const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const int size = outw * outh;
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-    const int elempack = 1;
-    const int out_elempack = 16;
-    const int wstep = out_elempack * elempack;
-    const float zeros[out_elempack] = {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
-    const float* zeros_ptr = zeros;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int h_col = 0; h_col < outh; h_col++)
-    {
-        for (int w_col = 0; w_col < outw; w_col++)
-        {
-            int h_in = h_col * stride_h - pad_top;
-            int w_in = w_col * stride_w - pad_left;
-            for (int oc = 0; oc < outch; oc++)
-            {
-                const float* kptr = weight_data_packed.channel(oc);
-                float* outptr = top_blob.channel(oc);
-                __m512 _sum = _mm512_setzero_ps();
-                if (bias_data_ptr)
-                    _sum = _mm512_loadu_ps(bias_data_ptr + oc * out_elempack);
-                for (int i = 0; i < kernel_h; i++)
-                {
-                    for (int j = 0; j < kernel_w; j++)
-                    {
-                        float offset_h = 0.f;
-                        float offset_w = 0.f;
-                        float mask_ = 1.f;
-                        if (offset_not_pack)
-                        {
-                            offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
-                            offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
-                        }
-                        else
-                        {
-                            const int y_c = (i * kernel_w + j) * 2;
-                            const int x_c = (i * kernel_w + j) * 2 + 1;
-                            offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
-                            offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
-                        }
-                        if (has_mask)
-                        {
-                            const Mat& mask = bottom_blobs[2];
-                            if (mask_not_pack)
-                            {
-                                mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
-                            }
-                            else
-                            {
-                                const int m_c = i * kernel_w + j;
-                                mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
-                            }
-                        }
-                        const float h_im = h_in + i * dilation_h + offset_h;
-                        const float w_im = w_in + j * dilation_w + offset_w;
-
-                        // Bilinear
-                        const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
-                        float w1 = 0.f;
-                        float w2 = 0.f;
-                        float w3 = 0.f;
-                        float w4 = 0.f;
-                        bool v1_cond = false;
-                        bool v2_cond = false;
-                        bool v3_cond = false;
-                        bool v4_cond = false;
-                        int v1_pos = 0;
-                        int v2_pos = 0;
-                        int v3_pos = 0;
-                        int v4_pos = 0;
-                        if (cond)
-                        {
-                            int h_low = (int)floorf(h_im);
-                            int w_low = (int)floorf(w_im);
-                            int h_high = h_low + 1;
-                            int w_high = w_low + 1;
-
-                            float lh = h_im - h_low;
-                            float lw = w_im - w_low;
-                            float hh = 1 - lh;
-                            float hw = 1 - lw;
-
-                            v1_cond = (h_low >= 0 && w_low >= 0);
-                            v2_cond = (h_low >= 0 && w_high <= w - 1);
-                            v3_cond = (h_high <= h - 1 && w_low >= 0);
-                            v4_cond = (h_high <= h - 1 && w_high <= w - 1);
-                            if (v1_cond)
-                                v1_pos = h_low * w + w_low;
-                            if (v2_cond)
-                                v2_pos = h_low * w + w_high;
-                            if (v3_cond)
-                                v3_pos = h_high * w + w_low;
-                            if (v4_cond)
-                                v4_pos = h_high * w + w_high;
-
-                            w1 = hh * hw;
-                            w2 = hh * lw;
-                            w3 = lh * hw;
-                            w4 = lh * lw;
-                        }
-                        const float w1s[out_elempack] = {w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1};
-                        const float* w1_ptr = w1s;
-                        const float w2s[out_elempack] = {w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2};
-                        const float* w2_ptr = w2s;
-                        const float w3s[out_elempack] = {w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3};
-                        const float* w3_ptr = w3s;
-                        const float w4s[out_elempack] = {w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4};
-                        const float* w4_ptr = w4s;
-                        const float masks[out_elempack] = {mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_};
-                        const float* mask_ptr = masks;
-
-                        for (int ic = 0; ic < inch; ic++)
-                        {
-                            const float* data_im_ptr = bottom_blob.channel(ic);
-                            __m512 _val_channel0 = _mm512_loadu_ps(zeros_ptr);
-                            if (cond)
-                            {
-                                __m512 _v1_channel0 = _val_channel0;
-                                __m512 _v2_channel0 = _val_channel0;
-                                __m512 _v3_channel0 = _val_channel0;
-                                __m512 _v4_channel0 = _val_channel0;
-                                if (v1_cond)
-                                {
-                                    _v1_channel0 = _mm512_set1_ps(data_im_ptr[v1_pos]);
-                                }
-                                if (v2_cond)
-                                {
-                                    _v2_channel0 = _mm512_set1_ps(data_im_ptr[v2_pos]);
-                                }
-                                if (v3_cond)
-                                {
-                                    _v3_channel0 = _mm512_set1_ps(data_im_ptr[v3_pos]);
-                                }
-                                if (v4_cond)
-                                {
-                                    _v4_channel0 = _mm512_set1_ps(data_im_ptr[v4_pos]);
-                                }
-                                __m512 _w1 = _mm512_loadu_ps(w1_ptr);
-                                __m512 _w2 = _mm512_loadu_ps(w2_ptr);
-                                __m512 _w3 = _mm512_loadu_ps(w3_ptr);
-                                __m512 _w4 = _mm512_loadu_ps(w4_ptr);
-                                _val_channel0 = _mm512_fmadd_ps(_v1_channel0, _w1, _val_channel0);
-                                _val_channel0 = _mm512_fmadd_ps(_v2_channel0, _w2, _val_channel0);
-                                _val_channel0 = _mm512_fmadd_ps(_v3_channel0, _w3, _val_channel0);
-                                _val_channel0 = _mm512_fmadd_ps(_v4_channel0, _w4, _val_channel0);
-                            }
-                            if (has_mask)
-                            {
-                                __m512 _mask = _mm512_loadu_ps(mask_ptr);
-                                _val_channel0 = _mm512_mul_ps(_val_channel0, _mask);
-                            }
-                            __m512 _conv_w0 = _mm512_load_ps(kptr);
-                            _sum = _mm512_fmadd_ps(_val_channel0, _conv_w0, _sum);
-                            kptr += wstep;
-                        }
-                    }
-                }
-                _sum = activation_avx512(_sum, activation_type, activation_params);
-                _mm512_store_ps(outptr + (h_col * outw + w_col) * out_elempack, _sum);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack1to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack1to4.h
deleted file mode 100644
index 512cf3e6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack1to4.h
+++ /dev/null
@@ -1,195 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deformableconv2d_pack1to4_sse(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const Mat& bottom_blob = bottom_blobs[0];
-    const Mat& offset = bottom_blobs[1];
-    const bool has_mask = (bottom_blobs.size() == 3);
-    const bool offset_not_pack = offset.elempack == 1;
-    const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const int size = outw * outh;
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-    const int elempack = 1;
-    const int out_elempack = 4;
-    const int wstep = out_elempack * elempack;
-    const float zeros[out_elempack] = {0.f, 0.f, 0.f, 0.f};
-    const float* zeros_ptr = zeros;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int h_col = 0; h_col < outh; h_col++)
-    {
-        for (int w_col = 0; w_col < outw; w_col++)
-        {
-            int h_in = h_col * stride_h - pad_top;
-            int w_in = w_col * stride_w - pad_left;
-            for (int oc = 0; oc < outch; oc++)
-            {
-                const float* kptr = weight_data_packed.channel(oc);
-                float* outptr = top_blob.channel(oc);
-                __m128 _sum = _mm_setzero_ps();
-                if (bias_data_ptr)
-                    _sum = _mm_loadu_ps(bias_data_ptr + oc * out_elempack);
-                for (int i = 0; i < kernel_h; i++)
-                {
-                    for (int j = 0; j < kernel_w; j++)
-                    {
-                        float offset_h = 0.f;
-                        float offset_w = 0.f;
-                        float mask_ = 1.f;
-                        if (offset_not_pack)
-                        {
-                            offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
-                            offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
-                        }
-                        else
-                        {
-                            const int y_c = (i * kernel_w + j) * 2;
-                            const int x_c = (i * kernel_w + j) * 2 + 1;
-                            offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
-                            offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
-                        }
-                        if (has_mask)
-                        {
-                            const Mat& mask = bottom_blobs[2];
-                            if (mask_not_pack)
-                            {
-                                mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
-                            }
-                            else
-                            {
-                                const int m_c = i * kernel_w + j;
-                                mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
-                            }
-                        }
-                        const float h_im = h_in + i * dilation_h + offset_h;
-                        const float w_im = w_in + j * dilation_w + offset_w;
-
-                        // Bilinear
-                        const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
-                        float w1 = 0.f;
-                        float w2 = 0.f;
-                        float w3 = 0.f;
-                        float w4 = 0.f;
-                        bool v1_cond = false;
-                        bool v2_cond = false;
-                        bool v3_cond = false;
-                        bool v4_cond = false;
-                        int v1_pos = 0;
-                        int v2_pos = 0;
-                        int v3_pos = 0;
-                        int v4_pos = 0;
-                        if (cond)
-                        {
-                            int h_low = (int)floorf(h_im);
-                            int w_low = (int)floorf(w_im);
-                            int h_high = h_low + 1;
-                            int w_high = w_low + 1;
-
-                            float lh = h_im - h_low;
-                            float lw = w_im - w_low;
-                            float hh = 1 - lh;
-                            float hw = 1 - lw;
-
-                            v1_cond = (h_low >= 0 && w_low >= 0);
-                            v2_cond = (h_low >= 0 && w_high <= w - 1);
-                            v3_cond = (h_high <= h - 1 && w_low >= 0);
-                            v4_cond = (h_high <= h - 1 && w_high <= w - 1);
-                            if (v1_cond)
-                                v1_pos = h_low * w + w_low;
-                            if (v2_cond)
-                                v2_pos = h_low * w + w_high;
-                            if (v3_cond)
-                                v3_pos = h_high * w + w_low;
-                            if (v4_cond)
-                                v4_pos = h_high * w + w_high;
-
-                            w1 = hh * hw;
-                            w2 = hh * lw;
-                            w3 = lh * hw;
-                            w4 = lh * lw;
-                        }
-                        const float w1s[out_elempack] = {w1, w1, w1, w1};
-                        const float* w1_ptr = w1s;
-                        const float w2s[out_elempack] = {w2, w2, w2, w2};
-                        const float* w2_ptr = w2s;
-                        const float w3s[out_elempack] = {w3, w3, w3, w3};
-                        const float* w3_ptr = w3s;
-                        const float w4s[out_elempack] = {w4, w4, w4, w4};
-                        const float* w4_ptr = w4s;
-                        const float masks[out_elempack] = {mask_, mask_, mask_, mask_};
-                        const float* mask_ptr = masks;
-
-                        for (int ic = 0; ic < inch; ic++)
-                        {
-                            const float* data_im_ptr = bottom_blob.channel(ic);
-                            __m128 _val_channel0 = _mm_loadu_ps(zeros_ptr);
-                            if (cond)
-                            {
-                                __m128 _v1_channel0 = _val_channel0;
-                                __m128 _v2_channel0 = _val_channel0;
-                                __m128 _v3_channel0 = _val_channel0;
-                                __m128 _v4_channel0 = _val_channel0;
-                                if (v1_cond)
-                                {
-                                    _v1_channel0 = _mm_load1_ps(data_im_ptr + v1_pos);
-                                }
-                                if (v2_cond)
-                                {
-                                    _v2_channel0 = _mm_load1_ps(data_im_ptr + v2_pos);
-                                }
-                                if (v3_cond)
-                                {
-                                    _v3_channel0 = _mm_load1_ps(data_im_ptr + v3_pos);
-                                }
-                                if (v4_cond)
-                                {
-                                    _v4_channel0 = _mm_load1_ps(data_im_ptr + v4_pos);
-                                }
-                                __m128 _w1 = _mm_loadu_ps(w1_ptr);
-                                __m128 _w2 = _mm_loadu_ps(w2_ptr);
-                                __m128 _w3 = _mm_loadu_ps(w3_ptr);
-                                __m128 _w4 = _mm_loadu_ps(w4_ptr);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v1_channel0, _w1, _val_channel0);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v2_channel0, _w2, _val_channel0);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v3_channel0, _w3, _val_channel0);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v4_channel0, _w4, _val_channel0);
-                            }
-                            if (has_mask)
-                            {
-                                __m128 _mask = _mm_loadu_ps(mask_ptr);
-                                _val_channel0 = _mm_mul_ps(_val_channel0, _mask);
-                            }
-                            __m128 _conv_w0 = _mm_load_ps(kptr);
-                            _sum = _mm_comp_fmadd_ps(_val_channel0, _conv_w0, _sum);
-                            kptr += wstep;
-                        }
-                    }
-                }
-                _sum = activation_sse(_sum, activation_type, activation_params);
-                _mm_storeu_ps(outptr + (h_col * outw + w_col) * out_elempack, _sum);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack1to8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack1to8.h
deleted file mode 100644
index 2db40aa6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack1to8.h
+++ /dev/null
@@ -1,195 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deformableconv2d_pack1to8_avx(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const Mat& bottom_blob = bottom_blobs[0];
-    const Mat& offset = bottom_blobs[1];
-    const bool has_mask = (bottom_blobs.size() == 3);
-    const bool offset_not_pack = offset.elempack == 1;
-    const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const int size = outw * outh;
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-    const int elempack = 1;
-    const int out_elempack = 8;
-    const int wstep = out_elempack * elempack;
-    const float zeros[out_elempack] = {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
-    const float* zeros_ptr = zeros;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int h_col = 0; h_col < outh; h_col++)
-    {
-        for (int w_col = 0; w_col < outw; w_col++)
-        {
-            int h_in = h_col * stride_h - pad_top;
-            int w_in = w_col * stride_w - pad_left;
-            for (int oc = 0; oc < outch; oc++)
-            {
-                const float* kptr = weight_data_packed.channel(oc);
-                float* outptr = top_blob.channel(oc);
-                __m256 _sum = _mm256_setzero_ps();
-                if (bias_data_ptr)
-                    _sum = _mm256_loadu_ps(bias_data_ptr + oc * out_elempack);
-                for (int i = 0; i < kernel_h; i++)
-                {
-                    for (int j = 0; j < kernel_w; j++)
-                    {
-                        float offset_h = 0.f;
-                        float offset_w = 0.f;
-                        float mask_ = 1.f;
-                        if (offset_not_pack)
-                        {
-                            offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
-                            offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
-                        }
-                        else
-                        {
-                            const int y_c = (i * kernel_w + j) * 2;
-                            const int x_c = (i * kernel_w + j) * 2 + 1;
-                            offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
-                            offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
-                        }
-                        if (has_mask)
-                        {
-                            const Mat& mask = bottom_blobs[2];
-                            if (mask_not_pack)
-                            {
-                                mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
-                            }
-                            else
-                            {
-                                const int m_c = i * kernel_w + j;
-                                mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
-                            }
-                        }
-                        const float h_im = h_in + i * dilation_h + offset_h;
-                        const float w_im = w_in + j * dilation_w + offset_w;
-
-                        // Bilinear
-                        const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
-                        float w1 = 0.f;
-                        float w2 = 0.f;
-                        float w3 = 0.f;
-                        float w4 = 0.f;
-                        bool v1_cond = false;
-                        bool v2_cond = false;
-                        bool v3_cond = false;
-                        bool v4_cond = false;
-                        int v1_pos = 0;
-                        int v2_pos = 0;
-                        int v3_pos = 0;
-                        int v4_pos = 0;
-                        if (cond)
-                        {
-                            int h_low = (int)floorf(h_im);
-                            int w_low = (int)floorf(w_im);
-                            int h_high = h_low + 1;
-                            int w_high = w_low + 1;
-
-                            float lh = h_im - h_low;
-                            float lw = w_im - w_low;
-                            float hh = 1 - lh;
-                            float hw = 1 - lw;
-
-                            v1_cond = (h_low >= 0 && w_low >= 0);
-                            v2_cond = (h_low >= 0 && w_high <= w - 1);
-                            v3_cond = (h_high <= h - 1 && w_low >= 0);
-                            v4_cond = (h_high <= h - 1 && w_high <= w - 1);
-                            if (v1_cond)
-                                v1_pos = h_low * w + w_low;
-                            if (v2_cond)
-                                v2_pos = h_low * w + w_high;
-                            if (v3_cond)
-                                v3_pos = h_high * w + w_low;
-                            if (v4_cond)
-                                v4_pos = h_high * w + w_high;
-
-                            w1 = hh * hw;
-                            w2 = hh * lw;
-                            w3 = lh * hw;
-                            w4 = lh * lw;
-                        }
-                        const float w1s[out_elempack] = {w1, w1, w1, w1, w1, w1, w1, w1};
-                        const float* w1_ptr = w1s;
-                        const float w2s[out_elempack] = {w2, w2, w2, w2, w2, w2, w2, w2};
-                        const float* w2_ptr = w2s;
-                        const float w3s[out_elempack] = {w3, w3, w3, w3, w3, w3, w3, w3};
-                        const float* w3_ptr = w3s;
-                        const float w4s[out_elempack] = {w4, w4, w4, w4, w4, w4, w4, w4};
-                        const float* w4_ptr = w4s;
-                        const float masks[out_elempack] = {mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_};
-                        const float* mask_ptr = masks;
-
-                        for (int ic = 0; ic < inch; ic++)
-                        {
-                            const float* data_im_ptr = bottom_blob.channel(ic);
-                            __m256 _val_channel0 = _mm256_loadu_ps(zeros_ptr);
-                            if (cond)
-                            {
-                                __m256 _v1_channel0 = _val_channel0;
-                                __m256 _v2_channel0 = _val_channel0;
-                                __m256 _v3_channel0 = _val_channel0;
-                                __m256 _v4_channel0 = _val_channel0;
-                                if (v1_cond)
-                                {
-                                    _v1_channel0 = _mm256_broadcast_ss(data_im_ptr + v1_pos);
-                                }
-                                if (v2_cond)
-                                {
-                                    _v2_channel0 = _mm256_broadcast_ss(data_im_ptr + v2_pos);
-                                }
-                                if (v3_cond)
-                                {
-                                    _v3_channel0 = _mm256_broadcast_ss(data_im_ptr + v3_pos);
-                                }
-                                if (v4_cond)
-                                {
-                                    _v4_channel0 = _mm256_broadcast_ss(data_im_ptr + v4_pos);
-                                }
-                                __m256 _w1 = _mm256_loadu_ps(w1_ptr);
-                                __m256 _w2 = _mm256_loadu_ps(w2_ptr);
-                                __m256 _w3 = _mm256_loadu_ps(w3_ptr);
-                                __m256 _w4 = _mm256_loadu_ps(w4_ptr);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v1_channel0, _w1, _val_channel0);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v2_channel0, _w2, _val_channel0);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v3_channel0, _w3, _val_channel0);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v4_channel0, _w4, _val_channel0);
-                            }
-                            if (has_mask)
-                            {
-                                __m256 _mask = _mm256_loadu_ps(mask_ptr);
-                                _val_channel0 = _mm256_mul_ps(_val_channel0, _mask);
-                            }
-                            __m256 _conv_w0 = _mm256_load_ps(kptr);
-                            _sum = _mm256_comp_fmadd_ps(_val_channel0, _conv_w0, _sum);
-                            kptr += wstep;
-                        }
-                    }
-                }
-                _sum = activation_avx(_sum, activation_type, activation_params);
-                _mm256_store_ps(outptr + (h_col * outw + w_col) * out_elempack, _sum);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4.h
deleted file mode 100644
index e9b38fd8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4.h
+++ /dev/null
@@ -1,243 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deformableconv2d_pack4_sse(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const Mat& bottom_blob = bottom_blobs[0];
-    const Mat& offset = bottom_blobs[1];
-    const bool has_mask = (bottom_blobs.size() == 3);
-    const bool offset_not_pack = offset.elempack == 1;
-    const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const int size = outw * outh;
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-    const int elempack = 4;
-    const int out_elempack = 4;
-    const int wstep = out_elempack * elempack;
-    const float zeros[out_elempack] = {0.f, 0.f, 0.f, 0.f};
-    const float* zeros_ptr = zeros;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int h_col = 0; h_col < outh; h_col++)
-    {
-        for (int w_col = 0; w_col < outw; w_col++)
-        {
-            int h_in = h_col * stride_h - pad_top;
-            int w_in = w_col * stride_w - pad_left;
-            for (int oc = 0; oc < outch; oc++)
-            {
-                const float* kptr = weight_data_packed.channel(oc);
-                float* outptr = top_blob.channel(oc);
-                __m128 _sum = _mm_setzero_ps();
-                if (bias_data_ptr)
-                    _sum = _mm_loadu_ps(bias_data_ptr + oc * out_elempack);
-                for (int i = 0; i < kernel_h; i++)
-                {
-                    for (int j = 0; j < kernel_w; j++)
-                    {
-                        float offset_h = 0.f;
-                        float offset_w = 0.f;
-                        float mask_ = 1.f;
-                        if (offset_not_pack)
-                        {
-                            offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
-                            offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
-                        }
-                        else
-                        {
-                            const int y_c = (i * kernel_w + j) * 2;
-                            const int x_c = (i * kernel_w + j) * 2 + 1;
-                            offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
-                            offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
-                        }
-                        if (has_mask)
-                        {
-                            const Mat& mask = bottom_blobs[2];
-                            if (mask_not_pack)
-                            {
-                                mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
-                            }
-                            else
-                            {
-                                const int m_c = i * kernel_w + j;
-                                mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
-                            }
-                        }
-                        const float h_im = h_in + i * dilation_h + offset_h;
-                        const float w_im = w_in + j * dilation_w + offset_w;
-
-                        // Bilinear
-                        const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
-                        float w1 = 0.f;
-                        float w2 = 0.f;
-                        float w3 = 0.f;
-                        float w4 = 0.f;
-                        bool v1_cond = false;
-                        bool v2_cond = false;
-                        bool v3_cond = false;
-                        bool v4_cond = false;
-                        int v1_pos = 0;
-                        int v2_pos = 0;
-                        int v3_pos = 0;
-                        int v4_pos = 0;
-                        if (cond)
-                        {
-                            int h_low = (int)floorf(h_im);
-                            int w_low = (int)floorf(w_im);
-                            int h_high = h_low + 1;
-                            int w_high = w_low + 1;
-
-                            float lh = h_im - h_low;
-                            float lw = w_im - w_low;
-                            float hh = 1 - lh;
-                            float hw = 1 - lw;
-
-                            v1_cond = (h_low >= 0 && w_low >= 0);
-                            v2_cond = (h_low >= 0 && w_high <= w - 1);
-                            v3_cond = (h_high <= h - 1 && w_low >= 0);
-                            v4_cond = (h_high <= h - 1 && w_high <= w - 1);
-                            if (v1_cond)
-                                v1_pos = h_low * w + w_low;
-                            if (v2_cond)
-                                v2_pos = h_low * w + w_high;
-                            if (v3_cond)
-                                v3_pos = h_high * w + w_low;
-                            if (v4_cond)
-                                v4_pos = h_high * w + w_high;
-
-                            w1 = hh * hw;
-                            w2 = hh * lw;
-                            w3 = lh * hw;
-                            w4 = lh * lw;
-                        }
-                        const float w1s[out_elempack] = {w1, w1, w1, w1};
-                        const float* w1_ptr = w1s;
-                        const float w2s[out_elempack] = {w2, w2, w2, w2};
-                        const float* w2_ptr = w2s;
-                        const float w3s[out_elempack] = {w3, w3, w3, w3};
-                        const float* w3_ptr = w3s;
-                        const float w4s[out_elempack] = {w4, w4, w4, w4};
-                        const float* w4_ptr = w4s;
-                        const float masks[out_elempack] = {mask_, mask_, mask_, mask_};
-                        const float* mask_ptr = masks;
-
-                        for (int ic = 0; ic < inch; ic++)
-                        {
-                            const float* data_im_ptr = bottom_blob.channel(ic);
-                            __m128 _val_channel0 = _mm_loadu_ps(zeros_ptr);
-                            __m128 _val_channel1 = _val_channel0;
-                            __m128 _val_channel2 = _val_channel0;
-                            __m128 _val_channel3 = _val_channel0;
-                            if (cond)
-                            {
-                                __m128 _v1_channel0 = _val_channel0;
-                                __m128 _v1_channel1 = _val_channel0;
-                                __m128 _v1_channel2 = _val_channel0;
-                                __m128 _v1_channel3 = _val_channel0;
-                                __m128 _v2_channel0 = _val_channel0;
-                                __m128 _v2_channel1 = _val_channel0;
-                                __m128 _v2_channel2 = _val_channel0;
-                                __m128 _v2_channel3 = _val_channel0;
-                                __m128 _v3_channel0 = _val_channel0;
-                                __m128 _v3_channel1 = _val_channel0;
-                                __m128 _v3_channel2 = _val_channel0;
-                                __m128 _v3_channel3 = _val_channel0;
-                                __m128 _v4_channel0 = _val_channel0;
-                                __m128 _v4_channel1 = _val_channel0;
-                                __m128 _v4_channel2 = _val_channel0;
-                                __m128 _v4_channel3 = _val_channel0;
-                                if (v1_cond)
-                                {
-                                    _v1_channel0 = _mm_load1_ps(data_im_ptr + v1_pos * elempack);
-                                    _v1_channel1 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 1);
-                                    _v1_channel2 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 2);
-                                    _v1_channel3 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 3);
-                                }
-                                if (v2_cond)
-                                {
-                                    _v2_channel0 = _mm_load1_ps(data_im_ptr + v2_pos * elempack);
-                                    _v2_channel1 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 1);
-                                    _v2_channel2 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 2);
-                                    _v2_channel3 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 3);
-                                }
-                                if (v3_cond)
-                                {
-                                    _v3_channel0 = _mm_load1_ps(data_im_ptr + v3_pos * elempack);
-                                    _v3_channel1 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 1);
-                                    _v3_channel2 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 2);
-                                    _v3_channel3 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 3);
-                                }
-                                if (v4_cond)
-                                {
-                                    _v4_channel0 = _mm_load1_ps(data_im_ptr + v4_pos * elempack);
-                                    _v4_channel1 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 1);
-                                    _v4_channel2 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 2);
-                                    _v4_channel3 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 3);
-                                }
-                                __m128 _w1 = _mm_loadu_ps(w1_ptr);
-                                __m128 _w2 = _mm_loadu_ps(w2_ptr);
-                                __m128 _w3 = _mm_loadu_ps(w3_ptr);
-                                __m128 _w4 = _mm_loadu_ps(w4_ptr);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v1_channel0, _w1, _val_channel0);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v2_channel0, _w2, _val_channel0);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v3_channel0, _w3, _val_channel0);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v4_channel0, _w4, _val_channel0);
-                                _val_channel1 = _mm_comp_fmadd_ps(_v1_channel1, _w1, _val_channel1);
-                                _val_channel1 = _mm_comp_fmadd_ps(_v2_channel1, _w2, _val_channel1);
-                                _val_channel1 = _mm_comp_fmadd_ps(_v3_channel1, _w3, _val_channel1);
-                                _val_channel1 = _mm_comp_fmadd_ps(_v4_channel1, _w4, _val_channel1);
-                                _val_channel2 = _mm_comp_fmadd_ps(_v1_channel2, _w1, _val_channel2);
-                                _val_channel2 = _mm_comp_fmadd_ps(_v2_channel2, _w2, _val_channel2);
-                                _val_channel2 = _mm_comp_fmadd_ps(_v3_channel2, _w3, _val_channel2);
-                                _val_channel2 = _mm_comp_fmadd_ps(_v4_channel2, _w4, _val_channel2);
-                                _val_channel3 = _mm_comp_fmadd_ps(_v1_channel3, _w1, _val_channel3);
-                                _val_channel3 = _mm_comp_fmadd_ps(_v2_channel3, _w2, _val_channel3);
-                                _val_channel3 = _mm_comp_fmadd_ps(_v3_channel3, _w3, _val_channel3);
-                                _val_channel3 = _mm_comp_fmadd_ps(_v4_channel3, _w4, _val_channel3);
-                            }
-                            if (has_mask)
-                            {
-                                __m128 _mask = _mm_loadu_ps(mask_ptr);
-                                _val_channel0 = _mm_mul_ps(_val_channel0, _mask);
-                                _val_channel1 = _mm_mul_ps(_val_channel1, _mask);
-                                _val_channel2 = _mm_mul_ps(_val_channel2, _mask);
-                                _val_channel3 = _mm_mul_ps(_val_channel3, _mask);
-                            }
-                            __m128 _conv_w0 = _mm_load_ps(kptr);
-                            __m128 _conv_w1 = _mm_load_ps(kptr + out_elempack); // 1 * out_elempack
-                            _sum = _mm_comp_fmadd_ps(_val_channel0, _conv_w0, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val_channel1, _conv_w1, _sum);
-                            __m128 _conv_w2 = _mm_load_ps(kptr + 8);  // 2 * out_elempack
-                            __m128 _conv_w3 = _mm_load_ps(kptr + 12); // 3 * out_elempack
-                            _sum = _mm_comp_fmadd_ps(_val_channel2, _conv_w2, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val_channel3, _conv_w3, _sum);
-                            kptr += wstep;
-                        }
-                    }
-                }
-                _sum = activation_sse(_sum, activation_type, activation_params);
-                _mm_storeu_ps(outptr + (h_col * outw + w_col) * out_elempack, _sum);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4to1.h
deleted file mode 100644
index 2910b16c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4to1.h
+++ /dev/null
@@ -1,211 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deformableconv2d_pack4to1_sse(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const Mat& bottom_blob = bottom_blobs[0];
-    const Mat& offset = bottom_blobs[1];
-    const bool has_mask = (bottom_blobs.size() == 3);
-    const bool offset_not_pack = offset.elempack == 1;
-    const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const int size = outw * outh;
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-    const int elempack = 4;
-    const int out_elempack = 1;
-    const int wstep = out_elempack * elempack;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int h_col = 0; h_col < outh; h_col++)
-    {
-        for (int w_col = 0; w_col < outw; w_col++)
-        {
-            int h_in = h_col * stride_h - pad_top;
-            int w_in = w_col * stride_w - pad_left;
-            for (int oc = 0; oc < outch; oc++)
-            {
-                const float* kptr = weight_data_packed.channel(oc);
-                float* outptr = top_blob.channel(oc);
-                float _sum = 0.f;
-                if (bias_data_ptr)
-                    _sum = *(bias_data_ptr + oc);
-                for (int i = 0; i < kernel_h; i++)
-                {
-                    for (int j = 0; j < kernel_w; j++)
-                    {
-                        float offset_h = 0.f;
-                        float offset_w = 0.f;
-                        float mask_ = 1.f;
-                        if (offset_not_pack)
-                        {
-                            offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
-                            offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
-                        }
-                        else
-                        {
-                            const int y_c = (i * kernel_w + j) * 2;
-                            const int x_c = (i * kernel_w + j) * 2 + 1;
-                            offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
-                            offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
-                        }
-                        if (has_mask)
-                        {
-                            const Mat& mask = bottom_blobs[2];
-                            if (mask_not_pack)
-                            {
-                                mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
-                            }
-                            else
-                            {
-                                const int m_c = i * kernel_w + j;
-                                mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
-                            }
-                        }
-                        const float h_im = h_in + i * dilation_h + offset_h;
-                        const float w_im = w_in + j * dilation_w + offset_w;
-
-                        // Bilinear
-                        const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
-                        float w1 = 0.f;
-                        float w2 = 0.f;
-                        float w3 = 0.f;
-                        float w4 = 0.f;
-                        bool v1_cond = false;
-                        bool v2_cond = false;
-                        bool v3_cond = false;
-                        bool v4_cond = false;
-                        int v1_pos = 0;
-                        int v2_pos = 0;
-                        int v3_pos = 0;
-                        int v4_pos = 0;
-                        if (cond)
-                        {
-                            int h_low = (int)floorf(h_im);
-                            int w_low = (int)floorf(w_im);
-                            int h_high = h_low + 1;
-                            int w_high = w_low + 1;
-
-                            float lh = h_im - h_low;
-                            float lw = w_im - w_low;
-                            float hh = 1 - lh;
-                            float hw = 1 - lw;
-
-                            v1_cond = (h_low >= 0 && w_low >= 0);
-                            v2_cond = (h_low >= 0 && w_high <= w - 1);
-                            v3_cond = (h_high <= h - 1 && w_low >= 0);
-                            v4_cond = (h_high <= h - 1 && w_high <= w - 1);
-                            if (v1_cond)
-                                v1_pos = h_low * w + w_low;
-                            if (v2_cond)
-                                v2_pos = h_low * w + w_high;
-                            if (v3_cond)
-                                v3_pos = h_high * w + w_low;
-                            if (v4_cond)
-                                v4_pos = h_high * w + w_high;
-
-                            w1 = hh * hw;
-                            w2 = hh * lw;
-                            w3 = lh * hw;
-                            w4 = lh * lw;
-                        }
-
-                        for (int ic = 0; ic < inch; ic++)
-                        {
-                            const float* data_im_ptr = bottom_blob.channel(ic);
-                            float _val_channel0 = 0.f;
-                            float _val_channel1 = _val_channel0;
-                            float _val_channel2 = _val_channel0;
-                            float _val_channel3 = _val_channel0;
-                            if (cond)
-                            {
-                                float _v1_channel0 = _val_channel0;
-                                float _v1_channel1 = _val_channel0;
-                                float _v1_channel2 = _val_channel0;
-                                float _v1_channel3 = _val_channel0;
-                                float _v2_channel0 = _val_channel0;
-                                float _v2_channel1 = _val_channel0;
-                                float _v2_channel2 = _val_channel0;
-                                float _v2_channel3 = _val_channel0;
-                                float _v3_channel0 = _val_channel0;
-                                float _v3_channel1 = _val_channel0;
-                                float _v3_channel2 = _val_channel0;
-                                float _v3_channel3 = _val_channel0;
-                                float _v4_channel0 = _val_channel0;
-                                float _v4_channel1 = _val_channel0;
-                                float _v4_channel2 = _val_channel0;
-                                float _v4_channel3 = _val_channel0;
-                                if (v1_cond)
-                                {
-                                    _v1_channel0 = *(data_im_ptr + v1_pos * elempack);
-                                    _v1_channel1 = *(data_im_ptr + v1_pos * elempack + 1);
-                                    _v1_channel2 = *(data_im_ptr + v1_pos * elempack + 2);
-                                    _v1_channel3 = *(data_im_ptr + v1_pos * elempack + 3);
-                                }
-                                if (v2_cond)
-                                {
-                                    _v2_channel0 = *(data_im_ptr + v2_pos * elempack);
-                                    _v2_channel1 = *(data_im_ptr + v2_pos * elempack + 1);
-                                    _v2_channel2 = *(data_im_ptr + v2_pos * elempack + 2);
-                                    _v2_channel3 = *(data_im_ptr + v2_pos * elempack + 3);
-                                }
-                                if (v3_cond)
-                                {
-                                    _v3_channel0 = *(data_im_ptr + v3_pos * elempack);
-                                    _v3_channel1 = *(data_im_ptr + v3_pos * elempack + 1);
-                                    _v3_channel2 = *(data_im_ptr + v3_pos * elempack + 2);
-                                    _v3_channel3 = *(data_im_ptr + v3_pos * elempack + 3);
-                                }
-                                if (v4_cond)
-                                {
-                                    _v4_channel0 = *(data_im_ptr + v4_pos * elempack);
-                                    _v4_channel1 = *(data_im_ptr + v4_pos * elempack + 1);
-                                    _v4_channel2 = *(data_im_ptr + v4_pos * elempack + 2);
-                                    _v4_channel3 = *(data_im_ptr + v4_pos * elempack + 3);
-                                }
-                                _val_channel0 = w1 * _v1_channel0 + w2 * _v2_channel0 + w3 * _v3_channel0 + w4 * _v4_channel0;
-                                _val_channel1 = w1 * _v1_channel1 + w2 * _v2_channel1 + w3 * _v3_channel1 + w4 * _v4_channel1;
-                                _val_channel2 = w1 * _v1_channel2 + w2 * _v2_channel2 + w3 * _v3_channel2 + w4 * _v4_channel2;
-                                _val_channel3 = w1 * _v1_channel3 + w2 * _v2_channel3 + w3 * _v3_channel3 + w4 * _v4_channel3;
-                            }
-                            if (has_mask)
-                            {
-                                _val_channel0 *= mask_;
-                                _val_channel1 *= mask_;
-                                _val_channel2 *= mask_;
-                                _val_channel3 *= mask_;
-                            }
-                            float _conv_w0 = *(kptr);
-                            float _conv_w1 = *(kptr + out_elempack); // 1 * out_elempack
-                            float _conv_w2 = *(kptr + 2);            // 2 * out_elempack
-                            float _conv_w3 = *(kptr + 3);            // 3 * out_elempack
-                            _sum += (_val_channel0 * _conv_w0 + _val_channel1 * _conv_w1 + _val_channel2 * _conv_w2 + _val_channel3 * _conv_w3);
-                            kptr += wstep;
-                        }
-                    }
-                }
-                _sum = activation_ss(_sum, activation_type, activation_params);
-                *(outptr + h_col * outw + w_col) = _sum;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4to16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4to16.h
deleted file mode 100644
index 6f0393e5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4to16.h
+++ /dev/null
@@ -1,243 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deformableconv2d_pack4to16_avx512(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const Mat& bottom_blob = bottom_blobs[0];
-    const Mat& offset = bottom_blobs[1];
-    const bool has_mask = (bottom_blobs.size() == 3);
-    const bool offset_not_pack = offset.elempack == 1;
-    const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const int size = outw * outh;
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-    const int elempack = 4;
-    const int out_elempack = 16;
-    const int wstep = out_elempack * elempack;
-    const float zeros[out_elempack] = {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
-    const float* zeros_ptr = zeros;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int h_col = 0; h_col < outh; h_col++)
-    {
-        for (int w_col = 0; w_col < outw; w_col++)
-        {
-            int h_in = h_col * stride_h - pad_top;
-            int w_in = w_col * stride_w - pad_left;
-            for (int oc = 0; oc < outch; oc++)
-            {
-                const float* kptr = weight_data_packed.channel(oc);
-                float* outptr = top_blob.channel(oc);
-                __m512 _sum = _mm512_setzero_ps();
-                if (bias_data_ptr)
-                    _sum = _mm512_loadu_ps(bias_data_ptr + oc * out_elempack);
-                for (int i = 0; i < kernel_h; i++)
-                {
-                    for (int j = 0; j < kernel_w; j++)
-                    {
-                        float offset_h = 0.f;
-                        float offset_w = 0.f;
-                        float mask_ = 1.f;
-                        if (offset_not_pack)
-                        {
-                            offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
-                            offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
-                        }
-                        else
-                        {
-                            const int y_c = (i * kernel_w + j) * 2;
-                            const int x_c = (i * kernel_w + j) * 2 + 1;
-                            offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
-                            offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
-                        }
-                        if (has_mask)
-                        {
-                            const Mat& mask = bottom_blobs[2];
-                            if (mask_not_pack)
-                            {
-                                mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
-                            }
-                            else
-                            {
-                                const int m_c = i * kernel_w + j;
-                                mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
-                            }
-                        }
-                        const float h_im = h_in + i * dilation_h + offset_h;
-                        const float w_im = w_in + j * dilation_w + offset_w;
-
-                        // Bilinear
-                        const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
-                        float w1 = 0.f;
-                        float w2 = 0.f;
-                        float w3 = 0.f;
-                        float w4 = 0.f;
-                        bool v1_cond = false;
-                        bool v2_cond = false;
-                        bool v3_cond = false;
-                        bool v4_cond = false;
-                        int v1_pos = 0;
-                        int v2_pos = 0;
-                        int v3_pos = 0;
-                        int v4_pos = 0;
-                        if (cond)
-                        {
-                            int h_low = (int)floorf(h_im);
-                            int w_low = (int)floorf(w_im);
-                            int h_high = h_low + 1;
-                            int w_high = w_low + 1;
-
-                            float lh = h_im - h_low;
-                            float lw = w_im - w_low;
-                            float hh = 1 - lh;
-                            float hw = 1 - lw;
-
-                            v1_cond = (h_low >= 0 && w_low >= 0);
-                            v2_cond = (h_low >= 0 && w_high <= w - 1);
-                            v3_cond = (h_high <= h - 1 && w_low >= 0);
-                            v4_cond = (h_high <= h - 1 && w_high <= w - 1);
-                            if (v1_cond)
-                                v1_pos = h_low * w + w_low;
-                            if (v2_cond)
-                                v2_pos = h_low * w + w_high;
-                            if (v3_cond)
-                                v3_pos = h_high * w + w_low;
-                            if (v4_cond)
-                                v4_pos = h_high * w + w_high;
-
-                            w1 = hh * hw;
-                            w2 = hh * lw;
-                            w3 = lh * hw;
-                            w4 = lh * lw;
-                        }
-                        const float w1s[out_elempack] = {w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1};
-                        const float* w1_ptr = w1s;
-                        const float w2s[out_elempack] = {w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2};
-                        const float* w2_ptr = w2s;
-                        const float w3s[out_elempack] = {w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3};
-                        const float* w3_ptr = w3s;
-                        const float w4s[out_elempack] = {w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4};
-                        const float* w4_ptr = w4s;
-                        const float masks[out_elempack] = {mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_};
-                        const float* mask_ptr = masks;
-
-                        for (int ic = 0; ic < inch; ic++)
-                        {
-                            const float* data_im_ptr = bottom_blob.channel(ic);
-                            __m512 _val_channel0 = _mm512_loadu_ps(zeros_ptr);
-                            __m512 _val_channel1 = _val_channel0;
-                            __m512 _val_channel2 = _val_channel0;
-                            __m512 _val_channel3 = _val_channel0;
-                            if (cond)
-                            {
-                                __m512 _v1_channel0 = _val_channel0;
-                                __m512 _v1_channel1 = _val_channel0;
-                                __m512 _v1_channel2 = _val_channel0;
-                                __m512 _v1_channel3 = _val_channel0;
-                                __m512 _v2_channel0 = _val_channel0;
-                                __m512 _v2_channel1 = _val_channel0;
-                                __m512 _v2_channel2 = _val_channel0;
-                                __m512 _v2_channel3 = _val_channel0;
-                                __m512 _v3_channel0 = _val_channel0;
-                                __m512 _v3_channel1 = _val_channel0;
-                                __m512 _v3_channel2 = _val_channel0;
-                                __m512 _v3_channel3 = _val_channel0;
-                                __m512 _v4_channel0 = _val_channel0;
-                                __m512 _v4_channel1 = _val_channel0;
-                                __m512 _v4_channel2 = _val_channel0;
-                                __m512 _v4_channel3 = _val_channel0;
-                                if (v1_cond)
-                                {
-                                    _v1_channel0 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack]);
-                                    _v1_channel1 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 1]);
-                                    _v1_channel2 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 2]);
-                                    _v1_channel3 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 3]);
-                                }
-                                if (v2_cond)
-                                {
-                                    _v2_channel0 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack]);
-                                    _v2_channel1 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 1]);
-                                    _v2_channel2 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 2]);
-                                    _v2_channel3 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 3]);
-                                }
-                                if (v3_cond)
-                                {
-                                    _v3_channel0 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack]);
-                                    _v3_channel1 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 1]);
-                                    _v3_channel2 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 2]);
-                                    _v3_channel3 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 3]);
-                                }
-                                if (v4_cond)
-                                {
-                                    _v4_channel0 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack]);
-                                    _v4_channel1 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 1]);
-                                    _v4_channel2 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 2]);
-                                    _v4_channel3 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 3]);
-                                }
-                                __m512 _w1 = _mm512_loadu_ps(w1_ptr);
-                                __m512 _w2 = _mm512_loadu_ps(w2_ptr);
-                                __m512 _w3 = _mm512_loadu_ps(w3_ptr);
-                                __m512 _w4 = _mm512_loadu_ps(w4_ptr);
-                                _val_channel0 = _mm512_fmadd_ps(_v1_channel0, _w1, _val_channel0);
-                                _val_channel0 = _mm512_fmadd_ps(_v2_channel0, _w2, _val_channel0);
-                                _val_channel0 = _mm512_fmadd_ps(_v3_channel0, _w3, _val_channel0);
-                                _val_channel0 = _mm512_fmadd_ps(_v4_channel0, _w4, _val_channel0);
-                                _val_channel1 = _mm512_fmadd_ps(_v1_channel1, _w1, _val_channel1);
-                                _val_channel1 = _mm512_fmadd_ps(_v2_channel1, _w2, _val_channel1);
-                                _val_channel1 = _mm512_fmadd_ps(_v3_channel1, _w3, _val_channel1);
-                                _val_channel1 = _mm512_fmadd_ps(_v4_channel1, _w4, _val_channel1);
-                                _val_channel2 = _mm512_fmadd_ps(_v1_channel2, _w1, _val_channel2);
-                                _val_channel2 = _mm512_fmadd_ps(_v2_channel2, _w2, _val_channel2);
-                                _val_channel2 = _mm512_fmadd_ps(_v3_channel2, _w3, _val_channel2);
-                                _val_channel2 = _mm512_fmadd_ps(_v4_channel2, _w4, _val_channel2);
-                                _val_channel3 = _mm512_fmadd_ps(_v1_channel3, _w1, _val_channel3);
-                                _val_channel3 = _mm512_fmadd_ps(_v2_channel3, _w2, _val_channel3);
-                                _val_channel3 = _mm512_fmadd_ps(_v3_channel3, _w3, _val_channel3);
-                                _val_channel3 = _mm512_fmadd_ps(_v4_channel3, _w4, _val_channel3);
-                            }
-                            if (has_mask)
-                            {
-                                __m512 _mask = _mm512_loadu_ps(mask_ptr);
-                                _val_channel0 = _mm512_mul_ps(_val_channel0, _mask);
-                                _val_channel1 = _mm512_mul_ps(_val_channel1, _mask);
-                                _val_channel2 = _mm512_mul_ps(_val_channel2, _mask);
-                                _val_channel3 = _mm512_mul_ps(_val_channel3, _mask);
-                            }
-                            __m512 _conv_w0 = _mm512_load_ps(kptr);
-                            __m512 _conv_w1 = _mm512_load_ps(kptr + out_elempack); // 1 * out_elempack
-                            _sum = _mm512_fmadd_ps(_val_channel0, _conv_w0, _sum);
-                            _sum = _mm512_fmadd_ps(_val_channel1, _conv_w1, _sum);
-                            __m512 _conv_w2 = _mm512_load_ps(kptr + 32); // 2 * out_elempack
-                            __m512 _conv_w3 = _mm512_load_ps(kptr + 48); // 3 * out_elempack
-                            _sum = _mm512_fmadd_ps(_val_channel2, _conv_w2, _sum);
-                            _sum = _mm512_fmadd_ps(_val_channel3, _conv_w3, _sum);
-                            kptr += wstep;
-                        }
-                    }
-                }
-                _sum = activation_avx512(_sum, activation_type, activation_params);
-                _mm512_store_ps(outptr + (h_col * outw + w_col) * out_elempack, _sum);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4to8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4to8.h
deleted file mode 100644
index 86cd0b75..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack4to8.h
+++ /dev/null
@@ -1,243 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deformableconv2d_pack4to8_avx(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const Mat& bottom_blob = bottom_blobs[0];
-    const Mat& offset = bottom_blobs[1];
-    const bool has_mask = (bottom_blobs.size() == 3);
-    const bool offset_not_pack = offset.elempack == 1;
-    const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const int size = outw * outh;
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-    const int elempack = 4;
-    const int out_elempack = 8;
-    const int wstep = out_elempack * elempack;
-    const float zeros[out_elempack] = {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
-    const float* zeros_ptr = zeros;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int h_col = 0; h_col < outh; h_col++)
-    {
-        for (int w_col = 0; w_col < outw; w_col++)
-        {
-            int h_in = h_col * stride_h - pad_top;
-            int w_in = w_col * stride_w - pad_left;
-            for (int oc = 0; oc < outch; oc++)
-            {
-                const float* kptr = weight_data_packed.channel(oc);
-                float* outptr = top_blob.channel(oc);
-                __m256 _sum = _mm256_setzero_ps();
-                if (bias_data_ptr)
-                    _sum = _mm256_loadu_ps(bias_data_ptr + oc * out_elempack);
-                for (int i = 0; i < kernel_h; i++)
-                {
-                    for (int j = 0; j < kernel_w; j++)
-                    {
-                        float offset_h = 0.f;
-                        float offset_w = 0.f;
-                        float mask_ = 1.f;
-                        if (offset_not_pack)
-                        {
-                            offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
-                            offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
-                        }
-                        else
-                        {
-                            const int y_c = (i * kernel_w + j) * 2;
-                            const int x_c = (i * kernel_w + j) * 2 + 1;
-                            offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
-                            offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
-                        }
-                        if (has_mask)
-                        {
-                            const Mat& mask = bottom_blobs[2];
-                            if (mask_not_pack)
-                            {
-                                mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
-                            }
-                            else
-                            {
-                                const int m_c = i * kernel_w + j;
-                                mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
-                            }
-                        }
-                        const float h_im = h_in + i * dilation_h + offset_h;
-                        const float w_im = w_in + j * dilation_w + offset_w;
-
-                        // Bilinear
-                        const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
-                        float w1 = 0.f;
-                        float w2 = 0.f;
-                        float w3 = 0.f;
-                        float w4 = 0.f;
-                        bool v1_cond = false;
-                        bool v2_cond = false;
-                        bool v3_cond = false;
-                        bool v4_cond = false;
-                        int v1_pos = 0;
-                        int v2_pos = 0;
-                        int v3_pos = 0;
-                        int v4_pos = 0;
-                        if (cond)
-                        {
-                            int h_low = (int)floorf(h_im);
-                            int w_low = (int)floorf(w_im);
-                            int h_high = h_low + 1;
-                            int w_high = w_low + 1;
-
-                            float lh = h_im - h_low;
-                            float lw = w_im - w_low;
-                            float hh = 1 - lh;
-                            float hw = 1 - lw;
-
-                            v1_cond = (h_low >= 0 && w_low >= 0);
-                            v2_cond = (h_low >= 0 && w_high <= w - 1);
-                            v3_cond = (h_high <= h - 1 && w_low >= 0);
-                            v4_cond = (h_high <= h - 1 && w_high <= w - 1);
-                            if (v1_cond)
-                                v1_pos = h_low * w + w_low;
-                            if (v2_cond)
-                                v2_pos = h_low * w + w_high;
-                            if (v3_cond)
-                                v3_pos = h_high * w + w_low;
-                            if (v4_cond)
-                                v4_pos = h_high * w + w_high;
-
-                            w1 = hh * hw;
-                            w2 = hh * lw;
-                            w3 = lh * hw;
-                            w4 = lh * lw;
-                        }
-                        const float w1s[out_elempack] = {w1, w1, w1, w1, w1, w1, w1, w1};
-                        const float* w1_ptr = w1s;
-                        const float w2s[out_elempack] = {w2, w2, w2, w2, w2, w2, w2, w2};
-                        const float* w2_ptr = w2s;
-                        const float w3s[out_elempack] = {w3, w3, w3, w3, w3, w3, w3, w3};
-                        const float* w3_ptr = w3s;
-                        const float w4s[out_elempack] = {w4, w4, w4, w4, w4, w4, w4, w4};
-                        const float* w4_ptr = w4s;
-                        const float masks[out_elempack] = {mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_};
-                        const float* mask_ptr = masks;
-
-                        for (int ic = 0; ic < inch; ic++)
-                        {
-                            const float* data_im_ptr = bottom_blob.channel(ic);
-                            __m256 _val_channel0 = _mm256_loadu_ps(zeros_ptr);
-                            __m256 _val_channel1 = _val_channel0;
-                            __m256 _val_channel2 = _val_channel0;
-                            __m256 _val_channel3 = _val_channel0;
-                            if (cond)
-                            {
-                                __m256 _v1_channel0 = _val_channel0;
-                                __m256 _v1_channel1 = _val_channel0;
-                                __m256 _v1_channel2 = _val_channel0;
-                                __m256 _v1_channel3 = _val_channel0;
-                                __m256 _v2_channel0 = _val_channel0;
-                                __m256 _v2_channel1 = _val_channel0;
-                                __m256 _v2_channel2 = _val_channel0;
-                                __m256 _v2_channel3 = _val_channel0;
-                                __m256 _v3_channel0 = _val_channel0;
-                                __m256 _v3_channel1 = _val_channel0;
-                                __m256 _v3_channel2 = _val_channel0;
-                                __m256 _v3_channel3 = _val_channel0;
-                                __m256 _v4_channel0 = _val_channel0;
-                                __m256 _v4_channel1 = _val_channel0;
-                                __m256 _v4_channel2 = _val_channel0;
-                                __m256 _v4_channel3 = _val_channel0;
-                                if (v1_cond)
-                                {
-                                    _v1_channel0 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack);
-                                    _v1_channel1 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 1);
-                                    _v1_channel2 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 2);
-                                    _v1_channel3 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 3);
-                                }
-                                if (v2_cond)
-                                {
-                                    _v2_channel0 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack);
-                                    _v2_channel1 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 1);
-                                    _v2_channel2 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 2);
-                                    _v2_channel3 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 3);
-                                }
-                                if (v3_cond)
-                                {
-                                    _v3_channel0 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack);
-                                    _v3_channel1 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 1);
-                                    _v3_channel2 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 2);
-                                    _v3_channel3 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 3);
-                                }
-                                if (v4_cond)
-                                {
-                                    _v4_channel0 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack);
-                                    _v4_channel1 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 1);
-                                    _v4_channel2 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 2);
-                                    _v4_channel3 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 3);
-                                }
-                                __m256 _w1 = _mm256_loadu_ps(w1_ptr);
-                                __m256 _w2 = _mm256_loadu_ps(w2_ptr);
-                                __m256 _w3 = _mm256_loadu_ps(w3_ptr);
-                                __m256 _w4 = _mm256_loadu_ps(w4_ptr);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v1_channel0, _w1, _val_channel0);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v2_channel0, _w2, _val_channel0);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v3_channel0, _w3, _val_channel0);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v4_channel0, _w4, _val_channel0);
-                                _val_channel1 = _mm256_comp_fmadd_ps(_v1_channel1, _w1, _val_channel1);
-                                _val_channel1 = _mm256_comp_fmadd_ps(_v2_channel1, _w2, _val_channel1);
-                                _val_channel1 = _mm256_comp_fmadd_ps(_v3_channel1, _w3, _val_channel1);
-                                _val_channel1 = _mm256_comp_fmadd_ps(_v4_channel1, _w4, _val_channel1);
-                                _val_channel2 = _mm256_comp_fmadd_ps(_v1_channel2, _w1, _val_channel2);
-                                _val_channel2 = _mm256_comp_fmadd_ps(_v2_channel2, _w2, _val_channel2);
-                                _val_channel2 = _mm256_comp_fmadd_ps(_v3_channel2, _w3, _val_channel2);
-                                _val_channel2 = _mm256_comp_fmadd_ps(_v4_channel2, _w4, _val_channel2);
-                                _val_channel3 = _mm256_comp_fmadd_ps(_v1_channel3, _w1, _val_channel3);
-                                _val_channel3 = _mm256_comp_fmadd_ps(_v2_channel3, _w2, _val_channel3);
-                                _val_channel3 = _mm256_comp_fmadd_ps(_v3_channel3, _w3, _val_channel3);
-                                _val_channel3 = _mm256_comp_fmadd_ps(_v4_channel3, _w4, _val_channel3);
-                            }
-                            if (has_mask)
-                            {
-                                __m256 _mask = _mm256_loadu_ps(mask_ptr);
-                                _val_channel0 = _mm256_mul_ps(_val_channel0, _mask);
-                                _val_channel1 = _mm256_mul_ps(_val_channel1, _mask);
-                                _val_channel2 = _mm256_mul_ps(_val_channel2, _mask);
-                                _val_channel3 = _mm256_mul_ps(_val_channel3, _mask);
-                            }
-                            __m256 _conv_w0 = _mm256_load_ps(kptr);
-                            __m256 _conv_w1 = _mm256_load_ps(kptr + out_elempack); // 1 * out_elempack
-                            _sum = _mm256_comp_fmadd_ps(_val_channel0, _conv_w0, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val_channel1, _conv_w1, _sum);
-                            __m256 _conv_w2 = _mm256_load_ps(kptr + 16); // 2 * out_elempack
-                            __m256 _conv_w3 = _mm256_load_ps(kptr + 24); // 3 * out_elempack
-                            _sum = _mm256_comp_fmadd_ps(_val_channel2, _conv_w2, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val_channel3, _conv_w3, _sum);
-                            kptr += wstep;
-                        }
-                    }
-                }
-                _sum = activation_avx(_sum, activation_type, activation_params);
-                _mm256_store_ps(outptr + (h_col * outw + w_col) * out_elempack, _sum);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8.h
deleted file mode 100644
index 93dcd682..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8.h
+++ /dev/null
@@ -1,307 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deformableconv2d_pack8_avx(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const Mat& bottom_blob = bottom_blobs[0];
-    const Mat& offset = bottom_blobs[1];
-    const bool has_mask = (bottom_blobs.size() == 3);
-    const bool offset_not_pack = offset.elempack == 1;
-    const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const int size = outw * outh;
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-    const int elempack = 8;
-    const int out_elempack = 8;
-    const int wstep = out_elempack * elempack;
-    const float zeros[out_elempack] = {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
-    const float* zeros_ptr = zeros;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int h_col = 0; h_col < outh; h_col++)
-    {
-        for (int w_col = 0; w_col < outw; w_col++)
-        {
-            int h_in = h_col * stride_h - pad_top;
-            int w_in = w_col * stride_w - pad_left;
-            for (int oc = 0; oc < outch; oc++)
-            {
-                const float* kptr = weight_data_packed.channel(oc);
-                float* outptr = top_blob.channel(oc);
-                __m256 _sum = _mm256_setzero_ps();
-                if (bias_data_ptr)
-                    _sum = _mm256_loadu_ps(bias_data_ptr + oc * out_elempack);
-                for (int i = 0; i < kernel_h; i++)
-                {
-                    for (int j = 0; j < kernel_w; j++)
-                    {
-                        float offset_h = 0.f;
-                        float offset_w = 0.f;
-                        float mask_ = 1.f;
-                        if (offset_not_pack)
-                        {
-                            offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
-                            offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
-                        }
-                        else
-                        {
-                            const int y_c = (i * kernel_w + j) * 2;
-                            const int x_c = (i * kernel_w + j) * 2 + 1;
-                            offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
-                            offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
-                        }
-                        if (has_mask)
-                        {
-                            const Mat& mask = bottom_blobs[2];
-                            if (mask_not_pack)
-                            {
-                                mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
-                            }
-                            else
-                            {
-                                const int m_c = i * kernel_w + j;
-                                mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
-                            }
-                        }
-                        const float h_im = h_in + i * dilation_h + offset_h;
-                        const float w_im = w_in + j * dilation_w + offset_w;
-
-                        // Bilinear
-                        const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
-                        float w1 = 0.f;
-                        float w2 = 0.f;
-                        float w3 = 0.f;
-                        float w4 = 0.f;
-                        bool v1_cond = false;
-                        bool v2_cond = false;
-                        bool v3_cond = false;
-                        bool v4_cond = false;
-                        int v1_pos = 0;
-                        int v2_pos = 0;
-                        int v3_pos = 0;
-                        int v4_pos = 0;
-                        if (cond)
-                        {
-                            int h_low = (int)floorf(h_im);
-                            int w_low = (int)floorf(w_im);
-                            int h_high = h_low + 1;
-                            int w_high = w_low + 1;
-
-                            float lh = h_im - h_low;
-                            float lw = w_im - w_low;
-                            float hh = 1 - lh;
-                            float hw = 1 - lw;
-
-                            v1_cond = (h_low >= 0 && w_low >= 0);
-                            v2_cond = (h_low >= 0 && w_high <= w - 1);
-                            v3_cond = (h_high <= h - 1 && w_low >= 0);
-                            v4_cond = (h_high <= h - 1 && w_high <= w - 1);
-                            if (v1_cond)
-                                v1_pos = h_low * w + w_low;
-                            if (v2_cond)
-                                v2_pos = h_low * w + w_high;
-                            if (v3_cond)
-                                v3_pos = h_high * w + w_low;
-                            if (v4_cond)
-                                v4_pos = h_high * w + w_high;
-
-                            w1 = hh * hw;
-                            w2 = hh * lw;
-                            w3 = lh * hw;
-                            w4 = lh * lw;
-                        }
-                        const float w1s[out_elempack] = {w1, w1, w1, w1, w1, w1, w1, w1};
-                        const float* w1_ptr = w1s;
-                        const float w2s[out_elempack] = {w2, w2, w2, w2, w2, w2, w2, w2};
-                        const float* w2_ptr = w2s;
-                        const float w3s[out_elempack] = {w3, w3, w3, w3, w3, w3, w3, w3};
-                        const float* w3_ptr = w3s;
-                        const float w4s[out_elempack] = {w4, w4, w4, w4, w4, w4, w4, w4};
-                        const float* w4_ptr = w4s;
-                        const float masks[out_elempack] = {mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_};
-                        const float* mask_ptr = masks;
-
-                        for (int ic = 0; ic < inch; ic++)
-                        {
-                            const float* data_im_ptr = bottom_blob.channel(ic);
-                            __m256 _val_channel0 = _mm256_loadu_ps(zeros_ptr);
-                            __m256 _val_channel1 = _val_channel0;
-                            __m256 _val_channel2 = _val_channel0;
-                            __m256 _val_channel3 = _val_channel0;
-                            __m256 _val_channel4 = _val_channel0;
-                            __m256 _val_channel5 = _val_channel0;
-                            __m256 _val_channel6 = _val_channel0;
-                            __m256 _val_channel7 = _val_channel0;
-                            if (cond)
-                            {
-                                __m256 _v1_channel0 = _val_channel0;
-                                __m256 _v1_channel1 = _val_channel0;
-                                __m256 _v1_channel2 = _val_channel0;
-                                __m256 _v1_channel3 = _val_channel0;
-                                __m256 _v1_channel4 = _val_channel0;
-                                __m256 _v1_channel5 = _val_channel0;
-                                __m256 _v1_channel6 = _val_channel0;
-                                __m256 _v1_channel7 = _val_channel0;
-                                __m256 _v2_channel0 = _val_channel0;
-                                __m256 _v2_channel1 = _val_channel0;
-                                __m256 _v2_channel2 = _val_channel0;
-                                __m256 _v2_channel3 = _val_channel0;
-                                __m256 _v2_channel4 = _val_channel0;
-                                __m256 _v2_channel5 = _val_channel0;
-                                __m256 _v2_channel6 = _val_channel0;
-                                __m256 _v2_channel7 = _val_channel0;
-                                __m256 _v3_channel0 = _val_channel0;
-                                __m256 _v3_channel1 = _val_channel0;
-                                __m256 _v3_channel2 = _val_channel0;
-                                __m256 _v3_channel3 = _val_channel0;
-                                __m256 _v3_channel4 = _val_channel0;
-                                __m256 _v3_channel5 = _val_channel0;
-                                __m256 _v3_channel6 = _val_channel0;
-                                __m256 _v3_channel7 = _val_channel0;
-                                __m256 _v4_channel0 = _val_channel0;
-                                __m256 _v4_channel1 = _val_channel0;
-                                __m256 _v4_channel2 = _val_channel0;
-                                __m256 _v4_channel3 = _val_channel0;
-                                __m256 _v4_channel4 = _val_channel0;
-                                __m256 _v4_channel5 = _val_channel0;
-                                __m256 _v4_channel6 = _val_channel0;
-                                __m256 _v4_channel7 = _val_channel0;
-                                if (v1_cond)
-                                {
-                                    _v1_channel0 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack);
-                                    _v1_channel1 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 1);
-                                    _v1_channel2 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 2);
-                                    _v1_channel3 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 3);
-                                    _v1_channel4 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 4);
-                                    _v1_channel5 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 5);
-                                    _v1_channel6 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 6);
-                                    _v1_channel7 = _mm256_broadcast_ss(data_im_ptr + v1_pos * elempack + 7);
-                                }
-                                if (v2_cond)
-                                {
-                                    _v2_channel0 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack);
-                                    _v2_channel1 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 1);
-                                    _v2_channel2 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 2);
-                                    _v2_channel3 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 3);
-                                    _v2_channel4 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 4);
-                                    _v2_channel5 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 5);
-                                    _v2_channel6 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 6);
-                                    _v2_channel7 = _mm256_broadcast_ss(data_im_ptr + v2_pos * elempack + 7);
-                                }
-                                if (v3_cond)
-                                {
-                                    _v3_channel0 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack);
-                                    _v3_channel1 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 1);
-                                    _v3_channel2 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 2);
-                                    _v3_channel3 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 3);
-                                    _v3_channel4 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 4);
-                                    _v3_channel5 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 5);
-                                    _v3_channel6 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 6);
-                                    _v3_channel7 = _mm256_broadcast_ss(data_im_ptr + v3_pos * elempack + 7);
-                                }
-                                if (v4_cond)
-                                {
-                                    _v4_channel0 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack);
-                                    _v4_channel1 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 1);
-                                    _v4_channel2 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 2);
-                                    _v4_channel3 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 3);
-                                    _v4_channel4 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 4);
-                                    _v4_channel5 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 5);
-                                    _v4_channel6 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 6);
-                                    _v4_channel7 = _mm256_broadcast_ss(data_im_ptr + v4_pos * elempack + 7);
-                                }
-                                __m256 _w1 = _mm256_loadu_ps(w1_ptr);
-                                __m256 _w2 = _mm256_loadu_ps(w2_ptr);
-                                __m256 _w3 = _mm256_loadu_ps(w3_ptr);
-                                __m256 _w4 = _mm256_loadu_ps(w4_ptr);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v1_channel0, _w1, _val_channel0);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v2_channel0, _w2, _val_channel0);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v3_channel0, _w3, _val_channel0);
-                                _val_channel0 = _mm256_comp_fmadd_ps(_v4_channel0, _w4, _val_channel0);
-                                _val_channel1 = _mm256_comp_fmadd_ps(_v1_channel1, _w1, _val_channel1);
-                                _val_channel1 = _mm256_comp_fmadd_ps(_v2_channel1, _w2, _val_channel1);
-                                _val_channel1 = _mm256_comp_fmadd_ps(_v3_channel1, _w3, _val_channel1);
-                                _val_channel1 = _mm256_comp_fmadd_ps(_v4_channel1, _w4, _val_channel1);
-                                _val_channel2 = _mm256_comp_fmadd_ps(_v1_channel2, _w1, _val_channel2);
-                                _val_channel2 = _mm256_comp_fmadd_ps(_v2_channel2, _w2, _val_channel2);
-                                _val_channel2 = _mm256_comp_fmadd_ps(_v3_channel2, _w3, _val_channel2);
-                                _val_channel2 = _mm256_comp_fmadd_ps(_v4_channel2, _w4, _val_channel2);
-                                _val_channel3 = _mm256_comp_fmadd_ps(_v1_channel3, _w1, _val_channel3);
-                                _val_channel3 = _mm256_comp_fmadd_ps(_v2_channel3, _w2, _val_channel3);
-                                _val_channel3 = _mm256_comp_fmadd_ps(_v3_channel3, _w3, _val_channel3);
-                                _val_channel3 = _mm256_comp_fmadd_ps(_v4_channel3, _w4, _val_channel3);
-                                _val_channel4 = _mm256_comp_fmadd_ps(_v1_channel4, _w1, _val_channel4);
-                                _val_channel4 = _mm256_comp_fmadd_ps(_v2_channel4, _w2, _val_channel4);
-                                _val_channel4 = _mm256_comp_fmadd_ps(_v3_channel4, _w3, _val_channel4);
-                                _val_channel4 = _mm256_comp_fmadd_ps(_v4_channel4, _w4, _val_channel4);
-                                _val_channel5 = _mm256_comp_fmadd_ps(_v1_channel5, _w1, _val_channel5);
-                                _val_channel5 = _mm256_comp_fmadd_ps(_v2_channel5, _w2, _val_channel5);
-                                _val_channel5 = _mm256_comp_fmadd_ps(_v3_channel5, _w3, _val_channel5);
-                                _val_channel5 = _mm256_comp_fmadd_ps(_v4_channel5, _w4, _val_channel5);
-                                _val_channel6 = _mm256_comp_fmadd_ps(_v1_channel6, _w1, _val_channel6);
-                                _val_channel6 = _mm256_comp_fmadd_ps(_v2_channel6, _w2, _val_channel6);
-                                _val_channel6 = _mm256_comp_fmadd_ps(_v3_channel6, _w3, _val_channel6);
-                                _val_channel6 = _mm256_comp_fmadd_ps(_v4_channel6, _w4, _val_channel6);
-                                _val_channel7 = _mm256_comp_fmadd_ps(_v1_channel7, _w1, _val_channel7);
-                                _val_channel7 = _mm256_comp_fmadd_ps(_v2_channel7, _w2, _val_channel7);
-                                _val_channel7 = _mm256_comp_fmadd_ps(_v3_channel7, _w3, _val_channel7);
-                                _val_channel7 = _mm256_comp_fmadd_ps(_v4_channel7, _w4, _val_channel7);
-                            }
-                            if (has_mask)
-                            {
-                                __m256 _mask = _mm256_loadu_ps(mask_ptr);
-                                _val_channel0 = _mm256_mul_ps(_val_channel0, _mask);
-                                _val_channel1 = _mm256_mul_ps(_val_channel1, _mask);
-                                _val_channel2 = _mm256_mul_ps(_val_channel2, _mask);
-                                _val_channel3 = _mm256_mul_ps(_val_channel3, _mask);
-                                _val_channel4 = _mm256_mul_ps(_val_channel4, _mask);
-                                _val_channel5 = _mm256_mul_ps(_val_channel5, _mask);
-                                _val_channel6 = _mm256_mul_ps(_val_channel6, _mask);
-                                _val_channel7 = _mm256_mul_ps(_val_channel7, _mask);
-                            }
-                            __m256 _conv_w0 = _mm256_load_ps(kptr);
-                            __m256 _conv_w1 = _mm256_load_ps(kptr + out_elempack); // 1 * out_elempack
-                            _sum = _mm256_comp_fmadd_ps(_val_channel0, _conv_w0, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val_channel1, _conv_w1, _sum);
-                            __m256 _conv_w2 = _mm256_load_ps(kptr + 16); // 2 * out_elempack
-                            __m256 _conv_w3 = _mm256_load_ps(kptr + 24); // 3 * out_elempack
-                            _sum = _mm256_comp_fmadd_ps(_val_channel2, _conv_w2, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val_channel3, _conv_w3, _sum);
-                            __m256 _conv_w4 = _mm256_load_ps(kptr + 32); // 4 * out_elempack
-                            __m256 _conv_w5 = _mm256_load_ps(kptr + 40); // 5 * out_elempack
-                            _sum = _mm256_comp_fmadd_ps(_val_channel4, _conv_w4, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val_channel5, _conv_w5, _sum);
-                            __m256 _conv_w6 = _mm256_load_ps(kptr + 48); // 6 * out_elempack
-                            __m256 _conv_w7 = _mm256_load_ps(kptr + 56); // 7 * out_elempack
-                            _sum = _mm256_comp_fmadd_ps(_val_channel6, _conv_w6, _sum);
-                            _sum = _mm256_comp_fmadd_ps(_val_channel7, _conv_w7, _sum);
-                            kptr += wstep;
-                        }
-                    }
-                }
-                _sum = activation_avx(_sum, activation_type, activation_params);
-                _mm256_store_ps(outptr + (h_col * outw + w_col) * out_elempack, _sum);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8to1.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8to1.h
deleted file mode 100644
index d0674ed0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8to1.h
+++ /dev/null
@@ -1,259 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deformableconv2d_pack8to1_avx(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const Mat& bottom_blob = bottom_blobs[0];
-    const Mat& offset = bottom_blobs[1];
-    const bool has_mask = (bottom_blobs.size() == 3);
-    const bool offset_not_pack = offset.elempack == 1;
-    const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const int size = outw * outh;
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-    const int elempack = 8;
-    const int out_elempack = 1;
-    const int wstep = out_elempack * elempack;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int h_col = 0; h_col < outh; h_col++)
-    {
-        for (int w_col = 0; w_col < outw; w_col++)
-        {
-            int h_in = h_col * stride_h - pad_top;
-            int w_in = w_col * stride_w - pad_left;
-            for (int oc = 0; oc < outch; oc++)
-            {
-                const float* kptr = weight_data_packed.channel(oc);
-                float* outptr = top_blob.channel(oc);
-                float _sum = 0.f;
-                if (bias_data_ptr)
-                    _sum = *(bias_data_ptr + oc);
-                for (int i = 0; i < kernel_h; i++)
-                {
-                    for (int j = 0; j < kernel_w; j++)
-                    {
-                        float offset_h = 0.f;
-                        float offset_w = 0.f;
-                        float mask_ = 1.f;
-                        if (offset_not_pack)
-                        {
-                            offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
-                            offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
-                        }
-                        else
-                        {
-                            const int y_c = (i * kernel_w + j) * 2;
-                            const int x_c = (i * kernel_w + j) * 2 + 1;
-                            offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
-                            offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
-                        }
-                        if (has_mask)
-                        {
-                            const Mat& mask = bottom_blobs[2];
-                            if (mask_not_pack)
-                            {
-                                mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
-                            }
-                            else
-                            {
-                                const int m_c = i * kernel_w + j;
-                                mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
-                            }
-                        }
-                        const float h_im = h_in + i * dilation_h + offset_h;
-                        const float w_im = w_in + j * dilation_w + offset_w;
-
-                        // Bilinear
-                        const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
-                        float w1 = 0.f;
-                        float w2 = 0.f;
-                        float w3 = 0.f;
-                        float w4 = 0.f;
-                        bool v1_cond = false;
-                        bool v2_cond = false;
-                        bool v3_cond = false;
-                        bool v4_cond = false;
-                        int v1_pos = 0;
-                        int v2_pos = 0;
-                        int v3_pos = 0;
-                        int v4_pos = 0;
-                        if (cond)
-                        {
-                            int h_low = (int)floorf(h_im);
-                            int w_low = (int)floorf(w_im);
-                            int h_high = h_low + 1;
-                            int w_high = w_low + 1;
-
-                            float lh = h_im - h_low;
-                            float lw = w_im - w_low;
-                            float hh = 1 - lh;
-                            float hw = 1 - lw;
-
-                            v1_cond = (h_low >= 0 && w_low >= 0);
-                            v2_cond = (h_low >= 0 && w_high <= w - 1);
-                            v3_cond = (h_high <= h - 1 && w_low >= 0);
-                            v4_cond = (h_high <= h - 1 && w_high <= w - 1);
-                            if (v1_cond)
-                                v1_pos = h_low * w + w_low;
-                            if (v2_cond)
-                                v2_pos = h_low * w + w_high;
-                            if (v3_cond)
-                                v3_pos = h_high * w + w_low;
-                            if (v4_cond)
-                                v4_pos = h_high * w + w_high;
-
-                            w1 = hh * hw;
-                            w2 = hh * lw;
-                            w3 = lh * hw;
-                            w4 = lh * lw;
-                        }
-
-                        for (int ic = 0; ic < inch; ic++)
-                        {
-                            const float* data_im_ptr = bottom_blob.channel(ic);
-                            float _val_channel0 = 0.f;
-                            float _val_channel1 = _val_channel0;
-                            float _val_channel2 = _val_channel0;
-                            float _val_channel3 = _val_channel0;
-                            float _val_channel4 = _val_channel0;
-                            float _val_channel5 = _val_channel0;
-                            float _val_channel6 = _val_channel0;
-                            float _val_channel7 = _val_channel0;
-                            if (cond)
-                            {
-                                float _v1_channel0 = _val_channel0;
-                                float _v1_channel1 = _val_channel0;
-                                float _v1_channel2 = _val_channel0;
-                                float _v1_channel3 = _val_channel0;
-                                float _v1_channel4 = _val_channel0;
-                                float _v1_channel5 = _val_channel0;
-                                float _v1_channel6 = _val_channel0;
-                                float _v1_channel7 = _val_channel0;
-                                float _v2_channel0 = _val_channel0;
-                                float _v2_channel1 = _val_channel0;
-                                float _v2_channel2 = _val_channel0;
-                                float _v2_channel3 = _val_channel0;
-                                float _v2_channel4 = _val_channel0;
-                                float _v2_channel5 = _val_channel0;
-                                float _v2_channel6 = _val_channel0;
-                                float _v2_channel7 = _val_channel0;
-                                float _v3_channel0 = _val_channel0;
-                                float _v3_channel1 = _val_channel0;
-                                float _v3_channel2 = _val_channel0;
-                                float _v3_channel3 = _val_channel0;
-                                float _v3_channel4 = _val_channel0;
-                                float _v3_channel5 = _val_channel0;
-                                float _v3_channel6 = _val_channel0;
-                                float _v3_channel7 = _val_channel0;
-                                float _v4_channel0 = _val_channel0;
-                                float _v4_channel1 = _val_channel0;
-                                float _v4_channel2 = _val_channel0;
-                                float _v4_channel3 = _val_channel0;
-                                float _v4_channel4 = _val_channel0;
-                                float _v4_channel5 = _val_channel0;
-                                float _v4_channel6 = _val_channel0;
-                                float _v4_channel7 = _val_channel0;
-                                if (v1_cond)
-                                {
-                                    _v1_channel0 = *(data_im_ptr + v1_pos * elempack);
-                                    _v1_channel1 = *(data_im_ptr + v1_pos * elempack + 1);
-                                    _v1_channel2 = *(data_im_ptr + v1_pos * elempack + 2);
-                                    _v1_channel3 = *(data_im_ptr + v1_pos * elempack + 3);
-                                    _v1_channel4 = *(data_im_ptr + v1_pos * elempack + 4);
-                                    _v1_channel5 = *(data_im_ptr + v1_pos * elempack + 5);
-                                    _v1_channel6 = *(data_im_ptr + v1_pos * elempack + 6);
-                                    _v1_channel7 = *(data_im_ptr + v1_pos * elempack + 7);
-                                }
-                                if (v2_cond)
-                                {
-                                    _v2_channel0 = *(data_im_ptr + v2_pos * elempack);
-                                    _v2_channel1 = *(data_im_ptr + v2_pos * elempack + 1);
-                                    _v2_channel2 = *(data_im_ptr + v2_pos * elempack + 2);
-                                    _v2_channel3 = *(data_im_ptr + v2_pos * elempack + 3);
-                                    _v2_channel4 = *(data_im_ptr + v2_pos * elempack + 4);
-                                    _v2_channel5 = *(data_im_ptr + v2_pos * elempack + 5);
-                                    _v2_channel6 = *(data_im_ptr + v2_pos * elempack + 6);
-                                    _v2_channel7 = *(data_im_ptr + v2_pos * elempack + 7);
-                                }
-                                if (v3_cond)
-                                {
-                                    _v3_channel0 = *(data_im_ptr + v3_pos * elempack);
-                                    _v3_channel1 = *(data_im_ptr + v3_pos * elempack + 1);
-                                    _v3_channel2 = *(data_im_ptr + v3_pos * elempack + 2);
-                                    _v3_channel3 = *(data_im_ptr + v3_pos * elempack + 3);
-                                    _v3_channel4 = *(data_im_ptr + v3_pos * elempack + 4);
-                                    _v3_channel5 = *(data_im_ptr + v3_pos * elempack + 5);
-                                    _v3_channel6 = *(data_im_ptr + v3_pos * elempack + 6);
-                                    _v3_channel7 = *(data_im_ptr + v3_pos * elempack + 7);
-                                }
-                                if (v4_cond)
-                                {
-                                    _v4_channel0 = *(data_im_ptr + v4_pos * elempack);
-                                    _v4_channel1 = *(data_im_ptr + v4_pos * elempack + 1);
-                                    _v4_channel2 = *(data_im_ptr + v4_pos * elempack + 2);
-                                    _v4_channel3 = *(data_im_ptr + v4_pos * elempack + 3);
-                                    _v4_channel4 = *(data_im_ptr + v4_pos * elempack + 4);
-                                    _v4_channel5 = *(data_im_ptr + v4_pos * elempack + 5);
-                                    _v4_channel6 = *(data_im_ptr + v4_pos * elempack + 6);
-                                    _v4_channel7 = *(data_im_ptr + v4_pos * elempack + 7);
-                                }
-                                _val_channel0 = w1 * _v1_channel0 + w2 * _v2_channel0 + w3 * _v3_channel0 + w4 * _v4_channel0;
-                                _val_channel1 = w1 * _v1_channel1 + w2 * _v2_channel1 + w3 * _v3_channel1 + w4 * _v4_channel1;
-                                _val_channel2 = w1 * _v1_channel2 + w2 * _v2_channel2 + w3 * _v3_channel2 + w4 * _v4_channel2;
-                                _val_channel3 = w1 * _v1_channel3 + w2 * _v2_channel3 + w3 * _v3_channel3 + w4 * _v4_channel3;
-                                _val_channel4 = w1 * _v1_channel4 + w2 * _v2_channel4 + w3 * _v3_channel4 + w4 * _v4_channel4;
-                                _val_channel5 = w1 * _v1_channel5 + w2 * _v2_channel5 + w3 * _v3_channel5 + w4 * _v4_channel5;
-                                _val_channel6 = w1 * _v1_channel6 + w2 * _v2_channel6 + w3 * _v3_channel6 + w4 * _v4_channel6;
-                                _val_channel7 = w1 * _v1_channel7 + w2 * _v2_channel7 + w3 * _v3_channel7 + w4 * _v4_channel7;
-                            }
-                            if (has_mask)
-                            {
-                                _val_channel0 *= mask_;
-                                _val_channel1 *= mask_;
-                                _val_channel2 *= mask_;
-                                _val_channel3 *= mask_;
-                                _val_channel4 *= mask_;
-                                _val_channel5 *= mask_;
-                                _val_channel6 *= mask_;
-                                _val_channel7 *= mask_;
-                            }
-                            float _conv_w0 = *(kptr);
-                            float _conv_w1 = *(kptr + out_elempack); // 1 * out_elempack
-                            float _conv_w2 = *(kptr + 2);            // 2 * out_elempack
-                            float _conv_w3 = *(kptr + 3);            // 3 * out_elempack
-                            float _conv_w4 = *(kptr + 4);            // 4 * out_elempack
-                            float _conv_w5 = *(kptr + 5);            // 5 * out_elempack
-                            float _conv_w6 = *(kptr + 6);            // 6 * out_elempack
-                            float _conv_w7 = *(kptr + 7);            // 7 * out_elempack
-                            _sum += (_val_channel0 * _conv_w0 + _val_channel1 * _conv_w1 + _val_channel2 * _conv_w2 + _val_channel3 * _conv_w3 + _val_channel4 * _conv_w4 + _val_channel5 * _conv_w5 + _val_channel6 * _conv_w6 + _val_channel7 * _conv_w7);
-                            kptr += wstep;
-                        }
-                    }
-                }
-                _sum = activation_ss(_sum, activation_type, activation_params);
-                *(outptr + h_col * outw + w_col) = _sum;
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8to16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8to16.h
deleted file mode 100644
index bcd63bed..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8to16.h
+++ /dev/null
@@ -1,307 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deformableconv2d_pack8to16_avx512(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const Mat& bottom_blob = bottom_blobs[0];
-    const Mat& offset = bottom_blobs[1];
-    const bool has_mask = (bottom_blobs.size() == 3);
-    const bool offset_not_pack = offset.elempack == 1;
-    const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const int size = outw * outh;
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-    const int elempack = 8;
-    const int out_elempack = 16;
-    const int wstep = out_elempack * elempack;
-    const float zeros[out_elempack] = {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
-    const float* zeros_ptr = zeros;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int h_col = 0; h_col < outh; h_col++)
-    {
-        for (int w_col = 0; w_col < outw; w_col++)
-        {
-            int h_in = h_col * stride_h - pad_top;
-            int w_in = w_col * stride_w - pad_left;
-            for (int oc = 0; oc < outch; oc++)
-            {
-                const float* kptr = weight_data_packed.channel(oc);
-                float* outptr = top_blob.channel(oc);
-                __m512 _sum = _mm512_setzero_ps();
-                if (bias_data_ptr)
-                    _sum = _mm512_loadu_ps(bias_data_ptr + oc * out_elempack);
-                for (int i = 0; i < kernel_h; i++)
-                {
-                    for (int j = 0; j < kernel_w; j++)
-                    {
-                        float offset_h = 0.f;
-                        float offset_w = 0.f;
-                        float mask_ = 1.f;
-                        if (offset_not_pack)
-                        {
-                            offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
-                            offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
-                        }
-                        else
-                        {
-                            const int y_c = (i * kernel_w + j) * 2;
-                            const int x_c = (i * kernel_w + j) * 2 + 1;
-                            offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
-                            offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
-                        }
-                        if (has_mask)
-                        {
-                            const Mat& mask = bottom_blobs[2];
-                            if (mask_not_pack)
-                            {
-                                mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
-                            }
-                            else
-                            {
-                                const int m_c = i * kernel_w + j;
-                                mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
-                            }
-                        }
-                        const float h_im = h_in + i * dilation_h + offset_h;
-                        const float w_im = w_in + j * dilation_w + offset_w;
-
-                        // Bilinear
-                        const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
-                        float w1 = 0.f;
-                        float w2 = 0.f;
-                        float w3 = 0.f;
-                        float w4 = 0.f;
-                        bool v1_cond = false;
-                        bool v2_cond = false;
-                        bool v3_cond = false;
-                        bool v4_cond = false;
-                        int v1_pos = 0;
-                        int v2_pos = 0;
-                        int v3_pos = 0;
-                        int v4_pos = 0;
-                        if (cond)
-                        {
-                            int h_low = (int)floorf(h_im);
-                            int w_low = (int)floorf(w_im);
-                            int h_high = h_low + 1;
-                            int w_high = w_low + 1;
-
-                            float lh = h_im - h_low;
-                            float lw = w_im - w_low;
-                            float hh = 1 - lh;
-                            float hw = 1 - lw;
-
-                            v1_cond = (h_low >= 0 && w_low >= 0);
-                            v2_cond = (h_low >= 0 && w_high <= w - 1);
-                            v3_cond = (h_high <= h - 1 && w_low >= 0);
-                            v4_cond = (h_high <= h - 1 && w_high <= w - 1);
-                            if (v1_cond)
-                                v1_pos = h_low * w + w_low;
-                            if (v2_cond)
-                                v2_pos = h_low * w + w_high;
-                            if (v3_cond)
-                                v3_pos = h_high * w + w_low;
-                            if (v4_cond)
-                                v4_pos = h_high * w + w_high;
-
-                            w1 = hh * hw;
-                            w2 = hh * lw;
-                            w3 = lh * hw;
-                            w4 = lh * lw;
-                        }
-                        const float w1s[out_elempack] = {w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1, w1};
-                        const float* w1_ptr = w1s;
-                        const float w2s[out_elempack] = {w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2, w2};
-                        const float* w2_ptr = w2s;
-                        const float w3s[out_elempack] = {w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3, w3};
-                        const float* w3_ptr = w3s;
-                        const float w4s[out_elempack] = {w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4, w4};
-                        const float* w4_ptr = w4s;
-                        const float masks[out_elempack] = {mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_, mask_};
-                        const float* mask_ptr = masks;
-
-                        for (int ic = 0; ic < inch; ic++)
-                        {
-                            const float* data_im_ptr = bottom_blob.channel(ic);
-                            __m512 _val_channel0 = _mm512_loadu_ps(zeros_ptr);
-                            __m512 _val_channel1 = _val_channel0;
-                            __m512 _val_channel2 = _val_channel0;
-                            __m512 _val_channel3 = _val_channel0;
-                            __m512 _val_channel4 = _val_channel0;
-                            __m512 _val_channel5 = _val_channel0;
-                            __m512 _val_channel6 = _val_channel0;
-                            __m512 _val_channel7 = _val_channel0;
-                            if (cond)
-                            {
-                                __m512 _v1_channel0 = _val_channel0;
-                                __m512 _v1_channel1 = _val_channel0;
-                                __m512 _v1_channel2 = _val_channel0;
-                                __m512 _v1_channel3 = _val_channel0;
-                                __m512 _v1_channel4 = _val_channel0;
-                                __m512 _v1_channel5 = _val_channel0;
-                                __m512 _v1_channel6 = _val_channel0;
-                                __m512 _v1_channel7 = _val_channel0;
-                                __m512 _v2_channel0 = _val_channel0;
-                                __m512 _v2_channel1 = _val_channel0;
-                                __m512 _v2_channel2 = _val_channel0;
-                                __m512 _v2_channel3 = _val_channel0;
-                                __m512 _v2_channel4 = _val_channel0;
-                                __m512 _v2_channel5 = _val_channel0;
-                                __m512 _v2_channel6 = _val_channel0;
-                                __m512 _v2_channel7 = _val_channel0;
-                                __m512 _v3_channel0 = _val_channel0;
-                                __m512 _v3_channel1 = _val_channel0;
-                                __m512 _v3_channel2 = _val_channel0;
-                                __m512 _v3_channel3 = _val_channel0;
-                                __m512 _v3_channel4 = _val_channel0;
-                                __m512 _v3_channel5 = _val_channel0;
-                                __m512 _v3_channel6 = _val_channel0;
-                                __m512 _v3_channel7 = _val_channel0;
-                                __m512 _v4_channel0 = _val_channel0;
-                                __m512 _v4_channel1 = _val_channel0;
-                                __m512 _v4_channel2 = _val_channel0;
-                                __m512 _v4_channel3 = _val_channel0;
-                                __m512 _v4_channel4 = _val_channel0;
-                                __m512 _v4_channel5 = _val_channel0;
-                                __m512 _v4_channel6 = _val_channel0;
-                                __m512 _v4_channel7 = _val_channel0;
-                                if (v1_cond)
-                                {
-                                    _v1_channel0 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack]);
-                                    _v1_channel1 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 1]);
-                                    _v1_channel2 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 2]);
-                                    _v1_channel3 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 3]);
-                                    _v1_channel4 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 4]);
-                                    _v1_channel5 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 5]);
-                                    _v1_channel6 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 6]);
-                                    _v1_channel7 = _mm512_set1_ps(data_im_ptr[v1_pos * elempack + 7]);
-                                }
-                                if (v2_cond)
-                                {
-                                    _v2_channel0 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack]);
-                                    _v2_channel1 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 1]);
-                                    _v2_channel2 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 2]);
-                                    _v2_channel3 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 3]);
-                                    _v2_channel4 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 4]);
-                                    _v2_channel5 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 5]);
-                                    _v2_channel6 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 6]);
-                                    _v2_channel7 = _mm512_set1_ps(data_im_ptr[v2_pos * elempack + 7]);
-                                }
-                                if (v3_cond)
-                                {
-                                    _v3_channel0 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack]);
-                                    _v3_channel1 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 1]);
-                                    _v3_channel2 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 2]);
-                                    _v3_channel3 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 3]);
-                                    _v3_channel4 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 4]);
-                                    _v3_channel5 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 5]);
-                                    _v3_channel6 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 6]);
-                                    _v3_channel7 = _mm512_set1_ps(data_im_ptr[v3_pos * elempack + 7]);
-                                }
-                                if (v4_cond)
-                                {
-                                    _v4_channel0 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack]);
-                                    _v4_channel1 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 1]);
-                                    _v4_channel2 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 2]);
-                                    _v4_channel3 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 3]);
-                                    _v4_channel4 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 4]);
-                                    _v4_channel5 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 5]);
-                                    _v4_channel6 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 6]);
-                                    _v4_channel7 = _mm512_set1_ps(data_im_ptr[v4_pos * elempack + 7]);
-                                }
-                                __m512 _w1 = _mm512_loadu_ps(w1_ptr);
-                                __m512 _w2 = _mm512_loadu_ps(w2_ptr);
-                                __m512 _w3 = _mm512_loadu_ps(w3_ptr);
-                                __m512 _w4 = _mm512_loadu_ps(w4_ptr);
-                                _val_channel0 = _mm512_fmadd_ps(_v1_channel0, _w1, _val_channel0);
-                                _val_channel0 = _mm512_fmadd_ps(_v2_channel0, _w2, _val_channel0);
-                                _val_channel0 = _mm512_fmadd_ps(_v3_channel0, _w3, _val_channel0);
-                                _val_channel0 = _mm512_fmadd_ps(_v4_channel0, _w4, _val_channel0);
-                                _val_channel1 = _mm512_fmadd_ps(_v1_channel1, _w1, _val_channel1);
-                                _val_channel1 = _mm512_fmadd_ps(_v2_channel1, _w2, _val_channel1);
-                                _val_channel1 = _mm512_fmadd_ps(_v3_channel1, _w3, _val_channel1);
-                                _val_channel1 = _mm512_fmadd_ps(_v4_channel1, _w4, _val_channel1);
-                                _val_channel2 = _mm512_fmadd_ps(_v1_channel2, _w1, _val_channel2);
-                                _val_channel2 = _mm512_fmadd_ps(_v2_channel2, _w2, _val_channel2);
-                                _val_channel2 = _mm512_fmadd_ps(_v3_channel2, _w3, _val_channel2);
-                                _val_channel2 = _mm512_fmadd_ps(_v4_channel2, _w4, _val_channel2);
-                                _val_channel3 = _mm512_fmadd_ps(_v1_channel3, _w1, _val_channel3);
-                                _val_channel3 = _mm512_fmadd_ps(_v2_channel3, _w2, _val_channel3);
-                                _val_channel3 = _mm512_fmadd_ps(_v3_channel3, _w3, _val_channel3);
-                                _val_channel3 = _mm512_fmadd_ps(_v4_channel3, _w4, _val_channel3);
-                                _val_channel4 = _mm512_fmadd_ps(_v1_channel4, _w1, _val_channel4);
-                                _val_channel4 = _mm512_fmadd_ps(_v2_channel4, _w2, _val_channel4);
-                                _val_channel4 = _mm512_fmadd_ps(_v3_channel4, _w3, _val_channel4);
-                                _val_channel4 = _mm512_fmadd_ps(_v4_channel4, _w4, _val_channel4);
-                                _val_channel5 = _mm512_fmadd_ps(_v1_channel5, _w1, _val_channel5);
-                                _val_channel5 = _mm512_fmadd_ps(_v2_channel5, _w2, _val_channel5);
-                                _val_channel5 = _mm512_fmadd_ps(_v3_channel5, _w3, _val_channel5);
-                                _val_channel5 = _mm512_fmadd_ps(_v4_channel5, _w4, _val_channel5);
-                                _val_channel6 = _mm512_fmadd_ps(_v1_channel6, _w1, _val_channel6);
-                                _val_channel6 = _mm512_fmadd_ps(_v2_channel6, _w2, _val_channel6);
-                                _val_channel6 = _mm512_fmadd_ps(_v3_channel6, _w3, _val_channel6);
-                                _val_channel6 = _mm512_fmadd_ps(_v4_channel6, _w4, _val_channel6);
-                                _val_channel7 = _mm512_fmadd_ps(_v1_channel7, _w1, _val_channel7);
-                                _val_channel7 = _mm512_fmadd_ps(_v2_channel7, _w2, _val_channel7);
-                                _val_channel7 = _mm512_fmadd_ps(_v3_channel7, _w3, _val_channel7);
-                                _val_channel7 = _mm512_fmadd_ps(_v4_channel7, _w4, _val_channel7);
-                            }
-                            if (has_mask)
-                            {
-                                __m512 _mask = _mm512_loadu_ps(mask_ptr);
-                                _val_channel0 = _mm512_mul_ps(_val_channel0, _mask);
-                                _val_channel1 = _mm512_mul_ps(_val_channel1, _mask);
-                                _val_channel2 = _mm512_mul_ps(_val_channel2, _mask);
-                                _val_channel3 = _mm512_mul_ps(_val_channel3, _mask);
-                                _val_channel4 = _mm512_mul_ps(_val_channel4, _mask);
-                                _val_channel5 = _mm512_mul_ps(_val_channel5, _mask);
-                                _val_channel6 = _mm512_mul_ps(_val_channel6, _mask);
-                                _val_channel7 = _mm512_mul_ps(_val_channel7, _mask);
-                            }
-                            __m512 _conv_w0 = _mm512_load_ps(kptr);
-                            __m512 _conv_w1 = _mm512_load_ps(kptr + out_elempack); // 1 * out_elempack
-                            _sum = _mm512_fmadd_ps(_val_channel0, _conv_w0, _sum);
-                            _sum = _mm512_fmadd_ps(_val_channel1, _conv_w1, _sum);
-                            __m512 _conv_w2 = _mm512_load_ps(kptr + 32); // 2 * out_elempack
-                            __m512 _conv_w3 = _mm512_load_ps(kptr + 48); // 3 * out_elempack
-                            _sum = _mm512_fmadd_ps(_val_channel2, _conv_w2, _sum);
-                            _sum = _mm512_fmadd_ps(_val_channel3, _conv_w3, _sum);
-                            __m512 _conv_w4 = _mm512_load_ps(kptr + 64); // 4 * out_elempack
-                            __m512 _conv_w5 = _mm512_load_ps(kptr + 80); // 5 * out_elempack
-                            _sum = _mm512_fmadd_ps(_val_channel4, _conv_w4, _sum);
-                            _sum = _mm512_fmadd_ps(_val_channel5, _conv_w5, _sum);
-                            __m512 _conv_w6 = _mm512_load_ps(kptr + 96);  // 6 * out_elempack
-                            __m512 _conv_w7 = _mm512_load_ps(kptr + 112); // 7 * out_elempack
-                            _sum = _mm512_fmadd_ps(_val_channel6, _conv_w6, _sum);
-                            _sum = _mm512_fmadd_ps(_val_channel7, _conv_w7, _sum);
-                            kptr += wstep;
-                        }
-                    }
-                }
-                _sum = activation_avx512(_sum, activation_type, activation_params);
-                _mm512_store_ps(outptr + (h_col * outw + w_col) * out_elempack, _sum);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8to4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8to4.h
deleted file mode 100644
index 849bbd7d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_pack8to4.h
+++ /dev/null
@@ -1,307 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void deformableconv2d_pack8to4_avx(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
-{
-    const Mat& bottom_blob = bottom_blobs[0];
-    const Mat& offset = bottom_blobs[1];
-    const bool has_mask = (bottom_blobs.size() == 3);
-    const bool offset_not_pack = offset.elempack == 1;
-    const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
-
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-    int outch = top_blob.c;
-    const int size = outw * outh;
-    const int maxk = kernel_w * kernel_h;
-
-    const float* bias_data_ptr = bias_data;
-    const int elempack = 8;
-    const int out_elempack = 4;
-    const int wstep = out_elempack * elempack;
-    const float zeros[out_elempack] = {0.f, 0.f, 0.f, 0.f};
-    const float* zeros_ptr = zeros;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int h_col = 0; h_col < outh; h_col++)
-    {
-        for (int w_col = 0; w_col < outw; w_col++)
-        {
-            int h_in = h_col * stride_h - pad_top;
-            int w_in = w_col * stride_w - pad_left;
-            for (int oc = 0; oc < outch; oc++)
-            {
-                const float* kptr = weight_data_packed.channel(oc);
-                float* outptr = top_blob.channel(oc);
-                __m128 _sum = _mm_setzero_ps();
-                if (bias_data_ptr)
-                    _sum = _mm_loadu_ps(bias_data_ptr + oc * out_elempack);
-                for (int i = 0; i < kernel_h; i++)
-                {
-                    for (int j = 0; j < kernel_w; j++)
-                    {
-                        float offset_h = 0.f;
-                        float offset_w = 0.f;
-                        float mask_ = 1.f;
-                        if (offset_not_pack)
-                        {
-                            offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
-                            offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
-                        }
-                        else
-                        {
-                            const int y_c = (i * kernel_w + j) * 2;
-                            const int x_c = (i * kernel_w + j) * 2 + 1;
-                            offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
-                            offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
-                        }
-                        if (has_mask)
-                        {
-                            const Mat& mask = bottom_blobs[2];
-                            if (mask_not_pack)
-                            {
-                                mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
-                            }
-                            else
-                            {
-                                const int m_c = i * kernel_w + j;
-                                mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
-                            }
-                        }
-                        const float h_im = h_in + i * dilation_h + offset_h;
-                        const float w_im = w_in + j * dilation_w + offset_w;
-
-                        // Bilinear
-                        const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
-                        float w1 = 0.f;
-                        float w2 = 0.f;
-                        float w3 = 0.f;
-                        float w4 = 0.f;
-                        bool v1_cond = false;
-                        bool v2_cond = false;
-                        bool v3_cond = false;
-                        bool v4_cond = false;
-                        int v1_pos = 0;
-                        int v2_pos = 0;
-                        int v3_pos = 0;
-                        int v4_pos = 0;
-                        if (cond)
-                        {
-                            int h_low = (int)floorf(h_im);
-                            int w_low = (int)floorf(w_im);
-                            int h_high = h_low + 1;
-                            int w_high = w_low + 1;
-
-                            float lh = h_im - h_low;
-                            float lw = w_im - w_low;
-                            float hh = 1 - lh;
-                            float hw = 1 - lw;
-
-                            v1_cond = (h_low >= 0 && w_low >= 0);
-                            v2_cond = (h_low >= 0 && w_high <= w - 1);
-                            v3_cond = (h_high <= h - 1 && w_low >= 0);
-                            v4_cond = (h_high <= h - 1 && w_high <= w - 1);
-                            if (v1_cond)
-                                v1_pos = h_low * w + w_low;
-                            if (v2_cond)
-                                v2_pos = h_low * w + w_high;
-                            if (v3_cond)
-                                v3_pos = h_high * w + w_low;
-                            if (v4_cond)
-                                v4_pos = h_high * w + w_high;
-
-                            w1 = hh * hw;
-                            w2 = hh * lw;
-                            w3 = lh * hw;
-                            w4 = lh * lw;
-                        }
-                        const float w1s[out_elempack] = {w1, w1, w1, w1};
-                        const float* w1_ptr = w1s;
-                        const float w2s[out_elempack] = {w2, w2, w2, w2};
-                        const float* w2_ptr = w2s;
-                        const float w3s[out_elempack] = {w3, w3, w3, w3};
-                        const float* w3_ptr = w3s;
-                        const float w4s[out_elempack] = {w4, w4, w4, w4};
-                        const float* w4_ptr = w4s;
-                        const float masks[out_elempack] = {mask_, mask_, mask_, mask_};
-                        const float* mask_ptr = masks;
-
-                        for (int ic = 0; ic < inch; ic++)
-                        {
-                            const float* data_im_ptr = bottom_blob.channel(ic);
-                            __m128 _val_channel0 = _mm_loadu_ps(zeros_ptr);
-                            __m128 _val_channel1 = _val_channel0;
-                            __m128 _val_channel2 = _val_channel0;
-                            __m128 _val_channel3 = _val_channel0;
-                            __m128 _val_channel4 = _val_channel0;
-                            __m128 _val_channel5 = _val_channel0;
-                            __m128 _val_channel6 = _val_channel0;
-                            __m128 _val_channel7 = _val_channel0;
-                            if (cond)
-                            {
-                                __m128 _v1_channel0 = _val_channel0;
-                                __m128 _v1_channel1 = _val_channel0;
-                                __m128 _v1_channel2 = _val_channel0;
-                                __m128 _v1_channel3 = _val_channel0;
-                                __m128 _v1_channel4 = _val_channel0;
-                                __m128 _v1_channel5 = _val_channel0;
-                                __m128 _v1_channel6 = _val_channel0;
-                                __m128 _v1_channel7 = _val_channel0;
-                                __m128 _v2_channel0 = _val_channel0;
-                                __m128 _v2_channel1 = _val_channel0;
-                                __m128 _v2_channel2 = _val_channel0;
-                                __m128 _v2_channel3 = _val_channel0;
-                                __m128 _v2_channel4 = _val_channel0;
-                                __m128 _v2_channel5 = _val_channel0;
-                                __m128 _v2_channel6 = _val_channel0;
-                                __m128 _v2_channel7 = _val_channel0;
-                                __m128 _v3_channel0 = _val_channel0;
-                                __m128 _v3_channel1 = _val_channel0;
-                                __m128 _v3_channel2 = _val_channel0;
-                                __m128 _v3_channel3 = _val_channel0;
-                                __m128 _v3_channel4 = _val_channel0;
-                                __m128 _v3_channel5 = _val_channel0;
-                                __m128 _v3_channel6 = _val_channel0;
-                                __m128 _v3_channel7 = _val_channel0;
-                                __m128 _v4_channel0 = _val_channel0;
-                                __m128 _v4_channel1 = _val_channel0;
-                                __m128 _v4_channel2 = _val_channel0;
-                                __m128 _v4_channel3 = _val_channel0;
-                                __m128 _v4_channel4 = _val_channel0;
-                                __m128 _v4_channel5 = _val_channel0;
-                                __m128 _v4_channel6 = _val_channel0;
-                                __m128 _v4_channel7 = _val_channel0;
-                                if (v1_cond)
-                                {
-                                    _v1_channel0 = _mm_load1_ps(data_im_ptr + v1_pos * elempack);
-                                    _v1_channel1 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 1);
-                                    _v1_channel2 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 2);
-                                    _v1_channel3 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 3);
-                                    _v1_channel4 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 4);
-                                    _v1_channel5 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 5);
-                                    _v1_channel6 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 6);
-                                    _v1_channel7 = _mm_load1_ps(data_im_ptr + v1_pos * elempack + 7);
-                                }
-                                if (v2_cond)
-                                {
-                                    _v2_channel0 = _mm_load1_ps(data_im_ptr + v2_pos * elempack);
-                                    _v2_channel1 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 1);
-                                    _v2_channel2 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 2);
-                                    _v2_channel3 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 3);
-                                    _v2_channel4 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 4);
-                                    _v2_channel5 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 5);
-                                    _v2_channel6 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 6);
-                                    _v2_channel7 = _mm_load1_ps(data_im_ptr + v2_pos * elempack + 7);
-                                }
-                                if (v3_cond)
-                                {
-                                    _v3_channel0 = _mm_load1_ps(data_im_ptr + v3_pos * elempack);
-                                    _v3_channel1 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 1);
-                                    _v3_channel2 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 2);
-                                    _v3_channel3 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 3);
-                                    _v3_channel4 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 4);
-                                    _v3_channel5 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 5);
-                                    _v3_channel6 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 6);
-                                    _v3_channel7 = _mm_load1_ps(data_im_ptr + v3_pos * elempack + 7);
-                                }
-                                if (v4_cond)
-                                {
-                                    _v4_channel0 = _mm_load1_ps(data_im_ptr + v4_pos * elempack);
-                                    _v4_channel1 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 1);
-                                    _v4_channel2 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 2);
-                                    _v4_channel3 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 3);
-                                    _v4_channel4 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 4);
-                                    _v4_channel5 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 5);
-                                    _v4_channel6 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 6);
-                                    _v4_channel7 = _mm_load1_ps(data_im_ptr + v4_pos * elempack + 7);
-                                }
-                                __m128 _w1 = _mm_loadu_ps(w1_ptr);
-                                __m128 _w2 = _mm_loadu_ps(w2_ptr);
-                                __m128 _w3 = _mm_loadu_ps(w3_ptr);
-                                __m128 _w4 = _mm_loadu_ps(w4_ptr);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v1_channel0, _w1, _val_channel0);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v2_channel0, _w2, _val_channel0);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v3_channel0, _w3, _val_channel0);
-                                _val_channel0 = _mm_comp_fmadd_ps(_v4_channel0, _w4, _val_channel0);
-                                _val_channel1 = _mm_comp_fmadd_ps(_v1_channel1, _w1, _val_channel1);
-                                _val_channel1 = _mm_comp_fmadd_ps(_v2_channel1, _w2, _val_channel1);
-                                _val_channel1 = _mm_comp_fmadd_ps(_v3_channel1, _w3, _val_channel1);
-                                _val_channel1 = _mm_comp_fmadd_ps(_v4_channel1, _w4, _val_channel1);
-                                _val_channel2 = _mm_comp_fmadd_ps(_v1_channel2, _w1, _val_channel2);
-                                _val_channel2 = _mm_comp_fmadd_ps(_v2_channel2, _w2, _val_channel2);
-                                _val_channel2 = _mm_comp_fmadd_ps(_v3_channel2, _w3, _val_channel2);
-                                _val_channel2 = _mm_comp_fmadd_ps(_v4_channel2, _w4, _val_channel2);
-                                _val_channel3 = _mm_comp_fmadd_ps(_v1_channel3, _w1, _val_channel3);
-                                _val_channel3 = _mm_comp_fmadd_ps(_v2_channel3, _w2, _val_channel3);
-                                _val_channel3 = _mm_comp_fmadd_ps(_v3_channel3, _w3, _val_channel3);
-                                _val_channel3 = _mm_comp_fmadd_ps(_v4_channel3, _w4, _val_channel3);
-                                _val_channel4 = _mm_comp_fmadd_ps(_v1_channel4, _w1, _val_channel4);
-                                _val_channel4 = _mm_comp_fmadd_ps(_v2_channel4, _w2, _val_channel4);
-                                _val_channel4 = _mm_comp_fmadd_ps(_v3_channel4, _w3, _val_channel4);
-                                _val_channel4 = _mm_comp_fmadd_ps(_v4_channel4, _w4, _val_channel4);
-                                _val_channel5 = _mm_comp_fmadd_ps(_v1_channel5, _w1, _val_channel5);
-                                _val_channel5 = _mm_comp_fmadd_ps(_v2_channel5, _w2, _val_channel5);
-                                _val_channel5 = _mm_comp_fmadd_ps(_v3_channel5, _w3, _val_channel5);
-                                _val_channel5 = _mm_comp_fmadd_ps(_v4_channel5, _w4, _val_channel5);
-                                _val_channel6 = _mm_comp_fmadd_ps(_v1_channel6, _w1, _val_channel6);
-                                _val_channel6 = _mm_comp_fmadd_ps(_v2_channel6, _w2, _val_channel6);
-                                _val_channel6 = _mm_comp_fmadd_ps(_v3_channel6, _w3, _val_channel6);
-                                _val_channel6 = _mm_comp_fmadd_ps(_v4_channel6, _w4, _val_channel6);
-                                _val_channel7 = _mm_comp_fmadd_ps(_v1_channel7, _w1, _val_channel7);
-                                _val_channel7 = _mm_comp_fmadd_ps(_v2_channel7, _w2, _val_channel7);
-                                _val_channel7 = _mm_comp_fmadd_ps(_v3_channel7, _w3, _val_channel7);
-                                _val_channel7 = _mm_comp_fmadd_ps(_v4_channel7, _w4, _val_channel7);
-                            }
-                            if (has_mask)
-                            {
-                                __m128 _mask = _mm_loadu_ps(mask_ptr);
-                                _val_channel0 = _mm_mul_ps(_val_channel0, _mask);
-                                _val_channel1 = _mm_mul_ps(_val_channel1, _mask);
-                                _val_channel2 = _mm_mul_ps(_val_channel2, _mask);
-                                _val_channel3 = _mm_mul_ps(_val_channel3, _mask);
-                                _val_channel4 = _mm_mul_ps(_val_channel4, _mask);
-                                _val_channel5 = _mm_mul_ps(_val_channel5, _mask);
-                                _val_channel6 = _mm_mul_ps(_val_channel6, _mask);
-                                _val_channel7 = _mm_mul_ps(_val_channel7, _mask);
-                            }
-                            __m128 _conv_w0 = _mm_load_ps(kptr);
-                            __m128 _conv_w1 = _mm_load_ps(kptr + out_elempack); // 1 * out_elempack
-                            _sum = _mm_comp_fmadd_ps(_val_channel0, _conv_w0, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val_channel1, _conv_w1, _sum);
-                            __m128 _conv_w2 = _mm_load_ps(kptr + 8);  // 2 * out_elempack
-                            __m128 _conv_w3 = _mm_load_ps(kptr + 12); // 3 * out_elempack
-                            _sum = _mm_comp_fmadd_ps(_val_channel2, _conv_w2, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val_channel3, _conv_w3, _sum);
-                            __m128 _conv_w4 = _mm_load_ps(kptr + 16); // 4 * out_elempack
-                            __m128 _conv_w5 = _mm_load_ps(kptr + 20); // 5 * out_elempack
-                            _sum = _mm_comp_fmadd_ps(_val_channel4, _conv_w4, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val_channel5, _conv_w5, _sum);
-                            __m128 _conv_w6 = _mm_load_ps(kptr + 24); // 6 * out_elempack
-                            __m128 _conv_w7 = _mm_load_ps(kptr + 28); // 7 * out_elempack
-                            _sum = _mm_comp_fmadd_ps(_val_channel6, _conv_w6, _sum);
-                            _sum = _mm_comp_fmadd_ps(_val_channel7, _conv_w7, _sum);
-                            kptr += wstep;
-                        }
-                    }
-                }
-                _sum = activation_sse(_sum, activation_type, activation_params);
-                _mm_storeu_ps(outptr + (h_col * outw + w_col) * out_elempack, _sum);
-            }
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_x86.h
deleted file mode 100644
index 66cce21a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/deformableconv2d_x86.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DEFORMABLECONV2D_X86_H
-#define LAYER_DEFORMABLECONV2D_X86_H
-
-#include "deformableconv2d.h"
-
-namespace ncnn {
-
-class DeformableConv2D_x86 : public DeformableConv2D
-{
-public:
-    DeformableConv2D_x86();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    Layer* activation;
-
-    Mat weight_data_tm;
-
-    Layer* gemm;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DEFORMABLECONV2D_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/dequantize_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/dequantize_x86.h
deleted file mode 100644
index 52bfcaed..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/dequantize_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DEQUANTIZE_X86_H
-#define LAYER_DEQUANTIZE_X86_H
-
-#include "dequantize.h"
-
-namespace ncnn {
-
-class Dequantize_x86 : public Dequantize
-{
-public:
-    Dequantize_x86();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DEQUANTIZE_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/dropout_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/dropout_x86.h
deleted file mode 100644
index d44a8987..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/dropout_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_DROPOUT_X86_H
-#define LAYER_DROPOUT_X86_H
-
-#include "dropout.h"
-
-namespace ncnn {
-
-class Dropout_x86 : public Dropout
-{
-public:
-    Dropout_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_DROPOUT_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/eltwise_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/eltwise_x86.h
deleted file mode 100644
index e941817a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/eltwise_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ELTWISE_X86_H
-#define LAYER_ELTWISE_X86_H
-
-#include "eltwise.h"
-
-namespace ncnn {
-
-class Eltwise_x86 : public Eltwise
-{
-public:
-    Eltwise_x86();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ELTWISE_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/elu_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/elu_x86.h
deleted file mode 100644
index 6da00490..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/elu_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ELU_X86_H
-#define LAYER_ELU_X86_H
-
-#include "elu.h"
-
-namespace ncnn {
-
-class ELU_x86 : public ELU
-{
-public:
-    ELU_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ELU_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/flatten_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/flatten_x86.h
deleted file mode 100644
index 29820121..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/flatten_x86.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_FLATTEN_X86_H
-#define LAYER_FLATTEN_X86_H
-
-#include "flatten.h"
-
-namespace ncnn {
-
-class Flatten_x86 : public Flatten
-{
-public:
-    Flatten_x86();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_FLATTEN_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gelu_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gelu_x86.h
deleted file mode 100644
index ba4b43e6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gelu_x86.h
+++ /dev/null
@@ -1,33 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GELU_X86_H
-#define LAYER_GELU_X86_H
-
-#include "gelu.h"
-
-namespace ncnn {
-
-class GELU_x86 : public GELU
-{
-public:
-    GELU_x86();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GELU_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gemm_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gemm_x86.h
deleted file mode 100644
index 6f8eb4a8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gemm_x86.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GEMM_X86_H
-#define LAYER_GEMM_X86_H
-
-#include "gemm.h"
-
-namespace ncnn {
-
-class Gemm_x86 : public Gemm
-{
-public:
-    Gemm_x86();
-
-    virtual int create_pipeline(const Option& opt);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    int nT;
-    Mat AT_data;
-    Mat BT_data;
-    Mat CT_data;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GEMM_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bicubic_apply_interpolation.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bicubic_apply_interpolation.h
deleted file mode 100644
index 0b7be771..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bicubic_apply_interpolation.h
+++ /dev/null
@@ -1,288 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-static void cubic_interp1d_p16(__m512& coeffs0, __m512& coeffs1, __m512& coeffs2, __m512& coeffs3, const __m512& tx)
-{
-    const __m512 A = _mm512_set1_ps(-0.75f);
-
-    const __m512 x0 = _mm512_add_ps(tx, _mm512_set1_ps(1.0f));
-    const __m512& x1 = tx;
-    const __m512 x2 = _mm512_sub_ps(_mm512_set1_ps(1.0f), tx);
-
-    coeffs0 = _mm512_sub_ps(_mm512_mul_ps(_mm512_add_ps(_mm512_mul_ps(_mm512_sub_ps(_mm512_mul_ps(A, x0), _mm512_mul_ps(_mm512_set1_ps(5.0f), A)), x0), _mm512_mul_ps(_mm512_set1_ps(8.0f), A)), x0), _mm512_mul_ps(_mm512_set1_ps(4), A));
-    coeffs1 = _mm512_add_ps(_mm512_mul_ps(_mm512_mul_ps(_mm512_sub_ps(_mm512_mul_ps(_mm512_add_ps(A, _mm512_set1_ps(2.0f)), x1), _mm512_add_ps(A, _mm512_set1_ps(3.0f))), x1), x1), _mm512_set1_ps(1.0f));
-    coeffs2 = _mm512_add_ps(_mm512_mul_ps(_mm512_mul_ps(_mm512_sub_ps(_mm512_mul_ps(_mm512_add_ps(A, _mm512_set1_ps(2.0f)), x2), _mm512_add_ps(A, _mm512_set1_ps(3.0f))), x2), x2), _mm512_set1_ps(1.0f));
-    coeffs3 = _mm512_sub_ps(_mm512_sub_ps(_mm512_sub_ps(_mm512_set1_ps(1.0f), coeffs0), coeffs1), coeffs2);
-}
-
-static void gridsample_2d_bicubic_apply_interpolation_p16(const Mat& src, Mat& dst, Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int grid_size = outw * outh;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const float* offset_value_ptr = offset_value.channel(0);
-
-        for (int i = 0; i < grid_size; i++)
-        {
-            __m512 x_coeffs0, x_coeffs1, x_coeffs2, x_coeffs3;
-            __m512 y_coeffs0, y_coeffs1, y_coeffs2, y_coeffs3;
-            __m512 value_f[4];
-            cubic_interp1d_p16(x_coeffs0, x_coeffs1, x_coeffs2, x_coeffs3, _mm512_set1_ps(offset_value_ptr[0]));
-            cubic_interp1d_p16(y_coeffs0, y_coeffs1, y_coeffs2, y_coeffs3, _mm512_set1_ps(offset_value_ptr[1]));
-
-            const int* offset_ptr = (int*)offset_value_ptr + 2;
-
-            for (int ii = 0; ii < 4; ii++)
-            {
-                __m512 x0_val = offset_ptr[0] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[0]) : _mm512_set1_ps(0);
-                __m512 x1_val = offset_ptr[1] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[1]) : _mm512_set1_ps(0);
-                __m512 x2_val = offset_ptr[2] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[2]) : _mm512_set1_ps(0);
-                __m512 x3_val = offset_ptr[3] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[3]) : _mm512_set1_ps(0);
-
-                value_f[ii] = _mm512_mul_ps(x_coeffs0, x0_val);
-                value_f[ii] = _mm512_fmadd_ps(x_coeffs1, x1_val, value_f[ii]);
-                value_f[ii] = _mm512_fmadd_ps(x_coeffs2, x2_val, value_f[ii]);
-                value_f[ii] = _mm512_fmadd_ps(x_coeffs3, x3_val, value_f[ii]);
-
-                offset_ptr += 4;
-            }
-
-            __m512 _v = _mm512_mul_ps(y_coeffs0, value_f[0]);
-            _v = _mm512_fmadd_ps(y_coeffs1, value_f[1], _v);
-            _v = _mm512_fmadd_ps(y_coeffs2, value_f[2], _v);
-            _v = _mm512_fmadd_ps(y_coeffs3, value_f[3], _v);
-            _mm512_storeu_ps(dstptr, _v);
-
-            dstptr += 16;
-            offset_value_ptr += 18;
-        }
-    }
-}
-
-#endif // __AVX512F__
-static void cubic_interp1d_p8(__m256& coeffs0, __m256& coeffs1, __m256& coeffs2, __m256& coeffs3, const __m256& tx)
-{
-    const __m256 A = _mm256_set1_ps(-0.75f);
-
-    const __m256 x0 = _mm256_add_ps(tx, _mm256_set1_ps(1));
-    const __m256& x1 = tx;
-    const __m256 x2 = _mm256_sub_ps(_mm256_set1_ps(1), tx);
-    //const __m256 x3 = _mm256_add_ps(x2, _mm256_set1_ps(1));
-
-    coeffs0 = _mm256_sub_ps(_mm256_mul_ps(_mm256_add_ps(_mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(A, x0), _mm256_mul_ps(_mm256_set1_ps(5.0f), A)), x0), _mm256_mul_ps(_mm256_set1_ps(8.0f), A)), x0), _mm256_mul_ps(_mm256_set1_ps(4), A));
-    coeffs1 = _mm256_add_ps(_mm256_mul_ps(_mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(_mm256_add_ps(A, _mm256_set1_ps(2.0f)), x1), _mm256_add_ps(A, _mm256_set1_ps(3.0f))), x1), x1), _mm256_set1_ps(1));
-    coeffs2 = _mm256_add_ps(_mm256_mul_ps(_mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(_mm256_add_ps(A, _mm256_set1_ps(2.0f)), x2), _mm256_add_ps(A, _mm256_set1_ps(3.0f))), x2), x2), _mm256_set1_ps(1));
-    coeffs3 = _mm256_sub_ps(_mm256_sub_ps(_mm256_sub_ps(_mm256_set1_ps(1), coeffs0), coeffs1), coeffs2);
-}
-
-static void gridsample_2d_bicubic_apply_interpolation_p8(const Mat& src, Mat& dst, Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int grid_size = outw * outh;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const float* offset_value_ptr = offset_value.channel(0);
-
-        for (int i = 0; i < grid_size; i++)
-        {
-            __m256 x_coeffs0, x_coeffs1, x_coeffs2, x_coeffs3;
-            __m256 y_coeffs0, y_coeffs1, y_coeffs2, y_coeffs3;
-            __m256 value_f[4];
-            cubic_interp1d_p8(x_coeffs0, x_coeffs1, x_coeffs2, x_coeffs3, _mm256_set1_ps(offset_value_ptr[0]));
-            cubic_interp1d_p8(y_coeffs0, y_coeffs1, y_coeffs2, y_coeffs3, _mm256_set1_ps(offset_value_ptr[1]));
-
-            const int* offset_ptr = (int*)offset_value_ptr + 2;
-
-            for (int ii = 0; ii < 4; ii++)
-            {
-                __m256 x0_val = offset_ptr[0] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[0]) : _mm256_set1_ps(0);
-                __m256 x1_val = offset_ptr[1] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[1]) : _mm256_set1_ps(0);
-                __m256 x2_val = offset_ptr[2] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[2]) : _mm256_set1_ps(0);
-                __m256 x3_val = offset_ptr[3] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[3]) : _mm256_set1_ps(0);
-
-                value_f[ii] = _mm256_mul_ps(x_coeffs0, x0_val);
-                value_f[ii] = _mm256_comp_fmadd_ps(x_coeffs1, x1_val, value_f[ii]);
-                value_f[ii] = _mm256_comp_fmadd_ps(x_coeffs2, x2_val, value_f[ii]);
-                value_f[ii] = _mm256_comp_fmadd_ps(x_coeffs3, x3_val, value_f[ii]);
-
-                offset_ptr += 4;
-            }
-
-            __m256 _v = _mm256_mul_ps(y_coeffs0, value_f[0]);
-            _v = _mm256_comp_fmadd_ps(y_coeffs1, value_f[1], _v);
-            _v = _mm256_comp_fmadd_ps(y_coeffs2, value_f[2], _v);
-            _v = _mm256_comp_fmadd_ps(y_coeffs3, value_f[3], _v);
-            _mm256_storeu_ps(dstptr, _v);
-
-            dstptr += 8;
-            offset_value_ptr += 18;
-        }
-    }
-}
-
-#endif // __AVX__
-static void cubic_interp1d_p4(__m128& coeffs0, __m128& coeffs1, __m128& coeffs2, __m128& coeffs3, const __m128& tx)
-{
-    const __m128 A = _mm_set_ps1(-0.75f);
-
-    const __m128 x0 = _mm_add_ps(tx, _mm_set_ps1(1.0f));
-    const __m128& x1 = tx;
-    const __m128 x2 = _mm_sub_ps(_mm_set_ps1(1.0f), tx);
-    //const __m128 x3 = _mm_add_ps(x2, _mm_set_ps1(1.0f));
-
-    coeffs0 = _mm_sub_ps(_mm_mul_ps(_mm_add_ps(_mm_mul_ps(_mm_sub_ps(_mm_mul_ps(A, x0), _mm_mul_ps(_mm_set_ps1(5.0f), A)), x0), _mm_mul_ps(_mm_set_ps1(8.0f), A)), x0), _mm_mul_ps(_mm_set_ps1(4), A));
-    coeffs1 = _mm_add_ps(_mm_mul_ps(_mm_mul_ps(_mm_sub_ps(_mm_mul_ps(_mm_add_ps(A, _mm_set_ps1(2.0f)), x1), _mm_add_ps(A, _mm_set_ps1(3.0f))), x1), x1), _mm_set_ps1(1.0f));
-    coeffs2 = _mm_add_ps(_mm_mul_ps(_mm_mul_ps(_mm_sub_ps(_mm_mul_ps(_mm_add_ps(A, _mm_set_ps1(2.0f)), x2), _mm_add_ps(A, _mm_set_ps1(3.0f))), x2), x2), _mm_set_ps1(1.0f));
-    coeffs3 = _mm_sub_ps(_mm_sub_ps(_mm_sub_ps(_mm_set_ps1(1.0f), coeffs0), coeffs1), coeffs2);
-}
-
-static void gridsample_2d_bicubic_apply_interpolation_p4(const Mat& src, Mat& dst, Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int grid_size = outw * outh;
-
-    __m128 x_coeffs0, x_coeffs1, x_coeffs2, x_coeffs3;
-    __m128 y_coeffs0, y_coeffs1, y_coeffs2, y_coeffs3;
-    __m128 value_f[4];
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const float* offset_value_ptr = offset_value.channel(0);
-
-        for (int i = 0; i < grid_size; i++)
-        {
-            cubic_interp1d_p4(x_coeffs0, x_coeffs1, x_coeffs2, x_coeffs3, _mm_set_ps1(offset_value_ptr[0]));
-            cubic_interp1d_p4(y_coeffs0, y_coeffs1, y_coeffs2, y_coeffs3, _mm_set_ps1(offset_value_ptr[1]));
-
-            const int* offset_ptr = (int*)offset_value_ptr + 2;
-
-            for (int ii = 0; ii < 4; ii++)
-            {
-                __m128 x0_val = offset_ptr[0] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[0]) : _mm_set1_ps(0);
-                __m128 x1_val = offset_ptr[1] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[1]) : _mm_set1_ps(0);
-                __m128 x2_val = offset_ptr[2] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[2]) : _mm_set1_ps(0);
-                __m128 x3_val = offset_ptr[3] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[3]) : _mm_set1_ps(0);
-
-                value_f[ii] = _mm_mul_ps(x_coeffs0, x0_val);
-                value_f[ii] = _mm_comp_fmadd_ps(x_coeffs1, x1_val, value_f[ii]);
-                value_f[ii] = _mm_comp_fmadd_ps(x_coeffs2, x2_val, value_f[ii]);
-                value_f[ii] = _mm_comp_fmadd_ps(x_coeffs3, x3_val, value_f[ii]);
-
-                offset_ptr += 4;
-            }
-
-            __m128 _v = _mm_mul_ps(y_coeffs0, value_f[0]);
-            _v = _mm_comp_fmadd_ps(y_coeffs1, value_f[1], _v);
-            _v = _mm_comp_fmadd_ps(y_coeffs2, value_f[2], _v);
-            _v = _mm_comp_fmadd_ps(y_coeffs3, value_f[3], _v);
-            _mm_storeu_ps(dstptr, _v);
-
-            dstptr += 4;
-            offset_value_ptr += 18;
-        }
-    }
-}
-
-#endif // __SSE2__
-
-static inline void cubic_interp1d(float& coeffs0, float& coeffs1, float& coeffs2, float& coeffs3, float fx)
-{
-    const float A = -0.75f;
-
-    float fx0 = fx + 1;
-    float fx1 = fx;
-    float fx2 = 1 - fx;
-    // float fx3 = 2 - fx;
-
-    coeffs0 = A * fx0 * fx0 * fx0 - 5 * A * fx0 * fx0 + 8 * A * fx0 - 4 * A;
-    coeffs1 = (A + 2) * fx1 * fx1 * fx1 - (A + 3) * fx1 * fx1 + 1;
-    coeffs2 = (A + 2) * fx2 * fx2 * fx2 - (A + 3) * fx2 * fx2 + 1;
-    coeffs3 = 1.f - coeffs0 - coeffs1 - coeffs2;
-}
-
-static void gridsample_2d_bicubic_apply_interpolation_p1(const Mat& src, Mat& dst, Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int grid_size = outw * outh;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const float* offset_value_ptr = offset_value.channel(0);
-
-        for (int x = 0; x < grid_size; x++)
-        {
-            float x_coeffs0, x_coeffs1, x_coeffs2, x_coeffs3;
-            float y_coeffs0, y_coeffs1, y_coeffs2, y_coeffs3;
-            float value_f[4];
-            cubic_interp1d(x_coeffs0, x_coeffs1, x_coeffs2, x_coeffs3, offset_value_ptr[0]);
-            cubic_interp1d(y_coeffs0, y_coeffs1, y_coeffs2, y_coeffs3, offset_value_ptr[1]);
-
-            const int* offset_ptr = (int*)offset_value_ptr + 2;
-
-            for (int ii = 0; ii < 4; ii++)
-            {
-                float x0_val = offset_ptr[0] >= 0 ? *(srcptr + offset_ptr[0]) : 0;
-                float x1_val = offset_ptr[1] >= 0 ? *(srcptr + offset_ptr[1]) : 0;
-                float x2_val = offset_ptr[2] >= 0 ? *(srcptr + offset_ptr[2]) : 0;
-                float x3_val = offset_ptr[3] >= 0 ? *(srcptr + offset_ptr[3]) : 0;
-
-                value_f[ii] = x_coeffs0 * x0_val;
-                value_f[ii] = x_coeffs1 * x1_val + value_f[ii];
-                value_f[ii] = x_coeffs2 * x2_val + value_f[ii];
-                value_f[ii] = x_coeffs3 * x3_val + value_f[ii];
-
-                offset_ptr += 4;
-            }
-
-            float _v = y_coeffs0 * value_f[0];
-            _v = y_coeffs1 * value_f[1] + _v;
-            _v = y_coeffs2 * value_f[2] + _v;
-            _v = y_coeffs3 * value_f[3] + _v;
-            *dstptr = _v;
-
-            dstptr++;
-            offset_value_ptr += 18;
-        }
-    }
-}
\ No newline at end of file
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bicubic_compute_blob.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bicubic_compute_blob.h
deleted file mode 100644
index 9006153d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bicubic_compute_blob.h
+++ /dev/null
@@ -1,299 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-template<GridSample::PaddingMode pd, bool align_corner>
-void gridsample_2d_bicubic_compute_blob(const Mat& src, const Mat& grid, Mat& offset_value, int permute_fusion)
-{
-    const int grid_size = grid.w * grid.h;
-
-    float* offset_value_ptr = offset_value.channel(0);
-
-    grid_sample_unormalize<align_corner> unormalize;
-    compute_coord<pd, align_corner> get_coord;
-
-    if (permute_fusion == 0)
-    {
-        for (int y = 0; y < grid.c; y++)
-        {
-            const float* gridptr = grid.channel(y);
-            int x = 0;
-#if __AVX__
-            for (; x + 15 < grid_size; x += 16)
-            {
-                __m256 gx = _mm256_loadu_ps(gridptr);
-                __m256 gy = _mm256_loadu_ps(gridptr + 8);
-
-                transpose2x8_ps(gx, gy);
-
-                gx = unormalize(_mm256_set1_ps(src.w), gx);
-                gy = unormalize(_mm256_set1_ps(src.h), gy);
-
-                __m256 gx_floor = _mm256_floor_ps(gx);
-                __m256 gy_floor = _mm256_floor_ps(gy);
-
-                __m256 tx = _mm256_sub_ps(gx, gx_floor);
-                __m256 ty = _mm256_sub_ps(gy, gy_floor);
-
-                __m256 gx0 = _mm256_add_ps(gx_floor, _mm256_set1_ps(-1));
-                __m256 gx1 = gx_floor;
-                __m256 gx2 = _mm256_add_ps(gx_floor, _mm256_set1_ps(1));
-                __m256 gx3 = _mm256_add_ps(gx2, _mm256_set1_ps(1));
-
-                gx0 = get_coord(_mm256_set1_ps(src.w), gx0);
-                gx1 = get_coord(_mm256_set1_ps(src.w), gx1);
-                gx2 = get_coord(_mm256_set1_ps(src.w), gx2);
-                gx3 = get_coord(_mm256_set1_ps(src.w), gx3);
-
-                __m256 x0_in_range = _mm256_and_ps(_mm256_cmp_ps(gx0, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), gx0, _CMP_GT_OS));
-                __m256 x1_in_range = _mm256_and_ps(_mm256_cmp_ps(gx1, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), gx1, _CMP_GT_OS));
-                __m256 x2_in_range = _mm256_and_ps(_mm256_cmp_ps(gx2, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), gx2, _CMP_GT_OS));
-                __m256 x3_in_range = _mm256_and_ps(_mm256_cmp_ps(gx3, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), gx3, _CMP_GT_OS));
-                __m256 v0_offset_f[4], v1_offset_f[4], v2_offset_f[4], v3_offset_f[4];
-                for (int i = 0; i < 4; i++)
-                {
-                    gy = _mm256_add_ps(gy_floor, _mm256_set1_ps(-1.0f + i));
-                    gy = get_coord(_mm256_set1_ps(src.h), gy);
-
-                    __m256 y_in_range = _mm256_and_ps(_mm256_cmp_ps(gy, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.h), gy, _CMP_GT_OS));
-
-                    __m256 gy_offset = _mm256_mul_ps(gy, _mm256_set1_ps(src.w));
-
-                    v0_offset_f[i] = _mm256_mul_ps(_mm256_add_ps(gy_offset, gx0), _mm256_set1_ps(src.elempack));
-                    v1_offset_f[i] = _mm256_mul_ps(_mm256_add_ps(gy_offset, gx1), _mm256_set1_ps(src.elempack));
-                    v2_offset_f[i] = _mm256_mul_ps(_mm256_add_ps(gy_offset, gx2), _mm256_set1_ps(src.elempack));
-                    v3_offset_f[i] = _mm256_mul_ps(_mm256_add_ps(gy_offset, gx3), _mm256_set1_ps(src.elempack));
-
-                    v0_offset_f[i] = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), v0_offset_f[i], _mm256_and_ps(x0_in_range, y_in_range));
-                    v1_offset_f[i] = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), v1_offset_f[i], _mm256_and_ps(x1_in_range, y_in_range));
-                    v2_offset_f[i] = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), v2_offset_f[i], _mm256_and_ps(x2_in_range, y_in_range));
-                    v3_offset_f[i] = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), v3_offset_f[i], _mm256_and_ps(x3_in_range, y_in_range));
-
-                    v0_offset_f[i] = _mm256_castsi256_ps(_mm256_cvtps_epi32(v0_offset_f[i]));
-                    v1_offset_f[i] = _mm256_castsi256_ps(_mm256_cvtps_epi32(v1_offset_f[i]));
-                    v2_offset_f[i] = _mm256_castsi256_ps(_mm256_cvtps_epi32(v2_offset_f[i]));
-                    v3_offset_f[i] = _mm256_castsi256_ps(_mm256_cvtps_epi32(v3_offset_f[i]));
-                }
-
-                transpose8x18_ps(tx, ty, v0_offset_f[0], v1_offset_f[0], v2_offset_f[0], v3_offset_f[0], v0_offset_f[1], v1_offset_f[1], v2_offset_f[1], v3_offset_f[1], v0_offset_f[2], v1_offset_f[2], v2_offset_f[2], v3_offset_f[2], v0_offset_f[3], v1_offset_f[3], v2_offset_f[3], v3_offset_f[3]);
-
-                _mm256_storeu_ps(offset_value_ptr, tx);
-                _mm256_storeu_ps(offset_value_ptr + 8, ty);
-                offset_value_ptr += 16;
-                for (int i = 0; i < 4; i++)
-                {
-                    _mm256_storeu_ps(offset_value_ptr, v0_offset_f[i]);
-                    _mm256_storeu_ps(offset_value_ptr + 8, v1_offset_f[i]);
-                    _mm256_storeu_ps(offset_value_ptr + 16, v2_offset_f[i]);
-                    _mm256_storeu_ps(offset_value_ptr + 24, v3_offset_f[i]);
-                    offset_value_ptr += 32;
-                }
-                gridptr += 16;
-            }
-
-#endif // __AVX__
-
-            for (; x < grid_size; x += 2)
-            {
-                float sample_x = *gridptr;
-                float sample_y = *(gridptr + 1);
-
-                sample_x = unormalize(src.w, sample_x);
-                sample_y = unormalize(src.h, sample_y);
-
-                int x1 = floorf(sample_x);
-                int y1 = floorf(sample_y);
-                int x0 = x1 - 1;
-                int x2 = x1 + 1;
-                int x3 = x1 + 2;
-
-                offset_value_ptr[0] = sample_x - static_cast<float>(x1);
-                offset_value_ptr[1] = sample_y - static_cast<float>(y1);
-
-                x1 = get_coord(src.w, x1);
-                x0 = get_coord(src.w, x0);
-                x2 = get_coord(src.w, x2);
-                x3 = get_coord(src.w, x3);
-
-                bool x1_in_range = (x1 > -1) & (x1 < src.w);
-                bool x0_in_range = (x0 > -1) & (x0 < src.w);
-                bool x2_in_range = (x2 > -1) & (x2 < src.w);
-                bool x3_in_range = (x3 > -1) & (x3 < src.w);
-
-                int* offset_ptr = (int*)offset_value_ptr + 2;
-
-                for (int i = 0; i < 4; i++)
-                {
-                    int gy = y1 + i - 1;
-                    gy = get_coord(src.h, gy);
-                    int offset_y = gy * src.w;
-
-                    bool y_in_range = (gy > -1) & (gy < src.h);
-
-                    bool v0_in_bound = (x0_in_range & y_in_range);
-                    bool v1_in_bound = (x1_in_range & y_in_range);
-                    bool v2_in_bound = (x2_in_range & y_in_range);
-                    bool v3_in_bound = (x3_in_range & y_in_range);
-
-                    offset_ptr[0] = v0_in_bound ? (offset_y + x0) * src.elempack : -1.0;
-                    offset_ptr[1] = v1_in_bound ? (offset_y + x1) * src.elempack : -1.0;
-                    offset_ptr[2] = v2_in_bound ? (offset_y + x2) * src.elempack : -1.0;
-                    offset_ptr[3] = v3_in_bound ? (offset_y + x3) * src.elempack : -1.0;
-
-                    offset_ptr += 4;
-                }
-
-                gridptr += 2;
-                offset_value_ptr += 18;
-            }
-        }
-    }
-    else
-    {
-        const float* gridptr_x = grid.channel(0);
-        const float* gridptr_y = grid.channel(1);
-
-        int x = 0;
-#if __AVX__
-        for (; x + 7 < grid_size; x += 8)
-        {
-            __m256 gx = _mm256_loadu_ps(gridptr_x);
-            __m256 gy = _mm256_loadu_ps(gridptr_y);
-
-            gx = unormalize(_mm256_set1_ps(src.w), gx);
-            gy = unormalize(_mm256_set1_ps(src.h), gy);
-
-            __m256 gx_floor = _mm256_floor_ps(gx);
-            __m256 gy_floor = _mm256_floor_ps(gy);
-
-            __m256 tx = _mm256_sub_ps(gx, gx_floor);
-            __m256 ty = _mm256_sub_ps(gy, gy_floor);
-
-            __m256 gx0 = _mm256_add_ps(gx_floor, _mm256_set1_ps(-1));
-            __m256 gx1 = gx_floor;
-            __m256 gx2 = _mm256_add_ps(gx_floor, _mm256_set1_ps(1));
-            __m256 gx3 = _mm256_add_ps(gx2, _mm256_set1_ps(1));
-
-            gx0 = get_coord(_mm256_set1_ps(src.w), gx0);
-            gx1 = get_coord(_mm256_set1_ps(src.w), gx1);
-            gx2 = get_coord(_mm256_set1_ps(src.w), gx2);
-            gx3 = get_coord(_mm256_set1_ps(src.w), gx3);
-
-            __m256 x0_in_range = _mm256_and_ps(_mm256_cmp_ps(gx0, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), gx0, _CMP_GT_OS));
-            __m256 x1_in_range = _mm256_and_ps(_mm256_cmp_ps(gx1, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), gx1, _CMP_GT_OS));
-            __m256 x2_in_range = _mm256_and_ps(_mm256_cmp_ps(gx2, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), gx2, _CMP_GT_OS));
-            __m256 x3_in_range = _mm256_and_ps(_mm256_cmp_ps(gx3, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), gx3, _CMP_GT_OS));
-
-            __m256 v0_offset_f[4], v1_offset_f[4], v2_offset_f[4], v3_offset_f[4];
-            for (int i = 0; i < 4; i++)
-            {
-                gy = _mm256_add_ps(gy_floor, _mm256_set1_ps(-1.0f + i));
-                gy = get_coord(_mm256_set1_ps(src.h), gy);
-
-                __m256 y_in_range = _mm256_and_ps(_mm256_cmp_ps(gy, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.h), gy, _CMP_GT_OS));
-
-                __m256 gy_offset = _mm256_mul_ps(gy, _mm256_set1_ps(src.w));
-
-                v0_offset_f[i] = _mm256_mul_ps(_mm256_add_ps(gy_offset, gx0), _mm256_set1_ps(src.elempack));
-                v1_offset_f[i] = _mm256_mul_ps(_mm256_add_ps(gy_offset, gx1), _mm256_set1_ps(src.elempack));
-                v2_offset_f[i] = _mm256_mul_ps(_mm256_add_ps(gy_offset, gx2), _mm256_set1_ps(src.elempack));
-                v3_offset_f[i] = _mm256_mul_ps(_mm256_add_ps(gy_offset, gx3), _mm256_set1_ps(src.elempack));
-
-                v0_offset_f[i] = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), v0_offset_f[i], _mm256_and_ps(x0_in_range, y_in_range));
-                v1_offset_f[i] = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), v1_offset_f[i], _mm256_and_ps(x1_in_range, y_in_range));
-                v2_offset_f[i] = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), v2_offset_f[i], _mm256_and_ps(x2_in_range, y_in_range));
-                v3_offset_f[i] = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), v3_offset_f[i], _mm256_and_ps(x3_in_range, y_in_range));
-
-                v0_offset_f[i] = _mm256_castsi256_ps(_mm256_cvtps_epi32(v0_offset_f[i]));
-                v1_offset_f[i] = _mm256_castsi256_ps(_mm256_cvtps_epi32(v1_offset_f[i]));
-                v2_offset_f[i] = _mm256_castsi256_ps(_mm256_cvtps_epi32(v2_offset_f[i]));
-                v3_offset_f[i] = _mm256_castsi256_ps(_mm256_cvtps_epi32(v3_offset_f[i]));
-            }
-
-            transpose8x18_ps(tx, ty, v0_offset_f[0], v1_offset_f[0], v2_offset_f[0], v3_offset_f[0], v0_offset_f[1], v1_offset_f[1], v2_offset_f[1], v3_offset_f[1], v0_offset_f[2], v1_offset_f[2], v2_offset_f[2], v3_offset_f[2], v0_offset_f[3], v1_offset_f[3], v2_offset_f[3], v3_offset_f[3]);
-
-            _mm256_storeu_ps(offset_value_ptr, tx);
-            _mm256_storeu_ps(offset_value_ptr + 8, ty);
-            offset_value_ptr += 16;
-            for (int i = 0; i < 4; i++)
-            {
-                _mm256_storeu_ps(offset_value_ptr, v0_offset_f[i]);
-                _mm256_storeu_ps(offset_value_ptr + 8, v1_offset_f[i]);
-                _mm256_storeu_ps(offset_value_ptr + 16, v2_offset_f[i]);
-                _mm256_storeu_ps(offset_value_ptr + 24, v3_offset_f[i]);
-                offset_value_ptr += 32;
-            }
-
-            gridptr_x += 8;
-            gridptr_y += 8;
-        }
-
-#endif // __AVX__
-
-        for (; x < grid_size; x++)
-        {
-            float sample_x = *gridptr_x;
-            float sample_y = *gridptr_y;
-
-            sample_x = unormalize(src.w, sample_x);
-            sample_y = unormalize(src.h, sample_y);
-
-            int x1 = floorf(sample_x);
-            int y1 = floorf(sample_y);
-            int x0 = x1 - 1;
-            int x2 = x1 + 1;
-            int x3 = x1 + 2;
-
-            offset_value_ptr[0] = sample_x - static_cast<float>(x1);
-            offset_value_ptr[1] = sample_y - static_cast<float>(y1);
-
-            x1 = get_coord(src.w, x1);
-            x0 = get_coord(src.w, x0);
-            x2 = get_coord(src.w, x2);
-            x3 = get_coord(src.w, x3);
-
-            bool x1_in_range = (x1 > -1) & (x1 < src.w);
-            bool x0_in_range = (x0 > -1) & (x0 < src.w);
-            bool x2_in_range = (x2 > -1) & (x2 < src.w);
-            bool x3_in_range = (x3 > -1) & (x3 < src.w);
-
-            int* offset_ptr = (int*)offset_value_ptr + 2;
-
-            for (int i = 0; i < 4; i++)
-            {
-                int gy = y1 + i - 1;
-                gy = get_coord(src.h, gy);
-                int offset_y = gy * src.w;
-
-                bool y_in_range = (gy > -1) & (gy < src.h);
-
-                bool v0_in_bound = (x0_in_range & y_in_range);
-                bool v1_in_bound = (x1_in_range & y_in_range);
-                bool v2_in_bound = (x2_in_range & y_in_range);
-                bool v3_in_bound = (x3_in_range & y_in_range);
-
-                offset_ptr[0] = v0_in_bound ? (offset_y + x0) * src.elempack : -1.0;
-                offset_ptr[1] = v1_in_bound ? (offset_y + x1) * src.elempack : -1.0;
-                offset_ptr[2] = v2_in_bound ? (offset_y + x2) * src.elempack : -1.0;
-                offset_ptr[3] = v3_in_bound ? (offset_y + x3) * src.elempack : -1.0;
-
-                offset_ptr += 4;
-            }
-
-            gridptr_x++;
-            gridptr_y++;
-
-            offset_value_ptr += 18;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bilinear_apply_interpolation.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bilinear_apply_interpolation.h
deleted file mode 100644
index 56f2767e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bilinear_apply_interpolation.h
+++ /dev/null
@@ -1,370 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-static void gridsample_2d_bilinear_apply_interpolation_p16(const Mat& src, Mat& dst, const Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int grid_size = outw * outh;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const float* offset_value_ptr = offset_value.channel(0);
-
-        for (int i = 0; i < grid_size; i++)
-        {
-            const int* offset_ptr = (int*)offset_value_ptr;
-            const float* value_ptr = offset_value_ptr + 4;
-
-            __m512 v00_val = offset_ptr[0] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[0]) : _mm512_set1_ps(0);
-            __m512 v01_val = offset_ptr[1] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[1]) : _mm512_set1_ps(0);
-            __m512 v10_val = offset_ptr[2] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[2]) : _mm512_set1_ps(0);
-            __m512 v11_val = offset_ptr[3] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[3]) : _mm512_set1_ps(0);
-
-            __m512 value1 = _mm512_set1_ps(value_ptr[0]);
-            __m512 v0 = _mm512_fmadd_ps(v01_val, value1, _mm512_fnmadd_ps(v00_val, value1, v00_val));
-            __m512 v1 = _mm512_fmadd_ps(v11_val, value1, _mm512_fnmadd_ps(v10_val, value1, v10_val));
-
-            __m512 value2 = _mm512_set1_ps(value_ptr[1]);
-            __m512 _v = _mm512_fmadd_ps(v1, value2, _mm512_fnmadd_ps(v0, value2, v0));
-            _mm512_storeu_ps(dstptr, _v);
-
-            dstptr += 16;
-            offset_value_ptr += 6;
-        }
-    }
-}
-
-static void gridsample_3d_bilinear_apply_interpolation_p16(const Mat& src, Mat& dst, const Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int outd = dst.d;
-    const int grid_size = outw * outh * outd;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const float* offset_value_ptr = offset_value.channel(0);
-
-        for (int i = 0; i < grid_size; i++)
-        {
-            const int* offset_ptr = (int*)offset_value_ptr;
-            const float* value_ptr = offset_value_ptr + 8;
-
-            __m512 v000_val = offset_ptr[0] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[0]) : _mm512_set1_ps(0);
-            __m512 v001_val = offset_ptr[1] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[1]) : _mm512_set1_ps(0);
-            __m512 v010_val = offset_ptr[2] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[2]) : _mm512_set1_ps(0);
-            __m512 v011_val = offset_ptr[3] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[3]) : _mm512_set1_ps(0);
-
-            __m512 v100_val = offset_ptr[4] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[4]) : _mm512_set1_ps(0);
-            __m512 v101_val = offset_ptr[5] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[5]) : _mm512_set1_ps(0);
-            __m512 v110_val = offset_ptr[6] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[6]) : _mm512_set1_ps(0);
-            __m512 v111_val = offset_ptr[7] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[7]) : _mm512_set1_ps(0);
-
-            __m512 value = _mm512_set1_ps(value_ptr[0]);
-            __m512 v00 = _mm512_fmadd_ps(v001_val, value, _mm512_fnmadd_ps(v000_val, value, v000_val));
-            __m512 v01 = _mm512_fmadd_ps(v011_val, value, _mm512_fnmadd_ps(v010_val, value, v010_val));
-            __m512 v10 = _mm512_fmadd_ps(v101_val, value, _mm512_fnmadd_ps(v100_val, value, v100_val));
-            __m512 v11 = _mm512_fmadd_ps(v111_val, value, _mm512_fnmadd_ps(v110_val, value, v110_val));
-
-            value = _mm512_set1_ps(value_ptr[1]);
-            __m512 v0 = _mm512_fmadd_ps(v01, value, _mm512_fnmadd_ps(v00, value, v00));
-            __m512 v1 = _mm512_fmadd_ps(v11, value, _mm512_fnmadd_ps(v10, value, v10));
-
-            value = _mm512_set1_ps(value_ptr[2]);
-            __m512 _v = _mm512_fmadd_ps(v1, value, _mm512_fnmadd_ps(v0, value, v0));
-            _mm512_storeu_ps(dstptr, _v);
-
-            dstptr += 16;
-            offset_value_ptr += 11;
-        }
-    }
-}
-
-#endif // __AVX512F__
-
-static void gridsample_2d_bilinear_apply_interpolation_p8(const Mat& src, Mat& dst, const Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int grid_size = outw * outh;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const float* offset_value_ptr = offset_value.channel(0);
-
-        for (int i = 0; i < grid_size; i++)
-        {
-            const int* offset_ptr = (int*)offset_value_ptr;
-            const float* value_ptr = offset_value_ptr + 4;
-
-            __m256 v00_val = offset_ptr[0] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[0]) : _mm256_set1_ps(0);
-            __m256 v01_val = offset_ptr[1] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[1]) : _mm256_set1_ps(0);
-            __m256 v10_val = offset_ptr[2] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[2]) : _mm256_set1_ps(0);
-            __m256 v11_val = offset_ptr[3] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[3]) : _mm256_set1_ps(0);
-
-            __m256 value1 = _mm256_set1_ps(value_ptr[0]);
-            __m256 v0 = _mm256_comp_fmadd_ps(v01_val, value1, _mm256_comp_fnmadd_ps(v00_val, value1, v00_val));
-            __m256 v1 = _mm256_comp_fmadd_ps(v11_val, value1, _mm256_comp_fnmadd_ps(v10_val, value1, v10_val));
-
-            __m256 value2 = _mm256_set1_ps(value_ptr[1]);
-            __m256 _v = _mm256_comp_fmadd_ps(v1, value2, _mm256_comp_fnmadd_ps(v0, value2, v0));
-            _mm256_storeu_ps(dstptr, _v);
-
-            dstptr += 8;
-            offset_value_ptr += 6;
-        }
-    }
-}
-static void gridsample_3d_bilinear_apply_interpolation_p8(const Mat& src, Mat& dst, const Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int outd = dst.d;
-    const int grid_size = outw * outh * outd;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const float* offset_value_ptr = offset_value.channel(0);
-
-        for (int i = 0; i < grid_size; i++)
-        {
-            const int* offset_ptr = (int*)offset_value_ptr;
-            const float* value_ptr = offset_value_ptr + 8;
-
-            __m256 v000_val = offset_ptr[0] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[0]) : _mm256_set1_ps(0);
-            __m256 v001_val = offset_ptr[1] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[1]) : _mm256_set1_ps(0);
-            __m256 v010_val = offset_ptr[2] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[2]) : _mm256_set1_ps(0);
-            __m256 v011_val = offset_ptr[3] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[3]) : _mm256_set1_ps(0);
-
-            __m256 v100_val = offset_ptr[4] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[4]) : _mm256_set1_ps(0);
-            __m256 v101_val = offset_ptr[5] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[5]) : _mm256_set1_ps(0);
-            __m256 v110_val = offset_ptr[6] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[6]) : _mm256_set1_ps(0);
-            __m256 v111_val = offset_ptr[7] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[7]) : _mm256_set1_ps(0);
-
-            __m256 value = _mm256_set1_ps(value_ptr[0]);
-            __m256 v00 = _mm256_comp_fmadd_ps(v001_val, value, _mm256_comp_fnmadd_ps(v000_val, value, v000_val));
-            __m256 v01 = _mm256_comp_fmadd_ps(v011_val, value, _mm256_comp_fnmadd_ps(v010_val, value, v010_val));
-            __m256 v10 = _mm256_comp_fmadd_ps(v101_val, value, _mm256_comp_fnmadd_ps(v100_val, value, v100_val));
-            __m256 v11 = _mm256_comp_fmadd_ps(v111_val, value, _mm256_comp_fnmadd_ps(v110_val, value, v110_val));
-
-            value = _mm256_set1_ps(value_ptr[1]);
-            __m256 v0 = _mm256_comp_fmadd_ps(v01, value, _mm256_comp_fnmadd_ps(v00, value, v00));
-            __m256 v1 = _mm256_comp_fmadd_ps(v11, value, _mm256_comp_fnmadd_ps(v10, value, v10));
-
-            value = _mm256_set1_ps(value_ptr[2]);
-            __m256 _v = _mm256_comp_fmadd_ps(v1, value, _mm256_comp_fnmadd_ps(v0, value, v0));
-            _mm256_storeu_ps(dstptr, _v);
-
-            dstptr += 8;
-            offset_value_ptr += 11;
-        }
-    }
-}
-#endif // __AVX__
-static void gridsample_2d_bilinear_apply_interpolation_p4(const Mat& src, Mat& dst, const Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int grid_size = outw * outh;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const float* offset_value_ptr = offset_value.channel(0);
-
-        for (int i = 0; i < grid_size; i++)
-        {
-            const int* offset_ptr = (int*)offset_value_ptr;
-            const float* value_ptr = offset_value_ptr + 4;
-
-            __m128 v00_val = offset_ptr[0] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[0]) : _mm_set1_ps(0);
-            __m128 v01_val = offset_ptr[1] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[1]) : _mm_set1_ps(0);
-            __m128 v10_val = offset_ptr[2] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[2]) : _mm_set1_ps(0);
-            __m128 v11_val = offset_ptr[3] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[3]) : _mm_set1_ps(0);
-
-            __m128 value1 = _mm_set1_ps(value_ptr[0]);
-            __m128 v0 = _mm_comp_fmadd_ps(v01_val, value1, _mm_comp_fnmadd_ps(v00_val, value1, v00_val));
-            __m128 v1 = _mm_comp_fmadd_ps(v11_val, value1, _mm_comp_fnmadd_ps(v10_val, value1, v10_val));
-
-            __m128 value2 = _mm_set1_ps(value_ptr[1]);
-            __m128 _v = _mm_comp_fmadd_ps(v1, value2, _mm_comp_fnmadd_ps(v0, value2, v0));
-            _mm_storeu_ps(dstptr, _v);
-
-            dstptr += 4;
-            offset_value_ptr += 6;
-        }
-    }
-}
-static void gridsample_3d_bilinear_apply_interpolation_p4(const Mat& src, Mat& dst, const Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int outd = dst.d;
-    const int grid_size = outw * outh * outd;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const float* offset_value_ptr = offset_value.channel(0);
-
-        for (int i = 0; i < grid_size; i++)
-        {
-            const int* offset_ptr = (int*)offset_value_ptr;
-            const float* value_ptr = offset_value_ptr + 8;
-
-            __m128 v000_val = offset_ptr[0] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[0]) : _mm_set1_ps(0);
-            __m128 v001_val = offset_ptr[1] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[1]) : _mm_set1_ps(0);
-            __m128 v010_val = offset_ptr[2] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[2]) : _mm_set1_ps(0);
-            __m128 v011_val = offset_ptr[3] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[3]) : _mm_set1_ps(0);
-
-            __m128 v100_val = offset_ptr[4] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[4]) : _mm_set1_ps(0);
-            __m128 v101_val = offset_ptr[5] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[5]) : _mm_set1_ps(0);
-            __m128 v110_val = offset_ptr[6] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[6]) : _mm_set1_ps(0);
-            __m128 v111_val = offset_ptr[7] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[7]) : _mm_set1_ps(0);
-
-            __m128 value = _mm_set1_ps(value_ptr[0]);
-            __m128 v00 = _mm_comp_fmadd_ps(v001_val, value, _mm_comp_fnmadd_ps(v000_val, value, v000_val));
-            __m128 v01 = _mm_comp_fmadd_ps(v011_val, value, _mm_comp_fnmadd_ps(v010_val, value, v010_val));
-            __m128 v10 = _mm_comp_fmadd_ps(v101_val, value, _mm_comp_fnmadd_ps(v100_val, value, v100_val));
-            __m128 v11 = _mm_comp_fmadd_ps(v111_val, value, _mm_comp_fnmadd_ps(v110_val, value, v110_val));
-
-            value = _mm_set1_ps(value_ptr[1]);
-            __m128 v0 = _mm_comp_fmadd_ps(v01, value, _mm_comp_fnmadd_ps(v00, value, v00));
-            __m128 v1 = _mm_comp_fmadd_ps(v11, value, _mm_comp_fnmadd_ps(v10, value, v10));
-
-            value = _mm_set1_ps(value_ptr[2]);
-            __m128 _v = _mm_comp_fmadd_ps(v1, value, _mm_comp_fnmadd_ps(v0, value, v0));
-            _mm_storeu_ps(dstptr, _v);
-
-            dstptr += 4;
-            offset_value_ptr += 11;
-        }
-    }
-}
-#endif // __SSE2__
-
-static void gridsample_2d_bilinear_apply_interpolation_p1(const Mat& src, Mat& dst, const Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int grid_size = outw * outh;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const float* offset_value_ptr = offset_value.channel(0);
-
-        for (int x = 0; x < grid_size; x++)
-        {
-            const int* offset_ptr = (int*)offset_value_ptr;
-            const float* value_ptr = offset_value_ptr + 4;
-
-            float v00 = offset_ptr[0] >= 0 ? *(srcptr + offset_ptr[0]) : 0;
-            float v01 = offset_ptr[1] >= 0 ? *(srcptr + offset_ptr[1]) : 0;
-            float v10 = offset_ptr[2] >= 0 ? *(srcptr + offset_ptr[2]) : 0;
-            float v11 = offset_ptr[3] >= 0 ? *(srcptr + offset_ptr[3]) : 0;
-
-            float v0 = v00 * (1 - value_ptr[0]) + v01 * value_ptr[0];
-            float v1 = v10 * (1 - value_ptr[0]) + v11 * value_ptr[0];
-
-            *dstptr = v0 * (1 - value_ptr[1]) + v1 * value_ptr[1];
-
-            dstptr++;
-            offset_value_ptr += 6;
-        }
-    }
-}
-
-static void gridsample_3d_bilinear_apply_interpolation_p1(const Mat& src, Mat& dst, const Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int outd = dst.d;
-    const int grid_size = outw * outh * outd;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const float* offset_value_ptr = offset_value.channel(0);
-
-        for (int x = 0; x < grid_size; x++)
-        {
-            const int* offset_ptr = (int*)offset_value_ptr;
-            const float* value_ptr = offset_value_ptr + 8;
-
-            float v000 = offset_ptr[0] >= 0 ? *(srcptr + offset_ptr[0]) : 0;
-            float v001 = offset_ptr[1] >= 0 ? *(srcptr + offset_ptr[1]) : 0;
-            float v010 = offset_ptr[2] >= 0 ? *(srcptr + offset_ptr[2]) : 0;
-            float v011 = offset_ptr[3] >= 0 ? *(srcptr + offset_ptr[3]) : 0;
-
-            float v100 = offset_ptr[4] >= 0 ? *(srcptr + offset_ptr[4]) : 0;
-            float v101 = offset_ptr[5] >= 0 ? *(srcptr + offset_ptr[5]) : 0;
-            float v110 = offset_ptr[6] >= 0 ? *(srcptr + offset_ptr[6]) : 0;
-            float v111 = offset_ptr[7] >= 0 ? *(srcptr + offset_ptr[7]) : 0;
-
-            float v00 = v000 * (1 - value_ptr[0]) + v001 * value_ptr[0];
-            float v01 = v010 * (1 - value_ptr[0]) + v011 * value_ptr[0];
-            float v10 = v100 * (1 - value_ptr[0]) + v101 * value_ptr[0];
-            float v11 = v110 * (1 - value_ptr[0]) + v111 * value_ptr[0];
-
-            float v0 = v00 * (1 - value_ptr[1]) + v01 * value_ptr[1];
-            float v1 = v10 * (1 - value_ptr[1]) + v11 * value_ptr[1];
-
-            *dstptr = v0 * (1 - value_ptr[2]) + v1 * value_ptr[2];
-
-            dstptr++;
-            offset_value_ptr += 11;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bilinear_compute_blob.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bilinear_compute_blob.h
deleted file mode 100644
index f78e017c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_bilinear_compute_blob.h
+++ /dev/null
@@ -1,623 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-template<GridSample::PaddingMode pd, bool align_corner>
-void gridsample_2d_bilinear_compute_blob(const Mat& src, const Mat& grid, Mat& offset_value, int permute_fusion)
-{
-    const int grid_size = grid.w * grid.h;
-
-    float* offset_value_ptr = offset_value.channel(0);
-
-    grid_sample_unormalize<align_corner> unormalize;
-    compute_coord<pd, align_corner> get_coord;
-
-    if (permute_fusion == 0)
-    {
-        for (int y = 0; y < grid.c; y++)
-        {
-            const float* gridptr = grid.channel(y);
-            int x = 0;
-#if __AVX__
-            for (; x + 15 < grid_size; x += 16)
-            {
-                __m256 gx = _mm256_loadu_ps(gridptr);
-                __m256 gy = _mm256_loadu_ps(gridptr + 8);
-
-                transpose2x8_ps(gx, gy);
-
-                gx = unormalize(_mm256_set1_ps(src.w), gx);
-                gx = get_coord(_mm256_set1_ps(src.w), gx);
-
-                gy = unormalize(_mm256_set1_ps(src.h), gy);
-                gy = get_coord(_mm256_set1_ps(src.h), gy);
-
-                __m256 x_w = _mm256_floor_ps(gx);
-                __m256 y_n = _mm256_floor_ps(gy);
-
-                __m256 x1 = _mm256_add_ps(x_w, _mm256_set1_ps(1));
-                __m256 y1 = _mm256_add_ps(y_n, _mm256_set1_ps(1));
-
-                __m256 x0_in_range = _mm256_and_ps(_mm256_cmp_ps(x_w, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), x_w, _CMP_GT_OS));
-                __m256 x1_in_range = _mm256_and_ps(_mm256_cmp_ps(x1, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), x1, _CMP_GT_OS));
-                __m256 y0_in_range = _mm256_and_ps(_mm256_cmp_ps(y_n, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.h), y_n, _CMP_GT_OS));
-                __m256 y1_in_range = _mm256_and_ps(_mm256_cmp_ps(y1, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.h), y1, _CMP_GT_OS));
-
-                __m256 v00_in_range = _mm256_and_ps(x0_in_range, y0_in_range);
-                __m256 v01_in_range = _mm256_and_ps(x1_in_range, y0_in_range);
-                __m256 v10_in_range = _mm256_and_ps(x0_in_range, y1_in_range);
-                __m256 v11_in_range = _mm256_and_ps(x1_in_range, y1_in_range);
-
-                __m256 nw_offset = _mm256_mul_ps(_mm256_comp_fmadd_ps(y_n, _mm256_set1_ps(src.w), x_w), _mm256_set1_ps(src.elempack));
-                __m256 ne_offset = _mm256_add_ps(nw_offset, _mm256_set1_ps(src.elempack));
-                __m256 sw_offset = _mm256_comp_fmadd_ps(_mm256_set1_ps(src.w), _mm256_set1_ps(src.elempack), nw_offset);
-                __m256 se_offset = _mm256_add_ps(sw_offset, _mm256_set1_ps(src.elempack));
-
-                nw_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), nw_offset, v00_in_range);
-                ne_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), ne_offset, v01_in_range);
-                sw_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), sw_offset, v10_in_range);
-                se_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), se_offset, v11_in_range);
-
-                nw_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(nw_offset));
-                ne_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(ne_offset));
-                sw_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(sw_offset));
-                se_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(se_offset));
-
-                __m256 alpha = _mm256_sub_ps(gx, x_w);
-                __m256 beta = _mm256_sub_ps(gy, y_n);
-
-                transpose8x6_ps(nw_offset, ne_offset, sw_offset, se_offset, alpha, beta);
-
-                _mm256_storeu_ps(offset_value_ptr, nw_offset);
-                _mm256_storeu_ps(offset_value_ptr + 8, ne_offset);
-                _mm256_storeu_ps(offset_value_ptr + 16, sw_offset);
-                _mm256_storeu_ps(offset_value_ptr + 24, se_offset);
-
-                _mm256_storeu_ps(offset_value_ptr + 32, alpha);
-                _mm256_storeu_ps(offset_value_ptr + 40, beta);
-
-                gridptr += 16;
-                offset_value_ptr += 48;
-            }
-#endif // __AVX__
-
-            for (; x < grid_size; x += 2)
-            {
-                float sample_x = *gridptr;
-                float sample_y = *(gridptr + 1);
-
-                sample_x = unormalize(src.w, sample_x);
-                sample_x = get_coord(src.w, sample_x);
-
-                sample_y = unormalize(src.h, sample_y);
-                sample_y = get_coord(src.h, sample_y);
-
-                int x0 = (int)floorf(sample_x);
-                int y0 = (int)floorf(sample_y);
-                int x1 = x0 + 1;
-                int y1 = y0 + 1;
-
-                bool x0_in_bound = (x0 > -1) & (x0 < src.w);
-                bool x1_in_bound = (x1 > -1) & (x1 < src.w);
-                bool y0_in_bound = (y0 > -1) & (y0 < src.h);
-                bool y1_in_bound = (y1 > -1) & (y1 < src.h);
-
-                bool in_bound_00 = x0_in_bound & y0_in_bound;
-                bool in_bound_01 = x1_in_bound & y0_in_bound;
-                bool in_bound_10 = x0_in_bound & y1_in_bound;
-                bool in_bound_11 = x1_in_bound & y1_in_bound;
-
-                int* offset_ptr = (int*)offset_value_ptr;
-                float* value_ptr = offset_value_ptr + 4;
-
-                offset_ptr[0] = in_bound_00 ? (x0 + y0 * src.w) * src.elempack : -1.0;
-                offset_ptr[1] = in_bound_01 ? (x1 + y0 * src.w) * src.elempack : -1.0;
-                offset_ptr[2] = in_bound_10 ? (x0 + y1 * src.w) * src.elempack : -1.0;
-                offset_ptr[3] = in_bound_11 ? (x1 + y1 * src.w) * src.elempack : -1.0;
-
-                value_ptr[0] = sample_x - x0;
-                value_ptr[1] = sample_y - y0;
-
-                gridptr += 2;
-                offset_value_ptr += 6;
-            }
-        }
-    }
-    else
-    {
-        const float* gridptr_x = grid.channel(0);
-        const float* gridptr_y = grid.channel(1);
-
-        int x = 0;
-#if __AVX__
-        for (; x + 7 < grid_size; x += 8)
-        {
-            __m256 gx = _mm256_loadu_ps(gridptr_x);
-            __m256 gy = _mm256_loadu_ps(gridptr_y);
-
-            gx = unormalize(_mm256_set1_ps(src.w), gx);
-            gx = get_coord(_mm256_set1_ps(src.w), gx);
-
-            gy = unormalize(_mm256_set1_ps(src.h), gy);
-            gy = get_coord(_mm256_set1_ps(src.h), gy);
-
-            __m256 x_w = _mm256_floor_ps(gx);
-            __m256 y_n = _mm256_floor_ps(gy);
-
-            __m256 x1 = _mm256_add_ps(x_w, _mm256_set1_ps(1));
-            __m256 y1 = _mm256_add_ps(y_n, _mm256_set1_ps(1));
-
-            __m256 x0_in_range = _mm256_and_ps(_mm256_cmp_ps(x_w, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), x_w, _CMP_GT_OS));
-            __m256 x1_in_range = _mm256_and_ps(_mm256_cmp_ps(x1, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), x1, _CMP_GT_OS));
-            __m256 y0_in_range = _mm256_and_ps(_mm256_cmp_ps(y_n, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.h), y_n, _CMP_GT_OS));
-            __m256 y1_in_range = _mm256_and_ps(_mm256_cmp_ps(y1, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.h), y1, _CMP_GT_OS));
-
-            __m256 v00_in_range = _mm256_and_ps(x0_in_range, y0_in_range);
-            __m256 v01_in_range = _mm256_and_ps(x1_in_range, y0_in_range);
-            __m256 v10_in_range = _mm256_and_ps(x0_in_range, y1_in_range);
-            __m256 v11_in_range = _mm256_and_ps(x1_in_range, y1_in_range);
-
-            __m256 nw_offset = _mm256_mul_ps(_mm256_comp_fmadd_ps(y_n, _mm256_set1_ps(src.w), x_w), _mm256_set1_ps(src.elempack));
-            __m256 ne_offset = _mm256_add_ps(nw_offset, _mm256_set1_ps(src.elempack));
-            __m256 sw_offset = _mm256_comp_fmadd_ps(_mm256_set1_ps(src.w), _mm256_set1_ps(src.elempack), nw_offset);
-            __m256 se_offset = _mm256_add_ps(sw_offset, _mm256_set1_ps(src.elempack));
-
-            nw_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), nw_offset, v00_in_range);
-            ne_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), ne_offset, v01_in_range);
-            sw_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), sw_offset, v10_in_range);
-            se_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), se_offset, v11_in_range);
-
-            nw_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(nw_offset));
-            ne_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(ne_offset));
-            sw_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(sw_offset));
-            se_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(se_offset));
-
-            __m256 alpha = _mm256_sub_ps(gx, x_w);
-            __m256 beta = _mm256_sub_ps(gy, y_n);
-
-            transpose8x6_ps(nw_offset, ne_offset, sw_offset, se_offset, alpha, beta);
-
-            _mm256_storeu_ps(offset_value_ptr, nw_offset);
-            _mm256_storeu_ps(offset_value_ptr + 8, ne_offset);
-            _mm256_storeu_ps(offset_value_ptr + 16, sw_offset);
-            _mm256_storeu_ps(offset_value_ptr + 24, se_offset);
-
-            _mm256_storeu_ps(offset_value_ptr + 32, alpha);
-            _mm256_storeu_ps(offset_value_ptr + 40, beta);
-
-            gridptr_x += 8;
-            gridptr_y += 8;
-            offset_value_ptr += 48;
-        }
-
-#endif // __AVX__
-
-        for (; x < grid_size; x++)
-        {
-            float sample_x = *gridptr_x;
-            float sample_y = *gridptr_y;
-
-            sample_x = unormalize(src.w, sample_x);
-            sample_x = get_coord(src.w, sample_x);
-
-            sample_y = unormalize(src.h, sample_y);
-            sample_y = get_coord(src.h, sample_y);
-
-            int x0 = (int)floorf(sample_x);
-            int y0 = (int)floorf(sample_y);
-            int x1 = x0 + 1;
-            int y1 = y0 + 1;
-
-            bool x0_in_bound = (x0 > -1) & (x0 < src.w);
-            bool x1_in_bound = (x1 > -1) & (x1 < src.w);
-            bool y0_in_bound = (y0 > -1) & (y0 < src.h);
-            bool y1_in_bound = (y1 > -1) & (y1 < src.h);
-
-            bool in_bound_00 = x0_in_bound & y0_in_bound;
-            bool in_bound_01 = x1_in_bound & y0_in_bound;
-            bool in_bound_10 = x0_in_bound & y1_in_bound;
-            bool in_bound_11 = x1_in_bound & y1_in_bound;
-
-            int* offset_ptr = (int*)offset_value_ptr;
-            float* value_ptr = offset_value_ptr + 4;
-
-            offset_ptr[0] = in_bound_00 ? (x0 + y0 * src.w) * src.elempack : -1.0;
-            offset_ptr[1] = in_bound_01 ? (x1 + y0 * src.w) * src.elempack : -1.0;
-            offset_ptr[2] = in_bound_10 ? (x0 + y1 * src.w) * src.elempack : -1.0;
-            offset_ptr[3] = in_bound_11 ? (x1 + y1 * src.w) * src.elempack : -1.0;
-
-            value_ptr[0] = sample_x - x0;
-            value_ptr[1] = sample_y - y0;
-
-            gridptr_x++;
-            gridptr_y++;
-            offset_value_ptr += 6;
-        }
-    }
-}
-
-template<GridSample::PaddingMode pd, bool align_corner>
-void gridsample_3d_bilinear_compute_blob(const Mat& src, const Mat& grid, Mat& offset_value, int permute_fusion)
-{
-    const int grid_size = grid.w * grid.h * grid.d;
-
-    float* offset_value_ptr = offset_value.channel(0);
-
-    grid_sample_unormalize<align_corner> unormalize;
-    compute_coord<pd, align_corner> get_coord;
-
-    if (permute_fusion == 0)
-    {
-        for (int y = 0; y < grid.c; y++)
-        {
-            const float* gridptr = grid.channel(y);
-            int x = 0;
-#if __AVX__
-            for (; x + 23 < grid_size; x += 24)
-            {
-                __m256 gx = _mm256_loadu_ps(gridptr);
-                __m256 gy = _mm256_loadu_ps(gridptr + 8);
-                __m256 gz = _mm256_loadu_ps(gridptr + 16);
-
-                transpose3x8_ps(gx, gy, gz);
-
-                gx = unormalize(_mm256_set1_ps(src.w), gx);
-                gx = get_coord(_mm256_set1_ps(src.w), gx);
-
-                gy = unormalize(_mm256_set1_ps(src.h), gy);
-                gy = get_coord(_mm256_set1_ps(src.h), gy);
-
-                gz = unormalize(_mm256_set1_ps(src.d), gz);
-                gz = get_coord(_mm256_set1_ps(src.d), gz);
-
-                __m256 x_w = _mm256_floor_ps(gx);
-                __m256 y_n = _mm256_floor_ps(gy);
-                __m256 z_t = _mm256_floor_ps(gz);
-
-                __m256 x1 = _mm256_add_ps(x_w, _mm256_set1_ps(1));
-                __m256 y1 = _mm256_add_ps(y_n, _mm256_set1_ps(1));
-                __m256 z1 = _mm256_add_ps(z_t, _mm256_set1_ps(1));
-
-                __m256 x0_in_range = _mm256_and_ps(_mm256_cmp_ps(x_w, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), x_w, _CMP_GT_OS));
-                __m256 x1_in_range = _mm256_and_ps(_mm256_cmp_ps(x1, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), x1, _CMP_GT_OS));
-                __m256 y0_in_range = _mm256_and_ps(_mm256_cmp_ps(y_n, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.h), y_n, _CMP_GT_OS));
-                __m256 y1_in_range = _mm256_and_ps(_mm256_cmp_ps(y1, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.h), y1, _CMP_GT_OS));
-                __m256 z0_in_range = _mm256_and_ps(_mm256_cmp_ps(z_t, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.d), z_t, _CMP_GT_OS));
-                __m256 z1_in_range = _mm256_and_ps(_mm256_cmp_ps(z1, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.d), z1, _CMP_GT_OS));
-
-                __m256 v000_in_range, v010_in_range, v100_in_range, v110_in_range, v001_in_range, v011_in_range, v101_in_range, v111_in_range;
-                {
-                    __m256 v00_in_range = _mm256_and_ps(x0_in_range, y0_in_range);
-                    __m256 v01_in_range = _mm256_and_ps(x1_in_range, y0_in_range);
-                    __m256 v10_in_range = _mm256_and_ps(x0_in_range, y1_in_range);
-                    __m256 v11_in_range = _mm256_and_ps(x1_in_range, y1_in_range);
-
-                    v000_in_range = _mm256_and_ps(v00_in_range, z0_in_range);
-                    v001_in_range = _mm256_and_ps(v01_in_range, z0_in_range);
-                    v010_in_range = _mm256_and_ps(v10_in_range, z0_in_range);
-                    v011_in_range = _mm256_and_ps(v11_in_range, z0_in_range);
-
-                    v100_in_range = _mm256_and_ps(v00_in_range, z1_in_range);
-                    v101_in_range = _mm256_and_ps(v01_in_range, z1_in_range);
-                    v110_in_range = _mm256_and_ps(v10_in_range, z1_in_range);
-                    v111_in_range = _mm256_and_ps(v11_in_range, z1_in_range);
-                }
-
-                __m256 tnw_offset = _mm256_mul_ps(_mm256_comp_fmadd_ps(_mm256_mul_ps(_mm256_set1_ps(src.w), _mm256_set1_ps(src.h)), z_t,
-                                                  _mm256_comp_fmadd_ps(y_n, _mm256_set1_ps(src.w), x_w)),
-                                                  _mm256_set1_ps(src.elempack));
-                __m256 tne_offset = _mm256_add_ps(tnw_offset, _mm256_set1_ps(src.elempack));
-                __m256 tsw_offset = _mm256_add_ps(tnw_offset, _mm256_mul_ps(_mm256_set1_ps(src.w), _mm256_set1_ps(src.elempack)));
-                __m256 tse_offset = _mm256_add_ps(tsw_offset, _mm256_set1_ps(src.elempack));
-
-                __m256 bnw_offset = _mm256_comp_fmadd_ps(_mm256_mul_ps(_mm256_set1_ps(src.w), _mm256_set1_ps(src.h)), _mm256_set1_ps(src.elempack), tnw_offset);
-                __m256 bne_offset = _mm256_add_ps(bnw_offset, _mm256_set1_ps(src.elempack));
-                __m256 bsw_offset = _mm256_add_ps(bnw_offset, _mm256_mul_ps(_mm256_set1_ps(src.w), _mm256_set1_ps(src.elempack)));
-                __m256 bse_offset = _mm256_add_ps(bsw_offset, _mm256_set1_ps(src.elempack));
-
-                tnw_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), tnw_offset, v000_in_range);
-                tne_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), tne_offset, v001_in_range);
-                tsw_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), tsw_offset, v010_in_range);
-                tse_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), tse_offset, v011_in_range);
-
-                bnw_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), bnw_offset, v100_in_range);
-                bne_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), bne_offset, v101_in_range);
-                bsw_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), bsw_offset, v110_in_range);
-                bse_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), bse_offset, v111_in_range);
-
-                tnw_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(tnw_offset));
-                tne_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(tne_offset));
-                tsw_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(tsw_offset));
-                tse_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(tse_offset));
-
-                bnw_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(bnw_offset));
-                bne_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(bne_offset));
-                bsw_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(bsw_offset));
-                bse_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(bse_offset));
-
-                __m256 alpha = _mm256_sub_ps(gx, x_w);
-                __m256 beta = _mm256_sub_ps(gy, y_n);
-                __m256 gamma = _mm256_sub_ps(gz, z_t);
-
-                transpose8x11_ps(tnw_offset, tne_offset, tsw_offset, tse_offset, bnw_offset, bne_offset, bsw_offset, bse_offset, alpha, beta, gamma);
-
-                _mm256_storeu_ps(offset_value_ptr, tnw_offset);
-                _mm256_storeu_ps(offset_value_ptr + 8, tne_offset);
-                _mm256_storeu_ps(offset_value_ptr + 16, tsw_offset);
-                _mm256_storeu_ps(offset_value_ptr + 24, tse_offset);
-
-                _mm256_storeu_ps(offset_value_ptr + 32, bnw_offset);
-                _mm256_storeu_ps(offset_value_ptr + 40, bne_offset);
-                _mm256_storeu_ps(offset_value_ptr + 48, bsw_offset);
-                _mm256_storeu_ps(offset_value_ptr + 56, bse_offset);
-
-                _mm256_storeu_ps(offset_value_ptr + 64, alpha);
-                _mm256_storeu_ps(offset_value_ptr + 72, beta);
-                _mm256_storeu_ps(offset_value_ptr + 80, gamma);
-
-                gridptr += 24;
-
-                offset_value_ptr += 88;
-            }
-#endif // __AVX__
-
-            for (; x < grid_size; x += 3)
-            {
-                float sample_x = *gridptr;
-                float sample_y = *(gridptr + 1);
-                float sample_z = *(gridptr + 2);
-
-                sample_x = unormalize(src.w, sample_x);
-                sample_x = get_coord(src.w, sample_x);
-
-                sample_y = unormalize(src.h, sample_y);
-                sample_y = get_coord(src.h, sample_y);
-
-                sample_z = unormalize(src.d, sample_z);
-                sample_z = get_coord(src.d, sample_z);
-
-                int x0 = (int)floorf(sample_x);
-                int y0 = (int)floorf(sample_y);
-                int z0 = (int)floorf(sample_z);
-                int x1 = x0 + 1;
-                int y1 = y0 + 1;
-                int z1 = z0 + 1;
-
-                bool x0_in_range = (x0 > -1) & (x0 < src.w);
-                bool y0_in_range = (y0 > -1) & (y0 < src.h);
-                bool z0_in_range = (z0 > -1) & (z0 < src.d);
-                bool x1_in_range = (x1 > -1) & (x1 < src.w);
-                bool y1_in_range = (y1 > -1) & (y1 < src.h);
-                bool z1_in_range = (z1 > -1) & (z1 < src.d);
-
-                bool v00_in_range = x0_in_range & y0_in_range;
-                bool v01_in_range = x1_in_range & y0_in_range;
-                bool v10_in_range = x0_in_range & y1_in_range;
-                bool v11_in_range = x1_in_range & y1_in_range;
-
-                bool in_bound_000 = v00_in_range & z0_in_range;
-                bool in_bound_001 = v01_in_range & z0_in_range;
-                bool in_bound_010 = v10_in_range & z0_in_range;
-                bool in_bound_011 = v11_in_range & z0_in_range;
-
-                bool in_bound_100 = v00_in_range & z1_in_range;
-                bool in_bound_101 = v01_in_range & z1_in_range;
-                bool in_bound_110 = v10_in_range & z1_in_range;
-                bool in_bound_111 = v11_in_range & z1_in_range;
-
-                int* offset_ptr = (int*)offset_value_ptr;
-                float* value_ptr = offset_value_ptr + 8;
-
-                offset_ptr[0] = in_bound_000 ? (x0 + y0 * src.w + z0 * src.w * src.h) * src.elempack : -1.0;
-                offset_ptr[1] = in_bound_001 ? (x1 + y0 * src.w + z0 * src.w * src.h) * src.elempack : -1.0;
-                offset_ptr[2] = in_bound_010 ? (x0 + y1 * src.w + z0 * src.w * src.h) * src.elempack : -1.0;
-                offset_ptr[3] = in_bound_011 ? (x1 + y1 * src.w + z0 * src.w * src.h) * src.elempack : -1.0;
-
-                offset_ptr[4] = in_bound_100 ? (x0 + y0 * src.w + z1 * src.w * src.h) * src.elempack : -1.0;
-                offset_ptr[5] = in_bound_101 ? (x1 + y0 * src.w + z1 * src.w * src.h) * src.elempack : -1.0;
-                offset_ptr[6] = in_bound_110 ? (x0 + y1 * src.w + z1 * src.w * src.h) * src.elempack : -1.0;
-                offset_ptr[7] = in_bound_111 ? (x1 + y1 * src.w + z1 * src.w * src.h) * src.elempack : -1.0;
-
-                value_ptr[0] = sample_x - x0;
-                value_ptr[1] = sample_y - y0;
-                value_ptr[2] = sample_z - z0;
-
-                gridptr += 3;
-                offset_value_ptr += 11;
-            }
-        }
-    }
-    else
-    {
-        const float* gridptr_x = grid.channel(0);
-        const float* gridptr_y = grid.channel(1);
-        const float* gridptr_z = grid.channel(2);
-
-        int x = 0;
-#if __AVX__
-        for (; x + 7 < grid_size; x += 8)
-        {
-            __m256 gx = _mm256_loadu_ps(gridptr_x);
-            __m256 gy = _mm256_loadu_ps(gridptr_y);
-            __m256 gz = _mm256_loadu_ps(gridptr_z);
-
-            gx = unormalize(_mm256_set1_ps(src.w), gx);
-            gx = get_coord(_mm256_set1_ps(src.w), gx);
-
-            gy = unormalize(_mm256_set1_ps(src.h), gy);
-            gy = get_coord(_mm256_set1_ps(src.h), gy);
-
-            gz = unormalize(_mm256_set1_ps(src.d), gz);
-            gz = get_coord(_mm256_set1_ps(src.d), gz);
-
-            __m256 x_w = _mm256_floor_ps(gx);
-            __m256 y_n = _mm256_floor_ps(gy);
-            __m256 z_t = _mm256_floor_ps(gz);
-
-            __m256 x1 = _mm256_add_ps(x_w, _mm256_set1_ps(1));
-            __m256 y1 = _mm256_add_ps(y_n, _mm256_set1_ps(1));
-            __m256 z1 = _mm256_add_ps(z_t, _mm256_set1_ps(1));
-
-            __m256 x0_in_range = _mm256_and_ps(_mm256_cmp_ps(x_w, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), x_w, _CMP_GT_OS));
-            __m256 x1_in_range = _mm256_and_ps(_mm256_cmp_ps(x1, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), x1, _CMP_GT_OS));
-            __m256 y0_in_range = _mm256_and_ps(_mm256_cmp_ps(y_n, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.h), y_n, _CMP_GT_OS));
-            __m256 y1_in_range = _mm256_and_ps(_mm256_cmp_ps(y1, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.h), y1, _CMP_GT_OS));
-            __m256 z0_in_range = _mm256_and_ps(_mm256_cmp_ps(z_t, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.d), z_t, _CMP_GT_OS));
-            __m256 z1_in_range = _mm256_and_ps(_mm256_cmp_ps(z1, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.d), z1, _CMP_GT_OS));
-
-            __m256 v000_in_range, v010_in_range, v100_in_range, v110_in_range, v001_in_range, v011_in_range, v101_in_range, v111_in_range;
-            {
-                __m256 v00_in_range = _mm256_and_ps(x0_in_range, y0_in_range);
-                __m256 v01_in_range = _mm256_and_ps(x1_in_range, y0_in_range);
-                __m256 v10_in_range = _mm256_and_ps(x0_in_range, y1_in_range);
-                __m256 v11_in_range = _mm256_and_ps(x1_in_range, y1_in_range);
-
-                v000_in_range = _mm256_and_ps(v00_in_range, z0_in_range);
-                v001_in_range = _mm256_and_ps(v01_in_range, z0_in_range);
-                v010_in_range = _mm256_and_ps(v10_in_range, z0_in_range);
-                v011_in_range = _mm256_and_ps(v11_in_range, z0_in_range);
-
-                v100_in_range = _mm256_and_ps(v00_in_range, z1_in_range);
-                v101_in_range = _mm256_and_ps(v01_in_range, z1_in_range);
-                v110_in_range = _mm256_and_ps(v10_in_range, z1_in_range);
-                v111_in_range = _mm256_and_ps(v11_in_range, z1_in_range);
-            }
-
-            __m256 tnw_offset = _mm256_mul_ps(_mm256_comp_fmadd_ps(_mm256_mul_ps(_mm256_set1_ps(src.w), _mm256_set1_ps(src.h)), z_t,
-                                              _mm256_comp_fmadd_ps(y_n, _mm256_set1_ps(src.w), x_w)),
-                                              _mm256_set1_ps(src.elempack));
-            __m256 tne_offset = _mm256_add_ps(tnw_offset, _mm256_set1_ps(src.elempack));
-            __m256 tsw_offset = _mm256_add_ps(tnw_offset, _mm256_mul_ps(_mm256_set1_ps(src.w), _mm256_set1_ps(src.elempack)));
-            __m256 tse_offset = _mm256_add_ps(tsw_offset, _mm256_set1_ps(src.elempack));
-
-            __m256 bnw_offset = _mm256_comp_fmadd_ps(_mm256_mul_ps(_mm256_set1_ps(src.w), _mm256_set1_ps(src.h)), _mm256_set1_ps(src.elempack), tnw_offset);
-            __m256 bne_offset = _mm256_add_ps(bnw_offset, _mm256_set1_ps(src.elempack));
-            __m256 bsw_offset = _mm256_add_ps(bnw_offset, _mm256_mul_ps(_mm256_set1_ps(src.w), _mm256_set1_ps(src.elempack)));
-            __m256 bse_offset = _mm256_add_ps(bsw_offset, _mm256_set1_ps(src.elempack));
-
-            tnw_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), tnw_offset, v000_in_range);
-            tne_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), tne_offset, v001_in_range);
-            tsw_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), tsw_offset, v010_in_range);
-            tse_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), tse_offset, v011_in_range);
-
-            bnw_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), bnw_offset, v100_in_range);
-            bne_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), bne_offset, v101_in_range);
-            bsw_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), bsw_offset, v110_in_range);
-            bse_offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), bse_offset, v111_in_range);
-
-            tnw_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(tnw_offset));
-            tne_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(tne_offset));
-            tsw_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(tsw_offset));
-            tse_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(tse_offset));
-
-            bnw_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(bnw_offset));
-            bne_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(bne_offset));
-            bsw_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(bsw_offset));
-            bse_offset = _mm256_castsi256_ps(_mm256_cvtps_epi32(bse_offset));
-
-            __m256 alpha = _mm256_sub_ps(gx, x_w);
-            __m256 beta = _mm256_sub_ps(gy, y_n);
-            __m256 gamma = _mm256_sub_ps(gz, z_t);
-
-            transpose8x11_ps(tnw_offset, tne_offset, tsw_offset, tse_offset, bnw_offset, bne_offset, bsw_offset, bse_offset, alpha, beta, gamma);
-
-            _mm256_storeu_ps(offset_value_ptr, tnw_offset);
-            _mm256_storeu_ps(offset_value_ptr + 8, tne_offset);
-            _mm256_storeu_ps(offset_value_ptr + 16, tsw_offset);
-            _mm256_storeu_ps(offset_value_ptr + 24, tse_offset);
-
-            _mm256_storeu_ps(offset_value_ptr + 32, bnw_offset);
-            _mm256_storeu_ps(offset_value_ptr + 40, bne_offset);
-            _mm256_storeu_ps(offset_value_ptr + 48, bsw_offset);
-            _mm256_storeu_ps(offset_value_ptr + 56, bse_offset);
-
-            _mm256_storeu_ps(offset_value_ptr + 64, alpha);
-            _mm256_storeu_ps(offset_value_ptr + 72, beta);
-            _mm256_storeu_ps(offset_value_ptr + 80, gamma);
-
-            gridptr_x += 8;
-            gridptr_y += 8;
-            gridptr_z += 8;
-
-            offset_value_ptr += 88;
-        }
-#endif // __AVX__
-
-        for (; x < grid_size; x++)
-        {
-            float sample_x = *gridptr_x;
-            float sample_y = *gridptr_y;
-            float sample_z = *gridptr_z;
-
-            sample_x = unormalize(src.w, sample_x);
-            sample_x = get_coord(src.w, sample_x);
-
-            sample_y = unormalize(src.h, sample_y);
-            sample_y = get_coord(src.h, sample_y);
-
-            sample_z = unormalize(src.d, sample_z);
-            sample_z = get_coord(src.d, sample_z);
-
-            int x0 = (int)floorf(sample_x);
-            int y0 = (int)floorf(sample_y);
-            int z0 = (int)floorf(sample_z);
-            int x1 = x0 + 1;
-            int y1 = y0 + 1;
-            int z1 = z0 + 1;
-
-            bool x0_in_range = (x0 > -1) & (x0 < src.w);
-            bool y0_in_range = (y0 > -1) & (y0 < src.h);
-            bool z0_in_range = (z0 > -1) & (z0 < src.d);
-            bool x1_in_range = (x1 > -1) & (x1 < src.w);
-            bool y1_in_range = (y1 > -1) & (y1 < src.h);
-            bool z1_in_range = (z1 > -1) & (z1 < src.d);
-
-            bool v00_in_range = x0_in_range & y0_in_range;
-            bool v01_in_range = x1_in_range & y0_in_range;
-            bool v10_in_range = x0_in_range & y1_in_range;
-            bool v11_in_range = x1_in_range & y1_in_range;
-
-            bool in_bound_000 = v00_in_range & z0_in_range;
-            bool in_bound_001 = v01_in_range & z0_in_range;
-            bool in_bound_010 = v10_in_range & z0_in_range;
-            bool in_bound_011 = v11_in_range & z0_in_range;
-
-            bool in_bound_100 = v00_in_range & z1_in_range;
-            bool in_bound_101 = v01_in_range & z1_in_range;
-            bool in_bound_110 = v10_in_range & z1_in_range;
-            bool in_bound_111 = v11_in_range & z1_in_range;
-
-            int* offset_ptr = (int*)offset_value_ptr;
-            float* value_ptr = offset_value_ptr + 8;
-
-            offset_ptr[0] = in_bound_000 ? (x0 + y0 * src.w + z0 * src.w * src.h) * src.elempack : -1.0;
-            offset_ptr[1] = in_bound_001 ? (x1 + y0 * src.w + z0 * src.w * src.h) * src.elempack : -1.0;
-            offset_ptr[2] = in_bound_010 ? (x0 + y1 * src.w + z0 * src.w * src.h) * src.elempack : -1.0;
-            offset_ptr[3] = in_bound_011 ? (x1 + y1 * src.w + z0 * src.w * src.h) * src.elempack : -1.0;
-
-            offset_ptr[4] = in_bound_100 ? (x0 + y0 * src.w + z1 * src.w * src.h) * src.elempack : -1.0;
-            offset_ptr[5] = in_bound_101 ? (x1 + y0 * src.w + z1 * src.w * src.h) * src.elempack : -1.0;
-            offset_ptr[6] = in_bound_110 ? (x0 + y1 * src.w + z1 * src.w * src.h) * src.elempack : -1.0;
-            offset_ptr[7] = in_bound_111 ? (x1 + y1 * src.w + z1 * src.w * src.h) * src.elempack : -1.0;
-
-            value_ptr[0] = sample_x - x0;
-            value_ptr[1] = sample_y - y0;
-            value_ptr[2] = sample_z - z0;
-
-            gridptr_x++;
-            gridptr_y++;
-            gridptr_z++;
-            offset_value_ptr += 11;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_compute_blob.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_compute_blob.h
deleted file mode 100644
index 1fc90366..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_compute_blob.h
+++ /dev/null
@@ -1,145 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#include "x86_usability.h"
-
-template<bool align_corner>
-struct grid_sample_unormalize;
-
-template<>
-struct grid_sample_unormalize</*align_corner*/ true>
-{
-#if __AVX__
-    __m256 operator()(__m256 length, __m256 coord)
-    {
-        return _mm256_mul_ps(_mm256_div_ps(_mm256_add_ps(coord, _mm256_set1_ps(1)), _mm256_set1_ps(2)), _mm256_sub_ps(length, _mm256_set1_ps(1)));
-    }
-#endif // __AVX__
-    float operator()(int length, float coord)
-    {
-        return (coord + 1) / 2.f * (length - 1);
-    }
-};
-
-template<>
-struct grid_sample_unormalize</*align_corner*/ false>
-{
-#if __AVX__
-    __m256 operator()(__m256 length, __m256 coord)
-    {
-        return _mm256_div_ps(_mm256_comp_fmsub_ps(_mm256_add_ps(coord, _mm256_set1_ps(1)), length, _mm256_set1_ps(1)), _mm256_set1_ps(2));
-    }
-#endif // __AVX__
-    float operator()(int length, float coord)
-    {
-        return ((coord + 1) * length - 1) / 2.f;
-    }
-};
-
-template<GridSample::PaddingMode pd, bool align_corner>
-struct compute_coord
-{
-#if __AVX__
-    __m256 operator()(__m256 /*length*/, __m256 coord)
-    {
-        return coord;
-    }
-#endif // __AVX__
-    float operator()(int /*length*/, float coord)
-    {
-        return coord;
-    }
-};
-
-template<bool align_corner>
-struct compute_coord<GridSample::Padding_BORDER, align_corner>
-{
-#if __AVX__
-    __m256 operator()(__m256 length, __m256 coord)
-    {
-        const __m256 border_x = _mm256_sub_ps(length, _mm256_set1_ps(1));
-
-        coord = _mm256_min_ps(border_x, _mm256_max_ps(coord, _mm256_setzero_ps()));
-
-        return coord;
-    }
-#endif // __AVX__
-    float operator()(int length, float coord)
-    {
-        return std::min(length - 1.0f, std::max(coord, 0.0f));
-    }
-};
-
-template<>
-struct compute_coord<GridSample::Padding_REFLECTION, /*align_corner*/ true>
-{
-#if __AVX__
-    __m256 operator()(__m256 length, __m256 coord)
-    {
-        const __m256 border_x = _mm256_sub_ps(length, _mm256_set1_ps(1));
-
-        coord = abs256_ps(coord);
-
-        __m256 reflectx_v = abs256_ps(_mm256_sub_ps(coord, border_x));
-        coord = _mm256_sub_ps(border_x, reflectx_v);
-
-        return coord;
-    }
-#endif // __AVX__
-    float operator()(int length, float coord)
-    {
-        coord = fabs(coord);
-        coord = (length - 1) - fabs(coord - (length - 1));
-
-        return std::min(length - 1.0f, std::max(coord, 0.0f));
-    }
-};
-
-template<>
-struct compute_coord<GridSample::Padding_REFLECTION, /*align_corner*/ false>
-{
-#if __AVX__
-    __m256 operator()(__m256 length, __m256 coord)
-    {
-        const __m256 border_x = _mm256_sub_ps(length, _mm256_set1_ps(1));
-
-        __m256 v0p5fp8 = _mm256_set1_ps(0.5f);
-        coord = _mm256_add_ps(coord, v0p5fp8);
-
-        coord = abs256_ps(coord);
-
-        __m256 reflectx_v = abs256_ps(_mm256_sub_ps(coord, length));
-        coord = _mm256_sub_ps(length, reflectx_v);
-
-        coord = _mm256_sub_ps(coord, v0p5fp8);
-
-        _mm256_sub_ps(coord, v0p5fp8);
-
-        coord = _mm256_min_ps(border_x, _mm256_max_ps(coord, _mm256_setzero_ps()));
-
-        return coord;
-    }
-#endif // __AVX__
-    float operator()(int length, float coord)
-    {
-        coord = fabs(coord + 0.5f);
-        coord = length - fabs(coord - length) - 0.5;
-
-        return std::min(length - 1.0f, std::max(coord, 0.0f));
-    }
-};
-
-#include "gridsample_bilinear_compute_blob.h"
-#include "gridsample_bicubic_compute_blob.h"
-#include "gridsample_nearest_compute_blob.h"
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_nearest_apply_interpolation.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_nearest_apply_interpolation.h
deleted file mode 100644
index e84cdc7d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_nearest_apply_interpolation.h
+++ /dev/null
@@ -1,126 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-static void gridsample_nearest_apply_interpolation_p16(const Mat& src, Mat& dst, const Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int outd = dst.d;
-    const int grid_size = outw * outh * outd;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const int* offset_ptr = offset_value.channel(0);
-
-        for (int i = 0; i < grid_size; i++)
-        {
-            __m512 _v = offset_ptr[0] >= 0 ? _mm512_loadu_ps(srcptr + offset_ptr[0]) : _mm512_set1_ps(0);
-            offset_ptr++;
-
-            _mm512_storeu_ps(dstptr, _v);
-            dstptr += 16;
-        }
-    }
-}
-#endif // __AVX512F__
-
-static void gridsample_nearest_apply_interpolation_p8(const Mat& src, Mat& dst, const Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int outd = dst.d;
-    const int grid_size = outw * outh * outd;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const int* offset_ptr = offset_value.channel(0);
-
-        for (int i = 0; i < grid_size; i++)
-        {
-            __m256 _v = offset_ptr[0] >= 0 ? _mm256_loadu_ps(srcptr + offset_ptr[0]) : _mm256_set1_ps(0);
-            offset_ptr++;
-
-            _mm256_storeu_ps(dstptr, _v);
-            dstptr += 8;
-        }
-    }
-}
-#endif // __AVX__
-static void gridsample_nearest_apply_interpolation_p4(const Mat& src, Mat& dst, const Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int outd = dst.d;
-    const int grid_size = outw * outh * outd;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const int* offset_ptr = offset_value.channel(0);
-
-        for (int i = 0; i < grid_size; i++)
-        {
-            __m128 _v = offset_ptr[0] >= 0 ? _mm_loadu_ps(srcptr + offset_ptr[0]) : _mm_set1_ps(0);
-            offset_ptr++;
-
-            _mm_storeu_ps(dstptr, _v);
-            dstptr += 4;
-        }
-    }
-}
-
-#endif // __SSE2__
-
-static void gridsample_nearest_apply_interpolation_p1(const Mat& src, Mat& dst, const Mat& offset_value, const Option& opt)
-{
-    const int channels = dst.c;
-    const int outw = dst.w;
-    const int outh = dst.h;
-    const int outd = dst.d;
-    const int grid_size = outw * outh * outd;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < channels; q++)
-    {
-        const float* srcptr = src.channel(q);
-        float* dstptr = dst.channel(q);
-
-        const int* offset_ptr = offset_value.channel(0);
-
-        for (int x = 0; x < grid_size; x++)
-        {
-            *dstptr = offset_ptr[0] >= 0 ? *(srcptr + offset_ptr[0]) : 0;
-
-            offset_ptr++;
-            dstptr++;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_nearest_compute_blob.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_nearest_compute_blob.h
deleted file mode 100644
index a7a12066..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_nearest_compute_blob.h
+++ /dev/null
@@ -1,315 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-template<GridSample::PaddingMode pd, bool align_corner>
-void gridsample_2d_nearest_compute_blob(const Mat& src, const Mat& grid, Mat& offset_value, int permute_fusion)
-{
-    const int grid_size = grid.w * grid.h;
-
-    float* offset_ptr = offset_value.channel(0);
-
-    grid_sample_unormalize<align_corner> unormalize;
-    compute_coord<pd, align_corner> get_coord;
-
-    if (permute_fusion == 0)
-    {
-        for (int y = 0; y < grid.c; y++)
-        {
-            const float* gridptr = grid.channel(y);
-            int x = 0;
-#if __AVX__
-            for (; x + 15 < grid_size; x += 16)
-            {
-                __m256 gx = _mm256_loadu_ps(gridptr);
-                __m256 gy = _mm256_loadu_ps(gridptr + 8);
-
-                transpose2x8_ps(gx, gy);
-
-                gx = unormalize(_mm256_set1_ps(src.w), gx);
-                gx = get_coord(_mm256_set1_ps(src.w), gx);
-
-                gy = unormalize(_mm256_set1_ps(src.h), gy);
-                gy = get_coord(_mm256_set1_ps(src.h), gy);
-
-                gx = _mm256_floor_ps(_mm256_add_ps(gx, _mm256_set1_ps(0.5f)));
-                gy = _mm256_floor_ps(_mm256_add_ps(gy, _mm256_set1_ps(0.5f)));
-
-                __m256 v_in_range = _mm256_and_ps(_mm256_and_ps(_mm256_cmp_ps(gx, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), gx, _CMP_GT_OS)),
-                                                  _mm256_and_ps(_mm256_cmp_ps(gy, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.h), gy, _CMP_GT_OS)));
-
-                __m256 offset = _mm256_mul_ps(_mm256_comp_fmadd_ps(gy, _mm256_set1_ps(src.w), gx), _mm256_set1_ps(src.elempack));
-
-                offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), _mm256_castsi256_ps(_mm256_cvtps_epi32(offset)), v_in_range);
-
-                _mm256_storeu_ps(offset_ptr, offset);
-
-                gridptr += 16;
-                offset_ptr += 8;
-            }
-
-#endif // __AVX__
-
-            for (; x < grid_size; x += 2)
-            {
-                float sample_x = *gridptr;
-                float sample_y = *(gridptr + 1);
-
-                sample_x = unormalize(src.w, sample_x);
-                sample_x = get_coord(src.w, sample_x);
-
-                sample_y = unormalize(src.h, sample_y);
-                sample_y = get_coord(src.h, sample_y);
-
-                int x0 = static_cast<int>(floorf(sample_x + 0.5f));
-                int y0 = static_cast<int>(floorf(sample_y + 0.5f));
-
-                bool in_bound = ((x0 > -1) & (x0 < src.w) & (y0 > -1) & (y0 < src.h));
-
-                int* iptr = (int*)offset_ptr;
-                *iptr = in_bound ? (x0 + y0 * src.w) * src.elempack : -1.0;
-
-                gridptr += 2;
-                offset_ptr++;
-            }
-        }
-    }
-    else
-    {
-        const float* gridptr_x = grid.channel(0);
-        const float* gridptr_y = grid.channel(1);
-
-        int x = 0;
-#if __AVX__
-        for (; x + 7 < grid_size; x += 8)
-        {
-            __m256 gx = _mm256_loadu_ps(gridptr_x);
-            __m256 gy = _mm256_loadu_ps(gridptr_y);
-
-            gx = unormalize(_mm256_set1_ps(src.w), gx);
-            gx = get_coord(_mm256_set1_ps(src.w), gx);
-
-            gy = unormalize(_mm256_set1_ps(src.h), gy);
-            gy = get_coord(_mm256_set1_ps(src.h), gy);
-
-            gx = _mm256_floor_ps(_mm256_add_ps(gx, _mm256_set1_ps(0.5f)));
-            gy = _mm256_floor_ps(_mm256_add_ps(gy, _mm256_set1_ps(0.5f)));
-
-            __m256 v_in_range = _mm256_and_ps(_mm256_and_ps(_mm256_cmp_ps(gx, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), gx, _CMP_GT_OS)),
-                                              _mm256_and_ps(_mm256_cmp_ps(gy, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.h), gy, _CMP_GT_OS)));
-
-            __m256 offset = _mm256_mul_ps(_mm256_comp_fmadd_ps(gy, _mm256_set1_ps(src.w), gx), _mm256_set1_ps(src.elempack));
-
-            offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), _mm256_castsi256_ps(_mm256_cvtps_epi32(offset)), v_in_range);
-
-            _mm256_storeu_ps(offset_ptr, offset);
-
-            gridptr_x += 8;
-            gridptr_y += 8;
-            offset_ptr += 8;
-        }
-
-#endif // __AVX__
-
-        for (; x < grid_size; x++)
-        {
-            float sample_x = *gridptr_x;
-            float sample_y = *gridptr_y;
-
-            sample_x = unormalize(src.w, sample_x);
-            sample_x = get_coord(src.w, sample_x);
-
-            sample_y = unormalize(src.h, sample_y);
-            sample_y = get_coord(src.h, sample_y);
-
-            int x0 = static_cast<int>(floorf(sample_x + 0.5f));
-            int y0 = static_cast<int>(floorf(sample_y + 0.5f));
-
-            bool in_bound = ((x0 > -1) & (x0 < src.w) & (y0 > -1) & (y0 < src.h));
-
-            int* iptr = (int*)offset_ptr;
-            *iptr = in_bound ? (x0 + y0 * src.w) * src.elempack : -1.0;
-
-            gridptr_x++;
-            gridptr_y++;
-
-            offset_ptr++;
-        }
-    }
-}
-
-template<GridSample::PaddingMode pd, bool align_corner>
-void gridsample_3d_nearest_compute_blob(const Mat& src, const Mat& grid, Mat& offset_value, int permute_fusion)
-{
-    const int grid_size = grid.w * grid.h * grid.d;
-
-    float* offset_ptr = offset_value.channel(0);
-
-    grid_sample_unormalize<align_corner> unormalize;
-    compute_coord<pd, align_corner> get_coord;
-
-    if (permute_fusion == 0)
-    {
-        for (int y = 0; y < grid.c; y++)
-        {
-            const float* gridptr = grid.channel(y);
-            int x = 0;
-#if __AVX__
-            for (; x + 23 < grid_size; x += 24)
-            {
-                __m256 gx = _mm256_loadu_ps(gridptr);
-                __m256 gy = _mm256_loadu_ps(gridptr + 8);
-                __m256 gz = _mm256_loadu_ps(gridptr + 16);
-
-                transpose3x8_ps(gx, gy, gz);
-
-                gx = unormalize(_mm256_set1_ps(src.w), gx);
-                gx = get_coord(_mm256_set1_ps(src.w), gx);
-
-                gy = unormalize(_mm256_set1_ps(src.h), gy);
-                gy = get_coord(_mm256_set1_ps(src.h), gy);
-
-                gz = unormalize(_mm256_set1_ps(src.d), gz);
-                gz = get_coord(_mm256_set1_ps(src.d), gz);
-
-                gx = _mm256_floor_ps(_mm256_add_ps(gx, _mm256_set1_ps(0.5f)));
-                gy = _mm256_floor_ps(_mm256_add_ps(gy, _mm256_set1_ps(0.5f)));
-                gz = _mm256_floor_ps(_mm256_add_ps(gz, _mm256_set1_ps(0.5f)));
-
-                __m256 v_in_range = _mm256_and_ps(_mm256_and_ps(_mm256_cmp_ps(gx, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), gx, _CMP_GT_OS)),
-                                                  _mm256_and_ps(_mm256_cmp_ps(gy, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.h), gy, _CMP_GT_OS)));
-                v_in_range = _mm256_and_ps(v_in_range, _mm256_and_ps(_mm256_cmp_ps(gz, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.d), gz, _CMP_GT_OS)));
-
-                __m256 offset = _mm256_mul_ps(_mm256_comp_fmadd_ps(_mm256_mul_ps(_mm256_set1_ps(src.w), _mm256_set1_ps(src.h)), gz,
-                                              _mm256_comp_fmadd_ps(gy, _mm256_set1_ps(src.w), gx)),
-                                              _mm256_set1_ps(src.elempack));
-
-                offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), _mm256_castsi256_ps(_mm256_cvtps_epi32(offset)), v_in_range);
-
-                _mm256_storeu_ps(offset_ptr, offset);
-
-                gridptr += 24;
-                offset_ptr += 8;
-            }
-
-#endif // __AVX__
-
-            for (; x < grid_size; x += 3)
-            {
-                float sample_x = *gridptr;
-                float sample_y = *(gridptr + 1);
-                float sample_z = *(gridptr + 2);
-
-                sample_x = unormalize(src.w, sample_x);
-                sample_x = get_coord(src.w, sample_x);
-
-                sample_y = unormalize(src.h, sample_y);
-                sample_y = get_coord(src.h, sample_y);
-
-                sample_z = unormalize(src.d, sample_z);
-                sample_z = get_coord(src.d, sample_z);
-
-                int x0 = static_cast<int>(floorf(sample_x + 0.5f));
-                int y0 = static_cast<int>(floorf(sample_y + 0.5f));
-                int z0 = static_cast<int>(floorf(sample_z + 0.5f));
-
-                bool in_bound = ((x0 > -1) & (x0 < src.w) & (y0 > -1) & (y0 < src.h) & (z0 > -1) & (z0 < src.d));
-
-                int* iptr = (int*)offset_ptr;
-                *iptr = in_bound ? (x0 + y0 * src.w + z0 * src.w * src.h) * src.elempack : -1.0;
-
-                gridptr += 3;
-                offset_ptr++;
-            }
-        }
-    }
-    else
-    {
-        const float* gridptr_x = grid.channel(0);
-        const float* gridptr_y = grid.channel(1);
-        const float* gridptr_z = grid.channel(2);
-
-        int x = 0;
-#if __AVX__
-        for (; x + 7 < grid_size; x += 8)
-        {
-            __m256 gx = _mm256_loadu_ps(gridptr_x);
-            __m256 gy = _mm256_loadu_ps(gridptr_y);
-            __m256 gz = _mm256_loadu_ps(gridptr_z);
-
-            gx = unormalize(_mm256_set1_ps(src.w), gx);
-            gx = get_coord(_mm256_set1_ps(src.w), gx);
-
-            gy = unormalize(_mm256_set1_ps(src.h), gy);
-            gy = get_coord(_mm256_set1_ps(src.h), gy);
-
-            gz = unormalize(_mm256_set1_ps(src.d), gz);
-            gz = get_coord(_mm256_set1_ps(src.d), gz);
-
-            gx = _mm256_floor_ps(_mm256_add_ps(gx, _mm256_set1_ps(0.5f)));
-            gy = _mm256_floor_ps(_mm256_add_ps(gy, _mm256_set1_ps(0.5f)));
-            gz = _mm256_floor_ps(_mm256_add_ps(gz, _mm256_set1_ps(0.5f)));
-
-            __m256 v_in_range = _mm256_and_ps(_mm256_and_ps(_mm256_cmp_ps(gx, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.w), gx, _CMP_GT_OS)),
-                                              _mm256_and_ps(_mm256_cmp_ps(gy, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.h), gy, _CMP_GT_OS)));
-            v_in_range = _mm256_and_ps(v_in_range, _mm256_and_ps(_mm256_cmp_ps(gz, _mm256_set1_ps(-1), _CMP_GT_OS), _mm256_cmp_ps(_mm256_set1_ps(src.d), gz, _CMP_GT_OS)));
-
-            __m256 offset = _mm256_mul_ps(_mm256_comp_fmadd_ps(_mm256_mul_ps(_mm256_set1_ps(src.w), _mm256_set1_ps(src.h)), gz,
-                                          _mm256_comp_fmadd_ps(gy, _mm256_set1_ps(src.w), gx)),
-                                          _mm256_set1_ps(src.elempack));
-
-            offset = _mm256_blendv_ps(_mm256_set1_ps(-1.0f), _mm256_castsi256_ps(_mm256_cvtps_epi32(offset)), v_in_range);
-
-            _mm256_storeu_ps(offset_ptr, offset);
-
-            gridptr_x += 8;
-            gridptr_y += 8;
-            gridptr_z += 8;
-
-            offset_ptr += 8;
-        }
-
-#endif // __AVX__
-
-        for (; x < grid_size; x++)
-        {
-            float sample_x = *gridptr_x;
-            float sample_y = *gridptr_y;
-            float sample_z = *gridptr_z;
-
-            sample_x = unormalize(src.w, sample_x);
-            sample_x = get_coord(src.w, sample_x);
-
-            sample_y = unormalize(src.h, sample_y);
-            sample_y = get_coord(src.h, sample_y);
-
-            sample_z = unormalize(src.d, sample_z);
-            sample_z = get_coord(src.d, sample_z);
-
-            int x0 = static_cast<int>(floorf(sample_x + 0.5f));
-            int y0 = static_cast<int>(floorf(sample_y + 0.5f));
-            int z0 = static_cast<int>(floorf(sample_z + 0.5f));
-
-            bool in_bound = ((x0 > -1) & (x0 < src.w) & (y0 > -1) & (y0 < src.h) & (z0 > -1) & (z0 < src.d));
-
-            int* iptr = (int*)offset_ptr;
-            *iptr = in_bound ? (x0 + y0 * src.w + z0 * src.w * src.h) * src.elempack : -1.0;
-
-            gridptr_x++;
-            gridptr_y++;
-            gridptr_z++;
-
-            offset_ptr++;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_x86.h
deleted file mode 100644
index caf7c7c5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/gridsample_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GRIDSAMPLE_X86_H
-#define LAYER_GRIDSAMPLE_X86_H
-
-#include "gridsample.h"
-
-namespace ncnn {
-
-class GridSample_x86 : public GridSample
-{
-public:
-    GridSample_x86();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GRIDSAMPLE_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/groupnorm_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/groupnorm_x86.h
deleted file mode 100644
index 151884e5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/groupnorm_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_GROUPNORM_X86_H
-#define LAYER_GROUPNORM_X86_H
-
-#include "groupnorm.h"
-
-namespace ncnn {
-
-class GroupNorm_x86 : public GroupNorm
-{
-public:
-    GroupNorm_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_GROUPNORM_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/hardsigmoid_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/hardsigmoid_x86.h
deleted file mode 100644
index 418a8dc9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/hardsigmoid_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_HARDSIGMOID_X86_H
-#define LAYER_HARDSIGMOID_X86_H
-
-#include "hardsigmoid.h"
-
-namespace ncnn {
-
-class HardSigmoid_x86 : public HardSigmoid
-{
-public:
-    HardSigmoid_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_HARDSIGMOID_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/hardswish_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/hardswish_x86.h
deleted file mode 100644
index 4fe521ea..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/hardswish_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_HARDSWISH_X86_H
-#define LAYER_HARDSWISH_X86_H
-
-#include "hardswish.h"
-
-namespace ncnn {
-
-class HardSwish_x86 : public HardSwish
-{
-public:
-    HardSwish_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_HARDSWISH_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/innerproduct_fp.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/innerproduct_fp.h
deleted file mode 100644
index 6edcd74d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/innerproduct_fp.h
+++ /dev/null
@@ -1,1401 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if NCNN_RUNTIME_CPU && NCNN_F16C && __AVX__ && !__F16C__
-void innerproduct_fp16s_sse_f16c(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt);
-void innerproduct_transform_kernel_fp16s_sse_f16c(const Mat& weight_data, Mat& weight_data_tm, int num_input, int num_output, const Option& opt);
-#endif
-
-#if NCNN_IMPL_FP16S
-static void innerproduct_fp16s_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt)
-#else
-static void innerproduct_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt)
-#endif
-{
-#if NCNN_RUNTIME_CPU && NCNN_IMPL_FP16S && NCNN_F16C && __AVX__ && !__F16C__
-    if (ncnn::cpu_support_x86_f16c())
-    {
-        innerproduct_fp16s_sse_f16c(bottom_blob, top_blob, weight_data_tm, bias_data, activation_type, activation_params, opt);
-        return;
-    }
-#else // NCNN_RUNTIME_CPU
-
-    const int num_input = bottom_blob.w * bottom_blob.elempack;
-    const int outw = top_blob.w;
-    const int out_elempack = top_blob.elempack;
-
-    const float* bias_data_ptr = bias_data;
-
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    if (out_elempack == 16)
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < outw; p++)
-        {
-            __m512 _sum0 = _mm512_setzero_ps();
-            __m512 _sum1 = _mm512_setzero_ps();
-            __m512 _sum2 = _mm512_setzero_ps();
-            __m512 _sum3 = _mm512_setzero_ps();
-            __m512 _sum4 = _mm512_setzero_ps();
-            __m512 _sum5 = _mm512_setzero_ps();
-            __m512 _sum6 = _mm512_setzero_ps();
-            __m512 _sum7 = _mm512_setzero_ps();
-
-            if (bias_data_ptr)
-            {
-                _sum0 = _mm512_loadu_ps(bias_data_ptr + p * 16);
-            }
-
-#if NCNN_IMPL_FP16S
-            const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-            const float* kptr = weight_data_tm.row(p);
-#endif
-            const float* sptr = bottom_blob;
-
-            int i = 0;
-            for (; i + 7 < num_input; i += 8)
-            {
-                __m512 _val0 = _mm512_set1_ps(sptr[0]);
-                __m512 _val1 = _mm512_set1_ps(sptr[1]);
-                __m512 _val2 = _mm512_set1_ps(sptr[2]);
-                __m512 _val3 = _mm512_set1_ps(sptr[3]);
-#if NCNN_IMPL_FP16S
-                __m512i _w01 = _mm512_loadu_si512(kptr);
-                __m512i _w23 = _mm512_loadu_si512(kptr + 32);
-                __m512 _w0 = _mm512_cvtph_ps(_mm512_extracti32x8_epi32(_w01, 0));
-                __m512 _w1 = _mm512_cvtph_ps(_mm512_extracti32x8_epi32(_w01, 1));
-                __m512 _w2 = _mm512_cvtph_ps(_mm512_extracti32x8_epi32(_w23, 0));
-                __m512 _w3 = _mm512_cvtph_ps(_mm512_extracti32x8_epi32(_w23, 1));
-#else
-                __m512 _w0 = _mm512_loadu_ps(kptr + 16 * 0);
-                __m512 _w1 = _mm512_loadu_ps(kptr + 16 * 1);
-                __m512 _w2 = _mm512_loadu_ps(kptr + 16 * 2);
-                __m512 _w3 = _mm512_loadu_ps(kptr + 16 * 3);
-#endif
-
-                _sum0 = _mm512_fmadd_ps(_val0, _w0, _sum0);
-                _sum1 = _mm512_fmadd_ps(_val1, _w1, _sum1);
-                _sum2 = _mm512_fmadd_ps(_val2, _w2, _sum2);
-                _sum3 = _mm512_fmadd_ps(_val3, _w3, _sum3);
-
-                __m512 _val4 = _mm512_set1_ps(sptr[4]);
-                __m512 _val5 = _mm512_set1_ps(sptr[5]);
-                __m512 _val6 = _mm512_set1_ps(sptr[6]);
-                __m512 _val7 = _mm512_set1_ps(sptr[7]);
-#if NCNN_IMPL_FP16S
-                __m512i _w45 = _mm512_loadu_si512(kptr + 64);
-                __m512i _w67 = _mm512_loadu_si512(kptr + 96);
-                __m512 _w4 = _mm512_cvtph_ps(_mm512_extracti32x8_epi32(_w45, 0));
-                __m512 _w5 = _mm512_cvtph_ps(_mm512_extracti32x8_epi32(_w45, 1));
-                __m512 _w6 = _mm512_cvtph_ps(_mm512_extracti32x8_epi32(_w67, 0));
-                __m512 _w7 = _mm512_cvtph_ps(_mm512_extracti32x8_epi32(_w67, 1));
-#else
-                __m512 _w4 = _mm512_loadu_ps(kptr + 16 * 4);
-                __m512 _w5 = _mm512_loadu_ps(kptr + 16 * 5);
-                __m512 _w6 = _mm512_loadu_ps(kptr + 16 * 6);
-                __m512 _w7 = _mm512_loadu_ps(kptr + 16 * 7);
-#endif
-
-                _sum4 = _mm512_fmadd_ps(_val4, _w4, _sum4);
-                _sum5 = _mm512_fmadd_ps(_val5, _w5, _sum5);
-                _sum6 = _mm512_fmadd_ps(_val6, _w6, _sum6);
-                _sum7 = _mm512_fmadd_ps(_val7, _w7, _sum7);
-
-                sptr += 8;
-                kptr += 128;
-            }
-            for (; i + 3 < num_input; i += 4)
-            {
-                __m512 _val0 = _mm512_set1_ps(sptr[0]);
-                __m512 _val1 = _mm512_set1_ps(sptr[1]);
-                __m512 _val2 = _mm512_set1_ps(sptr[2]);
-                __m512 _val3 = _mm512_set1_ps(sptr[3]);
-#if NCNN_IMPL_FP16S
-                __m512i _w01 = _mm512_loadu_si512(kptr);
-                __m512i _w23 = _mm512_loadu_si512(kptr + 32);
-                __m512 _w0 = _mm512_cvtph_ps(_mm512_extracti32x8_epi32(_w01, 0));
-                __m512 _w1 = _mm512_cvtph_ps(_mm512_extracti32x8_epi32(_w01, 1));
-                __m512 _w2 = _mm512_cvtph_ps(_mm512_extracti32x8_epi32(_w23, 0));
-                __m512 _w3 = _mm512_cvtph_ps(_mm512_extracti32x8_epi32(_w23, 1));
-#else
-                __m512 _w0 = _mm512_loadu_ps(kptr);
-                __m512 _w1 = _mm512_loadu_ps(kptr + 16);
-                __m512 _w2 = _mm512_loadu_ps(kptr + 32);
-                __m512 _w3 = _mm512_loadu_ps(kptr + 48);
-#endif
-
-                _sum0 = _mm512_fmadd_ps(_val0, _w0, _sum0);
-                _sum1 = _mm512_fmadd_ps(_val1, _w1, _sum1);
-                _sum2 = _mm512_fmadd_ps(_val2, _w2, _sum2);
-                _sum3 = _mm512_fmadd_ps(_val3, _w3, _sum3);
-
-                sptr += 4;
-                kptr += 64;
-            }
-            for (; i < num_input; i++)
-            {
-                __m512 _val = _mm512_set1_ps(sptr[0]);
-#if NCNN_IMPL_FP16S
-                __m512 _w = _mm512_cvtph_ps(_mm256_lddqu_si256((const __m256i*)kptr));
-#else
-                __m512 _w = _mm512_loadu_ps(kptr);
-#endif
-                _sum0 = _mm512_fmadd_ps(_val, _w, _sum0);
-
-                sptr += 1;
-                kptr += 16;
-            }
-
-            _sum0 = _mm512_add_ps(_sum0, _sum1);
-            _sum2 = _mm512_add_ps(_sum2, _sum3);
-            _sum4 = _mm512_add_ps(_sum4, _sum5);
-            _sum6 = _mm512_add_ps(_sum6, _sum7);
-            _sum0 = _mm512_add_ps(_sum0, _sum2);
-            _sum4 = _mm512_add_ps(_sum4, _sum6);
-            _sum0 = _mm512_add_ps(_sum0, _sum4);
-
-            _sum0 = activation_avx512(_sum0, activation_type, activation_params);
-
-            float* outptr = top_blob;
-            _mm512_storeu_ps(outptr + p * 16, _sum0);
-        }
-    }
-#endif // __AVX512F__
-
-    if (out_elempack == 8)
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < outw; p++)
-        {
-            __m256 _sum0 = _mm256_setzero_ps();
-            __m256 _sum1 = _mm256_setzero_ps();
-            __m256 _sum2 = _mm256_setzero_ps();
-            __m256 _sum3 = _mm256_setzero_ps();
-            __m256 _sum4 = _mm256_setzero_ps();
-            __m256 _sum5 = _mm256_setzero_ps();
-            __m256 _sum6 = _mm256_setzero_ps();
-            __m256 _sum7 = _mm256_setzero_ps();
-
-            if (bias_data_ptr)
-            {
-                _sum0 = _mm256_loadu_ps(bias_data_ptr + p * 8);
-            }
-
-#if NCNN_IMPL_FP16S
-            const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-            const float* kptr = weight_data_tm.row(p);
-#endif
-            const float* sptr = bottom_blob;
-
-            int i = 0;
-            for (; i + 7 < num_input; i += 8)
-            {
-                __m256 _val0 = _mm256_broadcast_ss(sptr);
-                __m256 _val1 = _mm256_broadcast_ss(sptr + 1);
-                __m256 _val2 = _mm256_broadcast_ss(sptr + 2);
-                __m256 _val3 = _mm256_broadcast_ss(sptr + 3);
-#if NCNN_IMPL_FP16S
-                __m256i _w01 = _mm256_lddqu_si256((const __m256i*)kptr);
-                __m256i _w23 = _mm256_lddqu_si256((const __m256i*)(kptr + 16));
-                __m256 _w0 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w01, 0));
-                __m256 _w1 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w01, 1));
-                __m256 _w2 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w23, 0));
-                __m256 _w3 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w23, 1));
-#else
-                __m256 _w0 = _mm256_loadu_ps(kptr);
-                __m256 _w1 = _mm256_loadu_ps(kptr + 8);
-                __m256 _w2 = _mm256_loadu_ps(kptr + 16);
-                __m256 _w3 = _mm256_loadu_ps(kptr + 24);
-#endif
-
-                _sum0 = _mm256_comp_fmadd_ps(_val0, _w0, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_val1, _w1, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_val2, _w2, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_val3, _w3, _sum3);
-
-                __m256 _val4 = _mm256_broadcast_ss(sptr + 4);
-                __m256 _val5 = _mm256_broadcast_ss(sptr + 5);
-                __m256 _val6 = _mm256_broadcast_ss(sptr + 6);
-                __m256 _val7 = _mm256_broadcast_ss(sptr + 7);
-#if NCNN_IMPL_FP16S
-                __m256i _w45 = _mm256_lddqu_si256((const __m256i*)(kptr + 32));
-                __m256i _w67 = _mm256_lddqu_si256((const __m256i*)(kptr + 48));
-                __m256 _w4 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w45, 0));
-                __m256 _w5 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w45, 1));
-                __m256 _w6 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w67, 0));
-                __m256 _w7 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w67, 1));
-#else
-                __m256 _w4 = _mm256_loadu_ps(kptr + 32);
-                __m256 _w5 = _mm256_loadu_ps(kptr + 40);
-                __m256 _w6 = _mm256_loadu_ps(kptr + 48);
-                __m256 _w7 = _mm256_loadu_ps(kptr + 56);
-#endif
-
-                _sum4 = _mm256_comp_fmadd_ps(_val4, _w4, _sum4);
-                _sum5 = _mm256_comp_fmadd_ps(_val5, _w5, _sum5);
-                _sum6 = _mm256_comp_fmadd_ps(_val6, _w6, _sum6);
-                _sum7 = _mm256_comp_fmadd_ps(_val7, _w7, _sum7);
-
-                sptr += 8;
-                kptr += 64;
-            }
-            for (; i + 3 < num_input; i += 4)
-            {
-                __m256 _val0 = _mm256_broadcast_ss(sptr);
-                __m256 _val1 = _mm256_broadcast_ss(sptr + 1);
-                __m256 _val2 = _mm256_broadcast_ss(sptr + 2);
-                __m256 _val3 = _mm256_broadcast_ss(sptr + 3);
-#if NCNN_IMPL_FP16S
-                __m256i _w01 = _mm256_lddqu_si256((const __m256i*)kptr);
-                __m256i _w23 = _mm256_lddqu_si256((const __m256i*)(kptr + 16));
-                __m256 _w0 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w01, 0));
-                __m256 _w1 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w01, 1));
-                __m256 _w2 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w23, 0));
-                __m256 _w3 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w23, 1));
-#else
-                __m256 _w0 = _mm256_loadu_ps(kptr);
-                __m256 _w1 = _mm256_loadu_ps(kptr + 8);
-                __m256 _w2 = _mm256_loadu_ps(kptr + 16);
-                __m256 _w3 = _mm256_loadu_ps(kptr + 24);
-#endif
-
-                _sum0 = _mm256_comp_fmadd_ps(_val0, _w0, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_val1, _w1, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_val2, _w2, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_val3, _w3, _sum3);
-
-                sptr += 4;
-                kptr += 32;
-            }
-            for (; i < num_input; i++)
-            {
-                __m256 _val = _mm256_set1_ps(sptr[0]);
-#if NCNN_IMPL_FP16S
-                __m256 _w = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)kptr));
-#else
-                __m256 _w = _mm256_loadu_ps(kptr);
-#endif
-                _sum0 = _mm256_comp_fmadd_ps(_val, _w, _sum0);
-
-                sptr += 1;
-                kptr += 8;
-            }
-
-            _sum0 = _mm256_add_ps(_sum0, _sum1);
-            _sum2 = _mm256_add_ps(_sum2, _sum3);
-            _sum4 = _mm256_add_ps(_sum4, _sum5);
-            _sum6 = _mm256_add_ps(_sum6, _sum7);
-            _sum0 = _mm256_add_ps(_sum0, _sum2);
-            _sum4 = _mm256_add_ps(_sum4, _sum6);
-            _sum0 = _mm256_add_ps(_sum0, _sum4);
-
-            _sum0 = activation_avx(_sum0, activation_type, activation_params);
-
-            float* outptr = top_blob;
-            _mm256_storeu_ps(outptr + p * 8, _sum0);
-        }
-    }
-#endif // __AVX__
-
-    if (out_elempack == 4)
-    {
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = 0; p < outw; p++)
-        {
-            __m128 _sum0 = _mm_setzero_ps();
-#if __AVX__
-            __m256 _sum01 = _mm256_setzero_ps();
-            __m256 _sum23 = _mm256_setzero_ps();
-            __m256 _sum45 = _mm256_setzero_ps();
-            __m256 _sum67 = _mm256_setzero_ps();
-#else
-            __m128 _sum1 = _mm_setzero_ps();
-            __m128 _sum2 = _mm_setzero_ps();
-            __m128 _sum3 = _mm_setzero_ps();
-#endif
-
-            if (bias_data_ptr)
-            {
-                _sum0 = _mm_loadu_ps(bias_data_ptr + p * 4);
-            }
-
-#if NCNN_IMPL_FP16S
-            const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-            const float* kptr = weight_data_tm.row(p);
-#endif
-            const float* sptr = bottom_blob;
-
-            int i = 0;
-#if __AVX__
-            for (; i + 7 < num_input; i += 8)
-            {
-                __m128 _val0 = _mm_broadcast_ss(sptr);
-                __m128 _val1 = _mm_broadcast_ss(sptr + 1);
-                __m128 _val2 = _mm_broadcast_ss(sptr + 2);
-                __m128 _val3 = _mm_broadcast_ss(sptr + 3);
-                __m128 _val4 = _mm_broadcast_ss(sptr + 4);
-                __m128 _val5 = _mm_broadcast_ss(sptr + 5);
-                __m128 _val6 = _mm_broadcast_ss(sptr + 6);
-                __m128 _val7 = _mm_broadcast_ss(sptr + 7);
-
-                __m256 _val01 = _mm256_insertf128_ps(_mm256_castps128_ps256(_val0), _val1, 1);
-                __m256 _val23 = _mm256_insertf128_ps(_mm256_castps128_ps256(_val2), _val3, 1);
-                __m256 _val45 = _mm256_insertf128_ps(_mm256_castps128_ps256(_val4), _val5, 1);
-                __m256 _val67 = _mm256_insertf128_ps(_mm256_castps128_ps256(_val6), _val7, 1);
-
-#if NCNN_IMPL_FP16S
-                __m256i _w0123 = _mm256_lddqu_si256((const __m256i*)kptr);
-                __m256i _w4567 = _mm256_lddqu_si256((const __m256i*)(kptr + 16));
-                __m256 _w01 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w0123, 0));
-                __m256 _w23 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w0123, 1));
-                __m256 _w45 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w4567, 0));
-                __m256 _w67 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w4567, 1));
-#else
-                __m256 _w01 = _mm256_loadu_ps(kptr);
-                __m256 _w23 = _mm256_loadu_ps(kptr + 8);
-                __m256 _w45 = _mm256_loadu_ps(kptr + 16);
-                __m256 _w67 = _mm256_loadu_ps(kptr + 24);
-#endif
-
-                _sum01 = _mm256_comp_fmadd_ps(_val01, _w01, _sum01);
-                _sum23 = _mm256_comp_fmadd_ps(_val23, _w23, _sum23);
-                _sum45 = _mm256_comp_fmadd_ps(_val45, _w45, _sum45);
-                _sum67 = _mm256_comp_fmadd_ps(_val67, _w67, _sum67);
-
-                sptr += 8;
-                kptr += 32;
-            }
-#endif
-            for (; i + 3 < num_input; i += 4)
-            {
-#if __AVX__
-                __m128 _val0 = _mm_broadcast_ss(sptr);
-                __m128 _val1 = _mm_broadcast_ss(sptr + 1);
-                __m128 _val2 = _mm_broadcast_ss(sptr + 2);
-                __m128 _val3 = _mm_broadcast_ss(sptr + 3);
-
-                __m256 _val01 = _mm256_insertf128_ps(_mm256_castps128_ps256(_val0), _val1, 1);
-                __m256 _val23 = _mm256_insertf128_ps(_mm256_castps128_ps256(_val2), _val3, 1);
-
-#if NCNN_IMPL_FP16S
-                __m256i _w0123 = _mm256_lddqu_si256((const __m256i*)kptr);
-                __m256 _w01 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w0123, 0));
-                __m256 _w23 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w0123, 1));
-#else
-                __m256 _w01 = _mm256_loadu_ps(kptr);
-                __m256 _w23 = _mm256_loadu_ps(kptr + 8);
-#endif
-
-                _sum01 = _mm256_comp_fmadd_ps(_val01, _w01, _sum01);
-                _sum23 = _mm256_comp_fmadd_ps(_val23, _w23, _sum23);
-#else
-                __m128 _val0 = _mm_set1_ps(sptr[0]);
-                __m128 _val1 = _mm_set1_ps(sptr[1]);
-                __m128 _val2 = _mm_set1_ps(sptr[2]);
-                __m128 _val3 = _mm_set1_ps(sptr[3]);
-
-                __m128 _w0 = _mm_loadu_ps(kptr);
-                __m128 _w1 = _mm_loadu_ps(kptr + 4);
-                __m128 _w2 = _mm_loadu_ps(kptr + 8);
-                __m128 _w3 = _mm_loadu_ps(kptr + 12);
-
-                _sum0 = _mm_comp_fmadd_ps(_val0, _w0, _sum0);
-                _sum1 = _mm_comp_fmadd_ps(_val1, _w1, _sum1);
-                _sum2 = _mm_comp_fmadd_ps(_val2, _w2, _sum2);
-                _sum3 = _mm_comp_fmadd_ps(_val3, _w3, _sum3);
-#endif
-
-                sptr += 4;
-                kptr += 16;
-            }
-            for (; i < num_input; i++)
-            {
-                __m128 _val = _mm_set1_ps(sptr[0]);
-#if NCNN_IMPL_FP16S
-                __m128 _w = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i*)kptr));
-#else
-                __m128 _w = _mm_loadu_ps(kptr);
-#endif
-                _sum0 = _mm_comp_fmadd_ps(_val, _w, _sum0);
-
-                sptr += 1;
-                kptr += 4;
-            }
-
-#if __AVX__
-            _sum01 = _mm256_add_ps(_sum01, _sum23);
-            _sum45 = _mm256_add_ps(_sum45, _sum67);
-            _sum01 = _mm256_add_ps(_sum01, _sum45);
-
-            _sum0 = _mm_add_ps(_sum0, _mm256_extractf128_ps(_sum01, 0));
-            _sum0 = _mm_add_ps(_sum0, _mm256_extractf128_ps(_sum01, 1));
-#else
-            _sum0 = _mm_add_ps(_sum0, _sum1);
-            _sum2 = _mm_add_ps(_sum2, _sum3);
-            _sum0 = _mm_add_ps(_sum0, _sum2);
-#endif
-
-            _sum0 = activation_sse(_sum0, activation_type, activation_params);
-
-            float* outptr = top_blob;
-            _mm_storeu_ps(outptr + p * 4, _sum0);
-        }
-    }
-#endif // __SSE2__
-
-    if (out_elempack == 1)
-    {
-#if __SSE2__
-#if __AVX__
-        int remain_outw_start = 0;
-        int nn_outw = outw >> 3;
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int pp = 0; pp < nn_outw; pp++)
-        {
-            int p = pp * 8;
-
-            float sums[8] = {0.0f};
-            if (bias_data_ptr)
-            {
-                sums[0] = bias_data_ptr[p];
-                sums[1] = bias_data_ptr[p + 1];
-                sums[2] = bias_data_ptr[p + 2];
-                sums[3] = bias_data_ptr[p + 3];
-                sums[4] = bias_data_ptr[p + 4];
-                sums[5] = bias_data_ptr[p + 5];
-                sums[6] = bias_data_ptr[p + 6];
-                sums[7] = bias_data_ptr[p + 7];
-            }
-
-#if NCNN_IMPL_FP16S
-            const unsigned short* w0 = weight_data_tm.row<const unsigned short>(p);
-            const unsigned short* w1 = weight_data_tm.row<const unsigned short>(p + 1);
-            const unsigned short* w2 = weight_data_tm.row<const unsigned short>(p + 2);
-            const unsigned short* w3 = weight_data_tm.row<const unsigned short>(p + 3);
-            const unsigned short* w4 = weight_data_tm.row<const unsigned short>(p + 4);
-            const unsigned short* w5 = weight_data_tm.row<const unsigned short>(p + 5);
-            const unsigned short* w6 = weight_data_tm.row<const unsigned short>(p + 6);
-            const unsigned short* w7 = weight_data_tm.row<const unsigned short>(p + 7);
-#else
-            const float* w0 = (const float*)weight_data_tm + num_input * p;
-            const float* w1 = (const float*)weight_data_tm + num_input * (p + 1);
-            const float* w2 = (const float*)weight_data_tm + num_input * (p + 2);
-            const float* w3 = (const float*)weight_data_tm + num_input * (p + 3);
-            const float* w4 = (const float*)weight_data_tm + num_input * (p + 4);
-            const float* w5 = (const float*)weight_data_tm + num_input * (p + 5);
-            const float* w6 = (const float*)weight_data_tm + num_input * (p + 6);
-            const float* w7 = (const float*)weight_data_tm + num_input * (p + 7);
-#endif
-            const float* m = bottom_blob;
-
-            __m256 _sum0 = _mm256_setzero_ps();
-            __m256 _sum1 = _mm256_setzero_ps();
-            __m256 _sum2 = _mm256_setzero_ps();
-            __m256 _sum3 = _mm256_setzero_ps();
-            __m256 _sum4 = _mm256_setzero_ps();
-            __m256 _sum5 = _mm256_setzero_ps();
-            __m256 _sum6 = _mm256_setzero_ps();
-            __m256 _sum7 = _mm256_setzero_ps();
-
-            int i = 0;
-            for (; i + 7 < num_input; i += 8)
-            {
-                __m256 _m = _mm256_loadu_ps(m);
-
-#if NCNN_IMPL_FP16S
-                __m256 _w0 = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)w0));
-                __m256 _w1 = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)w1));
-                __m256 _w2 = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)w2));
-                __m256 _w3 = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)w3));
-#else
-                __m256 _w0 = _mm256_loadu_ps(w0);
-                __m256 _w1 = _mm256_loadu_ps(w1);
-                __m256 _w2 = _mm256_loadu_ps(w2);
-                __m256 _w3 = _mm256_loadu_ps(w3);
-#endif
-
-                _sum0 = _mm256_comp_fmadd_ps(_m, _w0, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_m, _w1, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_m, _w2, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_m, _w3, _sum3);
-
-#if NCNN_IMPL_FP16S
-                __m256 _w4 = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)w4));
-                __m256 _w5 = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)w5));
-                __m256 _w6 = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)w6));
-                __m256 _w7 = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)w7));
-#else
-                __m256 _w4 = _mm256_loadu_ps(w4);
-                __m256 _w5 = _mm256_loadu_ps(w5);
-                __m256 _w6 = _mm256_loadu_ps(w6);
-                __m256 _w7 = _mm256_loadu_ps(w7);
-#endif
-
-                _sum4 = _mm256_comp_fmadd_ps(_m, _w4, _sum4);
-                _sum5 = _mm256_comp_fmadd_ps(_m, _w5, _sum5);
-                _sum6 = _mm256_comp_fmadd_ps(_m, _w6, _sum6);
-                _sum7 = _mm256_comp_fmadd_ps(_m, _w7, _sum7);
-
-                m += 8;
-                w0 += 8;
-                w1 += 8;
-                w2 += 8;
-                w3 += 8;
-                w4 += 8;
-                w5 += 8;
-                w6 += 8;
-                w7 += 8;
-            }
-            for (; i < num_input; i++)
-            {
-#if NCNN_IMPL_FP16S
-                sums[0] += *m * float16_to_float32(*w0);
-                sums[1] += *m * float16_to_float32(*w1);
-                sums[2] += *m * float16_to_float32(*w2);
-                sums[3] += *m * float16_to_float32(*w3);
-                sums[4] += *m * float16_to_float32(*w4);
-                sums[5] += *m * float16_to_float32(*w5);
-                sums[6] += *m * float16_to_float32(*w6);
-                sums[7] += *m * float16_to_float32(*w7);
-#else
-                sums[0] += *m * *w0;
-                sums[1] += *m * *w1;
-                sums[2] += *m * *w2;
-                sums[3] += *m * *w3;
-                sums[4] += *m * *w4;
-                sums[5] += *m * *w5;
-                sums[6] += *m * *w6;
-                sums[7] += *m * *w7;
-#endif
-
-                m++;
-                w0++;
-                w1++;
-                w2++;
-                w3++;
-                w4++;
-                w5++;
-                w6++;
-                w7++;
-            }
-
-            __m256 _sums = HorizontalSums(_sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7);
-            __m256 _sums_f = _mm256_loadu_ps(sums);
-            _sums = _mm256_add_ps(_sums_f, _sums);
-            _sums = activation_avx(_sums, activation_type, activation_params);
-
-            float* outptr = top_blob;
-            _mm256_storeu_ps(outptr + p, _sums);
-        }
-
-        remain_outw_start += (nn_outw << 3);
-        nn_outw = (outw - remain_outw_start) >> 2;
-#else
-        int remain_outw_start = 0;
-        int nn_outw = outw >> 2;
-#endif // __AVX__
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int pp = 0; pp < nn_outw; pp++)
-        {
-            int p = remain_outw_start + (pp * 4);
-
-            float sums[4] = {0.0f};
-            if (bias_data_ptr)
-            {
-                sums[0] = bias_data_ptr[p];
-                sums[1] = bias_data_ptr[p + 1];
-                sums[2] = bias_data_ptr[p + 2];
-                sums[3] = bias_data_ptr[p + 3];
-            }
-
-#if NCNN_IMPL_FP16S
-            const unsigned short* w0 = weight_data_tm.row<const unsigned short>(p);
-            const unsigned short* w1 = weight_data_tm.row<const unsigned short>(p + 1);
-            const unsigned short* w2 = weight_data_tm.row<const unsigned short>(p + 2);
-            const unsigned short* w3 = weight_data_tm.row<const unsigned short>(p + 3);
-#else
-            const float* w0 = (const float*)weight_data_tm + num_input * p;
-            const float* w1 = (const float*)weight_data_tm + num_input * (p + 1);
-            const float* w2 = (const float*)weight_data_tm + num_input * (p + 2);
-            const float* w3 = (const float*)weight_data_tm + num_input * (p + 3);
-#endif
-            const float* m = bottom_blob;
-
-            int i = 0;
-#if __AVX__
-            __m256 _sum0 = _mm256_setzero_ps();
-            __m256 _sum1 = _mm256_setzero_ps();
-            __m256 _sum2 = _mm256_setzero_ps();
-            __m256 _sum3 = _mm256_setzero_ps();
-            for (; i + 7 < num_input; i += 8)
-            {
-                __m256 _m = _mm256_loadu_ps(m);
-
-#if NCNN_IMPL_FP16S
-                __m256 _w0 = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)w0));
-                __m256 _w1 = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)w1));
-                __m256 _w2 = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)w2));
-                __m256 _w3 = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)w3));
-#else
-                __m256 _w0 = _mm256_loadu_ps(w0);
-                __m256 _w1 = _mm256_loadu_ps(w1);
-                __m256 _w2 = _mm256_loadu_ps(w2);
-                __m256 _w3 = _mm256_loadu_ps(w3);
-#endif
-
-                _sum0 = _mm256_comp_fmadd_ps(_m, _w0, _sum0);
-                _sum1 = _mm256_comp_fmadd_ps(_m, _w1, _sum1);
-                _sum2 = _mm256_comp_fmadd_ps(_m, _w2, _sum2);
-                _sum3 = _mm256_comp_fmadd_ps(_m, _w3, _sum3);
-
-                m += 8;
-                w0 += 8;
-                w1 += 8;
-                w2 += 8;
-                w3 += 8;
-            }
-#endif // __AVX__
-
-            __m128 _sum0l = _mm_setzero_ps();
-            __m128 _sum1l = _mm_setzero_ps();
-            __m128 _sum2l = _mm_setzero_ps();
-            __m128 _sum3l = _mm_setzero_ps();
-            for (; i + 3 < num_input; i += 4)
-            {
-                __m128 _m = _mm_loadu_ps(m);
-
-#if NCNN_IMPL_FP16S
-                __m128 _w0 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i*)w0));
-                __m128 _w1 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i*)w1));
-                __m128 _w2 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i*)w2));
-                __m128 _w3 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i*)w3));
-#else
-                __m128 _w0 = _mm_loadu_ps(w0);
-                __m128 _w1 = _mm_loadu_ps(w1);
-                __m128 _w2 = _mm_loadu_ps(w2);
-                __m128 _w3 = _mm_loadu_ps(w3);
-#endif
-
-                _sum0l = _mm_comp_fmadd_ps(_m, _w0, _sum0l);
-                _sum1l = _mm_comp_fmadd_ps(_m, _w1, _sum1l);
-                _sum2l = _mm_comp_fmadd_ps(_m, _w2, _sum2l);
-                _sum3l = _mm_comp_fmadd_ps(_m, _w3, _sum3l);
-
-                m += 4;
-                w0 += 4;
-                w1 += 4;
-                w2 += 4;
-                w3 += 4;
-            }
-            for (; i < num_input; i++)
-            {
-#if NCNN_IMPL_FP16S
-                sums[0] += *m * float16_to_float32(*w0);
-                sums[1] += *m * float16_to_float32(*w1);
-                sums[2] += *m * float16_to_float32(*w2);
-                sums[3] += *m * float16_to_float32(*w3);
-#else
-                sums[0] += *m * *w0;
-                sums[1] += *m * *w1;
-                sums[2] += *m * *w2;
-                sums[3] += *m * *w3;
-#endif
-
-                m++;
-                w0++;
-                w1++;
-                w2++;
-                w3++;
-            }
-
-            __m128 _sums = _mm_loadu_ps(sums);
-#if __AVX__
-            _sums = _mm_add_ps(HorizontalSums(_sum0, _sum1, _sum2, _sum3), _sums);
-#endif
-            _MM_TRANSPOSE4_PS(_sum0l, _sum1l, _sum2l, _sum3l);
-            _sums = _mm_add_ps(_sum0l, _sums);
-            _sums = _mm_add_ps(_sum1l, _sums);
-            _sums = _mm_add_ps(_sum2l, _sums);
-            _sums = _mm_add_ps(_sum3l, _sums);
-            _sums = activation_sse(_sums, activation_type, activation_params);
-
-            float* outptr = top_blob;
-            _mm_storeu_ps(outptr + p, _sums);
-        }
-
-        remain_outw_start += (nn_outw << 2);
-#else
-        int remain_outw_start = 0;
-#endif // __SSE2__
-
-        #pragma omp parallel for num_threads(opt.num_threads)
-        for (int p = remain_outw_start; p < outw; p++)
-        {
-            float sum = 0.f;
-
-            if (bias_data_ptr)
-                sum = bias_data_ptr[p];
-
-#if NCNN_IMPL_FP16S
-            const unsigned short* w = weight_data_tm.row<const unsigned short>(p);
-#else
-            const float* w = (const float*)weight_data_tm + num_input * p;
-#endif
-            const float* m = bottom_blob;
-
-            int i = 0;
-#if __SSE2__
-#if __AVX__
-            __m256 _sum = _mm256_setzero_ps();
-            for (; i + 7 < num_input; i += 8)
-            {
-                __m256 _m = _mm256_loadu_ps(m);
-#if NCNN_IMPL_FP16S
-                __m256 _w = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)w));
-#else
-                __m256 _w = _mm256_loadu_ps(w);
-#endif
-                _sum = _mm256_comp_fmadd_ps(_m, _w, _sum);
-
-                m += 8;
-                w += 8;
-            }
-#endif // __AVX__
-            __m128 _suml = _mm_setzero_ps();
-            for (; i + 3 < num_input; i += 4)
-            {
-                __m128 _m = _mm_loadu_ps(m);
-#if NCNN_IMPL_FP16S
-                __m128 _w = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i*)w));
-#else
-                __m128 _w = _mm_loadu_ps(w);
-#endif
-                _suml = _mm_comp_fmadd_ps(_m, _w, _suml);
-
-                m += 4;
-                w += 4;
-            }
-#endif // __SSE2__
-            for (; i < num_input; i++)
-            {
-#if NCNN_IMPL_FP16S
-                sum += *m * float16_to_float32(*w);
-#else
-                sum += *m * *w;
-#endif
-                m++;
-                w++;
-            }
-
-#if __SSE2__
-#if __AVX__
-            _suml = _mm_add_ps(_suml, _mm256_extractf128_ps(_sum, 1));
-            _suml = _mm_add_ps(_suml, _mm256_castps256_ps128(_sum));
-#endif // __AVX__
-            sum += _mm_reduce_add_ps(_suml);
-#endif // __SSE2__
-
-            sum = activation_ss(sum, activation_type, activation_params);
-
-            float* outptr = top_blob;
-            outptr[p] = sum;
-        }
-    }
-#endif // NCNN_RUNTIME_CPU
-}
-
-#if NCNN_IMPL_FP16S
-static void innerproduct_transform_kernel_fp16s_sse(const Mat& weight_data, Mat& weight_data_tm, int num_input, int num_output, const Option& opt)
-#else
-static void innerproduct_transform_kernel_sse(const Mat& weight_data, Mat& weight_data_tm, int num_input, int num_output, const Option& opt)
-#endif
-{
-#if NCNN_RUNTIME_CPU && NCNN_IMPL_FP16S && NCNN_F16C && __AVX__ && !__F16C__
-    if (ncnn::cpu_support_x86_f16c())
-    {
-        innerproduct_transform_kernel_fp16s_sse_f16c(weight_data, weight_data_tm, num_input, num_output, opt);
-        return;
-    }
-#else // NCNN_RUNTIME_CPU
-
-    int out_elempack = 1;
-#if __SSE2__
-    if (opt.use_packing_layout)
-    {
-#if __AVX512F__
-        out_elempack = num_output % 16 == 0 ? 16 : num_output % 8 == 0 ? 8 : num_output % 4 == 0 ? 4 : 1;
-#elif __AVX__
-        out_elempack = num_output % 8 == 0 ? 8 : num_output % 4 == 0 ? 4 : 1;
-#else
-        out_elempack = num_output % 4 == 0 ? 4 : 1;
-#endif
-    }
-#endif // __SSE2__
-
-    // src = inch-outch
-    // dst = pb-inch-outch/pb
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    if (out_elempack == 16)
-    {
-        Mat weight_data_r2 = weight_data.reshape(num_input, num_output);
-
-#if NCNN_IMPL_FP16S
-        weight_data_tm.create(num_input, num_output / 16, (size_t)32u, 16);
-#else
-        weight_data_tm.create(num_input, num_output / 16, (size_t)64u, 16);
-#endif
-
-        for (int q = 0; q + 15 < num_output; q += 16)
-        {
-#if NCNN_IMPL_FP16S
-            unsigned short* g0 = weight_data_tm.row<unsigned short>(q / 16);
-#else
-            float* g0 = weight_data_tm.row(q / 16);
-#endif
-
-            const float* k0 = weight_data_r2.row(q);
-            const float* k1 = weight_data_r2.row(q + 1);
-            const float* k2 = weight_data_r2.row(q + 2);
-            const float* k3 = weight_data_r2.row(q + 3);
-            const float* k4 = weight_data_r2.row(q + 4);
-            const float* k5 = weight_data_r2.row(q + 5);
-            const float* k6 = weight_data_r2.row(q + 6);
-            const float* k7 = weight_data_r2.row(q + 7);
-            const float* k8 = weight_data_r2.row(q + 8);
-            const float* k9 = weight_data_r2.row(q + 9);
-            const float* ka = weight_data_r2.row(q + 10);
-            const float* kb = weight_data_r2.row(q + 11);
-            const float* kc = weight_data_r2.row(q + 12);
-            const float* kd = weight_data_r2.row(q + 13);
-            const float* ke = weight_data_r2.row(q + 14);
-            const float* kf = weight_data_r2.row(q + 15);
-
-            int p = 0;
-            for (; p + 15 < num_input; p += 16)
-            {
-                // transpose 16x16
-#if NCNN_IMPL_FP16S
-                __m256i _r0 = _mm512_cvtps_ph(_mm512_loadu_ps(k0), _MM_FROUND_TRUNC);
-                __m256i _r1 = _mm512_cvtps_ph(_mm512_loadu_ps(k1), _MM_FROUND_TRUNC);
-                __m256i _r2 = _mm512_cvtps_ph(_mm512_loadu_ps(k2), _MM_FROUND_TRUNC);
-                __m256i _r3 = _mm512_cvtps_ph(_mm512_loadu_ps(k3), _MM_FROUND_TRUNC);
-                __m256i _r4 = _mm512_cvtps_ph(_mm512_loadu_ps(k4), _MM_FROUND_TRUNC);
-                __m256i _r5 = _mm512_cvtps_ph(_mm512_loadu_ps(k5), _MM_FROUND_TRUNC);
-                __m256i _r6 = _mm512_cvtps_ph(_mm512_loadu_ps(k6), _MM_FROUND_TRUNC);
-                __m256i _r7 = _mm512_cvtps_ph(_mm512_loadu_ps(k7), _MM_FROUND_TRUNC);
-                __m256i _r8 = _mm512_cvtps_ph(_mm512_loadu_ps(k8), _MM_FROUND_TRUNC);
-                __m256i _r9 = _mm512_cvtps_ph(_mm512_loadu_ps(k9), _MM_FROUND_TRUNC);
-                __m256i _ra = _mm512_cvtps_ph(_mm512_loadu_ps(ka), _MM_FROUND_TRUNC);
-                __m256i _rb = _mm512_cvtps_ph(_mm512_loadu_ps(kb), _MM_FROUND_TRUNC);
-                __m256i _rc = _mm512_cvtps_ph(_mm512_loadu_ps(kc), _MM_FROUND_TRUNC);
-                __m256i _rd = _mm512_cvtps_ph(_mm512_loadu_ps(kd), _MM_FROUND_TRUNC);
-                __m256i _re = _mm512_cvtps_ph(_mm512_loadu_ps(ke), _MM_FROUND_TRUNC);
-                __m256i _rf = _mm512_cvtps_ph(_mm512_loadu_ps(kf), _MM_FROUND_TRUNC);
-
-                transpose16x16_epi16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7, _r8, _r9, _ra, _rb, _rc, _rd, _re, _rf);
-
-                _mm256_storeu_si256((__m256i*)g0, _r0);
-                _mm256_storeu_si256((__m256i*)(g0 + 16), _r1);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 2), _r2);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 3), _r3);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 4), _r4);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 5), _r5);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 6), _r6);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 7), _r7);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 8), _r8);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 9), _r9);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 10), _ra);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 11), _rb);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 12), _rc);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 13), _rd);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 14), _re);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 15), _rf);
-#else
-                __m512 _r0 = _mm512_loadu_ps(k0);
-                __m512 _r1 = _mm512_loadu_ps(k1);
-                __m512 _r2 = _mm512_loadu_ps(k2);
-                __m512 _r3 = _mm512_loadu_ps(k3);
-                __m512 _r4 = _mm512_loadu_ps(k4);
-                __m512 _r5 = _mm512_loadu_ps(k5);
-                __m512 _r6 = _mm512_loadu_ps(k6);
-                __m512 _r7 = _mm512_loadu_ps(k7);
-                __m512 _r8 = _mm512_loadu_ps(k8);
-                __m512 _r9 = _mm512_loadu_ps(k9);
-                __m512 _ra = _mm512_loadu_ps(ka);
-                __m512 _rb = _mm512_loadu_ps(kb);
-                __m512 _rc = _mm512_loadu_ps(kc);
-                __m512 _rd = _mm512_loadu_ps(kd);
-                __m512 _re = _mm512_loadu_ps(ke);
-                __m512 _rf = _mm512_loadu_ps(kf);
-
-                transpose16x16_ps(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7, _r8, _r9, _ra, _rb, _rc, _rd, _re, _rf);
-
-                _mm512_storeu_ps(g0, _r0);
-                _mm512_storeu_ps(g0 + 16, _r1);
-                _mm512_storeu_ps(g0 + 16 * 2, _r2);
-                _mm512_storeu_ps(g0 + 16 * 3, _r3);
-                _mm512_storeu_ps(g0 + 16 * 4, _r4);
-                _mm512_storeu_ps(g0 + 16 * 5, _r5);
-                _mm512_storeu_ps(g0 + 16 * 6, _r6);
-                _mm512_storeu_ps(g0 + 16 * 7, _r7);
-                _mm512_storeu_ps(g0 + 16 * 8, _r8);
-                _mm512_storeu_ps(g0 + 16 * 9, _r9);
-                _mm512_storeu_ps(g0 + 16 * 10, _ra);
-                _mm512_storeu_ps(g0 + 16 * 11, _rb);
-                _mm512_storeu_ps(g0 + 16 * 12, _rc);
-                _mm512_storeu_ps(g0 + 16 * 13, _rd);
-                _mm512_storeu_ps(g0 + 16 * 14, _re);
-                _mm512_storeu_ps(g0 + 16 * 15, _rf);
-#endif
-
-                k0 += 16;
-                k1 += 16;
-                k2 += 16;
-                k3 += 16;
-                k4 += 16;
-                k5 += 16;
-                k6 += 16;
-                k7 += 16;
-                k8 += 16;
-                k9 += 16;
-                ka += 16;
-                kb += 16;
-                kc += 16;
-                kd += 16;
-                ke += 16;
-                kf += 16;
-                g0 += 256;
-            }
-            for (; p + 7 < num_input; p += 8)
-            {
-                // transpose 8x16
-#if NCNN_IMPL_FP16S
-                __m128i _r0 = _mm256_cvtps_ph(_mm256_loadu_ps(k0), _MM_FROUND_TRUNC);
-                __m128i _r1 = _mm256_cvtps_ph(_mm256_loadu_ps(k1), _MM_FROUND_TRUNC);
-                __m128i _r2 = _mm256_cvtps_ph(_mm256_loadu_ps(k2), _MM_FROUND_TRUNC);
-                __m128i _r3 = _mm256_cvtps_ph(_mm256_loadu_ps(k3), _MM_FROUND_TRUNC);
-                __m128i _r4 = _mm256_cvtps_ph(_mm256_loadu_ps(k4), _MM_FROUND_TRUNC);
-                __m128i _r5 = _mm256_cvtps_ph(_mm256_loadu_ps(k5), _MM_FROUND_TRUNC);
-                __m128i _r6 = _mm256_cvtps_ph(_mm256_loadu_ps(k6), _MM_FROUND_TRUNC);
-                __m128i _r7 = _mm256_cvtps_ph(_mm256_loadu_ps(k7), _MM_FROUND_TRUNC);
-                __m128i _r8 = _mm256_cvtps_ph(_mm256_loadu_ps(k8), _MM_FROUND_TRUNC);
-                __m128i _r9 = _mm256_cvtps_ph(_mm256_loadu_ps(k9), _MM_FROUND_TRUNC);
-                __m128i _ra = _mm256_cvtps_ph(_mm256_loadu_ps(ka), _MM_FROUND_TRUNC);
-                __m128i _rb = _mm256_cvtps_ph(_mm256_loadu_ps(kb), _MM_FROUND_TRUNC);
-                __m128i _rc = _mm256_cvtps_ph(_mm256_loadu_ps(kc), _MM_FROUND_TRUNC);
-                __m128i _rd = _mm256_cvtps_ph(_mm256_loadu_ps(kd), _MM_FROUND_TRUNC);
-                __m128i _re = _mm256_cvtps_ph(_mm256_loadu_ps(ke), _MM_FROUND_TRUNC);
-                __m128i _rf = _mm256_cvtps_ph(_mm256_loadu_ps(kf), _MM_FROUND_TRUNC);
-
-                __m256i _r08 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r0), _r8, 1);
-                __m256i _r19 = _mm256_inserti128_si256(_mm256_castsi128_si256(_r1), _r9, 1);
-                __m256i _r2a = _mm256_inserti128_si256(_mm256_castsi128_si256(_r2), _ra, 1);
-                __m256i _r3b = _mm256_inserti128_si256(_mm256_castsi128_si256(_r3), _rb, 1);
-                __m256i _r4c = _mm256_inserti128_si256(_mm256_castsi128_si256(_r4), _rc, 1);
-                __m256i _r5d = _mm256_inserti128_si256(_mm256_castsi128_si256(_r5), _rd, 1);
-                __m256i _r6e = _mm256_inserti128_si256(_mm256_castsi128_si256(_r6), _re, 1);
-                __m256i _r7f = _mm256_inserti128_si256(_mm256_castsi128_si256(_r7), _rf, 1);
-
-                __m256i _tmp0 = _mm256_unpacklo_epi16(_r08, _r19);
-                __m256i _tmp1 = _mm256_unpackhi_epi16(_r08, _r19);
-                __m256i _tmp2 = _mm256_unpacklo_epi16(_r2a, _r3b);
-                __m256i _tmp3 = _mm256_unpackhi_epi16(_r2a, _r3b);
-                __m256i _tmp4 = _mm256_unpacklo_epi16(_r4c, _r5d);
-                __m256i _tmp5 = _mm256_unpackhi_epi16(_r4c, _r5d);
-                __m256i _tmp6 = _mm256_unpacklo_epi16(_r6e, _r7f);
-                __m256i _tmp7 = _mm256_unpackhi_epi16(_r6e, _r7f);
-
-                __m256i _tmpg = _mm256_unpacklo_epi32(_tmp0, _tmp2);
-                __m256i _tmph = _mm256_unpackhi_epi32(_tmp0, _tmp2);
-                __m256i _tmpi = _mm256_unpacklo_epi32(_tmp1, _tmp3);
-                __m256i _tmpj = _mm256_unpackhi_epi32(_tmp1, _tmp3);
-                __m256i _tmpk = _mm256_unpacklo_epi32(_tmp4, _tmp6);
-                __m256i _tmpl = _mm256_unpackhi_epi32(_tmp4, _tmp6);
-                __m256i _tmpm = _mm256_unpacklo_epi32(_tmp5, _tmp7);
-                __m256i _tmpn = _mm256_unpackhi_epi32(_tmp5, _tmp7);
-
-                _r08 = _mm256_unpacklo_epi64(_tmpg, _tmpk);
-                _r19 = _mm256_unpackhi_epi64(_tmpg, _tmpk);
-                _r2a = _mm256_unpacklo_epi64(_tmph, _tmpl);
-                _r3b = _mm256_unpackhi_epi64(_tmph, _tmpl);
-                _r4c = _mm256_unpacklo_epi64(_tmpi, _tmpm);
-                _r5d = _mm256_unpackhi_epi64(_tmpi, _tmpm);
-                _r6e = _mm256_unpacklo_epi64(_tmpj, _tmpn);
-                _r7f = _mm256_unpackhi_epi64(_tmpj, _tmpn);
-
-                _mm256_storeu_si256((__m256i*)g0, _r08);
-                _mm256_storeu_si256((__m256i*)(g0 + 16), _r19);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 2), _r2a);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 3), _r3b);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 4), _r4c);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 5), _r5d);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 6), _r6e);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 7), _r7f);
-#else
-                __m256 _r0 = _mm256_loadu_ps(k0);
-                __m256 _r1 = _mm256_loadu_ps(k1);
-                __m256 _r2 = _mm256_loadu_ps(k2);
-                __m256 _r3 = _mm256_loadu_ps(k3);
-                __m256 _r4 = _mm256_loadu_ps(k4);
-                __m256 _r5 = _mm256_loadu_ps(k5);
-                __m256 _r6 = _mm256_loadu_ps(k6);
-                __m256 _r7 = _mm256_loadu_ps(k7);
-                __m256 _r8 = _mm256_loadu_ps(k8);
-                __m256 _r9 = _mm256_loadu_ps(k9);
-                __m256 _ra = _mm256_loadu_ps(ka);
-                __m256 _rb = _mm256_loadu_ps(kb);
-                __m256 _rc = _mm256_loadu_ps(kc);
-                __m256 _rd = _mm256_loadu_ps(kd);
-                __m256 _re = _mm256_loadu_ps(ke);
-                __m256 _rf = _mm256_loadu_ps(kf);
-
-                transpose8x16_ps(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7, _r8, _r9, _ra, _rb, _rc, _rd, _re, _rf);
-
-                _mm256_storeu_ps(g0, _r0);
-                _mm256_storeu_ps(g0 + 8, _r1);
-                _mm256_storeu_ps(g0 + 8 * 2, _r2);
-                _mm256_storeu_ps(g0 + 8 * 3, _r3);
-                _mm256_storeu_ps(g0 + 8 * 4, _r4);
-                _mm256_storeu_ps(g0 + 8 * 5, _r5);
-                _mm256_storeu_ps(g0 + 8 * 6, _r6);
-                _mm256_storeu_ps(g0 + 8 * 7, _r7);
-                _mm256_storeu_ps(g0 + 8 * 8, _r8);
-                _mm256_storeu_ps(g0 + 8 * 9, _r9);
-                _mm256_storeu_ps(g0 + 8 * 10, _ra);
-                _mm256_storeu_ps(g0 + 8 * 11, _rb);
-                _mm256_storeu_ps(g0 + 8 * 12, _rc);
-                _mm256_storeu_ps(g0 + 8 * 13, _rd);
-                _mm256_storeu_ps(g0 + 8 * 14, _re);
-                _mm256_storeu_ps(g0 + 8 * 15, _rf);
-#endif
-
-                k0 += 8;
-                k1 += 8;
-                k2 += 8;
-                k3 += 8;
-                k4 += 8;
-                k5 += 8;
-                k6 += 8;
-                k7 += 8;
-                k8 += 8;
-                k9 += 8;
-                ka += 8;
-                kb += 8;
-                kc += 8;
-                kd += 8;
-                ke += 8;
-                kf += 8;
-                g0 += 128;
-            }
-            for (; p < num_input; p++)
-            {
-#if NCNN_IMPL_FP16S
-                g0[0] = float32_to_float16(*k0++);
-                g0[1] = float32_to_float16(*k1++);
-                g0[2] = float32_to_float16(*k2++);
-                g0[3] = float32_to_float16(*k3++);
-                g0[4] = float32_to_float16(*k4++);
-                g0[5] = float32_to_float16(*k5++);
-                g0[6] = float32_to_float16(*k6++);
-                g0[7] = float32_to_float16(*k7++);
-                g0[8] = float32_to_float16(*k8++);
-                g0[9] = float32_to_float16(*k9++);
-                g0[10] = float32_to_float16(*ka++);
-                g0[11] = float32_to_float16(*kb++);
-                g0[12] = float32_to_float16(*kc++);
-                g0[13] = float32_to_float16(*kd++);
-                g0[14] = float32_to_float16(*ke++);
-                g0[15] = float32_to_float16(*kf++);
-#else
-                g0[0] = *k0++;
-                g0[1] = *k1++;
-                g0[2] = *k2++;
-                g0[3] = *k3++;
-                g0[4] = *k4++;
-                g0[5] = *k5++;
-                g0[6] = *k6++;
-                g0[7] = *k7++;
-                g0[8] = *k8++;
-                g0[9] = *k9++;
-                g0[10] = *ka++;
-                g0[11] = *kb++;
-                g0[12] = *kc++;
-                g0[13] = *kd++;
-                g0[14] = *ke++;
-                g0[15] = *kf++;
-#endif
-                g0 += 16;
-            }
-        }
-    }
-#endif // __AVX512F__
-
-    if (out_elempack == 8)
-    {
-        Mat weight_data_r2 = weight_data.reshape(num_input, num_output);
-
-#if NCNN_IMPL_FP16S
-        weight_data_tm.create(num_input, num_output / 8, (size_t)16u, 8);
-#else
-        weight_data_tm.create(num_input, num_output / 8, (size_t)32u, 8);
-#endif
-
-        for (int q = 0; q + 7 < num_output; q += 8)
-        {
-#if NCNN_IMPL_FP16S
-            unsigned short* g0 = weight_data_tm.row<unsigned short>(q / 8);
-#else
-            float* g0 = weight_data_tm.row(q / 8);
-#endif
-
-            const float* k0 = weight_data_r2.row(q);
-            const float* k1 = weight_data_r2.row(q + 1);
-            const float* k2 = weight_data_r2.row(q + 2);
-            const float* k3 = weight_data_r2.row(q + 3);
-            const float* k4 = weight_data_r2.row(q + 4);
-            const float* k5 = weight_data_r2.row(q + 5);
-            const float* k6 = weight_data_r2.row(q + 6);
-            const float* k7 = weight_data_r2.row(q + 7);
-
-            int p = 0;
-#if __AVX512F__
-            for (; p + 15 < num_input; p += 16)
-            {
-                // transpose 16x8
-#if NCNN_IMPL_FP16S
-                __m256i _r0 = _mm512_cvtps_ph(_mm512_loadu_ps(k0), _MM_FROUND_TRUNC);
-                __m256i _r1 = _mm512_cvtps_ph(_mm512_loadu_ps(k1), _MM_FROUND_TRUNC);
-                __m256i _r2 = _mm512_cvtps_ph(_mm512_loadu_ps(k2), _MM_FROUND_TRUNC);
-                __m256i _r3 = _mm512_cvtps_ph(_mm512_loadu_ps(k3), _MM_FROUND_TRUNC);
-                __m256i _r4 = _mm512_cvtps_ph(_mm512_loadu_ps(k4), _MM_FROUND_TRUNC);
-                __m256i _r5 = _mm512_cvtps_ph(_mm512_loadu_ps(k5), _MM_FROUND_TRUNC);
-                __m256i _r6 = _mm512_cvtps_ph(_mm512_loadu_ps(k6), _MM_FROUND_TRUNC);
-                __m256i _r7 = _mm512_cvtps_ph(_mm512_loadu_ps(k7), _MM_FROUND_TRUNC);
-
-                transpose16x8_epi16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-
-                _mm256_storeu_si256((__m256i*)g0, _r0);
-                _mm256_storeu_si256((__m256i*)(g0 + 16), _r1);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 2), _r2);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 3), _r3);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 4), _r4);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 5), _r5);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 6), _r6);
-                _mm256_storeu_si256((__m256i*)(g0 + 16 * 7), _r7);
-#else
-                __m512 _r0 = _mm512_loadu_ps(k0);
-                __m512 _r1 = _mm512_loadu_ps(k1);
-                __m512 _r2 = _mm512_loadu_ps(k2);
-                __m512 _r3 = _mm512_loadu_ps(k3);
-                __m512 _r4 = _mm512_loadu_ps(k4);
-                __m512 _r5 = _mm512_loadu_ps(k5);
-                __m512 _r6 = _mm512_loadu_ps(k6);
-                __m512 _r7 = _mm512_loadu_ps(k7);
-
-                transpose16x8_ps(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-
-                _mm512_storeu_ps(g0, _r0);
-                _mm512_storeu_ps(g0 + 16, _r1);
-                _mm512_storeu_ps(g0 + 16 * 2, _r2);
-                _mm512_storeu_ps(g0 + 16 * 3, _r3);
-                _mm512_storeu_ps(g0 + 16 * 4, _r4);
-                _mm512_storeu_ps(g0 + 16 * 5, _r5);
-                _mm512_storeu_ps(g0 + 16 * 6, _r6);
-                _mm512_storeu_ps(g0 + 16 * 7, _r7);
-#endif
-
-                k0 += 16;
-                k1 += 16;
-                k2 += 16;
-                k3 += 16;
-                k4 += 16;
-                k5 += 16;
-                k6 += 16;
-                k7 += 16;
-                g0 += 128;
-            }
-#endif // __AVX512F__
-            for (; p + 7 < num_input; p += 8)
-            {
-                // transpose 8x8
-#if NCNN_IMPL_FP16S
-                __m128i _r0 = _mm256_cvtps_ph(_mm256_loadu_ps(k0), _MM_FROUND_TRUNC);
-                __m128i _r1 = _mm256_cvtps_ph(_mm256_loadu_ps(k1), _MM_FROUND_TRUNC);
-                __m128i _r2 = _mm256_cvtps_ph(_mm256_loadu_ps(k2), _MM_FROUND_TRUNC);
-                __m128i _r3 = _mm256_cvtps_ph(_mm256_loadu_ps(k3), _MM_FROUND_TRUNC);
-                __m128i _r4 = _mm256_cvtps_ph(_mm256_loadu_ps(k4), _MM_FROUND_TRUNC);
-                __m128i _r5 = _mm256_cvtps_ph(_mm256_loadu_ps(k5), _MM_FROUND_TRUNC);
-                __m128i _r6 = _mm256_cvtps_ph(_mm256_loadu_ps(k6), _MM_FROUND_TRUNC);
-                __m128i _r7 = _mm256_cvtps_ph(_mm256_loadu_ps(k7), _MM_FROUND_TRUNC);
-
-                transpose8x8_epi16(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-
-                _mm_storeu_si128((__m128i*)g0, _r0);
-                _mm_storeu_si128((__m128i*)(g0 + 8), _r1);
-                _mm_storeu_si128((__m128i*)(g0 + 16), _r2);
-                _mm_storeu_si128((__m128i*)(g0 + 24), _r3);
-                _mm_storeu_si128((__m128i*)(g0 + 32), _r4);
-                _mm_storeu_si128((__m128i*)(g0 + 40), _r5);
-                _mm_storeu_si128((__m128i*)(g0 + 48), _r6);
-                _mm_storeu_si128((__m128i*)(g0 + 56), _r7);
-#else
-                __m256 _r0 = _mm256_loadu_ps(k0);
-                __m256 _r1 = _mm256_loadu_ps(k1);
-                __m256 _r2 = _mm256_loadu_ps(k2);
-                __m256 _r3 = _mm256_loadu_ps(k3);
-                __m256 _r4 = _mm256_loadu_ps(k4);
-                __m256 _r5 = _mm256_loadu_ps(k5);
-                __m256 _r6 = _mm256_loadu_ps(k6);
-                __m256 _r7 = _mm256_loadu_ps(k7);
-
-                transpose8x8_ps(_r0, _r1, _r2, _r3, _r4, _r5, _r6, _r7);
-
-                _mm256_storeu_ps(g0, _r0);
-                _mm256_storeu_ps(g0 + 8, _r1);
-                _mm256_storeu_ps(g0 + 16, _r2);
-                _mm256_storeu_ps(g0 + 24, _r3);
-                _mm256_storeu_ps(g0 + 32, _r4);
-                _mm256_storeu_ps(g0 + 40, _r5);
-                _mm256_storeu_ps(g0 + 48, _r6);
-                _mm256_storeu_ps(g0 + 56, _r7);
-#endif
-
-                k0 += 8;
-                k1 += 8;
-                k2 += 8;
-                k3 += 8;
-                k4 += 8;
-                k5 += 8;
-                k6 += 8;
-                k7 += 8;
-                g0 += 64;
-            }
-            for (; p < num_input; p++)
-            {
-#if NCNN_IMPL_FP16S
-                g0[0] = float32_to_float16(*k0++);
-                g0[1] = float32_to_float16(*k1++);
-                g0[2] = float32_to_float16(*k2++);
-                g0[3] = float32_to_float16(*k3++);
-                g0[4] = float32_to_float16(*k4++);
-                g0[5] = float32_to_float16(*k5++);
-                g0[6] = float32_to_float16(*k6++);
-                g0[7] = float32_to_float16(*k7++);
-#else
-                g0[0] = *k0++;
-                g0[1] = *k1++;
-                g0[2] = *k2++;
-                g0[3] = *k3++;
-                g0[4] = *k4++;
-                g0[5] = *k5++;
-                g0[6] = *k6++;
-                g0[7] = *k7++;
-#endif
-                g0 += 8;
-            }
-        }
-    }
-#endif // __AVX__
-
-    if (out_elempack == 4)
-    {
-        Mat weight_data_r2 = weight_data.reshape(num_input, num_output);
-
-#if NCNN_IMPL_FP16S
-        weight_data_tm.create(num_input, num_output / 4, (size_t)8u, 4);
-#else
-        weight_data_tm.create(num_input, num_output / 4, (size_t)16u, 4);
-#endif
-
-        for (int q = 0; q + 3 < num_output; q += 4)
-        {
-#if NCNN_IMPL_FP16S
-            unsigned short* g0 = weight_data_tm.row<unsigned short>(q / 4);
-#else
-            float* g0 = weight_data_tm.row(q / 4);
-#endif
-
-            const float* k0 = weight_data_r2.row(q);
-            const float* k1 = weight_data_r2.row(q + 1);
-            const float* k2 = weight_data_r2.row(q + 2);
-            const float* k3 = weight_data_r2.row(q + 3);
-
-            int p = 0;
-            for (; p + 3 < num_input; p += 4)
-            {
-                // transpose 4x4
-                __m128 _r0 = _mm_loadu_ps(k0);
-                __m128 _r1 = _mm_loadu_ps(k1);
-                __m128 _r2 = _mm_loadu_ps(k2);
-                __m128 _r3 = _mm_loadu_ps(k3);
-                _MM_TRANSPOSE4_PS(_r0, _r1, _r2, _r3);
-#if NCNN_IMPL_FP16S
-                __m256 _r01 = _mm256_insertf128_ps(_mm256_castps128_ps256(_r0), _r1, 1);
-                __m256 _r23 = _mm256_insertf128_ps(_mm256_castps128_ps256(_r2), _r3, 1);
-                __m128i _r01_fp16 = _mm256_cvtps_ph(_r01, _MM_FROUND_TRUNC);
-                __m128i _r23_fp16 = _mm256_cvtps_ph(_r23, _MM_FROUND_TRUNC);
-                _mm_storeu_si128((__m128i*)g0, _r01_fp16);
-                _mm_storeu_si128((__m128i*)(g0 + 8), _r23_fp16);
-#else
-                _mm_storeu_ps(g0, _r0);
-                _mm_storeu_ps(g0 + 4, _r1);
-                _mm_storeu_ps(g0 + 8, _r2);
-                _mm_storeu_ps(g0 + 12, _r3);
-#endif
-
-                k0 += 4;
-                k1 += 4;
-                k2 += 4;
-                k3 += 4;
-                g0 += 16;
-            }
-            for (; p < num_input; p++)
-            {
-#if NCNN_IMPL_FP16S
-                g0[0] = float32_to_float16(*k0++);
-                g0[1] = float32_to_float16(*k1++);
-                g0[2] = float32_to_float16(*k2++);
-                g0[3] = float32_to_float16(*k3++);
-#else
-                g0[0] = *k0++;
-                g0[1] = *k1++;
-                g0[2] = *k2++;
-                g0[3] = *k3++;
-#endif
-                g0 += 4;
-            }
-        }
-    }
-#endif // __SSE2__
-
-    if (out_elempack == 1)
-    {
-#if NCNN_IMPL_FP16S
-        Mat weight_data_r2 = weight_data.reshape(num_input, num_output);
-        ncnn::cast_float32_to_float16(weight_data_r2, weight_data_tm, opt);
-#else
-        weight_data_tm = weight_data;
-#endif
-    }
-#endif // NCNN_RUNTIME_CPU
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/innerproduct_gemm_fp.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/innerproduct_gemm_fp.h
deleted file mode 100644
index 68a1a37d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/innerproduct_gemm_fp.h
+++ /dev/null
@@ -1,1285 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#if NCNN_RUNTIME_CPU && NCNN_F16C && __AVX__ && !__F16C__
-void innerproduct_gemm_fp16s_sse_f16c(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt);
-#endif
-
-#if NCNN_IMPL_FP16S
-static void innerproduct_gemm_fp16s_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt)
-#else
-static void innerproduct_gemm_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_tm, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt)
-#endif
-{
-#if NCNN_RUNTIME_CPU && NCNN_IMPL_FP16S && NCNN_F16C && __AVX__ && !__F16C__
-    if (ncnn::cpu_support_x86_f16c())
-    {
-        innerproduct_gemm_fp16s_sse_f16c(bottom_blob, top_blob, weight_data_tm, bias_data, activation_type, activation_params, opt);
-        return;
-    }
-#else // NCNN_RUNTIME_CPU
-
-    const int num_input = bottom_blob.w;
-    const int elempack = bottom_blob.elempack;
-    const int num_output = top_blob.w;
-    const int h = bottom_blob.h;
-
-    const float* bias_data_ptr = bias_data;
-
-    int num_output_elempack = 1;
-#if __SSE2__
-    if (opt.use_packing_layout)
-    {
-#if __AVX512F__
-        num_output_elempack = num_output % 16 == 0 ? 16 : num_output % 8 == 0 ? 8 : num_output % 4 == 0 ? 4 : 1;
-#elif __AVX__
-        num_output_elempack = num_output % 8 == 0 ? 8 : num_output % 4 == 0 ? 4 : 1;
-#else
-        num_output_elempack = num_output % 4 == 0 ? 4 : 1;
-#endif
-    }
-#endif // __SSE2__
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int j = 0; j < h; j++)
-    {
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-        if (elempack == 16 && num_output_elempack == 16)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output / num_output_elempack; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = weight_data_tm.row(p);
-#endif
-                const float* m = bottom_blob.row(j);
-
-                __m512 _sum0 = _mm512_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm512_loadu_ps(bias_data_ptr + p * 16);
-                }
-
-                __m512 _sum1 = _sum0;
-                __m512 _sum2 = _sum0;
-                __m512 _sum3 = _sum0;
-                __m512 _sum4 = _sum0;
-                __m512 _sum5 = _sum0;
-                __m512 _sum6 = _sum0;
-                __m512 _sum7 = _sum0;
-                __m512 _sum8 = _sum0;
-                __m512 _sum9 = _sum0;
-                __m512 _suma = _sum0;
-                __m512 _sumb = _sum0;
-                __m512 _sumc = _sum0;
-                __m512 _sumd = _sum0;
-                __m512 _sume = _sum0;
-                __m512 _sumf = _sum0;
-
-                for (int i = 0; i < num_input; i++)
-                {
-                    __m512 _val0 = _mm512_set1_ps(m[0]);
-                    __m512 _val1 = _mm512_set1_ps(m[1]);
-                    __m512 _val2 = _mm512_set1_ps(m[2]);
-                    __m512 _val3 = _mm512_set1_ps(m[3]);
-                    __m512 _val4 = _mm512_set1_ps(m[4]);
-                    __m512 _val5 = _mm512_set1_ps(m[5]);
-                    __m512 _val6 = _mm512_set1_ps(m[6]);
-                    __m512 _val7 = _mm512_set1_ps(m[7]);
-                    __m512 _val8 = _mm512_set1_ps(m[8]);
-                    __m512 _val9 = _mm512_set1_ps(m[9]);
-                    __m512 _vala = _mm512_set1_ps(m[10]);
-                    __m512 _valb = _mm512_set1_ps(m[11]);
-                    __m512 _valc = _mm512_set1_ps(m[12]);
-                    __m512 _vald = _mm512_set1_ps(m[13]);
-                    __m512 _vale = _mm512_set1_ps(m[14]);
-                    __m512 _valf = _mm512_set1_ps(m[15]);
-
-#if NCNN_IMPL_FP16S
-                    __m512 _w = _mm512_cvtph_ps(_mm256_lddqu_si256((const __m256i*)kptr));
-#else
-                    __m512 _w = _mm512_loadu_ps(kptr);
-#endif
-
-                    _sum0 = _mm512_fmadd_ps(_val0, _w, _sum0);
-                    _sum1 = _mm512_fmadd_ps(_val1, _w, _sum1);
-                    _sum2 = _mm512_fmadd_ps(_val2, _w, _sum2);
-                    _sum3 = _mm512_fmadd_ps(_val3, _w, _sum3);
-                    _sum4 = _mm512_fmadd_ps(_val4, _w, _sum4);
-                    _sum5 = _mm512_fmadd_ps(_val5, _w, _sum5);
-                    _sum6 = _mm512_fmadd_ps(_val6, _w, _sum6);
-                    _sum7 = _mm512_fmadd_ps(_val7, _w, _sum7);
-                    _sum8 = _mm512_fmadd_ps(_val8, _w, _sum8);
-                    _sum9 = _mm512_fmadd_ps(_val9, _w, _sum9);
-                    _suma = _mm512_fmadd_ps(_vala, _w, _suma);
-                    _sumb = _mm512_fmadd_ps(_valb, _w, _sumb);
-                    _sumc = _mm512_fmadd_ps(_valc, _w, _sumc);
-                    _sumd = _mm512_fmadd_ps(_vald, _w, _sumd);
-                    _sume = _mm512_fmadd_ps(_vale, _w, _sume);
-                    _sumf = _mm512_fmadd_ps(_valf, _w, _sumf);
-
-                    m += 16;
-                    kptr += 16;
-                }
-
-                _sum0 = activation_avx512(_sum0, activation_type, activation_params);
-                _sum1 = activation_avx512(_sum1, activation_type, activation_params);
-                _sum2 = activation_avx512(_sum2, activation_type, activation_params);
-                _sum3 = activation_avx512(_sum3, activation_type, activation_params);
-                _sum4 = activation_avx512(_sum4, activation_type, activation_params);
-                _sum5 = activation_avx512(_sum5, activation_type, activation_params);
-                _sum6 = activation_avx512(_sum6, activation_type, activation_params);
-                _sum7 = activation_avx512(_sum7, activation_type, activation_params);
-                _sum8 = activation_avx512(_sum8, activation_type, activation_params);
-                _sum9 = activation_avx512(_sum9, activation_type, activation_params);
-                _suma = activation_avx512(_suma, activation_type, activation_params);
-                _sumb = activation_avx512(_sumb, activation_type, activation_params);
-                _sumc = activation_avx512(_sumc, activation_type, activation_params);
-                _sumd = activation_avx512(_sumd, activation_type, activation_params);
-                _sume = activation_avx512(_sume, activation_type, activation_params);
-                _sumf = activation_avx512(_sumf, activation_type, activation_params);
-
-                transpose16x16_ps(_sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7, _sum8, _sum9, _suma, _sumb, _sumc, _sumd, _sume, _sumf);
-
-                _mm512_storeu_ps(outptr, _sum0);
-                _mm512_storeu_ps(outptr + 16, _sum1);
-                _mm512_storeu_ps(outptr + 16 * 2, _sum2);
-                _mm512_storeu_ps(outptr + 16 * 3, _sum3);
-                _mm512_storeu_ps(outptr + 16 * 4, _sum4);
-                _mm512_storeu_ps(outptr + 16 * 5, _sum5);
-                _mm512_storeu_ps(outptr + 16 * 6, _sum6);
-                _mm512_storeu_ps(outptr + 16 * 7, _sum7);
-                _mm512_storeu_ps(outptr + 16 * 8, _sum8);
-                _mm512_storeu_ps(outptr + 16 * 9, _sum9);
-                _mm512_storeu_ps(outptr + 16 * 10, _suma);
-                _mm512_storeu_ps(outptr + 16 * 11, _sumb);
-                _mm512_storeu_ps(outptr + 16 * 12, _sumc);
-                _mm512_storeu_ps(outptr + 16 * 13, _sumd);
-                _mm512_storeu_ps(outptr + 16 * 14, _sume);
-                _mm512_storeu_ps(outptr + 16 * 15, _sumf);
-                outptr += 256;
-            }
-        }
-
-        if (elempack == 1 && num_output_elempack == 16)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output / num_output_elempack; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = weight_data_tm.row(p);
-#endif
-                const float* m = bottom_blob.row(j);
-
-                __m512 _sum = _mm512_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum = _mm512_loadu_ps(bias_data_ptr + p * 16);
-                }
-
-                int i = 0;
-                for (; i < num_input; i++)
-                {
-                    __m512 _val = _mm512_set1_ps(m[0]);
-#if NCNN_IMPL_FP16S
-                    __m512 _w = _mm512_cvtph_ps(_mm256_lddqu_si256((const __m256i*)kptr));
-#else
-                    __m512 _w = _mm512_loadu_ps(kptr);
-#endif
-
-                    _sum = _mm512_fmadd_ps(_val, _w, _sum);
-
-                    m += 1;
-                    kptr += 16;
-                }
-
-                _sum = activation_avx512(_sum, activation_type, activation_params);
-
-                _mm512_storeu_ps(outptr, _sum);
-                outptr += 16;
-            }
-        }
-
-        if (elempack == 4 && num_output_elempack == 16)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output / num_output_elempack; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = weight_data_tm.row(p);
-#endif
-                const float* m = bottom_blob.row(j);
-
-                __m512 _sum0 = _mm512_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm512_loadu_ps(bias_data_ptr + p * 16);
-                }
-
-                __m512 _sum1 = _sum0;
-                __m512 _sum2 = _sum0;
-                __m512 _sum3 = _sum0;
-
-                int i = 0;
-                for (; i < num_input; i++)
-                {
-                    __m512 _val0 = _mm512_set1_ps(m[0]);
-                    __m512 _val1 = _mm512_set1_ps(m[1]);
-                    __m512 _val2 = _mm512_set1_ps(m[2]);
-                    __m512 _val3 = _mm512_set1_ps(m[3]);
-#if NCNN_IMPL_FP16S
-                    __m512 _w = _mm512_cvtph_ps(_mm256_lddqu_si256((const __m256i*)kptr));
-#else
-                    __m512 _w = _mm512_loadu_ps(kptr);
-#endif
-
-                    _sum0 = _mm512_fmadd_ps(_val0, _w, _sum0);
-                    _sum1 = _mm512_fmadd_ps(_val1, _w, _sum1);
-                    _sum2 = _mm512_fmadd_ps(_val2, _w, _sum2);
-                    _sum3 = _mm512_fmadd_ps(_val3, _w, _sum3);
-
-                    m += 4;
-                    kptr += 16;
-                }
-
-                _sum0 = activation_avx512(_sum0, activation_type, activation_params);
-                _sum1 = activation_avx512(_sum1, activation_type, activation_params);
-                _sum2 = activation_avx512(_sum2, activation_type, activation_params);
-                _sum3 = activation_avx512(_sum3, activation_type, activation_params);
-
-                transpose16x4_ps(_sum0, _sum1, _sum2, _sum3);
-
-                _mm512_storeu_ps(outptr, _sum0);
-                _mm512_storeu_ps(outptr + 16, _sum1);
-                _mm512_storeu_ps(outptr + 32, _sum2);
-                _mm512_storeu_ps(outptr + 48, _sum3);
-                outptr += 64;
-            }
-        }
-
-        if (elempack == 8 && num_output_elempack == 16)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output / num_output_elempack; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = weight_data_tm.row(p);
-#endif
-                const float* m = bottom_blob.row(j);
-
-                __m512 _sum0 = _mm512_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm512_loadu_ps(bias_data_ptr + p * 16);
-                }
-
-                __m512 _sum1 = _sum0;
-                __m512 _sum2 = _sum0;
-                __m512 _sum3 = _sum0;
-                __m512 _sum4 = _sum0;
-                __m512 _sum5 = _sum0;
-                __m512 _sum6 = _sum0;
-                __m512 _sum7 = _sum0;
-
-                int i = 0;
-                for (; i < num_input; i++)
-                {
-                    __m512 _val0 = _mm512_set1_ps(m[0]);
-                    __m512 _val1 = _mm512_set1_ps(m[1]);
-                    __m512 _val2 = _mm512_set1_ps(m[2]);
-                    __m512 _val3 = _mm512_set1_ps(m[3]);
-                    __m512 _val4 = _mm512_set1_ps(m[4]);
-                    __m512 _val5 = _mm512_set1_ps(m[5]);
-                    __m512 _val6 = _mm512_set1_ps(m[6]);
-                    __m512 _val7 = _mm512_set1_ps(m[7]);
-#if NCNN_IMPL_FP16S
-                    __m512 _w = _mm512_cvtph_ps(_mm256_lddqu_si256((const __m256i*)kptr));
-#else
-                    __m512 _w = _mm512_loadu_ps(kptr);
-#endif
-
-                    _sum0 = _mm512_fmadd_ps(_val0, _w, _sum0);
-                    _sum1 = _mm512_fmadd_ps(_val1, _w, _sum1);
-                    _sum2 = _mm512_fmadd_ps(_val2, _w, _sum2);
-                    _sum3 = _mm512_fmadd_ps(_val3, _w, _sum3);
-                    _sum4 = _mm512_fmadd_ps(_val4, _w, _sum4);
-                    _sum5 = _mm512_fmadd_ps(_val5, _w, _sum5);
-                    _sum6 = _mm512_fmadd_ps(_val6, _w, _sum6);
-                    _sum7 = _mm512_fmadd_ps(_val7, _w, _sum7);
-
-                    m += 8;
-                    kptr += 16;
-                }
-
-                _sum0 = activation_avx512(_sum0, activation_type, activation_params);
-                _sum1 = activation_avx512(_sum1, activation_type, activation_params);
-                _sum2 = activation_avx512(_sum2, activation_type, activation_params);
-                _sum3 = activation_avx512(_sum3, activation_type, activation_params);
-                _sum4 = activation_avx512(_sum4, activation_type, activation_params);
-                _sum5 = activation_avx512(_sum5, activation_type, activation_params);
-                _sum6 = activation_avx512(_sum6, activation_type, activation_params);
-                _sum7 = activation_avx512(_sum7, activation_type, activation_params);
-
-                transpose16x8_ps(_sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7);
-
-                _mm512_storeu_ps(outptr, _sum0);
-                _mm512_storeu_ps(outptr + 16, _sum1);
-                _mm512_storeu_ps(outptr + 16 * 2, _sum2);
-                _mm512_storeu_ps(outptr + 16 * 3, _sum3);
-                _mm512_storeu_ps(outptr + 16 * 4, _sum4);
-                _mm512_storeu_ps(outptr + 16 * 5, _sum5);
-                _mm512_storeu_ps(outptr + 16 * 6, _sum6);
-                _mm512_storeu_ps(outptr + 16 * 7, _sum7);
-                outptr += 128;
-            }
-        }
-
-        if (elempack == 16 && num_output_elempack == 1)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = (const float*)weight_data_tm + num_input * p;
-#endif
-                const float* m = bottom_blob.row(j);
-
-                __m512 _sum0 = _mm512_setzero_ps();
-                __m512 _sum1 = _mm512_setzero_ps();
-                __m512 _sum2 = _mm512_setzero_ps();
-                __m512 _sum3 = _mm512_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm512_set1_ps(bias_data_ptr[p]);
-                }
-
-                int i = 0;
-                for (; i + 3 < num_input; i += 4)
-                {
-                    __m512 _val0 = _mm512_loadu_ps(m);
-                    __m512 _val1 = _mm512_loadu_ps(m + 16);
-                    __m512 _val2 = _mm512_loadu_ps(m + 32);
-                    __m512 _val3 = _mm512_loadu_ps(m + 48);
-#if NCNN_IMPL_FP16S
-                    __m128 _w = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i*)kptr));
-                    __m256 _ww = _mm256_insertf128_ps(_mm256_castps128_ps256(_w), _w, 1);
-                    __m512 _www = _mm512_insertf32x8(_mm512_castps256_ps512(_ww), _ww, 1);
-
-                    __m512 _w0 = _mm512_permute_ps(_www, _MM_SHUFFLE(0, 0, 0, 0));
-                    __m512 _w1 = _mm512_permute_ps(_www, _MM_SHUFFLE(1, 1, 1, 1));
-                    __m512 _w2 = _mm512_permute_ps(_www, _MM_SHUFFLE(2, 2, 2, 2));
-                    __m512 _w3 = _mm512_permute_ps(_www, _MM_SHUFFLE(3, 3, 3, 3));
-#else
-                    __m512 _w0 = _mm512_set1_ps(kptr[0]);
-                    __m512 _w1 = _mm512_set1_ps(kptr[1]);
-                    __m512 _w2 = _mm512_set1_ps(kptr[2]);
-                    __m512 _w3 = _mm512_set1_ps(kptr[3]);
-#endif
-
-                    _sum0 = _mm512_fmadd_ps(_val0, _w0, _sum0);
-                    _sum1 = _mm512_fmadd_ps(_val1, _w1, _sum1);
-                    _sum2 = _mm512_fmadd_ps(_val2, _w2, _sum2);
-                    _sum3 = _mm512_fmadd_ps(_val3, _w3, _sum3);
-
-                    m += 64;
-                    kptr += 4;
-                }
-                for (; i < num_input; i++)
-                {
-                    __m512 _val = _mm512_loadu_ps(m);
-#if NCNN_IMPL_FP16S
-                    __m512 _w = _mm512_set1_ps(float16_to_float32(kptr[0]));
-#else
-                    __m512 _w = _mm512_set1_ps(kptr[0]);
-#endif
-                    _sum0 = _mm512_fmadd_ps(_val, _w, _sum0);
-
-                    m += 16;
-                    kptr += 1;
-                }
-
-                _sum0 = _mm512_add_ps(_sum0, _sum1);
-                _sum2 = _mm512_add_ps(_sum2, _sum3);
-                _sum0 = _mm512_add_ps(_sum0, _sum2);
-
-                _sum0 = activation_avx512(_sum0, activation_type, activation_params);
-
-                _mm512_storeu_ps(outptr, _sum0);
-                outptr += 16;
-            }
-        }
-
-        if (elempack == 16 && num_output_elempack == 4)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output / num_output_elempack; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = weight_data_tm.row(p);
-#endif
-                const float* m = bottom_blob.row(j);
-
-                __m512 _sum0 = _mm512_setzero_ps();
-                __m512 _sum1 = _mm512_setzero_ps();
-                __m512 _sum2 = _mm512_setzero_ps();
-                __m512 _sum3 = _mm512_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm512_set1_ps(bias_data_ptr[p * 4 + 0]);
-                    _sum1 = _mm512_set1_ps(bias_data_ptr[p * 4 + 1]);
-                    _sum2 = _mm512_set1_ps(bias_data_ptr[p * 4 + 2]);
-                    _sum3 = _mm512_set1_ps(bias_data_ptr[p * 4 + 3]);
-                }
-
-                int i = 0;
-                for (; i < num_input; i++)
-                {
-                    __m512 _val = _mm512_loadu_ps(m);
-#if NCNN_IMPL_FP16S
-                    __m128 _w = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i*)kptr));
-                    __m256 _ww = _mm256_insertf128_ps(_mm256_castps128_ps256(_w), _w, 1);
-                    __m512 _www = _mm512_insertf32x8(_mm512_castps256_ps512(_ww), _ww, 1);
-
-                    __m512 _w0 = _mm512_permute_ps(_www, _MM_SHUFFLE(0, 0, 0, 0));
-                    __m512 _w1 = _mm512_permute_ps(_www, _MM_SHUFFLE(1, 1, 1, 1));
-                    __m512 _w2 = _mm512_permute_ps(_www, _MM_SHUFFLE(2, 2, 2, 2));
-                    __m512 _w3 = _mm512_permute_ps(_www, _MM_SHUFFLE(3, 3, 3, 3));
-#else
-                    __m512 _w0 = _mm512_set1_ps(kptr[0]);
-                    __m512 _w1 = _mm512_set1_ps(kptr[1]);
-                    __m512 _w2 = _mm512_set1_ps(kptr[2]);
-                    __m512 _w3 = _mm512_set1_ps(kptr[3]);
-#endif
-
-                    _sum0 = _mm512_fmadd_ps(_val, _w0, _sum0);
-                    _sum1 = _mm512_fmadd_ps(_val, _w1, _sum1);
-                    _sum2 = _mm512_fmadd_ps(_val, _w2, _sum2);
-                    _sum3 = _mm512_fmadd_ps(_val, _w3, _sum3);
-
-                    m += 16;
-                    kptr += 4;
-                }
-
-                _sum0 = activation_avx512(_sum0, activation_type, activation_params);
-                _sum1 = activation_avx512(_sum1, activation_type, activation_params);
-                _sum2 = activation_avx512(_sum2, activation_type, activation_params);
-                _sum3 = activation_avx512(_sum3, activation_type, activation_params);
-
-                _mm512_storeu_ps(outptr, _sum0);
-                _mm512_storeu_ps(outptr + 16, _sum1);
-                _mm512_storeu_ps(outptr + 32, _sum2);
-                _mm512_storeu_ps(outptr + 48, _sum3);
-                outptr += 64;
-            }
-        }
-
-        if (elempack == 16 && num_output_elempack == 8)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output / num_output_elempack; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = weight_data_tm.row(p);
-#endif
-                const float* m = bottom_blob.row(j);
-
-                __m512 _sum0 = _mm512_setzero_ps();
-                __m512 _sum1 = _mm512_setzero_ps();
-                __m512 _sum2 = _mm512_setzero_ps();
-                __m512 _sum3 = _mm512_setzero_ps();
-                __m512 _sum4 = _mm512_setzero_ps();
-                __m512 _sum5 = _mm512_setzero_ps();
-                __m512 _sum6 = _mm512_setzero_ps();
-                __m512 _sum7 = _mm512_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm512_set1_ps(bias_data_ptr[p * 8 + 0]);
-                    _sum1 = _mm512_set1_ps(bias_data_ptr[p * 8 + 1]);
-                    _sum2 = _mm512_set1_ps(bias_data_ptr[p * 8 + 2]);
-                    _sum3 = _mm512_set1_ps(bias_data_ptr[p * 8 + 3]);
-                    _sum4 = _mm512_set1_ps(bias_data_ptr[p * 8 + 4]);
-                    _sum5 = _mm512_set1_ps(bias_data_ptr[p * 8 + 5]);
-                    _sum6 = _mm512_set1_ps(bias_data_ptr[p * 8 + 6]);
-                    _sum7 = _mm512_set1_ps(bias_data_ptr[p * 8 + 7]);
-                }
-
-                int i = 0;
-                for (; i < num_input; i++)
-                {
-                    __m512 _val = _mm512_loadu_ps(m);
-#if NCNN_IMPL_FP16S
-                    __m256 _w = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)kptr));
-                    __m512 _ww = _mm512_castps256_ps512(_w);
-                    __m512 _www0 = _mm512_shuffle_f32x4(_ww, _ww, _MM_SHUFFLE(0, 0, 0, 0));
-                    __m512 _www1 = _mm512_shuffle_f32x4(_ww, _ww, _MM_SHUFFLE(1, 1, 1, 1));
-
-                    __m512 _w0 = _mm512_permute_ps(_www0, _MM_SHUFFLE(0, 0, 0, 0));
-                    __m512 _w1 = _mm512_permute_ps(_www0, _MM_SHUFFLE(1, 1, 1, 1));
-                    __m512 _w2 = _mm512_permute_ps(_www0, _MM_SHUFFLE(2, 2, 2, 2));
-                    __m512 _w3 = _mm512_permute_ps(_www0, _MM_SHUFFLE(3, 3, 3, 3));
-                    __m512 _w4 = _mm512_permute_ps(_www1, _MM_SHUFFLE(0, 0, 0, 0));
-                    __m512 _w5 = _mm512_permute_ps(_www1, _MM_SHUFFLE(1, 1, 1, 1));
-                    __m512 _w6 = _mm512_permute_ps(_www1, _MM_SHUFFLE(2, 2, 2, 2));
-                    __m512 _w7 = _mm512_permute_ps(_www1, _MM_SHUFFLE(3, 3, 3, 3));
-#else
-                    __m512 _w0 = _mm512_set1_ps(kptr[0]);
-                    __m512 _w1 = _mm512_set1_ps(kptr[1]);
-                    __m512 _w2 = _mm512_set1_ps(kptr[2]);
-                    __m512 _w3 = _mm512_set1_ps(kptr[3]);
-                    __m512 _w4 = _mm512_set1_ps(kptr[4]);
-                    __m512 _w5 = _mm512_set1_ps(kptr[5]);
-                    __m512 _w6 = _mm512_set1_ps(kptr[6]);
-                    __m512 _w7 = _mm512_set1_ps(kptr[7]);
-#endif
-
-                    _sum0 = _mm512_fmadd_ps(_val, _w0, _sum0);
-                    _sum1 = _mm512_fmadd_ps(_val, _w1, _sum1);
-                    _sum2 = _mm512_fmadd_ps(_val, _w2, _sum2);
-                    _sum3 = _mm512_fmadd_ps(_val, _w3, _sum3);
-                    _sum4 = _mm512_fmadd_ps(_val, _w4, _sum4);
-                    _sum5 = _mm512_fmadd_ps(_val, _w5, _sum5);
-                    _sum6 = _mm512_fmadd_ps(_val, _w6, _sum6);
-                    _sum7 = _mm512_fmadd_ps(_val, _w7, _sum7);
-
-                    m += 16;
-                    kptr += 8;
-                }
-
-                _sum0 = activation_avx512(_sum0, activation_type, activation_params);
-                _sum1 = activation_avx512(_sum1, activation_type, activation_params);
-                _sum2 = activation_avx512(_sum2, activation_type, activation_params);
-                _sum3 = activation_avx512(_sum3, activation_type, activation_params);
-                _sum4 = activation_avx512(_sum4, activation_type, activation_params);
-                _sum5 = activation_avx512(_sum5, activation_type, activation_params);
-                _sum6 = activation_avx512(_sum6, activation_type, activation_params);
-                _sum7 = activation_avx512(_sum7, activation_type, activation_params);
-
-                _mm512_storeu_ps(outptr, _sum0);
-                _mm512_storeu_ps(outptr + 16, _sum1);
-                _mm512_storeu_ps(outptr + 16 * 2, _sum2);
-                _mm512_storeu_ps(outptr + 16 * 3, _sum3);
-                _mm512_storeu_ps(outptr + 16 * 4, _sum4);
-                _mm512_storeu_ps(outptr + 16 * 5, _sum5);
-                _mm512_storeu_ps(outptr + 16 * 6, _sum6);
-                _mm512_storeu_ps(outptr + 16 * 7, _sum7);
-                outptr += 128;
-            }
-        }
-
-#endif // __AVX512F__
-
-        if (elempack == 8 && num_output_elempack == 8)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output / num_output_elempack; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = weight_data_tm.row(p);
-#endif
-                const float* m = bottom_blob.row(j);
-
-                __m256 _sum0 = _mm256_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm256_loadu_ps(bias_data_ptr + p * 8);
-                }
-
-                __m256 _sum1 = _sum0;
-                __m256 _sum2 = _sum0;
-                __m256 _sum3 = _sum0;
-                __m256 _sum4 = _sum0;
-                __m256 _sum5 = _sum0;
-                __m256 _sum6 = _sum0;
-                __m256 _sum7 = _sum0;
-
-                for (int i = 0; i < num_input; i++)
-                {
-                    __m256 _val0 = _mm256_broadcast_ss(m);
-                    __m256 _val1 = _mm256_broadcast_ss(m + 1);
-                    __m256 _val2 = _mm256_broadcast_ss(m + 2);
-                    __m256 _val3 = _mm256_broadcast_ss(m + 3);
-                    __m256 _val4 = _mm256_broadcast_ss(m + 4);
-                    __m256 _val5 = _mm256_broadcast_ss(m + 5);
-                    __m256 _val6 = _mm256_broadcast_ss(m + 6);
-                    __m256 _val7 = _mm256_broadcast_ss(m + 7);
-#if NCNN_IMPL_FP16S
-                    __m256 _w = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)kptr));
-#else
-                    __m256 _w = _mm256_loadu_ps(kptr);
-#endif
-
-                    _sum0 = _mm256_comp_fmadd_ps(_val0, _w, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_val1, _w, _sum1);
-                    _sum2 = _mm256_comp_fmadd_ps(_val2, _w, _sum2);
-                    _sum3 = _mm256_comp_fmadd_ps(_val3, _w, _sum3);
-                    _sum4 = _mm256_comp_fmadd_ps(_val4, _w, _sum4);
-                    _sum5 = _mm256_comp_fmadd_ps(_val5, _w, _sum5);
-                    _sum6 = _mm256_comp_fmadd_ps(_val6, _w, _sum6);
-                    _sum7 = _mm256_comp_fmadd_ps(_val7, _w, _sum7);
-
-                    m += 8;
-                    kptr += 8;
-                }
-
-                _sum0 = activation_avx(_sum0, activation_type, activation_params);
-                _sum1 = activation_avx(_sum1, activation_type, activation_params);
-                _sum2 = activation_avx(_sum2, activation_type, activation_params);
-                _sum3 = activation_avx(_sum3, activation_type, activation_params);
-                _sum4 = activation_avx(_sum4, activation_type, activation_params);
-                _sum5 = activation_avx(_sum5, activation_type, activation_params);
-                _sum6 = activation_avx(_sum6, activation_type, activation_params);
-                _sum7 = activation_avx(_sum7, activation_type, activation_params);
-
-                transpose8x8_ps(_sum0, _sum1, _sum2, _sum3, _sum4, _sum5, _sum6, _sum7);
-
-                _mm256_storeu_ps(outptr, _sum0);
-                _mm256_storeu_ps(outptr + 8, _sum1);
-                _mm256_storeu_ps(outptr + 16, _sum2);
-                _mm256_storeu_ps(outptr + 24, _sum3);
-                _mm256_storeu_ps(outptr + 32, _sum4);
-                _mm256_storeu_ps(outptr + 40, _sum5);
-                _mm256_storeu_ps(outptr + 48, _sum6);
-                _mm256_storeu_ps(outptr + 56, _sum7);
-                outptr += 64;
-            }
-        }
-
-        if (elempack == 1 && num_output_elempack == 8)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output / num_output_elempack; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = weight_data_tm.row(p);
-#endif
-                const float* m = bottom_blob.row(j);
-
-                __m256 _sum0 = _mm256_setzero_ps();
-                __m256 _sum1 = _mm256_setzero_ps();
-                __m256 _sum2 = _mm256_setzero_ps();
-                __m256 _sum3 = _mm256_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm256_loadu_ps(bias_data_ptr + p * 8);
-                }
-
-                int i = 0;
-                for (; i + 7 < num_input; i += 8)
-                {
-                    __m256 _val0 = _mm256_broadcast_ss(m);
-                    __m256 _val1 = _mm256_broadcast_ss(m + 1);
-                    __m256 _val2 = _mm256_broadcast_ss(m + 2);
-                    __m256 _val3 = _mm256_broadcast_ss(m + 3);
-#if NCNN_IMPL_FP16S
-                    __m256i _w01 = _mm256_lddqu_si256((const __m256i*)kptr);
-                    __m256i _w23 = _mm256_lddqu_si256((const __m256i*)(kptr + 16));
-                    __m256 _w0 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w01, 0));
-                    __m256 _w1 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w01, 1));
-                    __m256 _w2 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w23, 0));
-                    __m256 _w3 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w23, 1));
-#else
-                    __m256 _w0 = _mm256_loadu_ps(kptr);
-                    __m256 _w1 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _w2 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _w3 = _mm256_loadu_ps(kptr + 24);
-#endif
-
-                    _sum0 = _mm256_comp_fmadd_ps(_val0, _w0, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_val1, _w1, _sum1);
-                    _sum2 = _mm256_comp_fmadd_ps(_val2, _w2, _sum2);
-                    _sum3 = _mm256_comp_fmadd_ps(_val3, _w3, _sum3);
-
-                    __m256 _val4 = _mm256_broadcast_ss(m + 4);
-                    __m256 _val5 = _mm256_broadcast_ss(m + 5);
-                    __m256 _val6 = _mm256_broadcast_ss(m + 6);
-                    __m256 _val7 = _mm256_broadcast_ss(m + 7);
-#if NCNN_IMPL_FP16S
-                    __m256i _w45 = _mm256_lddqu_si256((const __m256i*)(kptr + 32));
-                    __m256i _w67 = _mm256_lddqu_si256((const __m256i*)(kptr + 48));
-                    __m256 _w4 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w45, 0));
-                    __m256 _w5 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w45, 1));
-                    __m256 _w6 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w67, 0));
-                    __m256 _w7 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w67, 1));
-#else
-                    __m256 _w4 = _mm256_loadu_ps(kptr + 32);
-                    __m256 _w5 = _mm256_loadu_ps(kptr + 40);
-                    __m256 _w6 = _mm256_loadu_ps(kptr + 48);
-                    __m256 _w7 = _mm256_loadu_ps(kptr + 56);
-#endif
-
-                    _sum0 = _mm256_comp_fmadd_ps(_val4, _w4, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_val5, _w5, _sum1);
-                    _sum2 = _mm256_comp_fmadd_ps(_val6, _w6, _sum2);
-                    _sum3 = _mm256_comp_fmadd_ps(_val7, _w7, _sum3);
-
-                    m += 8;
-                    kptr += 64;
-                }
-                for (; i + 3 < num_input; i += 4)
-                {
-                    __m256 _val0 = _mm256_broadcast_ss(m);
-                    __m256 _val1 = _mm256_broadcast_ss(m + 1);
-                    __m256 _val2 = _mm256_broadcast_ss(m + 2);
-                    __m256 _val3 = _mm256_broadcast_ss(m + 3);
-#if NCNN_IMPL_FP16S
-                    __m256i _w01 = _mm256_lddqu_si256((const __m256i*)kptr);
-                    __m256i _w23 = _mm256_lddqu_si256((const __m256i*)(kptr + 16));
-                    __m256 _w0 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w01, 0));
-                    __m256 _w1 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w01, 1));
-                    __m256 _w2 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w23, 0));
-                    __m256 _w3 = _mm256_cvtph_ps(_mm256_extractf128_si256(_w23, 1));
-#else
-                    __m256 _w0 = _mm256_loadu_ps(kptr);
-                    __m256 _w1 = _mm256_loadu_ps(kptr + 8);
-                    __m256 _w2 = _mm256_loadu_ps(kptr + 16);
-                    __m256 _w3 = _mm256_loadu_ps(kptr + 24);
-#endif
-
-                    _sum0 = _mm256_comp_fmadd_ps(_val0, _w0, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_val1, _w1, _sum1);
-                    _sum2 = _mm256_comp_fmadd_ps(_val2, _w2, _sum2);
-                    _sum3 = _mm256_comp_fmadd_ps(_val3, _w3, _sum3);
-
-                    m += 4;
-                    kptr += 32;
-                }
-                for (; i < num_input; i++)
-                {
-                    __m256 _val = _mm256_set1_ps(m[0]);
-#if NCNN_IMPL_FP16S
-                    __m256 _w = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)kptr));
-#else
-                    __m256 _w = _mm256_loadu_ps(kptr);
-#endif
-                    _sum0 = _mm256_comp_fmadd_ps(_val, _w, _sum0);
-
-                    m += 1;
-                    kptr += 8;
-                }
-
-                _sum0 = _mm256_add_ps(_sum0, _sum1);
-                _sum2 = _mm256_add_ps(_sum2, _sum3);
-                _sum0 = _mm256_add_ps(_sum0, _sum2);
-
-                _sum0 = activation_avx(_sum0, activation_type, activation_params);
-
-                _mm256_storeu_ps(outptr, _sum0);
-                outptr += 8;
-            }
-        }
-
-        if (elempack == 4 && num_output_elempack == 8)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output / num_output_elempack; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = weight_data_tm.row(p);
-#endif
-                const float* m = bottom_blob.row(j);
-
-                __m256 _sum0 = _mm256_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm256_loadu_ps(bias_data_ptr + p * 8);
-                }
-
-                __m256 _sum1 = _sum0;
-                __m256 _sum2 = _sum0;
-                __m256 _sum3 = _sum0;
-
-                int i = 0;
-                for (; i < num_input; i++)
-                {
-                    __m256 _val0 = _mm256_broadcast_ss(m);
-                    __m256 _val1 = _mm256_broadcast_ss(m + 1);
-                    __m256 _val2 = _mm256_broadcast_ss(m + 2);
-                    __m256 _val3 = _mm256_broadcast_ss(m + 3);
-#if NCNN_IMPL_FP16S
-                    __m256 _w = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)kptr));
-#else
-                    __m256 _w = _mm256_loadu_ps(kptr);
-#endif
-
-                    _sum0 = _mm256_comp_fmadd_ps(_val0, _w, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_val1, _w, _sum1);
-                    _sum2 = _mm256_comp_fmadd_ps(_val2, _w, _sum2);
-                    _sum3 = _mm256_comp_fmadd_ps(_val3, _w, _sum3);
-
-                    m += 4;
-                    kptr += 8;
-                }
-
-                _sum0 = activation_avx(_sum0, activation_type, activation_params);
-                _sum1 = activation_avx(_sum1, activation_type, activation_params);
-                _sum2 = activation_avx(_sum2, activation_type, activation_params);
-                _sum3 = activation_avx(_sum3, activation_type, activation_params);
-
-                transpose8x4_ps(_sum0, _sum1, _sum2, _sum3);
-
-                _mm256_storeu_ps(outptr, _sum0);
-                _mm256_storeu_ps(outptr + 8, _sum1);
-                _mm256_storeu_ps(outptr + 16, _sum2);
-                _mm256_storeu_ps(outptr + 24, _sum3);
-                outptr += 32;
-            }
-        }
-
-        if (elempack == 8 && num_output_elempack == 1)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = (const float*)weight_data_tm + num_input * p;
-#endif
-                const float* m = bottom_blob.row(j);
-
-                __m256 _sum0 = _mm256_setzero_ps();
-                __m256 _sum1 = _mm256_setzero_ps();
-                __m256 _sum2 = _mm256_setzero_ps();
-                __m256 _sum3 = _mm256_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm256_set1_ps(bias_data_ptr[p]);
-                }
-
-                int i = 0;
-                for (; i + 3 < num_input; i += 4)
-                {
-                    __m256 _val0 = _mm256_loadu_ps(m);
-                    __m256 _val1 = _mm256_loadu_ps(m + 8);
-                    __m256 _val2 = _mm256_loadu_ps(m + 16);
-                    __m256 _val3 = _mm256_loadu_ps(m + 24);
-#if NCNN_IMPL_FP16S
-                    __m128 _w = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i*)kptr));
-                    __m256 _ww = _mm256_insertf128_ps(_mm256_castps128_ps256(_w), _w, 1);
-
-                    __m256 _w0 = _mm256_permute_ps(_ww, _MM_SHUFFLE(0, 0, 0, 0));
-                    __m256 _w1 = _mm256_permute_ps(_ww, _MM_SHUFFLE(1, 1, 1, 1));
-                    __m256 _w2 = _mm256_permute_ps(_ww, _MM_SHUFFLE(2, 2, 2, 2));
-                    __m256 _w3 = _mm256_permute_ps(_ww, _MM_SHUFFLE(3, 3, 3, 3));
-#else
-                    __m256 _w0 = _mm256_set1_ps(kptr[0]);
-                    __m256 _w1 = _mm256_set1_ps(kptr[1]);
-                    __m256 _w2 = _mm256_set1_ps(kptr[2]);
-                    __m256 _w3 = _mm256_set1_ps(kptr[3]);
-#endif
-
-                    _sum0 = _mm256_comp_fmadd_ps(_val0, _w0, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_val1, _w1, _sum1);
-                    _sum2 = _mm256_comp_fmadd_ps(_val2, _w2, _sum2);
-                    _sum3 = _mm256_comp_fmadd_ps(_val3, _w3, _sum3);
-
-                    m += 32;
-                    kptr += 4;
-                }
-                for (; i < num_input; i++)
-                {
-                    __m256 _val = _mm256_loadu_ps(m);
-#if NCNN_IMPL_FP16S
-                    __m256 _w = _mm256_set1_ps(float16_to_float32(kptr[0]));
-#else
-                    __m256 _w = _mm256_set1_ps(kptr[0]);
-#endif
-                    _sum0 = _mm256_comp_fmadd_ps(_val, _w, _sum0);
-
-                    m += 8;
-                    kptr += 1;
-                }
-
-                _sum0 = _mm256_add_ps(_sum0, _sum1);
-                _sum2 = _mm256_add_ps(_sum2, _sum3);
-                _sum0 = _mm256_add_ps(_sum0, _sum2);
-
-                _sum0 = activation_avx(_sum0, activation_type, activation_params);
-
-                _mm256_storeu_ps(outptr, _sum0);
-                outptr += 8;
-            }
-        }
-
-        if (elempack == 8 && num_output_elempack == 4)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output / num_output_elempack; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = weight_data_tm.row(p);
-#endif
-                const float* m = bottom_blob.row(j);
-
-                __m256 _sum0 = _mm256_setzero_ps();
-                __m256 _sum1 = _mm256_setzero_ps();
-                __m256 _sum2 = _mm256_setzero_ps();
-                __m256 _sum3 = _mm256_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm256_set1_ps(bias_data_ptr[p * 4 + 0]);
-                    _sum1 = _mm256_set1_ps(bias_data_ptr[p * 4 + 1]);
-                    _sum2 = _mm256_set1_ps(bias_data_ptr[p * 4 + 2]);
-                    _sum3 = _mm256_set1_ps(bias_data_ptr[p * 4 + 3]);
-                }
-
-                int i = 0;
-                for (; i < num_input; i++)
-                {
-                    __m256 _val = _mm256_loadu_ps(m);
-#if NCNN_IMPL_FP16S
-                    __m128 _w = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i*)kptr));
-                    __m256 _ww = _mm256_insertf128_ps(_mm256_castps128_ps256(_w), _w, 1);
-
-                    __m256 _w0 = _mm256_permute_ps(_ww, _MM_SHUFFLE(0, 0, 0, 0));
-                    __m256 _w1 = _mm256_permute_ps(_ww, _MM_SHUFFLE(1, 1, 1, 1));
-                    __m256 _w2 = _mm256_permute_ps(_ww, _MM_SHUFFLE(2, 2, 2, 2));
-                    __m256 _w3 = _mm256_permute_ps(_ww, _MM_SHUFFLE(3, 3, 3, 3));
-#else
-                    __m256 _w0 = _mm256_set1_ps(kptr[0]);
-                    __m256 _w1 = _mm256_set1_ps(kptr[1]);
-                    __m256 _w2 = _mm256_set1_ps(kptr[2]);
-                    __m256 _w3 = _mm256_set1_ps(kptr[3]);
-#endif
-
-                    _sum0 = _mm256_comp_fmadd_ps(_val, _w0, _sum0);
-                    _sum1 = _mm256_comp_fmadd_ps(_val, _w1, _sum1);
-                    _sum2 = _mm256_comp_fmadd_ps(_val, _w2, _sum2);
-                    _sum3 = _mm256_comp_fmadd_ps(_val, _w3, _sum3);
-
-                    m += 8;
-                    kptr += 4;
-                }
-
-                _sum0 = activation_avx(_sum0, activation_type, activation_params);
-                _sum1 = activation_avx(_sum1, activation_type, activation_params);
-                _sum2 = activation_avx(_sum2, activation_type, activation_params);
-                _sum3 = activation_avx(_sum3, activation_type, activation_params);
-
-                _mm256_storeu_ps(outptr, _sum0);
-                _mm256_storeu_ps(outptr + 8, _sum1);
-                _mm256_storeu_ps(outptr + 16, _sum2);
-                _mm256_storeu_ps(outptr + 24, _sum3);
-                outptr += 32;
-            }
-        }
-#endif // __AVX__
-
-        if (elempack == 4 && num_output_elempack == 4)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output / num_output_elempack; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = weight_data_tm.row(p);
-#endif
-                const float* m = bottom_blob.row(j);
-
-                __m128 _sum0 = _mm_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm_loadu_ps(bias_data_ptr + p * 4);
-                }
-
-                __m128 _sum1 = _sum0;
-                __m128 _sum2 = _sum0;
-                __m128 _sum3 = _sum0;
-
-                int i = 0;
-                for (; i < num_input; i++)
-                {
-                    __m128 _val0 = _mm_set1_ps(m[0]);
-                    __m128 _val1 = _mm_set1_ps(m[1]);
-                    __m128 _val2 = _mm_set1_ps(m[2]);
-                    __m128 _val3 = _mm_set1_ps(m[3]);
-#if NCNN_IMPL_FP16S
-                    __m128 _w = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i*)kptr));
-#else
-                    __m128 _w = _mm_loadu_ps(kptr);
-#endif
-                    _sum0 = _mm_comp_fmadd_ps(_val0, _w, _sum0);
-                    _sum1 = _mm_comp_fmadd_ps(_val1, _w, _sum1);
-                    _sum2 = _mm_comp_fmadd_ps(_val2, _w, _sum2);
-                    _sum3 = _mm_comp_fmadd_ps(_val3, _w, _sum3);
-
-                    m += 4;
-                    kptr += 4;
-                }
-
-                _sum0 = activation_sse(_sum0, activation_type, activation_params);
-                _sum1 = activation_sse(_sum1, activation_type, activation_params);
-                _sum2 = activation_sse(_sum2, activation_type, activation_params);
-                _sum3 = activation_sse(_sum3, activation_type, activation_params);
-
-                _MM_TRANSPOSE4_PS(_sum0, _sum1, _sum2, _sum3);
-
-                _mm_storeu_ps(outptr, _sum0);
-                _mm_storeu_ps(outptr + 4, _sum1);
-                _mm_storeu_ps(outptr + 8, _sum2);
-                _mm_storeu_ps(outptr + 12, _sum3);
-                outptr += 16;
-            }
-        }
-
-        if (elempack == 1 && num_output_elempack == 4)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output / num_output_elempack; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = weight_data_tm.row(p);
-#endif
-                const float* m = bottom_blob.row(j);
-
-                __m128 _sum = _mm_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum = _mm_loadu_ps(bias_data_ptr + p * 4);
-                }
-
-                int i = 0;
-                for (; i < num_input; i++)
-                {
-                    __m128 _val = _mm_set1_ps(m[0]);
-#if NCNN_IMPL_FP16S
-                    __m128 _w = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i*)kptr));
-#else
-                    __m128 _w = _mm_loadu_ps(kptr);
-#endif
-                    _sum = _mm_comp_fmadd_ps(_val, _w, _sum);
-
-                    m += 1;
-                    kptr += 4;
-                }
-
-                _sum = activation_sse(_sum, activation_type, activation_params);
-
-                _mm_storeu_ps(outptr, _sum);
-                outptr += 4;
-            }
-        }
-
-        if (elempack == 4 && num_output_elempack == 1)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = (const float*)weight_data_tm + num_input * p;
-#endif
-                const float* m = bottom_blob.row(j);
-
-                __m128 _sum0 = _mm_setzero_ps();
-                __m128 _sum1 = _mm_setzero_ps();
-                __m128 _sum2 = _mm_setzero_ps();
-                __m128 _sum3 = _mm_setzero_ps();
-
-                if (bias_data_ptr)
-                {
-                    _sum0 = _mm_set1_ps(bias_data_ptr[p]);
-                }
-
-                int i = 0;
-                for (; i + 3 < num_input; i += 4)
-                {
-                    __m128 _val0 = _mm_loadu_ps(m);
-                    __m128 _val1 = _mm_loadu_ps(m + 4);
-                    __m128 _val2 = _mm_loadu_ps(m + 8);
-                    __m128 _val3 = _mm_loadu_ps(m + 12);
-#if NCNN_IMPL_FP16S
-                    __m128 _w = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i*)kptr));
-
-                    __m128 _w0 = _mm_permute_ps(_w, _MM_SHUFFLE(0, 0, 0, 0));
-                    __m128 _w1 = _mm_permute_ps(_w, _MM_SHUFFLE(1, 1, 1, 1));
-                    __m128 _w2 = _mm_permute_ps(_w, _MM_SHUFFLE(2, 2, 2, 2));
-                    __m128 _w3 = _mm_permute_ps(_w, _MM_SHUFFLE(3, 3, 3, 3));
-#else
-                    __m128 _w0 = _mm_set1_ps(kptr[0]);
-                    __m128 _w1 = _mm_set1_ps(kptr[1]);
-                    __m128 _w2 = _mm_set1_ps(kptr[2]);
-                    __m128 _w3 = _mm_set1_ps(kptr[3]);
-#endif
-
-                    _sum0 = _mm_comp_fmadd_ps(_val0, _w0, _sum0);
-                    _sum1 = _mm_comp_fmadd_ps(_val1, _w1, _sum1);
-                    _sum2 = _mm_comp_fmadd_ps(_val2, _w2, _sum2);
-                    _sum3 = _mm_comp_fmadd_ps(_val3, _w3, _sum3);
-
-                    m += 16;
-                    kptr += 4;
-                }
-                for (; i < num_input; i++)
-                {
-                    __m128 _val = _mm_loadu_ps(m);
-#if NCNN_IMPL_FP16S
-                    __m128 _w = _mm_set1_ps(float16_to_float32(kptr[0]));
-#else
-                    __m128 _w = _mm_set1_ps(kptr[0]);
-#endif
-                    _sum0 = _mm_comp_fmadd_ps(_val, _w, _sum0);
-
-                    m += 4;
-                    kptr += 1;
-                }
-
-                _sum0 = _mm_add_ps(_sum0, _sum1);
-                _sum2 = _mm_add_ps(_sum2, _sum3);
-                _sum0 = _mm_add_ps(_sum0, _sum2);
-
-                _sum0 = activation_sse(_sum0, activation_type, activation_params);
-
-                _mm_storeu_ps(outptr, _sum0);
-                outptr += 4;
-            }
-        }
-#endif // __SSE2__
-
-        if (elempack == 1 && num_output_elempack == 1)
-        {
-            float* outptr = top_blob.row(j);
-
-            for (int p = 0; p < num_output; p++)
-            {
-#if NCNN_IMPL_FP16S
-                const unsigned short* kptr = weight_data_tm.row<const unsigned short>(p);
-#else
-                const float* kptr = (const float*)weight_data_tm + num_input * p;
-#endif
-                const float* m = bottom_blob.row(j);
-
-                float sum = 0.f;
-
-                if (bias_data_ptr)
-                {
-                    sum = bias_data_ptr[p];
-                }
-
-                int i = 0;
-#if __SSE2__
-#if __AVX__
-                __m256 _sum = _mm256_setzero_ps();
-                for (; i + 7 < num_input; i += 8)
-                {
-                    __m256 _m = _mm256_loadu_ps(m);
-#if NCNN_IMPL_FP16S
-                    __m256 _w = _mm256_cvtph_ps(_mm_lddqu_si128((const __m128i*)kptr));
-#else
-                    __m256 _w = _mm256_loadu_ps(kptr);
-#endif
-                    _sum = _mm256_comp_fmadd_ps(_m, _w, _sum);
-
-                    m += 8;
-                    kptr += 8;
-                }
-#endif // __AVX__
-                __m128 _suml = _mm_setzero_ps();
-                for (; i + 3 < num_input; i += 4)
-                {
-                    __m128 _val = _mm_loadu_ps(m);
-#if NCNN_IMPL_FP16S
-                    __m128 _w = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i*)kptr));
-#else
-                    __m128 _w = _mm_loadu_ps(kptr);
-#endif
-                    _suml = _mm_comp_fmadd_ps(_val, _w, _suml);
-
-                    m += 4;
-                    kptr += 4;
-                }
-#endif // __SSE2__
-                for (; i < num_input; i++)
-                {
-#if NCNN_IMPL_FP16S
-                    sum += *m++ * float16_to_float32(*kptr++);
-#else
-                    sum += *m++ * *kptr++;
-#endif
-                }
-
-#if __SSE2__
-#if __AVX__
-                _suml = _mm_add_ps(_suml, _mm256_extractf128_ps(_sum, 1));
-                _suml = _mm_add_ps(_suml, _mm256_castps256_ps128(_sum));
-#endif // __AVX__
-                sum += _mm_reduce_add_ps(_suml);
-#endif // __SSE2__
-
-                sum = activation_ss(sum, activation_type, activation_params);
-
-                outptr[0] = sum;
-                outptr += 1;
-            }
-        }
-    }
-#endif // NCNN_RUNTIME_CPU
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/innerproduct_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/innerproduct_x86.h
deleted file mode 100644
index 19da245f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/innerproduct_x86.h
+++ /dev/null
@@ -1,54 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INNERPRODUCT_X86_H
-#define LAYER_INNERPRODUCT_X86_H
-
-#include "innerproduct.h"
-
-namespace ncnn {
-
-class InnerProduct_x86 : public InnerProduct
-{
-public:
-    InnerProduct_x86();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-#if NCNN_F16C && __AVX__
-    int create_pipeline_fp16s(const Option& opt);
-    int forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-#if NCNN_INT8
-    int create_pipeline_int8_x86(const Option& opt);
-    int forward_int8_x86(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-#endif
-
-public:
-    Layer* flatten;
-
-    Mat weight_data_tm;
-
-#if NCNN_INT8
-    Mat scale_in_data;
-#endif
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INNERPRODUCT_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic.h
deleted file mode 100644
index e327df71..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic.h
+++ /dev/null
@@ -1,310 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static inline void interpolate_cubic(float fx, float* coeffs)
-{
-    const float A = -0.75f;
-
-    float fx0 = fx + 1;
-    float fx1 = fx;
-    float fx2 = 1 - fx;
-    // float fx3 = 2 - fx;
-
-    coeffs[0] = A * fx0 * fx0 * fx0 - 5 * A * fx0 * fx0 + 8 * A * fx0 - 4 * A;
-    coeffs[1] = (A + 2) * fx1 * fx1 * fx1 - (A + 3) * fx1 * fx1 + 1;
-    coeffs[2] = (A + 2) * fx2 * fx2 * fx2 - (A + 3) * fx2 * fx2 + 1;
-    coeffs[3] = 1.f - coeffs[0] - coeffs[1] - coeffs[2];
-}
-
-static void cubic_coeffs(int w, int outw, int* xofs, float* alpha, int align_corner)
-{
-    double scale = (double)w / outw;
-    if (align_corner)
-    {
-        scale = (double)(w - 1) / (outw - 1);
-    }
-
-    for (int dx = 0; dx < outw; dx++)
-    {
-        float fx = (float)((dx + 0.5) * scale - 0.5);
-        if (align_corner)
-        {
-            fx = (float)(dx * scale);
-        }
-
-        int sx = static_cast<int>(floor(fx));
-        fx -= sx;
-
-        interpolate_cubic(fx, alpha + dx * 4);
-
-        if (sx <= -1)
-        {
-            sx = 1;
-            alpha[dx * 4 + 0] = 1.f - alpha[dx * 4 + 3];
-            alpha[dx * 4 + 1] = alpha[dx * 4 + 3];
-            alpha[dx * 4 + 2] = 0.f;
-            alpha[dx * 4 + 3] = 0.f;
-        }
-        if (sx == 0)
-        {
-            sx = 1;
-            alpha[dx * 4 + 0] = alpha[dx * 4 + 0] + alpha[dx * 4 + 1];
-            alpha[dx * 4 + 1] = alpha[dx * 4 + 2];
-            alpha[dx * 4 + 2] = alpha[dx * 4 + 3];
-            alpha[dx * 4 + 3] = 0.f;
-        }
-        if (sx == w - 2)
-        {
-            sx = w - 3;
-            alpha[dx * 4 + 3] = alpha[dx * 4 + 2] + alpha[dx * 4 + 3];
-            alpha[dx * 4 + 2] = alpha[dx * 4 + 1];
-            alpha[dx * 4 + 1] = alpha[dx * 4 + 0];
-            alpha[dx * 4 + 0] = 0.f;
-        }
-        if (sx >= w - 1)
-        {
-            sx = w - 3;
-            alpha[dx * 4 + 3] = 1.f - alpha[dx * 4 + 0];
-            alpha[dx * 4 + 2] = alpha[dx * 4 + 0];
-            alpha[dx * 4 + 1] = 0.f;
-            alpha[dx * 4 + 0] = 0.f;
-        }
-
-        xofs[dx] = sx;
-    }
-}
-
-static void resize_bicubic_image(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    Mat rowsbuf2(w);
-    Mat rowsbuf3(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows1p[dx] = S1p[-1] * a0 + S1p[0] * a1 + S1p[1] * a2 + S1p[2] * a3;
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const float* S0 = src.row(sy - 1);
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                float a2 = alphap[2];
-                float a3 = alphap[3];
-                rows0p[dx] = S0p[-1] * a0 + S0p[0] * a1 + S0p[1] * a2 + S0p[2] * a3;
-                rows1p[dx] = S1p[-1] * a0 + S1p[0] * a1 + S1p[1] * a2 + S1p[2] * a3;
-                rows2p[dx] = S2p[-1] * a0 + S2p[0] * a1 + S2p[1] * a2 + S2p[2] * a3;
-                rows3p[dx] = S3p[-1] * a0 + S3p[0] * a1 + S3p[1] * a2 + S3p[2] * a3;
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-        float b2 = beta[2];
-        float b3 = beta[3];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        float* Dp = dst.row(dy);
-
-        int dx = 0;
-#if __SSE2__
-#if __AVX__
-        __m256 _b0_256 = _mm256_set1_ps(b0);
-        __m256 _b1_256 = _mm256_set1_ps(b1);
-        __m256 _b2_256 = _mm256_set1_ps(b2);
-        __m256 _b3_256 = _mm256_set1_ps(b3);
-        for (; dx + 7 < w; dx += 8)
-        {
-            __m256 _rows0 = _mm256_loadu_ps(rows0p);
-            __m256 _rows1 = _mm256_loadu_ps(rows1p);
-            __m256 _rows2 = _mm256_loadu_ps(rows2p);
-            __m256 _rows3 = _mm256_loadu_ps(rows3p);
-            __m256 _Dp = _mm256_mul_ps(_rows0, _b0_256);
-            _Dp = _mm256_comp_fmadd_ps(_rows1, _b1_256, _Dp);
-            _Dp = _mm256_comp_fmadd_ps(_rows2, _b2_256, _Dp);
-            _Dp = _mm256_comp_fmadd_ps(_rows3, _b3_256, _Dp);
-            _mm256_storeu_ps(Dp, _Dp);
-
-            Dp += 8;
-            rows0p += 8;
-            rows1p += 8;
-            rows2p += 8;
-            rows3p += 8;
-        }
-#endif // __AVX__
-        __m128 _b0_128 = _mm_set1_ps(b0);
-        __m128 _b1_128 = _mm_set1_ps(b1);
-        __m128 _b2_128 = _mm_set1_ps(b2);
-        __m128 _b3_128 = _mm_set1_ps(b3);
-        for (; dx + 3 < w; dx += 4)
-        {
-            __m128 _rows0 = _mm_loadu_ps(rows0p);
-            __m128 _rows1 = _mm_loadu_ps(rows1p);
-            __m128 _rows2 = _mm_loadu_ps(rows2p);
-            __m128 _rows3 = _mm_loadu_ps(rows3p);
-            __m128 _Dp = _mm_mul_ps(_rows0, _b0_128);
-            _Dp = _mm_comp_fmadd_ps(_rows1, _b1_128, _Dp);
-            _Dp = _mm_comp_fmadd_ps(_rows2, _b2_128, _Dp);
-            _Dp = _mm_comp_fmadd_ps(_rows3, _b3_128, _Dp);
-            _mm_storeu_ps(Dp, _Dp);
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-            rows2p += 4;
-            rows3p += 4;
-        }
-#endif // __SSE2__
-        for (; dx < w; dx++)
-        {
-            *Dp++ = *rows0p++ * b0 + *rows1p++ * b1 + *rows2p++ * b2 + *rows3p++ * b3;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic_pack16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic_pack16.h
deleted file mode 100644
index 4c6eb14e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic_pack16.h
+++ /dev/null
@@ -1,286 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bicubic_image_pack16(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)16 * 4u, 16);
-    Mat rowsbuf1(w, (size_t)16 * 4u, 16);
-    Mat rowsbuf2(w, (size_t)16 * 4u, 16);
-    Mat rowsbuf3(w, (size_t)16 * 4u, 16);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 16;
-                const float* S3p = S3 + sx;
-
-                __m512 _a0 = _mm512_set1_ps(alphap[0]);
-                __m512 _a1 = _mm512_set1_ps(alphap[1]);
-                __m512 _a2 = _mm512_set1_ps(alphap[2]);
-                __m512 _a3 = _mm512_set1_ps(alphap[3]);
-
-                __m512 _S30 = _mm512_load_ps(S3p - 16);
-                __m512 _S31 = _mm512_load_ps(S3p + 0);
-                __m512 _S32 = _mm512_load_ps(S3p + 16);
-                __m512 _S33 = _mm512_load_ps(S3p + 32);
-                __m512 _rows3 = _mm512_mul_ps(_S30, _a0);
-                _rows3 = _mm512_fmadd_ps(_S31, _a1, _rows3);
-                _rows3 = _mm512_fmadd_ps(_S32, _a2, _rows3);
-                _rows3 = _mm512_fmadd_ps(_S33, _a3, _rows3);
-                _mm512_store_ps(rows3p + dx * 16, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 16;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                __m512 _a0 = _mm512_set1_ps(alphap[0]);
-                __m512 _a1 = _mm512_set1_ps(alphap[1]);
-                __m512 _a2 = _mm512_set1_ps(alphap[2]);
-                __m512 _a3 = _mm512_set1_ps(alphap[3]);
-
-                __m512 _S20 = _mm512_load_ps(S2p - 16);
-                __m512 _S21 = _mm512_load_ps(S2p + 0);
-                __m512 _S22 = _mm512_load_ps(S2p + 16);
-                __m512 _S23 = _mm512_load_ps(S2p + 32);
-                __m512 _S30 = _mm512_load_ps(S3p - 16);
-                __m512 _S31 = _mm512_load_ps(S3p + 0);
-                __m512 _S32 = _mm512_load_ps(S3p + 16);
-                __m512 _S33 = _mm512_load_ps(S3p + 32);
-                __m512 _rows2 = _mm512_mul_ps(_S20, _a0);
-                __m512 _rows3 = _mm512_mul_ps(_S30, _a0);
-                _rows2 = _mm512_fmadd_ps(_S21, _a1, _rows2);
-                _rows3 = _mm512_fmadd_ps(_S31, _a1, _rows3);
-                _rows2 = _mm512_fmadd_ps(_S22, _a2, _rows2);
-                _rows3 = _mm512_fmadd_ps(_S32, _a2, _rows3);
-                _rows2 = _mm512_fmadd_ps(_S23, _a3, _rows2);
-                _rows3 = _mm512_fmadd_ps(_S33, _a3, _rows3);
-                _mm512_store_ps(rows2p + dx * 16, _rows2);
-                _mm512_store_ps(rows3p + dx * 16, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 16;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                __m512 _a0 = _mm512_set1_ps(alphap[0]);
-                __m512 _a1 = _mm512_set1_ps(alphap[1]);
-                __m512 _a2 = _mm512_set1_ps(alphap[2]);
-                __m512 _a3 = _mm512_set1_ps(alphap[3]);
-
-                __m512 _S10 = _mm512_load_ps(S1p - 16);
-                __m512 _S11 = _mm512_load_ps(S1p + 0);
-                __m512 _S12 = _mm512_load_ps(S1p + 16);
-                __m512 _S13 = _mm512_load_ps(S1p + 32);
-                __m512 _S20 = _mm512_load_ps(S2p - 16);
-                __m512 _S21 = _mm512_load_ps(S2p + 0);
-                __m512 _S22 = _mm512_load_ps(S2p + 16);
-                __m512 _S23 = _mm512_load_ps(S2p + 32);
-                __m512 _S30 = _mm512_load_ps(S3p - 16);
-                __m512 _S31 = _mm512_load_ps(S3p + 0);
-                __m512 _S32 = _mm512_load_ps(S3p + 16);
-                __m512 _S33 = _mm512_load_ps(S3p + 32);
-                __m512 _rows1 = _mm512_mul_ps(_S10, _a0);
-                __m512 _rows2 = _mm512_mul_ps(_S20, _a0);
-                __m512 _rows3 = _mm512_mul_ps(_S30, _a0);
-                _rows1 = _mm512_fmadd_ps(_S11, _a1, _rows1);
-                _rows2 = _mm512_fmadd_ps(_S21, _a1, _rows2);
-                _rows3 = _mm512_fmadd_ps(_S31, _a1, _rows3);
-                _rows1 = _mm512_fmadd_ps(_S12, _a2, _rows1);
-                _rows2 = _mm512_fmadd_ps(_S22, _a2, _rows2);
-                _rows3 = _mm512_fmadd_ps(_S32, _a2, _rows3);
-                _rows1 = _mm512_fmadd_ps(_S13, _a3, _rows1);
-                _rows2 = _mm512_fmadd_ps(_S23, _a3, _rows2);
-                _rows3 = _mm512_fmadd_ps(_S33, _a3, _rows3);
-                _mm512_store_ps(rows1p + dx * 16, _rows1);
-                _mm512_store_ps(rows2p + dx * 16, _rows2);
-                _mm512_store_ps(rows3p + dx * 16, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const float* S0 = src.row(sy - 1);
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 16;
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                __m512 _a0 = _mm512_set1_ps(alphap[0]);
-                __m512 _a1 = _mm512_set1_ps(alphap[1]);
-                __m512 _a2 = _mm512_set1_ps(alphap[2]);
-                __m512 _a3 = _mm512_set1_ps(alphap[3]);
-
-                __m512 _S00 = _mm512_load_ps(S0p - 16);
-                __m512 _S01 = _mm512_load_ps(S0p + 0);
-                __m512 _S02 = _mm512_load_ps(S0p + 16);
-                __m512 _S03 = _mm512_load_ps(S0p + 32);
-                __m512 _S10 = _mm512_load_ps(S1p - 16);
-                __m512 _S11 = _mm512_load_ps(S1p + 0);
-                __m512 _S12 = _mm512_load_ps(S1p + 16);
-                __m512 _S13 = _mm512_load_ps(S1p + 32);
-                __m512 _S20 = _mm512_load_ps(S2p - 16);
-                __m512 _S21 = _mm512_load_ps(S2p + 0);
-                __m512 _S22 = _mm512_load_ps(S2p + 16);
-                __m512 _S23 = _mm512_load_ps(S2p + 32);
-                __m512 _S30 = _mm512_load_ps(S3p - 16);
-                __m512 _S31 = _mm512_load_ps(S3p + 0);
-                __m512 _S32 = _mm512_load_ps(S3p + 16);
-                __m512 _S33 = _mm512_load_ps(S3p + 32);
-                __m512 _rows0 = _mm512_mul_ps(_S00, _a0);
-                __m512 _rows1 = _mm512_mul_ps(_S10, _a0);
-                __m512 _rows2 = _mm512_mul_ps(_S20, _a0);
-                __m512 _rows3 = _mm512_mul_ps(_S30, _a0);
-                _rows0 = _mm512_fmadd_ps(_S01, _a1, _rows0);
-                _rows1 = _mm512_fmadd_ps(_S11, _a1, _rows1);
-                _rows2 = _mm512_fmadd_ps(_S21, _a1, _rows2);
-                _rows3 = _mm512_fmadd_ps(_S31, _a1, _rows3);
-                _rows0 = _mm512_fmadd_ps(_S02, _a2, _rows0);
-                _rows1 = _mm512_fmadd_ps(_S12, _a2, _rows1);
-                _rows2 = _mm512_fmadd_ps(_S22, _a2, _rows2);
-                _rows3 = _mm512_fmadd_ps(_S32, _a2, _rows3);
-                _rows0 = _mm512_fmadd_ps(_S03, _a3, _rows0);
-                _rows1 = _mm512_fmadd_ps(_S13, _a3, _rows1);
-                _rows2 = _mm512_fmadd_ps(_S23, _a3, _rows2);
-                _rows3 = _mm512_fmadd_ps(_S33, _a3, _rows3);
-                _mm512_store_ps(rows0p + dx * 16, _rows0);
-                _mm512_store_ps(rows1p + dx * 16, _rows1);
-                _mm512_store_ps(rows2p + dx * 16, _rows2);
-                _mm512_store_ps(rows3p + dx * 16, _rows3);
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        __m512 _b0 = _mm512_set1_ps(beta[0]);
-        __m512 _b1 = _mm512_set1_ps(beta[1]);
-        __m512 _b2 = _mm512_set1_ps(beta[2]);
-        __m512 _b3 = _mm512_set1_ps(beta[3]);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        float* Dp = dst.row(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            __m512 _rows0 = _mm512_load_ps(rows0p);
-            __m512 _rows1 = _mm512_load_ps(rows1p);
-            __m512 _rows2 = _mm512_load_ps(rows2p);
-            __m512 _rows3 = _mm512_load_ps(rows3p);
-            __m512 _Dp = _mm512_mul_ps(_rows0, _b0);
-            _Dp = _mm512_fmadd_ps(_rows1, _b1, _Dp);
-            _Dp = _mm512_fmadd_ps(_rows2, _b2, _Dp);
-            _Dp = _mm512_fmadd_ps(_rows3, _b3, _Dp);
-            _mm512_store_ps(Dp, _Dp);
-
-            Dp += 16;
-            rows0p += 16;
-            rows1p += 16;
-            rows2p += 16;
-            rows3p += 16;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic_pack4.h
deleted file mode 100644
index de165a79..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic_pack4.h
+++ /dev/null
@@ -1,286 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bicubic_image_pack4(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf1(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf2(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf3(w, (size_t)4 * 4u, 4);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S3p = S3 + sx;
-
-                __m128 _a0 = _mm_set1_ps(alphap[0]);
-                __m128 _a1 = _mm_set1_ps(alphap[1]);
-                __m128 _a2 = _mm_set1_ps(alphap[2]);
-                __m128 _a3 = _mm_set1_ps(alphap[3]);
-
-                __m128 _S30 = _mm_load_ps(S3p - 4);
-                __m128 _S31 = _mm_load_ps(S3p + 0);
-                __m128 _S32 = _mm_load_ps(S3p + 4);
-                __m128 _S33 = _mm_load_ps(S3p + 8);
-                __m128 _rows3 = _mm_mul_ps(_S30, _a0);
-                _rows3 = _mm_comp_fmadd_ps(_S31, _a1, _rows3);
-                _rows3 = _mm_comp_fmadd_ps(_S32, _a2, _rows3);
-                _rows3 = _mm_comp_fmadd_ps(_S33, _a3, _rows3);
-                _mm_store_ps(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                __m128 _a0 = _mm_set1_ps(alphap[0]);
-                __m128 _a1 = _mm_set1_ps(alphap[1]);
-                __m128 _a2 = _mm_set1_ps(alphap[2]);
-                __m128 _a3 = _mm_set1_ps(alphap[3]);
-
-                __m128 _S20 = _mm_load_ps(S2p - 4);
-                __m128 _S21 = _mm_load_ps(S2p + 0);
-                __m128 _S22 = _mm_load_ps(S2p + 4);
-                __m128 _S23 = _mm_load_ps(S2p + 8);
-                __m128 _S30 = _mm_load_ps(S3p - 4);
-                __m128 _S31 = _mm_load_ps(S3p + 0);
-                __m128 _S32 = _mm_load_ps(S3p + 4);
-                __m128 _S33 = _mm_load_ps(S3p + 8);
-                __m128 _rows2 = _mm_mul_ps(_S20, _a0);
-                __m128 _rows3 = _mm_mul_ps(_S30, _a0);
-                _rows2 = _mm_comp_fmadd_ps(_S21, _a1, _rows2);
-                _rows3 = _mm_comp_fmadd_ps(_S31, _a1, _rows3);
-                _rows2 = _mm_comp_fmadd_ps(_S22, _a2, _rows2);
-                _rows3 = _mm_comp_fmadd_ps(_S32, _a2, _rows3);
-                _rows2 = _mm_comp_fmadd_ps(_S23, _a3, _rows2);
-                _rows3 = _mm_comp_fmadd_ps(_S33, _a3, _rows3);
-                _mm_store_ps(rows2p + dx * 4, _rows2);
-                _mm_store_ps(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                __m128 _a0 = _mm_set1_ps(alphap[0]);
-                __m128 _a1 = _mm_set1_ps(alphap[1]);
-                __m128 _a2 = _mm_set1_ps(alphap[2]);
-                __m128 _a3 = _mm_set1_ps(alphap[3]);
-
-                __m128 _S10 = _mm_load_ps(S1p - 4);
-                __m128 _S11 = _mm_load_ps(S1p + 0);
-                __m128 _S12 = _mm_load_ps(S1p + 4);
-                __m128 _S13 = _mm_load_ps(S1p + 8);
-                __m128 _S20 = _mm_load_ps(S2p - 4);
-                __m128 _S21 = _mm_load_ps(S2p + 0);
-                __m128 _S22 = _mm_load_ps(S2p + 4);
-                __m128 _S23 = _mm_load_ps(S2p + 8);
-                __m128 _S30 = _mm_load_ps(S3p - 4);
-                __m128 _S31 = _mm_load_ps(S3p + 0);
-                __m128 _S32 = _mm_load_ps(S3p + 4);
-                __m128 _S33 = _mm_load_ps(S3p + 8);
-                __m128 _rows1 = _mm_mul_ps(_S10, _a0);
-                __m128 _rows2 = _mm_mul_ps(_S20, _a0);
-                __m128 _rows3 = _mm_mul_ps(_S30, _a0);
-                _rows1 = _mm_comp_fmadd_ps(_S11, _a1, _rows1);
-                _rows2 = _mm_comp_fmadd_ps(_S21, _a1, _rows2);
-                _rows3 = _mm_comp_fmadd_ps(_S31, _a1, _rows3);
-                _rows1 = _mm_comp_fmadd_ps(_S12, _a2, _rows1);
-                _rows2 = _mm_comp_fmadd_ps(_S22, _a2, _rows2);
-                _rows3 = _mm_comp_fmadd_ps(_S32, _a2, _rows3);
-                _rows1 = _mm_comp_fmadd_ps(_S13, _a3, _rows1);
-                _rows2 = _mm_comp_fmadd_ps(_S23, _a3, _rows2);
-                _rows3 = _mm_comp_fmadd_ps(_S33, _a3, _rows3);
-                _mm_store_ps(rows1p + dx * 4, _rows1);
-                _mm_store_ps(rows2p + dx * 4, _rows2);
-                _mm_store_ps(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const float* S0 = src.row(sy - 1);
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                __m128 _a0 = _mm_set1_ps(alphap[0]);
-                __m128 _a1 = _mm_set1_ps(alphap[1]);
-                __m128 _a2 = _mm_set1_ps(alphap[2]);
-                __m128 _a3 = _mm_set1_ps(alphap[3]);
-
-                __m128 _S00 = _mm_load_ps(S0p - 4);
-                __m128 _S01 = _mm_load_ps(S0p + 0);
-                __m128 _S02 = _mm_load_ps(S0p + 4);
-                __m128 _S03 = _mm_load_ps(S0p + 8);
-                __m128 _S10 = _mm_load_ps(S1p - 4);
-                __m128 _S11 = _mm_load_ps(S1p + 0);
-                __m128 _S12 = _mm_load_ps(S1p + 4);
-                __m128 _S13 = _mm_load_ps(S1p + 8);
-                __m128 _S20 = _mm_load_ps(S2p - 4);
-                __m128 _S21 = _mm_load_ps(S2p + 0);
-                __m128 _S22 = _mm_load_ps(S2p + 4);
-                __m128 _S23 = _mm_load_ps(S2p + 8);
-                __m128 _S30 = _mm_load_ps(S3p - 4);
-                __m128 _S31 = _mm_load_ps(S3p + 0);
-                __m128 _S32 = _mm_load_ps(S3p + 4);
-                __m128 _S33 = _mm_load_ps(S3p + 8);
-                __m128 _rows0 = _mm_mul_ps(_S00, _a0);
-                __m128 _rows1 = _mm_mul_ps(_S10, _a0);
-                __m128 _rows2 = _mm_mul_ps(_S20, _a0);
-                __m128 _rows3 = _mm_mul_ps(_S30, _a0);
-                _rows0 = _mm_comp_fmadd_ps(_S01, _a1, _rows0);
-                _rows1 = _mm_comp_fmadd_ps(_S11, _a1, _rows1);
-                _rows2 = _mm_comp_fmadd_ps(_S21, _a1, _rows2);
-                _rows3 = _mm_comp_fmadd_ps(_S31, _a1, _rows3);
-                _rows0 = _mm_comp_fmadd_ps(_S02, _a2, _rows0);
-                _rows1 = _mm_comp_fmadd_ps(_S12, _a2, _rows1);
-                _rows2 = _mm_comp_fmadd_ps(_S22, _a2, _rows2);
-                _rows3 = _mm_comp_fmadd_ps(_S32, _a2, _rows3);
-                _rows0 = _mm_comp_fmadd_ps(_S03, _a3, _rows0);
-                _rows1 = _mm_comp_fmadd_ps(_S13, _a3, _rows1);
-                _rows2 = _mm_comp_fmadd_ps(_S23, _a3, _rows2);
-                _rows3 = _mm_comp_fmadd_ps(_S33, _a3, _rows3);
-                _mm_store_ps(rows0p + dx * 4, _rows0);
-                _mm_store_ps(rows1p + dx * 4, _rows1);
-                _mm_store_ps(rows2p + dx * 4, _rows2);
-                _mm_store_ps(rows3p + dx * 4, _rows3);
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        __m128 _b0 = _mm_set1_ps(beta[0]);
-        __m128 _b1 = _mm_set1_ps(beta[1]);
-        __m128 _b2 = _mm_set1_ps(beta[2]);
-        __m128 _b3 = _mm_set1_ps(beta[3]);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        float* Dp = dst.row(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            __m128 _rows0 = _mm_load_ps(rows0p);
-            __m128 _rows1 = _mm_load_ps(rows1p);
-            __m128 _rows2 = _mm_load_ps(rows2p);
-            __m128 _rows3 = _mm_load_ps(rows3p);
-            __m128 _Dp = _mm_mul_ps(_rows0, _b0);
-            _Dp = _mm_comp_fmadd_ps(_rows1, _b1, _Dp);
-            _Dp = _mm_comp_fmadd_ps(_rows2, _b2, _Dp);
-            _Dp = _mm_comp_fmadd_ps(_rows3, _b3, _Dp);
-            _mm_store_ps(Dp, _Dp);
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-            rows2p += 4;
-            rows3p += 4;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic_pack8.h
deleted file mode 100644
index e8cea178..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bicubic_pack8.h
+++ /dev/null
@@ -1,286 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bicubic_image_pack8(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)8 * 4u, 8);
-    Mat rowsbuf1(w, (size_t)8 * 4u, 8);
-    Mat rowsbuf2(w, (size_t)8 * 4u, 8);
-    Mat rowsbuf3(w, (size_t)8 * 4u, 8);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-    float* rows2 = rowsbuf2;
-    float* rows3 = rowsbuf3;
-
-    int prev_sy1 = -3;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows2;
-            rows2 = rows3;
-            rows3 = rows0_old;
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 8;
-                const float* S3p = S3 + sx;
-
-                __m256 _a0 = _mm256_set1_ps(alphap[0]);
-                __m256 _a1 = _mm256_set1_ps(alphap[1]);
-                __m256 _a2 = _mm256_set1_ps(alphap[2]);
-                __m256 _a3 = _mm256_set1_ps(alphap[3]);
-
-                __m256 _S30 = _mm256_load_ps(S3p - 8);
-                __m256 _S31 = _mm256_load_ps(S3p + 0);
-                __m256 _S32 = _mm256_load_ps(S3p + 8);
-                __m256 _S33 = _mm256_load_ps(S3p + 16);
-                __m256 _rows3 = _mm256_mul_ps(_S30, _a0);
-                _rows3 = _mm256_comp_fmadd_ps(_S31, _a1, _rows3);
-                _rows3 = _mm256_comp_fmadd_ps(_S32, _a2, _rows3);
-                _rows3 = _mm256_comp_fmadd_ps(_S33, _a3, _rows3);
-                _mm256_store_ps(rows3p + dx * 8, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 2)
-        {
-            // hresize two rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            rows0 = rows2;
-            rows1 = rows3;
-            rows2 = rows0_old;
-            rows3 = rows1_old;
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 8;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                __m256 _a0 = _mm256_set1_ps(alphap[0]);
-                __m256 _a1 = _mm256_set1_ps(alphap[1]);
-                __m256 _a2 = _mm256_set1_ps(alphap[2]);
-                __m256 _a3 = _mm256_set1_ps(alphap[3]);
-
-                __m256 _S20 = _mm256_load_ps(S2p - 8);
-                __m256 _S21 = _mm256_load_ps(S2p + 0);
-                __m256 _S22 = _mm256_load_ps(S2p + 8);
-                __m256 _S23 = _mm256_load_ps(S2p + 16);
-                __m256 _S30 = _mm256_load_ps(S3p - 8);
-                __m256 _S31 = _mm256_load_ps(S3p + 0);
-                __m256 _S32 = _mm256_load_ps(S3p + 8);
-                __m256 _S33 = _mm256_load_ps(S3p + 16);
-                __m256 _rows2 = _mm256_mul_ps(_S20, _a0);
-                __m256 _rows3 = _mm256_mul_ps(_S30, _a0);
-                _rows2 = _mm256_comp_fmadd_ps(_S21, _a1, _rows2);
-                _rows3 = _mm256_comp_fmadd_ps(_S31, _a1, _rows3);
-                _rows2 = _mm256_comp_fmadd_ps(_S22, _a2, _rows2);
-                _rows3 = _mm256_comp_fmadd_ps(_S32, _a2, _rows3);
-                _rows2 = _mm256_comp_fmadd_ps(_S23, _a3, _rows2);
-                _rows3 = _mm256_comp_fmadd_ps(_S33, _a3, _rows3);
-                _mm256_store_ps(rows2p + dx * 8, _rows2);
-                _mm256_store_ps(rows3p + dx * 8, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else if (sy == prev_sy1 + 3)
-        {
-            // hresize three rows
-            float* rows0_old = rows0;
-            float* rows1_old = rows1;
-            float* rows2_old = rows2;
-            rows0 = rows3;
-            rows1 = rows0_old;
-            rows2 = rows1_old;
-            rows3 = rows2_old;
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 8;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                __m256 _a0 = _mm256_set1_ps(alphap[0]);
-                __m256 _a1 = _mm256_set1_ps(alphap[1]);
-                __m256 _a2 = _mm256_set1_ps(alphap[2]);
-                __m256 _a3 = _mm256_set1_ps(alphap[3]);
-
-                __m256 _S10 = _mm256_load_ps(S1p - 8);
-                __m256 _S11 = _mm256_load_ps(S1p + 0);
-                __m256 _S12 = _mm256_load_ps(S1p + 8);
-                __m256 _S13 = _mm256_load_ps(S1p + 16);
-                __m256 _S20 = _mm256_load_ps(S2p - 8);
-                __m256 _S21 = _mm256_load_ps(S2p + 0);
-                __m256 _S22 = _mm256_load_ps(S2p + 8);
-                __m256 _S23 = _mm256_load_ps(S2p + 16);
-                __m256 _S30 = _mm256_load_ps(S3p - 8);
-                __m256 _S31 = _mm256_load_ps(S3p + 0);
-                __m256 _S32 = _mm256_load_ps(S3p + 8);
-                __m256 _S33 = _mm256_load_ps(S3p + 16);
-                __m256 _rows1 = _mm256_mul_ps(_S10, _a0);
-                __m256 _rows2 = _mm256_mul_ps(_S20, _a0);
-                __m256 _rows3 = _mm256_mul_ps(_S30, _a0);
-                _rows1 = _mm256_comp_fmadd_ps(_S11, _a1, _rows1);
-                _rows2 = _mm256_comp_fmadd_ps(_S21, _a1, _rows2);
-                _rows3 = _mm256_comp_fmadd_ps(_S31, _a1, _rows3);
-                _rows1 = _mm256_comp_fmadd_ps(_S12, _a2, _rows1);
-                _rows2 = _mm256_comp_fmadd_ps(_S22, _a2, _rows2);
-                _rows3 = _mm256_comp_fmadd_ps(_S32, _a2, _rows3);
-                _rows1 = _mm256_comp_fmadd_ps(_S13, _a3, _rows1);
-                _rows2 = _mm256_comp_fmadd_ps(_S23, _a3, _rows2);
-                _rows3 = _mm256_comp_fmadd_ps(_S33, _a3, _rows3);
-                _mm256_store_ps(rows1p + dx * 8, _rows1);
-                _mm256_store_ps(rows2p + dx * 8, _rows2);
-                _mm256_store_ps(rows3p + dx * 8, _rows3);
-
-                alphap += 4;
-            }
-        }
-        else
-        {
-            // hresize four rows
-            const float* S0 = src.row(sy - 1);
-            const float* S1 = src.row(sy);
-            const float* S2 = src.row(sy + 1);
-            const float* S3 = src.row(sy + 2);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            float* rows2p = rows2;
-            float* rows3p = rows3;
-            for (int dx = 0; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 8;
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-                const float* S2p = S2 + sx;
-                const float* S3p = S3 + sx;
-
-                __m256 _a0 = _mm256_set1_ps(alphap[0]);
-                __m256 _a1 = _mm256_set1_ps(alphap[1]);
-                __m256 _a2 = _mm256_set1_ps(alphap[2]);
-                __m256 _a3 = _mm256_set1_ps(alphap[3]);
-
-                __m256 _S00 = _mm256_load_ps(S0p - 8);
-                __m256 _S01 = _mm256_load_ps(S0p + 0);
-                __m256 _S02 = _mm256_load_ps(S0p + 8);
-                __m256 _S03 = _mm256_load_ps(S0p + 16);
-                __m256 _S10 = _mm256_load_ps(S1p - 8);
-                __m256 _S11 = _mm256_load_ps(S1p + 0);
-                __m256 _S12 = _mm256_load_ps(S1p + 8);
-                __m256 _S13 = _mm256_load_ps(S1p + 16);
-                __m256 _S20 = _mm256_load_ps(S2p - 8);
-                __m256 _S21 = _mm256_load_ps(S2p + 0);
-                __m256 _S22 = _mm256_load_ps(S2p + 8);
-                __m256 _S23 = _mm256_load_ps(S2p + 16);
-                __m256 _S30 = _mm256_load_ps(S3p - 8);
-                __m256 _S31 = _mm256_load_ps(S3p + 0);
-                __m256 _S32 = _mm256_load_ps(S3p + 8);
-                __m256 _S33 = _mm256_load_ps(S3p + 16);
-                __m256 _rows0 = _mm256_mul_ps(_S00, _a0);
-                __m256 _rows1 = _mm256_mul_ps(_S10, _a0);
-                __m256 _rows2 = _mm256_mul_ps(_S20, _a0);
-                __m256 _rows3 = _mm256_mul_ps(_S30, _a0);
-                _rows0 = _mm256_comp_fmadd_ps(_S01, _a1, _rows0);
-                _rows1 = _mm256_comp_fmadd_ps(_S11, _a1, _rows1);
-                _rows2 = _mm256_comp_fmadd_ps(_S21, _a1, _rows2);
-                _rows3 = _mm256_comp_fmadd_ps(_S31, _a1, _rows3);
-                _rows0 = _mm256_comp_fmadd_ps(_S02, _a2, _rows0);
-                _rows1 = _mm256_comp_fmadd_ps(_S12, _a2, _rows1);
-                _rows2 = _mm256_comp_fmadd_ps(_S22, _a2, _rows2);
-                _rows3 = _mm256_comp_fmadd_ps(_S32, _a2, _rows3);
-                _rows0 = _mm256_comp_fmadd_ps(_S03, _a3, _rows0);
-                _rows1 = _mm256_comp_fmadd_ps(_S13, _a3, _rows1);
-                _rows2 = _mm256_comp_fmadd_ps(_S23, _a3, _rows2);
-                _rows3 = _mm256_comp_fmadd_ps(_S33, _a3, _rows3);
-                _mm256_store_ps(rows0p + dx * 8, _rows0);
-                _mm256_store_ps(rows1p + dx * 8, _rows1);
-                _mm256_store_ps(rows2p + dx * 8, _rows2);
-                _mm256_store_ps(rows3p + dx * 8, _rows3);
-
-                alphap += 4;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        __m256 _b0 = _mm256_set1_ps(beta[0]);
-        __m256 _b1 = _mm256_set1_ps(beta[1]);
-        __m256 _b2 = _mm256_set1_ps(beta[2]);
-        __m256 _b3 = _mm256_set1_ps(beta[3]);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* rows2p = rows2;
-        float* rows3p = rows3;
-        float* Dp = dst.row(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            __m256 _rows0 = _mm256_load_ps(rows0p);
-            __m256 _rows1 = _mm256_load_ps(rows1p);
-            __m256 _rows2 = _mm256_load_ps(rows2p);
-            __m256 _rows3 = _mm256_load_ps(rows3p);
-            __m256 _Dp = _mm256_mul_ps(_rows0, _b0);
-            _Dp = _mm256_comp_fmadd_ps(_rows1, _b1, _Dp);
-            _Dp = _mm256_comp_fmadd_ps(_rows2, _b2, _Dp);
-            _Dp = _mm256_comp_fmadd_ps(_rows3, _b3, _Dp);
-            _mm256_store_ps(Dp, _Dp);
-
-            Dp += 8;
-            rows0p += 8;
-            rows1p += 8;
-            rows2p += 8;
-            rows3p += 8;
-        }
-
-        beta += 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear.h
deleted file mode 100644
index 39d2432a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear.h
+++ /dev/null
@@ -1,171 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void linear_coeffs(int w, int outw, int* xofs, float* alpha, int align_corner)
-{
-    double scale = (double)w / outw;
-    if (align_corner)
-    {
-        scale = (double)(w - 1) / (outw - 1);
-    }
-
-    for (int dx = 0; dx < outw; dx++)
-    {
-        float fx = (float)((dx + 0.5) * scale - 0.5);
-        if (align_corner)
-        {
-            fx = (float)(dx * scale);
-        }
-
-        int sx = (int)floorf(fx);
-        fx -= sx;
-
-        if (sx < 0)
-        {
-            sx = 0;
-            fx = 0.f;
-        }
-        if (sx >= w - 1)
-        {
-            sx = w - 2;
-            fx = 1.f;
-        }
-
-        xofs[dx] = sx;
-
-        alpha[dx * 2] = 1.f - fx;
-        alpha[dx * 2 + 1] = fx;
-    }
-}
-
-static void resize_bilinear_image(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w);
-    Mat rowsbuf1(w);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows1p[dx] = S1p[0] * a0 + S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const float* S0 = src.row(sy);
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx];
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-
-                float a0 = alphap[0];
-                float a1 = alphap[1];
-                rows0p[dx] = S0p[0] * a0 + S0p[1] * a1;
-                rows1p[dx] = S1p[0] * a0 + S1p[1] * a1;
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        float b0 = beta[0];
-        float b1 = beta[1];
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* Dp = dst.row(dy);
-
-        int dx = 0;
-#if __SSE2__
-#if __AVX__
-        __m256 _b0_256 = _mm256_set1_ps(b0);
-        __m256 _b1_256 = _mm256_set1_ps(b1);
-        for (; dx + 7 < w; dx += 8)
-        {
-            __m256 _rows0 = _mm256_loadu_ps(rows0p);
-            __m256 _rows1 = _mm256_loadu_ps(rows1p);
-            __m256 _Dp = _mm256_mul_ps(_rows0, _b0_256);
-            _Dp = _mm256_comp_fmadd_ps(_rows1, _b1_256, _Dp);
-            _mm256_storeu_ps(Dp, _Dp);
-
-            Dp += 8;
-            rows0p += 8;
-            rows1p += 8;
-        }
-#endif // __AVX__
-        __m128 _b0_128 = _mm_set1_ps(b0);
-        __m128 _b1_128 = _mm_set1_ps(b1);
-        for (; dx + 3 < w; dx += 4)
-        {
-            __m128 _rows0 = _mm_loadu_ps(rows0p);
-            __m128 _rows1 = _mm_loadu_ps(rows1p);
-            __m128 _Dp = _mm_mul_ps(_rows0, _b0_128);
-            _Dp = _mm_comp_fmadd_ps(_rows1, _b1_128, _Dp);
-            _mm_storeu_ps(Dp, _Dp);
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-        }
-#endif // __SSE2__
-        for (; dx < w; dx++)
-        {
-            *Dp++ = *rows0p++ * b0 + *rows1p++ * b1;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear_pack16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear_pack16.h
deleted file mode 100644
index 5157bbe5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear_pack16.h
+++ /dev/null
@@ -1,123 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bilinear_image_pack16(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)16 * 4u, 16);
-    Mat rowsbuf1(w, (size_t)16 * 4u, 16);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 16;
-                const float* S1p = S1 + sx;
-
-                __m512 _a0 = _mm512_set1_ps(alphap[0]);
-                __m512 _a1 = _mm512_set1_ps(alphap[1]);
-
-                __m512 _S10 = _mm512_load_ps(S1p);
-                __m512 _S11 = _mm512_load_ps(S1p + 16);
-                __m512 _rows1 = _mm512_mul_ps(_S10, _a0);
-                _rows1 = _mm512_fmadd_ps(_S11, _a1, _rows1);
-                _mm512_store_ps(rows1p + dx * 16, _rows1);
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const float* S0 = src.row(sy);
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 16;
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-
-                __m512 _a0 = _mm512_set1_ps(alphap[0]);
-                __m512 _a1 = _mm512_set1_ps(alphap[1]);
-
-                __m512 _S00 = _mm512_load_ps(S0p);
-                __m512 _S01 = _mm512_load_ps(S0p + 16);
-                __m512 _S10 = _mm512_load_ps(S1p);
-                __m512 _S11 = _mm512_load_ps(S1p + 16);
-                __m512 _rows0 = _mm512_mul_ps(_S00, _a0);
-                __m512 _rows1 = _mm512_mul_ps(_S10, _a0);
-                _rows0 = _mm512_fmadd_ps(_S01, _a1, _rows0);
-                _rows1 = _mm512_fmadd_ps(_S11, _a1, _rows1);
-                _mm512_store_ps(rows0p + dx * 16, _rows0);
-                _mm512_store_ps(rows1p + dx * 16, _rows1);
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        __m512 _b0 = _mm512_set1_ps(beta[0]);
-        __m512 _b1 = _mm512_set1_ps(beta[1]);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* Dp = dst.row(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            __m512 _rows0 = _mm512_load_ps(rows0p);
-            __m512 _rows1 = _mm512_load_ps(rows1p);
-            __m512 _Dp = _mm512_mul_ps(_rows0, _b0);
-            _Dp = _mm512_fmadd_ps(_rows1, _b1, _Dp);
-            _mm512_store_ps(Dp, _Dp);
-
-            Dp += 16;
-            rows0p += 16;
-            rows1p += 16;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear_pack4.h
deleted file mode 100644
index 3f1c7a1a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear_pack4.h
+++ /dev/null
@@ -1,123 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bilinear_image_pack4(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)4 * 4u, 4);
-    Mat rowsbuf1(w, (size_t)4 * 4u, 4);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S1p = S1 + sx;
-
-                __m128 _a0 = _mm_set1_ps(alphap[0]);
-                __m128 _a1 = _mm_set1_ps(alphap[1]);
-
-                __m128 _S10 = _mm_load_ps(S1p);
-                __m128 _S11 = _mm_load_ps(S1p + 4);
-                __m128 _rows1 = _mm_mul_ps(_S10, _a0);
-                _rows1 = _mm_comp_fmadd_ps(_S11, _a1, _rows1);
-                _mm_store_ps(rows1p + dx * 4, _rows1);
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const float* S0 = src.row(sy);
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 4;
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-
-                __m128 _a0 = _mm_set1_ps(alphap[0]);
-                __m128 _a1 = _mm_set1_ps(alphap[1]);
-
-                __m128 _S00 = _mm_load_ps(S0p);
-                __m128 _S01 = _mm_load_ps(S0p + 4);
-                __m128 _S10 = _mm_load_ps(S1p);
-                __m128 _S11 = _mm_load_ps(S1p + 4);
-                __m128 _rows0 = _mm_mul_ps(_S00, _a0);
-                __m128 _rows1 = _mm_mul_ps(_S10, _a0);
-                _rows0 = _mm_comp_fmadd_ps(_S01, _a1, _rows0);
-                _rows1 = _mm_comp_fmadd_ps(_S11, _a1, _rows1);
-                _mm_store_ps(rows0p + dx * 4, _rows0);
-                _mm_store_ps(rows1p + dx * 4, _rows1);
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        __m128 _b0 = _mm_set1_ps(beta[0]);
-        __m128 _b1 = _mm_set1_ps(beta[1]);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* Dp = dst.row(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            __m128 _rows0 = _mm_load_ps(rows0p);
-            __m128 _rows1 = _mm_load_ps(rows1p);
-            __m128 _Dp = _mm_mul_ps(_rows0, _b0);
-            _Dp = _mm_comp_fmadd_ps(_rows1, _b1, _Dp);
-            _mm_store_ps(Dp, _Dp);
-
-            Dp += 4;
-            rows0p += 4;
-            rows1p += 4;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear_pack8.h
deleted file mode 100644
index 41e7071d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_bilinear_pack8.h
+++ /dev/null
@@ -1,123 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void resize_bilinear_image_pack8(const Mat& src, Mat& dst, float* alpha, int* xofs, float* beta, int* yofs)
-{
-    int w = dst.w;
-    int h = dst.h;
-
-    // loop body
-    Mat rowsbuf0(w, (size_t)8 * 4u, 8);
-    Mat rowsbuf1(w, (size_t)8 * 4u, 8);
-    float* rows0 = rowsbuf0;
-    float* rows1 = rowsbuf1;
-
-    int prev_sy1 = -2;
-
-    for (int dy = 0; dy < h; dy++)
-    {
-        int sy = yofs[dy];
-
-        if (sy == prev_sy1)
-        {
-            // reuse all rows
-        }
-        else if (sy == prev_sy1 + 1)
-        {
-            // hresize one row
-            float* rows0_old = rows0;
-            rows0 = rows1;
-            rows1 = rows0_old;
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 8;
-                const float* S1p = S1 + sx;
-
-                __m256 _a0 = _mm256_set1_ps(alphap[0]);
-                __m256 _a1 = _mm256_set1_ps(alphap[1]);
-
-                __m256 _S10 = _mm256_load_ps(S1p);
-                __m256 _S11 = _mm256_load_ps(S1p + 8);
-                __m256 _rows1 = _mm256_mul_ps(_S10, _a0);
-                _rows1 = _mm256_comp_fmadd_ps(_S11, _a1, _rows1);
-                _mm256_store_ps(rows1p + dx * 8, _rows1);
-
-                alphap += 2;
-            }
-        }
-        else
-        {
-            // hresize two rows
-            const float* S0 = src.row(sy);
-            const float* S1 = src.row(sy + 1);
-
-            const float* alphap = alpha;
-            float* rows0p = rows0;
-            float* rows1p = rows1;
-            int dx = 0;
-            for (; dx < w; dx++)
-            {
-                int sx = xofs[dx] * 8;
-                const float* S0p = S0 + sx;
-                const float* S1p = S1 + sx;
-
-                __m256 _a0 = _mm256_set1_ps(alphap[0]);
-                __m256 _a1 = _mm256_set1_ps(alphap[1]);
-
-                __m256 _S00 = _mm256_load_ps(S0p);
-                __m256 _S01 = _mm256_load_ps(S0p + 8);
-                __m256 _S10 = _mm256_load_ps(S1p);
-                __m256 _S11 = _mm256_load_ps(S1p + 8);
-                __m256 _rows0 = _mm256_mul_ps(_S00, _a0);
-                __m256 _rows1 = _mm256_mul_ps(_S10, _a0);
-                _rows0 = _mm256_comp_fmadd_ps(_S01, _a1, _rows0);
-                _rows1 = _mm256_comp_fmadd_ps(_S11, _a1, _rows1);
-                _mm256_store_ps(rows0p + dx * 8, _rows0);
-                _mm256_store_ps(rows1p + dx * 8, _rows1);
-
-                alphap += 2;
-            }
-        }
-
-        prev_sy1 = sy;
-
-        // vresize
-        __m256 _b0 = _mm256_set1_ps(beta[0]);
-        __m256 _b1 = _mm256_set1_ps(beta[1]);
-
-        float* rows0p = rows0;
-        float* rows1p = rows1;
-        float* Dp = dst.row(dy);
-
-        for (int dx = 0; dx < w; dx++)
-        {
-            __m256 _rows0 = _mm256_load_ps(rows0p);
-            __m256 _rows1 = _mm256_load_ps(rows1p);
-            __m256 _Dp = _mm256_mul_ps(_rows0, _b0);
-            _Dp = _mm256_comp_fmadd_ps(_rows1, _b1, _Dp);
-            _mm256_store_ps(Dp, _Dp);
-
-            Dp += 8;
-            rows0p += 8;
-            rows1p += 8;
-        }
-
-        beta += 2;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_x86.h
deleted file mode 100644
index 46fcde6f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/interp_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_INTERP_X86_H
-#define LAYER_INTERP_X86_H
-
-#include "interp.h"
-
-namespace ncnn {
-
-class Interp_x86 : public Interp
-{
-public:
-    Interp_x86();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_INTERP_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/layernorm_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/layernorm_x86.h
deleted file mode 100644
index 7e8ec058..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/layernorm_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_LAYERNORM_X86_H
-#define LAYER_LAYERNORM_X86_H
-
-#include "layernorm.h"
-
-namespace ncnn {
-
-class LayerNorm_x86 : public LayerNorm
-{
-public:
-    LayerNorm_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_LAYERNORM_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/lrn_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/lrn_x86.h
deleted file mode 100644
index 9aa85367..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/lrn_x86.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_LRN_X86_H
-#define LAYER_LRN_X86_H
-
-#include "lrn.h"
-
-namespace ncnn {
-
-class LRN_x86 : public LRN
-{
-public:
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_LRN_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/lstm_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/lstm_x86.h
deleted file mode 100644
index 1dc56d45..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/lstm_x86.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_LSTM_X86_H
-#define LAYER_LSTM_X86_H
-
-#include "lstm.h"
-
-namespace ncnn {
-
-class LSTM_x86 : public LSTM
-{
-public:
-    LSTM_x86();
-
-    virtual int create_pipeline(const Option& opt);
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    Mat weight_xc_data_packed;
-    Mat bias_c_data_packed;
-    Mat weight_hc_data_packed;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_LSTM_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/matmul_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/matmul_x86.h
deleted file mode 100644
index afbb85a7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/matmul_x86.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MATMUL_X86_H
-#define LAYER_MATMUL_X86_H
-
-#include "matmul.h"
-
-namespace ncnn {
-
-class MatMul_x86 : public MatMul
-{
-public:
-    MatMul_x86();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    Layer* gemm;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MATMUL_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/mish_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/mish_x86.h
deleted file mode 100644
index dce8823c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/mish_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MISH_X86_H
-#define LAYER_MISH_X86_H
-
-#include "mish.h"
-
-namespace ncnn {
-
-class Mish_x86 : public Mish
-{
-public:
-    Mish_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MISH_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/multiheadattention_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/multiheadattention_x86.h
deleted file mode 100644
index 55ea4178..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/multiheadattention_x86.h
+++ /dev/null
@@ -1,46 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_MULTIHEADATTENTION_X86_H
-#define LAYER_MULTIHEADATTENTION_X86_H
-
-#include "multiheadattention.h"
-
-namespace ncnn {
-
-class MultiHeadAttention_x86 : public MultiHeadAttention
-{
-public:
-    MultiHeadAttention_x86();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    Layer* q_gemm;
-    Layer* k_gemm;
-    Layer* v_gemm;
-    Layer* o_gemm;
-
-    Layer* qk_gemm;
-    Layer* qkv_gemm;
-
-    Layer* qk_softmax;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_MULTIHEADATTENTION_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/packing_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/packing_x86.h
deleted file mode 100644
index 9f8f3680..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/packing_x86.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PACKING_X86_H
-#define LAYER_PACKING_X86_H
-
-#include "packing.h"
-
-namespace ncnn {
-
-class Packing_x86 : public Packing
-{
-public:
-    Packing_x86();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PACKING_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack16.h
deleted file mode 100644
index e5e4017a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack16.h
+++ /dev/null
@@ -1,212 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void padding_constant_pack16_avx512(const Mat& src, Mat& dst, int top, int bottom, int left, int right, __m512 v)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-    int top_size = top * dst.w;
-    int bottom_size = bottom * dst.w;
-
-    // fill top
-    for (int y = 0; y < top_size; y++)
-    {
-        _mm512_store_ps(outptr, v);
-        outptr += 16;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            _mm512_store_ps(outptr, v);
-            outptr += 16;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _mm512_store_ps(outptr, _mm512_load_ps(ptr));
-            ptr += 16;
-            outptr += 16;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            _mm512_store_ps(outptr, v);
-            outptr += 16;
-        }
-    }
-    // fill top
-    for (int y = 0; y < bottom_size; y++)
-    {
-        _mm512_store_ps(outptr, v);
-        outptr += 16;
-    }
-}
-
-static void padding_replicate_pack16_avx512(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-
-    // fill top
-    for (int y = 0; y < top; y++)
-    {
-        const float* ptr0 = ptr;
-        __m512 _p = _mm512_load_ps(ptr0);
-        for (int x = 0; x < left; x++)
-        {
-            _mm512_store_ps(outptr, _p);
-            outptr += 16;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = _mm512_load_ps(ptr0);
-            _mm512_store_ps(outptr, _p);
-            ptr0 += 16;
-            outptr += 16;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            _mm512_store_ps(outptr, _p);
-            outptr += 16;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        __m512 _p = _mm512_load_ps(ptr);
-        for (int x = 0; x < left; x++)
-        {
-            _mm512_store_ps(outptr, _p);
-            outptr += 16;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = _mm512_load_ps(ptr);
-            _mm512_store_ps(outptr, _p);
-            ptr += 16;
-            outptr += 16;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            _mm512_store_ps(outptr, _p);
-            outptr += 16;
-        }
-    }
-    // fill bottom
-    ptr -= src.w * 16;
-    for (int y = 0; y < bottom; y++)
-    {
-        const float* ptr0 = ptr;
-        __m512 _p = _mm512_load_ps(ptr0);
-        for (int x = 0; x < left; x++)
-        {
-            _mm512_store_ps(outptr, _p);
-            outptr += 16;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = _mm512_load_ps(ptr0);
-            _mm512_store_ps(outptr, _p);
-            ptr0 += 16;
-            outptr += 16;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            _mm512_store_ps(outptr, _p);
-            outptr += 16;
-        }
-    }
-}
-
-static void padding_reflect_pack16_avx512(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-
-    // fill top
-    ptr += top * src.w * 16;
-    for (int y = 0; y < top; y++)
-    {
-        const float* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            __m512 _p = _mm512_load_ps(ptr0 + (left - x) * 16);
-            _mm512_store_ps(outptr, _p);
-            outptr += 16;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            __m512 _p = _mm512_load_ps(ptr0);
-            _mm512_store_ps(outptr, _p);
-            ptr0 += 16;
-            outptr += 16;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __m512 _p = _mm512_load_ps(ptr0 - 32 - x * 16);
-            _mm512_store_ps(outptr, _p);
-            outptr += 16;
-        }
-        ptr -= src.w * 16;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            __m512 _p = _mm512_load_ps(ptr + (left - x) * 16);
-            _mm512_store_ps(outptr, _p);
-            outptr += 16;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            __m512 _p = _mm512_load_ps(ptr);
-            _mm512_store_ps(outptr, _p);
-            ptr += 16;
-            outptr += 16;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __m512 _p = _mm512_load_ps(ptr - 32 - x * 16);
-            _mm512_store_ps(outptr, _p);
-            outptr += 16;
-        }
-    }
-    // fill bottom
-    ptr -= 2 * src.w * 16;
-    for (int y = 0; y < bottom; y++)
-    {
-        const float* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            __m512 _p = _mm512_load_ps(ptr0 + (left - x) * 16);
-            _mm512_store_ps(outptr, _p);
-            outptr += 16;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            __m512 _p = _mm512_load_ps(ptr0);
-            _mm512_store_ps(outptr, _p);
-            ptr0 += 16;
-            outptr += 16;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __m512 _p = _mm512_load_ps(ptr0 - 32 - x * 16);
-            _mm512_store_ps(outptr, _p);
-            outptr += 16;
-        }
-        ptr -= src.w * 16;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack4.h
deleted file mode 100644
index 7f4cad8e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack4.h
+++ /dev/null
@@ -1,212 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void padding_constant_pack4_sse(const Mat& src, Mat& dst, int top, int bottom, int left, int right, __m128 v)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-    int top_size = top * dst.w;
-    int bottom_size = bottom * dst.w;
-
-    // fill top
-    for (int y = 0; y < top_size; y++)
-    {
-        _mm_store_ps(outptr, v);
-        outptr += 4;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            _mm_store_ps(outptr, v);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _mm_store_ps(outptr, _mm_load_ps(ptr));
-            ptr += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            _mm_store_ps(outptr, v);
-            outptr += 4;
-        }
-    }
-    // fill top
-    for (int y = 0; y < bottom_size; y++)
-    {
-        _mm_store_ps(outptr, v);
-        outptr += 4;
-    }
-}
-
-static void padding_replicate_pack4_sse(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-
-    // fill top
-    for (int y = 0; y < top; y++)
-    {
-        const float* ptr0 = ptr;
-        __m128 _p = _mm_load_ps(ptr0);
-        for (int x = 0; x < left; x++)
-        {
-            _mm_store_ps(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = _mm_load_ps(ptr0);
-            _mm_store_ps(outptr, _p);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            _mm_store_ps(outptr, _p);
-            outptr += 4;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        __m128 _p = _mm_load_ps(ptr);
-        for (int x = 0; x < left; x++)
-        {
-            _mm_store_ps(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = _mm_load_ps(ptr);
-            _mm_store_ps(outptr, _p);
-            ptr += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            _mm_store_ps(outptr, _p);
-            outptr += 4;
-        }
-    }
-    // fill bottom
-    ptr -= src.w * 4;
-    for (int y = 0; y < bottom; y++)
-    {
-        const float* ptr0 = ptr;
-        __m128 _p = _mm_load_ps(ptr0);
-        for (int x = 0; x < left; x++)
-        {
-            _mm_store_ps(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = _mm_load_ps(ptr0);
-            _mm_store_ps(outptr, _p);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            _mm_store_ps(outptr, _p);
-            outptr += 4;
-        }
-    }
-}
-
-static void padding_reflect_pack4_sse(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-
-    // fill top
-    ptr += top * src.w * 4;
-    for (int y = 0; y < top; y++)
-    {
-        const float* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            __m128 _p = _mm_load_ps(ptr0 + (left - x) * 4);
-            _mm_store_ps(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            __m128 _p = _mm_load_ps(ptr0);
-            _mm_store_ps(outptr, _p);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __m128 _p = _mm_load_ps(ptr0 - 8 - x * 4);
-            _mm_store_ps(outptr, _p);
-            outptr += 4;
-        }
-        ptr -= src.w * 4;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            __m128 _p = _mm_load_ps(ptr + (left - x) * 4);
-            _mm_store_ps(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            __m128 _p = _mm_load_ps(ptr);
-            _mm_store_ps(outptr, _p);
-            ptr += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __m128 _p = _mm_load_ps(ptr - 8 - x * 4);
-            _mm_store_ps(outptr, _p);
-            outptr += 4;
-        }
-    }
-    // fill bottom
-    ptr -= 2 * src.w * 4;
-    for (int y = 0; y < bottom; y++)
-    {
-        const float* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            __m128 _p = _mm_load_ps(ptr0 + (left - x) * 4);
-            _mm_store_ps(outptr, _p);
-            outptr += 4;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            __m128 _p = _mm_load_ps(ptr0);
-            _mm_store_ps(outptr, _p);
-            ptr0 += 4;
-            outptr += 4;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __m128 _p = _mm_load_ps(ptr0 - 8 - x * 4);
-            _mm_store_ps(outptr, _p);
-            outptr += 4;
-        }
-        ptr -= src.w * 4;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack8.h
deleted file mode 100644
index 533f0e3e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack8.h
+++ /dev/null
@@ -1,212 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void padding_constant_pack8_avx(const Mat& src, Mat& dst, int top, int bottom, int left, int right, __m256 v)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-    int top_size = top * dst.w;
-    int bottom_size = bottom * dst.w;
-
-    // fill top
-    for (int y = 0; y < top_size; y++)
-    {
-        _mm256_store_ps(outptr, v);
-        outptr += 8;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            _mm256_store_ps(outptr, v);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _mm256_store_ps(outptr, _mm256_load_ps(ptr));
-            ptr += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            _mm256_store_ps(outptr, v);
-            outptr += 8;
-        }
-    }
-    // fill top
-    for (int y = 0; y < bottom_size; y++)
-    {
-        _mm256_store_ps(outptr, v);
-        outptr += 8;
-    }
-}
-
-static void padding_replicate_pack8_avx(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-
-    // fill top
-    for (int y = 0; y < top; y++)
-    {
-        const float* ptr0 = ptr;
-        __m256 _p = _mm256_load_ps(ptr0);
-        for (int x = 0; x < left; x++)
-        {
-            _mm256_store_ps(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = _mm256_load_ps(ptr0);
-            _mm256_store_ps(outptr, _p);
-            ptr0 += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            _mm256_store_ps(outptr, _p);
-            outptr += 8;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        __m256 _p = _mm256_load_ps(ptr);
-        for (int x = 0; x < left; x++)
-        {
-            _mm256_store_ps(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = _mm256_load_ps(ptr);
-            _mm256_store_ps(outptr, _p);
-            ptr += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            _mm256_store_ps(outptr, _p);
-            outptr += 8;
-        }
-    }
-    // fill bottom
-    ptr -= src.w * 8;
-    for (int y = 0; y < bottom; y++)
-    {
-        const float* ptr0 = ptr;
-        __m256 _p = _mm256_load_ps(ptr0);
-        for (int x = 0; x < left; x++)
-        {
-            _mm256_store_ps(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            _p = _mm256_load_ps(ptr0);
-            _mm256_store_ps(outptr, _p);
-            ptr0 += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            _mm256_store_ps(outptr, _p);
-            outptr += 8;
-        }
-    }
-}
-
-static void padding_reflect_pack8_avx(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const float* ptr = src;
-    float* outptr = dst;
-
-    // fill top
-    ptr += top * src.w * 8;
-    for (int y = 0; y < top; y++)
-    {
-        const float* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            __m256 _p = _mm256_load_ps(ptr0 + (left - x) * 8);
-            _mm256_store_ps(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            __m256 _p = _mm256_load_ps(ptr0);
-            _mm256_store_ps(outptr, _p);
-            ptr0 += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __m256 _p = _mm256_load_ps(ptr0 - 16 - x * 8);
-            _mm256_store_ps(outptr, _p);
-            outptr += 8;
-        }
-        ptr -= src.w * 8;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            __m256 _p = _mm256_load_ps(ptr + (left - x) * 8);
-            _mm256_store_ps(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            __m256 _p = _mm256_load_ps(ptr);
-            _mm256_store_ps(outptr, _p);
-            ptr += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __m256 _p = _mm256_load_ps(ptr - 16 - x * 8);
-            _mm256_store_ps(outptr, _p);
-            outptr += 8;
-        }
-    }
-    // fill bottom
-    ptr -= 2 * src.w * 8;
-    for (int y = 0; y < bottom; y++)
-    {
-        const float* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            __m256 _p = _mm256_load_ps(ptr0 + (left - x) * 8);
-            _mm256_store_ps(outptr, _p);
-            outptr += 8;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            __m256 _p = _mm256_load_ps(ptr0);
-            _mm256_store_ps(outptr, _p);
-            ptr0 += 8;
-            outptr += 8;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            __m256 _p = _mm256_load_ps(ptr0 - 16 - x * 8);
-            _mm256_store_ps(outptr, _p);
-            outptr += 8;
-        }
-        ptr -= src.w * 8;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack8_int8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack8_int8.h
deleted file mode 100644
index e2eb86ea..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_pack8_int8.h
+++ /dev/null
@@ -1,171 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void padding_constant_pack8_int8_sse(const Mat& src, Mat& dst, int top, int bottom, int left, int right, int64_t _v)
-{
-    const int64_t* ptr = src;
-    int64_t* outptr = dst;
-
-    // fill top
-    for (int y = 0; y < top; y++)
-    {
-        for (int x = 0; x < dst.w; x++)
-        {
-            *outptr++ = _v;
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = _v;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = _v;
-        }
-    }
-    // fill bottom
-    for (int y = 0; y < bottom; y++)
-    {
-        for (int x = 0; x < dst.w; x++)
-        {
-            *outptr++ = _v;
-        }
-    }
-}
-
-static void padding_replicate_pack8_int8_sse(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const int64_t* ptr = src;
-    int64_t* outptr = dst;
-
-    // fill top
-    for (int y = 0; y < top; y++)
-    {
-        const int64_t* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = *ptr0;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr0++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr0[-1];
-        }
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = *ptr;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr[-1];
-        }
-    }
-    // fill bottom
-    ptr -= src.w;
-    for (int y = 0; y < bottom; y++)
-    {
-        const int64_t* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = *ptr0;
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr0++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr0[-1];
-        }
-    }
-}
-
-static void padding_reflect_pack8_int8_sse(const Mat& src, Mat& dst, int top, int bottom, int left, int right)
-{
-    const int64_t* ptr = src;
-    int64_t* outptr = dst;
-
-    // fill top
-    ptr += top * src.w;
-    for (int y = 0; y < top; y++)
-    {
-        const int64_t* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = ptr0[left - x];
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr0++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr0[-2 - x];
-        }
-        ptr -= src.w;
-    }
-    // fill center
-    for (int y = 0; y < src.h; y++)
-    {
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = ptr[left - x];
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr[-2 - x];
-        }
-    }
-    // fill bottom
-    ptr -= 2 * src.w;
-    for (int y = 0; y < bottom; y++)
-    {
-        const int64_t* ptr0 = ptr;
-        for (int x = 0; x < left; x++)
-        {
-            *outptr++ = ptr0[left - x];
-        }
-        for (int x = 0; x < src.w; x++)
-        {
-            *outptr++ = *ptr0++;
-        }
-        for (int x = 0; x < right; x++)
-        {
-            *outptr++ = ptr0[-2 - x];
-        }
-        ptr -= src.w;
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_x86.h
deleted file mode 100644
index 8772fe30..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/padding_x86.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PADDING_X86_H
-#define LAYER_PADDING_X86_H
-
-#include "padding.h"
-
-namespace ncnn {
-
-class Padding_x86 : public Padding
-{
-public:
-    Padding_x86();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-
-protected:
-    int forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PADDING_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2.h
deleted file mode 100644
index d062f9b1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2.h
+++ /dev/null
@@ -1,83 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn
-// available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this
-// file except in compliance with the License. You may obtain a copy of the
-// License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
-// WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
-// License for the specific language governing permissions and limitations under
-// the License.
-
-static void pooling2x2s2_max_avx(const Mat& bottom_blob, Mat& top_blob,
-                                 const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = w - 2 * outw + w;
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const float* img0 = bottom_blob.channel(q);
-        float* outptr = top_blob.channel(q);
-        int outcount = 0;
-        const float* r0 = img0;
-        const float* r1 = img0 + w;
-#if __AVX2__
-        __m256i permute_mask = _mm256_setr_epi32(0, 2, 4, 6, 1, 3, 5, 7);
-#endif // __AVX__
-
-        for (int i = 0; i < outh; i++)
-        {
-#if __AVX2__
-            int nn = outw >> 2;
-            int remain = outw - (nn << 2);
-#else
-            int remain = outw;
-#endif // __AVX__
-
-#if __AVX2__
-            for (; nn > 0; nn--)
-            {
-                __m256 _r0 = _mm256_loadu_ps(r0);
-                __m256 _r1 = _mm256_loadu_ps(r1);
-                __m256 _max_r0_r1 = _mm256_max_ps(_r0, _r1);
-                _max_r0_r1 = _mm256_castsi256_ps(_mm256_permutevar8x32_epi32(
-                                                     _mm256_castps_si256(_max_r0_r1), permute_mask));
-                __m128 _max_0 = _mm256_extractf128_ps(_max_r0_r1, 0);
-                __m128 _max_1 = _mm256_extractf128_ps(_max_r0_r1, 1);
-                __m128 _max = _mm_max_ps(_max_0, _max_1);
-                _mm_storeu_ps(outptr, _max);
-                r0 += 8;
-                r1 += 8;
-                outptr += 4;
-                outcount += 4;
-            }
-#endif // __AVX__
-            for (; remain > 0; remain--)
-            {
-                float max0 = std::max(r0[0], r0[1]);
-                float max1 = std::max(r1[0], r1[1]);
-
-                *outptr = std::max(max0, max1);
-
-                r0 += 2;
-                r1 += 2;
-                outptr++;
-                outcount++;
-            }
-            r0 += tailstep;
-            r1 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2_pack16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2_pack16.h
deleted file mode 100644
index 5ce81d0e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2_pack16.h
+++ /dev/null
@@ -1,60 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pooling2x2s2_max_pack16_avx512(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * 16;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        float* outptr = top_blob.channel(q);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j < outw; j++)
-            {
-                __m512 _r00 = _mm512_loadu_ps(r0);
-                __m512 _r01 = _mm512_loadu_ps(r0 + 16);
-                __m512 _r10 = _mm512_loadu_ps(r1);
-                __m512 _r11 = _mm512_loadu_ps(r1 + 16);
-
-                __m512 _max0 = _mm512_max_ps(_r00, _r01);
-                __m512 _max1 = _mm512_max_ps(_r10, _r11);
-                __m512 _max = _mm512_max_ps(_max0, _max1);
-
-                _mm512_storeu_ps(outptr, _max);
-
-                r0 += 32;
-                r1 += 32;
-                outptr += 16;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2_pack4.h
deleted file mode 100644
index 3f3a0e5f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2_pack4.h
+++ /dev/null
@@ -1,61 +0,0 @@
-
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pooling2x2s2_max_pack4_sse(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        float* outptr = top_blob.channel(q);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j < outw; j++)
-            {
-                __m128 _r00 = _mm_loadu_ps(r0);
-                __m128 _r01 = _mm_loadu_ps(r0 + 4);
-                __m128 _r10 = _mm_loadu_ps(r1);
-                __m128 _r11 = _mm_loadu_ps(r1 + 4);
-
-                __m128 _max0 = _mm_max_ps(_r00, _r01);
-                __m128 _max1 = _mm_max_ps(_r10, _r11);
-                __m128 _max = _mm_max_ps(_max0, _max1);
-
-                _mm_storeu_ps(outptr, _max);
-
-                r0 += 8;
-                r1 += 8;
-                outptr += 4;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2_pack8.h
deleted file mode 100644
index eeb9f934..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_2x2_pack8.h
+++ /dev/null
@@ -1,61 +0,0 @@
-
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pooling2x2s2_max_pack8_avx(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * 8;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        float* outptr = top_blob.channel(q);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-
-            for (; j < outw; j++)
-            {
-                __m256 _r00 = _mm256_loadu_ps(r0);
-                __m256 _r01 = _mm256_loadu_ps(r0 + 8);
-                __m256 _r10 = _mm256_loadu_ps(r1);
-                __m256 _r11 = _mm256_loadu_ps(r1 + 8);
-
-                __m256 _max0 = _mm256_max_ps(_r00, _r01);
-                __m256 _max1 = _mm256_max_ps(_r10, _r11);
-                __m256 _max = _mm256_max_ps(_max0, _max1);
-
-                _mm256_storeu_ps(outptr, _max);
-
-                r0 += 16;
-                r1 += 16;
-                outptr += 8;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_3x3_pack16.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_3x3_pack16.h
deleted file mode 100644
index 70000b08..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_3x3_pack16.h
+++ /dev/null
@@ -1,193 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pooling3x3s2_max_pack16_avx512(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * 16;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        float* outptr = top_blob.channel(q);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                __m512 _r00 = _mm512_loadu_ps(r0);
-                __m512 _r01 = _mm512_loadu_ps(r0 + 16);
-                __m512 _r02 = _mm512_loadu_ps(r0 + 32);
-                __m512 _r10 = _mm512_loadu_ps(r1);
-                __m512 _r11 = _mm512_loadu_ps(r1 + 16);
-                __m512 _r12 = _mm512_loadu_ps(r1 + 32);
-                __m512 _r20 = _mm512_loadu_ps(r2);
-                __m512 _r21 = _mm512_loadu_ps(r2 + 16);
-                __m512 _r22 = _mm512_loadu_ps(r2 + 32);
-
-                __m512 _max00 = _mm512_max_ps(_r00, _r01);
-                _max00 = _mm512_max_ps(_max00, _r02);
-                _max00 = _mm512_max_ps(_max00, _r10);
-                _max00 = _mm512_max_ps(_max00, _r11);
-                __m512 _max01 = _mm512_max_ps(_r12, _r20);
-                _max01 = _mm512_max_ps(_max01, _r21);
-                _max01 = _mm512_max_ps(_max01, _r22);
-
-                __m512 _r03 = _mm512_loadu_ps(r0 + 48);
-                __m512 _r04 = _mm512_loadu_ps(r0 + 64);
-                __m512 _r13 = _mm512_loadu_ps(r1 + 48);
-                __m512 _r14 = _mm512_loadu_ps(r1 + 64);
-                __m512 _r23 = _mm512_loadu_ps(r2 + 48);
-                __m512 _r24 = _mm512_loadu_ps(r2 + 64);
-
-                _mm512_storeu_ps(outptr, _mm512_max_ps(_max00, _max01));
-
-                __m512 _max10 = _mm512_max_ps(_r03, _r04);
-                _max10 = _mm512_max_ps(_max10, _r02);
-                _max10 = _mm512_max_ps(_max10, _r13);
-                _max10 = _mm512_max_ps(_max10, _r14);
-                __m512 _max11 = _mm512_max_ps(_r12, _r23);
-                _max10 = _mm512_max_ps(_max10, _r24);
-                _max10 = _mm512_max_ps(_max10, _r22);
-
-                __m512 _r05 = _mm512_loadu_ps(r0 + 80);
-                __m512 _r06 = _mm512_loadu_ps(r0 + 96);
-                __m512 _r15 = _mm512_loadu_ps(r1 + 80);
-                __m512 _r16 = _mm512_loadu_ps(r1 + 96);
-                __m512 _r25 = _mm512_loadu_ps(r2 + 80);
-                __m512 _r26 = _mm512_loadu_ps(r2 + 96);
-
-                _mm512_storeu_ps(outptr + 16, _mm512_max_ps(_max10, _max11));
-
-                __m512 _max20 = _mm512_max_ps(_r05, _r06);
-                _max20 = _mm512_max_ps(_max20, _r04);
-                _max20 = _mm512_max_ps(_max20, _r15);
-                _max20 = _mm512_max_ps(_max20, _r16);
-                __m512 _max21 = _mm512_max_ps(_r14, _r25);
-                _max20 = _mm512_max_ps(_max20, _r26);
-                _max20 = _mm512_max_ps(_max20, _r24);
-
-                __m512 _r07 = _mm512_loadu_ps(r0 + 112);
-                __m512 _r08 = _mm512_loadu_ps(r0 + 128);
-                __m512 _r17 = _mm512_loadu_ps(r1 + 112);
-                __m512 _r18 = _mm512_loadu_ps(r1 + 128);
-                __m512 _r27 = _mm512_loadu_ps(r2 + 112);
-                __m512 _r28 = _mm512_loadu_ps(r2 + 128);
-
-                _mm512_storeu_ps(outptr + 32, _mm512_max_ps(_max20, _max21));
-
-                __m512 _max30 = _mm512_max_ps(_r07, _r08);
-                _max30 = _mm512_max_ps(_max30, _r06);
-                _max30 = _mm512_max_ps(_max30, _r17);
-                _max30 = _mm512_max_ps(_max30, _r18);
-                __m512 _max31 = _mm512_max_ps(_r16, _r27);
-                _max30 = _mm512_max_ps(_max30, _r28);
-                _max30 = _mm512_max_ps(_max30, _r26);
-
-                _mm512_storeu_ps(outptr + 48, _mm512_max_ps(_max30, _max31));
-
-                r0 += 128;
-                r1 += 128;
-                r2 += 128;
-                outptr += 64;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                __m512 _r00 = _mm512_loadu_ps(r0);
-                __m512 _r01 = _mm512_loadu_ps(r0 + 16);
-                __m512 _r02 = _mm512_loadu_ps(r0 + 32);
-                __m512 _r10 = _mm512_loadu_ps(r1);
-                __m512 _r11 = _mm512_loadu_ps(r1 + 16);
-                __m512 _r12 = _mm512_loadu_ps(r1 + 32);
-                __m512 _r20 = _mm512_loadu_ps(r2);
-                __m512 _r21 = _mm512_loadu_ps(r2 + 16);
-                __m512 _r22 = _mm512_loadu_ps(r2 + 32);
-
-                __m512 _max00 = _mm512_max_ps(_r00, _r01);
-                _max00 = _mm512_max_ps(_max00, _r02);
-                _max00 = _mm512_max_ps(_max00, _r10);
-                _max00 = _mm512_max_ps(_max00, _r11);
-                __m512 _max01 = _mm512_max_ps(_r12, _r20);
-                _max01 = _mm512_max_ps(_max01, _r21);
-                _max01 = _mm512_max_ps(_max01, _r22);
-
-                __m512 _r03 = _mm512_loadu_ps(r0 + 48);
-                __m512 _r04 = _mm512_loadu_ps(r0 + 64);
-                __m512 _r13 = _mm512_loadu_ps(r1 + 48);
-                __m512 _r14 = _mm512_loadu_ps(r1 + 64);
-                __m512 _r23 = _mm512_loadu_ps(r2 + 48);
-                __m512 _r24 = _mm512_loadu_ps(r2 + 64);
-
-                _mm512_storeu_ps(outptr, _mm512_max_ps(_max00, _max01));
-
-                __m512 _max10 = _mm512_max_ps(_r03, _r04);
-                _max10 = _mm512_max_ps(_max10, _r02);
-                _max10 = _mm512_max_ps(_max10, _r13);
-                _max10 = _mm512_max_ps(_max10, _r14);
-                __m512 _max11 = _mm512_max_ps(_r12, _r23);
-                _max10 = _mm512_max_ps(_max10, _r24);
-                _max10 = _mm512_max_ps(_max10, _r22);
-
-                _mm512_storeu_ps(outptr + 16, _mm512_max_ps(_max10, _max11));
-
-                r0 += 64;
-                r1 += 64;
-                r2 += 64;
-                outptr += 32;
-            }
-
-            for (; j < outw; j++)
-            {
-                __m512 _r00 = _mm512_loadu_ps(r0);
-                __m512 _r01 = _mm512_loadu_ps(r0 + 16);
-                __m512 _r02 = _mm512_loadu_ps(r0 + 32);
-                __m512 _r10 = _mm512_loadu_ps(r1);
-                __m512 _r11 = _mm512_loadu_ps(r1 + 16);
-                __m512 _r12 = _mm512_loadu_ps(r1 + 32);
-                __m512 _r20 = _mm512_loadu_ps(r2);
-                __m512 _r21 = _mm512_loadu_ps(r2 + 16);
-                __m512 _r22 = _mm512_loadu_ps(r2 + 32);
-
-                __m512 _max0 = _mm512_max_ps(_r00, _r01);
-                _max0 = _mm512_max_ps(_max0, _r02);
-                _max0 = _mm512_max_ps(_max0, _r10);
-                _max0 = _mm512_max_ps(_max0, _r11);
-                __m512 _max1 = _mm512_max_ps(_r12, _r20);
-                _max1 = _mm512_max_ps(_max1, _r21);
-                _max1 = _mm512_max_ps(_max1, _r22);
-
-                _mm512_storeu_ps(outptr, _mm512_max_ps(_max0, _max1));
-
-                r0 += 32;
-                r1 += 32;
-                r2 += 32;
-                outptr += 16;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_3x3_pack4.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_3x3_pack4.h
deleted file mode 100644
index ed84c1a2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_3x3_pack4.h
+++ /dev/null
@@ -1,116 +0,0 @@
-
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pooling3x3s2_max_pack4_sse(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * 4;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        float* outptr = top_blob.channel(q);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 1 < outw; j += 2)
-            {
-                __m128 _r00 = _mm_loadu_ps(r0);
-                __m128 _r01 = _mm_loadu_ps(r0 + 4);
-                __m128 _r02 = _mm_loadu_ps(r0 + 8);
-                __m128 _r10 = _mm_loadu_ps(r1);
-                __m128 _r11 = _mm_loadu_ps(r1 + 4);
-                __m128 _r12 = _mm_loadu_ps(r1 + 8);
-                __m128 _r20 = _mm_loadu_ps(r2);
-                __m128 _r21 = _mm_loadu_ps(r2 + 4);
-                __m128 _r22 = _mm_loadu_ps(r2 + 8);
-
-                __m128 _max00 = _mm_max_ps(_r00, _r01);
-                _max00 = _mm_max_ps(_max00, _r02);
-                _max00 = _mm_max_ps(_max00, _r10);
-                _max00 = _mm_max_ps(_max00, _r11);
-                __m128 _max01 = _mm_max_ps(_r12, _r20);
-                _max01 = _mm_max_ps(_max01, _r21);
-                _max01 = _mm_max_ps(_max01, _r22);
-
-                __m128 _r03 = _mm_loadu_ps(r0 + 12);
-                __m128 _r04 = _mm_loadu_ps(r0 + 16);
-                __m128 _r13 = _mm_loadu_ps(r1 + 12);
-                __m128 _r14 = _mm_loadu_ps(r1 + 16);
-                __m128 _r23 = _mm_loadu_ps(r2 + 12);
-                __m128 _r24 = _mm_loadu_ps(r2 + 16);
-
-                _mm_storeu_ps(outptr, _mm_max_ps(_max00, _max01));
-
-                __m128 _max10 = _mm_max_ps(_r03, _r04);
-                _max10 = _mm_max_ps(_max10, _r02);
-                _max10 = _mm_max_ps(_max10, _r13);
-                _max10 = _mm_max_ps(_max10, _r14);
-                __m128 _max11 = _mm_max_ps(_r12, _r23);
-                _max10 = _mm_max_ps(_max10, _r24);
-                _max10 = _mm_max_ps(_max10, _r22);
-
-                _mm_storeu_ps(outptr + 4, _mm_max_ps(_max10, _max11));
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                outptr += 8;
-            }
-
-            for (; j < outw; j++)
-            {
-                __m128 _r00 = _mm_loadu_ps(r0);
-                __m128 _r01 = _mm_loadu_ps(r0 + 4);
-                __m128 _r02 = _mm_loadu_ps(r0 + 8);
-                __m128 _r10 = _mm_loadu_ps(r1);
-                __m128 _r11 = _mm_loadu_ps(r1 + 4);
-                __m128 _r12 = _mm_loadu_ps(r1 + 8);
-                __m128 _r20 = _mm_loadu_ps(r2);
-                __m128 _r21 = _mm_loadu_ps(r2 + 4);
-                __m128 _r22 = _mm_loadu_ps(r2 + 8);
-
-                __m128 _max0 = _mm_max_ps(_r00, _r01);
-                _max0 = _mm_max_ps(_max0, _r02);
-                _max0 = _mm_max_ps(_max0, _r10);
-                _max0 = _mm_max_ps(_max0, _r11);
-                __m128 _max1 = _mm_max_ps(_r12, _r20);
-                _max1 = _mm_max_ps(_max1, _r21);
-                _max1 = _mm_max_ps(_max1, _r22);
-
-                _mm_storeu_ps(outptr, _mm_max_ps(_max0, _max1));
-
-                r0 += 8;
-                r1 += 8;
-                r2 += 8;
-                outptr += 4;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_3x3_pack8.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_3x3_pack8.h
deleted file mode 100644
index 0c5f6877..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_3x3_pack8.h
+++ /dev/null
@@ -1,194 +0,0 @@
-
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void pooling3x3s2_max_pack8_avx(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
-{
-    int w = bottom_blob.w;
-    int inch = bottom_blob.c;
-
-    int outw = top_blob.w;
-    int outh = top_blob.h;
-
-    const int tailstep = (w - 2 * outw + w) * 8;
-
-    #pragma omp parallel for num_threads(opt.num_threads)
-    for (int q = 0; q < inch; q++)
-    {
-        const Mat img0 = bottom_blob.channel(q);
-        float* outptr = top_blob.channel(q);
-
-        const float* r0 = img0.row(0);
-        const float* r1 = img0.row(1);
-        const float* r2 = img0.row(2);
-        for (int i = 0; i < outh; i++)
-        {
-            int j = 0;
-            for (; j + 3 < outw; j += 4)
-            {
-                __m256 _r00 = _mm256_loadu_ps(r0);
-                __m256 _r01 = _mm256_loadu_ps(r0 + 8);
-                __m256 _r02 = _mm256_loadu_ps(r0 + 16);
-                __m256 _r10 = _mm256_loadu_ps(r1);
-                __m256 _r11 = _mm256_loadu_ps(r1 + 8);
-                __m256 _r12 = _mm256_loadu_ps(r1 + 16);
-                __m256 _r20 = _mm256_loadu_ps(r2);
-                __m256 _r21 = _mm256_loadu_ps(r2 + 8);
-                __m256 _r22 = _mm256_loadu_ps(r2 + 16);
-
-                __m256 _max00 = _mm256_max_ps(_r00, _r01);
-                _max00 = _mm256_max_ps(_max00, _r02);
-                _max00 = _mm256_max_ps(_max00, _r10);
-                _max00 = _mm256_max_ps(_max00, _r11);
-                __m256 _max01 = _mm256_max_ps(_r12, _r20);
-                _max01 = _mm256_max_ps(_max01, _r21);
-                _max01 = _mm256_max_ps(_max01, _r22);
-
-                __m256 _r03 = _mm256_loadu_ps(r0 + 24);
-                __m256 _r04 = _mm256_loadu_ps(r0 + 32);
-                __m256 _r13 = _mm256_loadu_ps(r1 + 24);
-                __m256 _r14 = _mm256_loadu_ps(r1 + 32);
-                __m256 _r23 = _mm256_loadu_ps(r2 + 24);
-                __m256 _r24 = _mm256_loadu_ps(r2 + 32);
-
-                _mm256_storeu_ps(outptr, _mm256_max_ps(_max00, _max01));
-
-                __m256 _max10 = _mm256_max_ps(_r03, _r04);
-                _max10 = _mm256_max_ps(_max10, _r02);
-                _max10 = _mm256_max_ps(_max10, _r13);
-                _max10 = _mm256_max_ps(_max10, _r14);
-                __m256 _max11 = _mm256_max_ps(_r12, _r23);
-                _max10 = _mm256_max_ps(_max10, _r24);
-                _max10 = _mm256_max_ps(_max10, _r22);
-
-                __m256 _r05 = _mm256_loadu_ps(r0 + 40);
-                __m256 _r06 = _mm256_loadu_ps(r0 + 48);
-                __m256 _r15 = _mm256_loadu_ps(r1 + 40);
-                __m256 _r16 = _mm256_loadu_ps(r1 + 48);
-                __m256 _r25 = _mm256_loadu_ps(r2 + 40);
-                __m256 _r26 = _mm256_loadu_ps(r2 + 48);
-
-                _mm256_storeu_ps(outptr + 8, _mm256_max_ps(_max10, _max11));
-
-                __m256 _max20 = _mm256_max_ps(_r05, _r06);
-                _max20 = _mm256_max_ps(_max20, _r04);
-                _max20 = _mm256_max_ps(_max20, _r15);
-                _max20 = _mm256_max_ps(_max20, _r16);
-                __m256 _max21 = _mm256_max_ps(_r14, _r25);
-                _max20 = _mm256_max_ps(_max20, _r26);
-                _max20 = _mm256_max_ps(_max20, _r24);
-
-                __m256 _r07 = _mm256_loadu_ps(r0 + 56);
-                __m256 _r08 = _mm256_loadu_ps(r0 + 64);
-                __m256 _r17 = _mm256_loadu_ps(r1 + 56);
-                __m256 _r18 = _mm256_loadu_ps(r1 + 64);
-                __m256 _r27 = _mm256_loadu_ps(r2 + 56);
-                __m256 _r28 = _mm256_loadu_ps(r2 + 64);
-
-                _mm256_storeu_ps(outptr + 16, _mm256_max_ps(_max20, _max21));
-
-                __m256 _max30 = _mm256_max_ps(_r07, _r08);
-                _max30 = _mm256_max_ps(_max30, _r06);
-                _max30 = _mm256_max_ps(_max30, _r17);
-                _max30 = _mm256_max_ps(_max30, _r18);
-                __m256 _max31 = _mm256_max_ps(_r16, _r27);
-                _max30 = _mm256_max_ps(_max30, _r28);
-                _max30 = _mm256_max_ps(_max30, _r26);
-
-                _mm256_storeu_ps(outptr + 24, _mm256_max_ps(_max30, _max31));
-
-                r0 += 64;
-                r1 += 64;
-                r2 += 64;
-                outptr += 32;
-            }
-            for (; j + 1 < outw; j += 2)
-            {
-                __m256 _r00 = _mm256_loadu_ps(r0);
-                __m256 _r01 = _mm256_loadu_ps(r0 + 8);
-                __m256 _r02 = _mm256_loadu_ps(r0 + 16);
-                __m256 _r10 = _mm256_loadu_ps(r1);
-                __m256 _r11 = _mm256_loadu_ps(r1 + 8);
-                __m256 _r12 = _mm256_loadu_ps(r1 + 16);
-                __m256 _r20 = _mm256_loadu_ps(r2);
-                __m256 _r21 = _mm256_loadu_ps(r2 + 8);
-                __m256 _r22 = _mm256_loadu_ps(r2 + 16);
-
-                __m256 _max00 = _mm256_max_ps(_r00, _r01);
-                _max00 = _mm256_max_ps(_max00, _r02);
-                _max00 = _mm256_max_ps(_max00, _r10);
-                _max00 = _mm256_max_ps(_max00, _r11);
-                __m256 _max01 = _mm256_max_ps(_r12, _r20);
-                _max01 = _mm256_max_ps(_max01, _r21);
-                _max01 = _mm256_max_ps(_max01, _r22);
-
-                __m256 _r03 = _mm256_loadu_ps(r0 + 24);
-                __m256 _r04 = _mm256_loadu_ps(r0 + 32);
-                __m256 _r13 = _mm256_loadu_ps(r1 + 24);
-                __m256 _r14 = _mm256_loadu_ps(r1 + 32);
-                __m256 _r23 = _mm256_loadu_ps(r2 + 24);
-                __m256 _r24 = _mm256_loadu_ps(r2 + 32);
-
-                _mm256_storeu_ps(outptr, _mm256_max_ps(_max00, _max01));
-
-                __m256 _max10 = _mm256_max_ps(_r03, _r04);
-                _max10 = _mm256_max_ps(_max10, _r02);
-                _max10 = _mm256_max_ps(_max10, _r13);
-                _max10 = _mm256_max_ps(_max10, _r14);
-                __m256 _max11 = _mm256_max_ps(_r12, _r23);
-                _max10 = _mm256_max_ps(_max10, _r24);
-                _max10 = _mm256_max_ps(_max10, _r22);
-
-                _mm256_storeu_ps(outptr + 8, _mm256_max_ps(_max10, _max11));
-
-                r0 += 32;
-                r1 += 32;
-                r2 += 32;
-                outptr += 16;
-            }
-
-            for (; j < outw; j++)
-            {
-                __m256 _r00 = _mm256_loadu_ps(r0);
-                __m256 _r01 = _mm256_loadu_ps(r0 + 8);
-                __m256 _r02 = _mm256_loadu_ps(r0 + 16);
-                __m256 _r10 = _mm256_loadu_ps(r1);
-                __m256 _r11 = _mm256_loadu_ps(r1 + 8);
-                __m256 _r12 = _mm256_loadu_ps(r1 + 16);
-                __m256 _r20 = _mm256_loadu_ps(r2);
-                __m256 _r21 = _mm256_loadu_ps(r2 + 8);
-                __m256 _r22 = _mm256_loadu_ps(r2 + 16);
-
-                __m256 _max0 = _mm256_max_ps(_r00, _r01);
-                _max0 = _mm256_max_ps(_max0, _r02);
-                _max0 = _mm256_max_ps(_max0, _r10);
-                _max0 = _mm256_max_ps(_max0, _r11);
-                __m256 _max1 = _mm256_max_ps(_r12, _r20);
-                _max1 = _mm256_max_ps(_max1, _r21);
-                _max1 = _mm256_max_ps(_max1, _r22);
-
-                _mm256_storeu_ps(outptr, _mm256_max_ps(_max0, _max1));
-
-                r0 += 16;
-                r1 += 16;
-                r2 += 16;
-                outptr += 8;
-            }
-
-            r0 += tailstep;
-            r1 += tailstep;
-            r2 += tailstep;
-        }
-    }
-}
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_x86.h
deleted file mode 100644
index 030964fc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/pooling_x86.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn
-// available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this
-// file except in compliance with the License. You may obtain a copy of the
-// License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
-// WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
-// License for the specific language governing permissions and limitations under
-// the License.
-
-#ifndef LAYER_POOLING_X86_H
-#define LAYER_POOLING_X86_H
-
-#include "pooling.h"
-
-namespace ncnn {
-
-class Pooling_x86 : public Pooling
-{
-public:
-    Pooling_x86();
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob,
-                        const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_POOLING_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/prelu_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/prelu_x86.h
deleted file mode 100644
index 17d60d4b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/prelu_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_PRELU_X86_H
-#define LAYER_PRELU_X86_H
-
-#include "prelu.h"
-
-namespace ncnn {
-
-class PReLU_x86 : public PReLU
-{
-public:
-    PReLU_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_PRELU_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/quantize_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/quantize_x86.h
deleted file mode 100644
index 5c743fe4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/quantize_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_QUANTIZE_X86_H
-#define LAYER_QUANTIZE_X86_H
-
-#include "quantize.h"
-
-namespace ncnn {
-
-class Quantize_x86 : public Quantize
-{
-public:
-    Quantize_x86();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_QUANTIZE_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/relu_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/relu_x86.h
deleted file mode 100644
index 9d0b5966..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/relu_x86.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_RELU_X86_H
-#define LAYER_RELU_X86_H
-
-#include "relu.h"
-
-namespace ncnn {
-
-class ReLU_x86 : public ReLU
-{
-public:
-    ReLU_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-
-protected:
-    int forward_inplace_int8(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_RELU_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/requantize_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/requantize_x86.h
deleted file mode 100644
index febc4186..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/requantize_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_REQUANTIZE_X86_H
-#define LAYER_REQUANTIZE_X86_H
-
-#include "requantize.h"
-
-namespace ncnn {
-
-class Requantize_x86 : public Requantize
-{
-public:
-    Requantize_x86();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_REQUANTIZE_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/reshape_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/reshape_x86.h
deleted file mode 100644
index 56c8ddfb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/reshape_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_RESHAPE_X86_H
-#define LAYER_RESHAPE_X86_H
-
-#include "reshape.h"
-
-namespace ncnn {
-
-class Reshape_x86 : public Reshape
-{
-public:
-    Reshape_x86();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_RESHAPE_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/roialign_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/roialign_x86.h
deleted file mode 100644
index 1b91c1a8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/roialign_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_ROIALIGN_X86_H
-#define LAYER_ROIALIGN_X86_H
-
-#include "roialign.h"
-
-namespace ncnn {
-
-class ROIAlign_x86 : public ROIAlign
-{
-public:
-    ROIAlign_x86();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_ROIALIGN_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/scale_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/scale_x86.h
deleted file mode 100644
index f06cf414..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/scale_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SCALE_X86_H
-#define LAYER_SCALE_X86_H
-
-#include "scale.h"
-
-namespace ncnn {
-
-class Scale_x86 : public Scale
-{
-public:
-    Scale_x86();
-
-    virtual int forward_inplace(std::vector<Mat>& bottom_top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SCALE_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/selu_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/selu_x86.h
deleted file mode 100644
index 7f4a78f8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/selu_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SELU_X86_H
-#define LAYER_SELU_X86_H
-
-#include "selu.h"
-
-namespace ncnn {
-
-class SELU_x86 : public SELU
-{
-public:
-    SELU_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SELU_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/shufflechannel_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/shufflechannel_x86.h
deleted file mode 100644
index 1e4328a2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/shufflechannel_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SHUFFLECHANNEL_X86_H
-#define LAYER_SHUFFLECHANNEL_X86_H
-
-#include "shufflechannel.h"
-
-namespace ncnn {
-
-class ShuffleChannel_x86 : public ShuffleChannel
-{
-public:
-    ShuffleChannel_x86();
-
-    virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SHUFFLECHANNEL_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/sigmoid_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/sigmoid_x86.h
deleted file mode 100644
index 52bf85d9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/sigmoid_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SIGMOID_X86_H
-#define LAYER_SIGMOID_X86_H
-
-#include "sigmoid.h"
-
-namespace ncnn {
-
-class Sigmoid_x86 : public Sigmoid
-{
-public:
-    Sigmoid_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SIGMOID_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/slice_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/slice_x86.h
deleted file mode 100644
index 0c9b266f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/slice_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SLICE_X86_H
-#define LAYER_SLICE_X86_H
-
-#include "slice.h"
-
-namespace ncnn {
-
-class Slice_x86 : public Slice
-{
-public:
-    Slice_x86();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SLICE_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/softmax_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/softmax_x86.h
deleted file mode 100644
index 3d1b733a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/softmax_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SOFTMAX_X86_H
-#define LAYER_SOFTMAX_X86_H
-
-#include "softmax.h"
-
-namespace ncnn {
-
-class Softmax_x86 : public Softmax
-{
-public:
-    Softmax_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SOFTMAX_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/sse_mathfun.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/sse_mathfun.h
deleted file mode 100644
index b7cecfb8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/sse_mathfun.h
+++ /dev/null
@@ -1,1160 +0,0 @@
-/* SIMD (SSE1+MMX or SSE2) implementation of sin, cos, exp and log
-
-   Inspired by Intel Approximate Math library, and based on the
-   corresponding algorithms of the cephes math library
-
-   The default is to use the SSE1 version. If you define USE_SSE2 the
-   the SSE2 intrinsics will be used in place of the MMX intrinsics. Do
-   not expect any significant performance improvement with SSE2.
-*/
-
-/* Copyright (C) 2007  Julien Pommier
-
-  This software is provided 'as-is', without any express or implied
-  warranty.  In no event will the authors be held liable for any damages
-  arising from the use of this software.
-
-  Permission is granted to anyone to use this software for any purpose,
-  including commercial applications, and to alter it and redistribute it
-  freely, subject to the following restrictions:
-
-  1. The origin of this software must not be misrepresented; you must not
-     claim that you wrote the original software. If you use this software
-     in a product, an acknowledgment in the product documentation would be
-     appreciated but is not required.
-  2. Altered source versions must be plainly marked as such, and must not be
-     misrepresented as being the original software.
-  3. This notice may not be removed or altered from any source distribution.
-
-  (this is the zlib license)
-*/
-
-#ifndef SSE_MATHFUN_H
-#define SSE_MATHFUN_H
-
-#define USE_SSE2 1
-
-#include <xmmintrin.h>
-#include <x86_usability.h>
-
-/* yes I know, the top of this file is quite ugly */
-
-#ifdef _MSC_VER /* visual c++ */
-#define ALIGN16_BEG __declspec(align(16))
-#define ALIGN16_END
-#else /* gcc or icc */
-#define ALIGN16_BEG
-#define ALIGN16_END __attribute__((aligned(16)))
-#endif
-
-/* __m128 is ugly to write */
-typedef __m128 v4sf; // vector of 4 float (sse1)
-
-#ifdef USE_SSE2
-#include <emmintrin.h>
-typedef __m128i v4si; // vector of 4 int (sse2)
-#else
-typedef __m64 v2si; // vector of 2 int (mmx)
-#endif
-
-/* declare some SSE constants -- why can't I figure a better way to do that? */
-#define _PS_CONST(Name, Val) \
-    static const ALIGN16_BEG float _ps_##Name[4] ALIGN16_END = {Val, Val, Val, Val}
-#define _PI32_CONST(Name, Val) \
-    static const ALIGN16_BEG int _pi32_##Name[4] ALIGN16_END = {Val, Val, Val, Val}
-#define _PS_CONST_TYPE(Name, Type, Val) \
-    static const ALIGN16_BEG Type _ps_##Name[4] ALIGN16_END = {Val, Val, Val, Val}
-
-_PS_CONST(1, 1.0f);
-_PS_CONST(0p5, 0.5f);
-/* the smallest non denormalized float number */
-_PS_CONST_TYPE(min_norm_pos, int, 0x00800000);
-_PS_CONST_TYPE(mant_mask, int, 0x7f800000);
-_PS_CONST_TYPE(inv_mant_mask, int, ~0x7f800000);
-
-_PS_CONST_TYPE(sign_mask, int, (int)0x80000000);
-_PS_CONST_TYPE(inv_sign_mask, int, ~0x80000000);
-
-_PI32_CONST(1, 1);
-_PI32_CONST(inv1, ~1);
-_PI32_CONST(2, 2);
-_PI32_CONST(4, 4);
-_PI32_CONST(0x7f, 0x7f);
-
-_PS_CONST(cephes_SQRTHF, 0.707106781186547524f);
-_PS_CONST(cephes_log_p0, 7.0376836292E-2f);
-_PS_CONST(cephes_log_p1, -1.1514610310E-1f);
-_PS_CONST(cephes_log_p2, 1.1676998740E-1f);
-_PS_CONST(cephes_log_p3, -1.2420140846E-1f);
-_PS_CONST(cephes_log_p4, +1.4249322787E-1f);
-_PS_CONST(cephes_log_p5, -1.6668057665E-1f);
-_PS_CONST(cephes_log_p6, +2.0000714765E-1f);
-_PS_CONST(cephes_log_p7, -2.4999993993E-1f);
-_PS_CONST(cephes_log_p8, +3.3333331174E-1f);
-_PS_CONST(cephes_log_q1, -2.12194440e-4f);
-_PS_CONST(cephes_log_q2, 0.693359375f);
-
-#ifndef USE_SSE2
-typedef union xmm_mm_union
-{
-    __m128 xmm;
-    __m64 mm[2];
-} xmm_mm_union;
-
-#define COPY_XMM_TO_MM(xmm_, mm0_, mm1_) \
-    {                                    \
-        xmm_mm_union u;                  \
-        u.xmm = xmm_;                    \
-        mm0_ = u.mm[0];                  \
-        mm1_ = u.mm[1];                  \
-    }
-
-#define COPY_MM_TO_XMM(mm0_, mm1_, xmm_) \
-    {                                    \
-        xmm_mm_union u;                  \
-        u.mm[0] = mm0_;                  \
-        u.mm[1] = mm1_;                  \
-        xmm_ = u.xmm;                    \
-    }
-
-#endif // USE_SSE2
-
-/* natural logarithm computed for 4 simultaneous float
-   return NaN for x <= 0
-*/
-static NCNN_FORCEINLINE v4sf log_ps(v4sf x)
-{
-#ifdef USE_SSE2
-    v4si emm0;
-#else
-    v2si mm0, mm1;
-#endif
-    v4sf one = *(v4sf*)_ps_1;
-
-    v4sf invalid_mask = _mm_cmple_ps(x, _mm_setzero_ps());
-
-    x = _mm_max_ps(x, *(v4sf*)_ps_min_norm_pos); /* cut off denormalized stuff */
-
-#ifndef USE_SSE2
-    /* part 1: x = frexpf(x, &e); */
-    COPY_XMM_TO_MM(x, mm0, mm1);
-    mm0 = _mm_srli_pi32(mm0, 23);
-    mm1 = _mm_srli_pi32(mm1, 23);
-#else
-    emm0 = _mm_srli_epi32(_mm_castps_si128(x), 23);
-#endif
-    /* keep only the fractional part */
-    x = _mm_and_ps(x, *(v4sf*)_ps_inv_mant_mask);
-    x = _mm_or_ps(x, *(v4sf*)_ps_0p5);
-
-#ifndef USE_SSE2
-    /* now e=mm0:mm1 contain the really base-2 exponent */
-    mm0 = _mm_sub_pi32(mm0, *(v2si*)_pi32_0x7f);
-    mm1 = _mm_sub_pi32(mm1, *(v2si*)_pi32_0x7f);
-    v4sf e = _mm_cvtpi32x2_ps(mm0, mm1);
-    _mm_empty(); /* bye bye mmx */
-#else
-    emm0 = _mm_sub_epi32(emm0, *(v4si*)_pi32_0x7f);
-    v4sf e = _mm_cvtepi32_ps(emm0);
-#endif
-
-    e = _mm_add_ps(e, one);
-
-    /* part2:
-       if( x < SQRTHF ) {
-         e -= 1;
-         x = x + x - 1.0;
-       } else { x = x - 1.0; }
-    */
-    v4sf mask = _mm_cmplt_ps(x, *(v4sf*)_ps_cephes_SQRTHF);
-    v4sf tmp = _mm_and_ps(x, mask);
-    x = _mm_sub_ps(x, one);
-    e = _mm_sub_ps(e, _mm_and_ps(one, mask));
-    x = _mm_add_ps(x, tmp);
-
-    v4sf z = _mm_mul_ps(x, x);
-
-    v4sf y = *(v4sf*)_ps_cephes_log_p0;
-    y = _mm_comp_fmadd_ps(y, x, *(v4sf*)_ps_cephes_log_p1);
-    y = _mm_comp_fmadd_ps(y, x, *(v4sf*)_ps_cephes_log_p2);
-    y = _mm_comp_fmadd_ps(y, x, *(v4sf*)_ps_cephes_log_p3);
-    y = _mm_comp_fmadd_ps(y, x, *(v4sf*)_ps_cephes_log_p4);
-    y = _mm_comp_fmadd_ps(y, x, *(v4sf*)_ps_cephes_log_p5);
-    y = _mm_comp_fmadd_ps(y, x, *(v4sf*)_ps_cephes_log_p6);
-    y = _mm_comp_fmadd_ps(y, x, *(v4sf*)_ps_cephes_log_p7);
-    y = _mm_comp_fmadd_ps(y, x, *(v4sf*)_ps_cephes_log_p8);
-    y = _mm_mul_ps(y, x);
-
-    y = _mm_mul_ps(y, z);
-
-    y = _mm_comp_fmadd_ps(e, *(v4sf*)_ps_cephes_log_q1, y);
-
-    //y = -z * 0.5 + y
-    y = _mm_comp_fnmadd_ps(z, *(v4sf*)_ps_0p5, y);
-
-    x = _mm_add_ps(x, y);
-    x = _mm_comp_fmadd_ps(e, *(v4sf*)_ps_cephes_log_q2, x);
-    x = _mm_or_ps(x, invalid_mask); // negative arg will be NAN
-    return x;
-}
-
-_PS_CONST(exp_hi, 88.3762626647949f);
-_PS_CONST(exp_lo, -88.3762626647949f);
-
-_PS_CONST(cephes_LOG2EF, 1.44269504088896341f);
-_PS_CONST(cephes_exp_C1, 0.693359375f);
-_PS_CONST(cephes_exp_C2, -2.12194440e-4f);
-
-_PS_CONST(cephes_exp_p0, 1.9875691500E-4f);
-_PS_CONST(cephes_exp_p1, 1.3981999507E-3f);
-_PS_CONST(cephes_exp_p2, 8.3334519073E-3f);
-_PS_CONST(cephes_exp_p3, 4.1665795894E-2f);
-_PS_CONST(cephes_exp_p4, 1.6666665459E-1f);
-_PS_CONST(cephes_exp_p5, 5.0000001201E-1f);
-
-static NCNN_FORCEINLINE v4sf exp_ps(v4sf x)
-{
-    v4sf tmp = _mm_setzero_ps(), fx;
-#ifdef USE_SSE2
-    v4si emm0;
-#else
-    v2si mm0, mm1;
-#endif
-    v4sf one = *(v4sf*)_ps_1;
-
-    x = _mm_min_ps(x, *(v4sf*)_ps_exp_hi);
-    x = _mm_max_ps(x, *(v4sf*)_ps_exp_lo);
-
-    /* express exp(x) as exp(g + n*log(2)) */
-    fx = _mm_mul_ps(x, *(v4sf*)_ps_cephes_LOG2EF);
-    fx = _mm_add_ps(fx, *(v4sf*)_ps_0p5);
-
-    /* how to perform a floorf with SSE: just below */
-#ifndef USE_SSE2
-    /* step 1 : cast to int */
-    tmp = _mm_movehl_ps(tmp, fx);
-    mm0 = _mm_cvttps_pi32(fx);
-    mm1 = _mm_cvttps_pi32(tmp);
-    /* step 2 : cast back to float */
-    tmp = _mm_cvtpi32x2_ps(mm0, mm1);
-#else
-    emm0 = _mm_cvttps_epi32(fx);
-    tmp = _mm_cvtepi32_ps(emm0);
-#endif
-    /* if greater, substract 1 */
-    v4sf mask = _mm_cmpgt_ps(tmp, fx);
-    mask = _mm_and_ps(mask, one);
-    fx = _mm_sub_ps(tmp, mask);
-
-    // x = x - fx * exp_C1
-    x = _mm_comp_fnmadd_ps(fx, *(v4sf*)_ps_cephes_exp_C1, x);
-    // x = x - fx * exp_C2
-    x = _mm_comp_fnmadd_ps(fx, *(v4sf*)_ps_cephes_exp_C2, x);
-
-    tmp = _mm_mul_ps(x, x);
-
-    v4sf y = *(v4sf*)_ps_cephes_exp_p0;
-    y = _mm_comp_fmadd_ps(y, x, *(v4sf*)_ps_cephes_exp_p1);
-    y = _mm_comp_fmadd_ps(y, x, *(v4sf*)_ps_cephes_exp_p2);
-    y = _mm_comp_fmadd_ps(y, x, *(v4sf*)_ps_cephes_exp_p3);
-    y = _mm_comp_fmadd_ps(y, x, *(v4sf*)_ps_cephes_exp_p4);
-    y = _mm_comp_fmadd_ps(y, x, *(v4sf*)_ps_cephes_exp_p5);
-    y = _mm_comp_fmadd_ps(y, tmp, x);
-
-    y = _mm_add_ps(y, one);
-
-    /* build 2^n */
-#ifndef USE_SSE2
-    z = _mm_movehl_ps(z, fx);
-    mm0 = _mm_cvttps_pi32(fx);
-    mm1 = _mm_cvttps_pi32(z);
-    mm0 = _mm_add_pi32(mm0, *(v2si*)_pi32_0x7f);
-    mm1 = _mm_add_pi32(mm1, *(v2si*)_pi32_0x7f);
-    mm0 = _mm_slli_pi32(mm0, 23);
-    mm1 = _mm_slli_pi32(mm1, 23);
-
-    v4sf pow2n;
-    COPY_MM_TO_XMM(mm0, mm1, pow2n);
-    _mm_empty();
-#else
-    emm0 = _mm_cvttps_epi32(fx);
-    emm0 = _mm_add_epi32(emm0, *(v4si*)_pi32_0x7f);
-    emm0 = _mm_slli_epi32(emm0, 23);
-    v4sf pow2n = _mm_castsi128_ps(emm0);
-#endif
-    y = _mm_mul_ps(y, pow2n);
-    return y;
-}
-
-_PS_CONST(tanh_hi, 9.0f);
-_PS_CONST(tanh_lo, -9.0f);
-
-_PS_CONST(cephes_tanh_p0, -2.76076847742355E-16f);
-_PS_CONST(cephes_tanh_p1, 2.00018790482477E-13f);
-_PS_CONST(cephes_tanh_p2, -8.60467152213735E-11f);
-_PS_CONST(cephes_tanh_p3, 5.12229709037114E-08f);
-_PS_CONST(cephes_tanh_p4, 1.48572235717979E-05f);
-_PS_CONST(cephes_tanh_p5, 6.37261928875436E-04f);
-_PS_CONST(cephes_tanh_p6, 4.89352455891786E-03f);
-_PS_CONST(cephes_tanh_p7, 1.19825839466702e-06f);
-_PS_CONST(cephes_tanh_p8, 1.18534705686654e-04f);
-_PS_CONST(cephes_tanh_p9, 2.26843463243900e-03f);
-
-// an approximation of tanh
-static inline v4sf tanh_ps(const v4sf x)
-{
-    v4sf value = x;
-    value = _mm_max_ps(*(v4sf*)_ps_tanh_lo, value);
-    value = _mm_min_ps(*(v4sf*)_ps_tanh_hi, value);
-
-    v4sf value_squared = _mm_mul_ps(value, value);
-
-    v4sf p;
-    p = _mm_comp_fmadd_ps(value_squared, *(v4sf*)_ps_cephes_tanh_p0, *(v4sf*)_ps_cephes_tanh_p1);
-    p = _mm_comp_fmadd_ps(p, value_squared, *(v4sf*)_ps_cephes_tanh_p2);
-    p = _mm_comp_fmadd_ps(p, value_squared, *(v4sf*)_ps_cephes_tanh_p3);
-    p = _mm_comp_fmadd_ps(p, value_squared, *(v4sf*)_ps_cephes_tanh_p4);
-    p = _mm_comp_fmadd_ps(p, value_squared, *(v4sf*)_ps_cephes_tanh_p5);
-    p = _mm_comp_fmadd_ps(p, value_squared, *(v4sf*)_ps_cephes_tanh_p6);
-    p = _mm_mul_ps(p, value);
-
-    v4sf q;
-    q = _mm_comp_fmadd_ps(value_squared, *(v4sf*)_ps_cephes_tanh_p7, *(v4sf*)_ps_cephes_tanh_p8);
-    q = _mm_comp_fmadd_ps(q, value_squared, *(v4sf*)_ps_cephes_tanh_p9);
-    q = _mm_comp_fmadd_ps(q, value_squared, *(v4sf*)_ps_cephes_tanh_p6);
-
-    v4sf dst = _mm_div_ps(p, q);
-    return dst;
-}
-
-_PS_CONST(minus_cephes_DP1, -0.78515625f);
-_PS_CONST(minus_cephes_DP2, -2.4187564849853515625e-4f);
-_PS_CONST(minus_cephes_DP3, -3.77489497744594108e-8f);
-_PS_CONST(sincof_p0, -1.9515295891E-4f);
-_PS_CONST(sincof_p1, 8.3321608736E-3f);
-_PS_CONST(sincof_p2, -1.6666654611E-1f);
-_PS_CONST(coscof_p0, 2.443315711809948E-005f);
-_PS_CONST(coscof_p1, -1.388731625493765E-003f);
-_PS_CONST(coscof_p2, 4.166664568298827E-002f);
-_PS_CONST(cephes_FOPI, 1.27323954473516f); // 4 / M_PI
-
-/* evaluation of 4 sines at onces, using only SSE1+MMX intrinsics so
-   it runs also on old athlons XPs and the pentium III of your grand
-   mother.
-
-   The code is the exact rewriting of the cephes sinf function.
-   Precision is excellent as long as x < 8192 (I did not bother to
-   take into account the special handling they have for greater values
-   -- it does not return garbage for arguments over 8192, though, but
-   the extra precision is missing).
-
-   Note that it is such that sinf((float)M_PI) = 8.74e-8, which is the
-   surprising but correct result.
-
-   Performance is also surprisingly good, 1.33 times faster than the
-   macos vsinf SSE2 function, and 1.5 times faster than the
-   __vrs4_sinf of amd's ACML (which is only available in 64 bits). Not
-   too bad for an SSE1 function (with no special tuning) !
-   However the latter libraries probably have a much better handling of NaN,
-   Inf, denormalized and other special arguments..
-
-   On my core 1 duo, the execution of this function takes approximately 95 cycles.
-
-   From what I have observed on the experiments with Intel AMath lib, switching to an
-   SSE2 version would improve the perf by only 10%.
-
-   Since it is based on SSE intrinsics, it has to be compiled at -O2 to
-   deliver full speed.
-*/
-static NCNN_FORCEINLINE v4sf sin_ps(v4sf x)
-{   // any x
-    v4sf xmm1, xmm2 = _mm_setzero_ps(), xmm3, sign_bit, y;
-
-#ifdef USE_SSE2
-    v4si emm0, emm2;
-#else
-    v2si mm0, mm1, mm2, mm3;
-#endif
-    sign_bit = x;
-    /* take the absolute value */
-    x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
-    /* extract the sign bit (upper one) */
-    sign_bit = _mm_and_ps(sign_bit, *(v4sf*)_ps_sign_mask);
-
-    /* scale by 4/Pi */
-    y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
-
-#ifdef USE_SSE2
-    /* store the integer part of y in mm0 */
-    emm2 = _mm_cvttps_epi32(y);
-    /* j=(j+1) & (~1) (see the cephes sources) */
-    emm2 = _mm_add_epi32(emm2, *(v4si*)_pi32_1);
-    emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_inv1);
-    y = _mm_cvtepi32_ps(emm2);
-
-    /* get the swap sign flag */
-    emm0 = _mm_and_si128(emm2, *(v4si*)_pi32_4);
-    emm0 = _mm_slli_epi32(emm0, 29);
-    /* get the polynom selection mask
-       there is one polynom for 0 <= x <= Pi/4
-       and another one for Pi/4<x<=Pi/2
-
-       Both branches will be computed.
-    */
-    emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_2);
-    emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
-
-    v4sf swap_sign_bit = _mm_castsi128_ps(emm0);
-    v4sf poly_mask = _mm_castsi128_ps(emm2);
-    sign_bit = _mm_xor_ps(sign_bit, swap_sign_bit);
-
-#else
-    /* store the integer part of y in mm0:mm1 */
-    xmm2 = _mm_movehl_ps(xmm2, y);
-    mm2 = _mm_cvttps_pi32(y);
-    mm3 = _mm_cvttps_pi32(xmm2);
-    /* j=(j+1) & (~1) (see the cephes sources) */
-    mm2 = _mm_add_pi32(mm2, *(v2si*)_pi32_1);
-    mm3 = _mm_add_pi32(mm3, *(v2si*)_pi32_1);
-    mm2 = _mm_and_si64(mm2, *(v2si*)_pi32_inv1);
-    mm3 = _mm_and_si64(mm3, *(v2si*)_pi32_inv1);
-    y = _mm_cvtpi32x2_ps(mm2, mm3);
-    /* get the swap sign flag */
-    mm0 = _mm_and_si64(mm2, *(v2si*)_pi32_4);
-    mm1 = _mm_and_si64(mm3, *(v2si*)_pi32_4);
-    mm0 = _mm_slli_pi32(mm0, 29);
-    mm1 = _mm_slli_pi32(mm1, 29);
-    /* get the polynom selection mask */
-    mm2 = _mm_and_si64(mm2, *(v2si*)_pi32_2);
-    mm3 = _mm_and_si64(mm3, *(v2si*)_pi32_2);
-    mm2 = _mm_cmpeq_pi32(mm2, _mm_setzero_si64());
-    mm3 = _mm_cmpeq_pi32(mm3, _mm_setzero_si64());
-    v4sf swap_sign_bit, poly_mask;
-    COPY_MM_TO_XMM(mm0, mm1, swap_sign_bit);
-    COPY_MM_TO_XMM(mm2, mm3, poly_mask);
-    sign_bit = _mm_xor_ps(sign_bit, swap_sign_bit);
-    _mm_empty(); /* good-bye mmx */
-#endif
-
-    /* The magic pass: "Extended precision modular arithmetic"
-       x = ((x - y * DP1) - y * DP2) - y * DP3; */
-    xmm1 = *(v4sf*)_ps_minus_cephes_DP1;
-    xmm2 = *(v4sf*)_ps_minus_cephes_DP2;
-    xmm3 = *(v4sf*)_ps_minus_cephes_DP3;
-    x = _mm_comp_fmadd_ps(y, xmm1, x);
-    x = _mm_comp_fmadd_ps(y, xmm2, x);
-    x = _mm_comp_fmadd_ps(y, xmm3, x);
-
-    /* Evaluate the first polynom  (0 <= x <= Pi/4) */
-    y = *(v4sf*)_ps_coscof_p0;
-    v4sf z = _mm_mul_ps(x, x);
-
-    y = _mm_comp_fmadd_ps(y, z, *(v4sf*)_ps_coscof_p1);
-    y = _mm_comp_fmadd_ps(y, z, *(v4sf*)_ps_coscof_p2);
-    y = _mm_mul_ps(y, z);
-    y = _mm_mul_ps(y, z);
-    y = _mm_comp_fnmadd_ps(z, *(v4sf*)_ps_0p5, y);
-    y = _mm_add_ps(y, *(v4sf*)_ps_1);
-
-    /* Evaluate the second polynom  (Pi/4 <= x <= 0) */
-
-    v4sf y2 = *(v4sf*)_ps_sincof_p0;
-    y2 = _mm_comp_fmadd_ps(y2, z, *(v4sf*)_ps_sincof_p1);
-    y2 = _mm_comp_fmadd_ps(y2, z, *(v4sf*)_ps_sincof_p2);
-    y2 = _mm_mul_ps(y2, z);
-    y2 = _mm_comp_fmadd_ps(y2, x, x);
-
-    /* select the correct result from the two polynoms */
-    xmm3 = poly_mask;
-    y2 = _mm_and_ps(xmm3, y2); //, xmm3);
-    y = _mm_andnot_ps(xmm3, y);
-    y = _mm_add_ps(y, y2);
-    /* update the sign */
-    y = _mm_xor_ps(y, sign_bit);
-    return y;
-}
-
-/* almost the same as sin_ps */
-static NCNN_FORCEINLINE v4sf cos_ps(v4sf x)
-{   // any x
-    v4sf xmm1, xmm2 = _mm_setzero_ps(), xmm3, y;
-#ifdef USE_SSE2
-    v4si emm0, emm2;
-#else
-    v2si mm0, mm1, mm2, mm3;
-#endif
-    /* take the absolute value */
-    x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
-
-    /* scale by 4/Pi */
-    y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
-
-#ifdef USE_SSE2
-    /* store the integer part of y in mm0 */
-    emm2 = _mm_cvttps_epi32(y);
-    /* j=(j+1) & (~1) (see the cephes sources) */
-    emm2 = _mm_add_epi32(emm2, *(v4si*)_pi32_1);
-    emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_inv1);
-    y = _mm_cvtepi32_ps(emm2);
-
-    emm2 = _mm_sub_epi32(emm2, *(v4si*)_pi32_2);
-
-    /* get the swap sign flag */
-    emm0 = _mm_andnot_si128(emm2, *(v4si*)_pi32_4);
-    emm0 = _mm_slli_epi32(emm0, 29);
-    /* get the polynom selection mask */
-    emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_2);
-    emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
-
-    v4sf sign_bit = _mm_castsi128_ps(emm0);
-    v4sf poly_mask = _mm_castsi128_ps(emm2);
-#else
-    /* store the integer part of y in mm0:mm1 */
-    xmm2 = _mm_movehl_ps(xmm2, y);
-    mm2 = _mm_cvttps_pi32(y);
-    mm3 = _mm_cvttps_pi32(xmm2);
-
-    /* j=(j+1) & (~1) (see the cephes sources) */
-    mm2 = _mm_add_pi32(mm2, *(v2si*)_pi32_1);
-    mm3 = _mm_add_pi32(mm3, *(v2si*)_pi32_1);
-    mm2 = _mm_and_si64(mm2, *(v2si*)_pi32_inv1);
-    mm3 = _mm_and_si64(mm3, *(v2si*)_pi32_inv1);
-
-    y = _mm_cvtpi32x2_ps(mm2, mm3);
-
-    mm2 = _mm_sub_pi32(mm2, *(v2si*)_pi32_2);
-    mm3 = _mm_sub_pi32(mm3, *(v2si*)_pi32_2);
-
-    /* get the swap sign flag in mm0:mm1 and the
-       polynom selection mask in mm2:mm3 */
-
-    mm0 = _mm_andnot_si64(mm2, *(v2si*)_pi32_4);
-    mm1 = _mm_andnot_si64(mm3, *(v2si*)_pi32_4);
-    mm0 = _mm_slli_pi32(mm0, 29);
-    mm1 = _mm_slli_pi32(mm1, 29);
-
-    mm2 = _mm_and_si64(mm2, *(v2si*)_pi32_2);
-    mm3 = _mm_and_si64(mm3, *(v2si*)_pi32_2);
-
-    mm2 = _mm_cmpeq_pi32(mm2, _mm_setzero_si64());
-    mm3 = _mm_cmpeq_pi32(mm3, _mm_setzero_si64());
-
-    v4sf sign_bit, poly_mask;
-    COPY_MM_TO_XMM(mm0, mm1, sign_bit);
-    COPY_MM_TO_XMM(mm2, mm3, poly_mask);
-    _mm_empty(); /* good-bye mmx */
-#endif
-    /* The magic pass: "Extended precision modular arithmetic"
-       x = ((x - y * DP1) - y * DP2) - y * DP3; */
-    xmm1 = *(v4sf*)_ps_minus_cephes_DP1;
-    xmm2 = *(v4sf*)_ps_minus_cephes_DP2;
-    xmm3 = *(v4sf*)_ps_minus_cephes_DP3;
-    x = _mm_comp_fmadd_ps(y, xmm1, x);
-    x = _mm_comp_fmadd_ps(y, xmm2, x);
-    x = _mm_comp_fmadd_ps(y, xmm3, x);
-
-    /* Evaluate the first polynom  (0 <= x <= Pi/4) */
-    y = *(v4sf*)_ps_coscof_p0;
-    v4sf z = _mm_mul_ps(x, x);
-
-    y = _mm_comp_fmadd_ps(y, z, *(v4sf*)_ps_coscof_p1);
-    y = _mm_comp_fmadd_ps(y, z, *(v4sf*)_ps_coscof_p2);
-    y = _mm_mul_ps(y, z);
-    y = _mm_mul_ps(y, z);
-    y = _mm_comp_fnmadd_ps(z, *(v4sf*)_ps_0p5, y);
-    y = _mm_add_ps(y, *(v4sf*)_ps_1);
-
-    /* Evaluate the second polynom  (Pi/4 <= x <= 0) */
-
-    v4sf y2 = *(v4sf*)_ps_sincof_p0;
-    y2 = _mm_comp_fmadd_ps(y2, z, *(v4sf*)_ps_sincof_p1);
-    y2 = _mm_comp_fmadd_ps(y2, z, *(v4sf*)_ps_sincof_p2);
-    y2 = _mm_mul_ps(y2, z);
-    y2 = _mm_comp_fmadd_ps(y2, x, x);
-
-    /* select the correct result from the two polynoms */
-    xmm3 = poly_mask;
-    y2 = _mm_and_ps(xmm3, y2); //, xmm3);
-    y = _mm_andnot_ps(xmm3, y);
-    y = _mm_add_ps(y, y2);
-    /* update the sign */
-    y = _mm_xor_ps(y, sign_bit);
-
-    return y;
-}
-
-/* since sin_ps and cos_ps are almost identical, sincos_ps could replace both of them..
-   it is almost as fast, and gives you a free cosine with your sine */
-static NCNN_FORCEINLINE void sincos_ps(v4sf x, v4sf* s, v4sf* c)
-{
-    v4sf xmm1, xmm2, xmm3 = _mm_setzero_ps(), sign_bit_sin, y;
-#ifdef USE_SSE2
-    v4si emm0, emm2, emm4;
-#else
-    v2si mm0, mm1, mm2, mm3, mm4, mm5;
-#endif
-    sign_bit_sin = x;
-    /* take the absolute value */
-    x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
-    /* extract the sign bit (upper one) */
-    sign_bit_sin = _mm_and_ps(sign_bit_sin, *(v4sf*)_ps_sign_mask);
-
-    /* scale by 4/Pi */
-    y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
-
-#ifdef USE_SSE2
-    /* store the integer part of y in emm2 */
-    emm2 = _mm_cvttps_epi32(y);
-
-    /* j=(j+1) & (~1) (see the cephes sources) */
-    emm2 = _mm_add_epi32(emm2, *(v4si*)_pi32_1);
-    emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_inv1);
-    y = _mm_cvtepi32_ps(emm2);
-
-    emm4 = emm2;
-
-    /* get the swap sign flag for the sine */
-    emm0 = _mm_and_si128(emm2, *(v4si*)_pi32_4);
-    emm0 = _mm_slli_epi32(emm0, 29);
-    v4sf swap_sign_bit_sin = _mm_castsi128_ps(emm0);
-
-    /* get the polynom selection mask for the sine*/
-    emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_2);
-    emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
-    v4sf poly_mask = _mm_castsi128_ps(emm2);
-#else
-    /* store the integer part of y in mm2:mm3 */
-    xmm3 = _mm_movehl_ps(xmm3, y);
-    mm2 = _mm_cvttps_pi32(y);
-    mm3 = _mm_cvttps_pi32(xmm3);
-
-    /* j=(j+1) & (~1) (see the cephes sources) */
-    mm2 = _mm_add_pi32(mm2, *(v2si*)_pi32_1);
-    mm3 = _mm_add_pi32(mm3, *(v2si*)_pi32_1);
-    mm2 = _mm_and_si64(mm2, *(v2si*)_pi32_inv1);
-    mm3 = _mm_and_si64(mm3, *(v2si*)_pi32_inv1);
-
-    y = _mm_cvtpi32x2_ps(mm2, mm3);
-
-    mm4 = mm2;
-    mm5 = mm3;
-
-    /* get the swap sign flag for the sine */
-    mm0 = _mm_and_si64(mm2, *(v2si*)_pi32_4);
-    mm1 = _mm_and_si64(mm3, *(v2si*)_pi32_4);
-    mm0 = _mm_slli_pi32(mm0, 29);
-    mm1 = _mm_slli_pi32(mm1, 29);
-    v4sf swap_sign_bit_sin;
-    COPY_MM_TO_XMM(mm0, mm1, swap_sign_bit_sin);
-
-    /* get the polynom selection mask for the sine */
-
-    mm2 = _mm_and_si64(mm2, *(v2si*)_pi32_2);
-    mm3 = _mm_and_si64(mm3, *(v2si*)_pi32_2);
-    mm2 = _mm_cmpeq_pi32(mm2, _mm_setzero_si64());
-    mm3 = _mm_cmpeq_pi32(mm3, _mm_setzero_si64());
-    v4sf poly_mask;
-    COPY_MM_TO_XMM(mm2, mm3, poly_mask);
-#endif
-
-    /* The magic pass: "Extended precision modular arithmetic"
-       x = ((x - y * DP1) - y * DP2) - y * DP3; */
-    xmm1 = *(v4sf*)_ps_minus_cephes_DP1;
-    xmm2 = *(v4sf*)_ps_minus_cephes_DP2;
-    xmm3 = *(v4sf*)_ps_minus_cephes_DP3;
-    x = _mm_comp_fmadd_ps(y, xmm1, x);
-    x = _mm_comp_fmadd_ps(y, xmm2, x);
-    x = _mm_comp_fmadd_ps(y, xmm3, x);
-
-#ifdef USE_SSE2
-    emm4 = _mm_sub_epi32(emm4, *(v4si*)_pi32_2);
-    emm4 = _mm_andnot_si128(emm4, *(v4si*)_pi32_4);
-    emm4 = _mm_slli_epi32(emm4, 29);
-    v4sf sign_bit_cos = _mm_castsi128_ps(emm4);
-#else
-    /* get the sign flag for the cosine */
-    mm4 = _mm_sub_pi32(mm4, *(v2si*)_pi32_2);
-    mm5 = _mm_sub_pi32(mm5, *(v2si*)_pi32_2);
-    mm4 = _mm_andnot_si64(mm4, *(v2si*)_pi32_4);
-    mm5 = _mm_andnot_si64(mm5, *(v2si*)_pi32_4);
-    mm4 = _mm_slli_pi32(mm4, 29);
-    mm5 = _mm_slli_pi32(mm5, 29);
-    v4sf sign_bit_cos;
-    COPY_MM_TO_XMM(mm4, mm5, sign_bit_cos);
-    _mm_empty(); /* good-bye mmx */
-#endif
-
-    sign_bit_sin = _mm_xor_ps(sign_bit_sin, swap_sign_bit_sin);
-
-    /* Evaluate the first polynom  (0 <= x <= Pi/4) */
-    v4sf z = _mm_mul_ps(x, x);
-    y = *(v4sf*)_ps_coscof_p0;
-
-    y = _mm_comp_fmadd_ps(y, z, *(v4sf*)_ps_coscof_p1);
-    y = _mm_comp_fmadd_ps(y, z, *(v4sf*)_ps_coscof_p2);
-    y = _mm_mul_ps(y, z);
-    y = _mm_mul_ps(y, z);
-    y = _mm_comp_fnmadd_ps(z, *(v4sf*)_ps_0p5, y);
-    y = _mm_add_ps(y, *(v4sf*)_ps_1);
-
-    /* Evaluate the second polynom  (Pi/4 <= x <= 0) */
-
-    v4sf y2 = *(v4sf*)_ps_sincof_p0;
-    y2 = _mm_comp_fmadd_ps(y2, z, *(v4sf*)_ps_sincof_p1);
-    y2 = _mm_comp_fmadd_ps(y2, z, *(v4sf*)_ps_sincof_p2);
-    y2 = _mm_mul_ps(y2, z);
-    y2 = _mm_comp_fmadd_ps(y2, x, x);
-
-    /* select the correct result from the two polynoms */
-    xmm3 = poly_mask;
-    v4sf ysin2 = _mm_and_ps(xmm3, y2);
-    v4sf ysin1 = _mm_andnot_ps(xmm3, y);
-    y2 = _mm_sub_ps(y2, ysin2);
-    y = _mm_sub_ps(y, ysin1);
-
-    xmm1 = _mm_add_ps(ysin1, ysin2);
-    xmm2 = _mm_add_ps(y, y2);
-
-    /* update the sign */
-    *s = _mm_xor_ps(xmm1, sign_bit_sin);
-    *c = _mm_xor_ps(xmm2, sign_bit_cos);
-}
-
-static NCNN_FORCEINLINE __m128 tan_ps(__m128 x)
-{
-    __m128 ysin, ycos;
-    __m128 eps = _mm_set1_ps(1E-8f);
-    sincos_ps(x, &ysin, &ycos);
-    __m128 mask = _mm_cmpeq_ps(ycos, _mm_setzero_ps());
-    __m128 _tmp = _mm_and_ps(eps, mask);
-    ycos = _mm_add_ps(ycos, _tmp);
-    __m128 ytan = _mm_div_ps(ysin, ycos);
-    return ytan;
-}
-
-static NCNN_FORCEINLINE __m128 pow_ps(__m128 a, __m128 b)
-{
-    // pow(x, m) = exp(m * log(x))
-    return exp_ps(_mm_mul_ps(b, log_ps(a)));
-}
-
-static NCNN_FORCEINLINE __m128 ceil_ps(__m128 x)
-{
-#if __SSE4_1__
-    return _mm_ceil_ps(x);
-#endif // __SSE4_1__
-
-    // Use negative zero as the sign bit mask.
-    const __m128 magic_negative_zero = _mm_set_ps1(-0.0f);
-
-    // The smallest float number that have no fractional part. (2^23)
-    const __m128 magic_smallest_no_fraction = _mm_set_ps1(8388608.0f);
-
-    // absolute = abs(x);
-    __m128 absolute = _mm_andnot_ps(magic_negative_zero, x);
-
-    // negative_mask = magic_negative_zero && x;
-    __m128 negative_mask = _mm_and_ps(magic_negative_zero, x);
-
-    // no_fraction = (magic_smallest_no_fraction < absolute);
-    __m128 no_fraction = _mm_cmplt_ps(magic_smallest_no_fraction, absolute);
-
-    // truncated = static_cast<float>(static_cast<uint32_t>(absolute));
-    __m128 truncated = _mm_cvtepi32_ps(_mm_cvttps_epi32(absolute));
-
-    // truncated_with_sign = (truncated || negative_mask);
-    __m128 truncated_with_sign = _mm_or_ps(truncated, negative_mask);
-
-    // positive_fix = ((x > -0.0f) && (x > truncated_with_sign) ? -1.0f : 0.0f);
-    __m128 positive_fix = _mm_and_ps(
-                              _mm_and_ps(
-                                  _mm_cmpgt_ps(x, magic_negative_zero),
-                                  _mm_cmpgt_ps(x, truncated_with_sign)),
-                              _mm_set_ps1(-1.0f));
-
-    // fixed_result = truncated_with_sign - positive_fix;
-    __m128 fixed_result = _mm_sub_ps(truncated_with_sign, positive_fix);
-
-    // return ((x && no_fraction) || (!no_fraction && fixed_result));
-    return _mm_or_ps(
-               _mm_and_ps(x, no_fraction),
-               _mm_andnot_ps(no_fraction, fixed_result));
-}
-
-static NCNN_FORCEINLINE __m128 floor_ps(__m128 x)
-{
-#if __SSE4_1__
-    return _mm_floor_ps(x);
-#endif // __SSE4_1__
-
-    // Use negative zero as the sign bit mask.
-    const __m128 magic_negative_zero = _mm_set_ps1(-0.0f);
-
-    // The smallest float number that have no fractional part. (2^23)
-    const __m128 magic_smallest_no_fraction = _mm_set_ps1(8388608.0f);
-
-    // absolute = abs(x);
-    __m128 absolute = _mm_andnot_ps(magic_negative_zero, x);
-
-    // negative_mask = magic_negative_zero && x;
-    __m128 negative_mask = _mm_and_ps(magic_negative_zero, x);
-
-    // no_fraction = (magic_smallest_no_fraction < absolute);
-    __m128 no_fraction = _mm_cmplt_ps(magic_smallest_no_fraction, absolute);
-
-    // truncated = static_cast<float>(static_cast<uint32_t>(absolute));
-    __m128 truncated = _mm_cvtepi32_ps(_mm_cvttps_epi32(absolute));
-
-    // truncated_with_sign = (truncated || negative_mask);
-    __m128 truncated_with_sign = _mm_or_ps(truncated, negative_mask);
-
-    // negative_fix = ((x < truncated_with_sign) ? 1.0f : 0.0f);
-    __m128 negative_fix = _mm_and_ps(
-                              _mm_cmplt_ps(x, truncated_with_sign),
-                              _mm_set_ps1(1.0f));
-
-    // fixed_result = truncated_with_sign - negative_fix;
-    __m128 fixed_result = _mm_sub_ps(truncated_with_sign, negative_fix);
-
-    // return ((x && no_fraction) || (!no_fraction && fixed_result));
-    return _mm_or_ps(
-               _mm_and_ps(x, no_fraction),
-               _mm_andnot_ps(no_fraction, fixed_result));
-}
-
-static NCNN_FORCEINLINE __m128 asin_ps(__m128 x)
-{
-    const __m128 magic_negative_zero = _mm_set_ps1(-0.0f);
-    const __m128 magic_half_one = _mm_set_ps1(0.5f);
-    const __m128 magic_one = _mm_set_ps1(1.0f);
-    const __m128 magic_a4 = _mm_set_ps1(0.023994016f);
-    const __m128 magic_a5 = _mm_set_ps1(0.042417344f);
-    const __m128 magic_a2 = _mm_set_ps1(0.07494697f);
-    const __m128 magic_a3 = _mm_set_ps1(0.045520633f);
-    const __m128 magic_a0 = _mm_set_ps1(1.0f);
-    const __m128 magic_a1 = _mm_set_ps1(0.166667819f);
-    const __m128 magic_half_pi = _mm_set_ps1(1.5707964f);
-    const __m128 magic_three = _mm_set_ps1(3.0f);
-
-    // negative_mask = magic_negative_zero && x;
-    __m128 negative_mask = _mm_and_ps(magic_negative_zero, x);
-
-    // absolute = abs(x);
-    __m128 absolute = _mm_andnot_ps(magic_negative_zero, x);
-
-    // Reference: https://en.wikipedia.org/wiki/Small-angle_approximation
-
-    // is_small_input = (absolute <= 0.5f);
-    __m128 is_small_input = _mm_cmple_ps(absolute, magic_half_one);
-
-    // is_big_input = (is_small_input ? 0.0f : 1.0f);
-    __m128 is_big_input = _mm_andnot_ps(is_small_input, magic_one);
-
-    // big_input_approx = sqrt(0.5f * (1 - absolute));
-    __m128 big_input_approx = _mm_sqrt_ps(_mm_mul_ps(
-            magic_half_one,
-            _mm_sub_ps(magic_one, absolute)));
-
-    // input_approx = (is_small_input ? absolute : big_input_approx);
-    __m128 input_approx = _mm_or_ps(
-                              _mm_and_ps(is_small_input, absolute),
-                              _mm_andnot_ps(is_small_input, big_input_approx));
-
-    // square_of_input_approx = input_approx * input_approx;
-    __m128 square_of_input_approx = _mm_mul_ps(input_approx, input_approx);
-
-    // fourth_power_of_input_approx =
-    //     square_of_input_approx * square_of_input_approx;
-    __m128 fourth_power_of_input_approx = _mm_mul_ps(
-            square_of_input_approx, square_of_input_approx);
-
-    // TODO: Need more explanations.
-    // x1 = ((fourth_power_of_input_approx * magic_a4) + magic_a2);
-    // x2 = ((fourth_power_of_input_approx * magic_a5) + magic_a3);
-    // x3 = ((fourth_power_of_input_approx * x1) + magic_a0);
-    // x4 = ((fourth_power_of_input_approx * x2) + magic_a1);
-    // output_approx = ((square_of_input_approx * x4) + x3);
-    __m128 output_approx = _mm_comp_fmadd_ps(
-                               square_of_input_approx,
-                               _mm_comp_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm_comp_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       magic_a5,
-                                       magic_a3),
-                                   magic_a1),
-                               _mm_comp_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm_comp_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       magic_a4,
-                                       magic_a2),
-                                   magic_a0));
-
-    // TODO: Need more explanations.
-    // x1 = ((0.5 * PI) * is_big_input);
-    // x2 = (output_approx * input_approx);
-    // x3 = (-(3.0f * is_big_input) + 1.0f);
-    // final_approx = ((x2 * x3) + x1);
-    __m128 final_approx = _mm_comp_fmadd_ps(
-                              _mm_mul_ps(output_approx, input_approx),
-                              _mm_comp_fnmadd_ps(magic_three, is_big_input, magic_one),
-                              _mm_mul_ps(magic_half_pi, is_big_input));
-
-    // return (final_approx || negative_mask);
-    return _mm_or_ps(final_approx, negative_mask);
-}
-
-static NCNN_FORCEINLINE __m128 acos_ps(__m128 x)
-{
-    const __m128 magic_negative_zero = _mm_set_ps1(-0.0f);
-    const __m128 magic_zero = _mm_set_ps1(0.0f);
-    const __m128 magic_half_one = _mm_set_ps1(0.5f);
-    const __m128 magic_one = _mm_set_ps1(1.0f);
-    const __m128 magic_a4 = _mm_set_ps1(0.023994016f);
-    const __m128 magic_a5 = _mm_set_ps1(0.042417344f);
-    const __m128 magic_a2 = _mm_set_ps1(0.07494697f);
-    const __m128 magic_a3 = _mm_set_ps1(0.045520633f);
-    const __m128 magic_a0 = _mm_set_ps1(1.0f);
-    const __m128 magic_a1 = _mm_set_ps1(0.166667819f);
-    const __m128 magic_half_pi = _mm_set_ps1(1.5707964f);
-    const __m128 magic_pi = _mm_set_ps1(3.1415927f);
-
-    // negative_mask = magic_negative_zero && x;
-    __m128 negative_mask = _mm_and_ps(magic_negative_zero, x);
-
-    // absolute = abs(x);
-    __m128 absolute = _mm_andnot_ps(magic_negative_zero, x);
-
-    // Reference: https://en.wikipedia.org/wiki/Small-angle_approximation
-
-    // is_small_input = (absolute <= 0.5f);
-    __m128 is_small_input = _mm_cmple_ps(absolute, magic_half_one);
-
-    // big_input_approx = sqrt(0.5f * (1 - absolute));
-    __m128 big_input_approx = _mm_sqrt_ps(_mm_mul_ps(
-            magic_half_one,
-            _mm_sub_ps(magic_one, absolute)));
-
-    // input_approx = (is_small_input ? absolute : big_input_approx);
-    __m128 input_approx = _mm_or_ps(
-                              _mm_and_ps(is_small_input, absolute),
-                              _mm_andnot_ps(is_small_input, big_input_approx));
-
-    // square_of_input_approx = input_approx * input_approx;
-    __m128 square_of_input_approx = _mm_mul_ps(input_approx, input_approx);
-
-    // fourth_power_of_input_approx =
-    //     square_of_input_approx * square_of_input_approx;
-    __m128 fourth_power_of_input_approx = _mm_mul_ps(
-            square_of_input_approx, square_of_input_approx);
-
-    // TODO: Need more explanations.
-    // x1 = ((fourth_power_of_input_approx * magic_a4) + magic_a2);
-    // x2 = ((fourth_power_of_input_approx * magic_a5) + magic_a3);
-    // x3 = ((fourth_power_of_input_approx * x1) + magic_a0);
-    // x4 = ((fourth_power_of_input_approx * x2) + magic_a1);
-    // output_approx = ((square_of_input_approx * x4) + x3);
-    __m128 output_approx = _mm_comp_fmadd_ps(
-                               square_of_input_approx,
-                               _mm_comp_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm_comp_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       magic_a5,
-                                       magic_a3),
-                                   magic_a1),
-                               _mm_comp_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm_comp_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       magic_a4,
-                                       magic_a2),
-                                   magic_a0));
-
-    // TODO: Need more explanations.
-    // x1 = (output_approx * input_approx);
-    __m128 x1 = _mm_mul_ps(output_approx, input_approx);
-
-    // TODO: Need more explanations.
-    // small_final_approx = ((0.5 * PI) - (x1 | negative_mask));
-    __m128 small_final_approx = _mm_sub_ps(
-                                    magic_half_pi,
-                                    _mm_or_ps(x1, negative_mask));
-
-    // TODO: Need more explanations.
-    // big_final_approx = (((x < 0.0f) & PI) + ((x1 * 2) | negative_mask));
-    __m128 big_final_approx = _mm_add_ps(
-                                  _mm_and_ps(_mm_cmplt_ps(x, magic_zero), magic_pi),
-                                  _mm_or_ps(_mm_add_ps(x1, x1), negative_mask));
-
-    // return (is_small_input ? small_final_approx : big_final_approx);
-    return _mm_or_ps(
-               _mm_and_ps(is_small_input, small_final_approx),
-               _mm_andnot_ps(is_small_input, big_final_approx));
-}
-
-static NCNN_FORCEINLINE __m128 atan_ps(__m128 x)
-{
-    const __m128 magic_negative_zero = _mm_set_ps1(-0.0f);
-    const __m128 magic_one = _mm_set_ps1(1.0f);
-    const __m128 magic_negative_one = _mm_set_ps1(-1.0f);
-    const __m128 magic_half_pi = _mm_set_ps1(1.5707964f);
-    const __m128 magic_a0 = _mm_set_ps1(1.0f);
-    const __m128 magic_a1 = _mm_set_ps1(-0.33333072f);
-    const __m128 magic_a2 = _mm_set_ps1(0.1999262f);
-    const __m128 magic_a3 = _mm_set_ps1(-0.14203644f);
-    const __m128 magic_a4 = _mm_set_ps1(0.10640934f);
-    const __m128 magic_a5 = _mm_set_ps1(-0.07504295f);
-    const __m128 magic_a6 = _mm_set_ps1(0.04269152f);
-    const __m128 magic_a7 = _mm_set_ps1(-0.01606863f);
-    const __m128 magic_a8 = _mm_set_ps1(0.0028498897f);
-
-    // negative_mask = magic_negative_zero && x;
-    __m128 negative_mask = _mm_and_ps(magic_negative_zero, x);
-
-    // absolute = abs(x);
-    __m128 absolute = _mm_andnot_ps(magic_negative_zero, x);
-
-    // Reference: https://en.wikipedia.org/wiki/Small-angle_approximation
-
-    // is_small_input = (1.0f < absolute);
-    __m128 is_small_input = _mm_cmplt_ps(magic_one, absolute);
-
-    // x1 = (is_small_input ? -1.0f : absolute);
-    // x2 = (is_small_input ? absolute : 1.0f)
-    // input_approx = x1 / x2;
-    __m128 input_approx = _mm_div_ps(
-                              _mm_or_ps(
-                                  _mm_and_ps(is_small_input, magic_negative_one),
-                                  _mm_andnot_ps(is_small_input, absolute)),
-                              _mm_or_ps(
-                                  _mm_and_ps(is_small_input, absolute),
-                                  _mm_andnot_ps(is_small_input, magic_one)));
-
-    // square_of_input_approx = input_approx * input_approx;
-    __m128 square_of_input_approx = _mm_mul_ps(input_approx, input_approx);
-
-    // fourth_power_of_input_approx =
-    //     square_of_input_approx * square_of_input_approx;
-    __m128 fourth_power_of_input_approx = _mm_mul_ps(
-            square_of_input_approx, square_of_input_approx);
-
-    // TODO: Need more explanations.
-    // x1 = ((fourth_power_of_input_approx * magic_a7) + magic_a5);
-    // x2 = ((fourth_power_of_input_approx * magic_a8) + magic_a6);
-    // x3 = ((fourth_power_of_input_approx * x1) + magic_a3);
-    // x4 = ((fourth_power_of_input_approx * x2) + magic_a4);
-    // x5 = ((fourth_power_of_input_approx * x3) + magic_a1);
-    // x6 = ((fourth_power_of_input_approx * x4) + magic_a2);
-    // x7 = ((fourth_power_of_input_approx * x6) + magic_a0);
-    // output_approx = ((square_of_input_approx * x5) + x7);
-    __m128 output_approx = _mm_comp_fmadd_ps(
-                               square_of_input_approx,
-                               _mm_comp_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm_comp_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       _mm_comp_fmadd_ps(
-                                           fourth_power_of_input_approx,
-                                           magic_a7,
-                                           magic_a5),
-                                       magic_a3),
-                                   magic_a1),
-                               _mm_comp_fmadd_ps(
-                                   fourth_power_of_input_approx,
-                                   _mm_comp_fmadd_ps(
-                                       fourth_power_of_input_approx,
-                                       _mm_comp_fmadd_ps(
-                                           fourth_power_of_input_approx,
-                                           _mm_comp_fmadd_ps(
-                                                   fourth_power_of_input_approx,
-                                                   magic_a8,
-                                                   magic_a6),
-                                           magic_a4),
-                                       magic_a2),
-                                   magic_a0));
-
-    // TODO: Need more explanations.
-    // x1 = (output_approx * input_approx);
-    // if (is_small_input) x1 += (0.5 * PI);
-    // return (negative_mask ? -x1 : x1);
-    return _mm_or_ps(
-               _mm_add_ps(
-                   _mm_mul_ps(output_approx, input_approx),
-                   _mm_and_ps(is_small_input, magic_half_pi)),
-               negative_mask);
-}
-
-static NCNN_FORCEINLINE __m128 atan2_ps(__m128 y, __m128 x)
-{
-    // Reference: https://mazzo.li/posts/vectorized-atan2.html
-
-    const __m128 magic_zero = _mm_set_ps1(0.0f);
-    const __m128 magic_negative_zero = _mm_set_ps1(-0.0f);
-    const __m128 magic_pi = _mm_set_ps1(3.1415927f);
-    const __m128 magic_half_pi = _mm_set_ps1(1.5707964f);
-
-    // not_equal_zero_x = (x != 0.0f);
-    __m128 not_equal_zero_x = _mm_cmpneq_ps(x, magic_zero);
-
-    // not_equal_zero_y = (y != 0.0f);
-    __m128 not_equal_zero_y = _mm_cmpneq_ps(y, magic_zero);
-
-    // normal_mode = ((x != 0.0f) & (y != 0.0f));
-    __m128 normal_mode = _mm_and_ps(not_equal_zero_x, not_equal_zero_y);
-
-    // negative_mask_x = magic_negative_zero && x;
-    __m128 negative_mask_x = _mm_and_ps(magic_negative_zero, x);
-
-    // negative_mask_y = magic_negative_zero && y;
-    __m128 negative_mask_y = _mm_and_ps(magic_negative_zero, y);
-
-    // pi_additions = ((x < 0.0f) ? ((y < 0.0f) ? -PI : PI) : 0.0f);
-    __m128 pi_additions = _mm_and_ps(
-                              _mm_cmplt_ps(x, magic_zero),
-                              _mm_or_ps(
-                                  _mm_and_ps(
-                                      _mm_cmplt_ps(y, magic_zero),
-                                      magic_negative_zero),
-                                  magic_pi));
-
-    // normal_result = (atan(y / x) + pi_additions);
-    __m128 normal_result = _mm_add_ps(
-                               atan_ps(_mm_div_ps(y, x)),
-                               pi_additions);
-
-    // negative_mask_full_x = ((negative_mask_x | PI) < 0.0f);
-    __m128 negative_mask_full_x = _mm_cmplt_ps(
-                                      _mm_or_ps(negative_mask_x, magic_pi),
-                                      magic_zero);
-
-    // x1 = (negative_mask_y ? -(0.5 * PI) : (0.5 * PI));
-    // x2 = (negative_mask_full_x ? PI : 0.0f);
-    // special_result = ((y != 0.0f) ? x1 : x2);
-    __m128 special_result = _mm_or_ps(
-                                _mm_and_ps(
-                                    not_equal_zero_y,
-                                    _mm_or_ps(negative_mask_y, magic_half_pi)),
-                                _mm_andnot_ps(
-                                    not_equal_zero_y,
-                                    _mm_or_ps(
-                                        _mm_and_ps(negative_mask_full_x, magic_pi),
-                                        _mm_andnot_ps(negative_mask_full_x, magic_zero))));
-
-    // return (normal_mode ? normal_result : special_result);
-    return _mm_or_ps(
-               _mm_and_ps(normal_mode, normal_result),
-               _mm_andnot_ps(normal_mode, special_result));
-}
-
-static NCNN_FORCEINLINE __m128 abs_ps(__m128 inputs)
-{
-    // Use negative zero as the sign bit mask.
-    const __m128 magic_negative_zero = _mm_set_ps1(-0.0f);
-
-    // return (!magic_negative_zero && x);
-    return _mm_andnot_ps(magic_negative_zero, inputs);
-}
-
-#endif // SSE_MATHFUN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/swish_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/swish_x86.h
deleted file mode 100644
index 76b7c3d8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/swish_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_SWISH_X86_H
-#define LAYER_SWISH_X86_H
-
-#include "swish.h"
-
-namespace ncnn {
-
-class Swish_x86 : public Swish
-{
-public:
-    Swish_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_SWISH_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/tanh_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/tanh_x86.h
deleted file mode 100644
index e4c4477b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/tanh_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_TANH_X86_H
-#define LAYER_TANH_X86_H
-
-#include "tanh.h"
-
-namespace ncnn {
-
-class TanH_x86 : public TanH
-{
-public:
-    TanH_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_TANH_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/unaryop_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/unaryop_x86.h
deleted file mode 100644
index 0e4a7ff5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/unaryop_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_UNARYOP_X86_H
-#define LAYER_UNARYOP_X86_H
-
-#include "unaryop.h"
-
-namespace ncnn {
-
-class UnaryOp_x86 : public UnaryOp
-{
-public:
-    UnaryOp_x86();
-
-    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_UNARYOP_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/x86_activation.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/x86_activation.h
deleted file mode 100644
index 691bc65e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/x86_activation.h
+++ /dev/null
@@ -1,327 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef X86_ACTIVATION_H
-#define X86_ACTIVATION_H
-
-#include "mat.h"
-#include "fused_activation.h"
-#include "x86_usability.h"
-
-#if __SSE2__
-#include <emmintrin.h>
-#include "sse_mathfun.h"
-
-static NCNN_FORCEINLINE __m128 sigmoid_sse(__m128 inputs)
-{
-    const __m128 one = _mm_set1_ps(1.0f);
-    return _mm_div_ps(one, _mm_add_ps(one, exp_ps(_mm_sub_ps(_mm_setzero_ps(), inputs))));
-}
-
-static NCNN_FORCEINLINE __m128 tanh_sse(__m128 inputs)
-{
-    const __m128 one = _mm_set1_ps(1.0f);
-    const __m128 two = _mm_set1_ps(2.0f);
-    return _mm_sub_ps(_mm_mul_ps(sigmoid_sse(_mm_mul_ps(inputs, two)), two), one);
-}
-
-static NCNN_FORCEINLINE __m128 mish_sse(__m128 inputs)
-{
-    return _mm_mul_ps(inputs, tanh_sse(log_ps(_mm_add_ps(exp_ps(inputs), _mm_set1_ps(1.f)))));
-}
-
-static NCNN_FORCEINLINE __m128 swish_sse(__m128 inputs)
-{
-    return _mm_mul_ps(inputs, sigmoid_sse(inputs));
-}
-
-static NCNN_FORCEINLINE __m128 hardswish_sse(__m128 inputs, __m128 a, __m128 b)
-{
-    const __m128 one = _mm_set1_ps(1.0f);
-    b = _mm_add_ps(_mm_mul_ps(inputs, a), b);
-    b = _mm_max_ps(b, _mm_setzero_ps());
-    b = _mm_min_ps(b, one);
-    return _mm_mul_ps(b, inputs);
-}
-
-static NCNN_FORCEINLINE __m128 lrelu_sse(__m128 inputs, float slope)
-{
-    __m128 pos = _mm_max_ps(_mm_setzero_ps(), inputs);
-    __m128 neg = _mm_min_ps(_mm_setzero_ps(), inputs);
-    return _mm_add_ps(pos, _mm_mul_ps(_mm_set1_ps(slope), neg));
-}
-
-static NCNN_FORCEINLINE __m128 prelu_sse(__m128 inputs, __m128 alphas)
-{
-    __m128 pos = _mm_max_ps(_mm_setzero_ps(), inputs);
-    __m128 neg = _mm_min_ps(_mm_setzero_ps(), inputs);
-    return _mm_add_ps(pos, _mm_mul_ps(alphas, neg));
-}
-
-static NCNN_FORCEINLINE __m128 elu_sse(__m128 inputs, __m128 alphas)
-{
-    __m128 pos = _mm_max_ps(_mm_setzero_ps(), inputs);
-    __m128 neg = _mm_min_ps(_mm_setzero_ps(), inputs);
-    neg = _mm_sub_ps(exp_ps(neg), _mm_set1_ps(1.f));
-    return _mm_add_ps(pos, _mm_mul_ps(alphas, neg));
-}
-
-static NCNN_FORCEINLINE __m128 activation_sse(__m128 _v, int activation_type, const ncnn::Mat& activation_params)
-{
-    // Process fused activations
-    switch (activation_type)
-    {
-    case 1:
-    {
-        // Relu
-        return _mm_max_ps(_v, _mm_setzero_ps());
-    }
-    case 2:
-    {
-        // Leaky relu
-        return lrelu_sse(_v, activation_params[0]);
-    }
-    case 3:
-    {
-        // min max clip
-        __m128 min = _mm_set1_ps(activation_params[0]);
-        __m128 max = _mm_set1_ps(activation_params[1]);
-        return _mm_min_ps(_mm_max_ps(_v, min), max);
-    }
-    case 4:
-    {
-        // Sigmoid
-        return sigmoid_sse(_v);
-    }
-    case 5:
-    {
-        return mish_sse(_v);
-    }
-    case 6:
-    {
-        __m128 _a = _mm_set1_ps(activation_params[0]);
-        __m128 _b = _mm_set1_ps(activation_params[1]);
-        return hardswish_sse(_v, _a, _b);
-    }
-    }
-
-    return _v;
-}
-
-#if __AVX__
-#include <immintrin.h>
-#include "avx_mathfun.h"
-
-static NCNN_FORCEINLINE __m256 sigmoid_avx(__m256 inputs)
-{
-    const __m256 one = _mm256_set1_ps(1.0f);
-    return _mm256_div_ps(one, _mm256_add_ps(one, exp256_ps(_mm256_sub_ps(_mm256_setzero_ps(), inputs))));
-}
-
-static NCNN_FORCEINLINE __m256 tanh_avx(__m256 inputs)
-{
-    const __m256 one = _mm256_set1_ps(1.0f);
-    const __m256 two = _mm256_set1_ps(2.0f);
-#if __FMA__
-    return _mm256_fmsub_ps(sigmoid_avx(_mm256_mul_ps(inputs, two)), two, one);
-#else
-    return _mm256_sub_ps(_mm256_mul_ps(sigmoid_avx(_mm256_mul_ps(inputs, two)), two), one);
-#endif
-}
-
-static NCNN_FORCEINLINE __m256 mish_avx(__m256 inputs)
-{
-    return _mm256_mul_ps(inputs, tanh_avx(log256_ps(_mm256_add_ps(exp256_ps(inputs), _mm256_set1_ps(1.f)))));
-}
-
-static NCNN_FORCEINLINE __m256 swish_avx(__m256 inputs)
-{
-    return _mm256_mul_ps(inputs, sigmoid_avx(inputs));
-}
-
-static NCNN_FORCEINLINE __m256 hardswish_avx(__m256 inputs, __m256 a, __m256 b)
-{
-    const __m256 one = _mm256_set1_ps(1.0f);
-    b = _mm256_comp_fmadd_ps(inputs, a, b);
-    b = _mm256_max_ps(b, _mm256_setzero_ps());
-    b = _mm256_min_ps(b, one);
-    return _mm256_mul_ps(b, inputs);
-}
-
-static NCNN_FORCEINLINE __m256 lrelu_avx(__m256 inputs, float slope)
-{
-    __m256 pos = _mm256_max_ps(_mm256_setzero_ps(), inputs);
-    __m256 neg = _mm256_min_ps(_mm256_setzero_ps(), inputs);
-    return _mm256_add_ps(pos, _mm256_mul_ps(_mm256_set1_ps(slope), neg));
-}
-
-static NCNN_FORCEINLINE __m256 prelu_avx(__m256 inputs, __m256 alphas)
-{
-    __m256 pos = _mm256_max_ps(_mm256_setzero_ps(), inputs);
-    __m256 neg = _mm256_min_ps(_mm256_setzero_ps(), inputs);
-    return _mm256_add_ps(pos, _mm256_mul_ps(alphas, neg));
-}
-
-static NCNN_FORCEINLINE __m256 elu_avx(__m256 inputs, __m256 alphas)
-{
-    __m256 pos = _mm256_max_ps(_mm256_setzero_ps(), inputs);
-    __m256 neg = _mm256_min_ps(_mm256_setzero_ps(), inputs);
-    neg = _mm256_sub_ps(exp256_ps(neg), _mm256_set1_ps(1.f));
-    return _mm256_add_ps(pos, _mm256_mul_ps(alphas, neg));
-}
-
-static NCNN_FORCEINLINE __m256 activation_avx(__m256 _v, int activation_type, const ncnn::Mat& activation_params)
-{
-    // Process fused activations
-    switch (activation_type)
-    {
-    case 1:
-    {
-        // Relu
-        return _mm256_max_ps(_v, _mm256_setzero_ps());
-    }
-    case 2:
-    {
-        // Leaky relu
-        return lrelu_avx(_v, activation_params[0]);
-    }
-    case 3:
-    {
-        // min max clip
-        __m256 min = _mm256_set1_ps(activation_params[0]);
-        __m256 max = _mm256_set1_ps(activation_params[1]);
-        return _mm256_min_ps(_mm256_max_ps(_v, min), max);
-    }
-    case 4:
-    {
-        // Sigmoid
-        return sigmoid_avx(_v);
-    }
-    case 5:
-    {
-        return mish_avx(_v);
-    }
-    case 6:
-    {
-        __m256 _a = _mm256_set1_ps(activation_params[0]);
-        __m256 _b = _mm256_set1_ps(activation_params[1]);
-        return hardswish_avx(_v, _a, _b);
-    }
-    }
-
-    return _v;
-}
-
-#if __AVX512F__
-#include "avx512_mathfun.h"
-
-static NCNN_FORCEINLINE __m512 sigmoid_avx512(__m512 inputs)
-{
-    const __m512 one = _mm512_set1_ps(1.0f);
-    return _mm512_div_ps(one, _mm512_add_ps(one, exp512_ps(_mm512_sub_ps(_mm512_setzero_ps(), inputs))));
-}
-
-static NCNN_FORCEINLINE __m512 tanh_avx512(__m512 inputs)
-{
-    const __m512 one = _mm512_set1_ps(1.0f);
-    const __m512 two = _mm512_set1_ps(2.0f);
-    return _mm512_fmsub_ps(sigmoid_avx512(_mm512_mul_ps(inputs, two)), two, one);
-}
-
-static NCNN_FORCEINLINE __m512 mish_avx512(__m512 inputs)
-{
-    return _mm512_mul_ps(inputs, tanh_avx512(log512_ps(_mm512_add_ps(exp512_ps(inputs), _mm512_set1_ps(1.f)))));
-}
-
-static NCNN_FORCEINLINE __m512 swish_avx512(__m512 inputs)
-{
-    return _mm512_mul_ps(inputs, sigmoid_avx512(inputs));
-}
-
-static NCNN_FORCEINLINE __m512 hardswish_avx512(__m512 inputs, __m512 a, __m512 b)
-{
-    const __m512 one = _mm512_set1_ps(1.0f);
-    b = _mm512_fmadd_ps(inputs, a, b);
-    b = _mm512_max_ps(b, _mm512_setzero_ps());
-    b = _mm512_min_ps(b, one);
-    return _mm512_mul_ps(b, inputs);
-}
-
-static NCNN_FORCEINLINE __m512 lrelu_avx512(__m512 inputs, float slope)
-{
-    __mmask16 _is_negative = _mm512_cmp_ps_mask(inputs, _mm512_setzero_ps(), _CMP_LT_OQ);
-    return _mm512_mask_mul_ps(inputs, _is_negative, inputs, _mm512_set1_ps(slope));
-}
-
-static NCNN_FORCEINLINE __m512 elu_avx512(__m512 inputs, __m512 alphas)
-{
-    __m512 pos = _mm512_max_ps(_mm512_setzero_ps(), inputs);
-    __m512 neg = _mm512_min_ps(_mm512_setzero_ps(), inputs);
-    neg = _mm512_sub_ps(exp512_ps(neg), _mm512_set1_ps(1.f));
-    return _mm512_add_ps(pos, _mm512_mul_ps(alphas, neg));
-}
-
-static NCNN_FORCEINLINE __m512 prelu_avx512(__m512 inputs, __m512 alphas)
-{
-    __m512 pos = _mm512_max_ps(_mm512_setzero_ps(), inputs);
-    __m512 neg = _mm512_min_ps(_mm512_setzero_ps(), inputs);
-    return _mm512_add_ps(pos, _mm512_mul_ps(alphas, neg));
-}
-
-static NCNN_FORCEINLINE __m512 activation_avx512(__m512 _v, int activation_type, const ncnn::Mat& activation_params)
-{
-    // Process fused activations
-    switch (activation_type)
-    {
-    case 1:
-    {
-        // Relu
-        return _mm512_max_ps(_v, _mm512_setzero_ps());
-    }
-    case 2:
-    {
-        // Leaky relu
-        return lrelu_avx512(_v, activation_params[0]);
-    }
-    case 3:
-    {
-        // min max clip
-        __m512 min = _mm512_set1_ps(activation_params[0]);
-        __m512 max = _mm512_set1_ps(activation_params[1]);
-        return _mm512_min_ps(_mm512_max_ps(_v, min), max);
-    }
-    case 4:
-    {
-        // Sigmoid
-        return sigmoid_avx512(_v);
-    }
-    case 5:
-    {
-        return mish_avx512(_v);
-    }
-    case 6:
-    {
-        __m512 _a = _mm512_set1_ps(activation_params[0]);
-        __m512 _b = _mm512_set1_ps(activation_params[1]);
-        return hardswish_avx512(_v, _a, _b);
-    }
-    }
-
-    return _v;
-}
-#endif // __AVX512F__
-#endif // __AVX__
-#endif // __SSE2__
-
-#endif // X86_ACTIVATION_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/x86_usability.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/x86_usability.h
deleted file mode 100644
index 583f4b10..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/x86_usability.h
+++ /dev/null
@@ -1,1344 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef X86_USABILITY_H
-#define X86_USABILITY_H
-
-#include <stdint.h>
-#if __SSE2__
-#include <emmintrin.h>
-#if __SSE4_1__
-#include <smmintrin.h>
-#if __AVX__
-#include <immintrin.h>
-#if __XOP__
-#ifdef _MSC_VER
-#include <ammintrin.h>
-#else
-#include <x86intrin.h>
-#endif
-#endif
-#endif
-#endif
-#endif // __SSE2__
-
-static NCNN_FORCEINLINE signed char float2int8(float v)
-{
-    int int32 = (int)round(v);
-    if (int32 > 127) return 127;
-    if (int32 < -127) return -127;
-    return (signed char)int32;
-}
-
-#if __SSE2__
-static NCNN_FORCEINLINE void transpose4x8_epi32(__m128i& _r0, __m128i& _r1, __m128i& _r2, __m128i& _r3, __m128i& _r4, __m128i& _r5, __m128i& _r6, __m128i& _r7)
-{
-    __m128i _tmp0 = _mm_unpacklo_epi32(_r0, _r1);
-    __m128i _tmp1 = _mm_unpackhi_epi32(_r0, _r1);
-    __m128i _tmp2 = _mm_unpacklo_epi32(_r2, _r3);
-    __m128i _tmp3 = _mm_unpackhi_epi32(_r2, _r3);
-    __m128i _tmp4 = _mm_unpacklo_epi32(_r4, _r5);
-    __m128i _tmp5 = _mm_unpackhi_epi32(_r4, _r5);
-    __m128i _tmp6 = _mm_unpacklo_epi32(_r6, _r7);
-    __m128i _tmp7 = _mm_unpackhi_epi32(_r6, _r7);
-
-    _r0 = _mm_unpacklo_epi64(_tmp0, _tmp2);
-    _r1 = _mm_unpacklo_epi64(_tmp4, _tmp6);
-    _r2 = _mm_unpackhi_epi64(_tmp0, _tmp2);
-    _r3 = _mm_unpackhi_epi64(_tmp4, _tmp6);
-    _r4 = _mm_unpacklo_epi64(_tmp1, _tmp3);
-    _r5 = _mm_unpacklo_epi64(_tmp5, _tmp7);
-    _r6 = _mm_unpackhi_epi64(_tmp1, _tmp3);
-    _r7 = _mm_unpackhi_epi64(_tmp5, _tmp7);
-}
-
-static NCNN_FORCEINLINE void transpose4x4_epi32(__m128i& _r0, __m128i& _r1, __m128i& _r2, __m128i& _r3)
-{
-    __m128i _tmp0 = _mm_unpacklo_epi32(_r0, _r1);
-    __m128i _tmp1 = _mm_unpackhi_epi32(_r0, _r1);
-    __m128i _tmp2 = _mm_unpacklo_epi32(_r2, _r3);
-    __m128i _tmp3 = _mm_unpackhi_epi32(_r2, _r3);
-
-    _r0 = _mm_unpacklo_epi64(_tmp0, _tmp2);
-    _r1 = _mm_unpackhi_epi64(_tmp0, _tmp2);
-    _r2 = _mm_unpacklo_epi64(_tmp1, _tmp3);
-    _r3 = _mm_unpackhi_epi64(_tmp1, _tmp3);
-}
-
-static NCNN_FORCEINLINE void transpose8x8_epi16(__m128i& _r0, __m128i& _r1, __m128i& _r2, __m128i& _r3, __m128i& _r4, __m128i& _r5, __m128i& _r6, __m128i& _r7)
-{
-    __m128i _tmp0 = _mm_unpacklo_epi16(_r0, _r1);
-    __m128i _tmp1 = _mm_unpackhi_epi16(_r0, _r1);
-    __m128i _tmp2 = _mm_unpacklo_epi16(_r2, _r3);
-    __m128i _tmp3 = _mm_unpackhi_epi16(_r2, _r3);
-    __m128i _tmp4 = _mm_unpacklo_epi16(_r4, _r5);
-    __m128i _tmp5 = _mm_unpackhi_epi16(_r4, _r5);
-    __m128i _tmp6 = _mm_unpacklo_epi16(_r6, _r7);
-    __m128i _tmp7 = _mm_unpackhi_epi16(_r6, _r7);
-
-    __m128i _tmp8 = _mm_unpacklo_epi32(_tmp0, _tmp2);
-    __m128i _tmp9 = _mm_unpackhi_epi32(_tmp0, _tmp2);
-    __m128i _tmpa = _mm_unpacklo_epi32(_tmp1, _tmp3);
-    __m128i _tmpb = _mm_unpackhi_epi32(_tmp1, _tmp3);
-    __m128i _tmpc = _mm_unpacklo_epi32(_tmp4, _tmp6);
-    __m128i _tmpd = _mm_unpackhi_epi32(_tmp4, _tmp6);
-    __m128i _tmpe = _mm_unpacklo_epi32(_tmp5, _tmp7);
-    __m128i _tmpf = _mm_unpackhi_epi32(_tmp5, _tmp7);
-
-    _r0 = _mm_unpacklo_epi64(_tmp8, _tmpc);
-    _r1 = _mm_unpackhi_epi64(_tmp8, _tmpc);
-    _r2 = _mm_unpacklo_epi64(_tmp9, _tmpd);
-    _r3 = _mm_unpackhi_epi64(_tmp9, _tmpd);
-    _r4 = _mm_unpacklo_epi64(_tmpa, _tmpe);
-    _r5 = _mm_unpackhi_epi64(_tmpa, _tmpe);
-    _r6 = _mm_unpacklo_epi64(_tmpb, _tmpf);
-    _r7 = _mm_unpackhi_epi64(_tmpb, _tmpf);
-}
-
-static NCNN_FORCEINLINE void transpose8x4_epi16(__m128i& _r0, __m128i& _r1, __m128i& _r2, __m128i& _r3)
-{
-    __m128i _tmp0 = _mm_unpacklo_epi16(_r0, _r1);
-    __m128i _tmp1 = _mm_unpackhi_epi16(_r0, _r1);
-    __m128i _tmp2 = _mm_unpacklo_epi16(_r2, _r3);
-    __m128i _tmp3 = _mm_unpackhi_epi16(_r2, _r3);
-
-    _r0 = _mm_unpacklo_epi32(_tmp0, _tmp2);
-    _r1 = _mm_unpackhi_epi32(_tmp0, _tmp2);
-    _r2 = _mm_unpacklo_epi32(_tmp1, _tmp3);
-    _r3 = _mm_unpackhi_epi32(_tmp1, _tmp3);
-}
-
-static NCNN_FORCEINLINE float _mm_reduce_add_ps(__m128 x128)
-{
-    const __m128 x64 = _mm_add_ps(x128, _mm_movehl_ps(x128, x128));
-    const __m128 x32 = _mm_add_ss(x64, _mm_shuffle_ps(x64, x64, 0x55));
-    return _mm_cvtss_f32(x32);
-}
-
-static NCNN_FORCEINLINE float _mm_reduce_max_ps(__m128 x128)
-{
-    const __m128 x64 = _mm_max_ps(x128, _mm_movehl_ps(x128, x128));
-    const __m128 x32 = _mm_max_ss(x64, _mm_shuffle_ps(x64, x64, 0x55));
-    return _mm_cvtss_f32(x32);
-}
-
-static NCNN_FORCEINLINE int _mm_reduce_add_epi32(__m128i x)
-{
-    __m128i hi64 = _mm_unpackhi_epi64(x, x);
-    __m128i sum64 = _mm_add_epi32(hi64, x);
-    __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
-    __m128i sum32 = _mm_add_epi32(sum64, hi32);
-    return _mm_cvtsi128_si32(sum32);
-}
-
-static NCNN_FORCEINLINE int32_t float2int8_sse(const __m128& _v0)
-{
-    // _MM_ROUND_NEAREST round to even
-    // simulate round to nearest via +/-0.5 with round to zero
-    __m128 _p5 = _mm_set1_ps(0.5f);
-    __m128 _signmask = _mm_castsi128_ps(_mm_set1_epi32(1 << 31));
-    __m128 _sign0 = _mm_and_ps(_v0, _signmask);
-    __m128 _v0_p5 = _mm_or_ps(_p5, _sign0);
-    __m128 _v0_adj = _mm_add_ps(_v0, _v0_p5);
-    __m128i _v0_i = _mm_cvttps_epi32(_v0_adj);
-
-    __m128i _v0_s16 = _mm_packs_epi32(_v0_i, _v0_i);
-
-    _v0_s16 = _mm_min_epi16(_v0_s16, _mm_set1_epi16(127));
-    _v0_s16 = _mm_max_epi16(_v0_s16, _mm_set1_epi16(-127));
-
-    __m128i _v8 = _mm_packs_epi16(_v0_s16, _v0_s16);
-
-#if defined(__x86_64__) || defined(_M_X64)
-    return (int32_t)_mm_cvtsi128_si64(_v8);
-#else
-    return _mm_cvtsi128_si32(_v8);
-#endif
-}
-
-static NCNN_FORCEINLINE int64_t float2int8_sse(const __m128& _v0, const __m128& _v1)
-{
-    // _MM_ROUND_NEAREST round to even
-    // simulate round to nearest via +/-0.5 with round to zero
-    __m128 _p5 = _mm_set1_ps(0.5f);
-    __m128 _signmask = _mm_castsi128_ps(_mm_set1_epi32(1 << 31));
-    __m128 _sign0 = _mm_and_ps(_v0, _signmask);
-    __m128 _sign1 = _mm_and_ps(_v1, _signmask);
-    __m128 _v0_p5 = _mm_or_ps(_p5, _sign0);
-    __m128 _v1_p5 = _mm_or_ps(_p5, _sign1);
-    __m128 _v0_adj = _mm_add_ps(_v0, _v0_p5);
-    __m128 _v1_adj = _mm_add_ps(_v1, _v1_p5);
-    __m128i _v0_i = _mm_cvttps_epi32(_v0_adj);
-    __m128i _v1_i = _mm_cvttps_epi32(_v1_adj);
-
-    __m128i _v01_s16 = _mm_packs_epi32(_v0_i, _v1_i);
-
-    _v01_s16 = _mm_min_epi16(_v01_s16, _mm_set1_epi16(127));
-    _v01_s16 = _mm_max_epi16(_v01_s16, _mm_set1_epi16(-127));
-
-    __m128i _v8 = _mm_packs_epi16(_v01_s16, _v01_s16);
-
-#if defined(__x86_64__) || defined(_M_X64)
-    return _mm_cvtsi128_si64(_v8);
-#else
-    int64_t v8[2];
-    _mm_storeu_si128((__m128i*)v8, _v8);
-    return v8[0];
-#endif
-}
-
-static NCNN_FORCEINLINE __m128i float2int8_sse(const __m128& _v0, const __m128& _v1, const __m128& _v2, const __m128& _v3)
-{
-    // _MM_ROUND_NEAREST round to even
-    // simulate round to nearest via +/-0.5 with round to zero
-    __m128 _p5 = _mm_set1_ps(0.5f);
-    __m128 _signmask = _mm_castsi128_ps(_mm_set1_epi32(1 << 31));
-    __m128 _sign0 = _mm_and_ps(_v0, _signmask);
-    __m128 _sign1 = _mm_and_ps(_v1, _signmask);
-    __m128 _sign2 = _mm_and_ps(_v2, _signmask);
-    __m128 _sign3 = _mm_and_ps(_v3, _signmask);
-    __m128 _v0_p5 = _mm_or_ps(_p5, _sign0);
-    __m128 _v1_p5 = _mm_or_ps(_p5, _sign1);
-    __m128 _v2_p5 = _mm_or_ps(_p5, _sign2);
-    __m128 _v3_p5 = _mm_or_ps(_p5, _sign3);
-    __m128 _v0_adj = _mm_add_ps(_v0, _v0_p5);
-    __m128 _v1_adj = _mm_add_ps(_v1, _v1_p5);
-    __m128 _v2_adj = _mm_add_ps(_v2, _v2_p5);
-    __m128 _v3_adj = _mm_add_ps(_v3, _v3_p5);
-    __m128i _v0_i = _mm_cvttps_epi32(_v0_adj);
-    __m128i _v1_i = _mm_cvttps_epi32(_v1_adj);
-    __m128i _v2_i = _mm_cvttps_epi32(_v2_adj);
-    __m128i _v3_i = _mm_cvttps_epi32(_v3_adj);
-
-    __m128i _v01_s16 = _mm_packs_epi32(_v0_i, _v1_i);
-    __m128i _v23_s16 = _mm_packs_epi32(_v2_i, _v3_i);
-
-    _v01_s16 = _mm_min_epi16(_v01_s16, _mm_set1_epi16(127));
-    _v23_s16 = _mm_min_epi16(_v23_s16, _mm_set1_epi16(127));
-    _v01_s16 = _mm_max_epi16(_v01_s16, _mm_set1_epi16(-127));
-    _v23_s16 = _mm_max_epi16(_v23_s16, _mm_set1_epi16(-127));
-
-    __m128i _v8 = _mm_packs_epi16(_v01_s16, _v23_s16);
-
-    return _v8;
-}
-
-static NCNN_FORCEINLINE __m128 bfloat2float_sse(const __m128i& v0)
-{
-    __m128i _zero = _mm_setzero_si128();
-    __m128i _a = _mm_unpacklo_epi16(_zero, v0);
-    __m128 _v = _mm_castsi128_ps(_a);
-    return _v;
-}
-
-static NCNN_FORCEINLINE __m128i float2bfloat_sse(const __m128& v0, const __m128& v1)
-{
-#if __AVX512BF16__
-    __m128i _v = (__m128i)_mm256_cvtneps_pbh(_mm256_insertf128_ps(_mm256_castps128_ps256(v0), v1, 1));
-#else
-    __m128i _a = _mm_castps_si128(v0);
-    __m128i _b = _mm_castps_si128(v1);
-#if __SSE4_1__
-    _a = _mm_srli_epi32(_a, 16);
-    _b = _mm_srli_epi32(_b, 16);
-    __m128i _v = _mm_packus_epi32(_a, _b);
-#else
-    _a = _mm_shufflelo_epi16(_a, _MM_SHUFFLE(2, 0, 3, 1));
-    _b = _mm_shufflelo_epi16(_b, _MM_SHUFFLE(2, 0, 3, 1));
-    _a = _mm_shufflehi_epi16(_a, _MM_SHUFFLE(2, 0, 3, 1));
-    _b = _mm_shufflehi_epi16(_b, _MM_SHUFFLE(2, 0, 3, 1));
-    __m128i _v = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(_a), _mm_castsi128_ps(_b), _MM_SHUFFLE(2, 0, 2, 0)));
-#endif
-#endif
-    return _v;
-}
-
-#ifndef __FMA__
-static NCNN_FORCEINLINE __m128 _mm_comp_fmadd_ps(const __m128& _a, const __m128& _b, const __m128& _c)
-{
-    return _mm_add_ps(_mm_mul_ps(_a, _b), _c);
-}
-static NCNN_FORCEINLINE __m128 _mm_comp_fnmadd_ps(const __m128& _a, const __m128& _b, const __m128& _c)
-{
-    return _mm_sub_ps(_c, _mm_mul_ps(_a, _b));
-}
-static NCNN_FORCEINLINE __m128 _mm_comp_fmsub_ps(const __m128& _a, const __m128& _b, const __m128& _c)
-{
-    return _mm_sub_ps(_mm_mul_ps(_a, _b), _c);
-}
-static NCNN_FORCEINLINE __m128 _mm_comp_fnmsub_ps(const __m128& _a, const __m128& _b, const __m128& _c)
-{
-    return _mm_sub_ps(_c, _mm_mul_ps(_mm_mul_ps(_a, _b), _mm_set1_ps(-1)));
-}
-#else
-static NCNN_FORCEINLINE __m128 _mm_comp_fmadd_ps(const __m128& _a, const __m128& _b, const __m128& _c)
-{
-    return _mm_fmadd_ps(_a, _b, _c);
-}
-static NCNN_FORCEINLINE __m128 _mm_comp_fnmadd_ps(const __m128& _a, const __m128& _b, const __m128& _c)
-{
-    // return -a * b + c
-    return _mm_fnmadd_ps(_a, _b, _c);
-}
-static NCNN_FORCEINLINE __m128 _mm_comp_fmsub_ps(const __m128& _a, const __m128& _b, const __m128& _c)
-{
-    return _mm_fmsub_ps(_a, _b, _c);
-}
-static NCNN_FORCEINLINE __m128 _mm_comp_fnmsub_ps(const __m128& _a, const __m128& _b, const __m128& _c)
-{
-    return _mm_fnmsub_ps(_a, _b, _c);
-}
-#endif // !__FMA__
-
-#if __AVX__
-#ifndef __FMA__
-static NCNN_FORCEINLINE __m256 _mm256_comp_fmadd_ps(const __m256& _a, const __m256& _b, const __m256& _c)
-{
-    return _mm256_add_ps(_mm256_mul_ps(_a, _b), _c);
-}
-static NCNN_FORCEINLINE __m256 _mm256_comp_fnmadd_ps(const __m256& _a, const __m256& _b, const __m256& _c)
-{
-    return _mm256_sub_ps(_c, _mm256_mul_ps(_a, _b));
-}
-static NCNN_FORCEINLINE __m256 _mm256_comp_fmsub_ps(const __m256& _a, const __m256& _b, const __m256& _c)
-{
-    return _mm256_sub_ps(_mm256_mul_ps(_a, _b), _c);
-}
-static NCNN_FORCEINLINE __m256 _mm256_comp_fnmsub_ps(const __m256& _a, const __m256& _b, const __m256& _c)
-{
-    return _mm256_sub_ps(_c, _mm256_mul_ps(_mm256_mul_ps(_a, _b), _mm256_set1_ps(-1)));
-}
-#else
-static NCNN_FORCEINLINE __m256 _mm256_comp_fmadd_ps(const __m256& _a, const __m256& _b, const __m256& _c)
-{
-    // return a * b + c
-    return _mm256_fmadd_ps(_a, _b, _c);
-}
-static NCNN_FORCEINLINE __m256 _mm256_comp_fnmadd_ps(const __m256& _a, const __m256& _b, const __m256& _c)
-{
-    // return -a * b + c
-    return _mm256_fnmadd_ps(_a, _b, _c);
-}
-static NCNN_FORCEINLINE __m256 _mm256_comp_fmsub_ps(const __m256& _a, const __m256& _b, const __m256& _c)
-{
-    // return a * b - c
-    return _mm256_fmsub_ps(_a, _b, _c);
-}
-static NCNN_FORCEINLINE __m256 _mm256_comp_fnmsub_ps(const __m256& _a, const __m256& _b, const __m256& _c)
-{
-    // return -(a * b) - c
-    return _mm256_fnmsub_ps(_a, _b, _c);
-}
-#endif
-
-static NCNN_FORCEINLINE __m256 _mm256_fmadd_1_ps(const __m256& a, const __m256& b, float c)
-{
-    return _mm256_comp_fmadd_ps(b, _mm256_set1_ps(c), a);
-}
-
-static NCNN_FORCEINLINE __m256 _mm256_fmrsub_1_ps(const __m256& a, const __m256& b, float c)
-{
-    // return a - b * c
-    return _mm256_comp_fnmadd_ps(b, _mm256_set1_ps(c), a);
-}
-
-static NCNN_FORCEINLINE void transpose8x12_ps(__m256& _r0, __m256& _r1, __m256& _r2, __m256& _r3, __m256& _r4, __m256& _r5, __m256& _r6, __m256& _r7,
-        __m256& _r8, __m256& _r9, __m256& _ra, __m256& _rb)
-{
-    __m256 _tmp0 = _mm256_unpacklo_ps(_r0, _r1);
-    __m256 _tmp1 = _mm256_unpackhi_ps(_r0, _r1);
-    __m256 _tmp2 = _mm256_unpacklo_ps(_r2, _r3);
-    __m256 _tmp3 = _mm256_unpackhi_ps(_r2, _r3);
-    __m256 _tmp4 = _mm256_unpacklo_ps(_r4, _r5);
-    __m256 _tmp5 = _mm256_unpackhi_ps(_r4, _r5);
-    __m256 _tmp6 = _mm256_unpacklo_ps(_r6, _r7);
-    __m256 _tmp7 = _mm256_unpackhi_ps(_r6, _r7);
-    __m256 _tmp8 = _mm256_unpacklo_ps(_r8, _r9);
-    __m256 _tmp9 = _mm256_unpackhi_ps(_r8, _r9);
-    __m256 _tmpa = _mm256_unpacklo_ps(_ra, _rb);
-    __m256 _tmpb = _mm256_unpackhi_ps(_ra, _rb);
-
-    __m256 _tmpc = _mm256_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpd = _mm256_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpe = _mm256_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpf = _mm256_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpg = _mm256_shuffle_ps(_tmp4, _tmp6, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmph = _mm256_shuffle_ps(_tmp4, _tmp6, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpi = _mm256_shuffle_ps(_tmp5, _tmp7, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpj = _mm256_shuffle_ps(_tmp5, _tmp7, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpk = _mm256_shuffle_ps(_tmp8, _tmpa, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpl = _mm256_shuffle_ps(_tmp8, _tmpa, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpm = _mm256_shuffle_ps(_tmp9, _tmpb, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpn = _mm256_shuffle_ps(_tmp9, _tmpb, _MM_SHUFFLE(3, 2, 3, 2));
-
-    _r0 = _mm256_permute2f128_ps(_tmpc, _tmpg, _MM_SHUFFLE(0, 2, 0, 0));
-    _r1 = _mm256_permute2f128_ps(_tmpk, _tmpd, _MM_SHUFFLE(0, 2, 0, 0));
-    _r2 = _mm256_permute2f128_ps(_tmph, _tmpl, _MM_SHUFFLE(0, 2, 0, 0));
-    _r3 = _mm256_permute2f128_ps(_tmpe, _tmpi, _MM_SHUFFLE(0, 2, 0, 0));
-    _r4 = _mm256_permute2f128_ps(_tmpm, _tmpf, _MM_SHUFFLE(0, 2, 0, 0));
-    _r5 = _mm256_permute2f128_ps(_tmpj, _tmpn, _MM_SHUFFLE(0, 2, 0, 0));
-    _r6 = _mm256_permute2f128_ps(_tmpc, _tmpg, _MM_SHUFFLE(0, 3, 0, 1));
-    _r7 = _mm256_permute2f128_ps(_tmpk, _tmpd, _MM_SHUFFLE(0, 3, 0, 1));
-    _r8 = _mm256_permute2f128_ps(_tmph, _tmpl, _MM_SHUFFLE(0, 3, 0, 1));
-    _r9 = _mm256_permute2f128_ps(_tmpe, _tmpi, _MM_SHUFFLE(0, 3, 0, 1));
-    _ra = _mm256_permute2f128_ps(_tmpm, _tmpf, _MM_SHUFFLE(0, 3, 0, 1));
-    _rb = _mm256_permute2f128_ps(_tmpj, _tmpn, _MM_SHUFFLE(0, 3, 0, 1));
-}
-
-static NCNN_FORCEINLINE void transpose8x8_ps(__m256& _r0, __m256& _r1, __m256& _r2, __m256& _r3, __m256& _r4, __m256& _r5, __m256& _r6, __m256& _r7)
-{
-    __m256 _tmp0 = _mm256_unpacklo_ps(_r0, _r1);
-    __m256 _tmp1 = _mm256_unpackhi_ps(_r0, _r1);
-    __m256 _tmp2 = _mm256_unpacklo_ps(_r2, _r3);
-    __m256 _tmp3 = _mm256_unpackhi_ps(_r2, _r3);
-    __m256 _tmp4 = _mm256_unpacklo_ps(_r4, _r5);
-    __m256 _tmp5 = _mm256_unpackhi_ps(_r4, _r5);
-    __m256 _tmp6 = _mm256_unpacklo_ps(_r6, _r7);
-    __m256 _tmp7 = _mm256_unpackhi_ps(_r6, _r7);
-
-    __m256 _tmp8 = _mm256_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmp9 = _mm256_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpa = _mm256_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpb = _mm256_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpc = _mm256_shuffle_ps(_tmp4, _tmp6, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpd = _mm256_shuffle_ps(_tmp4, _tmp6, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpe = _mm256_shuffle_ps(_tmp5, _tmp7, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpf = _mm256_shuffle_ps(_tmp5, _tmp7, _MM_SHUFFLE(3, 2, 3, 2));
-
-    _r0 = _mm256_permute2f128_ps(_tmp8, _tmpc, _MM_SHUFFLE(0, 2, 0, 0));
-    _r1 = _mm256_permute2f128_ps(_tmp9, _tmpd, _MM_SHUFFLE(0, 2, 0, 0));
-    _r2 = _mm256_permute2f128_ps(_tmpa, _tmpe, _MM_SHUFFLE(0, 2, 0, 0));
-    _r3 = _mm256_permute2f128_ps(_tmpb, _tmpf, _MM_SHUFFLE(0, 2, 0, 0));
-    _r4 = _mm256_permute2f128_ps(_tmp8, _tmpc, _MM_SHUFFLE(0, 3, 0, 1));
-    _r5 = _mm256_permute2f128_ps(_tmp9, _tmpd, _MM_SHUFFLE(0, 3, 0, 1));
-    _r6 = _mm256_permute2f128_ps(_tmpa, _tmpe, _MM_SHUFFLE(0, 3, 0, 1));
-    _r7 = _mm256_permute2f128_ps(_tmpb, _tmpf, _MM_SHUFFLE(0, 3, 0, 1));
-}
-
-static NCNN_FORCEINLINE void transpose8x4_ps(__m256& _r0, __m256& _r1, __m256& _r2, __m256& _r3)
-{
-    __m256 _tmp0 = _mm256_unpacklo_ps(_r0, _r1);
-    __m256 _tmp1 = _mm256_unpackhi_ps(_r0, _r1);
-    __m256 _tmp2 = _mm256_unpacklo_ps(_r2, _r3);
-    __m256 _tmp3 = _mm256_unpackhi_ps(_r2, _r3);
-
-    __m256 _tmp4 = _mm256_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmp5 = _mm256_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmp6 = _mm256_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmp7 = _mm256_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(3, 2, 3, 2));
-
-    _r0 = _mm256_permute2f128_ps(_tmp4, _tmp5, _MM_SHUFFLE(0, 2, 0, 0));
-    _r1 = _mm256_permute2f128_ps(_tmp6, _tmp7, _MM_SHUFFLE(0, 2, 0, 0));
-    _r2 = _mm256_permute2f128_ps(_tmp4, _tmp5, _MM_SHUFFLE(0, 3, 0, 1));
-    _r3 = _mm256_permute2f128_ps(_tmp6, _tmp7, _MM_SHUFFLE(0, 3, 0, 1));
-}
-
-static NCNN_FORCEINLINE void transpose8x2_ps(__m256& _r0, __m256& _r1)
-{
-    __m256 _tmp0 = _mm256_unpacklo_ps(_r0, _r1);
-    __m256 _tmp1 = _mm256_unpackhi_ps(_r0, _r1);
-
-    _r0 = _mm256_permute2f128_ps(_tmp0, _tmp1, _MM_SHUFFLE(0, 2, 0, 0));
-    _r1 = _mm256_permute2f128_ps(_tmp0, _tmp1, _MM_SHUFFLE(0, 3, 0, 1));
-}
-
-static NCNN_FORCEINLINE void transpose2x8_ps(__m256& _r0, __m256& _r1)
-{
-    __m256 _tmp0 = _mm256_permute2f128_ps(_r0, _r1, _MM_SHUFFLE(0, 2, 0, 0));
-    __m256 _tmp1 = _mm256_permute2f128_ps(_r0, _r1, _MM_SHUFFLE(0, 3, 0, 1));
-
-    _r0 = _mm256_shuffle_ps(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-    _r1 = _mm256_shuffle_ps(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-}
-
-static NCNN_FORCEINLINE void transpose3x8_ps(__m256& _r0, __m256& _r1, __m256& _r2)
-{
-    __m256 _tmp0 = _mm256_permute2f128_ps(_r0, _r1, _MM_SHUFFLE(0, 3, 0, 0));
-    __m256 _tmp1 = _mm256_permute2f128_ps(_r0, _r2, _MM_SHUFFLE(0, 2, 0, 1));
-    __m256 _tmp2 = _mm256_permute2f128_ps(_r1, _r2, _MM_SHUFFLE(0, 3, 0, 0));
-
-    __m256 _tmp4 = _mm256_shuffle_ps(_tmp0, _tmp1, _MM_SHUFFLE(1, 0, 2, 1));
-    __m256 _tmp5 = _mm256_shuffle_ps(_tmp1, _tmp2, _MM_SHUFFLE(2, 1, 3, 2));
-
-    _r0 = _mm256_shuffle_ps(_tmp0, _tmp5, _MM_SHUFFLE(2, 0, 3, 0));
-    _r1 = _mm256_shuffle_ps(_tmp4, _tmp5, _MM_SHUFFLE(3, 1, 2, 0));
-    _r2 = _mm256_shuffle_ps(_tmp4, _tmp2, _MM_SHUFFLE(3, 0, 3, 1));
-}
-
-static NCNN_FORCEINLINE void transpose8x6_ps(__m256& _r0, __m256& _r1, __m256& _r2, __m256& _r3, __m256& _r4, __m256& _r5)
-{
-    __m256 _tmp0 = _mm256_unpacklo_ps(_r0, _r1);
-    __m256 _tmp1 = _mm256_unpackhi_ps(_r0, _r1);
-    __m256 _tmp2 = _mm256_unpacklo_ps(_r2, _r3);
-    __m256 _tmp3 = _mm256_unpackhi_ps(_r2, _r3);
-    __m256 _tmp4 = _mm256_unpacklo_ps(_r4, _r5);
-    __m256 _tmp5 = _mm256_unpackhi_ps(_r4, _r5);
-
-    __m256 _tmp6 = _mm256_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmp7 = _mm256_shuffle_ps(_tmp4, _tmp0, _MM_SHUFFLE(3, 2, 1, 0));
-    __m256 _tmp8 = _mm256_shuffle_ps(_tmp2, _tmp4, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmp9 = _mm256_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpa = _mm256_shuffle_ps(_tmp5, _tmp1, _MM_SHUFFLE(3, 2, 1, 0));
-    __m256 _tmpb = _mm256_shuffle_ps(_tmp3, _tmp5, _MM_SHUFFLE(3, 2, 3, 2));
-
-    _r0 = _mm256_permute2f128_ps(_tmp6, _tmp7, _MM_SHUFFLE(0, 2, 0, 0));
-    _r1 = _mm256_permute2f128_ps(_tmp8, _tmp9, _MM_SHUFFLE(0, 2, 0, 0));
-    _r2 = _mm256_permute2f128_ps(_tmpa, _tmpb, _MM_SHUFFLE(0, 2, 0, 0));
-    _r3 = _mm256_permute2f128_ps(_tmp6, _tmp7, _MM_SHUFFLE(0, 3, 0, 1));
-    _r4 = _mm256_permute2f128_ps(_tmp8, _tmp9, _MM_SHUFFLE(0, 3, 0, 1));
-    _r5 = _mm256_permute2f128_ps(_tmpa, _tmpb, _MM_SHUFFLE(0, 3, 0, 1));
-}
-
-static NCNN_FORCEINLINE void transpose8x11_ps(__m256& _r0, __m256& _r1, __m256& _r2, __m256& _r3, __m256& _r4, __m256& _r5, __m256& _r6, __m256& _r7, __m256& _r8, __m256& _r9, __m256& _ra)
-{
-    __m256 _tmp0 = _mm256_unpacklo_ps(_r0, _r1);
-    __m256 _tmp1 = _mm256_unpackhi_ps(_r0, _r1);
-    __m256 _tmp2 = _mm256_unpacklo_ps(_r2, _r3);
-    __m256 _tmp3 = _mm256_unpackhi_ps(_r2, _r3);
-    __m256 _tmp4 = _mm256_unpacklo_ps(_r4, _r5);
-    __m256 _tmp5 = _mm256_unpackhi_ps(_r4, _r5);
-    __m256 _tmp6 = _mm256_unpacklo_ps(_r6, _r7);
-    __m256 _tmp7 = _mm256_unpackhi_ps(_r6, _r7);
-    __m256 _tmp8 = _mm256_unpacklo_ps(_r8, _r9);
-    __m256 _tmp9 = _mm256_unpackhi_ps(_r8, _r9);
-    __m256 _tmpa = _mm256_unpacklo_ps(_ra, _r0);
-    __m256 _tmpb = _mm256_shuffle_ps(_ra, _tmp1, _MM_SHUFFLE(3, 2, 1, 2));
-    __m256 _tmpc = _mm256_unpacklo_ps(_r1, _r2);
-    __m256 _tmpd = _mm256_unpackhi_ps(_r1, _r2);
-    __m256 _tmpe = _mm256_unpacklo_ps(_r3, _r4);
-    __m256 _tmpf = _mm256_unpackhi_ps(_r3, _r4);
-    __m256 _tmpg = _mm256_unpacklo_ps(_r5, _r6);
-    __m256 _tmph = _mm256_unpackhi_ps(_r5, _r6);
-    __m256 _tmpi = _mm256_unpacklo_ps(_r7, _r8);
-    __m256 _tmpj = _mm256_unpackhi_ps(_r7, _r8);
-    __m256 _tmpk = _mm256_unpacklo_ps(_r9, _ra);
-    __m256 _tmpl = _mm256_unpackhi_ps(_r9, _ra);
-
-    __m256 _tmpm = _mm256_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpn = _mm256_shuffle_ps(_tmp4, _tmp6, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpo = _mm256_shuffle_ps(_tmp8, _tmpa, _MM_SHUFFLE(3, 0, 1, 0));
-    __m256 _tmpp = _mm256_shuffle_ps(_tmpc, _tmpe, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpq = _mm256_shuffle_ps(_tmpg, _tmpi, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpr = _mm256_shuffle_ps(_tmpk, _tmp1, _MM_SHUFFLE(1, 0, 3, 2));
-    __m256 _tmps = _mm256_shuffle_ps(_tmp3, _tmp5, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpt = _mm256_shuffle_ps(_tmp7, _tmp9, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpu = _mm256_shuffle_ps(_tmpb, _tmpd, _MM_SHUFFLE(3, 2, 2, 0));
-    __m256 _tmpv = _mm256_shuffle_ps(_tmpf, _tmph, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpw = _mm256_shuffle_ps(_tmpj, _tmpl, _MM_SHUFFLE(3, 2, 3, 2));
-
-    _r0 = _mm256_permute2f128_ps(_tmpm, _tmpn, _MM_SHUFFLE(0, 2, 0, 0));
-    _r1 = _mm256_permute2f128_ps(_tmpo, _tmpp, _MM_SHUFFLE(0, 2, 0, 0));
-    _r2 = _mm256_permute2f128_ps(_tmpq, _tmpr, _MM_SHUFFLE(0, 2, 0, 0));
-    _r3 = _mm256_permute2f128_ps(_tmps, _tmpt, _MM_SHUFFLE(0, 2, 0, 0));
-    _r4 = _mm256_permute2f128_ps(_tmpu, _tmpv, _MM_SHUFFLE(0, 2, 0, 0));
-    _r5 = _mm256_permute2f128_ps(_tmpw, _tmpm, _MM_SHUFFLE(0, 3, 0, 0));
-    _r6 = _mm256_permute2f128_ps(_tmpn, _tmpo, _MM_SHUFFLE(0, 3, 0, 1));
-    _r7 = _mm256_permute2f128_ps(_tmpp, _tmpq, _MM_SHUFFLE(0, 3, 0, 1));
-    _r8 = _mm256_permute2f128_ps(_tmpr, _tmps, _MM_SHUFFLE(0, 3, 0, 1));
-    _r9 = _mm256_permute2f128_ps(_tmpt, _tmpu, _MM_SHUFFLE(0, 3, 0, 1));
-    _ra = _mm256_permute2f128_ps(_tmpv, _tmpw, _MM_SHUFFLE(0, 3, 0, 1));
-}
-
-static void transpose8x18_ps(__m256& _r0, __m256& _r1, __m256& _r2, __m256& _r3, __m256& _r4, __m256& _r5, __m256& _r6, __m256& _r7, __m256& _r8, __m256& _r9, __m256& _ra, __m256& _rb, __m256& _rc, __m256& _rd, __m256& _re, __m256& _rf, __m256& _rg, __m256& _rh)
-{
-    __m256 _tmp0 = _mm256_unpacklo_ps(_r0, _r1);
-    __m256 _tmp1 = _mm256_unpackhi_ps(_r0, _r1);
-    __m256 _tmp2 = _mm256_unpacklo_ps(_r2, _r3);
-    __m256 _tmp3 = _mm256_unpackhi_ps(_r2, _r3);
-    __m256 _tmp4 = _mm256_unpacklo_ps(_r4, _r5);
-    __m256 _tmp5 = _mm256_unpackhi_ps(_r4, _r5);
-    __m256 _tmp6 = _mm256_unpacklo_ps(_r6, _r7);
-    __m256 _tmp7 = _mm256_unpackhi_ps(_r6, _r7);
-    __m256 _tmp8 = _mm256_unpacklo_ps(_r8, _r9);
-    __m256 _tmp9 = _mm256_unpackhi_ps(_r8, _r9);
-    __m256 _tmpa = _mm256_unpacklo_ps(_ra, _rb);
-    __m256 _tmpb = _mm256_unpackhi_ps(_ra, _rb);
-    __m256 _tmpc = _mm256_unpacklo_ps(_rc, _rd);
-    __m256 _tmpd = _mm256_unpackhi_ps(_rc, _rd);
-    __m256 _tmpe = _mm256_unpacklo_ps(_re, _rf);
-    __m256 _tmpf = _mm256_unpackhi_ps(_re, _rf);
-    __m256 _tmpg = _mm256_unpacklo_ps(_rg, _rh);
-    __m256 _tmph = _mm256_unpackhi_ps(_rg, _rh);
-
-    __m256 _tmpi = _mm256_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpj = _mm256_shuffle_ps(_tmp4, _tmp6, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpk = _mm256_shuffle_ps(_tmp8, _tmpa, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpl = _mm256_shuffle_ps(_tmpc, _tmpe, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpm = _mm256_shuffle_ps(_tmpg, _tmp0, _MM_SHUFFLE(3, 2, 1, 0));
-    __m256 _tmpn = _mm256_shuffle_ps(_tmp2, _tmp4, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpo = _mm256_shuffle_ps(_tmp6, _tmp8, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpp = _mm256_shuffle_ps(_tmpa, _tmpc, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpq = _mm256_shuffle_ps(_tmpe, _tmpg, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpr = _mm256_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmps = _mm256_shuffle_ps(_tmp5, _tmp7, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpt = _mm256_shuffle_ps(_tmp9, _tmpb, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpu = _mm256_shuffle_ps(_tmpd, _tmpf, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpv = _mm256_shuffle_ps(_tmph, _tmp1, _MM_SHUFFLE(3, 2, 1, 0));
-    __m256 _tmpw = _mm256_shuffle_ps(_tmp3, _tmp5, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpx = _mm256_shuffle_ps(_tmp7, _tmp9, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpy = _mm256_shuffle_ps(_tmpb, _tmpd, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpz = _mm256_shuffle_ps(_tmpf, _tmph, _MM_SHUFFLE(3, 2, 3, 2));
-
-    _r0 = _mm256_permute2f128_ps(_tmpi, _tmpj, _MM_SHUFFLE(0, 2, 0, 0));
-    _r1 = _mm256_permute2f128_ps(_tmpk, _tmpl, _MM_SHUFFLE(0, 2, 0, 0));
-    _r2 = _mm256_permute2f128_ps(_tmpm, _tmpn, _MM_SHUFFLE(0, 2, 0, 0));
-    _r3 = _mm256_permute2f128_ps(_tmpo, _tmpp, _MM_SHUFFLE(0, 2, 0, 0));
-    _r4 = _mm256_permute2f128_ps(_tmpq, _tmpr, _MM_SHUFFLE(0, 2, 0, 0));
-    _r5 = _mm256_permute2f128_ps(_tmps, _tmpt, _MM_SHUFFLE(0, 2, 0, 0));
-    _r6 = _mm256_permute2f128_ps(_tmpu, _tmpv, _MM_SHUFFLE(0, 2, 0, 0));
-    _r7 = _mm256_permute2f128_ps(_tmpw, _tmpx, _MM_SHUFFLE(0, 2, 0, 0));
-    _r8 = _mm256_permute2f128_ps(_tmpy, _tmpz, _MM_SHUFFLE(0, 2, 0, 0));
-    _r9 = _mm256_permute2f128_ps(_tmpi, _tmpj, _MM_SHUFFLE(0, 3, 0, 1));
-    _ra = _mm256_permute2f128_ps(_tmpk, _tmpl, _MM_SHUFFLE(0, 3, 0, 1));
-    _rb = _mm256_permute2f128_ps(_tmpm, _tmpn, _MM_SHUFFLE(0, 3, 0, 1));
-    _rc = _mm256_permute2f128_ps(_tmpo, _tmpp, _MM_SHUFFLE(0, 3, 0, 1));
-    _rd = _mm256_permute2f128_ps(_tmpq, _tmpr, _MM_SHUFFLE(0, 3, 0, 1));
-    _re = _mm256_permute2f128_ps(_tmps, _tmpt, _MM_SHUFFLE(0, 3, 0, 1));
-    _rf = _mm256_permute2f128_ps(_tmpu, _tmpv, _MM_SHUFFLE(0, 3, 0, 1));
-    _rg = _mm256_permute2f128_ps(_tmpw, _tmpx, _MM_SHUFFLE(0, 3, 0, 1));
-    _rh = _mm256_permute2f128_ps(_tmpy, _tmpz, _MM_SHUFFLE(0, 3, 0, 1));
-}
-
-static NCNN_FORCEINLINE __m256 HorizontalSums(__m256& v0, __m256& v1, __m256& v2, __m256& v3, __m256& v4, __m256& v5, __m256& v6, __m256& v7)
-{
-    const __m256 s01 = _mm256_hadd_ps(v0, v1);
-    const __m256 s23 = _mm256_hadd_ps(v2, v3);
-    const __m256 s45 = _mm256_hadd_ps(v4, v5);
-    const __m256 s67 = _mm256_hadd_ps(v6, v7);
-    const __m256 s0123 = _mm256_hadd_ps(s01, s23);
-    const __m256 s4556 = _mm256_hadd_ps(s45, s67);
-
-    // inter-lane shuffle
-    const __m256 vb0 = _mm256_blend_ps(s0123, s4556, 0xF0);
-    const __m256 vb1 = _mm256_permute2f128_ps(s0123, s4556, 0x21);
-
-    return _mm256_add_ps(vb0, vb1);
-}
-
-static NCNN_FORCEINLINE __m128 HorizontalSums(__m256& v0, __m256& v1, __m256& v2, __m256& v3)
-{
-    const __m256 s01 = _mm256_hadd_ps(v0, v1);
-    const __m256 s23 = _mm256_hadd_ps(v2, v3);
-    const __m256 s0123 = _mm256_hadd_ps(s01, s23);
-
-    return _mm_add_ps(_mm256_extractf128_ps(s0123, 1),
-                      _mm256_castps256_ps128(s0123));
-}
-
-static NCNN_FORCEINLINE __m128 HorizontalSums(__m256& v0, __m256& v1, __m256& v2)
-{
-    const __m256 v3 = _mm256_set1_ps(0.0f);
-    const __m256 s01 = _mm256_hadd_ps(v0, v1);
-    const __m256 s23 = _mm256_hadd_ps(v2, v3);
-    const __m256 s0123 = _mm256_hadd_ps(s01, s23);
-
-    return _mm_add_ps(_mm256_extractf128_ps(s0123, 1),
-                      _mm256_castps256_ps128(s0123));
-}
-
-static NCNN_FORCEINLINE float _mm256_reduce_add_ps(__m256 x)
-{
-    /* ( x3+x7, x2+x6, x1+x5, x0+x4 ) */
-    const __m128 x128 = _mm_add_ps(_mm256_extractf128_ps(x, 1), _mm256_castps256_ps128(x));
-    /* ( -, -, x1+x3+x5+x7, x0+x2+x4+x6 ) */
-    const __m128 x64 = _mm_add_ps(x128, _mm_movehl_ps(x128, x128));
-    /* ( -, -, -, x0+x1+x2+x3+x4+x5+x6+x7 ) */
-    const __m128 x32 = _mm_add_ss(x64, _mm_shuffle_ps(x64, x64, 0x55));
-    /* Conversion to float is a no-op on x86-64 */
-    return _mm_cvtss_f32(x32);
-}
-
-static NCNN_FORCEINLINE float _mm256_reduce_max_ps(__m256 x)
-{
-    const __m128 x128 = _mm_max_ps(_mm256_extractf128_ps(x, 1), _mm256_castps256_ps128(x));
-    const __m128 x64 = _mm_max_ps(x128, _mm_movehl_ps(x128, x128));
-    const __m128 x32 = _mm_max_ss(x64, _mm_shuffle_ps(x64, x64, 0x55));
-    return _mm_cvtss_f32(x32);
-}
-
-static NCNN_FORCEINLINE int64_t float2int8_avx(const __m256& _v0)
-{
-    // _MM_FROUND_TO_NEAREST_INT round to even
-    // simulate round to nearest via +/-0.5 with round to zero
-    __m256 _p5 = _mm256_set1_ps(0.5f);
-    __m256 _signmask = _mm256_castsi256_ps(_mm256_set1_epi32(1 << 31));
-    __m256 _sign = _mm256_and_ps(_v0, _signmask);
-    __m256 _v0_p5 = _mm256_or_ps(_p5, _sign);
-    __m256 _v0_adj = _mm256_add_ps(_v0, _v0_p5);
-    __m256i _v0_i = _mm256_cvttps_epi32(_v0_adj);
-
-#if __AVX2__
-    __m256i _v01_s16 = _mm256_packs_epi32(_v0_i, _v0_i);
-    _v01_s16 = _mm256_permute4x64_epi64(_v01_s16, 0xd8);
-
-    __m128i _v01_s16low = _mm256_extractf128_si256(_v01_s16, 0);
-#else  // __AVX2__
-    __m128i _v0_i_low = _mm256_extractf128_si256(_v0_i, 0);
-    __m128i _v0_i_high = _mm256_extractf128_si256(_v0_i, 1);
-
-    __m128i _v01_s16low = _mm_packs_epi32(_v0_i_low, _v0_i_high);
-#endif // __AVX2__
-
-    _v01_s16low = _mm_min_epi16(_v01_s16low, _mm_set1_epi16(127));
-    _v01_s16low = _mm_max_epi16(_v01_s16low, _mm_set1_epi16(-127));
-
-    __m128i _v8 = _mm_packs_epi16(_v01_s16low, _v01_s16low);
-
-#if defined(__x86_64__) || defined(_M_X64)
-    return _mm_cvtsi128_si64(_v8);
-#else
-    int64_t v8[2];
-    _mm_storeu_si128((__m128i*)v8, _v8);
-    return v8[0];
-#endif
-}
-
-static NCNN_FORCEINLINE __m128i float2int8_avx(const __m256& _v0, const __m256& _v1)
-{
-    // _MM_FROUND_TO_NEAREST_INT round to even
-    // simulate round to nearest via +/-0.5 with round to zero
-    __m256 _p5 = _mm256_set1_ps(0.5f);
-    __m256 _signmask = _mm256_castsi256_ps(_mm256_set1_epi32(1 << 31));
-    __m256 _sign0 = _mm256_and_ps(_v0, _signmask);
-    __m256 _sign1 = _mm256_and_ps(_v1, _signmask);
-    __m256 _v0_p5 = _mm256_or_ps(_p5, _sign0);
-    __m256 _v1_p5 = _mm256_or_ps(_p5, _sign1);
-    __m256 _v0_adj = _mm256_add_ps(_v0, _v0_p5);
-    __m256 _v1_adj = _mm256_add_ps(_v1, _v1_p5);
-    __m256i _v0_i = _mm256_cvttps_epi32(_v0_adj);
-    __m256i _v1_i = _mm256_cvttps_epi32(_v1_adj);
-
-#if __AVX2__
-    __m256i _v01_s16 = _mm256_packs_epi32(_v0_i, _v1_i);
-    _v01_s16 = _mm256_permute4x64_epi64(_v01_s16, 0xd8);
-
-    _v01_s16 = _mm256_min_epi16(_v01_s16, _mm256_set1_epi16(127));
-    _v01_s16 = _mm256_max_epi16(_v01_s16, _mm256_set1_epi16(-127));
-
-    __m256i _v8 = _mm256_packs_epi16(_v01_s16, _v01_s16);
-    _v8 = _mm256_permute4x64_epi64(_v8, 0xd8);
-
-    return _mm256_extractf128_si256(_v8, 0);
-#else  // __AVX2__
-    __m128i _v0_i_low = _mm256_extractf128_si256(_v0_i, 0);
-    __m128i _v0_i_high = _mm256_extractf128_si256(_v0_i, 1);
-    __m128i _v1_i_low = _mm256_extractf128_si256(_v1_i, 0);
-    __m128i _v1_i_high = _mm256_extractf128_si256(_v1_i, 1);
-
-    __m128i _v01_s16low = _mm_packs_epi32(_v0_i_low, _v0_i_high);
-    __m128i _v01_s16high = _mm_packs_epi32(_v1_i_low, _v1_i_high);
-
-    _v01_s16low = _mm_min_epi16(_v01_s16low, _mm_set1_epi16(127));
-    _v01_s16high = _mm_min_epi16(_v01_s16high, _mm_set1_epi16(127));
-    _v01_s16low = _mm_max_epi16(_v01_s16low, _mm_set1_epi16(-127));
-    _v01_s16high = _mm_max_epi16(_v01_s16high, _mm_set1_epi16(-127));
-
-    __m128i _v8 = _mm_packs_epi16(_v01_s16low, _v01_s16high);
-    return _v8;
-#endif // __AVX2__
-}
-
-static NCNN_FORCEINLINE void _mm256_comp_fmadd_ps4(__m256& _sum,
-        const __m256& _w0, const __m256& _w1, const __m256& _w2, const __m256& _w3,
-        const __m256& _v0, const __m256& _v1, const __m256& _v2, const __m256& _v3)
-{
-    __m256 _mul0 = _mm256_mul_ps(_w0, _v0);
-    __m256 _mul1 = _mm256_mul_ps(_w1, _v1);
-    __m256 _sum01 = _mm256_add_ps(_mul0, _mul1);
-    __m256 _mul2 = _mm256_mul_ps(_w2, _v2);
-    __m256 _mul3 = _mm256_mul_ps(_w3, _v3);
-    __m256 _sum23 = _mm256_add_ps(_mul2, _mul3);
-    __m256 _sum0123 = _mm256_add_ps(_sum01, _sum23);
-    _sum = _mm256_add_ps(_sum, _sum0123);
-}
-
-static NCNN_FORCEINLINE void _mm256_comp_fmadd_ps8(__m256& _sum,
-        const __m256& _w0, const __m256& _w1, const __m256& _w2, const __m256& _w3, const __m256& _w4, const __m256& _w5, const __m256& _w6, const __m256& _w7,
-        const __m256& _v0, const __m256& _v1, const __m256& _v2, const __m256& _v3, const __m256& _v4, const __m256& _v5, const __m256& _v6, const __m256& _v7)
-{
-    _mm256_comp_fmadd_ps4(_sum, _w0, _w1, _w2, _w3, _v0, _v1, _v2, _v3);
-
-    _mm256_comp_fmadd_ps4(_sum, _w4, _w5, _w6, _w7, _v4, _v5, _v6, _v7);
-}
-
-static NCNN_FORCEINLINE __m256 bfloat2float_avx(const __m128i& v0)
-{
-#if __AVX512BF16__
-    __m256 _v = _mm256_cvtpbh_ps((__m128bh)v0);
-#else
-    __m128i _zero = _mm_setzero_si128();
-    __m128i _a = _mm_unpacklo_epi16(_zero, v0);
-    __m128i _b = _mm_unpackhi_epi16(_zero, v0);
-    __m256 _v = _mm256_castsi256_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(_a), _b, 1));
-#endif
-    return _v;
-}
-
-static NCNN_FORCEINLINE __m128i float2bfloat_avx(const __m256& v0)
-{
-#if __AVX512BF16__
-    __m128i _v = (__m128i)_mm256_cvtneps_pbh(v0);
-#else
-    __m256i _ab = _mm256_castps_si256(v0);
-#if __AVX2__
-    _ab = _mm256_srli_epi32(_ab, 16);
-    __m128i _a = _mm256_extractf128_si256(_ab, 0);
-    __m128i _b = _mm256_extractf128_si256(_ab, 1);
-#else
-    __m128i _a = _mm256_extractf128_si256(_ab, 0);
-    __m128i _b = _mm256_extractf128_si256(_ab, 1);
-    _a = _mm_srli_epi32(_a, 16);
-    _b = _mm_srli_epi32(_b, 16);
-#endif
-    __m128i _v = _mm_packus_epi32(_a, _b);
-#endif
-    return _v;
-}
-
-static NCNN_FORCEINLINE __m256i float2bfloat_avx(const __m256& v0, const __m256& v1)
-{
-#if __AVX512BF16__
-    __m128i _v0 = (__m128i)_mm256_cvtneps_pbh(v0);
-    __m128i _v1 = (__m128i)_mm256_cvtneps_pbh(v1);
-    __m256i _v = _mm256_insertf128_si256(_mm256_castsi128_si256(_v0), _v1, 1);
-#else
-    __m256i _a = _mm256_castps_si256(v0);
-    __m256i _b = _mm256_castps_si256(v1);
-#if __AVX2__
-    _a = _mm256_srli_epi32(_a, 16);
-    _b = _mm256_srli_epi32(_b, 16);
-    __m256i _v = _mm256_packus_epi32(_a, _b);
-    _v = _mm256_permute4x64_epi64(_v, _MM_SHUFFLE(3, 1, 2, 0));
-#else
-    __m128i _a0 = _mm256_extractf128_si256(_a, 0);
-    __m128i _a1 = _mm256_extractf128_si256(_a, 1);
-    __m128i _b0 = _mm256_extractf128_si256(_b, 0);
-    __m128i _b1 = _mm256_extractf128_si256(_b, 1);
-    _a0 = _mm_srli_epi32(_a0, 16);
-    _a1 = _mm_srli_epi32(_a1, 16);
-    _b0 = _mm_srli_epi32(_b0, 16);
-    _b1 = _mm_srli_epi32(_b1, 16);
-    __m128i _v0 = _mm_packus_epi32(_a0, _a1);
-    __m128i _v1 = _mm_packus_epi32(_b0, _b1);
-    __m256i _v = _mm256_insertf128_si256(_mm256_castsi128_si256(_v0), _v1, 1);
-#endif
-#endif
-    return _v;
-}
-
-#if __AVX2__
-static NCNN_FORCEINLINE void transpose8x2_epi32(__m256i& _r0, __m256i& _r1)
-{
-    __m256i _tmp0 = _mm256_unpacklo_epi32(_r0, _r1);
-    __m256i _tmp1 = _mm256_unpackhi_epi32(_r0, _r1);
-
-    _r0 = _mm256_permute2x128_si256(_tmp0, _tmp1, _MM_SHUFFLE(0, 2, 0, 0));
-    _r1 = _mm256_permute2x128_si256(_tmp0, _tmp1, _MM_SHUFFLE(0, 3, 0, 1));
-}
-
-static NCNN_FORCEINLINE void transpose16x8_epi16(__m256i& _r0, __m256i& _r1, __m256i& _r2, __m256i& _r3, __m256i& _r4, __m256i& _r5, __m256i& _r6, __m256i& _r7)
-{
-    __m256i _tmp0 = _mm256_unpacklo_epi16(_r0, _r1);
-    __m256i _tmp1 = _mm256_unpackhi_epi16(_r0, _r1);
-    __m256i _tmp2 = _mm256_unpacklo_epi16(_r2, _r3);
-    __m256i _tmp3 = _mm256_unpackhi_epi16(_r2, _r3);
-    __m256i _tmp4 = _mm256_unpacklo_epi16(_r4, _r5);
-    __m256i _tmp5 = _mm256_unpackhi_epi16(_r4, _r5);
-    __m256i _tmp6 = _mm256_unpacklo_epi16(_r6, _r7);
-    __m256i _tmp7 = _mm256_unpackhi_epi16(_r6, _r7);
-
-    __m256i _tmpg = _mm256_unpacklo_epi32(_tmp0, _tmp2);
-    __m256i _tmph = _mm256_unpackhi_epi32(_tmp0, _tmp2);
-    __m256i _tmpi = _mm256_unpacklo_epi32(_tmp1, _tmp3);
-    __m256i _tmpj = _mm256_unpackhi_epi32(_tmp1, _tmp3);
-    __m256i _tmpk = _mm256_unpacklo_epi32(_tmp4, _tmp6);
-    __m256i _tmpl = _mm256_unpackhi_epi32(_tmp4, _tmp6);
-    __m256i _tmpm = _mm256_unpacklo_epi32(_tmp5, _tmp7);
-    __m256i _tmpn = _mm256_unpackhi_epi32(_tmp5, _tmp7);
-
-    _tmp0 = _mm256_unpacklo_epi64(_tmpg, _tmpk);
-    _tmp1 = _mm256_unpackhi_epi64(_tmpg, _tmpk);
-    _tmp2 = _mm256_unpacklo_epi64(_tmph, _tmpl);
-    _tmp3 = _mm256_unpackhi_epi64(_tmph, _tmpl);
-    _tmp4 = _mm256_unpacklo_epi64(_tmpi, _tmpm);
-    _tmp5 = _mm256_unpackhi_epi64(_tmpi, _tmpm);
-    _tmp6 = _mm256_unpacklo_epi64(_tmpj, _tmpn);
-    _tmp7 = _mm256_unpackhi_epi64(_tmpj, _tmpn);
-
-    _r0 = _mm256_permute2x128_si256(_tmp0, _tmp1, _MM_SHUFFLE(0, 2, 0, 0));
-    _r1 = _mm256_permute2x128_si256(_tmp2, _tmp3, _MM_SHUFFLE(0, 2, 0, 0));
-    _r2 = _mm256_permute2x128_si256(_tmp4, _tmp5, _MM_SHUFFLE(0, 2, 0, 0));
-    _r3 = _mm256_permute2x128_si256(_tmp6, _tmp7, _MM_SHUFFLE(0, 2, 0, 0));
-    _r4 = _mm256_permute2x128_si256(_tmp0, _tmp1, _MM_SHUFFLE(0, 3, 0, 1));
-    _r5 = _mm256_permute2x128_si256(_tmp2, _tmp3, _MM_SHUFFLE(0, 3, 0, 1));
-    _r6 = _mm256_permute2x128_si256(_tmp4, _tmp5, _MM_SHUFFLE(0, 3, 0, 1));
-    _r7 = _mm256_permute2x128_si256(_tmp6, _tmp7, _MM_SHUFFLE(0, 3, 0, 1));
-}
-
-#if __AVX512F__
-static NCNN_FORCEINLINE void transpose16x16_ps(__m512& _r0, __m512& _r1, __m512& _r2, __m512& _r3, __m512& _r4, __m512& _r5, __m512& _r6, __m512& _r7,
-        __m512& _r8, __m512& _r9, __m512& _ra, __m512& _rb, __m512& _rc, __m512& _rd, __m512& _re, __m512& _rf)
-{
-    __m512 _tmp0 = _mm512_unpacklo_ps(_r0, _r1);
-    __m512 _tmp1 = _mm512_unpackhi_ps(_r0, _r1);
-    __m512 _tmp2 = _mm512_unpacklo_ps(_r2, _r3);
-    __m512 _tmp3 = _mm512_unpackhi_ps(_r2, _r3);
-    __m512 _tmp4 = _mm512_unpacklo_ps(_r4, _r5);
-    __m512 _tmp5 = _mm512_unpackhi_ps(_r4, _r5);
-    __m512 _tmp6 = _mm512_unpacklo_ps(_r6, _r7);
-    __m512 _tmp7 = _mm512_unpackhi_ps(_r6, _r7);
-    __m512 _tmp8 = _mm512_unpacklo_ps(_r8, _r9);
-    __m512 _tmp9 = _mm512_unpackhi_ps(_r8, _r9);
-    __m512 _tmpa = _mm512_unpacklo_ps(_ra, _rb);
-    __m512 _tmpb = _mm512_unpackhi_ps(_ra, _rb);
-    __m512 _tmpc = _mm512_unpacklo_ps(_rc, _rd);
-    __m512 _tmpd = _mm512_unpackhi_ps(_rc, _rd);
-    __m512 _tmpe = _mm512_unpacklo_ps(_re, _rf);
-    __m512 _tmpf = _mm512_unpackhi_ps(_re, _rf);
-
-    __m512 _tmpg = _mm512_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmph = _mm512_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmpi = _mm512_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmpj = _mm512_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmpk = _mm512_shuffle_ps(_tmp4, _tmp6, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmpl = _mm512_shuffle_ps(_tmp4, _tmp6, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmpm = _mm512_shuffle_ps(_tmp5, _tmp7, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmpn = _mm512_shuffle_ps(_tmp5, _tmp7, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmpo = _mm512_shuffle_ps(_tmp8, _tmpa, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmpp = _mm512_shuffle_ps(_tmp8, _tmpa, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmpq = _mm512_shuffle_ps(_tmp9, _tmpb, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmpr = _mm512_shuffle_ps(_tmp9, _tmpb, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmps = _mm512_shuffle_ps(_tmpc, _tmpe, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmpt = _mm512_shuffle_ps(_tmpc, _tmpe, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmpu = _mm512_shuffle_ps(_tmpd, _tmpf, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmpv = _mm512_shuffle_ps(_tmpd, _tmpf, _MM_SHUFFLE(3, 2, 3, 2));
-
-    _tmp0 = _mm512_shuffle_f32x4(_tmpg, _tmpk, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp1 = _mm512_shuffle_f32x4(_tmpo, _tmps, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp2 = _mm512_shuffle_f32x4(_tmph, _tmpl, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp3 = _mm512_shuffle_f32x4(_tmpp, _tmpt, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp4 = _mm512_shuffle_f32x4(_tmpi, _tmpm, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp5 = _mm512_shuffle_f32x4(_tmpq, _tmpu, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp6 = _mm512_shuffle_f32x4(_tmpj, _tmpn, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp7 = _mm512_shuffle_f32x4(_tmpr, _tmpv, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp8 = _mm512_shuffle_f32x4(_tmpg, _tmpk, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmp9 = _mm512_shuffle_f32x4(_tmpo, _tmps, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmpa = _mm512_shuffle_f32x4(_tmph, _tmpl, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmpb = _mm512_shuffle_f32x4(_tmpp, _tmpt, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmpc = _mm512_shuffle_f32x4(_tmpi, _tmpm, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmpd = _mm512_shuffle_f32x4(_tmpq, _tmpu, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmpe = _mm512_shuffle_f32x4(_tmpj, _tmpn, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmpf = _mm512_shuffle_f32x4(_tmpr, _tmpv, _MM_SHUFFLE(3, 1, 3, 1));
-
-    _r0 = _mm512_shuffle_f32x4(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-    _r1 = _mm512_shuffle_f32x4(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-    _r2 = _mm512_shuffle_f32x4(_tmp4, _tmp5, _MM_SHUFFLE(2, 0, 2, 0));
-    _r3 = _mm512_shuffle_f32x4(_tmp6, _tmp7, _MM_SHUFFLE(2, 0, 2, 0));
-    _r4 = _mm512_shuffle_f32x4(_tmp8, _tmp9, _MM_SHUFFLE(2, 0, 2, 0));
-    _r5 = _mm512_shuffle_f32x4(_tmpa, _tmpb, _MM_SHUFFLE(2, 0, 2, 0));
-    _r6 = _mm512_shuffle_f32x4(_tmpc, _tmpd, _MM_SHUFFLE(2, 0, 2, 0));
-    _r7 = _mm512_shuffle_f32x4(_tmpe, _tmpf, _MM_SHUFFLE(2, 0, 2, 0));
-    _r8 = _mm512_shuffle_f32x4(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-    _r9 = _mm512_shuffle_f32x4(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-    _ra = _mm512_shuffle_f32x4(_tmp4, _tmp5, _MM_SHUFFLE(3, 1, 3, 1));
-    _rb = _mm512_shuffle_f32x4(_tmp6, _tmp7, _MM_SHUFFLE(3, 1, 3, 1));
-    _rc = _mm512_shuffle_f32x4(_tmp8, _tmp9, _MM_SHUFFLE(3, 1, 3, 1));
-    _rd = _mm512_shuffle_f32x4(_tmpa, _tmpb, _MM_SHUFFLE(3, 1, 3, 1));
-    _re = _mm512_shuffle_f32x4(_tmpc, _tmpd, _MM_SHUFFLE(3, 1, 3, 1));
-    _rf = _mm512_shuffle_f32x4(_tmpe, _tmpf, _MM_SHUFFLE(3, 1, 3, 1));
-}
-
-static NCNN_FORCEINLINE void transpose16x12_ps(__m512& _r0, __m512& _r1, __m512& _r2, __m512& _r3, __m512& _r4, __m512& _r5, __m512& _r6, __m512& _r7,
-        __m512& _r8, __m512& _r9, __m512& _ra, __m512& _rb)
-{
-    __m512 _tmp0 = _mm512_unpacklo_ps(_r0, _r1);
-    __m512 _tmp1 = _mm512_unpackhi_ps(_r0, _r1);
-    __m512 _tmp2 = _mm512_unpacklo_ps(_r2, _r3);
-    __m512 _tmp3 = _mm512_unpackhi_ps(_r2, _r3);
-    __m512 _tmp4 = _mm512_unpacklo_ps(_r4, _r5);
-    __m512 _tmp5 = _mm512_unpackhi_ps(_r4, _r5);
-    __m512 _tmp6 = _mm512_unpacklo_ps(_r6, _r7);
-    __m512 _tmp7 = _mm512_unpackhi_ps(_r6, _r7);
-    __m512 _tmp8 = _mm512_unpacklo_ps(_r8, _r9);
-    __m512 _tmp9 = _mm512_unpackhi_ps(_r8, _r9);
-    __m512 _tmpa = _mm512_unpacklo_ps(_ra, _rb);
-    __m512 _tmpb = _mm512_unpackhi_ps(_ra, _rb);
-
-    __m512 _tmpc = _mm512_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmpd = _mm512_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmpe = _mm512_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmpf = _mm512_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmpg = _mm512_shuffle_ps(_tmp4, _tmp6, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmph = _mm512_shuffle_ps(_tmp4, _tmp6, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmpi = _mm512_shuffle_ps(_tmp5, _tmp7, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmpj = _mm512_shuffle_ps(_tmp5, _tmp7, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmpk = _mm512_shuffle_ps(_tmp8, _tmpa, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmpl = _mm512_shuffle_ps(_tmp8, _tmpa, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmpm = _mm512_shuffle_ps(_tmp9, _tmpb, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmpn = _mm512_shuffle_ps(_tmp9, _tmpb, _MM_SHUFFLE(3, 2, 3, 2));
-
-    _tmp0 = _mm512_shuffle_f32x4(_tmpc, _tmpg, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp1 = _mm512_shuffle_f32x4(_tmpk, _tmpd, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp2 = _mm512_shuffle_f32x4(_tmph, _tmpl, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp3 = _mm512_shuffle_f32x4(_tmpe, _tmpi, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp4 = _mm512_shuffle_f32x4(_tmpm, _tmpf, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp5 = _mm512_shuffle_f32x4(_tmpj, _tmpn, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp6 = _mm512_shuffle_f32x4(_tmpc, _tmpg, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmp7 = _mm512_shuffle_f32x4(_tmpk, _tmpd, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmp8 = _mm512_shuffle_f32x4(_tmph, _tmpl, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmp9 = _mm512_shuffle_f32x4(_tmpe, _tmpi, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmpa = _mm512_shuffle_f32x4(_tmpm, _tmpf, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmpb = _mm512_shuffle_f32x4(_tmpj, _tmpn, _MM_SHUFFLE(3, 1, 3, 1));
-
-    _r0 = _mm512_shuffle_f32x4(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-    _r1 = _mm512_shuffle_f32x4(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-    _r2 = _mm512_shuffle_f32x4(_tmp4, _tmp5, _MM_SHUFFLE(2, 0, 2, 0));
-    _r3 = _mm512_shuffle_f32x4(_tmp6, _tmp7, _MM_SHUFFLE(2, 0, 2, 0));
-    _r4 = _mm512_shuffle_f32x4(_tmp8, _tmp9, _MM_SHUFFLE(2, 0, 2, 0));
-    _r5 = _mm512_shuffle_f32x4(_tmpa, _tmpb, _MM_SHUFFLE(2, 0, 2, 0));
-    _r6 = _mm512_shuffle_f32x4(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-    _r7 = _mm512_shuffle_f32x4(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-    _r8 = _mm512_shuffle_f32x4(_tmp4, _tmp5, _MM_SHUFFLE(3, 1, 3, 1));
-    _r9 = _mm512_shuffle_f32x4(_tmp6, _tmp7, _MM_SHUFFLE(3, 1, 3, 1));
-    _ra = _mm512_shuffle_f32x4(_tmp8, _tmp9, _MM_SHUFFLE(3, 1, 3, 1));
-    _rb = _mm512_shuffle_f32x4(_tmpa, _tmpb, _MM_SHUFFLE(3, 1, 3, 1));
-}
-
-static NCNN_FORCEINLINE void transpose16x8_ps(__m512& _r0, __m512& _r1, __m512& _r2, __m512& _r3, __m512& _r4, __m512& _r5, __m512& _r6, __m512& _r7)
-{
-    __m512 _tmp0 = _mm512_unpacklo_ps(_r0, _r1);
-    __m512 _tmp1 = _mm512_unpackhi_ps(_r0, _r1);
-    __m512 _tmp2 = _mm512_unpacklo_ps(_r2, _r3);
-    __m512 _tmp3 = _mm512_unpackhi_ps(_r2, _r3);
-    __m512 _tmp4 = _mm512_unpacklo_ps(_r4, _r5);
-    __m512 _tmp5 = _mm512_unpackhi_ps(_r4, _r5);
-    __m512 _tmp6 = _mm512_unpacklo_ps(_r6, _r7);
-    __m512 _tmp7 = _mm512_unpackhi_ps(_r6, _r7);
-
-    __m512 _tmp8 = _mm512_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmp9 = _mm512_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmpa = _mm512_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmpb = _mm512_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmpc = _mm512_shuffle_ps(_tmp4, _tmp6, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmpd = _mm512_shuffle_ps(_tmp4, _tmp6, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmpe = _mm512_shuffle_ps(_tmp5, _tmp7, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmpf = _mm512_shuffle_ps(_tmp5, _tmp7, _MM_SHUFFLE(3, 2, 3, 2));
-
-    _tmp0 = _mm512_shuffle_f32x4(_tmp8, _tmpc, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp1 = _mm512_shuffle_f32x4(_tmp9, _tmpd, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp2 = _mm512_shuffle_f32x4(_tmpa, _tmpe, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp3 = _mm512_shuffle_f32x4(_tmpb, _tmpf, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp4 = _mm512_shuffle_f32x4(_tmp8, _tmpc, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmp5 = _mm512_shuffle_f32x4(_tmp9, _tmpd, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmp6 = _mm512_shuffle_f32x4(_tmpa, _tmpe, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmp7 = _mm512_shuffle_f32x4(_tmpb, _tmpf, _MM_SHUFFLE(3, 1, 3, 1));
-
-    _r0 = _mm512_shuffle_f32x4(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-    _r1 = _mm512_shuffle_f32x4(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-    _r2 = _mm512_shuffle_f32x4(_tmp4, _tmp5, _MM_SHUFFLE(2, 0, 2, 0));
-    _r3 = _mm512_shuffle_f32x4(_tmp6, _tmp7, _MM_SHUFFLE(2, 0, 2, 0));
-    _r4 = _mm512_shuffle_f32x4(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-    _r5 = _mm512_shuffle_f32x4(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-    _r6 = _mm512_shuffle_f32x4(_tmp4, _tmp5, _MM_SHUFFLE(3, 1, 3, 1));
-    _r7 = _mm512_shuffle_f32x4(_tmp6, _tmp7, _MM_SHUFFLE(3, 1, 3, 1));
-}
-
-static NCNN_FORCEINLINE void transpose16x4_ps(__m512& _r0, __m512& _r1, __m512& _r2, __m512& _r3)
-{
-    __m512 _tmp0 = _mm512_unpacklo_ps(_r0, _r1);
-    __m512 _tmp1 = _mm512_unpackhi_ps(_r0, _r1);
-    __m512 _tmp2 = _mm512_unpacklo_ps(_r2, _r3);
-    __m512 _tmp3 = _mm512_unpackhi_ps(_r2, _r3);
-
-    __m512 _tmp4 = _mm512_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmp5 = _mm512_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(3, 2, 3, 2));
-    __m512 _tmp6 = _mm512_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(1, 0, 1, 0));
-    __m512 _tmp7 = _mm512_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(3, 2, 3, 2));
-
-    _tmp0 = _mm512_shuffle_f32x4(_tmp4, _tmp5, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp1 = _mm512_shuffle_f32x4(_tmp6, _tmp7, _MM_SHUFFLE(2, 0, 2, 0));
-    _tmp2 = _mm512_shuffle_f32x4(_tmp4, _tmp5, _MM_SHUFFLE(3, 1, 3, 1));
-    _tmp3 = _mm512_shuffle_f32x4(_tmp6, _tmp7, _MM_SHUFFLE(3, 1, 3, 1));
-
-    _r0 = _mm512_shuffle_f32x4(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-    _r1 = _mm512_shuffle_f32x4(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-    _r2 = _mm512_shuffle_f32x4(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-    _r3 = _mm512_shuffle_f32x4(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-}
-
-static NCNN_FORCEINLINE void transpose16x2_ps(__m512& _r0, __m512& _r1)
-{
-    __m512 _tmp0 = _mm512_unpacklo_ps(_r0, _r1);
-    __m512 _tmp1 = _mm512_unpackhi_ps(_r0, _r1);
-
-    __m512 _tmp2 = _mm512_shuffle_f32x4(_tmp0, _tmp1, _MM_SHUFFLE(2, 0, 2, 0));
-    __m512 _tmp3 = _mm512_shuffle_f32x4(_tmp0, _tmp1, _MM_SHUFFLE(3, 1, 3, 1));
-
-    _r0 = _mm512_shuffle_f32x4(_tmp2, _tmp3, _MM_SHUFFLE(2, 0, 2, 0));
-    _r1 = _mm512_shuffle_f32x4(_tmp2, _tmp3, _MM_SHUFFLE(3, 1, 3, 1));
-}
-
-static NCNN_FORCEINLINE void transpose8x16_ps(__m256& _r0, __m256& _r1, __m256& _r2, __m256& _r3, __m256& _r4, __m256& _r5, __m256& _r6, __m256& _r7,
-        __m256& _r8, __m256& _r9, __m256& _ra, __m256& _rb, __m256& _rc, __m256& _rd, __m256& _re, __m256& _rf)
-{
-    __m256 _tmp0 = _mm256_unpacklo_ps(_r0, _r1);
-    __m256 _tmp1 = _mm256_unpackhi_ps(_r0, _r1);
-    __m256 _tmp2 = _mm256_unpacklo_ps(_r2, _r3);
-    __m256 _tmp3 = _mm256_unpackhi_ps(_r2, _r3);
-    __m256 _tmp4 = _mm256_unpacklo_ps(_r4, _r5);
-    __m256 _tmp5 = _mm256_unpackhi_ps(_r4, _r5);
-    __m256 _tmp6 = _mm256_unpacklo_ps(_r6, _r7);
-    __m256 _tmp7 = _mm256_unpackhi_ps(_r6, _r7);
-    __m256 _tmp8 = _mm256_unpacklo_ps(_r8, _r9);
-    __m256 _tmp9 = _mm256_unpackhi_ps(_r8, _r9);
-    __m256 _tmpa = _mm256_unpacklo_ps(_ra, _rb);
-    __m256 _tmpb = _mm256_unpackhi_ps(_ra, _rb);
-    __m256 _tmpc = _mm256_unpacklo_ps(_rc, _rd);
-    __m256 _tmpd = _mm256_unpackhi_ps(_rc, _rd);
-    __m256 _tmpe = _mm256_unpacklo_ps(_re, _rf);
-    __m256 _tmpf = _mm256_unpackhi_ps(_re, _rf);
-
-    __m256 _tmpg = _mm256_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmph = _mm256_shuffle_ps(_tmp0, _tmp2, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpi = _mm256_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpj = _mm256_shuffle_ps(_tmp1, _tmp3, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpk = _mm256_shuffle_ps(_tmp4, _tmp6, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpl = _mm256_shuffle_ps(_tmp4, _tmp6, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpm = _mm256_shuffle_ps(_tmp5, _tmp7, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpn = _mm256_shuffle_ps(_tmp5, _tmp7, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpo = _mm256_shuffle_ps(_tmp8, _tmpa, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpp = _mm256_shuffle_ps(_tmp8, _tmpa, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpq = _mm256_shuffle_ps(_tmp9, _tmpb, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpr = _mm256_shuffle_ps(_tmp9, _tmpb, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmps = _mm256_shuffle_ps(_tmpc, _tmpe, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpt = _mm256_shuffle_ps(_tmpc, _tmpe, _MM_SHUFFLE(3, 2, 3, 2));
-    __m256 _tmpu = _mm256_shuffle_ps(_tmpd, _tmpf, _MM_SHUFFLE(1, 0, 1, 0));
-    __m256 _tmpv = _mm256_shuffle_ps(_tmpd, _tmpf, _MM_SHUFFLE(3, 2, 3, 2));
-
-    _r0 = _mm256_permute2f128_ps(_tmpg, _tmpk, _MM_SHUFFLE(0, 2, 0, 0));
-    _r1 = _mm256_permute2f128_ps(_tmpo, _tmps, _MM_SHUFFLE(0, 2, 0, 0));
-    _r2 = _mm256_permute2f128_ps(_tmph, _tmpl, _MM_SHUFFLE(0, 2, 0, 0));
-    _r3 = _mm256_permute2f128_ps(_tmpp, _tmpt, _MM_SHUFFLE(0, 2, 0, 0));
-    _r4 = _mm256_permute2f128_ps(_tmpi, _tmpm, _MM_SHUFFLE(0, 2, 0, 0));
-    _r5 = _mm256_permute2f128_ps(_tmpq, _tmpu, _MM_SHUFFLE(0, 2, 0, 0));
-    _r6 = _mm256_permute2f128_ps(_tmpj, _tmpn, _MM_SHUFFLE(0, 2, 0, 0));
-    _r7 = _mm256_permute2f128_ps(_tmpr, _tmpv, _MM_SHUFFLE(0, 2, 0, 0));
-    _r8 = _mm256_permute2f128_ps(_tmpg, _tmpk, _MM_SHUFFLE(0, 3, 0, 1));
-    _r9 = _mm256_permute2f128_ps(_tmpo, _tmps, _MM_SHUFFLE(0, 3, 0, 1));
-    _ra = _mm256_permute2f128_ps(_tmph, _tmpl, _MM_SHUFFLE(0, 3, 0, 1));
-    _rb = _mm256_permute2f128_ps(_tmpp, _tmpt, _MM_SHUFFLE(0, 3, 0, 1));
-    _rc = _mm256_permute2f128_ps(_tmpi, _tmpm, _MM_SHUFFLE(0, 3, 0, 1));
-    _rd = _mm256_permute2f128_ps(_tmpq, _tmpu, _MM_SHUFFLE(0, 3, 0, 1));
-    _re = _mm256_permute2f128_ps(_tmpj, _tmpn, _MM_SHUFFLE(0, 3, 0, 1));
-    _rf = _mm256_permute2f128_ps(_tmpr, _tmpv, _MM_SHUFFLE(0, 3, 0, 1));
-}
-
-static NCNN_FORCEINLINE void transpose16x16_epi16(__m256i& _r0, __m256i& _r1, __m256i& _r2, __m256i& _r3, __m256i& _r4, __m256i& _r5, __m256i& _r6, __m256i& _r7,
-        __m256i& _r8, __m256i& _r9, __m256i& _ra, __m256i& _rb, __m256i& _rc, __m256i& _rd, __m256i& _re, __m256i& _rf)
-{
-    __m256i _tmp0 = _mm256_unpacklo_epi16(_r0, _r1);
-    __m256i _tmp1 = _mm256_unpackhi_epi16(_r0, _r1);
-    __m256i _tmp2 = _mm256_unpacklo_epi16(_r2, _r3);
-    __m256i _tmp3 = _mm256_unpackhi_epi16(_r2, _r3);
-    __m256i _tmp4 = _mm256_unpacklo_epi16(_r4, _r5);
-    __m256i _tmp5 = _mm256_unpackhi_epi16(_r4, _r5);
-    __m256i _tmp6 = _mm256_unpacklo_epi16(_r6, _r7);
-    __m256i _tmp7 = _mm256_unpackhi_epi16(_r6, _r7);
-    __m256i _tmp8 = _mm256_unpacklo_epi16(_r8, _r9);
-    __m256i _tmp9 = _mm256_unpackhi_epi16(_r8, _r9);
-    __m256i _tmpa = _mm256_unpacklo_epi16(_ra, _rb);
-    __m256i _tmpb = _mm256_unpackhi_epi16(_ra, _rb);
-    __m256i _tmpc = _mm256_unpacklo_epi16(_rc, _rd);
-    __m256i _tmpd = _mm256_unpackhi_epi16(_rc, _rd);
-    __m256i _tmpe = _mm256_unpacklo_epi16(_re, _rf);
-    __m256i _tmpf = _mm256_unpackhi_epi16(_re, _rf);
-
-    __m256i _tmpg = _mm256_unpacklo_epi32(_tmp0, _tmp2);
-    __m256i _tmph = _mm256_unpackhi_epi32(_tmp0, _tmp2);
-    __m256i _tmpi = _mm256_unpacklo_epi32(_tmp1, _tmp3);
-    __m256i _tmpj = _mm256_unpackhi_epi32(_tmp1, _tmp3);
-    __m256i _tmpk = _mm256_unpacklo_epi32(_tmp4, _tmp6);
-    __m256i _tmpl = _mm256_unpackhi_epi32(_tmp4, _tmp6);
-    __m256i _tmpm = _mm256_unpacklo_epi32(_tmp5, _tmp7);
-    __m256i _tmpn = _mm256_unpackhi_epi32(_tmp5, _tmp7);
-    __m256i _tmpo = _mm256_unpacklo_epi32(_tmp8, _tmpa);
-    __m256i _tmpp = _mm256_unpackhi_epi32(_tmp8, _tmpa);
-    __m256i _tmpq = _mm256_unpacklo_epi32(_tmp9, _tmpb);
-    __m256i _tmpr = _mm256_unpackhi_epi32(_tmp9, _tmpb);
-    __m256i _tmps = _mm256_unpacklo_epi32(_tmpc, _tmpe);
-    __m256i _tmpt = _mm256_unpackhi_epi32(_tmpc, _tmpe);
-    __m256i _tmpu = _mm256_unpacklo_epi32(_tmpd, _tmpf);
-    __m256i _tmpv = _mm256_unpackhi_epi32(_tmpd, _tmpf);
-
-    _tmp0 = _mm256_unpacklo_epi64(_tmpg, _tmpk);
-    _tmp1 = _mm256_unpackhi_epi64(_tmpg, _tmpk);
-    _tmp2 = _mm256_unpacklo_epi64(_tmph, _tmpl);
-    _tmp3 = _mm256_unpackhi_epi64(_tmph, _tmpl);
-    _tmp4 = _mm256_unpacklo_epi64(_tmpi, _tmpm);
-    _tmp5 = _mm256_unpackhi_epi64(_tmpi, _tmpm);
-    _tmp6 = _mm256_unpacklo_epi64(_tmpj, _tmpn);
-    _tmp7 = _mm256_unpackhi_epi64(_tmpj, _tmpn);
-    _tmp8 = _mm256_unpacklo_epi64(_tmpo, _tmps);
-    _tmp9 = _mm256_unpackhi_epi64(_tmpo, _tmps);
-    _tmpa = _mm256_unpacklo_epi64(_tmpp, _tmpt);
-    _tmpb = _mm256_unpackhi_epi64(_tmpp, _tmpt);
-    _tmpc = _mm256_unpacklo_epi64(_tmpq, _tmpu);
-    _tmpd = _mm256_unpackhi_epi64(_tmpq, _tmpu);
-    _tmpe = _mm256_unpacklo_epi64(_tmpr, _tmpv);
-    _tmpf = _mm256_unpackhi_epi64(_tmpr, _tmpv);
-
-    _r0 = _mm256_permute2x128_si256(_tmp0, _tmp8, _MM_SHUFFLE(0, 2, 0, 0));
-    _r1 = _mm256_permute2x128_si256(_tmp1, _tmp9, _MM_SHUFFLE(0, 2, 0, 0));
-    _r2 = _mm256_permute2x128_si256(_tmp2, _tmpa, _MM_SHUFFLE(0, 2, 0, 0));
-    _r3 = _mm256_permute2x128_si256(_tmp3, _tmpb, _MM_SHUFFLE(0, 2, 0, 0));
-    _r4 = _mm256_permute2x128_si256(_tmp4, _tmpc, _MM_SHUFFLE(0, 2, 0, 0));
-    _r5 = _mm256_permute2x128_si256(_tmp5, _tmpd, _MM_SHUFFLE(0, 2, 0, 0));
-    _r6 = _mm256_permute2x128_si256(_tmp6, _tmpe, _MM_SHUFFLE(0, 2, 0, 0));
-    _r7 = _mm256_permute2x128_si256(_tmp7, _tmpf, _MM_SHUFFLE(0, 2, 0, 0));
-    _r8 = _mm256_permute2x128_si256(_tmp0, _tmp8, _MM_SHUFFLE(0, 3, 0, 1));
-    _r9 = _mm256_permute2x128_si256(_tmp1, _tmp9, _MM_SHUFFLE(0, 3, 0, 1));
-    _ra = _mm256_permute2x128_si256(_tmp2, _tmpa, _MM_SHUFFLE(0, 3, 0, 1));
-    _rb = _mm256_permute2x128_si256(_tmp3, _tmpb, _MM_SHUFFLE(0, 3, 0, 1));
-    _rc = _mm256_permute2x128_si256(_tmp4, _tmpc, _MM_SHUFFLE(0, 3, 0, 1));
-    _rd = _mm256_permute2x128_si256(_tmp5, _tmpd, _MM_SHUFFLE(0, 3, 0, 1));
-    _re = _mm256_permute2x128_si256(_tmp6, _tmpe, _MM_SHUFFLE(0, 3, 0, 1));
-    _rf = _mm256_permute2x128_si256(_tmp7, _tmpf, _MM_SHUFFLE(0, 3, 0, 1));
-}
-
-static NCNN_FORCEINLINE void transpose8x16_epi16(__m128i& _r0, __m128i& _r1, __m128i& _r2, __m128i& _r3, __m128i& _r4, __m128i& _r5, __m128i& _r6, __m128i& _r7, __m128i& _r8, __m128i& _r9, __m128i& _ra, __m128i& _rb, __m128i& _rc, __m128i& _rd, __m128i& _re, __m128i& _rf)
-{
-    __m128i _tmp0 = _mm_unpacklo_epi16(_r0, _r1);
-    __m128i _tmp1 = _mm_unpackhi_epi16(_r0, _r1);
-    __m128i _tmp2 = _mm_unpacklo_epi16(_r2, _r3);
-    __m128i _tmp3 = _mm_unpackhi_epi16(_r2, _r3);
-    __m128i _tmp4 = _mm_unpacklo_epi16(_r4, _r5);
-    __m128i _tmp5 = _mm_unpackhi_epi16(_r4, _r5);
-    __m128i _tmp6 = _mm_unpacklo_epi16(_r6, _r7);
-    __m128i _tmp7 = _mm_unpackhi_epi16(_r6, _r7);
-    __m128i _tmp8 = _mm_unpacklo_epi16(_r8, _r9);
-    __m128i _tmp9 = _mm_unpackhi_epi16(_r8, _r9);
-    __m128i _tmpa = _mm_unpacklo_epi16(_ra, _rb);
-    __m128i _tmpb = _mm_unpackhi_epi16(_ra, _rb);
-    __m128i _tmpc = _mm_unpacklo_epi16(_rc, _rd);
-    __m128i _tmpd = _mm_unpackhi_epi16(_rc, _rd);
-    __m128i _tmpe = _mm_unpacklo_epi16(_re, _rf);
-    __m128i _tmpf = _mm_unpackhi_epi16(_re, _rf);
-
-    __m128i _tmpg = _mm_unpacklo_epi32(_tmp0, _tmp2);
-    __m128i _tmph = _mm_unpackhi_epi32(_tmp0, _tmp2);
-    __m128i _tmpi = _mm_unpacklo_epi32(_tmp1, _tmp3);
-    __m128i _tmpj = _mm_unpackhi_epi32(_tmp1, _tmp3);
-    __m128i _tmpk = _mm_unpacklo_epi32(_tmp4, _tmp6);
-    __m128i _tmpl = _mm_unpackhi_epi32(_tmp4, _tmp6);
-    __m128i _tmpm = _mm_unpacklo_epi32(_tmp5, _tmp7);
-    __m128i _tmpn = _mm_unpackhi_epi32(_tmp5, _tmp7);
-    __m128i _tmpo = _mm_unpacklo_epi32(_tmp8, _tmpa);
-    __m128i _tmpp = _mm_unpackhi_epi32(_tmp8, _tmpa);
-    __m128i _tmpq = _mm_unpacklo_epi32(_tmp9, _tmpb);
-    __m128i _tmpr = _mm_unpackhi_epi32(_tmp9, _tmpb);
-    __m128i _tmps = _mm_unpacklo_epi32(_tmpc, _tmpe);
-    __m128i _tmpt = _mm_unpackhi_epi32(_tmpc, _tmpe);
-    __m128i _tmpu = _mm_unpacklo_epi32(_tmpd, _tmpf);
-    __m128i _tmpv = _mm_unpackhi_epi32(_tmpd, _tmpf);
-
-    _r0 = _mm_unpacklo_epi64(_tmpg, _tmpk);
-    _r1 = _mm_unpacklo_epi64(_tmpo, _tmps);
-    _r2 = _mm_unpackhi_epi64(_tmpg, _tmpk);
-    _r3 = _mm_unpackhi_epi64(_tmpo, _tmps);
-    _r4 = _mm_unpacklo_epi64(_tmph, _tmpl);
-    _r5 = _mm_unpacklo_epi64(_tmpp, _tmpt);
-    _r6 = _mm_unpackhi_epi64(_tmph, _tmpl);
-    _r7 = _mm_unpackhi_epi64(_tmpp, _tmpt);
-    _r8 = _mm_unpacklo_epi64(_tmpi, _tmpm);
-    _r9 = _mm_unpacklo_epi64(_tmpq, _tmpu);
-    _ra = _mm_unpackhi_epi64(_tmpi, _tmpm);
-    _rb = _mm_unpackhi_epi64(_tmpq, _tmpu);
-    _rc = _mm_unpacklo_epi64(_tmpj, _tmpn);
-    _rd = _mm_unpacklo_epi64(_tmpr, _tmpv);
-    _re = _mm_unpackhi_epi64(_tmpj, _tmpn);
-    _rf = _mm_unpackhi_epi64(_tmpr, _tmpv);
-}
-
-static NCNN_FORCEINLINE float _mm512_comp_reduce_add_ps(__m512 x)
-{
-    const __m256 x256 = _mm256_add_ps(_mm512_castps512_ps256(x), _mm512_extractf32x8_ps(x, 1));
-    const __m128 x128 = _mm_add_ps(_mm256_castps256_ps128(x256), _mm256_extractf128_ps(x256, 1));
-    const __m128 x64 = _mm_add_ps(x128, _mm_movehl_ps(x128, x128));
-    const __m128 x32 = _mm_add_ss(x64, _mm_shuffle_ps(x64, x64, 0x55));
-    return _mm_cvtss_f32(x32);
-}
-
-static NCNN_FORCEINLINE float _mm512_comp_reduce_max_ps(__m512 x)
-{
-    const __m256 x256 = _mm256_max_ps(_mm512_castps512_ps256(x), _mm512_extractf32x8_ps(x, 1));
-    const __m128 x128 = _mm_max_ps(_mm256_castps256_ps128(x256), _mm256_extractf128_ps(x256, 1));
-    const __m128 x64 = _mm_max_ps(x128, _mm_movehl_ps(x128, x128));
-    const __m128 x32 = _mm_max_ss(x64, _mm_shuffle_ps(x64, x64, 0x55));
-    return _mm_cvtss_f32(x32);
-}
-
-static NCNN_FORCEINLINE __m512 bfloat2float_avx512(const __m256i& v0)
-{
-#if __AVX512BF16__
-    __m512 _v = _mm512_cvtpbh_ps((__m256bh)v0);
-#else
-    __m256i _zero = _mm256_setzero_si256();
-    __m256i _a = _mm256_unpacklo_epi16(_zero, v0);
-    __m256i _b = _mm256_unpackhi_epi16(_zero, v0);
-    __m256i _c = _mm256_permute2x128_si256(_a, _b, _MM_SHUFFLE(0, 2, 0, 0));
-    __m256i _d = _mm256_permute2x128_si256(_a, _b, _MM_SHUFFLE(0, 3, 0, 1));
-    __m512 _v = _mm512_castsi512_ps(_mm512_inserti32x8(_mm512_castsi256_si512(_c), _d, 1));
-#endif
-    return _v;
-}
-
-static NCNN_FORCEINLINE __m256i float2bfloat_avx512(const __m512& v0)
-{
-#if __AVX512BF16__
-    __m256i _v = (__m256i)_mm512_cvtneps_pbh(v0);
-#else
-    __m512i _ab = _mm512_castps_si512(v0);
-    _ab = _mm512_srli_epi32(_ab, 16);
-    __m256i _a = _mm512_extracti32x8_epi32(_ab, 0);
-    __m256i _b = _mm512_extracti32x8_epi32(_ab, 1);
-    __m256i _v = _mm256_packus_epi32(_a, _b);
-    _v = _mm256_permute4x64_epi64(_v, _MM_SHUFFLE(3, 1, 2, 0));
-#endif
-    return _v;
-}
-
-static NCNN_FORCEINLINE __m512i float2bfloat_avx512(const __m512& v0, const __m512& v1)
-{
-#if __AVX512BF16__
-    __m256bh _v0 = _mm512_cvtneps_pbh(v0);
-    __m256bh _v1 = _mm512_cvtneps_pbh(v1);
-    __m512i _v = _mm512_inserti32x8(_mm512_castsi256_si512((__m256i)_v0), (__m256i)_v1, 1);
-#else
-    __m512i _a = _mm512_castps_si512(v0);
-    __m512i _b = _mm512_castps_si512(v1);
-    _a = _mm512_srli_epi32(_a, 16);
-    _b = _mm512_srli_epi32(_b, 16);
-    __m512i _v = _mm512_packus_epi32(_a, _b);
-    _v = _mm512_permutex_epi64(_v, _MM_SHUFFLE(3, 1, 2, 0));
-    _v = _mm512_shuffle_i32x4(_v, _v, _MM_SHUFFLE(3, 1, 2, 0));
-#endif
-    return _v;
-}
-
-#endif // __AVX512F__
-#endif // __AVX2__
-#endif // __AVX__
-#endif // __SSE2__
-
-#endif // X86_USABILITY_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/yolov3detectionoutput_x86.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/yolov3detectionoutput_x86.h
deleted file mode 100644
index c378b582..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/x86/yolov3detectionoutput_x86.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_YOLOV3DETECTIONOUTPUT_X86_H
-#define LAYER_YOLOV3DETECTIONOUTPUT_X86_H
-
-#include "yolov3detectionoutput.h"
-
-namespace ncnn {
-
-class Yolov3DetectionOutput_x86 : public Yolov3DetectionOutput
-{
-public:
-    Yolov3DetectionOutput_x86();
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_YOLOV3DETECTIONOUTPUT_X86_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/yolodetectionoutput.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/yolodetectionoutput.h
deleted file mode 100644
index 5dc35afb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/yolodetectionoutput.h
+++ /dev/null
@@ -1,46 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2018 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_YOLODETECTIONOUTPUT_H
-#define LAYER_YOLODETECTIONOUTPUT_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class YoloDetectionOutput : public Layer
-{
-public:
-    YoloDetectionOutput();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int create_pipeline(const Option& opt);
-    virtual int destroy_pipeline(const Option& opt);
-
-    virtual int forward_inplace(std::vector<Mat>& bottom_top_blobs, const Option& opt) const;
-
-public:
-    int num_class;
-    int num_box;
-    float confidence_threshold;
-    float nms_threshold;
-    Mat biases;
-
-    ncnn::Layer* softmax;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_YOLODETECTIONOUTPUT_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/yolov3detectionoutput.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/yolov3detectionoutput.h
deleted file mode 100644
index fbaf339e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer/yolov3detectionoutput.h
+++ /dev/null
@@ -1,62 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2018 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef LAYER_YOLOV3DETECTIONOUTPUT_H
-#define LAYER_YOLOV3DETECTIONOUTPUT_H
-
-#include "layer.h"
-
-namespace ncnn {
-
-class Yolov3DetectionOutput : public Layer
-{
-public:
-    Yolov3DetectionOutput();
-    ~Yolov3DetectionOutput();
-
-    virtual int load_param(const ParamDict& pd);
-
-    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
-
-public:
-    int num_class;
-    int num_box;
-    float confidence_threshold;
-    float nms_threshold;
-    Mat biases;
-    Mat mask;
-    Mat anchors_scale;
-    int mask_group_num;
-    ncnn::Layer* softmax;
-
-public:
-    struct BBoxRect
-    {
-        float score;
-        float xmin;
-        float ymin;
-        float xmax;
-        float ymax;
-        float area;
-        int label;
-    };
-
-    void qsort_descent_inplace(std::vector<BBoxRect>& datas, int left, int right) const;
-    void qsort_descent_inplace(std::vector<BBoxRect>& datas) const;
-    void nms_sorted_bboxes(std::vector<BBoxRect>& bboxes, std::vector<size_t>& picked, float nms_threshold) const;
-};
-
-} // namespace ncnn
-
-#endif // LAYER_YOLODETECTIONOUTPUT_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_declaration.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_declaration.h
deleted file mode 100644
index f55fdbbe..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_declaration.h
+++ /dev/null
@@ -1,125 +0,0 @@
-// Layer Declaration header
-//
-// This file is auto-generated by cmake, don't edit it.
-
-#include "layer/concat.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Concat) }
-#include "layer/arm/concat_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Concat_arm) }
-#include "layer/convolution.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Convolution) }
-#include "layer/arm/convolution_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Convolution_arm) }
-#include "layer/crop.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Crop) }
-#include "layer/arm/crop_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Crop_arm) }
-#include "layer/dropout.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Dropout) }
-#include "layer/arm/dropout_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Dropout_arm) }
-#include "layer/embed.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Embed) }
-#include "layer/flatten.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Flatten) }
-#include "layer/arm/flatten_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Flatten_arm) }
-#include "layer/innerproduct.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(InnerProduct) }
-#include "layer/arm/innerproduct_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(InnerProduct_arm) }
-#include "layer/input.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Input) }
-#include "layer/memorydata.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(MemoryData) }
-#include "layer/reduction.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Reduction) }
-#include "layer/relu.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(ReLU) }
-#include "layer/arm/relu_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(ReLU_arm) }
-#include "layer/reshape.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Reshape) }
-#include "layer/arm/reshape_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Reshape_arm) }
-#include "layer/sigmoid.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Sigmoid) }
-#include "layer/arm/sigmoid_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Sigmoid_arm) }
-#include "layer/slice.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Slice) }
-#include "layer/arm/slice_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Slice_arm) }
-#include "layer/softmax.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Softmax) }
-#include "layer/arm/softmax_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Softmax_arm) }
-#include "layer/split.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Split) }
-#include "layer/tanh.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(TanH) }
-#include "layer/arm/tanh_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(TanH_arm) }
-#include "layer/rnn.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(RNN) }
-#include "layer/arm/rnn_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(RNN_arm) }
-#include "layer/lstm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(LSTM) }
-#include "layer/arm/lstm_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(LSTM_arm) }
-#include "layer/binaryop.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(BinaryOp) }
-#include "layer/arm/binaryop_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(BinaryOp_arm) }
-#include "layer/unaryop.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(UnaryOp) }
-#include "layer/arm/unaryop_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(UnaryOp_arm) }
-#include "layer/padding.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Padding) }
-#include "layer/arm/padding_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Padding_arm) }
-#include "layer/expanddims.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(ExpandDims) }
-#include "layer/permute.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Permute) }
-#include "layer/quantize.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Quantize) }
-#include "layer/arm/quantize_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Quantize_arm) }
-#include "layer/dequantize.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Dequantize) }
-#include "layer/arm/dequantize_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Dequantize_arm) }
-#include "layer/packing.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Packing) }
-#include "layer/arm/packing_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Packing_arm) }
-#include "layer/requantize.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Requantize) }
-#include "layer/arm/requantize_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Requantize_arm) }
-#include "layer/cast.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Cast) }
-#include "layer/arm/cast_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Cast_arm) }
-#include "layer/gemm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Gemm) }
-#include "layer/arm/gemm_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Gemm_arm) }
-#include "layer/convolution1d.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Convolution1D) }
-#include "layer/arm/convolution1d_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(Convolution1D_arm) }
-#include "layer/convolutiondepthwise1d.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(ConvolutionDepthWise1D) }
-#include "layer/matmul.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(MatMul) }
-#include "layer/arm/matmul_arm.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(MatMul_arm) }
-#include "layer/deconvolutiondepthwise1d.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(DeconvolutionDepthWise1D) }
-#include "layer/glu.h"
-namespace ncnn { DEFINE_LAYER_CREATOR(GLU) }
-
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_registry.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_registry.h
deleted file mode 100644
index 1cd706f6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_registry.h
+++ /dev/null
@@ -1,1619 +0,0 @@
-// Layer Registry header
-//
-// This file is auto-generated by cmake, don't edit it.
-
-static const layer_registry_entry layer_registry[] = {
-#if NCNN_STRING
-{"AbsVal", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ArgMax", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"BatchNorm", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Bias", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"BNLL", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Concat", Concat_layer_creator},
-#else
-{Concat_layer_creator},
-#endif
-#if NCNN_STRING
-{"Convolution", Convolution_layer_creator},
-#else
-{Convolution_layer_creator},
-#endif
-#if NCNN_STRING
-{"Crop", Crop_layer_creator},
-#else
-{Crop_layer_creator},
-#endif
-#if NCNN_STRING
-{"Deconvolution", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Dropout", Dropout_layer_creator},
-#else
-{Dropout_layer_creator},
-#endif
-#if NCNN_STRING
-{"Eltwise", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ELU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Embed", Embed_layer_creator},
-#else
-{Embed_layer_creator},
-#endif
-#if NCNN_STRING
-{"Exp", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Flatten", Flatten_layer_creator},
-#else
-{Flatten_layer_creator},
-#endif
-#if NCNN_STRING
-{"InnerProduct", InnerProduct_layer_creator},
-#else
-{InnerProduct_layer_creator},
-#endif
-#if NCNN_STRING
-{"Input", Input_layer_creator},
-#else
-{Input_layer_creator},
-#endif
-#if NCNN_STRING
-{"Log", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"LRN", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"MemoryData", MemoryData_layer_creator},
-#else
-{MemoryData_layer_creator},
-#endif
-#if NCNN_STRING
-{"MVN", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Pooling", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Power", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"PReLU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Proposal", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Reduction", Reduction_layer_creator},
-#else
-{Reduction_layer_creator},
-#endif
-#if NCNN_STRING
-{"ReLU", ReLU_layer_creator},
-#else
-{ReLU_layer_creator},
-#endif
-#if NCNN_STRING
-{"Reshape", Reshape_layer_creator},
-#else
-{Reshape_layer_creator},
-#endif
-#if NCNN_STRING
-{"ROIPooling", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Scale", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Sigmoid", Sigmoid_layer_creator},
-#else
-{Sigmoid_layer_creator},
-#endif
-#if NCNN_STRING
-{"Slice", Slice_layer_creator},
-#else
-{Slice_layer_creator},
-#endif
-#if NCNN_STRING
-{"Softmax", Softmax_layer_creator},
-#else
-{Softmax_layer_creator},
-#endif
-#if NCNN_STRING
-{"Split", Split_layer_creator},
-#else
-{Split_layer_creator},
-#endif
-#if NCNN_STRING
-{"SPP", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"TanH", TanH_layer_creator},
-#else
-{TanH_layer_creator},
-#endif
-#if NCNN_STRING
-{"Threshold", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Tile", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"RNN", RNN_layer_creator},
-#else
-{RNN_layer_creator},
-#endif
-#if NCNN_STRING
-{"LSTM", LSTM_layer_creator},
-#else
-{LSTM_layer_creator},
-#endif
-#if NCNN_STRING
-{"BinaryOp", BinaryOp_layer_creator},
-#else
-{BinaryOp_layer_creator},
-#endif
-#if NCNN_STRING
-{"UnaryOp", UnaryOp_layer_creator},
-#else
-{UnaryOp_layer_creator},
-#endif
-#if NCNN_STRING
-{"ConvolutionDepthWise", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Padding", Padding_layer_creator},
-#else
-{Padding_layer_creator},
-#endif
-#if NCNN_STRING
-{"Squeeze", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ExpandDims", ExpandDims_layer_creator},
-#else
-{ExpandDims_layer_creator},
-#endif
-#if NCNN_STRING
-{"Normalize", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Permute", Permute_layer_creator},
-#else
-{Permute_layer_creator},
-#endif
-#if NCNN_STRING
-{"PriorBox", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DetectionOutput", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Interp", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DeconvolutionDepthWise", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ShuffleChannel", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"InstanceNorm", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Clip", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Reorg", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"YoloDetectionOutput", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Quantize", Quantize_layer_creator},
-#else
-{Quantize_layer_creator},
-#endif
-#if NCNN_STRING
-{"Dequantize", Dequantize_layer_creator},
-#else
-{Dequantize_layer_creator},
-#endif
-#if NCNN_STRING
-{"Yolov3DetectionOutput", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"PSROIPooling", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ROIAlign", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Packing", Packing_layer_creator},
-#else
-{Packing_layer_creator},
-#endif
-#if NCNN_STRING
-{"Requantize", Requantize_layer_creator},
-#else
-{Requantize_layer_creator},
-#endif
-#if NCNN_STRING
-{"Cast", Cast_layer_creator},
-#else
-{Cast_layer_creator},
-#endif
-#if NCNN_STRING
-{"HardSigmoid", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"SELU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"HardSwish", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Noop", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"PixelShuffle", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DeepCopy", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Mish", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"StatisticsPooling", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Swish", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Gemm", Gemm_layer_creator},
-#else
-{Gemm_layer_creator},
-#endif
-#if NCNN_STRING
-{"GroupNorm", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"LayerNorm", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Softplus", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"GRU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"MultiHeadAttention", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"GELU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Convolution1D", Convolution1D_layer_creator},
-#else
-{Convolution1D_layer_creator},
-#endif
-#if NCNN_STRING
-{"Pooling1D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ConvolutionDepthWise1D", ConvolutionDepthWise1D_layer_creator},
-#else
-{ConvolutionDepthWise1D_layer_creator},
-#endif
-#if NCNN_STRING
-{"Convolution3D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ConvolutionDepthWise3D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Pooling3D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"MatMul", MatMul_layer_creator},
-#else
-{MatMul_layer_creator},
-#endif
-#if NCNN_STRING
-{"Deconvolution1D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DeconvolutionDepthWise1D", DeconvolutionDepthWise1D_layer_creator},
-#else
-{DeconvolutionDepthWise1D_layer_creator},
-#endif
-#if NCNN_STRING
-{"Deconvolution3D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DeconvolutionDepthWise3D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Einsum", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DeformableConv2D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"GLU", GLU_layer_creator},
-#else
-{GLU_layer_creator},
-#endif
-#if NCNN_STRING
-{"Fold", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Unfold", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"GridSample", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"CumulativeSum", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"CopyTo", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Erf", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Diag", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"CELU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Shrink", 0},
-#else
-{0},
-#endif
-
-};
-
-static const layer_registry_entry layer_registry_arch[] = {
-#if NCNN_STRING
-{"AbsVal", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ArgMax", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"BatchNorm", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Bias", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"BNLL", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Concat", Concat_arm_layer_creator},
-#else
-{Concat_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Convolution", Convolution_arm_layer_creator},
-#else
-{Convolution_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Crop", Crop_arm_layer_creator},
-#else
-{Crop_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Deconvolution", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Dropout", Dropout_arm_layer_creator},
-#else
-{Dropout_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Eltwise", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ELU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Embed", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Exp", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Flatten", Flatten_arm_layer_creator},
-#else
-{Flatten_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"InnerProduct", InnerProduct_arm_layer_creator},
-#else
-{InnerProduct_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Input", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Log", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"LRN", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"MemoryData", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"MVN", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Pooling", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Power", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"PReLU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Proposal", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Reduction", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ReLU", ReLU_arm_layer_creator},
-#else
-{ReLU_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Reshape", Reshape_arm_layer_creator},
-#else
-{Reshape_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"ROIPooling", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Scale", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Sigmoid", Sigmoid_arm_layer_creator},
-#else
-{Sigmoid_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Slice", Slice_arm_layer_creator},
-#else
-{Slice_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Softmax", Softmax_arm_layer_creator},
-#else
-{Softmax_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Split", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"SPP", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"TanH", TanH_arm_layer_creator},
-#else
-{TanH_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Threshold", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Tile", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"RNN", RNN_arm_layer_creator},
-#else
-{RNN_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"LSTM", LSTM_arm_layer_creator},
-#else
-{LSTM_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"BinaryOp", BinaryOp_arm_layer_creator},
-#else
-{BinaryOp_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"UnaryOp", UnaryOp_arm_layer_creator},
-#else
-{UnaryOp_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"ConvolutionDepthWise", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Padding", Padding_arm_layer_creator},
-#else
-{Padding_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Squeeze", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ExpandDims", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Normalize", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Permute", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"PriorBox", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DetectionOutput", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Interp", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DeconvolutionDepthWise", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ShuffleChannel", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"InstanceNorm", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Clip", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Reorg", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"YoloDetectionOutput", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Quantize", Quantize_arm_layer_creator},
-#else
-{Quantize_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Dequantize", Dequantize_arm_layer_creator},
-#else
-{Dequantize_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Yolov3DetectionOutput", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"PSROIPooling", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ROIAlign", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Packing", Packing_arm_layer_creator},
-#else
-{Packing_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Requantize", Requantize_arm_layer_creator},
-#else
-{Requantize_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Cast", Cast_arm_layer_creator},
-#else
-{Cast_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"HardSigmoid", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"SELU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"HardSwish", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Noop", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"PixelShuffle", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DeepCopy", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Mish", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"StatisticsPooling", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Swish", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Gemm", Gemm_arm_layer_creator},
-#else
-{Gemm_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"GroupNorm", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"LayerNorm", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Softplus", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"GRU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"MultiHeadAttention", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"GELU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Convolution1D", Convolution1D_arm_layer_creator},
-#else
-{Convolution1D_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Pooling1D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ConvolutionDepthWise1D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Convolution3D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ConvolutionDepthWise3D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Pooling3D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"MatMul", MatMul_arm_layer_creator},
-#else
-{MatMul_arm_layer_creator},
-#endif
-#if NCNN_STRING
-{"Deconvolution1D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DeconvolutionDepthWise1D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Deconvolution3D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DeconvolutionDepthWise3D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Einsum", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DeformableConv2D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"GLU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Fold", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Unfold", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"GridSample", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"CumulativeSum", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"CopyTo", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Erf", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Diag", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"CELU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Shrink", 0},
-#else
-{0},
-#endif
-
-};
-
-#if NCNN_RUNTIME_CPU && NCNN_AVX512
-static const layer_registry_entry layer_registry_avx512[] = {
-
-};
-#endif // NCNN_RUNTIME_CPU && NCNN_AVX512
-
-#if NCNN_RUNTIME_CPU && NCNN_FMA
-static const layer_registry_entry layer_registry_fma[] = {
-
-};
-#endif // NCNN_RUNTIME_CPU && NCNN_FMA
-
-#if NCNN_RUNTIME_CPU && NCNN_AVX
-static const layer_registry_entry layer_registry_avx[] = {
-
-};
-#endif // NCNN_RUNTIME_CPU && NCNN_AVX
-
-#if NCNN_RUNTIME_CPU && NCNN_MSA
-static const layer_registry_entry layer_registry_msa[] = {
-
-};
-#endif // NCNN_RUNTIME_CPU && NCNN_MSA
-
-#if NCNN_RUNTIME_CPU && NCNN_LASX
-static const layer_registry_entry layer_registry_lasx[] = {
-
-};
-#endif // NCNN_RUNTIME_CPU && NCNN_LASX
-
-#if NCNN_RUNTIME_CPU && NCNN_LSX
-static const layer_registry_entry layer_registry_lsx[] = {
-
-};
-#endif // NCNN_RUNTIME_CPU && NCNN_LSX
-
-#if NCNN_RUNTIME_CPU && NCNN_RVV
-static const layer_registry_entry layer_registry_rvv[] = {
-
-};
-#endif // NCNN_RUNTIME_CPU && NCNN_RVV
-
-#if NCNN_VULKAN
-static const layer_registry_entry layer_registry_vulkan[] = {
-#if NCNN_STRING
-{"AbsVal", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ArgMax", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"BatchNorm", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Bias", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"BNLL", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Concat", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Convolution", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Crop", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Deconvolution", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Dropout", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Eltwise", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ELU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Embed", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Exp", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Flatten", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"InnerProduct", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Input", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Log", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"LRN", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"MemoryData", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"MVN", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Pooling", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Power", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"PReLU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Proposal", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Reduction", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ReLU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Reshape", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ROIPooling", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Scale", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Sigmoid", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Slice", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Softmax", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Split", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"SPP", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"TanH", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Threshold", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Tile", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"RNN", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"LSTM", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"BinaryOp", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"UnaryOp", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ConvolutionDepthWise", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Padding", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Squeeze", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ExpandDims", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Normalize", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Permute", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"PriorBox", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DetectionOutput", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Interp", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DeconvolutionDepthWise", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ShuffleChannel", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"InstanceNorm", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Clip", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Reorg", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"YoloDetectionOutput", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Quantize", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Dequantize", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Yolov3DetectionOutput", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"PSROIPooling", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ROIAlign", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Packing", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Requantize", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Cast", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"HardSigmoid", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"SELU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"HardSwish", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Noop", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"PixelShuffle", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DeepCopy", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Mish", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"StatisticsPooling", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Swish", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Gemm", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"GroupNorm", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"LayerNorm", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Softplus", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"GRU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"MultiHeadAttention", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"GELU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Convolution1D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Pooling1D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ConvolutionDepthWise1D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Convolution3D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"ConvolutionDepthWise3D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Pooling3D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"MatMul", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Deconvolution1D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DeconvolutionDepthWise1D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Deconvolution3D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DeconvolutionDepthWise3D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Einsum", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"DeformableConv2D", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"GLU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Fold", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Unfold", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"GridSample", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"CumulativeSum", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"CopyTo", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Erf", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Diag", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"CELU", 0},
-#else
-{0},
-#endif
-#if NCNN_STRING
-{"Shrink", 0},
-#else
-{0},
-#endif
-
-};
-#endif // NCNN_VULKAN
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_registry.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_registry.h
deleted file mode 100644
index a9c67e06..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_registry.h
+++ /dev/null
@@ -1,5 +0,0 @@
-// Layer Shader Registry header
-//
-// This file is auto-generated by cmake, don't edit it.
-
-
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_spv_data.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_spv_data.h
deleted file mode 100644
index c8949837..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_spv_data.h
+++ /dev/null
@@ -1,5 +0,0 @@
-// Layer Shader Spv Data header
-//
-// This file is auto-generated by cmake, don't edit it.
-
-
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_type.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_type.h
deleted file mode 100644
index c143e7de..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_type.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_LAYER_SHADER_TYPE_H
-#define NCNN_LAYER_SHADER_TYPE_H
-
-namespace ncnn {
-
-namespace LayerShaderType {
-enum LayerShaderType
-{
-#include "layer_shader_type_enum.h"
-};
-} // namespace LayerShaderType
-
-} // namespace ncnn
-
-#endif // NCNN_LAYER_SHADER_TYPE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_type_enum.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_type_enum.h
deleted file mode 100644
index aac8803b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_shader_type_enum.h
+++ /dev/null
@@ -1,5 +0,0 @@
-// Layer Shader Enum header
-//
-// This file is auto-generated by cmake, don't edit it.
-
-
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_type.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_type.h
deleted file mode 100644
index 511c714b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_type.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_LAYER_TYPE_H
-#define NCNN_LAYER_TYPE_H
-
-namespace ncnn {
-
-namespace LayerType {
-enum LayerType
-{
-#include "layer_type_enum.h"
-    CustomBit = (1 << 8),
-};
-} // namespace LayerType
-
-} // namespace ncnn
-
-#endif // NCNN_LAYER_TYPE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_type_enum.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_type_enum.h
deleted file mode 100644
index 97153edb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/layer_type_enum.h
+++ /dev/null
@@ -1,109 +0,0 @@
-// Layer Type Enum header
-//
-// This file is auto-generated by cmake, don't edit it.
-
-AbsVal = 0,
-ArgMax = 1,
-BatchNorm = 2,
-Bias = 3,
-BNLL = 4,
-Concat = 5,
-Convolution = 6,
-Crop = 7,
-Deconvolution = 8,
-Dropout = 9,
-Eltwise = 10,
-ELU = 11,
-Embed = 12,
-Exp = 13,
-Flatten = 14,
-InnerProduct = 15,
-Input = 16,
-Log = 17,
-LRN = 18,
-MemoryData = 19,
-MVN = 20,
-Pooling = 21,
-Power = 22,
-PReLU = 23,
-Proposal = 24,
-Reduction = 25,
-ReLU = 26,
-Reshape = 27,
-ROIPooling = 28,
-Scale = 29,
-Sigmoid = 30,
-Slice = 31,
-Softmax = 32,
-Split = 33,
-SPP = 34,
-TanH = 35,
-Threshold = 36,
-Tile = 37,
-RNN = 38,
-LSTM = 39,
-BinaryOp = 40,
-UnaryOp = 41,
-ConvolutionDepthWise = 42,
-Padding = 43,
-Squeeze = 44,
-ExpandDims = 45,
-Normalize = 46,
-Permute = 47,
-PriorBox = 48,
-DetectionOutput = 49,
-Interp = 50,
-DeconvolutionDepthWise = 51,
-ShuffleChannel = 52,
-InstanceNorm = 53,
-Clip = 54,
-Reorg = 55,
-YoloDetectionOutput = 56,
-Quantize = 57,
-Dequantize = 58,
-Yolov3DetectionOutput = 59,
-PSROIPooling = 60,
-ROIAlign = 61,
-Packing = 62,
-Requantize = 63,
-Cast = 64,
-HardSigmoid = 65,
-SELU = 66,
-HardSwish = 67,
-Noop = 68,
-PixelShuffle = 69,
-DeepCopy = 70,
-Mish = 71,
-StatisticsPooling = 72,
-Swish = 73,
-Gemm = 74,
-GroupNorm = 75,
-LayerNorm = 76,
-Softplus = 77,
-GRU = 78,
-MultiHeadAttention = 79,
-GELU = 80,
-Convolution1D = 81,
-Pooling1D = 82,
-ConvolutionDepthWise1D = 83,
-Convolution3D = 84,
-ConvolutionDepthWise3D = 85,
-Pooling3D = 86,
-MatMul = 87,
-Deconvolution1D = 88,
-DeconvolutionDepthWise1D = 89,
-Deconvolution3D = 90,
-DeconvolutionDepthWise3D = 91,
-Einsum = 92,
-DeformableConv2D = 93,
-GLU = 94,
-Fold = 95,
-Unfold = 96,
-GridSample = 97,
-CumulativeSum = 98,
-CopyTo = 99,
-Erf = 100,
-Diag = 101,
-CELU = 102,
-Shrink = 103,
-
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/mat.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/mat.h
deleted file mode 100644
index fdf5cc59..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/mat.h
+++ /dev/null
@@ -1,1847 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_MAT_H
-#define NCNN_MAT_H
-
-#include <stdlib.h>
-#include <string.h>
-#if __ARM_NEON
-#include <arm_neon.h>
-#endif
-#if __SSE2__
-#include <emmintrin.h>
-#if __AVX__
-#include <immintrin.h>
-#endif
-#endif
-#if __mips_msa
-#include <msa.h>
-#endif
-#if __loongarch_sx
-#include <lsxintrin.h>
-#endif
-#if __riscv_vector
-#include <riscv_vector.h>
-#include "cpu.h" // cpu_riscv_vlenb()
-#endif
-
-#include "allocator.h"
-#include "option.h"
-#include "platform.h"
-
-#if NCNN_PIXEL
-#if NCNN_PLATFORM_API
-#if __ANDROID_API__ >= 9
-#include <android/bitmap.h>
-#include <jni.h>
-#endif // __ANDROID_API__ >= 9
-#endif // NCNN_PLATFORM_API
-#endif // NCNN_PIXEL
-
-namespace ncnn {
-
-#if NCNN_VULKAN
-class VkMat;
-class VkImageMat;
-#endif // NCNN_VULKAN
-
-// the three dimension matrix
-class NCNN_EXPORT Mat
-{
-public:
-    // empty
-    Mat();
-    // vec
-    Mat(int w, size_t elemsize = 4u, Allocator* allocator = 0);
-    // image
-    Mat(int w, int h, size_t elemsize = 4u, Allocator* allocator = 0);
-    // dim
-    Mat(int w, int h, int c, size_t elemsize = 4u, Allocator* allocator = 0);
-    // cube
-    Mat(int w, int h, int d, int c, size_t elemsize = 4u, Allocator* allocator = 0);
-    // packed vec
-    Mat(int w, size_t elemsize, int elempack, Allocator* allocator = 0);
-    // packed image
-    Mat(int w, int h, size_t elemsize, int elempack, Allocator* allocator = 0);
-    // packed dim
-    Mat(int w, int h, int c, size_t elemsize, int elempack, Allocator* allocator = 0);
-    // packed cube
-    Mat(int w, int h, int d, int c, size_t elemsize, int elempack, Allocator* allocator = 0);
-    // copy
-    Mat(const Mat& m);
-    // external vec
-    Mat(int w, void* data, size_t elemsize = 4u, Allocator* allocator = 0);
-    // external image
-    Mat(int w, int h, void* data, size_t elemsize = 4u, Allocator* allocator = 0);
-    // external dim
-    Mat(int w, int h, int c, void* data, size_t elemsize = 4u, Allocator* allocator = 0);
-    // external cube
-    Mat(int w, int h, int d, int c, void* data, size_t elemsize = 4u, Allocator* allocator = 0);
-    // external packed vec
-    Mat(int w, void* data, size_t elemsize, int elempack, Allocator* allocator = 0);
-    // external packed image
-    Mat(int w, int h, void* data, size_t elemsize, int elempack, Allocator* allocator = 0);
-    // external packed dim
-    Mat(int w, int h, int c, void* data, size_t elemsize, int elempack, Allocator* allocator = 0);
-    // external packed cube
-    Mat(int w, int h, int d, int c, void* data, size_t elemsize, int elempack, Allocator* allocator = 0);
-    // release
-    ~Mat();
-    // assign
-    Mat& operator=(const Mat& m);
-    // set all
-    void fill(float v);
-    void fill(int v);
-#if __ARM_NEON
-    void fill(float32x4_t _v);
-    void fill(uint16x4_t _v);
-#if !defined(_MSC_VER)
-    void fill(int32x4_t _v);
-#endif
-    void fill(int32x4_t _v0, int32x4_t _v1);
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-#if !defined(_MSC_VER)
-    void fill(float16x4_t _v);
-    void fill(float16x8_t _v);
-#endif
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-#endif // __ARM_NEON
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-    void fill(__m512 _v);
-#endif // __AVX512F__
-    void fill(__m256 _v, int i = 0);
-#endif // __AVX__
-    void fill(__m128 _v);
-    void fill(__m128i _v);
-#endif // __SSE2__
-#if __mips_msa
-    void fill(v4f32 _v);
-#endif // __mips_msa
-#if __loongarch_sx
-    void fill(__m128 _v);
-#endif //__loongarch_sx
-#if __riscv_vector
-    void fill(vfloat32m1_t _v);
-    void fill(vuint16m1_t _v);
-    void fill(vint8m1_t _v);
-#if __riscv_zfh
-    void fill(vfloat16m1_t _v);
-#endif // __riscv_zfh
-#endif // __riscv_vector
-    template<typename T>
-    void fill(T v);
-    // deep copy
-    Mat clone(Allocator* allocator = 0) const;
-    // deep copy from other mat, inplace
-    void clone_from(const ncnn::Mat& mat, Allocator* allocator = 0);
-    // reshape vec
-    Mat reshape(int w, Allocator* allocator = 0) const;
-    // reshape image
-    Mat reshape(int w, int h, Allocator* allocator = 0) const;
-    // reshape dim
-    Mat reshape(int w, int h, int c, Allocator* allocator = 0) const;
-    // reshape cube
-    Mat reshape(int w, int h, int d, int c, Allocator* allocator = 0) const;
-    // allocate vec
-    void create(int w, size_t elemsize = 4u, Allocator* allocator = 0);
-    // allocate image
-    void create(int w, int h, size_t elemsize = 4u, Allocator* allocator = 0);
-    // allocate dim
-    void create(int w, int h, int c, size_t elemsize = 4u, Allocator* allocator = 0);
-    // allocate cube
-    void create(int w, int h, int d, int c, size_t elemsize = 4u, Allocator* allocator = 0);
-    // allocate packed vec
-    void create(int w, size_t elemsize, int elempack, Allocator* allocator = 0);
-    // allocate packed image
-    void create(int w, int h, size_t elemsize, int elempack, Allocator* allocator = 0);
-    // allocate packed dim
-    void create(int w, int h, int c, size_t elemsize, int elempack, Allocator* allocator = 0);
-    // allocate packed cube
-    void create(int w, int h, int d, int c, size_t elemsize, int elempack, Allocator* allocator = 0);
-    // allocate like
-    void create_like(const Mat& m, Allocator* allocator = 0);
-#if NCNN_VULKAN
-    // allocate like
-    void create_like(const VkMat& m, Allocator* allocator = 0);
-    // allocate like
-    void create_like(const VkImageMat& im, Allocator* allocator = 0);
-#endif // NCNN_VULKAN
-    // refcount++
-    void addref();
-    // refcount--
-    void release();
-
-    bool empty() const;
-    size_t total() const;
-
-    // bits per element
-    int elembits() const;
-
-    // shape only
-    Mat shape() const;
-
-    // data reference
-    Mat channel(int c);
-    const Mat channel(int c) const;
-    Mat depth(int z);
-    const Mat depth(int z) const;
-    float* row(int y);
-    const float* row(int y) const;
-    template<typename T>
-    T* row(int y);
-    template<typename T>
-    const T* row(int y) const;
-
-    // range reference
-    Mat channel_range(int c, int channels);
-    const Mat channel_range(int c, int channels) const;
-    Mat depth_range(int z, int depths);
-    const Mat depth_range(int z, int depths) const;
-    Mat row_range(int y, int rows);
-    const Mat row_range(int y, int rows) const;
-    Mat range(int x, int n);
-    const Mat range(int x, int n) const;
-
-    // access raw data
-    template<typename T>
-    operator T*();
-    template<typename T>
-    operator const T*() const;
-
-    // convenient access float vec element
-    float& operator[](size_t i);
-    const float& operator[](size_t i) const;
-
-#if NCNN_PIXEL
-    enum PixelType
-    {
-        PIXEL_CONVERT_SHIFT = 16,
-        PIXEL_FORMAT_MASK = 0x0000ffff,
-        PIXEL_CONVERT_MASK = 0xffff0000,
-
-        PIXEL_RGB = 1,
-        PIXEL_BGR = 2,
-        PIXEL_GRAY = 3,
-        PIXEL_RGBA = 4,
-        PIXEL_BGRA = 5,
-
-        PIXEL_RGB2BGR = PIXEL_RGB | (PIXEL_BGR << PIXEL_CONVERT_SHIFT),
-        PIXEL_RGB2GRAY = PIXEL_RGB | (PIXEL_GRAY << PIXEL_CONVERT_SHIFT),
-        PIXEL_RGB2RGBA = PIXEL_RGB | (PIXEL_RGBA << PIXEL_CONVERT_SHIFT),
-        PIXEL_RGB2BGRA = PIXEL_RGB | (PIXEL_BGRA << PIXEL_CONVERT_SHIFT),
-
-        PIXEL_BGR2RGB = PIXEL_BGR | (PIXEL_RGB << PIXEL_CONVERT_SHIFT),
-        PIXEL_BGR2GRAY = PIXEL_BGR | (PIXEL_GRAY << PIXEL_CONVERT_SHIFT),
-        PIXEL_BGR2RGBA = PIXEL_BGR | (PIXEL_RGBA << PIXEL_CONVERT_SHIFT),
-        PIXEL_BGR2BGRA = PIXEL_BGR | (PIXEL_BGRA << PIXEL_CONVERT_SHIFT),
-
-        PIXEL_GRAY2RGB = PIXEL_GRAY | (PIXEL_RGB << PIXEL_CONVERT_SHIFT),
-        PIXEL_GRAY2BGR = PIXEL_GRAY | (PIXEL_BGR << PIXEL_CONVERT_SHIFT),
-        PIXEL_GRAY2RGBA = PIXEL_GRAY | (PIXEL_RGBA << PIXEL_CONVERT_SHIFT),
-        PIXEL_GRAY2BGRA = PIXEL_GRAY | (PIXEL_BGRA << PIXEL_CONVERT_SHIFT),
-
-        PIXEL_RGBA2RGB = PIXEL_RGBA | (PIXEL_RGB << PIXEL_CONVERT_SHIFT),
-        PIXEL_RGBA2BGR = PIXEL_RGBA | (PIXEL_BGR << PIXEL_CONVERT_SHIFT),
-        PIXEL_RGBA2GRAY = PIXEL_RGBA | (PIXEL_GRAY << PIXEL_CONVERT_SHIFT),
-        PIXEL_RGBA2BGRA = PIXEL_RGBA | (PIXEL_BGRA << PIXEL_CONVERT_SHIFT),
-
-        PIXEL_BGRA2RGB = PIXEL_BGRA | (PIXEL_RGB << PIXEL_CONVERT_SHIFT),
-        PIXEL_BGRA2BGR = PIXEL_BGRA | (PIXEL_BGR << PIXEL_CONVERT_SHIFT),
-        PIXEL_BGRA2GRAY = PIXEL_BGRA | (PIXEL_GRAY << PIXEL_CONVERT_SHIFT),
-        PIXEL_BGRA2RGBA = PIXEL_BGRA | (PIXEL_RGBA << PIXEL_CONVERT_SHIFT),
-    };
-    // convenient construct from pixel data
-    static Mat from_pixels(const unsigned char* pixels, int type, int w, int h, Allocator* allocator = 0);
-    // convenient construct from pixel data with stride(bytes-per-row) parameter
-    static Mat from_pixels(const unsigned char* pixels, int type, int w, int h, int stride, Allocator* allocator = 0);
-    // convenient construct from pixel data and resize to specific size
-    static Mat from_pixels_resize(const unsigned char* pixels, int type, int w, int h, int target_width, int target_height, Allocator* allocator = 0);
-    // convenient construct from pixel data and resize to specific size with stride(bytes-per-row) parameter
-    static Mat from_pixels_resize(const unsigned char* pixels, int type, int w, int h, int stride, int target_width, int target_height, Allocator* allocator = 0);
-    // convenient construct from pixel data roi
-    static Mat from_pixels_roi(const unsigned char* pixels, int type, int w, int h, int roix, int roiy, int roiw, int roih, Allocator* allocator = 0);
-    // convenient construct from pixel data roi with stride(bytes-per-row) parameter
-    static Mat from_pixels_roi(const unsigned char* pixels, int type, int w, int h, int stride, int roix, int roiy, int roiw, int roih, Allocator* allocator = 0);
-    // convenient construct from pixel data roi and resize to specific size
-    static Mat from_pixels_roi_resize(const unsigned char* pixels, int type, int w, int h, int roix, int roiy, int roiw, int roih, int target_width, int target_height, Allocator* allocator = 0);
-    // convenient construct from pixel data roi and resize to specific size with stride(bytes-per-row) parameter
-    static Mat from_pixels_roi_resize(const unsigned char* pixels, int type, int w, int h, int stride, int roix, int roiy, int roiw, int roih, int target_width, int target_height, Allocator* allocator = 0);
-
-    // convenient export to pixel data
-    void to_pixels(unsigned char* pixels, int type) const;
-    // convenient export to pixel data with stride(bytes-per-row) parameter
-    void to_pixels(unsigned char* pixels, int type, int stride) const;
-    // convenient export to pixel data and resize to specific size
-    void to_pixels_resize(unsigned char* pixels, int type, int target_width, int target_height) const;
-    // convenient export to pixel data and resize to specific size with stride(bytes-per-row) parameter
-    void to_pixels_resize(unsigned char* pixels, int type, int target_width, int target_height, int target_stride) const;
-
-#if NCNN_PLATFORM_API
-#if __ANDROID_API__ >= 9
-    // convenient construct from android Bitmap
-    static Mat from_android_bitmap(JNIEnv* env, jobject bitmap, int type_to, Allocator* allocator = 0);
-    // convenient construct from android Bitmap and resize to specific size
-    static Mat from_android_bitmap_resize(JNIEnv* env, jobject bitmap, int type_to, int target_width, int target_height, Allocator* allocator = 0);
-    // convenient construct from android Bitmap roi
-    static Mat from_android_bitmap_roi(JNIEnv* env, jobject bitmap, int type_to, int roix, int roiy, int roiw, int roih, Allocator* allocator = 0);
-    // convenient construct from android Bitmap roi and resize to specific size
-    static Mat from_android_bitmap_roi_resize(JNIEnv* env, jobject bitmap, int type_to, int roix, int roiy, int roiw, int roih, int target_width, int target_height, Allocator* allocator = 0);
-    // convenient export to android Bitmap and resize to the android Bitmap size
-    void to_android_bitmap(JNIEnv* env, jobject bitmap, int type_from) const;
-#endif // __ANDROID_API__ >= 9
-#endif // NCNN_PLATFORM_API
-#endif // NCNN_PIXEL
-
-    // substract channel-wise mean values, then multiply by normalize values, pass 0 to skip
-    void substract_mean_normalize(const float* mean_vals, const float* norm_vals);
-
-    // convenient construct from half precision floating point data
-    static Mat from_float16(const unsigned short* data, int size);
-
-    // pointer to the data
-    void* data;
-
-    // pointer to the reference counter
-    // when points to user-allocated data, the pointer is NULL
-    int* refcount;
-
-    // element size in bytes
-    // 4 = float32/int32
-    // 2 = float16
-    // 1 = int8/uint8
-    // 0 = empty
-    size_t elemsize;
-
-    // packed count inside element
-    // c/1-d-h-w-1  c/1-h-w-1  h/1-w-1  w/1-1  scalar
-    // c/4-d-h-w-4  c/4-h-w-4  h/4-w-4  w/4-4  sse/neon
-    // c/8-d-h-w-8  c/8-h-w-8  h/8-w-8  w/8-8  avx/fp16
-    int elempack;
-
-    // the allocator
-    Allocator* allocator;
-
-    // the dimension rank
-    int dims;
-
-    int w;
-    int h;
-    int d;
-    int c;
-
-    size_t cstep;
-};
-
-#if NCNN_VULKAN
-
-// the three dimension matrix, vulkan version
-class NCNN_EXPORT VkMat
-{
-public:
-    // empty
-    VkMat();
-    // vec
-    VkMat(int w, size_t elemsize, VkAllocator* allocator);
-    // image
-    VkMat(int w, int h, size_t elemsize, VkAllocator* allocator);
-    // dim
-    VkMat(int w, int h, int c, size_t elemsize, VkAllocator* allocator);
-    // cube
-    VkMat(int w, int h, int d, int c, size_t elemsize, VkAllocator* allocator);
-    // packed vec
-    VkMat(int w, size_t elemsize, int elempack, VkAllocator* allocator);
-    // packed image
-    VkMat(int w, int h, size_t elemsize, int elempack, VkAllocator* allocator);
-    // packed dim
-    VkMat(int w, int h, int c, size_t elemsize, int elempack, VkAllocator* allocator);
-    // packed cube
-    VkMat(int w, int h, int d, int c, size_t elemsize, int elempack, VkAllocator* allocator);
-    // copy
-    VkMat(const VkMat& m);
-    // external vec
-    VkMat(int w, VkBufferMemory* data, size_t elemsize, VkAllocator* allocator);
-    // external image
-    VkMat(int w, int h, VkBufferMemory* data, size_t elemsize, VkAllocator* allocator);
-    // external dim
-    VkMat(int w, int h, int c, VkBufferMemory* data, size_t elemsize, VkAllocator* allocator);
-    // external cube
-    VkMat(int w, int h, int d, int c, VkBufferMemory* data, size_t elemsize, VkAllocator* allocator);
-    // external packed vec
-    VkMat(int w, VkBufferMemory* data, size_t elemsize, int elempack, VkAllocator* allocator);
-    // external packed image
-    VkMat(int w, int h, VkBufferMemory* data, size_t elemsize, int elempack, VkAllocator* allocator);
-    // external packed dim
-    VkMat(int w, int h, int c, VkBufferMemory* data, size_t elemsize, int elempack, VkAllocator* allocator);
-    // external packed cube
-    VkMat(int w, int h, int d, int c, VkBufferMemory* data, size_t elemsize, int elempack, VkAllocator* allocator);
-    // release
-    ~VkMat();
-    // assign
-    VkMat& operator=(const VkMat& m);
-    // allocate vec
-    void create(int w, size_t elemsize, VkAllocator* allocator);
-    // allocate image
-    void create(int w, int h, size_t elemsize, VkAllocator* allocator);
-    // allocate dim
-    void create(int w, int h, int c, size_t elemsize, VkAllocator* allocator);
-    // allocate cube
-    void create(int w, int h, int d, int c, size_t elemsize, VkAllocator* allocator);
-    // allocate packed vec
-    void create(int w, size_t elemsize, int elempack, VkAllocator* allocator);
-    // allocate packed image
-    void create(int w, int h, size_t elemsize, int elempack, VkAllocator* allocator);
-    // allocate packed dim
-    void create(int w, int h, int c, size_t elemsize, int elempack, VkAllocator* allocator);
-    // allocate packed cube
-    void create(int w, int h, int d, int c, size_t elemsize, int elempack, VkAllocator* allocator);
-    // allocate like
-    void create_like(const Mat& m, VkAllocator* allocator);
-    // allocate like
-    void create_like(const VkMat& m, VkAllocator* allocator);
-    // allocate like
-    void create_like(const VkImageMat& im, VkAllocator* allocator);
-
-    // mapped
-    Mat mapped() const;
-    void* mapped_ptr() const;
-
-    // refcount++
-    void addref();
-    // refcount--
-    void release();
-
-    bool empty() const;
-    size_t total() const;
-
-    // bits per element
-    int elembits() const;
-
-    // shape only
-    Mat shape() const;
-
-    // low-level reference
-    VkBuffer buffer() const;
-    size_t buffer_offset() const;
-    size_t buffer_capacity() const;
-
-    // device buffer
-    VkBufferMemory* data;
-
-    // pointer to the reference counter
-    // when points to user-allocated data, the pointer is NULL
-    int* refcount;
-
-    // element size in bytes
-    // 4 = float32/int32
-    // 2 = float16
-    // 1 = int8/uint8
-    // 0 = empty
-    size_t elemsize;
-
-    // packed count inside element
-    // c/1-d-h-w-1  c/1-h-w-1  h/1-w-1  w/1-1  scalar
-    // c/4-d-h-w-4  c/4-h-w-4  h/4-w-4  w/4-4  sse/neon
-    // c/8-d-h-w-8  c/8-h-w-8  h/8-w-8  w/8-8  avx/fp16
-    int elempack;
-
-    // the allocator
-    VkAllocator* allocator;
-
-    // the dimension rank
-    int dims;
-
-    int w;
-    int h;
-    int d;
-    int c;
-
-    size_t cstep;
-};
-
-class NCNN_EXPORT VkImageMat
-{
-public:
-    // empty
-    VkImageMat();
-    // vec
-    VkImageMat(int w, size_t elemsize, VkAllocator* allocator);
-    // image
-    VkImageMat(int w, int h, size_t elemsize, VkAllocator* allocator);
-    // dim
-    VkImageMat(int w, int h, int c, size_t elemsize, VkAllocator* allocator);
-    // cube
-    VkImageMat(int w, int h, int d, int c, size_t elemsize, VkAllocator* allocator);
-    // packed vec
-    VkImageMat(int w, size_t elemsize, int elempack, VkAllocator* allocator);
-    // packed image
-    VkImageMat(int w, int h, size_t elemsize, int elempack, VkAllocator* allocator);
-    // packed dim
-    VkImageMat(int w, int h, int c, size_t elemsize, int elempack, VkAllocator* allocator);
-    // packed cube
-    VkImageMat(int w, int h, int d, int c, size_t elemsize, int elempack, VkAllocator* allocator);
-    // copy
-    VkImageMat(const VkImageMat& m);
-    // external vec
-    VkImageMat(int w, VkImageMemory* data, size_t elemsize, VkAllocator* allocator);
-    // external image
-    VkImageMat(int w, int h, VkImageMemory* data, size_t elemsize, VkAllocator* allocator);
-    // external dim
-    VkImageMat(int w, int h, int c, VkImageMemory* data, size_t elemsize, VkAllocator* allocator);
-    // external cube
-    VkImageMat(int w, int h, int d, int c, VkImageMemory* data, size_t elemsize, VkAllocator* allocator);
-    // external packed vec
-    VkImageMat(int w, VkImageMemory* data, size_t elemsize, int elempack, VkAllocator* allocator);
-    // external packed image
-    VkImageMat(int w, int h, VkImageMemory* data, size_t elemsize, int elempack, VkAllocator* allocator);
-    // external packed dim
-    VkImageMat(int w, int h, int c, VkImageMemory* data, size_t elemsize, int elempack, VkAllocator* allocator);
-    // external packed cube
-    VkImageMat(int w, int h, int d, int c, VkImageMemory* data, size_t elemsize, int elempack, VkAllocator* allocator);
-    // release
-    ~VkImageMat();
-    // assign
-    VkImageMat& operator=(const VkImageMat& m);
-    // allocate vec
-    void create(int w, size_t elemsize, VkAllocator* allocator);
-    // allocate image
-    void create(int w, int h, size_t elemsize, VkAllocator* allocator);
-    // allocate dim
-    void create(int w, int h, int c, size_t elemsize, VkAllocator* allocator);
-    // allocate cube
-    void create(int w, int h, int d, int c, size_t elemsize, VkAllocator* allocator);
-    // allocate packed vec
-    void create(int w, size_t elemsize, int elempack, VkAllocator* allocator);
-    // allocate packed image
-    void create(int w, int h, size_t elemsize, int elempack, VkAllocator* allocator);
-    // allocate packed dim
-    void create(int w, int h, int c, size_t elemsize, int elempack, VkAllocator* allocator);
-    // allocate packed cube
-    void create(int w, int h, int d, int c, size_t elemsize, int elempack, VkAllocator* allocator);
-    // allocate like
-    void create_like(const Mat& m, VkAllocator* allocator);
-    // allocate like
-    void create_like(const VkMat& m, VkAllocator* allocator);
-    // allocate like
-    void create_like(const VkImageMat& im, VkAllocator* allocator);
-
-    // mapped
-    Mat mapped() const;
-    void* mapped_ptr() const;
-
-    // refcount++
-    void addref();
-    // refcount--
-    void release();
-
-    bool empty() const;
-    size_t total() const;
-
-    // bits per element
-    int elembits() const;
-
-    // shape only
-    Mat shape() const;
-
-    // low-level reference
-    VkImage image() const;
-    VkImageView imageview() const;
-
-#if NCNN_PLATFORM_API
-#if __ANDROID_API__ >= 26
-    // convenient construct from android hardware buffer
-    static VkImageMat from_android_hardware_buffer(VkAndroidHardwareBufferImageAllocator* allocator);
-#endif // __ANDROID_API__ >= 26
-#endif // NCNN_PLATFORM_API
-
-    // device image
-    VkImageMemory* data;
-
-    // pointer to the reference counter
-    // when points to user-allocated data, the pointer is NULL
-    int* refcount;
-
-    // element size in bytes
-    // 4 = float32/int32
-    // 2 = float16
-    // 1 = int8/uint8
-    // 0 = empty
-    size_t elemsize;
-
-    // packed count inside element
-    // c/1-d-h-w-1  c/1-h-w-1  h/1-w-1  w/1-1  scalar
-    // c/4-d-h-w-4  c/4-h-w-4  h/4-w-4  w/4-4  sse/neon
-    // c/8-d-h-w-8  c/8-h-w-8  h/8-w-8  w/8-8  avx/fp16
-    int elempack;
-
-    // the allocator
-    VkAllocator* allocator;
-
-    // the dimension rank
-    int dims;
-
-    int w;
-    int h;
-    int d;
-    int c;
-};
-
-// type for vulkan specialization constant and push constant
-union vk_specialization_type
-{
-    int i;
-    float f;
-    uint32_t u32;
-};
-union vk_constant_type
-{
-    int i;
-    float f;
-};
-#endif // NCNN_VULKAN
-
-// misc function
-#if NCNN_PIXEL
-// convert yuv420sp(nv21) to rgb, the fast approximate version
-NCNN_EXPORT void yuv420sp2rgb(const unsigned char* yuv420sp, int w, int h, unsigned char* rgb);
-// convert yuv420sp(nv12) to rgb, the fast approximate version
-NCNN_EXPORT void yuv420sp2rgb_nv12(const unsigned char* yuv420sp, int w, int h, unsigned char* rgb);
-// convert yuv420sp(nv21) to rgb with half resize, the faster approximate version
-NCNN_EXPORT void yuv420sp2rgb_half(const unsigned char* yuv420sp, int w, int h, unsigned char* rgb);
-// image pixel bilinear resize
-NCNN_EXPORT void resize_bilinear_c1(const unsigned char* src, int srcw, int srch, unsigned char* dst, int w, int h);
-NCNN_EXPORT void resize_bilinear_c2(const unsigned char* src, int srcw, int srch, unsigned char* dst, int w, int h);
-NCNN_EXPORT void resize_bilinear_c3(const unsigned char* src, int srcw, int srch, unsigned char* dst, int w, int h);
-NCNN_EXPORT void resize_bilinear_c4(const unsigned char* src, int srcw, int srch, unsigned char* dst, int w, int h);
-// image pixel bilinear resize with stride(bytes-per-row) parameter
-NCNN_EXPORT void resize_bilinear_c1(const unsigned char* src, int srcw, int srch, int srcstride, unsigned char* dst, int w, int h, int stride);
-NCNN_EXPORT void resize_bilinear_c2(const unsigned char* src, int srcw, int srch, int srcstride, unsigned char* dst, int w, int h, int stride);
-NCNN_EXPORT void resize_bilinear_c3(const unsigned char* src, int srcw, int srch, int srcstride, unsigned char* dst, int w, int h, int stride);
-NCNN_EXPORT void resize_bilinear_c4(const unsigned char* src, int srcw, int srch, int srcstride, unsigned char* dst, int w, int h, int stride);
-// image pixel bilinear resize, convenient wrapper for yuv420sp(nv21/nv12)
-NCNN_EXPORT void resize_bilinear_yuv420sp(const unsigned char* src, int srcw, int srch, unsigned char* dst, int w, int h);
-#endif // NCNN_PIXEL
-#if NCNN_PIXEL_ROTATE
-// type is the from type, 6 means rotating from 6 to 1
-//
-//     1        2       3      4         5            6           7          8
-//
-//   888888  888888      88  88      8888888888  88                  88  8888888888
-//   88          88      88  88      88  88      88  88          88  88      88  88
-//   8888      8888    8888  8888    88          8888888888  8888888888          88
-//   88          88      88  88
-//   88          88  888888  888888
-//
-// ref http://sylvana.net/jpegcrop/exif_orientation.html
-// image pixel kanna rotate
-NCNN_EXPORT void kanna_rotate_c1(const unsigned char* src, int srcw, int srch, unsigned char* dst, int w, int h, int type);
-NCNN_EXPORT void kanna_rotate_c2(const unsigned char* src, int srcw, int srch, unsigned char* dst, int w, int h, int type);
-NCNN_EXPORT void kanna_rotate_c3(const unsigned char* src, int srcw, int srch, unsigned char* dst, int w, int h, int type);
-NCNN_EXPORT void kanna_rotate_c4(const unsigned char* src, int srcw, int srch, unsigned char* dst, int w, int h, int type);
-// image pixel kanna rotate with stride(bytes-per-row) parameter
-NCNN_EXPORT void kanna_rotate_c1(const unsigned char* src, int srcw, int srch, int srcstride, unsigned char* dst, int w, int h, int stride, int type);
-NCNN_EXPORT void kanna_rotate_c2(const unsigned char* src, int srcw, int srch, int srcstride, unsigned char* dst, int w, int h, int stride, int type);
-NCNN_EXPORT void kanna_rotate_c3(const unsigned char* src, int srcw, int srch, int srcstride, unsigned char* dst, int w, int h, int stride, int type);
-NCNN_EXPORT void kanna_rotate_c4(const unsigned char* src, int srcw, int srch, int srcstride, unsigned char* dst, int w, int h, int stride, int type);
-// image pixel kanna rotate, convenient wrapper for yuv420sp(nv21/nv12)
-NCNN_EXPORT void kanna_rotate_yuv420sp(const unsigned char* src, int srcw, int srch, unsigned char* dst, int w, int h, int type);
-#endif // NCNN_PIXEL_ROTATE
-#if NCNN_PIXEL_AFFINE
-// resolve affine transform matrix from rotation angle, scale factor and x y offset
-NCNN_EXPORT void get_rotation_matrix(float angle, float scale, float dx, float dy, float* tm);
-// resolve affine transform matrix from two set of points, num_point must be >= 2
-NCNN_EXPORT void get_affine_transform(const float* points_from, const float* points_to, int num_point, float* tm);
-// resolve the inversion affine transform matrix
-NCNN_EXPORT void invert_affine_transform(const float* tm, float* tm_inv);
-// image pixel bilinear warpaffine inverse transform, set -233 for transparent border color, the color RGBA is little-endian encoded
-NCNN_EXPORT void warpaffine_bilinear_c1(const unsigned char* src, int srcw, int srch, unsigned char* dst, int w, int h, const float* tm, int type = 0, unsigned int v = 0);
-NCNN_EXPORT void warpaffine_bilinear_c2(const unsigned char* src, int srcw, int srch, unsigned char* dst, int w, int h, const float* tm, int type = 0, unsigned int v = 0);
-NCNN_EXPORT void warpaffine_bilinear_c3(const unsigned char* src, int srcw, int srch, unsigned char* dst, int w, int h, const float* tm, int type = 0, unsigned int v = 0);
-NCNN_EXPORT void warpaffine_bilinear_c4(const unsigned char* src, int srcw, int srch, unsigned char* dst, int w, int h, const float* tm, int type = 0, unsigned int v = 0);
-// image pixel bilinear warpaffine inverse transform with stride(bytes-per-row) parameter, set -233 for transparent border color, the color RGBA is little-endian encoded
-NCNN_EXPORT void warpaffine_bilinear_c1(const unsigned char* src, int srcw, int srch, int srcstride, unsigned char* dst, int w, int h, int stride, const float* tm, int type = 0, unsigned int v = 0);
-NCNN_EXPORT void warpaffine_bilinear_c2(const unsigned char* src, int srcw, int srch, int srcstride, unsigned char* dst, int w, int h, int stride, const float* tm, int type = 0, unsigned int v = 0);
-NCNN_EXPORT void warpaffine_bilinear_c3(const unsigned char* src, int srcw, int srch, int srcstride, unsigned char* dst, int w, int h, int stride, const float* tm, int type = 0, unsigned int v = 0);
-NCNN_EXPORT void warpaffine_bilinear_c4(const unsigned char* src, int srcw, int srch, int srcstride, unsigned char* dst, int w, int h, int stride, const float* tm, int type = 0, unsigned int v = 0);
-// image pixel bilinear warpaffine, convenient wrapper for yuv420sp(nv21/nv12), set -233 for transparent border color, the color YUV_ is little-endian encoded
-NCNN_EXPORT void warpaffine_bilinear_yuv420sp(const unsigned char* src, int srcw, int srch, unsigned char* dst, int w, int h, const float* tm, int type = 0, unsigned int v = 0);
-#endif // NCNN_PIXEL_AFFINE
-#if NCNN_PIXEL_DRAWING
-// draw rectangle, set thickness -1 for filled rectangle, the color RGBA is little-endian encoded
-NCNN_EXPORT void draw_rectangle_c1(unsigned char* pixels, int w, int h, int rx, int ry, int rw, int rh, unsigned int color, int thickness);
-NCNN_EXPORT void draw_rectangle_c2(unsigned char* pixels, int w, int h, int rx, int ry, int rw, int rh, unsigned int color, int thickness);
-NCNN_EXPORT void draw_rectangle_c3(unsigned char* pixels, int w, int h, int rx, int ry, int rw, int rh, unsigned int color, int thickness);
-NCNN_EXPORT void draw_rectangle_c4(unsigned char* pixels, int w, int h, int rx, int ry, int rw, int rh, unsigned int color, int thickness);
-// draw rectangle with stride(bytes-per-row) parameter, set thickness -1 for filled rectangle, the color RGBA is little-endian encoded
-NCNN_EXPORT void draw_rectangle_c1(unsigned char* pixels, int w, int h, int stride, int rx, int ry, int rw, int rh, unsigned int color, int thickness);
-NCNN_EXPORT void draw_rectangle_c2(unsigned char* pixels, int w, int h, int stride, int rx, int ry, int rw, int rh, unsigned int color, int thickness);
-NCNN_EXPORT void draw_rectangle_c3(unsigned char* pixels, int w, int h, int stride, int rx, int ry, int rw, int rh, unsigned int color, int thickness);
-NCNN_EXPORT void draw_rectangle_c4(unsigned char* pixels, int w, int h, int stride, int rx, int ry, int rw, int rh, unsigned int color, int thickness);
-// draw rectangle, convenient wrapper for yuv420sp(nv21/nv12), set thickness -1 for filled rectangle, the color YUV_ is little-endian encoded
-NCNN_EXPORT void draw_rectangle_yuv420sp(unsigned char* yuv420sp, int w, int h, int rx, int ry, int rw, int rh, unsigned int color, int thickness);
-// draw circle, set thickness -1 for filled circle, the color RGBA is little-endian encoded
-NCNN_EXPORT void draw_circle_c1(unsigned char* pixels, int w, int h, int cx, int cy, int radius, unsigned int color, int thickness);
-NCNN_EXPORT void draw_circle_c2(unsigned char* pixels, int w, int h, int cx, int cy, int radius, unsigned int color, int thickness);
-NCNN_EXPORT void draw_circle_c3(unsigned char* pixels, int w, int h, int cx, int cy, int radius, unsigned int color, int thickness);
-NCNN_EXPORT void draw_circle_c4(unsigned char* pixels, int w, int h, int cx, int cy, int radius, unsigned int color, int thickness);
-// draw circle with stride(bytes-per-row) parameter, set thickness -1 for filled circle, the color RGBA is little-endian encoded
-NCNN_EXPORT void draw_circle_c1(unsigned char* pixels, int w, int h, int stride, int cx, int cy, int radius, unsigned int color, int thickness);
-NCNN_EXPORT void draw_circle_c2(unsigned char* pixels, int w, int h, int stride, int cx, int cy, int radius, unsigned int color, int thickness);
-NCNN_EXPORT void draw_circle_c3(unsigned char* pixels, int w, int h, int stride, int cx, int cy, int radius, unsigned int color, int thickness);
-NCNN_EXPORT void draw_circle_c4(unsigned char* pixels, int w, int h, int stride, int cx, int cy, int radius, unsigned int color, int thickness);
-// draw circle, convenient wrapper for yuv420sp(nv21/nv12), set thickness -1 for filled circle, the color YUV_ is little-endian encoded
-NCNN_EXPORT void draw_circle_yuv420sp(unsigned char* yuv420sp, int w, int h, int cx, int cy, int radius, unsigned int color, int thickness);
-// draw line, the color RGBA is little-endian encoded
-NCNN_EXPORT void draw_line_c1(unsigned char* pixels, int w, int h, int x0, int y0, int x1, int y1, unsigned int color, int thickness);
-NCNN_EXPORT void draw_line_c2(unsigned char* pixels, int w, int h, int x0, int y0, int x1, int y1, unsigned int color, int thickness);
-NCNN_EXPORT void draw_line_c3(unsigned char* pixels, int w, int h, int x0, int y0, int x1, int y1, unsigned int color, int thickness);
-NCNN_EXPORT void draw_line_c4(unsigned char* pixels, int w, int h, int x0, int y0, int x1, int y1, unsigned int color, int thickness);
-// draw line with stride(bytes-per-row) parameter, the color RGBA is little-endian encoded
-NCNN_EXPORT void draw_line_c1(unsigned char* pixels, int w, int h, int stride, int x0, int y0, int x1, int y1, unsigned int color, int thickness);
-NCNN_EXPORT void draw_line_c2(unsigned char* pixels, int w, int h, int stride, int x0, int y0, int x1, int y1, unsigned int color, int thickness);
-NCNN_EXPORT void draw_line_c3(unsigned char* pixels, int w, int h, int stride, int x0, int y0, int x1, int y1, unsigned int color, int thickness);
-NCNN_EXPORT void draw_line_c4(unsigned char* pixels, int w, int h, int stride, int x0, int y0, int x1, int y1, unsigned int color, int thickness);
-// draw line, convenient wrapper for yuv420sp(nv21/nv12), the color YUV_ is little-endian encoded
-NCNN_EXPORT void draw_line_yuv420sp(unsigned char* yuv420sp, int w, int h, int x0, int y0, int x1, int y1, unsigned int color, int thickness);
-// resolve text bounding box size
-NCNN_EXPORT void get_text_drawing_size(const char* text, int fontpixelsize, int* w, int* h);
-// draw ascii printables and newline, the color RGBA is little-endian encoded
-NCNN_EXPORT void draw_text_c1(unsigned char* pixels, int w, int h, const char* text, int x, int y, int fontpixelsize, unsigned int color);
-NCNN_EXPORT void draw_text_c2(unsigned char* pixels, int w, int h, const char* text, int x, int y, int fontpixelsize, unsigned int color);
-NCNN_EXPORT void draw_text_c3(unsigned char* pixels, int w, int h, const char* text, int x, int y, int fontpixelsize, unsigned int color);
-NCNN_EXPORT void draw_text_c4(unsigned char* pixels, int w, int h, const char* text, int x, int y, int fontpixelsize, unsigned int color);
-// draw ascii printables and newline with stride(bytes-per-row) parameter, the color RGBA is little-endian encoded
-NCNN_EXPORT void draw_text_c1(unsigned char* pixels, int w, int h, int stride, const char* text, int x, int y, int fontpixelsize, unsigned int color);
-NCNN_EXPORT void draw_text_c2(unsigned char* pixels, int w, int h, int stride, const char* text, int x, int y, int fontpixelsize, unsigned int color);
-NCNN_EXPORT void draw_text_c3(unsigned char* pixels, int w, int h, int stride, const char* text, int x, int y, int fontpixelsize, unsigned int color);
-NCNN_EXPORT void draw_text_c4(unsigned char* pixels, int w, int h, int stride, const char* text, int x, int y, int fontpixelsize, unsigned int color);
-// draw ascii printables and newline, convenient wrapper for yuv420sp(nv21/nv12), the color YUV_ is little-endian encoded
-NCNN_EXPORT void draw_text_yuv420sp(unsigned char* yuv420sp, int w, int h, const char* text, int x, int y, int fontpixelsize, unsigned int color);
-#endif // NCNN_PIXEL_DRAWING
-
-// type conversion
-// convert float to half precision floating point
-NCNN_EXPORT unsigned short float32_to_float16(float value);
-// convert half precision floating point to float
-NCNN_EXPORT float float16_to_float32(unsigned short value);
-// convert float to brain half
-NCNN_EXPORT NCNN_FORCEINLINE unsigned short float32_to_bfloat16(float value)
-{
-    // 16 : 16
-    union
-    {
-        unsigned int u;
-        float f;
-    } tmp;
-    tmp.f = value;
-    return tmp.u >> 16;
-}
-// convert brain half to float
-NCNN_EXPORT NCNN_FORCEINLINE float bfloat16_to_float32(unsigned short value)
-{
-    // 16 : 16
-    union
-    {
-        unsigned int u;
-        float f;
-    } tmp;
-    tmp.u = value << 16;
-    return tmp.f;
-}
-
-// mat process
-enum BorderType
-{
-    BORDER_CONSTANT = 0,
-    BORDER_REPLICATE = 1,
-    BORDER_REFLECT = 2,
-    BORDER_TRANSPARENT = -233,
-};
-NCNN_EXPORT void copy_make_border(const Mat& src, Mat& dst, int top, int bottom, int left, int right, int type, float v, const Option& opt = Option());
-NCNN_EXPORT void copy_make_border_3d(const Mat& src, Mat& dst, int top, int bottom, int left, int right, int front, int behind, int type, float v, const Option& opt = Option());
-NCNN_EXPORT void copy_cut_border(const Mat& src, Mat& dst, int top, int bottom, int left, int right, const Option& opt = Option());
-NCNN_EXPORT void copy_cut_border_3d(const Mat& src, Mat& dst, int top, int bottom, int left, int right, int front, int behind, const Option& opt = Option());
-NCNN_EXPORT void resize_nearest(const Mat& src, Mat& dst, int w, int h, const Option& opt = Option());
-NCNN_EXPORT void resize_bilinear(const Mat& src, Mat& dst, int w, int h, const Option& opt = Option());
-NCNN_EXPORT void resize_bicubic(const Mat& src, Mat& dst, int w, int h, const Option& opt = Option());
-NCNN_EXPORT void convert_packing(const Mat& src, Mat& dst, int elempack, const Option& opt = Option());
-NCNN_EXPORT void flatten(const Mat& src, Mat& dst, const Option& opt = Option());
-NCNN_EXPORT void cast_float32_to_float16(const Mat& src, Mat& dst, const Option& opt = Option());
-NCNN_EXPORT void cast_float16_to_float32(const Mat& src, Mat& dst, const Option& opt = Option());
-NCNN_EXPORT void cast_int8_to_float32(const Mat& src, Mat& dst, const Option& opt = Option());
-NCNN_EXPORT void cast_float32_to_bfloat16(const Mat& src, Mat& dst, const Option& opt = Option());
-NCNN_EXPORT void cast_bfloat16_to_float32(const Mat& src, Mat& dst, const Option& opt = Option());
-NCNN_EXPORT void quantize_to_int8(const Mat& src, Mat& dst, const Mat& scale_data, const Option& opt = Option());
-NCNN_EXPORT void dequantize_from_int32(const Mat& src, Mat& dst, const Mat& scale_data, const Mat& bias_data, const Option& opt = Option());
-NCNN_EXPORT void requantize_from_int32_to_int8(const Mat& src, Mat& dst, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, int activation_type, const Mat& activation_params, const Option& opt = Option());
-
-NCNN_FORCEINLINE Mat::Mat()
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, size_t _elemsize, Allocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _elemsize, _allocator);
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, int _h, size_t _elemsize, Allocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _h, _elemsize, _allocator);
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, int _h, int _c, size_t _elemsize, Allocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _h, _c, _elemsize, _allocator);
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, int _h, int _d, int _c, size_t _elemsize, Allocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _h, _d, _c, _elemsize, _allocator);
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, size_t _elemsize, int _elempack, Allocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _elemsize, _elempack, _allocator);
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, int _h, size_t _elemsize, int _elempack, Allocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _h, _elemsize, _elempack, _allocator);
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, int _h, int _c, size_t _elemsize, int _elempack, Allocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _h, _c, _elemsize, _elempack, _allocator);
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, int _h, int _d, int _c, size_t _elemsize, int _elempack, Allocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _h, _d, _c, _elemsize, _elempack, _allocator);
-}
-
-NCNN_FORCEINLINE Mat::Mat(const Mat& m)
-    : data(m.data), refcount(m.refcount), elemsize(m.elemsize), elempack(m.elempack), allocator(m.allocator), dims(m.dims), w(m.w), h(m.h), d(m.d), c(m.c), cstep(m.cstep)
-{
-    addref();
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, void* _data, size_t _elemsize, Allocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(1), allocator(_allocator), dims(1), w(_w), h(1), d(1), c(1)
-{
-    cstep = w;
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, int _h, void* _data, size_t _elemsize, Allocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(1), allocator(_allocator), dims(2), w(_w), h(_h), d(1), c(1)
-{
-    cstep = (size_t)w * h;
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, int _h, int _c, void* _data, size_t _elemsize, Allocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(1), allocator(_allocator), dims(3), w(_w), h(_h), d(1), c(_c)
-{
-    cstep = alignSize((size_t)w * h * elemsize, 16) / elemsize;
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, int _h, int _d, int _c, void* _data, size_t _elemsize, Allocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(1), allocator(_allocator), dims(4), w(_w), h(_h), d(_d), c(_c)
-{
-    cstep = alignSize((size_t)w * h * d * elemsize, 16) / elemsize;
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, void* _data, size_t _elemsize, int _elempack, Allocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(_elempack), allocator(_allocator), dims(1), w(_w), h(1), d(1), c(1)
-{
-    cstep = w;
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, int _h, void* _data, size_t _elemsize, int _elempack, Allocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(_elempack), allocator(_allocator), dims(2), w(_w), h(_h), d(1), c(1)
-{
-    cstep = (size_t)w * h;
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, int _h, int _c, void* _data, size_t _elemsize, int _elempack, Allocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(_elempack), allocator(_allocator), dims(3), w(_w), h(_h), d(1), c(_c)
-{
-    cstep = alignSize((size_t)w * h * elemsize, 16) / elemsize;
-}
-
-NCNN_FORCEINLINE Mat::Mat(int _w, int _h, int _d, int _c, void* _data, size_t _elemsize, int _elempack, Allocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(_elempack), allocator(_allocator), dims(4), w(_w), h(_h), d(_d), c(_c)
-{
-    cstep = alignSize((size_t)w * h * d * elemsize, 16) / elemsize;
-}
-
-NCNN_FORCEINLINE Mat::~Mat()
-{
-    release();
-}
-
-NCNN_FORCEINLINE void Mat::fill(float _v)
-{
-    int size = (int)total();
-    float* ptr = (float*)data;
-
-    int i = 0;
-#if __ARM_NEON
-    float32x4_t _c = vdupq_n_f32(_v);
-    for (; i + 3 < size; i += 4)
-    {
-        vst1q_f32(ptr, _c);
-        ptr += 4;
-    }
-#endif // __ARM_NEON
-    for (; i < size; i++)
-    {
-        *ptr++ = _v;
-    }
-}
-
-NCNN_FORCEINLINE void Mat::fill(int _v)
-{
-    int size = (int)total();
-    int* ptr = (int*)data;
-
-    int i = 0;
-#if __ARM_NEON
-    int32x4_t _c = vdupq_n_s32(_v);
-    for (; i + 3 < size; i += 4)
-    {
-        vst1q_s32(ptr, _c);
-        ptr += 4;
-    }
-#endif // __ARM_NEON
-    for (; i < size; i++)
-    {
-        *ptr++ = _v;
-    }
-}
-
-#if __ARM_NEON
-NCNN_FORCEINLINE void Mat::fill(float32x4_t _v)
-{
-    int size = (int)total();
-    float* ptr = (float*)data;
-    for (int i = 0; i < size; i++)
-    {
-        vst1q_f32(ptr, _v);
-        ptr += 4;
-    }
-}
-
-NCNN_FORCEINLINE void Mat::fill(uint16x4_t _v)
-{
-    int size = (int)total();
-    unsigned short* ptr = (unsigned short*)data;
-    for (int i = 0; i < size; i++)
-    {
-        vst1_u16(ptr, _v);
-        ptr += 4;
-    }
-}
-
-#if !defined(_MSC_VER)
-NCNN_FORCEINLINE void Mat::fill(int32x4_t _v)
-{
-    int size = (int)total();
-    int* ptr = (int*)data;
-    for (int i = 0; i < size; i++)
-    {
-        vst1q_s32(ptr, _v);
-        ptr += 4;
-    }
-}
-#endif
-
-NCNN_FORCEINLINE void Mat::fill(int32x4_t _v0, int32x4_t _v1)
-{
-    int size = (int)total();
-    int* ptr = (int*)data;
-    for (int i = 0; i < size; i++)
-    {
-        vst1q_s32(ptr, _v0);
-        vst1q_s32(ptr + 4, _v1);
-        ptr += 8;
-    }
-}
-#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-#if !defined(_MSC_VER)
-NCNN_FORCEINLINE void Mat::fill(float16x4_t _v)
-{
-    int size = (int)total();
-    __fp16* ptr = (__fp16*)data;
-    for (int i = 0; i < size; i++)
-    {
-        vst1_f16(ptr, _v);
-        ptr += 4;
-    }
-}
-
-NCNN_FORCEINLINE void Mat::fill(float16x8_t _v)
-{
-    int size = (int)total();
-    __fp16* ptr = (__fp16*)data;
-    for (int i = 0; i < size; i++)
-    {
-        vst1q_f16(ptr, _v);
-        ptr += 8;
-    }
-}
-#endif
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-#endif // __ARM_NEON
-
-#if __SSE2__
-#if __AVX__
-#if __AVX512F__
-NCNN_FORCEINLINE void Mat::fill(__m512 _v)
-{
-    int size = (int)total();
-    float* ptr = (float*)data;
-    for (int i = 0; i < size; i++)
-    {
-        _mm512_storeu_ps(ptr, _v);
-        ptr += 16;
-    }
-}
-#endif // __AVX512F__
-NCNN_FORCEINLINE void Mat::fill(__m256 _v, int _i)
-{
-    // old gcc cannot overload __m128 and __m256 type
-    // add a dummy int parameter for different mangled function symbol
-    (void)_i;
-    int size = (int)total();
-    float* ptr = (float*)data;
-    for (int i = 0; i < size; i++)
-    {
-        _mm256_storeu_ps(ptr, _v);
-        ptr += 8;
-    }
-}
-#endif // __AVX__
-NCNN_FORCEINLINE void Mat::fill(__m128 _v)
-{
-    int size = (int)total();
-    float* ptr = (float*)data;
-    for (int i = 0; i < size; i++)
-    {
-        _mm_storeu_ps(ptr, _v);
-        ptr += 4;
-    }
-}
-NCNN_FORCEINLINE void Mat::fill(__m128i _v)
-{
-    int size = (int)total();
-    unsigned short* ptr = (unsigned short*)data;
-    for (int i = 0; i < size; i++)
-    {
-        _mm_store_si128((__m128i*)ptr, _v);
-        ptr += 8;
-    }
-}
-#endif // __SSE2__
-
-#if __mips_msa
-NCNN_FORCEINLINE void Mat::fill(v4f32 _v)
-{
-    int size = (int)total();
-    float* ptr = (float*)data;
-    for (int i = 0; i < size; i++)
-    {
-        __msa_st_w((v4i32)_v, ptr, 0);
-        ptr += 4;
-    }
-}
-#endif // __mips_msa
-
-#if __loongarch_sx
-NCNN_FORCEINLINE void Mat::fill(__m128 _v)
-{
-    int size = (int)total();
-    float* ptr = (float*)data;
-    for (int i = 0; i < size; i++)
-    {
-        __lsx_vst(_v, ptr, 0);
-        ptr += 4;
-    }
-}
-#endif // __loongarch_sx
-#if __riscv_vector
-NCNN_FORCEINLINE void Mat::fill(vfloat32m1_t _v)
-{
-    const int packn = cpu_riscv_vlenb() / 4;
-    const size_t vl = vsetvl_e32m1(packn);
-
-    int size = (int)total();
-    float* ptr = (float*)data;
-    for (int i = 0; i < size; i++)
-    {
-        vse32_v_f32m1(ptr, _v, vl);
-        ptr += packn;
-    }
-}
-
-NCNN_FORCEINLINE void Mat::fill(vuint16m1_t _v)
-{
-    const int packn = cpu_riscv_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int size = (int)total();
-    unsigned short* ptr = (unsigned short*)data;
-    for (int i = 0; i < size; i++)
-    {
-        vse16_v_u16m1(ptr, _v, vl);
-        ptr += packn;
-    }
-}
-
-NCNN_FORCEINLINE void Mat::fill(vint8m1_t _v)
-{
-    const int packn = cpu_riscv_vlenb() / 1;
-    const size_t vl = vsetvl_e8m1(packn);
-
-    int size = (int)total();
-    signed char* ptr = (signed char*)data;
-    for (int i = 0; i < size; i++)
-    {
-        vse8_v_i8m1(ptr, _v, vl);
-        ptr += packn;
-    }
-}
-#if __riscv_zfh
-NCNN_FORCEINLINE void Mat::fill(vfloat16m1_t _v)
-{
-    const int packn = cpu_riscv_vlenb() / 2;
-    const size_t vl = vsetvl_e16m1(packn);
-
-    int size = (int)total();
-    __fp16* ptr = (__fp16*)data;
-    for (int i = 0; i < size; i++)
-    {
-        vse16_v_f16m1(ptr, _v, vl);
-        ptr += packn;
-    }
-}
-#endif // __riscv_zfh
-#endif // __riscv_vector
-
-template<typename T>
-NCNN_FORCEINLINE void Mat::fill(T _v)
-{
-    int size = (int)total();
-    T* ptr = (T*)data;
-    for (int i = 0; i < size; i++)
-    {
-        ptr[i] = _v;
-    }
-}
-
-NCNN_FORCEINLINE Mat& Mat::operator=(const Mat& m)
-{
-    if (this == &m)
-        return *this;
-
-    if (m.refcount)
-        NCNN_XADD(m.refcount, 1);
-
-    release();
-
-    data = m.data;
-    refcount = m.refcount;
-    elemsize = m.elemsize;
-    elempack = m.elempack;
-    allocator = m.allocator;
-
-    dims = m.dims;
-    w = m.w;
-    h = m.h;
-    d = m.d;
-    c = m.c;
-
-    cstep = m.cstep;
-
-    return *this;
-}
-
-NCNN_FORCEINLINE void Mat::addref()
-{
-    if (refcount)
-        NCNN_XADD(refcount, 1);
-}
-
-NCNN_FORCEINLINE void Mat::release()
-{
-    if (refcount && NCNN_XADD(refcount, -1) == 1)
-    {
-        if (allocator)
-            allocator->fastFree(data);
-        else
-            fastFree(data);
-    }
-
-    data = 0;
-
-    elemsize = 0;
-    elempack = 0;
-
-    dims = 0;
-    w = 0;
-    h = 0;
-    d = 0;
-    c = 0;
-
-    cstep = 0;
-
-    refcount = 0;
-}
-
-NCNN_FORCEINLINE bool Mat::empty() const
-{
-    return data == 0 || total() == 0;
-}
-
-NCNN_FORCEINLINE size_t Mat::total() const
-{
-    return cstep * c;
-}
-
-NCNN_FORCEINLINE int Mat::elembits() const
-{
-    return elempack ? static_cast<int>(elemsize * 8) / elempack : 0;
-}
-
-NCNN_FORCEINLINE Mat Mat::shape() const
-{
-    if (dims == 1)
-        return Mat(w * elempack, (void*)0);
-    if (dims == 2)
-        return Mat(w, h * elempack, (void*)0);
-    if (dims == 3)
-        return Mat(w, h, c * elempack, (void*)0);
-    if (dims == 4)
-        return Mat(w, h, d, c * elempack, (void*)0);
-
-    return Mat();
-}
-
-NCNN_FORCEINLINE Mat Mat::channel(int _c)
-{
-    Mat m(w, h, d, (unsigned char*)data + cstep * _c * elemsize, elemsize, elempack, allocator);
-    m.dims = dims - 1;
-    if (dims == 4)
-        m.cstep = (size_t)w * h;
-    return m;
-}
-
-NCNN_FORCEINLINE const Mat Mat::channel(int _c) const
-{
-    Mat m(w, h, d, (unsigned char*)data + cstep * _c * elemsize, elemsize, elempack, allocator);
-    m.dims = dims - 1;
-    if (dims == 4)
-        m.cstep = (size_t)w * h;
-    return m;
-}
-
-NCNN_FORCEINLINE Mat Mat::depth(int z)
-{
-    return Mat(w, h, (unsigned char*)data + (size_t)w * h * z * elemsize, elemsize, elempack, allocator);
-}
-
-NCNN_FORCEINLINE const Mat Mat::depth(int z) const
-{
-    return Mat(w, h, (unsigned char*)data + (size_t)w * h * z * elemsize, elemsize, elempack, allocator);
-}
-
-NCNN_FORCEINLINE float* Mat::row(int y)
-{
-    return (float*)((unsigned char*)data + (size_t)w * y * elemsize);
-}
-
-NCNN_FORCEINLINE const float* Mat::row(int y) const
-{
-    return (const float*)((unsigned char*)data + (size_t)w * y * elemsize);
-}
-
-template<typename T>
-NCNN_FORCEINLINE T* Mat::row(int y)
-{
-    return (T*)((unsigned char*)data + (size_t)w * y * elemsize);
-}
-
-template<typename T>
-NCNN_FORCEINLINE const T* Mat::row(int y) const
-{
-    return (const T*)((unsigned char*)data + (size_t)w * y * elemsize);
-}
-
-NCNN_FORCEINLINE Mat Mat::channel_range(int _c, int channels)
-{
-    Mat m(w, h, d, channels, (unsigned char*)data + cstep * _c * elemsize, elemsize, elempack, allocator);
-    m.dims = dims;
-    return m;
-}
-
-NCNN_FORCEINLINE const Mat Mat::channel_range(int _c, int channels) const
-{
-    Mat m(w, h, d, channels, (unsigned char*)data + cstep * _c * elemsize, elemsize, elempack, allocator);
-    m.dims = dims;
-    return m;
-}
-
-NCNN_FORCEINLINE Mat Mat::depth_range(int z, int depths)
-{
-    Mat m(w, h, depths, (unsigned char*)data + (size_t)w * h * z * elemsize, elemsize, elempack, allocator);
-    m.cstep = (size_t)w * h;
-    return m;
-}
-
-NCNN_FORCEINLINE const Mat Mat::depth_range(int z, int depths) const
-{
-    Mat m(w, h, depths, (unsigned char*)data + (size_t)w * h * z * elemsize, elemsize, elempack, allocator);
-    m.cstep = (size_t)w * h;
-    return m;
-}
-
-NCNN_FORCEINLINE Mat Mat::row_range(int y, int rows)
-{
-    return Mat(w, rows, (unsigned char*)data + (size_t)w * y * elemsize, elemsize, elempack, allocator);
-}
-
-NCNN_FORCEINLINE const Mat Mat::row_range(int y, int rows) const
-{
-    return Mat(w, rows, (unsigned char*)data + (size_t)w * y * elemsize, elemsize, elempack, allocator);
-}
-
-NCNN_FORCEINLINE Mat Mat::range(int x, int n)
-{
-    return Mat(n, (unsigned char*)data + x * elemsize, elemsize, elempack, allocator);
-}
-
-NCNN_FORCEINLINE const Mat Mat::range(int x, int n) const
-{
-    return Mat(n, (unsigned char*)data + x * elemsize, elemsize, elempack, allocator);
-}
-
-template<typename T>
-NCNN_FORCEINLINE Mat::operator T*()
-{
-    return (T*)data;
-}
-
-template<typename T>
-NCNN_FORCEINLINE Mat::operator const T*() const
-{
-    return (const T*)data;
-}
-
-NCNN_FORCEINLINE float& Mat::operator[](size_t i)
-{
-    return ((float*)data)[i];
-}
-
-NCNN_FORCEINLINE const float& Mat::operator[](size_t i) const
-{
-    return ((const float*)data)[i];
-}
-
-#if NCNN_VULKAN
-
-NCNN_FORCEINLINE VkMat::VkMat()
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, size_t _elemsize, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _elemsize, _allocator);
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, int _h, size_t _elemsize, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _h, _elemsize, _allocator);
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, int _h, int _c, size_t _elemsize, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _h, _c, _elemsize, _allocator);
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, int _h, int _d, int _c, size_t _elemsize, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _h, _d, _c, _elemsize, _allocator);
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _elemsize, _elempack, _allocator);
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, int _h, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _h, _elemsize, _elempack, _allocator);
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, int _h, int _c, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _h, _c, _elemsize, _elempack, _allocator);
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, int _h, int _d, int _c, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0), cstep(0)
-{
-    create(_w, _h, _d, _c, _elemsize, _elempack, _allocator);
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(const VkMat& m)
-    : data(m.data), refcount(m.refcount), elemsize(m.elemsize), elempack(m.elempack), allocator(m.allocator), dims(m.dims), w(m.w), h(m.h), d(m.d), c(m.c)
-{
-    addref();
-
-    cstep = m.cstep;
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, VkBufferMemory* _data, size_t _elemsize, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(1), allocator(_allocator), dims(1), w(_w), h(1), d(1), c(1)
-{
-    cstep = w;
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, int _h, VkBufferMemory* _data, size_t _elemsize, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(1), allocator(_allocator), dims(2), w(_w), h(_h), d(1), c(1)
-{
-    cstep = w * h;
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, int _h, int _c, VkBufferMemory* _data, size_t _elemsize, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(1), allocator(_allocator), dims(3), w(_w), h(_h), d(1), c(_c)
-{
-    cstep = alignSize(w * h * elemsize, 16) / elemsize;
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, int _h, int _d, int _c, VkBufferMemory* _data, size_t _elemsize, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(1), allocator(_allocator), dims(4), w(_w), h(_h), d(_d), c(_c)
-{
-    cstep = alignSize(w * h * d * elemsize, 16) / elemsize;
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, VkBufferMemory* _data, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(_elempack), allocator(_allocator), dims(1), w(_w), h(1), d(1), c(1)
-{
-    cstep = w;
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, int _h, VkBufferMemory* _data, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(_elempack), allocator(_allocator), dims(2), w(_w), h(_h), d(1), c(1)
-{
-    cstep = w * h;
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, int _h, int _c, VkBufferMemory* _data, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(_elempack), allocator(_allocator), dims(3), w(_w), h(_h), d(1), c(_c)
-{
-    cstep = alignSize(w * h * elemsize, 16) / elemsize;
-}
-
-NCNN_FORCEINLINE VkMat::VkMat(int _w, int _h, int _d, int _c, VkBufferMemory* _data, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(_elempack), allocator(_allocator), dims(4), w(_w), h(_h), d(_d), c(_c)
-{
-    cstep = alignSize(w * h * d * elemsize, 16) / elemsize;
-}
-
-NCNN_FORCEINLINE VkMat::~VkMat()
-{
-    release();
-}
-
-NCNN_FORCEINLINE VkMat& VkMat::operator=(const VkMat& m)
-{
-    if (this == &m)
-        return *this;
-
-    if (m.refcount)
-        NCNN_XADD(m.refcount, 1);
-
-    release();
-
-    data = m.data;
-    refcount = m.refcount;
-    elemsize = m.elemsize;
-    elempack = m.elempack;
-    allocator = m.allocator;
-
-    dims = m.dims;
-    w = m.w;
-    h = m.h;
-    d = m.d;
-    c = m.c;
-
-    cstep = m.cstep;
-
-    return *this;
-}
-
-NCNN_FORCEINLINE Mat VkMat::mapped() const
-{
-    if (!allocator->mappable)
-        return Mat();
-
-    if (dims == 1)
-        return Mat(w, mapped_ptr(), elemsize, elempack, 0);
-
-    if (dims == 2)
-        return Mat(w, h, mapped_ptr(), elemsize, elempack, 0);
-
-    if (dims == 3)
-        return Mat(w, h, c, mapped_ptr(), elemsize, elempack, 0);
-
-    if (dims == 4)
-        return Mat(w, h, d, c, mapped_ptr(), elemsize, elempack, 0);
-
-    return Mat();
-}
-
-NCNN_FORCEINLINE void* VkMat::mapped_ptr() const
-{
-    if (!allocator->mappable)
-        return 0;
-
-    return (unsigned char*)data->mapped_ptr + data->offset;
-}
-
-NCNN_FORCEINLINE void VkMat::addref()
-{
-    if (refcount)
-        NCNN_XADD(refcount, 1);
-}
-
-NCNN_FORCEINLINE void VkMat::release()
-{
-    if (refcount && NCNN_XADD(refcount, -1) == 1)
-    {
-        if (allocator && data)
-        {
-            allocator->fastFree(data);
-        }
-    }
-
-    data = 0;
-
-    elemsize = 0;
-    elempack = 0;
-
-    dims = 0;
-    w = 0;
-    h = 0;
-    d = 0;
-    c = 0;
-
-    cstep = 0;
-
-    refcount = 0;
-}
-
-NCNN_FORCEINLINE bool VkMat::empty() const
-{
-    return data == 0 || total() == 0;
-}
-
-NCNN_FORCEINLINE size_t VkMat::total() const
-{
-    return cstep * c;
-}
-
-NCNN_FORCEINLINE int VkMat::elembits() const
-{
-    return elempack ? static_cast<int>(elemsize) * 8 / elempack : 0;
-}
-
-NCNN_FORCEINLINE Mat VkMat::shape() const
-{
-    if (dims == 1)
-        return Mat(w * elempack, (void*)0);
-    if (dims == 2)
-        return Mat(w, h * elempack, (void*)0);
-    if (dims == 3)
-        return Mat(w, h, c * elempack, (void*)0);
-    if (dims == 4)
-        return Mat(w, h, d, c * elempack, (void*)0);
-
-    return Mat();
-}
-
-NCNN_FORCEINLINE VkBuffer VkMat::buffer() const
-{
-    return data->buffer;
-}
-
-NCNN_FORCEINLINE size_t VkMat::buffer_offset() const
-{
-    return data->offset;
-}
-
-NCNN_FORCEINLINE size_t VkMat::buffer_capacity() const
-{
-    return data->capacity;
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat()
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0)
-{
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, size_t _elemsize, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0)
-{
-    create(_w, _elemsize, _allocator);
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, int _h, size_t _elemsize, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0)
-{
-    create(_w, _h, _elemsize, _allocator);
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, int _h, int _c, size_t _elemsize, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0)
-{
-    create(_w, _h, _c, _elemsize, _allocator);
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, int _h, int _d, int _c, size_t _elemsize, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0)
-{
-    create(_w, _h, _d, _c, _elemsize, _allocator);
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0)
-{
-    create(_w, _elemsize, _elempack, _allocator);
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, int _h, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0)
-{
-    create(_w, _h, _elemsize, _elempack, _allocator);
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, int _h, int _c, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0)
-{
-    create(_w, _h, _c, _elemsize, _elempack, _allocator);
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, int _h, int _d, int _c, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(0), refcount(0), elemsize(0), elempack(0), allocator(0), dims(0), w(0), h(0), d(0), c(0)
-{
-    create(_w, _h, _d, _c, _elemsize, _elempack, _allocator);
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(const VkImageMat& m)
-    : data(m.data), refcount(m.refcount), elemsize(m.elemsize), elempack(m.elempack), allocator(m.allocator), dims(m.dims), w(m.w), h(m.h), d(m.d), c(m.c)
-{
-    addref();
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, VkImageMemory* _data, size_t _elemsize, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(1), allocator(_allocator), dims(1), w(_w), h(1), d(1), c(1)
-{
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, int _h, VkImageMemory* _data, size_t _elemsize, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(1), allocator(_allocator), dims(2), w(_w), h(_h), d(1), c(1)
-{
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, int _h, int _c, VkImageMemory* _data, size_t _elemsize, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(1), allocator(_allocator), dims(3), w(_w), h(_h), d(1), c(_c)
-{
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, int _h, int _d, int _c, VkImageMemory* _data, size_t _elemsize, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(1), allocator(_allocator), dims(4), w(_w), h(_h), d(_d), c(_c)
-{
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, VkImageMemory* _data, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(_elempack), allocator(_allocator), dims(1), w(_w), h(1), d(1), c(1)
-{
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, int _h, VkImageMemory* _data, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(_elempack), allocator(_allocator), dims(2), w(_w), h(_h), d(1), c(1)
-{
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, int _h, int _c, VkImageMemory* _data, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(_elempack), allocator(_allocator), dims(3), w(_w), h(_h), d(1), c(_c)
-{
-}
-
-NCNN_FORCEINLINE VkImageMat::VkImageMat(int _w, int _h, int _d, int _c, VkImageMemory* _data, size_t _elemsize, int _elempack, VkAllocator* _allocator)
-    : data(_data), refcount(0), elemsize(_elemsize), elempack(_elempack), allocator(_allocator), dims(4), w(_w), h(_h), d(_d), c(_c)
-{
-}
-
-NCNN_FORCEINLINE VkImageMat::~VkImageMat()
-{
-    release();
-}
-
-NCNN_FORCEINLINE VkImageMat& VkImageMat::operator=(const VkImageMat& m)
-{
-    if (this == &m)
-        return *this;
-
-    if (m.refcount)
-        NCNN_XADD(m.refcount, 1);
-
-    release();
-
-    data = m.data;
-    refcount = m.refcount;
-    elemsize = m.elemsize;
-    elempack = m.elempack;
-    allocator = m.allocator;
-
-    dims = m.dims;
-    w = m.w;
-    h = m.h;
-    d = m.d;
-    c = m.c;
-
-    return *this;
-}
-
-NCNN_FORCEINLINE Mat VkImageMat::mapped() const
-{
-    if (!allocator->mappable || !data->mapped_ptr)
-        return Mat();
-
-    if (dims == 1)
-        return Mat(w, mapped_ptr(), elemsize, elempack, 0);
-
-    if (dims == 2)
-        return Mat(w, h, mapped_ptr(), elemsize, elempack, 0);
-
-    if (dims == 3)
-        return Mat(w, h, c, mapped_ptr(), elemsize, elempack, 0);
-
-    if (dims == 4)
-        return Mat(w, h, d, c, mapped_ptr(), elemsize, elempack, 0);
-
-    return Mat();
-}
-
-NCNN_FORCEINLINE void* VkImageMat::mapped_ptr() const
-{
-    if (!allocator->mappable || !data->mapped_ptr)
-        return 0;
-
-    return (unsigned char*)data->mapped_ptr + data->bind_offset;
-}
-
-NCNN_FORCEINLINE void VkImageMat::addref()
-{
-    if (refcount)
-        NCNN_XADD(refcount, 1);
-}
-
-NCNN_FORCEINLINE void VkImageMat::release()
-{
-    if (refcount && NCNN_XADD(refcount, -1) == 1)
-    {
-        if (allocator && data)
-        {
-            allocator->fastFree(data);
-        }
-    }
-
-    data = 0;
-
-    elemsize = 0;
-    elempack = 0;
-
-    dims = 0;
-    w = 0;
-    h = 0;
-    d = 0;
-    c = 0;
-
-    refcount = 0;
-}
-
-NCNN_FORCEINLINE bool VkImageMat::empty() const
-{
-    return data == 0 || total() == 0;
-}
-
-NCNN_FORCEINLINE size_t VkImageMat::total() const
-{
-    return w * h * d * c;
-}
-
-NCNN_FORCEINLINE int VkImageMat::elembits() const
-{
-    return elempack ? static_cast<int>(elemsize) * 8 / elempack : 0;
-}
-
-NCNN_FORCEINLINE Mat VkImageMat::shape() const
-{
-    if (dims == 1)
-        return Mat(w * elempack, (void*)0);
-    if (dims == 2)
-        return Mat(w, h * elempack, (void*)0);
-    if (dims == 3)
-        return Mat(w, h, c * elempack, (void*)0);
-    if (dims == 4)
-        return Mat(w, h, d, c * elempack, (void*)0);
-
-    return Mat();
-}
-
-NCNN_FORCEINLINE VkImage VkImageMat::image() const
-{
-    return data->image;
-}
-
-NCNN_FORCEINLINE VkImageView VkImageMat::imageview() const
-{
-    return data->imageview;
-}
-
-#endif // NCNN_VULKAN
-
-} // namespace ncnn
-
-#endif // NCNN_MAT_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/mat_pixel_drawing_font.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/mat_pixel_drawing_font.h
deleted file mode 100644
index 47dd3866..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/mat_pixel_drawing_font.h
+++ /dev/null
@@ -1,3870 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static const unsigned char mono_font_data[94][40 * 10] = {
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 249, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 248, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 248, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 247, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 246, 223, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 207, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 207, 0, 0, 0, 0,
-        0, 0, 0, 0, 244, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 243, 175, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 159, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 159, 0, 0, 0, 0,
-        0, 0, 0, 0, 241, 143, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 127, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 224, 95, 0, 0, 0, 0,
-        0, 0, 0, 0, 208, 79, 0, 0, 0, 0,
-        0, 0, 0, 0, 160, 61, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 145, 91, 0, 0, 0, 0,
-        0, 0, 0, 0, 251, 255, 1, 0, 0, 0,
-        0, 0, 0, 0, 254, 255, 4, 0, 0, 0,
-        0, 0, 0, 0, 252, 255, 2, 0, 0, 0,
-        0, 0, 0, 0, 212, 143, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 243, 255, 10, 48, 255, 175, 0, 0,
-        0, 0, 242, 255, 8, 32, 255, 143, 0, 0,
-        0, 0, 241, 255, 7, 16, 255, 127, 0, 0,
-        0, 0, 240, 255, 6, 0, 255, 111, 0, 0,
-        0, 0, 224, 255, 5, 0, 254, 95, 0, 0,
-        0, 0, 208, 255, 3, 0, 253, 63, 0, 0,
-        0, 0, 192, 255, 2, 0, 252, 47, 0, 0,
-        0, 0, 176, 255, 1, 0, 251, 31, 0, 0,
-        0, 0, 144, 255, 0, 0, 249, 15, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 96, 255, 0, 0, 241, 79, 0,
-        0, 0, 0, 144, 223, 0, 0, 244, 31, 0,
-        0, 0, 0, 192, 175, 0, 0, 247, 14, 0,
-        0, 0, 0, 240, 127, 0, 0, 250, 11, 0,
-        0, 0, 0, 242, 79, 0, 0, 253, 8, 0,
-        0, 0, 0, 245, 31, 0, 0, 255, 6, 0,
-        0, 0, 0, 247, 15, 0, 48, 255, 3, 0,
-        0, 85, 85, 252, 93, 85, 133, 255, 85, 53,
-        0, 255, 255, 255, 255, 255, 255, 255, 255, 175,
-        0, 255, 255, 255, 255, 255, 255, 255, 255, 175,
-        0, 0, 48, 255, 3, 0, 240, 127, 0, 0,
-        0, 0, 96, 255, 0, 0, 242, 79, 0, 0,
-        0, 0, 144, 223, 0, 0, 245, 31, 0, 0,
-        0, 0, 192, 175, 0, 0, 248, 14, 0, 0,
-        0, 0, 240, 127, 0, 0, 251, 11, 0, 0,
-        81, 85, 246, 143, 85, 85, 254, 91, 85, 2,
-        244, 255, 255, 255, 255, 255, 255, 255, 255, 7,
-        244, 255, 255, 255, 255, 255, 255, 255, 255, 7,
-        0, 0, 252, 11, 0, 112, 255, 0, 0, 0,
-        0, 0, 254, 8, 0, 160, 207, 0, 0, 0,
-        0, 16, 255, 5, 0, 208, 143, 0, 0, 0,
-        0, 64, 255, 2, 0, 241, 95, 0, 0, 0,
-        0, 112, 239, 0, 0, 244, 47, 0, 0, 0,
-        0, 160, 191, 0, 0, 247, 15, 0, 0, 0,
-        0, 208, 143, 0, 0, 250, 12, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 224, 95, 0, 0, 0, 0,
-        0, 0, 0, 0, 224, 95, 0, 0, 0, 0,
-        0, 0, 0, 0, 224, 95, 0, 0, 0, 0,
-        0, 0, 0, 147, 253, 255, 206, 73, 0, 0,
-        0, 0, 160, 255, 255, 255, 255, 255, 11, 0,
-        0, 0, 251, 255, 252, 191, 235, 255, 6, 0,
-        0, 64, 255, 62, 224, 95, 0, 148, 0, 0,
-        0, 144, 255, 6, 224, 95, 0, 0, 0, 0,
-        0, 160, 255, 4, 224, 95, 0, 0, 0, 0,
-        0, 144, 255, 5, 224, 95, 0, 0, 0, 0,
-        0, 96, 255, 29, 224, 95, 0, 0, 0, 0,
-        0, 0, 254, 239, 230, 95, 0, 0, 0, 0,
-        0, 0, 226, 255, 255, 127, 0, 0, 0, 0,
-        0, 0, 16, 249, 255, 255, 126, 0, 0, 0,
-        0, 0, 0, 16, 247, 255, 255, 78, 0, 0,
-        0, 0, 0, 0, 224, 207, 255, 255, 4, 0,
-        0, 0, 0, 0, 224, 95, 177, 255, 13, 0,
-        0, 0, 0, 0, 224, 95, 0, 254, 47, 0,
-        0, 0, 0, 0, 224, 95, 0, 250, 79, 0,
-        0, 0, 0, 0, 224, 95, 0, 251, 79, 0,
-        0, 80, 0, 0, 224, 95, 48, 255, 31, 0,
-        0, 240, 158, 37, 224, 143, 248, 255, 9, 0,
-        0, 240, 255, 255, 255, 255, 255, 191, 0, 0,
-        0, 128, 254, 255, 255, 255, 223, 6, 0, 0,
-        0, 0, 48, 117, 248, 159, 2, 0, 0, 0,
-        0, 0, 0, 0, 224, 95, 0, 0, 0, 0,
-        0, 0, 0, 0, 224, 95, 0, 0, 0, 0,
-        0, 0, 0, 0, 224, 95, 0, 0, 0, 0,
-        0, 0, 0, 0, 112, 40, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 215, 255, 59, 0, 0, 0, 245, 79, 0,
-        160, 255, 255, 255, 5, 0, 0, 252, 13, 0,
-        245, 159, 50, 252, 31, 0, 64, 255, 5, 0,
-        251, 13, 0, 242, 127, 0, 192, 223, 0, 0,
-        254, 9, 0, 224, 175, 0, 244, 95, 0, 0,
-        255, 8, 0, 208, 191, 0, 251, 13, 0, 0,
-        254, 10, 0, 224, 175, 48, 255, 6, 0, 0,
-        251, 13, 0, 242, 127, 176, 239, 0, 0, 0,
-        245, 143, 33, 252, 31, 243, 111, 0, 0, 0,
-        160, 255, 255, 255, 6, 250, 14, 0, 0, 0,
-        0, 214, 255, 76, 32, 255, 7, 0, 0, 0,
-        0, 0, 0, 0, 160, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 127, 0, 0, 0, 0,
-        0, 0, 0, 0, 249, 31, 129, 253, 158, 1,
-        0, 0, 0, 16, 255, 24, 253, 255, 255, 29,
-        0, 0, 0, 144, 255, 145, 255, 22, 245, 175,
-        0, 0, 0, 241, 159, 240, 159, 0, 144, 255,
-        0, 0, 0, 248, 31, 243, 95, 0, 64, 255,
-        0, 0, 16, 255, 9, 244, 63, 0, 48, 255,
-        0, 0, 128, 255, 2, 244, 63, 0, 48, 255,
-        0, 0, 224, 175, 0, 243, 95, 0, 64, 255,
-        0, 0, 247, 47, 0, 224, 159, 0, 144, 255,
-        0, 0, 254, 10, 0, 128, 255, 22, 245, 175,
-        0, 112, 255, 2, 0, 0, 252, 255, 255, 29,
-        0, 224, 191, 0, 0, 0, 128, 254, 158, 1,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 16, 200, 255, 190, 5, 0, 0, 0,
-        0, 0, 227, 255, 255, 255, 191, 0, 0, 0,
-        0, 16, 254, 255, 154, 252, 255, 9, 0, 0,
-        0, 112, 255, 29, 0, 112, 255, 15, 0, 0,
-        0, 160, 255, 6, 0, 0, 253, 63, 0, 0,
-        0, 176, 255, 4, 0, 0, 252, 79, 0, 0,
-        0, 144, 255, 6, 0, 0, 254, 31, 0, 0,
-        0, 64, 255, 12, 0, 112, 255, 10, 0, 0,
-        0, 0, 253, 95, 0, 246, 239, 1, 0, 0,
-        0, 0, 245, 239, 179, 255, 62, 0, 0, 0,
-        0, 0, 144, 255, 255, 207, 2, 0, 0, 0,
-        0, 0, 96, 255, 255, 7, 0, 0, 0, 0,
-        0, 16, 251, 255, 255, 9, 0, 0, 0, 0,
-        0, 209, 255, 142, 255, 111, 0, 0, 201, 76,
-        0, 253, 223, 1, 247, 255, 3, 0, 255, 31,
-        96, 255, 47, 0, 160, 255, 30, 80, 255, 12,
-        176, 255, 8, 0, 0, 252, 207, 208, 255, 5,
-        224, 255, 4, 0, 0, 225, 255, 252, 223, 0,
-        240, 255, 3, 0, 0, 48, 255, 255, 95, 0,
-        224, 255, 6, 0, 0, 0, 246, 255, 11, 0,
-        176, 255, 13, 0, 0, 0, 251, 255, 30, 0,
-        80, 255, 191, 1, 0, 229, 255, 255, 191, 0,
-        0, 252, 255, 207, 252, 255, 207, 226, 255, 7,
-        0, 193, 255, 255, 255, 255, 9, 64, 255, 79,
-        0, 0, 181, 254, 206, 40, 0, 0, 249, 239,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 251, 255, 2, 0, 0, 0,
-        0, 0, 0, 0, 250, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 249, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 248, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 246, 223, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 244, 175, 0, 0, 0, 0,
-        0, 0, 0, 0, 243, 159, 0, 0, 0, 0,
-        0, 0, 0, 0, 241, 143, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 64, 255, 29, 0, 0,
-        0, 0, 0, 0, 0, 245, 207, 1, 0, 0,
-        0, 0, 0, 0, 64, 255, 29, 0, 0, 0,
-        0, 0, 0, 0, 242, 255, 2, 0, 0, 0,
-        0, 0, 0, 0, 252, 111, 0, 0, 0, 0,
-        0, 0, 0, 112, 255, 11, 0, 0, 0, 0,
-        0, 0, 0, 241, 255, 2, 0, 0, 0, 0,
-        0, 0, 0, 248, 175, 0, 0, 0, 0, 0,
-        0, 0, 0, 254, 79, 0, 0, 0, 0, 0,
-        0, 0, 64, 255, 14, 0, 0, 0, 0, 0,
-        0, 0, 128, 255, 9, 0, 0, 0, 0, 0,
-        0, 0, 176, 255, 6, 0, 0, 0, 0, 0,
-        0, 0, 224, 255, 3, 0, 0, 0, 0, 0,
-        0, 0, 240, 255, 1, 0, 0, 0, 0, 0,
-        0, 0, 241, 255, 0, 0, 0, 0, 0, 0,
-        0, 0, 241, 255, 0, 0, 0, 0, 0, 0,
-        0, 0, 240, 255, 1, 0, 0, 0, 0, 0,
-        0, 0, 240, 255, 2, 0, 0, 0, 0, 0,
-        0, 0, 208, 255, 4, 0, 0, 0, 0, 0,
-        0, 0, 160, 255, 7, 0, 0, 0, 0, 0,
-        0, 0, 96, 255, 11, 0, 0, 0, 0, 0,
-        0, 0, 32, 255, 31, 0, 0, 0, 0, 0,
-        0, 0, 0, 252, 127, 0, 0, 0, 0, 0,
-        0, 0, 0, 245, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 192, 255, 6, 0, 0, 0, 0,
-        0, 0, 0, 48, 255, 30, 0, 0, 0, 0,
-        0, 0, 0, 0, 248, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 176, 255, 7, 0, 0, 0,
-        0, 0, 0, 0, 0, 252, 95, 0, 0, 0,
-        0, 0, 0, 0, 0, 193, 255, 6, 0, 0,
-        0, 0, 0, 0, 0, 0, 135, 23, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 128, 255, 10, 0, 0, 0, 0, 0,
-        0, 0, 0, 247, 191, 0, 0, 0, 0, 0,
-        0, 0, 0, 144, 255, 10, 0, 0, 0, 0,
-        0, 0, 0, 0, 251, 143, 0, 0, 0, 0,
-        0, 0, 0, 0, 225, 255, 3, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 13, 0, 0, 0,
-        0, 0, 0, 0, 0, 252, 111, 0, 0, 0,
-        0, 0, 0, 0, 0, 244, 239, 0, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 4, 0, 0,
-        0, 0, 0, 0, 0, 128, 255, 10, 0, 0,
-        0, 0, 0, 0, 0, 48, 255, 14, 0, 0,
-        0, 0, 0, 0, 0, 0, 255, 47, 0, 0,
-        0, 0, 0, 0, 0, 0, 253, 79, 0, 0,
-        0, 0, 0, 0, 0, 0, 251, 111, 0, 0,
-        0, 0, 0, 0, 0, 0, 250, 127, 0, 0,
-        0, 0, 0, 0, 0, 0, 250, 127, 0, 0,
-        0, 0, 0, 0, 0, 0, 251, 111, 0, 0,
-        0, 0, 0, 0, 0, 0, 252, 95, 0, 0,
-        0, 0, 0, 0, 0, 0, 254, 63, 0, 0,
-        0, 0, 0, 0, 0, 16, 255, 31, 0, 0,
-        0, 0, 0, 0, 0, 80, 255, 13, 0, 0,
-        0, 0, 0, 0, 0, 160, 255, 8, 0, 0,
-        0, 0, 0, 0, 0, 241, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 247, 191, 0, 0, 0,
-        0, 0, 0, 0, 16, 255, 63, 0, 0, 0,
-        0, 0, 0, 0, 144, 255, 10, 0, 0, 0,
-        0, 0, 0, 0, 244, 223, 0, 0, 0, 0,
-        0, 0, 0, 32, 254, 63, 0, 0, 0, 0,
-        0, 0, 0, 209, 255, 4, 0, 0, 0, 0,
-        0, 0, 16, 253, 79, 0, 0, 0, 0, 0,
-        0, 0, 96, 136, 3, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 247, 223, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 244, 159, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 127, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 16, 0,
-        0, 246, 107, 1, 240, 79, 0, 148, 221, 0,
-        0, 249, 255, 223, 232, 143, 251, 255, 255, 0,
-        0, 251, 255, 255, 255, 255, 255, 255, 255, 1,
-        0, 33, 83, 118, 253, 255, 104, 69, 19, 0,
-        0, 0, 0, 64, 255, 255, 9, 0, 0, 0,
-        0, 0, 0, 225, 191, 247, 95, 0, 0, 0,
-        0, 0, 0, 251, 79, 241, 255, 2, 0, 0,
-        0, 0, 96, 255, 13, 128, 255, 12, 0, 0,
-        0, 0, 242, 255, 6, 16, 255, 143, 0, 0,
-        0, 0, 162, 255, 0, 0, 249, 93, 0, 0,
-        0, 0, 0, 99, 0, 0, 113, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 176, 75, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 247, 255, 255, 255, 255, 255, 255, 223, 0,
-        0, 247, 255, 255, 255, 255, 255, 255, 223, 0,
-        0, 132, 136, 136, 249, 191, 136, 136, 120, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 80, 85, 5, 0, 0, 0,
-        0, 0, 0, 0, 240, 255, 14, 0, 0, 0,
-        0, 0, 0, 0, 242, 255, 10, 0, 0, 0,
-        0, 0, 0, 0, 245, 255, 5, 0, 0, 0,
-        0, 0, 0, 0, 248, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 251, 175, 0, 0, 0, 0,
-        0, 0, 0, 0, 254, 79, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 80, 255, 7, 0, 0, 0, 0,
-        0, 0, 0, 144, 255, 1, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 211, 221, 221, 221, 221, 157, 0, 0,
-        0, 0, 244, 255, 255, 255, 255, 175, 0, 0,
-        0, 0, 244, 255, 255, 255, 255, 175, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 129, 73, 0, 0, 0, 0,
-        0, 0, 0, 0, 254, 255, 5, 0, 0, 0,
-        0, 0, 0, 48, 255, 255, 10, 0, 0, 0,
-        0, 0, 0, 64, 255, 255, 10, 0, 0, 0,
-        0, 0, 0, 16, 255, 255, 7, 0, 0, 0,
-        0, 0, 0, 0, 230, 175, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 64, 255, 10, 0,
-        0, 0, 0, 0, 0, 0, 176, 255, 3, 0,
-        0, 0, 0, 0, 0, 0, 242, 223, 0, 0,
-        0, 0, 0, 0, 0, 0, 248, 111, 0, 0,
-        0, 0, 0, 0, 0, 0, 254, 15, 0, 0,
-        0, 0, 0, 0, 0, 96, 255, 9, 0, 0,
-        0, 0, 0, 0, 0, 192, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 243, 191, 0, 0, 0,
-        0, 0, 0, 0, 0, 250, 95, 0, 0, 0,
-        0, 0, 0, 0, 16, 255, 14, 0, 0, 0,
-        0, 0, 0, 0, 112, 255, 7, 0, 0, 0,
-        0, 0, 0, 0, 208, 255, 1, 0, 0, 0,
-        0, 0, 0, 0, 244, 175, 0, 0, 0, 0,
-        0, 0, 0, 0, 251, 63, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 13, 0, 0, 0, 0,
-        0, 0, 0, 128, 255, 6, 0, 0, 0, 0,
-        0, 0, 0, 224, 255, 0, 0, 0, 0, 0,
-        0, 0, 0, 246, 159, 0, 0, 0, 0, 0,
-        0, 0, 0, 252, 47, 0, 0, 0, 0, 0,
-        0, 0, 48, 255, 11, 0, 0, 0, 0, 0,
-        0, 0, 160, 255, 5, 0, 0, 0, 0, 0,
-        0, 0, 241, 239, 0, 0, 0, 0, 0, 0,
-        0, 0, 247, 127, 0, 0, 0, 0, 0, 0,
-        0, 0, 253, 31, 0, 0, 0, 0, 0, 0,
-        0, 64, 255, 10, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 113, 236, 223, 41, 0, 0, 0,
-        0, 0, 64, 254, 255, 255, 255, 7, 0, 0,
-        0, 0, 242, 255, 191, 235, 255, 111, 0, 0,
-        0, 0, 252, 191, 1, 0, 247, 255, 2, 0,
-        0, 64, 255, 13, 0, 0, 144, 255, 9, 0,
-        0, 160, 255, 6, 0, 0, 16, 255, 15, 0,
-        0, 224, 255, 1, 0, 0, 0, 251, 79, 0,
-        0, 242, 239, 0, 0, 0, 0, 248, 143, 0,
-        0, 244, 191, 0, 0, 0, 0, 245, 175, 0,
-        0, 246, 175, 0, 0, 0, 0, 243, 207, 0,
-        0, 247, 159, 0, 0, 0, 0, 242, 223, 0,
-        0, 248, 143, 0, 0, 0, 0, 242, 239, 0,
-        0, 248, 143, 0, 0, 0, 0, 241, 255, 0,
-        0, 248, 143, 0, 0, 0, 0, 242, 239, 0,
-        0, 247, 159, 0, 0, 0, 0, 242, 223, 0,
-        0, 246, 175, 0, 0, 0, 0, 243, 207, 0,
-        0, 244, 191, 0, 0, 0, 0, 245, 175, 0,
-        0, 242, 239, 0, 0, 0, 0, 248, 143, 0,
-        0, 224, 255, 1, 0, 0, 0, 251, 95, 0,
-        0, 144, 255, 6, 0, 0, 16, 255, 15, 0,
-        0, 48, 255, 13, 0, 0, 128, 255, 10, 0,
-        0, 0, 251, 191, 1, 0, 247, 255, 3, 0,
-        0, 0, 226, 255, 191, 234, 255, 143, 0, 0,
-        0, 0, 48, 254, 255, 255, 255, 8, 0, 0,
-        0, 0, 0, 113, 236, 223, 57, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 128, 255, 9, 0, 0, 0,
-        0, 0, 0, 16, 251, 255, 9, 0, 0, 0,
-        0, 0, 0, 210, 255, 255, 9, 0, 0, 0,
-        0, 0, 80, 255, 191, 255, 9, 0, 0, 0,
-        0, 0, 248, 255, 55, 255, 9, 0, 0, 0,
-        0, 16, 254, 111, 64, 255, 9, 0, 0, 0,
-        0, 0, 227, 4, 64, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 80, 255, 9, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 165, 253, 223, 57, 0, 0, 0,
-        0, 0, 213, 255, 255, 255, 255, 8, 0, 0,
-        0, 144, 255, 255, 172, 235, 255, 143, 0, 0,
-        0, 160, 255, 25, 0, 0, 249, 255, 2, 0,
-        0, 0, 58, 0, 0, 0, 192, 255, 7, 0,
-        0, 0, 0, 0, 0, 0, 96, 255, 11, 0,
-        0, 0, 0, 0, 0, 0, 64, 255, 12, 0,
-        0, 0, 0, 0, 0, 0, 80, 255, 10, 0,
-        0, 0, 0, 0, 0, 0, 112, 255, 8, 0,
-        0, 0, 0, 0, 0, 0, 192, 255, 3, 0,
-        0, 0, 0, 0, 0, 0, 244, 239, 0, 0,
-        0, 0, 0, 0, 0, 0, 253, 111, 0, 0,
-        0, 0, 0, 0, 0, 160, 255, 11, 0, 0,
-        0, 0, 0, 0, 0, 247, 239, 1, 0, 0,
-        0, 0, 0, 0, 64, 255, 63, 0, 0, 0,
-        0, 0, 0, 0, 243, 255, 5, 0, 0, 0,
-        0, 0, 0, 32, 254, 127, 0, 0, 0, 0,
-        0, 0, 0, 225, 255, 9, 0, 0, 0, 0,
-        0, 0, 16, 253, 175, 0, 0, 0, 0, 0,
-        0, 0, 192, 255, 11, 0, 0, 0, 0, 0,
-        0, 0, 251, 207, 0, 0, 0, 0, 0, 0,
-        0, 144, 255, 29, 0, 0, 0, 0, 0, 0,
-        0, 245, 255, 238, 238, 238, 238, 238, 174, 0,
-        0, 246, 255, 255, 255, 255, 255, 255, 207, 0,
-        0, 246, 255, 255, 255, 255, 255, 255, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 64, 217, 254, 206, 41, 0, 0, 0,
-        0, 64, 253, 255, 255, 255, 255, 9, 0, 0,
-        0, 246, 255, 239, 171, 252, 255, 159, 0, 0,
-        0, 225, 159, 2, 0, 16, 249, 255, 3, 0,
-        0, 48, 2, 0, 0, 0, 176, 255, 8, 0,
-        0, 0, 0, 0, 0, 0, 96, 255, 11, 0,
-        0, 0, 0, 0, 0, 0, 96, 255, 11, 0,
-        0, 0, 0, 0, 0, 0, 144, 255, 9, 0,
-        0, 0, 0, 0, 0, 0, 242, 255, 4, 0,
-        0, 0, 0, 0, 0, 80, 254, 175, 0, 0,
-        0, 0, 16, 153, 169, 254, 255, 10, 0, 0,
-        0, 0, 32, 255, 255, 239, 57, 0, 0, 0,
-        0, 0, 32, 255, 255, 255, 157, 3, 0, 0,
-        0, 0, 0, 0, 16, 115, 254, 159, 0, 0,
-        0, 0, 0, 0, 0, 0, 176, 255, 9, 0,
-        0, 0, 0, 0, 0, 0, 16, 255, 31, 0,
-        0, 0, 0, 0, 0, 0, 0, 251, 111, 0,
-        0, 0, 0, 0, 0, 0, 0, 249, 127, 0,
-        0, 0, 0, 0, 0, 0, 0, 251, 127, 0,
-        0, 0, 0, 0, 0, 0, 0, 254, 79, 0,
-        0, 1, 0, 0, 0, 0, 128, 255, 15, 0,
-        0, 172, 4, 0, 0, 16, 249, 255, 8, 0,
-        0, 253, 255, 172, 169, 252, 255, 191, 0, 0,
-        0, 251, 255, 255, 255, 255, 255, 8, 0, 0,
-        0, 48, 184, 254, 255, 190, 23, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 64, 255, 47, 0, 0,
-        0, 0, 0, 0, 0, 224, 255, 47, 0, 0,
-        0, 0, 0, 0, 0, 249, 255, 47, 0, 0,
-        0, 0, 0, 0, 48, 255, 252, 47, 0, 0,
-        0, 0, 0, 0, 208, 207, 249, 47, 0, 0,
-        0, 0, 0, 0, 248, 95, 250, 47, 0, 0,
-        0, 0, 0, 32, 255, 12, 250, 47, 0, 0,
-        0, 0, 0, 192, 255, 2, 251, 47, 0, 0,
-        0, 0, 0, 247, 127, 0, 251, 47, 0, 0,
-        0, 0, 16, 255, 12, 0, 252, 47, 0, 0,
-        0, 0, 176, 255, 3, 0, 252, 47, 0, 0,
-        0, 0, 245, 143, 0, 0, 252, 47, 0, 0,
-        0, 16, 254, 13, 0, 0, 252, 47, 0, 0,
-        0, 160, 255, 3, 0, 0, 252, 47, 0, 0,
-        0, 244, 159, 0, 0, 0, 252, 47, 0, 0,
-        0, 254, 14, 0, 0, 0, 252, 47, 0, 0,
-        144, 255, 4, 0, 0, 0, 252, 47, 0, 0,
-        240, 255, 255, 255, 255, 255, 255, 255, 255, 11,
-        240, 255, 255, 255, 255, 255, 255, 255, 255, 11,
-        192, 221, 221, 221, 221, 221, 255, 223, 221, 8,
-        0, 0, 0, 0, 0, 0, 252, 47, 0, 0,
-        0, 0, 0, 0, 0, 0, 252, 47, 0, 0,
-        0, 0, 0, 0, 0, 0, 252, 47, 0, 0,
-        0, 0, 0, 0, 0, 0, 252, 47, 0, 0,
-        0, 0, 0, 0, 0, 0, 252, 47, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 252, 255, 255, 255, 255, 191, 0, 0,
-        0, 0, 253, 255, 255, 255, 255, 191, 0, 0,
-        0, 0, 254, 223, 221, 221, 221, 157, 0, 0,
-        0, 0, 255, 13, 0, 0, 0, 0, 0, 0,
-        0, 16, 255, 12, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 10, 0, 0, 0, 0, 0, 0,
-        0, 48, 255, 9, 0, 0, 0, 0, 0, 0,
-        0, 80, 255, 7, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 6, 0, 0, 0, 0, 0, 0,
-        0, 112, 255, 120, 152, 121, 3, 0, 0, 0,
-        0, 128, 255, 255, 255, 255, 223, 4, 0, 0,
-        0, 96, 255, 255, 255, 255, 255, 127, 0, 0,
-        0, 0, 83, 19, 0, 114, 254, 255, 5, 0,
-        0, 0, 0, 0, 0, 0, 192, 255, 13, 0,
-        0, 0, 0, 0, 0, 0, 32, 255, 63, 0,
-        0, 0, 0, 0, 0, 0, 0, 252, 111, 0,
-        0, 0, 0, 0, 0, 0, 0, 250, 127, 0,
-        0, 0, 0, 0, 0, 0, 0, 250, 127, 0,
-        0, 0, 0, 0, 0, 0, 0, 251, 95, 0,
-        0, 0, 0, 0, 0, 0, 16, 255, 31, 0,
-        0, 0, 0, 0, 0, 0, 144, 255, 12, 0,
-        0, 226, 22, 0, 0, 16, 249, 255, 4, 0,
-        0, 242, 255, 189, 170, 252, 255, 127, 0, 0,
-        0, 225, 255, 255, 255, 255, 239, 5, 0, 0,
-        0, 0, 166, 253, 255, 190, 6, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 97, 219, 254, 206, 0, 0,
-        0, 0, 0, 112, 255, 255, 255, 255, 0, 0,
-        0, 0, 0, 251, 255, 191, 153, 202, 0, 0,
-        0, 0, 160, 255, 110, 0, 0, 0, 0, 0,
-        0, 0, 245, 223, 2, 0, 0, 0, 0, 0,
-        0, 0, 253, 63, 0, 0, 0, 0, 0, 0,
-        0, 64, 255, 10, 0, 0, 0, 0, 0, 0,
-        0, 144, 255, 5, 0, 0, 0, 0, 0, 0,
-        0, 208, 255, 1, 0, 0, 0, 0, 0, 0,
-        0, 240, 223, 16, 216, 255, 157, 2, 0, 0,
-        0, 243, 207, 226, 255, 255, 255, 111, 0, 0,
-        0, 244, 191, 253, 141, 135, 253, 255, 5, 0,
-        0, 245, 255, 111, 0, 0, 128, 255, 14, 0,
-        0, 246, 255, 7, 0, 0, 0, 252, 95, 0,
-        0, 246, 223, 0, 0, 0, 0, 245, 175, 0,
-        0, 245, 175, 0, 0, 0, 0, 242, 207, 0,
-        0, 244, 159, 0, 0, 0, 0, 241, 223, 0,
-        0, 242, 191, 0, 0, 0, 0, 242, 207, 0,
-        0, 240, 239, 0, 0, 0, 0, 244, 191, 0,
-        0, 160, 255, 5, 0, 0, 0, 248, 127, 0,
-        0, 64, 255, 30, 0, 0, 16, 254, 47, 0,
-        0, 0, 251, 207, 2, 0, 194, 255, 11, 0,
-        0, 0, 225, 255, 191, 186, 255, 239, 1, 0,
-        0, 0, 32, 253, 255, 255, 255, 45, 0, 0,
-        0, 0, 0, 96, 235, 239, 107, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 250, 255, 255, 255, 255, 255, 255, 255, 0,
-        0, 250, 255, 255, 255, 255, 255, 255, 255, 0,
-        0, 216, 221, 221, 221, 221, 221, 253, 239, 0,
-        0, 0, 0, 0, 0, 0, 0, 246, 143, 0,
-        0, 0, 0, 0, 0, 0, 0, 253, 47, 0,
-        0, 0, 0, 0, 0, 0, 48, 255, 11, 0,
-        0, 0, 0, 0, 0, 0, 160, 255, 5, 0,
-        0, 0, 0, 0, 0, 0, 241, 239, 0, 0,
-        0, 0, 0, 0, 0, 0, 248, 143, 0, 0,
-        0, 0, 0, 0, 0, 0, 254, 31, 0, 0,
-        0, 0, 0, 0, 0, 80, 255, 11, 0, 0,
-        0, 0, 0, 0, 0, 192, 255, 4, 0, 0,
-        0, 0, 0, 0, 0, 242, 223, 0, 0, 0,
-        0, 0, 0, 0, 0, 249, 127, 0, 0, 0,
-        0, 0, 0, 0, 16, 255, 31, 0, 0, 0,
-        0, 0, 0, 0, 112, 255, 10, 0, 0, 0,
-        0, 0, 0, 0, 208, 255, 3, 0, 0, 0,
-        0, 0, 0, 0, 244, 223, 0, 0, 0, 0,
-        0, 0, 0, 0, 251, 111, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 15, 0, 0, 0, 0,
-        0, 0, 0, 128, 255, 9, 0, 0, 0, 0,
-        0, 0, 0, 224, 255, 3, 0, 0, 0, 0,
-        0, 0, 0, 246, 207, 0, 0, 0, 0, 0,
-        0, 0, 0, 253, 111, 0, 0, 0, 0, 0,
-        0, 0, 48, 255, 14, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 146, 253, 239, 91, 0, 0, 0,
-        0, 0, 128, 255, 255, 255, 255, 45, 0, 0,
-        0, 0, 248, 255, 173, 202, 255, 239, 1, 0,
-        0, 32, 255, 127, 0, 0, 211, 255, 8, 0,
-        0, 112, 255, 10, 0, 0, 64, 255, 13, 0,
-        0, 144, 255, 6, 0, 0, 0, 255, 15, 0,
-        0, 128, 255, 6, 0, 0, 0, 255, 15, 0,
-        0, 80, 255, 9, 0, 0, 48, 255, 12, 0,
-        0, 0, 254, 31, 0, 0, 176, 255, 6, 0,
-        0, 0, 246, 223, 2, 0, 250, 191, 0, 0,
-        0, 0, 128, 255, 143, 213, 255, 10, 0, 0,
-        0, 0, 0, 245, 255, 255, 95, 0, 0, 0,
-        0, 0, 16, 250, 255, 255, 191, 2, 0, 0,
-        0, 0, 227, 255, 8, 179, 255, 79, 0, 0,
-        0, 32, 254, 95, 0, 0, 246, 255, 4, 0,
-        0, 176, 255, 7, 0, 0, 64, 255, 30, 0,
-        0, 241, 239, 0, 0, 0, 0, 248, 127, 0,
-        0, 245, 175, 0, 0, 0, 0, 243, 191, 0,
-        0, 246, 159, 0, 0, 0, 0, 242, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 245, 191, 0,
-        0, 241, 255, 1, 0, 0, 0, 251, 127, 0,
-        0, 176, 255, 28, 0, 0, 144, 255, 31, 0,
-        0, 32, 255, 255, 155, 185, 255, 255, 5, 0,
-        0, 0, 211, 255, 255, 255, 255, 94, 0, 0,
-        0, 0, 0, 165, 254, 239, 123, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 163, 253, 223, 40, 0, 0, 0,
-        0, 0, 160, 255, 255, 255, 255, 6, 0, 0,
-        0, 0, 250, 255, 173, 234, 255, 111, 0, 0,
-        0, 80, 255, 95, 0, 0, 247, 255, 2, 0,
-        0, 208, 255, 5, 0, 0, 112, 255, 10, 0,
-        0, 241, 239, 0, 0, 0, 0, 253, 31, 0,
-        0, 245, 175, 0, 0, 0, 0, 247, 95, 0,
-        0, 246, 143, 0, 0, 0, 0, 244, 143, 0,
-        0, 246, 143, 0, 0, 0, 0, 243, 191, 0,
-        0, 245, 159, 0, 0, 0, 0, 244, 207, 0,
-        0, 242, 239, 0, 0, 0, 0, 251, 223, 0,
-        0, 208, 255, 7, 0, 0, 112, 255, 207, 0,
-        0, 80, 255, 159, 2, 64, 251, 253, 207, 0,
-        0, 0, 248, 255, 255, 255, 223, 245, 191, 0,
-        0, 0, 80, 253, 255, 255, 26, 246, 159, 0,
-        0, 0, 0, 48, 101, 20, 0, 248, 111, 0,
-        0, 0, 0, 0, 0, 0, 0, 251, 79, 0,
-        0, 0, 0, 0, 0, 0, 0, 255, 15, 0,
-        0, 0, 0, 0, 0, 0, 96, 255, 10, 0,
-        0, 0, 0, 0, 0, 0, 224, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 250, 191, 0, 0,
-        0, 0, 0, 0, 0, 195, 255, 46, 0, 0,
-        0, 0, 168, 137, 218, 255, 255, 3, 0, 0,
-        0, 0, 249, 255, 255, 255, 43, 0, 0, 0,
-        0, 0, 231, 255, 206, 56, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 212, 143, 0, 0, 0, 0,
-        0, 0, 0, 0, 252, 255, 2, 0, 0, 0,
-        0, 0, 0, 0, 254, 255, 4, 0, 0, 0,
-        0, 0, 0, 0, 251, 255, 1, 0, 0, 0,
-        0, 0, 0, 0, 145, 90, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 145, 91, 0, 0, 0, 0,
-        0, 0, 0, 0, 251, 255, 1, 0, 0, 0,
-        0, 0, 0, 0, 254, 255, 4, 0, 0, 0,
-        0, 0, 0, 0, 252, 255, 2, 0, 0, 0,
-        0, 0, 0, 0, 212, 143, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 210, 175, 0, 0, 0, 0,
-        0, 0, 0, 0, 250, 255, 4, 0, 0, 0,
-        0, 0, 0, 0, 252, 255, 6, 0, 0, 0,
-        0, 0, 0, 0, 249, 255, 3, 0, 0, 0,
-        0, 0, 0, 0, 145, 107, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 249, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 252, 175, 0, 0, 0, 0,
-        0, 0, 0, 0, 255, 111, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 31, 0, 0, 0, 0,
-        0, 0, 0, 80, 255, 11, 0, 0, 0, 0,
-        0, 0, 0, 128, 255, 5, 0, 0, 0, 0,
-        0, 0, 0, 192, 239, 0, 0, 0, 0, 0,
-        0, 0, 0, 240, 143, 0, 0, 0, 0, 0,
-        0, 0, 0, 129, 24, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 130, 0,
-        0, 0, 0, 0, 0, 0, 0, 162, 223, 0,
-        0, 0, 0, 0, 0, 0, 162, 255, 223, 0,
-        0, 0, 0, 0, 0, 162, 255, 255, 25, 0,
-        0, 0, 0, 0, 162, 255, 255, 24, 0, 0,
-        0, 0, 0, 162, 255, 255, 24, 0, 0, 0,
-        0, 0, 162, 255, 239, 24, 0, 0, 0, 0,
-        0, 162, 255, 239, 7, 0, 0, 0, 0, 0,
-        0, 247, 255, 8, 0, 0, 0, 0, 0, 0,
-        0, 245, 255, 143, 1, 0, 0, 0, 0, 0,
-        0, 16, 249, 255, 143, 1, 0, 0, 0, 0,
-        0, 0, 16, 249, 255, 159, 1, 0, 0, 0,
-        0, 0, 0, 16, 249, 255, 159, 2, 0, 0,
-        0, 0, 0, 0, 16, 249, 255, 159, 2, 0,
-        0, 0, 0, 0, 0, 16, 249, 255, 159, 0,
-        0, 0, 0, 0, 0, 0, 16, 249, 223, 0,
-        0, 0, 0, 0, 0, 0, 0, 16, 217, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 16, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 131, 136, 136, 136, 136, 136, 136, 120, 0,
-        0, 247, 255, 255, 255, 255, 255, 255, 223, 0,
-        0, 247, 255, 255, 255, 255, 255, 255, 223, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 132, 136, 136, 136, 136, 136, 136, 120, 0,
-        0, 247, 255, 255, 255, 255, 255, 255, 223, 0,
-        0, 247, 255, 255, 255, 255, 255, 255, 223, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 85, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 247, 93, 0, 0, 0, 0, 0, 0, 0,
-        0, 247, 255, 93, 0, 0, 0, 0, 0, 0,
-        0, 96, 253, 255, 93, 0, 0, 0, 0, 0,
-        0, 0, 80, 253, 255, 93, 0, 0, 0, 0,
-        0, 0, 0, 80, 253, 255, 93, 0, 0, 0,
-        0, 0, 0, 0, 80, 253, 255, 93, 0, 0,
-        0, 0, 0, 0, 0, 64, 252, 255, 77, 0,
-        0, 0, 0, 0, 0, 0, 64, 254, 223, 0,
-        0, 0, 0, 0, 0, 0, 213, 255, 175, 0,
-        0, 0, 0, 0, 0, 214, 255, 207, 4, 0,
-        0, 0, 0, 0, 214, 255, 207, 4, 0, 0,
-        0, 0, 0, 230, 255, 207, 4, 0, 0, 0,
-        0, 0, 230, 255, 207, 4, 0, 0, 0, 0,
-        0, 228, 255, 207, 4, 0, 0, 0, 0, 0,
-        0, 247, 207, 4, 0, 0, 0, 0, 0, 0,
-        0, 199, 4, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 81, 201, 254, 239, 107, 1, 0, 0,
-        0, 160, 255, 255, 255, 255, 255, 94, 0, 0,
-        0, 208, 255, 255, 205, 236, 255, 255, 6, 0,
-        0, 96, 142, 4, 0, 0, 179, 255, 31, 0,
-        0, 0, 0, 0, 0, 0, 0, 252, 111, 0,
-        0, 0, 0, 0, 0, 0, 0, 246, 143, 0,
-        0, 0, 0, 0, 0, 0, 0, 245, 143, 0,
-        0, 0, 0, 0, 0, 0, 0, 248, 111, 0,
-        0, 0, 0, 0, 0, 0, 0, 254, 31, 0,
-        0, 0, 0, 0, 0, 0, 176, 255, 10, 0,
-        0, 0, 0, 0, 0, 16, 252, 207, 0, 0,
-        0, 0, 0, 0, 0, 228, 255, 11, 0, 0,
-        0, 0, 0, 0, 96, 255, 143, 0, 0, 0,
-        0, 0, 0, 0, 244, 255, 5, 0, 0, 0,
-        0, 0, 0, 0, 252, 79, 0, 0, 0, 0,
-        0, 0, 0, 16, 255, 11, 0, 0, 0, 0,
-        0, 0, 0, 48, 255, 7, 0, 0, 0, 0,
-        0, 0, 0, 32, 153, 3, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 16, 185, 5, 0, 0, 0, 0,
-        0, 0, 0, 176, 255, 31, 0, 0, 0, 0,
-        0, 0, 0, 224, 255, 79, 0, 0, 0, 0,
-        0, 0, 0, 192, 255, 47, 0, 0, 0, 0,
-        0, 0, 0, 64, 253, 8, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 80, 219, 239, 141, 2, 0, 0,
-        0, 0, 64, 254, 255, 255, 255, 143, 0, 0,
-        0, 0, 246, 255, 123, 101, 232, 255, 10, 0,
-        0, 80, 255, 44, 0, 0, 0, 249, 143, 0,
-        0, 241, 207, 0, 0, 0, 0, 160, 255, 2,
-        0, 249, 30, 0, 0, 0, 0, 0, 254, 10,
-        16, 255, 7, 0, 0, 34, 0, 0, 246, 15,
-        96, 255, 0, 16, 249, 255, 223, 9, 241, 95,
-        176, 175, 0, 209, 255, 255, 255, 31, 192, 143,
-        240, 111, 0, 250, 111, 16, 250, 15, 144, 191,
-        242, 63, 48, 255, 9, 0, 249, 15, 112, 207,
-        244, 15, 128, 255, 3, 0, 250, 15, 96, 223,
-        245, 15, 192, 239, 0, 0, 251, 14, 96, 239,
-        246, 14, 224, 207, 0, 0, 252, 13, 96, 223,
-        247, 13, 224, 191, 0, 0, 253, 12, 112, 207,
-        246, 14, 224, 207, 0, 0, 255, 12, 128, 175,
-        246, 15, 192, 223, 0, 32, 255, 12, 160, 127,
-        244, 31, 128, 255, 2, 128, 255, 14, 224, 63,
-        242, 79, 32, 255, 76, 246, 158, 127, 248, 13,
-        224, 127, 0, 248, 255, 255, 68, 255, 255, 5,
-        176, 223, 0, 112, 253, 76, 0, 231, 110, 0,
-        80, 255, 4, 0, 0, 0, 0, 0, 0, 0,
-        0, 254, 13, 0, 0, 0, 0, 0, 0, 0,
-        0, 245, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 160, 255, 44, 0, 0, 0, 32, 7, 0,
-        0, 0, 250, 255, 124, 86, 150, 252, 14, 0,
-        0, 0, 112, 255, 255, 255, 255, 255, 11, 0,
-        0, 0, 0, 97, 219, 255, 206, 56, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 249, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 254, 255, 4, 0, 0, 0,
-        0, 0, 0, 48, 255, 255, 9, 0, 0, 0,
-        0, 0, 0, 144, 255, 254, 14, 0, 0, 0,
-        0, 0, 0, 224, 223, 248, 63, 0, 0, 0,
-        0, 0, 0, 243, 159, 244, 159, 0, 0, 0,
-        0, 0, 0, 248, 95, 240, 239, 0, 0, 0,
-        0, 0, 0, 253, 31, 176, 255, 3, 0, 0,
-        0, 0, 32, 255, 11, 96, 255, 8, 0, 0,
-        0, 0, 112, 255, 7, 16, 255, 13, 0, 0,
-        0, 0, 192, 255, 2, 0, 252, 47, 0, 0,
-        0, 0, 242, 223, 0, 0, 247, 143, 0, 0,
-        0, 0, 247, 143, 0, 0, 242, 223, 0, 0,
-        0, 0, 252, 63, 0, 0, 208, 255, 2, 0,
-        0, 16, 255, 207, 204, 204, 236, 255, 7, 0,
-        0, 96, 255, 255, 255, 255, 255, 255, 12, 0,
-        0, 176, 255, 255, 255, 255, 255, 255, 31, 0,
-        0, 241, 255, 0, 0, 0, 0, 249, 127, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 251, 95, 0, 0, 0, 0, 240, 255, 1,
-        0, 255, 31, 0, 0, 0, 0, 160, 255, 6,
-        80, 255, 12, 0, 0, 0, 0, 96, 255, 11,
-        160, 255, 7, 0, 0, 0, 0, 16, 255, 31,
-        240, 255, 2, 0, 0, 0, 0, 0, 252, 95,
-        244, 223, 0, 0, 0, 0, 0, 0, 247, 191,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 252, 255, 255, 255, 239, 156, 4, 0, 0,
-        0, 252, 255, 255, 255, 255, 255, 207, 2, 0,
-        0, 252, 207, 187, 187, 219, 255, 255, 30, 0,
-        0, 252, 95, 0, 0, 0, 129, 255, 159, 0,
-        0, 252, 95, 0, 0, 0, 0, 248, 239, 0,
-        0, 252, 95, 0, 0, 0, 0, 242, 255, 1,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 1,
-        0, 252, 95, 0, 0, 0, 0, 243, 239, 0,
-        0, 252, 95, 0, 0, 0, 0, 251, 127, 0,
-        0, 252, 95, 0, 0, 16, 197, 255, 10, 0,
-        0, 252, 255, 255, 255, 255, 255, 109, 0, 0,
-        0, 252, 255, 255, 255, 255, 207, 23, 0, 0,
-        0, 252, 207, 170, 170, 186, 254, 255, 9, 0,
-        0, 252, 95, 0, 0, 0, 64, 253, 191, 0,
-        0, 252, 95, 0, 0, 0, 0, 225, 255, 5,
-        0, 252, 95, 0, 0, 0, 0, 144, 255, 11,
-        0, 252, 95, 0, 0, 0, 0, 96, 255, 13,
-        0, 252, 95, 0, 0, 0, 0, 112, 255, 13,
-        0, 252, 95, 0, 0, 0, 0, 160, 255, 11,
-        0, 252, 95, 0, 0, 0, 0, 242, 255, 7,
-        0, 252, 95, 0, 0, 0, 64, 254, 239, 1,
-        0, 252, 207, 187, 187, 203, 254, 255, 95, 0,
-        0, 252, 255, 255, 255, 255, 255, 239, 4, 0,
-        0, 252, 255, 255, 255, 255, 190, 6, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 97, 219, 254, 222, 106, 1,
-        0, 0, 0, 128, 255, 255, 255, 255, 255, 95,
-        0, 0, 16, 253, 255, 255, 221, 254, 255, 31,
-        0, 0, 209, 255, 207, 4, 0, 16, 181, 9,
-        0, 0, 251, 255, 7, 0, 0, 0, 0, 0,
-        0, 80, 255, 143, 0, 0, 0, 0, 0, 0,
-        0, 192, 255, 12, 0, 0, 0, 0, 0, 0,
-        0, 242, 255, 5, 0, 0, 0, 0, 0, 0,
-        0, 246, 255, 0, 0, 0, 0, 0, 0, 0,
-        0, 249, 191, 0, 0, 0, 0, 0, 0, 0,
-        0, 251, 143, 0, 0, 0, 0, 0, 0, 0,
-        0, 252, 127, 0, 0, 0, 0, 0, 0, 0,
-        0, 253, 111, 0, 0, 0, 0, 0, 0, 0,
-        0, 252, 127, 0, 0, 0, 0, 0, 0, 0,
-        0, 251, 143, 0, 0, 0, 0, 0, 0, 0,
-        0, 250, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 247, 223, 0, 0, 0, 0, 0, 0, 0,
-        0, 243, 255, 3, 0, 0, 0, 0, 0, 0,
-        0, 224, 255, 10, 0, 0, 0, 0, 0, 0,
-        0, 128, 255, 79, 0, 0, 0, 0, 0, 0,
-        0, 16, 254, 255, 3, 0, 0, 0, 0, 0,
-        0, 0, 244, 255, 159, 3, 0, 0, 99, 6,
-        0, 0, 80, 255, 255, 255, 205, 253, 255, 10,
-        0, 0, 0, 195, 255, 255, 255, 255, 255, 10,
-        0, 0, 0, 0, 148, 236, 255, 206, 106, 1,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 252, 255, 255, 223, 124, 3, 0, 0, 0,
-        0, 252, 255, 255, 255, 255, 191, 2, 0, 0,
-        0, 252, 207, 187, 236, 255, 255, 79, 0, 0,
-        0, 252, 95, 0, 0, 114, 255, 255, 4, 0,
-        0, 252, 95, 0, 0, 0, 209, 255, 30, 0,
-        0, 252, 95, 0, 0, 0, 16, 254, 159, 0,
-        0, 252, 95, 0, 0, 0, 0, 246, 255, 1,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 6,
-        0, 252, 95, 0, 0, 0, 0, 160, 255, 10,
-        0, 252, 95, 0, 0, 0, 0, 112, 255, 13,
-        0, 252, 95, 0, 0, 0, 0, 80, 255, 14,
-        0, 252, 95, 0, 0, 0, 0, 64, 255, 15,
-        0, 252, 95, 0, 0, 0, 0, 64, 255, 15,
-        0, 252, 95, 0, 0, 0, 0, 64, 255, 15,
-        0, 252, 95, 0, 0, 0, 0, 80, 255, 14,
-        0, 252, 95, 0, 0, 0, 0, 128, 255, 12,
-        0, 252, 95, 0, 0, 0, 0, 192, 255, 8,
-        0, 252, 95, 0, 0, 0, 0, 241, 255, 5,
-        0, 252, 95, 0, 0, 0, 0, 249, 239, 0,
-        0, 252, 95, 0, 0, 0, 64, 255, 143, 0,
-        0, 252, 95, 0, 0, 0, 244, 255, 13, 0,
-        0, 252, 95, 0, 0, 180, 255, 255, 3, 0,
-        0, 252, 223, 203, 253, 255, 255, 62, 0, 0,
-        0, 252, 255, 255, 255, 255, 159, 1, 0, 0,
-        0, 252, 255, 255, 222, 106, 1, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 255, 255, 255, 255, 255, 159, 0,
-        0, 96, 255, 255, 255, 255, 255, 255, 159, 0,
-        0, 96, 255, 206, 204, 204, 204, 204, 124, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 255, 255, 255, 255, 255, 15, 0,
-        0, 96, 255, 255, 255, 255, 255, 255, 15, 0,
-        0, 96, 255, 206, 204, 204, 204, 204, 11, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 96, 255, 206, 204, 204, 204, 204, 124, 0,
-        0, 96, 255, 255, 255, 255, 255, 255, 159, 0,
-        0, 96, 255, 255, 255, 255, 255, 255, 159, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 255, 255, 255, 255, 255, 255, 0,
-        0, 0, 255, 255, 255, 255, 255, 255, 255, 0,
-        0, 0, 255, 207, 204, 204, 204, 204, 204, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 255, 255, 255, 255, 255, 111, 0,
-        0, 0, 255, 255, 255, 255, 255, 255, 111, 0,
-        0, 0, 255, 207, 204, 204, 204, 204, 76, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 255, 47, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 182, 253, 239, 123, 2, 0,
-        0, 0, 0, 229, 255, 255, 255, 255, 127, 0,
-        0, 0, 128, 255, 255, 223, 253, 255, 79, 0,
-        0, 0, 247, 255, 77, 0, 0, 147, 12, 0,
-        0, 32, 255, 191, 0, 0, 0, 0, 0, 0,
-        0, 176, 255, 13, 0, 0, 0, 0, 0, 0,
-        0, 241, 255, 5, 0, 0, 0, 0, 0, 0,
-        0, 246, 239, 0, 0, 0, 0, 0, 0, 0,
-        0, 250, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 253, 127, 0, 0, 0, 0, 0, 0, 0,
-        0, 255, 79, 0, 0, 0, 0, 0, 0, 0,
-        0, 255, 79, 0, 0, 160, 204, 204, 204, 1,
-        0, 255, 63, 0, 0, 208, 255, 255, 255, 2,
-        0, 255, 63, 0, 0, 208, 255, 255, 255, 2,
-        0, 255, 79, 0, 0, 0, 0, 224, 255, 2,
-        0, 253, 111, 0, 0, 0, 0, 224, 255, 2,
-        0, 250, 159, 0, 0, 0, 0, 224, 255, 2,
-        0, 247, 223, 0, 0, 0, 0, 224, 255, 2,
-        0, 242, 255, 2, 0, 0, 0, 224, 255, 2,
-        0, 208, 255, 11, 0, 0, 0, 224, 255, 2,
-        0, 64, 255, 127, 0, 0, 0, 224, 255, 2,
-        0, 0, 251, 255, 42, 0, 0, 226, 255, 2,
-        0, 0, 192, 255, 255, 206, 253, 255, 255, 2,
-        0, 0, 16, 250, 255, 255, 255, 255, 239, 1,
-        0, 0, 0, 48, 217, 254, 223, 139, 4, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 255, 255, 255, 255, 255, 255, 255, 2,
-        0, 252, 255, 255, 255, 255, 255, 255, 255, 2,
-        0, 252, 223, 204, 204, 204, 204, 252, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 2,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 64, 255, 255, 255, 255, 255, 255, 10, 0,
-        0, 64, 255, 255, 255, 255, 255, 255, 10, 0,
-        0, 16, 68, 68, 248, 207, 68, 68, 2, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 16, 68, 68, 248, 207, 68, 68, 2, 0,
-        0, 64, 255, 255, 255, 255, 255, 255, 10, 0,
-        0, 64, 255, 255, 255, 255, 255, 255, 10, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 3, 0,
-        0, 0, 0, 0, 0, 0, 240, 255, 1, 0,
-        0, 0, 0, 0, 0, 0, 245, 239, 0, 0,
-        0, 0, 0, 0, 0, 32, 254, 143, 0, 0,
-        0, 104, 2, 0, 0, 230, 255, 30, 0, 0,
-        0, 251, 255, 205, 253, 255, 255, 3, 0, 0,
-        0, 251, 255, 255, 255, 255, 61, 0, 0, 0,
-        0, 113, 235, 255, 223, 90, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 128, 255, 9, 0, 0, 0, 64, 255, 63,
-        0, 128, 255, 9, 0, 0, 0, 225, 255, 6,
-        0, 128, 255, 9, 0, 0, 0, 252, 159, 0,
-        0, 128, 255, 9, 0, 0, 144, 255, 12, 0,
-        0, 128, 255, 9, 0, 0, 246, 239, 1, 0,
-        0, 128, 255, 9, 0, 48, 255, 63, 0, 0,
-        0, 128, 255, 9, 0, 225, 255, 6, 0, 0,
-        0, 128, 255, 9, 0, 251, 175, 0, 0, 0,
-        0, 128, 255, 9, 128, 255, 12, 0, 0, 0,
-        0, 128, 255, 9, 244, 239, 1, 0, 0, 0,
-        0, 128, 255, 25, 254, 79, 0, 0, 0, 0,
-        0, 128, 255, 185, 255, 95, 0, 0, 0, 0,
-        0, 128, 255, 254, 255, 239, 0, 0, 0, 0,
-        0, 128, 255, 255, 198, 255, 9, 0, 0, 0,
-        0, 128, 255, 78, 48, 255, 63, 0, 0, 0,
-        0, 128, 255, 9, 0, 248, 223, 0, 0, 0,
-        0, 128, 255, 9, 0, 224, 255, 7, 0, 0,
-        0, 128, 255, 9, 0, 64, 255, 47, 0, 0,
-        0, 128, 255, 9, 0, 0, 250, 207, 0, 0,
-        0, 128, 255, 9, 0, 0, 241, 255, 6, 0,
-        0, 128, 255, 9, 0, 0, 96, 255, 30, 0,
-        0, 128, 255, 9, 0, 0, 0, 252, 175, 0,
-        0, 128, 255, 9, 0, 0, 0, 243, 255, 4,
-        0, 128, 255, 9, 0, 0, 0, 128, 255, 14,
-        0, 128, 255, 9, 0, 0, 0, 0, 254, 159,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 15, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 223, 221, 221, 221, 221, 109, 0,
-        0, 32, 255, 255, 255, 255, 255, 255, 143, 0,
-        0, 32, 255, 255, 255, 255, 255, 255, 143, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        32, 255, 255, 1, 0, 0, 0, 252, 255, 8,
-        32, 255, 255, 4, 0, 0, 0, 255, 255, 8,
-        32, 255, 255, 7, 0, 0, 32, 255, 255, 8,
-        32, 255, 253, 11, 0, 0, 80, 255, 253, 8,
-        32, 255, 250, 14, 0, 0, 144, 207, 253, 8,
-        32, 255, 248, 31, 0, 0, 192, 159, 254, 8,
-        32, 255, 245, 79, 0, 0, 240, 95, 254, 8,
-        32, 255, 213, 127, 0, 0, 243, 47, 255, 8,
-        32, 255, 165, 191, 0, 0, 246, 14, 255, 8,
-        32, 255, 118, 239, 0, 0, 249, 11, 255, 8,
-        32, 255, 54, 255, 1, 0, 252, 8, 255, 8,
-        32, 255, 6, 255, 4, 0, 255, 4, 255, 8,
-        32, 255, 6, 252, 8, 48, 255, 1, 255, 8,
-        32, 255, 6, 249, 11, 96, 223, 0, 255, 8,
-        32, 255, 6, 246, 14, 144, 175, 0, 255, 8,
-        32, 255, 6, 242, 31, 208, 111, 0, 255, 8,
-        32, 255, 6, 240, 79, 240, 63, 0, 255, 8,
-        32, 255, 6, 192, 143, 243, 15, 0, 255, 8,
-        32, 255, 6, 128, 191, 246, 12, 0, 255, 8,
-        32, 255, 6, 80, 239, 250, 9, 0, 255, 8,
-        32, 255, 6, 16, 255, 254, 5, 0, 255, 8,
-        32, 255, 6, 0, 254, 255, 2, 0, 255, 8,
-        32, 255, 6, 0, 251, 239, 0, 0, 255, 8,
-        32, 255, 6, 0, 247, 191, 0, 0, 255, 8,
-        32, 255, 6, 0, 244, 127, 0, 0, 255, 8,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 252, 255, 1, 0, 0, 0, 192, 255, 2,
-        0, 252, 255, 7, 0, 0, 0, 192, 255, 2,
-        0, 252, 255, 14, 0, 0, 0, 192, 255, 2,
-        0, 252, 255, 111, 0, 0, 0, 192, 255, 2,
-        0, 252, 238, 223, 0, 0, 0, 192, 255, 2,
-        0, 252, 141, 255, 5, 0, 0, 192, 255, 2,
-        0, 252, 30, 255, 12, 0, 0, 192, 255, 2,
-        0, 252, 15, 249, 79, 0, 0, 192, 255, 2,
-        0, 252, 15, 242, 191, 0, 0, 192, 255, 2,
-        0, 252, 15, 160, 255, 3, 0, 192, 255, 2,
-        0, 252, 31, 48, 255, 10, 0, 192, 255, 2,
-        0, 252, 31, 0, 251, 47, 0, 192, 255, 2,
-        0, 252, 31, 0, 244, 159, 0, 192, 255, 2,
-        0, 252, 31, 0, 192, 255, 1, 192, 255, 2,
-        0, 252, 31, 0, 80, 255, 8, 192, 255, 2,
-        0, 252, 31, 0, 0, 253, 15, 176, 255, 2,
-        0, 252, 31, 0, 0, 246, 127, 176, 255, 2,
-        0, 252, 31, 0, 0, 224, 239, 160, 255, 2,
-        0, 252, 31, 0, 0, 112, 255, 166, 255, 2,
-        0, 252, 31, 0, 0, 0, 255, 157, 255, 2,
-        0, 252, 31, 0, 0, 0, 248, 223, 255, 2,
-        0, 252, 31, 0, 0, 0, 241, 255, 255, 2,
-        0, 252, 31, 0, 0, 0, 144, 255, 255, 2,
-        0, 252, 31, 0, 0, 0, 32, 255, 255, 2,
-        0, 252, 31, 0, 0, 0, 0, 250, 255, 2,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 147, 237, 223, 91, 0, 0, 0,
-        0, 0, 161, 255, 255, 255, 255, 61, 0, 0,
-        0, 0, 252, 255, 223, 237, 255, 255, 2, 0,
-        0, 144, 255, 175, 2, 0, 229, 255, 13, 0,
-        0, 242, 255, 9, 0, 0, 32, 255, 127, 0,
-        0, 249, 239, 0, 0, 0, 0, 247, 239, 0,
-        0, 254, 143, 0, 0, 0, 0, 240, 255, 4,
-        32, 255, 63, 0, 0, 0, 0, 176, 255, 8,
-        64, 255, 15, 0, 0, 0, 0, 128, 255, 11,
-        112, 255, 13, 0, 0, 0, 0, 96, 255, 13,
-        128, 255, 11, 0, 0, 0, 0, 64, 255, 15,
-        144, 255, 11, 0, 0, 0, 0, 64, 255, 15,
-        144, 255, 10, 0, 0, 0, 0, 48, 255, 15,
-        144, 255, 11, 0, 0, 0, 0, 64, 255, 15,
-        128, 255, 11, 0, 0, 0, 0, 64, 255, 15,
-        112, 255, 13, 0, 0, 0, 0, 96, 255, 13,
-        64, 255, 15, 0, 0, 0, 0, 128, 255, 11,
-        32, 255, 63, 0, 0, 0, 0, 176, 255, 8,
-        0, 254, 143, 0, 0, 0, 0, 240, 255, 3,
-        0, 249, 239, 0, 0, 0, 0, 247, 239, 0,
-        0, 242, 255, 9, 0, 0, 32, 255, 127, 0,
-        0, 144, 255, 175, 1, 0, 228, 255, 13, 0,
-        0, 0, 252, 255, 223, 237, 255, 239, 2, 0,
-        0, 0, 161, 255, 255, 255, 255, 44, 0, 0,
-        0, 0, 0, 147, 237, 223, 91, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 241, 255, 255, 255, 239, 123, 1, 0, 0,
-        0, 241, 255, 255, 255, 255, 255, 143, 0, 0,
-        0, 241, 255, 187, 187, 236, 255, 255, 10, 0,
-        0, 241, 255, 0, 0, 0, 212, 255, 95, 0,
-        0, 241, 255, 0, 0, 0, 0, 253, 191, 0,
-        0, 241, 255, 0, 0, 0, 0, 245, 255, 0,
-        0, 241, 255, 0, 0, 0, 0, 242, 255, 1,
-        0, 241, 255, 0, 0, 0, 0, 241, 255, 2,
-        0, 241, 255, 0, 0, 0, 0, 242, 255, 0,
-        0, 241, 255, 0, 0, 0, 0, 246, 223, 0,
-        0, 241, 255, 0, 0, 0, 16, 253, 143, 0,
-        0, 241, 255, 0, 0, 0, 212, 255, 31, 0,
-        0, 241, 255, 187, 187, 252, 255, 255, 4, 0,
-        0, 241, 255, 255, 255, 255, 255, 61, 0, 0,
-        0, 241, 255, 255, 255, 206, 73, 0, 0, 0,
-        0, 241, 255, 0, 0, 0, 0, 0, 0, 0,
-        0, 241, 255, 0, 0, 0, 0, 0, 0, 0,
-        0, 241, 255, 0, 0, 0, 0, 0, 0, 0,
-        0, 241, 255, 0, 0, 0, 0, 0, 0, 0,
-        0, 241, 255, 0, 0, 0, 0, 0, 0, 0,
-        0, 241, 255, 0, 0, 0, 0, 0, 0, 0,
-        0, 241, 255, 0, 0, 0, 0, 0, 0, 0,
-        0, 241, 255, 0, 0, 0, 0, 0, 0, 0,
-        0, 241, 255, 0, 0, 0, 0, 0, 0, 0,
-        0, 241, 255, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 147, 237, 223, 91, 0, 0, 0,
-        0, 0, 161, 255, 255, 255, 255, 61, 0, 0,
-        0, 0, 252, 255, 223, 237, 255, 239, 2, 0,
-        0, 144, 255, 175, 2, 0, 229, 255, 13, 0,
-        0, 242, 255, 9, 0, 0, 32, 255, 127, 0,
-        0, 249, 239, 0, 0, 0, 0, 247, 239, 0,
-        0, 254, 143, 0, 0, 0, 0, 240, 255, 3,
-        32, 255, 63, 0, 0, 0, 0, 176, 255, 8,
-        64, 255, 15, 0, 0, 0, 0, 128, 255, 11,
-        112, 255, 13, 0, 0, 0, 0, 96, 255, 13,
-        128, 255, 11, 0, 0, 0, 0, 64, 255, 15,
-        144, 255, 11, 0, 0, 0, 0, 64, 255, 15,
-        144, 255, 10, 0, 0, 0, 0, 48, 255, 15,
-        144, 255, 11, 0, 0, 0, 0, 64, 255, 15,
-        128, 255, 11, 0, 0, 0, 0, 64, 255, 15,
-        112, 255, 13, 0, 0, 0, 0, 96, 255, 13,
-        64, 255, 15, 0, 0, 0, 0, 128, 255, 11,
-        32, 255, 63, 0, 0, 0, 0, 176, 255, 8,
-        0, 254, 143, 0, 0, 0, 0, 240, 255, 4,
-        0, 249, 239, 0, 0, 0, 0, 247, 239, 0,
-        0, 242, 255, 9, 0, 0, 32, 255, 143, 0,
-        0, 144, 255, 175, 1, 0, 228, 255, 30, 0,
-        0, 0, 252, 255, 223, 237, 255, 255, 3, 0,
-        0, 0, 161, 255, 255, 255, 255, 62, 0, 0,
-        0, 0, 0, 147, 237, 255, 255, 8, 0, 0,
-        0, 0, 0, 0, 0, 48, 255, 47, 0, 0,
-        0, 0, 0, 0, 0, 0, 251, 223, 0, 0,
-        0, 0, 0, 0, 0, 0, 242, 255, 11, 0,
-        0, 0, 0, 0, 0, 0, 96, 255, 207, 1,
-        0, 0, 0, 0, 0, 0, 0, 248, 255, 10,
-        0, 0, 0, 0, 0, 0, 0, 112, 223, 1,
-        0, 0, 0, 0, 0, 0, 0, 0, 37, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 224, 255, 255, 255, 206, 73, 0, 0, 0,
-        0, 224, 255, 255, 255, 255, 255, 61, 0, 0,
-        0, 224, 255, 188, 187, 253, 255, 255, 3, 0,
-        0, 224, 255, 3, 0, 0, 246, 255, 13, 0,
-        0, 224, 255, 3, 0, 0, 64, 255, 63, 0,
-        0, 224, 255, 3, 0, 0, 0, 253, 127, 0,
-        0, 224, 255, 3, 0, 0, 0, 249, 159, 0,
-        0, 224, 255, 3, 0, 0, 0, 249, 159, 0,
-        0, 224, 255, 3, 0, 0, 0, 251, 143, 0,
-        0, 224, 255, 3, 0, 0, 0, 254, 79, 0,
-        0, 224, 255, 3, 0, 0, 144, 255, 13, 0,
-        0, 224, 255, 3, 0, 81, 252, 255, 3, 0,
-        0, 224, 255, 255, 255, 255, 255, 78, 0, 0,
-        0, 224, 255, 255, 255, 255, 143, 1, 0, 0,
-        0, 224, 255, 187, 187, 255, 79, 0, 0, 0,
-        0, 224, 255, 3, 0, 248, 223, 0, 0, 0,
-        0, 224, 255, 3, 0, 224, 255, 6, 0, 0,
-        0, 224, 255, 3, 0, 96, 255, 30, 0, 0,
-        0, 224, 255, 3, 0, 0, 253, 159, 0, 0,
-        0, 224, 255, 3, 0, 0, 244, 255, 3, 0,
-        0, 224, 255, 3, 0, 0, 176, 255, 12, 0,
-        0, 224, 255, 3, 0, 0, 32, 255, 95, 0,
-        0, 224, 255, 3, 0, 0, 0, 249, 239, 0,
-        0, 224, 255, 3, 0, 0, 0, 241, 255, 8,
-        0, 224, 255, 3, 0, 0, 0, 112, 255, 47,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 148, 237, 255, 206, 89, 1, 0,
-        0, 0, 194, 255, 255, 255, 255, 255, 95, 0,
-        0, 32, 254, 255, 223, 220, 255, 255, 31, 0,
-        0, 176, 255, 94, 0, 0, 32, 182, 11, 0,
-        0, 241, 255, 2, 0, 0, 0, 0, 0, 0,
-        0, 244, 191, 0, 0, 0, 0, 0, 0, 0,
-        0, 245, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 243, 223, 0, 0, 0, 0, 0, 0, 0,
-        0, 240, 255, 5, 0, 0, 0, 0, 0, 0,
-        0, 128, 255, 127, 0, 0, 0, 0, 0, 0,
-        0, 0, 252, 255, 126, 1, 0, 0, 0, 0,
-        0, 0, 160, 255, 255, 191, 4, 0, 0, 0,
-        0, 0, 0, 196, 255, 255, 223, 6, 0, 0,
-        0, 0, 0, 0, 146, 254, 255, 207, 1, 0,
-        0, 0, 0, 0, 0, 80, 253, 255, 13, 0,
-        0, 0, 0, 0, 0, 0, 96, 255, 127, 0,
-        0, 0, 0, 0, 0, 0, 0, 246, 223, 0,
-        0, 0, 0, 0, 0, 0, 0, 241, 255, 0,
-        0, 0, 0, 0, 0, 0, 0, 240, 255, 0,
-        0, 0, 0, 0, 0, 0, 0, 242, 239, 0,
-        0, 0, 0, 0, 0, 0, 0, 250, 175, 0,
-        0, 186, 38, 0, 0, 0, 179, 255, 63, 0,
-        0, 250, 255, 223, 188, 252, 255, 255, 7, 0,
-        0, 249, 255, 255, 255, 255, 255, 94, 0, 0,
-        0, 64, 184, 253, 255, 222, 90, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        80, 255, 255, 255, 255, 255, 255, 255, 255, 10,
-        80, 255, 255, 255, 255, 255, 255, 255, 255, 10,
-        48, 204, 204, 204, 253, 255, 204, 204, 204, 8,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 254, 63, 0, 0, 0, 0, 208, 255, 4,
-        0, 253, 79, 0, 0, 0, 0, 224, 255, 3,
-        0, 251, 127, 0, 0, 0, 0, 241, 255, 1,
-        0, 247, 223, 0, 0, 0, 0, 247, 239, 0,
-        0, 242, 255, 6, 0, 0, 16, 254, 143, 0,
-        0, 176, 255, 127, 1, 0, 212, 255, 31, 0,
-        0, 16, 254, 255, 223, 236, 255, 255, 5, 0,
-        0, 0, 210, 255, 255, 255, 255, 94, 0, 0,
-        0, 0, 0, 165, 254, 239, 123, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        244, 239, 0, 0, 0, 0, 0, 0, 249, 191,
-        240, 255, 4, 0, 0, 0, 0, 0, 254, 95,
-        160, 255, 9, 0, 0, 0, 0, 48, 255, 31,
-        80, 255, 14, 0, 0, 0, 0, 128, 255, 11,
-        0, 255, 47, 0, 0, 0, 0, 208, 255, 6,
-        0, 250, 127, 0, 0, 0, 0, 242, 255, 1,
-        0, 245, 207, 0, 0, 0, 0, 247, 191, 0,
-        0, 240, 255, 1, 0, 0, 0, 252, 111, 0,
-        0, 176, 255, 6, 0, 0, 16, 255, 31, 0,
-        0, 96, 255, 11, 0, 0, 80, 255, 12, 0,
-        0, 16, 255, 31, 0, 0, 160, 255, 7, 0,
-        0, 0, 251, 95, 0, 0, 240, 255, 1, 0,
-        0, 0, 246, 175, 0, 0, 244, 207, 0, 0,
-        0, 0, 241, 255, 0, 0, 249, 127, 0, 0,
-        0, 0, 192, 255, 4, 0, 254, 47, 0, 0,
-        0, 0, 96, 255, 9, 48, 255, 13, 0, 0,
-        0, 0, 16, 255, 14, 128, 255, 7, 0, 0,
-        0, 0, 0, 252, 47, 192, 255, 2, 0, 0,
-        0, 0, 0, 247, 111, 241, 223, 0, 0, 0,
-        0, 0, 0, 242, 175, 245, 143, 0, 0, 0,
-        0, 0, 0, 208, 239, 249, 63, 0, 0, 0,
-        0, 0, 0, 112, 255, 254, 13, 0, 0, 0,
-        0, 0, 0, 32, 255, 255, 8, 0, 0, 0,
-        0, 0, 0, 0, 253, 255, 3, 0, 0, 0,
-        0, 0, 0, 0, 248, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        254, 14, 0, 0, 0, 0, 0, 0, 128, 255,
-        252, 15, 0, 0, 0, 0, 0, 0, 144, 255,
-        250, 47, 0, 0, 0, 0, 0, 0, 176, 255,
-        247, 79, 0, 0, 0, 0, 0, 0, 208, 255,
-        245, 111, 0, 0, 0, 0, 0, 0, 224, 223,
-        243, 143, 0, 0, 0, 0, 0, 0, 240, 191,
-        241, 191, 0, 0, 0, 0, 0, 0, 241, 159,
-        240, 223, 0, 0, 48, 35, 0, 0, 243, 127,
-        208, 255, 0, 0, 246, 239, 0, 0, 245, 95,
-        176, 255, 1, 0, 251, 255, 2, 0, 246, 63,
-        128, 255, 3, 0, 255, 254, 7, 0, 248, 31,
-        96, 255, 5, 48, 255, 248, 11, 0, 250, 15,
-        64, 255, 7, 112, 207, 244, 15, 0, 251, 13,
-        32, 255, 9, 192, 159, 240, 79, 0, 253, 11,
-        0, 255, 10, 240, 79, 192, 143, 0, 255, 9,
-        0, 254, 12, 244, 15, 128, 207, 0, 255, 7,
-        0, 251, 13, 248, 12, 64, 255, 33, 255, 5,
-        0, 249, 15, 252, 8, 0, 255, 69, 255, 3,
-        0, 247, 31, 255, 4, 0, 251, 89, 255, 1,
-        0, 245, 111, 255, 0, 0, 246, 125, 255, 0,
-        0, 243, 175, 191, 0, 0, 242, 175, 223, 0,
-        0, 241, 239, 127, 0, 0, 208, 239, 191, 0,
-        0, 224, 255, 63, 0, 0, 144, 255, 159, 0,
-        0, 192, 255, 15, 0, 0, 80, 255, 127, 0,
-        0, 160, 255, 11, 0, 0, 0, 255, 95, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        80, 255, 30, 0, 0, 0, 0, 128, 255, 9,
-        0, 251, 143, 0, 0, 0, 0, 242, 239, 1,
-        0, 242, 255, 2, 0, 0, 0, 251, 111, 0,
-        0, 128, 255, 11, 0, 0, 64, 255, 13, 0,
-        0, 0, 254, 79, 0, 0, 208, 255, 3, 0,
-        0, 0, 246, 223, 0, 0, 247, 175, 0, 0,
-        0, 0, 192, 255, 6, 16, 255, 31, 0, 0,
-        0, 0, 48, 255, 30, 144, 255, 7, 0, 0,
-        0, 0, 0, 249, 159, 243, 223, 0, 0, 0,
-        0, 0, 0, 241, 255, 253, 79, 0, 0, 0,
-        0, 0, 0, 112, 255, 255, 11, 0, 0, 0,
-        0, 0, 0, 0, 253, 255, 2, 0, 0, 0,
-        0, 0, 0, 0, 254, 255, 4, 0, 0, 0,
-        0, 0, 0, 112, 255, 255, 13, 0, 0, 0,
-        0, 0, 0, 241, 223, 251, 127, 0, 0, 0,
-        0, 0, 0, 250, 95, 242, 239, 1, 0, 0,
-        0, 0, 48, 255, 12, 160, 255, 9, 0, 0,
-        0, 0, 192, 255, 3, 16, 255, 47, 0, 0,
-        0, 0, 246, 175, 0, 0, 248, 191, 0, 0,
-        0, 0, 254, 31, 0, 0, 225, 255, 5, 0,
-        0, 128, 255, 8, 0, 0, 112, 255, 13, 0,
-        0, 242, 239, 0, 0, 0, 0, 254, 127, 0,
-        0, 250, 111, 0, 0, 0, 0, 245, 255, 1,
-        64, 255, 13, 0, 0, 0, 0, 208, 255, 9,
-        208, 255, 4, 0, 0, 0, 0, 64, 255, 63,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        243, 255, 2, 0, 0, 0, 0, 0, 252, 143,
-        160, 255, 10, 0, 0, 0, 0, 64, 255, 31,
-        32, 255, 47, 0, 0, 0, 0, 192, 255, 8,
-        0, 249, 175, 0, 0, 0, 0, 244, 239, 0,
-        0, 241, 255, 2, 0, 0, 0, 252, 127, 0,
-        0, 144, 255, 10, 0, 0, 64, 255, 14, 0,
-        0, 16, 255, 47, 0, 0, 192, 255, 6, 0,
-        0, 0, 248, 175, 0, 0, 244, 223, 0, 0,
-        0, 0, 225, 255, 2, 0, 252, 95, 0, 0,
-        0, 0, 112, 255, 10, 64, 255, 12, 0, 0,
-        0, 0, 0, 254, 47, 192, 255, 4, 0, 0,
-        0, 0, 0, 246, 175, 244, 191, 0, 0, 0,
-        0, 0, 0, 224, 255, 253, 63, 0, 0, 0,
-        0, 0, 0, 96, 255, 255, 10, 0, 0, 0,
-        0, 0, 0, 0, 253, 255, 2, 0, 0, 0,
-        0, 0, 0, 0, 246, 207, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 245, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 255, 255, 255, 255, 255, 255, 255, 255, 5,
-        0, 255, 255, 255, 255, 255, 255, 255, 255, 5,
-        0, 220, 221, 221, 221, 221, 221, 253, 255, 3,
-        0, 0, 0, 0, 0, 0, 0, 249, 175, 0,
-        0, 0, 0, 0, 0, 0, 64, 255, 30, 0,
-        0, 0, 0, 0, 0, 0, 224, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 249, 175, 0, 0,
-        0, 0, 0, 0, 0, 64, 255, 30, 0, 0,
-        0, 0, 0, 0, 0, 224, 255, 4, 0, 0,
-        0, 0, 0, 0, 0, 249, 175, 0, 0, 0,
-        0, 0, 0, 0, 64, 255, 30, 0, 0, 0,
-        0, 0, 0, 0, 224, 255, 4, 0, 0, 0,
-        0, 0, 0, 0, 249, 159, 0, 0, 0, 0,
-        0, 0, 0, 64, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 224, 255, 4, 0, 0, 0, 0,
-        0, 0, 0, 249, 159, 0, 0, 0, 0, 0,
-        0, 0, 64, 255, 14, 0, 0, 0, 0, 0,
-        0, 0, 224, 255, 4, 0, 0, 0, 0, 0,
-        0, 0, 249, 159, 0, 0, 0, 0, 0, 0,
-        0, 64, 255, 14, 0, 0, 0, 0, 0, 0,
-        0, 224, 255, 4, 0, 0, 0, 0, 0, 0,
-        0, 250, 159, 0, 0, 0, 0, 0, 0, 0,
-        48, 255, 223, 221, 221, 221, 221, 221, 221, 8,
-        80, 255, 255, 255, 255, 255, 255, 255, 255, 10,
-        80, 255, 255, 255, 255, 255, 255, 255, 255, 10,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 255, 255, 255, 159, 0, 0,
-        0, 0, 0, 242, 255, 255, 255, 159, 0, 0,
-        0, 0, 0, 242, 239, 136, 136, 88, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 255, 255, 255, 159, 0, 0,
-        0, 0, 0, 242, 255, 255, 255, 159, 0, 0,
-        0, 0, 0, 129, 136, 136, 136, 88, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 64, 255, 10, 0, 0, 0, 0, 0, 0,
-        0, 0, 253, 31, 0, 0, 0, 0, 0, 0,
-        0, 0, 247, 127, 0, 0, 0, 0, 0, 0,
-        0, 0, 241, 239, 0, 0, 0, 0, 0, 0,
-        0, 0, 160, 255, 5, 0, 0, 0, 0, 0,
-        0, 0, 48, 255, 11, 0, 0, 0, 0, 0,
-        0, 0, 0, 252, 47, 0, 0, 0, 0, 0,
-        0, 0, 0, 246, 159, 0, 0, 0, 0, 0,
-        0, 0, 0, 224, 255, 0, 0, 0, 0, 0,
-        0, 0, 0, 128, 255, 6, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 13, 0, 0, 0, 0,
-        0, 0, 0, 0, 251, 63, 0, 0, 0, 0,
-        0, 0, 0, 0, 244, 175, 0, 0, 0, 0,
-        0, 0, 0, 0, 208, 255, 1, 0, 0, 0,
-        0, 0, 0, 0, 112, 255, 7, 0, 0, 0,
-        0, 0, 0, 0, 16, 255, 14, 0, 0, 0,
-        0, 0, 0, 0, 0, 250, 95, 0, 0, 0,
-        0, 0, 0, 0, 0, 243, 191, 0, 0, 0,
-        0, 0, 0, 0, 0, 192, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 96, 255, 9, 0, 0,
-        0, 0, 0, 0, 0, 0, 254, 15, 0, 0,
-        0, 0, 0, 0, 0, 0, 248, 111, 0, 0,
-        0, 0, 0, 0, 0, 0, 242, 223, 0, 0,
-        0, 0, 0, 0, 0, 0, 176, 255, 3, 0,
-        0, 0, 0, 0, 0, 0, 64, 255, 10, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 243, 255, 255, 255, 159, 0, 0, 0,
-        0, 0, 243, 255, 255, 255, 159, 0, 0, 0,
-        0, 0, 129, 136, 136, 251, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 0, 0, 0, 247, 159, 0, 0, 0,
-        0, 0, 243, 255, 255, 255, 159, 0, 0, 0,
-        0, 0, 243, 255, 255, 255, 159, 0, 0, 0,
-        0, 0, 129, 136, 136, 136, 72, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 129, 5, 0, 0, 0, 0,
-        0, 0, 0, 0, 249, 31, 0, 0, 0, 0,
-        0, 0, 0, 16, 255, 159, 0, 0, 0, 0,
-        0, 0, 0, 112, 255, 255, 2, 0, 0, 0,
-        0, 0, 0, 224, 111, 253, 10, 0, 0, 0,
-        0, 0, 0, 246, 14, 245, 47, 0, 0, 0,
-        0, 0, 0, 253, 8, 208, 175, 0, 0, 0,
-        0, 0, 64, 255, 2, 80, 255, 3, 0, 0,
-        0, 0, 192, 191, 0, 0, 253, 11, 0, 0,
-        0, 0, 243, 79, 0, 0, 245, 63, 0, 0,
-        0, 0, 250, 13, 0, 0, 208, 191, 0, 0,
-        0, 32, 255, 6, 0, 0, 80, 255, 3, 0,
-        0, 144, 255, 0, 0, 0, 0, 253, 12, 0,
-        0, 241, 143, 0, 0, 0, 0, 245, 79, 0,
-        0, 247, 47, 0, 0, 0, 0, 208, 207, 0,
-        0, 254, 11, 0, 0, 0, 0, 80, 255, 4,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        102, 102, 102, 102, 102, 102, 102, 102, 102, 102,
-        255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-        255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 112, 136, 5, 0, 0, 0, 0,
-        0, 0, 0, 144, 255, 31, 0, 0, 0, 0,
-        0, 0, 0, 0, 252, 159, 0, 0, 0, 0,
-        0, 0, 0, 0, 226, 255, 2, 0, 0, 0,
-        0, 0, 0, 0, 32, 254, 12, 0, 0, 0,
-        0, 0, 0, 0, 0, 226, 95, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 48, 184, 254, 239, 123, 1, 0, 0,
-        0, 48, 252, 255, 255, 255, 255, 95, 0, 0,
-        0, 16, 255, 239, 155, 201, 255, 255, 2, 0,
-        0, 0, 153, 3, 0, 0, 227, 255, 9, 0,
-        0, 0, 0, 0, 0, 0, 80, 255, 13, 0,
-        0, 0, 0, 0, 0, 0, 0, 255, 15, 0,
-        0, 0, 0, 0, 0, 0, 0, 255, 15, 0,
-        0, 0, 0, 0, 0, 16, 17, 254, 15, 0,
-        0, 0, 48, 201, 254, 255, 255, 255, 15, 0,
-        0, 16, 251, 255, 255, 255, 255, 255, 15, 0,
-        0, 208, 255, 143, 20, 0, 0, 255, 15, 0,
-        0, 246, 255, 2, 0, 0, 0, 255, 15, 0,
-        0, 250, 143, 0, 0, 0, 0, 255, 15, 0,
-        0, 251, 111, 0, 0, 0, 64, 255, 15, 0,
-        0, 250, 143, 0, 0, 0, 192, 255, 15, 0,
-        0, 247, 239, 2, 0, 16, 251, 255, 15, 0,
-        0, 241, 255, 175, 136, 251, 191, 251, 15, 0,
-        0, 64, 255, 255, 255, 255, 10, 247, 15, 0,
-        0, 0, 146, 253, 239, 74, 0, 243, 15, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 159, 16, 216, 255, 141, 1, 0, 0,
-        0, 246, 159, 228, 255, 255, 255, 78, 0, 0,
-        0, 246, 191, 255, 174, 185, 255, 255, 3, 0,
-        0, 246, 255, 111, 0, 0, 193, 255, 12, 0,
-        0, 246, 255, 8, 0, 0, 16, 255, 79, 0,
-        0, 246, 255, 2, 0, 0, 0, 249, 175, 0,
-        0, 246, 239, 0, 0, 0, 0, 244, 223, 0,
-        0, 246, 191, 0, 0, 0, 0, 242, 255, 0,
-        0, 246, 175, 0, 0, 0, 0, 240, 255, 1,
-        0, 246, 175, 0, 0, 0, 0, 240, 255, 1,
-        0, 246, 175, 0, 0, 0, 0, 241, 255, 1,
-        0, 246, 207, 0, 0, 0, 0, 242, 255, 0,
-        0, 246, 239, 0, 0, 0, 0, 245, 223, 0,
-        0, 246, 255, 2, 0, 0, 0, 249, 159, 0,
-        0, 246, 255, 9, 0, 0, 16, 255, 79, 0,
-        0, 246, 255, 127, 0, 0, 193, 255, 12, 0,
-        0, 246, 223, 255, 174, 185, 255, 255, 3, 0,
-        0, 246, 63, 247, 255, 255, 255, 78, 0, 0,
-        0, 246, 15, 32, 217, 255, 141, 1, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 16, 183, 254, 223, 156, 5, 0,
-        0, 0, 0, 248, 255, 255, 255, 255, 111, 0,
-        0, 0, 176, 255, 255, 206, 220, 255, 15, 0,
-        0, 0, 248, 255, 42, 0, 0, 80, 6, 0,
-        0, 16, 255, 143, 0, 0, 0, 0, 0, 0,
-        0, 128, 255, 13, 0, 0, 0, 0, 0, 0,
-        0, 192, 255, 6, 0, 0, 0, 0, 0, 0,
-        0, 240, 255, 2, 0, 0, 0, 0, 0, 0,
-        0, 241, 255, 0, 0, 0, 0, 0, 0, 0,
-        0, 241, 255, 0, 0, 0, 0, 0, 0, 0,
-        0, 241, 255, 0, 0, 0, 0, 0, 0, 0,
-        0, 240, 255, 2, 0, 0, 0, 0, 0, 0,
-        0, 208, 255, 6, 0, 0, 0, 0, 0, 0,
-        0, 144, 255, 12, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 111, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 255, 25, 0, 0, 32, 55, 0,
-        0, 0, 192, 255, 255, 189, 203, 255, 79, 0,
-        0, 0, 0, 249, 255, 255, 255, 255, 79, 0,
-        0, 0, 0, 32, 200, 254, 239, 156, 4, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 243, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 243, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 243, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 243, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 243, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 243, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 243, 207, 0,
-        0, 0, 0, 181, 254, 174, 4, 243, 207, 0,
-        0, 0, 176, 255, 255, 255, 159, 242, 207, 0,
-        0, 0, 251, 255, 156, 201, 255, 250, 207, 0,
-        0, 96, 255, 95, 0, 0, 227, 255, 207, 0,
-        0, 208, 255, 7, 0, 0, 48, 255, 207, 0,
-        0, 243, 255, 0, 0, 0, 0, 252, 207, 0,
-        0, 246, 191, 0, 0, 0, 0, 248, 207, 0,
-        0, 249, 143, 0, 0, 0, 0, 245, 207, 0,
-        0, 250, 127, 0, 0, 0, 0, 244, 207, 0,
-        0, 251, 111, 0, 0, 0, 0, 243, 207, 0,
-        0, 250, 127, 0, 0, 0, 0, 244, 207, 0,
-        0, 249, 143, 0, 0, 0, 0, 245, 207, 0,
-        0, 247, 191, 0, 0, 0, 0, 247, 207, 0,
-        0, 243, 255, 0, 0, 0, 0, 251, 207, 0,
-        0, 224, 255, 7, 0, 0, 32, 255, 207, 0,
-        0, 112, 255, 95, 0, 0, 210, 255, 207, 0,
-        0, 0, 252, 255, 156, 201, 255, 252, 207, 0,
-        0, 0, 193, 255, 255, 255, 191, 224, 207, 0,
-        0, 0, 0, 181, 254, 190, 5, 176, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 113, 236, 239, 124, 0, 0, 0,
-        0, 0, 80, 254, 255, 255, 255, 78, 0, 0,
-        0, 0, 246, 255, 173, 169, 254, 255, 4, 0,
-        0, 48, 255, 95, 0, 0, 128, 255, 31, 0,
-        0, 176, 255, 5, 0, 0, 0, 251, 143, 0,
-        0, 242, 223, 0, 0, 0, 0, 245, 223, 0,
-        0, 246, 143, 0, 0, 0, 0, 242, 255, 1,
-        0, 250, 191, 136, 136, 136, 136, 248, 255, 2,
-        0, 251, 255, 255, 255, 255, 255, 255, 255, 3,
-        0, 252, 255, 255, 255, 255, 255, 255, 255, 3,
-        0, 251, 111, 0, 0, 0, 0, 0, 0, 0,
-        0, 250, 127, 0, 0, 0, 0, 0, 0, 0,
-        0, 247, 191, 0, 0, 0, 0, 0, 0, 0,
-        0, 242, 255, 1, 0, 0, 0, 0, 0, 0,
-        0, 176, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 32, 255, 191, 2, 0, 0, 81, 59, 0,
-        0, 0, 245, 255, 223, 187, 236, 255, 63, 0,
-        0, 0, 48, 253, 255, 255, 255, 255, 47, 0,
-        0, 0, 0, 80, 218, 255, 206, 89, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 181, 254, 239, 157, 3,
-        0, 0, 0, 0, 192, 255, 255, 255, 255, 3,
-        0, 0, 0, 0, 248, 255, 156, 153, 219, 0,
-        0, 0, 0, 0, 254, 127, 0, 0, 0, 0,
-        0, 0, 0, 16, 255, 15, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 163, 219, 255, 255, 255, 255, 255, 95, 0,
-        0, 246, 255, 255, 255, 255, 255, 255, 95, 0,
-        0, 82, 85, 101, 255, 94, 85, 85, 21, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 181, 254, 255, 255, 255, 255, 11,
-        0, 0, 193, 255, 255, 255, 255, 255, 255, 11,
-        0, 0, 252, 239, 71, 133, 255, 223, 105, 3,
-        0, 80, 255, 13, 0, 0, 243, 239, 0, 0,
-        0, 160, 255, 5, 0, 0, 176, 255, 4, 0,
-        0, 208, 255, 2, 0, 0, 128, 255, 7, 0,
-        0, 208, 255, 2, 0, 0, 112, 255, 7, 0,
-        0, 176, 255, 4, 0, 0, 160, 255, 5, 0,
-        0, 96, 255, 12, 0, 0, 242, 255, 1, 0,
-        0, 0, 253, 207, 37, 99, 254, 143, 0, 0,
-        0, 0, 210, 255, 255, 255, 255, 9, 0, 0,
-        0, 0, 16, 254, 255, 223, 58, 0, 0, 0,
-        0, 0, 209, 175, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 14, 0, 0, 0, 0, 0, 0,
-        0, 0, 253, 12, 0, 0, 0, 0, 0, 0,
-        0, 0, 253, 63, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 255, 205, 204, 188, 73, 0, 0,
-        0, 0, 192, 255, 255, 255, 255, 255, 28, 0,
-        0, 48, 251, 255, 255, 255, 255, 255, 207, 0,
-        0, 245, 207, 20, 0, 0, 32, 246, 255, 3,
-        0, 254, 30, 0, 0, 0, 0, 128, 255, 6,
-        48, 255, 10, 0, 0, 0, 0, 96, 255, 6,
-        64, 255, 9, 0, 0, 0, 0, 176, 255, 4,
-        16, 255, 47, 0, 0, 0, 0, 249, 239, 0,
-        0, 249, 255, 122, 85, 134, 251, 255, 63, 0,
-        0, 160, 255, 255, 255, 255, 255, 207, 3, 0,
-        0, 0, 131, 236, 255, 223, 139, 3, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 245, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 245, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 245, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 245, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 245, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 245, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 245, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 245, 175, 16, 199, 254, 190, 6, 0, 0,
-        0, 245, 159, 227, 255, 255, 255, 207, 1, 0,
-        0, 245, 175, 254, 191, 169, 253, 255, 11, 0,
-        0, 245, 255, 191, 1, 0, 112, 255, 63, 0,
-        0, 245, 255, 12, 0, 0, 0, 252, 127, 0,
-        0, 245, 255, 4, 0, 0, 0, 246, 175, 0,
-        0, 245, 255, 0, 0, 0, 0, 244, 191, 0,
-        0, 245, 207, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 191, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 176, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 243, 255, 3, 0, 0, 0,
-        0, 0, 0, 0, 242, 255, 2, 0, 0, 0,
-        0, 0, 0, 0, 144, 141, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 247, 255, 255, 255, 0, 0, 0, 0,
-        0, 0, 247, 255, 255, 255, 0, 0, 0, 0,
-        0, 0, 49, 100, 247, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 255, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 255, 0, 0, 0, 0,
-        0, 48, 68, 68, 244, 255, 68, 68, 52, 0,
-        0, 176, 255, 255, 255, 255, 255, 255, 223, 0,
-        0, 176, 255, 255, 255, 255, 255, 255, 223, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 193, 143, 0, 0, 0,
-        0, 0, 0, 0, 0, 245, 255, 0, 0, 0,
-        0, 0, 0, 0, 0, 245, 255, 0, 0, 0,
-        0, 0, 0, 0, 0, 160, 125, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 253, 255, 255, 255, 255, 2, 0, 0,
-        0, 0, 253, 255, 255, 255, 255, 2, 0, 0,
-        0, 0, 67, 68, 68, 228, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 208, 255, 2, 0, 0,
-        0, 0, 0, 0, 0, 224, 255, 1, 0, 0,
-        0, 0, 0, 0, 0, 243, 239, 0, 0, 0,
-        0, 1, 0, 0, 48, 253, 159, 0, 0, 0,
-        0, 236, 171, 153, 252, 255, 30, 0, 0, 0,
-        0, 252, 255, 255, 255, 223, 2, 0, 0, 0,
-        0, 165, 237, 255, 190, 6, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 112, 255, 9, 0, 0, 0, 0, 0, 0,
-        0, 112, 255, 9, 0, 0, 0, 0, 0, 0,
-        0, 112, 255, 9, 0, 0, 0, 0, 0, 0,
-        0, 112, 255, 9, 0, 0, 0, 0, 0, 0,
-        0, 112, 255, 9, 0, 0, 0, 0, 0, 0,
-        0, 112, 255, 9, 0, 0, 0, 0, 0, 0,
-        0, 112, 255, 9, 0, 0, 0, 0, 0, 0,
-        0, 112, 255, 9, 0, 0, 0, 245, 255, 5,
-        0, 112, 255, 9, 0, 0, 64, 255, 95, 0,
-        0, 112, 255, 9, 0, 0, 244, 255, 5, 0,
-        0, 112, 255, 9, 0, 64, 255, 95, 0, 0,
-        0, 112, 255, 8, 0, 243, 255, 5, 0, 0,
-        0, 112, 255, 8, 48, 255, 95, 0, 0, 0,
-        0, 112, 255, 7, 226, 255, 5, 0, 0, 0,
-        0, 112, 255, 22, 254, 95, 0, 0, 0, 0,
-        0, 112, 255, 198, 255, 13, 0, 0, 0, 0,
-        0, 112, 255, 254, 255, 175, 0, 0, 0, 0,
-        0, 112, 255, 255, 248, 255, 7, 0, 0, 0,
-        0, 112, 255, 44, 80, 255, 79, 0, 0, 0,
-        0, 112, 255, 9, 0, 248, 239, 2, 0, 0,
-        0, 112, 255, 9, 0, 160, 255, 13, 0, 0,
-        0, 112, 255, 9, 0, 0, 253, 175, 0, 0,
-        0, 112, 255, 9, 0, 0, 226, 255, 7, 0,
-        0, 112, 255, 9, 0, 0, 64, 255, 79, 0,
-        0, 112, 255, 9, 0, 0, 0, 247, 255, 2,
-        0, 112, 255, 9, 0, 0, 0, 160, 255, 13,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 255, 255, 239, 0, 0, 0, 0,
-        0, 0, 249, 255, 255, 239, 0, 0, 0, 0,
-        0, 0, 16, 67, 247, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 0, 0, 0, 242, 239, 0, 0, 0, 0,
-        0, 48, 68, 68, 246, 239, 68, 68, 52, 0,
-        0, 224, 255, 255, 255, 255, 255, 255, 191, 0,
-        0, 224, 255, 255, 255, 255, 255, 255, 191, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        112, 191, 32, 251, 109, 0, 161, 238, 26, 0,
-        112, 223, 225, 255, 255, 6, 253, 255, 223, 0,
-        112, 255, 234, 201, 255, 126, 191, 250, 255, 5,
-        112, 255, 47, 0, 252, 255, 8, 64, 255, 9,
-        112, 255, 10, 0, 246, 255, 0, 0, 255, 12,
-        112, 255, 6, 0, 243, 207, 0, 0, 253, 13,
-        112, 255, 5, 0, 242, 175, 0, 0, 252, 14,
-        112, 255, 4, 0, 242, 159, 0, 0, 252, 14,
-        112, 255, 3, 0, 242, 159, 0, 0, 252, 14,
-        112, 255, 3, 0, 242, 159, 0, 0, 252, 14,
-        112, 255, 3, 0, 242, 159, 0, 0, 252, 14,
-        112, 255, 3, 0, 242, 159, 0, 0, 252, 14,
-        112, 255, 3, 0, 242, 159, 0, 0, 252, 14,
-        112, 255, 3, 0, 242, 159, 0, 0, 252, 14,
-        112, 255, 3, 0, 242, 159, 0, 0, 252, 14,
-        112, 255, 3, 0, 242, 159, 0, 0, 252, 14,
-        112, 255, 3, 0, 242, 159, 0, 0, 252, 14,
-        112, 255, 3, 0, 242, 159, 0, 0, 252, 14,
-        112, 255, 3, 0, 242, 159, 0, 0, 252, 14,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 245, 47, 16, 199, 254, 190, 6, 0, 0,
-        0, 245, 95, 227, 255, 255, 255, 207, 1, 0,
-        0, 245, 159, 254, 191, 169, 253, 255, 11, 0,
-        0, 245, 255, 191, 1, 0, 112, 255, 63, 0,
-        0, 245, 255, 12, 0, 0, 0, 252, 127, 0,
-        0, 245, 255, 4, 0, 0, 0, 246, 175, 0,
-        0, 245, 255, 0, 0, 0, 0, 244, 191, 0,
-        0, 245, 207, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 191, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 147, 237, 239, 90, 0, 0, 0,
-        0, 0, 144, 255, 255, 255, 255, 28, 0, 0,
-        0, 0, 251, 255, 157, 185, 255, 239, 2, 0,
-        0, 128, 255, 78, 0, 0, 177, 255, 12, 0,
-        0, 241, 255, 3, 0, 0, 0, 253, 111, 0,
-        0, 247, 175, 0, 0, 0, 0, 245, 223, 0,
-        0, 252, 95, 0, 0, 0, 0, 240, 255, 1,
-        0, 255, 47, 0, 0, 0, 0, 192, 255, 5,
-        0, 255, 31, 0, 0, 0, 0, 176, 255, 6,
-        16, 255, 15, 0, 0, 0, 0, 160, 255, 7,
-        0, 255, 31, 0, 0, 0, 0, 176, 255, 6,
-        0, 254, 47, 0, 0, 0, 0, 192, 255, 4,
-        0, 251, 95, 0, 0, 0, 0, 240, 255, 1,
-        0, 247, 191, 0, 0, 0, 0, 245, 223, 0,
-        0, 241, 255, 3, 0, 0, 0, 253, 111, 0,
-        0, 96, 255, 78, 0, 0, 177, 255, 13, 0,
-        0, 0, 250, 255, 157, 185, 255, 239, 2, 0,
-        0, 0, 128, 255, 255, 255, 255, 45, 0, 0,
-        0, 0, 0, 130, 253, 239, 90, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 47, 16, 216, 255, 141, 1, 0, 0,
-        0, 246, 79, 227, 255, 255, 255, 78, 0, 0,
-        0, 246, 159, 254, 173, 185, 255, 255, 3, 0,
-        0, 246, 255, 111, 0, 0, 193, 255, 12, 0,
-        0, 246, 255, 8, 0, 0, 16, 255, 79, 0,
-        0, 246, 255, 1, 0, 0, 0, 249, 175, 0,
-        0, 246, 223, 0, 0, 0, 0, 244, 223, 0,
-        0, 246, 191, 0, 0, 0, 0, 242, 255, 0,
-        0, 246, 175, 0, 0, 0, 0, 240, 255, 1,
-        0, 246, 175, 0, 0, 0, 0, 240, 255, 1,
-        0, 246, 175, 0, 0, 0, 0, 241, 255, 1,
-        0, 246, 207, 0, 0, 0, 0, 242, 255, 0,
-        0, 246, 239, 0, 0, 0, 0, 245, 223, 0,
-        0, 246, 255, 2, 0, 0, 0, 249, 159, 0,
-        0, 246, 255, 9, 0, 0, 16, 255, 79, 0,
-        0, 246, 255, 127, 0, 0, 193, 255, 12, 0,
-        0, 246, 191, 255, 174, 185, 255, 255, 3, 0,
-        0, 246, 143, 228, 255, 255, 255, 78, 0, 0,
-        0, 246, 159, 16, 216, 255, 141, 1, 0, 0,
-        0, 246, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 175, 0, 0, 0, 0, 0, 0, 0,
-        0, 246, 175, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 181, 254, 174, 5, 192, 207, 0,
-        0, 0, 176, 255, 255, 255, 191, 224, 207, 0,
-        0, 0, 251, 255, 156, 201, 255, 252, 207, 0,
-        0, 96, 255, 95, 0, 0, 227, 255, 207, 0,
-        0, 208, 255, 7, 0, 0, 48, 255, 207, 0,
-        0, 243, 255, 0, 0, 0, 0, 252, 207, 0,
-        0, 246, 191, 0, 0, 0, 0, 248, 207, 0,
-        0, 249, 143, 0, 0, 0, 0, 245, 207, 0,
-        0, 250, 127, 0, 0, 0, 0, 244, 207, 0,
-        0, 251, 111, 0, 0, 0, 0, 243, 207, 0,
-        0, 250, 127, 0, 0, 0, 0, 244, 207, 0,
-        0, 249, 143, 0, 0, 0, 0, 245, 207, 0,
-        0, 247, 191, 0, 0, 0, 0, 247, 207, 0,
-        0, 243, 255, 0, 0, 0, 0, 251, 207, 0,
-        0, 224, 255, 7, 0, 0, 32, 255, 207, 0,
-        0, 112, 255, 95, 0, 0, 210, 255, 207, 0,
-        0, 0, 252, 255, 156, 201, 255, 249, 207, 0,
-        0, 0, 193, 255, 255, 255, 143, 242, 207, 0,
-        0, 0, 0, 181, 254, 174, 3, 243, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 243, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 243, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 243, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 243, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 243, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 243, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 243, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 243, 207, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 14, 0, 181, 254, 206, 40, 0,
-        0, 0, 249, 15, 176, 255, 255, 255, 95, 0,
-        0, 0, 249, 31, 251, 255, 237, 255, 31, 0,
-        0, 0, 249, 175, 223, 5, 0, 48, 6, 0,
-        0, 0, 249, 255, 12, 0, 0, 0, 0, 0,
-        0, 0, 249, 255, 2, 0, 0, 0, 0, 0,
-        0, 0, 249, 191, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 143, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 111, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 111, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 111, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 111, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 111, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 111, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 111, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 111, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 111, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 111, 0, 0, 0, 0, 0, 0,
-        0, 0, 249, 111, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 147, 253, 255, 190, 55, 0, 0,
-        0, 0, 160, 255, 255, 255, 255, 255, 8, 0,
-        0, 0, 250, 255, 172, 169, 252, 255, 3, 0,
-        0, 32, 255, 62, 0, 0, 0, 149, 0, 0,
-        0, 80, 255, 9, 0, 0, 0, 0, 0, 0,
-        0, 64, 255, 11, 0, 0, 0, 0, 0, 0,
-        0, 16, 255, 143, 0, 0, 0, 0, 0, 0,
-        0, 0, 247, 255, 142, 2, 0, 0, 0, 0,
-        0, 0, 96, 255, 255, 207, 6, 0, 0, 0,
-        0, 0, 0, 129, 254, 255, 239, 6, 0, 0,
-        0, 0, 0, 0, 80, 251, 255, 175, 0, 0,
-        0, 0, 0, 0, 0, 32, 250, 255, 7, 0,
-        0, 0, 0, 0, 0, 0, 112, 255, 13, 0,
-        0, 0, 0, 0, 0, 0, 0, 255, 15, 0,
-        0, 0, 0, 0, 0, 0, 16, 255, 15, 0,
-        0, 128, 40, 0, 0, 0, 160, 255, 10, 0,
-        0, 160, 255, 190, 137, 185, 255, 255, 2, 0,
-        0, 144, 255, 255, 255, 255, 255, 62, 0, 0,
-        0, 0, 148, 236, 255, 223, 106, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 16, 207, 0, 0, 0, 0, 0,
-        0, 0, 0, 96, 207, 0, 0, 0, 0, 0,
-        0, 0, 0, 160, 207, 0, 0, 0, 0, 0,
-        0, 0, 0, 224, 207, 0, 0, 0, 0, 0,
-        0, 0, 0, 243, 207, 0, 0, 0, 0, 0,
-        0, 115, 218, 255, 255, 255, 255, 255, 15, 0,
-        0, 251, 255, 255, 255, 255, 255, 255, 15, 0,
-        0, 83, 85, 248, 223, 85, 85, 85, 5, 0,
-        0, 0, 0, 244, 207, 0, 0, 0, 0, 0,
-        0, 0, 0, 244, 207, 0, 0, 0, 0, 0,
-        0, 0, 0, 244, 207, 0, 0, 0, 0, 0,
-        0, 0, 0, 244, 207, 0, 0, 0, 0, 0,
-        0, 0, 0, 244, 207, 0, 0, 0, 0, 0,
-        0, 0, 0, 244, 207, 0, 0, 0, 0, 0,
-        0, 0, 0, 244, 207, 0, 0, 0, 0, 0,
-        0, 0, 0, 244, 207, 0, 0, 0, 0, 0,
-        0, 0, 0, 244, 207, 0, 0, 0, 0, 0,
-        0, 0, 0, 244, 207, 0, 0, 0, 0, 0,
-        0, 0, 0, 243, 223, 0, 0, 0, 0, 0,
-        0, 0, 0, 242, 255, 0, 0, 0, 0, 0,
-        0, 0, 0, 224, 255, 8, 0, 0, 0, 0,
-        0, 0, 0, 128, 255, 223, 137, 185, 12, 0,
-        0, 0, 0, 0, 251, 255, 255, 255, 15, 0,
-        0, 0, 0, 0, 80, 235, 255, 190, 8, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 244, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 245, 207, 0,
-        0, 245, 175, 0, 0, 0, 0, 246, 207, 0,
-        0, 245, 191, 0, 0, 0, 0, 248, 207, 0,
-        0, 244, 223, 0, 0, 0, 0, 253, 207, 0,
-        0, 241, 255, 2, 0, 0, 64, 255, 207, 0,
-        0, 192, 255, 28, 0, 0, 244, 255, 207, 0,
-        0, 64, 255, 255, 154, 218, 255, 248, 207, 0,
-        0, 0, 247, 255, 255, 255, 143, 224, 207, 0,
-        0, 0, 48, 217, 255, 157, 3, 192, 207, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        112, 255, 10, 0, 0, 0, 0, 64, 255, 13,
-        16, 255, 15, 0, 0, 0, 0, 160, 255, 7,
-        0, 251, 95, 0, 0, 0, 0, 240, 255, 1,
-        0, 245, 191, 0, 0, 0, 0, 245, 191, 0,
-        0, 240, 255, 1, 0, 0, 0, 251, 111, 0,
-        0, 160, 255, 6, 0, 0, 16, 255, 15, 0,
-        0, 64, 255, 12, 0, 0, 96, 255, 10, 0,
-        0, 0, 254, 47, 0, 0, 192, 255, 4, 0,
-        0, 0, 248, 127, 0, 0, 241, 239, 0, 0,
-        0, 0, 243, 223, 0, 0, 247, 143, 0, 0,
-        0, 0, 208, 255, 3, 0, 253, 63, 0, 0,
-        0, 0, 112, 255, 8, 32, 255, 13, 0, 0,
-        0, 0, 16, 255, 14, 112, 255, 7, 0, 0,
-        0, 0, 0, 251, 63, 192, 255, 1, 0, 0,
-        0, 0, 0, 245, 143, 241, 191, 0, 0, 0,
-        0, 0, 0, 240, 207, 246, 111, 0, 0, 0,
-        0, 0, 0, 160, 255, 251, 15, 0, 0, 0,
-        0, 0, 0, 64, 255, 255, 10, 0, 0, 0,
-        0, 0, 0, 0, 254, 255, 4, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        255, 11, 0, 0, 247, 255, 0, 0, 64, 255,
-        252, 14, 0, 0, 251, 255, 2, 0, 96, 255,
-        249, 15, 0, 0, 254, 255, 5, 0, 144, 255,
-        246, 63, 0, 32, 255, 251, 9, 0, 192, 207,
-        243, 111, 0, 80, 207, 248, 12, 0, 240, 159,
-        240, 159, 0, 144, 159, 244, 15, 0, 242, 111,
-        192, 191, 0, 192, 111, 241, 63, 0, 245, 63,
-        144, 239, 0, 240, 63, 224, 111, 0, 248, 15,
-        96, 255, 1, 243, 31, 176, 175, 0, 251, 12,
-        48, 255, 4, 247, 13, 128, 223, 0, 253, 9,
-        0, 255, 6, 250, 10, 64, 255, 0, 255, 6,
-        0, 253, 9, 253, 6, 16, 255, 51, 255, 3,
-        0, 250, 11, 255, 3, 0, 253, 86, 255, 0,
-        0, 246, 61, 255, 0, 0, 250, 121, 223, 0,
-        0, 243, 111, 207, 0, 0, 247, 172, 159, 0,
-        0, 240, 191, 143, 0, 0, 243, 207, 111, 0,
-        0, 208, 255, 95, 0, 0, 240, 255, 63, 0,
-        0, 160, 255, 31, 0, 0, 208, 255, 15, 0,
-        0, 112, 255, 14, 0, 0, 144, 255, 13, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 251, 191, 0, 0, 0, 0, 244, 255, 2,
-        0, 225, 255, 6, 0, 0, 16, 254, 111, 0,
-        0, 64, 255, 31, 0, 0, 160, 255, 11, 0,
-        0, 0, 249, 191, 0, 0, 245, 239, 1, 0,
-        0, 0, 208, 255, 6, 16, 254, 79, 0, 0,
-        0, 0, 48, 255, 47, 176, 255, 9, 0, 0,
-        0, 0, 0, 247, 207, 246, 223, 0, 0, 0,
-        0, 0, 0, 192, 255, 255, 47, 0, 0, 0,
-        0, 0, 0, 32, 255, 255, 7, 0, 0, 0,
-        0, 0, 0, 0, 251, 255, 2, 0, 0, 0,
-        0, 0, 0, 96, 255, 255, 12, 0, 0, 0,
-        0, 0, 0, 242, 255, 253, 127, 0, 0, 0,
-        0, 0, 0, 252, 127, 242, 255, 3, 0, 0,
-        0, 0, 128, 255, 12, 96, 255, 13, 0, 0,
-        0, 0, 243, 255, 2, 0, 251, 159, 0, 0,
-        0, 0, 254, 111, 0, 0, 225, 255, 4, 0,
-        0, 160, 255, 11, 0, 0, 80, 255, 30, 0,
-        0, 245, 255, 1, 0, 0, 0, 250, 191, 0,
-        16, 254, 95, 0, 0, 0, 0, 225, 255, 6,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        112, 255, 10, 0, 0, 0, 0, 64, 255, 13,
-        16, 255, 31, 0, 0, 0, 0, 160, 255, 7,
-        0, 250, 127, 0, 0, 0, 0, 240, 255, 1,
-        0, 243, 223, 0, 0, 0, 0, 246, 191, 0,
-        0, 208, 255, 3, 0, 0, 0, 251, 95, 0,
-        0, 96, 255, 10, 0, 0, 16, 255, 14, 0,
-        0, 0, 255, 31, 0, 0, 112, 255, 9, 0,
-        0, 0, 249, 111, 0, 0, 208, 255, 3, 0,
-        0, 0, 243, 207, 0, 0, 242, 223, 0, 0,
-        0, 0, 192, 255, 3, 0, 248, 127, 0, 0,
-        0, 0, 96, 255, 9, 0, 254, 31, 0, 0,
-        0, 0, 0, 255, 15, 48, 255, 11, 0, 0,
-        0, 0, 0, 249, 95, 144, 255, 5, 0, 0,
-        0, 0, 0, 243, 191, 224, 239, 0, 0, 0,
-        0, 0, 0, 192, 255, 243, 159, 0, 0, 0,
-        0, 0, 0, 96, 255, 251, 63, 0, 0, 0,
-        0, 0, 0, 0, 255, 255, 13, 0, 0, 0,
-        0, 0, 0, 0, 249, 255, 7, 0, 0, 0,
-        0, 0, 0, 0, 242, 255, 1, 0, 0, 0,
-        0, 0, 0, 0, 244, 191, 0, 0, 0, 0,
-        0, 0, 0, 0, 251, 95, 0, 0, 0, 0,
-        0, 0, 0, 16, 255, 14, 0, 0, 0, 0,
-        0, 0, 0, 144, 255, 8, 0, 0, 0, 0,
-        0, 0, 0, 246, 239, 1, 0, 0, 0, 0,
-        64, 154, 217, 255, 95, 0, 0, 0, 0, 0,
-        80, 255, 255, 255, 8, 0, 0, 0, 0, 0,
-        64, 253, 239, 75, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 112, 255, 255, 255, 255, 255, 255, 15, 0,
-        0, 112, 255, 255, 255, 255, 255, 255, 15, 0,
-        0, 32, 85, 85, 85, 85, 133, 255, 14, 0,
-        0, 0, 0, 0, 0, 0, 208, 255, 4, 0,
-        0, 0, 0, 0, 0, 0, 250, 143, 0, 0,
-        0, 0, 0, 0, 0, 96, 255, 11, 0, 0,
-        0, 0, 0, 0, 0, 243, 239, 1, 0, 0,
-        0, 0, 0, 0, 0, 253, 63, 0, 0, 0,
-        0, 0, 0, 0, 160, 255, 7, 0, 0, 0,
-        0, 0, 0, 0, 246, 191, 0, 0, 0, 0,
-        0, 0, 0, 48, 255, 30, 0, 0, 0, 0,
-        0, 0, 0, 209, 255, 3, 0, 0, 0, 0,
-        0, 0, 0, 250, 127, 0, 0, 0, 0, 0,
-        0, 0, 112, 255, 11, 0, 0, 0, 0, 0,
-        0, 0, 243, 223, 1, 0, 0, 0, 0, 0,
-        0, 16, 254, 63, 0, 0, 0, 0, 0, 0,
-        0, 176, 255, 91, 85, 85, 85, 85, 21, 0,
-        0, 240, 255, 255, 255, 255, 255, 255, 95, 0,
-        0, 240, 255, 255, 255, 255, 255, 255, 95, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 166, 253, 9, 0,
-        0, 0, 0, 0, 0, 227, 255, 255, 10, 0,
-        0, 0, 0, 0, 16, 254, 255, 139, 5, 0,
-        0, 0, 0, 0, 96, 255, 30, 0, 0, 0,
-        0, 0, 0, 0, 144, 255, 8, 0, 0, 0,
-        0, 0, 0, 0, 160, 255, 6, 0, 0, 0,
-        0, 0, 0, 0, 160, 255, 6, 0, 0, 0,
-        0, 0, 0, 0, 160, 255, 6, 0, 0, 0,
-        0, 0, 0, 0, 160, 255, 6, 0, 0, 0,
-        0, 0, 0, 0, 160, 255, 6, 0, 0, 0,
-        0, 0, 0, 0, 160, 255, 5, 0, 0, 0,
-        0, 0, 0, 0, 192, 255, 4, 0, 0, 0,
-        0, 0, 0, 0, 244, 255, 1, 0, 0, 0,
-        0, 0, 16, 148, 255, 159, 0, 0, 0, 0,
-        0, 64, 255, 255, 239, 7, 0, 0, 0, 0,
-        0, 64, 255, 255, 77, 0, 0, 0, 0, 0,
-        0, 32, 186, 254, 255, 45, 0, 0, 0, 0,
-        0, 0, 0, 48, 252, 223, 0, 0, 0, 0,
-        0, 0, 0, 0, 225, 255, 3, 0, 0, 0,
-        0, 0, 0, 0, 176, 255, 5, 0, 0, 0,
-        0, 0, 0, 0, 160, 255, 6, 0, 0, 0,
-        0, 0, 0, 0, 160, 255, 6, 0, 0, 0,
-        0, 0, 0, 0, 160, 255, 6, 0, 0, 0,
-        0, 0, 0, 0, 160, 255, 6, 0, 0, 0,
-        0, 0, 0, 0, 160, 255, 6, 0, 0, 0,
-        0, 0, 0, 0, 160, 255, 7, 0, 0, 0,
-        0, 0, 0, 0, 112, 255, 30, 0, 0, 0,
-        0, 0, 0, 0, 16, 255, 255, 155, 5, 0,
-        0, 0, 0, 0, 0, 228, 255, 255, 10, 0,
-        0, 0, 0, 0, 0, 0, 182, 253, 9, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 240, 111, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 48, 239, 140, 2, 0, 0, 0, 0, 0,
-        0, 48, 255, 255, 143, 0, 0, 0, 0, 0,
-        0, 16, 168, 254, 255, 6, 0, 0, 0, 0,
-        0, 0, 0, 160, 255, 13, 0, 0, 0, 0,
-        0, 0, 0, 32, 255, 15, 0, 0, 0, 0,
-        0, 0, 0, 0, 255, 31, 0, 0, 0, 0,
-        0, 0, 0, 0, 255, 31, 0, 0, 0, 0,
-        0, 0, 0, 0, 255, 31, 0, 0, 0, 0,
-        0, 0, 0, 0, 255, 31, 0, 0, 0, 0,
-        0, 0, 0, 0, 255, 31, 0, 0, 0, 0,
-        0, 0, 0, 0, 255, 31, 0, 0, 0, 0,
-        0, 0, 0, 0, 254, 47, 0, 0, 0, 0,
-        0, 0, 0, 0, 251, 175, 0, 0, 0, 0,
-        0, 0, 0, 0, 243, 255, 109, 2, 0, 0,
-        0, 0, 0, 0, 48, 252, 255, 255, 10, 0,
-        0, 0, 0, 0, 0, 146, 255, 255, 10, 0,
-        0, 0, 0, 0, 128, 255, 255, 188, 7, 0,
-        0, 0, 0, 0, 246, 255, 7, 0, 0, 0,
-        0, 0, 0, 0, 252, 111, 0, 0, 0, 0,
-        0, 0, 0, 0, 255, 31, 0, 0, 0, 0,
-        0, 0, 0, 0, 255, 31, 0, 0, 0, 0,
-        0, 0, 0, 0, 255, 31, 0, 0, 0, 0,
-        0, 0, 0, 0, 255, 31, 0, 0, 0, 0,
-        0, 0, 0, 0, 255, 31, 0, 0, 0, 0,
-        0, 0, 0, 0, 255, 31, 0, 0, 0, 0,
-        0, 0, 0, 0, 255, 15, 0, 0, 0, 0,
-        0, 0, 0, 64, 255, 15, 0, 0, 0, 0,
-        0, 0, 32, 230, 255, 10, 0, 0, 0, 0,
-        0, 48, 255, 255, 239, 2, 0, 0, 0, 0,
-        0, 48, 255, 255, 26, 0, 0, 0, 0, 0,
-        0, 16, 103, 4, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-    {   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 97, 120, 3, 0, 0, 0, 32, 0,
-        0, 128, 255, 255, 239, 56, 0, 0, 214, 0,
-        0, 246, 255, 255, 255, 255, 173, 217, 223, 0,
-        0, 247, 40, 48, 232, 255, 255, 255, 127, 0,
-        0, 54, 0, 0, 0, 148, 253, 190, 4, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-    },
-};
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/modelbin.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/modelbin.h
deleted file mode 100644
index aada5f61..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/modelbin.h
+++ /dev/null
@@ -1,80 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_MODELBIN_H
-#define NCNN_MODELBIN_H
-
-#include "mat.h"
-
-namespace ncnn {
-
-class DataReader;
-class NCNN_EXPORT ModelBin
-{
-public:
-    ModelBin();
-    virtual ~ModelBin();
-    // element type
-    // 0 = auto
-    // 1 = float32
-    // 2 = float16
-    // 3 = int8
-    // load vec
-    virtual Mat load(int w, int type) const;
-    // load image
-    virtual Mat load(int w, int h, int type) const;
-    // load dim
-    virtual Mat load(int w, int h, int c, int type) const;
-    // load cube
-    virtual Mat load(int w, int h, int d, int c, int type) const;
-};
-
-class ModelBinFromDataReaderPrivate;
-class NCNN_EXPORT ModelBinFromDataReader : public ModelBin
-{
-public:
-    explicit ModelBinFromDataReader(const DataReader& dr);
-    virtual ~ModelBinFromDataReader();
-
-    virtual Mat load(int w, int type) const;
-
-private:
-    ModelBinFromDataReader(const ModelBinFromDataReader&);
-    ModelBinFromDataReader& operator=(const ModelBinFromDataReader&);
-
-private:
-    ModelBinFromDataReaderPrivate* const d;
-};
-
-class ModelBinFromMatArrayPrivate;
-class NCNN_EXPORT ModelBinFromMatArray : public ModelBin
-{
-public:
-    // construct from weight blob array
-    explicit ModelBinFromMatArray(const Mat* weights);
-    virtual ~ModelBinFromMatArray();
-
-    virtual Mat load(int w, int type) const;
-
-private:
-    ModelBinFromMatArray(const ModelBinFromMatArray&);
-    ModelBinFromMatArray& operator=(const ModelBinFromMatArray&);
-
-private:
-    ModelBinFromMatArrayPrivate* const d;
-};
-
-} // namespace ncnn
-
-#endif // NCNN_MODELBIN_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/ncnn_export.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/ncnn_export.h
deleted file mode 100644
index e2f5fdee..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/ncnn_export.h
+++ /dev/null
@@ -1,42 +0,0 @@
-
-#ifndef NCNN_EXPORT_H
-#define NCNN_EXPORT_H
-
-#ifdef NCNN_STATIC_DEFINE
-#  define NCNN_EXPORT
-#  define NCNN_NO_EXPORT
-#else
-#  ifndef NCNN_EXPORT
-#    ifdef ncnn_EXPORTS
-        /* We are building this library */
-#      define NCNN_EXPORT __attribute__((visibility("default")))
-#    else
-        /* We are using this library */
-#      define NCNN_EXPORT __attribute__((visibility("default")))
-#    endif
-#  endif
-
-#  ifndef NCNN_NO_EXPORT
-#    define NCNN_NO_EXPORT __attribute__((visibility("hidden")))
-#  endif
-#endif
-
-#ifndef NCNN_DEPRECATED
-#  define NCNN_DEPRECATED __attribute__ ((__deprecated__))
-#endif
-
-#ifndef NCNN_DEPRECATED_EXPORT
-#  define NCNN_DEPRECATED_EXPORT NCNN_EXPORT NCNN_DEPRECATED
-#endif
-
-#ifndef NCNN_DEPRECATED_NO_EXPORT
-#  define NCNN_DEPRECATED_NO_EXPORT NCNN_NO_EXPORT NCNN_DEPRECATED
-#endif
-
-#if 0 /* DEFINE_NO_DEPRECATED */
-#  ifndef NCNN_NO_DEPRECATED
-#    define NCNN_NO_DEPRECATED
-#  endif
-#endif
-
-#endif /* NCNN_EXPORT_H */
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/net.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/net.h
deleted file mode 100644
index d69443bb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/net.h
+++ /dev/null
@@ -1,275 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_NET_H
-#define NCNN_NET_H
-
-#include "blob.h"
-#include "layer.h"
-#include "mat.h"
-#include "option.h"
-#include "platform.h"
-
-#if NCNN_PLATFORM_API
-#if __ANDROID_API__ >= 9
-#include <android/asset_manager.h>
-#endif // __ANDROID_API__ >= 9
-#endif // NCNN_PLATFORM_API
-
-namespace ncnn {
-
-#if NCNN_VULKAN
-class VkCompute;
-#endif // NCNN_VULKAN
-class DataReader;
-class Extractor;
-class NetPrivate;
-class NCNN_EXPORT Net
-{
-public:
-    // empty init
-    Net();
-    // clear and destroy
-    virtual ~Net();
-
-public:
-    // option can be changed before loading
-    Option opt;
-
-#if NCNN_VULKAN
-    // set gpu device by index
-    void set_vulkan_device(int device_index);
-
-    // set gpu device by device handle, no owner transfer
-    void set_vulkan_device(const VulkanDevice* vkdev);
-
-    const VulkanDevice* vulkan_device() const;
-#endif // NCNN_VULKAN
-
-#if NCNN_STRING
-    // register custom layer or overwrite built-in layer by layer type name
-    // return 0 if success
-    int register_custom_layer(const char* type, layer_creator_func creator, layer_destroyer_func destroyer = 0, void* userdata = 0);
-    virtual int custom_layer_to_index(const char* type);
-#endif // NCNN_STRING
-    // register custom layer or overwrite built-in layer by layer type
-    // return 0 if success
-    int register_custom_layer(int index, layer_creator_func creator, layer_destroyer_func destroyer = 0, void* userdata = 0);
-
-#if NCNN_STRING
-    int load_param(const DataReader& dr);
-#endif // NCNN_STRING
-
-    int load_param_bin(const DataReader& dr);
-
-    int load_model(const DataReader& dr);
-
-#if NCNN_STDIO
-#if NCNN_STRING
-    // load network structure from plain param file
-    // return 0 if success
-    int load_param(FILE* fp);
-    int load_param(const char* protopath);
-    int load_param_mem(const char* mem);
-#endif // NCNN_STRING
-    // load network structure from binary param file
-    // return 0 if success
-    int load_param_bin(FILE* fp);
-    int load_param_bin(const char* protopath);
-
-    // load network weight data from model file
-    // return 0 if success
-    int load_model(FILE* fp);
-    int load_model(const char* modelpath);
-#endif // NCNN_STDIO
-
-    // load network structure from external memory
-    // memory pointer must be 32-bit aligned
-    // return bytes consumed
-    int load_param(const unsigned char* mem);
-
-    // reference network weight data from external memory
-    // weight data is not copied but referenced
-    // so external memory should be retained when used
-    // memory pointer must be 32-bit aligned
-    // return bytes consumed
-    int load_model(const unsigned char* mem);
-
-#if NCNN_PLATFORM_API
-#if __ANDROID_API__ >= 9
-#if NCNN_STRING
-    // convenient load network structure from android asset plain param file
-    int load_param(AAsset* asset);
-    int load_param(AAssetManager* mgr, const char* assetpath);
-#endif // NCNN_STRING
-    // convenient load network structure from android asset binary param file
-    int load_param_bin(AAsset* asset);
-    int load_param_bin(AAssetManager* mgr, const char* assetpath);
-
-    // convenient load network weight data from android asset model file
-    int load_model(AAsset* asset);
-    int load_model(AAssetManager* mgr, const char* assetpath);
-#endif // __ANDROID_API__ >= 9
-#endif // NCNN_PLATFORM_API
-
-    // unload network structure and weight data
-    void clear();
-
-    // construct an Extractor from network
-    Extractor create_extractor() const;
-
-    // get input/output indexes/names
-    const std::vector<int>& input_indexes() const;
-    const std::vector<int>& output_indexes() const;
-#if NCNN_STRING
-    const std::vector<const char*>& input_names() const;
-    const std::vector<const char*>& output_names() const;
-#endif
-
-    const std::vector<Blob>& blobs() const;
-    const std::vector<Layer*>& layers() const;
-
-    std::vector<Blob>& mutable_blobs();
-    std::vector<Layer*>& mutable_layers();
-
-protected:
-    friend class Extractor;
-#if NCNN_STRING
-    int find_blob_index_by_name(const char* name) const;
-    int find_layer_index_by_name(const char* name) const;
-    virtual Layer* create_custom_layer(const char* type);
-    virtual Layer* create_overwrite_builtin_layer(const char* type);
-#endif // NCNN_STRING
-    virtual Layer* create_custom_layer(int index);
-    virtual Layer* create_overwrite_builtin_layer(int typeindex);
-
-private:
-    Net(const Net&);
-    Net& operator=(const Net&);
-
-private:
-    NetPrivate* const d;
-};
-
-class ExtractorPrivate;
-class NCNN_EXPORT Extractor
-{
-public:
-    virtual ~Extractor();
-
-    // copy
-    Extractor(const Extractor&);
-
-    // assign
-    Extractor& operator=(const Extractor&);
-
-    // clear blob mats and alloctors
-    void clear();
-
-    // enable light mode
-    // intermediate blob will be recycled when enabled
-    // enabled by default
-    void set_light_mode(bool enable);
-
-    // deprecated, no-op
-    // instead, set net.opt.num_threads before net.load_param()
-    void set_num_threads(int num_threads);
-
-    // set blob memory allocator
-    void set_blob_allocator(Allocator* allocator);
-
-    // set workspace memory allocator
-    void set_workspace_allocator(Allocator* allocator);
-
-#if NCNN_VULKAN
-    // deprecated, no-op
-    // instead, set net.opt.use_vulkan_compute before net.load_param()
-    void set_vulkan_compute(bool enable);
-
-    void set_blob_vkallocator(VkAllocator* allocator);
-
-    void set_workspace_vkallocator(VkAllocator* allocator);
-
-    void set_staging_vkallocator(VkAllocator* allocator);
-#endif // NCNN_VULKAN
-
-#if NCNN_STRING
-    // set input by blob name
-    // return 0 if success
-    int input(const char* blob_name, const Mat& in);
-
-    // get result by blob name
-    // return 0 if success
-    // type = 0, default
-    // type = 1, do not convert fp16/bf16 or / and packing
-    int extract(const char* blob_name, Mat& feat, int type = 0);
-#endif // NCNN_STRING
-
-    // set input by blob index
-    // return 0 if success
-    int input(int blob_index, const Mat& in);
-
-    // get result by blob index
-    // return 0 if success
-    // type = 0, default
-    // type = 1, do not convert fp16/bf16 or / and packing
-    int extract(int blob_index, Mat& feat, int type = 0);
-
-#if NCNN_VULKAN
-#if NCNN_STRING
-    // set input by blob name
-    // return 0 if success
-    int input(const char* blob_name, const VkMat& in);
-
-    // get result by blob name
-    // return 0 if success
-    int extract(const char* blob_name, VkMat& feat, VkCompute& cmd);
-
-    // set input by blob name
-    // return 0 if success
-    int input(const char* blob_name, const VkImageMat& in);
-
-    // get result by blob name
-    // return 0 if success
-    int extract(const char* blob_name, VkImageMat& feat, VkCompute& cmd);
-#endif // NCNN_STRING
-
-    // set input by blob index
-    // return 0 if success
-    int input(int blob_index, const VkMat& in);
-
-    // get result by blob index
-    // return 0 if success
-    int extract(int blob_index, VkMat& feat, VkCompute& cmd);
-
-    // set input by blob index
-    // return 0 if success
-    int input(int blob_index, const VkImageMat& in);
-
-    // get result by blob index
-    // return 0 if success
-    int extract(int blob_index, VkImageMat& feat, VkCompute& cmd);
-#endif // NCNN_VULKAN
-
-protected:
-    friend Extractor Net::create_extractor() const;
-    Extractor(const Net* net, size_t blob_count);
-
-private:
-    ExtractorPrivate* const d;
-};
-
-} // namespace ncnn
-
-#endif // NCNN_NET_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/option.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/option.h
deleted file mode 100644
index eb2a5a7d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/option.h
+++ /dev/null
@@ -1,158 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_OPTION_H
-#define NCNN_OPTION_H
-
-#include "platform.h"
-
-namespace ncnn {
-
-#if NCNN_VULKAN
-class VkAllocator;
-class PipelineCache;
-#endif // NCNN_VULKAN
-
-class Allocator;
-class NCNN_EXPORT Option
-{
-public:
-    // default option
-    Option();
-
-public:
-    // light mode
-    // intermediate blob will be recycled when enabled
-    // enabled by default
-    bool lightmode;
-
-    // thread count
-    // default value is the one returned by get_cpu_count()
-    int num_threads;
-
-    // blob memory allocator
-    Allocator* blob_allocator;
-
-    // workspace memory allocator
-    Allocator* workspace_allocator;
-
-#if NCNN_VULKAN
-    // blob memory allocator
-    VkAllocator* blob_vkallocator;
-
-    // workspace memory allocator
-    VkAllocator* workspace_vkallocator;
-
-    // staging memory allocator
-    VkAllocator* staging_vkallocator;
-
-    // pipeline cache
-    PipelineCache* pipeline_cache;
-#endif // NCNN_VULKAN
-
-    // the time openmp threads busy-wait for more work before going to sleep
-    // default value is 20ms to keep the cores enabled
-    // without too much extra power consumption afterwards
-    int openmp_blocktime;
-
-    // enable winograd convolution optimization
-    // improve convolution 3x3 stride1 performance, may consume more memory
-    // changes should be applied before loading network structure and weight
-    // enabled by default
-    bool use_winograd_convolution;
-
-    // enable sgemm convolution optimization
-    // improve convolution 1x1 stride1 performance, may consume more memory
-    // changes should be applied before loading network structure and weight
-    // enabled by default
-    bool use_sgemm_convolution;
-
-    // enable quantized int8 inference
-    // use low-precision int8 path for quantized model
-    // changes should be applied before loading network structure and weight
-    // enabled by default
-    bool use_int8_inference;
-
-    // enable vulkan compute
-    bool use_vulkan_compute;
-
-    // enable bf16 data type for storage
-    // improve most operator performance on all arm devices, may consume more memory
-    bool use_bf16_storage;
-
-    // enable options for gpu inference
-    bool use_fp16_packed;
-    bool use_fp16_storage;
-    bool use_fp16_arithmetic;
-    bool use_int8_packed;
-    bool use_int8_storage;
-    bool use_int8_arithmetic;
-
-    // enable simd-friendly packed memory layout
-    // improve all operator performance on all arm devices, will consume more memory
-    // changes should be applied before loading network structure and weight
-    // enabled by default
-    bool use_packing_layout;
-
-    bool use_shader_pack8;
-
-    // subgroup option
-    bool use_subgroup_basic;
-    bool use_subgroup_vote;
-    bool use_subgroup_ballot;
-    bool use_subgroup_shuffle;
-
-    // turn on for adreno
-    bool use_image_storage;
-    bool use_tensor_storage;
-
-    bool use_reserved_0;
-
-    // enable DAZ(Denormals-Are-Zero) and FTZ(Flush-To-Zero)
-    // default value is 3
-    // 0 = DAZ OFF, FTZ OFF
-    // 1 = DAZ ON , FTZ OFF
-    // 2 = DAZ OFF, FTZ ON
-    // 3 = DAZ ON,  FTZ ON
-    int flush_denormals;
-
-    bool use_local_pool_allocator;
-
-    // enable local memory optimization for gpu inference
-    bool use_shader_local_memory;
-
-    // enable cooperative matrix optimization for gpu inference
-    bool use_cooperative_matrix;
-
-    // more fine-grained control of winograd convolution
-    bool use_winograd23_convolution;
-    bool use_winograd43_convolution;
-    bool use_winograd63_convolution;
-
-    // this option is turned on for A53/A55 automatically
-    // but you can force this on/off if you wish
-    bool use_a53_a55_optimized_kernel;
-
-    // enable options for shared variables in gpu shader
-    bool use_fp16_uniform;
-    bool use_int8_uniform;
-
-    bool use_reserved_9;
-    bool use_reserved_10;
-    bool use_reserved_11;
-};
-
-} // namespace ncnn
-
-#endif // NCNN_OPTION_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/paramdict.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/paramdict.h
deleted file mode 100644
index c2ef1606..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/paramdict.h
+++ /dev/null
@@ -1,73 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_PARAMDICT_H
-#define NCNN_PARAMDICT_H
-
-#include "mat.h"
-
-// at most 32 parameters
-#define NCNN_MAX_PARAM_COUNT 32
-
-namespace ncnn {
-
-class DataReader;
-class Net;
-class ParamDictPrivate;
-class NCNN_EXPORT ParamDict
-{
-public:
-    // empty
-    ParamDict();
-
-    virtual ~ParamDict();
-
-    // copy
-    ParamDict(const ParamDict&);
-
-    // assign
-    ParamDict& operator=(const ParamDict&);
-
-    // get type
-    int type(int id) const;
-
-    // get int
-    int get(int id, int def) const;
-    // get float
-    float get(int id, float def) const;
-    // get array
-    Mat get(int id, const Mat& def) const;
-
-    // set int
-    void set(int id, int i);
-    // set float
-    void set(int id, float f);
-    // set array
-    void set(int id, const Mat& v);
-
-protected:
-    friend class Net;
-
-    void clear();
-
-    int load_param(const DataReader& dr);
-    int load_param_bin(const DataReader& dr);
-
-private:
-    ParamDictPrivate* const d;
-};
-
-} // namespace ncnn
-
-#endif // NCNN_PARAMDICT_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/pipeline.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/pipeline.h
deleted file mode 100644
index 8772e665..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/pipeline.h
+++ /dev/null
@@ -1,111 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_PIPELINE_H
-#define NCNN_PIPELINE_H
-
-#include "mat.h"
-#include "platform.h"
-#if NCNN_VULKAN
-#include "gpu.h"
-#endif // NCNN_VULKAN
-
-namespace ncnn {
-
-#if NCNN_VULKAN
-class Option;
-class PipelinePrivate;
-class NCNN_EXPORT Pipeline
-{
-public:
-    explicit Pipeline(const VulkanDevice* vkdev);
-    virtual ~Pipeline();
-
-public:
-    void set_optimal_local_size_xyz(int w = 4, int h = 4, int c = 4);
-    void set_optimal_local_size_xyz(const Mat& local_size_xyz);
-    void set_local_size_xyz(int w, int h, int c);
-
-    int create(const uint32_t* spv_data, size_t spv_data_size, const std::vector<vk_specialization_type>& specializations);
-
-    int create(int shader_type_index, const Option& opt, const std::vector<vk_specialization_type>& specializations);
-
-public:
-    VkShaderModule shader_module() const;
-    VkDescriptorSetLayout descriptorset_layout() const;
-    VkPipelineLayout pipeline_layout() const;
-    VkPipeline pipeline() const;
-    VkDescriptorUpdateTemplateKHR descriptor_update_template() const;
-
-    const ShaderInfo& shader_info() const;
-
-    uint32_t local_size_x() const;
-    uint32_t local_size_y() const;
-    uint32_t local_size_z() const;
-
-protected:
-    void set_shader_module(VkShaderModule shader_module);
-    void set_descriptorset_layout(VkDescriptorSetLayout descriptorset_layout);
-    void set_pipeline_layout(VkPipelineLayout pipeline_layout);
-    void set_pipeline(VkPipeline pipeline);
-    void set_descriptor_update_template(VkDescriptorUpdateTemplateKHR descriptor_update_template);
-
-    void set_shader_info(const ShaderInfo& shader_info);
-
-public:
-    const VulkanDevice* vkdev;
-
-private:
-    Pipeline(const Pipeline&);
-    Pipeline& operator=(const Pipeline&);
-
-private:
-    PipelinePrivate* const d;
-};
-
-#if NCNN_PLATFORM_API
-#if __ANDROID_API__ >= 26
-class VkCompute;
-class NCNN_EXPORT ImportAndroidHardwareBufferPipeline : private Pipeline
-{
-public:
-    explicit ImportAndroidHardwareBufferPipeline(const VulkanDevice* vkdev);
-    virtual ~ImportAndroidHardwareBufferPipeline();
-
-    int create(VkAndroidHardwareBufferImageAllocator* ahb_im_allocator, int type_to, int rotate_from, const Option& opt);
-    int create(VkAndroidHardwareBufferImageAllocator* ahb_im_allocator, int type_to, int rotate_from, int target_width, int target_height, const Option& opt);
-    void destroy();
-
-    friend class VkCompute;
-
-protected:
-    int create_shader_module(const Option& opt);
-    int create_sampler(VkAndroidHardwareBufferImageAllocator* ahb_im_allocator);
-    int create_descriptorset_layout();
-
-public:
-    int type_to;
-    int rotate_from;
-    bool need_resize;
-
-    VkSampler sampler;
-};
-#endif // __ANDROID_API__ >= 26
-#endif // NCNN_PLATFORM_API
-
-#endif // NCNN_VULKAN
-
-} // namespace ncnn
-
-#endif // NCNN_PIPELINE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/pipelinecache.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/pipelinecache.h
deleted file mode 100644
index 1af4c890..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/pipelinecache.h
+++ /dev/null
@@ -1,81 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_PIPELINECACHE_H
-#define NCNN_PIPELINECACHE_H
-
-#include "platform.h"
-
-#include "mat.h"
-#include "gpu.h"
-
-namespace ncnn {
-
-#if NCNN_VULKAN
-
-class VulkanDevice;
-class PipelineCachePrivate;
-class NCNN_EXPORT PipelineCache
-{
-public:
-    explicit PipelineCache(const VulkanDevice* _vkdev);
-
-    virtual ~PipelineCache();
-
-    void clear();
-
-    int get_pipeline(const uint32_t* spv_data, size_t spv_data_size, const std::vector<vk_specialization_type>& specializations,
-                     uint32_t local_size_x, uint32_t local_size_y, uint32_t local_size_z,
-                     VkShaderModule* shader_module,
-                     VkDescriptorSetLayout* descriptorset_layout,
-                     VkPipelineLayout* pipeline_layout,
-                     VkPipeline* pipeline,
-                     VkDescriptorUpdateTemplateKHR* descriptor_update_template,
-                     ShaderInfo& shader_info) const;
-
-    int get_pipeline(int shader_type_index, const Option& opt, const std::vector<vk_specialization_type>& specializations,
-                     uint32_t local_size_x, uint32_t local_size_y, uint32_t local_size_z,
-                     VkShaderModule* shader_module,
-                     VkDescriptorSetLayout* descriptorset_layout,
-                     VkPipelineLayout* pipeline_layout,
-                     VkPipeline* pipeline,
-                     VkDescriptorUpdateTemplateKHR* descriptor_update_template,
-                     ShaderInfo& shader_info) const;
-
-protected:
-    int create_shader_module(int shader_type_index, const Option& opt, uint32_t local_size_x, uint32_t local_size_y, uint32_t local_size_z,
-                             VkShaderModule* _shader_module, ShaderInfo& si) const;
-
-    int new_pipeline(VkShaderModule shader_module, const ShaderInfo& shader_info, const std::vector<vk_specialization_type>& specializations,
-                     VkDescriptorSetLayout* descriptorset_layout,
-                     VkPipelineLayout* pipeline_layout,
-                     VkPipeline* pipeline,
-                     VkDescriptorUpdateTemplateKHR* descriptor_update_template) const;
-
-protected:
-    const VulkanDevice* vkdev;
-
-private:
-    PipelineCache(const PipelineCache&);
-    PipelineCache& operator=(const PipelineCache&);
-
-private:
-    PipelineCachePrivate* const d;
-};
-
-#endif // NCNN_VULKAN
-
-} // namespace ncnn
-
-#endif // NCNN_PIPELINECACHE_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/platform.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/platform.h
deleted file mode 100644
index 3bf58e29..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/platform.h
+++ /dev/null
@@ -1,325 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_PLATFORM_H
-#define NCNN_PLATFORM_H
-
-#define NCNN_STDIO 1
-#define NCNN_STRING 1
-#define NCNN_SIMPLEOCV 0
-#define NCNN_SIMPLEOMP 0
-#define NCNN_SIMPLESTL 0
-#define NCNN_SIMPLEMATH 0
-#define NCNN_THREADS 1
-#define NCNN_BENCHMARK 0
-#define NCNN_C_API 1
-#define NCNN_PLATFORM_API 1
-#define NCNN_PIXEL 0
-#define NCNN_PIXEL_ROTATE 0
-#define NCNN_PIXEL_AFFINE 0
-#define NCNN_PIXEL_DRAWING 0
-#define NCNN_VULKAN 0
-#define NCNN_SIMPLEVK 1
-#define NCNN_SYSTEM_GLSLANG 0
-#define NCNN_RUNTIME_CPU 1
-#define NCNN_GNU_INLINE_ASM 1
-#define NCNN_AVX 0
-#define NCNN_XOP 0
-#define NCNN_FMA 0
-#define NCNN_F16C 0
-#define NCNN_AVX2 0
-#define NCNN_AVXVNNI 0
-#define NCNN_AVX512 0
-#define NCNN_AVX512VNNI 0
-#define NCNN_AVX512BF16 0
-#define NCNN_AVX512FP16 0
-#define NCNN_VFPV4 1
-#define NCNN_ARM82 1
-#define NCNN_ARM82DOT 1
-#define NCNN_ARM82FP16FML 0
-#define NCNN_ARM84BF16 0
-#define NCNN_ARM84I8MM 0
-#define NCNN_ARM86SVE 0
-#define NCNN_ARM86SVE2 0
-#define NCNN_ARM86SVEBF16 0
-#define NCNN_ARM86SVEI8MM 0
-#define NCNN_ARM86SVEF32MM 0
-#define NCNN_MSA 0
-#define NCNN_LSX 0
-#define NCNN_MMI 0
-#define NCNN_RVV 0
-#define NCNN_INT8 1
-#define NCNN_BF16 1
-#define NCNN_FORCE_INLINE 1
-
-#define NCNN_VERSION_STRING "1.0.20240807"
-
-#include "ncnn_export.h"
-
-#ifdef __cplusplus
-
-#if NCNN_THREADS
-#if (defined _WIN32 && !(defined __MINGW32__))
-#define WIN32_LEAN_AND_MEAN
-#include <windows.h>
-#include <process.h>
-#else
-#include <pthread.h>
-#endif
-#endif // NCNN_THREADS
-
-#if __ANDROID_API__ >= 26
-#define VK_USE_PLATFORM_ANDROID_KHR
-#endif // __ANDROID_API__ >= 26
-
-namespace ncnn {
-
-#if NCNN_THREADS
-#if (defined _WIN32 && !(defined __MINGW32__))
-class NCNN_EXPORT Mutex
-{
-public:
-    Mutex() { InitializeSRWLock(&srwlock); }
-    ~Mutex() {}
-    void lock() { AcquireSRWLockExclusive(&srwlock); }
-    void unlock() { ReleaseSRWLockExclusive(&srwlock); }
-private:
-    friend class ConditionVariable;
-    // NOTE SRWLock is available from windows vista
-    SRWLOCK srwlock;
-};
-
-class NCNN_EXPORT ConditionVariable
-{
-public:
-    ConditionVariable() { InitializeConditionVariable(&condvar); }
-    ~ConditionVariable() {}
-    void wait(Mutex& mutex) { SleepConditionVariableSRW(&condvar, &mutex.srwlock, INFINITE, 0); }
-    void broadcast() { WakeAllConditionVariable(&condvar); }
-    void signal() { WakeConditionVariable(&condvar); }
-private:
-    CONDITION_VARIABLE condvar;
-};
-
-static unsigned __stdcall start_wrapper(void* args);
-class NCNN_EXPORT Thread
-{
-public:
-    Thread(void* (*start)(void*), void* args = 0) { _start = start; _args = args; handle = (HANDLE)_beginthreadex(0, 0, start_wrapper, this, 0, 0); }
-    ~Thread() {}
-    void join() { WaitForSingleObject(handle, INFINITE); CloseHandle(handle); }
-private:
-    friend unsigned __stdcall start_wrapper(void* args)
-    {
-        Thread* t = (Thread*)args;
-        t->_start(t->_args);
-        return 0;
-    }
-    HANDLE handle;
-    void* (*_start)(void*);
-    void* _args;
-};
-
-class NCNN_EXPORT ThreadLocalStorage
-{
-public:
-    ThreadLocalStorage() { key = TlsAlloc(); }
-    ~ThreadLocalStorage() { TlsFree(key); }
-    void set(void* value) { TlsSetValue(key, (LPVOID)value); }
-    void* get() { return (void*)TlsGetValue(key); }
-private:
-    DWORD key;
-};
-#else // (defined _WIN32 && !(defined __MINGW32__))
-class NCNN_EXPORT Mutex
-{
-public:
-    Mutex() { pthread_mutex_init(&mutex, 0); }
-    ~Mutex() { pthread_mutex_destroy(&mutex); }
-    void lock() { pthread_mutex_lock(&mutex); }
-    void unlock() { pthread_mutex_unlock(&mutex); }
-private:
-    friend class ConditionVariable;
-    pthread_mutex_t mutex;
-};
-
-class NCNN_EXPORT ConditionVariable
-{
-public:
-    ConditionVariable() { pthread_cond_init(&cond, 0); }
-    ~ConditionVariable() { pthread_cond_destroy(&cond); }
-    void wait(Mutex& mutex) { pthread_cond_wait(&cond, &mutex.mutex); }
-    void broadcast() { pthread_cond_broadcast(&cond); }
-    void signal() { pthread_cond_signal(&cond); }
-private:
-    pthread_cond_t cond;
-};
-
-class NCNN_EXPORT Thread
-{
-public:
-    Thread(void* (*start)(void*), void* args = 0) { pthread_create(&t, 0, start, args); }
-    ~Thread() {}
-    void join() { pthread_join(t, 0); }
-private:
-    pthread_t t;
-};
-
-class NCNN_EXPORT ThreadLocalStorage
-{
-public:
-    ThreadLocalStorage() { pthread_key_create(&key, 0); }
-    ~ThreadLocalStorage() { pthread_key_delete(key); }
-    void set(void* value) { pthread_setspecific(key, value); }
-    void* get() { return pthread_getspecific(key); }
-private:
-    pthread_key_t key;
-};
-#endif // (defined _WIN32 && !(defined __MINGW32__))
-#else // NCNN_THREADS
-class NCNN_EXPORT Mutex
-{
-public:
-    Mutex() {}
-    ~Mutex() {}
-    void lock() {}
-    void unlock() {}
-};
-
-class NCNN_EXPORT ConditionVariable
-{
-public:
-    ConditionVariable() {}
-    ~ConditionVariable() {}
-    void wait(Mutex& /*mutex*/) {}
-    void broadcast() {}
-    void signal() {}
-};
-
-class NCNN_EXPORT Thread
-{
-public:
-    Thread(void* (*/*start*/)(void*), void* /*args*/ = 0) {}
-    ~Thread() {}
-    void join() {}
-};
-
-class NCNN_EXPORT ThreadLocalStorage
-{
-public:
-    ThreadLocalStorage() { data = 0; }
-    ~ThreadLocalStorage() {}
-    void set(void* value) { data = value; }
-    void* get() { return data; }
-private:
-    void* data;
-};
-#endif // NCNN_THREADS
-
-class NCNN_EXPORT MutexLockGuard
-{
-public:
-    MutexLockGuard(Mutex& _mutex) : mutex(_mutex) { mutex.lock(); }
-    ~MutexLockGuard() { mutex.unlock(); }
-private:
-    Mutex& mutex;
-};
-
-static inline void swap_endianness_16(void* x)
-{
-    unsigned char* xx = (unsigned char*)x;
-    unsigned char x0 = xx[0];
-    unsigned char x1 = xx[1];
-    xx[0] = x1;
-    xx[1] = x0;
-}
-
-static inline void swap_endianness_32(void* x)
-{
-    unsigned char* xx = (unsigned char*)x;
-    unsigned char x0 = xx[0];
-    unsigned char x1 = xx[1];
-    unsigned char x2 = xx[2];
-    unsigned char x3 = xx[3];
-    xx[0] = x3;
-    xx[1] = x2;
-    xx[2] = x1;
-    xx[3] = x0;
-}
-
-} // namespace ncnn
-
-#if NCNN_SIMPLESTL
-#include "simplestl.h"
-#else
-#include <algorithm>
-#include <list>
-#include <vector>
-#include <string>
-#endif
-
-// simplemath
-#if NCNN_SIMPLEMATH
-#include "simplemath.h"
-#else
-#include <math.h>
-#include <fenv.h>
-#endif
-
-#if NCNN_VULKAN
-#if NCNN_SIMPLEVK
-#include "simplevk.h"
-#else
-#include <vulkan/vulkan.h>
-#endif
-#include "vulkan_header_fix.h"
-#endif // NCNN_VULKAN
-
-#endif // __cplusplus
-
-#if NCNN_STDIO
-#if NCNN_PLATFORM_API && __ANDROID_API__ >= 8
-#include <android/log.h>
-#define NCNN_LOGE(...) do { \
-    fprintf(stderr, ##__VA_ARGS__); fprintf(stderr, "\n"); \
-    __android_log_print(ANDROID_LOG_WARN, "ncnn", ##__VA_ARGS__); } while(0)
-#else // NCNN_PLATFORM_API && __ANDROID_API__ >= 8
-#include <stdio.h>
-#define NCNN_LOGE(...) do { \
-    fprintf(stderr, ##__VA_ARGS__); fprintf(stderr, "\n"); } while(0)
-#endif // NCNN_PLATFORM_API && __ANDROID_API__ >= 8
-#else
-#define NCNN_LOGE(...)
-#endif
-
-
-#if NCNN_FORCE_INLINE
-#ifdef _MSC_VER
-    #define NCNN_FORCEINLINE __forceinline
-#elif defined(__GNUC__)
-    #define NCNN_FORCEINLINE inline __attribute__((__always_inline__))
-#elif defined(__CLANG__)
-    #if __has_attribute(__always_inline__)
-        #define NCNN_FORCEINLINE inline __attribute__((__always_inline__))
-    #else
-        #define NCNN_FORCEINLINE inline
-    #endif
-#else
-    #define NCNN_FORCEINLINE inline
-#endif
-#else
-    #define NCNN_FORCEINLINE inline
-#endif
-
-#endif // NCNN_PLATFORM_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/ruapu.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/ruapu.h
deleted file mode 100644
index ff3c19e2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/ruapu.h
+++ /dev/null
@@ -1,302 +0,0 @@
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2024 nihui (https://github.com/nihui)
-// Copyright (c) 2024 kernelbin (https://github.com/kernelbin)
-//
-// ruapu --- detect cpu isa features with single-file
-
-#ifndef RUAPU_H
-#define RUAPU_H
-
-void ruapu_init();
-
-int ruapu_supports(const char* isa);
-
-#ifdef RUAPU_IMPLEMENTATION
-
-#include <setjmp.h>
-#include <string.h>
-
-#if defined _WIN32
-
-#include <windows.h>
-
-#if WINAPI_FAMILY == WINAPI_FAMILY_APP
-static int ruapu_detect_isa(const void* some_inst)
-{
-    // uwp does not support seh  :(
-    (void)some_inst;
-    return 0;
-}
-#else // WINAPI_FAMILY == WINAPI_FAMILY_APP
-static int g_ruapu_sigill_caught = 0;
-static jmp_buf g_ruapu_jmpbuf;
-
-typedef const void* ruapu_some_inst;
-
-static LONG CALLBACK ruapu_catch_sigill(struct _EXCEPTION_POINTERS* ExceptionInfo)
-{
-    if (ExceptionInfo->ExceptionRecord->ExceptionCode == EXCEPTION_ILLEGAL_INSTRUCTION)
-    {
-        g_ruapu_sigill_caught = 1;
-        longjmp(g_ruapu_jmpbuf, -1);
-    }
-
-    return EXCEPTION_CONTINUE_SEARCH;
-}
-
-static int ruapu_detect_isa(const void* some_inst)
-{
-    g_ruapu_sigill_caught = 0;
-
-    PVOID eh = AddVectoredExceptionHandler(1, ruapu_catch_sigill);
-
-    if (setjmp(g_ruapu_jmpbuf) == 0)
-    {
-        ((void (*)())some_inst)();
-    }
-
-    RemoveVectoredExceptionHandler(eh);
-
-    return g_ruapu_sigill_caught ? 0 : 1;
-}
-#endif // WINAPI_FAMILY == WINAPI_FAMILY_APP
-
-#if defined(__i386__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64)
-#ifdef _MSC_VER
-#define RUAPU_INSTCODE(isa, ...) __pragma(section(".text")) __declspec(allocate(".text")) static unsigned char ruapu_some_##isa[] = { __VA_ARGS__, 0xc3 };
-#else
-#define RUAPU_INSTCODE(isa, ...) __attribute__((section(".text"))) static unsigned char ruapu_some_##isa[] = { __VA_ARGS__, 0xc3 };
-#endif
-
-#elif __aarch64__ || defined(_M_ARM64)
-#ifdef _MSC_VER
-#define RUAPU_INSTCODE(isa, ...) __pragma(section(".text")) __declspec(allocate(".text")) static unsigned int ruapu_some_##isa[] = { __VA_ARGS__, 0xd65f03c0 };
-#else
-#define RUAPU_INSTCODE(isa, ...) __attribute__((section(".text"))) static unsigned int ruapu_some_##isa[] = { __VA_ARGS__, 0xd65f03c0 };
-#endif
-
-#elif __arm__ || defined(_M_ARM)
-#if __thumb__
-#ifdef _MSC_VER
-#define RUAPU_INSTCODE(isa, ...) __pragma(section(".text")) __declspec(allocate(".text")) static unsigned int ruapu_some_##isa[] = { __VA_ARGS__, 0x4770 };
-#else
-#define RUAPU_INSTCODE(isa, ...) __attribute__((section(".text"))) static unsigned int ruapu_some_##isa[] = { __VA_ARGS__, 0x4770 };
-#endif
-#else
-#ifdef _MSC_VER
-#define RUAPU_INSTCODE(isa, ...) __pragma(section(".text")) __declspec(allocate(".text")) static unsigned int ruapu_some_##isa[] = { __VA_ARGS__, 0xe12fff1e };
-#else
-#define RUAPU_INSTCODE(isa, ...) __attribute__((section(".text"))) static unsigned int ruapu_some_##isa[] = { __VA_ARGS__, 0xe12fff1e };
-#endif
-#endif
-
-#endif
-
-#elif defined __ANDROID__ || defined __linux__ || defined __APPLE__
-#include <signal.h>
-
-static int g_ruapu_sigill_caught = 0;
-static sigjmp_buf g_ruapu_jmpbuf;
-
-typedef void (*ruapu_some_inst)();
-
-static void ruapu_catch_sigill(int signo, siginfo_t* si, void* data)
-{
-    (void)signo;
-    (void)si;
-    (void)data;
-
-    g_ruapu_sigill_caught = 1;
-    siglongjmp(g_ruapu_jmpbuf, -1);
-}
-
-static int ruapu_detect_isa(ruapu_some_inst some_inst)
-{
-    g_ruapu_sigill_caught = 0;
-
-    struct sigaction sa = { 0 };
-    struct sigaction old_sa;
-    sa.sa_flags = SA_ONSTACK | SA_RESTART | SA_SIGINFO;
-    sa.sa_sigaction = ruapu_catch_sigill;
-    sigaction(SIGILL, &sa, &old_sa);
-
-    if (sigsetjmp(g_ruapu_jmpbuf, 1) == 0)
-    {
-        some_inst();
-    }
-
-    sigaction(SIGILL, &old_sa, NULL);
-
-    return g_ruapu_sigill_caught ? 0 : 1;
-}
-
-#if defined(__i386__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64)
-#define RUAPU_INSTCODE(isa, ...) static void ruapu_some_##isa() { asm volatile(".byte " #__VA_ARGS__ : : : ); }
-#elif __aarch64__
-#define RUAPU_INSTCODE(isa, ...) static void ruapu_some_##isa() { asm volatile(".word " #__VA_ARGS__ : : : ); }
-#elif __arm__
-#define RUAPU_INSTCODE(isa, ...) static void ruapu_some_##isa() { asm volatile(".word " #__VA_ARGS__ : : : ); }
-#endif
-
-#else // defined _WIN32 || defined __ANDROID__ || defined __linux__ || defined __APPLE__
-typedef const void* ruapu_some_inst;
-static int ruapu_detect_isa(const void* some_inst)
-{
-    // unknown platform, bare metal os ?
-    (void)some_inst;
-    return 0;
-}
-
-#define RUAPU_INSTCODE(isa, ...) static void ruapu_some_##isa() { }
-#endif // defined _WIN32 || defined __ANDROID__ || defined __linux__ || defined __APPLE__
-
-struct ruapu_isa_entry
-{
-    const char* isa;
-    ruapu_some_inst inst;
-    int capable;
-};
-
-#if defined(__i386__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64)
-RUAPU_INSTCODE(mmx, 0x0f, 0xdb, 0xc0) // pand mm0,mm0
-RUAPU_INSTCODE(sse, 0x0f, 0x54, 0xc0) // andps xmm0,xmm0
-RUAPU_INSTCODE(sse2, 0x66, 0x0f, 0xfe, 0xc0) // paddd xmm0,xmm0
-RUAPU_INSTCODE(sse3, 0xf2, 0x0f, 0x7c, 0xc0) // haddps xmm0,xmm0
-RUAPU_INSTCODE(ssse3, 0x66, 0x0f, 0x38, 0x06, 0xc0) // phsubd xmm0,xmm0
-RUAPU_INSTCODE(sse41, 0x66, 0x0f, 0x38, 0x3d, 0xc0) // pmaxsd xmm0,xmm0
-RUAPU_INSTCODE(sse42, 0x66, 0x0f, 0x38, 0x37, 0xc0) // pcmpgtq xmm0,xmm0
-RUAPU_INSTCODE(sse4a, 0x66, 0x0f, 0x79, 0xc0) // extrq xmm0,xmm0
-RUAPU_INSTCODE(xop, 0x8f, 0xe8, 0x78, 0xb6, 0xc0, 0x00)  // vpmadcswd xmm0,xmm0,xmm0,xmm0
-RUAPU_INSTCODE(avx, 0xc5, 0xfc, 0x54, 0xc0) // vandps ymm0,ymm0,ymm0
-RUAPU_INSTCODE(f16c, 0xc4, 0xe2, 0x7d, 0x13, 0xc0) // vcvtph2ps ymm0,xmm0
-RUAPU_INSTCODE(fma, 0xc4, 0xe2, 0x7d, 0x98, 0xc0) // vfmadd132ps ymm0,ymm0,ymm0
-RUAPU_INSTCODE(fma4, 0xc4, 0xe3, 0xfd, 0x68, 0xc0, 0x00) // vfmaddps ymm0,ymm0,ymm0,ymm0
-RUAPU_INSTCODE(avx2, 0xc5, 0xfd, 0xfe, 0xc0) // vpaddd ymm0,ymm0,ymm0
-RUAPU_INSTCODE(avx512f, 0x62, 0xf1, 0x7c, 0x48, 0x58, 0xc0) // vaddps zmm0,zmm0,zmm0
-RUAPU_INSTCODE(avx512bw, 0x62, 0xf1, 0x7d, 0x48, 0xfd, 0xc0) // vpaddw zmm0,zmm0,zmm0
-RUAPU_INSTCODE(avx512cd, 0x62, 0xf2, 0xfd, 0x48, 0x44, 0xc0) // vplzcntq zmm0,zmm0
-RUAPU_INSTCODE(avx512dq, 0x62, 0xf1, 0x7c, 0x48, 0x54, 0xc0) // vandps zmm0,zmm0,zmm0
-RUAPU_INSTCODE(avx512vl, 0x62, 0xf2, 0xfd, 0x28, 0x1f, 0xc0) // vpabsq ymm0,ymm0
-RUAPU_INSTCODE(avx512vnni, 0x62, 0xf2, 0x7d, 0x48, 0x52, 0xc0) // vpdpwssd zmm0,zmm0,zmm0
-RUAPU_INSTCODE(avx512bf16, 0x62, 0xf2, 0x7e, 0x48, 0x52, 0xc0) // vdpbf16ps zmm0,zmm0,zmm0
-RUAPU_INSTCODE(avx512ifma, 0x62, 0xf2, 0xfd, 0x48, 0xb4, 0xc0) // vpmadd52luq zmm0,zmm0,zmm0
-RUAPU_INSTCODE(avx512vbmi, 0x62, 0xf2, 0x7d, 0x48, 0x75, 0xc0) // vpermi2b zmm0,zmm0,zmm0
-RUAPU_INSTCODE(avx512vbmi2, 0x62, 0xf2, 0x7d, 0x48, 0x71, 0xc0) // vpshldvd zmm0,zmm0,zmm0
-RUAPU_INSTCODE(avx512fp16, 0x62, 0xf6, 0x7d, 0x48, 0x98, 0xc0) // vfmadd132ph zmm0,zmm0,zmm0
-RUAPU_INSTCODE(avxvnni, 0xc4, 0xe2, 0x7d, 0x52, 0xc0) // vpdpwssd ymm0,ymm0,ymm0
-RUAPU_INSTCODE(avxvnniint8, 0xc4, 0xe2, 0x7f, 0x50, 0xc0) // vpdpbssd ymm0,ymm0,ymm0
-RUAPU_INSTCODE(avxifma, 0xc4, 0xe2, 0xfd, 0xb4, 0xc0) // vpmadd52luq ymm0,ymm0,ymm0
-
-#elif __aarch64__ || defined(_M_ARM64)
-RUAPU_INSTCODE(neon, 0x4e20d400) // fadd v0.4s,v0.4s,v0.4s
-RUAPU_INSTCODE(vfpv4, 0x0e216800) // fcvtn v0.4h,v0.4s
-RUAPU_INSTCODE(cpuid, 0xd5380000) // mrs x0,midr_el1
-RUAPU_INSTCODE(asimdhp, 0x0e401400) // fadd v0.4h,v0.4h,v0.4h
-RUAPU_INSTCODE(asimddp, 0x4e809400) // sdot v0.4h,v0.16b,v0.16b
-RUAPU_INSTCODE(asimdfhm, 0x4e20ec00) // fmlal v0.4s,v0.4h,v0.4h
-RUAPU_INSTCODE(bf16, 0x6e40ec00) // bfmmla v0.4h,v0.8h,v0.8h
-RUAPU_INSTCODE(i8mm, 0x4e80a400) // smmla v0.4h,v0.16b,v0.16b
-RUAPU_INSTCODE(sve, 0x65608000) // fmad z0.h,p0/m,z0.h,z0.h
-RUAPU_INSTCODE(sve2, 0x44405000) // smlslb z0.h,z0.b,z0.b
-RUAPU_INSTCODE(svebf16, 0x6460e400) // bfmmla z0.s,z0.h,z0.h
-RUAPU_INSTCODE(svei8mm, 0x45009800) // smmla z0.s,z0.b,z0.b
-RUAPU_INSTCODE(svef32mm, 0x64a0e400) // fmmla z0.s,z0.s,z0.s
-
-#elif __arm__ || defined(_M_ARM)
-#if __thumb__
-RUAPU_INSTCODE(edsp, 0xfb20, 0x0000) // smlad r0,r0,r0,r0
-RUAPU_INSTCODE(neon, 0xef00, 0x0d40) // vadd.f32 q0,q0,q0
-RUAPU_INSTCODE(vfpv4, 0xffb6, 0x0600) // vcvt.f16.f32 d0,q0
-#else
-RUAPU_INSTCODE(edsp, 0xe7000010) // smlad r0,r0,r0,r0
-RUAPU_INSTCODE(neon, 0xf2000d40) // vadd.f32 q0,q0,q0
-RUAPU_INSTCODE(vfpv4, 0xf3b60600) // vcvt.f16.f32 d0,q0
-#endif
-
-#endif
-
-#undef RUAPU_INSTCODE
-
-#define RUAPU_ISAENTRY(isa) { #isa, (ruapu_some_inst)ruapu_some_##isa, 0 },
-
-struct ruapu_isa_entry g_ruapu_isa_map[] = {
-
-#if defined(__i386__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64)
-RUAPU_ISAENTRY(mmx)
-RUAPU_ISAENTRY(sse)
-RUAPU_ISAENTRY(sse2)
-RUAPU_ISAENTRY(sse3)
-RUAPU_ISAENTRY(ssse3)
-RUAPU_ISAENTRY(sse41)
-RUAPU_ISAENTRY(sse42)
-RUAPU_ISAENTRY(sse4a)
-RUAPU_ISAENTRY(xop)
-RUAPU_ISAENTRY(avx)
-RUAPU_ISAENTRY(f16c)
-RUAPU_ISAENTRY(fma)
-RUAPU_ISAENTRY(fma4)
-RUAPU_ISAENTRY(avx2)
-RUAPU_ISAENTRY(avx512f)
-RUAPU_ISAENTRY(avx512bw)
-RUAPU_ISAENTRY(avx512cd)
-RUAPU_ISAENTRY(avx512dq)
-RUAPU_ISAENTRY(avx512vl)
-RUAPU_ISAENTRY(avx512vnni)
-RUAPU_ISAENTRY(avx512bf16)
-RUAPU_ISAENTRY(avx512ifma)
-RUAPU_ISAENTRY(avx512vbmi)
-RUAPU_ISAENTRY(avx512vbmi2)
-RUAPU_ISAENTRY(avx512fp16)
-RUAPU_ISAENTRY(avxvnni)
-RUAPU_ISAENTRY(avxvnniint8)
-RUAPU_ISAENTRY(avxifma)
-
-#elif __aarch64__ || defined(_M_ARM64)
-RUAPU_ISAENTRY(neon)
-RUAPU_ISAENTRY(vfpv4)
-RUAPU_ISAENTRY(cpuid)
-RUAPU_ISAENTRY(asimdhp)
-RUAPU_ISAENTRY(asimddp)
-RUAPU_ISAENTRY(asimdfhm)
-RUAPU_ISAENTRY(bf16)
-RUAPU_ISAENTRY(i8mm)
-RUAPU_ISAENTRY(sve)
-RUAPU_ISAENTRY(sve2)
-RUAPU_ISAENTRY(svebf16)
-RUAPU_ISAENTRY(svei8mm)
-RUAPU_ISAENTRY(svef32mm)
-
-#elif __arm__ || defined(_M_ARM)
-RUAPU_ISAENTRY(edsp)
-RUAPU_ISAENTRY(neon)
-RUAPU_ISAENTRY(vfpv4)
-
-#endif
-};
-
-#undef RUAPU_ISAENTRY
-
-void ruapu_init()
-{
-    for (size_t i = 0; i < sizeof(g_ruapu_isa_map) / sizeof(g_ruapu_isa_map[0]); i++)
-    {
-        g_ruapu_isa_map[i].capable = ruapu_detect_isa(g_ruapu_isa_map[i].inst);
-    }
-}
-
-int ruapu_supports(const char* isa)
-{
-    for (size_t i = 0; i < sizeof(g_ruapu_isa_map) / sizeof(g_ruapu_isa_map[0]); i++)
-    {
-        if (strcmp(g_ruapu_isa_map[i].isa, isa) == 0)
-        {
-            return g_ruapu_isa_map[i].capable;
-        }
-    }
-
-    return 0;
-}
-
-#endif // RUAPU_IMPLEMENTATION
-
-#endif // RUAPU_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/simplemath.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/simplemath.h
deleted file mode 100644
index c3269d8b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/simplemath.h
+++ /dev/null
@@ -1,103 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_SIMPLEMATH_H
-#define NCNN_SIMPLEMATH_H
-
-#include "platform.h"
-
-#if NCNN_SIMPLEMATH
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-/*
-* ====================================================
-* discrete functions
-* ====================================================
-*/
-NCNN_EXPORT float fabs(float);
-NCNN_EXPORT float fabsf(float);
-NCNN_EXPORT float fmod(float, float);
-NCNN_EXPORT float floor(float);
-NCNN_EXPORT float floorf(float);
-NCNN_EXPORT float round(float);
-NCNN_EXPORT float roundf(float);
-NCNN_EXPORT float ceil(float);
-NCNN_EXPORT float ceilf(float);
-NCNN_EXPORT float fmaxf(float, float);
-NCNN_EXPORT float truncf(float);
-NCNN_EXPORT float frac(float);
-/*
-* ====================================================
-* trigonometric functions
-* ====================================================
-*/
-NCNN_EXPORT float sinf(float);
-NCNN_EXPORT float cosf(float);
-NCNN_EXPORT float tanf(float);
-NCNN_EXPORT float asinf(float);
-NCNN_EXPORT float acosf(float);
-NCNN_EXPORT float atanf(float);
-NCNN_EXPORT float atan2f(float, float);
-NCNN_EXPORT float tanhf(float);
-
-/*
-* ====================================================
-* power functions
-* ====================================================
-*/
-NCNN_EXPORT float sqrtf(float);
-NCNN_EXPORT float sqrt(float);
-NCNN_EXPORT float powf(float, float);
-
-/*
-* ====================================================
-* exponential and logarithm functions
-* ====================================================
-*/
-NCNN_EXPORT float expf(float);
-NCNN_EXPORT float frexp(float, int*);
-NCNN_EXPORT float logf(float);
-NCNN_EXPORT float log(float);
-NCNN_EXPORT float log10f(float);
-
-/*
-* ====================================================
-* probability functions
-* ====================================================
-*/
-NCNN_EXPORT float erf(float);
-NCNN_EXPORT float erff(float);
-NCNN_EXPORT float erfcf(float);
-
-/*
-* ====================================================
-* other functions
-* ====================================================
-*/
-NCNN_EXPORT int msb(unsigned int);
-NCNN_EXPORT float fmaf(float, float, float);
-NCNN_EXPORT float copysignf(float, float);
-NCNN_EXPORT void fesetround(int);
-NCNN_EXPORT int fegetround();
-NCNN_EXPORT float nearbyintf(float);
-
-#ifdef __cplusplus
-} // extern "C"
-#endif
-
-#endif // NCNN_SIMPLEMATH
-
-#endif // NCNN_SIMPLEMATH_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/simpleocv.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/simpleocv.h
deleted file mode 100644
index 015e48b4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/simpleocv.h
+++ /dev/null
@@ -1,503 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_SIMPLEOCV_H
-#define NCNN_SIMPLEOCV_H
-
-#include "platform.h"
-
-#if NCNN_SIMPLEOCV
-
-#include <limits.h>
-#include <string.h>
-#include "allocator.h"
-#include "mat.h"
-
-#if defined(_MSC_VER) || defined(__GNUC__)
-#pragma push_macro("min")
-#pragma push_macro("max")
-#undef min
-#undef max
-#endif
-
-#ifndef NCNN_XADD
-using ncnn::NCNN_XADD;
-#endif
-
-typedef unsigned char uchar;
-typedef unsigned short ushort;
-typedef unsigned int uint;
-
-enum
-{
-    CV_LOAD_IMAGE_UNCHANGED = -1,
-    CV_LOAD_IMAGE_GRAYSCALE = 0,
-    CV_LOAD_IMAGE_COLOR = 1,
-};
-
-enum
-{
-    CV_IMWRITE_JPEG_QUALITY = 1
-};
-
-// minimal opencv style data structure implementation
-namespace cv {
-
-template<typename _Tp>
-static inline _Tp saturate_cast(int v)
-{
-    return _Tp(v);
-}
-template<>
-inline uchar saturate_cast<uchar>(int v)
-{
-    return (uchar)((unsigned)v <= UCHAR_MAX ? v : v > 0 ? UCHAR_MAX : 0);
-}
-
-template<typename _Tp>
-struct Scalar_
-{
-    Scalar_()
-    {
-        v[0] = 0;
-        v[1] = 0;
-        v[2] = 0;
-        v[3] = 0;
-    }
-    Scalar_(_Tp _v0)
-    {
-        v[0] = _v0;
-        v[1] = 0;
-        v[2] = 0;
-        v[3] = 0;
-    }
-    Scalar_(_Tp _v0, _Tp _v1, _Tp _v2)
-    {
-        v[0] = _v0;
-        v[1] = _v1;
-        v[2] = _v2;
-        v[3] = 0;
-    }
-    Scalar_(_Tp _v0, _Tp _v1, _Tp _v2, _Tp _v3)
-    {
-        v[0] = _v0;
-        v[1] = _v1;
-        v[2] = _v2;
-        v[3] = _v3;
-    }
-
-    const _Tp operator[](const int i) const
-    {
-        return v[i];
-    }
-
-    _Tp operator[](const int i)
-    {
-        return v[i];
-    }
-
-    _Tp v[4];
-};
-
-typedef Scalar_<uchar> Scalar;
-
-template<typename _Tp>
-struct Point_
-{
-    Point_()
-        : x(0), y(0)
-    {
-    }
-    Point_(_Tp _x, _Tp _y)
-        : x(_x), y(_y)
-    {
-    }
-
-    template<typename _Tp2>
-    operator Point_<_Tp2>() const
-    {
-        return Point_<_Tp2>(saturate_cast<_Tp2>(x), saturate_cast<_Tp2>(y));
-    }
-
-    _Tp x;
-    _Tp y;
-};
-
-typedef Point_<int> Point;
-typedef Point_<float> Point2f;
-
-template<typename _Tp>
-struct Size_
-{
-    Size_()
-        : width(0), height(0)
-    {
-    }
-    Size_(_Tp _w, _Tp _h)
-        : width(_w), height(_h)
-    {
-    }
-
-    template<typename _Tp2>
-    operator Size_<_Tp2>() const
-    {
-        return Size_<_Tp2>(saturate_cast<_Tp2>(width), saturate_cast<_Tp2>(height));
-    }
-
-    _Tp width;
-    _Tp height;
-};
-
-typedef Size_<int> Size;
-typedef Size_<float> Size2f;
-
-template<typename _Tp>
-struct Rect_
-{
-    Rect_()
-        : x(0), y(0), width(0), height(0)
-    {
-    }
-    Rect_(_Tp _x, _Tp _y, _Tp _w, _Tp _h)
-        : x(_x), y(_y), width(_w), height(_h)
-    {
-    }
-    Rect_(Point_<_Tp> _p, Size_<_Tp> _size)
-        : x(_p.x), y(_p.y), width(_size.width), height(_size.height)
-    {
-    }
-
-    template<typename _Tp2>
-    operator Rect_<_Tp2>() const
-    {
-        return Rect_<_Tp2>(saturate_cast<_Tp2>(x), saturate_cast<_Tp2>(y), saturate_cast<_Tp2>(width), saturate_cast<_Tp2>(height));
-    }
-
-    _Tp x;
-    _Tp y;
-    _Tp width;
-    _Tp height;
-
-    // area
-    _Tp area() const
-    {
-        return width * height;
-    }
-};
-
-template<typename _Tp>
-static inline Rect_<_Tp>& operator&=(Rect_<_Tp>& a, const Rect_<_Tp>& b)
-{
-    _Tp x1 = std::max(a.x, b.x), y1 = std::max(a.y, b.y);
-    a.width = std::min(a.x + a.width, b.x + b.width) - x1;
-    a.height = std::min(a.y + a.height, b.y + b.height) - y1;
-    a.x = x1;
-    a.y = y1;
-    if (a.width <= 0 || a.height <= 0)
-        a = Rect_<_Tp>();
-    return a;
-}
-
-template<typename _Tp>
-static inline Rect_<_Tp>& operator|=(Rect_<_Tp>& a, const Rect_<_Tp>& b)
-{
-    _Tp x1 = std::min(a.x, b.x), y1 = std::min(a.y, b.y);
-    a.width = std::max(a.x + a.width, b.x + b.width) - x1;
-    a.height = std::max(a.y + a.height, b.y + b.height) - y1;
-    a.x = x1;
-    a.y = y1;
-    return a;
-}
-
-template<typename _Tp>
-static inline Rect_<_Tp> operator&(const Rect_<_Tp>& a, const Rect_<_Tp>& b)
-{
-    Rect_<_Tp> c = a;
-    return c &= b;
-}
-
-template<typename _Tp>
-static inline Rect_<_Tp> operator|(const Rect_<_Tp>& a, const Rect_<_Tp>& b)
-{
-    Rect_<_Tp> c = a;
-    return c |= b;
-}
-
-typedef Rect_<int> Rect;
-typedef Rect_<float> Rect2f;
-
-#define CV_8UC1  1
-#define CV_8UC3  3
-#define CV_8UC4  4
-#define CV_32FC1 4
-
-struct NCNN_EXPORT Mat
-{
-    Mat()
-        : data(0), refcount(0), rows(0), cols(0), c(0)
-    {
-    }
-
-    Mat(int _rows, int _cols, int flags)
-        : data(0), refcount(0)
-    {
-        create(_rows, _cols, flags);
-    }
-
-    // copy
-    Mat(const Mat& m)
-        : data(m.data), refcount(m.refcount)
-    {
-        if (refcount)
-            NCNN_XADD(refcount, 1);
-
-        rows = m.rows;
-        cols = m.cols;
-        c = m.c;
-    }
-
-    Mat(int _rows, int _cols, int flags, void* _data)
-        : data((unsigned char*)_data), refcount(0)
-    {
-        rows = _rows;
-        cols = _cols;
-        c = flags;
-    }
-
-    ~Mat()
-    {
-        release();
-    }
-
-    // assign
-    Mat& operator=(const Mat& m)
-    {
-        if (this == &m)
-            return *this;
-
-        if (m.refcount)
-            NCNN_XADD(m.refcount, 1);
-
-        release();
-
-        data = m.data;
-        refcount = m.refcount;
-
-        rows = m.rows;
-        cols = m.cols;
-        c = m.c;
-
-        return *this;
-    }
-
-    Mat& operator=(const Scalar& s)
-    {
-        if (total() > 0)
-        {
-            uchar* p = data;
-            for (int i = 0; i < cols * rows; i++)
-            {
-                for (int j = 0; j < c; j++)
-                {
-                    *p++ = s[j];
-                }
-            }
-        }
-
-        return *this;
-    }
-
-    void create(int _rows, int _cols, int flags)
-    {
-        release();
-
-        rows = _rows;
-        cols = _cols;
-        c = flags;
-
-        if (total() > 0)
-        {
-            // refcount address must be aligned, so we expand totalsize here
-            size_t totalsize = (total() + 3) >> 2 << 2;
-            data = (uchar*)ncnn::fastMalloc(totalsize + (int)sizeof(*refcount));
-            refcount = (int*)(((uchar*)data) + totalsize);
-            *refcount = 1;
-        }
-    }
-
-    void release()
-    {
-        if (refcount && NCNN_XADD(refcount, -1) == 1)
-            ncnn::fastFree(data);
-
-        data = 0;
-
-        rows = 0;
-        cols = 0;
-        c = 0;
-
-        refcount = 0;
-    }
-
-    Mat clone() const
-    {
-        if (empty())
-            return Mat();
-
-        Mat m(rows, cols, c);
-
-        if (total() > 0)
-        {
-            memcpy(m.data, data, total());
-        }
-
-        return m;
-    }
-
-    bool empty() const
-    {
-        return data == 0 || total() == 0;
-    }
-
-    int channels() const
-    {
-        return c;
-    }
-
-    int type() const
-    {
-        return c;
-    }
-
-    size_t total() const
-    {
-        return cols * rows * c;
-    }
-
-    const uchar* ptr(int y) const
-    {
-        return data + y * cols * c;
-    }
-
-    uchar* ptr(int y)
-    {
-        return data + y * cols * c;
-    }
-
-    template<typename _Tp>
-    const _Tp* ptr(int y) const
-    {
-        return (const _Tp*)(data + y * cols * c);
-    }
-
-    template<typename _Tp>
-    _Tp* ptr(int y)
-    {
-        return (_Tp*)(data + y * cols * c);
-    }
-
-    // roi
-    Mat operator()(const Rect& roi) const
-    {
-        if (empty())
-            return Mat();
-
-        Mat m(roi.height, roi.width, c);
-
-        int sy = roi.y;
-        for (int y = 0; y < roi.height; y++)
-        {
-            const uchar* sptr = ptr(sy) + roi.x * c;
-            uchar* dptr = m.ptr(y);
-            memcpy(dptr, sptr, roi.width * c);
-            sy++;
-        }
-
-        return m;
-    }
-
-    uchar* data;
-
-    // pointer to the reference counter;
-    // when points to user-allocated data, the pointer is NULL
-    int* refcount;
-
-    int rows;
-    int cols;
-
-    int c;
-};
-
-enum ImreadModes
-{
-    IMREAD_UNCHANGED = -1,
-    IMREAD_GRAYSCALE = 0,
-    IMREAD_COLOR = 1
-};
-
-NCNN_EXPORT Mat imread(const std::string& path, int flags = IMREAD_COLOR);
-
-NCNN_EXPORT Mat imdecode(const std::vector<uchar>& buf, int flags = IMREAD_COLOR);
-
-enum ImwriteFlags
-{
-    IMWRITE_JPEG_QUALITY = 1
-};
-
-NCNN_EXPORT bool imwrite(const std::string& path, const Mat& m, const std::vector<int>& params = std::vector<int>());
-
-NCNN_EXPORT void imshow(const std::string& name, const Mat& m);
-
-NCNN_EXPORT int waitKey(int delay = 0);
-
-#if NCNN_PIXEL
-NCNN_EXPORT void resize(const Mat& src, Mat& dst, const Size& size, float sw = 0.f, float sh = 0.f, int flags = 0);
-#endif // NCNN_PIXEL
-
-#if NCNN_PIXEL_DRAWING
-
-enum
-{
-    FILLED = -1
-};
-
-NCNN_EXPORT void rectangle(Mat& img, Point pt1, Point pt2, const Scalar& color, int thickness = 1);
-
-NCNN_EXPORT void rectangle(Mat& img, Rect rec, const Scalar& color, int thickness = 1);
-
-NCNN_EXPORT void circle(Mat& img, Point center, int radius, const Scalar& color, int thickness = 1);
-
-NCNN_EXPORT void line(Mat& img, Point p0, Point p1, const Scalar& color, int thickness = 1);
-
-enum
-{
-    FONT_HERSHEY_SIMPLEX = 0
-};
-
-NCNN_EXPORT void putText(Mat& img, const std::string& text, Point org, int fontFace, double fontScale, Scalar color, int thickness = 1);
-
-NCNN_EXPORT Size getTextSize(const std::string& text, int fontFace, double fontScale, int thickness, int* baseLine);
-
-#endif // NCNN_PIXEL_DRAWING
-
-} // namespace cv
-
-#if defined(_MSC_VER) || defined(__GNUC__)
-#pragma pop_macro("min")
-#pragma pop_macro("max")
-#endif
-
-#endif // NCNN_SIMPLEOCV
-
-#endif // NCNN_SIMPLEOCV_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/simpleomp.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/simpleomp.h
deleted file mode 100644
index 13e24529..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/simpleomp.h
+++ /dev/null
@@ -1,53 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_SIMPLEOMP_H
-#define NCNN_SIMPLEOMP_H
-
-#include "platform.h"
-
-#if NCNN_SIMPLEOMP
-
-#include <stdint.h>
-
-// This minimal openmp runtime implementation only supports the llvm openmp abi
-// and only supports #pragma omp parallel for num_threads(X)
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-NCNN_EXPORT int omp_get_max_threads();
-
-NCNN_EXPORT void omp_set_num_threads(int num_threads);
-
-NCNN_EXPORT int omp_get_dynamic();
-
-NCNN_EXPORT void omp_set_dynamic(int dynamic);
-
-NCNN_EXPORT int omp_get_num_threads();
-
-NCNN_EXPORT int omp_get_thread_num();
-
-NCNN_EXPORT int kmp_get_blocktime();
-
-NCNN_EXPORT void kmp_set_blocktime(int blocktime);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif // NCNN_SIMPLEOMP
-
-#endif // NCNN_SIMPLEOMP_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/simplestl.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/simplestl.h
deleted file mode 100644
index 00ff4680..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/simplestl.h
+++ /dev/null
@@ -1,565 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_SIMPLESTL_H
-#define NCNN_SIMPLESTL_H
-
-#include <stddef.h>
-#include <stdint.h>
-#include <string.h>
-
-#if !NCNN_SIMPLESTL
-
-#include <new>
-
-#else
-
-// allocation functions
-NCNN_EXPORT void* operator new(size_t size);
-NCNN_EXPORT void* operator new[](size_t size);
-// placement allocation functions
-NCNN_EXPORT void* operator new(size_t size, void* ptr);
-NCNN_EXPORT void* operator new[](size_t size, void* ptr);
-// deallocation functions
-NCNN_EXPORT void operator delete(void* ptr);
-NCNN_EXPORT void operator delete[](void* ptr);
-// deallocation functions since c++14
-#if __cplusplus >= 201402L
-NCNN_EXPORT void operator delete(void* ptr, size_t sz);
-NCNN_EXPORT void operator delete[](void* ptr, size_t sz);
-#endif
-// placement deallocation functions
-NCNN_EXPORT void operator delete(void* ptr, void* voidptr2);
-NCNN_EXPORT void operator delete[](void* ptr, void* voidptr2);
-
-#endif
-
-// minimal stl data structure implementation
-namespace std {
-
-template<typename T>
-const T& max(const T& a, const T& b)
-{
-    return (a < b) ? b : a;
-}
-
-template<typename T>
-const T& min(const T& a, const T& b)
-{
-    return (a > b) ? b : a;
-}
-
-template<typename T>
-void swap(T& a, T& b)
-{
-    T temp(a);
-    a = b;
-    b = temp;
-}
-
-template<typename T1, typename T2>
-struct pair
-{
-    pair()
-        : first(), second()
-    {
-    }
-    pair(const T1& t1, const T2& t2)
-        : first(t1), second(t2)
-    {
-    }
-
-    T1 first;
-    T2 second;
-};
-
-template<typename T1, typename T2>
-bool operator==(const pair<T1, T2>& x, const pair<T1, T2>& y)
-{
-    return (x.first == y.first && x.second == y.second);
-}
-template<typename T1, typename T2>
-bool operator<(const pair<T1, T2>& x, const pair<T1, T2>& y)
-{
-    return x.first < y.first || (!(y.first < x.first) && x.second < y.second);
-}
-template<typename T1, typename T2>
-bool operator!=(const pair<T1, T2>& x, const pair<T1, T2>& y)
-{
-    return !(x == y);
-}
-template<typename T1, typename T2>
-bool operator>(const pair<T1, T2>& x, const pair<T1, T2>& y)
-{
-    return y < x;
-}
-template<typename T1, typename T2>
-bool operator<=(const pair<T1, T2>& x, const pair<T1, T2>& y)
-{
-    return !(y < x);
-}
-template<typename T1, typename T2>
-bool operator>=(const pair<T1, T2>& x, const pair<T1, T2>& y)
-{
-    return !(x < y);
-}
-
-template<typename T1, typename T2>
-pair<T1, T2> make_pair(const T1& t1, const T2& t2)
-{
-    return pair<T1, T2>(t1, t2);
-}
-
-template<typename T>
-struct node
-{
-    node* prev_;
-    node* next_;
-    T data_;
-
-    node()
-        : prev_(0), next_(0), data_()
-    {
-    }
-    node(const T& t)
-        : prev_(0), next_(0), data_(t)
-    {
-    }
-};
-
-template<typename T>
-struct iter_list
-{
-    iter_list()
-        : curr_(0)
-    {
-    }
-    iter_list(node<T>* n)
-        : curr_(n)
-    {
-    }
-    iter_list(const iter_list& i)
-        : curr_(i.curr_)
-    {
-    }
-    ~iter_list()
-    {
-    }
-
-    iter_list& operator=(const iter_list& i)
-    {
-        curr_ = i.curr_;
-        return *this;
-    }
-
-    T& operator*()
-    {
-        return curr_->data_;
-    }
-    T* operator->()
-    {
-        return &(curr_->data_);
-    }
-
-    bool operator==(const iter_list& i)
-    {
-        return curr_ == i.curr_;
-    }
-    bool operator!=(const iter_list& i)
-    {
-        return curr_ != i.curr_;
-    }
-
-    iter_list& operator++()
-    {
-        curr_ = curr_->next_;
-        return *this;
-    }
-    iter_list& operator--()
-    {
-        curr_ = curr_->prev_;
-        return *this;
-    }
-
-    node<T>* curr_;
-};
-
-template<typename T>
-struct list
-{
-    typedef iter_list<T> iterator;
-
-    list()
-    {
-        head_ = new node<T>();
-        tail_ = head_;
-        count_ = 0;
-    }
-    ~list()
-    {
-        clear();
-        delete head_;
-    }
-    list(const list& l)
-    {
-        head_ = new node<T>();
-        tail_ = head_;
-        count_ = 0;
-
-        for (iter_list<T> i = l.begin(); i != l.end(); ++i)
-        {
-            push_back(*i);
-        }
-    }
-
-    list& operator=(const list& l)
-    {
-        if (this == &l)
-        {
-            return *this;
-        }
-        clear();
-
-        for (iter_list<T> i = l.begin(); i != l.end(); ++i)
-        {
-            push_back(*i);
-        }
-        return *this;
-    }
-
-    void clear()
-    {
-        while (count_ > 0)
-        {
-            pop_front();
-        }
-    }
-
-    void pop_front()
-    {
-        if (count_ > 0)
-        {
-            head_ = head_->next_;
-            delete head_->prev_;
-            head_->prev_ = 0;
-            --count_;
-        }
-    }
-
-    size_t size() const
-    {
-        return count_;
-    }
-    iter_list<T> begin() const
-    {
-        return iter_list<T>(head_);
-    }
-    iter_list<T> end() const
-    {
-        return iter_list<T>(tail_);
-    }
-    bool empty() const
-    {
-        return count_ == 0;
-    }
-
-    void push_back(const T& t)
-    {
-        if (count_ == 0)
-        {
-            head_ = new node<T>(t);
-            head_->prev_ = 0;
-            head_->next_ = tail_;
-            tail_->prev_ = head_;
-            count_ = 1;
-        }
-        else
-        {
-            node<T>* temp = new node<T>(t);
-            temp->prev_ = tail_->prev_;
-            temp->next_ = tail_;
-            tail_->prev_->next_ = temp;
-            tail_->prev_ = temp;
-            ++count_;
-        }
-    }
-
-    iter_list<T> erase(iter_list<T> pos)
-    {
-        if (pos != end())
-        {
-            node<T>* temp = pos.curr_;
-            if (temp == head_)
-            {
-                ++pos;
-                temp->next_->prev_ = 0;
-                head_ = temp->next_;
-            }
-            else
-            {
-                --pos;
-                temp->next_->prev_ = temp->prev_;
-                temp->prev_->next_ = temp->next_;
-                ++pos;
-            }
-            delete temp;
-            --count_;
-        }
-        return pos;
-    }
-
-protected:
-    node<T>* head_;
-    node<T>* tail_;
-    size_t count_;
-};
-
-template<typename T>
-struct greater
-{
-    bool operator()(const T& x, const T& y) const
-    {
-        return (x > y);
-    }
-};
-
-template<typename T>
-struct less
-{
-    bool operator()(const T& x, const T& y) const
-    {
-        return (x < y);
-    }
-};
-
-template<typename RandomAccessIter, typename Compare>
-void partial_sort(RandomAccessIter first, RandomAccessIter middle, RandomAccessIter last, Compare comp)
-{
-    // [TODO] heap sort should be used here, but we simply use bubble sort now
-    for (RandomAccessIter i = first; i < middle; ++i)
-    {
-        // bubble sort
-        for (RandomAccessIter j = last - 1; j > first; --j)
-        {
-            if (comp(*j, *(j - 1)))
-            {
-                swap(*j, *(j - 1));
-            }
-        }
-    }
-}
-
-template<typename T>
-struct vector
-{
-    vector()
-        : data_(0), size_(0), capacity_(0)
-    {
-    }
-    vector(const size_t new_size, const T& value = T())
-        : data_(0), size_(0), capacity_(0)
-    {
-        resize(new_size, value);
-    }
-    ~vector()
-    {
-        clear();
-    }
-    vector(const vector& v)
-        : data_(0), size_(0), capacity_(0)
-    {
-        resize(v.size());
-        for (size_t i = 0; i < size_; i++)
-        {
-            data_[i] = v.data_[i];
-        }
-    }
-
-    vector& operator=(const vector& v)
-    {
-        if (this == &v)
-        {
-            return *this;
-        }
-        resize(0);
-        resize(v.size());
-        for (size_t i = 0; i < size_; i++)
-        {
-            data_[i] = v.data_[i];
-        }
-        return *this;
-    }
-
-    void resize(const size_t new_size, const T& value = T())
-    {
-        try_alloc(new_size);
-        if (new_size > size_)
-        {
-            for (size_t i = size_; i < new_size; i++)
-            {
-                new (&data_[i]) T(value);
-            }
-        }
-        else if (new_size < size_)
-        {
-            for (size_t i = new_size; i < size_; i++)
-            {
-                data_[i].~T();
-            }
-        }
-        size_ = new_size;
-    }
-
-    void clear()
-    {
-        for (size_t i = 0; i < size_; i++)
-        {
-            data_[i].~T();
-        }
-        delete[](char*) data_;
-        data_ = 0;
-        size_ = 0;
-        capacity_ = 0;
-    }
-
-    T* data() const
-    {
-        return data_;
-    }
-    size_t size() const
-    {
-        return size_;
-    }
-    T& operator[](size_t i) const
-    {
-        return data_[i];
-    }
-    T* begin() const
-    {
-        return &data_[0];
-    }
-    T* end() const
-    {
-        return &data_[size_];
-    }
-    bool empty() const
-    {
-        return size_ == 0;
-    }
-
-    void push_back(const T& t)
-    {
-        try_alloc(size_ + 1);
-        new (&data_[size_]) T(t);
-        size_++;
-    }
-
-    void insert(T* pos, T* b, T* e)
-    {
-        vector* v = 0;
-        if (b >= begin() && b < end())
-        {
-            //the same vector
-            v = new vector(*this);
-            b = v->begin() + (b - begin());
-            e = v->begin() + (e - begin());
-        }
-        size_t diff = pos - begin();
-        try_alloc(size_ + (e - b));
-        pos = begin() + diff;
-        memmove(pos + (e - b), pos, (end() - pos) * sizeof(T));
-        size_t len = e - b;
-        size_ += len;
-        for (size_t i = 0; i < len; i++)
-        {
-            *pos = *b;
-            pos++;
-            b++;
-        }
-        delete v;
-    }
-
-    T* erase(T* pos)
-    {
-        pos->~T();
-        memmove(pos, pos + 1, (end() - pos - 1) * sizeof(T));
-        size_--;
-        return pos;
-    }
-
-protected:
-    T* data_;
-    size_t size_;
-    size_t capacity_;
-    void try_alloc(size_t new_size)
-    {
-        if (new_size * 3 / 2 > capacity_ / 2)
-        {
-            capacity_ = new_size * 2;
-            T* new_data = (T*)new char[capacity_ * sizeof(T)];
-            memset(static_cast<void*>(new_data), 0, capacity_ * sizeof(T));
-            if (data_)
-            {
-                memmove(new_data, data_, sizeof(T) * size_);
-                delete[](char*) data_;
-            }
-            data_ = new_data;
-        }
-    }
-};
-
-struct NCNN_EXPORT string : public vector<char>
-{
-    string()
-    {
-    }
-    string(const char* str)
-    {
-        size_t len = strlen(str);
-        resize(len);
-        memcpy(data_, str, len);
-    }
-    const char* c_str() const
-    {
-        return (const char*)data_;
-    }
-    bool operator==(const string& str2) const
-    {
-        return strcmp(data_, str2.data_) == 0;
-    }
-    bool operator==(const char* str2) const
-    {
-        return strcmp(data_, str2) == 0;
-    }
-    bool operator!=(const char* str2) const
-    {
-        return strcmp(data_, str2) != 0;
-    }
-    string& operator+=(const string& str1)
-    {
-        insert(end(), str1.begin(), str1.end());
-        return *this;
-    }
-};
-
-inline string operator+(const string& str1, const string& str2)
-{
-    string str(str1);
-    str.insert(str.end(), str2.begin(), str2.end());
-    return str;
-}
-
-} // namespace std
-
-#endif // NCNN_SIMPLESTL_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/simplevk.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/simplevk.h
deleted file mode 100644
index 6146c813..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/simplevk.h
+++ /dev/null
@@ -1,2207 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2023 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_SIMPLEVK_H
-#define NCNN_SIMPLEVK_H
-
-#include "platform.h"
-
-#if NCNN_VULKAN
-#if NCNN_SIMPLEVK
-
-#define VK_VERSION_1_0 1
-
-#ifdef __cplusplus
-extern "C" {
-#endif // __cplusplus
-
-/*
-***************************************************************************************************
-*   Platform-specific directives and type declarations
-***************************************************************************************************
-*/
-
-/* Platform-specific calling convention macros.
- *
- * Platforms should define these so that Vulkan clients call Vulkan commands
- * with the same calling conventions that the Vulkan implementation expects.
- *
- * VKAPI_ATTR - Placed before the return type in function declarations.
- *              Useful for C++11 and GCC/Clang-style function attribute syntax.
- * VKAPI_CALL - Placed after the return type in function declarations.
- *              Useful for MSVC-style calling convention syntax.
- * VKAPI_PTR  - Placed between the '(' and '*' in function pointer types.
- *
- * Function declaration:  VKAPI_ATTR void VKAPI_CALL vkCommand(void);
- * Function pointer type: typedef void (VKAPI_PTR *PFN_vkCommand)(void);
- */
-#if defined(_WIN32)
-// On Windows, Vulkan commands use the stdcall convention
-#define VKAPI_ATTR
-#define VKAPI_CALL __stdcall
-#define VKAPI_PTR  VKAPI_CALL
-#elif defined(__ANDROID__) && defined(__ARM_ARCH) && __ARM_ARCH < 7
-#error "Vulkan isn't supported for the 'armeabi' NDK ABI"
-#elif defined(__ANDROID__) && defined(__ARM_ARCH) && __ARM_ARCH >= 7 && defined(__ARM_32BIT_STATE)
-// On Android 32-bit ARM targets, Vulkan functions use the "hardfloat"
-// calling convention, i.e. float parameters are passed in registers. This
-// is true even if the rest of the application passes floats on the stack,
-// as it does by default when compiling for the armeabi-v7a NDK ABI.
-#define VKAPI_ATTR __attribute__((pcs("aapcs-vfp")))
-#define VKAPI_CALL
-#define VKAPI_PTR VKAPI_ATTR
-#else
-// On other platforms, use the default calling convention
-#define VKAPI_ATTR
-#define VKAPI_CALL
-#define VKAPI_PTR
-#endif
-
-#include <stddef.h>
-
-#if !defined(VK_NO_STDINT_H)
-#if defined(_MSC_VER) && (_MSC_VER < 1600)
-typedef signed __int8 int8_t;
-typedef unsigned __int8 uint8_t;
-typedef signed __int16 int16_t;
-typedef unsigned __int16 uint16_t;
-typedef signed __int32 int32_t;
-typedef unsigned __int32 uint32_t;
-typedef signed __int64 int64_t;
-typedef unsigned __int64 uint64_t;
-#else
-#include <stdint.h>
-#endif
-#endif // !defined(VK_NO_STDINT_H)
-
-#ifdef __cplusplus
-} // extern "C"
-#endif // __cplusplus
-
-#define VK_MAKE_VERSION(major, minor, patch) (((major) << 22) | ((minor) << 12) | (patch))
-
-// Vulkan 1.0 version number
-#define VK_API_VERSION_1_0 VK_MAKE_VERSION(1, 0, 0) // Patch version should always be set to 0
-
-#define VK_VERSION_MAJOR(version) ((uint32_t)(version) >> 22)
-#define VK_VERSION_MINOR(version) (((uint32_t)(version) >> 12) & 0x3ff)
-#define VK_VERSION_PATCH(version) ((uint32_t)(version)&0xfff)
-// Version of this file
-#define VK_HEADER_VERSION 0
-
-#define VK_NULL_HANDLE 0
-
-#define VK_DEFINE_HANDLE(object) typedef struct object##_T* object;
-
-#if !defined(VK_DEFINE_NON_DISPATCHABLE_HANDLE)
-#if defined(__LP64__) || defined(_WIN64) || (defined(__x86_64__) && !defined(__ILP32__)) || defined(_M_X64) || defined(__ia64) || defined(_M_IA64) || defined(__aarch64__) || defined(__powerpc64__)
-#define VK_DEFINE_NON_DISPATCHABLE_HANDLE(object) typedef struct object##_T* object;
-#else
-#define VK_DEFINE_NON_DISPATCHABLE_HANDLE(object) typedef uint64_t object;
-#endif
-#endif
-
-typedef uint32_t VkBool32;
-typedef uint64_t VkDeviceSize;
-typedef uint32_t VkFlags;
-typedef uint32_t VkSampleMask;
-
-VK_DEFINE_HANDLE(VkCommandBuffer)
-VK_DEFINE_HANDLE(VkDevice)
-VK_DEFINE_HANDLE(VkInstance)
-VK_DEFINE_HANDLE(VkPhysicalDevice)
-VK_DEFINE_HANDLE(VkQueue)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkBuffer)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkBufferView)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkCommandPool)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkDescriptorPool)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkDescriptorSet)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkDescriptorSetLayout)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkDeviceMemory)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkFence)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkFramebuffer)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkImage)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkImageView)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkPipeline)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkPipelineCache)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkPipelineLayout)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkQueryPool)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkRenderPass)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkSampler)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkSemaphore)
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkShaderModule)
-#define VK_UUID_SIZE                     16U
-#define VK_ATTACHMENT_UNUSED             (~0U)
-#define VK_FALSE                         0U
-#define VK_LOD_CLAMP_NONE                1000.0F
-#define VK_QUEUE_FAMILY_IGNORED          (~0U)
-#define VK_REMAINING_ARRAY_LAYERS        (~0U)
-#define VK_REMAINING_MIP_LEVELS          (~0U)
-#define VK_SUBPASS_EXTERNAL              (~0U)
-#define VK_TRUE                          1U
-#define VK_WHOLE_SIZE                    (~0ULL)
-#define VK_MAX_MEMORY_TYPES              32U
-#define VK_MAX_MEMORY_HEAPS              16U
-#define VK_MAX_PHYSICAL_DEVICE_NAME_SIZE 256U
-#define VK_MAX_EXTENSION_NAME_SIZE       256U
-#define VK_MAX_DESCRIPTION_SIZE          256U
-
-typedef enum VkPipelineCacheHeaderVersion
-{
-    VK_PIPELINE_CACHE_HEADER_VERSION_ONE = 1,
-    VK_PIPELINE_CACHE_HEADER_VERSION_MAX_ENUM = 0x7FFFFFFF
-} VkPipelineCacheHeaderVersion;
-
-typedef enum VkResult
-{
-    VK_SUCCESS = 0,
-    VK_NOT_READY = 1,
-    VK_TIMEOUT = 2,
-    VK_EVENT_SET = 3,
-    VK_EVENT_RESET = 4,
-    VK_INCOMPLETE = 5,
-    VK_ERROR_OUT_OF_HOST_MEMORY = -1,
-    VK_ERROR_OUT_OF_DEVICE_MEMORY = -2,
-    VK_ERROR_INITIALIZATION_FAILED = -3,
-    VK_ERROR_DEVICE_LOST = -4,
-    VK_ERROR_MEMORY_MAP_FAILED = -5,
-    VK_ERROR_LAYER_NOT_PRESENT = -6,
-    VK_ERROR_EXTENSION_NOT_PRESENT = -7,
-    VK_ERROR_FEATURE_NOT_PRESENT = -8,
-    VK_ERROR_INCOMPATIBLE_DRIVER = -9,
-    VK_ERROR_TOO_MANY_OBJECTS = -10,
-    VK_ERROR_FORMAT_NOT_SUPPORTED = -11,
-    VK_ERROR_FRAGMENTED_POOL = -12,
-    VK_ERROR_SURFACE_LOST_KHR = -1000000000,
-    VK_ERROR_NATIVE_WINDOW_IN_USE_KHR = -1000000001,
-    VK_SUBOPTIMAL_KHR = 1000001003,
-    VK_ERROR_OUT_OF_DATE_KHR = -1000001004,
-    VK_ERROR_INCOMPATIBLE_DISPLAY_KHR = -1000003001,
-    VK_ERROR_VALIDATION_FAILED_EXT = -1000011001,
-    VK_ERROR_INVALID_SHADER_NV = -1000012000,
-    VK_ERROR_OUT_OF_POOL_MEMORY_KHR = -1000069000,
-    VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR = -1000072003,
-    VK_RESULT_MAX_ENUM = 0x7FFFFFFF
-} VkResult;
-
-typedef enum VkStructureType
-{
-    VK_STRUCTURE_TYPE_APPLICATION_INFO = 0,
-    VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO = 1,
-    VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO = 2,
-    VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO = 3,
-    VK_STRUCTURE_TYPE_SUBMIT_INFO = 4,
-    VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO = 5,
-    VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE = 6,
-    VK_STRUCTURE_TYPE_BIND_SPARSE_INFO = 7,
-    VK_STRUCTURE_TYPE_FENCE_CREATE_INFO = 8,
-    VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO = 9,
-    VK_STRUCTURE_TYPE_EVENT_CREATE_INFO = 10,
-    VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO = 11,
-    VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO = 12,
-    VK_STRUCTURE_TYPE_BUFFER_VIEW_CREATE_INFO = 13,
-    VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO = 14,
-    VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO = 15,
-    VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO = 16,
-    VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO = 17,
-    VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO = 18,
-    VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO = 19,
-    VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO = 20,
-    VK_STRUCTURE_TYPE_PIPELINE_TESSELLATION_STATE_CREATE_INFO = 21,
-    VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO = 22,
-    VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO = 23,
-    VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO = 24,
-    VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO = 25,
-    VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO = 26,
-    VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO = 27,
-    VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO = 28,
-    VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO = 29,
-    VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO = 30,
-    VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO = 31,
-    VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO = 32,
-    VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO = 33,
-    VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO = 34,
-    VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET = 35,
-    VK_STRUCTURE_TYPE_COPY_DESCRIPTOR_SET = 36,
-    VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO = 37,
-    VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO = 38,
-    VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO = 39,
-    VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO = 40,
-    VK_STRUCTURE_TYPE_COMMAND_BUFFER_INHERITANCE_INFO = 41,
-    VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO = 42,
-    VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO = 43,
-    VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER = 44,
-    VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER = 45,
-    VK_STRUCTURE_TYPE_MEMORY_BARRIER = 46,
-    VK_STRUCTURE_TYPE_LOADER_INSTANCE_CREATE_INFO = 47,
-    VK_STRUCTURE_TYPE_LOADER_DEVICE_CREATE_INFO = 48,
-    VK_STRUCTURE_TYPE_SWAPCHAIN_CREATE_INFO_KHR = 1000001000,
-    VK_STRUCTURE_TYPE_PRESENT_INFO_KHR = 1000001001,
-    VK_STRUCTURE_TYPE_DISPLAY_MODE_CREATE_INFO_KHR = 1000002000,
-    VK_STRUCTURE_TYPE_DISPLAY_SURFACE_CREATE_INFO_KHR = 1000002001,
-    VK_STRUCTURE_TYPE_DISPLAY_PRESENT_INFO_KHR = 1000003000,
-    VK_STRUCTURE_TYPE_XLIB_SURFACE_CREATE_INFO_KHR = 1000004000,
-    VK_STRUCTURE_TYPE_XCB_SURFACE_CREATE_INFO_KHR = 1000005000,
-    VK_STRUCTURE_TYPE_WAYLAND_SURFACE_CREATE_INFO_KHR = 1000006000,
-    VK_STRUCTURE_TYPE_MIR_SURFACE_CREATE_INFO_KHR = 1000007000,
-    VK_STRUCTURE_TYPE_ANDROID_SURFACE_CREATE_INFO_KHR = 1000008000,
-    VK_STRUCTURE_TYPE_WIN32_SURFACE_CREATE_INFO_KHR = 1000009000,
-    VK_STRUCTURE_TYPE_DEBUG_REPORT_CALLBACK_CREATE_INFO_EXT = 1000011000,
-    VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_RASTERIZATION_ORDER_AMD = 1000018000,
-    VK_STRUCTURE_TYPE_DEBUG_MARKER_OBJECT_NAME_INFO_EXT = 1000022000,
-    VK_STRUCTURE_TYPE_DEBUG_MARKER_OBJECT_TAG_INFO_EXT = 1000022001,
-    VK_STRUCTURE_TYPE_DEBUG_MARKER_MARKER_INFO_EXT = 1000022002,
-    VK_STRUCTURE_TYPE_MAX_ENUM = 0x7FFFFFFF
-} VkStructureType;
-
-typedef enum VkSystemAllocationScope
-{
-    VK_SYSTEM_ALLOCATION_SCOPE_COMMAND = 0,
-    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT = 1,
-    VK_SYSTEM_ALLOCATION_SCOPE_CACHE = 2,
-    VK_SYSTEM_ALLOCATION_SCOPE_DEVICE = 3,
-    VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE = 4,
-    VK_SYSTEM_ALLOCATION_SCOPE_MAX_ENUM = 0x7FFFFFFF
-} VkSystemAllocationScope;
-
-typedef enum VkInternalAllocationType
-{
-    VK_INTERNAL_ALLOCATION_TYPE_EXECUTABLE = 0,
-    VK_INTERNAL_ALLOCATION_TYPE_MAX_ENUM = 0x7FFFFFFF
-} VkInternalAllocationType;
-
-typedef enum VkFormat
-{
-    VK_FORMAT_UNDEFINED = 0,
-    VK_FORMAT_R4G4_UNORM_PACK8 = 1,
-    VK_FORMAT_R4G4B4A4_UNORM_PACK16 = 2,
-    VK_FORMAT_B4G4R4A4_UNORM_PACK16 = 3,
-    VK_FORMAT_R5G6B5_UNORM_PACK16 = 4,
-    VK_FORMAT_B5G6R5_UNORM_PACK16 = 5,
-    VK_FORMAT_R5G5B5A1_UNORM_PACK16 = 6,
-    VK_FORMAT_B5G5R5A1_UNORM_PACK16 = 7,
-    VK_FORMAT_A1R5G5B5_UNORM_PACK16 = 8,
-    VK_FORMAT_R8_UNORM = 9,
-    VK_FORMAT_R8_SNORM = 10,
-    VK_FORMAT_R8_USCALED = 11,
-    VK_FORMAT_R8_SSCALED = 12,
-    VK_FORMAT_R8_UINT = 13,
-    VK_FORMAT_R8_SINT = 14,
-    VK_FORMAT_R8_SRGB = 15,
-    VK_FORMAT_R8G8_UNORM = 16,
-    VK_FORMAT_R8G8_SNORM = 17,
-    VK_FORMAT_R8G8_USCALED = 18,
-    VK_FORMAT_R8G8_SSCALED = 19,
-    VK_FORMAT_R8G8_UINT = 20,
-    VK_FORMAT_R8G8_SINT = 21,
-    VK_FORMAT_R8G8_SRGB = 22,
-    VK_FORMAT_R8G8B8_UNORM = 23,
-    VK_FORMAT_R8G8B8_SNORM = 24,
-    VK_FORMAT_R8G8B8_USCALED = 25,
-    VK_FORMAT_R8G8B8_SSCALED = 26,
-    VK_FORMAT_R8G8B8_UINT = 27,
-    VK_FORMAT_R8G8B8_SINT = 28,
-    VK_FORMAT_R8G8B8_SRGB = 29,
-    VK_FORMAT_B8G8R8_UNORM = 30,
-    VK_FORMAT_B8G8R8_SNORM = 31,
-    VK_FORMAT_B8G8R8_USCALED = 32,
-    VK_FORMAT_B8G8R8_SSCALED = 33,
-    VK_FORMAT_B8G8R8_UINT = 34,
-    VK_FORMAT_B8G8R8_SINT = 35,
-    VK_FORMAT_B8G8R8_SRGB = 36,
-    VK_FORMAT_R8G8B8A8_UNORM = 37,
-    VK_FORMAT_R8G8B8A8_SNORM = 38,
-    VK_FORMAT_R8G8B8A8_USCALED = 39,
-    VK_FORMAT_R8G8B8A8_SSCALED = 40,
-    VK_FORMAT_R8G8B8A8_UINT = 41,
-    VK_FORMAT_R8G8B8A8_SINT = 42,
-    VK_FORMAT_R8G8B8A8_SRGB = 43,
-    VK_FORMAT_B8G8R8A8_UNORM = 44,
-    VK_FORMAT_B8G8R8A8_SNORM = 45,
-    VK_FORMAT_B8G8R8A8_USCALED = 46,
-    VK_FORMAT_B8G8R8A8_SSCALED = 47,
-    VK_FORMAT_B8G8R8A8_UINT = 48,
-    VK_FORMAT_B8G8R8A8_SINT = 49,
-    VK_FORMAT_B8G8R8A8_SRGB = 50,
-    VK_FORMAT_A8B8G8R8_UNORM_PACK32 = 51,
-    VK_FORMAT_A8B8G8R8_SNORM_PACK32 = 52,
-    VK_FORMAT_A8B8G8R8_USCALED_PACK32 = 53,
-    VK_FORMAT_A8B8G8R8_SSCALED_PACK32 = 54,
-    VK_FORMAT_A8B8G8R8_UINT_PACK32 = 55,
-    VK_FORMAT_A8B8G8R8_SINT_PACK32 = 56,
-    VK_FORMAT_A8B8G8R8_SRGB_PACK32 = 57,
-    VK_FORMAT_A2R10G10B10_UNORM_PACK32 = 58,
-    VK_FORMAT_A2R10G10B10_SNORM_PACK32 = 59,
-    VK_FORMAT_A2R10G10B10_USCALED_PACK32 = 60,
-    VK_FORMAT_A2R10G10B10_SSCALED_PACK32 = 61,
-    VK_FORMAT_A2R10G10B10_UINT_PACK32 = 62,
-    VK_FORMAT_A2R10G10B10_SINT_PACK32 = 63,
-    VK_FORMAT_A2B10G10R10_UNORM_PACK32 = 64,
-    VK_FORMAT_A2B10G10R10_SNORM_PACK32 = 65,
-    VK_FORMAT_A2B10G10R10_USCALED_PACK32 = 66,
-    VK_FORMAT_A2B10G10R10_SSCALED_PACK32 = 67,
-    VK_FORMAT_A2B10G10R10_UINT_PACK32 = 68,
-    VK_FORMAT_A2B10G10R10_SINT_PACK32 = 69,
-    VK_FORMAT_R16_UNORM = 70,
-    VK_FORMAT_R16_SNORM = 71,
-    VK_FORMAT_R16_USCALED = 72,
-    VK_FORMAT_R16_SSCALED = 73,
-    VK_FORMAT_R16_UINT = 74,
-    VK_FORMAT_R16_SINT = 75,
-    VK_FORMAT_R16_SFLOAT = 76,
-    VK_FORMAT_R16G16_UNORM = 77,
-    VK_FORMAT_R16G16_SNORM = 78,
-    VK_FORMAT_R16G16_USCALED = 79,
-    VK_FORMAT_R16G16_SSCALED = 80,
-    VK_FORMAT_R16G16_UINT = 81,
-    VK_FORMAT_R16G16_SINT = 82,
-    VK_FORMAT_R16G16_SFLOAT = 83,
-    VK_FORMAT_R16G16B16_UNORM = 84,
-    VK_FORMAT_R16G16B16_SNORM = 85,
-    VK_FORMAT_R16G16B16_USCALED = 86,
-    VK_FORMAT_R16G16B16_SSCALED = 87,
-    VK_FORMAT_R16G16B16_UINT = 88,
-    VK_FORMAT_R16G16B16_SINT = 89,
-    VK_FORMAT_R16G16B16_SFLOAT = 90,
-    VK_FORMAT_R16G16B16A16_UNORM = 91,
-    VK_FORMAT_R16G16B16A16_SNORM = 92,
-    VK_FORMAT_R16G16B16A16_USCALED = 93,
-    VK_FORMAT_R16G16B16A16_SSCALED = 94,
-    VK_FORMAT_R16G16B16A16_UINT = 95,
-    VK_FORMAT_R16G16B16A16_SINT = 96,
-    VK_FORMAT_R16G16B16A16_SFLOAT = 97,
-    VK_FORMAT_R32_UINT = 98,
-    VK_FORMAT_R32_SINT = 99,
-    VK_FORMAT_R32_SFLOAT = 100,
-    VK_FORMAT_R32G32_UINT = 101,
-    VK_FORMAT_R32G32_SINT = 102,
-    VK_FORMAT_R32G32_SFLOAT = 103,
-    VK_FORMAT_R32G32B32_UINT = 104,
-    VK_FORMAT_R32G32B32_SINT = 105,
-    VK_FORMAT_R32G32B32_SFLOAT = 106,
-    VK_FORMAT_R32G32B32A32_UINT = 107,
-    VK_FORMAT_R32G32B32A32_SINT = 108,
-    VK_FORMAT_R32G32B32A32_SFLOAT = 109,
-    VK_FORMAT_R64_UINT = 110,
-    VK_FORMAT_R64_SINT = 111,
-    VK_FORMAT_R64_SFLOAT = 112,
-    VK_FORMAT_R64G64_UINT = 113,
-    VK_FORMAT_R64G64_SINT = 114,
-    VK_FORMAT_R64G64_SFLOAT = 115,
-    VK_FORMAT_R64G64B64_UINT = 116,
-    VK_FORMAT_R64G64B64_SINT = 117,
-    VK_FORMAT_R64G64B64_SFLOAT = 118,
-    VK_FORMAT_R64G64B64A64_UINT = 119,
-    VK_FORMAT_R64G64B64A64_SINT = 120,
-    VK_FORMAT_R64G64B64A64_SFLOAT = 121,
-    VK_FORMAT_B10G11R11_UFLOAT_PACK32 = 122,
-    VK_FORMAT_E5B9G9R9_UFLOAT_PACK32 = 123,
-    VK_FORMAT_D16_UNORM = 124,
-    VK_FORMAT_X8_D24_UNORM_PACK32 = 125,
-    VK_FORMAT_D32_SFLOAT = 126,
-    VK_FORMAT_S8_UINT = 127,
-    VK_FORMAT_D16_UNORM_S8_UINT = 128,
-    VK_FORMAT_D24_UNORM_S8_UINT = 129,
-    VK_FORMAT_D32_SFLOAT_S8_UINT = 130,
-    VK_FORMAT_BC1_RGB_UNORM_BLOCK = 131,
-    VK_FORMAT_BC1_RGB_SRGB_BLOCK = 132,
-    VK_FORMAT_BC1_RGBA_UNORM_BLOCK = 133,
-    VK_FORMAT_BC1_RGBA_SRGB_BLOCK = 134,
-    VK_FORMAT_BC2_UNORM_BLOCK = 135,
-    VK_FORMAT_BC2_SRGB_BLOCK = 136,
-    VK_FORMAT_BC3_UNORM_BLOCK = 137,
-    VK_FORMAT_BC3_SRGB_BLOCK = 138,
-    VK_FORMAT_BC4_UNORM_BLOCK = 139,
-    VK_FORMAT_BC4_SNORM_BLOCK = 140,
-    VK_FORMAT_BC5_UNORM_BLOCK = 141,
-    VK_FORMAT_BC5_SNORM_BLOCK = 142,
-    VK_FORMAT_BC6H_UFLOAT_BLOCK = 143,
-    VK_FORMAT_BC6H_SFLOAT_BLOCK = 144,
-    VK_FORMAT_BC7_UNORM_BLOCK = 145,
-    VK_FORMAT_BC7_SRGB_BLOCK = 146,
-    VK_FORMAT_ETC2_R8G8B8_UNORM_BLOCK = 147,
-    VK_FORMAT_ETC2_R8G8B8_SRGB_BLOCK = 148,
-    VK_FORMAT_ETC2_R8G8B8A1_UNORM_BLOCK = 149,
-    VK_FORMAT_ETC2_R8G8B8A1_SRGB_BLOCK = 150,
-    VK_FORMAT_ETC2_R8G8B8A8_UNORM_BLOCK = 151,
-    VK_FORMAT_ETC2_R8G8B8A8_SRGB_BLOCK = 152,
-    VK_FORMAT_EAC_R11_UNORM_BLOCK = 153,
-    VK_FORMAT_EAC_R11_SNORM_BLOCK = 154,
-    VK_FORMAT_EAC_R11G11_UNORM_BLOCK = 155,
-    VK_FORMAT_EAC_R11G11_SNORM_BLOCK = 156,
-    VK_FORMAT_ASTC_4x4_UNORM_BLOCK = 157,
-    VK_FORMAT_ASTC_4x4_SRGB_BLOCK = 158,
-    VK_FORMAT_ASTC_5x4_UNORM_BLOCK = 159,
-    VK_FORMAT_ASTC_5x4_SRGB_BLOCK = 160,
-    VK_FORMAT_ASTC_5x5_UNORM_BLOCK = 161,
-    VK_FORMAT_ASTC_5x5_SRGB_BLOCK = 162,
-    VK_FORMAT_ASTC_6x5_UNORM_BLOCK = 163,
-    VK_FORMAT_ASTC_6x5_SRGB_BLOCK = 164,
-    VK_FORMAT_ASTC_6x6_UNORM_BLOCK = 165,
-    VK_FORMAT_ASTC_6x6_SRGB_BLOCK = 166,
-    VK_FORMAT_ASTC_8x5_UNORM_BLOCK = 167,
-    VK_FORMAT_ASTC_8x5_SRGB_BLOCK = 168,
-    VK_FORMAT_ASTC_8x6_UNORM_BLOCK = 169,
-    VK_FORMAT_ASTC_8x6_SRGB_BLOCK = 170,
-    VK_FORMAT_ASTC_8x8_UNORM_BLOCK = 171,
-    VK_FORMAT_ASTC_8x8_SRGB_BLOCK = 172,
-    VK_FORMAT_ASTC_10x5_UNORM_BLOCK = 173,
-    VK_FORMAT_ASTC_10x5_SRGB_BLOCK = 174,
-    VK_FORMAT_ASTC_10x6_UNORM_BLOCK = 175,
-    VK_FORMAT_ASTC_10x6_SRGB_BLOCK = 176,
-    VK_FORMAT_ASTC_10x8_UNORM_BLOCK = 177,
-    VK_FORMAT_ASTC_10x8_SRGB_BLOCK = 178,
-    VK_FORMAT_ASTC_10x10_UNORM_BLOCK = 179,
-    VK_FORMAT_ASTC_10x10_SRGB_BLOCK = 180,
-    VK_FORMAT_ASTC_12x10_UNORM_BLOCK = 181,
-    VK_FORMAT_ASTC_12x10_SRGB_BLOCK = 182,
-    VK_FORMAT_ASTC_12x12_UNORM_BLOCK = 183,
-    VK_FORMAT_ASTC_12x12_SRGB_BLOCK = 184,
-    VK_FORMAT_PVRTC1_2BPP_UNORM_BLOCK_IMG = 1000054000,
-    VK_FORMAT_PVRTC1_4BPP_UNORM_BLOCK_IMG = 1000054001,
-    VK_FORMAT_PVRTC2_2BPP_UNORM_BLOCK_IMG = 1000054002,
-    VK_FORMAT_PVRTC2_4BPP_UNORM_BLOCK_IMG = 1000054003,
-    VK_FORMAT_PVRTC1_2BPP_SRGB_BLOCK_IMG = 1000054004,
-    VK_FORMAT_PVRTC1_4BPP_SRGB_BLOCK_IMG = 1000054005,
-    VK_FORMAT_PVRTC2_2BPP_SRGB_BLOCK_IMG = 1000054006,
-    VK_FORMAT_PVRTC2_4BPP_SRGB_BLOCK_IMG = 1000054007,
-    VK_FORMAT_G8B8G8R8_422_UNORM_KHR = 1000156000,
-    VK_FORMAT_B8G8R8G8_422_UNORM_KHR = 1000156001,
-    VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM_KHR = 1000156002,
-    VK_FORMAT_G8_B8R8_2PLANE_420_UNORM_KHR = 1000156003,
-    VK_FORMAT_G8_B8_R8_3PLANE_422_UNORM_KHR = 1000156004,
-    VK_FORMAT_G8_B8R8_2PLANE_422_UNORM_KHR = 1000156005,
-    VK_FORMAT_G8_B8_R8_3PLANE_444_UNORM_KHR = 1000156006,
-    VK_FORMAT_R10X6_UNORM_PACK16_KHR = 1000156007,
-    VK_FORMAT_R10X6G10X6_UNORM_2PACK16_KHR = 1000156008,
-    VK_FORMAT_R10X6G10X6B10X6A10X6_UNORM_4PACK16_KHR = 1000156009,
-    VK_FORMAT_G10X6B10X6G10X6R10X6_422_UNORM_4PACK16_KHR = 1000156010,
-    VK_FORMAT_B10X6G10X6R10X6G10X6_422_UNORM_4PACK16_KHR = 1000156011,
-    VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_420_UNORM_3PACK16_KHR = 1000156012,
-    VK_FORMAT_G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16_KHR = 1000156013,
-    VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_422_UNORM_3PACK16_KHR = 1000156014,
-    VK_FORMAT_G10X6_B10X6R10X6_2PLANE_422_UNORM_3PACK16_KHR = 1000156015,
-    VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_444_UNORM_3PACK16_KHR = 1000156016,
-    VK_FORMAT_R12X4_UNORM_PACK16_KHR = 1000156017,
-    VK_FORMAT_R12X4G12X4_UNORM_2PACK16_KHR = 1000156018,
-    VK_FORMAT_R12X4G12X4B12X4A12X4_UNORM_4PACK16_KHR = 1000156019,
-    VK_FORMAT_G12X4B12X4G12X4R12X4_422_UNORM_4PACK16_KHR = 1000156020,
-    VK_FORMAT_B12X4G12X4R12X4G12X4_422_UNORM_4PACK16_KHR = 1000156021,
-    VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_420_UNORM_3PACK16_KHR = 1000156022,
-    VK_FORMAT_G12X4_B12X4R12X4_2PLANE_420_UNORM_3PACK16_KHR = 1000156023,
-    VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_422_UNORM_3PACK16_KHR = 1000156024,
-    VK_FORMAT_G12X4_B12X4R12X4_2PLANE_422_UNORM_3PACK16_KHR = 1000156025,
-    VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_444_UNORM_3PACK16_KHR = 1000156026,
-    VK_FORMAT_G16B16G16R16_422_UNORM_KHR = 1000156027,
-    VK_FORMAT_B16G16R16G16_422_UNORM_KHR = 1000156028,
-    VK_FORMAT_G16_B16_R16_3PLANE_420_UNORM_KHR = 1000156029,
-    VK_FORMAT_G16_B16R16_2PLANE_420_UNORM_KHR = 1000156030,
-    VK_FORMAT_G16_B16_R16_3PLANE_422_UNORM_KHR = 1000156031,
-    VK_FORMAT_G16_B16R16_2PLANE_422_UNORM_KHR = 1000156032,
-    VK_FORMAT_G16_B16_R16_3PLANE_444_UNORM_KHR = 1000156033,
-    VK_FORMAT_MAX_ENUM = 0x7FFFFFFF
-} VkFormat;
-
-typedef enum VkImageType
-{
-    VK_IMAGE_TYPE_1D = 0,
-    VK_IMAGE_TYPE_2D = 1,
-    VK_IMAGE_TYPE_3D = 2,
-    VK_IMAGE_TYPE_MAX_ENUM = 0x7FFFFFFF
-} VkImageType;
-
-typedef enum VkImageTiling
-{
-    VK_IMAGE_TILING_OPTIMAL = 0,
-    VK_IMAGE_TILING_LINEAR = 1,
-    VK_IMAGE_TILING_MAX_ENUM = 0x7FFFFFFF
-} VkImageTiling;
-
-typedef enum VkPhysicalDeviceType
-{
-    VK_PHYSICAL_DEVICE_TYPE_OTHER = 0,
-    VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU = 1,
-    VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU = 2,
-    VK_PHYSICAL_DEVICE_TYPE_VIRTUAL_GPU = 3,
-    VK_PHYSICAL_DEVICE_TYPE_CPU = 4,
-    VK_PHYSICAL_DEVICE_TYPE_MAX_ENUM = 0x7FFFFFFF
-} VkPhysicalDeviceType;
-
-typedef enum VkQueryType
-{
-    VK_QUERY_TYPE_OCCLUSION = 0,
-    VK_QUERY_TYPE_PIPELINE_STATISTICS = 1,
-    VK_QUERY_TYPE_TIMESTAMP = 2,
-    VK_QUERY_TYPE_MAX_ENUM = 0x7FFFFFFF
-} VkQueryType;
-
-typedef enum VkSharingMode
-{
-    VK_SHARING_MODE_EXCLUSIVE = 0,
-    VK_SHARING_MODE_CONCURRENT = 1,
-    VK_SHARING_MODE_MAX_ENUM = 0x7FFFFFFF
-} VkSharingMode;
-
-typedef enum VkImageLayout
-{
-    VK_IMAGE_LAYOUT_UNDEFINED = 0,
-    VK_IMAGE_LAYOUT_GENERAL = 1,
-    VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL = 2,
-    VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL = 3,
-    VK_IMAGE_LAYOUT_DEPTH_STENCIL_READ_ONLY_OPTIMAL = 4,
-    VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL = 5,
-    VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL = 6,
-    VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL = 7,
-    VK_IMAGE_LAYOUT_PREINITIALIZED = 8,
-    VK_IMAGE_LAYOUT_PRESENT_SRC_KHR = 1000001002,
-    VK_IMAGE_LAYOUT_SHARED_PRESENT_KHR = 1000111000,
-    VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL_KHR = 1000117000,
-    VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_STENCIL_READ_ONLY_OPTIMAL_KHR = 1000117001,
-    VK_IMAGE_LAYOUT_MAX_ENUM = 0x7FFFFFFF
-} VkImageLayout;
-
-typedef enum VkImageViewType
-{
-    VK_IMAGE_VIEW_TYPE_1D = 0,
-    VK_IMAGE_VIEW_TYPE_2D = 1,
-    VK_IMAGE_VIEW_TYPE_3D = 2,
-    VK_IMAGE_VIEW_TYPE_CUBE = 3,
-    VK_IMAGE_VIEW_TYPE_1D_ARRAY = 4,
-    VK_IMAGE_VIEW_TYPE_2D_ARRAY = 5,
-    VK_IMAGE_VIEW_TYPE_CUBE_ARRAY = 6,
-    VK_IMAGE_VIEW_TYPE_MAX_ENUM = 0x7FFFFFFF
-} VkImageViewType;
-
-typedef enum VkComponentSwizzle
-{
-    VK_COMPONENT_SWIZZLE_IDENTITY = 0,
-    VK_COMPONENT_SWIZZLE_ZERO = 1,
-    VK_COMPONENT_SWIZZLE_ONE = 2,
-    VK_COMPONENT_SWIZZLE_R = 3,
-    VK_COMPONENT_SWIZZLE_G = 4,
-    VK_COMPONENT_SWIZZLE_B = 5,
-    VK_COMPONENT_SWIZZLE_A = 6,
-    VK_COMPONENT_SWIZZLE_MAX_ENUM = 0x7FFFFFFF
-} VkComponentSwizzle;
-
-typedef enum VkPrimitiveTopology
-{
-    VK_PRIMITIVE_TOPOLOGY_POINT_LIST = 0,
-    VK_PRIMITIVE_TOPOLOGY_LINE_LIST = 1,
-    VK_PRIMITIVE_TOPOLOGY_LINE_STRIP = 2,
-    VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST = 3,
-    VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP = 4,
-    VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN = 5,
-    VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY = 6,
-    VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY = 7,
-    VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY = 8,
-    VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY = 9,
-    VK_PRIMITIVE_TOPOLOGY_PATCH_LIST = 10,
-    VK_PRIMITIVE_TOPOLOGY_MAX_ENUM = 0x7FFFFFFF
-} VkPrimitiveTopology;
-
-typedef enum VkPolygonMode
-{
-    VK_POLYGON_MODE_FILL = 0,
-    VK_POLYGON_MODE_LINE = 1,
-    VK_POLYGON_MODE_POINT = 2,
-    VK_POLYGON_MODE_FILL_RECTANGLE_NV = 1000153000,
-    VK_POLYGON_MODE_MAX_ENUM = 0x7FFFFFFF
-} VkPolygonMode;
-
-typedef enum VkCompareOp
-{
-    VK_COMPARE_OP_NEVER = 0,
-    VK_COMPARE_OP_LESS = 1,
-    VK_COMPARE_OP_EQUAL = 2,
-    VK_COMPARE_OP_LESS_OR_EQUAL = 3,
-    VK_COMPARE_OP_GREATER = 4,
-    VK_COMPARE_OP_NOT_EQUAL = 5,
-    VK_COMPARE_OP_GREATER_OR_EQUAL = 6,
-    VK_COMPARE_OP_ALWAYS = 7,
-    VK_COMPARE_OP_MAX_ENUM = 0x7FFFFFFF
-} VkCompareOp;
-
-typedef enum VkLogicOp
-{
-    VK_LOGIC_OP_CLEAR = 0,
-    VK_LOGIC_OP_AND = 1,
-    VK_LOGIC_OP_AND_REVERSE = 2,
-    VK_LOGIC_OP_COPY = 3,
-    VK_LOGIC_OP_AND_INVERTED = 4,
-    VK_LOGIC_OP_NO_OP = 5,
-    VK_LOGIC_OP_XOR = 6,
-    VK_LOGIC_OP_OR = 7,
-    VK_LOGIC_OP_NOR = 8,
-    VK_LOGIC_OP_EQUIVALENT = 9,
-    VK_LOGIC_OP_INVERT = 10,
-    VK_LOGIC_OP_OR_REVERSE = 11,
-    VK_LOGIC_OP_COPY_INVERTED = 12,
-    VK_LOGIC_OP_OR_INVERTED = 13,
-    VK_LOGIC_OP_NAND = 14,
-    VK_LOGIC_OP_SET = 15,
-    VK_LOGIC_OP_MAX_ENUM = 0x7FFFFFFF
-} VkLogicOp;
-
-typedef enum VkFilter
-{
-    VK_FILTER_NEAREST = 0,
-    VK_FILTER_LINEAR = 1,
-    VK_FILTER_CUBIC_IMG = 1000015000,
-    VK_FILTER_MAX_ENUM = 0x7FFFFFFF
-} VkFilter;
-
-typedef enum VkSamplerMipmapMode
-{
-    VK_SAMPLER_MIPMAP_MODE_NEAREST = 0,
-    VK_SAMPLER_MIPMAP_MODE_LINEAR = 1,
-    VK_SAMPLER_MIPMAP_MODE_MAX_ENUM = 0x7FFFFFFF
-} VkSamplerMipmapMode;
-
-typedef enum VkSamplerAddressMode
-{
-    VK_SAMPLER_ADDRESS_MODE_REPEAT = 0,
-    VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT = 1,
-    VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE = 2,
-    VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER = 3,
-    VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE = 4,
-    VK_SAMPLER_ADDRESS_MODE_MAX_ENUM = 0x7FFFFFFF
-} VkSamplerAddressMode;
-
-typedef enum VkBorderColor
-{
-    VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK = 0,
-    VK_BORDER_COLOR_INT_TRANSPARENT_BLACK = 1,
-    VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK = 2,
-    VK_BORDER_COLOR_INT_OPAQUE_BLACK = 3,
-    VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE = 4,
-    VK_BORDER_COLOR_INT_OPAQUE_WHITE = 5,
-    VK_BORDER_COLOR_MAX_ENUM = 0x7FFFFFFF
-} VkBorderColor;
-
-typedef enum VkDescriptorType
-{
-    VK_DESCRIPTOR_TYPE_SAMPLER = 0,
-    VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER = 1,
-    VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE = 2,
-    VK_DESCRIPTOR_TYPE_STORAGE_IMAGE = 3,
-    VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER = 4,
-    VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER = 5,
-    VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER = 6,
-    VK_DESCRIPTOR_TYPE_STORAGE_BUFFER = 7,
-    VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC = 8,
-    VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC = 9,
-    VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT = 10,
-    VK_DESCRIPTOR_TYPE_MAX_ENUM = 0x7FFFFFFF
-} VkDescriptorType;
-
-typedef enum VkPipelineBindPoint
-{
-    VK_PIPELINE_BIND_POINT_GRAPHICS = 0,
-    VK_PIPELINE_BIND_POINT_COMPUTE = 1,
-    VK_PIPELINE_BIND_POINT_MAX_ENUM = 0x7FFFFFFF
-} VkPipelineBindPoint;
-
-typedef enum VkCommandBufferLevel
-{
-    VK_COMMAND_BUFFER_LEVEL_PRIMARY = 0,
-    VK_COMMAND_BUFFER_LEVEL_SECONDARY = 1,
-    VK_COMMAND_BUFFER_LEVEL_MAX_ENUM = 0x7FFFFFFF
-} VkCommandBufferLevel;
-
-typedef enum VkIndexType
-{
-    VK_INDEX_TYPE_UINT16 = 0,
-    VK_INDEX_TYPE_UINT32 = 1,
-    VK_INDEX_TYPE_MAX_ENUM = 0x7FFFFFFF
-} VkIndexType;
-
-typedef enum VkObjectType
-{
-    VK_OBJECT_TYPE_UNKNOWN = 0,
-    VK_OBJECT_TYPE_INSTANCE = 1,
-    VK_OBJECT_TYPE_PHYSICAL_DEVICE = 2,
-    VK_OBJECT_TYPE_DEVICE = 3,
-    VK_OBJECT_TYPE_QUEUE = 4,
-    VK_OBJECT_TYPE_SEMAPHORE = 5,
-    VK_OBJECT_TYPE_COMMAND_BUFFER = 6,
-    VK_OBJECT_TYPE_FENCE = 7,
-    VK_OBJECT_TYPE_DEVICE_MEMORY = 8,
-    VK_OBJECT_TYPE_BUFFER = 9,
-    VK_OBJECT_TYPE_IMAGE = 10,
-    VK_OBJECT_TYPE_EVENT = 11,
-    VK_OBJECT_TYPE_QUERY_POOL = 12,
-    VK_OBJECT_TYPE_BUFFER_VIEW = 13,
-    VK_OBJECT_TYPE_IMAGE_VIEW = 14,
-    VK_OBJECT_TYPE_SHADER_MODULE = 15,
-    VK_OBJECT_TYPE_PIPELINE_CACHE = 16,
-    VK_OBJECT_TYPE_PIPELINE_LAYOUT = 17,
-    VK_OBJECT_TYPE_RENDER_PASS = 18,
-    VK_OBJECT_TYPE_PIPELINE = 19,
-    VK_OBJECT_TYPE_DESCRIPTOR_SET_LAYOUT = 20,
-    VK_OBJECT_TYPE_SAMPLER = 21,
-    VK_OBJECT_TYPE_DESCRIPTOR_POOL = 22,
-    VK_OBJECT_TYPE_DESCRIPTOR_SET = 23,
-    VK_OBJECT_TYPE_FRAMEBUFFER = 24,
-    VK_OBJECT_TYPE_COMMAND_POOL = 25,
-    VK_OBJECT_TYPE_SURFACE_KHR = 1000000000,
-    VK_OBJECT_TYPE_SWAPCHAIN_KHR = 1000001000,
-    VK_OBJECT_TYPE_DISPLAY_KHR = 1000002000,
-    VK_OBJECT_TYPE_DISPLAY_MODE_KHR = 1000002001,
-    VK_OBJECT_TYPE_DEBUG_REPORT_CALLBACK_EXT = 1000011000,
-    VK_OBJECT_TYPE_DESCRIPTOR_UPDATE_TEMPLATE_KHR = 1000085000,
-    VK_OBJECT_TYPE_OBJECT_TABLE_NVX = 1000086000,
-    VK_OBJECT_TYPE_INDIRECT_COMMANDS_LAYOUT_NVX = 1000086001,
-    VK_OBJECT_TYPE_SAMPLER_YCBCR_CONVERSION_KHR = 1000156000,
-    VK_OBJECT_TYPE_VALIDATION_CACHE_EXT = 1000160000,
-    VK_OBJECT_TYPE_MAX_ENUM = 0x7FFFFFFF
-} VkObjectType;
-
-typedef VkFlags VkInstanceCreateFlags;
-
-typedef enum VkFormatFeatureFlagBits
-{
-    VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT = 0x00000001,
-    VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT = 0x00000002,
-    VK_FORMAT_FEATURE_STORAGE_IMAGE_ATOMIC_BIT = 0x00000004,
-    VK_FORMAT_FEATURE_UNIFORM_TEXEL_BUFFER_BIT = 0x00000008,
-    VK_FORMAT_FEATURE_STORAGE_TEXEL_BUFFER_BIT = 0x00000010,
-    VK_FORMAT_FEATURE_STORAGE_TEXEL_BUFFER_ATOMIC_BIT = 0x00000020,
-    VK_FORMAT_FEATURE_VERTEX_BUFFER_BIT = 0x00000040,
-    VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT = 0x00000080,
-    VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BLEND_BIT = 0x00000100,
-    VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT = 0x00000200,
-    VK_FORMAT_FEATURE_BLIT_SRC_BIT = 0x00000400,
-    VK_FORMAT_FEATURE_BLIT_DST_BIT = 0x00000800,
-    VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_LINEAR_BIT = 0x00001000,
-    VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_CUBIC_BIT_IMG = 0x00002000,
-    VK_FORMAT_FEATURE_TRANSFER_SRC_BIT_KHR = 0x00004000,
-    VK_FORMAT_FEATURE_TRANSFER_DST_BIT_KHR = 0x00008000,
-    VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_MINMAX_BIT_EXT = 0x00010000,
-    VK_FORMAT_FEATURE_MIDPOINT_CHROMA_SAMPLES_BIT_KHR = 0x00020000,
-    VK_FORMAT_FEATURE_SAMPLED_IMAGE_YCBCR_CONVERSION_LINEAR_FILTER_BIT_KHR = 0x00040000,
-    VK_FORMAT_FEATURE_SAMPLED_IMAGE_YCBCR_CONVERSION_SEPARATE_RECONSTRUCTION_FILTER_BIT_KHR = 0x00080000,
-    VK_FORMAT_FEATURE_SAMPLED_IMAGE_YCBCR_CONVERSION_CHROMA_RECONSTRUCTION_EXPLICIT_BIT_KHR = 0x00100000,
-    VK_FORMAT_FEATURE_SAMPLED_IMAGE_YCBCR_CONVERSION_CHROMA_RECONSTRUCTION_EXPLICIT_FORCEABLE_BIT_KHR = 0x00200000,
-    VK_FORMAT_FEATURE_DISJOINT_BIT_KHR = 0x00400000,
-    VK_FORMAT_FEATURE_COSITED_CHROMA_SAMPLES_BIT_KHR = 0x00800000,
-    VK_FORMAT_FEATURE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkFormatFeatureFlagBits;
-typedef VkFlags VkFormatFeatureFlags;
-
-typedef enum VkImageUsageFlagBits
-{
-    VK_IMAGE_USAGE_TRANSFER_SRC_BIT = 0x00000001,
-    VK_IMAGE_USAGE_TRANSFER_DST_BIT = 0x00000002,
-    VK_IMAGE_USAGE_SAMPLED_BIT = 0x00000004,
-    VK_IMAGE_USAGE_STORAGE_BIT = 0x00000008,
-    VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT = 0x00000010,
-    VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT = 0x00000020,
-    VK_IMAGE_USAGE_TRANSIENT_ATTACHMENT_BIT = 0x00000040,
-    VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT = 0x00000080,
-    VK_IMAGE_USAGE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkImageUsageFlagBits;
-typedef VkFlags VkImageUsageFlags;
-
-typedef enum VkImageCreateFlagBits
-{
-    VK_IMAGE_CREATE_SPARSE_BINDING_BIT = 0x00000001,
-    VK_IMAGE_CREATE_SPARSE_RESIDENCY_BIT = 0x00000002,
-    VK_IMAGE_CREATE_SPARSE_ALIASED_BIT = 0x00000004,
-    VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT = 0x00000008,
-    VK_IMAGE_CREATE_CUBE_COMPATIBLE_BIT = 0x00000010,
-    VK_IMAGE_CREATE_BIND_SFR_BIT_KHX = 0x00000040,
-    VK_IMAGE_CREATE_2D_ARRAY_COMPATIBLE_BIT_KHR = 0x00000020,
-    VK_IMAGE_CREATE_BLOCK_TEXEL_VIEW_COMPATIBLE_BIT_KHR = 0x00000080,
-    VK_IMAGE_CREATE_EXTENDED_USAGE_BIT_KHR = 0x00000100,
-    VK_IMAGE_CREATE_SAMPLE_LOCATIONS_COMPATIBLE_DEPTH_BIT_EXT = 0x00001000,
-    VK_IMAGE_CREATE_DISJOINT_BIT_KHR = 0x00000200,
-    VK_IMAGE_CREATE_ALIAS_BIT_KHR = 0x00000400,
-    VK_IMAGE_CREATE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkImageCreateFlagBits;
-typedef VkFlags VkImageCreateFlags;
-
-typedef enum VkSampleCountFlagBits
-{
-    VK_SAMPLE_COUNT_1_BIT = 0x00000001,
-    VK_SAMPLE_COUNT_2_BIT = 0x00000002,
-    VK_SAMPLE_COUNT_4_BIT = 0x00000004,
-    VK_SAMPLE_COUNT_8_BIT = 0x00000008,
-    VK_SAMPLE_COUNT_16_BIT = 0x00000010,
-    VK_SAMPLE_COUNT_32_BIT = 0x00000020,
-    VK_SAMPLE_COUNT_64_BIT = 0x00000040,
-    VK_SAMPLE_COUNT_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkSampleCountFlagBits;
-typedef VkFlags VkSampleCountFlags;
-
-typedef enum VkQueueFlagBits
-{
-    VK_QUEUE_GRAPHICS_BIT = 0x00000001,
-    VK_QUEUE_COMPUTE_BIT = 0x00000002,
-    VK_QUEUE_TRANSFER_BIT = 0x00000004,
-    VK_QUEUE_SPARSE_BINDING_BIT = 0x00000008,
-    VK_QUEUE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkQueueFlagBits;
-typedef VkFlags VkQueueFlags;
-
-typedef enum VkMemoryPropertyFlagBits
-{
-    VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT = 0x00000001,
-    VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT = 0x00000002,
-    VK_MEMORY_PROPERTY_HOST_COHERENT_BIT = 0x00000004,
-    VK_MEMORY_PROPERTY_HOST_CACHED_BIT = 0x00000008,
-    VK_MEMORY_PROPERTY_LAZILY_ALLOCATED_BIT = 0x00000010,
-    VK_MEMORY_PROPERTY_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkMemoryPropertyFlagBits;
-typedef VkFlags VkMemoryPropertyFlags;
-
-typedef enum VkMemoryHeapFlagBits
-{
-    VK_MEMORY_HEAP_DEVICE_LOCAL_BIT = 0x00000001,
-    VK_MEMORY_HEAP_MULTI_INSTANCE_BIT_KHX = 0x00000002,
-    VK_MEMORY_HEAP_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkMemoryHeapFlagBits;
-typedef VkFlags VkMemoryHeapFlags;
-typedef VkFlags VkDeviceCreateFlags;
-typedef VkFlags VkDeviceQueueCreateFlags;
-
-typedef enum VkPipelineStageFlagBits
-{
-    VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT = 0x00000001,
-    VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT = 0x00000002,
-    VK_PIPELINE_STAGE_VERTEX_INPUT_BIT = 0x00000004,
-    VK_PIPELINE_STAGE_VERTEX_SHADER_BIT = 0x00000008,
-    VK_PIPELINE_STAGE_TESSELLATION_CONTROL_SHADER_BIT = 0x00000010,
-    VK_PIPELINE_STAGE_TESSELLATION_EVALUATION_SHADER_BIT = 0x00000020,
-    VK_PIPELINE_STAGE_GEOMETRY_SHADER_BIT = 0x00000040,
-    VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT = 0x00000080,
-    VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT = 0x00000100,
-    VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT = 0x00000200,
-    VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT = 0x00000400,
-    VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT = 0x00000800,
-    VK_PIPELINE_STAGE_TRANSFER_BIT = 0x00001000,
-    VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT = 0x00002000,
-    VK_PIPELINE_STAGE_HOST_BIT = 0x00004000,
-    VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT = 0x00008000,
-    VK_PIPELINE_STAGE_ALL_COMMANDS_BIT = 0x00010000,
-    VK_PIPELINE_STAGE_COMMAND_PROCESS_BIT_NVX = 0x00020000,
-    VK_PIPELINE_STAGE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkPipelineStageFlagBits;
-typedef VkFlags VkPipelineStageFlags;
-typedef VkFlags VkMemoryMapFlags;
-
-typedef enum VkImageAspectFlagBits
-{
-    VK_IMAGE_ASPECT_COLOR_BIT = 0x00000001,
-    VK_IMAGE_ASPECT_DEPTH_BIT = 0x00000002,
-    VK_IMAGE_ASPECT_STENCIL_BIT = 0x00000004,
-    VK_IMAGE_ASPECT_METADATA_BIT = 0x00000008,
-    VK_IMAGE_ASPECT_PLANE_0_BIT_KHR = 0x00000010,
-    VK_IMAGE_ASPECT_PLANE_1_BIT_KHR = 0x00000020,
-    VK_IMAGE_ASPECT_PLANE_2_BIT_KHR = 0x00000040,
-    VK_IMAGE_ASPECT_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkImageAspectFlagBits;
-typedef VkFlags VkImageAspectFlags;
-
-typedef enum VkFenceCreateFlagBits
-{
-    VK_FENCE_CREATE_SIGNALED_BIT = 0x00000001,
-    VK_FENCE_CREATE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkFenceCreateFlagBits;
-typedef VkFlags VkFenceCreateFlags;
-typedef VkFlags VkSemaphoreCreateFlags;
-typedef VkFlags VkEventCreateFlags;
-typedef VkFlags VkQueryPoolCreateFlags;
-
-typedef enum VkQueryPipelineStatisticFlagBits
-{
-    VK_QUERY_PIPELINE_STATISTIC_INPUT_ASSEMBLY_VERTICES_BIT = 0x00000001,
-    VK_QUERY_PIPELINE_STATISTIC_INPUT_ASSEMBLY_PRIMITIVES_BIT = 0x00000002,
-    VK_QUERY_PIPELINE_STATISTIC_VERTEX_SHADER_INVOCATIONS_BIT = 0x00000004,
-    VK_QUERY_PIPELINE_STATISTIC_GEOMETRY_SHADER_INVOCATIONS_BIT = 0x00000008,
-    VK_QUERY_PIPELINE_STATISTIC_GEOMETRY_SHADER_PRIMITIVES_BIT = 0x00000010,
-    VK_QUERY_PIPELINE_STATISTIC_CLIPPING_INVOCATIONS_BIT = 0x00000020,
-    VK_QUERY_PIPELINE_STATISTIC_CLIPPING_PRIMITIVES_BIT = 0x00000040,
-    VK_QUERY_PIPELINE_STATISTIC_FRAGMENT_SHADER_INVOCATIONS_BIT = 0x00000080,
-    VK_QUERY_PIPELINE_STATISTIC_TESSELLATION_CONTROL_SHADER_PATCHES_BIT = 0x00000100,
-    VK_QUERY_PIPELINE_STATISTIC_TESSELLATION_EVALUATION_SHADER_INVOCATIONS_BIT = 0x00000200,
-    VK_QUERY_PIPELINE_STATISTIC_COMPUTE_SHADER_INVOCATIONS_BIT = 0x00000400,
-    VK_QUERY_PIPELINE_STATISTIC_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkQueryPipelineStatisticFlagBits;
-typedef VkFlags VkQueryPipelineStatisticFlags;
-
-typedef enum VkQueryResultFlagBits
-{
-    VK_QUERY_RESULT_64_BIT = 0x00000001,
-    VK_QUERY_RESULT_WAIT_BIT = 0x00000002,
-    VK_QUERY_RESULT_WITH_AVAILABILITY_BIT = 0x00000004,
-    VK_QUERY_RESULT_PARTIAL_BIT = 0x00000008,
-    VK_QUERY_RESULT_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkQueryResultFlagBits;
-typedef VkFlags VkQueryResultFlags;
-
-typedef enum VkBufferCreateFlagBits
-{
-    VK_BUFFER_CREATE_SPARSE_BINDING_BIT = 0x00000001,
-    VK_BUFFER_CREATE_SPARSE_RESIDENCY_BIT = 0x00000002,
-    VK_BUFFER_CREATE_SPARSE_ALIASED_BIT = 0x00000004,
-    VK_BUFFER_CREATE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkBufferCreateFlagBits;
-typedef VkFlags VkBufferCreateFlags;
-
-typedef enum VkBufferUsageFlagBits
-{
-    VK_BUFFER_USAGE_TRANSFER_SRC_BIT = 0x00000001,
-    VK_BUFFER_USAGE_TRANSFER_DST_BIT = 0x00000002,
-    VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT = 0x00000004,
-    VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT = 0x00000008,
-    VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT = 0x00000010,
-    VK_BUFFER_USAGE_STORAGE_BUFFER_BIT = 0x00000020,
-    VK_BUFFER_USAGE_INDEX_BUFFER_BIT = 0x00000040,
-    VK_BUFFER_USAGE_VERTEX_BUFFER_BIT = 0x00000080,
-    VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT = 0x00000100,
-    VK_BUFFER_USAGE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkBufferUsageFlagBits;
-typedef VkFlags VkBufferUsageFlags;
-typedef VkFlags VkBufferViewCreateFlags;
-typedef VkFlags VkImageViewCreateFlags;
-typedef VkFlags VkShaderModuleCreateFlags;
-typedef VkFlags VkPipelineCacheCreateFlags;
-
-typedef enum VkPipelineCreateFlagBits
-{
-    VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT = 0x00000001,
-    VK_PIPELINE_CREATE_ALLOW_DERIVATIVES_BIT = 0x00000002,
-    VK_PIPELINE_CREATE_DERIVATIVE_BIT = 0x00000004,
-    VK_PIPELINE_CREATE_VIEW_INDEX_FROM_DEVICE_INDEX_BIT_KHX = 0x00000008,
-    VK_PIPELINE_CREATE_DISPATCH_BASE_KHX = 0x00000010,
-    VK_PIPELINE_CREATE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkPipelineCreateFlagBits;
-typedef VkFlags VkPipelineCreateFlags;
-typedef VkFlags VkPipelineShaderStageCreateFlags;
-
-typedef enum VkShaderStageFlagBits
-{
-    VK_SHADER_STAGE_VERTEX_BIT = 0x00000001,
-    VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT = 0x00000002,
-    VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT = 0x00000004,
-    VK_SHADER_STAGE_GEOMETRY_BIT = 0x00000008,
-    VK_SHADER_STAGE_FRAGMENT_BIT = 0x00000010,
-    VK_SHADER_STAGE_COMPUTE_BIT = 0x00000020,
-    VK_SHADER_STAGE_ALL_GRAPHICS = 0x0000001F,
-    VK_SHADER_STAGE_ALL = 0x7FFFFFFF,
-    VK_SHADER_STAGE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkShaderStageFlagBits;
-typedef VkFlags VkPipelineVertexInputStateCreateFlags;
-typedef VkFlags VkPipelineInputAssemblyStateCreateFlags;
-typedef VkFlags VkPipelineTessellationStateCreateFlags;
-typedef VkFlags VkPipelineViewportStateCreateFlags;
-typedef VkFlags VkPipelineRasterizationStateCreateFlags;
-
-typedef VkFlags VkPipelineMultisampleStateCreateFlags;
-typedef VkFlags VkPipelineDepthStencilStateCreateFlags;
-typedef VkFlags VkPipelineColorBlendStateCreateFlags;
-
-typedef VkFlags VkPipelineDynamicStateCreateFlags;
-typedef VkFlags VkPipelineLayoutCreateFlags;
-typedef VkFlags VkShaderStageFlags;
-typedef VkFlags VkSamplerCreateFlags;
-
-typedef enum VkDescriptorSetLayoutCreateFlagBits
-{
-    VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR = 0x00000001,
-    VK_DESCRIPTOR_SET_LAYOUT_CREATE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkDescriptorSetLayoutCreateFlagBits;
-typedef VkFlags VkDescriptorSetLayoutCreateFlags;
-
-typedef enum VkDescriptorPoolCreateFlagBits
-{
-    VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT = 0x00000001,
-    VK_DESCRIPTOR_POOL_CREATE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkDescriptorPoolCreateFlagBits;
-typedef VkFlags VkDescriptorPoolCreateFlags;
-typedef VkFlags VkDescriptorPoolResetFlags;
-typedef VkFlags VkFramebufferCreateFlags;
-typedef VkFlags VkRenderPassCreateFlags;
-
-typedef enum VkAccessFlagBits
-{
-    VK_ACCESS_INDIRECT_COMMAND_READ_BIT = 0x00000001,
-    VK_ACCESS_INDEX_READ_BIT = 0x00000002,
-    VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT = 0x00000004,
-    VK_ACCESS_UNIFORM_READ_BIT = 0x00000008,
-    VK_ACCESS_INPUT_ATTACHMENT_READ_BIT = 0x00000010,
-    VK_ACCESS_SHADER_READ_BIT = 0x00000020,
-    VK_ACCESS_SHADER_WRITE_BIT = 0x00000040,
-    VK_ACCESS_COLOR_ATTACHMENT_READ_BIT = 0x00000080,
-    VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT = 0x00000100,
-    VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT = 0x00000200,
-    VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT = 0x00000400,
-    VK_ACCESS_TRANSFER_READ_BIT = 0x00000800,
-    VK_ACCESS_TRANSFER_WRITE_BIT = 0x00001000,
-    VK_ACCESS_HOST_READ_BIT = 0x00002000,
-    VK_ACCESS_HOST_WRITE_BIT = 0x00004000,
-    VK_ACCESS_MEMORY_READ_BIT = 0x00008000,
-    VK_ACCESS_MEMORY_WRITE_BIT = 0x00010000,
-    VK_ACCESS_COMMAND_PROCESS_READ_BIT_NVX = 0x00020000,
-    VK_ACCESS_COMMAND_PROCESS_WRITE_BIT_NVX = 0x00040000,
-    VK_ACCESS_COLOR_ATTACHMENT_READ_NONCOHERENT_BIT_EXT = 0x00080000,
-    VK_ACCESS_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkAccessFlagBits;
-typedef VkFlags VkAccessFlags;
-
-typedef enum VkDependencyFlagBits
-{
-    VK_DEPENDENCY_BY_REGION_BIT = 0x00000001,
-    VK_DEPENDENCY_VIEW_LOCAL_BIT_KHX = 0x00000002,
-    VK_DEPENDENCY_DEVICE_GROUP_BIT_KHX = 0x00000004,
-    VK_DEPENDENCY_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkDependencyFlagBits;
-typedef VkFlags VkDependencyFlags;
-
-typedef enum VkCommandPoolCreateFlagBits
-{
-    VK_COMMAND_POOL_CREATE_TRANSIENT_BIT = 0x00000001,
-    VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT = 0x00000002,
-    VK_COMMAND_POOL_CREATE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkCommandPoolCreateFlagBits;
-typedef VkFlags VkCommandPoolCreateFlags;
-
-typedef enum VkCommandPoolResetFlagBits
-{
-    VK_COMMAND_POOL_RESET_RELEASE_RESOURCES_BIT = 0x00000001,
-    VK_COMMAND_POOL_RESET_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkCommandPoolResetFlagBits;
-typedef VkFlags VkCommandPoolResetFlags;
-
-typedef enum VkCommandBufferUsageFlagBits
-{
-    VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT = 0x00000001,
-    VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT = 0x00000002,
-    VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT = 0x00000004,
-    VK_COMMAND_BUFFER_USAGE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkCommandBufferUsageFlagBits;
-typedef VkFlags VkCommandBufferUsageFlags;
-
-typedef enum VkQueryControlFlagBits
-{
-    VK_QUERY_CONTROL_PRECISE_BIT = 0x00000001,
-    VK_QUERY_CONTROL_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkQueryControlFlagBits;
-typedef VkFlags VkQueryControlFlags;
-
-typedef enum VkCommandBufferResetFlagBits
-{
-    VK_COMMAND_BUFFER_RESET_RELEASE_RESOURCES_BIT = 0x00000001,
-    VK_COMMAND_BUFFER_RESET_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkCommandBufferResetFlagBits;
-typedef VkFlags VkCommandBufferResetFlags;
-
-typedef struct VkApplicationInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    const char* pApplicationName;
-    uint32_t applicationVersion;
-    const char* pEngineName;
-    uint32_t engineVersion;
-    uint32_t apiVersion;
-} VkApplicationInfo;
-
-typedef struct VkInstanceCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkInstanceCreateFlags flags;
-    const VkApplicationInfo* pApplicationInfo;
-    uint32_t enabledLayerCount;
-    const char* const* ppEnabledLayerNames;
-    uint32_t enabledExtensionCount;
-    const char* const* ppEnabledExtensionNames;
-} VkInstanceCreateInfo;
-
-typedef void*(VKAPI_PTR* PFN_vkAllocationFunction)(
-    void* pUserData,
-    size_t size,
-    size_t alignment,
-    VkSystemAllocationScope allocationScope);
-
-typedef void*(VKAPI_PTR* PFN_vkReallocationFunction)(
-    void* pUserData,
-    void* pOriginal,
-    size_t size,
-    size_t alignment,
-    VkSystemAllocationScope allocationScope);
-
-typedef void(VKAPI_PTR* PFN_vkFreeFunction)(
-    void* pUserData,
-    void* pMemory);
-
-typedef void(VKAPI_PTR* PFN_vkInternalAllocationNotification)(
-    void* pUserData,
-    size_t size,
-    VkInternalAllocationType allocationType,
-    VkSystemAllocationScope allocationScope);
-
-typedef void(VKAPI_PTR* PFN_vkInternalFreeNotification)(
-    void* pUserData,
-    size_t size,
-    VkInternalAllocationType allocationType,
-    VkSystemAllocationScope allocationScope);
-
-typedef struct VkAllocationCallbacks
-{
-    void* pUserData;
-    PFN_vkAllocationFunction pfnAllocation;
-    PFN_vkReallocationFunction pfnReallocation;
-    PFN_vkFreeFunction pfnFree;
-    PFN_vkInternalAllocationNotification pfnInternalAllocation;
-    PFN_vkInternalFreeNotification pfnInternalFree;
-} VkAllocationCallbacks;
-
-typedef struct VkPhysicalDeviceFeatures
-{
-    VkBool32 robustBufferAccess;
-    VkBool32 fullDrawIndexUint32;
-    VkBool32 imageCubeArray;
-    VkBool32 independentBlend;
-    VkBool32 geometryShader;
-    VkBool32 tessellationShader;
-    VkBool32 sampleRateShading;
-    VkBool32 dualSrcBlend;
-    VkBool32 logicOp;
-    VkBool32 multiDrawIndirect;
-    VkBool32 drawIndirectFirstInstance;
-    VkBool32 depthClamp;
-    VkBool32 depthBiasClamp;
-    VkBool32 fillModeNonSolid;
-    VkBool32 depthBounds;
-    VkBool32 wideLines;
-    VkBool32 largePoints;
-    VkBool32 alphaToOne;
-    VkBool32 multiViewport;
-    VkBool32 samplerAnisotropy;
-    VkBool32 textureCompressionETC2;
-    VkBool32 textureCompressionASTC_LDR;
-    VkBool32 textureCompressionBC;
-    VkBool32 occlusionQueryPrecise;
-    VkBool32 pipelineStatisticsQuery;
-    VkBool32 vertexPipelineStoresAndAtomics;
-    VkBool32 fragmentStoresAndAtomics;
-    VkBool32 shaderTessellationAndGeometryPointSize;
-    VkBool32 shaderImageGatherExtended;
-    VkBool32 shaderStorageImageExtendedFormats;
-    VkBool32 shaderStorageImageMultisample;
-    VkBool32 shaderStorageImageReadWithoutFormat;
-    VkBool32 shaderStorageImageWriteWithoutFormat;
-    VkBool32 shaderUniformBufferArrayDynamicIndexing;
-    VkBool32 shaderSampledImageArrayDynamicIndexing;
-    VkBool32 shaderStorageBufferArrayDynamicIndexing;
-    VkBool32 shaderStorageImageArrayDynamicIndexing;
-    VkBool32 shaderClipDistance;
-    VkBool32 shaderCullDistance;
-    VkBool32 shaderFloat64;
-    VkBool32 shaderInt64;
-    VkBool32 shaderInt16;
-    VkBool32 shaderResourceResidency;
-    VkBool32 shaderResourceMinLod;
-    VkBool32 sparseBinding;
-    VkBool32 sparseResidencyBuffer;
-    VkBool32 sparseResidencyImage2D;
-    VkBool32 sparseResidencyImage3D;
-    VkBool32 sparseResidency2Samples;
-    VkBool32 sparseResidency4Samples;
-    VkBool32 sparseResidency8Samples;
-    VkBool32 sparseResidency16Samples;
-    VkBool32 sparseResidencyAliased;
-    VkBool32 variableMultisampleRate;
-    VkBool32 inheritedQueries;
-} VkPhysicalDeviceFeatures;
-
-typedef struct VkFormatProperties
-{
-    VkFormatFeatureFlags linearTilingFeatures;
-    VkFormatFeatureFlags optimalTilingFeatures;
-    VkFormatFeatureFlags bufferFeatures;
-} VkFormatProperties;
-
-typedef struct VkExtent3D
-{
-    uint32_t width;
-    uint32_t height;
-    uint32_t depth;
-} VkExtent3D;
-
-typedef struct VkImageFormatProperties
-{
-    VkExtent3D maxExtent;
-    uint32_t maxMipLevels;
-    uint32_t maxArrayLayers;
-    VkSampleCountFlags sampleCounts;
-    VkDeviceSize maxResourceSize;
-} VkImageFormatProperties;
-
-typedef struct VkPhysicalDeviceLimits
-{
-    uint32_t maxImageDimension1D;
-    uint32_t maxImageDimension2D;
-    uint32_t maxImageDimension3D;
-    uint32_t maxImageDimensionCube;
-    uint32_t maxImageArrayLayers;
-    uint32_t maxTexelBufferElements;
-    uint32_t maxUniformBufferRange;
-    uint32_t maxStorageBufferRange;
-    uint32_t maxPushConstantsSize;
-    uint32_t maxMemoryAllocationCount;
-    uint32_t maxSamplerAllocationCount;
-    VkDeviceSize bufferImageGranularity;
-    VkDeviceSize sparseAddressSpaceSize;
-    uint32_t maxBoundDescriptorSets;
-    uint32_t maxPerStageDescriptorSamplers;
-    uint32_t maxPerStageDescriptorUniformBuffers;
-    uint32_t maxPerStageDescriptorStorageBuffers;
-    uint32_t maxPerStageDescriptorSampledImages;
-    uint32_t maxPerStageDescriptorStorageImages;
-    uint32_t maxPerStageDescriptorInputAttachments;
-    uint32_t maxPerStageResources;
-    uint32_t maxDescriptorSetSamplers;
-    uint32_t maxDescriptorSetUniformBuffers;
-    uint32_t maxDescriptorSetUniformBuffersDynamic;
-    uint32_t maxDescriptorSetStorageBuffers;
-    uint32_t maxDescriptorSetStorageBuffersDynamic;
-    uint32_t maxDescriptorSetSampledImages;
-    uint32_t maxDescriptorSetStorageImages;
-    uint32_t maxDescriptorSetInputAttachments;
-    uint32_t maxVertexInputAttributes;
-    uint32_t maxVertexInputBindings;
-    uint32_t maxVertexInputAttributeOffset;
-    uint32_t maxVertexInputBindingStride;
-    uint32_t maxVertexOutputComponents;
-    uint32_t maxTessellationGenerationLevel;
-    uint32_t maxTessellationPatchSize;
-    uint32_t maxTessellationControlPerVertexInputComponents;
-    uint32_t maxTessellationControlPerVertexOutputComponents;
-    uint32_t maxTessellationControlPerPatchOutputComponents;
-    uint32_t maxTessellationControlTotalOutputComponents;
-    uint32_t maxTessellationEvaluationInputComponents;
-    uint32_t maxTessellationEvaluationOutputComponents;
-    uint32_t maxGeometryShaderInvocations;
-    uint32_t maxGeometryInputComponents;
-    uint32_t maxGeometryOutputComponents;
-    uint32_t maxGeometryOutputVertices;
-    uint32_t maxGeometryTotalOutputComponents;
-    uint32_t maxFragmentInputComponents;
-    uint32_t maxFragmentOutputAttachments;
-    uint32_t maxFragmentDualSrcAttachments;
-    uint32_t maxFragmentCombinedOutputResources;
-    uint32_t maxComputeSharedMemorySize;
-    uint32_t maxComputeWorkGroupCount[3];
-    uint32_t maxComputeWorkGroupInvocations;
-    uint32_t maxComputeWorkGroupSize[3];
-    uint32_t subPixelPrecisionBits;
-    uint32_t subTexelPrecisionBits;
-    uint32_t mipmapPrecisionBits;
-    uint32_t maxDrawIndexedIndexValue;
-    uint32_t maxDrawIndirectCount;
-    float maxSamplerLodBias;
-    float maxSamplerAnisotropy;
-    uint32_t maxViewports;
-    uint32_t maxViewportDimensions[2];
-    float viewportBoundsRange[2];
-    uint32_t viewportSubPixelBits;
-    size_t minMemoryMapAlignment;
-    VkDeviceSize minTexelBufferOffsetAlignment;
-    VkDeviceSize minUniformBufferOffsetAlignment;
-    VkDeviceSize minStorageBufferOffsetAlignment;
-    int32_t minTexelOffset;
-    uint32_t maxTexelOffset;
-    int32_t minTexelGatherOffset;
-    uint32_t maxTexelGatherOffset;
-    float minInterpolationOffset;
-    float maxInterpolationOffset;
-    uint32_t subPixelInterpolationOffsetBits;
-    uint32_t maxFramebufferWidth;
-    uint32_t maxFramebufferHeight;
-    uint32_t maxFramebufferLayers;
-    VkSampleCountFlags framebufferColorSampleCounts;
-    VkSampleCountFlags framebufferDepthSampleCounts;
-    VkSampleCountFlags framebufferStencilSampleCounts;
-    VkSampleCountFlags framebufferNoAttachmentsSampleCounts;
-    uint32_t maxColorAttachments;
-    VkSampleCountFlags sampledImageColorSampleCounts;
-    VkSampleCountFlags sampledImageIntegerSampleCounts;
-    VkSampleCountFlags sampledImageDepthSampleCounts;
-    VkSampleCountFlags sampledImageStencilSampleCounts;
-    VkSampleCountFlags storageImageSampleCounts;
-    uint32_t maxSampleMaskWords;
-    VkBool32 timestampComputeAndGraphics;
-    float timestampPeriod;
-    uint32_t maxClipDistances;
-    uint32_t maxCullDistances;
-    uint32_t maxCombinedClipAndCullDistances;
-    uint32_t discreteQueuePriorities;
-    float pointSizeRange[2];
-    float lineWidthRange[2];
-    float pointSizeGranularity;
-    float lineWidthGranularity;
-    VkBool32 strictLines;
-    VkBool32 standardSampleLocations;
-    VkDeviceSize optimalBufferCopyOffsetAlignment;
-    VkDeviceSize optimalBufferCopyRowPitchAlignment;
-    VkDeviceSize nonCoherentAtomSize;
-} VkPhysicalDeviceLimits;
-
-typedef struct VkPhysicalDeviceSparseProperties
-{
-    VkBool32 residencyStandard2DBlockShape;
-    VkBool32 residencyStandard2DMultisampleBlockShape;
-    VkBool32 residencyStandard3DBlockShape;
-    VkBool32 residencyAlignedMipSize;
-    VkBool32 residencyNonResidentStrict;
-} VkPhysicalDeviceSparseProperties;
-
-typedef struct VkPhysicalDeviceProperties
-{
-    uint32_t apiVersion;
-    uint32_t driverVersion;
-    uint32_t vendorID;
-    uint32_t deviceID;
-    VkPhysicalDeviceType deviceType;
-    char deviceName[VK_MAX_PHYSICAL_DEVICE_NAME_SIZE];
-    uint8_t pipelineCacheUUID[VK_UUID_SIZE];
-    VkPhysicalDeviceLimits limits;
-    VkPhysicalDeviceSparseProperties sparseProperties;
-} VkPhysicalDeviceProperties;
-
-typedef struct VkQueueFamilyProperties
-{
-    VkQueueFlags queueFlags;
-    uint32_t queueCount;
-    uint32_t timestampValidBits;
-    VkExtent3D minImageTransferGranularity;
-} VkQueueFamilyProperties;
-
-typedef struct VkMemoryType
-{
-    VkMemoryPropertyFlags propertyFlags;
-    uint32_t heapIndex;
-} VkMemoryType;
-
-typedef struct VkMemoryHeap
-{
-    VkDeviceSize size;
-    VkMemoryHeapFlags flags;
-} VkMemoryHeap;
-
-typedef struct VkPhysicalDeviceMemoryProperties
-{
-    uint32_t memoryTypeCount;
-    VkMemoryType memoryTypes[VK_MAX_MEMORY_TYPES];
-    uint32_t memoryHeapCount;
-    VkMemoryHeap memoryHeaps[VK_MAX_MEMORY_HEAPS];
-} VkPhysicalDeviceMemoryProperties;
-
-typedef void(VKAPI_PTR* PFN_vkVoidFunction)(void);
-typedef struct VkDeviceQueueCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkDeviceQueueCreateFlags flags;
-    uint32_t queueFamilyIndex;
-    uint32_t queueCount;
-    const float* pQueuePriorities;
-} VkDeviceQueueCreateInfo;
-
-typedef struct VkDeviceCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkDeviceCreateFlags flags;
-    uint32_t queueCreateInfoCount;
-    const VkDeviceQueueCreateInfo* pQueueCreateInfos;
-    uint32_t enabledLayerCount;
-    const char* const* ppEnabledLayerNames;
-    uint32_t enabledExtensionCount;
-    const char* const* ppEnabledExtensionNames;
-    const VkPhysicalDeviceFeatures* pEnabledFeatures;
-} VkDeviceCreateInfo;
-
-typedef struct VkExtensionProperties
-{
-    char extensionName[VK_MAX_EXTENSION_NAME_SIZE];
-    uint32_t specVersion;
-} VkExtensionProperties;
-
-typedef struct VkLayerProperties
-{
-    char layerName[VK_MAX_EXTENSION_NAME_SIZE];
-    uint32_t specVersion;
-    uint32_t implementationVersion;
-    char description[VK_MAX_DESCRIPTION_SIZE];
-} VkLayerProperties;
-
-typedef struct VkSubmitInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    uint32_t waitSemaphoreCount;
-    const VkSemaphore* pWaitSemaphores;
-    const VkPipelineStageFlags* pWaitDstStageMask;
-    uint32_t commandBufferCount;
-    const VkCommandBuffer* pCommandBuffers;
-    uint32_t signalSemaphoreCount;
-    const VkSemaphore* pSignalSemaphores;
-} VkSubmitInfo;
-
-typedef struct VkMemoryAllocateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkDeviceSize allocationSize;
-    uint32_t memoryTypeIndex;
-} VkMemoryAllocateInfo;
-
-typedef struct VkMappedMemoryRange
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkDeviceMemory memory;
-    VkDeviceSize offset;
-    VkDeviceSize size;
-} VkMappedMemoryRange;
-
-typedef struct VkMemoryRequirements
-{
-    VkDeviceSize size;
-    VkDeviceSize alignment;
-    uint32_t memoryTypeBits;
-} VkMemoryRequirements;
-
-typedef struct VkImageSubresource
-{
-    VkImageAspectFlags aspectMask;
-    uint32_t mipLevel;
-    uint32_t arrayLayer;
-} VkImageSubresource;
-
-typedef struct VkOffset3D
-{
-    int32_t x;
-    int32_t y;
-    int32_t z;
-} VkOffset3D;
-
-typedef struct VkFenceCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkFenceCreateFlags flags;
-} VkFenceCreateInfo;
-
-typedef struct VkSemaphoreCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkSemaphoreCreateFlags flags;
-} VkSemaphoreCreateInfo;
-
-typedef struct VkQueryPoolCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkQueryPoolCreateFlags flags;
-    VkQueryType queryType;
-    uint32_t queryCount;
-    VkQueryPipelineStatisticFlags pipelineStatistics;
-} VkQueryPoolCreateInfo;
-
-typedef struct VkBufferCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkBufferCreateFlags flags;
-    VkDeviceSize size;
-    VkBufferUsageFlags usage;
-    VkSharingMode sharingMode;
-    uint32_t queueFamilyIndexCount;
-    const uint32_t* pQueueFamilyIndices;
-} VkBufferCreateInfo;
-
-typedef struct VkBufferViewCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkBufferViewCreateFlags flags;
-    VkBuffer buffer;
-    VkFormat format;
-    VkDeviceSize offset;
-    VkDeviceSize range;
-} VkBufferViewCreateInfo;
-
-typedef struct VkImageCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkImageCreateFlags flags;
-    VkImageType imageType;
-    VkFormat format;
-    VkExtent3D extent;
-    uint32_t mipLevels;
-    uint32_t arrayLayers;
-    VkSampleCountFlagBits samples;
-    VkImageTiling tiling;
-    VkImageUsageFlags usage;
-    VkSharingMode sharingMode;
-    uint32_t queueFamilyIndexCount;
-    const uint32_t* pQueueFamilyIndices;
-    VkImageLayout initialLayout;
-} VkImageCreateInfo;
-
-typedef struct VkSubresourceLayout
-{
-    VkDeviceSize offset;
-    VkDeviceSize size;
-    VkDeviceSize rowPitch;
-    VkDeviceSize arrayPitch;
-    VkDeviceSize depthPitch;
-} VkSubresourceLayout;
-
-typedef struct VkComponentMapping
-{
-    VkComponentSwizzle r;
-    VkComponentSwizzle g;
-    VkComponentSwizzle b;
-    VkComponentSwizzle a;
-} VkComponentMapping;
-
-typedef struct VkImageSubresourceRange
-{
-    VkImageAspectFlags aspectMask;
-    uint32_t baseMipLevel;
-    uint32_t levelCount;
-    uint32_t baseArrayLayer;
-    uint32_t layerCount;
-} VkImageSubresourceRange;
-
-typedef struct VkImageViewCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkImageViewCreateFlags flags;
-    VkImage image;
-    VkImageViewType viewType;
-    VkFormat format;
-    VkComponentMapping components;
-    VkImageSubresourceRange subresourceRange;
-} VkImageViewCreateInfo;
-
-typedef struct VkShaderModuleCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkShaderModuleCreateFlags flags;
-    size_t codeSize;
-    const uint32_t* pCode;
-} VkShaderModuleCreateInfo;
-
-typedef struct VkPipelineCacheCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkPipelineCacheCreateFlags flags;
-    size_t initialDataSize;
-    const void* pInitialData;
-} VkPipelineCacheCreateInfo;
-
-typedef struct VkSpecializationMapEntry
-{
-    uint32_t constantID;
-    uint32_t offset;
-    size_t size;
-} VkSpecializationMapEntry;
-
-typedef struct VkSpecializationInfo
-{
-    uint32_t mapEntryCount;
-    const VkSpecializationMapEntry* pMapEntries;
-    size_t dataSize;
-    const void* pData;
-} VkSpecializationInfo;
-
-typedef struct VkPipelineShaderStageCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkPipelineShaderStageCreateFlags flags;
-    VkShaderStageFlagBits stage;
-    VkShaderModule module;
-    const char* pName;
-    const VkSpecializationInfo* pSpecializationInfo;
-} VkPipelineShaderStageCreateInfo;
-
-typedef struct VkViewport
-{
-    float x;
-    float y;
-    float width;
-    float height;
-    float minDepth;
-    float maxDepth;
-} VkViewport;
-
-typedef struct VkOffset2D
-{
-    int32_t x;
-    int32_t y;
-} VkOffset2D;
-
-typedef struct VkExtent2D
-{
-    uint32_t width;
-    uint32_t height;
-} VkExtent2D;
-
-typedef struct VkRect2D
-{
-    VkOffset2D offset;
-    VkExtent2D extent;
-} VkRect2D;
-
-typedef struct VkComputePipelineCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkPipelineCreateFlags flags;
-    VkPipelineShaderStageCreateInfo stage;
-    VkPipelineLayout layout;
-    VkPipeline basePipelineHandle;
-    int32_t basePipelineIndex;
-} VkComputePipelineCreateInfo;
-
-typedef struct VkPushConstantRange
-{
-    VkShaderStageFlags stageFlags;
-    uint32_t offset;
-    uint32_t size;
-} VkPushConstantRange;
-
-typedef struct VkPipelineLayoutCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkPipelineLayoutCreateFlags flags;
-    uint32_t setLayoutCount;
-    const VkDescriptorSetLayout* pSetLayouts;
-    uint32_t pushConstantRangeCount;
-    const VkPushConstantRange* pPushConstantRanges;
-} VkPipelineLayoutCreateInfo;
-
-typedef struct VkSamplerCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkSamplerCreateFlags flags;
-    VkFilter magFilter;
-    VkFilter minFilter;
-    VkSamplerMipmapMode mipmapMode;
-    VkSamplerAddressMode addressModeU;
-    VkSamplerAddressMode addressModeV;
-    VkSamplerAddressMode addressModeW;
-    float mipLodBias;
-    VkBool32 anisotropyEnable;
-    float maxAnisotropy;
-    VkBool32 compareEnable;
-    VkCompareOp compareOp;
-    float minLod;
-    float maxLod;
-    VkBorderColor borderColor;
-    VkBool32 unnormalizedCoordinates;
-} VkSamplerCreateInfo;
-
-typedef struct VkDescriptorSetLayoutBinding
-{
-    uint32_t binding;
-    VkDescriptorType descriptorType;
-    uint32_t descriptorCount;
-    VkShaderStageFlags stageFlags;
-    const VkSampler* pImmutableSamplers;
-} VkDescriptorSetLayoutBinding;
-
-typedef struct VkDescriptorSetLayoutCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkDescriptorSetLayoutCreateFlags flags;
-    uint32_t bindingCount;
-    const VkDescriptorSetLayoutBinding* pBindings;
-} VkDescriptorSetLayoutCreateInfo;
-
-typedef struct VkDescriptorPoolSize
-{
-    VkDescriptorType type;
-    uint32_t descriptorCount;
-} VkDescriptorPoolSize;
-
-typedef struct VkDescriptorPoolCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkDescriptorPoolCreateFlags flags;
-    uint32_t maxSets;
-    uint32_t poolSizeCount;
-    const VkDescriptorPoolSize* pPoolSizes;
-} VkDescriptorPoolCreateInfo;
-
-typedef struct VkDescriptorSetAllocateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkDescriptorPool descriptorPool;
-    uint32_t descriptorSetCount;
-    const VkDescriptorSetLayout* pSetLayouts;
-} VkDescriptorSetAllocateInfo;
-
-typedef struct VkDescriptorImageInfo
-{
-    VkSampler sampler;
-    VkImageView imageView;
-    VkImageLayout imageLayout;
-} VkDescriptorImageInfo;
-
-typedef struct VkDescriptorBufferInfo
-{
-    VkBuffer buffer;
-    VkDeviceSize offset;
-    VkDeviceSize range;
-} VkDescriptorBufferInfo;
-
-typedef struct VkWriteDescriptorSet
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkDescriptorSet dstSet;
-    uint32_t dstBinding;
-    uint32_t dstArrayElement;
-    uint32_t descriptorCount;
-    VkDescriptorType descriptorType;
-    const VkDescriptorImageInfo* pImageInfo;
-    const VkDescriptorBufferInfo* pBufferInfo;
-    const VkBufferView* pTexelBufferView;
-} VkWriteDescriptorSet;
-
-typedef struct VkCopyDescriptorSet
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkDescriptorSet srcSet;
-    uint32_t srcBinding;
-    uint32_t srcArrayElement;
-    VkDescriptorSet dstSet;
-    uint32_t dstBinding;
-    uint32_t dstArrayElement;
-    uint32_t descriptorCount;
-} VkCopyDescriptorSet;
-
-typedef struct VkCommandPoolCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkCommandPoolCreateFlags flags;
-    uint32_t queueFamilyIndex;
-} VkCommandPoolCreateInfo;
-
-typedef struct VkCommandBufferAllocateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkCommandPool commandPool;
-    VkCommandBufferLevel level;
-    uint32_t commandBufferCount;
-} VkCommandBufferAllocateInfo;
-
-typedef struct VkCommandBufferInheritanceInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkRenderPass renderPass;
-    uint32_t subpass;
-    VkFramebuffer framebuffer;
-    VkBool32 occlusionQueryEnable;
-    VkQueryControlFlags queryFlags;
-    VkQueryPipelineStatisticFlags pipelineStatistics;
-} VkCommandBufferInheritanceInfo;
-
-typedef struct VkCommandBufferBeginInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkCommandBufferUsageFlags flags;
-    const VkCommandBufferInheritanceInfo* pInheritanceInfo;
-} VkCommandBufferBeginInfo;
-
-typedef struct VkBufferCopy
-{
-    VkDeviceSize srcOffset;
-    VkDeviceSize dstOffset;
-    VkDeviceSize size;
-} VkBufferCopy;
-
-typedef struct VkImageSubresourceLayers
-{
-    VkImageAspectFlags aspectMask;
-    uint32_t mipLevel;
-    uint32_t baseArrayLayer;
-    uint32_t layerCount;
-} VkImageSubresourceLayers;
-
-typedef struct VkImageCopy
-{
-    VkImageSubresourceLayers srcSubresource;
-    VkOffset3D srcOffset;
-    VkImageSubresourceLayers dstSubresource;
-    VkOffset3D dstOffset;
-    VkExtent3D extent;
-} VkImageCopy;
-
-typedef struct VkBufferImageCopy
-{
-    VkDeviceSize bufferOffset;
-    uint32_t bufferRowLength;
-    uint32_t bufferImageHeight;
-    VkImageSubresourceLayers imageSubresource;
-    VkOffset3D imageOffset;
-    VkExtent3D imageExtent;
-} VkBufferImageCopy;
-
-typedef struct VkImageResolve
-{
-    VkImageSubresourceLayers srcSubresource;
-    VkOffset3D srcOffset;
-    VkImageSubresourceLayers dstSubresource;
-    VkOffset3D dstOffset;
-    VkExtent3D extent;
-} VkImageResolve;
-
-typedef struct VkMemoryBarrier
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkAccessFlags srcAccessMask;
-    VkAccessFlags dstAccessMask;
-} VkMemoryBarrier;
-
-typedef struct VkBufferMemoryBarrier
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkAccessFlags srcAccessMask;
-    VkAccessFlags dstAccessMask;
-    uint32_t srcQueueFamilyIndex;
-    uint32_t dstQueueFamilyIndex;
-    VkBuffer buffer;
-    VkDeviceSize offset;
-    VkDeviceSize size;
-} VkBufferMemoryBarrier;
-
-typedef struct VkImageMemoryBarrier
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkAccessFlags srcAccessMask;
-    VkAccessFlags dstAccessMask;
-    VkImageLayout oldLayout;
-    VkImageLayout newLayout;
-    uint32_t srcQueueFamilyIndex;
-    uint32_t dstQueueFamilyIndex;
-    VkImage image;
-    VkImageSubresourceRange subresourceRange;
-} VkImageMemoryBarrier;
-
-typedef struct VkDispatchIndirectCommand
-{
-    uint32_t x;
-    uint32_t y;
-    uint32_t z;
-} VkDispatchIndirectCommand;
-
-// vulkan base functions
-typedef VkResult(VKAPI_PTR* PFN_vkEnumerateInstanceExtensionProperties)(const char* pLayerName, uint32_t* pPropertyCount, VkExtensionProperties* pProperties);
-typedef VkResult(VKAPI_PTR* PFN_vkCreateInstance)(const VkInstanceCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkInstance* pInstance);
-typedef PFN_vkVoidFunction(VKAPI_PTR* PFN_vkGetInstanceProcAddr)(VkInstance instance, const char* pName);
-
-typedef VkResult(VKAPI_PTR* PFN_vkAllocateCommandBuffers)(VkDevice device, const VkCommandBufferAllocateInfo* pAllocateInfo, VkCommandBuffer* pCommandBuffers);
-typedef VkResult(VKAPI_PTR* PFN_vkAllocateDescriptorSets)(VkDevice device, const VkDescriptorSetAllocateInfo* pAllocateInfo, VkDescriptorSet* pDescriptorSets);
-typedef VkResult(VKAPI_PTR* PFN_vkAllocateMemory)(VkDevice device, const VkMemoryAllocateInfo* pAllocateInfo, const VkAllocationCallbacks* pAllocator, VkDeviceMemory* pMemory);
-typedef VkResult(VKAPI_PTR* PFN_vkBeginCommandBuffer)(VkCommandBuffer commandBuffer, const VkCommandBufferBeginInfo* pBeginInfo);
-typedef VkResult(VKAPI_PTR* PFN_vkBindBufferMemory)(VkDevice device, VkBuffer buffer, VkDeviceMemory memory, VkDeviceSize memoryOffset);
-typedef VkResult(VKAPI_PTR* PFN_vkBindImageMemory)(VkDevice device, VkImage image, VkDeviceMemory memory, VkDeviceSize memoryOffset);
-typedef void(VKAPI_PTR* PFN_vkCmdBeginQuery)(VkCommandBuffer commandBuffer, VkQueryPool queryPool, uint32_t query, VkQueryControlFlags flags);
-typedef void(VKAPI_PTR* PFN_vkCmdBindDescriptorSets)(VkCommandBuffer commandBuffer, VkPipelineBindPoint pipelineBindPoint, VkPipelineLayout layout, uint32_t firstSet, uint32_t descriptorSetCount, const VkDescriptorSet* pDescriptorSets, uint32_t dynamicOffsetCount, const uint32_t* pDynamicOffsets);
-typedef void(VKAPI_PTR* PFN_vkCmdBindIndexBuffer)(VkCommandBuffer commandBuffer, VkBuffer buffer, VkDeviceSize offset, VkIndexType indexType);
-typedef void(VKAPI_PTR* PFN_vkCmdBindPipeline)(VkCommandBuffer commandBuffer, VkPipelineBindPoint pipelineBindPoint, VkPipeline pipeline);
-typedef void(VKAPI_PTR* PFN_vkCmdCopyBuffer)(VkCommandBuffer commandBuffer, VkBuffer srcBuffer, VkBuffer dstBuffer, uint32_t regionCount, const VkBufferCopy* pRegions);
-typedef void(VKAPI_PTR* PFN_vkCmdCopyBufferToImage)(VkCommandBuffer commandBuffer, VkBuffer srcBuffer, VkImage dstImage, VkImageLayout dstImageLayout, uint32_t regionCount, const VkBufferImageCopy* pRegions);
-typedef void(VKAPI_PTR* PFN_vkCmdCopyImage)(VkCommandBuffer commandBuffer, VkImage srcImage, VkImageLayout srcImageLayout, VkImage dstImage, VkImageLayout dstImageLayout, uint32_t regionCount, const VkImageCopy* pRegions);
-typedef void(VKAPI_PTR* PFN_vkCmdCopyImageToBuffer)(VkCommandBuffer commandBuffer, VkImage srcImage, VkImageLayout srcImageLayout, VkBuffer dstBuffer, uint32_t regionCount, const VkBufferImageCopy* pRegions);
-typedef void(VKAPI_PTR* PFN_vkCmdCopyQueryPoolResults)(VkCommandBuffer commandBuffer, VkQueryPool queryPool, uint32_t firstQuery, uint32_t queryCount, VkBuffer dstBuffer, VkDeviceSize dstOffset, VkDeviceSize stride, VkQueryResultFlags flags);
-typedef void(VKAPI_PTR* PFN_vkCmdDispatch)(VkCommandBuffer commandBuffer, uint32_t groupCountX, uint32_t groupCountY, uint32_t groupCountZ);
-typedef void(VKAPI_PTR* PFN_vkCmdDispatchIndirect)(VkCommandBuffer commandBuffer, VkBuffer buffer, VkDeviceSize offset);
-typedef void(VKAPI_PTR* PFN_vkCmdEndQuery)(VkCommandBuffer commandBuffer, VkQueryPool queryPool, uint32_t query);
-typedef void(VKAPI_PTR* PFN_vkCmdExecuteCommands)(VkCommandBuffer commandBuffer, uint32_t commandBufferCount, const VkCommandBuffer* pCommandBuffers);
-typedef void(VKAPI_PTR* PFN_vkCmdFillBuffer)(VkCommandBuffer commandBuffer, VkBuffer dstBuffer, VkDeviceSize dstOffset, VkDeviceSize size, uint32_t data);
-typedef void(VKAPI_PTR* PFN_vkCmdPipelineBarrier)(VkCommandBuffer commandBuffer, VkPipelineStageFlags srcStageMask, VkPipelineStageFlags dstStageMask, VkDependencyFlags dependencyFlags, uint32_t memoryBarrierCount, const VkMemoryBarrier* pMemoryBarriers, uint32_t bufferMemoryBarrierCount, const VkBufferMemoryBarrier* pBufferMemoryBarriers, uint32_t imageMemoryBarrierCount, const VkImageMemoryBarrier* pImageMemoryBarriers);
-typedef void(VKAPI_PTR* PFN_vkCmdPushConstants)(VkCommandBuffer commandBuffer, VkPipelineLayout layout, VkShaderStageFlags stageFlags, uint32_t offset, uint32_t size, const void* pValues);
-typedef void(VKAPI_PTR* PFN_vkCmdResetQueryPool)(VkCommandBuffer commandBuffer, VkQueryPool queryPool, uint32_t firstQuery, uint32_t queryCount);
-typedef void(VKAPI_PTR* PFN_vkCmdResolveImage)(VkCommandBuffer commandBuffer, VkImage srcImage, VkImageLayout srcImageLayout, VkImage dstImage, VkImageLayout dstImageLayout, uint32_t regionCount, const VkImageResolve* pRegions);
-typedef void(VKAPI_PTR* PFN_vkCmdUpdateBuffer)(VkCommandBuffer commandBuffer, VkBuffer dstBuffer, VkDeviceSize dstOffset, VkDeviceSize dataSize, const void* pData);
-typedef void(VKAPI_PTR* PFN_vkCmdWriteTimestamp)(VkCommandBuffer commandBuffer, VkPipelineStageFlagBits pipelineStage, VkQueryPool queryPool, uint32_t query);
-typedef VkResult(VKAPI_PTR* PFN_vkCreateBuffer)(VkDevice device, const VkBufferCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkBuffer* pBuffer);
-typedef VkResult(VKAPI_PTR* PFN_vkCreateBufferView)(VkDevice device, const VkBufferViewCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkBufferView* pView);
-typedef VkResult(VKAPI_PTR* PFN_vkCreateCommandPool)(VkDevice device, const VkCommandPoolCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkCommandPool* pCommandPool);
-typedef VkResult(VKAPI_PTR* PFN_vkCreateComputePipelines)(VkDevice device, VkPipelineCache pipelineCache, uint32_t createInfoCount, const VkComputePipelineCreateInfo* pCreateInfos, const VkAllocationCallbacks* pAllocator, VkPipeline* pPipelines);
-typedef VkResult(VKAPI_PTR* PFN_vkCreateDescriptorPool)(VkDevice device, const VkDescriptorPoolCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkDescriptorPool* pDescriptorPool);
-typedef VkResult(VKAPI_PTR* PFN_vkCreateDescriptorSetLayout)(VkDevice device, const VkDescriptorSetLayoutCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkDescriptorSetLayout* pSetLayout);
-typedef VkResult(VKAPI_PTR* PFN_vkCreateDevice)(VkPhysicalDevice physicalDevice, const VkDeviceCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkDevice* pDevice);
-typedef VkResult(VKAPI_PTR* PFN_vkCreateFence)(VkDevice device, const VkFenceCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkFence* pFence);
-typedef VkResult(VKAPI_PTR* PFN_vkCreateImage)(VkDevice device, const VkImageCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkImage* pImage);
-typedef VkResult(VKAPI_PTR* PFN_vkCreateImageView)(VkDevice device, const VkImageViewCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkImageView* pView);
-typedef VkResult(VKAPI_PTR* PFN_vkCreatePipelineCache)(VkDevice device, const VkPipelineCacheCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkPipelineCache* pPipelineCache);
-typedef VkResult(VKAPI_PTR* PFN_vkCreatePipelineLayout)(VkDevice device, const VkPipelineLayoutCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkPipelineLayout* pPipelineLayout);
-typedef VkResult(VKAPI_PTR* PFN_vkCreateQueryPool)(VkDevice device, const VkQueryPoolCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkQueryPool* pQueryPool);
-typedef VkResult(VKAPI_PTR* PFN_vkCreateSampler)(VkDevice device, const VkSamplerCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSampler* pSampler);
-typedef VkResult(VKAPI_PTR* PFN_vkCreateSemaphore)(VkDevice device, const VkSemaphoreCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSemaphore* pSemaphore);
-typedef VkResult(VKAPI_PTR* PFN_vkCreateShaderModule)(VkDevice device, const VkShaderModuleCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkShaderModule* pShaderModule);
-typedef void(VKAPI_PTR* PFN_vkDestroyBuffer)(VkDevice device, VkBuffer buffer, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroyBufferView)(VkDevice device, VkBufferView bufferView, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroyCommandPool)(VkDevice device, VkCommandPool commandPool, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroyDescriptorPool)(VkDevice device, VkDescriptorPool descriptorPool, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroyDescriptorSetLayout)(VkDevice device, VkDescriptorSetLayout descriptorSetLayout, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroyDevice)(VkDevice device, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroyFence)(VkDevice device, VkFence fence, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroyImage)(VkDevice device, VkImage image, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroyImageView)(VkDevice device, VkImageView imageView, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroyInstance)(VkInstance instance, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroyPipeline)(VkDevice device, VkPipeline pipeline, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroyPipelineCache)(VkDevice device, VkPipelineCache pipelineCache, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroyPipelineLayout)(VkDevice device, VkPipelineLayout pipelineLayout, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroyQueryPool)(VkDevice device, VkQueryPool queryPool, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroySampler)(VkDevice device, VkSampler sampler, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroySemaphore)(VkDevice device, VkSemaphore semaphore, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkDestroyShaderModule)(VkDevice device, VkShaderModule shaderModule, const VkAllocationCallbacks* pAllocator);
-typedef VkResult(VKAPI_PTR* PFN_vkDeviceWaitIdle)(VkDevice device);
-typedef VkResult(VKAPI_PTR* PFN_vkEndCommandBuffer)(VkCommandBuffer commandBuffer);
-typedef VkResult(VKAPI_PTR* PFN_vkEnumerateDeviceExtensionProperties)(VkPhysicalDevice physicalDevice, const char* pLayerName, uint32_t* pPropertyCount, VkExtensionProperties* pProperties);
-typedef VkResult(VKAPI_PTR* PFN_vkEnumerateDeviceLayerProperties)(VkPhysicalDevice physicalDevice, uint32_t* pPropertyCount, VkLayerProperties* pProperties);
-typedef VkResult(VKAPI_PTR* PFN_vkEnumerateInstanceLayerProperties)(uint32_t* pPropertyCount, VkLayerProperties* pProperties);
-typedef VkResult(VKAPI_PTR* PFN_vkEnumeratePhysicalDevices)(VkInstance instance, uint32_t* pPhysicalDeviceCount, VkPhysicalDevice* pPhysicalDevices);
-typedef VkResult(VKAPI_PTR* PFN_vkFlushMappedMemoryRanges)(VkDevice device, uint32_t memoryRangeCount, const VkMappedMemoryRange* pMemoryRanges);
-typedef void(VKAPI_PTR* PFN_vkFreeCommandBuffers)(VkDevice device, VkCommandPool commandPool, uint32_t commandBufferCount, const VkCommandBuffer* pCommandBuffers);
-typedef VkResult(VKAPI_PTR* PFN_vkFreeDescriptorSets)(VkDevice device, VkDescriptorPool descriptorPool, uint32_t descriptorSetCount, const VkDescriptorSet* pDescriptorSets);
-typedef void(VKAPI_PTR* PFN_vkFreeMemory)(VkDevice device, VkDeviceMemory memory, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkGetBufferMemoryRequirements)(VkDevice device, VkBuffer buffer, VkMemoryRequirements* pMemoryRequirements);
-typedef void(VKAPI_PTR* PFN_vkGetDeviceMemoryCommitment)(VkDevice device, VkDeviceMemory memory, VkDeviceSize* pCommittedMemoryInBytes);
-typedef void(VKAPI_PTR* PFN_vkGetDeviceQueue)(VkDevice device, uint32_t queueFamilyIndex, uint32_t queueIndex, VkQueue* pQueue);
-typedef PFN_vkVoidFunction(VKAPI_PTR* PFN_vkGetDeviceProcAddr)(VkDevice device, const char* pName);
-typedef VkResult(VKAPI_PTR* PFN_vkGetFenceStatus)(VkDevice device, VkFence fence);
-typedef void(VKAPI_PTR* PFN_vkGetImageMemoryRequirements)(VkDevice device, VkImage image, VkMemoryRequirements* pMemoryRequirements);
-typedef void(VKAPI_PTR* PFN_vkGetImageSubresourceLayout)(VkDevice device, VkImage image, const VkImageSubresource* pSubresource, VkSubresourceLayout* pLayout);
-typedef void(VKAPI_PTR* PFN_vkGetPhysicalDeviceFeatures)(VkPhysicalDevice physicalDevice, VkPhysicalDeviceFeatures* pFeatures);
-typedef void(VKAPI_PTR* PFN_vkGetPhysicalDeviceFormatProperties)(VkPhysicalDevice physicalDevice, VkFormat format, VkFormatProperties* pFormatProperties);
-typedef VkResult(VKAPI_PTR* PFN_vkGetPhysicalDeviceImageFormatProperties)(VkPhysicalDevice physicalDevice, VkFormat format, VkImageType type, VkImageTiling tiling, VkImageUsageFlags usage, VkImageCreateFlags flags, VkImageFormatProperties* pImageFormatProperties);
-typedef void(VKAPI_PTR* PFN_vkGetPhysicalDeviceMemoryProperties)(VkPhysicalDevice physicalDevice, VkPhysicalDeviceMemoryProperties* pMemoryProperties);
-typedef void(VKAPI_PTR* PFN_vkGetPhysicalDeviceProperties)(VkPhysicalDevice physicalDevice, VkPhysicalDeviceProperties* pProperties);
-typedef void(VKAPI_PTR* PFN_vkGetPhysicalDeviceQueueFamilyProperties)(VkPhysicalDevice physicalDevice, uint32_t* pQueueFamilyPropertyCount, VkQueueFamilyProperties* pQueueFamilyProperties);
-typedef VkResult(VKAPI_PTR* PFN_vkGetPipelineCacheData)(VkDevice device, VkPipelineCache pipelineCache, size_t* pDataSize, void* pData);
-typedef VkResult(VKAPI_PTR* PFN_vkGetQueryPoolResults)(VkDevice device, VkQueryPool queryPool, uint32_t firstQuery, uint32_t queryCount, size_t dataSize, void* pData, VkDeviceSize stride, VkQueryResultFlags flags);
-typedef VkResult(VKAPI_PTR* PFN_vkInvalidateMappedMemoryRanges)(VkDevice device, uint32_t memoryRangeCount, const VkMappedMemoryRange* pMemoryRanges);
-typedef VkResult(VKAPI_PTR* PFN_vkMapMemory)(VkDevice device, VkDeviceMemory memory, VkDeviceSize offset, VkDeviceSize size, VkMemoryMapFlags flags, void** ppData);
-typedef VkResult(VKAPI_PTR* PFN_vkMergePipelineCaches)(VkDevice device, VkPipelineCache dstCache, uint32_t srcCacheCount, const VkPipelineCache* pSrcCaches);
-typedef VkResult(VKAPI_PTR* PFN_vkQueueSubmit)(VkQueue queue, uint32_t submitCount, const VkSubmitInfo* pSubmits, VkFence fence);
-typedef VkResult(VKAPI_PTR* PFN_vkQueueWaitIdle)(VkQueue queue);
-typedef VkResult(VKAPI_PTR* PFN_vkResetCommandBuffer)(VkCommandBuffer commandBuffer, VkCommandBufferResetFlags flags);
-typedef VkResult(VKAPI_PTR* PFN_vkResetCommandPool)(VkDevice device, VkCommandPool commandPool, VkCommandPoolResetFlags flags);
-typedef VkResult(VKAPI_PTR* PFN_vkResetDescriptorPool)(VkDevice device, VkDescriptorPool descriptorPool, VkDescriptorPoolResetFlags flags);
-typedef VkResult(VKAPI_PTR* PFN_vkResetFences)(VkDevice device, uint32_t fenceCount, const VkFence* pFences);
-typedef void(VKAPI_PTR* PFN_vkUnmapMemory)(VkDevice device, VkDeviceMemory memory);
-typedef void(VKAPI_PTR* PFN_vkUpdateDescriptorSets)(VkDevice device, uint32_t descriptorWriteCount, const VkWriteDescriptorSet* pDescriptorWrites, uint32_t descriptorCopyCount, const VkCopyDescriptorSet* pDescriptorCopies);
-typedef VkResult(VKAPI_PTR* PFN_vkWaitForFences)(VkDevice device, uint32_t fenceCount, const VkFence* pFences, VkBool32 waitAll, uint64_t timeout);
-
-#ifndef VK_KHR_surface
-#define VK_KHR_surface 1
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkSurfaceKHR)
-#define VK_COLORSPACE_SRGB_NONLINEAR_KHR VK_COLOR_SPACE_SRGB_NONLINEAR_KHR
-typedef enum VkColorSpaceKHR
-{
-    VK_COLOR_SPACE_SRGB_NONLINEAR_KHR = 0,
-    VK_COLOR_SPACE_DISPLAY_P3_NONLINEAR_EXT = 1000104001,
-    VK_COLOR_SPACE_EXTENDED_SRGB_LINEAR_EXT = 1000104002,
-    VK_COLOR_SPACE_DCI_P3_LINEAR_EXT = 1000104003,
-    VK_COLOR_SPACE_DCI_P3_NONLINEAR_EXT = 1000104004,
-    VK_COLOR_SPACE_BT709_LINEAR_EXT = 1000104005,
-    VK_COLOR_SPACE_BT709_NONLINEAR_EXT = 1000104006,
-    VK_COLOR_SPACE_BT2020_LINEAR_EXT = 1000104007,
-    VK_COLOR_SPACE_HDR10_ST2084_EXT = 1000104008,
-    VK_COLOR_SPACE_DOLBYVISION_EXT = 1000104009,
-    VK_COLOR_SPACE_HDR10_HLG_EXT = 1000104010,
-    VK_COLOR_SPACE_ADOBERGB_LINEAR_EXT = 1000104011,
-    VK_COLOR_SPACE_ADOBERGB_NONLINEAR_EXT = 1000104012,
-    VK_COLOR_SPACE_PASS_THROUGH_EXT = 1000104013,
-    VK_COLOR_SPACE_EXTENDED_SRGB_NONLINEAR_EXT = 1000104014,
-    VK_COLOR_SPACE_MAX_ENUM_KHR = 0x7FFFFFFF
-} VkColorSpaceKHR;
-typedef enum VkPresentModeKHR
-{
-    VK_PRESENT_MODE_IMMEDIATE_KHR = 0,
-    VK_PRESENT_MODE_MAILBOX_KHR = 1,
-    VK_PRESENT_MODE_FIFO_KHR = 2,
-    VK_PRESENT_MODE_FIFO_RELAXED_KHR = 3,
-    VK_PRESENT_MODE_SHARED_DEMAND_REFRESH_KHR = 1000111000,
-    VK_PRESENT_MODE_SHARED_CONTINUOUS_REFRESH_KHR = 1000111001,
-    VK_PRESENT_MODE_MAX_ENUM_KHR = 0x7FFFFFFF
-} VkPresentModeKHR;
-typedef enum VkSurfaceTransformFlagBitsKHR
-{
-    VK_SURFACE_TRANSFORM_IDENTITY_BIT_KHR = 0x00000001,
-    VK_SURFACE_TRANSFORM_ROTATE_90_BIT_KHR = 0x00000002,
-    VK_SURFACE_TRANSFORM_ROTATE_180_BIT_KHR = 0x00000004,
-    VK_SURFACE_TRANSFORM_ROTATE_270_BIT_KHR = 0x00000008,
-    VK_SURFACE_TRANSFORM_HORIZONTAL_MIRROR_BIT_KHR = 0x00000010,
-    VK_SURFACE_TRANSFORM_HORIZONTAL_MIRROR_ROTATE_90_BIT_KHR = 0x00000020,
-    VK_SURFACE_TRANSFORM_HORIZONTAL_MIRROR_ROTATE_180_BIT_KHR = 0x00000040,
-    VK_SURFACE_TRANSFORM_HORIZONTAL_MIRROR_ROTATE_270_BIT_KHR = 0x00000080,
-    VK_SURFACE_TRANSFORM_INHERIT_BIT_KHR = 0x00000100,
-    VK_SURFACE_TRANSFORM_FLAG_BITS_MAX_ENUM_KHR = 0x7FFFFFFF
-} VkSurfaceTransformFlagBitsKHR;
-typedef VkFlags VkSurfaceTransformFlagsKHR;
-typedef enum VkCompositeAlphaFlagBitsKHR
-{
-    VK_COMPOSITE_ALPHA_OPAQUE_BIT_KHR = 0x00000001,
-    VK_COMPOSITE_ALPHA_PRE_MULTIPLIED_BIT_KHR = 0x00000002,
-    VK_COMPOSITE_ALPHA_POST_MULTIPLIED_BIT_KHR = 0x00000004,
-    VK_COMPOSITE_ALPHA_INHERIT_BIT_KHR = 0x00000008,
-    VK_COMPOSITE_ALPHA_FLAG_BITS_MAX_ENUM_KHR = 0x7FFFFFFF
-} VkCompositeAlphaFlagBitsKHR;
-typedef VkFlags VkCompositeAlphaFlagsKHR;
-typedef struct VkSurfaceCapabilitiesKHR
-{
-    uint32_t minImageCount;
-    uint32_t maxImageCount;
-    VkExtent2D currentExtent;
-    VkExtent2D minImageExtent;
-    VkExtent2D maxImageExtent;
-    uint32_t maxImageArrayLayers;
-    VkSurfaceTransformFlagsKHR supportedTransforms;
-    VkSurfaceTransformFlagBitsKHR currentTransform;
-    VkCompositeAlphaFlagsKHR supportedCompositeAlpha;
-    VkImageUsageFlags supportedUsageFlags;
-} VkSurfaceCapabilitiesKHR;
-typedef struct VkSurfaceFormatKHR
-{
-    VkFormat format;
-    VkColorSpaceKHR colorSpace;
-} VkSurfaceFormatKHR;
-typedef void(VKAPI_PTR* PFN_vkDestroySurfaceKHR)(VkInstance instance, VkSurfaceKHR surface, const VkAllocationCallbacks* pAllocator);
-typedef VkResult(VKAPI_PTR* PFN_vkGetPhysicalDeviceSurfaceSupportKHR)(VkPhysicalDevice physicalDevice, uint32_t queueFamilyIndex, VkSurfaceKHR surface, VkBool32* pSupported);
-typedef VkResult(VKAPI_PTR* PFN_vkGetPhysicalDeviceSurfaceCapabilitiesKHR)(VkPhysicalDevice physicalDevice, VkSurfaceKHR surface, VkSurfaceCapabilitiesKHR* pSurfaceCapabilities);
-typedef VkResult(VKAPI_PTR* PFN_vkGetPhysicalDeviceSurfaceFormatsKHR)(VkPhysicalDevice physicalDevice, VkSurfaceKHR surface, uint32_t* pSurfaceFormatCount, VkSurfaceFormatKHR* pSurfaceFormats);
-typedef VkResult(VKAPI_PTR* PFN_vkGetPhysicalDeviceSurfacePresentModesKHR)(VkPhysicalDevice physicalDevice, VkSurfaceKHR surface, uint32_t* pPresentModeCount, VkPresentModeKHR* pPresentModes);
-#endif // VK_KHR_surface
-
-#ifndef VK_KHR_swapchain
-#define VK_KHR_swapchain 1
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkSwapchainKHR)
-typedef enum VkSwapchainCreateFlagBitsKHR
-{
-    VK_SWAPCHAIN_CREATE_BIND_SFR_BIT_KHX = 0x00000001,
-    VK_SWAPCHAIN_CREATE_FLAG_BITS_MAX_ENUM_KHR = 0x7FFFFFFF
-} VkSwapchainCreateFlagBitsKHR;
-typedef VkFlags VkSwapchainCreateFlagsKHR;
-typedef struct VkSwapchainCreateInfoKHR
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkSwapchainCreateFlagsKHR flags;
-    VkSurfaceKHR surface;
-    uint32_t minImageCount;
-    VkFormat imageFormat;
-    VkColorSpaceKHR imageColorSpace;
-    VkExtent2D imageExtent;
-    uint32_t imageArrayLayers;
-    VkImageUsageFlags imageUsage;
-    VkSharingMode imageSharingMode;
-    uint32_t queueFamilyIndexCount;
-    const uint32_t* pQueueFamilyIndices;
-    VkSurfaceTransformFlagBitsKHR preTransform;
-    VkCompositeAlphaFlagBitsKHR compositeAlpha;
-    VkPresentModeKHR presentMode;
-    VkBool32 clipped;
-    VkSwapchainKHR oldSwapchain;
-} VkSwapchainCreateInfoKHR;
-typedef struct VkPresentInfoKHR
-{
-    VkStructureType sType;
-    const void* pNext;
-    uint32_t waitSemaphoreCount;
-    const VkSemaphore* pWaitSemaphores;
-    uint32_t swapchainCount;
-    const VkSwapchainKHR* pSwapchains;
-    const uint32_t* pImageIndices;
-    VkResult* pResults;
-} VkPresentInfoKHR;
-typedef VkResult(VKAPI_PTR* PFN_vkCreateSwapchainKHR)(VkDevice device, const VkSwapchainCreateInfoKHR* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSwapchainKHR* pSwapchain);
-typedef void(VKAPI_PTR* PFN_vkDestroySwapchainKHR)(VkDevice device, VkSwapchainKHR swapchain, const VkAllocationCallbacks* pAllocator);
-typedef VkResult(VKAPI_PTR* PFN_vkGetSwapchainImagesKHR)(VkDevice device, VkSwapchainKHR swapchain, uint32_t* pSwapchainImageCount, VkImage* pSwapchainImages);
-typedef VkResult(VKAPI_PTR* PFN_vkAcquireNextImageKHR)(VkDevice device, VkSwapchainKHR swapchain, uint64_t timeout, VkSemaphore semaphore, VkFence fence, uint32_t* pImageIndex);
-typedef VkResult(VKAPI_PTR* PFN_vkQueuePresentKHR)(VkQueue queue, const VkPresentInfoKHR* pPresentInfo);
-#endif // VK_KHR_swapchain
-
-#ifndef VK_KHR_android_surface
-#define VK_KHR_android_surface 1
-struct ANativeWindow;
-typedef VkFlags VkAndroidSurfaceCreateFlagsKHR;
-typedef struct VkAndroidSurfaceCreateInfoKHR
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkAndroidSurfaceCreateFlagsKHR flags;
-    struct ANativeWindow* window;
-} VkAndroidSurfaceCreateInfoKHR;
-typedef VkResult(VKAPI_PTR* PFN_vkCreateAndroidSurfaceKHR)(VkInstance instance, const VkAndroidSurfaceCreateInfoKHR* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSurfaceKHR* pSurface);
-#endif // VK_KHR_android_surface
-
-namespace ncnn {
-
-// vulkan loader entrypoint
-extern PFN_vkGetInstanceProcAddr vkGetInstanceProcAddr;
-
-// vulkan global functions
-extern PFN_vkEnumerateInstanceExtensionProperties vkEnumerateInstanceExtensionProperties;
-extern PFN_vkCreateInstance vkCreateInstance;
-extern PFN_vkEnumerateInstanceLayerProperties vkEnumerateInstanceLayerProperties;
-
-// Load vulkan driver dynamic library when driver_path is provided
-// Load from env VK_ICD_FILENAMES and NCNN_VULKAN_DRIVER otherwise
-// If failed, it will try system libvulkan and well-known library names
-// The three vulkan base function pointers will be set on success
-int load_vulkan_driver(const char* driver_path = 0);
-
-// Unload vulkan driver library and clear
-void unload_vulkan_driver();
-
-} // namespace ncnn
-
-#endif // NCNN_SIMPLEVK
-#endif // NCNN_VULKAN
-
-#endif // NCNN_SIMPLEVK_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/stb_image.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/stb_image.h
deleted file mode 100644
index 1c2096a3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/stb_image.h
+++ /dev/null
@@ -1,8362 +0,0 @@
-/* stb_image - v2.28 - public domain image loader - http://nothings.org/stb
-                                  no warranty implied; use at your own risk
-
-   Do this:
-      #define STB_IMAGE_IMPLEMENTATION
-   before you include this file in *one* C or C++ file to create the implementation.
-
-   // i.e. it should look like this:
-   #include ...
-   #include ...
-   #include ...
-   #define STB_IMAGE_IMPLEMENTATION
-   #include "stb_image.h"
-
-   You can #define STBI_ASSERT(x) before the #include to avoid using assert.h.
-   And #define STBI_MALLOC, STBI_REALLOC, and STBI_FREE to avoid using malloc,realloc,free
-
-
-   QUICK NOTES:
-      Primarily of interest to game developers and other people who can
-          avoid problematic images and only need the trivial interface
-
-      JPEG baseline & progressive (12 bpc/arithmetic not supported, same as stock IJG lib)
-      PNG 1/2/4/8/16-bit-per-channel
-
-      TGA (not sure what subset, if a subset)
-      BMP non-1bpp, non-RLE
-      PSD (composited view only, no extra channels, 8/16 bit-per-channel)
-
-      GIF (*comp always reports as 4-channel)
-      HDR (radiance rgbE format)
-      PIC (Softimage PIC)
-      PNM (PPM and PGM binary only)
-
-      Animated GIF still needs a proper API, but here's one way to do it:
-          http://gist.github.com/urraka/685d9a6340b26b830d49
-
-      - decode from memory or through FILE (define STBI_NO_STDIO to remove code)
-      - decode from arbitrary I/O callbacks
-      - SIMD acceleration on x86/x64 (SSE2) and ARM (NEON)
-
-   Full documentation under "DOCUMENTATION" below.
-
-
-LICENSE
-
-  See end of file for license information.
-
-RECENT REVISION HISTORY:
-
-      2.28  (2023-01-29) many error fixes, security errors, just tons of stuff
-      2.27  (2021-07-11) document stbi_info better, 16-bit PNM support, bug fixes
-      2.26  (2020-07-13) many minor fixes
-      2.25  (2020-02-02) fix warnings
-      2.24  (2020-02-02) fix warnings; thread-local failure_reason and flip_vertically
-      2.23  (2019-08-11) fix clang static analysis warning
-      2.22  (2019-03-04) gif fixes, fix warnings
-      2.21  (2019-02-25) fix typo in comment
-      2.20  (2019-02-07) support utf8 filenames in Windows; fix warnings and platform ifdefs
-      2.19  (2018-02-11) fix warning
-      2.18  (2018-01-30) fix warnings
-      2.17  (2018-01-29) bugfix, 1-bit BMP, 16-bitness query, fix warnings
-      2.16  (2017-07-23) all functions have 16-bit variants; optimizations; bugfixes
-      2.15  (2017-03-18) fix png-1,2,4; all Imagenet JPGs; no runtime SSE detection on GCC
-      2.14  (2017-03-03) remove deprecated STBI_JPEG_OLD; fixes for Imagenet JPGs
-      2.13  (2016-12-04) experimental 16-bit API, only for PNG so far; fixes
-      2.12  (2016-04-02) fix typo in 2.11 PSD fix that caused crashes
-      2.11  (2016-04-02) 16-bit PNGS; enable SSE2 in non-gcc x64
-                         RGB-format JPEG; remove white matting in PSD;
-                         allocate large structures on the stack;
-                         correct channel count for PNG & BMP
-      2.10  (2016-01-22) avoid warning introduced in 2.09
-      2.09  (2016-01-16) 16-bit TGA; comments in PNM files; STBI_REALLOC_SIZED
-
-   See end of file for full revision history.
-
-
- ============================    Contributors    =========================
-
- Image formats                          Extensions, features
-    Sean Barrett (jpeg, png, bmp)          Jetro Lauha (stbi_info)
-    Nicolas Schulz (hdr, psd)              Martin "SpartanJ" Golini (stbi_info)
-    Jonathan Dummer (tga)                  James "moose2000" Brown (iPhone PNG)
-    Jean-Marc Lienher (gif)                Ben "Disch" Wenger (io callbacks)
-    Tom Seddon (pic)                       Omar Cornut (1/2/4-bit PNG)
-    Thatcher Ulrich (psd)                  Nicolas Guillemot (vertical flip)
-    Ken Miller (pgm, ppm)                  Richard Mitton (16-bit PSD)
-    github:urraka (animated gif)           Junggon Kim (PNM comments)
-    Christopher Forseth (animated gif)     Daniel Gibson (16-bit TGA)
-                                           socks-the-fox (16-bit PNG)
-                                           Jeremy Sawicki (handle all ImageNet JPGs)
- Optimizations & bugfixes                  Mikhail Morozov (1-bit BMP)
-    Fabian "ryg" Giesen                    Anael Seghezzi (is-16-bit query)
-    Arseny Kapoulkine                      Simon Breuss (16-bit PNM)
-    John-Mark Allen
-    Carmelo J Fdez-Aguera
-
- Bug & warning fixes
-    Marc LeBlanc            David Woo          Guillaume George     Martins Mozeiko
-    Christpher Lloyd        Jerry Jansson      Joseph Thomson       Blazej Dariusz Roszkowski
-    Phil Jordan                                Dave Moore           Roy Eltham
-    Hayaki Saito            Nathan Reed        Won Chun
-    Luke Graham             Johan Duparc       Nick Verigakis       the Horde3D community
-    Thomas Ruf              Ronny Chevalier                         github:rlyeh
-    Janez Zemva             John Bartholomew   Michal Cichon        github:romigrou
-    Jonathan Blow           Ken Hamada         Tero Hanninen        github:svdijk
-    Eugene Golushkov        Laurent Gomila     Cort Stratton        github:snagar
-    Aruelien Pocheville     Sergio Gonzalez    Thibault Reuille     github:Zelex
-    Cass Everitt            Ryamond Barbiero                        github:grim210
-    Paul Du Bois            Engin Manap        Aldo Culquicondor    github:sammyhw
-    Philipp Wiesemann       Dale Weiler        Oriol Ferrer Mesia   github:phprus
-    Josh Tobin              Neil Bickford      Matthew Gregan       github:poppolopoppo
-    Julian Raschke          Gregory Mullen     Christian Floisand   github:darealshinji
-    Baldur Karlsson         Kevin Schmidt      JR Smith             github:Michaelangel007
-                            Brad Weinberger    Matvey Cherevko      github:mosra
-    Luca Sas                Alexander Veselov  Zack Middleton       [reserved]
-    Ryan C. Gordon          [reserved]                              [reserved]
-                     DO NOT ADD YOUR NAME HERE
-
-                     Jacko Dirks
-
-  To add your name to the credits, pick a random blank space in the middle and fill it.
-  80% of merge conflicts on stb PRs are due to people adding their name at the end
-  of the credits.
-*/
-
-#ifndef STBI_INCLUDE_STB_IMAGE_H
-#define STBI_INCLUDE_STB_IMAGE_H
-
-// DOCUMENTATION
-//
-// Limitations:
-//    - no 12-bit-per-channel JPEG
-//    - no JPEGs with arithmetic coding
-//    - GIF always returns *comp=4
-//
-// Basic usage (see HDR discussion below for HDR usage):
-//    int x,y,n;
-//    unsigned char *data = stbi_load(filename, &x, &y, &n, 0);
-//    // ... process data if not NULL ...
-//    // ... x = width, y = height, n = # 8-bit components per pixel ...
-//    // ... replace '0' with '1'..'4' to force that many components per pixel
-//    // ... but 'n' will always be the number that it would have been if you said 0
-//    stbi_image_free(data);
-//
-// Standard parameters:
-//    int *x                 -- outputs image width in pixels
-//    int *y                 -- outputs image height in pixels
-//    int *channels_in_file  -- outputs # of image components in image file
-//    int desired_channels   -- if non-zero, # of image components requested in result
-//
-// The return value from an image loader is an 'unsigned char *' which points
-// to the pixel data, or NULL on an allocation failure or if the image is
-// corrupt or invalid. The pixel data consists of *y scanlines of *x pixels,
-// with each pixel consisting of N interleaved 8-bit components; the first
-// pixel pointed to is top-left-most in the image. There is no padding between
-// image scanlines or between pixels, regardless of format. The number of
-// components N is 'desired_channels' if desired_channels is non-zero, or
-// *channels_in_file otherwise. If desired_channels is non-zero,
-// *channels_in_file has the number of components that _would_ have been
-// output otherwise. E.g. if you set desired_channels to 4, you will always
-// get RGBA output, but you can check *channels_in_file to see if it's trivially
-// opaque because e.g. there were only 3 channels in the source image.
-//
-// An output image with N components has the following components interleaved
-// in this order in each pixel:
-//
-//     N=#comp     components
-//       1           grey
-//       2           grey, alpha
-//       3           red, green, blue
-//       4           red, green, blue, alpha
-//
-// If image loading fails for any reason, the return value will be NULL,
-// and *x, *y, *channels_in_file will be unchanged. The function
-// stbi_failure_reason() can be queried for an extremely brief, end-user
-// unfriendly explanation of why the load failed. Define STBI_NO_FAILURE_STRINGS
-// to avoid compiling these strings at all, and STBI_FAILURE_USERMSG to get slightly
-// more user-friendly ones.
-//
-// Paletted PNG, BMP, GIF, and PIC images are automatically depalettized.
-//
-// To query the width, height and component count of an image without having to
-// decode the full file, you can use the stbi_info family of functions:
-//
-//   int x,y,n,ok;
-//   ok = stbi_info(filename, &x, &y, &n);
-//   // returns ok=1 and sets x, y, n if image is a supported format,
-//   // 0 otherwise.
-//
-// Note that stb_image pervasively uses ints in its public API for sizes,
-// including sizes of memory buffers. This is now part of the API and thus
-// hard to change without causing breakage. As a result, the various image
-// loaders all have certain limits on image size; these differ somewhat
-// by format but generally boil down to either just under 2GB or just under
-// 1GB. When the decoded image would be larger than this, stb_image decoding
-// will fail.
-//
-// Additionally, stb_image will reject image files that have any of their
-// dimensions set to a larger value than the configurable STBI_MAX_DIMENSIONS,
-// which defaults to 2**24 = 16777216 pixels. Due to the above memory limit,
-// the only way to have an image with such dimensions load correctly
-// is for it to have a rather extreme aspect ratio. Either way, the
-// assumption here is that such larger images are likely to be malformed
-// or malicious. If you do need to load an image with individual dimensions
-// larger than that, and it still fits in the overall size limit, you can
-// #define STBI_MAX_DIMENSIONS on your own to be something larger.
-//
-// ===========================================================================
-//
-// UNICODE:
-//
-//   If compiling for Windows and you wish to use Unicode filenames, compile
-//   with
-//       #define STBI_WINDOWS_UTF8
-//   and pass utf8-encoded filenames. Call stbi_convert_wchar_to_utf8 to convert
-//   Windows wchar_t filenames to utf8.
-//
-// ===========================================================================
-//
-// Philosophy
-//
-// stb libraries are designed with the following priorities:
-//
-//    1. easy to use
-//    2. easy to maintain
-//    3. good performance
-//
-// Sometimes I let "good performance" creep up in priority over "easy to maintain",
-// and for best performance I may provide less-easy-to-use APIs that give higher
-// performance, in addition to the easy-to-use ones. Nevertheless, it's important
-// to keep in mind that from the standpoint of you, a client of this library,
-// all you care about is #1 and #3, and stb libraries DO NOT emphasize #3 above all.
-//
-// Some secondary priorities arise directly from the first two, some of which
-// provide more explicit reasons why performance can't be emphasized.
-//
-//    - Portable ("ease of use")
-//    - Small source code footprint ("easy to maintain")
-//    - No dependencies ("ease of use")
-//
-// ===========================================================================
-//
-// I/O callbacks
-//
-// I/O callbacks allow you to read from arbitrary sources, like packaged
-// files or some other source. Data read from callbacks are processed
-// through a small internal buffer (currently 128 bytes) to try to reduce
-// overhead.
-//
-// The three functions you must define are "read" (reads some bytes of data),
-// "skip" (skips some bytes of data), "eof" (reports if the stream is at the end).
-//
-// ===========================================================================
-//
-// SIMD support
-//
-// The JPEG decoder will try to automatically use SIMD kernels on x86 when
-// supported by the compiler. For ARM Neon support, you must explicitly
-// request it.
-//
-// (The old do-it-yourself SIMD API is no longer supported in the current
-// code.)
-//
-// On x86, SSE2 will automatically be used when available based on a run-time
-// test; if not, the generic C versions are used as a fall-back. On ARM targets,
-// the typical path is to have separate builds for NEON and non-NEON devices
-// (at least this is true for iOS and Android). Therefore, the NEON support is
-// toggled by a build flag: define STBI_NEON to get NEON loops.
-//
-// If for some reason you do not want to use any of SIMD code, or if
-// you have issues compiling it, you can disable it entirely by
-// defining STBI_NO_SIMD.
-//
-// ===========================================================================
-//
-// HDR image support   (disable by defining STBI_NO_HDR)
-//
-// stb_image supports loading HDR images in general, and currently the Radiance
-// .HDR file format specifically. You can still load any file through the existing
-// interface; if you attempt to load an HDR file, it will be automatically remapped
-// to LDR, assuming gamma 2.2 and an arbitrary scale factor defaulting to 1;
-// both of these constants can be reconfigured through this interface:
-//
-//     stbi_hdr_to_ldr_gamma(2.2f);
-//     stbi_hdr_to_ldr_scale(1.0f);
-//
-// (note, do not use _inverse_ constants; stbi_image will invert them
-// appropriately).
-//
-// Additionally, there is a new, parallel interface for loading files as
-// (linear) floats to preserve the full dynamic range:
-//
-//    float *data = stbi_loadf(filename, &x, &y, &n, 0);
-//
-// If you load LDR images through this interface, those images will
-// be promoted to floating point values, run through the inverse of
-// constants corresponding to the above:
-//
-//     stbi_ldr_to_hdr_scale(1.0f);
-//     stbi_ldr_to_hdr_gamma(2.2f);
-//
-// Finally, given a filename (or an open file or memory block--see header
-// file for details) containing image data, you can query for the "most
-// appropriate" interface to use (that is, whether the image is HDR or
-// not), using:
-//
-//     stbi_is_hdr(char *filename);
-//
-// ===========================================================================
-//
-// iPhone PNG support:
-//
-// We optionally support converting iPhone-formatted PNGs (which store
-// premultiplied BGRA) back to RGB, even though they're internally encoded
-// differently. To enable this conversion, call
-// stbi_convert_iphone_png_to_rgb(1).
-//
-// Call stbi_set_unpremultiply_on_load(1) as well to force a divide per
-// pixel to remove any premultiplied alpha *only* if the image file explicitly
-// says there's premultiplied data (currently only happens in iPhone images,
-// and only if iPhone convert-to-rgb processing is on).
-//
-// ===========================================================================
-//
-// ADDITIONAL CONFIGURATION
-//
-//  - You can suppress implementation of any of the decoders to reduce
-//    your code footprint by #defining one or more of the following
-//    symbols before creating the implementation.
-//
-//        STBI_NO_JPEG
-//        STBI_NO_PNG
-//        STBI_NO_BMP
-//        STBI_NO_PSD
-//        STBI_NO_TGA
-//        STBI_NO_GIF
-//        STBI_NO_HDR
-//        STBI_NO_PIC
-//        STBI_NO_PNM   (.ppm and .pgm)
-//
-//  - You can request *only* certain decoders and suppress all other ones
-//    (this will be more forward-compatible, as addition of new decoders
-//    doesn't require you to disable them explicitly):
-//
-//        STBI_ONLY_JPEG
-//        STBI_ONLY_PNG
-//        STBI_ONLY_BMP
-//        STBI_ONLY_PSD
-//        STBI_ONLY_TGA
-//        STBI_ONLY_GIF
-//        STBI_ONLY_HDR
-//        STBI_ONLY_PIC
-//        STBI_ONLY_PNM   (.ppm and .pgm)
-//
-//   - If you use STBI_NO_PNG (or _ONLY_ without PNG), and you still
-//     want the zlib decoder to be available, #define STBI_SUPPORT_ZLIB
-//
-//  - If you define STBI_MAX_DIMENSIONS, stb_image will reject images greater
-//    than that size (in either width or height) without further processing.
-//    This is to let programs in the wild set an upper bound to prevent
-//    denial-of-service attacks on untrusted data, as one could generate a
-//    valid image of gigantic dimensions and force stb_image to allocate a
-//    huge block of memory and spend disproportionate time decoding it. By
-//    default this is set to (1 << 24), which is 16777216, but that's still
-//    very big.
-
-#ifndef STBI_NO_STDIO
-#include <stdio.h>
-#endif // STBI_NO_STDIO
-
-#define STBI_VERSION 1
-
-enum
-{
-   STBI_default = 0, // only used for desired_channels
-
-   STBI_grey       = 1,
-   STBI_grey_alpha = 2,
-   STBI_rgb        = 3,
-   STBI_rgb_alpha  = 4
-};
-
-#include <stdlib.h>
-typedef unsigned char stbi_uc;
-typedef unsigned short stbi_us;
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#ifndef STBIDEF
-#ifdef STB_IMAGE_STATIC
-#define STBIDEF static
-#else
-#define STBIDEF extern
-#endif
-#endif
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// PRIMARY API - works on images of any type
-//
-
-//
-// load image by filename, open file, or memory buffer
-//
-
-typedef struct
-{
-   int      (*read)  (void *user,char *data,int size);   // fill 'data' with 'size' bytes.  return number of bytes actually read
-   void     (*skip)  (void *user,int n);                 // skip the next 'n' bytes, or 'unget' the last -n bytes if negative
-   int      (*eof)   (void *user);                       // returns nonzero if we are at end of file/data
-} stbi_io_callbacks;
-
-////////////////////////////////////
-//
-// 8-bits-per-channel interface
-//
-
-STBIDEF stbi_uc *stbi_load_from_memory   (stbi_uc           const *buffer, int len   , int *x, int *y, int *channels_in_file, int desired_channels);
-STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk  , void *user, int *x, int *y, int *channels_in_file, int desired_channels);
-
-#ifndef STBI_NO_STDIO
-STBIDEF stbi_uc *stbi_load            (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels);
-STBIDEF stbi_uc *stbi_load_from_file  (FILE *f, int *x, int *y, int *channels_in_file, int desired_channels);
-// for stbi_load_from_file, file pointer is left pointing immediately after image
-#endif
-
-#ifndef STBI_NO_GIF
-STBIDEF stbi_uc *stbi_load_gif_from_memory(stbi_uc const *buffer, int len, int **delays, int *x, int *y, int *z, int *comp, int req_comp);
-#endif
-
-#ifdef STBI_WINDOWS_UTF8
-STBIDEF int stbi_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input);
-#endif
-
-////////////////////////////////////
-//
-// 16-bits-per-channel interface
-//
-
-STBIDEF stbi_us *stbi_load_16_from_memory   (stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels);
-STBIDEF stbi_us *stbi_load_16_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels);
-
-#ifndef STBI_NO_STDIO
-STBIDEF stbi_us *stbi_load_16          (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels);
-STBIDEF stbi_us *stbi_load_from_file_16(FILE *f, int *x, int *y, int *channels_in_file, int desired_channels);
-#endif
-
-////////////////////////////////////
-//
-// float-per-channel interface
-//
-#ifndef STBI_NO_LINEAR
-   STBIDEF float *stbi_loadf_from_memory     (stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels);
-   STBIDEF float *stbi_loadf_from_callbacks  (stbi_io_callbacks const *clbk, void *user, int *x, int *y,  int *channels_in_file, int desired_channels);
-
-   #ifndef STBI_NO_STDIO
-   STBIDEF float *stbi_loadf            (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels);
-   STBIDEF float *stbi_loadf_from_file  (FILE *f, int *x, int *y, int *channels_in_file, int desired_channels);
-   #endif
-#endif
-
-#ifndef STBI_NO_HDR
-   STBIDEF void   stbi_hdr_to_ldr_gamma(float gamma);
-   STBIDEF void   stbi_hdr_to_ldr_scale(float scale);
-#endif // STBI_NO_HDR
-
-#ifndef STBI_NO_LINEAR
-   STBIDEF void   stbi_ldr_to_hdr_gamma(float gamma);
-   STBIDEF void   stbi_ldr_to_hdr_scale(float scale);
-#endif // STBI_NO_LINEAR
-
-// stbi_is_hdr is always defined, but always returns false if STBI_NO_HDR
-STBIDEF int    stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user);
-STBIDEF int    stbi_is_hdr_from_memory(stbi_uc const *buffer, int len);
-#ifndef STBI_NO_STDIO
-STBIDEF int      stbi_is_hdr          (char const *filename);
-STBIDEF int      stbi_is_hdr_from_file(FILE *f);
-#endif // STBI_NO_STDIO
-
-
-// get a VERY brief reason for failure
-// on most compilers (and ALL modern mainstream compilers) this is threadsafe
-STBIDEF const char *stbi_failure_reason  (void);
-
-// free the loaded image -- this is just free()
-STBIDEF void     stbi_image_free      (void *retval_from_stbi_load);
-
-// get image dimensions & components without fully decoding
-STBIDEF int      stbi_info_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp);
-STBIDEF int      stbi_info_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp);
-STBIDEF int      stbi_is_16_bit_from_memory(stbi_uc const *buffer, int len);
-STBIDEF int      stbi_is_16_bit_from_callbacks(stbi_io_callbacks const *clbk, void *user);
-
-#ifndef STBI_NO_STDIO
-STBIDEF int      stbi_info               (char const *filename,     int *x, int *y, int *comp);
-STBIDEF int      stbi_info_from_file     (FILE *f,                  int *x, int *y, int *comp);
-STBIDEF int      stbi_is_16_bit          (char const *filename);
-STBIDEF int      stbi_is_16_bit_from_file(FILE *f);
-#endif
-
-
-
-// for image formats that explicitly notate that they have premultiplied alpha,
-// we just return the colors as stored in the file. set this flag to force
-// unpremultiplication. results are undefined if the unpremultiply overflow.
-STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply);
-
-// indicate whether we should process iphone images back to canonical format,
-// or just pass them through "as-is"
-STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert);
-
-// flip the image vertically, so the first pixel in the output array is the bottom left
-STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip);
-
-// as above, but only applies to images loaded on the thread that calls the function
-// this function is only available if your compiler supports thread-local variables;
-// calling it will fail to link if your compiler doesn't
-STBIDEF void stbi_set_unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply);
-STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert);
-STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_flip);
-
-// ZLIB client - used by PNG, available for other purposes
-
-STBIDEF char *stbi_zlib_decode_malloc_guesssize(const char *buffer, int len, int initial_size, int *outlen);
-STBIDEF char *stbi_zlib_decode_malloc_guesssize_headerflag(const char *buffer, int len, int initial_size, int *outlen, int parse_header);
-STBIDEF char *stbi_zlib_decode_malloc(const char *buffer, int len, int *outlen);
-STBIDEF int   stbi_zlib_decode_buffer(char *obuffer, int olen, const char *ibuffer, int ilen);
-
-STBIDEF char *stbi_zlib_decode_noheader_malloc(const char *buffer, int len, int *outlen);
-STBIDEF int   stbi_zlib_decode_noheader_buffer(char *obuffer, int olen, const char *ibuffer, int ilen);
-
-
-#ifdef __cplusplus
-}
-#endif
-
-//
-//
-////   end header file   /////////////////////////////////////////////////////
-#endif // STBI_INCLUDE_STB_IMAGE_H
-
-#ifdef STB_IMAGE_IMPLEMENTATION
-
-#if defined(STBI_ONLY_JPEG) || defined(STBI_ONLY_PNG) || defined(STBI_ONLY_BMP) \
-  || defined(STBI_ONLY_TGA) || defined(STBI_ONLY_GIF) || defined(STBI_ONLY_PSD) \
-  || defined(STBI_ONLY_HDR) || defined(STBI_ONLY_PIC) || defined(STBI_ONLY_PNM) \
-  || defined(STBI_ONLY_ZLIB)
-   #ifndef STBI_ONLY_JPEG
-   #define STBI_NO_JPEG
-   #endif
-   #ifndef STBI_ONLY_PNG
-   #define STBI_NO_PNG
-   #endif
-   #ifndef STBI_ONLY_BMP
-   #define STBI_NO_BMP
-   #endif
-   #ifndef STBI_ONLY_PSD
-   #define STBI_NO_PSD
-   #endif
-   #ifndef STBI_ONLY_TGA
-   #define STBI_NO_TGA
-   #endif
-   #ifndef STBI_ONLY_GIF
-   #define STBI_NO_GIF
-   #endif
-   #ifndef STBI_ONLY_HDR
-   #define STBI_NO_HDR
-   #endif
-   #ifndef STBI_ONLY_PIC
-   #define STBI_NO_PIC
-   #endif
-   #ifndef STBI_ONLY_PNM
-   #define STBI_NO_PNM
-   #endif
-#endif
-
-#if defined(STBI_NO_PNG) && !defined(STBI_SUPPORT_ZLIB) && !defined(STBI_NO_ZLIB)
-#define STBI_NO_ZLIB
-#endif
-
-
-#include <stdarg.h>
-#include <stddef.h> // ptrdiff_t on osx
-#include <stdlib.h>
-#include <string.h>
-#include <limits.h>
-
-#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR)
-  // ldexp, pow
-#endif
-
-#ifndef STBI_NO_STDIO
-#include <stdio.h>
-#endif
-
-#ifndef STBI_ASSERT
-#include <assert.h>
-#define STBI_ASSERT(x) assert(x)
-#endif
-
-#ifdef __cplusplus
-#define STBI_EXTERN extern "C"
-#else
-#define STBI_EXTERN extern
-#endif
-
-
-#ifndef _MSC_VER
-   #ifdef __cplusplus
-   #define stbi_inline inline
-   #else
-   #define stbi_inline
-   #endif
-#else
-   #define stbi_inline __forceinline
-#endif
-
-#ifndef STBI_NO_THREAD_LOCALS
-   #if defined(__cplusplus) &&  __cplusplus >= 201103L
-      #define STBI_THREAD_LOCAL       thread_local
-   #elif defined(__GNUC__) && __GNUC__ < 5
-      #define STBI_THREAD_LOCAL       __thread
-   #elif defined(_MSC_VER)
-      #define STBI_THREAD_LOCAL       __declspec(thread)
-   #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 201112L && !defined(__STDC_NO_THREADS__)
-      #define STBI_THREAD_LOCAL       _Thread_local
-   #endif
-
-   #ifndef STBI_THREAD_LOCAL
-      #if defined(__GNUC__)
-        #define STBI_THREAD_LOCAL       __thread
-      #endif
-   #endif
-#endif
-
-#if defined(_MSC_VER) || defined(__SYMBIAN32__)
-typedef unsigned short stbi__uint16;
-typedef   signed short stbi__int16;
-typedef unsigned int   stbi__uint32;
-typedef   signed int   stbi__int32;
-#else
-#include <stdint.h>
-typedef uint16_t stbi__uint16;
-typedef int16_t  stbi__int16;
-typedef uint32_t stbi__uint32;
-typedef int32_t  stbi__int32;
-#endif
-
-// should produce compiler error if size is wrong
-typedef unsigned char validate_uint32[sizeof(stbi__uint32)==4 ? 1 : -1];
-
-#ifdef _MSC_VER
-#define STBI_NOTUSED(v)  (void)(v)
-#else
-#define STBI_NOTUSED(v)  (void)sizeof(v)
-#endif
-
-#ifdef _MSC_VER
-#define STBI_HAS_LROTL
-#endif
-
-#ifdef STBI_HAS_LROTL
-   #define stbi_lrot(x,y)  _lrotl(x,y)
-#else
-   #define stbi_lrot(x,y)  (((x) << (y)) | ((x) >> (-(y) & 31)))
-#endif
-
-#if defined(STBI_MALLOC) && defined(STBI_FREE) && (defined(STBI_REALLOC) || defined(STBI_REALLOC_SIZED))
-// ok
-#elif !defined(STBI_MALLOC) && !defined(STBI_FREE) && !defined(STBI_REALLOC) && !defined(STBI_REALLOC_SIZED)
-// ok
-#else
-#error "Must define all or none of STBI_MALLOC, STBI_FREE, and STBI_REALLOC (or STBI_REALLOC_SIZED)."
-#endif
-
-#ifndef STBI_MALLOC
-#define STBI_MALLOC(sz)           malloc(sz)
-#define STBI_REALLOC(p,newsz)     realloc(p,newsz)
-#define STBI_FREE(p)              free(p)
-#endif
-
-#ifndef STBI_REALLOC_SIZED
-#define STBI_REALLOC_SIZED(p,oldsz,newsz) STBI_REALLOC(p,newsz)
-#endif
-
-// x86/x64 detection
-#if defined(__x86_64__) || defined(_M_X64)
-#define STBI__X64_TARGET
-#elif defined(__i386) || defined(_M_IX86)
-#define STBI__X86_TARGET
-#endif
-
-#if defined(__GNUC__) && defined(STBI__X86_TARGET) && !defined(__SSE2__) && !defined(STBI_NO_SIMD)
-// gcc doesn't support sse2 intrinsics unless you compile with -msse2,
-// which in turn means it gets to use SSE2 everywhere. This is unfortunate,
-// but previous attempts to provide the SSE2 functions with runtime
-// detection caused numerous issues. The way architecture extensions are
-// exposed in GCC/Clang is, sadly, not really suited for one-file libs.
-// New behavior: if compiled with -msse2, we use SSE2 without any
-// detection; if not, we don't use it at all.
-#define STBI_NO_SIMD
-#endif
-
-#if defined(__MINGW32__) && defined(STBI__X86_TARGET) && !defined(STBI_MINGW_ENABLE_SSE2) && !defined(STBI_NO_SIMD)
-// Note that __MINGW32__ doesn't actually mean 32-bit, so we have to avoid STBI__X64_TARGET
-//
-// 32-bit MinGW wants ESP to be 16-byte aligned, but this is not in the
-// Windows ABI and VC++ as well as Windows DLLs don't maintain that invariant.
-// As a result, enabling SSE2 on 32-bit MinGW is dangerous when not
-// simultaneously enabling "-mstackrealign".
-//
-// See https://github.com/nothings/stb/issues/81 for more information.
-//
-// So default to no SSE2 on 32-bit MinGW. If you've read this far and added
-// -mstackrealign to your build settings, feel free to #define STBI_MINGW_ENABLE_SSE2.
-#define STBI_NO_SIMD
-#endif
-
-#if !defined(STBI_NO_SIMD) && (defined(STBI__X86_TARGET) || defined(STBI__X64_TARGET))
-#define STBI_SSE2
-#include <emmintrin.h>
-
-#ifdef _MSC_VER
-
-#if _MSC_VER >= 1400  // not VC6
-#include <intrin.h> // __cpuid
-static int stbi__cpuid3(void)
-{
-   int info[4];
-   __cpuid(info,1);
-   return info[3];
-}
-#else
-static int stbi__cpuid3(void)
-{
-   int res;
-   __asm {
-      mov  eax,1
-      cpuid
-      mov  res,edx
-   }
-   return res;
-}
-#endif
-
-#define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name
-
-#if !defined(STBI_NO_JPEG) && defined(STBI_SSE2)
-static int stbi__sse2_available(void)
-{
-   int info3 = stbi__cpuid3();
-   return ((info3 >> 26) & 1) != 0;
-}
-#endif
-
-#else // assume GCC-style if not VC++
-#define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16)))
-
-#if !defined(STBI_NO_JPEG) && defined(STBI_SSE2)
-static int stbi__sse2_available(void)
-{
-   // If we're even attempting to compile this on GCC/Clang, that means
-   // -msse2 is on, which means the compiler is allowed to use SSE2
-   // instructions at will, and so are we.
-   return 1;
-}
-#endif
-
-#endif
-#endif
-
-// ARM NEON
-#if defined(STBI_NO_SIMD) && defined(STBI_NEON)
-#undef STBI_NEON
-#endif
-
-#ifdef STBI_NEON
-#include <arm_neon.h>
-#ifdef _MSC_VER
-#define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name
-#else
-#define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16)))
-#endif
-#endif
-
-// RISC-V VECTOR
-#if defined(STBI_NO_SIMD) && defined(STBI_RVV)
-#undef STBI_RVV
-#endif
-
-#ifdef STBI_RVV
-#include <riscv_vector.h>
-#ifdef _MSC_VER
-#define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name
-#else
-#define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16)))
-#endif
-#endif
-
-#ifndef STBI_SIMD_ALIGN
-#define STBI_SIMD_ALIGN(type, name) type name
-#endif
-
-#ifndef STBI_MAX_DIMENSIONS
-#define STBI_MAX_DIMENSIONS (1 << 24)
-#endif
-
-///////////////////////////////////////////////
-//
-//  stbi__context struct and start_xxx functions
-
-// stbi__context structure is our basic context used by all images, so it
-// contains all the IO context, plus some basic image information
-typedef struct
-{
-   stbi__uint32 img_x, img_y;
-   int img_n, img_out_n;
-
-   stbi_io_callbacks io;
-   void *io_user_data;
-
-   int read_from_callbacks;
-   int buflen;
-   stbi_uc buffer_start[128];
-   int callback_already_read;
-
-   stbi_uc *img_buffer, *img_buffer_end;
-   stbi_uc *img_buffer_original, *img_buffer_original_end;
-} stbi__context;
-
-
-static void stbi__refill_buffer(stbi__context *s);
-
-// initialize a memory-decode context
-static void stbi__start_mem(stbi__context *s, stbi_uc const *buffer, int len)
-{
-   s->io.read = NULL;
-   s->read_from_callbacks = 0;
-   s->callback_already_read = 0;
-   s->img_buffer = s->img_buffer_original = (stbi_uc *) buffer;
-   s->img_buffer_end = s->img_buffer_original_end = (stbi_uc *) buffer+len;
-}
-
-// initialize a callback-based context
-static void stbi__start_callbacks(stbi__context *s, stbi_io_callbacks *c, void *user)
-{
-   s->io = *c;
-   s->io_user_data = user;
-   s->buflen = sizeof(s->buffer_start);
-   s->read_from_callbacks = 1;
-   s->callback_already_read = 0;
-   s->img_buffer = s->img_buffer_original = s->buffer_start;
-   stbi__refill_buffer(s);
-   s->img_buffer_original_end = s->img_buffer_end;
-}
-
-#ifndef STBI_NO_STDIO
-
-static int stbi__stdio_read(void *user, char *data, int size)
-{
-   return (int) fread(data,1,size,(FILE*) user);
-}
-
-static void stbi__stdio_skip(void *user, int n)
-{
-   int ch;
-   fseek((FILE*) user, n, SEEK_CUR);
-   ch = fgetc((FILE*) user);  /* have to read a byte to reset feof()'s flag */
-   if (ch != EOF) {
-      ungetc(ch, (FILE *) user);  /* push byte back onto stream if valid. */
-   }
-}
-
-static int stbi__stdio_eof(void *user)
-{
-   return feof((FILE*) user) || ferror((FILE *) user);
-}
-
-static stbi_io_callbacks stbi__stdio_callbacks =
-{
-   stbi__stdio_read,
-   stbi__stdio_skip,
-   stbi__stdio_eof,
-};
-
-static void stbi__start_file(stbi__context *s, FILE *f)
-{
-   stbi__start_callbacks(s, &stbi__stdio_callbacks, (void *) f);
-}
-
-//static void stop_file(stbi__context *s) { }
-
-#endif // !STBI_NO_STDIO
-
-static void stbi__rewind(stbi__context *s)
-{
-   // conceptually rewind SHOULD rewind to the beginning of the stream,
-   // but we just rewind to the beginning of the initial buffer, because
-   // we only use it after doing 'test', which only ever looks at at most 92 bytes
-   s->img_buffer = s->img_buffer_original;
-   s->img_buffer_end = s->img_buffer_original_end;
-}
-
-enum
-{
-   STBI_ORDER_RGB,
-   STBI_ORDER_BGR
-};
-
-typedef struct
-{
-   int bits_per_channel;
-   int num_channels;
-   int channel_order;
-} stbi__result_info;
-
-#ifndef STBI_NO_JPEG
-static int      stbi__jpeg_test(stbi__context *s);
-static void    *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
-static int      stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp);
-#endif
-
-#ifndef STBI_NO_PNG
-static int      stbi__png_test(stbi__context *s);
-static void    *stbi__png_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
-static int      stbi__png_info(stbi__context *s, int *x, int *y, int *comp);
-static int      stbi__png_is16(stbi__context *s);
-#endif
-
-#ifndef STBI_NO_BMP
-static int      stbi__bmp_test(stbi__context *s);
-static void    *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
-static int      stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp);
-#endif
-
-#ifndef STBI_NO_TGA
-static int      stbi__tga_test(stbi__context *s);
-static void    *stbi__tga_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
-static int      stbi__tga_info(stbi__context *s, int *x, int *y, int *comp);
-#endif
-
-#ifndef STBI_NO_PSD
-static int      stbi__psd_test(stbi__context *s);
-static void    *stbi__psd_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc);
-static int      stbi__psd_info(stbi__context *s, int *x, int *y, int *comp);
-static int      stbi__psd_is16(stbi__context *s);
-#endif
-
-#ifndef STBI_NO_HDR
-static int      stbi__hdr_test(stbi__context *s);
-static float   *stbi__hdr_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
-static int      stbi__hdr_info(stbi__context *s, int *x, int *y, int *comp);
-#endif
-
-#ifndef STBI_NO_PIC
-static int      stbi__pic_test(stbi__context *s);
-static void    *stbi__pic_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
-static int      stbi__pic_info(stbi__context *s, int *x, int *y, int *comp);
-#endif
-
-#ifndef STBI_NO_GIF
-static int      stbi__gif_test(stbi__context *s);
-static void    *stbi__gif_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
-static void    *stbi__load_gif_main(stbi__context *s, int **delays, int *x, int *y, int *z, int *comp, int req_comp);
-static int      stbi__gif_info(stbi__context *s, int *x, int *y, int *comp);
-#endif
-
-#ifndef STBI_NO_PNM
-static int      stbi__pnm_test(stbi__context *s);
-static void    *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
-static int      stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp);
-static int      stbi__pnm_is16(stbi__context *s);
-#endif
-
-static
-#ifdef STBI_THREAD_LOCAL
-STBI_THREAD_LOCAL
-#endif
-const char *stbi__g_failure_reason;
-
-STBIDEF const char *stbi_failure_reason(void)
-{
-   return stbi__g_failure_reason;
-}
-
-#ifndef STBI_NO_FAILURE_STRINGS
-static int stbi__err(const char *str)
-{
-   stbi__g_failure_reason = str;
-   return 0;
-}
-#endif
-
-static void *stbi__malloc(size_t size)
-{
-    return STBI_MALLOC(size);
-}
-
-// stb_image uses ints pervasively, including for offset calculations.
-// therefore the largest decoded image size we can support with the
-// current code, even on 64-bit targets, is INT_MAX. this is not a
-// significant limitation for the intended use case.
-//
-// we do, however, need to make sure our size calculations don't
-// overflow. hence a few helper functions for size calculations that
-// multiply integers together, making sure that they're non-negative
-// and no overflow occurs.
-
-// return 1 if the sum is valid, 0 on overflow.
-// negative terms are considered invalid.
-static int stbi__addsizes_valid(int a, int b)
-{
-   if (b < 0) return 0;
-   // now 0 <= b <= INT_MAX, hence also
-   // 0 <= INT_MAX - b <= INTMAX.
-   // And "a + b <= INT_MAX" (which might overflow) is the
-   // same as a <= INT_MAX - b (no overflow)
-   return a <= INT_MAX - b;
-}
-
-// returns 1 if the product is valid, 0 on overflow.
-// negative factors are considered invalid.
-static int stbi__mul2sizes_valid(int a, int b)
-{
-   if (a < 0 || b < 0) return 0;
-   if (b == 0) return 1; // mul-by-0 is always safe
-   // portable way to check for no overflows in a*b
-   return a <= INT_MAX/b;
-}
-
-#if !defined(STBI_NO_JPEG) || !defined(STBI_NO_PNG) || !defined(STBI_NO_TGA) || !defined(STBI_NO_HDR)
-// returns 1 if "a*b + add" has no negative terms/factors and doesn't overflow
-static int stbi__mad2sizes_valid(int a, int b, int add)
-{
-   return stbi__mul2sizes_valid(a, b) && stbi__addsizes_valid(a*b, add);
-}
-#endif
-
-// returns 1 if "a*b*c + add" has no negative terms/factors and doesn't overflow
-static int stbi__mad3sizes_valid(int a, int b, int c, int add)
-{
-   return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a*b, c) &&
-      stbi__addsizes_valid(a*b*c, add);
-}
-
-// returns 1 if "a*b*c*d + add" has no negative terms/factors and doesn't overflow
-#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM)
-static int stbi__mad4sizes_valid(int a, int b, int c, int d, int add)
-{
-   return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a*b, c) &&
-      stbi__mul2sizes_valid(a*b*c, d) && stbi__addsizes_valid(a*b*c*d, add);
-}
-#endif
-
-#if !defined(STBI_NO_JPEG) || !defined(STBI_NO_PNG) || !defined(STBI_NO_TGA) || !defined(STBI_NO_HDR)
-// mallocs with size overflow checking
-static void *stbi__malloc_mad2(int a, int b, int add)
-{
-   if (!stbi__mad2sizes_valid(a, b, add)) return NULL;
-   return stbi__malloc(a*b + add);
-}
-#endif
-
-static void *stbi__malloc_mad3(int a, int b, int c, int add)
-{
-   if (!stbi__mad3sizes_valid(a, b, c, add)) return NULL;
-   return stbi__malloc(a*b*c + add);
-}
-
-#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM)
-static void *stbi__malloc_mad4(int a, int b, int c, int d, int add)
-{
-   if (!stbi__mad4sizes_valid(a, b, c, d, add)) return NULL;
-   return stbi__malloc(a*b*c*d + add);
-}
-#endif
-
-// returns 1 if the sum of two signed ints is valid (between -2^31 and 2^31-1 inclusive), 0 on overflow.
-static int stbi__addints_valid(int a, int b)
-{
-   if ((a >= 0) != (b >= 0)) return 1; // a and b have different signs, so no overflow
-   if (a < 0 && b < 0) return a >= INT_MIN - b; // same as a + b >= INT_MIN; INT_MIN - b cannot overflow since b < 0.
-   return a <= INT_MAX - b;
-}
-
-// returns 1 if the product of two signed shorts is valid, 0 on overflow.
-static int stbi__mul2shorts_valid(short a, short b)
-{
-   if (b == 0 || b == -1) return 1; // multiplication by 0 is always 0; check for -1 so SHRT_MIN/b doesn't overflow
-   if ((a >= 0) == (b >= 0)) return a <= SHRT_MAX/b; // product is positive, so similar to mul2sizes_valid
-   if (b < 0) return a <= SHRT_MIN / b; // same as a * b >= SHRT_MIN
-   return a >= SHRT_MIN / b;
-}
-
-// stbi__err - error
-// stbi__errpf - error returning pointer to float
-// stbi__errpuc - error returning pointer to unsigned char
-
-#ifdef STBI_NO_FAILURE_STRINGS
-   #define stbi__err(x,y)  0
-#elif defined(STBI_FAILURE_USERMSG)
-   #define stbi__err(x,y)  stbi__err(y)
-#else
-   #define stbi__err(x,y)  stbi__err(x)
-#endif
-
-#define stbi__errpf(x,y)   ((float *)(size_t) (stbi__err(x,y)?NULL:NULL))
-#define stbi__errpuc(x,y)  ((unsigned char *)(size_t) (stbi__err(x,y)?NULL:NULL))
-
-STBIDEF void stbi_image_free(void *retval_from_stbi_load)
-{
-   STBI_FREE(retval_from_stbi_load);
-}
-
-#ifndef STBI_NO_LINEAR
-static float   *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp);
-#endif
-
-#ifndef STBI_NO_HDR
-static stbi_uc *stbi__hdr_to_ldr(float   *data, int x, int y, int comp);
-#endif
-
-static int stbi__vertically_flip_on_load_global = 0;
-
-STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip)
-{
-   stbi__vertically_flip_on_load_global = flag_true_if_should_flip;
-}
-
-#ifndef STBI_THREAD_LOCAL
-#define stbi__vertically_flip_on_load  stbi__vertically_flip_on_load_global
-#else
-static STBI_THREAD_LOCAL int stbi__vertically_flip_on_load_local, stbi__vertically_flip_on_load_set;
-
-STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_flip)
-{
-   stbi__vertically_flip_on_load_local = flag_true_if_should_flip;
-   stbi__vertically_flip_on_load_set = 1;
-}
-
-#define stbi__vertically_flip_on_load  (stbi__vertically_flip_on_load_set       \
-                                         ? stbi__vertically_flip_on_load_local  \
-                                         : stbi__vertically_flip_on_load_global)
-#endif // STBI_THREAD_LOCAL
-
-static void *stbi__load_main(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc)
-{
-   memset(ri, 0, sizeof(*ri)); // make sure it's initialized if we add new fields
-   ri->bits_per_channel = 8; // default is 8 so most paths don't have to be changed
-   ri->channel_order = STBI_ORDER_RGB; // all current input & output are this, but this is here so we can add BGR order
-   ri->num_channels = 0;
-
-   // test the formats with a very explicit header first (at least a FOURCC
-   // or distinctive magic number first)
-   #ifndef STBI_NO_PNG
-   if (stbi__png_test(s))  return stbi__png_load(s,x,y,comp,req_comp, ri);
-   #endif
-   #ifndef STBI_NO_BMP
-   if (stbi__bmp_test(s))  return stbi__bmp_load(s,x,y,comp,req_comp, ri);
-   #endif
-   #ifndef STBI_NO_GIF
-   if (stbi__gif_test(s))  return stbi__gif_load(s,x,y,comp,req_comp, ri);
-   #endif
-   #ifndef STBI_NO_PSD
-   if (stbi__psd_test(s))  return stbi__psd_load(s,x,y,comp,req_comp, ri, bpc);
-   #else
-   STBI_NOTUSED(bpc);
-   #endif
-   #ifndef STBI_NO_PIC
-   if (stbi__pic_test(s))  return stbi__pic_load(s,x,y,comp,req_comp, ri);
-   #endif
-
-   // then the formats that can end up attempting to load with just 1 or 2
-   // bytes matching expectations; these are prone to false positives, so
-   // try them later
-   #ifndef STBI_NO_JPEG
-   if (stbi__jpeg_test(s)) return stbi__jpeg_load(s,x,y,comp,req_comp, ri);
-   #endif
-   #ifndef STBI_NO_PNM
-   if (stbi__pnm_test(s))  return stbi__pnm_load(s,x,y,comp,req_comp, ri);
-   #endif
-
-   #ifndef STBI_NO_HDR
-   if (stbi__hdr_test(s)) {
-      float *hdr = stbi__hdr_load(s, x,y,comp,req_comp, ri);
-      return stbi__hdr_to_ldr(hdr, *x, *y, req_comp ? req_comp : *comp);
-   }
-   #endif
-
-   #ifndef STBI_NO_TGA
-   // test tga last because it's a crappy test!
-   if (stbi__tga_test(s))
-      return stbi__tga_load(s,x,y,comp,req_comp, ri);
-   #endif
-
-   return stbi__errpuc("unknown image type", "Image not of any known type, or corrupt");
-}
-
-static stbi_uc *stbi__convert_16_to_8(stbi__uint16 *orig, int w, int h, int channels)
-{
-   int i;
-   int img_len = w * h * channels;
-   stbi_uc *reduced;
-
-   reduced = (stbi_uc *) stbi__malloc(img_len);
-   if (reduced == NULL) return stbi__errpuc("outofmem", "Out of memory");
-
-   for (i = 0; i < img_len; ++i)
-      reduced[i] = (stbi_uc)((orig[i] >> 8) & 0xFF); // top half of each byte is sufficient approx of 16->8 bit scaling
-
-   STBI_FREE(orig);
-   return reduced;
-}
-
-static stbi__uint16 *stbi__convert_8_to_16(stbi_uc *orig, int w, int h, int channels)
-{
-   int i;
-   int img_len = w * h * channels;
-   stbi__uint16 *enlarged;
-
-   enlarged = (stbi__uint16 *) stbi__malloc(img_len*2);
-   if (enlarged == NULL) return (stbi__uint16 *) stbi__errpuc("outofmem", "Out of memory");
-
-   for (i = 0; i < img_len; ++i)
-      enlarged[i] = (stbi__uint16)((orig[i] << 8) + orig[i]); // replicate to high and low byte, maps 0->0, 255->0xffff
-
-   STBI_FREE(orig);
-   return enlarged;
-}
-
-static void stbi__vertical_flip(void *image, int w, int h, int bytes_per_pixel)
-{
-   int row;
-   size_t bytes_per_row = (size_t)w * bytes_per_pixel;
-   stbi_uc temp[2048];
-   stbi_uc *bytes = (stbi_uc *)image;
-
-   for (row = 0; row < (h>>1); row++) {
-      stbi_uc *row0 = bytes + row*bytes_per_row;
-      stbi_uc *row1 = bytes + (h - row - 1)*bytes_per_row;
-      // swap row0 with row1
-      size_t bytes_left = bytes_per_row;
-      while (bytes_left) {
-         size_t bytes_copy = (bytes_left < sizeof(temp)) ? bytes_left : sizeof(temp);
-         memcpy(temp, row0, bytes_copy);
-         memcpy(row0, row1, bytes_copy);
-         memcpy(row1, temp, bytes_copy);
-         row0 += bytes_copy;
-         row1 += bytes_copy;
-         bytes_left -= bytes_copy;
-      }
-   }
-}
-
-#ifndef STBI_NO_GIF
-static void stbi__vertical_flip_slices(void *image, int w, int h, int z, int bytes_per_pixel)
-{
-   int slice;
-   int slice_size = w * h * bytes_per_pixel;
-
-   stbi_uc *bytes = (stbi_uc *)image;
-   for (slice = 0; slice < z; ++slice) {
-      stbi__vertical_flip(bytes, w, h, bytes_per_pixel);
-      bytes += slice_size;
-   }
-}
-#endif
-
-static unsigned char *stbi__load_and_postprocess_8bit(stbi__context *s, int *x, int *y, int *comp, int req_comp)
-{
-   stbi__result_info ri;
-   void *result = stbi__load_main(s, x, y, comp, req_comp, &ri, 8);
-
-   if (result == NULL)
-      return NULL;
-
-   // it is the responsibility of the loaders to make sure we get either 8 or 16 bit.
-   STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16);
-
-   if (ri.bits_per_channel != 8) {
-      result = stbi__convert_16_to_8((stbi__uint16 *) result, *x, *y, req_comp == 0 ? *comp : req_comp);
-      ri.bits_per_channel = 8;
-   }
-
-   // @TODO: move stbi__convert_format to here
-
-   if (stbi__vertically_flip_on_load) {
-      int channels = req_comp ? req_comp : *comp;
-      stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi_uc));
-   }
-
-   return (unsigned char *) result;
-}
-
-static stbi__uint16 *stbi__load_and_postprocess_16bit(stbi__context *s, int *x, int *y, int *comp, int req_comp)
-{
-   stbi__result_info ri;
-   void *result = stbi__load_main(s, x, y, comp, req_comp, &ri, 16);
-
-   if (result == NULL)
-      return NULL;
-
-   // it is the responsibility of the loaders to make sure we get either 8 or 16 bit.
-   STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16);
-
-   if (ri.bits_per_channel != 16) {
-      result = stbi__convert_8_to_16((stbi_uc *) result, *x, *y, req_comp == 0 ? *comp : req_comp);
-      ri.bits_per_channel = 16;
-   }
-
-   // @TODO: move stbi__convert_format16 to here
-   // @TODO: special case RGB-to-Y (and RGBA-to-YA) for 8-bit-to-16-bit case to keep more precision
-
-   if (stbi__vertically_flip_on_load) {
-      int channels = req_comp ? req_comp : *comp;
-      stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi__uint16));
-   }
-
-   return (stbi__uint16 *) result;
-}
-
-#if !defined(STBI_NO_HDR) && !defined(STBI_NO_LINEAR)
-static void stbi__float_postprocess(float *result, int *x, int *y, int *comp, int req_comp)
-{
-   if (stbi__vertically_flip_on_load && result != NULL) {
-      int channels = req_comp ? req_comp : *comp;
-      stbi__vertical_flip(result, *x, *y, channels * sizeof(float));
-   }
-}
-#endif
-
-#ifndef STBI_NO_STDIO
-
-#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
-STBI_EXTERN __declspec(dllimport) int __stdcall MultiByteToWideChar(unsigned int cp, unsigned long flags, const char *str, int cbmb, wchar_t *widestr, int cchwide);
-STBI_EXTERN __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int cp, unsigned long flags, const wchar_t *widestr, int cchwide, char *str, int cbmb, const char *defchar, int *used_default);
-#endif
-
-#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
-STBIDEF int stbi_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input)
-{
-	return WideCharToMultiByte(65001 /* UTF8 */, 0, input, -1, buffer, (int) bufferlen, NULL, NULL);
-}
-#endif
-
-static FILE *stbi__fopen(char const *filename, char const *mode)
-{
-   FILE *f;
-#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
-   wchar_t wMode[64];
-   wchar_t wFilename[1024];
-	if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename)/sizeof(*wFilename)))
-      return 0;
-
-	if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode)/sizeof(*wMode)))
-      return 0;
-
-#if defined(_MSC_VER) && _MSC_VER >= 1400
-	if (0 != _wfopen_s(&f, wFilename, wMode))
-		f = 0;
-#else
-   f = _wfopen(wFilename, wMode);
-#endif
-
-#elif defined(_MSC_VER) && _MSC_VER >= 1400
-   if (0 != fopen_s(&f, filename, mode))
-      f=0;
-#else
-   f = fopen(filename, mode);
-#endif
-   return f;
-}
-
-
-STBIDEF stbi_uc *stbi_load(char const *filename, int *x, int *y, int *comp, int req_comp)
-{
-   FILE *f = stbi__fopen(filename, "rb");
-   unsigned char *result;
-   if (!f) return stbi__errpuc("can't fopen", "Unable to open file");
-   result = stbi_load_from_file(f,x,y,comp,req_comp);
-   fclose(f);
-   return result;
-}
-
-STBIDEF stbi_uc *stbi_load_from_file(FILE *f, int *x, int *y, int *comp, int req_comp)
-{
-   unsigned char *result;
-   stbi__context s;
-   stbi__start_file(&s,f);
-   result = stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp);
-   if (result) {
-      // need to 'unget' all the characters in the IO buffer
-      fseek(f, - (int) (s.img_buffer_end - s.img_buffer), SEEK_CUR);
-   }
-   return result;
-}
-
-STBIDEF stbi__uint16 *stbi_load_from_file_16(FILE *f, int *x, int *y, int *comp, int req_comp)
-{
-   stbi__uint16 *result;
-   stbi__context s;
-   stbi__start_file(&s,f);
-   result = stbi__load_and_postprocess_16bit(&s,x,y,comp,req_comp);
-   if (result) {
-      // need to 'unget' all the characters in the IO buffer
-      fseek(f, - (int) (s.img_buffer_end - s.img_buffer), SEEK_CUR);
-   }
-   return result;
-}
-
-STBIDEF stbi_us *stbi_load_16(char const *filename, int *x, int *y, int *comp, int req_comp)
-{
-   FILE *f = stbi__fopen(filename, "rb");
-   stbi__uint16 *result;
-   if (!f) return (stbi_us *) stbi__errpuc("can't fopen", "Unable to open file");
-   result = stbi_load_from_file_16(f,x,y,comp,req_comp);
-   fclose(f);
-   return result;
-}
-
-
-#endif //!STBI_NO_STDIO
-
-STBIDEF stbi_us *stbi_load_16_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels)
-{
-   stbi__context s;
-   stbi__start_mem(&s,buffer,len);
-   return stbi__load_and_postprocess_16bit(&s,x,y,channels_in_file,desired_channels);
-}
-
-STBIDEF stbi_us *stbi_load_16_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels)
-{
-   stbi__context s;
-   stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user);
-   return stbi__load_and_postprocess_16bit(&s,x,y,channels_in_file,desired_channels);
-}
-
-STBIDEF stbi_uc *stbi_load_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp)
-{
-   stbi__context s;
-   stbi__start_mem(&s,buffer,len);
-   return stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp);
-}
-
-STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp)
-{
-   stbi__context s;
-   stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user);
-   return stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp);
-}
-
-#ifndef STBI_NO_GIF
-STBIDEF stbi_uc *stbi_load_gif_from_memory(stbi_uc const *buffer, int len, int **delays, int *x, int *y, int *z, int *comp, int req_comp)
-{
-   unsigned char *result;
-   stbi__context s;
-   stbi__start_mem(&s,buffer,len);
-
-   result = (unsigned char*) stbi__load_gif_main(&s, delays, x, y, z, comp, req_comp);
-   if (stbi__vertically_flip_on_load) {
-      stbi__vertical_flip_slices( result, *x, *y, *z, *comp );
-   }
-
-   return result;
-}
-#endif
-
-#ifndef STBI_NO_LINEAR
-static float *stbi__loadf_main(stbi__context *s, int *x, int *y, int *comp, int req_comp)
-{
-   unsigned char *data;
-   #ifndef STBI_NO_HDR
-   if (stbi__hdr_test(s)) {
-      stbi__result_info ri;
-      float *hdr_data = stbi__hdr_load(s,x,y,comp,req_comp, &ri);
-      if (hdr_data)
-         stbi__float_postprocess(hdr_data,x,y,comp,req_comp);
-      return hdr_data;
-   }
-   #endif
-   data = stbi__load_and_postprocess_8bit(s, x, y, comp, req_comp);
-   if (data)
-      return stbi__ldr_to_hdr(data, *x, *y, req_comp ? req_comp : *comp);
-   return stbi__errpf("unknown image type", "Image not of any known type, or corrupt");
-}
-
-STBIDEF float *stbi_loadf_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp)
-{
-   stbi__context s;
-   stbi__start_mem(&s,buffer,len);
-   return stbi__loadf_main(&s,x,y,comp,req_comp);
-}
-
-STBIDEF float *stbi_loadf_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp)
-{
-   stbi__context s;
-   stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user);
-   return stbi__loadf_main(&s,x,y,comp,req_comp);
-}
-
-#ifndef STBI_NO_STDIO
-STBIDEF float *stbi_loadf(char const *filename, int *x, int *y, int *comp, int req_comp)
-{
-   float *result;
-   FILE *f = stbi__fopen(filename, "rb");
-   if (!f) return stbi__errpf("can't fopen", "Unable to open file");
-   result = stbi_loadf_from_file(f,x,y,comp,req_comp);
-   fclose(f);
-   return result;
-}
-
-STBIDEF float *stbi_loadf_from_file(FILE *f, int *x, int *y, int *comp, int req_comp)
-{
-   stbi__context s;
-   stbi__start_file(&s,f);
-   return stbi__loadf_main(&s,x,y,comp,req_comp);
-}
-#endif // !STBI_NO_STDIO
-
-#endif // !STBI_NO_LINEAR
-
-// these is-hdr-or-not is defined independent of whether STBI_NO_LINEAR is
-// defined, for API simplicity; if STBI_NO_LINEAR is defined, it always
-// reports false!
-
-STBIDEF int stbi_is_hdr_from_memory(stbi_uc const *buffer, int len)
-{
-   #ifndef STBI_NO_HDR
-   stbi__context s;
-   stbi__start_mem(&s,buffer,len);
-   return stbi__hdr_test(&s);
-   #else
-   STBI_NOTUSED(buffer);
-   STBI_NOTUSED(len);
-   return 0;
-   #endif
-}
-
-#ifndef STBI_NO_STDIO
-STBIDEF int      stbi_is_hdr          (char const *filename)
-{
-   FILE *f = stbi__fopen(filename, "rb");
-   int result=0;
-   if (f) {
-      result = stbi_is_hdr_from_file(f);
-      fclose(f);
-   }
-   return result;
-}
-
-STBIDEF int stbi_is_hdr_from_file(FILE *f)
-{
-   #ifndef STBI_NO_HDR
-   long pos = ftell(f);
-   int res;
-   stbi__context s;
-   stbi__start_file(&s,f);
-   res = stbi__hdr_test(&s);
-   fseek(f, pos, SEEK_SET);
-   return res;
-   #else
-   STBI_NOTUSED(f);
-   return 0;
-   #endif
-}
-#endif // !STBI_NO_STDIO
-
-STBIDEF int      stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user)
-{
-   #ifndef STBI_NO_HDR
-   stbi__context s;
-   stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user);
-   return stbi__hdr_test(&s);
-   #else
-   STBI_NOTUSED(clbk);
-   STBI_NOTUSED(user);
-   return 0;
-   #endif
-}
-
-#ifndef STBI_NO_LINEAR
-static float stbi__l2h_gamma=2.2f, stbi__l2h_scale=1.0f;
-
-STBIDEF void   stbi_ldr_to_hdr_gamma(float gamma) { stbi__l2h_gamma = gamma; }
-STBIDEF void   stbi_ldr_to_hdr_scale(float scale) { stbi__l2h_scale = scale; }
-#endif
-
-static float stbi__h2l_gamma_i=1.0f/2.2f, stbi__h2l_scale_i=1.0f;
-
-STBIDEF void   stbi_hdr_to_ldr_gamma(float gamma) { stbi__h2l_gamma_i = 1/gamma; }
-STBIDEF void   stbi_hdr_to_ldr_scale(float scale) { stbi__h2l_scale_i = 1/scale; }
-
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// Common code used by all image loaders
-//
-
-enum
-{
-   STBI__SCAN_load=0,
-   STBI__SCAN_type,
-   STBI__SCAN_header
-};
-
-static void stbi__refill_buffer(stbi__context *s)
-{
-   int n = (s->io.read)(s->io_user_data,(char*)s->buffer_start,s->buflen);
-   s->callback_already_read += (int) (s->img_buffer - s->img_buffer_original);
-   if (n == 0) {
-      // at end of file, treat same as if from memory, but need to handle case
-      // where s->img_buffer isn't pointing to safe memory, e.g. 0-byte file
-      s->read_from_callbacks = 0;
-      s->img_buffer = s->buffer_start;
-      s->img_buffer_end = s->buffer_start+1;
-      *s->img_buffer = 0;
-   } else {
-      s->img_buffer = s->buffer_start;
-      s->img_buffer_end = s->buffer_start + n;
-   }
-}
-
-stbi_inline static stbi_uc stbi__get8(stbi__context *s)
-{
-   if (s->img_buffer < s->img_buffer_end)
-      return *s->img_buffer++;
-   if (s->read_from_callbacks) {
-      stbi__refill_buffer(s);
-      return *s->img_buffer++;
-   }
-   return 0;
-}
-
-#if defined(STBI_NO_JPEG) && defined(STBI_NO_HDR) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM)
-// nothing
-#else
-stbi_inline static int stbi__at_eof(stbi__context *s)
-{
-   if (s->io.read) {
-      if (!(s->io.eof)(s->io_user_data)) return 0;
-      // if feof() is true, check if buffer = end
-      // special case: we've only got the special 0 character at the end
-      if (s->read_from_callbacks == 0) return 1;
-   }
-
-   return s->img_buffer >= s->img_buffer_end;
-}
-#endif
-
-#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC)
-// nothing
-#else
-static void stbi__skip(stbi__context *s, int n)
-{
-   if (n == 0) return;  // already there!
-   if (n < 0) {
-      s->img_buffer = s->img_buffer_end;
-      return;
-   }
-   if (s->io.read) {
-      int blen = (int) (s->img_buffer_end - s->img_buffer);
-      if (blen < n) {
-         s->img_buffer = s->img_buffer_end;
-         (s->io.skip)(s->io_user_data, n - blen);
-         return;
-      }
-   }
-   s->img_buffer += n;
-}
-#endif
-
-#if defined(STBI_NO_PNG) && defined(STBI_NO_TGA) && defined(STBI_NO_HDR) && defined(STBI_NO_PNM)
-// nothing
-#else
-static int stbi__getn(stbi__context *s, stbi_uc *buffer, int n)
-{
-   if (s->io.read) {
-      int blen = (int) (s->img_buffer_end - s->img_buffer);
-      if (blen < n) {
-         int res, count;
-
-         memcpy(buffer, s->img_buffer, blen);
-
-         count = (s->io.read)(s->io_user_data, (char*) buffer + blen, n - blen);
-         res = (count == (n-blen));
-         s->img_buffer = s->img_buffer_end;
-         return res;
-      }
-   }
-
-   if (s->img_buffer+n <= s->img_buffer_end) {
-      memcpy(buffer, s->img_buffer, n);
-      s->img_buffer += n;
-      return 1;
-   } else
-      return 0;
-}
-#endif
-
-#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_PSD) && defined(STBI_NO_PIC)
-// nothing
-#else
-static int stbi__get16be(stbi__context *s)
-{
-   int z = stbi__get8(s);
-   return (z << 8) + stbi__get8(s);
-}
-#endif
-
-#if defined(STBI_NO_PNG) && defined(STBI_NO_PSD) && defined(STBI_NO_PIC)
-// nothing
-#else
-static stbi__uint32 stbi__get32be(stbi__context *s)
-{
-   stbi__uint32 z = stbi__get16be(s);
-   return (z << 16) + stbi__get16be(s);
-}
-#endif
-
-#if defined(STBI_NO_BMP) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF)
-// nothing
-#else
-static int stbi__get16le(stbi__context *s)
-{
-   int z = stbi__get8(s);
-   return z + (stbi__get8(s) << 8);
-}
-#endif
-
-#ifndef STBI_NO_BMP
-static stbi__uint32 stbi__get32le(stbi__context *s)
-{
-   stbi__uint32 z = stbi__get16le(s);
-   z += (stbi__uint32)stbi__get16le(s) << 16;
-   return z;
-}
-#endif
-
-#define STBI__BYTECAST(x)  ((stbi_uc) ((x) & 255))  // truncate int to byte without warnings
-
-#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM)
-// nothing
-#else
-//////////////////////////////////////////////////////////////////////////////
-//
-//  generic converter from built-in img_n to req_comp
-//    individual types do this automatically as much as possible (e.g. jpeg
-//    does all cases internally since it needs to colorspace convert anyway,
-//    and it never has alpha, so very few cases ). png can automatically
-//    interleave an alpha=255 channel, but falls back to this for other cases
-//
-//  assume data buffer is malloced, so malloc a new one and free that one
-//  only failure mode is malloc failing
-
-static stbi_uc stbi__compute_y(int r, int g, int b)
-{
-   return (stbi_uc) (((r*77) + (g*150) +  (29*b)) >> 8);
-}
-#endif
-
-#if defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM)
-// nothing
-#else
-static unsigned char *stbi__convert_format(unsigned char *data, int img_n, int req_comp, unsigned int x, unsigned int y)
-{
-   int i,j;
-   unsigned char *good;
-
-   if (req_comp == img_n) return data;
-   STBI_ASSERT(req_comp >= 1 && req_comp <= 4);
-
-   good = (unsigned char *) stbi__malloc_mad3(req_comp, x, y, 0);
-   if (good == NULL) {
-      STBI_FREE(data);
-      return stbi__errpuc("outofmem", "Out of memory");
-   }
-
-   for (j=0; j < (int) y; ++j) {
-      unsigned char *src  = data + j * x * img_n   ;
-      unsigned char *dest = good + j * x * req_comp;
-
-      #define STBI__COMBO(a,b)  ((a)*8+(b))
-      #define STBI__CASE(a,b)   case STBI__COMBO(a,b): for(i=x-1; i >= 0; --i, src += a, dest += b)
-      // convert source image with img_n components to one with req_comp components;
-      // avoid switch per pixel, so use switch per scanline and massive macros
-      switch (STBI__COMBO(img_n, req_comp)) {
-         STBI__CASE(1,2) { dest[0]=src[0]; dest[1]=255;                                     } break;
-         STBI__CASE(1,3) { dest[0]=dest[1]=dest[2]=src[0];                                  } break;
-         STBI__CASE(1,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=255;                     } break;
-         STBI__CASE(2,1) { dest[0]=src[0];                                                  } break;
-         STBI__CASE(2,3) { dest[0]=dest[1]=dest[2]=src[0];                                  } break;
-         STBI__CASE(2,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=src[1];                  } break;
-         STBI__CASE(3,4) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];dest[3]=255;        } break;
-         STBI__CASE(3,1) { dest[0]=stbi__compute_y(src[0],src[1],src[2]);                   } break;
-         STBI__CASE(3,2) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); dest[1] = 255;    } break;
-         STBI__CASE(4,1) { dest[0]=stbi__compute_y(src[0],src[1],src[2]);                   } break;
-         STBI__CASE(4,2) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); dest[1] = src[3]; } break;
-         STBI__CASE(4,3) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];                    } break;
-         default: STBI_ASSERT(0); STBI_FREE(data); STBI_FREE(good); return stbi__errpuc("unsupported", "Unsupported format conversion");
-      }
-      #undef STBI__CASE
-   }
-
-   STBI_FREE(data);
-   return good;
-}
-#endif
-
-#if defined(STBI_NO_PNG) && defined(STBI_NO_PSD)
-// nothing
-#else
-static stbi__uint16 stbi__compute_y_16(int r, int g, int b)
-{
-   return (stbi__uint16) (((r*77) + (g*150) +  (29*b)) >> 8);
-}
-#endif
-
-#if defined(STBI_NO_PNG) && defined(STBI_NO_PSD)
-// nothing
-#else
-static stbi__uint16 *stbi__convert_format16(stbi__uint16 *data, int img_n, int req_comp, unsigned int x, unsigned int y)
-{
-   int i,j;
-   stbi__uint16 *good;
-
-   if (req_comp == img_n) return data;
-   STBI_ASSERT(req_comp >= 1 && req_comp <= 4);
-
-   good = (stbi__uint16 *) stbi__malloc(req_comp * x * y * 2);
-   if (good == NULL) {
-      STBI_FREE(data);
-      return (stbi__uint16 *) stbi__errpuc("outofmem", "Out of memory");
-   }
-
-   for (j=0; j < (int) y; ++j) {
-      stbi__uint16 *src  = data + j * x * img_n   ;
-      stbi__uint16 *dest = good + j * x * req_comp;
-
-      #define STBI__COMBO(a,b)  ((a)*8+(b))
-      #define STBI__CASE(a,b)   case STBI__COMBO(a,b): for(i=x-1; i >= 0; --i, src += a, dest += b)
-      // convert source image with img_n components to one with req_comp components;
-      // avoid switch per pixel, so use switch per scanline and massive macros
-      switch (STBI__COMBO(img_n, req_comp)) {
-         STBI__CASE(1,2) { dest[0]=src[0]; dest[1]=0xffff;                                     } break;
-         STBI__CASE(1,3) { dest[0]=dest[1]=dest[2]=src[0];                                     } break;
-         STBI__CASE(1,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=0xffff;                     } break;
-         STBI__CASE(2,1) { dest[0]=src[0];                                                     } break;
-         STBI__CASE(2,3) { dest[0]=dest[1]=dest[2]=src[0];                                     } break;
-         STBI__CASE(2,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=src[1];                     } break;
-         STBI__CASE(3,4) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];dest[3]=0xffff;        } break;
-         STBI__CASE(3,1) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]);                   } break;
-         STBI__CASE(3,2) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); dest[1] = 0xffff; } break;
-         STBI__CASE(4,1) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]);                   } break;
-         STBI__CASE(4,2) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); dest[1] = src[3]; } break;
-         STBI__CASE(4,3) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];                       } break;
-         default: STBI_ASSERT(0); STBI_FREE(data); STBI_FREE(good); return (stbi__uint16*) stbi__errpuc("unsupported", "Unsupported format conversion");
-      }
-      #undef STBI__CASE
-   }
-
-   STBI_FREE(data);
-   return good;
-}
-#endif
-
-#ifndef STBI_NO_LINEAR
-static float   *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp)
-{
-   int i,k,n;
-   float *output;
-   if (!data) return NULL;
-   output = (float *) stbi__malloc_mad4(x, y, comp, sizeof(float), 0);
-   if (output == NULL) { STBI_FREE(data); return stbi__errpf("outofmem", "Out of memory"); }
-   // compute number of non-alpha components
-   if (comp & 1) n = comp; else n = comp-1;
-   for (i=0; i < x*y; ++i) {
-      for (k=0; k < n; ++k) {
-         output[i*comp + k] = powf(data[i*comp+k]/255.0f, stbi__l2h_gamma) * stbi__l2h_scale;
-      }
-   }
-   if (n < comp) {
-      for (i=0; i < x*y; ++i) {
-         output[i*comp + n] = data[i*comp + n]/255.0f;
-      }
-   }
-   STBI_FREE(data);
-   return output;
-}
-#endif
-
-#ifndef STBI_NO_HDR
-#define stbi__float2int(x)   ((int) (x))
-static stbi_uc *stbi__hdr_to_ldr(float   *data, int x, int y, int comp)
-{
-   int i,k,n;
-   stbi_uc *output;
-   if (!data) return NULL;
-   output = (stbi_uc *) stbi__malloc_mad3(x, y, comp, 0);
-   if (output == NULL) { STBI_FREE(data); return stbi__errpuc("outofmem", "Out of memory"); }
-   // compute number of non-alpha components
-   if (comp & 1) n = comp; else n = comp-1;
-   for (i=0; i < x*y; ++i) {
-      for (k=0; k < n; ++k) {
-         float z = (float) pow(data[i*comp+k]*stbi__h2l_scale_i, stbi__h2l_gamma_i) * 255 + 0.5f;
-         if (z < 0) z = 0;
-         if (z > 255) z = 255;
-         output[i*comp + k] = (stbi_uc) stbi__float2int(z);
-      }
-      if (k < comp) {
-         float z = data[i*comp+k] * 255 + 0.5f;
-         if (z < 0) z = 0;
-         if (z > 255) z = 255;
-         output[i*comp + k] = (stbi_uc) stbi__float2int(z);
-      }
-   }
-   STBI_FREE(data);
-   return output;
-}
-#endif
-
-//////////////////////////////////////////////////////////////////////////////
-//
-//  "baseline" JPEG/JFIF decoder
-//
-//    simple implementation
-//      - doesn't support delayed output of y-dimension
-//      - simple interface (only one output format: 8-bit interleaved RGB)
-//      - doesn't try to recover corrupt jpegs
-//      - doesn't allow partial loading, loading multiple at once
-//      - still fast on x86 (copying globals into locals doesn't help x86)
-//      - allocates lots of intermediate memory (full size of all components)
-//        - non-interleaved case requires this anyway
-//        - allows good upsampling (see next)
-//    high-quality
-//      - upsampled channels are bilinearly interpolated, even across blocks
-//      - quality integer IDCT derived from IJG's 'slow'
-//    performance
-//      - fast huffman; reasonable integer IDCT
-//      - some SIMD kernels for common paths on targets with SSE2/NEON
-//      - uses a lot of intermediate memory, could cache poorly
-
-#ifndef STBI_NO_JPEG
-
-// huffman decoding acceleration
-#define FAST_BITS   9  // larger handles more cases; smaller stomps less cache
-
-typedef struct
-{
-   stbi_uc  fast[1 << FAST_BITS];
-   // weirdly, repacking this into AoS is a 10% speed loss, instead of a win
-   stbi__uint16 code[256];
-   stbi_uc  values[256];
-   stbi_uc  size[257];
-   unsigned int maxcode[18];
-   int    delta[17];   // old 'firstsymbol' - old 'firstcode'
-} stbi__huffman;
-
-typedef struct
-{
-   stbi__context *s;
-   stbi__huffman huff_dc[4];
-   stbi__huffman huff_ac[4];
-   stbi__uint16 dequant[4][64];
-   stbi__int16 fast_ac[4][1 << FAST_BITS];
-
-// sizes for components, interleaved MCUs
-   int img_h_max, img_v_max;
-   int img_mcu_x, img_mcu_y;
-   int img_mcu_w, img_mcu_h;
-
-// definition of jpeg image component
-   struct
-   {
-      int id;
-      int h,v;
-      int tq;
-      int hd,ha;
-      int dc_pred;
-
-      int x,y,w2,h2;
-      stbi_uc *data;
-      void *raw_data, *raw_coeff;
-      stbi_uc *linebuf;
-      short   *coeff;   // progressive only
-      int      coeff_w, coeff_h; // number of 8x8 coefficient blocks
-   } img_comp[4];
-
-   stbi__uint32   code_buffer; // jpeg entropy-coded buffer
-   int            code_bits;   // number of valid bits
-   unsigned char  marker;      // marker seen while filling entropy buffer
-   int            nomore;      // flag if we saw a marker so must stop
-
-   int            progressive;
-   int            spec_start;
-   int            spec_end;
-   int            succ_high;
-   int            succ_low;
-   int            eob_run;
-   int            jfif;
-   int            app14_color_transform; // Adobe APP14 tag
-   int            rgb;
-
-   int scan_n, order[4];
-   int restart_interval, todo;
-
-// kernels
-   void (*idct_block_kernel)(stbi_uc *out, int out_stride, short data[64]);
-   void (*YCbCr_to_RGB_kernel)(stbi_uc *out, const stbi_uc *y, const stbi_uc *pcb, const stbi_uc *pcr, int count, int step);
-   stbi_uc *(*resample_row_hv_2_kernel)(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs);
-} stbi__jpeg;
-
-static int stbi__build_huffman(stbi__huffman *h, int *count)
-{
-   int i,j,k=0;
-   unsigned int code;
-   // build size list for each symbol (from JPEG spec)
-   for (i=0; i < 16; ++i) {
-      for (j=0; j < count[i]; ++j) {
-         h->size[k++] = (stbi_uc) (i+1);
-         if(k >= 257) return stbi__err("bad size list","Corrupt JPEG");
-      }
-   }
-   h->size[k] = 0;
-
-   // compute actual symbols (from jpeg spec)
-   code = 0;
-   k = 0;
-   for(j=1; j <= 16; ++j) {
-      // compute delta to add to code to compute symbol id
-      h->delta[j] = k - code;
-      if (h->size[k] == j) {
-         while (h->size[k] == j)
-            h->code[k++] = (stbi__uint16) (code++);
-         if (code-1 >= (1u << j)) return stbi__err("bad code lengths","Corrupt JPEG");
-      }
-      // compute largest code + 1 for this size, preshifted as needed later
-      h->maxcode[j] = code << (16-j);
-      code <<= 1;
-   }
-   h->maxcode[j] = 0xffffffff;
-
-   // build non-spec acceleration table; 255 is flag for not-accelerated
-   memset(h->fast, 255, 1 << FAST_BITS);
-   for (i=0; i < k; ++i) {
-      int s = h->size[i];
-      if (s <= FAST_BITS) {
-         int c = h->code[i] << (FAST_BITS-s);
-         int m = 1 << (FAST_BITS-s);
-         for (j=0; j < m; ++j) {
-            h->fast[c+j] = (stbi_uc) i;
-         }
-      }
-   }
-   return 1;
-}
-
-// build a table that decodes both magnitude and value of small ACs in
-// one go.
-static void stbi__build_fast_ac(stbi__int16 *fast_ac, stbi__huffman *h)
-{
-   int i;
-   for (i=0; i < (1 << FAST_BITS); ++i) {
-      stbi_uc fast = h->fast[i];
-      fast_ac[i] = 0;
-      if (fast < 255) {
-         int rs = h->values[fast];
-         int run = (rs >> 4) & 15;
-         int magbits = rs & 15;
-         int len = h->size[fast];
-
-         if (magbits && len + magbits <= FAST_BITS) {
-            // magnitude code followed by receive_extend code
-            int k = ((i << len) & ((1 << FAST_BITS) - 1)) >> (FAST_BITS - magbits);
-            int m = 1 << (magbits - 1);
-            if (k < m) k += (~0U << magbits) + 1;
-            // if the result is small enough, we can fit it in fast_ac table
-            if (k >= -128 && k <= 127)
-               fast_ac[i] = (stbi__int16) ((k * 256) + (run * 16) + (len + magbits));
-         }
-      }
-   }
-}
-
-static void stbi__grow_buffer_unsafe(stbi__jpeg *j)
-{
-   do {
-      unsigned int b = j->nomore ? 0 : stbi__get8(j->s);
-      if (b == 0xff) {
-         int c = stbi__get8(j->s);
-         while (c == 0xff) c = stbi__get8(j->s); // consume fill bytes
-         if (c != 0) {
-            j->marker = (unsigned char) c;
-            j->nomore = 1;
-            return;
-         }
-      }
-      j->code_buffer |= b << (24 - j->code_bits);
-      j->code_bits += 8;
-   } while (j->code_bits <= 24);
-}
-
-// (1 << n) - 1
-static const stbi__uint32 stbi__bmask[17]={0,1,3,7,15,31,63,127,255,511,1023,2047,4095,8191,16383,32767,65535};
-
-// decode a jpeg huffman value from the bitstream
-stbi_inline static int stbi__jpeg_huff_decode(stbi__jpeg *j, stbi__huffman *h)
-{
-   unsigned int temp;
-   int c,k;
-
-   if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
-
-   // look at the top FAST_BITS and determine what symbol ID it is,
-   // if the code is <= FAST_BITS
-   c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1);
-   k = h->fast[c];
-   if (k < 255) {
-      int s = h->size[k];
-      if (s > j->code_bits)
-         return -1;
-      j->code_buffer <<= s;
-      j->code_bits -= s;
-      return h->values[k];
-   }
-
-   // naive test is to shift the code_buffer down so k bits are
-   // valid, then test against maxcode. To speed this up, we've
-   // preshifted maxcode left so that it has (16-k) 0s at the
-   // end; in other words, regardless of the number of bits, it
-   // wants to be compared against something shifted to have 16;
-   // that way we don't need to shift inside the loop.
-   temp = j->code_buffer >> 16;
-   for (k=FAST_BITS+1 ; ; ++k)
-      if (temp < h->maxcode[k])
-         break;
-   if (k == 17) {
-      // error! code not found
-      j->code_bits -= 16;
-      return -1;
-   }
-
-   if (k > j->code_bits)
-      return -1;
-
-   // convert the huffman code to the symbol id
-   c = ((j->code_buffer >> (32 - k)) & stbi__bmask[k]) + h->delta[k];
-   if(c < 0 || c >= 256) // symbol id out of bounds!
-       return -1;
-   STBI_ASSERT((((j->code_buffer) >> (32 - h->size[c])) & stbi__bmask[h->size[c]]) == h->code[c]);
-
-   // convert the id to a symbol
-   j->code_bits -= k;
-   j->code_buffer <<= k;
-   return h->values[c];
-}
-
-// bias[n] = (-1<<n) + 1
-static const int stbi__jbias[16] = {0,-1,-3,-7,-15,-31,-63,-127,-255,-511,-1023,-2047,-4095,-8191,-16383,-32767};
-
-// combined JPEG 'receive' and JPEG 'extend', since baseline
-// always extends everything it receives.
-stbi_inline static int stbi__extend_receive(stbi__jpeg *j, int n)
-{
-   unsigned int k;
-   int sgn;
-   if (j->code_bits < n) stbi__grow_buffer_unsafe(j);
-   if (j->code_bits < n) return 0; // ran out of bits from stream, return 0s intead of continuing
-
-   sgn = j->code_buffer >> 31; // sign bit always in MSB; 0 if MSB clear (positive), 1 if MSB set (negative)
-   k = stbi_lrot(j->code_buffer, n);
-   j->code_buffer = k & ~stbi__bmask[n];
-   k &= stbi__bmask[n];
-   j->code_bits -= n;
-   return k + (stbi__jbias[n] & (sgn - 1));
-}
-
-// get some unsigned bits
-stbi_inline static int stbi__jpeg_get_bits(stbi__jpeg *j, int n)
-{
-   unsigned int k;
-   if (j->code_bits < n) stbi__grow_buffer_unsafe(j);
-   if (j->code_bits < n) return 0; // ran out of bits from stream, return 0s intead of continuing
-   k = stbi_lrot(j->code_buffer, n);
-   j->code_buffer = k & ~stbi__bmask[n];
-   k &= stbi__bmask[n];
-   j->code_bits -= n;
-   return k;
-}
-
-stbi_inline static int stbi__jpeg_get_bit(stbi__jpeg *j)
-{
-   unsigned int k;
-   if (j->code_bits < 1) stbi__grow_buffer_unsafe(j);
-   if (j->code_bits < 1) return 0; // ran out of bits from stream, return 0s intead of continuing
-   k = j->code_buffer;
-   j->code_buffer <<= 1;
-   --j->code_bits;
-   return k & 0x80000000;
-}
-
-// given a value that's at position X in the zigzag stream,
-// where does it appear in the 8x8 matrix coded as row-major?
-static const stbi_uc stbi__jpeg_dezigzag[64+15] =
-{
-    0,  1,  8, 16,  9,  2,  3, 10,
-   17, 24, 32, 25, 18, 11,  4,  5,
-   12, 19, 26, 33, 40, 48, 41, 34,
-   27, 20, 13,  6,  7, 14, 21, 28,
-   35, 42, 49, 56, 57, 50, 43, 36,
-   29, 22, 15, 23, 30, 37, 44, 51,
-   58, 59, 52, 45, 38, 31, 39, 46,
-   53, 60, 61, 54, 47, 55, 62, 63,
-   // let corrupt input sample past end
-   63, 63, 63, 63, 63, 63, 63, 63,
-   63, 63, 63, 63, 63, 63, 63
-};
-
-// decode one 64-entry block--
-static int stbi__jpeg_decode_block(stbi__jpeg *j, short data[64], stbi__huffman *hdc, stbi__huffman *hac, stbi__int16 *fac, int b, stbi__uint16 *dequant)
-{
-   int diff,dc,k;
-   int t;
-
-   if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
-   t = stbi__jpeg_huff_decode(j, hdc);
-   if (t < 0 || t > 15) return stbi__err("bad huffman code","Corrupt JPEG");
-
-   // 0 all the ac values now so we can do it 32-bits at a time
-   memset(data,0,64*sizeof(data[0]));
-
-   diff = t ? stbi__extend_receive(j, t) : 0;
-   if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) return stbi__err("bad delta","Corrupt JPEG");
-   dc = j->img_comp[b].dc_pred + diff;
-   j->img_comp[b].dc_pred = dc;
-   if (!stbi__mul2shorts_valid(dc, dequant[0])) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
-   data[0] = (short) (dc * dequant[0]);
-
-   // decode AC components, see JPEG spec
-   k = 1;
-   do {
-      unsigned int zig;
-      int c,r,s;
-      if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
-      c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1);
-      r = fac[c];
-      if (r) { // fast-AC path
-         k += (r >> 4) & 15; // run
-         s = r & 15; // combined length
-         if (s > j->code_bits) return stbi__err("bad huffman code", "Combined length longer than code bits available");
-         j->code_buffer <<= s;
-         j->code_bits -= s;
-         // decode into unzigzag'd location
-         zig = stbi__jpeg_dezigzag[k++];
-         data[zig] = (short) ((r >> 8) * dequant[zig]);
-      } else {
-         int rs = stbi__jpeg_huff_decode(j, hac);
-         if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
-         s = rs & 15;
-         r = rs >> 4;
-         if (s == 0) {
-            if (rs != 0xf0) break; // end block
-            k += 16;
-         } else {
-            k += r;
-            // decode into unzigzag'd location
-            zig = stbi__jpeg_dezigzag[k++];
-            data[zig] = (short) (stbi__extend_receive(j,s) * dequant[zig]);
-         }
-      }
-   } while (k < 64);
-   return 1;
-}
-
-static int stbi__jpeg_decode_block_prog_dc(stbi__jpeg *j, short data[64], stbi__huffman *hdc, int b)
-{
-   int diff,dc;
-   int t;
-   if (j->spec_end != 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
-
-   if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
-
-   if (j->succ_high == 0) {
-      // first scan for DC coefficient, must be first
-      memset(data,0,64*sizeof(data[0])); // 0 all the ac values now
-      t = stbi__jpeg_huff_decode(j, hdc);
-      if (t < 0 || t > 15) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
-      diff = t ? stbi__extend_receive(j, t) : 0;
-
-      if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) return stbi__err("bad delta", "Corrupt JPEG");
-      dc = j->img_comp[b].dc_pred + diff;
-      j->img_comp[b].dc_pred = dc;
-      if (!stbi__mul2shorts_valid(dc, 1 << j->succ_low)) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
-      data[0] = (short) (dc * (1 << j->succ_low));
-   } else {
-      // refinement scan for DC coefficient
-      if (stbi__jpeg_get_bit(j))
-         data[0] += (short) (1 << j->succ_low);
-   }
-   return 1;
-}
-
-// @OPTIMIZE: store non-zigzagged during the decode passes,
-// and only de-zigzag when dequantizing
-static int stbi__jpeg_decode_block_prog_ac(stbi__jpeg *j, short data[64], stbi__huffman *hac, stbi__int16 *fac)
-{
-   int k;
-   if (j->spec_start == 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
-
-   if (j->succ_high == 0) {
-      int shift = j->succ_low;
-
-      if (j->eob_run) {
-         --j->eob_run;
-         return 1;
-      }
-
-      k = j->spec_start;
-      do {
-         unsigned int zig;
-         int c,r,s;
-         if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
-         c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1);
-         r = fac[c];
-         if (r) { // fast-AC path
-            k += (r >> 4) & 15; // run
-            s = r & 15; // combined length
-            if (s > j->code_bits) return stbi__err("bad huffman code", "Combined length longer than code bits available");
-            j->code_buffer <<= s;
-            j->code_bits -= s;
-            zig = stbi__jpeg_dezigzag[k++];
-            data[zig] = (short) ((r >> 8) * (1 << shift));
-         } else {
-            int rs = stbi__jpeg_huff_decode(j, hac);
-            if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
-            s = rs & 15;
-            r = rs >> 4;
-            if (s == 0) {
-               if (r < 15) {
-                  j->eob_run = (1 << r);
-                  if (r)
-                     j->eob_run += stbi__jpeg_get_bits(j, r);
-                  --j->eob_run;
-                  break;
-               }
-               k += 16;
-            } else {
-               k += r;
-               zig = stbi__jpeg_dezigzag[k++];
-               data[zig] = (short) (stbi__extend_receive(j,s) * (1 << shift));
-            }
-         }
-      } while (k <= j->spec_end);
-   } else {
-      // refinement scan for these AC coefficients
-
-      short bit = (short) (1 << j->succ_low);
-
-      if (j->eob_run) {
-         --j->eob_run;
-         for (k = j->spec_start; k <= j->spec_end; ++k) {
-            short *p = &data[stbi__jpeg_dezigzag[k]];
-            if (*p != 0)
-               if (stbi__jpeg_get_bit(j))
-                  if ((*p & bit)==0) {
-                     if (*p > 0)
-                        *p += bit;
-                     else
-                        *p -= bit;
-                  }
-         }
-      } else {
-         k = j->spec_start;
-         do {
-            int r,s;
-            int rs = stbi__jpeg_huff_decode(j, hac); // @OPTIMIZE see if we can use the fast path here, advance-by-r is so slow, eh
-            if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
-            s = rs & 15;
-            r = rs >> 4;
-            if (s == 0) {
-               if (r < 15) {
-                  j->eob_run = (1 << r) - 1;
-                  if (r)
-                     j->eob_run += stbi__jpeg_get_bits(j, r);
-                  r = 64; // force end of block
-               } else {
-                  // r=15 s=0 should write 16 0s, so we just do
-                  // a run of 15 0s and then write s (which is 0),
-                  // so we don't have to do anything special here
-               }
-            } else {
-               if (s != 1) return stbi__err("bad huffman code", "Corrupt JPEG");
-               // sign bit
-               if (stbi__jpeg_get_bit(j))
-                  s = bit;
-               else
-                  s = -bit;
-            }
-
-            // advance by r
-            while (k <= j->spec_end) {
-               short *p = &data[stbi__jpeg_dezigzag[k++]];
-               if (*p != 0) {
-                  if (stbi__jpeg_get_bit(j))
-                     if ((*p & bit)==0) {
-                        if (*p > 0)
-                           *p += bit;
-                        else
-                           *p -= bit;
-                     }
-               } else {
-                  if (r == 0) {
-                     *p = (short) s;
-                     break;
-                  }
-                  --r;
-               }
-            }
-         } while (k <= j->spec_end);
-      }
-   }
-   return 1;
-}
-
-// take a -128..127 value and stbi__clamp it and convert to 0..255
-stbi_inline static stbi_uc stbi__clamp(int x)
-{
-   // trick to use a single test to catch both cases
-   if ((unsigned int) x > 255) {
-      if (x < 0) return 0;
-      if (x > 255) return 255;
-   }
-   return (stbi_uc) x;
-}
-
-#define stbi__f2f(x)  ((int) (((x) * 4096 + 0.5)))
-#define stbi__fsh(x)  ((x) * 4096)
-
-// derived from jidctint -- DCT_ISLOW
-#define STBI__IDCT_1D(s0,s1,s2,s3,s4,s5,s6,s7) \
-   int t0,t1,t2,t3,p1,p2,p3,p4,p5,x0,x1,x2,x3; \
-   p2 = s2;                                    \
-   p3 = s6;                                    \
-   p1 = (p2+p3) * stbi__f2f(0.5411961f);       \
-   t2 = p1 + p3*stbi__f2f(-1.847759065f);      \
-   t3 = p1 + p2*stbi__f2f( 0.765366865f);      \
-   p2 = s0;                                    \
-   p3 = s4;                                    \
-   t0 = stbi__fsh(p2+p3);                      \
-   t1 = stbi__fsh(p2-p3);                      \
-   x0 = t0+t3;                                 \
-   x3 = t0-t3;                                 \
-   x1 = t1+t2;                                 \
-   x2 = t1-t2;                                 \
-   t0 = s7;                                    \
-   t1 = s5;                                    \
-   t2 = s3;                                    \
-   t3 = s1;                                    \
-   p3 = t0+t2;                                 \
-   p4 = t1+t3;                                 \
-   p1 = t0+t3;                                 \
-   p2 = t1+t2;                                 \
-   p5 = (p3+p4)*stbi__f2f( 1.175875602f);      \
-   t0 = t0*stbi__f2f( 0.298631336f);           \
-   t1 = t1*stbi__f2f( 2.053119869f);           \
-   t2 = t2*stbi__f2f( 3.072711026f);           \
-   t3 = t3*stbi__f2f( 1.501321110f);           \
-   p1 = p5 + p1*stbi__f2f(-0.899976223f);      \
-   p2 = p5 + p2*stbi__f2f(-2.562915447f);      \
-   p3 = p3*stbi__f2f(-1.961570560f);           \
-   p4 = p4*stbi__f2f(-0.390180644f);           \
-   t3 += p1+p4;                                \
-   t2 += p2+p3;                                \
-   t1 += p2+p4;                                \
-   t0 += p1+p3;
-
-static void stbi__idct_block(stbi_uc *out, int out_stride, short data[64])
-{
-   int i,val[64],*v=val;
-   stbi_uc *o;
-   short *d = data;
-
-   // columns
-   for (i=0; i < 8; ++i,++d, ++v) {
-      // if all zeroes, shortcut -- this avoids dequantizing 0s and IDCTing
-      if (d[ 8]==0 && d[16]==0 && d[24]==0 && d[32]==0
-           && d[40]==0 && d[48]==0 && d[56]==0) {
-         //    no shortcut                 0     seconds
-         //    (1|2|3|4|5|6|7)==0          0     seconds
-         //    all separate               -0.047 seconds
-         //    1 && 2|3 && 4|5 && 6|7:    -0.047 seconds
-         int dcterm = d[0]*4;
-         v[0] = v[8] = v[16] = v[24] = v[32] = v[40] = v[48] = v[56] = dcterm;
-      } else {
-         STBI__IDCT_1D(d[ 0],d[ 8],d[16],d[24],d[32],d[40],d[48],d[56])
-         // constants scaled things up by 1<<12; let's bring them back
-         // down, but keep 2 extra bits of precision
-         x0 += 512; x1 += 512; x2 += 512; x3 += 512;
-         v[ 0] = (x0+t3) >> 10;
-         v[56] = (x0-t3) >> 10;
-         v[ 8] = (x1+t2) >> 10;
-         v[48] = (x1-t2) >> 10;
-         v[16] = (x2+t1) >> 10;
-         v[40] = (x2-t1) >> 10;
-         v[24] = (x3+t0) >> 10;
-         v[32] = (x3-t0) >> 10;
-      }
-   }
-
-   for (i=0, v=val, o=out; i < 8; ++i,v+=8,o+=out_stride) {
-      // no fast case since the first 1D IDCT spread components out
-      STBI__IDCT_1D(v[0],v[1],v[2],v[3],v[4],v[5],v[6],v[7])
-      // constants scaled things up by 1<<12, plus we had 1<<2 from first
-      // loop, plus horizontal and vertical each scale by sqrt(8) so together
-      // we've got an extra 1<<3, so 1<<17 total we need to remove.
-      // so we want to round that, which means adding 0.5 * 1<<17,
-      // aka 65536. Also, we'll end up with -128 to 127 that we want
-      // to encode as 0..255 by adding 128, so we'll add that before the shift
-      x0 += 65536 + (128<<17);
-      x1 += 65536 + (128<<17);
-      x2 += 65536 + (128<<17);
-      x3 += 65536 + (128<<17);
-      // tried computing the shifts into temps, or'ing the temps to see
-      // if any were out of range, but that was slower
-      o[0] = stbi__clamp((x0+t3) >> 17);
-      o[7] = stbi__clamp((x0-t3) >> 17);
-      o[1] = stbi__clamp((x1+t2) >> 17);
-      o[6] = stbi__clamp((x1-t2) >> 17);
-      o[2] = stbi__clamp((x2+t1) >> 17);
-      o[5] = stbi__clamp((x2-t1) >> 17);
-      o[3] = stbi__clamp((x3+t0) >> 17);
-      o[4] = stbi__clamp((x3-t0) >> 17);
-   }
-}
-
-#ifdef STBI_SSE2
-// sse2 integer IDCT. not the fastest possible implementation but it
-// produces bit-identical results to the generic C version so it's
-// fully "transparent".
-static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64])
-{
-   // This is constructed to match our regular (generic) integer IDCT exactly.
-   __m128i row0, row1, row2, row3, row4, row5, row6, row7;
-   __m128i tmp;
-
-   // dot product constant: even elems=x, odd elems=y
-   #define dct_const(x,y)  _mm_setr_epi16((x),(y),(x),(y),(x),(y),(x),(y))
-
-   // out(0) = c0[even]*x + c0[odd]*y   (c0, x, y 16-bit, out 32-bit)
-   // out(1) = c1[even]*x + c1[odd]*y
-   #define dct_rot(out0,out1, x,y,c0,c1) \
-      __m128i c0##lo = _mm_unpacklo_epi16((x),(y)); \
-      __m128i c0##hi = _mm_unpackhi_epi16((x),(y)); \
-      __m128i out0##_l = _mm_madd_epi16(c0##lo, c0); \
-      __m128i out0##_h = _mm_madd_epi16(c0##hi, c0); \
-      __m128i out1##_l = _mm_madd_epi16(c0##lo, c1); \
-      __m128i out1##_h = _mm_madd_epi16(c0##hi, c1)
-
-   // out = in << 12  (in 16-bit, out 32-bit)
-   #define dct_widen(out, in) \
-      __m128i out##_l = _mm_srai_epi32(_mm_unpacklo_epi16(_mm_setzero_si128(), (in)), 4); \
-      __m128i out##_h = _mm_srai_epi32(_mm_unpackhi_epi16(_mm_setzero_si128(), (in)), 4)
-
-   // wide add
-   #define dct_wadd(out, a, b) \
-      __m128i out##_l = _mm_add_epi32(a##_l, b##_l); \
-      __m128i out##_h = _mm_add_epi32(a##_h, b##_h)
-
-   // wide sub
-   #define dct_wsub(out, a, b) \
-      __m128i out##_l = _mm_sub_epi32(a##_l, b##_l); \
-      __m128i out##_h = _mm_sub_epi32(a##_h, b##_h)
-
-   // butterfly a/b, add bias, then shift by "s" and pack
-   #define dct_bfly32o(out0, out1, a,b,bias,s) \
-      { \
-         __m128i abiased_l = _mm_add_epi32(a##_l, bias); \
-         __m128i abiased_h = _mm_add_epi32(a##_h, bias); \
-         dct_wadd(sum, abiased, b); \
-         dct_wsub(dif, abiased, b); \
-         out0 = _mm_packs_epi32(_mm_srai_epi32(sum_l, s), _mm_srai_epi32(sum_h, s)); \
-         out1 = _mm_packs_epi32(_mm_srai_epi32(dif_l, s), _mm_srai_epi32(dif_h, s)); \
-      }
-
-   // 8-bit interleave step (for transposes)
-   #define dct_interleave8(a, b) \
-      tmp = a; \
-      a = _mm_unpacklo_epi8(a, b); \
-      b = _mm_unpackhi_epi8(tmp, b)
-
-   // 16-bit interleave step (for transposes)
-   #define dct_interleave16(a, b) \
-      tmp = a; \
-      a = _mm_unpacklo_epi16(a, b); \
-      b = _mm_unpackhi_epi16(tmp, b)
-
-   #define dct_pass(bias,shift) \
-      { \
-         /* even part */ \
-         dct_rot(t2e,t3e, row2,row6, rot0_0,rot0_1); \
-         __m128i sum04 = _mm_add_epi16(row0, row4); \
-         __m128i dif04 = _mm_sub_epi16(row0, row4); \
-         dct_widen(t0e, sum04); \
-         dct_widen(t1e, dif04); \
-         dct_wadd(x0, t0e, t3e); \
-         dct_wsub(x3, t0e, t3e); \
-         dct_wadd(x1, t1e, t2e); \
-         dct_wsub(x2, t1e, t2e); \
-         /* odd part */ \
-         dct_rot(y0o,y2o, row7,row3, rot2_0,rot2_1); \
-         dct_rot(y1o,y3o, row5,row1, rot3_0,rot3_1); \
-         __m128i sum17 = _mm_add_epi16(row1, row7); \
-         __m128i sum35 = _mm_add_epi16(row3, row5); \
-         dct_rot(y4o,y5o, sum17,sum35, rot1_0,rot1_1); \
-         dct_wadd(x4, y0o, y4o); \
-         dct_wadd(x5, y1o, y5o); \
-         dct_wadd(x6, y2o, y5o); \
-         dct_wadd(x7, y3o, y4o); \
-         dct_bfly32o(row0,row7, x0,x7,bias,shift); \
-         dct_bfly32o(row1,row6, x1,x6,bias,shift); \
-         dct_bfly32o(row2,row5, x2,x5,bias,shift); \
-         dct_bfly32o(row3,row4, x3,x4,bias,shift); \
-      }
-
-   __m128i rot0_0 = dct_const(stbi__f2f(0.5411961f), stbi__f2f(0.5411961f) + stbi__f2f(-1.847759065f));
-   __m128i rot0_1 = dct_const(stbi__f2f(0.5411961f) + stbi__f2f( 0.765366865f), stbi__f2f(0.5411961f));
-   __m128i rot1_0 = dct_const(stbi__f2f(1.175875602f) + stbi__f2f(-0.899976223f), stbi__f2f(1.175875602f));
-   __m128i rot1_1 = dct_const(stbi__f2f(1.175875602f), stbi__f2f(1.175875602f) + stbi__f2f(-2.562915447f));
-   __m128i rot2_0 = dct_const(stbi__f2f(-1.961570560f) + stbi__f2f( 0.298631336f), stbi__f2f(-1.961570560f));
-   __m128i rot2_1 = dct_const(stbi__f2f(-1.961570560f), stbi__f2f(-1.961570560f) + stbi__f2f( 3.072711026f));
-   __m128i rot3_0 = dct_const(stbi__f2f(-0.390180644f) + stbi__f2f( 2.053119869f), stbi__f2f(-0.390180644f));
-   __m128i rot3_1 = dct_const(stbi__f2f(-0.390180644f), stbi__f2f(-0.390180644f) + stbi__f2f( 1.501321110f));
-
-   // rounding biases in column/row passes, see stbi__idct_block for explanation.
-   __m128i bias_0 = _mm_set1_epi32(512);
-   __m128i bias_1 = _mm_set1_epi32(65536 + (128<<17));
-
-   // load
-   row0 = _mm_load_si128((const __m128i *) (data + 0*8));
-   row1 = _mm_load_si128((const __m128i *) (data + 1*8));
-   row2 = _mm_load_si128((const __m128i *) (data + 2*8));
-   row3 = _mm_load_si128((const __m128i *) (data + 3*8));
-   row4 = _mm_load_si128((const __m128i *) (data + 4*8));
-   row5 = _mm_load_si128((const __m128i *) (data + 5*8));
-   row6 = _mm_load_si128((const __m128i *) (data + 6*8));
-   row7 = _mm_load_si128((const __m128i *) (data + 7*8));
-
-   // column pass
-   dct_pass(bias_0, 10);
-
-   {
-      // 16bit 8x8 transpose pass 1
-      dct_interleave16(row0, row4);
-      dct_interleave16(row1, row5);
-      dct_interleave16(row2, row6);
-      dct_interleave16(row3, row7);
-
-      // transpose pass 2
-      dct_interleave16(row0, row2);
-      dct_interleave16(row1, row3);
-      dct_interleave16(row4, row6);
-      dct_interleave16(row5, row7);
-
-      // transpose pass 3
-      dct_interleave16(row0, row1);
-      dct_interleave16(row2, row3);
-      dct_interleave16(row4, row5);
-      dct_interleave16(row6, row7);
-   }
-
-   // row pass
-   dct_pass(bias_1, 17);
-
-   {
-      // pack
-      __m128i p0 = _mm_packus_epi16(row0, row1); // a0a1a2a3...a7b0b1b2b3...b7
-      __m128i p1 = _mm_packus_epi16(row2, row3);
-      __m128i p2 = _mm_packus_epi16(row4, row5);
-      __m128i p3 = _mm_packus_epi16(row6, row7);
-
-      // 8bit 8x8 transpose pass 1
-      dct_interleave8(p0, p2); // a0e0a1e1...
-      dct_interleave8(p1, p3); // c0g0c1g1...
-
-      // transpose pass 2
-      dct_interleave8(p0, p1); // a0c0e0g0...
-      dct_interleave8(p2, p3); // b0d0f0h0...
-
-      // transpose pass 3
-      dct_interleave8(p0, p2); // a0b0c0d0...
-      dct_interleave8(p1, p3); // a4b4c4d4...
-
-      // store
-      _mm_storel_epi64((__m128i *) out, p0); out += out_stride;
-      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p0, 0x4e)); out += out_stride;
-      _mm_storel_epi64((__m128i *) out, p2); out += out_stride;
-      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p2, 0x4e)); out += out_stride;
-      _mm_storel_epi64((__m128i *) out, p1); out += out_stride;
-      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p1, 0x4e)); out += out_stride;
-      _mm_storel_epi64((__m128i *) out, p3); out += out_stride;
-      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p3, 0x4e));
-   }
-
-#undef dct_const
-#undef dct_rot
-#undef dct_widen
-#undef dct_wadd
-#undef dct_wsub
-#undef dct_bfly32o
-#undef dct_interleave8
-#undef dct_interleave16
-#undef dct_pass
-}
-
-#endif // STBI_SSE2
-
-#ifdef STBI_NEON
-
-// NEON integer IDCT. should produce bit-identical
-// results to the generic C version.
-static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64])
-{
-   int16x8_t row0, row1, row2, row3, row4, row5, row6, row7;
-
-   int16x4_t rot0_0 = vdup_n_s16(stbi__f2f(0.5411961f));
-   int16x4_t rot0_1 = vdup_n_s16(stbi__f2f(-1.847759065f));
-   int16x4_t rot0_2 = vdup_n_s16(stbi__f2f( 0.765366865f));
-   int16x4_t rot1_0 = vdup_n_s16(stbi__f2f( 1.175875602f));
-   int16x4_t rot1_1 = vdup_n_s16(stbi__f2f(-0.899976223f));
-   int16x4_t rot1_2 = vdup_n_s16(stbi__f2f(-2.562915447f));
-   int16x4_t rot2_0 = vdup_n_s16(stbi__f2f(-1.961570560f));
-   int16x4_t rot2_1 = vdup_n_s16(stbi__f2f(-0.390180644f));
-   int16x4_t rot3_0 = vdup_n_s16(stbi__f2f( 0.298631336f));
-   int16x4_t rot3_1 = vdup_n_s16(stbi__f2f( 2.053119869f));
-   int16x4_t rot3_2 = vdup_n_s16(stbi__f2f( 3.072711026f));
-   int16x4_t rot3_3 = vdup_n_s16(stbi__f2f( 1.501321110f));
-
-#define dct_long_mul(out, inq, coeff) \
-   int32x4_t out##_l = vmull_s16(vget_low_s16(inq), coeff); \
-   int32x4_t out##_h = vmull_s16(vget_high_s16(inq), coeff)
-
-#define dct_long_mac(out, acc, inq, coeff) \
-   int32x4_t out##_l = vmlal_s16(acc##_l, vget_low_s16(inq), coeff); \
-   int32x4_t out##_h = vmlal_s16(acc##_h, vget_high_s16(inq), coeff)
-
-#define dct_widen(out, inq) \
-   int32x4_t out##_l = vshll_n_s16(vget_low_s16(inq), 12); \
-   int32x4_t out##_h = vshll_n_s16(vget_high_s16(inq), 12)
-
-// wide add
-#define dct_wadd(out, a, b) \
-   int32x4_t out##_l = vaddq_s32(a##_l, b##_l); \
-   int32x4_t out##_h = vaddq_s32(a##_h, b##_h)
-
-// wide sub
-#define dct_wsub(out, a, b) \
-   int32x4_t out##_l = vsubq_s32(a##_l, b##_l); \
-   int32x4_t out##_h = vsubq_s32(a##_h, b##_h)
-
-// butterfly a/b, then shift using "shiftop" by "s" and pack
-#define dct_bfly32o(out0,out1, a,b,shiftop,s) \
-   { \
-      dct_wadd(sum, a, b); \
-      dct_wsub(dif, a, b); \
-      out0 = vcombine_s16(shiftop(sum_l, s), shiftop(sum_h, s)); \
-      out1 = vcombine_s16(shiftop(dif_l, s), shiftop(dif_h, s)); \
-   }
-
-#define dct_pass(shiftop, shift) \
-   { \
-      /* even part */ \
-      int16x8_t sum26 = vaddq_s16(row2, row6); \
-      dct_long_mul(p1e, sum26, rot0_0); \
-      dct_long_mac(t2e, p1e, row6, rot0_1); \
-      dct_long_mac(t3e, p1e, row2, rot0_2); \
-      int16x8_t sum04 = vaddq_s16(row0, row4); \
-      int16x8_t dif04 = vsubq_s16(row0, row4); \
-      dct_widen(t0e, sum04); \
-      dct_widen(t1e, dif04); \
-      dct_wadd(x0, t0e, t3e); \
-      dct_wsub(x3, t0e, t3e); \
-      dct_wadd(x1, t1e, t2e); \
-      dct_wsub(x2, t1e, t2e); \
-      /* odd part */ \
-      int16x8_t sum15 = vaddq_s16(row1, row5); \
-      int16x8_t sum17 = vaddq_s16(row1, row7); \
-      int16x8_t sum35 = vaddq_s16(row3, row5); \
-      int16x8_t sum37 = vaddq_s16(row3, row7); \
-      int16x8_t sumodd = vaddq_s16(sum17, sum35); \
-      dct_long_mul(p5o, sumodd, rot1_0); \
-      dct_long_mac(p1o, p5o, sum17, rot1_1); \
-      dct_long_mac(p2o, p5o, sum35, rot1_2); \
-      dct_long_mul(p3o, sum37, rot2_0); \
-      dct_long_mul(p4o, sum15, rot2_1); \
-      dct_wadd(sump13o, p1o, p3o); \
-      dct_wadd(sump24o, p2o, p4o); \
-      dct_wadd(sump23o, p2o, p3o); \
-      dct_wadd(sump14o, p1o, p4o); \
-      dct_long_mac(x4, sump13o, row7, rot3_0); \
-      dct_long_mac(x5, sump24o, row5, rot3_1); \
-      dct_long_mac(x6, sump23o, row3, rot3_2); \
-      dct_long_mac(x7, sump14o, row1, rot3_3); \
-      dct_bfly32o(row0,row7, x0,x7,shiftop,shift); \
-      dct_bfly32o(row1,row6, x1,x6,shiftop,shift); \
-      dct_bfly32o(row2,row5, x2,x5,shiftop,shift); \
-      dct_bfly32o(row3,row4, x3,x4,shiftop,shift); \
-   }
-
-   // load
-   row0 = vld1q_s16(data + 0*8);
-   row1 = vld1q_s16(data + 1*8);
-   row2 = vld1q_s16(data + 2*8);
-   row3 = vld1q_s16(data + 3*8);
-   row4 = vld1q_s16(data + 4*8);
-   row5 = vld1q_s16(data + 5*8);
-   row6 = vld1q_s16(data + 6*8);
-   row7 = vld1q_s16(data + 7*8);
-
-   // add DC bias
-   row0 = vaddq_s16(row0, vsetq_lane_s16(1024, vdupq_n_s16(0), 0));
-
-   // column pass
-   dct_pass(vrshrn_n_s32, 10);
-
-   // 16bit 8x8 transpose
-   {
-// these three map to a single VTRN.16, VTRN.32, and VSWP, respectively.
-// whether compilers actually get this is another story, sadly.
-#define dct_trn16(x, y) { int16x8x2_t t = vtrnq_s16(x, y); x = t.val[0]; y = t.val[1]; }
-#define dct_trn32(x, y) { int32x4x2_t t = vtrnq_s32(vreinterpretq_s32_s16(x), vreinterpretq_s32_s16(y)); x = vreinterpretq_s16_s32(t.val[0]); y = vreinterpretq_s16_s32(t.val[1]); }
-#define dct_trn64(x, y) { int16x8_t x0 = x; int16x8_t y0 = y; x = vcombine_s16(vget_low_s16(x0), vget_low_s16(y0)); y = vcombine_s16(vget_high_s16(x0), vget_high_s16(y0)); }
-
-      // pass 1
-      dct_trn16(row0, row1); // a0b0a2b2a4b4a6b6
-      dct_trn16(row2, row3);
-      dct_trn16(row4, row5);
-      dct_trn16(row6, row7);
-
-      // pass 2
-      dct_trn32(row0, row2); // a0b0c0d0a4b4c4d4
-      dct_trn32(row1, row3);
-      dct_trn32(row4, row6);
-      dct_trn32(row5, row7);
-
-      // pass 3
-      dct_trn64(row0, row4); // a0b0c0d0e0f0g0h0
-      dct_trn64(row1, row5);
-      dct_trn64(row2, row6);
-      dct_trn64(row3, row7);
-
-#undef dct_trn16
-#undef dct_trn32
-#undef dct_trn64
-   }
-
-   // row pass
-   // vrshrn_n_s32 only supports shifts up to 16, we need
-   // 17. so do a non-rounding shift of 16 first then follow
-   // up with a rounding shift by 1.
-   dct_pass(vshrn_n_s32, 16);
-
-   {
-      // pack and round
-      uint8x8_t p0 = vqrshrun_n_s16(row0, 1);
-      uint8x8_t p1 = vqrshrun_n_s16(row1, 1);
-      uint8x8_t p2 = vqrshrun_n_s16(row2, 1);
-      uint8x8_t p3 = vqrshrun_n_s16(row3, 1);
-      uint8x8_t p4 = vqrshrun_n_s16(row4, 1);
-      uint8x8_t p5 = vqrshrun_n_s16(row5, 1);
-      uint8x8_t p6 = vqrshrun_n_s16(row6, 1);
-      uint8x8_t p7 = vqrshrun_n_s16(row7, 1);
-
-      // again, these can translate into one instruction, but often don't.
-#define dct_trn8_8(x, y) { uint8x8x2_t t = vtrn_u8(x, y); x = t.val[0]; y = t.val[1]; }
-#define dct_trn8_16(x, y) { uint16x4x2_t t = vtrn_u16(vreinterpret_u16_u8(x), vreinterpret_u16_u8(y)); x = vreinterpret_u8_u16(t.val[0]); y = vreinterpret_u8_u16(t.val[1]); }
-#define dct_trn8_32(x, y) { uint32x2x2_t t = vtrn_u32(vreinterpret_u32_u8(x), vreinterpret_u32_u8(y)); x = vreinterpret_u8_u32(t.val[0]); y = vreinterpret_u8_u32(t.val[1]); }
-
-      // sadly can't use interleaved stores here since we only write
-      // 8 bytes to each scan line!
-
-      // 8x8 8-bit transpose pass 1
-      dct_trn8_8(p0, p1);
-      dct_trn8_8(p2, p3);
-      dct_trn8_8(p4, p5);
-      dct_trn8_8(p6, p7);
-
-      // pass 2
-      dct_trn8_16(p0, p2);
-      dct_trn8_16(p1, p3);
-      dct_trn8_16(p4, p6);
-      dct_trn8_16(p5, p7);
-
-      // pass 3
-      dct_trn8_32(p0, p4);
-      dct_trn8_32(p1, p5);
-      dct_trn8_32(p2, p6);
-      dct_trn8_32(p3, p7);
-
-      // store
-      vst1_u8(out, p0); out += out_stride;
-      vst1_u8(out, p1); out += out_stride;
-      vst1_u8(out, p2); out += out_stride;
-      vst1_u8(out, p3); out += out_stride;
-      vst1_u8(out, p4); out += out_stride;
-      vst1_u8(out, p5); out += out_stride;
-      vst1_u8(out, p6); out += out_stride;
-      vst1_u8(out, p7);
-
-#undef dct_trn8_8
-#undef dct_trn8_16
-#undef dct_trn8_32
-   }
-
-#undef dct_long_mul
-#undef dct_long_mac
-#undef dct_widen
-#undef dct_wadd
-#undef dct_wsub
-#undef dct_bfly32o
-#undef dct_pass
-}
-
-#endif // STBI_NEON
-
-#ifdef STBI_RVV
-
-// risc-v vector integer IDCT. should produce bit-identical
-// results to the generic C version.
-static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64])
-{
-   const short rot0_0 = stbi__f2f(0.5411961f);
-   const short rot0_1 = stbi__f2f(-1.847759065f);
-   const short rot0_2 = stbi__f2f( 0.765366865f);
-   const short rot1_0 = stbi__f2f( 1.175875602f);
-   const short rot1_1 = stbi__f2f(-0.899976223f);
-   const short rot1_2 = stbi__f2f(-2.562915447f);
-   const short rot2_0 = stbi__f2f(-1.961570560f);
-   const short rot2_1 = stbi__f2f(-0.390180644f);
-   const short rot3_0 = stbi__f2f( 0.298631336f);
-   const short rot3_1 = stbi__f2f( 2.053119869f);
-   const short rot3_2 = stbi__f2f( 3.072711026f);
-   const short rot3_3 = stbi__f2f( 1.501321110f);
-
-   // scratch buffer for data transpose
-   short tmp[64];
-
-   const size_t vl = vsetvl_e16m1(8);
-
-   // column pass
-   {
-      vint16m1_t row0 = vle16_v_i16m1(data + 0*8, vl);
-      vint16m1_t row1 = vle16_v_i16m1(data + 1*8, vl);
-      vint16m1_t row2 = vle16_v_i16m1(data + 2*8, vl);
-      vint16m1_t row3 = vle16_v_i16m1(data + 3*8, vl);
-      vint16m1_t row4 = vle16_v_i16m1(data + 4*8, vl);
-      vint16m1_t row5 = vle16_v_i16m1(data + 5*8, vl);
-      vint16m1_t row6 = vle16_v_i16m1(data + 6*8, vl);
-      vint16m1_t row7 = vle16_v_i16m1(data + 7*8, vl);
-
-      // even part
-      vint16m1_t sum26 = vadd_vv_i16m1(row2, row6, vl);
-      vint32m2_t p1e = vwmul_vx_i32m2(sum26, rot0_0, vl);
-      vint32m2_t t2e = vwmacc_vx_i32m2(p1e, rot0_1, row6, vl);
-      vint32m2_t t3e = vwmacc_vx_i32m2(p1e, rot0_2, row2, vl);
-      vint32m2_t t0e = vsll_vx_i32m2(vwadd_vv_i32m2(row0, row4, vl), 12, vl);
-      vint32m2_t t1e = vsll_vx_i32m2(vwsub_vv_i32m2(row0, row4, vl), 12, vl);
-      vint32m2_t x0 = vadd_vv_i32m2(t0e, t3e, vl);
-      vint32m2_t x3 = vsub_vv_i32m2(t0e, t3e, vl);
-      vint32m2_t x1 = vadd_vv_i32m2(t1e, t2e, vl);
-      vint32m2_t x2 = vsub_vv_i32m2(t1e, t2e, vl);
-
-      // odd part
-      vint16m1_t sum15 = vadd_vv_i16m1(row1, row5, vl);
-      vint16m1_t sum17 = vadd_vv_i16m1(row1, row7, vl);
-      vint16m1_t sum35 = vadd_vv_i16m1(row3, row5, vl);
-      vint16m1_t sum37 = vadd_vv_i16m1(row3, row7, vl);
-      vint32m2_t p5o = vwmul_vx_i32m2(vadd_vv_i16m1(sum17, sum35, vl), rot1_0, vl);
-      vint32m2_t p1o = vwmacc_vx_i32m2(p5o, rot1_1, sum17, vl);
-      vint32m2_t p2o = vwmacc_vx_i32m2(p5o, rot1_2, sum35, vl);
-      vint32m2_t p3o = vwmul_vx_i32m2(sum37, rot2_0, vl);
-      vint32m2_t p4o = vwmul_vx_i32m2(sum15, rot2_1, vl);
-      vint32m2_t sump13o = vadd_vv_i32m2(p1o, p3o, vl);
-      vint32m2_t sump24o = vadd_vv_i32m2(p2o, p4o, vl);
-      vint32m2_t sump23o = vadd_vv_i32m2(p2o, p3o, vl);
-      vint32m2_t sump14o = vadd_vv_i32m2(p1o, p4o, vl);
-      vint32m2_t x4 = vwmacc_vx_i32m2(sump13o, rot3_0, row7, vl);
-      vint32m2_t x5 = vwmacc_vx_i32m2(sump24o, rot3_1, row5, vl);
-      vint32m2_t x6 = vwmacc_vx_i32m2(sump23o, rot3_2, row3, vl);
-      vint32m2_t x7 = vwmacc_vx_i32m2(sump14o, rot3_3, row1, vl);
-
-      // bfly32
-      x0 = vadd_vx_i32m2(x0, 512, vl);
-      x1 = vadd_vx_i32m2(x1, 512, vl);
-      x2 = vadd_vx_i32m2(x2, 512, vl);
-      x3 = vadd_vx_i32m2(x3, 512, vl);
-      vint16m1_t out0 = vnsra_wx_i16m1(vadd_vv_i32m2(x0, x7, vl), 10, vl);
-      vint16m1_t out7 = vnsra_wx_i16m1(vsub_vv_i32m2(x0, x7, vl), 10, vl);
-      vint16m1_t out1 = vnsra_wx_i16m1(vadd_vv_i32m2(x1, x6, vl), 10, vl);
-      vint16m1_t out6 = vnsra_wx_i16m1(vsub_vv_i32m2(x1, x6, vl), 10, vl);
-      vint16m1_t out2 = vnsra_wx_i16m1(vadd_vv_i32m2(x2, x5, vl), 10, vl);
-      vint16m1_t out5 = vnsra_wx_i16m1(vsub_vv_i32m2(x2, x5, vl), 10, vl);
-      vint16m1_t out3 = vnsra_wx_i16m1(vadd_vv_i32m2(x3, x4, vl), 10, vl);
-      vint16m1_t out4 = vnsra_wx_i16m1(vsub_vv_i32m2(x3, x4, vl), 10, vl);
-
-      // 8x8 transpose
-      // I would prefer to implement this transpose in register without save+load,
-      // but rvv does not have shuffle/zip instructions among multiple registers.
-      // what a pity :(   --- nihui
-      vsse16_v_i16m1(tmp + 0, 8 * 2, out0, vl);
-      vsse16_v_i16m1(tmp + 1, 8 * 2, out1, vl);
-      vsse16_v_i16m1(tmp + 2, 8 * 2, out2, vl);
-      vsse16_v_i16m1(tmp + 3, 8 * 2, out3, vl);
-      vsse16_v_i16m1(tmp + 4, 8 * 2, out4, vl);
-      vsse16_v_i16m1(tmp + 5, 8 * 2, out5, vl);
-      vsse16_v_i16m1(tmp + 6, 8 * 2, out6, vl);
-      vsse16_v_i16m1(tmp + 7, 8 * 2, out7, vl);
-   }
-
-   // row pass
-   {
-      vint16m1_t row0 = vle16_v_i16m1(tmp + 0*8, vl);
-      vint16m1_t row1 = vle16_v_i16m1(tmp + 1*8, vl);
-      vint16m1_t row2 = vle16_v_i16m1(tmp + 2*8, vl);
-      vint16m1_t row3 = vle16_v_i16m1(tmp + 3*8, vl);
-      vint16m1_t row4 = vle16_v_i16m1(tmp + 4*8, vl);
-      vint16m1_t row5 = vle16_v_i16m1(tmp + 5*8, vl);
-      vint16m1_t row6 = vle16_v_i16m1(tmp + 6*8, vl);
-      vint16m1_t row7 = vle16_v_i16m1(tmp + 7*8, vl);
-
-      // even part
-      vint16m1_t sum26 = vadd_vv_i16m1(row2, row6, vl);
-      vint32m2_t p1e = vwmul_vx_i32m2(sum26, rot0_0, vl);
-      vint32m2_t t2e = vwmacc_vx_i32m2(p1e, rot0_1, row6, vl);
-      vint32m2_t t3e = vwmacc_vx_i32m2(p1e, rot0_2, row2, vl);
-      vint32m2_t t0e = vsll_vx_i32m2(vwadd_vv_i32m2(row0, row4, vl), 12, vl);
-      vint32m2_t t1e = vsll_vx_i32m2(vwsub_vv_i32m2(row0, row4, vl), 12, vl);
-      vint32m2_t x0 = vadd_vv_i32m2(t0e, t3e, vl);
-      vint32m2_t x3 = vsub_vv_i32m2(t0e, t3e, vl);
-      vint32m2_t x1 = vadd_vv_i32m2(t1e, t2e, vl);
-      vint32m2_t x2 = vsub_vv_i32m2(t1e, t2e, vl);
-
-      // odd part
-      vint16m1_t sum15 = vadd_vv_i16m1(row1, row5, vl);
-      vint16m1_t sum17 = vadd_vv_i16m1(row1, row7, vl);
-      vint16m1_t sum35 = vadd_vv_i16m1(row3, row5, vl);
-      vint16m1_t sum37 = vadd_vv_i16m1(row3, row7, vl);
-      vint32m2_t p5o = vwmul_vx_i32m2(vadd_vv_i16m1(sum17, sum35, vl), rot1_0, vl);
-      vint32m2_t p1o = vwmacc_vx_i32m2(p5o, rot1_1, sum17, vl);
-      vint32m2_t p2o = vwmacc_vx_i32m2(p5o, rot1_2, sum35, vl);
-      vint32m2_t p3o = vwmul_vx_i32m2(sum37, rot2_0, vl);
-      vint32m2_t p4o = vwmul_vx_i32m2(sum15, rot2_1, vl);
-      vint32m2_t sump13o = vadd_vv_i32m2(p1o, p3o, vl);
-      vint32m2_t sump24o = vadd_vv_i32m2(p2o, p4o, vl);
-      vint32m2_t sump23o = vadd_vv_i32m2(p2o, p3o, vl);
-      vint32m2_t sump14o = vadd_vv_i32m2(p1o, p4o, vl);
-      vint32m2_t x4 = vwmacc_vx_i32m2(sump13o, rot3_0, row7, vl);
-      vint32m2_t x5 = vwmacc_vx_i32m2(sump24o, rot3_1, row5, vl);
-      vint32m2_t x6 = vwmacc_vx_i32m2(sump23o, rot3_2, row3, vl);
-      vint32m2_t x7 = vwmacc_vx_i32m2(sump14o, rot3_3, row1, vl);
-
-      // bfly32
-      x0 = vadd_vx_i32m2(x0, (int)(65536 + (128<<17)), vl);
-      x1 = vadd_vx_i32m2(x1, (int)(65536 + (128<<17)), vl);
-      x2 = vadd_vx_i32m2(x2, (int)(65536 + (128<<17)), vl);
-      x3 = vadd_vx_i32m2(x3, (int)(65536 + (128<<17)), vl);
-      vint16m1_t out0 = vnsra_wx_i16m1(vadd_vv_i32m2(x0, x7, vl), 17, vl);
-      vint16m1_t out7 = vnsra_wx_i16m1(vsub_vv_i32m2(x0, x7, vl), 17, vl);
-      vint16m1_t out1 = vnsra_wx_i16m1(vadd_vv_i32m2(x1, x6, vl), 17, vl);
-      vint16m1_t out6 = vnsra_wx_i16m1(vsub_vv_i32m2(x1, x6, vl), 17, vl);
-      vint16m1_t out2 = vnsra_wx_i16m1(vadd_vv_i32m2(x2, x5, vl), 17, vl);
-      vint16m1_t out5 = vnsra_wx_i16m1(vsub_vv_i32m2(x2, x5, vl), 17, vl);
-      vint16m1_t out3 = vnsra_wx_i16m1(vadd_vv_i32m2(x3, x4, vl), 17, vl);
-      vint16m1_t out4 = vnsra_wx_i16m1(vsub_vv_i32m2(x3, x4, vl), 17, vl);
-
-      // clamp 0~255
-      vuint8m1_t out0u8 = vnclipu_wx_u8m1(vreinterpret_v_i16m1_u16m2(vmax_vx_i16m1(out0, 0, vl)), 0, vl);
-      vuint8m1_t out7u8 = vnclipu_wx_u8m1(vreinterpret_v_i16m1_u16m2(vmax_vx_i16m1(out7, 0, vl)), 0, vl);
-      vuint8m1_t out1u8 = vnclipu_wx_u8m1(vreinterpret_v_i16m1_u16m2(vmax_vx_i16m1(out1, 0, vl)), 0, vl);
-      vuint8m1_t out6u8 = vnclipu_wx_u8m1(vreinterpret_v_i16m1_u16m2(vmax_vx_i16m1(out6, 0, vl)), 0, vl);
-      vuint8m1_t out2u8 = vnclipu_wx_u8m1(vreinterpret_v_i16m1_u16m2(vmax_vx_i16m1(out2, 0, vl)), 0, vl);
-      vuint8m1_t out5u8 = vnclipu_wx_u8m1(vreinterpret_v_i16m1_u16m2(vmax_vx_i16m1(out5, 0, vl)), 0, vl);
-      vuint8m1_t out3u8 = vnclipu_wx_u8m1(vreinterpret_v_i16m1_u16m2(vmax_vx_i16m1(out3, 0, vl)), 0, vl);
-      vuint8m1_t out4u8 = vnclipu_wx_u8m1(vreinterpret_v_i16m1_u16m2(vmax_vx_i16m1(out4, 0, vl)), 0, vl);
-
-      // 8x8 transpose
-      vsse8_v_u8m1(out + 0, out_stride, out0u8, vl);
-      vsse8_v_u8m1(out + 1, out_stride, out1u8, vl);
-      vsse8_v_u8m1(out + 2, out_stride, out2u8, vl);
-      vsse8_v_u8m1(out + 3, out_stride, out3u8, vl);
-      vsse8_v_u8m1(out + 4, out_stride, out4u8, vl);
-      vsse8_v_u8m1(out + 5, out_stride, out5u8, vl);
-      vsse8_v_u8m1(out + 6, out_stride, out6u8, vl);
-      vsse8_v_u8m1(out + 7, out_stride, out7u8, vl);
-   }
-}
-
-#endif // STBI_RVV
-
-#define STBI__MARKER_none  0xff
-// if there's a pending marker from the entropy stream, return that
-// otherwise, fetch from the stream and get a marker. if there's no
-// marker, return 0xff, which is never a valid marker value
-static stbi_uc stbi__get_marker(stbi__jpeg *j)
-{
-   stbi_uc x;
-   if (j->marker != STBI__MARKER_none) { x = j->marker; j->marker = STBI__MARKER_none; return x; }
-   x = stbi__get8(j->s);
-   if (x != 0xff) return STBI__MARKER_none;
-   while (x == 0xff)
-      x = stbi__get8(j->s); // consume repeated 0xff fill bytes
-   return x;
-}
-
-// in each scan, we'll have scan_n components, and the order
-// of the components is specified by order[]
-#define STBI__RESTART(x)     ((x) >= 0xd0 && (x) <= 0xd7)
-
-// after a restart interval, stbi__jpeg_reset the entropy decoder and
-// the dc prediction
-static void stbi__jpeg_reset(stbi__jpeg *j)
-{
-   j->code_bits = 0;
-   j->code_buffer = 0;
-   j->nomore = 0;
-   j->img_comp[0].dc_pred = j->img_comp[1].dc_pred = j->img_comp[2].dc_pred = j->img_comp[3].dc_pred = 0;
-   j->marker = STBI__MARKER_none;
-   j->todo = j->restart_interval ? j->restart_interval : 0x7fffffff;
-   j->eob_run = 0;
-   // no more than 1<<31 MCUs if no restart_interal? that's plenty safe,
-   // since we don't even allow 1<<30 pixels
-}
-
-static int stbi__parse_entropy_coded_data(stbi__jpeg *z)
-{
-   stbi__jpeg_reset(z);
-   if (!z->progressive) {
-      if (z->scan_n == 1) {
-         int i,j;
-         STBI_SIMD_ALIGN(short, data[64]);
-         int n = z->order[0];
-         // non-interleaved data, we just need to process one block at a time,
-         // in trivial scanline order
-         // number of blocks to do just depends on how many actual "pixels" this
-         // component has, independent of interleaved MCU blocking and such
-         int w = (z->img_comp[n].x+7) >> 3;
-         int h = (z->img_comp[n].y+7) >> 3;
-         for (j=0; j < h; ++j) {
-            for (i=0; i < w; ++i) {
-               int ha = z->img_comp[n].ha;
-               if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0;
-               z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*j*8+i*8, z->img_comp[n].w2, data);
-               // every data block is an MCU, so countdown the restart interval
-               if (--z->todo <= 0) {
-                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
-                  // if it's NOT a restart, then just bail, so we get corrupt data
-                  // rather than no data
-                  if (!STBI__RESTART(z->marker)) return 1;
-                  stbi__jpeg_reset(z);
-               }
-            }
-         }
-         return 1;
-      } else { // interleaved
-         int i,j,k,x,y;
-         STBI_SIMD_ALIGN(short, data[64]);
-         for (j=0; j < z->img_mcu_y; ++j) {
-            for (i=0; i < z->img_mcu_x; ++i) {
-               // scan an interleaved mcu... process scan_n components in order
-               for (k=0; k < z->scan_n; ++k) {
-                  int n = z->order[k];
-                  // scan out an mcu's worth of this component; that's just determined
-                  // by the basic H and V specified for the component
-                  for (y=0; y < z->img_comp[n].v; ++y) {
-                     for (x=0; x < z->img_comp[n].h; ++x) {
-                        int x2 = (i*z->img_comp[n].h + x)*8;
-                        int y2 = (j*z->img_comp[n].v + y)*8;
-                        int ha = z->img_comp[n].ha;
-                        if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0;
-                        z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*y2+x2, z->img_comp[n].w2, data);
-                     }
-                  }
-               }
-               // after all interleaved components, that's an interleaved MCU,
-               // so now count down the restart interval
-               if (--z->todo <= 0) {
-                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
-                  if (!STBI__RESTART(z->marker)) return 1;
-                  stbi__jpeg_reset(z);
-               }
-            }
-         }
-         return 1;
-      }
-   } else {
-      if (z->scan_n == 1) {
-         int i,j;
-         int n = z->order[0];
-         // non-interleaved data, we just need to process one block at a time,
-         // in trivial scanline order
-         // number of blocks to do just depends on how many actual "pixels" this
-         // component has, independent of interleaved MCU blocking and such
-         int w = (z->img_comp[n].x+7) >> 3;
-         int h = (z->img_comp[n].y+7) >> 3;
-         for (j=0; j < h; ++j) {
-            for (i=0; i < w; ++i) {
-               short *data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w);
-               if (z->spec_start == 0) {
-                  if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n))
-                     return 0;
-               } else {
-                  int ha = z->img_comp[n].ha;
-                  if (!stbi__jpeg_decode_block_prog_ac(z, data, &z->huff_ac[ha], z->fast_ac[ha]))
-                     return 0;
-               }
-               // every data block is an MCU, so countdown the restart interval
-               if (--z->todo <= 0) {
-                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
-                  if (!STBI__RESTART(z->marker)) return 1;
-                  stbi__jpeg_reset(z);
-               }
-            }
-         }
-         return 1;
-      } else { // interleaved
-         int i,j,k,x,y;
-         for (j=0; j < z->img_mcu_y; ++j) {
-            for (i=0; i < z->img_mcu_x; ++i) {
-               // scan an interleaved mcu... process scan_n components in order
-               for (k=0; k < z->scan_n; ++k) {
-                  int n = z->order[k];
-                  // scan out an mcu's worth of this component; that's just determined
-                  // by the basic H and V specified for the component
-                  for (y=0; y < z->img_comp[n].v; ++y) {
-                     for (x=0; x < z->img_comp[n].h; ++x) {
-                        int x2 = (i*z->img_comp[n].h + x);
-                        int y2 = (j*z->img_comp[n].v + y);
-                        short *data = z->img_comp[n].coeff + 64 * (x2 + y2 * z->img_comp[n].coeff_w);
-                        if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n))
-                           return 0;
-                     }
-                  }
-               }
-               // after all interleaved components, that's an interleaved MCU,
-               // so now count down the restart interval
-               if (--z->todo <= 0) {
-                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
-                  if (!STBI__RESTART(z->marker)) return 1;
-                  stbi__jpeg_reset(z);
-               }
-            }
-         }
-         return 1;
-      }
-   }
-}
-
-static void stbi__jpeg_dequantize(short *data, stbi__uint16 *dequant)
-{
-   int i;
-   for (i=0; i < 64; ++i)
-      data[i] *= dequant[i];
-}
-
-static void stbi__jpeg_finish(stbi__jpeg *z)
-{
-   if (z->progressive) {
-      // dequantize and idct the data
-      int i,j,n;
-      for (n=0; n < z->s->img_n; ++n) {
-         int w = (z->img_comp[n].x+7) >> 3;
-         int h = (z->img_comp[n].y+7) >> 3;
-         for (j=0; j < h; ++j) {
-            for (i=0; i < w; ++i) {
-               short *data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w);
-               stbi__jpeg_dequantize(data, z->dequant[z->img_comp[n].tq]);
-               z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*j*8+i*8, z->img_comp[n].w2, data);
-            }
-         }
-      }
-   }
-}
-
-static int stbi__process_marker(stbi__jpeg *z, int m)
-{
-   int L;
-   switch (m) {
-      case STBI__MARKER_none: // no marker found
-         return stbi__err("expected marker","Corrupt JPEG");
-
-      case 0xDD: // DRI - specify restart interval
-         if (stbi__get16be(z->s) != 4) return stbi__err("bad DRI len","Corrupt JPEG");
-         z->restart_interval = stbi__get16be(z->s);
-         return 1;
-
-      case 0xDB: // DQT - define quantization table
-         L = stbi__get16be(z->s)-2;
-         while (L > 0) {
-            int q = stbi__get8(z->s);
-            int p = q >> 4, sixteen = (p != 0);
-            int t = q & 15,i;
-            if (p != 0 && p != 1) return stbi__err("bad DQT type","Corrupt JPEG");
-            if (t > 3) return stbi__err("bad DQT table","Corrupt JPEG");
-
-            for (i=0; i < 64; ++i)
-               z->dequant[t][stbi__jpeg_dezigzag[i]] = (stbi__uint16)(sixteen ? stbi__get16be(z->s) : stbi__get8(z->s));
-            L -= (sixteen ? 129 : 65);
-         }
-         return L==0;
-
-      case 0xC4: // DHT - define huffman table
-         L = stbi__get16be(z->s)-2;
-         while (L > 0) {
-            stbi_uc *v;
-            int sizes[16],i,n=0;
-            int q = stbi__get8(z->s);
-            int tc = q >> 4;
-            int th = q & 15;
-            if (tc > 1 || th > 3) return stbi__err("bad DHT header","Corrupt JPEG");
-            for (i=0; i < 16; ++i) {
-               sizes[i] = stbi__get8(z->s);
-               n += sizes[i];
-            }
-            if(n > 256) return stbi__err("bad DHT header","Corrupt JPEG"); // Loop over i < n would write past end of values!
-            L -= 17;
-            if (tc == 0) {
-               if (!stbi__build_huffman(z->huff_dc+th, sizes)) return 0;
-               v = z->huff_dc[th].values;
-            } else {
-               if (!stbi__build_huffman(z->huff_ac+th, sizes)) return 0;
-               v = z->huff_ac[th].values;
-            }
-            for (i=0; i < n; ++i)
-               v[i] = stbi__get8(z->s);
-            if (tc != 0)
-               stbi__build_fast_ac(z->fast_ac[th], z->huff_ac + th);
-            L -= n;
-         }
-         return L==0;
-   }
-
-   // check for comment block or APP blocks
-   if ((m >= 0xE0 && m <= 0xEF) || m == 0xFE) {
-      L = stbi__get16be(z->s);
-      if (L < 2) {
-         if (m == 0xFE)
-            return stbi__err("bad COM len","Corrupt JPEG");
-         else
-            return stbi__err("bad APP len","Corrupt JPEG");
-      }
-      L -= 2;
-
-      if (m == 0xE0 && L >= 5) { // JFIF APP0 segment
-         static const unsigned char tag[5] = {'J','F','I','F','\0'};
-         int ok = 1;
-         int i;
-         for (i=0; i < 5; ++i)
-            if (stbi__get8(z->s) != tag[i])
-               ok = 0;
-         L -= 5;
-         if (ok)
-            z->jfif = 1;
-      } else if (m == 0xEE && L >= 12) { // Adobe APP14 segment
-         static const unsigned char tag[6] = {'A','d','o','b','e','\0'};
-         int ok = 1;
-         int i;
-         for (i=0; i < 6; ++i)
-            if (stbi__get8(z->s) != tag[i])
-               ok = 0;
-         L -= 6;
-         if (ok) {
-            stbi__get8(z->s); // version
-            stbi__get16be(z->s); // flags0
-            stbi__get16be(z->s); // flags1
-            z->app14_color_transform = stbi__get8(z->s); // color transform
-            L -= 6;
-         }
-      }
-
-      stbi__skip(z->s, L);
-      return 1;
-   }
-
-   return stbi__err("unknown marker","Corrupt JPEG");
-}
-
-// after we see SOS
-static int stbi__process_scan_header(stbi__jpeg *z)
-{
-   int i;
-   int Ls = stbi__get16be(z->s);
-   z->scan_n = stbi__get8(z->s);
-   if (z->scan_n < 1 || z->scan_n > 4 || z->scan_n > (int) z->s->img_n) return stbi__err("bad SOS component count","Corrupt JPEG");
-   if (Ls != 6+2*z->scan_n) return stbi__err("bad SOS len","Corrupt JPEG");
-   for (i=0; i < z->scan_n; ++i) {
-      int id = stbi__get8(z->s), which;
-      int q = stbi__get8(z->s);
-      for (which = 0; which < z->s->img_n; ++which)
-         if (z->img_comp[which].id == id)
-            break;
-      if (which == z->s->img_n) return 0; // no match
-      z->img_comp[which].hd = q >> 4;   if (z->img_comp[which].hd > 3) return stbi__err("bad DC huff","Corrupt JPEG");
-      z->img_comp[which].ha = q & 15;   if (z->img_comp[which].ha > 3) return stbi__err("bad AC huff","Corrupt JPEG");
-      z->order[i] = which;
-   }
-
-   {
-      int aa;
-      z->spec_start = stbi__get8(z->s);
-      z->spec_end   = stbi__get8(z->s); // should be 63, but might be 0
-      aa = stbi__get8(z->s);
-      z->succ_high = (aa >> 4);
-      z->succ_low  = (aa & 15);
-      if (z->progressive) {
-         if (z->spec_start > 63 || z->spec_end > 63  || z->spec_start > z->spec_end || z->succ_high > 13 || z->succ_low > 13)
-            return stbi__err("bad SOS", "Corrupt JPEG");
-      } else {
-         if (z->spec_start != 0) return stbi__err("bad SOS","Corrupt JPEG");
-         if (z->succ_high != 0 || z->succ_low != 0) return stbi__err("bad SOS","Corrupt JPEG");
-         z->spec_end = 63;
-      }
-   }
-
-   return 1;
-}
-
-static int stbi__free_jpeg_components(stbi__jpeg *z, int ncomp, int why)
-{
-   int i;
-   for (i=0; i < ncomp; ++i) {
-      if (z->img_comp[i].raw_data) {
-         STBI_FREE(z->img_comp[i].raw_data);
-         z->img_comp[i].raw_data = NULL;
-         z->img_comp[i].data = NULL;
-      }
-      if (z->img_comp[i].raw_coeff) {
-         STBI_FREE(z->img_comp[i].raw_coeff);
-         z->img_comp[i].raw_coeff = 0;
-         z->img_comp[i].coeff = 0;
-      }
-      if (z->img_comp[i].linebuf) {
-         STBI_FREE(z->img_comp[i].linebuf);
-         z->img_comp[i].linebuf = NULL;
-      }
-   }
-   return why;
-}
-
-static int stbi__process_frame_header(stbi__jpeg *z, int scan)
-{
-   stbi__context *s = z->s;
-   int Lf,p,i,q, h_max=1,v_max=1,c;
-   Lf = stbi__get16be(s);         if (Lf < 11) return stbi__err("bad SOF len","Corrupt JPEG"); // JPEG
-   p  = stbi__get8(s);            if (p != 8) return stbi__err("only 8-bit","JPEG format not supported: 8-bit only"); // JPEG baseline
-   s->img_y = stbi__get16be(s);   if (s->img_y == 0) return stbi__err("no header height", "JPEG format not supported: delayed height"); // Legal, but we don't handle it--but neither does IJG
-   s->img_x = stbi__get16be(s);   if (s->img_x == 0) return stbi__err("0 width","Corrupt JPEG"); // JPEG requires
-   if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
-   if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
-   c = stbi__get8(s);
-   if (c != 3 && c != 1 && c != 4) return stbi__err("bad component count","Corrupt JPEG");
-   s->img_n = c;
-   for (i=0; i < c; ++i) {
-      z->img_comp[i].data = NULL;
-      z->img_comp[i].linebuf = NULL;
-   }
-
-   if (Lf != 8+3*s->img_n) return stbi__err("bad SOF len","Corrupt JPEG");
-
-   z->rgb = 0;
-   for (i=0; i < s->img_n; ++i) {
-      static const unsigned char rgb[3] = { 'R', 'G', 'B' };
-      z->img_comp[i].id = stbi__get8(s);
-      if (s->img_n == 3 && z->img_comp[i].id == rgb[i])
-         ++z->rgb;
-      q = stbi__get8(s);
-      z->img_comp[i].h = (q >> 4);  if (!z->img_comp[i].h || z->img_comp[i].h > 4) return stbi__err("bad H","Corrupt JPEG");
-      z->img_comp[i].v = q & 15;    if (!z->img_comp[i].v || z->img_comp[i].v > 4) return stbi__err("bad V","Corrupt JPEG");
-      z->img_comp[i].tq = stbi__get8(s);  if (z->img_comp[i].tq > 3) return stbi__err("bad TQ","Corrupt JPEG");
-   }
-
-   if (scan != STBI__SCAN_load) return 1;
-
-   if (!stbi__mad3sizes_valid(s->img_x, s->img_y, s->img_n, 0)) return stbi__err("too large", "Image too large to decode");
-
-   for (i=0; i < s->img_n; ++i) {
-      if (z->img_comp[i].h > h_max) h_max = z->img_comp[i].h;
-      if (z->img_comp[i].v > v_max) v_max = z->img_comp[i].v;
-   }
-
-   // check that plane subsampling factors are integer ratios; our resamplers can't deal with fractional ratios
-   // and I've never seen a non-corrupted JPEG file actually use them
-   for (i=0; i < s->img_n; ++i) {
-      if (h_max % z->img_comp[i].h != 0) return stbi__err("bad H","Corrupt JPEG");
-      if (v_max % z->img_comp[i].v != 0) return stbi__err("bad V","Corrupt JPEG");
-   }
-
-   // compute interleaved mcu info
-   z->img_h_max = h_max;
-   z->img_v_max = v_max;
-   z->img_mcu_w = h_max * 8;
-   z->img_mcu_h = v_max * 8;
-   // these sizes can't be more than 17 bits
-   z->img_mcu_x = (s->img_x + z->img_mcu_w-1) / z->img_mcu_w;
-   z->img_mcu_y = (s->img_y + z->img_mcu_h-1) / z->img_mcu_h;
-
-   for (i=0; i < s->img_n; ++i) {
-      // number of effective pixels (e.g. for non-interleaved MCU)
-      z->img_comp[i].x = (s->img_x * z->img_comp[i].h + h_max-1) / h_max;
-      z->img_comp[i].y = (s->img_y * z->img_comp[i].v + v_max-1) / v_max;
-      // to simplify generation, we'll allocate enough memory to decode
-      // the bogus oversized data from using interleaved MCUs and their
-      // big blocks (e.g. a 16x16 iMCU on an image of width 33); we won't
-      // discard the extra data until colorspace conversion
-      //
-      // img_mcu_x, img_mcu_y: <=17 bits; comp[i].h and .v are <=4 (checked earlier)
-      // so these muls can't overflow with 32-bit ints (which we require)
-      z->img_comp[i].w2 = z->img_mcu_x * z->img_comp[i].h * 8;
-      z->img_comp[i].h2 = z->img_mcu_y * z->img_comp[i].v * 8;
-      z->img_comp[i].coeff = 0;
-      z->img_comp[i].raw_coeff = 0;
-      z->img_comp[i].linebuf = NULL;
-      z->img_comp[i].raw_data = stbi__malloc_mad2(z->img_comp[i].w2, z->img_comp[i].h2, 15);
-      if (z->img_comp[i].raw_data == NULL)
-         return stbi__free_jpeg_components(z, i+1, stbi__err("outofmem", "Out of memory"));
-      // align blocks for idct using mmx/sse
-      z->img_comp[i].data = (stbi_uc*) (((size_t) z->img_comp[i].raw_data + 15) & ~15);
-      if (z->progressive) {
-         // w2, h2 are multiples of 8 (see above)
-         z->img_comp[i].coeff_w = z->img_comp[i].w2 / 8;
-         z->img_comp[i].coeff_h = z->img_comp[i].h2 / 8;
-         z->img_comp[i].raw_coeff = stbi__malloc_mad3(z->img_comp[i].w2, z->img_comp[i].h2, sizeof(short), 15);
-         if (z->img_comp[i].raw_coeff == NULL)
-            return stbi__free_jpeg_components(z, i+1, stbi__err("outofmem", "Out of memory"));
-         z->img_comp[i].coeff = (short*) (((size_t) z->img_comp[i].raw_coeff + 15) & ~15);
-      }
-   }
-
-   return 1;
-}
-
-// use comparisons since in some cases we handle more than one case (e.g. SOF)
-#define stbi__DNL(x)         ((x) == 0xdc)
-#define stbi__SOI(x)         ((x) == 0xd8)
-#define stbi__EOI(x)         ((x) == 0xd9)
-#define stbi__SOF(x)         ((x) == 0xc0 || (x) == 0xc1 || (x) == 0xc2)
-#define stbi__SOS(x)         ((x) == 0xda)
-
-#define stbi__SOF_progressive(x)   ((x) == 0xc2)
-
-static int stbi__decode_jpeg_header(stbi__jpeg *z, int scan)
-{
-   int m;
-   z->jfif = 0;
-   z->app14_color_transform = -1; // valid values are 0,1,2
-   z->marker = STBI__MARKER_none; // initialize cached marker to empty
-   m = stbi__get_marker(z);
-   if (!stbi__SOI(m)) return stbi__err("no SOI","Corrupt JPEG");
-   if (scan == STBI__SCAN_type) return 1;
-   m = stbi__get_marker(z);
-   while (!stbi__SOF(m)) {
-      if (!stbi__process_marker(z,m)) return 0;
-      m = stbi__get_marker(z);
-      while (m == STBI__MARKER_none) {
-         // some files have extra padding after their blocks, so ok, we'll scan
-         if (stbi__at_eof(z->s)) return stbi__err("no SOF", "Corrupt JPEG");
-         m = stbi__get_marker(z);
-      }
-   }
-   z->progressive = stbi__SOF_progressive(m);
-   if (!stbi__process_frame_header(z, scan)) return 0;
-   return 1;
-}
-
-static int stbi__skip_jpeg_junk_at_end(stbi__jpeg *j)
-{
-   // some JPEGs have junk at end, skip over it but if we find what looks
-   // like a valid marker, resume there
-   while (!stbi__at_eof(j->s)) {
-      int x = stbi__get8(j->s);
-      while (x == 255) { // might be a marker
-         if (stbi__at_eof(j->s)) return STBI__MARKER_none;
-         x = stbi__get8(j->s);
-         if (x != 0x00 && x != 0xff) {
-            // not a stuffed zero or lead-in to another marker, looks
-            // like an actual marker, return it
-            return x;
-         }
-         // stuffed zero has x=0 now which ends the loop, meaning we go
-         // back to regular scan loop.
-         // repeated 0xff keeps trying to read the next byte of the marker.
-      }
-   }
-   return STBI__MARKER_none;
-}
-
-// decode image to YCbCr format
-static int stbi__decode_jpeg_image(stbi__jpeg *j)
-{
-   int m;
-   for (m = 0; m < 4; m++) {
-      j->img_comp[m].raw_data = NULL;
-      j->img_comp[m].raw_coeff = NULL;
-   }
-   j->restart_interval = 0;
-   if (!stbi__decode_jpeg_header(j, STBI__SCAN_load)) return 0;
-   m = stbi__get_marker(j);
-   while (!stbi__EOI(m)) {
-      if (stbi__SOS(m)) {
-         if (!stbi__process_scan_header(j)) return 0;
-         if (!stbi__parse_entropy_coded_data(j)) return 0;
-         if (j->marker == STBI__MARKER_none ) {
-         j->marker = stbi__skip_jpeg_junk_at_end(j);
-            // if we reach eof without hitting a marker, stbi__get_marker() below will fail and we'll eventually return 0
-         }
-         m = stbi__get_marker(j);
-         if (STBI__RESTART(m))
-            m = stbi__get_marker(j);
-      } else if (stbi__DNL(m)) {
-         int Ld = stbi__get16be(j->s);
-         stbi__uint32 NL = stbi__get16be(j->s);
-         if (Ld != 4) return stbi__err("bad DNL len", "Corrupt JPEG");
-         if (NL != j->s->img_y) return stbi__err("bad DNL height", "Corrupt JPEG");
-         m = stbi__get_marker(j);
-      } else {
-         if (!stbi__process_marker(j, m)) return 1;
-         m = stbi__get_marker(j);
-      }
-   }
-   if (j->progressive)
-      stbi__jpeg_finish(j);
-   return 1;
-}
-
-// static jfif-centered resampling (across block boundaries)
-
-typedef stbi_uc *(*resample_row_func)(stbi_uc *out, stbi_uc *in0, stbi_uc *in1,
-                                    int w, int hs);
-
-#define stbi__div4(x) ((stbi_uc) ((x) >> 2))
-
-static stbi_uc *resample_row_1(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
-{
-   STBI_NOTUSED(out);
-   STBI_NOTUSED(in_far);
-   STBI_NOTUSED(w);
-   STBI_NOTUSED(hs);
-   return in_near;
-}
-
-static stbi_uc* stbi__resample_row_v_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
-{
-   // need to generate two samples vertically for every one in input
-   int i;
-   STBI_NOTUSED(hs);
-   for (i=0; i < w; ++i)
-      out[i] = stbi__div4(3*in_near[i] + in_far[i] + 2);
-   return out;
-}
-
-static stbi_uc*  stbi__resample_row_h_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
-{
-   // need to generate two samples horizontally for every one in input
-   int i;
-   stbi_uc *input = in_near;
-
-   if (w == 1) {
-      // if only one sample, can't do any interpolation
-      out[0] = out[1] = input[0];
-      return out;
-   }
-
-   out[0] = input[0];
-   out[1] = stbi__div4(input[0]*3 + input[1] + 2);
-   for (i=1; i < w-1; ++i) {
-      int n = 3*input[i]+2;
-      out[i*2+0] = stbi__div4(n+input[i-1]);
-      out[i*2+1] = stbi__div4(n+input[i+1]);
-   }
-   out[i*2+0] = stbi__div4(input[w-2]*3 + input[w-1] + 2);
-   out[i*2+1] = input[w-1];
-
-   STBI_NOTUSED(in_far);
-   STBI_NOTUSED(hs);
-
-   return out;
-}
-
-#define stbi__div16(x) ((stbi_uc) ((x) >> 4))
-
-static stbi_uc *stbi__resample_row_hv_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
-{
-   // need to generate 2x2 samples for every one in input
-   int i,t0,t1;
-   if (w == 1) {
-      out[0] = out[1] = stbi__div4(3*in_near[0] + in_far[0] + 2);
-      return out;
-   }
-
-   t1 = 3*in_near[0] + in_far[0];
-   out[0] = stbi__div4(t1+2);
-   for (i=1; i < w; ++i) {
-      t0 = t1;
-      t1 = 3*in_near[i]+in_far[i];
-      out[i*2-1] = stbi__div16(3*t0 + t1 + 8);
-      out[i*2  ] = stbi__div16(3*t1 + t0 + 8);
-   }
-   out[w*2-1] = stbi__div4(t1+2);
-
-   STBI_NOTUSED(hs);
-
-   return out;
-}
-
-#if defined(STBI_SSE2) || defined(STBI_NEON) || defined(STBI_RVV)
-static stbi_uc *stbi__resample_row_hv_2_simd(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
-{
-   // need to generate 2x2 samples for every one in input
-   int i=0,t0,t1;
-
-   if (w == 1) {
-      out[0] = out[1] = stbi__div4(3*in_near[0] + in_far[0] + 2);
-      return out;
-   }
-
-   t1 = 3*in_near[0] + in_far[0];
-#if defined(STBI_RVV)
-   // process groups of vl*4 pixels for as long as we can.
-   // note we can't handle the last pixel in a row in this loop
-   // because we need to handle the filter boundary conditions.
-   int n = w-1;
-   while (n > 0) {
-      size_t vl = vsetvl_e8m4(n);
-
-      // load and perform the vertical filtering pass
-      vuint8m4_t farb  = vle8_v_u8m4(in_far + i, vl);
-      vuint8m4_t nearb = vle8_v_u8m4(in_near + i, vl);
-      vuint16m8_t curr  = vadd_vv_u16m8(vwmulu_vx_u16m8(nearb, 3, vl), vwcvtu_x_x_v_u16m8(farb, vl), vl); // current row
-
-      // horizontal filter works the same based on shifted vers of current
-      // row. "prev" is current row shifted right by 1 pixel; we need to
-      // insert the previous pixel value (from t1).
-      // "next" is current row shifted left by 1 pixel, with first pixel
-      // of next block of vl*4 pixels added in.
-      vuint16m8_t prev = vslide1up_vx_u16m8(curr, t1, vl);
-      vuint16m8_t next = vslide1down_vx_u16m8(curr, 3*in_near[i+vl] + in_far[i+vl], vl);
-
-      // horizontal filter, polyphase implementation since it's convenient:
-      // even pixels = 3*cur + prev + 8
-      // odd  pixels = 3*cur + next + 8
-      // note the shared term.
-      vuint16m8_t curs = vmacc_vx_u16m8(vmv_v_x_u16m8(8, vl), 3, curr, vl);
-      vuint16m8_t even = vadd_vv_u16m8(curs, prev, vl);
-      vuint16m8_t odd  = vadd_vv_u16m8(curs, next, vl);
-
-      // undo scaling and round, then store with even/odd phases interleaved
-      vuint8m4_t evenu8 = vnclipu_wx_u8m4(vsrl_vx_u16m8(even, 4, vl), 0, vl);
-      vuint8m4_t oddu8 = vnclipu_wx_u8m4(vsrl_vx_u16m8(odd, 4, vl), 0, vl);
-      vuint8m4x2_t o = vcreate_u8m4x2(evenu8, oddu8);
-      vsseg2e8_v_u8m4x2(out + i*2, o, vl);
-
-      // "previous" value for next iter
-      t1 = 3*in_near[i+(vl-1)] + in_far[i+(vl-1)];
-
-      i += vl;
-      n -= vl;
-   }
-#else
-   // process groups of 8 pixels for as long as we can.
-   // note we can't handle the last pixel in a row in this loop
-   // because we need to handle the filter boundary conditions.
-   for (; i < ((w-1) & ~7); i += 8) {
-#if defined(STBI_SSE2)
-      // load and perform the vertical filtering pass
-      // this uses 3*x + y = 4*x + (y - x)
-      __m128i zero  = _mm_setzero_si128();
-      __m128i farb  = _mm_loadl_epi64((__m128i *) (in_far + i));
-      __m128i nearb = _mm_loadl_epi64((__m128i *) (in_near + i));
-      __m128i farw  = _mm_unpacklo_epi8(farb, zero);
-      __m128i nearw = _mm_unpacklo_epi8(nearb, zero);
-      __m128i diff  = _mm_sub_epi16(farw, nearw);
-      __m128i nears = _mm_slli_epi16(nearw, 2);
-      __m128i curr  = _mm_add_epi16(nears, diff); // current row
-
-      // horizontal filter works the same based on shifted vers of current
-      // row. "prev" is current row shifted right by 1 pixel; we need to
-      // insert the previous pixel value (from t1).
-      // "next" is current row shifted left by 1 pixel, with first pixel
-      // of next block of 8 pixels added in.
-      __m128i prv0 = _mm_slli_si128(curr, 2);
-      __m128i nxt0 = _mm_srli_si128(curr, 2);
-      __m128i prev = _mm_insert_epi16(prv0, t1, 0);
-      __m128i next = _mm_insert_epi16(nxt0, 3*in_near[i+8] + in_far[i+8], 7);
-
-      // horizontal filter, polyphase implementation since it's convenient:
-      // even pixels = 3*cur + prev = cur*4 + (prev - cur)
-      // odd  pixels = 3*cur + next = cur*4 + (next - cur)
-      // note the shared term.
-      __m128i bias  = _mm_set1_epi16(8);
-      __m128i curs = _mm_slli_epi16(curr, 2);
-      __m128i prvd = _mm_sub_epi16(prev, curr);
-      __m128i nxtd = _mm_sub_epi16(next, curr);
-      __m128i curb = _mm_add_epi16(curs, bias);
-      __m128i even = _mm_add_epi16(prvd, curb);
-      __m128i odd  = _mm_add_epi16(nxtd, curb);
-
-      // interleave even and odd pixels, then undo scaling.
-      __m128i int0 = _mm_unpacklo_epi16(even, odd);
-      __m128i int1 = _mm_unpackhi_epi16(even, odd);
-      __m128i de0  = _mm_srli_epi16(int0, 4);
-      __m128i de1  = _mm_srli_epi16(int1, 4);
-
-      // pack and write output
-      __m128i outv = _mm_packus_epi16(de0, de1);
-      _mm_storeu_si128((__m128i *) (out + i*2), outv);
-#elif defined(STBI_NEON)
-      // load and perform the vertical filtering pass
-      // this uses 3*x + y = 4*x + (y - x)
-      uint8x8_t farb  = vld1_u8(in_far + i);
-      uint8x8_t nearb = vld1_u8(in_near + i);
-      int16x8_t diff  = vreinterpretq_s16_u16(vsubl_u8(farb, nearb));
-      int16x8_t nears = vreinterpretq_s16_u16(vshll_n_u8(nearb, 2));
-      int16x8_t curr  = vaddq_s16(nears, diff); // current row
-
-      // horizontal filter works the same based on shifted vers of current
-      // row. "prev" is current row shifted right by 1 pixel; we need to
-      // insert the previous pixel value (from t1).
-      // "next" is current row shifted left by 1 pixel, with first pixel
-      // of next block of 8 pixels added in.
-      int16x8_t prv0 = vextq_s16(curr, curr, 7);
-      int16x8_t nxt0 = vextq_s16(curr, curr, 1);
-      int16x8_t prev = vsetq_lane_s16(t1, prv0, 0);
-      int16x8_t next = vsetq_lane_s16(3*in_near[i+8] + in_far[i+8], nxt0, 7);
-
-      // horizontal filter, polyphase implementation since it's convenient:
-      // even pixels = 3*cur + prev = cur*4 + (prev - cur)
-      // odd  pixels = 3*cur + next = cur*4 + (next - cur)
-      // note the shared term.
-      int16x8_t curs = vshlq_n_s16(curr, 2);
-      int16x8_t prvd = vsubq_s16(prev, curr);
-      int16x8_t nxtd = vsubq_s16(next, curr);
-      int16x8_t even = vaddq_s16(curs, prvd);
-      int16x8_t odd  = vaddq_s16(curs, nxtd);
-
-      // undo scaling and round, then store with even/odd phases interleaved
-      uint8x8x2_t o;
-      o.val[0] = vqrshrun_n_s16(even, 4);
-      o.val[1] = vqrshrun_n_s16(odd,  4);
-      vst2_u8(out + i*2, o);
-#endif
-
-      // "previous" value for next iter
-      t1 = 3*in_near[i+7] + in_far[i+7];
-   }
-#endif
-
-   t0 = t1;
-   t1 = 3*in_near[i] + in_far[i];
-   out[i*2] = stbi__div16(3*t1 + t0 + 8);
-
-   for (++i; i < w; ++i) {
-      t0 = t1;
-      t1 = 3*in_near[i]+in_far[i];
-      out[i*2-1] = stbi__div16(3*t0 + t1 + 8);
-      out[i*2  ] = stbi__div16(3*t1 + t0 + 8);
-   }
-   out[w*2-1] = stbi__div4(t1+2);
-
-   STBI_NOTUSED(hs);
-
-   return out;
-}
-#endif
-
-static stbi_uc *stbi__resample_row_generic(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
-{
-   // resample with nearest-neighbor
-   int i,j;
-   STBI_NOTUSED(in_far);
-   for (i=0; i < w; ++i)
-      for (j=0; j < hs; ++j)
-         out[i*hs+j] = in_near[i];
-   return out;
-}
-
-// this is a reduced-precision calculation of YCbCr-to-RGB introduced
-// to make sure the code produces the same results in both SIMD and scalar
-#define stbi__float2fixed(x)  (((int) ((x) * 4096.0f + 0.5f)) << 8)
-static void stbi__YCbCr_to_RGB_row(stbi_uc *out, const stbi_uc *y, const stbi_uc *pcb, const stbi_uc *pcr, int count, int step)
-{
-   int i;
-   for (i=0; i < count; ++i) {
-      int y_fixed = (y[i] << 20) + (1<<19); // rounding
-      int r,g,b;
-      int cr = pcr[i] - 128;
-      int cb = pcb[i] - 128;
-      r = y_fixed +  cr* stbi__float2fixed(1.40200f);
-      g = y_fixed + (cr*-stbi__float2fixed(0.71414f)) + ((cb*-stbi__float2fixed(0.34414f)) & 0xffff0000);
-      b = y_fixed                                     +   cb* stbi__float2fixed(1.77200f);
-      r >>= 20;
-      g >>= 20;
-      b >>= 20;
-      if ((unsigned) r > 255) { if (r < 0) r = 0; else r = 255; }
-      if ((unsigned) g > 255) { if (g < 0) g = 0; else g = 255; }
-      if ((unsigned) b > 255) { if (b < 0) b = 0; else b = 255; }
-      out[0] = (stbi_uc)r;
-      out[1] = (stbi_uc)g;
-      out[2] = (stbi_uc)b;
-      out[3] = 255;
-      out += step;
-   }
-}
-
-#if defined(STBI_SSE2) || defined(STBI_NEON) || defined(STBI_RVV)
-static void stbi__YCbCr_to_RGB_simd(stbi_uc *out, stbi_uc const *y, stbi_uc const *pcb, stbi_uc const *pcr, int count, int step)
-{
-   int i = 0;
-
-#ifdef STBI_SSE2
-   // step == 3 is pretty ugly on the final interleave, and i'm not convinced
-   // it's useful in practice (you wouldn't use it for textures, for example).
-   // so just accelerate step == 4 case.
-   if (step == 4) {
-      // this is a fairly straightforward implementation and not super-optimized.
-      __m128i signflip  = _mm_set1_epi8(-0x80);
-      __m128i cr_const0 = _mm_set1_epi16(   (short) ( 1.40200f*4096.0f+0.5f));
-      __m128i cr_const1 = _mm_set1_epi16( - (short) ( 0.71414f*4096.0f+0.5f));
-      __m128i cb_const0 = _mm_set1_epi16( - (short) ( 0.34414f*4096.0f+0.5f));
-      __m128i cb_const1 = _mm_set1_epi16(   (short) ( 1.77200f*4096.0f+0.5f));
-      __m128i y_bias = _mm_set1_epi8((char) (unsigned char) 128);
-      __m128i xw = _mm_set1_epi16(255); // alpha channel
-
-      for (; i+7 < count; i += 8) {
-         // load
-         __m128i y_bytes = _mm_loadl_epi64((__m128i *) (y+i));
-         __m128i cr_bytes = _mm_loadl_epi64((__m128i *) (pcr+i));
-         __m128i cb_bytes = _mm_loadl_epi64((__m128i *) (pcb+i));
-         __m128i cr_biased = _mm_xor_si128(cr_bytes, signflip); // -128
-         __m128i cb_biased = _mm_xor_si128(cb_bytes, signflip); // -128
-
-         // unpack to short (and left-shift cr, cb by 8)
-         __m128i yw  = _mm_unpacklo_epi8(y_bias, y_bytes);
-         __m128i crw = _mm_unpacklo_epi8(_mm_setzero_si128(), cr_biased);
-         __m128i cbw = _mm_unpacklo_epi8(_mm_setzero_si128(), cb_biased);
-
-         // color transform
-         __m128i yws = _mm_srli_epi16(yw, 4);
-         __m128i cr0 = _mm_mulhi_epi16(cr_const0, crw);
-         __m128i cb0 = _mm_mulhi_epi16(cb_const0, cbw);
-         __m128i cb1 = _mm_mulhi_epi16(cbw, cb_const1);
-         __m128i cr1 = _mm_mulhi_epi16(crw, cr_const1);
-         __m128i rws = _mm_add_epi16(cr0, yws);
-         __m128i gwt = _mm_add_epi16(cb0, yws);
-         __m128i bws = _mm_add_epi16(yws, cb1);
-         __m128i gws = _mm_add_epi16(gwt, cr1);
-
-         // descale
-         __m128i rw = _mm_srai_epi16(rws, 4);
-         __m128i bw = _mm_srai_epi16(bws, 4);
-         __m128i gw = _mm_srai_epi16(gws, 4);
-
-         // back to byte, set up for transpose
-         __m128i brb = _mm_packus_epi16(rw, bw);
-         __m128i gxb = _mm_packus_epi16(gw, xw);
-
-         // transpose to interleave channels
-         __m128i t0 = _mm_unpacklo_epi8(brb, gxb);
-         __m128i t1 = _mm_unpackhi_epi8(brb, gxb);
-         __m128i o0 = _mm_unpacklo_epi16(t0, t1);
-         __m128i o1 = _mm_unpackhi_epi16(t0, t1);
-
-         // store
-         _mm_storeu_si128((__m128i *) (out + 0), o0);
-         _mm_storeu_si128((__m128i *) (out + 16), o1);
-         out += 32;
-      }
-   }
-#endif
-
-#ifdef STBI_NEON
-   if (step == 3) {
-      // this is a fairly straightforward implementation and not super-optimized.
-      uint8x8_t signflip = vdup_n_u8(0x80);
-      int16x8_t cr_const0 = vdupq_n_s16(   (short) ( 1.40200f*4096.0f+0.5f));
-      int16x8_t cr_const1 = vdupq_n_s16( - (short) ( 0.71414f*4096.0f+0.5f));
-      int16x8_t cb_const0 = vdupq_n_s16( - (short) ( 0.34414f*4096.0f+0.5f));
-      int16x8_t cb_const1 = vdupq_n_s16(   (short) ( 1.77200f*4096.0f+0.5f));
-
-      for (; i+7 < count; i += 8) {
-         // load
-         uint8x8_t y_bytes  = vld1_u8(y + i);
-         uint8x8_t cr_bytes = vld1_u8(pcr + i);
-         uint8x8_t cb_bytes = vld1_u8(pcb + i);
-         int8x8_t cr_biased = vreinterpret_s8_u8(vsub_u8(cr_bytes, signflip));
-         int8x8_t cb_biased = vreinterpret_s8_u8(vsub_u8(cb_bytes, signflip));
-
-         // expand to s16
-         int16x8_t yws = vreinterpretq_s16_u16(vshll_n_u8(y_bytes, 4));
-         int16x8_t crw = vshll_n_s8(cr_biased, 7);
-         int16x8_t cbw = vshll_n_s8(cb_biased, 7);
-
-         // color transform
-         int16x8_t cr0 = vqdmulhq_s16(crw, cr_const0);
-         int16x8_t cb0 = vqdmulhq_s16(cbw, cb_const0);
-         int16x8_t cr1 = vqdmulhq_s16(crw, cr_const1);
-         int16x8_t cb1 = vqdmulhq_s16(cbw, cb_const1);
-         int16x8_t rws = vaddq_s16(yws, cr0);
-         int16x8_t gws = vaddq_s16(vaddq_s16(yws, cb0), cr1);
-         int16x8_t bws = vaddq_s16(yws, cb1);
-
-         // undo scaling, round, convert to byte
-         uint8x8x3_t o;
-         o.val[0] = vqrshrun_n_s16(rws, 4);
-         o.val[1] = vqrshrun_n_s16(gws, 4);
-         o.val[2] = vqrshrun_n_s16(bws, 4);
-
-         // store, interleaving r/g/b
-         vst3_u8(out, o);
-         out += 8*3;
-      }
-   }
-   if (step == 4) {
-      // this is a fairly straightforward implementation and not super-optimized.
-      uint8x8_t signflip = vdup_n_u8(0x80);
-      int16x8_t cr_const0 = vdupq_n_s16(   (short) ( 1.40200f*4096.0f+0.5f));
-      int16x8_t cr_const1 = vdupq_n_s16( - (short) ( 0.71414f*4096.0f+0.5f));
-      int16x8_t cb_const0 = vdupq_n_s16( - (short) ( 0.34414f*4096.0f+0.5f));
-      int16x8_t cb_const1 = vdupq_n_s16(   (short) ( 1.77200f*4096.0f+0.5f));
-
-      for (; i+7 < count; i += 8) {
-         // load
-         uint8x8_t y_bytes  = vld1_u8(y + i);
-         uint8x8_t cr_bytes = vld1_u8(pcr + i);
-         uint8x8_t cb_bytes = vld1_u8(pcb + i);
-         int8x8_t cr_biased = vreinterpret_s8_u8(vsub_u8(cr_bytes, signflip));
-         int8x8_t cb_biased = vreinterpret_s8_u8(vsub_u8(cb_bytes, signflip));
-
-         // expand to s16
-         int16x8_t yws = vreinterpretq_s16_u16(vshll_n_u8(y_bytes, 4));
-         int16x8_t crw = vshll_n_s8(cr_biased, 7);
-         int16x8_t cbw = vshll_n_s8(cb_biased, 7);
-
-         // color transform
-         int16x8_t cr0 = vqdmulhq_s16(crw, cr_const0);
-         int16x8_t cb0 = vqdmulhq_s16(cbw, cb_const0);
-         int16x8_t cr1 = vqdmulhq_s16(crw, cr_const1);
-         int16x8_t cb1 = vqdmulhq_s16(cbw, cb_const1);
-         int16x8_t rws = vaddq_s16(yws, cr0);
-         int16x8_t gws = vaddq_s16(vaddq_s16(yws, cb0), cr1);
-         int16x8_t bws = vaddq_s16(yws, cb1);
-
-         // undo scaling, round, convert to byte
-         uint8x8x4_t o;
-         o.val[0] = vqrshrun_n_s16(rws, 4);
-         o.val[1] = vqrshrun_n_s16(gws, 4);
-         o.val[2] = vqrshrun_n_s16(bws, 4);
-         o.val[3] = vdup_n_u8(255);
-
-         // store, interleaving r/g/b/a
-         vst4_u8(out, o);
-         out += 8*4;
-      }
-   }
-#endif
-
-#ifdef STBI_RVV
-   if (step == 3) {
-      // this is a fairly straightforward implementation and not super-optimized.
-      const unsigned char signflip = 0x80;
-      const short cr_const0 =    (short) ( 1.40200f*4096.0f+0.5f);
-      const short cr_const1 =  - (short) ( 0.71414f*4096.0f+0.5f);
-      const short cb_const0 =  - (short) ( 0.34414f*4096.0f+0.5f);
-      const short cb_const1 =    (short) ( 1.77200f*4096.0f+0.5f);
-
-      int n = count;
-      while (n > 0) {
-         size_t vl = vsetvl_e8m2(n);
-
-         // load
-         vuint8m2_t y_bytes  = vle8_v_u8m2(y + i, vl);
-         vuint8m2_t cr_bytes = vle8_v_u8m2(pcr + i, vl);
-         vuint8m2_t cb_bytes = vle8_v_u8m2(pcb + i, vl);
-         vint8m2_t cr_biased = vreinterpret_v_u8m2_i8m2(vsub_vx_u8m2(cr_bytes, signflip, vl));
-         vint8m2_t cb_biased = vreinterpret_v_u8m2_i8m2(vsub_vx_u8m2(cb_bytes, signflip, vl));
-
-         // expand to s16
-         vint16m4_t yws = vadd_vx_i16m4(vsll_vx_i16m4(vreinterpret_v_u16m4_i16m4(vwcvtu_x_x_v_u16m4(y_bytes, vl)), 4, vl), 8, vl);
-         vint16m4_t crw = vsll_vx_i16m4(vwcvt_x_x_v_i16m4(cr_biased, vl), 8, vl);
-         vint16m4_t cbw = vsll_vx_i16m4(vwcvt_x_x_v_i16m4(cb_biased, vl), 8, vl);
-
-         // color transform
-         vint16m4_t cr0 = vnclip_wx_i16m4(vsra_vx_i32m8(vwmul_vx_i32m8(crw, cr_const0, vl), 16, vl), 0, vl);
-         vint16m4_t cb0 = vnclip_wx_i16m4(vsra_vx_i32m8(vwmul_vx_i32m8(cbw, cb_const0, vl), 16, vl), 0, vl);
-         vint16m4_t cr1 = vnclip_wx_i16m4(vsra_vx_i32m8(vwmul_vx_i32m8(crw, cr_const1, vl), 16, vl), 0, vl);
-         vint16m4_t cb1 = vnclip_wx_i16m4(vsra_vx_i32m8(vwmul_vx_i32m8(cbw, cb_const1, vl), 16, vl), 0, vl);
-         vint16m4_t rws = vadd_vv_i16m4(yws, cr0, vl);
-         vint16m4_t gws = vadd_vv_i16m4(vadd_vv_i16m4(yws, cb0, vl), cr1, vl);
-         vint16m4_t bws = vadd_vv_i16m4(yws, cb1, vl);
-
-         // undo scaling, round, convert to byte
-         vuint8m2_t rb = vnclipu_wx_u8m2(vreinterpret_v_i16m4_u16m4(vmax_vx_i16m4(vsra_vx_i16m4(rws, 4, vl), 0, vl)), 0, vl);
-         vuint8m2_t gb = vnclipu_wx_u8m2(vreinterpret_v_i16m4_u16m4(vmax_vx_i16m4(vsra_vx_i16m4(gws, 4, vl), 0, vl)), 0, vl);
-         vuint8m2_t bb = vnclipu_wx_u8m2(vreinterpret_v_i16m4_u16m4(vmax_vx_i16m4(vsra_vx_i16m4(bws, 4, vl), 0, vl)), 0, vl);
-
-         // store, interleaving r/g/b
-         vuint8m2x3_t o = vcreate_u8m2x3(rb, gb, bb);
-         vsseg3e8_v_u8m2x3(out, o, vl);
-         out += vl*3;
-
-         i += vl;
-         n -= vl;
-      }
-   }
-   if (step == 4) {
-      // this is a fairly straightforward implementation and not super-optimized.
-      const unsigned char signflip = 128;
-      const short cr_const0 =    (short) ( 1.40200f*4096.0f+0.5f);
-      const short cr_const1 =  - (short) ( 0.71414f*4096.0f+0.5f);
-      const short cb_const0 =  - (short) ( 0.34414f*4096.0f+0.5f);
-      const short cb_const1 =    (short) ( 1.77200f*4096.0f+0.5f);
-
-      int n = count;
-      while (n > 0) {
-         size_t vl = vsetvl_e8m1(n);
-
-         // load
-         vuint8m2_t y_bytes  = vle8_v_u8m2(y + i, vl);
-         vuint8m2_t cr_bytes = vle8_v_u8m2(pcr + i, vl);
-         vuint8m2_t cb_bytes = vle8_v_u8m2(pcb + i, vl);
-         vint8m2_t cr_biased = vreinterpret_v_u8m2_i8m2(vsub_vx_u8m2(cr_bytes, signflip, vl));
-         vint8m2_t cb_biased = vreinterpret_v_u8m2_i8m2(vsub_vx_u8m2(cb_bytes, signflip, vl));
-
-         // expand to s16
-         vint16m4_t yws = vadd_vx_i16m4(vsll_vx_i16m4(vreinterpret_v_u16m4_i16m4(vwcvtu_x_x_v_u16m4(y_bytes, vl)), 4, vl), 8, vl);
-         vint16m4_t crw = vsll_vx_i16m4(vwcvt_x_x_v_i16m4(cr_biased, vl), 8, vl);
-         vint16m4_t cbw = vsll_vx_i16m4(vwcvt_x_x_v_i16m4(cb_biased, vl), 8, vl);
-
-         // color transform
-         vint16m4_t cr0 = vnclip_wx_i16m4(vsra_vx_i32m8(vwmul_vx_i32m8(crw, cr_const0, vl), 16, vl), 0, vl);
-         vint16m4_t cb0 = vnclip_wx_i16m4(vsra_vx_i32m8(vwmul_vx_i32m8(cbw, cb_const0, vl), 16, vl), 0, vl);
-         vint16m4_t cr1 = vnclip_wx_i16m4(vsra_vx_i32m8(vwmul_vx_i32m8(crw, cr_const1, vl), 16, vl), 0, vl);
-         vint16m4_t cb1 = vnclip_wx_i16m4(vsra_vx_i32m8(vwmul_vx_i32m8(cbw, cb_const1, vl), 16, vl), 0, vl);
-         vint16m4_t rws = vadd_vv_i16m4(yws, cr0, vl);
-         vint16m4_t gws = vadd_vv_i16m4(vadd_vv_i16m4(yws, cb0, vl), cr1, vl);
-         vint16m4_t bws = vadd_vv_i16m4(yws, cb1, vl);
-
-         // undo scaling, round, convert to byte
-         vuint8m2_t rb = vnclipu_wx_u8m2(vreinterpret_v_i16m4_u16m4(vmax_vx_i16m4(vsra_vx_i16m4(rws, 4, vl), 0, vl)), 0, vl);
-         vuint8m2_t gb = vnclipu_wx_u8m2(vreinterpret_v_i16m4_u16m4(vmax_vx_i16m4(vsra_vx_i16m4(gws, 4, vl), 0, vl)), 0, vl);
-         vuint8m2_t bb = vnclipu_wx_u8m2(vreinterpret_v_i16m4_u16m4(vmax_vx_i16m4(vsra_vx_i16m4(bws, 4, vl), 0, vl)), 0, vl);
-         vuint8m2_t ab = vmv_v_x_u8m2(255, vl);
-
-         // store, interleaving r/g/b/a
-         vuint8m2x4_t o = vcreate_u8m2x4(rb, gb, bb, ab);
-         vsseg4e8_v_u8m2x4(out, o, vl);
-         out += vl*4;
-
-         i += vl;
-         n -= vl;
-      }
-   }
-#endif
-
-   for (; i < count; ++i) {
-      int y_fixed = (y[i] << 20) + (1<<19); // rounding
-      int r,g,b;
-      int cr = pcr[i] - 128;
-      int cb = pcb[i] - 128;
-      r = y_fixed + cr* stbi__float2fixed(1.40200f);
-      g = y_fixed + cr*-stbi__float2fixed(0.71414f) + ((cb*-stbi__float2fixed(0.34414f)) & 0xffff0000);
-      b = y_fixed                                   +   cb* stbi__float2fixed(1.77200f);
-      r >>= 20;
-      g >>= 20;
-      b >>= 20;
-      if ((unsigned) r > 255) { if (r < 0) r = 0; else r = 255; }
-      if ((unsigned) g > 255) { if (g < 0) g = 0; else g = 255; }
-      if ((unsigned) b > 255) { if (b < 0) b = 0; else b = 255; }
-      out[0] = (stbi_uc)r;
-      out[1] = (stbi_uc)g;
-      out[2] = (stbi_uc)b;
-      out[3] = 255;
-      out += step;
-   }
-}
-#endif
-
-// set up the kernels
-static void stbi__setup_jpeg(stbi__jpeg *j)
-{
-   j->idct_block_kernel = stbi__idct_block;
-   j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_row;
-   j->resample_row_hv_2_kernel = stbi__resample_row_hv_2;
-
-#ifdef STBI_SSE2
-   if (stbi__sse2_available()) {
-      j->idct_block_kernel = stbi__idct_simd;
-      j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd;
-      j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd;
-   }
-#endif
-
-#ifdef STBI_NEON
-   j->idct_block_kernel = stbi__idct_simd;
-   j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd;
-   j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd;
-#endif
-
-#ifdef STBI_RVV
-   j->idct_block_kernel = stbi__idct_simd;
-   j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd;
-   j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd;
-#endif
-}
-
-// clean up the temporary component buffers
-static void stbi__cleanup_jpeg(stbi__jpeg *j)
-{
-   stbi__free_jpeg_components(j, j->s->img_n, 0);
-}
-
-typedef struct
-{
-   resample_row_func resample;
-   stbi_uc *line0,*line1;
-   int hs,vs;   // expansion factor in each axis
-   int w_lores; // horizontal pixels pre-expansion
-   int ystep;   // how far through vertical expansion we are
-   int ypos;    // which pre-expansion row we're on
-} stbi__resample;
-
-// fast 0..255 * 0..255 => 0..255 rounded multiplication
-static stbi_uc stbi__blinn_8x8(stbi_uc x, stbi_uc y)
-{
-   unsigned int t = x*y + 128;
-   return (stbi_uc) ((t + (t >>8)) >> 8);
-}
-
-static stbi_uc *load_jpeg_image(stbi__jpeg *z, int *out_x, int *out_y, int *comp, int req_comp)
-{
-   int n, decode_n, is_rgb;
-   z->s->img_n = 0; // make stbi__cleanup_jpeg safe
-
-   // validate req_comp
-   if (req_comp < 0 || req_comp > 4) return stbi__errpuc("bad req_comp", "Internal error");
-
-   // load a jpeg image from whichever source, but leave in YCbCr format
-   if (!stbi__decode_jpeg_image(z)) { stbi__cleanup_jpeg(z); return NULL; }
-
-   // determine actual number of components to generate
-   n = req_comp ? req_comp : z->s->img_n >= 3 ? 3 : 1;
-
-   is_rgb = z->s->img_n == 3 && (z->rgb == 3 || (z->app14_color_transform == 0 && !z->jfif));
-
-   if (z->s->img_n == 3 && n < 3 && !is_rgb)
-      decode_n = 1;
-   else
-      decode_n = z->s->img_n;
-
-   // nothing to do if no components requested; check this now to avoid
-   // accessing uninitialized coutput[0] later
-   if (decode_n <= 0) { stbi__cleanup_jpeg(z); return NULL; }
-
-   // resample and color-convert
-   {
-      int k;
-      unsigned int i,j;
-      stbi_uc *output;
-      stbi_uc *coutput[4] = { NULL, NULL, NULL, NULL };
-
-      stbi__resample res_comp[4];
-
-      for (k=0; k < decode_n; ++k) {
-         stbi__resample *r = &res_comp[k];
-
-         // allocate line buffer big enough for upsampling off the edges
-         // with upsample factor of 4
-         z->img_comp[k].linebuf = (stbi_uc *) stbi__malloc(z->s->img_x + 3);
-         if (!z->img_comp[k].linebuf) { stbi__cleanup_jpeg(z); return stbi__errpuc("outofmem", "Out of memory"); }
-
-         r->hs      = z->img_h_max / z->img_comp[k].h;
-         r->vs      = z->img_v_max / z->img_comp[k].v;
-         r->ystep   = r->vs >> 1;
-         r->w_lores = (z->s->img_x + r->hs-1) / r->hs;
-         r->ypos    = 0;
-         r->line0   = r->line1 = z->img_comp[k].data;
-
-         if      (r->hs == 1 && r->vs == 1) r->resample = resample_row_1;
-         else if (r->hs == 1 && r->vs == 2) r->resample = stbi__resample_row_v_2;
-         else if (r->hs == 2 && r->vs == 1) r->resample = stbi__resample_row_h_2;
-         else if (r->hs == 2 && r->vs == 2) r->resample = z->resample_row_hv_2_kernel;
-         else                               r->resample = stbi__resample_row_generic;
-      }
-
-      // can't error after this so, this is safe
-      output = (stbi_uc *) stbi__malloc_mad3(n, z->s->img_x, z->s->img_y, 1);
-      if (!output) { stbi__cleanup_jpeg(z); return stbi__errpuc("outofmem", "Out of memory"); }
-
-      // now go ahead and resample
-      for (j=0; j < z->s->img_y; ++j) {
-         stbi_uc *out = output + n * z->s->img_x * j;
-         for (k=0; k < decode_n; ++k) {
-            stbi__resample *r = &res_comp[k];
-            int y_bot = r->ystep >= (r->vs >> 1);
-            coutput[k] = r->resample(z->img_comp[k].linebuf,
-                                     y_bot ? r->line1 : r->line0,
-                                     y_bot ? r->line0 : r->line1,
-                                     r->w_lores, r->hs);
-            if (++r->ystep >= r->vs) {
-               r->ystep = 0;
-               r->line0 = r->line1;
-               if (++r->ypos < z->img_comp[k].y)
-                  r->line1 += z->img_comp[k].w2;
-            }
-         }
-         if (n >= 3) {
-            stbi_uc *y = coutput[0];
-            if (z->s->img_n == 3) {
-               if (is_rgb) {
-                  for (i=0; i < z->s->img_x; ++i) {
-                     out[0] = y[i];
-                     out[1] = coutput[1][i];
-                     out[2] = coutput[2][i];
-                     out[3] = 255;
-                     out += n;
-                  }
-               } else {
-                  z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n);
-               }
-            } else if (z->s->img_n == 4) {
-               if (z->app14_color_transform == 0) { // CMYK
-                  for (i=0; i < z->s->img_x; ++i) {
-                     stbi_uc m = coutput[3][i];
-                     out[0] = stbi__blinn_8x8(coutput[0][i], m);
-                     out[1] = stbi__blinn_8x8(coutput[1][i], m);
-                     out[2] = stbi__blinn_8x8(coutput[2][i], m);
-                     out[3] = 255;
-                     out += n;
-                  }
-               } else if (z->app14_color_transform == 2) { // YCCK
-                  z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n);
-                  for (i=0; i < z->s->img_x; ++i) {
-                     stbi_uc m = coutput[3][i];
-                     out[0] = stbi__blinn_8x8(255 - out[0], m);
-                     out[1] = stbi__blinn_8x8(255 - out[1], m);
-                     out[2] = stbi__blinn_8x8(255 - out[2], m);
-                     out += n;
-                  }
-               } else { // YCbCr + alpha?  Ignore the fourth channel for now
-                  z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n);
-               }
-            } else
-               for (i=0; i < z->s->img_x; ++i) {
-                  out[0] = out[1] = out[2] = y[i];
-                  out[3] = 255; // not used if n==3
-                  out += n;
-               }
-         } else {
-            if (is_rgb) {
-               if (n == 1)
-                  for (i=0; i < z->s->img_x; ++i)
-                     *out++ = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]);
-               else {
-                  for (i=0; i < z->s->img_x; ++i, out += 2) {
-                     out[0] = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]);
-                     out[1] = 255;
-                  }
-               }
-            } else if (z->s->img_n == 4 && z->app14_color_transform == 0) {
-               for (i=0; i < z->s->img_x; ++i) {
-                  stbi_uc m = coutput[3][i];
-                  stbi_uc r = stbi__blinn_8x8(coutput[0][i], m);
-                  stbi_uc g = stbi__blinn_8x8(coutput[1][i], m);
-                  stbi_uc b = stbi__blinn_8x8(coutput[2][i], m);
-                  out[0] = stbi__compute_y(r, g, b);
-                  out[1] = 255;
-                  out += n;
-               }
-            } else if (z->s->img_n == 4 && z->app14_color_transform == 2) {
-               for (i=0; i < z->s->img_x; ++i) {
-                  out[0] = stbi__blinn_8x8(255 - coutput[0][i], coutput[3][i]);
-                  out[1] = 255;
-                  out += n;
-               }
-            } else {
-               stbi_uc *y = coutput[0];
-               if (n == 1)
-                  for (i=0; i < z->s->img_x; ++i) out[i] = y[i];
-               else
-                  for (i=0; i < z->s->img_x; ++i) { *out++ = y[i]; *out++ = 255; }
-            }
-         }
-      }
-      stbi__cleanup_jpeg(z);
-      *out_x = z->s->img_x;
-      *out_y = z->s->img_y;
-      if (comp) *comp = z->s->img_n >= 3 ? 3 : 1; // report original components, not output
-      return output;
-   }
-}
-
-static void *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
-{
-   unsigned char* result;
-   stbi__jpeg* j = (stbi__jpeg*) stbi__malloc(sizeof(stbi__jpeg));
-   if (!j) return stbi__errpuc("outofmem", "Out of memory");
-   memset(j, 0, sizeof(stbi__jpeg));
-   STBI_NOTUSED(ri);
-   j->s = s;
-   stbi__setup_jpeg(j);
-   result = load_jpeg_image(j, x,y,comp,req_comp);
-   STBI_FREE(j);
-   return result;
-}
-
-static int stbi__jpeg_test(stbi__context *s)
-{
-   int r;
-   stbi__jpeg* j = (stbi__jpeg*)stbi__malloc(sizeof(stbi__jpeg));
-   if (!j) return stbi__err("outofmem", "Out of memory");
-   memset(j, 0, sizeof(stbi__jpeg));
-   j->s = s;
-   stbi__setup_jpeg(j);
-   r = stbi__decode_jpeg_header(j, STBI__SCAN_type);
-   stbi__rewind(s);
-   STBI_FREE(j);
-   return r;
-}
-
-static int stbi__jpeg_info_raw(stbi__jpeg *j, int *x, int *y, int *comp)
-{
-   if (!stbi__decode_jpeg_header(j, STBI__SCAN_header)) {
-      stbi__rewind( j->s );
-      return 0;
-   }
-   if (x) *x = j->s->img_x;
-   if (y) *y = j->s->img_y;
-   if (comp) *comp = j->s->img_n >= 3 ? 3 : 1;
-   return 1;
-}
-
-static int stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp)
-{
-   int result;
-   stbi__jpeg* j = (stbi__jpeg*) (stbi__malloc(sizeof(stbi__jpeg)));
-   if (!j) return stbi__err("outofmem", "Out of memory");
-   memset(j, 0, sizeof(stbi__jpeg));
-   j->s = s;
-   result = stbi__jpeg_info_raw(j, x, y, comp);
-   STBI_FREE(j);
-   return result;
-}
-#endif
-
-// public domain zlib decode    v0.2  Sean Barrett 2006-11-18
-//    simple implementation
-//      - all input must be provided in an upfront buffer
-//      - all output is written to a single output buffer (can malloc/realloc)
-//    performance
-//      - fast huffman
-
-#ifndef STBI_NO_ZLIB
-
-// fast-way is faster to check than jpeg huffman, but slow way is slower
-#define STBI__ZFAST_BITS  9 // accelerate all cases in default tables
-#define STBI__ZFAST_MASK  ((1 << STBI__ZFAST_BITS) - 1)
-#define STBI__ZNSYMS 288 // number of symbols in literal/length alphabet
-
-// zlib-style huffman encoding
-// (jpegs packs from left, zlib from right, so can't share code)
-typedef struct
-{
-   stbi__uint16 fast[1 << STBI__ZFAST_BITS];
-   stbi__uint16 firstcode[16];
-   int maxcode[17];
-   stbi__uint16 firstsymbol[16];
-   stbi_uc  size[STBI__ZNSYMS];
-   stbi__uint16 value[STBI__ZNSYMS];
-} stbi__zhuffman;
-
-stbi_inline static int stbi__bitreverse16(int n)
-{
-  n = ((n & 0xAAAA) >>  1) | ((n & 0x5555) << 1);
-  n = ((n & 0xCCCC) >>  2) | ((n & 0x3333) << 2);
-  n = ((n & 0xF0F0) >>  4) | ((n & 0x0F0F) << 4);
-  n = ((n & 0xFF00) >>  8) | ((n & 0x00FF) << 8);
-  return n;
-}
-
-stbi_inline static int stbi__bit_reverse(int v, int bits)
-{
-   STBI_ASSERT(bits <= 16);
-   // to bit reverse n bits, reverse 16 and shift
-   // e.g. 11 bits, bit reverse and shift away 5
-   return stbi__bitreverse16(v) >> (16-bits);
-}
-
-static int stbi__zbuild_huffman(stbi__zhuffman *z, const stbi_uc *sizelist, int num)
-{
-   int i,k=0;
-   int code, next_code[16], sizes[17];
-
-   // DEFLATE spec for generating codes
-   memset(sizes, 0, sizeof(sizes));
-   memset(z->fast, 0, sizeof(z->fast));
-   for (i=0; i < num; ++i)
-      ++sizes[sizelist[i]];
-   sizes[0] = 0;
-   for (i=1; i < 16; ++i)
-      if (sizes[i] > (1 << i))
-         return stbi__err("bad sizes", "Corrupt PNG");
-   code = 0;
-   for (i=1; i < 16; ++i) {
-      next_code[i] = code;
-      z->firstcode[i] = (stbi__uint16) code;
-      z->firstsymbol[i] = (stbi__uint16) k;
-      code = (code + sizes[i]);
-      if (sizes[i])
-         if (code-1 >= (1 << i)) return stbi__err("bad codelengths","Corrupt PNG");
-      z->maxcode[i] = code << (16-i); // preshift for inner loop
-      code <<= 1;
-      k += sizes[i];
-   }
-   z->maxcode[16] = 0x10000; // sentinel
-   for (i=0; i < num; ++i) {
-      int s = sizelist[i];
-      if (s) {
-         int c = next_code[s] - z->firstcode[s] + z->firstsymbol[s];
-         stbi__uint16 fastv = (stbi__uint16) ((s << 9) | i);
-         z->size [c] = (stbi_uc     ) s;
-         z->value[c] = (stbi__uint16) i;
-         if (s <= STBI__ZFAST_BITS) {
-            int j = stbi__bit_reverse(next_code[s],s);
-            while (j < (1 << STBI__ZFAST_BITS)) {
-               z->fast[j] = fastv;
-               j += (1 << s);
-            }
-         }
-         ++next_code[s];
-      }
-   }
-   return 1;
-}
-
-// zlib-from-memory implementation for PNG reading
-//    because PNG allows splitting the zlib stream arbitrarily,
-//    and it's annoying structurally to have PNG call ZLIB call PNG,
-//    we require PNG read all the IDATs and combine them into a single
-//    memory buffer
-
-typedef struct
-{
-   stbi_uc *zbuffer, *zbuffer_end;
-   int num_bits;
-   stbi__uint32 code_buffer;
-
-   char *zout;
-   char *zout_start;
-   char *zout_end;
-   int   z_expandable;
-
-   stbi__zhuffman z_length, z_distance;
-} stbi__zbuf;
-
-stbi_inline static int stbi__zeof(stbi__zbuf *z)
-{
-   return (z->zbuffer >= z->zbuffer_end);
-}
-
-stbi_inline static stbi_uc stbi__zget8(stbi__zbuf *z)
-{
-   return stbi__zeof(z) ? 0 : *z->zbuffer++;
-}
-
-static void stbi__fill_bits(stbi__zbuf *z)
-{
-   do {
-      if (z->code_buffer >= (1U << z->num_bits)) {
-        z->zbuffer = z->zbuffer_end;  /* treat this as EOF so we fail. */
-        return;
-      }
-      z->code_buffer |= (unsigned int) stbi__zget8(z) << z->num_bits;
-      z->num_bits += 8;
-   } while (z->num_bits <= 24);
-}
-
-stbi_inline static unsigned int stbi__zreceive(stbi__zbuf *z, int n)
-{
-   unsigned int k;
-   if (z->num_bits < n) stbi__fill_bits(z);
-   k = z->code_buffer & ((1 << n) - 1);
-   z->code_buffer >>= n;
-   z->num_bits -= n;
-   return k;
-}
-
-static int stbi__zhuffman_decode_slowpath(stbi__zbuf *a, stbi__zhuffman *z)
-{
-   int b,s,k;
-   // not resolved by fast table, so compute it the slow way
-   // use jpeg approach, which requires MSbits at top
-   k = stbi__bit_reverse(a->code_buffer, 16);
-   for (s=STBI__ZFAST_BITS+1; ; ++s)
-      if (k < z->maxcode[s])
-         break;
-   if (s >= 16) return -1; // invalid code!
-   // code size is s, so:
-   b = (k >> (16-s)) - z->firstcode[s] + z->firstsymbol[s];
-   if (b >= STBI__ZNSYMS) return -1; // some data was corrupt somewhere!
-   if (z->size[b] != s) return -1;  // was originally an assert, but report failure instead.
-   a->code_buffer >>= s;
-   a->num_bits -= s;
-   return z->value[b];
-}
-
-stbi_inline static int stbi__zhuffman_decode(stbi__zbuf *a, stbi__zhuffman *z)
-{
-   int b,s;
-   if (a->num_bits < 16) {
-      if (stbi__zeof(a)) {
-         return -1;   /* report error for unexpected end of data. */
-      }
-      stbi__fill_bits(a);
-   }
-   b = z->fast[a->code_buffer & STBI__ZFAST_MASK];
-   if (b) {
-      s = b >> 9;
-      a->code_buffer >>= s;
-      a->num_bits -= s;
-      return b & 511;
-   }
-   return stbi__zhuffman_decode_slowpath(a, z);
-}
-
-static int stbi__zexpand(stbi__zbuf *z, char *zout, int n)  // need to make room for n bytes
-{
-   char *q;
-   unsigned int cur, limit, old_limit;
-   z->zout = zout;
-   if (!z->z_expandable) return stbi__err("output buffer limit","Corrupt PNG");
-   cur   = (unsigned int) (z->zout - z->zout_start);
-   limit = old_limit = (unsigned) (z->zout_end - z->zout_start);
-   if (UINT_MAX - cur < (unsigned) n) return stbi__err("outofmem", "Out of memory");
-   while (cur + n > limit) {
-      if(limit > UINT_MAX / 2) return stbi__err("outofmem", "Out of memory");
-      limit *= 2;
-   }
-   q = (char *) STBI_REALLOC_SIZED(z->zout_start, old_limit, limit);
-   STBI_NOTUSED(old_limit);
-   if (q == NULL) return stbi__err("outofmem", "Out of memory");
-   z->zout_start = q;
-   z->zout       = q + cur;
-   z->zout_end   = q + limit;
-   return 1;
-}
-
-static const int stbi__zlength_base[31] = {
-   3,4,5,6,7,8,9,10,11,13,
-   15,17,19,23,27,31,35,43,51,59,
-   67,83,99,115,131,163,195,227,258,0,0 };
-
-static const int stbi__zlength_extra[31]=
-{ 0,0,0,0,0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3,4,4,4,4,5,5,5,5,0,0,0 };
-
-static const int stbi__zdist_base[32] = { 1,2,3,4,5,7,9,13,17,25,33,49,65,97,129,193,
-257,385,513,769,1025,1537,2049,3073,4097,6145,8193,12289,16385,24577,0,0};
-
-static const int stbi__zdist_extra[32] =
-{ 0,0,0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,8,8,9,9,10,10,11,11,12,12,13,13};
-
-static int stbi__parse_huffman_block(stbi__zbuf *a)
-{
-   char *zout = a->zout;
-   for(;;) {
-      int z = stbi__zhuffman_decode(a, &a->z_length);
-      if (z < 256) {
-         if (z < 0) return stbi__err("bad huffman code","Corrupt PNG"); // error in huffman codes
-         if (zout >= a->zout_end) {
-            if (!stbi__zexpand(a, zout, 1)) return 0;
-            zout = a->zout;
-         }
-         *zout++ = (char) z;
-      } else {
-         stbi_uc *p;
-         int len,dist;
-         if (z == 256) {
-            a->zout = zout;
-            return 1;
-         }
-         if (z >= 286) return stbi__err("bad huffman code","Corrupt PNG"); // per DEFLATE, length codes 286 and 287 must not appear in compressed data
-         z -= 257;
-         len = stbi__zlength_base[z];
-         if (stbi__zlength_extra[z]) len += stbi__zreceive(a, stbi__zlength_extra[z]);
-         z = stbi__zhuffman_decode(a, &a->z_distance);
-         if (z < 0 || z >= 30) return stbi__err("bad huffman code","Corrupt PNG"); // per DEFLATE, distance codes 30 and 31 must not appear in compressed data
-         dist = stbi__zdist_base[z];
-         if (stbi__zdist_extra[z]) dist += stbi__zreceive(a, stbi__zdist_extra[z]);
-         if (zout - a->zout_start < dist) return stbi__err("bad dist","Corrupt PNG");
-         if (zout + len > a->zout_end) {
-            if (!stbi__zexpand(a, zout, len)) return 0;
-            zout = a->zout;
-         }
-         p = (stbi_uc *) (zout - dist);
-         if (dist == 1) { // run of one byte; common in images.
-            stbi_uc v = *p;
-            if (len) { do *zout++ = v; while (--len); }
-         } else {
-            if (len) { do *zout++ = *p++; while (--len); }
-         }
-      }
-   }
-}
-
-static int stbi__compute_huffman_codes(stbi__zbuf *a)
-{
-   static const stbi_uc length_dezigzag[19] = { 16,17,18,0,8,7,9,6,10,5,11,4,12,3,13,2,14,1,15 };
-   stbi__zhuffman z_codelength;
-   stbi_uc lencodes[286+32+137];//padding for maximum single op
-   stbi_uc codelength_sizes[19];
-   int i,n;
-
-   int hlit  = stbi__zreceive(a,5) + 257;
-   int hdist = stbi__zreceive(a,5) + 1;
-   int hclen = stbi__zreceive(a,4) + 4;
-   int ntot  = hlit + hdist;
-
-   memset(codelength_sizes, 0, sizeof(codelength_sizes));
-   for (i=0; i < hclen; ++i) {
-      int s = stbi__zreceive(a,3);
-      codelength_sizes[length_dezigzag[i]] = (stbi_uc) s;
-   }
-   if (!stbi__zbuild_huffman(&z_codelength, codelength_sizes, 19)) return 0;
-
-   n = 0;
-   while (n < ntot) {
-      int c = stbi__zhuffman_decode(a, &z_codelength);
-      if (c < 0 || c >= 19) return stbi__err("bad codelengths", "Corrupt PNG");
-      if (c < 16)
-         lencodes[n++] = (stbi_uc) c;
-      else {
-         stbi_uc fill = 0;
-         if (c == 16) {
-            c = stbi__zreceive(a,2)+3;
-            if (n == 0) return stbi__err("bad codelengths", "Corrupt PNG");
-            fill = lencodes[n-1];
-         } else if (c == 17) {
-            c = stbi__zreceive(a,3)+3;
-         } else if (c == 18) {
-            c = stbi__zreceive(a,7)+11;
-         } else {
-            return stbi__err("bad codelengths", "Corrupt PNG");
-         }
-         if (ntot - n < c) return stbi__err("bad codelengths", "Corrupt PNG");
-         memset(lencodes+n, fill, c);
-         n += c;
-      }
-   }
-   if (n != ntot) return stbi__err("bad codelengths","Corrupt PNG");
-   if (!stbi__zbuild_huffman(&a->z_length, lencodes, hlit)) return 0;
-   if (!stbi__zbuild_huffman(&a->z_distance, lencodes+hlit, hdist)) return 0;
-   return 1;
-}
-
-static int stbi__parse_uncompressed_block(stbi__zbuf *a)
-{
-   stbi_uc header[4];
-   int len,nlen,k;
-   if (a->num_bits & 7)
-      stbi__zreceive(a, a->num_bits & 7); // discard
-   // drain the bit-packed data into header
-   k = 0;
-   while (a->num_bits > 0) {
-      header[k++] = (stbi_uc) (a->code_buffer & 255); // suppress MSVC run-time check
-      a->code_buffer >>= 8;
-      a->num_bits -= 8;
-   }
-   if (a->num_bits < 0) return stbi__err("zlib corrupt","Corrupt PNG");
-   // now fill header the normal way
-   while (k < 4)
-      header[k++] = stbi__zget8(a);
-   len  = header[1] * 256 + header[0];
-   nlen = header[3] * 256 + header[2];
-   if (nlen != (len ^ 0xffff)) return stbi__err("zlib corrupt","Corrupt PNG");
-   if (a->zbuffer + len > a->zbuffer_end) return stbi__err("read past buffer","Corrupt PNG");
-   if (a->zout + len > a->zout_end)
-      if (!stbi__zexpand(a, a->zout, len)) return 0;
-   memcpy(a->zout, a->zbuffer, len);
-   a->zbuffer += len;
-   a->zout += len;
-   return 1;
-}
-
-static int stbi__parse_zlib_header(stbi__zbuf *a)
-{
-   int cmf   = stbi__zget8(a);
-   int cm    = cmf & 15;
-   /* int cinfo = cmf >> 4; */
-   int flg   = stbi__zget8(a);
-   if (stbi__zeof(a)) return stbi__err("bad zlib header","Corrupt PNG"); // zlib spec
-   if ((cmf*256+flg) % 31 != 0) return stbi__err("bad zlib header","Corrupt PNG"); // zlib spec
-   if (flg & 32) return stbi__err("no preset dict","Corrupt PNG"); // preset dictionary not allowed in png
-   if (cm != 8) return stbi__err("bad compression","Corrupt PNG"); // DEFLATE required for png
-   // window = 1 << (8 + cinfo)... but who cares, we fully buffer output
-   return 1;
-}
-
-static const stbi_uc stbi__zdefault_length[STBI__ZNSYMS] =
-{
-   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
-   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
-   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
-   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
-   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
-   9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
-   9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
-   9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
-   7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7, 7,7,7,7,7,7,7,7,8,8,8,8,8,8,8,8
-};
-static const stbi_uc stbi__zdefault_distance[32] =
-{
-   5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5
-};
-/*
-Init algorithm:
-{
-   int i;   // use <= to match clearly with spec
-   for (i=0; i <= 143; ++i)     stbi__zdefault_length[i]   = 8;
-   for (   ; i <= 255; ++i)     stbi__zdefault_length[i]   = 9;
-   for (   ; i <= 279; ++i)     stbi__zdefault_length[i]   = 7;
-   for (   ; i <= 287; ++i)     stbi__zdefault_length[i]   = 8;
-
-   for (i=0; i <=  31; ++i)     stbi__zdefault_distance[i] = 5;
-}
-*/
-
-static int stbi__parse_zlib(stbi__zbuf *a, int parse_header)
-{
-   int final, type;
-   if (parse_header)
-      if (!stbi__parse_zlib_header(a)) return 0;
-   a->num_bits = 0;
-   a->code_buffer = 0;
-   do {
-      final = stbi__zreceive(a,1);
-      type = stbi__zreceive(a,2);
-      if (type == 0) {
-         if (!stbi__parse_uncompressed_block(a)) return 0;
-      } else if (type == 3) {
-         return 0;
-      } else {
-         if (type == 1) {
-            // use fixed code lengths
-            if (!stbi__zbuild_huffman(&a->z_length  , stbi__zdefault_length  , STBI__ZNSYMS)) return 0;
-            if (!stbi__zbuild_huffman(&a->z_distance, stbi__zdefault_distance,  32)) return 0;
-         } else {
-            if (!stbi__compute_huffman_codes(a)) return 0;
-         }
-         if (!stbi__parse_huffman_block(a)) return 0;
-      }
-   } while (!final);
-   return 1;
-}
-
-static int stbi__do_zlib(stbi__zbuf *a, char *obuf, int olen, int exp, int parse_header)
-{
-   a->zout_start = obuf;
-   a->zout       = obuf;
-   a->zout_end   = obuf + olen;
-   a->z_expandable = exp;
-
-   return stbi__parse_zlib(a, parse_header);
-}
-
-STBIDEF char *stbi_zlib_decode_malloc_guesssize(const char *buffer, int len, int initial_size, int *outlen)
-{
-   stbi__zbuf a;
-   char *p = (char *) stbi__malloc(initial_size);
-   if (p == NULL) return NULL;
-   a.zbuffer = (stbi_uc *) buffer;
-   a.zbuffer_end = (stbi_uc *) buffer + len;
-   if (stbi__do_zlib(&a, p, initial_size, 1, 1)) {
-      if (outlen) *outlen = (int) (a.zout - a.zout_start);
-      return a.zout_start;
-   } else {
-      STBI_FREE(a.zout_start);
-      return NULL;
-   }
-}
-
-STBIDEF char *stbi_zlib_decode_malloc(char const *buffer, int len, int *outlen)
-{
-   return stbi_zlib_decode_malloc_guesssize(buffer, len, 16384, outlen);
-}
-
-STBIDEF char *stbi_zlib_decode_malloc_guesssize_headerflag(const char *buffer, int len, int initial_size, int *outlen, int parse_header)
-{
-   stbi__zbuf a;
-   char *p = (char *) stbi__malloc(initial_size);
-   if (p == NULL) return NULL;
-   a.zbuffer = (stbi_uc *) buffer;
-   a.zbuffer_end = (stbi_uc *) buffer + len;
-   if (stbi__do_zlib(&a, p, initial_size, 1, parse_header)) {
-      if (outlen) *outlen = (int) (a.zout - a.zout_start);
-      return a.zout_start;
-   } else {
-      STBI_FREE(a.zout_start);
-      return NULL;
-   }
-}
-
-STBIDEF int stbi_zlib_decode_buffer(char *obuffer, int olen, char const *ibuffer, int ilen)
-{
-   stbi__zbuf a;
-   a.zbuffer = (stbi_uc *) ibuffer;
-   a.zbuffer_end = (stbi_uc *) ibuffer + ilen;
-   if (stbi__do_zlib(&a, obuffer, olen, 0, 1))
-      return (int) (a.zout - a.zout_start);
-   else
-      return -1;
-}
-
-STBIDEF char *stbi_zlib_decode_noheader_malloc(char const *buffer, int len, int *outlen)
-{
-   stbi__zbuf a;
-   char *p = (char *) stbi__malloc(16384);
-   if (p == NULL) return NULL;
-   a.zbuffer = (stbi_uc *) buffer;
-   a.zbuffer_end = (stbi_uc *) buffer+len;
-   if (stbi__do_zlib(&a, p, 16384, 1, 0)) {
-      if (outlen) *outlen = (int) (a.zout - a.zout_start);
-      return a.zout_start;
-   } else {
-      STBI_FREE(a.zout_start);
-      return NULL;
-   }
-}
-
-STBIDEF int stbi_zlib_decode_noheader_buffer(char *obuffer, int olen, const char *ibuffer, int ilen)
-{
-   stbi__zbuf a;
-   a.zbuffer = (stbi_uc *) ibuffer;
-   a.zbuffer_end = (stbi_uc *) ibuffer + ilen;
-   if (stbi__do_zlib(&a, obuffer, olen, 0, 0))
-      return (int) (a.zout - a.zout_start);
-   else
-      return -1;
-}
-#endif
-
-// public domain "baseline" PNG decoder   v0.10  Sean Barrett 2006-11-18
-//    simple implementation
-//      - only 8-bit samples
-//      - no CRC checking
-//      - allocates lots of intermediate memory
-//        - avoids problem of streaming data between subsystems
-//        - avoids explicit window management
-//    performance
-//      - uses stb_zlib, a PD zlib implementation with fast huffman decoding
-
-#ifndef STBI_NO_PNG
-typedef struct
-{
-   stbi__uint32 length;
-   stbi__uint32 type;
-} stbi__pngchunk;
-
-static stbi__pngchunk stbi__get_chunk_header(stbi__context *s)
-{
-   stbi__pngchunk c;
-   c.length = stbi__get32be(s);
-   c.type   = stbi__get32be(s);
-   return c;
-}
-
-static int stbi__check_png_header(stbi__context *s)
-{
-   static const stbi_uc png_sig[8] = { 137,80,78,71,13,10,26,10 };
-   int i;
-   for (i=0; i < 8; ++i)
-      if (stbi__get8(s) != png_sig[i]) return stbi__err("bad png sig","Not a PNG");
-   return 1;
-}
-
-typedef struct
-{
-   stbi__context *s;
-   stbi_uc *idata, *expanded, *out;
-   int depth;
-} stbi__png;
-
-
-enum {
-   STBI__F_none=0,
-   STBI__F_sub=1,
-   STBI__F_up=2,
-   STBI__F_avg=3,
-   STBI__F_paeth=4,
-   // synthetic filters used for first scanline to avoid needing a dummy row of 0s
-   STBI__F_avg_first,
-   STBI__F_paeth_first
-};
-
-static stbi_uc first_row_filter[5] =
-{
-   STBI__F_none,
-   STBI__F_sub,
-   STBI__F_none,
-   STBI__F_avg_first,
-   STBI__F_paeth_first
-};
-
-static int stbi__paeth(int a, int b, int c)
-{
-   int p = a + b - c;
-   int pa = abs(p-a);
-   int pb = abs(p-b);
-   int pc = abs(p-c);
-   if (pa <= pb && pa <= pc) return a;
-   if (pb <= pc) return b;
-   return c;
-}
-
-static const stbi_uc stbi__depth_scale_table[9] = { 0, 0xff, 0x55, 0, 0x11, 0,0,0, 0x01 };
-
-// create the png data from post-deflated data
-static int stbi__create_png_image_raw(stbi__png *a, stbi_uc *raw, stbi__uint32 raw_len, int out_n, stbi__uint32 x, stbi__uint32 y, int depth, int color)
-{
-   int bytes = (depth == 16? 2 : 1);
-   stbi__context *s = a->s;
-   stbi__uint32 i,j,stride = x*out_n*bytes;
-   stbi__uint32 img_len, img_width_bytes;
-   int k;
-   int img_n = s->img_n; // copy it into a local for later
-
-   int output_bytes = out_n*bytes;
-   int filter_bytes = img_n*bytes;
-   int width = x;
-
-   STBI_ASSERT(out_n == s->img_n || out_n == s->img_n+1);
-   a->out = (stbi_uc *) stbi__malloc_mad3(x, y, output_bytes, 0); // extra bytes to write off the end into
-   if (!a->out) return stbi__err("outofmem", "Out of memory");
-
-   if (!stbi__mad3sizes_valid(img_n, x, depth, 7)) return stbi__err("too large", "Corrupt PNG");
-   img_width_bytes = (((img_n * x * depth) + 7) >> 3);
-   img_len = (img_width_bytes + 1) * y;
-
-   // we used to check for exact match between raw_len and img_len on non-interlaced PNGs,
-   // but issue #276 reported a PNG in the wild that had extra data at the end (all zeros),
-   // so just check for raw_len < img_len always.
-   if (raw_len < img_len) return stbi__err("not enough pixels","Corrupt PNG");
-
-   for (j=0; j < y; ++j) {
-      stbi_uc *cur = a->out + stride*j;
-      stbi_uc *prior;
-      int filter = *raw++;
-
-      if (filter > 4)
-         return stbi__err("invalid filter","Corrupt PNG");
-
-      if (depth < 8) {
-         if (img_width_bytes > x) return stbi__err("invalid width","Corrupt PNG");
-         cur += x*out_n - img_width_bytes; // store output to the rightmost img_len bytes, so we can decode in place
-         filter_bytes = 1;
-         width = img_width_bytes;
-      }
-      prior = cur - stride; // bugfix: need to compute this after 'cur +=' computation above
-
-      // if first row, use special filter that doesn't sample previous row
-      if (j == 0) filter = first_row_filter[filter];
-
-      // handle first byte explicitly
-      for (k=0; k < filter_bytes; ++k) {
-         switch (filter) {
-            case STBI__F_none       : cur[k] = raw[k]; break;
-            case STBI__F_sub        : cur[k] = raw[k]; break;
-            case STBI__F_up         : cur[k] = STBI__BYTECAST(raw[k] + prior[k]); break;
-            case STBI__F_avg        : cur[k] = STBI__BYTECAST(raw[k] + (prior[k]>>1)); break;
-            case STBI__F_paeth      : cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(0,prior[k],0)); break;
-            case STBI__F_avg_first  : cur[k] = raw[k]; break;
-            case STBI__F_paeth_first: cur[k] = raw[k]; break;
-         }
-      }
-
-      if (depth == 8) {
-         if (img_n != out_n)
-            cur[img_n] = 255; // first pixel
-         raw += img_n;
-         cur += out_n;
-         prior += out_n;
-      } else if (depth == 16) {
-         if (img_n != out_n) {
-            cur[filter_bytes]   = 255; // first pixel top byte
-            cur[filter_bytes+1] = 255; // first pixel bottom byte
-         }
-         raw += filter_bytes;
-         cur += output_bytes;
-         prior += output_bytes;
-      } else {
-         raw += 1;
-         cur += 1;
-         prior += 1;
-      }
-
-      // this is a little gross, so that we don't switch per-pixel or per-component
-      if (depth < 8 || img_n == out_n) {
-         int nk = (width - 1)*filter_bytes;
-         #define STBI__CASE(f) \
-             case f:     \
-                for (k=0; k < nk; ++k)
-         switch (filter) {
-            // "none" filter turns into a memcpy here; make that explicit.
-            case STBI__F_none:         memcpy(cur, raw, nk); break;
-            STBI__CASE(STBI__F_sub)          { cur[k] = STBI__BYTECAST(raw[k] + cur[k-filter_bytes]); } break;
-            STBI__CASE(STBI__F_up)           { cur[k] = STBI__BYTECAST(raw[k] + prior[k]); } break;
-            STBI__CASE(STBI__F_avg)          { cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k-filter_bytes])>>1)); } break;
-            STBI__CASE(STBI__F_paeth)        { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k-filter_bytes],prior[k],prior[k-filter_bytes])); } break;
-            STBI__CASE(STBI__F_avg_first)    { cur[k] = STBI__BYTECAST(raw[k] + (cur[k-filter_bytes] >> 1)); } break;
-            STBI__CASE(STBI__F_paeth_first)  { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k-filter_bytes],0,0)); } break;
-         }
-         #undef STBI__CASE
-         raw += nk;
-      } else {
-         STBI_ASSERT(img_n+1 == out_n);
-         #define STBI__CASE(f) \
-             case f:     \
-                for (i=x-1; i >= 1; --i, cur[filter_bytes]=255,raw+=filter_bytes,cur+=output_bytes,prior+=output_bytes) \
-                   for (k=0; k < filter_bytes; ++k)
-         switch (filter) {
-            STBI__CASE(STBI__F_none)         { cur[k] = raw[k]; } break;
-            STBI__CASE(STBI__F_sub)          { cur[k] = STBI__BYTECAST(raw[k] + cur[k- output_bytes]); } break;
-            STBI__CASE(STBI__F_up)           { cur[k] = STBI__BYTECAST(raw[k] + prior[k]); } break;
-            STBI__CASE(STBI__F_avg)          { cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k- output_bytes])>>1)); } break;
-            STBI__CASE(STBI__F_paeth)        { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k- output_bytes],prior[k],prior[k- output_bytes])); } break;
-            STBI__CASE(STBI__F_avg_first)    { cur[k] = STBI__BYTECAST(raw[k] + (cur[k- output_bytes] >> 1)); } break;
-            STBI__CASE(STBI__F_paeth_first)  { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k- output_bytes],0,0)); } break;
-         }
-         #undef STBI__CASE
-
-         // the loop above sets the high byte of the pixels' alpha, but for
-         // 16 bit png files we also need the low byte set. we'll do that here.
-         if (depth == 16) {
-            cur = a->out + stride*j; // start at the beginning of the row again
-            for (i=0; i < x; ++i,cur+=output_bytes) {
-               cur[filter_bytes+1] = 255;
-            }
-         }
-      }
-   }
-
-   // we make a separate pass to expand bits to pixels; for performance,
-   // this could run two scanlines behind the above code, so it won't
-   // intefere with filtering but will still be in the cache.
-   if (depth < 8) {
-      for (j=0; j < y; ++j) {
-         stbi_uc *cur = a->out + stride*j;
-         stbi_uc *in  = a->out + stride*j + x*out_n - img_width_bytes;
-         // unpack 1/2/4-bit into a 8-bit buffer. allows us to keep the common 8-bit path optimal at minimal cost for 1/2/4-bit
-         // png guarante byte alignment, if width is not multiple of 8/4/2 we'll decode dummy trailing data that will be skipped in the later loop
-         stbi_uc scale = (color == 0) ? stbi__depth_scale_table[depth] : 1; // scale grayscale values to 0..255 range
-
-         // note that the final byte might overshoot and write more data than desired.
-         // we can allocate enough data that this never writes out of memory, but it
-         // could also overwrite the next scanline. can it overwrite non-empty data
-         // on the next scanline? yes, consider 1-pixel-wide scanlines with 1-bit-per-pixel.
-         // so we need to explicitly clamp the final ones
-
-         if (depth == 4) {
-            for (k=x*img_n; k >= 2; k-=2, ++in) {
-               *cur++ = scale * ((*in >> 4)       );
-               *cur++ = scale * ((*in     ) & 0x0f);
-            }
-            if (k > 0) *cur++ = scale * ((*in >> 4)       );
-         } else if (depth == 2) {
-            for (k=x*img_n; k >= 4; k-=4, ++in) {
-               *cur++ = scale * ((*in >> 6)       );
-               *cur++ = scale * ((*in >> 4) & 0x03);
-               *cur++ = scale * ((*in >> 2) & 0x03);
-               *cur++ = scale * ((*in     ) & 0x03);
-            }
-            if (k > 0) *cur++ = scale * ((*in >> 6)       );
-            if (k > 1) *cur++ = scale * ((*in >> 4) & 0x03);
-            if (k > 2) *cur++ = scale * ((*in >> 2) & 0x03);
-         } else if (depth == 1) {
-            for (k=x*img_n; k >= 8; k-=8, ++in) {
-               *cur++ = scale * ((*in >> 7)       );
-               *cur++ = scale * ((*in >> 6) & 0x01);
-               *cur++ = scale * ((*in >> 5) & 0x01);
-               *cur++ = scale * ((*in >> 4) & 0x01);
-               *cur++ = scale * ((*in >> 3) & 0x01);
-               *cur++ = scale * ((*in >> 2) & 0x01);
-               *cur++ = scale * ((*in >> 1) & 0x01);
-               *cur++ = scale * ((*in     ) & 0x01);
-            }
-            if (k > 0) *cur++ = scale * ((*in >> 7)       );
-            if (k > 1) *cur++ = scale * ((*in >> 6) & 0x01);
-            if (k > 2) *cur++ = scale * ((*in >> 5) & 0x01);
-            if (k > 3) *cur++ = scale * ((*in >> 4) & 0x01);
-            if (k > 4) *cur++ = scale * ((*in >> 3) & 0x01);
-            if (k > 5) *cur++ = scale * ((*in >> 2) & 0x01);
-            if (k > 6) *cur++ = scale * ((*in >> 1) & 0x01);
-         }
-         if (img_n != out_n) {
-            int q;
-            // insert alpha = 255
-            cur = a->out + stride*j;
-            if (img_n == 1) {
-               for (q=x-1; q >= 0; --q) {
-                  cur[q*2+1] = 255;
-                  cur[q*2+0] = cur[q];
-               }
-            } else {
-               STBI_ASSERT(img_n == 3);
-               for (q=x-1; q >= 0; --q) {
-                  cur[q*4+3] = 255;
-                  cur[q*4+2] = cur[q*3+2];
-                  cur[q*4+1] = cur[q*3+1];
-                  cur[q*4+0] = cur[q*3+0];
-               }
-            }
-         }
-      }
-   } else if (depth == 16) {
-      // force the image data from big-endian to platform-native.
-      // this is done in a separate pass due to the decoding relying
-      // on the data being untouched, but could probably be done
-      // per-line during decode if care is taken.
-      stbi_uc *cur = a->out;
-      stbi__uint16 *cur16 = (stbi__uint16*)cur;
-
-      for(i=0; i < x*y*out_n; ++i,cur16++,cur+=2) {
-         *cur16 = (cur[0] << 8) | cur[1];
-      }
-   }
-
-   return 1;
-}
-
-static int stbi__create_png_image(stbi__png *a, stbi_uc *image_data, stbi__uint32 image_data_len, int out_n, int depth, int color, int interlaced)
-{
-   int bytes = (depth == 16 ? 2 : 1);
-   int out_bytes = out_n * bytes;
-   stbi_uc *final;
-   int p;
-   if (!interlaced)
-      return stbi__create_png_image_raw(a, image_data, image_data_len, out_n, a->s->img_x, a->s->img_y, depth, color);
-
-   // de-interlacing
-   final = (stbi_uc *) stbi__malloc_mad3(a->s->img_x, a->s->img_y, out_bytes, 0);
-   if (!final) return stbi__err("outofmem", "Out of memory");
-   for (p=0; p < 7; ++p) {
-      int xorig[] = { 0,4,0,2,0,1,0 };
-      int yorig[] = { 0,0,4,0,2,0,1 };
-      int xspc[]  = { 8,8,4,4,2,2,1 };
-      int yspc[]  = { 8,8,8,4,4,2,2 };
-      int i,j,x,y;
-      // pass1_x[4] = 0, pass1_x[5] = 1, pass1_x[12] = 1
-      x = (a->s->img_x - xorig[p] + xspc[p]-1) / xspc[p];
-      y = (a->s->img_y - yorig[p] + yspc[p]-1) / yspc[p];
-      if (x && y) {
-         stbi__uint32 img_len = ((((a->s->img_n * x * depth) + 7) >> 3) + 1) * y;
-         if (!stbi__create_png_image_raw(a, image_data, image_data_len, out_n, x, y, depth, color)) {
-            STBI_FREE(final);
-            return 0;
-         }
-         for (j=0; j < y; ++j) {
-            for (i=0; i < x; ++i) {
-               int out_y = j*yspc[p]+yorig[p];
-               int out_x = i*xspc[p]+xorig[p];
-               memcpy(final + out_y*a->s->img_x*out_bytes + out_x*out_bytes,
-                      a->out + (j*x+i)*out_bytes, out_bytes);
-            }
-         }
-         STBI_FREE(a->out);
-         image_data += img_len;
-         image_data_len -= img_len;
-      }
-   }
-   a->out = final;
-
-   return 1;
-}
-
-static int stbi__compute_transparency(stbi__png *z, stbi_uc tc[3], int out_n)
-{
-   stbi__context *s = z->s;
-   stbi__uint32 i, pixel_count = s->img_x * s->img_y;
-   stbi_uc *p = z->out;
-
-   // compute color-based transparency, assuming we've
-   // already got 255 as the alpha value in the output
-   STBI_ASSERT(out_n == 2 || out_n == 4);
-
-   if (out_n == 2) {
-      for (i=0; i < pixel_count; ++i) {
-         p[1] = (p[0] == tc[0] ? 0 : 255);
-         p += 2;
-      }
-   } else {
-      for (i=0; i < pixel_count; ++i) {
-         if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2])
-            p[3] = 0;
-         p += 4;
-      }
-   }
-   return 1;
-}
-
-static int stbi__compute_transparency16(stbi__png *z, stbi__uint16 tc[3], int out_n)
-{
-   stbi__context *s = z->s;
-   stbi__uint32 i, pixel_count = s->img_x * s->img_y;
-   stbi__uint16 *p = (stbi__uint16*) z->out;
-
-   // compute color-based transparency, assuming we've
-   // already got 65535 as the alpha value in the output
-   STBI_ASSERT(out_n == 2 || out_n == 4);
-
-   if (out_n == 2) {
-      for (i = 0; i < pixel_count; ++i) {
-         p[1] = (p[0] == tc[0] ? 0 : 65535);
-         p += 2;
-      }
-   } else {
-      for (i = 0; i < pixel_count; ++i) {
-         if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2])
-            p[3] = 0;
-         p += 4;
-      }
-   }
-   return 1;
-}
-
-static int stbi__expand_png_palette(stbi__png *a, stbi_uc *palette, int len, int pal_img_n)
-{
-   stbi__uint32 i, pixel_count = a->s->img_x * a->s->img_y;
-   stbi_uc *p, *temp_out, *orig = a->out;
-
-   p = (stbi_uc *) stbi__malloc_mad2(pixel_count, pal_img_n, 0);
-   if (p == NULL) return stbi__err("outofmem", "Out of memory");
-
-   // between here and free(out) below, exitting would leak
-   temp_out = p;
-
-   if (pal_img_n == 3) {
-      for (i=0; i < pixel_count; ++i) {
-         int n = orig[i]*4;
-         p[0] = palette[n  ];
-         p[1] = palette[n+1];
-         p[2] = palette[n+2];
-         p += 3;
-      }
-   } else {
-      for (i=0; i < pixel_count; ++i) {
-         int n = orig[i]*4;
-         p[0] = palette[n  ];
-         p[1] = palette[n+1];
-         p[2] = palette[n+2];
-         p[3] = palette[n+3];
-         p += 4;
-      }
-   }
-   STBI_FREE(a->out);
-   a->out = temp_out;
-
-   STBI_NOTUSED(len);
-
-   return 1;
-}
-
-static int stbi__unpremultiply_on_load_global = 0;
-static int stbi__de_iphone_flag_global = 0;
-
-STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply)
-{
-   stbi__unpremultiply_on_load_global = flag_true_if_should_unpremultiply;
-}
-
-STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert)
-{
-   stbi__de_iphone_flag_global = flag_true_if_should_convert;
-}
-
-#ifndef STBI_THREAD_LOCAL
-#define stbi__unpremultiply_on_load  stbi__unpremultiply_on_load_global
-#define stbi__de_iphone_flag  stbi__de_iphone_flag_global
-#else
-static STBI_THREAD_LOCAL int stbi__unpremultiply_on_load_local, stbi__unpremultiply_on_load_set;
-static STBI_THREAD_LOCAL int stbi__de_iphone_flag_local, stbi__de_iphone_flag_set;
-
-STBIDEF void stbi_set_unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply)
-{
-   stbi__unpremultiply_on_load_local = flag_true_if_should_unpremultiply;
-   stbi__unpremultiply_on_load_set = 1;
-}
-
-STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert)
-{
-   stbi__de_iphone_flag_local = flag_true_if_should_convert;
-   stbi__de_iphone_flag_set = 1;
-}
-
-#define stbi__unpremultiply_on_load  (stbi__unpremultiply_on_load_set           \
-                                       ? stbi__unpremultiply_on_load_local      \
-                                       : stbi__unpremultiply_on_load_global)
-#define stbi__de_iphone_flag  (stbi__de_iphone_flag_set                         \
-                                ? stbi__de_iphone_flag_local                    \
-                                : stbi__de_iphone_flag_global)
-#endif // STBI_THREAD_LOCAL
-
-static void stbi__de_iphone(stbi__png *z)
-{
-   stbi__context *s = z->s;
-   stbi__uint32 i, pixel_count = s->img_x * s->img_y;
-   stbi_uc *p = z->out;
-
-   if (s->img_out_n == 3) {  // convert bgr to rgb
-      for (i=0; i < pixel_count; ++i) {
-         stbi_uc t = p[0];
-         p[0] = p[2];
-         p[2] = t;
-         p += 3;
-      }
-   } else {
-      STBI_ASSERT(s->img_out_n == 4);
-      if (stbi__unpremultiply_on_load) {
-         // convert bgr to rgb and unpremultiply
-         for (i=0; i < pixel_count; ++i) {
-            stbi_uc a = p[3];
-            stbi_uc t = p[0];
-            if (a) {
-               stbi_uc half = a / 2;
-               p[0] = (p[2] * 255 + half) / a;
-               p[1] = (p[1] * 255 + half) / a;
-               p[2] = ( t   * 255 + half) / a;
-            } else {
-               p[0] = p[2];
-               p[2] = t;
-            }
-            p += 4;
-         }
-      } else {
-         // convert bgr to rgb
-         for (i=0; i < pixel_count; ++i) {
-            stbi_uc t = p[0];
-            p[0] = p[2];
-            p[2] = t;
-            p += 4;
-         }
-      }
-   }
-}
-
-#define STBI__PNG_TYPE(a,b,c,d)  (((unsigned) (a) << 24) + ((unsigned) (b) << 16) + ((unsigned) (c) << 8) + (unsigned) (d))
-
-static int stbi__parse_png_file(stbi__png *z, int scan, int req_comp)
-{
-   stbi_uc palette[1024], pal_img_n=0;
-   stbi_uc has_trans=0, tc[3]={0};
-   stbi__uint16 tc16[3];
-   stbi__uint32 ioff=0, idata_limit=0, i, pal_len=0;
-   int first=1,k,interlace=0, color=0, is_iphone=0;
-   stbi__context *s = z->s;
-
-   z->expanded = NULL;
-   z->idata = NULL;
-   z->out = NULL;
-
-   if (!stbi__check_png_header(s)) return 0;
-
-   if (scan == STBI__SCAN_type) return 1;
-
-   for (;;) {
-      stbi__pngchunk c = stbi__get_chunk_header(s);
-      switch (c.type) {
-         case STBI__PNG_TYPE('C','g','B','I'):
-            is_iphone = 1;
-            stbi__skip(s, c.length);
-            break;
-         case STBI__PNG_TYPE('I','H','D','R'): {
-            int comp,filter;
-            if (!first) return stbi__err("multiple IHDR","Corrupt PNG");
-            first = 0;
-            if (c.length != 13) return stbi__err("bad IHDR len","Corrupt PNG");
-            s->img_x = stbi__get32be(s);
-            s->img_y = stbi__get32be(s);
-            if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
-            if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
-            z->depth = stbi__get8(s);  if (z->depth != 1 && z->depth != 2 && z->depth != 4 && z->depth != 8 && z->depth != 16)  return stbi__err("1/2/4/8/16-bit only","PNG not supported: 1/2/4/8/16-bit only");
-            color = stbi__get8(s);  if (color > 6)         return stbi__err("bad ctype","Corrupt PNG");
-            if (color == 3 && z->depth == 16)                  return stbi__err("bad ctype","Corrupt PNG");
-            if (color == 3) pal_img_n = 3; else if (color & 1) return stbi__err("bad ctype","Corrupt PNG");
-            comp  = stbi__get8(s);  if (comp) return stbi__err("bad comp method","Corrupt PNG");
-            filter= stbi__get8(s);  if (filter) return stbi__err("bad filter method","Corrupt PNG");
-            interlace = stbi__get8(s); if (interlace>1) return stbi__err("bad interlace method","Corrupt PNG");
-            if (!s->img_x || !s->img_y) return stbi__err("0-pixel image","Corrupt PNG");
-            if (!pal_img_n) {
-               s->img_n = (color & 2 ? 3 : 1) + (color & 4 ? 1 : 0);
-               if ((1 << 30) / s->img_x / s->img_n < s->img_y) return stbi__err("too large", "Image too large to decode");
-            } else {
-               // if paletted, then pal_n is our final components, and
-               // img_n is # components to decompress/filter.
-               s->img_n = 1;
-               if ((1 << 30) / s->img_x / 4 < s->img_y) return stbi__err("too large","Corrupt PNG");
-            }
-            // even with SCAN_header, have to scan to see if we have a tRNS
-            break;
-         }
-
-         case STBI__PNG_TYPE('P','L','T','E'):  {
-            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
-            if (c.length > 256*3) return stbi__err("invalid PLTE","Corrupt PNG");
-            pal_len = c.length / 3;
-            if (pal_len * 3 != c.length) return stbi__err("invalid PLTE","Corrupt PNG");
-            for (i=0; i < pal_len; ++i) {
-               palette[i*4+0] = stbi__get8(s);
-               palette[i*4+1] = stbi__get8(s);
-               palette[i*4+2] = stbi__get8(s);
-               palette[i*4+3] = 255;
-            }
-            break;
-         }
-
-         case STBI__PNG_TYPE('t','R','N','S'): {
-            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
-            if (z->idata) return stbi__err("tRNS after IDAT","Corrupt PNG");
-            if (pal_img_n) {
-               if (scan == STBI__SCAN_header) { s->img_n = 4; return 1; }
-               if (pal_len == 0) return stbi__err("tRNS before PLTE","Corrupt PNG");
-               if (c.length > pal_len) return stbi__err("bad tRNS len","Corrupt PNG");
-               pal_img_n = 4;
-               for (i=0; i < c.length; ++i)
-                  palette[i*4+3] = stbi__get8(s);
-            } else {
-               if (!(s->img_n & 1)) return stbi__err("tRNS with alpha","Corrupt PNG");
-               if (c.length != (stbi__uint32) s->img_n*2) return stbi__err("bad tRNS len","Corrupt PNG");
-               has_trans = 1;
-               // non-paletted with tRNS = constant alpha. if header-scanning, we can stop now.
-               if (scan == STBI__SCAN_header) { ++s->img_n; return 1; }
-               if (z->depth == 16) {
-                  for (k = 0; k < s->img_n; ++k) tc16[k] = (stbi__uint16)stbi__get16be(s); // copy the values as-is
-               } else {
-                  for (k = 0; k < s->img_n; ++k) tc[k] = (stbi_uc)(stbi__get16be(s) & 255) * stbi__depth_scale_table[z->depth]; // non 8-bit images will be larger
-               }
-            }
-            break;
-         }
-
-         case STBI__PNG_TYPE('I','D','A','T'): {
-            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
-            if (pal_img_n && !pal_len) return stbi__err("no PLTE","Corrupt PNG");
-            if (scan == STBI__SCAN_header) {
-               // header scan definitely stops at first IDAT
-               if (pal_img_n)
-                  s->img_n = pal_img_n;
-               return 1;
-            }
-            if (c.length > (1u << 30)) return stbi__err("IDAT size limit", "IDAT section larger than 2^30 bytes");
-            if ((int)(ioff + c.length) < (int)ioff) return 0;
-            if (ioff + c.length > idata_limit) {
-               stbi__uint32 idata_limit_old = idata_limit;
-               stbi_uc *p;
-               if (idata_limit == 0) idata_limit = c.length > 4096 ? c.length : 4096;
-               while (ioff + c.length > idata_limit)
-                  idata_limit *= 2;
-               STBI_NOTUSED(idata_limit_old);
-               p = (stbi_uc *) STBI_REALLOC_SIZED(z->idata, idata_limit_old, idata_limit); if (p == NULL) return stbi__err("outofmem", "Out of memory");
-               z->idata = p;
-            }
-            if (!stbi__getn(s, z->idata+ioff,c.length)) return stbi__err("outofdata","Corrupt PNG");
-            ioff += c.length;
-            break;
-         }
-
-         case STBI__PNG_TYPE('I','E','N','D'): {
-            stbi__uint32 raw_len, bpl;
-            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
-            if (scan != STBI__SCAN_load) return 1;
-            if (z->idata == NULL) return stbi__err("no IDAT","Corrupt PNG");
-            // initial guess for decoded data size to avoid unnecessary reallocs
-            bpl = (s->img_x * z->depth + 7) / 8; // bytes per line, per component
-            raw_len = bpl * s->img_y * s->img_n /* pixels */ + s->img_y /* filter mode per row */;
-            z->expanded = (stbi_uc *) stbi_zlib_decode_malloc_guesssize_headerflag((char *) z->idata, ioff, raw_len, (int *) &raw_len, !is_iphone);
-            if (z->expanded == NULL) return 0; // zlib should set error
-            STBI_FREE(z->idata); z->idata = NULL;
-            if ((req_comp == s->img_n+1 && req_comp != 3 && !pal_img_n) || has_trans)
-               s->img_out_n = s->img_n+1;
-            else
-               s->img_out_n = s->img_n;
-            if (!stbi__create_png_image(z, z->expanded, raw_len, s->img_out_n, z->depth, color, interlace)) return 0;
-            if (has_trans) {
-               if (z->depth == 16) {
-                  if (!stbi__compute_transparency16(z, tc16, s->img_out_n)) return 0;
-               } else {
-                  if (!stbi__compute_transparency(z, tc, s->img_out_n)) return 0;
-               }
-            }
-            if (is_iphone && stbi__de_iphone_flag && s->img_out_n > 2)
-               stbi__de_iphone(z);
-            if (pal_img_n) {
-               // pal_img_n == 3 or 4
-               s->img_n = pal_img_n; // record the actual colors we had
-               s->img_out_n = pal_img_n;
-               if (req_comp >= 3) s->img_out_n = req_comp;
-               if (!stbi__expand_png_palette(z, palette, pal_len, s->img_out_n))
-                  return 0;
-            } else if (has_trans) {
-               // non-paletted image with tRNS -> source image has (constant) alpha
-               ++s->img_n;
-            }
-            STBI_FREE(z->expanded); z->expanded = NULL;
-            // end of PNG chunk, read and skip CRC
-            stbi__get32be(s);
-            return 1;
-         }
-
-         default:
-            // if critical, fail
-            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
-            if ((c.type & (1 << 29)) == 0) {
-               #ifndef STBI_NO_FAILURE_STRINGS
-               // not threadsafe
-               static char invalid_chunk[] = "XXXX PNG chunk not known";
-               invalid_chunk[0] = STBI__BYTECAST(c.type >> 24);
-               invalid_chunk[1] = STBI__BYTECAST(c.type >> 16);
-               invalid_chunk[2] = STBI__BYTECAST(c.type >>  8);
-               invalid_chunk[3] = STBI__BYTECAST(c.type >>  0);
-               #endif
-               return stbi__err(invalid_chunk, "PNG not supported: unknown PNG chunk type");
-            }
-            stbi__skip(s, c.length);
-            break;
-      }
-      // end of PNG chunk, read and skip CRC
-      stbi__get32be(s);
-   }
-}
-
-static void *stbi__do_png(stbi__png *p, int *x, int *y, int *n, int req_comp, stbi__result_info *ri)
-{
-   void *result=NULL;
-   if (req_comp < 0 || req_comp > 4) return stbi__errpuc("bad req_comp", "Internal error");
-   if (stbi__parse_png_file(p, STBI__SCAN_load, req_comp)) {
-      if (p->depth <= 8)
-         ri->bits_per_channel = 8;
-      else if (p->depth == 16)
-         ri->bits_per_channel = 16;
-      else
-         return stbi__errpuc("bad bits_per_channel", "PNG not supported: unsupported color depth");
-      result = p->out;
-      p->out = NULL;
-      if (req_comp && req_comp != p->s->img_out_n) {
-         if (ri->bits_per_channel == 8)
-            result = stbi__convert_format((unsigned char *) result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y);
-         else
-            result = stbi__convert_format16((stbi__uint16 *) result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y);
-         p->s->img_out_n = req_comp;
-         if (result == NULL) return result;
-      }
-      *x = p->s->img_x;
-      *y = p->s->img_y;
-      if (n) *n = p->s->img_n;
-   }
-   STBI_FREE(p->out);      p->out      = NULL;
-   STBI_FREE(p->expanded); p->expanded = NULL;
-   STBI_FREE(p->idata);    p->idata    = NULL;
-
-   return result;
-}
-
-static void *stbi__png_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
-{
-   stbi__png p;
-   p.s = s;
-   return stbi__do_png(&p, x,y,comp,req_comp, ri);
-}
-
-static int stbi__png_test(stbi__context *s)
-{
-   int r;
-   r = stbi__check_png_header(s);
-   stbi__rewind(s);
-   return r;
-}
-
-static int stbi__png_info_raw(stbi__png *p, int *x, int *y, int *comp)
-{
-   if (!stbi__parse_png_file(p, STBI__SCAN_header, 0)) {
-      stbi__rewind( p->s );
-      return 0;
-   }
-   if (x) *x = p->s->img_x;
-   if (y) *y = p->s->img_y;
-   if (comp) *comp = p->s->img_n;
-   return 1;
-}
-
-static int stbi__png_info(stbi__context *s, int *x, int *y, int *comp)
-{
-   stbi__png p;
-   p.s = s;
-   return stbi__png_info_raw(&p, x, y, comp);
-}
-
-static int stbi__png_is16(stbi__context *s)
-{
-   stbi__png p;
-   p.s = s;
-   if (!stbi__png_info_raw(&p, NULL, NULL, NULL))
-	   return 0;
-   if (p.depth != 16) {
-      stbi__rewind(p.s);
-      return 0;
-   }
-   return 1;
-}
-#endif
-
-// Microsoft/Windows BMP image
-
-#ifndef STBI_NO_BMP
-static int stbi__bmp_test_raw(stbi__context *s)
-{
-   int r;
-   int sz;
-   if (stbi__get8(s) != 'B') return 0;
-   if (stbi__get8(s) != 'M') return 0;
-   stbi__get32le(s); // discard filesize
-   stbi__get16le(s); // discard reserved
-   stbi__get16le(s); // discard reserved
-   stbi__get32le(s); // discard data offset
-   sz = stbi__get32le(s);
-   r = (sz == 12 || sz == 40 || sz == 56 || sz == 108 || sz == 124);
-   return r;
-}
-
-static int stbi__bmp_test(stbi__context *s)
-{
-   int r = stbi__bmp_test_raw(s);
-   stbi__rewind(s);
-   return r;
-}
-
-
-// returns 0..31 for the highest set bit
-static int stbi__high_bit(unsigned int z)
-{
-   int n=0;
-   if (z == 0) return -1;
-   if (z >= 0x10000) { n += 16; z >>= 16; }
-   if (z >= 0x00100) { n +=  8; z >>=  8; }
-   if (z >= 0x00010) { n +=  4; z >>=  4; }
-   if (z >= 0x00004) { n +=  2; z >>=  2; }
-   if (z >= 0x00002) { n +=  1;/* >>=  1;*/ }
-   return n;
-}
-
-static int stbi__bitcount(unsigned int a)
-{
-   a = (a & 0x55555555) + ((a >>  1) & 0x55555555); // max 2
-   a = (a & 0x33333333) + ((a >>  2) & 0x33333333); // max 4
-   a = (a + (a >> 4)) & 0x0f0f0f0f; // max 8 per 4, now 8 bits
-   a = (a + (a >> 8)); // max 16 per 8 bits
-   a = (a + (a >> 16)); // max 32 per 8 bits
-   return a & 0xff;
-}
-
-// extract an arbitrarily-aligned N-bit value (N=bits)
-// from v, and then make it 8-bits long and fractionally
-// extend it to full full range.
-static int stbi__shiftsigned(unsigned int v, int shift, int bits)
-{
-   static unsigned int mul_table[9] = {
-      0,
-      0xff/*0b11111111*/, 0x55/*0b01010101*/, 0x49/*0b01001001*/, 0x11/*0b00010001*/,
-      0x21/*0b00100001*/, 0x41/*0b01000001*/, 0x81/*0b10000001*/, 0x01/*0b00000001*/,
-   };
-   static unsigned int shift_table[9] = {
-      0, 0,0,1,0,2,4,6,0,
-   };
-   if (shift < 0)
-      v <<= -shift;
-   else
-      v >>= shift;
-   STBI_ASSERT(v < 256);
-   v >>= (8-bits);
-   STBI_ASSERT(bits >= 0 && bits <= 8);
-   return (int) ((unsigned) v * mul_table[bits]) >> shift_table[bits];
-}
-
-typedef struct
-{
-   int bpp, offset, hsz;
-   unsigned int mr,mg,mb,ma, all_a;
-   int extra_read;
-} stbi__bmp_data;
-
-static int stbi__bmp_set_mask_defaults(stbi__bmp_data *info, int compress)
-{
-   // BI_BITFIELDS specifies masks explicitly, don't override
-   if (compress == 3)
-      return 1;
-
-   if (compress == 0) {
-      if (info->bpp == 16) {
-         info->mr = 31u << 10;
-         info->mg = 31u <<  5;
-         info->mb = 31u <<  0;
-      } else if (info->bpp == 32) {
-         info->mr = 0xffu << 16;
-         info->mg = 0xffu <<  8;
-         info->mb = 0xffu <<  0;
-         info->ma = 0xffu << 24;
-         info->all_a = 0; // if all_a is 0 at end, then we loaded alpha channel but it was all 0
-      } else {
-         // otherwise, use defaults, which is all-0
-         info->mr = info->mg = info->mb = info->ma = 0;
-      }
-      return 1;
-   }
-   return 0; // error
-}
-
-static void *stbi__bmp_parse_header(stbi__context *s, stbi__bmp_data *info)
-{
-   int hsz;
-   if (stbi__get8(s) != 'B' || stbi__get8(s) != 'M') return stbi__errpuc("not BMP", "Corrupt BMP");
-   stbi__get32le(s); // discard filesize
-   stbi__get16le(s); // discard reserved
-   stbi__get16le(s); // discard reserved
-   info->offset = stbi__get32le(s);
-   info->hsz = hsz = stbi__get32le(s);
-   info->mr = info->mg = info->mb = info->ma = 0;
-   info->extra_read = 14;
-
-   if (info->offset < 0) return stbi__errpuc("bad BMP", "bad BMP");
-
-   if (hsz != 12 && hsz != 40 && hsz != 56 && hsz != 108 && hsz != 124) return stbi__errpuc("unknown BMP", "BMP type not supported: unknown");
-   if (hsz == 12) {
-      s->img_x = stbi__get16le(s);
-      s->img_y = stbi__get16le(s);
-   } else {
-      s->img_x = stbi__get32le(s);
-      s->img_y = stbi__get32le(s);
-   }
-   if (stbi__get16le(s) != 1) return stbi__errpuc("bad BMP", "bad BMP");
-   info->bpp = stbi__get16le(s);
-   if (hsz != 12) {
-      int compress = stbi__get32le(s);
-      if (compress == 1 || compress == 2) return stbi__errpuc("BMP RLE", "BMP type not supported: RLE");
-      if (compress >= 4) return stbi__errpuc("BMP JPEG/PNG", "BMP type not supported: unsupported compression"); // this includes PNG/JPEG modes
-      if (compress == 3 && info->bpp != 16 && info->bpp != 32) return stbi__errpuc("bad BMP", "bad BMP"); // bitfields requires 16 or 32 bits/pixel
-      stbi__get32le(s); // discard sizeof
-      stbi__get32le(s); // discard hres
-      stbi__get32le(s); // discard vres
-      stbi__get32le(s); // discard colorsused
-      stbi__get32le(s); // discard max important
-      if (hsz == 40 || hsz == 56) {
-         if (hsz == 56) {
-            stbi__get32le(s);
-            stbi__get32le(s);
-            stbi__get32le(s);
-            stbi__get32le(s);
-         }
-         if (info->bpp == 16 || info->bpp == 32) {
-            if (compress == 0) {
-               stbi__bmp_set_mask_defaults(info, compress);
-            } else if (compress == 3) {
-               info->mr = stbi__get32le(s);
-               info->mg = stbi__get32le(s);
-               info->mb = stbi__get32le(s);
-               info->extra_read += 12;
-               // not documented, but generated by photoshop and handled by mspaint
-               if (info->mr == info->mg && info->mg == info->mb) {
-                  // ?!?!?
-                  return stbi__errpuc("bad BMP", "bad BMP");
-               }
-            } else
-               return stbi__errpuc("bad BMP", "bad BMP");
-         }
-      } else {
-         // V4/V5 header
-         int i;
-         if (hsz != 108 && hsz != 124)
-            return stbi__errpuc("bad BMP", "bad BMP");
-         info->mr = stbi__get32le(s);
-         info->mg = stbi__get32le(s);
-         info->mb = stbi__get32le(s);
-         info->ma = stbi__get32le(s);
-         if (compress != 3) // override mr/mg/mb unless in BI_BITFIELDS mode, as per docs
-            stbi__bmp_set_mask_defaults(info, compress);
-         stbi__get32le(s); // discard color space
-         for (i=0; i < 12; ++i)
-            stbi__get32le(s); // discard color space parameters
-         if (hsz == 124) {
-            stbi__get32le(s); // discard rendering intent
-            stbi__get32le(s); // discard offset of profile data
-            stbi__get32le(s); // discard size of profile data
-            stbi__get32le(s); // discard reserved
-         }
-      }
-   }
-   return (void *) 1;
-}
-
-
-static void *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
-{
-   stbi_uc *out;
-   unsigned int mr=0,mg=0,mb=0,ma=0, all_a;
-   stbi_uc pal[256][4];
-   int psize=0,i,j,width;
-   int flip_vertically, pad, target;
-   stbi__bmp_data info;
-   STBI_NOTUSED(ri);
-
-   info.all_a = 255;
-   if (stbi__bmp_parse_header(s, &info) == NULL)
-      return NULL; // error code already set
-
-   flip_vertically = ((int) s->img_y) > 0;
-   s->img_y = abs((int) s->img_y);
-
-   if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
-   if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
-
-   mr = info.mr;
-   mg = info.mg;
-   mb = info.mb;
-   ma = info.ma;
-   all_a = info.all_a;
-
-   if (info.hsz == 12) {
-      if (info.bpp < 24)
-         psize = (info.offset - info.extra_read - 24) / 3;
-   } else {
-      if (info.bpp < 16)
-         psize = (info.offset - info.extra_read - info.hsz) >> 2;
-   }
-   if (psize == 0) {
-      // accept some number of extra bytes after the header, but if the offset points either to before
-      // the header ends or implies a large amount of extra data, reject the file as malformed
-      int bytes_read_so_far = s->callback_already_read + (int)(s->img_buffer - s->img_buffer_original);
-      int header_limit = 1024; // max we actually read is below 256 bytes currently.
-      int extra_data_limit = 256*4; // what ordinarily goes here is a palette; 256 entries*4 bytes is its max size.
-      if (bytes_read_so_far <= 0 || bytes_read_so_far > header_limit) {
-         return stbi__errpuc("bad header", "Corrupt BMP");
-      }
-      // we established that bytes_read_so_far is positive and sensible.
-      // the first half of this test rejects offsets that are either too small positives, or
-      // negative, and guarantees that info.offset >= bytes_read_so_far > 0. this in turn
-      // ensures the number computed in the second half of the test can't overflow.
-      if (info.offset < bytes_read_so_far || info.offset - bytes_read_so_far > extra_data_limit) {
-         return stbi__errpuc("bad offset", "Corrupt BMP");
-      } else {
-         stbi__skip(s, info.offset - bytes_read_so_far);
-      }
-   }
-
-   if (info.bpp == 24 && ma == 0xff000000)
-      s->img_n = 3;
-   else
-      s->img_n = ma ? 4 : 3;
-   if (req_comp && req_comp >= 3) // we can directly decode 3 or 4
-      target = req_comp;
-   else
-      target = s->img_n; // if they want monochrome, we'll post-convert
-
-   // sanity-check size
-   if (!stbi__mad3sizes_valid(target, s->img_x, s->img_y, 0))
-      return stbi__errpuc("too large", "Corrupt BMP");
-
-   out = (stbi_uc *) stbi__malloc_mad3(target, s->img_x, s->img_y, 0);
-   if (!out) return stbi__errpuc("outofmem", "Out of memory");
-   if (info.bpp < 16) {
-      int z=0;
-      if (psize == 0 || psize > 256) { STBI_FREE(out); return stbi__errpuc("invalid", "Corrupt BMP"); }
-      for (i=0; i < psize; ++i) {
-         pal[i][2] = stbi__get8(s);
-         pal[i][1] = stbi__get8(s);
-         pal[i][0] = stbi__get8(s);
-         if (info.hsz != 12) stbi__get8(s);
-         pal[i][3] = 255;
-      }
-      stbi__skip(s, info.offset - info.extra_read - info.hsz - psize * (info.hsz == 12 ? 3 : 4));
-      if (info.bpp == 1) width = (s->img_x + 7) >> 3;
-      else if (info.bpp == 4) width = (s->img_x + 1) >> 1;
-      else if (info.bpp == 8) width = s->img_x;
-      else { STBI_FREE(out); return stbi__errpuc("bad bpp", "Corrupt BMP"); }
-      pad = (-width)&3;
-      if (info.bpp == 1) {
-         for (j=0; j < (int) s->img_y; ++j) {
-            int bit_offset = 7, v = stbi__get8(s);
-            for (i=0; i < (int) s->img_x; ++i) {
-               int color = (v>>bit_offset)&0x1;
-               out[z++] = pal[color][0];
-               out[z++] = pal[color][1];
-               out[z++] = pal[color][2];
-               if (target == 4) out[z++] = 255;
-               if (i+1 == (int) s->img_x) break;
-               if((--bit_offset) < 0) {
-                  bit_offset = 7;
-                  v = stbi__get8(s);
-               }
-            }
-            stbi__skip(s, pad);
-         }
-      } else {
-         for (j=0; j < (int) s->img_y; ++j) {
-            for (i=0; i < (int) s->img_x; i += 2) {
-               int v=stbi__get8(s),v2=0;
-               if (info.bpp == 4) {
-                  v2 = v & 15;
-                  v >>= 4;
-               }
-               out[z++] = pal[v][0];
-               out[z++] = pal[v][1];
-               out[z++] = pal[v][2];
-               if (target == 4) out[z++] = 255;
-               if (i+1 == (int) s->img_x) break;
-               v = (info.bpp == 8) ? stbi__get8(s) : v2;
-               out[z++] = pal[v][0];
-               out[z++] = pal[v][1];
-               out[z++] = pal[v][2];
-               if (target == 4) out[z++] = 255;
-            }
-            stbi__skip(s, pad);
-         }
-      }
-   } else {
-      int rshift=0,gshift=0,bshift=0,ashift=0,rcount=0,gcount=0,bcount=0,acount=0;
-      int z = 0;
-      int easy=0;
-      stbi__skip(s, info.offset - info.extra_read - info.hsz);
-      if (info.bpp == 24) width = 3 * s->img_x;
-      else if (info.bpp == 16) width = 2*s->img_x;
-      else /* bpp = 32 and pad = 0 */ width=0;
-      pad = (-width) & 3;
-      if (info.bpp == 24) {
-         easy = 1;
-      } else if (info.bpp == 32) {
-         if (mb == 0xff && mg == 0xff00 && mr == 0x00ff0000 && ma == 0xff000000)
-            easy = 2;
-      }
-      if (!easy) {
-         if (!mr || !mg || !mb) { STBI_FREE(out); return stbi__errpuc("bad masks", "Corrupt BMP"); }
-         // right shift amt to put high bit in position #7
-         rshift = stbi__high_bit(mr)-7; rcount = stbi__bitcount(mr);
-         gshift = stbi__high_bit(mg)-7; gcount = stbi__bitcount(mg);
-         bshift = stbi__high_bit(mb)-7; bcount = stbi__bitcount(mb);
-         ashift = stbi__high_bit(ma)-7; acount = stbi__bitcount(ma);
-         if (rcount > 8 || gcount > 8 || bcount > 8 || acount > 8) { STBI_FREE(out); return stbi__errpuc("bad masks", "Corrupt BMP"); }
-      }
-      for (j=0; j < (int) s->img_y; ++j) {
-         if (easy) {
-            for (i=0; i < (int) s->img_x; ++i) {
-               unsigned char a;
-               out[z+2] = stbi__get8(s);
-               out[z+1] = stbi__get8(s);
-               out[z+0] = stbi__get8(s);
-               z += 3;
-               a = (easy == 2 ? stbi__get8(s) : 255);
-               all_a |= a;
-               if (target == 4) out[z++] = a;
-            }
-         } else {
-            int bpp = info.bpp;
-            for (i=0; i < (int) s->img_x; ++i) {
-               stbi__uint32 v = (bpp == 16 ? (stbi__uint32) stbi__get16le(s) : stbi__get32le(s));
-               unsigned int a;
-               out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mr, rshift, rcount));
-               out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mg, gshift, gcount));
-               out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mb, bshift, bcount));
-               a = (ma ? stbi__shiftsigned(v & ma, ashift, acount) : 255);
-               all_a |= a;
-               if (target == 4) out[z++] = STBI__BYTECAST(a);
-            }
-         }
-         stbi__skip(s, pad);
-      }
-   }
-
-   // if alpha channel is all 0s, replace with all 255s
-   if (target == 4 && all_a == 0)
-      for (i=4*s->img_x*s->img_y-1; i >= 0; i -= 4)
-         out[i] = 255;
-
-   if (flip_vertically) {
-      stbi_uc t;
-      for (j=0; j < (int) s->img_y>>1; ++j) {
-         stbi_uc *p1 = out +      j     *s->img_x*target;
-         stbi_uc *p2 = out + (s->img_y-1-j)*s->img_x*target;
-         for (i=0; i < (int) s->img_x*target; ++i) {
-            t = p1[i]; p1[i] = p2[i]; p2[i] = t;
-         }
-      }
-   }
-
-   if (req_comp && req_comp != target) {
-      out = stbi__convert_format(out, target, req_comp, s->img_x, s->img_y);
-      if (out == NULL) return out; // stbi__convert_format frees input on failure
-   }
-
-   *x = s->img_x;
-   *y = s->img_y;
-   if (comp) *comp = s->img_n;
-   return out;
-}
-#endif
-
-// Targa Truevision - TGA
-// by Jonathan Dummer
-#ifndef STBI_NO_TGA
-// returns STBI_rgb or whatever, 0 on error
-static int stbi__tga_get_comp(int bits_per_pixel, int is_grey, int* is_rgb16)
-{
-   // only RGB or RGBA (incl. 16bit) or grey allowed
-   if (is_rgb16) *is_rgb16 = 0;
-   switch(bits_per_pixel) {
-      case 8:  return STBI_grey;
-      case 16: if(is_grey) return STBI_grey_alpha;
-               // fallthrough
-      case 15: if(is_rgb16) *is_rgb16 = 1;
-               return STBI_rgb;
-      case 24: // fallthrough
-      case 32: return bits_per_pixel/8;
-      default: return 0;
-   }
-}
-
-static int stbi__tga_info(stbi__context *s, int *x, int *y, int *comp)
-{
-    int tga_w, tga_h, tga_comp, tga_image_type, tga_bits_per_pixel, tga_colormap_bpp;
-    int sz, tga_colormap_type;
-    stbi__get8(s);                   // discard Offset
-    tga_colormap_type = stbi__get8(s); // colormap type
-    if( tga_colormap_type > 1 ) {
-        stbi__rewind(s);
-        return 0;      // only RGB or indexed allowed
-    }
-    tga_image_type = stbi__get8(s); // image type
-    if ( tga_colormap_type == 1 ) { // colormapped (paletted) image
-        if (tga_image_type != 1 && tga_image_type != 9) {
-            stbi__rewind(s);
-            return 0;
-        }
-        stbi__skip(s,4);       // skip index of first colormap entry and number of entries
-        sz = stbi__get8(s);    //   check bits per palette color entry
-        if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) {
-            stbi__rewind(s);
-            return 0;
-        }
-        stbi__skip(s,4);       // skip image x and y origin
-        tga_colormap_bpp = sz;
-    } else { // "normal" image w/o colormap - only RGB or grey allowed, +/- RLE
-        if ( (tga_image_type != 2) && (tga_image_type != 3) && (tga_image_type != 10) && (tga_image_type != 11) ) {
-            stbi__rewind(s);
-            return 0; // only RGB or grey allowed, +/- RLE
-        }
-        stbi__skip(s,9); // skip colormap specification and image x/y origin
-        tga_colormap_bpp = 0;
-    }
-    tga_w = stbi__get16le(s);
-    if( tga_w < 1 ) {
-        stbi__rewind(s);
-        return 0;   // test width
-    }
-    tga_h = stbi__get16le(s);
-    if( tga_h < 1 ) {
-        stbi__rewind(s);
-        return 0;   // test height
-    }
-    tga_bits_per_pixel = stbi__get8(s); // bits per pixel
-    stbi__get8(s); // ignore alpha bits
-    if (tga_colormap_bpp != 0) {
-        if((tga_bits_per_pixel != 8) && (tga_bits_per_pixel != 16)) {
-            // when using a colormap, tga_bits_per_pixel is the size of the indexes
-            // I don't think anything but 8 or 16bit indexes makes sense
-            stbi__rewind(s);
-            return 0;
-        }
-        tga_comp = stbi__tga_get_comp(tga_colormap_bpp, 0, NULL);
-    } else {
-        tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3) || (tga_image_type == 11), NULL);
-    }
-    if(!tga_comp) {
-      stbi__rewind(s);
-      return 0;
-    }
-    if (x) *x = tga_w;
-    if (y) *y = tga_h;
-    if (comp) *comp = tga_comp;
-    return 1;                   // seems to have passed everything
-}
-
-static int stbi__tga_test(stbi__context *s)
-{
-   int res = 0;
-   int sz, tga_color_type;
-   stbi__get8(s);      //   discard Offset
-   tga_color_type = stbi__get8(s);   //   color type
-   if ( tga_color_type > 1 ) goto errorEnd;   //   only RGB or indexed allowed
-   sz = stbi__get8(s);   //   image type
-   if ( tga_color_type == 1 ) { // colormapped (paletted) image
-      if (sz != 1 && sz != 9) goto errorEnd; // colortype 1 demands image type 1 or 9
-      stbi__skip(s,4);       // skip index of first colormap entry and number of entries
-      sz = stbi__get8(s);    //   check bits per palette color entry
-      if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) goto errorEnd;
-      stbi__skip(s,4);       // skip image x and y origin
-   } else { // "normal" image w/o colormap
-      if ( (sz != 2) && (sz != 3) && (sz != 10) && (sz != 11) ) goto errorEnd; // only RGB or grey allowed, +/- RLE
-      stbi__skip(s,9); // skip colormap specification and image x/y origin
-   }
-   if ( stbi__get16le(s) < 1 ) goto errorEnd;      //   test width
-   if ( stbi__get16le(s) < 1 ) goto errorEnd;      //   test height
-   sz = stbi__get8(s);   //   bits per pixel
-   if ( (tga_color_type == 1) && (sz != 8) && (sz != 16) ) goto errorEnd; // for colormapped images, bpp is size of an index
-   if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) goto errorEnd;
-
-   res = 1; // if we got this far, everything's good and we can return 1 instead of 0
-
-errorEnd:
-   stbi__rewind(s);
-   return res;
-}
-
-// read 16bit value and convert to 24bit RGB
-static void stbi__tga_read_rgb16(stbi__context *s, stbi_uc* out)
-{
-   stbi__uint16 px = (stbi__uint16)stbi__get16le(s);
-   stbi__uint16 fiveBitMask = 31;
-   // we have 3 channels with 5bits each
-   int r = (px >> 10) & fiveBitMask;
-   int g = (px >> 5) & fiveBitMask;
-   int b = px & fiveBitMask;
-   // Note that this saves the data in RGB(A) order, so it doesn't need to be swapped later
-   out[0] = (stbi_uc)((r * 255)/31);
-   out[1] = (stbi_uc)((g * 255)/31);
-   out[2] = (stbi_uc)((b * 255)/31);
-
-   // some people claim that the most significant bit might be used for alpha
-   // (possibly if an alpha-bit is set in the "image descriptor byte")
-   // but that only made 16bit test images completely translucent..
-   // so let's treat all 15 and 16bit TGAs as RGB with no alpha.
-}
-
-static void *stbi__tga_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
-{
-   //   read in the TGA header stuff
-   int tga_offset = stbi__get8(s);
-   int tga_indexed = stbi__get8(s);
-   int tga_image_type = stbi__get8(s);
-   int tga_is_RLE = 0;
-   int tga_palette_start = stbi__get16le(s);
-   int tga_palette_len = stbi__get16le(s);
-   int tga_palette_bits = stbi__get8(s);
-   int tga_x_origin = stbi__get16le(s);
-   int tga_y_origin = stbi__get16le(s);
-   int tga_width = stbi__get16le(s);
-   int tga_height = stbi__get16le(s);
-   int tga_bits_per_pixel = stbi__get8(s);
-   int tga_comp, tga_rgb16=0;
-   int tga_inverted = stbi__get8(s);
-   // int tga_alpha_bits = tga_inverted & 15; // the 4 lowest bits - unused (useless?)
-   //   image data
-   unsigned char *tga_data;
-   unsigned char *tga_palette = NULL;
-   int i, j;
-   unsigned char raw_data[4] = {0};
-   int RLE_count = 0;
-   int RLE_repeating = 0;
-   int read_next_pixel = 1;
-   STBI_NOTUSED(ri);
-   STBI_NOTUSED(tga_x_origin); // @TODO
-   STBI_NOTUSED(tga_y_origin); // @TODO
-
-   if (tga_height > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
-   if (tga_width > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
-
-   //   do a tiny bit of precessing
-   if ( tga_image_type >= 8 )
-   {
-      tga_image_type -= 8;
-      tga_is_RLE = 1;
-   }
-   tga_inverted = 1 - ((tga_inverted >> 5) & 1);
-
-   //   If I'm paletted, then I'll use the number of bits from the palette
-   if ( tga_indexed ) tga_comp = stbi__tga_get_comp(tga_palette_bits, 0, &tga_rgb16);
-   else tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3), &tga_rgb16);
-
-   if(!tga_comp) // shouldn't really happen, stbi__tga_test() should have ensured basic consistency
-      return stbi__errpuc("bad format", "Can't find out TGA pixelformat");
-
-   //   tga info
-   *x = tga_width;
-   *y = tga_height;
-   if (comp) *comp = tga_comp;
-
-   if (!stbi__mad3sizes_valid(tga_width, tga_height, tga_comp, 0))
-      return stbi__errpuc("too large", "Corrupt TGA");
-
-   tga_data = (unsigned char*)stbi__malloc_mad3(tga_width, tga_height, tga_comp, 0);
-   if (!tga_data) return stbi__errpuc("outofmem", "Out of memory");
-
-   // skip to the data's starting position (offset usually = 0)
-   stbi__skip(s, tga_offset );
-
-   if ( !tga_indexed && !tga_is_RLE && !tga_rgb16 ) {
-      for (i=0; i < tga_height; ++i) {
-         int row = tga_inverted ? tga_height -i - 1 : i;
-         stbi_uc *tga_row = tga_data + row*tga_width*tga_comp;
-         stbi__getn(s, tga_row, tga_width * tga_comp);
-      }
-   } else  {
-      //   do I need to load a palette?
-      if ( tga_indexed)
-      {
-         if (tga_palette_len == 0) {  /* you have to have at least one entry! */
-            STBI_FREE(tga_data);
-            return stbi__errpuc("bad palette", "Corrupt TGA");
-         }
-
-         //   any data to skip? (offset usually = 0)
-         stbi__skip(s, tga_palette_start );
-         //   load the palette
-         tga_palette = (unsigned char*)stbi__malloc_mad2(tga_palette_len, tga_comp, 0);
-         if (!tga_palette) {
-            STBI_FREE(tga_data);
-            return stbi__errpuc("outofmem", "Out of memory");
-         }
-         if (tga_rgb16) {
-            stbi_uc *pal_entry = tga_palette;
-            STBI_ASSERT(tga_comp == STBI_rgb);
-            for (i=0; i < tga_palette_len; ++i) {
-               stbi__tga_read_rgb16(s, pal_entry);
-               pal_entry += tga_comp;
-            }
-         } else if (!stbi__getn(s, tga_palette, tga_palette_len * tga_comp)) {
-               STBI_FREE(tga_data);
-               STBI_FREE(tga_palette);
-               return stbi__errpuc("bad palette", "Corrupt TGA");
-         }
-      }
-      //   load the data
-      for (i=0; i < tga_width * tga_height; ++i)
-      {
-         //   if I'm in RLE mode, do I need to get a RLE stbi__pngchunk?
-         if ( tga_is_RLE )
-         {
-            if ( RLE_count == 0 )
-            {
-               //   yep, get the next byte as a RLE command
-               int RLE_cmd = stbi__get8(s);
-               RLE_count = 1 + (RLE_cmd & 127);
-               RLE_repeating = RLE_cmd >> 7;
-               read_next_pixel = 1;
-            } else if ( !RLE_repeating )
-            {
-               read_next_pixel = 1;
-            }
-         } else
-         {
-            read_next_pixel = 1;
-         }
-         //   OK, if I need to read a pixel, do it now
-         if ( read_next_pixel )
-         {
-            //   load however much data we did have
-            if ( tga_indexed )
-            {
-               // read in index, then perform the lookup
-               int pal_idx = (tga_bits_per_pixel == 8) ? stbi__get8(s) : stbi__get16le(s);
-               if ( pal_idx >= tga_palette_len ) {
-                  // invalid index
-                  pal_idx = 0;
-               }
-               pal_idx *= tga_comp;
-               for (j = 0; j < tga_comp; ++j) {
-                  raw_data[j] = tga_palette[pal_idx+j];
-               }
-            } else if(tga_rgb16) {
-               STBI_ASSERT(tga_comp == STBI_rgb);
-               stbi__tga_read_rgb16(s, raw_data);
-            } else {
-               //   read in the data raw
-               for (j = 0; j < tga_comp; ++j) {
-                  raw_data[j] = stbi__get8(s);
-               }
-            }
-            //   clear the reading flag for the next pixel
-            read_next_pixel = 0;
-         } // end of reading a pixel
-
-         // copy data
-         for (j = 0; j < tga_comp; ++j)
-           tga_data[i*tga_comp+j] = raw_data[j];
-
-         //   in case we're in RLE mode, keep counting down
-         --RLE_count;
-      }
-      //   do I need to invert the image?
-      if ( tga_inverted )
-      {
-         for (j = 0; j*2 < tga_height; ++j)
-         {
-            int index1 = j * tga_width * tga_comp;
-            int index2 = (tga_height - 1 - j) * tga_width * tga_comp;
-            for (i = tga_width * tga_comp; i > 0; --i)
-            {
-               unsigned char temp = tga_data[index1];
-               tga_data[index1] = tga_data[index2];
-               tga_data[index2] = temp;
-               ++index1;
-               ++index2;
-            }
-         }
-      }
-      //   clear my palette, if I had one
-      if ( tga_palette != NULL )
-      {
-         STBI_FREE( tga_palette );
-      }
-   }
-
-   // swap RGB - if the source data was RGB16, it already is in the right order
-   if (tga_comp >= 3 && !tga_rgb16)
-   {
-      unsigned char* tga_pixel = tga_data;
-      for (i=0; i < tga_width * tga_height; ++i)
-      {
-         unsigned char temp = tga_pixel[0];
-         tga_pixel[0] = tga_pixel[2];
-         tga_pixel[2] = temp;
-         tga_pixel += tga_comp;
-      }
-   }
-
-   // convert to target component count
-   if (req_comp && req_comp != tga_comp)
-      tga_data = stbi__convert_format(tga_data, tga_comp, req_comp, tga_width, tga_height);
-
-   //   the things I do to get rid of an error message, and yet keep
-   //   Microsoft's C compilers happy... [8^(
-   tga_palette_start = tga_palette_len = tga_palette_bits =
-         tga_x_origin = tga_y_origin = 0;
-   STBI_NOTUSED(tga_palette_start);
-   //   OK, done
-   return tga_data;
-}
-#endif
-
-// *************************************************************************************************
-// Photoshop PSD loader -- PD by Thatcher Ulrich, integration by Nicolas Schulz, tweaked by STB
-
-#ifndef STBI_NO_PSD
-static int stbi__psd_test(stbi__context *s)
-{
-   int r = (stbi__get32be(s) == 0x38425053);
-   stbi__rewind(s);
-   return r;
-}
-
-static int stbi__psd_decode_rle(stbi__context *s, stbi_uc *p, int pixelCount)
-{
-   int count, nleft, len;
-
-   count = 0;
-   while ((nleft = pixelCount - count) > 0) {
-      len = stbi__get8(s);
-      if (len == 128) {
-         // No-op.
-      } else if (len < 128) {
-         // Copy next len+1 bytes literally.
-         len++;
-         if (len > nleft) return 0; // corrupt data
-         count += len;
-         while (len) {
-            *p = stbi__get8(s);
-            p += 4;
-            len--;
-         }
-      } else if (len > 128) {
-         stbi_uc   val;
-         // Next -len+1 bytes in the dest are replicated from next source byte.
-         // (Interpret len as a negative 8-bit int.)
-         len = 257 - len;
-         if (len > nleft) return 0; // corrupt data
-         val = stbi__get8(s);
-         count += len;
-         while (len) {
-            *p = val;
-            p += 4;
-            len--;
-         }
-      }
-   }
-
-   return 1;
-}
-
-static void *stbi__psd_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc)
-{
-   int pixelCount;
-   int channelCount, compression;
-   int channel, i;
-   int bitdepth;
-   int w,h;
-   stbi_uc *out;
-   STBI_NOTUSED(ri);
-
-   // Check identifier
-   if (stbi__get32be(s) != 0x38425053)   // "8BPS"
-      return stbi__errpuc("not PSD", "Corrupt PSD image");
-
-   // Check file type version.
-   if (stbi__get16be(s) != 1)
-      return stbi__errpuc("wrong version", "Unsupported version of PSD image");
-
-   // Skip 6 reserved bytes.
-   stbi__skip(s, 6 );
-
-   // Read the number of channels (R, G, B, A, etc).
-   channelCount = stbi__get16be(s);
-   if (channelCount < 0 || channelCount > 16)
-      return stbi__errpuc("wrong channel count", "Unsupported number of channels in PSD image");
-
-   // Read the rows and columns of the image.
-   h = stbi__get32be(s);
-   w = stbi__get32be(s);
-
-   if (h > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
-   if (w > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
-
-   // Make sure the depth is 8 bits.
-   bitdepth = stbi__get16be(s);
-   if (bitdepth != 8 && bitdepth != 16)
-      return stbi__errpuc("unsupported bit depth", "PSD bit depth is not 8 or 16 bit");
-
-   // Make sure the color mode is RGB.
-   // Valid options are:
-   //   0: Bitmap
-   //   1: Grayscale
-   //   2: Indexed color
-   //   3: RGB color
-   //   4: CMYK color
-   //   7: Multichannel
-   //   8: Duotone
-   //   9: Lab color
-   if (stbi__get16be(s) != 3)
-      return stbi__errpuc("wrong color format", "PSD is not in RGB color format");
-
-   // Skip the Mode Data.  (It's the palette for indexed color; other info for other modes.)
-   stbi__skip(s,stbi__get32be(s) );
-
-   // Skip the image resources.  (resolution, pen tool paths, etc)
-   stbi__skip(s, stbi__get32be(s) );
-
-   // Skip the reserved data.
-   stbi__skip(s, stbi__get32be(s) );
-
-   // Find out if the data is compressed.
-   // Known values:
-   //   0: no compression
-   //   1: RLE compressed
-   compression = stbi__get16be(s);
-   if (compression > 1)
-      return stbi__errpuc("bad compression", "PSD has an unknown compression format");
-
-   // Check size
-   if (!stbi__mad3sizes_valid(4, w, h, 0))
-      return stbi__errpuc("too large", "Corrupt PSD");
-
-   // Create the destination image.
-
-   if (!compression && bitdepth == 16 && bpc == 16) {
-      out = (stbi_uc *) stbi__malloc_mad3(8, w, h, 0);
-      ri->bits_per_channel = 16;
-   } else
-      out = (stbi_uc *) stbi__malloc(4 * w*h);
-
-   if (!out) return stbi__errpuc("outofmem", "Out of memory");
-   pixelCount = w*h;
-
-   // Initialize the data to zero.
-   //memset( out, 0, pixelCount * 4 );
-
-   // Finally, the image data.
-   if (compression) {
-      // RLE as used by .PSD and .TIFF
-      // Loop until you get the number of unpacked bytes you are expecting:
-      //     Read the next source byte into n.
-      //     If n is between 0 and 127 inclusive, copy the next n+1 bytes literally.
-      //     Else if n is between -127 and -1 inclusive, copy the next byte -n+1 times.
-      //     Else if n is 128, noop.
-      // Endloop
-
-      // The RLE-compressed data is preceded by a 2-byte data count for each row in the data,
-      // which we're going to just skip.
-      stbi__skip(s, h * channelCount * 2 );
-
-      // Read the RLE data by channel.
-      for (channel = 0; channel < 4; channel++) {
-         stbi_uc *p;
-
-         p = out+channel;
-         if (channel >= channelCount) {
-            // Fill this channel with default data.
-            for (i = 0; i < pixelCount; i++, p += 4)
-               *p = (channel == 3 ? 255 : 0);
-         } else {
-            // Read the RLE data.
-            if (!stbi__psd_decode_rle(s, p, pixelCount)) {
-               STBI_FREE(out);
-               return stbi__errpuc("corrupt", "bad RLE data");
-            }
-         }
-      }
-
-   } else {
-      // We're at the raw image data.  It's each channel in order (Red, Green, Blue, Alpha, ...)
-      // where each channel consists of an 8-bit (or 16-bit) value for each pixel in the image.
-
-      // Read the data by channel.
-      for (channel = 0; channel < 4; channel++) {
-         if (channel >= channelCount) {
-            // Fill this channel with default data.
-            if (bitdepth == 16 && bpc == 16) {
-               stbi__uint16 *q = ((stbi__uint16 *) out) + channel;
-               stbi__uint16 val = channel == 3 ? 65535 : 0;
-               for (i = 0; i < pixelCount; i++, q += 4)
-                  *q = val;
-            } else {
-               stbi_uc *p = out+channel;
-               stbi_uc val = channel == 3 ? 255 : 0;
-               for (i = 0; i < pixelCount; i++, p += 4)
-                  *p = val;
-            }
-         } else {
-            if (ri->bits_per_channel == 16) {    // output bpc
-               stbi__uint16 *q = ((stbi__uint16 *) out) + channel;
-               for (i = 0; i < pixelCount; i++, q += 4)
-                  *q = (stbi__uint16) stbi__get16be(s);
-            } else {
-               stbi_uc *p = out+channel;
-               if (bitdepth == 16) {  // input bpc
-                  for (i = 0; i < pixelCount; i++, p += 4)
-                     *p = (stbi_uc) (stbi__get16be(s) >> 8);
-               } else {
-                  for (i = 0; i < pixelCount; i++, p += 4)
-                     *p = stbi__get8(s);
-               }
-            }
-         }
-      }
-   }
-
-   // remove weird white matte from PSD
-   if (channelCount >= 4) {
-      if (ri->bits_per_channel == 16) {
-         for (i=0; i < w*h; ++i) {
-            stbi__uint16 *pixel = (stbi__uint16 *) out + 4*i;
-            if (pixel[3] != 0 && pixel[3] != 65535) {
-               float a = pixel[3] / 65535.0f;
-               float ra = 1.0f / a;
-               float inv_a = 65535.0f * (1 - ra);
-               pixel[0] = (stbi__uint16) (pixel[0]*ra + inv_a);
-               pixel[1] = (stbi__uint16) (pixel[1]*ra + inv_a);
-               pixel[2] = (stbi__uint16) (pixel[2]*ra + inv_a);
-            }
-         }
-      } else {
-         for (i=0; i < w*h; ++i) {
-            unsigned char *pixel = out + 4*i;
-            if (pixel[3] != 0 && pixel[3] != 255) {
-               float a = pixel[3] / 255.0f;
-               float ra = 1.0f / a;
-               float inv_a = 255.0f * (1 - ra);
-               pixel[0] = (unsigned char) (pixel[0]*ra + inv_a);
-               pixel[1] = (unsigned char) (pixel[1]*ra + inv_a);
-               pixel[2] = (unsigned char) (pixel[2]*ra + inv_a);
-            }
-         }
-      }
-   }
-
-   // convert to desired output format
-   if (req_comp && req_comp != 4) {
-      if (ri->bits_per_channel == 16)
-         out = (stbi_uc *) stbi__convert_format16((stbi__uint16 *) out, 4, req_comp, w, h);
-      else
-         out = stbi__convert_format(out, 4, req_comp, w, h);
-      if (out == NULL) return out; // stbi__convert_format frees input on failure
-   }
-
-   if (comp) *comp = 4;
-   *y = h;
-   *x = w;
-
-   return out;
-}
-#endif
-
-// *************************************************************************************************
-// Softimage PIC loader
-// by Tom Seddon
-//
-// See http://softimage.wiki.softimage.com/index.php/INFO:_PIC_file_format
-// See http://ozviz.wasp.uwa.edu.au/~pbourke/dataformats/softimagepic/
-
-#ifndef STBI_NO_PIC
-static int stbi__pic_is4(stbi__context *s,const char *str)
-{
-   int i;
-   for (i=0; i<4; ++i)
-      if (stbi__get8(s) != (stbi_uc)str[i])
-         return 0;
-
-   return 1;
-}
-
-static int stbi__pic_test_core(stbi__context *s)
-{
-   int i;
-
-   if (!stbi__pic_is4(s,"\x53\x80\xF6\x34"))
-      return 0;
-
-   for(i=0;i<84;++i)
-      stbi__get8(s);
-
-   if (!stbi__pic_is4(s,"PICT"))
-      return 0;
-
-   return 1;
-}
-
-typedef struct
-{
-   stbi_uc size,type,channel;
-} stbi__pic_packet;
-
-static stbi_uc *stbi__readval(stbi__context *s, int channel, stbi_uc *dest)
-{
-   int mask=0x80, i;
-
-   for (i=0; i<4; ++i, mask>>=1) {
-      if (channel & mask) {
-         if (stbi__at_eof(s)) return stbi__errpuc("bad file","PIC file too short");
-         dest[i]=stbi__get8(s);
-      }
-   }
-
-   return dest;
-}
-
-static void stbi__copyval(int channel,stbi_uc *dest,const stbi_uc *src)
-{
-   int mask=0x80,i;
-
-   for (i=0;i<4; ++i, mask>>=1)
-      if (channel&mask)
-         dest[i]=src[i];
-}
-
-static stbi_uc *stbi__pic_load_core(stbi__context *s,int width,int height,int *comp, stbi_uc *result)
-{
-   int act_comp=0,num_packets=0,y,chained;
-   stbi__pic_packet packets[10];
-
-   // this will (should...) cater for even some bizarre stuff like having data
-    // for the same channel in multiple packets.
-   do {
-      stbi__pic_packet *packet;
-
-      if (num_packets==sizeof(packets)/sizeof(packets[0]))
-         return stbi__errpuc("bad format","too many packets");
-
-      packet = &packets[num_packets++];
-
-      chained = stbi__get8(s);
-      packet->size    = stbi__get8(s);
-      packet->type    = stbi__get8(s);
-      packet->channel = stbi__get8(s);
-
-      act_comp |= packet->channel;
-
-      if (stbi__at_eof(s))          return stbi__errpuc("bad file","file too short (reading packets)");
-      if (packet->size != 8)  return stbi__errpuc("bad format","packet isn't 8bpp");
-   } while (chained);
-
-   *comp = (act_comp & 0x10 ? 4 : 3); // has alpha channel?
-
-   for(y=0; y<height; ++y) {
-      int packet_idx;
-
-      for(packet_idx=0; packet_idx < num_packets; ++packet_idx) {
-         stbi__pic_packet *packet = &packets[packet_idx];
-         stbi_uc *dest = result+y*width*4;
-
-         switch (packet->type) {
-            default:
-               return stbi__errpuc("bad format","packet has bad compression type");
-
-            case 0: {//uncompressed
-               int x;
-
-               for(x=0;x<width;++x, dest+=4)
-                  if (!stbi__readval(s,packet->channel,dest))
-                     return 0;
-               break;
-            }
-
-            case 1://Pure RLE
-               {
-                  int left=width, i;
-
-                  while (left>0) {
-                     stbi_uc count,value[4];
-
-                     count=stbi__get8(s);
-                     if (stbi__at_eof(s))   return stbi__errpuc("bad file","file too short (pure read count)");
-
-                     if (count > left)
-                        count = (stbi_uc) left;
-
-                     if (!stbi__readval(s,packet->channel,value))  return 0;
-
-                     for(i=0; i<count; ++i,dest+=4)
-                        stbi__copyval(packet->channel,dest,value);
-                     left -= count;
-                  }
-               }
-               break;
-
-            case 2: {//Mixed RLE
-               int left=width;
-               while (left>0) {
-                  int count = stbi__get8(s), i;
-                  if (stbi__at_eof(s))  return stbi__errpuc("bad file","file too short (mixed read count)");
-
-                  if (count >= 128) { // Repeated
-                     stbi_uc value[4];
-
-                     if (count==128)
-                        count = stbi__get16be(s);
-                     else
-                        count -= 127;
-                     if (count > left)
-                        return stbi__errpuc("bad file","scanline overrun");
-
-                     if (!stbi__readval(s,packet->channel,value))
-                        return 0;
-
-                     for(i=0;i<count;++i, dest += 4)
-                        stbi__copyval(packet->channel,dest,value);
-                  } else { // Raw
-                     ++count;
-                     if (count>left) return stbi__errpuc("bad file","scanline overrun");
-
-                     for(i=0;i<count;++i, dest+=4)
-                        if (!stbi__readval(s,packet->channel,dest))
-                           return 0;
-                  }
-                  left-=count;
-               }
-               break;
-            }
-         }
-      }
-   }
-
-   return result;
-}
-
-static void *stbi__pic_load(stbi__context *s,int *px,int *py,int *comp,int req_comp, stbi__result_info *ri)
-{
-   stbi_uc *result;
-   int i, x,y, internal_comp;
-   STBI_NOTUSED(ri);
-
-   if (!comp) comp = &internal_comp;
-
-   for (i=0; i<92; ++i)
-      stbi__get8(s);
-
-   x = stbi__get16be(s);
-   y = stbi__get16be(s);
-
-   if (y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
-   if (x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
-
-   if (stbi__at_eof(s))  return stbi__errpuc("bad file","file too short (pic header)");
-   if (!stbi__mad3sizes_valid(x, y, 4, 0)) return stbi__errpuc("too large", "PIC image too large to decode");
-
-   stbi__get32be(s); //skip `ratio'
-   stbi__get16be(s); //skip `fields'
-   stbi__get16be(s); //skip `pad'
-
-   // intermediate buffer is RGBA
-   result = (stbi_uc *) stbi__malloc_mad3(x, y, 4, 0);
-   if (!result) return stbi__errpuc("outofmem", "Out of memory");
-   memset(result, 0xff, x*y*4);
-
-   if (!stbi__pic_load_core(s,x,y,comp, result)) {
-      STBI_FREE(result);
-      result=0;
-   }
-   *px = x;
-   *py = y;
-   if (req_comp == 0) req_comp = *comp;
-   result=stbi__convert_format(result,4,req_comp,x,y);
-
-   return result;
-}
-
-static int stbi__pic_test(stbi__context *s)
-{
-   int r = stbi__pic_test_core(s);
-   stbi__rewind(s);
-   return r;
-}
-#endif
-
-// *************************************************************************************************
-// GIF loader -- public domain by Jean-Marc Lienher -- simplified/shrunk by stb
-
-#ifndef STBI_NO_GIF
-typedef struct
-{
-   stbi__int16 prefix;
-   stbi_uc first;
-   stbi_uc suffix;
-} stbi__gif_lzw;
-
-typedef struct
-{
-   int w,h;
-   stbi_uc *out;                 // output buffer (always 4 components)
-   stbi_uc *background;          // The current "background" as far as a gif is concerned
-   stbi_uc *history;
-   int flags, bgindex, ratio, transparent, eflags;
-   stbi_uc  pal[256][4];
-   stbi_uc lpal[256][4];
-   stbi__gif_lzw codes[8192];
-   stbi_uc *color_table;
-   int parse, step;
-   int lflags;
-   int start_x, start_y;
-   int max_x, max_y;
-   int cur_x, cur_y;
-   int line_size;
-   int delay;
-} stbi__gif;
-
-static int stbi__gif_test_raw(stbi__context *s)
-{
-   int sz;
-   if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8') return 0;
-   sz = stbi__get8(s);
-   if (sz != '9' && sz != '7') return 0;
-   if (stbi__get8(s) != 'a') return 0;
-   return 1;
-}
-
-static int stbi__gif_test(stbi__context *s)
-{
-   int r = stbi__gif_test_raw(s);
-   stbi__rewind(s);
-   return r;
-}
-
-static void stbi__gif_parse_colortable(stbi__context *s, stbi_uc pal[256][4], int num_entries, int transp)
-{
-   int i;
-   for (i=0; i < num_entries; ++i) {
-      pal[i][2] = stbi__get8(s);
-      pal[i][1] = stbi__get8(s);
-      pal[i][0] = stbi__get8(s);
-      pal[i][3] = transp == i ? 0 : 255;
-   }
-}
-
-static int stbi__gif_header(stbi__context *s, stbi__gif *g, int *comp, int is_info)
-{
-   stbi_uc version;
-   if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8')
-      return stbi__err("not GIF", "Corrupt GIF");
-
-   version = stbi__get8(s);
-   if (version != '7' && version != '9')    return stbi__err("not GIF", "Corrupt GIF");
-   if (stbi__get8(s) != 'a')                return stbi__err("not GIF", "Corrupt GIF");
-
-   stbi__g_failure_reason = "";
-   g->w = stbi__get16le(s);
-   g->h = stbi__get16le(s);
-   g->flags = stbi__get8(s);
-   g->bgindex = stbi__get8(s);
-   g->ratio = stbi__get8(s);
-   g->transparent = -1;
-
-   if (g->w > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
-   if (g->h > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
-
-   if (comp != 0) *comp = 4;  // can't actually tell whether it's 3 or 4 until we parse the comments
-
-   if (is_info) return 1;
-
-   if (g->flags & 0x80)
-      stbi__gif_parse_colortable(s,g->pal, 2 << (g->flags & 7), -1);
-
-   return 1;
-}
-
-static int stbi__gif_info_raw(stbi__context *s, int *x, int *y, int *comp)
-{
-   stbi__gif* g = (stbi__gif*) stbi__malloc(sizeof(stbi__gif));
-   if (!g) return stbi__err("outofmem", "Out of memory");
-   if (!stbi__gif_header(s, g, comp, 1)) {
-      STBI_FREE(g);
-      stbi__rewind( s );
-      return 0;
-   }
-   if (x) *x = g->w;
-   if (y) *y = g->h;
-   STBI_FREE(g);
-   return 1;
-}
-
-static void stbi__out_gif_code(stbi__gif *g, stbi__uint16 code)
-{
-   stbi_uc *p, *c;
-   int idx;
-
-   // recurse to decode the prefixes, since the linked-list is backwards,
-   // and working backwards through an interleaved image would be nasty
-   if (g->codes[code].prefix >= 0)
-      stbi__out_gif_code(g, g->codes[code].prefix);
-
-   if (g->cur_y >= g->max_y) return;
-
-   idx = g->cur_x + g->cur_y;
-   p = &g->out[idx];
-   g->history[idx / 4] = 1;
-
-   c = &g->color_table[g->codes[code].suffix * 4];
-   if (c[3] > 128) { // don't render transparent pixels;
-      p[0] = c[2];
-      p[1] = c[1];
-      p[2] = c[0];
-      p[3] = c[3];
-   }
-   g->cur_x += 4;
-
-   if (g->cur_x >= g->max_x) {
-      g->cur_x = g->start_x;
-      g->cur_y += g->step;
-
-      while (g->cur_y >= g->max_y && g->parse > 0) {
-         g->step = (1 << g->parse) * g->line_size;
-         g->cur_y = g->start_y + (g->step >> 1);
-         --g->parse;
-      }
-   }
-}
-
-static stbi_uc *stbi__process_gif_raster(stbi__context *s, stbi__gif *g)
-{
-   stbi_uc lzw_cs;
-   stbi__int32 len, init_code;
-   stbi__uint32 first;
-   stbi__int32 codesize, codemask, avail, oldcode, bits, valid_bits, clear;
-   stbi__gif_lzw *p;
-
-   lzw_cs = stbi__get8(s);
-   if (lzw_cs > 12) return NULL;
-   clear = 1 << lzw_cs;
-   first = 1;
-   codesize = lzw_cs + 1;
-   codemask = (1 << codesize) - 1;
-   bits = 0;
-   valid_bits = 0;
-   for (init_code = 0; init_code < clear; init_code++) {
-      g->codes[init_code].prefix = -1;
-      g->codes[init_code].first = (stbi_uc) init_code;
-      g->codes[init_code].suffix = (stbi_uc) init_code;
-   }
-
-   // support no starting clear code
-   avail = clear+2;
-   oldcode = -1;
-
-   len = 0;
-   for(;;) {
-      if (valid_bits < codesize) {
-         if (len == 0) {
-            len = stbi__get8(s); // start new block
-            if (len == 0)
-               return g->out;
-         }
-         --len;
-         bits |= (stbi__int32) stbi__get8(s) << valid_bits;
-         valid_bits += 8;
-      } else {
-         stbi__int32 code = bits & codemask;
-         bits >>= codesize;
-         valid_bits -= codesize;
-         // @OPTIMIZE: is there some way we can accelerate the non-clear path?
-         if (code == clear) {  // clear code
-            codesize = lzw_cs + 1;
-            codemask = (1 << codesize) - 1;
-            avail = clear + 2;
-            oldcode = -1;
-            first = 0;
-         } else if (code == clear + 1) { // end of stream code
-            stbi__skip(s, len);
-            while ((len = stbi__get8(s)) > 0)
-               stbi__skip(s,len);
-            return g->out;
-         } else if (code <= avail) {
-            if (first) {
-               return stbi__errpuc("no clear code", "Corrupt GIF");
-            }
-
-            if (oldcode >= 0) {
-               p = &g->codes[avail++];
-               if (avail > 8192) {
-                  return stbi__errpuc("too many codes", "Corrupt GIF");
-               }
-
-               p->prefix = (stbi__int16) oldcode;
-               p->first = g->codes[oldcode].first;
-               p->suffix = (code == avail) ? p->first : g->codes[code].first;
-            } else if (code == avail)
-               return stbi__errpuc("illegal code in raster", "Corrupt GIF");
-
-            stbi__out_gif_code(g, (stbi__uint16) code);
-
-            if ((avail & codemask) == 0 && avail <= 0x0FFF) {
-               codesize++;
-               codemask = (1 << codesize) - 1;
-            }
-
-            oldcode = code;
-         } else {
-            return stbi__errpuc("illegal code in raster", "Corrupt GIF");
-         }
-      }
-   }
-}
-
-// this function is designed to support animated gifs, although stb_image doesn't support it
-// two back is the image from two frames ago, used for a very specific disposal format
-static stbi_uc *stbi__gif_load_next(stbi__context *s, stbi__gif *g, int *comp, int req_comp, stbi_uc *two_back)
-{
-   int dispose;
-   int first_frame;
-   int pi;
-   int pcount;
-   STBI_NOTUSED(req_comp);
-
-   // on first frame, any non-written pixels get the background colour (non-transparent)
-   first_frame = 0;
-   if (g->out == 0) {
-      if (!stbi__gif_header(s, g, comp,0)) return 0; // stbi__g_failure_reason set by stbi__gif_header
-      if (!stbi__mad3sizes_valid(4, g->w, g->h, 0))
-         return stbi__errpuc("too large", "GIF image is too large");
-      pcount = g->w * g->h;
-      g->out = (stbi_uc *) stbi__malloc(4 * pcount);
-      g->background = (stbi_uc *) stbi__malloc(4 * pcount);
-      g->history = (stbi_uc *) stbi__malloc(pcount);
-      if (!g->out || !g->background || !g->history)
-         return stbi__errpuc("outofmem", "Out of memory");
-
-      // image is treated as "transparent" at the start - ie, nothing overwrites the current background;
-      // background colour is only used for pixels that are not rendered first frame, after that "background"
-      // color refers to the color that was there the previous frame.
-      memset(g->out, 0x00, 4 * pcount);
-      memset(g->background, 0x00, 4 * pcount); // state of the background (starts transparent)
-      memset(g->history, 0x00, pcount);        // pixels that were affected previous frame
-      first_frame = 1;
-   } else {
-      // second frame - how do we dispose of the previous one?
-      dispose = (g->eflags & 0x1C) >> 2;
-      pcount = g->w * g->h;
-
-      if ((dispose == 3) && (two_back == 0)) {
-         dispose = 2; // if I don't have an image to revert back to, default to the old background
-      }
-
-      if (dispose == 3) { // use previous graphic
-         for (pi = 0; pi < pcount; ++pi) {
-            if (g->history[pi]) {
-               memcpy( &g->out[pi * 4], &two_back[pi * 4], 4 );
-            }
-         }
-      } else if (dispose == 2) {
-         // restore what was changed last frame to background before that frame;
-         for (pi = 0; pi < pcount; ++pi) {
-            if (g->history[pi]) {
-               memcpy( &g->out[pi * 4], &g->background[pi * 4], 4 );
-            }
-         }
-      } else {
-         // This is a non-disposal case eithe way, so just
-         // leave the pixels as is, and they will become the new background
-         // 1: do not dispose
-         // 0:  not specified.
-      }
-
-      // background is what out is after the undoing of the previou frame;
-      memcpy( g->background, g->out, 4 * g->w * g->h );
-   }
-
-   // clear my history;
-   memset( g->history, 0x00, g->w * g->h );        // pixels that were affected previous frame
-
-   for (;;) {
-      int tag = stbi__get8(s);
-      switch (tag) {
-         case 0x2C: /* Image Descriptor */
-         {
-            stbi__int32 x, y, w, h;
-            stbi_uc *o;
-
-            x = stbi__get16le(s);
-            y = stbi__get16le(s);
-            w = stbi__get16le(s);
-            h = stbi__get16le(s);
-            if (((x + w) > (g->w)) || ((y + h) > (g->h)))
-               return stbi__errpuc("bad Image Descriptor", "Corrupt GIF");
-
-            g->line_size = g->w * 4;
-            g->start_x = x * 4;
-            g->start_y = y * g->line_size;
-            g->max_x   = g->start_x + w * 4;
-            g->max_y   = g->start_y + h * g->line_size;
-            g->cur_x   = g->start_x;
-            g->cur_y   = g->start_y;
-
-            // if the width of the specified rectangle is 0, that means
-            // we may not see *any* pixels or the image is malformed;
-            // to make sure this is caught, move the current y down to
-            // max_y (which is what out_gif_code checks).
-            if (w == 0)
-               g->cur_y = g->max_y;
-
-            g->lflags = stbi__get8(s);
-
-            if (g->lflags & 0x40) {
-               g->step = 8 * g->line_size; // first interlaced spacing
-               g->parse = 3;
-            } else {
-               g->step = g->line_size;
-               g->parse = 0;
-            }
-
-            if (g->lflags & 0x80) {
-               stbi__gif_parse_colortable(s,g->lpal, 2 << (g->lflags & 7), g->eflags & 0x01 ? g->transparent : -1);
-               g->color_table = (stbi_uc *) g->lpal;
-            } else if (g->flags & 0x80) {
-               g->color_table = (stbi_uc *) g->pal;
-            } else
-               return stbi__errpuc("missing color table", "Corrupt GIF");
-
-            o = stbi__process_gif_raster(s, g);
-            if (!o) return NULL;
-
-            // if this was the first frame,
-            pcount = g->w * g->h;
-            if (first_frame && (g->bgindex > 0)) {
-               // if first frame, any pixel not drawn to gets the background color
-               for (pi = 0; pi < pcount; ++pi) {
-                  if (g->history[pi] == 0) {
-                     g->pal[g->bgindex][3] = 255; // just in case it was made transparent, undo that; It will be reset next frame if need be;
-                     memcpy( &g->out[pi * 4], &g->pal[g->bgindex], 4 );
-                  }
-               }
-            }
-
-            return o;
-         }
-
-         case 0x21: // Comment Extension.
-         {
-            int len;
-            int ext = stbi__get8(s);
-            if (ext == 0xF9) { // Graphic Control Extension.
-               len = stbi__get8(s);
-               if (len == 4) {
-                  g->eflags = stbi__get8(s);
-                  g->delay = 10 * stbi__get16le(s); // delay - 1/100th of a second, saving as 1/1000ths.
-
-                  // unset old transparent
-                  if (g->transparent >= 0) {
-                     g->pal[g->transparent][3] = 255;
-                  }
-                  if (g->eflags & 0x01) {
-                     g->transparent = stbi__get8(s);
-                     if (g->transparent >= 0) {
-                        g->pal[g->transparent][3] = 0;
-                     }
-                  } else {
-                     // don't need transparent
-                     stbi__skip(s, 1);
-                     g->transparent = -1;
-                  }
-               } else {
-                  stbi__skip(s, len);
-                  break;
-               }
-            }
-            while ((len = stbi__get8(s)) != 0) {
-               stbi__skip(s, len);
-            }
-            break;
-         }
-
-         case 0x3B: // gif stream termination code
-            return (stbi_uc *) s; // using '1' causes warning on some compilers
-
-         default:
-            return stbi__errpuc("unknown code", "Corrupt GIF");
-      }
-   }
-}
-
-static void *stbi__load_gif_main_outofmem(stbi__gif *g, stbi_uc *out, int **delays)
-{
-   STBI_FREE(g->out);
-   STBI_FREE(g->history);
-   STBI_FREE(g->background);
-
-   if (out) STBI_FREE(out);
-   if (delays && *delays) STBI_FREE(*delays);
-   return stbi__errpuc("outofmem", "Out of memory");
-}
-
-static void *stbi__load_gif_main(stbi__context *s, int **delays, int *x, int *y, int *z, int *comp, int req_comp)
-{
-   if (stbi__gif_test(s)) {
-      int layers = 0;
-      stbi_uc *u = 0;
-      stbi_uc *out = 0;
-      stbi_uc *two_back = 0;
-      stbi__gif g;
-      int stride;
-      int out_size = 0;
-      int delays_size = 0;
-
-      STBI_NOTUSED(out_size);
-      STBI_NOTUSED(delays_size);
-
-      memset(&g, 0, sizeof(g));
-      if (delays) {
-         *delays = 0;
-      }
-
-      do {
-         u = stbi__gif_load_next(s, &g, comp, req_comp, two_back);
-         if (u == (stbi_uc *) s) u = 0;  // end of animated gif marker
-
-         if (u) {
-            *x = g.w;
-            *y = g.h;
-            ++layers;
-            stride = g.w * g.h * 4;
-
-            if (out) {
-               void *tmp = (stbi_uc*) STBI_REALLOC_SIZED( out, out_size, layers * stride );
-               if (!tmp)
-                  return stbi__load_gif_main_outofmem(&g, out, delays);
-               else {
-                   out = (stbi_uc*) tmp;
-                   out_size = layers * stride;
-               }
-
-               if (delays) {
-                  int *new_delays = (int*) STBI_REALLOC_SIZED( *delays, delays_size, sizeof(int) * layers );
-                  if (!new_delays)
-                     return stbi__load_gif_main_outofmem(&g, out, delays);
-                  *delays = new_delays;
-                  delays_size = layers * sizeof(int);
-               }
-            } else {
-               out = (stbi_uc*)stbi__malloc( layers * stride );
-               if (!out)
-                  return stbi__load_gif_main_outofmem(&g, out, delays);
-               out_size = layers * stride;
-               if (delays) {
-                  *delays = (int*) stbi__malloc( layers * sizeof(int) );
-                  if (!*delays)
-                     return stbi__load_gif_main_outofmem(&g, out, delays);
-                  delays_size = layers * sizeof(int);
-               }
-            }
-            memcpy( out + ((layers - 1) * stride), u, stride );
-            if (layers >= 2) {
-               two_back = out - 2 * stride;
-            }
-
-            if (delays) {
-               (*delays)[layers - 1U] = g.delay;
-            }
-         }
-      } while (u != 0);
-
-      // free temp buffer;
-      STBI_FREE(g.out);
-      STBI_FREE(g.history);
-      STBI_FREE(g.background);
-
-      // do the final conversion after loading everything;
-      if (req_comp && req_comp != 4)
-         out = stbi__convert_format(out, 4, req_comp, layers * g.w, g.h);
-
-      *z = layers;
-      return out;
-   } else {
-      return stbi__errpuc("not GIF", "Image was not as a gif type.");
-   }
-}
-
-static void *stbi__gif_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
-{
-   stbi_uc *u = 0;
-   stbi__gif g;
-   memset(&g, 0, sizeof(g));
-   STBI_NOTUSED(ri);
-
-   u = stbi__gif_load_next(s, &g, comp, req_comp, 0);
-   if (u == (stbi_uc *) s) u = 0;  // end of animated gif marker
-   if (u) {
-      *x = g.w;
-      *y = g.h;
-
-      // moved conversion to after successful load so that the same
-      // can be done for multiple frames.
-      if (req_comp && req_comp != 4)
-         u = stbi__convert_format(u, 4, req_comp, g.w, g.h);
-   } else if (g.out) {
-      // if there was an error and we allocated an image buffer, free it!
-      STBI_FREE(g.out);
-   }
-
-   // free buffers needed for multiple frame loading;
-   STBI_FREE(g.history);
-   STBI_FREE(g.background);
-
-   return u;
-}
-
-static int stbi__gif_info(stbi__context *s, int *x, int *y, int *comp)
-{
-   return stbi__gif_info_raw(s,x,y,comp);
-}
-#endif
-
-// *************************************************************************************************
-// Radiance RGBE HDR loader
-// originally by Nicolas Schulz
-#ifndef STBI_NO_HDR
-static int stbi__hdr_test_core(stbi__context *s, const char *signature)
-{
-   int i;
-   for (i=0; signature[i]; ++i)
-      if (stbi__get8(s) != signature[i])
-          return 0;
-   stbi__rewind(s);
-   return 1;
-}
-
-static int stbi__hdr_test(stbi__context* s)
-{
-   int r = stbi__hdr_test_core(s, "#?RADIANCE\n");
-   stbi__rewind(s);
-   if(!r) {
-       r = stbi__hdr_test_core(s, "#?RGBE\n");
-       stbi__rewind(s);
-   }
-   return r;
-}
-
-#define STBI__HDR_BUFLEN  1024
-static char *stbi__hdr_gettoken(stbi__context *z, char *buffer)
-{
-   int len=0;
-   char c = '\0';
-
-   c = (char) stbi__get8(z);
-
-   while (!stbi__at_eof(z) && c != '\n') {
-      buffer[len++] = c;
-      if (len == STBI__HDR_BUFLEN-1) {
-         // flush to end of line
-         while (!stbi__at_eof(z) && stbi__get8(z) != '\n')
-            ;
-         break;
-      }
-      c = (char) stbi__get8(z);
-   }
-
-   buffer[len] = 0;
-   return buffer;
-}
-
-static void stbi__hdr_convert(float *output, stbi_uc *input, int req_comp)
-{
-   if ( input[3] != 0 ) {
-      float f1;
-      // Exponent
-      f1 = (float) ldexp(1.0f, input[3] - (int)(128 + 8));
-      if (req_comp <= 2)
-         output[0] = (input[0] + input[1] + input[2]) * f1 / 3;
-      else {
-         output[0] = input[0] * f1;
-         output[1] = input[1] * f1;
-         output[2] = input[2] * f1;
-      }
-      if (req_comp == 2) output[1] = 1;
-      if (req_comp == 4) output[3] = 1;
-   } else {
-      switch (req_comp) {
-         case 4: output[3] = 1; /* fallthrough */
-         case 3: output[0] = output[1] = output[2] = 0;
-                 break;
-         case 2: output[1] = 1; /* fallthrough */
-         case 1: output[0] = 0;
-                 break;
-      }
-   }
-}
-
-static float *stbi__hdr_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
-{
-   char buffer[STBI__HDR_BUFLEN];
-   char *token;
-   int valid = 0;
-   int width, height;
-   stbi_uc *scanline;
-   float *hdr_data;
-   int len;
-   unsigned char count, value;
-   int i, j, k, c1,c2, z;
-   const char *headerToken;
-   STBI_NOTUSED(ri);
-
-   // Check identifier
-   headerToken = stbi__hdr_gettoken(s,buffer);
-   if (strcmp(headerToken, "#?RADIANCE") != 0 && strcmp(headerToken, "#?RGBE") != 0)
-      return stbi__errpf("not HDR", "Corrupt HDR image");
-
-   // Parse header
-   for(;;) {
-      token = stbi__hdr_gettoken(s,buffer);
-      if (token[0] == 0) break;
-      if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) valid = 1;
-   }
-
-   if (!valid)    return stbi__errpf("unsupported format", "Unsupported HDR format");
-
-   // Parse width and height
-   // can't use sscanf() if we're not using stdio!
-   token = stbi__hdr_gettoken(s,buffer);
-   if (strncmp(token, "-Y ", 3))  return stbi__errpf("unsupported data layout", "Unsupported HDR format");
-   token += 3;
-   height = (int) strtol(token, &token, 10);
-   while (*token == ' ') ++token;
-   if (strncmp(token, "+X ", 3))  return stbi__errpf("unsupported data layout", "Unsupported HDR format");
-   token += 3;
-   width = (int) strtol(token, NULL, 10);
-
-   if (height > STBI_MAX_DIMENSIONS) return stbi__errpf("too large","Very large image (corrupt?)");
-   if (width > STBI_MAX_DIMENSIONS) return stbi__errpf("too large","Very large image (corrupt?)");
-
-   *x = width;
-   *y = height;
-
-   if (comp) *comp = 3;
-   if (req_comp == 0) req_comp = 3;
-
-   if (!stbi__mad4sizes_valid(width, height, req_comp, sizeof(float), 0))
-      return stbi__errpf("too large", "HDR image is too large");
-
-   // Read data
-   hdr_data = (float *) stbi__malloc_mad4(width, height, req_comp, sizeof(float), 0);
-   if (!hdr_data)
-      return stbi__errpf("outofmem", "Out of memory");
-
-   // Load image data
-   // image data is stored as some number of sca
-   if ( width < 8 || width >= 32768) {
-      // Read flat data
-      for (j=0; j < height; ++j) {
-         for (i=0; i < width; ++i) {
-            stbi_uc rgbe[4];
-           main_decode_loop:
-            stbi__getn(s, rgbe, 4);
-            stbi__hdr_convert(hdr_data + j * width * req_comp + i * req_comp, rgbe, req_comp);
-         }
-      }
-   } else {
-      // Read RLE-encoded data
-      scanline = NULL;
-
-      for (j = 0; j < height; ++j) {
-         c1 = stbi__get8(s);
-         c2 = stbi__get8(s);
-         len = stbi__get8(s);
-         if (c1 != 2 || c2 != 2 || (len & 0x80)) {
-            // not run-length encoded, so we have to actually use THIS data as a decoded
-            // pixel (note this can't be a valid pixel--one of RGB must be >= 128)
-            stbi_uc rgbe[4];
-            rgbe[0] = (stbi_uc) c1;
-            rgbe[1] = (stbi_uc) c2;
-            rgbe[2] = (stbi_uc) len;
-            rgbe[3] = (stbi_uc) stbi__get8(s);
-            stbi__hdr_convert(hdr_data, rgbe, req_comp);
-            i = 1;
-            j = 0;
-            STBI_FREE(scanline);
-            goto main_decode_loop; // yes, this makes no sense
-         }
-         len <<= 8;
-         len |= stbi__get8(s);
-         if (len != width) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("invalid decoded scanline length", "corrupt HDR"); }
-         if (scanline == NULL) {
-            scanline = (stbi_uc *) stbi__malloc_mad2(width, 4, 0);
-            if (!scanline) {
-               STBI_FREE(hdr_data);
-               return stbi__errpf("outofmem", "Out of memory");
-            }
-         }
-
-         for (k = 0; k < 4; ++k) {
-            int nleft;
-            i = 0;
-            while ((nleft = width - i) > 0) {
-               count = stbi__get8(s);
-               if (count > 128) {
-                  // Run
-                  value = stbi__get8(s);
-                  count -= 128;
-                  if ((count == 0) || (count > nleft)) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("corrupt", "bad RLE data in HDR"); }
-                  for (z = 0; z < count; ++z)
-                     scanline[i++ * 4 + k] = value;
-               } else {
-                  // Dump
-                  if ((count == 0) || (count > nleft)) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("corrupt", "bad RLE data in HDR"); }
-                  for (z = 0; z < count; ++z)
-                     scanline[i++ * 4 + k] = stbi__get8(s);
-               }
-            }
-         }
-         for (i=0; i < width; ++i)
-            stbi__hdr_convert(hdr_data+(j*width + i)*req_comp, scanline + i*4, req_comp);
-      }
-      if (scanline)
-         STBI_FREE(scanline);
-   }
-
-   return hdr_data;
-}
-
-static int stbi__hdr_info(stbi__context *s, int *x, int *y, int *comp)
-{
-   char buffer[STBI__HDR_BUFLEN];
-   char *token;
-   int valid = 0;
-   int dummy;
-
-   if (!x) x = &dummy;
-   if (!y) y = &dummy;
-   if (!comp) comp = &dummy;
-
-   if (stbi__hdr_test(s) == 0) {
-       stbi__rewind( s );
-       return 0;
-   }
-
-   for(;;) {
-      token = stbi__hdr_gettoken(s,buffer);
-      if (token[0] == 0) break;
-      if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) valid = 1;
-   }
-
-   if (!valid) {
-       stbi__rewind( s );
-       return 0;
-   }
-   token = stbi__hdr_gettoken(s,buffer);
-   if (strncmp(token, "-Y ", 3)) {
-       stbi__rewind( s );
-       return 0;
-   }
-   token += 3;
-   *y = (int) strtol(token, &token, 10);
-   while (*token == ' ') ++token;
-   if (strncmp(token, "+X ", 3)) {
-       stbi__rewind( s );
-       return 0;
-   }
-   token += 3;
-   *x = (int) strtol(token, NULL, 10);
-   *comp = 3;
-   return 1;
-}
-#endif // STBI_NO_HDR
-
-#ifndef STBI_NO_BMP
-static int stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp)
-{
-   void *p;
-   stbi__bmp_data info;
-
-   info.all_a = 255;
-   p = stbi__bmp_parse_header(s, &info);
-   if (p == NULL) {
-      stbi__rewind( s );
-      return 0;
-   }
-   if (x) *x = s->img_x;
-   if (y) *y = s->img_y;
-   if (comp) {
-      if (info.bpp == 24 && info.ma == 0xff000000)
-         *comp = 3;
-      else
-         *comp = info.ma ? 4 : 3;
-   }
-   return 1;
-}
-#endif
-
-#ifndef STBI_NO_PSD
-static int stbi__psd_info(stbi__context *s, int *x, int *y, int *comp)
-{
-   int channelCount, dummy, depth;
-   if (!x) x = &dummy;
-   if (!y) y = &dummy;
-   if (!comp) comp = &dummy;
-   if (stbi__get32be(s) != 0x38425053) {
-       stbi__rewind( s );
-       return 0;
-   }
-   if (stbi__get16be(s) != 1) {
-       stbi__rewind( s );
-       return 0;
-   }
-   stbi__skip(s, 6);
-   channelCount = stbi__get16be(s);
-   if (channelCount < 0 || channelCount > 16) {
-       stbi__rewind( s );
-       return 0;
-   }
-   *y = stbi__get32be(s);
-   *x = stbi__get32be(s);
-   depth = stbi__get16be(s);
-   if (depth != 8 && depth != 16) {
-       stbi__rewind( s );
-       return 0;
-   }
-   if (stbi__get16be(s) != 3) {
-       stbi__rewind( s );
-       return 0;
-   }
-   *comp = 4;
-   return 1;
-}
-
-static int stbi__psd_is16(stbi__context *s)
-{
-   int channelCount, depth;
-   if (stbi__get32be(s) != 0x38425053) {
-       stbi__rewind( s );
-       return 0;
-   }
-   if (stbi__get16be(s) != 1) {
-       stbi__rewind( s );
-       return 0;
-   }
-   stbi__skip(s, 6);
-   channelCount = stbi__get16be(s);
-   if (channelCount < 0 || channelCount > 16) {
-       stbi__rewind( s );
-       return 0;
-   }
-   STBI_NOTUSED(stbi__get32be(s));
-   STBI_NOTUSED(stbi__get32be(s));
-   depth = stbi__get16be(s);
-   if (depth != 16) {
-       stbi__rewind( s );
-       return 0;
-   }
-   return 1;
-}
-#endif
-
-#ifndef STBI_NO_PIC
-static int stbi__pic_info(stbi__context *s, int *x, int *y, int *comp)
-{
-   int act_comp=0,num_packets=0,chained,dummy;
-   stbi__pic_packet packets[10];
-
-   if (!x) x = &dummy;
-   if (!y) y = &dummy;
-   if (!comp) comp = &dummy;
-
-   if (!stbi__pic_is4(s,"\x53\x80\xF6\x34")) {
-      stbi__rewind(s);
-      return 0;
-   }
-
-   stbi__skip(s, 88);
-
-   *x = stbi__get16be(s);
-   *y = stbi__get16be(s);
-   if (stbi__at_eof(s)) {
-      stbi__rewind( s);
-      return 0;
-   }
-   if ( (*x) != 0 && (1 << 28) / (*x) < (*y)) {
-      stbi__rewind( s );
-      return 0;
-   }
-
-   stbi__skip(s, 8);
-
-   do {
-      stbi__pic_packet *packet;
-
-      if (num_packets==sizeof(packets)/sizeof(packets[0]))
-         return 0;
-
-      packet = &packets[num_packets++];
-      chained = stbi__get8(s);
-      packet->size    = stbi__get8(s);
-      packet->type    = stbi__get8(s);
-      packet->channel = stbi__get8(s);
-      act_comp |= packet->channel;
-
-      if (stbi__at_eof(s)) {
-          stbi__rewind( s );
-          return 0;
-      }
-      if (packet->size != 8) {
-          stbi__rewind( s );
-          return 0;
-      }
-   } while (chained);
-
-   *comp = (act_comp & 0x10 ? 4 : 3);
-
-   return 1;
-}
-#endif
-
-// *************************************************************************************************
-// Portable Gray Map and Portable Pixel Map loader
-// by Ken Miller
-//
-// PGM: http://netpbm.sourceforge.net/doc/pgm.html
-// PPM: http://netpbm.sourceforge.net/doc/ppm.html
-//
-// Known limitations:
-//    Does not support comments in the header section
-//    Does not support ASCII image data (formats P2 and P3)
-
-#ifndef STBI_NO_PNM
-
-static int      stbi__pnm_test(stbi__context *s)
-{
-   char p, t;
-   p = (char) stbi__get8(s);
-   t = (char) stbi__get8(s);
-   if (p != 'P' || (t != '5' && t != '6')) {
-       stbi__rewind( s );
-       return 0;
-   }
-   return 1;
-}
-
-static void *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
-{
-   stbi_uc *out;
-   STBI_NOTUSED(ri);
-
-   ri->bits_per_channel = stbi__pnm_info(s, (int *)&s->img_x, (int *)&s->img_y, (int *)&s->img_n);
-   if (ri->bits_per_channel == 0)
-      return 0;
-
-   if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
-   if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
-
-   *x = s->img_x;
-   *y = s->img_y;
-   if (comp) *comp = s->img_n;
-
-   if (!stbi__mad4sizes_valid(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0))
-      return stbi__errpuc("too large", "PNM too large");
-
-   out = (stbi_uc *) stbi__malloc_mad4(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0);
-   if (!out) return stbi__errpuc("outofmem", "Out of memory");
-   if (!stbi__getn(s, out, s->img_n * s->img_x * s->img_y * (ri->bits_per_channel / 8))) {
-      STBI_FREE(out);
-      return stbi__errpuc("bad PNM", "PNM file truncated");
-   }
-
-   if (req_comp && req_comp != s->img_n) {
-      if (ri->bits_per_channel == 16) {
-         out = (stbi_uc *) stbi__convert_format16((stbi__uint16 *) out, s->img_n, req_comp, s->img_x, s->img_y);
-      } else {
-         out = stbi__convert_format(out, s->img_n, req_comp, s->img_x, s->img_y);
-      }
-      if (out == NULL) return out; // stbi__convert_format frees input on failure
-   }
-   return out;
-}
-
-static int      stbi__pnm_isspace(char c)
-{
-   return c == ' ' || c == '\t' || c == '\n' || c == '\v' || c == '\f' || c == '\r';
-}
-
-static void     stbi__pnm_skip_whitespace(stbi__context *s, char *c)
-{
-   for (;;) {
-      while (!stbi__at_eof(s) && stbi__pnm_isspace(*c))
-         *c = (char) stbi__get8(s);
-
-      if (stbi__at_eof(s) || *c != '#')
-         break;
-
-      while (!stbi__at_eof(s) && *c != '\n' && *c != '\r' )
-         *c = (char) stbi__get8(s);
-   }
-}
-
-static int      stbi__pnm_isdigit(char c)
-{
-   return c >= '0' && c <= '9';
-}
-
-static int      stbi__pnm_getinteger(stbi__context *s, char *c)
-{
-   int value = 0;
-
-   while (!stbi__at_eof(s) && stbi__pnm_isdigit(*c)) {
-      value = value*10 + (*c - '0');
-      *c = (char) stbi__get8(s);
-      if((value > 214748364) || (value == 214748364 && *c > '7'))
-          return stbi__err("integer parse overflow", "Parsing an integer in the PPM header overflowed a 32-bit int");
-   }
-
-   return value;
-}
-
-static int      stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp)
-{
-   int maxv, dummy;
-   char c, p, t;
-
-   if (!x) x = &dummy;
-   if (!y) y = &dummy;
-   if (!comp) comp = &dummy;
-
-   stbi__rewind(s);
-
-   // Get identifier
-   p = (char) stbi__get8(s);
-   t = (char) stbi__get8(s);
-   if (p != 'P' || (t != '5' && t != '6')) {
-       stbi__rewind(s);
-       return 0;
-   }
-
-   *comp = (t == '6') ? 3 : 1;  // '5' is 1-component .pgm; '6' is 3-component .ppm
-
-   c = (char) stbi__get8(s);
-   stbi__pnm_skip_whitespace(s, &c);
-
-   *x = stbi__pnm_getinteger(s, &c); // read width
-   if(*x == 0)
-       return stbi__err("invalid width", "PPM image header had zero or overflowing width");
-   stbi__pnm_skip_whitespace(s, &c);
-
-   *y = stbi__pnm_getinteger(s, &c); // read height
-   if (*y == 0)
-       return stbi__err("invalid width", "PPM image header had zero or overflowing width");
-   stbi__pnm_skip_whitespace(s, &c);
-
-   maxv = stbi__pnm_getinteger(s, &c);  // read max value
-   if (maxv > 65535)
-      return stbi__err("max value > 65535", "PPM image supports only 8-bit and 16-bit images");
-   else if (maxv > 255)
-      return 16;
-   else
-      return 8;
-}
-
-static int stbi__pnm_is16(stbi__context *s)
-{
-   if (stbi__pnm_info(s, NULL, NULL, NULL) == 16)
-	   return 1;
-   return 0;
-}
-#endif
-
-static int stbi__info_main(stbi__context *s, int *x, int *y, int *comp)
-{
-   #ifndef STBI_NO_JPEG
-   if (stbi__jpeg_info(s, x, y, comp)) return 1;
-   #endif
-
-   #ifndef STBI_NO_PNG
-   if (stbi__png_info(s, x, y, comp))  return 1;
-   #endif
-
-   #ifndef STBI_NO_GIF
-   if (stbi__gif_info(s, x, y, comp))  return 1;
-   #endif
-
-   #ifndef STBI_NO_BMP
-   if (stbi__bmp_info(s, x, y, comp))  return 1;
-   #endif
-
-   #ifndef STBI_NO_PSD
-   if (stbi__psd_info(s, x, y, comp))  return 1;
-   #endif
-
-   #ifndef STBI_NO_PIC
-   if (stbi__pic_info(s, x, y, comp))  return 1;
-   #endif
-
-   #ifndef STBI_NO_PNM
-   if (stbi__pnm_info(s, x, y, comp))  return 1;
-   #endif
-
-   #ifndef STBI_NO_HDR
-   if (stbi__hdr_info(s, x, y, comp))  return 1;
-   #endif
-
-   // test tga last because it's a crappy test!
-   #ifndef STBI_NO_TGA
-   if (stbi__tga_info(s, x, y, comp))
-       return 1;
-   #endif
-   return stbi__err("unknown image type", "Image not of any known type, or corrupt");
-}
-
-static int stbi__is_16_main(stbi__context *s)
-{
-   #ifndef STBI_NO_PNG
-   if (stbi__png_is16(s))  return 1;
-   #endif
-
-   #ifndef STBI_NO_PSD
-   if (stbi__psd_is16(s))  return 1;
-   #endif
-
-   #ifndef STBI_NO_PNM
-   if (stbi__pnm_is16(s))  return 1;
-   #endif
-   return 0;
-}
-
-#ifndef STBI_NO_STDIO
-STBIDEF int stbi_info(char const *filename, int *x, int *y, int *comp)
-{
-    FILE *f = stbi__fopen(filename, "rb");
-    int result;
-    if (!f) return stbi__err("can't fopen", "Unable to open file");
-    result = stbi_info_from_file(f, x, y, comp);
-    fclose(f);
-    return result;
-}
-
-STBIDEF int stbi_info_from_file(FILE *f, int *x, int *y, int *comp)
-{
-   int r;
-   stbi__context s;
-   long pos = ftell(f);
-   stbi__start_file(&s, f);
-   r = stbi__info_main(&s,x,y,comp);
-   fseek(f,pos,SEEK_SET);
-   return r;
-}
-
-STBIDEF int stbi_is_16_bit(char const *filename)
-{
-    FILE *f = stbi__fopen(filename, "rb");
-    int result;
-    if (!f) return stbi__err("can't fopen", "Unable to open file");
-    result = stbi_is_16_bit_from_file(f);
-    fclose(f);
-    return result;
-}
-
-STBIDEF int stbi_is_16_bit_from_file(FILE *f)
-{
-   int r;
-   stbi__context s;
-   long pos = ftell(f);
-   stbi__start_file(&s, f);
-   r = stbi__is_16_main(&s);
-   fseek(f,pos,SEEK_SET);
-   return r;
-}
-#endif // !STBI_NO_STDIO
-
-STBIDEF int stbi_info_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp)
-{
-   stbi__context s;
-   stbi__start_mem(&s,buffer,len);
-   return stbi__info_main(&s,x,y,comp);
-}
-
-STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const *c, void *user, int *x, int *y, int *comp)
-{
-   stbi__context s;
-   stbi__start_callbacks(&s, (stbi_io_callbacks *) c, user);
-   return stbi__info_main(&s,x,y,comp);
-}
-
-STBIDEF int stbi_is_16_bit_from_memory(stbi_uc const *buffer, int len)
-{
-   stbi__context s;
-   stbi__start_mem(&s,buffer,len);
-   return stbi__is_16_main(&s);
-}
-
-STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const *c, void *user)
-{
-   stbi__context s;
-   stbi__start_callbacks(&s, (stbi_io_callbacks *) c, user);
-   return stbi__is_16_main(&s);
-}
-
-#endif // STB_IMAGE_IMPLEMENTATION
-
-/*
-   revision history:
-      2.20  (2019-02-07) support utf8 filenames in Windows; fix warnings and platform ifdefs
-      2.19  (2018-02-11) fix warning
-      2.18  (2018-01-30) fix warnings
-      2.17  (2018-01-29) change sbti__shiftsigned to avoid clang -O2 bug
-                         1-bit BMP
-                         *_is_16_bit api
-                         avoid warnings
-      2.16  (2017-07-23) all functions have 16-bit variants;
-                         STBI_NO_STDIO works again;
-                         compilation fixes;
-                         fix rounding in unpremultiply;
-                         optimize vertical flip;
-                         disable raw_len validation;
-                         documentation fixes
-      2.15  (2017-03-18) fix png-1,2,4 bug; now all Imagenet JPGs decode;
-                         warning fixes; disable run-time SSE detection on gcc;
-                         uniform handling of optional "return" values;
-                         thread-safe initialization of zlib tables
-      2.14  (2017-03-03) remove deprecated STBI_JPEG_OLD; fixes for Imagenet JPGs
-      2.13  (2016-11-29) add 16-bit API, only supported for PNG right now
-      2.12  (2016-04-02) fix typo in 2.11 PSD fix that caused crashes
-      2.11  (2016-04-02) allocate large structures on the stack
-                         remove white matting for transparent PSD
-                         fix reported channel count for PNG & BMP
-                         re-enable SSE2 in non-gcc 64-bit
-                         support RGB-formatted JPEG
-                         read 16-bit PNGs (only as 8-bit)
-      2.10  (2016-01-22) avoid warning introduced in 2.09 by STBI_REALLOC_SIZED
-      2.09  (2016-01-16) allow comments in PNM files
-                         16-bit-per-pixel TGA (not bit-per-component)
-                         info() for TGA could break due to .hdr handling
-                         info() for BMP to shares code instead of sloppy parse
-                         can use STBI_REALLOC_SIZED if allocator doesn't support realloc
-                         code cleanup
-      2.08  (2015-09-13) fix to 2.07 cleanup, reading RGB PSD as RGBA
-      2.07  (2015-09-13) fix compiler warnings
-                         partial animated GIF support
-                         limited 16-bpc PSD support
-                         #ifdef unused functions
-                         bug with < 92 byte PIC,PNM,HDR,TGA
-      2.06  (2015-04-19) fix bug where PSD returns wrong '*comp' value
-      2.05  (2015-04-19) fix bug in progressive JPEG handling, fix warning
-      2.04  (2015-04-15) try to re-enable SIMD on MinGW 64-bit
-      2.03  (2015-04-12) extra corruption checking (mmozeiko)
-                         stbi_set_flip_vertically_on_load (nguillemot)
-                         fix NEON support; fix mingw support
-      2.02  (2015-01-19) fix incorrect assert, fix warning
-      2.01  (2015-01-17) fix various warnings; suppress SIMD on gcc 32-bit without -msse2
-      2.00b (2014-12-25) fix STBI_MALLOC in progressive JPEG
-      2.00  (2014-12-25) optimize JPG, including x86 SSE2 & NEON SIMD (ryg)
-                         progressive JPEG (stb)
-                         PGM/PPM support (Ken Miller)
-                         STBI_MALLOC,STBI_REALLOC,STBI_FREE
-                         GIF bugfix -- seemingly never worked
-                         STBI_NO_*, STBI_ONLY_*
-      1.48  (2014-12-14) fix incorrectly-named assert()
-      1.47  (2014-12-14) 1/2/4-bit PNG support, both direct and paletted (Omar Cornut & stb)
-                         optimize PNG (ryg)
-                         fix bug in interlaced PNG with user-specified channel count (stb)
-      1.46  (2014-08-26)
-              fix broken tRNS chunk (colorkey-style transparency) in non-paletted PNG
-      1.45  (2014-08-16)
-              fix MSVC-ARM internal compiler error by wrapping malloc
-      1.44  (2014-08-07)
-              various warning fixes from Ronny Chevalier
-      1.43  (2014-07-15)
-              fix MSVC-only compiler problem in code changed in 1.42
-      1.42  (2014-07-09)
-              don't define _CRT_SECURE_NO_WARNINGS (affects user code)
-              fixes to stbi__cleanup_jpeg path
-              added STBI_ASSERT to avoid requiring assert.h
-      1.41  (2014-06-25)
-              fix search&replace from 1.36 that messed up comments/error messages
-      1.40  (2014-06-22)
-              fix gcc struct-initialization warning
-      1.39  (2014-06-15)
-              fix to TGA optimization when req_comp != number of components in TGA;
-              fix to GIF loading because BMP wasn't rewinding (whoops, no GIFs in my test suite)
-              add support for BMP version 5 (more ignored fields)
-      1.38  (2014-06-06)
-              suppress MSVC warnings on integer casts truncating values
-              fix accidental rename of 'skip' field of I/O
-      1.37  (2014-06-04)
-              remove duplicate typedef
-      1.36  (2014-06-03)
-              convert to header file single-file library
-              if de-iphone isn't set, load iphone images color-swapped instead of returning NULL
-      1.35  (2014-05-27)
-              various warnings
-              fix broken STBI_SIMD path
-              fix bug where stbi_load_from_file no longer left file pointer in correct place
-              fix broken non-easy path for 32-bit BMP (possibly never used)
-              TGA optimization by Arseny Kapoulkine
-      1.34  (unknown)
-              use STBI_NOTUSED in stbi__resample_row_generic(), fix one more leak in tga failure case
-      1.33  (2011-07-14)
-              make stbi_is_hdr work in STBI_NO_HDR (as specified), minor compiler-friendly improvements
-      1.32  (2011-07-13)
-              support for "info" function for all supported filetypes (SpartanJ)
-      1.31  (2011-06-20)
-              a few more leak fixes, bug in PNG handling (SpartanJ)
-      1.30  (2011-06-11)
-              added ability to load files via callbacks to accomidate custom input streams (Ben Wenger)
-              removed deprecated format-specific test/load functions
-              removed support for installable file formats (stbi_loader) -- would have been broken for IO callbacks anyway
-              error cases in bmp and tga give messages and don't leak (Raymond Barbiero, grisha)
-              fix inefficiency in decoding 32-bit BMP (David Woo)
-      1.29  (2010-08-16)
-              various warning fixes from Aurelien Pocheville
-      1.28  (2010-08-01)
-              fix bug in GIF palette transparency (SpartanJ)
-      1.27  (2010-08-01)
-              cast-to-stbi_uc to fix warnings
-      1.26  (2010-07-24)
-              fix bug in file buffering for PNG reported by SpartanJ
-      1.25  (2010-07-17)
-              refix trans_data warning (Won Chun)
-      1.24  (2010-07-12)
-              perf improvements reading from files on platforms with lock-heavy fgetc()
-              minor perf improvements for jpeg
-              deprecated type-specific functions so we'll get feedback if they're needed
-              attempt to fix trans_data warning (Won Chun)
-      1.23    fixed bug in iPhone support
-      1.22  (2010-07-10)
-              removed image *writing* support
-              stbi_info support from Jetro Lauha
-              GIF support from Jean-Marc Lienher
-              iPhone PNG-extensions from James Brown
-              warning-fixes from Nicolas Schulz and Janez Zemva (i.stbi__err. Janez (U+017D)emva)
-      1.21    fix use of 'stbi_uc' in header (reported by jon blow)
-      1.20    added support for Softimage PIC, by Tom Seddon
-      1.19    bug in interlaced PNG corruption check (found by ryg)
-      1.18  (2008-08-02)
-              fix a threading bug (local mutable static)
-      1.17    support interlaced PNG
-      1.16    major bugfix - stbi__convert_format converted one too many pixels
-      1.15    initialize some fields for thread safety
-      1.14    fix threadsafe conversion bug
-              header-file-only version (#define STBI_HEADER_FILE_ONLY before including)
-      1.13    threadsafe
-      1.12    const qualifiers in the API
-      1.11    Support installable IDCT, colorspace conversion routines
-      1.10    Fixes for 64-bit (don't use "unsigned long")
-              optimized upsampling by Fabian "ryg" Giesen
-      1.09    Fix format-conversion for PSD code (bad global variables!)
-      1.08    Thatcher Ulrich's PSD code integrated by Nicolas Schulz
-      1.07    attempt to fix C++ warning/errors again
-      1.06    attempt to fix C++ warning/errors again
-      1.05    fix TGA loading to return correct *comp and use good luminance calc
-      1.04    default float alpha is 1, not 255; use 'void *' for stbi_image_free
-      1.03    bugfixes to STBI_NO_STDIO, STBI_NO_HDR
-      1.02    support for (subset of) HDR files, float interface for preferred access to them
-      1.01    fix bug: possible bug in handling right-side up bmps... not sure
-              fix bug: the stbi__bmp_load() and stbi__tga_load() functions didn't work at all
-      1.00    interface to zlib that skips zlib header
-      0.99    correct handling of alpha in palette
-      0.98    TGA loader by lonesock; dynamically add loaders (untested)
-      0.97    jpeg errors on too large a file; also catch another malloc failure
-      0.96    fix detection of invalid v value - particleman@mollyrocket forum
-      0.95    during header scan, seek to markers in case of padding
-      0.94    STBI_NO_STDIO to disable stdio usage; rename all #defines the same
-      0.93    handle jpegtran output; verbose errors
-      0.92    read 4,8,16,24,32-bit BMP files of several formats
-      0.91    output 24-bit Windows 3.0 BMP files
-      0.90    fix a few more warnings; bump version number to approach 1.0
-      0.61    bugfixes due to Marc LeBlanc, Christopher Lloyd
-      0.60    fix compiling as c++
-      0.59    fix warnings: merge Dave Moore's -Wall fixes
-      0.58    fix bug: zlib uncompressed mode len/nlen was wrong endian
-      0.57    fix bug: jpg last huffman symbol before marker was >9 bits but less than 16 available
-      0.56    fix bug: zlib uncompressed mode len vs. nlen
-      0.55    fix bug: restart_interval not initialized to 0
-      0.54    allow NULL for 'int *comp'
-      0.53    fix bug in png 3->4; speedup png decoding
-      0.52    png handles req_comp=3,4 directly; minor cleanup; jpeg comments
-      0.51    obey req_comp requests, 1-component jpegs return as 1-component,
-              on 'test' only check type, not whether we support this variant
-      0.50  (2006-11-19)
-              first released version
-*/
-
-
-/*
-------------------------------------------------------------------------------
-This software is available under 2 licenses -- choose whichever you prefer.
-------------------------------------------------------------------------------
-ALTERNATIVE A - MIT License
-Copyright (c) 2017 Sean Barrett
-Permission is hereby granted, free of charge, to any person obtaining a copy of
-this software and associated documentation files (the "Software"), to deal in
-the Software without restriction, including without limitation the rights to
-use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
-of the Software, and to permit persons to whom the Software is furnished to do
-so, subject to the following conditions:
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
-------------------------------------------------------------------------------
-ALTERNATIVE B - Public Domain (www.unlicense.org)
-This is free and unencumbered software released into the public domain.
-Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
-software, either in source code form or as a compiled binary, for any purpose,
-commercial or non-commercial, and by any means.
-In jurisdictions that recognize copyright laws, the author or authors of this
-software dedicate any and all copyright interest in the software to the public
-domain. We make this dedication for the benefit of the public at large and to
-the detriment of our heirs and successors. We intend this dedication to be an
-overt act of relinquishment in perpetuity of all present and future rights to
-this software under copyright law.
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
-ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
-WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
-------------------------------------------------------------------------------
-*/
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/stb_image_write.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/stb_image_write.h
deleted file mode 100644
index aa397c09..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/stb_image_write.h
+++ /dev/null
@@ -1,1724 +0,0 @@
-/* stb_image_write - v1.16 - public domain - http://nothings.org/stb
-   writes out PNG/BMP/TGA/JPEG/HDR images to C stdio - Sean Barrett 2010-2015
-                                     no warranty implied; use at your own risk
-
-   Before #including,
-
-       #define STB_IMAGE_WRITE_IMPLEMENTATION
-
-   in the file that you want to have the implementation.
-
-   Will probably not work correctly with strict-aliasing optimizations.
-
-ABOUT:
-
-   This header file is a library for writing images to C stdio or a callback.
-
-   The PNG output is not optimal; it is 20-50% larger than the file
-   written by a decent optimizing implementation; though providing a custom
-   zlib compress function (see STBIW_ZLIB_COMPRESS) can mitigate that.
-   This library is designed for source code compactness and simplicity,
-   not optimal image file size or run-time performance.
-
-BUILDING:
-
-   You can #define STBIW_ASSERT(x) before the #include to avoid using assert.h.
-   You can #define STBIW_MALLOC(), STBIW_REALLOC(), and STBIW_FREE() to replace
-   malloc,realloc,free.
-   You can #define STBIW_MEMMOVE() to replace memmove()
-   You can #define STBIW_ZLIB_COMPRESS to use a custom zlib-style compress function
-   for PNG compression (instead of the builtin one), it must have the following signature:
-   unsigned char * my_compress(unsigned char *data, int data_len, int *out_len, int quality);
-   The returned data will be freed with STBIW_FREE() (free() by default),
-   so it must be heap allocated with STBIW_MALLOC() (malloc() by default),
-
-UNICODE:
-
-   If compiling for Windows and you wish to use Unicode filenames, compile
-   with
-       #define STBIW_WINDOWS_UTF8
-   and pass utf8-encoded filenames. Call stbiw_convert_wchar_to_utf8 to convert
-   Windows wchar_t filenames to utf8.
-
-USAGE:
-
-   There are five functions, one for each image file format:
-
-     int stbi_write_png(char const *filename, int w, int h, int comp, const void *data, int stride_in_bytes);
-     int stbi_write_bmp(char const *filename, int w, int h, int comp, const void *data);
-     int stbi_write_tga(char const *filename, int w, int h, int comp, const void *data);
-     int stbi_write_jpg(char const *filename, int w, int h, int comp, const void *data, int quality);
-     int stbi_write_hdr(char const *filename, int w, int h, int comp, const float *data);
-
-     void stbi_flip_vertically_on_write(int flag); // flag is non-zero to flip data vertically
-
-   There are also five equivalent functions that use an arbitrary write function. You are
-   expected to open/close your file-equivalent before and after calling these:
-
-     int stbi_write_png_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data, int stride_in_bytes);
-     int stbi_write_bmp_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
-     int stbi_write_tga_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
-     int stbi_write_hdr_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const float *data);
-     int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int quality);
-
-   where the callback is:
-      void stbi_write_func(void *context, void *data, int size);
-
-   You can configure it with these global variables:
-      int stbi_write_tga_with_rle;             // defaults to true; set to 0 to disable RLE
-      int stbi_write_png_compression_level;    // defaults to 8; set to higher for more compression
-      int stbi_write_force_png_filter;         // defaults to -1; set to 0..5 to force a filter mode
-
-
-   You can define STBI_WRITE_NO_STDIO to disable the file variant of these
-   functions, so the library will not use stdio.h at all. However, this will
-   also disable HDR writing, because it requires stdio for formatted output.
-
-   Each function returns 0 on failure and non-0 on success.
-
-   The functions create an image file defined by the parameters. The image
-   is a rectangle of pixels stored from left-to-right, top-to-bottom.
-   Each pixel contains 'comp' channels of data stored interleaved with 8-bits
-   per channel, in the following order: 1=Y, 2=YA, 3=RGB, 4=RGBA. (Y is
-   monochrome color.) The rectangle is 'w' pixels wide and 'h' pixels tall.
-   The *data pointer points to the first byte of the top-left-most pixel.
-   For PNG, "stride_in_bytes" is the distance in bytes from the first byte of
-   a row of pixels to the first byte of the next row of pixels.
-
-   PNG creates output files with the same number of components as the input.
-   The BMP format expands Y to RGB in the file format and does not
-   output alpha.
-
-   PNG supports writing rectangles of data even when the bytes storing rows of
-   data are not consecutive in memory (e.g. sub-rectangles of a larger image),
-   by supplying the stride between the beginning of adjacent rows. The other
-   formats do not. (Thus you cannot write a native-format BMP through the BMP
-   writer, both because it is in BGR order and because it may have padding
-   at the end of the line.)
-
-   PNG allows you to set the deflate compression level by setting the global
-   variable 'stbi_write_png_compression_level' (it defaults to 8).
-
-   HDR expects linear float data. Since the format is always 32-bit rgb(e)
-   data, alpha (if provided) is discarded, and for monochrome data it is
-   replicated across all three channels.
-
-   TGA supports RLE or non-RLE compressed data. To use non-RLE-compressed
-   data, set the global variable 'stbi_write_tga_with_rle' to 0.
-
-   JPEG does ignore alpha channels in input data; quality is between 1 and 100.
-   Higher quality looks better but results in a bigger image.
-   JPEG baseline (no JPEG progressive).
-
-CREDITS:
-
-
-   Sean Barrett           -    PNG/BMP/TGA
-   Baldur Karlsson        -    HDR
-   Jean-Sebastien Guay    -    TGA monochrome
-   Tim Kelsey             -    misc enhancements
-   Alan Hickman           -    TGA RLE
-   Emmanuel Julien        -    initial file IO callback implementation
-   Jon Olick              -    original jo_jpeg.cpp code
-   Daniel Gibson          -    integrate JPEG, allow external zlib
-   Aarni Koskela          -    allow choosing PNG filter
-
-   bugfixes:
-      github:Chribba
-      Guillaume Chereau
-      github:jry2
-      github:romigrou
-      Sergio Gonzalez
-      Jonas Karlsson
-      Filip Wasil
-      Thatcher Ulrich
-      github:poppolopoppo
-      Patrick Boettcher
-      github:xeekworx
-      Cap Petschulat
-      Simon Rodriguez
-      Ivan Tikhonov
-      github:ignotion
-      Adam Schackart
-      Andrew Kensler
-
-LICENSE
-
-  See end of file for license information.
-
-*/
-
-#ifndef INCLUDE_STB_IMAGE_WRITE_H
-#define INCLUDE_STB_IMAGE_WRITE_H
-
-#include <stdlib.h>
-
-// if STB_IMAGE_WRITE_STATIC causes problems, try defining STBIWDEF to 'inline' or 'static inline'
-#ifndef STBIWDEF
-#ifdef STB_IMAGE_WRITE_STATIC
-#define STBIWDEF  static
-#else
-#ifdef __cplusplus
-#define STBIWDEF  extern "C"
-#else
-#define STBIWDEF  extern
-#endif
-#endif
-#endif
-
-#ifndef STB_IMAGE_WRITE_STATIC  // C++ forbids static forward declarations
-STBIWDEF int stbi_write_tga_with_rle;
-STBIWDEF int stbi_write_png_compression_level;
-STBIWDEF int stbi_write_force_png_filter;
-#endif
-
-#ifndef STBI_WRITE_NO_STDIO
-STBIWDEF int stbi_write_png(char const *filename, int w, int h, int comp, const void  *data, int stride_in_bytes);
-STBIWDEF int stbi_write_bmp(char const *filename, int w, int h, int comp, const void  *data);
-STBIWDEF int stbi_write_tga(char const *filename, int w, int h, int comp, const void  *data);
-STBIWDEF int stbi_write_hdr(char const *filename, int w, int h, int comp, const float *data);
-STBIWDEF int stbi_write_jpg(char const *filename, int x, int y, int comp, const void  *data, int quality);
-
-#ifdef STBIW_WINDOWS_UTF8
-STBIWDEF int stbiw_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input);
-#endif
-#endif
-
-typedef void stbi_write_func(void *context, void *data, int size);
-
-STBIWDEF int stbi_write_png_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data, int stride_in_bytes);
-STBIWDEF int stbi_write_bmp_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
-STBIWDEF int stbi_write_tga_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
-STBIWDEF int stbi_write_hdr_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const float *data);
-STBIWDEF int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void  *data, int quality);
-
-STBIWDEF void stbi_flip_vertically_on_write(int flip_boolean);
-
-#endif//INCLUDE_STB_IMAGE_WRITE_H
-
-#ifdef STB_IMAGE_WRITE_IMPLEMENTATION
-
-#ifdef _WIN32
-   #ifndef _CRT_SECURE_NO_WARNINGS
-   #define _CRT_SECURE_NO_WARNINGS
-   #endif
-   #ifndef _CRT_NONSTDC_NO_DEPRECATE
-   #define _CRT_NONSTDC_NO_DEPRECATE
-   #endif
-#endif
-
-#ifndef STBI_WRITE_NO_STDIO
-#include <stdio.h>
-#endif // STBI_WRITE_NO_STDIO
-
-#include <stdarg.h>
-#include <stdlib.h>
-#include <string.h>
-
-
-#if defined(STBIW_MALLOC) && defined(STBIW_FREE) && (defined(STBIW_REALLOC) || defined(STBIW_REALLOC_SIZED))
-// ok
-#elif !defined(STBIW_MALLOC) && !defined(STBIW_FREE) && !defined(STBIW_REALLOC) && !defined(STBIW_REALLOC_SIZED)
-// ok
-#else
-#error "Must define all or none of STBIW_MALLOC, STBIW_FREE, and STBIW_REALLOC (or STBIW_REALLOC_SIZED)."
-#endif
-
-#ifndef STBIW_MALLOC
-#define STBIW_MALLOC(sz)        malloc(sz)
-#define STBIW_REALLOC(p,newsz)  realloc(p,newsz)
-#define STBIW_FREE(p)           free(p)
-#endif
-
-#ifndef STBIW_REALLOC_SIZED
-#define STBIW_REALLOC_SIZED(p,oldsz,newsz) STBIW_REALLOC(p,newsz)
-#endif
-
-
-#ifndef STBIW_MEMMOVE
-#define STBIW_MEMMOVE(a,b,sz) memmove(a,b,sz)
-#endif
-
-
-#ifndef STBIW_ASSERT
-#include <assert.h>
-#define STBIW_ASSERT(x) assert(x)
-#endif
-
-#define STBIW_UCHAR(x) (unsigned char) ((x) & 0xff)
-
-#ifdef STB_IMAGE_WRITE_STATIC
-static int stbi_write_png_compression_level = 8;
-static int stbi_write_tga_with_rle = 1;
-static int stbi_write_force_png_filter = -1;
-#else
-int stbi_write_png_compression_level = 8;
-int stbi_write_tga_with_rle = 1;
-int stbi_write_force_png_filter = -1;
-#endif
-
-static int stbi__flip_vertically_on_write = 0;
-
-STBIWDEF void stbi_flip_vertically_on_write(int flag)
-{
-   stbi__flip_vertically_on_write = flag;
-}
-
-typedef struct
-{
-   stbi_write_func *func;
-   void *context;
-   unsigned char buffer[64];
-   int buf_used;
-} stbi__write_context;
-
-// initialize a callback-based context
-static void stbi__start_write_callbacks(stbi__write_context *s, stbi_write_func *c, void *context)
-{
-   s->func    = c;
-   s->context = context;
-}
-
-#ifndef STBI_WRITE_NO_STDIO
-
-static void stbi__stdio_write(void *context, void *data, int size)
-{
-   fwrite(data,1,size,(FILE*) context);
-}
-
-#if defined(_WIN32) && defined(STBIW_WINDOWS_UTF8)
-#ifdef __cplusplus
-#define STBIW_EXTERN extern "C"
-#else
-#define STBIW_EXTERN extern
-#endif
-STBIW_EXTERN __declspec(dllimport) int __stdcall MultiByteToWideChar(unsigned int cp, unsigned long flags, const char *str, int cbmb, wchar_t *widestr, int cchwide);
-STBIW_EXTERN __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int cp, unsigned long flags, const wchar_t *widestr, int cchwide, char *str, int cbmb, const char *defchar, int *used_default);
-
-STBIWDEF int stbiw_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input)
-{
-   return WideCharToMultiByte(65001 /* UTF8 */, 0, input, -1, buffer, (int) bufferlen, NULL, NULL);
-}
-#endif
-
-static FILE *stbiw__fopen(char const *filename, char const *mode)
-{
-   FILE *f;
-#if defined(_WIN32) && defined(STBIW_WINDOWS_UTF8)
-   wchar_t wMode[64];
-   wchar_t wFilename[1024];
-   if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename)/sizeof(*wFilename)))
-      return 0;
-
-   if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode)/sizeof(*wMode)))
-      return 0;
-
-#if defined(_MSC_VER) && _MSC_VER >= 1400
-   if (0 != _wfopen_s(&f, wFilename, wMode))
-      f = 0;
-#else
-   f = _wfopen(wFilename, wMode);
-#endif
-
-#elif defined(_MSC_VER) && _MSC_VER >= 1400
-   if (0 != fopen_s(&f, filename, mode))
-      f=0;
-#else
-   f = fopen(filename, mode);
-#endif
-   return f;
-}
-
-static int stbi__start_write_file(stbi__write_context *s, const char *filename)
-{
-   FILE *f = stbiw__fopen(filename, "wb");
-   stbi__start_write_callbacks(s, stbi__stdio_write, (void *) f);
-   return f != NULL;
-}
-
-static void stbi__end_write_file(stbi__write_context *s)
-{
-   fclose((FILE *)s->context);
-}
-
-#endif // !STBI_WRITE_NO_STDIO
-
-typedef unsigned int stbiw_uint32;
-typedef int stb_image_write_test[sizeof(stbiw_uint32)==4 ? 1 : -1];
-
-static void stbiw__writefv(stbi__write_context *s, const char *fmt, va_list v)
-{
-   while (*fmt) {
-      switch (*fmt++) {
-         case ' ': break;
-         case '1': { unsigned char x = STBIW_UCHAR(va_arg(v, int));
-                     s->func(s->context,&x,1);
-                     break; }
-         case '2': { int x = va_arg(v,int);
-                     unsigned char b[2];
-                     b[0] = STBIW_UCHAR(x);
-                     b[1] = STBIW_UCHAR(x>>8);
-                     s->func(s->context,b,2);
-                     break; }
-         case '4': { stbiw_uint32 x = va_arg(v,int);
-                     unsigned char b[4];
-                     b[0]=STBIW_UCHAR(x);
-                     b[1]=STBIW_UCHAR(x>>8);
-                     b[2]=STBIW_UCHAR(x>>16);
-                     b[3]=STBIW_UCHAR(x>>24);
-                     s->func(s->context,b,4);
-                     break; }
-         default:
-            STBIW_ASSERT(0);
-            return;
-      }
-   }
-}
-
-static void stbiw__writef(stbi__write_context *s, const char *fmt, ...)
-{
-   va_list v;
-   va_start(v, fmt);
-   stbiw__writefv(s, fmt, v);
-   va_end(v);
-}
-
-static void stbiw__write_flush(stbi__write_context *s)
-{
-   if (s->buf_used) {
-      s->func(s->context, &s->buffer, s->buf_used);
-      s->buf_used = 0;
-   }
-}
-
-static void stbiw__putc(stbi__write_context *s, unsigned char c)
-{
-   s->func(s->context, &c, 1);
-}
-
-static void stbiw__write1(stbi__write_context *s, unsigned char a)
-{
-   if ((size_t)s->buf_used + 1 > sizeof(s->buffer))
-      stbiw__write_flush(s);
-   s->buffer[s->buf_used++] = a;
-}
-
-static void stbiw__write3(stbi__write_context *s, unsigned char a, unsigned char b, unsigned char c)
-{
-   int n;
-   if ((size_t)s->buf_used + 3 > sizeof(s->buffer))
-      stbiw__write_flush(s);
-   n = s->buf_used;
-   s->buf_used = n+3;
-   s->buffer[n+0] = a;
-   s->buffer[n+1] = b;
-   s->buffer[n+2] = c;
-}
-
-static void stbiw__write_pixel(stbi__write_context *s, int rgb_dir, int comp, int write_alpha, int expand_mono, unsigned char *d)
-{
-   unsigned char bg[3] = { 255, 0, 255}, px[3];
-   int k;
-
-   if (write_alpha < 0)
-      stbiw__write1(s, d[comp - 1]);
-
-   switch (comp) {
-      case 2: // 2 pixels = mono + alpha, alpha is written separately, so same as 1-channel case
-      case 1:
-         if (expand_mono)
-            stbiw__write3(s, d[0], d[0], d[0]); // monochrome bmp
-         else
-            stbiw__write1(s, d[0]);  // monochrome TGA
-         break;
-      case 4:
-         if (!write_alpha) {
-            // composite against pink background
-            for (k = 0; k < 3; ++k)
-               px[k] = bg[k] + ((d[k] - bg[k]) * d[3]) / 255;
-            stbiw__write3(s, px[1 - rgb_dir], px[1], px[1 + rgb_dir]);
-            break;
-         }
-         /* FALLTHROUGH */
-      case 3:
-         stbiw__write3(s, d[1 - rgb_dir], d[1], d[1 + rgb_dir]);
-         break;
-   }
-   if (write_alpha > 0)
-      stbiw__write1(s, d[comp - 1]);
-}
-
-static void stbiw__write_pixels(stbi__write_context *s, int rgb_dir, int vdir, int x, int y, int comp, void *data, int write_alpha, int scanline_pad, int expand_mono)
-{
-   stbiw_uint32 zero = 0;
-   int i,j, j_end;
-
-   if (y <= 0)
-      return;
-
-   if (stbi__flip_vertically_on_write)
-      vdir *= -1;
-
-   if (vdir < 0) {
-      j_end = -1; j = y-1;
-   } else {
-      j_end =  y; j = 0;
-   }
-
-   for (; j != j_end; j += vdir) {
-      for (i=0; i < x; ++i) {
-         unsigned char *d = (unsigned char *) data + (j*x+i)*comp;
-         stbiw__write_pixel(s, rgb_dir, comp, write_alpha, expand_mono, d);
-      }
-      stbiw__write_flush(s);
-      s->func(s->context, &zero, scanline_pad);
-   }
-}
-
-static int stbiw__outfile(stbi__write_context *s, int rgb_dir, int vdir, int x, int y, int comp, int expand_mono, void *data, int alpha, int pad, const char *fmt, ...)
-{
-   if (y < 0 || x < 0) {
-      return 0;
-   } else {
-      va_list v;
-      va_start(v, fmt);
-      stbiw__writefv(s, fmt, v);
-      va_end(v);
-      stbiw__write_pixels(s,rgb_dir,vdir,x,y,comp,data,alpha,pad, expand_mono);
-      return 1;
-   }
-}
-
-static int stbi_write_bmp_core(stbi__write_context *s, int x, int y, int comp, const void *data)
-{
-   if (comp != 4) {
-      // write RGB bitmap
-      int pad = (-x*3) & 3;
-      return stbiw__outfile(s,-1,-1,x,y,comp,1,(void *) data,0,pad,
-              "11 4 22 4" "4 44 22 444444",
-              'B', 'M', 14+40+(x*3+pad)*y, 0,0, 14+40,  // file header
-               40, x,y, 1,24, 0,0,0,0,0,0);             // bitmap header
-   } else {
-      // RGBA bitmaps need a v4 header
-      // use BI_BITFIELDS mode with 32bpp and alpha mask
-      // (straight BI_RGB with alpha mask doesn't work in most readers)
-      return stbiw__outfile(s,-1,-1,x,y,comp,1,(void *)data,1,0,
-         "11 4 22 4" "4 44 22 444444 4444 4 444 444 444 444",
-         'B', 'M', 14+108+x*y*4, 0, 0, 14+108, // file header
-         108, x,y, 1,32, 3,0,0,0,0,0, 0xff0000,0xff00,0xff,0xff000000u, 0, 0,0,0, 0,0,0, 0,0,0, 0,0,0); // bitmap V4 header
-   }
-}
-
-STBIWDEF int stbi_write_bmp_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data)
-{
-   stbi__write_context s = { 0 };
-   stbi__start_write_callbacks(&s, func, context);
-   return stbi_write_bmp_core(&s, x, y, comp, data);
-}
-
-#ifndef STBI_WRITE_NO_STDIO
-STBIWDEF int stbi_write_bmp(char const *filename, int x, int y, int comp, const void *data)
-{
-   stbi__write_context s = { 0 };
-   if (stbi__start_write_file(&s,filename)) {
-      int r = stbi_write_bmp_core(&s, x, y, comp, data);
-      stbi__end_write_file(&s);
-      return r;
-   } else
-      return 0;
-}
-#endif //!STBI_WRITE_NO_STDIO
-
-static int stbi_write_tga_core(stbi__write_context *s, int x, int y, int comp, void *data)
-{
-   int has_alpha = (comp == 2 || comp == 4);
-   int colorbytes = has_alpha ? comp-1 : comp;
-   int format = colorbytes < 2 ? 3 : 2; // 3 color channels (RGB/RGBA) = 2, 1 color channel (Y/YA) = 3
-
-   if (y < 0 || x < 0)
-      return 0;
-
-   if (!stbi_write_tga_with_rle) {
-      return stbiw__outfile(s, -1, -1, x, y, comp, 0, (void *) data, has_alpha, 0,
-         "111 221 2222 11", 0, 0, format, 0, 0, 0, 0, 0, x, y, (colorbytes + has_alpha) * 8, has_alpha * 8);
-   } else {
-      int i,j,k;
-      int jend, jdir;
-
-      stbiw__writef(s, "111 221 2222 11", 0,0,format+8, 0,0,0, 0,0,x,y, (colorbytes + has_alpha) * 8, has_alpha * 8);
-
-      if (stbi__flip_vertically_on_write) {
-         j = 0;
-         jend = y;
-         jdir = 1;
-      } else {
-         j = y-1;
-         jend = -1;
-         jdir = -1;
-      }
-      for (; j != jend; j += jdir) {
-         unsigned char *row = (unsigned char *) data + j * x * comp;
-         int len;
-
-         for (i = 0; i < x; i += len) {
-            unsigned char *begin = row + i * comp;
-            int diff = 1;
-            len = 1;
-
-            if (i < x - 1) {
-               ++len;
-               diff = memcmp(begin, row + (i + 1) * comp, comp);
-               if (diff) {
-                  const unsigned char *prev = begin;
-                  for (k = i + 2; k < x && len < 128; ++k) {
-                     if (memcmp(prev, row + k * comp, comp)) {
-                        prev += comp;
-                        ++len;
-                     } else {
-                        --len;
-                        break;
-                     }
-                  }
-               } else {
-                  for (k = i + 2; k < x && len < 128; ++k) {
-                     if (!memcmp(begin, row + k * comp, comp)) {
-                        ++len;
-                     } else {
-                        break;
-                     }
-                  }
-               }
-            }
-
-            if (diff) {
-               unsigned char header = STBIW_UCHAR(len - 1);
-               stbiw__write1(s, header);
-               for (k = 0; k < len; ++k) {
-                  stbiw__write_pixel(s, -1, comp, has_alpha, 0, begin + k * comp);
-               }
-            } else {
-               unsigned char header = STBIW_UCHAR(len - 129);
-               stbiw__write1(s, header);
-               stbiw__write_pixel(s, -1, comp, has_alpha, 0, begin);
-            }
-         }
-      }
-      stbiw__write_flush(s);
-   }
-   return 1;
-}
-
-STBIWDEF int stbi_write_tga_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data)
-{
-   stbi__write_context s = { 0 };
-   stbi__start_write_callbacks(&s, func, context);
-   return stbi_write_tga_core(&s, x, y, comp, (void *) data);
-}
-
-#ifndef STBI_WRITE_NO_STDIO
-STBIWDEF int stbi_write_tga(char const *filename, int x, int y, int comp, const void *data)
-{
-   stbi__write_context s = { 0 };
-   if (stbi__start_write_file(&s,filename)) {
-      int r = stbi_write_tga_core(&s, x, y, comp, (void *) data);
-      stbi__end_write_file(&s);
-      return r;
-   } else
-      return 0;
-}
-#endif
-
-// *************************************************************************************************
-// Radiance RGBE HDR writer
-// by Baldur Karlsson
-
-#define stbiw__max(a, b)  ((a) > (b) ? (a) : (b))
-
-#ifndef STBI_WRITE_NO_STDIO
-
-static void stbiw__linear_to_rgbe(unsigned char *rgbe, float *linear)
-{
-   int exponent;
-   float maxcomp = stbiw__max(linear[0], stbiw__max(linear[1], linear[2]));
-
-   if (maxcomp < 1e-32f) {
-      rgbe[0] = rgbe[1] = rgbe[2] = rgbe[3] = 0;
-   } else {
-      float normalize = (float) frexp(maxcomp, &exponent) * 256.0f/maxcomp;
-
-      rgbe[0] = (unsigned char)(linear[0] * normalize);
-      rgbe[1] = (unsigned char)(linear[1] * normalize);
-      rgbe[2] = (unsigned char)(linear[2] * normalize);
-      rgbe[3] = (unsigned char)(exponent + 128);
-   }
-}
-
-static void stbiw__write_run_data(stbi__write_context *s, int length, unsigned char databyte)
-{
-   unsigned char lengthbyte = STBIW_UCHAR(length+128);
-   STBIW_ASSERT(length+128 <= 255);
-   s->func(s->context, &lengthbyte, 1);
-   s->func(s->context, &databyte, 1);
-}
-
-static void stbiw__write_dump_data(stbi__write_context *s, int length, unsigned char *data)
-{
-   unsigned char lengthbyte = STBIW_UCHAR(length);
-   STBIW_ASSERT(length <= 128); // inconsistent with spec but consistent with official code
-   s->func(s->context, &lengthbyte, 1);
-   s->func(s->context, data, length);
-}
-
-static void stbiw__write_hdr_scanline(stbi__write_context *s, int width, int ncomp, unsigned char *scratch, float *scanline)
-{
-   unsigned char scanlineheader[4] = { 2, 2, 0, 0 };
-   unsigned char rgbe[4];
-   float linear[3];
-   int x;
-
-   scanlineheader[2] = (width&0xff00)>>8;
-   scanlineheader[3] = (width&0x00ff);
-
-   /* skip RLE for images too small or large */
-   if (width < 8 || width >= 32768) {
-      for (x=0; x < width; x++) {
-         switch (ncomp) {
-            case 4: /* fallthrough */
-            case 3: linear[2] = scanline[x*ncomp + 2];
-                    linear[1] = scanline[x*ncomp + 1];
-                    linear[0] = scanline[x*ncomp + 0];
-                    break;
-            default:
-                    linear[0] = linear[1] = linear[2] = scanline[x*ncomp + 0];
-                    break;
-         }
-         stbiw__linear_to_rgbe(rgbe, linear);
-         s->func(s->context, rgbe, 4);
-      }
-   } else {
-      int c,r;
-      /* encode into scratch buffer */
-      for (x=0; x < width; x++) {
-         switch(ncomp) {
-            case 4: /* fallthrough */
-            case 3: linear[2] = scanline[x*ncomp + 2];
-                    linear[1] = scanline[x*ncomp + 1];
-                    linear[0] = scanline[x*ncomp + 0];
-                    break;
-            default:
-                    linear[0] = linear[1] = linear[2] = scanline[x*ncomp + 0];
-                    break;
-         }
-         stbiw__linear_to_rgbe(rgbe, linear);
-         scratch[x + width*0] = rgbe[0];
-         scratch[x + width*1] = rgbe[1];
-         scratch[x + width*2] = rgbe[2];
-         scratch[x + width*3] = rgbe[3];
-      }
-
-      s->func(s->context, scanlineheader, 4);
-
-      /* RLE each component separately */
-      for (c=0; c < 4; c++) {
-         unsigned char *comp = &scratch[width*c];
-
-         x = 0;
-         while (x < width) {
-            // find first run
-            r = x;
-            while (r+2 < width) {
-               if (comp[r] == comp[r+1] && comp[r] == comp[r+2])
-                  break;
-               ++r;
-            }
-            if (r+2 >= width)
-               r = width;
-            // dump up to first run
-            while (x < r) {
-               int len = r-x;
-               if (len > 128) len = 128;
-               stbiw__write_dump_data(s, len, &comp[x]);
-               x += len;
-            }
-            // if there's a run, output it
-            if (r+2 < width) { // same test as what we break out of in search loop, so only true if we break'd
-               // find next byte after run
-               while (r < width && comp[r] == comp[x])
-                  ++r;
-               // output run up to r
-               while (x < r) {
-                  int len = r-x;
-                  if (len > 127) len = 127;
-                  stbiw__write_run_data(s, len, comp[x]);
-                  x += len;
-               }
-            }
-         }
-      }
-   }
-}
-
-static int stbi_write_hdr_core(stbi__write_context *s, int x, int y, int comp, float *data)
-{
-   if (y <= 0 || x <= 0 || data == NULL)
-      return 0;
-   else {
-      // Each component is stored separately. Allocate scratch space for full output scanline.
-      unsigned char *scratch = (unsigned char *) STBIW_MALLOC(x*4);
-      int i, len;
-      char buffer[128];
-      char header[] = "#?RADIANCE\n# Written by stb_image_write.h\nFORMAT=32-bit_rle_rgbe\n";
-      s->func(s->context, header, sizeof(header)-1);
-
-#ifdef __STDC_LIB_EXT1__
-      len = sprintf_s(buffer, sizeof(buffer), "EXPOSURE=          1.0000000000000\n\n-Y %d +X %d\n", y, x);
-#else
-      len = sprintf(buffer, "EXPOSURE=          1.0000000000000\n\n-Y %d +X %d\n", y, x);
-#endif
-      s->func(s->context, buffer, len);
-
-      for(i=0; i < y; i++)
-         stbiw__write_hdr_scanline(s, x, comp, scratch, data + comp*x*(stbi__flip_vertically_on_write ? y-1-i : i));
-      STBIW_FREE(scratch);
-      return 1;
-   }
-}
-
-STBIWDEF int stbi_write_hdr_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const float *data)
-{
-   stbi__write_context s = { 0 };
-   stbi__start_write_callbacks(&s, func, context);
-   return stbi_write_hdr_core(&s, x, y, comp, (float *) data);
-}
-
-STBIWDEF int stbi_write_hdr(char const *filename, int x, int y, int comp, const float *data)
-{
-   stbi__write_context s = { 0 };
-   if (stbi__start_write_file(&s,filename)) {
-      int r = stbi_write_hdr_core(&s, x, y, comp, (float *) data);
-      stbi__end_write_file(&s);
-      return r;
-   } else
-      return 0;
-}
-#endif // STBI_WRITE_NO_STDIO
-
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// PNG writer
-//
-
-#ifndef STBIW_ZLIB_COMPRESS
-// stretchy buffer; stbiw__sbpush() == vector<>::push_back() -- stbiw__sbcount() == vector<>::size()
-#define stbiw__sbraw(a) ((int *) (void *) (a) - 2)
-#define stbiw__sbm(a)   stbiw__sbraw(a)[0]
-#define stbiw__sbn(a)   stbiw__sbraw(a)[1]
-
-#define stbiw__sbneedgrow(a,n)  ((a)==0 || stbiw__sbn(a)+n >= stbiw__sbm(a))
-#define stbiw__sbmaybegrow(a,n) (stbiw__sbneedgrow(a,(n)) ? stbiw__sbgrow(a,n) : 0)
-#define stbiw__sbgrow(a,n)  stbiw__sbgrowf((void **) &(a), (n), sizeof(*(a)))
-
-#define stbiw__sbpush(a, v)      (stbiw__sbmaybegrow(a,1), (a)[stbiw__sbn(a)++] = (v))
-#define stbiw__sbcount(a)        ((a) ? stbiw__sbn(a) : 0)
-#define stbiw__sbfree(a)         ((a) ? STBIW_FREE(stbiw__sbraw(a)),0 : 0)
-
-static void *stbiw__sbgrowf(void **arr, int increment, int itemsize)
-{
-   int m = *arr ? 2*stbiw__sbm(*arr)+increment : increment+1;
-   void *p = STBIW_REALLOC_SIZED(*arr ? stbiw__sbraw(*arr) : 0, *arr ? (stbiw__sbm(*arr)*itemsize + sizeof(int)*2) : 0, itemsize * m + sizeof(int)*2);
-   STBIW_ASSERT(p);
-   if (p) {
-      if (!*arr) ((int *) p)[1] = 0;
-      *arr = (void *) ((int *) p + 2);
-      stbiw__sbm(*arr) = m;
-   }
-   return *arr;
-}
-
-static unsigned char *stbiw__zlib_flushf(unsigned char *data, unsigned int *bitbuffer, int *bitcount)
-{
-   while (*bitcount >= 8) {
-      stbiw__sbpush(data, STBIW_UCHAR(*bitbuffer));
-      *bitbuffer >>= 8;
-      *bitcount -= 8;
-   }
-   return data;
-}
-
-static int stbiw__zlib_bitrev(int code, int codebits)
-{
-   int res=0;
-   while (codebits--) {
-      res = (res << 1) | (code & 1);
-      code >>= 1;
-   }
-   return res;
-}
-
-static unsigned int stbiw__zlib_countm(unsigned char *a, unsigned char *b, int limit)
-{
-   int i;
-   for (i=0; i < limit && i < 258; ++i)
-      if (a[i] != b[i]) break;
-   return i;
-}
-
-static unsigned int stbiw__zhash(unsigned char *data)
-{
-   stbiw_uint32 hash = data[0] + (data[1] << 8) + (data[2] << 16);
-   hash ^= hash << 3;
-   hash += hash >> 5;
-   hash ^= hash << 4;
-   hash += hash >> 17;
-   hash ^= hash << 25;
-   hash += hash >> 6;
-   return hash;
-}
-
-#define stbiw__zlib_flush() (out = stbiw__zlib_flushf(out, &bitbuf, &bitcount))
-#define stbiw__zlib_add(code,codebits) \
-      (bitbuf |= (code) << bitcount, bitcount += (codebits), stbiw__zlib_flush())
-#define stbiw__zlib_huffa(b,c)  stbiw__zlib_add(stbiw__zlib_bitrev(b,c),c)
-// default huffman tables
-#define stbiw__zlib_huff1(n)  stbiw__zlib_huffa(0x30 + (n), 8)
-#define stbiw__zlib_huff2(n)  stbiw__zlib_huffa(0x190 + (n)-144, 9)
-#define stbiw__zlib_huff3(n)  stbiw__zlib_huffa(0 + (n)-256,7)
-#define stbiw__zlib_huff4(n)  stbiw__zlib_huffa(0xc0 + (n)-280,8)
-#define stbiw__zlib_huff(n)  ((n) <= 143 ? stbiw__zlib_huff1(n) : (n) <= 255 ? stbiw__zlib_huff2(n) : (n) <= 279 ? stbiw__zlib_huff3(n) : stbiw__zlib_huff4(n))
-#define stbiw__zlib_huffb(n) ((n) <= 143 ? stbiw__zlib_huff1(n) : stbiw__zlib_huff2(n))
-
-#define stbiw__ZHASH   16384
-
-#endif // STBIW_ZLIB_COMPRESS
-
-STBIWDEF unsigned char * stbi_zlib_compress(unsigned char *data, int data_len, int *out_len, int quality)
-{
-#ifdef STBIW_ZLIB_COMPRESS
-   // user provided a zlib compress implementation, use that
-   return STBIW_ZLIB_COMPRESS(data, data_len, out_len, quality);
-#else // use builtin
-   static unsigned short lengthc[] = { 3,4,5,6,7,8,9,10,11,13,15,17,19,23,27,31,35,43,51,59,67,83,99,115,131,163,195,227,258, 259 };
-   static unsigned char  lengtheb[]= { 0,0,0,0,0,0,0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4,  4,  5,  5,  5,  5,  0 };
-   static unsigned short distc[]   = { 1,2,3,4,5,7,9,13,17,25,33,49,65,97,129,193,257,385,513,769,1025,1537,2049,3073,4097,6145,8193,12289,16385,24577, 32768 };
-   static unsigned char  disteb[]  = { 0,0,0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,8,8,9,9,10,10,11,11,12,12,13,13 };
-   unsigned int bitbuf=0;
-   int i,j, bitcount=0;
-   unsigned char *out = NULL;
-   unsigned char ***hash_table = (unsigned char***) STBIW_MALLOC(stbiw__ZHASH * sizeof(unsigned char**));
-   if (hash_table == NULL)
-      return NULL;
-   if (quality < 5) quality = 5;
-
-   stbiw__sbpush(out, 0x78);   // DEFLATE 32K window
-   stbiw__sbpush(out, 0x5e);   // FLEVEL = 1
-   stbiw__zlib_add(1,1);  // BFINAL = 1
-   stbiw__zlib_add(1,2);  // BTYPE = 1 -- fixed huffman
-
-   for (i=0; i < stbiw__ZHASH; ++i)
-      hash_table[i] = NULL;
-
-   i=0;
-   while (i < data_len-3) {
-      // hash next 3 bytes of data to be compressed
-      int h = stbiw__zhash(data+i)&(stbiw__ZHASH-1), best=3;
-      unsigned char *bestloc = 0;
-      unsigned char **hlist = hash_table[h];
-      int n = stbiw__sbcount(hlist);
-      for (j=0; j < n; ++j) {
-         if (hlist[j]-data > i-32768) { // if entry lies within window
-            int d = stbiw__zlib_countm(hlist[j], data+i, data_len-i);
-            if (d >= best) { best=d; bestloc=hlist[j]; }
-         }
-      }
-      // when hash table entry is too long, delete half the entries
-      if (hash_table[h] && stbiw__sbn(hash_table[h]) == 2*quality) {
-         STBIW_MEMMOVE(hash_table[h], hash_table[h]+quality, sizeof(hash_table[h][0])*quality);
-         stbiw__sbn(hash_table[h]) = quality;
-      }
-      stbiw__sbpush(hash_table[h],data+i);
-
-      if (bestloc) {
-         // "lazy matching" - check match at *next* byte, and if it's better, do cur byte as literal
-         h = stbiw__zhash(data+i+1)&(stbiw__ZHASH-1);
-         hlist = hash_table[h];
-         n = stbiw__sbcount(hlist);
-         for (j=0; j < n; ++j) {
-            if (hlist[j]-data > i-32767) {
-               int e = stbiw__zlib_countm(hlist[j], data+i+1, data_len-i-1);
-               if (e > best) { // if next match is better, bail on current match
-                  bestloc = NULL;
-                  break;
-               }
-            }
-         }
-      }
-
-      if (bestloc) {
-         int d = (int) (data+i - bestloc); // distance back
-         STBIW_ASSERT(d <= 32767 && best <= 258);
-         for (j=0; best > lengthc[j+1]-1; ++j);
-         stbiw__zlib_huff(j+257);
-         if (lengtheb[j]) stbiw__zlib_add(best - lengthc[j], lengtheb[j]);
-         for (j=0; d > distc[j+1]-1; ++j);
-         stbiw__zlib_add(stbiw__zlib_bitrev(j,5),5);
-         if (disteb[j]) stbiw__zlib_add(d - distc[j], disteb[j]);
-         i += best;
-      } else {
-         stbiw__zlib_huffb(data[i]);
-         ++i;
-      }
-   }
-   // write out final bytes
-   for (;i < data_len; ++i)
-      stbiw__zlib_huffb(data[i]);
-   stbiw__zlib_huff(256); // end of block
-   // pad with 0 bits to byte boundary
-   while (bitcount)
-      stbiw__zlib_add(0,1);
-
-   for (i=0; i < stbiw__ZHASH; ++i)
-      (void) stbiw__sbfree(hash_table[i]);
-   STBIW_FREE(hash_table);
-
-   // store uncompressed instead if compression was worse
-   if (stbiw__sbn(out) > data_len + 2 + ((data_len+32766)/32767)*5) {
-      stbiw__sbn(out) = 2;  // truncate to DEFLATE 32K window and FLEVEL = 1
-      for (j = 0; j < data_len;) {
-         int blocklen = data_len - j;
-         if (blocklen > 32767) blocklen = 32767;
-         stbiw__sbpush(out, data_len - j == blocklen); // BFINAL = ?, BTYPE = 0 -- no compression
-         stbiw__sbpush(out, STBIW_UCHAR(blocklen)); // LEN
-         stbiw__sbpush(out, STBIW_UCHAR(blocklen >> 8));
-         stbiw__sbpush(out, STBIW_UCHAR(~blocklen)); // NLEN
-         stbiw__sbpush(out, STBIW_UCHAR(~blocklen >> 8));
-         memcpy(out+stbiw__sbn(out), data+j, blocklen);
-         stbiw__sbn(out) += blocklen;
-         j += blocklen;
-      }
-   }
-
-   {
-      // compute adler32 on input
-      unsigned int s1=1, s2=0;
-      int blocklen = (int) (data_len % 5552);
-      j=0;
-      while (j < data_len) {
-         for (i=0; i < blocklen; ++i) { s1 += data[j+i]; s2 += s1; }
-         s1 %= 65521; s2 %= 65521;
-         j += blocklen;
-         blocklen = 5552;
-      }
-      stbiw__sbpush(out, STBIW_UCHAR(s2 >> 8));
-      stbiw__sbpush(out, STBIW_UCHAR(s2));
-      stbiw__sbpush(out, STBIW_UCHAR(s1 >> 8));
-      stbiw__sbpush(out, STBIW_UCHAR(s1));
-   }
-   *out_len = stbiw__sbn(out);
-   // make returned pointer freeable
-   STBIW_MEMMOVE(stbiw__sbraw(out), out, *out_len);
-   return (unsigned char *) stbiw__sbraw(out);
-#endif // STBIW_ZLIB_COMPRESS
-}
-
-static unsigned int stbiw__crc32(unsigned char *buffer, int len)
-{
-#ifdef STBIW_CRC32
-    return STBIW_CRC32(buffer, len);
-#else
-   static unsigned int crc_table[256] =
-   {
-      0x00000000, 0x77073096, 0xEE0E612C, 0x990951BA, 0x076DC419, 0x706AF48F, 0xE963A535, 0x9E6495A3,
-      0x0eDB8832, 0x79DCB8A4, 0xE0D5E91E, 0x97D2D988, 0x09B64C2B, 0x7EB17CBD, 0xE7B82D07, 0x90BF1D91,
-      0x1DB71064, 0x6AB020F2, 0xF3B97148, 0x84BE41DE, 0x1ADAD47D, 0x6DDDE4EB, 0xF4D4B551, 0x83D385C7,
-      0x136C9856, 0x646BA8C0, 0xFD62F97A, 0x8A65C9EC, 0x14015C4F, 0x63066CD9, 0xFA0F3D63, 0x8D080DF5,
-      0x3B6E20C8, 0x4C69105E, 0xD56041E4, 0xA2677172, 0x3C03E4D1, 0x4B04D447, 0xD20D85FD, 0xA50AB56B,
-      0x35B5A8FA, 0x42B2986C, 0xDBBBC9D6, 0xACBCF940, 0x32D86CE3, 0x45DF5C75, 0xDCD60DCF, 0xABD13D59,
-      0x26D930AC, 0x51DE003A, 0xC8D75180, 0xBFD06116, 0x21B4F4B5, 0x56B3C423, 0xCFBA9599, 0xB8BDA50F,
-      0x2802B89E, 0x5F058808, 0xC60CD9B2, 0xB10BE924, 0x2F6F7C87, 0x58684C11, 0xC1611DAB, 0xB6662D3D,
-      0x76DC4190, 0x01DB7106, 0x98D220BC, 0xEFD5102A, 0x71B18589, 0x06B6B51F, 0x9FBFE4A5, 0xE8B8D433,
-      0x7807C9A2, 0x0F00F934, 0x9609A88E, 0xE10E9818, 0x7F6A0DBB, 0x086D3D2D, 0x91646C97, 0xE6635C01,
-      0x6B6B51F4, 0x1C6C6162, 0x856530D8, 0xF262004E, 0x6C0695ED, 0x1B01A57B, 0x8208F4C1, 0xF50FC457,
-      0x65B0D9C6, 0x12B7E950, 0x8BBEB8EA, 0xFCB9887C, 0x62DD1DDF, 0x15DA2D49, 0x8CD37CF3, 0xFBD44C65,
-      0x4DB26158, 0x3AB551CE, 0xA3BC0074, 0xD4BB30E2, 0x4ADFA541, 0x3DD895D7, 0xA4D1C46D, 0xD3D6F4FB,
-      0x4369E96A, 0x346ED9FC, 0xAD678846, 0xDA60B8D0, 0x44042D73, 0x33031DE5, 0xAA0A4C5F, 0xDD0D7CC9,
-      0x5005713C, 0x270241AA, 0xBE0B1010, 0xC90C2086, 0x5768B525, 0x206F85B3, 0xB966D409, 0xCE61E49F,
-      0x5EDEF90E, 0x29D9C998, 0xB0D09822, 0xC7D7A8B4, 0x59B33D17, 0x2EB40D81, 0xB7BD5C3B, 0xC0BA6CAD,
-      0xEDB88320, 0x9ABFB3B6, 0x03B6E20C, 0x74B1D29A, 0xEAD54739, 0x9DD277AF, 0x04DB2615, 0x73DC1683,
-      0xE3630B12, 0x94643B84, 0x0D6D6A3E, 0x7A6A5AA8, 0xE40ECF0B, 0x9309FF9D, 0x0A00AE27, 0x7D079EB1,
-      0xF00F9344, 0x8708A3D2, 0x1E01F268, 0x6906C2FE, 0xF762575D, 0x806567CB, 0x196C3671, 0x6E6B06E7,
-      0xFED41B76, 0x89D32BE0, 0x10DA7A5A, 0x67DD4ACC, 0xF9B9DF6F, 0x8EBEEFF9, 0x17B7BE43, 0x60B08ED5,
-      0xD6D6A3E8, 0xA1D1937E, 0x38D8C2C4, 0x4FDFF252, 0xD1BB67F1, 0xA6BC5767, 0x3FB506DD, 0x48B2364B,
-      0xD80D2BDA, 0xAF0A1B4C, 0x36034AF6, 0x41047A60, 0xDF60EFC3, 0xA867DF55, 0x316E8EEF, 0x4669BE79,
-      0xCB61B38C, 0xBC66831A, 0x256FD2A0, 0x5268E236, 0xCC0C7795, 0xBB0B4703, 0x220216B9, 0x5505262F,
-      0xC5BA3BBE, 0xB2BD0B28, 0x2BB45A92, 0x5CB36A04, 0xC2D7FFA7, 0xB5D0CF31, 0x2CD99E8B, 0x5BDEAE1D,
-      0x9B64C2B0, 0xEC63F226, 0x756AA39C, 0x026D930A, 0x9C0906A9, 0xEB0E363F, 0x72076785, 0x05005713,
-      0x95BF4A82, 0xE2B87A14, 0x7BB12BAE, 0x0CB61B38, 0x92D28E9B, 0xE5D5BE0D, 0x7CDCEFB7, 0x0BDBDF21,
-      0x86D3D2D4, 0xF1D4E242, 0x68DDB3F8, 0x1FDA836E, 0x81BE16CD, 0xF6B9265B, 0x6FB077E1, 0x18B74777,
-      0x88085AE6, 0xFF0F6A70, 0x66063BCA, 0x11010B5C, 0x8F659EFF, 0xF862AE69, 0x616BFFD3, 0x166CCF45,
-      0xA00AE278, 0xD70DD2EE, 0x4E048354, 0x3903B3C2, 0xA7672661, 0xD06016F7, 0x4969474D, 0x3E6E77DB,
-      0xAED16A4A, 0xD9D65ADC, 0x40DF0B66, 0x37D83BF0, 0xA9BCAE53, 0xDEBB9EC5, 0x47B2CF7F, 0x30B5FFE9,
-      0xBDBDF21C, 0xCABAC28A, 0x53B39330, 0x24B4A3A6, 0xBAD03605, 0xCDD70693, 0x54DE5729, 0x23D967BF,
-      0xB3667A2E, 0xC4614AB8, 0x5D681B02, 0x2A6F2B94, 0xB40BBE37, 0xC30C8EA1, 0x5A05DF1B, 0x2D02EF8D
-   };
-
-   unsigned int crc = ~0u;
-   int i;
-   for (i=0; i < len; ++i)
-      crc = (crc >> 8) ^ crc_table[buffer[i] ^ (crc & 0xff)];
-   return ~crc;
-#endif
-}
-
-#define stbiw__wpng4(o,a,b,c,d) ((o)[0]=STBIW_UCHAR(a),(o)[1]=STBIW_UCHAR(b),(o)[2]=STBIW_UCHAR(c),(o)[3]=STBIW_UCHAR(d),(o)+=4)
-#define stbiw__wp32(data,v) stbiw__wpng4(data, (v)>>24,(v)>>16,(v)>>8,(v));
-#define stbiw__wptag(data,s) stbiw__wpng4(data, s[0],s[1],s[2],s[3])
-
-static void stbiw__wpcrc(unsigned char **data, int len)
-{
-   unsigned int crc = stbiw__crc32(*data - len - 4, len+4);
-   stbiw__wp32(*data, crc);
-}
-
-static unsigned char stbiw__paeth(int a, int b, int c)
-{
-   int p = a + b - c, pa = abs(p-a), pb = abs(p-b), pc = abs(p-c);
-   if (pa <= pb && pa <= pc) return STBIW_UCHAR(a);
-   if (pb <= pc) return STBIW_UCHAR(b);
-   return STBIW_UCHAR(c);
-}
-
-// @OPTIMIZE: provide an option that always forces left-predict or paeth predict
-static void stbiw__encode_png_line(unsigned char *pixels, int stride_bytes, int width, int height, int y, int n, int filter_type, signed char *line_buffer)
-{
-   static int mapping[] = { 0,1,2,3,4 };
-   static int firstmap[] = { 0,1,0,5,6 };
-   int *mymap = (y != 0) ? mapping : firstmap;
-   int i;
-   int type = mymap[filter_type];
-   unsigned char *z = pixels + stride_bytes * (stbi__flip_vertically_on_write ? height-1-y : y);
-   int signed_stride = stbi__flip_vertically_on_write ? -stride_bytes : stride_bytes;
-
-   if (type==0) {
-      memcpy(line_buffer, z, width*n);
-      return;
-   }
-
-   // first loop isn't optimized since it's just one pixel
-   for (i = 0; i < n; ++i) {
-      switch (type) {
-         case 1: line_buffer[i] = z[i]; break;
-         case 2: line_buffer[i] = z[i] - z[i-signed_stride]; break;
-         case 3: line_buffer[i] = z[i] - (z[i-signed_stride]>>1); break;
-         case 4: line_buffer[i] = (signed char) (z[i] - stbiw__paeth(0,z[i-signed_stride],0)); break;
-         case 5: line_buffer[i] = z[i]; break;
-         case 6: line_buffer[i] = z[i]; break;
-      }
-   }
-   switch (type) {
-      case 1: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - z[i-n]; break;
-      case 2: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - z[i-signed_stride]; break;
-      case 3: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - ((z[i-n] + z[i-signed_stride])>>1); break;
-      case 4: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - stbiw__paeth(z[i-n], z[i-signed_stride], z[i-signed_stride-n]); break;
-      case 5: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - (z[i-n]>>1); break;
-      case 6: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - stbiw__paeth(z[i-n], 0,0); break;
-   }
-}
-
-STBIWDEF unsigned char *stbi_write_png_to_mem(const unsigned char *pixels, int stride_bytes, int x, int y, int n, int *out_len)
-{
-   int force_filter = stbi_write_force_png_filter;
-   int ctype[5] = { -1, 0, 4, 2, 6 };
-   unsigned char sig[8] = { 137,80,78,71,13,10,26,10 };
-   unsigned char *out,*o, *filt, *zlib;
-   signed char *line_buffer;
-   int j,zlen;
-
-   if (stride_bytes == 0)
-      stride_bytes = x * n;
-
-   if (force_filter >= 5) {
-      force_filter = -1;
-   }
-
-   filt = (unsigned char *) STBIW_MALLOC((x*n+1) * y); if (!filt) return 0;
-   line_buffer = (signed char *) STBIW_MALLOC(x * n); if (!line_buffer) { STBIW_FREE(filt); return 0; }
-   for (j=0; j < y; ++j) {
-      int filter_type;
-      if (force_filter > -1) {
-         filter_type = force_filter;
-         stbiw__encode_png_line((unsigned char*)(pixels), stride_bytes, x, y, j, n, force_filter, line_buffer);
-      } else { // Estimate the best filter by running through all of them:
-         int best_filter = 0, best_filter_val = 0x7fffffff, est, i;
-         for (filter_type = 0; filter_type < 5; filter_type++) {
-            stbiw__encode_png_line((unsigned char*)(pixels), stride_bytes, x, y, j, n, filter_type, line_buffer);
-
-            // Estimate the entropy of the line using this filter; the less, the better.
-            est = 0;
-            for (i = 0; i < x*n; ++i) {
-               est += abs((signed char) line_buffer[i]);
-            }
-            if (est < best_filter_val) {
-               best_filter_val = est;
-               best_filter = filter_type;
-            }
-         }
-         if (filter_type != best_filter) {  // If the last iteration already got us the best filter, don't redo it
-            stbiw__encode_png_line((unsigned char*)(pixels), stride_bytes, x, y, j, n, best_filter, line_buffer);
-            filter_type = best_filter;
-         }
-      }
-      // when we get here, filter_type contains the filter type, and line_buffer contains the data
-      filt[j*(x*n+1)] = (unsigned char) filter_type;
-      STBIW_MEMMOVE(filt+j*(x*n+1)+1, line_buffer, x*n);
-   }
-   STBIW_FREE(line_buffer);
-   zlib = stbi_zlib_compress(filt, y*( x*n+1), &zlen, stbi_write_png_compression_level);
-   STBIW_FREE(filt);
-   if (!zlib) return 0;
-
-   // each tag requires 12 bytes of overhead
-   out = (unsigned char *) STBIW_MALLOC(8 + 12+13 + 12+zlen + 12);
-   if (!out) return 0;
-   *out_len = 8 + 12+13 + 12+zlen + 12;
-
-   o=out;
-   STBIW_MEMMOVE(o,sig,8); o+= 8;
-   stbiw__wp32(o, 13); // header length
-   stbiw__wptag(o, "IHDR");
-   stbiw__wp32(o, x);
-   stbiw__wp32(o, y);
-   *o++ = 8;
-   *o++ = STBIW_UCHAR(ctype[n]);
-   *o++ = 0;
-   *o++ = 0;
-   *o++ = 0;
-   stbiw__wpcrc(&o,13);
-
-   stbiw__wp32(o, zlen);
-   stbiw__wptag(o, "IDAT");
-   STBIW_MEMMOVE(o, zlib, zlen);
-   o += zlen;
-   STBIW_FREE(zlib);
-   stbiw__wpcrc(&o, zlen);
-
-   stbiw__wp32(o,0);
-   stbiw__wptag(o, "IEND");
-   stbiw__wpcrc(&o,0);
-
-   STBIW_ASSERT(o == out + *out_len);
-
-   return out;
-}
-
-#ifndef STBI_WRITE_NO_STDIO
-STBIWDEF int stbi_write_png(char const *filename, int x, int y, int comp, const void *data, int stride_bytes)
-{
-   FILE *f;
-   int len;
-   unsigned char *png = stbi_write_png_to_mem((const unsigned char *) data, stride_bytes, x, y, comp, &len);
-   if (png == NULL) return 0;
-
-   f = stbiw__fopen(filename, "wb");
-   if (!f) { STBIW_FREE(png); return 0; }
-   fwrite(png, 1, len, f);
-   fclose(f);
-   STBIW_FREE(png);
-   return 1;
-}
-#endif
-
-STBIWDEF int stbi_write_png_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int stride_bytes)
-{
-   int len;
-   unsigned char *png = stbi_write_png_to_mem((const unsigned char *) data, stride_bytes, x, y, comp, &len);
-   if (png == NULL) return 0;
-   func(context, png, len);
-   STBIW_FREE(png);
-   return 1;
-}
-
-
-/* ***************************************************************************
- *
- * JPEG writer
- *
- * This is based on Jon Olick's jo_jpeg.cpp:
- * public domain Simple, Minimalistic JPEG writer - http://www.jonolick.com/code.html
- */
-
-static const unsigned char stbiw__jpg_ZigZag[] = { 0,1,5,6,14,15,27,28,2,4,7,13,16,26,29,42,3,8,12,17,25,30,41,43,9,11,18,
-      24,31,40,44,53,10,19,23,32,39,45,52,54,20,22,33,38,46,51,55,60,21,34,37,47,50,56,59,61,35,36,48,49,57,58,62,63 };
-
-static void stbiw__jpg_writeBits(stbi__write_context *s, int *bitBufP, int *bitCntP, const unsigned short *bs) {
-   int bitBuf = *bitBufP, bitCnt = *bitCntP;
-   bitCnt += bs[1];
-   bitBuf |= bs[0] << (24 - bitCnt);
-   while(bitCnt >= 8) {
-      unsigned char c = (bitBuf >> 16) & 255;
-      stbiw__putc(s, c);
-      if(c == 255) {
-         stbiw__putc(s, 0);
-      }
-      bitBuf <<= 8;
-      bitCnt -= 8;
-   }
-   *bitBufP = bitBuf;
-   *bitCntP = bitCnt;
-}
-
-static void stbiw__jpg_DCT(float *d0p, float *d1p, float *d2p, float *d3p, float *d4p, float *d5p, float *d6p, float *d7p) {
-   float d0 = *d0p, d1 = *d1p, d2 = *d2p, d3 = *d3p, d4 = *d4p, d5 = *d5p, d6 = *d6p, d7 = *d7p;
-   float z1, z2, z3, z4, z5, z11, z13;
-
-   float tmp0 = d0 + d7;
-   float tmp7 = d0 - d7;
-   float tmp1 = d1 + d6;
-   float tmp6 = d1 - d6;
-   float tmp2 = d2 + d5;
-   float tmp5 = d2 - d5;
-   float tmp3 = d3 + d4;
-   float tmp4 = d3 - d4;
-
-   // Even part
-   float tmp10 = tmp0 + tmp3;   // phase 2
-   float tmp13 = tmp0 - tmp3;
-   float tmp11 = tmp1 + tmp2;
-   float tmp12 = tmp1 - tmp2;
-
-   d0 = tmp10 + tmp11;       // phase 3
-   d4 = tmp10 - tmp11;
-
-   z1 = (tmp12 + tmp13) * 0.707106781f; // c4
-   d2 = tmp13 + z1;       // phase 5
-   d6 = tmp13 - z1;
-
-   // Odd part
-   tmp10 = tmp4 + tmp5;       // phase 2
-   tmp11 = tmp5 + tmp6;
-   tmp12 = tmp6 + tmp7;
-
-   // The rotator is modified from fig 4-8 to avoid extra negations.
-   z5 = (tmp10 - tmp12) * 0.382683433f; // c6
-   z2 = tmp10 * 0.541196100f + z5; // c2-c6
-   z4 = tmp12 * 1.306562965f + z5; // c2+c6
-   z3 = tmp11 * 0.707106781f; // c4
-
-   z11 = tmp7 + z3;      // phase 5
-   z13 = tmp7 - z3;
-
-   *d5p = z13 + z2;         // phase 6
-   *d3p = z13 - z2;
-   *d1p = z11 + z4;
-   *d7p = z11 - z4;
-
-   *d0p = d0;  *d2p = d2;  *d4p = d4;  *d6p = d6;
-}
-
-static void stbiw__jpg_calcBits(int val, unsigned short bits[2]) {
-   int tmp1 = val < 0 ? -val : val;
-   val = val < 0 ? val-1 : val;
-   bits[1] = 1;
-   while(tmp1 >>= 1) {
-      ++bits[1];
-   }
-   bits[0] = val & ((1<<bits[1])-1);
-}
-
-static int stbiw__jpg_processDU(stbi__write_context *s, int *bitBuf, int *bitCnt, float *CDU, int du_stride, float *fdtbl, int DC, const unsigned short HTDC[256][2], const unsigned short HTAC[256][2]) {
-   const unsigned short EOB[2] = { HTAC[0x00][0], HTAC[0x00][1] };
-   const unsigned short M16zeroes[2] = { HTAC[0xF0][0], HTAC[0xF0][1] };
-   int dataOff, i, j, n, diff, end0pos, x, y;
-   int DU[64];
-
-   // DCT rows
-   for(dataOff=0, n=du_stride*8; dataOff<n; dataOff+=du_stride) {
-      stbiw__jpg_DCT(&CDU[dataOff], &CDU[dataOff+1], &CDU[dataOff+2], &CDU[dataOff+3], &CDU[dataOff+4], &CDU[dataOff+5], &CDU[dataOff+6], &CDU[dataOff+7]);
-   }
-   // DCT columns
-   for(dataOff=0; dataOff<8; ++dataOff) {
-      stbiw__jpg_DCT(&CDU[dataOff], &CDU[dataOff+du_stride], &CDU[dataOff+du_stride*2], &CDU[dataOff+du_stride*3], &CDU[dataOff+du_stride*4],
-                     &CDU[dataOff+du_stride*5], &CDU[dataOff+du_stride*6], &CDU[dataOff+du_stride*7]);
-   }
-   // Quantize/descale/zigzag the coefficients
-   for(y = 0, j=0; y < 8; ++y) {
-      for(x = 0; x < 8; ++x,++j) {
-         float v;
-         i = y*du_stride+x;
-         v = CDU[i]*fdtbl[j];
-         // DU[stbiw__jpg_ZigZag[j]] = (int)(v < 0 ? ceilf(v - 0.5f) : floorf(v + 0.5f));
-         // ceilf() and floorf() are C99, not C89, but I /think/ they're not needed here anyway?
-         DU[stbiw__jpg_ZigZag[j]] = (int)(v < 0 ? v - 0.5f : v + 0.5f);
-      }
-   }
-
-   // Encode DC
-   diff = DU[0] - DC;
-   if (diff == 0) {
-      stbiw__jpg_writeBits(s, bitBuf, bitCnt, HTDC[0]);
-   } else {
-      unsigned short bits[2];
-      stbiw__jpg_calcBits(diff, bits);
-      stbiw__jpg_writeBits(s, bitBuf, bitCnt, HTDC[bits[1]]);
-      stbiw__jpg_writeBits(s, bitBuf, bitCnt, bits);
-   }
-   // Encode ACs
-   end0pos = 63;
-   for(; (end0pos>0)&&(DU[end0pos]==0); --end0pos) {
-   }
-   // end0pos = first element in reverse order !=0
-   if(end0pos == 0) {
-      stbiw__jpg_writeBits(s, bitBuf, bitCnt, EOB);
-      return DU[0];
-   }
-   for(i = 1; i <= end0pos; ++i) {
-      int startpos = i;
-      int nrzeroes;
-      unsigned short bits[2];
-      for (; DU[i]==0 && i<=end0pos; ++i) {
-      }
-      nrzeroes = i-startpos;
-      if ( nrzeroes >= 16 ) {
-         int lng = nrzeroes>>4;
-         int nrmarker;
-         for (nrmarker=1; nrmarker <= lng; ++nrmarker)
-            stbiw__jpg_writeBits(s, bitBuf, bitCnt, M16zeroes);
-         nrzeroes &= 15;
-      }
-      stbiw__jpg_calcBits(DU[i], bits);
-      stbiw__jpg_writeBits(s, bitBuf, bitCnt, HTAC[(nrzeroes<<4)+bits[1]]);
-      stbiw__jpg_writeBits(s, bitBuf, bitCnt, bits);
-   }
-   if(end0pos != 63) {
-      stbiw__jpg_writeBits(s, bitBuf, bitCnt, EOB);
-   }
-   return DU[0];
-}
-
-static int stbi_write_jpg_core(stbi__write_context *s, int width, int height, int comp, const void* data, int quality) {
-   // Constants that don't pollute global namespace
-   static const unsigned char std_dc_luminance_nrcodes[] = {0,0,1,5,1,1,1,1,1,1,0,0,0,0,0,0,0};
-   static const unsigned char std_dc_luminance_values[] = {0,1,2,3,4,5,6,7,8,9,10,11};
-   static const unsigned char std_ac_luminance_nrcodes[] = {0,0,2,1,3,3,2,4,3,5,5,4,4,0,0,1,0x7d};
-   static const unsigned char std_ac_luminance_values[] = {
-      0x01,0x02,0x03,0x00,0x04,0x11,0x05,0x12,0x21,0x31,0x41,0x06,0x13,0x51,0x61,0x07,0x22,0x71,0x14,0x32,0x81,0x91,0xa1,0x08,
-      0x23,0x42,0xb1,0xc1,0x15,0x52,0xd1,0xf0,0x24,0x33,0x62,0x72,0x82,0x09,0x0a,0x16,0x17,0x18,0x19,0x1a,0x25,0x26,0x27,0x28,
-      0x29,0x2a,0x34,0x35,0x36,0x37,0x38,0x39,0x3a,0x43,0x44,0x45,0x46,0x47,0x48,0x49,0x4a,0x53,0x54,0x55,0x56,0x57,0x58,0x59,
-      0x5a,0x63,0x64,0x65,0x66,0x67,0x68,0x69,0x6a,0x73,0x74,0x75,0x76,0x77,0x78,0x79,0x7a,0x83,0x84,0x85,0x86,0x87,0x88,0x89,
-      0x8a,0x92,0x93,0x94,0x95,0x96,0x97,0x98,0x99,0x9a,0xa2,0xa3,0xa4,0xa5,0xa6,0xa7,0xa8,0xa9,0xaa,0xb2,0xb3,0xb4,0xb5,0xb6,
-      0xb7,0xb8,0xb9,0xba,0xc2,0xc3,0xc4,0xc5,0xc6,0xc7,0xc8,0xc9,0xca,0xd2,0xd3,0xd4,0xd5,0xd6,0xd7,0xd8,0xd9,0xda,0xe1,0xe2,
-      0xe3,0xe4,0xe5,0xe6,0xe7,0xe8,0xe9,0xea,0xf1,0xf2,0xf3,0xf4,0xf5,0xf6,0xf7,0xf8,0xf9,0xfa
-   };
-   static const unsigned char std_dc_chrominance_nrcodes[] = {0,0,3,1,1,1,1,1,1,1,1,1,0,0,0,0,0};
-   static const unsigned char std_dc_chrominance_values[] = {0,1,2,3,4,5,6,7,8,9,10,11};
-   static const unsigned char std_ac_chrominance_nrcodes[] = {0,0,2,1,2,4,4,3,4,7,5,4,4,0,1,2,0x77};
-   static const unsigned char std_ac_chrominance_values[] = {
-      0x00,0x01,0x02,0x03,0x11,0x04,0x05,0x21,0x31,0x06,0x12,0x41,0x51,0x07,0x61,0x71,0x13,0x22,0x32,0x81,0x08,0x14,0x42,0x91,
-      0xa1,0xb1,0xc1,0x09,0x23,0x33,0x52,0xf0,0x15,0x62,0x72,0xd1,0x0a,0x16,0x24,0x34,0xe1,0x25,0xf1,0x17,0x18,0x19,0x1a,0x26,
-      0x27,0x28,0x29,0x2a,0x35,0x36,0x37,0x38,0x39,0x3a,0x43,0x44,0x45,0x46,0x47,0x48,0x49,0x4a,0x53,0x54,0x55,0x56,0x57,0x58,
-      0x59,0x5a,0x63,0x64,0x65,0x66,0x67,0x68,0x69,0x6a,0x73,0x74,0x75,0x76,0x77,0x78,0x79,0x7a,0x82,0x83,0x84,0x85,0x86,0x87,
-      0x88,0x89,0x8a,0x92,0x93,0x94,0x95,0x96,0x97,0x98,0x99,0x9a,0xa2,0xa3,0xa4,0xa5,0xa6,0xa7,0xa8,0xa9,0xaa,0xb2,0xb3,0xb4,
-      0xb5,0xb6,0xb7,0xb8,0xb9,0xba,0xc2,0xc3,0xc4,0xc5,0xc6,0xc7,0xc8,0xc9,0xca,0xd2,0xd3,0xd4,0xd5,0xd6,0xd7,0xd8,0xd9,0xda,
-      0xe2,0xe3,0xe4,0xe5,0xe6,0xe7,0xe8,0xe9,0xea,0xf2,0xf3,0xf4,0xf5,0xf6,0xf7,0xf8,0xf9,0xfa
-   };
-   // Huffman tables
-   static const unsigned short YDC_HT[256][2] = { {0,2},{2,3},{3,3},{4,3},{5,3},{6,3},{14,4},{30,5},{62,6},{126,7},{254,8},{510,9}};
-   static const unsigned short UVDC_HT[256][2] = { {0,2},{1,2},{2,2},{6,3},{14,4},{30,5},{62,6},{126,7},{254,8},{510,9},{1022,10},{2046,11}};
-   static const unsigned short YAC_HT[256][2] = {
-      {10,4},{0,2},{1,2},{4,3},{11,4},{26,5},{120,7},{248,8},{1014,10},{65410,16},{65411,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {12,4},{27,5},{121,7},{502,9},{2038,11},{65412,16},{65413,16},{65414,16},{65415,16},{65416,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {28,5},{249,8},{1015,10},{4084,12},{65417,16},{65418,16},{65419,16},{65420,16},{65421,16},{65422,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {58,6},{503,9},{4085,12},{65423,16},{65424,16},{65425,16},{65426,16},{65427,16},{65428,16},{65429,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {59,6},{1016,10},{65430,16},{65431,16},{65432,16},{65433,16},{65434,16},{65435,16},{65436,16},{65437,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {122,7},{2039,11},{65438,16},{65439,16},{65440,16},{65441,16},{65442,16},{65443,16},{65444,16},{65445,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {123,7},{4086,12},{65446,16},{65447,16},{65448,16},{65449,16},{65450,16},{65451,16},{65452,16},{65453,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {250,8},{4087,12},{65454,16},{65455,16},{65456,16},{65457,16},{65458,16},{65459,16},{65460,16},{65461,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {504,9},{32704,15},{65462,16},{65463,16},{65464,16},{65465,16},{65466,16},{65467,16},{65468,16},{65469,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {505,9},{65470,16},{65471,16},{65472,16},{65473,16},{65474,16},{65475,16},{65476,16},{65477,16},{65478,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {506,9},{65479,16},{65480,16},{65481,16},{65482,16},{65483,16},{65484,16},{65485,16},{65486,16},{65487,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {1017,10},{65488,16},{65489,16},{65490,16},{65491,16},{65492,16},{65493,16},{65494,16},{65495,16},{65496,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {1018,10},{65497,16},{65498,16},{65499,16},{65500,16},{65501,16},{65502,16},{65503,16},{65504,16},{65505,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {2040,11},{65506,16},{65507,16},{65508,16},{65509,16},{65510,16},{65511,16},{65512,16},{65513,16},{65514,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {65515,16},{65516,16},{65517,16},{65518,16},{65519,16},{65520,16},{65521,16},{65522,16},{65523,16},{65524,16},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {2041,11},{65525,16},{65526,16},{65527,16},{65528,16},{65529,16},{65530,16},{65531,16},{65532,16},{65533,16},{65534,16},{0,0},{0,0},{0,0},{0,0},{0,0}
-   };
-   static const unsigned short UVAC_HT[256][2] = {
-      {0,2},{1,2},{4,3},{10,4},{24,5},{25,5},{56,6},{120,7},{500,9},{1014,10},{4084,12},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {11,4},{57,6},{246,8},{501,9},{2038,11},{4085,12},{65416,16},{65417,16},{65418,16},{65419,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {26,5},{247,8},{1015,10},{4086,12},{32706,15},{65420,16},{65421,16},{65422,16},{65423,16},{65424,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {27,5},{248,8},{1016,10},{4087,12},{65425,16},{65426,16},{65427,16},{65428,16},{65429,16},{65430,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {58,6},{502,9},{65431,16},{65432,16},{65433,16},{65434,16},{65435,16},{65436,16},{65437,16},{65438,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {59,6},{1017,10},{65439,16},{65440,16},{65441,16},{65442,16},{65443,16},{65444,16},{65445,16},{65446,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {121,7},{2039,11},{65447,16},{65448,16},{65449,16},{65450,16},{65451,16},{65452,16},{65453,16},{65454,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {122,7},{2040,11},{65455,16},{65456,16},{65457,16},{65458,16},{65459,16},{65460,16},{65461,16},{65462,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {249,8},{65463,16},{65464,16},{65465,16},{65466,16},{65467,16},{65468,16},{65469,16},{65470,16},{65471,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {503,9},{65472,16},{65473,16},{65474,16},{65475,16},{65476,16},{65477,16},{65478,16},{65479,16},{65480,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {504,9},{65481,16},{65482,16},{65483,16},{65484,16},{65485,16},{65486,16},{65487,16},{65488,16},{65489,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {505,9},{65490,16},{65491,16},{65492,16},{65493,16},{65494,16},{65495,16},{65496,16},{65497,16},{65498,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {506,9},{65499,16},{65500,16},{65501,16},{65502,16},{65503,16},{65504,16},{65505,16},{65506,16},{65507,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {2041,11},{65508,16},{65509,16},{65510,16},{65511,16},{65512,16},{65513,16},{65514,16},{65515,16},{65516,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {16352,14},{65517,16},{65518,16},{65519,16},{65520,16},{65521,16},{65522,16},{65523,16},{65524,16},{65525,16},{0,0},{0,0},{0,0},{0,0},{0,0},
-      {1018,10},{32707,15},{65526,16},{65527,16},{65528,16},{65529,16},{65530,16},{65531,16},{65532,16},{65533,16},{65534,16},{0,0},{0,0},{0,0},{0,0},{0,0}
-   };
-   static const int YQT[] = {16,11,10,16,24,40,51,61,12,12,14,19,26,58,60,55,14,13,16,24,40,57,69,56,14,17,22,29,51,87,80,62,18,22,
-                             37,56,68,109,103,77,24,35,55,64,81,104,113,92,49,64,78,87,103,121,120,101,72,92,95,98,112,100,103,99};
-   static const int UVQT[] = {17,18,24,47,99,99,99,99,18,21,26,66,99,99,99,99,24,26,56,99,99,99,99,99,47,66,99,99,99,99,99,99,
-                              99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99};
-   static const float aasf[] = { 1.0f * 2.828427125f, 1.387039845f * 2.828427125f, 1.306562965f * 2.828427125f, 1.175875602f * 2.828427125f,
-                                 1.0f * 2.828427125f, 0.785694958f * 2.828427125f, 0.541196100f * 2.828427125f, 0.275899379f * 2.828427125f };
-
-   int row, col, i, k, subsample;
-   float fdtbl_Y[64], fdtbl_UV[64];
-   unsigned char YTable[64], UVTable[64];
-
-   if(!data || !width || !height || comp > 4 || comp < 1) {
-      return 0;
-   }
-
-   quality = quality ? quality : 90;
-   subsample = quality <= 90 ? 1 : 0;
-   quality = quality < 1 ? 1 : quality > 100 ? 100 : quality;
-   quality = quality < 50 ? 5000 / quality : 200 - quality * 2;
-
-   for(i = 0; i < 64; ++i) {
-      int uvti, yti = (YQT[i]*quality+50)/100;
-      YTable[stbiw__jpg_ZigZag[i]] = (unsigned char) (yti < 1 ? 1 : yti > 255 ? 255 : yti);
-      uvti = (UVQT[i]*quality+50)/100;
-      UVTable[stbiw__jpg_ZigZag[i]] = (unsigned char) (uvti < 1 ? 1 : uvti > 255 ? 255 : uvti);
-   }
-
-   for(row = 0, k = 0; row < 8; ++row) {
-      for(col = 0; col < 8; ++col, ++k) {
-         fdtbl_Y[k]  = 1 / (YTable [stbiw__jpg_ZigZag[k]] * aasf[row] * aasf[col]);
-         fdtbl_UV[k] = 1 / (UVTable[stbiw__jpg_ZigZag[k]] * aasf[row] * aasf[col]);
-      }
-   }
-
-   // Write Headers
-   {
-      static const unsigned char head0[] = { 0xFF,0xD8,0xFF,0xE0,0,0x10,'J','F','I','F',0,1,1,0,0,1,0,1,0,0,0xFF,0xDB,0,0x84,0 };
-      static const unsigned char head2[] = { 0xFF,0xDA,0,0xC,3,1,0,2,0x11,3,0x11,0,0x3F,0 };
-      const unsigned char head1[] = { 0xFF,0xC0,0,0x11,8,(unsigned char)(height>>8),STBIW_UCHAR(height),(unsigned char)(width>>8),STBIW_UCHAR(width),
-                                      3,1,(unsigned char)(subsample?0x22:0x11),0,2,0x11,1,3,0x11,1,0xFF,0xC4,0x01,0xA2,0 };
-      s->func(s->context, (void*)head0, sizeof(head0));
-      s->func(s->context, (void*)YTable, sizeof(YTable));
-      stbiw__putc(s, 1);
-      s->func(s->context, UVTable, sizeof(UVTable));
-      s->func(s->context, (void*)head1, sizeof(head1));
-      s->func(s->context, (void*)(std_dc_luminance_nrcodes+1), sizeof(std_dc_luminance_nrcodes)-1);
-      s->func(s->context, (void*)std_dc_luminance_values, sizeof(std_dc_luminance_values));
-      stbiw__putc(s, 0x10); // HTYACinfo
-      s->func(s->context, (void*)(std_ac_luminance_nrcodes+1), sizeof(std_ac_luminance_nrcodes)-1);
-      s->func(s->context, (void*)std_ac_luminance_values, sizeof(std_ac_luminance_values));
-      stbiw__putc(s, 1); // HTUDCinfo
-      s->func(s->context, (void*)(std_dc_chrominance_nrcodes+1), sizeof(std_dc_chrominance_nrcodes)-1);
-      s->func(s->context, (void*)std_dc_chrominance_values, sizeof(std_dc_chrominance_values));
-      stbiw__putc(s, 0x11); // HTUACinfo
-      s->func(s->context, (void*)(std_ac_chrominance_nrcodes+1), sizeof(std_ac_chrominance_nrcodes)-1);
-      s->func(s->context, (void*)std_ac_chrominance_values, sizeof(std_ac_chrominance_values));
-      s->func(s->context, (void*)head2, sizeof(head2));
-   }
-
-   // Encode 8x8 macroblocks
-   {
-      static const unsigned short fillBits[] = {0x7F, 7};
-      int DCY=0, DCU=0, DCV=0;
-      int bitBuf=0, bitCnt=0;
-      // comp == 2 is grey+alpha (alpha is ignored)
-      int ofsG = comp > 2 ? 1 : 0, ofsB = comp > 2 ? 2 : 0;
-      const unsigned char *dataR = (const unsigned char *)data;
-      const unsigned char *dataG = dataR + ofsG;
-      const unsigned char *dataB = dataR + ofsB;
-      int x, y, pos;
-      if(subsample) {
-         for(y = 0; y < height; y += 16) {
-            for(x = 0; x < width; x += 16) {
-               float Y[256], U[256], V[256];
-               for(row = y, pos = 0; row < y+16; ++row) {
-                  // row >= height => use last input row
-                  int clamped_row = (row < height) ? row : height - 1;
-                  int base_p = (stbi__flip_vertically_on_write ? (height-1-clamped_row) : clamped_row)*width*comp;
-                  for(col = x; col < x+16; ++col, ++pos) {
-                     // if col >= width => use pixel from last input column
-                     int p = base_p + ((col < width) ? col : (width-1))*comp;
-                     float r = dataR[p], g = dataG[p], b = dataB[p];
-                     Y[pos]= +0.29900f*r + 0.58700f*g + 0.11400f*b - 128;
-                     U[pos]= -0.16874f*r - 0.33126f*g + 0.50000f*b;
-                     V[pos]= +0.50000f*r - 0.41869f*g - 0.08131f*b;
-                  }
-               }
-               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+0,   16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
-               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+8,   16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
-               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+128, 16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
-               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+136, 16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
-
-               // subsample U,V
-               {
-                  float subU[64], subV[64];
-                  int yy, xx;
-                  for(yy = 0, pos = 0; yy < 8; ++yy) {
-                     for(xx = 0; xx < 8; ++xx, ++pos) {
-                        int j = yy*32+xx*2;
-                        subU[pos] = (U[j+0] + U[j+1] + U[j+16] + U[j+17]) * 0.25f;
-                        subV[pos] = (V[j+0] + V[j+1] + V[j+16] + V[j+17]) * 0.25f;
-                     }
-                  }
-                  DCU = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, subU, 8, fdtbl_UV, DCU, UVDC_HT, UVAC_HT);
-                  DCV = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, subV, 8, fdtbl_UV, DCV, UVDC_HT, UVAC_HT);
-               }
-            }
-         }
-      } else {
-         for(y = 0; y < height; y += 8) {
-            for(x = 0; x < width; x += 8) {
-               float Y[64], U[64], V[64];
-               for(row = y, pos = 0; row < y+8; ++row) {
-                  // row >= height => use last input row
-                  int clamped_row = (row < height) ? row : height - 1;
-                  int base_p = (stbi__flip_vertically_on_write ? (height-1-clamped_row) : clamped_row)*width*comp;
-                  for(col = x; col < x+8; ++col, ++pos) {
-                     // if col >= width => use pixel from last input column
-                     int p = base_p + ((col < width) ? col : (width-1))*comp;
-                     float r = dataR[p], g = dataG[p], b = dataB[p];
-                     Y[pos]= +0.29900f*r + 0.58700f*g + 0.11400f*b - 128;
-                     U[pos]= -0.16874f*r - 0.33126f*g + 0.50000f*b;
-                     V[pos]= +0.50000f*r - 0.41869f*g - 0.08131f*b;
-                  }
-               }
-
-               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y, 8, fdtbl_Y,  DCY, YDC_HT, YAC_HT);
-               DCU = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, U, 8, fdtbl_UV, DCU, UVDC_HT, UVAC_HT);
-               DCV = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, V, 8, fdtbl_UV, DCV, UVDC_HT, UVAC_HT);
-            }
-         }
-      }
-
-      // Do the bit alignment of the EOI marker
-      stbiw__jpg_writeBits(s, &bitBuf, &bitCnt, fillBits);
-   }
-
-   // EOI
-   stbiw__putc(s, 0xFF);
-   stbiw__putc(s, 0xD9);
-
-   return 1;
-}
-
-STBIWDEF int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int quality)
-{
-   stbi__write_context s = { 0 };
-   stbi__start_write_callbacks(&s, func, context);
-   return stbi_write_jpg_core(&s, x, y, comp, (void *) data, quality);
-}
-
-
-#ifndef STBI_WRITE_NO_STDIO
-STBIWDEF int stbi_write_jpg(char const *filename, int x, int y, int comp, const void *data, int quality)
-{
-   stbi__write_context s = { 0 };
-   if (stbi__start_write_file(&s,filename)) {
-      int r = stbi_write_jpg_core(&s, x, y, comp, data, quality);
-      stbi__end_write_file(&s);
-      return r;
-   } else
-      return 0;
-}
-#endif
-
-#endif // STB_IMAGE_WRITE_IMPLEMENTATION
-
-/* Revision history
-      1.16  (2021-07-11)
-             make Deflate code emit uncompressed blocks when it would otherwise expand
-             support writing BMPs with alpha channel
-      1.15  (2020-07-13) unknown
-      1.14  (2020-02-02) updated JPEG writer to downsample chroma channels
-      1.13
-      1.12
-      1.11  (2019-08-11)
-
-      1.10  (2019-02-07)
-             support utf8 filenames in Windows; fix warnings and platform ifdefs
-      1.09  (2018-02-11)
-             fix typo in zlib quality API, improve STB_I_W_STATIC in C++
-      1.08  (2018-01-29)
-             add stbi__flip_vertically_on_write, external zlib, zlib quality, choose PNG filter
-      1.07  (2017-07-24)
-             doc fix
-      1.06 (2017-07-23)
-             writing JPEG (using Jon Olick's code)
-      1.05   ???
-      1.04 (2017-03-03)
-             monochrome BMP expansion
-      1.03   ???
-      1.02 (2016-04-02)
-             avoid allocating large structures on the stack
-      1.01 (2016-01-16)
-             STBIW_REALLOC_SIZED: support allocators with no realloc support
-             avoid race-condition in crc initialization
-             minor compile issues
-      1.00 (2015-09-14)
-             installable file IO function
-      0.99 (2015-09-13)
-             warning fixes; TGA rle support
-      0.98 (2015-04-08)
-             added STBIW_MALLOC, STBIW_ASSERT etc
-      0.97 (2015-01-18)
-             fixed HDR asserts, rewrote HDR rle logic
-      0.96 (2015-01-17)
-             add HDR output
-             fix monochrome BMP
-      0.95 (2014-08-17)
-             add monochrome TGA output
-      0.94 (2014-05-31)
-             rename private functions to avoid conflicts with stb_image.h
-      0.93 (2014-05-27)
-             warning fixes
-      0.92 (2010-08-01)
-             casts to unsigned char to fix warnings
-      0.91 (2010-07-17)
-             first public release
-      0.90   first internal release
-*/
-
-/*
-------------------------------------------------------------------------------
-This software is available under 2 licenses -- choose whichever you prefer.
-------------------------------------------------------------------------------
-ALTERNATIVE A - MIT License
-Copyright (c) 2017 Sean Barrett
-Permission is hereby granted, free of charge, to any person obtaining a copy of
-this software and associated documentation files (the "Software"), to deal in
-the Software without restriction, including without limitation the rights to
-use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
-of the Software, and to permit persons to whom the Software is furnished to do
-so, subject to the following conditions:
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
-------------------------------------------------------------------------------
-ALTERNATIVE B - Public Domain (www.unlicense.org)
-This is free and unencumbered software released into the public domain.
-Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
-software, either in source code form or as a compiled binary, for any purpose,
-commercial or non-commercial, and by any means.
-In jurisdictions that recognize copyright laws, the author or authors of this
-software dedicate any and all copyright interest in the software to the public
-domain. We make this dedication for the benefit of the public at large and to
-the detriment of our heirs and successors. We intend this dedication to be an
-overt act of relinquishment in perpetuity of all present and future rights to
-this software under copyright law.
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
-ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
-WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
-------------------------------------------------------------------------------
-*/
diff --git a/projects/llm_framework/include/sherpa/sherpa-ncnn/vulkan_header_fix.h b/projects/llm_framework/include/sherpa/sherpa-ncnn/vulkan_header_fix.h
deleted file mode 100644
index 1402bba2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-ncnn/vulkan_header_fix.h
+++ /dev/null
@@ -1,1222 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-#ifndef NCNN_VULKAN_HEADER_FIX_H
-#define NCNN_VULKAN_HEADER_FIX_H
-
-// #include "platform.h"
-
-// This header contains new structure and function declearation to fix build with old vulkan sdk
-
-#ifndef VK_KHR_maintenance1
-#define VK_KHR_maintenance1 1
-typedef VkFlags VkCommandPoolTrimFlags;
-typedef VkCommandPoolTrimFlags VkCommandPoolTrimFlagsKHR;
-typedef void(VKAPI_PTR* PFN_vkTrimCommandPool)(VkDevice device, VkCommandPool commandPool, VkCommandPoolTrimFlags flags);
-typedef PFN_vkTrimCommandPool PFN_vkTrimCommandPoolKHR;
-#endif // VK_KHR_maintenance1
-
-#ifndef VK_KHR_get_physical_device_properties2
-#define VK_KHR_get_physical_device_properties2                    1
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2              (VkStructureType)1000059000
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2            (VkStructureType)1000059001
-#define VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_2                     (VkStructureType)1000059002
-#define VK_STRUCTURE_TYPE_IMAGE_FORMAT_PROPERTIES_2               (VkStructureType)1000059003
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2     (VkStructureType)1000059004
-#define VK_STRUCTURE_TYPE_QUEUE_FAMILY_PROPERTIES_2               (VkStructureType)1000059005
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_PROPERTIES_2     (VkStructureType)1000059006
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2_KHR          VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR        VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2
-#define VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_2_KHR                 VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_2
-#define VK_STRUCTURE_TYPE_IMAGE_FORMAT_PROPERTIES_2_KHR           VK_STRUCTURE_TYPE_IMAGE_FORMAT_PROPERTIES_2
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2_KHR VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2
-#define VK_STRUCTURE_TYPE_QUEUE_FAMILY_PROPERTIES_2_KHR           VK_STRUCTURE_TYPE_QUEUE_FAMILY_PROPERTIES_2
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_PROPERTIES_2_KHR VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_PROPERTIES_2
-typedef struct VkPhysicalDeviceFeatures2
-{
-    VkStructureType sType;
-    void* pNext;
-    VkPhysicalDeviceFeatures features;
-} VkPhysicalDeviceFeatures2;
-typedef struct VkPhysicalDeviceProperties2
-{
-    VkStructureType sType;
-    void* pNext;
-    VkPhysicalDeviceProperties properties;
-} VkPhysicalDeviceProperties2;
-typedef struct VkFormatProperties2
-{
-    VkStructureType sType;
-    void* pNext;
-    VkFormatProperties formatProperties;
-} VkFormatProperties2;
-typedef struct VkImageFormatProperties2
-{
-    VkStructureType sType;
-    void* pNext;
-    VkImageFormatProperties imageFormatProperties;
-} VkImageFormatProperties2;
-typedef struct VkPhysicalDeviceImageFormatInfo2
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkFormat format;
-    VkImageType type;
-    VkImageTiling tiling;
-    VkImageUsageFlags usage;
-    VkImageCreateFlags flags;
-} VkPhysicalDeviceImageFormatInfo2;
-typedef struct VkQueueFamilyProperties2
-{
-    VkStructureType sType;
-    void* pNext;
-    VkQueueFamilyProperties queueFamilyProperties;
-} VkQueueFamilyProperties2;
-typedef struct VkPhysicalDeviceMemoryProperties2
-{
-    VkStructureType sType;
-    void* pNext;
-    VkPhysicalDeviceMemoryProperties memoryProperties;
-} VkPhysicalDeviceMemoryProperties2;
-typedef VkPhysicalDeviceFeatures2 VkPhysicalDeviceFeatures2KHR;
-typedef VkPhysicalDeviceProperties2 VkPhysicalDeviceProperties2KHR;
-typedef VkFormatProperties2 VkFormatProperties2KHR;
-typedef VkImageFormatProperties2 VkImageFormatProperties2KHR;
-typedef VkPhysicalDeviceImageFormatInfo2 VkPhysicalDeviceImageFormatInfo2KHR;
-typedef VkQueueFamilyProperties2 VkQueueFamilyProperties2KHR;
-typedef VkPhysicalDeviceMemoryProperties2 VkPhysicalDeviceMemoryProperties2KHR;
-typedef void(VKAPI_PTR* PFN_vkGetPhysicalDeviceFeatures2)(VkPhysicalDevice physicalDevice, VkPhysicalDeviceFeatures2* pFeatures);
-typedef void(VKAPI_PTR* PFN_vkGetPhysicalDeviceProperties2)(VkPhysicalDevice physicalDevice, VkPhysicalDeviceProperties2* pProperties);
-typedef void(VKAPI_PTR* PFN_vkGetPhysicalDeviceFormatProperties2)(VkPhysicalDevice physicalDevice, VkFormat format, VkFormatProperties2* pFormatProperties);
-typedef VkResult(VKAPI_PTR* PFN_vkGetPhysicalDeviceImageFormatProperties2)(VkPhysicalDevice physicalDevice, const VkPhysicalDeviceImageFormatInfo2* pImageFormatInfo, VkImageFormatProperties2* pImageFormatProperties);
-typedef void(VKAPI_PTR* PFN_vkGetPhysicalDeviceQueueFamilyProperties2)(VkPhysicalDevice physicalDevice, uint32_t* pQueueFamilyPropertyCount, VkQueueFamilyProperties2* pQueueFamilyProperties);
-typedef void(VKAPI_PTR* PFN_vkGetPhysicalDeviceMemoryProperties2)(VkPhysicalDevice physicalDevice, VkPhysicalDeviceMemoryProperties2* pMemoryProperties);
-typedef PFN_vkGetPhysicalDeviceFeatures2 PFN_vkGetPhysicalDeviceFeatures2KHR;
-typedef PFN_vkGetPhysicalDeviceProperties2 PFN_vkGetPhysicalDeviceProperties2KHR;
-typedef PFN_vkGetPhysicalDeviceFormatProperties2 PFN_vkGetPhysicalDeviceFormatProperties2KHR;
-typedef PFN_vkGetPhysicalDeviceImageFormatProperties2 PFN_vkGetPhysicalDeviceImageFormatProperties2KHR;
-typedef PFN_vkGetPhysicalDeviceQueueFamilyProperties2 PFN_vkGetPhysicalDeviceQueueFamilyProperties2KHR;
-typedef PFN_vkGetPhysicalDeviceMemoryProperties2 PFN_vkGetPhysicalDeviceMemoryProperties2KHR;
-#endif // VK_KHR_get_physical_device_properties2
-
-#ifndef VK_KHR_external_memory_capabilities
-#define VK_KHR_external_memory_capabilities                              1
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO     (VkStructureType)1000071000
-#define VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES               (VkStructureType)1000071001
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_BUFFER_INFO           (VkStructureType)1000071002
-#define VK_STRUCTURE_TYPE_EXTERNAL_BUFFER_PROPERTIES                     (VkStructureType)1000071003
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES                  (VkStructureType)1000071004
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO_KHR VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO
-#define VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES_KHR           VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_BUFFER_INFO_KHR       VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_BUFFER_INFO
-#define VK_STRUCTURE_TYPE_EXTERNAL_BUFFER_PROPERTIES_KHR                 VK_STRUCTURE_TYPE_EXTERNAL_BUFFER_PROPERTIES
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR              VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
-#define VK_LUID_SIZE                                                     8
-#define VK_LUID_SIZE_KHR                                                 VK_LUID_SIZE
-typedef enum VkExternalMemoryHandleTypeFlagBits
-{
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT = 0x00000001,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_BIT = 0x00000002,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT_BIT = 0x00000004,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_TEXTURE_BIT = 0x00000008,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_TEXTURE_KMT_BIT = 0x00000010,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_HEAP_BIT = 0x00000020,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE_BIT = 0x00000040,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT = 0x00000200,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID = 0x00000400,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT = 0x00000080,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT = 0x00000100,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_ZIRCON_VMO_BIT_FUCHSIA = 0x00000800,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_RDMA_ADDRESS_BIT_NV = 0x00001000,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_BIT_KHR = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_BIT,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT_BIT_KHR = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT_BIT,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_TEXTURE_BIT_KHR = VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_TEXTURE_BIT,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_TEXTURE_KMT_BIT_KHR = VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_TEXTURE_KMT_BIT,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_HEAP_BIT_KHR = VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_HEAP_BIT,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE_BIT_KHR = VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE_BIT,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkExternalMemoryHandleTypeFlagBits;
-typedef VkFlags VkExternalMemoryHandleTypeFlags;
-typedef enum VkExternalMemoryFeatureFlagBits
-{
-    VK_EXTERNAL_MEMORY_FEATURE_DEDICATED_ONLY_BIT = 0x00000001,
-    VK_EXTERNAL_MEMORY_FEATURE_EXPORTABLE_BIT = 0x00000002,
-    VK_EXTERNAL_MEMORY_FEATURE_IMPORTABLE_BIT = 0x00000004,
-    VK_EXTERNAL_MEMORY_FEATURE_DEDICATED_ONLY_BIT_KHR = VK_EXTERNAL_MEMORY_FEATURE_DEDICATED_ONLY_BIT,
-    VK_EXTERNAL_MEMORY_FEATURE_EXPORTABLE_BIT_KHR = VK_EXTERNAL_MEMORY_FEATURE_EXPORTABLE_BIT,
-    VK_EXTERNAL_MEMORY_FEATURE_IMPORTABLE_BIT_KHR = VK_EXTERNAL_MEMORY_FEATURE_IMPORTABLE_BIT,
-    VK_EXTERNAL_MEMORY_FEATURE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkExternalMemoryFeatureFlagBits;
-typedef VkFlags VkExternalMemoryFeatureFlags;
-typedef struct VkExternalMemoryProperties
-{
-    VkExternalMemoryFeatureFlags externalMemoryFeatures;
-    VkExternalMemoryHandleTypeFlags exportFromImportedHandleTypes;
-    VkExternalMemoryHandleTypeFlags compatibleHandleTypes;
-} VkExternalMemoryProperties;
-typedef struct VkPhysicalDeviceExternalImageFormatInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkExternalMemoryHandleTypeFlagBits handleType;
-} VkPhysicalDeviceExternalImageFormatInfo;
-typedef struct VkExternalImageFormatProperties
-{
-    VkStructureType sType;
-    void* pNext;
-    VkExternalMemoryProperties externalMemoryProperties;
-} VkExternalImageFormatProperties;
-typedef struct VkPhysicalDeviceExternalBufferInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkBufferCreateFlags flags;
-    VkBufferUsageFlags usage;
-    VkExternalMemoryHandleTypeFlagBits handleType;
-} VkPhysicalDeviceExternalBufferInfo;
-typedef struct VkExternalBufferProperties
-{
-    VkStructureType sType;
-    void* pNext;
-    VkExternalMemoryProperties externalMemoryProperties;
-} VkExternalBufferProperties;
-typedef struct VkPhysicalDeviceIDProperties
-{
-    VkStructureType sType;
-    void* pNext;
-    uint8_t deviceUUID[VK_UUID_SIZE];
-    uint8_t driverUUID[VK_UUID_SIZE];
-    uint8_t deviceLUID[VK_LUID_SIZE];
-    uint32_t deviceNodeMask;
-    VkBool32 deviceLUIDValid;
-} VkPhysicalDeviceIDProperties;
-typedef VkExternalMemoryHandleTypeFlags VkExternalMemoryHandleTypeFlagsKHR;
-typedef VkExternalMemoryHandleTypeFlagBits VkExternalMemoryHandleTypeFlagBitsKHR;
-typedef VkExternalMemoryFeatureFlags VkExternalMemoryFeatureFlagsKHR;
-typedef VkExternalMemoryFeatureFlagBits VkExternalMemoryFeatureFlagBitsKHR;
-typedef VkExternalMemoryProperties VkExternalMemoryPropertiesKHR;
-typedef VkPhysicalDeviceExternalImageFormatInfo VkPhysicalDeviceExternalImageFormatInfoKHR;
-typedef VkExternalImageFormatProperties VkExternalImageFormatPropertiesKHR;
-typedef VkPhysicalDeviceExternalBufferInfo VkPhysicalDeviceExternalBufferInfoKHR;
-typedef VkExternalBufferProperties VkExternalBufferPropertiesKHR;
-typedef VkPhysicalDeviceIDProperties VkPhysicalDeviceIDPropertiesKHR;
-typedef void(VKAPI_PTR* PFN_vkGetPhysicalDeviceExternalBufferProperties)(VkPhysicalDevice physicalDevice, const VkPhysicalDeviceExternalBufferInfo* pExternalBufferInfo, VkExternalBufferProperties* pExternalBufferProperties);
-typedef PFN_vkGetPhysicalDeviceExternalBufferProperties PFN_vkGetPhysicalDeviceExternalBufferPropertiesKHR;
-#endif // VK_KHR_external_memory_capabilities
-
-#ifndef VK_KHR_external_memory
-#define VK_KHR_external_memory                                   1
-#define VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_BUFFER_CREATE_INFO     (VkStructureType)1000072000
-#define VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO      (VkStructureType)1000072001
-#define VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO            (VkStructureType)1000072002
-#define VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_BUFFER_CREATE_INFO_KHR VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_BUFFER_CREATE_INFO
-#define VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO_KHR  VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO
-#define VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO_KHR        VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO
-#define VK_QUEUE_FAMILY_EXTERNAL_KHR                             (~0U - 1)
-typedef struct VkExternalMemoryImageCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkExternalMemoryHandleTypeFlags handleTypes;
-} VkExternalMemoryImageCreateInfo;
-typedef struct VkExternalMemoryBufferCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkExternalMemoryHandleTypeFlags handleTypes;
-} VkExternalMemoryBufferCreateInfo;
-typedef struct VkExportMemoryAllocateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkExternalMemoryHandleTypeFlags handleTypes;
-} VkExportMemoryAllocateInfo;
-typedef VkExternalMemoryImageCreateInfo VkExternalMemoryImageCreateInfoKHR;
-typedef VkExternalMemoryBufferCreateInfo VkExternalMemoryBufferCreateInfoKHR;
-typedef VkExportMemoryAllocateInfo VkExportMemoryAllocateInfoKHR;
-#endif // VK_KHR_external_memory
-
-#ifndef VK_KHR_descriptor_update_template
-#define VK_KHR_descriptor_update_template                            1
-#define VK_STRUCTURE_TYPE_DESCRIPTOR_UPDATE_TEMPLATE_CREATE_INFO     (VkStructureType)1000085000
-#define VK_STRUCTURE_TYPE_DESCRIPTOR_UPDATE_TEMPLATE_CREATE_INFO_KHR VK_STRUCTURE_TYPE_DESCRIPTOR_UPDATE_TEMPLATE_CREATE_INFO
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkDescriptorUpdateTemplate)
-typedef VkDescriptorUpdateTemplate VkDescriptorUpdateTemplateKHR;
-typedef enum VkDescriptorUpdateTemplateType
-{
-    VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_DESCRIPTOR_SET = 0,
-    VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_PUSH_DESCRIPTORS_KHR = 1,
-    VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_DESCRIPTOR_SET_KHR = VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_DESCRIPTOR_SET,
-    VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_MAX_ENUM = 0x7FFFFFFF
-} VkDescriptorUpdateTemplateType;
-typedef VkFlags VkDescriptorUpdateTemplateCreateFlags;
-typedef struct VkDescriptorUpdateTemplateEntry
-{
-    uint32_t dstBinding;
-    uint32_t dstArrayElement;
-    uint32_t descriptorCount;
-    VkDescriptorType descriptorType;
-    size_t offset;
-    size_t stride;
-} VkDescriptorUpdateTemplateEntry;
-typedef struct VkDescriptorUpdateTemplateCreateInfo
-{
-    VkStructureType sType;
-    void* pNext;
-    VkDescriptorUpdateTemplateCreateFlags flags;
-    uint32_t descriptorUpdateEntryCount;
-    const VkDescriptorUpdateTemplateEntry* pDescriptorUpdateEntries;
-    VkDescriptorUpdateTemplateType templateType;
-    VkDescriptorSetLayout descriptorSetLayout;
-    VkPipelineBindPoint pipelineBindPoint;
-    VkPipelineLayout pipelineLayout;
-    uint32_t set;
-} VkDescriptorUpdateTemplateCreateInfo;
-typedef VkDescriptorUpdateTemplateType VkDescriptorUpdateTemplateTypeKHR;
-typedef VkDescriptorUpdateTemplateCreateFlags VkDescriptorUpdateTemplateCreateFlagsKHR;
-typedef VkDescriptorUpdateTemplateEntry VkDescriptorUpdateTemplateEntryKHR;
-typedef VkDescriptorUpdateTemplateCreateInfo VkDescriptorUpdateTemplateCreateInfoKHR;
-typedef VkResult(VKAPI_PTR* PFN_vkCreateDescriptorUpdateTemplate)(VkDevice device, const VkDescriptorUpdateTemplateCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkDescriptorUpdateTemplate* pDescriptorUpdateTemplate);
-typedef void(VKAPI_PTR* PFN_vkDestroyDescriptorUpdateTemplate)(VkDevice device, VkDescriptorUpdateTemplate descriptorUpdateTemplate, const VkAllocationCallbacks* pAllocator);
-typedef void(VKAPI_PTR* PFN_vkUpdateDescriptorSetWithTemplate)(VkDevice device, VkDescriptorSet descriptorSet, VkDescriptorUpdateTemplate descriptorUpdateTemplate, const void* pData);
-typedef PFN_vkCreateDescriptorUpdateTemplate PFN_vkCreateDescriptorUpdateTemplateKHR;
-typedef PFN_vkDestroyDescriptorUpdateTemplate PFN_vkDestroyDescriptorUpdateTemplateKHR;
-typedef PFN_vkUpdateDescriptorSetWithTemplate PFN_vkUpdateDescriptorSetWithTemplateKHR;
-#endif // VK_KHR_descriptor_update_template
-
-#ifndef VK_KHR_push_descriptor
-#define VK_KHR_push_descriptor                                           1
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR (VkStructureType)1000080000
-typedef struct VkPhysicalDevicePushDescriptorPropertiesKHR
-{
-    VkStructureType sType;
-    void* pNext;
-    uint32_t maxPushDescriptors;
-} VkPhysicalDevicePushDescriptorPropertiesKHR;
-typedef void(VKAPI_PTR* PFN_vkCmdPushDescriptorSetKHR)(VkCommandBuffer commandBuffer, VkPipelineBindPoint pipelineBindPoint, VkPipelineLayout layout, uint32_t set, uint32_t descriptorWriteCount, const VkWriteDescriptorSet* pDescriptorWrites);
-typedef void(VKAPI_PTR* PFN_vkCmdPushDescriptorSetWithTemplateKHR)(VkCommandBuffer commandBuffer, VkDescriptorUpdateTemplate descriptorUpdateTemplate, VkPipelineLayout layout, uint32_t set, const void* pData);
-#endif // VK_KHR_push_descriptor
-
-#ifndef VK_KHR_get_surface_capabilities2
-#define VK_KHR_get_surface_capabilities2                     1
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SURFACE_INFO_2_KHR (VkStructureType)1000119000
-#define VK_STRUCTURE_TYPE_SURFACE_CAPABILITIES_2_KHR         (VkStructureType)1000119001
-#define VK_STRUCTURE_TYPE_SURFACE_FORMAT_2_KHR               (VkStructureType)1000119002
-typedef struct VkPhysicalDeviceSurfaceInfo2KHR
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkSurfaceKHR surface;
-} VkPhysicalDeviceSurfaceInfo2KHR;
-typedef struct VkSurfaceCapabilities2KHR
-{
-    VkStructureType sType;
-    void* pNext;
-    VkSurfaceCapabilitiesKHR surfaceCapabilities;
-} VkSurfaceCapabilities2KHR;
-typedef struct VkSurfaceFormat2KHR
-{
-    VkStructureType sType;
-    void* pNext;
-    VkSurfaceFormatKHR surfaceFormat;
-} VkSurfaceFormat2KHR;
-typedef VkResult(VKAPI_PTR* PFN_vkGetPhysicalDeviceSurfaceCapabilities2KHR)(VkPhysicalDevice physicalDevice, const VkPhysicalDeviceSurfaceInfo2KHR* pSurfaceInfo, VkSurfaceCapabilities2KHR* pSurfaceCapabilities);
-typedef VkResult(VKAPI_PTR* PFN_vkGetPhysicalDeviceSurfaceFormats2KHR)(VkPhysicalDevice physicalDevice, const VkPhysicalDeviceSurfaceInfo2KHR* pSurfaceInfo, uint32_t* pSurfaceFormatCount, VkSurfaceFormat2KHR* pSurfaceFormats);
-#endif // VK_KHR_get_surface_capabilities2
-
-#ifndef VK_KHR_external_memory_capabilities
-#define VK_KHR_external_memory_capabilities                              1
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO     (VkStructureType)1000071000
-#define VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES               (VkStructureType)1000071001
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_BUFFER_INFO           (VkStructureType)1000071002
-#define VK_STRUCTURE_TYPE_EXTERNAL_BUFFER_PROPERTIES                     (VkStructureType)1000071003
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES                  (VkStructureType)1000071004
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO_KHR VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO
-#define VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES_KHR           VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_BUFFER_INFO_KHR       VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_BUFFER_INFO
-#define VK_STRUCTURE_TYPE_EXTERNAL_BUFFER_PROPERTIES_KHR                 VK_STRUCTURE_TYPE_EXTERNAL_BUFFER_PROPERTIES
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR              VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
-#define VK_LUID_SIZE                                                     8U
-#define VK_LUID_SIZE_KHR                                                 VK_LUID_SIZE
-typedef enum VkExternalMemoryHandleTypeFlagBits
-{
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT = 0x00000001,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_BIT = 0x00000002,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT_BIT = 0x00000004,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_TEXTURE_BIT = 0x00000008,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_TEXTURE_KMT_BIT = 0x00000010,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_HEAP_BIT = 0x00000020,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE_BIT = 0x00000040,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT = 0x00000200,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID = 0x00000400,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT = 0x00000080,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT = 0x00000100,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_ZIRCON_VMO_BIT_FUCHSIA = 0x00000800,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_RDMA_ADDRESS_BIT_NV = 0x00001000,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_BIT_KHR = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_BIT,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT_BIT_KHR = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT_BIT,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_TEXTURE_BIT_KHR = VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_TEXTURE_BIT,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_TEXTURE_KMT_BIT_KHR = VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_TEXTURE_KMT_BIT,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_HEAP_BIT_KHR = VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_HEAP_BIT,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE_BIT_KHR = VK_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE_BIT,
-    VK_EXTERNAL_MEMORY_HANDLE_TYPE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkExternalMemoryHandleTypeFlagBits;
-typedef VkFlags VkExternalMemoryHandleTypeFlags;
-typedef enum VkExternalMemoryFeatureFlagBits
-{
-    VK_EXTERNAL_MEMORY_FEATURE_DEDICATED_ONLY_BIT = 0x00000001,
-    VK_EXTERNAL_MEMORY_FEATURE_EXPORTABLE_BIT = 0x00000002,
-    VK_EXTERNAL_MEMORY_FEATURE_IMPORTABLE_BIT = 0x00000004,
-    VK_EXTERNAL_MEMORY_FEATURE_DEDICATED_ONLY_BIT_KHR = VK_EXTERNAL_MEMORY_FEATURE_DEDICATED_ONLY_BIT,
-    VK_EXTERNAL_MEMORY_FEATURE_EXPORTABLE_BIT_KHR = VK_EXTERNAL_MEMORY_FEATURE_EXPORTABLE_BIT,
-    VK_EXTERNAL_MEMORY_FEATURE_IMPORTABLE_BIT_KHR = VK_EXTERNAL_MEMORY_FEATURE_IMPORTABLE_BIT,
-    VK_EXTERNAL_MEMORY_FEATURE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkExternalMemoryFeatureFlagBits;
-typedef VkFlags VkExternalMemoryFeatureFlags;
-typedef struct VkExternalMemoryProperties
-{
-    VkExternalMemoryFeatureFlags externalMemoryFeatures;
-    VkExternalMemoryHandleTypeFlags exportFromImportedHandleTypes;
-    VkExternalMemoryHandleTypeFlags compatibleHandleTypes;
-} VkExternalMemoryProperties;
-typedef struct VkPhysicalDeviceExternalImageFormatInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkExternalMemoryHandleTypeFlagBits handleType;
-} VkPhysicalDeviceExternalImageFormatInfo;
-typedef struct VkExternalImageFormatProperties
-{
-    VkStructureType sType;
-    void* pNext;
-    VkExternalMemoryProperties externalMemoryProperties;
-} VkExternalImageFormatProperties;
-typedef struct VkPhysicalDeviceExternalBufferInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkBufferCreateFlags flags;
-    VkBufferUsageFlags usage;
-    VkExternalMemoryHandleTypeFlagBits handleType;
-} VkPhysicalDeviceExternalBufferInfo;
-typedef struct VkExternalBufferProperties
-{
-    VkStructureType sType;
-    void* pNext;
-    VkExternalMemoryProperties externalMemoryProperties;
-} VkExternalBufferProperties;
-typedef struct VkPhysicalDeviceIDProperties
-{
-    VkStructureType sType;
-    void* pNext;
-    uint8_t deviceUUID[VK_UUID_SIZE];
-    uint8_t driverUUID[VK_UUID_SIZE];
-    uint8_t deviceLUID[VK_LUID_SIZE];
-    uint32_t deviceNodeMask;
-    VkBool32 deviceLUIDValid;
-} VkPhysicalDeviceIDProperties;
-typedef VkExternalMemoryHandleTypeFlags VkExternalMemoryHandleTypeFlagsKHR;
-typedef VkExternalMemoryHandleTypeFlagBits VkExternalMemoryHandleTypeFlagBitsKHR;
-typedef VkExternalMemoryFeatureFlags VkExternalMemoryFeatureFlagsKHR;
-typedef VkExternalMemoryFeatureFlagBits VkExternalMemoryFeatureFlagBitsKHR;
-typedef VkExternalMemoryProperties VkExternalMemoryPropertiesKHR;
-typedef VkPhysicalDeviceExternalImageFormatInfo VkPhysicalDeviceExternalImageFormatInfoKHR;
-typedef VkExternalImageFormatProperties VkExternalImageFormatPropertiesKHR;
-typedef VkPhysicalDeviceExternalBufferInfo VkPhysicalDeviceExternalBufferInfoKHR;
-typedef VkExternalBufferProperties VkExternalBufferPropertiesKHR;
-typedef VkPhysicalDeviceIDProperties VkPhysicalDeviceIDPropertiesKHR;
-typedef void(VKAPI_PTR* PFN_vkGetPhysicalDeviceExternalBufferProperties)(VkPhysicalDevice physicalDevice, const VkPhysicalDeviceExternalBufferInfo* pExternalBufferInfo, VkExternalBufferProperties* pExternalBufferProperties);
-typedef PFN_vkGetPhysicalDeviceExternalBufferProperties PFN_vkGetPhysicalDeviceExternalBufferPropertiesKHR;
-#endif // VK_KHR_external_memory_capabilities
-
-#ifndef VK_KHR_get_memory_requirements2
-#define VK_KHR_get_memory_requirements2                         1
-#define VK_STRUCTURE_TYPE_BUFFER_MEMORY_REQUIREMENTS_INFO_2     (VkStructureType)1000146000
-#define VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2      (VkStructureType)1000146001
-#define VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2                 (VkStructureType)1000146003
-#define VK_STRUCTURE_TYPE_BUFFER_MEMORY_REQUIREMENTS_INFO_2_KHR VK_STRUCTURE_TYPE_BUFFER_MEMORY_REQUIREMENTS_INFO_2
-#define VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2_KHR  VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2
-#define VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2_KHR             VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2
-typedef struct VkBufferMemoryRequirementsInfo2
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkBuffer buffer;
-} VkBufferMemoryRequirementsInfo2;
-typedef struct VkImageMemoryRequirementsInfo2
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkImage image;
-} VkImageMemoryRequirementsInfo2;
-typedef struct VkMemoryRequirements2
-{
-    VkStructureType sType;
-    void* pNext;
-    VkMemoryRequirements memoryRequirements;
-} VkMemoryRequirements2;
-typedef VkBufferMemoryRequirementsInfo2 VkBufferMemoryRequirementsInfo2KHR;
-typedef VkImageMemoryRequirementsInfo2 VkImageMemoryRequirementsInfo2KHR;
-typedef VkMemoryRequirements2 VkMemoryRequirements2KHR;
-typedef void(VKAPI_PTR* PFN_vkGetImageMemoryRequirements2)(VkDevice device, const VkImageMemoryRequirementsInfo2* pInfo, VkMemoryRequirements2* pMemoryRequirements);
-typedef void(VKAPI_PTR* PFN_vkGetBufferMemoryRequirements2)(VkDevice device, const VkBufferMemoryRequirementsInfo2* pInfo, VkMemoryRequirements2* pMemoryRequirements);
-typedef PFN_vkGetImageMemoryRequirements2 PFN_vkGetImageMemoryRequirements2KHR;
-typedef PFN_vkGetBufferMemoryRequirements2 PFN_vkGetBufferMemoryRequirements2KHR;
-#endif // VK_KHR_get_memory_requirements2
-
-#ifndef VK_KHR_dedicated_allocation
-#define VK_KHR_dedicated_allocation                          1
-#define VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS      (VkStructureType)1000127000
-#define VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO     (VkStructureType)1000127001
-#define VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR  VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
-#define VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO_KHR VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO
-typedef struct VkMemoryDedicatedRequirements
-{
-    VkStructureType sType;
-    void* pNext;
-    VkBool32 prefersDedicatedAllocation;
-    VkBool32 requiresDedicatedAllocation;
-} VkMemoryDedicatedRequirements;
-typedef struct VkMemoryDedicatedAllocateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkImage image;
-    VkBuffer buffer;
-} VkMemoryDedicatedAllocateInfo;
-typedef VkMemoryDedicatedRequirements VkMemoryDedicatedRequirementsKHR;
-typedef VkMemoryDedicatedAllocateInfo VkMemoryDedicatedAllocateInfoKHR;
-#endif // VK_KHR_dedicated_allocation
-
-#ifndef VK_KHR_16bit_storage
-#define VK_KHR_16bit_storage                                         1
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES     (VkStructureType)1000083000
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
-typedef struct VkPhysicalDevice16BitStorageFeatures
-{
-    VkStructureType sType;
-    void* pNext;
-    VkBool32 storageBuffer16BitAccess;
-    VkBool32 uniformAndStorageBuffer16BitAccess;
-    VkBool32 storagePushConstant16;
-    VkBool32 storageInputOutput16;
-} VkPhysicalDevice16BitStorageFeatures;
-typedef VkPhysicalDevice16BitStorageFeatures VkPhysicalDevice16BitStorageFeaturesKHR;
-#endif // VK_KHR_16bit_storage
-
-#ifndef VK_KHR_bind_memory2
-#define VK_KHR_bind_memory2                           1
-#define VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO     (VkStructureType)1000157000
-#define VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO      (VkStructureType)1000157001
-#define VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
-#define VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO_KHR  VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO
-typedef struct VkBindBufferMemoryInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkBuffer buffer;
-    VkDeviceMemory memory;
-    VkDeviceSize memoryOffset;
-} VkBindBufferMemoryInfo;
-typedef struct VkBindImageMemoryInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkImage image;
-    VkDeviceMemory memory;
-    VkDeviceSize memoryOffset;
-} VkBindImageMemoryInfo;
-typedef VkBindBufferMemoryInfo VkBindBufferMemoryInfoKHR;
-typedef VkBindImageMemoryInfo VkBindImageMemoryInfoKHR;
-typedef VkResult(VKAPI_PTR* PFN_vkBindBufferMemory2)(VkDevice device, uint32_t bindInfoCount, const VkBindBufferMemoryInfo* pBindInfos);
-typedef VkResult(VKAPI_PTR* PFN_vkBindImageMemory2)(VkDevice device, uint32_t bindInfoCount, const VkBindImageMemoryInfo* pBindInfos);
-typedef PFN_vkBindBufferMemory2 PFN_vkBindBufferMemory2KHR;
-typedef PFN_vkBindImageMemory2 PFN_vkBindImageMemory2KHR;
-#endif // VK_KHR_bind_memory2
-
-#ifndef VK_KHR_sampler_ycbcr_conversion
-#define VK_KHR_sampler_ycbcr_conversion                                         1
-#define VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_CREATE_INFO                  (VkStructureType)1000156000
-#define VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_INFO                         (VkStructureType)1000156001
-#define VK_STRUCTURE_TYPE_BIND_IMAGE_PLANE_MEMORY_INFO                          (VkStructureType)1000156002
-#define VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO                  (VkStructureType)1000156003
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES     (VkStructureType)1000156004
-#define VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_IMAGE_FORMAT_PROPERTIES      (VkStructureType)1000156005
-#define VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_CREATE_INFO_KHR              VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_CREATE_INFO
-#define VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_INFO_KHR                     VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_INFO
-#define VK_STRUCTURE_TYPE_BIND_IMAGE_PLANE_MEMORY_INFO_KHR                      VK_STRUCTURE_TYPE_BIND_IMAGE_PLANE_MEMORY_INFO
-#define VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO_KHR              VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES_KHR VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
-#define VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_IMAGE_FORMAT_PROPERTIES_KHR  VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_IMAGE_FORMAT_PROPERTIES
-VK_DEFINE_NON_DISPATCHABLE_HANDLE(VkSamplerYcbcrConversion)
-typedef VkSamplerYcbcrConversion VkSamplerYcbcrConversionKHR;
-typedef enum VkSamplerYcbcrModelConversion
-{
-    VK_SAMPLER_YCBCR_MODEL_CONVERSION_RGB_IDENTITY = 0,
-    VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_IDENTITY = 1,
-    VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_709 = 2,
-    VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_601 = 3,
-    VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_2020 = 4,
-    VK_SAMPLER_YCBCR_MODEL_CONVERSION_RGB_IDENTITY_KHR = VK_SAMPLER_YCBCR_MODEL_CONVERSION_RGB_IDENTITY,
-    VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_IDENTITY_KHR = VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_IDENTITY,
-    VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_709_KHR = VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_709,
-    VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_601_KHR = VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_601,
-    VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_2020_KHR = VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_2020,
-    VK_SAMPLER_YCBCR_MODEL_CONVERSION_MAX_ENUM = 0x7FFFFFFF
-} VkSamplerYcbcrModelConversion;
-typedef enum VkSamplerYcbcrRange
-{
-    VK_SAMPLER_YCBCR_RANGE_ITU_FULL = 0,
-    VK_SAMPLER_YCBCR_RANGE_ITU_NARROW = 1,
-    VK_SAMPLER_YCBCR_RANGE_ITU_FULL_KHR = VK_SAMPLER_YCBCR_RANGE_ITU_FULL,
-    VK_SAMPLER_YCBCR_RANGE_ITU_NARROW_KHR = VK_SAMPLER_YCBCR_RANGE_ITU_NARROW,
-    VK_SAMPLER_YCBCR_RANGE_MAX_ENUM = 0x7FFFFFFF
-} VkSamplerYcbcrRange;
-typedef enum VkChromaLocation
-{
-    VK_CHROMA_LOCATION_COSITED_EVEN = 0,
-    VK_CHROMA_LOCATION_MIDPOINT = 1,
-    VK_CHROMA_LOCATION_COSITED_EVEN_KHR = VK_CHROMA_LOCATION_COSITED_EVEN,
-    VK_CHROMA_LOCATION_MIDPOINT_KHR = VK_CHROMA_LOCATION_MIDPOINT,
-    VK_CHROMA_LOCATION_MAX_ENUM = 0x7FFFFFFF
-} VkChromaLocation;
-typedef struct VkSamplerYcbcrConversionCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkFormat format;
-    VkSamplerYcbcrModelConversion ycbcrModel;
-    VkSamplerYcbcrRange ycbcrRange;
-    VkComponentMapping components;
-    VkChromaLocation xChromaOffset;
-    VkChromaLocation yChromaOffset;
-    VkFilter chromaFilter;
-    VkBool32 forceExplicitReconstruction;
-} VkSamplerYcbcrConversionCreateInfo;
-typedef struct VkSamplerYcbcrConversionInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkSamplerYcbcrConversion conversion;
-} VkSamplerYcbcrConversionInfo;
-typedef struct VkBindImagePlaneMemoryInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkImageAspectFlagBits planeAspect;
-} VkBindImagePlaneMemoryInfo;
-typedef struct VkImagePlaneMemoryRequirementsInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkImageAspectFlagBits planeAspect;
-} VkImagePlaneMemoryRequirementsInfo;
-typedef struct VkPhysicalDeviceSamplerYcbcrConversionFeatures
-{
-    VkStructureType sType;
-    void* pNext;
-    VkBool32 samplerYcbcrConversion;
-} VkPhysicalDeviceSamplerYcbcrConversionFeatures;
-typedef struct VkSamplerYcbcrConversionImageFormatProperties
-{
-    VkStructureType sType;
-    void* pNext;
-    uint32_t combinedImageSamplerDescriptorCount;
-} VkSamplerYcbcrConversionImageFormatProperties;
-typedef VkSamplerYcbcrModelConversion VkSamplerYcbcrModelConversionKHR;
-typedef VkSamplerYcbcrRange VkSamplerYcbcrRangeKHR;
-typedef VkChromaLocation VkChromaLocationKHR;
-typedef VkSamplerYcbcrConversionCreateInfo VkSamplerYcbcrConversionCreateInfoKHR;
-typedef VkSamplerYcbcrConversionInfo VkSamplerYcbcrConversionInfoKHR;
-typedef VkBindImagePlaneMemoryInfo VkBindImagePlaneMemoryInfoKHR;
-typedef VkImagePlaneMemoryRequirementsInfo VkImagePlaneMemoryRequirementsInfoKHR;
-typedef VkPhysicalDeviceSamplerYcbcrConversionFeatures VkPhysicalDeviceSamplerYcbcrConversionFeaturesKHR;
-typedef VkSamplerYcbcrConversionImageFormatProperties VkSamplerYcbcrConversionImageFormatPropertiesKHR;
-typedef VkResult(VKAPI_PTR* PFN_vkCreateSamplerYcbcrConversion)(VkDevice device, const VkSamplerYcbcrConversionCreateInfo* pCreateInfo, const VkAllocationCallbacks* pAllocator, VkSamplerYcbcrConversion* pYcbcrConversion);
-typedef void(VKAPI_PTR* PFN_vkDestroySamplerYcbcrConversion)(VkDevice device, VkSamplerYcbcrConversion ycbcrConversion, const VkAllocationCallbacks* pAllocator);
-typedef PFN_vkCreateSamplerYcbcrConversion PFN_vkCreateSamplerYcbcrConversionKHR;
-typedef PFN_vkDestroySamplerYcbcrConversion PFN_vkDestroySamplerYcbcrConversionKHR;
-#endif // VK_KHR_sampler_ycbcr_conversion
-
-#if VK_HEADER_VERSION < 70
-// the unpublished VK_KHR_subgroup
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES (VkStructureType)1000094000
-typedef enum VkSubgroupFeatureFlagBits
-{
-    VK_SUBGROUP_FEATURE_BASIC_BIT = 0x00000001,
-    VK_SUBGROUP_FEATURE_VOTE_BIT = 0x00000002,
-    VK_SUBGROUP_FEATURE_ARITHMETIC_BIT = 0x00000004,
-    VK_SUBGROUP_FEATURE_BALLOT_BIT = 0x00000008,
-    VK_SUBGROUP_FEATURE_SHUFFLE_BIT = 0x00000010,
-    VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT = 0x00000020,
-    VK_SUBGROUP_FEATURE_CLUSTERED_BIT = 0x00000040,
-    VK_SUBGROUP_FEATURE_QUAD_BIT = 0x00000080,
-    VK_SUBGROUP_FEATURE_PARTITIONED_BIT_NV = 0x00000100,
-    VK_SUBGROUP_FEATURE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkSubgroupFeatureFlagBits;
-typedef VkFlags VkSubgroupFeatureFlags;
-typedef struct VkPhysicalDeviceSubgroupProperties
-{
-    VkStructureType sType;
-    void* pNext;
-    uint32_t subgroupSize;
-    VkShaderStageFlags supportedStages;
-    VkSubgroupFeatureFlags supportedOperations;
-    VkBool32 quadOperationsInAllStages;
-} VkPhysicalDeviceSubgroupProperties;
-
-// the unpublished VK_KHR_protected_memory
-#define VK_STRUCTURE_TYPE_PROTECTED_SUBMIT_INFO                       (VkStructureType)1000145000
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES   (VkStructureType)1000145001
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES (VkStructureType)1000145002
-#define VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2                         (VkStructureType)1000145003
-#define VK_MEMORY_PROPERTY_PROTECTED_BIT                              (VkMemoryPropertyFlagBits)0x00000020
-#define VK_IMAGE_CREATE_PROTECTED_BIT                                 (VkImageCreateFlagBits)0x00000800
-#define VK_BUFFER_CREATE_PROTECTED_BIT                                (VkBufferCreateFlagBits)0x00000008
-#define VK_QUEUE_PROTECTED_BIT                                        (VkQueueFlagBits)0x00000010
-#define VK_COMMAND_POOL_CREATE_PROTECTED_BIT                          (VkCommandPoolCreateFlagBits)0x00000004
-typedef enum VkDeviceQueueCreateFlagBits
-{
-    VK_DEVICE_QUEUE_CREATE_PROTECTED_BIT = 0x00000001,
-    VK_DEVICE_QUEUE_CREATE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkDeviceQueueCreateFlagBits;
-typedef VkFlags VkDeviceQueueCreateFlags;
-typedef struct VkPhysicalDeviceProtectedMemoryFeatures
-{
-    VkStructureType sType;
-    void* pNext;
-    VkBool32 protectedMemory;
-} VkPhysicalDeviceProtectedMemoryFeatures;
-typedef struct VkPhysicalDeviceProtectedMemoryProperties
-{
-    VkStructureType sType;
-    void* pNext;
-    VkBool32 protectedNoFault;
-} VkPhysicalDeviceProtectedMemoryProperties;
-typedef struct VkProtectedSubmitInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkBool32 protectedSubmit;
-} VkProtectedSubmitInfo;
-#endif // VK_HEADER_VERSION < 70
-
-#if VK_HEADER_VERSION < 204
-typedef uint64_t VkFlags64;
-#endif // VK_HEADER_VERSION < 204
-
-#ifndef VK_KHR_format_feature_flags2
-#define VK_KHR_format_feature_flags2              1
-#define VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_3     (VkStructureType)1000360000
-#define VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_3_KHR VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_3
-typedef VkFlags64 VkFormatFeatureFlags2;
-typedef VkFlags64 VkFormatFeatureFlagBits2;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_BIT = 0x00000001ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_BIT_KHR = 0x00000001ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_STORAGE_IMAGE_BIT = 0x00000002ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_STORAGE_IMAGE_BIT_KHR = 0x00000002ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_STORAGE_IMAGE_ATOMIC_BIT = 0x00000004ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_STORAGE_IMAGE_ATOMIC_BIT_KHR = 0x00000004ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_UNIFORM_TEXEL_BUFFER_BIT = 0x00000008ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_UNIFORM_TEXEL_BUFFER_BIT_KHR = 0x00000008ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_STORAGE_TEXEL_BUFFER_BIT = 0x00000010ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_STORAGE_TEXEL_BUFFER_BIT_KHR = 0x00000010ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_STORAGE_TEXEL_BUFFER_ATOMIC_BIT = 0x00000020ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_STORAGE_TEXEL_BUFFER_ATOMIC_BIT_KHR = 0x00000020ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_VERTEX_BUFFER_BIT = 0x00000040ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_VERTEX_BUFFER_BIT_KHR = 0x00000040ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_COLOR_ATTACHMENT_BIT = 0x00000080ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_COLOR_ATTACHMENT_BIT_KHR = 0x00000080ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_COLOR_ATTACHMENT_BLEND_BIT = 0x00000100ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_COLOR_ATTACHMENT_BLEND_BIT_KHR = 0x00000100ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_DEPTH_STENCIL_ATTACHMENT_BIT = 0x00000200ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_DEPTH_STENCIL_ATTACHMENT_BIT_KHR = 0x00000200ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_BLIT_SRC_BIT = 0x00000400ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_BLIT_SRC_BIT_KHR = 0x00000400ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_BLIT_DST_BIT = 0x00000800ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_BLIT_DST_BIT_KHR = 0x00000800ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_FILTER_LINEAR_BIT = 0x00001000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_FILTER_LINEAR_BIT_KHR = 0x00001000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_FILTER_CUBIC_BIT = 0x00002000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_FILTER_CUBIC_BIT_EXT = 0x00002000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_TRANSFER_SRC_BIT = 0x00004000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_TRANSFER_SRC_BIT_KHR = 0x00004000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_TRANSFER_DST_BIT = 0x00008000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_TRANSFER_DST_BIT_KHR = 0x00008000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_FILTER_MINMAX_BIT = 0x00010000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_FILTER_MINMAX_BIT_KHR = 0x00010000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_MIDPOINT_CHROMA_SAMPLES_BIT = 0x00020000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_MIDPOINT_CHROMA_SAMPLES_BIT_KHR = 0x00020000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_YCBCR_CONVERSION_LINEAR_FILTER_BIT = 0x00040000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_YCBCR_CONVERSION_LINEAR_FILTER_BIT_KHR = 0x00040000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_YCBCR_CONVERSION_SEPARATE_RECONSTRUCTION_FILTER_BIT = 0x00080000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_YCBCR_CONVERSION_SEPARATE_RECONSTRUCTION_FILTER_BIT_KHR = 0x00080000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_YCBCR_CONVERSION_CHROMA_RECONSTRUCTION_EXPLICIT_BIT = 0x00100000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_YCBCR_CONVERSION_CHROMA_RECONSTRUCTION_EXPLICIT_BIT_KHR = 0x00100000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_YCBCR_CONVERSION_CHROMA_RECONSTRUCTION_EXPLICIT_FORCEABLE_BIT = 0x00200000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_YCBCR_CONVERSION_CHROMA_RECONSTRUCTION_EXPLICIT_FORCEABLE_BIT_KHR = 0x00200000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_DISJOINT_BIT = 0x00400000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_DISJOINT_BIT_KHR = 0x00400000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_COSITED_CHROMA_SAMPLES_BIT = 0x00800000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_COSITED_CHROMA_SAMPLES_BIT_KHR = 0x00800000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_STORAGE_READ_WITHOUT_FORMAT_BIT = 0x80000000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_STORAGE_READ_WITHOUT_FORMAT_BIT_KHR = 0x80000000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_STORAGE_WRITE_WITHOUT_FORMAT_BIT = 0x100000000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_STORAGE_WRITE_WITHOUT_FORMAT_BIT_KHR = 0x100000000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_DEPTH_COMPARISON_BIT = 0x200000000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_DEPTH_COMPARISON_BIT_KHR = 0x200000000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_ACCELERATION_STRUCTURE_VERTEX_BUFFER_BIT_KHR = 0x20000000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_FRAGMENT_DENSITY_MAP_BIT_EXT = 0x01000000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_FRAGMENT_SHADING_RATE_ATTACHMENT_BIT_KHR = 0x40000000ULL;
-static const VkFormatFeatureFlagBits2 VK_FORMAT_FEATURE_2_LINEAR_COLOR_ATTACHMENT_BIT_NV = 0x4000000000ULL;
-typedef struct VkFormatProperties3
-{
-    VkStructureType sType;
-    void* pNext;
-    VkFormatFeatureFlags2 linearTilingFeatures;
-    VkFormatFeatureFlags2 optimalTilingFeatures;
-    VkFormatFeatureFlags2 bufferFeatures;
-} VkFormatProperties3;
-typedef VkFormatFeatureFlags2 VkFormatFeatureFlags2KHR;
-typedef VkFormatFeatureFlagBits2 VkFormatFeatureFlagBits2KHR;
-typedef VkFormatProperties3 VkFormatProperties3KHR;
-#endif // VK_KHR_format_feature_flags2
-
-#ifndef VK_ANDROID_external_memory_android_hardware_buffer
-#define VK_ANDROID_external_memory_android_hardware_buffer                    1
-#define VK_STRUCTURE_TYPE_ANDROID_HARDWARE_BUFFER_USAGE_ANDROID               (VkStructureType)1000129000
-#define VK_STRUCTURE_TYPE_ANDROID_HARDWARE_BUFFER_PROPERTIES_ANDROID          (VkStructureType)1000129001
-#define VK_STRUCTURE_TYPE_ANDROID_HARDWARE_BUFFER_FORMAT_PROPERTIES_ANDROID   (VkStructureType)1000129002
-#define VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID         (VkStructureType)1000129003
-#define VK_STRUCTURE_TYPE_MEMORY_GET_ANDROID_HARDWARE_BUFFER_INFO_ANDROID     (VkStructureType)1000129004
-#define VK_STRUCTURE_TYPE_EXTERNAL_FORMAT_ANDROID                             (VkStructureType)1000129005
-#define VK_STRUCTURE_TYPE_ANDROID_HARDWARE_BUFFER_FORMAT_PROPERTIES_2_ANDROID (VkStructureType)1000129006
-struct AHardwareBuffer;
-typedef struct VkAndroidHardwareBufferUsageANDROID
-{
-    VkStructureType sType;
-    void* pNext;
-    uint64_t androidHardwareBufferUsage;
-} VkAndroidHardwareBufferUsageANDROID;
-typedef struct VkAndroidHardwareBufferPropertiesANDROID
-{
-    VkStructureType sType;
-    void* pNext;
-    VkDeviceSize allocationSize;
-    uint32_t memoryTypeBits;
-} VkAndroidHardwareBufferPropertiesANDROID;
-typedef struct VkAndroidHardwareBufferFormatPropertiesANDROID
-{
-    VkStructureType sType;
-    void* pNext;
-    VkFormat format;
-    uint64_t externalFormat;
-    VkFormatFeatureFlags formatFeatures;
-    VkComponentMapping samplerYcbcrConversionComponents;
-    VkSamplerYcbcrModelConversion suggestedYcbcrModel;
-    VkSamplerYcbcrRange suggestedYcbcrRange;
-    VkChromaLocation suggestedXChromaOffset;
-    VkChromaLocation suggestedYChromaOffset;
-} VkAndroidHardwareBufferFormatPropertiesANDROID;
-typedef struct VkImportAndroidHardwareBufferInfoANDROID
-{
-    VkStructureType sType;
-    const void* pNext;
-    struct AHardwareBuffer* buffer;
-} VkImportAndroidHardwareBufferInfoANDROID;
-typedef struct VkMemoryGetAndroidHardwareBufferInfoANDROID
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkDeviceMemory memory;
-} VkMemoryGetAndroidHardwareBufferInfoANDROID;
-typedef struct VkExternalFormatANDROID
-{
-    VkStructureType sType;
-    void* pNext;
-    uint64_t externalFormat;
-} VkExternalFormatANDROID;
-typedef struct VkAndroidHardwareBufferFormatProperties2ANDROID
-{
-    VkStructureType sType;
-    void* pNext;
-    VkFormat format;
-    uint64_t externalFormat;
-    VkFormatFeatureFlags2 formatFeatures;
-    VkComponentMapping samplerYcbcrConversionComponents;
-    VkSamplerYcbcrModelConversion suggestedYcbcrModel;
-    VkSamplerYcbcrRange suggestedYcbcrRange;
-    VkChromaLocation suggestedXChromaOffset;
-    VkChromaLocation suggestedYChromaOffset;
-} VkAndroidHardwareBufferFormatProperties2ANDROID;
-typedef VkResult(VKAPI_PTR* PFN_vkGetAndroidHardwareBufferPropertiesANDROID)(VkDevice device, const struct AHardwareBuffer* buffer, VkAndroidHardwareBufferPropertiesANDROID* pProperties);
-typedef VkResult(VKAPI_PTR* PFN_vkGetMemoryAndroidHardwareBufferANDROID)(VkDevice device, const VkMemoryGetAndroidHardwareBufferInfoANDROID* pInfo, struct AHardwareBuffer** pBuffer);
-#endif // VK_ANDROID_external_memory_android_hardware_buffer
-
-#ifndef VK_KHR_maintenance3
-#define VK_KHR_maintenance3                                        1
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES (VkStructureType)1000168000
-#define VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_SUPPORT            (VkStructureType)1000168001
-typedef struct VkPhysicalDeviceMaintenance3Properties
-{
-    VkStructureType sType;
-    void* pNext;
-    uint32_t maxPerSetDescriptors;
-    VkDeviceSize maxMemoryAllocationSize;
-} VkPhysicalDeviceMaintenance3Properties;
-typedef struct VkDescriptorSetLayoutSupport
-{
-    VkStructureType sType;
-    void* pNext;
-    VkBool32 supported;
-} VkDescriptorSetLayoutSupport;
-typedef VkPhysicalDeviceMaintenance3Properties VkPhysicalDeviceMaintenance3PropertiesKHR;
-typedef VkDescriptorSetLayoutSupport VkDescriptorSetLayoutSupportKHR;
-typedef void(VKAPI_PTR* PFN_vkGetDescriptorSetLayoutSupport)(VkDevice device, const VkDescriptorSetLayoutCreateInfo* pCreateInfo, VkDescriptorSetLayoutSupport* pSupport);
-typedef PFN_vkGetDescriptorSetLayoutSupport PFN_vkGetDescriptorSetLayoutSupportKHR;
-#endif // VK_KHR_maintenance3
-
-#ifndef VK_KHR_8bit_storage
-#define VK_KHR_8bit_storage                                         1
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES     (VkStructureType)1000177000
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES
-typedef struct VkPhysicalDevice8BitStorageFeatures
-{
-    VkStructureType sType;
-    void* pNext;
-    VkBool32 storageBuffer8BitAccess;
-    VkBool32 uniformAndStorageBuffer8BitAccess;
-    VkBool32 storagePushConstant8;
-} VkPhysicalDevice8BitStorageFeatures;
-typedef VkPhysicalDevice8BitStorageFeatures VkPhysicalDevice8BitStorageFeaturesKHR;
-#endif // VK_KHR_8bit_storage
-
-#ifndef VK_KHR_shader_float16_int8
-#define VK_KHR_shader_float16_int8                                  1
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES     (VkStructureType)1000082000
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES
-typedef struct VkPhysicalDeviceFloat16Int8FeaturesKHR
-{
-    VkStructureType sType;
-    void* pNext;
-    VkBool32 shaderFloat16;
-    VkBool32 shaderInt8;
-} VkPhysicalDeviceFloat16Int8FeaturesKHR;
-#endif // VK_KHR_shader_float16_int8
-
-#ifndef VK_EXT_memory_budget
-#define VK_EXT_memory_budget                                           1
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT (VkStructureType)1000237000
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_PRIORITY_FEATURES_EXT (VkStructureType)1000238000
-#define VK_STRUCTURE_TYPE_MEMORY_PRIORITY_ALLOCATE_INFO_EXT            (VkStructureType)1000238001
-typedef uint64_t VkDeviceAddress;
-typedef struct VkPhysicalDeviceMemoryBudgetPropertiesEXT
-{
-    VkStructureType sType;
-    void* pNext;
-    VkDeviceSize heapBudget[VK_MAX_MEMORY_HEAPS];
-    VkDeviceSize heapUsage[VK_MAX_MEMORY_HEAPS];
-} VkPhysicalDeviceMemoryBudgetPropertiesEXT;
-typedef struct VkPhysicalDeviceMemoryPriorityFeaturesEXT
-{
-    VkStructureType sType;
-    void* pNext;
-    VkBool32 memoryPriority;
-} VkPhysicalDeviceMemoryPriorityFeaturesEXT;
-typedef struct VkMemoryPriorityAllocateInfoEXT
-{
-    VkStructureType sType;
-    const void* pNext;
-    float priority;
-} VkMemoryPriorityAllocateInfoEXT;
-#endif // VK_EXT_memory_budget
-
-#ifndef VK_EXT_buffer_device_address
-#define VK_EXT_buffer_device_address                                  1
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_ADDRESS_FEATURES_EXT (VkStructureType)1000244000
-#define VK_STRUCTURE_TYPE_BUFFER_DEVICE_ADDRESS_INFO_EXT              (VkStructureType)1000244001
-#define VK_STRUCTURE_TYPE_BUFFER_DEVICE_ADDRESS_CREATE_INFO_EXT       (VkStructureType)1000244002
-#define VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT_EXT                 (VkBufferUsageFlagBits)0x00020000
-typedef struct VkPhysicalDeviceBufferAddressFeaturesEXT
-{
-    VkStructureType sType;
-    void* pNext;
-    VkBool32 bufferDeviceAddress;
-    VkBool32 bufferDeviceAddressCaptureReplay;
-    VkBool32 bufferDeviceAddressMultiDevice;
-} VkPhysicalDeviceBufferAddressFeaturesEXT;
-typedef struct VkBufferDeviceAddressInfoEXT
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkBuffer buffer;
-} VkBufferDeviceAddressInfoEXT;
-typedef struct VkBufferDeviceAddressCreateInfoEXT
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkDeviceSize deviceAddress;
-} VkBufferDeviceAddressCreateInfoEXT;
-typedef VkDeviceAddress(VKAPI_PTR* PFN_vkGetBufferDeviceAddressEXT)(VkDevice device, const VkBufferDeviceAddressInfoEXT* pInfo);
-#endif // VK_EXT_buffer_device_address
-
-#ifndef VK_EXT_validation_features
-#define VK_EXT_validation_features                             1
-#define VK_STRUCTURE_TYPE_VALIDATION_FEATURES_EXT              (VkStructureType)1000247000
-#define VK_BUFFER_CREATE_DEVICE_ADDRESS_CAPTURE_REPLAY_BIT_EXT (VkBufferCreateFlagBits)0x00020000
-typedef enum VkValidationFeatureEnableEXT
-{
-    VK_VALIDATION_FEATURE_ENABLE_GPU_ASSISTED_EXT = 0,
-    VK_VALIDATION_FEATURE_ENABLE_GPU_ASSISTED_RESERVE_BINDING_SLOT_EXT = 1,
-    VK_VALIDATION_FEATURE_ENABLE_BEGIN_RANGE_EXT = VK_VALIDATION_FEATURE_ENABLE_GPU_ASSISTED_EXT,
-    VK_VALIDATION_FEATURE_ENABLE_END_RANGE_EXT = VK_VALIDATION_FEATURE_ENABLE_GPU_ASSISTED_RESERVE_BINDING_SLOT_EXT,
-    VK_VALIDATION_FEATURE_ENABLE_RANGE_SIZE_EXT = (VK_VALIDATION_FEATURE_ENABLE_GPU_ASSISTED_RESERVE_BINDING_SLOT_EXT - VK_VALIDATION_FEATURE_ENABLE_GPU_ASSISTED_EXT + 1),
-    VK_VALIDATION_FEATURE_ENABLE_MAX_ENUM_EXT = 0x7FFFFFFF
-} VkValidationFeatureEnableEXT;
-typedef enum VkValidationFeatureDisableEXT
-{
-    VK_VALIDATION_FEATURE_DISABLE_ALL_EXT = 0,
-    VK_VALIDATION_FEATURE_DISABLE_SHADERS_EXT = 1,
-    VK_VALIDATION_FEATURE_DISABLE_THREAD_SAFETY_EXT = 2,
-    VK_VALIDATION_FEATURE_DISABLE_API_PARAMETERS_EXT = 3,
-    VK_VALIDATION_FEATURE_DISABLE_OBJECT_LIFETIMES_EXT = 4,
-    VK_VALIDATION_FEATURE_DISABLE_CORE_CHECKS_EXT = 5,
-    VK_VALIDATION_FEATURE_DISABLE_UNIQUE_HANDLES_EXT = 6,
-    VK_VALIDATION_FEATURE_DISABLE_BEGIN_RANGE_EXT = VK_VALIDATION_FEATURE_DISABLE_ALL_EXT,
-    VK_VALIDATION_FEATURE_DISABLE_END_RANGE_EXT = VK_VALIDATION_FEATURE_DISABLE_UNIQUE_HANDLES_EXT,
-    VK_VALIDATION_FEATURE_DISABLE_RANGE_SIZE_EXT = (VK_VALIDATION_FEATURE_DISABLE_UNIQUE_HANDLES_EXT - VK_VALIDATION_FEATURE_DISABLE_ALL_EXT + 1),
-    VK_VALIDATION_FEATURE_DISABLE_MAX_ENUM_EXT = 0x7FFFFFFF
-} VkValidationFeatureDisableEXT;
-typedef struct VkValidationFeaturesEXT
-{
-    VkStructureType sType;
-    const void* pNext;
-    uint32_t enabledValidationFeatureCount;
-    const VkValidationFeatureEnableEXT* pEnabledValidationFeatures;
-    uint32_t disabledValidationFeatureCount;
-    const VkValidationFeatureDisableEXT* pDisabledValidationFeatures;
-} VkValidationFeaturesEXT;
-#endif // VK_EXT_validation_features
-
-#ifndef VK_NV_cooperative_matrix
-#define VK_NV_cooperative_matrix                                           1
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_NV   (VkStructureType)1000249000
-#define VK_STRUCTURE_TYPE_COOPERATIVE_MATRIX_PROPERTIES_NV                 (VkStructureType)1000249001
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_PROPERTIES_NV (VkStructureType)1000249002
-typedef enum VkComponentTypeNV
-{
-    VK_COMPONENT_TYPE_FLOAT16_NV = 0,
-    VK_COMPONENT_TYPE_FLOAT32_NV = 1,
-    VK_COMPONENT_TYPE_FLOAT64_NV = 2,
-    VK_COMPONENT_TYPE_SINT8_NV = 3,
-    VK_COMPONENT_TYPE_SINT16_NV = 4,
-    VK_COMPONENT_TYPE_SINT32_NV = 5,
-    VK_COMPONENT_TYPE_SINT64_NV = 6,
-    VK_COMPONENT_TYPE_UINT8_NV = 7,
-    VK_COMPONENT_TYPE_UINT16_NV = 8,
-    VK_COMPONENT_TYPE_UINT32_NV = 9,
-    VK_COMPONENT_TYPE_UINT64_NV = 10,
-    VK_COMPONENT_TYPE_BEGIN_RANGE_NV = VK_COMPONENT_TYPE_FLOAT16_NV,
-    VK_COMPONENT_TYPE_END_RANGE_NV = VK_COMPONENT_TYPE_UINT64_NV,
-    VK_COMPONENT_TYPE_RANGE_SIZE_NV = (VK_COMPONENT_TYPE_UINT64_NV - VK_COMPONENT_TYPE_FLOAT16_NV + 1),
-    VK_COMPONENT_TYPE_MAX_ENUM_NV = 0x7FFFFFFF
-} VkComponentTypeNV;
-typedef enum VkScopeNV
-{
-    VK_SCOPE_DEVICE_NV = 1,
-    VK_SCOPE_WORKGROUP_NV = 2,
-    VK_SCOPE_SUBGROUP_NV = 3,
-    VK_SCOPE_QUEUE_FAMILY_NV = 5,
-    VK_SCOPE_BEGIN_RANGE_NV = VK_SCOPE_DEVICE_NV,
-    VK_SCOPE_END_RANGE_NV = VK_SCOPE_QUEUE_FAMILY_NV,
-    VK_SCOPE_RANGE_SIZE_NV = (VK_SCOPE_QUEUE_FAMILY_NV - VK_SCOPE_DEVICE_NV + 1),
-    VK_SCOPE_MAX_ENUM_NV = 0x7FFFFFFF
-} VkScopeNV;
-typedef struct VkCooperativeMatrixPropertiesNV
-{
-    VkStructureType sType;
-    void* pNext;
-    uint32_t MSize;
-    uint32_t NSize;
-    uint32_t KSize;
-    VkComponentTypeNV AType;
-    VkComponentTypeNV BType;
-    VkComponentTypeNV CType;
-    VkComponentTypeNV DType;
-    VkScopeNV scope;
-} VkCooperativeMatrixPropertiesNV;
-typedef struct VkPhysicalDeviceCooperativeMatrixFeaturesNV
-{
-    VkStructureType sType;
-    void* pNext;
-    VkBool32 cooperativeMatrix;
-    VkBool32 cooperativeMatrixRobustBufferAccess;
-} VkPhysicalDeviceCooperativeMatrixFeaturesNV;
-typedef struct VkPhysicalDeviceCooperativeMatrixPropertiesNV
-{
-    VkStructureType sType;
-    void* pNext;
-    VkShaderStageFlags cooperativeMatrixSupportedStages;
-} VkPhysicalDeviceCooperativeMatrixPropertiesNV;
-typedef VkResult(VKAPI_PTR* PFN_vkGetPhysicalDeviceCooperativeMatrixPropertiesNV)(VkPhysicalDevice physicalDevice, uint32_t* pPropertyCount, VkCooperativeMatrixPropertiesNV* pProperties);
-#endif // VK_NV_cooperative_matrix
-
-#ifndef VK_AMD_device_coherent_memory
-#define VK_AMD_device_coherent_memory                                  1
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COHERENT_MEMORY_FEATURES_AMD (VkStructureType)1000229000
-#define VK_MEMORY_PROPERTY_DEVICE_COHERENT_BIT_AMD                     (VkMemoryPropertyFlagBits)0x00000040
-#define VK_MEMORY_PROPERTY_DEVICE_UNCACHED_BIT_AMD                     (VkMemoryPropertyFlagBits)0x00000040
-typedef struct VkPhysicalDeviceCoherentMemoryFeaturesAMD
-{
-    VkStructureType sType;
-    void* pNext;
-    VkBool32 deviceCoherentMemory;
-} VkPhysicalDeviceCoherentMemoryFeaturesAMD;
-#endif // VK_AMD_device_coherent_memory
-
-#ifndef VK_KHR_buffer_device_address
-#define VK_KHR_buffer_device_address                                         1
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES     (VkStructureType)1000257000
-#define VK_STRUCTURE_TYPE_BUFFER_DEVICE_ADDRESS_INFO                         (VkStructureType)1000244001
-#define VK_STRUCTURE_TYPE_BUFFER_OPAQUE_CAPTURE_ADDRESS_CREATE_INFO          (VkStructureType)1000257002
-#define VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO        (VkStructureType)1000257003
-#define VK_STRUCTURE_TYPE_DEVICE_MEMORY_OPAQUE_CAPTURE_ADDRESS_INFO          (VkStructureType)1000257004
-#define VK_BUFFER_CREATE_DEVICE_ADDRESS_CAPTURE_REPLAY_BIT                   (VkBufferCreateFlagBits)0x00020000
-#define VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT                            (VkBufferUsageFlagBits)0x00020000
-#define VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT                                (VkMemoryAllocateFlagBits)0x00000002
-#define VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_CAPTURE_REPLAY_BIT                 (VkMemoryAllocateFlagBits)0x00000004
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_KHR VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES
-#define VK_STRUCTURE_TYPE_BUFFER_DEVICE_ADDRESS_INFO_KHR                     VK_STRUCTURE_TYPE_BUFFER_DEVICE_ADDRESS_INFO
-#define VK_STRUCTURE_TYPE_BUFFER_OPAQUE_CAPTURE_ADDRESS_CREATE_INFO_KHR      VK_STRUCTURE_TYPE_BUFFER_OPAQUE_CAPTURE_ADDRESS_CREATE_INFO
-#define VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR    VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO
-#define VK_STRUCTURE_TYPE_DEVICE_MEMORY_OPAQUE_CAPTURE_ADDRESS_INFO_KHR      VK_STRUCTURE_TYPE_DEVICE_MEMORY_OPAQUE_CAPTURE_ADDRESS_INFO
-#define VK_BUFFER_CREATE_DEVICE_ADDRESS_CAPTURE_REPLAY_BIT_KHR               VK_BUFFER_CREATE_DEVICE_ADDRESS_CAPTURE_REPLAY_BIT
-#define VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT_KHR                        VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT
-#define VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR                            VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT
-#define VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_CAPTURE_REPLAY_BIT_KHR             VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_CAPTURE_REPLAY_BIT
-typedef struct VkPhysicalDeviceBufferDeviceAddressFeatures
-{
-    VkStructureType sType;
-    void* pNext;
-    VkBool32 bufferDeviceAddress;
-    VkBool32 bufferDeviceAddressCaptureReplay;
-    VkBool32 bufferDeviceAddressMultiDevice;
-} VkPhysicalDeviceBufferDeviceAddressFeatures;
-typedef struct VkBufferDeviceAddressInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkBuffer buffer;
-} VkBufferDeviceAddressInfo;
-typedef struct VkBufferOpaqueCaptureAddressCreateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    uint64_t opaqueCaptureAddress;
-} VkBufferOpaqueCaptureAddressCreateInfo;
-typedef struct VkMemoryOpaqueCaptureAddressAllocateInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    uint64_t opaqueCaptureAddress;
-} VkMemoryOpaqueCaptureAddressAllocateInfo;
-typedef struct VkDeviceMemoryOpaqueCaptureAddressInfo
-{
-    VkStructureType sType;
-    const void* pNext;
-    VkDeviceMemory memory;
-} VkDeviceMemoryOpaqueCaptureAddressInfo;
-typedef VkPhysicalDeviceBufferDeviceAddressFeatures VkPhysicalDeviceBufferDeviceAddressFeaturesKHR;
-typedef VkBufferDeviceAddressInfo VkBufferDeviceAddressInfoKHR;
-typedef VkBufferOpaqueCaptureAddressCreateInfo VkBufferOpaqueCaptureAddressCreateInfoKHR;
-typedef VkMemoryOpaqueCaptureAddressAllocateInfo VkMemoryOpaqueCaptureAddressAllocateInfoKHR;
-typedef VkDeviceMemoryOpaqueCaptureAddressInfo VkDeviceMemoryOpaqueCaptureAddressInfoKHR;
-typedef VkDeviceAddress(VKAPI_PTR* PFN_vkGetBufferDeviceAddress)(VkDevice device, const VkBufferDeviceAddressInfo* pInfo);
-typedef uint64_t(VKAPI_PTR* PFN_vkGetBufferOpaqueCaptureAddress)(VkDevice device, const VkBufferDeviceAddressInfo* pInfo);
-typedef uint64_t(VKAPI_PTR* PFN_vkGetDeviceMemoryOpaqueCaptureAddress)(VkDevice device, const VkDeviceMemoryOpaqueCaptureAddressInfo* pInfo);
-typedef PFN_vkGetBufferDeviceAddress PFN_vkGetBufferDeviceAddressKHR;
-typedef PFN_vkGetBufferOpaqueCaptureAddress PFN_vkGetBufferOpaqueCaptureAddressKHR;
-typedef PFN_vkGetDeviceMemoryOpaqueCaptureAddress PFN_vkGetDeviceMemoryOpaqueCaptureAddressKHR;
-#endif // VK_KHR_buffer_device_address
-
-#ifndef VK_KHR_portability_enumeration
-#define VK_KHR_portability_enumeration 1
-typedef enum VkInstanceCreateFlagBits
-{
-    VK_INSTANCE_CREATE_ENUMERATE_PORTABILITY_BIT_KHR = 0x00000001,
-    VK_INSTANCE_CREATE_FLAG_BITS_MAX_ENUM = 0x7FFFFFFF
-} VkInstanceCreateFlagBits;
-#endif // VK_KHR_portability_enumeration
-
-#ifndef VK_KHR_cooperative_matrix
-#define VK_KHR_cooperative_matrix                                           1
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR   (VkStructureType)1000506000
-#define VK_STRUCTURE_TYPE_COOPERATIVE_MATRIX_PROPERTIES_KHR                 (VkStructureType)1000506001
-#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_PROPERTIES_KHR (VkStructureType)1000506002
-typedef enum VkComponentTypeKHR
-{
-    VK_COMPONENT_TYPE_FLOAT16_KHR = 0,
-    VK_COMPONENT_TYPE_FLOAT32_KHR = 1,
-    VK_COMPONENT_TYPE_FLOAT64_KHR = 2,
-    VK_COMPONENT_TYPE_SINT8_KHR = 3,
-    VK_COMPONENT_TYPE_SINT16_KHR = 4,
-    VK_COMPONENT_TYPE_SINT32_KHR = 5,
-    VK_COMPONENT_TYPE_SINT64_KHR = 6,
-    VK_COMPONENT_TYPE_UINT8_KHR = 7,
-    VK_COMPONENT_TYPE_UINT16_KHR = 8,
-    VK_COMPONENT_TYPE_UINT32_KHR = 9,
-    VK_COMPONENT_TYPE_UINT64_KHR = 10,
-    VK_COMPONENT_TYPE_MAX_ENUM_KHR = 0x7FFFFFFF
-} VkComponentTypeKHR;
-typedef enum VkScopeKHR
-{
-    VK_SCOPE_DEVICE_KHR = 1,
-    VK_SCOPE_WORKGROUP_KHR = 2,
-    VK_SCOPE_SUBGROUP_KHR = 3,
-    VK_SCOPE_QUEUE_FAMILY_KHR = 5,
-    VK_SCOPE_MAX_ENUM_KHR = 0x7FFFFFFF
-} VkScopeKHR;
-typedef struct VkCooperativeMatrixPropertiesKHR
-{
-    VkStructureType sType;
-    void* pNext;
-    uint32_t MSize;
-    uint32_t NSize;
-    uint32_t KSize;
-    VkComponentTypeKHR AType;
-    VkComponentTypeKHR BType;
-    VkComponentTypeKHR CType;
-    VkComponentTypeKHR ResultType;
-    VkBool32 saturatingAccumulation;
-    VkScopeKHR scope;
-} VkCooperativeMatrixPropertiesKHR;
-typedef struct VkPhysicalDeviceCooperativeMatrixFeaturesKHR
-{
-    VkStructureType sType;
-    void* pNext;
-    VkBool32 cooperativeMatrix;
-    VkBool32 cooperativeMatrixRobustBufferAccess;
-} VkPhysicalDeviceCooperativeMatrixFeaturesKHR;
-typedef struct VkPhysicalDeviceCooperativeMatrixPropertiesKHR
-{
-    VkStructureType sType;
-    void* pNext;
-    VkShaderStageFlags cooperativeMatrixSupportedStages;
-} VkPhysicalDeviceCooperativeMatrixPropertiesKHR;
-typedef VkResult(VKAPI_PTR* PFN_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR)(VkPhysicalDevice physicalDevice, uint32_t* pPropertyCount, VkCooperativeMatrixPropertiesKHR* pProperties);
-#endif // VK_KHR_cooperative_matrix
-
-#endif // NCNN_VULKAN_HEADER_FIX_H
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/alsa-play.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/alsa-play.h
deleted file mode 100644
index de324ca0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/alsa-play.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// sherpa-onnx/csrc/alsa-play.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ALSA_PLAY_H_
-#define SHERPA_ONNX_CSRC_ALSA_PLAY_H_
-
-#include <cstdint>
-#include <memory>
-#include <vector>
-
-#include "alsa/asoundlib.h"
-#include "sherpa-onnx/csrc/resample.h"
-
-namespace sherpa_onnx {
-
-class AlsaPlay {
- public:
-  AlsaPlay(const char *device_name, int32_t sample_rate);
-  ~AlsaPlay();
-  void Play(const std::vector<float> &samples);
-
-  // wait for all the samples to be played
-  void Drain();
-
- private:
-  void SetParameters(int32_t sample_rate);
-
- private:
-  snd_pcm_t *handle_ = nullptr;
-  std::unique_ptr<LinearResample> resampler_;
-  std::vector<int16_t> buf_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ALSA_PLAY_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/alsa.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/alsa.h
deleted file mode 100644
index 7a9ca7a7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/alsa.h
+++ /dev/null
@@ -1,46 +0,0 @@
-// sherpa-onnx/csrc/sherpa-alsa.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ALSA_H_
-#define SHERPA_ONNX_CSRC_ALSA_H_
-
-#include <memory>
-#include <vector>
-
-#include "alsa/asoundlib.h"
-#include "sherpa-onnx/csrc/resample.h"
-
-namespace sherpa_onnx {
-
-class Alsa {
- public:
-  explicit Alsa(const char *device_name);
-  ~Alsa();
-
-  // This is a blocking read.
-  //
-  // @param num_samples  Number of samples to read.
-  //
-  // The returned value is valid until the next call to Read().
-  const std::vector<float> &Read(int32_t num_samples);
-
-  int32_t GetExpectedSampleRate() const { return expected_sample_rate_; }
-  int32_t GetActualSampleRate() const { return actual_sample_rate_; }
-
- private:
-  snd_pcm_t *capture_handle_;
-  int32_t expected_sample_rate_ = 16000;
-  int32_t actual_sample_rate_;
-
-  int32_t actual_channel_count_ = 1;
-
-  std::unique_ptr<LinearResample> resampler_;
-  std::vector<int16_t> samples_;  // directly from the microphone
-  std::vector<float> samples1_;   // normalized version of samples_
-  std::vector<float> samples2_;   // possibly resampled from samples1_
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ALSA_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-ced-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-ced-impl.h
deleted file mode 100644
index a1f94348..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-ced-impl.h
+++ /dev/null
@@ -1,111 +0,0 @@
-// sherpa-onnx/csrc/audio-tagging-ced-impl.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_AUDIO_TAGGING_CED_IMPL_H_
-#define SHERPA_ONNX_CSRC_AUDIO_TAGGING_CED_IMPL_H_
-
-#include <assert.h>
-
-#include <memory>
-#include <utility>
-#include <vector>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "sherpa-onnx/csrc/audio-tagging-impl.h"
-#include "sherpa-onnx/csrc/audio-tagging-label-file.h"
-#include "sherpa-onnx/csrc/audio-tagging.h"
-#include "sherpa-onnx/csrc/macros.h"
-#include "sherpa-onnx/csrc/math.h"
-#include "sherpa-onnx/csrc/offline-ced-model.h"
-
-namespace sherpa_onnx {
-
-class AudioTaggingCEDImpl : public AudioTaggingImpl {
- public:
-  explicit AudioTaggingCEDImpl(const AudioTaggingConfig &config)
-      : config_(config), model_(config.model), labels_(config.labels) {
-    if (model_.NumEventClasses() != labels_.NumEventClasses()) {
-      SHERPA_ONNX_LOGE("number of classes: %d (model) != %d (label file)",
-                       model_.NumEventClasses(), labels_.NumEventClasses());
-      exit(-1);
-    }
-  }
-
-#if __ANDROID_API__ >= 9
-  explicit AudioTaggingCEDImpl(AAssetManager *mgr,
-                               const AudioTaggingConfig &config)
-      : config_(config),
-        model_(mgr, config.model),
-        labels_(mgr, config.labels) {
-    if (model_.NumEventClasses() != labels_.NumEventClasses()) {
-      SHERPA_ONNX_LOGE("number of classes: %d (model) != %d (label file)",
-                       model_.NumEventClasses(), labels_.NumEventClasses());
-      exit(-1);
-    }
-  }
-#endif
-
-  std::unique_ptr<OfflineStream> CreateStream() const override {
-    return std::make_unique<OfflineStream>(CEDTag{});
-  }
-
-  std::vector<AudioEvent> Compute(OfflineStream *s,
-                                  int32_t top_k = -1) const override {
-    if (top_k < 0) {
-      top_k = config_.top_k;
-    }
-
-    int32_t num_event_classes = model_.NumEventClasses();
-
-    if (top_k > num_event_classes) {
-      top_k = num_event_classes;
-    }
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    // WARNING(fangjun): It is fixed to 64 for CED models
-    int32_t feat_dim = 64;
-    std::vector<float> f = s->GetFrames();
-
-    int32_t num_frames = f.size() / feat_dim;
-    assert(feat_dim * num_frames == static_cast<int32_t>(f.size()));
-
-    std::array<int64_t, 3> shape = {1, num_frames, feat_dim};
-
-    Ort::Value x = Ort::Value::CreateTensor(memory_info, f.data(), f.size(),
-                                            shape.data(), shape.size());
-
-    Ort::Value probs = model_.Forward(std::move(x));
-
-    const float *p = probs.GetTensorData<float>();
-
-    std::vector<int32_t> top_k_indexes = TopkIndex(p, num_event_classes, top_k);
-
-    std::vector<AudioEvent> ans(top_k);
-
-    int32_t i = 0;
-
-    for (int32_t index : top_k_indexes) {
-      ans[i].name = labels_.GetEventName(index);
-      ans[i].index = index;
-      ans[i].prob = p[index];
-      i += 1;
-    }
-
-    return ans;
-  }
-
- private:
-  AudioTaggingConfig config_;
-  OfflineCEDModel model_;
-  AudioTaggingLabels labels_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_AUDIO_TAGGING_CED_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-impl.h
deleted file mode 100644
index ac6f2e50..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-impl.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// sherpa-onnx/csrc/audio-tagging-impl.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_AUDIO_TAGGING_IMPL_H_
-#define SHERPA_ONNX_CSRC_AUDIO_TAGGING_IMPL_H_
-
-#include <memory>
-#include <vector>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "sherpa-onnx/csrc/audio-tagging.h"
-
-namespace sherpa_onnx {
-
-class AudioTaggingImpl {
- public:
-  virtual ~AudioTaggingImpl() = default;
-
-  static std::unique_ptr<AudioTaggingImpl> Create(
-      const AudioTaggingConfig &config);
-
-#if __ANDROID_API__ >= 9
-  static std::unique_ptr<AudioTaggingImpl> Create(
-      AAssetManager *mgr, const AudioTaggingConfig &config);
-#endif
-
-  virtual std::unique_ptr<OfflineStream> CreateStream() const = 0;
-
-  virtual std::vector<AudioEvent> Compute(OfflineStream *s,
-                                          int32_t top_k = -1) const = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_AUDIO_TAGGING_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-label-file.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-label-file.h
deleted file mode 100644
index c366972e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-label-file.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// sherpa-onnx/csrc/audio-tagging-label-file.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_AUDIO_TAGGING_LABEL_FILE_H_
-#define SHERPA_ONNX_CSRC_AUDIO_TAGGING_LABEL_FILE_H_
-
-#include <istream>
-#include <string>
-#include <vector>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-namespace sherpa_onnx {
-
-class AudioTaggingLabels {
- public:
-  explicit AudioTaggingLabels(const std::string &filename);
-#if __ANDROID_API__ >= 9
-  AudioTaggingLabels(AAssetManager *mgr, const std::string &filename);
-#endif
-
-  // Return the event name for the given index.
-  // The returned reference is valid as long as this object is alive
-  const std::string &GetEventName(int32_t index) const;
-  int32_t NumEventClasses() const { return names_.size(); }
-
- private:
-  void Init(std::istream &is);
-
- private:
-  std::vector<std::string> names_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_AUDIO_TAGGING_LABEL_FILE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-model-config.h
deleted file mode 100644
index 2f31ff17..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-model-config.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// sherpa-onnx/csrc/audio-tagging-model-config.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_AUDIO_TAGGING_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_AUDIO_TAGGING_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/offline-zipformer-audio-tagging-model-config.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct AudioTaggingModelConfig {
-  struct OfflineZipformerAudioTaggingModelConfig zipformer;
-  std::string ced;
-
-  int32_t num_threads = 1;
-  bool debug = false;
-  std::string provider = "cpu";
-
-  AudioTaggingModelConfig() = default;
-
-  AudioTaggingModelConfig(
-      const OfflineZipformerAudioTaggingModelConfig &zipformer,
-      const std::string &ced, int32_t num_threads, bool debug,
-      const std::string &provider)
-      : zipformer(zipformer),
-        ced(ced),
-        num_threads(num_threads),
-        debug(debug),
-        provider(provider) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_AUDIO_TAGGING_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-zipformer-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-zipformer-impl.h
deleted file mode 100644
index 5f7af265..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging-zipformer-impl.h
+++ /dev/null
@@ -1,118 +0,0 @@
-// sherpa-onnx/csrc/audio-tagging-zipformer-impl.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_AUDIO_TAGGING_ZIPFORMER_IMPL_H_
-#define SHERPA_ONNX_CSRC_AUDIO_TAGGING_ZIPFORMER_IMPL_H_
-
-#include <assert.h>
-
-#include <memory>
-#include <utility>
-#include <vector>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "sherpa-onnx/csrc/audio-tagging-impl.h"
-#include "sherpa-onnx/csrc/audio-tagging-label-file.h"
-#include "sherpa-onnx/csrc/audio-tagging.h"
-#include "sherpa-onnx/csrc/macros.h"
-#include "sherpa-onnx/csrc/math.h"
-#include "sherpa-onnx/csrc/offline-zipformer-audio-tagging-model.h"
-
-namespace sherpa_onnx {
-
-class AudioTaggingZipformerImpl : public AudioTaggingImpl {
- public:
-  explicit AudioTaggingZipformerImpl(const AudioTaggingConfig &config)
-      : config_(config), model_(config.model), labels_(config.labels) {
-    if (model_.NumEventClasses() != labels_.NumEventClasses()) {
-      SHERPA_ONNX_LOGE("number of classes: %d (model) != %d (label file)",
-                       model_.NumEventClasses(), labels_.NumEventClasses());
-      exit(-1);
-    }
-  }
-
-#if __ANDROID_API__ >= 9
-  explicit AudioTaggingZipformerImpl(AAssetManager *mgr,
-                                     const AudioTaggingConfig &config)
-      : config_(config),
-        model_(mgr, config.model),
-        labels_(mgr, config.labels) {
-    if (model_.NumEventClasses() != labels_.NumEventClasses()) {
-      SHERPA_ONNX_LOGE("number of classes: %d (model) != %d (label file)",
-                       model_.NumEventClasses(), labels_.NumEventClasses());
-      exit(-1);
-    }
-  }
-#endif
-
-  std::unique_ptr<OfflineStream> CreateStream() const override {
-    return std::make_unique<OfflineStream>();
-  }
-
-  std::vector<AudioEvent> Compute(OfflineStream *s,
-                                  int32_t top_k = -1) const override {
-    if (top_k < 0) {
-      top_k = config_.top_k;
-    }
-
-    int32_t num_event_classes = model_.NumEventClasses();
-
-    if (top_k > num_event_classes) {
-      top_k = num_event_classes;
-    }
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    // WARNING(fangjun): It is fixed to 80 for all models from icefall
-    int32_t feat_dim = 80;
-    std::vector<float> f = s->GetFrames();
-
-    int32_t num_frames = f.size() / feat_dim;
-
-    assert(feat_dim * num_frames == static_cast<int32_t>(f.size()));
-
-    std::array<int64_t, 3> shape = {1, num_frames, feat_dim};
-
-    Ort::Value x = Ort::Value::CreateTensor(memory_info, f.data(), f.size(),
-                                            shape.data(), shape.size());
-
-    int64_t x_length_scalar = num_frames;
-    std::array<int64_t, 1> x_length_shape = {1};
-    Ort::Value x_length =
-        Ort::Value::CreateTensor(memory_info, &x_length_scalar, 1,
-                                 x_length_shape.data(), x_length_shape.size());
-
-    Ort::Value probs = model_.Forward(std::move(x), std::move(x_length));
-
-    const float *p = probs.GetTensorData<float>();
-
-    std::vector<int32_t> top_k_indexes = TopkIndex(p, num_event_classes, top_k);
-
-    std::vector<AudioEvent> ans(top_k);
-
-    int32_t i = 0;
-
-    for (int32_t index : top_k_indexes) {
-      ans[i].name = labels_.GetEventName(index);
-      ans[i].index = index;
-      ans[i].prob = p[index];
-      i += 1;
-    }
-
-    return ans;
-  }
-
- private:
-  AudioTaggingConfig config_;
-  OfflineZipformerAudioTaggingModel model_;
-  AudioTaggingLabels labels_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_AUDIO_TAGGING_ZIPFORMER_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging.h
deleted file mode 100644
index 6f68e90f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/audio-tagging.h
+++ /dev/null
@@ -1,74 +0,0 @@
-// sherpa-onnx/csrc/audio-tagging.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_AUDIO_TAGGING_H_
-#define SHERPA_ONNX_CSRC_AUDIO_TAGGING_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "sherpa-onnx/csrc/audio-tagging-model-config.h"
-#include "sherpa-onnx/csrc/offline-stream.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct AudioTaggingConfig {
-  AudioTaggingModelConfig model;
-  std::string labels;
-
-  int32_t top_k = 5;
-
-  AudioTaggingConfig() = default;
-
-  AudioTaggingConfig(const AudioTaggingModelConfig &model,
-                     const std::string &labels, int32_t top_k)
-      : model(model), labels(labels), top_k(top_k) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-struct AudioEvent {
-  std::string name;  // name of the event
-  int32_t index;     // index of the event in the label file
-  float prob;        // probability of the event
-
-  std::string ToString() const;
-};
-
-class AudioTaggingImpl;
-
-class AudioTagging {
- public:
-  explicit AudioTagging(const AudioTaggingConfig &config);
-
-#if __ANDROID_API__ >= 9
-  AudioTagging(AAssetManager *mgr, const AudioTaggingConfig &config);
-#endif
-
-  ~AudioTagging();
-
-  std::unique_ptr<OfflineStream> CreateStream() const;
-
-  // If top_k is -1, then config.top_k is used.
-  // Otherwise, config.top_k is ignored
-  //
-  // Return top_k AudioEvent. ans[0].prob is the largest of all returned events.
-  std::vector<AudioEvent> Compute(OfflineStream *s, int32_t top_k = -1) const;
-
- private:
-  std::unique_ptr<AudioTaggingImpl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_AUDIO_TAGGING_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/base64-decode.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/base64-decode.h
deleted file mode 100644
index 3b2f9a34..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/base64-decode.h
+++ /dev/null
@@ -1,19 +0,0 @@
-// sherpa-onnx/csrc/base64-decode.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_BASE64_DECODE_H_
-#define SHERPA_ONNX_CSRC_BASE64_DECODE_H_
-
-#include <string>
-
-namespace sherpa_onnx {
-
-/** @param s A base64 encoded string.
- *  @return Return the decoded string.
- */
-std::string Base64Decode(const std::string &s);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_BASE64_DECODE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/cat.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/cat.h
deleted file mode 100644
index 92ea8ba9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/cat.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// sherpa-onnx/csrc/cat.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_CAT_H_
-#define SHERPA_ONNX_CSRC_CAT_H_
-
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-
-namespace sherpa_onnx {
-
-/** Cat a list of tensors along the given dim.
- *
- * @param allocator Allocator to allocate space for the returned tensor
- * @param values  Pointer to a list of tensors. The shape of the tensor must
- *                be the same except on the dim to be concatenated.
- * @param dim  The dim along which to concatenate the input tensors
- *
- * @return Return the concatenated tensor
- */
-template <typename T = float>
-Ort::Value Cat(OrtAllocator *allocator,
-               const std::vector<const Ort::Value *> &values, int32_t dim);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_CAT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/circular-buffer.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/circular-buffer.h
deleted file mode 100644
index 0f668952..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/circular-buffer.h
+++ /dev/null
@@ -1,61 +0,0 @@
-// sherpa-onnx/csrc/circular-buffer.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_CIRCULAR_BUFFER_H_
-#define SHERPA_ONNX_CSRC_CIRCULAR_BUFFER_H_
-
-#include <cstdint>
-#include <vector>
-
-namespace sherpa_onnx {
-
-class CircularBuffer {
- public:
-  // Capacity of this buffer. Should be large enough.
-  // If it is full, we just print a message and exit the program.
-  explicit CircularBuffer(int32_t capacity);
-
-  // Push an array
-  //
-  // @param p Pointer to the start address of the array
-  // @param n Number of elements in the array
-  //
-  // Note: If n + Size() > capacity, we print an error message and exit.
-  void Push(const float *p, int32_t n);
-
-  // @param start_index Should in the range [head_, tail_)
-  // @param n Number of elements to get
-  // @return Return a vector of size n containing the requested elements
-  std::vector<float> Get(int32_t start_index, int32_t n) const;
-
-  // Remove n elements from the buffer
-  //
-  // @param n Should be in the range [0, size_]
-  void Pop(int32_t n);
-
-  // Number of elements in the buffer.
-  int32_t Size() const { return tail_ - head_; }
-
-  // Current position of the head
-  int32_t Head() const { return head_; }
-
-  // Current position of the tail
-  int32_t Tail() const { return tail_; }
-
-  void Reset() {
-    head_ = 0;
-    tail_ = 0;
-  }
-
-  void Resize(int32_t new_capacity);
-
- private:
-  std::vector<float> buffer_;
-
-  int32_t head_ = 0;  // linear index; always increasing; never wraps around
-  int32_t tail_ = 0;  // linear index, always increasing; never wraps around.
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_CIRCULAR_BUFFER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/context-graph.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/context-graph.h
deleted file mode 100644
index e16fc4d3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/context-graph.h
+++ /dev/null
@@ -1,92 +0,0 @@
-// sherpa-onnx/csrc/context-graph.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_CONTEXT_GRAPH_H_
-#define SHERPA_ONNX_CSRC_CONTEXT_GRAPH_H_
-
-#include <memory>
-#include <string>
-#include <tuple>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include "sherpa-onnx/csrc/log.h"
-
-namespace sherpa_onnx {
-
-class ContextGraph;
-using ContextGraphPtr = std::shared_ptr<ContextGraph>;
-
-struct ContextState {
-  int32_t token;
-  float token_score;
-  float node_score;
-  float output_score;
-  int32_t level;
-  float ac_threshold;
-  bool is_end;
-  std::string phrase;
-  std::unordered_map<int32_t, std::unique_ptr<ContextState>> next;
-  const ContextState *fail = nullptr;
-  const ContextState *output = nullptr;
-
-  ContextState() = default;
-  ContextState(int32_t token, float token_score, float node_score,
-               float output_score, int32_t level = 0, float ac_threshold = 0.0f,
-               bool is_end = false, const std::string &phrase = {})
-      : token(token),
-        token_score(token_score),
-        node_score(node_score),
-        output_score(output_score),
-        level(level),
-        ac_threshold(ac_threshold),
-        is_end(is_end),
-        phrase(phrase) {}
-};
-
-class ContextGraph {
- public:
-  ContextGraph() = default;
-  ContextGraph(const std::vector<std::vector<int32_t>> &token_ids,
-               float context_score, float ac_threshold,
-               const std::vector<float> &scores = {},
-               const std::vector<std::string> &phrases = {},
-               const std::vector<float> &ac_thresholds = {})
-      : context_score_(context_score), ac_threshold_(ac_threshold) {
-    root_ = std::make_unique<ContextState>(-1, 0, 0, 0);
-    root_->fail = root_.get();
-    Build(token_ids, scores, phrases, ac_thresholds);
-  }
-
-  ContextGraph(const std::vector<std::vector<int32_t>> &token_ids,
-               float context_score, const std::vector<float> &scores = {})
-      : ContextGraph(token_ids, context_score, 0.0f, scores,
-                     std::vector<std::string>(), std::vector<float>()) {}
-
-  std::tuple<float, const ContextState *, const ContextState *> ForwardOneStep(
-      const ContextState *state, int32_t token_id,
-      bool strict_mode = true) const;
-
-  std::pair<bool, const ContextState *> IsMatched(
-      const ContextState *state) const;
-
-  std::pair<float, const ContextState *> Finalize(
-      const ContextState *state) const;
-
-  const ContextState *Root() const { return root_.get(); }
-
- private:
-  float context_score_;
-  float ac_threshold_;
-  std::unique_ptr<ContextState> root_;
-  void Build(const std::vector<std::vector<int32_t>> &token_ids,
-             const std::vector<float> &scores,
-             const std::vector<std::string> &phrases,
-             const std::vector<float> &ac_thresholds) const;
-  void FillFailOutput() const;
-};
-
-}  // namespace sherpa_onnx
-#endif  // SHERPA_ONNX_CSRC_CONTEXT_GRAPH_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/display.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/display.h
deleted file mode 100644
index a366a4f8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/display.h
+++ /dev/null
@@ -1,88 +0,0 @@
-// sherpa-onnx/csrc/display.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_DISPLAY_H_
-#define SHERPA_ONNX_CSRC_DISPLAY_H_
-#include <stdio.h>
-
-#include <string>
-
-namespace sherpa_onnx {
-
-class Display {
- public:
-  explicit Display(int32_t max_word_per_line = 60)
-      : max_word_per_line_(max_word_per_line) {}
-
-  void Print(int32_t segment_id, const std::string &s) {
-#ifdef _MSC_VER
-    if (segment_id != -1) {
-      fprintf(stderr, "%d:%s\n", segment_id, s.c_str());
-    } else {
-      fprintf(stderr, "%s\n", s.c_str());
-    }
-    return;
-#endif
-    if (last_segment_ == segment_id) {
-      Clear();
-    } else {
-      if (last_segment_ != -1) {
-        fprintf(stderr, "\n\r");
-      }
-      last_segment_ = segment_id;
-      num_previous_lines_ = 0;
-    }
-
-    if (segment_id != -1) {
-      fprintf(stderr, "\r%d:", segment_id);
-    }
-
-    int32_t i = 0;
-    for (size_t n = 0; n < s.size();) {
-      if (s[n] > 0 && s[n] < 0x7f) {
-        fprintf(stderr, "%c", s[n]);
-        ++n;
-      } else {
-        // Each Chinese character occupies 3 bytes for UTF-8 encoding.
-        std::string tmp(s.begin() + n, s.begin() + n + 3);
-        fprintf(stderr, "%s", tmp.data());
-        n += 3;
-      }
-
-      ++i;
-      if (i >= max_word_per_line_ && n + 1 < s.size() &&
-          (s[n] == ' ' || s[n] < 0)) {
-        fprintf(stderr, "\n\r ");
-        ++num_previous_lines_;
-        i = 0;
-      }
-    }
-  }
-
- private:
-  // Clear the output for the current segment
-  void Clear() {
-    ClearCurrentLine();
-    while (num_previous_lines_ > 0) {
-      GoUpOneLine();
-      ClearCurrentLine();
-      --num_previous_lines_;
-    }
-  }
-
-  // Clear the current line
-  void ClearCurrentLine() const { fprintf(stderr, "\33[2K\r"); }
-
-  // Move the cursor to the previous line
-  void GoUpOneLine() const { fprintf(stderr, "\033[1A\r"); }
-
- private:
-  int32_t max_word_per_line_;
-  int32_t num_previous_lines_ = 0;
-  int32_t last_segment_ = -1;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_DISPLAY_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/endpoint.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/endpoint.h
deleted file mode 100644
index aea5d7d2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/endpoint.h
+++ /dev/null
@@ -1,76 +0,0 @@
-// sherpa-onnx/csrc/endpoint.h
-//
-// Copyright (c)  2022  (authors: Pingfeng Luo)
-//                2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ENDPOINT_H_
-#define SHERPA_ONNX_CSRC_ENDPOINT_H_
-
-#include <string>
-#include <vector>
-
-namespace sherpa_onnx {
-
-struct EndpointRule {
-  // If True, for this endpointing rule to apply there must
-  // be nonsilence in the best-path traceback.
-  // For decoding, a non-blank token is considered as non-silence
-  bool must_contain_nonsilence = true;
-  // This endpointing rule requires duration of trailing silence
-  // (in seconds) to be >= this value.
-  float min_trailing_silence = 2.0;
-  // This endpointing rule requires utterance-length (in seconds)
-  // to be >= this value.
-  float min_utterance_length = 0.0f;
-
-  EndpointRule() = default;
-
-  EndpointRule(bool must_contain_nonsilence, float min_trailing_silence,
-               float min_utterance_length)
-      : must_contain_nonsilence(must_contain_nonsilence),
-        min_trailing_silence(min_trailing_silence),
-        min_utterance_length(min_utterance_length) {}
-
-  std::string ToString() const;
-};
-
-class ParseOptions;
-
-struct EndpointConfig {
-  // For default setting,
-  // rule1 times out after 2.4 seconds of silence, even if we decoded nothing.
-  // rule2 times out after 1.2 seconds of silence after decoding something.
-  // rule3 times out after the utterance is 20 seconds long, regardless of
-  // anything else.
-  EndpointRule rule1;
-  EndpointRule rule2;
-  EndpointRule rule3;
-
-  void Register(ParseOptions *po);
-
-  EndpointConfig()
-      : rule1{false, 2.4, 0}, rule2{true, 1.2, 0}, rule3{false, 0, 20} {}
-
-  EndpointConfig(const EndpointRule &rule1, const EndpointRule &rule2,
-                 const EndpointRule &rule3)
-      : rule1(rule1), rule2(rule2), rule3(rule3) {}
-
-  std::string ToString() const;
-};
-
-class Endpoint {
- public:
-  explicit Endpoint(const EndpointConfig &config) : config_(config) {}
-
-  /// This function returns true if this set of endpointing rules thinks we
-  /// should terminate decoding.
-  bool IsEndpoint(int32_t num_frames_decoded, int32_t trailing_silence_frames,
-                  float frame_shift_in_seconds) const;
-
- private:
-  EndpointConfig config_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ENDPOINT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/fast-clustering-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/fast-clustering-config.h
deleted file mode 100644
index 4abf2b12..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/fast-clustering-config.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// sherpa-onnx/csrc/fast-clustering-config.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_FAST_CLUSTERING_CONFIG_H_
-#define SHERPA_ONNX_CSRC_FAST_CLUSTERING_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct FastClusteringConfig {
-  // If greater than 0, then threshold is ignored.
-  //
-  // We strongly recommend that you set it if you know the number of clusters
-  // in advance
-  int32_t num_clusters = -1;
-
-  // distance threshold.
-  //
-  // The smaller, the more clusters it will generate.
-  // The larger, the fewer clusters it will generate.
-  float threshold = 0.5;
-
-  FastClusteringConfig() = default;
-
-  FastClusteringConfig(int32_t num_clusters, float threshold)
-      : num_clusters(num_clusters), threshold(threshold) {}
-
-  std::string ToString() const;
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-};
-
-}  // namespace sherpa_onnx
-#endif  // SHERPA_ONNX_CSRC_FAST_CLUSTERING_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/fast-clustering.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/fast-clustering.h
deleted file mode 100644
index d6ffd847..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/fast-clustering.h
+++ /dev/null
@@ -1,43 +0,0 @@
-// sherpa-onnx/csrc/fast-clustering.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_FAST_CLUSTERING_H_
-#define SHERPA_ONNX_CSRC_FAST_CLUSTERING_H_
-
-#include <memory>
-#include <vector>
-
-#include "sherpa-onnx/csrc/fast-clustering-config.h"
-
-namespace sherpa_onnx {
-
-class FastClustering {
- public:
-  explicit FastClustering(const FastClusteringConfig &config);
-  ~FastClustering();
-
-  /**
-   * @param features Pointer to a 2-D feature matrix in row major. Each row
-   *                 is a feature frame. It is changed in-place. We will
-   *                 convert each feature frame to a normalized vector.
-   *                 That is, the L2-norm of each vector will be equal to 1.
-   *                 It uses cosine dissimilarity,
-   *                 which is 1 - (cosine similarity)
-   * @param num_rows Number of feature frames
-   * @param num-cols The feature dimension.
-   *
-   * @return Return a vector of size num_rows. ans[i] contains the label
-   *         for the i-th feature frame, i.e., the i-th row of the feature
-   *         matrix.
-   */
-  std::vector<int32_t> Cluster(float *features, int32_t num_rows,
-                               int32_t num_cols) const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-#endif  // SHERPA_ONNX_CSRC_FAST_CLUSTERING_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/features.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/features.h
deleted file mode 100644
index fb5ff2fe..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/features.h
+++ /dev/null
@@ -1,137 +0,0 @@
-// sherpa-onnx/csrc/features.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_FEATURES_H_
-#define SHERPA_ONNX_CSRC_FEATURES_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct FeatureExtractorConfig {
-  // Sampling rate used by the feature extractor. If it is different from
-  // the sampling rate of the input waveform, we will do resampling inside.
-  int32_t sampling_rate = 16000;
-
-  // num_mel_bins
-  //
-  // Note: for mfcc, this value is also for num_mel_bins.
-  // The actual feature dimension is actuall num_ceps
-  int32_t feature_dim = 80;
-
-  // minimal frequency for Mel-filterbank, in Hz
-  float low_freq = 20.0f;
-
-  // maximal frequency of Mel-filterbank
-  // in Hz; negative value is subtracted from Nyquist freq.:
-  // i.e. for sampling_rate 16000 / 2 - 400 = 7600Hz
-  //
-  // Please see
-  // https://github.com/lhotse-speech/lhotse/blob/master/lhotse/features/fbank.py#L27
-  // and
-  // https://github.com/k2-fsa/sherpa-onnx/issues/514
-  float high_freq = -400.0f;
-
-  // dithering constant, useful for signals with hard-zeroes in non-speech parts
-  // this prevents large negative values in log-mel filterbanks
-  //
-  // In k2, audio samples are in range [-1..+1], in kaldi the range was
-  // [-32k..+32k], so the value 0.00003 is equivalent to kaldi default 1.0
-  //
-  float dither = 0.0f;  // dithering disabled by default
-
-  // Set internally by some models, e.g., paraformer sets it to false.
-  // This parameter is not exposed to users from the commandline
-  // If true, the feature extractor expects inputs to be normalized to
-  // the range [-1, 1].
-  // If false, we will multiply the inputs by 32768
-  bool normalize_samples = true;
-
-  bool snip_edges = false;
-  float frame_shift_ms = 10.0f;   // in milliseconds.
-  float frame_length_ms = 25.0f;  // in milliseconds.
-  bool is_librosa = false;
-  bool remove_dc_offset = true;       // Subtract mean of wave before FFT.
-  float preemph_coeff = 0.97f;        // Preemphasis coefficient.
-  std::string window_type = "povey";  // e.g. Hamming window
-
-  // For models from NeMo
-  // This option is not exposed and is set internally when loading models.
-  // Possible values:
-  // - per_feature
-  // - all_features (not implemented yet)
-  // - fixed_mean (not implemented)
-  // - fixed_std (not implemented)
-  // - or just leave it to empty
-  // See
-  // https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/parts/preprocessing/features.py#L59
-  // for details
-  std::string nemo_normalize_type;
-
-  // for MFCC
-  int32_t num_ceps = 13;
-  bool use_energy = true;
-
-  bool is_mfcc = false;
-
-  std::string ToString() const;
-
-  void Register(ParseOptions *po);
-};
-
-class FeatureExtractor {
- public:
-  explicit FeatureExtractor(const FeatureExtractorConfig &config = {});
-  ~FeatureExtractor();
-
-  /**
-     @param sampling_rate The sampling_rate of the input waveform. If it does
-                          not equal to  config.sampling_rate, we will do
-                          resampling inside.
-     @param waveform Pointer to a 1-D array of size n. It must be normalized to
-                     the range [-1, 1].
-     @param n Number of entries in waveform
-   */
-  void AcceptWaveform(int32_t sampling_rate, const float *waveform,
-                      int32_t n) const;
-
-  /**
-   * InputFinished() tells the class you won't be providing any
-   * more waveform.  This will help flush out the last frame or two
-   * of features, in the case where snip-edges == false; it also
-   * affects the return value of IsLastFrame().
-   */
-  void InputFinished() const;
-
-  int32_t NumFramesReady() const;
-
-  /** Note: IsLastFrame() will only ever return true if you have called
-   * InputFinished() (and this frame is the last frame).
-   */
-  bool IsLastFrame(int32_t frame) const;
-
-  /** Get n frames starting from the given frame index.
-   *
-   * @param frame_index  The starting frame index
-   * @param n  Number of frames to get.
-   * @return Return a 2-D tensor of shape (n, feature_dim).
-   *         which is flattened into a 1-D vector (flattened in row major)
-   */
-  std::vector<float> GetFrames(int32_t frame_index, int32_t n) const;
-
-  /// Return feature dim of this extractor
-  int32_t FeatureDim() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_FEATURES_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/file-utils.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/file-utils.h
deleted file mode 100644
index a41f6c9c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/file-utils.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// sherpa-onnx/csrc/file-utils.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_FILE_UTILS_H_
-#define SHERPA_ONNX_CSRC_FILE_UTILS_H_
-
-#include <fstream>
-#include <string>
-
-namespace sherpa_onnx {
-
-/** Check whether a given path is a file or not
- *
- * @param filename Path to check.
- * @return Return true if the given path is a file; return false otherwise.
- */
-bool FileExists(const std::string &filename);
-
-/** Abort if the file does not exist.
- *
- * @param filename The file to check.
- */
-void AssertFileExists(const std::string &filename);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_FILE_UTILS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/fst-utils.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/fst-utils.h
deleted file mode 100644
index 92e6a7aa..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/fst-utils.h
+++ /dev/null
@@ -1,18 +0,0 @@
-// sherpa-onnx/csrc/fst-utils.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_FST_UTILS_H_
-#define SHERPA_ONNX_CSRC_FST_UTILS_H_
-
-#include <string>
-
-#include "fst/fstlib.h"
-
-namespace sherpa_onnx {
-
-fst::Fst<fst::StdArc> *ReadGraph(const std::string &filename);
-
-}
-
-#endif  // SHERPA_ONNX_CSRC_FST_UTILS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/hypothesis.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/hypothesis.h
deleted file mode 100644
index 6a49bad3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/hypothesis.h
+++ /dev/null
@@ -1,166 +0,0 @@
-/**
- * Copyright (c)  2023  Xiaomi Corporation
- * Copyright (c)  2023  Pingfeng Luo
- *
- */
-
-#ifndef SHERPA_ONNX_CSRC_HYPOTHESIS_H_
-#define SHERPA_ONNX_CSRC_HYPOTHESIS_H_
-
-#include <sstream>
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/context-graph.h"
-#include "sherpa-onnx/csrc/math.h"
-#include "sherpa-onnx/csrc/onnx-utils.h"
-
-namespace sherpa_onnx {
-
-struct Hypothesis {
-  // The predicted tokens so far. Newly predicated tokens are appended.
-  std::vector<int64_t> ys;
-
-  // timestamps[i] contains the frame number after subsampling
-  // on which ys[i] is decoded.
-  std::vector<int32_t> timestamps;
-
-  // The acoustic probability for each token in ys.
-  // Used for keyword spotting task.
-  // For transducer mofified beam-search and greedy-search,
-  // this is filled with log_posterior scores.
-  std::vector<float> ys_probs;
-
-  // lm_probs[i] contains the lm score for each token in ys.
-  // Used only in transducer mofified beam-search.
-  // Elements filled only if LM is used.
-  std::vector<float> lm_probs;
-
-  // context_scores[i] contains the context-graph score for each token in ys.
-  // Used only in transducer mofified beam-search.
-  // Elements filled only if `ContextGraph` is used.
-  std::vector<float> context_scores;
-
-  // The total score of ys in log space.
-  // It contains only acoustic scores
-  double log_prob = 0;
-
-  // LM log prob if any.
-  double lm_log_prob = 0;
-
-  // the nn lm score for next token given the current ys,
-  // when using shallow fusion
-  CopyableOrtValue nn_lm_scores;
-
-  // cur scored tokens by RNN LM, when rescoring
-  int32_t cur_scored_pos = 0;
-
-  // the nn lm states
-  std::vector<CopyableOrtValue> nn_lm_states;
-
-  const ContextState *context_state;
-
-  // TODO(fangjun): Make it configurable
-  // the minimum of tokens in a chunk for streaming RNN LM
-  int32_t lm_rescore_min_chunk = 2;  // a const
-
-  int32_t num_trailing_blanks = 0;
-
-  Hypothesis() = default;
-  Hypothesis(const std::vector<int64_t> &ys, double log_prob,
-             const ContextState *context_state = nullptr)
-      : ys(ys), log_prob(log_prob), context_state(context_state) {}
-
-  double TotalLogProb() const { return log_prob + lm_log_prob; }
-
-  // If two Hypotheses have the same `Key`, then they contain
-  // the same token sequence.
-  std::string Key() const {
-    // TODO(fangjun): Use a hash function?
-    std::ostringstream os;
-    std::string sep;
-    for (auto i : ys) {
-      os << sep << i;
-      sep = "-";
-    }
-    return os.str();
-  }
-
-  // For debugging
-  std::string ToString() const {
-    std::ostringstream os;
-    os << "(" << Key() << ", " << log_prob << ")";
-    return os.str();
-  }
-};
-
-class Hypotheses {
- public:
-  Hypotheses() = default;
-
-  explicit Hypotheses(std::vector<Hypothesis> hyps) {
-    for (auto &h : hyps) {
-      hyps_dict_[h.Key()] = std::move(h);
-    }
-  }
-
-  explicit Hypotheses(std::unordered_map<std::string, Hypothesis> hyps_dict)
-      : hyps_dict_(std::move(hyps_dict)) {}
-
-  // Add hyp to this object. If it already exists, its log_prob
-  // is updated with the given hyp using log-sum-exp.
-  void Add(Hypothesis hyp);
-
-  // Get the hyp that has the largest log_prob.
-  // If length_norm is true, hyp's log_prob is divided by
-  // len(hyp.ys) before comparison.
-  Hypothesis GetMostProbable(bool length_norm) const;
-
-  // Get the k hyps that have the largest log_prob.
-  // If length_norm is true, hyp's log_prob is divided by
-  // len(hyp.ys) before comparison.
-  std::vector<Hypothesis> GetTopK(int32_t k, bool length_norm) const;
-
-  int32_t Size() const { return hyps_dict_.size(); }
-
-  std::string ToString() const {
-    std::ostringstream os;
-    for (const auto &p : hyps_dict_) {
-      os << p.second.ToString() << "\n";
-    }
-    return os.str();
-  }
-
-  auto begin() const { return hyps_dict_.begin(); }
-  auto end() const { return hyps_dict_.end(); }
-
-  auto begin() { return hyps_dict_.begin(); }
-  auto end() { return hyps_dict_.end(); }
-
-  void Clear() { hyps_dict_.clear(); }
-
- private:
-  // Return a list of hyps contained in this object.
-  std::vector<Hypothesis> Vec() const {
-    std::vector<Hypothesis> ans;
-    ans.reserve(hyps_dict_.size());
-    for (const auto &p : hyps_dict_) {
-      ans.push_back(p.second);
-    }
-    return ans;
-  }
-
- private:
-  using Map = std ::unordered_map<std::string, Hypothesis>;
-  Map hyps_dict_;
-};
-
-const std::vector<int32_t> GetHypsRowSplits(
-    const std::vector<Hypotheses> &hyps);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_HYPOTHESIS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/jieba-lexicon.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/jieba-lexicon.h
deleted file mode 100644
index d02e0ee5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/jieba-lexicon.h
+++ /dev/null
@@ -1,36 +0,0 @@
-// sherpa-onnx/csrc/jieba-lexicon.h
-//
-// Copyright (c)  2022-2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_JIEBA_LEXICON_H_
-#define SHERPA_ONNX_CSRC_JIEBA_LEXICON_H_
-
-#include <memory>
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-tts-frontend.h"
-#include "sherpa-onnx/csrc/offline-tts-vits-model-metadata.h"
-
-namespace sherpa_onnx {
-
-class JiebaLexicon : public OfflineTtsFrontend {
- public:
-  ~JiebaLexicon() override;
-  JiebaLexicon(const std::string &lexicon, const std::string &tokens,
-               const std::string &dict_dir,
-               const OfflineTtsVitsModelMetaData &meta_data, bool debug);
-
-  std::vector<TokenIDs> ConvertTextToTokenIds(
-      const std::string &text,
-      const std::string &unused_voice = "") const override;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_JIEBA_LEXICON_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/keyword-spotter-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/keyword-spotter-impl.h
deleted file mode 100644
index ded735ff..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/keyword-spotter-impl.h
+++ /dev/null
@@ -1,48 +0,0 @@
-// sherpa-onnx/csrc/keyword-spotter-impl.h
-//
-// Copyright (c)  2023-2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_KEYWORD_SPOTTER_IMPL_H_
-#define SHERPA_ONNX_CSRC_KEYWORD_SPOTTER_IMPL_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "sherpa-onnx/csrc/keyword-spotter.h"
-#include "sherpa-onnx/csrc/online-stream.h"
-
-namespace sherpa_onnx {
-
-class KeywordSpotterImpl {
- public:
-  static std::unique_ptr<KeywordSpotterImpl> Create(
-      const KeywordSpotterConfig &config);
-
-#if __ANDROID_API__ >= 9
-  static std::unique_ptr<KeywordSpotterImpl> Create(
-      AAssetManager *mgr, const KeywordSpotterConfig &config);
-#endif
-
-  virtual ~KeywordSpotterImpl() = default;
-
-  virtual std::unique_ptr<OnlineStream> CreateStream() const = 0;
-
-  virtual std::unique_ptr<OnlineStream> CreateStream(
-      const std::string &keywords) const = 0;
-
-  virtual bool IsReady(OnlineStream *s) const = 0;
-
-  virtual void DecodeStreams(OnlineStream **ss, int32_t n) const = 0;
-
-  virtual KeywordResult GetResult(OnlineStream *s) const = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_KEYWORD_SPOTTER_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/keyword-spotter-transducer-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/keyword-spotter-transducer-impl.h
deleted file mode 100644
index 75963918..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/keyword-spotter-transducer-impl.h
+++ /dev/null
@@ -1,350 +0,0 @@
-// sherpa-onnx/csrc/keyword-spotter-transducer-impl.h
-//
-// Copyright (c)  2023-2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_KEYWORD_SPOTTER_TRANSDUCER_IMPL_H_
-#define SHERPA_ONNX_CSRC_KEYWORD_SPOTTER_TRANSDUCER_IMPL_H_
-
-#include <algorithm>
-#include <memory>
-#include <regex>  // NOLINT
-#include <string>
-#include <utility>
-#include <vector>
-
-#if __ANDROID_API__ >= 9
-#include <strstream>
-
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "sherpa-onnx/csrc/file-utils.h"
-#include "sherpa-onnx/csrc/keyword-spotter-impl.h"
-#include "sherpa-onnx/csrc/keyword-spotter.h"
-#include "sherpa-onnx/csrc/macros.h"
-#include "sherpa-onnx/csrc/online-transducer-model.h"
-#include "sherpa-onnx/csrc/symbol-table.h"
-#include "sherpa-onnx/csrc/transducer-keyword-decoder.h"
-#include "sherpa-onnx/csrc/utils.h"
-
-namespace sherpa_onnx {
-
-static KeywordResult Convert(const TransducerKeywordResult &src,
-                             const SymbolTable &sym_table, float frame_shift_ms,
-                             int32_t subsampling_factor,
-                             int32_t frames_since_start) {
-  KeywordResult r;
-  r.tokens.reserve(src.tokens.size());
-  r.timestamps.reserve(src.tokens.size());
-  r.keyword = src.keyword;
-  bool from_tokens = src.keyword.empty();
-
-  for (auto i : src.tokens) {
-    auto sym = sym_table[i];
-    if (from_tokens) {
-      r.keyword.append(sym);
-    }
-    r.tokens.push_back(std::move(sym));
-  }
-  if (from_tokens && r.keyword.size()) {
-    r.keyword = r.keyword.substr(1);
-  }
-
-  float frame_shift_s = frame_shift_ms / 1000. * subsampling_factor;
-  for (auto t : src.timestamps) {
-    float time = frame_shift_s * t;
-    r.timestamps.push_back(time);
-  }
-
-  r.start_time = frames_since_start * frame_shift_ms / 1000.;
-
-  return r;
-}
-
-class KeywordSpotterTransducerImpl : public KeywordSpotterImpl {
- public:
-  explicit KeywordSpotterTransducerImpl(const KeywordSpotterConfig &config)
-      : config_(config),
-        model_(OnlineTransducerModel::Create(config.model_config)) {
-    if (!config.model_config.tokens_buf.empty()) {
-      sym_ = SymbolTable(config.model_config.tokens_buf, false);
-    } else {
-      /// assuming tokens_buf and tokens are guaranteed not being both empty
-      sym_ = SymbolTable(config.model_config.tokens, true);
-    }
-
-    if (sym_.Contains("<unk>")) {
-      unk_id_ = sym_["<unk>"];
-    }
-
-    model_->SetFeatureDim(config.feat_config.feature_dim);
-
-    if (config.keywords_buf.empty()) {
-      InitKeywords();
-    } else {
-      InitKeywordsFromBufStr();
-    }
-
-    decoder_ = std::make_unique<TransducerKeywordDecoder>(
-        model_.get(), config_.max_active_paths, config_.num_trailing_blanks,
-        unk_id_);
-  }
-
-#if __ANDROID_API__ >= 9
-  KeywordSpotterTransducerImpl(AAssetManager *mgr,
-                               const KeywordSpotterConfig &config)
-      : config_(config),
-        model_(OnlineTransducerModel::Create(mgr, config.model_config)),
-        sym_(mgr, config.model_config.tokens) {
-    if (sym_.Contains("<unk>")) {
-      unk_id_ = sym_["<unk>"];
-    }
-
-    model_->SetFeatureDim(config.feat_config.feature_dim);
-
-    InitKeywords(mgr);
-
-    decoder_ = std::make_unique<TransducerKeywordDecoder>(
-        model_.get(), config_.max_active_paths, config_.num_trailing_blanks,
-        unk_id_);
-  }
-#endif
-
-  std::unique_ptr<OnlineStream> CreateStream() const override {
-    auto stream =
-        std::make_unique<OnlineStream>(config_.feat_config, keywords_graph_);
-    InitOnlineStream(stream.get());
-    return stream;
-  }
-
-  std::unique_ptr<OnlineStream> CreateStream(
-      const std::string &keywords) const override {
-    auto kws = std::regex_replace(keywords, std::regex("/"), "\n");
-    std::istringstream is(kws);
-
-    std::vector<std::vector<int32_t>> current_ids;
-    std::vector<std::string> current_kws;
-    std::vector<float> current_scores;
-    std::vector<float> current_thresholds;
-
-    if (!EncodeKeywords(is, sym_, &current_ids, &current_kws, &current_scores,
-                        &current_thresholds)) {
-      SHERPA_ONNX_LOGE("Encode keywords %s failed.", keywords.c_str());
-      return nullptr;
-    }
-
-    int32_t num_kws = current_ids.size();
-    int32_t num_default_kws = keywords_id_.size();
-
-    current_ids.insert(current_ids.end(), keywords_id_.begin(),
-                       keywords_id_.end());
-
-    if (!current_kws.empty() && !keywords_.empty()) {
-      current_kws.insert(current_kws.end(), keywords_.begin(), keywords_.end());
-    } else if (!current_kws.empty() && keywords_.empty()) {
-      current_kws.insert(current_kws.end(), num_default_kws, std::string());
-    } else if (current_kws.empty() && !keywords_.empty()) {
-      current_kws.insert(current_kws.end(), num_kws, std::string());
-      current_kws.insert(current_kws.end(), keywords_.begin(), keywords_.end());
-    } else {
-      // Do nothing.
-    }
-
-    if (!current_scores.empty() && !boost_scores_.empty()) {
-      current_scores.insert(current_scores.end(), boost_scores_.begin(),
-                            boost_scores_.end());
-    } else if (!current_scores.empty() && boost_scores_.empty()) {
-      current_scores.insert(current_scores.end(), num_default_kws,
-                            config_.keywords_score);
-    } else if (current_scores.empty() && !boost_scores_.empty()) {
-      current_scores.insert(current_scores.end(), num_kws,
-                            config_.keywords_score);
-      current_scores.insert(current_scores.end(), boost_scores_.begin(),
-                            boost_scores_.end());
-    } else {
-      // Do nothing.
-    }
-
-    if (!current_thresholds.empty() && !thresholds_.empty()) {
-      current_thresholds.insert(current_thresholds.end(), thresholds_.begin(),
-                                thresholds_.end());
-    } else if (!current_thresholds.empty() && thresholds_.empty()) {
-      current_thresholds.insert(current_thresholds.end(), num_default_kws,
-                                config_.keywords_threshold);
-    } else if (current_thresholds.empty() && !thresholds_.empty()) {
-      current_thresholds.insert(current_thresholds.end(), num_kws,
-                                config_.keywords_threshold);
-      current_thresholds.insert(current_thresholds.end(), thresholds_.begin(),
-                                thresholds_.end());
-    } else {
-      // Do nothing.
-    }
-
-    auto keywords_graph = std::make_shared<ContextGraph>(
-        current_ids, config_.keywords_score, config_.keywords_threshold,
-        current_scores, current_kws, current_thresholds);
-
-    auto stream =
-        std::make_unique<OnlineStream>(config_.feat_config, keywords_graph);
-    InitOnlineStream(stream.get());
-    return stream;
-  }
-
-  bool IsReady(OnlineStream *s) const override {
-    return s->GetNumProcessedFrames() + model_->ChunkSize() <
-           s->NumFramesReady();
-  }
-
-  void DecodeStreams(OnlineStream **ss, int32_t n) const override {
-    int32_t chunk_size = model_->ChunkSize();
-    int32_t chunk_shift = model_->ChunkShift();
-
-    int32_t feature_dim = ss[0]->FeatureDim();
-
-    std::vector<TransducerKeywordResult> results(n);
-    std::vector<float> features_vec(n * chunk_size * feature_dim);
-    std::vector<std::vector<Ort::Value>> states_vec(n);
-    std::vector<int64_t> all_processed_frames(n);
-
-    for (int32_t i = 0; i != n; ++i) {
-      SHERPA_ONNX_CHECK(ss[i]->GetContextGraph() != nullptr);
-
-      const auto num_processed_frames = ss[i]->GetNumProcessedFrames();
-      std::vector<float> features =
-          ss[i]->GetFrames(num_processed_frames, chunk_size);
-
-      // Question: should num_processed_frames include chunk_shift?
-      ss[i]->GetNumProcessedFrames() += chunk_shift;
-
-      std::copy(features.begin(), features.end(),
-                features_vec.data() + i * chunk_size * feature_dim);
-
-      results[i] = std::move(ss[i]->GetKeywordResult());
-      states_vec[i] = std::move(ss[i]->GetStates());
-      all_processed_frames[i] = num_processed_frames;
-    }
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    std::array<int64_t, 3> x_shape{n, chunk_size, feature_dim};
-
-    Ort::Value x = Ort::Value::CreateTensor(memory_info, features_vec.data(),
-                                            features_vec.size(), x_shape.data(),
-                                            x_shape.size());
-
-    std::array<int64_t, 1> processed_frames_shape{
-        static_cast<int64_t>(all_processed_frames.size())};
-
-    Ort::Value processed_frames = Ort::Value::CreateTensor(
-        memory_info, all_processed_frames.data(), all_processed_frames.size(),
-        processed_frames_shape.data(), processed_frames_shape.size());
-
-    auto states = model_->StackStates(states_vec);
-
-    auto pair = model_->RunEncoder(std::move(x), std::move(states),
-                                   std::move(processed_frames));
-
-    decoder_->Decode(std::move(pair.first), ss, &results);
-
-    std::vector<std::vector<Ort::Value>> next_states =
-        model_->UnStackStates(pair.second);
-
-    for (int32_t i = 0; i != n; ++i) {
-      ss[i]->SetKeywordResult(results[i]);
-      ss[i]->SetStates(std::move(next_states[i]));
-    }
-  }
-
-  KeywordResult GetResult(OnlineStream *s) const override {
-    TransducerKeywordResult decoder_result = s->GetKeywordResult(true);
-
-    // TODO(fangjun): Remember to change these constants if needed
-    int32_t frame_shift_ms = 10;
-    int32_t subsampling_factor = 4;
-    return Convert(decoder_result, sym_, frame_shift_ms, subsampling_factor,
-                   s->GetNumFramesSinceStart());
-  }
-
- private:
-  void InitKeywords(std::istream &is) {
-    if (!EncodeKeywords(is, sym_, &keywords_id_, &keywords_, &boost_scores_,
-                        &thresholds_)) {
-      SHERPA_ONNX_LOGE("Encode keywords failed.");
-      exit(-1);
-    }
-    keywords_graph_ = std::make_shared<ContextGraph>(
-        keywords_id_, config_.keywords_score, config_.keywords_threshold,
-        boost_scores_, keywords_, thresholds_);
-  }
-
-  void InitKeywords() {
-#ifdef SHERPA_ONNX_ENABLE_WASM_KWS
-    // Due to the limitations of the wasm file system,
-    // the keyword_file variable is directly parsed as a string of keywords
-    // if WASM KWS on
-    std::istringstream is(config_.keywords_file);
-    InitKeywords(is);
-#else
-    // each line in keywords_file contains space-separated words
-    std::ifstream is(config_.keywords_file);
-    if (!is) {
-      SHERPA_ONNX_LOGE("Open keywords file failed: %s",
-                       config_.keywords_file.c_str());
-      exit(-1);
-    }
-    InitKeywords(is);
-#endif
-  }
-
-#if __ANDROID_API__ >= 9
-  void InitKeywords(AAssetManager *mgr) {
-    // each line in keywords_file contains space-separated words
-
-    auto buf = ReadFile(mgr, config_.keywords_file);
-
-    std::istrstream is(buf.data(), buf.size());
-
-    if (!is) {
-      SHERPA_ONNX_LOGE("Open keywords file failed: %s",
-                       config_.keywords_file.c_str());
-      exit(-1);
-    }
-    InitKeywords(is);
-  }
-#endif
-
-  void InitKeywordsFromBufStr() {
-    // keywords_buf's content is supposed to be same as the keywords_file's
-    std::istringstream is(config_.keywords_buf);
-    InitKeywords(is);
-  }
-
-  void InitOnlineStream(OnlineStream *stream) const {
-    auto r = decoder_->GetEmptyResult();
-    SHERPA_ONNX_CHECK_EQ(r.hyps.Size(), 1);
-
-    SHERPA_ONNX_CHECK(stream->GetContextGraph() != nullptr);
-    r.hyps.begin()->second.context_state = stream->GetContextGraph()->Root();
-
-    stream->SetKeywordResult(r);
-    stream->SetStates(model_->GetEncoderInitStates());
-  }
-
- private:
-  KeywordSpotterConfig config_;
-  std::vector<std::vector<int32_t>> keywords_id_;
-  std::vector<float> boost_scores_;
-  std::vector<float> thresholds_;
-  std::vector<std::string> keywords_;
-  ContextGraphPtr keywords_graph_;
-  std::unique_ptr<OnlineTransducerModel> model_;
-  std::unique_ptr<TransducerKeywordDecoder> decoder_;
-  SymbolTable sym_;
-  int32_t unk_id_ = -1;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_KEYWORD_SPOTTER_TRANSDUCER_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/keyword-spotter.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/keyword-spotter.h
deleted file mode 100644
index f0c31bdb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/keyword-spotter.h
+++ /dev/null
@@ -1,153 +0,0 @@
-// sherpa-onnx/csrc/keyword-spotter.h
-//
-// Copyright (c)  2023-2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_KEYWORD_SPOTTER_H_
-#define SHERPA_ONNX_CSRC_KEYWORD_SPOTTER_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "sherpa-onnx/csrc/features.h"
-#include "sherpa-onnx/csrc/online-model-config.h"
-#include "sherpa-onnx/csrc/online-stream.h"
-#include "sherpa-onnx/csrc/online-transducer-model-config.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct KeywordResult {
-  /// The triggered keyword.
-  /// For English, it consists of space separated words.
-  /// For Chinese, it consists of Chinese words without spaces.
-  /// Example 1: "hello world"
-  /// Example 2: "你好世界"
-  std::string keyword;
-
-  /// Decoded results at the token level.
-  /// For instance, for BPE-based models it consists of a list of BPE tokens.
-  std::vector<std::string> tokens;
-
-  /// timestamps.size() == tokens.size()
-  /// timestamps[i] records the time in seconds when tokens[i] is decoded.
-  std::vector<float> timestamps;
-
-  /// Starting time of this segment.
-  /// When an endpoint is detected, it will change
-  float start_time = 0;
-
-  /** Return a json string.
-   *
-   * The returned string contains:
-   *   {
-   *     "keyword": "The triggered keyword",
-   *     "tokens": [x, x, x],
-   *     "timestamps": [x, x, x],
-   *     "start_time": x,
-   *   }
-   */
-  std::string AsJsonString() const;
-};
-
-struct KeywordSpotterConfig {
-  FeatureExtractorConfig feat_config;
-  OnlineModelConfig model_config;
-
-  int32_t max_active_paths = 4;
-
-  int32_t num_trailing_blanks = 1;
-
-  float keywords_score = 1.0;
-
-  float keywords_threshold = 0.25;
-
-  std::string keywords_file;
-
-  /// if keywords_buf is non-empty,
-  /// the keywords will be loaded from the buffer instead of from the
-  /// "keywrods_file"
-  std::string keywords_buf;
-
-  KeywordSpotterConfig() = default;
-
-  KeywordSpotterConfig(const FeatureExtractorConfig &feat_config,
-                       const OnlineModelConfig &model_config,
-                       int32_t max_active_paths, int32_t num_trailing_blanks,
-                       float keywords_score, float keywords_threshold,
-                       const std::string &keywords_file)
-      : feat_config(feat_config),
-        model_config(model_config),
-        max_active_paths(max_active_paths),
-        num_trailing_blanks(num_trailing_blanks),
-        keywords_score(keywords_score),
-        keywords_threshold(keywords_threshold),
-        keywords_file(keywords_file) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-class KeywordSpotterImpl;
-
-class KeywordSpotter {
- public:
-  explicit KeywordSpotter(const KeywordSpotterConfig &config);
-
-#if __ANDROID_API__ >= 9
-  KeywordSpotter(AAssetManager *mgr, const KeywordSpotterConfig &config);
-#endif
-
-  ~KeywordSpotter();
-
-  /** Create a stream for decoding.
-   *
-   */
-  std::unique_ptr<OnlineStream> CreateStream() const;
-
-  /** Create a stream for decoding.
-   *
-   *  @param The keywords for this string, it might contain several keywords,
-   *         the keywords are separated by "/". In each of the keywords, there
-   *         are cjkchars or bpes, the bpe/cjkchar are separated by space (" ").
-   *         For example, keywords I LOVE YOU and HELLO WORLD, looks like:
-   *
-   *         "▁I ▁LOVE ▁YOU/▁HE LL O ▁WORLD"
-   */
-  std::unique_ptr<OnlineStream> CreateStream(const std::string &keywords) const;
-
-  /**
-   * Return true if the given stream has enough frames for decoding.
-   * Return false otherwise
-   */
-  bool IsReady(OnlineStream *s) const;
-
-  /** Decode a single stream. */
-  void DecodeStream(OnlineStream *s) const {
-    OnlineStream *ss[1] = {s};
-    DecodeStreams(ss, 1);
-  }
-
-  /** Decode multiple streams in parallel
-   *
-   * @param ss Pointer array containing streams to be decoded.
-   * @param n Number of streams in `ss`.
-   */
-  void DecodeStreams(OnlineStream **ss, int32_t n) const;
-
-  KeywordResult GetResult(OnlineStream *s) const;
-
- private:
-  std::unique_ptr<KeywordSpotterImpl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_KEYWORD_SPOTTER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/lexicon.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/lexicon.h
deleted file mode 100644
index 39329694..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/lexicon.h
+++ /dev/null
@@ -1,66 +0,0 @@
-// sherpa-onnx/csrc/lexicon.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_LEXICON_H_
-#define SHERPA_ONNX_CSRC_LEXICON_H_
-
-#include <cstdint>
-#include <istream>
-#include <memory>
-#include <string>
-#include <unordered_map>
-#include <unordered_set>
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-tts-frontend.h"
-
-namespace sherpa_onnx {
-
-class Lexicon : public OfflineTtsFrontend {
- public:
-  Lexicon() = default;  // for subclasses
-                        //
-  // Note: for models from piper, we won't use this class.
-  Lexicon(const std::string &lexicon, const std::string &tokens,
-          const std::string &punctuations, const std::string &language,
-          bool debug = false);
-
-  template <typename Manager>
-  Lexicon(Manager *mgr, const std::string &lexicon, const std::string &tokens,
-          const std::string &punctuations, const std::string &language,
-          bool debug = false);
-
-  std::vector<TokenIDs> ConvertTextToTokenIds(
-      const std::string &text, const std::string &voice = "") const override;
-
- private:
-  std::vector<TokenIDs> ConvertTextToTokenIdsNotChinese(
-      const std::string &text) const;
-
-  std::vector<TokenIDs> ConvertTextToTokenIdsChinese(
-      const std::string &text) const;
-
-  void InitLanguage(const std::string &lang);
-  void InitTokens(std::istream &is);
-  void InitLexicon(std::istream &is);
-  void InitPunctuations(const std::string &punctuations);
-
- private:
-  enum class Language {
-    kNotChinese,
-    kChinese,
-    kUnknown,
-  };
-
- private:
-  std::unordered_map<std::string, std::vector<int32_t>> word2ids_;
-  std::unordered_set<std::string> punctuations_;
-  std::unordered_map<std::string, int32_t> token2id_;
-  Language language_ = Language::kUnknown;
-  bool debug_ = false;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_LEXICON_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/log.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/log.h
deleted file mode 100644
index d2d29fe0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/log.h
+++ /dev/null
@@ -1,378 +0,0 @@
-// sherpa-onnx/csrc/log.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_LOG_H_
-#define SHERPA_ONNX_CSRC_LOG_H_
-
-#include <stdio.h>
-
-#include <mutex>  // NOLINT
-#include <sstream>
-#include <string>
-
-namespace sherpa_onnx {
-
-#if SHERPA_ONNX_ENABLE_CHECK
-
-#if defined(NDEBUG)
-constexpr bool kDisableDebug = true;
-#else
-constexpr bool kDisableDebug = false;
-#endif
-
-enum class LogLevel {
-  kTrace = 0,
-  kDebug = 1,
-  kInfo = 2,
-  kWarning = 3,
-  kError = 4,
-  kFatal = 5,  // print message and abort the program
-};
-
-// They are used in SHERPA_ONNX_LOG(xxx), so their names
-// do not follow the google c++ code style
-//
-// You can use them in the following way:
-//
-//  SHERPA_ONNX_LOG(TRACE) << "some message";
-//  SHERPA_ONNX_LOG(DEBUG) << "some message";
-#ifndef _MSC_VER
-constexpr LogLevel TRACE = LogLevel::kTrace;
-constexpr LogLevel DEBUG = LogLevel::kDebug;
-constexpr LogLevel INFO = LogLevel::kInfo;
-constexpr LogLevel WARNING = LogLevel::kWarning;
-constexpr LogLevel ERROR = LogLevel::kError;
-constexpr LogLevel FATAL = LogLevel::kFatal;
-#else
-#define TRACE LogLevel::kTrace
-#define DEBUG LogLevel::kDebug
-#define INFO LogLevel::kInfo
-#define WARNING LogLevel::kWarning
-#define ERROR LogLevel::kError
-#define FATAL LogLevel::kFatal
-#endif
-
-std::string GetStackTrace();
-
-/* Return the current log level.
-
-
-   If the current log level is TRACE, then all logged messages are printed out.
-
-   If the current log level is DEBUG, log messages with "TRACE" level are not
-   shown and all other levels are printed out.
-
-   Similarly, if the current log level is INFO, log message with "TRACE" and
-   "DEBUG" are not shown and all other levels are printed out.
-
-   If it is FATAL, then only FATAL messages are shown.
- */
-inline LogLevel GetCurrentLogLevel() {
-  static LogLevel log_level = INFO;
-  static std::once_flag init_flag;
-  std::call_once(init_flag, []() {
-    const char *env_log_level = std::getenv("SHERPA_ONNX_LOG_LEVEL");
-    if (env_log_level == nullptr) return;
-
-    std::string s = env_log_level;
-    if (s == "TRACE")
-      log_level = TRACE;
-    else if (s == "DEBUG")
-      log_level = DEBUG;
-    else if (s == "INFO")
-      log_level = INFO;
-    else if (s == "WARNING")
-      log_level = WARNING;
-    else if (s == "ERROR")
-      log_level = ERROR;
-    else if (s == "FATAL")
-      log_level = FATAL;
-    else
-      fprintf(stderr,
-              "Unknown SHERPA_ONNX_LOG_LEVEL: %s"
-              "\nSupported values are: "
-              "TRACE, DEBUG, INFO, WARNING, ERROR, FATAL",
-              s.c_str());
-  });
-  return log_level;
-}
-
-inline bool EnableAbort() {
-  static std::once_flag init_flag;
-  static bool enable_abort = false;
-  std::call_once(init_flag, []() {
-    enable_abort = (std::getenv("SHERPA_ONNX_ABORT") != nullptr);
-  });
-  return enable_abort;
-}
-
-class Logger {
- public:
-  Logger(const char *filename, const char *func_name, uint32_t line_num,
-         LogLevel level)
-      : filename_(filename),
-        func_name_(func_name),
-        line_num_(line_num),
-        level_(level) {
-    cur_level_ = GetCurrentLogLevel();
-    switch (level) {
-      case TRACE:
-        if (cur_level_ <= TRACE) fprintf(stderr, "[T] ");
-        break;
-      case DEBUG:
-        if (cur_level_ <= DEBUG) fprintf(stderr, "[D] ");
-        break;
-      case INFO:
-        if (cur_level_ <= INFO) fprintf(stderr, "[I] ");
-        break;
-      case WARNING:
-        if (cur_level_ <= WARNING) fprintf(stderr, "[W] ");
-        break;
-      case ERROR:
-        if (cur_level_ <= ERROR) fprintf(stderr, "[E] ");
-        break;
-      case FATAL:
-        if (cur_level_ <= FATAL) fprintf(stderr, "[F] ");
-        break;
-    }
-
-    if (cur_level_ <= level_) {
-      fprintf(stderr, "%s:%u:%s ", filename, line_num, func_name);
-    }
-  }
-
-  ~Logger() noexcept(false) {
-    static constexpr const char *kErrMsg = R"(
-    Some bad things happened. Please read the above error messages and stack
-    trace. If you are using Python, the following command may be helpful:
-
-      gdb --args python /path/to/your/code.py
-
-    (You can use `gdb` to debug the code. Please consider compiling
-    a debug version of sherpa_onnx.).
-
-    If you are unable to fix it, please open an issue at:
-
-      https://github.com/csukuangfj/kaldi-native-fbank/issues/new
-    )";
-    if (level_ == FATAL) {
-      fprintf(stderr, "\n");
-      std::string stack_trace = GetStackTrace();
-      if (!stack_trace.empty()) {
-        fprintf(stderr, "\n\n%s\n", stack_trace.c_str());
-      }
-
-      fflush(nullptr);
-
-#ifndef __ANDROID_API__
-      if (EnableAbort()) {
-        // NOTE: abort() will terminate the program immediately without
-        // printing the Python stack backtrace.
-        abort();
-      }
-
-      throw std::runtime_error(kErrMsg);
-#else
-      abort();
-#endif
-    }
-  }
-
-  const Logger &operator<<(bool b) const {
-    if (cur_level_ <= level_) {
-      fprintf(stderr, b ? "true" : "false");
-    }
-    return *this;
-  }
-
-  const Logger &operator<<(int8_t i) const {
-    if (cur_level_ <= level_) fprintf(stderr, "%d", i);
-    return *this;
-  }
-
-  const Logger &operator<<(const char *s) const {
-    if (cur_level_ <= level_) fprintf(stderr, "%s", s);
-    return *this;
-  }
-
-  const Logger &operator<<(int32_t i) const {
-    if (cur_level_ <= level_) fprintf(stderr, "%d", i);
-    return *this;
-  }
-
-  const Logger &operator<<(uint32_t i) const {
-    if (cur_level_ <= level_) fprintf(stderr, "%u", i);
-    return *this;
-  }
-
-  const Logger &operator<<(uint64_t i) const {
-    if (cur_level_ <= level_)
-      fprintf(stderr, "%llu", (long long unsigned int)i);  // NOLINT
-    return *this;
-  }
-
-  const Logger &operator<<(int64_t i) const {
-    if (cur_level_ <= level_)
-      fprintf(stderr, "%lli", (long long int)i);  // NOLINT
-    return *this;
-  }
-
-  const Logger &operator<<(float f) const {
-    if (cur_level_ <= level_) fprintf(stderr, "%f", f);
-    return *this;
-  }
-
-  const Logger &operator<<(double d) const {
-    if (cur_level_ <= level_) fprintf(stderr, "%f", d);
-    return *this;
-  }
-
-  template <typename T>
-  const Logger &operator<<(const T &t) const {
-    // require T overloads operator<<
-    std::ostringstream os;
-    os << t;
-    return *this << os.str().c_str();
-  }
-
-  // specialization to fix compile error: `stringstream << nullptr` is ambiguous
-  const Logger &operator<<(const std::nullptr_t &null) const {
-    if (cur_level_ <= level_) *this << "(null)";
-    return *this;
-  }
-
- private:
-  const char *filename_;
-  const char *func_name_;
-  uint32_t line_num_;
-  LogLevel level_;
-  LogLevel cur_level_;
-};
-#endif  // SHERPA_ONNX_ENABLE_CHECK
-
-class Voidifier {
- public:
-#if SHERPA_ONNX_ENABLE_CHECK
-  void operator&(const Logger &) const {}
-#endif
-};
-#if !defined(SHERPA_ONNX_ENABLE_CHECK)
-template <typename T>
-const Voidifier &operator<<(const Voidifier &v, T &&) {
-  return v;
-}
-#endif
-
-}  // namespace sherpa_onnx
-
-#define SHERPA_ONNX_STATIC_ASSERT(x) static_assert(x, "")
-
-#ifdef SHERPA_ONNX_ENABLE_CHECK
-
-#if defined(__clang__) || defined(__GNUC__) || defined(__GNUG__) || \
-    defined(__PRETTY_FUNCTION__)
-// for clang and GCC
-#define SHERPA_ONNX_FUNC __PRETTY_FUNCTION__
-#else
-// for other compilers
-#define SHERPA_ONNX_FUNC __func__
-#endif
-
-#define SHERPA_ONNX_CHECK(x)                                            \
-  (x) ? (void)0                                                         \
-      : ::sherpa_onnx::Voidifier() &                                    \
-            ::sherpa_onnx::Logger(__FILE__, SHERPA_ONNX_FUNC, __LINE__, \
-                                  ::sherpa_onnx::FATAL)                 \
-                << "Check failed: " << #x << " "
-
-// WARNING: x and y may be evaluated multiple times, but this happens only
-// when the check fails. Since the program aborts if it fails, we don't think
-// the extra evaluation of x and y matters.
-//
-// CAUTION: we recommend the following use case:
-//
-//      auto x = Foo();
-//      auto y = Bar();
-//      SHERPA_ONNX_CHECK_EQ(x, y) << "Some message";
-//
-//  And please avoid
-//
-//      SHERPA_ONNX_CHECK_EQ(Foo(), Bar());
-//
-//  if `Foo()` or `Bar()` causes some side effects, e.g., changing some
-//  local static variables or global variables.
-#define _SHERPA_ONNX_CHECK_OP(x, y, op)                                        \
-  ((x)op(y)) ? (void)0                                                         \
-             : ::sherpa_onnx::Voidifier() &                                    \
-                   ::sherpa_onnx::Logger(__FILE__, SHERPA_ONNX_FUNC, __LINE__, \
-                                         ::sherpa_onnx::FATAL)                 \
-                       << "Check failed: " << #x << " " << #op << " " << #y    \
-                       << " (" << (x) << " vs. " << (y) << ") "
-
-#define SHERPA_ONNX_CHECK_EQ(x, y) _SHERPA_ONNX_CHECK_OP(x, y, ==)
-#define SHERPA_ONNX_CHECK_NE(x, y) _SHERPA_ONNX_CHECK_OP(x, y, !=)
-#define SHERPA_ONNX_CHECK_LT(x, y) _SHERPA_ONNX_CHECK_OP(x, y, <)
-#define SHERPA_ONNX_CHECK_LE(x, y) _SHERPA_ONNX_CHECK_OP(x, y, <=)
-#define SHERPA_ONNX_CHECK_GT(x, y) _SHERPA_ONNX_CHECK_OP(x, y, >)
-#define SHERPA_ONNX_CHECK_GE(x, y) _SHERPA_ONNX_CHECK_OP(x, y, >=)
-
-#define SHERPA_ONNX_LOG(x) \
-  ::sherpa_onnx::Logger(__FILE__, SHERPA_ONNX_FUNC, __LINE__, ::sherpa_onnx::x)
-
-// ------------------------------------------------------------
-//       For debug check
-// ------------------------------------------------------------
-// If you define the macro "-D NDEBUG" while compiling kaldi-native-fbank,
-// the following macros are in fact empty and does nothing.
-
-#define SHERPA_ONNX_DCHECK(x) \
-  ::sherpa_onnx::kDisableDebug ? (void)0 : SHERPA_ONNX_CHECK(x)
-
-#define SHERPA_ONNX_DCHECK_EQ(x, y) \
-  ::sherpa_onnx::kDisableDebug ? (void)0 : SHERPA_ONNX_CHECK_EQ(x, y)
-
-#define SHERPA_ONNX_DCHECK_NE(x, y) \
-  ::sherpa_onnx::kDisableDebug ? (void)0 : SHERPA_ONNX_CHECK_NE(x, y)
-
-#define SHERPA_ONNX_DCHECK_LT(x, y) \
-  ::sherpa_onnx::kDisableDebug ? (void)0 : SHERPA_ONNX_CHECK_LT(x, y)
-
-#define SHERPA_ONNX_DCHECK_LE(x, y) \
-  ::sherpa_onnx::kDisableDebug ? (void)0 : SHERPA_ONNX_CHECK_LE(x, y)
-
-#define SHERPA_ONNX_DCHECK_GT(x, y) \
-  ::sherpa_onnx::kDisableDebug ? (void)0 : SHERPA_ONNX_CHECK_GT(x, y)
-
-#define SHERPA_ONNX_DCHECK_GE(x, y) \
-  ::sherpa_onnx::kDisableDebug ? (void)0 : SHERPA_ONNX_CHECK_GE(x, y)
-
-#define SHERPA_ONNX_DLOG(x)    \
-  ::sherpa_onnx::kDisableDebug \
-      ? (void)0                \
-      : ::sherpa_onnx::Voidifier() & SHERPA_ONNX_LOG(x)
-
-#else
-
-#define SHERPA_ONNX_CHECK(x) ::sherpa_onnx::Voidifier()
-#define SHERPA_ONNX_LOG(x) ::sherpa_onnx::Voidifier()
-
-#define SHERPA_ONNX_CHECK_EQ(x, y) ::sherpa_onnx::Voidifier()
-#define SHERPA_ONNX_CHECK_NE(x, y) ::sherpa_onnx::Voidifier()
-#define SHERPA_ONNX_CHECK_LT(x, y) ::sherpa_onnx::Voidifier()
-#define SHERPA_ONNX_CHECK_LE(x, y) ::sherpa_onnx::Voidifier()
-#define SHERPA_ONNX_CHECK_GT(x, y) ::sherpa_onnx::Voidifier()
-#define SHERPA_ONNX_CHECK_GE(x, y) ::sherpa_onnx::Voidifier()
-
-#define SHERPA_ONNX_DCHECK(x) ::sherpa_onnx::Voidifier()
-#define SHERPA_ONNX_DLOG(x) ::sherpa_onnx::Voidifier()
-#define SHERPA_ONNX_DCHECK_EQ(x, y) ::sherpa_onnx::Voidifier()
-#define SHERPA_ONNX_DCHECK_NE(x, y) ::sherpa_onnx::Voidifier()
-#define SHERPA_ONNX_DCHECK_LT(x, y) ::sherpa_onnx::Voidifier()
-#define SHERPA_ONNX_DCHECK_LE(x, y) ::sherpa_onnx::Voidifier()
-#define SHERPA_ONNX_DCHECK_GT(x, y) ::sherpa_onnx::Voidifier()
-#define SHERPA_ONNX_DCHECK_GE(x, y) ::sherpa_onnx::Voidifier()
-
-#endif  // SHERPA_ONNX_CHECK_NE
-
-#endif  // SHERPA_ONNX_CSRC_LOG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/macros.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/macros.h
deleted file mode 100644
index ac11d6a7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/macros.h
+++ /dev/null
@@ -1,188 +0,0 @@
-// sherpa-onnx/csrc/macros.h
-//
-// Copyright      2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_MACROS_H_
-#define SHERPA_ONNX_CSRC_MACROS_H_
-#include <stdio.h>
-#include <stdlib.h>
-
-#include <utility>
-#if __OHOS__
-#include "hilog/log.h"
-
-#undef LOG_DOMAIN
-#undef LOG_TAG
-
-// https://gitee.com/openharmony/docs/blob/145a084f0b742e4325915e32f8184817927d1251/en/contribute/OpenHarmony-Log-guide.md#hilog-api-usage-specifications
-#define LOG_DOMAIN 0x6666
-#define LOG_TAG "sherpa_onnx"
-#endif
-
-#if __ANDROID_API__ >= 8
-#include "android/log.h"
-#define SHERPA_ONNX_LOGE(...)                                            \
-  do {                                                                   \
-    fprintf(stderr, "%s:%s:%d ", __FILE__, __func__,                     \
-            static_cast<int>(__LINE__));                                 \
-    fprintf(stderr, ##__VA_ARGS__);                                      \
-    fprintf(stderr, "\n");                                               \
-    __android_log_print(ANDROID_LOG_WARN, "sherpa-onnx", ##__VA_ARGS__); \
-  } while (0)
-#elif defined(__OHOS__)
-#define SHERPA_ONNX_LOGE(...) OH_LOG_INFO(LOG_APP, ##__VA_ARGS__)
-#elif SHERPA_ONNX_ENABLE_WASM
-#define SHERPA_ONNX_LOGE(...)                        \
-  do {                                               \
-    fprintf(stdout, "%s:%s:%d ", __FILE__, __func__, \
-            static_cast<int>(__LINE__));             \
-    fprintf(stdout, ##__VA_ARGS__);                  \
-    fprintf(stdout, "\n");                           \
-  } while (0)
-#else
-#define SHERPA_ONNX_LOGE(...)                        \
-  do {                                               \
-    fprintf(stderr, "%s:%s:%d ", __FILE__, __func__, \
-            static_cast<int>(__LINE__));             \
-    fprintf(stderr, ##__VA_ARGS__);                  \
-    fprintf(stderr, "\n");                           \
-  } while (0)
-#endif
-
-// Read an integer
-#define SHERPA_ONNX_READ_META_DATA(dst, src_key)                           \
-  do {                                                                     \
-    auto value = LookupCustomModelMetaData(meta_data, src_key, allocator); \
-    if (value.empty()) {                                                   \
-      SHERPA_ONNX_LOGE("'%s' does not exist in the metadata", src_key);    \
-      exit(-1);                                                            \
-    }                                                                      \
-                                                                           \
-    dst = atoi(value.c_str());                                             \
-    if (dst < 0) {                                                         \
-      SHERPA_ONNX_LOGE("Invalid value %d for '%s'", dst, src_key);         \
-      exit(-1);                                                            \
-    }                                                                      \
-  } while (0)
-
-#define SHERPA_ONNX_READ_META_DATA_WITH_DEFAULT(dst, src_key, default_value) \
-  do {                                                                       \
-    auto value = LookupCustomModelMetaData(meta_data, src_key, allocator);   \
-    if (value.empty()) {                                                     \
-      dst = default_value;                                                   \
-    } else {                                                                 \
-      dst = atoi(value.c_str());                                             \
-      if (dst < 0) {                                                         \
-        SHERPA_ONNX_LOGE("Invalid value %d for '%s'", dst, src_key);         \
-        exit(-1);                                                            \
-      }                                                                      \
-    }                                                                        \
-  } while (0)
-
-// read a vector of integers
-#define SHERPA_ONNX_READ_META_DATA_VEC(dst, src_key)                           \
-  do {                                                                         \
-    auto value = LookupCustomModelMetaData(meta_data, src_key, allocator);     \
-    if (value.empty()) {                                                       \
-      SHERPA_ONNX_LOGE("'%s' does not exist in the metadata", src_key);        \
-      exit(-1);                                                                \
-    }                                                                          \
-                                                                               \
-    bool ret = SplitStringToIntegers(value.c_str(), ",", true, &dst);          \
-    if (!ret) {                                                                \
-      SHERPA_ONNX_LOGE("Invalid value '%s' for '%s'", value.c_str(), src_key); \
-      exit(-1);                                                                \
-    }                                                                          \
-  } while (0)
-
-// read a vector of floats
-#define SHERPA_ONNX_READ_META_DATA_VEC_FLOAT(dst, src_key)                     \
-  do {                                                                         \
-    auto value = LookupCustomModelMetaData(meta_data, src_key, allocator);     \
-    if (value.empty()) {                                                       \
-      SHERPA_ONNX_LOGE("%s does not exist in the metadata", src_key);          \
-      exit(-1);                                                                \
-    }                                                                          \
-                                                                               \
-    bool ret = SplitStringToFloats(value.c_str(), ",", true, &dst);            \
-    if (!ret) {                                                                \
-      SHERPA_ONNX_LOGE("Invalid value '%s' for '%s'", value.c_str(), src_key); \
-      exit(-1);                                                                \
-    }                                                                          \
-  } while (0)
-
-// read a vector of strings
-#define SHERPA_ONNX_READ_META_DATA_VEC_STRING(dst, src_key)                \
-  do {                                                                     \
-    auto value = LookupCustomModelMetaData(meta_data, src_key, allocator); \
-    if (value.empty()) {                                                   \
-      SHERPA_ONNX_LOGE("'%s' does not exist in the metadata", src_key);    \
-      exit(-1);                                                            \
-    }                                                                      \
-    SplitStringToVector(value.c_str(), ",", false, &dst);                  \
-                                                                           \
-    if (dst.empty()) {                                                     \
-      SHERPA_ONNX_LOGE("Invalid value '%s' for '%s'. Empty vector!",       \
-                       value.c_str(), src_key);                            \
-      exit(-1);                                                            \
-    }                                                                      \
-  } while (0)
-
-// read a vector of strings separated by sep
-#define SHERPA_ONNX_READ_META_DATA_VEC_STRING_SEP(dst, src_key, sep)       \
-  do {                                                                     \
-    auto value = LookupCustomModelMetaData(meta_data, src_key, allocator); \
-    if (value.empty()) {                                                   \
-      SHERPA_ONNX_LOGE("'%s' does not exist in the metadata", src_key);    \
-      exit(-1);                                                            \
-    }                                                                      \
-    SplitStringToVector(value.c_str(), sep, false, &dst);                  \
-                                                                           \
-    if (dst.empty()) {                                                     \
-      SHERPA_ONNX_LOGE("Invalid value '%s' for '%s'. Empty vector!",       \
-                       value.c_str(), src_key);                            \
-      exit(-1);                                                            \
-    }                                                                      \
-  } while (0)
-
-// Read a string
-#define SHERPA_ONNX_READ_META_DATA_STR(dst, src_key)                       \
-  do {                                                                     \
-    auto value = LookupCustomModelMetaData(meta_data, src_key, allocator); \
-    if (value.empty()) {                                                   \
-      SHERPA_ONNX_LOGE("'%s' does not exist in the metadata", src_key);    \
-      exit(-1);                                                            \
-    }                                                                      \
-                                                                           \
-    dst = std::move(value);                                                \
-    if (dst.empty()) {                                                     \
-      SHERPA_ONNX_LOGE("Invalid value for '%s'\n", src_key);               \
-      exit(-1);                                                            \
-    }                                                                      \
-  } while (0)
-
-#define SHERPA_ONNX_READ_META_DATA_STR_ALLOW_EMPTY(dst, src_key)           \
-  do {                                                                     \
-    auto value = LookupCustomModelMetaData(meta_data, src_key, allocator); \
-                                                                           \
-    dst = std::move(value);                                                \
-  } while (0)
-
-#define SHERPA_ONNX_READ_META_DATA_STR_WITH_DEFAULT(dst, src_key,          \
-                                                    default_value)         \
-  do {                                                                     \
-    auto value = LookupCustomModelMetaData(meta_data, src_key, allocator); \
-    if (value.empty()) {                                                   \
-      dst = default_value;                                                 \
-    } else {                                                               \
-      dst = std::move(value);                                              \
-      if (dst.empty()) {                                                   \
-        SHERPA_ONNX_LOGE("Invalid value for '%s'\n", src_key);             \
-        exit(-1);                                                          \
-      }                                                                    \
-    }                                                                      \
-  } while (0)
-
-#define SHERPA_ONNX_EXIT(code) exit(code)
-
-#endif  // SHERPA_ONNX_CSRC_MACROS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/math.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/math.h
deleted file mode 100644
index 121a05ae..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/math.h
+++ /dev/null
@@ -1,123 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Daniel Povey)
- * Copyright (c)  2023                     (Pingfeng Luo)
- *
- */
-// This file is copied from k2/csrc/utils.h
-#ifndef SHERPA_ONNX_CSRC_MATH_H_
-#define SHERPA_ONNX_CSRC_MATH_H_
-
-#include <algorithm>
-#include <cassert>
-#include <cmath>
-#include <numeric>
-#include <vector>
-
-namespace sherpa_onnx {
-
-// logf(FLT_EPSILON)
-#define SHERPA_ONNX_MIN_LOG_DIFF_FLOAT -15.9423847198486328125f
-
-// log(DBL_EPSILON)
-#define SHERPA_ONNX_MIN_LOG_DIFF_DOUBLE \
-  -36.0436533891171535515240975655615329742431640625
-
-template <typename T>
-struct LogAdd;
-
-template <>
-struct LogAdd<double> {
-  double operator()(double x, double y) const {
-    double diff;
-
-    if (x < y) {
-      diff = x - y;
-      x = y;
-    } else {
-      diff = y - x;
-    }
-    // diff is negative.  x is now the larger one.
-
-    if (diff >= SHERPA_ONNX_MIN_LOG_DIFF_DOUBLE) {
-      double res;
-      res = x + log1p(exp(diff));
-      return res;
-    }
-
-    return x;  // return the larger one.
-  }
-};
-
-template <>
-struct LogAdd<float> {
-  float operator()(float x, float y) const {
-    float diff;
-
-    if (x < y) {
-      diff = x - y;
-      x = y;
-    } else {
-      diff = y - x;
-    }
-    // diff is negative.  x is now the larger one.
-
-    if (diff >= SHERPA_ONNX_MIN_LOG_DIFF_DOUBLE) {
-      float res;
-      res = x + log1pf(expf(diff));
-      return res;
-    }
-
-    return x;  // return the larger one.
-  }
-};
-
-template <class T>
-void LogSoftmax(T *input, int32_t input_len) {
-  assert(input);
-
-  T m = *std::max_element(input, input + input_len);
-
-  T sum = 0.0;
-  for (int32_t i = 0; i < input_len; i++) {
-    sum += exp(input[i] - m);
-  }
-
-  T offset = m + log(sum);
-  for (int32_t i = 0; i < input_len; i++) {
-    input[i] -= offset;
-  }
-}
-
-template <typename T>
-void LogSoftmax(T *in, int32_t w, int32_t h) {
-  for (int32_t i = 0; i != h; ++i) {
-    LogSoftmax(in, w);
-    in += w;
-  }
-}
-
-template <typename T>
-void SubtractBlank(T *in, int32_t w, int32_t h, int32_t blank_idx,
-                   float blank_penalty) {
-  for (int32_t i = 0; i != h; ++i) {
-    in[blank_idx] -= blank_penalty;
-    in += w;
-  }
-}
-
-template <class T>
-std::vector<int32_t> TopkIndex(const T *vec, int32_t size, int32_t topk) {
-  std::vector<int32_t> vec_index(size);
-  std::iota(vec_index.begin(), vec_index.end(), 0);
-
-  std::partial_sort(vec_index.begin(), vec_index.begin() + topk,
-                    vec_index.end(), [vec](int32_t index_1, int32_t index_2) {
-                      return vec[index_1] > vec[index_2];
-                    });
-
-  int32_t k_num = std::min<int32_t>(size, topk);
-  return {vec_index.begin(), vec_index.begin() + k_num};
-}
-
-}  // namespace sherpa_onnx
-#endif  // SHERPA_ONNX_CSRC_MATH_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/melo-tts-lexicon.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/melo-tts-lexicon.h
deleted file mode 100644
index e91cf33f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/melo-tts-lexicon.h
+++ /dev/null
@@ -1,49 +0,0 @@
-// sherpa-onnx/csrc/melo-tts-lexicon.h
-//
-// Copyright (c)  2022-2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_MELO_TTS_LEXICON_H_
-#define SHERPA_ONNX_CSRC_MELO_TTS_LEXICON_H_
-
-#include <memory>
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-tts-frontend.h"
-#include "sherpa-onnx/csrc/offline-tts-vits-model-metadata.h"
-
-namespace sherpa_onnx {
-
-class MeloTtsLexicon : public OfflineTtsFrontend {
- public:
-  ~MeloTtsLexicon() override;
-  MeloTtsLexicon(const std::string &lexicon, const std::string &tokens,
-                 const std::string &dict_dir,
-                 const OfflineTtsVitsModelMetaData &meta_data, bool debug);
-
-  MeloTtsLexicon(const std::string &lexicon, const std::string &tokens,
-                 const OfflineTtsVitsModelMetaData &meta_data, bool debug);
-
-  template <typename Manager>
-  MeloTtsLexicon(Manager *mgr, const std::string &lexicon,
-                 const std::string &tokens, const std::string &dict_dir,
-                 const OfflineTtsVitsModelMetaData &meta_data, bool debug);
-
-  template <typename Manager>
-  MeloTtsLexicon(Manager *mgr, const std::string &lexicon,
-                 const std::string &tokens,
-                 const OfflineTtsVitsModelMetaData &meta_data, bool debug);
-
-  std::vector<TokenIDs> ConvertTextToTokenIds(
-      const std::string &text,
-      const std::string &unused_voice = "") const override;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_MELO_TTS_LEXICON_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/microphone.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/microphone.h
deleted file mode 100644
index 61cedbf3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/microphone.h
+++ /dev/null
@@ -1,18 +0,0 @@
-// sherpa-onnx/csrc/microphone.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_MICROPHONE_H_
-#define SHERPA_ONNX_CSRC_MICROPHONE_H_
-
-namespace sherpa_onnx {
-
-class Microphone {
- public:
-  Microphone();
-  ~Microphone();
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_MICROPHONE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ced-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ced-model.h
deleted file mode 100644
index bbaac579..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ced-model.h
+++ /dev/null
@@ -1,56 +0,0 @@
-// sherpa-onnx/csrc/offline-ced-model.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_CED_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_CED_MODEL_H_
-#include <memory>
-#include <utility>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/audio-tagging-model-config.h"
-
-namespace sherpa_onnx {
-
-/** This class implements the CED model from
- * https://github.com/RicherMans/CED/blob/main/export_onnx.py
- */
-class OfflineCEDModel {
- public:
-  explicit OfflineCEDModel(const AudioTaggingModelConfig &config);
-
-#if __ANDROID_API__ >= 9
-  OfflineCEDModel(AAssetManager *mgr, const AudioTaggingModelConfig &config);
-#endif
-
-  ~OfflineCEDModel();
-
-  /** Run the forward method of the model.
-   *
-   * @param features  A tensor of shape (N, T, C).
-   *
-   * @return Return a tensor
-   *  - probs: A 2-D tensor of shape (N, num_event_classes).
-   */
-  Ort::Value Forward(Ort::Value features) const;
-
-  /** Return the number of event classes of the model
-   */
-  int32_t NumEventClasses() const;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_CED_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ct-transformer-model-meta-data.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ct-transformer-model-meta-data.h
deleted file mode 100644
index d94e24a3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ct-transformer-model-meta-data.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// sherpa-onnx/csrc/offline-ct-transformer-model-meta-data.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_CT_TRANSFORMER_MODEL_META_DATA_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_CT_TRANSFORMER_MODEL_META_DATA_H_
-
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-namespace sherpa_onnx {
-
-struct OfflineCtTransformerModelMetaData {
-  std::unordered_map<std::string, int32_t> token2id;
-  std::unordered_map<std::string, int32_t> punct2id;
-  std::vector<std::string> id2punct;
-
-  int32_t unk_id;
-  int32_t dot_id;
-  int32_t comma_id;
-  int32_t quest_id;
-  int32_t pause_id;
-  int32_t underline_id;
-  int32_t num_punctuations;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_CT_TRANSFORMER_MODEL_META_DATA_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ct-transformer-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ct-transformer-model.h
deleted file mode 100644
index 06e14ec7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ct-transformer-model.h
+++ /dev/null
@@ -1,59 +0,0 @@
-// sherpa-onnx/csrc/offline-ct-transformer-model.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_CT_TRANSFORMER_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_CT_TRANSFORMER_MODEL_H_
-#include <memory>
-#include <utility>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-ct-transformer-model-meta-data.h"
-#include "sherpa-onnx/csrc/offline-punctuation-model-config.h"
-
-namespace sherpa_onnx {
-
-/** This class implements
- * https://github.com/alibaba-damo-academy/FunASR/blob/main/runtime/python/onnxruntime/funasr_onnx/punc_bin.py#L17
- * from FunASR
- */
-class OfflineCtTransformerModel {
- public:
-  explicit OfflineCtTransformerModel(
-      const OfflinePunctuationModelConfig &config);
-
-#if __ANDROID_API__ >= 9
-  OfflineCtTransformerModel(AAssetManager *mgr,
-                            const OfflinePunctuationModelConfig &config);
-#endif
-
-  ~OfflineCtTransformerModel();
-
-  /** Run the forward method of the model.
-   *
-   * @param text  A tensor of shape (N, T) of dtype int32.
-   * @param text  A tensor of shape (N) of dtype int32.
-   *
-   * @return Return a tensor
-   *  - punctuation_ids: A 2-D tensor of shape (N, T).
-   */
-  Ort::Value Forward(Ort::Value text, Ort::Value text_len) const;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const;
-
-  const OfflineCtTransformerModelMetaData &GetModelMetadata() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_CT_TRANSFORMER_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-decoder.h
deleted file mode 100644
index c9d1b36f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-decoder.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// sherpa-onnx/csrc/offline-ctc-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_CTC_DECODER_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_CTC_DECODER_H_
-
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-
-namespace sherpa_onnx {
-
-struct OfflineCtcDecoderResult {
-  /// The decoded token IDs
-  std::vector<int64_t> tokens;
-
-  /// The decoded word IDs
-  /// Note: tokens.size() is usually not equal to words.size()
-  /// words is empty for greedy search decoding.
-  /// it is not empty when an HLG graph or an HLG graph is used.
-  std::vector<int32_t> words;
-
-  /// timestamps[i] contains the output frame index where tokens[i] is decoded.
-  /// Note: The index is after subsampling
-  ///
-  /// tokens.size() == timestamps.size()
-  std::vector<int32_t> timestamps;
-};
-
-class OfflineCtcDecoder {
- public:
-  virtual ~OfflineCtcDecoder() = default;
-
-  /** Run CTC decoding given the output from the encoder model.
-   *
-   * @param log_probs A 3-D tensor of shape (N, T, vocab_size) containing
-   *                  lob_probs.
-   * @param log_probs_length A 1-D tensor of shape (N,) containing number
-   *                         of valid frames in log_probs before padding.
-   *
-   * @return Return a vector of size `N` containing the decoded results.
-   */
-  virtual std::vector<OfflineCtcDecoderResult> Decode(
-      Ort::Value log_probs, Ort::Value log_probs_length) = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_CTC_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-fst-decoder-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-fst-decoder-config.h
deleted file mode 100644
index b87fe89e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-fst-decoder-config.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// sherpa-onnx/csrc/offline-ctc-fst-decoder-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_CTC_FST_DECODER_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_CTC_FST_DECODER_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineCtcFstDecoderConfig {
-  // Path to H.fst, HL.fst or HLG.fst
-  std::string graph;
-  int32_t max_active = 3000;
-
-  OfflineCtcFstDecoderConfig() = default;
-
-  OfflineCtcFstDecoderConfig(const std::string &graph, int32_t max_active)
-      : graph(graph), max_active(max_active) {}
-
-  std::string ToString() const;
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_CTC_FST_DECODER_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-fst-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-fst-decoder.h
deleted file mode 100644
index 2b33c14e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-fst-decoder.h
+++ /dev/null
@@ -1,33 +0,0 @@
-// sherpa-onnx/csrc/offline-ctc-fst-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_CTC_FST_DECODER_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_CTC_FST_DECODER_H_
-
-#include <memory>
-#include <vector>
-
-#include "fst/fst.h"
-#include "sherpa-onnx/csrc/offline-ctc-decoder.h"
-#include "sherpa-onnx/csrc/offline-ctc-fst-decoder-config.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-class OfflineCtcFstDecoder : public OfflineCtcDecoder {
- public:
-  explicit OfflineCtcFstDecoder(const OfflineCtcFstDecoderConfig &config);
-
-  std::vector<OfflineCtcDecoderResult> Decode(
-      Ort::Value log_probs, Ort::Value log_probs_length) override;
-
- private:
-  OfflineCtcFstDecoderConfig config_;
-
-  std::unique_ptr<fst::Fst<fst::StdArc>> fst_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_CTC_FST_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-greedy-search-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-greedy-search-decoder.h
deleted file mode 100644
index ccc2f728..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-greedy-search-decoder.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// sherpa-onnx/csrc/offline-ctc-greedy-search-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_CTC_GREEDY_SEARCH_DECODER_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_CTC_GREEDY_SEARCH_DECODER_H_
-
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-ctc-decoder.h"
-
-namespace sherpa_onnx {
-
-class OfflineCtcGreedySearchDecoder : public OfflineCtcDecoder {
- public:
-  explicit OfflineCtcGreedySearchDecoder(int32_t blank_id)
-      : blank_id_(blank_id) {}
-
-  std::vector<OfflineCtcDecoderResult> Decode(
-      Ort::Value log_probs, Ort::Value log_probs_length) override;
-
- private:
-  int32_t blank_id_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_CTC_GREEDY_SEARCH_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-model.h
deleted file mode 100644
index 5ad4fcdc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-ctc-model.h
+++ /dev/null
@@ -1,71 +0,0 @@
-// sherpa-onnx/csrc/offline-ctc-model.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_CTC_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_CTC_MODEL_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-model-config.h"
-
-namespace sherpa_onnx {
-
-class OfflineCtcModel {
- public:
-  virtual ~OfflineCtcModel() = default;
-
-  static std::unique_ptr<OfflineCtcModel> Create(
-      const OfflineModelConfig &config);
-
-  template <typename Manager>
-  static std::unique_ptr<OfflineCtcModel> Create(
-      Manager *mgr, const OfflineModelConfig &config);
-
-  /** Run the forward method of the model.
-   *
-   * @param features  A tensor of shape (N, T, C).
-   * @param features_length  A 1-D tensor of shape (N,) containing number of
-   *                         valid frames in `features` before padding.
-   *                         Its dtype is int64_t.
-   *
-   * @return Return a vector containing:
-   *  - log_probs: A 3-D tensor of shape (N, T', vocab_size).
-   *  - log_probs_length A 1-D tensor of shape (N,). Its dtype is int64_t
-   */
-  virtual std::vector<Ort::Value> Forward(Ort::Value features,
-                                          Ort::Value features_length) = 0;
-
-  /** Return the vocabulary size of the model
-   */
-  virtual int32_t VocabSize() const = 0;
-
-  /** SubsamplingFactor of the model
-   *
-   * For NeMo Citrinet, the subsampling factor is usually 4.
-   * For NeMo Conformer CTC, the subsampling factor is usually 8.
-   */
-  virtual int32_t SubsamplingFactor() const { return 1; }
-
-  /** Return an allocator for allocating memory
-   */
-  virtual OrtAllocator *Allocator() const = 0;
-
-  /** For some models, e.g., those from NeMo, they require some preprocessing
-   * for the features.
-   */
-  virtual std::string FeatureNormalizationMethod() const { return {}; }
-
-  // Return true if the model supports batch size > 1
-  virtual bool SupportBatchProcessing() const { return true; }
-
-  // return true for models from https://github.com/salute-developers/GigaAM
-  // return false otherwise
-  virtual bool IsGigaAM() const { return false; }
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_CTC_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-lm-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-lm-config.h
deleted file mode 100644
index 3468c58a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-lm-config.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// sherpa-onnx/csrc/offline-lm-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_LM_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_LM_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineLMConfig {
-  // path to the onnx model
-  std::string model;
-
-  // LM scale
-  float scale = 0.5;
-  int32_t lm_num_threads = 1;
-  std::string lm_provider = "cpu";
-
-  OfflineLMConfig() = default;
-
-  OfflineLMConfig(const std::string &model, float scale, int32_t lm_num_threads,
-                  const std::string &lm_provider)
-      : model(model),
-        scale(scale),
-        lm_num_threads(lm_num_threads),
-        lm_provider(lm_provider) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_LM_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-lm.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-lm.h
deleted file mode 100644
index a9af8202..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-lm.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// sherpa-onnx/csrc/offline-lm.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_LM_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_LM_H_
-
-#include <memory>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/hypothesis.h"
-#include "sherpa-onnx/csrc/offline-lm-config.h"
-
-namespace sherpa_onnx {
-
-class OfflineLM {
- public:
-  virtual ~OfflineLM() = default;
-
-  static std::unique_ptr<OfflineLM> Create(const OfflineLMConfig &config);
-
-  template <typename Manager>
-  static std::unique_ptr<OfflineLM> Create(Manager *mgr,
-                                           const OfflineLMConfig &config);
-
-  /** Rescore a batch of sentences.
-   *
-   * @param x A 2-D tensor of shape (N, L) with data type int64.
-   * @param x_lens A 1-D tensor of shape (N,) with data type int64.
-   *               It contains number of valid tokens in x before padding.
-   * @return Return a 1-D tensor of shape (N,) containing the negative log
-   *         likelihood of each utterance. Its data type is float32.
-   *
-   * Caution: It returns negative log likelihood (nll), not log likelihood
-   */
-  virtual Ort::Value Rescore(Ort::Value x, Ort::Value x_lens) = 0;
-
-  // This function updates hyp.lm_lob_prob of hyps.
-  //
-  // @param scale LM score
-  // @param context_size Context size of the transducer decoder model
-  // @param hyps It is changed in-place.
-  void ComputeLMScore(float scale, int32_t context_size,
-                      std::vector<Hypotheses> *hyps);
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_LM_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-model-config.h
deleted file mode 100644
index cfff5eed..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-model-config.h
+++ /dev/null
@@ -1,93 +0,0 @@
-// sherpa-onnx/csrc/offline-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/offline-moonshine-model-config.h"
-#include "sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model-config.h"
-#include "sherpa-onnx/csrc/offline-paraformer-model-config.h"
-#include "sherpa-onnx/csrc/offline-sense-voice-model-config.h"
-#include "sherpa-onnx/csrc/offline-tdnn-model-config.h"
-#include "sherpa-onnx/csrc/offline-transducer-model-config.h"
-#include "sherpa-onnx/csrc/offline-wenet-ctc-model-config.h"
-#include "sherpa-onnx/csrc/offline-whisper-model-config.h"
-#include "sherpa-onnx/csrc/offline-zipformer-ctc-model-config.h"
-
-namespace sherpa_onnx {
-
-struct OfflineModelConfig {
-  OfflineTransducerModelConfig transducer;
-  OfflineParaformerModelConfig paraformer;
-  OfflineNemoEncDecCtcModelConfig nemo_ctc;
-  OfflineWhisperModelConfig whisper;
-  OfflineTdnnModelConfig tdnn;
-  OfflineZipformerCtcModelConfig zipformer_ctc;
-  OfflineWenetCtcModelConfig wenet_ctc;
-  OfflineSenseVoiceModelConfig sense_voice;
-  OfflineMoonshineModelConfig moonshine;
-  std::string telespeech_ctc;
-
-  std::string tokens;
-  int32_t num_threads = 2;
-  bool debug = false;
-  std::string provider = "cpu";
-
-  // With the help of this field, we only need to load the model once
-  // instead of twice; and therefore it reduces initialization time.
-  //
-  // Valid values:
-  //  - transducer. The given model is from icefall
-  //  - paraformer. It is a paraformer model
-  //  - nemo_ctc. It is a NeMo CTC model.
-  //
-  // All other values are invalid and lead to loading the model twice.
-  std::string model_type;
-
-  std::string modeling_unit = "cjkchar";
-  std::string bpe_vocab;
-
-  OfflineModelConfig() = default;
-  OfflineModelConfig(const OfflineTransducerModelConfig &transducer,
-                     const OfflineParaformerModelConfig &paraformer,
-                     const OfflineNemoEncDecCtcModelConfig &nemo_ctc,
-                     const OfflineWhisperModelConfig &whisper,
-                     const OfflineTdnnModelConfig &tdnn,
-                     const OfflineZipformerCtcModelConfig &zipformer_ctc,
-                     const OfflineWenetCtcModelConfig &wenet_ctc,
-                     const OfflineSenseVoiceModelConfig &sense_voice,
-                     const OfflineMoonshineModelConfig &moonshine,
-                     const std::string &telespeech_ctc,
-                     const std::string &tokens, int32_t num_threads, bool debug,
-                     const std::string &provider, const std::string &model_type,
-                     const std::string &modeling_unit,
-                     const std::string &bpe_vocab)
-      : transducer(transducer),
-        paraformer(paraformer),
-        nemo_ctc(nemo_ctc),
-        whisper(whisper),
-        tdnn(tdnn),
-        zipformer_ctc(zipformer_ctc),
-        wenet_ctc(wenet_ctc),
-        sense_voice(sense_voice),
-        moonshine(moonshine),
-        telespeech_ctc(telespeech_ctc),
-        tokens(tokens),
-        num_threads(num_threads),
-        debug(debug),
-        provider(provider),
-        model_type(model_type),
-        modeling_unit(modeling_unit),
-        bpe_vocab(bpe_vocab) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-decoder.h
deleted file mode 100644
index 4d0b9ac9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-decoder.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// sherpa-onnx/csrc/offline-moonshine-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_MOONSHINE_DECODER_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_MOONSHINE_DECODER_H_
-
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-
-namespace sherpa_onnx {
-
-struct OfflineMoonshineDecoderResult {
-  /// The decoded token IDs
-  std::vector<int32_t> tokens;
-};
-
-class OfflineMoonshineDecoder {
- public:
-  virtual ~OfflineMoonshineDecoder() = default;
-
-  /** Run beam search given the output from the moonshine encoder model.
-   *
-   * @param encoder_out A 3-D tensor of shape (batch_size, T, dim)
-   * @return Return a vector of size `N` containing the decoded results.
-   */
-  virtual std::vector<OfflineMoonshineDecoderResult> Decode(
-      Ort::Value encoder_out) = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_MOONSHINE_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-greedy-search-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-greedy-search-decoder.h
deleted file mode 100644
index b215405d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-greedy-search-decoder.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// sherpa-onnx/csrc/offline-moonshine-greedy-search-decoder.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_MOONSHINE_GREEDY_SEARCH_DECODER_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_MOONSHINE_GREEDY_SEARCH_DECODER_H_
-
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-moonshine-decoder.h"
-#include "sherpa-onnx/csrc/offline-moonshine-model.h"
-
-namespace sherpa_onnx {
-
-class OfflineMoonshineGreedySearchDecoder : public OfflineMoonshineDecoder {
- public:
-  explicit OfflineMoonshineGreedySearchDecoder(OfflineMoonshineModel *model)
-      : model_(model) {}
-
-  std::vector<OfflineMoonshineDecoderResult> Decode(
-      Ort::Value encoder_out) override;
-
- private:
-  OfflineMoonshineModel *model_;  // not owned
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_MOONSHINE_GREEDY_SEARCH_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-model-config.h
deleted file mode 100644
index 829ca520..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-model-config.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// sherpa-onnx/csrc/offline-moonshine-model-config.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_MOONSHINE_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_MOONSHINE_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineMoonshineModelConfig {
-  std::string preprocessor;
-  std::string encoder;
-  std::string uncached_decoder;
-  std::string cached_decoder;
-
-  OfflineMoonshineModelConfig() = default;
-  OfflineMoonshineModelConfig(const std::string &preprocessor,
-                              const std::string &encoder,
-                              const std::string &uncached_decoder,
-                              const std::string &cached_decoder)
-      : preprocessor(preprocessor),
-        encoder(encoder),
-        uncached_decoder(uncached_decoder),
-        cached_decoder(cached_decoder) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_MOONSHINE_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-model.h
deleted file mode 100644
index a8f9b408..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-moonshine-model.h
+++ /dev/null
@@ -1,87 +0,0 @@
-// sherpa-onnx/csrc/offline-moonshine-model.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_MOONSHINE_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_MOONSHINE_MODEL_H_
-
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-model-config.h"
-
-namespace sherpa_onnx {
-
-// please see
-// https://github.com/k2-fsa/sherpa-onnx/blob/master/scripts/moonshine/test.py
-class OfflineMoonshineModel {
- public:
-  explicit OfflineMoonshineModel(const OfflineModelConfig &config);
-
-  template <typename Manager>
-  OfflineMoonshineModel(Manager *mgr, const OfflineModelConfig &config);
-
-  ~OfflineMoonshineModel();
-
-  /** Run the preprocessor model.
-   *
-   * @param audio A float32 tensor of shape (batch_size, num_samples)
-   *
-   * @return Return a float32 tensor of shape (batch_size, T, dim) that
-   *         can be used as the input of ForwardEncoder()
-   */
-  Ort::Value ForwardPreprocessor(Ort::Value audio) const;
-
-  /** Run the encoder model.
-   *
-   * @param features A float32 tensor of shape (batch_size, T, dim)
-   * @param features_len A int32 tensor of shape (batch_size,)
-   * @returns A float32 tensor of shape (batch_size, T, dim).
-   */
-  Ort::Value ForwardEncoder(Ort::Value features, Ort::Value features_len) const;
-
-  /** Run the uncached decoder.
-   *
-   * @param token A int32 tensor of shape (batch_size, num_tokens)
-   * @param seq_len A int32 tensor of shape (batch_size,) containing number
-   *                of predicted tokens so far
-   * @param encoder_out A float32 tensor of shape (batch_size, T, dim)
-   *
-   * @returns Return a pair:
-   *
-   *          - logits, a float32 tensor of shape (batch_size, 1, dim)
-   *          - states, a list of states
-   */
-  std::pair<Ort::Value, std::vector<Ort::Value>> ForwardUnCachedDecoder(
-      Ort::Value token, Ort::Value seq_len, Ort::Value encoder_out) const;
-
-  /** Run the cached decoder.
-   *
-   * @param token A int32 tensor of shape (batch_size, num_tokens)
-   * @param seq_len A int32 tensor of shape (batch_size,) containing number
-   *                of predicted tokens so far
-   * @param encoder_out A float32 tensor of shape (batch_size, T, dim)
-   * @param states A list of previous states
-   *
-   * @returns Return a pair:
-   *          - logits, a float32 tensor of shape (batch_size, 1, dim)
-   *          - states, a list of new states
-   */
-  std::pair<Ort::Value, std::vector<Ort::Value>> ForwardCachedDecoder(
-      Ort::Value token, Ort::Value seq_len, Ort::Value encoder_out,
-      std::vector<Ort::Value> states) const;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_MOONSHINE_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model-config.h
deleted file mode 100644
index 9ef7b54a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model-config.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_NEMO_ENC_DEC_CTC_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_NEMO_ENC_DEC_CTC_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineNemoEncDecCtcModelConfig {
-  std::string model;
-
-  OfflineNemoEncDecCtcModelConfig() = default;
-  explicit OfflineNemoEncDecCtcModelConfig(const std::string &model)
-      : model(model) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_NEMO_ENC_DEC_CTC_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model.h
deleted file mode 100644
index 08ff068c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model.h
+++ /dev/null
@@ -1,83 +0,0 @@
-// sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_NEMO_ENC_DEC_CTC_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_NEMO_ENC_DEC_CTC_MODEL_H_
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-ctc-model.h"
-#include "sherpa-onnx/csrc/offline-model-config.h"
-
-namespace sherpa_onnx {
-
-/** This class implements the EncDecCTCModelBPE model from NeMo.
- *
- * See
- * https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/models/ctc_bpe_models.py
- * https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/models/ctc_models.py
- */
-class OfflineNemoEncDecCtcModel : public OfflineCtcModel {
- public:
-  explicit OfflineNemoEncDecCtcModel(const OfflineModelConfig &config);
-
-  template <typename Manager>
-  OfflineNemoEncDecCtcModel(Manager *mgr, const OfflineModelConfig &config);
-
-  ~OfflineNemoEncDecCtcModel() override;
-
-  /** Run the forward method of the model.
-   *
-   * @param features  A tensor of shape (N, T, C).
-   * @param features_length  A 1-D tensor of shape (N,) containing number of
-   *                         valid frames in `features` before padding.
-   *                         Its dtype is int64_t.
-   *
-   * @return Return a vector containing:
-   *  - log_probs: A 3-D tensor of shape (N, T', vocab_size).
-   *  - log_probs_length A 1-D tensor of shape (N,). Its dtype is int64_t
-   */
-  std::vector<Ort::Value> Forward(Ort::Value features,
-                                  Ort::Value features_length) override;
-
-  /** Return the vocabulary size of the model
-   */
-  int32_t VocabSize() const override;
-
-  /** SubsamplingFactor of the model
-   *
-   * For Citrinet, the subsampling factor is usually 4.
-   * For Conformer CTC, the subsampling factor is usually 8.
-   */
-  int32_t SubsamplingFactor() const override;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const override;
-
-  // Possible values:
-  // - per_feature
-  // - all_features (not implemented yet)
-  // - fixed_mean (not implemented)
-  // - fixed_std (not implemented)
-  // - or just leave it to empty
-  // See
-  // https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/parts/preprocessing/features.py#L59
-  // for details
-  std::string FeatureNormalizationMethod() const override;
-
-  bool IsGigaAM() const override;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-using OfflineNemoEncDecHybridRNNTCTCBPEModel = OfflineNemoEncDecCtcModel;
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_NEMO_ENC_DEC_CTC_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-decoder.h
deleted file mode 100644
index 2effdfe1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-decoder.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// sherpa-onnx/csrc/offline-paraformer-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_PARAFORMER_DECODER_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_PARAFORMER_DECODER_H_
-
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-
-namespace sherpa_onnx {
-
-struct OfflineParaformerDecoderResult {
-  /// The decoded token IDs
-  std::vector<int64_t> tokens;
-
-  // it contains the start time of each token in seconds
-  //
-  // len(timestamps) == len(tokens)
-  std::vector<float> timestamps;
-};
-
-class OfflineParaformerDecoder {
- public:
-  virtual ~OfflineParaformerDecoder() = default;
-
-  /** Run beam search given the output from the paraformer model.
-   *
-   * @param log_probs A 3-D tensor of shape (N, T, vocab_size)
-   * @param token_num A 1-D tensor of shape (N). token_num equals to T.
-   *
-   * @return Return a vector of size `N` containing the decoded results.
-   */
-  virtual std::vector<OfflineParaformerDecoderResult> Decode(
-      Ort::Value log_probs, Ort::Value token_num,
-      Ort::Value us_cif_peak = Ort::Value(nullptr)) = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_PARAFORMER_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-greedy-search-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-greedy-search-decoder.h
deleted file mode 100644
index eba3fc04..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-greedy-search-decoder.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// sherpa-onnx/csrc/offline-paraformer-greedy-search-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_PARAFORMER_GREEDY_SEARCH_DECODER_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_PARAFORMER_GREEDY_SEARCH_DECODER_H_
-
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-paraformer-decoder.h"
-
-namespace sherpa_onnx {
-
-class OfflineParaformerGreedySearchDecoder : public OfflineParaformerDecoder {
- public:
-  explicit OfflineParaformerGreedySearchDecoder(int32_t eos_id)
-      : eos_id_(eos_id) {}
-
-  std::vector<OfflineParaformerDecoderResult> Decode(
-      Ort::Value log_probs, Ort::Value token_num,
-      Ort::Value us_cif_peak = Ort::Value(nullptr)) override;
-
- private:
-  int32_t eos_id_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_PARAFORMER_GREEDY_SEARCH_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-model-config.h
deleted file mode 100644
index f0420dcb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-model-config.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// sherpa-onnx/csrc/offline-paraformer-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_PARAFORMER_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_PARAFORMER_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineParaformerModelConfig {
-  std::string model;
-
-  OfflineParaformerModelConfig() = default;
-  explicit OfflineParaformerModelConfig(const std::string &model)
-      : model(model) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_PARAFORMER_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-model.h
deleted file mode 100644
index 219f8079..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-paraformer-model.h
+++ /dev/null
@@ -1,74 +0,0 @@
-// sherpa-onnx/csrc/offline-paraformer-model.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_PARAFORMER_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_PARAFORMER_MODEL_H_
-
-#include <memory>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-model-config.h"
-
-namespace sherpa_onnx {
-
-class OfflineParaformerModel {
- public:
-  explicit OfflineParaformerModel(const OfflineModelConfig &config);
-
-  template <typename Manager>
-  OfflineParaformerModel(Manager *mgr, const OfflineModelConfig &config);
-
-  ~OfflineParaformerModel();
-
-  /** Run the forward method of the model.
-   *
-   * @param features  A tensor of shape (N, T, C). It is changed in-place.
-   * @param features_length  A 1-D tensor of shape (N,) containing number of
-   *                         valid frames in `features` before padding.
-   *                         Its dtype is int32_t.
-   *
-   * @return Return a vector containing:
-   *  - log_probs: A 3-D tensor of shape (N, T', vocab_size)
-   *  - token_num: A 1-D tensor of shape (N, T') containing number
-   *               of valid tokens in each utterance. Its dtype is int64_t.
-   *  If it is a model supporting timestamps, then there are additional two
-   *  outputs:
-   *   - us_alphas
-   *   - us_cif_peak
-   */
-  std::vector<Ort::Value> Forward(Ort::Value features,
-                                  Ort::Value features_length);
-
-  /** Return the vocabulary size of the model
-   */
-  int32_t VocabSize() const;
-
-  /** It is lfr_m in config.yaml
-   */
-  int32_t LfrWindowSize() const;
-
-  /** It is lfr_n in config.yaml
-   */
-  int32_t LfrWindowShift() const;
-
-  /** Return negative mean for CMVN
-   */
-  const std::vector<float> &NegativeMean() const;
-
-  /** Return inverse stddev for CMVN
-   */
-  const std::vector<float> &InverseStdDev() const;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_PARAFORMER_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation-ct-transformer-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation-ct-transformer-impl.h
deleted file mode 100644
index eb2c46d6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation-ct-transformer-impl.h
+++ /dev/null
@@ -1,197 +0,0 @@
-// sherpa-onnx/csrc/offline-punctuation-ct-transformer-impl.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_PUNCTUATION_CT_TRANSFORMER_IMPL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_PUNCTUATION_CT_TRANSFORMER_IMPL_H_
-
-#include <math.h>
-
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "sherpa-onnx/csrc/macros.h"
-#include "sherpa-onnx/csrc/math.h"
-#include "sherpa-onnx/csrc/offline-ct-transformer-model.h"
-#include "sherpa-onnx/csrc/offline-punctuation-impl.h"
-#include "sherpa-onnx/csrc/offline-punctuation.h"
-#include "sherpa-onnx/csrc/text-utils.h"
-
-namespace sherpa_onnx {
-
-class OfflinePunctuationCtTransformerImpl : public OfflinePunctuationImpl {
- public:
-  explicit OfflinePunctuationCtTransformerImpl(
-      const OfflinePunctuationConfig &config)
-      : config_(config), model_(config.model) {}
-
-#if __ANDROID_API__ >= 9
-  OfflinePunctuationCtTransformerImpl(AAssetManager *mgr,
-                                      const OfflinePunctuationConfig &config)
-      : config_(config), model_(mgr, config.model) {}
-#endif
-
-  std::string AddPunctuation(const std::string &text) const override {
-    if (text.empty()) {
-      return {};
-    }
-
-    std::vector<std::string> tokens = SplitUtf8(text);
-    std::vector<int32_t> token_ids;
-    token_ids.reserve(tokens.size());
-
-    const auto &meta_data = model_.GetModelMetadata();
-
-    for (const auto &t : tokens) {
-      std::string token = ToLowerCase(t);
-      if (meta_data.token2id.count(token)) {
-        token_ids.push_back(meta_data.token2id.at(token));
-      } else {
-        token_ids.push_back(meta_data.unk_id);
-      }
-    }
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    int32_t segment_size = 20;
-    int32_t max_len = 200;
-    int32_t num_segments =
-        ceil((static_cast<float>(token_ids.size()) + segment_size - 1) /
-             segment_size);
-
-    std::vector<int32_t> punctuations;
-    int32_t last = -1;
-    for (int32_t i = 0; i != num_segments; ++i) {
-      int32_t this_start = i * segment_size;         // included
-      int32_t this_end = this_start + segment_size;  // not included
-      if (this_end > static_cast<int32_t>(token_ids.size())) {
-        this_end = token_ids.size();
-      }
-
-      if (last != -1) {
-        this_start = last;
-      }
-      // token_ids[this_start:this_end] is sent to the model
-
-      std::array<int64_t, 2> x_shape = {1, this_end - this_start};
-      Ort::Value x =
-          Ort::Value::CreateTensor(memory_info, token_ids.data() + this_start,
-                                   x_shape[1], x_shape.data(), x_shape.size());
-
-      int64_t len_shape = 1;
-      int32_t len = x_shape[1];
-      Ort::Value x_len =
-          Ort::Value::CreateTensor(memory_info, &len, 1, &len_shape, 1);
-
-      Ort::Value out = model_.Forward(std::move(x), std::move(x_len));
-
-      // [N, T, num_punctuations]
-      std::vector<int64_t> out_shape =
-          out.GetTensorTypeAndShapeInfo().GetShape();
-
-      assert(out_shape[0] == 1);
-      assert(out_shape[1] == len);
-      assert(out_shape[2] == meta_data.num_punctuations);
-
-      std::vector<int32_t> this_punctuations;
-      this_punctuations.reserve(len);
-
-      const float *p = out.GetTensorData<float>();
-      for (int32_t k = 0; k != len; ++k, p += meta_data.num_punctuations) {
-        auto index = static_cast<int32_t>(std::distance(
-            p, std::max_element(p, p + meta_data.num_punctuations)));
-        this_punctuations.push_back(index);
-      }  // for (int32_t k = 0; k != len; ++k, p += meta_data.num_punctuations)
-
-      int32_t dot_index = -1;
-      int32_t comma_index = -1;
-
-      for (int32_t m = static_cast<int32_t>(this_punctuations.size()) - 2;
-           m >= 1; --m) {
-        int32_t punct_id = this_punctuations[m];
-
-        if (punct_id == meta_data.dot_id || punct_id == meta_data.quest_id) {
-          dot_index = m;
-          break;
-        }
-
-        if (comma_index == -1 && punct_id == meta_data.comma_id) {
-          comma_index = m;
-        }
-      }  // for (int32_t k = this_punctuations.size() - 1; k >= 1; --k)
-
-      if (dot_index == -1 && len >= max_len && comma_index != -1) {
-        dot_index = comma_index;
-        this_punctuations[dot_index] = meta_data.dot_id;
-      }
-
-      if (dot_index == -1) {
-        if (last == -1) {
-          last = this_start;
-        }
-
-        if (i == num_segments - 1) {
-          dot_index = static_cast<int32_t>(this_punctuations.size()) - 1;
-        }
-      } else {
-        last = this_start + dot_index + 1;
-      }
-
-      if (dot_index != -1) {
-        punctuations.insert(punctuations.end(), this_punctuations.begin(),
-                            this_punctuations.begin() + (dot_index + 1));
-      }
-    }  // for (int32_t i = 0; i != num_segments; ++i)
-
-    if (punctuations.empty()) {
-      return text + meta_data.id2punct[meta_data.dot_id];
-    }
-    std::vector<std::string> words_punct;
-
-    for (int32_t i = 0; i != static_cast<int32_t>(punctuations.size()); ++i) {
-      if (i >= static_cast<int32_t>(tokens.size())) {
-        break;
-      }
-      std::string &w = tokens[i];
-      if (i > 0 && !(words_punct.back()[0] & 0x80) && !(w[0] & 0x80)) {
-        words_punct.push_back(" ");
-      }
-      words_punct.push_back(std::move(w));
-
-      if (punctuations[i] != meta_data.underline_id) {
-        words_punct.push_back(meta_data.id2punct[punctuations[i]]);
-      }
-    }
-
-    if (words_punct.back() == meta_data.id2punct[meta_data.comma_id] ||
-        words_punct.back() == meta_data.id2punct[meta_data.pause_id]) {
-      words_punct.back() = meta_data.id2punct[meta_data.dot_id];
-    }
-
-    if (words_punct.back() != meta_data.id2punct[meta_data.dot_id] &&
-        words_punct.back() != meta_data.id2punct[meta_data.quest_id]) {
-      words_punct.push_back(meta_data.id2punct[meta_data.dot_id]);
-    }
-
-    std::string ans;
-    for (const auto &w : words_punct) {
-      ans.append(w);
-    }
-    return ans;
-  }
-
- private:
-  OfflinePunctuationConfig config_;
-  OfflineCtTransformerModel model_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_PUNCTUATION_CT_TRANSFORMER_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation-impl.h
deleted file mode 100644
index d0538aab..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation-impl.h
+++ /dev/null
@@ -1,36 +0,0 @@
-// sherpa-onnx/csrc/offline-punctuation-impl.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_PUNCTUATION_IMPL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_PUNCTUATION_IMPL_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "sherpa-onnx/csrc/offline-punctuation.h"
-
-namespace sherpa_onnx {
-
-class OfflinePunctuationImpl {
- public:
-  virtual ~OfflinePunctuationImpl() = default;
-
-  static std::unique_ptr<OfflinePunctuationImpl> Create(
-      const OfflinePunctuationConfig &config);
-
-#if __ANDROID_API__ >= 9
-  static std::unique_ptr<OfflinePunctuationImpl> Create(
-      AAssetManager *mgr, const OfflinePunctuationConfig &config);
-#endif
-
-  virtual std::string AddPunctuation(const std::string &text) const = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_PUNCTUATION_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation-model-config.h
deleted file mode 100644
index aa294f3f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation-model-config.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// sherpa-onnx/csrc/offline-punctuation-model-config.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_PUNCTUATION_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_PUNCTUATION_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflinePunctuationModelConfig {
-  std::string ct_transformer;
-
-  int32_t num_threads = 1;
-  bool debug = false;
-  std::string provider = "cpu";
-
-  OfflinePunctuationModelConfig() = default;
-
-  OfflinePunctuationModelConfig(const std::string &ct_transformer,
-                                int32_t num_threads, bool debug,
-                                const std::string &provider)
-      : ct_transformer(ct_transformer),
-        num_threads(num_threads),
-        debug(debug),
-        provider(provider) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_PUNCTUATION_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation.h
deleted file mode 100644
index 831e7cc9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-punctuation.h
+++ /dev/null
@@ -1,57 +0,0 @@
-// sherpa-onnx/csrc/offline-punctuation.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_PUNCTUATION_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_PUNCTUATION_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "sherpa-onnx/csrc/offline-punctuation-model-config.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflinePunctuationConfig {
-  OfflinePunctuationModelConfig model;
-
-  OfflinePunctuationConfig() = default;
-
-  explicit OfflinePunctuationConfig(const OfflinePunctuationModelConfig &model)
-      : model(model) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-class OfflinePunctuationImpl;
-
-class OfflinePunctuation {
- public:
-  explicit OfflinePunctuation(const OfflinePunctuationConfig &config);
-
-#if __ANDROID_API__ >= 9
-  OfflinePunctuation(AAssetManager *mgr,
-                     const OfflinePunctuationConfig &config);
-#endif
-
-  ~OfflinePunctuation();
-
-  // Add punctuation to the input text and return it.
-  std::string AddPunctuation(const std::string &text) const;
-
- private:
-  std::unique_ptr<OfflinePunctuationImpl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_PUNCTUATION_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-ctc-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-ctc-impl.h
deleted file mode 100644
index 2721ecdf..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-ctc-impl.h
+++ /dev/null
@@ -1,269 +0,0 @@
-// sherpa-onnx/csrc/offline-recognizer-ctc-impl.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_CTC_IMPL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_CTC_IMPL_H_
-
-#include <ios>
-#include <memory>
-#include <sstream>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-ctc-decoder.h"
-#include "sherpa-onnx/csrc/offline-ctc-fst-decoder.h"
-#include "sherpa-onnx/csrc/offline-ctc-greedy-search-decoder.h"
-#include "sherpa-onnx/csrc/offline-ctc-model.h"
-#include "sherpa-onnx/csrc/offline-recognizer-impl.h"
-#include "sherpa-onnx/csrc/pad-sequence.h"
-#include "sherpa-onnx/csrc/symbol-table.h"
-
-namespace sherpa_onnx {
-
-static OfflineRecognitionResult Convert(const OfflineCtcDecoderResult &src,
-                                        const SymbolTable &sym_table,
-                                        int32_t frame_shift_ms,
-                                        int32_t subsampling_factor) {
-  OfflineRecognitionResult r;
-  r.tokens.reserve(src.tokens.size());
-  r.timestamps.reserve(src.timestamps.size());
-
-  std::string text;
-
-  for (int32_t i = 0; i != src.tokens.size(); ++i) {
-    if (sym_table.Contains("SIL") && src.tokens[i] == sym_table["SIL"]) {
-      // tdnn models from yesno have a SIL token, we should remove it.
-      continue;
-    }
-    auto sym = sym_table[src.tokens[i]];
-    text.append(sym);
-
-    if (sym.size() == 1 && (sym[0] < 0x20 || sym[0] > 0x7e)) {
-      // for byte bpe models
-      // (but don't rewrite printable characters 0x20..0x7e,
-      //  which collide with standard BPE units)
-      std::ostringstream os;
-      os << "<0x" << std::hex << std::uppercase
-         << (static_cast<int32_t>(sym[0]) & 0xff) << ">";
-      sym = os.str();
-    }
-
-    r.tokens.push_back(std::move(sym));
-  }
-  r.text = std::move(text);
-
-  float frame_shift_s = frame_shift_ms / 1000. * subsampling_factor;
-  for (auto t : src.timestamps) {
-    float time = frame_shift_s * t;
-    r.timestamps.push_back(time);
-  }
-
-  r.words = std::move(src.words);
-
-  return r;
-}
-
-class OfflineRecognizerCtcImpl : public OfflineRecognizerImpl {
- public:
-  explicit OfflineRecognizerCtcImpl(const OfflineRecognizerConfig &config)
-      : OfflineRecognizerImpl(config),
-        config_(config),
-        symbol_table_(config_.model_config.tokens),
-        model_(OfflineCtcModel::Create(config_.model_config)) {
-    Init();
-  }
-
-  template <typename Manager>
-  OfflineRecognizerCtcImpl(Manager *mgr, const OfflineRecognizerConfig &config)
-      : OfflineRecognizerImpl(mgr, config),
-        config_(config),
-        symbol_table_(mgr, config_.model_config.tokens),
-        model_(OfflineCtcModel::Create(mgr, config_.model_config)) {
-    Init();
-  }
-
-  void Init() {
-    if (!config_.model_config.telespeech_ctc.empty()) {
-      config_.feat_config.snip_edges = true;
-      config_.feat_config.num_ceps = 40;
-      config_.feat_config.feature_dim = 40;
-      config_.feat_config.low_freq = 40;
-      config_.feat_config.high_freq = -200;
-      config_.feat_config.use_energy = false;
-      config_.feat_config.normalize_samples = false;
-      config_.feat_config.is_mfcc = true;
-    }
-
-    if (!config_.model_config.nemo_ctc.model.empty()) {
-      if (model_->IsGigaAM()) {
-        config_.feat_config.low_freq = 0;
-        config_.feat_config.high_freq = 8000;
-        config_.feat_config.remove_dc_offset = false;
-        config_.feat_config.preemph_coeff = 0;
-        config_.feat_config.window_type = "hann";
-        config_.feat_config.feature_dim = 64;
-      } else {
-        config_.feat_config.low_freq = 0;
-        config_.feat_config.high_freq = 0;
-        config_.feat_config.is_librosa = true;
-        config_.feat_config.remove_dc_offset = false;
-        config_.feat_config.window_type = "hann";
-      }
-    }
-
-    if (!config_.model_config.wenet_ctc.model.empty()) {
-      // WeNet CTC models assume input samples are in the range
-      // [-32768, 32767], so we set normalize_samples to false
-      config_.feat_config.normalize_samples = false;
-    }
-
-    config_.feat_config.nemo_normalize_type =
-        model_->FeatureNormalizationMethod();
-
-    if (!config_.ctc_fst_decoder_config.graph.empty()) {
-      // TODO(fangjun): Support android to read the graph from
-      // asset_manager
-      decoder_ = std::make_unique<OfflineCtcFstDecoder>(
-          config_.ctc_fst_decoder_config);
-    } else if (config_.decoding_method == "greedy_search") {
-      if (!symbol_table_.Contains("<blk>") &&
-          !symbol_table_.Contains("<eps>") &&
-          !symbol_table_.Contains("<blank>")) {
-        SHERPA_ONNX_LOGE(
-            "We expect that tokens.txt contains "
-            "the symbol <blk> or <eps> or <blank> and its ID.");
-        exit(-1);
-      }
-
-      int32_t blank_id = 0;
-      if (symbol_table_.Contains("<blk>")) {
-        blank_id = symbol_table_["<blk>"];
-      } else if (symbol_table_.Contains("<eps>")) {
-        // for tdnn models of the yesno recipe from icefall
-        blank_id = symbol_table_["<eps>"];
-      } else if (symbol_table_.Contains("<blank>")) {
-        // for Wenet CTC models
-        blank_id = symbol_table_["<blank>"];
-      }
-
-      decoder_ = std::make_unique<OfflineCtcGreedySearchDecoder>(blank_id);
-    } else {
-      SHERPA_ONNX_LOGE("Only greedy_search is supported at present. Given %s",
-                       config_.decoding_method.c_str());
-      exit(-1);
-    }
-  }
-
-  std::unique_ptr<OfflineStream> CreateStream() const override {
-    return std::make_unique<OfflineStream>(config_.feat_config);
-  }
-
-  void DecodeStreams(OfflineStream **ss, int32_t n) const override {
-    if (!model_->SupportBatchProcessing()) {
-      // If the model does not support batch process,
-      // we process each stream independently.
-      for (int32_t i = 0; i != n; ++i) {
-        DecodeStream(ss[i]);
-      }
-      return;
-    }
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    int32_t feat_dim = config_.feat_config.feature_dim;
-
-    std::vector<Ort::Value> features;
-    features.reserve(n);
-
-    std::vector<std::vector<float>> features_vec(n);
-    std::vector<int64_t> features_length_vec(n);
-
-    for (int32_t i = 0; i != n; ++i) {
-      std::vector<float> f = ss[i]->GetFrames();
-
-      int32_t num_frames = f.size() / feat_dim;
-      features_vec[i] = std::move(f);
-
-      features_length_vec[i] = num_frames;
-
-      std::array<int64_t, 2> shape = {num_frames, feat_dim};
-
-      Ort::Value x = Ort::Value::CreateTensor(
-          memory_info, features_vec[i].data(), features_vec[i].size(),
-          shape.data(), shape.size());
-      features.push_back(std::move(x));
-    }  // for (int32_t i = 0; i != n; ++i)
-
-    std::vector<const Ort::Value *> features_pointer(n);
-    for (int32_t i = 0; i != n; ++i) {
-      features_pointer[i] = &features[i];
-    }
-
-    std::array<int64_t, 1> features_length_shape = {n};
-    Ort::Value x_length = Ort::Value::CreateTensor(
-        memory_info, features_length_vec.data(), n,
-        features_length_shape.data(), features_length_shape.size());
-
-    Ort::Value x = PadSequence(model_->Allocator(), features_pointer,
-                               -23.025850929940457f);
-    auto t = model_->Forward(std::move(x), std::move(x_length));
-
-    auto results = decoder_->Decode(std::move(t[0]), std::move(t[1]));
-
-    int32_t frame_shift_ms = 10;
-    for (int32_t i = 0; i != n; ++i) {
-      auto r = Convert(results[i], symbol_table_, frame_shift_ms,
-                       model_->SubsamplingFactor());
-      r.text = ApplyInverseTextNormalization(std::move(r.text));
-      ss[i]->SetResult(r);
-    }
-  }
-
-  OfflineRecognizerConfig GetConfig() const override { return config_; }
-
- private:
-  // Decode a single stream.
-  // Some models do not support batch size > 1, e.g., WeNet CTC models.
-  void DecodeStream(OfflineStream *s) const {
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    int32_t feat_dim = config_.feat_config.feature_dim;
-    std::vector<float> f = s->GetFrames();
-
-    int32_t num_frames = f.size() / feat_dim;
-
-    std::array<int64_t, 3> shape = {1, num_frames, feat_dim};
-
-    Ort::Value x = Ort::Value::CreateTensor(memory_info, f.data(), f.size(),
-                                            shape.data(), shape.size());
-
-    int64_t x_length_scalar = num_frames;
-    std::array<int64_t, 1> x_length_shape = {1};
-    Ort::Value x_length =
-        Ort::Value::CreateTensor(memory_info, &x_length_scalar, 1,
-                                 x_length_shape.data(), x_length_shape.size());
-
-    auto t = model_->Forward(std::move(x), std::move(x_length));
-    auto results = decoder_->Decode(std::move(t[0]), std::move(t[1]));
-    int32_t frame_shift_ms = 10;
-
-    auto r = Convert(results[0], symbol_table_, frame_shift_ms,
-                     model_->SubsamplingFactor());
-    r.text = ApplyInverseTextNormalization(std::move(r.text));
-    s->SetResult(r);
-  }
-
- private:
-  OfflineRecognizerConfig config_;
-  SymbolTable symbol_table_;
-  std::unique_ptr<OfflineCtcModel> model_;
-  std::unique_ptr<OfflineCtcDecoder> decoder_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_CTC_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-impl.h
deleted file mode 100644
index 8a6e6fcc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-impl.h
+++ /dev/null
@@ -1,61 +0,0 @@
-// sherpa-onnx/csrc/offline-recognizer-impl.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_IMPL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_IMPL_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "kaldifst/csrc/text-normalizer.h"
-#include "sherpa-onnx/csrc/macros.h"
-#include "sherpa-onnx/csrc/offline-recognizer.h"
-#include "sherpa-onnx/csrc/offline-stream.h"
-
-namespace sherpa_onnx {
-
-class OfflineRecognizerImpl {
- public:
-  explicit OfflineRecognizerImpl(const OfflineRecognizerConfig &config);
-
-  static std::unique_ptr<OfflineRecognizerImpl> Create(
-      const OfflineRecognizerConfig &config);
-
-  template <typename Manager>
-  OfflineRecognizerImpl(Manager *mgr, const OfflineRecognizerConfig &config);
-
-  template <typename Manager>
-  static std::unique_ptr<OfflineRecognizerImpl> Create(
-      Manager *mgr, const OfflineRecognizerConfig &config);
-
-  virtual ~OfflineRecognizerImpl() = default;
-
-  virtual std::unique_ptr<OfflineStream> CreateStream(
-      const std::string &hotwords) const {
-    SHERPA_ONNX_LOGE("Only transducer models support contextual biasing.");
-    exit(-1);
-  }
-
-  virtual std::unique_ptr<OfflineStream> CreateStream() const = 0;
-
-  virtual void DecodeStreams(OfflineStream **ss, int32_t n) const = 0;
-
-  virtual void SetConfig(const OfflineRecognizerConfig &config);
-
-  virtual OfflineRecognizerConfig GetConfig() const = 0;
-
-  std::string ApplyInverseTextNormalization(std::string text) const;
-
- private:
-  OfflineRecognizerConfig config_;
-  // for inverse text normalization. Used only if
-  // config.rule_fsts is not empty or
-  // config.rule_fars is not empty
-  std::vector<std::unique_ptr<kaldifst::TextNormalizer>> itn_list_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-moonshine-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-moonshine-impl.h
deleted file mode 100644
index deec9852..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-moonshine-impl.h
+++ /dev/null
@@ -1,143 +0,0 @@
-// sherpa-onnx/csrc/offline-recognizer-moonshine-impl.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_MOONSHINE_IMPL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_MOONSHINE_IMPL_H_
-
-#include <algorithm>
-#include <cmath>
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-model-config.h"
-#include "sherpa-onnx/csrc/offline-moonshine-decoder.h"
-#include "sherpa-onnx/csrc/offline-moonshine-greedy-search-decoder.h"
-#include "sherpa-onnx/csrc/offline-moonshine-model.h"
-#include "sherpa-onnx/csrc/offline-recognizer-impl.h"
-#include "sherpa-onnx/csrc/offline-recognizer.h"
-#include "sherpa-onnx/csrc/symbol-table.h"
-#include "sherpa-onnx/csrc/transpose.h"
-
-namespace sherpa_onnx {
-
-static OfflineRecognitionResult Convert(
-    const OfflineMoonshineDecoderResult &src, const SymbolTable &sym_table) {
-  OfflineRecognitionResult r;
-  r.tokens.reserve(src.tokens.size());
-
-  std::string text;
-  for (auto i : src.tokens) {
-    if (!sym_table.Contains(i)) {
-      continue;
-    }
-
-    const auto &s = sym_table[i];
-    text += s;
-    r.tokens.push_back(s);
-  }
-
-  r.text = text;
-
-  return r;
-}
-
-class OfflineRecognizerMoonshineImpl : public OfflineRecognizerImpl {
- public:
-  explicit OfflineRecognizerMoonshineImpl(const OfflineRecognizerConfig &config)
-      : OfflineRecognizerImpl(config),
-        config_(config),
-        symbol_table_(config_.model_config.tokens),
-        model_(std::make_unique<OfflineMoonshineModel>(config.model_config)) {
-    Init();
-  }
-
-  template <typename Manager>
-  OfflineRecognizerMoonshineImpl(Manager *mgr,
-                                 const OfflineRecognizerConfig &config)
-      : OfflineRecognizerImpl(mgr, config),
-        config_(config),
-        symbol_table_(mgr, config_.model_config.tokens),
-        model_(
-            std::make_unique<OfflineMoonshineModel>(mgr, config.model_config)) {
-    Init();
-  }
-
-  void Init() {
-    if (config_.decoding_method == "greedy_search") {
-      decoder_ =
-          std::make_unique<OfflineMoonshineGreedySearchDecoder>(model_.get());
-    } else {
-      SHERPA_ONNX_LOGE(
-          "Only greedy_search is supported at present for moonshine. Given %s",
-          config_.decoding_method.c_str());
-      exit(-1);
-    }
-  }
-
-  std::unique_ptr<OfflineStream> CreateStream() const override {
-    MoonshineTag tag;
-    return std::make_unique<OfflineStream>(tag);
-  }
-
-  void DecodeStreams(OfflineStream **ss, int32_t n) const override {
-    // batch decoding is not implemented yet
-    for (int32_t i = 0; i != n; ++i) {
-      DecodeStream(ss[i]);
-    }
-  }
-
-  OfflineRecognizerConfig GetConfig() const override { return config_; }
-
- private:
-  void DecodeStream(OfflineStream *s) const {
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    std::vector<float> audio = s->GetFrames();
-
-    try {
-      std::array<int64_t, 2> shape{1, static_cast<int64_t>(audio.size())};
-
-      Ort::Value audio_tensor = Ort::Value::CreateTensor(
-          memory_info, audio.data(), audio.size(), shape.data(), shape.size());
-
-      Ort::Value features =
-          model_->ForwardPreprocessor(std::move(audio_tensor));
-
-      int32_t features_len = features.GetTensorTypeAndShapeInfo().GetShape()[1];
-
-      int64_t features_shape = 1;
-
-      Ort::Value features_len_tensor = Ort::Value::CreateTensor(
-          memory_info, &features_len, 1, &features_shape, 1);
-
-      Ort::Value encoder_out = model_->ForwardEncoder(
-          std::move(features), std::move(features_len_tensor));
-
-      auto results = decoder_->Decode(std::move(encoder_out));
-
-      auto r = Convert(results[0], symbol_table_);
-      r.text = ApplyInverseTextNormalization(std::move(r.text));
-      s->SetResult(r);
-    } catch (const Ort::Exception &ex) {
-      SHERPA_ONNX_LOGE(
-          "\n\nCaught exception:\n\n%s\n\nReturn an empty result. Number of "
-          "audio samples: %d",
-          ex.what(), static_cast<int32_t>(audio.size()));
-      return;
-    }
-  }
-
- private:
-  OfflineRecognizerConfig config_;
-  SymbolTable symbol_table_;
-  std::unique_ptr<OfflineMoonshineModel> model_;
-  std::unique_ptr<OfflineMoonshineDecoder> decoder_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_MOONSHINE_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-paraformer-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-paraformer-impl.h
deleted file mode 100644
index 5b80c99e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-paraformer-impl.h
+++ /dev/null
@@ -1,267 +0,0 @@
-// sherpa-onnx/csrc/offline-recognizer-paraformer-impl.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_PARAFORMER_IMPL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_PARAFORMER_IMPL_H_
-
-#include <algorithm>
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-model-config.h"
-#include "sherpa-onnx/csrc/offline-paraformer-decoder.h"
-#include "sherpa-onnx/csrc/offline-paraformer-greedy-search-decoder.h"
-#include "sherpa-onnx/csrc/offline-paraformer-model.h"
-#include "sherpa-onnx/csrc/offline-recognizer-impl.h"
-#include "sherpa-onnx/csrc/offline-recognizer.h"
-#include "sherpa-onnx/csrc/pad-sequence.h"
-#include "sherpa-onnx/csrc/symbol-table.h"
-
-namespace sherpa_onnx {
-
-static OfflineRecognitionResult Convert(
-    const OfflineParaformerDecoderResult &src, const SymbolTable &sym_table) {
-  OfflineRecognitionResult r;
-  r.tokens.reserve(src.tokens.size());
-  r.timestamps = src.timestamps;
-
-  std::string text;
-
-  // When the current token ends with "@@" we set mergeable to true
-  bool mergeable = false;
-
-  for (int32_t i = 0; i != src.tokens.size(); ++i) {
-    auto sym = sym_table[src.tokens[i]];
-    r.tokens.push_back(sym);
-
-    if ((sym.back() != '@') || (sym.size() > 2 && sym[sym.size() - 2] != '@')) {
-      // sym does not end with "@@"
-      const uint8_t *p = reinterpret_cast<const uint8_t *>(sym.c_str());
-      if (p[0] < 0x80) {
-        // an ascii
-        if (mergeable) {
-          mergeable = false;
-          text.append(sym);
-        } else {
-          text.append(" ");
-          text.append(sym);
-        }
-      } else {
-        // not an ascii
-        mergeable = false;
-
-        if (i > 0) {
-          const uint8_t p = reinterpret_cast<const uint8_t *>(
-              sym_table[src.tokens[i - 1]].c_str())[0];
-          if (p < 0x80) {
-            // put a space between ascii and non-ascii
-            text.append(" ");
-          }
-        }
-        text.append(sym);
-      }
-    } else {
-      // this sym ends with @@
-      sym = std::string(sym.data(), sym.size() - 2);
-      if (mergeable) {
-        text.append(sym);
-      } else {
-        text.append(" ");
-        text.append(sym);
-        mergeable = true;
-      }
-    }
-  }
-  r.text = std::move(text);
-
-  return r;
-}
-
-class OfflineRecognizerParaformerImpl : public OfflineRecognizerImpl {
- public:
-  explicit OfflineRecognizerParaformerImpl(
-      const OfflineRecognizerConfig &config)
-      : OfflineRecognizerImpl(config),
-        config_(config),
-        symbol_table_(config_.model_config.tokens),
-        model_(std::make_unique<OfflineParaformerModel>(config.model_config)) {
-    if (config.decoding_method == "greedy_search") {
-      int32_t eos_id = symbol_table_["</s>"];
-      decoder_ = std::make_unique<OfflineParaformerGreedySearchDecoder>(eos_id);
-    } else {
-      SHERPA_ONNX_LOGE("Only greedy_search is supported at present. Given %s",
-                       config.decoding_method.c_str());
-      exit(-1);
-    }
-
-    InitFeatConfig();
-  }
-
-  template <typename Manager>
-  OfflineRecognizerParaformerImpl(Manager *mgr,
-                                  const OfflineRecognizerConfig &config)
-      : OfflineRecognizerImpl(mgr, config),
-        config_(config),
-        symbol_table_(mgr, config_.model_config.tokens),
-        model_(std::make_unique<OfflineParaformerModel>(mgr,
-                                                        config.model_config)) {
-    if (config.decoding_method == "greedy_search") {
-      int32_t eos_id = symbol_table_["</s>"];
-      decoder_ = std::make_unique<OfflineParaformerGreedySearchDecoder>(eos_id);
-    } else {
-      SHERPA_ONNX_LOGE("Only greedy_search is supported at present. Given %s",
-                       config.decoding_method.c_str());
-      exit(-1);
-    }
-
-    InitFeatConfig();
-  }
-
-  std::unique_ptr<OfflineStream> CreateStream() const override {
-    return std::make_unique<OfflineStream>(config_.feat_config);
-  }
-
-  void DecodeStreams(OfflineStream **ss, int32_t n) const override {
-    // 1. Apply LFR
-    // 2. Apply CMVN
-    //
-    // Please refer to
-    // https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45555.pdf
-    // for what LFR means
-    //
-    // "Lower Frame Rate Neural Network Acoustic Models"
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    std::vector<Ort::Value> features;
-    features.reserve(n);
-
-    int32_t feat_dim =
-        config_.feat_config.feature_dim * model_->LfrWindowSize();
-
-    std::vector<std::vector<float>> features_vec(n);
-    std::vector<int32_t> features_length_vec(n);
-    for (int32_t i = 0; i != n; ++i) {
-      std::vector<float> f = ss[i]->GetFrames();
-
-      f = ApplyLFR(f);
-      ApplyCMVN(&f);
-
-      int32_t num_frames = f.size() / feat_dim;
-      features_vec[i] = std::move(f);
-
-      features_length_vec[i] = num_frames;
-
-      std::array<int64_t, 2> shape = {num_frames, feat_dim};
-
-      Ort::Value x = Ort::Value::CreateTensor(
-          memory_info, features_vec[i].data(), features_vec[i].size(),
-          shape.data(), shape.size());
-      features.push_back(std::move(x));
-    }
-
-    std::vector<const Ort::Value *> features_pointer(n);
-    for (int32_t i = 0; i != n; ++i) {
-      features_pointer[i] = &features[i];
-    }
-
-    std::array<int64_t, 1> features_length_shape = {n};
-    Ort::Value x_length = Ort::Value::CreateTensor(
-        memory_info, features_length_vec.data(), n,
-        features_length_shape.data(), features_length_shape.size());
-
-    // Caution(fangjun): We cannot pad it with log(eps),
-    // i.e., -23.025850929940457f
-    Ort::Value x = PadSequence(model_->Allocator(), features_pointer, 0);
-
-    std::vector<Ort::Value> t;
-    try {
-      t = model_->Forward(std::move(x), std::move(x_length));
-    } catch (const Ort::Exception &ex) {
-      SHERPA_ONNX_LOGE("\n\nCaught exception:\n\n%s\n\nReturn an empty result",
-                       ex.what());
-      return;
-    }
-
-    std::vector<OfflineParaformerDecoderResult> results;
-    if (t.size() == 2) {
-      results = decoder_->Decode(std::move(t[0]), std::move(t[1]));
-    } else {
-      results =
-          decoder_->Decode(std::move(t[0]), std::move(t[1]), std::move(t[3]));
-    }
-
-    for (int32_t i = 0; i != n; ++i) {
-      auto r = Convert(results[i], symbol_table_);
-      r.text = ApplyInverseTextNormalization(std::move(r.text));
-      ss[i]->SetResult(r);
-    }
-  }
-
-  OfflineRecognizerConfig GetConfig() const override { return config_; }
-
- private:
-  void InitFeatConfig() {
-    // Paraformer models assume input samples are in the range
-    // [-32768, 32767], so we set normalize_samples to false
-    config_.feat_config.normalize_samples = false;
-    config_.feat_config.window_type = "hamming";
-    config_.feat_config.high_freq = 0;
-    config_.feat_config.snip_edges = true;
-  }
-
-  std::vector<float> ApplyLFR(const std::vector<float> &in) const {
-    int32_t lfr_window_size = model_->LfrWindowSize();
-    int32_t lfr_window_shift = model_->LfrWindowShift();
-    int32_t in_feat_dim = config_.feat_config.feature_dim;
-
-    int32_t in_num_frames = in.size() / in_feat_dim;
-    int32_t out_num_frames =
-        (in_num_frames - lfr_window_size) / lfr_window_shift + 1;
-    int32_t out_feat_dim = in_feat_dim * lfr_window_size;
-
-    std::vector<float> out(out_num_frames * out_feat_dim);
-
-    const float *p_in = in.data();
-    float *p_out = out.data();
-
-    for (int32_t i = 0; i != out_num_frames; ++i) {
-      std::copy(p_in, p_in + out_feat_dim, p_out);
-
-      p_out += out_feat_dim;
-      p_in += lfr_window_shift * in_feat_dim;
-    }
-
-    return out;
-  }
-
-  void ApplyCMVN(std::vector<float> *v) const {
-    const std::vector<float> &neg_mean = model_->NegativeMean();
-    const std::vector<float> &inv_stddev = model_->InverseStdDev();
-
-    int32_t dim = neg_mean.size();
-    int32_t num_frames = v->size() / dim;
-
-    float *p = v->data();
-
-    for (int32_t i = 0; i != num_frames; ++i) {
-      for (int32_t k = 0; k != dim; ++k) {
-        p[k] = (p[k] + neg_mean[k]) * inv_stddev[k];
-      }
-
-      p += dim;
-    }
-  }
-
-  OfflineRecognizerConfig config_;
-  SymbolTable symbol_table_;
-  std::unique_ptr<OfflineParaformerModel> model_;
-  std::unique_ptr<OfflineParaformerDecoder> decoder_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_PARAFORMER_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-sense-voice-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-sense-voice-impl.h
deleted file mode 100644
index 7ee5e41c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-sense-voice-impl.h
+++ /dev/null
@@ -1,364 +0,0 @@
-// sherpa-onnx/csrc/offline-recognizer-sense-voice-impl.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_SENSE_VOICE_IMPL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_SENSE_VOICE_IMPL_H_
-
-#include <algorithm>
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-ctc-greedy-search-decoder.h"
-#include "sherpa-onnx/csrc/offline-model-config.h"
-#include "sherpa-onnx/csrc/offline-recognizer-impl.h"
-#include "sherpa-onnx/csrc/offline-recognizer.h"
-#include "sherpa-onnx/csrc/offline-sense-voice-model.h"
-#include "sherpa-onnx/csrc/pad-sequence.h"
-#include "sherpa-onnx/csrc/symbol-table.h"
-
-namespace sherpa_onnx {
-
-static OfflineRecognitionResult ConvertSenseVoiceResult(
-    const OfflineCtcDecoderResult &src, const SymbolTable &sym_table,
-    int32_t frame_shift_ms, int32_t subsampling_factor) {
-  OfflineRecognitionResult r;
-  r.tokens.reserve(src.tokens.size());
-  r.timestamps.reserve(src.timestamps.size());
-
-  std::string text;
-
-  for (int32_t i = 4; i < src.tokens.size(); ++i) {
-    auto sym = sym_table[src.tokens[i]];
-    text.append(sym);
-
-    r.tokens.push_back(std::move(sym));
-  }
-  r.text = std::move(text);
-
-  float frame_shift_s = frame_shift_ms / 1000. * subsampling_factor;
-
-  for (int32_t i = 4; i < src.timestamps.size(); ++i) {
-    float time = frame_shift_s * (src.timestamps[i] - 4);
-    r.timestamps.push_back(time);
-  }
-
-  r.words = std::move(src.words);
-
-  // parse lang, emotion and event from tokens.
-  if (src.tokens.size() >= 3) {
-    r.lang = sym_table[src.tokens[0]];
-    r.emotion = sym_table[src.tokens[1]];
-    r.event = sym_table[src.tokens[2]];
-  }
-
-  return r;
-}
-
-class OfflineRecognizerSenseVoiceImpl : public OfflineRecognizerImpl {
- public:
-  explicit OfflineRecognizerSenseVoiceImpl(
-      const OfflineRecognizerConfig &config)
-      : OfflineRecognizerImpl(config),
-        config_(config),
-        symbol_table_(config_.model_config.tokens),
-        model_(std::make_unique<OfflineSenseVoiceModel>(config.model_config)) {
-    const auto &meta_data = model_->GetModelMetadata();
-    if (config.decoding_method == "greedy_search") {
-      decoder_ =
-          std::make_unique<OfflineCtcGreedySearchDecoder>(meta_data.blank_id);
-    } else {
-      SHERPA_ONNX_LOGE("Only greedy_search is supported at present. Given %s",
-                       config.decoding_method.c_str());
-      exit(-1);
-    }
-
-    InitFeatConfig();
-  }
-
-  template <typename Manager>
-  OfflineRecognizerSenseVoiceImpl(Manager *mgr,
-                                  const OfflineRecognizerConfig &config)
-      : OfflineRecognizerImpl(mgr, config),
-        config_(config),
-        symbol_table_(mgr, config_.model_config.tokens),
-        model_(std::make_unique<OfflineSenseVoiceModel>(mgr,
-                                                        config.model_config)) {
-    const auto &meta_data = model_->GetModelMetadata();
-    if (config.decoding_method == "greedy_search") {
-      decoder_ =
-          std::make_unique<OfflineCtcGreedySearchDecoder>(meta_data.blank_id);
-    } else {
-      SHERPA_ONNX_LOGE("Only greedy_search is supported at present. Given %s",
-                       config.decoding_method.c_str());
-      exit(-1);
-    }
-
-    InitFeatConfig();
-  }
-
-  std::unique_ptr<OfflineStream> CreateStream() const override {
-    return std::make_unique<OfflineStream>(config_.feat_config);
-  }
-
-  void DecodeStreams(OfflineStream **ss, int32_t n) const override {
-    if (n == 1) {
-      DecodeOneStream(ss[0]);
-      return;
-    }
-
-    const auto &meta_data = model_->GetModelMetadata();
-    // 1. Apply LFR
-    // 2. Apply CMVN
-    //
-    // Please refer to
-    // https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45555.pdf
-    // for what LFR means
-    //
-    // "Lower Frame Rate Neural Network Acoustic Models"
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    std::vector<Ort::Value> features;
-    features.reserve(n);
-
-    int32_t feat_dim = config_.feat_config.feature_dim * meta_data.window_size;
-
-    std::vector<std::vector<float>> features_vec(n);
-    std::vector<int32_t> features_length_vec(n);
-    for (int32_t i = 0; i != n; ++i) {
-      std::vector<float> f = ss[i]->GetFrames();
-
-      f = ApplyLFR(f);
-      ApplyCMVN(&f);
-
-      int32_t num_frames = f.size() / feat_dim;
-      features_vec[i] = std::move(f);
-
-      features_length_vec[i] = num_frames;
-
-      std::array<int64_t, 2> shape = {num_frames, feat_dim};
-
-      Ort::Value x = Ort::Value::CreateTensor(
-          memory_info, features_vec[i].data(), features_vec[i].size(),
-          shape.data(), shape.size());
-      features.push_back(std::move(x));
-    }
-
-    std::vector<const Ort::Value *> features_pointer(n);
-    for (int32_t i = 0; i != n; ++i) {
-      features_pointer[i] = &features[i];
-    }
-
-    std::array<int64_t, 1> features_length_shape = {n};
-    Ort::Value x_length = Ort::Value::CreateTensor(
-        memory_info, features_length_vec.data(), n,
-        features_length_shape.data(), features_length_shape.size());
-
-    // Caution(fangjun): We cannot pad it with log(eps),
-    // i.e., -23.025850929940457f
-    Ort::Value x = PadSequence(model_->Allocator(), features_pointer, 0);
-
-    int32_t language = 0;
-    if (config_.model_config.sense_voice.language.empty()) {
-      language = 0;
-    } else if (meta_data.lang2id.count(
-                   config_.model_config.sense_voice.language)) {
-      language =
-          meta_data.lang2id.at(config_.model_config.sense_voice.language);
-    } else {
-      SHERPA_ONNX_LOGE("Unknown language: %s. Use 0 instead.",
-                       config_.model_config.sense_voice.language.c_str());
-    }
-
-    std::vector<int32_t> language_array(n);
-    std::fill(language_array.begin(), language_array.end(), language);
-
-    std::vector<int32_t> text_norm_array(n);
-    std::fill(text_norm_array.begin(), text_norm_array.end(),
-              config_.model_config.sense_voice.use_itn
-                  ? meta_data.with_itn_id
-                  : meta_data.without_itn_id);
-
-    Ort::Value language_tensor = Ort::Value::CreateTensor(
-        memory_info, language_array.data(), n, features_length_shape.data(),
-        features_length_shape.size());
-
-    Ort::Value text_norm_tensor = Ort::Value::CreateTensor(
-        memory_info, text_norm_array.data(), n, features_length_shape.data(),
-        features_length_shape.size());
-
-    Ort::Value logits{nullptr};
-    try {
-      logits = model_->Forward(std::move(x), std::move(x_length),
-                               std::move(language_tensor),
-                               std::move(text_norm_tensor));
-    } catch (const Ort::Exception &ex) {
-      SHERPA_ONNX_LOGE("\n\nCaught exception:\n\n%s\n\nReturn an empty result",
-                       ex.what());
-      return;
-    }
-
-    // decoder_->Decode() requires that logits_length is of dtype int64
-    std::vector<int64_t> features_length_vec_64;
-    features_length_vec_64.reserve(n);
-    for (auto i : features_length_vec) {
-      i += 4;
-      features_length_vec_64.push_back(i);
-    }
-
-    Ort::Value logits_length = Ort::Value::CreateTensor(
-        memory_info, features_length_vec_64.data(), n,
-        features_length_shape.data(), features_length_shape.size());
-
-    auto results =
-        decoder_->Decode(std::move(logits), std::move(logits_length));
-
-    int32_t frame_shift_ms = 10;
-    int32_t subsampling_factor = meta_data.window_shift;
-    for (int32_t i = 0; i != n; ++i) {
-      auto r = ConvertSenseVoiceResult(results[i], symbol_table_,
-                                       frame_shift_ms, subsampling_factor);
-      r.text = ApplyInverseTextNormalization(std::move(r.text));
-      ss[i]->SetResult(r);
-    }
-  }
-
-  OfflineRecognizerConfig GetConfig() const override { return config_; }
-
- private:
-  void DecodeOneStream(OfflineStream *s) const {
-    const auto &meta_data = model_->GetModelMetadata();
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    int32_t feat_dim = config_.feat_config.feature_dim * meta_data.window_size;
-    std::vector<float> f = s->GetFrames();
-    f = ApplyLFR(f);
-    ApplyCMVN(&f);
-    int32_t num_frames = f.size() / feat_dim;
-    std::array<int64_t, 3> shape = {1, num_frames, feat_dim};
-    Ort::Value x = Ort::Value::CreateTensor(memory_info, f.data(), f.size(),
-                                            shape.data(), shape.size());
-
-    int64_t scale_shape = 1;
-
-    Ort::Value x_length =
-        Ort::Value::CreateTensor(memory_info, &num_frames, 1, &scale_shape, 1);
-
-    int32_t language = 0;
-    if (config_.model_config.sense_voice.language.empty()) {
-      language = 0;
-    } else if (meta_data.lang2id.count(
-                   config_.model_config.sense_voice.language)) {
-      language =
-          meta_data.lang2id.at(config_.model_config.sense_voice.language);
-    } else {
-      SHERPA_ONNX_LOGE("Unknown language: %s. Use 0 instead.",
-                       config_.model_config.sense_voice.language.c_str());
-    }
-
-    int32_t text_norm = config_.model_config.sense_voice.use_itn
-                            ? meta_data.with_itn_id
-                            : meta_data.without_itn_id;
-
-    Ort::Value language_tensor =
-        Ort::Value::CreateTensor(memory_info, &language, 1, &scale_shape, 1);
-
-    Ort::Value text_norm_tensor =
-        Ort::Value::CreateTensor(memory_info, &text_norm, 1, &scale_shape, 1);
-
-    Ort::Value logits{nullptr};
-    try {
-      logits = model_->Forward(std::move(x), std::move(x_length),
-                               std::move(language_tensor),
-                               std::move(text_norm_tensor));
-    } catch (const Ort::Exception &ex) {
-      SHERPA_ONNX_LOGE("\n\nCaught exception:\n\n%s\n\nReturn an empty result",
-                       ex.what());
-      return;
-    }
-
-    int64_t new_num_frames = num_frames + 4;
-    Ort::Value logits_length = Ort::Value::CreateTensor(
-        memory_info, &new_num_frames, 1, &scale_shape, 1);
-
-    auto results =
-        decoder_->Decode(std::move(logits), std::move(logits_length));
-
-    int32_t frame_shift_ms = 10;
-    int32_t subsampling_factor = meta_data.window_shift;
-    auto r = ConvertSenseVoiceResult(results[0], symbol_table_, frame_shift_ms,
-                                     subsampling_factor);
-
-    r.text = ApplyInverseTextNormalization(std::move(r.text));
-    s->SetResult(r);
-  }
-
-  void InitFeatConfig() {
-    const auto &meta_data = model_->GetModelMetadata();
-
-    config_.feat_config.normalize_samples = meta_data.normalize_samples;
-    config_.feat_config.window_type = "hamming";
-    config_.feat_config.high_freq = 0;
-    config_.feat_config.snip_edges = true;
-  }
-  std::vector<float> ApplyLFR(const std::vector<float> &in) const {
-    const auto &meta_data = model_->GetModelMetadata();
-
-    int32_t lfr_window_size = meta_data.window_size;
-    int32_t lfr_window_shift = meta_data.window_shift;
-    int32_t in_feat_dim = config_.feat_config.feature_dim;
-
-    int32_t in_num_frames = in.size() / in_feat_dim;
-    int32_t out_num_frames =
-        (in_num_frames - lfr_window_size) / lfr_window_shift + 1;
-    int32_t out_feat_dim = in_feat_dim * lfr_window_size;
-
-    std::vector<float> out(out_num_frames * out_feat_dim);
-
-    const float *p_in = in.data();
-    float *p_out = out.data();
-
-    for (int32_t i = 0; i != out_num_frames; ++i) {
-      std::copy(p_in, p_in + out_feat_dim, p_out);
-
-      p_out += out_feat_dim;
-      p_in += lfr_window_shift * in_feat_dim;
-    }
-
-    return out;
-  }
-
-  void ApplyCMVN(std::vector<float> *v) const {
-    const auto &meta_data = model_->GetModelMetadata();
-
-    const std::vector<float> &neg_mean = meta_data.neg_mean;
-    const std::vector<float> &inv_stddev = meta_data.inv_stddev;
-
-    int32_t dim = neg_mean.size();
-    int32_t num_frames = v->size() / dim;
-
-    float *p = v->data();
-
-    for (int32_t i = 0; i != num_frames; ++i) {
-      for (int32_t k = 0; k != dim; ++k) {
-        p[k] = (p[k] + neg_mean[k]) * inv_stddev[k];
-      }
-
-      p += dim;
-    }
-  }
-
-  OfflineRecognizerConfig config_;
-  SymbolTable symbol_table_;
-  std::unique_ptr<OfflineSenseVoiceModel> model_;
-  std::unique_ptr<OfflineCtcDecoder> decoder_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_SENSE_VOICE_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-transducer-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-transducer-impl.h
deleted file mode 100644
index 64f3798f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-transducer-impl.h
+++ /dev/null
@@ -1,304 +0,0 @@
-// sherpa-onnx/csrc/offline-recognizer-transducer-impl.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_TRANSDUCER_IMPL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_TRANSDUCER_IMPL_H_
-
-#include <fstream>
-#include <ios>
-#include <memory>
-#include <regex>  // NOLINT
-#include <sstream>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "sherpa-onnx/csrc/context-graph.h"
-#include "sherpa-onnx/csrc/log.h"
-#include "sherpa-onnx/csrc/macros.h"
-#include "sherpa-onnx/csrc/offline-recognizer-impl.h"
-#include "sherpa-onnx/csrc/offline-recognizer.h"
-#include "sherpa-onnx/csrc/offline-transducer-decoder.h"
-#include "sherpa-onnx/csrc/offline-transducer-greedy-search-decoder.h"
-#include "sherpa-onnx/csrc/offline-transducer-model.h"
-#include "sherpa-onnx/csrc/offline-transducer-modified-beam-search-decoder.h"
-#include "sherpa-onnx/csrc/pad-sequence.h"
-#include "sherpa-onnx/csrc/symbol-table.h"
-#include "sherpa-onnx/csrc/utils.h"
-#include "ssentencepiece/csrc/ssentencepiece.h"
-
-namespace sherpa_onnx {
-
-static OfflineRecognitionResult Convert(
-    const OfflineTransducerDecoderResult &src, const SymbolTable &sym_table,
-    int32_t frame_shift_ms, int32_t subsampling_factor) {
-  OfflineRecognitionResult r;
-  r.tokens.reserve(src.tokens.size());
-  r.timestamps.reserve(src.timestamps.size());
-
-  std::string text;
-  for (auto i : src.tokens) {
-    auto sym = sym_table[i];
-    text.append(sym);
-
-    if (sym.size() == 1 && (sym[0] < 0x20 || sym[0] > 0x7e)) {
-      // for byte bpe models,
-      // (but don't rewrite printable characters 0x20..0x7e,
-      //  which collide with standard BPE units)
-      std::ostringstream os;
-      os << "<0x" << std::hex << std::uppercase
-         << (static_cast<int32_t>(sym[0]) & 0xff) << ">";
-      sym = os.str();
-    }
-
-    r.tokens.push_back(std::move(sym));
-  }
-  r.text = std::move(text);
-
-  float frame_shift_s = frame_shift_ms / 1000. * subsampling_factor;
-  for (auto t : src.timestamps) {
-    float time = frame_shift_s * t;
-    r.timestamps.push_back(time);
-  }
-
-  return r;
-}
-
-class OfflineRecognizerTransducerImpl : public OfflineRecognizerImpl {
- public:
-  explicit OfflineRecognizerTransducerImpl(
-      const OfflineRecognizerConfig &config)
-      : OfflineRecognizerImpl(config),
-        config_(config),
-        symbol_table_(config_.model_config.tokens),
-        model_(std::make_unique<OfflineTransducerModel>(config_.model_config)) {
-    if (symbol_table_.Contains("<unk>")) {
-      unk_id_ = symbol_table_["<unk>"];
-    }
-
-    if (config_.decoding_method == "greedy_search") {
-      decoder_ = std::make_unique<OfflineTransducerGreedySearchDecoder>(
-          model_.get(), unk_id_, config_.blank_penalty);
-    } else if (config_.decoding_method == "modified_beam_search") {
-      if (!config_.lm_config.model.empty()) {
-        lm_ = OfflineLM::Create(config.lm_config);
-      }
-
-      if (!config_.model_config.bpe_vocab.empty()) {
-        bpe_encoder_ = std::make_unique<ssentencepiece::Ssentencepiece>(
-            config_.model_config.bpe_vocab);
-      }
-
-      if (!config_.hotwords_file.empty()) {
-        InitHotwords();
-      }
-
-      decoder_ = std::make_unique<OfflineTransducerModifiedBeamSearchDecoder>(
-          model_.get(), lm_.get(), config_.max_active_paths,
-          config_.lm_config.scale, unk_id_, config_.blank_penalty);
-    } else {
-      SHERPA_ONNX_LOGE("Unsupported decoding method: %s",
-                       config_.decoding_method.c_str());
-      exit(-1);
-    }
-  }
-
-  template <typename Manager>
-  explicit OfflineRecognizerTransducerImpl(
-      Manager *mgr, const OfflineRecognizerConfig &config)
-      : OfflineRecognizerImpl(mgr, config),
-        config_(config),
-        symbol_table_(mgr, config_.model_config.tokens),
-        model_(std::make_unique<OfflineTransducerModel>(mgr,
-                                                        config_.model_config)) {
-    if (symbol_table_.Contains("<unk>")) {
-      unk_id_ = symbol_table_["<unk>"];
-    }
-
-    if (config_.decoding_method == "greedy_search") {
-      decoder_ = std::make_unique<OfflineTransducerGreedySearchDecoder>(
-          model_.get(), unk_id_, config_.blank_penalty);
-    } else if (config_.decoding_method == "modified_beam_search") {
-      if (!config_.lm_config.model.empty()) {
-        lm_ = OfflineLM::Create(mgr, config.lm_config);
-      }
-
-      if (!config_.model_config.bpe_vocab.empty()) {
-        auto buf = ReadFile(mgr, config_.model_config.bpe_vocab);
-        std::istringstream iss(std::string(buf.begin(), buf.end()));
-        bpe_encoder_ = std::make_unique<ssentencepiece::Ssentencepiece>(iss);
-      }
-
-      if (!config_.hotwords_file.empty()) {
-        InitHotwords(mgr);
-      }
-
-      decoder_ = std::make_unique<OfflineTransducerModifiedBeamSearchDecoder>(
-          model_.get(), lm_.get(), config_.max_active_paths,
-          config_.lm_config.scale, unk_id_, config_.blank_penalty);
-    } else {
-      SHERPA_ONNX_LOGE("Unsupported decoding method: %s",
-                       config_.decoding_method.c_str());
-      exit(-1);
-    }
-  }
-
-  std::unique_ptr<OfflineStream> CreateStream(
-      const std::string &hotwords) const override {
-    auto hws = std::regex_replace(hotwords, std::regex("/"), "\n");
-    std::istringstream is(hws);
-    std::vector<std::vector<int32_t>> current;
-    std::vector<float> current_scores;
-    if (!EncodeHotwords(is, config_.model_config.modeling_unit, symbol_table_,
-                        bpe_encoder_.get(), &current, &current_scores)) {
-      SHERPA_ONNX_LOGE("Encode hotwords failed, skipping, hotwords are : %s",
-                       hotwords.c_str());
-    }
-
-    int32_t num_default_hws = hotwords_.size();
-    int32_t num_hws = current.size();
-
-    current.insert(current.end(), hotwords_.begin(), hotwords_.end());
-
-    if (!current_scores.empty() && !boost_scores_.empty()) {
-      current_scores.insert(current_scores.end(), boost_scores_.begin(),
-                            boost_scores_.end());
-    } else if (!current_scores.empty() && boost_scores_.empty()) {
-      current_scores.insert(current_scores.end(), num_default_hws,
-                            config_.hotwords_score);
-    } else if (current_scores.empty() && !boost_scores_.empty()) {
-      current_scores.insert(current_scores.end(), num_hws,
-                            config_.hotwords_score);
-      current_scores.insert(current_scores.end(), boost_scores_.begin(),
-                            boost_scores_.end());
-    } else {
-      // Do nothing.
-    }
-
-    auto context_graph = std::make_shared<ContextGraph>(
-        current, config_.hotwords_score, current_scores);
-    return std::make_unique<OfflineStream>(config_.feat_config, context_graph);
-  }
-
-  std::unique_ptr<OfflineStream> CreateStream() const override {
-    return std::make_unique<OfflineStream>(config_.feat_config,
-                                           hotwords_graph_);
-  }
-
-  void DecodeStreams(OfflineStream **ss, int32_t n) const override {
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    int32_t feat_dim = ss[0]->FeatureDim();
-
-    std::vector<Ort::Value> features;
-
-    features.reserve(n);
-
-    std::vector<std::vector<float>> features_vec(n);
-    std::vector<int64_t> features_length_vec(n);
-    for (int32_t i = 0; i != n; ++i) {
-      auto f = ss[i]->GetFrames();
-      int32_t num_frames = f.size() / feat_dim;
-
-      features_length_vec[i] = num_frames;
-      features_vec[i] = std::move(f);
-
-      std::array<int64_t, 2> shape = {num_frames, feat_dim};
-
-      Ort::Value x = Ort::Value::CreateTensor(
-          memory_info, features_vec[i].data(), features_vec[i].size(),
-          shape.data(), shape.size());
-      features.push_back(std::move(x));
-    }
-
-    std::vector<const Ort::Value *> features_pointer(n);
-    for (int32_t i = 0; i != n; ++i) {
-      features_pointer[i] = &features[i];
-    }
-
-    std::array<int64_t, 1> features_length_shape = {n};
-    Ort::Value x_length = Ort::Value::CreateTensor(
-        memory_info, features_length_vec.data(), n,
-        features_length_shape.data(), features_length_shape.size());
-
-    Ort::Value x = PadSequence(model_->Allocator(), features_pointer,
-                               -23.025850929940457f);
-
-    auto t = model_->RunEncoder(std::move(x), std::move(x_length));
-    auto results =
-        decoder_->Decode(std::move(t.first), std::move(t.second), ss, n);
-
-    int32_t frame_shift_ms = 10;
-    for (int32_t i = 0; i != n; ++i) {
-      auto r = Convert(results[i], symbol_table_, frame_shift_ms,
-                       model_->SubsamplingFactor());
-      r.text = ApplyInverseTextNormalization(std::move(r.text));
-
-      ss[i]->SetResult(r);
-    }
-  }
-
-  OfflineRecognizerConfig GetConfig() const override { return config_; }
-
-  void InitHotwords() {
-    // each line in hotwords_file contains space-separated words
-
-    std::ifstream is(config_.hotwords_file);
-    if (!is) {
-      SHERPA_ONNX_LOGE("Open hotwords file failed: %s",
-                       config_.hotwords_file.c_str());
-      exit(-1);
-    }
-
-    if (!EncodeHotwords(is, config_.model_config.modeling_unit, symbol_table_,
-                        bpe_encoder_.get(), &hotwords_, &boost_scores_)) {
-      SHERPA_ONNX_LOGE(
-          "Failed to encode some hotwords, skip them already, see logs above "
-          "for details.");
-    }
-    hotwords_graph_ = std::make_shared<ContextGraph>(
-        hotwords_, config_.hotwords_score, boost_scores_);
-  }
-
-  template <typename Manager>
-  void InitHotwords(Manager *mgr) {
-    // each line in hotwords_file contains space-separated words
-
-    auto buf = ReadFile(mgr, config_.hotwords_file);
-
-    std::istringstream is(std::string(buf.begin(), buf.end()));
-
-    if (!is) {
-      SHERPA_ONNX_LOGE("Open hotwords file failed: %s",
-                       config_.hotwords_file.c_str());
-      exit(-1);
-    }
-
-    if (!EncodeHotwords(is, config_.model_config.modeling_unit, symbol_table_,
-                        bpe_encoder_.get(), &hotwords_, &boost_scores_)) {
-      SHERPA_ONNX_LOGE(
-          "Failed to encode some hotwords, skip them already, see logs above "
-          "for details.");
-    }
-    hotwords_graph_ = std::make_shared<ContextGraph>(
-        hotwords_, config_.hotwords_score, boost_scores_);
-  }
-
- private:
-  OfflineRecognizerConfig config_;
-  SymbolTable symbol_table_;
-  std::vector<std::vector<int32_t>> hotwords_;
-  std::vector<float> boost_scores_;
-  ContextGraphPtr hotwords_graph_;
-  std::unique_ptr<ssentencepiece::Ssentencepiece> bpe_encoder_;
-  std::unique_ptr<OfflineTransducerModel> model_;
-  std::unique_ptr<OfflineTransducerDecoder> decoder_;
-  std::unique_ptr<OfflineLM> lm_;
-  int32_t unk_id_ = -1;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_TRANSDUCER_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-transducer-nemo-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-transducer-nemo-impl.h
deleted file mode 100644
index 167d1021..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-transducer-nemo-impl.h
+++ /dev/null
@@ -1,189 +0,0 @@
-// sherpa-onnx/csrc/offline-recognizer-transducer-nemo-impl.h
-//
-// Copyright (c)  2022-2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_TRANSDUCER_NEMO_IMPL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_TRANSDUCER_NEMO_IMPL_H_
-
-#include <fstream>
-#include <ios>
-#include <memory>
-#include <regex>  // NOLINT
-#include <sstream>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "sherpa-onnx/csrc/macros.h"
-#include "sherpa-onnx/csrc/offline-recognizer-impl.h"
-#include "sherpa-onnx/csrc/offline-recognizer.h"
-#include "sherpa-onnx/csrc/offline-transducer-greedy-search-nemo-decoder.h"
-#include "sherpa-onnx/csrc/offline-transducer-nemo-model.h"
-#include "sherpa-onnx/csrc/pad-sequence.h"
-#include "sherpa-onnx/csrc/symbol-table.h"
-#include "sherpa-onnx/csrc/transpose.h"
-#include "sherpa-onnx/csrc/utils.h"
-
-namespace sherpa_onnx {
-
-// defined in ./offline-recognizer-transducer-impl.h
-OfflineRecognitionResult Convert(const OfflineTransducerDecoderResult &src,
-                                 const SymbolTable &sym_table,
-                                 int32_t frame_shift_ms,
-                                 int32_t subsampling_factor);
-
-class OfflineRecognizerTransducerNeMoImpl : public OfflineRecognizerImpl {
- public:
-  explicit OfflineRecognizerTransducerNeMoImpl(
-      const OfflineRecognizerConfig &config)
-      : OfflineRecognizerImpl(config),
-        config_(config),
-        symbol_table_(config_.model_config.tokens),
-        model_(std::make_unique<OfflineTransducerNeMoModel>(
-            config_.model_config)) {
-    if (config_.decoding_method == "greedy_search") {
-      decoder_ = std::make_unique<OfflineTransducerGreedySearchNeMoDecoder>(
-          model_.get(), config_.blank_penalty);
-    } else {
-      SHERPA_ONNX_LOGE("Unsupported decoding method: %s",
-                       config_.decoding_method.c_str());
-      exit(-1);
-    }
-    PostInit();
-  }
-
-  template <typename Manager>
-  explicit OfflineRecognizerTransducerNeMoImpl(
-      Manager *mgr, const OfflineRecognizerConfig &config)
-      : OfflineRecognizerImpl(mgr, config),
-        config_(config),
-        symbol_table_(mgr, config_.model_config.tokens),
-        model_(std::make_unique<OfflineTransducerNeMoModel>(
-            mgr, config_.model_config)) {
-    if (config_.decoding_method == "greedy_search") {
-      decoder_ = std::make_unique<OfflineTransducerGreedySearchNeMoDecoder>(
-          model_.get(), config_.blank_penalty);
-    } else {
-      SHERPA_ONNX_LOGE("Unsupported decoding method: %s",
-                       config_.decoding_method.c_str());
-      exit(-1);
-    }
-
-    PostInit();
-  }
-
-  std::unique_ptr<OfflineStream> CreateStream() const override {
-    return std::make_unique<OfflineStream>(config_.feat_config);
-  }
-
-  void DecodeStreams(OfflineStream **ss, int32_t n) const override {
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    int32_t feat_dim = ss[0]->FeatureDim();
-
-    std::vector<Ort::Value> features;
-
-    features.reserve(n);
-
-    std::vector<std::vector<float>> features_vec(n);
-    std::vector<int64_t> features_length_vec(n);
-    for (int32_t i = 0; i != n; ++i) {
-      auto f = ss[i]->GetFrames();
-      int32_t num_frames = f.size() / feat_dim;
-
-      features_length_vec[i] = num_frames;
-      features_vec[i] = std::move(f);
-
-      std::array<int64_t, 2> shape = {num_frames, feat_dim};
-
-      Ort::Value x = Ort::Value::CreateTensor(
-          memory_info, features_vec[i].data(), features_vec[i].size(),
-          shape.data(), shape.size());
-      features.push_back(std::move(x));
-    }
-
-    std::vector<const Ort::Value *> features_pointer(n);
-    for (int32_t i = 0; i != n; ++i) {
-      features_pointer[i] = &features[i];
-    }
-
-    std::array<int64_t, 1> features_length_shape = {n};
-    Ort::Value x_length = Ort::Value::CreateTensor(
-        memory_info, features_length_vec.data(), n,
-        features_length_shape.data(), features_length_shape.size());
-
-    Ort::Value x = PadSequence(model_->Allocator(), features_pointer, 0);
-
-    auto t = model_->RunEncoder(std::move(x), std::move(x_length));
-    // t[0] encoder_out, float tensor, (batch_size, dim, T)
-    // t[1] encoder_out_length, int64 tensor, (batch_size,)
-
-    Ort::Value encoder_out = Transpose12(model_->Allocator(), &t[0]);
-
-    auto results = decoder_->Decode(std::move(encoder_out), std::move(t[1]));
-
-    int32_t frame_shift_ms = 10;
-    for (int32_t i = 0; i != n; ++i) {
-      auto r = Convert(results[i], symbol_table_, frame_shift_ms,
-                       model_->SubsamplingFactor());
-      r.text = ApplyInverseTextNormalization(std::move(r.text));
-
-      ss[i]->SetResult(r);
-    }
-  }
-
-  OfflineRecognizerConfig GetConfig() const override { return config_; }
-
- private:
-  void PostInit() {
-    config_.feat_config.nemo_normalize_type =
-        model_->FeatureNormalizationMethod();
-
-    config_.feat_config.dither = 0;
-
-    if (model_->IsGigaAM()) {
-      config_.feat_config.low_freq = 0;
-      config_.feat_config.high_freq = 8000;
-      config_.feat_config.remove_dc_offset = false;
-      config_.feat_config.preemph_coeff = 0;
-      config_.feat_config.window_type = "hann";
-      config_.feat_config.feature_dim = 64;
-    } else {
-      config_.feat_config.low_freq = 0;
-      // config_.feat_config.high_freq = 8000;
-      config_.feat_config.is_librosa = true;
-      config_.feat_config.remove_dc_offset = false;
-      // config_.feat_config.window_type = "hann";
-    }
-
-    int32_t vocab_size = model_->VocabSize();
-
-    // check the blank ID
-    if (!symbol_table_.Contains("<blk>")) {
-      SHERPA_ONNX_LOGE("tokens.txt does not include the blank token <blk>");
-      exit(-1);
-    }
-
-    if (symbol_table_["<blk>"] != vocab_size - 1) {
-      SHERPA_ONNX_LOGE("<blk> is not the last token!");
-      exit(-1);
-    }
-
-    if (symbol_table_.NumSymbols() != vocab_size) {
-      SHERPA_ONNX_LOGE("number of lines in tokens.txt %d != %d (vocab_size)",
-                       symbol_table_.NumSymbols(), vocab_size);
-      exit(-1);
-    }
-  }
-
- private:
-  OfflineRecognizerConfig config_;
-  SymbolTable symbol_table_;
-  std::unique_ptr<OfflineTransducerNeMoModel> model_;
-  std::unique_ptr<OfflineTransducerDecoder> decoder_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_TRANSDUCER_NEMO_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-whisper-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-whisper-impl.h
deleted file mode 100644
index fc6cc74e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer-whisper-impl.h
+++ /dev/null
@@ -1,181 +0,0 @@
-// sherpa-onnx/csrc/offline-recognizer-whisper-impl.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_WHISPER_IMPL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_WHISPER_IMPL_H_
-
-#include <algorithm>
-#include <cmath>
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-model-config.h"
-#include "sherpa-onnx/csrc/offline-recognizer-impl.h"
-#include "sherpa-onnx/csrc/offline-recognizer.h"
-#include "sherpa-onnx/csrc/offline-whisper-decoder.h"
-#include "sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.h"
-#include "sherpa-onnx/csrc/offline-whisper-model.h"
-#include "sherpa-onnx/csrc/symbol-table.h"
-#include "sherpa-onnx/csrc/transpose.h"
-
-namespace sherpa_onnx {
-
-static OfflineRecognitionResult Convert(const OfflineWhisperDecoderResult &src,
-                                        const SymbolTable &sym_table) {
-  OfflineRecognitionResult r;
-  r.tokens.reserve(src.tokens.size());
-
-  std::string text;
-  for (auto i : src.tokens) {
-    if (!sym_table.Contains(i)) {
-      continue;
-    }
-
-    const auto &s = sym_table[i];
-    text += s;
-    r.tokens.push_back(s);
-  }
-
-  r.text = text;
-  r.lang = src.lang;
-
-  return r;
-}
-
-class OfflineRecognizerWhisperImpl : public OfflineRecognizerImpl {
- public:
-  explicit OfflineRecognizerWhisperImpl(const OfflineRecognizerConfig &config)
-      : OfflineRecognizerImpl(config),
-        config_(config),
-        symbol_table_(config_.model_config.tokens),
-        model_(std::make_unique<OfflineWhisperModel>(config.model_config)) {
-    Init();
-  }
-
-  template <typename Manager>
-  OfflineRecognizerWhisperImpl(Manager *mgr,
-                               const OfflineRecognizerConfig &config)
-      : OfflineRecognizerImpl(mgr, config),
-        config_(config),
-        symbol_table_(mgr, config_.model_config.tokens),
-        model_(
-            std::make_unique<OfflineWhisperModel>(mgr, config.model_config)) {
-    Init();
-  }
-
-  void Init() {
-    // tokens.txt from whisper is base64 encoded, so we need to decode it
-    symbol_table_.ApplyBase64Decode();
-
-    if (config_.decoding_method == "greedy_search") {
-      decoder_ = std::make_unique<OfflineWhisperGreedySearchDecoder>(
-          config_.model_config.whisper, model_.get());
-    } else {
-      SHERPA_ONNX_LOGE(
-          "Only greedy_search is supported at present for whisper. Given %s",
-          config_.decoding_method.c_str());
-      exit(-1);
-    }
-  }
-
-  std::unique_ptr<OfflineStream> CreateStream() const override {
-    WhisperTag tag;
-    tag.dim = model_->FeatureDim();
-    return std::make_unique<OfflineStream>(tag);
-  }
-
-  void DecodeStreams(OfflineStream **ss, int32_t n) const override {
-    // batch decoding is not implemented yet
-    for (int32_t i = 0; i != n; ++i) {
-      DecodeStream(ss[i]);
-    }
-  }
-
-  void SetConfig(const OfflineRecognizerConfig &config) override {
-    config_.model_config.whisper = config.model_config.whisper;
-  }
-
-  OfflineRecognizerConfig GetConfig() const override { return config_; }
-
- private:
-  void DecodeStream(OfflineStream *s) const {
-    decoder_->SetConfig(config_.model_config.whisper);
-
-    int32_t max_num_frames = 3000;
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    int32_t feat_dim = s->FeatureDim();
-    std::vector<float> f = s->GetFrames();
-    int32_t num_frames = f.size() / feat_dim;
-
-    // we use 50 here so that there will be some zero tail paddings
-    if (num_frames >= max_num_frames - 50) {
-      SHERPA_ONNX_LOGE(
-          "Only waves less than 30 seconds are supported. We process only the "
-          "first 30 seconds and discard the remaining data");
-      num_frames = max_num_frames - 50;
-    }
-
-    model_->NormalizeFeatures(f.data(), num_frames, feat_dim);
-
-    // note that 1000 is an experience-value.
-    // You can replace 1000 by other values, say, 100.
-    //
-    // Since we have removed the 30 seconds constraint, we need
-    // tail_padding_frames so that whisper is able to detect the eot token.
-    int32_t tail_padding_frames = 1000;
-
-    if (config_.model_config.whisper.tail_paddings > 0) {
-      tail_padding_frames = config_.model_config.whisper.tail_paddings;
-    }
-
-    int32_t actual_frames =
-        std::min(num_frames + tail_padding_frames, max_num_frames);
-
-    std::array<int64_t, 3> shape{1, actual_frames, feat_dim};
-
-    Ort::Value mel = Ort::Value::CreateTensor<float>(
-        model_->Allocator(), shape.data(), shape.size());
-
-    float *p_mel = mel.GetTensorMutableData<float>();
-    std::copy(f.data(), f.data() + num_frames * feat_dim, p_mel);
-
-    std::fill_n(p_mel + num_frames * feat_dim,
-                (actual_frames - num_frames) * feat_dim, 0);
-
-    mel = Transpose12(model_->Allocator(), &mel);
-
-    try {
-      auto cross_kv = model_->ForwardEncoder(std::move(mel));
-
-      auto results = decoder_->Decode(std::move(cross_kv.first),
-                                      std::move(cross_kv.second));
-
-      auto r = Convert(results[0], symbol_table_);
-      r.text = ApplyInverseTextNormalization(std::move(r.text));
-      s->SetResult(r);
-    } catch (const Ort::Exception &ex) {
-      SHERPA_ONNX_LOGE(
-          "\n\nCaught exception:\n\n%s\n\nReturn an empty result. Number of "
-          "input frames: %d, Current tail "
-          "paddings: %d. If you see a lot of such exceptions, please consider "
-          "using a larger --whisper-tail-paddings",
-          ex.what(), num_frames, tail_padding_frames);
-      return;
-    }
-  }
-
- private:
-  OfflineRecognizerConfig config_;
-  SymbolTable symbol_table_;
-  std::unique_ptr<OfflineWhisperModel> model_;
-  std::unique_ptr<OfflineWhisperDecoder> decoder_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_WHISPER_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer.h
deleted file mode 100644
index 3c78ea9b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-recognizer.h
+++ /dev/null
@@ -1,131 +0,0 @@
-// sherpa-onnx/csrc/offline-recognizer.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "sherpa-onnx/csrc/features.h"
-#include "sherpa-onnx/csrc/offline-ctc-fst-decoder-config.h"
-#include "sherpa-onnx/csrc/offline-lm-config.h"
-#include "sherpa-onnx/csrc/offline-model-config.h"
-#include "sherpa-onnx/csrc/offline-stream.h"
-#include "sherpa-onnx/csrc/offline-transducer-model-config.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineRecognitionResult;
-
-struct OfflineRecognizerConfig {
-  FeatureExtractorConfig feat_config;
-  OfflineModelConfig model_config;
-  OfflineLMConfig lm_config;
-  OfflineCtcFstDecoderConfig ctc_fst_decoder_config;
-
-  std::string decoding_method = "greedy_search";
-  int32_t max_active_paths = 4;
-
-  std::string hotwords_file;
-  float hotwords_score = 1.5;
-
-  float blank_penalty = 0.0;
-
-  // If there are multiple rules, they are applied from left to right.
-  std::string rule_fsts;
-
-  // If there are multiple FST archives, they are applied from left to right.
-  std::string rule_fars;
-
-  // only greedy_search is implemented
-  // TODO(fangjun): Implement modified_beam_search
-
-  OfflineRecognizerConfig() = default;
-  OfflineRecognizerConfig(
-      const FeatureExtractorConfig &feat_config,
-      const OfflineModelConfig &model_config, const OfflineLMConfig &lm_config,
-      const OfflineCtcFstDecoderConfig &ctc_fst_decoder_config,
-      const std::string &decoding_method, int32_t max_active_paths,
-      const std::string &hotwords_file, float hotwords_score,
-      float blank_penalty, const std::string &rule_fsts,
-      const std::string &rule_fars)
-      : feat_config(feat_config),
-        model_config(model_config),
-        lm_config(lm_config),
-        ctc_fst_decoder_config(ctc_fst_decoder_config),
-        decoding_method(decoding_method),
-        max_active_paths(max_active_paths),
-        hotwords_file(hotwords_file),
-        hotwords_score(hotwords_score),
-        blank_penalty(blank_penalty),
-        rule_fsts(rule_fsts),
-        rule_fars(rule_fars) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-class OfflineRecognizerImpl;
-
-class OfflineRecognizer {
- public:
-  ~OfflineRecognizer();
-
-  template <typename Manager>
-  OfflineRecognizer(Manager *mgr, const OfflineRecognizerConfig &config);
-
-  explicit OfflineRecognizer(const OfflineRecognizerConfig &config);
-
-  /// Create a stream for decoding.
-  std::unique_ptr<OfflineStream> CreateStream() const;
-
-  /** Create a stream for decoding.
-   *
-   *  @param The hotwords for this string, it might contain several hotwords,
-   *         the hotwords are separated by "/". In each of the hotwords, there
-   *         are cjkchars or bpes, the bpe/cjkchar are separated by space (" ").
-   *         For example, hotwords I LOVE YOU and HELLO WORLD, looks like:
-   *
-   *         "▁I ▁LOVE ▁YOU/▁HE LL O ▁WORLD"
-   */
-  std::unique_ptr<OfflineStream> CreateStream(
-      const std::string &hotwords) const;
-
-  /** Decode a single stream
-   *
-   * @param s The stream to decode.
-   */
-  void DecodeStream(OfflineStream *s) const {
-    OfflineStream *ss[1] = {s};
-    DecodeStreams(ss, 1);
-  }
-
-  /** Decode a list of streams.
-   *
-   * @param ss Pointer to an array of streams.
-   * @param n  Size of the input array.
-   */
-  void DecodeStreams(OfflineStream **ss, int32_t n) const;
-
-  /** Onnxruntime Session objects are not affected by this method.
-   * The exact behavior can be defined by a specific recognizer impl.
-   * For instance, for the whisper recognizer, you can retrieve the language and
-   * task from the config and ignore any remaining fields in `config`.
-   */
-  void SetConfig(const OfflineRecognizerConfig &config);
-
-  OfflineRecognizerConfig GetConfig() const;
-
- private:
-  std::unique_ptr<OfflineRecognizerImpl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-rnn-lm.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-rnn-lm.h
deleted file mode 100644
index 2d16a19e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-rnn-lm.h
+++ /dev/null
@@ -1,44 +0,0 @@
-// sherpa-onnx/csrc/offline-rnn-lm.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_RNN_LM_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_RNN_LM_H_
-
-#include <memory>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-lm-config.h"
-#include "sherpa-onnx/csrc/offline-lm.h"
-
-namespace sherpa_onnx {
-
-class OfflineRnnLM : public OfflineLM {
- public:
-  ~OfflineRnnLM() override;
-
-  explicit OfflineRnnLM(const OfflineLMConfig &config);
-
-  template <typename Manager>
-  OfflineRnnLM(Manager *mgr, const OfflineLMConfig &config);
-
-  /** Rescore a batch of sentences.
-   *
-   * @param x A 2-D tensor of shape (N, L) with data type int64.
-   * @param x_lens A 1-D tensor of shape (N,) with data type int64.
-   *               It contains number of valid tokens in x before padding.
-   * @return Return a 1-D tensor of shape (N,) containing the log likelihood
-   *         of each utterance. Its data type is float32.
-   *
-   * Caution: It returns log likelihood, not negative log likelihood (nll).
-   */
-  Ort::Value Rescore(Ort::Value x, Ort::Value x_lens) override;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_RNN_LM_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-sense-voice-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-sense-voice-model-config.h
deleted file mode 100644
index 2f724e44..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-sense-voice-model-config.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// sherpa-onnx/csrc/offline-sense-voice-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_SENSE_VOICE_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_SENSE_VOICE_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineSenseVoiceModelConfig {
-  std::string model;
-
-  // "" or "auto" to let the model recognize the language
-  // valid values:
-  //  zh, en, ja, ko, yue, auto
-  std::string language = "auto";
-
-  // true to use inverse text normalization
-  // false to not use inverse text normalization
-  bool use_itn = false;
-
-  OfflineSenseVoiceModelConfig() = default;
-  explicit OfflineSenseVoiceModelConfig(const std::string &model,
-                                        const std::string &language,
-                                        bool use_itn)
-      : model(model), language(language), use_itn(use_itn) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_SENSE_VOICE_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-sense-voice-model-meta-data.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-sense-voice-model-meta-data.h
deleted file mode 100644
index 6065c93c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-sense-voice-model-meta-data.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// sherpa-onnx/csrc/offline-sense-voice-model-meta-data.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_SENSE_VOICE_MODEL_META_DATA_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_SENSE_VOICE_MODEL_META_DATA_H_
-
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-namespace sherpa_onnx {
-
-struct OfflineSenseVoiceModelMetaData {
-  // ID for using inverse text normalization
-  int32_t with_itn_id;
-
-  // ID for not using inverse text normalization
-  int32_t without_itn_id;
-
-  int32_t window_size;   // lfr_m
-  int32_t window_shift;  // lfr_n
-  int32_t vocab_size;
-
-  int32_t subsampling_factor = 1;
-
-  // Usually 0 for SenseVoice models.
-  // 0 means samples are scaled to [-32768, 32767] before are sent to the
-  // feature extractor
-  int32_t normalize_samples = 0;
-
-  int32_t blank_id = 0;
-
-  // possible values:
-  // zh, en, ja, ko, yue, auto
-  // where
-  //  zh is Chinese (Mandarin)
-  //  en is English
-  //  ja is Japanese
-  //  ko is Korean
-  //  yue is Cantonese
-  //  auto is to let the model recognize the language
-  std::unordered_map<std::string, int32_t> lang2id;
-
-  std::vector<float> neg_mean;
-  std::vector<float> inv_stddev;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_SENSE_VOICE_MODEL_META_DATA_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-sense-voice-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-sense-voice-model.h
deleted file mode 100644
index e82680c5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-sense-voice-model.h
+++ /dev/null
@@ -1,55 +0,0 @@
-// sherpa-onnx/csrc/offline-sense-voice-model.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_SENSE_VOICE_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_SENSE_VOICE_MODEL_H_
-
-#include <memory>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-model-config.h"
-#include "sherpa-onnx/csrc/offline-sense-voice-model-meta-data.h"
-
-namespace sherpa_onnx {
-
-class OfflineSenseVoiceModel {
- public:
-  explicit OfflineSenseVoiceModel(const OfflineModelConfig &config);
-
-  template <typename Manager>
-  OfflineSenseVoiceModel(Manager *mgr, const OfflineModelConfig &config);
-
-  ~OfflineSenseVoiceModel();
-
-  /** Run the forward method of the model.
-   *
-   * @param features  A tensor of shape (N, T, C). It is changed in-place.
-   * @param features_length  A 1-D tensor of shape (N,) containing number of
-   *                         valid frames in `features` before padding.
-   *                         Its dtype is int32_t.
-   * @param language A 1-D tensor of shape (N,) with dtype int32_t
-   * @param text_norm A 1-D tensor of shape (N,) with dtype int32_t
-   *
-   * @return Return logits of shape (N, T, C) with dtype float
-   *
-   * Note: The subsampling factor is 1 for SenseVoice, so there is
-   *       no need to output logits_length.
-   */
-  Ort::Value Forward(Ort::Value features, Ort::Value features_length,
-                     Ort::Value language, Ort::Value text_norm) const;
-
-  const OfflineSenseVoiceModelMetaData &GetModelMetadata() const;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_SENSE_VOICE_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization-impl.h
deleted file mode 100644
index d2cbdebd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization-impl.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// sherpa-onnx/csrc/offline-speaker-diarization-impl.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_DIARIZATION_IMPL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_DIARIZATION_IMPL_H_
-
-#include <functional>
-#include <memory>
-
-#include "sherpa-onnx/csrc/offline-speaker-diarization.h"
-namespace sherpa_onnx {
-
-class OfflineSpeakerDiarizationImpl {
- public:
-  static std::unique_ptr<OfflineSpeakerDiarizationImpl> Create(
-      const OfflineSpeakerDiarizationConfig &config);
-
-  template <typename Manager>
-  static std::unique_ptr<OfflineSpeakerDiarizationImpl> Create(
-      Manager *mgr, const OfflineSpeakerDiarizationConfig &config);
-
-  virtual ~OfflineSpeakerDiarizationImpl() = default;
-
-  virtual int32_t SampleRate() const = 0;
-
-  // Note: Only config.clustering is used. All other fields in config are
-  // ignored
-  virtual void SetConfig(const OfflineSpeakerDiarizationConfig &config) = 0;
-
-  virtual OfflineSpeakerDiarizationResult Process(
-      const float *audio, int32_t n,
-      OfflineSpeakerDiarizationProgressCallback callback = nullptr,
-      void *callback_arg = nullptr) const = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_DIARIZATION_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization-pyannote-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization-pyannote-impl.h
deleted file mode 100644
index e8228d47..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization-pyannote-impl.h
+++ /dev/null
@@ -1,743 +0,0 @@
-// sherpa-onnx/csrc/offline-speaker-diarization-pyannote-impl.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_DIARIZATION_PYANNOTE_IMPL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_DIARIZATION_PYANNOTE_IMPL_H_
-
-#include <algorithm>
-#include <cmath>
-#include <memory>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include "Eigen/Dense"
-#include "sherpa-onnx/csrc/fast-clustering.h"
-#include "sherpa-onnx/csrc/math.h"
-#include "sherpa-onnx/csrc/offline-speaker-diarization-impl.h"
-#include "sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model.h"
-#include "sherpa-onnx/csrc/speaker-embedding-extractor.h"
-
-namespace sherpa_onnx {
-
-namespace {  // NOLINT
-
-// copied from https://github.com/k2-fsa/k2/blob/master/k2/csrc/host/util.h#L41
-template <class T>
-inline void hash_combine(std::size_t *seed, const T &v) {  // NOLINT
-  std::hash<T> hasher;
-  *seed ^= hasher(v) + 0x9e3779b9 + ((*seed) << 6) + ((*seed) >> 2);  // NOLINT
-}
-
-// copied from https://github.com/k2-fsa/k2/blob/master/k2/csrc/host/util.h#L47
-struct PairHash {
-  template <class T1, class T2>
-  std::size_t operator()(const std::pair<T1, T2> &pair) const {
-    std::size_t result = 0;
-    hash_combine(&result, pair.first);
-    hash_combine(&result, pair.second);
-    return result;
-  }
-};
-}  // namespace
-
-using Matrix2D =
-    Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>;
-
-using Matrix2DInt32 =
-    Eigen::Matrix<int32_t, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>;
-
-using FloatRowVector = Eigen::Matrix<float, 1, Eigen::Dynamic>;
-using Int32RowVector = Eigen::Matrix<int32_t, 1, Eigen::Dynamic>;
-
-using Int32Pair = std::pair<int32_t, int32_t>;
-
-class OfflineSpeakerDiarizationPyannoteImpl
-    : public OfflineSpeakerDiarizationImpl {
- public:
-  ~OfflineSpeakerDiarizationPyannoteImpl() override = default;
-
-  explicit OfflineSpeakerDiarizationPyannoteImpl(
-      const OfflineSpeakerDiarizationConfig &config)
-      : config_(config),
-        segmentation_model_(config_.segmentation),
-        embedding_extractor_(config_.embedding),
-        clustering_(std::make_unique<FastClustering>(config_.clustering)) {
-    Init();
-  }
-
-  template <typename Manager>
-  OfflineSpeakerDiarizationPyannoteImpl(
-      Manager *mgr, const OfflineSpeakerDiarizationConfig &config)
-      : config_(config),
-        segmentation_model_(mgr, config_.segmentation),
-        embedding_extractor_(mgr, config_.embedding),
-        clustering_(std::make_unique<FastClustering>(config_.clustering)) {
-    Init();
-  }
-
-  int32_t SampleRate() const override {
-    const auto &meta_data = segmentation_model_.GetModelMetaData();
-
-    return meta_data.sample_rate;
-  }
-
-  void SetConfig(const OfflineSpeakerDiarizationConfig &config) override {
-    if (!config.clustering.Validate()) {
-      SHERPA_ONNX_LOGE("Invalid clustering config. Skip it");
-      return;
-    }
-    clustering_ = std::make_unique<FastClustering>(config.clustering);
-    config_.clustering = config.clustering;
-  }
-
-  OfflineSpeakerDiarizationResult Process(
-      const float *audio, int32_t n,
-      OfflineSpeakerDiarizationProgressCallback callback = nullptr,
-      void *callback_arg = nullptr) const override {
-    std::vector<Matrix2D> segmentations = RunSpeakerSegmentationModel(audio, n);
-    // segmentations[i] is for chunk_i
-    // Each matrix is of shape (num_frames, num_powerset_classes)
-    if (segmentations.empty()) {
-      return {};
-    }
-
-    std::vector<Matrix2DInt32> labels;
-    labels.reserve(segmentations.size());
-
-    for (const auto &m : segmentations) {
-      labels.push_back(ToMultiLabel(m));
-    }
-
-    segmentations.clear();
-
-    if (labels.size() == 1) {
-      if (callback) {
-        callback(1, 1, callback_arg);
-      }
-
-      return HandleOneChunkSpecialCase(labels[0], n);
-    }
-
-    // labels[i] is a 0-1 matrix of shape (num_frames, num_speakers)
-
-    // speaker count per frame
-    Int32RowVector speakers_per_frame = ComputeSpeakersPerFrame(labels);
-
-    if (speakers_per_frame.maxCoeff() == 0) {
-      SHERPA_ONNX_LOGE("No speakers found in the audio samples");
-      return {};
-    }
-
-    auto chunk_speaker_samples_list_pair = GetChunkSpeakerSampleIndexes(labels);
-
-    // The embedding model may output NaN. valid_indexes contains indexes
-    // in chunk_speaker_samples_list_pair.second that don't lead to
-    // NaN embeddings.
-    std::vector<int32_t> valid_indexes;
-    valid_indexes.reserve(chunk_speaker_samples_list_pair.second.size());
-
-    Matrix2D embeddings =
-        ComputeEmbeddings(audio, n, chunk_speaker_samples_list_pair.second,
-                          &valid_indexes, std::move(callback), callback_arg);
-
-    if (valid_indexes.size() != chunk_speaker_samples_list_pair.second.size()) {
-      std::vector<Int32Pair> chunk_speaker_pair;
-      std::vector<std::vector<Int32Pair>> sample_indexes;
-
-      chunk_speaker_pair.reserve(valid_indexes.size());
-      sample_indexes.reserve(valid_indexes.size());
-      for (auto i : valid_indexes) {
-        chunk_speaker_pair.push_back(chunk_speaker_samples_list_pair.first[i]);
-        sample_indexes.push_back(
-            std::move(chunk_speaker_samples_list_pair.second[i]));
-      }
-
-      chunk_speaker_samples_list_pair.first = std::move(chunk_speaker_pair);
-      chunk_speaker_samples_list_pair.second = std::move(sample_indexes);
-    }
-
-    std::vector<int32_t> cluster_labels = clustering_->Cluster(
-        &embeddings(0, 0), embeddings.rows(), embeddings.cols());
-
-    int32_t max_cluster_index =
-        *std::max_element(cluster_labels.begin(), cluster_labels.end());
-
-    auto chunk_speaker_to_cluster = ConvertChunkSpeakerToCluster(
-        chunk_speaker_samples_list_pair.first, cluster_labels);
-
-    auto new_labels =
-        ReLabel(labels, max_cluster_index, chunk_speaker_to_cluster);
-
-    Matrix2DInt32 speaker_count = ComputeSpeakerCount(new_labels, n);
-
-    Matrix2DInt32 final_labels =
-        FinalizeLabels(speaker_count, speakers_per_frame);
-
-    auto result = ComputeResult(final_labels);
-
-    return result;
-  }
-
- private:
-  void Init() { InitPowersetMapping(); }
-
-  // see also
-  // https://github.com/pyannote/pyannote-audio/blob/develop/pyannote/audio/utils/powerset.py#L68
-  void InitPowersetMapping() {
-    const auto &meta_data = segmentation_model_.GetModelMetaData();
-    int32_t num_classes = meta_data.num_classes;
-    int32_t powerset_max_classes = meta_data.powerset_max_classes;
-    int32_t num_speakers = meta_data.num_speakers;
-
-    powerset_mapping_ = Matrix2DInt32(num_classes, num_speakers);
-    powerset_mapping_.setZero();
-
-    int32_t k = 1;
-    for (int32_t i = 1; i <= powerset_max_classes; ++i) {
-      if (i == 1) {
-        for (int32_t j = 0; j != num_speakers; ++j, ++k) {
-          powerset_mapping_(k, j) = 1;
-        }
-      } else if (i == 2) {
-        for (int32_t j = 0; j != num_speakers; ++j) {
-          for (int32_t m = j + 1; m < num_speakers; ++m, ++k) {
-            powerset_mapping_(k, j) = 1;
-            powerset_mapping_(k, m) = 1;
-          }
-        }
-      } else {
-#if __OHOS__
-        SHERPA_ONNX_LOGE(
-            "powerset_max_classes = %{public}d is currently not supported!", i);
-#else
-        SHERPA_ONNX_LOGE(
-            "powerset_max_classes = %d is currently not supported!", i);
-#endif
-        SHERPA_ONNX_EXIT(-1);
-      }
-    }
-  }
-
-  std::vector<Matrix2D> RunSpeakerSegmentationModel(const float *audio,
-                                                    int32_t n) const {
-    std::vector<Matrix2D> ans;
-
-    const auto &meta_data = segmentation_model_.GetModelMetaData();
-    int32_t window_size = meta_data.window_size;
-    int32_t window_shift = meta_data.window_shift;
-
-    if (n <= 0) {
-#if __OHOS__
-      SHERPA_ONNX_LOGE(
-          "number of audio samples is %{public}d (<= 0). Please provide a "
-          "positive number",
-          n);
-#else
-      SHERPA_ONNX_LOGE(
-          "number of audio samples is %d (<= 0). Please provide a positive "
-          "number",
-          n);
-#endif
-      return {};
-    }
-
-    if (n <= window_size) {
-      std::vector<float> buf(window_size);
-      // NOTE: buf is zero initialized by default
-
-      std::copy(audio, audio + n, buf.data());
-
-      Matrix2D m = ProcessChunk(buf.data());
-
-      ans.push_back(std::move(m));
-
-      return ans;
-    }
-
-    int32_t num_chunks = (n - window_size) / window_shift + 1;
-    bool has_last_chunk = ((n - window_size) % window_shift) > 0;
-
-    ans.reserve(num_chunks + has_last_chunk);
-
-    const float *p = audio;
-
-    for (int32_t i = 0; i != num_chunks; ++i, p += window_shift) {
-      Matrix2D m = ProcessChunk(p);
-
-      ans.push_back(std::move(m));
-    }
-
-    if (has_last_chunk) {
-      std::vector<float> buf(window_size);
-      std::copy(p, audio + n, buf.data());
-
-      Matrix2D m = ProcessChunk(buf.data());
-
-      ans.push_back(std::move(m));
-    }
-
-    return ans;
-  }
-
-  Matrix2D ProcessChunk(const float *p) const {
-    const auto &meta_data = segmentation_model_.GetModelMetaData();
-    int32_t window_size = meta_data.window_size;
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    std::array<int64_t, 3> shape = {1, 1, window_size};
-
-    Ort::Value x =
-        Ort::Value::CreateTensor(memory_info, const_cast<float *>(p),
-                                 window_size, shape.data(), shape.size());
-
-    Ort::Value out = segmentation_model_.Forward(std::move(x));
-    std::vector<int64_t> out_shape = out.GetTensorTypeAndShapeInfo().GetShape();
-    Matrix2D m(out_shape[1], out_shape[2]);
-    std::copy(out.GetTensorData<float>(), out.GetTensorData<float>() + m.size(),
-              &m(0, 0));
-    return m;
-  }
-
-  Matrix2DInt32 ToMultiLabel(const Matrix2D &m) const {
-    int32_t num_rows = m.rows();
-    Matrix2DInt32 ans(num_rows, powerset_mapping_.cols());
-
-    std::ptrdiff_t col_id;
-
-    for (int32_t i = 0; i != num_rows; ++i) {
-      m.row(i).maxCoeff(&col_id);
-      ans.row(i) = powerset_mapping_.row(col_id);
-    }
-
-    return ans;
-  }
-
-  // See also
-  // https://github.com/pyannote/pyannote-audio/blob/develop/pyannote/audio/pipelines/utils/diarization.py#L122
-  Int32RowVector ComputeSpeakersPerFrame(
-      const std::vector<Matrix2DInt32> &labels) const {
-    const auto &meta_data = segmentation_model_.GetModelMetaData();
-    int32_t window_size = meta_data.window_size;
-    int32_t window_shift = meta_data.window_shift;
-    int32_t receptive_field_shift = meta_data.receptive_field_shift;
-
-    int32_t num_chunks = labels.size();
-
-    int32_t num_frames = (window_size + (num_chunks - 1) * window_shift) /
-                             receptive_field_shift +
-                         1;
-
-    FloatRowVector count(num_frames);
-    FloatRowVector weight(num_frames);
-    count.setZero();
-    weight.setZero();
-
-    for (int32_t i = 0; i != num_chunks; ++i) {
-      int32_t start =
-          static_cast<float>(i) * window_shift / receptive_field_shift + 0.5;
-
-      auto seq = Eigen::seqN(start, labels[i].rows());
-
-      count(seq).array() += labels[i].rowwise().sum().array().cast<float>();
-
-      weight(seq).array() += 1;
-    }
-
-    return ((count.array() / (weight.array() + 1e-12f)) + 0.5).cast<int32_t>();
-  }
-
-  // ans.first: a list of (chunk_id, speaker_id)
-  // ans.second: a list of list of (start_sample_index, end_sample_index)
-  //
-  // ans.first[i] corresponds to ans.second[i]
-  std::pair<std::vector<Int32Pair>, std::vector<std::vector<Int32Pair>>>
-  GetChunkSpeakerSampleIndexes(const std::vector<Matrix2DInt32> &labels) const {
-    auto new_labels = ExcludeOverlap(labels);
-
-    std::vector<Int32Pair> chunk_speaker_list;
-    std::vector<std::vector<Int32Pair>> samples_index_list;
-
-    const auto &meta_data = segmentation_model_.GetModelMetaData();
-    int32_t window_size = meta_data.window_size;
-    int32_t window_shift = meta_data.window_shift;
-    int32_t receptive_field_shift = meta_data.receptive_field_shift;
-    int32_t num_speakers = meta_data.num_speakers;
-
-    int32_t chunk_index = 0;
-    for (const auto &label : new_labels) {
-      Matrix2DInt32 tmp = label.transpose();
-      // tmp: (num_speakers, num_frames)
-      int32_t num_frames = tmp.cols();
-
-      int32_t sample_offset = chunk_index * window_shift;
-
-      for (int32_t speaker_index = 0; speaker_index != num_speakers;
-           ++speaker_index) {
-        auto d = tmp.row(speaker_index);
-        if (d.sum() < 10) {
-          // skip segments less than 10 frames
-          continue;
-        }
-
-        Int32Pair this_chunk_speaker = {chunk_index, speaker_index};
-        std::vector<Int32Pair> this_speaker_samples;
-
-        bool is_active = false;
-        int32_t start_index;
-
-        for (int32_t k = 0; k != num_frames; ++k) {
-          if (d[k] != 0) {
-            if (!is_active) {
-              is_active = true;
-              start_index = k;
-            }
-          } else if (is_active) {
-            is_active = false;
-
-            int32_t start_samples =
-                static_cast<float>(start_index) / num_frames * window_size +
-                sample_offset;
-            int32_t end_samples =
-                static_cast<float>(k) / num_frames * window_size +
-                sample_offset;
-
-            this_speaker_samples.emplace_back(start_samples, end_samples);
-          }
-        }
-
-        if (is_active) {
-          int32_t start_samples =
-              static_cast<float>(start_index) / num_frames * window_size +
-              sample_offset;
-          int32_t end_samples =
-              static_cast<float>(num_frames - 1) / num_frames * window_size +
-              sample_offset;
-          this_speaker_samples.emplace_back(start_samples, end_samples);
-        }
-
-        chunk_speaker_list.push_back(std::move(this_chunk_speaker));
-        samples_index_list.push_back(std::move(this_speaker_samples));
-      }  // for (int32_t speaker_index = 0;
-      chunk_index += 1;
-    }  // for (const auto &label : new_labels)
-
-    return {chunk_speaker_list, samples_index_list};
-  }
-
-  // If there are multiple speakers at a frame, then this frame is excluded.
-  std::vector<Matrix2DInt32> ExcludeOverlap(
-      const std::vector<Matrix2DInt32> &labels) const {
-    int32_t num_chunks = labels.size();
-    std::vector<Matrix2DInt32> ans;
-    ans.reserve(num_chunks);
-
-    for (const auto &label : labels) {
-      Matrix2DInt32 new_label(label.rows(), label.cols());
-      new_label.setZero();
-      Int32RowVector v = label.rowwise().sum();
-
-      for (int32_t i = 0; i != v.cols(); ++i) {
-        if (v[i] < 2) {
-          new_label.row(i) = label.row(i);
-        }
-      }
-
-      ans.push_back(std::move(new_label));
-    }
-
-    return ans;
-  }
-
-  /**
-   * @param sample_indexes[i] contains the sample segment start and end indexes
-   *                          for the i-th (chunk, speaker) pair
-   * @return Return a matrix of shape (sample_indexes.size(), embedding_dim)
-   *         where ans.row[i] contains the embedding for the
-   *         i-th (chunk, speaker) pair
-   */
-  Matrix2D ComputeEmbeddings(
-      const float *audio, int32_t n,
-      const std::vector<std::vector<Int32Pair>> &sample_indexes,
-      std::vector<int32_t> *valid_indexes,
-      OfflineSpeakerDiarizationProgressCallback callback,
-      void *callback_arg) const {
-    const auto &meta_data = segmentation_model_.GetModelMetaData();
-    int32_t sample_rate = meta_data.sample_rate;
-    Matrix2D ans(sample_indexes.size(), embedding_extractor_.Dim());
-
-    auto IsNaNWrapper = [](float f) -> bool { return std::isnan(f); };
-
-    int32_t k = 0;
-    int32_t cur_row_index = 0;
-    for (const auto &v : sample_indexes) {
-      auto stream = embedding_extractor_.CreateStream();
-      for (const auto &p : v) {
-        int32_t end = (p.second <= n) ? p.second : n;
-        int32_t num_samples = end - p.first;
-
-        if (num_samples > 0) {
-          stream->AcceptWaveform(sample_rate, audio + p.first, num_samples);
-        }
-      }
-
-      stream->InputFinished();
-      if (!embedding_extractor_.IsReady(stream.get())) {
-        SHERPA_ONNX_LOGE(
-            "This segment is too short, which should not happen since we have "
-            "already filtered short segments");
-        SHERPA_ONNX_EXIT(-1);
-      }
-
-      std::vector<float> embedding = embedding_extractor_.Compute(stream.get());
-
-      if (std::none_of(embedding.begin(), embedding.end(), IsNaNWrapper)) {
-        // a valid embedding
-        std::copy(embedding.begin(), embedding.end(), &ans(cur_row_index, 0));
-        cur_row_index += 1;
-        valid_indexes->push_back(k);
-      }
-
-      k += 1;
-
-      if (callback) {
-        callback(k, ans.rows(), callback_arg);
-      }
-    }
-
-    if (k != cur_row_index) {
-      auto seq = Eigen::seqN(0, cur_row_index);
-      ans = ans(seq, Eigen::all);
-    }
-
-    return ans;
-  }
-
-  std::unordered_map<Int32Pair, int32_t, PairHash> ConvertChunkSpeakerToCluster(
-      const std::vector<Int32Pair> &chunk_speaker_pair,
-      const std::vector<int32_t> &cluster_labels) const {
-    std::unordered_map<Int32Pair, int32_t, PairHash> ans;
-
-    int32_t k = 0;
-    for (const auto &p : chunk_speaker_pair) {
-      ans[p] = cluster_labels[k];
-      k += 1;
-    }
-
-    return ans;
-  }
-
-  std::vector<Matrix2DInt32> ReLabel(
-      const std::vector<Matrix2DInt32> &labels, int32_t max_cluster_index,
-      std::unordered_map<Int32Pair, int32_t, PairHash> chunk_speaker_to_cluster)
-      const {
-    std::vector<Matrix2DInt32> new_labels;
-    new_labels.reserve(labels.size());
-
-    int32_t chunk_index = 0;
-    for (const auto &label : labels) {
-      Matrix2DInt32 new_label(label.rows(), max_cluster_index + 1);
-      new_label.setZero();
-
-      Matrix2DInt32 t = label.transpose();
-      // t: (num_speakers, num_frames)
-
-      for (int32_t speaker_index = 0; speaker_index != t.rows();
-           ++speaker_index) {
-        if (chunk_speaker_to_cluster.count({chunk_index, speaker_index}) == 0) {
-          continue;
-        }
-
-        int32_t new_speaker_index =
-            chunk_speaker_to_cluster.at({chunk_index, speaker_index});
-
-        for (int32_t k = 0; k != t.cols(); ++k) {
-          if (t(speaker_index, k) == 1) {
-            new_label(k, new_speaker_index) = 1;
-          }
-        }
-      }
-
-      new_labels.push_back(std::move(new_label));
-
-      chunk_index += 1;
-    }
-
-    return new_labels;
-  }
-
-  Matrix2DInt32 ComputeSpeakerCount(const std::vector<Matrix2DInt32> &labels,
-                                    int32_t num_samples) const {
-    const auto &meta_data = segmentation_model_.GetModelMetaData();
-    int32_t window_size = meta_data.window_size;
-    int32_t window_shift = meta_data.window_shift;
-    int32_t receptive_field_shift = meta_data.receptive_field_shift;
-
-    int32_t num_chunks = labels.size();
-
-    int32_t num_frames = (window_size + (num_chunks - 1) * window_shift) /
-                             receptive_field_shift +
-                         1;
-
-    Matrix2DInt32 count(num_frames, labels[0].cols());
-    count.setZero();
-
-    for (int32_t i = 0; i != num_chunks; ++i) {
-      int32_t start =
-          static_cast<float>(i) * window_shift / receptive_field_shift + 0.5;
-
-      auto seq = Eigen::seqN(start, labels[i].rows());
-
-      count(seq, Eigen::all).array() += labels[i].array();
-    }
-
-    bool has_last_chunk = ((num_samples - window_size) % window_shift) > 0;
-
-    if (!has_last_chunk) {
-      return count;
-    }
-
-    int32_t last_frame = num_samples / receptive_field_shift;
-    return count(Eigen::seq(0, last_frame), Eigen::all);
-  }
-
-  Matrix2DInt32 FinalizeLabels(const Matrix2DInt32 &count,
-                               const Int32RowVector &speakers_per_frame) const {
-    int32_t num_rows = count.rows();
-    int32_t num_cols = count.cols();
-
-    Matrix2DInt32 ans(num_rows, num_cols);
-    ans.setZero();
-
-    for (int32_t i = 0; i != num_rows; ++i) {
-      int32_t k = speakers_per_frame[i];
-      if (k == 0) {
-        continue;
-      }
-      auto top_k = TopkIndex(&count(i, 0), num_cols, k);
-
-      for (int32_t m : top_k) {
-        ans(i, m) = 1;
-      }
-    }
-
-    return ans;
-  }
-
-  OfflineSpeakerDiarizationResult ComputeResult(
-      const Matrix2DInt32 &final_labels) const {
-    Matrix2DInt32 final_labels_t = final_labels.transpose();
-    int32_t num_speakers = final_labels_t.rows();
-    int32_t num_frames = final_labels_t.cols();
-
-    const auto &meta_data = segmentation_model_.GetModelMetaData();
-    int32_t window_size = meta_data.window_size;
-    int32_t window_shift = meta_data.window_shift;
-    int32_t receptive_field_shift = meta_data.receptive_field_shift;
-    int32_t receptive_field_size = meta_data.receptive_field_size;
-    int32_t sample_rate = meta_data.sample_rate;
-
-    float scale = static_cast<float>(receptive_field_shift) / sample_rate;
-    float scale_offset = 0.5 * receptive_field_size / sample_rate;
-
-    OfflineSpeakerDiarizationResult ans;
-
-    for (int32_t speaker_index = 0; speaker_index != num_speakers;
-         ++speaker_index) {
-      std::vector<OfflineSpeakerDiarizationSegment> this_speaker;
-
-      bool is_active = final_labels_t(speaker_index, 0) > 0;
-      int32_t start_index = is_active ? 0 : -1;
-
-      for (int32_t frame_index = 1; frame_index != num_frames; ++frame_index) {
-        if (is_active) {
-          if (final_labels_t(speaker_index, frame_index) == 0) {
-            float start_time = start_index * scale + scale_offset;
-            float end_time = frame_index * scale + scale_offset;
-
-            OfflineSpeakerDiarizationSegment segment(start_time, end_time,
-                                                     speaker_index);
-            this_speaker.push_back(segment);
-
-            is_active = false;
-          }
-        } else if (final_labels_t(speaker_index, frame_index) == 1) {
-          is_active = true;
-          start_index = frame_index;
-        }
-      }
-
-      if (is_active) {
-        float start_time = start_index * scale + scale_offset;
-        float end_time = (num_frames - 1) * scale + scale_offset;
-
-        OfflineSpeakerDiarizationSegment segment(start_time, end_time,
-                                                 speaker_index);
-        this_speaker.push_back(segment);
-      }
-
-      // merge segments if the gap between them is less than min_duration_off
-      MergeSegments(&this_speaker);
-
-      for (const auto &seg : this_speaker) {
-        if (seg.Duration() > config_.min_duration_on) {
-          ans.Add(seg);
-        }
-      }
-    }  // for (int32_t speaker_index = 0; speaker_index != num_speakers;
-
-    return ans;
-  }
-
-  OfflineSpeakerDiarizationResult HandleOneChunkSpecialCase(
-      const Matrix2DInt32 &final_labels, int32_t num_samples) const {
-    const auto &meta_data = segmentation_model_.GetModelMetaData();
-    int32_t window_size = meta_data.window_size;
-    int32_t window_shift = meta_data.window_shift;
-    int32_t receptive_field_shift = meta_data.receptive_field_shift;
-
-    bool has_last_chunk = (num_samples - window_size) % window_shift > 0;
-    if (!has_last_chunk) {
-      return ComputeResult(final_labels);
-    }
-
-    int32_t num_frames = final_labels.rows();
-
-    int32_t new_num_frames = num_samples / receptive_field_shift;
-
-    num_frames = (new_num_frames <= num_frames) ? new_num_frames : num_frames;
-
-    return ComputeResult(final_labels(Eigen::seq(0, num_frames), Eigen::all));
-  }
-
-  void MergeSegments(
-      std::vector<OfflineSpeakerDiarizationSegment> *segments) const {
-    float min_duration_off = config_.min_duration_off;
-    bool changed = true;
-    while (changed) {
-      changed = false;
-      for (int32_t i = 0; i < static_cast<int32_t>(segments->size()) - 1; ++i) {
-        auto s = (*segments)[i].Merge((*segments)[i + 1], min_duration_off);
-        if (s) {
-          (*segments)[i] = s.value();
-          segments->erase(segments->begin() + i + 1);
-
-          changed = true;
-          break;
-        }
-      }
-    }
-  }
-
- private:
-  OfflineSpeakerDiarizationConfig config_;
-  OfflineSpeakerSegmentationPyannoteModel segmentation_model_;
-  SpeakerEmbeddingExtractor embedding_extractor_;
-  std::unique_ptr<FastClustering> clustering_;
-  Matrix2DInt32 powerset_mapping_;
-};
-
-}  // namespace sherpa_onnx
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_DIARIZATION_PYANNOTE_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization-result.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization-result.h
deleted file mode 100644
index 6298a87c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization-result.h
+++ /dev/null
@@ -1,67 +0,0 @@
-// sherpa-onnx/csrc/offline-speaker-diarization-result.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_DIARIZATION_RESULT_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_DIARIZATION_RESULT_H_
-
-#include <cstdint>
-#include <optional>
-#include <string>
-#include <vector>
-
-namespace sherpa_onnx {
-
-class OfflineSpeakerDiarizationSegment {
- public:
-  OfflineSpeakerDiarizationSegment(float start, float end, int32_t speaker,
-                                   const std::string &text = {});
-
-  // If the gap between the two segments is less than the given gap, then we
-  // merge them and return a new segment. Otherwise, it returns null.
-  std::optional<OfflineSpeakerDiarizationSegment> Merge(
-      const OfflineSpeakerDiarizationSegment &other, float gap) const;
-
-  float Start() const { return start_; }
-  float End() const { return end_; }
-  int32_t Speaker() const { return speaker_; }
-  const std::string &Text() const { return text_; }
-  float Duration() const { return end_ - start_; }
-
-  void SetText(const std::string &text) { text_ = text; }
-
-  std::string ToString() const;
-
- private:
-  float start_;       // in seconds
-  float end_;         // in seconds
-  int32_t speaker_;   // ID of the speaker, starting from 0
-  std::string text_;  // If not empty, it contains the speech recognition result
-                      // of this segment
-};
-
-class OfflineSpeakerDiarizationResult {
- public:
-  // Add a new segment
-  void Add(const OfflineSpeakerDiarizationSegment &segment);
-
-  // Number of distinct speakers contained in this object at this point
-  int32_t NumSpeakers() const;
-
-  int32_t NumSegments() const;
-
-  // Return a list of segments sorted by segment.start time
-  std::vector<OfflineSpeakerDiarizationSegment> SortByStartTime() const;
-
-  // ans.size() == NumSpeakers().
-  // ans[i] is for speaker_i and is sorted by start time
-  std::vector<std::vector<OfflineSpeakerDiarizationSegment>> SortBySpeaker()
-      const;
-
- private:
-  std::vector<OfflineSpeakerDiarizationSegment> segments_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_DIARIZATION_RESULT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization.h
deleted file mode 100644
index acbb6f52..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-diarization.h
+++ /dev/null
@@ -1,84 +0,0 @@
-// sherpa-onnx/csrc/offline-speaker-diarization.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_DIARIZATION_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_DIARIZATION_H_
-
-#include <functional>
-#include <memory>
-#include <string>
-
-#include "sherpa-onnx/csrc/fast-clustering-config.h"
-#include "sherpa-onnx/csrc/offline-speaker-diarization-result.h"
-#include "sherpa-onnx/csrc/offline-speaker-segmentation-model-config.h"
-#include "sherpa-onnx/csrc/speaker-embedding-extractor.h"
-
-namespace sherpa_onnx {
-
-struct OfflineSpeakerDiarizationConfig {
-  OfflineSpeakerSegmentationModelConfig segmentation;
-  SpeakerEmbeddingExtractorConfig embedding;
-  FastClusteringConfig clustering;
-
-  // if a segment is less than this value, then it is discarded
-  float min_duration_on = 0.3;  // in seconds
-
-  // if the gap between to segments of the same speaker is less than this value,
-  // then these two segments are merged into a single segment.
-  // We do this recursively.
-  float min_duration_off = 0.5;  // in seconds
-
-  OfflineSpeakerDiarizationConfig() = default;
-
-  OfflineSpeakerDiarizationConfig(
-      const OfflineSpeakerSegmentationModelConfig &segmentation,
-      const SpeakerEmbeddingExtractorConfig &embedding,
-      const FastClusteringConfig &clustering, float min_duration_on,
-      float min_duration_off)
-      : segmentation(segmentation),
-        embedding(embedding),
-        clustering(clustering),
-        min_duration_on(min_duration_on),
-        min_duration_off(min_duration_off) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-  std::string ToString() const;
-};
-
-class OfflineSpeakerDiarizationImpl;
-
-using OfflineSpeakerDiarizationProgressCallback = std::function<int32_t(
-    int32_t processed_chunks, int32_t num_chunks, void *arg)>;
-
-class OfflineSpeakerDiarization {
- public:
-  explicit OfflineSpeakerDiarization(
-      const OfflineSpeakerDiarizationConfig &config);
-
-  template <typename Manager>
-  OfflineSpeakerDiarization(Manager *mgr,
-                            const OfflineSpeakerDiarizationConfig &config);
-
-  ~OfflineSpeakerDiarization();
-
-  // Expected sample rate of the input audio samples
-  int32_t SampleRate() const;
-
-  // Note: Only config.clustering is used. All other fields in config are
-  // ignored
-  void SetConfig(const OfflineSpeakerDiarizationConfig &config);
-
-  OfflineSpeakerDiarizationResult Process(
-      const float *audio, int32_t n,
-      OfflineSpeakerDiarizationProgressCallback callback = nullptr,
-      void *callback_arg = nullptr) const;
-
- private:
-  std::unique_ptr<OfflineSpeakerDiarizationImpl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_DIARIZATION_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-model-config.h
deleted file mode 100644
index 8e9e4a96..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-model-config.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// sherpa-onnx/csrc/offline-speaker-segmentation-model-config.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_SEGMENTATION_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_SEGMENTATION_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model-config.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineSpeakerSegmentationModelConfig {
-  OfflineSpeakerSegmentationPyannoteModelConfig pyannote;
-
-  int32_t num_threads = 1;
-  bool debug = false;
-  std::string provider = "cpu";
-
-  OfflineSpeakerSegmentationModelConfig() = default;
-
-  explicit OfflineSpeakerSegmentationModelConfig(
-      const OfflineSpeakerSegmentationPyannoteModelConfig &pyannote,
-      int32_t num_threads, bool debug, const std::string &provider)
-      : pyannote(pyannote),
-        num_threads(num_threads),
-        debug(debug),
-        provider(provider) {}
-
-  void Register(ParseOptions *po);
-
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_SEGMENTATION_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model-config.h
deleted file mode 100644
index fb5ca4a4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model-config.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model-config.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_SEGMENTATION_PYANNOTE_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_SEGMENTATION_PYANNOTE_MODEL_CONFIG_H_
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineSpeakerSegmentationPyannoteModelConfig {
-  std::string model;
-
-  OfflineSpeakerSegmentationPyannoteModelConfig() = default;
-
-  explicit OfflineSpeakerSegmentationPyannoteModelConfig(
-      const std::string &model)
-      : model(model) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_SEGMENTATION_PYANNOTE_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model-meta-data.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model-meta-data.h
deleted file mode 100644
index 728ed7ff..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model-meta-data.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model-meta-data.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_SEGMENTATION_PYANNOTE_MODEL_META_DATA_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_SEGMENTATION_PYANNOTE_MODEL_META_DATA_H_
-
-#include <cstdint>
-#include <string>
-
-namespace sherpa_onnx {
-
-// If you are not sure what each field means, please
-// have a look of the Python file in the model directory that
-// you have downloaded.
-struct OfflineSpeakerSegmentationPyannoteModelMetaData {
-  int32_t sample_rate = 0;
-  int32_t window_size = 0;            // in samples
-  int32_t window_shift = 0;           // in samples
-  int32_t receptive_field_size = 0;   // in samples
-  int32_t receptive_field_shift = 0;  // in samples
-  int32_t num_speakers = 0;
-  int32_t powerset_max_classes = 0;
-  int32_t num_classes = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_SEGMENTATION_PYANNOTE_MODEL_META_DATA_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model.h
deleted file mode 100644
index a3cc7ed3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model.h
+++ /dev/null
@@ -1,44 +0,0 @@
-// sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_SEGMENTATION_PYANNOTE_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_SEGMENTATION_PYANNOTE_MODEL_H_
-
-#include <memory>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-speaker-segmentation-model-config.h"
-#include "sherpa-onnx/csrc/offline-speaker-segmentation-pyannote-model-meta-data.h"
-
-namespace sherpa_onnx {
-
-class OfflineSpeakerSegmentationPyannoteModel {
- public:
-  explicit OfflineSpeakerSegmentationPyannoteModel(
-      const OfflineSpeakerSegmentationModelConfig &config);
-
-  template <typename Manager>
-  OfflineSpeakerSegmentationPyannoteModel(
-      Manager *mgr, const OfflineSpeakerSegmentationModelConfig &config);
-
-  ~OfflineSpeakerSegmentationPyannoteModel();
-
-  const OfflineSpeakerSegmentationPyannoteModelMetaData &GetModelMetaData()
-      const;
-
-  /**
-   * @param x A 3-D float tensor of shape (batch_size, 1, num_samples)
-   * @return Return a float tensor of
-   *         shape (batch_size, num_frames, num_speakers). Note that
-   *         num_speakers here uses powerset encoding.
-   */
-  Ort::Value Forward(Ort::Value x) const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_SPEAKER_SEGMENTATION_PYANNOTE_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-stream.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-stream.h
deleted file mode 100644
index e4bed111..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-stream.h
+++ /dev/null
@@ -1,106 +0,0 @@
-// sherpa-onnx/csrc/offline-stream.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_STREAM_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_STREAM_H_
-#include <stdint.h>
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "sherpa-onnx/csrc/context-graph.h"
-#include "sherpa-onnx/csrc/features.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineRecognitionResult {
-  // Recognition results.
-  // For English, it consists of space separated words.
-  // For Chinese, it consists of Chinese words without spaces.
-  std::string text;
-
-  // Decoded results at the token level.
-  // For instance, for BPE-based models it consists of a list of BPE tokens.
-  std::vector<std::string> tokens;
-
-  std::string lang;
-
-  // emotion target of the audio.
-  std::string emotion;
-
-  // event target of the audio.
-  std::string event;
-
-  /// timestamps.size() == tokens.size()
-  /// timestamps[i] records the time in seconds when tokens[i] is decoded.
-  std::vector<float> timestamps;
-
-  std::vector<int32_t> words;
-
-  std::string AsJsonString() const;
-};
-
-struct WhisperTag {
-  int32_t dim = 80;
-};
-
-struct CEDTag {};
-
-// It uses a neural network model, a preprocessor, to convert
-// audio samples to features
-struct MoonshineTag {};
-
-class OfflineStream {
- public:
-  explicit OfflineStream(const FeatureExtractorConfig &config = {},
-                         ContextGraphPtr context_graph = {});
-
-  explicit OfflineStream(WhisperTag tag);
-  explicit OfflineStream(CEDTag tag);
-  explicit OfflineStream(MoonshineTag tag);
-  ~OfflineStream();
-
-  /**
-     @param sampling_rate The sampling_rate of the input waveform. If it does
-                          not equal to  config.sampling_rate, we will do
-                          resampling inside.
-     @param waveform Pointer to a 1-D array of size n. It must be normalized to
-                     the range [-1, 1].
-     @param n Number of entries in waveform
-
-     Caution: You can only invoke this function once so you have to input
-              all the samples at once
-   */
-  void AcceptWaveform(int32_t sampling_rate, const float *waveform,
-                      int32_t n) const;
-
-  /// Return feature dim of this extractor.
-  ///
-  /// Note: if it is Moonshine, then it returns the number of audio samples
-  /// currently received.
-  int32_t FeatureDim() const;
-
-  // Get all the feature frames of this stream in a 1-D array, which is
-  // flattened from a 2-D array of shape (num_frames, feat_dim).
-  std::vector<float> GetFrames() const;
-
-  /** Set the recognition result for this stream. */
-  void SetResult(const OfflineRecognitionResult &r);
-
-  /** Get the recognition result of this stream */
-  const OfflineRecognitionResult &GetResult() const;
-
-  /** Get the ContextGraph of this stream */
-  const ContextGraphPtr &GetContextGraph() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_STREAM_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tdnn-ctc-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tdnn-ctc-model.h
deleted file mode 100644
index 0c2e43af..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tdnn-ctc-model.h
+++ /dev/null
@@ -1,59 +0,0 @@
-// sherpa-onnx/csrc/offline-tdnn-ctc-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TDNN_CTC_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TDNN_CTC_MODEL_H_
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-ctc-model.h"
-#include "sherpa-onnx/csrc/offline-model-config.h"
-
-namespace sherpa_onnx {
-
-/** This class implements the tdnn model of the yesno recipe from icefall.
- *
- * See
- * https://github.com/k2-fsa/icefall/tree/master/egs/yesno/ASR/tdnn
- */
-class OfflineTdnnCtcModel : public OfflineCtcModel {
- public:
-  explicit OfflineTdnnCtcModel(const OfflineModelConfig &config);
-
-  template <typename Manager>
-  OfflineTdnnCtcModel(Manager *mgr, const OfflineModelConfig &config);
-
-  ~OfflineTdnnCtcModel() override;
-
-  /** Run the forward method of the model.
-   *
-   * @param features  A tensor of shape (N, T, C).
-   * @param features_length  A 1-D tensor of shape (N,) containing number of
-   *                         valid frames in `features` before padding.
-   *                         Its dtype is int64_t.
-   *
-   * @return Return a pair containing:
-   *  - log_probs: A 3-D tensor of shape (N, T', vocab_size).
-   *  - log_probs_length A 1-D tensor of shape (N,). Its dtype is int64_t
-   */
-  std::vector<Ort::Value> Forward(Ort::Value features,
-                                  Ort::Value /*features_length*/) override;
-
-  /** Return the vocabulary size of the model
-   */
-  int32_t VocabSize() const override;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const override;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TDNN_CTC_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tdnn-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tdnn-model-config.h
deleted file mode 100644
index bddea551..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tdnn-model-config.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// sherpa-onnx/csrc/offline-tdnn-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TDNN_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TDNN_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-// for https://github.com/k2-fsa/icefall/tree/master/egs/yesno/ASR/tdnn
-struct OfflineTdnnModelConfig {
-  std::string model;
-
-  OfflineTdnnModelConfig() = default;
-  explicit OfflineTdnnModelConfig(const std::string &model) : model(model) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TDNN_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-telespeech-ctc-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-telespeech-ctc-model.h
deleted file mode 100644
index 3fe701a0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-telespeech-ctc-model.h
+++ /dev/null
@@ -1,74 +0,0 @@
-// sherpa-onnx/csrc/offline-telespeech-ctc-model.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TELESPEECH_CTC_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TELESPEECH_CTC_MODEL_H_
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-ctc-model.h"
-#include "sherpa-onnx/csrc/offline-model-config.h"
-
-namespace sherpa_onnx {
-
-/** This class implements the CTC model from
- * https://github.com/Tele-AI/TeleSpeech-ASR.
- *
- * See
- * https://github.com/lovemefan/telespeech-asr-python/blob/main/telespeechasr/onnx/onnx_infer.py
- * and
- * https://github.com/k2-fsa/sherpa-onnx/blob/master/scripts/tele-speech/test.py
- */
-class OfflineTeleSpeechCtcModel : public OfflineCtcModel {
- public:
-  explicit OfflineTeleSpeechCtcModel(const OfflineModelConfig &config);
-
-  template <typename Manager>
-  OfflineTeleSpeechCtcModel(Manager *mgr, const OfflineModelConfig &config);
-
-  ~OfflineTeleSpeechCtcModel() override;
-
-  /** Run the forward method of the model.
-   *
-   * @param features  A tensor of shape (N, T, C).
-   * @param features_length  A 1-D tensor of shape (N,) containing number of
-   *                         valid frames in `features` before padding.
-   *                         Its dtype is int64_t.
-   *
-   * @return Return a vector containing:
-   *  - log_probs: A 3-D tensor of shape (N, T', vocab_size).
-   *  - log_probs_length A 1-D tensor of shape (N,). Its dtype is int64_t
-   */
-  std::vector<Ort::Value> Forward(Ort::Value features,
-                                  Ort::Value features_length) override;
-
-  /** Return the vocabulary size of the model
-   */
-  int32_t VocabSize() const override;
-
-  /** SubsamplingFactor of the model
-   */
-  int32_t SubsamplingFactor() const override;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const override;
-
-  // TeleSpeech CTC models do not support batch size > 1
-  bool SupportBatchProcessing() const override { return false; }
-
-  std::string FeatureNormalizationMethod() const override {
-    return "per_feature";
-  }
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TELESPEECH_CTC_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-decoder.h
deleted file mode 100644
index 36d93e44..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-decoder.h
+++ /dev/null
@@ -1,43 +0,0 @@
-// sherpa-onnx/csrc/offline-transducer-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_DECODER_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_DECODER_H_
-
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-stream.h"
-
-namespace sherpa_onnx {
-
-struct OfflineTransducerDecoderResult {
-  /// The decoded token IDs
-  std::vector<int64_t> tokens;
-
-  /// timestamps[i] contains the output frame index where tokens[i] is decoded.
-  /// Note: The index is after subsampling
-  std::vector<int32_t> timestamps;
-};
-
-class OfflineTransducerDecoder {
- public:
-  virtual ~OfflineTransducerDecoder() = default;
-
-  /** Run transducer beam search given the output from the encoder model.
-   *
-   * @param encoder_out A 3-D tensor of shape (N, T, joiner_dim)
-   * @param encoder_out_length A 1-D tensor of shape (N,) containing number
-   *                           of valid frames in encoder_out before padding.
-   *
-   * @return Return a vector of size `N` containing the decoded results.
-   */
-  virtual std::vector<OfflineTransducerDecoderResult> Decode(
-      Ort::Value encoder_out, Ort::Value encoder_out_length,
-      OfflineStream **ss = nullptr, int32_t n = 0) = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-greedy-search-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-greedy-search-decoder.h
deleted file mode 100644
index 79109e60..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-greedy-search-decoder.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// sherpa-onnx/csrc/offline-transducer-greedy-search-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_GREEDY_SEARCH_DECODER_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_GREEDY_SEARCH_DECODER_H_
-
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-transducer-decoder.h"
-#include "sherpa-onnx/csrc/offline-transducer-model.h"
-
-namespace sherpa_onnx {
-
-class OfflineTransducerGreedySearchDecoder : public OfflineTransducerDecoder {
- public:
-  OfflineTransducerGreedySearchDecoder(OfflineTransducerModel *model,
-                                       int32_t unk_id,
-                                       float blank_penalty)
-      : model_(model), unk_id_(unk_id), blank_penalty_(blank_penalty) {}
-
-  std::vector<OfflineTransducerDecoderResult> Decode(
-      Ort::Value encoder_out, Ort::Value encoder_out_length,
-      OfflineStream **ss = nullptr, int32_t n = 0) override;
-
- private:
-  OfflineTransducerModel *model_;  // Not owned
-  int32_t unk_id_;
-  float blank_penalty_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_GREEDY_SEARCH_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-greedy-search-nemo-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-greedy-search-nemo-decoder.h
deleted file mode 100644
index ab0ca422..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-greedy-search-nemo-decoder.h
+++ /dev/null
@@ -1,33 +0,0 @@
-// sherpa-onnx/csrc/offline-transducer-greedy-search-nemo-decoder.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_GREEDY_SEARCH_NEMO_DECODER_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_GREEDY_SEARCH_NEMO_DECODER_H_
-
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-transducer-decoder.h"
-#include "sherpa-onnx/csrc/offline-transducer-nemo-model.h"
-
-namespace sherpa_onnx {
-
-class OfflineTransducerGreedySearchNeMoDecoder
-    : public OfflineTransducerDecoder {
- public:
-  OfflineTransducerGreedySearchNeMoDecoder(OfflineTransducerNeMoModel *model,
-                                           float blank_penalty)
-      : model_(model), blank_penalty_(blank_penalty) {}
-
-  std::vector<OfflineTransducerDecoderResult> Decode(
-      Ort::Value encoder_out, Ort::Value encoder_out_length,
-      OfflineStream **ss = nullptr, int32_t n = 0) override;
-
- private:
-  OfflineTransducerNeMoModel *model_;  // Not owned
-  float blank_penalty_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_GREEDY_SEARCH_NEMO_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-model-config.h
deleted file mode 100644
index 1b51f104..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-model-config.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// sherpa-onnx/csrc/offline-transducer-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineTransducerModelConfig {
-  std::string encoder_filename;
-  std::string decoder_filename;
-  std::string joiner_filename;
-
-  OfflineTransducerModelConfig() = default;
-  OfflineTransducerModelConfig(const std::string &encoder_filename,
-                               const std::string &decoder_filename,
-                               const std::string &joiner_filename)
-      : encoder_filename(encoder_filename),
-        decoder_filename(decoder_filename),
-        joiner_filename(joiner_filename) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-model.h
deleted file mode 100644
index e8ca93a3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-model.h
+++ /dev/null
@@ -1,104 +0,0 @@
-// sherpa-onnx/csrc/offline-transducer-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_MODEL_H_
-
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/hypothesis.h"
-#include "sherpa-onnx/csrc/offline-model-config.h"
-
-namespace sherpa_onnx {
-
-struct OfflineTransducerDecoderResult;
-
-class OfflineTransducerModel {
- public:
-  explicit OfflineTransducerModel(const OfflineModelConfig &config);
-
-  template <typename Manager>
-  OfflineTransducerModel(Manager *mgr, const OfflineModelConfig &config);
-
-  ~OfflineTransducerModel();
-
-  /** Run the encoder.
-   *
-   * @param features  A tensor of shape (N, T, C). It is changed in-place.
-   * @param features_length  A 1-D tensor of shape (N,) containing number of
-   *                         valid frames in `features` before padding.
-   *                         Its dtype is int64_t.
-   *
-   * @return Return a pair containing:
-   *  - encoder_out: A 3-D tensor of shape (N, T', encoder_dim)
-   *  - encoder_out_length: A 1-D tensor of shape (N,) containing number
-   *                        of frames in `encoder_out` before padding.
-   */
-  std::pair<Ort::Value, Ort::Value> RunEncoder(Ort::Value features,
-                                               Ort::Value features_length);
-
-  /** Run the decoder network.
-   *
-   * Caution: We assume there are no recurrent connections in the decoder and
-   *          the decoder is stateless. See
-   * https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless2/decoder.py
-   *          for an example
-   *
-   * @param decoder_input It is usually of shape (N, context_size)
-   * @return Return a tensor of shape (N, decoder_dim).
-   */
-  Ort::Value RunDecoder(Ort::Value decoder_input);
-
-  /** Run the joint network.
-   *
-   * @param encoder_out Output of the encoder network. A tensor of shape
-   *                    (N, joiner_dim).
-   * @param decoder_out Output of the decoder network. A tensor of shape
-   *                    (N, joiner_dim).
-   * @return Return a tensor of shape (N, vocab_size). In icefall, the last
-   *         last layer of the joint network is `nn.Linear`,
-   *         not `nn.LogSoftmax`.
-   */
-  Ort::Value RunJoiner(Ort::Value encoder_out, Ort::Value decoder_out);
-
-  /** Return the vocabulary size of the model
-   */
-  int32_t VocabSize() const;
-
-  /** Return the context_size of the decoder model.
-   */
-  int32_t ContextSize() const;
-
-  /** Return the subsampling factor of the model.
-   */
-  int32_t SubsamplingFactor() const;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const;
-
-  /** Build decoder_input from the current results.
-   *
-   * @param results Current decoded results.
-   * @param end_index We only use results[0:end_index] to build
-   *                  the decoder_input. results[end_index] is not used.
-   * @return Return a tensor of shape (results.size(), ContextSize())
-   */
-  Ort::Value BuildDecoderInput(
-      const std::vector<OfflineTransducerDecoderResult> &results,
-      int32_t end_index) const;
-
-  Ort::Value BuildDecoderInput(const std::vector<Hypothesis> &results,
-                               int32_t end_index) const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-modified-beam-search-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-modified-beam-search-decoder.h
deleted file mode 100644
index 2e67cd71..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-modified-beam-search-decoder.h
+++ /dev/null
@@ -1,47 +0,0 @@
-// sherpa-onnx/csrc/offline-transducer-modified-beam-search-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_MODIFIED_BEAM_SEARCH_DECODER_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_MODIFIED_BEAM_SEARCH_DECODER_H_
-
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-lm.h"
-#include "sherpa-onnx/csrc/offline-transducer-decoder.h"
-#include "sherpa-onnx/csrc/offline-transducer-model.h"
-
-namespace sherpa_onnx {
-
-class OfflineTransducerModifiedBeamSearchDecoder
-    : public OfflineTransducerDecoder {
- public:
-  OfflineTransducerModifiedBeamSearchDecoder(OfflineTransducerModel *model,
-                                             OfflineLM *lm,
-                                             int32_t max_active_paths,
-                                             float lm_scale, int32_t unk_id,
-                                             float blank_penalty)
-      : model_(model),
-        lm_(lm),
-        max_active_paths_(max_active_paths),
-        lm_scale_(lm_scale),
-        unk_id_(unk_id),
-        blank_penalty_(blank_penalty) {}
-
-  std::vector<OfflineTransducerDecoderResult> Decode(
-      Ort::Value encoder_out, Ort::Value encoder_out_length,
-      OfflineStream **ss = nullptr, int32_t n = 0) override;
-
- private:
-  OfflineTransducerModel *model_;  // Not owned
-  OfflineLM *lm_;                  // Not owned; may be nullptr
-
-  int32_t max_active_paths_;
-  float lm_scale_;  // used only when lm_ is not nullptr
-  int32_t unk_id_;
-  float blank_penalty_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_MODIFIED_BEAM_SEARCH_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-nemo-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-nemo-model.h
deleted file mode 100644
index 697f749e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-transducer-nemo-model.h
+++ /dev/null
@@ -1,98 +0,0 @@
-// sherpa-onnx/csrc/offline-transducer-nemo-model.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_NEMO_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_NEMO_MODEL_H_
-
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-model-config.h"
-
-namespace sherpa_onnx {
-
-// see
-// https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/models/hybrid_rnnt_ctc_bpe_models.py#L40
-// Its decoder is stateful, not stateless.
-class OfflineTransducerNeMoModel {
- public:
-  explicit OfflineTransducerNeMoModel(const OfflineModelConfig &config);
-
-  template <typename Manager>
-  OfflineTransducerNeMoModel(Manager *mgr, const OfflineModelConfig &config);
-
-  ~OfflineTransducerNeMoModel();
-
-  /** Run the encoder.
-   *
-   * @param features  A tensor of shape (N, T, C). It is changed in-place.
-   * @param features_length  A 1-D tensor of shape (N,) containing number of
-   *                         valid frames in `features` before padding.
-   *                         Its dtype is int64_t.
-   *
-   * @return Return a vector containing:
-   *  - encoder_out: A 3-D tensor of shape (N, T', encoder_dim)
-   *  - encoder_out_length: A 1-D tensor of shape (N,) containing number
-   *                        of frames in `encoder_out` before padding.
-   */
-  std::vector<Ort::Value> RunEncoder(Ort::Value features,
-                                     Ort::Value features_length) const;
-
-  /** Run the decoder network.
-   *
-   * @param targets A int32 tensor of shape (batch_size, 1)
-   * @param targets_length A int32 tensor of shape (batch_size,)
-   * @param states The states for the decoder model.
-   * @return Return a vector:
-   *           - ans[0] is the decoder_out (a float tensor)
-   *           - ans[1] is the decoder_out_length (a int32 tensor)
-   *           - ans[2:] is the states_next
-   */
-  std::pair<Ort::Value, std::vector<Ort::Value>> RunDecoder(
-      Ort::Value targets, Ort::Value targets_length,
-      std::vector<Ort::Value> states) const;
-
-  std::vector<Ort::Value> GetDecoderInitStates(int32_t batch_size) const;
-
-  /** Run the joint network.
-   *
-   * @param encoder_out Output of the encoder network.
-   * @param decoder_out Output of the decoder network.
-   * @return Return a tensor of shape (N, 1, 1, vocab_size) containing logits.
-   */
-  Ort::Value RunJoiner(Ort::Value encoder_out, Ort::Value decoder_out) const;
-
-  /** Return the subsampling factor of the model.
-   */
-  int32_t SubsamplingFactor() const;
-
-  int32_t VocabSize() const;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const;
-
-  // Possible values:
-  // - per_feature
-  // - all_features (not implemented yet)
-  // - fixed_mean (not implemented)
-  // - fixed_std (not implemented)
-  // - or just leave it to empty
-  // See
-  // https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/parts/preprocessing/features.py#L59
-  // for details
-  std::string FeatureNormalizationMethod() const;
-
-  bool IsGigaAM() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TRANSDUCER_NEMO_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-character-frontend.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-character-frontend.h
deleted file mode 100644
index fcd2f6dd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-character-frontend.h
+++ /dev/null
@@ -1,48 +0,0 @@
-// sherpa-onnx/csrc/offline-tts-character-frontend.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TTS_CHARACTER_FRONTEND_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TTS_CHARACTER_FRONTEND_H_
-#include <cstdint>
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-tts-frontend.h"
-#include "sherpa-onnx/csrc/offline-tts-vits-model-metadata.h"
-
-namespace sherpa_onnx {
-
-class OfflineTtsCharacterFrontend : public OfflineTtsFrontend {
- public:
-  OfflineTtsCharacterFrontend(const std::string &tokens,
-                              const OfflineTtsVitsModelMetaData &meta_data);
-
-  template <typename Manager>
-  OfflineTtsCharacterFrontend(Manager *mgr, const std::string &tokens,
-                              const OfflineTtsVitsModelMetaData &meta_data);
-
-  /** Convert a string to token IDs.
-   *
-   * @param text The input text.
-   *             Example 1: "This is the first sample sentence; this is the
-   *             second one." Example 2: "这是第一句。这是第二句。"
-   * @param voice Optional. It is for espeak-ng.
-   *
-   * @return Return a vector-of-vector of token IDs. Each subvector contains
-   *         a sentence that can be processed independently.
-   *         If a frontend does not support splitting the text into
-   * sentences, the resulting vector contains only one subvector.
-   */
-  std::vector<TokenIDs> ConvertTextToTokenIds(
-      const std::string &text, const std::string &voice = "") const override;
-
- private:
-  OfflineTtsVitsModelMetaData meta_data_;
-  std::unordered_map<char32_t, int32_t> token2id_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TTS_CHARACTER_FRONTEND_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-frontend.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-frontend.h
deleted file mode 100644
index cba50e36..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-frontend.h
+++ /dev/null
@@ -1,56 +0,0 @@
-// sherpa-onnx/csrc/offline-tts-frontend.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TTS_FRONTEND_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TTS_FRONTEND_H_
-#include <cstdint>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "sherpa-onnx/csrc/macros.h"
-
-namespace sherpa_onnx {
-
-struct TokenIDs {
-  TokenIDs() = default;
-
-  /*implicit*/ TokenIDs(std::vector<int64_t> tokens)  // NOLINT
-      : tokens{std::move(tokens)} {}
-
-  TokenIDs(std::vector<int64_t> tokens,  // NOLINT
-           std::vector<int64_t> tones)   // NOLINT
-      : tokens{std::move(tokens)}, tones{std::move(tones)} {}
-
-  std::string ToString() const;
-
-  std::vector<int64_t> tokens;
-
-  // Used only in MeloTTS
-  std::vector<int64_t> tones;
-};
-
-class OfflineTtsFrontend {
- public:
-  virtual ~OfflineTtsFrontend() = default;
-
-  /** Convert a string to token IDs.
-   *
-   * @param text The input text.
-   *             Example 1: "This is the first sample sentence; this is the
-   *             second one." Example 2: "这是第一句。这是第二句。"
-   * @param voice Optional. It is for espeak-ng.
-   *
-   * @return Return a vector-of-vector of token IDs. Each subvector contains
-   *         a sentence that can be processed independently.
-   *         If a frontend does not support splitting the text into sentences,
-   *         the resulting vector contains only one subvector.
-   */
-  virtual std::vector<TokenIDs> ConvertTextToTokenIds(
-      const std::string &text, const std::string &voice = "") const = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TTS_FRONTEND_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-impl.h
deleted file mode 100644
index db8b7162..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-impl.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// sherpa-onnx/csrc/offline-tts-impl.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TTS_IMPL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TTS_IMPL_H_
-
-#include <memory>
-#include <string>
-
-#include "sherpa-onnx/csrc/offline-tts.h"
-
-namespace sherpa_onnx {
-
-class OfflineTtsImpl {
- public:
-  virtual ~OfflineTtsImpl() = default;
-
-  static std::unique_ptr<OfflineTtsImpl> Create(const OfflineTtsConfig &config);
-
-  template <typename Manager>
-  static std::unique_ptr<OfflineTtsImpl> Create(Manager *mgr,
-                                                const OfflineTtsConfig &config);
-
-  virtual GeneratedAudio Generate(
-      const std::string &text, int64_t sid = 0, float speed = 1.0,
-      GeneratedAudioCallback callback = nullptr) const = 0;
-
-  // Return the sample rate of the generated audio
-  virtual int32_t SampleRate() const = 0;
-
-  // Number of supported speakers.
-  // If it supports only a single speaker, then it return 0 or 1.
-  virtual int32_t NumSpeakers() const = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TTS_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-model-config.h
deleted file mode 100644
index bee50ba1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-model-config.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// sherpa-onnx/csrc/offline-tts-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TTS_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TTS_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/offline-tts-vits-model-config.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineTtsModelConfig {
-  OfflineTtsVitsModelConfig vits;
-
-  int32_t num_threads = 1;
-  bool debug = false;
-  std::string provider = "cpu";
-
-  OfflineTtsModelConfig() = default;
-
-  OfflineTtsModelConfig(const OfflineTtsVitsModelConfig &vits,
-                        int32_t num_threads, bool debug,
-                        const std::string &provider)
-      : vits(vits),
-        num_threads(num_threads),
-        debug(debug),
-        provider(provider) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TTS_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-impl.h
deleted file mode 100644
index 56057635..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-impl.h
+++ /dev/null
@@ -1,493 +0,0 @@
-// sherpa-onnx/csrc/offline-tts-vits-impl.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TTS_VITS_IMPL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TTS_VITS_IMPL_H_
-
-#include <memory>
-#include <string>
-#include <strstream>
-#include <utility>
-#include <vector>
-
-#include "fst/extensions/far/far.h"
-#include "kaldifst/csrc/kaldi-fst-io.h"
-#include "kaldifst/csrc/text-normalizer.h"
-#include "sherpa-onnx/csrc/jieba-lexicon.h"
-#include "sherpa-onnx/csrc/lexicon.h"
-#include "sherpa-onnx/csrc/macros.h"
-#include "sherpa-onnx/csrc/melo-tts-lexicon.h"
-#include "sherpa-onnx/csrc/offline-tts-character-frontend.h"
-#include "sherpa-onnx/csrc/offline-tts-frontend.h"
-#include "sherpa-onnx/csrc/offline-tts-impl.h"
-#include "sherpa-onnx/csrc/offline-tts-vits-model.h"
-#include "sherpa-onnx/csrc/onnx-utils.h"
-#include "sherpa-onnx/csrc/piper-phonemize-lexicon.h"
-#include "sherpa-onnx/csrc/text-utils.h"
-
-namespace sherpa_onnx {
-
-class OfflineTtsVitsImpl : public OfflineTtsImpl {
- public:
-  explicit OfflineTtsVitsImpl(const OfflineTtsConfig &config)
-      : config_(config),
-        model_(std::make_unique<OfflineTtsVitsModel>(config.model)) {
-    InitFrontend();
-
-    if (!config.rule_fsts.empty()) {
-      std::vector<std::string> files;
-      SplitStringToVector(config.rule_fsts, ",", false, &files);
-      tn_list_.reserve(files.size());
-      for (const auto &f : files) {
-        if (config.model.debug) {
-#if __OHOS__
-          SHERPA_ONNX_LOGE("rule fst: %{public}s", f.c_str());
-#else
-          SHERPA_ONNX_LOGE("rule fst: %s", f.c_str());
-#endif
-        }
-        tn_list_.push_back(std::make_unique<kaldifst::TextNormalizer>(f));
-      }
-    }
-
-    if (!config.rule_fars.empty()) {
-      if (config.model.debug) {
-        SHERPA_ONNX_LOGE("Loading FST archives");
-      }
-      std::vector<std::string> files;
-      SplitStringToVector(config.rule_fars, ",", false, &files);
-
-      tn_list_.reserve(files.size() + tn_list_.size());
-
-      for (const auto &f : files) {
-        if (config.model.debug) {
-#if __OHOS__
-          SHERPA_ONNX_LOGE("rule far: %{public}s", f.c_str());
-#else
-          SHERPA_ONNX_LOGE("rule far: %s", f.c_str());
-#endif
-        }
-        std::unique_ptr<fst::FarReader<fst::StdArc>> reader(
-            fst::FarReader<fst::StdArc>::Open(f));
-        for (; !reader->Done(); reader->Next()) {
-          std::unique_ptr<fst::StdConstFst> r(
-              fst::CastOrConvertToConstFst(reader->GetFst()->Copy()));
-
-          tn_list_.push_back(
-              std::make_unique<kaldifst::TextNormalizer>(std::move(r)));
-        }
-      }
-
-      if (config.model.debug) {
-        SHERPA_ONNX_LOGE("FST archives loaded!");
-      }
-    }
-  }
-
-  template <typename Manager>
-  OfflineTtsVitsImpl(Manager *mgr, const OfflineTtsConfig &config)
-      : config_(config),
-        model_(std::make_unique<OfflineTtsVitsModel>(mgr, config.model)) {
-    InitFrontend(mgr);
-
-    if (!config.rule_fsts.empty()) {
-      std::vector<std::string> files;
-      SplitStringToVector(config.rule_fsts, ",", false, &files);
-      tn_list_.reserve(files.size());
-      for (const auto &f : files) {
-        if (config.model.debug) {
-#if __OHOS__
-          SHERPA_ONNX_LOGE("rule fst: %{public}s", f.c_str());
-#else
-          SHERPA_ONNX_LOGE("rule fst: %s", f.c_str());
-#endif
-        }
-        auto buf = ReadFile(mgr, f);
-        std::istrstream is(buf.data(), buf.size());
-        tn_list_.push_back(std::make_unique<kaldifst::TextNormalizer>(is));
-      }
-    }
-
-    if (!config.rule_fars.empty()) {
-      std::vector<std::string> files;
-      SplitStringToVector(config.rule_fars, ",", false, &files);
-      tn_list_.reserve(files.size() + tn_list_.size());
-
-      for (const auto &f : files) {
-        if (config.model.debug) {
-#if __OHOS__
-          SHERPA_ONNX_LOGE("rule far: %{public}s", f.c_str());
-#else
-          SHERPA_ONNX_LOGE("rule far: %s", f.c_str());
-#endif
-        }
-
-        auto buf = ReadFile(mgr, f);
-
-        std::unique_ptr<std::istream> s(
-            new std::istrstream(buf.data(), buf.size()));
-
-        std::unique_ptr<fst::FarReader<fst::StdArc>> reader(
-            fst::FarReader<fst::StdArc>::Open(std::move(s)));
-
-        for (; !reader->Done(); reader->Next()) {
-          std::unique_ptr<fst::StdConstFst> r(
-              fst::CastOrConvertToConstFst(reader->GetFst()->Copy()));
-
-          tn_list_.push_back(
-              std::make_unique<kaldifst::TextNormalizer>(std::move(r)));
-        }  // for (; !reader->Done(); reader->Next())
-      }    // for (const auto &f : files)
-    }      // if (!config.rule_fars.empty())
-  }
-
-  int32_t SampleRate() const override {
-    return model_->GetMetaData().sample_rate;
-  }
-
-  int32_t NumSpeakers() const override {
-    return model_->GetMetaData().num_speakers;
-  }
-
-  GeneratedAudio Generate(
-      const std::string &_text, int64_t sid = 0, float speed = 1.0,
-      GeneratedAudioCallback callback = nullptr) const override {
-    const auto &meta_data = model_->GetMetaData();
-    int32_t num_speakers = meta_data.num_speakers;
-
-    if (num_speakers == 0 && sid != 0) {
-      SHERPA_ONNX_LOGE(
-          "This is a single-speaker model and supports only sid 0. Given sid: "
-          "%d. sid is ignored",
-          static_cast<int32_t>(sid));
-    }
-
-    if (num_speakers != 0 && (sid >= num_speakers || sid < 0)) {
-      SHERPA_ONNX_LOGE(
-          "This model contains only %d speakers. sid should be in the range "
-          "[%d, %d]. Given: %d. Use sid=0",
-          num_speakers, 0, num_speakers - 1, static_cast<int32_t>(sid));
-      sid = 0;
-    }
-
-    std::string text = _text;
-    if (config_.model.debug) {
-#if __OHOS__
-      SHERPA_ONNX_LOGE("Raw text: %{public}s", text.c_str());
-#else
-      SHERPA_ONNX_LOGE("Raw text: %s", text.c_str());
-#endif
-    }
-
-    if (!tn_list_.empty()) {
-      for (const auto &tn : tn_list_) {
-        text = tn->Normalize(text);
-        if (config_.model.debug) {
-#if __OHOS__
-          SHERPA_ONNX_LOGE("After normalizing: %{public}s", text.c_str());
-#else
-          SHERPA_ONNX_LOGE("After normalizing: %s", text.c_str());
-#endif
-        }
-      }
-    }
-
-    std::vector<TokenIDs> token_ids =
-        frontend_->ConvertTextToTokenIds(text, meta_data.voice);
-
-    if (token_ids.empty() ||
-        (token_ids.size() == 1 && token_ids[0].tokens.empty())) {
-      SHERPA_ONNX_LOGE("Failed to convert %s to token IDs", text.c_str());
-      return {};
-    }
-
-    std::vector<std::vector<int64_t>> x;
-    std::vector<std::vector<int64_t>> tones;
-
-    x.reserve(token_ids.size());
-
-    for (auto &i : token_ids) {
-      x.push_back(std::move(i.tokens));
-    }
-
-    if (!token_ids[0].tones.empty()) {
-      tones.reserve(token_ids.size());
-      for (auto &i : token_ids) {
-        tones.push_back(std::move(i.tones));
-      }
-    }
-
-    // TODO(fangjun): add blank inside the frontend, not here
-    if (meta_data.add_blank && config_.model.vits.data_dir.empty() &&
-        meta_data.frontend != "characters") {
-      for (auto &k : x) {
-        k = AddBlank(k);
-      }
-
-      for (auto &k : tones) {
-        k = AddBlank(k);
-      }
-    }
-
-    int32_t x_size = static_cast<int32_t>(x.size());
-
-    if (config_.max_num_sentences <= 0 || x_size <= config_.max_num_sentences) {
-      auto ans = Process(x, tones, sid, speed);
-      if (callback) {
-        callback(ans.samples.data(), ans.samples.size(), 1.0);
-      }
-      return ans;
-    }
-
-    // the input text is too long, we process sentences within it in batches
-    // to avoid OOM. Batch size is config_.max_num_sentences
-    std::vector<std::vector<int64_t>> batch_x;
-    std::vector<std::vector<int64_t>> batch_tones;
-
-    int32_t batch_size = config_.max_num_sentences;
-    batch_x.reserve(config_.max_num_sentences);
-    batch_tones.reserve(config_.max_num_sentences);
-    int32_t num_batches = x_size / batch_size;
-
-    if (config_.model.debug) {
-#if __OHOS__
-      SHERPA_ONNX_LOGE(
-          "Text is too long. Split it into %{public}d batches. batch size: "
-          "%{public}d. Number of sentences: %{public}d",
-          num_batches, batch_size, x_size);
-#else
-      SHERPA_ONNX_LOGE(
-          "Text is too long. Split it into %d batches. batch size: %d. Number "
-          "of sentences: %d",
-          num_batches, batch_size, x_size);
-#endif
-    }
-
-    GeneratedAudio ans;
-
-    int32_t should_continue = 1;
-
-    int32_t k = 0;
-
-    for (int32_t b = 0; b != num_batches && should_continue; ++b) {
-      batch_x.clear();
-      batch_tones.clear();
-      for (int32_t i = 0; i != batch_size; ++i, ++k) {
-        batch_x.push_back(std::move(x[k]));
-
-        if (!tones.empty()) {
-          batch_tones.push_back(std::move(tones[k]));
-        }
-      }
-
-      auto audio = Process(batch_x, batch_tones, sid, speed);
-      ans.sample_rate = audio.sample_rate;
-      ans.samples.insert(ans.samples.end(), audio.samples.begin(),
-                         audio.samples.end());
-      if (callback) {
-        should_continue = callback(audio.samples.data(), audio.samples.size(),
-                                   (b + 1) * 1.0 / num_batches);
-        // Caution(fangjun): audio is freed when the callback returns, so users
-        // should copy the data if they want to access the data after
-        // the callback returns to avoid segmentation fault.
-      }
-    }
-
-    batch_x.clear();
-    batch_tones.clear();
-    while (k < static_cast<int32_t>(x.size()) && should_continue) {
-      batch_x.push_back(std::move(x[k]));
-      if (!tones.empty()) {
-        batch_tones.push_back(std::move(tones[k]));
-      }
-
-      ++k;
-    }
-
-    if (!batch_x.empty()) {
-      auto audio = Process(batch_x, batch_tones, sid, speed);
-      ans.sample_rate = audio.sample_rate;
-      ans.samples.insert(ans.samples.end(), audio.samples.begin(),
-                         audio.samples.end());
-      if (callback) {
-        callback(audio.samples.data(), audio.samples.size(), 1.0);
-        // Caution(fangjun): audio is freed when the callback returns, so users
-        // should copy the data if they want to access the data after
-        // the callback returns to avoid segmentation fault.
-      }
-    }
-
-    return ans;
-  }
-
- private:
-  template <typename Manager>
-  void InitFrontend(Manager *mgr) {
-    const auto &meta_data = model_->GetMetaData();
-
-    if (meta_data.frontend == "characters") {
-      frontend_ = std::make_unique<OfflineTtsCharacterFrontend>(
-          mgr, config_.model.vits.tokens, meta_data);
-    } else if (meta_data.jieba && !config_.model.vits.dict_dir.empty() &&
-               meta_data.is_melo_tts) {
-      frontend_ = std::make_unique<MeloTtsLexicon>(
-          mgr, config_.model.vits.lexicon, config_.model.vits.tokens,
-          config_.model.vits.dict_dir, model_->GetMetaData(),
-          config_.model.debug);
-    } else if (meta_data.is_melo_tts && meta_data.language == "English") {
-      frontend_ = std::make_unique<MeloTtsLexicon>(
-          mgr, config_.model.vits.lexicon, config_.model.vits.tokens,
-          model_->GetMetaData(), config_.model.debug);
-    } else if ((meta_data.is_piper || meta_data.is_coqui ||
-                meta_data.is_icefall) &&
-               !config_.model.vits.data_dir.empty()) {
-      frontend_ = std::make_unique<PiperPhonemizeLexicon>(
-          mgr, config_.model.vits.tokens, config_.model.vits.data_dir,
-          meta_data);
-    } else {
-      if (config_.model.vits.lexicon.empty()) {
-        SHERPA_ONNX_LOGE(
-            "Not a model using characters as modeling unit. Please provide "
-            "--vits-lexicon if you leave --vits-data-dir empty");
-        exit(-1);
-      }
-
-      frontend_ = std::make_unique<Lexicon>(
-          mgr, config_.model.vits.lexicon, config_.model.vits.tokens,
-          meta_data.punctuations, meta_data.language, config_.model.debug);
-    }
-  }
-
-  void InitFrontend() {
-    const auto &meta_data = model_->GetMetaData();
-
-    if (meta_data.jieba && config_.model.vits.dict_dir.empty()) {
-      SHERPA_ONNX_LOGE(
-          "Please provide --vits-dict-dir for Chinese TTS models using jieba");
-      exit(-1);
-    }
-
-    if (!meta_data.jieba && !config_.model.vits.dict_dir.empty()) {
-      SHERPA_ONNX_LOGE(
-          "Current model is not using jieba but you provided --vits-dict-dir");
-      exit(-1);
-    }
-
-    if (meta_data.frontend == "characters") {
-      frontend_ = std::make_unique<OfflineTtsCharacterFrontend>(
-          config_.model.vits.tokens, meta_data);
-    } else if (meta_data.jieba && !config_.model.vits.dict_dir.empty() &&
-               meta_data.is_melo_tts) {
-      frontend_ = std::make_unique<MeloTtsLexicon>(
-          config_.model.vits.lexicon, config_.model.vits.tokens,
-          config_.model.vits.dict_dir, model_->GetMetaData(),
-          config_.model.debug);
-    } else if (meta_data.is_melo_tts && meta_data.language == "English") {
-      frontend_ = std::make_unique<MeloTtsLexicon>(
-          config_.model.vits.lexicon, config_.model.vits.tokens,
-          model_->GetMetaData(), config_.model.debug);
-    } else if (meta_data.jieba && !config_.model.vits.dict_dir.empty()) {
-      frontend_ = std::make_unique<JiebaLexicon>(
-          config_.model.vits.lexicon, config_.model.vits.tokens,
-          config_.model.vits.dict_dir, model_->GetMetaData(),
-          config_.model.debug);
-    } else if ((meta_data.is_piper || meta_data.is_coqui ||
-                meta_data.is_icefall) &&
-               !config_.model.vits.data_dir.empty()) {
-      frontend_ = std::make_unique<PiperPhonemizeLexicon>(
-          config_.model.vits.tokens, config_.model.vits.data_dir,
-          model_->GetMetaData());
-    } else {
-      if (config_.model.vits.lexicon.empty()) {
-        SHERPA_ONNX_LOGE(
-            "Not a model using characters as modeling unit. Please provide "
-            "--vits-lexicon if you leave --vits-data-dir empty");
-        exit(-1);
-      }
-      frontend_ = std::make_unique<Lexicon>(
-          config_.model.vits.lexicon, config_.model.vits.tokens,
-          meta_data.punctuations, meta_data.language, config_.model.debug);
-    }
-  }
-
-  std::vector<int64_t> AddBlank(const std::vector<int64_t> &x) const {
-    // we assume the blank ID is 0
-    std::vector<int64_t> buffer(x.size() * 2 + 1);
-    int32_t i = 1;
-    for (auto k : x) {
-      buffer[i] = k;
-      i += 2;
-    }
-    return buffer;
-  }
-
-  GeneratedAudio Process(const std::vector<std::vector<int64_t>> &tokens,
-                         const std::vector<std::vector<int64_t>> &tones,
-                         int32_t sid, float speed) const {
-    int32_t num_tokens = 0;
-    for (const auto &k : tokens) {
-      num_tokens += k.size();
-    }
-
-    std::vector<int64_t> x;
-    x.reserve(num_tokens);
-    for (const auto &k : tokens) {
-      x.insert(x.end(), k.begin(), k.end());
-    }
-
-    std::vector<int64_t> tone_list;
-    if (!tones.empty()) {
-      tone_list.reserve(num_tokens);
-      for (const auto &k : tones) {
-        tone_list.insert(tone_list.end(), k.begin(), k.end());
-      }
-    }
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    std::array<int64_t, 2> x_shape = {1, static_cast<int32_t>(x.size())};
-    Ort::Value x_tensor = Ort::Value::CreateTensor(
-        memory_info, x.data(), x.size(), x_shape.data(), x_shape.size());
-
-    Ort::Value tones_tensor{nullptr};
-    if (!tones.empty()) {
-      tones_tensor = Ort::Value::CreateTensor(memory_info, tone_list.data(),
-                                              tone_list.size(), x_shape.data(),
-                                              x_shape.size());
-    }
-
-    Ort::Value audio{nullptr};
-    if (tones.empty()) {
-      audio = model_->Run(std::move(x_tensor), sid, speed);
-    } else {
-      audio =
-          model_->Run(std::move(x_tensor), std::move(tones_tensor), sid, speed);
-    }
-
-    std::vector<int64_t> audio_shape =
-        audio.GetTensorTypeAndShapeInfo().GetShape();
-
-    int64_t total = 1;
-    // The output shape may be (1, 1, total) or (1, total) or (total,)
-    for (auto i : audio_shape) {
-      total *= i;
-    }
-
-    const float *p = audio.GetTensorData<float>();
-
-    GeneratedAudio ans;
-    ans.sample_rate = model_->GetMetaData().sample_rate;
-    ans.samples = std::vector<float>(p, p + total);
-    return ans;
-  }
-
- private:
-  OfflineTtsConfig config_;
-  std::unique_ptr<OfflineTtsVitsModel> model_;
-  std::vector<std::unique_ptr<kaldifst::TextNormalizer>> tn_list_;
-  std::unique_ptr<OfflineTtsFrontend> frontend_;
-};
-
-}  // namespace sherpa_onnx
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TTS_VITS_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-model-config.h
deleted file mode 100644
index 09f00b8b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-model-config.h
+++ /dev/null
@@ -1,59 +0,0 @@
-// sherpa-onnx/csrc/offline-tts-vits-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TTS_VITS_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TTS_VITS_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineTtsVitsModelConfig {
-  std::string model;
-  std::string lexicon;
-  std::string tokens;
-
-  // If data_dir is given, lexicon is ignored
-  // data_dir is for piper-phonemize, which uses espeak-ng
-  std::string data_dir;
-
-  // Used for Chinese TTS models using jieba
-  std::string dict_dir;
-
-  float noise_scale = 0.667;
-  float noise_scale_w = 0.8;
-  float length_scale = 1;
-
-  // used only for multi-speaker models, e.g, vctk speech dataset.
-  // Not applicable for single-speaker models, e.g., ljspeech dataset
-
-  OfflineTtsVitsModelConfig() = default;
-
-  OfflineTtsVitsModelConfig(const std::string &model,
-                            const std::string &lexicon,
-                            const std::string &tokens,
-                            const std::string &data_dir,
-                            const std::string &dict_dir,
-                            float noise_scale = 0.667,
-                            float noise_scale_w = 0.8, float length_scale = 1)
-      : model(model),
-        lexicon(lexicon),
-        tokens(tokens),
-        data_dir(data_dir),
-        dict_dir(dict_dir),
-        noise_scale(noise_scale),
-        noise_scale_w(noise_scale_w),
-        length_scale(length_scale) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TTS_VITS_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-model-metadata.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-model-metadata.h
deleted file mode 100644
index 5ce00d74..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-model-metadata.h
+++ /dev/null
@@ -1,49 +0,0 @@
-// sherpa-onnx/csrc/offline-tts-vits-model-metadata.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TTS_VITS_MODEL_METADATA_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TTS_VITS_MODEL_METADATA_H_
-
-#include <cstdint>
-#include <string>
-
-namespace sherpa_onnx {
-
-// If you are not sure what each field means, please
-// have a look of the Python file in the model directory that
-// you have downloaded.
-struct OfflineTtsVitsModelMetaData {
-  int32_t sample_rate = 0;
-  int32_t add_blank = 0;
-  int32_t num_speakers = 0;
-
-  bool is_piper = false;
-  bool is_coqui = false;
-  bool is_icefall = false;
-  bool is_melo_tts = false;
-
-  // for Chinese TTS models from
-  // https://github.com/Plachtaa/VITS-fast-fine-tuning
-  int32_t jieba = 0;
-
-  // the following options are for models from coqui-ai/TTS
-  int32_t blank_id = 0;
-  int32_t bos_id = 0;
-  int32_t eos_id = 0;
-  int32_t use_eos_bos = 0;
-  int32_t pad_id = 0;
-
-  // for melo tts
-  int32_t speaker_id = 0;
-  int32_t version = 0;
-
-  std::string punctuations;
-  std::string language;
-  std::string voice;
-  std::string frontend;  // characters
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TTS_VITS_MODEL_METADATA_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-model.h
deleted file mode 100644
index a880934e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts-vits-model.h
+++ /dev/null
@@ -1,51 +0,0 @@
-// sherpa-onnx/csrc/offline-tts-vits-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TTS_VITS_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TTS_VITS_MODEL_H_
-
-#include <memory>
-#include <string>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-tts-model-config.h"
-#include "sherpa-onnx/csrc/offline-tts-vits-model-metadata.h"
-
-namespace sherpa_onnx {
-
-class OfflineTtsVitsModel {
- public:
-  ~OfflineTtsVitsModel();
-
-  explicit OfflineTtsVitsModel(const OfflineTtsModelConfig &config);
-
-  template <typename Manager>
-  OfflineTtsVitsModel(Manager *mgr, const OfflineTtsModelConfig &config);
-
-  /** Run the model.
-   *
-   * @param x A int64 tensor of shape (1, num_tokens)
-  // @param sid Speaker ID. Used only for multi-speaker models, e.g., models
-  //            trained using the VCTK dataset. It is not used for
-  //            single-speaker models, e.g., models trained using the ljspeech
-  //            dataset.
-   * @return Return a float32 tensor containing audio samples. You can flatten
-   *         it to a 1-D tensor.
-   */
-  Ort::Value Run(Ort::Value x, int64_t sid = 0, float speed = 1.0);
-
-  // This is for MeloTTS
-  Ort::Value Run(Ort::Value x, Ort::Value tones, int64_t sid = 0,
-                 float speed = 1.0);
-
-  const OfflineTtsVitsModelMetaData &GetMetaData() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TTS_VITS_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts.h
deleted file mode 100644
index 8399443c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-tts.h
+++ /dev/null
@@ -1,100 +0,0 @@
-// sherpa-onnx/csrc/offline-tts.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_TTS_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_TTS_H_
-
-#include <cstdint>
-#include <functional>
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-tts-model-config.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineTtsConfig {
-  OfflineTtsModelConfig model;
-  // If not empty, it contains a list of rule FST filenames.
-  // Filenames are separated by a comma.
-  // Example value: rule1.fst,rule2,fst,rule3.fst
-  //
-  // If there are multiple rules, they are applied from left to right.
-  std::string rule_fsts;
-
-  // If there are multiple FST archives, they are applied from left to right.
-  std::string rule_fars;
-
-  // Maximum number of sentences that we process at a time.
-  // This is to avoid OOM for very long input text.
-  // If you set it to -1, then we process all sentences in a single batch.
-  int32_t max_num_sentences = 2;
-
-  OfflineTtsConfig() = default;
-  OfflineTtsConfig(const OfflineTtsModelConfig &model,
-                   const std::string &rule_fsts, const std::string &rule_fars,
-                   int32_t max_num_sentences)
-      : model(model),
-        rule_fsts(rule_fsts),
-        rule_fars(rule_fars),
-        max_num_sentences(max_num_sentences) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-struct GeneratedAudio {
-  std::vector<float> samples;
-  int32_t sample_rate;
-};
-
-class OfflineTtsImpl;
-
-// If the callback returns 0, then it stop generating
-// if the callback returns 1, then it keeps generating
-using GeneratedAudioCallback = std::function<int32_t(
-    const float * /*samples*/, int32_t /*n*/, float /*progress*/)>;
-
-class OfflineTts {
- public:
-  ~OfflineTts();
-  explicit OfflineTts(const OfflineTtsConfig &config);
-
-  template <typename Manager>
-  OfflineTts(Manager *mgr, const OfflineTtsConfig &config);
-
-  // @param text A string containing words separated by spaces
-  // @param sid Speaker ID. Used only for multi-speaker models, e.g., models
-  //            trained using the VCTK dataset. It is not used for
-  //            single-speaker models, e.g., models trained using the ljspeech
-  //            dataset.
-  // @param speed The speed for the generated speech. E.g., 2 means 2x faster.
-  // @param callback If not NULL, it is called whenever config.max_num_sentences
-  //                 sentences have been processed. Note that the passed
-  //                 pointer `samples` for the callback might be invalidated
-  //                 after the callback is returned, so the caller should not
-  //                 keep a reference to it. The caller can copy the data if
-  //                 he/she wants to access the samples after the callback
-  //                 returns. The callback is called in the current thread.
-  GeneratedAudio Generate(const std::string &text, int64_t sid = 0,
-                          float speed = 1.0,
-                          GeneratedAudioCallback callback = nullptr) const;
-
-  // Return the sample rate of the generated audio
-  int32_t SampleRate() const;
-
-  // Number of supported speakers.
-  // If it supports only a single speaker, then it return 0 or 1.
-  int32_t NumSpeakers() const;
-
- private:
-  std::unique_ptr<OfflineTtsImpl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_TTS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-websocket-server-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-websocket-server-impl.h
deleted file mode 100644
index 9ea88d53..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-websocket-server-impl.h
+++ /dev/null
@@ -1,205 +0,0 @@
-// sherpa-onnx/csrc/offline-websocket-server-impl.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_WEBSOCKET_SERVER_IMPL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_WEBSOCKET_SERVER_IMPL_H_
-
-#include <deque>
-#include <fstream>
-#include <map>
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-recognizer.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-#include "sherpa-onnx/csrc/tee-stream.h"
-#include "websocketpp/config/asio_no_tls.hpp"  // TODO(fangjun): support TLS
-#include "websocketpp/server.hpp"
-
-using server = websocketpp::server<websocketpp::config::asio>;
-using connection_hdl = websocketpp::connection_hdl;
-
-namespace sherpa_onnx {
-
-/** Communication protocol
- *
- * The client sends a byte stream to the server. The first 4 bytes in little
- * endian indicates the sample rate of the audio data that the client will send.
- * The next 4 bytes in little endian indicates the total samples in bytes the
- * client will send. The remaining bytes represent audio samples. Each audio
- * sample is a float occupying 4 bytes and is normalized into the range
- * [-1, 1].
- *
- * The byte stream can be broken into arbitrary number of messages.
- * We require that the first message has to be at least 8 bytes so that
- * we can get `sample_rate` and `expected_byte_size` from the first message.
- */
-struct ConnectionData {
-  // Sample rate of the audio samples the client
-  int32_t sample_rate;
-
-  // Number of expected bytes sent from the client
-  int32_t expected_byte_size = 0;
-
-  // Number of bytes received so far
-  int32_t cur = 0;
-
-  // It saves the received samples from the client.
-  // We will **reinterpret_cast** it to float.
-  // We expect that data.size() == expected_byte_size
-  std::vector<int8_t> data;
-
-  void Clear() {
-    sample_rate = 0;
-    expected_byte_size = 0;
-    cur = 0;
-    data.clear();
-  }
-};
-
-using ConnectionDataPtr = std::shared_ptr<ConnectionData>;
-
-struct OfflineWebsocketDecoderConfig {
-  OfflineRecognizerConfig recognizer_config;
-
-  int32_t max_batch_size = 5;
-
-  float max_utterance_length = 300;  // seconds
-
-  void Register(ParseOptions *po);
-  void Validate() const;
-};
-
-class OfflineWebsocketServer;
-
-class OfflineWebsocketDecoder {
- public:
-  /**
-   * @param config Configuration for the decoder.
-   * @param server **Borrowed** from outside.
-   */
-  explicit OfflineWebsocketDecoder(OfflineWebsocketServer *server);
-
-  /** Insert received data to the queue for decoding.
-   *
-   * @param hdl A handle to the connection. We can use it to send the result
-   *            back to the client once it finishes decoding.
-   * @param d  The received data
-   */
-  void Push(connection_hdl hdl, ConnectionDataPtr d);
-
-  /** It is called by one of the work thread.
-   */
-  void Decode();
-
-  const OfflineWebsocketDecoderConfig &GetConfig() const { return config_; }
-
- private:
-  OfflineWebsocketDecoderConfig config_;
-
-  /** When we have received all the data from the client, we put it into
-   * this queue; the worker threads will get items from this queue for
-   * decoding.
-   *
-   * Number of items to take from this queue is determined by
-   * `--max-batch-size`. If there are not enough items in the queue, we won't
-   * wait and take whatever we have for decoding.
-   */
-  std::mutex mutex_;
-  std::deque<std::pair<connection_hdl, ConnectionDataPtr>> streams_;
-
-  OfflineWebsocketServer *server_;  // Not owned
-  OfflineRecognizer recognizer_;
-};
-
-struct OfflineWebsocketServerConfig {
-  OfflineWebsocketDecoderConfig decoder_config;
-  std::string log_file = "./log.txt";
-
-  void Register(ParseOptions *po);
-  void Validate() const;
-};
-
-class OfflineWebsocketServer {
- public:
-  OfflineWebsocketServer(asio::io_context &io_conn,  // NOLINT
-                         asio::io_context &io_work,  // NOLINT
-                         const OfflineWebsocketServerConfig &config);
-
-  asio::io_context &GetConnectionContext() { return io_conn_; }
-  server &GetServer() { return server_; }
-
-  void Run(uint16_t port);
-
-  const OfflineWebsocketServerConfig &GetConfig() const { return config_; }
-
- private:
-  void SetupLog();
-
-  // When a websocket client is connected, it will invoke this method
-  // (Not for HTTP)
-  void OnOpen(connection_hdl hdl);
-
-  // When a websocket client is disconnected, it will invoke this method
-  void OnClose(connection_hdl hdl);
-
-  // When a message is received from a websocket client, this method will
-  // be invoked.
-  //
-  // The protocol between the client and the server is as follows:
-  //
-  // (1) The client connects to the server
-  // (2) The client starts to send binary byte stream to the server.
-  //     The byte stream can be broken into multiple messages or it can
-  //     be put into a single message.
-  //     The first message has to contain at least 8 bytes. The first
-  //     4 bytes in little endian contains a int32_t indicating the
-  //     sampling rate. The next 4 bytes in little endian contains a int32_t
-  //     indicating total number of bytes of samples the client will send.
-  //     We assume each sample is a float containing 4 bytes and has been
-  //     normalized to the range [-1, 1].
-  // (4) When the server receives all the samples from the client, it will
-  //     start to decode them. Once decoded, the server sends a text message
-  //     to the client containing the decoded results
-  // (5) After receiving the decoded results from the server, if the client has
-  //     another audio file to send, it repeats (2), (3), (4)
-  // (6) If the client has no more audio files to decode, the client sends a
-  //     text message containing "Done" to the server and closes the connection
-  // (7) The server receives a text message "Done" and closes the connection
-  //
-  // Note:
-  //  (a) All models in icefall use features extracted from audio samples
-  //      normalized to the range [-1, 1]. Please send normalized audio samples
-  //      if you use models from icefall.
-  //  (b) Only sound files with a single channel is supported
-  //  (c) Only audio samples are sent. For instance, if we want to decode
-  //      a WAVE file, the RIFF header of the WAVE is not sent.
-  void OnMessage(connection_hdl hdl, server::message_ptr msg);
-
-  // Close a websocket connection with given code and reason
-  void Close(connection_hdl hdl, websocketpp::close::status::value code,
-             const std::string &reason);
-
- private:
-  asio::io_context &io_conn_;
-  asio::io_context &io_work_;
-  server server_;
-
-  std::map<connection_hdl, ConnectionDataPtr, std::owner_less<connection_hdl>>
-      connections_;
-  std::mutex mutex_;
-
-  OfflineWebsocketServerConfig config_;
-
-  std::ofstream log_;
-  TeeStream tee_;
-
-  OfflineWebsocketDecoder decoder_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_WEBSOCKET_SERVER_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-wenet-ctc-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-wenet-ctc-model-config.h
deleted file mode 100644
index 4a9b30b8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-wenet-ctc-model-config.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// sherpa-onnx/csrc/offline-wenet-ctc-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_WENET_CTC_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_WENET_CTC_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineWenetCtcModelConfig {
-  std::string model;
-
-  OfflineWenetCtcModelConfig() = default;
-  explicit OfflineWenetCtcModelConfig(const std::string &model)
-      : model(model) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_WENET_CTC_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-wenet-ctc-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-wenet-ctc-model.h
deleted file mode 100644
index ba5f2124..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-wenet-ctc-model.h
+++ /dev/null
@@ -1,73 +0,0 @@
-// sherpa-onnx/csrc/offline-wenet-ctc-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_WENET_CTC_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_WENET_CTC_MODEL_H_
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-ctc-model.h"
-#include "sherpa-onnx/csrc/offline-model-config.h"
-
-namespace sherpa_onnx {
-
-/** This class implements the CTC model from WeNet.
- *
- * See
- * https://github.com/k2-fsa/sherpa-onnx/blob/master/scripts/wenet/export-onnx.py
- * https://github.com/k2-fsa/sherpa-onnx/blob/master/scripts/wenet/test-onnx.py
- * https://github.com/k2-fsa/sherpa-onnx/blob/master/scripts/wenet/run.sh
- *
- */
-class OfflineWenetCtcModel : public OfflineCtcModel {
- public:
-  explicit OfflineWenetCtcModel(const OfflineModelConfig &config);
-
-  template <typename Manager>
-  OfflineWenetCtcModel(Manager *mgr, const OfflineModelConfig &config);
-
-  ~OfflineWenetCtcModel() override;
-
-  /** Run the forward method of the model.
-   *
-   * @param features  A tensor of shape (N, T, C).
-   * @param features_length  A 1-D tensor of shape (N,) containing number of
-   *                         valid frames in `features` before padding.
-   *                         Its dtype is int64_t.
-   *
-   * @return Return a vector containing:
-   *  - log_probs: A 3-D tensor of shape (N, T', vocab_size).
-   *  - log_probs_length A 1-D tensor of shape (N,). Its dtype is int64_t
-   */
-  std::vector<Ort::Value> Forward(Ort::Value features,
-                                  Ort::Value features_length) override;
-
-  /** Return the vocabulary size of the model
-   */
-  int32_t VocabSize() const override;
-
-  /** SubsamplingFactor of the model
-   *
-   * For Citrinet, the subsampling factor is usually 4.
-   * For Conformer CTC, the subsampling factor is usually 8.
-   */
-  int32_t SubsamplingFactor() const override;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const override;
-
-  // WeNet CTC models do not support batch size > 1
-  bool SupportBatchProcessing() const override { return false; }
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_WENET_CTC_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-decoder.h
deleted file mode 100644
index 9cb5088e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-decoder.h
+++ /dev/null
@@ -1,43 +0,0 @@
-// sherpa-onnx/csrc/offline-whisper-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_WHISPER_DECODER_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_WHISPER_DECODER_H_
-
-#include <string>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-whisper-model-config.h"
-
-namespace sherpa_onnx {
-
-struct OfflineWhisperDecoderResult {
-  /// The decoded token IDs
-  std::vector<int32_t> tokens;
-  std::string lang;
-};
-
-class OfflineWhisperDecoder {
- public:
-  virtual ~OfflineWhisperDecoder() = default;
-
-  /** Run beam search given the output from the whisper encoder model.
-   *
-   * @param n_layer_cross_k       A 4-D tensor of shape
-   *                              (n_text_layer, N, n_audio_ctx, n_text_state).
-   * @param n_layer_cross_v       A 4-D tensor of shape
-   *                              (n_text_layer, N, n_audio_ctx, n_text_state).
-   *
-   * @return Return a vector of size `N` containing the decoded results.
-   */
-  virtual std::vector<OfflineWhisperDecoderResult> Decode(
-      Ort::Value n_layer_cross_k, Ort::Value n_layer_cross_v) = 0;
-
-  virtual void SetConfig(const OfflineWhisperModelConfig &config) = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_WHISPER_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.h
deleted file mode 100644
index 9692d90d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.h
+++ /dev/null
@@ -1,33 +0,0 @@
-// sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_WHISPER_GREEDY_SEARCH_DECODER_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_WHISPER_GREEDY_SEARCH_DECODER_H_
-
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-whisper-decoder.h"
-#include "sherpa-onnx/csrc/offline-whisper-model.h"
-
-namespace sherpa_onnx {
-
-class OfflineWhisperGreedySearchDecoder : public OfflineWhisperDecoder {
- public:
-  OfflineWhisperGreedySearchDecoder(const OfflineWhisperModelConfig &config,
-                                    OfflineWhisperModel *model)
-      : config_(config), model_(model) {}
-
-  std::vector<OfflineWhisperDecoderResult> Decode(Ort::Value cross_k,
-                                                  Ort::Value cross_v) override;
-
-  void SetConfig(const OfflineWhisperModelConfig &config) override;
-
- private:
-  OfflineWhisperModelConfig config_;
-  OfflineWhisperModel *model_;  // not owned
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_WHISPER_GREEDY_SEARCH_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-model-config.h
deleted file mode 100644
index a612a46d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-model-config.h
+++ /dev/null
@@ -1,60 +0,0 @@
-// sherpa-onnx/csrc/offline-whisper-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_WHISPER_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_WHISPER_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineWhisperModelConfig {
-  std::string encoder;
-  std::string decoder;
-
-  // Available languages can be found at
-  // https://github.com/openai/whisper/blob/main/whisper/tokenizer.py#L10
-  //
-  // Note: For non-multilingual models, it supports only "en"
-  //
-  // If empty, we will infer it from the input audio file when
-  // the model is multilingual.
-  std::string language;
-
-  // Valid values are transcribe and translate
-  //
-  // Note: For non-multilingual models, it supports only "transcribe"
-  std::string task = "transcribe";
-
-  // Number of tail padding frames.
-  //
-  // Since we remove the 30-second constraint, we need to add some paddings
-  // at the end.
-  //
-  // Recommended values:
-  //   - 50 for English models
-  //   - 300 for multilingual models
-  int32_t tail_paddings = -1;
-
-  OfflineWhisperModelConfig() = default;
-  OfflineWhisperModelConfig(const std::string &encoder,
-                            const std::string &decoder,
-                            const std::string &language,
-                            const std::string &task, int32_t tail_paddings)
-      : encoder(encoder),
-        decoder(decoder),
-        language(language),
-        task(task),
-        tail_paddings(tail_paddings) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_WHISPER_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-model.h
deleted file mode 100644
index 60d76928..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-whisper-model.h
+++ /dev/null
@@ -1,115 +0,0 @@
-// sherpa-onnx/csrc/offline-whisper-model.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_WHISPER_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_WHISPER_MODEL_H_
-
-#include <memory>
-#include <string>
-#include <tuple>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-model-config.h"
-#include "sherpa-onnx/csrc/spoken-language-identification.h"
-
-namespace sherpa_onnx {
-
-class OfflineWhisperModel {
- public:
-  explicit OfflineWhisperModel(const OfflineModelConfig &config);
-
-  explicit OfflineWhisperModel(
-      const SpokenLanguageIdentificationConfig &config);
-
-  template <typename Manager>
-  OfflineWhisperModel(Manager *mgr, const OfflineModelConfig &config);
-
-  template <typename Manager>
-  OfflineWhisperModel(Manager *mgr,
-                      const SpokenLanguageIdentificationConfig &config);
-
-  ~OfflineWhisperModel();
-
-  /** Run the encoder model.
-   *
-   * @param features  A tensor of shape (N, C, T). It is changed in-place.
-   *                  C is 80 and T is 3000.
-   *
-   * @return Return a pair containing:
-   *  - n_layer_cross_k: A 4-D tensor of shape
-   *                     (n_text_layer, N, n_audio_ctx, n_text_state)
-   *  - n_layer_cross_v: A 4-D tensor of shape
-   *                     (n_text_layer, N, n_audio_ctx, n_text_state)
-   */
-  std::pair<Ort::Value, Ort::Value> ForwardEncoder(Ort::Value features) const;
-
-  /** Run the decoder model.
-   *
-   * @param tokens A int64 tensor of shape (N, num_words)
-   * @param n_layer_self_k_cache  A 4-D tensor of shape
-   *                              (n_text_layer, N, n_text_ctx, n_text_state).
-   * @param n_layer_self_v_cache  A 4-D tensor of shape
-   *                              (n_text_layer, N, n_text_ctx, n_text_state).
-   * @param n_layer_cross_k       A 4-D tensor of shape
-   *                              (n_text_layer, N, n_audio_ctx, n_text_state).
-   * @param n_layer_cross_v       A 4-D tensor of shape
-   *                              (n_text_layer, N, n_audio_ctx, n_text_state).
-   * @param offset A int64 tensor of shape (N,)
-   *
-   * @return Return a tuple containing 6 tensors:
-   *
-   *  - logits A 3-D tensor of shape (N, num_words, vocab_size)
-   *  - out_n_layer_self_k_cache Same shape as n_layer_self_k_cache
-   *  - out_n_layer_self_v_cache Same shape as n_layer_self_v_cache
-   *  - out_n_layer_cross_k Same as n_layer_cross_k
-   *  - out_n_layer_cross_v Same as n_layer_cross_v
-   *  - out_offset Same as offset
-   */
-  std::tuple<Ort::Value, Ort::Value, Ort::Value, Ort::Value, Ort::Value,
-             Ort::Value>
-  ForwardDecoder(Ort::Value tokens, Ort::Value n_layer_self_k_cache,
-                 Ort::Value n_layer_self_v_cache, Ort::Value n_layer_cross_k,
-                 Ort::Value n_layer_cross_v, Ort::Value offset) const;
-
-  int32_t DetectLanguage(Ort::Value &cross_k,   // NOLINT
-                         Ort::Value &cross_v);  // NOLINT
-
-  /** Return the initial self kv cache in a pair
-   *  - n_layer_self_k_cache A 4-D tensor of shape
-   *                         (n_text_layer, N, n_audio_ctx, n_text_state).
-   *  - n_layer_self_v_cache A 4-D tensor of shape
-   *                         (n_text_layer, N, n_audio_ctx, n_text_state).
-   */
-  std::pair<Ort::Value, Ort::Value> GetInitialSelfKVCache() const;
-  const std::vector<int64_t> &GetInitialTokens() const;
-  const std::vector<int32_t> &GetAllLanguageIDs() const;
-  const std::unordered_map<std::string, int32_t> &GetLang2ID() const;
-  const std::unordered_map<int32_t, std::string> &GetID2Lang() const;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const;
-
-  int32_t NoTimeStampsToken() const;
-  int32_t EOT() const;
-  int32_t SOT() const;
-  int32_t TextCtx() const;
-  int32_t VocabSize() const;
-  int32_t FeatureDim() const;
-  int32_t Translate() const;
-  bool IsMultiLingual() const;
-
-  static void NormalizeFeatures(float *features, int32_t num_frames,
-                                int32_t feat_dim);
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_WHISPER_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-audio-tagging-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-audio-tagging-model-config.h
deleted file mode 100644
index 4f60e832..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-audio-tagging-model-config.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// sherpa-onnx/csrc/offline-zipformer-audio-tagging-model-config.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_ZIPFORMER_AUDIO_TAGGING_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_ZIPFORMER_AUDIO_TAGGING_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OfflineZipformerAudioTaggingModelConfig {
-  std::string model;
-
-  OfflineZipformerAudioTaggingModelConfig() = default;
-
-  explicit OfflineZipformerAudioTaggingModelConfig(const std::string &model)
-      : model(model) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_ZIPFORMER_AUDIO_TAGGING_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-audio-tagging-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-audio-tagging-model.h
deleted file mode 100644
index 28282349..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-audio-tagging-model.h
+++ /dev/null
@@ -1,64 +0,0 @@
-// sherpa-onnx/csrc/offline-zipformer-audio-tagging-model.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_ZIPFORMER_AUDIO_TAGGING_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_ZIPFORMER_AUDIO_TAGGING_MODEL_H_
-#include <memory>
-#include <utility>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/audio-tagging-model-config.h"
-
-namespace sherpa_onnx {
-
-/** This class implements the zipformer CTC model of the librispeech recipe
- * from icefall.
- *
- * See
- * https://github.com/k2-fsa/icefall/blob/master/egs/audioset/AT/zipformer/export-onnx.py
- */
-class OfflineZipformerAudioTaggingModel {
- public:
-  explicit OfflineZipformerAudioTaggingModel(
-      const AudioTaggingModelConfig &config);
-
-#if __ANDROID_API__ >= 9
-  OfflineZipformerAudioTaggingModel(AAssetManager *mgr,
-                                    const AudioTaggingModelConfig &config);
-#endif
-
-  ~OfflineZipformerAudioTaggingModel();
-
-  /** Run the forward method of the model.
-   *
-   * @param features  A tensor of shape (N, T, C).
-   * @param features_length  A 1-D tensor of shape (N,) containing number of
-   *                         valid frames in `features` before padding.
-   *                         Its dtype is int64_t.
-   *
-   * @return Return a tensor
-   *  - probs: A 2-D tensor of shape (N, num_event_classes).
-   */
-  Ort::Value Forward(Ort::Value features, Ort::Value features_length) const;
-
-  /** Return the number of event classes of the model
-   */
-  int32_t NumEventClasses() const;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_ZIPFORMER_AUDIO_TAGGING_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-ctc-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-ctc-model-config.h
deleted file mode 100644
index 702575e7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-ctc-model-config.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// sherpa-onnx/csrc/offline-zipformer-ctc-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_ZIPFORMER_CTC_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_ZIPFORMER_CTC_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-// for
-// https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/zipformer/export-onnx-ctc.py
-struct OfflineZipformerCtcModelConfig {
-  std::string model;
-
-  OfflineZipformerCtcModelConfig() = default;
-
-  explicit OfflineZipformerCtcModelConfig(const std::string &model)
-      : model(model) {}
-
-  void Register(ParseOptions *po);
-
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_ZIPFORMER_CTC_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-ctc-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-ctc-model.h
deleted file mode 100644
index f76f96d5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/offline-zipformer-ctc-model.h
+++ /dev/null
@@ -1,62 +0,0 @@
-// sherpa-onnx/csrc/offline-zipformer-ctc-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_OFFLINE_ZIPFORMER_CTC_MODEL_H_
-#define SHERPA_ONNX_CSRC_OFFLINE_ZIPFORMER_CTC_MODEL_H_
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-ctc-model.h"
-#include "sherpa-onnx/csrc/offline-model-config.h"
-
-namespace sherpa_onnx {
-
-/** This class implements the zipformer CTC model of the librispeech recipe
- * from icefall.
- *
- * See
- * https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/zipformer/export-onnx-ctc.py
- */
-class OfflineZipformerCtcModel : public OfflineCtcModel {
- public:
-  explicit OfflineZipformerCtcModel(const OfflineModelConfig &config);
-
-  template <typename Manager>
-  OfflineZipformerCtcModel(Manager *mgr, const OfflineModelConfig &config);
-
-  ~OfflineZipformerCtcModel() override;
-
-  /** Run the forward method of the model.
-   *
-   * @param features  A tensor of shape (N, T, C).
-   * @param features_length  A 1-D tensor of shape (N,) containing number of
-   *                         valid frames in `features` before padding.
-   *                         Its dtype is int64_t.
-   *
-   * @return Return a vector containing:
-   *  - log_probs: A 3-D tensor of shape (N, T', vocab_size).
-   *  - log_probs_length A 1-D tensor of shape (N,). Its dtype is int64_t
-   */
-  std::vector<Ort::Value> Forward(Ort::Value features,
-                                  Ort::Value features_length) override;
-
-  /** Return the vocabulary size of the model
-   */
-  int32_t VocabSize() const override;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const override;
-
-  int32_t SubsamplingFactor() const override;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_OFFLINE_ZIPFORMER_CTC_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-cnn-bilstm-model-meta-data.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-cnn-bilstm-model-meta-data.h
deleted file mode 100644
index 4e8531f6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-cnn-bilstm-model-meta-data.h
+++ /dev/null
@@ -1,25 +0,0 @@
-// sherpa-onnx/csrc/online-cnn-bilstm-model-meta-data.h
-//
-// Copyright (c) 2024 Jian You (jianyou@cisco.com, Cisco Systems)
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_CNN_BILSTM_MODEL_META_DATA_H_
-#define SHERPA_ONNX_CSRC_ONLINE_CNN_BILSTM_MODEL_META_DATA_H_
-
-namespace sherpa_onnx {
-
-struct OnlineCNNBiLSTMModelMetaData {
-  int32_t comma_id = -1;
-  int32_t period_id = -1;
-  int32_t quest_id = -1;
-
-  int32_t upper_id = -1;
-  int32_t cap_id = -1;
-  int32_t mix_case_id = -1;
-
-  int32_t num_cases = -1;
-  int32_t num_punctuations = -1;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_CNN_BILSTM_MODEL_META_DATA_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-cnn-bilstm-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-cnn-bilstm-model.h
deleted file mode 100644
index 25886107..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-cnn-bilstm-model.h
+++ /dev/null
@@ -1,62 +0,0 @@
-// sherpa-onnx/csrc/online-cnn-bilstm-model.h
-//
-// Copyright (c) 2024 Jian You (jianyou@cisco.com, Cisco Systems)
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_CNN_BILSTM_MODEL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_CNN_BILSTM_MODEL_H_
-#include <memory>
-#include <utility>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/online-cnn-bilstm-model-meta-data.h"
-#include "sherpa-onnx/csrc/online-punctuation-model-config.h"
-
-namespace sherpa_onnx {
-
-/** This class implements
- *  https://github.com/frankyoujian/Edge-Punct-Casing/blob/main/onnx_decode_sentence.py
- */
-class OnlineCNNBiLSTMModel {
- public:
-  explicit OnlineCNNBiLSTMModel(const OnlinePunctuationModelConfig &config);
-
-#if __ANDROID_API__ >= 9
-  OnlineCNNBiLSTMModel(AAssetManager *mgr,
-                       const OnlinePunctuationModelConfig &config);
-#endif
-
-  ~OnlineCNNBiLSTMModel();
-
-  /** Run the forward method of the model.
-   *
-   * @param token_ids  A tensor of shape (N, T) of dtype int32.
-   * @param valid_ids  A tensor of shape (N, T) of dtype int32.
-   * @param label_lens A tensor of shape (N) of dtype int32.
-   *
-   * @return Return a pair of tensors
-   *  - case_logits:  A 2-D tensor of shape (T', num_cases).
-   *  - punct_logits: A 2-D tensor of shape (T', num_puncts).
-   */
-  std::pair<Ort::Value, Ort::Value> Forward(Ort::Value token_ids,
-                                            Ort::Value valid_ids,
-                                            Ort::Value label_lens) const;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const;
-
-  const OnlineCNNBiLSTMModelMetaData &GetModelMetadata() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_CNN_BILSTM_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-conformer-transducer-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-conformer-transducer-model.h
deleted file mode 100644
index 5c901b87..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-conformer-transducer-model.h
+++ /dev/null
@@ -1,100 +0,0 @@
-// sherpa-onnx/csrc/online-conformer-transducer-model.h
-//
-// Copyright (c) 2023 Jingzhao Ou (jingzhao.ou@gmail.com)
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_CONFORMER_TRANSDUCER_MODEL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_CONFORMER_TRANSDUCER_MODEL_H_
-
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/online-model-config.h"
-#include "sherpa-onnx/csrc/online-transducer-model.h"
-
-namespace sherpa_onnx {
-
-class OnlineConformerTransducerModel : public OnlineTransducerModel {
- public:
-  explicit OnlineConformerTransducerModel(const OnlineModelConfig &config);
-
-  template <typename Manager>
-  OnlineConformerTransducerModel(Manager *mgr, const OnlineModelConfig &config);
-
-  std::vector<Ort::Value> StackStates(
-      const std::vector<std::vector<Ort::Value>> &states) const override;
-
-  std::vector<std::vector<Ort::Value>> UnStackStates(
-      const std::vector<Ort::Value> &states) const override;
-
-  std::vector<Ort::Value> GetEncoderInitStates() override;
-
-  std::pair<Ort::Value, std::vector<Ort::Value>> RunEncoder(
-      Ort::Value features, std::vector<Ort::Value> states,
-      Ort::Value processed_frames) override;
-
-  Ort::Value RunDecoder(Ort::Value decoder_input) override;
-
-  Ort::Value RunJoiner(Ort::Value encoder_out, Ort::Value decoder_out) override;
-
-  int32_t ContextSize() const override { return context_size_; }
-
-  int32_t ChunkSize() const override { return T_; }
-
-  int32_t ChunkShift() const override { return decode_chunk_len_; }
-
-  int32_t VocabSize() const override { return vocab_size_; }
-  OrtAllocator *Allocator() override { return allocator_; }
-
- private:
-  void InitEncoder(void *model_data, size_t model_data_length);
-  void InitDecoder(void *model_data, size_t model_data_length);
-  void InitJoiner(void *model_data, size_t model_data_length);
-
- private:
-  Ort::Env env_;
-  Ort::SessionOptions sess_opts_;
-  Ort::AllocatorWithDefaultOptions allocator_;
-
-  std::unique_ptr<Ort::Session> encoder_sess_;
-  std::unique_ptr<Ort::Session> decoder_sess_;
-  std::unique_ptr<Ort::Session> joiner_sess_;
-
-  std::vector<std::string> encoder_input_names_;
-  std::vector<const char *> encoder_input_names_ptr_;
-
-  std::vector<std::string> encoder_output_names_;
-  std::vector<const char *> encoder_output_names_ptr_;
-
-  std::vector<std::string> decoder_input_names_;
-  std::vector<const char *> decoder_input_names_ptr_;
-
-  std::vector<std::string> decoder_output_names_;
-  std::vector<const char *> decoder_output_names_ptr_;
-
-  std::vector<std::string> joiner_input_names_;
-  std::vector<const char *> joiner_input_names_ptr_;
-
-  std::vector<std::string> joiner_output_names_;
-  std::vector<const char *> joiner_output_names_ptr_;
-
-  OnlineModelConfig config_;
-
-  int32_t num_encoder_layers_ = 0;
-  int32_t T_ = 0;
-  int32_t decode_chunk_len_ = 0;
-  int32_t cnn_module_kernel_ = 0;
-  int32_t context_size_ = 0;
-  int32_t left_context_ = 0;
-  // TODO(jingzhaoou): to retrieve from model medadata
-  int32_t right_context_ = 4;
-  int32_t encoder_dim_ = 0;
-  int32_t pad_length_ = 0;
-  int32_t vocab_size_ = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_CONFORMER_TRANSDUCER_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-decoder.h
deleted file mode 100644
index 65305e6a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-decoder.h
+++ /dev/null
@@ -1,63 +0,0 @@
-// sherpa-onnx/csrc/online-ctc-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_CTC_DECODER_H_
-#define SHERPA_ONNX_CSRC_ONLINE_CTC_DECODER_H_
-
-#include <memory>
-#include <vector>
-
-#include "kaldi-decoder/csrc/faster-decoder.h"
-#include "onnxruntime_cxx_api.h"  // NOLINT
-
-namespace sherpa_onnx {
-
-class OnlineStream;
-
-struct OnlineCtcDecoderResult {
-  /// Number of frames after subsampling we have decoded so far
-  int32_t frame_offset = 0;
-
-  /// The decoded token IDs
-  std::vector<int64_t> tokens;
-
-  /// The decoded word IDs
-  /// Note: tokens.size() is usually not equal to words.size()
-  /// words is empty for greedy search decoding.
-  /// it is not empty when an HLG graph or an HLG graph is used.
-  std::vector<int32_t> words;
-
-  /// timestamps[i] contains the output frame index where tokens[i] is decoded.
-  /// Note: The index is after subsampling
-  ///
-  /// tokens.size() == timestamps.size()
-  std::vector<int32_t> timestamps;
-
-  int32_t num_trailing_blanks = 0;
-};
-
-class OnlineCtcDecoder {
- public:
-  virtual ~OnlineCtcDecoder() = default;
-
-  /** Run streaming CTC decoding given the output from the encoder model.
-   *
-   * @param log_probs A 3-D tensor of shape (N, T, vocab_size) containing
-   *                  lob_probs.
-   *
-   * @param  results Input & Output parameters..
-   */
-  virtual void Decode(Ort::Value log_probs,
-                      std::vector<OnlineCtcDecoderResult> *results,
-                      OnlineStream **ss = nullptr, int32_t n = 0) = 0;
-
-  virtual std::unique_ptr<kaldi_decoder::FasterDecoder> CreateFasterDecoder()
-      const {
-    return nullptr;
-  }
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_CTC_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-fst-decoder-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-fst-decoder-config.h
deleted file mode 100644
index 6f9e5b15..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-fst-decoder-config.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// sherpa-onnx/csrc/online-ctc-fst-decoder-config.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_CTC_FST_DECODER_CONFIG_H_
-#define SHERPA_ONNX_CSRC_ONLINE_CTC_FST_DECODER_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OnlineCtcFstDecoderConfig {
-  // Path to H.fst, HL.fst or HLG.fst
-  std::string graph;
-  int32_t max_active = 3000;
-
-  OnlineCtcFstDecoderConfig() = default;
-
-  OnlineCtcFstDecoderConfig(const std::string &graph, int32_t max_active)
-      : graph(graph), max_active(max_active) {}
-
-  std::string ToString() const;
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_CTC_FST_DECODER_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-fst-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-fst-decoder.h
deleted file mode 100644
index 992276d6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-fst-decoder.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// sherpa-onnx/csrc/online-ctc-fst-decoder.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_CTC_FST_DECODER_H_
-#define SHERPA_ONNX_CSRC_ONLINE_CTC_FST_DECODER_H_
-
-#include <memory>
-#include <vector>
-
-#include "fst/fst.h"
-#include "sherpa-onnx/csrc/online-ctc-decoder.h"
-#include "sherpa-onnx/csrc/online-ctc-fst-decoder-config.h"
-
-namespace sherpa_onnx {
-
-class OnlineCtcFstDecoder : public OnlineCtcDecoder {
- public:
-  OnlineCtcFstDecoder(const OnlineCtcFstDecoderConfig &config,
-                      int32_t blank_id);
-
-  void Decode(Ort::Value log_probs,
-              std::vector<OnlineCtcDecoderResult> *results,
-              OnlineStream **ss = nullptr, int32_t n = 0) override;
-
-  std::unique_ptr<kaldi_decoder::FasterDecoder> CreateFasterDecoder()
-      const override;
-
- private:
-  OnlineCtcFstDecoderConfig config_;
-  kaldi_decoder::FasterDecoderOptions options_;
-
-  std::unique_ptr<fst::Fst<fst::StdArc>> fst_;
-  int32_t blank_id_ = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_CTC_FST_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-greedy-search-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-greedy-search-decoder.h
deleted file mode 100644
index 0af37593..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-greedy-search-decoder.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// sherpa-onnx/csrc/online-ctc-greedy-search-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_CTC_GREEDY_SEARCH_DECODER_H_
-#define SHERPA_ONNX_CSRC_ONLINE_CTC_GREEDY_SEARCH_DECODER_H_
-
-#include <vector>
-
-#include "sherpa-onnx/csrc/online-ctc-decoder.h"
-
-namespace sherpa_onnx {
-
-class OnlineCtcGreedySearchDecoder : public OnlineCtcDecoder {
- public:
-  explicit OnlineCtcGreedySearchDecoder(int32_t blank_id)
-      : blank_id_(blank_id) {}
-
-  void Decode(Ort::Value log_probs,
-              std::vector<OnlineCtcDecoderResult> *results,
-              OnlineStream **ss = nullptr, int32_t n = 0) override;
-
- private:
-  int32_t blank_id_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_CTC_GREEDY_SEARCH_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-model.h
deleted file mode 100644
index bd01bc54..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-ctc-model.h
+++ /dev/null
@@ -1,84 +0,0 @@
-// sherpa-onnx/csrc/online-ctc-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_CTC_MODEL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_CTC_MODEL_H_
-
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/online-model-config.h"
-
-namespace sherpa_onnx {
-
-class OnlineCtcModel {
- public:
-  virtual ~OnlineCtcModel() = default;
-
-  static std::unique_ptr<OnlineCtcModel> Create(
-      const OnlineModelConfig &config);
-
-  template <typename Manager>
-  static std::unique_ptr<OnlineCtcModel> Create(
-      Manager *mgr, const OnlineModelConfig &config);
-
-  // Return a list of tensors containing the initial states
-  virtual std::vector<Ort::Value> GetInitStates() const = 0;
-
-  /** Stack a list of individual states into a batch.
-   *
-   * It is the inverse operation of `UnStackStates`.
-   *
-   * @param states states[i] contains the state for the i-th utterance.
-   * @return Return a single value representing the batched state.
-   */
-  virtual std::vector<Ort::Value> StackStates(
-      std::vector<std::vector<Ort::Value>> states) const = 0;
-
-  /** Unstack a batch state into a list of individual states.
-   *
-   * It is the inverse operation of `StackStates`.
-   *
-   * @param states A batched state.
-   * @return ans[i] contains the state for the i-th utterance.
-   */
-  virtual std::vector<std::vector<Ort::Value>> UnStackStates(
-      std::vector<Ort::Value> states) const = 0;
-
-  /**
-   *
-   * @param x A 3-D tensor of shape (N, T, C). N has to be 1.
-   * @param states  It is from GetInitStates() or returned from this method.
-   *
-   * @return Return a list of tensors
-   *    - ans[0] contains log_probs, of shape (N, T, C)
-   *    - ans[1:] contains next_states
-   */
-  virtual std::vector<Ort::Value> Forward(
-      Ort::Value x, std::vector<Ort::Value> states) const = 0;
-
-  /** Return the vocabulary size of the model
-   */
-  virtual int32_t VocabSize() const = 0;
-
-  /** Return an allocator for allocating memory
-   */
-  virtual OrtAllocator *Allocator() const = 0;
-
-  // The model accepts this number of frames before subsampling as input
-  virtual int32_t ChunkLength() const = 0;
-
-  // Similar to frame_shift in feature extractor, after processing
-  // ChunkLength() frames, we advance by ChunkShift() frames
-  // before we process the next chunk.
-  virtual int32_t ChunkShift() const = 0;
-
-  // Return true if the model supports batch size > 1
-  virtual bool SupportBatchProcessing() const { return true; }
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_CTC_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-lm-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-lm-config.h
deleted file mode 100644
index 8d5b1670..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-lm-config.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// sherpa-onnx/csrc/online-lm-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_LM_CONFIG_H_
-#define SHERPA_ONNX_CSRC_ONLINE_LM_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OnlineLMConfig {
-  // path to the onnx model
-  std::string model;
-
-  // LM scale
-  float scale = 0.5;
-  int32_t lm_num_threads = 1;
-  std::string lm_provider = "cpu";
-  // enable shallow fusion
-  bool shallow_fusion = true;
-
-  OnlineLMConfig() = default;
-
-  OnlineLMConfig(const std::string &model, float scale, int32_t lm_num_threads,
-                 const std::string &lm_provider, bool shallow_fusion)
-      : model(model),
-        scale(scale),
-        lm_num_threads(lm_num_threads),
-        lm_provider(lm_provider),
-        shallow_fusion(shallow_fusion) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_LM_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-lm.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-lm.h
deleted file mode 100644
index ffd48c56..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-lm.h
+++ /dev/null
@@ -1,63 +0,0 @@
-// sherpa-onnx/csrc/online-lm.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_LM_H_
-#define SHERPA_ONNX_CSRC_ONLINE_LM_H_
-
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/hypothesis.h"
-#include "sherpa-onnx/csrc/online-lm-config.h"
-
-namespace sherpa_onnx {
-
-class OnlineLM {
- public:
-  virtual ~OnlineLM() = default;
-
-  static std::unique_ptr<OnlineLM> Create(const OnlineLMConfig &config);
-
-  // init states for classic rescore
-  virtual std::vector<Ort::Value> GetInitStates() = 0;
-
-  // init states for shallow fusion
-  virtual std::pair<Ort::Value, std::vector<Ort::Value>> GetInitStatesSF() = 0;
-
-   /** ScoreToken a batch of sentences (shallow fusion).
-   *
-   * @param x A 2-D tensor of shape (N, 1) with data type int64.
-   * @param states It contains the states for the LM model
-   * @return Return a pair containing
-   *          - log_prob of NN LM
-   *          - updated states
-   *
-   */
-  virtual std::pair<Ort::Value, std::vector<Ort::Value>> ScoreToken(
-      Ort::Value x, std::vector<Ort::Value> states) = 0;
-
-  /** This function updates hyp.lm_log_prob of hyps (classic rescore).
-   *
-   * @param scale LM score
-   * @param context_size Context size of the transducer decoder model
-   * @param hyps It is changed in-place.
-   *
-   */
-  virtual void ComputeLMScore(float scale, int32_t context_size,
-                      std::vector<Hypotheses> *hyps) = 0;
-
-  /** This function updates lm_log_prob and nn_lm_scores of hyp (shallow fusion).
-   *
-   * @param scale LM score
-   * @param hyps It is changed in-place.
-   *
-   */
-  virtual void ComputeLMScoreSF(float scale, Hypothesis *hyp) = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_LM_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-lstm-transducer-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-lstm-transducer-model.h
deleted file mode 100644
index 64ade36d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-lstm-transducer-model.h
+++ /dev/null
@@ -1,95 +0,0 @@
-// sherpa-onnx/csrc/online-lstm-transducer-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_LSTM_TRANSDUCER_MODEL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_LSTM_TRANSDUCER_MODEL_H_
-
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/online-model-config.h"
-#include "sherpa-onnx/csrc/online-transducer-model.h"
-
-namespace sherpa_onnx {
-
-class OnlineLstmTransducerModel : public OnlineTransducerModel {
- public:
-  explicit OnlineLstmTransducerModel(const OnlineModelConfig &config);
-
-  template <typename Manager>
-  OnlineLstmTransducerModel(Manager *mgr, const OnlineModelConfig &config);
-
-  std::vector<Ort::Value> StackStates(
-      const std::vector<std::vector<Ort::Value>> &states) const override;
-
-  std::vector<std::vector<Ort::Value>> UnStackStates(
-      const std::vector<Ort::Value> &states) const override;
-
-  std::vector<Ort::Value> GetEncoderInitStates() override;
-
-  std::pair<Ort::Value, std::vector<Ort::Value>> RunEncoder(
-      Ort::Value features, std::vector<Ort::Value> states,
-      Ort::Value processed_frames) override;
-
-  Ort::Value RunDecoder(Ort::Value decoder_input) override;
-
-  Ort::Value RunJoiner(Ort::Value encoder_out, Ort::Value decoder_out) override;
-
-  int32_t ContextSize() const override { return context_size_; }
-
-  int32_t ChunkSize() const override { return T_; }
-
-  int32_t ChunkShift() const override { return decode_chunk_len_; }
-
-  int32_t VocabSize() const override { return vocab_size_; }
-  OrtAllocator *Allocator() override { return allocator_; }
-
- private:
-  void InitEncoder(void *model_data, size_t model_data_length);
-  void InitDecoder(void *model_data, size_t model_data_length);
-  void InitJoiner(void *model_data, size_t model_data_length);
-
- private:
-  Ort::Env env_;
-  Ort::SessionOptions sess_opts_;
-  Ort::AllocatorWithDefaultOptions allocator_;
-
-  std::unique_ptr<Ort::Session> encoder_sess_;
-  std::unique_ptr<Ort::Session> decoder_sess_;
-  std::unique_ptr<Ort::Session> joiner_sess_;
-
-  std::vector<std::string> encoder_input_names_;
-  std::vector<const char *> encoder_input_names_ptr_;
-
-  std::vector<std::string> encoder_output_names_;
-  std::vector<const char *> encoder_output_names_ptr_;
-
-  std::vector<std::string> decoder_input_names_;
-  std::vector<const char *> decoder_input_names_ptr_;
-
-  std::vector<std::string> decoder_output_names_;
-  std::vector<const char *> decoder_output_names_ptr_;
-
-  std::vector<std::string> joiner_input_names_;
-  std::vector<const char *> joiner_input_names_ptr_;
-
-  std::vector<std::string> joiner_output_names_;
-  std::vector<const char *> joiner_output_names_ptr_;
-
-  OnlineModelConfig config_;
-
-  int32_t num_encoder_layers_ = 0;
-  int32_t T_ = 0;
-  int32_t decode_chunk_len_ = 0;
-  int32_t rnn_hidden_size_ = 0;
-  int32_t d_model_ = 0;
-  int32_t context_size_ = 0;
-  int32_t vocab_size_ = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_LSTM_TRANSDUCER_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-model-config.h
deleted file mode 100644
index a920512d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-model-config.h
+++ /dev/null
@@ -1,86 +0,0 @@
-// sherpa-onnx/csrc/online-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_ONLINE_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/online-nemo-ctc-model-config.h"
-#include "sherpa-onnx/csrc/online-paraformer-model-config.h"
-#include "sherpa-onnx/csrc/online-transducer-model-config.h"
-#include "sherpa-onnx/csrc/online-wenet-ctc-model-config.h"
-#include "sherpa-onnx/csrc/online-zipformer2-ctc-model-config.h"
-#include "sherpa-onnx/csrc/provider-config.h"
-
-namespace sherpa_onnx {
-
-struct OnlineModelConfig {
-  OnlineTransducerModelConfig transducer;
-  OnlineParaformerModelConfig paraformer;
-  OnlineWenetCtcModelConfig wenet_ctc;
-  OnlineZipformer2CtcModelConfig zipformer2_ctc;
-  OnlineNeMoCtcModelConfig nemo_ctc;
-  ProviderConfig provider_config;
-  std::string tokens;
-  int32_t num_threads = 1;
-  int32_t warm_up = 0;
-  bool debug = false;
-
-  // Valid values:
-  //  - conformer, conformer transducer from icefall
-  //  - lstm, lstm transducer from icefall
-  //  - zipformer, zipformer transducer from icefall
-  //  - zipformer2, zipformer2 transducer or CTC from icefall
-  //  - wenet_ctc, wenet CTC model
-  //  - nemo_ctc, NeMo CTC model
-  //
-  // All other values are invalid and lead to loading the model twice.
-  std::string model_type;
-
-  // Valid values:
-  //  - cjkchar
-  //  - bpe
-  //  - cjkchar+bpe
-  std::string modeling_unit = "cjkchar";
-  std::string bpe_vocab;
-
-  /// if tokens_buf is non-empty,
-  /// the tokens will be loaded from the buffer instead of from the
-  /// "tokens" file
-  std::string tokens_buf;
-
-  OnlineModelConfig() = default;
-  OnlineModelConfig(const OnlineTransducerModelConfig &transducer,
-                    const OnlineParaformerModelConfig &paraformer,
-                    const OnlineWenetCtcModelConfig &wenet_ctc,
-                    const OnlineZipformer2CtcModelConfig &zipformer2_ctc,
-                    const OnlineNeMoCtcModelConfig &nemo_ctc,
-                    const ProviderConfig &provider_config,
-                    const std::string &tokens, int32_t num_threads,
-                    int32_t warm_up, bool debug, const std::string &model_type,
-                    const std::string &modeling_unit,
-                    const std::string &bpe_vocab)
-      : transducer(transducer),
-        paraformer(paraformer),
-        wenet_ctc(wenet_ctc),
-        zipformer2_ctc(zipformer2_ctc),
-        nemo_ctc(nemo_ctc),
-        provider_config(provider_config),
-        tokens(tokens),
-        num_threads(num_threads),
-        warm_up(warm_up),
-        debug(debug),
-        model_type(model_type),
-        modeling_unit(modeling_unit),
-        bpe_vocab(bpe_vocab) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-nemo-ctc-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-nemo-ctc-model-config.h
deleted file mode 100644
index 4fb1de0a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-nemo-ctc-model-config.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// sherpa-onnx/csrc/online-nemo-ctc-model-config.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_NEMO_CTC_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_ONLINE_NEMO_CTC_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OnlineNeMoCtcModelConfig {
-  std::string model;
-
-  OnlineNeMoCtcModelConfig() = default;
-
-  explicit OnlineNeMoCtcModelConfig(const std::string &model) : model(model) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_NEMO_CTC_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-nemo-ctc-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-nemo-ctc-model.h
deleted file mode 100644
index 4e5f820b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-nemo-ctc-model.h
+++ /dev/null
@@ -1,75 +0,0 @@
-// sherpa-onnx/csrc/online-nemo-ctc-model.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_NEMO_CTC_MODEL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_NEMO_CTC_MODEL_H_
-
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/online-ctc-model.h"
-#include "sherpa-onnx/csrc/online-model-config.h"
-
-namespace sherpa_onnx {
-
-class OnlineNeMoCtcModel : public OnlineCtcModel {
- public:
-  explicit OnlineNeMoCtcModel(const OnlineModelConfig &config);
-
-  template <typename Manager>
-  OnlineNeMoCtcModel(Manager *mgr, const OnlineModelConfig &config);
-
-  ~OnlineNeMoCtcModel() override;
-
-  // A list of 3 tensors:
-  //  - cache_last_channel
-  //  - cache_last_time
-  //  - cache_last_channel_len
-  std::vector<Ort::Value> GetInitStates() const override;
-
-  std::vector<Ort::Value> StackStates(
-      std::vector<std::vector<Ort::Value>> states) const override;
-
-  std::vector<std::vector<Ort::Value>> UnStackStates(
-      std::vector<Ort::Value> states) const override;
-
-  /**
-   *
-   * @param x A 3-D tensor of shape (N, T, C). N has to be 1.
-   * @param states  It is from GetInitStates() or returned from this method.
-   *
-   * @return Return a list of tensors
-   *    - ans[0] contains log_probs, of shape (N, T, C)
-   *    - ans[1:] contains next_states
-   */
-  std::vector<Ort::Value> Forward(
-      Ort::Value x, std::vector<Ort::Value> states) const override;
-
-  /** Return the vocabulary size of the model
-   */
-  int32_t VocabSize() const override;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const override;
-
-  // The model accepts this number of frames before subsampling as input
-  int32_t ChunkLength() const override;
-
-  // Similar to frame_shift in feature extractor, after processing
-  // ChunkLength() frames, we advance by ChunkShift() frames
-  // before we process the next chunk.
-  int32_t ChunkShift() const override;
-
-  bool SupportBatchProcessing() const override { return true; }
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_NEMO_CTC_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-paraformer-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-paraformer-decoder.h
deleted file mode 100644
index 9f675275..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-paraformer-decoder.h
+++ /dev/null
@@ -1,23 +0,0 @@
-// sherpa-onnx/csrc/online-paraformer-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_PARAFORMER_DECODER_H_
-#define SHERPA_ONNX_CSRC_ONLINE_PARAFORMER_DECODER_H_
-
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-
-namespace sherpa_onnx {
-
-struct OnlineParaformerDecoderResult {
-  /// The decoded token IDs
-  std::vector<int32_t> tokens;
-
-  int32_t last_non_blank_frame_index = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_PARAFORMER_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-paraformer-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-paraformer-model-config.h
deleted file mode 100644
index 29f33e45..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-paraformer-model-config.h
+++ /dev/null
@@ -1,31 +0,0 @@
-// sherpa-onnx/csrc/online-paraformer-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_PARAFORMER_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_ONLINE_PARAFORMER_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OnlineParaformerModelConfig {
-  std::string encoder;
-  std::string decoder;
-
-  OnlineParaformerModelConfig() = default;
-
-  OnlineParaformerModelConfig(const std::string &encoder,
-                              const std::string &decoder)
-      : encoder(encoder), decoder(decoder) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_PARAFORMER_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-paraformer-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-paraformer-model.h
deleted file mode 100644
index cbf2d715..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-paraformer-model.h
+++ /dev/null
@@ -1,70 +0,0 @@
-// sherpa-onnx/csrc/online-paraformer-model.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_PARAFORMER_MODEL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_PARAFORMER_MODEL_H_
-
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/online-model-config.h"
-
-namespace sherpa_onnx {
-
-class OnlineParaformerModel {
- public:
-  explicit OnlineParaformerModel(const OnlineModelConfig &config);
-
-  template <typename Manager>
-  OnlineParaformerModel(Manager *mgr, const OnlineModelConfig &config);
-
-  ~OnlineParaformerModel();
-
-  std::vector<Ort::Value> ForwardEncoder(Ort::Value features,
-                                         Ort::Value features_length) const;
-
-  std::vector<Ort::Value> ForwardDecoder(Ort::Value encoder_out,
-                                         Ort::Value encoder_out_length,
-                                         Ort::Value acoustic_embedding,
-                                         Ort::Value acoustic_embedding_length,
-                                         std::vector<Ort::Value> states) const;
-
-  /** Return the vocabulary size of the model
-   */
-  int32_t VocabSize() const;
-
-  /** It is lfr_m in config.yaml
-   */
-  int32_t LfrWindowSize() const;
-
-  /** It is lfr_n in config.yaml
-   */
-  int32_t LfrWindowShift() const;
-
-  int32_t EncoderOutputSize() const;
-
-  int32_t DecoderKernelSize() const;
-  int32_t DecoderNumBlocks() const;
-
-  /** Return negative mean for CMVN
-   */
-  const std::vector<float> &NegativeMean() const;
-
-  /** Return inverse stddev for CMVN
-   */
-  const std::vector<float> &InverseStdDev() const;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_PARAFORMER_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation-cnn-bilstm-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation-cnn-bilstm-impl.h
deleted file mode 100644
index e586ccd0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation-cnn-bilstm-impl.h
+++ /dev/null
@@ -1,278 +0,0 @@
-// sherpa-onnx/csrc/online-punctuation-cnn-bilstm-impl.h
-//
-// Copyright (c) 2024 Jian You (jianyou@cisco.com, Cisco Systems)
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_PUNCTUATION_CNN_BILSTM_IMPL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_PUNCTUATION_CNN_BILSTM_IMPL_H_
-
-#include <math.h>
-
-#include <algorithm>
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include <chrono>  // NOLINT
-
-#include "sherpa-onnx/csrc/macros.h"
-#include "sherpa-onnx/csrc/math.h"
-#include "sherpa-onnx/csrc/online-cnn-bilstm-model-meta-data.h"
-#include "sherpa-onnx/csrc/online-cnn-bilstm-model.h"
-#include "sherpa-onnx/csrc/online-punctuation-impl.h"
-#include "sherpa-onnx/csrc/online-punctuation.h"
-#include "sherpa-onnx/csrc/onnx-utils.h"
-#include "sherpa-onnx/csrc/text-utils.h"
-#include "ssentencepiece/csrc/ssentencepiece.h"
-
-namespace sherpa_onnx {
-
-static const int32_t kMaxSeqLen = 200;
-
-class OnlinePunctuationCNNBiLSTMImpl : public OnlinePunctuationImpl {
- public:
-  explicit OnlinePunctuationCNNBiLSTMImpl(const OnlinePunctuationConfig &config)
-      : config_(config), model_(config.model) {
-    if (!config_.model.bpe_vocab.empty()) {
-      bpe_encoder_ = std::make_unique<ssentencepiece::Ssentencepiece>(
-          config_.model.bpe_vocab);
-    }
-  }
-
-#if __ANDROID_API__ >= 9
-  OnlinePunctuationCNNBiLSTMImpl(AAssetManager *mgr,
-                                 const OnlinePunctuationConfig &config)
-      : config_(config), model_(mgr, config.model) {
-    if (!config_.model.bpe_vocab.empty()) {
-      auto buf = ReadFile(mgr, config_.model.bpe_vocab);
-      std::istringstream iss(std::string(buf.begin(), buf.end()));
-      bpe_encoder_ = std::make_unique<ssentencepiece::Ssentencepiece>(iss);
-    }
-  }
-#endif
-
-  std::string AddPunctuationWithCase(const std::string &text) const override {
-    if (text.empty()) {
-      return {};
-    }
-
-    std::vector<int32_t> tokens_list;     // N * kMaxSeqLen
-    std::vector<int32_t> valids_list;     // N * kMaxSeqLen
-    std::vector<int32_t> label_len_list;  // N
-
-    EncodeSentences(text, tokens_list, valids_list, label_len_list);
-
-    const auto &meta_data = model_.GetModelMetadata();
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    int32_t n = label_len_list.size();
-
-    std::array<int64_t, 2> token_ids_shape = {n, kMaxSeqLen};
-    Ort::Value token_ids = Ort::Value::CreateTensor(
-        memory_info, tokens_list.data(), tokens_list.size(),
-        token_ids_shape.data(), token_ids_shape.size());
-
-    std::array<int64_t, 2> valid_ids_shape = {n, kMaxSeqLen};
-    Ort::Value valid_ids = Ort::Value::CreateTensor(
-        memory_info, valids_list.data(), valids_list.size(),
-        valid_ids_shape.data(), valid_ids_shape.size());
-
-    std::array<int64_t, 1> label_len_shape = {n};
-    Ort::Value label_len = Ort::Value::CreateTensor(
-        memory_info, label_len_list.data(), label_len_list.size(),
-        label_len_shape.data(), label_len_shape.size());
-
-    auto pair = model_.Forward(std::move(token_ids), std::move(valid_ids),
-                               std::move(label_len));
-
-    std::vector<int32_t> case_pred;
-    std::vector<int32_t> punct_pred;
-    const float *active_case_logits = pair.first.GetTensorData<float>();
-    const float *active_punct_logits = pair.second.GetTensorData<float>();
-    std::vector<int64_t> case_logits_shape =
-        pair.first.GetTensorTypeAndShapeInfo().GetShape();
-
-    for (int32_t i = 0; i < case_logits_shape[0]; ++i) {
-      const float *p_cur_case = active_case_logits + i * meta_data.num_cases;
-      auto index_case = static_cast<int32_t>(std::distance(
-          p_cur_case,
-          std::max_element(p_cur_case, p_cur_case + meta_data.num_cases)));
-      case_pred.push_back(index_case);
-
-      const float *p_cur_punct =
-          active_punct_logits + i * meta_data.num_punctuations;
-      auto index_punct = static_cast<int32_t>(std::distance(
-          p_cur_punct,
-          std::max_element(p_cur_punct,
-                           p_cur_punct + meta_data.num_punctuations)));
-      punct_pred.push_back(index_punct);
-    }
-
-    std::string ans = DecodeSentences(text, case_pred, punct_pred);
-
-    return ans;
-  }
-
- private:
-  void EncodeSentences(const std::string &text,
-                       std::vector<int32_t> &tokens_list,             // NOLINT
-                       std::vector<int32_t> &valids_list,             // NOLINT
-                       std::vector<int32_t> &label_len_list) const {  // NOLINT
-    std::vector<int32_t> tokens;
-    std::vector<int32_t> valids;
-    int32_t label_len = 0;
-
-    tokens.push_back(1);  // hardcode 1 now, 1 - <s>
-    valids.push_back(1);
-
-    std::stringstream ss(text);
-    std::string word;
-    while (ss >> word) {
-      std::vector<int32_t> word_tokens;
-      bpe_encoder_->Encode(word, &word_tokens);
-
-      int32_t seq_len = tokens.size() + word_tokens.size();
-      if (seq_len > kMaxSeqLen - 1) {
-        tokens.push_back(2);  // hardcode 2 now, 2 - </s>
-        valids.push_back(1);
-
-        label_len = std::count(valids.begin(), valids.end(), 1);
-
-        if (tokens.size() < kMaxSeqLen) {
-          tokens.resize(kMaxSeqLen, 0);
-          valids.resize(kMaxSeqLen, 0);
-        }
-
-        assert(tokens.size() == kMaxSeqLen);
-        assert(valids.size() == kMaxSeqLen);
-
-        tokens_list.insert(tokens_list.end(), tokens.begin(), tokens.end());
-        valids_list.insert(valids_list.end(), valids.begin(), valids.end());
-        label_len_list.push_back(label_len);
-
-        std::vector<int32_t>().swap(tokens);
-        std::vector<int32_t>().swap(valids);
-        label_len = 0;
-        tokens.push_back(1);  // hardcode 1 now, 1 - <s>
-        valids.push_back(1);
-      }
-
-      tokens.insert(tokens.end(), word_tokens.begin(), word_tokens.end());
-      valids.push_back(1);  // only the first sub word is valid
-      int32_t remaining_size = static_cast<int32_t>(word_tokens.size()) - 1;
-      if (remaining_size > 0) {
-        int32_t valids_cur_size = static_cast<int32_t>(valids.size());
-        valids.resize(valids_cur_size + remaining_size, 0);
-      }
-    }
-
-    if (tokens.size() > 0) {
-      tokens.push_back(2);  // hardcode 2 now, 2 - </s>
-      valids.push_back(1);
-
-      label_len = std::count(valids.begin(), valids.end(), 1);
-
-      if (tokens.size() < kMaxSeqLen) {
-        tokens.resize(kMaxSeqLen, 0);
-        valids.resize(kMaxSeqLen, 0);
-      }
-
-      assert(tokens.size() == kMaxSeqLen);
-      assert(valids.size() == kMaxSeqLen);
-
-      tokens_list.insert(tokens_list.end(), tokens.begin(), tokens.end());
-      valids_list.insert(valids_list.end(), valids.begin(), valids.end());
-      label_len_list.push_back(label_len);
-    }
-  }
-
-  std::string DecodeSentences(const std::string &raw_text,
-                              const std::vector<int32_t> &case_pred,
-                              const std::vector<int32_t> &punct_pred) const {
-    std::string result_text;
-    std::istringstream iss(raw_text);
-    std::vector<std::string> words;
-    std::string word;
-
-    while (iss >> word) {
-      words.emplace_back(word);
-    }
-
-    assert(words.size() == case_pred.size());
-    assert(words.size() == punct_pred.size());
-
-    for (int32_t i = 0; i < words.size(); ++i) {
-      std::string prefix = ((i != 0) ? " " : "");
-      result_text += prefix;
-      switch (case_pred[i]) {
-        case 1:  // upper
-        {
-          std::transform(words[i].begin(), words[i].end(), words[i].begin(),
-                         [](auto c) { return std::toupper(c); });
-          result_text += words[i];
-          break;
-        }
-        case 2:  // cap
-        {
-          words[i][0] = std::toupper(words[i][0]);
-          result_text += words[i];
-          break;
-        }
-        case 3:  // mix case
-        {
-          // TODO(frankyoujian):
-          // Need to add a map containing supported mix case words so that we
-          // can fetch the predicted word from the map e.g. mcdonald's ->
-          // McDonald's
-          result_text += words[i];
-          break;
-        }
-        default: {
-          result_text += words[i];
-          break;
-        }
-      }
-
-      std::string suffix;
-      switch (punct_pred[i]) {
-        case 1:  // comma
-        {
-          suffix = ",";
-          break;
-        }
-        case 2:  // period
-        {
-          suffix = ".";
-          break;
-        }
-        case 3:  // question
-        {
-          suffix = "?";
-          break;
-        }
-        default:
-          break;
-      }
-
-      result_text += suffix;
-    }
-
-    return result_text;
-  }
-
- private:
-  OnlinePunctuationConfig config_;
-  OnlineCNNBiLSTMModel model_;
-  std::unique_ptr<ssentencepiece::Ssentencepiece> bpe_encoder_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_PUNCTUATION_CNN_BILSTM_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation-impl.h
deleted file mode 100644
index 456f594d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation-impl.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// sherpa-onnx/csrc/online-punctuation-impl.h
-//
-// Copyright (c) 2024 Jian You (jianyou@cisco.com, Cisco Systems)
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_PUNCTUATION_IMPL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_PUNCTUATION_IMPL_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "sherpa-onnx/csrc/online-punctuation.h"
-
-namespace sherpa_onnx {
-
-class OnlinePunctuationImpl {
- public:
-  virtual ~OnlinePunctuationImpl() = default;
-
-  static std::unique_ptr<OnlinePunctuationImpl> Create(
-      const OnlinePunctuationConfig &config);
-
-#if __ANDROID_API__ >= 9
-  static std::unique_ptr<OnlinePunctuationImpl> Create(
-      AAssetManager *mgr, const OnlinePunctuationConfig &config);
-#endif
-
-  virtual std::string AddPunctuationWithCase(const std::string &text) const = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_PUNCTUATION_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation-model-config.h
deleted file mode 100644
index b2d26ce9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation-model-config.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// sherpa-onnx/csrc/online-punctuation-model-config.h
-//
-// Copyright (c) 2024 Jian You (jianyou@cisco.com, Cisco Systems)
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_PUNCTUATION_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_ONLINE_PUNCTUATION_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OnlinePunctuationModelConfig {
-  std::string cnn_bilstm;
-  std::string bpe_vocab;
-
-  int32_t num_threads = 1;
-  bool debug = false;
-  std::string provider = "cpu";
-
-  OnlinePunctuationModelConfig() = default;
-
-  OnlinePunctuationModelConfig(const std::string &cnn_bilstm,
-                               const std::string &bpe_vocab,
-                               int32_t num_threads, bool debug,
-                               const std::string &provider)
-      : cnn_bilstm(cnn_bilstm),
-        bpe_vocab(bpe_vocab),
-        num_threads(num_threads),
-        debug(debug),
-        provider(provider) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_PUNCTUATION_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation.h
deleted file mode 100644
index f5d2a144..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-punctuation.h
+++ /dev/null
@@ -1,57 +0,0 @@
-// sherpa-onnx/csrc/online-punctuation.h
-//
-// Copyright (c) 2024 Jian You (jianyou@cisco.com, Cisco Systems)
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_PUNCTUATION_H_
-#define SHERPA_ONNX_CSRC_ONLINE_PUNCTUATION_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "sherpa-onnx/csrc/online-punctuation-model-config.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OnlinePunctuationConfig {
-  OnlinePunctuationModelConfig model;
-
-  OnlinePunctuationConfig() = default;
-
-  explicit OnlinePunctuationConfig(const OnlinePunctuationModelConfig &model)
-      : model(model) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-class OnlinePunctuationImpl;
-
-class OnlinePunctuation {
- public:
-  explicit OnlinePunctuation(const OnlinePunctuationConfig &config);
-
-#if __ANDROID_API__ >= 9
-  OnlinePunctuation(AAssetManager *mgr, const OnlinePunctuationConfig &config);
-#endif
-
-  ~OnlinePunctuation();
-
-  // Add punctuation and casing to the input text and return it.
-  std::string AddPunctuationWithCase(const std::string &text) const;
-
- private:
-  std::unique_ptr<OnlinePunctuationImpl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_PUNCTUATION_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-ctc-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-ctc-impl.h
deleted file mode 100644
index 3560b1ab..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-ctc-impl.h
+++ /dev/null
@@ -1,316 +0,0 @@
-// sherpa-onnx/csrc/online-recognizer-ctc-impl.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_CTC_IMPL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_CTC_IMPL_H_
-
-#include <algorithm>
-#include <ios>
-#include <memory>
-#include <sstream>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "sherpa-onnx/csrc/file-utils.h"
-#include "sherpa-onnx/csrc/macros.h"
-#include "sherpa-onnx/csrc/online-ctc-decoder.h"
-#include "sherpa-onnx/csrc/online-ctc-fst-decoder.h"
-#include "sherpa-onnx/csrc/online-ctc-greedy-search-decoder.h"
-#include "sherpa-onnx/csrc/online-ctc-model.h"
-#include "sherpa-onnx/csrc/online-recognizer-impl.h"
-#include "sherpa-onnx/csrc/symbol-table.h"
-
-namespace sherpa_onnx {
-
-static OnlineRecognizerResult Convert(const OnlineCtcDecoderResult &src,
-                                      const SymbolTable &sym_table,
-                                      float frame_shift_ms,
-                                      int32_t subsampling_factor,
-                                      int32_t segment,
-                                      int32_t frames_since_start) {
-  OnlineRecognizerResult r;
-  r.tokens.reserve(src.tokens.size());
-  r.timestamps.reserve(src.tokens.size());
-
-  for (auto i : src.tokens) {
-    auto sym = sym_table[i];
-
-    r.text.append(sym);
-
-    if (sym.size() == 1 && (sym[0] < 0x20 || sym[0] > 0x7e)) {
-      // for byte bpe models
-      // (but don't rewrite printable characters 0x20..0x7e,
-      //  which collide with standard BPE units)
-      std::ostringstream os;
-      os << "<0x" << std::hex << std::uppercase
-         << (static_cast<int32_t>(sym[0]) & 0xff) << ">";
-      sym = os.str();
-    }
-
-    r.tokens.push_back(std::move(sym));
-  }
-
-  float frame_shift_s = frame_shift_ms / 1000. * subsampling_factor;
-  for (auto t : src.timestamps) {
-    float time = frame_shift_s * t;
-    r.timestamps.push_back(time);
-  }
-
-  r.segment = segment;
-  r.words = std::move(src.words);
-  r.start_time = frames_since_start * frame_shift_ms / 1000.;
-
-  return r;
-}
-
-class OnlineRecognizerCtcImpl : public OnlineRecognizerImpl {
- public:
-  explicit OnlineRecognizerCtcImpl(const OnlineRecognizerConfig &config)
-      : OnlineRecognizerImpl(config),
-        config_(config),
-        model_(OnlineCtcModel::Create(config.model_config)),
-        endpoint_(config_.endpoint_config) {
-    if (!config.model_config.tokens_buf.empty()) {
-      sym_ = SymbolTable(config.model_config.tokens_buf, false);
-    } else {
-      /// assuming tokens_buf and tokens are guaranteed not being both empty
-      sym_ = SymbolTable(config.model_config.tokens, true);
-    }
-
-    if (!config.model_config.wenet_ctc.model.empty()) {
-      // WeNet CTC models assume input samples are in the range
-      // [-32768, 32767], so we set normalize_samples to false
-      config_.feat_config.normalize_samples = false;
-    }
-
-    InitDecoder();
-  }
-
-  template <typename Manager>
-  explicit OnlineRecognizerCtcImpl(Manager *mgr,
-                                   const OnlineRecognizerConfig &config)
-      : OnlineRecognizerImpl(mgr, config),
-        config_(config),
-        model_(OnlineCtcModel::Create(mgr, config.model_config)),
-        sym_(mgr, config.model_config.tokens),
-        endpoint_(config_.endpoint_config) {
-    if (!config.model_config.wenet_ctc.model.empty()) {
-      // WeNet CTC models assume input samples are in the range
-      // [-32768, 32767], so we set normalize_samples to false
-      config_.feat_config.normalize_samples = false;
-    }
-
-    InitDecoder();
-  }
-
-  std::unique_ptr<OnlineStream> CreateStream() const override {
-    auto stream = std::make_unique<OnlineStream>(config_.feat_config);
-    stream->SetStates(model_->GetInitStates());
-    stream->SetFasterDecoder(decoder_->CreateFasterDecoder());
-
-    return stream;
-  }
-
-  bool IsReady(OnlineStream *s) const override {
-    return s->GetNumProcessedFrames() + model_->ChunkLength() <
-           s->NumFramesReady();
-  }
-
-  void DecodeStreams(OnlineStream **ss, int32_t n) const override {
-    if (n == 1 || !model_->SupportBatchProcessing()) {
-      for (int32_t i = 0; i != n; ++i) {
-        DecodeStream(ss[i]);
-      }
-      return;
-    }
-
-    // batch processing
-    int32_t chunk_length = model_->ChunkLength();
-    int32_t chunk_shift = model_->ChunkShift();
-
-    int32_t feat_dim = ss[0]->FeatureDim();
-
-    std::vector<OnlineCtcDecoderResult> results(n);
-    std::vector<float> features_vec(n * chunk_length * feat_dim);
-    std::vector<std::vector<Ort::Value>> states_vec(n);
-    std::vector<int64_t> all_processed_frames(n);
-
-    for (int32_t i = 0; i != n; ++i) {
-      const auto num_processed_frames = ss[i]->GetNumProcessedFrames();
-      std::vector<float> features =
-          ss[i]->GetFrames(num_processed_frames, chunk_length);
-
-      // Question: should num_processed_frames include chunk_shift?
-      ss[i]->GetNumProcessedFrames() += chunk_shift;
-
-      std::copy(features.begin(), features.end(),
-                features_vec.data() + i * chunk_length * feat_dim);
-
-      results[i] = std::move(ss[i]->GetCtcResult());
-      states_vec[i] = std::move(ss[i]->GetStates());
-      all_processed_frames[i] = num_processed_frames;
-    }
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    std::array<int64_t, 3> x_shape{n, chunk_length, feat_dim};
-
-    Ort::Value x = Ort::Value::CreateTensor(memory_info, features_vec.data(),
-                                            features_vec.size(), x_shape.data(),
-                                            x_shape.size());
-
-    auto states = model_->StackStates(std::move(states_vec));
-    int32_t num_states = states.size();
-    auto out = model_->Forward(std::move(x), std::move(states));
-    std::vector<Ort::Value> out_states;
-    out_states.reserve(num_states);
-
-    for (int32_t k = 1; k != num_states + 1; ++k) {
-      out_states.push_back(std::move(out[k]));
-    }
-
-    std::vector<std::vector<Ort::Value>> next_states =
-        model_->UnStackStates(std::move(out_states));
-
-    decoder_->Decode(std::move(out[0]), &results, ss, n);
-
-    for (int32_t k = 0; k != n; ++k) {
-      ss[k]->SetCtcResult(results[k]);
-      ss[k]->SetStates(std::move(next_states[k]));
-    }
-  }
-
-  OnlineRecognizerResult GetResult(OnlineStream *s) const override {
-    OnlineCtcDecoderResult decoder_result = s->GetCtcResult();
-
-    // TODO(fangjun): Remember to change these constants if needed
-    int32_t frame_shift_ms = 10;
-    int32_t subsampling_factor = 4;
-    auto r = Convert(decoder_result, sym_, frame_shift_ms, subsampling_factor,
-                     s->GetCurrentSegment(), s->GetNumFramesSinceStart());
-    r.text = ApplyInverseTextNormalization(r.text);
-    return r;
-  }
-
-  bool IsEndpoint(OnlineStream *s) const override {
-    if (!config_.enable_endpoint) {
-      return false;
-    }
-
-    int32_t num_processed_frames = s->GetNumProcessedFrames();
-
-    // frame shift is 10 milliseconds
-    float frame_shift_in_seconds = 0.01;
-
-    // subsampling factor is 4
-    int32_t trailing_silence_frames = s->GetCtcResult().num_trailing_blanks * 4;
-
-    return endpoint_.IsEndpoint(num_processed_frames, trailing_silence_frames,
-                                frame_shift_in_seconds);
-  }
-
-  void Reset(OnlineStream *s) const override {
-    // segment is incremented only when the last
-    // result is not empty
-    const auto &r = s->GetCtcResult();
-    if (!r.tokens.empty()) {
-      s->GetCurrentSegment() += 1;
-    }
-
-    // clear result
-    s->SetCtcResult({});
-
-    // clear states
-    s->SetStates(model_->GetInitStates());
-
-    s->GetFasterDecoderProcessedFrames() = 0;
-
-    // Note: We only update counters. The underlying audio samples
-    // are not discarded.
-    s->Reset();
-  }
-
- private:
-  void InitDecoder() {
-    if (!sym_.Contains("<blk>") && !sym_.Contains("<eps>") &&
-        !sym_.Contains("<blank>")) {
-      SHERPA_ONNX_LOGE(
-          "We expect that tokens.txt contains "
-          "the symbol <blk> or <eps> or <blank> and its ID.");
-      exit(-1);
-    }
-
-    int32_t blank_id = 0;
-    if (sym_.Contains("<blk>")) {
-      blank_id = sym_["<blk>"];
-    } else if (sym_.Contains("<eps>")) {
-      // for tdnn models of the yesno recipe from icefall
-      blank_id = sym_["<eps>"];
-    } else if (sym_.Contains("<blank>")) {
-      // for WeNet CTC models
-      blank_id = sym_["<blank>"];
-    }
-
-    if (!config_.ctc_fst_decoder_config.graph.empty()) {
-      decoder_ = std::make_unique<OnlineCtcFstDecoder>(
-          config_.ctc_fst_decoder_config, blank_id);
-    } else if (config_.decoding_method == "greedy_search") {
-      decoder_ = std::make_unique<OnlineCtcGreedySearchDecoder>(blank_id);
-    } else {
-      SHERPA_ONNX_LOGE(
-          "Unsupported decoding method: %s for streaming CTC models",
-          config_.decoding_method.c_str());
-      exit(-1);
-    }
-  }
-
-  void DecodeStream(OnlineStream *s) const {
-    int32_t chunk_length = model_->ChunkLength();
-    int32_t chunk_shift = model_->ChunkShift();
-
-    int32_t feat_dim = s->FeatureDim();
-
-    const auto num_processed_frames = s->GetNumProcessedFrames();
-    std::vector<float> frames =
-        s->GetFrames(num_processed_frames, chunk_length);
-    s->GetNumProcessedFrames() += chunk_shift;
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    std::array<int64_t, 3> x_shape{1, chunk_length, feat_dim};
-    Ort::Value x =
-        Ort::Value::CreateTensor(memory_info, frames.data(), frames.size(),
-                                 x_shape.data(), x_shape.size());
-    auto out = model_->Forward(std::move(x), std::move(s->GetStates()));
-    int32_t num_states = static_cast<int32_t>(out.size()) - 1;
-
-    std::vector<Ort::Value> states;
-    states.reserve(num_states);
-
-    for (int32_t i = 0; i != num_states; ++i) {
-      states.push_back(std::move(out[i + 1]));
-    }
-    s->SetStates(std::move(states));
-
-    std::vector<OnlineCtcDecoderResult> results(1);
-    results[0] = std::move(s->GetCtcResult());
-
-    decoder_->Decode(std::move(out[0]), &results, &s, 1);
-    s->SetCtcResult(results[0]);
-  }
-
- private:
-  OnlineRecognizerConfig config_;
-  std::unique_ptr<OnlineCtcModel> model_;
-  std::unique_ptr<OnlineCtcDecoder> decoder_;
-  SymbolTable sym_;
-  Endpoint endpoint_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_CTC_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-impl.h
deleted file mode 100644
index b7bda786..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-impl.h
+++ /dev/null
@@ -1,71 +0,0 @@
-// sherpa-onnx/csrc/online-recognizer-impl.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_IMPL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_IMPL_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "kaldifst/csrc/text-normalizer.h"
-#include "sherpa-onnx/csrc/macros.h"
-#include "sherpa-onnx/csrc/online-recognizer.h"
-#include "sherpa-onnx/csrc/online-stream.h"
-
-namespace sherpa_onnx {
-
-class OnlineRecognizerImpl {
- public:
-  explicit OnlineRecognizerImpl(const OnlineRecognizerConfig &config);
-
-  static std::unique_ptr<OnlineRecognizerImpl> Create(
-      const OnlineRecognizerConfig &config);
-
-  template <typename Manager>
-  OnlineRecognizerImpl(Manager *mgr, const OnlineRecognizerConfig &config);
-
-  template <typename Manager>
-  static std::unique_ptr<OnlineRecognizerImpl> Create(
-      Manager *mgr, const OnlineRecognizerConfig &config);
-
-  virtual ~OnlineRecognizerImpl() = default;
-
-  virtual std::unique_ptr<OnlineStream> CreateStream() const = 0;
-
-  virtual std::unique_ptr<OnlineStream> CreateStream(
-      const std::string &hotwords) const {
-    SHERPA_ONNX_LOGE("Only transducer models support contextual biasing.");
-    exit(-1);
-  }
-
-  virtual bool IsReady(OnlineStream *s) const = 0;
-
-  virtual void WarmpUpRecognizer(int32_t warmup, int32_t mbs) const {
-    // ToDo extending to other  models
-    SHERPA_ONNX_LOGE("Only zipformer2 model supports Warm up for now.");
-    exit(-1);
-  }
-
-  virtual void DecodeStreams(OnlineStream **ss, int32_t n) const = 0;
-
-  virtual OnlineRecognizerResult GetResult(OnlineStream *s) const = 0;
-
-  virtual bool IsEndpoint(OnlineStream *s) const = 0;
-
-  virtual void Reset(OnlineStream *s) const = 0;
-
-  std::string ApplyInverseTextNormalization(std::string text) const;
-
- private:
-  OnlineRecognizerConfig config_;
-  // for inverse text normalization. Used only if
-  // config.rule_fsts is not empty or
-  // config.rule_fars is not empty
-  std::vector<std::unique_ptr<kaldifst::TextNormalizer>> itn_list_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-paraformer-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-paraformer-impl.h
deleted file mode 100644
index 1e02fe51..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-paraformer-impl.h
+++ /dev/null
@@ -1,475 +0,0 @@
-// sherpa-onnx/csrc/online-recognizer-paraformer-impl.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_PARAFORMER_IMPL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_PARAFORMER_IMPL_H_
-
-#include <algorithm>
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "sherpa-onnx/csrc/file-utils.h"
-#include "sherpa-onnx/csrc/macros.h"
-#include "sherpa-onnx/csrc/online-lm.h"
-#include "sherpa-onnx/csrc/online-paraformer-decoder.h"
-#include "sherpa-onnx/csrc/online-paraformer-model.h"
-#include "sherpa-onnx/csrc/online-recognizer-impl.h"
-#include "sherpa-onnx/csrc/online-recognizer.h"
-#include "sherpa-onnx/csrc/symbol-table.h"
-
-namespace sherpa_onnx {
-
-static OnlineRecognizerResult Convert(const OnlineParaformerDecoderResult &src,
-                                      const SymbolTable &sym_table) {
-  OnlineRecognizerResult r;
-  r.tokens.reserve(src.tokens.size());
-
-  std::string text;
-
-  // When the current token ends with "@@" we set mergeable to true
-  bool mergeable = false;
-
-  for (int32_t i = 0; i != src.tokens.size(); ++i) {
-    auto sym = sym_table[src.tokens[i]];
-    r.tokens.push_back(sym);
-
-    if ((sym.back() != '@') || (sym.size() > 2 && sym[sym.size() - 2] != '@')) {
-      // sym does not end with "@@"
-      const uint8_t *p = reinterpret_cast<const uint8_t *>(sym.c_str());
-      if (p[0] < 0x80) {
-        // an ascii
-        if (mergeable) {
-          mergeable = false;
-          text.append(sym);
-        } else {
-          text.append(" ");
-          text.append(sym);
-        }
-      } else {
-        // not an ascii
-        mergeable = false;
-
-        if (i > 0) {
-          const uint8_t p = reinterpret_cast<const uint8_t *>(
-              sym_table[src.tokens[i - 1]].c_str())[0];
-          if (p < 0x80) {
-            // put a space between ascii and non-ascii
-            text.append(" ");
-          }
-        }
-        text.append(sym);
-      }
-    } else {
-      // this sym ends with @@
-      sym = std::string(sym.data(), sym.size() - 2);
-      if (mergeable) {
-        text.append(sym);
-      } else {
-        text.append(" ");
-        text.append(sym);
-        mergeable = true;
-      }
-    }
-  }
-  r.text = std::move(text);
-
-  return r;
-}
-
-// y[i] += x[i] * scale
-static void ScaleAddInPlace(const float *x, int32_t n, float scale, float *y) {
-  for (int32_t i = 0; i != n; ++i) {
-    y[i] += x[i] * scale;
-  }
-}
-
-// y[i] = x[i] * scale
-static void Scale(const float *x, int32_t n, float scale, float *y) {
-  for (int32_t i = 0; i != n; ++i) {
-    y[i] = x[i] * scale;
-  }
-}
-
-class OnlineRecognizerParaformerImpl : public OnlineRecognizerImpl {
- public:
-  explicit OnlineRecognizerParaformerImpl(const OnlineRecognizerConfig &config)
-      : OnlineRecognizerImpl(config),
-        config_(config),
-        model_(config.model_config),
-        endpoint_(config_.endpoint_config) {
-    if (!config.model_config.tokens_buf.empty()) {
-      sym_ = SymbolTable(config.model_config.tokens_buf, false);
-    } else {
-      /// assuming tokens_buf and tokens are guaranteed not being both empty
-      sym_ = SymbolTable(config.model_config.tokens, true);
-    }
-
-    if (config.decoding_method != "greedy_search") {
-      SHERPA_ONNX_LOGE(
-          "Unsupported decoding method: %s. Support only greedy_search at "
-          "present",
-          config.decoding_method.c_str());
-      exit(-1);
-    }
-
-    // Paraformer models assume input samples are in the range
-    // [-32768, 32767], so we set normalize_samples to false
-    config_.feat_config.normalize_samples = false;
-  }
-
-  template <typename Manager>
-  explicit OnlineRecognizerParaformerImpl(Manager *mgr,
-                                          const OnlineRecognizerConfig &config)
-      : OnlineRecognizerImpl(mgr, config),
-        config_(config),
-        model_(mgr, config.model_config),
-        sym_(mgr, config.model_config.tokens),
-        endpoint_(config_.endpoint_config) {
-    if (config.decoding_method != "greedy_search") {
-      SHERPA_ONNX_LOGE("Unsupported decoding method: %s",
-                       config.decoding_method.c_str());
-      exit(-1);
-    }
-
-    // Paraformer models assume input samples are in the range
-    // [-32768, 32767], so we set normalize_samples to false
-    config_.feat_config.normalize_samples = false;
-  }
-
-  OnlineRecognizerParaformerImpl(const OnlineRecognizerParaformerImpl &) =
-      delete;
-
-  OnlineRecognizerParaformerImpl operator=(
-      const OnlineRecognizerParaformerImpl &) = delete;
-
-  std::unique_ptr<OnlineStream> CreateStream() const override {
-    auto stream = std::make_unique<OnlineStream>(config_.feat_config);
-
-    OnlineParaformerDecoderResult r;
-    stream->SetParaformerResult(r);
-
-    return stream;
-  }
-
-  bool IsReady(OnlineStream *s) const override {
-    return s->GetNumProcessedFrames() + chunk_size_ < s->NumFramesReady();
-  }
-
-  void DecodeStreams(OnlineStream **ss, int32_t n) const override {
-    // TODO(fangjun): Support batch size > 1
-    for (int32_t i = 0; i != n; ++i) {
-      DecodeStream(ss[i]);
-    }
-  }
-
-  OnlineRecognizerResult GetResult(OnlineStream *s) const override {
-    auto decoder_result = s->GetParaformerResult();
-
-    auto r = Convert(decoder_result, sym_);
-    r.text = ApplyInverseTextNormalization(r.text);
-    return r;
-  }
-
-  bool IsEndpoint(OnlineStream *s) const override {
-    if (!config_.enable_endpoint) {
-      return false;
-    }
-
-    const auto &result = s->GetParaformerResult();
-
-    int32_t num_processed_frames = s->GetNumProcessedFrames();
-
-    // frame shift is 10 milliseconds
-    float frame_shift_in_seconds = 0.01;
-
-    int32_t trailing_silence_frames =
-        num_processed_frames - result.last_non_blank_frame_index;
-
-    return endpoint_.IsEndpoint(num_processed_frames, trailing_silence_frames,
-                                frame_shift_in_seconds);
-  }
-
-  void Reset(OnlineStream *s) const override {
-    OnlineParaformerDecoderResult r;
-    s->SetParaformerResult(r);
-
-    s->GetStates().clear();
-    s->GetParaformerEncoderOutCache().clear();
-    s->GetParaformerAlphaCache().clear();
-
-    // s->GetParaformerFeatCache().clear();
-
-    // Note: We only update counters. The underlying audio samples
-    // are not discarded.
-    s->Reset();
-  }
-
- private:
-  void DecodeStream(OnlineStream *s) const {
-    const auto num_processed_frames = s->GetNumProcessedFrames();
-    std::vector<float> frames = s->GetFrames(num_processed_frames, chunk_size_);
-    s->GetNumProcessedFrames() += chunk_size_ - 1;
-
-    frames = ApplyLFR(frames);
-    ApplyCMVN(&frames);
-    PositionalEncoding(&frames, num_processed_frames / model_.LfrWindowShift());
-
-    int32_t feat_dim = model_.NegativeMean().size();
-
-    // We have scaled inv_stddev by sqrt(encoder_output_size)
-    // so the following line can be commented out
-    // frames *= encoder_output_size ** 0.5
-
-    // add overlap chunk
-    std::vector<float> &feat_cache = s->GetParaformerFeatCache();
-    if (feat_cache.empty()) {
-      int32_t n = (left_chunk_size_ + right_chunk_size_) * feat_dim;
-      feat_cache.resize(n, 0);
-    }
-
-    frames.insert(frames.begin(), feat_cache.begin(), feat_cache.end());
-    std::copy(frames.end() - feat_cache.size(), frames.end(),
-              feat_cache.begin());
-
-    int32_t num_frames = frames.size() / feat_dim;
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    std::array<int64_t, 3> x_shape{1, num_frames, feat_dim};
-    Ort::Value x =
-        Ort::Value::CreateTensor(memory_info, frames.data(), frames.size(),
-                                 x_shape.data(), x_shape.size());
-
-    int64_t x_len_shape = 1;
-    int32_t x_len_val = num_frames;
-
-    Ort::Value x_length =
-        Ort::Value::CreateTensor(memory_info, &x_len_val, 1, &x_len_shape, 1);
-
-    auto encoder_out_vec =
-        model_.ForwardEncoder(std::move(x), std::move(x_length));
-
-    // CIF search
-    auto &encoder_out = encoder_out_vec[0];
-    auto &encoder_out_len = encoder_out_vec[1];
-    auto &alpha = encoder_out_vec[2];
-
-    float *p_alpha = alpha.GetTensorMutableData<float>();
-
-    std::vector<int64_t> alpha_shape =
-        alpha.GetTensorTypeAndShapeInfo().GetShape();
-
-    std::fill(p_alpha, p_alpha + left_chunk_size_, 0);
-    std::fill(p_alpha + alpha_shape[1] - right_chunk_size_,
-              p_alpha + alpha_shape[1], 0);
-
-    const float *p_encoder_out = encoder_out.GetTensorData<float>();
-
-    std::vector<int64_t> encoder_out_shape =
-        encoder_out.GetTensorTypeAndShapeInfo().GetShape();
-
-    std::vector<float> &initial_hidden = s->GetParaformerEncoderOutCache();
-    if (initial_hidden.empty()) {
-      initial_hidden.resize(encoder_out_shape[2]);
-    }
-
-    std::vector<float> &alpha_cache = s->GetParaformerAlphaCache();
-    if (alpha_cache.empty()) {
-      alpha_cache.resize(1);
-    }
-
-    std::vector<float> acoustic_embedding;
-    acoustic_embedding.reserve(encoder_out_shape[1] * encoder_out_shape[2]);
-
-    float threshold = 1.0;
-
-    float integrate = alpha_cache[0];
-
-    for (int32_t i = 0; i != encoder_out_shape[1]; ++i) {
-      float this_alpha = p_alpha[i];
-      if (integrate + this_alpha < threshold) {
-        integrate += this_alpha;
-        ScaleAddInPlace(p_encoder_out + i * encoder_out_shape[2],
-                        encoder_out_shape[2], this_alpha,
-                        initial_hidden.data());
-        continue;
-      }
-
-      // fire
-      ScaleAddInPlace(p_encoder_out + i * encoder_out_shape[2],
-                      encoder_out_shape[2], threshold - integrate,
-                      initial_hidden.data());
-      acoustic_embedding.insert(acoustic_embedding.end(),
-                                initial_hidden.begin(), initial_hidden.end());
-      integrate += this_alpha - threshold;
-
-      Scale(p_encoder_out + i * encoder_out_shape[2], encoder_out_shape[2],
-            integrate, initial_hidden.data());
-    }
-
-    alpha_cache[0] = integrate;
-
-    if (acoustic_embedding.empty()) {
-      return;
-    }
-
-    auto &states = s->GetStates();
-    if (states.empty()) {
-      states.reserve(model_.DecoderNumBlocks());
-
-      std::array<int64_t, 3> shape{1, model_.EncoderOutputSize(),
-                                   model_.DecoderKernelSize() - 1};
-
-      int32_t num_bytes = sizeof(float) * shape[0] * shape[1] * shape[2];
-
-      for (int32_t i = 0; i != model_.DecoderNumBlocks(); ++i) {
-        Ort::Value this_state = Ort::Value::CreateTensor<float>(
-            model_.Allocator(), shape.data(), shape.size());
-
-        memset(this_state.GetTensorMutableData<float>(), 0, num_bytes);
-
-        states.push_back(std::move(this_state));
-      }
-    }
-
-    int32_t num_tokens = acoustic_embedding.size() / initial_hidden.size();
-    std::array<int64_t, 3> acoustic_embedding_shape{
-        1, num_tokens, static_cast<int32_t>(initial_hidden.size())};
-
-    Ort::Value acoustic_embedding_tensor = Ort::Value::CreateTensor(
-        memory_info, acoustic_embedding.data(), acoustic_embedding.size(),
-        acoustic_embedding_shape.data(), acoustic_embedding_shape.size());
-
-    std::array<int64_t, 1> acoustic_embedding_length_shape{1};
-    Ort::Value acoustic_embedding_length_tensor = Ort::Value::CreateTensor(
-        memory_info, &num_tokens, 1, acoustic_embedding_length_shape.data(),
-        acoustic_embedding_length_shape.size());
-
-    auto decoder_out_vec = model_.ForwardDecoder(
-        std::move(encoder_out), std::move(encoder_out_len),
-        std::move(acoustic_embedding_tensor),
-        std::move(acoustic_embedding_length_tensor), std::move(states));
-
-    states.reserve(model_.DecoderNumBlocks());
-    for (int32_t i = 2; i != decoder_out_vec.size(); ++i) {
-      // TODO(fangjun): When we change chunk_size_, we need to
-      // slice decoder_out_vec[i] accordingly.
-      states.push_back(std::move(decoder_out_vec[i]));
-    }
-
-    const auto &sample_ids = decoder_out_vec[1];
-    const int64_t *p_sample_ids = sample_ids.GetTensorData<int64_t>();
-
-    bool non_blank_detected = false;
-
-    auto &result = s->GetParaformerResult();
-
-    for (int32_t i = 0; i != num_tokens; ++i) {
-      int32_t t = p_sample_ids[i];
-      if (t == 0) {
-        continue;
-      }
-
-      non_blank_detected = true;
-      result.tokens.push_back(t);
-    }
-
-    if (non_blank_detected) {
-      result.last_non_blank_frame_index = num_processed_frames;
-    }
-  }
-
-  std::vector<float> ApplyLFR(const std::vector<float> &in) const {
-    int32_t lfr_window_size = model_.LfrWindowSize();
-    int32_t lfr_window_shift = model_.LfrWindowShift();
-    int32_t in_feat_dim = config_.feat_config.feature_dim;
-
-    int32_t in_num_frames = in.size() / in_feat_dim;
-    int32_t out_num_frames =
-        (in_num_frames - lfr_window_size) / lfr_window_shift + 1;
-    int32_t out_feat_dim = in_feat_dim * lfr_window_size;
-
-    std::vector<float> out(out_num_frames * out_feat_dim);
-
-    const float *p_in = in.data();
-    float *p_out = out.data();
-
-    for (int32_t i = 0; i != out_num_frames; ++i) {
-      std::copy(p_in, p_in + out_feat_dim, p_out);
-
-      p_out += out_feat_dim;
-      p_in += lfr_window_shift * in_feat_dim;
-    }
-
-    return out;
-  }
-
-  void ApplyCMVN(std::vector<float> *v) const {
-    const std::vector<float> &neg_mean = model_.NegativeMean();
-    const std::vector<float> &inv_stddev = model_.InverseStdDev();
-
-    int32_t dim = neg_mean.size();
-    int32_t num_frames = v->size() / dim;
-
-    float *p = v->data();
-
-    for (int32_t i = 0; i != num_frames; ++i) {
-      for (int32_t k = 0; k != dim; ++k) {
-        p[k] = (p[k] + neg_mean[k]) * inv_stddev[k];
-      }
-
-      p += dim;
-    }
-  }
-
-  void PositionalEncoding(std::vector<float> *v, int32_t t_offset) const {
-    int32_t lfr_window_size = model_.LfrWindowSize();
-    int32_t in_feat_dim = config_.feat_config.feature_dim;
-
-    int32_t feat_dim = in_feat_dim * lfr_window_size;
-    int32_t T = v->size() / feat_dim;
-
-    // log(10000)/(7*80/2-1) == 0.03301197265941284
-    // 7 is lfr_window_size
-    // 80 is in_feat_dim
-    // 7*80 is feat_dim
-    constexpr float kScale = -0.03301197265941284;
-
-    for (int32_t t = 0; t != T; ++t) {
-      float *p = v->data() + t * feat_dim;
-
-      int32_t offset = t + 1 + t_offset;
-
-      for (int32_t d = 0; d < feat_dim / 2; ++d) {
-        float inv_timescale = offset * std::exp(d * kScale);
-
-        float sin_d = std::sin(inv_timescale);
-        float cos_d = std::cos(inv_timescale);
-
-        p[d] += sin_d;
-        p[d + feat_dim / 2] += cos_d;
-      }
-    }
-  }
-
- private:
-  OnlineRecognizerConfig config_;
-  OnlineParaformerModel model_;
-  SymbolTable sym_;
-  Endpoint endpoint_;
-
-  // 0.61 seconds
-  int32_t chunk_size_ = 61;
-  // (61 - 7) / 6 + 1 = 10
-
-  int32_t left_chunk_size_ = 5;
-  int32_t right_chunk_size_ = 3;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_PARAFORMER_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-transducer-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-transducer-impl.h
deleted file mode 100644
index 2eac3cf8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-transducer-impl.h
+++ /dev/null
@@ -1,495 +0,0 @@
-// sherpa-onnx/csrc/online-recognizer-transducer-impl.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_TRANSDUCER_IMPL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_TRANSDUCER_IMPL_H_
-
-#include <algorithm>
-#include <ios>
-#include <memory>
-#include <regex>  // NOLINT
-#include <sstream>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "sherpa-onnx/csrc/file-utils.h"
-#include "sherpa-onnx/csrc/macros.h"
-#include "sherpa-onnx/csrc/online-lm.h"
-#include "sherpa-onnx/csrc/online-recognizer-impl.h"
-#include "sherpa-onnx/csrc/online-recognizer.h"
-#include "sherpa-onnx/csrc/online-transducer-decoder.h"
-#include "sherpa-onnx/csrc/online-transducer-greedy-search-decoder.h"
-#include "sherpa-onnx/csrc/online-transducer-model.h"
-#include "sherpa-onnx/csrc/online-transducer-modified-beam-search-decoder.h"
-#include "sherpa-onnx/csrc/onnx-utils.h"
-#include "sherpa-onnx/csrc/symbol-table.h"
-#include "sherpa-onnx/csrc/utils.h"
-#include "ssentencepiece/csrc/ssentencepiece.h"
-
-namespace sherpa_onnx {
-
-OnlineRecognizerResult Convert(const OnlineTransducerDecoderResult &src,
-                               const SymbolTable &sym_table,
-                               float frame_shift_ms, int32_t subsampling_factor,
-                               int32_t segment, int32_t frames_since_start) {
-  OnlineRecognizerResult r;
-  r.tokens.reserve(src.tokens.size());
-  r.timestamps.reserve(src.tokens.size());
-
-  for (auto i : src.tokens) {
-    auto sym = sym_table[i];
-
-    r.text.append(sym);
-
-    if (sym.size() == 1 && (sym[0] < 0x20 || sym[0] > 0x7e)) {
-      // for byte bpe models
-      // (but don't rewrite printable characters 0x20..0x7e,
-      //  which collide with standard BPE units)
-      std::ostringstream os;
-      os << "<0x" << std::hex << std::uppercase
-         << (static_cast<int32_t>(sym[0]) & 0xff) << ">";
-      sym = os.str();
-    }
-
-    r.tokens.push_back(std::move(sym));
-  }
-
-  float frame_shift_s = frame_shift_ms / 1000. * subsampling_factor;
-  for (auto t : src.timestamps) {
-    float time = frame_shift_s * t;
-    r.timestamps.push_back(time);
-  }
-
-  r.ys_probs = std::move(src.ys_probs);
-  r.lm_probs = std::move(src.lm_probs);
-  r.context_scores = std::move(src.context_scores);
-
-  r.segment = segment;
-  r.start_time = frames_since_start * frame_shift_ms / 1000.;
-
-  return r;
-}
-
-class OnlineRecognizerTransducerImpl : public OnlineRecognizerImpl {
- public:
-  explicit OnlineRecognizerTransducerImpl(const OnlineRecognizerConfig &config)
-      : OnlineRecognizerImpl(config),
-        config_(config),
-        model_(OnlineTransducerModel::Create(config.model_config)),
-        endpoint_(config_.endpoint_config) {
-    if (!config.model_config.tokens_buf.empty()) {
-      sym_ = SymbolTable(config.model_config.tokens_buf, false);
-    } else {
-      /// assuming tokens_buf and tokens are guaranteed not being both empty
-      sym_ = SymbolTable(config.model_config.tokens, true);
-    }
-
-    if (sym_.Contains("<unk>")) {
-      unk_id_ = sym_["<unk>"];
-    }
-
-    model_->SetFeatureDim(config.feat_config.feature_dim);
-
-    if (config.decoding_method == "modified_beam_search") {
-      if (!config_.model_config.bpe_vocab.empty()) {
-        bpe_encoder_ = std::make_unique<ssentencepiece::Ssentencepiece>(
-            config_.model_config.bpe_vocab);
-      }
-
-      if (!config_.hotwords_buf.empty()) {
-        InitHotwordsFromBufStr();
-      } else if (!config_.hotwords_file.empty()) {
-        InitHotwords();
-      }
-
-      if (!config_.lm_config.model.empty()) {
-        lm_ = OnlineLM::Create(config.lm_config);
-      }
-
-      decoder_ = std::make_unique<OnlineTransducerModifiedBeamSearchDecoder>(
-          model_.get(), lm_.get(), config_.max_active_paths,
-          config_.lm_config.scale, config_.lm_config.shallow_fusion, unk_id_,
-          config_.blank_penalty, config_.temperature_scale);
-
-    } else if (config.decoding_method == "greedy_search") {
-      decoder_ = std::make_unique<OnlineTransducerGreedySearchDecoder>(
-          model_.get(), unk_id_, config_.blank_penalty,
-          config_.temperature_scale);
-
-    } else {
-      SHERPA_ONNX_LOGE("Unsupported decoding method: %s",
-                       config.decoding_method.c_str());
-      exit(-1);
-    }
-  }
-
-  template <typename Manager>
-  explicit OnlineRecognizerTransducerImpl(Manager *mgr,
-                                          const OnlineRecognizerConfig &config)
-      : OnlineRecognizerImpl(mgr, config),
-        config_(config),
-        model_(OnlineTransducerModel::Create(mgr, config.model_config)),
-        sym_(mgr, config.model_config.tokens),
-        endpoint_(config_.endpoint_config) {
-    if (sym_.Contains("<unk>")) {
-      unk_id_ = sym_["<unk>"];
-    }
-
-    model_->SetFeatureDim(config.feat_config.feature_dim);
-
-    if (config.decoding_method == "modified_beam_search") {
-#if 0
-      // TODO(fangjun): Implement it
-      if (!config_.lm_config.model.empty()) {
-        lm_ = OnlineLM::Create(mgr, config.lm_config);
-      }
-#endif
-
-      if (!config_.model_config.bpe_vocab.empty()) {
-        auto buf = ReadFile(mgr, config_.model_config.bpe_vocab);
-        std::istringstream iss(std::string(buf.begin(), buf.end()));
-        bpe_encoder_ = std::make_unique<ssentencepiece::Ssentencepiece>(iss);
-      }
-
-      if (!config_.hotwords_file.empty()) {
-        InitHotwords(mgr);
-      }
-
-      decoder_ = std::make_unique<OnlineTransducerModifiedBeamSearchDecoder>(
-          model_.get(), lm_.get(), config_.max_active_paths,
-          config_.lm_config.scale, config_.lm_config.shallow_fusion, unk_id_,
-          config_.blank_penalty, config_.temperature_scale);
-
-    } else if (config.decoding_method == "greedy_search") {
-      decoder_ = std::make_unique<OnlineTransducerGreedySearchDecoder>(
-          model_.get(), unk_id_, config_.blank_penalty,
-          config_.temperature_scale);
-
-    } else {
-      SHERPA_ONNX_LOGE("Unsupported decoding method: %s",
-                       config.decoding_method.c_str());
-      exit(-1);
-    }
-  }
-
-  std::unique_ptr<OnlineStream> CreateStream() const override {
-    auto stream =
-        std::make_unique<OnlineStream>(config_.feat_config, hotwords_graph_);
-    InitOnlineStream(stream.get());
-    return stream;
-  }
-
-  std::unique_ptr<OnlineStream> CreateStream(
-      const std::string &hotwords) const override {
-    auto hws = std::regex_replace(hotwords, std::regex("/"), "\n");
-    std::istringstream is(hws);
-    std::vector<std::vector<int32_t>> current;
-    std::vector<float> current_scores;
-    if (!EncodeHotwords(is, config_.model_config.modeling_unit, sym_,
-                        bpe_encoder_.get(), &current, &current_scores)) {
-      SHERPA_ONNX_LOGE("Encode hotwords failed, skipping, hotwords are : %s",
-                       hotwords.c_str());
-    }
-
-    int32_t num_default_hws = hotwords_.size();
-    int32_t num_hws = current.size();
-
-    current.insert(current.end(), hotwords_.begin(), hotwords_.end());
-
-    if (!current_scores.empty() && !boost_scores_.empty()) {
-      current_scores.insert(current_scores.end(), boost_scores_.begin(),
-                            boost_scores_.end());
-    } else if (!current_scores.empty() && boost_scores_.empty()) {
-      current_scores.insert(current_scores.end(), num_default_hws,
-                            config_.hotwords_score);
-    } else if (current_scores.empty() && !boost_scores_.empty()) {
-      current_scores.insert(current_scores.end(), num_hws,
-                            config_.hotwords_score);
-      current_scores.insert(current_scores.end(), boost_scores_.begin(),
-                            boost_scores_.end());
-    } else {
-      // Do nothing.
-    }
-
-    auto context_graph = std::make_shared<ContextGraph>(
-        current, config_.hotwords_score, current_scores);
-    auto stream =
-        std::make_unique<OnlineStream>(config_.feat_config, context_graph);
-    InitOnlineStream(stream.get());
-    return stream;
-  }
-
-  bool IsReady(OnlineStream *s) const override {
-    return s->GetNumProcessedFrames() + model_->ChunkSize() <
-           s->NumFramesReady();
-  }
-
-  // Warmping up engine with wp: warm_up count and max-batch-size
-  void WarmpUpRecognizer(int32_t warmup, int32_t mbs) const override {
-    auto max_batch_size = mbs;
-    if (warmup <= 0 || warmup > 100) {
-      return;
-    }
-    int32_t chunk_size = model_->ChunkSize();
-    int32_t chunk_shift = model_->ChunkShift();
-    int32_t feature_dim = 80;
-    std::vector<OnlineTransducerDecoderResult> results(max_batch_size);
-    std::vector<float> features_vec(max_batch_size * chunk_size * feature_dim);
-    std::vector<std::vector<Ort::Value>> states_vec(max_batch_size);
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    std::array<int64_t, 3> x_shape{max_batch_size, chunk_size, feature_dim};
-
-    for (int32_t i = 0; i != max_batch_size; ++i) {
-      states_vec[i] = model_->GetEncoderInitStates();
-      results[i] = decoder_->GetEmptyResult();
-    }
-
-    for (int32_t i = 0; i != warmup; ++i) {
-      auto states = model_->StackStates(states_vec);
-      Ort::Value x = Ort::Value::CreateTensor(memory_info, features_vec.data(),
-                                              features_vec.size(),
-                                              x_shape.data(), x_shape.size());
-      auto x_copy = Clone(model_->Allocator(), &x);
-      auto pair = model_->RunEncoder(std::move(x), std::move(states),
-                                     std::move(x_copy));
-      decoder_->Decode(std::move(pair.first), &results);
-    }
-  }
-
-  void DecodeStreams(OnlineStream **ss, int32_t n) const override {
-    int32_t chunk_size = model_->ChunkSize();
-    int32_t chunk_shift = model_->ChunkShift();
-
-    int32_t feature_dim = ss[0]->FeatureDim();
-
-    std::vector<OnlineTransducerDecoderResult> results(n);
-    std::vector<float> features_vec(n * chunk_size * feature_dim);
-    std::vector<std::vector<Ort::Value>> states_vec(n);
-    std::vector<int64_t> all_processed_frames(n);
-    bool has_context_graph = false;
-
-    for (int32_t i = 0; i != n; ++i) {
-      if (!has_context_graph && ss[i]->GetContextGraph()) {
-        has_context_graph = true;
-      }
-
-      const auto num_processed_frames = ss[i]->GetNumProcessedFrames();
-      std::vector<float> features =
-          ss[i]->GetFrames(num_processed_frames, chunk_size);
-
-      // Question: should num_processed_frames include chunk_shift?
-      ss[i]->GetNumProcessedFrames() += chunk_shift;
-
-      std::copy(features.begin(), features.end(),
-                features_vec.data() + i * chunk_size * feature_dim);
-
-      results[i] = std::move(ss[i]->GetResult());
-      states_vec[i] = std::move(ss[i]->GetStates());
-      all_processed_frames[i] = num_processed_frames;
-    }
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    std::array<int64_t, 3> x_shape{n, chunk_size, feature_dim};
-
-    Ort::Value x = Ort::Value::CreateTensor(memory_info, features_vec.data(),
-                                            features_vec.size(), x_shape.data(),
-                                            x_shape.size());
-
-    std::array<int64_t, 1> processed_frames_shape{
-        static_cast<int64_t>(all_processed_frames.size())};
-
-    Ort::Value processed_frames = Ort::Value::CreateTensor(
-        memory_info, all_processed_frames.data(), all_processed_frames.size(),
-        processed_frames_shape.data(), processed_frames_shape.size());
-
-    auto states = model_->StackStates(states_vec);
-
-    auto pair = model_->RunEncoder(std::move(x), std::move(states),
-                                   std::move(processed_frames));
-
-    if (has_context_graph) {
-      decoder_->Decode(std::move(pair.first), ss, &results);
-    } else {
-      decoder_->Decode(std::move(pair.first), &results);
-    }
-
-    std::vector<std::vector<Ort::Value>> next_states =
-        model_->UnStackStates(pair.second);
-
-    for (int32_t i = 0; i != n; ++i) {
-      ss[i]->SetResult(results[i]);
-      ss[i]->SetStates(std::move(next_states[i]));
-    }
-  }
-
-  OnlineRecognizerResult GetResult(OnlineStream *s) const override {
-    OnlineTransducerDecoderResult decoder_result = s->GetResult();
-    decoder_->StripLeadingBlanks(&decoder_result);
-
-    // TODO(fangjun): Remember to change these constants if needed
-    int32_t frame_shift_ms = 10;
-    int32_t subsampling_factor = 4;
-    auto r = Convert(decoder_result, sym_, frame_shift_ms, subsampling_factor,
-                     s->GetCurrentSegment(), s->GetNumFramesSinceStart());
-    r.text = ApplyInverseTextNormalization(std::move(r.text));
-    return r;
-  }
-
-  bool IsEndpoint(OnlineStream *s) const override {
-    if (!config_.enable_endpoint) {
-      return false;
-    }
-
-    int32_t num_processed_frames = s->GetNumProcessedFrames();
-
-    // frame shift is 10 milliseconds
-    float frame_shift_in_seconds = 0.01;
-
-    // subsampling factor is 4
-    int32_t trailing_silence_frames = s->GetResult().num_trailing_blanks * 4;
-
-    return endpoint_.IsEndpoint(num_processed_frames, trailing_silence_frames,
-                                frame_shift_in_seconds);
-  }
-
-  void Reset(OnlineStream *s) const override {
-    int32_t context_size = model_->ContextSize();
-
-    {
-      // segment is incremented only when the last
-      // result is not empty, contains non-blanks and longer than context_size)
-      const auto &r = s->GetResult();
-      if (!r.tokens.empty() && r.tokens.back() != 0 &&
-          r.tokens.size() > context_size) {
-        s->GetCurrentSegment() += 1;
-      }
-    }
-
-    // reset encoder states
-    // s->SetStates(model_->GetEncoderInitStates());
-
-    auto r = decoder_->GetEmptyResult();
-    auto last_result = s->GetResult();
-    // if last result is not empty, then
-    // preserve last tokens as the context for next result
-    if (static_cast<int32_t>(last_result.tokens.size()) > context_size) {
-      std::vector<int64_t> context(last_result.tokens.end() - context_size,
-                                   last_result.tokens.end());
-
-      Hypotheses context_hyp({{context, 0}});
-      r.hyps = std::move(context_hyp);
-      r.tokens = std::move(context);
-    }
-
-    if (config_.decoding_method == "modified_beam_search" &&
-        nullptr != s->GetContextGraph()) {
-      for (auto it = r.hyps.begin(); it != r.hyps.end(); ++it) {
-        it->second.context_state = s->GetContextGraph()->Root();
-      }
-    }
-
-    s->SetResult(r);
-
-    // Note: We only update counters. The underlying audio samples
-    // are not discarded.
-    s->Reset();
-  }
-
- private:
-  void InitHotwords() {
-    // each line in hotwords_file contains space-separated words
-
-    std::ifstream is(config_.hotwords_file);
-    if (!is) {
-      SHERPA_ONNX_LOGE("Open hotwords file failed: %s",
-                       config_.hotwords_file.c_str());
-      exit(-1);
-    }
-
-    if (!EncodeHotwords(is, config_.model_config.modeling_unit, sym_,
-                        bpe_encoder_.get(), &hotwords_, &boost_scores_)) {
-      SHERPA_ONNX_LOGE(
-          "Failed to encode some hotwords, skip them already, see logs above "
-          "for details.");
-    }
-    hotwords_graph_ = std::make_shared<ContextGraph>(
-        hotwords_, config_.hotwords_score, boost_scores_);
-  }
-
-  template <typename Manager>
-  void InitHotwords(Manager *mgr) {
-    // each line in hotwords_file contains space-separated words
-
-    auto buf = ReadFile(mgr, config_.hotwords_file);
-
-    std::istringstream is(std::string(buf.begin(), buf.end()));
-
-    if (!is) {
-      SHERPA_ONNX_LOGE("Open hotwords file failed: %s",
-                       config_.hotwords_file.c_str());
-      exit(-1);
-    }
-
-    if (!EncodeHotwords(is, config_.model_config.modeling_unit, sym_,
-                        bpe_encoder_.get(), &hotwords_, &boost_scores_)) {
-      SHERPA_ONNX_LOGE(
-          "Failed to encode some hotwords, skip them already, see logs above "
-          "for details.");
-    }
-    hotwords_graph_ = std::make_shared<ContextGraph>(
-        hotwords_, config_.hotwords_score, boost_scores_);
-  }
-
-  void InitHotwordsFromBufStr() {
-    // each line in hotwords_file contains space-separated words
-
-    std::istringstream iss(config_.hotwords_buf);
-    if (!EncodeHotwords(iss, config_.model_config.modeling_unit, sym_,
-                        bpe_encoder_.get(), &hotwords_, &boost_scores_)) {
-      SHERPA_ONNX_LOGE(
-          "Failed to encode some hotwords, skip them already, see logs above "
-          "for details.");
-    }
-    hotwords_graph_ = std::make_shared<ContextGraph>(
-        hotwords_, config_.hotwords_score, boost_scores_);
-  }
-
-  void InitOnlineStream(OnlineStream *stream) const {
-    auto r = decoder_->GetEmptyResult();
-
-    if (config_.decoding_method == "modified_beam_search" &&
-        nullptr != stream->GetContextGraph()) {
-      // r.hyps has only one element.
-      for (auto it = r.hyps.begin(); it != r.hyps.end(); ++it) {
-        it->second.context_state = stream->GetContextGraph()->Root();
-      }
-    }
-
-    stream->SetResult(r);
-    stream->SetStates(model_->GetEncoderInitStates());
-  }
-
- private:
-  OnlineRecognizerConfig config_;
-  std::vector<std::vector<int32_t>> hotwords_;
-  std::vector<float> boost_scores_;
-  ContextGraphPtr hotwords_graph_;
-  std::unique_ptr<ssentencepiece::Ssentencepiece> bpe_encoder_;
-  std::unique_ptr<OnlineTransducerModel> model_;
-  std::unique_ptr<OnlineLM> lm_;
-  std::unique_ptr<OnlineTransducerDecoder> decoder_;
-  SymbolTable sym_;
-  Endpoint endpoint_;
-  int32_t unk_id_ = -1;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_TRANSDUCER_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-transducer-nemo-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-transducer-nemo-impl.h
deleted file mode 100644
index a3f2756c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer-transducer-nemo-impl.h
+++ /dev/null
@@ -1,251 +0,0 @@
-// sherpa-onnx/csrc/online-recognizer-transducer-nemo-impl.h
-//
-// Copyright (c)  2022-2024  Xiaomi Corporation
-// Copyright (c)  2024  Sangeet Sagar
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_TRANSDUCER_NEMO_IMPL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_TRANSDUCER_NEMO_IMPL_H_
-
-#include <algorithm>
-#include <fstream>
-#include <ios>
-#include <memory>
-#include <regex>  // NOLINT
-#include <sstream>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "sherpa-onnx/csrc/macros.h"
-#include "sherpa-onnx/csrc/online-recognizer-impl.h"
-#include "sherpa-onnx/csrc/online-recognizer.h"
-#include "sherpa-onnx/csrc/online-transducer-greedy-search-nemo-decoder.h"
-#include "sherpa-onnx/csrc/online-transducer-nemo-model.h"
-#include "sherpa-onnx/csrc/symbol-table.h"
-#include "sherpa-onnx/csrc/transpose.h"
-#include "sherpa-onnx/csrc/utils.h"
-
-namespace sherpa_onnx {
-
-// defined in ./online-recognizer-transducer-impl.h
-OnlineRecognizerResult Convert(const OnlineTransducerDecoderResult &src,
-                               const SymbolTable &sym_table,
-                               float frame_shift_ms, int32_t subsampling_factor,
-                               int32_t segment, int32_t frames_since_start);
-
-class OnlineRecognizerTransducerNeMoImpl : public OnlineRecognizerImpl {
- public:
-  explicit OnlineRecognizerTransducerNeMoImpl(
-      const OnlineRecognizerConfig &config)
-      : OnlineRecognizerImpl(config),
-        config_(config),
-        endpoint_(config_.endpoint_config),
-        model_(
-            std::make_unique<OnlineTransducerNeMoModel>(config.model_config)) {
-    if (!config.model_config.tokens_buf.empty()) {
-      symbol_table_ = SymbolTable(config.model_config.tokens_buf, false);
-    } else {
-      /// assuming tokens_buf and tokens are guaranteed not being both empty
-      symbol_table_ = SymbolTable(config.model_config.tokens, true);
-    }
-
-    if (config.decoding_method == "greedy_search") {
-      decoder_ = std::make_unique<OnlineTransducerGreedySearchNeMoDecoder>(
-          model_.get(), config_.blank_penalty);
-    } else {
-      SHERPA_ONNX_LOGE("Unsupported decoding method: %s",
-                       config.decoding_method.c_str());
-      exit(-1);
-    }
-    PostInit();
-  }
-
-  template <typename Manager>
-  explicit OnlineRecognizerTransducerNeMoImpl(
-      Manager *mgr, const OnlineRecognizerConfig &config)
-      : OnlineRecognizerImpl(mgr, config),
-        config_(config),
-        symbol_table_(mgr, config.model_config.tokens),
-        endpoint_(config_.endpoint_config),
-        model_(std::make_unique<OnlineTransducerNeMoModel>(
-            mgr, config.model_config)) {
-    if (config.decoding_method == "greedy_search") {
-      decoder_ = std::make_unique<OnlineTransducerGreedySearchNeMoDecoder>(
-          model_.get(), config_.blank_penalty);
-    } else {
-      SHERPA_ONNX_LOGE("Unsupported decoding method: %s",
-                       config.decoding_method.c_str());
-      exit(-1);
-    }
-
-    PostInit();
-  }
-
-  std::unique_ptr<OnlineStream> CreateStream() const override {
-    auto stream = std::make_unique<OnlineStream>(config_.feat_config);
-    InitOnlineStream(stream.get());
-    return stream;
-  }
-
-  bool IsReady(OnlineStream *s) const override {
-    return s->GetNumProcessedFrames() + model_->ChunkSize() <
-           s->NumFramesReady();
-  }
-
-  OnlineRecognizerResult GetResult(OnlineStream *s) const override {
-    // TODO(fangjun): Remember to change these constants if needed
-    int32_t frame_shift_ms = 10;
-    int32_t subsampling_factor = model_->SubsamplingFactor();
-    auto r = Convert(s->GetResult(), symbol_table_, frame_shift_ms,
-                     subsampling_factor, s->GetCurrentSegment(),
-                     s->GetNumFramesSinceStart());
-    r.text = ApplyInverseTextNormalization(std::move(r.text));
-    return r;
-  }
-
-  bool IsEndpoint(OnlineStream *s) const override {
-    if (!config_.enable_endpoint) {
-      return false;
-    }
-
-    int32_t num_processed_frames = s->GetNumProcessedFrames();
-
-    // frame shift is 10 milliseconds
-    float frame_shift_in_seconds = 0.01;
-
-    int32_t trailing_silence_frames =
-        s->GetResult().num_trailing_blanks * model_->SubsamplingFactor();
-
-    return endpoint_.IsEndpoint(num_processed_frames, trailing_silence_frames,
-                                frame_shift_in_seconds);
-  }
-
-  void Reset(OnlineStream *s) const override {
-    {
-      // segment is incremented only when the last
-      // result is not empty
-      const auto &r = s->GetResult();
-      if (!r.tokens.empty()) {
-        s->GetCurrentSegment() += 1;
-      }
-    }
-
-    s->SetResult({});
-
-    s->SetStates(model_->GetEncoderInitStates());
-
-    s->SetNeMoDecoderStates(model_->GetDecoderInitStates());
-
-    // Note: We only update counters. The underlying audio samples
-    // are not discarded.
-    s->Reset();
-  }
-
-  void DecodeStreams(OnlineStream **ss, int32_t n) const override {
-    int32_t chunk_size = model_->ChunkSize();
-    int32_t chunk_shift = model_->ChunkShift();
-
-    int32_t feature_dim = ss[0]->FeatureDim();
-
-    std::vector<float> features_vec(n * chunk_size * feature_dim);
-    std::vector<std::vector<Ort::Value>> encoder_states(n);
-
-    for (int32_t i = 0; i != n; ++i) {
-      const auto num_processed_frames = ss[i]->GetNumProcessedFrames();
-      std::vector<float> features =
-          ss[i]->GetFrames(num_processed_frames, chunk_size);
-
-      // Question: should num_processed_frames include chunk_shift?
-      ss[i]->GetNumProcessedFrames() += chunk_shift;
-
-      std::copy(features.begin(), features.end(),
-                features_vec.data() + i * chunk_size * feature_dim);
-
-      encoder_states[i] = std::move(ss[i]->GetStates());
-    }
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    std::array<int64_t, 3> x_shape{n, chunk_size, feature_dim};
-
-    Ort::Value x = Ort::Value::CreateTensor(memory_info, features_vec.data(),
-                                            features_vec.size(), x_shape.data(),
-                                            x_shape.size());
-
-    auto states = model_->StackStates(std::move(encoder_states));
-    int32_t num_states = states.size();  // num_states = 3
-    auto t = model_->RunEncoder(std::move(x), std::move(states));
-    // t[0] encoder_out, float tensor, (batch_size, dim, T)
-    // t[1] next states
-
-    std::vector<Ort::Value> out_states;
-    out_states.reserve(num_states);
-
-    for (int32_t k = 1; k != num_states + 1; ++k) {
-      out_states.push_back(std::move(t[k]));
-    }
-
-    auto unstacked_states = model_->UnStackStates(std::move(out_states));
-    for (int32_t i = 0; i != n; ++i) {
-      ss[i]->SetStates(std::move(unstacked_states[i]));
-    }
-
-    Ort::Value encoder_out = Transpose12(model_->Allocator(), &t[0]);
-
-    decoder_->Decode(std::move(encoder_out), ss, n);
-  }
-
-  void InitOnlineStream(OnlineStream *stream) const {
-    // set encoder states
-    stream->SetStates(model_->GetEncoderInitStates());
-
-    // set decoder states
-    stream->SetNeMoDecoderStates(model_->GetDecoderInitStates());
-  }
-
- private:
-  void PostInit() {
-    config_.feat_config.nemo_normalize_type =
-        model_->FeatureNormalizationMethod();
-
-    config_.feat_config.low_freq = 0;
-    // config_.feat_config.high_freq = 8000;
-    config_.feat_config.is_librosa = true;
-    config_.feat_config.remove_dc_offset = false;
-    // config_.feat_config.window_type = "hann";
-    config_.feat_config.dither = 0;
-    config_.feat_config.nemo_normalize_type =
-        model_->FeatureNormalizationMethod();
-
-    int32_t vocab_size = model_->VocabSize();
-
-    // check the blank ID
-    if (!symbol_table_.Contains("<blk>")) {
-      SHERPA_ONNX_LOGE("tokens.txt does not include the blank token <blk>");
-      exit(-1);
-    }
-
-    if (symbol_table_["<blk>"] != vocab_size - 1) {
-      SHERPA_ONNX_LOGE("<blk> is not the last token!");
-      exit(-1);
-    }
-
-    if (symbol_table_.NumSymbols() != vocab_size) {
-      SHERPA_ONNX_LOGE("number of lines in tokens.txt %d != %d (vocab_size)",
-                       symbol_table_.NumSymbols(), vocab_size);
-      exit(-1);
-    }
-  }
-
- private:
-  OnlineRecognizerConfig config_;
-  SymbolTable symbol_table_;
-  std::unique_ptr<OnlineTransducerNeMoModel> model_;
-  std::unique_ptr<OnlineTransducerGreedySearchNeMoDecoder> decoder_;
-  Endpoint endpoint_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_TRANSDUCER_NEMO_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer.h
deleted file mode 100644
index 8854fbd2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-recognizer.h
+++ /dev/null
@@ -1,211 +0,0 @@
-// sherpa-onnx/csrc/online-recognizer.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_H_
-#define SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "sherpa-onnx/csrc/endpoint.h"
-#include "sherpa-onnx/csrc/features.h"
-#include "sherpa-onnx/csrc/online-ctc-fst-decoder-config.h"
-#include "sherpa-onnx/csrc/online-lm-config.h"
-#include "sherpa-onnx/csrc/online-model-config.h"
-#include "sherpa-onnx/csrc/online-stream.h"
-#include "sherpa-onnx/csrc/online-transducer-model-config.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OnlineRecognizerResult {
-  /// Recognition results.
-  /// For English, it consists of space separated words.
-  /// For Chinese, it consists of Chinese words without spaces.
-  /// Example 1: "hello world"
-  /// Example 2: "你好世界"
-  std::string text;
-
-  /// Decoded results at the token level.
-  /// For instance, for BPE-based models it consists of a list of BPE tokens.
-  std::vector<std::string> tokens;
-
-  /// timestamps.size() == tokens.size()
-  /// timestamps[i] records the time in seconds when tokens[i] is decoded.
-  std::vector<float> timestamps;
-
-  std::vector<float> ys_probs;  //< log-prob scores from ASR model
-  std::vector<float> lm_probs;  //< log-prob scores from language model
-                                //
-  /// log-domain scores from "hot-phrase" contextual boosting
-  std::vector<float> context_scores;
-
-  std::vector<int32_t> words;
-
-  /// ID of this segment
-  /// When an endpoint is detected, it is incremented
-  int32_t segment = 0;
-
-  /// Starting time of this segment.
-  /// When an endpoint is detected, it will change
-  float start_time = 0;
-
-  /// True if the end of this segment is reached
-  bool is_final = false;
-
-  /** Return a json string.
-   *
-   * The returned string contains:
-   *   {
-   *     "text": "The recognition result",
-   *     "tokens": [x, x, x],
-   *     "timestamps": [x, x, x],
-   *     "ys_probs": [x, x, x],
-   *     "lm_probs": [x, x, x],
-   *     "context_scores": [x, x, x],
-   *     "segment": x,
-   *     "start_time": x,
-   *     "is_final": true|false
-   *   }
-   */
-  std::string AsJsonString() const;
-};
-
-struct OnlineRecognizerConfig {
-  FeatureExtractorConfig feat_config;
-  OnlineModelConfig model_config;
-  OnlineLMConfig lm_config;
-  EndpointConfig endpoint_config;
-  OnlineCtcFstDecoderConfig ctc_fst_decoder_config;
-  bool enable_endpoint = true;
-
-  std::string decoding_method = "greedy_search";
-  // now support modified_beam_search and greedy_search
-
-  // used only for modified_beam_search
-  int32_t max_active_paths = 4;
-
-  /// used only for modified_beam_search
-  std::string hotwords_file;
-  float hotwords_score = 1.5;
-
-  float blank_penalty = 0.0;
-
-  float temperature_scale = 2.0;
-
-  // If there are multiple rules, they are applied from left to right.
-  std::string rule_fsts;
-
-  // If there are multiple FST archives, they are applied from left to right.
-  std::string rule_fars;
-
-  /// used only for modified_beam_search, if hotwords_buf is non-empty,
-  /// the hotwords will be loaded from the buffered string instead of from the
-  /// "hotwords_file"
-  std::string hotwords_buf;
-
-  OnlineRecognizerConfig() = default;
-
-  OnlineRecognizerConfig(
-      const FeatureExtractorConfig &feat_config,
-      const OnlineModelConfig &model_config, const OnlineLMConfig &lm_config,
-      const EndpointConfig &endpoint_config,
-      const OnlineCtcFstDecoderConfig &ctc_fst_decoder_config,
-      bool enable_endpoint, const std::string &decoding_method,
-      int32_t max_active_paths, const std::string &hotwords_file,
-      float hotwords_score, float blank_penalty, float temperature_scale,
-      const std::string &rule_fsts, const std::string &rule_fars)
-      : feat_config(feat_config),
-        model_config(model_config),
-        lm_config(lm_config),
-        endpoint_config(endpoint_config),
-        ctc_fst_decoder_config(ctc_fst_decoder_config),
-        enable_endpoint(enable_endpoint),
-        decoding_method(decoding_method),
-        max_active_paths(max_active_paths),
-        hotwords_file(hotwords_file),
-        hotwords_score(hotwords_score),
-        blank_penalty(blank_penalty),
-        temperature_scale(temperature_scale),
-        rule_fsts(rule_fsts),
-        rule_fars(rule_fars) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-class OnlineRecognizerImpl;
-
-class OnlineRecognizer {
- public:
-  explicit OnlineRecognizer(const OnlineRecognizerConfig &config);
-
-  template <typename Manager>
-  OnlineRecognizer(Manager *mgr, const OnlineRecognizerConfig &config);
-
-  ~OnlineRecognizer();
-
-  /// Create a stream for decoding.
-  std::unique_ptr<OnlineStream> CreateStream() const;
-
-  /** Create a stream for decoding.
-   *
-   *  @param The hotwords for this string, it might contain several hotwords,
-   *         the hotwords are separated by "/". In each of the hotwords, there
-   *         are cjkchars or bpes, the bpe/cjkchar are separated by space (" ").
-   *         For example, hotwords I LOVE YOU and HELLO WORLD, looks like:
-   *
-   *         "▁I ▁LOVE ▁YOU/▁HE LL O ▁WORLD"
-   */
-  std::unique_ptr<OnlineStream> CreateStream(const std::string &hotwords) const;
-
-  /**
-   * Return true if the given stream has enough frames for decoding.
-   * Return false otherwise
-   */
-  bool IsReady(OnlineStream *s) const;
-
-  /** Decode a single stream. */
-  void DecodeStream(OnlineStream *s) const {
-    OnlineStream *ss[1] = {s};
-    DecodeStreams(ss, 1);
-  }
-
-  /**
-   * Warmups up onnxruntime sessions by apply optimization and
-   * allocating memory prior
-   *
-   * @param warmup Number of warmups.
-   * @param mbs : max-batch-size Max batch size for the models
-   */
-  void WarmpUpRecognizer(int32_t warmup, int32_t mbs) const;
-
-  /** Decode multiple streams in parallel
-   *
-   * @param ss Pointer array containing streams to be decoded.
-   * @param n Number of streams in `ss`.
-   */
-  void DecodeStreams(OnlineStream **ss, int32_t n) const;
-
-  OnlineRecognizerResult GetResult(OnlineStream *s) const;
-
-  // Return true if we detect an endpoint for this stream.
-  // Note: If this function returns true, you usually want to
-  // invoke Reset(s).
-  bool IsEndpoint(OnlineStream *s) const;
-
-  // Clear the state of this stream. If IsEndpoint(s) returns true,
-  // after calling this function, IsEndpoint(s) will return false
-  void Reset(OnlineStream *s) const;
-
- private:
-  std::unique_ptr<OnlineRecognizerImpl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-rnn-lm.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-rnn-lm.h
deleted file mode 100644
index 2e7ad70e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-rnn-lm.h
+++ /dev/null
@@ -1,68 +0,0 @@
-// sherpa-onnx/csrc/online-rnn-lm.h
-//
-// Copyright (c)  2023  Pingfeng Luo
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_RNN_LM_H_
-#define SHERPA_ONNX_CSRC_ONLINE_RNN_LM_H_
-
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/online-lm-config.h"
-#include "sherpa-onnx/csrc/online-lm.h"
-
-namespace sherpa_onnx {
-
-class OnlineRnnLM : public OnlineLM {
- public:
-  ~OnlineRnnLM() override;
-
-  explicit OnlineRnnLM(const OnlineLMConfig &config);
-
-  // init scores for classic rescore
-  std::vector<Ort::Value> GetInitStates() override;
-
-  // init scores for shallow fusion
-  std::pair<Ort::Value, std::vector<Ort::Value>> GetInitStatesSF() override;
-
-   /** ScoreToken a batch of sentences (shallow fusion).
-   *
-   * @param x A 2-D tensor of shape (N, L) with data type int64.
-   * @param states It contains the states for the LM model
-   * @return Return a pair containing
-   *          - log_prob of NN LM
-   *          - updated states
-   *
-   */
-  std::pair<Ort::Value, std::vector<Ort::Value>> ScoreToken(
-      Ort::Value x, std::vector<Ort::Value> states) override;
-
-   /** This function updates hyp.lm_lob_prob of hyps (classic rescore).
-   *
-   * @param scale LM score
-   * @param context_size Context size of the transducer decoder model
-   * @param hyps It is changed in-place.
-   *
-   */
-  void ComputeLMScore(float scale, int32_t context_size,
-                              std::vector<Hypotheses> *hyps) override;
-
-   /** This function updates lm_lob_prob and nn_lm_scores of hyp (shallow fusion).
-   *
-   * @param scale LM score
-   * @param hyps It is changed in-place.
-   *
-   */
-  void ComputeLMScoreSF(float scale, Hypothesis *hyp) override;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_RNN_LM_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-stream.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-stream.h
deleted file mode 100644
index d1183b29..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-stream.h
+++ /dev/null
@@ -1,120 +0,0 @@
-// sherpa-onnx/csrc/online-stream.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_STREAM_H_
-#define SHERPA_ONNX_CSRC_ONLINE_STREAM_H_
-
-#include <memory>
-#include <vector>
-
-#include "kaldi-decoder/csrc/faster-decoder.h"
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/context-graph.h"
-#include "sherpa-onnx/csrc/features.h"
-#include "sherpa-onnx/csrc/online-ctc-decoder.h"
-#include "sherpa-onnx/csrc/online-paraformer-decoder.h"
-#include "sherpa-onnx/csrc/online-transducer-decoder.h"
-
-namespace sherpa_onnx {
-
-struct TransducerKeywordResult;
-class OnlineStream {
- public:
-  explicit OnlineStream(const FeatureExtractorConfig &config = {},
-                        ContextGraphPtr context_graph = nullptr);
-  ~OnlineStream();
-
-  /**
-     @param sampling_rate The sampling_rate of the input waveform. If it does
-                          not equal to  config.sampling_rate, we will do
-                          resampling inside.
-     @param waveform Pointer to a 1-D array of size n. It must be normalized to
-                     the range [-1, 1].
-     @param n Number of entries in waveform
-   */
-  void AcceptWaveform(int32_t sampling_rate, const float *waveform,
-                      int32_t n) const;
-
-  /**
-   * InputFinished() tells the class you won't be providing any
-   * more waveform.  This will help flush out the last frame or two
-   * of features, in the case where snip-edges == false; it also
-   * affects the return value of IsLastFrame().
-   */
-  void InputFinished() const;
-
-  int32_t NumFramesReady() const;
-
-  /** Note: IsLastFrame() will only ever return true if you have called
-   * InputFinished() (and this frame is the last frame).
-   */
-  bool IsLastFrame(int32_t frame) const;
-
-  /** Get n frames starting from the given frame index.
-   *
-   * @param frame_index  The starting frame index
-   * @param n  Number of frames to get.
-   * @return Return a 2-D tensor of shape (n, feature_dim).
-   *         which is flattened into a 1-D vector (flattened in row major)
-   */
-  std::vector<float> GetFrames(int32_t frame_index, int32_t n) const;
-
-  void Reset();
-
-  int32_t FeatureDim() const;
-
-  // Return a reference to the number of processed frames so far
-  // before subsampling..
-  // Initially, it is 0. It is always less than NumFramesReady().
-  //
-  // The returned reference is valid as long as this object is alive.
-  int32_t &GetNumProcessedFrames();  // It's reset after calling Reset()
-
-  int32_t GetNumFramesSinceStart() const;
-
-  int32_t &GetCurrentSegment();
-
-  void SetResult(const OnlineTransducerDecoderResult &r);
-  OnlineTransducerDecoderResult &GetResult();
-
-  void SetKeywordResult(const TransducerKeywordResult &r);
-  TransducerKeywordResult &GetKeywordResult(bool remove_duplicates = false);
-
-  void SetCtcResult(const OnlineCtcDecoderResult &r);
-  OnlineCtcDecoderResult &GetCtcResult();
-
-  void SetParaformerResult(const OnlineParaformerDecoderResult &r);
-  OnlineParaformerDecoderResult &GetParaformerResult();
-
-  void SetStates(std::vector<Ort::Value> states);
-  std::vector<Ort::Value> &GetStates();
-
-  void SetNeMoDecoderStates(std::vector<Ort::Value> decoder_states);
-  std::vector<Ort::Value> &GetNeMoDecoderStates();
-
-  /**
-   * Get the context graph corresponding to this stream.
-   *
-   * @return Return the context graph for this stream.
-   */
-  const ContextGraphPtr &GetContextGraph() const;
-
-  // for online ctc decoder
-  void SetFasterDecoder(std::unique_ptr<kaldi_decoder::FasterDecoder> decoder);
-  kaldi_decoder::FasterDecoder *GetFasterDecoder() const;
-  int32_t &GetFasterDecoderProcessedFrames();
-
-  // for streaming paraformer
-  std::vector<float> &GetParaformerFeatCache();
-  std::vector<float> &GetParaformerEncoderOutCache();
-  std::vector<float> &GetParaformerAlphaCache();
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_STREAM_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-decoder.h
deleted file mode 100644
index e507a0fc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-decoder.h
+++ /dev/null
@@ -1,111 +0,0 @@
-// sherpa-onnx/csrc/online-transducer-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_DECODER_H_
-#define SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_DECODER_H_
-
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/hypothesis.h"
-#include "sherpa-onnx/csrc/macros.h"
-
-namespace sherpa_onnx {
-
-struct OnlineTransducerDecoderResult {
-  /// Number of frames after subsampling we have decoded so far
-  int32_t frame_offset = 0;
-
-  /// The decoded token IDs so far
-  std::vector<int64_t> tokens;
-
-  /// number of trailing blank frames decoded so far
-  int32_t num_trailing_blanks = 0;
-
-  /// timestamps[i] contains the output frame index where tokens[i] is decoded.
-  std::vector<int32_t> timestamps;
-
-  std::vector<float> ys_probs;
-  std::vector<float> lm_probs;
-  std::vector<float> context_scores;
-
-  // Cache decoder_out for endpointing
-  Ort::Value decoder_out;
-
-  // used only in modified beam_search
-  Hypotheses hyps;
-
-  OnlineTransducerDecoderResult()
-      : tokens{}, num_trailing_blanks(0), decoder_out{nullptr}, hyps{} {}
-
-  OnlineTransducerDecoderResult(const OnlineTransducerDecoderResult &other);
-
-  OnlineTransducerDecoderResult &operator=(
-      const OnlineTransducerDecoderResult &other);
-
-  OnlineTransducerDecoderResult(OnlineTransducerDecoderResult &&other) noexcept;
-
-  OnlineTransducerDecoderResult &operator=(
-      OnlineTransducerDecoderResult &&other) noexcept;
-};
-
-class OnlineStream;
-class OnlineTransducerDecoder {
- public:
-  virtual ~OnlineTransducerDecoder() = default;
-
-  /* Return an empty result.
-   *
-   * To simplify the decoding code, we add `context_size` blanks
-   * to the beginning of the decoding result, which will be
-   * stripped by calling `StripPrecedingBlanks()`.
-   */
-  virtual OnlineTransducerDecoderResult GetEmptyResult() const = 0;
-
-  /** Strip blanks added by `GetEmptyResult()`.
-   *
-   * @param r It is changed in-place.
-   */
-  virtual void StripLeadingBlanks(OnlineTransducerDecoderResult * /*r*/) const {
-  }
-
-  /** Run transducer beam search given the output from the encoder model.
-   *
-   * @param encoder_out A 3-D tensor of shape (N, T, joiner_dim)
-   * @param result  It is modified in-place.
-   *
-   * @note There is no need to pass encoder_out_length here since for the
-   * online decoding case, each utterance has the same number of frames
-   * and there are no paddings.
-   */
-  virtual void Decode(Ort::Value encoder_out,
-                      std::vector<OnlineTransducerDecoderResult> *result) = 0;
-
-  /** Run transducer beam search given the output from the encoder model.
-   *
-   * Note: Currently this interface is for contextual-biasing feature which
-   *       needs a ContextGraph owned by the OnlineStream.
-   *
-   * @param encoder_out A 3-D tensor of shape (N, T, joiner_dim)
-   * @param ss  A list of OnlineStreams.
-   * @param result  It is modified in-place.
-   *
-   * @note There is no need to pass encoder_out_length here since for the
-   * online decoding case, each utterance has the same number of frames
-   * and there are no paddings.
-   */
-  virtual void Decode(Ort::Value /*encoder_out*/, OnlineStream ** /*ss*/,
-                      std::vector<OnlineTransducerDecoderResult> * /*result*/) {
-    SHERPA_ONNX_LOGE(
-        "This interface is for OnlineTransducerModifiedBeamSearchDecoder.");
-    exit(-1);
-  }
-
-  // used for endpointing. We need to keep decoder_out after reset
-  virtual void UpdateDecoderOut(OnlineTransducerDecoderResult * /*result*/) {}
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-greedy-search-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-greedy-search-decoder.h
deleted file mode 100644
index 716f8848..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-greedy-search-decoder.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// sherpa-onnx/csrc/online-transducer-greedy-search-decoder.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_GREEDY_SEARCH_DECODER_H_
-#define SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_GREEDY_SEARCH_DECODER_H_
-
-#include <vector>
-
-#include "sherpa-onnx/csrc/online-transducer-decoder.h"
-#include "sherpa-onnx/csrc/online-transducer-model.h"
-
-namespace sherpa_onnx {
-
-class OnlineTransducerGreedySearchDecoder : public OnlineTransducerDecoder {
- public:
-  OnlineTransducerGreedySearchDecoder(OnlineTransducerModel *model,
-                                      int32_t unk_id,
-                                      float blank_penalty,
-                                      float temperature_scale)
-      : model_(model),
-      unk_id_(unk_id),
-      blank_penalty_(blank_penalty),
-      temperature_scale_(temperature_scale) {}
-
-  OnlineTransducerDecoderResult GetEmptyResult() const override;
-
-  void StripLeadingBlanks(OnlineTransducerDecoderResult *r) const override;
-
-  void Decode(Ort::Value encoder_out,
-              std::vector<OnlineTransducerDecoderResult> *result) override;
-
- private:
-  OnlineTransducerModel *model_;  // Not owned
-  int32_t unk_id_;
-  float blank_penalty_;
-  float temperature_scale_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_GREEDY_SEARCH_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-greedy-search-nemo-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-greedy-search-nemo-decoder.h
deleted file mode 100644
index 212008fd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-greedy-search-nemo-decoder.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// sherpa-onnx/csrc/online-transducer-greedy-search-nemo-decoder.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-// Copyright (c)  2024  Sangeet Sagar
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_GREEDY_SEARCH_NEMO_DECODER_H_
-#define SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_GREEDY_SEARCH_NEMO_DECODER_H_
-
-#include <vector>
-
-#include "sherpa-onnx/csrc/online-transducer-decoder.h"
-#include "sherpa-onnx/csrc/online-transducer-nemo-model.h"
-
-namespace sherpa_onnx {
-
-class OnlineStream;
-
-class OnlineTransducerGreedySearchNeMoDecoder {
- public:
-  OnlineTransducerGreedySearchNeMoDecoder(OnlineTransducerNeMoModel *model,
-                                          float blank_penalty)
-      : model_(model), blank_penalty_(blank_penalty) {}
-
-  // @param n number of elements in ss
-  void Decode(Ort::Value encoder_out, OnlineStream **ss, int32_t n) const;
-
- private:
-  OnlineTransducerNeMoModel *model_;  // Not owned
-  float blank_penalty_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_GREEDY_SEARCH_NEMO_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-model-config.h
deleted file mode 100644
index 5d79e25b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-model-config.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// sherpa-onnx/csrc/online-transducer-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OnlineTransducerModelConfig {
-  std::string encoder;
-  std::string decoder;
-  std::string joiner;
-
-  OnlineTransducerModelConfig() = default;
-  OnlineTransducerModelConfig(const std::string &encoder,
-                              const std::string &decoder,
-                              const std::string &joiner)
-      : encoder(encoder), decoder(decoder), joiner(joiner) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-model.h
deleted file mode 100644
index f6404ecc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-model.h
+++ /dev/null
@@ -1,145 +0,0 @@
-// sherpa-onnx/csrc/online-transducer-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_MODEL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_MODEL_H_
-
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/hypothesis.h"
-#include "sherpa-onnx/csrc/online-model-config.h"
-#include "sherpa-onnx/csrc/online-transducer-decoder.h"
-#include "sherpa-onnx/csrc/online-transducer-model-config.h"
-
-namespace sherpa_onnx {
-
-struct OnlineTransducerDecoderResult;
-
-class OnlineTransducerModel {
- public:
-  virtual ~OnlineTransducerModel() = default;
-
-  static std::unique_ptr<OnlineTransducerModel> Create(
-      const OnlineModelConfig &config);
-
-  template <typename Manager>
-  static std::unique_ptr<OnlineTransducerModel> Create(
-      Manager *mgr, const OnlineModelConfig &config);
-
-  /** Stack a list of individual states into a batch.
-   *
-   * It is the inverse operation of `UnStackStates`.
-   *
-   * @param states states[i] contains the state for the i-th utterance.
-   * @return Return a single value representing the batched state.
-   */
-  virtual std::vector<Ort::Value> StackStates(
-      const std::vector<std::vector<Ort::Value>> &states) const = 0;
-
-  /** Unstack a batch state into a list of individual states.
-   *
-   * It is the inverse operation of `StackStates`.
-   *
-   * @param states A batched state.
-   * @return ans[i] contains the state for the i-th utterance.
-   */
-  virtual std::vector<std::vector<Ort::Value>> UnStackStates(
-      const std::vector<Ort::Value> &states) const = 0;
-
-  /** Get the initial encoder states.
-   *
-   * @return Return the initial encoder state.
-   */
-  virtual std::vector<Ort::Value> GetEncoderInitStates() = 0;
-
-  /** Set feature dim.
-   *
-   * This is used in `OnlineZipformer2TransducerModel`,
-   * to pass `feature_dim` for `GetEncoderInitStates()`.
-   *
-   * This has to be called before GetEncoderInitStates(), so the `encoder_embed`
-   * init state has the correct `embed_dim` of its output.
-   */
-  virtual void SetFeatureDim(int32_t /*feature_dim*/) {}
-
-  /** Run the encoder.
-   *
-   * @param features  A tensor of shape (N, T, C). It is changed in-place.
-   * @param states  Encoder state of the previous chunk. It is changed in-place.
-   * @param processed_frames  Processed frames before subsampling. It is a 1-D
-   * tensor with data type int64_t.
-   *
-   * @return Return a tuple containing:
-   *           - encoder_out, a tensor of shape (N, T', encoder_out_dim)
-   *           - next_states  Encoder state for the next chunk.
-   */
-  virtual std::pair<Ort::Value, std::vector<Ort::Value>> RunEncoder(
-      Ort::Value features, std::vector<Ort::Value> states,
-      Ort::Value processed_frames) = 0;  // NOLINT
-
-  /** Run the decoder network.
-   *
-   * Caution: We assume there are no recurrent connections in the decoder and
-   *          the decoder is stateless. See
-   * https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless2/decoder.py
-   *          for an example
-   *
-   * @param decoder_input It is usually of shape (N, context_size)
-   * @return Return a tensor of shape (N, decoder_dim).
-   */
-  virtual Ort::Value RunDecoder(Ort::Value decoder_input) = 0;
-
-  /** Run the joint network.
-   *
-   * @param encoder_out Output of the encoder network. A tensor of shape
-   *                    (N, joiner_dim).
-   * @param decoder_out Output of the decoder network. A tensor of shape
-   *                    (N, joiner_dim).
-   * @return Return a tensor of shape (N, vocab_size). In icefall, the last
-   *         last layer of the joint network is `nn.Linear`,
-   *         not `nn.LogSoftmax`.
-   */
-  virtual Ort::Value RunJoiner(Ort::Value encoder_out,
-                               Ort::Value decoder_out) = 0;
-
-  /** If we are using a stateless decoder and if it contains a
-   *  Conv1D, this function returns the kernel size of the convolution layer.
-   */
-  virtual int32_t ContextSize() const = 0;
-
-  /** We send this number of feature frames to the encoder at a time. */
-  virtual int32_t ChunkSize() const = 0;
-
-  /** Number of input frames to discard after each call to RunEncoder.
-   *
-   * For instance, if we have 30 frames, chunk_size=8, chunk_shift=6.
-   *
-   * In the first call of RunEncoder, we use frames 0~7 since chunk_size is 8.
-   * Then we discard frame 0~5 since chunk_shift is 6.
-   * In the second call of RunEncoder, we use frames 6~13; and then we discard
-   * frames 6~11.
-   * In the third call of RunEncoder, we use frames 12~19; and then we discard
-   * frames 12~16.
-   *
-   * Note: ChunkSize() - ChunkShift() == right context size
-   */
-  virtual int32_t ChunkShift() const = 0;
-
-  virtual int32_t VocabSize() const = 0;
-
-  virtual int32_t SubsamplingFactor() const { return 4; }
-
-  virtual OrtAllocator *Allocator() = 0;
-
-  Ort::Value BuildDecoderInput(
-      const std::vector<OnlineTransducerDecoderResult> &results);
-
-  Ort::Value BuildDecoderInput(const std::vector<Hypothesis> &hyps);
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-modified-beam-search-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-modified-beam-search-decoder.h
deleted file mode 100644
index 6dea71be..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-modified-beam-search-decoder.h
+++ /dev/null
@@ -1,64 +0,0 @@
-// sherpa-onnx/csrc/online-transducer-modified_beam-search-decoder.h
-//
-// Copyright (c)  2023  Pingfeng Luo
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_MODIFIED_BEAM_SEARCH_DECODER_H_
-#define SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_MODIFIED_BEAM_SEARCH_DECODER_H_
-
-#include <vector>
-
-#include "sherpa-onnx/csrc/online-lm.h"
-#include "sherpa-onnx/csrc/online-stream.h"
-#include "sherpa-onnx/csrc/online-transducer-decoder.h"
-#include "sherpa-onnx/csrc/online-transducer-model.h"
-
-namespace sherpa_onnx {
-
-class OnlineTransducerModifiedBeamSearchDecoder
-    : public OnlineTransducerDecoder {
- public:
-  OnlineTransducerModifiedBeamSearchDecoder(OnlineTransducerModel *model,
-                                            OnlineLM *lm,
-                                            int32_t max_active_paths,
-                                            float lm_scale,
-                                            bool shallow_fusion,
-                                            int32_t unk_id,
-                                            float blank_penalty,
-                                            float temperature_scale)
-      : model_(model),
-        lm_(lm),
-        max_active_paths_(max_active_paths),
-        lm_scale_(lm_scale),
-        shallow_fusion_(shallow_fusion),
-        unk_id_(unk_id),
-        blank_penalty_(blank_penalty),
-        temperature_scale_(temperature_scale) {}
-
-  OnlineTransducerDecoderResult GetEmptyResult() const override;
-
-  void StripLeadingBlanks(OnlineTransducerDecoderResult *r) const override;
-
-  void Decode(Ort::Value encoder_out,
-              std::vector<OnlineTransducerDecoderResult> *result) override;
-
-  void Decode(Ort::Value encoder_out, OnlineStream **ss,
-              std::vector<OnlineTransducerDecoderResult> *result) override;
-
-  void UpdateDecoderOut(OnlineTransducerDecoderResult *result) override;
-
- private:
-  OnlineTransducerModel *model_;  // Not owned
-  OnlineLM *lm_;                  // Not owned
-
-  int32_t max_active_paths_;
-  float lm_scale_;  // used only when lm_ is not nullptr
-  bool shallow_fusion_;  // used only when lm_ is not nullptr
-  int32_t unk_id_;
-  float blank_penalty_;
-  float temperature_scale_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_MODIFIED_BEAM_SEARCH_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-nemo-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-nemo-model.h
deleted file mode 100644
index 98390bb1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-transducer-nemo-model.h
+++ /dev/null
@@ -1,124 +0,0 @@
-// sherpa-onnx/csrc/online-transducer-nemo-model.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-// Copyright (c)  2024  Sangeet Sagar
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_NEMO_MODEL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_NEMO_MODEL_H_
-
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/online-model-config.h"
-
-namespace sherpa_onnx {
-
-// see
-// https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/models/hybrid_rnnt_ctc_bpe_models.py#L40
-// Its decoder is stateful, not stateless.
-class OnlineTransducerNeMoModel {
- public:
-  explicit OnlineTransducerNeMoModel(const OnlineModelConfig &config);
-
-  template <typename Manager>
-  OnlineTransducerNeMoModel(Manager *mgr, const OnlineModelConfig &config);
-
-  ~OnlineTransducerNeMoModel();
-  // A list of 3 tensors:
-  //  - cache_last_channel
-  //  - cache_last_time
-  //  - cache_last_channel_len
-  std::vector<Ort::Value> GetEncoderInitStates() const;
-
-  // stack encoder states
-  std::vector<Ort::Value> StackStates(
-      std::vector<std::vector<Ort::Value>> states) const;
-
-  // unstack encoder states
-  std::vector<std::vector<Ort::Value>> UnStackStates(
-      std::vector<Ort::Value> states) const;
-
-  /** Run the encoder.
-   *
-   * @param features  A tensor of shape (N, T, C). It is changed in-place.
-   * @param states  It is from GetEncoderInitStates() or returned from this
-   *                method.
-   *
-   * @return Return a tuple containing:
-   *           - ans[0]: encoder_out, a tensor of shape (N, encoder_out_dim, T')
-   *           - ans[1:]: contains next states
-   */
-  std::vector<Ort::Value> RunEncoder(
-      Ort::Value features, std::vector<Ort::Value> states) const;  // NOLINT
-
-  /** Run the decoder network.
-   *
-   * @param targets A int32 tensor of shape (batch_size, 1)
-   * @param states The states for the decoder model.
-   * @return Return a vector:
-   *           - ans[0] is the decoder_out (a float tensor)
-   *           - ans[1:] is the next states
-   */
-  std::pair<Ort::Value, std::vector<Ort::Value>> RunDecoder(
-      Ort::Value targets, std::vector<Ort::Value> states) const;
-
-  std::vector<Ort::Value> GetDecoderInitStates() const;
-
-  /** Run the joint network.
-   *
-   * @param encoder_out Output of the encoder network.
-   * @param decoder_out Output of the decoder network.
-   * @return Return a tensor of shape (N, 1, 1, vocab_size) containing logits.
-   */
-  Ort::Value RunJoiner(Ort::Value encoder_out, Ort::Value decoder_out) const;
-
-  /** We send this number of feature frames to the encoder at a time. */
-  int32_t ChunkSize() const;
-
-  /** Number of input frames to discard after each call to RunEncoder.
-   *
-   * For instance, if we have 30 frames, chunk_size=8, chunk_shift=6.
-   *
-   * In the first call of RunEncoder, we use frames 0~7 since chunk_size is 8.
-   * Then we discard frame 0~5 since chunk_shift is 6.
-   * In the second call of RunEncoder, we use frames 6~13; and then we discard
-   * frames 6~11.
-   * In the third call of RunEncoder, we use frames 12~19; and then we discard
-   * frames 12~16.
-   *
-   * Note: ChunkSize() - ChunkShift() == right context size
-   */
-  int32_t ChunkShift() const;
-
-  /** Return the subsampling factor of the model.
-   */
-  int32_t SubsamplingFactor() const;
-
-  int32_t VocabSize() const;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const;
-
-  // Possible values:
-  // - per_feature
-  // - all_features (not implemented yet)
-  // - fixed_mean (not implemented)
-  // - fixed_std (not implemented)
-  // - or just leave it to empty
-  // See
-  // https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/parts/preprocessing/features.py#L59
-  // for details
-  std::string FeatureNormalizationMethod() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_TRANSDUCER_NEMO_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-websocket-server-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-websocket-server-impl.h
deleted file mode 100644
index 4e0582db..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-websocket-server-impl.h
+++ /dev/null
@@ -1,181 +0,0 @@
-// sherpa-onnx/csrc/online-websocket-server-impl.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_ONLINE_WEBSOCKET_SERVER_IMPL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_WEBSOCKET_SERVER_IMPL_H_
-
-#include <deque>
-#include <fstream>
-#include <map>
-#include <memory>
-#include <mutex>  // NOLINT
-#include <set>
-#include <string>
-#include <unordered_set>
-#include <utility>
-#include <vector>
-
-#include "asio.hpp"
-#include "sherpa-onnx/csrc/online-recognizer.h"
-#include "sherpa-onnx/csrc/online-stream.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-#include "sherpa-onnx/csrc/tee-stream.h"
-#include "websocketpp/config/asio_no_tls.hpp"  // TODO(fangjun): support TLS
-#include "websocketpp/server.hpp"
-using server = websocketpp::server<websocketpp::config::asio>;
-using connection_hdl = websocketpp::connection_hdl;
-
-namespace sherpa_onnx {
-
-struct Connection {
-  // handle to the connection. We can use it to send messages to the client
-  connection_hdl hdl;
-  std::shared_ptr<OnlineStream> s;
-
-  // set it to true when InputFinished() is called
-  bool eof = false;
-
-  // The last time we received a message from the client
-  // TODO(fangjun): Use it to disconnect from a client if it is inactive
-  // for a specified time.
-  std::chrono::steady_clock::time_point last_active;
-
-  std::mutex mutex;  // protect samples
-
-  // Audio samples received from the client.
-  //
-  // The I/O threads receive audio samples into this queue
-  // and invoke work threads to compute features
-  std::deque<std::vector<float>> samples;
-
-  Connection() = default;
-  Connection(connection_hdl hdl, std::shared_ptr<OnlineStream> s)
-      : hdl(hdl), s(s), last_active(std::chrono::steady_clock::now()) {}
-};
-
-struct OnlineWebsocketDecoderConfig {
-  OnlineRecognizerConfig recognizer_config;
-
-  // It determines how often the decoder loop runs.
-  int32_t loop_interval_ms = 10;
-
-  int32_t max_batch_size = 5;
-
-  float end_tail_padding = 0.8;
-
-  void Register(ParseOptions *po);
-  void Validate() const;
-};
-
-class OnlineWebsocketServer;
-
-class OnlineWebsocketDecoder {
- public:
-  /**
-   * @param server  Not owned.
-   */
-  explicit OnlineWebsocketDecoder(OnlineWebsocketServer *server);
-
-  std::shared_ptr<Connection> GetOrCreateConnection(connection_hdl hdl);
-
-  // Compute features for a stream given audio samples
-  void AcceptWaveform(std::shared_ptr<Connection> c);
-
-  // signal that there will be no more audio samples for a stream
-  void InputFinished(std::shared_ptr<Connection> c);
-
-  void Warmup() const;
-
-  void Run();
-
- private:
-  void ProcessConnections(const asio::error_code &ec);
-
-  /** It is called by one of the worker thread.
-   */
-  void Decode();
-
- private:
-  OnlineWebsocketServer *server_;  // not owned
-  std::unique_ptr<OnlineRecognizer> recognizer_;
-  OnlineWebsocketDecoderConfig config_;
-  asio::steady_timer timer_;
-
-  // It protects `connections_`, `ready_connections_`, and `active_`
-  std::mutex mutex_;
-
-  std::map<connection_hdl, std::shared_ptr<Connection>,
-           std::owner_less<connection_hdl>>
-      connections_;
-
-  // Whenever a connection has enough feature frames for decoding, we put
-  // it in this queue
-  std::deque<std::shared_ptr<Connection>> ready_connections_;
-
-  // If we are decoding a stream, we put it in the active_ set so that
-  // only one thread can decode a stream at a time.
-  std::set<connection_hdl, std::owner_less<connection_hdl>> active_;
-};
-
-struct OnlineWebsocketServerConfig {
-  OnlineWebsocketDecoderConfig decoder_config;
-
-  std::string log_file = "./log.txt";
-
-  void Register(sherpa_onnx::ParseOptions *po);
-  void Validate() const;
-};
-
-class OnlineWebsocketServer {
- public:
-  explicit OnlineWebsocketServer(asio::io_context &io_conn,  // NOLINT
-                                 asio::io_context &io_work,  // NOLINT
-                                 const OnlineWebsocketServerConfig &config);
-
-  void Run(uint16_t port);
-
-  const OnlineWebsocketServerConfig &GetConfig() const { return config_; }
-  asio::io_context &GetConnectionContext() { return io_conn_; }
-  asio::io_context &GetWorkContext() { return io_work_; }
-  server &GetServer() { return server_; }
-
-  void Send(connection_hdl hdl, const std::string &text);
-
-  bool Contains(connection_hdl hdl) const;
-
- private:
-  void SetupLog();
-
-  // When a websocket client is connected, it will invoke this method
-  // (Not for HTTP)
-  void OnOpen(connection_hdl hdl);
-
-  // When a websocket client is disconnected, it will invoke this method
-  void OnClose(connection_hdl hdl);
-
-  void OnMessage(connection_hdl hdl, server::message_ptr msg);
-
-  // Close a websocket connection with given code and reason
-  void Close(connection_hdl hdl, websocketpp::close::status::value code,
-             const std::string &reason);
-
- private:
-  OnlineWebsocketServerConfig config_;
-  asio::io_context &io_conn_;
-  asio::io_context &io_work_;
-  server server_;
-
-  std::ofstream log_;
-  sherpa_onnx::TeeStream tee_;
-
-  OnlineWebsocketDecoder decoder_;
-
-  mutable std::mutex mutex_;
-
-  std::set<connection_hdl, std::owner_less<connection_hdl>> connections_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_WEBSOCKET_SERVER_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-wenet-ctc-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-wenet-ctc-model-config.h
deleted file mode 100644
index b90295ea..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-wenet-ctc-model-config.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// sherpa-onnx/csrc/online-wenet-ctc-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_WENET_CTC_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_ONLINE_WENET_CTC_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OnlineWenetCtcModelConfig {
-  std::string model;
-
-  // --chunk_size from wenet
-  int32_t chunk_size = 16;
-
-  // --num_left_chunks from wenet
-  int32_t num_left_chunks = 4;
-
-  OnlineWenetCtcModelConfig() = default;
-
-  OnlineWenetCtcModelConfig(const std::string &model, int32_t chunk_size,
-                            int32_t num_left_chunks)
-      : model(model),
-        chunk_size(chunk_size),
-        num_left_chunks(num_left_chunks) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_WENET_CTC_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-wenet-ctc-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-wenet-ctc-model.h
deleted file mode 100644
index 28458b68..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-wenet-ctc-model.h
+++ /dev/null
@@ -1,75 +0,0 @@
-// sherpa-onnx/csrc/online-wenet-ctc-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_WENET_CTC_MODEL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_WENET_CTC_MODEL_H_
-
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/online-ctc-model.h"
-#include "sherpa-onnx/csrc/online-model-config.h"
-
-namespace sherpa_onnx {
-
-class OnlineWenetCtcModel : public OnlineCtcModel {
- public:
-  explicit OnlineWenetCtcModel(const OnlineModelConfig &config);
-
-  template <typename Manager>
-  OnlineWenetCtcModel(Manager *mgr, const OnlineModelConfig &config);
-
-  ~OnlineWenetCtcModel() override;
-
-  // A list of 3 tensors:
-  //  - attn_cache
-  //  - conv_cache
-  //  - offset
-  std::vector<Ort::Value> GetInitStates() const override;
-
-  std::vector<Ort::Value> StackStates(
-      std::vector<std::vector<Ort::Value>> states) const override;
-
-  std::vector<std::vector<Ort::Value>> UnStackStates(
-      std::vector<Ort::Value> states) const override;
-
-  /**
-   *
-   * @param x A 3-D tensor of shape (N, T, C). N has to be 1.
-   * @param states  It is from GetInitStates() or returned from this method.
-   *
-   * @return Return a list of tensors
-   *    - ans[0] contains log_probs, of shape (N, T, C)
-   *    - ans[1:] contains next_states
-   */
-  std::vector<Ort::Value> Forward(
-      Ort::Value x, std::vector<Ort::Value> states) const override;
-
-  /** Return the vocabulary size of the model
-   */
-  int32_t VocabSize() const override;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const override;
-
-  // The model accepts this number of frames before subsampling as input
-  int32_t ChunkLength() const override;
-
-  // Similar to frame_shift in feature extractor, after processing
-  // ChunkLength() frames, we advance by ChunkShift() frames
-  // before we process the next chunk.
-  int32_t ChunkShift() const override;
-
-  bool SupportBatchProcessing() const override { return false; }
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_WENET_CTC_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer-transducer-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer-transducer-model.h
deleted file mode 100644
index 9e4368a6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer-transducer-model.h
+++ /dev/null
@@ -1,99 +0,0 @@
-// sherpa-onnx/csrc/online-zipformer-transducer-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_ZIPFORMER_TRANSDUCER_MODEL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_ZIPFORMER_TRANSDUCER_MODEL_H_
-
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/online-model-config.h"
-#include "sherpa-onnx/csrc/online-transducer-model.h"
-
-namespace sherpa_onnx {
-
-class OnlineZipformerTransducerModel : public OnlineTransducerModel {
- public:
-  explicit OnlineZipformerTransducerModel(const OnlineModelConfig &config);
-
-  template <typename Manager>
-  OnlineZipformerTransducerModel(Manager *mgr, const OnlineModelConfig &config);
-
-  std::vector<Ort::Value> StackStates(
-      const std::vector<std::vector<Ort::Value>> &states) const override;
-
-  std::vector<std::vector<Ort::Value>> UnStackStates(
-      const std::vector<Ort::Value> &states) const override;
-
-  std::vector<Ort::Value> GetEncoderInitStates() override;
-
-  std::pair<Ort::Value, std::vector<Ort::Value>> RunEncoder(
-      Ort::Value features, std::vector<Ort::Value> states,
-      Ort::Value processed_frames) override;
-
-  Ort::Value RunDecoder(Ort::Value decoder_input) override;
-
-  Ort::Value RunJoiner(Ort::Value encoder_out, Ort::Value decoder_out) override;
-
-  int32_t ContextSize() const override { return context_size_; }
-
-  int32_t ChunkSize() const override { return T_; }
-
-  int32_t ChunkShift() const override { return decode_chunk_len_; }
-
-  int32_t VocabSize() const override { return vocab_size_; }
-  OrtAllocator *Allocator() override { return allocator_; }
-
- private:
-  void InitEncoder(void *model_data, size_t model_data_length);
-  void InitDecoder(void *model_data, size_t model_data_length);
-  void InitJoiner(void *model_data, size_t model_data_length);
-
- private:
-  Ort::Env env_;
-  Ort::SessionOptions sess_opts_;
-  Ort::AllocatorWithDefaultOptions allocator_;
-
-  std::unique_ptr<Ort::Session> encoder_sess_;
-  std::unique_ptr<Ort::Session> decoder_sess_;
-  std::unique_ptr<Ort::Session> joiner_sess_;
-
-  std::vector<std::string> encoder_input_names_;
-  std::vector<const char *> encoder_input_names_ptr_;
-
-  std::vector<std::string> encoder_output_names_;
-  std::vector<const char *> encoder_output_names_ptr_;
-
-  std::vector<std::string> decoder_input_names_;
-  std::vector<const char *> decoder_input_names_ptr_;
-
-  std::vector<std::string> decoder_output_names_;
-  std::vector<const char *> decoder_output_names_ptr_;
-
-  std::vector<std::string> joiner_input_names_;
-  std::vector<const char *> joiner_input_names_ptr_;
-
-  std::vector<std::string> joiner_output_names_;
-  std::vector<const char *> joiner_output_names_ptr_;
-
-  OnlineModelConfig config_;
-
-  std::vector<int32_t> encoder_dims_;
-  std::vector<int32_t> attention_dims_;
-  std::vector<int32_t> num_encoder_layers_;
-  std::vector<int32_t> cnn_module_kernels_;
-  std::vector<int32_t> left_context_len_;
-
-  int32_t T_ = 0;
-  int32_t decode_chunk_len_ = 0;
-
-  int32_t context_size_ = 0;
-  int32_t vocab_size_ = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_ZIPFORMER_TRANSDUCER_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer2-ctc-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer2-ctc-model-config.h
deleted file mode 100644
index 18115c8f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer2-ctc-model-config.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// sherpa-onnx/csrc/online-zipformer2-ctc-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_ZIPFORMER2_CTC_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_ONLINE_ZIPFORMER2_CTC_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct OnlineZipformer2CtcModelConfig {
-  std::string model;
-
-  OnlineZipformer2CtcModelConfig() = default;
-
-  explicit OnlineZipformer2CtcModelConfig(const std::string &model)
-      : model(model) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_ZIPFORMER2_CTC_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer2-ctc-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer2-ctc-model.h
deleted file mode 100644
index 32ddf212..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer2-ctc-model.h
+++ /dev/null
@@ -1,74 +0,0 @@
-// sherpa-onnx/csrc/online-zipformer2-ctc-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_ZIPFORMER2_CTC_MODEL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_ZIPFORMER2_CTC_MODEL_H_
-
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/online-ctc-model.h"
-#include "sherpa-onnx/csrc/online-model-config.h"
-
-namespace sherpa_onnx {
-
-class OnlineZipformer2CtcModel : public OnlineCtcModel {
- public:
-  explicit OnlineZipformer2CtcModel(const OnlineModelConfig &config);
-
-  template <typename Manager>
-  OnlineZipformer2CtcModel(Manager *mgr, const OnlineModelConfig &config);
-
-  ~OnlineZipformer2CtcModel() override;
-
-  // A list of tensors.
-  // See also
-  // https://github.com/k2-fsa/icefall/pull/1413
-  // and
-  // https://github.com/k2-fsa/icefall/pull/1415
-  std::vector<Ort::Value> GetInitStates() const override;
-
-  std::vector<Ort::Value> StackStates(
-      std::vector<std::vector<Ort::Value>> states) const override;
-
-  std::vector<std::vector<Ort::Value>> UnStackStates(
-      std::vector<Ort::Value> states) const override;
-
-  /**
-   *
-   * @param x A 3-D tensor of shape (N, T, C). N has to be 1.
-   * @param states  It is from GetInitStates() or returned from this method.
-   *
-   * @return Return a list of tensors
-   *    - ans[0] contains log_probs, of shape (N, T, C)
-   *    - ans[1:] contains next_states
-   */
-  std::vector<Ort::Value> Forward(
-      Ort::Value x, std::vector<Ort::Value> states) const override;
-
-  /** Return the vocabulary size of the model
-   */
-  int32_t VocabSize() const override;
-
-  /** Return an allocator for allocating memory
-   */
-  OrtAllocator *Allocator() const override;
-
-  // The model accepts this number of frames before subsampling as input
-  int32_t ChunkLength() const override;
-
-  // Similar to frame_shift in feature extractor, after processing
-  // ChunkLength() frames, we advance by ChunkShift() frames
-  // before we process the next chunk.
-  int32_t ChunkShift() const override;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_ZIPFORMER2_CTC_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer2-transducer-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer2-transducer-model.h
deleted file mode 100644
index 93b124f6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/online-zipformer2-transducer-model.h
+++ /dev/null
@@ -1,110 +0,0 @@
-// sherpa-onnx/csrc/online-zipformer2-transducer-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_ONLINE_ZIPFORMER2_TRANSDUCER_MODEL_H_
-#define SHERPA_ONNX_CSRC_ONLINE_ZIPFORMER2_TRANSDUCER_MODEL_H_
-
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/online-model-config.h"
-#include "sherpa-onnx/csrc/online-transducer-model.h"
-
-namespace sherpa_onnx {
-
-class OnlineZipformer2TransducerModel : public OnlineTransducerModel {
- public:
-  explicit OnlineZipformer2TransducerModel(const OnlineModelConfig &config);
-
-  template <typename Manager>
-  OnlineZipformer2TransducerModel(Manager *mgr,
-                                  const OnlineModelConfig &config);
-
-  std::vector<Ort::Value> StackStates(
-      const std::vector<std::vector<Ort::Value>> &states) const override;
-
-  std::vector<std::vector<Ort::Value>> UnStackStates(
-      const std::vector<Ort::Value> &states) const override;
-
-  std::vector<Ort::Value> GetEncoderInitStates() override;
-
-  void SetFeatureDim(int32_t feature_dim) override {
-    feature_dim_ = feature_dim;
-  }
-
-  std::pair<Ort::Value, std::vector<Ort::Value>> RunEncoder(
-      Ort::Value features, std::vector<Ort::Value> states,
-      Ort::Value processed_frames) override;
-
-  Ort::Value RunDecoder(Ort::Value decoder_input) override;
-
-  Ort::Value RunJoiner(Ort::Value encoder_out, Ort::Value decoder_out) override;
-
-  int32_t ContextSize() const override { return context_size_; }
-
-  int32_t ChunkSize() const override { return T_; }
-
-  int32_t ChunkShift() const override { return decode_chunk_len_; }
-
-  int32_t VocabSize() const override { return vocab_size_; }
-  OrtAllocator *Allocator() override { return allocator_; }
-
- private:
-  void InitEncoder(void *model_data, size_t model_data_length);
-  void InitDecoder(void *model_data, size_t model_data_length);
-  void InitJoiner(void *model_data, size_t model_data_length);
-
- private:
-  Ort::Env env_;
-  Ort::SessionOptions encoder_sess_opts_;
-  Ort::SessionOptions decoder_sess_opts_;
-  Ort::SessionOptions joiner_sess_opts_;
-
-  Ort::AllocatorWithDefaultOptions allocator_;
-
-  std::unique_ptr<Ort::Session> encoder_sess_;
-  std::unique_ptr<Ort::Session> decoder_sess_;
-  std::unique_ptr<Ort::Session> joiner_sess_;
-
-  std::vector<std::string> encoder_input_names_;
-  std::vector<const char *> encoder_input_names_ptr_;
-
-  std::vector<std::string> encoder_output_names_;
-  std::vector<const char *> encoder_output_names_ptr_;
-
-  std::vector<std::string> decoder_input_names_;
-  std::vector<const char *> decoder_input_names_ptr_;
-
-  std::vector<std::string> decoder_output_names_;
-  std::vector<const char *> decoder_output_names_ptr_;
-
-  std::vector<std::string> joiner_input_names_;
-  std::vector<const char *> joiner_input_names_ptr_;
-
-  std::vector<std::string> joiner_output_names_;
-  std::vector<const char *> joiner_output_names_ptr_;
-
-  OnlineModelConfig config_;
-
-  std::vector<int32_t> encoder_dims_;
-  std::vector<int32_t> query_head_dims_;
-  std::vector<int32_t> value_head_dims_;
-  std::vector<int32_t> num_heads_;
-  std::vector<int32_t> num_encoder_layers_;
-  std::vector<int32_t> cnn_module_kernels_;
-  std::vector<int32_t> left_context_len_;
-
-  int32_t T_ = 0;
-  int32_t decode_chunk_len_ = 0;
-
-  int32_t context_size_ = 0;
-  int32_t vocab_size_ = 0;
-  int32_t feature_dim_ = 80;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONLINE_ZIPFORMER2_TRANSDUCER_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/onnx-utils.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/onnx-utils.h
deleted file mode 100644
index c978fbb7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/onnx-utils.h
+++ /dev/null
@@ -1,142 +0,0 @@
-// sherpa-onnx/csrc/onnx-utils.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-// Copyright (c)  2023  Pingfeng Luo
-#ifndef SHERPA_ONNX_CSRC_ONNX_UTILS_H_
-#define SHERPA_ONNX_CSRC_ONNX_UTILS_H_
-
-#ifdef _MSC_VER
-// For ToWide() below
-#include <codecvt>
-#include <locale>
-#endif
-
-#include <cassert>
-#include <ostream>
-#include <string>
-#include <utility>
-#include <vector>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#if __OHOS__
-#include "rawfile/raw_file_manager.h"
-#endif
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-
-namespace sherpa_onnx {
-
-/**
- * Get the input names of a model.
- *
- * @param sess An onnxruntime session.
- * @param input_names. On return, it contains the input names of the model.
- * @param input_names_ptr. On return, input_names_ptr[i] contains
- *                         input_names[i].c_str()
- */
-void GetInputNames(Ort::Session *sess, std::vector<std::string> *input_names,
-                   std::vector<const char *> *input_names_ptr);
-
-/**
- * Get the output names of a model.
- *
- * @param sess An onnxruntime session.
- * @param output_names. On return, it contains the output names of the model.
- * @param output_names_ptr. On return, output_names_ptr[i] contains
- *                         output_names[i].c_str()
- */
-void GetOutputNames(Ort::Session *sess, std::vector<std::string> *output_names,
-                    std::vector<const char *> *output_names_ptr);
-
-/**
- * Get the output frame of Encoder
- *
- * @param allocator allocator of onnxruntime
- * @param encoder_out encoder out tensor
- * @param t frame_index
- *
- */
-Ort::Value GetEncoderOutFrame(OrtAllocator *allocator, Ort::Value *encoder_out,
-                              int32_t t);
-
-std::string LookupCustomModelMetaData(const Ort::ModelMetadata &meta_data,
-                                      const char *key, OrtAllocator *allocator);
-
-void PrintModelMetadata(std::ostream &os,
-                        const Ort::ModelMetadata &meta_data);  // NOLINT
-
-// Return a deep copy of v
-Ort::Value Clone(OrtAllocator *allocator, const Ort::Value *v);
-
-// Return a shallow copy
-Ort::Value View(Ort::Value *v);
-
-float ComputeSum(const Ort::Value *v, int32_t n = -1);
-float ComputeMean(const Ort::Value *v, int32_t n = -1);
-
-// Print a 1-D tensor to stderr
-template <typename T = float>
-void Print1D(const Ort::Value *v);
-
-// Print a 2-D tensor to stderr
-template <typename T = float>
-void Print2D(const Ort::Value *v);
-
-// Print a 3-D tensor to stderr
-void Print3D(const Ort::Value *v);
-
-// Print a 4-D tensor to stderr
-void Print4D(const Ort::Value *v);
-
-void PrintShape(const Ort::Value *v);
-
-template <typename T = float>
-void Fill(Ort::Value *tensor, T value) {
-  auto n = tensor->GetTypeInfo().GetTensorTypeAndShapeInfo().GetElementCount();
-  auto p = tensor->GetTensorMutableData<T>();
-  std::fill(p, p + n, value);
-}
-
-std::vector<char> ReadFile(const std::string &filename);
-
-#if __ANDROID_API__ >= 9
-std::vector<char> ReadFile(AAssetManager *mgr, const std::string &filename);
-#endif
-
-#if __OHOS__
-std::vector<char> ReadFile(NativeResourceManager *mgr,
-                           const std::string &filename);
-#endif
-
-// TODO(fangjun): Document it
-Ort::Value Repeat(OrtAllocator *allocator, Ort::Value *cur_encoder_out,
-                  const std::vector<int32_t> &hyps_num_split);
-
-struct CopyableOrtValue {
-  Ort::Value value{nullptr};
-
-  CopyableOrtValue() = default;
-
-  /*explicit*/ CopyableOrtValue(Ort::Value v)  // NOLINT
-      : value(std::move(v)) {}
-
-  CopyableOrtValue(const CopyableOrtValue &other);
-
-  CopyableOrtValue &operator=(const CopyableOrtValue &other);
-
-  CopyableOrtValue(CopyableOrtValue &&other) noexcept;
-
-  CopyableOrtValue &operator=(CopyableOrtValue &&other) noexcept;
-};
-
-std::vector<CopyableOrtValue> Convert(std::vector<Ort::Value> values);
-
-std::vector<Ort::Value> Convert(std::vector<CopyableOrtValue> values);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_ONNX_UTILS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/packed-sequence.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/packed-sequence.h
deleted file mode 100644
index 203f47c3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/packed-sequence.h
+++ /dev/null
@@ -1,52 +0,0 @@
-// sherpa-onnx/csrc/packed-sequence.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_PACKED_SEQUENCE_H_
-#define SHERPA_ONNX_CSRC_PACKED_SEQUENCE_H_
-
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-
-namespace sherpa_onnx {
-
-struct PackedSequence {
-  std::vector<int32_t> sorted_indexes;
-  std::vector<int32_t> batch_sizes;
-
-  // data is a 2-D tensor of shape (sum(batch_sizes), channels)
-  Ort::Value data{nullptr};
-
-  // Return a shallow copy of data[start:start+size, :]
-  Ort::Value Get(int32_t start, int32_t size) {
-    auto shape = data.GetTensorTypeAndShapeInfo().GetShape();
-
-    std::array<int64_t, 2> ans_shape{size, shape[1]};
-
-    float *p = data.GetTensorMutableData<float>();
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    // a shallow copy
-    return Ort::Value::CreateTensor(memory_info, p + start * shape[1],
-                                    size * shape[1], ans_shape.data(),
-                                    ans_shape.size());
-  }
-};
-
-/** Similar to torch.nn.utils.rnn.pad_sequence but it supports only
- * batch_first=true.
- *
- * @param allocator
- * @param value  A 3-D tensor of shape (B, T, C). Its dtype is float.
- * @param length A 1-D tensor of shape (B,). Its dtype is int64_t. Each
- *               element in it specifies the valid length of the corresponding
- *               entry in value before padding.
- */
-PackedSequence PackPaddedSequence(OrtAllocator *allocator,
-                                  const Ort::Value *value, Ort::Value *length);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_PACKED_SEQUENCE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/pad-sequence.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/pad-sequence.h
deleted file mode 100644
index d44370ca..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/pad-sequence.h
+++ /dev/null
@@ -1,31 +0,0 @@
-// sherpa-onnx/csrc/pad-sequence.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_PAD_SEQUENCE_H_
-#define SHERPA_ONNX_CSRC_PAD_SEQUENCE_H_
-
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-
-namespace sherpa_onnx {
-
-/** Similar to torch.nn.utils.rnn.pad_sequence but it supports only
- * batch_first=true.
- *
- * @param allocator
- * @param values A list of 2-D tensors. Each tensor's second dimension
- *               must be the same and the data type of each tensor should
- *               be float.
- * @param padding_value Value used for padding. For log-fbank, you usually use
- *                      -23.025850929940457f as the padding value.
- *
- * @return Return a 3-D tensor of shape (B, max_T, C).
- */
-Ort::Value PadSequence(OrtAllocator *allocator,
-                       const std::vector<const Ort::Value *> &values,
-                       float padding_value);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_PAD_SEQUENCE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/parse-options.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/parse-options.h
deleted file mode 100644
index 615ee7d3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/parse-options.h
+++ /dev/null
@@ -1,259 +0,0 @@
-// sherpa-onnx/csrc/parse-options.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-//
-// This file is copied and modified from kaldi/src/util/parse-options.h
-
-#ifndef SHERPA_ONNX_CSRC_PARSE_OPTIONS_H_
-#define SHERPA_ONNX_CSRC_PARSE_OPTIONS_H_
-
-#include <cstdint>
-#include <sstream>
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-namespace sherpa_onnx {
-
-class ParseOptions {
- public:
-  explicit ParseOptions(const char *usage)
-      : print_args_(true),
-        help_(false),
-        usage_(usage),
-        argc_(0),
-        argv_(nullptr),
-        prefix_(""),
-        other_parser_(nullptr) {
-#if !defined(_MSC_VER) && !defined(__CYGWIN__)
-    // This is just a convenient place to set the stderr to line
-    // buffering mode, since it's called at program start.
-    // This helps ensure different programs' output is not mixed up.
-    setlinebuf(stderr);
-#endif
-    RegisterStandard("config", &config_,
-                     "Configuration file to read (this "
-                     "option may be repeated)");
-    RegisterStandard("print-args", &print_args_,
-                     "Print the command line arguments (to stderr)");
-    RegisterStandard("help", &help_, "Print out usage message");
-  }
-
-  /**
-    This is a constructor for the special case where some options are
-    registered with a prefix to avoid conflicts.  The object thus created will
-    only be used temporarily to register an options class with the original
-    options parser (which is passed as the *other pointer) using the given
-    prefix.  It should not be used for any other purpose, and the prefix must
-    not be the empty string.  It seems to be the least bad way of implementing
-    options with prefixes at this point.
-    Example of usage is:
-     ParseOptions po;  // original ParseOptions object
-     ParseOptions po_mfcc("mfcc", &po); // object with prefix.
-     MfccOptions mfcc_opts;
-     mfcc_opts.Register(&po_mfcc);
-    The options will now get registered as, e.g., --mfcc.frame-shift=10.0
-    instead of just --frame-shift=10.0
-   */
-  ParseOptions(const std::string &prefix, ParseOptions *other);
-
-  ParseOptions(const ParseOptions &) = delete;
-  ParseOptions &operator=(const ParseOptions &) = delete;
-  ~ParseOptions() = default;
-
-  void Register(const std::string &name, bool *ptr, const std::string &doc);
-  void Register(const std::string &name, int32_t *ptr, const std::string &doc);
-  void Register(const std::string &name, int64_t *ptr, const std::string &doc);
-  void Register(const std::string &name, uint32_t *ptr, const std::string &doc);
-  void Register(const std::string &name, float *ptr, const std::string &doc);
-  void Register(const std::string &name, double *ptr, const std::string &doc);
-  void Register(const std::string &name, std::string *ptr,
-                const std::string &doc);
-
-  /// If called after registering an option and before calling
-  /// Read(), disables that option from being used.  Will crash
-  /// at runtime if that option had not been registered.
-  void DisableOption(const std::string &name);
-
-  /// This one is used for registering standard parameters of all the programs
-  template <typename T>
-  void RegisterStandard(const std::string &name, T *ptr,
-                        const std::string &doc);
-
-  /**
-    Parses the command line options and fills the ParseOptions-registered
-    variables. This must be called after all the variables were registered!!!
-
-    Initially the variables have implicit values,
-    then the config file values are set-up,
-    finally the command line values given.
-    Returns the first position in argv that was not used.
-    [typically not useful: use NumParams() and GetParam(). ]
-   */
-  int Read(int argc, const char *const *argv);
-
-  /// Prints the usage documentation [provided in the constructor].
-  void PrintUsage(bool print_command_line = false) const;
-
-  /// Prints the actual configuration of all the registered variables
-  void PrintConfig(std::ostream &os) const;
-
-  /// Reads the options values from a config file.  Must be called after
-  /// registering all options.  This is usually used internally after the
-  /// standard --config option is used, but it may also be called from a
-  /// program.
-  void ReadConfigFile(const std::string &filename);
-
-  /// Number of positional parameters (c.f. argc-1).
-  int NumArgs() const;
-
-  /// Returns one of the positional parameters; 1-based indexing for argc/argv
-  /// compatibility. Will crash if param is not >=1 and <=NumArgs().
-  ///
-  /// Note: Index is 1 based.
-  std::string GetArg(int param) const;
-
-  std::string GetOptArg(int param) const {
-    return (param <= NumArgs() ? GetArg(param) : "");
-  }
-
-  /// The following function will return a possibly quoted and escaped
-  /// version of "str", according to the current shell.  Currently
-  /// this is just hardwired to bash.  It's useful for debug output.
-  static std::string Escape(const std::string &str);
-
- private:
-  /// Template to register various variable types,
-  /// used for program-specific parameters
-  template <typename T>
-  void RegisterTmpl(const std::string &name, T *ptr, const std::string &doc);
-
-  // Following functions do just the datatype-specific part of the job
-  /// Register boolean variable
-  void RegisterSpecific(const std::string &name, const std::string &idx,
-                        bool *b, const std::string &doc, bool is_standard);
-  /// Register int32_t variable
-  void RegisterSpecific(const std::string &name, const std::string &idx,
-                        int32_t *i, const std::string &doc, bool is_standard);
-  /// Register int64_t variable
-  void RegisterSpecific(const std::string &name, const std::string &idx,
-                        int64_t *i, const std::string &doc, bool is_standard);
-  /// Register unsigned  int32_t variable
-  void RegisterSpecific(const std::string &name, const std::string &idx,
-                        uint32_t *u, const std::string &doc, bool is_standard);
-  /// Register float variable
-  void RegisterSpecific(const std::string &name, const std::string &idx,
-                        float *f, const std::string &doc, bool is_standard);
-  /// Register double variable [useful as we change BaseFloat type].
-  void RegisterSpecific(const std::string &name, const std::string &idx,
-                        double *f, const std::string &doc, bool is_standard);
-  /// Register string variable
-  void RegisterSpecific(const std::string &name, const std::string &idx,
-                        std::string *s, const std::string &doc,
-                        bool is_standard);
-
-  /// Does the actual job for both kinds of parameters
-  /// Does the common part of the job for all datatypes,
-  /// then calls RegisterSpecific
-  template <typename T>
-  void RegisterCommon(const std::string &name, T *ptr, const std::string &doc,
-                      bool is_standard);
-
-  /// Set option with name "key" to "value"; will crash if can't do it.
-  /// "has_equal_sign" is used to allow --x for a boolean option x,
-  /// and --y=, for a string option y.
-  bool SetOption(const std::string &key, const std::string &value,
-                 bool has_equal_sign);
-
-  bool ToBool(std::string str) const;
-  int32_t ToInt(const std::string &str) const;
-  int64_t ToInt64(const std::string &str) const;
-  uint32_t ToUint(const std::string &str) const;
-  float ToFloat(const std::string &str) const;
-  double ToDouble(const std::string &str) const;
-
-  // maps for option variables
-  std::unordered_map<std::string, bool *> bool_map_;
-  std::unordered_map<std::string, int32_t *> int_map_;
-  std::unordered_map<std::string, int64_t *> int64_map_;
-  std::unordered_map<std::string, uint32_t *> uint_map_;
-  std::unordered_map<std::string, float *> float_map_;
-  std::unordered_map<std::string, double *> double_map_;
-  std::unordered_map<std::string, std::string *> string_map_;
-
-  /**
-     Structure for options' documentation
-   */
-  struct DocInfo {
-    DocInfo() = default;
-    DocInfo(const std::string &name, const std::string &usemsg)
-        : name_(name), use_msg_(usemsg), is_standard_(false) {}
-    DocInfo(const std::string &name, const std::string &usemsg,
-            bool is_standard)
-        : name_(name), use_msg_(usemsg), is_standard_(is_standard) {}
-
-    std::string name_;
-    std::string use_msg_;
-    bool is_standard_;
-  };
-  using DocMapType = std::unordered_map<std::string, DocInfo>;
-  DocMapType doc_map_;  ///< map for the documentation
-
-  bool print_args_;     ///< variable for the implicit --print-args parameter
-  bool help_;           ///< variable for the implicit --help parameter
-  std::string config_;  ///< variable for the implicit --config parameter
-  std::vector<std::string> positional_args_;
-  const char *usage_;
-  int argc_;
-  const char *const *argv_;
-
-  /// These members are not normally used. They are only used when the object
-  /// is constructed with a prefix
-  std::string prefix_;
-  ParseOptions *other_parser_;
-
- protected:
-  /// SplitLongArg parses an argument of the form --a=b, --a=, or --a,
-  /// and sets "has_equal_sign" to true if an equals-sign was parsed..
-  /// this is needed in order to correctly allow --x for a boolean option
-  /// x, and --y= for a string option y, and to disallow --x= and --y.
-  void SplitLongArg(const std::string &in, std::string *key, std::string *value,
-                    bool *has_equal_sign) const;
-
-  void NormalizeArgName(std::string *str) const;
-
-  /// Removes the beginning and trailing whitespaces from a string
-  void Trim(std::string *str) const;
-};
-
-/// This template is provided for convenience in reading config classes from
-/// files; this is not the standard way to read configuration options, but may
-/// occasionally be needed.  This function assumes the config has a function
-/// "void Register(ParseOptions *opts)" which it can call to register the
-/// ParseOptions object.
-template <class C>
-void ReadConfigFromFile(const std::string &config_filename, C *c) {
-  std::ostringstream usage_str;
-  usage_str << "Parsing config from "
-            << "from '" << config_filename << "'";
-  ParseOptions po(usage_str.str().c_str());
-  c->Register(&po);
-  po.ReadConfigFile(config_filename);
-}
-
-/// This variant of the template ReadConfigFromFile is for if you need to read
-/// two config classes from the same file.
-template <class C1, class C2>
-void ReadConfigsFromFile(const std::string &conf, C1 *c1, C2 *c2) {
-  std::ostringstream usage_str;
-  usage_str << "Parsing config from "
-            << "from '" << conf << "'";
-  ParseOptions po(usage_str.str().c_str());
-  c1->Register(&po);
-  c2->Register(&po);
-  po.ReadConfigFile(conf);
-}
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_PARSE_OPTIONS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/piper-phonemize-lexicon.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/piper-phonemize-lexicon.h
deleted file mode 100644
index ccd790a9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/piper-phonemize-lexicon.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// sherpa-onnx/csrc/piper-phonemize-lexicon.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_PIPER_PHONEMIZE_LEXICON_H_
-#define SHERPA_ONNX_CSRC_PIPER_PHONEMIZE_LEXICON_H_
-
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-#include "sherpa-onnx/csrc/offline-tts-frontend.h"
-#include "sherpa-onnx/csrc/offline-tts-vits-model-metadata.h"
-
-namespace sherpa_onnx {
-
-class PiperPhonemizeLexicon : public OfflineTtsFrontend {
- public:
-  PiperPhonemizeLexicon(const std::string &tokens, const std::string &data_dir,
-                        const OfflineTtsVitsModelMetaData &meta_data);
-
-  template <typename Manager>
-  PiperPhonemizeLexicon(Manager *mgr, const std::string &tokens,
-                        const std::string &data_dir,
-                        const OfflineTtsVitsModelMetaData &meta_data);
-
-  std::vector<TokenIDs> ConvertTextToTokenIds(
-      const std::string &text, const std::string &voice = "") const override;
-
- private:
-  // map unicode codepoint to an integer ID
-  std::unordered_map<char32_t, int32_t> token2id_;
-  OfflineTtsVitsModelMetaData meta_data_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_PIPER_PHONEMIZE_LEXICON_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/provider-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/provider-config.h
deleted file mode 100644
index fdc875e0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/provider-config.h
+++ /dev/null
@@ -1,95 +0,0 @@
-// sherpa-onnx/csrc/provider-config.h
-//
-// Copyright (c)  2024  Uniphore (Author: Manickavela)
-
-#ifndef SHERPA_ONNX_CSRC_PROVIDER_CONFIG_H_
-#define SHERPA_ONNX_CSRC_PROVIDER_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-#include "sherpa-onnx/csrc/macros.h"
-#include "onnxruntime_cxx_api.h"  // NOLINT
-
-namespace sherpa_onnx {
-
-struct CudaConfig {
-  int32_t cudnn_conv_algo_search = OrtCudnnConvAlgoSearchHeuristic;
-
-  CudaConfig() = default;
-  explicit CudaConfig(int32_t cudnn_conv_algo_search)
-      : cudnn_conv_algo_search(cudnn_conv_algo_search) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-struct TensorrtConfig {
-  int64_t trt_max_workspace_size = 2147483647;
-  int32_t trt_max_partition_iterations = 10;
-  int32_t trt_min_subgraph_size = 5;
-  bool trt_fp16_enable = true;
-  bool trt_detailed_build_log = false;
-  bool trt_engine_cache_enable = true;
-  bool trt_timing_cache_enable = true;
-  std::string trt_engine_cache_path = ".";
-  std::string trt_timing_cache_path = ".";
-  bool trt_dump_subgraphs = false;
-
-  TensorrtConfig() = default;
-  TensorrtConfig(int64_t trt_max_workspace_size,
-                int32_t trt_max_partition_iterations,
-                int32_t trt_min_subgraph_size,
-                bool trt_fp16_enable,
-                bool trt_detailed_build_log,
-                bool trt_engine_cache_enable,
-                bool trt_timing_cache_enable,
-                const std::string &trt_engine_cache_path,
-                const std::string &trt_timing_cache_path,
-                bool trt_dump_subgraphs)
-      : trt_max_workspace_size(trt_max_workspace_size),
-      trt_max_partition_iterations(trt_max_partition_iterations),
-      trt_min_subgraph_size(trt_min_subgraph_size),
-      trt_fp16_enable(trt_fp16_enable),
-      trt_detailed_build_log(trt_detailed_build_log),
-      trt_engine_cache_enable(trt_engine_cache_enable),
-      trt_timing_cache_enable(trt_timing_cache_enable),
-      trt_engine_cache_path(trt_engine_cache_path),
-      trt_timing_cache_path(trt_timing_cache_path),
-      trt_dump_subgraphs(trt_dump_subgraphs) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-struct ProviderConfig {
-  TensorrtConfig trt_config;
-  CudaConfig cuda_config;
-  std::string provider = "cpu";
-  int32_t device = 0;
-  // device only used for cuda and trt
-
-  ProviderConfig() = default;
-  ProviderConfig(const std::string &provider,
-                int32_t device)
-      : provider(provider), device(device) {}
-  ProviderConfig(const TensorrtConfig &trt_config,
-                const CudaConfig &cuda_config,
-                const std::string &provider,
-                int32_t device)
-      : trt_config(trt_config), cuda_config(cuda_config),
-      provider(provider), device(device) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_PROVIDER_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/provider.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/provider.h
deleted file mode 100644
index 2b85b8a2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/provider.h
+++ /dev/null
@@ -1,36 +0,0 @@
-// sherpa-onnx/csrc/provider.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_PROVIDER_H_
-#define SHERPA_ONNX_CSRC_PROVIDER_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/provider-config.h"
-namespace sherpa_onnx {
-
-// Please refer to
-// https://github.com/microsoft/onnxruntime/blob/main/java/src/main/java/ai/onnxruntime/OrtProvider.java
-// for a list of available providers
-enum class Provider {
-  kCPU = 0,       // CPUExecutionProvider
-  kCUDA = 1,      // CUDAExecutionProvider
-  kCoreML = 2,    // CoreMLExecutionProvider
-  kXnnpack = 3,   // XnnpackExecutionProvider
-  kNNAPI = 4,     // NnapiExecutionProvider
-  kTRT = 5,       // TensorRTExecutionProvider
-  kDirectML = 6,  // DmlExecutionProvider
-};
-
-/**
- * Convert a string to an enum.
- *
- * @param s We will convert it to lowercase before comparing.
- * @return Return an instance of Provider.
- */
-Provider StringToProvider(std::string s);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_PROVIDER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/resample.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/resample.h
deleted file mode 100644
index d6e22a19..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/resample.h
+++ /dev/null
@@ -1,144 +0,0 @@
-/**
- * Copyright     2013  Pegah Ghahremani
- *               2014  IMSL, PKU-HKUST (author: Wei Shi)
- *               2014  Yanqing Sun, Junjie Wang
- *               2014  Johns Hopkins University (author: Daniel Povey)
- * Copyright     2023  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-// this file is copied and modified from
-// kaldi/src/feat/resample.h
-#ifndef SHERPA_ONNX_CSRC_RESAMPLE_H_
-#define SHERPA_ONNX_CSRC_RESAMPLE_H_
-
-#include <cstdint>
-#include <vector>
-
-namespace sherpa_onnx {
-
-/*
-   We require that the input and output sampling rate be specified as
-   integers, as this is an easy way to specify that their ratio be rational.
-*/
-
-class LinearResample {
- public:
-  /// Constructor.  We make the input and output sample rates integers, because
-  /// we are going to need to find a common divisor.  This should just remind
-  /// you that they need to be integers.  The filter cutoff needs to be less
-  /// than samp_rate_in_hz/2 and less than samp_rate_out_hz/2.  num_zeros
-  /// controls the sharpness of the filter, more == sharper but less efficient.
-  /// We suggest around 4 to 10 for normal use.
-  LinearResample(int32_t samp_rate_in_hz, int32_t samp_rate_out_hz,
-                 float filter_cutoff_hz, int32_t num_zeros);
-
-  /// Calling the function Reset() resets the state of the object prior to
-  /// processing a new signal; it is only necessary if you have called
-  /// Resample(x, x_size, false, y) for some signal, leading to a remainder of
-  /// the signal being called, but then abandon processing the signal before
-  /// calling Resample(x, x_size, true, y) for the last piece.  Call it
-  /// unnecessarily between signals will not do any harm.
-  void Reset();
-
-  /// This function does the resampling.  If you call it with flush == true and
-  /// you have never called it with flush == false, it just resamples the input
-  /// signal (it resizes the output to a suitable number of samples).
-  ///
-  /// You can also use this function to process a signal a piece at a time.
-  /// suppose you break it into piece1, piece2, ... pieceN.  You can call
-  /// \code{.cc}
-  /// Resample(piece1, piece1_size, false, &output1);
-  /// Resample(piece2, piece2_size, false, &output2);
-  /// Resample(piece3, piece3_size, true, &output3);
-  /// \endcode
-  /// If you call it with flush == false, it won't output the last few samples
-  /// but will remember them, so that if you later give it a second piece of
-  /// the input signal it can process it correctly.
-  /// If your most recent call to the object was with flush == false, it will
-  /// have internal state; you can remove this by calling Reset().
-  /// Empty input is acceptable.
-  void Resample(const float *input, int32_t input_dim, bool flush,
-                std::vector<float> *output);
-
-  //// Return the input and output sampling rates (for checks, for example)
-  int32_t GetInputSamplingRate() const { return samp_rate_in_; }
-  int32_t GetOutputSamplingRate() const { return samp_rate_out_; }
-
- private:
-  void SetIndexesAndWeights();
-
-  float FilterFunc(float) const;
-
-  /// This function outputs the number of output samples we will output
-  /// for a signal with "input_num_samp" input samples.  If flush == true,
-  /// we return the largest n such that
-  /// (n/samp_rate_out_) is in the interval [ 0, input_num_samp/samp_rate_in_ ),
-  /// and note that the interval is half-open.  If flush == false,
-  /// define window_width as num_zeros / (2.0 * filter_cutoff_);
-  /// we return the largest n such that (n/samp_rate_out_) is in the interval
-  /// [ 0, input_num_samp/samp_rate_in_ - window_width ).
-  int64_t GetNumOutputSamples(int64_t input_num_samp, bool flush) const;
-
-  /// Given an output-sample index, this function outputs to *first_samp_in the
-  /// first input-sample index that we have a weight on (may be negative),
-  /// and to *samp_out_wrapped the index into weights_ where we can get the
-  /// corresponding weights on the input.
-  inline void GetIndexes(int64_t samp_out, int64_t *first_samp_in,
-                         int32_t *samp_out_wrapped) const;
-
-  void SetRemainder(const float *input, int32_t input_dim);
-
- private:
-  // The following variables are provided by the user.
-  int32_t samp_rate_in_;
-  int32_t samp_rate_out_;
-  float filter_cutoff_;
-  int32_t num_zeros_;
-
-  int32_t input_samples_in_unit_;  ///< The number of input samples in the
-                                   ///< smallest repeating unit: num_samp_in_ =
-                                   ///< samp_rate_in_hz / Gcd(samp_rate_in_hz,
-                                   ///< samp_rate_out_hz)
-
-  int32_t output_samples_in_unit_;  ///< The number of output samples in the
-                                    ///< smallest repeating unit: num_samp_out_
-                                    ///< = samp_rate_out_hz /
-                                    ///< Gcd(samp_rate_in_hz, samp_rate_out_hz)
-
-  /// The first input-sample index that we sum over, for this output-sample
-  /// index.  May be negative; any truncation at the beginning is handled
-  /// separately.  This is just for the first few output samples, but we can
-  /// extrapolate the correct input-sample index for arbitrary output samples.
-  std::vector<int32_t> first_index_;
-
-  /// Weights on the input samples, for this output-sample index.
-  std::vector<std::vector<float>> weights_;
-
-  // the following variables keep track of where we are in a particular signal,
-  // if it is being provided over multiple calls to Resample().
-
-  int64_t input_sample_offset_ = 0;   ///< The number of input samples we have
-                                      ///< already received for this signal
-                                      ///< (including anything in remainder_)
-  int64_t output_sample_offset_ = 0;  ///< The number of samples we have already
-                                      ///< output for this signal.
-  std::vector<float> input_remainder_;  ///< A small trailing part of the
-                                        ///< previously seen input signal.
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_RESAMPLE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/session.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/session.h
deleted file mode 100644
index e19db6c2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/session.h
+++ /dev/null
@@ -1,36 +0,0 @@
-// sherpa-onnx/csrc/session.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_SESSION_H_
-#define SHERPA_ONNX_CSRC_SESSION_H_
-
-#include <string>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/offline-lm-config.h"
-#include "sherpa-onnx/csrc/online-lm-config.h"
-#include "sherpa-onnx/csrc/online-model-config.h"
-
-namespace sherpa_onnx {
-
-Ort::SessionOptions GetSessionOptionsImpl(
-    int32_t num_threads, const std::string &provider_str,
-    const ProviderConfig *provider_config = nullptr);
-
-Ort::SessionOptions GetSessionOptions(const OfflineLMConfig &config);
-Ort::SessionOptions GetSessionOptions(const OnlineLMConfig &config);
-
-Ort::SessionOptions GetSessionOptions(const OnlineModelConfig &config);
-
-Ort::SessionOptions GetSessionOptions(const OnlineModelConfig &config,
-                                      const std::string &model_type);
-
-template <typename T>
-Ort::SessionOptions GetSessionOptions(const T &config) {
-  return GetSessionOptionsImpl(config.num_threads, config.provider);
-}
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_SESSION_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/silero-vad-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/silero-vad-model-config.h
deleted file mode 100644
index 5ae06f5f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/silero-vad-model-config.h
+++ /dev/null
@@ -1,46 +0,0 @@
-// sherpa-onnx/csrc/silero-vad-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_SILERO_VAD_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_SILERO_VAD_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct SileroVadModelConfig {
-  std::string model;
-
-  // threshold to classify a segment as speech
-  //
-  // If the predicted probability of a segment is larger than this
-  // value, then it is classified as speech.
-  float threshold = 0.5;
-
-  float min_silence_duration = 0.5;  // in seconds
-
-  float min_speech_duration = 0.25;  // in seconds
-
-  // 512, 1024, 1536 samples for 16000 Hz
-  // 256, 512, 768 samples for 800 Hz
-  int32_t window_size = 512;  // in samples
-
-  // If a speech segment is longer than this value, then we increase
-  // the threshold to 0.9. After finishing detecting the segment,
-  // the threshold value is reset to its original value.
-  float max_speech_duration = 20;  // in seconds
-
-  SileroVadModelConfig() = default;
-
-  void Register(ParseOptions *po);
-
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_SILERO_VAD_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/silero-vad-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/silero-vad-model.h
deleted file mode 100644
index 83913999..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/silero-vad-model.h
+++ /dev/null
@@ -1,55 +0,0 @@
-// sherpa-onnx/csrc/silero-vad-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_SILERO_VAD_MODEL_H_
-#define SHERPA_ONNX_CSRC_SILERO_VAD_MODEL_H_
-
-#include <memory>
-
-#include "sherpa-onnx/csrc/vad-model.h"
-
-namespace sherpa_onnx {
-
-class SileroVadModel : public VadModel {
- public:
-  explicit SileroVadModel(const VadModelConfig &config);
-
-  template <typename Manager>
-  SileroVadModel(Manager *mgr, const VadModelConfig &config);
-
-  ~SileroVadModel() override;
-
-  // reset the internal model states
-  void Reset() override;
-
-  /**
-   * @param samples Pointer to a 1-d array containing audio samples.
-   *                Each sample should be normalized to the range [-1, 1].
-   * @param n Number of samples.
-   *
-   * @return Return true if speech is detected. Return false otherwise.
-   */
-  bool IsSpeech(const float *samples, int32_t n) override;
-
-  // For silero vad V4, it is WindowShift().
-  // For silero vad V5, it is WindowShift()+64 for 16kHz and
-  //                          WindowShift()+32 for 8kHz
-  int32_t WindowSize() const override;
-
-  // 512
-  int32_t WindowShift() const override;
-
-  int32_t MinSilenceDurationSamples() const override;
-  int32_t MinSpeechDurationSamples() const override;
-
-  void SetMinSilenceDuration(float s) override;
-  void SetThreshold(float threshold) override;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_SILERO_VAD_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/slice.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/slice.h
deleted file mode 100644
index 42ebf6ce..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/slice.h
+++ /dev/null
@@ -1,48 +0,0 @@
-// sherpa-onnx/csrc/slice.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_SLICE_H_
-#define SHERPA_ONNX_CSRC_SLICE_H_
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-
-namespace sherpa_onnx {
-
-/** Get a deep copy by slicing a 3-D tensor v.
- *
- * It returns v[dim0_start:dim0_end, dim1_start:dim1_end, :]
- *
- * @param allocator
- * @param v A 3-D tensor. Its data type is T.
- * @param dim0_start  Start index of the first dimension..
- * @param dim0_end    End index of the first dimension..
- * @param dim1_start Start index of the second dimension.
- * @param dim1_end  End index of the second dimension.
- *
- * @return Return a 3-D tensor of shape
- *         (dim0_end-dim0_start, dim1_end-dim1_start, v.shape[2])
- */
-template <typename T = float>
-Ort::Value Slice(OrtAllocator *allocator, const Ort::Value *v,
-                 int32_t dim0_start, int32_t dim0_end, int32_t dim1_start,
-                 int32_t dim1_end);
-
-/** Get a deep copy by slicing a 2-D tensor v.
- *
- * It returns v[dim0_start:dim0_end, :]
- *
- * @param allocator
- * @param v A 2-D tensor. Its data type is T.
- * @param dim0_start  Start index of the first dimension..
- * @param dim0_end    End index of the first dimension..
- *
- * @return Return a 2-D tensor of shape
- *         (dim0_end-dim0_start, v.shape[1])
- */
-template <typename T = float>
-Ort::Value Slice(OrtAllocator *allocator, const Ort::Value *v,
-                 int32_t dim0_start, int32_t dim0_end);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_SLICE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-general-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-general-impl.h
deleted file mode 100644
index 8a884e85..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-general-impl.h
+++ /dev/null
@@ -1,117 +0,0 @@
-// sherpa-onnx/csrc/speaker-embedding-extractor-general-impl.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_GENERAL_IMPL_H_
-#define SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_GENERAL_IMPL_H_
-#include <algorithm>
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include "Eigen/Dense"
-#include "sherpa-onnx/csrc/speaker-embedding-extractor-impl.h"
-#include "sherpa-onnx/csrc/speaker-embedding-extractor-model.h"
-
-namespace sherpa_onnx {
-
-class SpeakerEmbeddingExtractorGeneralImpl
-    : public SpeakerEmbeddingExtractorImpl {
- public:
-  explicit SpeakerEmbeddingExtractorGeneralImpl(
-      const SpeakerEmbeddingExtractorConfig &config)
-      : model_(config) {}
-
-  template <typename Manager>
-  SpeakerEmbeddingExtractorGeneralImpl(
-      Manager *mgr, const SpeakerEmbeddingExtractorConfig &config)
-      : model_(mgr, config) {}
-
-  int32_t Dim() const override { return model_.GetMetaData().output_dim; }
-
-  std::unique_ptr<OnlineStream> CreateStream() const override {
-    FeatureExtractorConfig feat_config;
-    const auto &meta_data = model_.GetMetaData();
-    feat_config.sampling_rate = meta_data.sample_rate;
-    feat_config.normalize_samples = meta_data.normalize_samples;
-
-    return std::make_unique<OnlineStream>(feat_config);
-  }
-
-  bool IsReady(OnlineStream *s) const override {
-    return s->GetNumProcessedFrames() < s->NumFramesReady();
-  }
-
-  std::vector<float> Compute(OnlineStream *s) const override {
-    int32_t num_frames = s->NumFramesReady() - s->GetNumProcessedFrames();
-    if (num_frames <= 0) {
-#if __OHOS__
-      SHERPA_ONNX_LOGE(
-          "Please make sure IsReady(s) returns true. num_frames: %{public}d",
-          num_frames);
-#else
-      SHERPA_ONNX_LOGE(
-          "Please make sure IsReady(s) returns true. num_frames: %d",
-          num_frames);
-#endif
-      return {};
-    }
-
-    std::vector<float> features =
-        s->GetFrames(s->GetNumProcessedFrames(), num_frames);
-
-    s->GetNumProcessedFrames() += num_frames;
-
-    int32_t feat_dim = features.size() / num_frames;
-
-    const auto &meta_data = model_.GetMetaData();
-    if (!meta_data.feature_normalize_type.empty()) {
-      if (meta_data.feature_normalize_type == "global-mean") {
-        SubtractGlobalMean(features.data(), num_frames, feat_dim);
-      } else {
-#if __OHOS__
-        SHERPA_ONNX_LOGE("Unsupported feature_normalize_type: %{public}s",
-                         meta_data.feature_normalize_type.c_str());
-#else
-        SHERPA_ONNX_LOGE("Unsupported feature_normalize_type: %s",
-                         meta_data.feature_normalize_type.c_str());
-#endif
-        exit(-1);
-      }
-    }
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    std::array<int64_t, 3> x_shape{1, num_frames, feat_dim};
-    Ort::Value x =
-        Ort::Value::CreateTensor(memory_info, features.data(), features.size(),
-                                 x_shape.data(), x_shape.size());
-    Ort::Value embedding = model_.Compute(std::move(x));
-    std::vector<int64_t> embedding_shape =
-        embedding.GetTensorTypeAndShapeInfo().GetShape();
-
-    std::vector<float> ans(embedding_shape[1]);
-    std::copy(embedding.GetTensorData<float>(),
-              embedding.GetTensorData<float>() + ans.size(), ans.begin());
-
-    return ans;
-  }
-
- private:
-  void SubtractGlobalMean(float *p, int32_t num_frames,
-                          int32_t feat_dim) const {
-    auto m = Eigen::Map<
-        Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>>(
-        p, num_frames, feat_dim);
-
-    m = m.rowwise() - m.colwise().mean();
-  }
-
- private:
-  SpeakerEmbeddingExtractorModel model_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_GENERAL_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-impl.h
deleted file mode 100644
index 6299dce4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-impl.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// sherpa-onnx/csrc/speaker-embedding-extractor-impl.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_IMPL_H_
-#define SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_IMPL_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "sherpa-onnx/csrc/speaker-embedding-extractor.h"
-
-namespace sherpa_onnx {
-
-class SpeakerEmbeddingExtractorImpl {
- public:
-  virtual ~SpeakerEmbeddingExtractorImpl() = default;
-
-  static std::unique_ptr<SpeakerEmbeddingExtractorImpl> Create(
-      const SpeakerEmbeddingExtractorConfig &config);
-
-  template <typename Manager>
-  static std::unique_ptr<SpeakerEmbeddingExtractorImpl> Create(
-      Manager *mgr, const SpeakerEmbeddingExtractorConfig &config);
-
-  virtual int32_t Dim() const = 0;
-
-  virtual std::unique_ptr<OnlineStream> CreateStream() const = 0;
-
-  virtual bool IsReady(OnlineStream *s) const = 0;
-
-  virtual std::vector<float> Compute(OnlineStream *s) const = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-model-meta-data.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-model-meta-data.h
deleted file mode 100644
index 530938b5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-model-meta-data.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// sherpa-onnx/csrc/speaker-embedding-extractor-model-meta-data.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_MODEL_META_DATA_H_
-#define SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_MODEL_META_DATA_H_
-
-#include <cstdint>
-#include <string>
-
-namespace sherpa_onnx {
-
-struct SpeakerEmbeddingExtractorModelMetaData {
-  int32_t output_dim = 0;
-  int32_t sample_rate = 0;
-
-  // for wespeaker models, it is 0;
-  // for 3d-speaker models, it is 1
-  int32_t normalize_samples = 1;
-
-  // Chinese, English, etc.
-  std::string language;
-
-  // for 3d-speaker, it is global-mean
-  std::string feature_normalize_type;
-};
-
-}  // namespace sherpa_onnx
-#endif  // SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_MODEL_META_DATA_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-model.h
deleted file mode 100644
index 6c6bdd01..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-model.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// sherpa-onnx/csrc/speaker-embedding-extractor-model.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_MODEL_H_
-#define SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_MODEL_H_
-
-#include <memory>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/speaker-embedding-extractor-model-meta-data.h"
-#include "sherpa-onnx/csrc/speaker-embedding-extractor.h"
-
-namespace sherpa_onnx {
-
-class SpeakerEmbeddingExtractorModel {
- public:
-  explicit SpeakerEmbeddingExtractorModel(
-      const SpeakerEmbeddingExtractorConfig &config);
-
-  template <typename Manager>
-  SpeakerEmbeddingExtractorModel(Manager *mgr,
-                                 const SpeakerEmbeddingExtractorConfig &config);
-
-  ~SpeakerEmbeddingExtractorModel();
-
-  const SpeakerEmbeddingExtractorModelMetaData &GetMetaData() const;
-
-  /**
-   * @param x A float32 tensor of shape (N, T, C)
-   * @return A float32 tensor of shape (N, C)
-   */
-  Ort::Value Compute(Ort::Value x) const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-nemo-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-nemo-impl.h
deleted file mode 100644
index ec1c44d6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-nemo-impl.h
+++ /dev/null
@@ -1,145 +0,0 @@
-// sherpa-onnx/csrc/speaker-embedding-extractor-nemo-impl.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_NEMO_IMPL_H_
-#define SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_NEMO_IMPL_H_
-#include <algorithm>
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include "Eigen/Dense"
-#include "sherpa-onnx/csrc/speaker-embedding-extractor-impl.h"
-#include "sherpa-onnx/csrc/speaker-embedding-extractor-nemo-model.h"
-#include "sherpa-onnx/csrc/transpose.h"
-
-namespace sherpa_onnx {
-
-class SpeakerEmbeddingExtractorNeMoImpl : public SpeakerEmbeddingExtractorImpl {
- public:
-  explicit SpeakerEmbeddingExtractorNeMoImpl(
-      const SpeakerEmbeddingExtractorConfig &config)
-      : model_(config) {}
-
-  template <typename Manager>
-  SpeakerEmbeddingExtractorNeMoImpl(
-      Manager *mgr, const SpeakerEmbeddingExtractorConfig &config)
-      : model_(mgr, config) {}
-
-  int32_t Dim() const override { return model_.GetMetaData().output_dim; }
-
-  std::unique_ptr<OnlineStream> CreateStream() const override {
-    FeatureExtractorConfig feat_config;
-    const auto &meta_data = model_.GetMetaData();
-    feat_config.sampling_rate = meta_data.sample_rate;
-    feat_config.feature_dim = meta_data.feat_dim;
-    feat_config.normalize_samples = true;
-    feat_config.snip_edges = true;
-    feat_config.frame_shift_ms = meta_data.window_stride_ms;
-    feat_config.frame_length_ms = meta_data.window_size_ms;
-    feat_config.low_freq = 0;
-    feat_config.is_librosa = true;
-    feat_config.remove_dc_offset = false;
-    feat_config.window_type = meta_data.window_type;
-
-    return std::make_unique<OnlineStream>(feat_config);
-  }
-
-  bool IsReady(OnlineStream *s) const override {
-    return s->GetNumProcessedFrames() < s->NumFramesReady();
-  }
-
-  std::vector<float> Compute(OnlineStream *s) const override {
-    int32_t num_frames = s->NumFramesReady() - s->GetNumProcessedFrames();
-    if (num_frames <= 0) {
-#if __OHOS__
-      SHERPA_ONNX_LOGE(
-          "Please make sure IsReady(s) returns true. num_frames: %{public}d",
-          num_frames);
-#else
-      SHERPA_ONNX_LOGE(
-          "Please make sure IsReady(s) returns true. num_frames: %d",
-          num_frames);
-#endif
-      return {};
-    }
-
-    std::vector<float> features =
-        s->GetFrames(s->GetNumProcessedFrames(), num_frames);
-
-    s->GetNumProcessedFrames() += num_frames;
-
-    int32_t feat_dim = features.size() / num_frames;
-
-    const auto &meta_data = model_.GetMetaData();
-    if (!meta_data.feature_normalize_type.empty()) {
-      if (meta_data.feature_normalize_type == "per_feature") {
-        NormalizePerFeature(features.data(), num_frames, feat_dim);
-      } else {
-#if __OHOS__
-        SHERPA_ONNX_LOGE("Unsupported feature_normalize_type: %{public}s",
-                         meta_data.feature_normalize_type.c_str());
-#else
-
-        SHERPA_ONNX_LOGE("Unsupported feature_normalize_type: %s",
-                         meta_data.feature_normalize_type.c_str());
-#endif
-        exit(-1);
-      }
-    }
-
-    if (num_frames % 16 != 0) {
-      int32_t pad = 16 - num_frames % 16;
-      features.resize((num_frames + pad) * feat_dim);
-    }
-
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    std::array<int64_t, 3> x_shape{1, num_frames, feat_dim};
-    Ort::Value x =
-        Ort::Value::CreateTensor(memory_info, features.data(), features.size(),
-                                 x_shape.data(), x_shape.size());
-
-    x = Transpose12(model_.Allocator(), &x);
-
-    int64_t x_lens = num_frames;
-    std::array<int64_t, 1> x_lens_shape{1};
-    Ort::Value x_lens_tensor = Ort::Value::CreateTensor(
-        memory_info, &x_lens, 1, x_lens_shape.data(), x_lens_shape.size());
-
-    Ort::Value embedding =
-        model_.Compute(std::move(x), std::move(x_lens_tensor));
-    std::vector<int64_t> embedding_shape =
-        embedding.GetTensorTypeAndShapeInfo().GetShape();
-
-    std::vector<float> ans(embedding_shape[1]);
-    std::copy(embedding.GetTensorData<float>(),
-              embedding.GetTensorData<float>() + ans.size(), ans.begin());
-
-    return ans;
-  }
-
- private:
-  void NormalizePerFeature(float *p, int32_t num_frames,
-                           int32_t feat_dim) const {
-    auto m = Eigen::Map<
-        Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>>(
-        p, num_frames, feat_dim);
-
-    auto EX = m.colwise().mean();
-    auto EX2 = m.array().pow(2).colwise().sum() / num_frames;
-    auto variance = EX2 - EX.array().pow(2);
-    auto stddev = variance.array().sqrt();
-
-    m = (m.rowwise() - EX).array().rowwise() / (stddev.array() + 1e-5);
-  }
-
- private:
-  SpeakerEmbeddingExtractorNeMoModel model_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_NEMO_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-nemo-model-meta-data.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-nemo-model-meta-data.h
deleted file mode 100644
index f0ff1f7b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-nemo-model-meta-data.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// sherpa-onnx/csrc/speaker-embedding-extractor-nemo-model-meta-data.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_NEMO_MODEL_META_DATA_H_
-#define SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_NEMO_MODEL_META_DATA_H_
-
-#include <cstdint>
-#include <string>
-
-namespace sherpa_onnx {
-
-struct SpeakerEmbeddingExtractorNeMoModelMetaData {
-  int32_t output_dim = 0;
-  int32_t feat_dim = 80;
-  int32_t sample_rate = 0;
-  int32_t window_size_ms = 25;
-  int32_t window_stride_ms = 25;
-
-  // Chinese, English, etc.
-  std::string language;
-
-  // for 3d-speaker, it is global-mean
-  std::string feature_normalize_type;
-  std::string window_type = "hann";
-};
-
-}  // namespace sherpa_onnx
-#endif  // SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_NEMO_MODEL_META_DATA_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-nemo-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-nemo-model.h
deleted file mode 100644
index ed61ee8b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor-nemo-model.h
+++ /dev/null
@@ -1,44 +0,0 @@
-// sherpa-onnx/csrc/speaker-embedding-extractor-nemo-model.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_NEMO_MODEL_H_
-#define SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_NEMO_MODEL_H_
-
-#include <memory>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-#include "sherpa-onnx/csrc/speaker-embedding-extractor-nemo-model-meta-data.h"
-#include "sherpa-onnx/csrc/speaker-embedding-extractor.h"
-
-namespace sherpa_onnx {
-
-class SpeakerEmbeddingExtractorNeMoModel {
- public:
-  explicit SpeakerEmbeddingExtractorNeMoModel(
-      const SpeakerEmbeddingExtractorConfig &config);
-
-  template <typename Manager>
-  SpeakerEmbeddingExtractorNeMoModel(
-      Manager *mgr, const SpeakerEmbeddingExtractorConfig &config);
-
-  ~SpeakerEmbeddingExtractorNeMoModel();
-
-  const SpeakerEmbeddingExtractorNeMoModelMetaData &GetMetaData() const;
-
-  /**
-   * @param x A float32 tensor of shape (N, C, T)
-   * @param x_len A int64 tensor of shape (N,)
-   * @return A float32 tensor of shape (N, C)
-   */
-  Ort::Value Compute(Ort::Value x, Ort::Value x_len) const;
-
-  OrtAllocator *Allocator() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_NEMO_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor.h
deleted file mode 100644
index 068e6b8d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-extractor.h
+++ /dev/null
@@ -1,71 +0,0 @@
-// sherpa-onnx/csrc/speaker-embedding-extractor.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_H_
-#define SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "sherpa-onnx/csrc/online-stream.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct SpeakerEmbeddingExtractorConfig {
-  std::string model;
-  int32_t num_threads = 1;
-  bool debug = false;
-  std::string provider = "cpu";
-
-  SpeakerEmbeddingExtractorConfig() = default;
-  SpeakerEmbeddingExtractorConfig(const std::string &model, int32_t num_threads,
-                                  bool debug, const std::string &provider)
-      : model(model),
-        num_threads(num_threads),
-        debug(debug),
-        provider(provider) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-  std::string ToString() const;
-};
-
-class SpeakerEmbeddingExtractorImpl;
-
-class SpeakerEmbeddingExtractor {
- public:
-  explicit SpeakerEmbeddingExtractor(
-      const SpeakerEmbeddingExtractorConfig &config);
-
-  template <typename Manager>
-  SpeakerEmbeddingExtractor(Manager *mgr,
-                            const SpeakerEmbeddingExtractorConfig &config);
-
-  ~SpeakerEmbeddingExtractor();
-
-  // Return the dimension of the embedding
-  int32_t Dim() const;
-
-  // Create a stream to accept audio samples and compute features
-  std::unique_ptr<OnlineStream> CreateStream() const;
-
-  // Return true if there are feature frames in OnlineStream that
-  // can be used to compute embeddings.
-  bool IsReady(OnlineStream *s) const;
-
-  // Compute the speaker embedding from the available unprocessed features
-  // of the given stream
-  //
-  // You have to ensure IsReady(s) returns true before you call this method.
-  std::vector<float> Compute(OnlineStream *s) const;
-
- private:
-  std::unique_ptr<SpeakerEmbeddingExtractorImpl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_EXTRACTOR_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-manager.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-manager.h
deleted file mode 100644
index 9490765c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/speaker-embedding-manager.h
+++ /dev/null
@@ -1,120 +0,0 @@
-// sherpa-onnx/csrc/speaker-embedding-manager.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_MANAGER_H_
-#define SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_MANAGER_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-struct SpeakerMatch {
-  const std::string name;
-  float score;
-};
-
-namespace sherpa_onnx {
-
-class SpeakerEmbeddingManager {
- public:
-  // @param dim Embedding dimension.
-  explicit SpeakerEmbeddingManager(int32_t dim);
-  ~SpeakerEmbeddingManager();
-
-  /* Add the embedding and name of a speaker to the manager.
-   *
-   * @param name Name of the speaker
-   * @param p Pointer to the embedding. Its length is `dim`.
-   * @return Return true if added successfully. Return false if it failed.
-   *         At present, the only reason for a failure is that there is already
-   *         a speaker with the same `name`.
-   */
-  bool Add(const std::string &name, const float *p) const;
-
-  /** Add a list of embeddings of a speaker.
-   *
-   * @param name Name of the speaker
-   * @param embedding_list A list of embeddings. Each entry should be of size
-   *                       `dim`. The average of the list is the final
-   *                       embedding.
-   * @return Return true if added successfully. Return false if it failed.
-   *         At present, the only reason for a failure is that there is already
-   *         a speaker with the same `name`.
-   */
-  bool Add(const std::string &name,
-           const std::vector<std::vector<float>> &embedding_list) const;
-
-  /* Remove a speaker by its name.
-   *
-   * @param name Name of the speaker to remove.
-   * @return Return true if it is removed successfully. Return false
-   *         if there is no such a speaker.
-   */
-  bool Remove(const std::string &name) const;
-
-  /** It is for speaker identification.
-   *
-   * It computes the cosine similarity between and given embedding and all
-   * other embeddings and find the embedding that has the largest score
-   * and the score is above or equal to threshold. Return the speaker
-   * name for the embedding if found; otherwise, it returns an empty string.
-   *
-   * @param p The input embedding.
-   * @param threshold A value between 0 and 1.
-   * @param If found, return the name of the speaker. Otherwise, return an
-   *        empty string.
-   */
-  std::string Search(const float *p, float threshold) const;
-
-  /**
-   * It is for speaker identification.
-   *
-   * It computes the cosine similarity between a given embedding and all
-   * other embeddings and finds the embeddings that have the largest scores
-   * and the scores are above or equal to the threshold. Returns a vector of
-   * SpeakerMatch structures containing the speaker names and scores for the
-   * embeddings if found; otherwise, returns an empty vector.
-   *
-   * @param p A pointer to the input embedding.
-   * @param threshold A value between 0 and 1.
-   * @param n The number of top matches to return.
-   * @return A vector of SpeakerMatch structures. If matches are found, the
-   *         vector contains the names and scores of the speakers. Otherwise,
-   *         it returns an empty vector.
-   */
-  std::vector<SpeakerMatch> GetBestMatches(const float *p, float threshold,
-                                           int32_t n) const;
-
-  /* Check whether the input embedding matches the embedding of the input
-   * speaker.
-   *
-   * It is for speaker verification.
-   *
-   * @param name The target speaker name.
-   * @param p The input embedding to check.
-   * @param threshold A value between 0 and 1.
-   * @return Return true if it matches. Otherwise, it returns false.
-   */
-  bool Verify(const std::string &name, const float *p, float threshold) const;
-
-  float Score(const std::string &name, const float *p) const;
-
-  // Return true if the given speaker already exists; return false otherwise.
-  bool Contains(const std::string &name) const;
-
-  int32_t NumSpeakers() const;
-
-  int32_t Dim() const;
-
-  // Return a list of speaker names
-  std::vector<std::string> GetAllSpeakers() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_SPEAKER_EMBEDDING_MANAGER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/spoken-language-identification-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/spoken-language-identification-impl.h
deleted file mode 100644
index bbbfedef..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/spoken-language-identification-impl.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// sherpa-onnx/csrc/spoken-language-identification-impl.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_SPOKEN_LANGUAGE_IDENTIFICATION_IMPL_H_
-#define SHERPA_ONNX_CSRC_SPOKEN_LANGUAGE_IDENTIFICATION_IMPL_H_
-
-#include <memory>
-#include <string>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "sherpa-onnx/csrc/spoken-language-identification.h"
-
-namespace sherpa_onnx {
-
-class SpokenLanguageIdentificationImpl {
- public:
-  virtual ~SpokenLanguageIdentificationImpl() = default;
-
-  static std::unique_ptr<SpokenLanguageIdentificationImpl> Create(
-      const SpokenLanguageIdentificationConfig &config);
-
-#if __ANDROID_API__ >= 9
-  static std::unique_ptr<SpokenLanguageIdentificationImpl> Create(
-      AAssetManager *mgr, const SpokenLanguageIdentificationConfig &config);
-#endif
-
-  virtual std::unique_ptr<OfflineStream> CreateStream() const = 0;
-
-  virtual std::string Compute(OfflineStream *s) const = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_SPOKEN_LANGUAGE_IDENTIFICATION_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/spoken-language-identification-whisper-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/spoken-language-identification-whisper-impl.h
deleted file mode 100644
index d82883b5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/spoken-language-identification-whisper-impl.h
+++ /dev/null
@@ -1,133 +0,0 @@
-// sherpa-onnx/csrc/spoken-language-identification-whisper-impl.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_SPOKEN_LANGUAGE_IDENTIFICATION_WHISPER_IMPL_H_
-#define SHERPA_ONNX_CSRC_SPOKEN_LANGUAGE_IDENTIFICATION_WHISPER_IMPL_H_
-
-#include <algorithm>
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "sherpa-onnx/csrc/offline-whisper-model.h"
-#include "sherpa-onnx/csrc/spoken-language-identification-impl.h"
-#include "sherpa-onnx/csrc/transpose.h"
-
-namespace sherpa_onnx {
-
-class SpokenLanguageIdentificationWhisperImpl
-    : public SpokenLanguageIdentificationImpl {
- public:
-  explicit SpokenLanguageIdentificationWhisperImpl(
-      const SpokenLanguageIdentificationConfig &config)
-      : config_(config), model_(std::make_unique<OfflineWhisperModel>(config)) {
-    Check();
-  }
-
-#if __ANDROID_API__ >= 9
-  SpokenLanguageIdentificationWhisperImpl(
-      AAssetManager *mgr, const SpokenLanguageIdentificationConfig &config)
-      : config_(config),
-        model_(std::make_unique<OfflineWhisperModel>(mgr, config)) {
-    Check();
-  }
-#endif
-
-  std::unique_ptr<OfflineStream> CreateStream() const override {
-    return std::make_unique<OfflineStream>(WhisperTag{});
-  }
-
-  std::string Compute(OfflineStream *s) const override {
-    int32_t max_num_frames = 3000;
-    auto memory_info =
-        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-    int32_t feat_dim = s->FeatureDim();
-    std::vector<float> f = s->GetFrames();
-    int32_t num_frames = f.size() / feat_dim;
-
-    // we use 50 here so that there will be some zero tail paddings
-    if (num_frames >= max_num_frames - 50) {
-      SHERPA_ONNX_LOGE(
-          "Only waves less than 30 seconds are supported. We process only the "
-          "first 30 seconds and discard the remaining data");
-      num_frames = max_num_frames - 50;
-    }
-
-    model_->NormalizeFeatures(f.data(), num_frames, feat_dim);
-
-    // note that 1000 is an experience-value.
-    // You can replace 1000 by other values, say, 100.
-    //
-    // Since we have removed the 30 seconds constraint, we need
-    // tail_padding_frames so that whisper is able to detect the eot token.
-    int32_t tail_padding_frames = 1000;
-
-    if (config_.whisper.tail_paddings > 0) {
-      tail_padding_frames = config_.whisper.tail_paddings;
-    }
-
-    int32_t actual_frames =
-        std::min(num_frames + tail_padding_frames, max_num_frames);
-
-    std::array<int64_t, 3> shape{1, actual_frames, feat_dim};
-
-    Ort::Value mel = Ort::Value::CreateTensor<float>(
-        model_->Allocator(), shape.data(), shape.size());
-
-    float *p_mel = mel.GetTensorMutableData<float>();
-    std::copy(f.data(), f.data() + num_frames * feat_dim, p_mel);
-
-    std::fill_n(p_mel + num_frames * feat_dim,
-                (actual_frames - num_frames) * feat_dim, 0);
-
-    mel = Transpose12(model_->Allocator(), &mel);
-
-    try {
-      auto cross_kv = model_->ForwardEncoder(std::move(mel));
-      int32_t lang_id = model_->DetectLanguage(cross_kv.first, cross_kv.second);
-      const auto &id2lang = model_->GetID2Lang();
-      if (id2lang.count(lang_id)) {
-        return id2lang.at(lang_id);
-      } else {
-        SHERPA_ONNX_LOGE("Unknown language ID: %d. Return an empty string.",
-                         lang_id);
-        return "";
-      }
-    } catch (const Ort::Exception &ex) {
-      SHERPA_ONNX_LOGE(
-          "\n\nCaught exception:\n\n%s\n\nReturn an empty result. Number of "
-          "input frames: %d, Current tail "
-          "paddings: %d. If you see a lot of such exceptions, please consider "
-          "using a larger --whisper-tail-paddings",
-          ex.what(), num_frames, tail_padding_frames);
-      return "";
-    }
-  }
-
- private:
-  void Check() const {
-    if (!model_->IsMultiLingual()) {
-      SHERPA_ONNX_LOGE(
-          "Only whisper multilingual models can be used for spoken language "
-          "identification. Given: %s,%s",
-          config_.whisper.encoder.c_str(), config_.whisper.decoder.c_str());
-      exit(-1);
-    }
-  }
-
- private:
-  SpokenLanguageIdentificationConfig config_;
-  std::unique_ptr<OfflineWhisperModel> model_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_SPOKEN_LANGUAGE_IDENTIFICATION_WHISPER_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/spoken-language-identification.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/spoken-language-identification.h
deleted file mode 100644
index 6adb25d2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/spoken-language-identification.h
+++ /dev/null
@@ -1,99 +0,0 @@
-// sherpa-onnx/csrc/spoken-language-identification.h
-//
-// Copyright (c)  2024  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_SPOKEN_LANGUAGE_IDENTIFICATION_H_
-#define SHERPA_ONNX_CSRC_SPOKEN_LANGUAGE_IDENTIFICATION_H_
-
-#include <memory>
-#include <string>
-
-#if __ANDROID_API__ >= 9
-#include "android/asset_manager.h"
-#include "android/asset_manager_jni.h"
-#endif
-
-#include "sherpa-onnx/csrc/offline-stream.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-
-namespace sherpa_onnx {
-
-struct SpokenLanguageIdentificationWhisperConfig {
-  // Requires a multi-lingual whisper model.
-  // That is, it supports only tiny, base, small, medium, large.
-  // Note: It does NOT support tiny.en, base.en, small.en, medium.en
-  std::string encoder;
-  std::string decoder;
-
-  // Number of tail padding frames.
-  //
-  // Since we remove the 30-second constraint, we need to add some paddings
-  // at the end.
-  //
-  // Recommended values:
-  //   - 50 for English models
-  //   - 300 for multilingual models
-  int32_t tail_paddings = -1;
-
-  SpokenLanguageIdentificationWhisperConfig() = default;
-
-  SpokenLanguageIdentificationWhisperConfig(const std::string &encoder,
-                                            const std::string &decoder,
-                                            int32_t tail_paddings)
-      : encoder(encoder), decoder(decoder), tail_paddings(tail_paddings) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-  std::string ToString() const;
-};
-
-struct SpokenLanguageIdentificationConfig {
-  SpokenLanguageIdentificationWhisperConfig whisper;
-
-  int32_t num_threads = 1;
-  bool debug = false;
-  std::string provider = "cpu";
-
-  SpokenLanguageIdentificationConfig() = default;
-
-  SpokenLanguageIdentificationConfig(
-      const SpokenLanguageIdentificationWhisperConfig &whisper,
-      int32_t num_threads, bool debug, const std::string &provider)
-      : whisper(whisper),
-        num_threads(num_threads),
-        debug(debug),
-        provider(provider) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-  std::string ToString() const;
-};
-
-class SpokenLanguageIdentificationImpl;
-
-class SpokenLanguageIdentification {
- public:
-  explicit SpokenLanguageIdentification(
-      const SpokenLanguageIdentificationConfig &config);
-
-#if __ANDROID_API__ >= 9
-  SpokenLanguageIdentification(
-      AAssetManager *mgr, const SpokenLanguageIdentificationConfig &config);
-#endif
-
-  ~SpokenLanguageIdentification();
-
-  // Create a stream to accept audio samples and compute features
-  std::unique_ptr<OfflineStream> CreateStream() const;
-
-  // Return a string containing the language, e.g., en, zh, de,
-  // etc.
-  // Note: en is for English, zh is for Chinese, de is for German, etc.
-  std::string Compute(OfflineStream *s) const;
-
- private:
-  std::unique_ptr<SpokenLanguageIdentificationImpl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_SPOKEN_LANGUAGE_IDENTIFICATION_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/stack.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/stack.h
deleted file mode 100644
index 55f199aa..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/stack.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// sherpa-onnx/csrc/stack.h
-//
-// Copyright (c) 2023 Jingzhao Ou (jingzhao.ou@gmail.com)
-
-#ifndef SHERPA_ONNX_CSRC_STACK_H_
-#define SHERPA_ONNX_CSRC_STACK_H_
-
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-
-namespace sherpa_onnx {
-
-/** Stack a list of tensors along the given dim.
- *
- * @param allocator Allocator to allocate space for the returned tensor
- * @param values  Pointer to a list of tensors. The shape of the tensor must
- *                be the same except on the dim to be stacked.
- * @param dim  The dim along which to concatenate the input tensors
- *
- * @return Return the stacked tensor
- */
-template <typename T = float>
-Ort::Value Stack(OrtAllocator *allocator,
-                 const std::vector<const Ort::Value *> &values, int32_t dim);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_STACK_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/symbol-table.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/symbol-table.h
deleted file mode 100644
index 20d8d206..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/symbol-table.h
+++ /dev/null
@@ -1,71 +0,0 @@
-// sherpa-onnx/csrc/symbol-table.h
-//
-// Copyright (c)  2022-2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_SYMBOL_TABLE_H_
-#define SHERPA_ONNX_CSRC_SYMBOL_TABLE_H_
-
-#include <istream>
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-namespace sherpa_onnx {
-
-// The same token can be mapped to different integer IDs, so
-// we need an id2token argument here.
-std::unordered_map<std::string, int32_t> ReadTokens(
-    std::istream &is,
-    std::unordered_map<int32_t, std::string> *id2token = nullptr);
-
-std::vector<int32_t> ConvertTokensToIds(
-    const std::unordered_map<std::string, int32_t> &token2id,
-    const std::vector<std::string> &tokens);
-
-/// It manages mapping between symbols and integer IDs.
-class SymbolTable {
- public:
-  SymbolTable() = default;
-  /// Construct a symbol table from a file or from a buffered string.
-  /// Each line in the file contains two fields:
-  ///
-  ///    sym ID
-  ///
-  /// Fields are separated by space(s).
-  explicit SymbolTable(const std::string &filename, bool is_file = true);
-
-  template <typename Manager>
-  SymbolTable(Manager *mgr, const std::string &filename);
-
-  /// Return a string representation of this symbol table
-  std::string ToString() const;
-
-  /// Return the symbol corresponding to the given ID.
-  const std::string operator[](int32_t id) const;
-  /// Return the ID corresponding to the given symbol.
-  int32_t operator[](const std::string &sym) const;
-
-  /// Return true if there is a symbol with the given ID.
-  bool Contains(int32_t id) const;
-
-  /// Return true if there is a given symbol in the symbol table.
-  bool Contains(const std::string &sym) const;
-
-  // for tokens.txt from Whisper
-  void ApplyBase64Decode();
-
-  int32_t NumSymbols() const { return id2sym_.size(); }
-
- private:
-  void Init(std::istream &is);
-
- private:
-  std::unordered_map<std::string, int32_t> sym2id_;
-  std::unordered_map<int32_t, std::string> id2sym_;
-};
-
-std::ostream &operator<<(std::ostream &os, const SymbolTable &symbol_table);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_SYMBOL_TABLE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/tee-stream.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/tee-stream.h
deleted file mode 100644
index 6fbd3d29..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/tee-stream.h
+++ /dev/null
@@ -1,61 +0,0 @@
-// Code in this file is copied and modified from
-// https://wordaligned.org/articles/cpp-streambufs
-
-#ifndef SHERPA_ONNX_CSRC_TEE_STREAM_H_
-#define SHERPA_ONNX_CSRC_TEE_STREAM_H_
-#include <ostream>
-#include <streambuf>
-#include <string>
-
-namespace sherpa_onnx {
-
-template <typename char_type, typename traits = std::char_traits<char_type>>
-class basic_teebuf : public std::basic_streambuf<char_type, traits> {
- public:
-  using int_type = typename traits::int_type;
-
-  basic_teebuf(std::basic_streambuf<char_type, traits> *sb1,
-               std::basic_streambuf<char_type, traits> *sb2)
-      : sb1(sb1), sb2(sb2) {}
-
- private:
-  int sync() override {
-    int const r1 = sb1->pubsync();
-    int const r2 = sb2->pubsync();
-    return r1 == 0 && r2 == 0 ? 0 : -1;
-  }
-
-  int_type overflow(int_type c) override {
-    int_type const eof = traits::eof();
-
-    if (traits::eq_int_type(c, eof)) {
-      return traits::not_eof(c);
-    } else {
-      char_type const ch = traits::to_char_type(c);
-      int_type const r1 = sb1->sputc(ch);
-      int_type const r2 = sb2->sputc(ch);
-
-      return traits::eq_int_type(r1, eof) || traits::eq_int_type(r2, eof) ? eof
-                                                                          : c;
-    }
-  }
-
- private:
-  std::basic_streambuf<char_type, traits> *sb1;
-  std::basic_streambuf<char_type, traits> *sb2;
-};
-
-using teebuf = basic_teebuf<char>;
-
-class TeeStream : public std::ostream {
- public:
-  TeeStream(std::ostream &o1, std::ostream &o2)
-      : std::ostream(&tbuf), tbuf(o1.rdbuf(), o2.rdbuf()) {}
-
- private:
-  teebuf tbuf;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_TEE_STREAM_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/text-utils.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/text-utils.h
deleted file mode 100644
index a0b968d8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/text-utils.h
+++ /dev/null
@@ -1,129 +0,0 @@
-// sherpa-onnx/csrc/text-utils.h
-//
-// Copyright 2009-2011  Saarland University;  Microsoft Corporation
-// Copyright      2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_TEXT_UTILS_H_
-#define SHERPA_ONNX_CSRC_TEXT_UTILS_H_
-#include <errno.h>
-#include <stdlib.h>
-
-#include <limits>
-#include <string>
-#include <type_traits>
-#include <vector>
-
-#ifdef _MSC_VER
-#define SHERPA_ONNX_STRTOLL(cur_cstr, end_cstr) \
-  _strtoi64(cur_cstr, end_cstr, 10);
-#else
-#define SHERPA_ONNX_STRTOLL(cur_cstr, end_cstr) strtoll(cur_cstr, end_cstr, 10);
-#endif
-
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/util/text-utils.h
-
-namespace sherpa_onnx {
-
-/// Converts a string into an integer via strtoll and returns false if there was
-/// any kind of problem (i.e. the string was not an integer or contained extra
-/// non-whitespace junk, or the integer was too large to fit into the type it is
-/// being converted into).  Only sets *out if everything was OK and it returns
-/// true.
-template <class Int>
-bool ConvertStringToInteger(const std::string &str, Int *out) {
-  // copied from kaldi/src/util/text-util.h
-  static_assert(std::is_integral<Int>::value, "");
-  const char *this_str = str.c_str();
-  char *end = nullptr;
-  errno = 0;
-  int64_t i = SHERPA_ONNX_STRTOLL(this_str, &end);
-  if (end != this_str) {
-    while (isspace(*end)) ++end;
-  }
-  if (end == this_str || *end != '\0' || errno != 0) return false;
-  Int iInt = static_cast<Int>(i);
-  if (static_cast<int64_t>(iInt) != i ||
-      (i < 0 && !std::numeric_limits<Int>::is_signed)) {
-    return false;
-  }
-  *out = iInt;
-  return true;
-}
-
-/// Split a string using any of the single character delimiters.
-/// If omit_empty_strings == true, the output will contain any
-/// nonempty strings after splitting on any of the
-/// characters in the delimiter.  If omit_empty_strings == false,
-/// the output will contain n+1 strings if there are n characters
-/// in the set "delim" within the input string.  In this case
-/// the empty string is split to a single empty string.
-void SplitStringToVector(const std::string &full, const char *delim,
-                         bool omit_empty_strings,
-                         std::vector<std::string> *out);
-
-/**
-  \brief Split a string (e.g. 1:2:3) into a vector of integers.
-
-  \param [in]  delim  String containing a list of characters, any of which
-                      is allowed as a delimiter.
-  \param [in] omit_empty_strings If true, empty strings between delimiters are
-                      allowed and will not produce an output integer; if false,
-                      instances of characters in 'delim' that are consecutive or
-                      at the start or end of the string would be an error.
-                      You'll normally want this to be true if 'delim' consists
-                      of spaces, and false otherwise.
-  \param [out] out   The output list of integers.
-*/
-template <class I>
-bool SplitStringToIntegers(const std::string &full, const char *delim,
-                           bool omit_empty_strings,  // typically false [but
-                                                     // should probably be true
-                                                     // if "delim" is spaces].
-                           std::vector<I> *out) {
-  static_assert(std::is_integral<I>::value, "");
-  if (*(full.c_str()) == '\0') {
-    out->clear();
-    return true;
-  }
-  std::vector<std::string> split;
-  SplitStringToVector(full, delim, omit_empty_strings, &split);
-  out->resize(split.size());
-  for (size_t i = 0; i < split.size(); i++) {
-    const char *this_str = split[i].c_str();
-    char *end = NULL;
-    int64_t j = 0;
-    j = SHERPA_ONNX_STRTOLL(this_str, &end);
-    if (end == this_str || *end != '\0') {
-      out->clear();
-      return false;
-    } else {
-      I jI = static_cast<I>(j);
-      if (static_cast<int64_t>(jI) != j) {
-        // output type cannot fit this integer.
-        out->clear();
-        return false;
-      }
-      (*out)[i] = jI;
-    }
-  }
-  return true;
-}
-
-// This is defined for F = float and double.
-template <class F>
-bool SplitStringToFloats(const std::string &full, const char *delim,
-                         bool omit_empty_strings,  // typically false
-                         std::vector<F> *out);
-
-// This is defined for F = float and double.
-template <typename T>
-bool ConvertStringToReal(const std::string &str, T *out);
-
-std::vector<std::string> SplitUtf8(const std::string &text);
-
-std::string ToLowerCase(const std::string &s);
-void ToLowerCase(std::string *in_out);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_TEXT_UTILS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/transducer-keyword-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/transducer-keyword-decoder.h
deleted file mode 100644
index a352abcc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/transducer-keyword-decoder.h
+++ /dev/null
@@ -1,62 +0,0 @@
-// sherpa-onnx/csrc/transducer-keywords-decoder.h
-//
-// Copyright (c)  2023-2024  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_TRANSDUCER_KEYWORD_DECODER_H_
-#define SHERPA_ONNX_CSRC_TRANSDUCER_KEYWORD_DECODER_H_
-
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "sherpa-onnx/csrc/online-stream.h"
-#include "sherpa-onnx/csrc/online-transducer-model.h"
-
-namespace sherpa_onnx {
-
-struct TransducerKeywordResult {
-  /// Number of frames after subsampling we have decoded so far
-  int32_t frame_offset = 0;
-
-  /// The decoded token IDs for keywords
-  std::vector<int64_t> tokens;
-
-  /// The triggered keyword
-  std::string keyword;
-
-  /// number of trailing blank frames decoded so far
-  int32_t num_trailing_blanks = 0;
-
-  /// timestamps[i] contains the output frame index where tokens[i] is decoded.
-  std::vector<int32_t> timestamps;
-
-  // used only in modified beam_search
-  Hypotheses hyps;
-};
-
-class TransducerKeywordDecoder {
- public:
-  TransducerKeywordDecoder(OnlineTransducerModel *model,
-                           int32_t max_active_paths,
-                           int32_t num_trailing_blanks, int32_t unk_id)
-      : model_(model),
-        max_active_paths_(max_active_paths),
-        num_trailing_blanks_(num_trailing_blanks),
-        unk_id_(unk_id) {}
-
-  TransducerKeywordResult GetEmptyResult() const;
-
-  void Decode(Ort::Value encoder_out, OnlineStream **ss,
-              std::vector<TransducerKeywordResult> *result);
-
- private:
-  OnlineTransducerModel *model_;  // Not owned
-
-  int32_t max_active_paths_;
-  int32_t num_trailing_blanks_;
-  int32_t unk_id_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_TRANSDUCER_KEYWORD_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/transpose.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/transpose.h
deleted file mode 100644
index aba7f44c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/transpose.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// sherpa-onnx/csrc/transpose.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_TRANSPOSE_H_
-#define SHERPA_ONNX_CSRC_TRANSPOSE_H_
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-
-namespace sherpa_onnx {
-/** Transpose a 3-D tensor from shape (B, T, C) to (T, B, C).
- *
- * @param allocator
- * @param v A 3-D tensor of shape (B, T, C). Its dataype is type.
- *
- * @return Return a 3-D tensor of shape (T, B, C). Its datatype is type.
- */
-template <typename type = float>
-Ort::Value Transpose01(OrtAllocator *allocator, const Ort::Value *v);
-
-/** Transpose a 3-D tensor from shape (B, T, C) to (B, C, T).
- *
- * @param allocator
- * @param v A 3-D tensor of shape (B, T, C). Its dataype is type.
- *
- * @return Return a 3-D tensor of shape (B, C, T). Its datatype is type.
- */
-template <typename type = float>
-Ort::Value Transpose12(OrtAllocator *allocator, const Ort::Value *v);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_TRANSPOSE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/unbind.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/unbind.h
deleted file mode 100644
index 9c799120..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/unbind.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// sherpa-onnx/csrc/unbind.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_UNBIND_H_
-#define SHERPA_ONNX_CSRC_UNBIND_H_
-
-#include <vector>
-
-#include "onnxruntime_cxx_api.h"  // NOLINT
-
-namespace sherpa_onnx {
-
-/** It is similar to torch.unbind() but we keep the unbind dim to 1 in
- * the output
- *
- * @param allocator Allocator to allocate space for the returned tensor
- * @param value  The tensor to unbind
- * @param dim  The dim along which to unbind the tensor
- *
- * @return Return a list of tensors
- */
-template <typename T = float>
-std::vector<Ort::Value> Unbind(OrtAllocator *allocator, const Ort::Value *value,
-                               int32_t dim);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_UNBIND_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/utils.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/utils.h
deleted file mode 100644
index a9d59e8a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/utils.h
+++ /dev/null
@@ -1,62 +0,0 @@
-// sherpa-onnx/csrc/utils.h
-//
-// Copyright      2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_UTILS_H_
-#define SHERPA_ONNX_CSRC_UTILS_H_
-
-#include <string>
-#include <vector>
-
-#include "sherpa-onnx/csrc/symbol-table.h"
-#include "ssentencepiece/csrc/ssentencepiece.h"
-
-namespace sherpa_onnx {
-
-/* Encode the hotwords in an input stream to be tokens ids.
- *
- * @param is The input stream, it contains several lines, one hotword for each
- *           line. For each hotword, the tokens (cjkchar or bpe) are separated
- *           by spaces.
- * @param symbol_table  The tokens table mapping symbols to ids. All the symbols
- *                      in the stream should be in the symbol_table, if not this
- *                      function returns fasle.
- *
- * @@param hotwords  The encoded ids to be written to.
- *
- * @return  If all the symbols from ``is`` are in the symbol_table, returns true
- *          otherwise returns false.
- */
-bool EncodeHotwords(std::istream &is, const std::string &modeling_unit,
-                    const SymbolTable &symbol_table,
-                    const ssentencepiece::Ssentencepiece *bpe_encoder_,
-                    std::vector<std::vector<int32_t>> *hotwords_id,
-                    std::vector<float> *boost_scores);
-
-/* Encode the keywords in an input stream to be tokens ids.
- *
- * @param is The input stream, it contains several lines, one hotword for each
- *           line. For each hotword, the tokens (cjkchar or bpe) are separated
- *           by spaces, it might contain boosting score (starting with :),
- *           triggering threshold (starting with #) and keyword string (starting
- *           with @) too.
- * @param symbol_table  The tokens table mapping symbols to ids. All the symbols
- *                      in the stream should be in the symbol_table, if not this
- *                      function returns fasle.
- *
- * @param keywords_id The encoded ids to be written to.
- * @param keywords The original keyword string to be written to.
- * @param boost_scores  The boosting score for each keyword to be written to.
- * @param threshold  The triggering threshold for each keyword to be written to.
- *
- * @return  If all the symbols from ``is`` are in the symbol_table, returns true
- *          otherwise returns false.
- */
-bool EncodeKeywords(std::istream &is, const SymbolTable &symbol_table,
-                    std::vector<std::vector<int32_t>> *keywords_id,
-                    std::vector<std::string> *keywords,
-                    std::vector<float> *boost_scores,
-                    std::vector<float> *threshold);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_UTILS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/vad-model-config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/vad-model-config.h
deleted file mode 100644
index 725ab340..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/vad-model-config.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// sherpa-onnx/csrc/vad-model-config.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_VAD_MODEL_CONFIG_H_
-#define SHERPA_ONNX_CSRC_VAD_MODEL_CONFIG_H_
-
-#include <string>
-
-#include "sherpa-onnx/csrc/parse-options.h"
-#include "sherpa-onnx/csrc/silero-vad-model-config.h"
-
-namespace sherpa_onnx {
-
-struct VadModelConfig {
-  SileroVadModelConfig silero_vad;
-
-  int32_t sample_rate = 16000;
-  int32_t num_threads = 1;
-  std::string provider = "cpu";
-
-  // true to show debug information when loading models
-  bool debug = false;
-
-  VadModelConfig() = default;
-
-  VadModelConfig(const SileroVadModelConfig &silero_vad, int32_t sample_rate,
-                 int32_t num_threads, const std::string &provider, bool debug)
-      : silero_vad(silero_vad),
-        sample_rate(sample_rate),
-        num_threads(num_threads),
-        provider(provider),
-        debug(debug) {}
-
-  void Register(ParseOptions *po);
-  bool Validate() const;
-
-  std::string ToString() const;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_VAD_MODEL_CONFIG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/vad-model.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/vad-model.h
deleted file mode 100644
index aac68bb1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/vad-model.h
+++ /dev/null
@@ -1,47 +0,0 @@
-// sherpa-onnx/csrc/vad-model.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_VAD_MODEL_H_
-#define SHERPA_ONNX_CSRC_VAD_MODEL_H_
-
-#include <memory>
-
-#include "sherpa-onnx/csrc/vad-model-config.h"
-
-namespace sherpa_onnx {
-
-class VadModel {
- public:
-  virtual ~VadModel() = default;
-
-  static std::unique_ptr<VadModel> Create(const VadModelConfig &config);
-
-  template <typename Manager>
-  static std::unique_ptr<VadModel> Create(Manager *mgr,
-                                          const VadModelConfig &config);
-
-  // reset the internal model states
-  virtual void Reset() = 0;
-
-  /**
-   * @param samples Pointer to a 1-d array containing audio samples.
-   *                Each sample should be normalized to the range [-1, 1].
-   * @param n Number of samples. Should be equal to WindowSize()
-   *
-   * @return Return true if speech is detected. Return false otherwise.
-   */
-  virtual bool IsSpeech(const float *samples, int32_t n) = 0;
-
-  virtual int32_t WindowSize() const = 0;
-
-  virtual int32_t WindowShift() const = 0;
-
-  virtual int32_t MinSilenceDurationSamples() const = 0;
-  virtual int32_t MinSpeechDurationSamples() const = 0;
-  virtual void SetMinSilenceDuration(float s) = 0;
-  virtual void SetThreshold(float threshold) = 0;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_VAD_MODEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/voice-activity-detector.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/voice-activity-detector.h
deleted file mode 100644
index 015cedf0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/voice-activity-detector.h
+++ /dev/null
@@ -1,53 +0,0 @@
-// sherpa-onnx/csrc/voice-activity-detector.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef SHERPA_ONNX_CSRC_VOICE_ACTIVITY_DETECTOR_H_
-#define SHERPA_ONNX_CSRC_VOICE_ACTIVITY_DETECTOR_H_
-
-#include <memory>
-#include <vector>
-
-#include "sherpa-onnx/csrc/vad-model-config.h"
-
-namespace sherpa_onnx {
-
-struct SpeechSegment {
-  int32_t start;  // in samples
-  std::vector<float> samples;
-};
-
-class VoiceActivityDetector {
- public:
-  explicit VoiceActivityDetector(const VadModelConfig &config,
-                                 float buffer_size_in_seconds = 60);
-
-  template <typename Manager>
-  VoiceActivityDetector(Manager *mgr, const VadModelConfig &config,
-                        float buffer_size_in_seconds = 60);
-
-  ~VoiceActivityDetector();
-
-  void AcceptWaveform(const float *samples, int32_t n);
-  bool Empty() const;
-  void Pop();
-  void Clear();
-  const SpeechSegment &Front() const;
-
-  bool IsSpeechDetected() const;
-
-  void Reset() const;
-
-  // At the end of the utterance, you can invoke this method so that
-  // the last speech segment can be detected.
-  void Flush() const;
-
-  const VadModelConfig &GetConfig() const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_VOICE_ACTIVITY_DETECTOR_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/wave-reader.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/wave-reader.h
deleted file mode 100644
index 98e956ab..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/wave-reader.h
+++ /dev/null
@@ -1,31 +0,0 @@
-// sherpa-onnx/csrc/wave-reader.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_WAVE_READER_H_
-#define SHERPA_ONNX_CSRC_WAVE_READER_H_
-
-#include <istream>
-#include <string>
-#include <vector>
-
-namespace sherpa_onnx {
-
-/** Read a wave file with expected sample rate.
-
-    @param filename Path to a wave file. It MUST be single channel, 16-bit
-                    PCM encoded.
-    @param sampling_rate  On return, it contains the sampling rate of the file.
-    @param is_ok On return it is true if the reading succeeded; false otherwise.
-
-    @return Return wave samples normalized to the range [-1, 1).
- */
-std::vector<float> ReadWave(const std::string &filename, int32_t *sampling_rate,
-                            bool *is_ok);
-
-std::vector<float> ReadWave(std::istream &is, int32_t *sampling_rate,
-                            bool *is_ok);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_WAVE_READER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/wave-writer.h b/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/wave-writer.h
deleted file mode 100644
index bae4c504..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/csrc/wave-writer.h
+++ /dev/null
@@ -1,27 +0,0 @@
-// sherpa-onnx/csrc/wave-writer.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef SHERPA_ONNX_CSRC_WAVE_WRITER_H_
-#define SHERPA_ONNX_CSRC_WAVE_WRITER_H_
-
-#include <cstdint>
-#include <string>
-
-namespace sherpa_onnx {
-
-// Write a single channel wave file.
-// Note that the input samples are in the range [-1, 1]. It will be multiplied
-// by 32767 and saved in int16_t format in the wave file.
-//
-// @param filename Path to save the samples.
-// @param sampling_rate Sample rate of the samples.
-// @param samples Pointer to the samples
-// @param n Number of samples
-// @return Return true if the write succeeds; return false otherwise.
-bool WriteWave(const std::string &filename, int32_t sampling_rate,
-               const float *samples, int32_t n);
-
-}  // namespace sherpa_onnx
-
-#endif  // SHERPA_ONNX_CSRC_WAVE_WRITER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/decodable-ctc.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/decodable-ctc.h
deleted file mode 100644
index 71a482cc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/decodable-ctc.h
+++ /dev/null
@@ -1,47 +0,0 @@
-// kaldi-decoder/csrc/decodable-ctc.h
-//
-// Copyright (c)  2023  Xiaomi Corporation
-
-#ifndef KALDI_DECODER_CSRC_DECODABLE_CTC_H_
-#define KALDI_DECODER_CSRC_DECODABLE_CTC_H_
-
-#include "kaldi-decoder/csrc/decodable-itf.h"
-#include "kaldi-decoder/csrc/eigen.h"
-
-namespace kaldi_decoder {
-
-class DecodableCtc : public DecodableInterface {
- public:
-  // It copies the input log_probs
-  explicit DecodableCtc(const FloatMatrix &log_probs, int32_t offset = 0);
-
-  // It shares the memory with the input array.
-  //
-  // @param p Pointer to a 2-d array of shape (num_rows, num_cols).
-  //          The array should be kept alive as long as this object is still
-  //          alive.
-  DecodableCtc(const float *p, int32_t num_rows, int32_t num_cols,
-               int32_t offset = 0);
-
-  float LogLikelihood(int32_t frame, int32_t index) override;
-
-  int32_t NumFramesReady() const override;
-
-  // Indices are one-based!  This is for compatibility with OpenFst.
-  int32_t NumIndices() const override;
-
-  bool IsLastFrame(int32_t frame) const override;
-
- private:
-  // it saves log_softmax output
-  FloatMatrix log_probs_;
-
-  const float *p_ = nullptr;  // pointer to a 2-d array
-  int32_t num_rows_;          // number of rows in the 2-d array
-  int32_t num_cols_;          // number of cols in the 2-d array
-  int32_t offset_ = 0;
-};
-
-}  // namespace kaldi_decoder
-
-#endif  // KALDI_DECODER_CSRC_DECODABLE_CTC_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/decodable-itf.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/decodable-itf.h
deleted file mode 100644
index 9f58ec9e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/decodable-itf.h
+++ /dev/null
@@ -1,106 +0,0 @@
-// kaldi-decoder/csrc/decodable-itf.h
-
-// Copyright 2009-2011  Microsoft Corporation;  Saarland University;
-//                      Mirko Hannemann;  Go Vivace Inc.;
-//                2013  Johns Hopkins University (author: Daniel Povey)
-// Copyright (c)  2023  Xiaomi Corporation
-// this file is copied and modified from
-// kaldi/src/itf/decodable-itf.h
-#ifndef KALDI_DECODER_CSRC_DECODABLE_ITF_H_
-#define KALDI_DECODER_CSRC_DECODABLE_ITF_H_
-#include "kaldi-decoder/csrc/log.h"
-
-namespace kaldi_decoder {
-
-/**
-    DecodableInterface provides a link between the (acoustic-modeling and
-    feature-processing) code and the decoder.  The idea is to make this
-    interface as small as possible, and to make it as agnostic as possible about
-    the form of the acoustic model (e.g. don't assume the probabilities are a
-    function of just a vector of floats), and about the decoder (e.g. don't
-    assume it accesses frames in strict left-to-right order).  For normal
-    models, without on-line operation, the "decodable" sub-class will just be a
-    wrapper around a matrix of features and an acoustic model, and it will
-    answer the question 'what is the acoustic likelihood for this index and this
-    frame?'.
-
-    For online decoding, where the features are coming in in real time, it is
-    important to understand the IsLastFrame() and NumFramesReady() functions.
-    There are two ways these are used: the old online-decoding code, in
-   ../online/, and the new online-decoding code, in ../online2/.  In the old
-   online-decoding code, the decoder would do: \code{.cc} for (int frame = 0;
-   !decodable.IsLastFrame(frame); frame++) {
-      // Process this frame
-    }
-    \endcode
-   and the call to IsLastFrame would block if the features had not arrived yet.
-   The decodable object would have to know when to terminate the decoding.  This
-   online-decoding mode is still supported, it is what happens when you call,
-   for example, LatticeFasterDecoder::Decode().
-
-   We realized that this "blocking" mode of decoding is not very convenient
-   because it forces the program to be multi-threaded and makes it complex to
-   control endpointing.  In the "new" decoding code, you don't call (for
-   example) LatticeFasterDecoder::Decode(), you call
-   LatticeFasterDecoder::InitDecoding(), and then each time you get more
-   features, you provide them to the decodable object, and you call
-   LatticeFasterDecoder::AdvanceDecoding(), which does something like this:
-   \code{.cc}
-   while (num_frames_decoded_ < decodable.NumFramesReady()) {
-     // Decode one more frame [increments num_frames_decoded_]
-   }
-   \endcode
-   So the decodable object never has IsLastFrame() called.  For decoding where
-   you are starting with a matrix of features, the NumFramesReady() function
-   will always just return the number of frames in the file, and IsLastFrame()
-   will return true for the last frame.
-
-   For truly online decoding, the "old" online decodable objects in ../online/
-   have a "blocking" IsLastFrame() and will crash if you call NumFramesReady().
-   The "new" online decodable objects in ../online2/ return the number of frames
-   currently accessible if you call NumFramesReady().  You will likely not need
-   to call IsLastFrame(), but we implement it to only return true for the last
-   frame of the file once we've decided to terminate decoding.
-*/
-class DecodableInterface {
- public:
-  virtual ~DecodableInterface() = default;
-
-  /// Returns the log likelihood, which will be negated in the decoder.
-  /// The "frame" starts from zero.  You should verify that NumFramesReady() >
-  /// frame before calling this.
-  virtual float LogLikelihood(int32_t frame, int32_t index) = 0;
-
-  /// Returns true if this is the last frame.  Frames are zero-based, so the
-  /// first frame is zero.  IsLastFrame(-1) will return false, unless the file
-  /// is empty (which is a case that I'm not sure all the code will handle, so
-  /// be careful).  Caution: the behavior of this function in an online setting
-  /// is being changed somewhat.  In future it may return false in cases where
-  /// we haven't yet decided to terminate decoding, but later true if we decide
-  /// to terminate decoding.  The plan in future is to rely more on
-  /// NumFramesReady(), and in future, IsLastFrame() would always return false
-  /// in an online-decoding setting, and would only return true in a
-  /// decoding-from-matrix setting where we want to allow the last delta or LDA
-  /// features to be flushed out for compatibility with the baseline setup.
-  virtual bool IsLastFrame(int32_t frame) const = 0;
-
-  /// The call NumFramesReady() will return the number of frames currently
-  /// available for this decodable object.  This is for use in setups where you
-  /// don't want the decoder to block while waiting for input.  This is newly
-  /// added as of Jan 2014, and I hope, going forward, to rely on this mechanism
-  /// more than IsLastFrame to know when to stop decoding.
-  virtual int32_t NumFramesReady() const {
-    KALDI_DECODER_ERR
-        << "NumFramesReady() not implemented for this decodable type.";
-    return -1;
-  }
-
-  /// Returns the number of states in the acoustic model
-  /// (they will be indexed one-based, i.e. from 1 to NumIndices();
-  /// this is for compatibility with OpenFst).
-  virtual int32_t NumIndices() const = 0;
-};
-
-}  // namespace kaldi_decoder
-
-#endif  // KALDI_DECODER_CSRC_DECODABLE_ITF_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/eigen.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/eigen.h
deleted file mode 100644
index 444ac9db..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/eigen.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// kaldi-decoder/csrc/eigen.h
-
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef KALDI_DECODER_CSRC_EIGEN_H_
-#define KALDI_DECODER_CSRC_EIGEN_H_
-
-#include "Eigen/Dense"
-
-namespace kaldi_decoder {
-
-using FloatMatrix =
-    Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>;
-
-using DoubleMatrix =
-    Eigen::Matrix<double, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>;
-
-using FloatVector = Eigen::Matrix<float, Eigen::Dynamic, 1>;
-
-using DoubleVector = Eigen::Matrix<double, Eigen::Dynamic, 1>;
-
-using FloatRowVector = Eigen::Matrix<float, 1, Eigen::Dynamic>;
-
-using DoubleRowVector = Eigen::Matrix<double, 1, Eigen::Dynamic>;
-
-float LogSumExp(const FloatVector &v);
-
-FloatVector Softmax(const FloatVector &v, float *log_sum_exp = nullptr);
-
-// A vector of normal distribution
-FloatVector RandnVector(int32_t n, float mean = 0, float stddev = 1);
-
-FloatMatrix RandnMatrix(int32_t rows, int32_t cols, float mean = 0,
-                        float stddev = 1);
-
-float Randn(float mean = 0, float stddev = 1);
-
-}  // namespace kaldi_decoder
-
-#endif  // KALDI_DECODER_CSRC_EIGEN_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/faster-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/faster-decoder.h
deleted file mode 100644
index ba41deef..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/faster-decoder.h
+++ /dev/null
@@ -1,204 +0,0 @@
-// kaldi-decoder/csrc/faster-decoder.h
-
-// Copyright 2009-2011  Microsoft Corporation
-//                2013  Johns Hopkins University (author: Daniel Povey)
-// Copyright (c)  2023  Xiaomi Corporation
-
-// this file is copied and modified from
-// kaldi/src/decoder/faster-decoder.h
-#ifndef KALDI_DECODER_CSRC_FASTER_DECODER_H_
-#define KALDI_DECODER_CSRC_FASTER_DECODER_H_
-
-#include <limits>
-#include <string>
-#include <vector>
-
-#include "fst/fst.h"
-#include "fst/fstlib.h"
-#include "kaldi-decoder/csrc/decodable-itf.h"
-#include "kaldi-decoder/csrc/hash-list.h"
-#include "kaldifst/csrc/lattice-weight.h"
-
-namespace kaldi_decoder {
-
-struct FasterDecoderOptions {
-  // Decoding beam.  Larger->slower, more accurate.
-  float beam;
-
-  // Decoder max active states.  Larger->slower; more accurate
-  int32_t max_active;
-
-  // Decoder min active states (don't prune if #active less than this).
-  int32_t min_active;
-
-  // Increment used in decoder [obscure setting]
-  float beam_delta;
-
-  // Setting used in decoder to control hash behavior
-  float hash_ratio;
-
-  /*implicit*/ FasterDecoderOptions(
-      float beam = 16.0,
-      int32_t max_active = std::numeric_limits<int32_t>::max(),
-      int32_t min_active = 20, float beam_delta = 0.5, float hash_ratio = 2.0)
-      : beam(beam),
-        max_active(max_active),
-        min_active(min_active),  // This decoder mostly used for
-                                 // alignment, use small default.
-        beam_delta(beam_delta),
-        hash_ratio(hash_ratio) {}
-
-  std::string ToString() const {
-    std::ostringstream os;
-
-    os << "FasterDecoderOptions(";
-    os << "beam=" << beam << ", ";
-    os << "max_active=" << max_active << ", ";
-    os << "min_active=" << min_active << ", ";
-    os << "beam_delta=" << beam_delta << ", ";
-    os << "hash_ratio=" << hash_ratio << ")";
-
-    return os.str();
-  }
-};
-
-class FasterDecoder {
- public:
-  typedef fst::StdArc Arc;
-  typedef Arc::Label Label;
-  typedef Arc::StateId StateId;
-  typedef Arc::Weight Weight;
-
-  FasterDecoder(const fst::Fst<fst::StdArc> &fst,
-                const FasterDecoderOptions &config);
-
-  FasterDecoder(const FasterDecoder &) = delete;
-  FasterDecoder &operator=(const FasterDecoder &) = delete;
-
-  void SetOptions(const FasterDecoderOptions &config) { config_ = config; }
-
-  ~FasterDecoder() { ClearToks(toks_.Clear()); }
-
-  void Decode(DecodableInterface *decodable);
-
-  /// Returns true if a final state was active on the last frame.
-  bool ReachedFinal() const;
-
-  /// GetBestPath gets the decoding traceback. If "use_final_probs" is true
-  /// AND we reached a final state, it limits itself to final states;
-  /// otherwise it gets the most likely token not taking into account
-  /// final-probs. Returns true if the output best path was not the empty
-  /// FST (will only return false in unusual circumstances where
-  /// no tokens survived).
-  bool GetBestPath(fst::MutableFst<fst::LatticeArc> *fst_out,
-                   bool use_final_probs = true);
-
-  /// As a new alternative to Decode(), you can call InitDecoding
-  /// and then (possibly multiple times) AdvanceDecoding().
-  void InitDecoding();
-
-  /// This will decode until there are no more frames ready in the decodable
-  /// object, but if max_num_frames is >= 0 it will decode no more than
-  /// that many frames.
-  void AdvanceDecoding(DecodableInterface *decodable,
-                       int32_t max_num_frames = -1);
-
-  /// Returns the number of frames already decoded.
-  int32_t NumFramesDecoded() const { return num_frames_decoded_; }
-
- protected:
-  class Token {
-   public:
-    Arc arc_;  // contains only the graph part of the cost;
-    // we can work out the acoustic part from difference between
-    // "cost_" and prev->cost_.
-    Token *prev_;
-    int32_t ref_count_;
-    // if you are looking for weight_ here, it was removed and now we just have
-    // cost_, which corresponds to ConvertToCost(weight_).
-    double cost_;
-    inline Token(const Arc &arc, float ac_cost, Token *prev)
-        : arc_(arc), prev_(prev), ref_count_(1) {
-      if (prev) {
-        prev->ref_count_++;
-        cost_ = prev->cost_ + arc.weight.Value() + ac_cost;
-      } else {
-        cost_ = arc.weight.Value() + ac_cost;
-      }
-    }
-
-    // for epsilon transitions, i.e., non-emitting arcs
-    inline Token(const Arc &arc, Token *prev)
-        : arc_(arc), prev_(prev), ref_count_(1) {
-      if (prev) {
-        prev->ref_count_++;
-        cost_ = prev->cost_ + arc.weight.Value();
-      } else {
-        cost_ = arc.weight.Value();
-      }
-    }
-
-    inline bool operator<(const Token &other) const {
-      return cost_ > other.cost_;
-    }
-
-    inline static void TokenDelete(Token *tok) {
-      while (--tok->ref_count_ == 0) {
-        Token *prev = tok->prev_;
-        delete tok;
-        if (prev == nullptr) {
-          return;
-        } else {
-          tok = prev;
-        }
-      }
-    }
-  };
-
-  using Elem = HashList<StateId, Token *>::Elem;
-
-  /// Gets the weight cutoff.  Also counts the active tokens.
-  double GetCutoff(Elem *list_head, size_t *tok_count, float *adaptive_beam,
-                   Elem **best_elem);
-
-  void PossiblyResizeHash(size_t num_toks);
-
-  // ProcessEmitting returns the likelihood cutoff used.
-  // It decodes the frame num_frames_decoded_ of the decodable object
-  // and then increments num_frames_decoded_
-  double ProcessEmitting(DecodableInterface *decodable);
-
-  // TODO(dan): first time we go through this, could avoid using the queue.
-  void ProcessNonemitting(double cutoff);
-
-  // HashList defined in ../hash-list.h.  It actually allows us to maintain
-  // more than one list (e.g. for current and previous frames), but only one of
-  // them at a time can be indexed by StateId.
-  HashList<StateId, Token *> toks_;
-
-  const fst::Fst<fst::StdArc> &fst_;
-
-  FasterDecoderOptions config_;
-
-  // temp variable used in ProcessNonemitting,
-  std::vector<const Elem *> queue_;
-
-  std::vector<float> tmp_array_;  // used in GetCutoff.
-  // make it class member to avoid internal new/delete.
-
-  // Keep track of the number of frames decoded in the current file.
-  int32_t num_frames_decoded_;
-
-  // It might seem unclear why we call ClearToks(toks_.Clear()).
-  // There are two separate cleanup tasks we need to do at when we start a new
-  // file. one is to delete the Token objects in the list; the other is to
-  // delete the Elem objects.  toks_.Clear() just clears them from the hash and
-  // gives ownership to the caller, who then has to call toks_.Delete(e) for
-  // each one.  It was designed this way for convenience in propagating tokens
-  // from one frame to the next.
-  void ClearToks(Elem *list);
-};
-
-}  // namespace kaldi_decoder
-
-#endif  // KALDI_DECODER_CSRC_FASTER_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/hash-list-inl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/hash-list-inl.h
deleted file mode 100644
index a58ae22b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/hash-list-inl.h
+++ /dev/null
@@ -1,207 +0,0 @@
-// kaldi-decoder/csrc/hash-list-inl.h
-
-// Copyright 2009-2011   Microsoft Corporation
-//                2013   Johns Hopkins University (author: Daniel Povey)
-// Copyright (c)  2023  Xiaomi Corporation
-
-// this file is copied and modified from
-// kaldi/src/utils/hash-list-inl.h
-
-#ifndef KALDI_DECODER_CSRC_HASH_LIST_INL_H_
-#define KALDI_DECODER_CSRC_HASH_LIST_INL_H_
-
-// Do not include this file directly.  It is included by fast-hash.h
-
-namespace kaldi_decoder {
-
-template <class I, class T>
-HashList<I, T>::HashList() {
-  list_head_ = nullptr;
-  bucket_list_tail_ = static_cast<size_t>(-1);  // invalid.
-  hash_size_ = 0;
-  freed_head_ = nullptr;
-}
-
-template <class I, class T>
-void HashList<I, T>::SetSize(size_t size) {
-  hash_size_ = size;
-  KALDI_DECODER_ASSERT(list_head_ == nullptr &&
-                       bucket_list_tail_ ==
-                           static_cast<size_t>(-1));  // make sure empty.
-
-  if (size > buckets_.size()) {
-    buckets_.resize(size, HashBucket(0, nullptr));
-  }
-}
-
-template <class I, class T>
-typename HashList<I, T>::Elem *HashList<I, T>::Clear() {
-  // Clears the hashtable and gives ownership of the currently contained list
-  // to the user.
-  for (size_t cur_bucket = bucket_list_tail_;
-       cur_bucket != static_cast<size_t>(-1);
-       cur_bucket = buckets_[cur_bucket].prev_bucket) {
-    buckets_[cur_bucket].last_elem =
-        nullptr;  // this is how we indicate "empty".
-  }
-  bucket_list_tail_ = static_cast<size_t>(-1);
-  Elem *ans = list_head_;
-  list_head_ = nullptr;
-  return ans;
-}
-
-template <class I, class T>
-const typename HashList<I, T>::Elem *HashList<I, T>::GetList() const {
-  return list_head_;
-}
-
-template <class I, class T>
-inline void HashList<I, T>::Delete(Elem *e) {
-  e->tail = freed_head_;
-  freed_head_ = e;
-}
-
-template <class I, class T>
-inline typename HashList<I, T>::Elem *HashList<I, T>::Find(I key) {
-  size_t index = (static_cast<size_t>(key) % hash_size_);
-  HashBucket &bucket = buckets_[index];
-  if (bucket.last_elem == nullptr) {
-    return nullptr;  // empty bucket.
-  } else {
-    Elem *head = (bucket.prev_bucket == static_cast<size_t>(-1)
-                      ? list_head_
-                      : buckets_[bucket.prev_bucket].last_elem->tail),
-         *tail = bucket.last_elem->tail;
-    for (Elem *e = head; e != tail; e = e->tail) {
-      if (e->key == key) {
-        return e;
-      }
-    }
-
-    return nullptr;  // Not found.
-  }
-}
-
-template <class I, class T>
-inline typename HashList<I, T>::Elem *HashList<I, T>::New() {
-  if (freed_head_) {
-    Elem *ans = freed_head_;
-    freed_head_ = freed_head_->tail;
-    return ans;
-  } else {
-    Elem *tmp = new Elem[allocate_block_size_];
-    for (size_t i = 0; i + 1 < allocate_block_size_; i++) {
-      tmp[i].tail = tmp + i + 1;
-    }
-
-    tmp[allocate_block_size_ - 1].tail = nullptr;
-    freed_head_ = tmp;
-    allocated_.push_back(tmp);
-    return this->New();
-  }
-}
-
-template <class I, class T>
-HashList<I, T>::~HashList() {
-  // First test whether we had any memory leak within the
-  // HashList, i.e. things for which the user did not call Delete().
-  size_t num_in_list = 0, num_allocated = 0;
-
-  for (Elem *e = freed_head_; e != nullptr; e = e->tail) {
-    num_in_list++;
-  }
-
-  for (size_t i = 0; i < allocated_.size(); i++) {
-    num_allocated += allocate_block_size_;
-    delete[] allocated_[i];
-  }
-
-  if (num_in_list != num_allocated) {
-    KALDI_DECODER_WARN << "Possible memory leak: " << num_in_list
-                       << " != " << num_allocated
-                       << ": you might have forgotten to call Delete on "
-                       << "some Elems";
-  }
-}
-
-template <class I, class T>
-inline typename HashList<I, T>::Elem *HashList<I, T>::Insert(I key, T val) {
-  size_t index = (static_cast<size_t>(key) % hash_size_);
-  HashBucket &bucket = buckets_[index];
-  // Check the element is existing or not.
-  if (bucket.last_elem != nullptr) {
-    Elem *head = (bucket.prev_bucket == static_cast<size_t>(-1)
-                      ? list_head_
-                      : buckets_[bucket.prev_bucket].last_elem->tail),
-         *tail = bucket.last_elem->tail;
-
-    for (Elem *e = head; e != tail; e = e->tail) {
-      if (e->key == key) {
-        return e;
-      }
-    }
-  }
-
-  // This is a new element. Insert it.
-  Elem *elem = New();
-  elem->key = key;
-  elem->val = val;
-
-  if (bucket.last_elem == nullptr) {  // Unoccupied bucket.  Insert at
-    // head of bucket list (which is tail of regular list, they go in
-    // opposite directions).
-    if (bucket_list_tail_ == static_cast<size_t>(-1)) {
-      // list was empty so this is the first elem.
-      KALDI_DECODER_ASSERT(list_head_ == nullptr);
-      list_head_ = elem;
-    } else {
-      // link in to the chain of Elems
-      buckets_[bucket_list_tail_].last_elem->tail = elem;
-    }
-    elem->tail = nullptr;
-    bucket.last_elem = elem;
-    bucket.prev_bucket = bucket_list_tail_;
-    bucket_list_tail_ = index;
-  } else {
-    // Already-occupied bucket.  Insert at tail of list of elements within
-    // the bucket.
-    elem->tail = bucket.last_elem->tail;
-    bucket.last_elem->tail = elem;
-    bucket.last_elem = elem;
-  }
-  return elem;
-}
-
-template <class I, class T>
-void HashList<I, T>::InsertMore(I key, T val) {
-  size_t index = (static_cast<size_t>(key) % hash_size_);
-  HashBucket &bucket = buckets_[index];
-  Elem *elem = New();
-  elem->key = key;
-  elem->val = val;
-
-  // assume one element is already here
-  KALDI_DECODER_ASSERT(bucket.last_elem != nullptr);
-
-  if (bucket.last_elem->key == key) {  // standard behavior: add as last element
-    elem->tail = bucket.last_elem->tail;
-    bucket.last_elem->tail = elem;
-    bucket.last_elem = elem;
-    return;
-  }
-  Elem *e = (bucket.prev_bucket == static_cast<size_t>(-1)
-                 ? list_head_
-                 : buckets_[bucket.prev_bucket].last_elem->tail);
-  // find place to insert in linked list
-  while (e != bucket.last_elem->tail && e->key != key) {
-    e = e->tail;
-  }
-
-  KALDI_DECODER_ASSERT(e->key == key);  // not found? - should not happen
-  elem->tail = e->tail;
-  e->tail = elem;
-}
-
-}  // namespace kaldi_decoder
-
-#endif  // KALDI_DECODER_CSRC_HASH_LIST_INL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/hash-list.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/hash-list.h
deleted file mode 100644
index d006ef9e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/hash-list.h
+++ /dev/null
@@ -1,133 +0,0 @@
-// kaldi-decoder/csrc/hash-list.h
-
-// Copyright 2009-2011   Microsoft Corporation
-//                2013   Johns Hopkins University (author: Daniel Povey)
-// Copyright (c)  2023   Xiaomi Corporation
-
-// this file is copied and modified from
-// kaldi/src/utils/hash-list.h
-
-#ifndef KALDI_DECODER_CSRC_HASH_LIST_H_
-#define KALDI_DECODER_CSRC_HASH_LIST_H_
-
-#include <algorithm>
-#include <cassert>
-#include <limits>
-#include <set>
-#include <vector>
-
-#include "kaldi-decoder/csrc/stl-utils.h"
-
-/* This header provides utilities for a structure that's used in a decoder (but
-   is quite generic in nature so we implement and test it separately).
-   Basically it's a singly-linked list, but implemented in such a way that we
-   can quickly search for elements in the list.  We give it a slightly richer
-   interface than just a hash and a list.  The idea is that we want to separate
-   the hash part and the list part: basically, in the decoder, we want to have a
-   single hash for the current frame and the next frame, because by the time we
-   need to access the hash for the next frame we no longer need the hash for the
-   previous frame.  So we have an operation that clears the hash but leaves the
-   list structure intact.  We also control memory management inside this object,
-   to avoid repeated new's/deletes.
-*/
-
-namespace kaldi_decoder {
-
-template <class I, class T>
-class HashList {
- public:
-  struct Elem {
-    I key;
-    T val;
-    Elem *tail;
-  };
-
-  /// Constructor takes no arguments.
-  /// Call SetSize to inform it of the likely size.
-  HashList();
-
-  /// Clears the hash and gives the head of the current list to the user;
-  /// ownership is transferred to the user (the user must call Delete()
-  /// for each element in the list, at his/her leisure).
-  Elem *Clear();
-
-  /// Gives the head of the current list to the user.  Ownership retained in the
-  /// class.  Caution: in December 2013 the return type was changed to const
-  /// Elem* and this function was made const.  You may need to change some types
-  /// of local Elem* variables to const if this produces compilation errors.
-  const Elem *GetList() const;
-
-  /// Think of this like delete().  It is to be called for each Elem in turn
-  /// after you "obtained ownership" by doing Clear().  This is not the opposite
-  /// of Insert, it is the opposite of New.  It's really a memory operation.
-  inline void Delete(Elem *e);
-
-  /// This should probably not be needed to be called directly by the user.
-  /// Think of it as opposite
-  /// to Delete();
-  inline Elem *New();
-
-  /// Find tries to find this element in the current list using the hashtable.
-  /// It returns NULL if not present.  The Elem it returns is not owned by the
-  /// user, it is part of the internal list owned by this object, but the user
-  /// is free to modify the "val" element.
-  inline Elem *Find(I key);
-
-  /// Insert inserts a new element into the hashtable/stored list.
-  /// Because element keys in a hashtable are unique, this operation checks
-  /// whether each inserted element has a key equivalent to the one of an
-  /// element already in the hashtable. If so, the element is not inserted,
-  /// returning an pointer to this existing element.
-  inline Elem *Insert(I key, T val);
-
-  /// Insert inserts another element with same key into the hashtable/
-  /// stored list.
-  /// By calling this, the user asserts that one element with that key is
-  /// already present.
-  /// We insert it that way, that all elements with the same key
-  /// follow each other.
-  /// Find() will return the first one of the elements with the same key.
-  inline void InsertMore(I key, T val);
-
-  /// SetSize tells the object how many hash buckets to allocate (should
-  /// typically be at least twice the number of objects we expect to go in the
-  /// structure, for fastest performance).  It must be called while the hash
-  /// is empty (e.g. after Clear() or after initializing the object, but before
-  /// adding anything to the hash.
-  void SetSize(size_t sz);
-
-  /// Returns current number of hash buckets.
-  inline size_t Size() const { return hash_size_; }
-
-  ~HashList();
-
- private:
-  struct HashBucket {
-    size_t prev_bucket;  // index to next bucket (-1 if list tail).  Note:
-    // list of buckets goes in opposite direction to list of Elems.
-    Elem *last_elem;  // pointer to last element in this bucket (NULL if empty)
-    inline HashBucket(size_t i, Elem *e) : prev_bucket(i), last_elem(e) {}
-  };
-
-  Elem *list_head_;          // head of currently stored list.
-  size_t bucket_list_tail_;  // tail of list of active hash buckets.
-
-  size_t hash_size_;  // number of hash buckets.
-
-  std::vector<HashBucket> buckets_;
-
-  Elem *freed_head_;  // head of list of currently freed elements. [ready for
-  // allocation]
-
-  std::vector<Elem *> allocated_;  // list of allocated blocks.
-
-  static const size_t allocate_block_size_ = 1024;  // Number of Elements to
-  // allocate in one block.  Must be largish so storing allocated_ doesn't
-  // become a problem.
-};
-
-}  // namespace kaldi_decoder
-
-#include "kaldi-decoder/csrc/hash-list-inl.h"
-
-#endif  // KALDI_DECODER_CSRC_HASH_LIST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/kaldi-math.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/kaldi-math.h
deleted file mode 100644
index 02fd4c00..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/kaldi-math.h
+++ /dev/null
@@ -1,48 +0,0 @@
-// kaldi-decoder/csrc/kaldi-math.h
-
-// Copyright 2009-2011  Ondrej Glembek;  Microsoft Corporation;  Yanmin Qian;
-//                      Jan Silovsky;  Saarland University
-//                2023  Xiaomi Corporation
-
-// this file is copied and modified from
-// kaldi/src/base/kaldi-math.h
-#ifndef KALDI_DECODER_CSRC_KALDI_MATH_H_
-#define KALDI_DECODER_CSRC_KALDI_MATH_H_
-
-#include <cmath>
-#include <limits>
-
-#ifndef DBL_EPSILON
-#define DBL_EPSILON 2.2204460492503131e-16
-#endif
-
-#ifndef FLT_EPSILON
-#define FLT_EPSILON 1.19209290e-7f
-#endif
-
-// M_LOG_2PI =  log(2*pi)
-#ifndef M_LOG_2PI
-#define M_LOG_2PI 1.8378770664093454835606594728112
-#endif
-
-#define KALDI_ISINF std::isinf
-#define KALDI_ISNAN std::isnan
-
-#include "kaldi-decoder/csrc/log.h"
-
-namespace kaldi_decoder {
-
-/// return abs(a - b) <= relative_tolerance * (abs(a)+abs(b)).
-static inline bool ApproxEqual(float a, float b,
-                               float relative_tolerance = 0.001) {
-  // a==b handles infinities.
-  if (a == b) return true;
-  float diff = std::abs(a - b);
-  if (diff == std::numeric_limits<float>::infinity() || diff != diff)
-    return false;  // diff is +inf or nan.
-  return (diff <= relative_tolerance * (std::abs(a) + std::abs(b)));
-}
-
-}  // namespace kaldi_decoder
-
-#endif  // KALDI_DECODER_CSRC_KALDI_MATH_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/lattice-faster-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/lattice-faster-decoder.h
deleted file mode 100644
index 07176601..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/lattice-faster-decoder.h
+++ /dev/null
@@ -1,274 +0,0 @@
-// kaldi-decoder/csrc/lattice-faster-decoder.h
-
-// Copyright 2009-2013  Microsoft Corporation;  Mirko Hannemann;
-//           2013-2014  Johns Hopkins University (Author: Daniel Povey)
-//                2014  Guoguo Chen
-//                2018  Zhehuai Chen
-// Copyright (c)  2023  Xiaomi Corporation
-
-// this file is copied and modified from
-// kaldi/src/decoder/lattice-faster-decoder.h
-#ifndef KALDI_DECODER_CSRC_LATTICE_FASTER_DECODER_H_
-#define KALDI_DECODER_CSRC_LATTICE_FASTER_DECODER_H_
-
-#include <limits>
-#include <string>
-
-#include "fst/fst.h"
-#include "fst/fstlib.h"
-#include "kaldi-decoder/csrc/log.h"
-
-namespace kaldi_decoder {
-
-struct LatticeFasterDecoderConfig {
-  float beam;
-  int32_t max_active;
-  int32_t min_active;
-  float lattice_beam;
-  int32_t prune_interval;
-  bool determinize_lattice;  // not inspected by this class... used in
-                             // command-line program.
-  float beam_delta;
-  float hash_ratio;
-  // Note: we don't make prune_scale configurable on the command line, it's not
-  // a very important parameter.  It affects the algorithm that prunes the
-  // tokens as we go.
-  float prune_scale;
-
-  // Number of elements in the block for Token and ForwardLink memory
-  // pool allocation.
-  int32_t memory_pool_tokens_block_size;
-  int32_t memory_pool_links_block_size;
-
-  // Most of the options inside det_opts are not actually queried by the
-  // LatticeFasterDecoder class itself, but by the code that calls it, for
-  // example in the function DecodeUtteranceLatticeFaster.
-  // fst::DeterminizeLatticePhonePrunedOptions det_opts;
-
-  LatticeFasterDecoderConfig(
-      float beam = 16.0,
-      int32_t max_active = std::numeric_limits<int32_t>::max(),
-      int32_t min_active = 200, float lattice_beam = 10.0,
-      int32_t prune_interval = 25, bool determinize_lattice = true,
-      float beam_delta = 0.5, float hash_ratio = 2.0, float prune_scale = 0.1,
-      int32_t memory_pool_tokens_block_size = 1 << 8,
-      int32_t memory_pool_links_block_size = 1 << 8)
-      : beam(beam),
-        max_active(max_active),
-        min_active(min_active),
-        lattice_beam(lattice_beam),
-        prune_interval(prune_interval),
-        determinize_lattice(determinize_lattice),
-        beam_delta(beam_delta),
-        hash_ratio(hash_ratio),
-        prune_scale(prune_scale),
-        memory_pool_tokens_block_size(memory_pool_tokens_block_size),
-        memory_pool_links_block_size(memory_pool_links_block_size) {}
-#if 0
-  void Register(OptionsItf *opts) {
-    det_opts.Register(opts);
-    opts->Register("beam", &beam,
-                   "Decoding beam.  Larger->slower, more accurate.");
-    opts->Register("max-active", &max_active,
-                   "Decoder max active states.  Larger->slower; "
-                   "more accurate");
-    opts->Register("min-active", &min_active,
-                   "Decoder minimum #active states.");
-    opts->Register("lattice-beam", &lattice_beam,
-                   "Lattice generation beam.  Larger->slower, "
-                   "and deeper lattices");
-    opts->Register("prune-interval", &prune_interval,
-                   "Interval (in frames) at "
-                   "which to prune tokens");
-    opts->Register(
-        "determinize-lattice", &determinize_lattice,
-        "If true, "
-        "determinize the lattice (lattice-determinization, keeping only "
-        "best pdf-sequence for each word-sequence).");
-    opts->Register(
-        "beam-delta", &beam_delta,
-        "Increment used in decoding-- this "
-        "parameter is obscure and relates to a speedup in the way the "
-        "max-active constraint is applied.  Larger is more accurate.");
-    opts->Register("hash-ratio", &hash_ratio,
-                   "Setting used in decoder to "
-                   "control hash behavior");
-    opts->Register(
-        "memory-pool-tokens-block-size", &memory_pool_tokens_block_size,
-        "Memory pool block size suggestion for storing tokens (in elements). "
-        "Smaller uses less memory but increases cache misses.");
-    opts->Register(
-        "memory-pool-links-block-size", &memory_pool_links_block_size,
-        "Memory pool block size suggestion for storing links (in elements). "
-        "Smaller uses less memory but increases cache misses.");
-  }
-#endif
-  std::string ToString() const {
-    std::ostringstream os;
-
-    os << "LatticeFasterDecoderConfig(";
-    os << "beam=" << beam << ", ";
-    os << "max_active=" << max_active << ", ";
-    os << "min_active=" << min_active << ", ";
-    os << "lattice_beam=" << lattice_beam << ", ";
-    os << "prune_interval=" << prune_interval << ", ";
-    os << "determinize_lattice=" << (determinize_lattice ? "True" : "False")
-       << ", ";
-
-    os << "beam_delta=" << beam_delta << ", ";
-    os << "hash_ratio=" << hash_ratio << ", ";
-    os << "prune_scale=" << prune_scale << ", ";
-    os << "memory_pool_tokens_block_size=" << memory_pool_tokens_block_size
-       << ", ";
-    os << "memory_pool_links_block_size=" << memory_pool_links_block_size
-       << ")";
-
-    return os.str();
-  }
-  void Check() const {
-    KALDI_DECODER_ASSERT(beam > 0.0 && max_active > 1 && lattice_beam > 0.0 &&
-                         min_active <= max_active && prune_interval > 0 &&
-                         beam_delta > 0.0 && hash_ratio >= 1.0 &&
-                         prune_scale > 0.0 && prune_scale < 1.0);
-  }
-};
-
-namespace decoder {
-// We will template the decoder on the token type as well as the FST type; this
-// is a mechanism so that we can use the same underlying decoder code for
-// versions of the decoder that support quickly getting the best path
-// (LatticeFasterOnlineDecoder, see lattice-faster-online-decoder.h) and also
-// those that do not (LatticeFasterDecoder).
-
-// ForwardLinks are the links from a token to a token on the next frame.
-// or sometimes on the current frame (for input-epsilon links).
-template <typename Token>
-struct ForwardLink {
-  using Label = fst::StdArc::Label;
-
-  Token *next_tok;      // the next token [or NULL if represents final-state]
-  Label ilabel;         // ilabel on arc
-  Label olabel;         // olabel on arc
-  float graph_cost;     // graph cost of traversing arc (contains LM, etc.)
-  float acoustic_cost;  // acoustic cost (pre-scaled) of traversing arc
-  ForwardLink *next;    // next in singly-linked list of forward arcs (arcs
-                        // in the state-level lattice) from a token.
-  ForwardLink(Token *next_tok, Label ilabel, Label olabel, float graph_cost,
-              float acoustic_cost, ForwardLink *next)
-      : next_tok(next_tok),
-        ilabel(ilabel),
-        olabel(olabel),
-        graph_cost(graph_cost),
-        acoustic_cost(acoustic_cost),
-        next(next) {}
-};
-
-struct StdToken {
-  using ForwardLinkT = ForwardLink<StdToken>;
-  using Token = StdToken;
-
-  // Standard token type for LatticeFasterDecoder.  Each active HCLG
-  // (decoding-graph) state on each frame has one token.
-
-  // tot_cost is the total (LM + acoustic) cost from the beginning of the
-  // utterance up to this point.  (but see cost_offset_, which is subtracted
-  // to keep it in a good numerical range).
-  float tot_cost;
-
-  // extra_cost is >= 0.  After calling PruneForwardLinks, this equals the
-  // minimum difference between the cost of the best path that this link is a
-  // part of, and the cost of the absolute best path, under the assumption that
-  // any of the currently active states at the decoding front may eventually
-  // succeed (e.g. if you were to take the currently active states one by one
-  // and compute this difference, and then take the minimum).
-  float extra_cost;
-
-  // 'links' is the head of singly-linked list of ForwardLinks, which is what we
-  // use for lattice generation.
-  ForwardLinkT *links;
-
-  //'next' is the next in the singly-linked list of tokens for this frame.
-  Token *next;
-
-  // This function does nothing and should be optimized out; it's needed
-  // so we can share the regular LatticeFasterDecoderTpl code and the code
-  // for LatticeFasterOnlineDecoder that supports fast traceback.
-  inline void SetBackpointer(Token *backpointer) {}
-
-  // This constructor just ignores the 'backpointer' argument.  That argument is
-  // needed so that we can use the same decoder code for LatticeFasterDecoderTpl
-  // and LatticeFasterOnlineDecoderTpl (which needs backpointers to support a
-  // fast way to obtain the best path).
-  StdToken(float tot_cost, float extra_cost, ForwardLinkT *links, Token *next,
-           Token *backpointer)
-      : tot_cost(tot_cost), extra_cost(extra_cost), links(links), next(next) {}
-};
-
-struct BackpointerToken {
-  using ForwardLinkT = ForwardLink<BackpointerToken>;
-  using Token = BackpointerToken;
-
-  // BackpointerToken is like Token but also
-  // Standard token type for LatticeFasterDecoder.  Each active HCLG
-  // (decoding-graph) state on each frame has one token.
-
-  // tot_cost is the total (LM + acoustic) cost from the beginning of the
-  // utterance up to this point.  (but see cost_offset_, which is subtracted
-  // to keep it in a good numerical range).
-  float tot_cost;
-
-  // exta_cost is >= 0.  After calling PruneForwardLinks, this equals
-  // the minimum difference between the cost of the best path, and the cost of
-  // this is on, and the cost of the absolute best path, under the assumption
-  // that any of the currently active states at the decoding front may
-  // eventually succeed (e.g. if you were to take the currently active states
-  // one by one and compute this difference, and then take the minimum).
-  float extra_cost;
-
-  // 'links' is the head of singly-linked list of ForwardLinks, which is what we
-  // use for lattice generation.
-  ForwardLinkT *links;
-
-  //'next' is the next in the singly-linked list of tokens for this frame.
-  BackpointerToken *next;
-
-  // Best preceding BackpointerToken (could be a on this frame, connected to
-  // this via an epsilon transition, or on a previous frame).  This is only
-  // required for an efficient GetBestPath function in
-  // LatticeFasterOnlineDecoderTpl; it plays no part in the lattice generation
-  // (the "links" list is what stores the forward links, for that).
-  Token *backpointer;
-
-  void SetBackpointer(Token *backpointer) { this->backpointer = backpointer; }
-
-  BackpointerToken(float tot_cost, float extra_cost, ForwardLinkT *links,
-                   Token *next, Token *backpointer)
-      : tot_cost(tot_cost),
-        extra_cost(extra_cost),
-        links(links),
-        next(next),
-        backpointer(backpointer) {}
-};
-
-}  // namespace decoder
-
-/** This is the "normal" lattice-generating decoder.
-    See \ref lattices_generation \ref decoders_faster and \ref decoders_simple
-     for more information.
-
-   The decoder is templated on the FST type and the token type.  The token type
-   will normally be StdToken, but also may be BackpointerToken which is to
-   support quick lookup of the current best path (see
-   lattice-faster-online-decoder.h)
-
-   The FST you invoke this decoder which is expected to equal
-   Fst::Fst<fst::StdArc>, a.k.a. StdFst, or GrammarFst.  If you invoke it with
-   FST == StdFst and it notices that the actual FST type is
-   fst::VectorFst<fst::StdArc> or fst::ConstFst<fst::StdArc>, the decoder object
-   will internally cast itself to one that is templated on those more specific
-   types; this is an optimization for speed.
- */
-
-}  // namespace kaldi_decoder
-
-#endif  // KALDI_DECODER_CSRC_LATTICE_FASTER_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/lattice-simple-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/lattice-simple-decoder.h
deleted file mode 100644
index b1269d58..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/lattice-simple-decoder.h
+++ /dev/null
@@ -1,324 +0,0 @@
-// kaldi-decoder/csrc/lattice-simple-decoder.h
-
-// Copyright 2009-2012  Microsoft Corporation
-//           2012-2014  Johns Hopkins University (Author: Daniel Povey)
-//                2014  Guoguo Chen
-// Copyright (c)  2023  Xiaomi Corporation
-
-// this file is copied and modified from
-// kaldi/src/decoder/lattice-simple-decoder.h
-#ifndef KALDI_DECODER_CSRC_LATTICE_SIMPLE_DECODER_H_
-#define KALDI_DECODER_CSRC_LATTICE_SIMPLE_DECODER_H_
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-#include "fst/fst.h"
-#include "fst/fstlib.h"
-#include "kaldi-decoder/csrc/decodable-itf.h"
-#include "kaldi-decoder/csrc/log.h"
-#include "kaldifst/csrc/lattice-weight.h"
-
-namespace kaldi_decoder {
-
-struct LatticeSimpleDecoderConfig {
-  float beam;
-  float lattice_beam;
-  int32_t prune_interval;
-  bool determinize_lattice;  // not inspected by this class... used in
-  // command-line program.
-  bool prune_lattice;
-  float beam_ratio;
-  float prune_scale;  // Note: we don't make this configurable on the command
-                      // line, it's not a very important parameter.  It affects
-                      // the algorithm that prunes the tokens as we go.
-  // fst::DeterminizeLatticePhonePrunedOptions det_opts;
-
-  LatticeSimpleDecoderConfig(float beam = 16.0, float lattice_beam = 10.0,
-                             int32_t prune_interval = 25,
-                             bool determinize_lattice = true,
-                             bool prune_lattice = true, float beam_ratio = 0.9,
-                             float prune_scale = 0.1)
-      : beam(beam),
-        lattice_beam(lattice_beam),
-        prune_interval(prune_interval),
-        determinize_lattice(determinize_lattice),
-        prune_lattice(prune_lattice),
-        beam_ratio(beam_ratio),
-        prune_scale(prune_scale) {}
-#if 0
-  void Register(OptionsItf *opts) {
-    det_opts.Register(opts);
-    opts->Register("beam", &beam, "Decoding beam.");
-    opts->Register("lattice-beam", &lattice_beam, "Lattice generation beam");
-    opts->Register("prune-interval", &prune_interval,
-                   "Interval (in frames) at "
-                   "which to prune tokens");
-    opts->Register("determinize-lattice", &determinize_lattice,
-                   "If true, "
-                   "determinize the lattice (in a special sense, keeping only "
-                   "best pdf-sequence for each word-sequence).");
-  }
-#endif
-  void Check() const {
-    KALDI_DECODER_ASSERT(beam > 0.0 && lattice_beam > 0.0 &&
-                         prune_interval > 0);
-  }
-
-  std::string ToString() const {
-    std::ostringstream os;
-
-    os << "LatticeSimpleDecoderConfig(";
-    os << "beam=" << beam << ", ";
-    os << "lattice_beam=" << lattice_beam << ", ";
-    os << "prune_interval=" << prune_interval << ", ";
-    os << "determinize_lattice=" << (determinize_lattice ? "True" : "False")
-       << ", ";
-
-    os << "prune_lattice=" << prune_lattice << ", ";
-    os << "beam_ratio=" << beam_ratio << ", ";
-    os << "prune_scale=" << prune_scale << ")";
-
-    return os.str();
-  }
-};
-
-/** Simplest possible decoder, included largely for didactic purposes and as a
-    means to debug more highly optimized decoders.  See \ref decoders_simple
-    for more information.
- */
-class LatticeSimpleDecoder {
- public:
-  using Arc = fst::StdArc;
-  using Label = Arc::Label;
-  using StateId = Arc::StateId;
-  using Weight = Arc::Weight;
-
-  // instantiate this class once for each thing you have to decode.
-  LatticeSimpleDecoder(const fst::Fst<fst::StdArc> &fst,
-                       const LatticeSimpleDecoderConfig &config)
-      : fst_(fst), config_(config), num_toks_(0) {
-    config.Check();
-  }
-
-  /// InitDecoding initializes the decoding, and should only be used if you
-  /// intend to call AdvanceDecoding().  If you call Decode(), you don't need
-  /// to call this.  You can call InitDecoding if you have already decoded an
-  /// utterance and want to start with a new utterance.
-  void InitDecoding();
-
-  /// This function may be optionally called after AdvanceDecoding(), when you
-  /// do not plan to decode any further.  It does an extra pruning step that
-  /// will help to prune the lattices output by GetLattice and (particularly)
-  /// GetRawLattice more accurately, particularly toward the end of the
-  /// utterance.  It does this by using the final-probs in pruning (if any
-  /// final-state survived); it also does a final pruning step that visits all
-  /// states (the pruning that is done during decoding may fail to prune states
-  /// that are within kPruningScale = 0.1 outside of the beam).  If you call
-  /// this, you cannot call AdvanceDecoding again (it will fail), and you
-  /// cannot call GetLattice() and related functions with use_final_probs =
-  /// false.
-  /// Used to be called PruneActiveTokensFinal().
-  void FinalizeDecoding();
-
-  // Outputs an FST corresponding to the single best path
-  // through the lattice.  Returns true if result is nonempty
-  // (using the return status is deprecated, it will become void).
-  // If "use_final_probs" is true AND we reached the final-state
-  // of the graph then it will include those as final-probs, else
-  // it will treat all final-probs as one.
-  bool GetBestPath(fst::Lattice *lat, bool use_final_probs = true) const;
-
-  // Outputs an FST corresponding to the raw, state-level
-  // tracebacks.  Returns true if result is nonempty
-  // (using the return status is deprecated, it will become void).
-  // If "use_final_probs" is true AND we reached the final-state
-  // of the graph then it will include those as final-probs, else
-  // it will treat all final-probs as one.
-  bool GetRawLattice(fst::Lattice *lat, bool use_final_probs = true) const;
-
-  ~LatticeSimpleDecoder() { ClearActiveTokens(); }
-
-  /// FinalRelativeCost() serves the same purpose as ReachedFinal(), but gives
-  /// more information.  It returns the difference between the best (final-cost
-  /// plus cost) of any token on the final frame, and the best cost of any token
-  /// on the final frame.  If it is infinity it means no final-states were
-  /// present on the final frame.  It will usually be nonnegative.  If it not
-  /// too positive (e.g. < 5 is my first guess, but this is not tested) you can
-  /// take it as a good indication that we reached the final-state with
-  /// reasonable likelihood.
-  float FinalRelativeCost() const;
-
-  const LatticeSimpleDecoderConfig &GetOptions() const { return config_; }
-
-  int32_t NumFramesDecoded() const { return active_toks_.size() - 1; }
-
-  // Returns true if any kind of traceback is available (not necessarily from
-  // a final state).
-  bool Decode(DecodableInterface *decodable);
-
- private:
-  struct Token;
-  // ForwardLinks are the links from a token to a token on the next frame.
-  // or sometimes on the current frame (for input-epsilon links).
-  struct ForwardLink {
-    Token *next_tok;      // the next token [or NULL if represents final-state]
-    Label ilabel;         // ilabel on link.
-    Label olabel;         // olabel on link.
-    float graph_cost;     // graph cost of traversing link (contains LM, etc.)
-    float acoustic_cost;  // acoustic cost (pre-scaled) of traversing link
-    ForwardLink *next;    // next in singly-linked list of forward links from a
-                          // token.
-    ForwardLink(Token *next_tok, Label ilabel, Label olabel, float graph_cost,
-                float acoustic_cost, ForwardLink *next)
-        : next_tok(next_tok),
-          ilabel(ilabel),
-          olabel(olabel),
-          graph_cost(graph_cost),
-          acoustic_cost(acoustic_cost),
-          next(next) {}
-  };
-
-  // Token is what's resident in a particular state at a particular time.
-  // In this decoder a Token actually contains *forward* links.
-  // When first created, a Token just has the (total) cost.    We add forward
-  // links from it when we process the next frame.
-  struct Token {
-    float tot_cost;    // would equal weight.Value()... cost up to this point.
-    float extra_cost;  // >= 0.  After calling PruneForwardLinks, this equals
-    // the minimum difference between the cost of the best path this is on,
-    // and the cost of the absolute best path, under the assumption
-    // that any of the currently active states at the decoding front may
-    // eventually succeed (e.g. if you were to take the currently active states
-    // one by one and compute this difference, and then take the minimum).
-
-    ForwardLink *links;  // Head of singly linked list of ForwardLinks
-
-    Token *next;  // Next in list of tokens for this frame.
-
-    Token(float tot_cost, float extra_cost, ForwardLink *links, Token *next)
-        : tot_cost(tot_cost),
-          extra_cost(extra_cost),
-          links(links),
-          next(next) {}
-
-    Token() = default;
-
-    void DeleteForwardLinks() {
-      ForwardLink *l = links;
-      ForwardLink *m;
-
-      while (l != nullptr) {
-        m = l->next;
-        delete l;
-        l = m;
-      }
-      links = nullptr;
-    }
-  };
-
-  // head and tail of per-frame list of Tokens (list is in topological order),
-  // and something saying whether we ever pruned it using PruneForwardLinks.
-  struct TokenList {
-    Token *toks;
-    bool must_prune_forward_links;
-    bool must_prune_tokens;
-    TokenList()
-        : toks(nullptr),
-          must_prune_forward_links(true),
-          must_prune_tokens(true) {}
-  };
-
-  // FindOrAddToken either locates a token in cur_toks_, or if necessary inserts
-  // a new, empty token (i.e. with no forward links) for the current frame.
-  // [note: it's inserted if necessary into cur_toks_ and also into the singly
-  // linked list of tokens active on this frame (whose head is at
-  // active_toks_[frame]).
-  //
-  // Returns the Token pointer.  Sets "changed" (if non-NULL) to true
-  // if the token was newly created or the cost changed.
-  inline Token *FindOrAddToken(StateId state, int32_t frame_plus_one,
-                               float tot_cost, bool emitting, bool *changed);
-
-  // delta is the amount by which the extra_costs must
-  // change before it sets "extra_costs_changed" to true.  If delta is larger,
-  // we'll tend to go back less far toward the beginning of the file.
-  void PruneForwardLinks(int32_t frame, bool *extra_costs_changed,
-                         bool *links_pruned, float delta);
-
-  // Prune away any tokens on this frame that have no forward links. [we don't
-  // do this in PruneForwardLinks because it would give us a problem with
-  // dangling pointers].
-  void PruneTokensForFrame(int32_t frame);
-
-  void ClearActiveTokens();  // a cleanup routine, at utt end/begin
-
-  void ProcessNonemitting();
-
-  // PruneForwardLinksFinal is a version of PruneForwardLinks that we call
-  // on the final frame.  If there are final tokens active, it uses the
-  // final-probs for pruning, otherwise it treats all tokens as final.
-  void PruneForwardLinksFinal();
-
-  // This function computes the final-costs for tokens active on the final
-  // frame.  It outputs to final-costs, if non-NULL, a map from the Token*
-  // pointer to the final-prob of the corresponding state, or zero for all
-  // states if none were final.  It outputs to final_relative_cost, if non-NULL,
-  // the difference between the best forward-cost including the final-prob cost,
-  // and the best forward-cost without including the final-prob cost (this will
-  // usually be positive), or infinity if there were no final-probs.  It outputs
-  // to final_best_cost, if non-NULL, the lowest for any token t active on the
-  // final frame, of t + final-cost[t], where final-cost[t] is the final-cost
-  // in the graph of the state corresponding to token t, or zero if there
-  // were no final-probs active on the final frame.
-  // You cannot call this after FinalizeDecoding() has been called; in that
-  // case you should get the answer from class-member variables.
-  void ComputeFinalCosts(std::unordered_map<Token *, float> *final_costs,
-                         float *final_relative_cost,
-                         float *final_best_cost) const;
-
-  // Go backwards through still-alive tokens, pruning them if the
-  // forward+backward cost is more than lat_beam away from the best path.  It's
-  // possible to prove that this is "correct" in the sense that we won't lose
-  // anything outside of lat_beam, regardless of what happens in the future.
-  // delta controls when it considers a cost to have changed enough to continue
-  // going backward and propagating the change.  larger delta -> will recurse
-  // less far.
-  void PruneActiveTokens(float delta);
-
-  // PruneCurrentTokens deletes the tokens from the "toks" map, but not
-  // from the active_toks_ list, which could cause dangling forward pointers
-  // (will delete it during regular pruning operation).
-  void PruneCurrentTokens(float beam,
-                          std::unordered_map<StateId, Token *> *toks);
-
-  void ProcessEmitting(DecodableInterface *decodable);
-
- private:
-  const fst::Fst<fst::StdArc> &fst_;
-  LatticeSimpleDecoderConfig config_;
-  int32_t num_toks_;  // current total #toks allocated...
-  bool warned_;
-
-  std::unordered_map<StateId, Token *> cur_toks_;
-  std::unordered_map<StateId, Token *> prev_toks_;
-  std::vector<TokenList> active_toks_;  // Lists of tokens, indexed by frame
-
-  /// decoding_finalized_ is true if someone called FinalizeDecoding().  [note,
-  /// calling this is optional].  If true, it's forbidden to decode more.  Also,
-  /// if this is set, then the output of ComputeFinalCosts() is in the next
-  /// three variables.  The reason we need to do this is that after
-  /// FinalizeDecoding() calls PruneTokensForFrame() for the final frame, some
-  /// of the tokens on the last frame are freed, so we free the list from
-  /// cur_toks_ to avoid having dangling pointers hanging around.
-  bool decoding_finalized_;
-  /// For the meaning of the next 3 variables, see the comment for
-  /// decoding_finalized_ above., and ComputeFinalCosts().
-  std::unordered_map<Token *, float> final_costs_;
-  float final_relative_cost_;
-  float final_best_cost_;
-};
-
-}  // namespace kaldi_decoder
-
-#endif  // KALDI_DECODER_CSRC_LATTICE_SIMPLE_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/log.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/log.h
deleted file mode 100644
index f3ad8067..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/log.h
+++ /dev/null
@@ -1,100 +0,0 @@
-// kaldi-decoder/csrc/log.h
-//
-// Copyright (c)  2022  Xiaomi Corporation
-
-#ifndef KALDI_DECODER_CSRC_LOG_H_
-#define KALDI_DECODER_CSRC_LOG_H_
-
-#include <cstdint>
-#include <iostream>
-#include <sstream>
-#include <stdexcept>
-
-namespace kaldi_decoder {
-
-enum class LogLevel {
-  kInfo = 0,
-  kWarn = 1,
-  kError = 2,  // abort the program
-};
-
-class Logger {
- public:
-  Logger(const char *filename, const char *func_name, uint32_t line_num,
-         LogLevel level)
-      : level_(level) {
-    os_ << filename << ":" << func_name << ":" << line_num << "\n";
-    switch (level_) {
-      case LogLevel::kInfo:
-        os_ << "[I] ";
-        break;
-      case LogLevel::kWarn:
-        os_ << "[W] ";
-        break;
-      case LogLevel::kError:
-        os_ << "[E] ";
-        break;
-    }
-  }
-
-  template <typename T>
-  Logger &operator<<(const T &val) {
-    os_ << val;
-    return *this;
-  }
-
-  ~Logger() noexcept(false) {
-    if (level_ == LogLevel::kError) {
-      // throw std::runtime_error(os_.str());
-      // abort();
-      throw std::runtime_error(os_.str());
-    }
-    // fprintf(stderr, "%s\n", os_.str().c_str());
-  }
-
- private:
-  std::ostringstream os_;
-  LogLevel level_;
-};
-
-class Voidifier {
- public:
-  void operator&(const Logger &) const {}
-};
-
-#if defined(__clang__) || defined(__GNUC__) || defined(__GNUG__) || \
-    defined(__PRETTY_FUNCTION__)
-// for clang and GCC
-#define KALDI_DECODER_FUNC __PRETTY_FUNCTION__
-#else
-// for other compilers
-#define KALDI_DECODER_FUNC __func__
-#endif
-
-#define KALDI_DECODER_LOG                                       \
-  kaldi_decoder::Logger(__FILE__, KALDI_DECODER_FUNC, __LINE__, \
-                        kaldi_decoder::LogLevel::kInfo)
-
-#define KALDI_DECODER_WARN                                      \
-  kaldi_decoder::Logger(__FILE__, KALDI_DECODER_FUNC, __LINE__, \
-                        kaldi_decoder::LogLevel::kWarn)
-
-#define KALDI_DECODER_ERR                                       \
-  kaldi_decoder::Logger(__FILE__, KALDI_DECODER_FUNC, __LINE__, \
-                        kaldi_decoder::LogLevel::kError)
-
-#define KALDI_DECODER_ASSERT(x)                                             \
-  (x) ? (void)0                                                             \
-      : kaldi_decoder::Voidifier() & KALDI_DECODER_ERR << "Check failed!\n" \
-                                                       << "x: " << #x
-
-#define KALDI_DECODER_PARANOID_ASSERT KALDI_DECODER_ASSERT
-
-#define KALDI_DECODER_DISALLOW_COPY_AND_ASSIGN(Class) \
- public:                                              \
-  Class(const Class &) = delete;                      \
-  Class &operator=(const Class &) = delete;
-
-}  // namespace kaldi_decoder
-
-#endif  // KALDI_DECODER_CSRC_LOG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/simple-decoder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/simple-decoder.h
deleted file mode 100644
index d45df763..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/simple-decoder.h
+++ /dev/null
@@ -1,138 +0,0 @@
-// kaldi-decoder/csrc/simple-decoder.h
-
-// Copyright 2009-2013  Microsoft Corporation;  Lukas Burget;
-//                      Saarland University (author: Arnab Ghoshal);
-//                      Johns Hopkins University (author: Daniel Povey)
-// Copyright (c)  2023  Xiaomi Corporation
-#ifndef KALDI_DECODER_CSRC_SIMPLE_DECODER_H_
-#define KALDI_DECODER_CSRC_SIMPLE_DECODER_H_
-
-#include <unordered_map>
-
-#include "fst/fst.h"
-#include "fst/fstlib.h"
-#include "kaldi-decoder/csrc/decodable-itf.h"
-#include "kaldi-decoder/csrc/log.h"
-#include "kaldifst/csrc/lattice-weight.h"
-
-namespace kaldi_decoder {
-
-/** Simplest possible decoder, included largely for didactic purposes and as a
-    means to debug more highly optimized decoders.  See \ref decoders_simple
-    for more information.
- */
-class SimpleDecoder {
- public:
-  using StdArc = fst::StdArc;
-  using StdWeight = StdArc::Weight;
-  using Label = StdArc::Label;
-  using StateId = StdArc::StateId;
-
-  SimpleDecoder(const fst::Fst<fst::StdArc> &fst, float beam)
-      : fst_(fst), beam_(beam) {}
-
-  ~SimpleDecoder();
-  SimpleDecoder(const SimpleDecoder &) = delete;
-  SimpleDecoder &operator=(const SimpleDecoder &) = delete;
-
-  /// Decode this utterance.
-  /// Returns true if any tokens reached the end of the file (regardless of
-  /// whether they are in a final state); query ReachedFinal() after Decode()
-  /// to see whether we reached a final state.
-  bool Decode(DecodableInterface *decodable);
-
-  bool ReachedFinal() const;
-
-  // GetBestPath gets the decoding traceback. If "use_final_probs" is true
-  // AND we reached a final state, it limits itself to final states;
-  // otherwise it gets the most likely token not taking into account
-  // final-probs. fst_out will be empty (Start() == kNoStateId) if nothing was
-  // available due to search error. If Decode() returned true, it is safe to
-  // assume GetBestPath will return true. It returns true if the output lattice
-  // was nonempty (i.e. had states in it); using the return value is deprecated.
-  bool GetBestPath(fst::Lattice *fst_out, bool use_final_probs = true) const;
-
-  /// *** The next functions are from the "new interface". ***
-
-  /// FinalRelativeCost() serves the same function as ReachedFinal(), but gives
-  /// more information.  It returns the difference between the best (final-cost
-  /// plus cost) of any token on the final frame, and the best cost of any token
-  /// on the final frame.  If it is infinity it means no final-states were
-  /// present on the final frame.  It will usually be nonnegative.
-  float FinalRelativeCost() const;
-
-  /// InitDecoding initializes the decoding, and should only be used if you
-  /// intend to call AdvanceDecoding().  If you call Decode(), you don't need
-  /// to call this.  You can call InitDecoding if you have already decoded an
-  /// utterance and want to start with a new utterance.
-  void InitDecoding();
-
-  /// This will decode until there are no more frames ready in the decodable
-  /// object, but if max_num_frames is >= 0 it will decode no more than
-  /// that many frames.  If it returns false, then no tokens are alive,
-  /// which is a kind of error state.
-  void AdvanceDecoding(DecodableInterface *decodable,
-                       int32_t max_num_frames = -1);
-
-  /// Returns the number of frames already decoded.
-  int32_t NumFramesDecoded() const { return num_frames_decoded_; }
-
- private:
-  class Token {
-   public:
-    fst::LatticeArc arc_;  // We use LatticeArc so that we can separately
-                           // store the acoustic and graph cost, in case
-                           // we need to produce lattice-formatted output.
-    Token *prev_;
-    int32_t ref_count_;
-    double cost_;  // accumulated total cost up to this point.
-    Token(const StdArc &arc, float acoustic_cost, Token *prev)
-        : prev_(prev), ref_count_(1) {
-      arc_.ilabel = arc.ilabel;
-      arc_.olabel = arc.olabel;
-      arc_.weight = fst::LatticeWeight(arc.weight.Value(), acoustic_cost);
-      arc_.nextstate = arc.nextstate;
-      if (prev) {
-        prev->ref_count_++;
-        cost_ = prev->cost_ + (arc.weight.Value() + acoustic_cost);
-      } else {
-        cost_ = arc.weight.Value() + acoustic_cost;
-      }
-    }
-    bool operator<(const Token &other) { return cost_ > other.cost_; }
-
-    static void TokenDelete(Token *tok) {
-      while (--tok->ref_count_ == 0) {
-        Token *prev = tok->prev_;
-        delete tok;
-        if (prev == nullptr) {
-          return;
-        } else {
-          tok = prev;
-        }
-      }
-      KALDI_DECODER_ASSERT(tok->ref_count_ > 0);
-    }
-  };
-
-  // ProcessEmitting decodes the frame num_frames_decoded_ of the
-  // decodable object, then increments num_frames_decoded_.
-  void ProcessEmitting(DecodableInterface *decodable);
-
-  void ProcessNonemitting();
-
-  std::unordered_map<StateId, Token *> cur_toks_;
-  std::unordered_map<StateId, Token *> prev_toks_;
-  const fst::Fst<fst::StdArc> &fst_;
-  float beam_;
-  // Keep track of the number of frames decoded in the current file.
-  int32_t num_frames_decoded_ = -1;
-
-  static void ClearToks(std::unordered_map<StateId, Token *> &toks);  // NOLINT
-
-  static void PruneToks(float beam, std::unordered_map<StateId, Token *> *toks);
-};
-
-}  // namespace kaldi_decoder
-
-#endif  // KALDI_DECODER_CSRC_SIMPLE_DECODER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/stl-utils.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/stl-utils.h
deleted file mode 100644
index b06ddf8b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldi-decoder/csrc/stl-utils.h
+++ /dev/null
@@ -1,209 +0,0 @@
-// kaldi-decoder/csrc/stl-utils.h
-//
-// Copyright (c)  2022  Xiaomi Corporation
-
-// this file is copied and modified from
-// kaldi/src/util/stl-utils.h
-
-#ifndef KALDI_DECODER_CSRC_STL_UTILS_H_
-#define KALDI_DECODER_CSRC_STL_UTILS_H_
-#include <algorithm>
-#include <istream>
-#include <type_traits>
-#include <utility>
-#include <vector>
-
-#include "kaldi-decoder/csrc/log.h"
-
-namespace kaldi_decoder {
-
-/// Returns true if the vector is sorted.
-template <typename T>
-inline bool IsSorted(const std::vector<T> &vec) {
-  typename std::vector<T>::const_iterator iter = vec.begin(), end = vec.end();
-  if (iter == end) return true;
-  while (1) {
-    typename std::vector<T>::const_iterator next_iter = iter;
-    ++next_iter;
-    if (next_iter == end) return true;  // end of loop and nothing out of order
-    if (*next_iter < *iter) return false;
-    iter = next_iter;
-  }
-}
-
-/// Sorts and uniq's (removes duplicates) from a vector.
-template <typename T>
-inline void SortAndUniq(std::vector<T> *vec) {
-  std::sort(vec->begin(), vec->end());
-  vec->erase(std::unique(vec->begin(), vec->end()), vec->end());
-}
-
-/// Returns true if the vector is sorted and contains each element
-/// only once.
-template <typename T>
-inline bool IsSortedAndUniq(const std::vector<T> &vec) {
-  typename std::vector<T>::const_iterator iter = vec.begin(), end = vec.end();
-  if (iter == end) return true;
-  while (1) {
-    typename std::vector<T>::const_iterator next_iter = iter;
-    ++next_iter;
-    if (next_iter == end) return true;  // end of loop and nothing out of order
-    if (*next_iter <= *iter) return false;
-    iter = next_iter;
-  }
-}
-
-template <class T>
-inline void WriteIntegerVector(std::ostream &os, bool binary,
-                               const std::vector<T> &v) {
-  // Compile time assertion that this is not called with a wrong type.
-  static_assert(std::is_integral<T>::value, "");
-  if (binary) {
-    char sz = sizeof(T);  // this is currently just a check.
-    os.write(&sz, 1);
-    int32_t vecsz = static_cast<int32_t>(v.size());
-    KALDI_DECODER_ASSERT((size_t)vecsz == v.size());
-
-    os.write(reinterpret_cast<const char *>(&vecsz), sizeof(vecsz));
-    if (vecsz != 0) {
-      os.write(reinterpret_cast<const char *>(&(v[0])), sizeof(T) * vecsz);
-    }
-  } else {
-    // focus here is on prettiness of text form rather than
-    // efficiency of reading-in.
-    // reading-in is dominated by low-level operations anyway:
-    // for efficiency use binary.
-    os << "[ ";
-    typename std::vector<T>::const_iterator iter = v.begin(), end = v.end();
-    for (; iter != end; ++iter) {
-      if (sizeof(T) == 1)
-        os << static_cast<int16_t>(*iter) << " ";
-      else
-        os << *iter << " ";
-    }
-    os << "]\n";
-  }
-  if (os.fail()) {
-    KALDI_DECODER_ERR << "Write failure in WriteIntegerVector.";
-  }
-}
-
-template <class T>
-inline void ReadIntegerVector(std::istream &is, bool binary,
-                              std::vector<T> *v) {
-  static_assert(std::is_integral<T>::value, "");
-  KALDI_DECODER_ASSERT(v != nullptr);
-  if (binary) {
-    int sz = is.peek();
-    if (sz == sizeof(T)) {
-      is.get();
-    } else {  // this is currently just a check.
-      KALDI_DECODER_ERR << "ReadIntegerVector: expected to see type of size "
-                        << sizeof(T) << ", saw instead " << sz
-                        << ", at file position " << is.tellg();
-    }
-    int32_t vecsz;
-    is.read(reinterpret_cast<char *>(&vecsz), sizeof(vecsz));
-    if (is.fail() || vecsz < 0) goto bad;
-
-    v->resize(vecsz);
-
-    if (vecsz > 0) {
-      is.read(reinterpret_cast<char *>(&((*v)[0])), sizeof(T) * vecsz);
-    }
-  } else {
-    std::vector<T> tmp_v;  // use temporary so v doesn't use extra memory
-                           // due to resizing.
-    is >> std::ws;
-    if (is.peek() != static_cast<int>('[')) {
-      KALDI_DECODER_ERR << "ReadIntegerVector: expected to see [, saw "
-                        << is.peek() << ", at file position " << is.tellg();
-    }
-    is.get();       // consume the '['.
-    is >> std::ws;  // consume whitespace.
-    while (is.peek() != static_cast<int>(']')) {
-      if (sizeof(T) == 1) {  // read/write chars as numbers.
-        int16_t next_t;
-        is >> next_t >> std::ws;
-        if (is.fail())
-          goto bad;
-        else
-          tmp_v.push_back((T)next_t);
-      } else {
-        T next_t;
-        is >> next_t >> std::ws;
-        if (is.fail())
-          goto bad;
-        else
-          tmp_v.push_back(next_t);
-      }
-    }
-    is.get();    // get the final ']'.
-    *v = tmp_v;  // could use std::swap to use less temporary memory, but this
-    // uses less permanent memory.
-  }
-  if (!is.fail()) return;
-bad:
-  KALDI_DECODER_ERR << "ReadIntegerVector: read failure at file position "
-                    << is.tellg();
-}
-
-/// Deletes any non-NULL pointers in the vector v, and sets
-/// the corresponding entries of v to NULL
-template <class A>
-void DeletePointers(std::vector<A *> *v) {
-  KALDI_DECODER_ASSERT(v != nullptr);
-  typename std::vector<A *>::iterator iter = v->begin(), end = v->end();
-  for (; iter != end; ++iter) {
-    if (*iter != nullptr) {
-      delete *iter;
-      *iter = nullptr;  // set to NULL for extra safety.
-    }
-  }
-}
-
-/// A hashing function-object for pairs of ints
-template <typename Int1, typename Int2 = Int1>
-struct PairHasher {  // hashing function for pair<int>
-  size_t operator()(const std::pair<Int1, Int2> &x) const noexcept {
-    // 7853 was chosen at random from a list of primes.
-    return x.first + x.second * 7853;
-  }
-  PairHasher() {  // Check we're instantiated with an integer type.
-    static_assert(std::is_integral<Int1>::value, "");
-    static_assert(std::is_integral<Int2>::value, "");
-  }
-};
-
-/// Returns true if the vector of pointers contains NULL pointers.
-template <class A>
-bool ContainsNullPointers(const std::vector<A *> &v) {
-  typename std::vector<A *>::const_iterator iter = v.begin(), end = v.end();
-  for (; iter != end; ++iter)
-    if (*iter == static_cast<A *>(nullptr)) return true;
-  return false;
-}
-
-/// A hashing function-object for vectors.
-template <typename Int>
-struct VectorHasher {  // hashing function for vector<Int>.
-  size_t operator()(const std::vector<Int> &x) const noexcept {
-    size_t ans = 0;
-    auto iter = x.begin(), end = x.end();
-    for (; iter != end; ++iter) {
-      ans *= kPrime;
-      ans += *iter;
-    }
-    return ans;
-  }
-  VectorHasher() {  // Check we're instantiated with an integer type.
-    static_assert(std::is_integral<Int>::value, "");
-  }
-
- private:
-  static const int kPrime = 7853;
-};
-
-}  // namespace kaldi_decoder
-
-#endif  // KALDI_DECODER_CSRC_STL_UTILS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/basic-filebuf.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/basic-filebuf.h
deleted file mode 100644
index 949482c0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/basic-filebuf.h
+++ /dev/null
@@ -1,999 +0,0 @@
-// Copyright (c) 2009-2014 by the contributors listed in CREDITS.TXT (included
-// below)
-// This file is copied from kaldi/src/util/basic-filebuf.h
-///////////////////////////////////////////////////////////////////////////////
-// This is a modified version of the std::basic_filebuf from libc++
-// (http://libcxx.llvm.org/).
-// It allows one to create basic_filebuf from an existing FILE* handle or file
-// descriptor.
-//
-// This file is dual licensed under the MIT and the University of Illinois Open
-// Source License licenses. See LICENSE.TXT for details (included at the
-// bottom).
-///////////////////////////////////////////////////////////////////////////////
-#ifndef KALDIFST_CSRC_BASIC_FILEBUF_H_
-#define KALDIFST_CSRC_BASIC_FILEBUF_H_
-
-///////////////////////////////////////////////////////////////////////////////
-#include <algorithm>
-#include <cstdio>
-#include <cstring>
-#include <fstream>
-#include <memory>
-#include <string>
-#include <utility>
-
-///////////////////////////////////////////////////////////////////////////////
-namespace kaldifst {
-///////////////////////////////////////////////////////////////////////////////
-template <typename CharT, typename Traits = std::char_traits<CharT> >
-class basic_filebuf : public std::basic_streambuf<CharT, Traits> {
- public:
-    typedef CharT                            char_type;
-    typedef Traits                           traits_type;
-    typedef typename traits_type::int_type   int_type;
-    typedef typename traits_type::pos_type   pos_type;
-    typedef typename traits_type::off_type   off_type;
-    typedef typename traits_type::state_type state_type;
-
-    basic_filebuf();
-    basic_filebuf(basic_filebuf&& rhs);
-    virtual ~basic_filebuf();
-
-    basic_filebuf& operator=(basic_filebuf&& rhs);
-    void swap(basic_filebuf& rhs);
-
-    bool is_open() const;
-    basic_filebuf* open(const char* s, std::ios_base::openmode mode);
-    basic_filebuf* open(const std::string& s, std::ios_base::openmode mode);
-    basic_filebuf* open(int fd, std::ios_base::openmode mode);
-    basic_filebuf* open(FILE* f, std::ios_base::openmode mode);
-    basic_filebuf* close();
-
-    FILE* file() { return this->_M_file; }
-    int fd() { return fileno(this->_M_file); }
-
- protected:
-    int_type underflow() override;
-    int_type pbackfail(int_type c = traits_type::eof()) override;
-    int_type overflow(int_type c = traits_type::eof()) override;
-    std::basic_streambuf<char_type, traits_type>*
-      setbuf(char_type* s, std::streamsize n) override;
-    pos_type seekoff(off_type off, std::ios_base::seekdir way,
-                     std::ios_base::openmode wch =
-                     std::ios_base::in | std::ios_base::out) override;
-    pos_type seekpos(pos_type sp,
-                     std::ios_base::openmode wch =
-                     std::ios_base::in | std::ios_base::out) override;
-    int sync() override;
-    void imbue(const std::locale& loc) override;
-
- protected:
-    char*       _M_extbuf;
-    const char* _M_extbufnext;
-    const char* _M_extbufend;
-    char _M_extbuf_min[8];
-    size_t _M_ebs;
-    char_type* _M_intbuf;
-    size_t _M_ibs;
-    FILE* _M_file;
-    const std::codecvt<char_type, char, state_type>* _M_cv;
-    state_type _M_st;
-    state_type _M_st_last;
-    std::ios_base::openmode _M_om;
-    std::ios_base::openmode _M_cm;
-    bool _M_owns_eb;
-    bool _M_owns_ib;
-    bool _M_always_noconv;
-
-    const char* _M_get_mode(std::ios_base::openmode mode);
-    bool _M_read_mode();
-    void _M_write_mode();
-};
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-basic_filebuf<CharT, Traits>::basic_filebuf()
-    : _M_extbuf(nullptr),
-      _M_extbufnext(nullptr),
-      _M_extbufend(nullptr),
-      _M_ebs(0),
-      _M_intbuf(nullptr),
-      _M_ibs(0),
-      _M_file(nullptr),
-      _M_cv(nullptr),
-      _M_st(),
-      _M_st_last(),
-      _M_om(std::ios_base::openmode(0)),
-      _M_cm(std::ios_base::openmode(0)),
-      _M_owns_eb(false),
-      _M_owns_ib(false),
-      _M_always_noconv(false) {
-    if (std::has_facet<std::codecvt<char_type, char, state_type> >
-        (this->getloc())) {
-        _M_cv = &std::use_facet<std::codecvt<char_type, char, state_type> >
-          (this->getloc());
-        _M_always_noconv = _M_cv->always_noconv();
-    }
-    setbuf(0, 4096);
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-basic_filebuf<CharT, Traits>::basic_filebuf(basic_filebuf&& rhs)
-    : std::basic_streambuf<CharT, Traits>(rhs) {
-    if (rhs._M_extbuf == rhs._M_extbuf_min) {
-        _M_extbuf = _M_extbuf_min;
-        _M_extbufnext = _M_extbuf + (rhs._M_extbufnext - rhs._M_extbuf);
-        _M_extbufend = _M_extbuf + (rhs._M_extbufend - rhs._M_extbuf);
-    } else {
-        _M_extbuf = rhs._M_extbuf;
-        _M_extbufnext = rhs._M_extbufnext;
-        _M_extbufend = rhs._M_extbufend;
-    }
-    _M_ebs = rhs._M_ebs;
-    _M_intbuf = rhs._M_intbuf;
-    _M_ibs = rhs._M_ibs;
-    _M_file = rhs._M_file;
-    _M_cv = rhs._M_cv;
-    _M_st = rhs._M_st;
-    _M_st_last = rhs._M_st_last;
-    _M_om = rhs._M_om;
-    _M_cm = rhs._M_cm;
-    _M_owns_eb = rhs._M_owns_eb;
-    _M_owns_ib = rhs._M_owns_ib;
-    _M_always_noconv = rhs._M_always_noconv;
-    if (rhs.pbase()) {
-        if (rhs.pbase() == rhs._M_intbuf)
-            this->setp(_M_intbuf, _M_intbuf + (rhs. epptr() - rhs.pbase()));
-        else
-            this->setp(reinterpret_cast<char_type*>(_M_extbuf),
-                       reinterpret_cast<char_type*>(_M_extbuf)
-                       + (rhs. epptr() - rhs.pbase()));
-        this->pbump(rhs. pptr() - rhs.pbase());
-    } else if (rhs.eback()) {
-        if (rhs.eback() == rhs._M_intbuf)
-            this->setg(_M_intbuf, _M_intbuf + (rhs.gptr() - rhs.eback()),
-                                  _M_intbuf + (rhs.egptr() - rhs.eback()));
-        else
-            this->setg(reinterpret_cast<char_type*>(_M_extbuf),
-                       reinterpret_cast<char_type*>(_M_extbuf) +
-                       (rhs.gptr() - rhs.eback()),
-                       reinterpret_cast<char_type*>(_M_extbuf) +
-                       (rhs.egptr() - rhs.eback()));
-    }
-    rhs._M_extbuf = nullptr;
-    rhs._M_extbufnext = nullptr;
-    rhs._M_extbufend = nullptr;
-    rhs._M_ebs = 0;
-    rhs._M_intbuf = nullptr;
-    rhs._M_ibs = 0;
-    rhs._M_file = nullptr;
-    rhs._M_st = state_type();
-    rhs._M_st_last = state_type();
-    rhs._M_om = std::ios_base::openmode(0);
-    rhs._M_cm = std::ios_base::openmode(0);
-    rhs._M_owns_eb = false;
-    rhs._M_owns_ib = false;
-    rhs.setg(0, 0, 0);
-    rhs.setp(0, 0);
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-inline
-basic_filebuf<CharT, Traits>&
-basic_filebuf<CharT, Traits>::operator=(basic_filebuf&& rhs) {
-    close();
-    swap(rhs);
-    return *this;
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-basic_filebuf<CharT, Traits>::~basic_filebuf() {
-    // try
-    // {
-    //     close();
-    // }
-    // catch (...)
-    // {
-    // }
-    if (_M_owns_eb)
-        delete [] _M_extbuf;
-    if (_M_owns_ib)
-        delete [] _M_intbuf;
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-void
-basic_filebuf<CharT, Traits>::swap(basic_filebuf& rhs) {
-    std::basic_streambuf<char_type, traits_type>::swap(rhs);
-    if (_M_extbuf != _M_extbuf_min && rhs._M_extbuf != rhs._M_extbuf_min) {
-        std::swap(_M_extbuf, rhs._M_extbuf);
-        std::swap(_M_extbufnext, rhs._M_extbufnext);
-        std::swap(_M_extbufend, rhs._M_extbufend);
-    } else {
-        ptrdiff_t ln = _M_extbufnext - _M_extbuf;
-        ptrdiff_t le = _M_extbufend - _M_extbuf;
-        ptrdiff_t rn = rhs._M_extbufnext - rhs._M_extbuf;
-        ptrdiff_t re = rhs._M_extbufend - rhs._M_extbuf;
-        if (_M_extbuf == _M_extbuf_min && rhs._M_extbuf != rhs._M_extbuf_min) {
-            _M_extbuf = rhs._M_extbuf;
-            rhs._M_extbuf = rhs._M_extbuf_min;
-        } else if (_M_extbuf != _M_extbuf_min &&
-            rhs._M_extbuf == rhs._M_extbuf_min) {
-            rhs._M_extbuf = _M_extbuf;
-            _M_extbuf = _M_extbuf_min;
-        }
-        _M_extbufnext = _M_extbuf + rn;
-        _M_extbufend = _M_extbuf + re;
-        rhs._M_extbufnext = rhs._M_extbuf + ln;
-        rhs._M_extbufend = rhs._M_extbuf + le;
-    }
-    std::swap(_M_ebs, rhs._M_ebs);
-    std::swap(_M_intbuf, rhs._M_intbuf);
-    std::swap(_M_ibs, rhs._M_ibs);
-    std::swap(_M_file, rhs._M_file);
-    std::swap(_M_cv, rhs._M_cv);
-    std::swap(_M_st, rhs._M_st);
-    std::swap(_M_st_last, rhs._M_st_last);
-    std::swap(_M_om, rhs._M_om);
-    std::swap(_M_cm, rhs._M_cm);
-    std::swap(_M_owns_eb, rhs._M_owns_eb);
-    std::swap(_M_owns_ib, rhs._M_owns_ib);
-    std::swap(_M_always_noconv, rhs._M_always_noconv);
-    if (this->eback() == reinterpret_cast<char_type*>(rhs._M_extbuf_min)) {
-        ptrdiff_t n = this->gptr() - this->eback();
-        ptrdiff_t e = this->egptr() - this->eback();
-        this->setg(reinterpret_cast<char_type*>(_M_extbuf_min),
-                   reinterpret_cast<char_type*>(_M_extbuf_min) + n,
-                   reinterpret_cast<char_type*>(_M_extbuf_min) + e);
-    } else if (this->pbase() ==
-               reinterpret_cast<char_type*>(rhs._M_extbuf_min)) {
-        ptrdiff_t n = this->pptr() - this->pbase();
-        ptrdiff_t e = this->epptr() - this->pbase();
-        this->setp(reinterpret_cast<char_type*>(_M_extbuf_min),
-                   reinterpret_cast<char_type*>(_M_extbuf_min) + e);
-        this->pbump(n);
-    }
-    if (rhs.eback() == reinterpret_cast<char_type*>(_M_extbuf_min)) {
-        ptrdiff_t n = rhs.gptr() - rhs.eback();
-        ptrdiff_t e = rhs.egptr() - rhs.eback();
-        rhs.setg(reinterpret_cast<char_type*>(rhs._M_extbuf_min),
-                 reinterpret_cast<char_type*>(rhs._M_extbuf_min) + n,
-                 reinterpret_cast<char_type*>(rhs._M_extbuf_min) + e);
-    } else if (rhs.pbase() == reinterpret_cast<char_type*>(_M_extbuf_min)) {
-        ptrdiff_t n = rhs.pptr() - rhs.pbase();
-        ptrdiff_t e = rhs.epptr() - rhs.pbase();
-        rhs.setp(reinterpret_cast<char_type*>(rhs._M_extbuf_min),
-                 reinterpret_cast<char_type*>(rhs._M_extbuf_min) + e);
-        rhs.pbump(n);
-    }
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-inline
-void
-swap(basic_filebuf<CharT, Traits>& x, basic_filebuf<CharT, Traits>& y) {
-    x.swap(y);
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-inline
-bool
-basic_filebuf<CharT, Traits>::is_open() const {
-    return _M_file != nullptr;
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-const char* basic_filebuf<CharT, Traits>::
-_M_get_mode(std::ios_base::openmode mode) {
-    switch ((mode & ~std::ios_base::ate) | 0) {
-    case std::ios_base::out:
-    case std::ios_base::out | std::ios_base::trunc:
-        return "w";
-    case std::ios_base::out | std::ios_base::app:
-    case std::ios_base::app:
-        return "a";
-        break;
-    case std::ios_base::in:
-        return "r";
-    case std::ios_base::in  | std::ios_base::out:
-        return "r+";
-    case std::ios_base::in  | std::ios_base::out | std::ios_base::trunc:
-        return "w+";
-    case std::ios_base::in  | std::ios_base::out | std::ios_base::app:
-    case std::ios_base::in  | std::ios_base::app:
-        return "a+";
-    case std::ios_base::out | std::ios_base::binary:
-    case std::ios_base::out | std::ios_base::trunc | std::ios_base::binary:
-        return "wb";
-    case std::ios_base::out | std::ios_base::app | std::ios_base::binary:
-    case std::ios_base::app | std::ios_base::binary:
-        return "ab";
-    case std::ios_base::in  | std::ios_base::binary:
-        return "rb";
-    case std::ios_base::in  | std::ios_base::out | std::ios_base::binary:
-        return "r+b";
-    case std::ios_base::in  | std::ios_base::out | std::ios_base::trunc |
-      std::ios_base::binary:
-        return "w+b";
-    case std::ios_base::in  | std::ios_base::out | std::ios_base::app |
-      std::ios_base::binary:
-    case std::ios_base::in  | std::ios_base::app | std::ios_base::binary:
-        return "a+b";
-    default:
-        return nullptr;
-    }
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-basic_filebuf<CharT, Traits>*
-basic_filebuf<CharT, Traits>::
-open(const char* s, std::ios_base::openmode mode) {
-    basic_filebuf<CharT, Traits>* rt = nullptr;
-    if (_M_file == nullptr) {
-        const char* md = _M_get_mode(mode);
-        if (md) {
-            _M_file = fopen(s, md);
-            if (_M_file) {
-                rt = this;
-                _M_om = mode;
-                if (mode & std::ios_base::ate) {
-                    if (fseek(_M_file, 0, SEEK_END)) {
-                        fclose(_M_file);
-                        _M_file = nullptr;
-                        rt = nullptr;
-                    }
-                }
-            }
-        }
-    }
-    return rt;
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-inline
-basic_filebuf<CharT, Traits>*
-basic_filebuf<CharT, Traits>::open(const std::string& s,
-    std::ios_base::openmode mode) {
-    return open(s.c_str(), mode);
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-basic_filebuf<CharT, Traits>*
-basic_filebuf<CharT, Traits>::open(int fd, std::ios_base::openmode mode) {
-    const char* md = this->_M_get_mode(mode);
-    if (md) {
-        this->_M_file = fdopen(fd, md);
-        this->_M_om = mode;
-        return this;
-    } else {
-      return nullptr;
-    }
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-basic_filebuf<CharT, Traits>*
-basic_filebuf<CharT, Traits>::open(FILE* f, std::ios_base::openmode mode) {
-    this->_M_file = f;
-    this->_M_om = mode;
-    return this;
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-basic_filebuf<CharT, Traits>*
-basic_filebuf<CharT, Traits>::close() {
-    basic_filebuf<CharT, Traits>* rt = nullptr;
-    if (_M_file) {
-        rt = this;
-        std::unique_ptr<FILE, int(*)(FILE*)> h(_M_file, fclose);
-        if (sync())
-            rt = nullptr;
-        if (fclose(h.release()) == 0)
-            _M_file = nullptr;
-        else
-            rt = nullptr;
-    }
-    return rt;
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-typename basic_filebuf<CharT, Traits>::int_type
-basic_filebuf<CharT, Traits>::underflow() {
-    if (_M_file == nullptr)
-        return traits_type::eof();
-    bool initial = _M_read_mode();
-    char_type buf;
-    if (this->gptr() == nullptr)
-        this->setg(&buf, &buf+1, &buf+1);
-    const size_t unget_sz = initial ? 0 : std::
-      min<size_t>((this->egptr() - this->eback()) / 2, 4);
-    int_type c = traits_type::eof();
-    if (this->gptr() == this->egptr()) {
-        memmove(this->eback(), this->egptr() - unget_sz,
-            unget_sz * sizeof(char_type));
-        if (_M_always_noconv) {
-            size_t nmemb = static_cast<size_t>
-              (this->egptr() - this->eback() - unget_sz);
-            nmemb = fread(this->eback() + unget_sz, 1, nmemb, _M_file);
-            if (nmemb != 0) {
-                this->setg(this->eback(),
-                           this->eback() + unget_sz,
-                           this->eback() + unget_sz + nmemb);
-                c = traits_type::to_int_type(*this->gptr());
-            }
-        } else {
-            memmove(_M_extbuf, _M_extbufnext, _M_extbufend - _M_extbufnext);
-            _M_extbufnext = _M_extbuf + (_M_extbufend - _M_extbufnext);
-            _M_extbufend = _M_extbuf +
-              (_M_extbuf == _M_extbuf_min ? sizeof(_M_extbuf_min) : _M_ebs);
-            size_t nmemb = std::min(static_cast<size_t>(_M_ibs - unget_sz),
-                                    static_cast<size_t>
-                                    (_M_extbufend - _M_extbufnext));
-            std::codecvt_base::result r;
-            _M_st_last = _M_st;
-            size_t nr = fread(
-                reinterpret_cast<void*>(const_cast<char_type*>(_M_extbufnext)),
-                1, nmemb, _M_file);
-            if (nr != 0) {
-                if (!_M_cv)
-                    throw std::bad_cast();
-                _M_extbufend = _M_extbufnext + nr;
-                char_type*  inext;
-                r = _M_cv->in(_M_st, _M_extbuf, _M_extbufend, _M_extbufnext,
-                              this->eback() + unget_sz,
-                              this->eback() + _M_ibs, inext);
-                if (r == std::codecvt_base::noconv) {
-                    this->setg(reinterpret_cast<char_type*>(_M_extbuf),
-                               reinterpret_cast<char_type*>(_M_extbuf),
-                               const_cast<char_type*>(_M_extbufend));
-                    c = traits_type::to_int_type(*this->gptr());
-                } else if (inext != this->eback() + unget_sz) {
-                    this->setg(this->eback(), this->eback() + unget_sz, inext);
-                    c = traits_type::to_int_type(*this->gptr());
-                }
-            }
-        }
-    } else {
-        c = traits_type::to_int_type(*this->gptr());
-    }
-    if (this->eback() == &buf)
-        this->setg(0, 0, 0);
-    return c;
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-typename basic_filebuf<CharT, Traits>::int_type
-basic_filebuf<CharT, Traits>::pbackfail(int_type c) {
-    if (_M_file && this->eback() < this->gptr()) {
-        if (traits_type::eq_int_type(c, traits_type::eof())) {
-            this->gbump(-1);
-            return traits_type::not_eof(c);
-        }
-        if ((_M_om & std::ios_base::out) ||
-            traits_type::eq(traits_type::to_char_type(c), this->gptr()[-1])) {
-            this->gbump(-1);
-            *this->gptr() = traits_type::to_char_type(c);
-            return c;
-        }
-    }
-    return traits_type::eof();
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-typename basic_filebuf<CharT, Traits>::int_type
-basic_filebuf<CharT, Traits>::overflow(int_type c) {
-    if (_M_file == nullptr)
-        return traits_type::eof();
-    _M_write_mode();
-    char_type buf;
-    char_type* pb_save = this->pbase();
-    char_type* epb_save = this->epptr();
-    if (!traits_type::eq_int_type(c, traits_type::eof())) {
-        if (this->pptr() == nullptr)
-            this->setp(&buf, &buf+1);
-        *this->pptr() = traits_type::to_char_type(c);
-        this->pbump(1);
-    }
-    if (this->pptr() != this->pbase()) {
-        if (_M_always_noconv) {
-            size_t nmemb = static_cast<size_t>(this->pptr() - this->pbase());
-            if (fwrite(this->pbase(), sizeof(char_type),
-                  nmemb, _M_file) != nmemb)
-                return traits_type::eof();
-        } else {
-            char* extbe = _M_extbuf;
-            std::codecvt_base::result r;
-            do {
-                if (!_M_cv)
-                    throw std::bad_cast();
-                const char_type* e;
-                r = _M_cv->out(_M_st, this->pbase(), this->pptr(), e,
-                               _M_extbuf, _M_extbuf + _M_ebs, extbe);
-                if (e == this->pbase())
-                    return traits_type::eof();
-                if (r == std::codecvt_base::noconv) {
-                    size_t nmemb = static_cast<size_t>
-                      (this->pptr() - this->pbase());
-                    if (fwrite(this->pbase(), 1, nmemb, _M_file) != nmemb)
-                        return traits_type::eof();
-                } else if (r == std::codecvt_base::ok ||
-                    r == std::codecvt_base::partial) {
-                    size_t nmemb = static_cast<size_t>(extbe - _M_extbuf);
-                    if (fwrite(_M_extbuf, 1, nmemb, _M_file) != nmemb)
-                        return traits_type::eof();
-                    if (r == std::codecvt_base::partial) {
-                        this->setp(const_cast<char_type*>(e),
-                            this->pptr());
-                        this->pbump(this->epptr() - this->pbase());
-                    }
-                } else {
-                    return traits_type::eof();
-                }
-            } while (r == std::codecvt_base::partial);
-        }
-        this->setp(pb_save, epb_save);
-    }
-    return traits_type::not_eof(c);
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-std::basic_streambuf<CharT, Traits>*
-basic_filebuf<CharT, Traits>::setbuf(char_type* s, std::streamsize n) {
-    this->setg(0, 0, 0);
-    this->setp(0, 0);
-    if (_M_owns_eb)
-        delete [] _M_extbuf;
-    if (_M_owns_ib)
-        delete [] _M_intbuf;
-    _M_ebs = n;
-    if (_M_ebs > sizeof(_M_extbuf_min)) {
-        if (_M_always_noconv && s) {
-            _M_extbuf = reinterpret_cast<char*>(s);
-            _M_owns_eb = false;
-        } else {
-            _M_extbuf = new char[_M_ebs];
-            _M_owns_eb = true;
-        }
-    } else {
-        _M_extbuf = _M_extbuf_min;
-        _M_ebs = sizeof(_M_extbuf_min);
-        _M_owns_eb = false;
-    }
-    if (!_M_always_noconv) {
-        _M_ibs = std::max<std::streamsize>(n, sizeof(_M_extbuf_min));
-        if (s && _M_ibs >= sizeof(_M_extbuf_min)) {
-            _M_intbuf = s;
-            _M_owns_ib = false;
-        } else {
-            _M_intbuf = new char_type[_M_ibs];
-            _M_owns_ib = true;
-        }
-    } else {
-        _M_ibs = 0;
-        _M_intbuf = 0;
-        _M_owns_ib = false;
-    }
-    return this;
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-typename basic_filebuf<CharT, Traits>::pos_type
-basic_filebuf<CharT, Traits>::seekoff(off_type off, std::ios_base::seekdir way,
-                                      std::ios_base::openmode) {
-    if (!_M_cv)
-        throw std::bad_cast();
-    int width = _M_cv->encoding();
-    if (_M_file == nullptr || (width <= 0 && off != 0) || sync())
-        return pos_type(off_type(-1));
-    // width > 0 || off == 0
-    int whence;
-    switch (way) {
-    case std::ios_base::beg:
-        whence = SEEK_SET;
-        break;
-    case std::ios_base::cur:
-        whence = SEEK_CUR;
-        break;
-    case std::ios_base::end:
-        whence = SEEK_END;
-        break;
-    default:
-        return pos_type(off_type(-1));
-    }
-#if _WIN32
-    if (fseek(_M_file, width > 0 ? width * off : 0, whence))
-        return pos_type(off_type(-1));
-    pos_type r = ftell(_M_file);
-#else
-    if (fseeko(_M_file, width > 0 ? width * off : 0, whence))
-        return pos_type(off_type(-1));
-    pos_type r = ftello(_M_file);
-#endif
-    r.state(_M_st);
-    return r;
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-typename basic_filebuf<CharT, Traits>::pos_type
-basic_filebuf<CharT, Traits>::seekpos(pos_type sp, std::ios_base::openmode) {
-    if (_M_file == nullptr || sync())
-        return pos_type(off_type(-1));
-#if _WIN32
-    if (fseek(_M_file, sp, SEEK_SET))
-        return pos_type(off_type(-1));
-#else
-    if (fseeko(_M_file, sp, SEEK_SET))
-        return pos_type(off_type(-1));
-#endif
-    _M_st = sp.state();
-    return sp;
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-int
-basic_filebuf<CharT, Traits>::sync() {
-    if (_M_file == nullptr)
-        return 0;
-    if (!_M_cv)
-        throw std::bad_cast();
-    if (_M_cm & std::ios_base::out) {
-        if (this->pptr() != this->pbase())
-            if (overflow() == traits_type::eof())
-                return -1;
-        std::codecvt_base::result r;
-        do {
-            char* extbe;
-            r = _M_cv->unshift(_M_st, _M_extbuf, _M_extbuf + _M_ebs, extbe);
-            size_t nmemb = static_cast<size_t>(extbe - _M_extbuf);
-            if (fwrite(_M_extbuf, 1, nmemb, _M_file) != nmemb)
-                return -1;
-        } while (r == std::codecvt_base::partial);
-        if (r == std::codecvt_base::error)
-            return -1;
-        if (fflush(_M_file))
-            return -1;
-    } else if (_M_cm & std::ios_base::in) {
-        off_type c;
-        state_type state = _M_st_last;
-        bool update_st = false;
-        if (_M_always_noconv) {
-            c = this->egptr() - this->gptr();
-        } else {
-            int width = _M_cv->encoding();
-            c = _M_extbufend - _M_extbufnext;
-            if (width > 0) {
-                c += width * (this->egptr() - this->gptr());
-            } else {
-                if (this->gptr() != this->egptr()) {
-                    const int off = _M_cv->length(state, _M_extbuf,
-                                                  _M_extbufnext,
-                                                  this->gptr() - this->eback());
-                    c += _M_extbufnext - _M_extbuf - off;
-                    update_st = true;
-                }
-            }
-        }
-#if _WIN32
-        if (fseek(_M_file_, -c, SEEK_CUR))
-            return -1;
-#else
-        if (fseeko(_M_file, -c, SEEK_CUR))
-            return -1;
-#endif
-        if (update_st)
-            _M_st = state;
-        _M_extbufnext = _M_extbufend = _M_extbuf;
-        this->setg(0, 0, 0);
-        _M_cm = std::ios_base::openmode(0);
-    }
-    return 0;
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-void
-basic_filebuf<CharT, Traits>::imbue(const std::locale& loc) {
-    sync();
-    _M_cv = &std::use_facet<std::codecvt<char_type, char, state_type> >(loc);
-    bool old_anc = _M_always_noconv;
-    _M_always_noconv = _M_cv->always_noconv();
-    if (old_anc != _M_always_noconv) {
-        this->setg(0, 0, 0);
-        this->setp(0, 0);
-        // invariant, char_type is char, else we couldn't get here
-        // need to dump _M_intbuf
-        if (_M_always_noconv) {
-            if (_M_owns_eb)
-                delete [] _M_extbuf;
-            _M_owns_eb = _M_owns_ib;
-            _M_ebs = _M_ibs;
-            _M_extbuf = reinterpret_cast<char*>(_M_intbuf);
-            _M_ibs = 0;
-            _M_intbuf = nullptr;
-            _M_owns_ib = false;
-        } else {  // need to obtain an _M_intbuf.
-         // If _M_extbuf is user-supplied, use it, else new _M_intbuf
-            if (!_M_owns_eb && _M_extbuf != _M_extbuf_min) {
-                _M_ibs = _M_ebs;
-                _M_intbuf = reinterpret_cast<char_type*>(_M_extbuf);
-                _M_owns_ib = false;
-                _M_extbuf = new char[_M_ebs];
-                _M_owns_eb = true;
-            } else {
-                _M_ibs = _M_ebs;
-                _M_intbuf = new char_type[_M_ibs];
-                _M_owns_ib = true;
-            }
-        }
-    }
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-bool
-basic_filebuf<CharT, Traits>::_M_read_mode() {
-    if (!(_M_cm & std::ios_base::in)) {
-        this->setp(0, 0);
-        if (_M_always_noconv)
-            this->setg(reinterpret_cast<char_type*>(_M_extbuf),
-                       reinterpret_cast<char_type*>(_M_extbuf) + _M_ebs,
-                       reinterpret_cast<char_type*>(_M_extbuf) + _M_ebs);
-        else
-            this->setg(_M_intbuf, _M_intbuf + _M_ibs, _M_intbuf + _M_ibs);
-        _M_cm = std::ios_base::in;
-        return true;
-    }
-    return false;
-}
-
-///////////////////////////////////////////////////////////////////////////////
-template <class CharT, class Traits>
-void
-basic_filebuf<CharT, Traits>::_M_write_mode() {
-    if (!(_M_cm & std::ios_base::out)) {
-        this->setg(0, 0, 0);
-        if (_M_ebs > sizeof(_M_extbuf_min)) {
-            if (_M_always_noconv)
-                this->setp(reinterpret_cast<char_type*>(_M_extbuf),
-                           reinterpret_cast<char_type*>(_M_extbuf) +
-                           (_M_ebs - 1));
-            else
-                this->setp(_M_intbuf, _M_intbuf + (_M_ibs - 1));
-        } else {
-            this->setp(0, 0);
-        }
-        _M_cm = std::ios_base::out;
-    }
-}
-
-///////////////////////////////////////////////////////////////////////////////
-}  // namespace kaldifst
-
-///////////////////////////////////////////////////////////////////////////////
-#endif  // KALDIFST_CSRC_BASIC_FILEBUF_H_
-
-///////////////////////////////////////////////////////////////////////////////
-
-/*
- * ============================================================================
- * libc++ License
- * ============================================================================
- *
- * The libc++ library is dual licensed under both the University of Illinois
- * "BSD-Like" license and the MIT license.  As a user of this code you may
- * choose to use it under either license.  As a contributor, you agree to allow
- * your code to be used under both.
- *
- * Full text of the relevant licenses is included below.
- *
- * ============================================================================
- *
- * University of Illinois/NCSA
- * Open Source License
- *
- * Copyright (c) 2009-2014 by the contributors listed in CREDITS.TXT (included below)
- *
- * All rights reserved.
- *
- * Developed by:
- *
- *     LLVM Team
- *
- *     University of Illinois at Urbana-Champaign
- *
- *     http://llvm.org
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy of
- * this software and associated documentation files (the "Software"), to deal with
- * the Software without restriction, including without limitation the rights to
- * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
- * of the Software, and to permit persons to whom the Software is furnished to do
- * so, subject to the following conditions:
- *
- *     * Redistributions of source code must retain the above copyright notice,
- *       this list of conditions and the following disclaimers.
- *
- *     * Redistributions in binary form must reproduce the above copyright notice,
- *       this list of conditions and the following disclaimers in the
- *       documentation and/or other materials provided with the distribution.
- *
- *     * Neither the names of the LLVM Team, University of Illinois at
- *       Urbana-Champaign, nor the names of its contributors may be used to
- *       endorse or promote products derived from this Software without specific
- *       prior written permission.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
- * FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
- * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH THE
- * SOFTWARE.
- *
- * ==============================================================================
- *
- * Copyright (c) 2009-2014 by the contributors listed in CREDITS.TXT (included below)
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in
- * all copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
- * THE SOFTWARE.
- *
- * ==============================================================================
- *
- * This file is a partial list of people who have contributed to the LLVM/libc++
- * project.  If you have contributed a patch or made some other contribution to
- * LLVM/libc++, please submit a patch to this file to add yourself, and it will be
- * done!
- *
- * The list is sorted by surname and formatted to allow easy grepping and
- * beautification by scripts.  The fields are: name (N), email (E), web-address
- * (W), PGP key ID and fingerprint (P), description (D), and snail-mail address
- * (S).
- *
- * N: Saleem Abdulrasool
- * E: compnerd@compnerd.org
- * D: Minor patches and Linux fixes.
- *
- * N: Dimitry Andric
- * E: dimitry@andric.com
- * D: Visibility fixes, minor FreeBSD portability patches.
- *
- * N: Holger Arnold
- * E: holgerar@gmail.com
- * D: Minor fix.
- *
- * N: Ruben Van Boxem
- * E: vanboxem dot ruben at gmail dot com
- * D: Initial Windows patches.
- *
- * N: David Chisnall
- * E: theraven at theravensnest dot org
- * D: FreeBSD and Solaris ports, libcxxrt support, some atomics work.
- *
- * N: Marshall Clow
- * E: mclow.lists@gmail.com
- * E: marshall@idio.com
- * D: C++14 support, patches and bug fixes.
- *
- * N: Bill Fisher
- * E: william.w.fisher@gmail.com
- * D: Regex bug fixes.
- *
- * N: Matthew Dempsky
- * E: matthew@dempsky.org
- * D: Minor patches and bug fixes.
- *
- * N: Google Inc.
- * D: Copyright owner and contributor of the CityHash algorithm
- *
- * N: Howard Hinnant
- * E: hhinnant@apple.com
- * D: Architect and primary author of libc++
- *
- * N: Hyeon-bin Jeong
- * E: tuhertz@gmail.com
- * D: Minor patches and bug fixes.
- *
- * N: Argyrios Kyrtzidis
- * E: kyrtzidis@apple.com
- * D: Bug fixes.
- *
- * N: Bruce Mitchener, Jr.
- * E: bruce.mitchener@gmail.com
- * D: Emscripten-related changes.
- *
- * N: Michel Morin
- * E: mimomorin@gmail.com
- * D: Minor patches to is_convertible.
- *
- * N: Andrew Morrow
- * E: andrew.c.morrow@gmail.com
- * D: Minor patches and Linux fixes.
- *
- * N: Arvid Picciani
- * E: aep at exys dot org
- * D: Minor patches and musl port.
- *
- * N: Bjorn Reese
- * E: breese@users.sourceforge.net
- * D: Initial regex prototype
- *
- * N: Nico Rieck
- * E: nico.rieck@gmail.com
- * D: Windows fixes
- *
- * N: Jonathan Sauer
- * D: Minor patches, mostly related to constexpr
- *
- * N: Craig Silverstein
- * E: csilvers@google.com
- * D: Implemented Cityhash as the string hash function on 64-bit machines
- *
- * N: Richard Smith
- * D: Minor patches.
- *
- * N: Joerg Sonnenberger
- * E: joerg@NetBSD.org
- * D: NetBSD port.
- *
- * N: Stephan Tolksdorf
- * E: st@quanttec.com
- * D: Minor <atomic> fix
- *
- * N: Michael van der Westhuizen
- * E: r1mikey at gmail dot com
- *
- * N: Klaas de Vries
- * E: klaas at klaasgaaf dot nl
- * D: Minor bug fix.
- *
- * N: Zhang Xiongpang
- * E: zhangxiongpang@gmail.com
- * D: Minor patches and bug fixes.
- *
- * N: Xing Xue
- * E: xingxue@ca.ibm.com
- * D: AIX port
- *
- * N: Zhihao Yuan
- * E: lichray@gmail.com
- * D: Standard compatibility fixes.
- *
- * N: Jeffrey Yasskin
- * E: jyasskin@gmail.com
- * E: jyasskin@google.com
- * D: Linux fixes.
- */
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/const-integer-set-inl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/const-integer-set-inl.h
deleted file mode 100644
index 42733958..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/const-integer-set-inl.h
+++ /dev/null
@@ -1,76 +0,0 @@
-// kaldifst/csrc/const-integer-set-inl.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/util/const-integer-set-inl.h
-//
-// Copyright 2009-2011     Microsoft Corporation
-
-#ifndef KALDIFST_CSRC_CONST_INTEGER_SET_INL_H_
-#define KALDIFST_CSRC_CONST_INTEGER_SET_INL_H_
-#include <type_traits>
-
-// Do not include this file directly.  It is included by const-integer-set.h
-
-namespace kaldifst {
-
-template <class I>
-void ConstIntegerSet<I>::InitInternal() {
-  static_assert(std::is_integral<I>::value, "");
-  quick_set_.clear();  // just in case we previously had data.
-  if (slow_set_.size() == 0) {
-    lowest_member_ = (I)1;
-    highest_member_ = (I)0;
-    contiguous_ = false;
-    quick_ = false;
-  } else {
-    lowest_member_ = slow_set_.front();
-    highest_member_ = slow_set_.back();
-    size_t range = highest_member_ + 1 - lowest_member_;
-    if (range == slow_set_.size()) {
-      contiguous_ = true;
-      quick_ = false;
-    } else {
-      contiguous_ = false;
-      // If it would be more compact to store as bool
-      if (range < slow_set_.size() * 8 * sizeof(I)) {
-        // (assuming 1 bit per element)...
-        quick_set_.resize(range, false);
-        for (size_t i = 0; i < slow_set_.size(); i++)
-          quick_set_[slow_set_[i] - lowest_member_] = true;
-        quick_ = true;
-      } else {
-        quick_ = false;
-      }
-    }
-  }
-}
-
-template <class I>
-int ConstIntegerSet<I>::count(I i) const {
-  if (i < lowest_member_ || i > highest_member_) {
-    return 0;
-  } else {
-    if (contiguous_) return true;
-    if (quick_) {
-      return (quick_set_[i - lowest_member_] ? 1 : 0);
-    } else {
-      bool ans = std::binary_search(slow_set_.begin(), slow_set_.end(), i);
-      return (ans ? 1 : 0);
-    }
-  }
-}
-
-template <class I>
-void ConstIntegerSet<I>::Write(std::ostream &os, bool binary) const {
-  WriteIntegerVector(os, binary, slow_set_);
-}
-
-template <class I>
-void ConstIntegerSet<I>::Read(std::istream &is, bool binary) {
-  ReadIntegerVector(is, binary, &slow_set_);
-  InitInternal();
-}
-
-}  // namespace kaldifst
-
-#endif  // KALDIFST_CSRC_CONST_INTEGER_SET_INL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/const-integer-set.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/const-integer-set.h
deleted file mode 100644
index b2533c92..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/const-integer-set.h
+++ /dev/null
@@ -1,85 +0,0 @@
-// kaldifst/csrc/const-integer-set.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/util/const-integer-set.h
-
-// Copyright 2009-2011  Microsoft Corporation;  Saarland University
-
-#ifndef KALDIFST_CSRC_CONST_INTEGER_SET_H_
-#define KALDIFST_CSRC_CONST_INTEGER_SET_H_
-#include <algorithm>
-#include <cassert>
-#include <limits>
-#include <set>
-#include <vector>
-
-#include "kaldifst/csrc/stl-utils.h"
-
-/* ConstIntegerSet is a way to efficiently test whether something is in a
-   supplied set of integers.  It can be initialized from a vector or set, but
-   never changed after that. It either uses a sorted vector or an array of
-   bool, depending on the input.  It behaves like a const version of an STL set,
-   with only a subset of the functionality, except all the member functions are
-   upper-case.
-
-   Note that we could get rid of the member slow_set_, but we'd have to
-   do more work to implement an iterator type.  This would save memory.
-*/
-
-namespace kaldifst {
-
-template <class I>
-class ConstIntegerSet {
- public:
-  ConstIntegerSet() : lowest_member_(1), highest_member_(0) {}
-
-  void Init(const std::vector<I> &input) {
-    slow_set_ = input;
-    SortAndUniq(&slow_set_);
-    InitInternal();
-  }
-
-  void Init(const std::set<I> &input) {
-    CopySetToVector(input, &slow_set_);
-    InitInternal();
-  }
-
-  explicit ConstIntegerSet(const std::vector<I> &input) : slow_set_(input) {
-    SortAndUniq(&slow_set_);
-    InitInternal();
-  }
-  explicit ConstIntegerSet(const std::set<I> &input) {
-    CopySetToVector(input, &slow_set_);
-    InitInternal();
-  }
-  explicit ConstIntegerSet(const ConstIntegerSet<I> &other)
-      : slow_set_(other.slow_set_) {
-    InitInternal();
-  }
-
-  int count(I i) const;  // returns 1 or 0.
-
-  typedef typename std::vector<I>::const_iterator iterator;
-  iterator begin() const { return slow_set_.begin(); }
-  iterator end() const { return slow_set_.end(); }
-  size_t size() const { return slow_set_.size(); }
-  bool empty() const { return slow_set_.empty(); }
-
-  void Write(std::ostream &os, bool binary) const;
-  void Read(std::istream &is, bool binary);
-
- private:
-  I lowest_member_;
-  I highest_member_;
-  bool contiguous_;
-  bool quick_;
-  std::vector<bool> quick_set_;
-  std::vector<I> slow_set_;
-  void InitInternal();
-};
-
-}  // namespace kaldifst
-
-#include "kaldifst/csrc/const-integer-set-inl.h"
-
-#endif  // KALDIFST_CSRC_CONST_INTEGER_SET_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/context-fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/context-fst.h
deleted file mode 100644
index 57cc65d3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/context-fst.h
+++ /dev/null
@@ -1,274 +0,0 @@
-// kaldifst/csrc/context-fst.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/context-fst.h
-
-// Copyright 2009-2011  Microsoft Corporation
-//                2018  Johns Hopkins University (author: Daniel Povey)
-#ifndef KALDIFST_CSRC_CONTEXT_FST_H_
-#define KALDIFST_CSRC_CONTEXT_FST_H_
-
-#include <unordered_map>
-#include <vector>
-
-#include "fst/fst-decl.h"
-#include "fst/fstlib.h"
-#include "kaldifst/csrc/const-integer-set.h"
-#include "kaldifst/csrc/deterministic-fst.h"
-
-namespace fst {
-
-/**
-  Modifies an FST so that it transduces the same paths, but the input side of
-  the paths can all have the subsequential symbol '$' appended to them any
-  number of times (we could easily specify the number of times, but accepting
-  any number of repetitions is just more convenient).  The actual way we do this
-  is for each final state, we add a transition with weight equal to the
-  final-weight of that state, with input-symbol '$' and output-symbols \<eps\>,
-  and ending in a new super-final state that has unit final-probability and a
-  unit-weight self-loop with '$' on its input and \<eps\> on its output.  The
-  reason we don't just add a loop to each final-state has to do with preserving
-  stochasticity (see \ref fst_algo_stochastic).  We keep the final-probability
-  in all the original final-states rather than setting them to zero, so the
-  resulting FST can accept zero '$' symbols at the end (in case we had no right
-  context).
-*/
-void AddSubsequentialLoop(StdArc::Label subseq_symbol, MutableFst<StdArc> *fst);
-
-/*
-   InverseContextFst represents the inverse of the context FST "C" (the "C" in
-   "HCLG") which transduces from symbols representing phone context windows
-   (e.g. "a, b, c") to individual phones, e.g. "a".  So InverseContextFst
-   transduces from phones to symbols representing phone context windows.  The
-   point is that the inverse is deterministic, so the DeterministicOnDemandFst
-   interface is applicable, which turns out to be a convenient way to implement
-   this.
-
-   This doesn't implement the full Fst interface, it implements the
-   DeterministicOnDemandFst interface which is much simpler and which is
-   sufficient for what we need to do with this.
-
-   Search for "hbka.pdf" ("Speech Recognition with Weighted Finite State
-   Transducers") by M. Mohri, for more context.
-*/
-
-class InverseContextFst : public DeterministicOnDemandFst<StdArc> {
- public:
-  typedef StdArc Arc;
-  typedef typename StdArc::StateId StateId;
-  typedef typename StdArc::Weight Weight;
-  typedef typename StdArc::Label Label;
-
-  /**
-     Constructor.
-        @param [in] subsequential_symbol   The integer id of the 'subsequential
-     symbol' (usually represented as '$') that terminates sequences on the
-                          output of C.fst (input of InverseContextFst).  Search
-     for "quential" in https://cs.nyu.edu/~mohri/pub/hbka.pdf. This may just be
-     the first unused integer id.  Must be nonzero.
-        @param [in] phones      List of integer ids of phones, as you would see
-     in phones.txt
-        @param [in] disambig_syms   List of integer ids of disambiguation
-     symbols, e.g. the ids of #0, #1, #2 in phones.txt
-        @param [in] context_width  Size of context window, e.g. 3 for triphone.
-        @param [in] central_position  Central position in context window
-     (zero-based), e.g. 1 for triphone. See \ref graph_context for more details.
-  */
-  InverseContextFst(Label subsequential_symbol,
-                    const std::vector<int32_t> &phones,
-                    const std::vector<int32_t> &disambig_syms,
-                    int32_t context_width, int32_t central_position);
-
-  StateId Start() override { return 0; }
-
-  Weight Final(StateId s) override;
-
-  /// Note: ilabel must not be epsilon.
-  bool GetArc(StateId s, Label ilabel, Arc *arc) override;
-
-  ~InverseContextFst() override = default;
-
-  // Returns a reference to a vector<vector<int32_t> > with information about
-  // all the input symbols of C (i.e. all the output symbols of this
-  // InverseContextFst).  See
-  // "http://kaldi-asr.org/doc/tree_externals.html#tree_ilabel".
-  const std::vector<std::vector<int32_t>> &IlabelInfo() const {
-    return ilabel_info_;
-  }
-
-  // A way to destructively obtain the ilabel-info.  Only do this if you
-  // are just about to destroy this object.
-  void SwapIlabelInfo(std::vector<std::vector<int32_t>> *vec) {
-    ilabel_info_.swap(*vec);
-  }
-
- private:
-  /// Returns the state-id corresponding to this vector of phones; creates the
-  /// state it if necessary.  Requires seq.size() == context_width_ - 1.
-  StateId FindState(const std::vector<int32_t> &seq);
-
-  /// Finds the label index corresponding to this context-window of phones
-  /// (likely of width context_width_).  Inserts it into the
-  /// ilabel_info_/ilabel_map_ tables if necessary.
-  Label FindLabel(const std::vector<int32_t> &label_info);
-
-  inline bool IsDisambigSymbol(Label lab) const {
-    return (disambig_syms_.count(lab) != 0);
-  }
-
-  inline bool IsPhoneSymbol(Label lab) const {
-    return (phone_syms_.count(lab) != 0);
-  }
-
-  /// Create disambiguation-symbol self-loop arc; where 'ilabel' must correspond
-  /// to a disambiguation symbol.  Called from CreateArc().
-  inline void CreateDisambigArc(StateId s, Label ilabel, Arc *arc);
-
-  /// Creates an arc, this function is to be called only when 'ilabel'
-  /// corresponds to a phone.  Called from CreateArc().  The olabel may end be
-  /// epsilon, instead of a phone-in-context, if the system has right context
-  /// and we are very near the beginning of the phone sequence.
-  inline void CreatePhoneOrEpsArc(StateId src, StateId dst, Label ilabel,
-                                  const std::vector<int32_t> &phone_seq,
-                                  Arc *arc);
-
-  /// If phone_seq is nonempty then this function it left by one and appends
-  /// 'label' to it, otherwise it does nothing.  We expect (but do not check)
-  /// that phone_seq->size() == context_width_ - 1.
-  inline void ShiftSequenceLeft(Label label, std::vector<int32_t> *phone_seq);
-
-  /// This utility function does something equivalent to the following 3 steps:
-  ///   *full_phone_sequence =  seq;
-  ///  full_phone_sequence->append(label)
-  ///  Replace any values equal to 'subsequential_symbol_' in
-  /// full_phone_sequence with zero (this is to avoid having to keep track of
-  /// the value of 'subsequential_symbol_' outside of this program).
-  /// This function assumes that seq.size() == context_width_ - 1, and also that
-  /// 'subsequential_symbol_' does not appear in positions 0 through
-  /// central_position_ of 'seq'.
-  inline void GetFullPhoneSequence(const std::vector<int32_t> &seq, Label label,
-                                   std::vector<int32_t> *full_phone_sequence);
-
-  // Map type to map from vectors of int32_t (representing phonetic contexts,
-  // which will be of dimension context_width - 1) to StateId (corresponding to
-  // the state index in this FST).
-  typedef unordered_map<std::vector<int32_t>, StateId,
-                        kaldifst::VectorHasher<int32_t>>
-      VectorToStateMap;
-
-  // Map type to map from vectors of int32_t (representing ilabel-info,
-  // see http://kaldi-asr.org/doc/tree_externals.html#tree_ilabel) to
-  // Label (the output label in this FST).
-  typedef unordered_map<std::vector<int32_t>, Label,
-                        kaldifst::VectorHasher<int32_t>>
-      VectorToLabelMap;
-
-  // Sometimes called N, context_width_ this is the width of the
-  // phonetic context, e.g. 3 for triphone, 2 for biphone, one for monophone.
-  // It is a user-specified value.
-  int32_t context_width_;
-
-  // Sometimes called P, central_position_ is is the (zero-based) "central
-  // position" in the context window, meaning the phone that is "in" a certain
-  // context.  The most widely used values of (context-width, central-position)
-  // are: (3,1) for triphone, (1,0) for monophone, and (2, 1) for left biphone.
-  // This is also specified by the user.  As an example, in the left-biphone
-  // [ 5, 6 ], we view it as "the phone numbered 6 with the phone numbered 5 as
-  // its left-context".
-  int32_t central_position_;
-
-  // The following three variables were also passed in by the caller:
-
-  // 'phone_syms_' are a set of phone-ids, typically 1, 2, .. num_phones.
-  kaldifst::ConstIntegerSet<Label> phone_syms_;
-
-  // disambig_syms_ is the set of integer ids of the disambiguation symbols,
-  // usually represented in text form as #0, #1, #2, etc.  These are inserted
-  // into the grammar (for #0) and the lexicon (for #1, #2, ...) in order to
-  // make the composed FSTs determinizable.  They are treated "specially" by the
-  // context FST in that they are not part of the context, they are just "passed
-  // through" via self-loops.  See the Mohri chapter mrentioned above for more
-  // information.
-  kaldifst::ConstIntegerSet<Label> disambig_syms_;
-
-  // subsequential_symbol_, represented as "$" in the Mohri chapter mentioned
-  // above, is something which terminates phonetic sequences to force out the
-  // last phones-in-context.  In our implementation it's added to det(LG) as a
-  // self-loop on final states before composing with C.
-  // (c.f. AddSubsequentialLoop()).
-  Label subsequential_symbol_;
-
-  // pseudo_eps_symbol_, which in printed form we refer to as "#-1", is a symbol
-  // that appears on the ilabels of the context transducer C, i.e. the olabels
-  // of this FST which is C's inverse.  It is a symbol we introduce to solve a
-  // special problem in systems with right-context (context_width_ >
-  // central_position_ + 1) that use disambiguation symbols.  It exists to
-  // prevent CLG from being nondeterminizable.
-  //
-  // The issue is that, in this case, the disambiguation symbols are shifted
-  // left w.r.t. the phones, and there becomes an ambiguity, if a disambiguation
-  // symbol appears at the start of a sequence on the input of CLG, about
-  // whether it was at the very start of the input of LG, or just after, say,
-  // the first real phone.  This can lead to determinization failure under
-  // certain circumstances.  What we do if we need pseudo_eps_symbol_ to be not
-  // epsilon, we create a special symbol with symbol-id 1 and sequence
-  // representation (ilabels entry) [ 0 ] .
-  int32_t pseudo_eps_symbol_;
-
-  // maps from vector<int32_t>, representing phonetic contexts of length
-  // context_width_ - 1, to StateId.  (The states of the "C" fst correspond to
-  // phonetic contexts, but we only create them as and when they are needed).
-  VectorToStateMap state_map_;
-
-  // The inverse of 'state_map_': gives us the phonetic context corresponding to
-  // each state-id.
-  std::vector<std::vector<int32_t>> state_seqs_;
-
-  // maps from vector<int32_t>, representing phonetic contexts of length
-  // context_width_ - 1, to Label.  These are actually the output labels of this
-  // InverseContextFst (because of the "Inverse" part), but for historical
-  // reasons and because we've used the term ilabels" in the documentation, we
-  // still call these "ilabels").
-  VectorToLabelMap ilabel_map_;
-
-  // ilabel_info_ is the reverse map of ilabel_map_.
-  // Indexed by olabel (although we call this ilabel_info_ for historical
-  // reasons and because is for the ilabels of C), ilabel_info_[i] gives
-  // information about the meaning of each symbol on the input of C
-  // aka the output of inv(C).
-  // See "http://kaldi-asr.org/doc/tree_externals.html#tree_ilabel".
-  std::vector<std::vector<int32_t>> ilabel_info_;
-};
-
-/**
-   Used in the command-line tool fstcomposecontext.  It creates a context FST
-   and composes it on the left with "ifst" to make "ofst".  It outputs the label
-   information to ilabels_out.  "ifst" is mutable because we need to add the
-   subsequential loop.
-
-    @param [in] disambig_syms  List of disambiguation symbols, e.g. the integer
-                 ids of #0, #1, #2 ... in the phones.txt.
-    @param [in] context_width  Size of context window, e.g. 3 for triphone.
-    @param [in] central_position  Central position in phonetic context window
-                  (zero-based index), e.g. 1 for triphone.
-    @param [in,out] ifst   The FST we are composing with C (e.g. LG.fst),
-   mustable because we need to add the subsequential loop to it.
-    @param [out] ofst   Composed output FST (would be CLG.fst).
-    @param [out] ilabels_out  Vector, indexed by ilabel of CLG.fst, providing
-   information about the meaning of that ilabel; see
-                  "http://kaldi-asr.org/doc/tree_externals.html#tree_ilabel".
-    @param [in] project_ifst  This is intended only to be set to true
-                  in the program 'fstmakecontextfst'... if true, it will
-                  project on the input after adding the subsequential loop
-                  to 'ifst', which allows us to reconstruct the context
-                  fst C.fst.
- */
-void ComposeContext(const std::vector<int32_t> &disambig_syms,
-                    int32_t context_width, int32_t central_position,
-                    VectorFst<StdArc> *ifst, VectorFst<StdArc> *ofst,
-                    std::vector<std::vector<int32_t>> *ilabels_out,
-                    bool project_ifst = false);
-
-}  // namespace fst
-
-#endif  // KALDIFST_CSRC_CONTEXT_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/deterministic-fst-inl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/deterministic-fst-inl.h
deleted file mode 100644
index 71189098..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/deterministic-fst-inl.h
+++ /dev/null
@@ -1,276 +0,0 @@
-// kaldifst/csrc/deterministic-fst-inl.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/deterministic-fst-inl.h
-
-// Copyright 2011-2012 Gilles Boulianne
-//                2014 Telepoint Global Hosting Service, LLC. (Author: David
-//                Snyder)
-//           2012-2015 Johns Hopkins University (author: Daniel Povey)
-#ifndef KALDIFST_CSRC_DETERMINISTIC_FST_INL_H_
-#define KALDIFST_CSRC_DETERMINISTIC_FST_INL_H_
-#include <algorithm>
-#include <queue>
-#include <unordered_map>
-#include <utility>
-
-#include "kaldifst/csrc/log.h"
-#include "kaldifst/csrc/stl-utils.h"
-
-namespace fst {
-
-template <class Arc>
-ComposeDeterministicOnDemandFst<Arc>::ComposeDeterministicOnDemandFst(
-    DeterministicOnDemandFst<Arc> *fst1, DeterministicOnDemandFst<Arc> *fst2)
-    : fst1_(fst1), fst2_(fst2) {
-  KALDIFST_ASSERT(fst1 != NULL && fst2 != NULL);
-  if (fst1_->Start() == -1 || fst2_->Start() == -1) {
-    start_state_ = -1;
-    next_state_ = 0;  // actually we don't care about this value.
-  } else {
-    start_state_ = 0;
-    std::pair<StateId, StateId> start_pair(fst1_->Start(), fst2_->Start());
-    state_map_[start_pair] = start_state_;
-    state_vec_.push_back(start_pair);
-    next_state_ = 1;
-  }
-}
-
-template <class Arc>
-typename Arc::Weight ComposeDeterministicOnDemandFst<Arc>::Final(StateId s) {
-  KALDIFST_ASSERT(s < static_cast<StateId>(state_vec_.size()));
-  const std::pair<StateId, StateId> &pr(state_vec_[s]);
-  return Times(fst1_->Final(pr.first), fst2_->Final(pr.second));
-}
-
-template <class Arc>
-bool ComposeDeterministicOnDemandFst<Arc>::GetArc(StateId s, Label ilabel,
-                                                  Arc *oarc) {
-  typedef typename MapType::iterator IterType;
-  KALDIFST_ASSERT(
-      ilabel != 0 &&
-      "This program expects epsilon-free compact lattices as input");
-  KALDIFST_ASSERT(s < static_cast<StateId>(state_vec_.size()));
-  const std::pair<StateId, StateId> pr(state_vec_[s]);
-
-  Arc arc1;
-  if (!fst1_->GetArc(pr.first, ilabel, &arc1)) return false;
-  if (arc1.olabel == 0) {  // There is no output label on the
-    // arc, so only the first state changes.
-    std::pair<const std::pair<StateId, StateId>, StateId> new_value(
-        std::pair<StateId, StateId>(arc1.nextstate, pr.second), next_state_);
-
-    std::pair<IterType, bool> result = state_map_.insert(new_value);
-    oarc->ilabel = ilabel;
-    oarc->olabel = 0;
-    oarc->nextstate = result.first->second;
-    oarc->weight = arc1.weight;
-    if (result.second == true) {  // was inserted
-      next_state_++;
-      const std::pair<StateId, StateId> &new_pair(new_value.first);
-      state_vec_.push_back(new_pair);
-    }
-    return true;
-  }
-  // There is an output label, so we need to traverse an arc on the
-  // second fst also.
-  Arc arc2;
-  if (!fst2_->GetArc(pr.second, arc1.olabel, &arc2)) return false;
-  std::pair<const std::pair<StateId, StateId>, StateId> new_value(
-      std::pair<StateId, StateId>(arc1.nextstate, arc2.nextstate), next_state_);
-  std::pair<IterType, bool> result = state_map_.insert(new_value);
-  oarc->ilabel = ilabel;
-  oarc->olabel = arc2.olabel;
-  oarc->nextstate = result.first->second;
-  oarc->weight = Times(arc1.weight, arc2.weight);
-  if (result.second == true) {  // was inserted
-    next_state_++;
-    const std::pair<StateId, StateId> &new_pair(new_value.first);
-    state_vec_.push_back(new_pair);
-  }
-  return true;
-}
-
-template <class Arc>
-void ComposeDeterministicOnDemand(const Fst<Arc> &fst1,
-                                  DeterministicOnDemandFst<Arc> *fst2,
-                                  MutableFst<Arc> *fst_composed) {
-  typedef typename Arc::Weight Weight;
-  typedef typename Arc::StateId StateId;
-  typedef std::pair<StateId, StateId> StatePair;
-  typedef unordered_map<StatePair, StateId, kaldifst::PairHasher<StateId>>
-      MapType;
-  typedef typename MapType::iterator IterType;
-
-  fst_composed->DeleteStates();
-
-  MapType state_map;
-  std::queue<StatePair> state_queue;
-
-  // Set start state in fst_composed.
-  StateId s1 = fst1.Start(), s2 = fst2->Start(),
-          start_state = fst_composed->AddState();
-  StatePair start_pair(s1, s2);
-  state_queue.push(start_pair);
-  fst_composed->SetStart(start_state);
-  // A mapping between pairs of states in fst1 and fst2 and the corresponding
-  // state in fst_composed.
-  std::pair<const StatePair, StateId> start_map(start_pair, start_state);
-  std::pair<IterType, bool> result = state_map.insert(start_map);
-  KALDIFST_ASSERT(result.second == true);
-
-  while (!state_queue.empty()) {
-    StatePair q = state_queue.front();
-    StateId q1 = q.first, q2 = q.second;
-    state_queue.pop();
-    // If the product of the final weights of the two fsts is non-zero then
-    // we can set a final-prob in fst_composed
-    Weight final_weight = Times(fst1.Final(q1), fst2->Final(q2));
-    if (final_weight != Weight::Zero()) {
-      KALDIFST_ASSERT(state_map.find(q) != state_map.end());
-      fst_composed->SetFinal(state_map[q], final_weight);
-    }
-
-    // for each pair of edges from fst1 and fst2 at q1 and q2.
-    for (ArcIterator<Fst<Arc>> aiter(fst1, q1); !aiter.Done(); aiter.Next()) {
-      const Arc &arc1 = aiter.Value();
-      Arc arc2;
-      StatePair next_pair;
-      StateId next_state1 = arc1.nextstate, next_state2, next_state;
-      // If there is an epsilon on the arc of fst1 we transition to the next
-      // state but keep fst2 at the current state.
-      if (arc1.olabel == 0) {
-        next_state2 = q2;
-      } else {
-        bool match = fst2->GetArc(q2, arc1.olabel, &arc2);
-        if (!match)  // There is no matching arc -> nothing to do.
-          continue;
-        next_state2 = arc2.nextstate;
-      }
-      next_pair = StatePair(next_state1, next_state2);
-      IterType sitr = state_map.find(next_pair);
-      // If sitr == state_map.end() then the state isn't in fst_composed yet.
-      if (sitr == state_map.end()) {
-        next_state = fst_composed->AddState();
-        std::pair<const StatePair, StateId> new_state(next_pair, next_state);
-        std::pair<IterType, bool> result = state_map.insert(new_state);
-        // Since we already checked if state_map contained new_state,
-        // it should always be added if we reach here.
-        KALDIFST_ASSERT(result.second == true);
-        state_queue.push(next_pair);
-        // If sitr != state_map.end() then the next state is already in
-        // the state_map.
-      } else {
-        next_state = sitr->second;
-      }
-      if (arc1.olabel == 0) {
-        fst_composed->AddArc(state_map[q],
-                             Arc(arc1.ilabel, 0, arc1.weight, next_state));
-      } else {
-        fst_composed->AddArc(state_map[q],
-                             Arc(arc1.ilabel, arc2.olabel,
-                                 Times(arc1.weight, arc2.weight), next_state));
-      }
-    }
-  }
-}
-
-// we are doing *fst_composed = Compose(Inverse(*left), right).
-template <class Arc>
-void ComposeDeterministicOnDemandInverse(const Fst<Arc> &right,
-                                         DeterministicOnDemandFst<Arc> *left,
-                                         MutableFst<Arc> *fst_composed) {
-  typedef typename Arc::Weight Weight;
-  typedef typename Arc::StateId StateId;
-  typedef std::pair<StateId, StateId> StatePair;
-  typedef unordered_map<StatePair, StateId, kaldifst::PairHasher<StateId>>
-      MapType;
-  typedef typename MapType::iterator IterType;
-
-  fst_composed->DeleteStates();
-
-  // the queue and map contain pairs (state-in-left, state-in-right)
-  MapType state_map;
-  std::queue<StatePair> state_queue;
-
-  // Set start state in fst_composed.
-  StateId s_left = left->Start(), s_right = right.Start();
-  if (s_left == kNoStateId || s_right == kNoStateId) return;  // Empty result.
-  StatePair start_pair(s_left, s_right);
-  StateId start_state = fst_composed->AddState();
-  state_queue.push(start_pair);
-  fst_composed->SetStart(start_state);
-  // A mapping between pairs of states in *left and right, and the corresponding
-  // state in fst_composed.
-  std::pair<const StatePair, StateId> start_map(start_pair, start_state);
-  std::pair<IterType, bool> result = state_map.insert(start_map);
-  KALDIFST_ASSERT(result.second == true);
-
-  while (!state_queue.empty()) {
-    StatePair q = state_queue.front();
-    StateId q_left = q.first, q_right = q.second;
-    state_queue.pop();
-    // If the product of the final weights of the two fsts is non-zero then
-    // we can set a final-prob in fst_composed
-    Weight final_weight = Times(left->Final(q_left), right.Final(q_right));
-    if (final_weight != Weight::Zero()) {
-      KALDIFST_ASSERT(state_map.find(q) != state_map.end());
-      fst_composed->SetFinal(state_map[q], final_weight);
-    }
-
-    for (ArcIterator<Fst<Arc>> aiter(right, q_right); !aiter.Done();
-         aiter.Next()) {
-      const Arc &arc_right = aiter.Value();
-      Arc arc_left;
-      StatePair next_pair;
-      StateId next_state_right = arc_right.nextstate, next_state_left,
-              next_state;
-      // If there is an epsilon on the input side of the rigth arc, we
-      // transition to the next state of the output but keep 'left' at the
-      // current state.
-      if (arc_right.ilabel == 0) {
-        next_state_left = q_left;
-      } else {
-        bool match = left->GetArc(q_left, arc_right.ilabel, &arc_left);
-        if (!match)  // There is no matching arc -> nothing to do.
-          continue;
-        // the next 'swap' is because we are composing with the inverse of
-        // *left.  Just removing the swap statement wouldn't let us compose
-        // with non-inverted *left though, because the GetArc function call
-        // above interprets the second argument as an ilabel not an olabel.
-        std::swap(arc_left.ilabel, arc_left.olabel);
-        next_state_left = arc_left.nextstate;
-      }
-      next_pair = StatePair(next_state_left, next_state_right);
-      IterType sitr = state_map.find(next_pair);
-      // If sitr == state_map.end() then the state isn't in fst_composed yet.
-      if (sitr == state_map.end()) {
-        next_state = fst_composed->AddState();
-        std::pair<const StatePair, StateId> new_state(next_pair, next_state);
-        std::pair<IterType, bool> result = state_map.insert(new_state);
-        // Since we already checked if state_map contained new_state,
-        // it should always be added if we reach here.
-        KALDIFST_ASSERT(result.second == true);
-        state_queue.push(next_pair);
-        // If sitr != state_map.end() then the next state is already in
-        // the state_map.
-      } else {
-        next_state = sitr->second;
-      }
-      if (arc_right.ilabel == 0) {
-        // we didn't get an actual arc from the left FST.
-        fst_composed->AddArc(state_map[q], Arc(0, arc_right.olabel,
-                                               arc_right.weight, next_state));
-      } else {
-        fst_composed->AddArc(
-            state_map[q],
-            Arc(arc_left.ilabel, arc_right.olabel,
-                Times(arc_left.weight, arc_right.weight), next_state));
-      }
-    }
-  }
-}
-
-}  // namespace fst
-
-#endif  // KALDIFST_CSRC_DETERMINISTIC_FST_INL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/deterministic-fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/deterministic-fst.h
deleted file mode 100644
index 16943f3d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/deterministic-fst.h
+++ /dev/null
@@ -1,121 +0,0 @@
-// kaldifst/csrc/deterministic-fst.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/deterministic-fst.h
-
-// Copyright 2011-2012 Gilles Boulianne
-//                2014 Telepoint Global Hosting Service, LLC. (Author: David
-//                Snyder)
-//           2012-2015 Johns Hopkins University (author: Daniel Povey)
-#ifndef KALDIFST_CSRC_DETERMINISTIC_FST_H_
-#define KALDIFST_CSRC_DETERMINISTIC_FST_H_
-
-/* This header defines the DeterministicOnDemand interface,
-   which is an FST with a special interface that allows
-   only a single arc with a non-epsilon input symbol
-   out of each state.
-*/
-
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include "fst/fst-decl.h"
-#include "fst/fstlib.h"
-
-namespace fst {
-
-/// \addtogroup deterministic_fst_group "Classes and functions related to
-/// on-demand deterministic FST's"
-/// @{
-
-/// class DeterministicOnDemandFst is an "FST-like" base-class.  It does not
-/// actually inherit from any Fst class because its interface is not exactly the
-/// same; it's much smaller.  It assumes that the FST can have only one arc for
-/// any given input symbol, which makes the GetArc function below possible.
-/// (The FST is also assumed to be free of input epsilons).  Note: we don't use
-/// "const" in this interface, because it creates problems when we do things
-/// like caching.
-template <class Arc>
-class DeterministicOnDemandFst {
- public:
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Weight Weight;
-  typedef typename Arc::Label Label;
-
-  virtual StateId Start() = 0;
-
-  virtual Weight Final(StateId s) = 0;
-
-  /// Note: ilabel must not be epsilon.
-  virtual bool GetArc(StateId s, Label ilabel, Arc *oarc) = 0;
-
-  virtual ~DeterministicOnDemandFst() = default;
-};
-
-template <class Arc>
-class ComposeDeterministicOnDemandFst : public DeterministicOnDemandFst<Arc> {
- public:
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Weight Weight;
-  typedef typename Arc::Label Label;
-
-  /// Note: constructor does not "take ownership" of the input fst's.  The input
-  /// fst's should be treated as const, in that their contents do not change,
-  /// but they are not const as the DeterministicOnDemandFst's data-access
-  /// functions are not const, for reasons relating to caching.
-  ComposeDeterministicOnDemandFst(DeterministicOnDemandFst<Arc> *fst1,
-                                  DeterministicOnDemandFst<Arc> *fst2);
-
-  StateId Start() override { return start_state_; }
-
-  Weight Final(StateId s) override;
-
-  bool GetArc(StateId s, Label ilabel, Arc *oarc) override;
-
- private:
-  DeterministicOnDemandFst<Arc> *fst1_;
-  DeterministicOnDemandFst<Arc> *fst2_;
-  typedef unordered_map<std::pair<StateId, StateId>, StateId,
-                        kaldifst::PairHasher<StateId>>
-      MapType;
-  MapType state_map_;
-  std::vector<std::pair<StateId, StateId>> state_vec_;  // maps from
-  // StateId to pair.
-  StateId next_state_;
-  StateId start_state_;
-};
-
-// Compose an FST (which may be a lattice) with a DeterministicOnDemandFst and
-// store the result in fst_composed.  This is mainly used for expanding lattice
-// n-gram histories, where fst1 is a lattice and fst2 is an UnweightedNgramFst.
-// This does not call Connect.
-template <class Arc>
-void ComposeDeterministicOnDemand(const Fst<Arc> &fst1,
-                                  DeterministicOnDemandFst<Arc> *fst2,
-                                  MutableFst<Arc> *fst_composed);
-
-/**
-   This function does
-   '*fst_composed = Compose(Inverse(*fst2), fst1)'
-   Note that the arguments are reversed; this is unfortunate but it's
-   because the fst2 argument needs to be non-const and non-const arguments
-   must follow const ones.
-   This is the counterpart to ComposeDeterministicOnDemand, used for
-   the case where the DeterministicOnDemandFst is on the left.  The
-   reason why we need to make the left-hand argument to compose the
-   inverse of 'fst2' (i.e. with the input and output symbols swapped),
-   is that the DeterministicOnDemandFst interface only supports lookup
-   by ilabel (see its function GetArc).
-   This does not call Connect().
-*/
-template <class Arc>
-void ComposeDeterministicOnDemandInverse(const Fst<Arc> &fst1,
-                                         DeterministicOnDemandFst<Arc> *fst2,
-                                         MutableFst<Arc> *fst_composed);
-
-}  // namespace fst
-
-#include "kaldifst/csrc/deterministic-fst-inl.h"
-
-#endif  // KALDIFST_CSRC_DETERMINISTIC_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/determinize-star-inl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/determinize-star-inl.h
deleted file mode 100644
index 96161876..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/determinize-star-inl.h
+++ /dev/null
@@ -1,1116 +0,0 @@
-// kaldifst/csrc/determinize-star-inl.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/determinize-star-inl.h
-
-// copyright 2009-2011  microsoft corporation;  jan silovsky
-//           2015 hainan xu
-
-#ifndef KALDIFST_CSRC_DETERMINIZE_STAR_INL_H_
-#define KALDIFST_CSRC_DETERMINIZE_STAR_INL_H_
-
-#include <algorithm>
-#include <deque>
-#include <limits>
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include "fst/fst.h"
-#include "kaldifst/csrc/log.h"
-
-namespace fst {
-
-// This class maps back and forth from/to integer id's to sequences of strings.
-// used in determinization algorithm.
-template <class Label, class StringId>
-class StringRepository {
-  // Label and StringId are both integer types, possibly the same.
-  // This is a utility that maps back and forth between a vector<Label> and
-  // StringId representation of sequences of Labels.  It is to save memory, and
-  // to save compute. We treat sequences of length zero and one separately, for
-  // efficiency.
-
- public:
-  class VectorKey {  // Hash function object.
-   public:
-    size_t operator()(const std::vector<Label> *vec) const {
-      assert(vec != NULL);
-      size_t hash = 0, factor = 1;
-      for (typename std::vector<Label>::const_iterator it = vec->begin();
-           it != vec->end(); it++) {
-        hash += factor * (*it);
-        factor *= 103333;  // just an arbitrary prime number.
-      }
-      return hash;
-    }
-  };
-  class VectorEqual {  // Equality-operator function object.
-   public:
-    size_t operator()(const std::vector<Label> *vec1,
-                      const std::vector<Label> *vec2) const {
-      return (*vec1 == *vec2);
-    }
-  };
-
-  typedef unordered_map<const std::vector<Label> *, StringId, VectorKey,
-                        VectorEqual>
-      MapType;
-
-  StringId IdOfEmpty() { return no_symbol; }
-
-  StringId IdOfLabel(Label l) {
-    if (l >= 0 && l <= (Label)single_symbol_range) {
-      return l + single_symbol_start;
-    } else {
-      // l is out of the allowed range so we have to treat it as a sequence of
-      // length one.  Should be v. rare.
-      std::vector<Label> v;
-      v.push_back(l);
-      return IdOfSeqInternal(v);
-    }
-  }
-
-  StringId IdOfSeq(
-      const std::vector<Label> &v) {  // also works for sizes 0 and 1.
-    size_t sz = v.size();
-    if (sz == 0)
-      return no_symbol;
-    else if (v.size() == 1)
-      return IdOfLabel(v[0]);
-    else
-      return IdOfSeqInternal(v);
-  }
-
-  inline bool IsEmptyString(StringId id) { return id == no_symbol; }
-  void SeqOfId(StringId id, std::vector<Label> *v) {
-    if (id == no_symbol) {
-      v->clear();
-    } else if (id >= single_symbol_start) {
-      v->resize(1);
-      (*v)[0] = id - single_symbol_start;
-    } else {
-      assert(static_cast<size_t>(id) < vec_.size());
-      *v = *(vec_[id]);
-    }
-  }
-  StringId RemovePrefix(StringId id, size_t prefix_len) {
-    if (prefix_len == 0) {
-      return id;
-    } else {
-      std::vector<Label> v;
-      SeqOfId(id, &v);
-      size_t sz = v.size();
-      assert(sz >= prefix_len);
-      std::vector<Label> v_noprefix(sz - prefix_len);
-      for (size_t i = 0; i < sz - prefix_len; i++)
-        v_noprefix[i] = v[i + prefix_len];
-      return IdOfSeq(v_noprefix);
-    }
-  }
-
-  StringRepository() {
-    // The following are really just constants but don't want to complicate
-    // compilation so make them class variables.  Due to the brokenness of
-    // <limits>, they can't be accessed as constants.
-    string_end = (std::numeric_limits<StringId>::max() / 2) -
-                 1;  // all hash values must be <= this.
-    no_symbol = (std::numeric_limits<StringId>::max() /
-                 2);  // reserved for empty sequence.
-    single_symbol_start = (std::numeric_limits<StringId>::max() / 2) + 1;
-    single_symbol_range =
-        std::numeric_limits<StringId>::max() - single_symbol_start;
-  }
-  void Destroy() {
-    for (typename std::vector<std::vector<Label> *>::iterator iter =
-             vec_.begin();
-         iter != vec_.end(); ++iter)
-      delete *iter;
-    std::vector<std::vector<Label> *> tmp_vec;
-    tmp_vec.swap(vec_);
-    MapType tmp_map;
-    tmp_map.swap(map_);
-  }
-  ~StringRepository() { Destroy(); }
-
-  KALDIFST_DISALLOW_COPY_AND_ASSIGN(StringRepository);
-
- private:
-  StringId IdOfSeqInternal(const std::vector<Label> &v) {
-    typename MapType::iterator iter = map_.find(&v);
-    if (iter != map_.end()) {
-      return iter->second;
-    } else {  // must add it to map.
-      StringId this_id = (StringId)vec_.size();
-      std::vector<Label> *v_new = new std::vector<Label>(v);
-      vec_.push_back(v_new);
-      map_[v_new] = this_id;
-      assert(this_id < string_end);  // or we used up the labels.
-      return this_id;
-    }
-  }
-
-  std::vector<std::vector<Label> *> vec_;
-  MapType map_;
-
-  static const StringId string_start =
-      (StringId)0;      // This must not change.  It's assumed.
-  StringId string_end;  // = (numeric_limits<StringId>::max() / 2) - 1; // all
-                        // hash values must be <= this.
-  StringId no_symbol;   // = (numeric_limits<StringId>::max() / 2); // reserved
-                        // for empty sequence.
-  StringId
-      single_symbol_start;  // =  (numeric_limits<StringId>::max() / 2) + 1;
-  StringId single_symbol_range;  // =  numeric_limits<StringId>::max() -
-                                 // single_symbol_start;
-};
-
-template <class F>
-class DeterminizerStar {
-  typedef typename F::Arc Arc;
-
- public:
-  KALDIFST_DISALLOW_COPY_AND_ASSIGN(DeterminizerStar);
-  // Output to standard FST.  We will create extra states to handle sequences of
-  // symbols on the output.  If destroy == true, release memory as we go (but we
-  // cannot output again).
-  void Output(MutableFst<Arc> *ofst, bool destroy = true);
-
-  // Initializer.  After initializing the object you will typically call
-  // Determinize() and then one of the Output functions.
-  DeterminizerStar(const Fst<Arc> &ifst, float delta = kDelta,
-                   int max_states = -1, bool allow_partial = false)
-      : ifst_(ifst.Copy()),
-        delta_(delta),
-        max_states_(max_states),
-        determinized_(false),
-        allow_partial_(allow_partial),
-        is_partial_(false),
-        equal_(delta),
-        // clang-format off
-        hash_(ifst.Properties(kExpanded, false)
-                  ? down_cast<const ExpandedFst<Arc> *, const Fst<Arc>>(&ifst)
-                                ->NumStates() / 2 +3
-                  : 20,
-              hasher_, equal_),
-        // clang-format on
-        epsilon_closure_(ifst_, max_states, &repository_, delta) {}
-
-  void Determinize(bool *debug_ptr) {
-    KALDIFST_ASSERT(!determinized_);
-    // This determinizes the input fst but leaves it in the "special format"
-    // in "output_arcs_".
-    InputStateId start_id = ifst_->Start();
-    if (start_id == kNoStateId) {
-      determinized_ = true;
-      // Nothing to do.
-      return;
-    } else {  // Insert start state into hash and queue.
-      Element elem;
-      elem.state = start_id;
-      elem.weight = Weight::One();
-      elem.string = repository_.IdOfEmpty();  // Id of empty sequence.
-      std::vector<Element> vec;
-      vec.push_back(elem);
-      OutputStateId cur_id = SubsetToStateId(vec);
-      assert(cur_id == 0 && "Do not call Determinize twice.");
-    }
-    while (!Q_.empty()) {
-      std::pair<std::vector<Element> *, OutputStateId> cur_pair = Q_.front();
-      Q_.pop_front();
-      ProcessSubset(cur_pair);
-      if (debug_ptr && *debug_ptr) Debug();  // will exit.
-      if (max_states_ > 0 && output_arcs_.size() > max_states_) {
-        if (allow_partial_ == false) {
-          KALDIFST_ERR << "Determinization aborted since passed " << max_states_
-                       << " states";
-        } else {
-          KALDIFST_WARN << "Determinization terminated since passed "
-                        << max_states_
-                        << " states, partial results will be generated";
-          is_partial_ = true;
-          break;
-        }
-      }
-    }
-    determinized_ = true;
-  }
-
-  bool IsPartial() { return is_partial_; }
-
-  // frees all except output_arcs_, which contains the important info
-  // we need to output.
-  void FreeMostMemory() {
-    if (ifst_) {
-      delete ifst_;
-      ifst_ = NULL;
-    }
-    for (typename SubsetHash::iterator iter = hash_.begin();
-         iter != hash_.end(); ++iter)
-      delete iter->first;
-    SubsetHash tmp;
-    tmp.swap(hash_);
-  }
-
-  ~DeterminizerStar() { FreeMostMemory(); }
-
- private:
-  typedef typename Arc::Label Label;
-  typedef typename Arc::Weight Weight;
-  typedef typename Arc::StateId InputStateId;
-  typedef typename Arc::StateId
-      OutputStateId;  // same as above but distinguish states in output Fst.
-  typedef typename Arc::Label StringId;  // Id type used in the StringRepository
-  typedef StringRepository<Label, StringId> StringRepositoryType;
-
-  // Element of a subset [of original states]
-
-  struct Element {
-    InputStateId state;
-    StringId string;
-    Weight weight;
-    bool operator!=(const Element &other) const {
-      return (state != other.state || string != other.string ||
-              weight != other.weight);
-    }
-  };
-
-  // Arcs in the format we temporarily create in this class (a representation,
-  // essentially of a Gallic Fst).
-  struct TempArc {
-    Label ilabel;
-    StringId ostring;  // Look it up in the StringRepository, it's a sequence of
-                       // Labels.
-    OutputStateId nextstate;  // or kNoState for final weights.
-    Weight weight;
-  };
-
-  // Hashing function used in hash of subsets.
-  // A subset is a pointer to vector<Element>.
-  // The Elements are in sorted order on state id, and without repeated states.
-  // Because the order of Elements is fixed, we can use a hashing function that
-  // is order-dependent.  However the weights are not included in the hashing
-  // function-- we hash subsets that differ only in weight to the same key. This
-  // is not optimal in terms of the O(N) performance but typically if we have a
-  // lot of determinized states that differ only in weight then the input
-  // probably was pathological in some way, or even non-determinizable.
-  //   We don't quantize the weights, in order to avoid inexactness in simple
-  //   cases.
-  // Instead we apply the delta when comparing subsets for equality, and allow a
-  // small difference.
-
-  class SubsetKey {
-   public:
-    size_t operator()(const std::vector<Element> *subset)
-        const {  // hashes only the state and string.
-      size_t hash = 0, factor = 1;
-      for (typename std::vector<Element>::const_iterator iter = subset->begin();
-           iter != subset->end(); ++iter) {
-        hash *= factor;
-        hash += iter->state + size_t(103333) * iter->string;
-        factor *= 23531;  // these numbers are primes.
-      }
-      return hash;
-    }
-  };
-
-  // This is the equality operator on subsets.  It checks for exact match on
-  // state-id and string, and approximate match on weights.
-  class SubsetEqual {
-   public:
-    bool operator()(const std::vector<Element> *s1,
-                    const std::vector<Element> *s2) const {
-      size_t sz = s1->size();
-      assert(sz >= 0);
-      if (sz != s2->size()) return false;
-      typename std::vector<Element>::const_iterator iter1 = s1->begin(),
-                                                    iter1_end = s1->end(),
-                                                    iter2 = s2->begin();
-      for (; iter1 < iter1_end; ++iter1, ++iter2) {
-        if (iter1->state != iter2->state || iter1->string != iter2->string ||
-            !ApproxEqual(iter1->weight, iter2->weight, delta_))
-          return false;
-      }
-      return true;
-    }
-    float delta_;
-    explicit SubsetEqual(float delta) : delta_(delta) {}
-    SubsetEqual() : delta_(kDelta) {}
-  };
-
-  // Operator that says whether two Elements have the same states.
-  // Used only for debug.
-  class SubsetEqualStates {
-   public:
-    bool operator()(const std::vector<Element> *s1,
-                    const std::vector<Element> *s2) const {
-      size_t sz = s1->size();
-      assert(sz >= 0);
-      if (sz != s2->size()) return false;
-      typename std::vector<Element>::const_iterator iter1 = s1->begin(),
-                                                    iter1_end = s1->end(),
-                                                    iter2 = s2->begin();
-      for (; iter1 < iter1_end; ++iter1, ++iter2) {
-        if (iter1->state != iter2->state) return false;
-      }
-      return true;
-    }
-  };
-
-  // Define the hash type we use to store subsets.
-  typedef std::unordered_map<const std::vector<Element> *, OutputStateId,
-                             SubsetKey, SubsetEqual>
-      SubsetHash;
-
-  class EpsilonClosure {
-   public:
-    EpsilonClosure(const Fst<Arc> *ifst, int max_states,
-                   StringRepository<Label, StringId> *repository, float delta)
-        : ifst_(ifst),
-          max_states_(max_states),
-          repository_(repository),
-          delta_(delta) {}
-
-    // This function computes epsilon closure of subset of states by following
-    // epsilon links. Called by ProcessSubset. Has no side effects except on the
-    // repository.
-    void GetEpsilonClosure(const std::vector<Element> &input_subset,
-                           std::vector<Element> *output_subset);
-
-   private:
-    struct EpsilonClosureInfo {
-      EpsilonClosureInfo() {}
-      EpsilonClosureInfo(const Element &e, const Weight &w, bool i)
-          : element(e), weight_to_process(w), in_queue(i) {}
-      // the weight in the Element struct is the total current weight
-      // that has been processed already
-      Element element;
-      // this stores the weight that we haven't processed (propagated)
-      Weight weight_to_process;
-      // whether "this" struct is in the queue
-      // we store the info here so that we don't have to look it up every time
-      bool in_queue;
-      bool operator<(const EpsilonClosureInfo &other) const {
-        return this->element.state < other.element.state;
-      }
-    };
-
-    // to further speed up EpsilonClosure() computation, we have 2 queues
-    // the 2nd queue is used when we first iterate over the input set -
-    // if queue_2_.empty() then we directly set output_set equal to input_set
-    // and return immediately
-    // Since Epsilon arcs are relatively rare, this way we could efficiently
-    // detect the epsilon-free case, without having to waste our computation
-    // e.g. allocating the EpsilonClosureInfo structure; this also lets us do a
-    // level-by-level traversal, which could avoid some (unfortunately not all)
-    // duplicate computation if epsilons form a DAG that is not a tree
-    //
-    // We put the queues here for better efficiency for memory allocation
-    std::deque<typename Arc::StateId> queue_;
-    std::vector<Element> queue_2_;
-
-    // the following 2 structures together form our *virtual "map"*
-    // basically we need a map from state_id to EpsilonClosureInfo that operates
-    // in O(1) time, while still takes relatively small mem, and this does it
-    // well for efficiency we don't clear id_to_index_ of its outdated
-    // information As a result each time we do a look-up, we need to check if
-    // (ecinfo_[id_to_index_[id]].element.state == id) Yet this is still faster
-    // than using a std::map<StateId, EpsilonClosureInfo>
-    std::vector<int> id_to_index_;
-    // unlike id_to_index_, we clear the content of ecinfo_ each time we call
-    // EpsilonClosure(). This needed because we need an efficient way to
-    // traverse the virtual map - it is just too costly to traverse the
-    // id_to_index_ vector.
-    std::vector<EpsilonClosureInfo> ecinfo_;
-
-    // Add one element (elem) into cur_subset
-    // it also adds the necessary stuff to queue_, set the correct weight
-    void AddOneElement(const Element &elem, const Weight &unprocessed_weight);
-
-    // Sub-routine that we call in EpsilonClosure()
-    // It takes the current "unprocessed_weight" and propagate it to the
-    // states accessible from elem.state by an epsilon arc
-    // and add the results to cur_subset.
-    // save_to_queue_2 is set true when we iterate over the initial subset
-    // - then we save it to queue_2 s.t. if it's empty, we directly return
-    // the input set
-    void ExpandOneElement(const Element &elem, bool sorted,
-                          const Weight &unprocessed_weight,
-                          bool save_to_queue_2 = false);
-
-    // no pointers below would take the ownership
-    const Fst<Arc> *ifst_;
-    int max_states_;
-    StringRepository<Label, StringId> *repository_;
-    float delta_;
-  };
-
-  // This function works out the final-weight of the determinized state.
-  // called by ProcessSubset.
-  // Has no side effects except on the variable repository_, and output_arcs_.
-
-  void ProcessFinal(const std::vector<Element> &closed_subset,
-                    OutputStateId state) {
-    // processes final-weights for this subset.
-    bool is_final = false;
-    StringId final_string = 0;  // = 0 to keep compiler happy.
-    Weight final_weight =
-        Weight::One();  // This value will never be accessed, and
-    // we just set it to avoid spurious compiler warnings.  We avoid setting it
-    // to Zero() because floating-point infinities can sometimes generate
-    // interrupts and slow things down.
-    typename std::vector<Element>::const_iterator iter = closed_subset.begin(),
-                                                  end = closed_subset.end();
-    for (; iter != end; ++iter) {
-      const Element &elem = *iter;
-      Weight this_final_weight = ifst_->Final(elem.state);
-      if (this_final_weight != Weight::Zero()) {
-        if (!is_final) {  // first final-weight
-          final_string = elem.string;
-          final_weight = Times(elem.weight, this_final_weight);
-          is_final = true;
-        } else {  // already have one.
-          if (final_string != elem.string) {
-            KALDIFST_ERR << "FST was not functional -> not determinizable";
-          }
-          final_weight =
-              Plus(final_weight, Times(elem.weight, this_final_weight));
-        }
-      }
-    }
-    if (is_final) {
-      // store final weights in TempArc structure, just like a transition.
-      TempArc temp_arc;
-      temp_arc.ilabel = 0;
-      temp_arc.nextstate =
-          kNoStateId;  // special marker meaning "final weight".
-      temp_arc.ostring = final_string;
-      temp_arc.weight = final_weight;
-      output_arcs_[state].push_back(temp_arc);
-    }
-  }
-
-  // ProcessTransition is called from "ProcessTransitions".  Broken out for
-  // clarity.  Has side effects on output_arcs_, and (via SubsetToStateId), Q_
-  // and hash_.
-  void ProcessTransition(OutputStateId state, Label ilabel,
-                         std::vector<Element> *subset);
-
-  // "less than" operator for pair<Label, Element>.   Used in
-  // ProcessTransitions. Lexicographical order, with comparing the state only
-  // for "Element".
-
-  class PairComparator {
-   public:
-    inline bool operator()(const std::pair<Label, Element> &p1,
-                           const std::pair<Label, Element> &p2) {
-      if (p1.first < p2.first) {
-        return true;
-      } else if (p1.first > p2.first) {
-        return false;
-      } else {
-        return p1.second.state < p2.second.state;
-      }
-    }
-  };
-
-  // ProcessTransitions handles transitions out of this subset of states.
-  // Ignores epsilon transitions (epsilon closure already handled that).
-  // Does not consider final states.  Breaks the transitions up by ilabel,
-  // and creates a new transition in determinized FST, for each ilabel.
-  // Does this by creating a big vector of pairs <Label, Element> and then
-  // sorting them using a lexicographical ordering, and calling
-  // ProcessTransition for each range with the same ilabel. Side effects on
-  // repository, and (via ProcessTransition) on Q_, hash_, and output_arcs_.
-  void ProcessTransitions(const std::vector<Element> &closed_subset,
-                          OutputStateId state) {
-    std::vector<std::pair<Label, Element>> all_elems;
-    {  // Push back into "all_elems", elements corresponding to all
-       // non-epsilon-input transitions
-      // out of all states in "closed_subset".
-      typename std::vector<Element>::const_iterator iter =
-                                                        closed_subset.begin(),
-                                                    end = closed_subset.end();
-      for (; iter != end; ++iter) {
-        const Element &elem = *iter;
-        for (ArcIterator<Fst<Arc>> aiter(*ifst_, elem.state); !aiter.Done();
-             aiter.Next()) {
-          const Arc &arc = aiter.Value();
-          if (arc.ilabel !=
-              0) {  // Non-epsilon transition -- ignore epsilons here.
-            std::pair<Label, Element> this_pr;
-            this_pr.first = arc.ilabel;
-            Element &next_elem(this_pr.second);
-            next_elem.state = arc.nextstate;
-            next_elem.weight = Times(elem.weight, arc.weight);
-            if (arc.olabel == 0) {  // output epsilon-- this is simple case so
-                                    // handle separately for efficiency
-              next_elem.string = elem.string;
-            } else {
-              std::vector<Label> seq;
-              repository_.SeqOfId(elem.string, &seq);
-              seq.push_back(arc.olabel);
-              next_elem.string = repository_.IdOfSeq(seq);
-            }
-            all_elems.push_back(this_pr);
-          }
-        }
-      }
-    }
-    PairComparator pc;
-    std::sort(all_elems.begin(), all_elems.end(), pc);
-    // now sorted first on input label, then on state.
-    typedef typename std::vector<std::pair<Label, Element>>::const_iterator
-        PairIter;
-    PairIter cur = all_elems.begin(), end = all_elems.end();
-    std::vector<Element> this_subset;
-    while (cur != end) {
-      // Process ranges that share the same input symbol.
-      Label ilabel = cur->first;
-      this_subset.clear();
-      while (cur != end && cur->first == ilabel) {
-        this_subset.push_back(cur->second);
-        cur++;
-      }
-      // We now have a subset for this ilabel.
-      ProcessTransition(state, ilabel, &this_subset);
-    }
-  }
-
-  // SubsetToStateId converts a subset (vector of Elements) to a StateId in the
-  // output fst.  This is a hash lookup; if no such state exists, it adds a new
-  // state to the hash and adds a new pair to the queue. Side effects on hash_
-  // and Q_, and on output_arcs_ [just affects the size].
-  OutputStateId SubsetToStateId(
-      const std::vector<Element> &subset) {  // may add the subset to the queue.
-    typedef typename SubsetHash::iterator IterType;
-    IterType iter = hash_.find(&subset);
-    if (iter == hash_.end()) {  // was not there.
-      std::vector<Element> *new_subset = new std::vector<Element>(subset);
-      OutputStateId new_state_id = (OutputStateId)output_arcs_.size();
-      bool ans =
-          hash_
-              .insert(std::pair<const std::vector<Element> *, OutputStateId>(
-                  new_subset, new_state_id))
-              .second;
-      assert(ans);
-      output_arcs_.push_back(std::vector<TempArc>());
-      if (allow_partial_ == false) {
-        // If --allow-partial is not requested, we do the old way.
-        Q_.push_front(std::pair<std::vector<Element> *, OutputStateId>(
-            new_subset, new_state_id));
-      } else {
-        // If --allow-partial is requested, we do breadth first search. This
-        // ensures that when we return partial results, we return the states
-        // that are reachable by the fewest steps from the start state.
-        Q_.push_back(std::pair<std::vector<Element> *, OutputStateId>(
-            new_subset, new_state_id));
-      }
-      return new_state_id;
-    } else {
-      return iter->second;  // the OutputStateId.
-    }
-  }
-
-  // ProcessSubset does the processing of a determinized state, i.e. it creates
-  // transitions out of it and adds new determinized states to the queue if
-  // necessary. The first stage is "EpsilonClosure" (follow epsilons to get a
-  // possibly larger set of (states, weights)).  After that we ignore epsilons.
-  // We process the final-weight of the state, and then handle transitions out
-  // (this may add more determinized states to the queue).
-  void ProcessSubset(
-      const std::pair<std::vector<Element> *, OutputStateId> &pair) {
-    const std::vector<Element> *subset = pair.first;
-    OutputStateId state = pair.second;
-
-    std::vector<Element> closed_subset;  // subset after epsilon closure.
-    epsilon_closure_.GetEpsilonClosure(*subset, &closed_subset);
-
-    // Now follow non-epsilon arcs [and also process final states]
-    ProcessFinal(closed_subset, state);
-
-    // Now handle transitions out of these states.
-    ProcessTransitions(closed_subset, state);
-  }
-
-  void Debug();
-
-  std::deque<std::pair<std::vector<Element> *, OutputStateId>>
-      Q_;  // queue of subsets to be processed.
-
-  std::vector<std::vector<TempArc>>
-      output_arcs_;  // essentially an FST in our format.
-
-  const Fst<Arc> *ifst_;
-  float delta_;
-  int max_states_;
-  bool determinized_;   // used to check usage.
-  bool allow_partial_;  // output paritial results or not
-  bool is_partial_;     // if we get partial results or not
-  SubsetKey hasher_;    // object that computes keys-- has no data members.
-  SubsetEqual
-      equal_;  // object that compares subsets-- only data member is delta_.
-  SubsetHash hash_;  // hash from Subset to StateId in final Fst.
-
-  StringRepository<Label, StringId>
-      repository_;  // associate integer id's with sequences of labels.
-  EpsilonClosure epsilon_closure_;
-};
-
-template <class F>
-void DeterminizerStar<F>::EpsilonClosure::GetEpsilonClosure(
-    const std::vector<Element> &input_subset,
-    std::vector<Element> *output_subset) {
-  ecinfo_.resize(0);
-  size_t size = input_subset.size();
-  // find whether input fst is known to be sorted in input label.
-  bool sorted =
-      ((ifst_->Properties(kILabelSorted, false) & kILabelSorted) != 0);
-
-  // size is still the input_subset.size()
-  for (size_t i = 0; i < size; i++) {
-    ExpandOneElement(input_subset[i], sorted, input_subset[i].weight, true);
-  }
-
-  size_t s = queue_2_.size();
-  if (s == 0) {
-    *output_subset = input_subset;
-    return;
-  } else {
-    // queue_2 not empty. Need to create the vector<info>
-    for (size_t i = 0; i < size; i++) {
-      // the weight has not been processed yet,
-      // so put all of them in the "weight_to_process"
-      ecinfo_.push_back(
-          EpsilonClosureInfo(input_subset[i], input_subset[i].weight, false));
-      ecinfo_.back().element.weight = Weight::Zero();  // clear the weight
-
-      if (id_to_index_.size() < input_subset[i].state + 1) {
-        id_to_index_.resize(2 * input_subset[i].state + 1, -1);
-      }
-      id_to_index_[input_subset[i].state] = ecinfo_.size() - 1;
-    }
-  }
-
-  {
-    Element elem;
-    elem.weight = Weight::Zero();
-    for (size_t i = 0; i < s; i++) {
-      elem.state = queue_2_[i].state;
-      elem.string = queue_2_[i].string;
-      AddOneElement(elem, queue_2_[i].weight);
-    }
-    queue_2_.resize(0);
-  }
-
-  int counter = 0;  // relates to max-states option, used for test.
-  while (!queue_.empty()) {
-    InputStateId id = queue_.front();
-
-    // no need to check validity of the index
-    // since anything in the queue we are sure they're in the "virtual set"
-    int index = id_to_index_[id];
-    EpsilonClosureInfo &info = ecinfo_[index];
-    Element &elem = info.element;
-    Weight unprocessed_weight = info.weight_to_process;
-
-    elem.weight = Plus(elem.weight, unprocessed_weight);
-    info.weight_to_process = Weight::Zero();
-
-    info.in_queue = false;
-    queue_.pop_front();
-
-    if (max_states_ > 0 && counter++ > max_states_) {
-      KALDIFST_ERR << "Determinization aborted since looped more than "
-                   << max_states_ << " times during epsilon closure";
-    }
-
-    // generally we need to be careful about iterator-invalidation problem
-    // here we pass a reference (elem), which could be an issue.
-    // In the beginning of ExpandOneElement, we make a copy of elem.string
-    // to avoid that issue
-    ExpandOneElement(elem, sorted, unprocessed_weight);
-  }
-
-  {
-    // this sorting is based on StateId
-    sort(ecinfo_.begin(), ecinfo_.end());
-
-    output_subset->clear();
-
-    size = ecinfo_.size();
-    output_subset->reserve(size);
-    for (size_t i = 0; i < size; i++) {
-      EpsilonClosureInfo &info = ecinfo_[i];
-      if (info.weight_to_process != Weight::Zero()) {
-        info.element.weight = Plus(info.element.weight, info.weight_to_process);
-      }
-      output_subset->push_back(info.element);
-    }
-  }
-}
-
-template <class F>
-void DeterminizerStar<F>::EpsilonClosure::AddOneElement(
-    const Element &elem, const Weight &unprocessed_weight) {
-  // first we try to find the element info in the ecinfo_ vector
-  int index = -1;
-  if (elem.state < id_to_index_.size()) {
-    index = id_to_index_[elem.state];
-  }
-  if (index != -1) {
-    if (index >= ecinfo_.size()) {
-      index = -1;
-      // since ecinfo_ might store outdated information, we need to check
-    } else if (ecinfo_[index].element.state != elem.state) {
-      index = -1;
-    }
-  }
-
-  if (index == -1) {
-    // was no such StateId: insert and add to queue.
-    ecinfo_.push_back(EpsilonClosureInfo(elem, unprocessed_weight, true));
-    size_t size = id_to_index_.size();
-    if (size < elem.state + 1) {
-      // double the size to reduce memory operations
-      id_to_index_.resize(2 * elem.state + 1, -1);
-    }
-    id_to_index_[elem.state] = ecinfo_.size() - 1;
-    queue_.push_back(elem.state);
-
-  } else {  // one is already there.  Add weights.
-    EpsilonClosureInfo &info = ecinfo_[index];
-    if (info.element.string != elem.string) {
-      // Non-functional FST.
-      std::ostringstream ss;
-      ss << "FST was not functional -> not determinizable.";
-      {  // Print some debugging information.  Can be helpful to debug
-        // the inputs when FSTs are mysteriously non-functional.
-        std::vector<Label> tmp_seq;
-        repository_->SeqOfId(info.element.string, &tmp_seq);
-        ss << "\nFirst string:";
-        for (size_t i = 0; i < tmp_seq.size(); i++) ss << ' ' << tmp_seq[i];
-        ss << "\nSecond string:";
-        repository_->SeqOfId(elem.string, &tmp_seq);
-        for (size_t i = 0; i < tmp_seq.size(); i++) ss << ' ' << tmp_seq[i];
-      }
-      KALDIFST_ERR << ss.str();
-    }
-
-    info.weight_to_process = Plus(info.weight_to_process, unprocessed_weight);
-
-    if (!info.in_queue) {
-      // this is because the code in "else" below: the
-      // iter->second.weight_to_process might not be Zero()
-      Weight weight = Plus(info.element.weight, info.weight_to_process);
-
-      // What is done below is, we propagate the weight (by adding them
-      // to the queue only when the change is big enough;
-      // otherwise we just store the weight, until before returning
-      // we add the element.weight and weight_to_process together
-      if (!ApproxEqual(weight, info.element.weight, delta_)) {
-        // add extra part of weight to queue.
-        info.in_queue = true;
-        queue_.push_back(elem.state);
-      }
-    }
-  }
-}
-
-template <class F>
-void DeterminizerStar<F>::EpsilonClosure::ExpandOneElement(
-    const Element &elem, bool sorted, const Weight &unprocessed_weight,
-    bool save_to_queue_2) {
-  StringId str =
-      elem.string;  // copy it here because there is an iterator-
-                    // - invalidation problem (it really happens for some FSTs)
-
-  // now we are going to propagate the "unprocessed_weight"
-  for (ArcIterator<Fst<Arc>> aiter(*ifst_, elem.state); !aiter.Done();
-       aiter.Next()) {
-    const Arc &arc = aiter.Value();
-    if (sorted && arc.ilabel > 0) {
-      break;
-      // Break from the loop: due to sorting there will be no
-      // more transitions with epsilons as input labels.
-    }
-    if (arc.ilabel != 0) {
-      continue;  // we only process epsilons here
-    }
-    Element next_elem;
-    next_elem.state = arc.nextstate;
-    next_elem.weight = Weight::Zero();
-    Weight next_unprocessed_weight = Times(unprocessed_weight, arc.weight);
-
-    // now must append strings
-    if (arc.olabel == 0) {
-      next_elem.string = str;
-    } else {
-      std::vector<Label> seq;
-      repository_->SeqOfId(str, &seq);
-      if (arc.olabel != 0) seq.push_back(arc.olabel);
-      next_elem.string = repository_->IdOfSeq(seq);
-    }
-    if (save_to_queue_2) {
-      next_elem.weight = next_unprocessed_weight;
-      queue_2_.push_back(next_elem);
-    } else {
-      AddOneElement(next_elem, next_unprocessed_weight);
-    }
-  }
-}
-
-template <class F>
-void DeterminizerStar<F>::Output(MutableFst<Arc> *ofst, bool destroy) {
-  assert(determinized_);
-  if (destroy) determinized_ = false;
-  // Outputs to standard fst.
-  OutputStateId num_states = static_cast<OutputStateId>(output_arcs_.size());
-  if (destroy) FreeMostMemory();
-  ofst->DeleteStates();
-  if (num_states == 0) {
-    ofst->SetStart(kNoStateId);
-    return;
-  }
-  // Add basic states-- but will add extra ones to account for strings on
-  // output.
-  for (OutputStateId s = 0; s < num_states; s++) {
-    OutputStateId news = ofst->AddState();
-    assert(news == s);
-  }
-  ofst->SetStart(0);
-  for (OutputStateId this_state = 0; this_state < num_states; this_state++) {
-    std::vector<TempArc> &this_vec(output_arcs_[this_state]);
-
-    typename std::vector<TempArc>::const_iterator iter = this_vec.begin(),
-                                                  end = this_vec.end();
-    for (; iter != end; ++iter) {
-      const TempArc &temp_arc(*iter);
-      std::vector<Label> seq;
-      repository_.SeqOfId(temp_arc.ostring, &seq);
-      if (temp_arc.nextstate == kNoStateId) {  // Really a final weight.
-        // Make a sequence of states going to a final state, with the strings as
-        // labels. Put the weight on the first arc.
-        OutputStateId cur_state = this_state;
-        for (size_t i = 0; i < seq.size(); i++) {
-          OutputStateId next_state = ofst->AddState();
-          Arc arc;
-          arc.nextstate = next_state;
-          arc.weight = (i == 0 ? temp_arc.weight : Weight::One());
-          arc.ilabel = 0;  // epsilon.
-          arc.olabel = seq[i];
-          ofst->AddArc(cur_state, arc);
-          cur_state = next_state;
-        }
-        ofst->SetFinal(cur_state,
-                       (seq.size() == 0 ? temp_arc.weight : Weight::One()));
-      } else {  // Really an arc.
-        OutputStateId cur_state = this_state;
-        // Have to be careful with this integer comparison (i+1 < seq.size())
-        // because unsigned. i < seq.size()-1 could fail for zero-length
-        // sequences.
-        for (size_t i = 0; i + 1 < seq.size(); i++) {
-          // for all but the last element of seq, create new state.
-          OutputStateId next_state = ofst->AddState();
-          Arc arc;
-          arc.nextstate = next_state;
-          arc.weight = (i == 0 ? temp_arc.weight : Weight::One());
-          arc.ilabel = (i == 0 ? temp_arc.ilabel
-                               : 0);  // put ilabel on first element of seq.
-          arc.olabel = seq[i];
-          ofst->AddArc(cur_state, arc);
-          cur_state = next_state;
-        }
-        // Add the final arc in the sequence.
-        Arc arc;
-        arc.nextstate = temp_arc.nextstate;
-        arc.weight = (seq.size() <= 1 ? temp_arc.weight : Weight::One());
-        arc.ilabel = (seq.size() <= 1 ? temp_arc.ilabel : 0);
-        arc.olabel = (seq.size() > 0 ? seq.back() : 0);
-        ofst->AddArc(cur_state, arc);
-      }
-    }
-    // Free up memory.  Do this inside the loop as ofst is also allocating
-    // memory
-    if (destroy) {
-      std::vector<TempArc> temp;
-      temp.swap(this_vec);
-    }
-  }
-  if (destroy) {
-    std::vector<std::vector<TempArc>> temp;
-    temp.swap(output_arcs_);
-    repository_.Destroy();
-  }
-}
-
-template <class F>
-void DeterminizerStar<F>::ProcessTransition(OutputStateId state, Label ilabel,
-                                            std::vector<Element> *subset) {
-  // At input, "subset" may contain duplicates for a given dest state (but in
-  // sorted order).  This function removes duplicates from "subset", normalizes
-  // it, and adds a transition to the dest. state (possibly affecting Q_ and
-  // hash_, if state did not exist).
-
-  typedef typename std::vector<Element>::iterator IterType;
-  {  // This block makes the subset have one unique Element per state, adding
-     // the weights.
-    IterType cur_in = subset->begin(), cur_out = cur_in, end = subset->end();
-    size_t num_out = 0;
-    // Merge elements with same state-id
-    while (cur_in != end) {  // while we have more elements to process.
-      // At this point, cur_out points to location of next place we want to put
-      // an element, cur_in points to location of next element we want to
-      // process.
-      if (cur_in != cur_out) *cur_out = *cur_in;
-      cur_in++;
-      while (cur_in != end &&
-             cur_in->state == cur_out->state) {  // merge elements.
-        if (cur_in->string != cur_out->string) {
-          KALDIFST_ERR << "FST was not functional -> not determinizable";
-        }
-        cur_out->weight = Plus(cur_out->weight, cur_in->weight);
-        cur_in++;
-      }
-      cur_out++;
-      num_out++;
-    }
-    subset->resize(num_out);
-  }
-
-  StringId common_str;
-  Weight tot_weight;
-  {  // This block computes common_str and tot_weight (essentially: the common
-     // divisor)
-    // and removes them from the elements.
-    std::vector<Label> seq;
-
-    IterType begin = subset->begin(), iter, end = subset->end();
-    {  // This block computes "seq", which is the common prefix, and
-       // "common_str",
-      // which is the StringId version of "seq".
-      std::vector<Label> tmp_seq;
-      for (iter = begin; iter != end; ++iter) {
-        if (iter == begin) {
-          repository_.SeqOfId(iter->string, &seq);
-        } else {
-          repository_.SeqOfId(iter->string, &tmp_seq);
-          if (tmp_seq.size() < seq.size())
-            seq.resize(tmp_seq.size());  // size of shortest one.
-          for (size_t i = 0; i < seq.size();
-               i++)  // seq.size() is the shorter one at this point.
-            if (tmp_seq[i] != seq[i]) seq.resize(i);
-        }
-        if (seq.size() == 0) break;  // will not get any prefix.
-      }
-      common_str = repository_.IdOfSeq(seq);
-    }
-
-    {  // This block computes "tot_weight".
-      iter = begin;
-      tot_weight = iter->weight;
-      for (++iter; iter != end; ++iter)
-        tot_weight = Plus(tot_weight, iter->weight);
-    }
-
-    // Now divide out common stuff from elements.
-    size_t prefix_len = seq.size();
-    for (iter = begin; iter != end; ++iter) {
-      iter->weight = Divide(iter->weight, tot_weight);
-      iter->string = repository_.RemovePrefix(iter->string, prefix_len);
-    }
-  }
-
-  // Now add an arc to the state that the subset represents.
-  // We may create a new state id for this (in SubsetToStateId).
-  TempArc temp_arc;
-  temp_arc.ilabel = ilabel;
-  temp_arc.nextstate =
-      SubsetToStateId(*subset);  // may or may not really add the subset.
-  temp_arc.ostring = common_str;
-  temp_arc.weight = tot_weight;
-  output_arcs_[state].push_back(temp_arc);  // record the arc.
-}
-
-template <class F>
-void DeterminizerStar<F>::Debug() {
-  // this function called if you send a signal
-  // SIGUSR1 to the process (and it's caught by the handler in
-  // fstdeterminizestar).  It prints out some traceback
-  // info and exits.
-
-  KALDIFST_WARN << "Debug function called (probably SIGUSR1 caught)";
-  // free up memory from the hash as we need a little memory
-  {
-    SubsetHash hash_tmp;
-    std::swap(hash_tmp, hash_);
-  }
-
-  if (output_arcs_.size() <= 2) {
-    KALDIFST_ERR << "Nothing to trace back";
-  }
-  size_t max_state = output_arcs_.size() - 2;  // don't take the last
-  // one as we might be halfway into constructing it.
-
-  std::vector<OutputStateId> predecessor(max_state + 1, kNoStateId);
-  for (size_t i = 0; i < max_state; i++) {
-    for (size_t j = 0; j < output_arcs_[i].size(); j++) {
-      OutputStateId nextstate = output_arcs_[i][j].nextstate;
-      // Always find an earlier-numbered predecessor; this
-      // is always possible because of the way the algorithm
-      // works.
-      if (nextstate <= max_state && nextstate > i) predecessor[nextstate] = i;
-    }
-  }
-  std::vector<std::pair<Label, StringId>> traceback;
-  // 'traceback' is a pair of (ilabel, olabel-seq).
-  OutputStateId cur_state = max_state;  // A recently constructed state.
-
-  while (cur_state != 0 && cur_state != kNoStateId) {
-    OutputStateId last_state = predecessor[cur_state];
-    std::pair<Label, StringId> p;
-    size_t i;
-    for (i = 0; i < output_arcs_[last_state].size(); i++) {
-      if (output_arcs_[last_state][i].nextstate == cur_state) {
-        p.first = output_arcs_[last_state][i].ilabel;
-        p.second = output_arcs_[last_state][i].ostring;
-        traceback.push_back(p);
-        break;
-      }
-    }
-    KALDIFST_ASSERT(i != output_arcs_[last_state].size());  // Or fell off loop.
-    cur_state = last_state;
-  }
-  if (cur_state == kNoStateId)
-    KALDIFST_WARN << "Traceback did not reach start state "
-                  << "(possibly debug-code error)";
-
-  std::stringstream ss;
-  ss << "Traceback follows in format "
-     << "ilabel (olabel olabel) ilabel (olabel) ... :";
-  for (ssize_t i = traceback.size() - 1; i >= 0; i--) {
-    ss << ' ' << traceback[i].first << " ( ";
-    std::vector<Label> seq;
-    repository_.SeqOfId(traceback[i].second, &seq);
-    for (size_t j = 0; j < seq.size(); j++) ss << seq[j] << ' ';
-    ss << ')';
-  }
-  KALDIFST_ERR << ss.str();
-}
-
-template <class F>
-bool DeterminizeStar(F &ifst,  // NOLINT
-                     MutableFst<typename F::Arc> *ofst, float delta,
-                     bool *debug_ptr, int max_states, bool allow_partial) {
-  ofst->SetOutputSymbols(ifst.OutputSymbols());
-  ofst->SetInputSymbols(ifst.InputSymbols());
-  DeterminizerStar<F> det(ifst, delta, max_states, allow_partial);
-  det.Determinize(debug_ptr);
-  det.Output(ofst);
-  return det.IsPartial();
-}
-
-}  // namespace fst
-
-#endif  // KALDIFST_CSRC_DETERMINIZE_STAR_INL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/determinize-star.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/determinize-star.h
deleted file mode 100644
index cbd60f96..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/determinize-star.h
+++ /dev/null
@@ -1,47 +0,0 @@
-// kaldifst/csrc/determinize-star.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/determinize-star.h
-
-// Copyright 2009-2011  Microsoft Corporation
-//                2014  Guoguo Chen
-//                2015  Hainan Xu
-
-#ifndef KALDIFST_CSRC_DETERMINIZE_STAR_H_
-#define KALDIFST_CSRC_DETERMINIZE_STAR_H_
-
-#include "fst/fst-decl.h"
-#include "fst/fst.h"
-
-namespace fst {
-
-// This algorithm will be slightly faster if you sort the input fst on input
-// label.
-
-/**
-    This function implements the normal version of DeterminizeStar, in which the
-    output strings are represented using sequences of arcs, where all but the
-    first one has an epsilon on the input side.  The debug_ptr argument is an
-    optional pointer to a bool that, if it becomes true while the algorithm is
-    executing, the algorithm will print a traceback and terminate (used in
-    fstdeterminizestar.cc debug non-terminating determinization).
-    If max_states is positive, it will stop determinization and throw an
-    exception as soon as the max-states is reached. This can be useful in test.
-    If allow_partial is true, the algorithm will output partial results when the
-    specified max_states is reached (when larger than zero), instead of throwing
-    out an error.
-
-    Caution, the return status is un-intuitive: this function will return false
-    if determinization completed normally, and true if it was stopped early by
-    reaching the 'max-states' limit, and a partial FST was generated.
-*/
-template <class F>
-bool DeterminizeStar(F &ifst,  // NOLINT
-                     MutableFst<typename F::Arc> *ofst, float delta = kDelta,
-                     bool *debug_ptr = NULL, int max_states = -1,
-                     bool allow_partial = false);
-}  // namespace fst
-
-#include "kaldifst/csrc/determinize-star-inl.h"
-
-#endif  // KALDIFST_CSRC_DETERMINIZE_STAR_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/fstext-utils-inl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/fstext-utils-inl.h
deleted file mode 100644
index 59701479..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/fstext-utils-inl.h
+++ /dev/null
@@ -1,612 +0,0 @@
-// kaldifst/csrc/fstext-utils-inl.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/fstext-utils-inl.h
-
-// Copyright 2009-2012  Microsoft Corporation  Johns Hopkins University
-//                      (Author: Daniel Povey)
-//
-//           2014  Telepoint Global Hosting Service, LLC.
-//                 (Author: David Snyder)
-
-#ifndef KALDIFST_CSRC_FSTEXT_UTILS_INL_H_
-#define KALDIFST_CSRC_FSTEXT_UTILS_INL_H_
-
-#include <algorithm>
-#include <map>
-#include <set>
-#include <string>
-#include <type_traits>
-#include <unordered_map>
-#include <unordered_set>
-#include <utility>
-#include <vector>
-
-#include "kaldifst/csrc/const-integer-set.h"
-#include "kaldifst/csrc/determinize-star.h"
-#include "kaldifst/csrc/kaldi-math.h"
-#include "kaldifst/csrc/stl-utils.h"
-
-namespace fst {
-
-inline bool DeterminizeStarInLog(VectorFst<StdArc> *fst, float delta,
-                                 bool *debug_ptr, int max_states) {
-  // DeterminizeStarInLog determinizes 'fst' in the log semiring, using
-  // the DeterminizeStar algorithm (which also removes epsilons).
-
-  ArcSort(fst, ILabelCompare<StdArc>());  // helps DeterminizeStar to be faster.
-  VectorFst<LogArc> *fst_log =
-      new VectorFst<LogArc>;  // Want to determinize in log semiring.
-  Cast(*fst, fst_log);
-  VectorFst<StdArc> tmp;
-  *fst = tmp;  // make fst empty to free up memory. [actually may make no
-               // difference..]
-  VectorFst<LogArc> *fst_det_log = new VectorFst<LogArc>;
-  bool status =
-      DeterminizeStar(*fst_log, fst_det_log, delta, debug_ptr, max_states);
-  Cast(*fst_det_log, fst);
-  delete fst_log;
-  delete fst_det_log;
-  return status;
-}
-
-template <class Arc, class I>
-void MakeLinearAcceptor(const std::vector<I> &labels, MutableFst<Arc> *ofst) {
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Weight Weight;
-
-  ofst->DeleteStates();
-  StateId cur_state = ofst->AddState();
-  ofst->SetStart(cur_state);
-  for (size_t i = 0; i < labels.size(); i++) {
-    StateId next_state = ofst->AddState();
-    Arc arc(labels[i], labels[i], Weight::One(), next_state);
-    ofst->AddArc(cur_state, arc);
-    cur_state = next_state;
-  }
-  ofst->SetFinal(cur_state, Weight::One());
-}
-
-template <class Arc>
-VectorFst<Arc> *MakeLoopFst(const std::vector<const ExpandedFst<Arc> *> &fsts) {
-  typedef typename Arc::Weight Weight;
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Label Label;
-
-  VectorFst<Arc> *ans = new VectorFst<Arc>;
-  StateId loop_state = ans->AddState();  // = 0.
-  ans->SetStart(loop_state);
-  ans->SetFinal(loop_state, Weight::One());
-
-  // "cache" is used as an optimization when some of the pointers in "fsts"
-  // may have the same value.
-  unordered_map<const ExpandedFst<Arc> *, Arc> cache;
-
-  for (Label i = 0; i < static_cast<Label>(fsts.size()); i++) {
-    const ExpandedFst<Arc> *fst = fsts[i];
-    if (fst == nullptr) continue;
-    {  // optimization with cache: helpful if some members of "fsts" may
-      // contain the same pointer value (e.g. in GetHTransducer).
-      auto iter = cache.find(fst);
-      if (iter != cache.end()) {
-        Arc arc = iter->second;
-        arc.olabel = i;
-        ans->AddArc(0, arc);
-        continue;
-      }
-    }
-
-    KALDIFST_ASSERT(fst->Properties(kAcceptor, true) ==
-                    kAcceptor);  // expect acceptor.
-
-    StateId fst_num_states = fst->NumStates();
-    StateId fst_start_state = fst->Start();
-
-    if (fst_start_state == kNoStateId) continue;  // empty fst.
-
-    bool share_start_state =
-        fst->Properties(kInitialAcyclic, true) == kInitialAcyclic &&
-        fst->NumArcs(fst_start_state) == 1 &&
-        fst->Final(fst_start_state) == Weight::Zero();
-
-    std::vector<StateId> state_map(fst_num_states);  // fst state -> ans state
-    for (StateId s = 0; s < fst_num_states; s++) {
-      if (s == fst_start_state && share_start_state)
-        state_map[s] = loop_state;
-      else
-        state_map[s] = ans->AddState();
-    }
-    if (!share_start_state) {
-      Arc arc(0, i, Weight::One(), state_map[fst_start_state]);
-      cache[fst] = arc;
-      ans->AddArc(0, arc);
-    }
-    for (StateId s = 0; s < fst_num_states; s++) {
-      // Add arcs out of state s.
-      for (ArcIterator<ExpandedFst<Arc>> aiter(*fst, s); !aiter.Done();
-           aiter.Next()) {
-        const Arc &arc = aiter.Value();
-        Label olabel = (s == fst_start_state && share_start_state ? i : 0);
-        Arc newarc(arc.ilabel, olabel, arc.weight, state_map[arc.nextstate]);
-        ans->AddArc(state_map[s], newarc);
-        if (s == fst_start_state && share_start_state) cache[fst] = newarc;
-      }
-      if (fst->Final(s) != Weight::Zero()) {
-        KALDIFST_ASSERT(!(s == fst_start_state && share_start_state));
-        ans->AddArc(state_map[s], Arc(0, 0, fst->Final(s), loop_state));
-      }
-    }
-  }
-  return ans;
-}
-
-template <class Arc>
-void ApplyProbabilityScale(float scale, MutableFst<Arc> *fst) {
-  typedef typename Arc::Weight Weight;
-  typedef typename Arc::StateId StateId;
-  for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
-       siter.Next()) {
-    StateId s = siter.Value();
-    for (MutableArcIterator<MutableFst<Arc>> aiter(fst, s); !aiter.Done();
-         aiter.Next()) {
-      Arc arc = aiter.Value();
-      arc.weight = Weight(arc.weight.Value() * scale);
-      aiter.SetValue(arc);
-    }
-    if (fst->Final(s) != Weight::Zero())
-      fst->SetFinal(s, Weight(fst->Final(s).Value() * scale));
-  }
-}
-
-template <class Arc, class I>
-class RemoveSomeInputSymbolsMapper {
- public:
-  Arc operator()(const Arc &arc_in) {
-    Arc ans = arc_in;
-    if (to_remove_set_.count(ans.ilabel) != 0)
-      ans.ilabel = 0;  // remove this symbol
-    return ans;
-  }
-  MapFinalAction FinalAction() { return MAP_NO_SUPERFINAL; }
-  MapSymbolsAction InputSymbolsAction() { return MAP_CLEAR_SYMBOLS; }
-  MapSymbolsAction OutputSymbolsAction() { return MAP_COPY_SYMBOLS; }
-  uint64_t Properties(uint64_t props) const {
-    // remove the following as we don't know now if any of them are true.
-    uint64_t to_remove = kAcceptor | kNotAcceptor | kIDeterministic |
-                         kNonIDeterministic | kNoEpsilons | kNoIEpsilons |
-                         kILabelSorted | kNotILabelSorted;
-    return props & ~to_remove;
-  }
-  explicit RemoveSomeInputSymbolsMapper(const std::vector<I> &to_remove)
-      : to_remove_set_(to_remove) {
-    static_assert(std::is_integral<I>::value, "");
-    assert(to_remove_set_.count(0) == 0);  // makes no sense to remove epsilon.
-  }
-
- private:
-  kaldifst::ConstIntegerSet<I> to_remove_set_;
-};
-
-template <class Arc, class I>
-void RemoveSomeInputSymbols(const std::vector<I> &to_remove,
-                            MutableFst<Arc> *fst) {
-  static_assert(std::is_integral<I>::value, "");
-  RemoveSomeInputSymbolsMapper<Arc, I> mapper(to_remove);
-  Map(fst, mapper);
-}
-
-template <class T>
-struct IdentityFunction {
-  typedef T Arg;
-  typedef T Result;
-  T operator()(const T &t) const { return t; }
-};
-
-template <class Arc, class F>
-void MakePrecedingInputSymbolsSameClass(bool start_is_epsilon,
-                                        MutableFst<Arc> *fst, const F &f) {
-  typedef typename F::Result ClassType;
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Weight Weight;
-  std::vector<ClassType> classes;
-  ClassType noClass = f(kNoLabel);
-  ClassType epsClass = f(0);
-  if (start_is_epsilon) {  // treat having-start-state as epsilon in-transition.
-    StateId start_state = fst->Start();
-    if (start_state < 0 || start_state == kNoStateId)  // empty FST.
-      return;
-    classes.resize(start_state + 1, noClass);
-    classes[start_state] = epsClass;
-  }
-
-  // Find bad states (states with multiple input-symbols into them).
-  std::set<StateId> bad_states;  // states that we need to change.
-  for (StateIterator<Fst<Arc>> siter(*fst); !siter.Done(); siter.Next()) {
-    StateId s = siter.Value();
-    for (ArcIterator<Fst<Arc>> aiter(*fst, s); !aiter.Done(); aiter.Next()) {
-      const Arc &arc = aiter.Value();
-      if (classes.size() <= static_cast<size_t>(arc.nextstate))
-        classes.resize(arc.nextstate + 1, noClass);
-      if (classes[arc.nextstate] == noClass)
-        classes[arc.nextstate] = f(arc.ilabel);
-      else if (classes[arc.nextstate] != f(arc.ilabel))
-        bad_states.insert(arc.nextstate);
-    }
-  }
-  if (bad_states.empty()) return;  // Nothing to do.
-  kaldifst::ConstIntegerSet<StateId> bad_states_ciset(
-      bad_states);  // faster lookup.
-
-  // Work out list of arcs we have to change as (state, arc-offset).
-  // Can't do the actual changes in this pass, since we have to add new
-  // states which invalidates the iterators.
-  std::vector<std::pair<StateId, size_t>> arcs_to_change;
-  for (StateIterator<Fst<Arc>> siter(*fst); !siter.Done(); siter.Next()) {
-    StateId s = siter.Value();
-    for (ArcIterator<Fst<Arc>> aiter(*fst, s); !aiter.Done(); aiter.Next()) {
-      const Arc &arc = aiter.Value();
-      if (arc.ilabel != 0 && bad_states_ciset.count(arc.nextstate) != 0)
-        arcs_to_change.push_back(std::make_pair(s, aiter.Position()));
-    }
-  }
-  KALDIFST_ASSERT(!arcs_to_change.empty());  // since !bad_states.empty().
-
-  std::map<std::pair<StateId, ClassType>, StateId> state_map;
-  // state_map is a map from (bad-state, input-symbol-class) to dummy-state.
-
-  for (size_t i = 0; i < arcs_to_change.size(); i++) {
-    StateId s = arcs_to_change[i].first;
-    ArcIterator<MutableFst<Arc>> aiter(*fst, s);
-    aiter.Seek(arcs_to_change[i].second);
-    Arc arc = aiter.Value();
-
-    // Transition is non-eps transition to "bad" state.  Introduce new state (or
-    // find existing one).
-    std::pair<StateId, ClassType> p(arc.nextstate, f(arc.ilabel));
-    if (state_map.count(p) == 0) {
-      StateId newstate = state_map[p] = fst->AddState();
-      fst->AddArc(newstate, Arc(0, 0, Weight::One(), arc.nextstate));
-    }
-    StateId dst_state = state_map[p];
-    arc.nextstate = dst_state;
-
-    // Initialize the MutableArcIterator only now, as the call to NewState()
-    // may have invalidated the first arc iterator.
-    MutableArcIterator<MutableFst<Arc>> maiter(fst, s);
-    maiter.Seek(arcs_to_change[i].second);
-    maiter.SetValue(arc);
-  }
-}
-
-template <class Arc>
-void MakeFollowingInputSymbolsSame(bool end_is_epsilon, MutableFst<Arc> *fst) {
-  IdentityFunction<typename Arc::Label> f;
-  MakeFollowingInputSymbolsSameClass(end_is_epsilon, fst, f);
-}
-
-template <class Arc, class F>
-void MakeFollowingInputSymbolsSameClass(bool end_is_epsilon,
-                                        MutableFst<Arc> *fst, const F &f) {
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Weight Weight;
-  typedef typename F::Result ClassType;
-  std::vector<StateId> bad_states;
-  ClassType noClass = f(kNoLabel);
-  ClassType epsClass = f(0);
-  for (StateIterator<Fst<Arc>> siter(*fst); !siter.Done(); siter.Next()) {
-    StateId s = siter.Value();
-    ClassType c = noClass;
-    bool bad = false;
-    for (ArcIterator<Fst<Arc>> aiter(*fst, s); !aiter.Done(); aiter.Next()) {
-      const Arc &arc = aiter.Value();
-      if (c == noClass) {
-        c = f(arc.ilabel);
-      } else if (c != f(arc.ilabel)) {
-        bad = true;
-        break;
-      }
-    }
-    if (end_is_epsilon && c != noClass && c != epsClass &&
-        fst->Final(s) != Weight::Zero())
-      bad = true;
-    if (bad) bad_states.push_back(s);
-  }
-  std::vector<Arc> my_arcs;
-  for (size_t i = 0; i < bad_states.size(); i++) {
-    StateId s = bad_states[i];
-    my_arcs.clear();
-    for (ArcIterator<MutableFst<Arc>> aiter(*fst, s); !aiter.Done();
-         aiter.Next())
-      my_arcs.push_back(aiter.Value());
-
-    for (size_t j = 0; j < my_arcs.size(); j++) {
-      Arc &arc = my_arcs[j];
-      if (arc.ilabel != 0) {
-        StateId newstate = fst->AddState();
-        // Create a new state for each non-eps arc in original FST, out of each
-        // bad state. Not as optimal as it could be, but does avoid some
-        // complicated weight-pushing issues in which, to maintain
-        // stochasticity, we would have to know which semiring we want to
-        // maintain stochasticity in.
-        fst->AddArc(newstate, Arc(arc.ilabel, 0, Weight::One(), arc.nextstate));
-        MutableArcIterator<MutableFst<Arc>> maiter(fst, s);
-        maiter.Seek(j);
-        maiter.SetValue(Arc(0, arc.olabel, arc.weight, newstate));
-      }
-    }
-  }
-}
-
-// return arc-offset of self-loop with ilabel (or -1 if none exists).
-// if more than one such self-loop, pick first one.
-template <class Arc>
-ssize_t FindSelfLoopWithILabel(const Fst<Arc> &fst, typename Arc::StateId s) {
-  for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next())
-    if (aiter.Value().nextstate == s && aiter.Value().ilabel != 0)
-      return static_cast<ssize_t>(aiter.Position());
-  return static_cast<ssize_t>(-1);
-}
-
-template <class Arc>
-bool EqualAlign(const Fst<Arc> &ifst, typename Arc::StateId length,
-                int rand_seed, MutableFst<Arc> *ofst, int num_retries) {
-  srand(rand_seed);
-  KALDIFST_ASSERT(ofst->NumStates() == 0);  // make sure ofst empty.
-  // make sure all states can reach final-state (or this algorithm may enter
-  // infinite loop.
-  KALDIFST_ASSERT(ifst.Properties(kCoAccessible, true) == kCoAccessible);
-
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Weight Weight;
-
-  if (ifst.Start() == kNoStateId) {
-    KALDIFST_WARN << "Empty input fst.";
-    return false;
-  }
-  // First select path through ifst.
-  std::vector<StateId> path;
-  std::vector<size_t> arc_offsets;  // arc taken out of each state.
-  std::vector<int> nof_ilabels;
-
-  StateId num_ilabels = 0;
-  int retry_no = 0;
-
-  // Under normal circumstances, this will be one-pass-only process
-  // Multiple tries might be needed in special cases, typically when
-  // the number of frames is close to number of transitions from
-  // the start node to the final node. It usually happens for really
-  // short utterances
-  do {
-    num_ilabels = 0;
-    arc_offsets.clear();
-    path.clear();
-    path.push_back(ifst.Start());
-
-    while (1) {
-      // Select either an arc or final-prob.
-      StateId s = path.back();
-      size_t num_arcs = ifst.NumArcs(s);
-      size_t num_arcs_tot = num_arcs;
-      if (ifst.Final(s) != Weight::Zero()) num_arcs_tot++;
-      // kaldifst::RandInt is a bit like Rand(), but gets around situations
-      // where RAND_MAX is very small.
-      // Change this to Rand() % num_arcs_tot if compile issues arise
-      size_t arc_offset =
-          static_cast<size_t>(kaldifst::RandInt(0, num_arcs_tot - 1));
-
-      if (arc_offset < num_arcs) {  // an actual arc.
-        ArcIterator<Fst<Arc>> aiter(ifst, s);
-        aiter.Seek(arc_offset);
-        const Arc &arc = aiter.Value();
-        if (arc.nextstate == s) {
-          continue;  // don't take this self-loop arc
-        } else {
-          arc_offsets.push_back(arc_offset);
-          path.push_back(arc.nextstate);
-          if (arc.ilabel != 0) num_ilabels++;
-        }
-      } else {
-        break;  // Chose final-prob.
-      }
-    }
-
-    nof_ilabels.push_back(num_ilabels);
-  } while ((++retry_no < num_retries) && (num_ilabels > length));
-
-  if (num_ilabels > length) {
-    std::stringstream ilabel_vec;
-    std::copy(nof_ilabels.begin(), nof_ilabels.end(),
-              std::ostream_iterator<int>(ilabel_vec, ","));
-    std::string s = ilabel_vec.str();
-    s.erase(s.end() - 1);
-    KALDIFST_WARN << "EqualAlign: the randomly constructed paths lengths: "
-                  << s;
-    KALDIFST_WARN << "EqualAlign: utterance has too few frames " << length
-                  << " to align.";
-    return false;  // can't make it shorter by adding self-loops!.
-  }
-
-  StateId num_self_loops = 0;
-  std::vector<ssize_t> self_loop_offsets(path.size());
-  for (size_t i = 0; i < path.size(); i++)
-    if ((self_loop_offsets[i] = FindSelfLoopWithILabel(ifst, path[i])) !=
-        static_cast<ssize_t>(-1))
-      num_self_loops++;
-
-  if (num_self_loops == 0 && num_ilabels < length) {
-    KALDIFST_WARN << "No self-loops on chosen path; cannot match length.";
-    return false;  // no self-loops to make it longer.
-  }
-
-  StateId num_extra = length - num_ilabels;  // Number of self-loops we need.
-
-  StateId min_num_loops = 0;
-  if (num_extra != 0)
-    min_num_loops = num_extra / num_self_loops;  // prevent div by zero.
-  StateId num_with_one_more_loop = num_extra - (min_num_loops * num_self_loops);
-  KALDIFST_ASSERT(num_with_one_more_loop < num_self_loops ||
-                  num_self_loops == 0);
-
-  ofst->AddState();
-  ofst->SetStart(0);
-  StateId cur_state = 0;
-  StateId counter = 0;  // tell us when we should stop adding one more loop.
-  for (size_t i = 0; i < path.size(); i++) {
-    // First, add any self-loops that are necessary.
-    StateId num_loops = 0;
-    if (self_loop_offsets[i] != static_cast<ssize_t>(-1)) {
-      num_loops = min_num_loops + (counter < num_with_one_more_loop ? 1 : 0);
-      counter++;
-    }
-    for (StateId j = 0; j < num_loops; j++) {
-      ArcIterator<Fst<Arc>> aiter(ifst, path[i]);
-      aiter.Seek(self_loop_offsets[i]);
-      Arc arc = aiter.Value();
-      KALDIFST_ASSERT(arc.nextstate == path[i] &&
-                      arc.ilabel != 0);  // make sure self-loop with ilabel.
-      StateId next_state = ofst->AddState();
-      ofst->AddArc(cur_state,
-                   Arc(arc.ilabel, arc.olabel, arc.weight, next_state));
-      cur_state = next_state;
-    }
-    if (i + 1 < path.size()) {  // add forward transition.
-      ArcIterator<Fst<Arc>> aiter(ifst, path[i]);
-      aiter.Seek(arc_offsets[i]);
-      Arc arc = aiter.Value();
-      KALDIFST_ASSERT(arc.nextstate == path[i + 1]);
-      StateId next_state = ofst->AddState();
-      ofst->AddArc(cur_state,
-                   Arc(arc.ilabel, arc.olabel, arc.weight, next_state));
-      cur_state = next_state;
-    } else {  // add final-prob.
-      Weight weight = ifst.Final(path[i]);
-      KALDIFST_ASSERT(weight != Weight::Zero());
-      ofst->SetFinal(cur_state, weight);
-    }
-  }
-  return true;
-}
-
-template <class Arc, class I>
-bool GetLinearSymbolSequence(const Fst<Arc> &fst, std::vector<I> *isymbols_out,
-                             std::vector<I> *osymbols_out,
-                             typename Arc::Weight *tot_weight_out) {
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Weight Weight;
-
-  Weight tot_weight = Weight::One();
-  std::vector<I> ilabel_seq;
-  std::vector<I> olabel_seq;
-
-  StateId cur_state = fst.Start();
-  if (cur_state == kNoStateId) {  // empty sequence.
-    if (isymbols_out != nullptr) isymbols_out->clear();
-    if (osymbols_out != nullptr) osymbols_out->clear();
-    if (tot_weight_out != nullptr) *tot_weight_out = Weight::Zero();
-    return true;
-  }
-  while (1) {
-    Weight w = fst.Final(cur_state);
-    if (w != Weight::Zero()) {  // is final..
-      tot_weight = Times(tot_weight, w);
-      if (fst.NumArcs(cur_state) != 0) return false;
-      if (isymbols_out != nullptr) *isymbols_out = ilabel_seq;
-      if (osymbols_out != nullptr) *osymbols_out = olabel_seq;
-      if (tot_weight_out != nullptr) *tot_weight_out = tot_weight;
-      return true;
-    } else {
-      if (fst.NumArcs(cur_state) != 1) return false;
-
-      ArcIterator<Fst<Arc>> iter(fst, cur_state);  // get the only arc.
-      const Arc &arc = iter.Value();
-      tot_weight = Times(tot_weight, arc.weight);
-      if (arc.ilabel != 0) ilabel_seq.push_back(arc.ilabel);
-      if (arc.olabel != 0) olabel_seq.push_back(arc.olabel);
-      cur_state = arc.nextstate;
-    }
-  }
-}
-
-template <class Arc, class I>
-void GetInputSymbols(const Fst<Arc> &fst, bool include_eps,
-                     std::vector<I> *symbols) {
-  static_assert(std::is_integral<I>::value, "");
-  std::unordered_set<I> all_syms;
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    typename Arc::StateId s = siter.Value();
-    for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-      const Arc &arc = aiter.Value();
-      all_syms.insert(arc.ilabel);
-    }
-  }
-  // Remove epsilon, if instructed.
-  if (!include_eps && all_syms.count(0) != 0) {
-    all_syms.erase(0);
-  }
-
-  KALDIFST_ASSERT(symbols != NULL);
-  kaldifst::CopySetToVector(all_syms, symbols);
-  std::sort(symbols->begin(), symbols->end());
-}
-
-// see fstext-utils.h for comment.
-template <class Arc>
-void ConvertNbestToVector(const Fst<Arc> &fst,
-                          std::vector<VectorFst<Arc>> *fsts_out) {
-  typedef typename Arc::Weight Weight;
-  typedef typename Arc::StateId StateId;
-  fsts_out->clear();
-  StateId start_state = fst.Start();
-
-  if (start_state == kNoStateId) {
-    return;  // No output.
-  }
-
-  size_t n_arcs = fst.NumArcs(start_state);
-  bool start_is_final = (fst.Final(start_state) != Weight::Zero());
-  fsts_out->reserve(n_arcs + (start_is_final ? 1 : 0));
-
-  if (start_is_final) {
-    fsts_out->resize(fsts_out->size() + 1);
-    StateId start_state_out = fsts_out->back().AddState();
-    fsts_out->back().SetFinal(start_state_out, fst.Final(start_state));
-  }
-
-  for (ArcIterator<Fst<Arc>> start_aiter(fst, start_state); !start_aiter.Done();
-       start_aiter.Next()) {
-    fsts_out->resize(fsts_out->size() + 1);
-    VectorFst<Arc> &ofst = fsts_out->back();
-    const Arc &first_arc = start_aiter.Value();
-    StateId cur_state = start_state, cur_ostate = ofst.AddState();
-    ofst.SetStart(cur_ostate);
-    StateId next_ostate = ofst.AddState();
-    ofst.AddArc(cur_ostate, Arc(first_arc.ilabel, first_arc.olabel,
-                                first_arc.weight, next_ostate));
-    cur_state = first_arc.nextstate;
-    cur_ostate = next_ostate;
-    while (1) {
-      size_t this_n_arcs = fst.NumArcs(cur_state);
-      KALDIFST_ASSERT(this_n_arcs <= 1);  // or it violates our assumptions
-                                          // about the input.
-      if (this_n_arcs == 1) {
-        KALDIFST_ASSERT(fst.Final(cur_state) == Weight::Zero());
-        // or problem with ShortestPath.
-        ArcIterator<Fst<Arc>> aiter(fst, cur_state);
-        const Arc &arc = aiter.Value();
-        next_ostate = ofst.AddState();
-        ofst.AddArc(cur_ostate,
-                    Arc(arc.ilabel, arc.olabel, arc.weight, next_ostate));
-        cur_state = arc.nextstate;
-        cur_ostate = next_ostate;
-      } else {
-        KALDIFST_ASSERT(fst.Final(cur_state) != Weight::Zero());
-        // or problem with ShortestPath.
-        ofst.SetFinal(cur_ostate, fst.Final(cur_state));
-        break;
-      }
-    }
-  }
-}
-
-}  // namespace fst
-#endif  // KALDIFST_CSRC_FSTEXT_UTILS_INL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/fstext-utils.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/fstext-utils.h
deleted file mode 100644
index 7ffb0dff..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/fstext-utils.h
+++ /dev/null
@@ -1,154 +0,0 @@
-// kaldifst/csrc/fstext-utils.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/fstext-utils.h
-
-// Copyright 2009-2011  Microsoft Corporation
-//           2012-2013  Johns Hopkins University (Author: Daniel Povey)
-//                2013  Guoguo Chen
-//                2014  Telepoint Global Hosting Service, LLC. (Author: David
-//                Snyder)
-
-#ifndef KALDIFST_CSRC_FSTEXT_UTILS_H_
-#define KALDIFST_CSRC_FSTEXT_UTILS_H_
-
-#include <vector>
-
-#include "fst/fst.h"
-#include "fst/map.h"
-#include "fst/minimize.h"
-
-namespace fst {
-
-// Minimizes after encoding; applicable to all FSTs.  It is like what you get
-// from the Minimize() function, except it will not push the weights, or the
-// symbols.  This is better for our recipes, as we avoid ever pushing the
-// weights.  However, it will only minimize optimally if your graphs are such
-// that the symbols are as far to the left as they can go, and the weights
-// in combinable paths are the same... hard to formalize this, but it's
-// something that is satisfied by our normal FSTs.
-template <class Arc>
-void MinimizeEncoded(MutableFst<Arc> *fst, float delta = kDelta) {
-  Map(fst, QuantizeMapper<Arc>(delta));
-  EncodeMapper<Arc> encoder(kEncodeLabels | kEncodeWeights, ENCODE);
-  Encode(fst, &encoder);
-  internal::AcceptorMinimize(fst);
-  Decode(fst, encoder);
-}
-
-inline bool DeterminizeStarInLog(VectorFst<StdArc> *fst, float delta = kDelta,
-                                 bool *debug_ptr = NULL, int max_states = -1);
-
-/// Creates unweighted linear acceptor from symbol sequence.
-template <class Arc, class I>
-void MakeLinearAcceptor(const std::vector<I> &labels, MutableFst<Arc> *ofst);
-
-/// MakeLoopFst creates an FST that has a state that is both initial and
-/// final (weight == Weight::One()), and for each non-NULL pointer fsts[i],
-/// it has an arc out whose output-symbol is i and which goes to a
-/// sub-graph whose input language is equivalent to fsts[i], where the
-/// final-state becomes a transition to the loop-state.  Each fst in "fsts"
-/// should be an acceptor.  The fst MakeLoopFst returns is output-deterministic,
-/// but not output-epsilon free necessarily, and arcs are sorted on output
-/// label. Note: if some of the pointers in the input vector "fsts" have the
-/// same value, "MakeLoopFst" uses this to speed up the computation.
-
-/// Formally: suppose I is the set of indexes i such that fsts[i] != NULL.
-/// Let L[i] be the language that the acceptor fsts[i] accepts.
-/// Let the language K be the set of input-output pairs i:l such
-/// that i in I and l in L[i].  Then the FST returned by MakeLoopFst
-/// accepts the language K*, where * is the Kleene closure (CLOSURE_STAR)
-/// of K.
-
-/// We could have implemented this via a combination of "project",
-/// "concat", "union" and "closure".  But that FST would have been
-/// less well optimized and would have a lot of final-states.
-
-template <class Arc>
-VectorFst<Arc> *MakeLoopFst(const std::vector<const ExpandedFst<Arc> *> &fsts);
-
-/// ApplyProbabilityScale is applicable to FSTs in the log or tropical semiring.
-/// It multiplies the arc and final weights by "scale" [this is not the Mul
-/// operation of the semiring, it's actual multiplication, which is equivalent
-/// to taking a power in the semiring].
-template <class Arc>
-void ApplyProbabilityScale(float scale, MutableFst<Arc> *fst);
-
-/// RemoveSomeInputSymbols removes any symbol that appears in "to_remove", from
-/// the input side of the FST, replacing them with epsilon.
-template <class Arc, class I>
-void RemoveSomeInputSymbols(const std::vector<I> &to_remove,
-                            MutableFst<Arc> *fst);
-
-/// As MakePrecedingInputSymbolsSame, but takes a functor object that maps
-/// labels to classes.
-template <class Arc, class F>
-void MakePrecedingInputSymbolsSameClass(bool start_is_epsilon,
-                                        MutableFst<Arc> *fst, const F &f);
-
-/// MakeFollowingInputSymbolsSame ensures that all arcs exiting any given fst
-/// state have the same input symbol.  It does this by detecting states that
-/// have differing input symbols on arcs that exit it, and inserting, for each
-/// of the following arcs with non-epsilon input symbol, a new dummy state that
-/// has an input-epsilon link from the fst state.  The output symbol and weight
-/// stay on the link to the dummy state (in order to keep the FST
-/// output-deterministic and stochastic, if it already was). If end_is_epsilon,
-/// treat "being a final-state" like having an epsilon output link.
-template <class Arc>
-void MakeFollowingInputSymbolsSame(bool end_is_epsilon, MutableFst<Arc> *fst);
-
-/// As MakeFollowingInputSymbolsSame, but takes a functor object that maps
-/// labels to classes.
-template <class Arc, class F>
-void MakeFollowingInputSymbolsSameClass(bool end_is_epsilon,
-                                        MutableFst<Arc> *fst, const F &f);
-
-/// EqualAlign is similar to RandGen, but it generates a sequence with exactly
-/// "length" input symbols.  It returns true on success, false on failure
-/// (failure is partly random but should never happen in practice for normal
-/// speech models.) It generates a random path through the input FST, finds out
-/// which subset of the states it visits along the way have self-loops with
-/// inupt symbols on them, and outputs a path with exactly enough self-loops to
-/// have the requested number of input symbols. Note that EqualAlign does not
-/// use the probabilities on the FST.  It just uses equal probabilities in the
-/// first stage of selection (since the output will anyway not be a truly random
-/// sample from the FST). The input fst "ifst" must be connected or this may
-/// enter an infinite loop.
-template <class Arc>
-bool EqualAlign(const Fst<Arc> &ifst, typename Arc::StateId length,
-                int rand_seed, MutableFst<Arc> *ofst, int num_retries = 10);
-
-/// GetLinearSymbolSequence gets the symbol sequence from a linear FST.
-/// If the FST is not just a linear sequence, it returns false.   If it is
-/// a linear sequence (including the empty FST), it returns true.  In this
-/// case it outputs the symbol
-/// sequences as "isymbols_out" and "osymbols_out" (removing epsilons), and
-/// the total weight as "tot_weight". The total weight will be Weight::Zero()
-/// if the FST is empty.  If any of the output pointers are NULL, it does not
-/// create that output.
-template <class Arc, class I>
-bool GetLinearSymbolSequence(const Fst<Arc> &fst, std::vector<I> *isymbols_out,
-                             std::vector<I> *osymbols_out,
-                             typename Arc::Weight *tot_weight_out);
-
-/// GetInputSymbols gets the list of symbols on the input of fst
-/// (including epsilon, if include_eps == true), as a sorted, unique
-/// list.
-template <class Arc, class I>
-void GetInputSymbols(const Fst<Arc> &fst, bool include_eps,
-                     std::vector<I> *symbols);
-
-/// This function converts an FST with a special structure, which is
-/// output by the OpenFst functions ShortestPath and RandGen, and converts
-/// them into a std::vector of separate FSTs.  This special structure is that
-/// the only state that has more than one (arcs-out or final-prob) is the
-/// start state.  fsts_out is resized to the appropriate size.
-template <class Arc>
-void ConvertNbestToVector(const Fst<Arc> &fst,
-                          std::vector<VectorFst<Arc>> *fsts_out);
-
-}  // namespace fst
-
-#include "kaldifst/csrc/fstext-utils-inl.h"
-
-#endif  // KALDIFST_CSRC_FSTEXT_UTILS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/grammar-context-fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/grammar-context-fst.h
deleted file mode 100644
index bcf5a2bc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/grammar-context-fst.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// kaldifst/csrc/grammar-context-fst.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/grammar-context-fst.h
-
-// Copyright 2009-2011  Microsoft Corporation
-
-#ifndef KALDIFST_CSRC_GRAMMAR_CONTEXT_FST_H_
-#define KALDIFST_CSRC_GRAMMAR_CONTEXT_FST_H_
-
-#include "fst/fst-decl.h"
-#include "fst/fstlib.h"
-
-namespace fst {
-
-enum NonterminalValues {
-  kNontermBos = 0,      // #nonterm_bos
-  kNontermBegin = 1,    // #nonterm_begin
-  kNontermEnd = 2,      // #nonterm_end
-  kNontermReenter = 3,  // #nonterm_reenter
-  kNontermUserDefined =
-      4,  // the lowest-numbered user-defined nonterminal, e.g. #nonterm:foo
-  // kNontermMediumNumber and kNontermBigNumber come into the encoding of
-  // nonterminal-related symbols in HCLG.fst.  The only hard constraint on them
-  // is that kNontermBigNumber must be bigger than the biggest transition-id in
-  // your system, and kNontermMediumNumber must be >0.  These values were chosen
-  // for ease of human inspection of numbers encoded with them.
-  kNontermMediumNumber = 1000,
-  kNontermBigNumber = 10000000
-};
-
-// Returns the smallest multiple of 1000 that is strictly greater than
-// nonterm_phones_offset.  Used in the encoding of special symbol in HCLG;
-// they are encoded as
-//  special_symbol =
-//     kNontermBigNumber + (nonterminal * encoding_multiple) + phone_index
-inline int32_t GetEncodingMultiple(int32_t nonterm_phones_offset) {
-  int32_t medium_number = static_cast<int32_t>(kNontermMediumNumber);
-  return medium_number *
-         ((nonterm_phones_offset + medium_number) / medium_number);
-}
-
-}  // namespace fst
-
-#endif  // KALDIFST_CSRC_GRAMMAR_CONTEXT_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/io-funcs-inl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/io-funcs-inl.h
deleted file mode 100644
index dd99acbe..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/io-funcs-inl.h
+++ /dev/null
@@ -1,54 +0,0 @@
-// kaldifst/csrc/io-funcs-inl.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/base/io-funcs-inl.h
-
-// Copyright 2009-2011  Microsoft Corporation;  Saarland University;
-//                      Jan Silovsky;   Yanmin Qian;
-//                      Johns Hopkins University (Author: Daniel Povey)
-//                2016  Xiaohui Zhang
-#ifndef KALDIFST_CSRC_IO_FUNCS_INL_H_
-#define KALDIFST_CSRC_IO_FUNCS_INL_H_
-
-#include <iostream>
-
-namespace kaldifst {
-
-// Initialize an opened stream for writing by writing an optional binary
-// header and modifying the floating-point precision.
-inline void InitKaldiOutputStream(std::ostream &os, bool binary) {
-  // This does not throw exceptions (does not check for errors).
-  if (binary) {
-    os.put('\0');
-    os.put('B');
-  }
-  // Note, in non-binary mode we may at some point want to mess with
-  // the precision a bit.
-  // 7 is a bit more than the precision of float..
-  if (os.precision() < 7) os.precision(7);
-}
-
-/// Initialize an opened stream for reading by detecting the binary header and
-// setting the "binary" value appropriately.
-inline bool InitKaldiInputStream(std::istream &is, bool *binary) {
-  // Sets the 'binary' variable.
-  // Throws exception in the very unusual situation that stream
-  // starts with '\0' but not then 'B'.
-
-  if (is.peek() == '\0') {  // seems to be binary
-    is.get();
-    if (is.peek() != 'B') {
-      return false;
-    }
-    is.get();
-    *binary = true;
-    return true;
-  } else {
-    *binary = false;
-    return true;
-  }
-}
-
-}  // namespace kaldifst
-
-#endif  // KALDIFST_CSRC_IO_FUNCS_INL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/io-funcs.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/io-funcs.h
deleted file mode 100644
index b282623e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/io-funcs.h
+++ /dev/null
@@ -1,27 +0,0 @@
-// kaldifst/csrc/io-funcs.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/base/io-funcs.h
-
-// Copyright 2009-2011  Microsoft Corporation;  Saarland University;
-//                      Jan Silovsky;   Yanmin Qian
-//                2016  Xiaohui Zhang
-#ifndef KALDIFST_CSRC_IO_FUNCS_H_
-#define KALDIFST_CSRC_IO_FUNCS_H_
-
-#include "kaldifst/csrc/io-funcs-inl.h"
-
-namespace kaldifst {
-
-/// InitKaldiOutputStream initializes an opened stream for writing by writing an
-/// optional binary header and modifying the floating-point precision; it will
-/// typically not be called by users directly.
-inline void InitKaldiOutputStream(std::ostream &os, bool binary);
-
-/// InitKaldiInputStream initializes an opened stream for reading by detecting
-/// the binary header and setting the "binary" value appropriately;
-/// It will typically not be called by users directly.
-inline bool InitKaldiInputStream(std::istream &is, bool *binary);
-
-}  // namespace kaldifst
-#endif  // KALDIFST_CSRC_IO_FUNCS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-fst-io-inl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-fst-io-inl.h
deleted file mode 100644
index a12a19e3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-fst-io-inl.h
+++ /dev/null
@@ -1,202 +0,0 @@
-// kaldifst/csrc/kaldi-fst-io-inl.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/kaldi-fst-io-inl.h
-
-// Copyright 2009-2011  Microsoft Corporation
-//           2012-2015  Johns Hopkins University (Author: Daniel Povey)
-//                2013  Guoguo Chen
-
-#ifndef KALDIFST_CSRC_KALDI_FST_IO_INL_H_
-#define KALDIFST_CSRC_KALDI_FST_IO_INL_H_
-
-#include <iostream>
-#include <sstream>
-#include <string>
-#include <vector>
-
-#include "fst/fst.h"
-#include "fst/script/print.h"
-#include "kaldifst/csrc/log.h"
-#include "kaldifst/csrc/text-utils.h"
-
-namespace fst {
-
-template <class Arc>
-void WriteFstKaldi(std::ostream &os, bool binary, const VectorFst<Arc> &t) {
-  bool ok;
-  if (binary) {
-    // Binary-mode writing.
-    ok = t.Write(os, FstWriteOptions());
-  } else {
-    // Text-mode output.  Note: we expect that t.InputSymbols() and
-    // t.OutputSymbols() would always return NULL.  The corresponding input
-    // routine would not work if the FST actually had symbols attached.  Write a
-    // newline to start the FST; in a table, the first line of the FST will
-    // appear on its own line.
-    os << '\n';
-    bool acceptor = false, write_one = false;
-    fst::FstPrinter<Arc> printer(t, t.InputSymbols(), t.OutputSymbols(), NULL,
-                                 acceptor, write_one, "\t");
-    printer.Print(&os, "<unknown>");
-    if (os.fail())
-      KALDIFST_ERR << "Stream failure detected writing FST to stream";
-    // Write another newline as a terminating character.  The read routine will
-    // detect this [this is a Kaldi mechanism, not something in the original
-    // OpenFst code].
-    os << '\n';
-    ok = os.good();
-  }
-  if (!ok) {
-    KALDIFST_ERR << "Error writing FST to stream";
-  }
-}
-
-// Utility function used in ReadFstKaldi
-template <class W>
-inline bool StrToWeight(const std::string &s, bool allow_zero, W *w) {
-  std::istringstream strm(s);
-  strm >> *w;
-  if (strm.fail() || (!allow_zero && *w == W::Zero())) {
-    return false;
-  }
-  return true;
-}
-
-template <class Arc>
-void ReadFstKaldi(std::istream &is, bool binary, VectorFst<Arc> *fst) {
-  typedef typename Arc::Weight Weight;
-  typedef typename Arc::StateId StateId;
-  if (binary) {
-    // We don't have access to the filename here, so write [unknown].
-    VectorFst<Arc> *ans =
-        VectorFst<Arc>::Read(is, FstReadOptions(std::string("[unknown]")));
-    if (ans == NULL) {
-      KALDIFST_ERR << "Error reading FST from stream.";
-    }
-    *fst = *ans;  // shallow copy.
-    delete ans;
-  } else {
-    // Consume the \r on Windows, the \n that the text-form FST format starts
-    // with, and any extra spaces that might have got in there somehow.
-    while (std::isspace(is.peek()) && is.peek() != '\n') is.get();
-    if (is.peek() == '\n') {
-      is.get();  // consume the newline.
-    } else {     // saw spaces but no newline.. this is not expected.
-      KALDIFST_ERR << "Reading FST: unexpected sequence of spaces "
-                   << " at file position " << is.tellg();
-    }
-    using kaldifst::ConvertStringToInteger;
-    using kaldifst::SplitStringToIntegers;
-    using std::string;
-    using std::vector;
-    fst->DeleteStates();
-    string line;
-    size_t nline = 0;
-    string separator = FLAGS_fst_field_separator + "\r\n";
-    while (std::getline(is, line)) {
-      nline++;
-      vector<string> col;
-      // on Windows we'll write in text and read in binary mode.
-      kaldifst::SplitStringToVector(line, separator.c_str(), true, &col);
-      if (col.size() == 0) break;  // Empty line is a signal to stop, in our
-      // archive format.
-      if (col.size() > 5) {
-        KALDIFST_ERR << "Bad line in FST: " << line;
-      }
-      StateId s;
-      if (!ConvertStringToInteger(col[0], &s)) {
-        KALDIFST_ERR << "Bad line in FST: " << line;
-      }
-      while (s >= fst->NumStates()) fst->AddState();
-      if (nline == 1) fst->SetStart(s);
-
-      bool ok = true;
-      Arc arc;
-      Weight w;
-      StateId d = s;
-      switch (col.size()) {
-        case 1:
-          fst->SetFinal(s, Weight::One());
-          break;
-        case 2:
-          if (!StrToWeight(col[1], true, &w))
-            ok = false;
-          else
-            fst->SetFinal(s, w);
-          break;
-        case 3:  // 3 columns not ok for Lattice format; it's not an acceptor.
-          ok = false;
-          break;
-        case 4:
-          ok = ConvertStringToInteger(col[1], &arc.nextstate) &&
-               ConvertStringToInteger(col[2], &arc.ilabel) &&
-               ConvertStringToInteger(col[3], &arc.olabel);
-          if (ok) {
-            d = arc.nextstate;
-            arc.weight = Weight::One();
-            fst->AddArc(s, arc);
-          }
-          break;
-        case 5:
-          ok = ConvertStringToInteger(col[1], &arc.nextstate) &&
-               ConvertStringToInteger(col[2], &arc.ilabel) &&
-               ConvertStringToInteger(col[3], &arc.olabel) &&
-               StrToWeight(col[4], false, &arc.weight);
-          if (ok) {
-            d = arc.nextstate;
-            fst->AddArc(s, arc);
-          }
-          break;
-        default:
-          ok = false;
-      }
-      while (d >= fst->NumStates()) fst->AddState();
-      if (!ok) KALDIFST_ERR << "Bad line in FST: " << line;
-    }
-  }
-}
-
-template <class Arc>  // static
-bool VectorFstTplHolder<Arc>::Write(std::ostream &os, bool binary, const T &t) {
-  try {
-    WriteFstKaldi(os, binary, t);
-    return true;
-  } catch (...) {
-    return false;
-  }
-}
-
-template <class Arc>
-bool VectorFstTplHolder<Arc>::Read(std::istream &is) {
-  Clear();
-  int c = is.peek();
-  if (c == -1) {
-    KALDIFST_WARN << "End of stream detected reading Fst";
-    return false;
-  } else if (isspace(c)) {  // The text form of the FST begins
-    // with space (normally, '\n'), so this means it's text (the binary form
-    // cannot begin with space because it starts with the FST Type() which is
-    // not space).
-    try {
-      t_ = new VectorFst<Arc>();
-      ReadFstKaldi(is, false, t_);
-    } catch (...) {
-      Clear();
-      return false;
-    }
-  } else {  // reading a binary FST.
-    try {
-      t_ = new VectorFst<Arc>();
-      ReadFstKaldi(is, true, t_);
-    } catch (...) {
-      Clear();
-      return false;
-    }
-  }
-  return true;
-}
-
-}  // namespace fst
-
-#endif  // KALDIFST_CSRC_KALDI_FST_IO_INL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-fst-io.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-fst-io.h
deleted file mode 100644
index 2e6a9e14..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-fst-io.h
+++ /dev/null
@@ -1,150 +0,0 @@
-// kaldifst/csrc/kaldi-fst-io.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/kaldi-fst-io.h
-
-// Copyright 2009-2011  Microsoft Corporation
-//           2012-2015  Johns Hopkins University (Author: Daniel Povey)
-//                2013  Guoguo Chen
-
-#ifndef KALDIFST_CSRC_KALDI_FST_IO_H_
-#define KALDIFST_CSRC_KALDI_FST_IO_H_
-
-#include <string>
-#include <utility>
-
-#include "fst/fst.h"
-#include "fst/vector-fst.h"
-#include "kaldifst/csrc/log.h"
-
-// Some functions for writing Fsts.
-// I/O for FSTs is a bit of a mess, and not very well integrated with Kaldi's
-// generic I/O mechanisms, because we want files containing just FSTs to
-// be readable by OpenFST's native binaries, which is not compatible
-// with the normal \0B header that identifies Kaldi files as containing
-// binary data.
-// So use the functions here with your eyes open, and with caution!
-namespace fst {
-
-// Read a binary FST using Kaldi I/O mechanisms (pipes, etc.)
-// On error returns NULL. Only supports VectorFst and exists
-// mainly for backward code compatibility.
-//
-// Caution: The user has to free the returned pointer using `delete`.
-VectorFst<StdArc> *ReadFstKaldi(std::string rxfilename);
-
-// Read a binary FST using Kaldi I/O mechanisms (pipes, etc.)
-// If it can't read the FST, if throw_on_err == true it throws using
-// KALDIFST_ERR; otherwise it prints a warning and returns. Note:this doesn't
-// support the text-mode option that we generally like to support. This version
-// currently supports ConstFst<StdArc> or VectorFst<StdArc> (const-fst can give
-// better performance for decoding). Other types could be also loaded if
-// registered inside OpenFst.
-Fst<StdArc> *ReadFstKaldiGeneric(std::string rxfilename,
-                                 bool throw_on_err = true);
-
-// This function attempts to dynamic_cast the pointer 'fst' (which will likely
-// have been returned by ReadFstGeneric()), to the more derived
-// type VectorFst<StdArc>. If this succeeds, it returns the same pointer;
-// if it fails, it converts the FST type (by creating a new VectorFst<stdArc>
-// initialized by 'fst'), prints a warning, and deletes 'fst'.
-VectorFst<StdArc> *CastOrConvertToVectorFst(Fst<StdArc> *fst);
-
-ConstFst<StdArc> *CastOrConvertToConstFst(Fst<StdArc> *fst);
-
-// Version of ReadFstKaldi() that writes to a pointer.  Assumes
-// the FST is binary with no binary marker.  Crashes on error.
-void ReadFstKaldi(std::string rxfilename, VectorFst<StdArc> *ofst);
-
-// Write an FST using Kaldi I/O mechanisms (pipes, etc.)
-// On error, throws using KALDIFST_ERR.  For use only in code in fstbin/,
-// as it doesn't support the text-mode option.
-void WriteFstKaldi(const VectorFst<StdArc> &fst, std::string wxfilename);
-
-// This is a more general Kaldi-type-IO mechanism of writing FSTs to
-// streams, supporting binary or text-mode writing.  (note: we just
-// write the integers, symbol tables are not supported).
-// On error, throws using KALDIFST_ERR.
-template <class Arc>
-void WriteFstKaldi(std::ostream &os, bool binary, const VectorFst<Arc> &fst);
-
-// A generic Kaldi-type-IO mechanism of reading FSTs from streams,
-// supporting binary or text-mode reading/writing.
-template <class Arc>
-void ReadFstKaldi(std::istream &is, bool binary, VectorFst<Arc> *fst);
-
-// Read an FST file for LM (G.fst) and make it an acceptor,
-// and make sure it is sorted on labels
-//
-// Caution: The user has to free the returned pointer using `delete`.
-VectorFst<StdArc> *ReadAndPrepareLmFst(const std::string &rxfilename);
-
-// This is a Holder class with T = VectorFst<Arc>, that meets the requirements
-// of a Holder class as described in ../util/kaldi-holder.h. This enables us to
-// read/write collections of FSTs indexed by strings, using the Table concept (
-// see ../util/kaldi-table.h).
-// Originally it was only templated on T = VectorFst<StdArc>, but as the keyword
-// spotting stuff introduced more types of FSTs, we made it also templated on
-// the arc.
-template <class Arc>
-class VectorFstTplHolder {
- public:
-  typedef VectorFst<Arc> T;
-
-  VectorFstTplHolder() : t_(NULL) {}
-
-  static bool Write(std::ostream &os, bool binary, const T &t);
-
-  void Copy(const T &t) {  // copies it into the holder.
-    Clear();
-    t_ = new T(t);
-  }
-
-  // Reads into the holder.
-  bool Read(std::istream &is);
-
-  // It's potentially a binary format, so must read in binary mode (linefeed
-  // translation will corrupt the file.  We don't know till we open the file if
-  // it's really binary, so we need to read in binary mode to be on the safe
-  // side.  Extra linefeeds won't matter, the text-mode reading code ignores
-  // them.
-  static bool IsReadInBinary() { return true; }
-
-  T &Value() {
-    // code error if !t_.
-    if (!t_) KALDIFST_ERR << "VectorFstTplHolder::Value() called wrongly.";
-    return *t_;
-  }
-
-  void Clear() {
-    if (t_) {
-      delete t_;
-      t_ = NULL;
-    }
-  }
-
-  void Swap(VectorFstTplHolder<Arc> *other) { std::swap(t_, other->t_); }
-
-  bool ExtractRange(const VectorFstTplHolder<Arc> &other,
-                    const std::string &range) {
-    KALDIFST_ERR << "ExtractRange is not defined for this type of holder.";
-    return false;
-  }
-
-  ~VectorFstTplHolder() { Clear(); }
-  // No destructor.  Assignment and
-  // copy constructor take their default implementations.
- private:
-  KALDIFST_DISALLOW_COPY_AND_ASSIGN(VectorFstTplHolder);
-  T *t_;
-};
-
-// Now make the original VectorFstHolder as the typedef of
-// VectorFstHolder<StdArc>.
-typedef VectorFstTplHolder<StdArc> VectorFstHolder;
-
-}  // namespace fst
-
-#include "kaldifst/csrc/kaldi-fst-io-inl.h"
-
-#endif  // KALDIFST_CSRC_KALDI_FST_IO_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-holder-inl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-holder-inl.h
deleted file mode 100644
index 6b5275fa..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-holder-inl.h
+++ /dev/null
@@ -1,76 +0,0 @@
-// kaldifst/csrc/kaldi-holder-inl.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/util/kaldi-holder-inl.h
-
-// Copyright 2009-2011     Microsoft Corporation
-//                2016     Johns Hopkins University (author: Daniel Povey)
-//                2016     Xiaohui Zhang
-//
-#ifndef KALDIFST_CSRC_KALDI_HOLDER_INL_H_
-#define KALDIFST_CSRC_KALDI_HOLDER_INL_H_
-
-#include <iostream>
-#include <string>
-
-#include "kaldifst/csrc/log.h"
-#include "kaldifst/csrc/text-utils.h"
-
-namespace kaldifst {
-
-// We define a Token as a nonempty, printable, whitespace-free std::string.
-// The binary and text formats here are the same (newline-terminated)
-// and as such we don't bother with the binary-mode headers.
-class TokenHolder {
- public:
-  typedef std::string T;
-
-  TokenHolder() {}
-
-  static bool Write(std::ostream &os, bool, const T &t) {  // ignore binary-mode
-    KALDIFST_ASSERT(IsToken(t));
-    os << t << '\n';
-    return os.good();
-  }
-
-  void Clear() { t_.clear(); }
-
-  // Reads into the holder.
-  bool Read(std::istream &is) {
-    is >> t_;
-    if (is.fail()) return false;
-    char c;
-    while (isspace(c = is.peek()) && c != '\n') is.get();
-    if (is.peek() != '\n') {
-      KALDIFST_WARN << "TokenHolder::Read, expected newline, got char "
-                    << CharToString(is.peek()) << ", at stream pos "
-                    << is.tellg();
-      return false;
-    }
-    is.get();  // get '\n'
-    return true;
-  }
-
-  // Since this is fundamentally a text format, read in text mode (would work
-  // fine either way, but doing it this way will exercise more of the code).
-  static bool IsReadInBinary() { return false; }
-
-  T &Value() { return t_; }
-
-  ~TokenHolder() {}
-
-  void Swap(TokenHolder *other) { t_.swap(other->t_); }
-
-  bool ExtractRange(const TokenHolder &other, const std::string &range) {
-    KALDIFST_ERR << "ExtractRange is not defined for this type of holder.";
-    return false;
-  }
-
- private:
-  KALDIFST_DISALLOW_COPY_AND_ASSIGN(TokenHolder);
-  T t_;
-};
-
-}  // namespace kaldifst
-
-#endif  // KALDIFST_CSRC_KALDI_HOLDER_INL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-holder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-holder.h
deleted file mode 100644
index e93ce727..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-holder.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// kaldifst/csrc/kaldi-holder.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/util/kaldi-holder.h
-
-// Copyright 2009-2011     Microsoft Corporation
-//                2016     Johns Hopkins University (author: Daniel Povey)
-//                2016     Xiaohui Zhang
-//
-#ifndef KALDIFST_CSRC_KALDI_HOLDER_H_
-#define KALDIFST_CSRC_KALDI_HOLDER_H_
-
-#include <string>
-
-namespace kaldifst {
-
-/// We define a Token (not a typedef, just a word) as a nonempty, printable,
-/// whitespace-free std::string.  The binary and text formats here are the same
-/// (newline-terminated) and as such we don't bother with the binary-mode
-/// headers.
-class TokenHolder;
-
-// In SequentialTableReaderScriptImpl and RandomAccessTableReaderScriptImpl, for
-// cases where the scp contained 'range specifiers' (things in square brackets
-// identifying parts of objects like matrices), use this function to separate
-// the input string 'rxfilename_with_range' (e.g "1.ark:100[1:2,2:10]") into the
-// data_rxfilename (e.g. "1.ark:100") and the optional range specifier which
-// will be everything inside the square brackets.  It returns true if everything
-// seems OK, and false if for example the string contained more than one '['.
-// This function should only be called if 'line' ends in ']', otherwise it is an
-// error.
-bool ExtractRangeSpecifier(const std::string &rxfilename_with_range,
-                           std::string *data_rxfilename, std::string *range);
-
-}  // namespace kaldifst
-
-#include "kaldifst/csrc/kaldi-holder-inl.h"
-
-#endif  // KALDIFST_CSRC_KALDI_HOLDER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-io-inl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-io-inl.h
deleted file mode 100644
index dda5b597..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-io-inl.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// kaldifst/csrc/util/kaldi-io-inl.h
-
-// Copyright 2009-2011 Microsoft Corporation
-
-// See ../../COPYING for clarification regarding multiple authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-
-//  http://www.apache.org/licenses/LICENSE-2.0
-
-// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
-// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
-// MERCHANTABLITY OR NON-INFRINGEMENT.
-// See the Apache 2 License for the specific language governing permissions and
-// limitations under the License.
-#ifndef KALDIFST_CSRC_KALDI_IO_INL_H_
-#define KALDIFST_CSRC_KALDI_IO_INL_H_
-
-#include <string>
-
-namespace kaldifst {
-
-bool Input::Open(const std::string &rxfilename, bool *binary) {
-  return OpenInternal(rxfilename, true, binary);
-}
-
-bool Input::OpenTextMode(const std::string &rxfilename) {
-  return OpenInternal(rxfilename, false, NULL);
-}
-
-bool Input::IsOpen() { return impl_ != NULL; }
-
-bool Output::IsOpen() { return impl_ != NULL; }
-
-}  // namespace kaldifst
-
-#endif  // KALDIFST_CSRC_KALDI_IO_INL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-io.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-io.h
deleted file mode 100644
index bdbbc735..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-io.h
+++ /dev/null
@@ -1,227 +0,0 @@
-// kaldifst/csrc/kaldi-io.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/util/kaldi-io.h
-
-// Copyright 2009-2011  Microsoft Corporation;  Jan Silovsky
-//                2016  Xiaohui Zhang
-
-#ifndef KALDIFST_CSRC_KALDI_IO_H_
-#define KALDIFST_CSRC_KALDI_IO_H_
-
-#include <string>
-
-#include "kaldifst/csrc/log.h"
-
-namespace kaldifst {
-
-class OutputImplBase;  // Forward decl; defined in a .cc file
-class InputImplBase;   // Forward decl; defined in a .cc file
-
-// The Output and Input classes handle stream-opening for "extended" filenames
-// that include actual files, standard-input/standard-output, pipes, and
-// offsets into actual files.  They also handle reading and writing the
-// binary-mode headers for Kaldi files, where applicable.  The classes have
-// versions of the Open routines that throw and do not throw, depending whether
-// the calling code wants to catch the errors or not; there are also versions
-// that write (or do not write) the Kaldi binary-mode header that says if it's
-// binary mode.  Generally files that contain Kaldi objects will have the header
-// on, so we know upon reading them whether they have the header.  So you would
-// use the OpenWithHeader routines for these (or the constructor); but other
-// types of objects (e.g. FSTs) would have files without a header so you would
-// use OpenNoHeader.
-
-// We now document the types of extended filenames that we use.
-//
-// A "wxfilename"  is an extended filename for writing. It can take three forms:
-// (1) Filename: e.g.    "/some/filename", "./a/b/c", "c:\Users\dpovey\My
-//                        Documents\\boo"
-//          (whatever the actual file-system interprets)
-// (2) Standard output:  "" or "-"
-// (3) A pipe: e.g. "| gzip -c > /tmp/abc.gz"
-//
-//
-// A "rxfilename" is an extended filename for reading.  It can take four forms:
-// (1) An actual filename, whatever the file-system can read, e.g. "/my/file".
-// (2) Standard input: "" or "-"
-// (3) A pipe: e.g.  "gunzip -c /tmp/abc.gz |"
-// (4) An offset into a file, e.g.: "/mnt/blah/data/1.ark:24871"
-//   [these are created by the Table and TableWriter classes; I may also write
-//    a program that creates them for arbitrary files]
-//
-
-// Typical usage:
-// ...
-// bool binary;
-// MyObject.Write(Output(some_filename, binary).Stream(), binary);
-//
-// ... more extensive example:
-// {
-//    Output ko(some_filename, binary);
-//    MyObject1.Write(ko.Stream(), binary);
-//    MyObject2.Write(ko.Stream(), binary);
-// }
-enum OutputType { kNoOutput, kFileOutput, kStandardOutput, kPipeOutput };
-
-/// ClassifyWxfilename interprets filenames as follows:
-///  - kNoOutput: invalid filenames (leading or trailing space, things that look
-///     like wspecifiers and rspecifiers or like pipes to read from with leading
-///     |.
-///  - kFileOutput: Normal filenames
-///  - kStandardOutput: The empty string or "-", interpreted as standard output
-///  - kPipeOutput: pipes, e.g. "| gzip -c > /tmp/abc.gz"
-OutputType ClassifyWxfilename(const std::string &wxfilename);
-
-enum InputType {
-  kNoInput,
-  kFileInput,
-  kStandardInput,
-  kOffsetFileInput,
-  kPipeInput
-};
-
-/// ClassifyRxfilenames interprets filenames for reading as follows:
-///  - kNoInput: invalid filenames (leading or trailing space, things that
-///       look like wspecifiers and rspecifiers or pipes to write to
-///       with trailing |.
-///  - kFileInput: normal filenames
-///  - kStandardInput: the empty string or "-"
-///  - kPipeInput: e.g. "gunzip -c /tmp/abc.gz |"
-///  - kOffsetFileInput: offsets into files, e.g.  /some/filename:12970
-InputType ClassifyRxfilename(const std::string &rxfilename);
-
-class Output {
- public:
-  // The normal constructor, provided for convenience.
-  // Equivalent to calling with default constructor then Open()
-  // with these arguments.
-  Output(const std::string &filename, bool binary, bool write_header = true);
-
-  Output() : impl_(NULL) {}
-
-  /// This opens the stream, with the given mode (binary or text).  It returns
-  /// true on success and false on failure.  However, it will throw if something
-  /// was already open and could not be closed (to avoid this, call Close()
-  /// first.  if write_header == true and binary == true, it writes the Kaldi
-  /// binary-mode header ('\0' then 'B').  You may call Open even if it is
-  /// already open; it will close the existing stream and reopen (however if
-  /// closing the old stream failed it will throw).
-  bool Open(const std::string &wxfilename, bool binary, bool write_header);
-
-  inline bool IsOpen();  // return true if we have an open stream.  Does not
-  // imply stream is good for writing.
-
-  std::ostream &Stream();  // will throw if not open; else returns stream.
-
-  // Close closes the stream. Calling Close is never necessary unless you
-  // want to avoid exceptions being thrown.  There are times when calling
-  // Close will hurt efficiency (basically, when using offsets into files,
-  // and using the same Input object),
-  // but most of the time the user won't be doing this directly, it will
-  // be done in kaldi-table.{h, cc}, so you don't have to worry about it.
-  bool Close();
-
-  // This will throw if stream could not be closed (to check error status,
-  // call Close()).
-  ~Output();
-
- private:
-  OutputImplBase *impl_;  // non-NULL if open.
-  std::string filename_;
-  KALDIFST_DISALLOW_COPY_AND_ASSIGN(Output);
-};
-
-// bool binary_in;
-// Input ki(some_filename, &binary_in);
-// MyObject.Read(ki.Stream(), binary_in);
-//
-// ... more extensive example:
-//
-// {
-//    bool binary_in;
-//    Input ki(some_filename, &binary_in);
-//    MyObject1.Read(ki.Stream(), &binary_in);
-//    MyObject2.Write(ki.Stream(), &binary_in);
-// }
-// Note that to catch errors you need to use try.. catch.
-// Input communicates errors by throwing exceptions.
-
-// Input interprets four kinds of filenames:
-//  (1) Normal filenames
-//  (2) The empty string or "-", interpreted as standard output
-//  (3) A pipe: e.g.  "gunzip -c /tmp/abc.gz |"
-//  (4) Offsets into [real] files, e.g. "/my/filename:12049"
-// The last one has no correspondence in Output.
-
-class Input {
- public:
-  /// The normal constructor.  Opens the stream in binary mode.
-  /// Equivalent to calling the default constructor followed by Open(); then, if
-  /// binary != NULL, it calls ReadHeader(), putting the output in "binary"; it
-  /// throws on error.
-  explicit Input(const std::string &rxfilename, bool *contents_binary = NULL);
-
-  Input() : impl_(NULL) {}
-
-  // Open opens the stream for reading (the mode, where relevant, is binary; use
-  // OpenTextMode for text-mode, we made this a separate function rather than a
-  // boolean argument, to avoid confusion with Kaldi's text/binary distinction,
-  // since reading in the file system's text mode is unusual.)  If
-  // contents_binary != NULL, it reads the binary-mode header and puts it in the
-  // "binary" variable.  Returns true on success.  If it returns false it will
-  // not be open.  You may call Open even if it is already open; it will close
-  // the existing stream and reopen (however if closing the old stream failed it
-  // will throw).
-  inline bool Open(const std::string &rxfilename, bool *contents_binary = NULL);
-
-  // As Open but (if the file system has text/binary modes) opens in text mode;
-  // you shouldn't ever have to use this as in Kaldi we read even text files in
-  // binary mode (and ignore the \r).
-  inline bool OpenTextMode(const std::string &rxfilename);
-
-  // Return true if currently open for reading and Stream() will
-  // succeed.  Does not guarantee that the stream is good.
-  inline bool IsOpen();
-
-  // It is never necessary or helpful to call Close, except if
-  // you are concerned about to many filehandles being open.
-  // Close does not throw. It returns the exit code as int32_t
-  // in the case of a pipe [kPipeInput], and always zero otherwise.
-  int32_t Close();
-
-  // Returns the underlying stream. Throws if !IsOpen()
-  std::istream &Stream();
-
-  // Destructor does not throw: input streams may legitimately fail so we
-  // don't worry about the status when we close them.
-  ~Input();
-
- private:
-  bool OpenInternal(const std::string &rxfilename, bool file_binary,
-                    bool *contents_binary);
-  InputImplBase *impl_;
-  KALDIFST_DISALLOW_COPY_AND_ASSIGN(Input);
-};
-
-template <class C>
-void ReadKaldiObject(const std::string &filename, C *c) {
-  bool binary_in;
-  Input ki(filename, &binary_in);
-  c->Read(ki.Stream(), binary_in);
-}
-
-/// PrintableRxfilename turns the rxfilename into a more human-readable
-/// form for error reporting, i.e. it does quoting and escaping and
-/// replaces "" or "-" with "standard input".
-std::string PrintableRxfilename(const std::string &rxfilename);
-
-/// PrintableWxfilename turns the wxfilename into a more human-readable
-/// form for error reporting, i.e. it does quoting and escaping and
-/// replaces "" or "-" with "standard output".
-std::string PrintableWxfilename(const std::string &wxfilename);
-
-}  // namespace kaldifst
-
-#include "kaldifst/csrc/kaldi-io-inl.h"
-
-#endif  // KALDIFST_CSRC_KALDI_IO_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-math.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-math.h
deleted file mode 100644
index f915cddc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-math.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// kaldifst/csrc/kaldi-math.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/base/kaldi-math.h
-//
-// Copyright 2009-2011  Ondrej Glembek;  Microsoft Corporation;  Yanmin Qian;
-//                      Jan Silovsky;  Saarland University
-//
-#ifndef KALDIFST_CSRC_KALDI_MATH_H_
-#define KALDIFST_CSRC_KALDI_MATH_H_
-
-#include <cstdint>
-
-namespace kaldifst {
-
-struct RandomState {
-  RandomState();
-  unsigned seed;
-};
-
-int32_t Rand(struct RandomState *state = nullptr);
-
-// Returns a random integer between first and last inclusive.
-int32_t RandInt(int32_t first, int32_t last,
-                struct RandomState *state = nullptr);
-
-}  // namespace kaldifst
-
-#endif  // KALDIFST_CSRC_KALDI_MATH_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-pipebuf.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-pipebuf.h
deleted file mode 100644
index 5915d788..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-pipebuf.h
+++ /dev/null
@@ -1,88 +0,0 @@
-// kaldifst/csrc/util/kaldi-pipebuf.h
-
-// Copyright 2009-2011  Ondrej Glembek
-
-// See ../../COPYING for clarification regarding multiple authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-
-//  http://www.apache.org/licenses/LICENSE-2.0
-
-// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
-// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
-// MERCHANTABLITY OR NON-INFRINGEMENT.
-// See the Apache 2 License for the specific language governing permissions and
-// limitations under the License.
-
-/** @file kaldi-pipebuf.h
- *  This is an Kaldi C++ Library header.
- */
-
-#ifndef KALDIFST_CSRC_KALDI_PIPEBUF_H_
-#define KALDIFST_CSRC_KALDI_PIPEBUF_H_
-
-#include <string>
-#if !defined(_LIBCPP_VERSION)  // libc++
-#include <fstream>
-#else
-#include "kaldifst/csrc/basic-filebuf.h"
-#endif
-
-#include "kaldifst/csrc/log.h"
-
-namespace kaldifst {
-// This class provides a way to initialize a filebuf with a FILE* pointer
-// directly; it will not close the file pointer when it is deleted.
-// The C++ standard does not allow implementations of C++ to provide
-// this constructor within basic_filebuf, which makes it hard to deal
-// with pipes using completely native C++.  This is a workaround
-
-#ifdef _MSC_VER
-#elif defined(_LIBCPP_VERSION)  // libc++
-template <class CharType, class Traits = std::char_traits<CharType>>
-class basic_pipebuf : public basic_filebuf<CharType, Traits> {
- public:
-  typedef basic_pipebuf<CharType, Traits> ThisType;
-
- public:
-  basic_pipebuf(FILE *fptr, std::ios_base::openmode mode)
-      : basic_filebuf<CharType, Traits>() {
-    this->open(fptr, mode);
-    if (!this->is_open()) {
-      KALDIFST_WARN << "Error initializing pipebuf";  // probably indicates
-      // code error, if the fptr was good.
-      return;
-    }
-  }
-};  // class basic_pipebuf
-#else
-template <class CharType, class Traits = std::char_traits<CharType>>
-class basic_pipebuf : public std::basic_filebuf<CharType, Traits> {
- public:
-  typedef basic_pipebuf<CharType, Traits> ThisType;
-
- public:
-  basic_pipebuf(FILE *fptr, std::ios_base::openmode mode)
-      : std::basic_filebuf<CharType, Traits>() {
-    this->_M_file.sys_open(fptr, mode);
-    if (!this->is_open()) {
-      KALDIFST_WARN << "Error initializing pipebuf";  // probably indicates
-      // code error, if the fptr was good.
-      return;
-    }
-    this->_M_mode = mode;
-    this->_M_buf_size = BUFSIZ;
-    this->_M_allocate_internal_buffer();
-    this->_M_reading = false;
-    this->_M_writing = false;
-    this->_M_set_buffer(-1);
-  }
-};  // class basic_pipebuf
-#endif  // _MSC_VER
-
-}  // namespace kaldifst
-
-#endif  // KALDIFST_CSRC_KALDI_PIPEBUF_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-semaphore.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-semaphore.h
deleted file mode 100644
index b3c90f62..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-semaphore.h
+++ /dev/null
@@ -1,51 +0,0 @@
-// kaldifst/csrc/kaldi-semaphore.h
-
-// Copyright 2012  Karel Vesely (Brno University of Technology)
-//           2017  Dogan Can (University of Southern California)
-
-// See ../../COPYING for clarification regarding multiple authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//  http://www.apache.org/licenses/LICENSE-2.0
-//
-// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
-// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
-// MERCHANTABLITY OR NON-INFRINGEMENT.
-// See the Apache 2 License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef KALDIFST_CSRC_KALDI_SEMAPHORE_H_
-#define KALDIFST_CSRC_KALDI_SEMAPHORE_H_
-
-#include <condition_variable>  // NOLINT
-#include <mutex>               // NOLINT
-
-#include "kaldifst/csrc/log.h"
-
-namespace kaldifst {
-
-class Semaphore {
- public:
-  explicit Semaphore(int32_t count = 0);
-
-  ~Semaphore();
-
-  bool TryWait();  ///< Returns true if Wait() goes through
-  void Wait();     ///< decrease the counter
-  void Signal();   ///< increase the counter
-
- private:
-  int32_t count_;  ///< the semaphore counter, 0 means block on Wait()
-
-  std::mutex mutex_;
-  std::condition_variable condition_variable_;
-  KALDIFST_DISALLOW_COPY_AND_ASSIGN(Semaphore);
-};
-
-}  // namespace kaldifst
-
-#endif  // KALDIFST_CSRC_KALDI_SEMAPHORE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-table-inl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-table-inl.h
deleted file mode 100644
index bbae1bd6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-table-inl.h
+++ /dev/null
@@ -1,2703 +0,0 @@
-// kaldifst/src/kaldi-table-inl.h
-
-// Copyright 2009-2011    Microsoft Corporation
-//                2013    Johns Hopkins University (author: Daniel Povey)
-//                2016    Xiaohui Zhang
-
-// See ../../COPYING for clarification regarding multiple authors
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//  http://www.apache.org/licenses/LICENSE-2.0
-//
-// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
-// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
-// MERCHANTABLITY OR NON-INFRINGEMENT.
-// See the Apache 2 License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef KALDIFST_CSRC_KALDI_TABLE_INL_H_
-#define KALDIFST_CSRC_KALDI_TABLE_INL_H_
-
-#include <errno.h>
-#include <string.h>
-
-#include <algorithm>
-#include <string>
-#include <thread>  // NOLINT
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include "kaldifst/csrc/kaldi-io.h"
-#include "kaldifst/csrc/log.h"
-// #include "util/kaldi-io.h"
-// #include "util/kaldi-holder.h"
-// #include "util/text-utils.h"
-#include "kaldifst/csrc/kaldi-semaphore.h"
-#include "kaldifst/csrc/stl-utils.h"  // for StringHasher.
-
-namespace kaldifst {
-
-/// \addtogroup table_impl_types
-/// @{
-
-template <class Holder>
-class SequentialTableReaderImplBase {
- public:
-  typedef typename Holder::T T;
-  // note that Open takes rxfilename not rspecifier.  Open will only be
-  // called on a just-allocated object.
-  virtual bool Open(const std::string &rxfilename) = 0;
-  // Done() should be called on a successfully opened, not-closed object.
-  // only throws if called at the wrong time (i.e. code error).
-  virtual bool Done() const = 0;
-  // Returns true if the reader is open [i.e. Open() succeeded and
-  // the user has not called Close()]
-  virtual bool IsOpen() const = 0;
-  // Returns the current key; it is valid to call this if Done() returned false.
-  // Only throws on code error (i.e. called at the wrong time).
-  virtual std::string Key() = 0;
-  // Returns the value associated with the current key.  Valid to call it if
-  // Done() returned false.  It throws if the value could not be read.  [However
-  // if you use the ,p modifier it will never throw, unless you call it at the
-  // wrong time, i.e. unless there is a code error.]
-  virtual T &Value() = 0;
-  virtual void FreeCurrent() = 0;
-  // move to the next object.  This won't throw unless called wrongly (e.g. on
-  // non-open archive.]
-  virtual void Next() = 0;
-  // Close the table.  Returns its status as bool so it won't throw, unless
-  // called wrongly [i.e. on non-open archive.]
-  virtual bool Close() = 0;
-  // SwapHolder() is not part of the public interface of SequentialTableReader.
-  // It should be called when it would be valid to call Value() or FreeCurrent()
-  // (i.e. when a value is stored), and after this it's not valid to get the
-  // value any more until you call Next().  It swaps the contents of
-  // this->holder_ with those of 'other_holder'.  It's needed as part of how
-  // we implement SequentialTableReaderBackgroundImpl.
-  virtual void SwapHolder(Holder *other_holder) = 0;
-  SequentialTableReaderImplBase() {}
-  virtual ~SequentialTableReaderImplBase() {}  // throws.
-
- private:
-  KALDIFST_DISALLOW_COPY_AND_ASSIGN(SequentialTableReaderImplBase);
-};
-
-// This is the implementation for SequentialTableReader
-// when it's actually a script file.
-template <class Holder>
-class SequentialTableReaderScriptImpl
-    : public SequentialTableReaderImplBase<Holder> {
- public:
-  typedef typename Holder::T T;
-
-  SequentialTableReaderScriptImpl() : state_(kUninitialized) {}
-
-  // You may call Open from states kUninitialized and kError.
-  // It may leave the object in any of the states.
-  virtual bool Open(const std::string &rspecifier) {
-    if (state_ != kUninitialized && state_ != kError)
-      if (!Close())  // call Close() yourself to suppress this exception.
-        KALDIFST_ERR << "Error closing previous input: "
-                     << "rspecifier was " << rspecifier_;
-    bool binary;
-    rspecifier_ = rspecifier;
-    RspecifierType rs =
-        ClassifyRspecifier(rspecifier, &script_rxfilename_, &opts_);
-    KALDIFST_ASSERT(rs == kScriptRspecifier);
-    if (!script_input_.Open(script_rxfilename_, &binary)) {  // Failure on Open
-      KALDIFST_WARN << "Failed to open script file "
-                    << PrintableRxfilename(script_rxfilename_);
-      state_ = kUninitialized;
-      return false;
-    } else {  // Open succeeded.
-      if (binary) {
-        KALDIFST_WARN << "Script file should not be binary file.";
-        SetErrorState();
-        return false;
-      } else {
-        state_ = kFileStart;
-        Next();
-        if (state_ == kError) return false;
-        // any other status, including kEof, is OK from the point of view of
-        // the 'open' function (empty scp file is not inherently an error).
-        return true;
-      }
-    }
-  }
-
-  virtual bool IsOpen() const {
-    switch (state_) {
-      case kEof:
-      case kHaveScpLine:
-      case kHaveObject:
-      case kHaveRange:
-        return true;
-      case kUninitialized:
-      case kError:
-        return false;
-      default:
-        KALDIFST_ERR << "IsOpen() called on invalid object.";
-        // note: kFileStart is not a valid state for the user to call a member
-        // function (we never return from a public function in this state).
-        return false;
-    }
-  }
-
-  virtual bool Done() const {
-    switch (state_) {
-      case kHaveScpLine:
-      case kHaveObject:
-      case kHaveRange:
-        return false;
-      case kEof:
-      case kError:
-        return true;  // Error condition, like Eof, counts
-        // as Done(); the destructor/Close() will inform the user of the error.
-      default:
-        KALDIFST_ERR << "Done() called on TableReader object at the wrong"
-                        " time.";
-        return false;
-    }
-  }
-
-  virtual std::string Key() {
-    // Valid to call this whenever Done() returns false.
-    switch (state_) {
-      case kHaveScpLine:
-      case kHaveObject:
-      case kHaveRange:
-        break;
-      default:
-        // coding error.
-        KALDIFST_ERR << "Key() called on TableReader object at the wrong time.";
-    }
-    return key_;
-  }
-
-  T &Value() {
-    if (!EnsureObjectLoaded())
-      KALDIFST_ERR << "Failed to load object from "
-                   << PrintableRxfilename(data_rxfilename_)
-                   << " (to suppress this error, add the permissive "
-                   << "(p, ) option to the rspecifier.";
-    // Because EnsureObjectLoaded() returned with success, we know
-    // that if range_ is nonempty (i.e. a range was requested), the
-    // state will be kHaveRange.
-    if (state_ == kHaveRange) {
-      return range_holder_.Value();
-    } else {
-      KALDIFST_ASSERT(state_ == kHaveObject);
-      return holder_.Value();
-    }
-  }
-
-  void FreeCurrent() {
-    if (state_ == kHaveObject) {
-      holder_.Clear();
-      state_ = kHaveScpLine;
-    } else if (state_ == kHaveRange) {
-      range_holder_.Clear();
-      state_ = kHaveObject;
-    } else {
-      KALDIFST_WARN << "FreeCurrent called at the wrong time.";
-    }
-  }
-
-  void SwapHolder(Holder *other_holder) {
-    // call Value() to ensure we have a value, and ignore its return value while
-    // suppressing compiler warnings by casting to void.  It will cause the
-    // program to die with KALDIFST_ERR if we couldn't get a value.
-    (void)Value();
-    // At this point we know that we successfully loaded an object,
-    // and if there was a range specified, it's in range_holder_.
-    if (state_ == kHaveObject) {
-      holder_.Swap(other_holder);
-      state_ = kHaveScpLine;
-    } else if (state_ == kHaveRange) {
-      range_holder_.Swap(other_holder);
-      state_ = kHaveObject;
-      // This indicates that we still have the base object (but no range).
-    } else {
-      KALDIFST_ERR << "Code error";
-    }
-    // Note: after this call there may be some junk left in range_holder_ or
-    // holder_, but it won't matter.  We avoid calling Clear() on them, as this
-    // function needs to be lightweight for the 'bg' feature to work well.
-  }
-
-  // Next goes to the next object.
-  // It can leave the object in most of the statuses, but
-  // the only circumstances under which it will return are:
-  //  either:
-  //  - if Done() returned true, i.e. kError or kEof.
-  //  or:
-  //  - in non-permissive mode, status kHaveScpLine or kHaveObjecct
-  //  - in permissive mode, only when we successfully have an object,
-  //    which means either (kHaveObject and range_.empty()), or
-  //    kHaveRange.
-  void Next() {
-    while (1) {
-      NextScpLine();
-      if (Done()) return;
-      if (opts_.permissive) {
-        // Permissive mode means, when reading scp files, we treat keys whose
-        // scp entry cannot be read as nonexistent.  This means trying to read.
-        if (EnsureObjectLoaded()) return;  // Success.
-        // else try the next scp line.
-      } else {
-        return;  // We go the next key; Value() will crash if we can't read the
-                 // object on the scp line.
-      }
-    }
-  }
-
-  // This function may be entered at in any state.  At exit, the object will be
-  // in state kUninitialized.  It only returns false in the situation where we
-  // were at the end of the stream (kEof) and the script_input_ was a pipe and
-  // it ended with error status; this is so that we can catch errors from
-  // programs that we invoked via a pipe.
-  virtual bool Close() {
-    int32_t status = 0;
-    if (script_input_.IsOpen()) status = script_input_.Close();
-    if (data_input_.IsOpen()) data_input_.Close();
-    range_holder_.Clear();
-    holder_.Clear();
-    if (!this->IsOpen())
-      KALDIFST_ERR << "Close() called on input that was not open.";
-    StateType old_state = state_;
-    state_ = kUninitialized;
-    if (old_state == kError || (old_state == kEof && status != 0)) {
-      if (opts_.permissive) {
-        KALDIFST_WARN
-            << "Close() called on scp file with read error, ignoring the"
-               " error because permissive mode specified.";
-        return true;
-      } else {
-        return false;  // User will do something with the error status.
-      }
-    } else {
-      return true;
-    }
-    // Possible states                                          Return value
-    // kLoadSucceeded/kRangeSucceeded/kRangeFailed              true
-    // kError (if opts_.permissive)                             true
-    // kError (if !opts_.permissive)                            false
-    // kEof (if script_input_.Close() && !opts.permissive)      false
-    // kEof (if !script_input_.Close() || opts.permissive)      true
-    // kUninitialized/kFileStart/kHaveScpLine                   true
-    // kUnitialized                                             true
-  }
-
-  virtual ~SequentialTableReaderScriptImpl() {
-    if (this->IsOpen() && !Close())
-      KALDIFST_ERR << "TableReader: reading script file failed: from scp "
-                   << PrintableRxfilename(script_rxfilename_);
-  }
-
- private:
-  // Function EnsureObjectLoaded() ensures that we have fully loaded any object
-  // (including object range) associated with the current key, and returns true
-  // on success (i.e. we have the object) and false on failure.
-  //
-  // Possible entry states: kHaveScpLine, kLoadSucceeded, kRangeSucceeded
-  //
-  // Possible exit states: kHaveScpLine, kLoadSucceeded, kRangeSucceeded.
-  //
-  // Note: the return status has information that cannot be deduced from
-  // just the exit state.  If the object could not be loaded we go to state
-  // kHaveScpLine but return false; and if the range was requested but
-  // could not be extracted, we go to state kLoadSucceeded but return false.
-  bool EnsureObjectLoaded() {
-    if (!(state_ == kHaveScpLine || state_ == kHaveObject ||
-          state_ == kHaveRange))
-      KALDIFST_ERR << "Invalid state (code error)";
-
-    if (state_ == kHaveScpLine) {  // need to load the object into holder_.
-      bool ans;
-      // note, NULL means it doesn't read the binary-mode header
-      if (Holder::IsReadInBinary()) {
-        ans = data_input_.Open(data_rxfilename_, NULL);
-      } else {
-        ans = data_input_.OpenTextMode(data_rxfilename_);
-      }
-      if (!ans) {
-        KALDIFST_WARN << "Failed to open file "
-                      << PrintableRxfilename(data_rxfilename_);
-        return false;
-      } else {
-        if (holder_.Read(data_input_.Stream())) {
-          state_ = kHaveObject;
-        } else {  // holder_ will not contain data.
-          KALDIFST_WARN << "Failed to load object from "
-                        << PrintableRxfilename(data_rxfilename_);
-          return false;
-        }
-      }
-    }
-    // OK, at this point the state must be either
-    // kHaveObject or kHaveRange.
-    if (range_.empty()) {
-      // if range_ is the empty string, we should not be in the state
-      // kHaveRange.
-      KALDIFST_ASSERT(state_ == kHaveObject);
-      return true;
-    }
-    // range_ is nonempty.
-    if (state_ == kHaveRange) {
-      // range was already extracted, so there nothing to do.
-      return true;
-    }
-    // OK, range_ is nonempty and state_ is kHaveObject.  We attempt to extract
-    // the range object.  Note: ExtractRange() will throw with KALDIFST_ERR if
-    // the object type doesn't support ranges.
-    if (!range_holder_.ExtractRange(holder_, range_)) {
-      KALDIFST_WARN << "Failed to load object from "
-                    << PrintableRxfilename(data_rxfilename_) << "[" << range_
-                    << "]";
-      return false;
-    } else {
-      state_ = kHaveRange;
-      return true;
-    }
-  }
-
-  void SetErrorState() {
-    state_ = kError;
-    script_input_.Close();
-    data_input_.Close();
-    holder_.Clear();
-    range_holder_.Clear();
-  }
-
-  // Reads the next line in the script file.
-  // Possible entry states: kHaveObject, kHaveRange, kHaveScpLine, kFileStart.
-  // Possible exit states: kEof, kError, kHaveScpLine, kHaveObject.
-  void NextScpLine() {
-    switch (state_) {  // Check and simplify the state.
-      case kHaveRange:
-        range_holder_.Clear();
-        state_ = kHaveObject;
-        break;
-      case kHaveScpLine:
-      case kHaveObject:
-      case kFileStart:
-        break;
-      default:
-        // No other states are valid to call Next() from.
-        KALDIFST_ERR << "Reading script file: Next called wrongly.";
-    }
-    // at this point the state will be kHaveObject, kHaveScpLine, or kFileStart.
-    std::string line;
-    if (getline(script_input_.Stream(), line)) {
-      // After extracting "key" from "line", we put the rest
-      // of "line" into "rest", and then extract data_rxfilename_
-      // (e.g. 1.ark:100) and possibly the range_ specifer
-      // (e.g. [1:2,2:10]) from "rest".
-      std::string data_rxfilename, rest;
-      SplitStringOnFirstSpace(line, &key_, &rest);
-      if (!key_.empty() && !rest.empty()) {
-        // Got a valid line.
-        if (rest[rest.size() - 1] == ']') {
-          if (!ExtractRangeSpecifier(rest, &data_rxfilename, &range_)) {
-            KALDIFST_WARN << "Reading rspecifier '" << rspecifier_
-                          << ", cannot make sense of scp line " << line;
-            SetErrorState();
-            return;
-          }
-        } else {
-          data_rxfilename = rest;
-          range_ = "";
-        }
-        bool filenames_equal = (data_rxfilename_ == data_rxfilename);
-        if (!filenames_equal) data_rxfilename_ = data_rxfilename;
-        if (state_ == kHaveObject) {
-          if (!filenames_equal) {
-            holder_.Clear();
-            state_ = kHaveScpLine;
-          }
-          // else leave state_ at kHaveObject and leave the object in the
-          // holder.
-        } else {
-          state_ = kHaveScpLine;
-        }
-      } else {
-        KALDIFST_WARN << "We got an invalid line in the scp file. "
-                      << "It should look like: some_key 1.ark:10, got: "
-                      << line;
-        SetErrorState();
-      }
-    } else {
-      state_ = kEof;  // there is nothing more in the scp file.  Might as well
-                      // close input streams as we don't need them.
-      script_input_.Close();
-      if (data_input_.IsOpen()) data_input_.Close();
-      holder_.Clear();        // clear the holder if it was nonempty.
-      range_holder_.Clear();  // clear the range holder if it was nonempty.
-    }
-  }
-
-  std::string rspecifier_;  // the rspecifier that this class was opened with.
-  RspecifierOptions opts_;  // options.
-  std::string script_rxfilename_;  // rxfilename of the script file.
-
-  Input script_input_;  // Input object for the .scp file
-  Input
-      data_input_;  // Input object for the entries in the script file;
-                    // we make this a class member instead of a local variable,
-                    // so that rspecifiers of the form filename:byte-offset,
-                    // e.g. foo.ark:12345, can be handled using fseek().
-
-  Holder holder_;        // Holds the object.
-  Holder range_holder_;  // Holds the partial object corresponding to the object
-                         // range specifier 'range_'; this is only used when
-                         // 'range_' is specified, i.e. when the .scp file
-                         // contains lines of the form rspecifier[range], like
-                         // foo.ark:242[0:9] (representing a row range of a
-                         // matrix).
-
-  std::string key_;  // the key of the current scp line we're processing
-  std::string
-      data_rxfilename_;  // the rxfilename corresponding to the current key
-  std::string
-      range_;  // the range of object corresponding to the current key, if an
-               // object range was specified in the script file, else "".
-
-  enum StateType {
-    //  Summary of the states this object can be in (state_).
-    //
-    //                (*) Does holder_ contain the object corresponding to
-    //                    data_rxfilename_ ?
-    //                    (*) Does range_holder_ contain a range object?
-    //                         (*) is script_input_ open?
-    //                             (*) are key_, data_rxfilename_ and range_ [if
-    //                             applicable] set?
-    //
-    kUninitialized,  // no  no  no  no            Uninitialized or closed
-                     // object.
-    kFileStart,      // no  no  yes no            We just opened the .scp file
-                     // (we'll never be in this
-    //                           state when a user-visible function
-    //                           is called.)
-    kEof,    // no  no  no  no            We did Next() and found eof in script
-             // file.
-    kError,  // no  no  no  no            Error reading or parsing script file.
-    kHaveScpLine,  // no  no  yes yes           Have a line of the script file
-                   // but nothing else.
-    kHaveObject,   // yes no  yes yes           holder_ contains an object but
-                   // range_holder_ does not.
-    kHaveRange,    // yes yes yes yes           we have the range object in
-                   // range_holder_ (implies
-                   //                           range_ nonempty).
-  } state_;
-};
-
-// This is the implementation for SequentialTableReader
-// when it's an archive.  Note that the archive format is:
-// key1 [space] object1 key2 [space]
-// object2 ... eof.
-// "object1" is the output of the Holder::Write function and will
-// typically contain a binary header (in binary mode) and then
-// the output of object.Write(os, binary).
-// The archive itself does not care whether it is in binary
-// or text mode, for reading purposes.
-
-template <class Holder>
-class SequentialTableReaderArchiveImpl
-    : public SequentialTableReaderImplBase<Holder> {
- public:
-  typedef typename Holder::T T;
-
-  SequentialTableReaderArchiveImpl() : state_(kUninitialized) {}
-
-  virtual bool Open(const std::string &rspecifier) {
-    if (state_ != kUninitialized) {
-      if (!Close()) {  // call Close() yourself to suppress this exception.
-        if (opts_.permissive)
-          KALDIFST_WARN << "Error closing previous input "
-                           "(only warning, since permissive mode).";
-        else
-          KALDIFST_ERR << "Error closing previous input.";
-      }
-    }
-    rspecifier_ = rspecifier;
-    RspecifierType rs =
-        ClassifyRspecifier(rspecifier, &archive_rxfilename_, &opts_);
-    KALDIFST_ASSERT(rs == kArchiveRspecifier);
-
-    bool ans;
-    // NULL means don't expect binary-mode header
-    if (Holder::IsReadInBinary())
-      ans = input_.Open(archive_rxfilename_, NULL);
-    else
-      ans = input_.OpenTextMode(archive_rxfilename_);
-    if (!ans) {  // header.
-      KALDIFST_WARN << "Failed to open stream "
-                    << PrintableRxfilename(archive_rxfilename_);
-      state_ = kUninitialized;  // Failure on Open
-      return false;             // User should print the error message.
-    }
-    state_ = kFileStart;
-    Next();
-    if (state_ == kError) {
-      KALDIFST_WARN
-          << "Error beginning to read archive file (wrong filename?): "
-          << PrintableRxfilename(archive_rxfilename_);
-      input_.Close();
-      state_ = kUninitialized;
-      return false;
-    }
-    KALDIFST_ASSERT(state_ == kHaveObject || state_ == kEof);
-    return true;
-  }
-
-  virtual void Next() {
-    switch (state_) {
-      case kHaveObject:
-        holder_.Clear();
-        break;
-      case kFileStart:
-      case kFreedObject:
-        break;
-      default:
-        KALDIFST_ERR << "Next() called wrongly.";
-    }
-    std::istream &is = input_.Stream();
-    is.clear();  // Clear any fail bits that may have been set... just in case
-    // this happened in the Read function.
-    is >> key_;  // This eats up any leading whitespace and gets the string.
-    if (is.eof()) {
-      state_ = kEof;
-      return;
-    }
-    if (is.fail()) {  // This shouldn't really happen, barring file-system
-                      // errors.
-      KALDIFST_WARN << "Error reading archive "
-                    << PrintableRxfilename(archive_rxfilename_);
-      state_ = kError;
-      return;
-    }
-    int c;
-    if ((c = is.peek()) != ' ' && c != '\t' && c != '\n') {  // We expect a
-      // space ' ' after the key.
-      // We also allow tab [which is consumed] and newline [which is not], just
-      // so we can read archives generated by scripts that may not be fully
-      // aware of how this format works.
-      KALDIFST_WARN << "Invalid archive file format: expected space after key "
-                    << key_ << ", got character "
-                    << CharToString(static_cast<char>(is.peek()))
-                    << ", reading " << PrintableRxfilename(archive_rxfilename_);
-      state_ = kError;
-      return;
-    }
-    if (c != '\n') is.get();  // Consume the space or tab.
-    if (holder_.Read(is)) {
-      state_ = kHaveObject;
-      return;
-    } else {
-      KALDIFST_WARN << "Object read failed, reading archive "
-                    << PrintableRxfilename(archive_rxfilename_);
-      state_ = kError;
-      return;
-    }
-  }
-
-  virtual bool IsOpen() const {
-    switch (state_) {
-      case kEof:
-      case kError:
-      case kHaveObject:
-      case kFreedObject:
-        return true;
-      case kUninitialized:
-        return false;
-      default:
-        KALDIFST_ERR << "IsOpen() called on invalid object.";  // kFileStart
-        // is not valid state for user to call something on.
-        return false;
-    }
-  }
-
-  virtual bool Done() const {
-    switch (state_) {
-      case kHaveObject:
-        return false;
-      case kEof:
-      case kError:
-        return true;  // Error-state counts as Done(), but destructor
-        // will fail (unless you check the status with Close()).
-      default:
-        KALDIFST_ERR
-            << "Done() called on TableReader object at the wrong time.";
-        return false;
-    }
-  }
-
-  virtual std::string Key() {
-    // Valid to call this whenever Done() returns false
-    switch (state_) {
-      case kHaveObject:
-        break;  // only valid case.
-      default:
-        // coding error.
-        KALDIFST_ERR << "Key() called on TableReader object at the wrong time.";
-    }
-    return key_;
-  }
-
-  T &Value() {
-    switch (state_) {
-      case kHaveObject:
-        break;  // only valid case.
-      default:
-        // coding error.
-        KALDIFST_ERR
-            << "Value() called on TableReader object at the wrong time.";
-    }
-    return holder_.Value();
-  }
-
-  virtual void FreeCurrent() {
-    if (state_ == kHaveObject) {
-      holder_.Clear();
-      state_ = kFreedObject;
-    } else {
-      KALDIFST_WARN << "FreeCurrent called at the wrong time.";
-    }
-  }
-
-  void SwapHolder(Holder *other_holder) {
-    // call Value() to ensure we have a value, and ignore its return value while
-    // suppressing compiler warnings by casting to void.
-    (void)Value();
-    if (state_ == kHaveObject) {
-      holder_.Swap(other_holder);
-      state_ = kFreedObject;
-    } else {
-      KALDIFST_ERR << "SwapHolder called at the wrong time "
-                      "(error related to ',bg' modifier).";
-    }
-  }
-
-  virtual bool Close() {
-    // To clean up, Close() also closes the Input object if
-    // it's open.  It will succeed if the stream was not in an error state,
-    // and the Input object isn't in an error state  we've found eof in the
-    // archive.
-    if (!this->IsOpen())
-      KALDIFST_ERR
-          << "Close() called on TableReader twice or otherwise wrongly.";
-    int32_t status = 0;
-    if (input_.IsOpen()) status = input_.Close();
-    if (state_ == kHaveObject) holder_.Clear();
-    StateType old_state = state_;
-    state_ = kUninitialized;
-    if (old_state == kError || (old_state == kEof && status != 0)) {
-      if (opts_.permissive) {
-        KALDIFST_WARN << "Error detected closing TableReader for archive "
-                      << PrintableRxfilename(archive_rxfilename_)
-                      << " but ignoring "
-                      << "it as permissive mode specified.";
-        return true;
-      } else {
-        return false;
-      }
-    } else {
-      return true;
-    }
-  }
-
-  virtual ~SequentialTableReaderArchiveImpl() {
-    if (this->IsOpen() && !Close())
-      KALDIFST_ERR << "TableReader: error detected closing archive "
-                   << PrintableRxfilename(archive_rxfilename_);
-  }
-
- private:
-  Input input_;    // Input object for the archive
-  Holder holder_;  // Holds the object.
-  std::string key_;
-  std::string rspecifier_;
-  std::string archive_rxfilename_;
-  RspecifierOptions opts_;
-  enum StateType {  //  [The state of the reading process]        [does holder_
-                    //  [is input_
-    //                                                     have object]   open]
-    kUninitialized,  // Uninitialized or closed.                  no         no
-    kFileStart,      // [state we use internally: just opened.]   no         yes
-    kEof,    // We did Next() and found eof in archive           no         no
-    kError,  // Some other error                                 no         no
-    kHaveObject,  // We read the key and the object after it.     yes        yes
-    kFreedObject,  // The user called FreeCurrent().              no         yes
-  } state_;
-};
-
-// this is for when someone adds the 'th' modifier; it wraps around the basic
-// implementation and allows it to do the reading in a background thread.
-template <class Holder>
-class SequentialTableReaderBackgroundImpl
-    : public SequentialTableReaderImplBase<Holder> {
- public:
-  typedef typename Holder::T T;
-
-  SequentialTableReaderBackgroundImpl(
-      SequentialTableReaderImplBase<Holder> *base_reader)
-      : base_reader_(base_reader) {}
-
-  // This function ignores the rxfilename argument.
-  // We use the same function signature as the regular Open(),
-  // for convenience.
-  virtual bool Open(const std::string &rxfilename) {
-    KALDIFST_ASSERT(base_reader_ != NULL &&
-                    base_reader_->IsOpen());  // or code error.
-    {
-      thread_ =
-          std::thread(SequentialTableReaderBackgroundImpl<Holder>::run, this);
-    }
-
-    if (!base_reader_->Done()) Next();
-    return true;
-  }
-
-  virtual bool IsOpen() const {
-    // Close() sets base_reader_ to NULL, and we never initialize this object
-    // with a non-open base_reader_, so no need to check if it's open.
-    return base_reader_ != NULL;
-  }
-
-  void RunInBackground() {
-    try {
-      // This function is called in the background thread.  The whole point of
-      // the background thread is that we don't want to do the actual reading
-      // (inside Next()) in the foreground.
-      while (base_reader_ != NULL && !base_reader_->Done()) {
-        consumer_sem_.Signal();
-        // Here is where the consumer process (parent thread) gets to do its
-        // stuff.  Principally it calls SwapHolder()-- a shallow swap that is
-        // cheap.
-        producer_sem_.Wait();
-        // we check that base_reader_ is not NULL in case Close() was
-        // called in the main thread.
-        if (base_reader_ != NULL)
-          base_reader_->Next();  //  here is where the work happens.
-      }
-      // this signal will be waited on in the Next() function of the foreground
-      // thread if it is still running, or Close() otherwise.
-      consumer_sem_.Signal();
-      // this signal may be waited on in Close().
-      consumer_sem_.Signal();
-    } catch (...) {
-      // There is nothing we called above that could potentially throw due to
-      // user data.  So we treat reaching this point as a code-error condition.
-      // Closing base_reader_ will trigger an exception in Next() in the main
-      // thread when it checks that base_reader_->IsOpen().
-      if (base_reader_->IsOpen()) {
-        base_reader_->Close();
-        delete base_reader_;
-        base_reader_ = NULL;
-      }
-      consumer_sem_.Signal();
-      return;
-    }
-  }
-  static void run(SequentialTableReaderBackgroundImpl<Holder> *object) {
-    object->RunInBackground();
-  }
-  virtual bool Done() const { return key_.empty(); }
-  virtual std::string Key() {
-    if (key_.empty()) KALDIFST_ERR << "Calling Key() at the wrong time.";
-    return key_;
-  }
-  virtual T &Value() {
-    if (key_.empty()) KALDIFST_ERR << "Calling Value() at the wrong time.";
-    return holder_.Value();
-  }
-  void SwapHolder(Holder *other_holder) {
-    KALDIFST_ERR << "SwapHolder() should not be called on this class.";
-  }
-  virtual void FreeCurrent() {
-    if (key_.empty())
-      KALDIFST_ERR << "Calling FreeCurrent() at the wrong time.";
-    // note: ideally a call to Value() should crash if you have just called
-    // FreeCurrent().  For typical holders such as KaldiObjectHolder this will
-    // happen inside the holder_.Value() call.  This won't be the case for all
-    // holders, but it's not a great loss (just a missed opportunity to spot a
-    // code error).
-    holder_.Clear();
-  }
-  virtual void Next() {
-    consumer_sem_.Wait();
-    if (base_reader_ == NULL || !base_reader_->IsOpen())
-      KALDIFST_ERR << "Error detected (likely code error) in background "
-                   << "reader (',bg' option)";
-    if (base_reader_->Done()) {
-      // there is nothing else to read.
-      key_ = "";
-    } else {
-      key_ = base_reader_->Key();
-      base_reader_->SwapHolder(&holder_);
-    }
-    // this Signal() tells the producer thread, in the background,
-    // that it's now safe to read the next value.
-    producer_sem_.Signal();
-  }
-
-  // note: we can be sure that Close() won't be called twice, as the TableReader
-  // object will delete this object after calling Close.
-  virtual bool Close() {
-    KALDIFST_ASSERT(base_reader_ != NULL && thread_.joinable());
-    // wait until the producer thread is idle.
-    consumer_sem_.Wait();
-    bool ans = true;
-    try {
-      ans = base_reader_->Close();
-    } catch (...) {
-      ans = false;
-    }
-    delete base_reader_;
-    // setting base_reader_ to NULL will cause the loop in the producer thread
-    // to exit.
-    base_reader_ = NULL;
-    producer_sem_.Signal();
-
-    thread_.join();
-    return ans;
-  }
-  ~SequentialTableReaderBackgroundImpl() {
-    if (base_reader_) {
-      if (!Close()) {
-        KALDIFST_ERR << "Error detected closing background reader "
-                     << "(relates to ',bg' modifier)";
-      }
-    }
-  }
-
- private:
-  std::string key_;
-  Holder holder_;
-  // I couldn't figure out what to call these semaphores.  consumer_sem_ is the
-  // one that the consumer (main thread) waits on; producer_sem_ is the one
-  // that the producer (background thread) waits on.
-  Semaphore consumer_sem_;
-  Semaphore producer_sem_;
-  std::thread thread_;
-  SequentialTableReaderImplBase<Holder> *base_reader_;
-};
-
-template <class Holder>
-SequentialTableReader<Holder>::SequentialTableReader(
-    const std::string &rspecifier)
-    : impl_(NULL) {
-  if (rspecifier != "" && !Open(rspecifier))
-    KALDIFST_ERR << "Error constructing TableReader: rspecifier is "
-                 << rspecifier;
-}
-
-template <class Holder>
-bool SequentialTableReader<Holder>::Open(const std::string &rspecifier) {
-  if (IsOpen())
-    if (!Close()) KALDIFST_ERR << "Could not close previously open object.";
-  // now impl_ will be NULL.
-
-  RspecifierOptions opts;
-  RspecifierType wt = ClassifyRspecifier(rspecifier, NULL, &opts);
-  switch (wt) {
-    case kArchiveRspecifier:
-      impl_ = new SequentialTableReaderArchiveImpl<Holder>();
-      break;
-    case kScriptRspecifier:
-      impl_ = new SequentialTableReaderScriptImpl<Holder>();
-      break;
-    case kNoRspecifier:
-    default:
-      KALDIFST_WARN << "Invalid rspecifier " << rspecifier;
-      return false;
-  }
-  if (!impl_->Open(rspecifier)) {
-    delete impl_;
-    impl_ = NULL;
-    return false;  // sub-object will have printed warnings.
-  }
-  if (opts.background) {
-    impl_ = new SequentialTableReaderBackgroundImpl<Holder>(impl_);
-    if (!impl_->Open("")) {
-      // the rxfilename is ignored in that Open() call.
-      // It should only return false on code error.
-      return false;
-    }
-  }
-  return true;
-}
-
-template <class Holder>
-bool SequentialTableReader<Holder>::Close() {
-  CheckImpl();
-  bool ans = impl_->Close();
-  delete impl_;  // We don't keep around empty impl_ objects.
-  impl_ = NULL;
-  return ans;
-}
-
-template <class Holder>
-bool SequentialTableReader<Holder>::IsOpen() const {
-  return (impl_ != NULL);  // Because we delete the object whenever
-  // that object is not open.  Thus, the IsOpen functions of the
-  // Impl objects are not really needed.
-}
-
-template <class Holder>
-std::string SequentialTableReader<Holder>::Key() {
-  CheckImpl();
-  return impl_->Key();  // this call may throw if called wrongly in other ways,
-  // e.g. eof.
-}
-
-template <class Holder>
-void SequentialTableReader<Holder>::FreeCurrent() {
-  CheckImpl();
-  impl_->FreeCurrent();
-}
-
-template <class Holder>
-typename SequentialTableReader<Holder>::T &
-SequentialTableReader<Holder>::Value() {
-  CheckImpl();
-  return impl_->Value();  // This may throw (if EnsureObjectLoaded() returned
-                          // false you are safe.).
-}
-
-template <class Holder>
-void SequentialTableReader<Holder>::Next() {
-  CheckImpl();
-  impl_->Next();
-}
-
-template <class Holder>
-bool SequentialTableReader<Holder>::Done() {
-  CheckImpl();
-  return impl_->Done();
-}
-
-template <class Holder>
-SequentialTableReader<Holder>::~SequentialTableReader() {
-  delete impl_;
-  // Destructor of impl_ may throw.
-}
-
-template <class Holder>
-class TableWriterImplBase {
- public:
-  typedef typename Holder::T T;
-
-  virtual bool Open(const std::string &wspecifier) = 0;
-
-  // Write returns true on success, false on failure, but
-  // some errors may not be detected until we call Close().
-  // It throws (via KALDIFST_ERR) if called wrongly.  We could
-  // have just thrown on all errors, since this is what
-  // TableWriter does; it was designed this way because originally
-  // TableWriter::Write returned an exit status.
-  virtual bool Write(const std::string &key, const T &value) = 0;
-
-  // Flush will flush any archive; it does not return error status,
-  //  any errors will be reported on the next Write or Close.
-  virtual void Flush() = 0;
-
-  virtual bool Close() = 0;
-
-  virtual bool IsOpen() const = 0;
-
-  // May throw on write error if Close was not called.
-  virtual ~TableWriterImplBase() {}
-
-  TableWriterImplBase() {}
-
- private:
-  KALDIFST_DISALLOW_COPY_AND_ASSIGN(TableWriterImplBase);
-};
-
-// The implementation of TableWriter we use when writing directly
-// to an archive with no associated scp.
-template <class Holder>
-class TableWriterArchiveImpl : public TableWriterImplBase<Holder> {
- public:
-  typedef typename Holder::T T;
-
-  virtual bool Open(const std::string &wspecifier) {
-    switch (state_) {
-      case kUninitialized:
-        break;
-      case kWriteError:
-        KALDIFST_ERR << "Opening stream, already open with write error.";
-      case kOpen:
-      default:
-        if (!Close())  // throw because this error may not have been previously
-          // detected by the user.
-          KALDIFST_ERR
-              << "Opening stream, error closing previously open stream.";
-    }
-    wspecifier_ = wspecifier;
-    WspecifierType ws =
-        ClassifyWspecifier(wspecifier, &archive_wxfilename_, NULL, &opts_);
-    KALDIFST_ASSERT(ws == kArchiveWspecifier);  // or wrongly called.
-
-    if (output_.Open(archive_wxfilename_, opts_.binary, false)) {  // false
-      // means no binary header.
-      state_ = kOpen;
-      return true;
-    } else {
-      // stream will not be open.  User will report this error
-      // (we return bool), so don't bother printing anything.
-      state_ = kUninitialized;
-      return false;
-    }
-  }
-
-  virtual bool IsOpen() const {
-    switch (state_) {
-      case kUninitialized:
-        return false;
-      case kOpen:
-      case kWriteError:
-        return true;
-      default:
-        KALDIFST_ERR << "IsOpen() called on TableWriter in invalid state.";
-    }
-    return false;
-  }
-
-  // Write returns true on success, false on failure, but
-  // some errors may not be detected till we call Close().
-  virtual bool Write(const std::string &key, const T &value) {
-    switch (state_) {
-      case kOpen:
-        break;
-      case kWriteError:
-        // user should have known from the last
-        // call to Write that there was a problem.
-        KALDIFST_WARN << "Attempting to write to invalid stream.";
-        return false;
-      case kUninitialized:
-      default:
-        KALDIFST_ERR << "Write called on invalid stream";
-    }
-    // state is now kOpen or kWriteError.
-    if (!IsToken(key))  // e.g. empty string or has spaces...
-      KALDIFST_ERR << "Using invalid key " << key;
-    output_.Stream() << key << ' ';
-    if (!Holder::Write(output_.Stream(), opts_.binary, value)) {
-      KALDIFST_WARN << "Write failure to "
-                    << PrintableWxfilename(archive_wxfilename_);
-      state_ = kWriteError;
-      return false;
-    }
-    if (state_ == kWriteError) return false;  // Even if this Write seems to
-    // have succeeded, we fail because a previous Write failed and the archive
-    // may be corrupted and unreadable.
-
-    if (opts_.flush) Flush();
-    return true;
-  }
-
-  // Flush will flush any archive; it does not return error status,
-  //  any errors will be reported on the next Write or Close.
-  virtual void Flush() {
-    switch (state_) {
-      case kWriteError:
-      case kOpen:
-        output_.Stream().flush();  // Don't check error status.
-        return;
-      default:
-        KALDIFST_WARN << "Flush called on not-open writer.";
-    }
-  }
-
-  virtual bool Close() {
-    if (!this->IsOpen() || !output_.IsOpen())
-      KALDIFST_ERR << "Close called on a stream that was not open."
-                   << this->IsOpen() << ", " << output_.IsOpen();
-    bool close_success = output_.Close();
-    if (!close_success) {
-      KALDIFST_WARN << "Error closing stream: wspecifier is " << wspecifier_;
-      state_ = kUninitialized;
-      return false;
-    }
-    if (state_ == kWriteError) {
-      KALDIFST_WARN << "Closing writer in error state: wspecifier is "
-                    << wspecifier_;
-      state_ = kUninitialized;
-      return false;
-    }
-    state_ = kUninitialized;
-    return true;
-  }
-
-  TableWriterArchiveImpl() : state_(kUninitialized) {}
-
-  // May throw on write error if Close was not called.
-  virtual ~TableWriterArchiveImpl() {
-    if (!IsOpen())
-      return;
-    else if (!Close())
-      KALDIFST_ERR << "At TableWriter destructor: Write failed or stream close "
-                   << "failed: wspecifier is " << wspecifier_;
-  }
-
- private:
-  Output output_;
-  WspecifierOptions opts_;
-  std::string wspecifier_;
-  std::string archive_wxfilename_;
-  enum {             // is stream open?
-    kUninitialized,  // no
-    kOpen,           // yes
-    kWriteError,     // yes
-  } state_;
-};
-
-// The implementation of TableWriter we use when writing to
-// individual files (more generally, wxfilenames) specified
-// in an scp file that we read.
-
-// Note: the code for this class is similar to
-// RandomAccessTableReaderScriptImpl; try to keep them in sync.
-
-template <class Holder>
-class TableWriterScriptImpl : public TableWriterImplBase<Holder> {
- public:
-  typedef typename Holder::T T;
-
-  TableWriterScriptImpl() : last_found_(0), state_(kUninitialized) {}
-
-  virtual bool Open(const std::string &wspecifier) {
-    switch (state_) {
-      case kReadScript:
-        KALDIFST_ERR << " Opening already open TableWriter: call Close first.";
-      case kUninitialized:
-      case kNotReadScript:
-        break;
-    }
-    wspecifier_ = wspecifier;
-    WspecifierType ws =
-        ClassifyWspecifier(wspecifier, NULL, &script_rxfilename_, &opts_);
-    KALDIFST_ASSERT(ws == kScriptWspecifier);  // or wrongly called.
-    KALDIFST_ASSERT(
-        script_.empty());  // no way it could be nonempty at this point.
-
-    if (!ReadScriptFile(script_rxfilename_,
-                        true,         // print any warnings
-                        &script_)) {  // error reading script file or invalid
-                                      // format
-      state_ = kNotReadScript;
-      return false;  // no need to print further warnings.  user gets the error.
-    }
-    std::sort(script_.begin(), script_.end());
-    for (size_t i = 0; i + 1 < script_.size(); i++) {
-      if (script_[i].first.compare(script_[i + 1].first) >= 0) {
-        // script[i] not < script[i+1] in lexical order...
-        KALDIFST_WARN << "Script file "
-                      << PrintableRxfilename(script_rxfilename_)
-                      << " contains duplicate key " << script_[i].first;
-        state_ = kNotReadScript;
-        return false;
-      }
-    }
-    state_ = kReadScript;
-    return true;
-  }
-
-  virtual bool IsOpen() const { return (state_ == kReadScript); }
-
-  virtual bool Close() {
-    if (!IsOpen())
-      KALDIFST_ERR << "Close() called on TableWriter that was not open.";
-    state_ = kUninitialized;
-    last_found_ = 0;
-    script_.clear();
-    return true;
-  }
-
-  // Write returns true on success, false on failure, but
-  // some errors may not be detected till we call Close().
-  virtual bool Write(const std::string &key, const T &value) {
-    if (!IsOpen()) KALDIFST_ERR << "Write called on invalid stream";
-
-    if (!IsToken(key))  // e.g. empty string or has spaces...
-      KALDIFST_ERR << "Using invalid key " << key;
-
-    std::string wxfilename;
-    if (!LookupFilename(key, &wxfilename)) {
-      if (opts_.permissive) {
-        return true;  // In permissive mode, it's as if we're writing to
-                      // /dev/null for missing keys.
-      } else {
-        KALDIFST_WARN << "Script file "
-                      << PrintableRxfilename(script_rxfilename_)
-                      << " has no entry for key " << key;
-        return false;
-      }
-    }
-    Output output;
-    if (!output.Open(wxfilename, opts_.binary, false)) {
-      // Open in the text/binary mode (on Windows) given by member var. "binary"
-      // (obtained from wspecifier), but do not put the binary-mode header (it
-      // will be written, if needed, by the Holder::Write function.)
-      KALDIFST_WARN << "Failed to open stream: "
-                    << PrintableWxfilename(wxfilename);
-      return false;
-    }
-    if (!Holder::Write(output.Stream(), opts_.binary, value) ||
-        !output.Close()) {
-      KALDIFST_WARN << "Failed to write data to "
-                    << PrintableWxfilename(wxfilename);
-
-      return false;
-    }
-    return true;
-  }
-
-  // Flush does nothing in this implementation, there is nothing to flush.
-  virtual void Flush() {}
-
-  virtual ~TableWriterScriptImpl() {
-    // Nothing to do in destructor.
-  }
-
- private:
-  // Note: this function is almost the same as in
-  // RandomAccessTableReaderScriptImpl.
-  bool LookupFilename(const std::string &key, std::string *wxfilename) {
-    // First, an optimization: if we're going consecutively, this will
-    // make the lookup very fast.
-    last_found_++;
-    if (last_found_ < script_.size() && script_[last_found_].first == key) {
-      *wxfilename = script_[last_found_].second;
-      return true;
-    }
-    std::pair<std::string, std::string> pr(key, "");  // Important that ""
-    // compares less than or equal to any string, so lower_bound points to the
-    // element that has the same key.
-    typedef typename std::vector<
-        std::pair<std::string, std::string>>::const_iterator IterType;
-    IterType iter = std::lower_bound(script_.begin(), script_.end(), pr);
-    if (iter != script_.end() && iter->first == key) {
-      last_found_ = iter - script_.begin();
-      *wxfilename = iter->second;
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  WspecifierOptions opts_;
-  std::string wspecifier_;
-  std::string script_rxfilename_;
-
-  // the script_ variable contains pairs of (key, filename), sorted using
-  // std::sort.  This can be used with binary_search to look up filenames for
-  // writing.  If this becomes inefficient we can use std::unordered_map (but I
-  // suspect this wouldn't be significantly faster & would use more memory).
-  // If memory becomes a problem here, the user should probably be passing
-  // only the relevant part of the scp file rather than expecting us to get too
-  // clever in the code.
-  std::vector<std::pair<std::string, std::string>> script_;
-  size_t last_found_;  // This is for an optimization used in LookupFilename.
-
-  enum {
-    kUninitialized,
-    kReadScript,
-    kNotReadScript,  // read of script failed.
-  } state_;
-};
-
-// The implementation of TableWriter we use when writing directly
-// to an archive plus an associated scp.
-template <class Holder>
-class TableWriterBothImpl : public TableWriterImplBase<Holder> {
- public:
-  typedef typename Holder::T T;
-
-  virtual bool Open(const std::string &wspecifier) {
-    switch (state_) {
-      case kUninitialized:
-        break;
-      case kWriteError:
-        KALDIFST_ERR << "Opening stream, already open with write error.";
-      case kOpen:
-      default:
-        if (!Close())  // throw because this error may not have been previously
-                       // detected by user.
-          KALDIFST_ERR
-              << "Opening stream, error closing previously open stream.";
-    }
-    wspecifier_ = wspecifier;
-    WspecifierType ws = ClassifyWspecifier(wspecifier, &archive_wxfilename_,
-                                           &script_wxfilename_, &opts_);
-    KALDIFST_ASSERT(ws == kBothWspecifier);  // or wrongly called.
-    if (ClassifyWxfilename(archive_wxfilename_) != kFileOutput)
-      KALDIFST_WARN
-          << "When writing to both archive and script, the script file "
-             "will generally not be interpreted correctly unless the archive "
-             "is "
-             "an actual file: wspecifier = "
-          << wspecifier;
-
-    if (!archive_output_.Open(archive_wxfilename_, opts_.binary, false)) {
-      // false means no binary header.
-      state_ = kUninitialized;
-      return false;
-    }
-    if (!script_output_.Open(script_wxfilename_, false, false)) {  // first
-      // false means text mode: script files always text-mode.   second false
-      //  means don't write header (doesn't matter for text mode).
-      archive_output_.Close();  // Don't care about status: error anyway.
-      state_ = kUninitialized;
-      return false;
-    }
-    state_ = kOpen;
-    return true;
-  }
-
-  virtual bool IsOpen() const {
-    switch (state_) {
-      case kUninitialized:
-        return false;
-      case kOpen:
-      case kWriteError:
-        return true;
-      default:
-        KALDIFST_ERR << "IsOpen() called on TableWriter in invalid state.";
-    }
-    return false;
-  }
-
-  void MakeFilename(typename std::ostream::pos_type streampos,
-                    std::string *output) const {
-    std::ostringstream ss;
-    ss << ':' << streampos;
-    KALDIFST_ASSERT(ss.str() != ":-1");
-    *output = archive_wxfilename_ + ss.str();
-
-    // e.g. /some/file:12302.
-    // Note that we warned if archive_wxfilename_ is not an actual filename;
-    // the philosophy is we give the user rope and if they want to hang
-    // themselves, with it, fine.
-  }
-
-  // Write returns true on success, false on failure, but
-  // some errors may not be detected till we call Close().
-  virtual bool Write(const std::string &key, const T &value) {
-    switch (state_) {
-      case kOpen:
-        break;
-      case kWriteError:
-        // user should have known from the last
-        // call to Write that there was a problem.  Warn about it.
-        KALDIFST_WARN << "Writing to non-open TableWriter object.";
-        return false;
-      case kUninitialized:
-      default:
-        KALDIFST_ERR << "Write called on invalid stream";
-    }
-    // state is now kOpen or kWriteError.
-    if (!IsToken(key))  // e.g. empty string or has spaces...
-      KALDIFST_ERR << "Using invalid key " << key;
-    std::ostream &archive_os = archive_output_.Stream();
-    archive_os << key << ' ';
-    typename std::ostream::pos_type archive_os_pos = archive_os.tellp();
-    // position at start of Write() to archive.  We will record this in the
-    // script file.
-    std::string offset_rxfilename;  // rxfilename with offset into the archive,
-    // e.g. some_archive_name.ark:431541423
-    MakeFilename(archive_os_pos, &offset_rxfilename);
-
-    // Write to the script file first.
-    // The idea is that we want to get all the information possible into the
-    // script file, to make it easier to unwind errors later.
-    std::ostream &script_os = script_output_.Stream();
-    script_output_.Stream() << key << ' ' << offset_rxfilename << '\n';
-
-    if (!Holder::Write(archive_output_.Stream(), opts_.binary, value)) {
-      KALDIFST_WARN << "Write failure to"
-                    << PrintableWxfilename(archive_wxfilename_);
-      state_ = kWriteError;
-      return false;
-    }
-
-    if (script_os.fail()) {
-      KALDIFST_WARN << "Write failure to script file detected: "
-                    << PrintableWxfilename(script_wxfilename_);
-      state_ = kWriteError;
-      return false;
-    }
-
-    if (archive_os.fail()) {
-      KALDIFST_WARN << "Write failure to archive file detected: "
-                    << PrintableWxfilename(archive_wxfilename_);
-      state_ = kWriteError;
-      return false;
-    }
-
-    if (state_ == kWriteError) return false;  // Even if this Write seems to
-    // have succeeded, we fail because a previous Write failed and the archive
-    // may be corrupted and unreadable.
-
-    if (opts_.flush) Flush();
-    return true;
-  }
-
-  // Flush will flush any archive; it does not return error status,
-  //  any errors will be reported on the next Write or Close.
-  virtual void Flush() {
-    switch (state_) {
-      case kWriteError:
-      case kOpen:
-        archive_output_.Stream().flush();  // Don't check error status.
-        script_output_.Stream().flush();   // Don't check error status.
-        return;
-      default:
-        KALDIFST_WARN << "Flush called on not-open writer.";
-    }
-  }
-
-  virtual bool Close() {
-    if (!this->IsOpen())
-      KALDIFST_ERR << "Close called on a stream that was not open.";
-    bool close_success = true;
-    if (archive_output_.IsOpen())
-      if (!archive_output_.Close()) close_success = false;
-    if (script_output_.IsOpen())
-      if (!script_output_.Close()) close_success = false;
-    bool ans = close_success && (state_ != kWriteError);
-    state_ = kUninitialized;
-    return ans;
-  }
-
-  TableWriterBothImpl() : state_(kUninitialized) {}
-
-  // May throw on write error if Close() was not called.
-  // User can get the error status by calling Close().
-  virtual ~TableWriterBothImpl() {
-    if (!IsOpen())
-      return;
-    else if (!Close())
-      KALDIFST_ERR << "Write failed or stream close failed: " << wspecifier_;
-  }
-
- private:
-  Output archive_output_;
-  Output script_output_;
-  WspecifierOptions opts_;
-  std::string archive_wxfilename_;
-  std::string script_wxfilename_;
-  std::string wspecifier_;
-  enum {             // is stream open?
-    kUninitialized,  // no
-    kOpen,           // yes
-    kWriteError,     // yes
-  } state_;
-};
-
-template <class Holder>
-TableWriter<Holder>::TableWriter(const std::string &wspecifier) : impl_(NULL) {
-  if (wspecifier != "" && !Open(wspecifier))
-    KALDIFST_ERR << "Failed to open table for writing with wspecifier: "
-                 << wspecifier
-                 << ": errno (in case it's relevant) is: " << strerror(errno);
-}
-
-template <class Holder>
-bool TableWriter<Holder>::IsOpen() const {
-  return (impl_ != NULL);
-}
-
-template <class Holder>
-bool TableWriter<Holder>::Open(const std::string &wspecifier) {
-  if (IsOpen()) {
-    if (!Close())  // call Close() yourself to suppress this exception.
-      KALDIFST_ERR << "Failed to close previously open writer.";
-  }
-  KALDIFST_ASSERT(impl_ == NULL);
-  WspecifierType wtype = ClassifyWspecifier(wspecifier, NULL, NULL, NULL);
-  switch (wtype) {
-    case kBothWspecifier:
-      impl_ = new TableWriterBothImpl<Holder>();
-      break;
-    case kArchiveWspecifier:
-      impl_ = new TableWriterArchiveImpl<Holder>();
-      break;
-    case kScriptWspecifier:
-      impl_ = new TableWriterScriptImpl<Holder>();
-      break;
-    case kNoWspecifier:
-    default:
-      KALDIFST_WARN << "ClassifyWspecifier: invalid wspecifier " << wspecifier;
-      return false;
-  }
-  if (impl_->Open(wspecifier)) {
-    return true;
-  } else {  // The class will have printed a more specific warning.
-    delete impl_;
-    impl_ = NULL;
-    return false;
-  }
-}
-
-template <class Holder>
-void TableWriter<Holder>::Write(const std::string &key, const T &value) const {
-  CheckImpl();
-  if (!impl_->Write(key, value)) KALDIFST_ERR << "Error in TableWriter::Write";
-  // More specific warning will have
-  // been printed in the Write function.
-}
-
-template <class Holder>
-void TableWriter<Holder>::Flush() {
-  CheckImpl();
-  impl_->Flush();
-}
-
-template <class Holder>
-bool TableWriter<Holder>::Close() {
-  CheckImpl();
-  bool ans = impl_->Close();
-  delete impl_;  // We don't keep around non-open impl_ objects
-                 // [c.f. definition of IsOpen()]
-  impl_ = NULL;
-  return ans;
-}
-
-template <class Holder>
-TableWriter<Holder>::~TableWriter() {
-  if (IsOpen() && !Close()) {
-    KALDIFST_ERR << "Error closing TableWriter [in destructor].";
-  }
-}
-
-// Types of RandomAccessTableReader:
-// In principle, we would like to have four types of RandomAccessTableReader:
-//  the 4 combinations  [scp, archive], [seekable, not-seekable],
-// where if something is seekable we only store a file offset.  However,
-// it seems sufficient for now to only implement two of these, in both
-// cases assuming it's not seekable so we never store file offsets and always
-// store either the scp line or the data in the archive.  The reasons are:
-// (1)
-// For scp files, storing the actual entry is not that much more expensive
-// than storing the file offsets (since the entries are just filenames), and
-// avoids a lot of fseek operations that might be expensive.
-// (2)
-// For archive files, there is no real reason, if you have the archive file
-// on disk somewhere, why you wouldn't access it via its associated scp.
-// [i.e. write it as ark, scp].  The main reason to read archives directly
-// is if they are part of a pipe, and in this case it's not seekable, so
-// we implement only this case.
-//
-// Note that we will rarely in practice have to keep in memory everything in
-// the archive, as long as things are only read once from the archive (the
-// "o, " or "once" option) and as long as we keep our keys in sorted order;
-// to take advantage of this we need the "s, " (sorted) option, so we would
-// read archives as e.g. "s, o, ark:-" (this is the rspecifier we would use if
-// it was the standard input and these conditions held).
-
-template <class Holder>
-class RandomAccessTableReaderImplBase {
- public:
-  typedef typename Holder::T T;
-
-  virtual bool Open(const std::string &rspecifier) = 0;
-
-  virtual bool HasKey(const std::string &key) = 0;
-
-  virtual const T &Value(const std::string &key) = 0;
-
-  virtual bool Close() = 0;
-
-  virtual ~RandomAccessTableReaderImplBase() {}
-};
-
-// Implementation of RandomAccessTableReader for a script file; for simplicity
-// we just read it in all in one go, as it's unlikely someone would generate
-// this from a pipe.  In principle we could read it on-demand as for the
-// archives, but this would probably be overkill.
-
-// Note: the code for this this class is similar to TableWriterScriptImpl:
-// try to keep them in sync.
-template <class Holder>
-class RandomAccessTableReaderScriptImpl
-    : public RandomAccessTableReaderImplBase<Holder> {
- public:
-  typedef typename Holder::T T;
-
-  RandomAccessTableReaderScriptImpl()
-      : last_found_(0), state_(kUninitialized) {}
-
-  virtual bool Open(const std::string &rspecifier) {
-    switch (state_) {
-      case kNotHaveObject:
-      case kHaveObject:
-      case kHaveRange:
-        KALDIFST_ERR << " Opening already open RandomAccessTableReader:"
-                        " call Close first.";
-      case kUninitialized:
-      case kNotReadScript:
-        break;
-    }
-    rspecifier_ = rspecifier;
-    RspecifierType rs =
-        ClassifyRspecifier(rspecifier, &script_rxfilename_, &opts_);
-    KALDIFST_ASSERT(rs == kScriptRspecifier);  // or wrongly called.
-    KALDIFST_ASSERT(
-        script_.empty());  // no way it could be nonempty at this point
-
-    if (!ReadScriptFile(script_rxfilename_,
-                        true,         // print any warnings
-                        &script_)) {  // error reading script file or invalid
-                                      // format
-      state_ = kNotReadScript;
-      return false;  // no need to print further warnings.  user gets the error.
-    }
-
-    rspecifier_ = rspecifier;
-    // If opts_.sorted, the user has asserted that the keys are already sorted.
-    // Although we could easily sort them, we want to let the user know of this
-    // mistake.  This same mistake could have serious effects if used with an
-    // archive rather than a script.
-    if (!opts_.sorted) std::sort(script_.begin(), script_.end());
-    for (size_t i = 0; i + 1 < script_.size(); i++) {
-      if (script_[i].first.compare(script_[i + 1].first) >= 0) {
-        // script[i] not < script[i+1] in lexical order...
-        bool same = (script_[i].first == script_[i + 1].first);
-        KALDIFST_WARN
-            << "Script file " << PrintableRxfilename(script_rxfilename_)
-            << (same ? " contains duplicate key: "
-                     : " is not sorted (remove s, option or add ns, option):"
-                       " key is ")
-            << script_[i].first;
-        state_ = kNotReadScript;
-        return false;
-      }
-    }
-    state_ = kNotHaveObject;
-    key_ = "";  // make sure we don't have a key set
-    return true;
-  }
-
-  virtual bool IsOpen() const {
-    return (state_ == kNotHaveObject || state_ == kHaveObject ||
-            state_ == kHaveRange);
-  }
-
-  virtual bool Close() {
-    if (!IsOpen())
-      KALDIFST_ERR << "Close() called on RandomAccessTableReader that was not"
-                      " open.";
-    holder_.Clear();
-    range_holder_.Clear();
-    state_ = kUninitialized;
-    last_found_ = 0;
-    script_.clear();
-    key_ = "";
-    range_ = "";
-    data_rxfilename_ = "";
-    // This cannot fail because any errors of a "global" nature would have been
-    // detected when we did Open().  With archives it's different.
-    return true;
-  }
-
-  virtual bool HasKey(const std::string &key) {
-    bool preload = opts_.permissive;
-    // In permissive mode, we have to check that we can read
-    // the scp entry before we assert that the key is there.
-    return HasKeyInternal(key, preload);
-  }
-
-  // Write returns true on success, false on failure, but
-  // some errors may not be detected till we call Close().
-  virtual const T &Value(const std::string &key) {
-    if (!HasKeyInternal(key, true))  // true == preload.
-      KALDIFST_ERR << "Could not get item for key " << key << ", rspecifier is "
-                   << rspecifier_ << " [to ignore this, "
-                   << "add the p, (permissive) option to the rspecifier.";
-    KALDIFST_ASSERT(key_ == key);
-    if (state_ == kHaveObject) {
-      return holder_.Value();
-    } else {
-      KALDIFST_ASSERT(state_ == kHaveRange);
-      return range_holder_.Value();
-    }
-  }
-
-  virtual ~RandomAccessTableReaderScriptImpl() {}
-
- private:
-  // HasKeyInternal when called with preload == false just tells us whether the
-  // key is in the scp.  With preload == true, which happens when the ,p
-  // (permissive) option is given in the rspecifier (or when called from
-  // Value()), it will also check that we can preload the object from disk
-  // (loading from the rxfilename in the scp), and only return true if we can.
-  // This function is called both from HasKey and from Value().
-  virtual bool HasKeyInternal(const std::string &key, bool preload) {
-    switch (state_) {
-      case kUninitialized:
-      case kNotReadScript:
-        KALDIFST_ERR
-            << "HasKey called on RandomAccessTableReader object that is"
-               " not open.";
-      case kHaveObject:
-        if (key == key_ && range_.empty()) return true;
-        break;
-      case kHaveRange:
-        if (key == key_) return true;
-        break;
-      case kNotHaveObject:
-      default:
-        break;
-    }
-    KALDIFST_ASSERT(IsToken(key));
-    size_t key_pos = 0;
-    if (!LookupKey(key, &key_pos)) {
-      return false;
-    } else {
-      if (!preload) {
-        return true;  // we have the key, and were not asked to verify that the
-                      // object could be read.
-      } else {  // preload specified, so we have to attempt to pre-load the
-                // object before returning.
-        std::string data_rxfilename, range;  // We will split
-        // script_[key_pos].second (e.g. "1.ark:100[0:2]" into data_rxfilename
-        // (e.g. "1.ark:100") and range (if any), e.g. "0:2".
-        if (script_[key_pos].second[script_[key_pos].second.size() - 1] ==
-            ']') {
-          if (!ExtractRangeSpecifier(script_[key_pos].second, &data_rxfilename,
-                                     &range)) {
-            KALDIFST_ERR << "TableReader: failed to parse range in '"
-                         << script_[key_pos].second << "'";
-          }
-        } else {
-          data_rxfilename = script_[key_pos].second;
-        }
-        if (state_ == kHaveRange) {
-          if (data_rxfilename_ == data_rxfilename && range_ == range) {
-            // the odd situation where two keys had the same rxfilename and
-            // range: just change the key and keep the object.
-            key_ = key;
-            return true;
-          } else {
-            range_holder_.Clear();
-            state_ = kHaveObject;
-          }
-        }
-        // OK, at this point the state will be kHaveObject or kNotHaveObject.
-        if (state_ == kHaveObject) {
-          if (data_rxfilename_ != data_rxfilename) {
-            // clear out the object.
-            state_ = kNotHaveObject;
-            holder_.Clear();
-          }
-        }
-        // At this point we can safely switch to the new key, data_rxfilename
-        // and range, and we know that if we have an object, it will already be
-        // the correct one.  The state is now kHaveObject or kNotHaveObject.
-        key_ = key;
-        data_rxfilename_ = data_rxfilename;
-        range_ = range;
-        if (state_ == kNotHaveObject) {
-          // we need to read the object.
-          if (!input_.Open(data_rxfilename)) {
-            KALDIFST_WARN << "Error opening stream "
-                          << PrintableRxfilename(data_rxfilename);
-            return false;
-          } else {
-            if (holder_.Read(input_.Stream())) {
-              state_ = kHaveObject;
-            } else {
-              KALDIFST_WARN << "Error reading object from "
-                               "stream "
-                            << PrintableRxfilename(data_rxfilename);
-              return false;
-            }
-          }
-        }
-        // At this point the state is kHaveObject.
-        if (range.empty()) return true;  // we're done: no range was requested.
-        if (range_holder_.ExtractRange(holder_, range)) {
-          state_ = kHaveRange;
-          return true;
-        } else {
-          KALDIFST_WARN << "Failed to load object from "
-                        << PrintableRxfilename(data_rxfilename) << "[" << range
-                        << "]";
-          // leave state at kHaveObject.
-          return false;
-        }
-      }
-    }
-  }
-
-  // This function attempts to look up the key "key" in the sorted array
-  // script_.  If it was found it returns true and puts the array offset into
-  // 'script_offset'; otherwise it returns false.
-  bool LookupKey(const std::string &key, size_t *script_offset) {
-    // First, an optimization: if we're going consecutively, this will
-    // make the lookup very fast.  Since we may call HasKey and then
-    // Value(), which both may look up the key, we test if either the
-    // current or next position are correct.
-    if (last_found_ < script_.size() && script_[last_found_].first == key) {
-      *script_offset = last_found_;
-      return true;
-    }
-    last_found_++;
-    if (last_found_ < script_.size() && script_[last_found_].first == key) {
-      *script_offset = last_found_;
-      return true;
-    }
-    std::pair<std::string, std::string> pr(key, "");  // Important that ""
-    // compares less than or equal to any string, so lower_bound points to the
-    // element that has the same key.
-    typedef typename std::vector<
-        std::pair<std::string, std::string>>::const_iterator IterType;
-    IterType iter = std::lower_bound(script_.begin(), script_.end(), pr);
-    if (iter != script_.end() && iter->first == key) {
-      last_found_ = *script_offset = iter - script_.begin();
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  Input input_;  // Use the same input_ object for reading each file, in case
-                 // the scp specifies offsets in an archive so we can keep the
-                 // same file open.
-  RspecifierOptions opts_;
-  std::string rspecifier_;  // rspecifier used to open this object; used in
-                            // debug messages
-  std::string script_rxfilename_;  // rxfilename of script file that we read.
-
-  std::string key_;  // The current key of the object that we have, but see the
-                     // notes regarding states_ for more explanation of the
-                     // semantics.
-
-  Holder holder_;
-  Holder range_holder_;  // Holds the partial object corresponding to the object
-                         // range specifier 'range_'. this is only used when
-                         // 'range_' is specified.
-  std::string range_;    // range within which we read the object from holder_.
-                         // If key_ is set, always correspond to the key.
-  std::string data_rxfilename_;  // the rxfilename corresponding to key_,
-                                 // always set when key_ is set.
-
-  // the script_ variable contains pairs of (key, filename), sorted using
-  // std::sort.  This can be used with binary_search to look up filenames for
-  // writing.  If this becomes inefficient we can use std::unordered_map (but I
-  // suspect this wouldn't be significantly faster & would use more memory).
-  // If memory becomes a problem here, the user should probably be passing
-  // only the relevant part of the scp file rather than expecting us to get too
-  // clever in the code.
-  std::vector<std::pair<std::string, std::string>> script_;
-  size_t last_found_;  // This is for an optimization used in FindFilename.
-
-  enum {
-    //                   (*) is script_ set up?
-    //                          (*) does holder_ contain an object?
-    //                               (*) does range_holder_ contain and object?
-    //
-    //
-    kUninitialized,  //    no    no    no
-    kNotReadScript,  //    no    no    no
-    kNotHaveObject,  //    yes   no    no
-    kHaveObject,     //    yes   yes   no
-    kHaveRange,      //    yes   yes   yes
-
-    // If we are in a state where holder_ contains an object, it always contains
-    // the object from 'key_', and the corresponding rxfilename is always
-    // 'data_rxfilename_'.  If range_holder_ contains an object, it always
-    // corresponds to the range 'range_' of the object in 'holder_', and always
-    // corresponds to the current key.
-  } state_;
-};
-
-// This is the base-class (with some implemented functions) for the
-// implementations of RandomAccessTableReader when it's an archive.  This
-// base-class handles opening the files, storing the state of the reading
-// process, and loading objects.  This is the only case in which we have
-// an intermediate class in the hierarchy between the virtual ImplBase
-// class and the actual Impl classes.
-// The child classes vary in the assumptions regarding sorting, etc.
-
-template <class Holder>
-class RandomAccessTableReaderArchiveImplBase
-    : public RandomAccessTableReaderImplBase<Holder> {
- public:
-  typedef typename Holder::T T;
-
-  RandomAccessTableReaderArchiveImplBase()
-      : holder_(NULL), state_(kUninitialized) {}
-
-  virtual bool Open(const std::string &rspecifier) {
-    if (state_ != kUninitialized) {
-      if (!this->Close())  // call Close() yourself to suppress this exception.
-        KALDIFST_ERR << "Error closing previous input.";
-    }
-    rspecifier_ = rspecifier;
-    RspecifierType rs =
-        ClassifyRspecifier(rspecifier, &archive_rxfilename_, &opts_);
-    KALDIFST_ASSERT(rs == kArchiveRspecifier);
-
-    // NULL means don't expect binary-mode header
-    bool ans;
-    if (Holder::IsReadInBinary())
-      ans = input_.Open(archive_rxfilename_, NULL);
-    else
-      ans = input_.OpenTextMode(archive_rxfilename_);
-    if (!ans) {  // header.
-      KALDIFST_WARN << "Failed to open stream "
-                    << PrintableRxfilename(archive_rxfilename_);
-      state_ = kUninitialized;  // Failure on Open
-      return false;             // User should print the error message.
-    } else {
-      state_ = kNoObject;
-    }
-    return true;
-  }
-
-  // ReadNextObject() requires that the state be kNoObject,
-  // and it will try read the next object.  If it succeeds,
-  // it sets the state to kHaveObject, and
-  // cur_key_ and holder_ have the key and value.  If it fails,
-  // it sets the state to kError or kEof.
-  void ReadNextObject() {
-    if (state_ != kNoObject)
-      KALDIFST_ERR << "ReadNextObject() called from wrong state.";
-    // Code error somewhere in this class or a child class.
-    std::istream &is = input_.Stream();
-    is.clear();  // Clear any fail bits that may have been set... just in case
-    // this happened in the Read function.
-    is >> cur_key_;  // This eats up any leading whitespace and gets the string.
-    if (is.eof()) {
-      state_ = kEof;
-      return;
-    }
-    if (is.fail()) {  // This shouldn't really happen, barring file-system
-                      // errors.
-      KALDIFST_WARN << "Error reading archive: rspecifier is " << rspecifier_;
-      state_ = kError;
-      return;
-    }
-    int c;
-    if ((c = is.peek()) != ' ' && c != '\t' && c != '\n') {  // We expect a
-      // space ' ' after the key.
-      // We also allow tab, just so we can read archives generated by scripts
-      // that may not be fully aware of how this format works.
-      KALDIFST_WARN << "Invalid archive file format: expected space after key "
-                    << cur_key_ << ", got character "
-                    << CharToString(static_cast<char>(is.peek()))
-                    << ", reading archive "
-                    << PrintableRxfilename(archive_rxfilename_);
-      state_ = kError;
-      return;
-    }
-    if (c != '\n') is.get();  // Consume the space or tab.
-    holder_ = new Holder;
-    if (holder_->Read(is)) {
-      state_ = kHaveObject;
-      return;
-    } else {
-      KALDIFST_WARN << "Object read failed, reading archive "
-                    << PrintableRxfilename(archive_rxfilename_);
-      state_ = kError;
-      delete holder_;
-      holder_ = NULL;
-      return;
-    }
-  }
-
-  virtual bool IsOpen() const {
-    switch (state_) {
-      case kEof:
-      case kError:
-      case kHaveObject:
-      case kNoObject:
-        return true;
-      case kUninitialized:
-        return false;
-      default:
-        KALDIFST_ERR << "IsOpen() called on invalid object.";
-        return false;
-    }
-  }
-
-  // Called by the child-class virutal Close() functions; does the
-  // shared parts of the cleanup.
-  bool CloseInternal() {
-    if (!this->IsOpen())
-      KALDIFST_ERR
-          << "Close() called on TableReader twice or otherwise wrongly.";
-    if (input_.IsOpen()) input_.Close();
-    if (state_ == kHaveObject) {
-      KALDIFST_ASSERT(holder_ != NULL);
-      delete holder_;
-      holder_ = NULL;
-    } else {
-      KALDIFST_ASSERT(holder_ == NULL);
-    }
-    bool ans = (state_ != kError);
-    state_ = kUninitialized;
-    if (!ans && opts_.permissive) {
-      KALDIFST_WARN << "Error state detected closing reader.  "
-                    << "Ignoring it because you specified permissive mode.";
-      return true;
-    }
-    return ans;
-  }
-
-  ~RandomAccessTableReaderArchiveImplBase() {
-    // The child class has the responsibility to call CloseInternal().
-    KALDIFST_ASSERT(state_ == kUninitialized && holder_ == NULL);
-  }
-
- private:
-  Input input_;  // Input object for the archive
- protected:
-  // The variables below are accessed by child classes.
-
-  std::string cur_key_;  // current key (if state == kHaveObject).
-  Holder *holder_;  // Holds the object we just read (if state == kHaveObject)
-
-  std::string rspecifier_;
-  std::string archive_rxfilename_;
-  RspecifierOptions opts_;
-
-  enum {  //  [The state of the reading process]        [does holder_ [is input_
-    //                                                      have object]   open]
-    kUninitialized,  // Uninitialized or closed                   no         no
-    kNoObject,    // Do not have object in holder_              no         yes
-    kHaveObject,  // Have object in holder_                     yes        yes
-    kEof,         // End of file                                no         yes
-    kError,       // Some kind of error-state in the reading.   no         yes
-  } state_;
-};
-
-// RandomAccessTableReaderDSortedArchiveImpl (DSorted for "doubly sorted") is
-// the implementation for random-access reading of archives when both the
-// archive, and the calling code, are in sorted order (i.e. we ask for the keys
-// in sorted order).  This is when the s and cs options are both given.  It only
-// ever has to keep one object in memory.  It inherits from
-// RandomAccessTableReaderArchiveImplBase which implements the common parts of
-// RandomAccessTableReader that are used when it's an archive we're reading from
-
-template <class Holder>
-class RandomAccessTableReaderDSortedArchiveImpl
-    : public RandomAccessTableReaderArchiveImplBase<Holder> {
-  using RandomAccessTableReaderArchiveImplBase<Holder>::kUninitialized;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::kHaveObject;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::kNoObject;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::kEof;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::kError;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::state_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::opts_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::cur_key_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::holder_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::rspecifier_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::archive_rxfilename_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::ReadNextObject;
-
- public:
-  typedef typename Holder::T T;
-
-  RandomAccessTableReaderDSortedArchiveImpl() {}
-
-  virtual bool Close() {
-    // We don't have anything additional to clean up, so just
-    // call generic base-class one.
-    return this->CloseInternal();
-  }
-
-  virtual bool HasKey(const std::string &key) { return FindKeyInternal(key); }
-  virtual const T &Value(const std::string &key) {
-    if (!FindKeyInternal(key)) {
-      KALDIFST_ERR << "Value() called but no such key " << key << " in archive "
-                   << PrintableRxfilename(archive_rxfilename_);
-    }
-    KALDIFST_ASSERT(this->state_ == kHaveObject && key == this->cur_key_ &&
-                    holder_ != NULL);
-    return this->holder_->Value();
-  }
-
-  virtual ~RandomAccessTableReaderDSortedArchiveImpl() {
-    if (this->IsOpen())
-      if (!Close())  // more specific warning will already have been printed.
-        // we are in some kind of error state & user did not find out by
-        // calling Close().
-        KALDIFST_ERR << "Error closing RandomAccessTableReader: rspecifier is "
-                     << rspecifier_;
-  }
-
- private:
-  // FindKeyInternal tries to find the key by calling "ReadNextObject()"
-  // as many times as necessary till we get to it.  It is called from
-  // both FindKey and Value().
-  bool FindKeyInternal(const std::string &key) {
-    // First check that the user is calling us right: should be
-    // in sorted order.  If not, error.
-    if (!last_requested_key_.empty()) {
-      if (key.compare(last_requested_key_) < 0) {  // key < last_requested_key_
-        KALDIFST_ERR << "You provided the \"cs\" option "
-                     << "but are not calling with keys in sorted order: " << key
-                     << " < " << last_requested_key_ << ": rspecifier is "
-                     << rspecifier_;
-      }
-    }
-    // last_requested_key_ is just for debugging of order of calling.
-    last_requested_key_ = key;
-
-    if (state_ == kNoObject) ReadNextObject();  // This can only happen
-    // once, the first time someone calls HasKey() or Value().  We don't
-    // do it in the initializer to stop the program hanging too soon,
-    // if reading from a pipe.
-
-    if (state_ == kEof || state_ == kError) return false;
-
-    if (state_ == kUninitialized)
-      KALDIFST_ERR
-          << "Trying to access a RandomAccessTableReader object that is"
-             " not open.";
-
-    std::string last_key_;  // To check that
-    // the archive we're reading is in sorted order.
-    while (1) {
-      KALDIFST_ASSERT(state_ == kHaveObject);
-      int compare = key.compare(cur_key_);
-      if (compare == 0) {        // key == key_
-        return true;             // we got it..
-      } else if (compare < 0) {  // key < cur_key_, so we already read past the
-        // place where we want to be.  This implies that we will never find it
-        // [due to the sorting etc., this means it just isn't in the archive].
-        return false;
-      } else {  // compare > 0, key > cur_key_.  We need to read further ahead.
-        last_key_ = cur_key_;
-        // read next object.. we have to set state to kNoObject first.
-        KALDIFST_ASSERT(holder_ != NULL);
-        delete holder_;
-        holder_ = NULL;
-        state_ = kNoObject;
-        ReadNextObject();
-        if (state_ != kHaveObject) return false;  // eof or read error.
-        if (cur_key_.compare(last_key_) <= 0) {
-          KALDIFST_ERR << "You provided the \"s\" option "
-                       << " (sorted order), but keys are out of order or"
-                          " duplicated: "
-                       << last_key_ << " is followed by " << cur_key_
-                       << ": rspecifier is " << rspecifier_;
-        }
-      }
-    }
-  }
-
-  /// Last string provided to HasKey() or Value();
-  std::string last_requested_key_;
-};
-
-// RandomAccessTableReaderSortedArchiveImpl is for random-access reading of
-// archives when the user specified the sorted (s) option but not the
-// called-sorted (cs) options.
-template <class Holder>
-class RandomAccessTableReaderSortedArchiveImpl
-    : public RandomAccessTableReaderArchiveImplBase<Holder> {
-  using RandomAccessTableReaderArchiveImplBase<Holder>::kUninitialized;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::kHaveObject;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::kNoObject;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::kEof;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::kError;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::state_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::opts_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::cur_key_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::holder_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::rspecifier_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::archive_rxfilename_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::ReadNextObject;
-
- public:
-  typedef typename Holder::T T;
-
-  RandomAccessTableReaderSortedArchiveImpl()
-      : last_found_index_(static_cast<size_t>(-1)),
-        pending_delete_(static_cast<size_t>(-1)) {}
-
-  virtual bool Close() {
-    for (size_t i = 0; i < seen_pairs_.size(); i++)
-      delete seen_pairs_[i].second;
-    seen_pairs_.clear();
-
-    pending_delete_ = static_cast<size_t>(-1);
-    last_found_index_ = static_cast<size_t>(-1);
-
-    return this->CloseInternal();
-  }
-  virtual bool HasKey(const std::string &key) {
-    HandlePendingDelete();
-    size_t index;
-    bool ans = FindKeyInternal(key, &index);
-    if (ans && opts_.once && seen_pairs_[index].second == NULL) {
-      // Just do a check RE the once option. "&&opts_.once" is for
-      // efficiency since this can only happen in that case.
-      KALDIFST_ERR << "Error: HasKey called after Value() already called for "
-                   << " that key, and once (o) option specified: rspecifier is "
-                   << rspecifier_;
-    }
-    return ans;
-  }
-  virtual const T &Value(const std::string &key) {
-    HandlePendingDelete();
-    size_t index;
-    if (!FindKeyInternal(key, &index)) {
-      KALDIFST_ERR << "Value() called but no such key " << key << " in archive "
-                   << PrintableRxfilename(archive_rxfilename_);
-    }
-    if (seen_pairs_[index].second == NULL) {  // can happen if opts.once_
-      KALDIFST_ERR << "Error: Value() called more than once for key " << key
-                   << " and once (o) option specified: rspecifier is "
-                   << rspecifier_;
-    }
-    if (opts_.once)
-      pending_delete_ = index;  // mark this index to be deleted on next call.
-    return seen_pairs_[index].second->Value();
-  }
-  virtual ~RandomAccessTableReaderSortedArchiveImpl() {
-    if (this->IsOpen())
-      if (!Close())  // more specific warning will already have been printed.
-        // we are in some kind of error state & user did not find out by
-        // calling Close().
-        KALDIFST_ERR << "Error closing RandomAccessTableReader: rspecifier is "
-                     << rspecifier_;
-  }
-
- private:
-  void HandlePendingDelete() {
-    const size_t npos = static_cast<size_t>(-1);
-    if (pending_delete_ != npos) {
-      KALDIFST_ASSERT(pending_delete_ < seen_pairs_.size());
-      KALDIFST_ASSERT(seen_pairs_[pending_delete_].second != NULL);
-      delete seen_pairs_[pending_delete_].second;
-      seen_pairs_[pending_delete_].second = NULL;
-      pending_delete_ = npos;
-    }
-  }
-
-  // FindKeyInternal tries to find the key in the array "seen_pairs_".
-  // If it is not already there, it reads ahead as far as necessary
-  // to determine whether we have the key or not.  On success it returns
-  // true and puts the index into the array seen_pairs_, into "index";
-  // on failure it returns false.
-  // It will leave the state as either kNoObject, kEof or kError.
-  // FindKeyInternal does not do any checking about whether you are asking
-  // about a key that has been already given (with the "once" option).
-  // That is the user's responsibility.
-
-  bool FindKeyInternal(const std::string &key, size_t *index) {
-    // First, an optimization in case the previous call was for the
-    // same key, and we found it.
-    if (last_found_index_ < seen_pairs_.size() &&
-        seen_pairs_[last_found_index_].first == key) {
-      *index = last_found_index_;
-      return true;
-    }
-
-    if (state_ == kUninitialized)
-      KALDIFST_ERR
-          << "Trying to access a RandomAccessTableReader object that is"
-             " not open.";
-
-    // Step one is to see whether we have to read ahead for the object..
-    // Note, the possible states right now are kNoObject, kEof or kError.
-    // We are never in the state kHaveObject except just after calling
-    // ReadNextObject().
-    bool looped = false;
-    while (state_ == kNoObject &&
-           (seen_pairs_.empty() || key.compare(seen_pairs_.back().first) > 0)) {
-      looped = true;
-      // Read this as:
-      //  while ( the stream is potentially good for reading &&
-      //        ([got no keys] || key > most_recent_key) ) { ...
-      //     Try to read a new object.
-      // Note that the keys in seen_pairs_ are ordered from least to greatest.
-      ReadNextObject();
-      if (state_ == kHaveObject) {   // Successfully read object.
-        if (!seen_pairs_.empty() &&  // This is just a check.
-            cur_key_.compare(seen_pairs_.back().first) <= 0) {
-          // read the expression above as: !( cur_key_ > previous_key).
-          // it means we are not in sorted order [the user specified that we
-          // are, or we would not be using this implementation].
-          KALDIFST_ERR
-              << "You provided the sorted (s) option but keys in archive "
-              << PrintableRxfilename(archive_rxfilename_) << " are not "
-              << "in sorted order: " << seen_pairs_.back().first
-              << " is followed by " << cur_key_;
-        }
-        KALDIFST_ASSERT(holder_ != NULL);
-        seen_pairs_.push_back(std::make_pair(cur_key_, holder_));
-        holder_ = NULL;
-        state_ = kNoObject;
-      }
-    }
-    if (looped) {  // We only need to check the last element of the seen_pairs_
-      // array, since we would not have read more after getting "key".
-      if (!seen_pairs_.empty() && seen_pairs_.back().first == key) {
-        last_found_index_ = *index = seen_pairs_.size() - 1;
-        return true;
-      } else {
-        return false;
-      }
-    }
-    // Now we have do an actual binary search in the seen_pairs_ array.
-    std::pair<std::string, Holder *> pr(key, static_cast<Holder *>(NULL));
-    typename std::vector<std::pair<std::string, Holder *>>::iterator iter =
-        std::lower_bound(seen_pairs_.begin(), seen_pairs_.end(), pr,
-                         PairCompare());
-    if (iter != seen_pairs_.end() && key == iter->first) {
-      last_found_index_ = *index = (iter - seen_pairs_.begin());
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  // These are the pairs of (key, object) we have read.  We keep all the keys we
-  // have read but the actual objects (if they are stored with pointers inside
-  // the Holder object) may be deallocated if once == true, and the Holder
-  // pointer set to NULL.
-  std::vector<std::pair<std::string, Holder *>> seen_pairs_;
-  size_t last_found_index_;  // An optimization s.t. if FindKeyInternal called
-  // twice with same key (as it often will), it doesn't have to do the key
-  // search twice.
-  size_t pending_delete_;  // If opts_.once == true, this is the index of
-  // element of seen_pairs_ that is pending deletion.
-  struct PairCompare {
-    // PairCompare is the Less-than operator for the pairs of(key, Holder).
-    // compares the keys.
-    inline bool operator()(const std::pair<std::string, Holder *> &pr1,
-                           const std::pair<std::string, Holder *> &pr2) {
-      return (pr1.first.compare(pr2.first) < 0);
-    }
-  };
-};
-
-// RandomAccessTableReaderUnsortedArchiveImpl is for random-access reading of
-// archives when the user does not specify the sorted (s) option (in this case
-// the called-sorted, or "cs" option, is ignored).  This is the least efficient
-// of the random access archive readers, in general, but it can be as efficient
-// as the others, in speed, memory and latency, if the "once" option is
-// specified and it happens that the keys of the archive are the same as the
-// keys the code is called with (to HasKey() and Value()), and in the same
-// order.  However, if you ask it for a key that's not present it will have to
-// read the archive till the end and store it all in memory.
-
-template <class Holder>
-class RandomAccessTableReaderUnsortedArchiveImpl
-    : public RandomAccessTableReaderArchiveImplBase<Holder> {
-  using RandomAccessTableReaderArchiveImplBase<Holder>::kUninitialized;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::kHaveObject;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::kNoObject;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::kEof;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::kError;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::state_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::opts_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::cur_key_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::holder_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::rspecifier_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::archive_rxfilename_;
-  using RandomAccessTableReaderArchiveImplBase<Holder>::ReadNextObject;
-
-  typedef typename Holder::T T;
-
- public:
-  RandomAccessTableReaderUnsortedArchiveImpl()
-      : to_delete_iter_(map_.end()), to_delete_iter_valid_(false) {
-    map_.max_load_factor(0.5);  // make it quite empty -> quite efficient.
-    // default seems to be 1.
-  }
-
-  virtual bool Close() {
-    for (typename MapType::iterator iter = map_.begin(); iter != map_.end();
-         ++iter) {
-      delete iter->second;
-    }
-    map_.clear();
-    first_deleted_string_ = "";
-    to_delete_iter_valid_ = false;
-    return this->CloseInternal();
-  }
-
-  virtual bool HasKey(const std::string &key) {
-    HandlePendingDelete();
-    return FindKeyInternal(key, NULL);
-  }
-  virtual const T &Value(const std::string &key) {
-    HandlePendingDelete();
-    const T *ans_ptr = NULL;
-    if (!FindKeyInternal(key, &ans_ptr))
-      KALDIFST_ERR << "Value() called but no such key " << key << " in archive "
-                   << PrintableRxfilename(archive_rxfilename_);
-    return *ans_ptr;
-  }
-  virtual ~RandomAccessTableReaderUnsortedArchiveImpl() {
-    if (this->IsOpen())
-      if (!Close())  // more specific warning will already have been printed.
-        // we are in some kind of error state & user did not find out by
-        // calling Close().
-        KALDIFST_ERR << "Error closing RandomAccessTableReader: rspecifier is "
-                     << rspecifier_;
-  }
-
- private:
-  void HandlePendingDelete() {
-    if (to_delete_iter_valid_) {
-      to_delete_iter_valid_ = false;
-      delete to_delete_iter_->second;  // Delete Holder object.
-      if (first_deleted_string_.length() == 0)
-        first_deleted_string_ = to_delete_iter_->first;
-      map_.erase(to_delete_iter_);  // delete that element.
-    }
-  }
-
-  // FindKeyInternal tries to find the key in the map "map_"
-  // If it is not already there, it reads ahead either until it finds the
-  // key, or until end of file.  If called with value_ptr == NULL,
-  // it assumes it's called from HasKey() and just returns true or false
-  // and doesn't otherwise have side effects.  If called with value_ptr !=
-  // NULL, it assumes it's called from Value().  Thus, it will crash
-  // if it cannot find the key.  If it can find it it puts its address in
-  // *value_ptr, and if opts_once == true it will mark that element of the
-  // map to be deleted.
-
-  bool FindKeyInternal(const std::string &key, const T **value_ptr = NULL) {
-    typename MapType::iterator iter = map_.find(key);
-    if (iter != map_.end()) {   // Found in the map...
-      if (value_ptr == NULL) {  // called from HasKey
-        return true;            // this is all we have to do.
-      } else {
-        *value_ptr = &(iter->second->Value());
-        if (opts_.once) {  // value won't be needed again, so mark
-          // for deletion.
-          to_delete_iter_ = iter;  // pending delete.
-          KALDIFST_ASSERT(!to_delete_iter_valid_);
-          to_delete_iter_valid_ = true;
-        }
-        return true;
-      }
-    }
-    while (state_ == kNoObject) {
-      ReadNextObject();
-      if (state_ == kHaveObject) {  // Successfully read object.
-        state_ = kNoObject;         // we are about to transfer ownership
-        // of the object in holder_ to map_.
-        // Insert it into map_.
-        std::pair<typename MapType::iterator, bool> pr =
-            map_.insert(typename MapType::value_type(cur_key_, holder_));
-
-        if (!pr.second) {  // Was not inserted-- previous element w/ same key
-          delete holder_;  // map was not changed, no ownership transferred.
-          holder_ = NULL;
-          KALDIFST_ERR << "Error in RandomAccessTableReader: duplicate key "
-                       << cur_key_ << " in archive " << archive_rxfilename_;
-        }
-        holder_ = NULL;             // ownership transferred to map_.
-        if (cur_key_ == key) {      // the one we wanted..
-          if (value_ptr == NULL) {  // called from HasKey
-            return true;
-          } else {                                      // called from Value()
-            *value_ptr = &(pr.first->second->Value());  // this gives us the
-            // Value() from the Holder in the map.
-            if (opts_.once) {  // mark for deletion, as won't be needed again.
-              to_delete_iter_ = pr.first;
-              KALDIFST_ASSERT(!to_delete_iter_valid_);
-              to_delete_iter_valid_ = true;
-            }
-            return true;
-          }
-        }
-      }
-    }
-    if (opts_.once && key == first_deleted_string_) {
-      KALDIFST_ERR << "You specified the once (o) option but "
-                   << "you are calling using key " << key
-                   << " more than once: rspecifier is " << rspecifier_;
-    }
-    return false;  // We read the entire archive (or got to error state) and
-    // didn't find it.
-  }
-
-  typedef std::unordered_map<std::string, Holder *, StringHasher> MapType;
-  MapType map_;
-
-  typename MapType::iterator to_delete_iter_;
-  bool to_delete_iter_valid_;
-
-  std::string first_deleted_string_;  // keep the first string we deleted
-  // from map_ (if opts_.once == true).  It's for an inexact spot-check that the
-  // "once" option isn't being used incorrectly.
-};
-
-template <class Holder>
-RandomAccessTableReader<Holder>::RandomAccessTableReader(
-    const std::string &rspecifier)
-    : impl_(NULL) {
-  if (rspecifier != "" && !Open(rspecifier))
-    KALDIFST_ERR << "Error opening RandomAccessTableReader object "
-                    " (rspecifier is: "
-                 << rspecifier << ")";
-}
-
-template <class Holder>
-bool RandomAccessTableReader<Holder>::Open(const std::string &rspecifier) {
-  if (IsOpen()) KALDIFST_ERR << "Already open.";
-  RspecifierOptions opts;
-  RspecifierType rs = ClassifyRspecifier(rspecifier, NULL, &opts);
-  switch (rs) {
-    case kScriptRspecifier:
-      impl_ = new RandomAccessTableReaderScriptImpl<Holder>();
-      break;
-    case kArchiveRspecifier:
-      if (opts.sorted) {
-        if (opts.called_sorted)  // "doubly" sorted case.
-          impl_ = new RandomAccessTableReaderDSortedArchiveImpl<Holder>();
-        else
-          impl_ = new RandomAccessTableReaderSortedArchiveImpl<Holder>();
-      } else {
-        impl_ = new RandomAccessTableReaderUnsortedArchiveImpl<Holder>();
-      }
-      break;
-    case kNoRspecifier:
-    default:
-      KALDIFST_WARN << "Invalid rspecifier: " << rspecifier;
-      return false;
-  }
-  if (!impl_->Open(rspecifier)) {
-    // A warning will already have been printed.
-    delete impl_;
-    impl_ = NULL;
-    return false;
-  }
-  return true;
-}
-
-template <class Holder>
-bool RandomAccessTableReader<Holder>::HasKey(const std::string &key) {
-  CheckImpl();
-  if (!IsToken(key)) KALDIFST_ERR << "Invalid key \"" << key << '"';
-  return impl_->HasKey(key);
-}
-
-template <class Holder>
-const typename RandomAccessTableReader<Holder>::T &
-RandomAccessTableReader<Holder>::Value(const std::string &key) {
-  CheckImpl();
-  return impl_->Value(key);
-}
-
-template <class Holder>
-bool RandomAccessTableReader<Holder>::Close() {
-  CheckImpl();
-  bool ans = impl_->Close();
-  delete impl_;
-  impl_ = NULL;
-  return ans;
-}
-
-template <class Holder>
-RandomAccessTableReader<Holder>::~RandomAccessTableReader() {
-  if (IsOpen() && !Close())  // call Close() yourself to stop this being thrown.
-    KALDIFST_ERR << "failure detected in destructor.";
-}
-
-template <class Holder>
-void SequentialTableReader<Holder>::CheckImpl() const {
-  if (!impl_) {
-    KALDIFST_ERR << "Trying to use empty SequentialTableReader (perhaps you "
-                 << "passed the empty string as an argument to a program?)";
-  }
-}
-
-template <class Holder>
-void RandomAccessTableReader<Holder>::CheckImpl() const {
-  if (!impl_) {
-    KALDIFST_ERR << "Trying to use empty RandomAccessTableReader (perhaps you "
-                 << "passed the empty string as an argument to a program?)";
-  }
-}
-
-template <class Holder>
-void TableWriter<Holder>::CheckImpl() const {
-  if (!impl_) {
-    KALDIFST_ERR << "Trying to use empty TableWriter (perhaps you "
-                 << "passed the empty string as an argument to a program?)";
-  }
-}
-
-template <class Holder>
-RandomAccessTableReaderMapped<Holder>::RandomAccessTableReaderMapped(
-    const std::string &table_rxfilename, const std::string &utt2spk_rxfilename)
-    : reader_(table_rxfilename),
-      token_reader_(table_rxfilename.empty() ? "" : utt2spk_rxfilename),
-      utt2spk_rxfilename_(utt2spk_rxfilename) {}
-
-template <class Holder>
-bool RandomAccessTableReaderMapped<Holder>::Open(
-    const std::string &table_rxfilename,
-    const std::string &utt2spk_rxfilename) {
-  if (reader_.IsOpen()) reader_.Close();
-  if (token_reader_.IsOpen()) token_reader_.Close();
-  KALDIFST_ASSERT(!table_rxfilename.empty());
-  if (!reader_.Open(table_rxfilename)) return false;  // will have printed
-  // warning internally, probably.
-  if (!utt2spk_rxfilename.empty()) {
-    if (!token_reader_.Open(utt2spk_rxfilename)) {
-      reader_.Close();
-      return false;
-    }
-  }
-  return true;
-}
-
-template <class Holder>
-bool RandomAccessTableReaderMapped<Holder>::HasKey(const std::string &utt) {
-  // We don't check IsOpen, we let the call go through to the member variable
-  // (reader_), which will crash with a more informative error message than
-  // we can give here, as we don't any longer know the rxfilename.
-  if (token_reader_.IsOpen()) {  // We need to map the key from utt to spk.
-    if (!token_reader_.HasKey(utt))
-      KALDIFST_ERR << "Attempting to read key " << utt
-                   << ", which is not present "
-                   << "in utt2spk map or similar map being read from "
-                   << PrintableRxfilename(utt2spk_rxfilename_);
-    const std::string &spk = token_reader_.Value(utt);
-    return reader_.HasKey(spk);
-  } else {
-    return reader_.HasKey(utt);
-  }
-}
-
-template <class Holder>
-const typename Holder::T &RandomAccessTableReaderMapped<Holder>::Value(
-    const std::string &utt) {
-  if (token_reader_.IsOpen()) {  // We need to map the key from utt to spk.
-    if (!token_reader_.HasKey(utt))
-      KALDIFST_ERR << "Attempting to read key " << utt
-                   << ", which is not present "
-                   << "in utt2spk map or similar map being read from "
-                   << PrintableRxfilename(utt2spk_rxfilename_);
-    const std::string &spk = token_reader_.Value(utt);
-    return reader_.Value(spk);
-  } else {
-    return reader_.Value(utt);
-  }
-}
-
-/// @}
-
-}  // namespace kaldifst
-
-#endif  // KALDIFST_CSRC_KALDI_TABLE_INL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-table.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-table.h
deleted file mode 100644
index 2b5b6787..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/kaldi-table.h
+++ /dev/null
@@ -1,451 +0,0 @@
-// kaldifst/csrc/kaldi-table.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/util/kaldi-table.h
-
-// Copyright 2009-2011    Microsoft Corporation
-//                2013    Johns Hopkins University (author: Daniel Povey)
-
-#ifndef KALDIFST_CSRC_KALDI_TABLE_H_
-#define KALDIFST_CSRC_KALDI_TABLE_H_
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "kaldifst/csrc/kaldi-holder.h"
-
-namespace kaldifst {
-
-// Forward declarations
-template <class Holder>
-class RandomAccessTableReaderImplBase;
-template <class Holder>
-class SequentialTableReaderImplBase;
-template <class Holder>
-class TableWriterImplBase;
-
-// This header defines the Table classes (RandomAccessTableReader,
-// SequentialTableReader and TableWriter) and explains what the Holder classes,
-// which the Table class requires as a template argument, are like.  It also
-// explains the "rspecifier" and "wspecifier" concepts (these are strings that
-// explain how to read/write objects via archives or scp files.  A table is
-// conceptually a collection of objects of a particular type T indexed by keys
-// of type std::string (these Keys additionally have an order within
-// each table).
-// The Table classes are templated on a type (call it Holder) such that
-// Holder::T is a typedef equal to T.
-
-// Documentation for "rspecifier"
-// "rspecifier" describes how we read a set of objects indexed by keys.
-// The possibilities are:
-//
-// ark:rxfilename
-// scp:rxfilename
-//
-// We also allow various modifiers:
-//   o   means the program will only ask for each key once, which enables
-//       the reader to discard already-asked-for values.
-//   s   means the keys are sorted on input (means we don't have to read till
-//       eof if someone asked for a key that wasn't there).
-//   cs  means that it is called in sorted order (we are generally asserting
-//       this based on knowledge of how the program works).
-//   p   means "permissive", and causes it to skip over keys whose corresponding
-//       scp-file entries cannot be read. [and to ignore errors in archives and
-//       script files, and just consider the "good" entries].
-//       We allow the negation of the options above, as in no, ns, np,
-//       but these aren't currently very useful (just equivalent to omitting the
-//       corresponding option).
-//       [any of the above options can be prefixed by n to negate them, e.g. no,
-//       ns, ncs, np; but these aren't currently useful as you could just omit
-//       the option].
-//   bg means "background".  It currently has no effect for random-access
-//   readers,
-//       but for sequential readers it will cause it to "read ahead" to the next
-//       value, in a background thread.  Recommended when reading larger objects
-//       such as neural-net training examples, especially when you want to
-//       maximize GPU usage.
-//
-//   b   is ignored [for scripting convenience]
-//   t   is ignored [for scripting convenience]
-//
-//
-//  So for instance the following would be a valid rspecifier:
-//
-//   "o, s, p, ark:gunzip -c foo.gz|"
-
-struct RspecifierOptions {
-  // These options only make a difference for the RandomAccessTableReader class.
-  bool once;    // we assert that the program will only ask for each key once.
-  bool sorted;  // we assert that the keys are sorted.
-  bool called_sorted;  // we assert that the (HasKey(), Value() functions will
-  // also be called in sorted order.  [this implies "once" but not vice versa].
-  bool permissive;  // If "permissive", when reading from scp files it treats
-  // scp files that can't be read as if the corresponding key were not there.
-  // For archive files it will suppress errors getting thrown if the archive
-  // is corrupted and can't be read to the end.
-  bool background;  // For sequential readers, if the background option ("bg")
-                    // is provided, it will read ahead to the next object in a
-                    // background thread.
-  RspecifierOptions()
-      : once(false),
-        sorted(false),
-        called_sorted(false),
-        permissive(false),
-        background(false) {}
-};
-
-enum RspecifierType {
-  kNoRspecifier,
-  kArchiveRspecifier,
-  kScriptRspecifier,
-};
-
-RspecifierType ClassifyRspecifier(const std::string &rspecifier,
-                                  std::string *rxfilename,
-                                  RspecifierOptions *opts);
-
-// Documentation for "wspecifier"
-// "wspecifier" describes how we write a set of objects indexed by keys.
-// The basic, unadorned wspecifiers are as follows:
-//
-//  ark:wxfilename
-//  scp:rxfilename
-//  ark,scp:filename,wxfilename
-//  ark,scp:filename,wxfilename
-//
-//
-//  We also allow the following modifiers:
-//  t means text mode.
-//  b means binary mode.
-//  f means flush the stream after writing each entry.
-//   (nf means don't flush, and the default is not to flush).
-//  p means permissive mode, when writing to an "scp" file only: will ignore
-//     missing scp entries, i.e. won't write anything for those files but will
-//     return success status).
-//
-//  So the following are valid wspecifiers:
-//  ark,b,f:foo
-//  "ark,b,b:| gzip -c > foo"
-//  "ark,scp,t,nf:foo.ark,|gzip -c > foo.scp.gz"
-//  ark,b:-
-//
-//  The meanings of rxfilename and wxfilename are as described in
-//  kaldi-io.h (they are filenames but include pipes, stdin/stdout
-//  and so on; filename is a regular filename.
-//
-
-//  The ark:wxfilename type of wspecifier instructs the class to
-//  write directly to an archive.  For small objects (e.g. lists of ints),
-//  the text archive format will generally be human readable with one line
-//  per entry in the archive.
-//
-//  The type "scp:xfilename" refers to an scp file which should
-//  already exist on disk, and tells us where to write the data for
-//  each key (usually an actual file); each line of the scp file
-//  would be:
-//   key xfilename
-//
-//  The type ark,scp:filename,wxfilename means
-//  we write both an archive and an scp file that specifies offsets into the
-//  archive, with lines like:
-//    key filename:12407
-//  where the number is the byte offset into the file.
-//  In this case we restrict the archive-filename to be an actual filename,
-//  as we can't see a situation where an extended filename would make sense
-//  for this (we can't fseek() in pipes).
-
-enum WspecifierType {
-  kNoWspecifier,
-  kArchiveWspecifier,
-  kScriptWspecifier,
-  kBothWspecifier
-};
-
-struct WspecifierOptions {
-  bool binary;
-  bool flush;
-  bool permissive;  // will ignore absent scp entries.
-  WspecifierOptions() : binary(true), flush(false), permissive(false) {}
-};
-
-// ClassifyWspecifier returns the type of the wspecifier string,
-// and (if pointers are non-NULL) outputs the extra information
-// about the options, and the script and archive
-// filenames.
-WspecifierType ClassifyWspecifier(const std::string &wspecifier,
-                                  std::string *archive_wxfilename,
-                                  std::string *script_wxfilename,
-                                  WspecifierOptions *opts);
-
-// ReadScriptFile reads an .scp file in its entirety, and appends it
-// (in order as it was in the scp file) in script_out_, which contains
-// pairs of (key, xfilename).  The .scp
-// file format is: on each line, key xfilename
-// where xfilename means rxfilename or wxfilename, and may contain internal
-// spaces (we trim away any leading or trailing space).  The key is space-free.
-// ReadScriptFile returns true if the format was valid (empty files
-// are valid).
-// If 'print_warnings', it will print out warning messages that explain what
-// kind of error there was.
-bool ReadScriptFile(
-    const std::string &rxfilename, bool print_warnings,
-    std::vector<std::pair<std::string, std::string>> *script_out);
-
-// This version of ReadScriptFile works from an istream.
-bool ReadScriptFile(
-    std::istream &is, bool print_warnings,
-    std::vector<std::pair<std::string, std::string>> *script_out);
-
-// Writes, for each entry in script, the first element, then ' ', then the
-// second element then '\n'.  Checks that the keys (first elements of pairs) are
-// valid tokens (nonempty, no whitespace), and the values (second elements of
-// pairs) are newline-free and contain no leading or trailing space.  Returns
-// true on success.
-bool WriteScriptFile(
-    const std::string &wxfilename,
-    const std::vector<std::pair<std::string, std::string>> &script);
-
-// This version writes to an ostream.
-bool WriteScriptFile(
-    std::ostream &os,
-    const std::vector<std::pair<std::string, std::string>> &script);
-
-/// Allows random access to a collection
-/// of objects in an archive or script file; see \ref io_sec_tables.
-template <class Holder>
-class RandomAccessTableReader {
- public:
-  typedef typename Holder::T T;
-
-  RandomAccessTableReader() : impl_(NULL) {}
-
-  // This constructor is equivalent to default constructor + "open", but
-  // throws on error.
-  explicit RandomAccessTableReader(const std::string &rspecifier);
-
-  // Opens the table.
-  bool Open(const std::string &rspecifier);
-
-  // Returns true if table is open.
-  bool IsOpen() const { return (impl_ != NULL); }
-
-  // Close() will close the table [throws if it was not open],
-  // and returns true on success (false if we were reading an
-  // archive and we discovered an error in the archive).
-  bool Close();
-
-  // Says if it has this key.
-  // If you are using the "permissive" (p) read option,
-  // it will return false for keys whose corresponding entry
-  // in the scp file cannot be read.
-
-  bool HasKey(const std::string &key);
-
-  // Value() may throw if you are reading an scp file, you
-  // do not have the "permissive" (p) option, and an entry
-  // in the scp file cannot be read.  Typically you won't
-  // want to catch this error.
-  const T &Value(const std::string &key);
-
-  ~RandomAccessTableReader();
-
-  // Allow copy-constructor only for non-opened readers (needed for inclusion in
-  // stl vector)
-  RandomAccessTableReader(const RandomAccessTableReader<Holder> &other)
-      : impl_(NULL) {
-    KALDI_ASSERT(other.impl_ == NULL);
-  }
-
- private:
-  // Disallow assignment.
-  RandomAccessTableReader &operator=(const RandomAccessTableReader<Holder> &);
-  void CheckImpl() const;  // Checks that impl_ is non-NULL; prints an error
-                           // message and dies (with KALDI_ERR) if NULL.
-  RandomAccessTableReaderImplBase<Holder> *impl_;
-};
-
-/// A templated class for reading objects sequentially from an archive or script
-/// file; see \ref io_sec_tables.
-template <class Holder>
-class SequentialTableReader {
- public:
-  typedef typename Holder::T T;
-
-  SequentialTableReader() : impl_(NULL) {}
-
-  // This constructor equivalent to default constructor + "open", but
-  // throws on error.
-  explicit SequentialTableReader(const std::string &rspecifier);
-
-  // Opens the table.  Returns exit status; but does throw if previously open
-  // stream was in error state.  You can call Close to prevent this; anyway,
-  // calling Open more than once is not usually needed.
-  bool Open(const std::string &rspecifier);
-
-  // Returns true if we're done.  It will also return true if there's some kind
-  // of error and we can't read any more; in this case, you can detect the
-  // error by calling Close and checking the return status; otherwise
-  // the destructor will throw.
-  inline bool Done();
-
-  // Only valid to call Key() if Done() returned false.
-  inline std::string Key();
-
-  // FreeCurrent() is provided as an optimization to save memory, for large
-  // objects.  It instructs the class to deallocate the current value. The
-  // reference Value() will be invalidated by this.
-  void FreeCurrent();
-
-  // Return reference to the current value.  It's only valid to call this if
-  // Done() returned false.  The reference is valid till next call to this
-  // object.  It will throw if you are reading an scp file, did not specify the
-  // "permissive" (p) option and the file cannot be read.  [The permissive
-  // option makes it behave as if that key does not even exist, if the
-  // corresponding file cannot be read.]  You probably wouldn't want to catch
-  // this exception; the user can just specify the p option in the rspecifier.
-  // We make this non-const to enable things like shallow swap on the held
-  // object in situations where this would avoid making a redundant copy.
-  T &Value();
-
-  // Next goes to the next key.  It will not throw; any error will
-  // result in Done() returning true, and then the destructor will
-  // throw unless you call Close().
-  void Next();
-
-  // Returns true if table is open for reading (does not imply
-  // stream is in good state).
-  bool IsOpen() const;
-
-  // Close() will return false (failure) if Done() became true
-  // because of an error/ condition rather than because we are
-  // really done [e.g. because of an error or early termination
-  // in the archive].
-  // If there is an error and you don't call Close(), the destructor
-  // will fail.
-  // Close()
-  bool Close();
-
-  // The destructor may throw.  This is the desired behaviour, as it's the way
-  // we signal the error to the user (to detect it, call Close().  The issue is
-  // that otherwise the user has no way to tell whether Done() returned true
-  // because we reached the end of the archive or script, or because there was
-  // an error that prevented further reading.
-  ~SequentialTableReader();
-
-  // Allow copy-constructor only for non-opened readers (needed for inclusion in
-  // stl vector)
-  SequentialTableReader(const SequentialTableReader<Holder> &other)
-      : impl_(NULL) {
-    KALDI_ASSERT(other.impl_ == NULL);
-  }
-
- private:
-  // Disallow assignment.
-  SequentialTableReader &operator=(const SequentialTableReader<Holder> &);
-  void CheckImpl() const;  // Checks that impl_ is non-NULL; prints an error
-                           // message and dies (with KALDI_ERR) if NULL.
-  SequentialTableReaderImplBase<Holder> *impl_;
-};
-
-/// A templated class for writing objects to an
-/// archive or script file; see \ref io_sec_tables.
-template <class Holder>
-class TableWriter {
- public:
-  typedef typename Holder::T T;
-
-  TableWriter() : impl_(NULL) {}
-
-  // This constructor equivalent to default constructor
-  // + "open", but throws on error.  See docs for
-  // wspecifier above.
-  explicit TableWriter(const std::string &wspecifier);
-
-  // Opens the table.  See docs for wspecifier above.
-  // If it returns true, it is open.
-  bool Open(const std::string &wspecifier);
-
-  // Returns true if open for writing.
-  bool IsOpen() const;
-
-  // Write the object. Throws KaldiFatalError on error via the KALDI_ERR macro.
-  inline void Write(const std::string &key, const T &value) const;
-
-  // Flush will flush any archive; it does not return error status
-  // or throw, any errors will be reported on the next Write or Close.
-  // Useful if we may be writing to a command in a pipe and want
-  // to ensure good CPU utilization.
-  void Flush();
-
-  // Close() is not necessary to call, as the destructor
-  // closes it; it's mainly useful if you want to handle
-  // error states because the destructor will throw on
-  // error if you do not call Close().
-  bool Close();
-
-  ~TableWriter();
-
-  // Allow copy-constructor only for non-opened writers (needed for inclusion in
-  // stl vector)
-  TableWriter(const TableWriter &other) : impl_(NULL) {
-    KALDI_ASSERT(other.impl_ == NULL);
-  }
-
- private:
-  TableWriter &operator=(const TableWriter &);  // Disallow assignment.
-
-  void CheckImpl() const;  // Checks that impl_ is non-NULL; prints an error
-                           // message and dies (with KALDI_ERR) if NULL.
-  TableWriterImplBase<Holder> *impl_;
-};
-
-/// This class is for when you are reading something in random access, but
-/// it may actually be stored per-speaker (or something similar) but the
-/// keys you're using are per utterance.  So you also provide an "rxfilename"
-/// for a file containing lines like
-/// utt1 spk1
-/// utt2 spk1
-/// utt3 spk1
-/// and so on.  Note: this is optional; if it is an empty string, we just won't
-/// do the mapping.  Also, "table_rxfilename" may be the empty string (as for
-/// a regular table), in which case the table just won't be opened.
-/// We provide only the most frequently used of the functions of
-/// RandomAccessTableReader.
-
-template <class Holder>
-class RandomAccessTableReaderMapped {
- public:
-  typedef typename Holder::T T;
-  /// Note: "utt2spk_rxfilename" will in the normal case be an rxfilename
-  /// for an utterance to speaker map, but this code is general; it accepts
-  /// a generic map.
-  RandomAccessTableReaderMapped(const std::string &table_rxfilename,
-                                const std::string &utt2spk_rxfilename);
-
-  RandomAccessTableReaderMapped() {}
-
-  /// Note: when calling Open, utt2spk_rxfilename may be empty.
-  bool Open(const std::string &table_rxfilename,
-            const std::string &utt2spk_rxfilename);
-
-  bool HasKey(const std::string &key);
-  const T &Value(const std::string &key);
-  inline bool IsOpen() const { return reader_.IsOpen(); }
-  inline bool Close() { return reader_.Close(); }
-
-  // The default copy-constructor will do what we want: it will crash for
-  // already-opened readers, by calling the member-variable copy-constructors.
- private:
-  // Disallow assignment.
-  RandomAccessTableReaderMapped &operator=(
-      const RandomAccessTableReaderMapped<Holder> &);
-  RandomAccessTableReader<Holder> reader_;
-  RandomAccessTableReader<TokenHolder> token_reader_;
-  std::string utt2spk_rxfilename_;  // Used only in diagnostic messages.
-};
-
-}  // namespace kaldifst
-
-#include "kaldifst/csrc/kaldi-table-inl.h"
-
-#endif  // KALDIFST_CSRC_KALDI_TABLE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/lattice-utils-inl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/lattice-utils-inl.h
deleted file mode 100644
index 024ffe51..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/lattice-utils-inl.h
+++ /dev/null
@@ -1,42 +0,0 @@
-// kaldifst/csrc/lattice-utils-inl.h
-//
-// Copyright 2009-2011  Microsoft Corporation
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/lattice-utils-inl.h
-#ifndef KALDIFST_CSRC_LATTICE_UTILS_INL_H_
-#define KALDIFST_CSRC_LATTICE_UTILS_INL_H_
-
-#include <assert.h>
-
-#include <vector>
-
-namespace fst {
-
-template <class Weight, class ScaleFloat>
-void ScaleLattice(const std::vector<std::vector<ScaleFloat>> &scale,
-                  MutableFst<ArcTpl<Weight>> *fst) {
-  assert(scale.size() == 2 && scale[0].size() == 2 && scale[1].size() == 2);
-  if (scale == DefaultLatticeScale()) {  // nothing to do.
-    return;
-  }
-
-  using Arc = ArcTpl<Weight>;
-  using Fst = MutableFst<Arc>;
-  using StateId = typename Arc::StateId;
-
-  StateId num_states = fst->NumStates();
-  for (StateId s = 0; s < num_states; s++) {
-    for (MutableArcIterator<Fst> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-      Arc arc = aiter.Value();
-      arc.weight = Weight(ScaleTupleWeight(arc.weight, scale));
-      aiter.SetValue(arc);
-    }
-    Weight final_weight = fst->Final(s);
-    if (final_weight != Weight::Zero())
-      fst->SetFinal(s, Weight(ScaleTupleWeight(final_weight, scale)));
-  }
-}
-
-}  // namespace fst
-#endif  // KALDIFST_CSRC_LATTICE_UTILS_INL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/lattice-utils.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/lattice-utils.h
deleted file mode 100644
index ee7480d1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/lattice-utils.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// kaldifst/csrc/lattice-utils.h
-//
-// Copyright 2009-2011  Microsoft Corporation
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/lattice-utils.h
-
-#ifndef KALDIFST_CSRC_LATTICE_UTILS_H_
-#define KALDIFST_CSRC_LATTICE_UTILS_H_
-
-#include <vector>
-
-#include "fst/fstlib.h"
-#include "kaldifst/csrc/lattice-weight.h"
-
-namespace fst {
-/** Returns a default 2x2 matrix scaling factor for LatticeWeight */
-inline std::vector<std::vector<double>> DefaultLatticeScale() {
-  std::vector<std::vector<double>> ans(2);
-  ans[0].resize(2, 0.0);
-  ans[1].resize(2, 0.0);
-  ans[0][0] = ans[1][1] = 1.0;
-  return ans;
-}
-
-inline std::vector<std::vector<double>> LatticeScale(double lmwt, double acwt) {
-  std::vector<std::vector<double>> ans(2);
-  ans[0].resize(2, 0.0);
-  ans[1].resize(2, 0.0);
-  ans[0][0] = lmwt;
-  ans[1][1] = acwt;
-  return ans;
-}
-
-/** Scales the pairs of weights in LatticeWeight or CompactLatticeWeight by
-    viewing the pair (a, b) as a 2-vector and pre-multiplying by the 2x2 matrix
-    in "scale".  E.g. typically scale would equal
-     [ 1   0;
-       0  acwt ]
-    if we want to scale the acoustics by "acwt".
- */
-template <class Weight, class ScaleFloat>
-void ScaleLattice(const std::vector<std::vector<ScaleFloat>> &scale,
-                  MutableFst<ArcTpl<Weight>> *fst);
-
-}  // namespace fst
-
-#include "kaldifst/csrc/lattice-utils-inl.h"
-
-#endif  // KALDIFST_CSRC_LATTICE_UTILS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/lattice-weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/lattice-weight.h
deleted file mode 100644
index ff78cf57..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/lattice-weight.h
+++ /dev/null
@@ -1,898 +0,0 @@
-// kaldifst/csrc/lattice-weight.h
-// Copyright 2009-2012  Microsoft Corporation
-//                      Johns Hopkins University (author: Daniel Povey)
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/lattice-weight.cc
-
-#ifndef KALDIFST_CSRC_LATTICE_WEIGHT_H_
-#define KALDIFST_CSRC_LATTICE_WEIGHT_H_
-
-#include <algorithm>
-#include <iostream>
-#include <string>
-#include <vector>
-
-#include "fst/fstlib.h"
-#include "kaldifst/csrc/log.h"
-
-namespace fst {
-
-// Declare weight type for lattice...  has two
-// members, value1_ and value2_, of type BaseFloat (normally equals float).  It
-// is basically the same as the tropical semiring on value1_+value2_, except it
-// keeps track of a and b separately.  More precisely, it is equivalent to the
-// lexicographic semiring on (value1_+value2_), (value1_-value2_)
-
-template <class FloatType>
-class LatticeWeightTpl;
-
-template <class FloatType>
-inline std::ostream &operator<<(std::ostream &strm,
-                                const LatticeWeightTpl<FloatType> &w);
-
-template <class FloatType>
-inline std::istream &operator>>(std::istream &strm,
-                                LatticeWeightTpl<FloatType> &w);
-
-template <class FloatType>
-class LatticeWeightTpl {
- public:
-  typedef FloatType T;  // normally float.
-  typedef LatticeWeightTpl ReverseWeight;
-  using ValueType = FloatType;
-
-  inline T Value1() const { return value1_; }  // usually graph cost
-
-  inline T Value2() const { return value2_; }  // usually acoustic cost
-
-  inline void SetValue1(T f) { value1_ = f; }
-
-  inline void SetValue2(T f) { value2_ = f; }
-
-  LatticeWeightTpl() : value1_{}, value2_{} {}
-
-  LatticeWeightTpl(T a, T b = 0) : value1_(a), value2_(b) {}  // NOLINT
-
-  LatticeWeightTpl(const LatticeWeightTpl &other)
-      : value1_(other.value1_), value2_(other.value2_) {}
-
-  LatticeWeightTpl &operator=(const LatticeWeightTpl &w) {
-    value1_ = w.value1_;
-    value2_ = w.value2_;
-    return *this;
-  }
-
-  LatticeWeightTpl<FloatType> Reverse() const { return *this; }
-
-  static const LatticeWeightTpl Zero() {
-    return LatticeWeightTpl(std::numeric_limits<T>::infinity(),
-                            std::numeric_limits<T>::infinity());
-  }
-
-  static const LatticeWeightTpl One() { return LatticeWeightTpl(0.0, 0.0); }
-
-  static const std::string &Type() {
-    static const std::string type = (sizeof(T) == 4 ? "lattice4" : "lattice8");
-    return type;
-  }
-
-  static const LatticeWeightTpl NoWeight() {
-    return LatticeWeightTpl(std::numeric_limits<FloatType>::quiet_NaN(),
-                            std::numeric_limits<FloatType>::quiet_NaN());
-  }
-
-  bool Member() const {
-    // value1_ == value1_ tests for NaN.
-    // also test for no -inf, and either both or neither
-    // must be +inf, and
-    if (value1_ != value1_ || value2_ != value2_) return false;  // NaN
-    if (value1_ == -std::numeric_limits<T>::infinity() ||
-        value2_ == -std::numeric_limits<T>::infinity())
-      return false;  // -infty not allowed
-    if (value1_ == std::numeric_limits<T>::infinity() ||
-        value2_ == std::numeric_limits<T>::infinity()) {
-      if (value1_ != std::numeric_limits<T>::infinity() ||
-          value2_ != std::numeric_limits<T>::infinity())
-        return false;  // both must be +infty;
-      // this is necessary so that the semiring has only one zero.
-    }
-    return true;
-  }
-
-  LatticeWeightTpl Quantize(float delta = kDelta) const {
-    if (value1_ + value2_ == -std::numeric_limits<T>::infinity()) {
-      return LatticeWeightTpl(-std::numeric_limits<T>::infinity(),
-                              -std::numeric_limits<T>::infinity());
-    } else if (value1_ + value2_ == std::numeric_limits<T>::infinity()) {
-      return LatticeWeightTpl(std::numeric_limits<T>::infinity(),
-                              std::numeric_limits<T>::infinity());
-    } else if (value1_ + value2_ != value1_ + value2_) {  // NaN
-      return LatticeWeightTpl(value1_ + value2_, value1_ + value2_);
-    } else {
-      return LatticeWeightTpl(floor(value1_ / delta + 0.5F) * delta,
-                              floor(value2_ / delta + 0.5F) * delta);
-    }
-  }
-  static constexpr uint64 Properties() {
-    return kLeftSemiring | kRightSemiring | kCommutative | kPath | kIdempotent;
-  }
-
-  // This is used in OpenFst for binary I/O.  This is OpenFst-style,
-  // not Kaldi-style, I/O.
-  std::istream &Read(std::istream &strm) {
-    // Always read/write as float, even if T is double,
-    // so we can use OpenFst-style read/write and still maintain
-    // compatibility when compiling with different FloatTypes
-    ReadType(strm, &value1_);
-    ReadType(strm, &value2_);
-    return strm;
-  }
-
-  // This is used in OpenFst for binary I/O.  This is OpenFst-style,
-  // not Kaldi-style, I/O.
-  std::ostream &Write(std::ostream &strm) const {
-    WriteType(strm, value1_);
-    WriteType(strm, value2_);
-    return strm;
-  }
-
-  size_t Hash() const {
-    size_t ans;
-    union {
-      T f;
-      size_t s;
-    } u;
-    u.s = 0;
-    u.f = value1_;
-    ans = u.s;
-    u.f = value2_;
-    ans += u.s;
-    return ans;
-  }
-
- protected:
-  inline static void WriteFloatType(std::ostream &strm, const T &f) {
-    if (f == std::numeric_limits<T>::infinity())
-      strm << "Infinity";
-    else if (f == -std::numeric_limits<T>::infinity())
-      strm << "-Infinity";
-    else if (f != f)
-      strm << "BadNumber";
-    else
-      strm << f;
-  }
-
-  // Internal helper function, used in ReadNoParen.
-  inline static void ReadFloatType(std::istream &strm, T &f) {  // NOLINT
-    std::string s;
-    strm >> s;
-    if (s == "Infinity") {
-      f = std::numeric_limits<T>::infinity();
-    } else if (s == "-Infinity") {
-      f = -std::numeric_limits<T>::infinity();
-    } else if (s == "BadNumber") {
-      f = std::numeric_limits<T>::quiet_NaN();
-    } else {
-      char *p;
-      f = strtod(s.c_str(), &p);
-      if (p < s.c_str() + s.size()) strm.clear(std::ios::badbit);
-    }
-  }
-
-  // Reads LatticeWeight when there are no parentheses around pair terms...
-  // currently the only form supported.
-  inline std::istream &ReadNoParen(std::istream &strm, char separator) {
-    int c;
-    do {
-      c = strm.get();
-    } while (isspace(c));
-
-    std::string s1;
-    while (c != separator) {
-      if (c == EOF) {
-        strm.clear(std::ios::badbit);
-        return strm;
-      }
-      s1 += c;
-      c = strm.get();
-    }
-    std::istringstream strm1(s1);
-    ReadFloatType(strm1, value1_);  // ReadFloatType is class member function
-    // read second element
-    ReadFloatType(strm, value2_);
-    return strm;
-  }
-
-  friend std::istream &operator>>
-      <FloatType>(std::istream &, LatticeWeightTpl<FloatType> &);
-  friend std::ostream &operator<<<FloatType>(
-      std::ostream &, const LatticeWeightTpl<FloatType> &);
-
- private:
-  T value1_;  // usually graph cost
-  T value2_;  // usually acoustic cost
-};
-
-/* ScaleTupleWeight is a function defined for LatticeWeightTpl and
-   CompactLatticeWeightTpl that mutliplies the pair (value1_, value2_) by a 2x2
-   matrix.  Used, for example, in applying acoustic scaling.
- */
-template <class FloatType, class ScaleFloatType>
-inline LatticeWeightTpl<FloatType> ScaleTupleWeight(
-    const LatticeWeightTpl<FloatType> &w,
-    const std::vector<std::vector<ScaleFloatType>> &scale) {
-  // Without the next special case we'd get NaNs from infinity * 0
-  if (w.Value1() == std::numeric_limits<FloatType>::infinity()) {
-    return LatticeWeightTpl<FloatType>::Zero();
-  }
-
-  return LatticeWeightTpl<FloatType>(
-      scale[0][0] * w.Value1() + scale[0][1] * w.Value2(),
-      scale[1][0] * w.Value1() + scale[1][1] * w.Value2());
-}
-
-/* For testing purposes and in case it's ever useful, we define a similar
-   function to apply to LexicographicWeight and the like, templated on
-   TropicalWeight<float> etc.; we use PairWeight which is the base class of
-   LexicographicWeight.
-*/
-template <class FloatType, class ScaleFloatType>
-inline PairWeight<TropicalWeightTpl<FloatType>, TropicalWeightTpl<FloatType>>
-ScaleTupleWeight(const PairWeight<TropicalWeightTpl<FloatType>,
-                                  TropicalWeightTpl<FloatType>> &w,
-                 const std::vector<std::vector<ScaleFloatType>> &scale) {
-  typedef TropicalWeightTpl<FloatType> BaseType;
-  typedef PairWeight<BaseType, BaseType> PairType;
-  const BaseType zero = BaseType::Zero();
-  // Without the next special case we'd get NaNs from infinity * 0
-  if (w.Value1() == zero || w.Value2() == zero) return PairType(zero, zero);
-  FloatType f1 = w.Value1().Value(), f2 = w.Value2().Value();
-  return PairType(BaseType(scale[0][0] * f1 + scale[0][1] * f2),
-                  BaseType(scale[1][0] * f1 + scale[1][1] * f2));
-}
-
-template <class FloatType>
-inline bool operator==(const LatticeWeightTpl<FloatType> &wa,
-                       const LatticeWeightTpl<FloatType> &wb) {
-  // Volatile qualifier thwarts over-aggressive compiler optimizations
-  // that lead to problems esp. with NaturalLess().
-  volatile FloatType va1 = wa.Value1(), va2 = wa.Value2(), vb1 = wb.Value1(),
-                     vb2 = wb.Value2();
-  return (va1 == vb1 && va2 == vb2);
-}
-
-template <class FloatType>
-inline bool operator!=(const LatticeWeightTpl<FloatType> &wa,
-                       const LatticeWeightTpl<FloatType> &wb) {
-  // Volatile qualifier thwarts over-aggressive compiler optimizations
-  // that lead to problems esp. with NaturalLess().
-  volatile FloatType va1 = wa.Value1(), va2 = wa.Value2(), vb1 = wb.Value1(),
-                     vb2 = wb.Value2();
-  return (va1 != vb1 || va2 != vb2);
-}
-
-// We define a Compare function LatticeWeightTpl even though it's
-// not required by the semiring standard-- it's just more efficient
-// to do it this way rather than using the NaturalLess template.
-
-/// Compare returns -1 if w1 < w2, +1 if w1 > w2, and 0 if w1 == w2.
-
-template <class FloatType>
-inline int Compare(const LatticeWeightTpl<FloatType> &w1,
-                   const LatticeWeightTpl<FloatType> &w2) {
-  FloatType f1 = w1.Value1() + w1.Value2(), f2 = w2.Value1() + w2.Value2();
-  if (f1 < f2) {
-    return 1;
-  }  // having smaller cost means you're larger
-  // in the semiring [higher probability]
-  else if (f1 > f2) {  // NOLINT
-    return -1;
-  }
-  // mathematically we should be comparing (w1.value1_-w1.value2_ <
-  // w2.value1_-w2.value2_) in the next line, but add w1.value1_+w1.value2_ =
-  // w2.value1_+w2.value2_ to both sides and divide by two, and we get the
-  // simpler equivalent form w1.value1_ < w2.value1_.
-  else if (w1.Value1() < w2.Value1()) {  // NOLINT
-    return 1;
-  } else if (w1.Value1() > w2.Value1()) {
-    return -1;
-  } else {
-    return 0;
-  }
-}
-
-template <class FloatType>
-inline LatticeWeightTpl<FloatType> Plus(const LatticeWeightTpl<FloatType> &w1,
-                                        const LatticeWeightTpl<FloatType> &w2) {
-  return (Compare(w1, w2) >= 0 ? w1 : w2);
-}
-
-// For efficiency, override the NaturalLess template class.
-template <class FloatType>
-class NaturalLess<LatticeWeightTpl<FloatType>> {
- public:
-  typedef LatticeWeightTpl<FloatType> Weight;
-
-  NaturalLess() {}
-
-  bool operator()(const Weight &w1, const Weight &w2) const {
-    // NaturalLess is a negative order (opposite to normal ordering).
-    // This operator () corresponds to "<" in the negative order, which
-    // corresponds to the ">" in the normal order.
-    return (Compare(w1, w2) == 1);
-  }
-};
-template <>
-class NaturalLess<LatticeWeightTpl<float>> {
- public:
-  typedef LatticeWeightTpl<float> Weight;
-
-  NaturalLess() {}
-
-  bool operator()(const Weight &w1, const Weight &w2) const {
-    // NaturalLess is a negative order (opposite to normal ordering).
-    // This operator () corresponds to "<" in the negative order, which
-    // corresponds to the ">" in the normal order.
-    return (Compare(w1, w2) == 1);
-  }
-};
-template <>
-class NaturalLess<LatticeWeightTpl<double>> {
- public:
-  typedef LatticeWeightTpl<double> Weight;
-
-  NaturalLess() {}
-
-  bool operator()(const Weight &w1, const Weight &w2) const {
-    // NaturalLess is a negative order (opposite to normal ordering).
-    // This operator () corresponds to "<" in the negative order, which
-    // corresponds to the ">" in the normal order.
-    return (Compare(w1, w2) == 1);
-  }
-};
-
-template <class FloatType>
-inline LatticeWeightTpl<FloatType> Times(
-    const LatticeWeightTpl<FloatType> &w1,
-    const LatticeWeightTpl<FloatType> &w2) {
-  return LatticeWeightTpl<FloatType>(w1.Value1() + w2.Value1(),
-                                     w1.Value2() + w2.Value2());
-}
-
-// divide w1 by w2 (on left/right/any doesn't matter as
-// commutative).
-template <class FloatType>
-inline LatticeWeightTpl<FloatType> Divide(const LatticeWeightTpl<FloatType> &w1,
-                                          const LatticeWeightTpl<FloatType> &w2,
-                                          DivideType typ = DIVIDE_ANY) {
-  typedef FloatType T;
-  T a = w1.Value1() - w2.Value1(), b = w1.Value2() - w2.Value2();
-  if (a != a || b != b || a == -std::numeric_limits<T>::infinity() ||
-      b == -std::numeric_limits<T>::infinity()) {
-    KALDIFST_WARN
-        << "LatticeWeightTpl::Divide, NaN or invalid number produced. "
-        << "[dividing by zero?]  Returning zero";
-    return LatticeWeightTpl<T>::Zero();
-  }
-  if (a == std::numeric_limits<T>::infinity() ||
-      b == std::numeric_limits<T>::infinity())  // NOLINT
-    return LatticeWeightTpl<T>::Zero();  // not a valid number if only one is
-                                         // infinite.
-  return LatticeWeightTpl<T>(a, b);
-}
-
-template <class FloatType>
-inline bool ApproxEqual(const LatticeWeightTpl<FloatType> &w1,
-                        const LatticeWeightTpl<FloatType> &w2,
-                        float delta = kDelta) {
-  if (w1.Value1() == w2.Value1() && w1.Value2() == w2.Value2())
-    return true;  // handles Zero().
-  return (fabs((w1.Value1() + w1.Value2()) - (w2.Value1() + w2.Value2())) <=
-          delta);
-}
-
-template <class FloatType>
-inline std::ostream &operator<<(std::ostream &strm,
-                                const LatticeWeightTpl<FloatType> &w) {
-  LatticeWeightTpl<FloatType>::WriteFloatType(strm, w.Value1());
-  CHECK_EQ(FLAGS_fst_weight_separator.size(), 1);
-  strm << FLAGS_fst_weight_separator[0];  // comma by default;
-  // may or may not be settable from Kaldi programs.
-  LatticeWeightTpl<FloatType>::WriteFloatType(strm, w.Value2());
-  return strm;
-}
-
-template <class FloatType>
-inline std::istream &operator>>(std::istream &strm,
-                                LatticeWeightTpl<FloatType> &w1) {
-  CHECK_EQ(FLAGS_fst_weight_separator.size(), 1);
-  // separator defaults to ','
-  return w1.ReadNoParen(strm, FLAGS_fst_weight_separator[0]);
-}
-
-// CompactLattice will be an acceptor (accepting the words/output-symbols),
-// with the weights and input-symbol-seqs on the arcs.
-// There must be a total order on W.  We assume for the sake of efficiency
-// that there is a function
-// Compare(W w1, W w2) that returns -1 if w1 < w2, +1 if w1 > w2, and
-// zero if w1 == w2, and Plus for type W returns (Compare(w1,w2) >= 0 ? w1 :
-// w2).
-
-template <class WeightType, class IntType>
-class CompactLatticeWeightTpl {
- public:
-  typedef WeightType W;
-
-  typedef CompactLatticeWeightTpl<WeightType, IntType> ReverseWeight;
-
-  // Plus is like LexicographicWeight on the pair (weight_, string_), but where
-  // we use standard lexicographic order on string_ [this is not the same as
-  // NaturalLess on the StringWeight equivalent, which does not define a
-  // total order].
-  // Times, Divide obvious... (support both left & right division..)
-  // CommonDivisor would need to be coded separately.
-
-  CompactLatticeWeightTpl() {}
-
-  CompactLatticeWeightTpl(const WeightType &w, const std::vector<IntType> &s)
-      : weight_(w), string_(s) {}
-
-  CompactLatticeWeightTpl(
-      const CompactLatticeWeightTpl &compactLatticeWeightTpl) = default;
-
-  CompactLatticeWeightTpl &operator=(const CompactLatticeWeightTpl &w) =
-      default;
-
-  const W &Weight() const { return weight_; }
-
-  const std::vector<IntType> &String() const { return string_; }
-
-  void SetWeight(const W &w) { weight_ = w; }
-
-  void SetString(const std::vector<IntType> &s) { string_ = s; }
-
-  static const CompactLatticeWeightTpl<WeightType, IntType> Zero() {
-    return CompactLatticeWeightTpl<WeightType, IntType>(WeightType::Zero(),
-                                                        std::vector<IntType>());
-  }
-
-  static const CompactLatticeWeightTpl<WeightType, IntType> One() {
-    return CompactLatticeWeightTpl<WeightType, IntType>(WeightType::One(),
-                                                        std::vector<IntType>());
-  }
-
-  inline static std::string GetIntSizeString() {
-    char buf[2];
-    buf[0] = '0' + sizeof(IntType);
-    buf[1] = '\0';
-    return buf;
-  }
-  static const std::string &Type() {
-    static const std::string type =
-        "compact" + WeightType::Type() + GetIntSizeString();
-    return type;
-  }
-
-  static const CompactLatticeWeightTpl<WeightType, IntType> NoWeight() {
-    return CompactLatticeWeightTpl<WeightType, IntType>(WeightType::NoWeight(),
-                                                        std::vector<IntType>());
-  }
-
-  CompactLatticeWeightTpl<WeightType, IntType> Reverse() const {
-    size_t s = string_.size();
-    std::vector<IntType> v(s);
-    for (size_t i = 0; i < s; i++) v[i] = string_[s - i - 1];
-    return CompactLatticeWeightTpl<WeightType, IntType>(weight_, v);
-  }
-
-  bool Member() const {
-    // a semiring has only one zero, this is the important property
-    // we're trying to maintain here.  So force string_ to be empty if
-    // w_ == zero.
-    if (!weight_.Member()) return false;
-    if (weight_ == WeightType::Zero())
-      return string_.empty();
-    else
-      return true;
-  }
-
-  CompactLatticeWeightTpl Quantize(float delta = kDelta) const {
-    return CompactLatticeWeightTpl(weight_.Quantize(delta), string_);
-  }
-
-  static constexpr uint64 Properties() {
-    return kLeftSemiring | kRightSemiring | kPath | kIdempotent;
-  }
-
-  // This is used in OpenFst for binary I/O.  This is OpenFst-style,
-  // not Kaldi-style, I/O.
-  std::istream &Read(std::istream &strm) {
-    weight_.Read(strm);
-    if (strm.fail()) {
-      return strm;
-    }
-    int32 sz;
-    ReadType(strm, &sz);
-    if (strm.fail()) {
-      return strm;
-    }
-    if (sz < 0) {
-      KALDIFST_WARN << "Negative string size!  Read failure";
-      strm.clear(std::ios::badbit);
-      return strm;
-    }
-    string_.resize(sz);
-    for (int32 i = 0; i < sz; i++) {
-      ReadType(strm, &(string_[i]));
-    }
-    return strm;
-  }
-
-  // This is used in OpenFst for binary I/O.  This is OpenFst-style,
-  // not Kaldi-style, I/O.
-  std::ostream &Write(std::ostream &strm) const {
-    weight_.Write(strm);
-    if (strm.fail()) {
-      return strm;
-    }
-    int32 sz = static_cast<int32>(string_.size());
-    WriteType(strm, sz);
-    for (int32 i = 0; i < sz; i++) WriteType(strm, string_[i]);
-    return strm;
-  }
-  size_t Hash() const {
-    size_t ans = weight_.Hash();
-    // any weird numbers here are largish primes
-    size_t sz = string_.size(), mult = 6967;
-    for (size_t i = 0; i < sz; i++) {
-      ans += string_[i] * mult;
-      mult *= 7499;
-    }
-    return ans;
-  }
-
- private:
-  W weight_;
-  std::vector<IntType> string_;
-};
-
-template <class WeightType, class IntType>
-inline bool operator==(const CompactLatticeWeightTpl<WeightType, IntType> &w1,
-                       const CompactLatticeWeightTpl<WeightType, IntType> &w2) {
-  return (w1.Weight() == w2.Weight() && w1.String() == w2.String());
-}
-
-template <class WeightType, class IntType>
-inline bool operator!=(const CompactLatticeWeightTpl<WeightType, IntType> &w1,
-                       const CompactLatticeWeightTpl<WeightType, IntType> &w2) {
-  return (w1.Weight() != w2.Weight() || w1.String() != w2.String());
-}
-
-template <class WeightType, class IntType>
-inline bool ApproxEqual(const CompactLatticeWeightTpl<WeightType, IntType> &w1,
-                        const CompactLatticeWeightTpl<WeightType, IntType> &w2,
-                        float delta = kDelta) {
-  return (ApproxEqual(w1.Weight(), w2.Weight(), delta) &&
-          w1.String() == w2.String());
-}
-
-// Compare is not part of the standard for weight types, but used internally for
-// efficiency.  The comparison here first compares the weight; if this is the
-// same, it compares the string.  The comparison on strings is: first compare
-// the length, if this is the same, use lexicographical order.  We can't just
-// use the lexicographical order because this would destroy the distributive
-// property of multiplication over addition, taking into account that addition
-// uses Compare.  The string element of "Compare" isn't super-important in
-// practical terms; it's only needed to ensure that Plus always give consistent
-// answers and is symmetric.  It's essentially for tie-breaking, but we need to
-// make sure all the semiring axioms are satisfied otherwise OpenFst might
-// break.
-
-template <class WeightType, class IntType>
-inline int Compare(const CompactLatticeWeightTpl<WeightType, IntType> &w1,
-                   const CompactLatticeWeightTpl<WeightType, IntType> &w2) {
-  int c1 = Compare(w1.Weight(), w2.Weight());
-  if (c1 != 0) return c1;
-  int l1 = w1.String().size(), l2 = w2.String().size();
-  // Use opposite order on the string lengths, so that if the costs are the
-  // same, the shorter string wins.
-  if (l1 > l2)
-    return -1;
-  else if (l1 < l2)
-    return 1;
-  for (int i = 0; i < l1; i++) {
-    if (w1.String()[i] < w2.String()[i])
-      return -1;
-    else if (w1.String()[i] > w2.String()[i])
-      return 1;
-  }
-  return 0;
-}
-
-// For efficiency, override the NaturalLess template class.
-template <class FloatType, class IntType>
-class NaturalLess<
-    CompactLatticeWeightTpl<LatticeWeightTpl<FloatType>, IntType>> {
- public:
-  typedef CompactLatticeWeightTpl<LatticeWeightTpl<FloatType>, IntType> Weight;
-
-  NaturalLess() {}
-
-  bool operator()(const Weight &w1, const Weight &w2) const {
-    // NaturalLess is a negative order (opposite to normal ordering).
-    // This operator () corresponds to "<" in the negative order, which
-    // corresponds to the ">" in the normal order.
-    return (Compare(w1, w2) == 1);
-  }
-};
-template <>
-class NaturalLess<CompactLatticeWeightTpl<LatticeWeightTpl<float>, int32>> {
- public:
-  typedef CompactLatticeWeightTpl<LatticeWeightTpl<float>, int32> Weight;
-
-  NaturalLess() {}
-
-  bool operator()(const Weight &w1, const Weight &w2) const {
-    // NaturalLess is a negative order (opposite to normal ordering).
-    // This operator () corresponds to "<" in the negative order, which
-    // corresponds to the ">" in the normal order.
-    return (Compare(w1, w2) == 1);
-  }
-};
-template <>
-class NaturalLess<CompactLatticeWeightTpl<LatticeWeightTpl<double>, int32>> {
- public:
-  typedef CompactLatticeWeightTpl<LatticeWeightTpl<double>, int32> Weight;
-
-  NaturalLess() {}
-
-  bool operator()(const Weight &w1, const Weight &w2) const {
-    // NaturalLess is a negative order (opposite to normal ordering).
-    // This operator () corresponds to "<" in the negative order, which
-    // corresponds to the ">" in the normal order.
-    return (Compare(w1, w2) == 1);
-  }
-};
-
-// Make sure Compare is defined for TropicalWeight, so everything works
-// if we substitute LatticeWeight for TropicalWeight.
-inline int Compare(const TropicalWeight &w1, const TropicalWeight &w2) {
-  float f1 = w1.Value(), f2 = w2.Value();
-  if (f1 == f2)
-    return 0;
-  else if (f1 > f2)
-    return -1;
-  else
-    return 1;
-}
-
-template <class WeightType, class IntType>
-inline CompactLatticeWeightTpl<WeightType, IntType> Plus(
-    const CompactLatticeWeightTpl<WeightType, IntType> &w1,
-    const CompactLatticeWeightTpl<WeightType, IntType> &w2) {
-  return (Compare(w1, w2) >= 0 ? w1 : w2);
-}
-
-template <class WeightType, class IntType>
-inline CompactLatticeWeightTpl<WeightType, IntType> Times(
-    const CompactLatticeWeightTpl<WeightType, IntType> &w1,
-    const CompactLatticeWeightTpl<WeightType, IntType> &w2) {
-  WeightType w = Times(w1.Weight(), w2.Weight());
-  if (w == WeightType::Zero()) {
-    return CompactLatticeWeightTpl<WeightType, IntType>::Zero();
-    // special case to ensure zero is unique
-  } else {
-    std::vector<IntType> v;
-    v.resize(w1.String().size() + w2.String().size());
-    typename std::vector<IntType>::iterator iter = v.begin();
-    iter = std::copy(w1.String().begin(), w1.String().end(),
-                     iter);  // returns end of first range.
-    std::copy(w2.String().begin(), w2.String().end(), iter);
-    return CompactLatticeWeightTpl<WeightType, IntType>(w, v);
-  }
-}
-
-template <class WeightType, class IntType>
-inline CompactLatticeWeightTpl<WeightType, IntType> Divide(
-    const CompactLatticeWeightTpl<WeightType, IntType> &w1,
-    const CompactLatticeWeightTpl<WeightType, IntType> &w2,
-    DivideType div = DIVIDE_ANY) {
-  if (w1.Weight() == WeightType::Zero()) {
-    if (w2.Weight() != WeightType::Zero()) {
-      return CompactLatticeWeightTpl<WeightType, IntType>::Zero();
-    } else {
-      KALDIFST_ERR << "Division by zero [0/0]";
-    }
-  } else if (w2.Weight() == WeightType::Zero()) {
-    KALDIFST_ERR << "Error: division by zero";
-  }
-  WeightType w = Divide(w1.Weight(), w2.Weight());
-
-  const std::vector<IntType> v1 = w1.String(), v2 = w2.String();
-  if (v2.size() > v1.size()) {
-    KALDIFST_ERR << "Cannot divide, length mismatch";
-  }
-  typename std::vector<IntType>::const_iterator v1b = v1.begin(),
-                                                v1e = v1.end(),
-                                                v2b = v2.begin(),
-                                                v2e = v2.end();
-  if (div == DIVIDE_LEFT) {
-    if (!std::equal(v2b, v2e,
-                    v1b)) {  // v2 must be identical to first part of v1.
-      KALDIFST_ERR << "Cannot divide, data mismatch";
-    }
-    return CompactLatticeWeightTpl<WeightType, IntType>(
-        w, std::vector<IntType>(v1b + (v2e - v2b),
-                                v1e));  // return last part of v1.
-  } else if (div == DIVIDE_RIGHT) {
-    if (!std::equal(
-            v2b, v2e,
-            v1e - (v2e - v2b))) {  // v2 must be identical to last part of v1.
-      KALDIFST_ERR << "Cannot divide, data mismatch";
-    }
-    return CompactLatticeWeightTpl<WeightType, IntType>(
-        w, std::vector<IntType>(
-               v1b, v1e - (v2e - v2b)));  // return first part of v1.
-
-  } else {
-    KALDIFST_ERR << "Cannot divide CompactLatticeWeightTpl with DIVIDE_ANY";
-  }
-  return CompactLatticeWeightTpl<WeightType,
-                                 IntType>::Zero();  // keep compiler happy.
-}
-
-template <class WeightType, class IntType>
-inline std::ostream &operator<<(
-    std::ostream &strm, const CompactLatticeWeightTpl<WeightType, IntType> &w) {
-  strm << w.Weight();
-  CHECK_EQ(FLAGS_fst_weight_separator.size(), 1);
-  strm << FLAGS_fst_weight_separator[0];  // comma by default.
-  for (size_t i = 0; i < w.String().size(); i++) {
-    strm << w.String()[i];
-    if (i + 1 < w.String().size())
-      strm << kStringSeparator;  // '_'; defined in string-weight.h in OpenFst
-                                 // code.
-  }
-  return strm;
-}
-
-template <class WeightType, class IntType>
-inline std::istream &operator>>(
-    std::istream &strm, CompactLatticeWeightTpl<WeightType, IntType> &w) {
-  std::string s;
-  strm >> s;
-  if (strm.fail()) {
-    return strm;
-  }
-  CHECK_EQ(FLAGS_fst_weight_separator.size(), 1);
-  size_t pos = s.find_last_of(FLAGS_fst_weight_separator);  // normally ","
-  if (pos == std::string::npos) {
-    strm.clear(std::ios::badbit);
-    return strm;
-  }
-  // get parts of str before and after the separator (default: ',');
-  std::string s1(s, 0, pos), s2(s, pos + 1);
-  std::istringstream strm1(s1);
-  WeightType weight;
-  strm1 >> weight;
-  w.SetWeight(weight);
-  if (strm1.fail() || !strm1.eof()) {
-    strm.clear(std::ios::badbit);
-    return strm;
-  }
-  // read string part.
-  std::vector<IntType> string;
-  const char *c = s2.c_str();
-  while (*c != '\0') {
-    if (*c == kStringSeparator)  // '_'
-      c++;
-    char *c2;
-    long int i = strtol(c, &c2, 10);  // NOLINT
-    if (c2 == c ||
-        static_cast<long int>(static_cast<IntType>(i)) != i) {  // NOLINT
-      strm.clear(std::ios::badbit);
-      return strm;
-    }
-    c = c2;
-    string.push_back(static_cast<IntType>(i));
-  }
-  w.SetString(string);
-  return strm;
-}
-
-template <class BaseWeightType, class IntType>
-class CompactLatticeWeightCommonDivisorTpl {
- public:
-  typedef CompactLatticeWeightTpl<BaseWeightType, IntType> Weight;
-
-  Weight operator()(const Weight &w1, const Weight &w2) const {
-    // First find longest common prefix of the strings.
-    typename std::vector<IntType>::const_iterator s1b = w1.String().begin(),
-                                                  s1e = w1.String().end(),
-                                                  s2b = w2.String().begin(),
-                                                  s2e = w2.String().end();
-    while (s1b < s1e && s2b < s2e && *s1b == *s2b) {
-      s1b++;
-      s2b++;
-    }
-    return Weight(Plus(w1.Weight(), w2.Weight()),
-                  std::vector<IntType>(w1.String().begin(), s1b));
-  }
-};
-
-/** Scales the pair (a, b) of floating-point weights inside a
-    CompactLatticeWeight by premultiplying it (viewed as a vector)
-    by a 2x2 matrix "scale".
-    Assumes there is a ScaleTupleWeight function that applies to "Weight";
-    this currently only works if Weight equals LatticeWeightTpl<FloatType>
-    for some FloatType.
-*/
-template <class Weight, class IntType, class ScaleFloatType>
-inline CompactLatticeWeightTpl<Weight, IntType> ScaleTupleWeight(
-    const CompactLatticeWeightTpl<Weight, IntType> &w,
-    const std::vector<std::vector<ScaleFloatType>> &scale) {
-  return CompactLatticeWeightTpl<Weight, IntType>(
-      Weight(ScaleTupleWeight(w.Weight(), scale)), w.String());
-}
-
-/** Define some ConvertLatticeWeight functions that are used in various lattice
-    conversions... make them all templates, some with no arguments, since some
-    must be templates.*/
-template <class Float1, class Float2>
-inline void ConvertLatticeWeight(const LatticeWeightTpl<Float1> &w_in,
-                                 LatticeWeightTpl<Float2> *w_out) {
-  w_out->SetValue1(w_in.Value1());
-  w_out->SetValue2(w_in.Value2());
-}
-
-template <class Float1, class Float2, class Int>
-inline void ConvertLatticeWeight(
-    const CompactLatticeWeightTpl<LatticeWeightTpl<Float1>, Int> &w_in,
-    CompactLatticeWeightTpl<LatticeWeightTpl<Float2>, Int> *w_out) {
-  LatticeWeightTpl<Float2> weight2(w_in.Weight().Value1(),
-                                   w_in.Weight().Value2());
-  w_out->SetWeight(weight2);
-  w_out->SetString(w_in.String());
-}
-
-// to convert from Lattice to standard FST
-template <class Float1, class Float2>
-inline void ConvertLatticeWeight(const LatticeWeightTpl<Float1> &w_in,
-                                 TropicalWeightTpl<Float2> *w_out) {
-  TropicalWeightTpl<Float2> w1(w_in.Value1());
-  TropicalWeightTpl<Float2> w2(w_in.Value2());
-  *w_out = Times(w1, w2);
-}
-
-template <class Float>
-inline double ConvertToCost(const LatticeWeightTpl<Float> &w) {
-  return static_cast<double>(w.Value1()) + static_cast<double>(w.Value2());
-}
-
-template <class Float, class Int>
-inline double ConvertToCost(
-    const CompactLatticeWeightTpl<LatticeWeightTpl<Float>, Int> &w) {
-  return static_cast<double>(w.Weight().Value1()) +
-         static_cast<double>(w.Weight().Value2());
-}
-
-template <class Float>
-inline double ConvertToCost(const TropicalWeightTpl<Float> &w) {
-  return w.Value();
-}
-
-typedef LatticeWeightTpl<float> LatticeWeight;
-typedef ArcTpl<LatticeWeight> LatticeArc;
-typedef VectorFst<LatticeArc> Lattice;
-
-// careful: kaldi::int32 is not always the same C type as fst::int32
-typedef CompactLatticeWeightTpl<LatticeWeight, int32_t> CompactLatticeWeight;
-typedef CompactLatticeWeightCommonDivisorTpl<LatticeWeight, int32_t>
-    CompactLatticeWeightCommonDivisor;
-
-typedef ArcTpl<CompactLatticeWeight> CompactLatticeArc;
-
-typedef VectorFst<CompactLatticeArc> CompactLattice;
-}  // namespace fst
-
-#endif  // KALDIFST_CSRC_LATTICE_WEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/log.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/log.h
deleted file mode 100644
index 8f5d2a06..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/log.h
+++ /dev/null
@@ -1,100 +0,0 @@
-// kaldifst/csrc/log.h
-//
-// Copyright (c)  2020  Xiaomi Corporation (authors: Fangjun Kuang)
-
-#ifndef KALDIFST_CSRC_LOG_H_
-#define KALDIFST_CSRC_LOG_H_
-#include <cstdint>
-#include <iostream>
-#include <sstream>
-#include <stdexcept>
-
-namespace kaldifst {
-
-enum class LogLevel {
-  kInfo = 0,
-  kWarn = 1,
-  kError = 2,  // abort the program
-};
-
-class Logger {
- public:
-  Logger(const char *filename, const char *func_name, uint32_t line_num,
-         LogLevel level)
-      : filename_(filename),
-        func_name_(func_name),
-        line_num_(line_num),
-        level_(level) {
-    os_ << filename << ":" << func_name << ":" << line_num << "\n";
-    switch (level_) {
-      case LogLevel::kInfo:
-        os_ << "[I] ";
-        break;
-      case LogLevel::kWarn:
-        os_ << "[W] ";
-        break;
-      case LogLevel::kError:
-        os_ << "[E] ";
-        break;
-    }
-  }
-
-  template <typename T>
-  Logger &operator<<(const T &val) {
-    os_ << val;
-    return *this;
-  }
-
-  ~Logger() noexcept(false) {
-    if (level_ == LogLevel::kError) {
-      throw std::runtime_error(os_.str());
-    }
-  }
-
- private:
-  std::ostringstream os_;
-  const char *filename_;
-  const char *func_name_;
-  uint32_t line_num_;
-  LogLevel level_;
-};
-
-class Voidifier {
- public:
-  void operator&(const Logger &) const {}
-};
-
-#if defined(__clang__) || defined(__GNUC__) || defined(__GNUG__) || \
-    defined(__PRETTY_FUNCTION__)
-// for clang and GCC
-#define KALDIFST_FUNC __PRETTY_FUNCTION__
-#else
-// for other compilers
-#define KALDIFST_FUNC __func__
-#endif
-
-#define KALDIFST_LOG \
-  kaldifst::Logger(__FILE__, KALDIFST_FUNC, __LINE__, kaldifst::LogLevel::kInfo)
-
-#define KALDIFST_WARN \
-  kaldifst::Logger(__FILE__, KALDIFST_FUNC, __LINE__, kaldifst::LogLevel::kWarn)
-
-#define KALDIFST_ERR                                  \
-  kaldifst::Logger(__FILE__, KALDIFST_FUNC, __LINE__, \
-                   kaldifst::LogLevel::kError)
-
-#define KALDIFST_ASSERT(x)                                        \
-  (x) ? (void)0                                                   \
-      : kaldifst::Voidifier() & KALDIFST_ERR << "Check failed!\n" \
-                                             << "x: " << #x
-
-#define KALDIFST_PARANOID_ASSERT KALDIFST_ASSERT
-
-#define KALDIFST_DISALLOW_COPY_AND_ASSIGN(Class) \
- public:                                         \
-  Class(const Class &) = delete;                 \
-  Class &operator=(const Class &) = delete;
-
-}  // namespace kaldifst
-
-#endif  // KALDIFST_CSRC_LOG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/parse-options.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/parse-options.h
deleted file mode 100644
index 614dfcdc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/parse-options.h
+++ /dev/null
@@ -1,26 +0,0 @@
-// kaldifst/csrc/parse-options.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/util/parse-options.h
-
-// Copyright 2009-2011  Karel Vesely;  Microsoft Corporation;
-//                      Saarland University (Author: Arnab Ghoshal);
-// Copyright 2012-2013  Frantisek Skala;  Arnab Ghoshal
-
-#ifndef KALDIFST_CSRC_PARSE_OPTIONS_H_
-#define KALDIFST_CSRC_PARSE_OPTIONS_H_
-
-#include <string>
-
-namespace kaldifst {
-
-class ParseOptions {
- public:
-  /// The following function will return a possibly quoted and escaped
-  /// version of "str", according to the current shell.  Currently
-  /// this is just hardwired to bash.  It's useful for debug output.
-  static std::string Escape(const std::string &str);
-};
-
-}  // namespace kaldifst
-#endif  // KALDIFST_CSRC_PARSE_OPTIONS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/pre-determinize-inl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/pre-determinize-inl.h
deleted file mode 100644
index d7fc3711..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/pre-determinize-inl.h
+++ /dev/null
@@ -1,94 +0,0 @@
-// kaldifst/csrc/pre-determinize-inl.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/pre-determinize-inl.h
-
-// Copyright 2009-2011  Microsoft Corporation
-
-#ifndef KALDIFST_CSRC_PRE_DETERMINIZE_INL_H_
-#define KALDIFST_CSRC_PRE_DETERMINIZE_INL_H_
-
-#include <assert.h>
-
-#include <algorithm>
-#include <set>
-#include <vector>
-
-namespace fst {
-
-// see pre-determinize.h for documentation.
-template <class Arc>
-void AddSelfLoops(MutableFst<Arc> *fst,
-                  const std::vector<typename Arc::Label> &isyms,
-                  const std::vector<typename Arc::Label> &osyms) {
-  assert(fst != nullptr);
-  assert(isyms.size() == osyms.size());
-  typedef typename Arc::Label Label;
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Weight Weight;
-  size_t n = isyms.size();
-  if (n == 0) return;  // Nothing to do.
-
-  // {
-  // the following declarations and statements are for quick detection of these
-  // symbols, which is purely for debugging/checking purposes.
-  Label isyms_min = *std::min_element(isyms.begin(), isyms.end()),
-        isyms_max = *std::max_element(isyms.begin(), isyms.end()),
-        osyms_min = *std::min_element(osyms.begin(), osyms.end()),
-        osyms_max = *std::max_element(osyms.begin(), osyms.end());
-  std::set<Label> isyms_set, osyms_set;
-  for (size_t i = 0; i < isyms.size(); i++) {
-    assert(isyms[i] > 0 &&
-           osyms[i] > 0);  // should not have epsilon or invalid symbols.
-    isyms_set.insert(isyms[i]);
-    osyms_set.insert(osyms[i]);
-  }
-  assert(isyms_set.size() == n && osyms_set.size() == n);
-  // } end block.
-
-  for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
-       siter.Next()) {
-    StateId state = siter.Value();
-    bool this_state_needs_self_loops = (fst->Final(state) != Weight::Zero());
-    for (ArcIterator<MutableFst<Arc>> aiter(*fst, state); !aiter.Done();
-         aiter.Next()) {
-      const Arc &arc = aiter.Value();
-      // If one of the following asserts fails, it means that the input FST
-      // already had the symbols we are inserting.  This is contrary to the
-      // preconditions of this algorithm.
-      assert(!(arc.ilabel >= isyms_min && arc.ilabel <= isyms_max &&
-               isyms_set.count(arc.ilabel) != 0));
-      assert(!(arc.olabel >= osyms_min && arc.olabel <= osyms_max &&
-               osyms_set.count(arc.olabel) != 0));
-      if (arc.olabel != 0)  // Has non-epsilon output label -> need self loops.
-        this_state_needs_self_loops = true;
-    }
-    if (this_state_needs_self_loops) {
-      for (size_t i = 0; i < n; i++) {
-        Arc arc;
-        arc.ilabel = isyms[i];
-        arc.olabel = osyms[i];
-        arc.weight = Weight::One();
-        arc.nextstate = state;
-        fst->AddArc(state, arc);
-      }
-    }
-  }
-}
-
-template <class T>
-void CopySetToVector(const std::set<T> s, std::vector<T> *v) {
-  // adds members of s to v, in sorted order from lowest to highest
-  // (because the set was in sorted order).
-  assert(v != nullptr);
-  v->resize(s.size());
-  typename std::set<T>::const_iterator siter = s.begin();
-  typename std::vector<T>::iterator viter = v->begin();
-  for (; siter != s.end(); ++siter, ++viter) {
-    assert(viter != v->end());
-    *viter = *siter;
-  }
-}
-
-}  // namespace fst
-#endif  // KALDIFST_CSRC_PRE_DETERMINIZE_INL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/pre-determinize.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/pre-determinize.h
deleted file mode 100644
index 9ecc92a4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/pre-determinize.h
+++ /dev/null
@@ -1,44 +0,0 @@
-// kaldifst/csrc/pre-determinize.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/pre-determinize.h
-
-// Copyright 2009-2011  Microsoft Corporation
-
-#ifndef KALDIFST_CSRC_PRE_DETERMINIZE_H_
-#define KALDIFST_CSRC_PRE_DETERMINIZE_H_
-
-#include <algorithm>
-#include <vector>
-
-#include "fst/fst-decl.h"
-#include "fst/fstlib.h"
-
-namespace fst {
-
-/** AddSelfLoops is a function you will probably want to use alongside
-   PreDeterminize, to add self-loops to any FSTs that you compose on the left
-   hand side of the one modified by PreDeterminize.
-
-    This function inserts loops with "special symbols" [e.g. \#0, \#1] into an
-   FST. This is done at each final state and each state with non-epsilon output
-   symbols on at least one arc out of it.  This is to ensure that these symbols,
-   when inserted into the input side of an FST we will compose with on the
-   right, can "pass through" this FST.
-
-    At input, isyms and osyms must be vectors of the same size n, corresponding
-    to symbols that currently do not exist in 'fst'.  For each state in n that
-   has non-epsilon symbols on the output side of arcs leaving it, or which is a
-   final state, this function inserts n self-loops with unit weight and one of
-   the n pairs of symbols on its input and output.
-*/
-template <class Arc>
-void AddSelfLoops(MutableFst<Arc> *fst,
-                  const std::vector<typename Arc::Label> &isyms,
-                  const std::vector<typename Arc::Label> &osyms);
-
-}  // namespace fst
-
-#include "kaldifst/csrc/pre-determinize-inl.h"
-
-#endif  // KALDIFST_CSRC_PRE_DETERMINIZE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/remove-eps-local-inl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/remove-eps-local-inl.h
deleted file mode 100644
index 7c3f9eea..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/remove-eps-local-inl.h
+++ /dev/null
@@ -1,306 +0,0 @@
-// kaldifst/csrc/remove-eps-local-inl.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/remove-eps-local-inl.h
-
-// Copyright 2009-2011  Microsoft Corporation
-//                2014  Johns Hopkins University (author: Daniel Povey
-
-#ifndef KALDIFST_CSRC_REMOVE_EPS_LOCAL_INL_H_
-#define KALDIFST_CSRC_REMOVE_EPS_LOCAL_INL_H_
-
-#include <vector>
-
-namespace fst {
-
-template <class Weight>
-struct ReweightPlusDefault {
-  inline Weight operator()(const Weight &a, const Weight &b) {
-    return Plus(a, b);
-  }
-};
-
-struct ReweightPlusLogArc {
-  inline TropicalWeight operator()(const TropicalWeight &a,
-                                   const TropicalWeight &b) {
-    LogWeight a_log(a.Value()), b_log(b.Value());
-    return TropicalWeight(Plus(a_log, b_log).Value());
-  }
-};
-
-template <class Arc,
-          class ReweightPlus = ReweightPlusDefault<typename Arc::Weight>>
-class RemoveEpsLocalClass {
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Label Label;
-  typedef typename Arc::Weight Weight;
-
- public:
-  RemoveEpsLocalClass(MutableFst<Arc> *fst) : fst_(fst) {  // NOLINT
-    if (fst_->Start() == kNoStateId) return;               // empty.
-    non_coacc_state_ = fst_->AddState();
-    InitNumArcs();
-    StateId num_states = fst_->NumStates();
-    for (StateId s = 0; s < num_states; s++)
-      for (size_t pos = 0; pos < fst_->NumArcs(s); pos++) RemoveEps(s, pos);
-    assert(CheckNumArcs());
-    Connect(fst);  // remove inaccessible states.
-  }
-
- private:
-  MutableFst<Arc> *fst_;
-  StateId non_coacc_state_;  //  use this to delete arcs: make it nextstate
-  std::vector<StateId> num_arcs_in_;  // The number of arcs into the state, plus
-                                      // one if it's the start state.
-  std::vector<StateId> num_arcs_out_;  // The number of arcs out of the state,
-                                       // plus one if it's a final state.
-  ReweightPlus reweight_plus_;
-
-  bool CanCombineArcs(const Arc &a, const Arc &b, Arc *c) {
-    if (a.ilabel != 0 && b.ilabel != 0) return false;
-    if (a.olabel != 0 && b.olabel != 0) return false;
-    c->weight = Times(a.weight, b.weight);
-    c->ilabel = (a.ilabel != 0 ? a.ilabel : b.ilabel);
-    c->olabel = (a.olabel != 0 ? a.olabel : b.olabel);
-    c->nextstate = b.nextstate;
-    return true;
-  }
-
-  static bool CanCombineFinal(const Arc &a, Weight final_prob,
-                              Weight *final_prob_out) {
-    if (a.ilabel != 0 || a.olabel != 0) {
-      return false;
-    } else {
-      *final_prob_out = Times(a.weight, final_prob);
-      return true;
-    }
-  }
-
-  void InitNumArcs() {  // init num transitions in/out of each state.
-    StateId num_states = fst_->NumStates();
-    num_arcs_in_.resize(num_states);
-    num_arcs_out_.resize(num_states);
-    num_arcs_in_[fst_->Start()]++;  // count start as trans in.
-    for (StateId s = 0; s < num_states; s++) {
-      if (fst_->Final(s) != Weight::Zero())
-        num_arcs_out_[s]++;  // count final as transition.
-      for (ArcIterator<MutableFst<Arc>> aiter(*fst_, s); !aiter.Done();
-           aiter.Next()) {
-        num_arcs_in_[aiter.Value().nextstate]++;
-        num_arcs_out_[s]++;
-      }
-    }
-  }
-
-  bool CheckNumArcs() {  // check num arcs in/out of each state, at end.  Debug.
-    num_arcs_in_[fst_->Start()]--;  // count start as trans in.
-    StateId num_states = fst_->NumStates();
-    for (StateId s = 0; s < num_states; s++) {
-      if (s == non_coacc_state_) continue;
-      if (fst_->Final(s) != Weight::Zero())
-        num_arcs_out_[s]--;  // count final as transition.
-      for (ArcIterator<MutableFst<Arc>> aiter(*fst_, s); !aiter.Done();
-           aiter.Next()) {
-        if (aiter.Value().nextstate == non_coacc_state_) continue;
-        num_arcs_in_[aiter.Value().nextstate]--;
-        num_arcs_out_[s]--;
-      }
-    }
-    for (StateId s = 0; s < num_states; s++) {
-      assert(num_arcs_in_[s] == 0);
-      assert(num_arcs_out_[s] == 0);
-    }
-    return true;  // always does this.  so we can assert it w/o warnings.
-  }
-
-  inline void GetArc(StateId s, size_t pos, Arc *arc) const {
-    ArcIterator<MutableFst<Arc>> aiter(*fst_, s);
-    aiter.Seek(pos);
-    *arc = aiter.Value();
-  }
-
-  inline void SetArc(StateId s, size_t pos, const Arc &arc) {
-    MutableArcIterator<MutableFst<Arc>> aiter(fst_, s);
-    aiter.Seek(pos);
-    aiter.SetValue(arc);
-  }
-
-  void Reweight(StateId s, size_t pos, Weight reweight) {
-    // Reweight is called from RemoveEpsPattern1; it is a step we
-    // do to preserve stochasticity.  This function multiplies the
-    // arc at (s, pos) by reweight and divides all the arcs [+final-prob]
-    // out of the next state by the same.  This is only valid if
-    // the next state has only one arc in and is not the start state.
-    assert(reweight != Weight::Zero());
-    MutableArcIterator<MutableFst<Arc>> aiter(fst_, s);
-    aiter.Seek(pos);
-    Arc arc = aiter.Value();
-    assert(num_arcs_in_[arc.nextstate] == 1);
-    arc.weight = Times(arc.weight, reweight);
-    aiter.SetValue(arc);
-
-    for (MutableArcIterator<MutableFst<Arc>> aiter_next(fst_, arc.nextstate);
-         !aiter_next.Done(); aiter_next.Next()) {
-      Arc nextarc = aiter_next.Value();
-      if (nextarc.nextstate != non_coacc_state_) {
-        nextarc.weight = Divide(nextarc.weight, reweight, DIVIDE_LEFT);
-        aiter_next.SetValue(nextarc);
-      }
-    }
-    Weight final = fst_->Final(arc.nextstate);
-    if (final != Weight::Zero()) {
-      fst_->SetFinal(arc.nextstate, Divide(final, reweight, DIVIDE_LEFT));
-    }
-  }
-
-  // RemoveEpsPattern1 applies where this arc, which is not a
-  // self-loop, enters a state which has only one input transition
-  // [and is not the start state], and has multiple output
-  // transitions [counting being the final-state as a final-transition].
-
-  void RemoveEpsPattern1(StateId s, size_t pos, Arc arc) {
-    const StateId nextstate = arc.nextstate;
-    Weight total_removed = Weight::Zero(),
-           total_kept = Weight::Zero();  // totals out of nextstate.
-    std::vector<Arc> arcs_to_add;        // to add to state s.
-    for (MutableArcIterator<MutableFst<Arc>> aiter_next(fst_, nextstate);
-         !aiter_next.Done(); aiter_next.Next()) {
-      Arc nextarc = aiter_next.Value();
-      if (nextarc.nextstate == non_coacc_state_) continue;  // deleted.
-      Arc combined;
-      if (CanCombineArcs(arc, nextarc, &combined)) {
-        total_removed = reweight_plus_(total_removed, nextarc.weight);
-        num_arcs_out_[nextstate]--;
-        num_arcs_in_[nextarc.nextstate]--;
-        nextarc.nextstate = non_coacc_state_;
-        aiter_next.SetValue(nextarc);
-        arcs_to_add.push_back(combined);
-      } else {
-        total_kept = reweight_plus_(total_kept, nextarc.weight);
-      }
-    }
-
-    {  // now final-state.
-      Weight next_final = fst_->Final(nextstate);
-      if (next_final != Weight::Zero()) {
-        Weight new_final;
-        if (CanCombineFinal(arc, next_final, &new_final)) {
-          total_removed = reweight_plus_(total_removed, next_final);
-          if (fst_->Final(s) == Weight::Zero())
-            num_arcs_out_[s]++;  // final is counted as arc.
-          fst_->SetFinal(s, Plus(fst_->Final(s), new_final));
-          num_arcs_out_[nextstate]--;
-          fst_->SetFinal(nextstate, Weight::Zero());
-        } else {
-          total_kept = reweight_plus_(total_kept, next_final);
-        }
-      }
-    }
-
-    if (total_removed != Weight::Zero()) {  // did something...
-      if (total_kept == Weight::Zero()) {   // removed everything: remove arc.
-        num_arcs_out_[s]--;
-        num_arcs_in_[arc.nextstate]--;
-        arc.nextstate = non_coacc_state_;
-        SetArc(s, pos, arc);
-      } else {
-        // Have to reweight.
-        Weight total = reweight_plus_(total_removed, total_kept);
-        Weight reweight = Divide(total_kept, total, DIVIDE_LEFT);  // <=1
-        Reweight(s, pos, reweight);
-      }
-    }
-    // Now add the arcs we were going to add.
-    for (size_t i = 0; i < arcs_to_add.size(); i++) {
-      num_arcs_out_[s]++;
-      num_arcs_in_[arcs_to_add[i].nextstate]++;
-      fst_->AddArc(s, arcs_to_add[i]);
-    }
-  }
-
-  void RemoveEpsPattern2(StateId s, size_t pos, Arc arc) {
-    // Pattern 2 is where "nextstate" has only one arc out, counting
-    // being-the-final-state as an arc, but possibly multiple arcs in.
-    // Also, nextstate != s.
-
-    const StateId nextstate = arc.nextstate;
-    bool can_delete_next = (num_arcs_in_[nextstate] == 1);  // if
-    // we combine, can delete the corresponding out-arc/final-prob
-    // of nextstate.
-    bool delete_arc = false;  // set to true if this arc to be deleted.
-
-    Weight next_final = fst_->Final(arc.nextstate);
-    if (next_final !=
-        Weight::Zero()) {  // nextstate has no actual arcs out, only final-prob.
-      Weight new_final;
-      if (CanCombineFinal(arc, next_final, &new_final)) {
-        if (fst_->Final(s) == Weight::Zero())
-          num_arcs_out_[s]++;  // final is counted as arc.
-        fst_->SetFinal(s, Plus(fst_->Final(s), new_final));
-        delete_arc = true;  // will delete "arc".
-        if (can_delete_next) {
-          num_arcs_out_[nextstate]--;
-          fst_->SetFinal(nextstate, Weight::Zero());
-        }
-      }
-    } else {  // has an arc but no final prob.
-      MutableArcIterator<MutableFst<Arc>> aiter_next(fst_, nextstate);
-      assert(!aiter_next.Done());
-      while (aiter_next.Value().nextstate == non_coacc_state_) {
-        aiter_next.Next();
-        assert(!aiter_next.Done());
-      }
-      // now aiter_next points to a real arc out of nextstate.
-      Arc nextarc = aiter_next.Value();
-      Arc combined;
-      if (CanCombineArcs(arc, nextarc, &combined)) {
-        delete_arc = true;
-        if (can_delete_next) {  // do it before we invalidate iterators
-          num_arcs_out_[nextstate]--;
-          num_arcs_in_[nextarc.nextstate]--;
-          nextarc.nextstate = non_coacc_state_;
-          aiter_next.SetValue(nextarc);
-        }
-        num_arcs_out_[s]++;
-        num_arcs_in_[combined.nextstate]++;
-        fst_->AddArc(s, combined);
-      }
-    }
-    if (delete_arc) {
-      num_arcs_out_[s]--;
-      num_arcs_in_[nextstate]--;
-      arc.nextstate = non_coacc_state_;
-      SetArc(s, pos, arc);
-    }
-  }
-
-  void RemoveEps(StateId s, size_t pos) {
-    // Tries to do local epsilon-removal for arc sequences starting with this
-    // arc
-    Arc arc;
-    GetArc(s, pos, &arc);
-    StateId nextstate = arc.nextstate;
-    if (nextstate == non_coacc_state_) return;  // deleted arc.
-    if (nextstate == s) return;  // don't handle self-loops: too complex.
-
-    if (num_arcs_in_[nextstate] == 1 && num_arcs_out_[nextstate] > 1) {
-      RemoveEpsPattern1(s, pos, arc);
-    } else if (num_arcs_out_[nextstate] == 1) {
-      RemoveEpsPattern2(s, pos, arc);
-    }
-  }
-};
-
-template <class Arc>
-void RemoveEpsLocal(MutableFst<Arc> *fst) {
-  RemoveEpsLocalClass<Arc> c(fst);  // work gets done in initializer.
-}
-
-void RemoveEpsLocalSpecial(MutableFst<StdArc> *fst) {
-  // work gets done in initializer.
-  RemoveEpsLocalClass<StdArc, ReweightPlusLogArc> c(fst);
-}
-
-}  // end namespace fst
-
-#endif  // KALDIFST_CSRC_REMOVE_EPS_LOCAL_INL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/remove-eps-local.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/remove-eps-local.h
deleted file mode 100644
index 2698d64b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/remove-eps-local.h
+++ /dev/null
@@ -1,45 +0,0 @@
-// kaldifst/csrc/remove-eps-local.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/remove-eps-local.h
-
-// Copyright 2009-2011  Microsoft Corporation
-//                2014  Johns Hopkins University (author: Daniel Povey)
-
-#ifndef KALDIFST_CSRC_REMOVE_EPS_LOCAL_H_
-#define KALDIFST_CSRC_REMOVE_EPS_LOCAL_H_
-
-#include "fst/fst-decl.h"
-#include "fst/fstlib.h"
-
-namespace fst {
-
-/// RemoveEpsLocal remove some (but not necessarily all) epsilons in an FST,
-/// using an algorithm that is guaranteed to never increase the number of arcs
-/// in the FST (and will also never increase the number of states).  The
-/// algorithm is not optimal but is reasonably clever.  It does not just remove
-/// epsilon arcs;it also combines pairs of input-epsilon and output-epsilon arcs
-/// into one.
-/// The algorithm preserves equivalence and stochasticity in the given semiring.
-/// If you want to preserve stochasticity in a different semiring (e.g. log),
-/// then use RemoveEpsLocalSpecial, which only works for StdArc but which
-/// preserves stochasticity, where possible (*) in the LogArc sense.  The reason
-/// that we can't just cast to a different semiring is that in that case we
-/// would no longer be able to guarantee equivalence in the original semiring
-/// (this arises from what happens when we combine identical arcs).
-/// (*) by "where possible".. there are situations where we wouldn't be able to
-/// preserve stochasticity in the LogArc sense while maintaining equivalence in
-/// the StdArc sense, so in these situations we maintain equivalence.
-
-template <class Arc>
-void RemoveEpsLocal(MutableFst<Arc> *fst);
-
-/// As RemoveEpsLocal but takes care to preserve stochasticity
-/// when cast to LogArc.
-inline void RemoveEpsLocalSpecial(MutableFst<StdArc> *fst);
-
-}  // namespace fst
-
-#include "kaldifst/csrc/remove-eps-local-inl.h"
-
-#endif  //  KALDIFST_CSRC_REMOVE_EPS_LOCAL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/stl-utils.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/stl-utils.h
deleted file mode 100644
index da77b386..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/stl-utils.h
+++ /dev/null
@@ -1,105 +0,0 @@
-// kaldifst/csrc/stl-utils.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/util/stl-utils.h
-
-// Copyright 2009-2011  Microsoft Corporation;  Saarland University
-
-#ifndef KALDIFST_CSRC_STL_UTILS_H_
-#define KALDIFST_CSRC_STL_UTILS_H_
-#include <algorithm>
-#include <set>
-#include <string>
-#include <type_traits>
-#include <unordered_set>
-#include <utility>
-#include <vector>
-
-#include "kaldifst/csrc/log.h"
-
-namespace kaldifst {
-
-/// A hashing function object for strings.
-struct StringHasher {  // hashing function for std::string
-  size_t operator()(const std::string &str) const noexcept {
-    size_t ans = 0, len = str.length();
-    const char *c = str.c_str(), *end = c + len;
-    for (; c != end; c++) {
-      ans *= kPrime;
-      ans += *c;
-    }
-    return ans;
-  }
-
- private:
-  static const int kPrime = 7853;
-};
-
-/// A hashing function-object for vectors.
-template <typename Int>
-struct VectorHasher {  // hashing function for vector<Int>.
-  size_t operator()(const std::vector<Int> &x) const noexcept {
-    size_t ans = 0;
-    typename std::vector<Int>::const_iterator iter = x.begin(), end = x.end();
-    for (; iter != end; ++iter) {
-      ans *= kPrime;
-      ans += *iter;
-    }
-    return ans;
-  }
-  VectorHasher() {  // Check we're instantiated with an integer type.
-    static_assert(std::is_integral<Int>::value, "");
-  }
-
- private:
-  static const int kPrime = 7853;
-};
-
-/// Sorts and uniq's (removes duplicates) from a vector.
-template <typename T>
-inline void SortAndUniq(std::vector<T> *vec) {
-  std::sort(vec->begin(), vec->end());
-  vec->erase(std::unique(vec->begin(), vec->end()), vec->end());
-}
-
-/// A hashing function-object for pairs of ints
-template <typename Int1, typename Int2 = Int1>
-struct PairHasher {  // hashing function for pair<int>
-  size_t operator()(const std::pair<Int1, Int2> &x) const noexcept {
-    // 7853 was chosen at random from a list of primes.
-    return x.first + x.second * 7853;
-  }
-  PairHasher() {  // Check we're instantiated with an integer type.
-    static_assert(std::is_integral<Int1>::value, "");
-    static_assert(std::is_integral<Int2>::value, "");
-  }
-};
-
-/// Copies the elements of a set to a vector.
-template <class T>
-void CopySetToVector(const std::set<T> &s, std::vector<T> *v) {
-  // copies members of s into v, in sorted order from lowest to highest
-  // (because the set was in sorted order).
-  KALDIFST_ASSERT(v != nullptr);
-  v->resize(s.size());
-  auto siter = s.begin(), send = s.end();
-  auto viter = v->begin();
-  for (; siter != send; ++siter, ++viter) {
-    *viter = *siter;
-  }
-}
-
-template <class T>
-void CopySetToVector(const std::unordered_set<T> &s, std::vector<T> *v) {
-  KALDIFST_ASSERT(v != nullptr);
-  v->resize(s.size());
-  auto siter = s.begin(), send = s.end();
-  auto viter = v->begin();
-  for (; siter != send; ++siter, ++viter) {
-    *viter = *siter;
-  }
-}
-
-}  // namespace kaldifst
-
-#endif  // KALDIFST_CSRC_STL_UTILS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/table-matcher.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/table-matcher.h
deleted file mode 100644
index e96d156c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/table-matcher.h
+++ /dev/null
@@ -1,381 +0,0 @@
-// kaldifst/csrc/table-matcher.h
-//
-// Copyright 2009-2011  Microsoft Corporation
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/fstext/table-matcher.h
-
-// Copyright 2009-2011  Microsoft Corporation
-
-#ifndef KALDIFST_CSRC_TABLE_MATCHER_H_
-#define KALDIFST_CSRC_TABLE_MATCHER_H_
-
-#include <memory>
-#include <vector>
-
-#include "fst/fst-decl.h"
-#include "fst/fstlib.h"
-
-namespace fst {
-
-/// TableMatcher is a matcher specialized for the case where the output
-/// side of the left FST always has either all-epsilons coming out of
-/// a state, or a majority of the symbol table.  Therefore we can
-/// either store nothing (for the all-epsilon case) or store a lookup
-/// table from Labels to arc offsets.  Since the TableMatcher has to
-/// iterate over all arcs in each left-hand state the first time it sees
-/// it, this matcher type is not efficient if you compose with
-/// something very small on the right-- unless you do it multiple
-/// times and keep the matcher around. To do this requires using the
-/// most advanced form of ComposeFst in Compose.h, that initializes
-/// with ComposeFstImplOptions.
-
-struct TableMatcherOptions {
-  float
-      table_ratio;  // we construct the table if it would be at least this full.
-  int min_table_size;
-  TableMatcherOptions() : table_ratio(0.25), min_table_size(4) {}
-};
-
-// Introducing an "impl" class for TableMatcher because
-// we need to do a shallow copy of the Matcher for when
-// we want to cache tables for multiple compositions.
-template <class F, class BackoffMatcher = SortedMatcher<F>>
-class TableMatcherImpl : public MatcherBase<typename F::Arc> {
- public:
-  typedef F FST;
-  typedef typename F::Arc Arc;
-  typedef typename Arc::Label Label;
-  typedef typename Arc::StateId StateId;
-  typedef StateId
-      ArcId;  // Use this type to store arc offsets [it's actually size_t
-  // in the Seek function of ArcIterator, but StateId should be big enough].
-  typedef typename Arc::Weight Weight;
-
- public:
-  TableMatcherImpl(const FST &fst, MatchType match_type,
-                   const TableMatcherOptions &opts = TableMatcherOptions())
-      : match_type_(match_type),
-        fst_(fst.Copy()),
-        loop_(match_type == MATCH_INPUT
-                  ? Arc(kNoLabel, 0, Weight::One(), kNoStateId)
-                  : Arc(0, kNoLabel, Weight::One(), kNoStateId)),
-        aiter_(NULL),
-        s_(kNoStateId),
-        opts_(opts),
-        backoff_matcher_(fst, match_type) {
-    assert(opts_.min_table_size > 0);
-    if (match_type == MATCH_INPUT)
-      assert(fst_->Properties(kILabelSorted, true) == kILabelSorted);
-    else if (match_type == MATCH_OUTPUT)
-      assert(fst_->Properties(kOLabelSorted, true) == kOLabelSorted);
-    else
-      assert(0 && "Invalid FST properties");
-  }
-
-  virtual const FST &GetFst() const { return *fst_; }
-
-  virtual ~TableMatcherImpl() {
-    std::vector<ArcId> *const empty =
-        ((std::vector<ArcId> *)(NULL)) + 1;  // special marker.
-    for (size_t i = 0; i < tables_.size(); i++) {
-      if (tables_[i] != NULL && tables_[i] != empty) delete tables_[i];
-    }
-    delete aiter_;
-    delete fst_;
-  }
-
-  virtual MatchType Type(bool test) const { return match_type_; }
-
-  void SetState(StateId s) {
-    if (aiter_) {
-      delete aiter_;
-      aiter_ = NULL;
-    }
-    if (match_type_ == MATCH_NONE) LOG(FATAL) << "TableMatcher: bad match type";
-    s_ = s;
-    std::vector<ArcId> *const empty =
-        ((std::vector<ArcId> *)(NULL)) + 1;  // special marker.
-    if (static_cast<size_t>(s) >= tables_.size()) {
-      assert(s >= 0);
-      tables_.resize(s + 1, NULL);
-    }
-    std::vector<ArcId> *&this_table_ = tables_[s];  // note: ref to ptr.
-    if (this_table_ == empty) {
-      backoff_matcher_.SetState(s);
-      return;
-    } else if (this_table_ == NULL) {  // NULL means has not been set.
-      ArcId num_arcs = fst_->NumArcs(s);
-      if (num_arcs == 0 || num_arcs < opts_.min_table_size) {
-        this_table_ = empty;
-        backoff_matcher_.SetState(s);
-        return;
-      }
-      ArcIterator<FST> aiter(*fst_, s);
-      aiter.SetFlags(
-          kArcNoCache |
-              (match_type_ == MATCH_OUTPUT ? kArcOLabelValue : kArcILabelValue),
-          kArcNoCache | kArcValueFlags);
-      // the statement above, says: "Don't cache stuff; and I only need the
-      // ilabel/olabel to be computed.
-      aiter.Seek(num_arcs - 1);
-      Label highest_label =
-          (match_type_ == MATCH_OUTPUT ? aiter.Value().olabel
-                                       : aiter.Value().ilabel);
-      if ((highest_label + 1) * opts_.table_ratio > num_arcs) {
-        this_table_ = empty;
-        backoff_matcher_.SetState(s);
-        return;  // table would be too sparse.
-      }
-      // OK, now we are creating the table.
-      this_table_ = new std::vector<ArcId>(highest_label + 1, kNoStateId);
-      ArcId pos = 0;
-      for (aiter.Seek(0); !aiter.Done(); aiter.Next(), pos++) {
-        Label label = (match_type_ == MATCH_OUTPUT ? aiter.Value().olabel
-                                                   : aiter.Value().ilabel);
-        assert((size_t)label <= (size_t)highest_label);  // also checks >= 0.
-        if ((*this_table_)[label] == kNoStateId) (*this_table_)[label] = pos;
-        // set this_table_[label] to first position where arc has this
-        // label.
-      }
-    }
-    // At this point in the code, this_table_ != NULL and != empty.
-    aiter_ = new ArcIterator<FST>(*fst_, s);
-    aiter_->SetFlags(kArcNoCache,
-                     kArcNoCache);  // don't need to cache arcs as may only
-    // need a small subset.
-    loop_.nextstate = s;
-    // aiter_ = NULL;
-    // backoff_matcher_.SetState(s);
-  }
-
-  bool Find(Label match_label) {
-    if (!aiter_) {
-      return backoff_matcher_.Find(match_label);
-    } else {
-      match_label_ = match_label;
-      current_loop_ = (match_label == 0);
-      // kNoLabel means the implicit loop on the other FST --
-      // matches real epsilons but not the self-loop.
-      match_label_ = (match_label_ == kNoLabel ? 0 : match_label_);
-      if (static_cast<size_t>(match_label_) < tables_[s_]->size() &&
-          (*(tables_[s_]))[match_label_] != kNoStateId) {
-        aiter_->Seek((*(tables_[s_]))[match_label_]);  // label exists.
-        return true;
-      }
-      return current_loop_;
-    }
-  }
-  const Arc &Value() const {
-    if (aiter_)
-      return current_loop_ ? loop_ : aiter_->Value();
-    else
-      return backoff_matcher_.Value();
-  }
-
-  void Next() {
-    if (aiter_) {
-      if (current_loop_) {
-        current_loop_ = false;
-      } else {
-        aiter_->Next();
-      }
-    } else {
-      backoff_matcher_.Next();
-    }
-  }
-
-  bool Done() const {
-    if (aiter_ != NULL) {
-      if (current_loop_) {
-        return false;
-      }
-      if (aiter_->Done()) {
-        return true;
-      }
-      Label label = (match_type_ == MATCH_OUTPUT ? aiter_->Value().olabel
-                                                 : aiter_->Value().ilabel);
-      return (label != match_label_);
-    } else {
-      return backoff_matcher_.Done();
-    }
-  }
-  const Arc &Value() {
-    if (aiter_ != NULL) {
-      return (current_loop_ ? loop_ : aiter_->Value());
-    } else {
-      return backoff_matcher_.Value();
-    }
-  }
-
-  virtual TableMatcherImpl<FST> *Copy(bool safe = false) const {
-    assert(0);  // shouldn't be called.  This is not a "real" matcher,
-    // although we derive from MatcherBase for convenience.
-    return NULL;
-  }
-
-  virtual uint64 Properties(uint64 props) const {
-    return props;
-  }  // simple matcher that does
-     // not change its FST, so properties are properties of FST it is applied to
-
- private:
-  virtual void SetState_(StateId s) { SetState(s); }
-  virtual bool Find_(Label label) { return Find(label); }
-  virtual bool Done_() const { return Done(); }
-  virtual const Arc &Value_() const { return Value(); }
-  virtual void Next_() { Next(); }
-
-  MatchType match_type_;
-  FST *fst_;
-  bool current_loop_;
-  Label match_label_;
-  Arc loop_;
-  ArcIterator<FST> *aiter_;
-  StateId s_;
-  std::vector<std::vector<ArcId> *> tables_;
-  TableMatcherOptions opts_;
-  BackoffMatcher backoff_matcher_;
-};
-
-template <class F, class BackoffMatcher = SortedMatcher<F>>
-class TableMatcher : public MatcherBase<typename F::Arc> {
- public:
-  typedef F FST;
-  typedef typename F::Arc Arc;
-  typedef typename Arc::Label Label;
-  typedef typename Arc::StateId StateId;
-  typedef StateId
-      ArcId;  // Use this type to store arc offsets [it's actually size_t
-  // in the Seek function of ArcIterator, but StateId should be big enough].
-  typedef typename Arc::Weight Weight;
-  typedef TableMatcherImpl<F, BackoffMatcher> Impl;
-
-  TableMatcher(const FST &fst, MatchType match_type,
-               const TableMatcherOptions &opts = TableMatcherOptions())
-      : impl_(std::make_shared<Impl>(fst, match_type, opts)) {}
-
-  TableMatcher(const TableMatcher<FST, BackoffMatcher> &matcher,
-               bool safe = false)
-      : impl_(matcher.impl_) {
-    if (safe == true) {
-      LOG(FATAL) << "TableMatcher: Safe copy not supported";
-    }
-  }
-
-  virtual const FST &GetFst() const { return impl_->GetFst(); }
-
-  virtual MatchType Type(bool test) const { return impl_->Type(test); }
-
-  void SetState(StateId s) { return impl_->SetState(s); }
-
-  bool Find(Label match_label) { return impl_->Find(match_label); }
-
-  const Arc &Value() const { return impl_->Value(); }
-
-  void Next() { return impl_->Next(); }
-
-  bool Done() const { return impl_->Done(); }
-
-  const Arc &Value() { return impl_->Value(); }
-
-  virtual TableMatcher<FST, BackoffMatcher> *Copy(bool safe = false) const {
-    return new TableMatcher<FST, BackoffMatcher>(*this, safe);
-  }
-
-  virtual uint64 Properties(uint64 props) const {
-    return impl_->Properties(props);
-  }  // simple matcher that does
-     // not change its FST, so properties are properties of FST it is applied to
- private:
-  std::shared_ptr<Impl> impl_;
-
-  virtual void SetState_(StateId s) { impl_->SetState(s); }
-  virtual bool Find_(Label label) { return impl_->Find(label); }
-  virtual bool Done_() const { return impl_->Done(); }
-  virtual const Arc &Value_() const { return impl_->Value(); }
-  virtual void Next_() { impl_->Next(); }
-
-  TableMatcher &operator=(const TableMatcher &) = delete;
-};
-
-struct TableComposeOptions : public TableMatcherOptions {
-  bool connect;               // Connect output
-  ComposeFilter filter_type;  // Which pre-defined filter to use
-  MatchType table_match_type;
-
-  explicit TableComposeOptions(const TableMatcherOptions &mo, bool c = true,
-                               ComposeFilter ft = SEQUENCE_FILTER,
-                               MatchType tms = MATCH_OUTPUT)
-      : TableMatcherOptions(mo),
-        connect(c),
-        filter_type(ft),
-        table_match_type(tms) {}
-  TableComposeOptions()
-      : connect(true),
-        filter_type(SEQUENCE_FILTER),
-        table_match_type(MATCH_OUTPUT) {}
-};
-
-template <class Arc>
-void TableCompose(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-                  MutableFst<Arc> *ofst,
-                  const TableComposeOptions &opts = TableComposeOptions()) {
-  typedef Fst<Arc> F;
-  CacheOptions nopts;
-  nopts.gc_limit = 0;  // Cache only the last state for fastest copy.
-  if (opts.table_match_type == MATCH_OUTPUT) {
-    // ComposeFstImplOptions templated on matcher for fst1, matcher for fst2.
-    ComposeFstImplOptions<TableMatcher<F>, SortedMatcher<F>> impl_opts(nopts);
-    impl_opts.matcher1 = new TableMatcher<F>(ifst1, MATCH_OUTPUT, opts);
-    *ofst = ComposeFst<Arc>(ifst1, ifst2, impl_opts);
-  } else {
-    assert(opts.table_match_type == MATCH_INPUT);
-    // ComposeFstImplOptions templated on matcher for fst1, matcher for fst2.
-    ComposeFstImplOptions<SortedMatcher<F>, TableMatcher<F>> impl_opts(nopts);
-    impl_opts.matcher2 = new TableMatcher<F>(ifst2, MATCH_INPUT, opts);
-    *ofst = ComposeFst<Arc>(ifst1, ifst2, impl_opts);
-  }
-  if (opts.connect) Connect(ofst);
-}
-
-/// TableComposeCache lets us do multiple compositions while caching the same
-/// matcher.
-template <class F>
-struct TableComposeCache {
-  TableMatcher<F> *matcher;
-  TableComposeOptions opts;
-  explicit TableComposeCache(
-      const TableComposeOptions &opts = TableComposeOptions())
-      : matcher(NULL), opts(opts) {}
-  ~TableComposeCache() { delete (matcher); }
-};
-
-template <class Arc>
-void TableCompose(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-                  MutableFst<Arc> *ofst, TableComposeCache<Fst<Arc>> *cache) {
-  typedef Fst<Arc> F;
-  assert(cache != NULL);
-  CacheOptions nopts;
-  nopts.gc_limit = 0;  // Cache only the last state for fastest copy.
-  if (cache->opts.table_match_type == MATCH_OUTPUT) {
-    ComposeFstImplOptions<TableMatcher<F>, SortedMatcher<F>> impl_opts(nopts);
-    if (cache->matcher == NULL)
-      cache->matcher = new TableMatcher<F>(ifst1, MATCH_OUTPUT, cache->opts);
-    impl_opts.matcher1 = cache->matcher->Copy();  // not passing "safe": may not
-    // be thread-safe-- anway I don't understand this part.
-    *ofst = ComposeFst<Arc>(ifst1, ifst2, impl_opts);
-  } else {
-    assert(cache->opts.table_match_type == MATCH_INPUT);
-    ComposeFstImplOptions<SortedMatcher<F>, TableMatcher<F>> impl_opts(nopts);
-    if (cache->matcher == NULL)
-      cache->matcher = new TableMatcher<F>(ifst2, MATCH_INPUT, cache->opts);
-    impl_opts.matcher2 = cache->matcher->Copy();
-    *ofst = ComposeFst<Arc>(ifst1, ifst2, impl_opts);
-  }
-  if (cache->opts.connect) Connect(ofst);
-}
-
-}  // namespace fst
-
-#endif  // KALDIFST_CSRC_TABLE_MATCHER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/text-normalizer.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/text-normalizer.h
deleted file mode 100644
index f162829b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/text-normalizer.h
+++ /dev/null
@@ -1,36 +0,0 @@
-// kaldifst/csrc/text-normalizer.h
-//
-// Copyright (c)  2023 Xiaomi Corporation
-
-#ifndef KALDIFST_CSRC_TEXT_NORMALIZER_H_
-#define KALDIFST_CSRC_TEXT_NORMALIZER_H_
-
-#include <memory>
-#include <string>
-
-#include "fst/fst.h"
-#include "fst/fstlib.h"
-
-namespace kaldifst {
-
-class TextNormalizer {
- public:
-  // Path to rule.fst
-  explicit TextNormalizer(const std::string &rule);
-  explicit TextNormalizer(std::istream &is);
-
-  explicit TextNormalizer(std::unique_ptr<fst::StdConstFst> rule);
-
-  // @param s The input text to be normalized
-  // @param remove_output_zero True to remove bytes whose value is 0 from the
-  //                           output.
-  std::string Normalize(const std::string &s,
-                        bool remove_output_zero = true) const;
-
- private:
-  std::unique_ptr<fst::StdConstFst> rule_;
-};
-
-}  // namespace kaldifst
-
-#endif  // KALDIFST_CSRC_TEXT_NORMALIZER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/text-utils.h b/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/text-utils.h
deleted file mode 100644
index 212b9552..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/kaldifst/csrc/text-utils.h
+++ /dev/null
@@ -1,127 +0,0 @@
-// kaldifst/csrc/text-utils.h
-//
-// This file is copied/modified from
-// https://github.com/kaldi-asr/kaldi/blob/master/src/util/text-utils.h
-
-// Copyright 2009-2011  Saarland University;  Microsoft Corporation
-
-#include <errno.h>
-
-#include <limits>
-#include <string>
-#include <type_traits>
-#include <vector>
-
-#ifndef KALDIFST_CSRC_TEXT_UTILS_H_
-#define KALDIFST_CSRC_TEXT_UTILS_H_
-
-#ifdef _MSC_VER
-#define KALDIFST_STRTOLL(cur_cstr, end_cstr) _strtoi64(cur_cstr, end_cstr, 10);
-#else
-#define KALDIFST_STRTOLL(cur_cstr, end_cstr) strtoll(cur_cstr, end_cstr, 10);
-#endif
-
-namespace kaldifst {
-
-/// Split a string using any of the single character delimiters.
-/// If omit_empty_strings == true, the output will contain any
-/// nonempty strings after splitting on any of the
-/// characters in the delimiter.  If omit_empty_strings == false,
-/// the output will contain n+1 strings if there are n characters
-/// in the set "delim" within the input string.  In this case
-/// the empty string is split to a single empty string.
-void SplitStringToVector(const std::string &full, const char *delim,
-                         bool omit_empty_strings,
-                         std::vector<std::string> *out);
-
-/// Converts a string into an integer via strtoll and returns false if there was
-/// any kind of problem (i.e. the string was not an integer or contained extra
-/// non-whitespace junk, or the integer was too large to fit into the type it is
-/// being converted into).  Only sets *out if everything was OK and it returns
-/// true.
-template <class Int>
-bool ConvertStringToInteger(const std::string &str, Int *out) {
-  static_assert(std::is_integral<Int>::value, "");
-  const char *this_str = str.c_str();
-  char *end = NULL;
-  errno = 0;
-  int64_t i = KALDIFST_STRTOLL(this_str, &end);
-  if (end != this_str)
-    while (isspace(*end)) end++;
-  if (end == this_str || *end != '\0' || errno != 0) return false;
-  Int iInt = static_cast<Int>(i);
-  if (static_cast<int64_t>(iInt) != i ||
-      (i < 0 && !std::numeric_limits<Int>::is_signed)) {
-    return false;
-  }
-  *out = iInt;
-  return true;
-}
-
-/// Returns true if "token" is nonempty, and all characters are
-/// printable and whitespace-free.
-bool IsToken(const std::string &token);
-
-// CharToString prints the character in a human-readable form, for debugging.
-std::string CharToString(const char &c);
-
-/// Removes leading and trailing white space from the string, then splits on the
-/// first section of whitespace found (if present), putting the part before the
-/// whitespace in "first" and the rest in "rest".  If there is no such space,
-/// everything that remains after removing leading and trailing whitespace goes
-/// in "first".
-void SplitStringOnFirstSpace(const std::string &line, std::string *first,
-                             std::string *rest);
-
-/**
-  \brief Split a string (e.g. 1:2:3) into a vector of integers.
-
-  \param [in]  delim  String containing a list of characters, any of which
-                      is allowed as a delimiter.
-  \param [in] omit_empty_strings If true, empty strings between delimiters are
-                      allowed and will not produce an output integer; if false,
-                      instances of characters in 'delim' that are consecutive or
-                      at the start or end of the string would be an error.
-                      You'll normally want this to be true if 'delim' consists
-                      of spaces, and false otherwise.
-  \param [out] out   The output list of integers.
-*/
-template <class I>
-bool SplitStringToIntegers(const std::string &full, const char *delim,
-                           bool omit_empty_strings,  // typically false [but
-                                                     // should probably be true
-                                                     // if "delim" is spaces].
-                           std::vector<I> *out) {
-  KALDIFST_ASSERT(out != NULL);
-  static_assert(std::is_integral<I>::value, "");
-  if (*(full.c_str()) == '\0') {
-    out->clear();
-    return true;
-  }
-  std::vector<std::string> split;
-  SplitStringToVector(full, delim, omit_empty_strings, &split);
-  out->resize(split.size());
-  for (size_t i = 0; i < split.size(); i++) {
-    const char *this_str = split[i].c_str();
-    char *end = NULL;
-    int64_t j = 0;
-    j = KALDIFST_STRTOLL(this_str, &end);
-    if (end == this_str || *end != '\0') {
-      out->clear();
-      return false;
-    } else {
-      I jI = static_cast<I>(j);
-      if (static_cast<int64_t>(jI) != j) {
-        // output type cannot fit this integer.
-        out->clear();
-        return false;
-      }
-      (*out)[i] = jI;
-    }
-  }
-  return true;
-}
-
-}  // namespace kaldifst
-
-#endif  // KALDIFST_CSRC_TEXT_UTILS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/cpu_provider_factory.h b/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/cpu_provider_factory.h
deleted file mode 100644
index 29267869..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/cpu_provider_factory.h
+++ /dev/null
@@ -1,19 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#include "onnxruntime_c_api.h"
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/**
- * \param use_arena zero: false. non-zero: true.
- */
-ORT_EXPORT
-ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_CPU, _In_ OrtSessionOptions* options, int use_arena)
-ORT_ALL_ARGS_NONNULL;
-
-#ifdef __cplusplus
-}
-#endif
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_c_api.h b/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_c_api.h
deleted file mode 100644
index 64095a31..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_c_api.h
+++ /dev/null
@@ -1,4717 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-// See docs\c_cxx\README.md on generating the Doxygen documentation from this file
-
-/** \mainpage ONNX Runtime
- *
- * ONNX Runtime is a high-performance inference and training graph execution engine for deep learning models.
- *
- * ONNX Runtime's C, C++ APIs offer an easy to use interface to onboard and execute onnx models.
- * - \subpage c_cpp_api "Core C, C++ APIs"
- * - \subpage training_c_cpp_api "Training C, C++ APIs for on-device training"
- *
- * \page c_cpp_api Core C, C++ APIs
- * <h1>C</h1>
- *
- * ::OrtApi - Click here to go to the structure with all C API functions.
- *
- * <h1>C++</h1>
- *
- * ::Ort - Click here to go to the namespace holding all of the C++ wrapper classes
- *
- * It is a set of header only wrapper classes around the C API. The goal is to turn the C style return value error codes into C++ exceptions, and to
- * automate memory management through standard C++ RAII principles.
- *
- * \addtogroup Global
- * ONNX Runtime C API
- * @{
- */
-
-#pragma once
-#include <stdbool.h>
-#include <stdint.h>
-#include <stdlib.h>
-#include <string.h>
-
-/** \brief The API version defined in this header
- *
- * This value is used by some API functions to behave as this version of the header expects.
- */
-#define ORT_API_VERSION 17
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-//! @}
-// SAL2 Definitions
-#ifndef _WIN32
-#define _In_
-#define _In_z_
-#define _In_opt_
-#define _In_opt_z_
-#define _Out_
-#define _Outptr_
-#define _Out_opt_
-#define _Inout_
-#define _Inout_opt_
-#define _Frees_ptr_opt_
-#define _Ret_maybenull_
-#define _Ret_notnull_
-#define _Check_return_
-#define _Outptr_result_maybenull_
-#define _In_reads_(X)
-#define _Inout_updates_(X)
-#define _Out_writes_(X)
-#define _Inout_updates_all_(X)
-#define _Out_writes_bytes_all_(X)
-#define _Out_writes_all_(X)
-#define _Success_(X)
-#define _Outptr_result_buffer_maybenull_(X)
-#define ORT_ALL_ARGS_NONNULL __attribute__((nonnull))
-#else
-#include <specstrings.h>
-#define ORT_ALL_ARGS_NONNULL
-#endif
-
-#ifdef _WIN32
-// Define ORT_DLL_IMPORT if your program is dynamically linked to Ort.
-// dllexport is not used, we use a .def file.
-#ifdef ORT_DLL_IMPORT
-#define ORT_EXPORT __declspec(dllimport)
-#else
-#define ORT_EXPORT
-#endif
-#define ORT_API_CALL _stdcall
-#define ORT_MUST_USE_RESULT
-#define ORTCHAR_T wchar_t
-#else
-// To make symbols visible on macOS/iOS
-#ifdef __APPLE__
-#define ORT_EXPORT __attribute__((visibility("default")))
-#else
-#define ORT_EXPORT
-#endif
-#define ORT_API_CALL
-#define ORT_MUST_USE_RESULT __attribute__((warn_unused_result))
-#define ORTCHAR_T char
-#endif
-
-/// ORTCHAR_T, ORT_TSTR are reserved specifically for path handling.
-/// All other strings are UTF-8 encoded, use char and std::string
-#ifndef ORT_TSTR
-#ifdef _WIN32
-#define ORT_TSTR(X) L##X
-// When X is a macro, L##X is not defined. In this case, we need to use ORT_TSTR_ON_MACRO.
-#define ORT_TSTR_ON_MACRO(X) L"" X
-#else
-#define ORT_TSTR(X) X
-#define ORT_TSTR_ON_MACRO(X) X
-#endif
-#endif
-
-// On Windows, ORT_FILE is a wchar_t version of the __FILE__ macro.
-// Otherwise, ORT_FILE is equivalent to __FILE__.
-#ifndef ORT_FILE
-#define ORT_FILE_INTERNAL(x) ORT_TSTR(x)
-#define ORT_FILE ORT_FILE_INTERNAL(__FILE__)
-#endif
-
-// Any pointer marked with _In_ or _Out_, cannot be NULL.
-
-// Windows users should use unicode paths when possible to bypass the MAX_PATH limitation
-// Every pointer marked with _In_ or _Out_, cannot be NULL. Caller should ensure that.
-// for ReleaseXXX(...) functions, they can accept NULL pointer.
-
-#ifdef __cplusplus
-// For any compiler with C++11 support, MSVC 2015 and greater, or Clang version supporting noexcept.
-// Such complex condition is needed because compilers set __cplusplus value differently.
-#ifndef __has_feature
-#define __has_feature(x) 0
-#endif
-#if ((__cplusplus >= 201103L) || (_MSC_VER >= 1900) || (defined(__has_feature) && __has_feature(cxx_noexcept)))
-#define NO_EXCEPTION noexcept
-#else
-#define NO_EXCEPTION throw()
-#endif
-#else
-#define NO_EXCEPTION
-#endif
-
-// __VA_ARGS__ on Windows and Linux are different
-#define ORT_API(RETURN_TYPE, NAME, ...) RETURN_TYPE ORT_API_CALL NAME(__VA_ARGS__) NO_EXCEPTION
-
-#define ORT_API_STATUS(NAME, ...)                                                                   \
-  _Success_(return == 0) _Check_return_ _Ret_maybenull_ OrtStatusPtr ORT_API_CALL NAME(__VA_ARGS__) \
-  NO_EXCEPTION ORT_MUST_USE_RESULT
-
-// XXX: Unfortunately, SAL annotations are known to not work with function pointers
-#define ORT_API2_STATUS(NAME, ...) \
-  _Check_return_ _Ret_maybenull_ OrtStatusPtr(ORT_API_CALL* NAME)(__VA_ARGS__) NO_EXCEPTION ORT_MUST_USE_RESULT
-
-// Used in *.cc files. Almost as same as ORT_API_STATUS, except without ORT_MUST_USE_RESULT and ORT_EXPORT
-#define ORT_API_STATUS_IMPL(NAME, ...) \
-  _Success_(return == 0) _Check_return_ _Ret_maybenull_ OrtStatusPtr ORT_API_CALL NAME(__VA_ARGS__) NO_EXCEPTION
-
-#define ORT_CLASS_RELEASE(X) void(ORT_API_CALL * Release##X)(_Frees_ptr_opt_ Ort##X * input)
-
-#ifdef __DOXYGEN__
-#undef ORT_API_STATUS
-#define ORT_API_STATUS(NAME, ...) OrtStatus* NAME(__VA_ARGS__)
-#undef ORT_API2_STATUS
-#define ORT_API2_STATUS(NAME, ...) OrtStatus* NAME(__VA_ARGS__)
-#undef ORT_CLASS_RELEASE
-#define ORT_CLASS_RELEASE(X) void Release##X(Ort##X* input)
-#undef NO_EXCEPTION
-#define NO_EXCEPTION
-#endif
-/** \addtogroup Global
- * ONNX Runtime C API
- * @{
- */
-
-/** Copied from TensorProto::DataType
- * Currently, Ort doesn't support complex64, complex128
- */
-typedef enum ONNXTensorElementDataType {
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_UNDEFINED,
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT,   // maps to c type float
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT8,   // maps to c type uint8_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_INT8,    // maps to c type int8_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT16,  // maps to c type uint16_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_INT16,   // maps to c type int16_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32,   // maps to c type int32_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64,   // maps to c type int64_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING,  // maps to c++ type std::string
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_BOOL,
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT16,
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_DOUBLE,      // maps to c type double
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT32,      // maps to c type uint32_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT64,      // maps to c type uint64_t
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_COMPLEX64,   // complex with float32 real and imaginary components
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_COMPLEX128,  // complex with float64 real and imaginary components
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_BFLOAT16,    // Non-IEEE floating-point format based on IEEE754 single-precision
-  // float 8 types were introduced in onnx 1.14, see https://onnx.ai/onnx/technical/float8.html
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E4M3FN,    // Non-IEEE floating-point format based on IEEE754 single-precision
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E4M3FNUZ,  // Non-IEEE floating-point format based on IEEE754 single-precision
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E5M2,      // Non-IEEE floating-point format based on IEEE754 single-precision
-  ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E5M2FNUZ   // Non-IEEE floating-point format based on IEEE754 single-precision
-} ONNXTensorElementDataType;
-
-// Synced with onnx TypeProto oneof
-typedef enum ONNXType {
-  ONNX_TYPE_UNKNOWN,
-  ONNX_TYPE_TENSOR,
-  ONNX_TYPE_SEQUENCE,
-  ONNX_TYPE_MAP,
-  ONNX_TYPE_OPAQUE,
-  ONNX_TYPE_SPARSETENSOR,
-  ONNX_TYPE_OPTIONAL
-} ONNXType;
-
-// These types are synced with internal
-// SparseFormatFlags
-typedef enum OrtSparseFormat {
-  ORT_SPARSE_UNDEFINED = 0,
-  ORT_SPARSE_COO = 0x1,
-  ORT_SPARSE_CSRC = 0x2,
-  ORT_SPARSE_BLOCK_SPARSE = 0x4
-} OrtSparseFormat;
-
-// Enum allows to query sparse tensor indices
-enum OrtSparseIndicesFormat {
-  ORT_SPARSE_COO_INDICES,
-  ORT_SPARSE_CSR_INNER_INDICES,
-  ORT_SPARSE_CSR_OUTER_INDICES,
-  ORT_SPARSE_BLOCK_SPARSE_INDICES
-};
-
-/** \brief Logging severity levels
- *
- * In typical API usage, specifying a logging severity level specifies the minimum severity of log messages to show.
- */
-typedef enum OrtLoggingLevel {
-  ORT_LOGGING_LEVEL_VERBOSE,  ///< Verbose informational messages (least severe).
-  ORT_LOGGING_LEVEL_INFO,     ///< Informational messages.
-  ORT_LOGGING_LEVEL_WARNING,  ///< Warning messages.
-  ORT_LOGGING_LEVEL_ERROR,    ///< Error messages.
-  ORT_LOGGING_LEVEL_FATAL,    ///< Fatal error messages (most severe).
-} OrtLoggingLevel;
-
-typedef enum OrtErrorCode {
-  ORT_OK,
-  ORT_FAIL,
-  ORT_INVALID_ARGUMENT,
-  ORT_NO_SUCHFILE,
-  ORT_NO_MODEL,
-  ORT_ENGINE_ERROR,
-  ORT_RUNTIME_EXCEPTION,
-  ORT_INVALID_PROTOBUF,
-  ORT_MODEL_LOADED,
-  ORT_NOT_IMPLEMENTED,
-  ORT_INVALID_GRAPH,
-  ORT_EP_FAIL,
-} OrtErrorCode;
-
-typedef enum OrtOpAttrType {
-  ORT_OP_ATTR_UNDEFINED = 0,
-  ORT_OP_ATTR_INT,
-  ORT_OP_ATTR_INTS,
-  ORT_OP_ATTR_FLOAT,
-  ORT_OP_ATTR_FLOATS,
-  ORT_OP_ATTR_STRING,
-  ORT_OP_ATTR_STRINGS,
-} OrtOpAttrType;
-
-//! @}
-#define ORT_RUNTIME_CLASS(X) \
-  struct Ort##X;             \
-  typedef struct Ort##X Ort##X;
-
-/** \addtogroup Global
- * ONNX Runtime C API
- * @{
- */
-// The actual types defined have an Ort prefix
-ORT_RUNTIME_CLASS(Env);
-ORT_RUNTIME_CLASS(Status);  // nullptr for Status* indicates success
-ORT_RUNTIME_CLASS(MemoryInfo);
-ORT_RUNTIME_CLASS(IoBinding);
-ORT_RUNTIME_CLASS(Session);  // Don't call ReleaseSession from Dllmain (because session owns a thread pool)
-ORT_RUNTIME_CLASS(Value);
-ORT_RUNTIME_CLASS(RunOptions);
-ORT_RUNTIME_CLASS(TypeInfo);
-ORT_RUNTIME_CLASS(TensorTypeAndShapeInfo);
-ORT_RUNTIME_CLASS(MapTypeInfo);
-ORT_RUNTIME_CLASS(SequenceTypeInfo);
-ORT_RUNTIME_CLASS(OptionalTypeInfo);
-ORT_RUNTIME_CLASS(SessionOptions);
-ORT_RUNTIME_CLASS(CustomOpDomain);
-ORT_RUNTIME_CLASS(ModelMetadata);
-ORT_RUNTIME_CLASS(ThreadPoolParams);
-ORT_RUNTIME_CLASS(ThreadingOptions);
-ORT_RUNTIME_CLASS(ArenaCfg);
-ORT_RUNTIME_CLASS(PrepackedWeightsContainer);
-ORT_RUNTIME_CLASS(TensorRTProviderOptionsV2);
-ORT_RUNTIME_CLASS(CUDAProviderOptionsV2);
-ORT_RUNTIME_CLASS(CANNProviderOptions);
-ORT_RUNTIME_CLASS(DnnlProviderOptions);
-ORT_RUNTIME_CLASS(Op);
-ORT_RUNTIME_CLASS(OpAttr);
-ORT_RUNTIME_CLASS(Logger);
-ORT_RUNTIME_CLASS(ShapeInferContext);
-
-#ifdef _WIN32
-typedef _Return_type_success_(return == 0) OrtStatus* OrtStatusPtr;
-#else
-typedef OrtStatus* OrtStatusPtr;
-#endif
-
-/** \brief Memory allocation interface
- *
- * Structure of function pointers that defines a memory allocator. This can be created and filled in by the user for custom allocators.
- *
- * When an allocator is passed to any function, be sure that the allocator object is not destroyed until the last allocated object using it is freed.
- */
-typedef struct OrtAllocator {
-  uint32_t version;                                                                   ///< Must be initialized to ORT_API_VERSION
-  void*(ORT_API_CALL* Alloc)(struct OrtAllocator* this_, size_t size);                ///< Returns a pointer to an allocated block of `size` bytes
-  void(ORT_API_CALL* Free)(struct OrtAllocator* this_, void* p);                      ///< Free a block of memory previously allocated with OrtAllocator::Alloc
-  const struct OrtMemoryInfo*(ORT_API_CALL* Info)(const struct OrtAllocator* this_);  ///< Return a pointer to an ::OrtMemoryInfo that describes this allocator
-} OrtAllocator;
-
-typedef void(ORT_API_CALL* OrtLoggingFunction)(
-    void* param, OrtLoggingLevel severity, const char* category, const char* logid, const char* code_location,
-    const char* message);
-
-/** \brief Graph optimization level
- *
- * Refer to https://www.onnxruntime.ai/docs/performance/graph-optimizations.html#graph-optimization-levels
- * for an in-depth understanding of the Graph Optimization Levels.
- */
-typedef enum GraphOptimizationLevel {
-  ORT_DISABLE_ALL = 0,
-  ORT_ENABLE_BASIC = 1,
-  ORT_ENABLE_EXTENDED = 2,
-  ORT_ENABLE_ALL = 99
-} GraphOptimizationLevel;
-
-typedef enum ExecutionMode {
-  ORT_SEQUENTIAL = 0,
-  ORT_PARALLEL = 1,
-} ExecutionMode;
-
-/** \brief Language projection identifiers
- * /see OrtApi::SetLanguageProjection
- */
-typedef enum OrtLanguageProjection {
-  ORT_PROJECTION_C = 0,
-  ORT_PROJECTION_CPLUSPLUS = 1,
-  ORT_PROJECTION_CSHARP = 2,
-  ORT_PROJECTION_PYTHON = 3,
-  ORT_PROJECTION_JAVA = 4,
-  ORT_PROJECTION_WINML = 5,
-  ORT_PROJECTION_NODEJS = 6,
-} OrtLanguageProjection;
-
-struct OrtKernelInfo;
-typedef struct OrtKernelInfo OrtKernelInfo;
-struct OrtKernelContext;
-typedef struct OrtKernelContext OrtKernelContext;
-struct OrtCustomOp;
-typedef struct OrtCustomOp OrtCustomOp;
-
-typedef enum OrtAllocatorType {
-  OrtInvalidAllocator = -1,
-  OrtDeviceAllocator = 0,
-  OrtArenaAllocator = 1
-} OrtAllocatorType;
-
-/** \brief Memory types for allocated memory, execution provider specific types should be extended in each provider.
- */
-// Whenever this struct is updated, please also update the MakeKey function in onnxruntime / core / framework / execution_provider.cc
-typedef enum OrtMemType {
-  OrtMemTypeCPUInput = -2,              ///< Any CPU memory used by non-CPU execution provider
-  OrtMemTypeCPUOutput = -1,             ///< CPU accessible memory outputted by non-CPU execution provider, i.e. CUDA_PINNED
-  OrtMemTypeCPU = OrtMemTypeCPUOutput,  ///< Temporary CPU accessible memory allocated by non-CPU execution provider, i.e. CUDA_PINNED
-  OrtMemTypeDefault = 0,                ///< The default allocator for execution provider
-} OrtMemType;
-
-/** \brief This mimics OrtDevice type constants so they can be returned in the API
- */
-typedef enum OrtMemoryInfoDeviceType {
-  OrtMemoryInfoDeviceType_CPU = 0,
-  OrtMemoryInfoDeviceType_GPU = 1,
-  OrtMemoryInfoDeviceType_FPGA = 2
-} OrtMemoryInfoDeviceType;
-
-/** \brief Algorithm to use for cuDNN Convolution Op
- */
-typedef enum OrtCudnnConvAlgoSearch {
-  OrtCudnnConvAlgoSearchExhaustive,  // expensive exhaustive benchmarking using cudnnFindConvolutionForwardAlgorithmEx
-  OrtCudnnConvAlgoSearchHeuristic,   // lightweight heuristic based search using cudnnGetConvolutionForwardAlgorithm_v7
-  OrtCudnnConvAlgoSearchDefault,     // default algorithm using CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM
-} OrtCudnnConvAlgoSearch;
-
-/** \brief CUDA Provider Options
- *
- * \see OrtApi::SessionOptionsAppendExecutionProvider_CUDA
- */
-typedef struct OrtCUDAProviderOptions {
-#ifdef __cplusplus
-  OrtCUDAProviderOptions()
-      : device_id{},
-        cudnn_conv_algo_search{OrtCudnnConvAlgoSearchExhaustive},
-        gpu_mem_limit{SIZE_MAX},
-        arena_extend_strategy{},
-        do_copy_in_default_stream{1},
-        has_user_compute_stream{},
-        user_compute_stream{},
-        default_memory_arena_cfg{},
-        tunable_op_enable{false},
-        tunable_op_tuning_enable{false},
-        tunable_op_max_tuning_duration_ms{} {}
-#endif
-
-  /** \brief CUDA device Id
-   *   Defaults to 0.
-   */
-  int device_id;
-
-  /** \brief CUDA Convolution algorithm search configuration.
-   *   See enum OrtCudnnConvAlgoSearch for more details.
-   *   Defaults to OrtCudnnConvAlgoSearchExhaustive.
-   */
-  OrtCudnnConvAlgoSearch cudnn_conv_algo_search;
-
-  /** \brief CUDA memory limit (To use all possible memory pass in maximum size_t)
-   *   Defaults to SIZE_MAX.
-   *   \note If a ::OrtArenaCfg has been applied, it will override this field
-   */
-  size_t gpu_mem_limit;
-
-  /** \brief Strategy used to grow the memory arena
-   *   0 = kNextPowerOfTwo<br>
-   *   1 = kSameAsRequested<br>
-   *   Defaults to 0.
-   *   \note If a ::OrtArenaCfg has been applied, it will override this field
-   */
-  int arena_extend_strategy;
-
-  /** \brief Flag indicating if copying needs to take place on the same stream as the compute stream in the CUDA EP
-   *   0 = Use separate streams for copying and compute.
-   *   1 = Use the same stream for copying and compute.
-   *   Defaults to 1.
-   *   WARNING: Setting this to 0 may result in data races for some models.
-   *   Please see issue #4829 for more details.
-   */
-  int do_copy_in_default_stream;
-
-  /** \brief Flag indicating if there is a user provided compute stream
-   *   Defaults to 0.
-   */
-  int has_user_compute_stream;
-
-  /** \brief User provided compute stream.
-   *   If provided, please set `has_user_compute_stream` to 1.
-   */
-  void* user_compute_stream;
-
-  /** \brief CUDA memory arena configuration parameters
-   */
-  OrtArenaCfg* default_memory_arena_cfg;
-
-  /** \brief Enable TunableOp for using.
-   *   Set it to 1/0 to enable/disable TunableOp. Otherwise, it is disabled by default.
-   *   This option can be overriden by environment variable ORT_CUDA_TUNABLE_OP_ENABLE.
-   */
-  int tunable_op_enable;
-
-  /** \brief Enable TunableOp for tuning.
-   *   Set it to 1/0 to enable/disable TunableOp tuning. Otherwise, it is disabled by default.
-   *   This option can be overriden by environment variable ORT_CUDA_TUNABLE_OP_TUNING_ENABLE.
-   */
-  int tunable_op_tuning_enable;
-
-  /** \brief Max tuning duration time limit for each instance of TunableOp.
-   *   Defaults to 0 to disable the limit.
-   */
-  int tunable_op_max_tuning_duration_ms;
-
-} OrtCUDAProviderOptions;
-
-/** \brief ROCM Provider Options
- *
- * \see OrtApi::SessionOptionsAppendExecutionProvider_ROCM
- */
-typedef struct OrtROCMProviderOptions {
-#ifdef __cplusplus
-  OrtROCMProviderOptions()
-      : device_id{},
-        miopen_conv_exhaustive_search{0},
-        gpu_mem_limit{SIZE_MAX},
-        arena_extend_strategy{},
-        do_copy_in_default_stream{1},
-        has_user_compute_stream{},
-        user_compute_stream{},
-        default_memory_arena_cfg{},
-        tunable_op_enable{false},
-        tunable_op_tuning_enable{false},
-        tunable_op_max_tuning_duration_ms{} {}
-#endif
-
-  /** \brief ROCM device Id
-   *   Defaults to 0.
-   */
-  int device_id;
-
-  /** \brief ROCM MIOpen Convolution algorithm exaustive search option.
-   *   Defaults to 0 (false).
-   */
-  int miopen_conv_exhaustive_search;
-
-  /** \brief ROCM memory limit (To use all possible memory pass in maximum size_t)
-   *   Defaults to SIZE_MAX.
-   *   \note If a ::OrtArenaCfg has been applied, it will override this field
-   */
-  size_t gpu_mem_limit;
-
-  /** \brief Strategy used to grow the memory arena
-   *   0 = kNextPowerOfTwo<br>
-   *   1 = kSameAsRequested<br>
-   *   Defaults to 0.
-   *   \note If a ::OrtArenaCfg has been applied, it will override this field
-   */
-  int arena_extend_strategy;
-
-  /** \brief Flag indicating if copying needs to take place on the same stream as the compute stream in the ROCM EP
-   *   0 = Use separate streams for copying and compute.
-   *   1 = Use the same stream for copying and compute.
-   *   Defaults to 1.
-   *   WARNING: Setting this to 0 may result in data races for some models.
-   *   Please see issue #4829 for more details.
-   */
-  int do_copy_in_default_stream;
-
-  /** \brief Flag indicating if there is a user provided compute stream
-   *   Defaults to 0.
-   */
-  int has_user_compute_stream;
-
-  /** \brief User provided compute stream.
-   *   If provided, please set `has_user_compute_stream` to 1.
-   */
-  void* user_compute_stream;
-
-  /** \brief ROCM memory arena configuration parameters
-   */
-  OrtArenaCfg* default_memory_arena_cfg;
-
-  /** \brief Enable TunableOp for using.
-   *   Set it to 1/0 to enable/disable TunableOp. Otherwise, it is disabled by default.
-   *   This option can be overriden by environment variable ORT_ROCM_TUNABLE_OP_ENABLE.
-   */
-  int tunable_op_enable;
-
-  /** \brief Enable TunableOp for tuning.
-   *   Set it to 1/0 to enable/disable TunableOp tuning. Otherwise, it is disabled by default.
-   *   This option can be overriden by environment variable ORT_ROCM_TUNABLE_OP_TUNING_ENABLE.
-   */
-  int tunable_op_tuning_enable;
-
-  /** \brief Max tuning duration time limit for each instance of TunableOp.
-   *   Defaults to 0 to disable the limit.
-   */
-  int tunable_op_max_tuning_duration_ms;
-
-} OrtROCMProviderOptions;
-
-/** \brief TensorRT Provider Options
- *
- * \see OrtApi::SessionOptionsAppendExecutionProvider_TensorRT
- */
-typedef struct OrtTensorRTProviderOptions {
-  int device_id;                                ///< CUDA device id (0 = default device)
-  int has_user_compute_stream;                  // indicator of user specified CUDA compute stream.
-  void* user_compute_stream;                    // user specified CUDA compute stream.
-  int trt_max_partition_iterations;             // maximum iterations for TensorRT parser to get capability
-  int trt_min_subgraph_size;                    // minimum size of TensorRT subgraphs
-  size_t trt_max_workspace_size;                // maximum workspace size for TensorRT.
-  int trt_fp16_enable;                          // enable TensorRT FP16 precision. Default 0 = false, nonzero = true
-  int trt_int8_enable;                          // enable TensorRT INT8 precision. Default 0 = false, nonzero = true
-  const char* trt_int8_calibration_table_name;  // TensorRT INT8 calibration table name.
-  int trt_int8_use_native_calibration_table;    // use native TensorRT generated calibration table. Default 0 = false, nonzero = true
-  int trt_dla_enable;                           // enable DLA. Default 0 = false, nonzero = true
-  int trt_dla_core;                             // DLA core number. Default 0
-  int trt_dump_subgraphs;                       // dump TRT subgraph. Default 0 = false, nonzero = true
-  int trt_engine_cache_enable;                  // enable engine caching. Default 0 = false, nonzero = true
-  const char* trt_engine_cache_path;            // specify engine cache path
-  int trt_engine_decryption_enable;             // enable engine decryption. Default 0 = false, nonzero = true
-  const char* trt_engine_decryption_lib_path;   // specify engine decryption library path
-  int trt_force_sequential_engine_build;        // force building TensorRT engine sequentially. Default 0 = false, nonzero = true
-  // This is the legacy struct and don't add new fields here.
-  // For new field that can be represented by string, please add it in include/onnxruntime/core/providers/tensorrt/tensorrt_provider_options.h
-  // For non-string field, need to create a new separate api to handle it.
-} OrtTensorRTProviderOptions;
-
-/** \brief MIGraphX Provider Options
- *
- * \see OrtApi::SessionOptionsAppendExecutionProvider_MIGraphX
- */
-typedef struct OrtMIGraphXProviderOptions {
-  int device_id;                                     // hip device id.
-  int migraphx_fp16_enable;                          // MIGraphX FP16 precision. Default 0 = false, nonzero = true
-  int migraphx_int8_enable;                          // MIGraphX INT8 precision. Default 0 = false, nonzero = true
-  int migraphx_use_native_calibration_table;         // MIGraphx INT8 cal table. Default 0 = false, noznero = true
-  const char* migraphx_int8_calibration_table_name;  // MIGraphx INT8 calibration table name
-} OrtMIGraphXProviderOptions;
-
-/** \brief OpenVINO Provider Options
- *
- * \see OrtApi::SessionOptionsAppendExecutionProvider_OpenVINO
- */
-typedef struct OrtOpenVINOProviderOptions {
-#ifdef __cplusplus
-  OrtOpenVINOProviderOptions() : device_type{},
-                                 enable_npu_fast_compile{},
-                                 device_id{},
-                                 num_of_threads{},
-                                 cache_dir{},
-                                 context{},
-                                 enable_opencl_throttling{},
-                                 enable_dynamic_shapes{} {}
-#endif
-  /** \brief Device type string
-   *
-   * Valid settings are one of: "CPU_FP32", "CPU_FP16", "GPU_FP32", "GPU_FP16"
-   */
-  const char* device_type;
-  unsigned char enable_npu_fast_compile;  ///< 0 = disabled, nonzero = enabled
-  const char* device_id;
-  size_t num_of_threads;  ///< 0 = Use default number of threads
-  const char* cache_dir;  // path is set to empty by default
-  void* context;
-  unsigned char enable_opencl_throttling;  ///< 0 = disabled, nonzero = enabled
-  unsigned char enable_dynamic_shapes;     ///< 0 = disabled, nonzero = enabled
-} OrtOpenVINOProviderOptions;
-
-struct OrtApi;
-typedef struct OrtApi OrtApi;
-
-struct OrtTrainingApi;
-typedef struct OrtTrainingApi OrtTrainingApi;
-
-/** \brief The helper interface to get the right version of OrtApi
- *
- * Get a pointer to this structure through ::OrtGetApiBase
- */
-struct OrtApiBase {
-  /** \brief Get a pointer to the requested version of the ::OrtApi
-   *
-   * \param[in] version Must be ::ORT_API_VERSION
-   * \return The ::OrtApi for the version requested, nullptr will be returned if this version is unsupported, for example when using a runtime
-   *   older than the version created with this header file.
-   *
-   * One can call GetVersionString() to get the version of the Onnxruntime library for logging
-   * and error reporting purposes.
-   */
-  const OrtApi*(ORT_API_CALL* GetApi)(uint32_t version)NO_EXCEPTION;
-
-  /** \brief Returns a null terminated string of the version of the Onnxruntime library (eg: "1.8.1")
-   *
-   *  \return UTF-8 encoded version string. Do not deallocate the returned buffer.
-   */
-  const char*(ORT_API_CALL* GetVersionString)(void)NO_EXCEPTION;
-};
-
-typedef struct OrtApiBase OrtApiBase;
-
-/** \brief The Onnxruntime library's entry point to access the C API
- *
- * Call this to get the a pointer to an ::OrtApiBase
- */
-ORT_EXPORT const OrtApiBase* ORT_API_CALL OrtGetApiBase(void) NO_EXCEPTION;
-
-/** \brief Thread work loop function
- *
- * Onnxruntime will provide the working loop on custom thread creation
- * Argument is an onnxruntime built-in type which will be provided when thread pool calls OrtCustomCreateThreadFn
- */
-typedef void (*OrtThreadWorkerFn)(void* ort_worker_fn_param);
-
-typedef const struct OrtCustomHandleType {
-  char __place_holder;
-}* OrtCustomThreadHandle;
-
-/** \brief Ort custom thread creation function
- *
- * The function should return a thread handle to be used in onnxruntime thread pools
- * Onnxruntime will throw exception on return value of nullptr or 0, indicating that the function failed to create a thread
- */
-typedef OrtCustomThreadHandle (*OrtCustomCreateThreadFn)(void* ort_custom_thread_creation_options, OrtThreadWorkerFn ort_thread_worker_fn, void* ort_worker_fn_param);
-
-/** \brief Custom thread join function
- *
- * Onnxruntime thread pool destructor will call the function to join a custom thread.
- * Argument ort_custom_thread_handle is the value returned by OrtCustomCreateThreadFn
- */
-typedef void (*OrtCustomJoinThreadFn)(OrtCustomThreadHandle ort_custom_thread_handle);
-
-typedef OrtStatus*(ORT_API_CALL* RegisterCustomOpsFn)(OrtSessionOptions* options, const OrtApiBase* api);
-
-/** \brief Callback function for RunAsync
- *
- * \param[in] user_data User specific data that passed back to the callback
- * \param[out] outputs On succeed, outputs host inference results, on error, the value will be nullptr
- * \param[out] num_outputs Number of outputs, on error, the value will be zero
- * \param[out] status On error, status will provide details
- */
-typedef void (*RunAsyncCallbackFn)(void* user_data, OrtValue** outputs, size_t num_outputs, OrtStatusPtr status);
-
-/** \brief The C API
- *
- * All C API functions are defined inside this structure as pointers to functions.
- * Call OrtApiBase::GetApi to get a pointer to it
- *
- * \nosubgrouping
- */
-struct OrtApi {
-  /// \name OrtStatus
-  /// @{
-
-  /**
-   * \brief Create an OrtStatus from a null terminated string
-   *
-   * \param[in] code
-   * \param[in] msg A null-terminated string. Its contents will be copied.
-   * \return A new OrtStatus object, must be destroyed with OrtApi::ReleaseStatus
-   */
-  OrtStatus*(ORT_API_CALL* CreateStatus)(OrtErrorCode code, _In_ const char* msg)NO_EXCEPTION ORT_ALL_ARGS_NONNULL;
-
-  /** \brief Get OrtErrorCode from OrtStatus
-   *
-   * \param[in] status
-   * \return OrtErrorCode that \p status was created with
-   */
-  OrtErrorCode(ORT_API_CALL* GetErrorCode)(_In_ const OrtStatus* status) NO_EXCEPTION ORT_ALL_ARGS_NONNULL;
-
-  /** \brief Get error string from OrtStatus
-   *
-   * \param[in] status
-   * \return The error message inside the `status`. Do not free the returned value.
-   */
-  const char*(ORT_API_CALL* GetErrorMessage)(_In_ const OrtStatus* status)NO_EXCEPTION ORT_ALL_ARGS_NONNULL;
-
-  /// @}
-  /// \name OrtEnv
-  /// @{
-
-  /** \brief Create an OrtEnv
-   *
-   * \note Invoking this function will return the same instance of the environment as that returned by a previous call
-   * to another env creation function; all arguments to this function will be ignored.
-   * \param[in] log_severity_level The log severity level.
-   * \param[in] logid The log identifier.
-   * \param[out] out Returned newly created OrtEnv. Must be freed with OrtApi::ReleaseEnv
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateEnv, OrtLoggingLevel log_severity_level, _In_ const char* logid, _Outptr_ OrtEnv** out);
-
-  /** \brief Create an OrtEnv
-   *
-   * \note Invoking this function will return the same instance of the environment as that returned by a previous call
-   * to another env creation function; all arguments to this function will be ignored. If you want to provide your
-   * own logging function, consider setting it using the SetUserLoggingFunction API instead.
-   * \param[in] logging_function A pointer to a logging function.
-   * \param[in] logger_param A pointer to arbitrary data passed as the ::OrtLoggingFunction `param` parameter to
-   *                         `logging_function`. This parameter is optional.
-   * \param[in] log_severity_level The log severity level.
-   * \param[in] logid The log identifier.
-   * \param[out] out Returned newly created OrtEnv. Must be freed with OrtApi::ReleaseEnv
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateEnvWithCustomLogger, _In_ OrtLoggingFunction logging_function, _In_opt_ void* logger_param,
-                  _In_ OrtLoggingLevel log_severity_level, _In_ const char* logid, _Outptr_ OrtEnv** out);
-
-  /** \brief Enable Telemetry
-   *
-   * \note Telemetry events are on by default since they are lightweight
-   * \param[in] env
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(EnableTelemetryEvents, _In_ const OrtEnv* env);
-  /** \brief Disable Telemetry
-   *
-   * \see OrtApi::EnableTelemetryEvents
-   * \param[in] env
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(DisableTelemetryEvents, _In_ const OrtEnv* env);
-
-  /// @}
-  /// \name OrtSession
-  /// @{
-
-  /** \brief Create an OrtSession from a model file
-   *
-   * \param[in] env
-   * \param[in] model_path
-   * \param[in] options
-   * \param[out] out Returned newly created OrtSession. Must be freed with OrtApi::ReleaseSession
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  // TODO: document the path separator convention? '/' vs '\'
-  // TODO: should specify the access characteristics of model_path. Is this read only during the
-  // execution of CreateSession, or does the OrtSession retain a handle to the file/directory
-  // and continue to access throughout the OrtSession lifetime?
-  //  What sort of access is needed to model_path : read or read/write?
-  ORT_API2_STATUS(CreateSession, _In_ const OrtEnv* env, _In_ const ORTCHAR_T* model_path,
-                  _In_ const OrtSessionOptions* options, _Outptr_ OrtSession** out);
-
-  /** \brief Create an OrtSession from memory
-   *
-   * \param[in] env
-   * \param[in] model_data
-   * \param[in] model_data_length
-   * \param[in] options
-   * \param[out] out Returned newly created OrtSession. Must be freed with OrtApi::ReleaseSession
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateSessionFromArray, _In_ const OrtEnv* env, _In_ const void* model_data, size_t model_data_length,
-                  _In_ const OrtSessionOptions* options, _Outptr_ OrtSession** out);
-
-  /** \brief Run the model in an ::OrtSession
-   *
-   * Will not return until the model run has completed. Multiple threads might be used to run the model based on
-   * the options in the ::OrtSession and settings used when creating the ::OrtEnv
-   *
-   * \param[in] session
-   * \param[in] run_options If nullptr, will use a default ::OrtRunOptions
-   * \param[in] input_names Array of null terminated UTF8 encoded strings of the input names
-   * \param[in] inputs Array of ::OrtValue%s of the input values
-   * \param[in] input_len Number of elements in the input_names and inputs arrays
-   * \param[in] output_names Array of null terminated UTF8 encoded strings of the output names
-   * \param[in] output_names_len Number of elements in the output_names and outputs array
-   * \param[out] outputs Array of ::OrtValue%s that the outputs are stored in. This can also be
-   *     an array of nullptr values, in this case ::OrtValue objects will be allocated and pointers
-   *     to them will be set into the `outputs` array.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(Run, _Inout_ OrtSession* session, _In_opt_ const OrtRunOptions* run_options,
-                  _In_reads_(input_len) const char* const* input_names,
-                  _In_reads_(input_len) const OrtValue* const* inputs, size_t input_len,
-                  _In_reads_(output_names_len) const char* const* output_names, size_t output_names_len,
-                  _Inout_updates_all_(output_names_len) OrtValue** outputs);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Create an ::OrtSessionOptions object
-   *
-   * To use additional providers, you must build ORT with the extra providers enabled. Then call one of these
-   * functions to enable them in the session:<br>
-   *   OrtSessionOptionsAppendExecutionProvider_CPU<br>
-   *   OrtSessionOptionsAppendExecutionProvider_CUDA<br>
-   *   OrtSessionOptionsAppendExecutionProvider_(remaining providers...)<br>
-   * The order they are called indicates the preference order as well. In other words call this method
-   * on your most preferred execution provider first followed by the less preferred ones.
-   * If none are called Ort will use its internal CPU execution provider.
-   *
-   * \param[out] options The newly created OrtSessionOptions. Must be freed with OrtApi::ReleaseSessionOptions
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateSessionOptions, _Outptr_ OrtSessionOptions** options);
-
-  /** \brief Set filepath to save optimized model after graph level transformations
-   *
-   * \param[in] options
-   * \param[in] optimized_model_filepath
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetOptimizedModelFilePath, _Inout_ OrtSessionOptions* options,
-                  _In_ const ORTCHAR_T* optimized_model_filepath);
-
-  /** \brief Create a copy of an existing ::OrtSessionOptions
-   *
-   * \param[in] in_options OrtSessionOptions to copy
-   * \param[out] out_options Returned newly created ::OrtSessionOptions. Must be freed with OrtApi::ReleaseSessionOptions
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CloneSessionOptions, _In_ const OrtSessionOptions* in_options,
-                  _Outptr_ OrtSessionOptions** out_options);
-
-  /** \brief Set execution mode
-   *
-   * Controls whether you want to execute operators in your graph sequentially or in parallel. Usually when the model
-   *  has many branches, setting this option to ExecutionMode.ORT_PARALLEL will give you better performance.
-   *  See [docs/ONNX_Runtime_Perf_Tuning.md] for more details.
-   *
-   * \param[in] options
-   * \param[in] execution_mode
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetSessionExecutionMode, _Inout_ OrtSessionOptions* options, ExecutionMode execution_mode);
-
-  /** \brief Enable profiling for a session
-   *
-   * \param[in] options
-   * \param[in] profile_file_prefix
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(EnableProfiling, _Inout_ OrtSessionOptions* options, _In_ const ORTCHAR_T* profile_file_prefix);
-
-  /** \brief Disable profiling for a session
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(DisableProfiling, _Inout_ OrtSessionOptions* options);
-
-  /** \brief Enable the memory pattern optimization
-   *
-   * The idea is if the input shapes are the same, we could trace the internal memory allocation
-   * and generate a memory pattern for future request. So next time we could just do one allocation
-   * with a big chunk for all the internal memory allocation.
-   * \note Memory pattern optimization is only available when Sequential Execution mode is enabled (see OrtApi::SetSessionExecutionMode)
-   *
-   * \see OrtApi::DisableMemPattern
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(EnableMemPattern, _Inout_ OrtSessionOptions* options);
-
-  /** \brief Disable the memory pattern optimization
-   *
-   * \see OrtApi::EnableMemPattern
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(DisableMemPattern, _Inout_ OrtSessionOptions* options);
-
-  /** \brief Enable the memory arena on CPU
-   *
-   * Arena may pre-allocate memory for future usage.
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(EnableCpuMemArena, _Inout_ OrtSessionOptions* options);
-
-  /** \brief Disable the memory arena on CPU
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(DisableCpuMemArena, _Inout_ OrtSessionOptions* options);
-
-  /** \brief Set session log id
-   *
-   * \param[in] options
-   * \param[in] logid The log identifier.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetSessionLogId, _Inout_ OrtSessionOptions* options, const char* logid);
-
-  /** \brief Set session log verbosity level
-   *
-   * Applies to session load, initialization, etc
-   *
-   * \param[in] options
-   * \param[in] session_log_verbosity_level \snippet{doc} snippets.dox Log Verbosity Level
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetSessionLogVerbosityLevel, _Inout_ OrtSessionOptions* options, int session_log_verbosity_level);
-
-  /** \brief Set session log severity level
-   *
-   * \param[in] options
-   * \param[in] session_log_severity_level The log severity level (refer to ::OrtLoggingLevel for possible values).
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetSessionLogSeverityLevel, _Inout_ OrtSessionOptions* options, int session_log_severity_level);
-
-  /** \brief Set the optimization level to apply when loading a graph
-   *
-   * Please see https://onnxruntime.ai/docs/performance/model-optimizations/graph-optimizations.html for an in-depth explanation
-   * \param[in,out] options The session options object
-   * \param[in] graph_optimization_level The optimization level
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetSessionGraphOptimizationLevel, _Inout_ OrtSessionOptions* options,
-                  GraphOptimizationLevel graph_optimization_level);
-
-  /** \brief Sets the number of threads used to parallelize the execution within nodes
-   *
-   * When running a single node operation, ex. add, this sets the maximum number of threads to use.
-   *
-   * \note If built with OpenMP, this has no effect on the number of threads used. In this case
-   *       use the OpenMP env variables to configure the number of intra op num threads.
-   *
-   * \param[in] options
-   * \param[in] intra_op_num_threads Number of threads to use<br>
-   *   A value of 0 will use the default number of threads<br>
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetIntraOpNumThreads, _Inout_ OrtSessionOptions* options, int intra_op_num_threads);
-
-  /** \brief Sets the number of threads used to parallelize the execution of the graph
-   *
-   * If nodes can be run in parallel, this sets the maximum number of threads to use to run them in parallel.
-   *
-   * \note If sequential execution is enabled this value is ignored, it acts as if it was set to 1.
-   *
-   * \param[in] options
-   * \param[in] inter_op_num_threads Number of threads to use<br>
-   *   A value of 0 will use the default number of threads<br>
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetInterOpNumThreads, _Inout_ OrtSessionOptions* options, int inter_op_num_threads);
-
-  /// @}
-  /// \name OrtCustomOpDomain
-  /// @{
-
-  /** \brief Create a custom op domain
-   *
-   * \param[in] domain
-   * \param[out] out Newly created domain. Must be freed with OrtApi::ReleaseCustomOpDomain
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateCustomOpDomain, _In_ const char* domain, _Outptr_ OrtCustomOpDomain** out);
-
-  /** \brief Add a custom op to a custom op domain
-   *
-   * \note The OrtCustomOp* pointer must remain valid until the ::OrtCustomOpDomain using it is released
-   *
-   * \param[in] custom_op_domain
-   * \param[in] op
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CustomOpDomain_Add, _Inout_ OrtCustomOpDomain* custom_op_domain, _In_ const OrtCustomOp* op);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Add custom op domain to a session options
-   *
-   * \note The OrtCustomOpDomain* must not be deleted until all sessions using it are released
-   *
-   * \param[in] options
-   * \param[in] custom_op_domain
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(AddCustomOpDomain, _Inout_ OrtSessionOptions* options, _In_ OrtCustomOpDomain* custom_op_domain);
-
-  /** \deprecated Use OrtApi::RegisterCustomOpsLibrary_V2.
-   *
-   * Registers custom ops from a shared library.
-   *
-   * Loads a shared library (dll on windows, so on linux, etc) named 'library_path' and looks for this entry point:
-   *		OrtStatus* RegisterCustomOps(OrtSessionOptions * options, const OrtApiBase* api);
-   * It then passes in the provided session options to this function along with the api base.
-   * The handle to the loaded library is returned in library_handle. It can be freed by the caller after all sessions using the passed in
-   * session options are destroyed, or if an error occurs and it is non null.
-   *
-   * \param[in] options
-   * \param[in] library_path
-   * \param[out] library_handle OS specific handle to the loaded library (Use FreeLibrary on Windows, dlclose on Linux, etc.. to unload)
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(RegisterCustomOpsLibrary, _Inout_ OrtSessionOptions* options, _In_ const char* library_path, _Outptr_ void** library_handle);
-
-  /// @}
-  /// \name OrtSession
-  /// @{
-
-  /** \brief Get input count for a session
-   *
-   * This number must also match the number of inputs passed to OrtApi::Run
-   *
-   * \see OrtApi::SessionGetInputTypeInfo, OrtApi::SessionGetInputName, OrtApi::Session
-   *
-   * \param[in] session
-   * \param[out] out Number of inputs
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetInputCount, _In_ const OrtSession* session, _Out_ size_t* out);
-
-  /** \brief Get output count for a session
-   *
-   * This number must also match the number of outputs returned by OrtApi::Run
-   *
-   * \see OrtApi::SessionGetOutputTypeInfo, OrtApi::SessionGetOutputName, OrtApi::Session
-   *
-   * \param[in] session
-   * \param[out] out Number of outputs
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetOutputCount, _In_ const OrtSession* session, _Out_ size_t* out);
-
-  /** \brief Get overridable initializer count
-   *
-   * \see OrtApi::SessionGetOverridableInitializerTypeInfo, OrtApi::SessionGetOverridableInitializerName
-   *
-   * \param[in] session
-   * \param[in] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetOverridableInitializerCount, _In_ const OrtSession* session, _Out_ size_t* out);
-
-  /** \brief Get input type information
-   *
-   * \param[in] session
-   * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetInputCount returns (exclusive)
-   * \param[out] type_info Must be freed with OrtApi::ReleaseTypeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetInputTypeInfo, _In_ const OrtSession* session, size_t index, _Outptr_ OrtTypeInfo** type_info);
-
-  /** \brief Get output type information
-   *
-   * \param[in] session
-   * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetOutputCount returns (exclusive)
-   * \param[out] type_info Must be freed with OrtApi::ReleaseTypeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetOutputTypeInfo, _In_ const OrtSession* session, size_t index, _Outptr_ OrtTypeInfo** type_info);
-
-  /** \brief Get overridable initializer type information
-   *
-   * \param[in] session
-   * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetOverridableInitializerCount returns (exclusive)
-   * \param[out] type_info Must be freed with OrtApi::ReleaseTypeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetOverridableInitializerTypeInfo, _In_ const OrtSession* session, size_t index, _Outptr_ OrtTypeInfo** type_info);
-
-  /** \brief Get input name
-   *
-   * \param[in] session
-   * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetInputCount returns (exclusive)
-   * \param[in] allocator
-   * \param[out] value Set to a null terminated UTF-8 encoded string allocated using `allocator`. Must be freed using `allocator`.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetInputName, _In_ const OrtSession* session, size_t index, _Inout_ OrtAllocator* allocator, _Outptr_ char** value);
-
-  /** \brief Get output name
-   *
-   * \param[in] session
-   * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetOutputCount returns (exclusive)
-   * \param[in] allocator
-   * \param[out] value Set to a null terminated UTF-8 encoded string allocated using `allocator`. Must be freed using `allocator`.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetOutputName, _In_ const OrtSession* session, size_t index, _Inout_ OrtAllocator* allocator, _Outptr_ char** value);
-
-  /** \brief Get overridable initializer name
-   *
-   * \param[in] session
-   * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetOverridableInitializerCount returns (exclusive)
-   * \param[in] allocator
-   * \param[out] value Set to a null terminated UTF-8 encoded string allocated using `allocator`. Must be freed using `allocator`.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetOverridableInitializerName, _In_ const OrtSession* session, size_t index,
-                  _Inout_ OrtAllocator* allocator, _Outptr_ char** value);
-
-  /// @}
-  /// \name OrtRunOptions
-  /// @{
-
-  /** \brief Create an OrtRunOptions
-   *
-   * \param[out] out Returned newly created ::OrtRunOptions. Must be freed with OrtApi::ReleaseRunOptions
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateRunOptions, _Outptr_ OrtRunOptions** out);
-
-  /** \brief Set per-run log verbosity level
-   *
-   * \see OrtApi::RunOptionsGetRunLogVerbosityLevel
-   *
-   * \param[in] options
-   * \param[in] log_verbosity_level \snippet{doc} snippets.dox Log Verbosity Level
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(RunOptionsSetRunLogVerbosityLevel, _Inout_ OrtRunOptions* options, int log_verbosity_level);
-
-  /** \brief Set per-run log severity level
-   *
-   * \see OrtApi::RunOptionsGetRunLogSeverityLevel
-   *
-   * \param[in] options
-   * \param[in] log_severity_level The log severity level (refer to ::OrtLoggingLevel for possible values).
-   */
-  ORT_API2_STATUS(RunOptionsSetRunLogSeverityLevel, _Inout_ OrtRunOptions* options, int log_severity_level);
-
-  /** \brief Set per-run tag
-   *
-   * This is used in a per-run log identifier.
-   *
-   * \see OrtApi::RunOptionsGetRunTag
-   *
-   * \param[in] options
-   * \param[in] run_tag The run tag.
-   */
-  ORT_API2_STATUS(RunOptionsSetRunTag, _Inout_ OrtRunOptions* options, _In_ const char* run_tag);
-
-  /** \brief Get per-run log verbosity level
-   *
-   * \see OrtApi::RunOptionsSetRunLogVerbosityLevel
-   *
-   * \param[in] options
-   * \param[out] log_verbosity_level \snippet{doc} snippets.dox Log Verbosity Level
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(RunOptionsGetRunLogVerbosityLevel, _In_ const OrtRunOptions* options,
-                  _Out_ int* log_verbosity_level);
-
-  /** \brief Get per-run log severity level
-   *
-   * \see OrtApi::RunOptionsSetRunLogSeverityLevel
-   *
-   * \param[in] options
-   * \param[out] log_severity_level The log severity level (refer to ::OrtLoggingLevel for possible values).
-   */
-  ORT_API2_STATUS(RunOptionsGetRunLogSeverityLevel, _In_ const OrtRunOptions* options, _Out_ int* log_severity_level);
-
-  /** \brief Get per-run tag
-   *
-   * This is used in a per-run log identifier.
-   *
-   * \see OrtApi::RunOptionsSetRunTag
-   *
-   * \param[in] options
-   * \param[out] run_tag The run tag.
-   *                     Do not free this value, it is owned by `options`. It will be invalidated if the run tag
-   *                     changes (i.e., with OrtApi::RunOptionsSetRunTag) or `options` is freed.
-   */
-  ORT_API2_STATUS(RunOptionsGetRunTag, _In_ const OrtRunOptions* options, _Out_ const char** run_tag);
-
-  /** \brief Set terminate flag
-   *
-   * If a currently executing session needs to be force terminated, this can be called from another thread to force it to fail with an error.
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(RunOptionsSetTerminate, _Inout_ OrtRunOptions* options);
-
-  /** \brief Clears the terminate flag
-   *
-   * Used so the OrtRunOptions instance can be used in a new OrtApi::Run call without it instantly terminating
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(RunOptionsUnsetTerminate, _Inout_ OrtRunOptions* options);
-
-  /// @}
-  /// \name OrtValue
-  /// @{
-
-  /** \brief Create a tensor
-   *
-   * Create a tensor using a supplied ::OrtAllocator
-   *
-   * \param[in] allocator
-   * \param[in] shape Pointer to the tensor shape dimensions.
-   * \param[in] shape_len The number of tensor shape dimensions.
-   * \param[in] type
-   * \param[out] out Returns newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateTensorAsOrtValue, _Inout_ OrtAllocator* allocator, _In_ const int64_t* shape, size_t shape_len,
-                  ONNXTensorElementDataType type, _Outptr_ OrtValue** out);
-
-  /** \brief Create a tensor backed by a user supplied buffer
-   *
-   * Create a tensor with user's buffer. You can fill the buffer either before calling this function or after.
-   * p_data is owned by caller. ReleaseValue won't release p_data.
-   *
-   * \param[in] info Memory description of where the p_data buffer resides (CPU vs GPU etc).
-   * \param[in] p_data Pointer to the data buffer.
-   * \param[in] p_data_len The number of bytes in the data buffer.
-   * \param[in] shape Pointer to the tensor shape dimensions.
-   * \param[in] shape_len The number of tensor shape dimensions.
-   * \param[in] type The data type.
-   * \param[out] out Returns newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateTensorWithDataAsOrtValue, _In_ const OrtMemoryInfo* info, _Inout_ void* p_data,
-                  size_t p_data_len, _In_ const int64_t* shape, size_t shape_len, ONNXTensorElementDataType type,
-                  _Outptr_ OrtValue** out);
-
-  /** \brief Return if an ::OrtValue is a tensor type
-   *
-   * \param[in] value A tensor type (string tensors are not supported)
-   * \param[out] out Set to 1 iff ::OrtValue is a tensor, 0 otherwise
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(IsTensor, _In_ const OrtValue* value, _Out_ int* out);
-
-  /** \brief Get a pointer to the raw data inside a tensor
-   *
-   * Used to read/write/modify the internal tensor data directly.
-   * \note The returned pointer is valid until the \p value is destroyed.
-   *
-   * \param[in] value A tensor type (string tensors are not supported)
-   * \param[out] out Filled in with a pointer to the internal storage
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetTensorMutableData, _In_ OrtValue* value, _Outptr_ void** out);
-
-  /** \brief Set all strings at once in a string tensor
-   *
-   * \param[in,out] value A tensor of type ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING
-   * \param[in] s An array of strings. Each string in this array must be null terminated.
-   * \param[in] s_len Count of strings in s (Must match the size of \p value's tensor shape)
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(FillStringTensor, _Inout_ OrtValue* value, _In_ const char* const* s, size_t s_len);
-
-  /** \brief Get total byte length for all strings in a string tensor
-   *
-   * Typically used with OrtApi::GetStringTensorContent
-   *
-   * \param[in] value A tensor of type ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING
-   * \param[out] len Total byte length of all strings (does not include trailing nulls)
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetStringTensorDataLength, _In_ const OrtValue* value, _Out_ size_t* len);
-
-  /** \brief Get all strings from a string tensor
-   *
-   * An example of the results:<br>
-   * Given \p value is a string tensor with the strings { "This" "is" "a" "test" }<br>
-   * \p s must have a size of 11 bytes<br>
-   * \p offsets must have 4 elements<br>
-   * After the call, these values will be filled in:<br>
-   * \p s will contain "Thisisatest"<br>
-   * \p offsets will contain { 0, 4, 6, 7 }<br>
-   * The length of the last string is just s_len - offsets[last]
-   *
-   * \param[in] value A tensor of type ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING
-   * \param[in] s Buffer to sequentially write all tensor strings to. Each string is NOT null-terminated.
-   * \param[in] s_len Number of bytes of buffer pointed to by \p s (Get it from OrtApi::GetStringTensorDataLength)
-   * \param[out] offsets Array of start offsets into the strings written to \p s
-   * \param[in] offsets_len Number of elements in offsets
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetStringTensorContent, _In_ const OrtValue* value, _Out_writes_bytes_all_(s_len) void* s,
-                  size_t s_len, _Out_writes_all_(offsets_len) size_t* offsets, size_t offsets_len);
-
-  /// @}
-  /// \name OrtTypeInfo
-  /// @{
-
-  /** \brief Get ::OrtTensorTypeAndShapeInfo from an ::OrtTypeInfo
-   *
-   * \param[in] type_info
-   * \param[out] out Do not free this value, it will be valid until type_info is freed.
-   *             If type_info does not represent tensor, this value will be set to nullptr.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CastTypeInfoToTensorInfo, _In_ const OrtTypeInfo* type_info,
-                  _Outptr_result_maybenull_ const OrtTensorTypeAndShapeInfo** out);
-
-  /** \brief Get ::ONNXType from ::OrtTypeInfo
-   *
-   * \param[in] type_info
-   * \param[out] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetOnnxTypeFromTypeInfo, _In_ const OrtTypeInfo* type_info, _Out_ enum ONNXType* out);
-
-  /// @}
-  /// \name OrtTensorTypeAndShapeInfo
-  /// @{
-
-  /** \brief Create an ::OrtTensorTypeAndShapeInfo object
-   *
-   * \param[out] out Returns newly created ::OrtTensorTypeAndShapeInfo. Must be freed with OrtApi::ReleaseTensorTypeAndShapeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateTensorTypeAndShapeInfo, _Outptr_ OrtTensorTypeAndShapeInfo** out);
-
-  /** \brief Set element type in ::OrtTensorTypeAndShapeInfo
-   *
-   * \param[in] info
-   * \param[in] type
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetTensorElementType, _Inout_ OrtTensorTypeAndShapeInfo* info, enum ONNXTensorElementDataType type);
-
-  /** \brief Set shape information in ::OrtTensorTypeAndShapeInfo
-   *
-   * \param[in] info
-   * \param[in] dim_values Array with `dim_count` elements. Can contain negative values.
-   * \param[in] dim_count Number of elements in `dim_values`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetDimensions, OrtTensorTypeAndShapeInfo* info, _In_ const int64_t* dim_values, size_t dim_count);
-
-  /** \brief Get element type in ::OrtTensorTypeAndShapeInfo
-   *
-   * \see OrtApi::SetTensorElementType
-   *
-   * \param[in] info
-   * \param[out] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetTensorElementType, _In_ const OrtTensorTypeAndShapeInfo* info,
-                  _Out_ enum ONNXTensorElementDataType* out);
-
-  /** \brief Get dimension count in ::OrtTensorTypeAndShapeInfo
-   *
-   * \see OrtApi::GetDimensions
-   *
-   * \param[in] info
-   * \param[out] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetDimensionsCount, _In_ const OrtTensorTypeAndShapeInfo* info, _Out_ size_t* out);
-
-  /** \brief Get dimensions in ::OrtTensorTypeAndShapeInfo
-   *
-   * \param[in] info
-   * \param[out] dim_values Array with `dim_values_length` elements. On return, filled with the dimensions stored in the ::OrtTensorTypeAndShapeInfo
-   * \param[in] dim_values_length Number of elements in `dim_values`. Use OrtApi::GetDimensionsCount to get this value
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetDimensions, _In_ const OrtTensorTypeAndShapeInfo* info, _Out_ int64_t* dim_values,
-                  size_t dim_values_length);
-
-  /** \brief Get symbolic dimension names in ::OrtTensorTypeAndShapeInfo
-   *
-   * \param[in] info
-   * \param[in] dim_params Array with `dim_params_length` elements. On return filled with pointers to null terminated strings of the dimension names
-   * \param[in] dim_params_length Number of elements in `dim_params`. Use OrtApi::GetDimensionsCount to get this value
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetSymbolicDimensions, _In_ const OrtTensorTypeAndShapeInfo* info,
-                  _Out_writes_all_(dim_params_length) const char* dim_params[], size_t dim_params_length);
-
-  /** \brief Get total number of elements in a tensor shape from an ::OrtTensorTypeAndShapeInfo
-   *
-   * Return the number of elements specified by the tensor shape (all dimensions multiplied by each other).
-   * For 0 dimensions, 1 is returned. If any dimension is less than 0, the result is always -1.
-   *
-   * Examples:<br>
-   * [] = 1<br>
-   * [1,3,4] = 12<br>
-   * [2,0,4] = 0<br>
-   * [-1,3,4] = -1<br>
-   *
-   * \param[in] info
-   * \param[out] out Number of elements
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetTensorShapeElementCount, _In_ const OrtTensorTypeAndShapeInfo* info, _Out_ size_t* out);
-
-  /// @}
-  /// \name OrtValue
-  /// @{
-
-  /** \brief Get type and shape information from a tensor ::OrtValue
-   *
-   * \param[in] value Must be a tensor (not a map/sequence/etc) or will return failure
-   * \param[out] out Newly created ::OrtTensorTypeAndShapeInfo. Must be freed with OrtApi::ReleaseTensorTypeAndShapeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetTensorTypeAndShape, _In_ const OrtValue* value, _Outptr_ OrtTensorTypeAndShapeInfo** out);
-
-  /** \brief Get type information of an OrtValue
-   *
-   * \param[in] value
-   * \param[out] out Newly created ::OrtTypeInfo. Must be freed with OrtApi::ReleaseTypeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetTypeInfo, _In_ const OrtValue* value, _Outptr_result_maybenull_ OrtTypeInfo** out);
-
-  /** \brief Get ONNXType of an ::OrtValue
-   *
-   * \param[in] value
-   * \param[out] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetValueType, _In_ const OrtValue* value, _Out_ enum ONNXType* out);
-
-  /// @}
-  /// \name OrtMemoryInfo
-  /// @{
-
-  /** \brief Create an ::OrtMemoryInfo
-   *
-   * \param[in] name
-   * \param[in] type
-   * \param[in] id
-   * \param[in] mem_type
-   * \param[out] out Newly created ::OrtMemoryInfo. Must be freed with OrtAPi::ReleaseMemoryInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateMemoryInfo, _In_ const char* name, enum OrtAllocatorType type, int id,
-                  enum OrtMemType mem_type, _Outptr_ OrtMemoryInfo** out);
-
-  /** \brief Create an ::OrtMemoryInfo for CPU memory
-   *
-   * Special case version of OrtApi::CreateMemoryInfo for CPU based memory. Same as using OrtApi::CreateMemoryInfo with name = "Cpu" and id = 0.
-   *
-   * \param[in] type
-   * \param[in] mem_type
-   * \param[out] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateCpuMemoryInfo, enum OrtAllocatorType type, enum OrtMemType mem_type,
-                  _Outptr_ OrtMemoryInfo** out);
-
-  /** \brief Compare ::OrtMemoryInfo objects for equality
-   *
-   * Compares all settings of each ::OrtMemoryInfo for equality
-   *
-   * \param[in] info1
-   * \param[in] info2
-   * \param[out] out Set to 0 if equal, -1 if not equal
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CompareMemoryInfo, _In_ const OrtMemoryInfo* info1, _In_ const OrtMemoryInfo* info2, _Out_ int* out);
-
-  /** \brief Get name from ::OrtMemoryInfo
-   *
-   * \param[in] ptr
-   * \param[out] out Writes null terminated string to this pointer. Do NOT free the returned pointer. It is valid for the lifetime of the ::OrtMemoryInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(MemoryInfoGetName, _In_ const OrtMemoryInfo* ptr, _Out_ const char** out);
-
-  /** \brief Get the id from ::OrtMemoryInfo
-   */
-  ORT_API2_STATUS(MemoryInfoGetId, _In_ const OrtMemoryInfo* ptr, _Out_ int* out);
-
-  /** \brief Get the ::OrtMemType from ::OrtMemoryInfo
-   */
-  ORT_API2_STATUS(MemoryInfoGetMemType, _In_ const OrtMemoryInfo* ptr, _Out_ OrtMemType* out);
-
-  /** \brief Get the ::OrtAllocatorType from ::OrtMemoryInfo
-   */
-  ORT_API2_STATUS(MemoryInfoGetType, _In_ const OrtMemoryInfo* ptr, _Out_ OrtAllocatorType* out);
-
-  /// @}
-  /// \name OrtAllocator
-  /// @{
-
-  /// \brief Calls OrtAllocator::Alloc function
-  ORT_API2_STATUS(AllocatorAlloc, _Inout_ OrtAllocator* ort_allocator, size_t size, _Outptr_ void** out);
-  /// \brief Calls OrtAllocator::Free function
-  ORT_API2_STATUS(AllocatorFree, _Inout_ OrtAllocator* ort_allocator, void* p);
-  /// \brief Calls OrtAllocator::Info function
-  ORT_API2_STATUS(AllocatorGetInfo, _In_ const OrtAllocator* ort_allocator, _Outptr_ const struct OrtMemoryInfo** out);
-
-  /** \brief Get the default allocator
-   *
-   * The default allocator is a CPU based, non-arena. Always returns the same pointer to the same default allocator.
-   *
-   * \param[out] out Returned value should NOT be freed
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetAllocatorWithDefaultOptions, _Outptr_ OrtAllocator** out);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Override session symbolic dimensions
-   *
-   * Override symbolic dimensions (by specific denotation strings) with actual values if known at session initialization time to enable
-   * optimizations that can take advantage of fixed values (such as memory planning, etc)
-   *
-   * \param[in] options
-   * \param[in] dim_denotation
-   * \param[in] dim_value
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(AddFreeDimensionOverride, _Inout_ OrtSessionOptions* options, _In_ const char* dim_denotation,
-                  _In_ int64_t dim_value);
-
-  /// @}
-  /// \name OrtValue
-  /// @{
-
-  /* Internal information (not seen in Doxygen)
-   *
-   * APIs to support non-tensor types - map and sequence.
-   * Currently only the following types are supported
-   * Note: the following types should be kept in sync with data_types.h
-   * Map types
-   * =========
-   * std::map<std::string, std::string>
-   * std::map<std::string, int64_t>
-   * std::map<std::string, float>
-   * std::map<std::string, double>
-   * std::map<int64_t, std::string>
-   * std::map<int64_t, int64_t>
-   * std::map<int64_t, float>
-   * std::map<int64_t, double>
-   *
-   * Sequence types
-   * ==============
-   * std::vector<std::string>
-   * std::vector<int64_t>
-   * std::vector<float>
-   * std::vector<double>
-   * std::vector<std::map<std::string, float>>
-   * std::vector<std::map<int64_t, float>
-   */
-
-  /** \brief Get non tensor data from an ::OrtValue
-   *
-   * If `value` is of type ONNX_TYPE_MAP, you need to retrieve the keys and values
-   * separately. Use index=0 to retrieve keys and index=1 to retrieve values.
-   * If `value` is of type ONNX_TYPE_SEQUENCE, use index to retrieve the index'th element
-   * of the sequence.
-   *
-   * \param[in] value
-   * \param[in] index See above for usage based on `value` type
-   * \param[in] allocator Allocator used to allocate ::OrtValue
-   * \param[out] out Created ::OrtValue that holds the element requested. Must be freed with OrtApi::ReleaseValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetValue, _In_ const OrtValue* value, int index, _Inout_ OrtAllocator* allocator,
-                  _Outptr_ OrtValue** out);
-
-  /** \brief Get non tensor value count from an ::OrtValue
-   *
-   * If `value` is of type ONNX_TYPE_MAP 2 will always be returned. For ONNX_TYPE_SEQUENCE
-   * the number of elements in the sequence will be returned
-   *
-   * \param[in] value
-   * \param[out] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetValueCount, _In_ const OrtValue* value, _Out_ size_t* out);
-
-  /** \brief Create a map or sequence ::OrtValue
-   *
-   * To construct a map (ONNX_TYPE_MAP), use num_values = 2 and `in` should be an array of 2 ::OrtValue%s
-   * representing keys and values.<br>
-   *
-   * To construct a sequence (ONNX_TYPE_SEQUENCE), use num_values = N where N is the number of the elements in the
-   * sequence. 'in' should be an array of N ::OrtValue%s.
-   *
-   * \param[in] in See above for details
-   * \param[in] num_values
-   * \param[in] value_type Must be either ONNX_TYPE_MAP or ONNX_TYPE_SEQUENCE
-   * \param[out] out Newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateValue, _In_reads_(num_values) const OrtValue* const* in, size_t num_values,
-                  enum ONNXType value_type, _Outptr_ OrtValue** out);
-
-  /** \brief Create an opaque (custom user defined type) ::OrtValue
-   *
-   * Constructs an ::OrtValue that contains a value of non-standard type created for
-   * experiments or while awaiting standardization. ::OrtValue in this case would contain
-   * an internal representation of the Opaque type. Opaque types are distinguished from
-   * each other by two strings 1) domain and 2) type name. The combination of the two
-   * must be unique, so the type representation is properly identified internally. The combination
-   * must be properly registered from within ORT at both compile/run time or by another API.
-   *
-   * To construct the ::OrtValue pass domain and type names, also a pointer to a data container
-   * the type of which must be known to both ORT and the client program. That data container may or may
-   * not match the internal representation of the Opaque type. The sizeof(data_container) is passed for
-   * verification purposes.
-   *
-   * \param[in] domain_name Null terminated string of the domain name
-   * \param[in] type_name Null terminated string of the type name
-   * \param[in] data_container User pointer Data to populate ::OrtValue
-   * \param[in] data_container_size Size in bytes of what `data_container` points to
-   * \param[out] out Newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateOpaqueValue, _In_z_ const char* domain_name, _In_z_ const char* type_name,
-                  _In_ const void* data_container, size_t data_container_size, _Outptr_ OrtValue** out);
-
-  /** \brief Get internal data from an opaque (custom user defined type) ::OrtValue
-   *
-   * Copies internal data from an opaque value into a user provided buffer
-   *
-   * \see OrtApi::CreateOpaqueValue
-   *
-   * \param[in] domain_name Null terminated string of the domain name
-   * \param[in] type_name Null terminated string of the type name
-   * \param[in] in The opaque ::OrtValue
-   * \param[out] data_container Buffer to copy data into
-   * \param[out] data_container_size Size in bytes of the buffer pointed to by data_container. Must match the size of the internal buffer.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetOpaqueValue, _In_ const char* domain_name, _In_ const char* type_name, _In_ const OrtValue* in,
-                  _Out_ void* data_container, size_t data_container_size);
-
-  /// @}
-  /// \name OrtKernelInfo
-  /// Custom operator APIs.
-  /// @{
-
-  /** \brief Get a float stored as an attribute in the graph node
-   *
-   * \param[in] info ::OrtKernelInfo instance
-   * \param[in] name Null terminated string of the name of the attribute
-   * \param[out] out Pointer to memory where the attribute will be stored
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(KernelInfoGetAttribute_float, _In_ const OrtKernelInfo* info, _In_ const char* name,
-                  _Out_ float* out);
-
-  /** \brief Fetch a 64-bit int stored as an attribute in the graph node
-   *
-   * \param[in] info ::OrtKernelInfo instance
-   * \param[in] name Null terminated string of the name of the attribute
-   * \param[out] out Pointer to memory where the attribute will be stored
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(KernelInfoGetAttribute_int64, _In_ const OrtKernelInfo* info, _In_ const char* name,
-                  _Out_ int64_t* out);
-
-  /** \brief Fetch a string stored as an attribute in the graph node
-   *
-   * If `out` is nullptr, the value of `size` is set to the true size of the string
-   * attribute, and a success status is returned.
-   *
-   * If the `size` parameter is greater than or equal to the actual string attribute's size,
-   * the value of `size` is set to the true size of the string attribute, the provided memory
-   * is filled with the attribute's contents, and a success status is returned.
-   *
-   * If the `size` parameter is less than the actual string attribute's size and `out`
-   * is not nullptr, the value of `size` is set to the true size of the string attribute
-   * and a failure status is returned.)
-   *
-   * \param[in] info ::OrtKernelInfo instance
-   * \param[in] name Null terminated string of the name of the attribute
-   * \param[out] out Pointer to memory where the attribute will be stored
-   * \param[in,out] size See above comments for details
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(KernelInfoGetAttribute_string, _In_ const OrtKernelInfo* info, _In_ const char* name, _Out_ char* out,
-                  _Inout_ size_t* size);
-
-  /// @}
-  /// \name OrtKernelContext
-  /// Custom operator APIs.
-  /// @{
-
-  /** \brief Used for custom operators, get the input count of a kernel
-   *
-   * \see ::OrtCustomOp
-   */
-  ORT_API2_STATUS(KernelContext_GetInputCount, _In_ const OrtKernelContext* context, _Out_ size_t* out);
-
-  /** \brief Used for custom operators, get the output count of a kernel
-   *
-   * \see ::OrtCustomOp
-   */
-  ORT_API2_STATUS(KernelContext_GetOutputCount, _In_ const OrtKernelContext* context, _Out_ size_t* out);
-
-  /** \brief Used for custom operators, get an input of a kernel
-   *
-   * \see ::OrtCustomOp
-   */
-  ORT_API2_STATUS(KernelContext_GetInput, _In_ const OrtKernelContext* context, _In_ size_t index,
-                  _Out_ const OrtValue** out);
-
-  /** \brief Used for custom operators, get an output of a kernel
-   *
-   * \see ::OrtCustomOp
-   */
-  ORT_API2_STATUS(KernelContext_GetOutput, _Inout_ OrtKernelContext* context, _In_ size_t index,
-                  _In_ const int64_t* dim_values, size_t dim_count, _Outptr_ OrtValue** out);
-
-  /// @}
-  /// \name OrtEnv
-  /// @{
-  ORT_CLASS_RELEASE(Env);
-  /// @}
-  /// \name OrtStatus
-  /// @{
-  ORT_CLASS_RELEASE(Status);
-  /// @}
-  /// \name OrtMemoryInfo
-  /// @{
-  ORT_CLASS_RELEASE(MemoryInfo);
-  /// @}
-  /// \name OrtSession
-  /// @{
-  ORT_CLASS_RELEASE(Session);  // Don't call ReleaseSession from Dllmain (because session owns a thread pool)
-  /// @}
-  /// \name OrtValue
-  /// @{
-  ORT_CLASS_RELEASE(Value);
-  /// @}
-  /// \name OrtRunOptions
-  /// @{
-  ORT_CLASS_RELEASE(RunOptions);
-  /// @}
-  /// \name OrtTypeInfo
-  /// @{
-  ORT_CLASS_RELEASE(TypeInfo);
-  /// @}
-  /// \name OrtTensorTypeAndShapeInfo
-  /// @{
-  ORT_CLASS_RELEASE(TensorTypeAndShapeInfo);
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-  ORT_CLASS_RELEASE(SessionOptions);
-  /// @}
-  /// \name OrtCustomOpDomain
-  /// @{
-  ORT_CLASS_RELEASE(CustomOpDomain);
-
-  /// @}
-  /// \name OrtTypeInfo
-  /// @{
-
-  /** \brief Get denotation from type information
-   *
-   * Augments ::OrtTypeInfo to return denotations on the type.
-   *
-   * This is used by WinML to determine if an input/output is intended to be an Image or a Tensor.
-   *
-   * \param[in] type_info
-   * \param[out] denotation Pointer to the null terminated denotation string is written to this pointer. This pointer is valid until the object is destroyed or the name is changed, do not free.
-   * \param[out] len Length in bytes of the string returned in `denotation`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetDenotationFromTypeInfo, _In_ const OrtTypeInfo* type_info, _Out_ const char** const denotation,
-                  _Out_ size_t* len);
-
-  /** \brief Get detailed map information from an ::OrtTypeInfo
-   *
-   * This augments ::OrtTypeInfo to return an ::OrtMapTypeInfo when the type is a map.
-   * The OrtMapTypeInfo has additional information about the map's key type and value type.
-   *
-   * This is used by WinML to support model reflection APIs.
-   *
-   * \param[out] type_info
-   * \param[out] out A pointer to the ::OrtMapTypeInfo. Do not free this value. If type_info
-   *             does not contain a map, this value will be set to nullptr.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CastTypeInfoToMapTypeInfo, _In_ const OrtTypeInfo* type_info,
-                  _Outptr_result_maybenull_ const OrtMapTypeInfo** out);
-
-  /** \brief Cast ::OrtTypeInfo to an ::OrtSequenceTypeInfo
-   *
-   * This api augments ::OrtTypeInfo to return an ::OrtSequenceTypeInfo when the type is a sequence.
-   * The ::OrtSequenceTypeInfo has additional information about the sequence's element type.
-   *
-   * This is used by WinML to support model reflection APIs.
-   *
-   * \param[in] type_info
-   * \param[out] out A pointer to the OrtSequenceTypeInfo. Do not free this value. If type_info
-   *             doesn not contain a sequence, this value will be set to nullptr.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CastTypeInfoToSequenceTypeInfo, _In_ const OrtTypeInfo* type_info,
-                  _Outptr_result_maybenull_ const OrtSequenceTypeInfo** out);
-
-  /// @}
-  /// \name OrtMapTypeInfo
-  /// @{
-
-  /** \brief Get key type from an ::OrtMapTypeInfo
-   *
-   * Key types are restricted to being scalar types.
-   *
-   * This is used by WinML to support model reflection APIs.
-   *
-   * \param[in] map_type_info
-   * \param[out] out
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetMapKeyType, _In_ const OrtMapTypeInfo* map_type_info, _Out_ enum ONNXTensorElementDataType* out);
-
-  /** \brief Get the value type from an ::OrtMapTypeInfo
-   *
-   * \param[in] map_type_info
-   * \param[out] type_info
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetMapValueType, _In_ const OrtMapTypeInfo* map_type_info, _Outptr_ OrtTypeInfo** type_info);
-
-  /// @}
-  /// \name OrtSequenceTypeInfo
-  /// @{
-
-  /** \brief Get element type from an ::OrtSequenceTypeInfo
-   *
-   * This is used by WinML to support model reflection APIs.
-   *
-   * \param[in] sequence_type_info
-   * \param[out] type_info
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetSequenceElementType, _In_ const OrtSequenceTypeInfo* sequence_type_info,
-                  _Outptr_ OrtTypeInfo** type_info);
-
-  /// @}
-  /// \name OrtMapTypeInfo
-  /// @{
-  ORT_CLASS_RELEASE(MapTypeInfo);
-  /// @}
-  /// \name OrtSequenceTypeInfo
-  /// @{
-  ORT_CLASS_RELEASE(SequenceTypeInfo);
-
-  /// @}
-  /// \name OrtSession
-  /// @{
-
-  /** \brief End profiling and return filename of the profile data
-   *
-   * Profiling is turned on through OrtApi::EnableProfiling
-   *
-   * \param[in] session
-   * \param[in] allocator
-   * \param[out] out Null terminated string of the filename, allocated using `allocator`. Must be freed using `allocator`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionEndProfiling, _In_ OrtSession* session, _Inout_ OrtAllocator* allocator, _Outptr_ char** out);
-
-  /** \brief Get ::OrtModelMetadata from an ::OrtSession
-   *
-   * \param[in] session
-   * \param[out] out Newly created ::OrtModelMetadata. Must be freed using OrtApi::ReleaseModelMetadata
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetModelMetadata, _In_ const OrtSession* session, _Outptr_ OrtModelMetadata** out);
-
-  /// @}
-  /// \name OrtModelMetadata
-  /// @{
-
-  /** \brief Get `producer name` from an ::OrtModelMetadata
-   *
-   * \param[in] model_metadata
-   * \param[in] allocator
-   * \param[out] value Set to a null terminated string allocated using `allocator`. Must be freed using `allocator`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataGetProducerName, _In_ const OrtModelMetadata* model_metadata,
-                  _Inout_ OrtAllocator* allocator, _Outptr_ char** value);
-
-  /** \brief Get `graph name` from an ::OrtModelMetadata
-   *
-   * \param[in] model_metadata
-   * \param[in] allocator
-   * \param[out] value Set to a null terminated string allocated using `allocator`. Must be freed using `allocator`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataGetGraphName, _In_ const OrtModelMetadata* model_metadata,
-                  _Inout_ OrtAllocator* allocator, _Outptr_ char** value);
-
-  /** \brief Get `domain` from an ::OrtModelMetadata
-   *
-   * \param[in] model_metadata
-   * \param[in] allocator
-   * \param[out] value Set to a null terminated string allocated using `allocator`. Must be freed using `allocator`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataGetDomain, _In_ const OrtModelMetadata* model_metadata, _Inout_ OrtAllocator* allocator,
-                  _Outptr_ char** value);
-
-  /** \brief Get `description` from an ::OrtModelMetadata
-   *
-   * \param[in] model_metadata
-   * \param[in] allocator
-   * \param[out] value Set to a null terminated string allocated using `allocator`. Must be freed using `allocator`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataGetDescription, _In_ const OrtModelMetadata* model_metadata,
-                  _Inout_ OrtAllocator* allocator, _Outptr_ char** value);
-
-  /** \brief Return data for a key in the custom metadata map in an ::OrtModelMetadata
-   *
-   * \param[in] model_metadata
-   * \param[in] allocator
-   * \param[in] key Null terminated string
-   * \param[out] value Set to a null terminated string allocated using `allocator`. Must be freed using `allocator`
-   * `value` will be set to nullptr if the given key is not found in the custom metadata map.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataLookupCustomMetadataMap, _In_ const OrtModelMetadata* model_metadata,
-                  _Inout_ OrtAllocator* allocator, _In_ const char* key, _Outptr_result_maybenull_ char** value);
-
-  /** \brief Get version number from an ::OrtModelMetadata
-   *
-   * \param[in] model_metadata
-   * \param[out] value Set to the version number
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataGetVersion, _In_ const OrtModelMetadata* model_metadata, _Out_ int64_t* value);
-
-  ORT_CLASS_RELEASE(ModelMetadata);
-
-  /// @}
-  /// \name OrtEnv
-  /// @{
-
-  /** \brief Create an OrtEnv
-   *
-   * Create an environment with global threadpools that will be shared across sessions.
-   * Use this in conjunction with OrtApi::DisablePerSessionThreads or else the session will use
-   * its own thread pools.
-   *
-   * \param[in] log_severity_level The log severity level.
-   * \param[in] logid The log identifier.
-   * \param[in] tp_options
-   * \param[out] out Returned newly created OrtEnv. Must be freed with OrtApi::ReleaseEnv
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateEnvWithGlobalThreadPools, OrtLoggingLevel log_severity_level, _In_ const char* logid,
-                  _In_ const OrtThreadingOptions* tp_options, _Outptr_ OrtEnv** out);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Use global thread pool on a session
-   *
-   * Disable using per session thread pool and use the shared global threadpool.
-   * This should be used in conjunction with OrtApi::CreateEnvWithGlobalThreadPools.
-   *
-   * \param[in] options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(DisablePerSessionThreads, _Inout_ OrtSessionOptions* options);
-
-  /// @}
-  /// \name OrtThreadingOptions
-  /// @{
-
-  /** \brief Create an ::OrtThreadingOptions
-   *
-   * \param[out] out Newly created ::OrtThreadingOptions. Must be freed with OrtApi::ReleaseThreadingOptions
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateThreadingOptions, _Outptr_ OrtThreadingOptions** out);
-
-  ORT_CLASS_RELEASE(ThreadingOptions);
-
-  /// @}
-  /// \name OrtModelMetadata
-  /// @{
-
-  /**
-   *
-   * \param[in] model_metadata
-   * \param[in] allocator
-   * \param[out] keys Array of null terminated strings (array count = num_keys) allocated using `allocator`.
-   *  The strings and the pointer array must be freed using `allocator`
-   *  `keys` will be set to nullptr if the custom metadata map is empty.
-   * \param[out] num_keys Set to the number of elements in the `keys` array
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataGetCustomMetadataMapKeys, _In_ const OrtModelMetadata* model_metadata,
-                  _Inout_ OrtAllocator* allocator, _Outptr_result_buffer_maybenull_(*num_keys) char*** keys, _Out_ int64_t* num_keys);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /**
-   *
-   * Override symbolic dimensions (by specific name strings) with actual values
-   * if known at session initialization time to enable optimizations that can
-   * take advantage of fixed values (such as memory planning, etc)
-   *
-   */
-  ORT_API2_STATUS(AddFreeDimensionOverrideByName,
-                  _Inout_ OrtSessionOptions* options, _In_ const char* dim_name,
-                  _In_ int64_t dim_value);
-
-  /// @}
-  /// \name Misc
-  /// @{
-
-  /** \brief Get the names of all available providers
-   *
-   * \note The providers in the list are not guaranteed to be usable. They may fail to load due to missing system dependencies.
-   *    For example, if the CUDA/cuDNN libraries are not installed, the CUDA provider will report an error when it is added to the session options.
-   *
-   * \param[out] out_ptr Set to a pointer to an array of null terminated strings of the available providers. The entries and the
-   *    array itself must be freed using OrtApi::ReleaseAvailableProviders
-   * \param[out] provider_length Set to the number of entries in the `out_ptr` array
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetAvailableProviders, _Outptr_ char*** out_ptr, _Out_ int* provider_length);
-
-  /** \brief Release data from OrtApi::GetAvailableProviders. This API will never fail
-   * so you can rely on it in a noexcept code.
-   *
-   * \param[in] ptr The `out_ptr` result from OrtApi::GetAvailableProviders.
-   * \param[in] providers_length The `provider_length` result from OrtApi::GetAvailableProviders
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ReleaseAvailableProviders, _In_ char** ptr,
-                  _In_ int providers_length);
-
-  /// @}
-  /// \name OrtValue
-  /// @{
-
-  /** \brief Get the length of a single string in a string tensor
-   *
-   * \param[in] value A string tensor
-   * \param[in] index Index of the string in the tensor
-   * \param[out] out Set to number of bytes of the string element
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetStringTensorElementLength, _In_ const OrtValue* value, size_t index, _Out_ size_t* out);
-
-  /** \brief Get a single string from a string tensor
-   *
-   * \param[in] value A string tensor
-   * \param[in] s_len Number of bytes in the `s` buffer. Must match the value returned by OrtApi::GetStringTensorElementLength.
-   * \param[in] index Index of the string in the tensor
-   * \param[out] s The string element contents in UTF-8 encoding. The string is NOT null-terminated.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetStringTensorElement, _In_ const OrtValue* value, size_t s_len, size_t index, _Out_writes_bytes_all_(s_len) void* s);
-
-  /** \brief Set a single string in a string tensor
-   *
-   * \param[in] value A string tensor
-   * \param[in] s A null terminated UTF-8 encoded string
-   * \param[in] index Index of the string in the tensor to set
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(FillStringTensorElement, _Inout_ OrtValue* value, _In_ const char* s, size_t index);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Set a session configuration entry as a pair of strings
-   *
-   * If a configuration with same key exists, this will overwrite the configuration with the given config_value.
-   *
-   * The config_key and the format of config_value are defined in onnxruntime_session_options_config_keys.h
-   *
-   * \param[in] options
-   * \param[in] config_key A null terminated string representation of the config key
-   * \param[in] config_value A null terminated string representation of the config value
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(AddSessionConfigEntry, _Inout_ OrtSessionOptions* options,
-                  _In_z_ const char* config_key, _In_z_ const char* config_value);
-
-  /// @}
-  /// \name OrtAllocator
-  /// @{
-
-  /** \brief Create an allocator for an ::OrtSession following an ::OrtMemoryInfo
-   *
-   * \param[in] session
-   * \param[in] mem_info valid ::OrtMemoryInfo instance
-   * \param[out] out Newly created ::OrtAllocator. Must be freed with OrtApi::ReleaseAllocator
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateAllocator, _In_ const OrtSession* session, _In_ const OrtMemoryInfo* mem_info,
-                  _Outptr_ OrtAllocator** out);
-
-  /** \brief Release an ::OrtAllocator obtained from OrtApi::CreateAllocator
-   */
-  ORT_CLASS_RELEASE(Allocator);
-
-  /// @}
-  /// \name OrtSession
-  /// @{
-
-  /** \brief Run a model using Io Bindings for the inputs & outputs
-   *
-   * \see OrtApi::Run
-   *
-   * \param[in] session
-   * \param[in] run_options
-   * \param[in] binding_ptr
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(RunWithBinding, _Inout_ OrtSession* session, _In_ const OrtRunOptions* run_options, _In_ const OrtIoBinding* binding_ptr);
-
-  /** \brief Create an ::OrtIoBinding instance
-   *
-   * An IoBinding object allows one to bind pre-allocated ::OrtValue%s to input names.
-   * Thus if you want to use a raw on device buffer as input or output you can avoid
-   * extra copy during runtime.
-   *
-   * \param[in] session
-   * \param[out] out Newly created ::OrtIoBinding. Must be freed with OrtApi::ReleaseIoBinding
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateIoBinding, _Inout_ OrtSession* session, _Outptr_ OrtIoBinding** out);
-
-  /// @}
-  /// \name OrtIoBinding
-  /// @{
-
-  /** \brief Release an ::OrtIoBinding obtained from OrtApi::CreateIoBinding
-   */
-  ORT_CLASS_RELEASE(IoBinding);
-
-  /** \brief Bind an ::OrtValue to an ::OrtIoBinding input
-   *
-   * When using OrtApi::RunWithBinding this value is used for the named input
-   *
-   * \param[in] binding_ptr
-   * \param[in] name Name for the model input
-   * \param[in] val_ptr ::OrtValue of Tensor type.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(BindInput, _Inout_ OrtIoBinding* binding_ptr, _In_ const char* name, _In_ const OrtValue* val_ptr);
-
-  /** \brief Bind an ::OrtValue to an ::OrtIoBinding output
-   *
-   * When using OrtApi::RunWithBinding this value is used for the named output
-   *
-   * \param[in] binding_ptr
-   * \param[in] name Null terminated string of the model output name
-   * \param[in] val_ptr ::OrtValue of Tensor type.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(BindOutput, _Inout_ OrtIoBinding* binding_ptr, _In_ const char* name, _In_ const OrtValue* val_ptr);
-
-  /** \brief Bind an ::OrtIoBinding output to a device
-   *
-   * Binds the ::OrtValue to a device which is specified by ::OrtMemoryInfo.
-   * You can either create an instance of ::OrtMemoryInfo with a device id or obtain one from the allocator that you have created/are using
-   * This is useful when one or more outputs have dynamic shapes and, it is hard to pre-allocate and bind a chunk of
-   * memory within ::OrtValue ahead of time.
-   *
-   * \see OrtApi::RunWithBinding
-   *
-   * \param[in] binding_ptr
-   * \param[in] name Null terminated string of the device name
-   * \param[in] mem_info_ptr
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(BindOutputToDevice, _Inout_ OrtIoBinding* binding_ptr, _In_ const char* name, _In_ const OrtMemoryInfo* mem_info_ptr);
-
-  /** \brief Get the names of an ::OrtIoBinding's outputs
-   *
-   * Returns the names of the outputs in the order they were bound. This is useful after running the model
-   * with bound outputs because the returned names are in order in which output ::OrtValue are returned. This is useful if
-   * the order of outputs and their names is not known.
-   *
-   * \param[in] binding_ptr
-   * \param[in] allocator Allocator used to allocate continuous buffers for output strings and lengths.
-   * \param[out] buffer Returns an array of non-null terminated UTF-8 strings. The number of strings stored is returned in the count parameter.
-   *   This buffer is allocated using `allocator` and must be freed using it.
-   * \param[out] lengths Returns an array of `count` lengths of the strings returned in `buffer`
-   *   This buffer is allocated using `allocator` and must be freed using it.
-   * \param[out] count Number of strings returned. If `binding_ptr` has no bound outputs, zero is returned,
-   *              no memory allocation is performed and buffer and lengths are set to nullptr.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetBoundOutputNames, _In_ const OrtIoBinding* binding_ptr, _In_ OrtAllocator* allocator,
-                  _Out_ char** buffer, _Out_writes_all_(count) size_t** lengths, _Out_ size_t* count);
-
-  /** \brief Get the output ::OrtValue objects from an ::OrtIoBinding
-   *
-   * Returns an array of pointers to individually allocated ::OrtValue%s that contain results of a model execution with OrtApi::RunWithBinding
-   * The array contains the same number of ::OrtValue%s and they are in the same order as they were bound with OrtApi::BindOutput
-   * or OrtApi::BindOutputToDevice.
-   *
-   * The returned ::OrtValue%s must be released using OrtApi::ReleaseValue after they are no longer needed.
-   * The array is allocated using the specified instance of the allocator and must be freed using the same allocator after
-   * all the ::OrtValue%s contained therein are individually released.
-   *
-   * \param[in] binding_ptr
-   * \param[in] allocator Allocator used to allocate output array
-   * \param[out] output Set to the allocated array of allocated ::OrtValue outputs. Set to nullptr if there are 0 outputs.
-   * \param[out] output_count Set to number of ::OrtValue%s returned
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetBoundOutputValues, _In_ const OrtIoBinding* binding_ptr, _In_ OrtAllocator* allocator,
-                  _Out_writes_all_(output_count) OrtValue*** output, _Out_ size_t* output_count);
-
-  /** \brief Clears any previously set Inputs for an ::OrtIoBinding
-   */
-  void(ORT_API_CALL* ClearBoundInputs)(_Inout_ OrtIoBinding* binding_ptr) NO_EXCEPTION ORT_ALL_ARGS_NONNULL;
-
-  /** \brief Clears any previously set Outputs for an ::OrtIoBinding
-   */
-  void(ORT_API_CALL* ClearBoundOutputs)(_Inout_ OrtIoBinding* binding_ptr) NO_EXCEPTION ORT_ALL_ARGS_NONNULL;
-
-  /// @}
-  /// \name OrtValue
-  /// @{
-
-  /** \brief Direct memory access to a specified tensor element
-   *
-   * For example, given a tensor with shape of [3,224,224], a pointer to the element at location [2,150,128] can be retrieved
-   *
-   * This function only works for numeric type tensors (No strings, etc).
-   * This is a no-copy method whose returned pointer is valid until the passed in ::OrtValue is free'd.
-   *
-   * \param[in] value
-   * \param[in] location_values Pointer to an array of index values that specify an element's location relative to its shape
-   * \param[in] location_values_count Number of elements in location_values. Must match the number of elements in the tensor's shape.
-   * \param[out] out Set to a pointer to the element specified
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(TensorAt, _Inout_ OrtValue* value, const int64_t* location_values, size_t location_values_count, _Outptr_ void** out);
-
-  /// @}
-  /// \name OrtEnv
-  /// @{
-
-  /** \brief Create an allocator and register it with the ::OrtEnv
-   *
-   * Enables sharing the allocator between multiple sessions that use the same env instance.
-   * Lifetime of the created allocator will be valid for the duration of the environment.
-   * Returns an error if an allocator with the same ::OrtMemoryInfo is already registered.
-   *
-   * See https://onnxruntime.ai/docs/get-started/with-c.html for details.
-   *
-   * \param[in] env ::OrtEnv instance
-   * \param[in] mem_info
-   * \param[in] arena_cfg Pass nullptr for defaults
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateAndRegisterAllocator, _Inout_ OrtEnv* env, _In_ const OrtMemoryInfo* mem_info,
-                  _In_ const OrtArenaCfg* arena_cfg);
-
-  /** \brief Set language projection
-   *
-   * Set the language projection for collecting telemetry data when Env is created.
-   *
-   * The default is ORT_PROJECTION_C, which means it will classify the language not in the list to C also.
-   *
-   * \param[in] ort_env
-   * \param[in] projection
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetLanguageProjection, _In_ const OrtEnv* ort_env, _In_ OrtLanguageProjection projection);
-
-  /// @}
-  /// \name OrtSession
-  /// @{
-
-  /** \brief Return the time that profiling was started
-   *
-   * \note The timer precision varies per platform. On Windows and MacOS, the precision will be ~100ns
-   *
-   * \param[in] session
-   * \param[out] out nanoseconds of profiling's start time
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionGetProfilingStartTimeNs, _In_ const OrtSession* session, _Outptr_ uint64_t* out);
-
-  /// @}
-  /// \name OrtThreadingOptions
-  /// @{
-
-  /** \brief Set global intra-op thread count
-   *
-   * This configures the global thread pool options to be used in the call to OrtApi::CreateEnvWithGlobalThreadPools
-   *
-   * \param[in] tp_options
-   * \param[in] intra_op_num_threads Number of threads, special values:<br>
-   *    0 = Use default thread count<br>
-   *    1 = The invoking thread will be used; no threads will be created in the thread pool.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetGlobalIntraOpNumThreads, _Inout_ OrtThreadingOptions* tp_options, int intra_op_num_threads);
-
-  /** \brief Set global inter-op thread count
-   *
-   * This configures the global thread pool options to be used in the call to OrtApi::CreateEnvWithGlobalThreadPools
-   *
-   * \param[in] tp_options
-   * \param[in] inter_op_num_threads Number of threads, special values:<br>
-   *    0 = Use default thread count<br>
-   *    1 = The invoking thread will be used; no threads will be created in the thread pool.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetGlobalInterOpNumThreads, _Inout_ OrtThreadingOptions* tp_options, int inter_op_num_threads);
-
-  /** \brief Set global spin control options
-   *
-   * This will configure the global thread pool options to be used in the call to OrtApi::CreateEnvWithGlobalThreadPools.
-   * Allow spinning of thread pools when their queues are empty. This will set the value for both
-   * inter_op and intra_op threadpools.
-   *
-   * \param[in] tp_options
-   * \param[in] allow_spinning Valid values are 0 or 1.<br>
-   *   0 = It won't spin (recommended if CPU usage is high)<br>
-   *   1 = Threadpool will spin to wait for queue to become non-empty
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetGlobalSpinControl, _Inout_ OrtThreadingOptions* tp_options, int allow_spinning);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Add a pre-allocated initializer to a session
-   *
-   * If a model contains an initializer with a name that is same as the name passed to this call,
-   * ORT will use this initializer instance instead of deserializing one from the model file. This
-   * is useful when you want to share the same initializer across sessions.
-   *
-   * \param[in] options
-   * \param[in] name Null terminated string of the initializer name
-   * \param[in] val ::OrtValue containing the initializer. Its lifetime and the underlying initializer buffer must be
-   *   managed by the user (created using the OrtApi::CreateTensorWithDataAsOrtValue) and it must outlive the session object
-   *   to which it is added.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(AddInitializer, _Inout_ OrtSessionOptions* options, _In_z_ const char* name,
-                  _In_ const OrtValue* val);
-
-  /// @}
-  /// \name OrtEnv
-  /// @{
-
-  /**
-   * Create a custom environment with global threadpools and logger that will be shared across sessions.
-   * Use this in conjunction with OrtApi::DisablePerSessionThreads or else the session will use
-   * its own thread pools.
-   *
-   * \param[in] logging_function A pointer to a logging function.
-   * \param[in] logger_param A pointer to arbitrary data passed as the ::OrtLoggingFunction `param` parameter to
-   *                         `logging_function`.
-   * \param[in] log_severity_level The log severity level.
-   * \param[in] logid The log identifier.
-   * \param[in] tp_options
-   * \param[out] out Newly created OrtEnv. Must be freed with OrtApi::ReleaseEnv
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateEnvWithCustomLoggerAndGlobalThreadPools, OrtLoggingFunction logging_function, _In_opt_ void* logger_param, OrtLoggingLevel log_severity_level,
-                  _In_ const char* logid, _In_ const struct OrtThreadingOptions* tp_options, _Outptr_ OrtEnv** out);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Append CUDA provider to session options
-   *
-   * If CUDA is not available (due to a non CUDA enabled build, or if CUDA is not installed on the system), this function will return failure.
-   *
-   * \param[in] options
-   * \param[in] cuda_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_CUDA,
-                  _In_ OrtSessionOptions* options, _In_ const OrtCUDAProviderOptions* cuda_options);
-
-  /** \brief Append ROCM execution provider to the session options
-   *
-   * If ROCM is not available (due to a non ROCM enabled build, or if ROCM is not installed on the system), this function will return failure.
-   *
-   * \param[in] options
-   * \param[in] rocm_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_ROCM,
-                  _In_ OrtSessionOptions* options, _In_ const OrtROCMProviderOptions* rocm_options);
-
-  /** \brief Append OpenVINO execution provider to the session options
-   *
-   * If OpenVINO is not available (due to a non OpenVINO enabled build, or if OpenVINO is not installed on the system), this function will fail.
-   *
-   * \param[in] options
-   * \param[in] provider_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_OpenVINO,
-                  _In_ OrtSessionOptions* options, _In_ const OrtOpenVINOProviderOptions* provider_options);
-
-  /// @}
-  /// \name OrtThreadingOptions
-  /// @{
-
-  /** \brief Set threading flush-to-zero and denormal-as-zero
-   *
-   * Sets global thread pool options to be used in the call to OrtApi::CreateEnvWithGlobalThreadPools.
-   * Flush-to-zero and denormal-as-zero are applied to threads in both intra and inter global thread pool.
-   * \note This option is not needed if the models used have no denormals. Having no denormals is recommended as this option may hurt model accuracy.
-   *
-   * \param[in] tp_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetGlobalDenormalAsZero, _Inout_ OrtThreadingOptions* tp_options);
-
-  /// @}
-  /// \name OrtArenaCfg
-  /// @{
-
-  /** \deprecated Use OrtApi::CreateArenaCfgV2
-   *
-   * This will create the configuration of an arena that can eventually be used to define an arena based allocator's behavior
-   *
-   * \param[in] max_mem Use 0 to allow ORT to choose the default
-   * \param[in] arena_extend_strategy Use -1 to allow ORT to choose the default, 0 = kNextPowerOfTwo, 1 = kSameAsRequested
-   * \param[in] initial_chunk_size_bytes Use -1 to allow ORT to choose the default
-   * \param[in] max_dead_bytes_per_chunk Use -1 to allow ORT to choose the default
-   * \param[in] out A pointer to an OrtArenaCfg instance
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateArenaCfg, _In_ size_t max_mem, int arena_extend_strategy, int initial_chunk_size_bytes,
-                  int max_dead_bytes_per_chunk, _Outptr_ OrtArenaCfg** out);
-
-  ORT_CLASS_RELEASE(ArenaCfg);
-
-  /// @}
-  /// \name OrtModelMetadata
-  /// @{
-
-  /**
-   * Use this to obtain the description of the graph present in the model
-   * (doc_string field of the GraphProto message within the ModelProto message).
-   * If it doesn't exist, an empty string will be returned.
-   *
-   * \param[in] model_metadata An instance of ::OrtModelMetadata
-   * \param[in] allocator Allocator used to allocate the string that will be returned back
-   * \param[out] value Set to a null terminated string allocated using `allocator`.  The caller is responsible for freeing it using `allocator`
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(ModelMetadataGetGraphDescription, _In_ const OrtModelMetadata* model_metadata,
-                  _Inout_ OrtAllocator* allocator, _Outptr_ char** value);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Append TensorRT provider to session options
-   *
-   * If TensorRT is not available (due to a non TensorRT enabled build, or if TensorRT is not installed on the system), this function will return failure.
-   *
-   * \param[in] options
-   * \param[in] tensorrt_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_TensorRT,
-                  _In_ OrtSessionOptions* options, _In_ const OrtTensorRTProviderOptions* tensorrt_options);
-
-  /// @}
-  /// \name Misc
-  /// @{
-
-  /** \brief Set current GPU device ID
-   *
-   * Set the current device id of the GPU execution provider (CUDA/tensorrt/rocm). The device id should be less
-   * than the total number of devices available. This is only useful when multiple-GPUs are installed and it is
-   * required to restrict execution to a single GPU.
-   *
-   * \param[in] device_id
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetCurrentGpuDeviceId, _In_ int device_id);
-
-  /** \brief Get current GPU device ID
-   *
-   * Get the current device id of the GPU execution provider (CUDA/tensorrt/rocm).
-   *
-   * \see OrtApi::SetCurrentGpuDeviceId
-   *
-   * \param[out] device_id
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetCurrentGpuDeviceId, _In_ int* device_id);
-
-  /// @}
-  /// \name OrtKernelInfo
-  /// Custom operator APIs.
-  /// @{
-
-  /** \brief Fetch an array of int64_t values stored as an attribute in the graph node
-   *
-   *
-   * If `out` is nullptr, the value of `size` is set to the true size of the attribute
-   * array's size, and a success status is returned.
-   *
-   * If the `size` parameter is greater than or equal to the actual attribute array's size,
-   * the value of `size` is set to the true size of the attribute array's size,
-   * the provided memory is filled with the attribute's contents,
-   * and a success status is returned.
-   *
-   * If the `size` parameter is less than the actual attribute array's size and `out`
-   * is not nullptr, the value of `size` is set to the true size of the attribute array's size
-   * and a failure status is returned.)
-   *
-   * \param[in] info instance
-   * \param[in] name name of the attribute to be parsed
-   * \param[out] out pointer to memory where the attribute's contents are to be stored
-   * \param[in, out] size actual size of attribute array
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(KernelInfoGetAttributeArray_float, _In_ const OrtKernelInfo* info, _In_ const char* name,
-                  _Out_ float* out, _Inout_ size_t* size);
-
-  /** \brief Fetch an array of int64_t values stored as an attribute in the graph node
-   *
-   * If `out` is nullptr, the value of `size` is set to the true size of the attribute
-   * array's size, and a success status is returned.
-   *
-   * If the `size` parameter is greater than or equal to the actual attribute array's size,
-   * the value of `size` is set to the true size of the attribute array's size,
-   * the provided memory is filled with the attribute's contents,
-   * and a success status is returned.
-   *
-   * If the `size` parameter is less than the actual attribute array's size and `out`
-   * is not nullptr, the value of `size` is set to the true size of the attribute array's size
-   * and a failure status is returned.)
-   *
-   * \param[in] info instance
-   * \param[in] name name of the attribute to be parsed
-   * \param[out] out pointer to memory where the attribute's contents are to be stored
-   * \param[in, out] size actual size of attribute array
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(KernelInfoGetAttributeArray_int64, _In_ const OrtKernelInfo* info, _In_ const char* name,
-                  _Out_ int64_t* out, _Inout_ size_t* size);
-
-  /// @}
-  /// \name OrtArenaCfg
-  /// @{
-
-  /** \brief Create an ::OrtArenaCfg
-   *
-   * Create the configuration of an arena that can eventually be used to define an arena based allocator's behavior.
-   *
-   * Supported keys are (See https://onnxruntime.ai/docs/get-started/with-c.html for details on what the
-   * following parameters mean and how to choose these values.):
-   * "max_mem": Maximum memory that can be allocated by the arena based allocator.
-   *  Use 0 for ORT to pick the best value. Default is 0.
-   * "arena_extend_strategy": 0 = kNextPowerOfTwo, 1 = kSameAsRequested.
-   *  Use -1 to allow ORT to choose the default.
-   * "initial_chunk_size_bytes": (Possible) Size of the first allocation in the arena.
-   *  Only relevant if arena strategy is `kNextPowerOfTwo`. Use -1 to allow ORT to choose the default.
-   *  Ultimately, the first allocation size is determined by the allocation memory request.
-   * "max_dead_bytes_per_chunk": Threshold of unused memory in an allocated chunk of arena memory after
-   *  crossing which the current chunk is chunked into 2.
-   * "initial_growth_chunk_size_bytes": (Possible) Size of the second allocation in the arena.
-   *  Only relevant if arena strategy is `kNextPowerOfTwo`. Use -1 to allow ORT to choose the default.
-   * "max_power_of_two_extend_bytes": The maximum enxtend size if arena strategy is `kNextPowerOfTwo`.
-   *  It is not an allocation limit, it is only a limit for extention when requested byte is less than the limit.
-   *  When requested bytes is more than the limit, allocator will still return as requested.
-   *  Use -1 to allow ORT to choose the default 1GB for max_power_of_two_extend_bytes.
-   *  Ultimately, the allocation size is determined by the allocation memory request.
-   *  Further allocation sizes are governed by the arena extend strategy.
-   *
-   * \param[in] arena_config_keys Keys to configure the arena
-   * \param[in] arena_config_values Values to configure the arena
-   * \param[in] num_keys Number of keys in `arena_config_keys` and `arena_config_values`
-   * \param[out] out Newly created ::OrtArenaCfg. Must be freed with OrtApi::ReleaseArenaCfg
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateArenaCfgV2, _In_reads_(num_keys) const char* const* arena_config_keys,
-                  _In_reads_(num_keys) const size_t* arena_config_values, _In_ size_t num_keys,
-                  _Outptr_ OrtArenaCfg** out);
-
-  /// @}
-  /// \name OrtRunOptions
-  /// @{
-
-  /** \brief Set a single run configuration entry as a pair of strings
-   *
-   * If a configuration with same key exists, this will overwrite the configuration with the given config_value
-   *
-   * The config_key and the format of config_value are defined in onnxruntime_run_options_config_keys.h
-   *
-   * \param[in] options
-   * \param[in] config_key A null terminated string representation of the config key
-   * \param[in] config_value  A null terminated string representation of the config value
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(AddRunConfigEntry, _Inout_ OrtRunOptions* options,
-                  _In_z_ const char* config_key, _In_z_ const char* config_value);
-
-  /// @}
-  /// \name OrtPrepackedWeightsContainer
-  /// @{
-
-  /** \brief Create an ::OrtPrepackedWeightsContainer
-   *
-   * This container will hold pre-packed buffers of shared initializers for sharing between sessions
-   * (i.e.) if there are shared initializers that can be shared between sessions, the pre-packed buffers
-   * of these (if any) may possibly be shared to provide memory footprint savings. Pass this container
-   * to sessions that you would like to share pre-packed buffers of shared initializers at session
-   * creation time.
-   *
-   *  \param[out] out Newly created ::OrtPrepackedWeightsContainer. Must be freed with OrtApi::ReleasePrepackedWeightsContainer
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreatePrepackedWeightsContainer, _Outptr_ OrtPrepackedWeightsContainer** out);
-
-  /** \brief Release OrtPrepackedWeightsContainer instance
-   *
-   * \note instance must not be released until the sessions using it are released
-   */
-  ORT_CLASS_RELEASE(PrepackedWeightsContainer);
-
-  /// @}
-  /// \name OrtSession
-  /// @{
-
-  /** \brief Create session with prepacked weights container
-   *
-   * Same functionality offered by OrtApi::CreateSession except that a container that contains
-   * pre-packed weights' buffers is written into/read from by the created session.
-   * This is useful when used in conjunction with OrtApi::AddInitializer which injects
-   * shared initializer info into sessions. Wherever possible, the pre-packed versions of these
-   * shared initializers are cached in this container so that multiple sessions can just re-use
-   * these instead of duplicating these in memory.
-   *
-   * \param[in] env OrtEnv instance instance
-   * \param[in] model_path Null terminated string of the path (wchar on Windows, char otherwise)
-   * \param[in] options
-   * \param[in] prepacked_weights_container
-   * \param[out] out Newly created ::OrtSession. Must be freed with OrtApi::ReleaseSession
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateSessionWithPrepackedWeightsContainer, _In_ const OrtEnv* env, _In_ const ORTCHAR_T* model_path,
-                  _In_ const OrtSessionOptions* options, _Inout_ OrtPrepackedWeightsContainer* prepacked_weights_container,
-                  _Outptr_ OrtSession** out);
-
-  /** \brief Create session from memory with prepacked weights container
-   *
-   * Same functionality offered by OrtApi::CreateSessionFromArray except that a container that contains
-   * pre-packed weights' buffers is written into/read from by the created session.
-   * This is useful when used in conjunction with OrtApi::AddInitializer which injects
-   * shared initializer info into sessions. Wherever possible, the pre-packed versions of these
-   * shared initializers are cached in this container so that multiple sessions can just re-use
-   * these instead of duplicating these in memory.
-   *
-   * \param[in] env
-   * \param[in] model_data Array of bytes holding the model
-   * \param[in] model_data_length Number of bytes in `model_data_model`
-   * \param[in] options
-   * \param[in] prepacked_weights_container
-   * \param[out] out Newly created ::OrtSession. Must be freed with OrtApi::ReleaseSession
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateSessionFromArrayWithPrepackedWeightsContainer, _In_ const OrtEnv* env,
-                  _In_ const void* model_data, size_t model_data_length,
-                  _In_ const OrtSessionOptions* options, _Inout_ OrtPrepackedWeightsContainer* prepacked_weights_container,
-                  _Outptr_ OrtSession** out);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Append TensorRT execution provider to the session options
-   *
-   * If TensorRT is not available (due to a non TensorRT enabled build), this function will return failure.
-   *
-   * This is slightly different from OrtApi::SessionOptionsAppendExecutionProvider_TensorRT, it takes an
-   * ::OrtTensorRTProviderOptions which is publicly defined. This takes an opaque ::OrtTensorRTProviderOptionsV2
-   * which must be created with OrtApi::CreateTensorRTProviderOptions.
-   *
-   * For OrtApi::SessionOptionsAppendExecutionProvider_TensorRT, the user needs to instantiate ::OrtTensorRTProviderOptions
-   * as well as allocate/release buffers for some members of ::OrtTensorRTProviderOptions.
-   * Here, OrtApi::CreateTensorRTProviderOptions and Ortapi::ReleaseTensorRTProviderOptions will do the memory management for you.
-   *
-   * \param[in] options
-   * \param[in] tensorrt_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_TensorRT_V2,
-                  _In_ OrtSessionOptions* options, _In_ const OrtTensorRTProviderOptionsV2* tensorrt_options);
-
-  /// @}
-  /// \name OrtTensorRTProviderOptionsV2
-  /// @{
-
-  /** \brief Create an OrtTensorRTProviderOptionsV2
-   *
-   * \param[out] out Newly created ::OrtTensorRTProviderOptionsV2. Must be released with OrtApi::ReleaseTensorRTProviderOptions
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateTensorRTProviderOptions, _Outptr_ OrtTensorRTProviderOptionsV2** out);
-
-  /** \brief Set options in a TensorRT Execution Provider.
-   *
-   * Please refer to https://onnxruntime.ai/docs/execution-providers/TensorRT-ExecutionProvider.html#cc
-   * to know the available keys and values. Key should be in null terminated string format of the member of ::OrtTensorRTProviderOptionsV2
-   * and value should be its related range.
-   *
-   * For example, key="trt_max_workspace_size" and value="2147483648"
-   *
-   * \param[in] tensorrt_options
-   * \param[in] provider_options_keys Array of UTF-8 null-terminated string for provider options keys
-   * \param[in] provider_options_values Array of UTF-8 null-terminated string for provider options values
-   * \param[in] num_keys Number of elements in the `provider_option_keys` and `provider_options_values` arrays
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(UpdateTensorRTProviderOptions, _Inout_ OrtTensorRTProviderOptionsV2* tensorrt_options,
-                  _In_reads_(num_keys) const char* const* provider_options_keys,
-                  _In_reads_(num_keys) const char* const* provider_options_values,
-                  _In_ size_t num_keys);
-
-  /** \brief Get serialized TensorRT provider options string.
-   *
-   * For example, "trt_max_workspace_size=2147483648;trt_max_partition_iterations=10;trt_int8_enable=1;......"
-   *
-   * \param tensorrt_options - OrtTensorRTProviderOptionsV2 instance
-   * \param allocator - a ptr to an instance of OrtAllocator obtained with OrtApi::CreateAllocator or OrtApi::GetAllocatorWithDefaultOptions
-   *                      the specified allocator will be used to allocate continuous buffers for output strings and lengths.
-   * \param ptr - is a UTF-8 null terminated string allocated using 'allocator'. The caller is responsible for using the same allocator to free it.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetTensorRTProviderOptionsAsString, _In_ const OrtTensorRTProviderOptionsV2* tensorrt_options, _Inout_ OrtAllocator* allocator, _Outptr_ char** ptr);
-
-  /** \brief Release an ::OrtTensorRTProviderOptionsV2
-   *
-   * \note This is an exception in the naming convention of other Release* functions, as the name of the method does not have the V2 suffix, but the type does
-   */
-  void(ORT_API_CALL* ReleaseTensorRTProviderOptions)(_Frees_ptr_opt_ OrtTensorRTProviderOptionsV2* input);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Enable custom operators
-   *
-   * See onnxruntime-extensions: https://github.com/microsoft/onnxruntime-extensions.git
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(EnableOrtCustomOps, _Inout_ OrtSessionOptions* options);
-
-  /// @}
-  /// \name OrtAllocator
-  /// @{
-
-  /** \brief Register a custom allocator
-   *
-   * Enables sharing between multiple sessions that use the same env instance.
-   * Returns an error if an allocator with the same ::OrtMemoryInfo is already registered.
-   *
-   * The behavior of this is exactly the same as OrtApi::CreateAndRegisterAllocator except
-   * instead of ORT creating an allocator based on provided info, in this case
-   * ORT uses the user-provided custom allocator.
-   * See https://onnxruntime.ai/docs/get-started/with-c.html for details.
-   *
-   * \param[in] env
-   * \param[in] allocator User provided allocator
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(RegisterAllocator, _Inout_ OrtEnv* env, _In_ OrtAllocator* allocator);
-
-  /** \brief Unregister a custom allocator
-   *
-   * It is an error if you provide an ::OrtMemoryInfo not corresponding to any
-   * registered allocators for sharing.
-   *
-   * \param[in] env
-   * \param[in] mem_info
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(UnregisterAllocator, _Inout_ OrtEnv* env,
-                  _In_ const OrtMemoryInfo* mem_info);
-
-  /// @}
-  /// \name OrtValue
-  /// @{
-
-  /** \brief Sets *out to 1 iff an ::OrtValue is a SparseTensor, and 0 otherwise
-   *
-   * \param[in] value existing ::OrtValue
-   * \param[out] out unless an error occurs, contains 1 iff the value contains an instance
-   *  of sparse tensor or 0 otherwise.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(IsSparseTensor, _In_ const OrtValue* value, _Out_ int* out);
-
-  /** \brief Create an ::OrtValue with a sparse tensor that is empty.
-   *
-   * Use FillSparseTensor<Format>() functions to populate sparse tensor with non-zero values and
-   * format specific indices data.
-   * Use ReleaseValue to destroy the sparse tensor, this will also release the buffer inside the output value
-   * if any was allocated.
-   * \param[in,out] allocator allocator to use when performing an allocation. Allocation will be performed
-   *   by FillSparseTensor<Format>() APIs. The lifespan of the allocator instance must eclipse the lifespan
-   *   this sparse tensor instance as the same allocator will be used to free memory.
-   * \param[in] dense_shape shape of the original dense tensor
-   * \param[in] dense_shape_len number of shape dimensions being passed
-   * \param[in] type must be one of TENSOR_ELEMENT_DATA_TYPE_xxxx
-   * \param[out] out Should be freed by calling ReleaseValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateSparseTensorAsOrtValue, _Inout_ OrtAllocator* allocator, _In_ const int64_t* dense_shape,
-                  size_t dense_shape_len, ONNXTensorElementDataType type, _Outptr_ OrtValue** out);
-
-  /**
-   * This fills populates an empty tensor that was created using OrtApi::CreateSparseTensorAsOrtValue.
-   * This will allocate required memory and copy the supplied NNZ values and COO indices into that memory allocation.
-   * Memory allocation is performed using the allocator that was specified with OrtApi::CreateSparseTensorAsOrtValue.
-   *
-   * \param[in,out] ort_value ::OrtValue to populate with data
-   * \param[in] data_mem_info serves to identify the location of the data to be copied. If the allocator specified
-   *  at the creation time has memory info that is not the same as mem_info argument to this function a X-device copy will be performed.
-   *  String data is assumed to be on CPU and will only be copied into a CPU allocated buffer.
-   * \param[in] values_shape pointer to values shape array
-   * \param[in] values_shape_len length of the values_shape
-   * \param[in] values pointer to an array of values. For strings, pass const char**.
-   * \param[in] indices_data pointer to a location of COO indices
-   * \param[in] indices_num number of COO indices
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(FillSparseTensorCoo, _Inout_ OrtValue* ort_value, _In_ const OrtMemoryInfo* data_mem_info,
-                  _In_ const int64_t* values_shape, size_t values_shape_len, _In_ const void* values,
-                  _In_ const int64_t* indices_data, size_t indices_num);
-
-  /**
-   * This fills populates an empty tensor that was created using OrtApi::CreateSparseTensorAsOrtValue.
-   * This will allocate required memory and copy the supplied NNZ values and CSR indices into that memory allocation.
-   * Memory allocation is performed using the allocator that was specified with OrtApi::CreateSparseTensorAsOrtValue.
-   *
-   * \param[in,out] ort_value ::OrtValue to populate with data
-   * \param[in] data_mem_info serves to identify the location of the data to be copied. If the allocator specified
-   *  at the creation time has memory info that is not the same as mem_info argument to this function a X-device copy will be performed.
-   *  String data is assumed to be on CPU and will only be copied into a CPU allocated buffer.
-   * \param[in] values_shape pointer to values shape array
-   * \param[in] values_shape_len length of the values_shape
-   * \param[in] values - pointer to an array of values. For strings, pass const char**.
-   * \param[in] inner_indices_data pointer to a location of CSR inner indices
-   * \param[in] inner_indices_num number of CSR inner indices
-   * \param[in] outer_indices_data pointer to a location of CSR outer indices
-   * \param[in] outer_indices_num number of CSR outer indices
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(FillSparseTensorCsr, _Inout_ OrtValue* ort_value, _In_ const OrtMemoryInfo* data_mem_info,
-                  _In_ const int64_t* values_shape, size_t values_shape_len, _In_ const void* values,
-                  _In_ const int64_t* inner_indices_data, size_t inner_indices_num,
-                  _In_ const int64_t* outer_indices_data, size_t outer_indices_num);
-
-  /**
-   * This fills populates an empty tensor that was created using OrtApi::CreateSparseTensorAsOrtValue.
-   * This will allocate required memory and copy the supplied NNZ values and BlockSparse indices into that memory allocation.
-   * Memory allocation is performed using the allocator that was specified with OrtApi::CreateSparseTensorAsOrtValue.
-   *
-   * \param[in,out] ort_value ::OrtValue to populate with data
-   * \param[in] data_mem_info serves to identify the location of the data to be copied. If the allocator specified
-   *  at the creation time has memory info that is not the same as mem_info argument to this function a X-device copy will be performed.
-   *  String data is assumed to be on CPU and will only be copied into a CPU allocated buffer.
-   * \param[in] values_shape
-   * \param[in] values_shape_len
-   * \param[in] values structure with values information
-   * \param[in] indices_shape_data pointer to a location of indices shape
-   * \param[in] indices_shape_len length of the block sparse indices shape
-   * \param[in] indices_data pointer to a location of indices data. Shape will determine the length of the indices data.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(FillSparseTensorBlockSparse, _Inout_ OrtValue* ort_value, _In_ const OrtMemoryInfo* data_mem_info,
-                  _In_ const int64_t* values_shape, size_t values_shape_len, _In_ const void* values,
-                  _In_ const int64_t* indices_shape_data, size_t indices_shape_len,
-                  _In_ const int32_t* indices_data);
-
-  /**
-   * Create an ::OrtValue with a sparse tensor. This is the first step.
-   * Next, use Use<Format>Indices() functions to supply sparse tensor with
-   * format specific indices data and set its sparse format to a specific enum value.
-   * This will not perform memory allocations. It will
-   * use supplied user buffer which should outlive the created sparse tensor.
-   * Use OrtApi::ReleaseValue to destroy the sparse tensor. It would not release the supplied values buffer.
-   * This function can not be used to map strings from the user allocated memory. Strings must always be copied
-   * and have UTF-8 encoding. Therefore, use OrtApi::CreateSparseTensorAsOrtValue above and then fill it with data
-   * using appropriate Make*() function.
-   *
-   * \param[in] info memory info where sparse values reside.
-   * \param[in,out] p_data pointer to a user allocated buffer with values. To create a full sparse tensor with no non-zero
-   *   values, pass nullptr
-   * \param[in] dense_shape shape of the original dense tensor
-   * \param[in] dense_shape_len number of shape dimensions being passed
-   * \param[in] values_shape shape of the values data. To create a fully sparse tensor with no non-zero values,
-   *   pass {0} shape.
-   * \param[in] values_shape_len number of values shape dimensions
-   * \param[in] type must be one of TENSOR_ELEMENT_DATA_TYPE_xxxx
-   * \param[out] out Should be freed by calling ReleaseValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(CreateSparseTensorWithValuesAsOrtValue, _In_ const OrtMemoryInfo* info, _Inout_ void* p_data,
-                  _In_ const int64_t* dense_shape, size_t dense_shape_len,
-                  _In_ const int64_t* values_shape, size_t values_shape_len,
-                  ONNXTensorElementDataType type, _Outptr_ OrtValue** out);
-
-  /**
-   * This assigns Coo format indices to the SparseTensor that was created by
-   * OrtApi::CreateSparseTensorWithValuesAsOrtValue above. It also sets OrtSparseFormat to
-   * ORT_SPARSE_COO. This will not allocate any additional memory for data. The life span of
-   * indices_data buffer should eclipse the life span of this ::OrtValue.
-   *
-   * \param[in,out] ort_value ::OrtValue instance constructed with OrtApi::CreateSparseTensorWithValuesAsOrtValue
-   * \param[in,out] indices_data pointer to a user pre-allocated buffer or nullptr for fully sparse tensors.
-   * \param[in] indices_num  number of COO indices. Should either be 0 for fully sparse tensors, be equal
-   *  to the number of nnz values specified to OrtApi::CreateSparseTensorWithValuesAsOrtValue for 1-D {nnz} indices or
-   *  be twice as number of nnz values for a  2-D indices {nnz, 2}
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(UseCooIndices, _Inout_ OrtValue* ort_value, _Inout_ int64_t* indices_data, size_t indices_num);
-
-  /**
-   * The assigns CSR format indices to the SparseTensor that was created by
-   * OrtApi::CreateSparseTensorWithValuesAsOrtValue above. It also sets OrtSparseFormat to
-   * ORT_SPARSE_CSRC. This will not allocate any additional memory for data. The life spans of
-   * inner_data and outer_data buffers should eclipse the life span of this ::OrtValue.
-   *
-   * \param[in,out] ort_value ::OrtValue instance constructed with OrtApi::CreateSparseTensorWithValuesAsOrtValue
-   * \param[in,out] inner_data pointer to a user pre-allocated buffer or nullptr for fully sparse tensors.
-   * \param[in] inner_num  number of inner CSR indices. Should either be 0 for fully sparse tensors or be equal
-   * to the number of nnz values specified to OrtApi::CreateSparseTensorWithValuesAsOrtValue.
-   * \param[in,out] outer_data pointer to user pre-allocated buffer or nullptr for fully sparse tensors.
-   * \param[in] outer_num number of CSR outer indices. Should either be 0 for fully sparse tensors or
-   * equal to rows + 1 of the dense shape.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(UseCsrIndices, _Inout_ OrtValue* ort_value, _Inout_ int64_t* inner_data, size_t inner_num,
-                  _Inout_ int64_t* outer_data, size_t outer_num);
-
-  /**
-   * The assigns BlockSparse format indices to the SparseTensor that was created by
-   * OrtApi::CreateSparseTensorWithValuesAsOrtValue above. It also sets OrtSparseFormat to
-   * ORT_SPARSE_BLOCK_SPARSE. This will not allocate any additional memory for data. The life span of
-   * indices_data buffer must eclipse the lifespan of this ::OrtValue.
-   *
-   * \param[in,out] ort_value OrtValue instance constructed with OrtApi::CreateSparseTensorWithValuesAsOrtValue
-   * \param[in] indices_shape pointer to indices shape. Use {0} for fully sparse tensors
-   * \param[in] indices_shape_len length of the indices shape
-   * \param[in,out] indices_data pointer to user pre-allocated buffer or nullptr for fully sparse tensors.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(UseBlockSparseIndices, _Inout_ OrtValue* ort_value, const int64_t* indices_shape, size_t indices_shape_len, _Inout_ int32_t* indices_data);
-
-  /** \brief Returns sparse tensor format enum iff a given ort value contains an instance of sparse tensor.
-   *
-   * \param[in] ort_value ::OrtValue that contains an instance of sparse tensor
-   * \param[out] out pointer to out parameter
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetSparseTensorFormat, _In_ const OrtValue* ort_value, _Out_ enum OrtSparseFormat* out);
-
-  /** \brief Returns data type and shape of sparse tensor values (nnz) iff ::OrtValue contains a SparseTensor.
-   *
-   * \param[in] ort_value An ::OrtValue that contains a fully constructed sparse tensor
-   * \param[out] out Must be freed by OrtApi::ReleaseTensorTypeAndShapeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetSparseTensorValuesTypeAndShape, _In_ const OrtValue* ort_value, _Outptr_ OrtTensorTypeAndShapeInfo** out);
-
-  /** \brief Returns numeric data for sparse tensor values (nnz). For string values use GetStringTensor*().
-   *
-   * \param[in] ort_value an instance of ::OrtValue containing sparse tensor
-   * \param[out] out returns a pointer to values data.  Do not attempt to free this ptr.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetSparseTensorValues, _In_ const OrtValue* ort_value, _Outptr_ const void** out);
-
-  /** \brief Returns data type, shape for the type of indices specified by indices_format.
-   *
-   * \param[in] ort_value ::OrtValue containing sparse tensor.
-   * \param[in] indices_format One of the indices formats. It is an error to request a format that the sparse
-   * tensor does not contain.
-   * \param[out] out an instance of ::OrtTensorTypeAndShapeInfo. Must be freed by OrtApi::ReleaseTensorTypeAndShapeInfo
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetSparseTensorIndicesTypeShape, _In_ const OrtValue* ort_value, enum OrtSparseIndicesFormat indices_format, _Outptr_ OrtTensorTypeAndShapeInfo** out);
-
-  /** \brief Returns indices data for the type of the indices specified by indices_format
-   *
-   * \param[in] ort_value ::OrtValue containing sparse tensor.
-   * \param[in] indices_format One of the indices formats. It is an error to request a format that the sparse tensor does not contain.
-   * \param[out] num_indices Pointer to where the number of indices entries is returned
-   * \param[out] indices Returned pointer to the indices data. Do not free the returned pointer as it refers to internal data owned by the ::OrtValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetSparseTensorIndices, _In_ const OrtValue* ort_value, enum OrtSparseIndicesFormat indices_format, _Out_ size_t* num_indices, _Outptr_ const void** indices);
-  /// @}
-  /// \name OrtSessionOptions
-  /// @{
-
-  /**
-   * \brief Sets out to 1 iff an optional type OrtValue has an element, 0 otherwise (OrtValue is None)
-   * Use this API to find if the optional type OrtValue is None or not.
-   * If the optional type OrtValue is not None, use the OrtValue just like any other OrtValue.
-   * For example, if you get an OrtValue that corresponds to Optional(tensor) and
-   * if HasValue() returns true, use it as tensor and so on.
-
-   * \param[in] value Input OrtValue.
-   * \param[out] out indicating if the input OrtValue contains data (1) or if it is a None (0)
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(HasValue, _In_ const OrtValue* value, _Out_ int* out);
-
-  /// @}
-  /// \name OrtKernelContext
-  /// Custom operator APIs.
-  /// @{
-
-  /** \brief Used for custom operators, gets the GPU compute stream to use to launch the custom a GPU kernel
-   *   \see ::OrtCustomOp
-   * \param[in]  context OrtKernelContext instance
-   * \param[out] out Returns pointer to a GPU compute stream that can be used to launch the custom GPU kernel.
-   *             If retrieving the GPU compute stream is not relevant (GPU not enabled in the build, kernel partitioned to
-   *             some other EP), then a nullptr is returned as the output param.
-   *             Do not free or mutate the returned pointer as it refers to internal data owned by the underlying session.
-   *             Only use it for custom kernel launching.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(KernelContext_GetGPUComputeStream, _In_ const OrtKernelContext* context, _Outptr_ void** out);
-
-  /// @}
-  /// \name GetTensorMemoryInfo
-  /// @{
-  /** \brief Returns a pointer to the ::OrtMemoryInfo of a Tensor
-   * \param[in] value ::OrtValue containing tensor.
-   * \param[out] mem_info ::OrtMemoryInfo of the tensor. Do NOT free the returned pointer. It is valid for the lifetime of the ::OrtValue
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetTensorMemoryInfo, _In_ const OrtValue* value, _Out_ const OrtMemoryInfo** mem_info);
-
-  /// @}
-  /// \name GetExecutionProviderApi
-  /// @{
-  /** \brief Get a pointer to the requested version of the Execution Provider specific
-   * API extensions to the OrtApi
-   * \param[in] provider_name The name of the execution provider name. Currently only the following
-   * values are supported: "DML".
-   * \param[in] version Must be ::ORT_API_VERSION.
-   * \param[out] provider_api A void pointer containing a reference to the execution provider versioned api structure.
-   * For example, the provider_api pointer can be cast to the OrtDmlApi* when the provider_name is "DML".
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetExecutionProviderApi, _In_ const char* provider_name, _In_ uint32_t version, _Outptr_ const void** provider_api);
-
-  /// @}
-
-  /// \name SessionOptions
-  /// @{
-  /** \brief Set custom thread creation function
-   *
-   * \param[in] options Session options
-   * \param[in] ort_custom_create_thread_fn Custom thread creation function
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsSetCustomCreateThreadFn, _Inout_ OrtSessionOptions* options, _In_ OrtCustomCreateThreadFn ort_custom_create_thread_fn);
-
-  /** \brief Set creation options for custom thread
-   *
-   * \param[in] options Session options
-   * \param[in] ort_custom_thread_creation_options Custom thread creation options (can be nullptr)
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsSetCustomThreadCreationOptions, _Inout_ OrtSessionOptions* options, _In_ void* ort_custom_thread_creation_options);
-
-  /** \brief Set custom thread join function
-   *
-   * \param[in] options Session options
-   * \param[in] ort_custom_join_thread_fn Custom join thread function, must not be nullptr when ort_custom_create_thread_fn is set
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsSetCustomJoinThreadFn, _Inout_ OrtSessionOptions* options, _In_ OrtCustomJoinThreadFn ort_custom_join_thread_fn);
-  /// @}
-
-  /// \name OrtThreadingOptions
-  /// @{
-  /** \brief Set custom thread creation function for global thread pools
-   *
-   * \param[inout] tp_options
-   * \param[in] ort_custom_create_thread_fn Custom thread creation function
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetGlobalCustomCreateThreadFn, _Inout_ OrtThreadingOptions* tp_options, _In_ OrtCustomCreateThreadFn ort_custom_create_thread_fn);
-
-  /** \brief Set custom thread creation options for global thread pools
-   *
-   * \param[inout] tp_options
-   * \param[in] ort_custom_thread_creation_options Custom thread creation options (can be nullptr)
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetGlobalCustomThreadCreationOptions, _Inout_ OrtThreadingOptions* tp_options, _In_ void* ort_custom_thread_creation_options);
-
-  /** \brief Set custom thread join function for global thread pools
-   *
-   * \param[inout] tp_options
-   * \param[in] ort_custom_join_thread_fn Custom thread join function, must not be nullptr when global ort_custom_create_thread_fn is set
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SetGlobalCustomJoinThreadFn, _Inout_ OrtThreadingOptions* tp_options, _In_ OrtCustomJoinThreadFn ort_custom_join_thread_fn);
-  /// @}
-
-  /** \brief Synchronize bound inputs. The call may be necessary for some providers, such as cuda,
-   *   in case the system that allocated bound memory operated on a different stream. However, the
-   *   operation is provider specific and could be a no-op.
-   *
-   * \param[inout] binding_ptr
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SynchronizeBoundInputs, _Inout_ OrtIoBinding* binding_ptr);
-
-  /** \brief Synchronize bound outputs. The call may be necessary for some providers, such as cuda,
-   *   in case the system that allocated bound memory operated on a different stream. However, the
-   *   operation is provider specific and could be a no-op.
-   *
-   * \param[inout] binding_ptr
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SynchronizeBoundOutputs, _Inout_ OrtIoBinding* binding_ptr);
-
-  /// \name OrtSessionOptions
-  /// @{
-
-  /** \brief Append CUDA execution provider to the session options
-   *
-   * If CUDA is not available (due to a non CUDA enabled build), this function will return failure.
-   *
-   * This is slightly different from OrtApi::SessionOptionsAppendExecutionProvider_CUDA, it takes an
-   * ::OrtCUDAProviderOptions which is publicly defined. This takes an opaque ::OrtCUDAProviderOptionsV2
-   * which must be created with OrtApi::CreateCUDAProviderOptions.
-   *
-   * For OrtApi::SessionOptionsAppendExecutionProvider_CUDA, the user needs to instantiate ::OrtCUDAProviderOptions
-   * as well as allocate/release buffers for some members of ::OrtCUDAProviderOptions.
-   * Here, OrtApi::CreateCUDAProviderOptions and Ortapi::ReleaseCUDAProviderOptions will do the memory management for you.
-   *
-   * \param[in] options
-   * \param[in] cuda_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.11.
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_CUDA_V2,
-                  _In_ OrtSessionOptions* options, _In_ const OrtCUDAProviderOptionsV2* cuda_options);
-
-  /// @}
-  /// \name OrtCUDAProviderOptionsV2
-  /// @{
-
-  /** \brief Create an OrtCUDAProviderOptionsV2
-   *
-   * \param[out] out Newly created ::OrtCUDAProviderOptionsV2. Must be released with OrtApi::ReleaseCudaProviderOptions
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.11.
-   */
-  ORT_API2_STATUS(CreateCUDAProviderOptions, _Outptr_ OrtCUDAProviderOptionsV2** out);
-
-  /** \brief Set options in a CUDA Execution Provider.
-   *
-   * Please refer to https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html#configuration-options
-   * to know the available keys and values. Key should be in null terminated string format of the member of ::OrtCUDAProviderOptionsV2
-   * and value should be its related range.
-   *
-   * For example, key="device_id" and value="0"
-   *
-   * \param[in] cuda_options
-   * \param[in] provider_options_keys Array of UTF-8 null-terminated string for provider options keys
-   * \param[in] provider_options_values Array of UTF-8 null-terminated string for provider options values
-   * \param[in] num_keys Number of elements in the `provider_option_keys` and `provider_options_values` arrays
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.11.
-   */
-  ORT_API2_STATUS(UpdateCUDAProviderOptions, _Inout_ OrtCUDAProviderOptionsV2* cuda_options,
-                  _In_reads_(num_keys) const char* const* provider_options_keys,
-                  _In_reads_(num_keys) const char* const* provider_options_values,
-                  _In_ size_t num_keys);
-
-  /**
-   * Get serialized CUDA provider options string.
-   *
-   * For example, "device_id=0;arena_extend_strategy=0;......"
-   *
-   * \param cuda_options - OrtCUDAProviderOptionsV2 instance
-   * \param allocator - a ptr to an instance of OrtAllocator obtained with CreateAllocator() or GetAllocatorWithDefaultOptions()
-   *                      the specified allocator will be used to allocate continuous buffers for output strings and lengths.
-   * \param ptr - is a UTF-8 null terminated string allocated using 'allocator'. The caller is responsible for using the same allocator to free it.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.11.
-   */
-  ORT_API2_STATUS(GetCUDAProviderOptionsAsString, _In_ const OrtCUDAProviderOptionsV2* cuda_options, _Inout_ OrtAllocator* allocator, _Outptr_ char** ptr);
-
-  /** \brief Release an ::OrtCUDAProviderOptionsV2
-   *
-   * \note This is an exception in the naming convention of other Release* functions, as the name of the method does not have the V2 suffix, but the type does
-   *
-   * \since Version 1.11.
-   */
-  void(ORT_API_CALL* ReleaseCUDAProviderOptions)(_Frees_ptr_opt_ OrtCUDAProviderOptionsV2* input);
-
-  /// @}
-
-  /** \brief Append MIGraphX provider to session options
-   *
-   * If MIGraphX is not available (due to a non MIGraphX enabled build, or if MIGraphX is not installed on the system), this function will return failure.
-   *
-   * \param[in] options
-   * \param[in] migraphx_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.11.
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_MIGraphX,
-                  _In_ OrtSessionOptions* options, _In_ const OrtMIGraphXProviderOptions* migraphx_options);
-
-  /** \brief Replace initialized Tensors with external data with the data provided in initializers.
-   *
-   * The function will find the initialized TensorProtos with external data in the graph with the provided names and
-   * replace them with the provided tensors. The API verifies that the TensorProto being replaced
-   * has an external data reference and has the same name, dimensions and data type as its replacement. The replacement
-   * will occur before any of the optimizations take place. The data will be copied into the graph
-   * since TensorProto can't refer to the user provided buffers.
-   *
-   * Once the model has been loaded, the OrtValue(s) added to SessionOptions instance will be removed
-   * from the internal SessionOptions copy to save memory, the user provided buffers can then be deallocated
-   * and the SessionOptions instance that refers to them can be destroyed.
-   *
-   * \param[in] options
-   * \param[in] initializer_names Array of null terminated UTF-8 encoded strings of the initializers names.
-   * \param[in] initializers Array of ::OrtValue type
-   * \param[in] initializers_num Number of elements in the initializer_names and initializers
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.12.
-   */
-  ORT_API2_STATUS(AddExternalInitializers, _In_ OrtSessionOptions* options,
-                  _In_reads_(input_len) const char* const* initializer_names,
-                  _In_reads_(input_len) const OrtValue* const* initializers, size_t initializers_num);
-
-  /** \brief: Create attribute of onnxruntime operator
-   *
-   * \param[in] name Name of the attribute
-   * \param[in] data Data content of the attribute
-   * \param[in] len Number of bytes stored in data
-   * \param[in] type Data type
-   * \param[out] op_attr Attribute that has been created, which must be released by OrtApi::ReleaseOpAttr
-   *
-   * \since Version 1.12.
-   */
-  ORT_API2_STATUS(CreateOpAttr,
-                  _In_ const char* name,
-                  _In_ const void* data,
-                  _In_ int len,
-                  _In_ OrtOpAttrType type,
-                  _Outptr_ OrtOpAttr** op_attr);
-
-  /* \brief: Release op attribute
-   *
-   * \param[in] opAttr Attribute created by OrtApi::CreateOpAttr
-   *
-   * \since Version 1.12.
-   */
-  ORT_CLASS_RELEASE(OpAttr);
-
-  /** \brief: Create onnxruntime native operator
-   *
-   * \param[in] info Kernel info
-   * \param[in] op_name Operator name
-   * \param[in] domain Operator domain
-   * \param[in] version Operator opset version
-   * \param[in] type_constraint_names Name of the type contraints, such as "T" or "T1"
-   * \param[in] type_constraint_values Type of each contraints
-   * \param[in] type_constraint_count Number of contraints
-   * \param[in] attr_values Attributes used to initialize the operator
-   * \param[in] attr_count Number of the attributes
-   * \param[in] input_count Number of inputs
-   * \param[in] output_count Number of outputs
-   * \param[out] ort_op Operator that has been created
-   *
-   * \since Version 1.12.
-   */
-  ORT_API2_STATUS(CreateOp,
-                  _In_ const OrtKernelInfo* info,
-                  _In_z_ const char* op_name,
-                  _In_z_ const char* domain,
-                  int version,
-                  _In_reads_(type_constraint_count) const char** type_constraint_names,
-                  _In_reads_(type_constraint_count) const ONNXTensorElementDataType* type_constraint_values,
-                  int type_constraint_count,
-                  _In_reads_(attr_count) const OrtOpAttr* const* attr_values,
-                  int attr_count,
-                  int input_count,
-                  int output_count,
-                  _Outptr_ OrtOp** ort_op);
-
-  /** \brief: Invoke the operator created by OrtApi::CreateOp
-   * The inputs must follow the order as specified in onnx specification
-   *
-   * \param[in] context Kernel context
-   * \param[in] ort_op Operator that has been created
-   * \param[in] input_values Array of inputs
-   * \param[in] input_count Number of inputs
-   * \param[in] output_values Array of outputs
-   * \param[in] output_count Number of outputs
-   *
-   * \since Version 1.12.
-   */
-  ORT_API2_STATUS(InvokeOp,
-                  _In_ const OrtKernelContext* context,
-                  _In_ const OrtOp* ort_op,
-                  _In_ const OrtValue* const* input_values,
-                  _In_ int input_count,
-                  _Inout_ OrtValue* const* output_values,
-                  _In_ int output_count);
-
-  /* \brief: Release an onnxruntime operator
-   *
-   * \param[in] Op Operator created by OrtApi::CreateOp
-   *
-   * \since Version 1.12.
-   */
-  ORT_CLASS_RELEASE(Op);
-
-  /** \brief: Append execution provider to the session options.
-   * \param[in] options
-   * \param[in] provider_name - provider to add.
-   * \param[in] provider_options_keys - keys to configure the provider options
-   * \param[in] provider_options_values - values to configure the provider options
-   * \param[in] num_keys - number of keys passed in
-   *
-   * Currently supported providers:
-   *   QNN
-   *   SNPE
-   *   XNNPACK
-   *
-   * Note: If an execution provider has a dedicated SessionOptionsAppendExecutionProvider_<provider name> function
-   *       that should be used to add it.
-   *
-   * QNN supported keys:
-   *   "backend_path": file path to QNN backend library.
-   *   "profiling_level": QNN profiling level, options: "off", "basic", "detailed". Default to off.
-   *   "rpc_control_latency": QNN RPC control latency.
-   *   "vtcm_mb": QNN VTCM size in MB. default to 0(not set).
-   *   "htp_performance_mode": QNN performance mode, options: "burst", "balanced", "default", "high_performance",
-   *   "high_power_saver", "low_balanced", "extreme_power_saver", "low_power_saver", "power_saver", "sustained_high_performance". Default to "default".
-   *   "qnn_saver_path": File path to the QNN Saver backend library. If specified, QNN Saver will be enabled and will
-   *   dump QNN API calls to disk for replay/debugging. QNN Saver produces incorrect model inference results and
-   *   may alter model/EP partitioning. Use only for debugging.
-   *   "qnn_context_priority": QNN context priority, options: "low", "normal", "normal_high", "high". Default to "normal".
-   *   "htp_graph_finalization_optimization_mode": Set the optimization mode for graph finalization on the HTP backend. Available options:
-   *     - "0": Default.
-   *     - "1": Faster preparation time, less optimal graph.
-   *     - "2": Longer preparation time, more optimal graph.
-   *     - "3": Longest preparation time, most likely even more optimal graph. See QNN SDK documentation for specific details.
-   *   "soc_model": The SoC model number. Refer to the QNN SDK documentation for valid values. Defaults to "0" (unknown).
-   *   "htp_arch": The minimum HTP architecture the driver will use to select compatible QNN operators. Available options:
-   *     - "0": Default (none).
-   *     - "68"
-   *     - "69"
-   *     - "73"
-   *     - "75"
-   *   "device_id": The ID of the device to use when setting 'htp_arch'. Defaults to "0" (for single device).
-   *
-   * SNPE supported keys:
-   *   "runtime": SNPE runtime engine, options: "CPU", "CPU_FLOAT32", "GPU", "GPU_FLOAT32_16_HYBRID", "GPU_FLOAT16",
-   *   "DSP", "DSP_FIXED8_TF", "AIP_FIXED_TF", "AIP_FIXED8_TF".
-   *   Mapping to SNPE Runtime_t definition: CPU, CPU_FLOAT32 => zdl::DlSystem::Runtime_t::CPU;
-   *   GPU, GPU_FLOAT32_16_HYBRID => zdl::DlSystem::Runtime_t::GPU;
-   *   GPU_FLOAT16 => zdl::DlSystem::Runtime_t::GPU_FLOAT16;
-   *   DSP, DSP_FIXED8_TF => zdl::DlSystem::Runtime_t::DSP.
-   *   AIP_FIXED_TF, AIP_FIXED8_TF => zdl::DlSystem::Runtime_t::AIP_FIXED_TF.
-   *   "priority": execution priority, options: "low", "normal".
-   *   "buffer_type": ITensor or user buffers, options: "ITENSOR", user buffer with different types - "TF8", "TF16", "UINT8", "FLOAT".
-   *   "ITENSOR" -- default, ITensor which is float only.
-   *   "TF8" -- quantized model required, "FLOAT" -- for both quantized or non-quantized model
-   *   "enable_init_cache": enable SNPE init caching feature, set to 1 to enabled it. Disabled by default.
-   *   If SNPE is not available (due to a non Snpe enabled build or its dependencies not being installed), this function will fail.
-   *
-   * XNNPACK supported keys:
-   *   "intra_op_num_threads": number of thread-pool size to use for XNNPACK execution provider.
-   *      default value is 0, which means to use the session thread-pool size.
-   *
-   * \since Version 1.12.
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider, _In_ OrtSessionOptions* options,
-                  _In_ const char* provider_name,
-                  _In_reads_(num_keys) const char* const* provider_options_keys,
-                  _In_reads_(num_keys) const char* const* provider_options_values,
-                  _In_ size_t num_keys);
-
-  /* \brief: Get a copy of kernel info
-   *
-   * \param[in] info Kernel info
-   * \param[out] info_copy Copy of kernel info
-   *
-   * \since Version 1.12.
-   */
-  ORT_API2_STATUS(CopyKernelInfo,
-                  _In_ const OrtKernelInfo* info,
-                  _Outptr_ OrtKernelInfo** info_copy);
-
-  /* \brief: Release kernel info
-   *
-   * \param[in] KernelInfo A copy of kernel info returned by CopyKernelInfo
-   *
-   * \since Version 1.12.
-   */
-  ORT_CLASS_RELEASE(KernelInfo);
-
-  /// \name Ort Training
-  /// @{
-  /** \brief Gets the Training C Api struct
-   *
-   * Call this function to access the ::OrtTrainingApi structure that holds pointers to functions that enable
-   * training with onnxruntime.
-   * \note A NULL pointer will be returned and no error message will be printed if the training api
-   * is not supported with this build. A NULL pointer will be returned and an error message will be
-   * printed if the provided version is unsupported, for example when using a runtime older than the
-   * version created with this header file.
-   *
-   * \param[in] version Must be ::ORT_API_VERSION
-   * \return The ::OrtTrainingApi struct for the version requested.
-   *
-   * \since Version 1.13
-   */
-  const OrtTrainingApi*(ORT_API_CALL* GetTrainingApi)(uint32_t version)NO_EXCEPTION;
-
-  /// @}
-
-  /** \brief Append CANN provider to session options
-   *
-   * If CANN is not available (due to a non CANN enabled build, or if CANN is not installed on the system), this function will return failure.
-   *
-   * \param[in] options
-   * \param[in] cann_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.13.
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_CANN,
-                  _In_ OrtSessionOptions* options, _In_ const OrtCANNProviderOptions* cann_options);
-
-  /** \brief Create an OrtCANNProviderOptions
-   *
-   * \param[out] out created ::OrtCANNProviderOptions. Must be released with OrtApi::ReleaseCANNProviderOptions
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.13.
-   */
-  ORT_API2_STATUS(CreateCANNProviderOptions, _Outptr_ OrtCANNProviderOptions** out);
-
-  /** \brief Set options in a CANN Execution Provider.
-   *
-   * \param[in] cann_options
-   * \param[in] provider_options_keys Array of UTF-8 null-terminated string for provider options keys
-   * \param[in] provider_options_values Array of UTF-8 null-terminated string for provider options values
-   * \param[in] num_keys Number of elements in the `provider_option_keys` and `provider_options_values` arrays
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.13.
-   */
-  ORT_API2_STATUS(UpdateCANNProviderOptions, _Inout_ OrtCANNProviderOptions* cann_options,
-                  _In_reads_(num_keys) const char* const* provider_options_keys,
-                  _In_reads_(num_keys) const char* const* provider_options_values,
-                  _In_ size_t num_keys);
-
-  /** \brief Get serialized CANN provider options string.
-   *
-   * \param[in] cann_options OrtCANNProviderOptions instance
-   * \param[in] allocator a ptr to an instance of OrtAllocator obtained with CreateAllocator()
-   *                      or GetAllocatorWithDefaultOptions(), the specified allocator will be used to allocate
-   *                      continuous buffers for output strings and lengths.
-   * \param[out] ptr is a UTF-8 null terminated string allocated using 'allocator'.
-   *                 The caller is responsible for using the same allocator to free it.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.13.
-   */
-  ORT_API2_STATUS(GetCANNProviderOptionsAsString, _In_ const OrtCANNProviderOptions* cann_options,
-                  _Inout_ OrtAllocator* allocator, _Outptr_ char** ptr);
-
-  /** \brief Release an OrtCANNProviderOptions
-   *
-   * \param[in] the pointer of OrtCANNProviderOptions which will been deleted
-   *
-   * \since Version 1.13.
-   */
-  void(ORT_API_CALL* ReleaseCANNProviderOptions)(_Frees_ptr_opt_ OrtCANNProviderOptions* input);
-
-  /*  \brief Get OrtDevice type from MemoryInfo
-   *
-   *  \since Version 1.14
-   */
-  void(ORT_API_CALL* MemoryInfoGetDeviceType)(_In_ const OrtMemoryInfo* ptr, _Out_ OrtMemoryInfoDeviceType* out);
-
-  /* \brief Update the OrtEnv instance with custom log severity level
-   *
-   * \param[in] ort_env The OrtEnv instance being used
-   * \param[in] log_severity_level The log severity level.
-   *
-   * \since Version 1.14.
-   */
-  ORT_API2_STATUS(UpdateEnvWithCustomLogLevel, _In_ OrtEnv* ort_env, OrtLoggingLevel log_severity_level);
-
-  /*  \brief Set affinities for intra op threads
-   *
-   * Affinity string follows format:
-   * logical_processor_id,logical_processor_id;logical_processor_id,logical_processor_id
-   * Semicolon isolates configurations among threads, while comma split processors where ith thread expected to attach to.
-   * e.g. 1,2,3;4,5
-   * specifies affinities for two threads, with the 1st thread attach to the 1st, 2nd, and 3rd processor, and 2nd thread to the 4th and 5th.
-   * To ease the configuration, an "interval" is also allowed:
-   * e.g. 1-8;8-16;17-24
-   * orders that the 1st thread runs on first eight processors, 2nd thread runs on next eight processors, and so forth.
-   * Note:
-   * 1. Once set, the number of thread affinities must equal to intra_op_num_threads - 1,
-   *    ort does not set affinity on the main thread which is started and managed by the calling app;
-   * 2. For windows, ort will infer the group id from a logical processor id, for example, assuming there are two groups with each has 64 logical processors,
-   *    an id of 64 will be inferred as the last processor of the 1st group, while 65 will be interpreted as the 1st processor of the second group.
-   *    Hence 64-65 is an invalid configuration, because a windows thread cannot be attached to processors across group boundary.
-   *
-   *  \since Version 1.14
-   */
-  ORT_API2_STATUS(SetGlobalIntraOpThreadAffinity, _Inout_ OrtThreadingOptions* tp_options, const char* affinity_string);
-
-  /** \brief Register custom ops from a shared library.
-   *
-   * Loads a shared library (.dll on windows, .so on linux, etc) named 'library_name' and looks for this entry point:
-   *		OrtStatus* RegisterCustomOps(OrtSessionOptions * options, const OrtApiBase* api);
-   * It then passes in the provided session options to this function along with the api base.
-   *
-   * The handle to the loaded library is automatically released by ORT when the last OrtSession that references the
-   * library handle is released. If no OrtSession is created, then the library handle is released when the provided
-   * OrtSessionOptions is released.
-   *
-   * \param[in] options The session options.
-   * \param[in] library_name The name of the shared library to load and register. Refer to OS-specific dynamic library
-   *                         loading utilities (e.g., LoadLibraryEx on Windows or dlopen on Linux/MacOS) for information
-   *                         on the format of library names and search paths.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(RegisterCustomOpsLibrary_V2, _Inout_ OrtSessionOptions* options, _In_ const ORTCHAR_T* library_name);
-
-  /** \brief Register custom ops by calling a RegisterCustomOpsFn function.
-   *
-   * Searches for registration_func_name and if found calls it.
-   *
-   * The library containing the function must either be linked against or previously loaded by the executable.
-   *
-   * If you want ONNX Runtime to load the library and manage its lifetime, use RegisterCustomOpsLibrary_V2.
-   *
-   * RegisterCustomOpsUsingFunction can be used in scenarios where it may not be possible for ONNX Runtime to load
-   * the library from a path. e.g. mobile platforms where the library must be linked into the app.
-   *
-   * The registration function must have the signature of RegisterCustomOpsFn:
-   *    OrtStatus* (*fn)(OrtSessionOptions* options, const OrtApiBase* api);
-   *
-   * See https://onnxruntime.ai/docs/reference/operators/add-custom-op.html for details on how the registration
-   * function should be implemented.
-   *
-   * \param[in] options OrtSessionOptions that is passed through as the first argument in the call to the
-   *                    registration function.
-   * \param[in] registration_func_name Name of registration function to use.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(RegisterCustomOpsUsingFunction, _Inout_ OrtSessionOptions* options,
-                  _In_ const char* registration_func_name);
-
-  /// \name OrtKernelInfo
-  /// Custom operator APIs.
-  /// @{
-
-  /** \brief Get the number of inputs from ::OrtKernelInfo.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to query the number of inputs
-   * during kernel/session creation.
-   *
-   * \param[in] info Instance of ::OrtKernelInfo.
-   * \param[out] out Pointer to variable assigned with the result on success.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(KernelInfo_GetInputCount, _In_ const OrtKernelInfo* info, _Out_ size_t* out);
-
-  /** \brief Get the number of outputs from ::OrtKernelInfo.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to query the number of outputs
-   * during kernel/session creation.
-   *
-   * \param[in] info Instance of ::OrtKernelInfo.
-   * \param[out] out Pointer to variable assigned with the result on success.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(KernelInfo_GetOutputCount, _In_ const OrtKernelInfo* info, _Out_ size_t* out);
-
-  /** \brief Get the name of a ::OrtKernelInfo's input.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to query an input's name
-   * during kernel/session creation.
-   *
-   * If `out` is nullptr, the value of `size` is set to the size of the name
-   * string (including null-terminator), and a success status is returned.
-   *
-   * If the `size` parameter is greater than or equal to the name string's size,
-   * the value of `size` is set to the true size of the string (including null-terminator),
-   * the provided memory is filled with the string's contents, and a success status is returned.
-   *
-   * If the `size` parameter is less than the actual string's size and `out`
-   * is not nullptr, the value of `size` is set to the true size of the string
-   * and a failure status is returned.
-   *
-   * \param[in] info An instance of ::OrtKernelInfo.
-   * \param[in] index The index of the input name to get. Returns a failure status if out-of-bounds.
-   * \param[out] out Memory location into which to write the UTF-8 null-terminated string representing the input's name.
-   * \param[in,out] size Pointer to the size of the `out` buffer. See above comments for details.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(KernelInfo_GetInputName, _In_ const OrtKernelInfo* info, size_t index, _Out_ char* out,
-                  _Inout_ size_t* size);
-
-  /** \brief Get the name of a ::OrtKernelInfo's output.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to query an output's name
-   * during kernel/session creation.
-   *
-   * If `out` is nullptr, the value of `size` is set to the size of the name
-   * string (including null-terminator), and a success status is returned.
-   *
-   * If the `size` parameter is greater than or equal to the name string's size,
-   * the value of `size` is set to the true size of the string (including null-terminator),
-   * the provided memory is filled with the string's contents, and a success status is returned.
-   *
-   * If the `size` parameter is less than the actual string's size and `out`
-   * is not nullptr, the value of `size` is set to the true size of the string
-   * and a failure status is returned.
-   *
-   * \param[in] info An instance of ::OrtKernelInfo.
-   * \param[in] index The index of the output name to get. Returns a failure status if out-of-bounds.
-   * \param[out] out Memory location into which to write the UTF-8 null-terminated string representing the output's
-   *                 name.
-   * \param[in,out] size Pointer to the size of the `out` buffer. See above comments for details.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(KernelInfo_GetOutputName, _In_ const OrtKernelInfo* info, size_t index, _Out_ char* out,
-                  _Inout_ size_t* size);
-
-  /** \brief Get the type information for a ::OrtKernelInfo's input.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to query the shape and type information
-   * of an input during kernel/session creation.
-   *
-   * \param[in] info An instance of ::OrtKernelInfo.
-   * \param[in] index Which input to get the type information for
-   * \param[out] type_info Pointer set to the resulting ::OrtTypeInfo. Must be freed with OrtApi::ReleaseTypeInfo.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(KernelInfo_GetInputTypeInfo, _In_ const OrtKernelInfo* info, size_t index,
-                  _Outptr_ OrtTypeInfo** type_info);
-
-  /** \brief Get the type information for a ::OrtKernelInfo's output.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to query the shape and type information
-   * of an output during kernel/session creation.
-   *
-   * \param[in] info An instance of ::OrtKernelInfo.
-   * \param[in] index Which input to get the type information for
-   * \param[out] type_info Pointer set to the resulting ::OrtTypeInfo. Must be freed with OrtApi::ReleaseTypeInfo.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(KernelInfo_GetOutputTypeInfo, _In_ const OrtKernelInfo* info, size_t index,
-                  _Outptr_ OrtTypeInfo** type_info);
-
-  /** \brief Get a ::OrtValue tensor stored as an attribute in the graph node.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to get a tensor attribute.
-   *
-   * \param[in] info ::OrtKernelInfo instance.
-   * \param[in] name UTF-8 null-terminated string representing the attribute's name.
-   * \param[in] allocator Allocator used to allocate the internal tensor state.
-   * \param[out] out Returns newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue,
-   *                 which will also free internal tensor state allocated with the provided allocator.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(KernelInfoGetAttribute_tensor, _In_ const OrtKernelInfo* info, _In_z_ const char* name,
-                  _Inout_ OrtAllocator* allocator, _Outptr_ OrtValue** out);
-
-  /// @}
-  /// \name OrtSessionOptions
-  /// Custom operator APIs
-  /// @{
-
-  /** \brief Checks if the given session configuration entry exists.
-   *
-   * The config_key formats are defined in onnxruntime_session_options_config_keys.h
-   *
-   * Can be used in a custom operator library to check for session configuration entries
-   * that target one or more custom operators in the library. Example: The config entry
-   * custom_op.myop.some_key targets a custom op named "myop".
-   *
-   * \param[in] options The ::OrtSessionOptions instance.
-   * \param[in] config_key A null-terminated UTF-8 string representation of the configuration key.
-   * \param[out] out Pointer set to 1 if the entry exists and 0 otherwise.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(HasSessionConfigEntry, _In_ const OrtSessionOptions* options,
-                  _In_z_ const char* config_key, _Out_ int* out);
-
-  /** \brief Get a session configuration value.
-   *
-   * Returns a failure status if the configuration key does not exist.
-   * The config_key and the format of config_value are defined in onnxruntime_session_options_config_keys.h
-   *
-   * If `config_value` is nullptr, the value of `size` is set to the true size of the string
-   * value (including null-terminator), and a success status is returned.
-   *
-   * If the `size` parameter is greater than or equal to the actual string value's size,
-   * the value of `size` is set to the true size of the string value, the provided memory
-   * is filled with the value's contents, and a success status is returned.
-   *
-   * If the `size` parameter is less than the actual string value's size and `config_value`
-   * is not nullptr, the value of `size` is set to the true size of the string value
-   * and a failure status is returned.
-   *
-   * Can be used in a custom operator library to get session configuration entries
-   * that target one or more custom operators in the library. Example: The config entry
-   * custom_op.myop.some_key targets a custom op named "myop".
-   *
-   * \param[in] options The session options.
-   * \param[in] config_key A null-terminated UTF-8 string representation of the config key.
-   * \param[in] config_value Pointer to memory where the null-terminated UTF-8 string value will be stored.
-   * \param[in,out] size Pointer to the size of the `config_value` buffer. See above comments for details.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.14
-   */
-  ORT_API2_STATUS(GetSessionConfigEntry, _In_ const OrtSessionOptions* options,
-                  _In_z_ const char* config_key, _Out_ char* config_value, _Inout_ size_t* size);
-
-  /// @}
-
-  /** \brief Append dnnl provider to session options
-   *
-   * If oneDNN is not available, this function will return failure.
-   *
-   * \param[in] options
-   * \param[in] dnnl_options
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.15.
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_Dnnl,
-                  _In_ OrtSessionOptions* options, _In_ const OrtDnnlProviderOptions* dnnl_options);
-
-  /** \brief Create an OrtDnnlProviderOptions
-   *
-   * \param[out] out Newly created ::OrtDnnlProviderOptions. Must be released with OrtApi::ReleaseDnnlProviderOptions
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.15.
-   */
-  ORT_API2_STATUS(CreateDnnlProviderOptions, _Outptr_ OrtDnnlProviderOptions** out);
-
-  /** \brief Set options in a oneDNN Execution Provider.
-   *
-   * Key should be in null terminated string format of the member of ::OrtDnnlProviderOptions
-   * and value should be its related range.
-   *
-   * For example, key="use_arena" and value="1"
-   *
-   * \param[in] dnnl_options
-   * \param[in] provider_options_keys Array of UTF-8 null-terminated string for provider options keys
-   * \param[in] provider_options_values Array of UTF-8 null-terminated string for provider options values
-   * \param[in] num_keys Number of elements in the `provider_option_keys` and `provider_options_values` arrays
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.15.
-   */
-  ORT_API2_STATUS(UpdateDnnlProviderOptions, _Inout_ OrtDnnlProviderOptions* dnnl_options,
-                  _In_reads_(num_keys) const char* const* provider_options_keys,
-                  _In_reads_(num_keys) const char* const* provider_options_values,
-                  _In_ size_t num_keys);
-
-  /**
-   * Get serialized oneDNN provider options string.
-   *
-   * For example, "use_arena=1;......"
-   *
-   * \param dnnl_options - OrtDnnlProviderOptions instance
-   * \param allocator - a ptr to an instance of OrtAllocator obtained with CreateAllocator() or GetAllocatorWithDefaultOptions()
-   *                      the specified allocator will be used to allocate continuous buffers for output strings and lengths.
-   * \param ptr - is a UTF-8 null terminated string allocated using 'allocator'. The caller is responsible for using the same allocator to free it.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.15.
-   */
-  ORT_API2_STATUS(GetDnnlProviderOptionsAsString, _In_ const OrtDnnlProviderOptions* dnnl_options, _Inout_ OrtAllocator* allocator, _Outptr_ char** ptr);
-
-  /** \brief Release an ::OrtDnnlProviderOptions
-   *
-   * \since Version 1.15.
-   */
-  void(ORT_API_CALL* ReleaseDnnlProviderOptions)(_Frees_ptr_opt_ OrtDnnlProviderOptions* input);
-
-  /// \name OrtKernelInfo
-  /// Custom operator APIs.
-  /// @{
-
-  /** \brief Get the graph node name from ::OrtKernelInfo.
-   *
-   * If `out` is nullptr, the value of `size` is set to the size of the name
-   * string (including null-terminator), and a success status is returned.
-   *
-   * If the `size` parameter is greater than or equal to the name string's size,
-   * the value of `size` is set to the true size of the string (including null-terminator),
-   * the provided memory is filled with the string's contents, and a success status is returned.
-   *
-   * If the `size` parameter is less than the actual string's size and `out`
-   * is not nullptr, the value of `size` is set to the true size of the string
-   * and a failure status is returned.
-   *
-   * Can be used in a custom operator's CreateKernel callback to get the name of the operator's node name in the graph.
-   *
-   * \param[in] info An instance of ::OrtKernelInfo.
-   * \param[out] out Memory location into which to write the UTF-8 null-terminated string representing the name.
-   * \param[in,out] size Pointer to the size of the `out` buffer. See above comments for details.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.15
-   */
-  ORT_API2_STATUS(KernelInfo_GetNodeName, _In_ const OrtKernelInfo* info, _Out_ char* out, _Inout_ size_t* size);
-
-  /** \brief Get the session logger from ::OrtKernelInfo.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to get a logger that can be used to log
-   * messages.
-   *
-   * \param[in] info An instance of ::OrtKernelInfo.
-   * \param[out] logger Pointer set to the session's ::OrtLogger. Owned by ONNX Runtime, so do not free.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.15
-   */
-  ORT_API2_STATUS(KernelInfo_GetLogger, _In_ const OrtKernelInfo* info, _Outptr_ const OrtLogger** logger);
-
-  /// @}
-  /// \name OrtKernelContext
-  /// Custom operator APIs.
-  /// @{
-
-  /** \brief Get the runtime logger from ::OrtKernelContext.
-   *
-   * Used in the KernelCompute callback of an OrtCustomOp to get a logger that can be used to log
-   * messages during inference.
-   *
-   * \param[in] context An instance of ::OrtKernelContext.
-   * \param[out] logger Pointer set to the kernel context's ::OrtLogger. Owned by ONNX Runtime, so do not free.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.15
-   */
-  ORT_API2_STATUS(KernelContext_GetLogger, _In_ const OrtKernelContext* context, _Outptr_ const OrtLogger** logger);
-
-  /// @}
-  /// \name OrtLogger
-  /// Custom operator APIs.
-  /// @{
-
-  /** \brief Logs a message at the given severity level using the provided ::OrtLogger.
-   *
-   * Only messages with a severity level equal or greater than the ::OrtLogger's logging severity level
-   * are logged. Use OrtApi::Logger_GetLoggingSeverityLevel to get the ::OrtLogger's logging severity
-   * level.
-   *
-   * Can be used in custom operators to log messages with the logger retrieved via OrtApi::KernelInfo_GetLogger.
-   *
-   * \param[in] logger The ::OrtLogger instance.
-   * \param[in] log_severity_level The message's severity level.
-   * \param[in] message The message to log.
-   * \param[in] file_path The filepath of the file in which the message is logged. Usually the value of ORT_FILE.
-   * \param[in] line_number The file line number in which the message is logged. Usually the value of __LINE__.
-   * \param[in] func_name The name of the function in which the message is logged. Usually the value of __FUNCTION__.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.15
-   */
-  ORT_API2_STATUS(Logger_LogMessage, _In_ const OrtLogger* logger, OrtLoggingLevel log_severity_level,
-                  _In_z_ const char* message, _In_z_ const ORTCHAR_T* file_path, int line_number,
-                  _In_z_ const char* func_name);
-
-  /** \brief Get the logging severity level of the ::OrtLogger.
-   *
-   * Can be used in a custom operator to get the logging serverity level of the ::OrtLogger associated with
-   * the ::OrtKernelInfo.
-   *
-   * \param[in] logger The ::OrtLogger instance.
-   * \param[out] out Pointer to variable assigned with the logging severity level on success.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   * \since Version 1.15
-   */
-  ORT_API2_STATUS(Logger_GetLoggingSeverityLevel, _In_ const OrtLogger* logger, _Out_ OrtLoggingLevel* out);
-
-  /// @}
-
-  /** \brief Get a ::OrtValue tensor stored as a constant initializer in the graph node.
-   *
-   * Used in the CreateKernel callback of an OrtCustomOp to get a tensor value.
-   *
-   * \param[in] info ::OrtKernelInfo instance.
-   * \param[in] index The node index.
-   * \param[out] is_constant Is it a constant node input or not.
-   * \param[out] out The OrtValue tensor value.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.15.
-   */
-  ORT_API2_STATUS(KernelInfoGetConstantInput_tensor, _In_ const OrtKernelInfo* info, size_t index, _Out_ int* is_constant, _Outptr_ const OrtValue** out);
-
-  /** \brief Get Optional Type information from an ::OrtTypeInfo
-   *
-   * This augments ::OrtTypeInfo to return an ::OrtOptionalTypeInfo when the type is optional.
-   * The OrtOptionalTypeInfo also has a nested ::OrtTypeInfo that describes the type of the optional value.
-   * ::OrtOptionalTypeInfo type can only appear within model metadata to describe inputs/outputs.
-   * The actual OrtValues that are supplied in place of optional type inputs should contain
-   * specific type that is described by ::OrtOptionalTypeInfo.
-   *
-   * So the picture: ::OrtTypeInfo -> ::OrtOptionalTypeInfo -> ::OrtTypeInfo (describes the type that can be supplied
-   * in place of the optional type when creating the actual ::OrtValue).
-   *
-   * \param[in] type_info
-   * \param[out] out A pointer to the ::OrtOptionalTypeInfo. Do not free this value,
-   *                 it is owned by OrtTypeInfo instance. When the type_info does not represent
-   *                 optional type, nullptr is returned in out.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.15.
-   */
-  ORT_API2_STATUS(CastTypeInfoToOptionalTypeInfo, _In_ const OrtTypeInfo* type_info,
-                  _Outptr_result_maybenull_ const OrtOptionalTypeInfo** out);
-
-  /** \brief Get OrtTypeInfo for the allowed contained type from an ::OrtOptionalTypeInfo.
-   *
-   * This augments ::OrtOptionalTypeInfo to return an ::OrtTypeInfo for the contained type.
-   * The OrtOptionalTypeInfo has a nested ::OrtTypeInfo that describes the type of the optional value.
-   * ::OrtOptionalTypeInfo type can only appear within model metadata to describe inputs/outputs.
-   * The actual OrtValues that are supplied in place of optional type inputs should contain
-   * specific type that is described by the returned ::OrtTypeInfo.
-   *
-   * \param[in] optional_type_info
-   * \param[out] out A pointer to the ::OrtTypeInfo for what the optional value could be.
-   * it is owned by OrtOptionalTypeInfo instance.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.15.
-   */
-  ORT_API2_STATUS(GetOptionalContainedTypeInfo, _In_ const OrtOptionalTypeInfo* optional_type_info,
-                  _Outptr_ OrtTypeInfo** out);
-
-  /** \brief Set a single string in a string tensor
-   *  Do not zero terminate the string data.
-   *
-   * \param[in] value A string tensor
-   * \param[in] index - flat index of the element
-   * \param[in] length_in_bytes length of the buffer in utf-8 bytes (without the null terminator)
-   * \param[inout] buffer - address of return value
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(GetResizedStringTensorElementBuffer, _Inout_ OrtValue* value, _In_ size_t index, _In_ size_t length_in_bytes, _Inout_ char** buffer);
-
-  /** \brief Get Allocator from KernelContext for a specific memoryInfo. Please use C API ReleaseAllocator to release out object
-   *
-   * \param[in] context OrtKernelContext instance
-   * \param[in] mem_info OrtMemoryInfo instance
-   * \param[out] out A pointer to OrtAllocator.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.15.
-   */
-  ORT_API2_STATUS(KernelContext_GetAllocator, _In_ const OrtKernelContext* context, _In_ const OrtMemoryInfo* mem_info, _Outptr_ OrtAllocator** out);
-
-  /** \brief Returns a null terminated string of the build info including git info and cxx flags
-   *
-   * \return UTF-8 encoded version string. Do not deallocate the returned buffer.
-   *
-   * \since Version 1.15.
-   */
-  const char*(ORT_API_CALL* GetBuildInfoString)(void);
-
-  /// \name OrtROCMProviderOptions
-  /// @{
-
-  /** \brief Create an OrtROCMProviderOptions
-   *
-   * \param[out] out Newly created ::OrtROCMProviderOptions. Must be released with OrtApi::ReleaseROCMProviderOptions
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.16.
-   */
-  ORT_API2_STATUS(CreateROCMProviderOptions, _Outptr_ OrtROCMProviderOptions** out);
-
-  /** \brief Set options in a ROCm Execution Provider.
-   *
-   * Please refer to https://onnxruntime.ai/docs/execution-providers/ROCm-ExecutionProvider.html
-   * to know the available keys and values. Key should be in null terminated string format of the member of
-   * ::OrtROCMProviderOptions and value should be its related range.
-   *
-   * For example, key="device_id" and value="0"
-   *
-   * \param[in] rocm_options
-   * \param[in] provider_options_keys Array of UTF-8 null-terminated string for provider options keys
-   * \param[in] provider_options_values Array of UTF-8 null-terminated string for provider options values
-   * \param[in] num_keys Number of elements in the `provider_option_keys` and `provider_options_values` arrays
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.16.
-   */
-  ORT_API2_STATUS(UpdateROCMProviderOptions, _Inout_ OrtROCMProviderOptions* rocm_options,
-                  _In_reads_(num_keys) const char* const* provider_options_keys,
-                  _In_reads_(num_keys) const char* const* provider_options_values,
-                  _In_ size_t num_keys);
-
-  /**
-   * Get serialized ROCm provider options string.
-   *
-   * For example, "device_id=0;arena_extend_strategy=0;......"
-   *
-   * \param rocm_options - OrtROCMProviderOptions instance
-   * \param allocator - a ptr to an instance of OrtAllocator obtained with CreateAllocator() or GetAllocatorWithDefaultOptions()
-   *                      the specified allocator will be used to allocate continuous buffers for output strings and lengths.
-   * \param ptr - is a UTF-8 null terminated string allocated using 'allocator'. The caller is responsible for using the same allocator to free it.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.16.
-   */
-  ORT_API2_STATUS(GetROCMProviderOptionsAsString, _In_ const OrtROCMProviderOptions* rocm_options, _Inout_ OrtAllocator* allocator, _Outptr_ char** ptr);
-
-  /** \brief Release an ::OrtROCMProviderOptions
-   *
-   * \note This is an exception in the naming convention of other Release* functions, as the name of the method does not have the V2 suffix, but the type does
-   *
-   * \since Version 1.16.
-   */
-  void(ORT_API_CALL* ReleaseROCMProviderOptions)(_Frees_ptr_opt_ OrtROCMProviderOptions* input);
-
-  /** \brief Create an allocator with specific type and register it with the ::OrtEnv
-   *  This API enhance CreateAndRegisterAllocator that it can create an allocator with specific type, not just CPU allocator
-   *  Enables sharing the allocator between multiple sessions that use the same env instance.
-   *  Lifetime of the created allocator will be valid for the duration of the environment.
-   *  Returns an error if an allocator with the same ::OrtMemoryInfo is already registered.
-   *  \param[in] env OrtEnv instance
-   *  \param[in] provider_type ExecutionProvider type
-   *  \param[in] mem_info OrtMemoryInfo instance
-   *  \param[in] arena_cfg Arena configuration
-   *  \param[in] provider_options_keys key of the provider options map
-   *  \param[in] provider_options_values value of the provider options map
-   *  \param[in] num_keys Length of the provider options map
-   */
-  ORT_API2_STATUS(CreateAndRegisterAllocatorV2, _Inout_ OrtEnv* env, _In_ const char* provider_type, _In_ const OrtMemoryInfo* mem_info, _In_ const OrtArenaCfg* arena_cfg,
-                  _In_reads_(num_keys) const char* const* provider_options_keys, _In_reads_(num_keys) const char* const* provider_options_values, _In_ size_t num_keys);
-
-  /** \brief Run the model asynchronously in a thread owned by intra op thread pool
-   *
-   * \param[in] session
-   * \param[in] run_options If nullptr, will use a default ::OrtRunOptions
-   * \param[in] input_names Array of null terminated UTF8 encoded strings of the input names
-   * \param[in] input Array of ::OrtValue%s of the input values
-   * \param[in] input_len Number of elements in the input_names and inputs arrays
-   * \param[in] output_names Array of null terminated UTF8 encoded strings of the output names
-   * \param[in] output_names_len Number of elements in the output_names and outputs array
-   * \param[out] output OrtValue* array of size output_names_len.
-   *             On calling RunAsync, output[i] could either be a null or a pointer to a preallocated OrtValue.
-   *             Later, the output array will be passed to run_async_callback with all null(s) filled with valid
-   *             OrtValue pointer(s) allocated by onnxruntime.
-   *             NOTE: it is customer's duty to finally release the output array and each of its member,
-   *             regardless of whether the member (OrtValue*) is allocated by onnxruntime or preallocated by the customer.
-   * \param[in] run_async_callback Callback function on model run completion
-   * \param[in] user_data User data that pass back to run_async_callback
-   */
-  ORT_API2_STATUS(RunAsync, _Inout_ OrtSession* session, _In_opt_ const OrtRunOptions* run_options,
-                  _In_reads_(input_len) const char* const* input_names,
-                  _In_reads_(input_len) const OrtValue* const* input, size_t input_len,
-                  _In_reads_(output_names_len) const char* const* output_names, size_t output_names_len,
-                  _Inout_updates_all_(output_names_len) OrtValue** output,
-                  _In_ RunAsyncCallbackFn run_async_callback, _In_opt_ void* user_data);
-
-  /**
-   * Update TensorRT EP provider option where its data type is pointer, for example 'user_compute_stream'.
-   * If the data type of the provider option can be represented by string please use UpdateTensorRTProviderOptions.
-   *
-   * Note: It's caller's responsibility to properly manage the lifetime of the instance pointed by this pointer.
-   *
-   * \param tensorrt_options - OrtTensorRTProviderOptionsV2 instance
-   * \param key - Name of the provider option
-   * \param value - A pointer to the instance that will be assigned to this provider option
-   *
-   * \since Version 1.16.
-   */
-  ORT_API2_STATUS(UpdateTensorRTProviderOptionsWithValue, _Inout_ OrtTensorRTProviderOptionsV2* tensorrt_options, _In_ const char* key, _In_ void* value);
-
-  /**
-   * Get TensorRT EP provider option where its data type is pointer.
-   * If the data type of the provider option can be represented by string please use GetTensorRTProviderOptionsAsString.
-   *
-   * \param tensorrt_options - OrtTensorRTProviderOptionsV2 instance
-   * \param key - Name of the provider option
-   * \param ptr - A pointer to the instance that is kept by the provider option
-   *
-   * \since Version 1.16.
-   */
-  ORT_API2_STATUS(GetTensorRTProviderOptionsByName, _In_ const OrtTensorRTProviderOptionsV2* tensorrt_options, _In_ const char* key, _Outptr_ void** ptr);
-
-  /**
-   * Update CUDA EP provider option where its data type is pointer, for example 'user_compute_stream'.
-   * If the data type of the provider option can be represented by string please use UpdateCUDAProviderOptions.
-   *
-   * Note: It's caller's responsibility to properly manage the lifetime of the instance pointed by this pointer.
-   *
-   * \param cuda_options - OrtCUDAProviderOptionsV2 instance
-   * \param key - Name of the provider option
-   * \param value - A pointer to the instance that will be assigned to this provider option
-   *
-   * \since Version 1.16.
-   */
-  ORT_API2_STATUS(UpdateCUDAProviderOptionsWithValue, _Inout_ OrtCUDAProviderOptionsV2* cuda_options, _In_ const char* key, _In_ void* value);
-
-  /**
-   * Get CUDA EP provider option where its data type is pointer.
-   * If the data type of the provider option can be represented by string please use GetCUDAProviderOptionsAsString.
-   *
-   * \param cuda_options - OrtCUDAProviderOptionsV2 instance
-   * \param key - Name of the provider option
-   * \param ptr - A pointer to the instance that is kept by the provider option
-   *
-   * \since Version 1.16.
-   */
-  ORT_API2_STATUS(GetCUDAProviderOptionsByName, _In_ const OrtCUDAProviderOptionsV2* cuda_options, _In_ const char* key, _Outptr_ void** ptr);
-
-  /**
-   * Get a EP resource.
-   * E.g. a cuda stream or a cublas handle
-   *
-   * \param context - Kernel context
-   * \param resouce_version - Version of the resource
-   * \param resource_id - Type of resource
-   * \param resource - A pointer to returned resource
-   *
-   * \since Version 1.16.
-   */
-  ORT_API2_STATUS(KernelContext_GetResource, _In_ const OrtKernelContext* context, _In_ int resouce_version, _In_ int resource_id, _Outptr_ void** resource);
-
-  /** \brief Set user logging function
-   *
-   *  By default the logger created by the CreateEnv* functions is used to create the session logger as well.
-   *  This function allows a user to override this default session logger with a logger of their own choosing. This way
-   *  the user doesn't have to create a separate environment with a custom logger. This addresses the problem when
-   *  the user already created an env but now wants to use a different logger for a specific session (for debugging or
-   *  other reasons).
-   *
-   * \param[in] options
-   * \param[in] user_logging_function A pointer to a logging function.
-   * \param[in] user_logging_param A pointer to arbitrary data passed as the ::OrtLoggingFunction `param` parameter to
-   *                         `user_logging_function`. This parameter is optional.
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   *
-   * \since Version 1.17.
-   */
-  ORT_API2_STATUS(SetUserLoggingFunction, _Inout_ OrtSessionOptions* options,
-                  _In_ OrtLoggingFunction user_logging_function, _In_opt_ void* user_logging_param);
-
-  /**
-   * Get number of input from OrtShapeInferContext
-   *
-   * \param[in] context
-   * \param[out] out The number of inputs
-   *
-   * \since Version 1.17.
-   */
-  ORT_API2_STATUS(ShapeInferContext_GetInputCount, _In_ const OrtShapeInferContext* context, _Out_ size_t* out);
-
-  /**
-   * Get type and shape info of an input
-   *
-   * \param[in] context
-   * \param[in] index The index of the input
-   * \param[out] info Type shape info of the input
-   *
-   * \since Version 1.17.
-   */
-  ORT_API2_STATUS(ShapeInferContext_GetInputTypeShape, _In_ const OrtShapeInferContext* context, _In_ size_t index, _Outptr_ OrtTensorTypeAndShapeInfo** info);
-
-  /**
-   * Get attribute from OrtShapeInferContext. Note that OrtShapeInferContext is a per-node context, one could only read attribute from current node.
-   *
-   * \param[in] context
-   * \param[in] attr_name Name of the attribute
-   * \param[out] attr Handle of the attribute fetched
-   *
-   * \since Version 1.17.
-   */
-  ORT_API2_STATUS(ShapeInferContext_GetAttribute, _In_ const OrtShapeInferContext* context, _In_ const char* attr_name, _Outptr_ const OrtOpAttr** attr);
-
-  /**
-   * Set type and shape info of an ouput
-   *
-   * \param[in] context
-   * \param[in] index The index of the ouput
-   * \param[out] info Type shape info of the output
-   *
-   * \since Version 1.17.
-   */
-  ORT_API2_STATUS(ShapeInferContext_SetOutputTypeShape, _In_ const OrtShapeInferContext* context, _In_ size_t index, _In_ const OrtTensorTypeAndShapeInfo* info);
-
-  /**
-   * Set symbolic shape to type shape info
-   *
-   * \param[in] info Type shape info
-   * \param[in] dim_params Symbolic strings
-   * \param[in] dim_params_length Number of strings
-   *
-   * \since Version 1.17.
-   */
-  ORT_API2_STATUS(SetSymbolicDimensions, _In_ OrtTensorTypeAndShapeInfo* info, _In_ const char* dim_params[], _In_ size_t dim_params_length);
-
-  /**
-   * Read contents of an attribute to data
-   *
-   * \param[in] op_attr
-   * \param[in] type Attribute type
-   * \param[out] data Memory address to save raw content of the attribute
-   * \param[in] len Number of bytes allowed to store in data
-   * \param[out] out Number of bytes required to save the data when the call failed, or the real number of bytes saved to data on success
-   *
-   * \since Version 1.17.
-   */
-  ORT_API2_STATUS(ReadOpAttr, _In_ const OrtOpAttr* op_attr, _In_ OrtOpAttrType type, _Inout_ void* data, _In_ size_t len, _Out_ size_t* out);
-
-  /** \brief Set whether to use deterministic compute.
-   *
-   * Default is false. If set to true, this will enable deterministic compute for GPU kernels where possible.
-   * Note that this most likely will have a performance cost.
-   *
-   * \param[in] options
-   * \param[in] value
-   *
-   * \since Version 1.17.
-   */
-  ORT_API2_STATUS(SetDeterministicCompute, _Inout_ OrtSessionOptions* options, bool value);
-
-  /**
-   * Run fn in parallel
-   *
-   * \param[in] context
-   * \param[in] fn Function accepting usr_data and an integer as iterator
-   * \param[in] total The number of times fn is to be invoked
-   * \param[in] num_batch Number of batches by which the "total" is to be divided in maximum. When zero, there is no limit
-   * \param[in] usr_data User data to be passed back to fn
-   *
-   * \since Version 1.17.
-   */
-  ORT_API2_STATUS(KernelContext_ParallelFor, _In_ const OrtKernelContext* context, _In_ void (*fn)(void*, size_t), _In_ size_t total, _In_ size_t num_batch, _In_ void* usr_data);
-
-  /** \brief Append OpenVINO execution provider to the session options
-   *
-   * If OpenVINO is not available (due to a non OpenVINO enabled build, or if OpenVINO is not installed on the system), this function will fail.
-   *
-   * \param[in] options
-   * \param[in] provider_options_keys
-   * \param[in] provider_options_values
-   * \param[in] num_keys
-   *
-   * \snippet{doc} snippets.dox OrtStatus Return Value
-   */
-  ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_OpenVINO_V2,
-                  _In_ OrtSessionOptions* options,
-                  _In_reads_(num_keys) const char* const* provider_options_keys,
-                  _In_reads_(num_keys) const char* const* provider_options_values,
-                  _In_ size_t num_keys);
-};
-
-/*
- * Steps to use a custom op:
- *   1 Create an OrtCustomOpDomain with the domain name used by the custom ops
- *   2 Create an OrtCustomOp structure for each op and add them to the domain
- *   3 Call OrtAddCustomOpDomain to add the custom domain of ops to the session options
- */
-
-// Specifies some characteristics of inputs/outputs of custom ops:
-// Specify if the inputs/outputs are one of:
-// 1) Non-optional (input/output must be present in the node)
-// 2) Optional (input/output may be absent in the node)
-// 3) Variadic: A variadic input or output specifies N (i.e., the minimum arity) or more operands.
-//              Only the last input or output of a custom op may be marked as variadic.
-//              The homogeneity of the variadic input or output determines whether all operands must be of the same
-//              tensor element type.
-typedef enum OrtCustomOpInputOutputCharacteristic {
-  INPUT_OUTPUT_REQUIRED = 0,
-  INPUT_OUTPUT_OPTIONAL,
-  INPUT_OUTPUT_VARIADIC,
-} OrtCustomOpInputOutputCharacteristic;
-
-/*
- * The OrtCustomOp structure defines a custom op's schema and its kernel callbacks. The callbacks are filled in by
- * the implementor of the custom op.
- */
-struct OrtCustomOp {
-  uint32_t version;  // Must be initialized to ORT_API_VERSION
-
-  // This callback creates the kernel, which is a user defined
-  // parameter that is passed to the Kernel* callbacks below. It is
-  // recommended to use CreateKernelV2 which allows for a safe error
-  // propagation by returning an OrtStatusPtr.
-  void*(ORT_API_CALL* CreateKernel)(_In_ const struct OrtCustomOp* op, _In_ const OrtApi* api,
-                                    _In_ const OrtKernelInfo* info);
-
-  // Returns the name of the op
-  const char*(ORT_API_CALL* GetName)(_In_ const struct OrtCustomOp* op);
-
-  // Returns the type of the execution provider, return nullptr to use CPU execution provider
-  const char*(ORT_API_CALL* GetExecutionProviderType)(_In_ const struct OrtCustomOp* op);
-
-  // Returns the count and types of the input & output tensors
-  ONNXTensorElementDataType(ORT_API_CALL* GetInputType)(_In_ const struct OrtCustomOp* op, _In_ size_t index);
-  size_t(ORT_API_CALL* GetInputTypeCount)(_In_ const struct OrtCustomOp* op);
-  ONNXTensorElementDataType(ORT_API_CALL* GetOutputType)(_In_ const struct OrtCustomOp* op, _In_ size_t index);
-  size_t(ORT_API_CALL* GetOutputTypeCount)(_In_ const struct OrtCustomOp* op);
-
-  // Perform a computation step.  It is recommended to use
-  // KernelComputeV2 which allows for a safe error propagation by
-  // returning an OrtStatusPtr.
-  void(ORT_API_CALL* KernelCompute)(_In_ void* op_kernel, _In_ OrtKernelContext* context);
-  void(ORT_API_CALL* KernelDestroy)(_In_ void* op_kernel);
-
-  // Returns the characteristics of the input & output tensors
-  OrtCustomOpInputOutputCharacteristic(ORT_API_CALL* GetInputCharacteristic)(_In_ const struct OrtCustomOp* op, _In_ size_t index);
-  OrtCustomOpInputOutputCharacteristic(ORT_API_CALL* GetOutputCharacteristic)(_In_ const struct OrtCustomOp* op, _In_ size_t index);
-
-  // Returns the memory type of the input tensors. This API allows the custom op
-  // to place the inputs on specific devices. By default, it returns
-  // OrtMemTypeDefault, which means the input is placed on the default device for
-  // the execution provider. If the inputs need to be with different memory tyeps,
-  // this function can be overridden to return the specific memory types.
-  OrtMemType(ORT_API_CALL* GetInputMemoryType)(_In_ const struct OrtCustomOp* op, _In_ size_t index);
-
-  // Returns the minimum number of input arguments expected for the variadic input.
-  // Applicable only for custom ops that have a variadic input.
-  int(ORT_API_CALL* GetVariadicInputMinArity)(_In_ const struct OrtCustomOp* op);
-
-  // Returns true (non-zero) if all arguments of a variadic input have to be of the same type (homogeneous),
-  // and false (zero) otherwise.
-  // Applicable only for custom ops that have a variadic input.
-  int(ORT_API_CALL* GetVariadicInputHomogeneity)(_In_ const struct OrtCustomOp* op);
-
-  // Returns the minimum number of output values expected for the variadic output.
-  // Applicable only for custom ops that have a variadic output.
-  int(ORT_API_CALL* GetVariadicOutputMinArity)(_In_ const struct OrtCustomOp* op);
-
-  // Returns true (non-zero) if all outputs values of a variadic output have to be of the same type (homogeneous),
-  // and false (zero) otherwise.
-  // Applicable only for custom ops that have a variadic output.
-  int(ORT_API_CALL* GetVariadicOutputHomogeneity)(_In_ const struct OrtCustomOp* op);
-
-  // Create the kernel state which is passed to each compute call.
-  OrtStatusPtr(ORT_API_CALL* CreateKernelV2)(_In_ const struct OrtCustomOp* op, _In_ const OrtApi* api,
-                                             _In_ const OrtKernelInfo* info,
-                                             _Out_ void** kernel);
-
-  // Perform the computation step.
-  OrtStatusPtr(ORT_API_CALL* KernelComputeV2)(_In_ void* op_kernel, _In_ OrtKernelContext* context);
-
-  OrtStatusPtr(ORT_API_CALL* InferOutputShapeFn)(_In_ const struct OrtCustomOp* op, _In_ OrtShapeInferContext*);
-
-  // Get start range
-  int(ORT_API_CALL* GetStartVersion)(_In_ const struct OrtCustomOp* op);
-  int(ORT_API_CALL* GetEndVersion)(_In_ const struct OrtCustomOp* op);
-};
-
-/*
- * This is the old way to add the CUDA provider to the session, please use SessionOptionsAppendExecutionProvider_CUDA above to access the latest functionality
- * This function always exists, but will only succeed if Onnxruntime was built with CUDA support and the CUDA provider shared library exists
- *
- * \param device_id CUDA device id, starts from zero.
- */
-ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_CUDA, _In_ OrtSessionOptions* options, int device_id);
-
-/*
- * This is the old way to add the ROCm provider to the session, please use
- * SessionOptionsAppendExecutionProvider_ROCM above to access the latest functionality
- * This function always exists, but will only succeed if Onnxruntime was built with
- * HIP support and the ROCm provider shared library exists
- *
- * \param device_id HIP device id, starts from zero.
- */
-ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_ROCM, _In_ OrtSessionOptions* options, int device_id);
-
-/*
- * This is the old way to add the MIGraphX provider to the session, please use
- * SessionOptionsAppendExecutionProvider_MIGraphX above to access the latest functionality
- * This function always exists, but will only succeed if Onnxruntime was built with
- * HIP support and the MIGraphX provider shared library exists
- *
- * \param device_id HIP device id, starts from zero.
- */
-ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_MIGraphX, _In_ OrtSessionOptions* options, int device_id);
-
-/*
- * This is the old way to add the oneDNN provider to the session, please use
- * SessionOptionsAppendExecutionProvider_oneDNN above to access the latest functionality
- * This function always exists, but will only succeed if Onnxruntime was built with
- * oneDNN support and the oneDNN provider shared library exists
- *
- * \param use_arena zero: false. non-zero: true.
- */
-ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_Dnnl, _In_ OrtSessionOptions* options, int use_arena);
-
-/*
- * This is the old way to add the TensorRT provider to the session, please use SessionOptionsAppendExecutionProvider_TensorRT_V2 above to access the latest functionality
- * This function always exists, but will only succeed if Onnxruntime was built with TensorRT support and the TensorRT provider shared library exists
- *
- * \param device_id CUDA device id, starts from zero.
- */
-ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_Tensorrt, _In_ OrtSessionOptions* options, int device_id);
-
-#ifdef __cplusplus
-}
-#endif
-/// @}
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_cxx_api.h b/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_cxx_api.h
deleted file mode 100644
index 7a553f9f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_cxx_api.h
+++ /dev/null
@@ -1,2372 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-// Summary: The Ort C++ API is a header only wrapper around the Ort C API.
-//
-// The C++ API simplifies usage by returning values directly instead of error codes, throwing exceptions on errors
-// and automatically releasing resources in the destructors. The primary purpose of C++ API is exception safety so
-// all the resources follow RAII and do not leak memory.
-//
-// Each of the C++ wrapper classes holds only a pointer to the C internal object. Treat them like smart pointers.
-// To create an empty object, pass 'nullptr' to the constructor (for example, Env e{nullptr};). However, you can't use them
-// until you assign an instance that actually holds an underlying object.
-//
-// For Ort objects only move assignment between objects is allowed, there are no copy constructors.
-// Some objects have explicit 'Clone' methods for this purpose.
-//
-// ConstXXXX types are copyable since they do not own the underlying C object, so you can pass them to functions as arguments
-// by value or by reference. ConstXXXX types are restricted to const only interfaces.
-//
-// UnownedXXXX are similar to ConstXXXX but also allow non-const interfaces.
-//
-// The lifetime of the corresponding owning object must eclipse the lifetimes of the ConstXXXX/UnownedXXXX types. They exists so you do not
-// have to fallback to C types and the API with the usual pitfalls. In general, do not use C API from your C++ code.
-
-#pragma once
-#include "onnxruntime_c_api.h"
-#include "onnxruntime_float16.h"
-
-#include <cstddef>
-#include <cstdio>
-#include <array>
-#include <memory>
-#include <stdexcept>
-#include <string>
-#include <vector>
-#include <unordered_map>
-#include <utility>
-#include <type_traits>
-
-#ifdef ORT_NO_EXCEPTIONS
-#include <iostream>
-#endif
-
-/** \brief All C++ Onnxruntime APIs are defined inside this namespace
- *
- */
-namespace Ort {
-
-/** \brief All C++ methods that can fail will throw an exception of this type
- *
- * If <tt>ORT_NO_EXCEPTIONS</tt> is defined, then any error will result in a call to abort()
- */
-struct Exception : std::exception {
-  Exception(std::string&& string, OrtErrorCode code) : message_{std::move(string)}, code_{code} {}
-
-  OrtErrorCode GetOrtErrorCode() const { return code_; }
-  const char* what() const noexcept override { return message_.c_str(); }
-
- private:
-  std::string message_;
-  OrtErrorCode code_;
-};
-
-#ifdef ORT_NO_EXCEPTIONS
-// The #ifndef is for the very special case where the user of this library wants to define their own way of handling errors.
-// NOTE: This header expects control flow to not continue after calling ORT_CXX_API_THROW
-#ifndef ORT_CXX_API_THROW
-#define ORT_CXX_API_THROW(string, code)       \
-  do {                                        \
-    std::cerr << Ort::Exception(string, code) \
-                     .what()                  \
-              << std::endl;                   \
-    abort();                                  \
-  } while (false)
-#endif
-#else
-#define ORT_CXX_API_THROW(string, code) \
-  throw Ort::Exception(string, code)
-#endif
-
-// This is used internally by the C++ API. This class holds the global variable that points to the OrtApi,
-//  it's in a template so that we can define a global variable in a header and make
-// it transparent to the users of the API.
-template <typename T>
-struct Global {
-  static const OrtApi* api_;
-};
-
-// If macro ORT_API_MANUAL_INIT is defined, no static initialization will be performed. Instead, user must call InitApi() before using it.
-template <typename T>
-#ifdef ORT_API_MANUAL_INIT
-const OrtApi* Global<T>::api_{};
-inline void InitApi() noexcept { Global<void>::api_ = OrtGetApiBase()->GetApi(ORT_API_VERSION); }
-
-// Used by custom operator libraries that are not linked to onnxruntime. Sets the global API object, which is
-// required by C++ APIs.
-//
-// Example mycustomop.cc:
-//
-// #define ORT_API_MANUAL_INIT
-// #include <onnxruntime_cxx_api.h>
-// #undef ORT_API_MANUAL_INIT
-//
-// OrtStatus* ORT_API_CALL RegisterCustomOps(OrtSessionOptions* options, const OrtApiBase* api_base) {
-//   Ort::InitApi(api_base->GetApi(ORT_API_VERSION));
-//   // ...
-// }
-//
-inline void InitApi(const OrtApi* api) noexcept { Global<void>::api_ = api; }
-#else
-#if defined(_MSC_VER) && !defined(__clang__)
-#pragma warning(push)
-// "Global initializer calls a non-constexpr function." Therefore you can't use ORT APIs in the other global initializers.
-// Please define ORT_API_MANUAL_INIT if it conerns you.
-#pragma warning(disable : 26426)
-#endif
-const OrtApi* Global<T>::api_ = OrtGetApiBase()->GetApi(ORT_API_VERSION);
-#if defined(_MSC_VER) && !defined(__clang__)
-#pragma warning(pop)
-#endif
-#endif
-
-/// This returns a reference to the OrtApi interface in use
-inline const OrtApi& GetApi() noexcept { return *Global<void>::api_; }
-
-/// <summary>
-/// This function returns the onnxruntime version string
-/// </summary>
-/// <returns>version string major.minor.rev</returns>
-std::string GetVersionString();
-
-/// <summary>
-/// This function returns the onnxruntime build information: including git branch,
-/// git commit id, build type(Debug/Release/RelWithDebInfo) and cmake cpp flags.
-/// </summary>
-/// <returns>string</returns>
-std::string GetBuildInfoString();
-
-/// <summary>
-/// This is a C++ wrapper for OrtApi::GetAvailableProviders() and
-/// returns a vector of strings representing the available execution providers.
-/// </summary>
-/// <returns>vector of strings</returns>
-std::vector<std::string> GetAvailableProviders();
-
-/** \brief IEEE 754 half-precision floating point data type
- *
- * \details This struct is used for converting float to float16 and back
- * so the user could feed inputs and fetch outputs using these type.
- *
- * The size of the structure should align with uint16_t and one can freely cast
- * uint16_t buffers to/from Ort::Float16_t to feed and retrieve data.
- *
- * \code{.unparsed}
- * // This example demonstrates converion from float to float16
- * constexpr float values[] = {1.f, 2.f, 3.f, 4.f, 5.f};
- * std::vector<Ort::Float16_t> fp16_values;
- * fp16_values.reserve(std::size(values));
- * std::transform(std::begin(values), std::end(values), std::back_inserter(fp16_values),
- *     [](float value) { return Ort::Float16_t(value); });
- *
- * \endcode
- */
-struct Float16_t : onnxruntime_float16::Float16Impl<Float16_t> {
- private:
-  /// <summary>
-  /// Constructor from a 16-bit representation of a float16 value
-  /// No conversion is done here.
-  /// </summary>
-  /// <param name="v">16-bit representation</param>
-  constexpr explicit Float16_t(uint16_t v) noexcept { val = v; }
-
- public:
-  using Base = onnxruntime_float16::Float16Impl<Float16_t>;
-
-  /// <summary>
-  /// Default constructor
-  /// </summary>
-  Float16_t() = default;
-
-  /// <summary>
-  /// Explicit conversion to uint16_t representation of float16.
-  /// </summary>
-  /// <param name="v">uint16_t bit representation of float16</param>
-  /// <returns>new instance of Float16_t</returns>
-  constexpr static Float16_t FromBits(uint16_t v) noexcept { return Float16_t(v); }
-
-  /// <summary>
-  /// __ctor from float. Float is converted into float16 16-bit representation.
-  /// </summary>
-  /// <param name="v">float value</param>
-  explicit Float16_t(float v) noexcept { val = Base::ToUint16Impl(v); }
-
-  /// <summary>
-  /// Converts float16 to float
-  /// </summary>
-  /// <returns>float representation of float16 value</returns>
-  float ToFloat() const noexcept { return Base::ToFloatImpl(); }
-
-  /// <summary>
-  /// Checks if the value is negative
-  /// </summary>
-  /// <returns>true if negative</returns>
-  using Base::IsNegative;
-
-  /// <summary>
-  /// Tests if the value is NaN
-  /// </summary>
-  /// <returns>true if NaN</returns>
-  using Base::IsNaN;
-
-  /// <summary>
-  /// Tests if the value is finite
-  /// </summary>
-  /// <returns>true if finite</returns>
-  using Base::IsFinite;
-
-  /// <summary>
-  /// Tests if the value represents positive infinity.
-  /// </summary>
-  /// <returns>true if positive infinity</returns>
-  using Base::IsPositiveInfinity;
-
-  /// <summary>
-  /// Tests if the value represents negative infinity
-  /// </summary>
-  /// <returns>true if negative infinity</returns>
-  using Base::IsNegativeInfinity;
-
-  /// <summary>
-  /// Tests if the value is either positive or negative infinity.
-  /// </summary>
-  /// <returns>True if absolute value is infinity</returns>
-  using Base::IsInfinity;
-
-  /// <summary>
-  /// Tests if the value is NaN or zero. Useful for comparisons.
-  /// </summary>
-  /// <returns>True if NaN or zero.</returns>
-  using Base::IsNaNOrZero;
-
-  /// <summary>
-  /// Tests if the value is normal (not zero, subnormal, infinite, or NaN).
-  /// </summary>
-  /// <returns>True if so</returns>
-  using Base::IsNormal;
-
-  /// <summary>
-  /// Tests if the value is subnormal (denormal).
-  /// </summary>
-  /// <returns>True if so</returns>
-  using Base::IsSubnormal;
-
-  /// <summary>
-  /// Creates an instance that represents absolute value.
-  /// </summary>
-  /// <returns>Absolute value</returns>
-  using Base::Abs;
-
-  /// <summary>
-  /// Creates a new instance with the sign flipped.
-  /// </summary>
-  /// <returns>Flipped sign instance</returns>
-  using Base::Negate;
-
-  /// <summary>
-  /// IEEE defines that positive and negative zero are equal, this gives us a quick equality check
-  /// for two values by or'ing the private bits together and stripping the sign. They are both zero,
-  /// and therefore equivalent, if the resulting value is still zero.
-  /// </summary>
-  /// <param name="lhs">first value</param>
-  /// <param name="rhs">second value</param>
-  /// <returns>True if both arguments represent zero</returns>
-  using Base::AreZero;
-
-  /// <summary>
-  /// User defined conversion operator. Converts Float16_t to float.
-  /// </summary>
-  explicit operator float() const noexcept { return ToFloat(); }
-
-  using Base::operator==;
-  using Base::operator!=;
-  using Base::operator<;
-};
-
-static_assert(sizeof(Float16_t) == sizeof(uint16_t), "Sizes must match");
-
-/** \brief bfloat16 (Brain Floating Point) data type
- *
- * \details This struct is used for converting float to bfloat16 and back
- * so the user could feed inputs and fetch outputs using these type.
- *
- * The size of the structure should align with uint16_t and one can freely cast
- * uint16_t buffers to/from Ort::BFloat16_t to feed and retrieve data.
- *
- * \code{.unparsed}
- * // This example demonstrates converion from float to float16
- * constexpr float values[] = {1.f, 2.f, 3.f, 4.f, 5.f};
- * std::vector<Ort::BFloat16_t> bfp16_values;
- * bfp16_values.reserve(std::size(values));
- * std::transform(std::begin(values), std::end(values), std::back_inserter(bfp16_values),
- *     [](float value) { return Ort::BFloat16_t(value); });
- *
- * \endcode
- */
-struct BFloat16_t : onnxruntime_float16::BFloat16Impl<BFloat16_t> {
- private:
-  /// <summary>
-  /// Constructor from a uint16_t representation of bfloat16
-  /// used in FromBits() to escape overload resolution issue with
-  /// constructor from float.
-  /// No conversion is done.
-  /// </summary>
-  /// <param name="v">16-bit bfloat16 value</param>
-  constexpr explicit BFloat16_t(uint16_t v) noexcept { val = v; }
-
- public:
-  using Base = onnxruntime_float16::BFloat16Impl<BFloat16_t>;
-
-  BFloat16_t() = default;
-
-  /// <summary>
-  /// Explicit conversion to uint16_t representation of bfloat16.
-  /// </summary>
-  /// <param name="v">uint16_t bit representation of bfloat16</param>
-  /// <returns>new instance of BFloat16_t</returns>
-  static constexpr BFloat16_t FromBits(uint16_t v) noexcept { return BFloat16_t(v); }
-
-  /// <summary>
-  /// __ctor from float. Float is converted into bfloat16 16-bit representation.
-  /// </summary>
-  /// <param name="v">float value</param>
-  explicit BFloat16_t(float v) noexcept { val = Base::ToUint16Impl(v); }
-
-  /// <summary>
-  /// Converts bfloat16 to float
-  /// </summary>
-  /// <returns>float representation of bfloat16 value</returns>
-  float ToFloat() const noexcept { return Base::ToFloatImpl(); }
-
-  /// <summary>
-  /// Checks if the value is negative
-  /// </summary>
-  /// <returns>true if negative</returns>
-  using Base::IsNegative;
-
-  /// <summary>
-  /// Tests if the value is NaN
-  /// </summary>
-  /// <returns>true if NaN</returns>
-  using Base::IsNaN;
-
-  /// <summary>
-  /// Tests if the value is finite
-  /// </summary>
-  /// <returns>true if finite</returns>
-  using Base::IsFinite;
-
-  /// <summary>
-  /// Tests if the value represents positive infinity.
-  /// </summary>
-  /// <returns>true if positive infinity</returns>
-  using Base::IsPositiveInfinity;
-
-  /// <summary>
-  /// Tests if the value represents negative infinity
-  /// </summary>
-  /// <returns>true if negative infinity</returns>
-  using Base::IsNegativeInfinity;
-
-  /// <summary>
-  /// Tests if the value is either positive or negative infinity.
-  /// </summary>
-  /// <returns>True if absolute value is infinity</returns>
-  using Base::IsInfinity;
-
-  /// <summary>
-  /// Tests if the value is NaN or zero. Useful for comparisons.
-  /// </summary>
-  /// <returns>True if NaN or zero.</returns>
-  using Base::IsNaNOrZero;
-
-  /// <summary>
-  /// Tests if the value is normal (not zero, subnormal, infinite, or NaN).
-  /// </summary>
-  /// <returns>True if so</returns>
-  using Base::IsNormal;
-
-  /// <summary>
-  /// Tests if the value is subnormal (denormal).
-  /// </summary>
-  /// <returns>True if so</returns>
-  using Base::IsSubnormal;
-
-  /// <summary>
-  /// Creates an instance that represents absolute value.
-  /// </summary>
-  /// <returns>Absolute value</returns>
-  using Base::Abs;
-
-  /// <summary>
-  /// Creates a new instance with the sign flipped.
-  /// </summary>
-  /// <returns>Flipped sign instance</returns>
-  using Base::Negate;
-
-  /// <summary>
-  /// IEEE defines that positive and negative zero are equal, this gives us a quick equality check
-  /// for two values by or'ing the private bits together and stripping the sign. They are both zero,
-  /// and therefore equivalent, if the resulting value is still zero.
-  /// </summary>
-  /// <param name="lhs">first value</param>
-  /// <param name="rhs">second value</param>
-  /// <returns>True if both arguments represent zero</returns>
-  using Base::AreZero;
-
-  /// <summary>
-  /// User defined conversion operator. Converts BFloat16_t to float.
-  /// </summary>
-  explicit operator float() const noexcept { return ToFloat(); }
-
-  // We do not have an inherited impl for the below operators
-  // as the internal class implements them a little differently
-  bool operator==(const BFloat16_t& rhs) const noexcept;
-  bool operator!=(const BFloat16_t& rhs) const noexcept { return !(*this == rhs); }
-  bool operator<(const BFloat16_t& rhs) const noexcept;
-};
-
-static_assert(sizeof(BFloat16_t) == sizeof(uint16_t), "Sizes must match");
-
-/** \brief float8e4m3fn (Float8 Floating Point) data type
- * \details It is necessary for type dispatching to make use of C++ API
- * The type is implicitly convertible to/from uint8_t.
- * See https://onnx.ai/onnx/technical/float8.html for further details.
- */
-struct Float8E4M3FN_t {
-  uint8_t value;
-  constexpr Float8E4M3FN_t() noexcept : value(0) {}
-  constexpr Float8E4M3FN_t(uint8_t v) noexcept : value(v) {}
-  constexpr operator uint8_t() const noexcept { return value; }
-  // nan values are treated like any other value for operator ==, !=
-  constexpr bool operator==(const Float8E4M3FN_t& rhs) const noexcept { return value == rhs.value; };
-  constexpr bool operator!=(const Float8E4M3FN_t& rhs) const noexcept { return value != rhs.value; };
-};
-
-static_assert(sizeof(Float8E4M3FN_t) == sizeof(uint8_t), "Sizes must match");
-
-/** \brief float8e4m3fnuz (Float8 Floating Point) data type
- * \details It is necessary for type dispatching to make use of C++ API
- * The type is implicitly convertible to/from uint8_t.
- * See https://onnx.ai/onnx/technical/float8.html for further details.
- */
-struct Float8E4M3FNUZ_t {
-  uint8_t value;
-  constexpr Float8E4M3FNUZ_t() noexcept : value(0) {}
-  constexpr Float8E4M3FNUZ_t(uint8_t v) noexcept : value(v) {}
-  constexpr operator uint8_t() const noexcept { return value; }
-  // nan values are treated like any other value for operator ==, !=
-  constexpr bool operator==(const Float8E4M3FNUZ_t& rhs) const noexcept { return value == rhs.value; };
-  constexpr bool operator!=(const Float8E4M3FNUZ_t& rhs) const noexcept { return value != rhs.value; };
-};
-
-static_assert(sizeof(Float8E4M3FNUZ_t) == sizeof(uint8_t), "Sizes must match");
-
-/** \brief float8e5m2 (Float8 Floating Point) data type
- * \details It is necessary for type dispatching to make use of C++ API
- * The type is implicitly convertible to/from uint8_t.
- * See https://onnx.ai/onnx/technical/float8.html for further details.
- */
-struct Float8E5M2_t {
-  uint8_t value;
-  constexpr Float8E5M2_t() noexcept : value(0) {}
-  constexpr Float8E5M2_t(uint8_t v) noexcept : value(v) {}
-  constexpr operator uint8_t() const noexcept { return value; }
-  // nan values are treated like any other value for operator ==, !=
-  constexpr bool operator==(const Float8E5M2_t& rhs) const noexcept { return value == rhs.value; };
-  constexpr bool operator!=(const Float8E5M2_t& rhs) const noexcept { return value != rhs.value; };
-};
-
-static_assert(sizeof(Float8E5M2_t) == sizeof(uint8_t), "Sizes must match");
-
-/** \brief float8e5m2fnuz (Float8 Floating Point) data type
- * \details It is necessary for type dispatching to make use of C++ API
- * The type is implicitly convertible to/from uint8_t.
- * See https://onnx.ai/onnx/technical/float8.html for further details.
- */
-struct Float8E5M2FNUZ_t {
-  uint8_t value;
-  constexpr Float8E5M2FNUZ_t() noexcept : value(0) {}
-  constexpr Float8E5M2FNUZ_t(uint8_t v) noexcept : value(v) {}
-  constexpr operator uint8_t() const noexcept { return value; }
-  // nan values are treated like any other value for operator ==, !=
-  constexpr bool operator==(const Float8E5M2FNUZ_t& rhs) const noexcept { return value == rhs.value; };
-  constexpr bool operator!=(const Float8E5M2FNUZ_t& rhs) const noexcept { return value != rhs.value; };
-};
-
-static_assert(sizeof(Float8E5M2FNUZ_t) == sizeof(uint8_t), "Sizes must match");
-
-namespace detail {
-// This is used internally by the C++ API. This macro is to make it easy to generate overloaded methods for all of the various OrtRelease* functions for every Ort* type
-// This can't be done in the C API since C doesn't have function overloading.
-#define ORT_DEFINE_RELEASE(NAME) \
-  inline void OrtRelease(Ort##NAME* ptr) { GetApi().Release##NAME(ptr); }
-
-ORT_DEFINE_RELEASE(Allocator);
-ORT_DEFINE_RELEASE(MemoryInfo);
-ORT_DEFINE_RELEASE(CustomOpDomain);
-ORT_DEFINE_RELEASE(ThreadingOptions);
-ORT_DEFINE_RELEASE(Env);
-ORT_DEFINE_RELEASE(RunOptions);
-ORT_DEFINE_RELEASE(Session);
-ORT_DEFINE_RELEASE(SessionOptions);
-ORT_DEFINE_RELEASE(TensorTypeAndShapeInfo);
-ORT_DEFINE_RELEASE(SequenceTypeInfo);
-ORT_DEFINE_RELEASE(MapTypeInfo);
-ORT_DEFINE_RELEASE(TypeInfo);
-ORT_DEFINE_RELEASE(Value);
-ORT_DEFINE_RELEASE(ModelMetadata);
-ORT_DEFINE_RELEASE(IoBinding);
-ORT_DEFINE_RELEASE(ArenaCfg);
-ORT_DEFINE_RELEASE(Status);
-ORT_DEFINE_RELEASE(OpAttr);
-ORT_DEFINE_RELEASE(Op);
-ORT_DEFINE_RELEASE(KernelInfo);
-
-#undef ORT_DEFINE_RELEASE
-
-/** \brief This is a tagging template type. Use it with Base<T> to indicate that the C++ interface object
- *   has no ownership of the underlying C object.
- */
-template <typename T>
-struct Unowned {
-  using Type = T;
-};
-
-/** \brief Used internally by the C++ API. C++ wrapper types inherit from this.
- *   This is a zero cost abstraction to wrap the C API objects and delete them on destruction.
- *
- * All of the C++ classes
- *  a) serve as containers for pointers to objects that are created by the underlying C API.
- *     Their size is just a pointer size, no need to dynamically allocate them. Use them by value.
- *  b) Each of struct XXXX, XXX instances function as smart pointers to the underlying C API objects.
- *     they would release objects owned automatically when going out of scope, they are move-only.
- *  c) ConstXXXX and UnownedXXX structs function as non-owning, copyable containers for the above pointers.
- *     ConstXXXX allow calling const interfaces only. They give access to objects that are owned by somebody else
- *     such as Onnxruntime or instances of XXXX classes.
- *  d) serve convenient interfaces that return C++ objects and further enhance exception and type safety so they can be used
- *     in C++ code.
- *
- */
-
-/// <summary>
-/// This is a non-const pointer holder that is move-only. Disposes of the pointer on destruction.
-/// </summary>
-template <typename T>
-struct Base {
-  using contained_type = T;
-
-  constexpr Base() = default;
-  constexpr explicit Base(contained_type* p) noexcept : p_{p} {}
-  ~Base() { OrtRelease(p_); }
-
-  Base(const Base&) = delete;
-  Base& operator=(const Base&) = delete;
-
-  Base(Base&& v) noexcept : p_{v.p_} { v.p_ = nullptr; }
-  Base& operator=(Base&& v) noexcept {
-    OrtRelease(p_);
-    p_ = v.release();
-    return *this;
-  }
-
-  constexpr operator contained_type*() const noexcept { return p_; }
-
-  /// \brief Relinquishes ownership of the contained C object pointer
-  /// The underlying object is not destroyed
-  contained_type* release() {
-    T* p = p_;
-    p_ = nullptr;
-    return p;
-  }
-
- protected:
-  contained_type* p_{};
-};
-
-// Undefined. For const types use Base<Unowned<const T>>
-template <typename T>
-struct Base<const T>;
-
-/// <summary>
-/// Covers unowned pointers owned by either the ORT
-/// or some other instance of CPP wrappers.
-/// Used for ConstXXX and UnownedXXXX types that are copyable.
-/// Also convenient to wrap raw OrtXX pointers .
-/// </summary>
-/// <typeparam name="T"></typeparam>
-template <typename T>
-struct Base<Unowned<T>> {
-  using contained_type = typename Unowned<T>::Type;
-
-  constexpr Base() = default;
-  constexpr explicit Base(contained_type* p) noexcept : p_{p} {}
-
-  ~Base() = default;
-
-  Base(const Base&) = default;
-  Base& operator=(const Base&) = default;
-
-  Base(Base&& v) noexcept : p_{v.p_} { v.p_ = nullptr; }
-  Base& operator=(Base&& v) noexcept {
-    p_ = nullptr;
-    std::swap(p_, v.p_);
-    return *this;
-  }
-
-  constexpr operator contained_type*() const noexcept { return p_; }
-
- protected:
-  contained_type* p_{};
-};
-
-// Light functor to release memory with OrtAllocator
-struct AllocatedFree {
-  OrtAllocator* allocator_;
-  explicit AllocatedFree(OrtAllocator* allocator)
-      : allocator_(allocator) {}
-  void operator()(void* ptr) const {
-    if (ptr) allocator_->Free(allocator_, ptr);
-  }
-};
-
-}  // namespace detail
-
-struct AllocatorWithDefaultOptions;
-struct Env;
-struct TypeInfo;
-struct Value;
-struct ModelMetadata;
-
-/** \brief unique_ptr typedef used to own strings allocated by OrtAllocators
- *  and release them at the end of the scope. The lifespan of the given allocator
- *  must eclipse the lifespan of AllocatedStringPtr instance
- */
-using AllocatedStringPtr = std::unique_ptr<char, detail::AllocatedFree>;
-
-/** \brief The Status that holds ownership of OrtStatus received from C API
- *  Use it to safely destroy OrtStatus* returned from the C API. Use appropriate
- *  constructors to construct an instance of a Status object from exceptions.
- */
-struct Status : detail::Base<OrtStatus> {
-  explicit Status(std::nullptr_t) noexcept {}               ///< Create an empty object, must be assigned a valid one to be used
-  explicit Status(OrtStatus* status) noexcept;              ///< Takes ownership of OrtStatus instance returned from the C API.
-  explicit Status(const Exception&) noexcept;               ///< Creates status instance out of exception
-  explicit Status(const std::exception&) noexcept;          ///< Creates status instance out of exception
-  Status(const char* message, OrtErrorCode code) noexcept;  ///< Creates status instance out of null-terminated string message.
-  std::string GetErrorMessage() const;
-  OrtErrorCode GetErrorCode() const;
-  bool IsOK() const noexcept;  ///< Returns true if instance represents an OK (non-error) status.
-};
-
-/** \brief The ThreadingOptions
- *
- * The ThreadingOptions used for set global threadpools' options of The Env.
- */
-struct ThreadingOptions : detail::Base<OrtThreadingOptions> {
-  /// \brief Wraps OrtApi::CreateThreadingOptions
-  ThreadingOptions();
-
-  /// \brief Wraps OrtApi::SetGlobalIntraOpNumThreads
-  ThreadingOptions& SetGlobalIntraOpNumThreads(int intra_op_num_threads);
-
-  /// \brief Wraps OrtApi::SetGlobalInterOpNumThreads
-  ThreadingOptions& SetGlobalInterOpNumThreads(int inter_op_num_threads);
-
-  /// \brief Wraps OrtApi::SetGlobalSpinControl
-  ThreadingOptions& SetGlobalSpinControl(int allow_spinning);
-
-  /// \brief Wraps OrtApi::SetGlobalDenormalAsZero
-  ThreadingOptions& SetGlobalDenormalAsZero();
-
-  /// \brief Wraps OrtApi::SetGlobalCustomCreateThreadFn
-  ThreadingOptions& SetGlobalCustomCreateThreadFn(OrtCustomCreateThreadFn ort_custom_create_thread_fn);
-
-  /// \brief Wraps OrtApi::SetGlobalCustomThreadCreationOptions
-  ThreadingOptions& SetGlobalCustomThreadCreationOptions(void* ort_custom_thread_creation_options);
-
-  /// \brief Wraps OrtApi::SetGlobalCustomJoinThreadFn
-  ThreadingOptions& SetGlobalCustomJoinThreadFn(OrtCustomJoinThreadFn ort_custom_join_thread_fn);
-};
-
-/** \brief The Env (Environment)
- *
- * The Env holds the logging state used by all other objects.
- * <b>Note:</b> One Env must be created before using any other Onnxruntime functionality
- */
-struct Env : detail::Base<OrtEnv> {
-  explicit Env(std::nullptr_t) {}  ///< Create an empty Env object, must be assigned a valid one to be used
-
-  /// \brief Wraps OrtApi::CreateEnv
-  Env(OrtLoggingLevel logging_level = ORT_LOGGING_LEVEL_WARNING, _In_ const char* logid = "");
-
-  /// \brief Wraps OrtApi::CreateEnvWithCustomLogger
-  Env(OrtLoggingLevel logging_level, const char* logid, OrtLoggingFunction logging_function, void* logger_param);
-
-  /// \brief Wraps OrtApi::CreateEnvWithGlobalThreadPools
-  Env(const OrtThreadingOptions* tp_options, OrtLoggingLevel logging_level = ORT_LOGGING_LEVEL_WARNING, _In_ const char* logid = "");
-
-  /// \brief Wraps OrtApi::CreateEnvWithCustomLoggerAndGlobalThreadPools
-  Env(const OrtThreadingOptions* tp_options, OrtLoggingFunction logging_function, void* logger_param,
-      OrtLoggingLevel logging_level = ORT_LOGGING_LEVEL_WARNING, _In_ const char* logid = "");
-
-  /// \brief C Interop Helper
-  explicit Env(OrtEnv* p) : Base<OrtEnv>{p} {}
-
-  Env& EnableTelemetryEvents();   ///< Wraps OrtApi::EnableTelemetryEvents
-  Env& DisableTelemetryEvents();  ///< Wraps OrtApi::DisableTelemetryEvents
-
-  Env& UpdateEnvWithCustomLogLevel(OrtLoggingLevel log_severity_level);  ///< Wraps OrtApi::UpdateEnvWithCustomLogLevel
-
-  Env& CreateAndRegisterAllocator(const OrtMemoryInfo* mem_info, const OrtArenaCfg* arena_cfg);  ///< Wraps OrtApi::CreateAndRegisterAllocator
-
-  Env& CreateAndRegisterAllocatorV2(const std::string& provider_type, const OrtMemoryInfo* mem_info, const std::unordered_map<std::string, std::string>& options, const OrtArenaCfg* arena_cfg);  ///< Wraps OrtApi::CreateAndRegisterAllocatorV2
-};
-
-/** \brief Custom Op Domain
- *
- */
-struct CustomOpDomain : detail::Base<OrtCustomOpDomain> {
-  explicit CustomOpDomain(std::nullptr_t) {}  ///< Create an empty CustomOpDomain object, must be assigned a valid one to be used
-
-  /// \brief Wraps OrtApi::CreateCustomOpDomain
-  explicit CustomOpDomain(const char* domain);
-
-  // This does not take ownership of the op, simply registers it.
-  void Add(const OrtCustomOp* op);  ///< Wraps CustomOpDomain_Add
-};
-
-/** \brief RunOptions
- *
- */
-struct RunOptions : detail::Base<OrtRunOptions> {
-  explicit RunOptions(std::nullptr_t) {}  ///< Create an empty RunOptions object, must be assigned a valid one to be used
-  RunOptions();                           ///< Wraps OrtApi::CreateRunOptions
-
-  RunOptions& SetRunLogVerbosityLevel(int);  ///< Wraps OrtApi::RunOptionsSetRunLogVerbosityLevel
-  int GetRunLogVerbosityLevel() const;       ///< Wraps OrtApi::RunOptionsGetRunLogVerbosityLevel
-
-  RunOptions& SetRunLogSeverityLevel(int);  ///< Wraps OrtApi::RunOptionsSetRunLogSeverityLevel
-  int GetRunLogSeverityLevel() const;       ///< Wraps OrtApi::RunOptionsGetRunLogSeverityLevel
-
-  RunOptions& SetRunTag(const char* run_tag);  ///< wraps OrtApi::RunOptionsSetRunTag
-  const char* GetRunTag() const;               ///< Wraps OrtApi::RunOptionsGetRunTag
-
-  RunOptions& AddConfigEntry(const char* config_key, const char* config_value);  ///< Wraps OrtApi::AddRunConfigEntry
-
-  /** \brief Terminates all currently executing Session::Run calls that were made using this RunOptions instance
-   *
-   * If a currently executing session needs to be force terminated, this can be called from another thread to force it to fail with an error
-   * Wraps OrtApi::RunOptionsSetTerminate
-   */
-  RunOptions& SetTerminate();
-
-  /** \brief Clears the terminate flag so this RunOptions instance can be used in a new Session::Run call without it instantly terminating
-   *
-   * Wraps OrtApi::RunOptionsUnsetTerminate
-   */
-  RunOptions& UnsetTerminate();
-};
-
-namespace detail {
-// Utility function that returns a SessionOption config entry key for a specific custom operator.
-// Ex: custom_op.[custom_op_name].[config]
-std::string MakeCustomOpConfigEntryKey(const char* custom_op_name, const char* config);
-}  // namespace detail
-
-/// <summary>
-/// Class that represents session configuration entries for one or more custom operators.
-///
-/// Example:
-///   Ort::CustomOpConfigs op_configs;
-///   op_configs.AddConfig("my_custom_op", "device_type", "CPU");
-///
-/// Passed to Ort::SessionOptions::RegisterCustomOpsLibrary.
-/// </summary>
-struct CustomOpConfigs {
-  CustomOpConfigs() = default;
-  ~CustomOpConfigs() = default;
-  CustomOpConfigs(const CustomOpConfigs&) = default;
-  CustomOpConfigs& operator=(const CustomOpConfigs&) = default;
-  CustomOpConfigs(CustomOpConfigs&& o) = default;
-  CustomOpConfigs& operator=(CustomOpConfigs&& o) = default;
-
-  /** \brief Adds a session configuration entry/value for a specific custom operator.
-   *
-   * \param custom_op_name The name of the custom operator for which to add a configuration entry.
-   *                       Must match the name returned by the CustomOp's GetName() method.
-   * \param config_key The name of the configuration entry.
-   * \param config_value The value of the configuration entry.
-   * \return A reference to this object to enable call chaining.
-   */
-  CustomOpConfigs& AddConfig(const char* custom_op_name, const char* config_key, const char* config_value);
-
-  /** \brief Returns a flattened map of custom operator configuration entries and their values.
-   *
-   * The keys has been flattened to include both the custom operator name and the configuration entry key name.
-   * For example, a prior call to AddConfig("my_op", "key", "value") corresponds to the flattened key/value pair
-   * {"my_op.key", "value"}.
-   *
-   * \return An unordered map of flattened configurations.
-   */
-  const std::unordered_map<std::string, std::string>& GetFlattenedConfigs() const;
-
- private:
-  std::unordered_map<std::string, std::string> flat_configs_;
-};
-
-/** \brief Options object used when creating a new Session object
- *
- * Wraps ::OrtSessionOptions object and methods
- */
-
-struct SessionOptions;
-
-namespace detail {
-// we separate const-only methods because passing const ptr to non-const methods
-// is only discovered when inline methods are compiled which is counter-intuitive
-template <typename T>
-struct ConstSessionOptionsImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  SessionOptions Clone() const;  ///< Creates and returns a copy of this SessionOptions object. Wraps OrtApi::CloneSessionOptions
-
-  std::string GetConfigEntry(const char* config_key) const;  ///< Wraps OrtApi::GetSessionConfigEntry
-  bool HasConfigEntry(const char* config_key) const;         ///< Wraps OrtApi::HasSessionConfigEntry
-  std::string GetConfigEntryOrDefault(const char* config_key, const std::string& def);
-};
-
-template <typename T>
-struct SessionOptionsImpl : ConstSessionOptionsImpl<T> {
-  using B = ConstSessionOptionsImpl<T>;
-  using B::B;
-
-  SessionOptionsImpl& SetIntraOpNumThreads(int intra_op_num_threads);                              ///< Wraps OrtApi::SetIntraOpNumThreads
-  SessionOptionsImpl& SetInterOpNumThreads(int inter_op_num_threads);                              ///< Wraps OrtApi::SetInterOpNumThreads
-  SessionOptionsImpl& SetGraphOptimizationLevel(GraphOptimizationLevel graph_optimization_level);  ///< Wraps OrtApi::SetSessionGraphOptimizationLevel
-  SessionOptionsImpl& SetDeterministicCompute(bool value);                                         ///< Wraps OrtApi::SetDeterministicCompute
-
-  SessionOptionsImpl& EnableCpuMemArena();   ///< Wraps OrtApi::EnableCpuMemArena
-  SessionOptionsImpl& DisableCpuMemArena();  ///< Wraps OrtApi::DisableCpuMemArena
-
-  SessionOptionsImpl& SetOptimizedModelFilePath(const ORTCHAR_T* optimized_model_file);  ///< Wraps OrtApi::SetOptimizedModelFilePath
-
-  SessionOptionsImpl& EnableProfiling(const ORTCHAR_T* profile_file_prefix);  ///< Wraps OrtApi::EnableProfiling
-  SessionOptionsImpl& DisableProfiling();                                     ///< Wraps OrtApi::DisableProfiling
-
-  SessionOptionsImpl& EnableOrtCustomOps();  ///< Wraps OrtApi::EnableOrtCustomOps
-
-  SessionOptionsImpl& EnableMemPattern();   ///< Wraps OrtApi::EnableMemPattern
-  SessionOptionsImpl& DisableMemPattern();  ///< Wraps OrtApi::DisableMemPattern
-
-  SessionOptionsImpl& SetExecutionMode(ExecutionMode execution_mode);  ///< Wraps OrtApi::SetSessionExecutionMode
-
-  SessionOptionsImpl& SetLogId(const char* logid);     ///< Wraps OrtApi::SetSessionLogId
-  SessionOptionsImpl& SetLogSeverityLevel(int level);  ///< Wraps OrtApi::SetSessionLogSeverityLevel
-
-  SessionOptionsImpl& Add(OrtCustomOpDomain* custom_op_domain);  ///< Wraps OrtApi::AddCustomOpDomain
-
-  SessionOptionsImpl& DisablePerSessionThreads();  ///< Wraps OrtApi::DisablePerSessionThreads
-
-  SessionOptionsImpl& AddConfigEntry(const char* config_key, const char* config_value);  ///< Wraps OrtApi::AddSessionConfigEntry
-
-  SessionOptionsImpl& AddInitializer(const char* name, const OrtValue* ort_val);                                             ///< Wraps OrtApi::AddInitializer
-  SessionOptionsImpl& AddExternalInitializers(const std::vector<std::string>& names, const std::vector<Value>& ort_values);  ///< Wraps OrtApi::AddExternalInitializers
-
-  SessionOptionsImpl& AppendExecutionProvider_CUDA(const OrtCUDAProviderOptions& provider_options);          ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_CUDA
-  SessionOptionsImpl& AppendExecutionProvider_CUDA_V2(const OrtCUDAProviderOptionsV2& provider_options);     ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_CUDA_V2
-  SessionOptionsImpl& AppendExecutionProvider_ROCM(const OrtROCMProviderOptions& provider_options);          ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_ROCM
-  SessionOptionsImpl& AppendExecutionProvider_OpenVINO(const OrtOpenVINOProviderOptions& provider_options);  ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_OpenVINO
-  ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_OpenVINO_V2
-  SessionOptionsImpl& AppendExecutionProvider_OpenVINO_V2(const std::unordered_map<std::string, std::string>& provider_options = {});
-  SessionOptionsImpl& AppendExecutionProvider_TensorRT(const OrtTensorRTProviderOptions& provider_options);       ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_TensorRT
-  SessionOptionsImpl& AppendExecutionProvider_TensorRT_V2(const OrtTensorRTProviderOptionsV2& provider_options);  ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_TensorRT
-  SessionOptionsImpl& AppendExecutionProvider_MIGraphX(const OrtMIGraphXProviderOptions& provider_options);       ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_MIGraphX
-  ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_CANN
-  SessionOptionsImpl& AppendExecutionProvider_CANN(const OrtCANNProviderOptions& provider_options);
-  ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_Dnnl
-  SessionOptionsImpl& AppendExecutionProvider_Dnnl(const OrtDnnlProviderOptions& provider_options);
-  /// Wraps OrtApi::SessionOptionsAppendExecutionProvider. Currently supports QNN, SNPE and XNNPACK.
-  SessionOptionsImpl& AppendExecutionProvider(const std::string& provider_name,
-                                              const std::unordered_map<std::string, std::string>& provider_options = {});
-
-  SessionOptionsImpl& SetCustomCreateThreadFn(OrtCustomCreateThreadFn ort_custom_create_thread_fn);  ///< Wraps OrtApi::SessionOptionsSetCustomCreateThreadFn
-  SessionOptionsImpl& SetCustomThreadCreationOptions(void* ort_custom_thread_creation_options);      ///< Wraps OrtApi::SessionOptionsSetCustomThreadCreationOptions
-  SessionOptionsImpl& SetCustomJoinThreadFn(OrtCustomJoinThreadFn ort_custom_join_thread_fn);        ///< Wraps OrtApi::SessionOptionsSetCustomJoinThreadFn
-
-  ///< Registers the custom operator from the specified shared library via OrtApi::RegisterCustomOpsLibrary_V2.
-  ///< The custom operator configurations are optional. If provided, custom operator configs are set via
-  ///< OrtApi::AddSessionConfigEntry.
-  SessionOptionsImpl& RegisterCustomOpsLibrary(const ORTCHAR_T* library_name, const CustomOpConfigs& custom_op_configs = {});
-
-  SessionOptionsImpl& RegisterCustomOpsUsingFunction(const char* function_name);  ///< Wraps OrtApi::RegisterCustomOpsUsingFunction
-};
-}  // namespace detail
-
-using UnownedSessionOptions = detail::SessionOptionsImpl<detail::Unowned<OrtSessionOptions>>;
-using ConstSessionOptions = detail::ConstSessionOptionsImpl<detail::Unowned<const OrtSessionOptions>>;
-
-/** \brief Wrapper around ::OrtSessionOptions
- *
- */
-struct SessionOptions : detail::SessionOptionsImpl<OrtSessionOptions> {
-  explicit SessionOptions(std::nullptr_t) {}                                                   ///< Create an empty SessionOptions object, must be assigned a valid one to be used
-  SessionOptions();                                                                            ///< Wraps OrtApi::CreateSessionOptions
-  explicit SessionOptions(OrtSessionOptions* p) : SessionOptionsImpl<OrtSessionOptions>{p} {}  ///< Used for interop with the C API
-  UnownedSessionOptions GetUnowned() const { return UnownedSessionOptions{this->p_}; }
-  ConstSessionOptions GetConst() const { return ConstSessionOptions{this->p_}; }
-};
-
-/** \brief Wrapper around ::OrtModelMetadata
- *
- */
-struct ModelMetadata : detail::Base<OrtModelMetadata> {
-  explicit ModelMetadata(std::nullptr_t) {}                                   ///< Create an empty ModelMetadata object, must be assigned a valid one to be used
-  explicit ModelMetadata(OrtModelMetadata* p) : Base<OrtModelMetadata>{p} {}  ///< Used for interop with the C API
-
-  /** \brief Returns a copy of the producer name.
-   *
-   * \param allocator to allocate memory for the copy of the name returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetProducerNameAllocated(OrtAllocator* allocator) const;  ///< Wraps OrtApi::ModelMetadataGetProducerName
-
-  /** \brief Returns a copy of the graph name.
-   *
-   * \param allocator to allocate memory for the copy of the name returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetGraphNameAllocated(OrtAllocator* allocator) const;  ///< Wraps OrtApi::ModelMetadataGetGraphName
-
-  /** \brief Returns a copy of the domain name.
-   *
-   * \param allocator to allocate memory for the copy of the name returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetDomainAllocated(OrtAllocator* allocator) const;  ///< Wraps OrtApi::ModelMetadataGetDomain
-
-  /** \brief Returns a copy of the description.
-   *
-   * \param allocator to allocate memory for the copy of the string returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetDescriptionAllocated(OrtAllocator* allocator) const;  ///< Wraps OrtApi::ModelMetadataGetDescription
-
-  /** \brief Returns a copy of the graph description.
-   *
-   * \param allocator to allocate memory for the copy of the string returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetGraphDescriptionAllocated(OrtAllocator* allocator) const;  ///< Wraps OrtApi::ModelMetadataGetGraphDescription
-
-  /** \brief Returns a vector of copies of the custom metadata keys.
-   *
-   * \param allocator to allocate memory for the copy of the string returned
-   * \return a instance std::vector of smart pointers that would deallocate the buffers when out of scope.
-   *  The OrtAllocator instance must be valid at the point of memory release.
-   */
-  std::vector<AllocatedStringPtr> GetCustomMetadataMapKeysAllocated(OrtAllocator* allocator) const;  ///< Wraps OrtApi::ModelMetadataGetCustomMetadataMapKeys
-
-  /** \brief Looks up a value by a key in the Custom Metadata map
-   *
-   * \param key zero terminated string key to lookup
-   * \param allocator to allocate memory for the copy of the string returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  maybe nullptr if key is not found.
-   *
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr LookupCustomMetadataMapAllocated(const char* key, OrtAllocator* allocator) const;  ///< Wraps OrtApi::ModelMetadataLookupCustomMetadataMap
-
-  int64_t GetVersion() const;  ///< Wraps OrtApi::ModelMetadataGetVersion
-};
-
-struct IoBinding;
-
-namespace detail {
-
-// we separate const-only methods because passing const ptr to non-const methods
-// is only discovered when inline methods are compiled which is counter-intuitive
-template <typename T>
-struct ConstSessionImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  size_t GetInputCount() const;                   ///< Returns the number of model inputs
-  size_t GetOutputCount() const;                  ///< Returns the number of model outputs
-  size_t GetOverridableInitializerCount() const;  ///< Returns the number of inputs that have defaults that can be overridden
-
-  /** \brief Returns a copy of input name at the specified index.
-   *
-   * \param index must less than the value returned by GetInputCount()
-   * \param allocator to allocate memory for the copy of the name returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetInputNameAllocated(size_t index, OrtAllocator* allocator) const;
-
-  /** \brief Returns a copy of output name at then specified index.
-   *
-   * \param index must less than the value returned by GetOutputCount()
-   * \param allocator to allocate memory for the copy of the name returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetOutputNameAllocated(size_t index, OrtAllocator* allocator) const;
-
-  /** \brief Returns a copy of the overridable initializer name at then specified index.
-   *
-   * \param index must less than the value returned by GetOverridableInitializerCount()
-   * \param allocator to allocate memory for the copy of the name returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr GetOverridableInitializerNameAllocated(size_t index, OrtAllocator* allocator) const;  ///< Wraps OrtApi::SessionGetOverridableInitializerName
-
-  uint64_t GetProfilingStartTimeNs() const;  ///< Wraps OrtApi::SessionGetProfilingStartTimeNs
-  ModelMetadata GetModelMetadata() const;    ///< Wraps OrtApi::SessionGetModelMetadata
-
-  TypeInfo GetInputTypeInfo(size_t index) const;                   ///< Wraps OrtApi::SessionGetInputTypeInfo
-  TypeInfo GetOutputTypeInfo(size_t index) const;                  ///< Wraps OrtApi::SessionGetOutputTypeInfo
-  TypeInfo GetOverridableInitializerTypeInfo(size_t index) const;  ///< Wraps OrtApi::SessionGetOverridableInitializerTypeInfo
-};
-
-template <typename T>
-struct SessionImpl : ConstSessionImpl<T> {
-  using B = ConstSessionImpl<T>;
-  using B::B;
-
-  /** \brief Run the model returning results in an Ort allocated vector.
-   *
-   * Wraps OrtApi::Run
-   *
-   * The caller provides a list of inputs and a list of the desired outputs to return.
-   *
-   * See the output logs for more information on warnings/errors that occur while processing the model.
-   * Common errors are.. (TODO)
-   *
-   * \param[in] run_options
-   * \param[in] input_names Array of null terminated strings of length input_count that is the list of input names
-   * \param[in] input_values Array of Value objects of length input_count that is the list of input values
-   * \param[in] input_count Number of inputs (the size of the input_names & input_values arrays)
-   * \param[in] output_names Array of C style strings of length output_count that is the list of output names
-   * \param[in] output_count Number of outputs (the size of the output_names array)
-   * \return A std::vector of Value objects that directly maps to the output_names array (eg. output_name[0] is the first entry of the returned vector)
-   */
-  std::vector<Value> Run(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count,
-                         const char* const* output_names, size_t output_count);
-
-  /** \brief Run the model returning results in user provided outputs
-   * Same as Run(const RunOptions&, const char* const*, const Value*, size_t,const char* const*, size_t)
-   */
-  void Run(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count,
-           const char* const* output_names, Value* output_values, size_t output_count);
-
-  void Run(const RunOptions& run_options, const IoBinding&);  ///< Wraps OrtApi::RunWithBinding
-
-  /** \brief Run the model asynchronously in a thread owned by intra op thread pool
-   *
-   * Wraps OrtApi::RunAsync
-   *
-   * \param[in] run_options
-   * \param[in] input_names Array of null terminated UTF8 encoded strings of the input names
-   * \param[in] input_values Array of Value objects of length input_count
-   * \param[in] input_count Number of elements in the input_names and inputs arrays
-   * \param[in] output_names Array of null terminated UTF8 encoded strings of the output names
-   * \param[out] output_values Array of provided Values to be filled with outputs.
-   *             On calling RunAsync, output_values[i] could either be initialized by a null pointer or a preallocated OrtValue*.
-   *             Later, on invoking the callback, each output_values[i] of null will be filled with an OrtValue* allocated by onnxruntime.
-   *             Then, an OrtValue** pointer will be casted from output_values, and pass to the callback.
-   *             NOTE: it is customer's duty to finally release output_values and each of its member,
-   *             regardless of whether the member (Ort::Value) is allocated by onnxruntime or preallocated by the customer.
-   * \param[in] output_count Number of elements in the output_names and outputs array
-   * \param[in] callback Callback function on model run completion
-   * \param[in] user_data User data that pass back to the callback
-   */
-  void RunAsync(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count,
-                const char* const* output_names, Value* output_values, size_t output_count, RunAsyncCallbackFn callback, void* user_data);
-
-  /** \brief End profiling and return a copy of the profiling file name.
-   *
-   * \param allocator to allocate memory for the copy of the string returned
-   * \return a instance of smart pointer that would deallocate the buffer when out of scope.
-   *  The OrtAllocator instances must be valid at the point of memory release.
-   */
-  AllocatedStringPtr EndProfilingAllocated(OrtAllocator* allocator);  ///< Wraps OrtApi::SessionEndProfiling
-};
-
-}  // namespace detail
-
-using ConstSession = detail::ConstSessionImpl<detail::Unowned<const OrtSession>>;
-using UnownedSession = detail::SessionImpl<detail::Unowned<OrtSession>>;
-
-/** \brief Wrapper around ::OrtSession
- *
- */
-struct Session : detail::SessionImpl<OrtSession> {
-  explicit Session(std::nullptr_t) {}                                                   ///< Create an empty Session object, must be assigned a valid one to be used
-  Session(const Env& env, const ORTCHAR_T* model_path, const SessionOptions& options);  ///< Wraps OrtApi::CreateSession
-  Session(const Env& env, const ORTCHAR_T* model_path, const SessionOptions& options,
-          OrtPrepackedWeightsContainer* prepacked_weights_container);                                        ///< Wraps OrtApi::CreateSessionWithPrepackedWeightsContainer
-  Session(const Env& env, const void* model_data, size_t model_data_length, const SessionOptions& options);  ///< Wraps OrtApi::CreateSessionFromArray
-  Session(const Env& env, const void* model_data, size_t model_data_length, const SessionOptions& options,
-          OrtPrepackedWeightsContainer* prepacked_weights_container);  ///< Wraps OrtApi::CreateSessionFromArrayWithPrepackedWeightsContainer
-
-  ConstSession GetConst() const { return ConstSession{this->p_}; }
-  UnownedSession GetUnowned() const { return UnownedSession{this->p_}; }
-};
-
-namespace detail {
-template <typename T>
-struct MemoryInfoImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  std::string GetAllocatorName() const;
-  OrtAllocatorType GetAllocatorType() const;
-  int GetDeviceId() const;
-  OrtMemoryInfoDeviceType GetDeviceType() const;
-  OrtMemType GetMemoryType() const;
-
-  template <typename U>
-  bool operator==(const MemoryInfoImpl<U>& o) const;
-};
-}  // namespace detail
-
-// Const object holder that does not own the underlying object
-using ConstMemoryInfo = detail::MemoryInfoImpl<detail::Unowned<const OrtMemoryInfo>>;
-
-/** \brief Wrapper around ::OrtMemoryInfo
- *
- */
-struct MemoryInfo : detail::MemoryInfoImpl<OrtMemoryInfo> {
-  static MemoryInfo CreateCpu(OrtAllocatorType type, OrtMemType mem_type1);
-  explicit MemoryInfo(std::nullptr_t) {}                                       ///< No instance is created
-  explicit MemoryInfo(OrtMemoryInfo* p) : MemoryInfoImpl<OrtMemoryInfo>{p} {}  ///< Take ownership of a pointer created by C Api
-  MemoryInfo(const char* name, OrtAllocatorType type, int id, OrtMemType mem_type);
-  ConstMemoryInfo GetConst() const { return ConstMemoryInfo{this->p_}; }
-};
-
-namespace detail {
-template <typename T>
-struct TensorTypeAndShapeInfoImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  ONNXTensorElementDataType GetElementType() const;  ///< Wraps OrtApi::GetTensorElementType
-  size_t GetElementCount() const;                    ///< Wraps OrtApi::GetTensorShapeElementCount
-
-  size_t GetDimensionsCount() const;  ///< Wraps OrtApi::GetDimensionsCount
-
-  /** \deprecated use GetShape() returning std::vector
-   * [[deprecated]]
-   * This interface is unsafe to use
-   */
-  [[deprecated("use GetShape()")]] void GetDimensions(int64_t* values, size_t values_count) const;  ///< Wraps OrtApi::GetDimensions
-
-  void GetSymbolicDimensions(const char** values, size_t values_count) const;  ///< Wraps OrtApi::GetSymbolicDimensions
-
-  std::vector<int64_t> GetShape() const;  ///< Uses GetDimensionsCount & GetDimensions to return a std::vector of the shape
-};
-
-}  // namespace detail
-
-using ConstTensorTypeAndShapeInfo = detail::TensorTypeAndShapeInfoImpl<detail::Unowned<const OrtTensorTypeAndShapeInfo>>;
-
-/** \brief Wrapper around ::OrtTensorTypeAndShapeInfo
- *
- */
-struct TensorTypeAndShapeInfo : detail::TensorTypeAndShapeInfoImpl<OrtTensorTypeAndShapeInfo> {
-  explicit TensorTypeAndShapeInfo(std::nullptr_t) {}                                                ///< Create an empty TensorTypeAndShapeInfo object, must be assigned a valid one to be used
-  explicit TensorTypeAndShapeInfo(OrtTensorTypeAndShapeInfo* p) : TensorTypeAndShapeInfoImpl{p} {}  ///< Used for interop with the C API
-  ConstTensorTypeAndShapeInfo GetConst() const { return ConstTensorTypeAndShapeInfo{this->p_}; }
-};
-
-namespace detail {
-template <typename T>
-struct SequenceTypeInfoImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-  TypeInfo GetSequenceElementType() const;  ///< Wraps OrtApi::GetSequenceElementType
-};
-
-}  // namespace detail
-
-using ConstSequenceTypeInfo = detail::SequenceTypeInfoImpl<detail::Unowned<const OrtSequenceTypeInfo>>;
-
-/** \brief Wrapper around ::OrtSequenceTypeInfo
- *
- */
-struct SequenceTypeInfo : detail::SequenceTypeInfoImpl<OrtSequenceTypeInfo> {
-  explicit SequenceTypeInfo(std::nullptr_t) {}                                                         ///< Create an empty SequenceTypeInfo object, must be assigned a valid one to be used
-  explicit SequenceTypeInfo(OrtSequenceTypeInfo* p) : SequenceTypeInfoImpl<OrtSequenceTypeInfo>{p} {}  ///< Used for interop with the C API
-  ConstSequenceTypeInfo GetConst() const { return ConstSequenceTypeInfo{this->p_}; }
-};
-
-namespace detail {
-template <typename T>
-struct OptionalTypeInfoImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-  TypeInfo GetOptionalElementType() const;  ///< Wraps OrtApi::CastOptionalTypeToContainedTypeInfo
-};
-
-}  // namespace detail
-
-// This is always owned by the TypeInfo and can only be obtained from it.
-using ConstOptionalTypeInfo = detail::OptionalTypeInfoImpl<detail::Unowned<const OrtOptionalTypeInfo>>;
-
-namespace detail {
-template <typename T>
-struct MapTypeInfoImpl : detail::Base<T> {
-  using B = Base<T>;
-  using B::B;
-  ONNXTensorElementDataType GetMapKeyType() const;  ///< Wraps OrtApi::GetMapKeyType
-  TypeInfo GetMapValueType() const;                 ///< Wraps OrtApi::GetMapValueType
-};
-
-}  // namespace detail
-
-using ConstMapTypeInfo = detail::MapTypeInfoImpl<detail::Unowned<const OrtMapTypeInfo>>;
-
-/** \brief Wrapper around ::OrtMapTypeInfo
- *
- */
-struct MapTypeInfo : detail::MapTypeInfoImpl<OrtMapTypeInfo> {
-  explicit MapTypeInfo(std::nullptr_t) {}                                          ///< Create an empty MapTypeInfo object, must be assigned a valid one to be used
-  explicit MapTypeInfo(OrtMapTypeInfo* p) : MapTypeInfoImpl<OrtMapTypeInfo>{p} {}  ///< Used for interop with the C API
-  ConstMapTypeInfo GetConst() const { return ConstMapTypeInfo{this->p_}; }
-};
-
-namespace detail {
-template <typename T>
-struct TypeInfoImpl : detail::Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  ConstTensorTypeAndShapeInfo GetTensorTypeAndShapeInfo() const;  ///< Wraps OrtApi::CastTypeInfoToTensorInfo
-  ConstSequenceTypeInfo GetSequenceTypeInfo() const;              ///< Wraps OrtApi::CastTypeInfoToSequenceTypeInfo
-  ConstMapTypeInfo GetMapTypeInfo() const;                        ///< Wraps OrtApi::CastTypeInfoToMapTypeInfo
-  ConstOptionalTypeInfo GetOptionalTypeInfo() const;              ///< wraps OrtApi::CastTypeInfoToOptionalTypeInfo
-
-  ONNXType GetONNXType() const;
-};
-}  // namespace detail
-
-/// <summary>
-/// Contains a constant, unowned OrtTypeInfo that can be copied and passed around by value.
-/// Provides access to const OrtTypeInfo APIs.
-/// </summary>
-using ConstTypeInfo = detail::TypeInfoImpl<detail::Unowned<const OrtTypeInfo>>;
-
-/// <summary>
-/// Type information that may contain either TensorTypeAndShapeInfo or
-/// the information about contained sequence or map depending on the ONNXType.
-/// </summary>
-struct TypeInfo : detail::TypeInfoImpl<OrtTypeInfo> {
-  explicit TypeInfo(std::nullptr_t) {}                                 ///< Create an empty TypeInfo object, must be assigned a valid one to be used
-  explicit TypeInfo(OrtTypeInfo* p) : TypeInfoImpl<OrtTypeInfo>{p} {}  ///< C API Interop
-
-  ConstTypeInfo GetConst() const { return ConstTypeInfo{this->p_}; }
-};
-
-namespace detail {
-// This structure is used to feed  sparse tensor values
-// information for use with FillSparseTensor<Format>() API
-// if the data type for the sparse tensor values is numeric
-// use data.p_data, otherwise, use data.str pointer to feed
-// values. data.str is an array of const char* that are zero terminated.
-// number of strings in the array must match shape size.
-// For fully sparse tensors use shape {0} and set p_data/str
-// to nullptr.
-struct OrtSparseValuesParam {
-  const int64_t* values_shape;
-  size_t values_shape_len;
-  union {
-    const void* p_data;
-    const char** str;
-  } data;
-};
-
-// Provides a way to pass shape in a single
-// argument
-struct Shape {
-  const int64_t* shape;
-  size_t shape_len;
-};
-
-template <typename T>
-struct ConstValueImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  /// <summary>
-  /// Obtains a pointer to a user defined data for experimental purposes
-  /// </summary>
-  template <typename R>
-  void GetOpaqueData(const char* domain, const char* type_name, R&) const;  ///< Wraps OrtApi::GetOpaqueValue
-
-  bool IsTensor() const;  ///< Returns true if Value is a tensor, false for other types like map/sequence/etc
-  bool HasValue() const;  /// < Return true if OrtValue contains data and returns false if the OrtValue is a None
-
-  size_t GetCount() const;  // If a non tensor, returns 2 for map and N for sequence, where N is the number of elements
-  Value GetValue(int index, OrtAllocator* allocator) const;
-
-  /// <summary>
-  /// This API returns a full length of string data contained within either a tensor or a sparse Tensor.
-  /// For sparse tensor it returns a full length of stored non-empty strings (values). The API is useful
-  /// for allocating necessary memory and calling GetStringTensorContent().
-  /// </summary>
-  /// <returns>total length of UTF-8 encoded bytes contained. No zero terminators counted.</returns>
-  size_t GetStringTensorDataLength() const;
-
-  /// <summary>
-  /// The API copies all of the UTF-8 encoded string data contained within a tensor or a sparse tensor
-  /// into a supplied buffer. Use GetStringTensorDataLength() to find out the length of the buffer to allocate.
-  /// The user must also allocate offsets buffer with the number of entries equal to that of the contained
-  /// strings.
-  ///
-  /// Strings are always assumed to be on CPU, no X-device copy.
-  /// </summary>
-  /// <param name="buffer">user allocated buffer</param>
-  /// <param name="buffer_length">length in bytes of the allocated buffer</param>
-  /// <param name="offsets">a pointer to the offsets user allocated buffer</param>
-  /// <param name="offsets_count">count of offsets, must be equal to the number of strings contained.
-  ///   that can be obtained from the shape of the tensor or from GetSparseTensorValuesTypeAndShapeInfo()
-  ///   for sparse tensors</param>
-  void GetStringTensorContent(void* buffer, size_t buffer_length, size_t* offsets, size_t offsets_count) const;
-
-  /// <summary>
-  /// Returns a const typed pointer to the tensor contained data.
-  /// No type checking is performed, the caller must ensure the type matches the tensor type.
-  /// </summary>
-  /// <typeparam name="T"></typeparam>
-  /// <returns>const pointer to data, no copies made</returns>
-  template <typename R>
-  const R* GetTensorData() const;  ///< Wraps OrtApi::GetTensorMutableData   /// <summary>
-
-  /// <summary>
-  /// Returns a non-typed pointer to a tensor contained data.
-  /// </summary>
-  /// <returns>const pointer to data, no copies made</returns>
-  const void* GetTensorRawData() const;
-
-  /// <summary>
-  /// The API returns type information for data contained in a tensor. For sparse
-  /// tensors it returns type information for contained non-zero values.
-  /// It returns dense shape for sparse tensors.
-  /// </summary>
-  /// <returns>TypeInfo</returns>
-  TypeInfo GetTypeInfo() const;
-
-  /// <summary>
-  /// The API returns type information for data contained in a tensor. For sparse
-  /// tensors it returns type information for contained non-zero values.
-  /// It returns dense shape for sparse tensors.
-  /// </summary>
-  /// <returns>TensorTypeAndShapeInfo</returns>
-  TensorTypeAndShapeInfo GetTensorTypeAndShapeInfo() const;
-
-  /// <summary>
-  /// This API returns information about the memory allocation used to hold data.
-  /// </summary>
-  /// <returns>Non owning instance of MemoryInfo</returns>
-  ConstMemoryInfo GetTensorMemoryInfo() const;
-
-  /// <summary>
-  /// The API copies UTF-8 encoded bytes for the requested string element
-  /// contained within a tensor or a sparse tensor into a provided buffer.
-  /// Use GetStringTensorElementLength() to obtain the length of the buffer to allocate.
-  /// </summary>
-  /// <param name="buffer_length"></param>
-  /// <param name="element_index"></param>
-  /// <param name="buffer"></param>
-  void GetStringTensorElement(size_t buffer_length, size_t element_index, void* buffer) const;
-
-  /// <summary>
-  /// Returns string tensor UTF-8 encoded string element.
-  /// Use of this API is recommended over GetStringTensorElement() that takes void* buffer pointer.
-  /// </summary>
-  /// <param name="element_index"></param>
-  /// <returns>std::string</returns>
-  std::string GetStringTensorElement(size_t element_index) const;
-
-  /// <summary>
-  /// The API returns a byte length of UTF-8 encoded string element
-  /// contained in either a tensor or a spare tensor values.
-  /// </summary>
-  /// <param name="element_index"></param>
-  /// <returns>byte length for the specified string element</returns>
-  size_t GetStringTensorElementLength(size_t element_index) const;
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-  /// <summary>
-  /// The API returns the sparse data format this OrtValue holds in a sparse tensor.
-  /// If the sparse tensor was not fully constructed, i.e. Use*() or Fill*() API were not used
-  /// the value returned is ORT_SPARSE_UNDEFINED.
-  /// </summary>
-  /// <returns>Format enum</returns>
-  OrtSparseFormat GetSparseFormat() const;
-
-  /// <summary>
-  /// The API returns type and shape information for stored non-zero values of the
-  /// sparse tensor. Use GetSparseTensorValues() to obtain values buffer pointer.
-  /// </summary>
-  /// <returns>TensorTypeAndShapeInfo values information</returns>
-  TensorTypeAndShapeInfo GetSparseTensorValuesTypeAndShapeInfo() const;
-
-  /// <summary>
-  /// The API returns type and shape information for the specified indices. Each supported
-  /// indices have their own enum values even if a give format has more than one kind of indices.
-  /// Use GetSparseTensorIndicesData() to obtain pointer to indices buffer.
-  /// </summary>
-  /// <param name="format">enum requested</param>
-  /// <returns>type and shape information</returns>
-  TensorTypeAndShapeInfo GetSparseTensorIndicesTypeShapeInfo(OrtSparseIndicesFormat format) const;
-
-  /// <summary>
-  /// The API retrieves a pointer to the internal indices buffer. The API merely performs
-  /// a convenience data type casting on the return type pointer. Make sure you are requesting
-  /// the right type, use GetSparseTensorIndicesTypeShapeInfo();
-  /// </summary>
-  /// <typeparam name="T">type to cast to</typeparam>
-  /// <param name="indices_format">requested indices kind</param>
-  /// <param name="num_indices">number of indices entries</param>
-  /// <returns>Pinter to the internal sparse tensor buffer containing indices. Do not free this pointer.</returns>
-  template <typename R>
-  const R* GetSparseTensorIndicesData(OrtSparseIndicesFormat indices_format, size_t& num_indices) const;
-
-  /// <summary>
-  /// Returns true if the OrtValue contains a sparse tensor
-  /// </summary>
-  /// <returns></returns>
-  bool IsSparseTensor() const;
-
-  /// <summary>
-  /// The API returns a pointer to an internal buffer of the sparse tensor
-  /// containing non-zero values. The API merely does casting. Make sure you
-  /// are requesting the right data type by calling GetSparseTensorValuesTypeAndShapeInfo()
-  /// first.
-  /// </summary>
-  /// <typeparam name="T">numeric data types only. Use GetStringTensor*() to retrieve strings.</typeparam>
-  /// <returns>a pointer to the internal values buffer. Do not free this pointer.</returns>
-  template <typename R>
-  const R* GetSparseTensorValues() const;
-
-#endif
-};
-
-template <typename T>
-struct ValueImpl : ConstValueImpl<T> {
-  using B = ConstValueImpl<T>;
-  using B::B;
-
-  /// <summary>
-  /// Returns a non-const typed pointer to an OrtValue/Tensor contained buffer
-  /// No type checking is performed, the caller must ensure the type matches the tensor type.
-  /// </summary>
-  /// <returns>non-const pointer to data, no copies made</returns>
-  template <typename R>
-  R* GetTensorMutableData();
-
-  /// <summary>
-  /// Returns a non-typed non-const pointer to a tensor contained data.
-  /// </summary>
-  /// <returns>pointer to data, no copies made</returns>
-  void* GetTensorMutableRawData();
-
-  /// <summary>
-  //  Obtain a reference to an element of data at the location specified
-  /// by the vector of dims.
-  /// </summary>
-  /// <typeparam name="R"></typeparam>
-  /// <param name="location">[in] expressed by a vecotr of dimensions offsets</param>
-  /// <returns></returns>
-  template <typename R>
-  R& At(const std::vector<int64_t>& location);
-
-  /// <summary>
-  /// Set all strings at once in a string tensor
-  /// </summary>
-  /// <param name="s">[in] An array of strings. Each string in this array must be null terminated.</param>
-  /// <param name="s_len">[in] Count of strings in s (Must match the size of \p value's tensor shape)</param>
-  void FillStringTensor(const char* const* s, size_t s_len);
-
-  /// <summary>
-  /// Set a single string in a string tensor
-  /// </summary>
-  /// <param name="s">[in] A null terminated UTF-8 encoded string</param>
-  /// <param name="index">[in] Index of the string in the tensor to set</param>
-  void FillStringTensorElement(const char* s, size_t index);
-
-  /// <summary>
-  /// Allocate if necessary and obtain a pointer to a UTF-8
-  /// encoded string element buffer indexed by the flat element index,
-  /// of the specified length.
-  ///
-  /// This API is for advanced usage. It avoids a need to construct
-  /// an auxiliary array of string pointers, and allows to write data directly
-  /// (do not zero terminate).
-  /// </summary>
-  /// <param name="index"></param>
-  /// <param name="buffer_length"></param>
-  /// <returns>a pointer to a writable buffer</returns>
-  char* GetResizedStringTensorElementBuffer(size_t index, size_t buffer_length);
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-  /// <summary>
-  /// Supplies COO format specific indices and marks the contained sparse tensor as being a COO format tensor.
-  /// Values are supplied with a CreateSparseTensor() API. The supplied indices are not copied and the user
-  /// allocated buffers lifespan must eclipse that of the OrtValue.
-  /// The location of the indices is assumed to be the same as specified by OrtMemoryInfo argument at the creation time.
-  /// </summary>
-  /// <param name="indices_data">pointer to the user allocated buffer with indices. Use nullptr for fully sparse tensors.</param>
-  /// <param name="indices_num">number of indices entries. Use 0 for fully sparse tensors</param>
-  void UseCooIndices(int64_t* indices_data, size_t indices_num);
-
-  /// <summary>
-  /// Supplies CSR format specific indices and marks the contained sparse tensor as being a CSR format tensor.
-  /// Values are supplied with a CreateSparseTensor() API. The supplied indices are not copied and the user
-  /// allocated buffers lifespan must eclipse that of the OrtValue.
-  /// The location of the indices is assumed to be the same as specified by OrtMemoryInfo argument at the creation time.
-  /// </summary>
-  /// <param name="inner_data">pointer to the user allocated buffer with inner indices or nullptr for fully sparse tensors</param>
-  /// <param name="inner_num">number of csr inner indices or 0 for fully sparse tensors</param>
-  /// <param name="outer_data">pointer to the user allocated buffer with outer indices or nullptr for fully sparse tensors</param>
-  /// <param name="outer_num">number of csr outer indices or 0 for fully sparse tensors</param>
-  void UseCsrIndices(int64_t* inner_data, size_t inner_num, int64_t* outer_data, size_t outer_num);
-
-  /// <summary>
-  /// Supplies BlockSparse format specific indices and marks the contained sparse tensor as being a BlockSparse format tensor.
-  /// Values are supplied with a CreateSparseTensor() API. The supplied indices are not copied and the user
-  /// allocated buffers lifespan must eclipse that of the OrtValue.
-  /// The location of the indices is assumed to be the same as specified by OrtMemoryInfo argument at the creation time.
-  /// </summary>
-  /// <param name="indices_shape">indices shape or a {0} for fully sparse</param>
-  /// <param name="indices_data">user allocated buffer with indices or nullptr for fully spare tensors</param>
-  void UseBlockSparseIndices(const Shape& indices_shape, int32_t* indices_data);
-
-  /// <summary>
-  /// The API will allocate memory using the allocator instance supplied to the CreateSparseTensor() API
-  /// and copy the values and COO indices into it. If data_mem_info specifies that the data is located
-  /// at difference device than the allocator, a X-device copy will be performed if possible.
-  /// </summary>
-  /// <param name="data_mem_info">specified buffer memory description</param>
-  /// <param name="values_param">values buffer information.</param>
-  /// <param name="indices_data">coo indices buffer or nullptr for fully sparse data</param>
-  /// <param name="indices_num">number of COO indices or 0 for fully sparse data</param>
-  void FillSparseTensorCoo(const OrtMemoryInfo* data_mem_info, const OrtSparseValuesParam& values_param,
-                           const int64_t* indices_data, size_t indices_num);
-
-  /// <summary>
-  /// The API will allocate memory using the allocator instance supplied to the CreateSparseTensor() API
-  /// and copy the values and CSR indices into it. If data_mem_info specifies that the data is located
-  /// at difference device than the allocator, a X-device copy will be performed if possible.
-  /// </summary>
-  /// <param name="data_mem_info">specified buffer memory description</param>
-  /// <param name="values">values buffer information</param>
-  /// <param name="inner_indices_data">csr inner indices pointer or nullptr for fully sparse tensors</param>
-  /// <param name="inner_indices_num">number of csr inner indices or 0 for fully sparse tensors</param>
-  /// <param name="outer_indices_data">pointer to csr indices data or nullptr for fully sparse tensors</param>
-  /// <param name="outer_indices_num">number of csr outer indices or 0</param>
-  void FillSparseTensorCsr(const OrtMemoryInfo* data_mem_info,
-                           const OrtSparseValuesParam& values,
-                           const int64_t* inner_indices_data, size_t inner_indices_num,
-                           const int64_t* outer_indices_data, size_t outer_indices_num);
-
-  /// <summary>
-  /// The API will allocate memory using the allocator instance supplied to the CreateSparseTensor() API
-  /// and copy the values and BlockSparse indices into it. If data_mem_info specifies that the data is located
-  /// at difference device than the allocator, a X-device copy will be performed if possible.
-  /// </summary>
-  /// <param name="data_mem_info">specified buffer memory description</param>
-  /// <param name="values">values buffer information</param>
-  /// <param name="indices_shape">indices shape. use {0} for fully sparse tensors</param>
-  /// <param name="indices_data">pointer to indices data or nullptr for fully sparse tensors</param>
-  void FillSparseTensorBlockSparse(const OrtMemoryInfo* data_mem_info,
-                                   const OrtSparseValuesParam& values,
-                                   const Shape& indices_shape,
-                                   const int32_t* indices_data);
-
-#endif
-};
-
-}  // namespace detail
-
-using ConstValue = detail::ConstValueImpl<detail::Unowned<const OrtValue>>;
-using UnownedValue = detail::ValueImpl<detail::Unowned<OrtValue>>;
-
-/** \brief Wrapper around ::OrtValue
- *
- */
-struct Value : detail::ValueImpl<OrtValue> {
-  using Base = detail::ValueImpl<OrtValue>;
-  using OrtSparseValuesParam = detail::OrtSparseValuesParam;
-  using Shape = detail::Shape;
-
-  explicit Value(std::nullptr_t) {}         ///< Create an empty Value object, must be assigned a valid one to be used
-  explicit Value(OrtValue* p) : Base{p} {}  ///< Used for interop with the C API
-  Value(Value&&) = default;
-  Value& operator=(Value&&) = default;
-
-  ConstValue GetConst() const { return ConstValue{this->p_}; }
-  UnownedValue GetUnowned() const { return UnownedValue{this->p_}; }
-
-  /** \brief Creates a tensor with a user supplied buffer. Wraps OrtApi::CreateTensorWithDataAsOrtValue.
-   * \tparam T The numeric datatype. This API is not suitable for strings.
-   * \param info Memory description of where the p_data buffer resides (CPU vs GPU etc).
-   * \param p_data Pointer to the data buffer.
-   * \param p_data_element_count The number of elements in the data buffer.
-   * \param shape Pointer to the tensor shape dimensions.
-   * \param shape_len The number of tensor shape dimensions.
-   */
-  template <typename T>
-  static Value CreateTensor(const OrtMemoryInfo* info, T* p_data, size_t p_data_element_count, const int64_t* shape, size_t shape_len);
-
-  /** \brief Creates a tensor with a user supplied buffer. Wraps OrtApi::CreateTensorWithDataAsOrtValue.
-   *
-   * \param info Memory description of where the p_data buffer resides (CPU vs GPU etc).
-   * \param p_data Pointer to the data buffer.
-   * \param p_data_byte_count The number of bytes in the data buffer.
-   * \param shape Pointer to the tensor shape dimensions.
-   * \param shape_len The number of tensor shape dimensions.
-   * \param type The data type.
-   */
-  static Value CreateTensor(const OrtMemoryInfo* info, void* p_data, size_t p_data_byte_count, const int64_t* shape, size_t shape_len,
-                            ONNXTensorElementDataType type);
-
-  /** \brief Creates an OrtValue with a tensor using a supplied OrtAllocator. Wraps OrtApi::CreateTensorAsOrtValue.
-   *         This overload will allocate the buffer for the tensor  according to the supplied shape and data type.
-   *         The allocated buffer will be owned by the returned OrtValue and will be freed when the OrtValue is released.
-   *         The input data would need to be copied into the allocated buffer.
-   *         This API is not suitable for strings.
-   *
-   * \tparam T The numeric datatype. This API is not suitable for strings.
-   * \param allocator The allocator to use.
-   * \param shape Pointer to the tensor shape dimensions.
-   * \param shape_len The number of tensor shape dimensions.
-   */
-  template <typename T>
-  static Value CreateTensor(OrtAllocator* allocator, const int64_t* shape, size_t shape_len);
-
-  /** \brief Creates an OrtValue with a tensor using the supplied OrtAllocator.
-   *   Wraps OrtApi::CreateTensorAsOrtValue.
-   *   The allocated buffer will be owned by the returned OrtValue and will be freed when the OrtValue is released.
-   *   The input data would need to be copied into the allocated buffer.
-   *   This API is not suitable for strings.
-   *
-   * \param allocator The allocator to use.
-   * \param shape Pointer to the tensor shape dimensions.
-   * \param shape_len The number of tensor shape dimensions.
-   * \param type The data type.
-   */
-  static Value CreateTensor(OrtAllocator* allocator, const int64_t* shape, size_t shape_len, ONNXTensorElementDataType type);
-
-  /** \brief Creates an OrtValue with a Map Onnx type representation.
-   *  The API would ref-count the supplied OrtValues and they will be released
-   *  when the returned OrtValue is released. The caller may release keys and values after the call
-   *  returns.
-   *
-   * \param keys an OrtValue containing a tensor with primitive data type keys.
-   * \param values an OrtValue that may contain a tensor. Ort currently supports only primitive data type values.
-   */
-  static Value CreateMap(const Value& keys, const Value& values);  ///< Wraps OrtApi::CreateValue
-
-  /** \brief Creates an OrtValue with a Sequence Onnx type representation.
-   *  The API would ref-count the supplied OrtValues and they will be released
-   *  when the returned OrtValue is released. The caller may release the values after the call
-   *  returns.
-   *
-   * \param values a vector of OrtValues that must have the same Onnx value type.
-   */
-  static Value CreateSequence(const std::vector<Value>& values);  ///< Wraps OrtApi::CreateValue
-
-  /** \brief Creates an OrtValue wrapping an Opaque type.
-   *  This is used for experimental support of non-tensor types.
-   *
-   * \tparam T - the type of the value.
-   * \param domain - zero terminated utf-8 string. Domain of the type.
-   * \param type_name - zero terminated utf-8 string. Name of the type.
-   * \param value - the value to be wrapped.
-   */
-  template <typename T>
-  static Value CreateOpaque(const char* domain, const char* type_name, const T& value);  ///< Wraps OrtApi::CreateOpaqueValue
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-  /// <summary>
-  /// This is a simple forwarding method to the other overload that helps deducing
-  /// data type enum value from the type of the buffer.
-  /// </summary>
-  /// <typeparam name="T">numeric datatype. This API is not suitable for strings.</typeparam>
-  /// <param name="info">Memory description where the user buffers reside (CPU vs GPU etc)</param>
-  /// <param name="p_data">pointer to the user supplied buffer, use nullptr for fully sparse tensors</param>
-  /// <param name="dense_shape">a would be dense shape of the tensor</param>
-  /// <param name="values_shape">non zero values shape. Use a single 0 shape for fully sparse tensors.</param>
-  /// <returns></returns>
-  template <typename T>
-  static Value CreateSparseTensor(const OrtMemoryInfo* info, T* p_data, const Shape& dense_shape,
-                                  const Shape& values_shape);
-
-  /// <summary>
-  /// Creates an OrtValue instance containing SparseTensor. This constructs
-  /// a sparse tensor that makes use of user allocated buffers. It does not make copies
-  /// of the user provided data and does not modify it. The lifespan of user provided buffers should
-  /// eclipse the life span of the resulting OrtValue. This call constructs an instance that only contain
-  /// a pointer to non-zero values. To fully populate the sparse tensor call Use<Format>Indices() API below
-  /// to supply a sparse format specific indices.
-  /// This API is not suitable for string data. Use CreateSparseTensor() with allocator specified so strings
-  /// can be properly copied into the allocated buffer.
-  /// </summary>
-  /// <param name="info">Memory description where the user buffers reside (CPU vs GPU etc)</param>
-  /// <param name="p_data">pointer to the user supplied buffer, use nullptr for fully sparse tensors</param>
-  /// <param name="dense_shape">a would be dense shape of the tensor</param>
-  /// <param name="values_shape">non zero values shape. Use a single 0 shape for fully sparse tensors.</param>
-  /// <param name="type">data type</param>
-  /// <returns>Ort::Value instance containing SparseTensor</returns>
-  static Value CreateSparseTensor(const OrtMemoryInfo* info, void* p_data, const Shape& dense_shape,
-                                  const Shape& values_shape, ONNXTensorElementDataType type);
-
-  /// <summary>
-  /// This is a simple forwarding method to the below CreateSparseTensor.
-  /// This helps to specify data type enum in terms of C++ data type.
-  /// Use CreateSparseTensor<T>
-  /// </summary>
-  /// <typeparam name="T">numeric data type only. String data enum must be specified explicitly.</typeparam>
-  /// <param name="allocator">allocator to use</param>
-  /// <param name="dense_shape">a would be dense shape of the tensor</param>
-  /// <returns>Ort::Value</returns>
-  template <typename T>
-  static Value CreateSparseTensor(OrtAllocator* allocator, const Shape& dense_shape);
-
-  /// <summary>
-  /// Creates an instance of OrtValue containing sparse tensor. The created instance has no data.
-  /// The data must be supplied by on of the FillSparseTensor<Format>() methods that take both non-zero values
-  /// and indices. The data will be copied into a buffer that would be allocated using the supplied allocator.
-  /// Use this API to create OrtValues that contain sparse tensors with all supported data types including
-  /// strings.
-  /// </summary>
-  /// <param name="allocator">allocator to use. The allocator lifespan must eclipse that of the resulting OrtValue</param>
-  /// <param name="dense_shape">a would be dense shape of the tensor</param>
-  /// <param name="type">data type</param>
-  /// <returns>an instance of Ort::Value</returns>
-  static Value CreateSparseTensor(OrtAllocator* allocator, const Shape& dense_shape, ONNXTensorElementDataType type);
-
-#endif  // !defined(DISABLE_SPARSE_TENSORS)
-};
-
-/// <summary>
-/// Represents native memory allocation coming from one of the
-/// OrtAllocators registered with OnnxRuntime.
-/// Use it to wrap an allocation made by an allocator
-/// so it can be automatically released when no longer needed.
-/// </summary>
-struct MemoryAllocation {
-  MemoryAllocation(OrtAllocator* allocator, void* p, size_t size);
-  ~MemoryAllocation();
-  MemoryAllocation(const MemoryAllocation&) = delete;
-  MemoryAllocation& operator=(const MemoryAllocation&) = delete;
-  MemoryAllocation(MemoryAllocation&&) noexcept;
-  MemoryAllocation& operator=(MemoryAllocation&&) noexcept;
-
-  void* get() { return p_; }
-  size_t size() const { return size_; }
-
- private:
-  OrtAllocator* allocator_;
-  void* p_;
-  size_t size_;
-};
-
-namespace detail {
-template <typename T>
-struct AllocatorImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  void* Alloc(size_t size);
-  MemoryAllocation GetAllocation(size_t size);
-  void Free(void* p);
-  ConstMemoryInfo GetInfo() const;
-};
-
-}  // namespace detail
-
-/** \brief Wrapper around ::OrtAllocator default instance that is owned by Onnxruntime
- *
- */
-struct AllocatorWithDefaultOptions : detail::AllocatorImpl<detail::Unowned<OrtAllocator>> {
-  explicit AllocatorWithDefaultOptions(std::nullptr_t) {}  ///< Convenience to create a class member and then replace with an instance
-  AllocatorWithDefaultOptions();
-};
-
-/** \brief Wrapper around ::OrtAllocator
- *
- */
-struct Allocator : detail::AllocatorImpl<OrtAllocator> {
-  explicit Allocator(std::nullptr_t) {}  ///< Convenience to create a class member and then replace with an instance
-  Allocator(const Session& session, const OrtMemoryInfo*);
-};
-
-using UnownedAllocator = detail::AllocatorImpl<detail::Unowned<OrtAllocator>>;
-
-namespace detail {
-namespace binding_utils {
-// Bring these out of template
-std::vector<std::string> GetOutputNamesHelper(const OrtIoBinding* binding, OrtAllocator*);
-std::vector<Value> GetOutputValuesHelper(const OrtIoBinding* binding, OrtAllocator*);
-}  // namespace binding_utils
-
-template <typename T>
-struct ConstIoBindingImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  std::vector<std::string> GetOutputNames() const;
-  std::vector<std::string> GetOutputNames(OrtAllocator*) const;
-  std::vector<Value> GetOutputValues() const;
-  std::vector<Value> GetOutputValues(OrtAllocator*) const;
-};
-
-template <typename T>
-struct IoBindingImpl : ConstIoBindingImpl<T> {
-  using B = ConstIoBindingImpl<T>;
-  using B::B;
-
-  void BindInput(const char* name, const Value&);
-  void BindOutput(const char* name, const Value&);
-  void BindOutput(const char* name, const OrtMemoryInfo*);
-  void ClearBoundInputs();
-  void ClearBoundOutputs();
-  void SynchronizeInputs();
-  void SynchronizeOutputs();
-};
-
-}  // namespace detail
-
-using ConstIoBinding = detail::ConstIoBindingImpl<detail::Unowned<const OrtIoBinding>>;
-using UnownedIoBinding = detail::IoBindingImpl<detail::Unowned<OrtIoBinding>>;
-
-/** \brief Wrapper around ::OrtIoBinding
- *
- */
-struct IoBinding : detail::IoBindingImpl<OrtIoBinding> {
-  explicit IoBinding(std::nullptr_t) {}  ///< Create an empty object for convenience. Sometimes, we want to initialize members later.
-  explicit IoBinding(Session& session);
-  ConstIoBinding GetConst() const { return ConstIoBinding{this->p_}; }
-  UnownedIoBinding GetUnowned() const { return UnownedIoBinding{this->p_}; }
-};
-
-/*! \struct Ort::ArenaCfg
- * \brief it is a structure that represents the configuration of an arena based allocator
- * \details Please see docs/C_API.md for details
- */
-struct ArenaCfg : detail::Base<OrtArenaCfg> {
-  explicit ArenaCfg(std::nullptr_t) {}  ///< Create an empty ArenaCfg object, must be assigned a valid one to be used
-  /**
-   * Wraps OrtApi::CreateArenaCfg
-   * \param max_mem - use 0 to allow ORT to choose the default
-   * \param arena_extend_strategy -  use -1 to allow ORT to choose the default, 0 = kNextPowerOfTwo, 1 = kSameAsRequested
-   * \param initial_chunk_size_bytes - use -1 to allow ORT to choose the default
-   * \param max_dead_bytes_per_chunk - use -1 to allow ORT to choose the default
-   * See docs/C_API.md for details on what the following parameters mean and how to choose these values
-   */
-  ArenaCfg(size_t max_mem, int arena_extend_strategy, int initial_chunk_size_bytes, int max_dead_bytes_per_chunk);
-};
-
-//
-// Custom OPs (only needed to implement custom OPs)
-//
-
-/// <summary>
-/// This struct provides life time management for custom op attribute
-/// </summary>
-struct OpAttr : detail::Base<OrtOpAttr> {
-  OpAttr(const char* name, const void* data, int len, OrtOpAttrType type);
-};
-
-/**
- * Macro that logs a message using the provided logger. Throws an exception if OrtApi::Logger_LogMessage fails.
- * Example: ORT_CXX_LOG(logger, ORT_LOGGING_LEVEL_INFO, "Log a message");
- *
- * \param logger The Ort::Logger instance to use. Must be a value or reference.
- * \param message_severity The logging severity level of the message.
- * \param message A null-terminated UTF-8 message to log.
- */
-#define ORT_CXX_LOG(logger, message_severity, message)                                       \
-  do {                                                                                       \
-    if (message_severity >= logger.GetLoggingSeverityLevel()) {                              \
-      Ort::ThrowOnError(logger.LogMessage(message_severity, ORT_FILE, __LINE__,              \
-                                          static_cast<const char*>(__FUNCTION__), message)); \
-    }                                                                                        \
-  } while (false)
-
-/**
- * Macro that logs a message using the provided logger. Can be used in noexcept code since errors are silently ignored.
- * Example: ORT_CXX_LOG_NOEXCEPT(logger, ORT_LOGGING_LEVEL_INFO, "Log a message");
- *
- * \param logger The Ort::Logger instance to use. Must be a value or reference.
- * \param message_severity The logging severity level of the message.
- * \param message A null-terminated UTF-8 message to log.
- */
-#define ORT_CXX_LOG_NOEXCEPT(logger, message_severity, message)                              \
-  do {                                                                                       \
-    if (message_severity >= logger.GetLoggingSeverityLevel()) {                              \
-      static_cast<void>(logger.LogMessage(message_severity, ORT_FILE, __LINE__,              \
-                                          static_cast<const char*>(__FUNCTION__), message)); \
-    }                                                                                        \
-  } while (false)
-
-/**
- * Macro that logs a printf-like formatted message using the provided logger. Throws an exception if
- * OrtApi::Logger_LogMessage fails or if a formatting error occurs.
- * Example: ORT_CXX_LOGF(logger, ORT_LOGGING_LEVEL_INFO, "Log an int: %d", 12);
- *
- * \param logger The Ort::Logger instance to use. Must be a value or reference.
- * \param message_severity The logging severity level of the message.
- * \param format A null-terminated UTF-8 format string forwarded to a printf-like function.
- *               Refer to https://en.cppreference.com/w/cpp/io/c/fprintf for information on valid formats.
- * \param ... Zero or more variadic arguments referenced by the format string.
- */
-#define ORT_CXX_LOGF(logger, message_severity, /*format,*/...)                                            \
-  do {                                                                                                    \
-    if (message_severity >= logger.GetLoggingSeverityLevel()) {                                           \
-      Ort::ThrowOnError(logger.LogFormattedMessage(message_severity, ORT_FILE, __LINE__,                  \
-                                                   static_cast<const char*>(__FUNCTION__), __VA_ARGS__)); \
-    }                                                                                                     \
-  } while (false)
-
-/**
- * Macro that logs a printf-like formatted message using the provided logger. Can be used in noexcept code since errors
- * are silently ignored.
- * Example: ORT_CXX_LOGF_NOEXCEPT(logger, ORT_LOGGING_LEVEL_INFO, "Log an int: %d", 12);
- *
- * \param logger The Ort::Logger instance to use. Must be a value or reference.
- * \param message_severity The logging severity level of the message.
- * \param format A null-terminated UTF-8 format string forwarded to a printf-like function.
- *               Refer to https://en.cppreference.com/w/cpp/io/c/fprintf for information on valid formats.
- * \param ... Zero or more variadic arguments referenced by the format string.
- */
-#define ORT_CXX_LOGF_NOEXCEPT(logger, message_severity, /*format,*/...)                                   \
-  do {                                                                                                    \
-    if (message_severity >= logger.GetLoggingSeverityLevel()) {                                           \
-      static_cast<void>(logger.LogFormattedMessage(message_severity, ORT_FILE, __LINE__,                  \
-                                                   static_cast<const char*>(__FUNCTION__), __VA_ARGS__)); \
-    }                                                                                                     \
-  } while (false)
-
-/// <summary>
-/// This class represents an ONNX Runtime logger that can be used to log information with an
-/// associated severity level and source code location (file path, line number, function name).
-///
-/// A Logger can be obtained from within custom operators by calling Ort::KernelInfo::GetLogger().
-/// Instances of Ort::Logger are the size of two pointers and can be passed by value.
-///
-/// Use the ORT_CXX_LOG macros to ensure the source code location is set properly from the callsite
-/// and to take advantage of a cached logging severity level that can bypass calls to the underlying C API.
-/// </summary>
-struct Logger {
-  /**
-   * Creates an empty Ort::Logger. Must be initialized from a valid Ort::Logger before use.
-   */
-  Logger() = default;
-
-  /**
-   * Creates an empty Ort::Logger. Must be initialized from a valid Ort::Logger before use.
-   */
-  explicit Logger(std::nullptr_t) {}
-
-  /**
-   * Creates a logger from an ::OrtLogger instance. Caches the logger's current severity level by calling
-   * OrtApi::Logger_GetLoggingSeverityLevel. Throws an exception if OrtApi::Logger_GetLoggingSeverityLevel fails.
-   *
-   * \param logger The ::OrtLogger to wrap.
-   */
-  explicit Logger(const OrtLogger* logger);
-
-  ~Logger() = default;
-
-  Logger(const Logger&) = default;
-  Logger& operator=(const Logger&) = default;
-
-  Logger(Logger&& v) noexcept = default;
-  Logger& operator=(Logger&& v) noexcept = default;
-
-  /**
-   * Returns the logger's current severity level from the cached member.
-   *
-   * \return The current ::OrtLoggingLevel.
-   */
-  OrtLoggingLevel GetLoggingSeverityLevel() const noexcept;
-
-  /**
-   * Logs the provided message via OrtApi::Logger_LogMessage. Use the ORT_CXX_LOG or ORT_CXX_LOG_NOEXCEPT
-   * macros to properly set the source code location and to use the cached severity level to potentially bypass
-   * calls to the underlying C API.
-   *
-   * \param log_severity_level The message's logging severity level.
-   * \param file_path The filepath of the file in which the message is logged. Usually the value of ORT_FILE.
-   * \param line_number The file line number in which the message is logged. Usually the value of __LINE__.
-   * \param func_name The name of the function in which the message is logged. Usually the value of __FUNCTION__.
-   * \param message The message to log.
-   * \return A Ort::Status value to indicate error or success.
-   */
-  Status LogMessage(OrtLoggingLevel log_severity_level, const ORTCHAR_T* file_path, int line_number,
-                    const char* func_name, const char* message) const noexcept;
-
-  /**
-   * Logs a printf-like formatted message via OrtApi::Logger_LogMessage. Use the ORT_CXX_LOGF or ORT_CXX_LOGF_NOEXCEPT
-   * macros to properly set the source code location and to use the cached severity level to potentially bypass
-   * calls to the underlying C API. Returns an error status if a formatting error occurs.
-   *
-   * \param log_severity_level The message's logging severity level.
-   * \param file_path The filepath of the file in which the message is logged. Usually the value of ORT_FILE.
-   * \param line_number The file line number in which the message is logged. Usually the value of __LINE__.
-   * \param func_name The name of the function in which the message is logged. Usually the value of __FUNCTION__.
-   * \param format A null-terminated UTF-8 format string forwarded to a printf-like function.
-   *               Refer to https://en.cppreference.com/w/cpp/io/c/fprintf for information on valid formats.
-   * \param args Zero or more variadic arguments referenced by the format string.
-   * \return A Ort::Status value to indicate error or success.
-   */
-  template <typename... Args>
-  Status LogFormattedMessage(OrtLoggingLevel log_severity_level, const ORTCHAR_T* file_path, int line_number,
-                             const char* func_name, const char* format, Args&&... args) const noexcept;
-
- private:
-  const OrtLogger* logger_{};
-  OrtLoggingLevel cached_severity_level_{};
-};
-
-/// <summary>
-/// This class wraps a raw pointer OrtKernelContext* that is being passed
-/// to the custom kernel Compute() method. Use it to safely access context
-/// attributes, input and output parameters with exception safety guarantees.
-/// See usage example in onnxruntime/test/testdata/custom_op_library/custom_op_library.cc
-/// </summary>
-struct KernelContext {
-  explicit KernelContext(OrtKernelContext* context);
-  size_t GetInputCount() const;
-  size_t GetOutputCount() const;
-  ConstValue GetInput(size_t index) const;
-  UnownedValue GetOutput(size_t index, const int64_t* dim_values, size_t dim_count) const;
-  UnownedValue GetOutput(size_t index, const std::vector<int64_t>& dims) const;
-  void* GetGPUComputeStream() const;
-  Logger GetLogger() const;
-  OrtAllocator* GetAllocator(const OrtMemoryInfo& memory_info) const;
-  OrtKernelContext* GetOrtKernelContext() const { return ctx_; }
-  void ParallelFor(void (*fn)(void*, size_t), size_t total, size_t num_batch, void* usr_data) const;
-
- private:
-  OrtKernelContext* ctx_;
-};
-
-struct KernelInfo;
-
-namespace detail {
-namespace attr_utils {
-void GetAttr(const OrtKernelInfo* p, const char* name, float&);
-void GetAttr(const OrtKernelInfo* p, const char* name, int64_t&);
-void GetAttr(const OrtKernelInfo* p, const char* name, std::string&);
-void GetAttrs(const OrtKernelInfo* p, const char* name, std::vector<float>&);
-void GetAttrs(const OrtKernelInfo* p, const char* name, std::vector<int64_t>&);
-}  // namespace attr_utils
-
-template <typename T>
-struct KernelInfoImpl : Base<T> {
-  using B = Base<T>;
-  using B::B;
-
-  KernelInfo Copy() const;
-
-  template <typename R>  // R is only implemented for float, int64_t, and string
-  R GetAttribute(const char* name) const {
-    R val;
-    attr_utils::GetAttr(this->p_, name, val);
-    return val;
-  }
-
-  template <typename R>  // R is only implemented for std::vector<float>, std::vector<int64_t>
-  std::vector<R> GetAttributes(const char* name) const {
-    std::vector<R> result;
-    attr_utils::GetAttrs(this->p_, name, result);
-    return result;
-  }
-
-  Value GetTensorAttribute(const char* name, OrtAllocator* allocator) const;
-
-  size_t GetInputCount() const;
-  size_t GetOutputCount() const;
-
-  std::string GetInputName(size_t index) const;
-  std::string GetOutputName(size_t index) const;
-
-  TypeInfo GetInputTypeInfo(size_t index) const;
-  TypeInfo GetOutputTypeInfo(size_t index) const;
-
-  ConstValue GetTensorConstantInput(size_t index, int* is_constant) const;
-
-  std::string GetNodeName() const;
-  Logger GetLogger() const;
-};
-
-}  // namespace detail
-
-using ConstKernelInfo = detail::KernelInfoImpl<detail::Unowned<const OrtKernelInfo>>;
-
-/// <summary>
-/// This struct owns the OrtKernInfo* pointer when a copy is made.
-/// For convenient wrapping of OrtKernelInfo* passed to kernel constructor
-/// and query attributes, warp the pointer with Ort::Unowned<KernelInfo> instance
-/// so it does not destroy the pointer the kernel does not own.
-/// </summary>
-struct KernelInfo : detail::KernelInfoImpl<OrtKernelInfo> {
-  explicit KernelInfo(std::nullptr_t) {}     ///< Create an empty instance to initialize later
-  explicit KernelInfo(OrtKernelInfo* info);  ///< Take ownership of the instance
-  ConstKernelInfo GetConst() const { return ConstKernelInfo{this->p_}; }
-};
-
-/// <summary>
-/// Create and own custom defined operation.
-/// </summary>
-struct Op : detail::Base<OrtOp> {
-  explicit Op(std::nullptr_t) {}  ///< Create an empty Operator object, must be assigned a valid one to be used
-
-  explicit Op(OrtOp*);  ///< Take ownership of the OrtOp
-
-  static Op Create(const OrtKernelInfo* info, const char* op_name, const char* domain,
-                   int version, const char** type_constraint_names,
-                   const ONNXTensorElementDataType* type_constraint_values,
-                   size_t type_constraint_count,
-                   const OpAttr* attr_values,
-                   size_t attr_count,
-                   size_t input_count, size_t output_count);
-
-  void Invoke(const OrtKernelContext* context,
-              const Value* input_values,
-              size_t input_count,
-              Value* output_values,
-              size_t output_count);
-
-  // For easier refactoring
-  void Invoke(const OrtKernelContext* context,
-              const OrtValue* const* input_values,
-              size_t input_count,
-              OrtValue* const* output_values,
-              size_t output_count);
-};
-
-/// <summary>
-/// Provide access to per-node attributes and input shapes, so one could compute and set output shapes.
-/// </summary>
-struct ShapeInferContext {
-  struct SymbolicInteger {
-    SymbolicInteger(int64_t i) : i_(i), is_int_(true){};
-    SymbolicInteger(const char* s) : s_(s), is_int_(false){};
-    SymbolicInteger(const SymbolicInteger&) = default;
-    SymbolicInteger(SymbolicInteger&&) = default;
-
-    SymbolicInteger& operator=(const SymbolicInteger&) = default;
-    SymbolicInteger& operator=(SymbolicInteger&&) = default;
-
-    bool operator==(const SymbolicInteger& dim) const {
-      if (is_int_ == dim.is_int_) {
-        if (is_int_) {
-          return i_ == dim.i_;
-        } else {
-          return std::string{s_} == std::string{dim.s_};
-        }
-      }
-      return false;
-    }
-
-    bool IsInt() const { return is_int_; }
-    int64_t AsInt() const { return i_; }
-    const char* AsSym() const { return s_; }
-
-    static constexpr int INVALID_INT_DIM = -2;
-
-   private:
-    union {
-      int64_t i_;
-      const char* s_;
-    };
-    bool is_int_;
-  };
-
-  using Shape = std::vector<SymbolicInteger>;
-
-  ShapeInferContext(const OrtApi* ort_api, OrtShapeInferContext* ctx);
-
-  const Shape& GetInputShape(size_t indice) const { return input_shapes_.at(indice); }
-
-  size_t GetInputCount() const { return input_shapes_.size(); }
-
-  Status SetOutputShape(size_t indice, const Shape& shape);
-
-  int64_t GetAttrInt(const char* attr_name);
-
-  using Ints = std::vector<int64_t>;
-  Ints GetAttrInts(const char* attr_name);
-
-  float GetAttrFloat(const char* attr_name);
-
-  using Floats = std::vector<float>;
-  Floats GetAttrFloats(const char* attr_name);
-
-  std::string GetAttrString(const char* attr_name);
-
-  using Strings = std::vector<std::string>;
-  Strings GetAttrStrings(const char* attr_name);
-
- private:
-  const OrtOpAttr* GetAttrHdl(const char* attr_name) const;
-  const OrtApi* ort_api_;
-  OrtShapeInferContext* ctx_;
-  std::vector<Shape> input_shapes_;
-};
-
-using ShapeInferFn = Ort::Status (*)(Ort::ShapeInferContext&);
-
-#define MAX_CUSTOM_OP_END_VER (1UL << 31) - 1
-
-template <typename TOp, typename TKernel, bool WithStatus = false>
-struct CustomOpBase : OrtCustomOp {
-  CustomOpBase() {
-    OrtCustomOp::version = ORT_API_VERSION;
-    OrtCustomOp::GetName = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetName(); };
-
-    OrtCustomOp::GetExecutionProviderType = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetExecutionProviderType(); };
-
-    OrtCustomOp::GetInputTypeCount = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetInputTypeCount(); };
-    OrtCustomOp::GetInputType = [](const OrtCustomOp* this_, size_t index) { return static_cast<const TOp*>(this_)->GetInputType(index); };
-    OrtCustomOp::GetInputMemoryType = [](const OrtCustomOp* this_, size_t index) { return static_cast<const TOp*>(this_)->GetInputMemoryType(index); };
-
-    OrtCustomOp::GetOutputTypeCount = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetOutputTypeCount(); };
-    OrtCustomOp::GetOutputType = [](const OrtCustomOp* this_, size_t index) { return static_cast<const TOp*>(this_)->GetOutputType(index); };
-
-#if defined(_MSC_VER) && !defined(__clang__)
-#pragma warning(push)
-#pragma warning(disable : 26409)
-#endif
-    OrtCustomOp::KernelDestroy = [](void* op_kernel) { delete static_cast<TKernel*>(op_kernel); };
-#if defined(_MSC_VER) && !defined(__clang__)
-#pragma warning(pop)
-#endif
-    OrtCustomOp::GetInputCharacteristic = [](const OrtCustomOp* this_, size_t index) { return static_cast<const TOp*>(this_)->GetInputCharacteristic(index); };
-    OrtCustomOp::GetOutputCharacteristic = [](const OrtCustomOp* this_, size_t index) { return static_cast<const TOp*>(this_)->GetOutputCharacteristic(index); };
-
-    OrtCustomOp::GetVariadicInputMinArity = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetVariadicInputMinArity(); };
-    OrtCustomOp::GetVariadicInputHomogeneity = [](const OrtCustomOp* this_) { return static_cast<int>(static_cast<const TOp*>(this_)->GetVariadicInputHomogeneity()); };
-    OrtCustomOp::GetVariadicOutputMinArity = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetVariadicOutputMinArity(); };
-    OrtCustomOp::GetVariadicOutputHomogeneity = [](const OrtCustomOp* this_) { return static_cast<int>(static_cast<const TOp*>(this_)->GetVariadicOutputHomogeneity()); };
-#ifdef __cpp_if_constexpr
-    if constexpr (WithStatus) {
-#else
-    if (WithStatus) {
-#endif
-      OrtCustomOp::CreateKernelV2 = [](const OrtCustomOp* this_, const OrtApi* api, const OrtKernelInfo* info, void** op_kernel) -> OrtStatusPtr {
-        return static_cast<const TOp*>(this_)->CreateKernelV2(*api, info, op_kernel);
-      };
-      OrtCustomOp::KernelComputeV2 = [](void* op_kernel, OrtKernelContext* context) -> OrtStatusPtr {
-        return static_cast<TKernel*>(op_kernel)->ComputeV2(context);
-      };
-    } else {
-      OrtCustomOp::CreateKernelV2 = nullptr;
-      OrtCustomOp::KernelComputeV2 = nullptr;
-
-      OrtCustomOp::CreateKernel = [](const OrtCustomOp* this_, const OrtApi* api, const OrtKernelInfo* info) { return static_cast<const TOp*>(this_)->CreateKernel(*api, info); };
-      OrtCustomOp::KernelCompute = [](void* op_kernel, OrtKernelContext* context) {
-        static_cast<TKernel*>(op_kernel)->Compute(context);
-      };
-    }
-
-    SetShapeInferFn<TOp>(0);
-
-    OrtCustomOp::GetStartVersion = [](const OrtCustomOp* this_) {
-      return static_cast<const TOp*>(this_)->start_ver_;
-    };
-
-    OrtCustomOp::GetEndVersion = [](const OrtCustomOp* this_) {
-      return static_cast<const TOp*>(this_)->end_ver_;
-    };
-  }
-
-  // Default implementation of GetExecutionProviderType that returns nullptr to default to the CPU provider
-  const char* GetExecutionProviderType() const { return nullptr; }
-
-  // Default implementations of GetInputCharacteristic() and GetOutputCharacteristic() below
-  // (inputs and outputs are required by default)
-  OrtCustomOpInputOutputCharacteristic GetInputCharacteristic(size_t /*index*/) const {
-    return OrtCustomOpInputOutputCharacteristic::INPUT_OUTPUT_REQUIRED;
-  }
-
-  OrtCustomOpInputOutputCharacteristic GetOutputCharacteristic(size_t /*index*/) const {
-    return OrtCustomOpInputOutputCharacteristic::INPUT_OUTPUT_REQUIRED;
-  }
-
-  // Default implemention of GetInputMemoryType() that returns OrtMemTypeDefault
-  OrtMemType GetInputMemoryType(size_t /*index*/) const {
-    return OrtMemTypeDefault;
-  }
-
-  // Default implementation of GetVariadicInputMinArity() returns 1 to specify that a variadic input
-  // should expect at least 1 argument.
-  int GetVariadicInputMinArity() const {
-    return 1;
-  }
-
-  // Default implementation of GetVariadicInputHomegeneity() returns true to specify that all arguments
-  // to a variadic input should be of the same type.
-  bool GetVariadicInputHomogeneity() const {
-    return true;
-  }
-
-  // Default implementation of GetVariadicOutputMinArity() returns 1 to specify that a variadic output
-  // should produce at least 1 output value.
-  int GetVariadicOutputMinArity() const {
-    return 1;
-  }
-
-  // Default implementation of GetVariadicOutputHomegeneity() returns true to specify that all output values
-  // produced by a variadic output should be of the same type.
-  bool GetVariadicOutputHomogeneity() const {
-    return true;
-  }
-
-  // Declare list of session config entries used by this Custom Op.
-  // Implement this function in order to get configs from CustomOpBase::GetSessionConfigs().
-  // This default implementation returns an empty vector of config entries.
-  std::vector<std::string> GetSessionConfigKeys() const {
-    return std::vector<std::string>{};
-  }
-
-  template <typename C>
-  decltype(&C::InferOutputShape) SetShapeInferFn(decltype(&C::InferOutputShape)) {
-    OrtCustomOp::InferOutputShapeFn = [](const OrtCustomOp*, OrtShapeInferContext* ort_ctx) -> OrtStatusPtr {
-      ShapeInferContext ctx(&GetApi(), ort_ctx);
-      return C::InferOutputShape(ctx);
-    };
-    return {};
-  }
-
-  template <typename C>
-  void SetShapeInferFn(...) {
-    OrtCustomOp::InferOutputShapeFn = {};
-  }
-
- protected:
-  // Helper function that returns a map of session config entries specified by CustomOpBase::GetSessionConfigKeys.
-  void GetSessionConfigs(std::unordered_map<std::string, std::string>& out, ConstSessionOptions options) const;
-
-  int start_ver_ = 1;
-  int end_ver_ = MAX_CUSTOM_OP_END_VER;
-};
-
-}  // namespace Ort
-
-#include "onnxruntime_cxx_inline.h"
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_cxx_inline.h b/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_cxx_inline.h
deleted file mode 100644
index 957e849c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_cxx_inline.h
+++ /dev/null
@@ -1,2075 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-// Do not include this file directly. Please include "onnxruntime_cxx_api.h" instead.
-// If interested in trying out features of the new experimental C++ API, include "experimental_onnxruntime_cxx_api.h" instead.
-//
-// These are the inline implementations of the C++ header APIs. They're in this separate file as to not clutter
-// the main C++ file with implementation details.
-
-#include <cstring>
-#include <functional>
-
-#define RETURN_ON_API_FAIL(expression) \
-  {                                    \
-    auto err = (expression);           \
-    if (err) {                         \
-      return Status(err);              \
-    }                                  \
-  }
-
-namespace Ort {
-
-namespace detail {
-inline void ThrowStatus(const Status& st) {
-  std::string error_message = st.GetErrorMessage();
-  OrtErrorCode error_code = st.GetErrorCode();
-  ORT_CXX_API_THROW(std::move(error_message), error_code);
-}
-}  // namespace detail
-
-inline void ThrowOnError(OrtStatus* ort_status) {
-  if (ort_status) {
-    Ort::Status st(ort_status);
-    detail::ThrowStatus(st);
-  }
-}
-
-inline void ThrowOnError(const Status& st) {
-  if (st) {
-    detail::ThrowStatus(st);
-  }
-}
-
-inline Status::Status(OrtStatus* status) noexcept : Base<OrtStatus>{status} {
-}
-
-inline Status::Status(const std::exception& e) noexcept {
-  p_ = GetApi().CreateStatus(ORT_FAIL, e.what());
-}
-
-inline Status::Status(const Exception& e) noexcept {
-  p_ = GetApi().CreateStatus(e.GetOrtErrorCode(), e.what());
-}
-
-inline Status::Status(const char* message, OrtErrorCode code) noexcept {
-  p_ = GetApi().CreateStatus(code, message);
-}
-
-inline std::string Status::GetErrorMessage() const {
-  std::string message(GetApi().GetErrorMessage(p_));
-  return message;
-}
-
-inline OrtErrorCode Status::GetErrorCode() const {
-  return GetApi().GetErrorCode(p_);
-}
-
-inline bool Status::IsOK() const noexcept {
-  return (p_ == nullptr);
-}
-
-// This template converts a C++ type into it's ONNXTensorElementDataType
-template <typename T>
-struct TypeToTensorType;
-template <>
-struct TypeToTensorType<float> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT;
-};
-template <>
-struct TypeToTensorType<Float16_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT16;
-};
-template <>
-struct TypeToTensorType<BFloat16_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_BFLOAT16;
-};
-template <>
-struct TypeToTensorType<double> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_DOUBLE;
-};
-template <>
-struct TypeToTensorType<int8_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_INT8;
-};
-template <>
-struct TypeToTensorType<int16_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_INT16;
-};
-template <>
-struct TypeToTensorType<int32_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32;
-};
-template <>
-struct TypeToTensorType<int64_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64;
-};
-template <>
-struct TypeToTensorType<uint8_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT8;
-};
-template <>
-struct TypeToTensorType<uint16_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT16;
-};
-template <>
-struct TypeToTensorType<uint32_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT32;
-};
-template <>
-struct TypeToTensorType<uint64_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT64;
-};
-template <>
-struct TypeToTensorType<bool> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_BOOL;
-};
-
-template <>
-struct TypeToTensorType<Float8E4M3FN_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E4M3FN;
-};
-template <>
-struct TypeToTensorType<Float8E4M3FNUZ_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E4M3FNUZ;
-};
-template <>
-struct TypeToTensorType<Float8E5M2_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E5M2;
-};
-template <>
-struct TypeToTensorType<Float8E5M2FNUZ_t> {
-  static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E5M2FNUZ;
-};
-
-inline bool BFloat16_t::operator==(const BFloat16_t& rhs) const noexcept {
-  if (IsNaN() || rhs.IsNaN()) {
-    // IEEE defines that NaN is not equal to anything, including itself.
-    return false;
-  }
-  return val == rhs.val;
-}
-
-inline bool BFloat16_t::operator<(const BFloat16_t& rhs) const noexcept {
-  if (IsNaN() || rhs.IsNaN()) {
-    // IEEE defines that NaN is unordered with respect to everything, including itself.
-    return false;
-  }
-
-  const bool left_is_negative = IsNegative();
-  if (left_is_negative != rhs.IsNegative()) {
-    // When the signs of left and right differ, we know that left is less than right if it is
-    // the negative value. The exception to this is if both values are zero, in which case IEEE
-    // says they should be equal, even if the signs differ.
-    return left_is_negative && !AreZero(*this, rhs);
-  }
-  return (val != rhs.val) && ((val < rhs.val) ^ left_is_negative);
-}
-
-inline MemoryAllocation::MemoryAllocation(OrtAllocator* allocator, void* p, size_t size)
-    : allocator_(allocator), p_(p), size_(size) {
-}
-
-inline MemoryAllocation::~MemoryAllocation() {
-  if (p_ != nullptr) {
-    // We do not throw out of destructor
-    auto ret = GetApi().AllocatorFree(allocator_, p_);
-    static_cast<void>(ret);
-  }
-}
-
-inline MemoryAllocation::MemoryAllocation(MemoryAllocation&& o) noexcept : allocator_(nullptr), p_(nullptr), size_(0) {
-  *this = std::move(o);
-}
-
-inline MemoryAllocation& MemoryAllocation::operator=(MemoryAllocation&& o) noexcept {
-  OrtAllocator* alloc = nullptr;
-  void* p = nullptr;
-  size_t sz = 0;
-
-  // Swap out this
-  std::swap(alloc, allocator_);
-  std::swap(p, p_);
-  std::swap(sz, size_);
-
-  // Swap with incoming
-  std::swap(allocator_, o.allocator_);
-  std::swap(p_, o.p_);
-  std::swap(size_, o.size_);
-
-  // Destroy this instance if needed
-  MemoryAllocation this_alloc(alloc, p, sz);
-  return *this;
-}
-
-namespace detail {
-
-template <typename T>
-inline void* AllocatorImpl<T>::Alloc(size_t size) {
-  void* out;
-  ThrowOnError(GetApi().AllocatorAlloc(this->p_, size, &out));
-  return out;
-}
-
-template <typename T>
-inline MemoryAllocation AllocatorImpl<T>::GetAllocation(size_t size) {
-  void* out;
-  ThrowOnError(GetApi().AllocatorAlloc(this->p_, size, &out));
-  MemoryAllocation result(this->p_, out, size);
-  return result;
-}
-
-template <typename T>
-inline void AllocatorImpl<T>::Free(void* p) {
-  ThrowOnError(GetApi().AllocatorFree(this->p_, p));
-}
-
-template <typename T>
-inline ConstMemoryInfo AllocatorImpl<T>::GetInfo() const {
-  const OrtMemoryInfo* out;
-  ThrowOnError(GetApi().AllocatorGetInfo(this->p_, &out));
-  return ConstMemoryInfo{out};
-}
-
-}  // namespace detail
-
-inline AllocatorWithDefaultOptions::AllocatorWithDefaultOptions() {
-  ThrowOnError(GetApi().GetAllocatorWithDefaultOptions(&this->p_));
-}
-
-inline Allocator::Allocator(const Session& sess, const OrtMemoryInfo* mem_info) {
-  ThrowOnError(GetApi().CreateAllocator(sess, mem_info, &this->p_));
-}
-
-namespace detail {
-
-template <typename T>
-inline std::string MemoryInfoImpl<T>::GetAllocatorName() const {
-  const char* name = nullptr;
-  ThrowOnError(GetApi().MemoryInfoGetName(this->p_, &name));
-  return std::string(name);
-}
-
-template <typename T>
-inline OrtAllocatorType MemoryInfoImpl<T>::GetAllocatorType() const {
-  OrtAllocatorType type;
-  ThrowOnError(GetApi().MemoryInfoGetType(this->p_, &type));
-  return type;
-}
-
-template <typename T>
-inline int MemoryInfoImpl<T>::GetDeviceId() const {
-  int id = 0;
-  ThrowOnError(GetApi().MemoryInfoGetId(this->p_, &id));
-  return id;
-}
-
-template <typename T>
-inline OrtMemoryInfoDeviceType MemoryInfoImpl<T>::GetDeviceType() const {
-  OrtMemoryInfoDeviceType type;
-  GetApi().MemoryInfoGetDeviceType(this->p_, &type);
-  return type;
-}
-
-template <typename T>
-inline OrtMemType MemoryInfoImpl<T>::GetMemoryType() const {
-  OrtMemType type;
-  ThrowOnError(GetApi().MemoryInfoGetMemType(this->p_, &type));
-  return type;
-}
-
-template <typename T>
-template <typename U>
-inline bool MemoryInfoImpl<T>::operator==(const MemoryInfoImpl<U>& o) const {
-  int comp_result = 0;
-  ThrowOnError(Ort::GetApi().CompareMemoryInfo(this->p_, o, &comp_result));
-  return comp_result == 0;
-}
-
-}  // namespace detail
-
-inline MemoryInfo MemoryInfo::CreateCpu(OrtAllocatorType type, OrtMemType mem_type) {
-  OrtMemoryInfo* p;
-  ThrowOnError(GetApi().CreateCpuMemoryInfo(type, mem_type, &p));
-  return MemoryInfo(p);
-}
-
-inline MemoryInfo::MemoryInfo(const char* name, OrtAllocatorType type, int id, OrtMemType mem_type) {
-  ThrowOnError(GetApi().CreateMemoryInfo(name, type, id, mem_type, &this->p_));
-}
-
-namespace detail {
-template <typename T>
-inline std::vector<std::string> ConstIoBindingImpl<T>::GetOutputNames() const {
-  AllocatorWithDefaultOptions allocator;
-  return binding_utils::GetOutputNamesHelper(this->p_, allocator);
-}
-
-template <typename T>
-inline std::vector<std::string> ConstIoBindingImpl<T>::GetOutputNames(OrtAllocator* allocator) const {
-  return binding_utils::GetOutputNamesHelper(this->p_, allocator);
-}
-
-template <typename T>
-inline std::vector<Value> ConstIoBindingImpl<T>::GetOutputValues() const {
-  AllocatorWithDefaultOptions allocator;
-  return binding_utils::GetOutputValuesHelper(this->p_, allocator);
-}
-
-template <typename T>
-inline std::vector<Value> ConstIoBindingImpl<T>::GetOutputValues(OrtAllocator* allocator) const {
-  return binding_utils::GetOutputValuesHelper(this->p_, allocator);
-}
-
-template <typename T>
-inline void IoBindingImpl<T>::BindInput(const char* name, const Value& value) {
-  ThrowOnError(GetApi().BindInput(this->p_, name, value));
-}
-
-template <typename T>
-inline void IoBindingImpl<T>::BindOutput(const char* name, const Value& value) {
-  ThrowOnError(GetApi().BindOutput(this->p_, name, value));
-}
-
-template <typename T>
-inline void IoBindingImpl<T>::BindOutput(const char* name, const OrtMemoryInfo* mem_info) {
-  ThrowOnError(GetApi().BindOutputToDevice(this->p_, name, mem_info));
-}
-
-template <typename T>
-inline void IoBindingImpl<T>::ClearBoundInputs() {
-  GetApi().ClearBoundInputs(this->p_);
-}
-
-template <typename T>
-inline void IoBindingImpl<T>::ClearBoundOutputs() {
-  GetApi().ClearBoundOutputs(this->p_);
-}
-
-template <typename T>
-inline void IoBindingImpl<T>::SynchronizeInputs() {
-  ThrowOnError(GetApi().SynchronizeBoundInputs(this->p_));
-}
-
-template <typename T>
-inline void IoBindingImpl<T>::SynchronizeOutputs() {
-  ThrowOnError(GetApi().SynchronizeBoundOutputs(this->p_));
-}
-
-namespace binding_utils {
-inline std::vector<std::string> GetOutputNamesHelper(const OrtIoBinding* binding, OrtAllocator* allocator) {
-  std::vector<std::string> result;
-  auto free_fn = detail::AllocatedFree(allocator);
-  using Ptr = std::unique_ptr<void, decltype(free_fn)>;
-
-  char* buffer = nullptr;
-  size_t* lengths = nullptr;
-  size_t count = 0;
-  ThrowOnError(GetApi().GetBoundOutputNames(binding, allocator, &buffer, &lengths, &count));
-
-  if (count == 0) {
-    return result;
-  }
-
-  Ptr buffer_g(buffer, free_fn);
-  Ptr lengths_g(lengths, free_fn);
-
-  result.reserve(count);
-  for (size_t i = 0; i < count; ++i) {
-    auto sz = *lengths;
-    result.emplace_back(buffer, sz);
-    buffer += sz;
-    ++lengths;
-  }
-  return result;
-}
-
-inline std::vector<Value> GetOutputValuesHelper(const OrtIoBinding* binding, OrtAllocator* allocator) {
-  std::vector<Value> result;
-  size_t owned = 0;
-  size_t output_count = 0;
-  // Lambda to release the buffer when no longer needed and
-  // make sure that we destroy all instances on exception
-  auto free_fn = [&owned, &output_count, allocator](OrtValue** buffer) {
-    if (buffer) {
-      while (owned < output_count) {
-        auto* p = buffer + owned++;
-        GetApi().ReleaseValue(*p);
-      }
-      allocator->Free(allocator, buffer);
-    }
-  };
-  using Ptr = std::unique_ptr<OrtValue*, decltype(free_fn)>;
-
-  OrtValue** output_buffer = nullptr;
-  ThrowOnError(GetApi().GetBoundOutputValues(binding, allocator, &output_buffer, &output_count));
-  if (output_count == 0) {
-    return result;
-  }
-
-  Ptr buffer_g(output_buffer, free_fn);
-
-  result.reserve(output_count);
-  for (size_t i = 0; i < output_count; ++i) {
-    result.emplace_back(output_buffer[i]);
-    ++owned;
-  }
-  return result;
-}
-
-}  // namespace binding_utils
-}  // namespace detail
-
-inline IoBinding::IoBinding(Session& session) {
-  ThrowOnError(GetApi().CreateIoBinding(session, &this->p_));
-}
-
-inline ArenaCfg::ArenaCfg(size_t max_mem, int arena_extend_strategy, int initial_chunk_size_bytes, int max_dead_bytes_per_chunk) {
-  ThrowOnError(GetApi().CreateArenaCfg(max_mem, arena_extend_strategy, initial_chunk_size_bytes, max_dead_bytes_per_chunk, &p_));
-}
-
-inline ThreadingOptions::ThreadingOptions() {
-  ThrowOnError(GetApi().CreateThreadingOptions(&p_));
-}
-
-inline ThreadingOptions& ThreadingOptions::SetGlobalIntraOpNumThreads(int intra_op_num_threads) {
-  ThrowOnError(GetApi().SetGlobalIntraOpNumThreads(p_, intra_op_num_threads));
-  return *this;
-}
-
-inline ThreadingOptions& ThreadingOptions::SetGlobalInterOpNumThreads(int inter_op_num_threads) {
-  ThrowOnError(GetApi().SetGlobalInterOpNumThreads(p_, inter_op_num_threads));
-  return *this;
-}
-
-inline ThreadingOptions& ThreadingOptions::SetGlobalSpinControl(int allow_spinning) {
-  ThrowOnError(GetApi().SetGlobalSpinControl(p_, allow_spinning));
-  return *this;
-}
-
-inline ThreadingOptions& ThreadingOptions::SetGlobalDenormalAsZero() {
-  ThrowOnError(GetApi().SetGlobalDenormalAsZero(p_));
-  return *this;
-}
-
-inline ThreadingOptions& ThreadingOptions::SetGlobalCustomCreateThreadFn(OrtCustomCreateThreadFn ort_custom_create_thread_fn) {
-  ThrowOnError(GetApi().SetGlobalCustomCreateThreadFn(p_, ort_custom_create_thread_fn));
-  return *this;
-}
-
-inline ThreadingOptions& ThreadingOptions::SetGlobalCustomThreadCreationOptions(void* ort_custom_thread_creation_options) {
-  ThrowOnError(GetApi().SetGlobalCustomThreadCreationOptions(p_, ort_custom_thread_creation_options));
-  return *this;
-}
-
-inline ThreadingOptions& ThreadingOptions::SetGlobalCustomJoinThreadFn(OrtCustomJoinThreadFn ort_custom_join_thread_fn) {
-  ThrowOnError(GetApi().SetGlobalCustomJoinThreadFn(p_, ort_custom_join_thread_fn));
-  return *this;
-}
-
-inline Env::Env(OrtLoggingLevel logging_level, _In_ const char* logid) {
-  ThrowOnError(GetApi().CreateEnv(logging_level, logid, &p_));
-  if (strcmp(logid, "onnxruntime-node") == 0) {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_NODEJS));
-  } else {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_CPLUSPLUS));
-  }
-}
-
-inline Env::Env(OrtLoggingLevel logging_level, const char* logid, OrtLoggingFunction logging_function, void* logger_param) {
-  ThrowOnError(GetApi().CreateEnvWithCustomLogger(logging_function, logger_param, logging_level, logid, &p_));
-  if (strcmp(logid, "onnxruntime-node") == 0) {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_NODEJS));
-  } else {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_CPLUSPLUS));
-  }
-}
-
-inline Env::Env(const OrtThreadingOptions* tp_options, OrtLoggingLevel logging_level, _In_ const char* logid) {
-  ThrowOnError(GetApi().CreateEnvWithGlobalThreadPools(logging_level, logid, tp_options, &p_));
-  if (strcmp(logid, "onnxruntime-node") == 0) {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_NODEJS));
-  } else {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_CPLUSPLUS));
-  }
-}
-
-inline Env::Env(const OrtThreadingOptions* tp_options, OrtLoggingFunction logging_function, void* logger_param,
-                OrtLoggingLevel logging_level, _In_ const char* logid) {
-  ThrowOnError(GetApi().CreateEnvWithCustomLoggerAndGlobalThreadPools(logging_function, logger_param, logging_level, logid, tp_options, &p_));
-  if (strcmp(logid, "onnxruntime-node") == 0) {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_NODEJS));
-  } else {
-    ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_CPLUSPLUS));
-  }
-}
-
-inline Env& Env::EnableTelemetryEvents() {
-  ThrowOnError(GetApi().EnableTelemetryEvents(p_));
-  return *this;
-}
-
-inline Env& Env::DisableTelemetryEvents() {
-  ThrowOnError(GetApi().DisableTelemetryEvents(p_));
-  return *this;
-}
-
-inline Env& Env::UpdateEnvWithCustomLogLevel(OrtLoggingLevel log_severity_level) {
-  ThrowOnError(GetApi().UpdateEnvWithCustomLogLevel(p_, log_severity_level));
-  return *this;
-}
-
-inline Env& Env::CreateAndRegisterAllocator(const OrtMemoryInfo* mem_info, const OrtArenaCfg* arena_cfg) {
-  ThrowOnError(GetApi().CreateAndRegisterAllocator(p_, mem_info, arena_cfg));
-  return *this;
-}
-
-inline Env& Env::CreateAndRegisterAllocatorV2(const std::string& provider_type, const OrtMemoryInfo* mem_info, const std::unordered_map<std::string, std::string>& options, const OrtArenaCfg* arena_cfg) {
-  std::vector<const char*> keys, values;
-  auto num_entries = options.size();
-  if (num_entries > 0) {
-    keys.reserve(num_entries);
-    values.reserve(num_entries);
-    for (const auto& entry : options) {
-      keys.push_back(entry.first.c_str());
-      values.push_back(entry.second.c_str());
-    }
-  }
-  ThrowOnError(GetApi().CreateAndRegisterAllocatorV2(p_, provider_type.c_str(), mem_info, arena_cfg, keys.data(), values.data(), num_entries));
-  return *this;
-}
-
-inline CustomOpDomain::CustomOpDomain(const char* domain) {
-  ThrowOnError(GetApi().CreateCustomOpDomain(domain, &p_));
-}
-
-inline void CustomOpDomain::Add(const OrtCustomOp* op) {
-  ThrowOnError(GetApi().CustomOpDomain_Add(p_, op));
-}
-
-inline RunOptions::RunOptions() {
-  ThrowOnError(GetApi().CreateRunOptions(&p_));
-}
-
-inline RunOptions& RunOptions::SetRunLogVerbosityLevel(int level) {
-  ThrowOnError(GetApi().RunOptionsSetRunLogVerbosityLevel(p_, level));
-  return *this;
-}
-
-inline RunOptions& RunOptions::SetRunLogSeverityLevel(int level) {
-  ThrowOnError(GetApi().RunOptionsSetRunLogSeverityLevel(p_, level));
-  return *this;
-}
-
-inline int RunOptions::GetRunLogVerbosityLevel() const {
-  int out;
-  ThrowOnError(GetApi().RunOptionsGetRunLogVerbosityLevel(p_, &out));
-  return out;
-}
-
-inline int RunOptions::GetRunLogSeverityLevel() const {
-  int out;
-  ThrowOnError(GetApi().RunOptionsGetRunLogSeverityLevel(p_, &out));
-  return out;
-}
-
-inline RunOptions& RunOptions::SetRunTag(const char* run_tag) {
-  ThrowOnError(GetApi().RunOptionsSetRunTag(p_, run_tag));
-  return *this;
-}
-
-inline const char* RunOptions::GetRunTag() const {
-  const char* out;
-  ThrowOnError(GetApi().RunOptionsGetRunTag(p_, &out));
-  return out;
-}
-
-inline RunOptions& RunOptions::AddConfigEntry(const char* config_key, const char* config_value) {
-  ThrowOnError(GetApi().AddRunConfigEntry(p_, config_key, config_value));
-  return *this;
-}
-
-inline RunOptions& RunOptions::SetTerminate() {
-  ThrowOnError(GetApi().RunOptionsSetTerminate(p_));
-  return *this;
-}
-
-inline RunOptions& RunOptions::UnsetTerminate() {
-  ThrowOnError(GetApi().RunOptionsUnsetTerminate(p_));
-  return *this;
-}
-
-namespace detail {
-
-template <typename T>
-inline Ort::SessionOptions ConstSessionOptionsImpl<T>::Clone() const {
-  OrtSessionOptions* out;
-  ThrowOnError(GetApi().CloneSessionOptions(this->p_, &out));
-  return SessionOptions{out};
-}
-
-template <typename T>
-inline std::string ConstSessionOptionsImpl<T>::GetConfigEntry(const char* config_key) const {
-  size_t size = 0;
-  // Feed nullptr for the data buffer to query the true size of the string value
-  Ort::ThrowOnError(GetApi().GetSessionConfigEntry(this->p_, config_key, nullptr, &size));
-
-  std::string out;
-  out.resize(size);
-  Ort::ThrowOnError(GetApi().GetSessionConfigEntry(this->p_, config_key, &out[0], &size));
-  out.resize(size - 1);  // remove the terminating character '\0'
-
-  return out;
-}
-
-template <typename T>
-inline bool ConstSessionOptionsImpl<T>::HasConfigEntry(const char* config_key) const {
-  int out = 0;
-  Ort::ThrowOnError(GetApi().HasSessionConfigEntry(this->p_, config_key, &out));
-  return static_cast<bool>(out);
-}
-
-template <typename T>
-inline std::string ConstSessionOptionsImpl<T>::GetConfigEntryOrDefault(const char* config_key, const std::string& def) {
-  if (!this->HasConfigEntry(config_key)) {
-    return def;
-  }
-
-  return this->GetConfigEntry(config_key);
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetIntraOpNumThreads(int intra_op_num_threads) {
-  ThrowOnError(GetApi().SetIntraOpNumThreads(this->p_, intra_op_num_threads));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetInterOpNumThreads(int inter_op_num_threads) {
-  ThrowOnError(GetApi().SetInterOpNumThreads(this->p_, inter_op_num_threads));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetGraphOptimizationLevel(GraphOptimizationLevel graph_optimization_level) {
-  ThrowOnError(GetApi().SetSessionGraphOptimizationLevel(this->p_, graph_optimization_level));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetDeterministicCompute(bool value) {
-  ThrowOnError(GetApi().SetDeterministicCompute(this->p_, value));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetOptimizedModelFilePath(const ORTCHAR_T* optimized_model_filepath) {
-  ThrowOnError(GetApi().SetOptimizedModelFilePath(this->p_, optimized_model_filepath));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::EnableProfiling(const ORTCHAR_T* profile_file_prefix) {
-  ThrowOnError(GetApi().EnableProfiling(this->p_, profile_file_prefix));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::DisableProfiling() {
-  ThrowOnError(GetApi().DisableProfiling(this->p_));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::EnableOrtCustomOps() {
-  ThrowOnError(GetApi().EnableOrtCustomOps(this->p_));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::EnableMemPattern() {
-  ThrowOnError(GetApi().EnableMemPattern(this->p_));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::DisableMemPattern() {
-  ThrowOnError(GetApi().DisableMemPattern(this->p_));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::EnableCpuMemArena() {
-  ThrowOnError(GetApi().EnableCpuMemArena(this->p_));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::DisableCpuMemArena() {
-  ThrowOnError(GetApi().DisableCpuMemArena(this->p_));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetExecutionMode(ExecutionMode execution_mode) {
-  ThrowOnError(GetApi().SetSessionExecutionMode(this->p_, execution_mode));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetLogId(const char* logid) {
-  ThrowOnError(GetApi().SetSessionLogId(this->p_, logid));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetLogSeverityLevel(int level) {
-  ThrowOnError(GetApi().SetSessionLogSeverityLevel(this->p_, level));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::Add(OrtCustomOpDomain* custom_op_domain) {
-  ThrowOnError(GetApi().AddCustomOpDomain(this->p_, custom_op_domain));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AddConfigEntry(const char* config_key, const char* config_value) {
-  ThrowOnError(GetApi().AddSessionConfigEntry(this->p_, config_key, config_value));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AddInitializer(const char* name, const OrtValue* ort_val) {
-  ThrowOnError(GetApi().AddInitializer(this->p_, name, ort_val));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::DisablePerSessionThreads() {
-  ThrowOnError(GetApi().DisablePerSessionThreads(this->p_));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AddExternalInitializers(const std::vector<std::string>& names,
-                                                                             const std::vector<Value>& ort_values) {
-  const size_t inputs_num = names.size();
-  if (inputs_num != ort_values.size()) {
-    ORT_CXX_API_THROW("Expecting names and ort_values to have the same length", ORT_INVALID_ARGUMENT);
-  }
-  std::vector<const char*> names_ptr;
-  std::vector<const OrtValue*> ort_values_ptrs;
-  names_ptr.reserve(inputs_num);
-  ort_values_ptrs.reserve(inputs_num);
-  for (size_t i = 0; i < inputs_num; ++i) {
-    names_ptr.push_back(names[i].c_str());
-    ort_values_ptrs.push_back(ort_values[i]);
-  }
-  ThrowOnError(GetApi().AddExternalInitializers(this->p_, names_ptr.data(), ort_values_ptrs.data(), inputs_num));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_CUDA(const OrtCUDAProviderOptions& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_CUDA(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_CUDA_V2(const OrtCUDAProviderOptionsV2& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_CUDA_V2(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_ROCM(const OrtROCMProviderOptions& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_ROCM(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_TensorRT(const OrtTensorRTProviderOptions& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_TensorRT(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_TensorRT_V2(const OrtTensorRTProviderOptionsV2& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_TensorRT_V2(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_MIGraphX(const OrtMIGraphXProviderOptions& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_MIGraphX(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_CANN(const OrtCANNProviderOptions& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_CANN(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_Dnnl(const OrtDnnlProviderOptions& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_Dnnl(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider(
-    const std::string& provider_name,
-    const std::unordered_map<std::string, std::string>& provider_options) {
-  auto num_entries = provider_options.size();
-  std::vector<const char*> keys, values;
-  if (num_entries > 0) {
-    keys.reserve(num_entries);
-    values.reserve(num_entries);
-
-    for (const auto& entry : provider_options) {
-      keys.push_back(entry.first.c_str());
-      values.push_back(entry.second.c_str());
-    }
-  }
-
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider(this->p_, provider_name.c_str(),
-                                                              keys.data(), values.data(), num_entries));
-
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetCustomCreateThreadFn(OrtCustomCreateThreadFn ort_custom_create_thread_fn) {
-  ThrowOnError(GetApi().SessionOptionsSetCustomCreateThreadFn(this->p_, ort_custom_create_thread_fn));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetCustomThreadCreationOptions(void* ort_custom_thread_creation_options) {
-  ThrowOnError(GetApi().SessionOptionsSetCustomThreadCreationOptions(this->p_, ort_custom_thread_creation_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::SetCustomJoinThreadFn(OrtCustomJoinThreadFn ort_custom_join_thread_fn) {
-  ThrowOnError(GetApi().SessionOptionsSetCustomJoinThreadFn(this->p_, ort_custom_join_thread_fn));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_OpenVINO(const OrtOpenVINOProviderOptions& provider_options) {
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_OpenVINO(this->p_, &provider_options));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::AppendExecutionProvider_OpenVINO_V2(const std::unordered_map<std::string, std::string>& provider_options) {
-  auto num_entries = provider_options.size();
-  std::vector<const char*> keys, values;
-  if (num_entries > 0) {
-    keys.reserve(num_entries);
-    values.reserve(num_entries);
-
-    for (const auto& entry : provider_options) {
-      keys.push_back(entry.first.c_str());
-      values.push_back(entry.second.c_str());
-    }
-  }
-
-  ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_OpenVINO_V2(this->p_,
-                                                                          keys.data(), values.data(), num_entries));
-
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::RegisterCustomOpsLibrary(const ORTCHAR_T* library_name,
-                                                                              const CustomOpConfigs& custom_op_configs) {
-  // Add custom op config entries before registering the custom op library. Otherwise, the config entries _may_ be ignored by
-  // the custom op library.
-  for (const auto& config_iter : custom_op_configs.GetFlattenedConfigs()) {
-    AddConfigEntry(config_iter.first.c_str(), config_iter.second.c_str());
-  }
-
-  ThrowOnError(GetApi().RegisterCustomOpsLibrary_V2(this->p_, library_name));
-  return *this;
-}
-
-template <typename T>
-inline SessionOptionsImpl<T>& SessionOptionsImpl<T>::RegisterCustomOpsUsingFunction(const char* registration_function_name) {
-  ThrowOnError(GetApi().RegisterCustomOpsUsingFunction(this->p_, registration_function_name));
-  return *this;
-}
-
-/// Session
-template <typename T>
-inline size_t ConstSessionImpl<T>::GetInputCount() const {
-  size_t out;
-  ThrowOnError(GetApi().SessionGetInputCount(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline size_t ConstSessionImpl<T>::GetOutputCount() const {
-  size_t out;
-  ThrowOnError(GetApi().SessionGetOutputCount(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline size_t ConstSessionImpl<T>::GetOverridableInitializerCount() const {
-  size_t out;
-  ThrowOnError(GetApi().SessionGetOverridableInitializerCount(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline AllocatedStringPtr ConstSessionImpl<T>::GetInputNameAllocated(size_t index, OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().SessionGetInputName(this->p_, index, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-template <typename T>
-inline AllocatedStringPtr ConstSessionImpl<T>::GetOutputNameAllocated(size_t index, OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().SessionGetOutputName(this->p_, index, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-template <typename T>
-inline AllocatedStringPtr ConstSessionImpl<T>::GetOverridableInitializerNameAllocated(size_t index, OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().SessionGetOverridableInitializerName(this->p_, index, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-template <typename T>
-inline uint64_t ConstSessionImpl<T>::GetProfilingStartTimeNs() const {
-  uint64_t out;
-  ThrowOnError(GetApi().SessionGetProfilingStartTimeNs(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline ModelMetadata ConstSessionImpl<T>::GetModelMetadata() const {
-  OrtModelMetadata* out;
-  ThrowOnError(GetApi().SessionGetModelMetadata(this->p_, &out));
-  return ModelMetadata{out};
-}
-
-template <typename T>
-inline TypeInfo ConstSessionImpl<T>::GetInputTypeInfo(size_t index) const {
-  OrtTypeInfo* out;
-  ThrowOnError(GetApi().SessionGetInputTypeInfo(this->p_, index, &out));
-  return TypeInfo{out};
-}
-
-template <typename T>
-inline TypeInfo ConstSessionImpl<T>::GetOutputTypeInfo(size_t index) const {
-  OrtTypeInfo* out;
-  ThrowOnError(GetApi().SessionGetOutputTypeInfo(this->p_, index, &out));
-  return TypeInfo{out};
-}
-
-template <typename T>
-inline TypeInfo ConstSessionImpl<T>::GetOverridableInitializerTypeInfo(size_t index) const {
-  OrtTypeInfo* out;
-  ThrowOnError(GetApi().SessionGetOverridableInitializerTypeInfo(this->p_, index, &out));
-  return TypeInfo{out};
-}
-
-template <typename T>
-inline std::vector<Value> SessionImpl<T>::Run(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count,
-                                              const char* const* output_names, size_t output_count) {
-  std::vector<Value> output_values;
-  output_values.reserve(output_count);
-  for (size_t i = 0; i < output_count; i++)
-    output_values.emplace_back(nullptr);
-  Run(run_options, input_names, input_values, input_count, output_names, output_values.data(), output_count);
-  return output_values;
-}
-
-template <typename T>
-inline void SessionImpl<T>::Run(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count,
-                                const char* const* output_names, Value* output_values, size_t output_count) {
-  static_assert(sizeof(Value) == sizeof(OrtValue*), "Value is really just an array of OrtValue* in memory, so we can reinterpret_cast safely");
-  auto ort_input_values = reinterpret_cast<const OrtValue* const*>(input_values);
-  auto ort_output_values = reinterpret_cast<OrtValue**>(output_values);
-  ThrowOnError(GetApi().Run(this->p_, run_options, input_names, ort_input_values, input_count, output_names, output_count, ort_output_values));
-}
-
-template <typename T>
-inline void SessionImpl<T>::Run(const RunOptions& run_options, const IoBinding& io_binding) {
-  ThrowOnError(GetApi().RunWithBinding(this->p_, run_options, io_binding));
-}
-
-template <typename T>
-inline void SessionImpl<T>::RunAsync(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count,
-                                     const char* const* output_names, Value* output_values, size_t output_count, RunAsyncCallbackFn callback, void* user_data) {
-  auto ort_input_values = reinterpret_cast<const OrtValue* const*>(input_values);
-  auto ort_output_values = reinterpret_cast<OrtValue**>(output_values);
-  ThrowOnError(GetApi().RunAsync(this->p_, run_options, input_names,
-                                 ort_input_values, input_count, output_names, output_count,
-                                 ort_output_values, callback, user_data));
-}
-
-template <typename T>
-inline AllocatedStringPtr SessionImpl<T>::EndProfilingAllocated(OrtAllocator* allocator) {
-  char* out = nullptr;
-  ThrowOnError(GetApi().SessionEndProfiling(this->p_, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-}  // namespace detail
-
-inline SessionOptions::SessionOptions() {
-  ThrowOnError(GetApi().CreateSessionOptions(&this->p_));
-}
-
-/// CustomOpConfigs
-inline std::string detail::MakeCustomOpConfigEntryKey(const char* custom_op_name, const char* config) {
-  std::string config_key = "custom_op.";
-
-  config_key += custom_op_name;
-  config_key += ".";
-  config_key += config;
-
-  return config_key;
-}
-
-inline CustomOpConfigs& CustomOpConfigs::AddConfig(const char* custom_op_name, const char* config_key, const char* config_value) {
-  const std::string full_flat_key = detail::MakeCustomOpConfigEntryKey(custom_op_name, config_key);
-  flat_configs_[full_flat_key] = config_value;
-  return *this;
-}
-
-inline const std::unordered_map<std::string, std::string>& CustomOpConfigs::GetFlattenedConfigs() const {
-  return flat_configs_;
-}
-
-inline Session::Session(const Env& env, const ORTCHAR_T* model_path, const SessionOptions& options) {
-  ThrowOnError(GetApi().CreateSession(env, model_path, options, &this->p_));
-}
-
-inline Session::Session(const Env& env, const ORTCHAR_T* model_path, const SessionOptions& options,
-                        OrtPrepackedWeightsContainer* prepacked_weights_container) {
-  ThrowOnError(GetApi().CreateSessionWithPrepackedWeightsContainer(env, model_path, options, prepacked_weights_container, &this->p_));
-}
-
-inline Session::Session(const Env& env, const void* model_data, size_t model_data_length, const SessionOptions& options) {
-  ThrowOnError(GetApi().CreateSessionFromArray(env, model_data, model_data_length, options, &this->p_));
-}
-
-inline Session::Session(const Env& env, const void* model_data, size_t model_data_length,
-                        const SessionOptions& options, OrtPrepackedWeightsContainer* prepacked_weights_container) {
-  ThrowOnError(GetApi().CreateSessionFromArrayWithPrepackedWeightsContainer(env, model_data, model_data_length, options,
-                                                                            prepacked_weights_container, &this->p_));
-}
-
-inline AllocatedStringPtr ModelMetadata::GetProducerNameAllocated(OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().ModelMetadataGetProducerName(p_, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-inline AllocatedStringPtr ModelMetadata::GetGraphNameAllocated(OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().ModelMetadataGetGraphName(p_, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-inline AllocatedStringPtr ModelMetadata::GetDomainAllocated(OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().ModelMetadataGetDomain(p_, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-inline AllocatedStringPtr Ort::ModelMetadata::GetDescriptionAllocated(OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().ModelMetadataGetDescription(p_, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-inline AllocatedStringPtr ModelMetadata::GetGraphDescriptionAllocated(OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().ModelMetadataGetGraphDescription(p_, allocator, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-inline AllocatedStringPtr ModelMetadata::LookupCustomMetadataMapAllocated(const char* key, OrtAllocator* allocator) const {
-  char* out;
-  ThrowOnError(GetApi().ModelMetadataLookupCustomMetadataMap(p_, allocator, key, &out));
-  return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
-}
-
-inline std::vector<AllocatedStringPtr> ModelMetadata::GetCustomMetadataMapKeysAllocated(OrtAllocator* allocator) const {
-  auto deletor = detail::AllocatedFree(allocator);
-  std::vector<AllocatedStringPtr> result;
-
-  char** out = nullptr;
-  int64_t num_keys = 0;
-  ThrowOnError(GetApi().ModelMetadataGetCustomMetadataMapKeys(p_, allocator, &out, &num_keys));
-  if (num_keys <= 0) {
-    return result;
-  }
-
-  // array of pointers will be freed
-  std::unique_ptr<void, decltype(deletor)> array_guard(out, deletor);
-  // reserve may throw
-  auto strings_deletor = [&deletor, num_keys](char** out) { for(int64_t i = 0; i < num_keys; ++i) deletor(out[i]); };
-  std::unique_ptr<char*, decltype(strings_deletor)> strings_guard(out, strings_deletor);
-  result.reserve(static_cast<size_t>(num_keys));
-  strings_guard.release();
-  for (int64_t i = 0; i < num_keys; ++i) {
-    result.push_back(AllocatedStringPtr(out[i], deletor));
-  }
-
-  return result;
-}
-
-inline int64_t ModelMetadata::GetVersion() const {
-  int64_t out;
-  ThrowOnError(GetApi().ModelMetadataGetVersion(p_, &out));
-  return out;
-}
-
-namespace detail {
-
-template <typename T>
-inline ONNXTensorElementDataType TensorTypeAndShapeInfoImpl<T>::GetElementType() const {
-  ONNXTensorElementDataType out;
-  ThrowOnError(GetApi().GetTensorElementType(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline size_t TensorTypeAndShapeInfoImpl<T>::GetElementCount() const {
-  size_t out;
-  ThrowOnError(GetApi().GetTensorShapeElementCount(this->p_, &out));
-  return static_cast<size_t>(out);
-}
-
-template <typename T>
-inline size_t TensorTypeAndShapeInfoImpl<T>::GetDimensionsCount() const {
-  size_t out;
-  ThrowOnError(GetApi().GetDimensionsCount(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline void TensorTypeAndShapeInfoImpl<T>::GetDimensions(int64_t* values, size_t values_count) const {
-  ThrowOnError(GetApi().GetDimensions(this->p_, values, values_count));
-}
-
-template <typename T>
-inline void TensorTypeAndShapeInfoImpl<T>::GetSymbolicDimensions(const char** values, size_t values_count) const {
-  ThrowOnError(GetApi().GetSymbolicDimensions(this->p_, values, values_count));
-}
-
-template <typename T>
-inline std::vector<int64_t> TensorTypeAndShapeInfoImpl<T>::GetShape() const {
-  std::vector<int64_t> out(GetDimensionsCount(), 0);
-  ThrowOnError(GetApi().GetDimensions(this->p_, out.data(), out.size()));
-  return out;
-}
-
-template <typename T>
-inline ConstTensorTypeAndShapeInfo TypeInfoImpl<T>::GetTensorTypeAndShapeInfo() const {
-  const OrtTensorTypeAndShapeInfo* out;
-  ThrowOnError(GetApi().CastTypeInfoToTensorInfo(this->p_, &out));
-  return ConstTensorTypeAndShapeInfo{out};
-}
-
-template <typename T>
-inline ConstSequenceTypeInfo TypeInfoImpl<T>::GetSequenceTypeInfo() const {
-  const OrtSequenceTypeInfo* out;
-  ThrowOnError(GetApi().CastTypeInfoToSequenceTypeInfo(this->p_, &out));
-  return ConstSequenceTypeInfo{out};
-}
-
-template <typename T>
-inline ConstMapTypeInfo TypeInfoImpl<T>::GetMapTypeInfo() const {
-  const OrtMapTypeInfo* out;
-  ThrowOnError(GetApi().CastTypeInfoToMapTypeInfo(this->p_, &out));
-  return ConstMapTypeInfo{out};
-}
-
-template <typename T>
-inline ONNXType TypeInfoImpl<T>::GetONNXType() const {
-  ONNXType out;
-  ThrowOnError(GetApi().GetOnnxTypeFromTypeInfo(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline TypeInfo SequenceTypeInfoImpl<T>::GetSequenceElementType() const {
-  OrtTypeInfo* output;
-  ThrowOnError(GetApi().GetSequenceElementType(this->p_, &output));
-  return TypeInfo{output};
-}
-
-template <typename T>
-inline TypeInfo OptionalTypeInfoImpl<T>::GetOptionalElementType() const {
-  OrtTypeInfo* info;
-  ThrowOnError(GetApi().GetOptionalContainedTypeInfo(this->p_, &info));
-  return TypeInfo{info};
-}
-
-template <typename T>
-inline ONNXTensorElementDataType MapTypeInfoImpl<T>::GetMapKeyType() const {
-  ONNXTensorElementDataType out;
-  ThrowOnError(GetApi().GetMapKeyType(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline TypeInfo MapTypeInfoImpl<T>::GetMapValueType() const {
-  OrtTypeInfo* output;
-  ThrowOnError(GetApi().GetMapValueType(this->p_, &output));
-  return TypeInfo{output};
-}
-
-template <typename T>
-inline ConstOptionalTypeInfo TypeInfoImpl<T>::GetOptionalTypeInfo() const {
-  const OrtOptionalTypeInfo* info;
-  ThrowOnError(GetApi().CastTypeInfoToOptionalTypeInfo(this->p_, &info));
-  return ConstOptionalTypeInfo{info};
-}
-
-}  // namespace detail
-
-namespace detail {
-
-template <typename T>
-template <typename R>
-inline void ConstValueImpl<T>::GetOpaqueData(const char* domain, const char* type_name, R& out) const {
-  ThrowOnError(GetApi().GetOpaqueValue(domain, type_name, this->p_, &out, sizeof(R)));
-}
-
-template <typename T>
-inline bool ConstValueImpl<T>::IsTensor() const {
-  int out;
-  ThrowOnError(GetApi().IsTensor(this->p_, &out));
-  return out != 0;
-}
-
-template <typename T>
-inline bool ConstValueImpl<T>::HasValue() const {
-  int out;
-  ThrowOnError(GetApi().HasValue(this->p_, &out));
-  return out != 0;
-}
-
-template <typename T>
-inline size_t ConstValueImpl<T>::GetCount() const {
-  size_t out;
-  ThrowOnError(GetApi().GetValueCount(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline Value ConstValueImpl<T>::GetValue(int index, OrtAllocator* allocator) const {
-  OrtValue* out;
-  ThrowOnError(GetApi().GetValue(this->p_, index, allocator, &out));
-  return Value{out};
-}
-
-template <typename T>
-inline size_t ConstValueImpl<T>::GetStringTensorDataLength() const {
-  size_t out;
-  ThrowOnError(GetApi().GetStringTensorDataLength(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline size_t ConstValueImpl<T>::GetStringTensorElementLength(size_t element_index) const {
-  size_t out;
-  ThrowOnError(GetApi().GetStringTensorElementLength(this->p_, element_index, &out));
-  return out;
-}
-
-template <typename T>
-template <typename R>
-inline const R* ConstValueImpl<T>::GetTensorData() const {
-  R* out;
-  ThrowOnError(GetApi().GetTensorMutableData(const_cast<OrtValue*>(this->p_), (void**)&out));
-  return out;
-}
-
-template <typename T>
-inline const void* ConstValueImpl<T>::GetTensorRawData() const {
-  void* out;
-  ThrowOnError(GetApi().GetTensorMutableData(const_cast<OrtValue*>(this->p_), &out));
-  return out;
-}
-
-template <typename T>
-inline TypeInfo ConstValueImpl<T>::GetTypeInfo() const {
-  OrtTypeInfo* output;
-  ThrowOnError(GetApi().GetTypeInfo(this->p_, &output));
-  return TypeInfo{output};
-}
-
-template <typename T>
-inline TensorTypeAndShapeInfo ConstValueImpl<T>::GetTensorTypeAndShapeInfo() const {
-  OrtTensorTypeAndShapeInfo* output;
-  ThrowOnError(GetApi().GetTensorTypeAndShape(this->p_, &output));
-  return TensorTypeAndShapeInfo{output};
-}
-
-template <typename T>
-inline ConstMemoryInfo ConstValueImpl<T>::GetTensorMemoryInfo() const {
-  const OrtMemoryInfo* mem_info;
-  ThrowOnError(GetApi().GetTensorMemoryInfo(this->p_, &mem_info));
-  return ConstMemoryInfo(mem_info);
-}
-
-template <typename T>
-inline void ConstValueImpl<T>::GetStringTensorElement(size_t buffer_length, size_t element_index, void* buffer) const {
-  ThrowOnError(GetApi().GetStringTensorElement(this->p_, buffer_length, element_index, buffer));
-}
-
-template <typename T>
-inline std::string ConstValueImpl<T>::GetStringTensorElement(size_t element_index) const {
-  size_t buffer_length;
-  ThrowOnError(GetApi().GetStringTensorElementLength(this->p_, element_index, &buffer_length));
-
-  std::string s;
-  s.resize(buffer_length);
-  ThrowOnError(GetApi().GetStringTensorElement(this->p_, buffer_length, element_index, &s[0]));
-  return s;
-}
-
-template <typename T>
-inline void ConstValueImpl<T>::GetStringTensorContent(void* buffer, size_t buffer_length, size_t* offsets, size_t offsets_count) const {
-  ThrowOnError(GetApi().GetStringTensorContent(this->p_, buffer, buffer_length, offsets, offsets_count));
-}
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-template <typename T>
-inline OrtSparseFormat ConstValueImpl<T>::GetSparseFormat() const {
-  OrtSparseFormat format;
-  ThrowOnError(GetApi().GetSparseTensorFormat(this->p_, &format));
-  return format;
-}
-
-template <typename T>
-inline TensorTypeAndShapeInfo ConstValueImpl<T>::GetSparseTensorValuesTypeAndShapeInfo() const {
-  OrtTensorTypeAndShapeInfo* output;
-  ThrowOnError(GetApi().GetSparseTensorValuesTypeAndShape(this->p_, &output));
-  return TensorTypeAndShapeInfo{output};
-}
-
-template <typename T>
-inline TensorTypeAndShapeInfo ConstValueImpl<T>::GetSparseTensorIndicesTypeShapeInfo(OrtSparseIndicesFormat indices_format) const {
-  OrtTensorTypeAndShapeInfo* output;
-  ThrowOnError(GetApi().GetSparseTensorIndicesTypeShape(this->p_, indices_format, &output));
-  return TensorTypeAndShapeInfo{output};
-}
-
-template <typename T>
-template <typename R>
-inline const R* ConstValueImpl<T>::GetSparseTensorIndicesData(OrtSparseIndicesFormat indices_format, size_t& num_indices) const {
-  const void* out;
-  ThrowOnError(GetApi().GetSparseTensorIndices(this->p_, indices_format, &num_indices, &out));
-  return reinterpret_cast<const R*>(out);
-}
-
-template <typename T>
-inline bool ConstValueImpl<T>::IsSparseTensor() const {
-  int out;
-  ThrowOnError(GetApi().IsSparseTensor(this->p_, &out));
-  return out != 0;
-}
-
-template <typename T>
-template <typename R>
-inline const R* ConstValueImpl<T>::GetSparseTensorValues() const {
-  const void* out;
-  ThrowOnError(GetApi().GetSparseTensorValues(this->p_, &out));
-  return reinterpret_cast<const R*>(out);
-}
-
-#endif
-
-template <typename T>
-void ValueImpl<T>::FillStringTensor(const char* const* s, size_t s_len) {
-  ThrowOnError(GetApi().FillStringTensor(this->p_, s, s_len));
-}
-
-template <typename T>
-void ValueImpl<T>::FillStringTensorElement(const char* s, size_t index) {
-  ThrowOnError(GetApi().FillStringTensorElement(this->p_, s, index));
-}
-
-template <typename T>
-inline char* ValueImpl<T>::GetResizedStringTensorElementBuffer(size_t index, size_t buffer_length) {
-  char* result;
-  ThrowOnError(GetApi().GetResizedStringTensorElementBuffer(this->p_, index, buffer_length, &result));
-  return result;
-}
-
-template <typename T>
-void* ValueImpl<T>::GetTensorMutableRawData() {
-  void* out;
-  ThrowOnError(GetApi().GetTensorMutableData(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-template <typename R>
-R* ValueImpl<T>::GetTensorMutableData() {
-  R* out;
-  ThrowOnError(GetApi().GetTensorMutableData(this->p_, (void**)&out));
-  return out;
-}
-
-template <typename T>
-template <typename R>
-R& ValueImpl<T>::At(const std::vector<int64_t>& location) {
-  static_assert(!std::is_same<T, std::string>::value, "this api does not support std::string");
-  R* out;
-  ThrowOnError(GetApi().TensorAt(this->p_, location.data(), location.size(), (void**)&out));
-  return *out;
-}
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-template <typename T>
-void ValueImpl<T>::UseCooIndices(int64_t* indices_data, size_t indices_num) {
-  ThrowOnError(GetApi().UseCooIndices(this->p_, indices_data, indices_num));
-}
-
-template <typename T>
-void ValueImpl<T>::UseCsrIndices(int64_t* inner_data, size_t inner_num, int64_t* outer_data, size_t outer_num) {
-  ThrowOnError(GetApi().UseCsrIndices(this->p_, inner_data, inner_num, outer_data, outer_num));
-}
-
-template <typename T>
-void ValueImpl<T>::UseBlockSparseIndices(const Shape& indices_shape, int32_t* indices_data) {
-  ThrowOnError(GetApi().UseBlockSparseIndices(this->p_, indices_shape.shape, indices_shape.shape_len, indices_data));
-}
-
-template <typename T>
-void ValueImpl<T>::FillSparseTensorCoo(const OrtMemoryInfo* mem_info, const OrtSparseValuesParam& values_param,
-                                       const int64_t* indices_data, size_t indices_num) {
-  ThrowOnError(GetApi().FillSparseTensorCoo(this->p_, mem_info, values_param.values_shape,
-                                            values_param.values_shape_len, values_param.data.p_data,
-                                            indices_data, indices_num));
-}
-
-template <typename T>
-void ValueImpl<T>::FillSparseTensorCsr(const OrtMemoryInfo* data_mem_info,
-                                       const OrtSparseValuesParam& values,
-                                       const int64_t* inner_indices_data, size_t inner_indices_num,
-                                       const int64_t* outer_indices_data, size_t outer_indices_num) {
-  ThrowOnError(GetApi().FillSparseTensorCsr(this->p_, data_mem_info, values.values_shape, values.values_shape_len, values.data.p_data,
-                                            inner_indices_data, inner_indices_num,
-                                            outer_indices_data, outer_indices_num));
-}
-
-template <typename T>
-void ValueImpl<T>::FillSparseTensorBlockSparse(const OrtMemoryInfo* data_mem_info,
-                                               const OrtSparseValuesParam& values,
-                                               const Shape& indices_shape,
-                                               const int32_t* indices_data) {
-  ThrowOnError(GetApi().FillSparseTensorBlockSparse(this->p_, data_mem_info, values.values_shape, values.values_shape_len, values.data.p_data,
-                                                    indices_shape.shape, indices_shape.shape_len,
-                                                    indices_data));
-}
-
-#endif  // !defined(DISABLE_SPARSE_TENSORS)
-
-}  // namespace detail
-
-template <typename T>
-inline Value Value::CreateTensor(const OrtMemoryInfo* info, T* p_data, size_t p_data_element_count, const int64_t* shape, size_t shape_len) {
-  return CreateTensor(info, p_data, p_data_element_count * sizeof(T), shape, shape_len, TypeToTensorType<T>::type);
-}
-
-inline Value Value::CreateTensor(const OrtMemoryInfo* info, void* p_data, size_t p_data_byte_count, const int64_t* shape, size_t shape_len,
-                                 ONNXTensorElementDataType type) {
-  OrtValue* out;
-  ThrowOnError(GetApi().CreateTensorWithDataAsOrtValue(info, p_data, p_data_byte_count, shape, shape_len, type, &out));
-  return Value{out};
-}
-
-template <typename T>
-inline Value Value::CreateTensor(OrtAllocator* allocator, const int64_t* shape, size_t shape_len) {
-  return CreateTensor(allocator, shape, shape_len, TypeToTensorType<T>::type);
-}
-
-inline Value Value::CreateTensor(OrtAllocator* allocator, const int64_t* shape, size_t shape_len, ONNXTensorElementDataType type) {
-  OrtValue* out;
-  ThrowOnError(GetApi().CreateTensorAsOrtValue(allocator, shape, shape_len, type, &out));
-  return Value{out};
-}
-
-#if !defined(DISABLE_SPARSE_TENSORS)
-
-template <typename T>
-inline Value Value::CreateSparseTensor(const OrtMemoryInfo* info, T* p_data, const Shape& dense_shape,
-                                       const Shape& values_shape) {
-  return CreateSparseTensor(info, p_data, dense_shape, values_shape, TypeToTensorType<T>::type);
-}
-
-inline Value Value::CreateSparseTensor(const OrtMemoryInfo* info, void* p_data, const Shape& dense_shape,
-                                       const Shape& values_shape, ONNXTensorElementDataType type) {
-  OrtValue* out;
-  ThrowOnError(GetApi().CreateSparseTensorWithValuesAsOrtValue(info, p_data, dense_shape.shape, dense_shape.shape_len,
-                                                               values_shape.shape, values_shape.shape_len, type, &out));
-  return Value{out};
-}
-
-template <typename T>
-inline Value Value::CreateSparseTensor(OrtAllocator* allocator, const Shape& dense_shape) {
-  return CreateSparseTensor(allocator, dense_shape, TypeToTensorType<T>::type);
-}
-
-inline Value Value::CreateSparseTensor(OrtAllocator* allocator, const Shape& dense_shape,
-                                       ONNXTensorElementDataType type) {
-  OrtValue* out;
-  ThrowOnError(GetApi().CreateSparseTensorAsOrtValue(allocator, dense_shape.shape, dense_shape.shape_len, type, &out));
-  return Value{out};
-}
-#endif  // !defined(DISABLE_SPARSE_TENSORS)
-
-inline Value Value::CreateMap(const Value& keys, const Value& values) {
-  OrtValue* out;
-  const OrtValue* inputs[2] = {keys, values};
-  ThrowOnError(GetApi().CreateValue(inputs, 2, ONNX_TYPE_MAP, &out));
-  return Value{out};
-}
-
-inline Value Value::CreateSequence(const std::vector<Value>& values) {
-  OrtValue* out;
-  std::vector<const OrtValue*> values_ort{values.data(), values.data() + values.size()};
-  ThrowOnError(GetApi().CreateValue(values_ort.data(), values_ort.size(), ONNX_TYPE_SEQUENCE, &out));
-  return Value{out};
-}
-
-template <typename T>
-inline Value Value::CreateOpaque(const char* domain, const char* type_name, const T& data_container) {
-  OrtValue* out;
-  ThrowOnError(GetApi().CreateOpaqueValue(domain, type_name, &data_container, sizeof(T), &out));
-  return Value{out};
-}
-
-//
-// Custom OP Inlines
-//
-inline Logger::Logger(const OrtLogger* logger) : logger_(logger) {
-  Ort::ThrowOnError(GetApi().Logger_GetLoggingSeverityLevel(this->logger_, &this->cached_severity_level_));
-}
-
-inline OrtLoggingLevel Logger::GetLoggingSeverityLevel() const noexcept {
-  return cached_severity_level_;
-}
-
-inline Status Logger::LogMessage(OrtLoggingLevel log_severity_level, const ORTCHAR_T* file_path, int line_number,
-                                 const char* func_name, const char* message) const noexcept {
-  OrtStatus* status = GetApi().Logger_LogMessage(logger_, log_severity_level, message, file_path, line_number,
-                                                 func_name);
-  return Status{status};
-}
-
-// Disable warnings about the format string not being a literal (-Wformat-nonliteral and -Wformat-security)
-// for gcc and clang. The alternative is to use actual C-style variadic parameters and apply
-// __attribute__(format(printf...)), which does not work with variadic templates.
-#if defined(__GNUC__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wformat-nonliteral"
-#pragma GCC diagnostic ignored "-Wformat-security"
-#elif defined(__clang__)
-#pragma clang diagnostic push
-#pragma clang diagnostic ignored "-Wformat-nonliteral"
-#pragma clang diagnostic ignored "-Wformat-security"
-#endif
-template <typename... Args>
-inline Status Logger::LogFormattedMessage(OrtLoggingLevel log_severity_level, const ORTCHAR_T* file_path,
-                                          int line_number, const char* func_name, const char* format,
-                                          Args&&... args) const noexcept {
-  int msg_len = std::snprintf(nullptr, 0U, format, std::forward<Args>(args)...);
-
-  if (msg_len < 0) {  // Formatting error
-    return Status("Failed to log message due to formatting error", OrtErrorCode::ORT_FAIL);
-  }
-
-  OrtStatus* status = nullptr;
-  const size_t buffer_size = static_cast<size_t>(msg_len) + 1U;
-
-  constexpr size_t kStackBufferSize = 1024;
-
-  if (buffer_size < kStackBufferSize) {
-    char buffer[kStackBufferSize];
-    snprintf(buffer, kStackBufferSize, format, std::forward<Args>(args)...);
-    status = GetApi().Logger_LogMessage(logger_, log_severity_level, buffer, file_path, line_number, func_name);
-  } else {
-    // std::make_unique is only supported starting at C++14.
-#if (__cplusplus >= 201402L) || (_MSC_VER >= 1900)
-    auto buffer = std::make_unique<char[]>(buffer_size);
-#else
-    std::unique_ptr<char[]> buffer(new char[buffer_size]);
-#endif
-    std::snprintf(buffer.get(), buffer_size, format, std::forward<Args>(args)...);
-    status = GetApi().Logger_LogMessage(logger_, log_severity_level, buffer.get(), file_path, line_number, func_name);
-  }
-
-  return Status{status};
-}
-// Re-enable -Wformat-nonliteral and -Wformat-security
-#if defined(__GNUC__)
-#pragma GCC diagnostic pop
-#elif defined(__clang__)
-#pragma clang diagnostic pop
-#endif
-
-inline KernelContext::KernelContext(OrtKernelContext* context) : ctx_(context) {
-}
-
-inline size_t KernelContext::GetInputCount() const {
-  size_t out = 0;
-  Ort::ThrowOnError(GetApi().KernelContext_GetInputCount(ctx_, &out));
-  return out;
-}
-
-inline size_t KernelContext::GetOutputCount() const {
-  size_t out = 0;
-  Ort::ThrowOnError(GetApi().KernelContext_GetOutputCount(ctx_, &out));
-  return out;
-}
-
-inline ConstValue KernelContext::GetInput(size_t index) const {
-  const OrtValue* out = nullptr;
-  Ort::ThrowOnError(GetApi().KernelContext_GetInput(ctx_, index, &out));
-  return ConstValue{out};
-}
-
-inline UnownedValue KernelContext::GetOutput(size_t index, const int64_t* dim_values, size_t dim_count) const {
-  OrtValue* out = nullptr;
-  Ort::ThrowOnError(GetApi().KernelContext_GetOutput(ctx_, index, dim_values, dim_count, &out));
-  return UnownedValue(out);
-}
-
-inline UnownedValue KernelContext::GetOutput(size_t index, const std::vector<int64_t>& dims) const {
-  OrtValue* out = nullptr;
-  Ort::ThrowOnError(GetApi().KernelContext_GetOutput(ctx_, index, dims.data(), dims.size(), &out));
-  return UnownedValue(out);
-}
-
-inline void* KernelContext::GetGPUComputeStream() const {
-  void* out = nullptr;
-  Ort::ThrowOnError(GetApi().KernelContext_GetGPUComputeStream(ctx_, &out));
-  return out;
-}
-
-inline OrtAllocator* KernelContext::GetAllocator(const OrtMemoryInfo& memory_info) const {
-  OrtAllocator* out = nullptr;
-  Ort::ThrowOnError(GetApi().KernelContext_GetAllocator(ctx_, &memory_info, &out));
-  return out;
-}
-
-inline Logger KernelContext::GetLogger() const {
-  const OrtLogger* out = nullptr;
-  ThrowOnError(GetApi().KernelContext_GetLogger(this->ctx_, &out));
-  return Logger{out};
-}
-
-inline void KernelContext::ParallelFor(void (*fn)(void*, size_t), size_t total, size_t num_batch, void* usr_data) const {
-  ThrowOnError(GetApi().KernelContext_ParallelFor(ctx_, fn, total, num_batch, usr_data));
-}
-
-inline OpAttr::OpAttr(const char* name, const void* data, int len, OrtOpAttrType type) {
-  Ort::ThrowOnError(GetApi().CreateOpAttr(name, data, len, type, &p_));
-}
-
-namespace detail {
-template <typename T>
-inline KernelInfo KernelInfoImpl<T>::Copy() const {
-  OrtKernelInfo* info_copy = nullptr;
-  Ort::ThrowOnError(GetApi().CopyKernelInfo(this->p_, &info_copy));
-  return KernelInfo{info_copy};
-}
-
-template <typename T>
-inline size_t KernelInfoImpl<T>::GetInputCount() const {
-  size_t out = 0;
-  ThrowOnError(GetApi().KernelInfo_GetInputCount(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline size_t KernelInfoImpl<T>::GetOutputCount() const {
-  size_t out = 0;
-  ThrowOnError(GetApi().KernelInfo_GetOutputCount(this->p_, &out));
-  return out;
-}
-
-template <typename T>
-inline std::string KernelInfoImpl<T>::GetInputName(size_t index) const {
-  size_t size = 0;
-
-  // Feed nullptr for the data buffer to query the true size of the string value
-  Ort::ThrowOnError(GetApi().KernelInfo_GetInputName(this->p_, index, nullptr, &size));
-
-  std::string out;
-  out.resize(size);
-  Ort::ThrowOnError(GetApi().KernelInfo_GetInputName(this->p_, index, &out[0], &size));
-  out.resize(size - 1);  // remove the terminating character '\0'
-
-  return out;
-}
-
-template <typename T>
-inline std::string KernelInfoImpl<T>::GetOutputName(size_t index) const {
-  size_t size = 0;
-
-  // Feed nullptr for the data buffer to query the true size of the string value
-  Ort::ThrowOnError(GetApi().KernelInfo_GetOutputName(this->p_, index, nullptr, &size));
-
-  std::string out;
-  out.resize(size);
-  Ort::ThrowOnError(GetApi().KernelInfo_GetOutputName(this->p_, index, &out[0], &size));
-  out.resize(size - 1);  // remove the terminating character '\0'
-
-  return out;
-}
-
-template <typename T>
-inline TypeInfo KernelInfoImpl<T>::GetInputTypeInfo(size_t index) const {
-  OrtTypeInfo* out = nullptr;
-  ThrowOnError(GetApi().KernelInfo_GetInputTypeInfo(this->p_, index, &out));
-  return TypeInfo{out};
-}
-
-template <typename T>
-inline TypeInfo KernelInfoImpl<T>::GetOutputTypeInfo(size_t index) const {
-  OrtTypeInfo* out = nullptr;
-  ThrowOnError(GetApi().KernelInfo_GetOutputTypeInfo(this->p_, index, &out));
-  return TypeInfo{out};
-}
-
-template <typename T>
-inline Value KernelInfoImpl<T>::GetTensorAttribute(const char* name, OrtAllocator* allocator) const {
-  OrtValue* out = nullptr;
-  ThrowOnError(GetApi().KernelInfoGetAttribute_tensor(this->p_, name, allocator, &out));
-  return Value{out};
-}
-
-template <typename T>
-inline ConstValue KernelInfoImpl<T>::GetTensorConstantInput(size_t index, int* is_constant) const {
-  const OrtValue* out = nullptr;
-  ThrowOnError(GetApi().KernelInfoGetConstantInput_tensor(this->p_, index, is_constant, &out));
-  return ConstValue{out};
-}
-
-template <typename T>
-inline std::string KernelInfoImpl<T>::GetNodeName() const {
-  size_t size = 0;
-
-  // Feed nullptr for the data buffer to query the true size of the string value
-  Ort::ThrowOnError(GetApi().KernelInfo_GetNodeName(this->p_, nullptr, &size));
-
-  std::string out;
-  out.resize(size);
-  Ort::ThrowOnError(GetApi().KernelInfo_GetNodeName(this->p_, &out[0], &size));
-  out.resize(size - 1);  // remove the terminating character '\0'
-
-  return out;
-}
-
-template <typename T>
-inline Logger KernelInfoImpl<T>::GetLogger() const {
-  const OrtLogger* out = nullptr;
-  ThrowOnError(GetApi().KernelInfo_GetLogger(this->p_, &out));
-  return Logger{out};
-}
-
-inline void attr_utils::GetAttr(const OrtKernelInfo* p, const char* name, float& out) {
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttribute_float(p, name, &out));
-}
-
-inline void attr_utils::GetAttr(const OrtKernelInfo* p, const char* name, int64_t& out) {
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttribute_int64(p, name, &out));
-}
-
-inline void attr_utils::GetAttr(const OrtKernelInfo* p, const char* name, std::string& result) {
-  size_t size = 0;
-  // Feed nullptr for the data buffer to query the true size of the string attribute
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttribute_string(p, name, nullptr, &size));
-
-  std::string out;
-  out.resize(size);
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttribute_string(p, name, &out[0], &size));
-  out.resize(size - 1);  // remove the terminating character '\0'
-  out.swap(result);
-}
-
-inline void attr_utils::GetAttrs(const OrtKernelInfo* p, const char* name, std::vector<float>& result) {
-  size_t size = 0;
-  // Feed nullptr for the data buffer to query the true size of the attribute
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttributeArray_float(p, name, nullptr, &size));
-
-  std::vector<float> out;
-  out.resize(size);
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttributeArray_float(p, name, out.data(), &size));
-  out.swap(result);
-}
-
-inline void attr_utils::GetAttrs(const OrtKernelInfo* p, const char* name, std::vector<int64_t>& result) {
-  size_t size = 0;
-
-  // Feed nullptr for the data buffer to query the true size of the attribute
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttributeArray_int64(p, name, nullptr, &size));
-
-  std::vector<int64_t> out;
-  out.resize(size);
-  Ort::ThrowOnError(GetApi().KernelInfoGetAttributeArray_int64(p, name, out.data(), &size));
-  out.swap(result);
-}
-}  // namespace detail
-
-inline KernelInfo::KernelInfo(OrtKernelInfo* info) : detail::KernelInfoImpl<OrtKernelInfo>{info} {}
-
-inline Op::Op(OrtOp* p) : Base<OrtOp>(p) {}
-
-inline Op Op::Create(const OrtKernelInfo* info, const char* op_name, const char* domain, int version,
-                     const char** type_constraint_names,
-                     const ONNXTensorElementDataType* type_constraint_values,
-                     size_t type_constraint_count,
-                     const OpAttr* attr_values, size_t attr_count,
-                     size_t input_count, size_t output_count) {
-  static_assert(sizeof(OpAttr) == sizeof(OrtOpAttr*),
-                "OpAttr's is expected to be just an array of OrtOpAttr in memory so we can reinterpret safely");
-  auto attr_input_values = reinterpret_cast<const OrtOpAttr* const*>(attr_values);
-  OrtOp* op;
-  Ort::ThrowOnError(GetApi().CreateOp(info, op_name, domain, version, type_constraint_names, type_constraint_values,
-                                      static_cast<int>(type_constraint_count),
-                                      attr_input_values,
-                                      static_cast<int>(attr_count),
-                                      static_cast<int>(input_count),
-                                      static_cast<int>(output_count), &op));
-  return Op{op};
-}
-
-inline void Op::Invoke(const OrtKernelContext* context,
-                       const Value* input_values,
-                       size_t input_count,
-                       Value* output_values,
-                       size_t output_count) {
-  static_assert(sizeof(Value) == sizeof(OrtValue*),
-                "Value is really just an array of OrtValue* in memory, so we can reinterpret_cast safely");
-  auto ort_input_values = reinterpret_cast<const OrtValue* const*>(input_values);
-  auto ort_output_values = reinterpret_cast<OrtValue**>(output_values);
-  Ort::ThrowOnError(GetApi().InvokeOp(context, p_, ort_input_values, static_cast<int>(input_count),
-                                      ort_output_values, static_cast<int>(output_count)));
-}
-
-inline void Op::Invoke(const OrtKernelContext* context,
-                       const OrtValue* const* input_values,
-                       size_t input_count,
-                       OrtValue* const* output_values,
-                       size_t output_count) {
-  Ort::ThrowOnError(GetApi().InvokeOp(context, p_, input_values, static_cast<int>(input_count),
-                                      output_values, static_cast<int>(output_count)));
-}
-
-inline std::string GetVersionString() {
-  return OrtGetApiBase()->GetVersionString();
-}
-
-inline std::string GetBuildInfoString() {
-  return GetApi().GetBuildInfoString();
-}
-
-inline std::vector<std::string> GetAvailableProviders() {
-  char** providers;
-  int len;
-
-  auto release_fn = [&len](char** providers) {
-    // This should always return nullptr.
-    ThrowOnError(GetApi().ReleaseAvailableProviders(providers, len));
-  };
-
-  ThrowOnError(GetApi().GetAvailableProviders(&providers, &len));
-  std::unique_ptr<char*, decltype(release_fn)> guard(providers, release_fn);
-  std::vector<std::string> available_providers;
-  available_providers.reserve(static_cast<size_t>(len));
-  for (int i = 0; i < len; ++i) {
-    available_providers.emplace_back(providers[i]);
-  }
-  return available_providers;
-}
-
-template <typename TOp, typename TKernel, bool WithStatus>
-void CustomOpBase<TOp, TKernel, WithStatus>::GetSessionConfigs(std::unordered_map<std::string, std::string>& out,
-                                                               ConstSessionOptions options) const {
-  const TOp* derived = static_cast<const TOp*>(this);
-  std::vector<std::string> keys = derived->GetSessionConfigKeys();
-
-  out.reserve(keys.size());
-
-  std::string config_entry_key = detail::MakeCustomOpConfigEntryKey(derived->GetName(), "");
-  const size_t prefix_size = config_entry_key.length();
-
-  for (const auto& key : keys) {
-    config_entry_key.resize(prefix_size);
-    config_entry_key.append(key);
-    out[key] = options.GetConfigEntryOrDefault(config_entry_key.c_str(), "");
-  }
-}
-
-inline ShapeInferContext::ShapeInferContext(const OrtApi* ort_api,
-                                            OrtShapeInferContext* ctx) : ort_api_(ort_api), ctx_(ctx) {
-  size_t input_count = 0;
-  Ort::ThrowOnError(ort_api_->ShapeInferContext_GetInputCount(ctx_, &input_count));
-  for (size_t ith_input = 0; ith_input < input_count; ++ith_input) {
-    OrtTensorTypeAndShapeInfo* info{};
-    Ort::ThrowOnError(ort_api_->ShapeInferContext_GetInputTypeShape(ctx, ith_input, &info));
-    TensorTypeAndShapeInfo type_shape_info(info);
-    auto integer_shape = type_shape_info.GetShape();
-    std::vector<const char*> symbolic_shape(integer_shape.size(), {});
-    type_shape_info.GetSymbolicDimensions(&symbolic_shape[0], integer_shape.size());
-    Shape shape;
-    for (size_t ith = 0; ith < integer_shape.size(); ++ith) {
-      if (symbolic_shape[ith] && std::string{symbolic_shape[ith]}.size() > 0) {
-        shape.emplace_back(symbolic_shape[ith]);
-      } else {
-        shape.emplace_back(integer_shape[ith]);
-      }
-    }
-    input_shapes_.push_back(std::move(shape));
-    type_shape_info.release();
-  }
-}
-
-inline Status ShapeInferContext::SetOutputShape(size_t indice, const Shape& shape) {
-  OrtTensorTypeAndShapeInfo* info = {};
-  RETURN_ON_API_FAIL(ort_api_->CreateTensorTypeAndShapeInfo(&info));
-
-  using InfoPtr = std::unique_ptr<OrtTensorTypeAndShapeInfo, std::function<void(OrtTensorTypeAndShapeInfo*)>>;
-
-  InfoPtr info_ptr(info, [this](OrtTensorTypeAndShapeInfo* obj) {
-    ort_api_->ReleaseTensorTypeAndShapeInfo(obj);
-  });
-
-  std::vector<int64_t> integer_dims;
-  std::vector<const char*> symbolic_dims;
-
-  for (const auto dim : shape) {
-    if (dim.IsInt()) {
-      integer_dims.push_back(dim.IsInt());
-      symbolic_dims.push_back("");
-    } else {
-      if (!dim.AsSym() || std::string{dim.AsSym()}.empty()) {
-        ORT_CXX_API_THROW("Symbolic dim must not be an empty string", ORT_INVALID_ARGUMENT);
-      }
-      integer_dims.push_back(SymbolicInteger::INVALID_INT_DIM);
-      symbolic_dims.push_back(dim.AsSym());
-    }
-  }
-
-  RETURN_ON_API_FAIL(ort_api_->SetDimensions(info, integer_dims.data(), integer_dims.size()));
-  RETURN_ON_API_FAIL(ort_api_->SetSymbolicDimensions(info, symbolic_dims.data(), symbolic_dims.size()));
-  RETURN_ON_API_FAIL(ort_api_->ShapeInferContext_SetOutputTypeShape(ctx_, indice, info));
-  return Status{nullptr};
-}
-
-inline int64_t ShapeInferContext::GetAttrInt(const char* attr_name) {
-  const auto* attr = GetAttrHdl(attr_name);
-  int64_t i = {};
-  size_t out = {};
-  Ort::ThrowOnError(ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_INT, &i, sizeof(i), &out));
-  return i;
-}
-
-inline ShapeInferContext::Ints ShapeInferContext::GetAttrInts(const char* attr_name) {
-  const auto* attr = GetAttrHdl(attr_name);
-  int64_t i = {};
-  size_t out = {};
-  // first call to get the bytes needed
-  auto status = ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_INTS, &i, sizeof(i), &out);
-  if (status) {
-    size_t num_i = out / sizeof(int64_t);
-    ShapeInferContext::Ints ints(num_i, 0);
-    Ort::ThrowOnError(ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_INTS, ints.data(), out, &out));
-    return ints;
-  } else {
-    return {i};
-  }
-}
-
-inline float ShapeInferContext::GetAttrFloat(const char* attr_name) {
-  const auto* attr = GetAttrHdl(attr_name);
-  float f = {};
-  size_t out = {};
-  Ort::ThrowOnError(ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_FLOAT, &f, sizeof(f), &out));
-  return f;
-}
-
-inline ShapeInferContext::Floats ShapeInferContext::GetAttrFloats(const char* attr_name) {
-  const auto* attr = GetAttrHdl(attr_name);
-  float f = {};
-  size_t out = {};
-  // first call to get the bytes needed
-  auto status = ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_FLOATS, &f, sizeof(f), &out);
-  if (status) {
-    size_t num_f = out / sizeof(float);
-    ShapeInferContext::Floats floats(num_f, 0);
-    Ort::ThrowOnError(ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_FLOATS, floats.data(), out, &out));
-    return floats;
-  } else {
-    return {f};
-  }
-}
-
-inline std::string ShapeInferContext::GetAttrString(const char* attr_name) {
-  const auto* attr = GetAttrHdl(attr_name);
-  char c = {};
-  size_t out = {};
-  // first call to get the bytes needed
-  auto status = ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_STRING, &c, sizeof(char), &out);
-  if (status) {
-    std::vector<char> chars(out, '\0');
-    Ort::ThrowOnError(ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_STRING, chars.data(), out, &out));
-    return {chars.data()};
-  } else {
-    return {c};
-  }
-}
-
-inline ShapeInferContext::Strings ShapeInferContext::GetAttrStrings(const char* attr_name) {
-  const auto* attr = GetAttrHdl(attr_name);
-  char c = {};
-  size_t out = {};
-  // first call to get the bytes needed
-  auto status = ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_STRINGS, &c, sizeof(char), &out);
-  if (status) {
-    std::vector<char> chars(out, '\0');
-    Ort::ThrowOnError(ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_STRINGS, chars.data(), out, &out));
-    ShapeInferContext::Strings strings;
-    char* char_st = chars.data();
-    char* char_ed = char_st + out;
-    while (char_st < char_ed) {
-      strings.emplace_back(char_st);
-      while (*char_st != '\0') {
-        char_st++;
-      }
-      char_st++;
-    }
-    return strings;
-  } else {
-    return {std::string{c}};
-  }
-}
-
-inline const OrtOpAttr* ShapeInferContext::GetAttrHdl(const char* attr_name) const {
-  const OrtOpAttr* attr_hdl = {};
-  Ort::ThrowOnError(ort_api_->ShapeInferContext_GetAttribute(ctx_, attr_name, &attr_hdl));
-  return attr_hdl;
-}
-
-}  // namespace Ort
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_float16.h b/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_float16.h
deleted file mode 100644
index 0b066a9c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_float16.h
+++ /dev/null
@@ -1,540 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <stdint.h>
-#include <cmath>
-#include <cstring>
-#include <limits>
-
-namespace onnxruntime_float16 {
-
-namespace detail {
-
-enum class endian {
-#if defined(_WIN32)
-  little = 0,
-  big = 1,
-  native = little,
-#elif defined(__GNUC__) || defined(__clang__)
-  little = __ORDER_LITTLE_ENDIAN__,
-  big = __ORDER_BIG_ENDIAN__,
-  native = __BYTE_ORDER__,
-#else
-#error onnxruntime_float16::detail::endian is not implemented in this environment.
-#endif
-};
-
-static_assert(
-    endian::native == endian::little || endian::native == endian::big,
-    "Only little-endian or big-endian native byte orders are supported.");
-
-}  // namespace detail
-
-/// <summary>
-/// Shared implementation between public and internal classes. CRTP pattern.
-/// </summary>
-template <class Derived>
-struct Float16Impl {
- protected:
-  /// <summary>
-  /// Converts from float to uint16_t float16 representation
-  /// </summary>
-  /// <param name="v"></param>
-  /// <returns></returns>
-  constexpr static uint16_t ToUint16Impl(float v) noexcept;
-
-  /// <summary>
-  /// Converts float16 to float
-  /// </summary>
-  /// <returns>float representation of float16 value</returns>
-  float ToFloatImpl() const noexcept;
-
-  /// <summary>
-  /// Creates an instance that represents absolute value.
-  /// </summary>
-  /// <returns>Absolute value</returns>
-  uint16_t AbsImpl() const noexcept {
-    return static_cast<uint16_t>(val & ~kSignMask);
-  }
-
-  /// <summary>
-  /// Creates a new instance with the sign flipped.
-  /// </summary>
-  /// <returns>Flipped sign instance</returns>
-  uint16_t NegateImpl() const noexcept {
-    return IsNaN() ? val : static_cast<uint16_t>(val ^ kSignMask);
-  }
-
- public:
-  // uint16_t special values
-  static constexpr uint16_t kSignMask = 0x8000U;
-  static constexpr uint16_t kBiasedExponentMask = 0x7C00U;
-  static constexpr uint16_t kPositiveInfinityBits = 0x7C00U;
-  static constexpr uint16_t kNegativeInfinityBits = 0xFC00U;
-  static constexpr uint16_t kPositiveQNaNBits = 0x7E00U;
-  static constexpr uint16_t kNegativeQNaNBits = 0xFE00U;
-  static constexpr uint16_t kEpsilonBits = 0x4170U;
-  static constexpr uint16_t kMinValueBits = 0xFBFFU;  // Minimum normal number
-  static constexpr uint16_t kMaxValueBits = 0x7BFFU;  // Largest normal number
-  static constexpr uint16_t kOneBits = 0x3C00U;
-  static constexpr uint16_t kMinusOneBits = 0xBC00U;
-
-  uint16_t val{0};
-
-  Float16Impl() = default;
-
-  /// <summary>
-  /// Checks if the value is negative
-  /// </summary>
-  /// <returns>true if negative</returns>
-  bool IsNegative() const noexcept {
-    return static_cast<int16_t>(val) < 0;
-  }
-
-  /// <summary>
-  /// Tests if the value is NaN
-  /// </summary>
-  /// <returns>true if NaN</returns>
-  bool IsNaN() const noexcept {
-    return AbsImpl() > kPositiveInfinityBits;
-  }
-
-  /// <summary>
-  /// Tests if the value is finite
-  /// </summary>
-  /// <returns>true if finite</returns>
-  bool IsFinite() const noexcept {
-    return AbsImpl() < kPositiveInfinityBits;
-  }
-
-  /// <summary>
-  /// Tests if the value represents positive infinity.
-  /// </summary>
-  /// <returns>true if positive infinity</returns>
-  bool IsPositiveInfinity() const noexcept {
-    return val == kPositiveInfinityBits;
-  }
-
-  /// <summary>
-  /// Tests if the value represents negative infinity
-  /// </summary>
-  /// <returns>true if negative infinity</returns>
-  bool IsNegativeInfinity() const noexcept {
-    return val == kNegativeInfinityBits;
-  }
-
-  /// <summary>
-  /// Tests if the value is either positive or negative infinity.
-  /// </summary>
-  /// <returns>True if absolute value is infinity</returns>
-  bool IsInfinity() const noexcept {
-    return AbsImpl() == kPositiveInfinityBits;
-  }
-
-  /// <summary>
-  /// Tests if the value is NaN or zero. Useful for comparisons.
-  /// </summary>
-  /// <returns>True if NaN or zero.</returns>
-  bool IsNaNOrZero() const noexcept {
-    auto abs = AbsImpl();
-    return (abs == 0 || abs > kPositiveInfinityBits);
-  }
-
-  /// <summary>
-  /// Tests if the value is normal (not zero, subnormal, infinite, or NaN).
-  /// </summary>
-  /// <returns>True if so</returns>
-  bool IsNormal() const noexcept {
-    auto abs = AbsImpl();
-    return (abs < kPositiveInfinityBits)           // is finite
-           && (abs != 0)                           // is not zero
-           && ((abs & kBiasedExponentMask) != 0);  // is not subnormal (has a non-zero exponent)
-  }
-
-  /// <summary>
-  /// Tests if the value is subnormal (denormal).
-  /// </summary>
-  /// <returns>True if so</returns>
-  bool IsSubnormal() const noexcept {
-    auto abs = AbsImpl();
-    return (abs < kPositiveInfinityBits)           // is finite
-           && (abs != 0)                           // is not zero
-           && ((abs & kBiasedExponentMask) == 0);  // is subnormal (has a zero exponent)
-  }
-
-  /// <summary>
-  /// Creates an instance that represents absolute value.
-  /// </summary>
-  /// <returns>Absolute value</returns>
-  Derived Abs() const noexcept { return Derived::FromBits(AbsImpl()); }
-
-  /// <summary>
-  /// Creates a new instance with the sign flipped.
-  /// </summary>
-  /// <returns>Flipped sign instance</returns>
-  Derived Negate() const noexcept { return Derived::FromBits(NegateImpl()); }
-
-  /// <summary>
-  /// IEEE defines that positive and negative zero are equal, this gives us a quick equality check
-  /// for two values by or'ing the private bits together and stripping the sign. They are both zero,
-  /// and therefore equivalent, if the resulting value is still zero.
-  /// </summary>
-  /// <param name="lhs">first value</param>
-  /// <param name="rhs">second value</param>
-  /// <returns>True if both arguments represent zero</returns>
-  static bool AreZero(const Float16Impl& lhs, const Float16Impl& rhs) noexcept {
-    return static_cast<uint16_t>((lhs.val | rhs.val) & ~kSignMask) == 0;
-  }
-
-  bool operator==(const Float16Impl& rhs) const noexcept {
-    if (IsNaN() || rhs.IsNaN()) {
-      // IEEE defines that NaN is not equal to anything, including itself.
-      return false;
-    }
-    return val == rhs.val;
-  }
-
-  bool operator!=(const Float16Impl& rhs) const noexcept { return !(*this == rhs); }
-
-  bool operator<(const Float16Impl& rhs) const noexcept {
-    if (IsNaN() || rhs.IsNaN()) {
-      // IEEE defines that NaN is unordered with respect to everything, including itself.
-      return false;
-    }
-
-    const bool left_is_negative = IsNegative();
-    if (left_is_negative != rhs.IsNegative()) {
-      // When the signs of left and right differ, we know that left is less than right if it is
-      // the negative value. The exception to this is if both values are zero, in which case IEEE
-      // says they should be equal, even if the signs differ.
-      return left_is_negative && !AreZero(*this, rhs);
-    }
-    return (val != rhs.val) && ((val < rhs.val) ^ left_is_negative);
-  }
-};
-
-// The following Float16_t conversions are based on the code from
-// Eigen library.
-
-// The conversion routines are Copyright (c) Fabian Giesen, 2016.
-// The original license follows:
-//
-// Copyright (c) Fabian Giesen, 2016
-// All rights reserved.
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted.
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-namespace detail {
-union float32_bits {
-  unsigned int u;
-  float f;
-};
-}  // namespace detail
-
-template <class Derived>
-inline constexpr uint16_t Float16Impl<Derived>::ToUint16Impl(float v) noexcept {
-  detail::float32_bits f{};
-  f.f = v;
-
-  constexpr detail::float32_bits f32infty = {255 << 23};
-  constexpr detail::float32_bits f16max = {(127 + 16) << 23};
-  constexpr detail::float32_bits denorm_magic = {((127 - 15) + (23 - 10) + 1) << 23};
-  constexpr unsigned int sign_mask = 0x80000000u;
-  uint16_t val = static_cast<uint16_t>(0x0u);
-
-  unsigned int sign = f.u & sign_mask;
-  f.u ^= sign;
-
-  // NOTE all the integer compares in this function can be safely
-  // compiled into signed compares since all operands are below
-  // 0x80000000. Important if you want fast straight SSE2 code
-  // (since there's no unsigned PCMPGTD).
-
-  if (f.u >= f16max.u) {                         // result is Inf or NaN (all exponent bits set)
-    val = (f.u > f32infty.u) ? 0x7e00 : 0x7c00;  // NaN->qNaN and Inf->Inf
-  } else {                                       // (De)normalized number or zero
-    if (f.u < (113 << 23)) {                     // resulting FP16 is subnormal or zero
-      // use a magic value to align our 10 mantissa bits at the bottom of
-      // the float. as long as FP addition is round-to-nearest-even this
-      // just works.
-      f.f += denorm_magic.f;
-
-      // and one integer subtract of the bias later, we have our final float!
-      val = static_cast<uint16_t>(f.u - denorm_magic.u);
-    } else {
-      unsigned int mant_odd = (f.u >> 13) & 1;  // resulting mantissa is odd
-
-      // update exponent, rounding bias part 1
-      // Equivalent to `f.u += ((unsigned int)(15 - 127) << 23) + 0xfff`, but
-      // without arithmetic overflow.
-      f.u += 0xc8000fffU;
-      // rounding bias part 2
-      f.u += mant_odd;
-      // take the bits!
-      val = static_cast<uint16_t>(f.u >> 13);
-    }
-  }
-
-  val |= static_cast<uint16_t>(sign >> 16);
-  return val;
-}
-
-template <class Derived>
-inline float Float16Impl<Derived>::ToFloatImpl() const noexcept {
-  constexpr detail::float32_bits magic = {113 << 23};
-  constexpr unsigned int shifted_exp = 0x7c00 << 13;  // exponent mask after shift
-  detail::float32_bits o{};
-
-  o.u = (val & 0x7fff) << 13;            // exponent/mantissa bits
-  unsigned int exp = shifted_exp & o.u;  // just the exponent
-  o.u += (127 - 15) << 23;               // exponent adjust
-
-  // handle exponent special cases
-  if (exp == shifted_exp) {   // Inf/NaN?
-    o.u += (128 - 16) << 23;  // extra exp adjust
-  } else if (exp == 0) {      // Zero/Denormal?
-    o.u += 1 << 23;           // extra exp adjust
-    o.f -= magic.f;           // re-normalize
-  }
-
-  // Attempt to workaround the Internal Compiler Error on ARM64
-  // for bitwise | operator, including std::bitset
-#if (defined _MSC_VER) && (defined _M_ARM || defined _M_ARM64 || defined _M_ARM64EC)
-  if (IsNegative()) {
-    return -o.f;
-  }
-#else
-  // original code:
-  o.u |= (val & 0x8000U) << 16U;  // sign bit
-#endif
-  return o.f;
-}
-
-/// Shared implementation between public and internal classes. CRTP pattern.
-template <class Derived>
-struct BFloat16Impl {
- protected:
-  /// <summary>
-  /// Converts from float to uint16_t float16 representation
-  /// </summary>
-  /// <param name="v"></param>
-  /// <returns></returns>
-  static uint16_t ToUint16Impl(float v) noexcept;
-
-  /// <summary>
-  /// Converts bfloat16 to float
-  /// </summary>
-  /// <returns>float representation of bfloat16 value</returns>
-  float ToFloatImpl() const noexcept;
-
-  /// <summary>
-  /// Creates an instance that represents absolute value.
-  /// </summary>
-  /// <returns>Absolute value</returns>
-  uint16_t AbsImpl() const noexcept {
-    return static_cast<uint16_t>(val & ~kSignMask);
-  }
-
-  /// <summary>
-  /// Creates a new instance with the sign flipped.
-  /// </summary>
-  /// <returns>Flipped sign instance</returns>
-  uint16_t NegateImpl() const noexcept {
-    return IsNaN() ? val : static_cast<uint16_t>(val ^ kSignMask);
-  }
-
- public:
-  // uint16_t special values
-  static constexpr uint16_t kSignMask = 0x8000U;
-  static constexpr uint16_t kBiasedExponentMask = 0x7F80U;
-  static constexpr uint16_t kPositiveInfinityBits = 0x7F80U;
-  static constexpr uint16_t kNegativeInfinityBits = 0xFF80U;
-  static constexpr uint16_t kPositiveQNaNBits = 0x7FC1U;
-  static constexpr uint16_t kNegativeQNaNBits = 0xFFC1U;
-  static constexpr uint16_t kSignaling_NaNBits = 0x7F80U;
-  static constexpr uint16_t kEpsilonBits = 0x0080U;
-  static constexpr uint16_t kMinValueBits = 0xFF7FU;
-  static constexpr uint16_t kMaxValueBits = 0x7F7FU;
-  static constexpr uint16_t kRoundToNearest = 0x7FFFU;
-  static constexpr uint16_t kOneBits = 0x3F80U;
-  static constexpr uint16_t kMinusOneBits = 0xBF80U;
-
-  uint16_t val{0};
-
-  BFloat16Impl() = default;
-
-  /// <summary>
-  /// Checks if the value is negative
-  /// </summary>
-  /// <returns>true if negative</returns>
-  bool IsNegative() const noexcept {
-    return static_cast<int16_t>(val) < 0;
-  }
-
-  /// <summary>
-  /// Tests if the value is NaN
-  /// </summary>
-  /// <returns>true if NaN</returns>
-  bool IsNaN() const noexcept {
-    return AbsImpl() > kPositiveInfinityBits;
-  }
-
-  /// <summary>
-  /// Tests if the value is finite
-  /// </summary>
-  /// <returns>true if finite</returns>
-  bool IsFinite() const noexcept {
-    return AbsImpl() < kPositiveInfinityBits;
-  }
-
-  /// <summary>
-  /// Tests if the value represents positive infinity.
-  /// </summary>
-  /// <returns>true if positive infinity</returns>
-  bool IsPositiveInfinity() const noexcept {
-    return val == kPositiveInfinityBits;
-  }
-
-  /// <summary>
-  /// Tests if the value represents negative infinity
-  /// </summary>
-  /// <returns>true if negative infinity</returns>
-  bool IsNegativeInfinity() const noexcept {
-    return val == kNegativeInfinityBits;
-  }
-
-  /// <summary>
-  /// Tests if the value is either positive or negative infinity.
-  /// </summary>
-  /// <returns>True if absolute value is infinity</returns>
-  bool IsInfinity() const noexcept {
-    return AbsImpl() == kPositiveInfinityBits;
-  }
-
-  /// <summary>
-  /// Tests if the value is NaN or zero. Useful for comparisons.
-  /// </summary>
-  /// <returns>True if NaN or zero.</returns>
-  bool IsNaNOrZero() const noexcept {
-    auto abs = AbsImpl();
-    return (abs == 0 || abs > kPositiveInfinityBits);
-  }
-
-  /// <summary>
-  /// Tests if the value is normal (not zero, subnormal, infinite, or NaN).
-  /// </summary>
-  /// <returns>True if so</returns>
-  bool IsNormal() const noexcept {
-    auto abs = AbsImpl();
-    return (abs < kPositiveInfinityBits)           // is finite
-           && (abs != 0)                           // is not zero
-           && ((abs & kBiasedExponentMask) != 0);  // is not subnormal (has a non-zero exponent)
-  }
-
-  /// <summary>
-  /// Tests if the value is subnormal (denormal).
-  /// </summary>
-  /// <returns>True if so</returns>
-  bool IsSubnormal() const noexcept {
-    auto abs = AbsImpl();
-    return (abs < kPositiveInfinityBits)           // is finite
-           && (abs != 0)                           // is not zero
-           && ((abs & kBiasedExponentMask) == 0);  // is subnormal (has a zero exponent)
-  }
-
-  /// <summary>
-  /// Creates an instance that represents absolute value.
-  /// </summary>
-  /// <returns>Absolute value</returns>
-  Derived Abs() const noexcept { return Derived::FromBits(AbsImpl()); }
-
-  /// <summary>
-  /// Creates a new instance with the sign flipped.
-  /// </summary>
-  /// <returns>Flipped sign instance</returns>
-  Derived Negate() const noexcept { return Derived::FromBits(NegateImpl()); }
-
-  /// <summary>
-  /// IEEE defines that positive and negative zero are equal, this gives us a quick equality check
-  /// for two values by or'ing the private bits together and stripping the sign. They are both zero,
-  /// and therefore equivalent, if the resulting value is still zero.
-  /// </summary>
-  /// <param name="lhs">first value</param>
-  /// <param name="rhs">second value</param>
-  /// <returns>True if both arguments represent zero</returns>
-  static bool AreZero(const BFloat16Impl& lhs, const BFloat16Impl& rhs) noexcept {
-    // IEEE defines that positive and negative zero are equal, this gives us a quick equality check
-    // for two values by or'ing the private bits together and stripping the sign. They are both zero,
-    // and therefore equivalent, if the resulting value is still zero.
-    return static_cast<uint16_t>((lhs.val | rhs.val) & ~kSignMask) == 0;
-  }
-};
-
-template <class Derived>
-inline uint16_t BFloat16Impl<Derived>::ToUint16Impl(float v) noexcept {
-  uint16_t result;
-  if (std::isnan(v)) {
-    result = kPositiveQNaNBits;
-  } else {
-    auto get_msb_half = [](float fl) {
-      uint16_t result;
-#ifdef __cpp_if_constexpr
-      if constexpr (detail::endian::native == detail::endian::little) {
-#else
-      if (detail::endian::native == detail::endian::little) {
-#endif
-        std::memcpy(&result, reinterpret_cast<char*>(&fl) + sizeof(uint16_t), sizeof(uint16_t));
-      } else {
-        std::memcpy(&result, &fl, sizeof(uint16_t));
-      }
-      return result;
-    };
-
-    uint16_t upper_bits = get_msb_half(v);
-    union {
-      uint32_t U32;
-      float F32;
-    };
-    F32 = v;
-    U32 += (upper_bits & 1) + kRoundToNearest;
-    result = get_msb_half(F32);
-  }
-  return result;
-}
-
-template <class Derived>
-inline float BFloat16Impl<Derived>::ToFloatImpl() const noexcept {
-  if (IsNaN()) {
-    return std::numeric_limits<float>::quiet_NaN();
-  }
-  float result;
-  char* const first = reinterpret_cast<char*>(&result);
-  char* const second = first + sizeof(uint16_t);
-#ifdef __cpp_if_constexpr
-  if constexpr (detail::endian::native == detail::endian::little) {
-#else
-  if (detail::endian::native == detail::endian::little) {
-#endif
-    std::memset(first, 0, sizeof(uint16_t));
-    std::memcpy(second, &val, sizeof(uint16_t));
-  } else {
-    std::memcpy(first, &val, sizeof(uint16_t));
-    std::memset(second, 0, sizeof(uint16_t));
-  }
-  return result;
-}
-
-}  // namespace onnxruntime_float16
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_run_options_config_keys.h b/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_run_options_config_keys.h
deleted file mode 100644
index 1f5fcd50..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_run_options_config_keys.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-/*
- * This file defines RunOptions Config Keys and format of the Config Values.
- *
- * The Naming Convention for a RunOptions Config Key,
- * "[Area][.[SubArea1].[SubArea2]...].[Keyname]"
- * Such as "ep.cuda.use_arena"
- * The Config Key cannot be empty
- * The maximum length of the Config Key is 128
- *
- * The string format of a RunOptions Config Value is defined individually for each Config.
- * The maximum length of the Config Value is 1024
- */
-
-// Key for enabling shrinkages of user listed device memory arenas.
-// Expects a list of semi-colon separated key value pairs separated by colon in the following format:
-// "device_0:device_id_0;device_1:device_id_1"
-// No white-spaces allowed in the provided list string.
-// Currently, the only supported devices are : "cpu", "gpu" (case sensitive).
-// If "cpu" is included in the list, DisableCpuMemArena() API must not be called (i.e.) arena for cpu should be enabled.
-// Example usage: "cpu:0;gpu:0" (or) "gpu:0"
-// By default, the value for this key is empty (i.e.) no memory arenas are shrunk
-static const char* const kOrtRunOptionsConfigEnableMemoryArenaShrinkage = "memory.enable_memory_arena_shrinkage";
-
-// Set to '1' to not synchronize execution providers with CPU at the end of session run.
-// Per default it will be set to '0'
-// Taking CUDA EP as an example, it omit triggering cudaStreamSynchronize on the compute stream.
-static const char* const kOrtRunOptionsConfigDisableSynchronizeExecutionProviders = "disable_synchronize_execution_providers";
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_session_options_config_keys.h b/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_session_options_config_keys.h
deleted file mode 100644
index b2824387..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/onnxruntime_session_options_config_keys.h
+++ /dev/null
@@ -1,258 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-/*
- * This file defines SessionOptions Config Keys and format of the Config Values.
- *
- * The Naming Convention for a SessionOptions Config Key,
- * "[Area][.[SubArea1].[SubArea2]...].[Keyname]"
- * Such as "ep.cuda.use_arena"
- * The Config Key cannot be empty
- * The maximum length of the Config Key is 128
- *
- * The string format of a SessionOptions Config Value is defined individually for each Config.
- * The maximum length of the Config Value is 1024
- */
-
-// Key for disable PrePacking,
-// If the config value is set to "1" then the prepacking is disabled, otherwise prepacking is enabled (default value)
-static const char* const kOrtSessionOptionsConfigDisablePrepacking = "session.disable_prepacking";
-
-// A value of "1" means allocators registered in the env will be used. "0" means the allocators created in the session
-// will be used. Use this to override the usage of env allocators on a per session level.
-static const char* const kOrtSessionOptionsConfigUseEnvAllocators = "session.use_env_allocators";
-
-// Set to 'ORT' (case sensitive) to load an ORT format model.
-// If unset, model type will default to ONNX unless inferred from filename ('.ort' == ORT format) or bytes to be ORT
-static const char* const kOrtSessionOptionsConfigLoadModelFormat = "session.load_model_format";
-
-// Set to 'ORT' (case sensitive) to save optimized model in ORT format when SessionOptions.optimized_model_path is set.
-// If unset, format will default to ONNX unless optimized_model_filepath ends in '.ort'.
-static const char* const kOrtSessionOptionsConfigSaveModelFormat = "session.save_model_format";
-
-// If a value is "1", flush-to-zero and denormal-as-zero are applied. The default is "0".
-// When multiple sessions are created, a main thread doesn't override changes from succeeding session options,
-// but threads in session thread pools follow option changes.
-// When ORT runs with OpenMP, the same rule is applied, i.e. the first session option to flush-to-zero and
-// denormal-as-zero is only applied to global OpenMP thread pool, which doesn't support per-session thread pool.
-// Note that an alternative way not using this option at runtime is to train and export a model without denormals
-// and that's recommended because turning this option on may hurt model accuracy.
-static const char* const kOrtSessionOptionsConfigSetDenormalAsZero = "session.set_denormal_as_zero";
-
-// It controls to run quantization model in QDQ (QuantizelinearDeQuantizelinear) format or not.
-// "0": enable. ORT does fusion logic for QDQ format.
-// "1": disable. ORT doesn't do fusion logic for QDQ format.
-// Its default value is "0" unless the DirectML execution provider is registered, in which case it defaults to "1".
-static const char* const kOrtSessionOptionsDisableQuantQDQ = "session.disable_quant_qdq";
-
-// It controls whether to enable Double QDQ remover and Identical Children Consolidation
-// "0": not to disable. ORT does remove the middle 2 Nodes from a Q->(QD->Q)->QD pairs
-// "1": disable. ORT doesn't remove the middle 2 Nodes from a Q->(QD->Q)->QD pairs
-// Its default value is "0"
-static const char* const kOrtSessionOptionsDisableDoubleQDQRemover = "session.disable_double_qdq_remover";
-
-// If set to "1", enables the removal of QuantizeLinear/DequantizeLinear node pairs once all QDQ handling has been
-// completed. e.g. If after all QDQ handling has completed and we have -> FloatOp -> Q -> DQ -> FloatOp -> the
-// Q -> DQ could potentially be removed. This will provide a performance benefit by avoiding going from float to
-// 8-bit and back to float, but could impact accuracy. The impact on accuracy will be model specific and depend on
-// other factors like whether the model was created using Quantization Aware Training or Post Training Quantization.
-// As such, it's best to test to determine if enabling this works well for your scenario.
-// The default value is "0"
-// Available since version 1.11.
-static const char* const kOrtSessionOptionsEnableQuantQDQCleanup = "session.enable_quant_qdq_cleanup";
-
-// Enable or disable gelu approximation in graph optimization. "0": disable; "1": enable. The default is "0".
-// GeluApproximation has side effects which may change the inference results. It is disabled by default due to this.
-static const char* const kOrtSessionOptionsEnableGeluApproximation = "optimization.enable_gelu_approximation";
-
-// This setting controls whether to enable AheadOfTime function inlining.
-// AOT function inlining examines the graph and attempts to inline as many locally defined functions in the model
-// as possible with the help of enabled execution providers.
-// This can reduce the number of function calls and improve performance because it is done before
-// Level1 optimizers and constant folding. However, under some circumstances, when the EPs are not available,
-// one can disable the AOT inlining, produce an optimized model and postpone AOT until run time.
-// "0": enable; "1": disable.
-// Its default value is "0".
-static const char* const kOrtSessionOptionsDisableAheadOfTimeFunctionInlining = "session.disable_aot_function_inlining";
-
-#ifdef ENABLE_TRAINING
-// Specifies a list of op types for memory footprint reduction.
-// The value should be a ","-delimited list of pair of
-// <subgraph string: optimization strategy: number of subgraph to apply>.
-// For example, "Gelu+Cast+:1:0,Dropout+:1:1".
-//   A valid "subgraph string" should be one subgraph representation output by ORT graph transformations.
-//   "optimization strategy" currently has valid values: 0 - disabled, 1 - recompute.
-//   "number of subgraph to apply" is used to control how many subgraphs to apply optimization, to avoid "oversaving"
-//   the memory.
-static const char* const kOrtSessionOptionsMemoryOptimizerEnabler = "optimization.memory_optimizer_config";
-
-// Specifies the config for detecting subgraphs for memory footprint reduction.
-// The value should be a string contains int separated using commas. The default value is "0:0".
-static const char* const kOrtSessionOptionsMemoryOptimizerProbeConfig = "optimization.enable_memory_probe_recompute_config";
-#endif
-
-// Enable or disable using device allocator for allocating initialized tensor memory. "1": enable; "0": disable. The default is "0".
-// Using device allocators means the memory allocation is made using malloc/new.
-static const char* const kOrtSessionOptionsUseDeviceAllocatorForInitializers = "session.use_device_allocator_for_initializers";
-
-// Configure whether to allow the inter_op/intra_op threads spinning a number of times before blocking
-// "0": thread will block if found no job to run
-// "1": default, thread will spin a number of times before blocking
-static const char* const kOrtSessionOptionsConfigAllowInterOpSpinning = "session.inter_op.allow_spinning";
-static const char* const kOrtSessionOptionsConfigAllowIntraOpSpinning = "session.intra_op.allow_spinning";
-
-// Key for using model bytes directly for ORT format
-// If a session is created using an input byte array contains the ORT format model data,
-// By default we will copy the model bytes at the time of session creation to ensure the model bytes
-// buffer is valid.
-// Setting this option to "1" will disable copy the model bytes, and use the model bytes directly. The caller
-// has to guarantee that the model bytes are valid until the ORT session using the model bytes is destroyed.
-static const char* const kOrtSessionOptionsConfigUseORTModelBytesDirectly = "session.use_ort_model_bytes_directly";
-
-/// <summary>
-/// Key for using the ORT format model flatbuffer bytes directly for initializers.
-/// This avoids copying the bytes and reduces peak memory usage during model loading and initialization.
-/// Requires `session.use_ort_model_bytes_directly` to be true.
-/// If set, the flatbuffer bytes provided when creating the InferenceSession MUST remain valid for the entire
-/// duration of the InferenceSession.
-/// </summary>
-static const char* const kOrtSessionOptionsConfigUseORTModelBytesForInitializers =
-    "session.use_ort_model_bytes_for_initializers";
-
-// This should only be specified when exporting an ORT format model for use on a different platform.
-// If the ORT format model will be used on ARM platforms set to "1". For other platforms set to "0"
-// Available since version 1.11.
-static const char* const kOrtSessionOptionsQDQIsInt8Allowed = "session.qdqisint8allowed";
-
-// x64 SSE4.1/AVX2/AVX512(with no VNNI) has overflow problem with quantizied matrix multiplication with U8S8.
-// To avoid this we need to use slower U8U8 matrix multiplication instead. This option, if
-// turned on, use slower U8U8 matrix multiplications. Only effective with AVX2 or AVX512
-// platforms.
-static const char* const kOrtSessionOptionsAvx2PrecisionMode = "session.x64quantprecision";
-
-// Specifies how minimal build graph optimizations are handled in a full build.
-// These optimizations are at the extended level or higher.
-// Possible values and their effects are:
-// "save": Save runtime optimizations when saving an ORT format model.
-// "apply": Only apply optimizations available in a minimal build.
-// ""/<unspecified>: Apply optimizations available in a full build.
-// Available since version 1.11.
-static const char* const kOrtSessionOptionsConfigMinimalBuildOptimizations =
-    "optimization.minimal_build_optimizations";
-
-// Note: The options specific to an EP should be specified prior to appending that EP to the session options object in
-// order for them to take effect.
-
-// Specifies a list of stop op types. Nodes of a type in the stop op types and nodes downstream from them will not be
-// run by the NNAPI EP.
-// The value should be a ","-delimited list of op types. For example, "Add,Sub".
-// If not specified, the default set of stop ops is used. To specify an empty stop ops types list and disable stop op
-// exclusion, set the value to "".
-static const char* const kOrtSessionOptionsConfigNnapiEpPartitioningStopOps = "ep.nnapi.partitioning_stop_ops";
-
-// Enabling dynamic block-sizing for multithreading.
-// With a positive value, thread pool will split a task of N iterations to blocks of size starting from:
-// N / (num_of_threads * dynamic_block_base)
-// As execution progresses, the size will decrease according to the diminishing residual of N,
-// meaning the task will be distributed in smaller granularity for better parallelism.
-// For some models, it helps to reduce the variance of E2E inference latency and boost performance.
-// The feature will not function by default, specify any positive integer, e.g. "4", to enable it.
-// Available since version 1.11.
-static const char* const kOrtSessionOptionsConfigDynamicBlockBase = "session.dynamic_block_base";
-
-// This option allows to decrease CPU usage between infrequent
-// requests and forces any TP threads spinning stop immediately when the last of
-// concurrent Run() call returns.
-// Spinning is restarted on the next Run() call.
-// Applies only to internal thread-pools
-static const char* const kOrtSessionOptionsConfigForceSpinningStop = "session.force_spinning_stop";
-
-// "1": all inconsistencies encountered during shape and type inference
-// will result in failures.
-// "0": in some cases warnings will be logged but processing will continue. The default.
-// May be useful to expose bugs in models.
-static const char* const kOrtSessionOptionsConfigStrictShapeTypeInference = "session.strict_shape_type_inference";
-
-// "1": every model using a more recent opset than the latest released one will fail
-// "0": the model may or may not work if onnxruntime cannot find an implementation, this option
-// is used for development purpose.
-static const char* const kOrtSessionOptionsConfigStrictAllowReleasedOpsetsOnly = "session.allow_released_opsets_only";
-
-// The file saves configuration for partitioning node among logic streams
-static const char* const kNodePartitionConfigFile = "session.node_partition_config_file";
-
-// This Option allows setting affinities for intra op threads.
-// Affinity string follows format:
-// logical_processor_id,logical_processor_id;logical_processor_id,logical_processor_id
-// Semicolon isolates configurations among threads, while comma split processors where ith thread expected to attach to.
-// e.g.1,2,3;4,5
-// specifies affinities for two threads, with the 1st thread attach to the 1st, 2nd, and 3rd processor, and 2nd thread to the 4th and 5th.
-// To ease the configuration, an "interval" is also allowed:
-// e.g. 1-8;8-16;17-24
-// orders that the 1st thread runs on first eight processors, 2nd thread runs on next eight processors, and so forth.
-// Note:
-// 1. Once set, the number of thread affinities must equal to intra_op_num_threads - 1, since ort does not set affinity on the main thread which
-//    is started and managed by the calling app;
-// 2. For windows, ort will infer the group id from a logical processor id, for example, assuming there are two groups with each has 64 logical processors,
-//    an id of 64 will be inferred as the last processor of the 1st group, while 65 will be interpreted as the 1st processor of the second group.
-//    Hence 64-65 is an invalid configuration, because a windows thread cannot be attached to processors across group boundary.
-static const char* const kOrtSessionOptionsConfigIntraOpThreadAffinities = "session.intra_op_thread_affinities";
-
-// This option will dump out the model to assist debugging any issues with layout transformation,
-// and is primarily intended for developer usage. It is only relevant if an execution provider that requests
-// NHWC layout is enabled such as NNAPI, XNNPACK or QNN.
-//
-// Default is off. Set to "1" to enable.
-//
-// If modified by layout transformation the model will be dumped after these steps:
-//   1) insertion of the layout transformation Transpose nodes
-//   2) after those are optimized using the transpose optimizer,
-//   3) after the L1 transformers are applied to the updated graph.
-// The model will be saved to filename post_layout_transform_step_<step_number>.onnx.
-static const char* const kDebugLayoutTransformation = "session.debug_layout_transformation";
-
-// Graph nodes that are not supported by the execution providers (EPs) explicitly added to the session are
-// assigned (i.e., "fallback") to the CPU EP by default.
-//
-// This option allows the user to disable the fallback of unsupported graph nodes to the CPU EP.
-// If this option is set to "1", session creation will fail if the execution providers other than the CPU EP cannot
-// fully support all of the nodes in the graph.
-//
-// It is invalid to set this option and explicitly add the CPU EP to the session. In this case, session creation
-// will also fail with an error.
-//
-// Option values:
-// - "0": CPU EP fallback is not disabled. [DEFAULT]
-// - "1": CPU EP fallback is disabled.
-static const char* const kOrtSessionOptionsDisableCPUEPFallback = "session.disable_cpu_ep_fallback";
-
-// Use this config when serializing a large model after optimization to specify an external initializers file
-static const char* const kOrtSessionOptionsOptimizedModelExternalInitializersFileName =
-    "session.optimized_model_external_initializers_file_name";
-
-// Use this config to control the minimum size of the initializer when externalizing it during serialization
-static const char* const kOrtSessionOptionsOptimizedModelExternalInitializersMinSizeInBytes =
-    "session.optimized_model_external_initializers_min_size_in_bytes";
-
-// Enable EP context feature to dump the partitioned graph which includes the EP context into Onnx file.
-// The dumped Onnx model with EP context can be used for future inference to avoid the EP graph partitioning/compile overhead.
-// "0": disable. (default)
-// "1": enable.
-static const char* const kOrtSessionOptionEpContextEnable = "ep.context_enable";
-
-// Specify the file path for the Onnx model which has EP context.
-// Default to original_file_name_ctx.onnx if not specified
-static const char* const kOrtSessionOptionEpContextFilePath = "ep.context_file_path";
-
-// Flag to specify whether to dump the EP context into the Onnx model.
-// "0": dump the EP context into separate file, keep the file name in the Onnx model.
-// "1": dump the EP context into the Onnx model. (default).
-static const char* const kOrtSessionOptionEpContextEmbedMode = "ep.context_embed_mode";
-
-// Gemm fastmath mode provides fp32 gemm acceleration with bfloat16 based matmul.
-// Option values:
-// - "0": Gemm FastMath mode is not enabled. [DEFAULT]
-// - "1": Gemm FastMath mode is enabled.
-static const char* const kOrtSessionOptionsMlasGemmFastMathArm64Bfloat16 = "mlas.enable_gemm_fastmath_arm64_bfloat16";
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/provider_options.h b/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/provider_options.h
deleted file mode 100644
index aab13e80..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/onnxruntime-src/provider_options.h
+++ /dev/null
@@ -1,18 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-namespace onnxruntime {
-
-// data types for execution provider options
-
-using ProviderOptions = std::unordered_map<std::string, std::string>;
-using ProviderOptionsVector = std::vector<ProviderOptions>;
-using ProviderOptionsMap = std::unordered_map<std::string, ProviderOptions>;
-
-}  // namespace onnxruntime
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/accumulator.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/accumulator.h
deleted file mode 100644
index 55d70c78..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/accumulator.h
+++ /dev/null
@@ -1,894 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes to accumulate arc weights. Useful for weight lookahead.
-
-#ifndef FST_ACCUMULATOR_H_
-#define FST_ACCUMULATOR_H_
-
-#include <algorithm>
-#include <functional>
-#include <unordered_map>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arcfilter.h>
-#include <fst/arcsort.h>
-#include <fst/dfs-visit.h>
-#include <fst/expanded-fst.h>
-#include <fst/replace.h>
-
-namespace fst {
-
-// This class accumulates arc weights using the semiring Plus().
-template <class A>
-class DefaultAccumulator {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  DefaultAccumulator() {}
-
-  DefaultAccumulator(const DefaultAccumulator &acc, bool safe = false) {}
-
-  void Init(const Fst<Arc> &fst, bool copy = false) {}
-
-  void SetState(StateId state) {}
-
-  Weight Sum(Weight w, Weight v) { return Plus(w, v); }
-
-  template <class ArcIter>
-  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) {
-    Adder<Weight> adder(w);  // maintains cumulative sum accurately
-    aiter->Seek(begin);
-    for (auto pos = begin; pos < end; aiter->Next(), ++pos)
-      adder.Add(aiter->Value().weight);
-    return adder.Sum();
-  }
-
-  constexpr bool Error() const { return false; }
-
- private:
-  DefaultAccumulator &operator=(const DefaultAccumulator &) = delete;
-};
-
-// This class accumulates arc weights using the log semiring Plus() assuming an
-// arc weight has a WeightConvert specialization to and from log64 weights.
-template <class A>
-class LogAccumulator {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  LogAccumulator() {}
-
-  LogAccumulator(const LogAccumulator &acc, bool safe = false) {}
-
-  void Init(const Fst<Arc> &fst, bool copy = false) {}
-
-  void SetState(StateId s) {}
-
-  Weight Sum(Weight w, Weight v) { return LogPlus(w, v); }
-
-  template <class ArcIter>
-  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) {
-    auto sum = w;
-    aiter->Seek(begin);
-    for (auto pos = begin; pos < end; aiter->Next(), ++pos) {
-      sum = LogPlus(sum, aiter->Value().weight);
-    }
-    return sum;
-  }
-
-  constexpr bool Error() const { return false; }
-
- private:
-  Weight LogPlus(Weight w, Weight v) {
-    const auto f1 = to_log_weight_(w).Value();
-    const auto f2 = to_log_weight_(v).Value();
-    if (f1 > f2) {
-      return to_weight_(Log64Weight(f2 - internal::LogPosExp(f1 - f2)));
-    } else {
-      return to_weight_(Log64Weight(f1 - internal::LogPosExp(f2 - f1)));
-    }
-  }
-
-  WeightConvert<Weight, Log64Weight> to_log_weight_;
-  WeightConvert<Log64Weight, Weight> to_weight_;
-
-  LogAccumulator &operator=(const LogAccumulator &) = delete;
-};
-
-// Interface for shareable data for fast log accumulator copies. Holds pointers
-// to data only, storage is provided by derived classes.
-class FastLogAccumulatorData {
- public:
-  FastLogAccumulatorData(int arc_limit, int arc_period)
-      : arc_limit_(arc_limit),
-        arc_period_(arc_period),
-        weights_ptr_(nullptr),
-        num_weights_(0),
-        weight_positions_ptr_(nullptr),
-        num_positions_(0) {}
-
-  virtual ~FastLogAccumulatorData() {}
-
-  // Cummulative weight per state for all states s.t. # of arcs > arc_limit_
-  // with arcs in order. The first element per state is Log64Weight::Zero().
-  const double *Weights() const { return weights_ptr_; }
-
-  int NumWeights() const { return num_weights_; }
-
-  // Maps from state to corresponding beginning weight position in weights_.
-  // osition -1 means no pre-computed weights for that state.
-  const int *WeightPositions() const { return weight_positions_ptr_; }
-
-  int NumPositions() const { return num_positions_; }
-
-  int ArcLimit() const { return arc_limit_; }
-
-  int ArcPeriod() const { return arc_period_; }
-
-  // Returns true if the data object is mutable and supports SetData().
-  virtual bool IsMutable() const = 0;
-
-  // Does not take ownership but may invalidate the contents of weights and
-  // weight_positions.
-  virtual void SetData(std::vector<double> *weights,
-                       std::vector<int> *weight_positions) = 0;
-
- protected:
-  void Init(int num_weights, const double *weights, int num_positions,
-            const int *weight_positions) {
-    weights_ptr_ = weights;
-    num_weights_ = num_weights;
-    weight_positions_ptr_ = weight_positions;
-    num_positions_ = num_positions;
-  }
-
- private:
-  const int arc_limit_;
-  const int arc_period_;
-  const double *weights_ptr_;
-  int num_weights_;
-  const int *weight_positions_ptr_;
-  int num_positions_;
-
-  FastLogAccumulatorData(const FastLogAccumulatorData &) = delete;
-  FastLogAccumulatorData &operator=(const FastLogAccumulatorData &) = delete;
-};
-
-// FastLogAccumulatorData with mutable storage; filled by
-// FastLogAccumulator::Init.
-class MutableFastLogAccumulatorData : public FastLogAccumulatorData {
- public:
-  MutableFastLogAccumulatorData(int arc_limit, int arc_period)
-      : FastLogAccumulatorData(arc_limit, arc_period) {}
-
-  bool IsMutable() const override { return true; }
-
-  void SetData(std::vector<double> *weights,
-               std::vector<int> *weight_positions) override {
-    weights_.swap(*weights);
-    weight_positions_.swap(*weight_positions);
-    Init(weights_.size(), weights_.data(), weight_positions_.size(),
-         weight_positions_.data());
-  }
-
- private:
-  std::vector<double> weights_;
-  std::vector<int> weight_positions_;
-
-  MutableFastLogAccumulatorData(const MutableFastLogAccumulatorData &) = delete;
-  MutableFastLogAccumulatorData &operator=(
-      const MutableFastLogAccumulatorData &) = delete;
-};
-
-// This class accumulates arc weights using the log semiring Plus() assuming an
-// arc weight has a WeightConvert specialization to and from log64 weights. The
-// member function Init(fst) has to be called to setup pre-computed weight
-// information.
-template <class A>
-class FastLogAccumulator {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit FastLogAccumulator(ssize_t arc_limit = 20, ssize_t arc_period = 10)
-      : to_log_weight_(),
-        to_weight_(),
-        arc_limit_(arc_limit),
-        arc_period_(arc_period),
-        data_(std::make_shared<MutableFastLogAccumulatorData>(arc_limit,
-                                                              arc_period)),
-        state_weights_(nullptr),
-        error_(false) {}
-
-  explicit FastLogAccumulator(std::shared_ptr<FastLogAccumulatorData> data)
-      : to_log_weight_(),
-        to_weight_(),
-        arc_limit_(data->ArcLimit()),
-        arc_period_(data->ArcPeriod()),
-        data_(data),
-        state_weights_(nullptr),
-        error_(false) {}
-
-  FastLogAccumulator(const FastLogAccumulator<Arc> &acc, bool safe = false)
-      : to_log_weight_(),
-        to_weight_(),
-        arc_limit_(acc.arc_limit_),
-        arc_period_(acc.arc_period_),
-        data_(acc.data_),
-        state_weights_(nullptr),
-        error_(acc.error_) {}
-
-  void SetState(StateId s) {
-    const auto *weights = data_->Weights();
-    const auto *weight_positions = data_->WeightPositions();
-    state_weights_ = nullptr;
-    if (s < data_->NumPositions()) {
-      const auto pos = weight_positions[s];
-      if (pos >= 0) state_weights_ = &(weights[pos]);
-    }
-  }
-
-  Weight Sum(Weight w, Weight v) const { return LogPlus(w, v); }
-
-  template <class ArcIter>
-  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) const {
-    if (error_) return Weight::NoWeight();
-    auto sum = w;
-    // Finds begin and end of pre-stored weights.
-    ssize_t index_begin = -1;
-    ssize_t index_end = -1;
-    ssize_t stored_begin = end;
-    ssize_t stored_end = end;
-    if (state_weights_) {
-      index_begin = begin > 0 ? (begin - 1) / arc_period_ + 1 : 0;
-      index_end = end / arc_period_;
-      stored_begin = index_begin * arc_period_;
-      stored_end = index_end * arc_period_;
-    }
-    // Computes sum before pre-stored weights.
-    if (begin < stored_begin) {
-      const auto pos_end = std::min(stored_begin, end);
-      aiter->Seek(begin);
-      for (auto pos = begin; pos < pos_end; aiter->Next(), ++pos) {
-        sum = LogPlus(sum, aiter->Value().weight);
-      }
-    }
-    // Computes sum between pre-stored weights.
-    if (stored_begin < stored_end) {
-      const auto f1 = state_weights_[index_end];
-      const auto f2 = state_weights_[index_begin];
-      if (f1 < f2) sum = LogPlus(sum, LogMinus(f1, f2));
-      // Commented out for efficiency; adds Zero().
-      /*
-      else {
-        // explicitly computes if cumulative sum lacks precision
-        aiter->Seek(stored_begin);
-        for (auto pos = stored_begin; pos < stored_end; aiter->Next(), ++pos)
-          sum = LogPlus(sum, aiter->Value().weight);
-      }
-      */
-    }
-    // Computes sum after pre-stored weights.
-    if (stored_end < end) {
-      const auto pos_start = std::max(stored_begin, stored_end);
-      aiter->Seek(pos_start);
-      for (auto pos = pos_start; pos < end; aiter->Next(), ++pos) {
-        sum = LogPlus(sum, aiter->Value().weight);
-      }
-    }
-    return sum;
-  }
-
-  template <class FST>
-  void Init(const FST &fst, bool copy = false) {
-    if (copy || !data_->IsMutable()) return;
-    if (data_->NumPositions() != 0 || arc_limit_ < arc_period_) {
-      FSTERROR() << "FastLogAccumulator: Initialization error";
-      error_ = true;
-      return;
-    }
-    std::vector<double> weights;
-    std::vector<int> weight_positions;
-    weight_positions.reserve(CountStates(fst));
-    for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      if (fst.NumArcs(s) >= arc_limit_) {
-        auto sum = FloatLimits<double>::PosInfinity();
-        if (weight_positions.size() <= s) weight_positions.resize(s + 1, -1);
-        weight_positions[s] = weights.size();
-        weights.push_back(sum);
-        size_t narcs = 0;
-        ArcIterator<FST> aiter(fst, s);
-        aiter.SetFlags(kArcWeightValue | kArcNoCache, kArcFlags);
-        for (; !aiter.Done(); aiter.Next()) {
-          const auto &arc = aiter.Value();
-          sum = LogPlus(sum, arc.weight);
-          // Stores cumulative weight distribution per arc_period_.
-          if (++narcs % arc_period_ == 0) weights.push_back(sum);
-        }
-      }
-    }
-    data_->SetData(&weights, &weight_positions);
-  }
-
-  bool Error() const { return error_; }
-
-  std::shared_ptr<FastLogAccumulatorData> GetData() const { return data_; }
-
- private:
-  static double LogPosExp(double x) {
-    return x == FloatLimits<double>::PosInfinity() ? 0.0
-                                                   : log(1.0F + exp(-x));
-  }
-
-  static double LogMinusExp(double x) {
-    return x == FloatLimits<double>::PosInfinity() ? 0.0
-                                                   : log(1.0F - exp(-x));
-  }
-
-  Weight LogPlus(Weight w, Weight v) const {
-    const auto f1 = to_log_weight_(w).Value();
-    const auto f2 = to_log_weight_(v).Value();
-    if (f1 > f2) {
-      return to_weight_(Log64Weight(f2 - LogPosExp(f1 - f2)));
-    } else {
-      return to_weight_(Log64Weight(f1 - LogPosExp(f2 - f1)));
-    }
-  }
-
-  double LogPlus(double f1, Weight v) const {
-    const auto f2 = to_log_weight_(v).Value();
-    if (f1 == FloatLimits<double>::PosInfinity()) {
-      return f2;
-    } else if (f1 > f2) {
-      return f2 - LogPosExp(f1 - f2);
-    } else {
-      return f1 - LogPosExp(f2 - f1);
-    }
-  }
-
-  // Assumes f1 < f2.
-  Weight LogMinus(double f1, double f2) const {
-    if (f2 == FloatLimits<double>::PosInfinity()) {
-      return to_weight_(Log64Weight(f1));
-    } else {
-      return to_weight_(Log64Weight(f1 - LogMinusExp(f2 - f1)));
-    }
-  }
-
-  const WeightConvert<Weight, Log64Weight> to_log_weight_;
-  const WeightConvert<Log64Weight, Weight> to_weight_;
-  const ssize_t arc_limit_;   // Minimum number of arcs to pre-compute state.
-  const ssize_t arc_period_;  // Saves cumulative weights per arc_period_.
-  std::shared_ptr<FastLogAccumulatorData> data_;
-  const double *state_weights_;
-  bool error_;
-
-  FastLogAccumulator &operator=(const FastLogAccumulator &) = delete;
-};
-
-// Stores shareable data for cache log accumulator copies. All copies share the
-// same cache.
-template <class Arc>
-class CacheLogAccumulatorData {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  CacheLogAccumulatorData(bool gc, size_t gc_limit)
-      : cache_gc_(gc), cache_limit_(gc_limit), cache_size_(0) {}
-
-  CacheLogAccumulatorData(const CacheLogAccumulatorData<Arc> &data)
-      : cache_gc_(data.cache_gc_),
-        cache_limit_(data.cache_limit_),
-        cache_size_(0) {}
-
-  bool CacheDisabled() const { return cache_gc_ && cache_limit_ == 0; }
-
-  std::vector<double> *GetWeights(StateId s) {
-    auto it = cache_.find(s);
-    if (it != cache_.end()) {
-      it->second.recent = true;
-      return it->second.weights.get();
-    } else {
-      return nullptr;
-    }
-  }
-
-  void AddWeights(StateId s, std::vector<double> *weights) {
-    if (cache_gc_ && cache_size_ >= cache_limit_) GC(false);
-    cache_.insert(std::make_pair(s, CacheState(weights, true)));
-    if (cache_gc_) cache_size_ += weights->capacity() * sizeof(double);
-  }
-
- private:
-  // Cached information for a given state.
-  struct CacheState {
-    std::unique_ptr<std::vector<double>> weights;  // Accumulated weights.
-    bool recent;  // Has this state been accessed since last GC?
-
-    CacheState(std::vector<double> *weights, bool recent)
-        : weights(weights), recent(recent) {}
-  };
-
-  // Garbage collect: Deletes from cache states that have not been accessed
-  // since the last GC ('free_recent = false') until 'cache_size_' is 2/3 of
-  // 'cache_limit_'. If it does not free enough memory, start deleting
-  // recently accessed states.
-  void GC(bool free_recent) {
-    auto cache_target = (2 * cache_limit_) / 3 + 1;
-    auto it = cache_.begin();
-    while (it != cache_.end() && cache_size_ > cache_target) {
-      auto &cs = it->second;
-      if (free_recent || !cs.recent) {
-        cache_size_ -= cs.weights->capacity() * sizeof(double);
-        cache_.erase(it++);
-      } else {
-        cs.recent = false;
-        ++it;
-      }
-    }
-    if (!free_recent && cache_size_ > cache_target) GC(true);
-  }
-
-  std::unordered_map<StateId, CacheState> cache_;  // Cache.
-  bool cache_gc_;       // Enables garbage collection.
-  size_t cache_limit_;  // # of bytes cached.
-  size_t cache_size_;   // # of bytes allowed before GC.
-
-  CacheLogAccumulatorData &operator=(const CacheLogAccumulatorData &) = delete;
-};
-
-// This class accumulates arc weights using the log semiring Plus() has a
-// WeightConvert specialization to and from log64 weights. It is similar to the
-// FastLogAccumator. However here, the accumulated weights are pre-computed and
-// stored only for the states that are visited. The member function Init(fst)
-// has to be called to setup this accumulator.
-template <class Arc>
-class CacheLogAccumulator {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit CacheLogAccumulator(ssize_t arc_limit = 10, bool gc = false,
-                               size_t gc_limit = 10 * 1024 * 1024)
-      : arc_limit_(arc_limit),
-        data_(std::make_shared<CacheLogAccumulatorData<Arc>>(gc, gc_limit)),
-        s_(kNoStateId),
-        error_(false) {}
-
-  CacheLogAccumulator(const CacheLogAccumulator<Arc> &acc, bool safe = false)
-      : arc_limit_(acc.arc_limit_),
-        fst_(acc.fst_ ? acc.fst_->Copy() : nullptr),
-        data_(safe ? std::make_shared<CacheLogAccumulatorData<Arc>>(*acc.data_)
-                   : acc.data_),
-        s_(kNoStateId),
-        error_(acc.error_) {}
-
-  // Argument arc_limit specifies the minimum number of arcs to pre-compute.
-  void Init(const Fst<Arc> &fst, bool copy = false) {
-    if (!copy && fst_) {
-      FSTERROR() << "CacheLogAccumulator: Initialization error";
-      error_ = true;
-      return;
-    }
-    fst_.reset(fst.Copy());
-  }
-
-  void SetState(StateId s, int depth = 0) {
-    if (s == s_) return;
-    s_ = s;
-    if (data_->CacheDisabled() || error_) {
-      weights_ = nullptr;
-      return;
-    }
-    if (!fst_) {
-      FSTERROR() << "CacheLogAccumulator::SetState: Incorrectly initialized";
-      error_ = true;
-      weights_ = nullptr;
-      return;
-    }
-    weights_ = data_->GetWeights(s);
-    if ((weights_ == nullptr) && (fst_->NumArcs(s) >= arc_limit_)) {
-      weights_ = new std::vector<double>;
-      weights_->reserve(fst_->NumArcs(s) + 1);
-      weights_->push_back(FloatLimits<double>::PosInfinity());
-      data_->AddWeights(s, weights_);
-    }
-  }
-
-  Weight Sum(Weight w, Weight v) { return LogPlus(w, v); }
-
-  template <class ArcIter>
-  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) {
-    if (weights_ == nullptr) {
-      auto sum = w;
-      aiter->Seek(begin);
-      for (auto pos = begin; pos < end; aiter->Next(), ++pos) {
-        sum = LogPlus(sum, aiter->Value().weight);
-      }
-      return sum;
-    } else {
-      Extend(end, aiter);
-      const auto &f1 = (*weights_)[end];
-      const auto &f2 = (*weights_)[begin];
-      if (f1 < f2) {
-        return LogPlus(w, LogMinus(f1, f2));
-      } else {
-        // Commented out for efficiency; adds Zero().
-        /*
-        auto sum = w;
-        // Explicitly computes if cumulative sum lacks precision.
-        aiter->Seek(begin);
-        for (auto pos = begin; pos < end; aiter->Next(), ++pos) {
-          sum = LogPlus(sum, aiter->Value().weight);
-        }
-        return sum;
-        */
-        return w;
-      }
-    }
-  }
-
-  // Returns first position from aiter->Position() whose accumulated
-  // value is greater or equal to w (w.r.t. Zero() < One()). The
-  // iterator may be repositioned.
-  template <class ArcIter>
-  size_t LowerBound(Weight w, ArcIter *aiter) {
-    const auto f = to_log_weight_(w).Value();
-    auto pos = aiter->Position();
-    if (weights_) {
-      Extend(fst_->NumArcs(s_), aiter);
-      return std::lower_bound(weights_->begin() + pos + 1, weights_->end(),
-                              f, std::greater<double>()) -
-          weights_->begin() - 1;
-    } else {
-      size_t n = 0;
-      auto x = FloatLimits<double>::PosInfinity();
-      for (aiter->Reset(); !aiter->Done(); aiter->Next(), ++n) {
-        x = LogPlus(x, aiter->Value().weight);
-        if (n >= pos && x <= f) break;
-      }
-      return n;
-    }
-  }
-
-  bool Error() const { return error_; }
-
- private:
-  double LogPosExp(double x) {
-    return x == FloatLimits<double>::PosInfinity() ? 0.0
-                                                   : log(1.0F + exp(-x));
-  }
-
-  double LogMinusExp(double x) {
-    return x == FloatLimits<double>::PosInfinity() ? 0.0
-                                                   : log(1.0F - exp(-x));
-  }
-
-  Weight LogPlus(Weight w, Weight v) {
-    const auto f1 = to_log_weight_(w).Value();
-    const auto f2 = to_log_weight_(v).Value();
-    if (f1 > f2) {
-      return to_weight_(Log64Weight(f2 - LogPosExp(f1 - f2)));
-    } else {
-      return to_weight_(Log64Weight(f1 - LogPosExp(f2 - f1)));
-    }
-  }
-
-  double LogPlus(double f1, Weight v) {
-    const auto f2 = to_log_weight_(v).Value();
-    if (f1 == FloatLimits<double>::PosInfinity()) {
-      return f2;
-    } else if (f1 > f2) {
-      return f2 - LogPosExp(f1 - f2);
-    } else {
-      return f1 - LogPosExp(f2 - f1);
-    }
-  }
-
-  // Assumes f1 < f2.
-  Weight LogMinus(double f1, double f2) {
-    if (f2 == FloatLimits<double>::PosInfinity()) {
-      return to_weight_(Log64Weight(f1));
-    } else {
-      return to_weight_(Log64Weight(f1 - LogMinusExp(f2 - f1)));
-    }
-  }
-
-  // Extends weights up to index 'end'.
-  template <class ArcIter>
-  void Extend(ssize_t end, ArcIter *aiter) {
-    if (weights_->size() <= end) {
-      for (aiter->Seek(weights_->size() - 1); weights_->size() <= end;
-           aiter->Next()) {
-        weights_->push_back(LogPlus(weights_->back(), aiter->Value().weight));
-      }
-    }
-  }
-
-
-  WeightConvert<Weight, Log64Weight> to_log_weight_;
-  WeightConvert<Log64Weight, Weight> to_weight_;
-  ssize_t arc_limit_;                    // Minimum # of arcs to cache a state.
-  std::vector<double> *weights_;         // Accumulated weights for cur. state.
-  std::unique_ptr<const Fst<Arc>> fst_;  // Input FST.
-  std::shared_ptr<CacheLogAccumulatorData<Arc>> data_;  // Cache data.
-  StateId s_;                                           // Current state.
-  bool error_;
-};
-
-// Stores shareable data for replace accumulator copies.
-template <class Accumulator, class T>
-class ReplaceAccumulatorData {
- public:
-  using Arc = typename Accumulator::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using StateTable = T;
-  using StateTuple = typename StateTable::StateTuple;
-
-  ReplaceAccumulatorData() : state_table_(nullptr) {}
-
-  explicit ReplaceAccumulatorData(
-      const std::vector<Accumulator *> &accumulators)
-      : state_table_(nullptr) {
-    accumulators_.reserve(accumulators.size());
-    for (const auto accumulator : accumulators) {
-      accumulators_.emplace_back(accumulator);
-    }
-  }
-
-  void Init(const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_tuples,
-            const StateTable *state_table) {
-    state_table_ = state_table;
-    accumulators_.resize(fst_tuples.size());
-    for (Label i = 0; i < accumulators_.size(); ++i) {
-      if (!accumulators_[i]) {
-        accumulators_[i].reset(new Accumulator());
-        accumulators_[i]->Init(*(fst_tuples[i].second));
-      }
-      fst_array_.emplace_back(fst_tuples[i].second->Copy());
-    }
-  }
-
-  const StateTuple &GetTuple(StateId s) const { return state_table_->Tuple(s); }
-
-  Accumulator *GetAccumulator(size_t i) { return accumulators_[i].get(); }
-
-  const Fst<Arc> *GetFst(size_t i) const { return fst_array_[i].get(); }
-
- private:
-  const StateTable *state_table_;
-  std::vector<std::unique_ptr<Accumulator>> accumulators_;
-  std::vector<std::unique_ptr<const Fst<Arc>>> fst_array_;
-};
-
-// This class accumulates weights in a ReplaceFst.  The 'Init' method takes as
-// input the argument used to build the ReplaceFst and the ReplaceFst state
-// table. It uses accumulators of type 'Accumulator' in the underlying FSTs.
-template <class Accumulator,
-          class T = DefaultReplaceStateTable<typename Accumulator::Arc>>
-class ReplaceAccumulator {
- public:
-  using Arc = typename Accumulator::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using StateTable = T;
-  using StateTuple = typename StateTable::StateTuple;
-  using Weight = typename Arc::Weight;
-
-  ReplaceAccumulator()
-      : init_(false),
-        data_(std::make_shared<
-              ReplaceAccumulatorData<Accumulator, StateTable>>()),
-        error_(false) {}
-
-  explicit ReplaceAccumulator(const std::vector<Accumulator *> &accumulators)
-      : init_(false),
-        data_(std::make_shared<ReplaceAccumulatorData<Accumulator, StateTable>>(
-            accumulators)),
-        error_(false) {}
-
-  ReplaceAccumulator(const ReplaceAccumulator<Accumulator, StateTable> &acc,
-                     bool safe = false)
-      : init_(acc.init_), data_(acc.data_), error_(acc.error_) {
-    if (!init_) {
-      FSTERROR() << "ReplaceAccumulator: Can't copy unintialized accumulator";
-    }
-    if (safe) FSTERROR() << "ReplaceAccumulator: Safe copy not supported";
-  }
-
-  // Does not take ownership of the state table, the state table is owned by
-  // the ReplaceFst.
-  void Init(const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_tuples,
-            const StateTable *state_table) {
-    init_ = true;
-    data_->Init(fst_tuples, state_table);
-  }
-
-  // Method required by LookAheadMatcher. However, ReplaceAccumulator needs to
-  // be initialized by calling the Init method above before being passed to
-  // LookAheadMatcher.
-  //
-  // TODO(allauzen): Revisit this. Consider creating a method
-  // Init(const ReplaceFst<A, T, C>&, bool) and using friendship to get access
-  // to the innards of ReplaceFst.
-  void Init(const Fst<Arc> &fst, bool copy = false) {
-    if (!init_) {
-      FSTERROR() << "ReplaceAccumulator::Init: Accumulator needs to be"
-                 << " initialized before being passed to LookAheadMatcher";
-      error_ = true;
-    }
-  }
-
-  void SetState(StateId s) {
-    if (!init_) {
-      FSTERROR() << "ReplaceAccumulator::SetState: Incorrectly initialized";
-      error_ = true;
-      return;
-    }
-    auto tuple = data_->GetTuple(s);
-    fst_id_ = tuple.fst_id - 1;  // Replace FST ID is 1-based.
-    data_->GetAccumulator(fst_id_)->SetState(tuple.fst_state);
-    if ((tuple.prefix_id != 0) &&
-        (data_->GetFst(fst_id_)->Final(tuple.fst_state) != Weight::Zero())) {
-      offset_ = 1;
-      offset_weight_ = data_->GetFst(fst_id_)->Final(tuple.fst_state);
-    } else {
-      offset_ = 0;
-      offset_weight_ = Weight::Zero();
-    }
-    aiter_.reset(
-        new ArcIterator<Fst<Arc>>(*data_->GetFst(fst_id_), tuple.fst_state));
-  }
-
-  Weight Sum(Weight w, Weight v) {
-    if (error_) return Weight::NoWeight();
-    return data_->GetAccumulator(fst_id_)->Sum(w, v);
-  }
-
-  template <class ArcIter>
-  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) {
-    if (error_) return Weight::NoWeight();
-    auto sum = begin == end ? Weight::Zero()
-                            : data_->GetAccumulator(fst_id_)->Sum(
-                                  w, aiter_.get(), begin ? begin - offset_ : 0,
-                                  end - offset_);
-    if (begin == 0 && end != 0 && offset_ > 0) sum = Sum(offset_weight_, sum);
-    return sum;
-  }
-
-  bool Error() const { return error_; }
-
- private:
-  bool init_;
-  std::shared_ptr<ReplaceAccumulatorData<Accumulator, StateTable>> data_;
-  Label fst_id_;
-  size_t offset_;
-  Weight offset_weight_;
-  std::unique_ptr<ArcIterator<Fst<Arc>>> aiter_;
-  bool error_;
-};
-
-// SafeReplaceAccumulator accumulates weights in a ReplaceFst and copies of it
-// are always thread-safe copies.
-template <class Accumulator, class T>
-class SafeReplaceAccumulator {
- public:
-  using Arc = typename Accumulator::Arc;
-  using StateId = typename Arc::StateId;
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-  using StateTable = T;
-  using StateTuple = typename StateTable::StateTuple;
-
-  SafeReplaceAccumulator() {}
-
-  SafeReplaceAccumulator(const SafeReplaceAccumulator &copy, bool safe)
-      : SafeReplaceAccumulator(copy) {}
-
-  explicit SafeReplaceAccumulator(
-      const std::vector<Accumulator> &accumulators) {
-    for (const auto &accumulator : accumulators) {
-      accumulators_.emplace_back(accumulator, true);
-    }
-  }
-
-  void Init(const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_tuples,
-            const StateTable *state_table) {
-    state_table_ = state_table;
-    for (Label i = 0; i < fst_tuples.size(); ++i) {
-      if (i == accumulators_.size()) {
-        accumulators_.resize(accumulators_.size() + 1);
-        accumulators_[i].Init(*(fst_tuples[i].second));
-      }
-      fst_array_.emplace_back(fst_tuples[i].second->Copy(true));
-    }
-    init_ = true;
-  }
-
-  void Init(const Fst<Arc> &fst, bool copy = false) {
-    if (!init_) {
-      FSTERROR() << "SafeReplaceAccumulator::Init: Accumulator needs to be"
-                 << " initialized before being passed to LookAheadMatcher";
-      error_ = true;
-    }
-  }
-
-  void SetState(StateId s) {
-    auto tuple = state_table_->Tuple(s);
-    fst_id_ = tuple.fst_id - 1;  // Replace FST ID is 1-based
-    GetAccumulator(fst_id_)->SetState(tuple.fst_state);
-    offset_ = 0;
-    offset_weight_ = Weight::Zero();
-    const auto final_weight = GetFst(fst_id_)->Final(tuple.fst_state);
-    if ((tuple.prefix_id != 0) && (final_weight != Weight::Zero())) {
-      offset_ = 1;
-      offset_weight_ = final_weight;
-    }
-    aiter_.Set(*GetFst(fst_id_), tuple.fst_state);
-  }
-
-  Weight Sum(Weight w, Weight v) {
-    if (error_) return Weight::NoWeight();
-    return GetAccumulator(fst_id_)->Sum(w, v);
-  }
-
-  template <class ArcIter>
-  Weight Sum(Weight w, ArcIter *aiter, ssize_t begin, ssize_t end) {
-    if (error_) return Weight::NoWeight();
-    if (begin == end) return Weight::Zero();
-    auto sum = GetAccumulator(fst_id_)->Sum(
-        w, aiter_.get(), begin ? begin - offset_ : 0, end - offset_);
-    if (begin == 0 && end != 0 && offset_ > 0) {
-      sum = Sum(offset_weight_, sum);
-    }
-    return sum;
-  }
-
-  bool Error() const { return error_; }
-
- private:
-  class ArcIteratorPtr {
-   public:
-    ArcIteratorPtr() {}
-
-    ArcIteratorPtr(const ArcIteratorPtr &copy) {}
-
-    void Set(const Fst<Arc> &fst, StateId state_id) {
-      ptr_.reset(new ArcIterator<Fst<Arc>>(fst, state_id));
-    }
-
-    ArcIterator<Fst<Arc>> *get() { return ptr_.get(); }
-
-   private:
-    std::unique_ptr<ArcIterator<Fst<Arc>>> ptr_;
-  };
-
-  Accumulator *GetAccumulator(size_t i) { return &accumulators_[i]; }
-
-  const Fst<Arc> *GetFst(size_t i) const { return fst_array_[i].get(); }
-
-  const StateTable *state_table_;
-  std::vector<Accumulator> accumulators_;
-  std::vector<std::shared_ptr<Fst<Arc>>> fst_array_;
-  ArcIteratorPtr aiter_;
-  bool init_ = false;
-  bool error_ = false;
-  Label fst_id_;
-  size_t offset_;
-  Weight offset_weight_;
-};
-
-}  // namespace fst
-
-#endif  // FST_ACCUMULATOR_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/add-on.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/add-on.h
deleted file mode 100644
index 4a95111f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/add-on.h
+++ /dev/null
@@ -1,248 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// FST implementation class to attach an arbitrary object with a read/write
-// method to an FST and its file representation. The FST is given a new type
-// name.
-
-#ifndef FST_ADD_ON_H_
-#define FST_ADD_ON_H_
-
-#include <stddef.h>
-#include <memory>
-#include <string>
-#include <utility>
-
-#include <fst/log.h>
-
-#include <fst/fst.h>
-
-
-namespace fst {
-
-// Identifies stream data as an add-on FST.
-static constexpr int32 kAddOnMagicNumber = 446681434;
-
-// Nothing to save.
-class NullAddOn {
- public:
-  NullAddOn() {}
-
-  static NullAddOn *Read(std::istream &strm, const FstReadOptions &opts) {
-    return new NullAddOn();
-  }
-
-  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
-    return true;
-  }
-};
-
-// Create a new add-on from a pair of add-ons.
-template <class A1, class A2>
-class AddOnPair {
- public:
-  // Argument reference count incremented.
-  AddOnPair(std::shared_ptr<A1> a1, std::shared_ptr<A2> a2)
-      : a1_(std::move(a1)), a2_(std::move(a2)) {}
-
-  const A1 *First() const { return a1_.get(); }
-
-  const A2 *Second() const { return a2_.get(); }
-
-  std::shared_ptr<A1> SharedFirst() const { return a1_; }
-
-  std::shared_ptr<A2> SharedSecond() const { return a2_; }
-
-  static AddOnPair<A1, A2> *Read(std::istream &istrm,
-                                 const FstReadOptions &opts) {
-    A1 *a1 = nullptr;
-    bool have_addon1 = false;
-    ReadType(istrm, &have_addon1);
-    if (have_addon1) a1 = A1::Read(istrm, opts);
-
-    A2 *a2 = nullptr;
-    bool have_addon2 = false;
-    ReadType(istrm, &have_addon2);
-    if (have_addon2) a2 = A2::Read(istrm, opts);
-
-    return new AddOnPair<A1, A2>(std::shared_ptr<A1>(a1),
-                                 std::shared_ptr<A2>(a2));
-  }
-
-  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
-    bool have_addon1 = a1_ != nullptr;
-    WriteType(ostrm, have_addon1);
-    if (have_addon1) a1_->Write(ostrm, opts);
-    bool have_addon2 = a2_ != nullptr;
-    WriteType(ostrm, have_addon2);
-    if (have_addon2) a2_->Write(ostrm, opts);
-    return true;
-  }
-
- private:
-  std::shared_ptr<A1> a1_;
-  std::shared_ptr<A2> a2_;
-};
-
-namespace internal {
-
-// Adds an object of type T to an FST. T must support:
-//
-//     T* Read(std::istream &);
-//     bool Write(std::ostream &);
-//
-// The resulting type is a new FST implementation.
-template <class FST, class T>
-class AddOnImpl : public FstImpl<typename FST::Arc> {
- public:
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::WriteHeader;
-
-  // We make a thread-safe copy of the FST by default since an FST
-  // implementation is expected to not share mutable data between objects.
-  AddOnImpl(const FST &fst, const string &type,
-            std::shared_ptr<T> t = std::shared_ptr<T>())
-      : fst_(fst, true), t_(std::move(t)) {
-    SetType(type);
-    SetProperties(fst_.Properties(kFstProperties, false));
-    SetInputSymbols(fst_.InputSymbols());
-    SetOutputSymbols(fst_.OutputSymbols());
-  }
-
-  // Conversion from const Fst<Arc> & to F always copies the underlying
-  // implementation.
-  AddOnImpl(const Fst<Arc> &fst, const string &type,
-            std::shared_ptr<T> t = std::shared_ptr<T>())
-      : fst_(fst), t_(std::move(t)) {
-    SetType(type);
-    SetProperties(fst_.Properties(kFstProperties, false));
-    SetInputSymbols(fst_.InputSymbols());
-    SetOutputSymbols(fst_.OutputSymbols());
-  }
-
-  // We make a thread-safe copy of the FST by default since an FST
-  // implementation is expected to not share mutable data between objects.
-  AddOnImpl(const AddOnImpl<FST, T> &impl)
-      : fst_(impl.fst_, true), t_(impl.t_) {
-    SetType(impl.Type());
-    SetProperties(fst_.Properties(kCopyProperties, false));
-    SetInputSymbols(fst_.InputSymbols());
-    SetOutputSymbols(fst_.OutputSymbols());
-  }
-
-  StateId Start() const { return fst_.Start(); }
-
-  Weight Final(StateId s) const { return fst_.Final(s); }
-
-  size_t NumArcs(StateId s) const { return fst_.NumArcs(s); }
-
-  size_t NumInputEpsilons(StateId s) const { return fst_.NumInputEpsilons(s); }
-
-  size_t NumOutputEpsilons(StateId s) const {
-    return fst_.NumOutputEpsilons(s);
-  }
-
-  size_t NumStates() const { return fst_.NumStates(); }
-
-  static AddOnImpl<FST, T> *Read(std::istream &strm,
-                                 const FstReadOptions &opts) {
-    FstReadOptions nopts(opts);
-    FstHeader hdr;
-    if (!nopts.header) {
-      hdr.Read(strm, nopts.source);
-      nopts.header = &hdr;
-    }
-    std::unique_ptr<AddOnImpl<FST, T>> impl(
-        new AddOnImpl<FST, T>(nopts.header->FstType()));
-    if (!impl->ReadHeader(strm, nopts, kMinFileVersion, &hdr)) return nullptr;
-    impl.reset();
-    int32 magic_number = 0;
-    ReadType(strm, &magic_number);  // Ensures this is an add-on FST.
-    if (magic_number != kAddOnMagicNumber) {
-      LOG(ERROR) << "AddOnImpl::Read: Bad add-on header: " << nopts.source;
-      return nullptr;
-    }
-    FstReadOptions fopts(opts);
-    fopts.header = nullptr;  // Contained header was written out.
-    std::unique_ptr<FST> fst(FST::Read(strm, fopts));
-    if (!fst) return nullptr;
-    std::shared_ptr<T> t;
-    bool have_addon = false;
-    ReadType(strm, &have_addon);
-    if (have_addon) {  // Reads add-on object if present.
-      t = std::shared_ptr<T>(T::Read(strm, fopts));
-      if (!t) return nullptr;
-    }
-    return new AddOnImpl<FST, T>(*fst, nopts.header->FstType(), t);
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
-    FstHeader hdr;
-    FstWriteOptions nopts(opts);
-    nopts.write_isymbols = false;  // Allows contained FST to hold any symbols.
-    nopts.write_osymbols = false;
-    WriteHeader(strm, nopts, kFileVersion, &hdr);
-    WriteType(strm, kAddOnMagicNumber);  // Ensures this is an add-on FST.
-    FstWriteOptions fopts(opts);
-    fopts.write_header = true;  // Forces writing contained header.
-    if (!fst_.Write(strm, fopts)) return false;
-    bool have_addon = !!t_;
-    WriteType(strm, have_addon);
-    // Writes add-on object if present.
-    if (have_addon) t_->Write(strm, opts);
-    return true;
-  }
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const {
-    fst_.InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const {
-    fst_.InitArcIterator(s, data);
-  }
-
-  FST &GetFst() { return fst_; }
-
-  const FST &GetFst() const { return fst_; }
-
-  const T *GetAddOn() const { return t_.get(); }
-
-  std::shared_ptr<T> GetSharedAddOn() const { return t_; }
-
-  void SetAddOn(std::shared_ptr<T> t) { t_ = t; }
-
- private:
-  explicit AddOnImpl(const string &type) : t_() {
-    SetType(type);
-    SetProperties(kExpanded);
-  }
-
-  // Current file format version.
-  static constexpr int kFileVersion = 1;
-  // Minimum file format version supported.
-  static constexpr int kMinFileVersion = 1;
-
-  FST fst_;
-  std::shared_ptr<T> t_;
-
-  AddOnImpl &operator=(const AddOnImpl &) = delete;
-};
-
-template <class FST, class T>
-constexpr int AddOnImpl<FST, T>::kFileVersion;
-
-template <class FST, class T>
-constexpr int AddOnImpl<FST, T>::kMinFileVersion;
-
-}  // namespace internal
-}  // namespace fst
-
-#endif  // FST_ADD_ON_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arc-arena.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arc-arena.h
deleted file mode 100644
index 13fe918a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arc-arena.h
+++ /dev/null
@@ -1,232 +0,0 @@
-#ifndef FST_ARC_ARENA_H_
-#define FST_ARC_ARENA_H_
-
-#include <deque>
-#include <memory>
-#include <utility>
-#include <fst/fst.h>
-#include <fst/memory.h>
-#include <unordered_map>
-
-namespace fst {
-
-// ArcArena is used for fast allocation of contiguous arrays of arcs.
-//
-// To create an arc array:
-//   for each state:
-//     for each arc:
-//       arena.PushArc();
-//     // Commits these arcs and returns pointer to them.
-//     Arc *arcs = arena.GetArcs();
-//
-//     OR
-//
-//     arena.DropArcs();  // Throws away current arcs, reuse the space.
-//
-// The arcs returned are guaranteed to be contiguous and the pointer returned
-// will never be invalidated until the arena is cleared for reuse.
-//
-// The contents of the arena can be released with a call to arena.Clear() after
-// which the arena will restart with an initial allocation capable of holding at
-// least all of the arcs requested in the last usage before Clear() making
-// subsequent uses of the Arena more efficient.
-//
-// The max_retained_size option can limit the amount of arc space requested on
-// Clear() to avoid excess growth from intermittent high usage.
-template <typename Arc>
-class ArcArena {
- public:
-  explicit ArcArena(size_t block_size = 256,
-                    size_t max_retained_size = 1e6)
-      : block_size_(block_size),
-        max_retained_size_(max_retained_size) {
-    blocks_.emplace_back(MakeSharedBlock(block_size_));
-    first_block_size_ = block_size_;
-    total_size_ = block_size_;
-    arcs_ = blocks_.back().get();
-    end_ = arcs_ + block_size_;
-    next_ = arcs_;
-  }
-
-  ArcArena(const ArcArena& copy)
-      : arcs_(copy.arcs_), next_(copy.next_), end_(copy.end_),
-        block_size_(copy.block_size_),
-        first_block_size_(copy.first_block_size_),
-        total_size_(copy.total_size_),
-        max_retained_size_(copy.max_retained_size_),
-        blocks_(copy.blocks_) {
-    NewBlock(block_size_);
-  }
-
-  void ReserveArcs(size_t n) {
-    if (next_ + n < end_) return;
-    NewBlock(n);
-  }
-
-  void PushArc(const Arc& arc) {
-    if (next_ == end_) {
-      size_t length = next_ - arcs_;
-      NewBlock(length * 2);
-    }
-    *next_ = arc;
-    ++next_;
-  }
-
-  const Arc* GetArcs() {
-    const auto *arcs = arcs_;
-    arcs_ = next_;
-    return arcs;
-  }
-
-  void DropArcs() { next_ = arcs_; }
-
-  size_t Size() { return total_size_; }
-
-  void Clear() {
-    blocks_.resize(1);
-    if (total_size_ > first_block_size_) {
-      first_block_size_ = std::min(max_retained_size_, total_size_);
-      blocks_.back() = MakeSharedBlock(first_block_size_);
-    }
-    total_size_ = first_block_size_;
-    arcs_ = blocks_.back().get();
-    end_ = arcs_ + first_block_size_;
-    next_ = arcs_;
-  }
-
- private:
-  // Allocates a new block with capacity of at least n or block_size,
-  // copying incomplete arc sequence from old block to new block.
-  void NewBlock(size_t n) {
-    const auto length = next_ - arcs_;
-    const auto new_block_size = std::max(n, block_size_);
-    total_size_ += new_block_size;
-    blocks_.emplace_back(MakeSharedBlock(new_block_size));
-    std::copy(arcs_, next_, blocks_.back().get());
-    arcs_ = blocks_.back().get();
-    next_ = arcs_ + length;
-    end_ = arcs_ + new_block_size;
-  }
-
-  std::shared_ptr<Arc> MakeSharedBlock(size_t size) {
-    return std::shared_ptr<Arc>(new Arc[size], std::default_delete<Arc[]>());
-  }
-
-  Arc *arcs_;
-  Arc *next_;
-  const Arc *end_;
-  size_t block_size_;
-  size_t first_block_size_;
-  size_t total_size_;
-  size_t max_retained_size_;
-  std::list<std::shared_ptr<Arc>> blocks_;
-};
-
-// ArcArenaStateStore uses a resusable ArcArena to store arc arrays and does not
-// require that the Expander call ReserveArcs first.
-//
-// TODO(tombagby): Make cache type configurable.
-// TODO(tombagby): Provide ThreadLocal/Concurrent configuration.
-template <class A>
-class ArcArenaStateStore {
- public:
-  using Arc = A;
-  using Weight = typename Arc::Weight;
-  using StateId = typename Arc::StateId;
-
-  ArcArenaStateStore() : arena_(64 * 1024) {
-  }
-
-  class State {
-   public:
-    Weight Final() const { return final_; }
-
-    size_t NumInputEpsilons() const { return niepsilons_; }
-
-    size_t NumOutputEpsilons() const { return noepsilons_; }
-
-    size_t NumArcs() const { return narcs_; }
-
-    const Arc &GetArc(size_t n) const { return arcs_[n]; }
-
-    const Arc *Arcs() const { return arcs_; }
-
-    int* MutableRefCount() const { return nullptr; }
-
-   private:
-    State(Weight weight, int32 niepsilons, int32 noepsilons, int32 narcs,
-          const Arc *arcs)
-        : final_(std::move(weight)),
-          niepsilons_(niepsilons),
-          noepsilons_(noepsilons),
-          narcs_(narcs),
-          arcs_(arcs) {}
-
-    Weight final_;
-    size_t niepsilons_;
-    size_t noepsilons_;
-    size_t narcs_;
-    const Arc *arcs_;
-
-    friend class ArcArenaStateStore<Arc>;
-  };
-
-  template <class Expander>
-  State *FindOrExpand(Expander &expander, StateId state_id) {  // NOLINT
-    auto it = cache_.insert(std::pair<StateId, State*>(state_id, nullptr));
-    if (!it.second) return it.first->second;
-    // Needs a new state.
-    StateBuilder builder(&arena_);
-    expander.Expand(state_id, &builder);
-    const auto arcs = arena_.GetArcs();
-    size_t narcs = builder.narcs_;
-    size_t niepsilons = 0;
-    size_t noepsilons = 0;
-    for (size_t i = 0; i < narcs; ++i) {
-      if (arcs[i].ilabel == 0) ++niepsilons;
-      if (arcs[i].olabel == 0) ++noepsilons;
-    }
-    states_.emplace_back(
-        State(builder.final_, niepsilons, noepsilons, narcs, arcs));
-    // Places it in the cache.
-    auto state = &states_.back();
-    it.first->second = state;
-    return state;
-  }
-
-  State *Find(StateId state_id) const {
-    auto it = cache_.find(state_id);
-    return (it == cache_.end()) ? nullptr : it->second;
-  }
-
- private:
-  class StateBuilder {
-   public:
-    explicit StateBuilder(ArcArena<Arc>* arena)
-       : arena_(arena), final_(Weight::Zero()), narcs_(0) {}
-
-    void SetFinal(Weight weight) { final_ = std::move(weight); }
-
-    void ReserveArcs(size_t n) { arena_->ReserveArcs(n); }
-
-    void AddArc(const Arc &arc) {
-      ++narcs_;
-      arena_->PushArc(arc);
-    }
-
-   private:
-    friend class ArcArenaStateStore<Arc>;
-
-    ArcArena<Arc> *arena_;
-    Weight final_;
-    size_t narcs_;
-  };
-
-  std::unordered_map<StateId, State *> cache_;
-  std::deque<State> states_;
-  ArcArena<Arc> arena_;
-};
-
-}  // namespace fst
-
-#endif  // FST_ARC_ARENA_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arc-map.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arc-map.h
deleted file mode 100644
index ae677c81..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arc-map.h
+++ /dev/null
@@ -1,1275 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to map over/transform arcs e.g., change semirings or
-// implement project/invert. Consider using when operation does
-// not change the number of arcs (except possibly superfinal arcs).
-
-#ifndef FST_ARC_MAP_H_
-#define FST_ARC_MAP_H_
-
-#include <string>
-#include <unordered_map>
-#include <utility>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-// Determines how final weights are mapped.
-enum MapFinalAction {
-  // A final weight is mapped into a final weight. An error is raised if this
-  // is not possible.
-  MAP_NO_SUPERFINAL,
-  // A final weight is mapped to an arc to the superfinal state when the result
-  // cannot be represented as a final weight. The superfinal state will be
-  // added only if it is needed.
-  MAP_ALLOW_SUPERFINAL,
-  // A final weight is mapped to an arc to the superfinal state unless the
-  // result can be represented as a final weight of weight Zero(). The
-  // superfinal state is always added (if the input is not the empty FST).
-  MAP_REQUIRE_SUPERFINAL
-};
-
-// Determines how symbol tables are mapped.
-enum MapSymbolsAction {
-  // Symbols should be cleared in the result by the map.
-  MAP_CLEAR_SYMBOLS,
-  // Symbols should be copied from the input FST by the map.
-  MAP_COPY_SYMBOLS,
-  // Symbols should not be modified in the result by the map itself.
-  // (They may set by the mapper).
-  MAP_NOOP_SYMBOLS
-};
-
-// The ArcMapper interfaces defines how arcs and final weights are mapped.
-// This is useful for implementing operations that do not change the number of
-// arcs (expect possibly superfinal arcs).
-//
-// template <class A, class B>
-// class ArcMapper {
-//  public:
-//   using FromArc = A;
-//   using ToArc = B;
-//
-//   // Maps an arc type FromArc to arc type ToArc.
-//   ToArc operator()(const FromArc &arc);
-//
-//   // Specifies final action the mapper requires (see above).
-//   // The mapper will be passed final weights as arcs of the form
-//   // Arc(0, 0, weight, kNoStateId).
-//   MapFinalAction FinalAction() const;
-//
-//   // Specifies input symbol table action the mapper requires (see above).
-//   MapSymbolsAction InputSymbolsAction() const;
-//
-//   // Specifies output symbol table action the mapper requires (see above).
-//   MapSymbolsAction OutputSymbolsAction() const;
-//
-//   // This specifies the known properties of an FST mapped by this mapper. It
-//   takes as argument the input FSTs's known properties.
-//   uint64 Properties(uint64 props) const;
-// };
-//
-// The ArcMap functions and classes below will use the FinalAction()
-// method of the mapper to determine how to treat final weights, e.g., whether
-// to add a superfinal state. They will use the Properties() method to set the
-// result FST properties.
-//
-// We include a various map versions below. One dimension of variation is
-// whether the mapping mutates its input, writes to a new result FST, or is an
-// on-the-fly FST. Another dimension is how we pass the mapper. We allow passing
-// the mapper by pointer for cases that we need to change the state of the
-// user's mapper.  This is the case with the EncodeMapper, which is reused
-// during decoding. We also include map versions that pass the mapper by value
-// or const reference when this suffices.
-
-// Maps an arc type A using a mapper function object C, passed
-// by pointer.  This version modifies its Fst input.
-template <class A, class C>
-void ArcMap(MutableFst<A> *fst, C *mapper) {
-  using FromArc = A;
-  using ToArc = A;
-  using StateId = typename FromArc::StateId;
-  using Weight = typename FromArc::Weight;
-  if (mapper->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    fst->SetInputSymbols(nullptr);
-  }
-  if (mapper->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    fst->SetOutputSymbols(nullptr);
-  }
-  if (fst->Start() == kNoStateId) return;
-  const auto props = fst->Properties(kFstProperties, false);
-  const auto final_action = mapper->FinalAction();
-  auto superfinal = kNoStateId;
-  if (final_action == MAP_REQUIRE_SUPERFINAL) {
-    superfinal = fst->AddState();
-    fst->SetFinal(superfinal, Weight::One());
-  }
-  for (StateIterator<MutableFst<FromArc>> siter(*fst); !siter.Done();
-       siter.Next()) {
-    const auto state = siter.Value();
-    for (MutableArcIterator<MutableFst<FromArc>> aiter(fst, state);
-         !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      aiter.SetValue((*mapper)(arc));
-    }
-    switch (final_action) {
-      case MAP_NO_SUPERFINAL:
-      default: {
-        const FromArc arc(0, 0, fst->Final(state), kNoStateId);
-        const auto final_arc = (*mapper)(arc);
-        if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
-          FSTERROR() << "ArcMap: Non-zero arc labels for superfinal arc";
-          fst->SetProperties(kError, kError);
-        }
-        fst->SetFinal(state, final_arc.weight);
-        break;
-      }
-      case MAP_ALLOW_SUPERFINAL: {
-        if (state != superfinal) {
-          const FromArc arc(0, 0, fst->Final(state), kNoStateId);
-          auto final_arc = (*mapper)(arc);
-          if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
-            // Add a superfinal state if not already done.
-            if (superfinal == kNoStateId) {
-              superfinal = fst->AddState();
-              fst->SetFinal(superfinal, Weight::One());
-            }
-            final_arc.nextstate = superfinal;
-            fst->AddArc(state, final_arc);
-            fst->SetFinal(state, Weight::Zero());
-          } else {
-            fst->SetFinal(state, final_arc.weight);
-          }
-        }
-        break;
-      }
-      case MAP_REQUIRE_SUPERFINAL: {
-        if (state != superfinal) {
-          const FromArc arc(0, 0, fst->Final(state), kNoStateId);
-          const auto final_arc = (*mapper)(arc);
-          if (final_arc.ilabel != 0 || final_arc.olabel != 0 ||
-              final_arc.weight != Weight::Zero()) {
-            fst->AddArc(state, ToArc(final_arc.ilabel, final_arc.olabel,
-                                     final_arc.weight, superfinal));
-          }
-          fst->SetFinal(state, Weight::Zero());
-        }
-        break;
-      }
-    }
-  }
-  fst->SetProperties(mapper->Properties(props), kFstProperties);
-}
-
-// Maps an arc type A using a mapper function object C, passed by value. This
-// version modifies its FST input.
-template <class A, class C>
-void ArcMap(MutableFst<A> *fst, C mapper) {
-  ArcMap(fst, &mapper);
-}
-
-// Maps an arc type A to an arc type B using mapper function object C,
-// passed by pointer. This version writes the mapped input FST to an
-// output MutableFst.
-template <class A, class B, class C>
-void ArcMap(const Fst<A> &ifst, MutableFst<B> *ofst, C *mapper) {
-  using FromArc = A;
-  using StateId = typename FromArc::StateId;
-  using Weight = typename FromArc::Weight;
-  ofst->DeleteStates();
-  if (mapper->InputSymbolsAction() == MAP_COPY_SYMBOLS) {
-    ofst->SetInputSymbols(ifst.InputSymbols());
-  } else if (mapper->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    ofst->SetInputSymbols(nullptr);
-  }
-  if (mapper->OutputSymbolsAction() == MAP_COPY_SYMBOLS) {
-    ofst->SetOutputSymbols(ifst.OutputSymbols());
-  } else if (mapper->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    ofst->SetOutputSymbols(nullptr);
-  }
-  const auto iprops = ifst.Properties(kCopyProperties, false);
-  if (ifst.Start() == kNoStateId) {
-    if (iprops & kError) ofst->SetProperties(kError, kError);
-    return;
-  }
-  const auto final_action = mapper->FinalAction();
-  if (ifst.Properties(kExpanded, false)) {
-    ofst->ReserveStates(
-        CountStates(ifst) + final_action == MAP_NO_SUPERFINAL ? 0 : 1);
-  }
-  // Adds all states.
-  for (StateIterator<Fst<A>> siter(ifst); !siter.Done(); siter.Next()) {
-    ofst->AddState();
-  }
-  StateId superfinal = kNoStateId;
-  if (final_action == MAP_REQUIRE_SUPERFINAL) {
-    superfinal = ofst->AddState();
-    ofst->SetFinal(superfinal, B::Weight::One());
-  }
-  for (StateIterator<Fst<A>> siter(ifst); !siter.Done(); siter.Next()) {
-    StateId s = siter.Value();
-    if (s == ifst.Start()) ofst->SetStart(s);
-    ofst->ReserveArcs(s, ifst.NumArcs(s));
-    for (ArcIterator<Fst<A>> aiter(ifst, s); !aiter.Done(); aiter.Next()) {
-      ofst->AddArc(s, (*mapper)(aiter.Value()));
-    }
-    switch (final_action) {
-      case MAP_NO_SUPERFINAL:
-      default: {
-        B final_arc = (*mapper)(A(0, 0, ifst.Final(s), kNoStateId));
-        if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
-          FSTERROR() << "ArcMap: Non-zero arc labels for superfinal arc";
-          ofst->SetProperties(kError, kError);
-        }
-        ofst->SetFinal(s, final_arc.weight);
-        break;
-      }
-      case MAP_ALLOW_SUPERFINAL: {
-        B final_arc = (*mapper)(A(0, 0, ifst.Final(s), kNoStateId));
-        if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
-          // Add a superfinal state if not already done.
-          if (superfinal == kNoStateId) {
-            superfinal = ofst->AddState();
-            ofst->SetFinal(superfinal, B::Weight::One());
-          }
-          final_arc.nextstate = superfinal;
-          ofst->AddArc(s, final_arc);
-          ofst->SetFinal(s, B::Weight::Zero());
-        } else {
-          ofst->SetFinal(s, final_arc.weight);
-        }
-        break;
-      }
-      case MAP_REQUIRE_SUPERFINAL: {
-        B final_arc = (*mapper)(A(0, 0, ifst.Final(s), kNoStateId));
-        if (final_arc.ilabel != 0 || final_arc.olabel != 0 ||
-            final_arc.weight != B::Weight::Zero()) {
-          ofst->AddArc(s, B(final_arc.ilabel, final_arc.olabel,
-                            final_arc.weight, superfinal));
-        }
-        ofst->SetFinal(s, B::Weight::Zero());
-        break;
-      }
-    }
-  }
-  const auto oprops = ofst->Properties(kFstProperties, false);
-  ofst->SetProperties(mapper->Properties(iprops) | oprops, kFstProperties);
-}
-
-// Maps an arc type A to an arc type B using mapper function
-// object C, passed by value. This version writes the mapped input
-// Fst to an output MutableFst.
-template <class A, class B, class C>
-void ArcMap(const Fst<A> &ifst, MutableFst<B> *ofst, C mapper) {
-  ArcMap(ifst, ofst, &mapper);
-}
-
-struct ArcMapFstOptions : public CacheOptions {
-  // ArcMapFst default caching behaviour is to do no caching. Most mappers are
-  // cheap and therefore we save memory by not doing caching.
-  ArcMapFstOptions() : CacheOptions(true, 0) {}
-
-  explicit ArcMapFstOptions(const CacheOptions &opts) : CacheOptions(opts) {}
-};
-
-template <class A, class B, class C>
-class ArcMapFst;
-
-namespace internal {
-
-// Implementation of delayed ArcMapFst.
-template <class A, class B, class C>
-class ArcMapFstImpl : public CacheImpl<B> {
- public:
-  using Arc = B;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<B>::SetType;
-  using FstImpl<B>::SetProperties;
-  using FstImpl<B>::SetInputSymbols;
-  using FstImpl<B>::SetOutputSymbols;
-
-  using CacheImpl<B>::PushArc;
-  using CacheImpl<B>::HasArcs;
-  using CacheImpl<B>::HasFinal;
-  using CacheImpl<B>::HasStart;
-  using CacheImpl<B>::SetArcs;
-  using CacheImpl<B>::SetFinal;
-  using CacheImpl<B>::SetStart;
-
-  friend class StateIterator<ArcMapFst<A, B, C>>;
-
-  ArcMapFstImpl(const Fst<A> &fst, const C &mapper,
-                const ArcMapFstOptions &opts)
-      : CacheImpl<B>(opts),
-        fst_(fst.Copy()),
-        mapper_(new C(mapper)),
-        own_mapper_(true),
-        superfinal_(kNoStateId),
-        nstates_(0) {
-    Init();
-  }
-
-  ArcMapFstImpl(const Fst<A> &fst, C *mapper, const ArcMapFstOptions &opts)
-      : CacheImpl<B>(opts),
-        fst_(fst.Copy()),
-        mapper_(mapper),
-        own_mapper_(false),
-        superfinal_(kNoStateId),
-        nstates_(0) {
-    Init();
-  }
-
-  ArcMapFstImpl(const ArcMapFstImpl<A, B, C> &impl)
-      : CacheImpl<B>(impl),
-        fst_(impl.fst_->Copy(true)),
-        mapper_(new C(*impl.mapper_)),
-        own_mapper_(true),
-        superfinal_(kNoStateId),
-        nstates_(0) {
-    Init();
-  }
-
-  ~ArcMapFstImpl() override {
-    if (own_mapper_) delete mapper_;
-  }
-
-  StateId Start() {
-    if (!HasStart()) SetStart(FindOState(fst_->Start()));
-    return CacheImpl<B>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) {
-      switch (final_action_) {
-        case MAP_NO_SUPERFINAL:
-        default: {
-          const auto final_arc =
-              (*mapper_)(A(0, 0, fst_->Final(FindIState(s)), kNoStateId));
-          if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
-            FSTERROR() << "ArcMapFst: Non-zero arc labels for superfinal arc";
-            SetProperties(kError, kError);
-          }
-          SetFinal(s, final_arc.weight);
-          break;
-        }
-        case MAP_ALLOW_SUPERFINAL: {
-          if (s == superfinal_) {
-            SetFinal(s, Weight::One());
-          } else {
-            const auto final_arc =
-                (*mapper_)(A(0, 0, fst_->Final(FindIState(s)), kNoStateId));
-            if (final_arc.ilabel == 0 && final_arc.olabel == 0) {
-              SetFinal(s, final_arc.weight);
-            } else {
-              SetFinal(s, Weight::Zero());
-            }
-          }
-          break;
-        }
-        case MAP_REQUIRE_SUPERFINAL: {
-          SetFinal(s, s == superfinal_ ? Weight::One() : Weight::Zero());
-          break;
-        }
-      }
-    }
-    return CacheImpl<B>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<B>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<B>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<B>::NumOutputEpsilons(s);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && (fst_->Properties(kError, false) ||
-                            (mapper_->Properties(0) & kError))) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<B> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<B>::InitArcIterator(s, data);
-  }
-
-  void Expand(StateId s) {
-    // Add exiting arcs.
-    if (s == superfinal_) {
-      SetArcs(s);
-      return;
-    }
-    for (ArcIterator<Fst<A>> aiter(*fst_, FindIState(s)); !aiter.Done();
-         aiter.Next()) {
-      auto aarc = aiter.Value();
-      aarc.nextstate = FindOState(aarc.nextstate);
-      const auto &barc = (*mapper_)(aarc);
-      PushArc(s, barc);
-    }
-
-    // Check for superfinal arcs.
-    if (!HasFinal(s) || Final(s) == Weight::Zero()) {
-      switch (final_action_) {
-        case MAP_NO_SUPERFINAL:
-        default:
-          break;
-        case MAP_ALLOW_SUPERFINAL: {
-          auto final_arc =
-              (*mapper_)(A(0, 0, fst_->Final(FindIState(s)), kNoStateId));
-          if (final_arc.ilabel != 0 || final_arc.olabel != 0) {
-            if (superfinal_ == kNoStateId) superfinal_ = nstates_++;
-            final_arc.nextstate = superfinal_;
-            PushArc(s, final_arc);
-          }
-          break;
-        }
-        case MAP_REQUIRE_SUPERFINAL: {
-          const auto final_arc =
-              (*mapper_)(A(0, 0, fst_->Final(FindIState(s)), kNoStateId));
-          if (final_arc.ilabel != 0 || final_arc.olabel != 0 ||
-              final_arc.weight != B::Weight::Zero()) {
-            PushArc(s, B(final_arc.ilabel, final_arc.olabel, final_arc.weight,
-                         superfinal_));
-          }
-          break;
-        }
-      }
-    }
-    SetArcs(s);
-  }
-
- private:
-  void Init() {
-    SetType("map");
-    if (mapper_->InputSymbolsAction() == MAP_COPY_SYMBOLS) {
-      SetInputSymbols(fst_->InputSymbols());
-    } else if (mapper_->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-      SetInputSymbols(nullptr);
-    }
-    if (mapper_->OutputSymbolsAction() == MAP_COPY_SYMBOLS) {
-      SetOutputSymbols(fst_->OutputSymbols());
-    } else if (mapper_->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-      SetOutputSymbols(nullptr);
-    }
-    if (fst_->Start() == kNoStateId) {
-      final_action_ = MAP_NO_SUPERFINAL;
-      SetProperties(kNullProperties);
-    } else {
-      final_action_ = mapper_->FinalAction();
-      uint64 props = fst_->Properties(kCopyProperties, false);
-      SetProperties(mapper_->Properties(props));
-      if (final_action_ == MAP_REQUIRE_SUPERFINAL) superfinal_ = 0;
-    }
-  }
-
-  // Maps from output state to input state.
-  StateId FindIState(StateId s) {
-    if (superfinal_ == kNoStateId || s < superfinal_) {
-      return s;
-    } else {
-      return s - 1;
-    }
-  }
-
-  // Maps from input state to output state.
-  StateId FindOState(StateId is) {
-    auto os = is;
-    if (!(superfinal_ == kNoStateId || is < superfinal_)) ++os;
-    if (os >= nstates_) nstates_ = os + 1;
-    return os;
-  }
-
-  std::unique_ptr<const Fst<A>> fst_;
-  C *mapper_;
-  const bool own_mapper_;
-  MapFinalAction final_action_;
-  StateId superfinal_;
-  StateId nstates_;
-};
-
-}  // namespace internal
-
-// Maps an arc type A to an arc type B using Mapper function object
-// C. This version is a delayed FST.
-template <class A, class B, class C>
-class ArcMapFst : public ImplToFst<internal::ArcMapFstImpl<A, B, C>> {
- public:
-  using Arc = B;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<B>;
-  using State = typename Store::State;
-  using Impl = internal::ArcMapFstImpl<A, B, C>;
-
-  friend class ArcIterator<ArcMapFst<A, B, C>>;
-  friend class StateIterator<ArcMapFst<A, B, C>>;
-
-  ArcMapFst(const Fst<A> &fst, const C &mapper, const ArcMapFstOptions &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, mapper, opts)) {}
-
-  ArcMapFst(const Fst<A> &fst, C *mapper, const ArcMapFstOptions &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, mapper, opts)) {}
-
-  ArcMapFst(const Fst<A> &fst, const C &mapper)
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, mapper, ArcMapFstOptions())) {}
-
-  ArcMapFst(const Fst<A> &fst, C *mapper)
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, mapper, ArcMapFstOptions())) {}
-
-  // See Fst<>::Copy() for doc.
-  ArcMapFst(const ArcMapFst<A, B, C> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this ArcMapFst. See Fst<>::Copy() for further doc.
-  ArcMapFst<A, B, C> *Copy(bool safe = false) const override {
-    return new ArcMapFst<A, B, C>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<B> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<B> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
- protected:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
- private:
-  ArcMapFst &operator=(const ArcMapFst &) = delete;
-};
-
-// Specialization for ArcMapFst.
-//
-// This may be derived from.
-template <class A, class B, class C>
-class StateIterator<ArcMapFst<A, B, C>> : public StateIteratorBase<B> {
- public:
-  using StateId = typename B::StateId;
-
-  explicit StateIterator(const ArcMapFst<A, B, C> &fst)
-      : impl_(fst.GetImpl()),
-        siter_(*impl_->fst_),
-        s_(0),
-        superfinal_(impl_->final_action_ == MAP_REQUIRE_SUPERFINAL) {
-    CheckSuperfinal();
-  }
-
-  bool Done() const final { return siter_.Done() && !superfinal_; }
-
-  StateId Value() const final { return s_; }
-
-  void Next() final {
-    ++s_;
-    if (!siter_.Done()) {
-      siter_.Next();
-      CheckSuperfinal();
-    } else if (superfinal_) {
-      superfinal_ = false;
-    }
-  }
-
-  void Reset() final {
-    s_ = 0;
-    siter_.Reset();
-    superfinal_ = impl_->final_action_ == MAP_REQUIRE_SUPERFINAL;
-    CheckSuperfinal();
-  }
-
- private:
-  void CheckSuperfinal() {
-    if (impl_->final_action_ != MAP_ALLOW_SUPERFINAL || superfinal_) return;
-    if (!siter_.Done()) {
-      const auto final_arc =
-          (*impl_->mapper_)(A(0, 0, impl_->fst_->Final(s_), kNoStateId));
-      if (final_arc.ilabel != 0 || final_arc.olabel != 0) superfinal_ = true;
-    }
-  }
-
-  const internal::ArcMapFstImpl<A, B, C> *impl_;
-  StateIterator<Fst<A>> siter_;
-  StateId s_;
-  bool superfinal_;  // True if there is a superfinal state and not done.
-};
-
-// Specialization for ArcMapFst.
-template <class A, class B, class C>
-class ArcIterator<ArcMapFst<A, B, C>>
-    : public CacheArcIterator<ArcMapFst<A, B, C>> {
- public:
-  using StateId = typename A::StateId;
-
-  ArcIterator(const ArcMapFst<A, B, C> &fst, StateId s)
-      : CacheArcIterator<ArcMapFst<A, B, C>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class A, class B, class C>
-inline void ArcMapFst<A, B, C>::InitStateIterator(
-    StateIteratorData<B> *data) const {
-  data->base = new StateIterator<ArcMapFst<A, B, C>>(*this);
-}
-
-// Utility Mappers.
-
-// Mapper that returns its input.
-template <class A>
-class IdentityArcMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-
-  ToArc operator()(const FromArc &arc) const { return arc; }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const { return props; }
-};
-
-// Mapper that converts all input symbols to epsilon.
-template <class A>
-class InputEpsilonMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-
-  ToArc operator()(const FromArc &arc) const {
-    return ToArc(0, arc.olabel, arc.weight, arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return (props & kSetArcProperties) | kIEpsilons;
-  }
-};
-
-// Mapper that converts all output symbols to epsilon.
-template <class A>
-class OutputEpsilonMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-
-  ToArc operator()(const FromArc &arc) const {
-    return ToArc(arc.ilabel, 0, arc.weight, arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return (props & kSetArcProperties) | kOEpsilons;
-  }
-};
-
-// Mapper that returns its input with final states redirected to a single
-// super-final state.
-template <class A>
-class SuperFinalMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-  using Label = typename FromArc::Label;
-  using Weight = typename FromArc::Weight;;
-
-  // Arg allows setting super-final label.
-  explicit SuperFinalMapper(Label final_label = 0)
-      : final_label_(final_label) {}
-
-  ToArc operator()(const FromArc &arc) const {
-    // Super-final arc.
-    if (arc.nextstate == kNoStateId && arc.weight != Weight::Zero()) {
-      return ToArc(final_label_, final_label_, arc.weight, kNoStateId);
-    } else {
-      return arc;
-    }
-  }
-
-  constexpr MapFinalAction FinalAction() const {
-    return MAP_REQUIRE_SUPERFINAL;
-  }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    if (final_label_ == 0) {
-      return props & kAddSuperFinalProperties;
-    } else {
-      return props & kAddSuperFinalProperties &
-          kILabelInvariantProperties & kOLabelInvariantProperties;
-    }
-  }
-
- private:
-  Label final_label_;
-};
-
-// Mapper that leaves labels and nextstate unchanged and constructs a new weight
-// from the underlying value of the arc weight. If no weight converter is
-// explictly specified, requires that there is a WeightConvert class
-// specialization that converts the weights.
-template <class A, class B,
-          class C = WeightConvert<typename A::Weight, typename B::Weight>>
-class WeightConvertMapper {
- public:
-  using FromArc = A;
-  using ToArc = B;
-  using Converter = C;
-  using FromWeight = typename FromArc::Weight;
-  using ToWeight = typename ToArc::Weight;
-
-  explicit WeightConvertMapper(const Converter &c = Converter())
-      : convert_weight_(c) {}
-
-  ToArc operator()(const FromArc &arc) const {
-    return ToArc(arc.ilabel, arc.olabel, convert_weight_(arc.weight),
-                 arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const { return props; }
-
- private:
-  Converter convert_weight_;
-};
-
-// Non-precision-changing weight conversions; consider using more efficient
-// Cast method instead.
-
-using StdToLogMapper = WeightConvertMapper<StdArc, LogArc>;
-
-using LogToStdMapper = WeightConvertMapper<LogArc, StdArc>;
-
-// Precision-changing weight conversions.
-
-using StdToLog64Mapper = WeightConvertMapper<StdArc, Log64Arc>;
-
-using LogToLog64Mapper = WeightConvertMapper<LogArc, Log64Arc>;
-
-using Log64ToStdMapper = WeightConvertMapper<Log64Arc, StdArc>;
-
-using Log64ToLogMapper = WeightConvertMapper<Log64Arc, LogArc>;
-
-// Mapper from A to GallicArc<A>.
-template <class A, GallicType G = GALLIC_LEFT>
-class ToGallicMapper {
- public:
-  using FromArc = A;
-  using ToArc = GallicArc<A, G>;
-
-  using SW = StringWeight<typename A::Label, GallicStringType(G)>;
-  using AW = typename FromArc::Weight;
-  using GW = typename ToArc::Weight;
-
-  ToArc operator()(const FromArc &arc) const {
-    // Super-final arc.
-    if (arc.nextstate == kNoStateId && arc.weight != AW::Zero()) {
-      return ToArc(0, 0, GW(SW::One(), arc.weight), kNoStateId);
-    // Super-non-final arc.
-    } else if (arc.nextstate == kNoStateId) {
-      return ToArc(0, 0, GW::Zero(), kNoStateId);
-    // Epsilon label.
-    } else if (arc.olabel == 0) {
-      return ToArc(arc.ilabel, arc.ilabel, GW(SW::One(), arc.weight),
-                   arc.nextstate);
-    // Regular label.
-    } else {
-      return ToArc(arc.ilabel, arc.ilabel, GW(SW(arc.olabel), arc.weight),
-                   arc.nextstate);
-    }
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return ProjectProperties(props, true) & kWeightInvariantProperties;
-  }
-};
-
-// Mapper from GallicArc<A> to A.
-template <class A, GallicType G = GALLIC_LEFT>
-class FromGallicMapper {
- public:
-  using FromArc = GallicArc<A, G>;
-  using ToArc = A;
-
-  using Label = typename ToArc::Label;
-  using AW = typename ToArc::Weight;
-  using GW = typename FromArc::Weight;
-
-  explicit FromGallicMapper(Label superfinal_label = 0)
-      : superfinal_label_(superfinal_label), error_(false) {}
-
-  ToArc operator()(const FromArc &arc) const {
-    // 'Super-non-final' arc.
-    if (arc.nextstate == kNoStateId && arc.weight == GW::Zero()) {
-      return A(arc.ilabel, 0, AW::Zero(), kNoStateId);
-    }
-    Label l = kNoLabel;
-    AW weight;
-    if (!Extract(arc.weight, &weight, &l) || arc.ilabel != arc.olabel) {
-      FSTERROR() << "FromGallicMapper: Unrepresentable weight: " << arc.weight
-                 << " for arc with ilabel = " << arc.ilabel
-                 << ", olabel = " << arc.olabel
-                 << ", nextstate = " << arc.nextstate;
-      error_ = true;
-    }
-    if (arc.ilabel == 0 && l != 0 && arc.nextstate == kNoStateId) {
-      return ToArc(superfinal_label_, l, weight, arc.nextstate);
-    } else {
-      return ToArc(arc.ilabel, l, weight, arc.nextstate);
-    }
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_ALLOW_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 inprops) const {
-    uint64 outprops = inprops & kOLabelInvariantProperties &
-                      kWeightInvariantProperties & kAddSuperFinalProperties;
-    if (error_) outprops |= kError;
-    return outprops;
-  }
-
- private:
-  template <GallicType GT>
-  static bool Extract(const GallicWeight<Label, AW, GT> &gallic_weight,
-                      typename A::Weight *weight, typename A::Label *label) {
-    using GW = StringWeight<Label, GallicStringType(GT)>;
-    const GW &w1 = gallic_weight.Value1();
-    const AW &w2 = gallic_weight.Value2();
-    typename GW::Iterator iter1(w1);
-    const Label l = w1.Size() == 1 ? iter1.Value() : 0;
-    if (l == kStringInfinity || l == kStringBad || w1.Size() > 1) return false;
-    *label = l;
-    *weight = w2;
-    return true;
-  }
-
-  static bool Extract(const GallicWeight<Label, AW, GALLIC> &gallic_weight,
-                      typename A::Weight *weight, typename A::Label *label) {
-    if (gallic_weight.Size() > 1) return false;
-    if (gallic_weight.Size() == 0) {
-      *label = 0;
-      *weight = A::Weight::Zero();
-      return true;
-    }
-    return Extract<GALLIC_RESTRICT>(gallic_weight.Back(), weight, label);
-  }
-
-  const Label superfinal_label_;
-  mutable bool error_;
-};
-
-// Mapper from GallicArc<A> to A.
-template <class A, GallicType G = GALLIC_LEFT>
-class GallicToNewSymbolsMapper {
- public:
-  using FromArc = GallicArc<A, G>;
-  using ToArc = A;
-
-  using Label = typename ToArc::Label;
-  using StateId = typename ToArc::StateId;
-  using AW = typename ToArc::Weight;
-  using GW = typename FromArc::Weight;
-  using SW = StringWeight<Label, GallicStringType(G)>;
-
-  explicit GallicToNewSymbolsMapper(MutableFst<ToArc> *fst)
-      : fst_(fst),
-        lmax_(0),
-        osymbols_(fst->OutputSymbols()),
-        isymbols_(nullptr),
-        error_(false) {
-    fst_->DeleteStates();
-    state_ = fst_->AddState();
-    fst_->SetStart(state_);
-    fst_->SetFinal(state_, AW::One());
-    if (osymbols_) {
-      string name = osymbols_->Name() + "_from_gallic";
-      fst_->SetInputSymbols(new SymbolTable(name));
-      isymbols_ = fst_->MutableInputSymbols();
-      const int64 zero = 0;
-      isymbols_->AddSymbol(osymbols_->Find(zero), 0);
-    } else {
-      fst_->SetInputSymbols(nullptr);
-    }
-  }
-
-  ToArc operator()(const FromArc &arc) {
-    // Super-non-final arc.
-    if (arc.nextstate == kNoStateId && arc.weight == GW::Zero()) {
-      return ToArc(arc.ilabel, 0, AW::Zero(), kNoStateId);
-    }
-    SW w1 = arc.weight.Value1();
-    AW w2 = arc.weight.Value2();
-    Label l;
-    if (w1.Size() == 0) {
-      l = 0;
-    } else {
-      auto insert_result = map_.insert(std::make_pair(w1, kNoLabel));
-      if (!insert_result.second) {
-        l = insert_result.first->second;
-      } else {
-        l = ++lmax_;
-        insert_result.first->second = l;
-        StringWeightIterator<SW> iter1(w1);
-        StateId n;
-        string s;
-        for (size_t i = 0, p = state_; i < w1.Size();
-             ++i, iter1.Next(), p = n) {
-          n = i == w1.Size() - 1 ? state_ : fst_->AddState();
-          fst_->AddArc(p, ToArc(i ? 0 : l, iter1.Value(), AW::One(), n));
-          if (isymbols_) {
-            if (i) s = s + "_";
-            s = s + osymbols_->Find(iter1.Value());
-          }
-        }
-        if (isymbols_) isymbols_->AddSymbol(s, l);
-      }
-    }
-    if (l == kStringInfinity || l == kStringBad || arc.ilabel != arc.olabel) {
-      FSTERROR() << "GallicToNewSymbolMapper: Unrepresentable weight: " << l;
-      error_ = true;
-    }
-    return ToArc(arc.ilabel, l, w2, arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_ALLOW_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 inprops) const {
-    uint64 outprops = inprops & kOLabelInvariantProperties &
-                      kWeightInvariantProperties & kAddSuperFinalProperties;
-    if (error_) outprops |= kError;
-    return outprops;
-  }
-
- private:
-  class StringKey {
-   public:
-    size_t operator()(const SW &x) const { return x.Hash(); }
-  };
-
-  using Map = std::unordered_map<SW, Label, StringKey>;
-
-  MutableFst<ToArc> *fst_;
-  Map map_;
-  Label lmax_;
-  StateId state_;
-  const SymbolTable *osymbols_;
-  SymbolTable *isymbols_;
-  mutable bool error_;
-};
-
-// Mapper to add a constant to all weights.
-template <class A>
-class PlusMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-  using Weight = typename FromArc::Weight;
-
-  explicit PlusMapper(Weight weight) : weight_(std::move(weight)) {}
-
-  ToArc operator()(const FromArc &arc) const {
-    if (arc.weight == Weight::Zero()) return arc;
-    return ToArc(arc.ilabel, arc.olabel, Plus(arc.weight, weight_),
-                 arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return props & kWeightInvariantProperties;
-  }
-
- private:
-  const Weight weight_;
-};
-
-// Mapper to (right) multiply a constant to all weights.
-template <class A>
-class TimesMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-  using Weight = typename FromArc::Weight;
-
-  explicit TimesMapper(Weight weight) : weight_(std::move(weight)) {}
-
-  ToArc operator()(const FromArc &arc) const {
-    if (arc.weight == Weight::Zero()) return arc;
-    return ToArc(arc.ilabel, arc.olabel, Times(arc.weight, weight_),
-                 arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return props & kWeightInvariantProperties;
-  }
-
- private:
-  const Weight weight_;
-};
-
-// Mapper to take all arc-weights to a fixed power.
-template <class A>
-class PowerMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-  using Weight = typename FromArc::Weight;
-
-  explicit PowerMapper(size_t power) : power_(power) {}
-
-  ToArc operator()(const FromArc &arc) const {
-    return ToArc(arc.ilabel, arc.olabel, Power(arc.weight, power_),
-                 arc.nextstate);
-  }
-
-  MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  MapSymbolsAction InputSymbolsAction() const { return MAP_COPY_SYMBOLS; }
-
-  MapSymbolsAction OutputSymbolsAction() const { return MAP_COPY_SYMBOLS; }
-
-  uint64 Properties(uint64 props) const {
-    return props & kWeightInvariantProperties;
-  }
-
- private:
-  size_t power_;
-};
-
-// Mapper to reciprocate all non-Zero() weights.
-template <class A>
-class InvertWeightMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-  using Weight = typename FromArc::Weight;
-
-  ToArc operator()(const FromArc &arc) const {
-    if (arc.weight == Weight::Zero()) return arc;
-    return ToArc(arc.ilabel, arc.olabel, Divide(Weight::One(), arc.weight),
-                 arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return props & kWeightInvariantProperties;
-  }
-};
-
-// Mapper to map all non-Zero() weights to One().
-template <class A, class B = A>
-class RmWeightMapper {
- public:
-  using FromArc = A;
-  using ToArc = B;
-  using FromWeight = typename FromArc::Weight;
-  using ToWeight = typename ToArc::Weight;
-
-  ToArc operator()(const FromArc &arc) const {
-    return ToArc(arc.ilabel, arc.olabel,
-                 arc.weight != FromWeight::Zero() ?
-                 ToWeight::One() : ToWeight::Zero(),
-                 arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return (props & kWeightInvariantProperties) | kUnweighted;
-  }
-};
-
-// Mapper to quantize all weights.
-template <class A, class B = A>
-class QuantizeMapper {
- public:
-  using FromArc = A;
-  using ToArc = B;
-  using FromWeight = typename FromArc::Weight;
-  using ToWeight = typename ToArc::Weight;
-
-  QuantizeMapper() : delta_(kDelta) {}
-
-  explicit QuantizeMapper(float d) : delta_(d) {}
-
-  ToArc operator()(const FromArc &arc) const {
-    return ToArc(arc.ilabel, arc.olabel, arc.weight.Quantize(delta_),
-                 arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return props & kWeightInvariantProperties;
-  }
-
- private:
-  const float delta_;
-};
-
-// Mapper from A to B under the assumption:
-//
-//    B::Weight = A::Weight::ReverseWeight
-//    B::Label == A::Label
-//    B::StateId == A::StateId
-//
-// The weight is reversed, while the label and nextstate are preserved.
-template <class A, class B>
-class ReverseWeightMapper {
- public:
-  using FromArc = A;
-  using ToArc = B;
-
-  ToArc operator()(const FromArc &arc) const {
-    return ToArc(arc.ilabel, arc.olabel, arc.weight.Reverse(), arc.nextstate);
-  }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const { return props; }
-};
-
-}  // namespace fst
-
-#endif  // FST_ARC_MAP_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arc.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arc.h
deleted file mode 100644
index 7bd0c47a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arc.h
+++ /dev/null
@@ -1,310 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Commonly used FST arc types.
-
-#ifndef FST_ARC_H_
-#define FST_ARC_H_
-
-#include <climits>
-#include <string>
-#include <utility>
-
-
-#include <fst/expectation-weight.h>
-#include <fst/float-weight.h>
-#include <fst/lexicographic-weight.h>
-#include <fst/power-weight.h>
-#include <fst/product-weight.h>
-#include <fst/signed-log-weight.h>
-#include <fst/sparse-power-weight.h>
-#include <fst/string-weight.h>
-
-
-namespace fst {
-
-template <class W>
-struct ArcTpl {
- public:
-  using Weight = W;
-  using Label = int;
-  using StateId = int;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  ArcTpl() {}
-
-  ArcTpl(Label ilabel, Label olabel, Weight weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::move(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const string *const type =
-        new string(Weight::Type() == "tropical" ? "standard" : Weight::Type());
-    return *type;
-  }
-};
-
-using StdArc = ArcTpl<TropicalWeight>;
-using LogArc = ArcTpl<LogWeight>;
-using Log64Arc = ArcTpl<Log64Weight>;
-using SignedLogArc = ArcTpl<SignedLogWeight>;
-using SignedLog64Arc = ArcTpl<SignedLog64Weight>;
-using MinMaxArc = ArcTpl<MinMaxWeight>;
-
-// Arc with integer labels and state IDs and string weights.
-template <StringType S = STRING_LEFT>
-struct StringArc {
- public:
-  using Label = int;
-  using Weight = StringWeight<int, S>;
-  using StateId = int;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  StringArc() = default;
-
-  StringArc(Label ilabel, Label olabel, Weight weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::move(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const string *const type =
-        new string(S == STRING_LEFT
-                       ? "left_standard_string"
-                       : (S == STRING_RIGHT ? "right_standard_string"
-                                            : "restricted_standard_string"));
-    return *type;
-  }
-};
-
-// Arc with label and state Id type the same as template arg and with
-// weights over the Gallic semiring w.r.t the output labels and weights of A.
-template <class A, GallicType G = GALLIC_LEFT>
-struct GallicArc {
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = GallicWeight<Label, typename Arc::Weight, G>;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  GallicArc() = default;
-
-  GallicArc(Label ilabel, Label olabel, Weight weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::move(weight)),
-        nextstate(nextstate) {}
-
-  explicit GallicArc(const Arc &arc)
-      : ilabel(arc.ilabel), olabel(arc.ilabel), weight(arc.olabel, arc.weight),
-        nextstate(arc.nextstate) {}
-
-  static const string &Type() {
-    static const string *const type =
-        new string(
-            (G == GALLIC_LEFT
-                 ? "left_gallic_"
-                 : (G == GALLIC_RIGHT
-                        ? "right_gallic_"
-                        : (G == GALLIC_RESTRICT
-                               ? "restricted_gallic_"
-                               : (G == GALLIC_MIN
-                                      ? "min_gallic_" : "gallic_")))) +
-            Arc::Type());
-    return *type;
-  }
-};
-
-// Arc with the reverse of the weight found in its template arg.
-template <class A>
-struct ReverseArc {
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using AWeight = typename Arc::Weight;
-  using Weight = typename AWeight::ReverseWeight;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  ReverseArc() = default;
-
-  ReverseArc(Label ilabel, Label olabel, Weight weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::move(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const string *const type = new string("reverse_" + Arc::Type());
-    return *type;
-  }
-};
-
-// Arc with integer labels and state IDs and lexicographic weights.
-template <class Weight1, class Weight2>
-struct LexicographicArc {
-  using Label = int;
-  using StateId = int;
-  using Weight = LexicographicWeight<Weight1, Weight2>;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  LexicographicArc() = default;
-
-  LexicographicArc(Label ilabel, Label olabel, Weight weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::move(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const string *const type = new string(Weight::Type());
-    return *type;
-  }
-};
-
-// Arc with integer labels and state IDs and product weights.
-template <class Weight1, class Weight2>
-struct ProductArc {
-  using Label = int;
-  using StateId = int;
-  using Weight = ProductWeight<Weight1, Weight2>;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  ProductArc() = default;
-
-  ProductArc(Label ilabel, Label olabel, Weight weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::move(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const string *const type = new string(Weight::Type());
-    return *type;
-  }
-};
-
-// Arc with label and state ID type the same as first template argument and with
-// weights over the n-th Cartesian power of the weight type of the template
-// argument.
-template <class A, unsigned int N>
-struct PowerArc {
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = PowerWeight<typename Arc::Weight, N>;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  PowerArc() = default;
-
-  PowerArc(Label ilabel, Label olabel, Weight weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::move(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const string *const type =
-        new string(Arc::Type() + "_^" + std::to_string(N));
-    return *type;
-  }
-};
-
-// Arc with label and state ID type the same as first template argument and with
-// weights over the arbitrary Cartesian power of the weight type.
-template <class A, class K = int>
-struct SparsePowerArc {
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::Label;
-  using Weight = SparsePowerWeight<typename Arc::Weight, K>;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  SparsePowerArc() = default;
-
-  SparsePowerArc(Label ilabel, Label olabel, Weight weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::move(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const string *const type = [] {
-      string type = Arc::Type() + "_^n";
-      if (sizeof(K) != sizeof(uint32)) {
-        type += "_" + std::to_string(CHAR_BIT * sizeof(K));
-      }
-      return new string(type);
-    }();
-    return *type;
-  }
-};
-
-// Arc with label and state ID type the same as first template argument and with
-// expectation weight over the first template argument's weight type and the
-// second template argument.
-template <class A, class X2>
-struct ExpectationArc {
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using X1 = typename Arc::Weight;
-  using Weight = ExpectationWeight<X1, X2>;
-
-  Label ilabel;
-  Label olabel;
-  Weight weight;
-  StateId nextstate;
-
-  ExpectationArc() = default;
-
-  ExpectationArc(Label ilabel, Label olabel, Weight weight, StateId nextstate)
-      : ilabel(ilabel),
-        olabel(olabel),
-        weight(std::move(weight)),
-        nextstate(nextstate) {}
-
-  static const string &Type() {
-    static const string *const type =
-        new string("expectation_" + Arc::Type() + "_" + X2::Type());
-    return *type;
-  }
-};
-
-}  // namespace fst
-
-#endif  // FST_ARC_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arcfilter.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arcfilter.h
deleted file mode 100644
index 598e543f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arcfilter.h
+++ /dev/null
@@ -1,93 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function objects to restrict which arcs are traversed in an FST.
-
-#ifndef FST_ARCFILTER_H_
-#define FST_ARCFILTER_H_
-
-
-#include <fst/fst.h>
-#include <fst/util.h>
-
-
-namespace fst {
-
-// True for all arcs.
-template <class Arc>
-class AnyArcFilter {
- public:
-  bool operator()(const Arc &arc) const { return true; }
-};
-
-// True for (input/output) epsilon arcs.
-template <class Arc>
-class EpsilonArcFilter {
- public:
-  bool operator()(const Arc &arc) const {
-    return arc.ilabel == 0 && arc.olabel == 0;
-  }
-};
-
-// True for input epsilon arcs.
-template <class Arc>
-class InputEpsilonArcFilter {
- public:
-  bool operator()(const Arc &arc) const { return arc.ilabel == 0; }
-};
-
-// True for output epsilon arcs.
-template <class Arc>
-class OutputEpsilonArcFilter {
- public:
-  bool operator()(const Arc &arc) const { return arc.olabel == 0; }
-};
-
-// True if specified label matches (doesn't match) when keep_match is
-// true (false).
-template <class Arc>
-class LabelArcFilter {
- public:
-  using Label = typename Arc::Label;
-
-  explicit LabelArcFilter(Label label, bool match_input = true,
-                          bool keep_match = true)
-      : label_(label), match_input_(match_input), keep_match_(keep_match) {}
-
-  bool operator()(const Arc &arc) const {
-    const bool match = (match_input_ ? arc.ilabel : arc.olabel) == label_;
-    return keep_match_ ? match : !match;
-  }
-
- private:
-  const Label label_;
-  const bool match_input_;
-  const bool keep_match_;
-};
-
-// True if specified labels match (don't match) when keep_match is true (false).
-template <class Arc>
-class MultiLabelArcFilter {
- public:
-  using Label = typename Arc::Label;
-
-  explicit MultiLabelArcFilter(bool match_input = true, bool keep_match = true)
-      : match_input_(match_input), keep_match_(keep_match) {}
-
-  bool operator()(const Arc &arc) const {
-    const Label label = match_input_ ? arc.ilabel : arc.olabel;
-    const bool match = labels_.Find(label) != labels_.End();
-    return keep_match_ ? match : !match;
-  }
-
-  void AddLabel(Label label) { labels_.Insert(label); }
-
- private:
-  CompactSet<Label, kNoLabel> labels_;
-  const bool match_input_;
-  const bool keep_match_;
-};
-
-}  // namespace fst
-
-#endif  // FST_ARCFILTER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arcsort.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arcsort.h
deleted file mode 100644
index b159757a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/arcsort.h
+++ /dev/null
@@ -1,211 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to sort arcs in an FST.
-
-#ifndef FST_ARCSORT_H_
-#define FST_ARCSORT_H_
-
-#include <algorithm>
-#include <string>
-#include <vector>
-
-#include <fst/cache.h>
-#include <fst/state-map.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-template <class Arc, class Compare>
-class ArcSortMapper {
- public:
-  using FromArc = Arc;
-  using ToArc = Arc;
-
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  ArcSortMapper(const Fst<Arc> &fst, const Compare &comp)
-      : fst_(fst), comp_(comp), i_(0) {}
-
-  // Allows updating Fst argument; pass only if changed.
-  ArcSortMapper(const ArcSortMapper<Arc, Compare> &mapper,
-                const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? *fst : mapper.fst_), comp_(mapper.comp_), i_(0) {}
-
-  StateId Start() { return fst_.Start(); }
-
-  Weight Final(StateId s) const { return fst_.Final(s); }
-
-  void SetState(StateId s) {
-    i_ = 0;
-    arcs_.clear();
-    arcs_.reserve(fst_.NumArcs(s));
-    for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
-      arcs_.push_back(aiter.Value());
-    }
-    std::sort(arcs_.begin(), arcs_.end(), comp_);
-  }
-
-  bool Done() const { return i_ >= arcs_.size(); }
-
-  const Arc &Value() const { return arcs_[i_]; }
-
-  void Next() { ++i_; }
-
-  MapSymbolsAction InputSymbolsAction() const { return MAP_COPY_SYMBOLS; }
-
-  MapSymbolsAction OutputSymbolsAction() const { return MAP_COPY_SYMBOLS; }
-
-  uint64 Properties(uint64 props) const { return comp_.Properties(props); }
-
- private:
-  const Fst<Arc> &fst_;
-  const Compare &comp_;
-  std::vector<Arc> arcs_;
-  ssize_t i_;  // current arc position
-
-  ArcSortMapper &operator=(const ArcSortMapper &) = delete;
-};
-
-// Sorts the arcs in an FST according to function object 'comp' of type Compare.
-// This version modifies its input. Comparison function objects ILabelCompare
-// and OLabelCompare are provided by the library. In general, Compare must meet
-// the requirements for a  comparison function object (e.g., similar to those
-// used by std::sort). It must also have a member Properties(uint64) that
-// specifies the known properties of the sorted FST; it takes as argument the
-// input FST's known properties before the sort.
-//
-// Complexity:
-//
-// - Time: O(v d log d)
-// - Space: O(d)
-//
-// where v = # of states and d = maximum out-degree.
-template <class Arc, class Compare>
-void ArcSort(MutableFst<Arc> *fst, Compare comp) {
-  ArcSortMapper<Arc, Compare> mapper(*fst, comp);
-  StateMap(fst, mapper);
-}
-
-using ArcSortFstOptions = CacheOptions;
-
-// Sorts the arcs in an FST according to function object 'comp' of type Compare.
-// This version is a delayed FST. Comparsion function objects ILabelCompare and
-// OLabelCompare are provided by the library. In general, Compare must meet the
-// requirements for a comparision function object (e.g., similar to those
-// used by std::sort). It must also have a member Properties(uint64) that
-// specifies the known properties of the sorted FST; it takes as argument the
-// input FST's known properties.
-//
-// Complexity:
-//
-// - Time: O(v d log d)
-// - Space: O(d)
-//
-// where v = # of states visited, d = maximum out-degree of states visited.
-// Constant time and space to visit an input state is assumed and exclusive of
-// caching.
-template <class Arc, class Compare>
-class ArcSortFst : public StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>> {
-  using StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>>::GetImpl;
-
- public:
-  using StateId = typename Arc::StateId;
-  using Mapper = ArcSortMapper<Arc, Compare>;
-
-  ArcSortFst(const Fst<Arc> &fst, const Compare &comp)
-      : StateMapFst<Arc, Arc, Mapper>(fst,
-                                      ArcSortMapper<Arc, Compare>(fst, comp)) {}
-
-  ArcSortFst(const Fst<Arc> &fst, const Compare &comp,
-             const ArcSortFstOptions &opts)
-      : StateMapFst<Arc, Arc, Mapper>(fst, Mapper(fst, comp), opts) {}
-
-  // See Fst<>::Copy() for doc.
-  ArcSortFst(const ArcSortFst<Arc, Compare> &fst, bool safe = false)
-      : StateMapFst<Arc, Arc, Mapper>(fst, safe) {}
-
-  // Gets a copy of this ArcSortFst. See Fst<>::Copy() for further doc.
-  ArcSortFst<Arc, Compare> *Copy(bool safe = false) const override {
-    return new ArcSortFst(*this, safe);
-  }
-
-  size_t NumArcs(StateId s) const override {
-    return GetImpl()->GetFst()->NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) const override {
-    return GetImpl()->GetFst()->NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) const override {
-    return GetImpl()->GetFst()->NumOutputEpsilons(s);
-  }
-};
-
-// Specialization for ArcSortFst.
-template <class Arc, class Compare>
-class StateIterator<ArcSortFst<Arc, Compare>>
-    : public StateIterator<StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>>> {
- public:
-  explicit StateIterator(const ArcSortFst<Arc, Compare> &fst)
-      : StateIterator<StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>>>(fst) {
-  }
-};
-
-// Specialization for ArcSortFst.
-template <class Arc, class Compare>
-class ArcIterator<ArcSortFst<Arc, Compare>>
-    : public ArcIterator<StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>>> {
- public:
-  ArcIterator(const ArcSortFst<Arc, Compare> &fst, typename Arc::StateId s)
-      : ArcIterator<StateMapFst<Arc, Arc, ArcSortMapper<Arc, Compare>>>(fst,
-                                                                        s) {}
-};
-
-// Compare class for comparing input labels of arcs.
-template <class Arc>
-class ILabelCompare {
- public:
-  ILabelCompare() {}
-
-  bool operator()(const Arc &arc1, const Arc &arc2) const {
-    return arc1.ilabel < arc2.ilabel;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return (props & kArcSortProperties) | kILabelSorted |
-           (props & kAcceptor ? kOLabelSorted : 0);
-  }
-};
-
-// Compare class for comparing output labels of arcs.
-template <class Arc>
-class OLabelCompare {
- public:
-  OLabelCompare() {}
-
-  bool operator()(const Arc &arc1, const Arc &arc2) const {
-    return arc1.olabel < arc2.olabel;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return (props & kArcSortProperties) | kOLabelSorted |
-           (props & kAcceptor ? kILabelSorted : 0);
-  }
-};
-
-// Useful aliases when using StdArc.
-
-template <class Compare>
-using StdArcSortFst = ArcSortFst<StdArc, Compare>;
-
-using StdILabelCompare = ILabelCompare<StdArc>;
-
-using StdOLabelCompare = OLabelCompare<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_ARCSORT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/bi-table.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/bi-table.h
deleted file mode 100644
index 101ea922..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/bi-table.h
+++ /dev/null
@@ -1,473 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes for representing a bijective mapping between an arbitrary entry
-// of type T and a signed integral ID.
-
-#ifndef FST_BI_TABLE_H_
-#define FST_BI_TABLE_H_
-
-#include <deque>
-#include <memory>
-#include <functional>
-#include <unordered_map>
-#include <unordered_set>
-#include <vector>
-
-#include <fst/log.h>
-#include <fst/memory.h>
-
-namespace fst {
-
-// Bitables model bijective mappings between entries of an arbitrary type T and
-// an signed integral ID of type I. The IDs are allocated starting from 0 in
-// order.
-//
-// template <class I, class T>
-// class BiTable {
-//  public:
-//
-//   // Required constructors.
-//   BiTable();
-//
-//   // Looks up integer ID from entry. If it doesn't exist and insert
-//   / is true, adds it; otherwise, returns -1.
-//   I FindId(const T &entry, bool insert = true);
-//
-//   // Looks up entry from integer ID.
-//   const T &FindEntry(I) const;
-//
-//   // Returns number of stored entries.
-//   I Size() const;
-// };
-
-// An implementation using a hash map for the entry to ID mapping. H is the
-// hash function and E is the equality function. If passed to the constructor,
-// ownership is given to this class.
-template <class I, class T, class H, class E = std::equal_to<T>>
-class HashBiTable {
- public:
-  // Reserves space for table_size elements. If passing H and E to the
-  // constructor, this class owns them.
-  explicit HashBiTable(size_t table_size = 0, H *h = nullptr, E *e = nullptr) :
-      hash_func_(h ? h : new H()), hash_equal_(e ? e : new E()),
-      entry2id_(table_size, *hash_func_, *hash_equal_) {
-    if (table_size) id2entry_.reserve(table_size);
-  }
-
-  HashBiTable(const HashBiTable<I, T, H, E> &table)
-      : hash_func_(new H(*table.hash_func_)),
-        hash_equal_(new E(*table.hash_equal_)),
-        entry2id_(table.entry2id_.begin(), table.entry2id_.end(),
-                  table.entry2id_.size(), *hash_func_, *hash_equal_),
-        id2entry_(table.id2entry_) {}
-
-  I FindId(const T &entry, bool insert = true) {
-    if (!insert) {
-      const auto it = entry2id_.find(entry);
-      return it == entry2id_.end() ? -1 : it->second - 1;
-    }
-    I &id_ref = entry2id_[entry];
-    if (id_ref == 0) {  // T not found; stores and assigns a new ID.
-      id2entry_.push_back(entry);
-      id_ref = id2entry_.size();
-    }
-    return id_ref - 1;  // NB: id_ref = ID + 1.
-  }
-
-  const T &FindEntry(I s) const { return id2entry_[s]; }
-
-  I Size() const { return id2entry_.size(); }
-
-  // TODO(riley): Add fancy clear-to-size, as in CompactHashBiTable.
-  void Clear() {
-    entry2id_.clear();
-    id2entry_.clear();
-  }
-
- private:
-  std::unique_ptr<H> hash_func_;
-  std::unique_ptr<E> hash_equal_;
-  std::unordered_map<T, I, H, E> entry2id_;
-  std::vector<T> id2entry_;
-};
-
-// Enables alternative hash set representations below.
-enum HSType { HS_STL = 0, HS_DENSE = 1, HS_SPARSE = 2, HS_FLAT = 3 };
-
-// Default hash set is STL hash_set.
-template <class K, class H, class E, HSType HS>
-struct HashSet : public std::unordered_set<K, H, E, PoolAllocator<K>> {
-  explicit HashSet(size_t n = 0, const H &h = H(), const E &e = E())
-      : std::unordered_set<K, H, E, PoolAllocator<K>>(n, h, e) {}
-
-  void rehash(size_t n) {}
-};
-
-// An implementation using a hash set for the entry to ID mapping. The hash set
-// holds keys which are either the ID or kCurrentKey. These keys can be mapped
-// to entries either by looking up in the entry vector or, if kCurrentKey, in
-// current_entry_. The hash and key equality functions map to entries first. H
-// is the hash function and E is the equality function. If passed to the
-// constructor, ownership is given to this class.
-template <class I, class T, class H, class E = std::equal_to<T>,
-          HSType HS = HS_FLAT>
-class CompactHashBiTable {
- public:
-  friend class HashFunc;
-  friend class HashEqual;
-
-  // Reserves space for table_size elements. If passing H and E to the
-  // constructor, this class owns them.
-  explicit CompactHashBiTable(size_t table_size = 0, H *h = nullptr,
-                              E *e = nullptr) :
-        hash_func_(h ? h : new H()), hash_equal_(e ? e : new E()),
-        compact_hash_func_(*this), compact_hash_equal_(*this),
-        keys_(table_size, compact_hash_func_, compact_hash_equal_) {
-    if (table_size) id2entry_.reserve(table_size);
-  }
-
-  CompactHashBiTable(const CompactHashBiTable<I, T, H, E, HS> &table)
-      : hash_func_(new H(*table.hash_func_)),
-        hash_equal_(new E(*table.hash_equal_)),
-        compact_hash_func_(*this), compact_hash_equal_(*this),
-        keys_(table.keys_.size(), compact_hash_func_, compact_hash_equal_),
-        id2entry_(table.id2entry_) {
-    keys_.insert(table.keys_.begin(), table.keys_.end());
-  }
-
-  I FindId(const T &entry, bool insert = true) {
-    current_entry_ = &entry;
-    if (insert) {
-      auto result = keys_.insert(kCurrentKey);
-      if (!result.second) return *result.first;  // Already exists.
-      // Overwrites kCurrentKey with a new key value; this is safe because it
-      // doesn't affect hashing or equality testing.
-      I key = id2entry_.size();
-      const_cast<I &>(*result.first) = key;
-      id2entry_.push_back(entry);
-      return key;
-    }
-    const auto it = keys_.find(kCurrentKey);
-    return it == keys_.end() ? -1 : *it;
-  }
-
-  const T &FindEntry(I s) const { return id2entry_[s]; }
-
-  I Size() const { return id2entry_.size(); }
-
-  // Clears content; with argument, erases last n IDs.
-  void Clear(ssize_t n = -1) {
-    if (n < 0 || n >= id2entry_.size()) {  // Clears completely.
-      keys_.clear();
-      id2entry_.clear();
-    } else if (n == id2entry_.size() - 1) {  // Leaves only key 0.
-      const T entry = FindEntry(0);
-      keys_.clear();
-      id2entry_.clear();
-      FindId(entry, true);
-    } else {
-      while (n-- > 0) {
-        I key = id2entry_.size() - 1;
-        keys_.erase(key);
-        id2entry_.pop_back();
-      }
-      keys_.rehash(0);
-    }
-  }
-
- private:
-  static constexpr I kCurrentKey = -1;
-  static constexpr I kEmptyKey = -2;
-  static constexpr I kDeletedKey = -3;
-
-  class HashFunc {
-   public:
-    explicit HashFunc(const CompactHashBiTable &ht) : ht_(&ht) {}
-
-    size_t operator()(I k) const {
-      if (k >= kCurrentKey) {
-        return (*ht_->hash_func_)(ht_->Key2Entry(k));
-      } else {
-        return 0;
-      }
-    }
-
-   private:
-    const CompactHashBiTable *ht_;
-  };
-
-  class HashEqual {
-   public:
-    explicit HashEqual(const CompactHashBiTable &ht) : ht_(&ht) {}
-
-    bool operator()(I k1, I k2) const {
-      if (k1 == k2) {
-        return true;
-      } else if (k1 >= kCurrentKey && k2 >= kCurrentKey) {
-        return (*ht_->hash_equal_)(ht_->Key2Entry(k1), ht_->Key2Entry(k2));
-      } else {
-        return false;
-      }
-    }
-
-   private:
-    const CompactHashBiTable *ht_;
-  };
-
-  using KeyHashSet = HashSet<I, HashFunc, HashEqual, HS>;
-
-  const T &Key2Entry(I k) const {
-    if (k == kCurrentKey) {
-      return *current_entry_;
-    } else {
-      return id2entry_[k];
-    }
-  }
-
-  std::unique_ptr<H> hash_func_;
-  std::unique_ptr<E> hash_equal_;
-  HashFunc compact_hash_func_;
-  HashEqual compact_hash_equal_;
-  KeyHashSet keys_;
-  std::vector<T> id2entry_;
-  const T *current_entry_;
-};
-
-template <class I, class T, class H, class E, HSType HS>
-constexpr I CompactHashBiTable<I, T, H, E, HS>::kCurrentKey;
-
-template <class I, class T, class H, class E, HSType HS>
-constexpr I CompactHashBiTable<I, T, H, E, HS>::kEmptyKey;
-
-template <class I, class T, class H, class E, HSType HS>
-constexpr I CompactHashBiTable<I, T, H, E, HS>::kDeletedKey;
-
-// An implementation using a vector for the entry to ID mapping. It is passed a
-// function object FP that should fingerprint entries uniquely to an integer
-// that can used as a vector index. Normally, VectorBiTable constructs the FP
-// object. The user can instead pass in this object; in that case, VectorBiTable
-// takes its ownership.
-template <class I, class T, class FP>
-class VectorBiTable {
- public:
-  // Reserves table_size cells of space. If passing FP argument to the
-  // constructor, this class owns it.
-  explicit VectorBiTable(FP *fp = nullptr, size_t table_size = 0) :
-      fp_(fp ? fp : new FP()) {
-    if (table_size) id2entry_.reserve(table_size);
-  }
-
-  VectorBiTable(const VectorBiTable<I, T, FP> &table)
-      : fp_(new FP(*table.fp_)), fp2id_(table.fp2id_),
-        id2entry_(table.id2entry_) {}
-
-  I FindId(const T &entry, bool insert = true) {
-    ssize_t fp = (*fp_)(entry);
-    if (fp >= fp2id_.size()) fp2id_.resize(fp + 1);
-    I &id_ref = fp2id_[fp];
-    if (id_ref == 0) {  // T not found.
-      if (insert) {     // Stores and assigns a new ID.
-        id2entry_.push_back(entry);
-        id_ref = id2entry_.size();
-      } else {
-        return -1;
-      }
-    }
-    return id_ref - 1;  // NB: id_ref = ID + 1.
-  }
-
-  const T &FindEntry(I s) const { return id2entry_[s]; }
-
-  I Size() const { return id2entry_.size(); }
-
-  const FP &Fingerprint() const { return *fp_; }
-
- private:
-  std::unique_ptr<FP> fp_;
-  std::vector<I> fp2id_;
-  std::vector<T> id2entry_;
-};
-
-// An implementation using a vector and a compact hash table. The selecting
-// functor S returns true for entries to be hashed in the vector. The
-// fingerprinting functor FP returns a unique fingerprint for each entry to be
-// hashed in the vector (these need to be suitable for indexing in a vector).
-// The hash functor H is used when hashing entry into the compact hash table.
-// If passed to the constructor, ownership is given to this class.
-template <class I, class T, class S, class FP, class H, HSType HS = HS_DENSE>
-class VectorHashBiTable {
- public:
-  friend class HashFunc;
-  friend class HashEqual;
-
-  explicit VectorHashBiTable(S *s, FP *fp, H *h, size_t vector_size = 0,
-                             size_t entry_size = 0)
-      : selector_(s), fp_(fp), h_(h), hash_func_(*this), hash_equal_(*this),
-        keys_(0, hash_func_, hash_equal_) {
-    if (vector_size) fp2id_.reserve(vector_size);
-    if (entry_size) id2entry_.reserve(entry_size);
-  }
-
-  VectorHashBiTable(const VectorHashBiTable<I, T, S, FP, H, HS> &table)
-      : selector_(new S(*table.selector_)), fp_(new FP(*table.fp_)),
-        h_(new H(*table.h_)), id2entry_(table.id2entry_),
-        fp2id_(table.fp2id_), hash_func_(*this), hash_equal_(*this),
-        keys_(table.keys_.size(), hash_func_, hash_equal_) {
-    keys_.insert(table.keys_.begin(), table.keys_.end());
-  }
-
-  I FindId(const T &entry, bool insert = true) {
-    if ((*selector_)(entry)) {  // Uses the vector if selector_(entry) == true.
-      uint64 fp = (*fp_)(entry);
-      if (fp2id_.size() <= fp) fp2id_.resize(fp + 1, 0);
-      if (fp2id_[fp] == 0) {  // T not found.
-        if (insert) {         // Stores and assigns a new ID.
-          id2entry_.push_back(entry);
-          fp2id_[fp] = id2entry_.size();
-        } else {
-          return -1;
-        }
-      }
-      return fp2id_[fp] - 1;  // NB: assoc_value = ID + 1.
-    } else {                  // Uses the hash table otherwise.
-      current_entry_ = &entry;
-      const auto it = keys_.find(kCurrentKey);
-      if (it == keys_.end()) {
-        if (insert) {
-          I key = id2entry_.size();
-          id2entry_.push_back(entry);
-          keys_.insert(key);
-          return key;
-        } else {
-          return -1;
-        }
-      } else {
-        return *it;
-      }
-    }
-  }
-
-  const T &FindEntry(I s) const { return id2entry_[s]; }
-
-  I Size() const { return id2entry_.size(); }
-
-  const S &Selector() const { return *selector_; }
-
-  const FP &Fingerprint() const { return *fp_; }
-
-  const H &Hash() const { return *h_; }
-
- private:
-  static constexpr I kCurrentKey = -1;
-  static constexpr I kEmptyKey = -2;
-
-  class HashFunc {
-   public:
-    explicit HashFunc(const VectorHashBiTable &ht) : ht_(&ht) {}
-
-    size_t operator()(I k) const {
-      if (k >= kCurrentKey) {
-        return (*(ht_->h_))(ht_->Key2Entry(k));
-      } else {
-        return 0;
-      }
-    }
-
-   private:
-    const VectorHashBiTable *ht_;
-  };
-
-  class HashEqual {
-   public:
-    explicit HashEqual(const VectorHashBiTable &ht) : ht_(&ht) {}
-
-    bool operator()(I k1, I k2) const {
-      if (k1 >= kCurrentKey && k2 >= kCurrentKey) {
-        return ht_->Key2Entry(k1) == ht_->Key2Entry(k2);
-      } else {
-        return k1 == k2;
-      }
-    }
-
-   private:
-    const VectorHashBiTable *ht_;
-  };
-
-  using KeyHashSet = HashSet<I, HashFunc, HashEqual, HS>;
-
-  const T &Key2Entry(I k) const {
-    if (k == kCurrentKey) {
-      return *current_entry_;
-    } else {
-      return id2entry_[k];
-    }
-  }
-
-  std::unique_ptr<S> selector_;  // True if entry hashed into vector.
-  std::unique_ptr<FP> fp_;       // Fingerprint used for hashing into vector.
-  std::unique_ptr<H> h_;         // Hash funcion used for hashing into hash_set.
-
-  std::vector<T> id2entry_;  // Maps state IDs to entry.
-  std::vector<I> fp2id_;     // Maps entry fingerprints to IDs.
-
-  // Compact implementation of the hash table mapping entries to state IDs
-  // using the hash function h_.
-  HashFunc hash_func_;
-  HashEqual hash_equal_;
-  KeyHashSet keys_;
-  const T *current_entry_;
-};
-
-template <class I, class T, class S, class FP, class H, HSType HS>
-constexpr I VectorHashBiTable<I, T, S, FP, H, HS>::kCurrentKey;
-
-template <class I, class T, class S, class FP, class H, HSType HS>
-constexpr I VectorHashBiTable<I, T, S, FP, H, HS>::kEmptyKey;
-
-// An implementation using a hash map for the entry to ID mapping. This version
-// permits erasing of arbitrary states. The entry T must have == defined and
-// its default constructor must produce a entry that will never be seen. F is
-// the hash function.
-template <class I, class T, class F>
-class ErasableBiTable {
- public:
-  ErasableBiTable() : first_(0) {}
-
-  I FindId(const T &entry, bool insert = true) {
-    I &id_ref = entry2id_[entry];
-    if (id_ref == 0) {  // T not found.
-      if (insert) {     // Stores and assigns a new ID.
-        id2entry_.push_back(entry);
-        id_ref = id2entry_.size() + first_;
-      } else {
-        return -1;
-      }
-    }
-    return id_ref - 1;  // NB: id_ref = ID + 1.
-  }
-
-  const T &FindEntry(I s) const { return id2entry_[s - first_]; }
-
-  I Size() const { return id2entry_.size(); }
-
-  void Erase(I s) {
-    auto &ref = id2entry_[s - first_];
-    entry2id_.erase(ref);
-    ref = empty_entry_;
-    while (!id2entry_.empty() && id2entry_.front() == empty_entry_) {
-      id2entry_.pop_front();
-      ++first_;
-    }
-  }
-
- private:
-  std::unordered_map<T, I, F> entry2id_;
-  std::deque<T> id2entry_;
-  const T empty_entry_;
-  I first_;  // I of first element in the deque.
-};
-
-}  // namespace fst
-
-#endif  // FST_BI_TABLE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/cache.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/cache.h
deleted file mode 100644
index 45d158f5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/cache.h
+++ /dev/null
@@ -1,1277 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// An FST implementation that caches FST elements of a delayed computation.
-
-#ifndef FST_CACHE_H_
-#define FST_CACHE_H_
-
-#include <functional>
-#include <unordered_map>
-using std::unordered_map;
-using std::unordered_multimap;
-#include <list>
-#include <vector>
-
-#include <fst/flags.h>
-#include <fst/log.h>
-
-#include <fst/vector-fst.h>
-
-
-DECLARE_bool(fst_default_cache_gc);
-DECLARE_int64(fst_default_cache_gc_limit);
-
-namespace fst {
-
-// Options for controlling caching behavior; higher level than CacheImplOptions.
-struct CacheOptions {
-  bool gc;          // Enables GC.
-  size_t gc_limit;  // Number of bytes allowed before GC.
-
-  explicit CacheOptions(bool gc = FLAGS_fst_default_cache_gc,
-                        size_t gc_limit = FLAGS_fst_default_cache_gc_limit)
-      : gc(gc), gc_limit(gc_limit) {}
-};
-
-// Options for controlling caching behavior, at a lower level than
-// CacheOptions; templated on the cache store and allows passing the store.
-template <class CacheStore>
-struct CacheImplOptions {
-  bool gc;            // Enables GC.
-  size_t gc_limit;    // Number of bytes allowed before GC.
-  CacheStore *store;  // Cache store.
-  bool own_store;     // Should CacheImpl takes ownership of the store?
-
-  explicit CacheImplOptions(bool gc = FLAGS_fst_default_cache_gc,
-                            size_t gc_limit = FLAGS_fst_default_cache_gc_limit,
-                            CacheStore *store = nullptr)
-      : gc(gc), gc_limit(gc_limit), store(store), own_store(true) {}
-
-  explicit CacheImplOptions(const CacheOptions &opts)
-      : gc(opts.gc), gc_limit(opts.gc_limit), store(nullptr), own_store(true) {}
-};
-
-// Cache flags.
-constexpr uint32 kCacheFinal = 0x0001;   // Final weight has been cached.
-constexpr uint32 kCacheArcs = 0x0002;    // Arcs have been cached.
-constexpr uint32 kCacheInit = 0x0004;    // Initialized by GC.
-constexpr uint32 kCacheRecent = 0x0008;  // Visited since GC.
-constexpr uint32 kCacheFlags =
-    kCacheFinal | kCacheArcs | kCacheInit | kCacheRecent;
-
-// Cache state, with arcs stored in a per-state std::vector.
-template <class A, class M = PoolAllocator<A>>
-class CacheState {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using ArcAllocator = M;
-  using StateAllocator =
-      typename ArcAllocator::template rebind<CacheState<A, M>>::other;
-
-  // Provides STL allocator for arcs.
-  explicit CacheState(const ArcAllocator &alloc)
-      : final_(Weight::Zero()),
-        niepsilons_(0),
-        noepsilons_(0),
-        arcs_(alloc),
-        flags_(0),
-        ref_count_(0) {}
-
-  CacheState(const CacheState<A> &state, const ArcAllocator &alloc)
-      : final_(state.Final()),
-        niepsilons_(state.NumInputEpsilons()),
-        noepsilons_(state.NumOutputEpsilons()),
-        arcs_(state.arcs_.begin(), state.arcs_.end(), alloc),
-        flags_(state.Flags()),
-        ref_count_(0) {}
-
-  void Reset() {
-    final_ = Weight::Zero();
-    niepsilons_ = 0;
-    noepsilons_ = 0;
-    ref_count_ = 0;
-    flags_ = 0;
-    arcs_.clear();
-  }
-
-  Weight Final() const { return final_; }
-
-  size_t NumInputEpsilons() const { return niepsilons_; }
-
-  size_t NumOutputEpsilons() const { return noepsilons_; }
-
-  size_t NumArcs() const { return arcs_.size(); }
-
-  const Arc &GetArc(size_t n) const { return arcs_[n]; }
-
-  // Used by the ArcIterator<Fst<Arc>> efficient implementation.
-  const Arc *Arcs() const { return !arcs_.empty() ? &arcs_[0] : nullptr; }
-
-  // Accesses flags; used by the caller.
-  uint32 Flags() const { return flags_; }
-
-  // Accesses ref count; used by the caller.
-  int RefCount() const { return ref_count_; }
-
-  void SetFinal(Weight weight) { final_ = std::move(weight); }
-
-  void ReserveArcs(size_t n) { arcs_.reserve(n); }
-
-  // Adds one arc at a time with all needed book-keeping; use PushArc and
-  // SetArcs for a more efficient alternative.
-  void AddArc(const Arc &arc) {
-    arcs_.push_back(arc);
-    if (arc.ilabel == 0) ++niepsilons_;
-    if (arc.olabel == 0) ++noepsilons_;
-  }
-
-  // Adds one arc at a time with delayed book-keeping; finalize with SetArcs().
-  void PushArc(const Arc &arc) { arcs_.push_back(arc); }
-
-  // Finalizes arcs book-keeping; call only once.
-  void SetArcs() {
-    for (const auto &arc : arcs_) {
-      if (arc.ilabel == 0) ++niepsilons_;
-      if (arc.olabel == 0) ++noepsilons_;
-    }
-  }
-
-  // Modifies nth arc.
-  void SetArc(const Arc &arc, size_t n) {
-    if (arcs_[n].ilabel == 0) --niepsilons_;
-    if (arcs_[n].olabel == 0) --noepsilons_;
-    if (arc.ilabel == 0) ++niepsilons_;
-    if (arc.olabel == 0) ++noepsilons_;
-    arcs_[n] = arc;
-  }
-
-  // Deletes all arcs.
-  void DeleteArcs() {
-    niepsilons_ = 0;
-    noepsilons_ = 0;
-    arcs_.clear();
-  }
-
-  void DeleteArcs(size_t n) {
-    for (size_t i = 0; i < n; ++i) {
-      if (arcs_.back().ilabel == 0) --niepsilons_;
-      if (arcs_.back().olabel == 0) --noepsilons_;
-      arcs_.pop_back();
-    }
-  }
-
-  // Sets status flags; used by the caller.
-  void SetFlags(uint32 flags, uint32 mask) const {
-    flags_ &= ~mask;
-    flags_ |= flags;
-  }
-
-  // Mutates reference counts; used by the caller.
-
-  int IncrRefCount() const { return ++ref_count_; }
-
-  int DecrRefCount() const { return --ref_count_; }
-
-  // Used by the ArcIterator<Fst<Arc>> efficient implementation.
-  int *MutableRefCount() const { return &ref_count_; }
-
-  // Used for state class allocation.
-  void *operator new(size_t size, StateAllocator *alloc) {
-    return alloc->allocate(1);
-  }
-
-  // For state destruction and memory freeing.
-  static void Destroy(CacheState<Arc> *state, StateAllocator *alloc) {
-    if (state) {
-      state->~CacheState<Arc>();
-      alloc->deallocate(state, 1);
-    }
-  }
-
- private:
-  Weight final_;                         // Final weight.
-  size_t niepsilons_;                    // # of input epsilons.
-  size_t noepsilons_;                    // # of output epsilons.
-  std::vector<Arc, ArcAllocator> arcs_;  // Arcs representation.
-  mutable uint32 flags_;
-  mutable int ref_count_;  // If 0, available for GC.
-};
-
-// Cache store, allocating and storing states, providing a mapping from state
-// IDs to cached states, and an iterator over these states. The state template
-// argument must implement the CacheState interface. The state for a StateId s
-// is constructed when requested by GetMutableState(s) if it is not yet stored.
-// Initially, a state has a reference count of zero, but the user may increment
-// or decrement this to control the time of destruction. In particular, a state
-// is destroyed when:
-//
-// 1. This instance is destroyed, or
-// 2. Clear() or Delete() is called, or
-// 3. Possibly (implementation-dependently) when:
-//    - Garbage collection is enabled (as defined by opts.gc),
-//    - The cache store size exceeds the limits (as defined by opts.gc_limits),
-//    - The state's reference count is zero, and
-//    - The state is not the most recently requested state.
-//
-// template <class S>
-// class CacheStore {
-//  public:
-//   using State = S;
-//   using Arc = typename State::Arc;
-//   using StateId = typename Arc::StateId;
-//
-//   // Required constructors/assignment operators.
-//   explicit CacheStore(const CacheOptions &opts);
-//
-//   // Returns nullptr if state is not stored.
-//   const State *GetState(StateId s);
-//
-//   // Creates state if state is not stored.
-//   State *GetMutableState(StateId s);
-//
-//   // Similar to State::AddArc() but updates cache store book-keeping.
-//   void AddArc(State *state, const Arc &arc);
-//
-//   // Similar to State::SetArcs() but updates cache store book-keeping; call
-//   // only once.
-//   void SetArcs(State *state);
-//
-//   // Similar to State::DeleteArcs() but updates cache store book-keeping.
-//
-//   void DeleteArcs(State *state);
-//
-//   void DeleteArcs(State *state, size_t n);
-//
-//   // Deletes all cached states.
-//   void Clear();
-//
-//   // Iterates over cached states (in an arbitrary order); only needed if
-//   // opts.gc is true.
-//   bool Done() const;      // End of iteration.
-//   StateId Value() const;  // Current state.
-//   void Next();            // Advances to next state (when !Done).
-//   void Reset();           // Returns to initial condition.
-//   void Delete();          // Deletes current state and advances to next.
-// };
-
-// Container cache stores.
-
-// This class uses a vector of pointers to states to store cached states.
-template <class S>
-class VectorCacheStore {
- public:
-  using State = S;
-  using Arc = typename State::Arc;
-  using StateId = typename Arc::StateId;
-  using StateList = std::list<StateId, PoolAllocator<StateId>>;
-
-  // Required constructors/assignment operators.
-  explicit VectorCacheStore(const CacheOptions &opts) : cache_gc_(opts.gc) {
-    Clear();
-    Reset();
-  }
-
-  VectorCacheStore(const VectorCacheStore<S> &store)
-      : cache_gc_(store.cache_gc_) {
-    CopyStates(store);
-    Reset();
-  }
-
-  ~VectorCacheStore() { Clear(); }
-
-  VectorCacheStore<State> &operator=(const VectorCacheStore<State> &store) {
-    if (this != &store) {
-      CopyStates(store);
-      Reset();
-    }
-    return *this;
-  }
-
-  // Returns nullptr if state is not stored.
-  const State *GetState(StateId s) const {
-    return s < state_vec_.size() ? state_vec_[s] : nullptr;
-  }
-
-  // Creates state if state is not stored.
-  State *GetMutableState(StateId s) {
-    State *state = nullptr;
-    if (s >= state_vec_.size()) {
-      state_vec_.resize(s + 1, nullptr);
-    } else {
-      state = state_vec_[s];
-    }
-    if (!state) {
-      state = new (&state_alloc_) State(arc_alloc_);
-      state_vec_[s] = state;
-      if (cache_gc_) state_list_.push_back(s);
-    }
-    return state;
-  }
-
-  // Similar to State::AddArc() but updates cache store book-keeping
-  void AddArc(State *state, const Arc &arc) { state->AddArc(arc); }
-
-  // Similar to State::SetArcs() but updates cache store book-keeping; call
-  // only once.
-  void SetArcs(State *state) { state->SetArcs(); }
-
-  // Deletes all arcs.
-  void DeleteArcs(State *state) { state->DeleteArcs(); }
-
-  // Deletes some arcs.
-  void DeleteArcs(State *state, size_t n) { state->DeleteArcs(n); }
-
-  // Deletes all cached states.
-  void Clear() {
-    for (StateId s = 0; s < state_vec_.size(); ++s) {
-      State::Destroy(state_vec_[s], &state_alloc_);
-    }
-    state_vec_.clear();
-    state_list_.clear();
-  }
-
-  // Iterates over cached states (in an arbitrary order); only works if GC is
-  // enabled (o.w. avoiding state_list_ overhead).
-  bool Done() const { return iter_ == state_list_.end(); }
-
-  StateId Value() const { return *iter_; }
-
-  void Next() { ++iter_; }
-
-  void Reset() { iter_ = state_list_.begin(); }
-
-  // Deletes current state and advances to next.
-  void Delete() {
-    State::Destroy(state_vec_[*iter_], &state_alloc_);
-    state_vec_[*iter_] = nullptr;
-    state_list_.erase(iter_++);
-  }
-
- private:
-  void CopyStates(const VectorCacheStore<State> &store) {
-    Clear();
-    state_vec_.reserve(store.state_vec_.size());
-    for (StateId s = 0; s < store.state_vec_.size(); ++s) {
-      State *state = nullptr;
-      const auto *store_state = store.state_vec_[s];
-      if (store_state) {
-        state = new (&state_alloc_) State(*store_state, arc_alloc_);
-        if (cache_gc_) state_list_.push_back(s);
-      }
-      state_vec_.push_back(state);
-    }
-  }
-
-  bool cache_gc_;                               // Supports iteration when true.
-  std::vector<State *> state_vec_;              // Vector of states (or null).
-  StateList state_list_;                        // List of states.
-  typename StateList::iterator iter_;           // State list iterator.
-  typename State::StateAllocator state_alloc_;  // For state allocation.
-  typename State::ArcAllocator arc_alloc_;      // For arc allocation.
-};
-
-// This class uses a hash map from state IDs to pointers to cached states.
-template <class S>
-class HashCacheStore {
- public:
-  using State = S;
-  using Arc = typename State::Arc;
-  using StateId = typename Arc::StateId;
-
-  using StateMap =
-      std::unordered_map<StateId, State *, std::hash<StateId>,
-                         std::equal_to<StateId>,
-                         PoolAllocator<std::pair<const StateId, State *>>>;
-
-  // Required constructors/assignment operators.
-  explicit HashCacheStore(const CacheOptions &opts) {
-    Clear();
-    Reset();
-  }
-
-  HashCacheStore(const HashCacheStore<S> &store) {
-    CopyStates(store);
-    Reset();
-  }
-
-  ~HashCacheStore() { Clear(); }
-
-  HashCacheStore<State> &operator=(const HashCacheStore<State> &store) {
-    if (this != &store) {
-      CopyStates(store);
-      Reset();
-    }
-    return *this;
-  }
-
-  // Returns nullptr if state is not stored.
-  const State *GetState(StateId s) const {
-    const auto it = state_map_.find(s);
-    return it != state_map_.end() ? it->second : nullptr;
-  }
-
-  // Creates state if state is not stored.
-  State *GetMutableState(StateId s) {
-    auto *&state = state_map_[s];
-    if (!state) state = new (&state_alloc_) State(arc_alloc_);
-    return state;
-  }
-
-  // Similar to State::AddArc() but updates cache store book-keeping.
-  void AddArc(State *state, const Arc &arc) { state->AddArc(arc); }
-
-  // Similar to State::SetArcs() but updates internal cache size; call only
-  // once.
-  void SetArcs(State *state) { state->SetArcs(); }
-
-  // Deletes all arcs.
-  void DeleteArcs(State *state) { state->DeleteArcs(); }
-
-  // Deletes some arcs.
-  void DeleteArcs(State *state, size_t n) { state->DeleteArcs(n); }
-
-  // Deletes all cached states.
-  void Clear() {
-    for (auto it = state_map_.begin(); it != state_map_.end(); ++it) {
-      State::Destroy(it->second, &state_alloc_);
-    }
-    state_map_.clear();
-  }
-
-  // Iterates over cached states (in an arbitrary order).
-  bool Done() const { return iter_ == state_map_.end(); }
-
-  StateId Value() const { return iter_->first; }
-
-  void Next() { ++iter_; }
-
-  void Reset() { iter_ = state_map_.begin(); }
-
-  // Deletes current state and advances to next.
-  void Delete() {
-    State::Destroy(iter_->second, &state_alloc_);
-    state_map_.erase(iter_++);
-  }
-
- private:
-  void CopyStates(const HashCacheStore<State> &store) {
-    Clear();
-    for (auto it = store.state_map_.begin(); it != store.state_map_.end();
-         ++it) {
-      state_map_[it->first] =
-          new (&state_alloc_) State(*it->second, arc_alloc_);
-    }
-  }
-
-  StateMap state_map_;                          // Map from state ID to state.
-  typename StateMap::iterator iter_;            // State map iterator.
-  typename State::StateAllocator state_alloc_;  // For state allocation.
-  typename State::ArcAllocator arc_alloc_;      // For arc allocation.
-};
-
-// Garbage-colllection cache stores.
-
-// This class implements a simple garbage collection scheme when
-// 'opts.gc_limit = 0'. In particular, the first cached state is reused for each
-// new state so long as the reference count is zero on the to-be-reused state.
-// Otherwise, the full underlying store is used. The caller can increment the
-// reference count to inhibit the GC of in-use states (e.g., in an ArcIterator).
-//
-// The typical use case for this optimization is when a single pass over a
-// cached
-// FST is performed with only one-state expanded at a time.
-template <class CacheStore>
-class FirstCacheStore {
- public:
-  using State = typename CacheStore::State;
-  using Arc = typename State::Arc;
-  using StateId = typename Arc::StateId;
-
-  // Required constructors/assignment operators.
-  explicit FirstCacheStore(const CacheOptions &opts)
-      : store_(opts),
-        cache_gc_(opts.gc_limit == 0),  // opts.gc ignored historically.
-        cache_first_state_id_(kNoStateId),
-        cache_first_state_(nullptr) {}
-
-  FirstCacheStore(const FirstCacheStore<CacheStore> &store)
-      : store_(store.store_),
-        cache_gc_(store.cache_gc_),
-        cache_first_state_id_(store.cache_first_state_id_),
-        cache_first_state_(store.cache_first_state_id_ != kNoStateId
-                               ? store_.GetMutableState(0)
-                               : nullptr) {}
-
-  FirstCacheStore<CacheStore> &operator=(
-      const FirstCacheStore<CacheStore> &store) {
-    if (this != &store) {
-      store_ = store.store_;
-      cache_gc_ = store.cache_gc_;
-      cache_first_state_id_ = store.cache_first_state_id_;
-      cache_first_state_ = store.cache_first_state_id_ != kNoStateId
-                               ? store_.GetMutableState(0)
-                               : nullptr;
-    }
-    return *this;
-  }
-
-  // Returns nullptr if state is not stored.
-  const State *GetState(StateId s) const {
-    // store_ state 0 may hold first cached state; the rest are shifted by 1.
-    return s == cache_first_state_id_ ? cache_first_state_
-                                      : store_.GetState(s + 1);
-  }
-
-  // Creates state if state is not stored.
-  State *GetMutableState(StateId s) {
-    // store_ state 0 used to hold first cached state; the rest are shifted by
-    // 1.
-    if (cache_first_state_id_ == s) {
-      return cache_first_state_;  // Request for first cached state.
-    }
-    if (cache_gc_) {
-      if (cache_first_state_id_ == kNoStateId) {
-        cache_first_state_id_ = s;  // Sets first cached state.
-        cache_first_state_ = store_.GetMutableState(0);
-        cache_first_state_->SetFlags(kCacheInit, kCacheInit);
-        cache_first_state_->ReserveArcs(2 * kAllocSize);
-        return cache_first_state_;
-      } else if (cache_first_state_->RefCount() == 0) {
-        cache_first_state_id_ = s;  // Updates first cached state.
-        cache_first_state_->Reset();
-        cache_first_state_->SetFlags(kCacheInit, kCacheInit);
-        return cache_first_state_;
-      } else {  // Keeps first cached state.
-        cache_first_state_->SetFlags(0, kCacheInit);  // Clears initialized bit.
-        cache_gc_ = false;                            // Disables GC.
-      }
-    }
-    auto *state = store_.GetMutableState(s + 1);
-    return state;
-  }
-
-  // Similar to State::AddArc() but updates cache store book-keeping.
-  void AddArc(State *state, const Arc &arc) { store_.AddArc(state, arc); }
-
-  // Similar to State::SetArcs() but updates internal cache size; call only
-  // once.
-  void SetArcs(State *state) { store_.SetArcs(state); }
-
-  // Deletes all arcs
-  void DeleteArcs(State *state) { store_.DeleteArcs(state); }
-
-  // Deletes some arcs
-  void DeleteArcs(State *state, size_t n) { store_.DeleteArcs(state, n); }
-
-  // Deletes all cached states
-  void Clear() {
-    store_.Clear();
-    cache_first_state_id_ = kNoStateId;
-    cache_first_state_ = nullptr;
-  }
-
-  // Iterates over cached states (in an arbitrary order). Only needed if GC is
-  // enabled.
-  bool Done() const { return store_.Done(); }
-
-  StateId Value() const {
-    // store_ state 0 may hold first cached state; rest shifted + 1.
-    const auto s = store_.Value();
-    return s ? s - 1 : cache_first_state_id_;
-  }
-
-  void Next() { store_.Next(); }
-
-  void Reset() { store_.Reset(); }
-
-  // Deletes current state and advances to next.
-  void Delete() {
-    if (Value() == cache_first_state_id_) {
-      cache_first_state_id_ = kNoStateId;
-      cache_first_state_ = nullptr;
-    }
-    store_.Delete();
-  }
-
- private:
-  CacheStore store_;              // Underlying store.
-  bool cache_gc_;                 // GC enabled.
-  StateId cache_first_state_id_;  // First cached state ID.
-  State *cache_first_state_;      // First cached state.
-};
-
-// This class implements mark-sweep garbage collection on an underlying cache
-// store. If GC is enabled, garbage collection of states is performed in a
-// rough approximation of LRU order once when 'gc_limit' bytes is reached. The
-// caller can increment the reference count to inhibit the GC of in-use state
-// (e.g., in an ArcIterator). With GC enabled, the 'gc_limit' parameter allows
-// the caller to trade-off time vs. space.
-template <class CacheStore>
-class GCCacheStore {
- public:
-  using State = typename CacheStore::State;
-  using Arc = typename State::Arc;
-  using StateId = typename Arc::StateId;
-
-  // Required constructors/assignment operators.
-  explicit GCCacheStore(const CacheOptions &opts)
-      : store_(opts),
-        cache_gc_request_(opts.gc),
-        cache_limit_(opts.gc_limit > kMinCacheLimit ? opts.gc_limit
-                                                    : kMinCacheLimit),
-        cache_gc_(false),
-        cache_size_(0) {}
-
-  // Returns 0 if state is not stored.
-  const State *GetState(StateId s) const { return store_.GetState(s); }
-
-  // Creates state if state is not stored
-  State *GetMutableState(StateId s) {
-    auto *state = store_.GetMutableState(s);
-    if (cache_gc_request_ && !(state->Flags() & kCacheInit)) {
-      state->SetFlags(kCacheInit, kCacheInit);
-      cache_size_ += sizeof(State) + state->NumArcs() * sizeof(Arc);
-      // GC is enabled once an uninited state (from underlying store) is seen.
-      cache_gc_ = true;
-      if (cache_size_ > cache_limit_) GC(state, false);
-    }
-    return state;
-  }
-
-  // Similar to State::AddArc() but updates cache store book-keeping.
-  void AddArc(State *state, const Arc &arc) {
-    store_.AddArc(state, arc);
-    if (cache_gc_ && (state->Flags() & kCacheInit)) {
-      cache_size_ += sizeof(Arc);
-      if (cache_size_ > cache_limit_) GC(state, false);
-    }
-  }
-
-  // Similar to State::SetArcs() but updates internal cache size; call only
-  // once.
-  void SetArcs(State *state) {
-    store_.SetArcs(state);
-    if (cache_gc_ && (state->Flags() & kCacheInit)) {
-      cache_size_ += state->NumArcs() * sizeof(Arc);
-      if (cache_size_ > cache_limit_) GC(state, false);
-    }
-  }
-
-  // Deletes all arcs.
-  void DeleteArcs(State *state) {
-    if (cache_gc_ && (state->Flags() & kCacheInit)) {
-      cache_size_ -= state->NumArcs() * sizeof(Arc);
-    }
-    store_.DeleteArcs(state);
-  }
-
-  // Deletes some arcs.
-  void DeleteArcs(State *state, size_t n) {
-    if (cache_gc_ && (state->Flags() & kCacheInit)) {
-      cache_size_ -= n * sizeof(Arc);
-    }
-    store_.DeleteArcs(state, n);
-  }
-
-  // Deletes all cached states.
-  void Clear() {
-    store_.Clear();
-    cache_size_ = 0;
-  }
-
-  // Iterates over cached states (in an arbitrary order); only needed if GC is
-  // enabled.
-  bool Done() const { return store_.Done(); }
-
-  StateId Value() const { return store_.Value(); }
-
-  void Next() { store_.Next(); }
-
-  void Reset() { store_.Reset(); }
-
-  // Deletes current state and advances to next.
-  void Delete() {
-    if (cache_gc_) {
-      const auto *state = store_.GetState(Value());
-      if (state->Flags() & kCacheInit) {
-        cache_size_ -= sizeof(State) + state->NumArcs() * sizeof(Arc);
-      }
-    }
-    store_.Delete();
-  }
-
-  // Removes from the cache store (not referenced-counted and not the current)
-  // states that have not been accessed since the last GC until at most
-  // cache_fraction * cache_limit_ bytes are cached. If that fails to free
-  // enough, attempts to uncaching recently visited states as well. If still
-  // unable to free enough memory, then widens cache_limit_.
-  void GC(const State *current, bool free_recent, float cache_fraction = 0.666);
-
-  // Returns the current cache size in bytes or 0 if GC is disabled.
-  size_t CacheSize() const { return cache_size_; }
-
-  // Returns the cache limit in bytes.
-  size_t CacheLimit() const { return cache_limit_; }
-
- private:
-  static constexpr size_t kMinCacheLimit = 8096;  // Minimum cache limit.
-
-  CacheStore store_;       // Underlying store.
-  bool cache_gc_request_;  // GC requested but possibly not yet enabled.
-  size_t cache_limit_;     // Number of bytes allowed before GC.
-  bool cache_gc_;          // GC enabled
-  size_t cache_size_;      // Number of bytes cached.
-};
-
-template <class CacheStore>
-void GCCacheStore<CacheStore>::GC(const State *current, bool free_recent,
-                                  float cache_fraction) {
-  if (!cache_gc_) return;
-  VLOG(2) << "GCCacheStore: Enter GC: object = "
-          << "(" << this << "), free recently cached = " << free_recent
-          << ", cache size = " << cache_size_
-          << ", cache frac = " << cache_fraction
-          << ", cache limit = " << cache_limit_ << "\n";
-  size_t cache_target = cache_fraction * cache_limit_;
-  store_.Reset();
-  while (!store_.Done()) {
-    auto *state = store_.GetMutableState(store_.Value());
-    if (cache_size_ > cache_target && state->RefCount() == 0 &&
-        (free_recent || !(state->Flags() & kCacheRecent)) && state != current) {
-      if (state->Flags() & kCacheInit) {
-        size_t size = sizeof(State) + state->NumArcs() * sizeof(Arc);
-        if (size < cache_size_) {
-          cache_size_ -= size;
-        }
-      }
-      store_.Delete();
-    } else {
-      state->SetFlags(0, kCacheRecent);
-      store_.Next();
-    }
-  }
-  if (!free_recent && cache_size_ > cache_target) {  // Recurses on recent.
-    GC(current, true, cache_fraction);
-  } else if (cache_target > 0) {  // Widens cache limit.
-    while (cache_size_ > cache_target) {
-      cache_limit_ *= 2;
-      cache_target *= 2;
-    }
-  } else if (cache_size_ > 0) {
-    FSTERROR() << "GCCacheStore:GC: Unable to free all cached states";
-  }
-  VLOG(2) << "GCCacheStore: Exit GC: object = "
-          << "(" << this << "), free recently cached = " << free_recent
-          << ", cache size = " << cache_size_
-          << ", cache frac = " << cache_fraction
-          << ", cache limit = " << cache_limit_ << "\n";
-}
-
-template <class CacheStore>
-constexpr size_t GCCacheStore<CacheStore>::kMinCacheLimit;
-
-// This class is the default cache state and store used by CacheBaseImpl.
-// It uses VectorCacheStore for storage decorated by FirstCacheStore
-// and GCCacheStore to do (optional) garbage collection.
-template <class Arc>
-class DefaultCacheStore
-    : public GCCacheStore<FirstCacheStore<VectorCacheStore<CacheState<Arc>>>> {
- public:
-  explicit DefaultCacheStore(const CacheOptions &opts)
-      : GCCacheStore<FirstCacheStore<VectorCacheStore<CacheState<Arc>>>>(opts) {
-  }
-};
-
-namespace internal {
-
-// This class is used to cache FST elements stored in states of type State
-// (see CacheState) with the flags used to indicate what has been cached. Use
-// HasStart(), HasFinal(), and HasArcs() to determine if cached and SetStart(),
-// SetFinal(), AddArc(), (or PushArc() and SetArcs()) to cache. Note that you
-// must set the final weight even if the state is non-final to mark it as
-// cached. The state storage method and any garbage collection policy are
-// determined by the cache store. If the store is passed in with the options,
-// CacheBaseImpl takes ownership.
-template <class State,
-          class CacheStore = DefaultCacheStore<typename State::Arc>>
-class CacheBaseImpl : public FstImpl<typename State::Arc> {
- public:
-  using Arc = typename State::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = CacheStore;
-
-  using FstImpl<Arc>::Type;
-  using FstImpl<Arc>::Properties;
-
-  explicit CacheBaseImpl(const CacheOptions &opts = CacheOptions())
-      : has_start_(false),
-        cache_start_(kNoStateId),
-        nknown_states_(0),
-        min_unexpanded_state_id_(0),
-        max_expanded_state_id_(-1),
-        cache_gc_(opts.gc),
-        cache_limit_(opts.gc_limit),
-        cache_store_(new CacheStore(opts)),
-        new_cache_store_(true),
-        own_cache_store_(true) {}
-
-  explicit CacheBaseImpl(const CacheImplOptions<CacheStore> &opts)
-      : has_start_(false),
-        cache_start_(kNoStateId),
-        nknown_states_(0),
-        min_unexpanded_state_id_(0),
-        max_expanded_state_id_(-1),
-        cache_gc_(opts.gc),
-        cache_limit_(opts.gc_limit),
-        cache_store_(opts.store ? opts.store : new CacheStore(CacheOptions(
-                                                   opts.gc, opts.gc_limit))),
-        new_cache_store_(!opts.store),
-        own_cache_store_(opts.store ? opts.own_store : true) {}
-
-  // Preserve gc parameters. If preserve_cache is true, also preserves
-  // cache data.
-  CacheBaseImpl(const CacheBaseImpl<State, CacheStore> &impl,
-                bool preserve_cache = false)
-      : FstImpl<Arc>(),
-        has_start_(false),
-        cache_start_(kNoStateId),
-        nknown_states_(0),
-        min_unexpanded_state_id_(0),
-        max_expanded_state_id_(-1),
-        cache_gc_(impl.cache_gc_),
-        cache_limit_(impl.cache_limit_),
-        cache_store_(new CacheStore(CacheOptions(cache_gc_, cache_limit_))),
-        new_cache_store_(impl.new_cache_store_ || !preserve_cache),
-        own_cache_store_(true) {
-    if (preserve_cache) {
-      *cache_store_ = *impl.cache_store_;
-      has_start_ = impl.has_start_;
-      cache_start_ = impl.cache_start_;
-      nknown_states_ = impl.nknown_states_;
-      expanded_states_ = impl.expanded_states_;
-      min_unexpanded_state_id_ = impl.min_unexpanded_state_id_;
-      max_expanded_state_id_ = impl.max_expanded_state_id_;
-    }
-  }
-
-  ~CacheBaseImpl() override { if (own_cache_store_) delete cache_store_; }
-
-  void SetStart(StateId s) {
-    cache_start_ = s;
-    has_start_ = true;
-    if (s >= nknown_states_) nknown_states_ = s + 1;
-  }
-
-  void SetFinal(StateId s, Weight weight) {
-    auto *state = cache_store_->GetMutableState(s);
-    state->SetFinal(std::move(weight));
-    static constexpr auto flags = kCacheFinal | kCacheRecent;
-    state->SetFlags(flags, flags);
-  }
-
-// Disabled to ensure PushArc not AddArc is used in existing code
-// TODO(sorenj): re-enable for backing store
-#if 0
-  // AddArc adds a single arc to a state and does incremental cache
-  // book-keeping. For efficiency, prefer PushArc and SetArcs below
-  // when possible.
-  void AddArc(StateId s, const Arc &arc) {
-    auto *state = cache_store_->GetMutableState(s);
-    cache_store_->AddArc(state, arc);
-    if (arc.nextstate >= nknown_states_)
-      nknown_states_ = arc.nextstate + 1;
-    SetExpandedState(s);
-    static constexpr auto flags = kCacheArcs | kCacheRecent;
-    state->SetFlags(flags, flags);
-  }
-#endif
-
-  // Adds a single arc to a state but delays cache book-keeping. SetArcs must
-  // be called when all PushArc calls at a state are complete. Do not mix with
-  // calls to AddArc.
-  void PushArc(StateId s, const Arc &arc) {
-    auto *state = cache_store_->GetMutableState(s);
-    state->PushArc(arc);
-  }
-
-  // Marks arcs of a state as cached and does cache book-keeping after all
-  // calls to PushArc have been completed. Do not mix with calls to AddArc.
-  void SetArcs(StateId s) {
-    auto *state = cache_store_->GetMutableState(s);
-    cache_store_->SetArcs(state);
-    const auto narcs = state->NumArcs();
-    for (size_t a = 0; a < narcs; ++a) {
-      const auto &arc = state->GetArc(a);
-      if (arc.nextstate >= nknown_states_) nknown_states_ = arc.nextstate + 1;
-    }
-    SetExpandedState(s);
-    static constexpr auto flags = kCacheArcs | kCacheRecent;
-    state->SetFlags(flags, flags);
-  }
-
-  void ReserveArcs(StateId s, size_t n) {
-    auto *state = cache_store_->GetMutableState(s);
-    state->ReserveArcs(n);
-  }
-
-  void DeleteArcs(StateId s) {
-    auto *state = cache_store_->GetMutableState(s);
-    cache_store_->DeleteArcs(state);
-  }
-
-  void DeleteArcs(StateId s, size_t n) {
-    auto *state = cache_store_->GetMutableState(s);
-    cache_store_->DeleteArcs(state, n);
-  }
-
-  void Clear() {
-    nknown_states_ = 0;
-    min_unexpanded_state_id_ = 0;
-    max_expanded_state_id_ = -1;
-    has_start_ = false;
-    cache_start_ = kNoStateId;
-    cache_store_->Clear();
-  }
-
-  // Is the start state cached?
-  bool HasStart() const {
-    if (!has_start_ && Properties(kError)) has_start_ = true;
-    return has_start_;
-  }
-
-  // Is the final weight of the state cached?
-  bool HasFinal(StateId s) const {
-    const auto *state = cache_store_->GetState(s);
-    if (state && state->Flags() & kCacheFinal) {
-      state->SetFlags(kCacheRecent, kCacheRecent);
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  // Are arcs of the state cached?
-  bool HasArcs(StateId s) const {
-    const auto *state = cache_store_->GetState(s);
-    if (state && state->Flags() & kCacheArcs) {
-      state->SetFlags(kCacheRecent, kCacheRecent);
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  StateId Start() const { return cache_start_; }
-
-  Weight Final(StateId s) const {
-    const auto *state = cache_store_->GetState(s);
-    return state->Final();
-  }
-
-  size_t NumArcs(StateId s) const {
-    const auto *state = cache_store_->GetState(s);
-    return state->NumArcs();
-  }
-
-  size_t NumInputEpsilons(StateId s) const {
-    const auto *state = cache_store_->GetState(s);
-    return state->NumInputEpsilons();
-  }
-
-  size_t NumOutputEpsilons(StateId s) const {
-    const auto *state = cache_store_->GetState(s);
-    return state->NumOutputEpsilons();
-  }
-
-  // Provides information needed for generic arc iterator.
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const {
-    const auto *state = cache_store_->GetState(s);
-    data->base = nullptr;
-    data->narcs = state->NumArcs();
-    data->arcs = state->Arcs();
-    data->ref_count = state->MutableRefCount();
-    state->IncrRefCount();
-  }
-
-  // Number of known states.
-  StateId NumKnownStates() const { return nknown_states_; }
-
-  // Updates number of known states, taking into account the passed state ID.
-  void UpdateNumKnownStates(StateId s) {
-    if (s >= nknown_states_) nknown_states_ = s + 1;
-  }
-
-  // Finds the mininum never-expanded state ID.
-  StateId MinUnexpandedState() const {
-    while (min_unexpanded_state_id_ <= max_expanded_state_id_ &&
-           ExpandedState(min_unexpanded_state_id_)) {
-      ++min_unexpanded_state_id_;
-    }
-    return min_unexpanded_state_id_;
-  }
-
-  // Returns maximum ever-expanded state ID.
-  StateId MaxExpandedState() const { return max_expanded_state_id_; }
-
-  void SetExpandedState(StateId s) {
-    if (s > max_expanded_state_id_) max_expanded_state_id_ = s;
-    if (s < min_unexpanded_state_id_) return;
-    if (s == min_unexpanded_state_id_) ++min_unexpanded_state_id_;
-    if (cache_gc_ || cache_limit_ == 0) {
-      if (expanded_states_.size() <= s) expanded_states_.resize(s + 1, false);
-      expanded_states_[s] = true;
-    }
-  }
-
-  bool ExpandedState(StateId s) const {
-    if (cache_gc_ || cache_limit_ == 0) {
-      return expanded_states_[s];
-    } else if (new_cache_store_) {
-      return cache_store_->GetState(s) != nullptr;
-    } else {
-      // If the cache was not created by this class, then the cached state needs
-      // to be inspected to update nknown_states_.
-      return false;
-    }
-  }
-
-  const CacheStore *GetCacheStore() const { return cache_store_; }
-
-  CacheStore *GetCacheStore() { return cache_store_; }
-
-  // Caching on/off switch, limit and size accessors.
-
-  bool GetCacheGc() const { return cache_gc_; }
-
-  size_t GetCacheLimit() const { return cache_limit_; }
-
- private:
-  mutable bool has_start_;                   // Is the start state cached?
-  StateId cache_start_;                      // ID of start state.
-  StateId nknown_states_;                    // Number of known states.
-  std::vector<bool> expanded_states_;        // States that have been expanded.
-  mutable StateId min_unexpanded_state_id_;  // Minimum never-expanded state ID
-  mutable StateId max_expanded_state_id_;    // Maximum ever-expanded state ID
-  bool cache_gc_;                            // GC enabled.
-  size_t cache_limit_;       // Number of bytes allowed before GC.
-  CacheStore *cache_store_;  // The store of cached states.
-  bool new_cache_store_;     // Was the store was created by class?
-  bool own_cache_store_;     // Is the store owned by class?
-
-  CacheBaseImpl &operator=(const CacheBaseImpl &impl) = delete;
-};
-
-// A CacheBaseImpl with the default cache state type.
-template <class Arc>
-class CacheImpl : public CacheBaseImpl<CacheState<Arc>> {
- public:
-  using State = CacheState<Arc>;
-
-  CacheImpl() {}
-
-  explicit CacheImpl(const CacheOptions &opts)
-      : CacheBaseImpl<CacheState<Arc>>(opts) {}
-
-  CacheImpl(const CacheImpl<Arc> &impl, bool preserve_cache = false)
-      : CacheBaseImpl<State>(impl, preserve_cache) {}
-
- private:
-  CacheImpl &operator=(const CacheImpl &impl) = delete;
-};
-
-}  // namespace internal
-
-// Use this to make a state iterator for a CacheBaseImpl-derived FST, which must
-// have Arc and Store types defined. Note this iterator only returns those
-// states reachable from the initial state, so consider implementing a
-// class-specific one.
-//
-// This class may be derived from.
-template <class FST>
-class CacheStateIterator : public StateIteratorBase<typename FST::Arc> {
- public:
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = typename FST::Store;
-  using State = typename Store::State;
-  using Impl = internal::CacheBaseImpl<State, Store>;
-
-  CacheStateIterator(const FST &fst, Impl *impl)
-      : fst_(fst), impl_(impl), s_(0) {
-    fst_.Start();  // Forces start state.
-  }
-
-  bool Done() const final {
-    if (s_ < impl_->NumKnownStates()) return false;
-    for (StateId u = impl_->MinUnexpandedState(); u < impl_->NumKnownStates();
-         u = impl_->MinUnexpandedState()) {
-      // Forces state expansion.
-      ArcIterator<FST> aiter(fst_, u);
-      aiter.SetFlags(kArcValueFlags, kArcValueFlags | kArcNoCache);
-      for (; !aiter.Done(); aiter.Next()) {
-        impl_->UpdateNumKnownStates(aiter.Value().nextstate);
-      }
-      impl_->SetExpandedState(u);
-      if (s_ < impl_->NumKnownStates()) return false;
-    }
-    return true;
-  }
-
-  StateId Value() const final { return s_; }
-
-  void Next() final { ++s_; }
-
-  void Reset() final { s_ = 0; }
-
- private:
-  const FST &fst_;
-  Impl *impl_;
-  StateId s_;
-};
-
-// Used to make an arc iterator for a CacheBaseImpl-derived FST, which must
-// have Arc and State types defined.
-template <class FST>
-class CacheArcIterator {
- public:
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = typename FST::Store;
-  using State = typename Store::State;
-  using Impl = internal::CacheBaseImpl<State, Store>;
-
-  CacheArcIterator(Impl *impl, StateId s) : i_(0) {
-    state_ = impl->GetCacheStore()->GetMutableState(s);
-    state_->IncrRefCount();
-  }
-
-  ~CacheArcIterator() { state_->DecrRefCount(); }
-
-  bool Done() const { return i_ >= state_->NumArcs(); }
-
-  const Arc &Value() const { return state_->GetArc(i_); }
-
-  void Next() { ++i_; }
-
-  size_t Position() const { return i_; }
-
-  void Reset() { i_ = 0; }
-
-  void Seek(size_t a) { i_ = a; }
-
-  constexpr uint32 Flags() const { return kArcValueFlags; }
-
-  void SetFlags(uint32 flags, uint32 mask) {}
-
- private:
-  const State *state_;
-  size_t i_;
-
-  CacheArcIterator(const CacheArcIterator &) = delete;
-  CacheArcIterator &operator=(const CacheArcIterator &) = delete;
-};
-
-// Use this to make a mutable arc iterator for a CacheBaseImpl-derived FST,
-// which must have types Arc and Store defined.
-template <class FST>
-class CacheMutableArcIterator
-    : public MutableArcIteratorBase<typename FST::Arc> {
- public:
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = typename FST::Store;
-  using State = typename Store::State;
-  using Impl = internal::CacheBaseImpl<State, Store>;
-
-  // User must call MutateCheck() in the constructor.
-  CacheMutableArcIterator(Impl *impl, StateId s) : i_(0), s_(s), impl_(impl) {
-    state_ = impl_->GetCacheStore()->GetMutableState(s_);
-    state_->IncrRefCount();
-  }
-
-  ~CacheMutableArcIterator() override { state_->DecrRefCount(); }
-
-  bool Done() const final { return i_ >= state_->NumArcs(); }
-
-  const Arc &Value() const final { return state_->GetArc(i_); }
-
-  void Next() final { ++i_; }
-
-  size_t Position() const final { return i_; }
-
-  void Reset() final { i_ = 0; }
-
-  void Seek(size_t a) final { i_ = a; }
-
-  void SetValue(const Arc &arc) final { state_->SetArc(arc, i_); }
-
-  uint32 Flags() const final { return kArcValueFlags; }
-
-  void SetFlags(uint32, uint32) final {}
-
- private:
-  size_t i_;
-  StateId s_;
-  Impl *impl_;
-  State *state_;
-
-  CacheMutableArcIterator(const CacheMutableArcIterator &) = delete;
-  CacheMutableArcIterator &operator=(const CacheMutableArcIterator &) = delete;
-};
-
-// Wrap existing CacheStore implementation to use with ExpanderFst.
-template <class CacheStore>
-class ExpanderCacheStore {
- public:
-  using State = typename CacheStore::State;
-  using Arc = typename CacheStore::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit ExpanderCacheStore(const CacheOptions &opts = CacheOptions())
-      : store_(opts) {}
-
-  template <class Expander>
-  State *FindOrExpand(Expander &expander, StateId s) {  // NOLINT
-    auto *state = store_.GetMutableState(s);
-    if (state->Flags()) {
-      state->SetFlags(kCacheRecent, kCacheRecent);
-    } else {
-      StateBuilder builder(state);
-      expander.Expand(s, &builder);
-      state->SetFlags(kCacheFlags, kCacheFlags);
-      store_.SetArcs(state);
-    }
-    return state;
-  }
-
- private:
-  CacheStore store_;
-
-  struct StateBuilder {
-    State *state;
-
-    explicit StateBuilder(State *state_) : state(state_) {}
-
-    void AddArc(const Arc &arc) { state->PushArc(arc); }
-
-    void SetFinal(Weight weight) { state->SetFinal(std::move(weight)); }
-  };
-};
-
-}  // namespace fst
-
-#endif  // FST_CACHE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/closure.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/closure.h
deleted file mode 100644
index 0e3df04c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/closure.h
+++ /dev/null
@@ -1,136 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to compute the concatenative closure of an FST.
-
-#ifndef FST_CLOSURE_H_
-#define FST_CLOSURE_H_
-
-#include <algorithm>
-#include <vector>
-
-#include <fst/mutable-fst.h>
-#include <fst/rational.h>
-
-
-namespace fst {
-
-// Computes the concatenative closure. This version modifies its
-// MutableFst input. If an FST transduces string x to y with weight a,
-// then its closure transduces x to y with weight a, xx to yy with
-// weight Times(a, a), xxx to yyy with with Times(Times(a, a), a),
-// etc. If closure_type == CLOSURE_STAR, then the empty string is
-// transduced to itself with weight Weight::One() as well.
-//
-// Complexity:
-//
-//   Time: O(V)
-//   Space: O(V)
-//
-// where V is the number of states.
-template <class Arc>
-void Closure(MutableFst<Arc> *fst, ClosureType closure_type) {
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  const auto props = fst->Properties(kFstProperties, false);
-  const auto start = fst->Start();
-  for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
-       siter.Next()) {
-    const auto s = siter.Value();
-    const auto weight = fst->Final(s);
-    if (weight != Weight::Zero()) fst->AddArc(s, Arc(0, 0, weight, start));
-  }
-  if (closure_type == CLOSURE_STAR) {
-    fst->ReserveStates(fst->NumStates() + 1);
-    const auto nstart = fst->AddState();
-    fst->SetStart(nstart);
-    fst->SetFinal(nstart, Weight::One());
-    if (start != kNoLabel) fst->AddArc(nstart, Arc(0, 0, Weight::One(), start));
-  }
-  fst->SetProperties(ClosureProperties(props, closure_type == CLOSURE_STAR),
-                     kFstProperties);
-}
-
-// Computes the concatenative closure. This version modifies its
-// RationalFst input.
-template <class Arc>
-void Closure(RationalFst<Arc> *fst, ClosureType closure_type) {
-  fst->GetMutableImpl()->AddClosure(closure_type);
-}
-
-struct ClosureFstOptions : RationalFstOptions {
-  ClosureType type;
-
-  ClosureFstOptions(const RationalFstOptions &opts,
-                    ClosureType type = CLOSURE_STAR)
-      : RationalFstOptions(opts), type(type) {}
-
-  explicit ClosureFstOptions(ClosureType type = CLOSURE_STAR) : type(type) {}
-};
-
-// Computes the concatenative closure. This version is a delayed FST. If an FST
-// transduces string x to y with weight a, then its closure transduces x to y
-// with weight a, xx to yy with weight Times(a, a), xxx to yyy with weight
-// Times(Times(a, a), a), etc. If closure_type == CLOSURE_STAR, then the empty
-// string is transduced to itself with weight Weight::One() as well.
-//
-// Complexity:
-//
-//   Time: O(v)
-//   Space: O(v)
-//
-// where v is the number of states visited. Constant time and space to visit an
-// input state or arc is assumed and exclusive of caching.
-template <class A>
-class ClosureFst : public RationalFst<A> {
- public:
-  using Arc = A;
-
-  ClosureFst(const Fst<Arc> &fst, ClosureType closure_type) {
-    GetMutableImpl()->InitClosure(fst, closure_type);
-  }
-
-  ClosureFst(const Fst<Arc> &fst, const ClosureFstOptions &opts)
-      : RationalFst<A>(opts) {
-    GetMutableImpl()->InitClosure(fst, opts.type);
-  }
-
-  // See Fst<>::Copy() for doc.
-  ClosureFst(const ClosureFst<Arc> &fst, bool safe = false)
-      : RationalFst<A>(fst, safe) {}
-
-  // Gets a copy of this ClosureFst. See Fst<>::Copy() for further doc.
-  ClosureFst<A> *Copy(bool safe = false) const override {
-    return new ClosureFst<A>(*this, safe);
-  }
-
- private:
-  using ImplToFst<internal::RationalFstImpl<Arc>>::GetImpl;
-  using ImplToFst<internal::RationalFstImpl<Arc>>::GetMutableImpl;
-};
-
-// Specialization for ClosureFst.
-template <class Arc>
-class StateIterator<ClosureFst<Arc>> : public StateIterator<RationalFst<Arc>> {
- public:
-  explicit StateIterator(const ClosureFst<Arc> &fst)
-      : StateIterator<RationalFst<Arc>>(fst) {}
-};
-
-// Specialization for ClosureFst.
-template <class Arc>
-class ArcIterator<ClosureFst<Arc>> : public ArcIterator<RationalFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const ClosureFst<Arc> &fst, StateId s)
-      : ArcIterator<RationalFst<Arc>>(fst, s) {}
-};
-
-// Useful alias when using StdArc.
-using StdClosureFst = ClosureFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_CLOSURE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compact-fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compact-fst.h
deleted file mode 100644
index df91895b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compact-fst.h
+++ /dev/null
@@ -1,1566 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// FST Class for memory-efficient representation of common types of
-// FSTs: linear automata, acceptors, unweighted FSTs, ...
-
-#ifndef FST_COMPACT_FST_H_
-#define FST_COMPACT_FST_H_
-
-#include <climits>
-#include <iterator>
-#include <memory>
-#include <tuple>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/expanded-fst.h>
-#include <fst/fst-decl.h>  // For optional argument declarations
-#include <fst/mapped-file.h>
-#include <fst/matcher.h>
-#include <fst/test-properties.h>
-#include <fst/util.h>
-
-
-namespace fst {
-
-struct CompactFstOptions : public CacheOptions {
-  // The default caching behaviour is to do no caching. Most compactors are
-  // cheap and therefore we save memory by not doing caching.
-  CompactFstOptions() : CacheOptions(true, 0) {}
-
-  explicit CompactFstOptions(const CacheOptions &opts) : CacheOptions(opts) {}
-};
-
-// New upcoming (Fst) Compactor interface - currently used internally
-// by CompactFstImpl.
-//
-// class Compactor {
-//  public:
-//   // Constructor from the Fst to be compacted.
-//   Compactor(const Fst<Arc> &fst, ...);
-//   // Copy constructor
-//   Compactor(const Compactor &compactor, bool safe = false)
-//   // Default constructor (optional, see comment below).
-//   Compactor();
-//
-//   // Returns the start state, number of states, and total number of arcs
-//   // of the compacted Fst
-//   StateId Start() const;
-//   StateId NumStates() const;
-//   size_t NumArcs() const;
-//
-//   // Accessor class for state attributes.
-//   class State {
-//    public:
-//     State();  // Required, corresponds to kNoStateId.
-//     State(const Compactor *c, StateId);  // Accessor for StateId 's'.
-//     StateId GetStateId() const;
-//     Weight Final() const;
-//     size_t NumArcs() const;
-//     Arc GetArc(size_t i) const;
-//   };
-//
-//   // Modifies 'state' accessor to provide access to state id 's'.
-//   void SetState(StateId s, State *state);
-//   // Tests whether 'fst' can be compacted by this compactor.
-//   bool IsCompatible(const Fst<A> &fst) const;
-//   // Return the properties that are always true for an fst
-//   // compacted using this compactor
-//   uint64 Properties() const;
-//   // Return a string identifying the type of compactor.
-//   static const string &Type();
-//   // Return true if an error has occured.
-//   bool Error() const;
-//   // Writes a compactor to a file.
-//   bool Write(std::ostream &strm, const FstWriteOptions &opts) const;
-//   // Reads a compactor from a file.
-//   static Compactor*Read(std::istream &strm, const FstReadOptions &opts,
-//                         const FstHeader &hdr);
-// };
-//
-
-// Old (Arc) Compactor Interface:
-//
-// The ArcCompactor class determines how arcs and final weights are compacted
-// and expanded.
-//
-// Final weights are treated as transitions to the superfinal state, i.e.,
-// ilabel = olabel = kNoLabel and nextstate = kNoStateId.
-//
-// There are two types of compactors:
-//
-// * Fixed out-degree compactors: 'compactor.Size()' returns a positive integer
-//   's'. An FST can be compacted by this compactor only if each state has
-//   exactly 's' outgoing transitions (counting a non-Zero() final weight as a
-//   transition). A typical example is a compactor for string FSTs, i.e.,
-//   's == 1'.
-//
-// * Variable out-degree compactors: 'compactor.Size() == -1'. There are no
-//   out-degree restrictions for these compactors.
-//
-// Interface:
-//
-// class ArcCompactor {
-//  public:
-//   // Element is the type of the compacted transitions.
-//   using Element = ...
-//
-//   // Returns the compacted representation of a transition 'arc'
-//   // at a state 's'.
-//   Element Compact(StateId s, const Arc &arc);
-//
-//   // Returns the transition at state 's' represented by the compacted
-//   // transition 'e'.
-//   Arc Expand(StateId s, const Element &e) const;
-//
-//   // Returns -1 for variable out-degree compactors, and the mandatory
-//   // out-degree otherwise.
-//   ssize_t Size() const;
-//
-//   // Tests whether an FST can be compacted by this compactor.
-//   bool Compatible(const Fst<A> &fst) const;
-//
-//   // Returns the properties that are always true for an FST compacted using
-//   // this compactor
-//   uint64 Properties() const;
-//
-//   // Returns a string identifying the type of compactor.
-//   static const string &Type();
-//
-//   // Writes a compactor to a file.
-//   bool Write(std::ostream &strm) const;
-//
-//   // Reads a compactor from a file.
-//   static ArcCompactor *Read(std::istream &strm);
-//
-//   // Default constructor (optional, see comment below).
-//   ArcCompactor();
-// };
-//
-// The default constructor is only required for FST_REGISTER to work (i.e.,
-// enabling Convert() and the command-line utilities to work with this new
-// compactor). However, a default constructor always needs to be specified for
-// this code to compile, but one can have it simply raise an error when called,
-// like so:
-//
-// Compactor::Compactor() {
-//   FSTERROR() << "Compactor: No default constructor";
-// }
-
-// Default implementation data for CompactFst, which can shared between
-// otherwise independent copies.
-//
-// The implementation contains two arrays: 'states_' and 'compacts_'.
-//
-// For fixed out-degree compactors, the 'states_' array is unallocated. The
-// 'compacts_' contains the compacted transitions. Its size is 'ncompacts_'.
-// The outgoing transitions at a given state are stored consecutively. For a
-// given state 's', its 'compactor.Size()' outgoing transitions (including
-// superfinal transition when 's' is final), are stored in position
-// ['s*compactor.Size()', '(s+1)*compactor.Size()').
-//
-// For variable out-degree compactors, the states_ array has size
-// 'nstates_ + 1' and contains pointers to positions into 'compacts_'. For a
-// given state 's', the compacted transitions of 's' are stored in positions
-// ['states_[s]', 'states_[s + 1]') in 'compacts_'. By convention,
-// 'states_[nstates_] == ncompacts_'.
-//
-// In both cases, the superfinal transitions (when 's' is final, i.e.,
-// 'Final(s) != Weight::Zero()') are stored first.
-//
-// The unsigned type U is used to represent indices into the compacts_ array.
-template <class Element, class Unsigned>
-class DefaultCompactStore {
- public:
-  DefaultCompactStore()
-      : states_(nullptr),
-        compacts_(nullptr),
-        nstates_(0),
-        ncompacts_(0),
-        narcs_(0),
-        start_(kNoStateId),
-        error_(false) {}
-
-  template <class Arc, class Compactor>
-  DefaultCompactStore(const Fst<Arc> &fst, const Compactor &compactor);
-
-  template <class Iterator, class Compactor>
-  DefaultCompactStore(const Iterator &begin, const Iterator &end,
-                      const Compactor &compactor);
-
-  ~DefaultCompactStore() {
-    if (!states_region_) delete[] states_;
-    if (!compacts_region_) delete[] compacts_;
-  }
-
-  template <class Compactor>
-  static DefaultCompactStore<Element, Unsigned> *Read(
-      std::istream &strm, const FstReadOptions &opts, const FstHeader &hdr,
-      const Compactor &compactor);
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const;
-
-  Unsigned States(ssize_t i) const { return states_[i]; }
-
-  const Element &Compacts(size_t i) const { return compacts_[i]; }
-
-  size_t NumStates() const { return nstates_; }
-
-  size_t NumCompacts() const { return ncompacts_; }
-
-  size_t NumArcs() const { return narcs_; }
-
-  ssize_t Start() const { return start_; }
-
-  bool Error() const { return error_; }
-
-  // Returns a string identifying the type of data storage container.
-  static const string &Type();
-
- private:
-  std::unique_ptr<MappedFile> states_region_;
-  std::unique_ptr<MappedFile> compacts_region_;
-  Unsigned *states_;
-  Element *compacts_;
-  size_t nstates_;
-  size_t ncompacts_;
-  size_t narcs_;
-  ssize_t start_;
-  bool error_;
-};
-
-template <class Element, class Unsigned>
-template <class Arc, class Compactor>
-DefaultCompactStore<Element, Unsigned>::DefaultCompactStore(
-    const Fst<Arc> &fst, const Compactor &compactor)
-    : states_(nullptr),
-      compacts_(nullptr),
-      nstates_(0),
-      ncompacts_(0),
-      narcs_(0),
-      start_(kNoStateId),
-      error_(false) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  start_ = fst.Start();
-  // Counts # of states and arcs.
-  StateId nfinals = 0;
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    ++nstates_;
-    const auto s = siter.Value();
-    for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-      ++narcs_;
-    }
-    if (fst.Final(s) != Weight::Zero()) ++nfinals;
-  }
-  if (compactor.Size() == -1) {
-    states_ = new Unsigned[nstates_ + 1];
-    ncompacts_ = narcs_ + nfinals;
-    compacts_ = new Element[ncompacts_];
-    states_[nstates_] = ncompacts_;
-  } else {
-    states_ = nullptr;
-    ncompacts_ = nstates_ * compactor.Size();
-    if ((narcs_ + nfinals) != ncompacts_) {
-      FSTERROR() << "DefaultCompactStore: Compactor incompatible with FST";
-      error_ = true;
-      return;
-    }
-    compacts_ = new Element[ncompacts_];
-  }
-  size_t pos = 0;
-  size_t fpos = 0;
-  for (size_t s = 0; s < nstates_; ++s) {
-    fpos = pos;
-    if (compactor.Size() == -1) states_[s] = pos;
-    if (fst.Final(s) != Weight::Zero()) {
-      compacts_[pos++] = compactor.Compact(
-          s, Arc(kNoLabel, kNoLabel, fst.Final(s), kNoStateId));
-    }
-    for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-      compacts_[pos++] = compactor.Compact(s, aiter.Value());
-    }
-    if ((compactor.Size() != -1) && ((pos - fpos) != compactor.Size())) {
-      FSTERROR() << "DefaultCompactStore: Compactor incompatible with FST";
-      error_ = true;
-      return;
-    }
-  }
-  if (pos != ncompacts_) {
-    FSTERROR() << "DefaultCompactStore: Compactor incompatible with FST";
-    error_ = true;
-    return;
-  }
-}
-
-template <class Element, class Unsigned>
-template <class Iterator, class Compactor>
-DefaultCompactStore<Element, Unsigned>::DefaultCompactStore(
-    const Iterator &begin, const Iterator &end, const Compactor &compactor)
-    : states_(nullptr),
-      compacts_(nullptr),
-      nstates_(0),
-      ncompacts_(0),
-      narcs_(0),
-      start_(kNoStateId),
-      error_(false) {
-  using Arc = typename Compactor::Arc;
-  using Weight = typename Arc::Weight;
-  if (compactor.Size() != -1) {
-    ncompacts_ = std::distance(begin, end);
-    if (compactor.Size() == 1) {
-      // For strings, allows implicit final weight. Empty input is the empty
-      // string.
-      if (ncompacts_ == 0) {
-        ++ncompacts_;
-      } else {
-        const auto arc =
-            compactor.Expand(ncompacts_ - 1, *(begin + (ncompacts_ - 1)));
-        if (arc.ilabel != kNoLabel) ++ncompacts_;
-      }
-    }
-    if (ncompacts_ % compactor.Size()) {
-      FSTERROR() << "DefaultCompactStore: Size of input container incompatible"
-                 << " with compactor";
-      error_ = true;
-      return;
-    }
-    if (ncompacts_ == 0) return;
-    start_ = 0;
-    nstates_ = ncompacts_ / compactor.Size();
-    compacts_ = new Element[ncompacts_];
-    size_t i = 0;
-    Iterator it = begin;
-    for (; it != end; ++it, ++i) {
-      compacts_[i] = *it;
-      if (compactor.Expand(i, *it).ilabel != kNoLabel) ++narcs_;
-    }
-    if (i < ncompacts_) {
-      compacts_[i] = compactor.Compact(
-          i, Arc(kNoLabel, kNoLabel, Weight::One(), kNoStateId));
-    }
-  } else {
-    if (std::distance(begin, end) == 0) return;
-    // Count # of states, arcs and compacts.
-    auto it = begin;
-    for (size_t i = 0; it != end; ++it, ++i) {
-      const auto arc = compactor.Expand(i, *it);
-      if (arc.ilabel != kNoLabel) {
-        ++narcs_;
-        ++ncompacts_;
-      } else {
-        ++nstates_;
-        if (arc.weight != Weight::Zero()) ++ncompacts_;
-      }
-    }
-    start_ = 0;
-    compacts_ = new Element[ncompacts_];
-    states_ = new Unsigned[nstates_ + 1];
-    states_[nstates_] = ncompacts_;
-    size_t i = 0;
-    size_t s = 0;
-    for (it = begin; it != end; ++it) {
-      const auto arc = compactor.Expand(i, *it);
-      if (arc.ilabel != kNoLabel) {
-        compacts_[i++] = *it;
-      } else {
-        states_[s++] = i;
-        if (arc.weight != Weight::Zero()) compacts_[i++] = *it;
-      }
-    }
-    if ((s != nstates_) || (i != ncompacts_)) {
-      FSTERROR() << "DefaultCompactStore: Ill-formed input container";
-      error_ = true;
-      return;
-    }
-  }
-}
-
-template <class Element, class Unsigned>
-template <class Compactor>
-DefaultCompactStore<Element, Unsigned>
-    *DefaultCompactStore<Element, Unsigned>::Read(std::istream &strm,
-                                                  const FstReadOptions &opts,
-                                                  const FstHeader &hdr,
-                                                  const Compactor &compactor) {
-  std::unique_ptr<DefaultCompactStore<Element, Unsigned>> data(
-      new DefaultCompactStore<Element, Unsigned>());
-  data->start_ = hdr.Start();
-  data->nstates_ = hdr.NumStates();
-  data->narcs_ = hdr.NumArcs();
-  if (compactor.Size() == -1) {
-    if ((hdr.GetFlags() & FstHeader::IS_ALIGNED) && !AlignInput(strm)) {
-      LOG(ERROR) << "DefaultCompactStore::Read: Alignment failed: "
-                 << opts.source;
-      return nullptr;
-    }
-    auto b = (data->nstates_ + 1) * sizeof(Unsigned);
-    data->states_region_.reset(MappedFile::Map(
-        &strm, opts.mode == FstReadOptions::MAP, opts.source, b));
-    if (!strm || !data->states_region_) {
-      LOG(ERROR) << "DefaultCompactStore::Read: Read failed: " << opts.source;
-      return nullptr;
-    }
-    data->states_ =
-        static_cast<Unsigned *>(data->states_region_->mutable_data());
-  } else {
-    data->states_ = nullptr;
-  }
-  data->ncompacts_ = compactor.Size() == -1 ? data->states_[data->nstates_]
-                                            : data->nstates_ * compactor.Size();
-  if ((hdr.GetFlags() & FstHeader::IS_ALIGNED) && !AlignInput(strm)) {
-    LOG(ERROR) << "DefaultCompactStore::Read: Alignment failed: "
-               << opts.source;
-    return nullptr;
-  }
-  size_t b = data->ncompacts_ * sizeof(Element);
-  data->compacts_region_.reset(
-      MappedFile::Map(&strm, opts.mode == FstReadOptions::MAP, opts.source, b));
-  if (!strm || !data->compacts_region_) {
-    LOG(ERROR) << "DefaultCompactStore::Read: Read failed: " << opts.source;
-    return nullptr;
-  }
-  data->compacts_ =
-      static_cast<Element *>(data->compacts_region_->mutable_data());
-  return data.release();
-}
-
-template <class Element, class Unsigned>
-bool DefaultCompactStore<Element, Unsigned>::Write(
-    std::ostream &strm, const FstWriteOptions &opts) const {
-  if (states_) {
-    if (opts.align && !AlignOutput(strm)) {
-      LOG(ERROR) << "DefaultCompactStore::Write: Alignment failed: "
-                 << opts.source;
-      return false;
-    }
-    strm.write(reinterpret_cast<char *>(states_),
-               (nstates_ + 1) * sizeof(Unsigned));
-  }
-  if (opts.align && !AlignOutput(strm)) {
-    LOG(ERROR) << "DefaultCompactStore::Write: Alignment failed: "
-               << opts.source;
-    return false;
-  }
-  strm.write(reinterpret_cast<char *>(compacts_), ncompacts_ * sizeof(Element));
-  strm.flush();
-  if (!strm) {
-    LOG(ERROR) << "DefaultCompactStore::Write: Write failed: " << opts.source;
-    return false;
-  }
-  return true;
-}
-
-template <class Element, class Unsigned>
-const string &DefaultCompactStore<Element, Unsigned>::Type() {
-  static const string *const type = new string("compact");
-  return *type;
-}
-
-template <class C, class U, class S> class DefaultCompactState;
-
-// Wraps an arc compactor and a compact store as a new Fst compactor.
-template <class C, class U,
-          class S = DefaultCompactStore<typename C::Element, U>>
-class DefaultCompactor {
- public:
-  using ArcCompactor = C;
-  using Unsigned = U;
-  using CompactStore = S;
-  using Element = typename C::Element;
-  using Arc = typename C::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using State = DefaultCompactState<C, U, S>;
-  friend State;
-
-  DefaultCompactor()
-      : arc_compactor_(nullptr), compact_store_(nullptr) {}
-
-  // Constructs from Fst.
-  DefaultCompactor(const Fst<Arc> &fst,
-                   std::shared_ptr<ArcCompactor> arc_compactor)
-      : arc_compactor_(std::move(arc_compactor)),
-        compact_store_(std::make_shared<S>(fst, *arc_compactor_)) {}
-
-  DefaultCompactor(const Fst<Arc> &fst,
-                   std::shared_ptr<DefaultCompactor<C, U, S>> compactor)
-      : arc_compactor_(compactor->arc_compactor_),
-        compact_store_(compactor->compact_store_ == nullptr ?
-                       std::make_shared<S>(fst, *arc_compactor_) :
-                       compactor->compact_store_) {}
-
-  // Constructs from CompactStore.
-  DefaultCompactor(std::shared_ptr<ArcCompactor> arc_compactor,
-                   std::shared_ptr<CompactStore> compact_store)
-      : arc_compactor_(std::move(arc_compactor)),
-        compact_store_(std::move(compact_store)) {}
-
-  // Constructs from set of compact elements (when arc_compactor.Size() != -1).
-  template <class Iterator>
-  DefaultCompactor(const Iterator &b, const Iterator &e,
-                   std::shared_ptr<C> arc_compactor)
-      : arc_compactor_(std::move(arc_compactor)),
-        compact_store_(std::make_shared<S>(b, e, *arc_compactor_)) {}
-
-  // Copy constructor.
-  DefaultCompactor(const DefaultCompactor<C, U, S> &compactor)
-      : arc_compactor_(std::make_shared<C>(*compactor.GetArcCompactor())),
-        compact_store_(compactor.SharedCompactStore()) {}
-
-  template <class OtherC>
-  explicit DefaultCompactor(const DefaultCompactor<OtherC, U, S> &compactor)
-      : arc_compactor_(std::make_shared<C>(*compactor.GetArcCompactor())),
-        compact_store_(compactor.SharedCompactStore()) {}
-
-  StateId Start() const { return compact_store_->Start(); }
-  StateId NumStates() const { return compact_store_->NumStates(); }
-  size_t NumArcs() const { return compact_store_->NumArcs(); }
-
-  void SetState(StateId s, State *state) const {
-    if (state->GetStateId() != s) state->Set(this, s);
-  }
-
-  static DefaultCompactor<C, U, S> *Read(std::istream &strm,
-                                         const FstReadOptions &opts,
-                                         const FstHeader &hdr) {
-    std::shared_ptr<C> arc_compactor(C::Read(strm));
-    if (arc_compactor == nullptr) return nullptr;
-    std::shared_ptr<S> compact_store(S::Read(strm, opts, hdr, *arc_compactor));
-    if (compact_store == nullptr) return nullptr;
-    return new DefaultCompactor<C, U, S>(arc_compactor, compact_store);
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
-    return arc_compactor_->Write(strm) && compact_store_->Write(strm, opts);
-  }
-
-  uint64 Properties() const { return arc_compactor_->Properties(); }
-
-  bool IsCompatible(const Fst<Arc> &fst) const {
-    return arc_compactor_->Compatible(fst);
-  }
-
-  bool Error() const { return compact_store_->Error(); }
-
-  bool HasFixedOutdegree() const { return arc_compactor_->Size() != -1; }
-
-  static const string &Type() {
-    static const string *const type = [] {
-      string type = "compact";
-      if (sizeof(U) != sizeof(uint32)) type += std::to_string(8 * sizeof(U));
-      type += "_";
-      type += C::Type();
-      if (CompactStore::Type() != "compact") {
-        type += "_";
-        type += CompactStore::Type();
-      }
-      return new string(type);
-    }();
-    return *type;
-  }
-
-  const ArcCompactor *GetArcCompactor() const { return arc_compactor_.get(); }
-  CompactStore *GetCompactStore() const { return compact_store_.get(); }
-
-  std::shared_ptr<ArcCompactor> SharedArcCompactor() const {
-    return arc_compactor_;
-  }
-
-  std::shared_ptr<CompactStore> SharedCompactStore() const {
-    return compact_store_;
-  }
-
-  // TODO(allauzen): remove dependencies on this method and make private.
-  Arc ComputeArc(StateId s, Unsigned i, uint32 f) const {
-    return arc_compactor_->Expand(s, compact_store_->Compacts(i), f);
-  }
-
- private:
-  std::pair<Unsigned, Unsigned> CompactsRange(StateId s) const {
-    std::pair<size_t, size_t> range;
-    if (HasFixedOutdegree()) {
-      range.first = s * arc_compactor_->Size();
-      range.second = arc_compactor_->Size();
-    } else {
-      range.first = compact_store_->States(s);
-      range.second = compact_store_->States(s + 1) - range.first;
-    }
-    return range;
-  }
-
- private:
-  std::shared_ptr<ArcCompactor> arc_compactor_;
-  std::shared_ptr<CompactStore> compact_store_;
-};
-
-// Default implementation of state attributes accessor class for
-// DefaultCompactor. Use of efficient specialization strongly encouraged.
-template <class C, class U, class S>
-class DefaultCompactState {
- public:
-  using Arc = typename C::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  DefaultCompactState() = default;
-
-  DefaultCompactState(const DefaultCompactor<C, U, S> *compactor, StateId s)
-      : compactor_(compactor),
-        s_(s),
-        range_(compactor->CompactsRange(s)),
-        has_final_(
-            range_.second != 0 &&
-            compactor->ComputeArc(s, range_.first,
-                                 kArcILabelValue).ilabel == kNoLabel) {
-    if (has_final_) {
-      ++range_.first;
-      --range_.second;
-    }
-  }
-
-  void Set(const DefaultCompactor<C, U, S> *compactor, StateId s) {
-    compactor_ = compactor;
-    s_ = s;
-    range_ = compactor->CompactsRange(s);
-    if (range_.second != 0 &&
-        compactor->ComputeArc(s, range_.first, kArcILabelValue).ilabel
-        == kNoLabel) {
-      has_final_ = true;
-      ++range_.first;
-      --range_.second;
-    } else {
-      has_final_ = false;
-    }
-  }
-
-  StateId GetStateId() const { return s_; }
-
-  Weight Final() const {
-    if (!has_final_) return Weight::Zero();
-    return compactor_->ComputeArc(s_, range_.first - 1, kArcWeightValue).weight;
-  }
-
-  size_t NumArcs() const { return range_.second; }
-
-  Arc GetArc(size_t i, uint32 f) const {
-    return compactor_->ComputeArc(s_, range_.first + i, f);
-  }
-
- private:
-  const DefaultCompactor<C, U, S> *compactor_ = nullptr;  // borrowed ref.
-  StateId s_ = kNoStateId;
-  std::pair<U, U> range_ = {0, 0};
-  bool has_final_ = false;
-};
-
-// Specialization for DefaultCompactStore.
-template <class C, class U>
-class DefaultCompactState<C, U, DefaultCompactStore<typename C::Element, U>> {
- public:
-  using Arc = typename C::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using CompactStore = DefaultCompactStore<typename C::Element, U>;
-
-  DefaultCompactState() = default;
-
-  DefaultCompactState(
-      const DefaultCompactor<C, U, CompactStore> *compactor, StateId s)
-      : arc_compactor_(compactor->GetArcCompactor()), s_(s) {
-    Init(compactor);
-  }
-
-  void Set(const DefaultCompactor<C, U, CompactStore> *compactor, StateId s) {
-    arc_compactor_ = compactor->GetArcCompactor();
-    s_ = s;
-    has_final_ = false;
-    Init(compactor);
-  }
-
-  StateId GetStateId() const { return s_; }
-
-  Weight Final() const {
-    if (!has_final_) return Weight::Zero();
-    return arc_compactor_->Expand(s_, *(compacts_ - 1), kArcWeightValue).weight;
-  }
-
-  size_t NumArcs() const { return num_arcs_; }
-
-  Arc GetArc(size_t i, uint32 f) const {
-    return arc_compactor_->Expand(s_, compacts_[i], f);
-  }
-
- private:
-  void Init(const DefaultCompactor<C, U, CompactStore> *compactor) {
-    const auto *store = compactor->GetCompactStore();
-    U offset;
-    if (!compactor->HasFixedOutdegree()) {  // Variable out-degree compactor.
-      offset = store->States(s_);
-      num_arcs_ = store->States(s_ + 1) - offset;
-    } else {  // Fixed out-degree compactor.
-      offset = s_ * arc_compactor_->Size();
-      num_arcs_ = arc_compactor_->Size();
-    }
-    if (num_arcs_ > 0) {
-      compacts_ = &(store->Compacts(offset));
-      if (arc_compactor_->Expand(s_, *compacts_, kArcILabelValue).ilabel
-          == kNoStateId) {
-        ++compacts_;
-        --num_arcs_;
-        has_final_ = true;
-      }
-    }
-  }
-
- private:
-  const C *arc_compactor_ = nullptr;               // Borrowed reference.
-  const typename C::Element *compacts_ = nullptr;  // Borrowed reference.
-  StateId s_ = kNoStateId;
-  U num_arcs_ = 0;
-  bool has_final_ = false;
-};
-
-template <class Arc, class ArcCompactor, class Unsigned, class CompactStore,
-          class CacheStore>
-class CompactFst;
-
-template <class F, class G>
-void Cast(const F &, G *);
-
-namespace internal {
-
-// Implementation class for CompactFst, which contains parametrizeable
-// Fst data storage (DefaultCompactStore by default) and Fst cache.
-template <class Arc, class C, class CacheStore = DefaultCacheStore<Arc>>
-class CompactFstImpl
-    : public CacheBaseImpl<typename CacheStore::State, CacheStore> {
- public:
-  using Weight = typename Arc::Weight;
-  using StateId = typename Arc::StateId;
-  using Compactor = C;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::Properties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-  using FstImpl<Arc>::WriteHeader;
-
-  using ImplBase = CacheBaseImpl<typename CacheStore::State, CacheStore>;
-  using ImplBase::PushArc;
-  using ImplBase::HasArcs;
-  using ImplBase::HasFinal;
-  using ImplBase::HasStart;
-  using ImplBase::SetArcs;
-  using ImplBase::SetFinal;
-  using ImplBase::SetStart;
-
-  CompactFstImpl()
-      : ImplBase(CompactFstOptions()),
-        compactor_() {
-    SetType(Compactor::Type());
-    SetProperties(kNullProperties | kStaticProperties);
-  }
-
-  CompactFstImpl(const Fst<Arc> &fst, std::shared_ptr<Compactor> compactor,
-                 const CompactFstOptions &opts)
-      : ImplBase(opts),
-        compactor_(std::make_shared<Compactor>(fst, compactor)) {
-    SetType(Compactor::Type());
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-    if (compactor_->Error()) SetProperties(kError, kError);
-    uint64 copy_properties = fst.Properties(kMutable, false) ?
-        fst.Properties(kCopyProperties, true):
-        CheckProperties(fst,
-                        kCopyProperties & ~kWeightedCycles & ~kUnweightedCycles,
-                        kCopyProperties);
-    if ((copy_properties & kError) || !compactor_->IsCompatible(fst)) {
-      FSTERROR() << "CompactFstImpl: Input Fst incompatible with compactor";
-      SetProperties(kError, kError);
-      return;
-    }
-    SetProperties(copy_properties | kStaticProperties);
-  }
-
-  CompactFstImpl(std::shared_ptr<Compactor> compactor,
-                 const CompactFstOptions &opts)
-      : ImplBase(opts),
-        compactor_(compactor) {
-    SetType(Compactor::Type());
-    SetProperties(kStaticProperties | compactor_->Properties());
-    if (compactor_->Error()) SetProperties(kError, kError);
-  }
-
-  CompactFstImpl(const CompactFstImpl<Arc, Compactor, CacheStore> &impl)
-      : ImplBase(impl),
-        compactor_(impl.compactor_ == nullptr ?
-                   std::make_shared<Compactor>() :
-                   std::make_shared<Compactor>(*impl.compactor_)) {
-    SetType(impl.Type());
-    SetProperties(impl.Properties());
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  // Allows to change the cache store from OtherI to I.
-  template <class OtherCacheStore>
-  CompactFstImpl(const CompactFstImpl<Arc, Compactor, OtherCacheStore> &impl)
-      : ImplBase(CacheOptions(impl.GetCacheGc(), impl.GetCacheLimit())),
-        compactor_(impl.compactor_ == nullptr ?
-                   std::make_shared<Compactor>() :
-                   std::make_shared<Compactor>(*impl.compactor_)) {
-    SetType(impl.Type());
-    SetProperties(impl.Properties());
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  StateId Start() {
-    if (!HasStart()) SetStart(compactor_->Start());
-    return ImplBase::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (HasFinal(s)) return ImplBase::Final(s);
-    compactor_->SetState(s, &state_);
-    return state_.Final();
-  }
-
-  StateId NumStates() const {
-    if (Properties(kError)) return 0;
-    return compactor_->NumStates();
-  }
-
-  size_t NumArcs(StateId s) {
-    if (HasArcs(s)) return ImplBase::NumArcs(s);
-    compactor_->SetState(s, &state_);
-    return state_.NumArcs();
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s) && !Properties(kILabelSorted)) Expand(s);
-    if (HasArcs(s)) return ImplBase::NumInputEpsilons(s);
-    return CountEpsilons(s, false);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s) && !Properties(kOLabelSorted)) Expand(s);
-    if (HasArcs(s)) return ImplBase::NumOutputEpsilons(s);
-    return CountEpsilons(s, true);
-  }
-
-  size_t CountEpsilons(StateId s, bool output_epsilons) {
-    compactor_->SetState(s, &state_);
-    const uint32 f = output_epsilons ? kArcOLabelValue : kArcILabelValue;
-    size_t num_eps = 0;
-    for (size_t i = 0; i < state_.NumArcs(); ++i) {
-      const auto& arc = state_.GetArc(i, f);
-      const auto label = output_epsilons ? arc.olabel : arc.ilabel;
-      if (label == 0)
-        ++num_eps;
-      else if (label > 0)
-        break;
-    }
-    return num_eps;
-  }
-
-  static CompactFstImpl<Arc, Compactor, CacheStore> *Read(
-      std::istream &strm, const FstReadOptions &opts) {
-    std::unique_ptr<CompactFstImpl<Arc, Compactor, CacheStore>> impl(
-      new CompactFstImpl<Arc, Compactor, CacheStore>());
-    FstHeader hdr;
-    if (!impl->ReadHeader(strm, opts, kMinFileVersion, &hdr)) {
-      return nullptr;
-    }
-    // Ensures compatibility.
-    if (hdr.Version() == kAlignedFileVersion) {
-      hdr.SetFlags(hdr.GetFlags() | FstHeader::IS_ALIGNED);
-    }
-    impl->compactor_ = std::shared_ptr<Compactor>(
-        Compactor::Read(strm, opts, hdr));
-    if (!impl->compactor_) {
-      return nullptr;
-    }
-    return impl.release();
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
-    FstHeader hdr;
-    hdr.SetStart(compactor_->Start());
-    hdr.SetNumStates(compactor_->NumStates());
-    hdr.SetNumArcs(compactor_->NumArcs());
-    // Ensures compatibility.
-    const auto file_version = opts.align ? kAlignedFileVersion : kFileVersion;
-    WriteHeader(strm, opts, file_version, &hdr);
-    return compactor_->Write(strm, opts);
-  }
-
-  // Provides information needed for generic state iterator.
-  void InitStateIterator(StateIteratorData<Arc> *data) const {
-    data->base = nullptr;
-    data->nstates = compactor_->NumStates();
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    ImplBase::InitArcIterator(s, data);
-  }
-
-  void Expand(StateId s) {
-    compactor_->SetState(s, &state_);
-    for (size_t i = 0; i < state_.NumArcs(); ++i)
-      PushArc(s, state_.GetArc(i, kArcValueFlags));
-    SetArcs(s);
-    if (!HasFinal(s)) SetFinal(s, state_.Final());
-  }
-
-  const Compactor *GetCompactor() const { return compactor_.get(); }
-  std::shared_ptr<Compactor> SharedCompactor() const { return compactor_; }
-  void SetCompactor(std::shared_ptr<Compactor> compactor) {
-    // TODO(allauzen): is this correct? is this needed?
-    // TODO(allauzen): consider removing and forcing this through direct calls
-    // to compactor.
-    compactor_ = compactor;
-  }
-
-  // Properties always true of this FST class.
-  static constexpr uint64 kStaticProperties = kExpanded;
-
- protected:
-  template <class OtherArc, class OtherCompactor, class OtherCacheStore>
-  explicit CompactFstImpl(
-    const CompactFstImpl<OtherArc, OtherCompactor, OtherCacheStore> &impl)
-    : compactor_(std::make_shared<Compactor>(*impl.GetCompactor())) {
-    SetType(impl.Type());
-    SetProperties(impl.Properties());
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
- private:
-  // Allows access during write.
-  template <class AnyArc, class ArcCompactor, class Unsigned,
-            class CompactStore, class AnyCacheStore>
-  friend class ::fst::CompactFst;  // allow access during write.
-
-  // Current unaligned file format version.
-  static constexpr int kFileVersion = 2;
-  // Current aligned file format version.
-  static constexpr int kAlignedFileVersion = 1;
-  // Minimum file format version supported.
-  static constexpr int kMinFileVersion = 1;
-
-  std::shared_ptr<Compactor> compactor_;
-  typename Compactor::State state_;
-};
-
-template <class Arc, class Compactor, class CacheStore>
-constexpr uint64 CompactFstImpl<Arc, Compactor, CacheStore>::kStaticProperties;
-
-template <class Arc, class Compactor, class CacheStore>
-constexpr int CompactFstImpl<Arc, Compactor, CacheStore>::kFileVersion;
-
-template <class Arc, class Compactor, class CacheStore>
-constexpr int CompactFstImpl<Arc, Compactor, CacheStore>::kAlignedFileVersion;
-
-template <class Arc, class Compactor, class CacheStore>
-constexpr int CompactFstImpl<Arc, Compactor, CacheStore>::kMinFileVersion;
-
-}  // namespace internal
-
-// This class attaches interface to implementation and handles reference
-// counting, delegating most methods to ImplToExpandedFst. The Unsigned type
-// is used to represent indices into the compact arc array. (Template
-// argument defaults are declared in fst-decl.h.)
-template <class A, class ArcCompactor, class Unsigned, class CompactStore,
-          class CacheStore>
-class CompactFst
-    : public ImplToExpandedFst<internal::CompactFstImpl<
-          A,
-          DefaultCompactor<ArcCompactor, Unsigned, CompactStore>,
-          CacheStore>> {
- public:
-  template <class F, class G>
-  void friend Cast(const F &, G *);
-
-  using Arc = A;
-  using StateId = typename A::StateId;
-  using Compactor = DefaultCompactor<ArcCompactor, Unsigned, CompactStore>;
-  using Impl = internal::CompactFstImpl<A, Compactor, CacheStore>;
-  using Store = CacheStore;  // for CacheArcIterator
-
-  friend class StateIterator<
-      CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>>;
-  friend class ArcIterator<
-      CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>>;
-
-  CompactFst() : ImplToExpandedFst<Impl>(std::make_shared<Impl>()) {}
-
-  // If data is not nullptr, it is assumed to be already initialized.
-  explicit CompactFst(
-      const Fst<A> &fst,
-      const ArcCompactor &compactor = ArcCompactor(),
-      const CompactFstOptions &opts = CompactFstOptions(),
-      std::shared_ptr<CompactStore> data = std::shared_ptr<CompactStore>())
-      : ImplToExpandedFst<Impl>(
-            std::make_shared<Impl>(
-                fst,
-                std::make_shared<Compactor>(
-                    std::make_shared<ArcCompactor>(compactor), data),
-                opts)) {}
-
-  // If data is not nullptr, it is assumed to be already initialized.
-  CompactFst(
-      const Fst<Arc> &fst,
-      std::shared_ptr<ArcCompactor> compactor,
-      const CompactFstOptions &opts = CompactFstOptions(),
-      std::shared_ptr<CompactStore> data = std::shared_ptr<CompactStore>())
-      : ImplToExpandedFst<Impl>(
-            std::make_shared<Impl>(fst,
-                                   std::make_shared<Compactor>(compactor, data),
-                                   opts)) {}
-
-  // The following 2 constructors take as input two iterators delimiting a set
-  // of (already) compacted transitions, starting with the transitions out of
-  // the initial state. The format of the input differs for fixed out-degree
-  // and variable out-degree compactors.
-  //
-  // - For fixed out-degree compactors, the final weight (encoded as a
-  // compacted transition) needs to be given only for final states. All strings
-  // (compactor of size 1) will be assume to be terminated by a final state
-  // even when the final state is not implicitely given.
-  //
-  // - For variable out-degree compactors, the final weight (encoded as a
-  // compacted transition) needs to be given for all states and must appeared
-  // first in the list (for state s, final weight of s, followed by outgoing
-  // transitons in s).
-  //
-  // These 2 constructors allows the direct construction of a CompactFst
-  // without first creating a more memory-hungry regular FST. This is useful
-  // when memory usage is severely constrained.
-  template <class Iterator>
-  explicit CompactFst(const Iterator &begin, const Iterator &end,
-                      const ArcCompactor &compactor = ArcCompactor(),
-                      const CompactFstOptions &opts = CompactFstOptions())
-      : ImplToExpandedFst<Impl>(
-            std::make_shared<Impl>(
-                std::make_shared<Compactor>(
-                    begin, end, std::make_shared<ArcCompactor>(compactor)),
-                opts)) {}
-
-  template <class Iterator>
-  CompactFst(const Iterator &begin, const Iterator &end,
-             std::shared_ptr<ArcCompactor> compactor,
-             const CompactFstOptions &opts = CompactFstOptions())
-      : ImplToExpandedFst<Impl>(
-            std::make_shared<Impl>(
-                std::make_shared<Compactor>(begin, end, compactor), opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  CompactFst(
-      const CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>
-      &fst,
-      bool safe = false)
-      : ImplToExpandedFst<Impl>(fst, safe) {}
-
-  // Get a copy of this CompactFst. See Fst<>::Copy() for further doc.
-  CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore> *Copy(
-      bool safe = false) const override {
-    return new CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>(
-        *this, safe);
-  }
-
-  // Read a CompactFst from an input stream; return nullptr on error
-  static CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore> *Read(
-      std::istream &strm, const FstReadOptions &opts) {
-    auto *impl = Impl::Read(strm, opts);
-    return impl ? new CompactFst<A, ArcCompactor, Unsigned, CompactStore,
-                                 CacheStore>(std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  // Read a CompactFst from a file; return nullptr on error
-  // Empty filename reads from standard input
-  static CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore> *Read(
-      const string &filename) {
-    auto *impl = ImplToExpandedFst<Impl>::Read(filename);
-    return impl ? new CompactFst<A, ArcCompactor, Unsigned, CompactStore,
-                                 CacheStore>(std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return GetImpl()->Write(strm, opts);
-  }
-
-  bool Write(const string &filename) const override {
-    return Fst<Arc>::WriteFile(filename);
-  }
-
-  template <class FST>
-  static bool WriteFst(const FST &fst, const ArcCompactor &compactor,
-                       std::ostream &strm, const FstWriteOptions &opts);
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const override {
-    GetImpl()->InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
-  MatcherBase<Arc> *InitMatcher(MatchType match_type) const override {
-    return new SortedMatcher<
-        CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>>(
-        *this, match_type);
-  }
-
-  template <class Iterator>
-  void SetCompactElements(const Iterator &b, const Iterator &e) {
-    GetMutableImpl()->SetCompactor(std::make_shared<Compactor>(
-        b, e, std::make_shared<ArcCompactor>()));
-  }
-
- private:
-  using ImplToFst<Impl, ExpandedFst<Arc>>::GetImpl;
-  using ImplToFst<Impl, ExpandedFst<Arc>>::GetMutableImpl;
-
-  explicit CompactFst(std::shared_ptr<Impl> impl)
-      : ImplToExpandedFst<Impl>(impl) {}
-
-  // Use overloading to extract the type of the argument.
-  static Impl *GetImplIfCompactFst(
-      const CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>
-          &compact_fst) {
-    return compact_fst.GetImpl();
-  }
-
-  // This does not give privileged treatment to subclasses of CompactFst.
-  template <typename NonCompactFst>
-  static Impl *GetImplIfCompactFst(const NonCompactFst &fst) {
-    return nullptr;
-  }
-
-  CompactFst &operator=(const CompactFst &fst) = delete;
-};
-
-// Writes FST in Compact format, with a possible pass over the machine before
-// writing to compute the number of states and arcs.
-template <class A, class ArcCompactor, class Unsigned, class CompactStore,
-          class CacheStore>
-template <class FST>
-bool CompactFst<A, ArcCompactor, Unsigned, CompactStore, CacheStore>::WriteFst(
-    const FST &fst, const ArcCompactor &compactor, std::ostream &strm,
-    const FstWriteOptions &opts) {
-  using Arc = A;
-  using Weight = typename A::Weight;
-  using Element = typename ArcCompactor::Element;
-  const auto file_version =
-      opts.align ? Impl::kAlignedFileVersion : Impl::kFileVersion;
-  size_t num_arcs = -1;
-  size_t num_states = -1;
-  auto first_pass_compactor = compactor;
-  if (auto *impl = GetImplIfCompactFst(fst)) {
-    num_arcs = impl->GetCompactor()->GetCompactStore()->NumArcs();
-    num_states = impl->GetCompactor()->GetCompactStore()->NumStates();
-    first_pass_compactor = *impl->GetCompactor()->GetArcCompactor();
-  } else {
-    // A first pass is needed to compute the state of the compactor, which
-    // is saved ahead of the rest of the data structures. This unfortunately
-    // means forcing a complete double compaction when writing in this format.
-    // TODO(allauzen): eliminate mutable state from compactors.
-    num_arcs = 0;
-    num_states = 0;
-    for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      ++num_states;
-      if (fst.Final(s) != Weight::Zero()) {
-        first_pass_compactor.Compact(
-            s, Arc(kNoLabel, kNoLabel, fst.Final(s), kNoStateId));
-      }
-      for (ArcIterator<FST> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-        ++num_arcs;
-        first_pass_compactor.Compact(s, aiter.Value());
-      }
-    }
-  }
-  FstHeader hdr;
-  hdr.SetStart(fst.Start());
-  hdr.SetNumStates(num_states);
-  hdr.SetNumArcs(num_arcs);
-  string type = "compact";
-  if (sizeof(Unsigned) != sizeof(uint32)) {
-    type += std::to_string(CHAR_BIT * sizeof(Unsigned));
-  }
-  type += "_";
-  type += ArcCompactor::Type();
-  if (CompactStore::Type() != "compact") {
-    type += "_";
-    type += CompactStore::Type();
-  }
-  const auto copy_properties = fst.Properties(kCopyProperties, true);
-  if ((copy_properties & kError) || !compactor.Compatible(fst)) {
-    FSTERROR() << "Fst incompatible with compactor";
-    return false;
-  }
-  uint64 properties = copy_properties | Impl::kStaticProperties;
-  internal::FstImpl<Arc>::WriteFstHeader(fst, strm, opts, file_version, type,
-                                         properties, &hdr);
-  first_pass_compactor.Write(strm);
-  if (first_pass_compactor.Size() == -1) {
-    if (opts.align && !AlignOutput(strm)) {
-      LOG(ERROR) << "CompactFst::Write: Alignment failed: " << opts.source;
-      return false;
-    }
-    Unsigned compacts = 0;
-    for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      strm.write(reinterpret_cast<const char *>(&compacts), sizeof(compacts));
-      if (fst.Final(s) != Weight::Zero()) {
-        ++compacts;
-      }
-      compacts += fst.NumArcs(s);
-    }
-    strm.write(reinterpret_cast<const char *>(&compacts), sizeof(compacts));
-  }
-  if (opts.align && !AlignOutput(strm)) {
-    LOG(ERROR) << "Could not align file during write after writing states";
-  }
-  const auto &second_pass_compactor = compactor;
-  Element element;
-  for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    if (fst.Final(s) != Weight::Zero()) {
-      element = second_pass_compactor.Compact(
-          s, A(kNoLabel, kNoLabel, fst.Final(s), kNoStateId));
-      strm.write(reinterpret_cast<const char *>(&element), sizeof(element));
-    }
-    for (ArcIterator<FST> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-      element = second_pass_compactor.Compact(s, aiter.Value());
-      strm.write(reinterpret_cast<const char *>(&element), sizeof(element));
-    }
-  }
-  strm.flush();
-  if (!strm) {
-    LOG(ERROR) << "CompactFst write failed: " << opts.source;
-    return false;
-  }
-  return true;
-}
-
-// Specialization for CompactFst; see generic version in fst.h for sample
-// usage (but use the CompactFst type!). This version should inline.
-template <class Arc, class ArcCompactor, class Unsigned, class CompactStore,
-          class CacheStore>
-class StateIterator<
-    CompactFst<Arc, ArcCompactor, Unsigned, CompactStore, CacheStore>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  explicit StateIterator(
-      const CompactFst<Arc, ArcCompactor, Unsigned, CompactStore,
-                       CacheStore> &fst)
-      : nstates_(fst.GetImpl()->NumStates()), s_(0) {}
-
-  bool Done() const { return s_ >= nstates_; }
-
-  StateId Value() const { return s_; }
-
-  void Next() { ++s_; }
-
-  void Reset() { s_ = 0; }
-
- private:
-  StateId nstates_;
-  StateId s_;
-};
-
-// Specialization for CompactFst. Never caches,
-// always iterates over the underlying compact elements.
-template <class Arc, class ArcCompactor, class Unsigned,
-          class CompactStore, class CacheStore>
-class ArcIterator<CompactFst<
-    Arc, ArcCompactor, Unsigned, CompactStore, CacheStore>> {
- public:
-  using StateId = typename Arc::StateId;
-  using Element = typename ArcCompactor::Element;
-  using Compactor = DefaultCompactor<ArcCompactor, Unsigned, CompactStore>;
-  using State = typename Compactor::State;
-
-  ArcIterator(const CompactFst<Arc, ArcCompactor, Unsigned, CompactStore,
-                               CacheStore> &fst,
-              StateId s)
-      : state_(fst.GetImpl()->GetCompactor(), s),
-        pos_(0),
-        flags_(kArcValueFlags) {}
-
-  bool Done() const { return pos_ >= state_.NumArcs(); }
-
-  const Arc &Value() const {
-    arc_ = state_.GetArc(pos_, flags_);
-    return arc_;
-  }
-
-  void Next() { ++pos_; }
-
-  size_t Position() const { return pos_; }
-
-  void Reset() { pos_ = 0; }
-
-  void Seek(size_t pos) { pos_ = pos; }
-
-  uint32 Flags() const { return flags_; }
-
-  void SetFlags(uint32 f, uint32 m) {
-    flags_ &= ~m;
-    flags_ |= (f & kArcValueFlags);
-  }
-
- private:
-  State state_;
-  size_t pos_;
-  mutable Arc arc_;
-  uint32 flags_;
-};
-
-// ArcCompactor for unweighted string FSTs.
-template <class A>
-class StringCompactor {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Element = Label;
-
-  Element Compact(StateId s, const Arc &arc) const { return arc.ilabel; }
-
-  Arc Expand(StateId s, const Element &p, uint32 f = kArcValueFlags) const {
-    return Arc(p, p, Weight::One(), p != kNoLabel ? s + 1 : kNoStateId);
-  }
-
-  constexpr ssize_t Size() const { return 1; }
-
-  constexpr uint64 Properties() const {
-    return kString | kAcceptor | kUnweighted;
-  }
-
-  bool Compatible(const Fst<Arc> &fst) const {
-    const auto props = Properties();
-    return fst.Properties(props, true) == props;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string("string");
-    return *type;
-  }
-
-  bool Write(std::ostream &strm) const { return true; }
-
-  static StringCompactor *Read(std::istream &strm) {
-    return new StringCompactor;
-  }
-};
-
-// ArcCompactor for weighted string FSTs.
-template <class A>
-class WeightedStringCompactor {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Element = std::pair<Label, Weight>;
-
-  Element Compact(StateId s, const Arc &arc) const {
-    return std::make_pair(arc.ilabel, arc.weight);
-  }
-
-  Arc Expand(StateId s, const Element &p, uint32 f = kArcValueFlags) const {
-    return Arc(p.first, p.first, p.second,
-               p.first != kNoLabel ? s + 1 : kNoStateId);
-  }
-
-  constexpr ssize_t Size() const { return 1; }
-
-  constexpr uint64 Properties() const { return kString | kAcceptor; }
-
-  bool Compatible(const Fst<Arc> &fst) const {
-    const auto props = Properties();
-    return fst.Properties(props, true) == props;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string("weighted_string");
-    return *type;
-  }
-
-  bool Write(std::ostream &strm) const { return true; }
-
-  static WeightedStringCompactor *Read(std::istream &strm) {
-    return new WeightedStringCompactor;
-  }
-};
-
-// ArcCompactor for unweighted acceptor FSTs.
-template <class A>
-class UnweightedAcceptorCompactor {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Element = std::pair<Label, StateId>;
-
-  Element Compact(StateId s, const Arc &arc) const {
-    return std::make_pair(arc.ilabel, arc.nextstate);
-  }
-
-  Arc Expand(StateId s, const Element &p, uint32 f = kArcValueFlags) const {
-    return Arc(p.first, p.first, Weight::One(), p.second);
-  }
-
-  constexpr ssize_t Size() const { return -1; }
-
-  constexpr uint64 Properties() const { return kAcceptor | kUnweighted; }
-
-  bool Compatible(const Fst<Arc> &fst) const {
-    const auto props = Properties();
-    return fst.Properties(props, true) == props;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string("unweighted_acceptor");
-    return *type;
-  }
-
-  bool Write(std::ostream &strm) const { return true; }
-
-  static UnweightedAcceptorCompactor *Read(std::istream &istrm) {
-    return new UnweightedAcceptorCompactor;
-  }
-};
-
-// ArcCompactor for weighted acceptor FSTs.
-template <class A>
-class AcceptorCompactor {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Element = std::pair<std::pair<Label, Weight>, StateId>;
-
-  Element Compact(StateId s, const Arc &arc) const {
-    return std::make_pair(std::make_pair(arc.ilabel, arc.weight),
-                          arc.nextstate);
-  }
-
-  Arc Expand(StateId s, const Element &p, uint32 f = kArcValueFlags) const {
-    return Arc(p.first.first, p.first.first, p.first.second, p.second);
-  }
-
-  constexpr ssize_t Size() const { return -1; }
-
-  constexpr uint64 Properties() const { return kAcceptor; }
-
-  bool Compatible(const Fst<Arc> &fst) const {
-    const auto props = Properties();
-    return fst.Properties(props, true) == props;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string("acceptor");
-    return *type;
-  }
-
-  bool Write(std::ostream &strm) const { return true; }
-
-  static AcceptorCompactor *Read(std::istream &strm) {
-    return new AcceptorCompactor;
-  }
-};
-
-// ArcCompactor for unweighted FSTs.
-template <class A>
-class UnweightedCompactor {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Element = std::pair<std::pair<Label, Label>, StateId>;
-
-  Element Compact(StateId s, const Arc &arc) const {
-    return std::make_pair(std::make_pair(arc.ilabel, arc.olabel),
-                          arc.nextstate);
-  }
-
-  Arc Expand(StateId s, const Element &p, uint32 f = kArcValueFlags) const {
-    return Arc(p.first.first, p.first.second, Weight::One(), p.second);
-  }
-
-  constexpr ssize_t Size() const { return -1; }
-
-  constexpr uint64 Properties() const { return kUnweighted; }
-
-  bool Compatible(const Fst<Arc> &fst) const {
-    const auto props = Properties();
-    return fst.Properties(props, true) == props;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string("unweighted");
-    return *type;
-  }
-
-  bool Write(std::ostream &strm) const { return true; }
-
-  static UnweightedCompactor *Read(std::istream &strm) {
-    return new UnweightedCompactor;
-  }
-};
-
-template <class Arc, class Unsigned /* = uint32 */>
-using CompactStringFst = CompactFst<Arc, StringCompactor<Arc>, Unsigned>;
-
-template <class Arc, class Unsigned /* = uint32 */>
-using CompactWeightedStringFst =
-    CompactFst<Arc, WeightedStringCompactor<Arc>, Unsigned>;
-
-template <class Arc, class Unsigned /* = uint32 */>
-using CompactAcceptorFst = CompactFst<Arc, AcceptorCompactor<Arc>, Unsigned>;
-
-template <class Arc, class Unsigned /* = uint32 */>
-using CompactUnweightedFst =
-    CompactFst<Arc, UnweightedCompactor<Arc>, Unsigned>;
-
-template <class Arc, class Unsigned /* = uint32 */>
-using CompactUnweightedAcceptorFst =
-    CompactFst<Arc, UnweightedAcceptorCompactor<Arc>, Unsigned>;
-
-using StdCompactStringFst = CompactStringFst<StdArc, uint32>;
-
-using StdCompactWeightedStringFst = CompactWeightedStringFst<StdArc, uint32>;
-
-using StdCompactAcceptorFst = CompactAcceptorFst<StdArc, uint32>;
-
-using StdCompactUnweightedFst = CompactUnweightedFst<StdArc, uint32>;
-
-using StdCompactUnweightedAcceptorFst =
-    CompactUnweightedAcceptorFst<StdArc, uint32>;
-
-}  // namespace fst
-
-#endif  // FST_COMPACT_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compat.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compat.h
deleted file mode 100644
index 73ed5737..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compat.h
+++ /dev/null
@@ -1,130 +0,0 @@
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_LIB_COMPAT_H_
-#define FST_LIB_COMPAT_H_
-
-#include <climits>
-#include <cstdlib>
-#include <cstring>
-#include <iostream>
-#include <string>
-#include <vector>
-
-// Makes copy constructor and operator= private
-// Deprecated: now just use =delete.
-#define DISALLOW_COPY_AND_ASSIGN(type)    \
-  type(const type&);                      \
-  void operator=(const type&)
-
-#if defined(__GNUC__) || defined(__clang__)
-#define OPENFST_DEPRECATED(message) __attribute__((deprecated(message)))
-#elif defined(_MSC_VER)
-#define OPENFST_DEPRECATED(message) __declspec(deprecated(message))
-#else
-#define OPENFST_DEPRECATED(message)
-#endif
-
-#include <fst/config.h>
-#include <fst/types.h>
-#include <fst/lock.h>
-#include <fst/flags.h>
-#include <fst/log.h>
-#include <fst/icu.h>
-
-using std::string;
-
-void FailedNewHandler();
-
-#ifdef _MSC_VER
-#include <intrin.h>
-const char* basename(const char* path);
-#define __builtin_popcount __popcnt
-
-#ifdef _M_X64
-// Using 64-bit MSVC intrinsics.
-#define __builtin_popcountll __popcnt64
-inline unsigned int __builtin_ctzll(std::uint64_t w) {
-  unsigned long v;
-  return _BitScanForward64(&v, std::uint32_t(w)) ? v : 0;
-}
-#else
-// Using 32-bit MSVC intrinsics.
-inline unsigned int __builtin_popcountll(std::uint64_t w) {
-  return __popcnt(std::uint32_t(w)) + __popcnt(std::uint32_t(w >> 32));
-}
-inline unsigned int __builtin_ctzll(std::uint64_t w) {
-  unsigned long v;
-  return (_BitScanForward(&v, std::uint32_t(w)) ? v :
-          _BitScanForward(&v, std::uint32_t(w >> 32)) ? v + 32 : 0);
-}
-#endif  // _M_X64
-#endif  // _MSC_VER
-
-namespace fst {
-
-// Downcasting.
-template<typename To, typename From>
-inline To down_cast(From* f) { return static_cast<To>(f); }
-
-// Bitcasting.
-template <class Dest, class Source>
-inline Dest bit_cast(const Source &source) {
-  static_assert(sizeof(Dest) == sizeof(Source),
-                "Bitcasting unsafe for specified types");
-  Dest dest;
-  memcpy(&dest, &source, sizeof(dest));
-  return dest;
-}
-
-// Check sums
-class CheckSummer {
- public:
-  CheckSummer() : count_(0) {
-    check_sum_.resize(kCheckSumLength, '\0');
-  }
-
-  void Reset() {
-    count_ = 0;
-    for (int i = 0; i < kCheckSumLength; ++i) check_sum_[i] = '\0';
-  }
-
-  void Update(void const *data, int size) {
-    const char *p = reinterpret_cast<const char *>(data);
-    for (int i = 0; i < size; ++i) {
-      check_sum_[(count_++) % kCheckSumLength] ^= p[i];
-    }
-  }
-
-  void Update(string const &data) {
-    for (int i = 0; i < data.size(); ++i) {
-      check_sum_[(count_++) % kCheckSumLength] ^= data[i];
-    }
-  }
-
-  string Digest() { return check_sum_; }
-
- private:
-  static const int kCheckSumLength = 32;
-  int count_;
-  string check_sum_;
-
-  CheckSummer(const CheckSummer &) = delete;
-  CheckSummer &operator=(const CheckSummer &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_LIB_COMPAT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/complement.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/complement.h
deleted file mode 100644
index 64eebc03..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/complement.h
+++ /dev/null
@@ -1,277 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to complement an FST.
-
-#ifndef FST_COMPLEMENT_H_
-#define FST_COMPLEMENT_H_
-
-#include <algorithm>
-#include <string>
-#include <vector>
-#include <fst/log.h>
-
-#include <fst/fst.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-template <class Arc>
-class ComplementFst;
-
-namespace internal {
-
-// Implementation of delayed ComplementFst. The algorithm used completes the
-// (deterministic) FSA and then exchanges final and non-final states.
-// Completion, i.e. ensuring that all labels can be read from every state, is
-// accomplished by using ρ-labels, which match all labels that are otherwise
-// not found leaving a state. The first state in the output is reserved to be a
-// new state that is the destination of all ρ-labels. Each remaining output
-// state s corresponds to input state s - 1. The first arc in the output at
-// these states is the ρ-label, the remaining arcs correspond to the input
-// arcs.
-template <class A>
-class ComplementFstImpl : public FstImpl<A> {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<A>::SetType;
-  using FstImpl<A>::SetProperties;
-  using FstImpl<A>::SetInputSymbols;
-  using FstImpl<A>::SetOutputSymbols;
-
-  friend class StateIterator<ComplementFst<Arc>>;
-  friend class ArcIterator<ComplementFst<Arc>>;
-
-  explicit ComplementFstImpl(const Fst<Arc> &fst) : fst_(fst.Copy()) {
-    SetType("complement");
-    uint64 props = fst.Properties(kILabelSorted, false);
-    SetProperties(ComplementProperties(props), kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  ComplementFstImpl(const ComplementFstImpl<Arc> &impl)
-      : fst_(impl.fst_->Copy()) {
-    SetType("complement");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  StateId Start() const {
-    if (Properties(kError)) return kNoStateId;
-    auto start = fst_->Start();
-    return start != kNoStateId ? start + 1 : 0;
-  }
-
-  // Exchange final and non-final states; makes ρ-destination state final.
-  Weight Final(StateId s) const {
-    if (s == 0 || fst_->Final(s - 1) == Weight::Zero()) {
-      return Weight::One();
-    } else {
-      return Weight::Zero();
-    }
-  }
-
-  size_t NumArcs(StateId s) const {
-    return s == 0 ? 1 : fst_->NumArcs(s - 1) + 1;
-  }
-
-  size_t NumInputEpsilons(StateId s) const {
-    return s == 0 ? 0 : fst_->NumInputEpsilons(s - 1);
-  }
-
-  size_t NumOutputEpsilons(StateId s) const {
-    return s == 0 ? 0 : fst_->NumOutputEpsilons(s - 1);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && fst_->Properties(kError, false)) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
- private:
-  std::unique_ptr<const Fst<Arc>> fst_;
-};
-
-}  // namespace internal
-
-// Complements an automaton. This is a library-internal operation that
-// introduces a (negative) ρ-label; use Difference/DifferenceFst in user code,
-// which will not see this label. This version is a delayed FST.
-//
-// This class attaches interface to implementation and handles
-// reference counting, delegating most methods to ImplToFst.
-template <class A>
-class ComplementFst : public ImplToFst<internal::ComplementFstImpl<A>> {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Impl = internal::ComplementFstImpl<Arc>;
-
-  friend class StateIterator<ComplementFst<Arc>>;
-  friend class ArcIterator<ComplementFst<Arc>>;
-
-  explicit ComplementFst(const Fst<Arc> &fst)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst)) {
-    static constexpr auto props =
-        kUnweighted | kNoEpsilons | kIDeterministic | kAcceptor;
-    if (fst.Properties(props, true) != props) {
-      FSTERROR() << "ComplementFst: Argument not an unweighted "
-                 << "epsilon-free deterministic acceptor";
-      GetImpl()->SetProperties(kError, kError);
-    }
-  }
-
-  // See Fst<>::Copy() for doc.
-  ComplementFst(const ComplementFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Gets a copy of this FST. See Fst<>::Copy() for further doc.
-  ComplementFst<Arc> *Copy(bool safe = false) const override {
-    return new ComplementFst<Arc>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  inline void InitArcIterator(StateId s,
-                              ArcIteratorData<Arc> *data) const override;
-
-  // Label that represents the ρ-transition; we use a negative value private to
-  // the library and which will preserve FST label sort order.
-  static const Label kRhoLabel = -2;
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-
-  ComplementFst &operator=(const ComplementFst &) = delete;
-};
-
-template <class Arc>
-const typename Arc::Label ComplementFst<Arc>::kRhoLabel;
-
-// Specialization for ComplementFst.
-template <class Arc>
-class StateIterator<ComplementFst<Arc>> : public StateIteratorBase<Arc> {
- public:
-  using StateId = typename Arc::StateId;
-
-  explicit StateIterator(const ComplementFst<Arc> &fst)
-      : siter_(*fst.GetImpl()->fst_), s_(0) {}
-
-  bool Done() const final { return s_ > 0 && siter_.Done(); }
-
-  StateId Value() const final { return s_; }
-
-  void Next() final {
-    if (s_ != 0) siter_.Next();
-    ++s_;
-  }
-
-  void Reset() final {
-    siter_.Reset();
-    s_ = 0;
-  }
-
- private:
-  StateIterator<Fst<Arc>> siter_;
-  StateId s_;
-};
-
-// Specialization for ComplementFst.
-template <class Arc>
-class ArcIterator<ComplementFst<Arc>> : public ArcIteratorBase<Arc> {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  ArcIterator(const ComplementFst<Arc> &fst, StateId s) : s_(s), pos_(0) {
-    if (s_ != 0) {
-      aiter_.reset(new ArcIterator<Fst<Arc>>(*fst.GetImpl()->fst_, s - 1));
-    }
-  }
-
-  bool Done() const final {
-    if (s_ != 0) {
-      return pos_ > 0 && aiter_->Done();
-    } else {
-      return pos_ > 0;
-    }
-  }
-
-  // Adds the ρ-label to the ρ destination state.
-  const Arc &Value() const final {
-    if (pos_ == 0) {
-      arc_.ilabel = arc_.olabel = ComplementFst<Arc>::kRhoLabel;
-      arc_.weight = Weight::One();
-      arc_.nextstate = 0;
-    } else {
-      arc_ = aiter_->Value();
-      ++arc_.nextstate;
-    }
-    return arc_;
-  }
-
-  void Next() final {
-    if (s_ != 0 && pos_ > 0) aiter_->Next();
-    ++pos_;
-  }
-
-  size_t Position() const final { return pos_; }
-
-  void Reset() final {
-    if (s_ != 0) aiter_->Reset();
-    pos_ = 0;
-  }
-
-  void Seek(size_t a) final {
-    if (s_ != 0) {
-      if (a == 0) {
-        aiter_->Reset();
-      } else {
-        aiter_->Seek(a - 1);
-      }
-    }
-    pos_ = a;
-  }
-
-  uint32 Flags() const final { return kArcValueFlags; }
-
-  void SetFlags(uint32, uint32) final {}
-
- private:
-  std::unique_ptr<ArcIterator<Fst<Arc>>> aiter_;
-  StateId s_;
-  size_t pos_;
-  mutable Arc arc_;
-};
-
-template <class Arc>
-inline void ComplementFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<ComplementFst<Arc>>(*this);
-}
-
-template <class Arc>
-inline void ComplementFst<Arc>::InitArcIterator(StateId s,
-    ArcIteratorData<Arc> *data) const {
-  data->base = new ArcIterator<ComplementFst<Arc>>(*this, s);
-}
-
-// Useful alias when using StdArc.
-using StdComplementFst = ComplementFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_COMPLEMENT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compose-filter.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compose-filter.h
deleted file mode 100644
index 72b79bce..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compose-filter.h
+++ /dev/null
@@ -1,508 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes for filtering the composition matches, e.g. for correct epsilon
-// handling.
-
-#ifndef FST_COMPOSE_FILTER_H_
-#define FST_COMPOSE_FILTER_H_
-
-#include <fst/filter-state.h>
-#include <fst/fst-decl.h>  // For optional argument declarations
-#include <fst/fst.h>
-#include <fst/matcher.h>
-
-
-namespace fst {
-
-// Composition filters determine which matches are allowed to proceed. The
-// filter's state is represeted by the type ComposeFilter::FilterState.
-// The basic filters handle correct epsilon matching. Their interface is:
-//
-// template <class M1, class M2>
-// class ComposeFilter {
-//  public:
-//   using Matcher1 = ...;
-//   using Matcher2 = ...;
-//   using FST1 = typename M1::FST;
-//   using FST2 = typename M2::FST;
-//   using FilterState = ...;
-//
-//   using Arc = typename FST1::Arc;
-//   using StateId = typename Arc::StateId;
-//   using Weight = typename Arc::Weight;
-//
-//   // Required constructor.
-//   ComposeFilter(const FST1 &fst1, const FST2 &fst2,
-//                 M1 *matcher1 = nullptr, M2 *matcher2 = nullptr);
-//
-//   // If safe=true, the copy is thread-safe. See Fst<>::Copy()
-//   // for further doc.
-//   ComposeFilter(const ComposeFilter<M1, M2> &filter,
-//                 bool safe = false);
-//
-//   // Return start state of filter.
-//   FilterState Start() const;
-//
-//   // Specifies current composition state.
-//   void SetState(StateId s1, StateId s2, const FilterState &fs);
-//
-//   // Apply filter at current composition state to these transitions. If an
-//   // arc label to be matched is kNolabel, then that side does not consume a
-//   // symbol. Returns the new filter state or, if disallowed,
-//   // FilterState::NoState(). The filter is permitted to modify its inputs
-//   // (e.g. for optimization reasons).
-//   FilterState FilterArc(Arc *arc1, Arc *arc2) const;
-
-//   // Apply filter at current composition state to these final weights
-//   // (cf. superfinal transitions). The filter may modify its inputs
-//   // (e.g. for optimization reasons).
-//   void FilterFinal(Weight *w1, Weight *w2) const;
-//
-//   // Return the respective matchers. Ownership stays with filter. These
-//   // methods allow the filter to access and possibly modify the compositio
-//   // matchers (useful, e.g., with lookahead).
-//
-//   Matcher1 *GetMatcher1();
-//
-//   Matcher2 *GetMatcher2();
-//
-//   // This specifies how the filter affects the composition result properties.
-//   It takes as argument the properties that would apply with a trivial
-//   // composition filter.
-//   uint64 Properties(uint64 props) const;
-// };
-//
-// This filter allows only exact matching of symbols from FST1 with on FST2;
-// e.g., no special interpretation of epsilons.
-template <class M1, class M2 /* = M1 */>
-class NullComposeFilter {
- public:
-  using Matcher1 = M1;
-  using Matcher2 = M2;
-  using FST1 = typename M1::FST;
-  using FST2 = typename M2::FST;
-  using FilterState = TrivialFilterState;
-
-  using Arc = typename FST1::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  NullComposeFilter(const FST1 &fst1, const FST2 &fst2,
-                    Matcher1 *matcher1 = nullptr, Matcher2 *matcher2 = nullptr)
-      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
-        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()) {}
-
-  NullComposeFilter(const NullComposeFilter<M1, M2> &filter, bool safe = false)
-      : matcher1_(filter.matcher1_->Copy(safe)),
-        matcher2_(filter.matcher2_->Copy(safe)),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()) {}
-
-  FilterState Start() const { return FilterState(true); }
-
-  void SetState(StateId, StateId, const FilterState &) {}
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    return (arc1->olabel == kNoLabel || arc2->ilabel == kNoLabel)
-               ? FilterState::NoState()
-               : FilterState(true);
-  }
-
-  void FilterFinal(Weight *, Weight *) const {}
-
-  Matcher1 *GetMatcher1() { return matcher1_.get(); }
-
-  Matcher2 *GetMatcher2() { return matcher2_.get(); }
-
-  uint64 Properties(uint64 props) const { return props; }
-
- private:
-  std::unique_ptr<Matcher1> matcher1_;
-  std::unique_ptr<Matcher2> matcher2_;
-  const FST1 &fst1_;
-  const FST2 &fst2_;
-};
-
-// This filter allows all epsilon matches, potentially resulting in redundant
-// epsilon paths. The use of this filter gives correct results iff one of the
-// following conditions hold:
-//
-//  (1) The semiring is idempotent,
-//  (2) the first FST is output-epsilon free, or
-//  (3) the second FST is input-epsilon free.
-//
-// For (1), redundant epsilon paths may be created but won't hurt correctness.
-// For (2) and (3), no redundant paths are created.
-template <class M1, class M2 /* = M1 */>
-class TrivialComposeFilter {
- public:
-  using Matcher1 = M1;
-  using Matcher2 = M2;
-  using FST1 = typename M1::FST;
-  using FST2 = typename M2::FST;
-  using FilterState = TrivialFilterState;
-
-  using Arc = typename FST1::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  TrivialComposeFilter(const FST1 &fst1, const FST2 &fst2,
-                       Matcher1 *matcher1 = nullptr,
-                       Matcher2 *matcher2 = nullptr)
-      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
-        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()) {}
-
-  TrivialComposeFilter(const TrivialComposeFilter<Matcher1, Matcher2> &filter,
-                       bool safe = false)
-      : matcher1_(filter.matcher1_->Copy(safe)),
-        matcher2_(filter.matcher2_->Copy(safe)),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()) {}
-
-  FilterState Start() const { return FilterState(true); }
-
-  void SetState(StateId, StateId, const FilterState &) {}
-
-  FilterState FilterArc(Arc *, Arc *) const { return FilterState(true); }
-
-  void FilterFinal(Weight *, Weight *) const {}
-
-  Matcher1 *GetMatcher1() { return matcher1_.get(); }
-
-  Matcher2 *GetMatcher2() { return matcher2_.get(); }
-
-  uint64 Properties(uint64 props) const { return props; }
-
- private:
-  std::unique_ptr<Matcher1> matcher1_;
-  std::unique_ptr<Matcher2> matcher2_;
-  const FST1 &fst1_;
-  const FST2 &fst2_;
-};
-
-// This filter requires epsilons on FST1 to be read before epsilons on FST2.
-template <class M1, class M2 /* = M1 */>
-class SequenceComposeFilter {
- public:
-  using Matcher1 = M1;
-  using Matcher2 = M2;
-  using FST1 = typename M1::FST;
-  using FST2 = typename M2::FST;
-  using FilterState = CharFilterState;
-
-  using Arc = typename FST1::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  SequenceComposeFilter(const FST1 &fst1, const FST2 &fst2,
-                        Matcher1 *matcher1 = nullptr,
-                        Matcher2 *matcher2 = nullptr)
-      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
-        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
-        fst1_(matcher1_->GetFst()),
-        s1_(kNoStateId),
-        s2_(kNoStateId),
-        fs_(kNoStateId) {}
-
-  SequenceComposeFilter(const SequenceComposeFilter<Matcher1, Matcher2> &filter,
-                        bool safe = false)
-      : matcher1_(filter.matcher1_->Copy(safe)),
-        matcher2_(filter.matcher2_->Copy(safe)),
-        fst1_(matcher1_->GetFst()),
-        s1_(kNoStateId),
-        s2_(kNoStateId),
-        fs_(kNoStateId) {}
-
-  FilterState Start() const { return FilterState(0); }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    if (s1_ == s1 && s2_ == s2 && fs == fs_) return;
-    s1_ = s1;
-    s2_ = s2;
-    fs_ = fs;
-    const auto na1 = internal::NumArcs(fst1_, s1);
-    const auto ne1 = internal::NumOutputEpsilons(fst1_, s1);
-    const bool fin1 = internal::Final(fst1_, s1) != Weight::Zero();
-    alleps1_ = na1 == ne1 && !fin1;
-    noeps1_ = ne1 == 0;
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    if (arc1->olabel == kNoLabel) {
-      return alleps1_ ? FilterState::NoState() : noeps1_ ? FilterState(0)
-                                                         : FilterState(1);
-    } else if (arc2->ilabel == kNoLabel) {
-      return fs_ != FilterState(0) ? FilterState::NoState() : FilterState(0);
-    } else {
-      return arc1->olabel == 0 ? FilterState::NoState() : FilterState(0);
-    }
-  }
-
-  void FilterFinal(Weight *, Weight *) const {}
-
-  Matcher1 *GetMatcher1() { return matcher1_.get(); }
-
-  Matcher2 *GetMatcher2() { return matcher2_.get(); }
-
-  uint64 Properties(uint64 props) const { return props; }
-
- private:
-  std::unique_ptr<Matcher1> matcher1_;
-  std::unique_ptr<Matcher2> matcher2_;
-  const FST1 &fst1_;
-  StateId s1_;      // Current fst1_ state.
-  StateId s2_;      // Current fst2_ state.
-  FilterState fs_;  // Current filter state.
-  bool alleps1_;   // Only epsilons (and non-final) leaving s1_?
-  bool noeps1_;    // No epsilons leaving s1_?
-};
-
-// This filter requires epsilons on FST2 to be read before epsilons on FST1.
-template <class M1, class M2 /* = M1 */>
-class AltSequenceComposeFilter {
- public:
-  using Matcher1 = M1;
-  using Matcher2 = M2;
-  using FST1 = typename M1::FST;
-  using FST2 = typename M2::FST;
-  using FilterState = CharFilterState;
-
-  using Arc = typename FST1::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  AltSequenceComposeFilter(const FST1 &fst1, const FST2 &fst2,
-                           Matcher1 *matcher1 = nullptr,
-                           Matcher2 *matcher2 = nullptr)
-      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
-        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
-        fst2_(matcher2_->GetFst()),
-        s1_(kNoStateId),
-        s2_(kNoStateId),
-        fs_(kNoStateId) {}
-
-  AltSequenceComposeFilter(
-      const AltSequenceComposeFilter<Matcher1, Matcher2> &filter,
-      bool safe = false)
-      : matcher1_(filter.matcher1_->Copy(safe)),
-        matcher2_(filter.matcher2_->Copy(safe)),
-        fst2_(matcher2_->GetFst()),
-        s1_(kNoStateId),
-        s2_(kNoStateId),
-        fs_(kNoStateId) {}
-
-  FilterState Start() const { return FilterState(0); }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    if (s1_ == s1 && s2_ == s2 && fs == fs_) return;
-    s1_ = s1;
-    s2_ = s2;
-    fs_ = fs;
-    const auto na2 = internal::NumArcs(fst2_, s2);
-    const auto ne2 = internal::NumInputEpsilons(fst2_, s2);
-    const bool fin2 = internal::Final(fst2_, s2) != Weight::Zero();
-    alleps2_ = na2 == ne2 && !fin2;
-    noeps2_ = ne2 == 0;
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    if (arc2->ilabel == kNoLabel) {
-      return alleps2_ ? FilterState::NoState() : noeps2_ ? FilterState(0)
-                                                         : FilterState(1);
-    } else if (arc1->olabel == kNoLabel) {
-      return fs_ == FilterState(1) ? FilterState::NoState() : FilterState(0);
-    } else {
-      return arc1->olabel == 0 ? FilterState::NoState() : FilterState(0);
-    }
-  }
-
-  void FilterFinal(Weight *, Weight *) const {}
-
-  Matcher1 *GetMatcher1() { return matcher1_.get(); }
-
-  Matcher2 *GetMatcher2() { return matcher2_.get(); }
-
-  uint64 Properties(uint64 props) const { return props; }
-
- private:
-  std::unique_ptr<Matcher1> matcher1_;
-  std::unique_ptr<Matcher2> matcher2_;
-  const FST2 &fst2_;
-  StateId s1_;      // Current fst1_ state.
-  StateId s2_;      // Current fst2_ state.
-  FilterState fs_;  // Current filter state.
-  bool alleps2_;    // Only epsilons (and non-final) leaving s2_?
-  bool noeps2_;     // No epsilons leaving s2_?
-};
-
-// This filter requires epsilons on FST1 to be matched with epsilons on FST2
-// whenever possible. (Template arg default declared in fst-decl.h.)
-template <class M1, class M2 /* = M1 */>
-class MatchComposeFilter {
- public:
-  using Matcher1 = M1;
-  using Matcher2 = M2;
-  using FST1 = typename M1::FST;
-  using FST2 = typename M2::FST;
-  using FilterState = CharFilterState;
-
-  using Arc = typename FST1::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  MatchComposeFilter(const FST1 &fst1, const FST2 &fst2,
-                     Matcher1 *matcher1 = nullptr, Matcher2 *matcher2 = nullptr)
-      : matcher1_(matcher1 ? matcher1 : new Matcher1(fst1, MATCH_OUTPUT)),
-        matcher2_(matcher2 ? matcher2 : new Matcher2(fst2, MATCH_INPUT)),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()),
-        s1_(kNoStateId),
-        s2_(kNoStateId),
-        fs_(kNoStateId) {}
-
-  MatchComposeFilter(const MatchComposeFilter<Matcher1, Matcher2> &filter,
-                     bool safe = false)
-      : matcher1_(filter.matcher1_->Copy(safe)),
-        matcher2_(filter.matcher2_->Copy(safe)),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()),
-        s1_(kNoStateId),
-        s2_(kNoStateId),
-        fs_(kNoStateId) {}
-
-  FilterState Start() const { return FilterState(0); }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    if (s1_ == s1 && s2_ == s2 && fs == fs_) return;
-    s1_ = s1;
-    s2_ = s2;
-    fs_ = fs;
-    size_t na1 = internal::NumArcs(fst1_, s1);
-    size_t ne1 = internal::NumOutputEpsilons(fst1_, s1);
-    bool f1 = internal::Final(fst1_, s1) != Weight::Zero();
-    alleps1_ = na1 == ne1 && !f1;
-    noeps1_ = ne1 == 0;
-    size_t na2 = internal::NumArcs(fst2_, s2);
-    size_t ne2 = internal::NumInputEpsilons(fst2_, s2);
-    bool f2 = internal::Final(fst2_, s2) != Weight::Zero();
-    alleps2_ = na2 == ne2 && !f2;
-    noeps2_ = ne2 == 0;
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    if (arc2->ilabel == kNoLabel) {  // Epsilon in FST1.
-      return fs_ == FilterState(0)
-                 ? (noeps2_
-                        ? FilterState(0)
-                        : (alleps2_ ? FilterState::NoState() : FilterState(1)))
-                 : (fs_ == FilterState(1) ? FilterState(1)
-                                          : FilterState::NoState());
-    } else if (arc1->olabel == kNoLabel) {  // Epsilon in FST2.
-      return fs_ == FilterState(0)
-                 ? (noeps1_
-                        ? FilterState(0)
-                        : (alleps1_ ? FilterState::NoState() : FilterState(2)))
-                 : (fs_ == FilterState(2) ? FilterState(2)
-                                          : FilterState::NoState());
-    } else if (arc1->olabel == 0) {  // Epsilon in both.
-      return fs_ == FilterState(0) ? FilterState(0) : FilterState::NoState();
-    } else {  // Both are non-epsilons.
-      return FilterState(0);
-    }
-  }
-
-  void FilterFinal(Weight *, Weight *) const {}
-
-  Matcher1 *GetMatcher1() { return matcher1_.get(); }
-
-  Matcher2 *GetMatcher2() { return matcher2_.get(); }
-
-  uint64 Properties(uint64 props) const { return props; }
-
- private:
-  std::unique_ptr<Matcher1> matcher1_;
-  std::unique_ptr<Matcher2> matcher2_;
-  const FST1 &fst1_;
-  const FST2 &fst2_;
-  StateId s1_;      // Current fst1_ state.
-  StateId s2_;      // Current fst2_ state.
-  FilterState fs_;  // Current filter state ID.
-  bool alleps1_;    // Only epsilson (and non-final) leaving s1?
-  bool alleps2_;    // Only epsilons (and non-final) leaving s2?
-  bool noeps1_;     // No epsilons leaving s1?
-  bool noeps2_;     // No epsilons leaving s2?
-};
-
-// This filter works with the MultiEpsMatcher to determine if multi-epsilons are
-// preserved in the composition output (rather than rewritten as 0) and
-// ensures correct properties.
-template <class Filter>
-class MultiEpsFilter {
- public:
-  using Matcher1 = typename Filter::Matcher1;
-  using Matcher2 = typename Filter::Matcher2;
-  using FST1 = typename Filter::FST1;
-  using FST2 = typename Filter::FST2;
-  using FilterState = typename Filter::FilterState;
-
-  using Arc = typename Filter::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  MultiEpsFilter(const FST1 &fst1, const FST2 &fst2,
-                 Matcher1 *matcher1 = nullptr, Matcher2 *matcher2 = nullptr,
-                 bool keep_multi_eps = false)
-      : filter_(fst1, fst2, matcher1, matcher2),
-        keep_multi_eps_(keep_multi_eps) {}
-
-  MultiEpsFilter(const MultiEpsFilter<Filter> &filter, bool safe = false)
-      : filter_(filter.filter_, safe),
-        keep_multi_eps_(filter.keep_multi_eps_) {}
-
-  FilterState Start() const { return filter_.Start(); }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    return filter_.SetState(s1, s2, fs);
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    const auto fs = filter_.FilterArc(arc1, arc2);
-    if (keep_multi_eps_) {
-      if (arc1->olabel == kNoLabel) arc1->ilabel = arc2->ilabel;
-      if (arc2->ilabel == kNoLabel) arc2->olabel = arc1->olabel;
-    }
-    return fs;
-  }
-
-  void FilterFinal(Weight *w1, Weight *w2) const {
-    return filter_.FilterFinal(w1, w2);
-  }
-
-  Matcher1 *GetMatcher1() { return filter_.GetMatcher1(); }
-
-  Matcher2 *GetMatcher2() { return filter_.GetMatcher2(); }
-
-  uint64 Properties(uint64 iprops) const {
-    const auto oprops = filter_.Properties(iprops);
-    return oprops & kILabelInvariantProperties & kOLabelInvariantProperties;
-  }
-
- private:
-  Filter filter_;
-  bool keep_multi_eps_;
-};
-
-}  // namespace fst
-
-#endif  // FST_COMPOSE_FILTER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compose.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compose.h
deleted file mode 100644
index a4863fd6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/compose.h
+++ /dev/null
@@ -1,1013 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to compute the composition of two FSTs.
-
-#ifndef FST_COMPOSE_H_
-#define FST_COMPOSE_H_
-
-#include <algorithm>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/compose-filter.h>
-#include <fst/fst-decl.h>  // For optional argument declarations
-#include <fst/lookahead-filter.h>
-#include <fst/matcher.h>
-#include <fst/state-table.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-// Delayed composition options templated on the arc type, the matcher,
-// the composition filter, and the composition state table.  By
-// default, the matchers, filter, and state table are constructed by
-// composition. If set below, the user can instead pass in these
-// objects; in that case, ComposeFst takes their ownership. This
-// version controls composition implemented between generic Fst<Arc>
-// types and a shared matcher type M for Fst<Arc>. This should be
-// adequate for most applications, giving a reasonable tradeoff
-// between efficiency and code sharing (but see ComposeFstImplOptions).
-template <class Arc, class M = Matcher<Fst<Arc>>,
-          class Filter = SequenceComposeFilter<M>,
-          class StateTable =
-              GenericComposeStateTable<Arc, typename Filter::FilterState>>
-struct ComposeFstOptions : public CacheOptions {
-  M *matcher1;              // FST1 matcher.
-  M *matcher2;              // FST2 matcher.
-  Filter *filter;           // Composition filter.
-  StateTable *state_table;  // Composition state table.
-
-  explicit ComposeFstOptions(const CacheOptions &opts = CacheOptions(),
-                             M *matcher1 = nullptr, M *matcher2 = nullptr,
-                             Filter *filter = nullptr,
-                             StateTable *state_table = nullptr)
-      : CacheOptions(opts),
-        matcher1(matcher1),
-        matcher2(matcher2),
-        filter(filter),
-        state_table(state_table) {}
-};
-
-// Forward declaration of ComposeFstMatcher.
-template <class C, class F, class T>
-class ComposeFstMatcher;
-
-// Delayed composition options templated on the two matcher types, the
-// composition filter, the composition state table and the cache store. By
-// default, the matchers, filter, state table and cache store are constructed
-// by composition. If set below, the user can instead pass in these objects; in
-// that case, ComposeFst takes their ownership. This version controls
-// composition implemented using arbitrary matchers (of the same arc type but
-// otherwise arbitrary FST type). The user must ensure the matchers are
-// compatible. These options permit the most efficient use, but shares the
-// least code. This is for advanced use only in the most demanding or
-// specialized applications that can benefit from it; otherwise, prefer
-// ComposeFstOptions).
-template <class M1, class M2, class Filter = SequenceComposeFilter<M1, M2>,
-          class StateTable = GenericComposeStateTable<
-              typename M1::Arc, typename Filter::FilterState>,
-          class CacheStore = DefaultCacheStore<typename M1::Arc>>
-struct ComposeFstImplOptions : public CacheImplOptions<CacheStore> {
-  M1 *matcher1;    // FST1 matcher (see matcher.h)....
-  M2 *matcher2;    // FST2 matcher.
-  Filter *filter;  // Composition filter (see compose-filter.h).
-  StateTable
-    *state_table;        // Composition state table (see compose-state-table.h).
-  bool own_state_table;   // ComposeFstImpl takes ownership of 'state_table'?
-  bool allow_noncommute;  // Allow non-commutative weights
-
-  explicit ComposeFstImplOptions(const CacheOptions &opts,
-                                 M1 *matcher1 = nullptr, M2 *matcher2 = nullptr,
-                                 Filter *filter = nullptr,
-                                 StateTable *state_table = nullptr)
-      : CacheImplOptions<CacheStore>(opts),
-        matcher1(matcher1),
-        matcher2(matcher2),
-        filter(filter),
-        state_table(state_table),
-        own_state_table(true),
-        allow_noncommute(false) {}
-
-  explicit ComposeFstImplOptions(const CacheImplOptions<CacheStore> &opts,
-                                 M1 *matcher1 = nullptr, M2 *matcher2 = nullptr,
-                                 Filter *filter = nullptr,
-                                 StateTable *state_table = nullptr)
-      : CacheImplOptions<CacheStore>(opts),
-        matcher1(matcher1),
-        matcher2(matcher2),
-        filter(filter),
-        state_table(state_table),
-        own_state_table(true),
-        allow_noncommute(false) {}
-
-  ComposeFstImplOptions()
-      : matcher1(nullptr),
-        matcher2(nullptr),
-        filter(nullptr),
-        state_table(nullptr),
-        own_state_table(true),
-        allow_noncommute(false) {}
-};
-
-namespace internal {
-
-// Implementation of delayed composition. This base class is common to the
-// variants with different matchers, composition filters and state tables.
-template <class Arc, class CacheStore = DefaultCacheStore<Arc>,
-          class F = ComposeFst<Arc, CacheStore>>
-class ComposeFstImplBase
-    : public CacheBaseImpl<typename CacheStore::State, CacheStore> {
- public:
-  using FST = F;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using State = typename CacheStore::State;
-  using CacheImpl = CacheBaseImpl<State, CacheStore>;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::Properties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheImpl::HasStart;
-  using CacheImpl::HasFinal;
-  using CacheImpl::HasArcs;
-  using CacheImpl::SetFinal;
-  using CacheImpl::SetStart;
-
-  ComposeFstImplBase(const CacheImplOptions<CacheStore> &opts)
-      : CacheImpl(opts) {}
-
-  ComposeFstImplBase(const CacheOptions &opts) : CacheImpl(opts) {}
-
-  ComposeFstImplBase(const ComposeFstImplBase &impl) : CacheImpl(impl, true) {
-    SetType(impl.Type());
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  virtual ComposeFstImplBase *Copy() const = 0;
-
-  ~ComposeFstImplBase() override {}
-
-  StateId Start() {
-    if (!HasStart()) {
-      const auto start = ComputeStart();
-      if (start != kNoStateId) SetStart(start);
-    }
-    return CacheImpl::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) SetFinal(s, ComputeFinal(s));
-    return CacheImpl::Final(s);
-  }
-
-  virtual void Expand(StateId s) = 0;
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl::NumOutputEpsilons(s);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl::InitArcIterator(s, data);
-  }
-
-  virtual MatcherBase<Arc> *InitMatcher(const F &fst,
-                                        MatchType match_type) const {
-    // Use the default matcher if no override is provided.
-    return nullptr;
-  }
-
- protected:
-  virtual StateId ComputeStart() = 0;
-  virtual Weight ComputeFinal(StateId s) = 0;
-};
-
-// Implementation of delayed composition templated on the matchers (see
-// matcher.h), composition filter (see compose-filter.h) and the composition
-// state table (see compose-state-table.h).
-template <class CacheStore, class Filter, class StateTable>
-class ComposeFstImpl
-    : public ComposeFstImplBase<typename CacheStore::Arc, CacheStore> {
- public:
-  using Matcher1 = typename Filter::Matcher1;
-  using Matcher2 = typename Filter::Matcher2;
-
-  using FST1 = typename Matcher1::FST;
-  using FST2 = typename Matcher2::FST;
-
-  using Arc = typename CacheStore::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FilterState = typename Filter::FilterState;
-  using State = typename CacheStore::State;
-
-  using CacheImpl = CacheBaseImpl<State, CacheStore>;
-
-  using StateTuple = typename StateTable::StateTuple;
-
-  friend class ComposeFstMatcher<CacheStore, Filter, StateTable>;
-
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-
-  template <class M1, class M2>
-  ComposeFstImpl(const FST1 &fst1, const FST2 &fst2,
-                 const ComposeFstImplOptions<M1, M2, Filter, StateTable,
-                                             CacheStore> &opts);
-
-  ComposeFstImpl(const ComposeFstImpl &impl)
-      : ComposeFstImplBase<Arc, CacheStore>(impl),
-        filter_(new Filter(*impl.filter_, true)),
-        matcher1_(filter_->GetMatcher1()),
-        matcher2_(filter_->GetMatcher2()),
-        fst1_(matcher1_->GetFst()),
-        fst2_(matcher2_->GetFst()),
-        state_table_(new StateTable(*impl.state_table_)),
-        own_state_table_(true),
-        match_type_(impl.match_type_) {}
-
-  ~ComposeFstImpl() override {
-    if (own_state_table_) delete state_table_;
-  }
-
-  ComposeFstImpl *Copy() const override { return new ComposeFstImpl(*this); }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) &&
-        (fst1_.Properties(kError, false) || fst2_.Properties(kError, false) ||
-         (matcher1_->Properties(0) & kError) ||
-         (matcher2_->Properties(0) & kError) |
-             (filter_->Properties(0) & kError) ||
-         state_table_->Error())) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  // Arranges it so that the first arg to OrderedExpand is the Fst
-  // that will be matched on.
-  void Expand(StateId s) override {
-    const auto &tuple = state_table_->Tuple(s);
-    const auto s1 = tuple.StateId1();
-    const auto s2 = tuple.StateId2();
-    filter_->SetState(s1, s2, tuple.GetFilterState());
-    if (MatchInput(s1, s2)) {
-      OrderedExpand(s, fst2_, s2, fst1_, s1, matcher2_, true);
-    } else {
-      OrderedExpand(s, fst1_, s1, fst2_, s2, matcher1_, false);
-    }
-  }
-
-  const FST1 &GetFst1() const { return fst1_; }
-
-  const FST2 &GetFst2() const { return fst2_; }
-
-  const Matcher1 *GetMatcher1() const { return matcher1_; }
-
-  Matcher1 *GetMatcher1() { return matcher1_; }
-
-  const Matcher2 *GetMatcher2() const { return matcher2_; }
-
-  Matcher2 *GetMatcher2() { return matcher2_; }
-
-  const Filter *GetFilter() const { return filter_.get(); }
-
-  Filter *GetFilter() { return filter_.get(); }
-
-  const StateTable *GetStateTable() const { return state_table_; }
-
-  StateTable *GetStateTable() { return state_table_; }
-
-  MatcherBase<Arc> *InitMatcher(const ComposeFst<Arc, CacheStore> &fst,
-                                MatchType match_type) const override {
-    const auto test_props = match_type == MATCH_INPUT
-                                ? kFstProperties & ~kILabelInvariantProperties
-                                : kFstProperties & ~kOLabelInvariantProperties;
-    // If both matchers support 'match_type' and we have a guarantee that a
-    // call to 'filter_->FilterArc(arc1, arc2)' will not modify the ilabel of
-    // arc1 when MATCH_INPUT or the olabel or arc2 when MATCH_OUTPUT, then
-    // ComposeFstMatcher can be used.
-    if ((matcher1_->Type(false) == match_type) &&
-        (matcher2_->Type(false) == match_type) &&
-        (filter_->Properties(test_props) == test_props)) {
-      return new ComposeFstMatcher<CacheStore, Filter, StateTable>(fst, this,
-                                                                   match_type);
-    }
-    return nullptr;
-  }
-
- private:
-  // This does that actual matching of labels in the composition. The
-  // arguments are ordered so matching is called on state 'sa' of
-  // 'fsta' for each arc leaving state 'sb' of 'fstb'. The 'match_input' arg
-  // determines whether the input or output label of arcs at 'sb' is
-  // the one to match on.
-  template <class FST, class Matcher>
-  void OrderedExpand(StateId s, const Fst<Arc> &, StateId sa, const FST &fstb,
-                     StateId sb, Matcher *matchera, bool match_input) {
-    matchera->SetState(sa);
-    // First processes non-consuming symbols (e.g., epsilons) on FSTA.
-    const Arc loop(match_input ? 0 : kNoLabel, match_input ? kNoLabel : 0,
-                   Weight::One(), sb);
-    MatchArc(s, matchera, loop, match_input);
-    // Then processes matches on FSTB.
-    for (ArcIterator<FST> iterb(fstb, sb); !iterb.Done(); iterb.Next()) {
-      MatchArc(s, matchera, iterb.Value(), match_input);
-    }
-    CacheImpl::SetArcs(s);
-  }
-
-  // Matches a single transition from 'fstb' against 'fata' at 's'.
-  template <class Matcher>
-  void MatchArc(StateId s, Matcher *matchera, const Arc &arc,
-                bool match_input) {
-    if (matchera->Find(match_input ? arc.olabel : arc.ilabel)) {
-      for (; !matchera->Done(); matchera->Next()) {
-        auto arca = matchera->Value();
-        auto arcb = arc;
-        if (match_input) {
-          const auto &fs = filter_->FilterArc(&arcb, &arca);
-          if (fs != FilterState::NoState()) AddArc(s, arcb, arca, fs);
-        } else {
-          const auto &fs = filter_->FilterArc(&arca, &arcb);
-          if (fs != FilterState::NoState()) AddArc(s, arca, arcb, fs);
-        }
-      }
-    }
-  }
-
-  // Add a matching transition at 's'.
-  void AddArc(StateId s, const Arc &arc1, const Arc &arc2,
-              const FilterState &f) {
-    const StateTuple tuple(arc1.nextstate, arc2.nextstate, f);
-    const Arc oarc(arc1.ilabel, arc2.olabel, Times(arc1.weight, arc2.weight),
-                   state_table_->FindState(tuple));
-    CacheImpl::PushArc(s, oarc);
-  }
-
-  StateId ComputeStart() override {
-    const auto s1 = fst1_.Start();
-    if (s1 == kNoStateId) return kNoStateId;
-    const auto s2 = fst2_.Start();
-    if (s2 == kNoStateId) return kNoStateId;
-    const auto &fs = filter_->Start();
-    const StateTuple tuple(s1, s2, fs);
-    return state_table_->FindState(tuple);
-  }
-
-  Weight ComputeFinal(StateId s) override {
-    const auto &tuple = state_table_->Tuple(s);
-    const auto s1 = tuple.StateId1();
-    auto final1 = matcher1_->Final(s1);
-    if (final1 == Weight::Zero()) return final1;
-    const auto s2 = tuple.StateId2();
-    auto final2 = matcher2_->Final(s2);
-    if (final2 == Weight::Zero()) return final2;
-    filter_->SetState(s1, s2, tuple.GetFilterState());
-    filter_->FilterFinal(&final1, &final2);
-    return Times(final1, final2);
-  }
-
-  // Determines which side to match on per composition state.
-  bool MatchInput(StateId s1, StateId s2) {
-    switch (match_type_) {
-      case MATCH_INPUT:
-        return true;
-      case MATCH_OUTPUT:
-        return false;
-      default:  // MATCH_BOTH
-        const auto priority1 = matcher1_->Priority(s1);
-        const auto priority2 = matcher2_->Priority(s2);
-        if (priority1 == kRequirePriority && priority2 == kRequirePriority) {
-          FSTERROR() << "ComposeFst: Both sides can't require match";
-          SetProperties(kError, kError);
-          return true;
-        }
-        if (priority1 == kRequirePriority) return false;
-        if (priority2 == kRequirePriority) {
-          return true;
-        }
-        return priority1 <= priority2;
-    }
-  }
-
-  // Identifies and verifies the capabilities of the matcher to be used for
-  // composition.
-  void SetMatchType();
-
-  std::unique_ptr<Filter> filter_;
-  Matcher1 *matcher1_;  // Borrowed reference.
-  Matcher2 *matcher2_;  // Borrowed reference.
-  const FST1 &fst1_;
-  const FST2 &fst2_;
-  StateTable *state_table_;
-  bool own_state_table_;
-
-  MatchType match_type_;
-};
-
-template <class CacheStore, class Filter, class StateTable>
-template <class M1, class M2>
-ComposeFstImpl<CacheStore, Filter, StateTable>::ComposeFstImpl(
-    const FST1 &fst1, const FST2 &fst2,
-    const ComposeFstImplOptions<M1, M2, Filter, StateTable, CacheStore> &opts)
-    : ComposeFstImplBase<Arc, CacheStore>(opts),
-      filter_(opts.filter
-                  ? opts.filter
-                  : new Filter(fst1, fst2, opts.matcher1, opts.matcher2)),
-      matcher1_(filter_->GetMatcher1()),
-      matcher2_(filter_->GetMatcher2()),
-      fst1_(matcher1_->GetFst()),
-      fst2_(matcher2_->GetFst()),
-      state_table_(opts.state_table ? opts.state_table
-                                    : new StateTable(fst1_, fst2_)),
-      own_state_table_(opts.state_table ? opts.own_state_table : true) {
-  SetType("compose");
-  if (!CompatSymbols(fst2.InputSymbols(), fst1.OutputSymbols())) {
-    FSTERROR() << "ComposeFst: Output symbol table of 1st argument "
-               << "does not match input symbol table of 2nd argument";
-    SetProperties(kError, kError);
-  }
-  SetInputSymbols(fst1_.InputSymbols());
-  SetOutputSymbols(fst2_.OutputSymbols());
-  SetMatchType();
-  VLOG(2) << "ComposeFstImpl: Match type: " << match_type_;
-  if (match_type_ == MATCH_NONE) SetProperties(kError, kError);
-  const auto fprops1 = fst1.Properties(kFstProperties, false);
-  const auto fprops2 = fst2.Properties(kFstProperties, false);
-  const auto mprops1 = matcher1_->Properties(fprops1);
-  const auto mprops2 = matcher2_->Properties(fprops2);
-  const auto cprops = ComposeProperties(mprops1, mprops2);
-  SetProperties(filter_->Properties(cprops), kCopyProperties);
-  if (state_table_->Error()) SetProperties(kError, kError);
-}
-
-template <class CacheStore, class Filter, class StateTable>
-void ComposeFstImpl<CacheStore, Filter, StateTable>::SetMatchType() {
-  // Ensures any required matching is possible and known.
-  if ((matcher1_->Flags() & kRequireMatch) &&
-      matcher1_->Type(true) != MATCH_OUTPUT) {
-    FSTERROR() << "ComposeFst: 1st argument cannot perform required matching "
-               << "(sort?).";
-    match_type_ = MATCH_NONE;
-    return;
-  }
-  if ((matcher2_->Flags() & kRequireMatch) &&
-      matcher2_->Type(true) != MATCH_INPUT) {
-    FSTERROR() << "ComposeFst: 2nd argument cannot perform required matching "
-               << "(sort?).";
-    match_type_ = MATCH_NONE;
-    return;
-  }
-  // Finds which sides to match on (favoring minimal testing of capabilities).
-  const auto type1 = matcher1_->Type(false);
-  const auto type2 = matcher2_->Type(false);
-  if (type1 == MATCH_OUTPUT && type2 == MATCH_INPUT) {
-    match_type_ = MATCH_BOTH;
-  } else if (type1 == MATCH_OUTPUT) {
-    match_type_ = MATCH_OUTPUT;
-  } else if (type2 == MATCH_INPUT) {
-    match_type_ = MATCH_INPUT;
-  } else if (matcher1_->Type(true) == MATCH_OUTPUT) {
-    match_type_ = MATCH_OUTPUT;
-  } else if (matcher2_->Type(true) == MATCH_INPUT) {
-    match_type_ = MATCH_INPUT;
-  } else {
-    FSTERROR() << "ComposeFst: 1st argument cannot match on output labels "
-               << "and 2nd argument cannot match on input labels (sort?).";
-    match_type_ = MATCH_NONE;
-  }
-}
-
-}  // namespace internal
-
-// Computes the composition of two transducers. This version is a delayed FST.
-// If FST1 transduces string x to y with weight a and FST2 transduces y to z
-// with weight b, then their composition transduces string x to z with weight
-// Times(x, z).
-//
-// The output labels of the first transducer or the input labels of the second
-// transducer must be sorted (with the default matcher). The weights need to
-// form a commutative semiring (valid for TropicalWeight and LogWeight).
-//
-// Complexity:
-//
-// Assuming the first FST is unsorted and the second is sorted,
-//
-//   Time: O(v1 v2 d1 (log d2 + m2)),
-//   Space: O(v1 v2)
-//
-// where vi = # of states visited, di = maximum out-degree, and mi the
-// maximum multiplicity of the states visited, for the ith FST. Constant time
-// and space to visit an input state or arc is assumed and exclusive of caching.
-//
-// Caveats:
-// - ComposeFst does not trim its output (since it is a delayed operation).
-// - The efficiency of composition can be strongly affected by several factors:
-//   - the choice of which transducer is sorted - prefer sorting the FST
-//     that has the greater average out-degree.
-//   - the amount of non-determinism
-//   - the presence and location of epsilon transitions - avoid epsilon
-//     transitions on the output side of the first transducer or
-//     the input side of the second transducer or prefer placing
-//     them later in a path since they delay matching and can
-//     introduce non-coaccessible states and transitions.
-//
-// This class attaches interface to implementation and handles reference
-// counting, delegating most methods to ImplToFst. The CacheStore specifies the
-// cache store (default declared in fst-decl.h).
-template <class A, class CacheStore /* = DefaultCacheStore<A> */>
-class ComposeFst
-    : public ImplToFst<internal::ComposeFstImplBase<A, CacheStore>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = CacheStore;
-  using State = typename CacheStore::State;
-
-  using Impl = internal::ComposeFstImplBase<A, CacheStore>;
-
-  friend class ArcIterator<ComposeFst<Arc, CacheStore>>;
-  friend class StateIterator<ComposeFst<Arc, CacheStore>>;
-
-  // Compose specifying only caching options.
-  ComposeFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-             const CacheOptions &opts = CacheOptions())
-      : ImplToFst<Impl>(CreateBase(fst1, fst2, opts)) {}
-
-  // Compose specifying one shared matcher type M. Requires that the input FSTs
-  // and matcher FST types be Fst<Arc>. Recommended for best code-sharing and
-  // matcher compatiblity.
-  template <class Matcher, class Filter, class StateTuple>
-  ComposeFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-             const ComposeFstOptions<Arc, Matcher, Filter, StateTuple> &opts)
-      : ImplToFst<Impl>(CreateBase1(fst1, fst2, opts)) {}
-
-  // Compose specifying two matcher types Matcher1 and Matcher2. Requires input
-  // FST (of the same Arc type, but o.w. arbitrary) match the corresponding
-  // matcher FST types). Recommended only for advanced use in demanding or
-  // specialized applications due to potential code bloat and matcher
-  // incompatibilities.
-  template <class Matcher1, class Matcher2, class Filter, class StateTuple>
-  ComposeFst(const typename Matcher1::FST &fst1,
-             const typename Matcher2::FST &fst2,
-             const ComposeFstImplOptions<Matcher1, Matcher2, Filter, StateTuple,
-                                         CacheStore> &opts)
-      : ImplToFst<Impl>(CreateBase2(fst1, fst2, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  ComposeFst(const ComposeFst<A, CacheStore> &fst, bool safe = false)
-      : ImplToFst<Impl>(safe ? std::shared_ptr<Impl>(fst.GetImpl()->Copy())
-                             : fst.GetSharedImpl()) {}
-
-  // Get a copy of this ComposeFst. See Fst<>::Copy() for further doc.
-  ComposeFst<A, CacheStore> *Copy(bool safe = false) const override {
-    return new ComposeFst<A, CacheStore>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
-  MatcherBase<Arc> *InitMatcher(MatchType match_type) const override {
-    return GetImpl()->InitMatcher(*this, match_type);
-  }
-
- protected:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  explicit ComposeFst(std::shared_ptr<Impl> impl) : ImplToFst<Impl>(impl) {}
-
-  // Create compose implementation specifying two matcher types.
-  template <class Matcher1, class Matcher2, class Filter, class StateTuple>
-  static std::shared_ptr<Impl> CreateBase2(
-      const typename Matcher1::FST &fst1, const typename Matcher2::FST &fst2,
-      const ComposeFstImplOptions<Matcher1, Matcher2, Filter, StateTuple,
-                                  CacheStore> &opts) {
-    auto impl = std::make_shared<
-        internal::ComposeFstImpl<CacheStore, Filter, StateTuple>>(fst1, fst2,
-                                                                  opts);
-    if (!(Weight::Properties() & kCommutative) && !opts.allow_noncommute) {
-      const auto props1 = fst1.Properties(kUnweighted, true);
-      const auto props2 = fst2.Properties(kUnweighted, true);
-      if (!(props1 & kUnweighted) && !(props2 & kUnweighted)) {
-        FSTERROR() << "ComposeFst: Weights must be a commutative semiring: "
-                   << Weight::Type();
-        impl->SetProperties(kError, kError);
-      }
-    }
-    return impl;
-  }
-
-  // Create compose implementation specifying one matcher type; requires that
-  // input and matcher FST types be Fst<Arc>.
-  template <class Matcher, class Filter, class StateTuple>
-  static std::shared_ptr<Impl> CreateBase1(
-      const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-      const ComposeFstOptions<Arc, Matcher, Filter, StateTuple> &opts) {
-    ComposeFstImplOptions<Matcher, Matcher, Filter, StateTuple, CacheStore>
-        nopts(opts, opts.matcher1, opts.matcher2, opts.filter,
-              opts.state_table);
-    return CreateBase2(fst1, fst2, nopts);
-  }
-
-  // Create compose implementation specifying no matcher type.
-  static std::shared_ptr<Impl> CreateBase(const Fst<Arc> &fst1,
-                                          const Fst<Arc> &fst2,
-                                          const CacheOptions &opts) {
-    switch (LookAheadMatchType(fst1, fst2)) {  // Check for lookahead matchers
-      default:
-      case MATCH_NONE: {  // Default composition (no look-ahead).
-        ComposeFstOptions<Arc> nopts(opts);
-        return CreateBase1(fst1, fst2, nopts);
-      }
-      case MATCH_OUTPUT: {  // Lookahead on fst1.
-        using M = typename DefaultLookAhead<Arc, MATCH_OUTPUT>::FstMatcher;
-        using F = typename DefaultLookAhead<Arc, MATCH_OUTPUT>::ComposeFilter;
-        ComposeFstOptions<Arc, M, F> nopts(opts);
-        return CreateBase1(fst1, fst2, nopts);
-      }
-      case MATCH_INPUT: {  // Lookahead on fst2
-        using M = typename DefaultLookAhead<Arc, MATCH_INPUT>::FstMatcher;
-        using F = typename DefaultLookAhead<Arc, MATCH_INPUT>::ComposeFilter;
-        ComposeFstOptions<Arc, M, F> nopts(opts);
-        return CreateBase1(fst1, fst2, nopts);
-      }
-    }
-  }
-
- private:
-  ComposeFst &operator=(const ComposeFst &fst) = delete;
-};
-
-// Specialization for ComposeFst.
-template <class Arc, class CacheStore>
-class StateIterator<ComposeFst<Arc, CacheStore>>
-    : public CacheStateIterator<ComposeFst<Arc, CacheStore>> {
- public:
-  explicit StateIterator(const ComposeFst<Arc, CacheStore> &fst)
-      : CacheStateIterator<ComposeFst<Arc, CacheStore>>(fst,
-                                                        fst.GetMutableImpl()) {}
-};
-
-// Specialization for ComposeFst.
-template <class Arc, class CacheStore>
-class ArcIterator<ComposeFst<Arc, CacheStore>>
-    : public CacheArcIterator<ComposeFst<Arc, CacheStore>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const ComposeFst<Arc, CacheStore> &fst, StateId s)
-      : CacheArcIterator<ComposeFst<Arc, CacheStore>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc, class CacheStore>
-inline void ComposeFst<Arc, CacheStore>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<ComposeFst<Arc, CacheStore>>(*this);
-}
-
-// Specialized matcher for ComposeFst. Supports MATCH_INPUT or MATCH_OUTPUT,
-// iff the underlying matchers for the two FSTS being composed support
-// MATCH_INPUT or MATCH_OUTPUT, respectively.
-template <class CacheStore, class Filter, class StateTable>
-class ComposeFstMatcher : public MatcherBase<typename CacheStore::Arc> {
- public:
-  using Arc = typename CacheStore::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Matcher1 = typename Filter::Matcher1;
-  using Matcher2 = typename Filter::Matcher2;
-  using FilterState = typename Filter::FilterState;
-
-  using StateTuple = typename StateTable::StateTuple;
-
-  ComposeFstMatcher(
-      const ComposeFst<Arc, CacheStore> &fst,
-      const internal::ComposeFstImpl<CacheStore, Filter, StateTable> *impl,
-      MatchType match_type)
-      : fst_(fst),
-        impl_(impl),
-        s_(kNoStateId),
-        match_type_(match_type),
-        matcher1_(impl->matcher1_->Copy()),
-        matcher2_(impl->matcher2_->Copy()),
-        current_loop_(false),
-        loop_(kNoLabel, 0, Weight::One(), kNoStateId),
-        error_(false) {
-    if (match_type_ == MATCH_OUTPUT) std::swap(loop_.ilabel, loop_.olabel);
-  }
-
-  ComposeFstMatcher(
-      const ComposeFstMatcher<CacheStore, Filter, StateTable> &matcher,
-      bool safe = false)
-      : fst_(matcher.fst_),
-        impl_(matcher.impl_),
-        s_(kNoStateId),
-        match_type_(matcher.match_type_),
-        matcher1_(matcher.matcher1_->Copy(safe)),
-        matcher2_(matcher.matcher2_->Copy(safe)),
-        current_loop_(false),
-        loop_(kNoLabel, 0, Weight::One(), kNoStateId),
-        error_(matcher.error_) {
-    if (safe == true) {
-      FSTERROR() << "ComposeFstMatcher: Safe copy not supported";
-      error_ = true;
-    }
-    if (match_type_ == MATCH_OUTPUT) std::swap(loop_.ilabel, loop_.olabel);
-  }
-
-  ComposeFstMatcher<CacheStore, Filter, StateTable> *Copy(
-      bool safe = false) const override {
-    return new ComposeFstMatcher<CacheStore, Filter, StateTable>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override {
-    if ((matcher1_->Type(test) == MATCH_NONE) ||
-        (matcher2_->Type(test) == MATCH_NONE)) {
-      return MATCH_NONE;
-    }
-    if (((matcher1_->Type(test) == MATCH_UNKNOWN) &&
-         (matcher2_->Type(test) == MATCH_UNKNOWN)) ||
-        ((matcher1_->Type(test) == MATCH_UNKNOWN) &&
-         (matcher2_->Type(test) == match_type_)) ||
-        ((matcher1_->Type(test) == match_type_) &&
-         (matcher2_->Type(test) == MATCH_UNKNOWN))) {
-      return MATCH_UNKNOWN;
-    }
-    if ((matcher1_->Type(test) == match_type_) &&
-        (matcher2_->Type(test) == match_type_)) {
-      return match_type_;
-    }
-    return MATCH_NONE;
-  }
-
-  const Fst<Arc> &GetFst() const override { return fst_; }
-
-  uint64 Properties(uint64 inprops) const override {
-    auto outprops = inprops;
-    if (error_) outprops |= kError;
-    return outprops;
-  }
-
-  void SetState(StateId s) final {
-    if (s_ == s) return;
-    s_ = s;
-    const auto &tuple = impl_->state_table_->Tuple(s);
-    matcher1_->SetState(tuple.StateId1());
-    matcher2_->SetState(tuple.StateId2());
-    loop_.nextstate = s_;
-  }
-
-  bool Find(Label label) final {
-    bool found = false;
-    current_loop_ = false;
-    if (label == 0) {
-      current_loop_ = true;
-      found = true;
-    }
-    if (match_type_ == MATCH_INPUT) {
-      found = found || FindLabel(label, matcher1_.get(), matcher2_.get());
-    } else {  // match_type_ == MATCH_OUTPUT
-      found = found || FindLabel(label, matcher2_.get(), matcher1_.get());
-    }
-    return found;
-  }
-
-  bool Done() const final {
-    return !current_loop_ && matcher1_->Done() && matcher2_->Done();
-  }
-
-  const Arc &Value() const final { return current_loop_ ? loop_ : arc_; }
-
-  void Next() final {
-    if (current_loop_) {
-      current_loop_ = false;
-    } else if (match_type_ == MATCH_INPUT) {
-      FindNext(matcher1_.get(), matcher2_.get());
-    } else {  // match_type_ == MATCH_OUTPUT
-      FindNext(matcher2_.get(), matcher1_.get());
-    }
-  }
-
-  ssize_t Priority(StateId s) final { return fst_.NumArcs(s); }
-
- private:
-  // Processes a match with the filter and creates resulting arc.
-  bool MatchArc(StateId s, Arc arc1,
-                Arc arc2) {  // FIXME(kbg): copy but not assignment.
-    const auto &fs = impl_->filter_->FilterArc(&arc1, &arc2);
-    if (fs == FilterState::NoState()) return false;
-    const StateTuple tuple(arc1.nextstate, arc2.nextstate, fs);
-    arc_.ilabel = arc1.ilabel;
-    arc_.olabel = arc2.olabel;
-    arc_.weight = Times(arc1.weight, arc2.weight);
-    arc_.nextstate = impl_->state_table_->FindState(tuple);
-    return true;
-  }
-
-  // Finds the first match allowed by the filter.
-  template <class MatcherA, class MatcherB>
-  bool FindLabel(Label label, MatcherA *matchera, MatcherB *matcherb) {
-    if (matchera->Find(label)) {
-      matcherb->Find(match_type_ == MATCH_INPUT ? matchera->Value().olabel
-                                                : matchera->Value().ilabel);
-      return FindNext(matchera, matcherb);
-    }
-    return false;
-  }
-
-  // Finds the next match allowed by the filter, returning true iff such a
-  // match is found.
-  template <class MatcherA, class MatcherB>
-  bool FindNext(MatcherA *matchera, MatcherB *matcherb) {
-    // State when entering this function:
-    // 'matchera' is pointed to a match x, y for label x, and a match for y was
-    // requested on 'matcherb'.
-    while (!matchera->Done() || !matcherb->Done()) {
-      if (matcherb->Done()) {
-        // If no more matches for y on 'matcherb', moves forward on 'matchera'
-        // until a match x, y' is found such that there is a match for y' on
-        // 'matcherb'.
-        matchera->Next();
-        while (!matchera->Done() &&
-               !matcherb->Find(match_type_ == MATCH_INPUT
-                                   ? matchera->Value().olabel
-                                   : matchera->Value().ilabel)) {
-          matchera->Next();
-        }
-      }
-      while (!matcherb->Done()) {
-        // 'matchera' is pointing to a match x, y' ('arca') and 'matcherb' is
-        // pointing to a match y', z' ('arcb'). If combining these two arcs is
-        // allowed by the filter (hence resulting in an arc x, z') return true.
-        // Position 'matcherb' on the next potential match for y' before
-        // returning.
-        const auto &arca = matchera->Value();
-        const auto &arcb = matcherb->Value();
-        // Position 'matcherb' on the next potential match for y'.
-        matcherb->Next();
-        // Returns true If combining these two arcs is allowed by the filter
-        // (hence resulting in an arc x, z'); otherwise consider next match
-        // for y' on 'matcherb'.
-        if (MatchArc(s_, match_type_ == MATCH_INPUT ? arca : arcb,
-                     match_type_ == MATCH_INPUT ? arcb : arca)) {
-          return true;
-        }
-      }
-    }
-    // Both 'matchera' and 'matcherb' are done, no more match to analyse.
-    return false;
-  }
-  const ComposeFst<Arc, CacheStore> &fst_;
-  const internal::ComposeFstImpl<CacheStore, Filter, StateTable> *impl_;
-  StateId s_;
-  MatchType match_type_;
-  std::unique_ptr<Matcher1> matcher1_;
-  std::unique_ptr<Matcher2> matcher2_;
-  bool current_loop_;
-  Arc loop_;
-  Arc arc_;
-  bool error_;
-};
-
-// Useful alias when using StdArc.
-using StdComposeFst = ComposeFst<StdArc>;
-
-enum ComposeFilter {
-  AUTO_FILTER,
-  NULL_FILTER,
-  TRIVIAL_FILTER,
-  SEQUENCE_FILTER,
-  ALT_SEQUENCE_FILTER,
-  MATCH_FILTER
-};
-
-struct ComposeOptions {
-  bool connect;               // Connect output?
-  ComposeFilter filter_type;  // Pre-defined filter to use.
-
-  explicit ComposeOptions(bool connect = true,
-                          ComposeFilter filter_type = AUTO_FILTER)
-      : connect(connect), filter_type(filter_type) {}
-};
-
-// Computes the composition of two transducers. This version writes
-// the composed FST into a MutableFst. If FST1 transduces string x to
-// y with weight a and FST2 transduces y to z with weight b, then
-// their composition transduces string x to z with weight
-// Times(x, z).
-//
-// The output labels of the first transducer or the input labels of
-// the second transducer must be sorted.  The weights need to form a
-// commutative semiring (valid for TropicalWeight and LogWeight).
-//
-// Complexity:
-//
-// Assuming the first FST is unsorted and the second is sorted:
-//
-//   Time: O(V1 V2 D1 (log D2 + M2)),
-//   Space: O(V1 V2 D1 M2)
-//
-// where Vi = # of states, Di = maximum out-degree, and Mi is the maximum
-// multiplicity, for the ith FST.
-//
-// Caveats:
-//
-// - Compose trims its output.
-// - The efficiency of composition can be strongly affected by several factors:
-//   - the choice of which transducer is sorted - prefer sorting the FST
-//     that has the greater average out-degree.
-//   - the amount of non-determinism
-//   - the presence and location of epsilon transitions - avoid epsilon
-//     transitions on the output side of the first transducer or
-//     the input side of the second transducer or prefer placing
-//     them later in a path since they delay matching and can
-//     introduce non-coaccessible states and transitions.
-template <class Arc>
-void Compose(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-             MutableFst<Arc> *ofst,
-             const ComposeOptions &opts = ComposeOptions()) {
-  using M = Matcher<Fst<Arc>>;
-  // In each case, we cache only the last state for fastest copy.
-  switch (opts.filter_type) {
-    case AUTO_FILTER: {
-      CacheOptions nopts;
-      nopts.gc_limit = 0;
-      *ofst = ComposeFst<Arc>(ifst1, ifst2, nopts);
-      break;
-    }
-    case NULL_FILTER: {
-      ComposeFstOptions<Arc, M, NullComposeFilter<M>> copts;
-      copts.gc_limit = 0;
-      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-      break;
-    }
-    case SEQUENCE_FILTER: {
-      ComposeFstOptions<Arc, M, SequenceComposeFilter<M>> copts;
-      copts.gc_limit = 0;
-      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-      break;
-    }
-    case ALT_SEQUENCE_FILTER: {
-      ComposeFstOptions<Arc, M, AltSequenceComposeFilter<M>> copts;
-      copts.gc_limit = 0;
-      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-      break;
-    }
-    case MATCH_FILTER: {
-      ComposeFstOptions<Arc, M, MatchComposeFilter<M>> copts;
-      copts.gc_limit = 0;
-      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-      break;
-    }
-    case TRIVIAL_FILTER: {
-      ComposeFstOptions<Arc, M, TrivialComposeFilter<M>> copts;
-      copts.gc_limit = 0;
-      *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-      break;
-    }
-  }
-  if (opts.connect) Connect(ofst);
-}
-
-}  // namespace fst
-
-#endif  // FST_COMPOSE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/concat.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/concat.h
deleted file mode 100644
index 8bf6e3e5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/concat.h
+++ /dev/null
@@ -1,223 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to compute the concatenation of two FSTs.
-
-#ifndef FST_CONCAT_H_
-#define FST_CONCAT_H_
-
-#include <algorithm>
-#include <vector>
-
-#include <fst/mutable-fst.h>
-#include <fst/rational.h>
-
-
-namespace fst {
-
-// Computes the concatenation (product) of two FSTs. If FST1 transduces string
-// x to y with weight a and FST2 transduces string w to v with weight b, then
-// their concatenation transduces string xw to yv with weight Times(a, b).
-//
-// This version modifies its MutableFst argument (in first position).
-//
-// Complexity:
-//
-//   Time: O(V1 + V2 + E2)
-//   Space: O(V1 + V2 + E2)
-//
-// where Vi is the number of states, and Ei is the number of arcs, of the ith
-// FST.
-template <class Arc>
-void Concat(MutableFst<Arc> *fst1, const Fst<Arc> &fst2) {
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  // Checks that the symbol table are compatible.
-  if (!CompatSymbols(fst1->InputSymbols(), fst2.InputSymbols()) ||
-      !CompatSymbols(fst1->OutputSymbols(), fst2.OutputSymbols())) {
-    FSTERROR() << "Concat: Input/output symbol tables of 1st argument "
-               << "does not match input/output symbol tables of 2nd argument";
-    fst1->SetProperties(kError, kError);
-    return;
-  }
-  const auto props1 = fst1->Properties(kFstProperties, false);
-  const auto props2 = fst2.Properties(kFstProperties, false);
-  const auto start1 = fst1->Start();
-  if (start1 == kNoStateId) {
-    if (props2 & kError) fst1->SetProperties(kError, kError);
-    return;
-  }
-  const auto numstates1 = fst1->NumStates();
-  if (fst2.Properties(kExpanded, false)) {
-    fst1->ReserveStates(numstates1 + CountStates(fst2));
-  }
-  for (StateIterator<Fst<Arc>> siter2(fst2); !siter2.Done(); siter2.Next()) {
-    const auto s1 = fst1->AddState();
-    const auto s2 = siter2.Value();
-    fst1->SetFinal(s1, fst2.Final(s2));
-    fst1->ReserveArcs(s1, fst2.NumArcs(s2));
-    for (ArcIterator<Fst<Arc>> aiter(fst2, s2); !aiter.Done(); aiter.Next()) {
-      auto arc = aiter.Value();
-      arc.nextstate += numstates1;
-      fst1->AddArc(s1, arc);
-    }
-  }
-  const auto start2 = fst2.Start();
-  for (StateId s1 = 0; s1 < numstates1; ++s1) {
-    const auto weight = fst1->Final(s1);
-    if (weight != Weight::Zero()) {
-      fst1->SetFinal(s1, Weight::Zero());
-      if (start2 != kNoStateId) {
-        fst1->AddArc(s1, Arc(0, 0, weight, start2 + numstates1));
-      }
-    }
-  }
-  if (start2 != kNoStateId) {
-    fst1->SetProperties(ConcatProperties(props1, props2), kFstProperties);
-  }
-}
-
-// Computes the concatentation of two FSTs.  This version modifies its
-// MutableFst argument (in second position).
-//
-// Complexity:
-//
-//   Time: O(V1 + E1)
-//   Space: O(V1 + E1)
-//
-// where Vi is the number of states, and Ei is the number of arcs, of the ith
-// FST.
-template <class Arc>
-void Concat(const Fst<Arc> &fst1, MutableFst<Arc> *fst2) {
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  // Checks that the symbol table are compatible.
-  if (!CompatSymbols(fst1.InputSymbols(), fst2->InputSymbols()) ||
-      !CompatSymbols(fst1.OutputSymbols(), fst2->OutputSymbols())) {
-    FSTERROR() << "Concat: Input/output symbol tables of 1st argument "
-               << "does not match input/output symbol tables of 2nd argument";
-    fst2->SetProperties(kError, kError);
-    return;
-  }
-  const auto props1 = fst1.Properties(kFstProperties, false);
-  const auto props2 = fst2->Properties(kFstProperties, false);
-  const auto start2 = fst2->Start();
-  if (start2 == kNoStateId) {
-    if (props1 & kError) fst2->SetProperties(kError, kError);
-    return;
-  }
-  const auto numstates2 = fst2->NumStates();
-  if (fst1.Properties(kExpanded, false)) {
-    fst2->ReserveStates(numstates2 + CountStates(fst1));
-  }
-  for (StateIterator<Fst<Arc>> siter(fst1); !siter.Done(); siter.Next()) {
-    const auto s1 = siter.Value();
-    const auto s2 = fst2->AddState();
-    const auto weight = fst1.Final(s1);
-    if (weight != Weight::Zero()) {
-      fst2->ReserveArcs(s2, fst1.NumArcs(s1) + 1);
-      fst2->AddArc(s2, Arc(0, 0, weight, start2));
-    } else {
-      fst2->ReserveArcs(s2, fst1.NumArcs(s1));
-    }
-    for (ArcIterator<Fst<Arc>> aiter(fst1, s1); !aiter.Done(); aiter.Next()) {
-      auto arc = aiter.Value();
-      arc.nextstate += numstates2;
-      fst2->AddArc(s2, arc);
-    }
-  }
-  const auto start1 = fst1.Start();
-  if (start1 != kNoStateId) {
-    fst2->SetStart(start1 + numstates2);
-    fst2->SetProperties(ConcatProperties(props1, props2), kFstProperties);
-  } else {
-    fst2->SetStart(fst2->AddState());
-  }
-}
-
-// Computes the concatentation of two FSTs. This version modifies its
-// RationalFst input (in first position).
-template <class Arc>
-void Concat(RationalFst<Arc> *fst1, const Fst<Arc> &fst2) {
-  fst1->GetMutableImpl()->AddConcat(fst2, true);
-}
-
-// Computes the concatentation of two FSTs. This version modifies its
-// RationalFst input (in second position).
-template <class Arc>
-void Concat(const Fst<Arc> &fst1, RationalFst<Arc> *fst2) {
-  fst2->GetMutableImpl()->AddConcat(fst1, false);
-}
-
-using ConcatFstOptions = RationalFstOptions;
-
-// Computes the concatenation (product) of two FSTs; this version is a delayed
-// FST. If FST1 transduces string x to y with weight a and FST2 transduces
-// string w to v with weight b, then their concatenation transduces string xw
-// to yv with Times(a, b).
-//
-// Complexity:
-//
-//   Time: O(v1 + e1 + v2 + e2),
-//   Space: O(v1 + v2)
-//
-// where vi is the number of states visited, and ei is the number of arcs
-// visited, of the ith FST. Constant time and space to visit an input state or
-// arc is assumed and exclusive of caching.
-template <class A>
-class ConcatFst : public RationalFst<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  ConcatFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {
-    GetMutableImpl()->InitConcat(fst1, fst2);
-  }
-
-  ConcatFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-            const ConcatFstOptions &opts)
-      : RationalFst<Arc>(opts) {
-    GetMutableImpl()->InitConcat(fst1, fst2);
-  }
-
-  // See Fst<>::Copy() for doc.
-  ConcatFst(const ConcatFst<Arc> &fst, bool safe = false)
-      : RationalFst<Arc>(fst, safe) {}
-
-  // Get a copy of this ConcatFst. See Fst<>::Copy() for further doc.
-  ConcatFst<Arc> *Copy(bool safe = false) const override {
-    return new ConcatFst<Arc>(*this, safe);
-  }
-
- private:
-  using ImplToFst<internal::RationalFstImpl<Arc>>::GetImpl;
-  using ImplToFst<internal::RationalFstImpl<Arc>>::GetMutableImpl;
-};
-
-// Specialization for ConcatFst.
-template <class Arc>
-class StateIterator<ConcatFst<Arc>> : public StateIterator<RationalFst<Arc>> {
- public:
-  explicit StateIterator(const ConcatFst<Arc> &fst)
-      : StateIterator<RationalFst<Arc>>(fst) {}
-};
-
-// Specialization for ConcatFst.
-template <class Arc>
-class ArcIterator<ConcatFst<Arc>> : public ArcIterator<RationalFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const ConcatFst<Arc> &fst, StateId s)
-      : ArcIterator<RationalFst<Arc>>(fst, s) {}
-};
-
-// Useful alias when using StdArc.
-using StdConcatFst = ConcatFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_CONCAT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/config.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/config.h
deleted file mode 100644
index 32f2a653..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/config.h
+++ /dev/null
@@ -1,3 +0,0 @@
-// Windows-specific OpenFst config file
-// No dynamic registration.
-#define FST_NO_DYNAMIC_LINKING 1
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/connect.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/connect.h
deleted file mode 100644
index 8b1ddaf4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/connect.h
+++ /dev/null
@@ -1,314 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes and functions to remove unsuccessful paths from an FST.
-
-#ifndef FST_CONNECT_H_
-#define FST_CONNECT_H_
-
-#include <vector>
-
-#include <fst/dfs-visit.h>
-#include <fst/mutable-fst.h>
-#include <fst/union-find.h>
-
-
-namespace fst {
-
-// Finds and returns connected components. Use with Visit().
-template <class Arc>
-class CcVisitor {
- public:
-  using Weight = typename Arc::Weight;
-  using StateId = typename Arc::StateId;
-
-  // cc[i]: connected component number for state i.
-  explicit CcVisitor(std::vector<StateId> *cc)
-      : comps_(new UnionFind<StateId>(0, kNoStateId)), cc_(cc), nstates_(0) {}
-
-  // comps: connected components equiv classes.
-  explicit CcVisitor(UnionFind<StateId> *comps)
-      : comps_(comps), cc_(nullptr), nstates_(0) {}
-
-  ~CcVisitor() {
-    if (cc_) delete comps_;
-  }
-
-  void InitVisit(const Fst<Arc> &fst) {}
-
-  bool InitState(StateId s, StateId root) {
-    ++nstates_;
-    if (comps_->FindSet(s) == kNoStateId) comps_->MakeSet(s);
-    return true;
-  }
-
-  bool WhiteArc(StateId s, const Arc &arc) {
-    comps_->MakeSet(arc.nextstate);
-    comps_->Union(s, arc.nextstate);
-    return true;
-  }
-
-  bool GreyArc(StateId s, const Arc &arc) {
-    comps_->Union(s, arc.nextstate);
-    return true;
-  }
-
-  bool BlackArc(StateId s, const Arc &arc) {
-    comps_->Union(s, arc.nextstate);
-    return true;
-  }
-
-  void FinishState(StateId s) {}
-
-  void FinishVisit() {
-    if (cc_) GetCcVector(cc_);
-  }
-
-  // Returns number of components.
-  // cc[i]: connected component number for state i.
-  int GetCcVector(std::vector<StateId> *cc) {
-    cc->clear();
-    cc->resize(nstates_, kNoStateId);
-    StateId ncomp = 0;
-    for (StateId s = 0; s < nstates_; ++s) {
-      const auto rep = comps_->FindSet(s);
-      auto &comp = (*cc)[rep];
-      if (comp == kNoStateId) {
-        comp = ncomp;
-        ++ncomp;
-      }
-      (*cc)[s] = comp;
-    }
-    return ncomp;
-  }
-
- private:
-  UnionFind<StateId> *comps_;  // Components.
-  std::vector<StateId> *cc_;   // State's cc number.
-  StateId nstates_;            // State count.
-};
-
-// Finds and returns strongly-connected components, accessible and
-// coaccessible states and related properties. Uses Tarjan's single
-// DFS SCC algorithm (see Aho, et al, "Design and Analysis of Computer
-// Algorithms", 189pp). Use with DfsVisit();
-template <class Arc>
-class SccVisitor {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // scc[i]: strongly-connected component number for state i.
-  //   SCC numbers will be in topological order for acyclic input.
-  // access[i]: accessibility of state i.
-  // coaccess[i]: coaccessibility of state i.
-  // Any of above can be NULL.
-  // props: related property bits (cyclicity, initial cyclicity,
-  //   accessibility, coaccessibility) set/cleared (o.w. unchanged).
-  SccVisitor(std::vector<StateId> *scc, std::vector<bool> *access,
-             std::vector<bool> *coaccess, uint64 *props)
-      : scc_(scc), access_(access), coaccess_(coaccess), props_(props) {}
-  explicit SccVisitor(uint64 *props)
-      : scc_(nullptr), access_(nullptr), coaccess_(nullptr), props_(props) {}
-
-  void InitVisit(const Fst<Arc> &fst);
-
-  bool InitState(StateId s, StateId root);
-
-  bool TreeArc(StateId s, const Arc &arc) { return true; }
-
-  bool BackArc(StateId s, const Arc &arc) {
-    const auto t = arc.nextstate;
-    if ((*dfnumber_)[t] < (*lowlink_)[s]) (*lowlink_)[s] = (*dfnumber_)[t];
-    if ((*coaccess_)[t]) (*coaccess_)[s] = true;
-    *props_ |= kCyclic;
-    *props_ &= ~kAcyclic;
-    if (t == start_) {
-      *props_ |= kInitialCyclic;
-      *props_ &= ~kInitialAcyclic;
-    }
-    return true;
-  }
-
-  bool ForwardOrCrossArc(StateId s, const Arc &arc) {
-    const auto t = arc.nextstate;
-    if ((*dfnumber_)[t] < (*dfnumber_)[s] /* cross edge */ && (*onstack_)[t] &&
-        (*dfnumber_)[t] < (*lowlink_)[s]) {
-      (*lowlink_)[s] = (*dfnumber_)[t];
-    }
-    if ((*coaccess_)[t]) (*coaccess_)[s] = true;
-    return true;
-  }
-
-  // Last argument always ignored, but required by the interface.
-  void FinishState(StateId state, StateId p, const Arc *);
-
-  void FinishVisit() {
-    // Numbers SCCs in topological order when acyclic.
-    if (scc_) {
-      for (StateId s = 0; s < scc_->size(); ++s) {
-        (*scc_)[s] = nscc_ - 1 - (*scc_)[s];
-      }
-    }
-    if (coaccess_internal_) delete coaccess_;
-    dfnumber_.reset();
-    lowlink_.reset();
-    onstack_.reset();
-    scc_stack_.reset();
-  }
-
- private:
-  std::vector<StateId> *scc_;    // State's scc number.
-  std::vector<bool> *access_;    // State's accessibility.
-  std::vector<bool> *coaccess_;  // State's coaccessibility.
-  uint64 *props_;
-  const Fst<Arc> *fst_;
-  StateId start_;
-  StateId nstates_;  // State count.
-  StateId nscc_;     // SCC count.
-  bool coaccess_internal_;
-  std::unique_ptr<std::vector<StateId>> dfnumber_;  // State discovery times.
-  std::unique_ptr<std::vector<StateId>>
-      lowlink_;  // lowlink[state] == dfnumber[state] => SCC root
-  std::unique_ptr<std::vector<bool>> onstack_;  // Is a state on the SCC stack?
-  std::unique_ptr<std::vector<StateId>>
-      scc_stack_;  // SCC stack, with random access.
-};
-
-template <class Arc>
-inline void SccVisitor<Arc>::InitVisit(const Fst<Arc> &fst) {
-  if (scc_) scc_->clear();
-  if (access_) access_->clear();
-  if (coaccess_) {
-    coaccess_->clear();
-    coaccess_internal_ = false;
-  } else {
-    coaccess_ = new std::vector<bool>;
-    coaccess_internal_ = true;
-  }
-  *props_ |= kAcyclic | kInitialAcyclic | kAccessible | kCoAccessible;
-  *props_ &= ~(kCyclic | kInitialCyclic | kNotAccessible | kNotCoAccessible);
-  fst_ = &fst;
-  start_ = fst.Start();
-  nstates_ = 0;
-  nscc_ = 0;
-  dfnumber_.reset(new std::vector<StateId>());
-  lowlink_.reset(new std::vector<StateId>());
-  onstack_.reset(new std::vector<bool>());
-  scc_stack_.reset(new std::vector<StateId>());
-}
-
-template <class Arc>
-inline bool SccVisitor<Arc>::InitState(StateId s, StateId root) {
-  scc_stack_->push_back(s);
-  while (dfnumber_->size() <= s) {
-    if (scc_) scc_->push_back(-1);
-    if (access_) access_->push_back(false);
-    coaccess_->push_back(false);
-    dfnumber_->push_back(-1);
-    lowlink_->push_back(-1);
-    onstack_->push_back(false);
-  }
-  (*dfnumber_)[s] = nstates_;
-  (*lowlink_)[s] = nstates_;
-  (*onstack_)[s] = true;
-  if (root == start_) {
-    if (access_) (*access_)[s] = true;
-  } else {
-    if (access_) (*access_)[s] = false;
-    *props_ |= kNotAccessible;
-    *props_ &= ~kAccessible;
-  }
-  ++nstates_;
-  return true;
-}
-
-template <class Arc>
-inline void SccVisitor<Arc>::FinishState(StateId s, StateId p, const Arc *) {
-  if (fst_->Final(s) != Weight::Zero()) (*coaccess_)[s] = true;
-  if ((*dfnumber_)[s] == (*lowlink_)[s]) {  // Root of new SCC.
-    bool scc_coaccess = false;
-    auto i = scc_stack_->size();
-    StateId t;
-    do {
-      t = (*scc_stack_)[--i];
-      if ((*coaccess_)[t]) scc_coaccess = true;
-    } while (s != t);
-    do {
-      t = scc_stack_->back();
-      if (scc_) (*scc_)[t] = nscc_;
-      if (scc_coaccess) (*coaccess_)[t] = true;
-      (*onstack_)[t] = false;
-      scc_stack_->pop_back();
-    } while (s != t);
-    if (!scc_coaccess) {
-      *props_ |= kNotCoAccessible;
-      *props_ &= ~kCoAccessible;
-    }
-    ++nscc_;
-  }
-  if (p != kNoStateId) {
-    if ((*coaccess_)[s]) (*coaccess_)[p] = true;
-    if ((*lowlink_)[s] < (*lowlink_)[p]) (*lowlink_)[p] = (*lowlink_)[s];
-  }
-}
-
-// Trims an FST, removing states and arcs that are not on successful paths.
-// This version modifies its input.
-//
-// Complexity:
-//
-//   Time:  O(V + E)
-//   Space: O(V + E)
-//
-// where V = # of states and E = # of arcs.
-template <class Arc>
-void Connect(MutableFst<Arc> *fst) {
-  using StateId = typename Arc::StateId;
-  std::vector<bool> access;
-  std::vector<bool> coaccess;
-  uint64 props = 0;
-  SccVisitor<Arc> scc_visitor(nullptr, &access, &coaccess, &props);
-  DfsVisit(*fst, &scc_visitor);
-  std::vector<StateId> dstates;
-  for (StateId s = 0; s < access.size(); ++s) {
-    if (!access[s] || !coaccess[s]) dstates.push_back(s);
-  }
-  fst->DeleteStates(dstates);
-  fst->SetProperties(kAccessible | kCoAccessible, kAccessible | kCoAccessible);
-}
-
-// Returns an acyclic FST where each SCC in the input FST has been condensed to
-// a single state with transitions between SCCs retained and within SCCs
-// dropped. Also populates 'scc' with a mapping from input to output states.
-template <class Arc>
-void Condense(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-              std::vector<typename Arc::StateId> *scc) {
-  using StateId = typename Arc::StateId;
-  ofst->DeleteStates();
-  uint64 props = 0;
-  SccVisitor<Arc> scc_visitor(scc, nullptr, nullptr, &props);
-  DfsVisit(ifst, &scc_visitor);
-  for (StateId s = 0; s < scc->size(); ++s) {
-    const auto c = (*scc)[s];
-    while (c >= ofst->NumStates()) ofst->AddState();
-    if (s == ifst.Start()) ofst->SetStart(c);
-    const auto weight = ifst.Final(s);
-    if (weight != Arc::Weight::Zero())
-      ofst->SetFinal(c, Plus(ofst->Final(c), weight));
-    for (ArcIterator<Fst<Arc>> aiter(ifst, s); !aiter.Done(); aiter.Next()) {
-      auto arc = aiter.Value();
-      const auto nextc = (*scc)[arc.nextstate];
-      if (nextc != c) {
-        while (nextc >= ofst->NumStates()) ofst->AddState();
-        arc.nextstate = nextc;
-        ofst->AddArc(c, arc);
-      }
-    }
-  }
-  ofst->SetProperties(kAcyclic | kInitialAcyclic, kAcyclic | kInitialAcyclic);
-}
-
-}  // namespace fst
-
-#endif  // FST_CONNECT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/const-fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/const-fst.h
deleted file mode 100644
index 1a2cba6a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/const-fst.h
+++ /dev/null
@@ -1,488 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Simple concrete immutable FST whose states and arcs are each stored in
-// single arrays.
-
-#ifndef FST_CONST_FST_H_
-#define FST_CONST_FST_H_
-
-#include <climits>
-#include <string>
-#include <vector>
-
-// Google-only...
-// ...Google-only
-#include <fst/log.h>
-
-#include <fst/expanded-fst.h>
-#include <fst/fst-decl.h>
-#include <fst/mapped-file.h>
-#include <fst/test-properties.h>
-#include <fst/util.h>
-
-
-namespace fst {
-
-template <class A, class Unsigned>
-class ConstFst;
-
-template <class F, class G>
-void Cast(const F &, G *);
-
-namespace internal {
-
-// States and arcs each implemented by single arrays, templated on the
-// Arc definition. Unsigned is used to represent indices into the arc array.
-template <class A, class Unsigned>
-class ConstFstImpl : public FstImpl<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<A>::SetInputSymbols;
-  using FstImpl<A>::SetOutputSymbols;
-  using FstImpl<A>::SetType;
-  using FstImpl<A>::SetProperties;
-  using FstImpl<A>::Properties;
-
-  ConstFstImpl()
-      : states_(nullptr),
-        arcs_(nullptr),
-        nstates_(0),
-        narcs_(0),
-        start_(kNoStateId) {
-    string type = "const";
-    if (sizeof(Unsigned) != sizeof(uint32)) {
-      type += std::to_string(CHAR_BIT * sizeof(Unsigned));
-    }
-    SetType(type);
-    SetProperties(kNullProperties | kStaticProperties);
-  }
-
-  explicit ConstFstImpl(const Fst<Arc> &fst);
-
-  StateId Start() const { return start_; }
-
-  Weight Final(StateId s) const { return states_[s].weight; }
-
-  StateId NumStates() const { return nstates_; }
-
-  size_t NumArcs(StateId s) const { return states_[s].narcs; }
-
-  size_t NumInputEpsilons(StateId s) const { return states_[s].niepsilons; }
-
-  size_t NumOutputEpsilons(StateId s) const { return states_[s].noepsilons; }
-
-  static ConstFstImpl<Arc, Unsigned> *Read(std::istream &strm,
-                                           const FstReadOptions &opts);
-
-  const Arc *Arcs(StateId s) const { return arcs_ + states_[s].pos; }
-
-  // Provide information needed for generic state iterator.
-  void InitStateIterator(StateIteratorData<Arc> *data) const {
-    data->base = nullptr;
-    data->nstates = nstates_;
-  }
-
-  // Provide information needed for the generic arc iterator.
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const {
-    data->base = nullptr;
-    data->arcs = arcs_ + states_[s].pos;
-    data->narcs = states_[s].narcs;
-    data->ref_count = nullptr;
-  }
-
- private:
-  // Used to find narcs_ and nstates_ in Write.
-  friend class ConstFst<Arc, Unsigned>;
-
-  // States implemented by array *states_ below, arcs by (single) *arcs_.
-  struct ConstState {
-    Weight weight;        // Final weight.
-    Unsigned pos;         // Start of state's arcs in *arcs_.
-    Unsigned narcs;       // Number of arcs (per state).
-    Unsigned niepsilons;  // Number of input epsilons.
-    Unsigned noepsilons;  // Number of output epsilons.
-
-    ConstState() : weight(Weight::Zero()) {}
-  };
-
-  // Properties always true of this FST class.
-  static constexpr uint64 kStaticProperties = kExpanded;
-  // Current unaligned file format version. The unaligned version was added and
-  // made the default since the aligned version does not work on pipes.
-  static constexpr int kFileVersion = 2;
-  // Current aligned file format version.
-  static constexpr int kAlignedFileVersion = 1;
-  // Minimum file format version supported.
-  static constexpr int kMinFileVersion = 1;
-
-  std::unique_ptr<MappedFile> states_region_;  // Mapped file for states.
-  std::unique_ptr<MappedFile> arcs_region_;    // Mapped file for arcs.
-  ConstState *states_;                         // States representation.
-  Arc *arcs_;                                  // Arcs representation.
-  StateId nstates_;                            // Number of states.
-  size_t narcs_;                               // Number of arcs.
-  StateId start_;                              // Initial state.
-
-  ConstFstImpl(const ConstFstImpl &) = delete;
-  ConstFstImpl &operator=(const ConstFstImpl &) = delete;
-};
-
-template <class Arc, class Unsigned>
-constexpr uint64 ConstFstImpl<Arc, Unsigned>::kStaticProperties;
-
-template <class Arc, class Unsigned>
-constexpr int ConstFstImpl<Arc, Unsigned>::kFileVersion;
-
-template <class Arc, class Unsigned>
-constexpr int ConstFstImpl<Arc, Unsigned>::kAlignedFileVersion;
-
-template <class Arc, class Unsigned>
-constexpr int ConstFstImpl<Arc, Unsigned>::kMinFileVersion;
-
-template <class Arc, class Unsigned>
-ConstFstImpl<Arc, Unsigned>::ConstFstImpl(const Fst<Arc> &fst)
-    : nstates_(0), narcs_(0) {
-  string type = "const";
-  if (sizeof(Unsigned) != sizeof(uint32)) {
-    type += std::to_string(CHAR_BIT * sizeof(Unsigned));
-  }
-  SetType(type);
-  SetInputSymbols(fst.InputSymbols());
-  SetOutputSymbols(fst.OutputSymbols());
-  start_ = fst.Start();
-  // Counts states and arcs.
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    ++nstates_;
-    for (ArcIterator<Fst<Arc>> aiter(fst, siter.Value()); !aiter.Done();
-         aiter.Next()) {
-      ++narcs_;
-    }
-  }
-  states_region_.reset(MappedFile::Allocate(nstates_ * sizeof(*states_)));
-  arcs_region_.reset(MappedFile::Allocate(narcs_ * sizeof(*arcs_)));
-  states_ = reinterpret_cast<ConstState *>(states_region_->mutable_data());
-  arcs_ = reinterpret_cast<Arc *>(arcs_region_->mutable_data());
-  size_t pos = 0;
-  for (StateId s = 0; s < nstates_; ++s) {
-    states_[s].weight = fst.Final(s);
-    states_[s].pos = pos;
-    states_[s].narcs = 0;
-    states_[s].niepsilons = 0;
-    states_[s].noepsilons = 0;
-    for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      ++states_[s].narcs;
-      if (arc.ilabel == 0) ++states_[s].niepsilons;
-      if (arc.olabel == 0) ++states_[s].noepsilons;
-      arcs_[pos] = arc;
-      ++pos;
-    }
-  }
-  const auto props =
-      fst.Properties(kMutable, false)
-          ? fst.Properties(kCopyProperties, true)
-          : CheckProperties(
-                fst, kCopyProperties & ~kWeightedCycles & ~kUnweightedCycles,
-                kCopyProperties);
-  SetProperties(props | kStaticProperties);
-}
-
-template <class Arc, class Unsigned>
-ConstFstImpl<Arc, Unsigned> *ConstFstImpl<Arc, Unsigned>::Read(
-    std::istream &strm, const FstReadOptions &opts) {
-  using ConstState = typename ConstFstImpl<Arc, Unsigned>::ConstState;
-  std::unique_ptr<ConstFstImpl<Arc, Unsigned>> impl(
-      new ConstFstImpl<Arc, Unsigned>());
-  FstHeader hdr;
-  if (!impl->ReadHeader(strm, opts, kMinFileVersion, &hdr)) return nullptr;
-  impl->start_ = hdr.Start();
-  impl->nstates_ = hdr.NumStates();
-  impl->narcs_ = hdr.NumArcs();
-  // Ensures compatibility.
-  if (hdr.Version() == kAlignedFileVersion) {
-    hdr.SetFlags(hdr.GetFlags() | FstHeader::IS_ALIGNED);
-  }
-  if ((hdr.GetFlags() & FstHeader::IS_ALIGNED) && !AlignInput(strm)) {
-    LOG(ERROR) << "ConstFst::Read: Alignment failed: " << opts.source;
-    return nullptr;
-  }
-  size_t b = impl->nstates_ * sizeof(ConstState);
-  impl->states_region_.reset(
-      MappedFile::Map(&strm, opts.mode == FstReadOptions::MAP, opts.source, b));
-  if (!strm || !impl->states_region_) {
-    LOG(ERROR) << "ConstFst::Read: Read failed: " << opts.source;
-    return nullptr;
-  }
-  impl->states_ =
-      reinterpret_cast<ConstState *>(impl->states_region_->mutable_data());
-  if ((hdr.GetFlags() & FstHeader::IS_ALIGNED) && !AlignInput(strm)) {
-    LOG(ERROR) << "ConstFst::Read: Alignment failed: " << opts.source;
-    return nullptr;
-  }
-  b = impl->narcs_ * sizeof(Arc);
-  impl->arcs_region_.reset(
-      MappedFile::Map(&strm, opts.mode == FstReadOptions::MAP, opts.source, b));
-  if (!strm || !impl->arcs_region_) {
-    LOG(ERROR) << "ConstFst::Read: Read failed: " << opts.source;
-    return nullptr;
-  }
-  impl->arcs_ = reinterpret_cast<Arc *>(impl->arcs_region_->mutable_data());
-  return impl.release();
-}
-
-}  // namespace internal
-
-// Simple concrete immutable FST. This class attaches interface to
-// implementation and handles reference counting, delegating most methods to
-// ImplToExpandedFst. The unsigned type U is used to represent indices into the
-// arc array (default declared in fst-decl.h).
-template <class A, class Unsigned>
-class ConstFst : public ImplToExpandedFst<internal::ConstFstImpl<A, Unsigned>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  using Impl = internal::ConstFstImpl<A, Unsigned>;
-  using ConstState = typename Impl::ConstState;
-
-  friend class StateIterator<ConstFst<Arc, Unsigned>>;
-  friend class ArcIterator<ConstFst<Arc, Unsigned>>;
-
-  template <class F, class G>
-  void friend Cast(const F &, G *);
-
-  ConstFst() : ImplToExpandedFst<Impl>(std::make_shared<Impl>()) {}
-
-  explicit ConstFst(const Fst<Arc> &fst)
-      : ImplToExpandedFst<Impl>(std::make_shared<Impl>(fst)) {}
-
-  ConstFst(const ConstFst<A, Unsigned> &fst, bool safe = false)
-      : ImplToExpandedFst<Impl>(fst) {}
-
-  // Gets a copy of this ConstFst. See Fst<>::Copy() for further doc.
-  ConstFst<A, Unsigned> *Copy(bool safe = false) const override {
-    return new ConstFst<A, Unsigned>(*this, safe);
-  }
-
-  // Reads a ConstFst from an input stream, returning nullptr on error.
-  static ConstFst<A, Unsigned> *Read(std::istream &strm,
-                                     const FstReadOptions &opts) {
-    auto *impl = Impl::Read(strm, opts);
-    return impl ? new ConstFst<A, Unsigned>(std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  // Read a ConstFst from a file; return nullptr on error; empty filename reads
-  // from standard input.
-  static ConstFst<A, Unsigned> *Read(const string &filename) {
-    auto *impl = ImplToExpandedFst<Impl>::Read(filename);
-    return impl ? new ConstFst<A, Unsigned>(std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return WriteFst(*this, strm, opts);
-  }
-
-  bool Write(const string &filename) const override {
-    return Fst<Arc>::WriteFile(filename);
-  }
-
-  template <class FST>
-  static bool WriteFst(const FST &fst, std::ostream &strm,
-                       const FstWriteOptions &opts);
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const override {
-    GetImpl()->InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetImpl()->InitArcIterator(s, data);
-  }
-
- private:
-  explicit ConstFst(std::shared_ptr<Impl> impl)
-      : ImplToExpandedFst<Impl>(impl) {}
-
-  using ImplToFst<Impl, ExpandedFst<Arc>>::GetImpl;
-
-  // Uses overloading to extract the type of the argument.
-  static const Impl *GetImplIfConstFst(const ConstFst &const_fst) {
-    return const_fst.GetImpl();
-  }
-
-  // NB: this does not give privileged treatment to subtypes of ConstFst.
-  template <typename FST>
-  static Impl *GetImplIfConstFst(const FST &fst) {
-    return nullptr;
-  }
-
-  ConstFst &operator=(const ConstFst &) = delete;
-};
-
-// Writes FST in Const format, potentially with a pass over the machine before
-// writing to compute number of states and arcs.
-template <class Arc, class Unsigned>
-template <class FST>
-bool ConstFst<Arc, Unsigned>::WriteFst(const FST &fst, std::ostream &strm,
-                                       const FstWriteOptions &opts) {
-  const auto file_version =
-      opts.align ? internal::ConstFstImpl<Arc, Unsigned>::kAlignedFileVersion
-                 : internal::ConstFstImpl<Arc, Unsigned>::kFileVersion;
-  size_t num_arcs = 0;    // To silence -Wsometimes-uninitialized warnings.
-  size_t num_states = 0;  // Ditto.
-  size_t start_offset = 0;
-  bool update_header = true;
-  if (const auto *impl = GetImplIfConstFst(fst)) {
-    num_arcs = impl->narcs_;
-    num_states = impl->nstates_;
-    update_header = false;
-  } else if (opts.stream_write || (start_offset = strm.tellp()) == -1) {
-    // precompute values needed for header when we cannot seek to rewrite it.
-    num_arcs = 0;
-    num_states = 0;
-    for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-      num_arcs += fst.NumArcs(siter.Value());
-      ++num_states;
-    }
-    update_header = false;
-  }
-  FstHeader hdr;
-  hdr.SetStart(fst.Start());
-  hdr.SetNumStates(num_states);
-  hdr.SetNumArcs(num_arcs);
-  string type = "const";
-  if (sizeof(Unsigned) != sizeof(uint32)) {
-    type += std::to_string(CHAR_BIT * sizeof(Unsigned));
-  }
-  const auto properties =
-      fst.Properties(kCopyProperties, true) |
-      internal::ConstFstImpl<Arc, Unsigned>::kStaticProperties;
-  internal::FstImpl<Arc>::WriteFstHeader(fst, strm, opts, file_version, type,
-                                         properties, &hdr);
-  if (opts.align && !AlignOutput(strm)) {
-    LOG(ERROR) << "Could not align file during write after header";
-    return false;
-  }
-  size_t pos = 0;
-  size_t states = 0;
-  typename ConstFst<Arc, Unsigned>::ConstState state;
-  for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    state.weight = fst.Final(s);
-    state.pos = pos;
-    state.narcs = fst.NumArcs(s);
-    state.niepsilons = fst.NumInputEpsilons(s);
-    state.noepsilons = fst.NumOutputEpsilons(s);
-    strm.write(reinterpret_cast<const char *>(&state), sizeof(state));
-    pos += state.narcs;
-    ++states;
-  }
-  hdr.SetNumStates(states);
-  hdr.SetNumArcs(pos);
-  if (opts.align && !AlignOutput(strm)) {
-    LOG(ERROR) << "Could not align file during write after writing states";
-  }
-  for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-    for (ArcIterator<FST> aiter(fst, siter.Value()); !aiter.Done();
-         aiter.Next()) {
-      const auto &arc = aiter.Value();
-// Google-only...
-#ifdef MEMORY_SANITIZER
-      // arc may contain padding which has unspecified contents. Tell MSAN to
-      // not complain about it when writing it to a file.
-      ANNOTATE_MEMORY_IS_INITIALIZED(reinterpret_cast<const char *>(&arc),
-                                     sizeof(arc));
-#endif
-      // ...Google-only
-      strm.write(reinterpret_cast<const char *>(&arc), sizeof(arc));
-    }
-  }
-  strm.flush();
-  if (!strm) {
-    LOG(ERROR) << "ConstFst::WriteFst: write failed: " << opts.source;
-    return false;
-  }
-  if (update_header) {
-    return internal::FstImpl<Arc>::UpdateFstHeader(
-        fst, strm, opts, file_version, type, properties, &hdr, start_offset);
-  } else {
-    if (hdr.NumStates() != num_states) {
-      LOG(ERROR) << "Inconsistent number of states observed during write";
-      return false;
-    }
-    if (hdr.NumArcs() != num_arcs) {
-      LOG(ERROR) << "Inconsistent number of arcs observed during write";
-      return false;
-    }
-  }
-  return true;
-}
-
-// Specialization for ConstFst; see generic version in fst.h for sample usage
-// (but use the ConstFst type instead). This version should inline.
-template <class Arc, class Unsigned>
-class StateIterator<ConstFst<Arc, Unsigned>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  explicit StateIterator(const ConstFst<Arc, Unsigned> &fst)
-      : nstates_(fst.GetImpl()->NumStates()), s_(0) {}
-
-  bool Done() const { return s_ >= nstates_; }
-
-  StateId Value() const { return s_; }
-
-  void Next() { ++s_; }
-
-  void Reset() { s_ = 0; }
-
- private:
-  const StateId nstates_;
-  StateId s_;
-};
-
-// Specialization for ConstFst; see generic version in fst.h for sample usage
-// (but use the ConstFst type instead). This version should inline.
-template <class Arc, class Unsigned>
-class ArcIterator<ConstFst<Arc, Unsigned>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const ConstFst<Arc, Unsigned> &fst, StateId s)
-      : arcs_(fst.GetImpl()->Arcs(s)),
-        narcs_(fst.GetImpl()->NumArcs(s)),
-        i_(0) {}
-
-  bool Done() const { return i_ >= narcs_; }
-
-  const Arc &Value() const { return arcs_[i_]; }
-
-  void Next() { ++i_; }
-
-  size_t Position() const { return i_; }
-
-  void Reset() { i_ = 0; }
-
-  void Seek(size_t a) { i_ = a; }
-
-  constexpr uint32 Flags() const { return kArcValueFlags; }
-
-  void SetFlags(uint32, uint32) {}
-
- private:
-  const Arc *arcs_;
-  size_t narcs_;
-  size_t i_;
-};
-
-// A useful alias when using StdArc.
-using StdConstFst = ConstFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_CONST_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/determinize.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/determinize.h
deleted file mode 100644
index 545938aa..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/determinize.h
+++ /dev/null
@@ -1,1086 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to determinize an FST.
-
-#ifndef FST_DETERMINIZE_H_
-#define FST_DETERMINIZE_H_
-
-#include <algorithm>
-#include <climits>
-#include <forward_list>
-#include <map>
-#include <string>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arc-map.h>
-#include <fst/bi-table.h>
-#include <fst/cache.h>
-#include <fst/factor-weight.h>
-#include <fst/filter-state.h>
-#include <fst/prune.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-// Common divisors are used in determinization to compute transition weights.
-// In the simplest case, it is the same as semiring Plus, but other choices
-// permit more efficient determinization when the output contains strings.
-
-// The default common divisor uses the semiring Plus.
-template <class W>
-struct DefaultCommonDivisor {
- public:
-  using Weight = W;
-
-  Weight operator()(const Weight &w1, const Weight &w2) const {
-    return Plus(w1, w2);
-  }
-};
-
-// The label common divisor for a (left) string semiring selects a single
-// letter common prefix or the empty string. This is used in the
-// determinization of output strings so that at most a single letter will
-// appear in the output of a transtion.
-template <typename Label, StringType S>
-struct LabelCommonDivisor {
- public:
-  using Weight = StringWeight<Label, S>;
-
-  Weight operator()(const Weight &w1, const Weight &w2) const {
-    typename Weight::Iterator iter1(w1);
-    typename Weight::Iterator iter2(w2);
-    if (!(StringWeight<Label, S>::Properties() & kLeftSemiring)) {
-      FSTERROR() << "LabelCommonDivisor: Weight needs to be left semiring";
-      return Weight::NoWeight();
-    } else if (w1.Size() == 0 || w2.Size() == 0) {
-      return Weight::One();
-    } else if (w1 == Weight::Zero()) {
-      return Weight(iter2.Value());
-    } else if (w2 == Weight::Zero()) {
-      return Weight(iter1.Value());
-    } else if (iter1.Value() == iter2.Value()) {
-      return Weight(iter1.Value());
-    } else {
-      return Weight::One();
-    }
-  }
-};
-
-// The gallic common divisor uses the label common divisor on the string
-// component and the common divisor on the weight component, which defaults to
-// the default common divisor.
-template <class Label, class W, GallicType G,
-          class CommonDivisor = DefaultCommonDivisor<W>>
-class GallicCommonDivisor {
- public:
-  using Weight = GallicWeight<Label, W, G>;
-
-  Weight operator()(const Weight &w1, const Weight &w2) const {
-    return Weight(label_common_divisor_(w1.Value1(), w2.Value1()),
-                  weight_common_divisor_(w1.Value2(), w2.Value2()));
-  }
-
- private:
-  LabelCommonDivisor<Label, GallicStringType(G)> label_common_divisor_;
-  CommonDivisor weight_common_divisor_;
-};
-
-// Specialization for general GALLIC weight.
-template <class Label, class W, class CommonDivisor>
-class GallicCommonDivisor<Label, W, GALLIC, CommonDivisor> {
- public:
-  using Weight = GallicWeight<Label, W, GALLIC>;
-  using GRWeight = GallicWeight<Label, W, GALLIC_RESTRICT>;
-  using Iterator =
-      UnionWeightIterator<GRWeight, GallicUnionWeightOptions<Label, W>>;
-
-  Weight operator()(const Weight &w1, const Weight &w2) const {
-    auto weight = GRWeight::Zero();
-    for (Iterator iter(w1); !iter.Done(); iter.Next()) {
-      weight = common_divisor_(weight, iter.Value());
-    }
-    for (Iterator iter(w2); !iter.Done(); iter.Next()) {
-      weight = common_divisor_(weight, iter.Value());
-    }
-    return weight == GRWeight::Zero() ? Weight::Zero() : Weight(weight);
-  }
-
- private:
-  GallicCommonDivisor<Label, W, GALLIC_RESTRICT, CommonDivisor> common_divisor_;
-};
-
-namespace internal {
-
-// Represents an element in a subset
-template <class Arc>
-struct DeterminizeElement {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  DeterminizeElement(StateId s, Weight weight)
-      : state_id(s), weight(std::move(weight)) {}
-
-  inline bool operator==(const DeterminizeElement<Arc> &element) const {
-    return state_id == element.state_id && weight == element.weight;
-  }
-
-  inline bool operator!=(const DeterminizeElement<Arc> &element) const {
-    return !(*this == element);
-  }
-
-  inline bool operator<(const DeterminizeElement<Arc> &element) const {
-    return state_id < element.state_id;
-  }
-
-  StateId state_id;  // Input state ID.
-  Weight weight;     // Residual weight.
-};
-
-// Represents a weighted subset and determinization filter state
-template <typename A, typename FilterState>
-struct DeterminizeStateTuple {
-  using Arc = A;
-  using Element = DeterminizeElement<Arc>;
-  using Subset = std::forward_list<Element>;
-
-  DeterminizeStateTuple() : filter_state(FilterState::NoState()) {}
-
-  inline bool operator==(
-      const DeterminizeStateTuple<Arc, FilterState> &tuple) const {
-    return (tuple.filter_state == filter_state) && (tuple.subset == subset);
-  }
-
-  inline bool operator!=(
-      const DeterminizeStateTuple<Arc, FilterState> &tuple) const {
-    return (tuple.filter_state != filter_state) || (tuple.subset != subset);
-  }
-
-  Subset subset;
-  FilterState filter_state;
-};
-
-// Proto-transition for determinization.
-template <class StateTuple>
-struct DeterminizeArc {
-  using Arc = typename StateTuple::Arc;
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-
-  DeterminizeArc()
-      : label(kNoLabel), weight(Weight::Zero()), dest_tuple(nullptr) {}
-
-  explicit DeterminizeArc(const Arc &arc)
-      : label(arc.ilabel), weight(Weight::Zero()), dest_tuple(new StateTuple) {}
-
-  Label label;             // Arc label.
-  Weight weight;           // Arc weight.
-  StateTuple *dest_tuple;  // Destination subset and filter state.
-};
-
-}  // namespace internal
-
-// Determinization filters are used to compute destination state tuples based
-// on the source tuple, transition, and destination element or on similar
-// super-final transition information. The filter operates on a map between a
-// label and the corresponding destination state tuples. It must define the map
-// type LabelMap. The default filter is used for weighted determinization.
-// A determinize filter for implementing weighted determinization.
-template <class Arc>
-class DefaultDeterminizeFilter {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FilterState = CharFilterState;
-  using Element = internal::DeterminizeElement<Arc>;
-  using StateTuple = internal::DeterminizeStateTuple<Arc, FilterState>;
-  using LabelMap = std::map<Label, internal::DeterminizeArc<StateTuple>>;
-
-  // This is needed e.g. to go into the gallic domain for transducers.
-  template <class A>
-  struct rebind {
-    using Other = DefaultDeterminizeFilter<A>;
-  };
-
-  explicit DefaultDeterminizeFilter(const Fst<Arc> &fst) : fst_(fst.Copy()) {}
-
-  // This is needed (e.g.) to go into the gallic domain for transducers.
-  // Ownership of the templated filter argument is given to this class.
-  template <class Filter>
-  DefaultDeterminizeFilter(const Fst<Arc> &fst, Filter *filter)
-      : fst_(fst.Copy()) {
-    delete filter;
-  }
-
-  // Copy constructor; the FST can be passed if it has been deep-copied.
-  DefaultDeterminizeFilter(const DefaultDeterminizeFilter<Arc> &filter,
-                           const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? fst->Copy() : filter.fst_->Copy()) {}
-
-  FilterState Start() const { return FilterState(0); }
-
-  // Does no work.
-  void SetState(StateId s, const StateTuple &tuple) {}
-
-  // Filters transition, possibly modifying label map. Returns true if arc is
-  // added to the label map.
-  bool FilterArc(const Arc &arc, const Element &src_element,
-                 const Element &dest_element, LabelMap *label_map) const {
-    // Adds element to unique state tuple for arc label.
-    auto &det_arc = (*label_map)[arc.ilabel];
-    if (det_arc.label == kNoLabel) {
-      det_arc = internal::DeterminizeArc<StateTuple>(arc);
-      det_arc.dest_tuple->filter_state = FilterState(0);
-    }
-    det_arc.dest_tuple->subset.push_front(dest_element);
-    return true;
-  }
-
-  // Filters super-final transition, returning new final weight.
-  Weight FilterFinal(Weight weight, const Element &element) { return weight; }
-
-  static uint64 Properties(uint64 props) { return props; }
-
- private:
-  std::unique_ptr<Fst<Arc>> fst_;
-};
-
-// Determinization state table interface:
-//
-// template <class Arc, class FilterState>
-// class DeterminizeStateTable {
-//  public:
-//   using StateId = typename Arc::StateId;
-//   using StateTuple = internal::DeterminizeStateTuple<Arc, FilterState>;
-//
-//   // Required sub-class. This is needed (e.g.) to go into the gallic domain.
-//   template <class B, class G>
-//   struct rebind {
-//     using Other = DeterminizeStateTable<B, G>;
-//   }
-//
-//   // Required constuctor.
-//   DeterminizeStateTable();
-//
-//   // Required copy constructor that does not copy state.
-//   DeterminizeStateTable(const DeterminizeStateTable<Arc, FilterState>
-//   &table);
-//
-//   // Looks up state ID by state tuple; if it doesn't exist, then adds it.
-//   // FindState takes ownership of the state tuple argument so that it
-//   // doesn't have to copy it if it creates a new state.
-//   StateId FindState(StateTuple *tuple);
-//
-//   // Looks up state tuple by ID.
-//   const StateTuple *Tuple(StateId id) const;
-// };
-
-// The default determinization state table based on the compact hash bi-table.
-template <class Arc, class FilterState>
-class DefaultDeterminizeStateTable {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using StateTuple = internal::DeterminizeStateTuple<Arc, FilterState>;
-  using Element = typename StateTuple::Element;
-  using Subset = typename StateTuple::Subset;
-
-  template <class B, class G>
-  struct rebind {
-    using Other = DefaultDeterminizeStateTable<B, G>;
-  };
-
-  explicit DefaultDeterminizeStateTable(size_t table_size = 0)
-      : table_size_(table_size), tuples_(table_size_) {}
-
-  DefaultDeterminizeStateTable(
-      const DefaultDeterminizeStateTable<Arc, FilterState> &table)
-      : table_size_(table.table_size_), tuples_(table_size_) {}
-
-  ~DefaultDeterminizeStateTable() {
-    for (StateId s = 0; s < tuples_.Size(); ++s) delete tuples_.FindEntry(s);
-  }
-
-  // Finds the state corresponding to a state tuple. Only creates a new state if
-  // the tuple is not found. FindState takes ownership of the tuple argument so
-  // that it doesn't have to copy it if it creates a new state.
-  StateId FindState(StateTuple *tuple) {
-    const StateId ns = tuples_.Size();
-    const auto s = tuples_.FindId(tuple);
-    if (s != ns) delete tuple;  // Tuple found.
-    return s;
-  }
-
-  const StateTuple *Tuple(StateId s) { return tuples_.FindEntry(s); }
-
- private:
-  // Comparison object for StateTuples.
-  class StateTupleEqual {
-   public:
-    bool operator()(const StateTuple *tuple1, const StateTuple *tuple2) const {
-      return *tuple1 == *tuple2;
-    }
-  };
-
-  // Hash function for StateTuples.
-  class StateTupleKey {
-   public:
-    size_t operator()(const StateTuple *tuple) const {
-      size_t h = tuple->filter_state.Hash();
-      for (auto it = tuple->subset.begin(); it != tuple->subset.end(); ++it) {
-        const size_t h1 = it->state_id;
-        static constexpr auto lshift = 5;
-        static constexpr auto rshift = CHAR_BIT * sizeof(size_t) - 5;
-        h ^= h << 1 ^ h1 << lshift ^ h1 >> rshift ^ it->weight.Hash();
-      }
-      return h;
-    }
-  };
-
-  size_t table_size_;
-  CompactHashBiTable<StateId, StateTuple *, StateTupleKey, StateTupleEqual,
-                     HS_STL>
-      tuples_;
-
-  DefaultDeterminizeStateTable &operator=(
-      const DefaultDeterminizeStateTable &) = delete;
-};
-
-// Determinization type.
-enum DeterminizeType {
-  // Input transducer is known to be functional (or error).
-  DETERMINIZE_FUNCTIONAL,  // Input transducer is functional (error if not).
-  // Input transducer is not known to be functional.
-  DETERMINIZE_NONFUNCTIONAL,
-  // Input transducer is not known to be functional but only keep the min of
-  // of ambiguous outputs.
-  DETERMINIZE_DISAMBIGUATE
-};
-
-// Options for finite-state transducer determinization templated on the arc
-// type, common divisor, the determinization filter and the state table.
-// DeterminizeFst takes ownership of the determinization filter and state table,
-// if provided.
-template <class Arc,
-          class CommonDivisor = DefaultCommonDivisor<typename Arc::Weight>,
-          class Filter = DefaultDeterminizeFilter<Arc>,
-          class StateTable =
-              DefaultDeterminizeStateTable<Arc, typename Filter::FilterState>>
-struct DeterminizeFstOptions : public CacheOptions {
-  using Label = typename Arc::Label;
-
-  float delta;                // Quantization delta for subset weights.
-  Label subsequential_label;  // Label used for residual final output
-                              // when producing subsequential transducers.
-  DeterminizeType type;       // Determinization type.
-  bool increment_subsequential_label;  // When creating several subsequential
-                                       // arcs at a given state, make their
-                                       // label distinct by incrementing.
-  Filter *filter;                      // Determinization filter;
-                                       // DeterminizeFst takes ownership.
-  StateTable *state_table;             // Determinization state table;
-                                       // DeterminizeFst takes ownership.
-
-  explicit DeterminizeFstOptions(const CacheOptions &opts, float delta = kDelta,
-                                 Label subsequential_label = 0,
-                                 DeterminizeType type = DETERMINIZE_FUNCTIONAL,
-                                 bool increment_subsequential_label = false,
-                                 Filter *filter = nullptr,
-                                 StateTable *state_table = nullptr)
-      : CacheOptions(opts),
-        delta(delta),
-        subsequential_label(subsequential_label),
-        type(type),
-        increment_subsequential_label(increment_subsequential_label),
-        filter(filter),
-        state_table(state_table) {}
-
-  explicit DeterminizeFstOptions(float delta = kDelta,
-                                 Label subsequential_label = 0,
-                                 DeterminizeType type = DETERMINIZE_FUNCTIONAL,
-                                 bool increment_subsequential_label = false,
-                                 Filter *filter = nullptr,
-                                 StateTable *state_table = nullptr)
-      : delta(delta),
-        subsequential_label(subsequential_label),
-        type(type),
-        increment_subsequential_label(increment_subsequential_label),
-        filter(filter),
-        state_table(state_table) {}
-};
-
-namespace internal {
-
-// Implementation of delayed DeterminizeFst. This base class is
-// common to the variants that implement acceptor and transducer
-// determinization.
-template <class Arc>
-class DeterminizeFstImplBase : public CacheImpl<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::Properties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheBaseImpl<CacheState<Arc>>::HasStart;
-  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
-  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
-  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
-  using CacheBaseImpl<CacheState<Arc>>::SetStart;
-
-  template <class CommonDivisor, class Filter, class StateTable>
-  DeterminizeFstImplBase(
-      const Fst<Arc> &fst,
-      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable> &opts)
-      : CacheImpl<Arc>(opts), fst_(fst.Copy()) {
-    SetType("determinize");
-    const auto iprops = fst.Properties(kFstProperties, false);
-    const auto dprops =
-        DeterminizeProperties(iprops, opts.subsequential_label != 0,
-                              opts.type == DETERMINIZE_NONFUNCTIONAL
-                                  ? opts.increment_subsequential_label
-                                  : true);
-    SetProperties(Filter::Properties(dprops), kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  DeterminizeFstImplBase(const DeterminizeFstImplBase<Arc> &impl)
-      : CacheImpl<Arc>(impl), fst_(impl.fst_->Copy(true)) {
-    SetType("determinize");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  virtual DeterminizeFstImplBase<Arc> *Copy() const = 0;
-
-  StateId Start() {
-    if (!HasStart()) {
-      const auto start = ComputeStart();
-      if (start != kNoStateId) SetStart(start);
-    }
-    return CacheImpl<Arc>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) SetFinal(s, ComputeFinal(s));
-    return CacheImpl<Arc>::Final(s);
-  }
-
-  virtual void Expand(StateId s) = 0;
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumOutputEpsilons(s);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<Arc>::InitArcIterator(s, data);
-  }
-
-  virtual StateId ComputeStart() = 0;
-
-  virtual Weight ComputeFinal(StateId s) = 0;
-
-  const Fst<Arc> &GetFst() const { return *fst_; }
-
- private:
-  std::unique_ptr<const Fst<Arc>> fst_;  // Input FST.
-};
-
-// Implementation of delayed determinization for weighted acceptors.
-template <class Arc, class CommonDivisor, class Filter, class StateTable>
-class DeterminizeFsaImpl : public DeterminizeFstImplBase<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FilterState = typename Filter::FilterState;
-  using StateTuple = internal::DeterminizeStateTuple<Arc, FilterState>;
-  using Element = typename StateTuple::Element;
-  using Subset = typename StateTuple::Subset;
-  using LabelMap = typename Filter::LabelMap;
-
-  using FstImpl<Arc>::SetProperties;
-  using DeterminizeFstImplBase<Arc>::GetFst;
-  using DeterminizeFstImplBase<Arc>::SetArcs;
-
-  DeterminizeFsaImpl(
-      const Fst<Arc> &fst, const std::vector<Weight> *in_dist,
-      std::vector<Weight> *out_dist,
-      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable> &opts)
-      : DeterminizeFstImplBase<Arc>(fst, opts),
-        delta_(opts.delta),
-        in_dist_(in_dist),
-        out_dist_(out_dist),
-        filter_(opts.filter ? opts.filter : new Filter(fst)),
-        state_table_(opts.state_table ? opts.state_table : new StateTable()) {
-    if (!fst.Properties(kAcceptor, true)) {
-      FSTERROR() << "DeterminizeFst: Argument not an acceptor";
-      SetProperties(kError, kError);
-    }
-    if (!(Weight::Properties() & kLeftSemiring)) {
-      FSTERROR() << "DeterminizeFst: Weight must be left distributive: "
-                 << Weight::Type();
-      SetProperties(kError, kError);
-    }
-    if (out_dist_) out_dist_->clear();
-  }
-
-  DeterminizeFsaImpl(
-      const DeterminizeFsaImpl<Arc, CommonDivisor, Filter, StateTable> &impl)
-      : DeterminizeFstImplBase<Arc>(impl),
-        delta_(impl.delta_),
-        in_dist_(nullptr),
-        out_dist_(nullptr),
-        filter_(new Filter(*impl.filter_, &GetFst())),
-        state_table_(new StateTable(*impl.state_table_)) {
-    if (impl.out_dist_) {
-      FSTERROR() << "DeterminizeFsaImpl: Cannot copy with out_dist vector";
-      SetProperties(kError, kError);
-    }
-  }
-
-  DeterminizeFsaImpl<Arc, CommonDivisor, Filter, StateTable> *Copy()
-      const override {
-    return new DeterminizeFsaImpl<Arc, CommonDivisor, Filter, StateTable>(
-        *this);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && (GetFst().Properties(kError, false))) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  StateId ComputeStart() override {
-    const auto s = GetFst().Start();
-    if (s == kNoStateId) return kNoStateId;
-    const Element element(s, Weight::One());
-    auto *tuple = new StateTuple;
-    tuple->subset.push_front(element);
-    tuple->filter_state = filter_->Start();
-    return FindState(tuple);
-  }
-
-  Weight ComputeFinal(StateId s) override {
-    const auto *tuple = state_table_->Tuple(s);
-    filter_->SetState(s, *tuple);
-    auto final_weight = Weight::Zero();
-    for (auto it = tuple->subset.begin(); it != tuple->subset.end(); ++it) {
-      const auto &element = *it;
-      final_weight =
-          Plus(final_weight,
-               Times(element.weight, GetFst().Final(element.state_id)));
-      final_weight = filter_->FilterFinal(final_weight, element);
-      if (!final_weight.Member()) SetProperties(kError, kError);
-    }
-    return final_weight;
-  }
-
-  StateId FindState(StateTuple *tuple) {
-    const auto s = state_table_->FindState(tuple);
-    if (in_dist_ && out_dist_->size() <= s) {
-      out_dist_->push_back(ComputeDistance(tuple->subset));
-    }
-    return s;
-  }
-
-  // Computes distance from a state to the final states in the DFA given the
-  // distances in the NFA.
-  Weight ComputeDistance(const Subset &subset) {
-    auto outd = Weight::Zero();
-    for (auto it = subset.begin(); it != subset.end(); ++it) {
-      const auto &element = *it;
-      const auto ind =
-          (element.state_id < in_dist_->size() ? (*in_dist_)[element.state_id]
-                                               : Weight::Zero());
-      outd = Plus(outd, Times(element.weight, ind));
-    }
-    return outd;
-  }
-
-  // Computes the outgoing transitions from a state, creating new destination
-  // states as needed.
-  void Expand(StateId s) override {
-    LabelMap label_map;
-    GetLabelMap(s, &label_map);
-    for (auto it = label_map.begin(); it != label_map.end(); ++it) {
-      AddArc(s, it->second);
-    }
-    SetArcs(s);
-  }
-
- private:
-  using DetArc = internal::DeterminizeArc<StateTuple>;
-
-  // Constructs proto-determinization transition, including destination subset,
-  // per label.
-  void GetLabelMap(StateId s, LabelMap *label_map) {
-    const auto *src_tuple = state_table_->Tuple(s);
-    filter_->SetState(s, *src_tuple);
-    for (auto it = src_tuple->subset.begin(); it != src_tuple->subset.end();
-         ++it) {
-      const auto &src_element = *it;
-      for (ArcIterator<Fst<Arc>> aiter(GetFst(), src_element.state_id);
-           !aiter.Done(); aiter.Next()) {
-        const auto &arc = aiter.Value();
-        const Element dest_element(arc.nextstate,
-                                   Times(src_element.weight, arc.weight));
-        filter_->FilterArc(arc, src_element, dest_element, label_map);
-      }
-    }
-    for (auto it = label_map->begin(); it != label_map->end(); ++it) {
-      NormArc(&it->second);
-    }
-  }
-
-  // Sorts subsets and removes duplicate elements, normalizing transition and
-  // subset weights.
-  void NormArc(DetArc *det_arc) {
-    auto *dest_tuple = det_arc->dest_tuple;
-    dest_tuple->subset.sort();
-    auto piter = dest_tuple->subset.begin();
-    for (auto diter = dest_tuple->subset.begin();
-         diter != dest_tuple->subset.end();) {
-      auto &dest_element = *diter;
-      auto &prev_element = *piter;
-      // Computes arc weight.
-      det_arc->weight = common_divisor_(det_arc->weight, dest_element.weight);
-      if (piter != diter && dest_element.state_id == prev_element.state_id) {
-        // Found duplicate state: sums state weight and deletes duplicate.
-        prev_element.weight = Plus(prev_element.weight, dest_element.weight);
-        if (!prev_element.weight.Member()) SetProperties(kError, kError);
-        ++diter;
-        dest_tuple->subset.erase_after(piter);
-      } else {
-        piter = diter;
-        ++diter;
-      }
-    }
-    // Divides out label weight from destination subset elements, quantizing to
-    // ensure comparisons are effective.
-    for (auto diter = dest_tuple->subset.begin();
-         diter != dest_tuple->subset.end(); ++diter) {
-      auto &dest_element = *diter;
-      dest_element.weight =
-          Divide(dest_element.weight, det_arc->weight, DIVIDE_LEFT);
-      dest_element.weight = dest_element.weight.Quantize(delta_);
-    }
-  }
-
-  // Adds an arc from state S to the destination state associated with state
-  // tuple in det_arc as created by GetLabelMap.
-  void AddArc(StateId s, const DetArc &det_arc) {
-    const Arc arc(det_arc.label, det_arc.label, det_arc.weight,
-                  FindState(det_arc.dest_tuple));
-    CacheImpl<Arc>::PushArc(s, arc);
-  }
-
-  float delta_;                         // Quantization delta for weights.
-  const std::vector<Weight> *in_dist_;  // Distance to final NFA states.
-  std::vector<Weight> *out_dist_;       // Distance to final DFA states.
-
-  // FIXME(kbg): Ought to be static const?
-  CommonDivisor common_divisor_;
-  std::unique_ptr<Filter> filter_;
-  std::unique_ptr<StateTable> state_table_;
-};
-
-// Implementation of delayed determinization for transducers. Transducer
-// determinization is implemented by mapping the input to the Gallic semiring as
-// an acceptor whose weights contain the output strings and using acceptor
-// determinization above to determinize that acceptor.
-template <class Arc, GallicType G, class CommonDivisor, class Filter,
-          class StateTable>
-class DeterminizeFstImpl : public DeterminizeFstImplBase<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using ToMapper = ToGallicMapper<Arc, G>;
-  using ToArc = typename ToMapper::ToArc;
-  using ToFst = ArcMapFst<Arc, ToArc, ToMapper>;
-  using FromMapper = FromGallicMapper<Arc, G>;
-  using FromFst = ArcMapFst<ToArc, Arc, FromMapper>;
-
-  using ToCommonDivisor = GallicCommonDivisor<Label, Weight, G, CommonDivisor>;
-  using ToFilter = typename Filter::template rebind<ToArc>::Other;
-  using ToFilterState = typename ToFilter::FilterState;
-  using ToStateTable =
-      typename StateTable::template rebind<ToArc, ToFilterState>::Other;
-  using FactorIterator = GallicFactor<Label, Weight, G>;
-
-  using FstImpl<Arc>::SetProperties;
-  using DeterminizeFstImplBase<Arc>::GetFst;
-  using CacheBaseImpl<CacheState<Arc>>::GetCacheGc;
-  using CacheBaseImpl<CacheState<Arc>>::GetCacheLimit;
-
-  DeterminizeFstImpl(
-      const Fst<Arc> &fst,
-      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable> &opts)
-      : DeterminizeFstImplBase<Arc>(fst, opts),
-        delta_(opts.delta),
-        subsequential_label_(opts.subsequential_label),
-        increment_subsequential_label_(opts.increment_subsequential_label) {
-    if (opts.state_table) {
-      FSTERROR() << "DeterminizeFst: "
-                 << "A state table can not be passed with transducer input";
-      SetProperties(kError, kError);
-      return;
-    }
-    Init(GetFst(), opts.filter);
-  }
-
-  DeterminizeFstImpl(
-      const DeterminizeFstImpl<Arc, G, CommonDivisor, Filter, StateTable> &impl)
-      : DeterminizeFstImplBase<Arc>(impl),
-        delta_(impl.delta_),
-        subsequential_label_(impl.subsequential_label_),
-        increment_subsequential_label_(impl.increment_subsequential_label_) {
-    Init(GetFst(), nullptr);
-  }
-
-  DeterminizeFstImpl<Arc, G, CommonDivisor, Filter, StateTable> *Copy()
-      const override {
-    return new DeterminizeFstImpl<Arc, G, CommonDivisor, Filter, StateTable>(
-        *this);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && (GetFst().Properties(kError, false) ||
-                            from_fst_->Properties(kError, false))) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  StateId ComputeStart() override { return from_fst_->Start(); }
-
-  Weight ComputeFinal(StateId s) override { return from_fst_->Final(s); }
-
-  void Expand(StateId s) override {
-    for (ArcIterator<FromFst> aiter(*from_fst_, s); !aiter.Done();
-         aiter.Next()) {
-      CacheImpl<Arc>::PushArc(s, aiter.Value());
-    }
-    CacheImpl<Arc>::SetArcs(s);
-  }
-
- private:
-  // Initialization of transducer determinization implementation, which is
-  // defined after DeterminizeFst since it calls it.
-  void Init(const Fst<Arc> &fst, Filter *filter);
-
-  float delta_;
-  Label subsequential_label_;
-  bool increment_subsequential_label_;
-  std::unique_ptr<FromFst> from_fst_;
-};
-
-}  // namespace internal
-
-// Determinizes a weighted transducer. This version is a delayed
-// FST. The result will be an equivalent FST that has the property
-// that no state has two transitions with the same input label.
-// For this algorithm, epsilon transitions are treated as regular
-// symbols (cf. RmEpsilon).
-//
-// The transducer must be functional. The weights must be (weakly) left
-// divisible (valid for TropicalWeight and LogWeight for instance) and be
-// zero-sum-free if for all a, b: (Plus(a, b) == 0) => a = b = 0.
-//
-// Complexity:
-//
-//   Determinizable: exponential (polynomial in the size of the output).
-//   Non-determinizable: does not terminate.
-//
-// The determinizable automata include all unweighted and all acyclic input.
-//
-// For more information, see:
-//
-// Mohri, M. 1997. Finite-state transducers in language and speech processing.
-// Computational Linguistics 23(2): 269-311.
-//
-// This class attaches interface to implementation and handles reference
-// counting, delegating most methods to ImplToFst.
-template <class A>
-class DeterminizeFst : public ImplToFst<internal::DeterminizeFstImplBase<A>> {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::DeterminizeFstImplBase<Arc>;
-
-  friend class ArcIterator<DeterminizeFst<Arc>>;
-  friend class StateIterator<DeterminizeFst<Arc>>;
-
-  template <class B, GallicType G, class CommonDivisor, class Filter,
-            class StateTable>
-  friend class DeterminizeFstImpl;
-
-  explicit DeterminizeFst(const Fst<A> &fst)
-      : ImplToFst<Impl>(CreateImpl(fst)) {}
-
-  template <class CommonDivisor, class Filter, class StateTable>
-  DeterminizeFst(
-      const Fst<Arc> &fst,
-      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable> &opts)
-      : ImplToFst<Impl>(CreateImpl(fst, opts)) {}
-
-  // This acceptor-only version additionally computes the distance to
-  // final states in the output if provided with those distances for the
-  // input. Useful for e.g., computing the k-shortest unique paths.
-  template <class D, class F, class T>
-  DeterminizeFst(const Fst<A> &fst, const std::vector<Weight> *in_dist,
-                 std::vector<Weight> *out_dist,
-                 const DeterminizeFstOptions<A, D, F, T> &opts)
-      : ImplToFst<Impl>(
-            std::make_shared<internal::DeterminizeFsaImpl<A, D, F, T>>(
-                fst, in_dist, out_dist, opts)) {
-    if (!fst.Properties(kAcceptor, true)) {
-      FSTERROR() << "DeterminizeFst: "
-                 << "Distance to final states computed for acceptors only";
-      GetMutableImpl()->SetProperties(kError, kError);
-    }
-  }
-
-  // See Fst<>::Copy() for doc.
-  DeterminizeFst(const DeterminizeFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(safe ? std::shared_ptr<Impl>(fst.GetImpl()->Copy())
-                             : fst.GetSharedImpl()) {}
-
-  // Get a copy of this DeterminizeFst. See Fst<>::Copy() for further doc.
-  DeterminizeFst<Arc> *Copy(bool safe = false) const override {
-    return new DeterminizeFst<Arc>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  static std::shared_ptr<Impl> CreateImpl(const Fst<Arc> &fst) {
-    using D = DefaultCommonDivisor<Weight>;
-    using F = DefaultDeterminizeFilter<Arc>;
-    using T = DefaultDeterminizeStateTable<Arc, typename F::FilterState>;
-    const DeterminizeFstOptions<Arc, D, F, T> opts;
-    return CreateImpl(fst, opts);
-  }
-
-  template <class CommonDivisor, class Filter, class StateTable>
-  static std::shared_ptr<Impl> CreateImpl(
-      const Fst<Arc> &fst,
-      const DeterminizeFstOptions<Arc, CommonDivisor, Filter, StateTable>
-          &opts) {
-    if (fst.Properties(kAcceptor, true)) {
-      // Calls implementation for acceptors.
-      return std::make_shared<
-          internal::DeterminizeFsaImpl<Arc, CommonDivisor, Filter, StateTable>>(
-          fst, nullptr, nullptr, opts);
-    } else if (opts.type == DETERMINIZE_DISAMBIGUATE) {
-      auto rv = std::make_shared<internal::DeterminizeFstImpl<
-          Arc, GALLIC_MIN, CommonDivisor, Filter, StateTable>>(fst, opts);
-      if (!(Weight::Properties() & kPath)) {
-        FSTERROR() << "DeterminizeFst: Weight needs to have the "
-                   << "path property to disambiguate output: "
-                   << Weight::Type();
-        rv->SetProperties(kError, kError);
-      }
-      // Calls disambiguating implementation for non-functional transducers.
-      return rv;
-    } else if (opts.type == DETERMINIZE_FUNCTIONAL) {
-      // Calls implementation for functional transducers.
-      return std::make_shared<internal::DeterminizeFstImpl<
-          Arc, GALLIC_RESTRICT, CommonDivisor, Filter, StateTable>>(fst, opts);
-    } else {  // opts.type == DETERMINIZE_NONFUNCTIONAL
-      // Calls implementation for non functional transducers;
-      return std::make_shared<internal::DeterminizeFstImpl<
-          Arc, GALLIC, CommonDivisor, Filter, StateTable>>(fst, opts);
-    }
-  }
-
-  DeterminizeFst &operator=(const DeterminizeFst &) = delete;
-};
-
-namespace internal {
-
-// Initialization of transducer determinization implementation, which is defined
-// after DeterminizeFst since it calls it.
-template <class Arc, GallicType G, class D, class Filter, class T>
-void DeterminizeFstImpl<Arc, G, D, Filter, T>::Init(const Fst<Arc> &fst, Filter *filter) {
-  // Mapper to an acceptor.
-  ToFst to_fst(fst, ToMapper());
-  auto *to_filter = filter ? new ToFilter(to_fst, filter) : nullptr;
-  // This recursive call terminates since it is to a (non-recursive)
-  // different constructor.
-  CacheOptions copts(GetCacheGc(), GetCacheLimit());
-  DeterminizeFstOptions<ToArc, ToCommonDivisor, ToFilter, ToStateTable> dopts(
-      copts, delta_, 0, DETERMINIZE_FUNCTIONAL, false, to_filter);
-  // Uses acceptor-only constructor to avoid template recursion.
-  DeterminizeFst<ToArc> det_fsa(to_fst, nullptr, nullptr, dopts);
-  // Mapper back to transducer.
-  FactorWeightOptions<ToArc> fopts(
-      CacheOptions(true, 0), delta_, kFactorFinalWeights, subsequential_label_,
-      subsequential_label_, increment_subsequential_label_,
-      increment_subsequential_label_);
-  FactorWeightFst<ToArc, FactorIterator> factored_fst(det_fsa, fopts);
-  from_fst_.reset(new FromFst(factored_fst, FromMapper(subsequential_label_)));
-}
-
-}  // namespace internal
-
-// Specialization for DeterminizeFst.
-template <class Arc>
-class StateIterator<DeterminizeFst<Arc>>
-    : public CacheStateIterator<DeterminizeFst<Arc>> {
- public:
-  explicit StateIterator(const DeterminizeFst<Arc> &fst)
-      : CacheStateIterator<DeterminizeFst<Arc>>(fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for DeterminizeFst.
-template <class Arc>
-class ArcIterator<DeterminizeFst<Arc>>
-    : public CacheArcIterator<DeterminizeFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const DeterminizeFst<Arc> &fst, StateId s)
-      : CacheArcIterator<DeterminizeFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc>
-inline void DeterminizeFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<DeterminizeFst<Arc>>(*this);
-}
-
-// Useful aliases when using StdArc.
-using StdDeterminizeFst = DeterminizeFst<StdArc>;
-
-template <class Arc>
-struct DeterminizeOptions {
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  float delta;                // Quantization delta for subset weights.
-  Weight weight_threshold;    // Pruning weight threshold.
-  StateId state_threshold;    // Pruning state threshold.
-  Label subsequential_label;  // Label used for residual final output.
-  DeterminizeType type;
-  bool increment_subsequential_label;  // When creating several subsequential
-                                       // arcs at a given state, make their
-                                       // label distinct by incrementation?
-
-  explicit DeterminizeOptions(float delta = kDelta,
-                              Weight weight_threshold = Weight::Zero(),
-                              StateId state_threshold = kNoStateId,
-                              Label subsequential_label = 0,
-                              DeterminizeType type = DETERMINIZE_FUNCTIONAL,
-                              bool increment_subsequential_label = false)
-      : delta(delta),
-        weight_threshold(std::move(weight_threshold)),
-        state_threshold(state_threshold),
-        subsequential_label(subsequential_label),
-        type(type),
-        increment_subsequential_label(increment_subsequential_label) {}
-};
-
-// Determinizes a weighted transducer. This version writes the
-// determinized Fst to an output MutableFst. The result will be an
-// equivalent FST that has the property that no state has two
-// transitions with the same input label.  For this algorithm, epsilon
-// transitions are treated as regular symbols (cf. RmEpsilon).
-//
-// The transducer must be functional. The weights must be (weakly)
-// left divisible (valid for TropicalWeight and LogWeight).
-//
-// Complexity:
-//
-//   Determinizable: exponential (polynomial in the size of the output)
-//   Non-determinizable: does not terminate
-//
-// The determinizable automata include all unweighted and all acyclic input.
-template <class Arc>
-void Determinize(
-    const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-    const DeterminizeOptions<Arc> &opts = DeterminizeOptions<Arc>()) {
-  using Weight = typename Arc::Weight;
-  DeterminizeFstOptions<Arc> nopts;
-  nopts.delta = opts.delta;
-  nopts.subsequential_label = opts.subsequential_label;
-  nopts.type = opts.type;
-  nopts.increment_subsequential_label = opts.increment_subsequential_label;
-  nopts.gc_limit = 0;  // Caches only the last state for fastest copy.
-  if (opts.weight_threshold != Weight::Zero() ||
-      opts.state_threshold != kNoStateId) {
-    if (ifst.Properties(kAcceptor, false)) {
-      std::vector<Weight> idistance;
-      std::vector<Weight> odistance;
-      ShortestDistance(ifst, &idistance, true);
-      DeterminizeFst<Arc> dfst(ifst, &idistance, &odistance, nopts);
-      PruneOptions<Arc, AnyArcFilter<Arc>> popts(
-          opts.weight_threshold, opts.state_threshold, AnyArcFilter<Arc>(),
-          &odistance);
-      Prune(dfst, ofst, popts);
-    } else {
-      *ofst = DeterminizeFst<Arc>(ifst, nopts);
-      Prune(ofst, opts.weight_threshold, opts.state_threshold);
-    }
-  } else {
-    *ofst = DeterminizeFst<Arc>(ifst, nopts);
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_DETERMINIZE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/dfs-visit.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/dfs-visit.h
deleted file mode 100644
index d138ab62..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/dfs-visit.h
+++ /dev/null
@@ -1,204 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Depth-first search visitation. See visit.h for more general search queue
-// disciplines.
-
-#ifndef FST_DFS_VISIT_H_
-#define FST_DFS_VISIT_H_
-
-#include <stack>
-#include <vector>
-
-#include <fst/arcfilter.h>
-#include <fst/fst.h>
-
-
-namespace fst {
-
-// Visitor Interface: class determining actions taken during a depth-first
-// search-style visit. If any of the boolean member functions return false, the
-// DFS is aborted by first calling FinishState() on all currently grey states
-// and then calling FinishVisit().
-//
-// This is similar to the more general visitor interface in visit.h, except
-// that FinishState returns additional information appropriate only for a DFS
-// and some methods names here are better suited to a DFS.
-//
-// template <class Arc>
-// class Visitor {
-//  public:
-//   using StateId = typename Arc::StateId;
-//
-//   Visitor(T *return_data);
-//
-//   // Invoked before DFS visit.
-//   void InitVisit(const Fst<Arc> &fst);
-//
-//   // Invoked when state discovered (2nd arg is DFS tree root).
-//   bool InitState(StateId s, StateId root);
-//
-//   // Invoked when tree arc to white/undiscovered state examined.
-//   bool TreeArc(StateId s, const Arc &arc);
-//
-//   // Invoked when back arc to grey/unfinished state examined.
-//   bool BackArc(StateId s, const Arc &arc);
-//
-//   // Invoked when forward or cross arc to black/finished state examined.
-//   bool ForwardOrCrossArc(StateId s, const Arc &arc);
-//
-//   // Invoked when state finished ('s' is tree root, 'parent' is kNoStateId,
-//   // and 'arc' is nullptr).
-//   void FinishState(StateId s, StateId parent, const Arc *arc);
-//
-//   // Invoked after DFS visit.
-//   void FinishVisit();
-// };
-
-namespace internal {
-
-// An FST state's DFS stack state.
-template <class FST>
-struct DfsState {
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-
-  DfsState(const FST &fst, StateId s) : state_id(s), arc_iter(fst, s) {}
-
-  void *operator new(size_t size, MemoryPool<DfsState<FST>> *pool) {
-    return pool->Allocate();
-  }
-
-  static void Destroy(DfsState<FST> *dfs_state,
-                      MemoryPool<DfsState<FST>> *pool) {
-    if (dfs_state) {
-      dfs_state->~DfsState<FST>();
-      pool->Free(dfs_state);
-    }
-  }
-
-  StateId state_id;           // FST state.
-  ArcIterator<FST> arc_iter;  // The corresponding arcs.
-};
-
-}  // namespace internal
-
-// Performs depth-first visitation. Visitor class argument determines actions
-// and contains any return data. ArcFilter determines arcs that are considered.
-// If 'access_only' is true, performs visitation only to states accessible from
-// the initial state.
-//
-// Note this is similar to Visit() in visit.h called with a LIFO queue, except
-// this version has a Visitor class specialized and augmented for a DFS.
-template <class FST, class Visitor, class ArcFilter>
-void DfsVisit(const FST &fst, Visitor *visitor, ArcFilter filter,
-              bool access_only = false) {
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-  visitor->InitVisit(fst);
-  const auto start = fst.Start();
-  if (start == kNoStateId) {
-    visitor->FinishVisit();
-    return;
-  }
-  // An FST state's DFS status
-  static constexpr uint8 kDfsWhite = 0;  // Undiscovered.
-  static constexpr uint8 kDfsGrey = 1;   // Discovered but unfinished.
-  static constexpr uint8 kDfsBlack = 2;  // Finished.
-  std::vector<uint8> state_color;
-  std::stack<internal::DfsState<FST> *> state_stack;  // DFS execution stack.
-  MemoryPool<internal::DfsState<FST>> state_pool;     // Pool for DFSStates.
-  auto nstates = start + 1;  // Number of known states in general case.
-  bool expanded = false;
-  if (fst.Properties(kExpanded, false)) {  // Tests if expanded case, then
-    nstates = CountStates(fst);            // uses ExpandedFst::NumStates().
-    expanded = true;
-  }
-  state_color.resize(nstates, kDfsWhite);
-  StateIterator<FST> siter(fst);
-  // Continue DFS while true.
-  bool dfs = true;
-  // Iterate over trees in DFS forest.
-  for (auto root = start; dfs && root < nstates;) {
-    state_color[root] = kDfsGrey;
-    state_stack.push(new (&state_pool) internal::DfsState<FST>(fst, root));
-    dfs = visitor->InitState(root, root);
-    while (!state_stack.empty()) {
-      auto *dfs_state = state_stack.top();
-      const auto s = dfs_state->state_id;
-      if (s >= state_color.size()) {
-        nstates = s + 1;
-        state_color.resize(nstates, kDfsWhite);
-      }
-      ArcIterator<FST> &aiter = dfs_state->arc_iter;
-      if (!dfs || aiter.Done()) {
-        state_color[s] = kDfsBlack;
-        internal::DfsState<FST>::Destroy(dfs_state, &state_pool);
-        state_stack.pop();
-        if (!state_stack.empty()) {
-          auto *parent_state = state_stack.top();
-          auto &piter = parent_state->arc_iter;
-          visitor->FinishState(s, parent_state->state_id, &piter.Value());
-          piter.Next();
-        } else {
-          visitor->FinishState(s, kNoStateId, nullptr);
-        }
-        continue;
-      }
-      const auto &arc = aiter.Value();
-      if (arc.nextstate >= state_color.size()) {
-        nstates = arc.nextstate + 1;
-        state_color.resize(nstates, kDfsWhite);
-      }
-      if (!filter(arc)) {
-        aiter.Next();
-        continue;
-      }
-      const auto next_color = state_color[arc.nextstate];
-      switch (next_color) {
-        default:
-        case kDfsWhite:
-          dfs = visitor->TreeArc(s, arc);
-          if (!dfs) break;
-          state_color[arc.nextstate] = kDfsGrey;
-          state_stack.push(new (&state_pool)
-                               internal::DfsState<FST>(fst, arc.nextstate));
-          dfs = visitor->InitState(arc.nextstate, root);
-          break;
-        case kDfsGrey:
-          dfs = visitor->BackArc(s, arc);
-          aiter.Next();
-          break;
-        case kDfsBlack:
-          dfs = visitor->ForwardOrCrossArc(s, arc);
-          aiter.Next();
-          break;
-      }
-    }
-    if (access_only) break;
-    // Finds next tree root.
-    for (root = root == start ? 0 : root + 1;
-         root < nstates && state_color[root] != kDfsWhite; ++root) {
-    }
-    // Checks for a state beyond the largest known state.
-    if (!expanded && root == nstates) {
-      for (; !siter.Done(); siter.Next()) {
-        if (siter.Value() == nstates) {
-          ++nstates;
-          state_color.push_back(kDfsWhite);
-          break;
-        }
-      }
-    }
-  }
-  visitor->FinishVisit();
-}
-
-template <class Arc, class Visitor>
-void DfsVisit(const Fst<Arc> &fst, Visitor *visitor) {
-  DfsVisit(fst, visitor, AnyArcFilter<Arc>());
-}
-
-}  // namespace fst
-
-#endif  // FST_DFS_VISIT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/difference.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/difference.h
deleted file mode 100644
index bfe13213..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/difference.h
+++ /dev/null
@@ -1,177 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to compute the difference between two FSAs.
-
-#ifndef FST_DIFFERENCE_H_
-#define FST_DIFFERENCE_H_
-
-#include <memory>
-
-
-#include <fst/cache.h>
-#include <fst/complement.h>
-#include <fst/compose.h>
-
-
-namespace fst {
-
-template <class Arc, class M = Matcher<Fst<Arc>>,
-          class Filter = SequenceComposeFilter<M>,
-          class StateTable =
-              GenericComposeStateTable<Arc, typename Filter::FilterState>>
-struct DifferenceFstOptions
-    : public ComposeFstOptions<Arc, M, Filter, StateTable> {
-  explicit DifferenceFstOptions(const CacheOptions &opts = CacheOptions(),
-                                M *matcher1 = nullptr, M *matcher2 = nullptr,
-                                Filter *filter = nullptr,
-                                StateTable *state_table = nullptr)
-      : ComposeFstOptions<Arc, M, Filter, StateTable>(opts, matcher1, matcher2,
-                                                      filter, state_table) {}
-};
-
-// Computes the difference between two FSAs. This version is a delayed FST.
-// Only strings that are in the first automaton but not in second are retained
-// in the result.
-//
-// The first argument must be an acceptor; the second argument must be an
-// unweighted, epsilon-free, deterministic acceptor. One of the arguments must
-// be label-sorted.
-//
-// Complexity: same as ComposeFst.
-//
-// Caveats: same as ComposeFst.
-template <class A>
-class DifferenceFst : public ComposeFst<A> {
- public:
-  using Arc = A;
-  using Weight = typename Arc::Weight;
-  using StateId = typename Arc::StateId;
-
-  using ComposeFst<Arc>::CreateBase1;
-
-  // A - B = A ^ B'.
-  DifferenceFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-                const CacheOptions &opts = CacheOptions())
-      : ComposeFst<Arc>(CreateDifferenceImplWithCacheOpts(fst1, fst2, opts)) {
-    if (!fst1.Properties(kAcceptor, true)) {
-      FSTERROR() << "DifferenceFst: 1st argument not an acceptor";
-      GetImpl()->SetProperties(kError, kError);
-    }
-  }
-
-  template <class Matcher, class Filter, class StateTable>
-  DifferenceFst(
-      const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-      const DifferenceFstOptions<Arc, Matcher, Filter, StateTable> &opts)
-      : ComposeFst<Arc>(
-            CreateDifferenceImplWithDifferenceOpts(fst1, fst2, opts)) {
-    if (!fst1.Properties(kAcceptor, true)) {
-      FSTERROR() << "DifferenceFst: 1st argument not an acceptor";
-      GetImpl()->SetProperties(kError, kError);
-    }
-  }
-
-  // See Fst<>::Copy() for doc.
-  DifferenceFst(const DifferenceFst<Arc> &fst, bool safe = false)
-      : ComposeFst<Arc>(fst, safe) {}
-
-  // Get a copy of this DifferenceFst. See Fst<>::Copy() for further doc.
-  DifferenceFst<Arc> *Copy(bool safe = false) const override {
-    return new DifferenceFst<Arc>(*this, safe);
-  }
-
- private:
-  using Impl = internal::ComposeFstImplBase<Arc>;
-  using ImplToFst<Impl>::GetImpl;
-
-  static std::shared_ptr<Impl> CreateDifferenceImplWithCacheOpts(
-      const Fst<Arc> &fst1, const Fst<Arc> &fst2, const CacheOptions &opts) {
-    using RM = RhoMatcher<Matcher<Fst<A>>>;
-    ComplementFst<Arc> cfst(fst2);
-    ComposeFstOptions<A, RM> copts(
-        CacheOptions(), new RM(fst1, MATCH_NONE),
-        new RM(cfst, MATCH_INPUT, ComplementFst<Arc>::kRhoLabel));
-    return CreateBase1(fst1, cfst, copts);
-  }
-
-  template <class Matcher, class Filter, class StateTable>
-  static std::shared_ptr<Impl> CreateDifferenceImplWithDifferenceOpts(
-      const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-      const DifferenceFstOptions<Arc, Matcher, Filter, StateTable> &opts) {
-    using RM = RhoMatcher<Matcher>;
-    ComplementFst<Arc> cfst(fst2);
-    ComposeFstOptions<Arc, RM> copts(opts);
-    copts.matcher1 = new RM(fst1, MATCH_NONE, kNoLabel, MATCHER_REWRITE_ALWAYS,
-                            opts.matcher1);
-    copts.matcher2 = new RM(cfst, MATCH_INPUT, ComplementFst<Arc>::kRhoLabel,
-                            MATCHER_REWRITE_ALWAYS, opts.matcher2);
-    return CreateBase1(fst1, cfst, copts);
-  }
-};
-
-// Specialization for DifferenceFst.
-template <class Arc>
-class StateIterator<DifferenceFst<Arc>>
-    : public StateIterator<ComposeFst<Arc>> {
- public:
-  explicit StateIterator(const DifferenceFst<Arc> &fst)
-      : StateIterator<ComposeFst<Arc>>(fst) {}
-};
-
-// Specialization for DifferenceFst.
-template <class Arc>
-class ArcIterator<DifferenceFst<Arc>> : public ArcIterator<ComposeFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const DifferenceFst<Arc> &fst, StateId s)
-      : ArcIterator<ComposeFst<Arc>>(fst, s) {}
-};
-
-using DifferenceOptions = ComposeOptions;
-
-// Useful alias when using StdArc.
-using StdDifferenceFst = DifferenceFst<StdArc>;
-
-using DifferenceOptions = ComposeOptions;
-
-// Computes the difference between two FSAs. This version writes the difference
-// to an output MutableFst. Only strings that are in the first automaton but not
-// in the second are retained in the result.
-//
-// The first argument must be an acceptor; the second argument must be an
-// unweighted, epsilon-free, deterministic acceptor. One of the arguments must
-// be label-sorted.
-//
-// Complexity: same as Compose.
-//
-// Caveats: same as Compose.
-template <class Arc>
-void Difference(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-                MutableFst<Arc> *ofst,
-                const DifferenceOptions &opts = DifferenceOptions()) {
-  using M = Matcher<Fst<Arc>>;
-  if (opts.filter_type == AUTO_FILTER) {
-    CacheOptions nopts;
-    nopts.gc_limit = 0;  // Cache only the last state for fastest copy.
-    *ofst = DifferenceFst<Arc>(ifst1, ifst2, nopts);
-  } else if (opts.filter_type == SEQUENCE_FILTER) {
-    DifferenceFstOptions<Arc> dopts;
-    dopts.gc_limit = 0;  // Cache only the last state for fastest copy.
-    *ofst = DifferenceFst<Arc>(ifst1, ifst2, dopts);
-  } else if (opts.filter_type == ALT_SEQUENCE_FILTER) {
-    DifferenceFstOptions<Arc, M, AltSequenceComposeFilter<M>> dopts;
-    dopts.gc_limit = 0;  // Cache only the last state for fastest copy.
-    *ofst = DifferenceFst<Arc>(ifst1, ifst2, dopts);
-  } else if (opts.filter_type == MATCH_FILTER) {
-    DifferenceFstOptions<Arc, M, MatchComposeFilter<M>> dopts;
-    dopts.gc_limit = 0;  // Cache only the last state for fastest copy.
-    *ofst = DifferenceFst<Arc>(ifst1, ifst2, dopts);
-  }
-  if (opts.connect) Connect(ofst);
-}
-
-}  // namespace fst
-
-#endif  // FST_DIFFERENCE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/disambiguate.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/disambiguate.h
deleted file mode 100644
index 7e3ba37c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/disambiguate.h
+++ /dev/null
@@ -1,564 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to disambiguate an FST.
-
-#ifndef FST_DISAMBIGUATE_H_
-#define FST_DISAMBIGUATE_H_
-
-#include <list>
-#include <map>
-#include <set>
-#include <utility>
-#include <vector>
-
-#include <fst/arcsort.h>
-#include <fst/compose.h>
-#include <fst/connect.h>
-#include <fst/determinize.h>
-#include <fst/dfs-visit.h>
-#include <fst/project.h>
-#include <fst/prune.h>
-#include <fst/state-map.h>
-#include <fst/state-table.h>
-#include <fst/union-find.h>
-#include <fst/verify.h>
-
-
-namespace fst {
-
-template <class Arc>
-struct DisambiguateOptions : public DeterminizeOptions<Arc> {
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit DisambiguateOptions(float delta = kDelta,
-                               Weight weight = Weight::Zero(),
-                               StateId n = kNoStateId, Label label = 0)
-      : DeterminizeOptions<Arc>(delta, std::move(weight), n, label,
-                                DETERMINIZE_FUNCTIONAL) {}
-};
-
-namespace internal {
-
-// A determinization filter based on a subset element relation. The relation is
-// assumed to be reflexive and symmetric.
-template <class Arc, class Relation>
-class RelationDeterminizeFilter {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FilterState = IntegerFilterState<StateId>;
-  using StateTuple = DeterminizeStateTuple<Arc, FilterState>;
-  using Subset = typename StateTuple::Subset;
-  using Element = typename StateTuple::Element;
-  using LabelMap = std::multimap<Label, DeterminizeArc<StateTuple>>;
-
-  // This is needed (e.g.) to go into the gallic domain for transducers; there
-  // is no need to rebind the relation since its use here only depends on the
-  // state IDs.
-  template <class A>
-  struct rebind {
-    using Other = RelationDeterminizeFilter<A, Relation>;
-  };
-
-  explicit RelationDeterminizeFilter(const Fst<Arc> &fst)
-      : fst_(fst.Copy()), r_(new Relation()), s_(kNoStateId), head_(nullptr) {}
-
-  // Ownership of the relation is given to this class.
-  RelationDeterminizeFilter(const Fst<Arc> &fst, Relation *r)
-      : fst_(fst.Copy()), r_(r), s_(kNoStateId), head_(0) {}
-
-  // Ownership of the relation is given to this class.
-  RelationDeterminizeFilter(const Fst<Arc> &fst, Relation *r,
-                            std::vector<StateId> *head)
-      : fst_(fst.Copy()), r_(r), s_(kNoStateId), head_(head) {}
-
-  // This is needed, e.g., to go into the gallic domain for transducers.
-  // Ownership of the templated filter argument is given to this class.
-  template <class Filter>
-  RelationDeterminizeFilter(const Fst<Arc> &fst, Filter *filter)
-      : fst_(fst.Copy()),
-        r_(new Relation(filter->GetRelation())),
-        s_(kNoStateId),
-        head_(filter->GetHeadStates()) {
-    delete filter;
-  }
-
-  // Copy constructor; the FST can be passed if it has been deep-copied.
-  RelationDeterminizeFilter(const RelationDeterminizeFilter &filter,
-                            const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? fst->Copy() : filter.fst_->Copy()),
-        r_(new Relation(*filter.r_)),
-        s_(kNoStateId),
-        head_() {}
-
-  FilterState Start() const { return FilterState(fst_->Start()); }
-
-  void SetState(StateId s, const StateTuple &tuple) {
-    if (s_ != s) {
-      s_ = s;
-      tuple_ = &tuple;
-      const auto head = tuple.filter_state.GetState();
-      is_final_ = fst_->Final(head) != Weight::Zero();
-      if (head_) {
-        if (head_->size() <= s) head_->resize(s + 1, kNoStateId);
-        (*head_)[s] = head;
-      }
-    }
-  }
-
-  // Filters transition, possibly modifying label map. Returns true if arc is
-  // added to label map.
-  bool FilterArc(const Arc &arc, const Element &src_element,
-                 const Element &dest_element, LabelMap *label_map) const;
-
-  // Filters super-final transition, returning new final weight.
-  Weight FilterFinal(const Weight final_weight, const Element &element) const {
-    return is_final_ ? final_weight : Weight::Zero();
-  }
-
-  static uint64 Properties(uint64 props) {
-    return props & ~(kIDeterministic | kODeterministic);
-  }
-
-  const Relation &GetRelation() { return *r_; }
-
-  std::vector<StateId> *GetHeadStates() { return head_; }
-
- private:
-  // Pairs arc labels with state tuples with possible heads and empty subsets.
-  void InitLabelMap(LabelMap *label_map) const;
-
-  std::unique_ptr<Fst<Arc>> fst_;  // Input FST.
-  std::unique_ptr<Relation> r_;    // Relation compatible with inv. trans. fnc.
-  StateId s_;                      // Current state.
-  const StateTuple *tuple_;        // Current tuple.
-  bool is_final_;                  // Is the current head state final?
-  std::vector<StateId> *head_;     // Head state for a given state,
-                                   // owned by the Disambiguator.
-};
-
-template <class Arc, class Relation>
-bool RelationDeterminizeFilter<Arc, Relation>::FilterArc(
-    const Arc &arc, const Element &src_element, const Element &dest_element,
-    LabelMap *label_map) const {
-  bool added = false;
-  if (label_map->empty()) InitLabelMap(label_map);
-  // Adds element to state tuple if element state is related to tuple head.
-  for (auto liter = label_map->lower_bound(arc.ilabel);
-       liter != label_map->end() && liter->first == arc.ilabel; ++liter) {
-    auto *dest_tuple = liter->second.dest_tuple;
-    const auto dest_head = dest_tuple->filter_state.GetState();
-    if ((*r_)(dest_element.state_id, dest_head)) {
-      dest_tuple->subset.push_front(dest_element);
-      added = true;
-    }
-  }
-  return added;
-}
-
-template <class Arc, class Relation>
-void RelationDeterminizeFilter<Arc, Relation>::InitLabelMap(
-    LabelMap *label_map) const {
-  const auto src_head = tuple_->filter_state.GetState();
-  Label label = kNoLabel;
-  StateId nextstate = kNoStateId;
-  for (ArcIterator<Fst<Arc>> aiter(*fst_, src_head); !aiter.Done();
-       aiter.Next()) {
-    const auto &arc = aiter.Value();
-    // Continues if multiarc.
-    if (arc.ilabel == label && arc.nextstate == nextstate) continue;
-    DeterminizeArc<StateTuple> det_arc(arc);
-    det_arc.dest_tuple->filter_state = FilterState(arc.nextstate);
-    label_map->insert(std::make_pair(arc.ilabel, det_arc));
-    label = arc.ilabel;
-    nextstate = arc.nextstate;
-  }
-}
-
-// Helper class to disambiguate an FST via Disambiguate().
-template <class Arc>
-class Disambiguator {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // IDs arcs with state ID and arc position. Arc position -1 indicates final
-  // (super-final transition).
-  using ArcId = std::pair<StateId, ssize_t>;
-
-  Disambiguator() : error_(false) {}
-
-  void Disambiguate(
-      const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-      const DisambiguateOptions<Arc> &opts = DisambiguateOptions<Arc>()) {
-    VectorFst<Arc> sfst(ifst);
-    Connect(&sfst);
-    ArcSort(&sfst, ArcCompare());
-    PreDisambiguate(sfst, ofst, opts);
-    ArcSort(ofst, ArcCompare());
-    FindAmbiguities(*ofst);
-    RemoveSplits(ofst);
-    MarkAmbiguities();
-    RemoveAmbiguities(ofst);
-    if (error_) ofst->SetProperties(kError, kError);
-  }
-
- private:
-  // Comparison functor for comparing input labels and next states of arcs. This
-  // sort order facilitates the predisambiguation.
-  class ArcCompare {
-   public:
-    bool operator()(const Arc &arc1, const Arc &arc2) const {
-      return arc1.ilabel < arc2.ilabel ||
-             (arc1.ilabel == arc2.ilabel && arc1.nextstate < arc2.nextstate);
-    }
-
-    uint64 Properties(uint64 props) const {
-      return (props & kArcSortProperties) | kILabelSorted |
-             (props & kAcceptor ? kOLabelSorted : 0);
-    }
-  };
-
-  // Comparison functor for comparing transitions represented by their arc ID.
-  // This sort order facilitates ambiguity detection.
-  class ArcIdCompare {
-   public:
-    explicit ArcIdCompare(const std::vector<StateId> &head) : head_(head) {}
-
-    bool operator()(const ArcId &a1, const ArcId &a2) const {
-      // Sort first by source head state...
-      const auto src1 = a1.first;
-      const auto src2 = a2.first;
-      const auto head1 = head_[src1];
-      const auto head2 = head_[src2];
-      if (head1 < head2) return true;
-      if (head2 < head1) return false;
-      // ...then by source state...
-      if (src1 < src2) return true;
-      if (src2 < src1) return false;
-      // ...then by position.
-      return a1.second < a2.second;
-    }
-
-   private:
-    const std::vector<StateId> &head_;
-  };
-
-  // A relation that determines if two states share a common future.
-  class CommonFuture {
-   public:
-    using StateTable = GenericComposeStateTable<Arc, TrivialFilterState>;
-    using StateTuple = typename StateTable::StateTuple;
-
-    // Needed for compilation with DeterminizeRelationFilter.
-    CommonFuture() {
-      FSTERROR() << "Disambiguate::CommonFuture: FST not provided";
-    }
-
-    explicit CommonFuture(const Fst<Arc> &ifst) {
-      using M = Matcher<Fst<Arc>>;
-      ComposeFstOptions<Arc, M, NullComposeFilter<M>> opts;
-      // Ensures composition is between acceptors.
-      const bool trans = ifst.Properties(kNotAcceptor, true);
-      const auto *fsa =
-          trans ? new ProjectFst<Arc>(ifst, PROJECT_INPUT) : &ifst;
-      opts.state_table = new StateTable(*fsa, *fsa);
-      const ComposeFst<Arc> cfst(*fsa, *fsa, opts);
-      std::vector<bool> coaccess;
-      uint64 props = 0;
-      SccVisitor<Arc> scc_visitor(nullptr, nullptr, &coaccess, &props);
-      DfsVisit(cfst, &scc_visitor);
-      for (StateId s = 0; s < coaccess.size(); ++s) {
-        if (coaccess[s]) {
-          related_.insert(opts.state_table->Tuple(s).StatePair());
-        }
-      }
-      if (trans) delete fsa;
-    }
-
-    bool operator()(const StateId s1, StateId s2) const {
-      return related_.count(std::make_pair(s1, s2)) > 0;
-    }
-
-   private:
-    // States s1 and s2 resp. are in this relation iff they there is a
-    // path from s1 to a final state that has the same label as some
-    // path from s2 to a final state.
-    std::set<std::pair<StateId, StateId>> related_;
-  };
-
-  using ArcIdMap = std::multimap<ArcId, ArcId, ArcIdCompare>;
-
-  // Inserts candidate into the arc ID map.
-  inline void InsertCandidate(StateId s1, StateId s2, const ArcId &a1,
-                              const ArcId &a2) {
-    candidates_->insert(head_[s1] > head_[s2] ? std::make_pair(a1, a2)
-                                              : std::make_pair(a2, a1));
-  }
-
-  // Returns the arc corresponding to ArcId a.
-  static Arc GetArc(const Fst<Arc> &fst, ArcId aid) {
-    if (aid.second == -1) {  // Returns super-final transition.
-      return Arc(kNoLabel, kNoLabel, fst.Final(aid.first), kNoStateId);
-    } else {
-      ArcIterator<Fst<Arc>> aiter(fst, aid.first);
-      aiter.Seek(aid.second);
-      return aiter.Value();
-    }
-  }
-
-  // Outputs an equivalent FST whose states are subsets of states that have a
-  // future path in common.
-  void PreDisambiguate(const ExpandedFst<Arc> &ifst, MutableFst<Arc> *ofst,
-                       const DisambiguateOptions<Arc> &opts);
-
-  // Finds transitions that are ambiguous candidates in the result of
-  // PreDisambiguate.
-  void FindAmbiguities(const ExpandedFst<Arc> &fst);
-
-  // Finds transition pairs that are ambiguous candidates from two specified
-  // source states.
-  void FindAmbiguousPairs(const ExpandedFst<Arc> &fst, StateId s1, StateId s2);
-
-  // Marks ambiguous transitions to be removed.
-  void MarkAmbiguities();
-
-  // Deletes spurious ambiguous transitions (due to quantization).
-  void RemoveSplits(MutableFst<Arc> *ofst);
-
-  // Deletes actual ambiguous transitions.
-  void RemoveAmbiguities(MutableFst<Arc> *ofst);
-
-  // States s1 and s2 are in this relation iff there is a path from the initial
-  // state to s1 that has the same label as some path from the initial state to
-  // s2. We store only state pairs s1, s2 such that s1 <= s2.
-  std::set<std::pair<StateId, StateId>> coreachable_;
-
-  // Queue of disambiguation-related states to be processed. We store only
-  // state pairs s1, s2 such that s1 <= s2.
-  std::list<std::pair<StateId, StateId>> queue_;
-
-  // Head state in the pre-disambiguation for a given state.
-  std::vector<StateId> head_;
-
-  // Maps from a candidate ambiguous arc A to each ambiguous candidate arc B
-  // with the same label and destination state as A, whose source state s' is
-  // coreachable with the source state s of A, and for which head(s') < head(s).
-  std::unique_ptr<ArcIdMap> candidates_;
-
-  // Set of ambiguous transitions to be removed.
-  std::set<ArcId> ambiguous_;
-
-  // States to merge due to quantization issues.
-  std::unique_ptr<UnionFind<StateId>> merge_;
-  // Marks error condition.
-  bool error_;
-
-  Disambiguator(const Disambiguator &) = delete;
-  Disambiguator &operator=(const Disambiguator &) = delete;
-};
-
-template <class Arc>
-void Disambiguator<Arc>::PreDisambiguate(const ExpandedFst<Arc> &ifst,
-                                         MutableFst<Arc> *ofst,
-                                         const DisambiguateOptions<Arc> &opts) {
-  using CommonDivisor = DefaultCommonDivisor<Weight>;
-  using Filter = RelationDeterminizeFilter<Arc, CommonFuture>;
-  // Subset elements with states s1 and s2 (resp.) are in this relation iff they
-  // there is a path from s1 to a final state that has the same label as some
-  // path from s2 to a final state.
-  auto *common_future = new CommonFuture(ifst);
-  DeterminizeFstOptions<Arc, CommonDivisor, Filter> nopts;
-  nopts.delta = opts.delta;
-  nopts.subsequential_label = opts.subsequential_label;
-  nopts.filter = new Filter(ifst, common_future, &head_);
-  // The filter takes ownership of 'common_future', and determinization takes
-  // ownership of the filter itself.
-  nopts.gc_limit = 0;  // Cache only the last state for fastest copy.
-  if (opts.weight_threshold != Weight::Zero() ||
-      opts.state_threshold != kNoStateId) {
-    /* TODO(riley): fails regression test; understand why
-    if (ifst.Properties(kAcceptor, true)) {
-      std::vector<Weight> idistance, odistance;
-      ShortestDistance(ifst, &idistance, true);
-      DeterminizeFst<Arc> dfst(ifst, &idistance, &odistance, nopts);
-      PruneOptions< Arc, AnyArcFilter<Arc>> popts(opts.weight_threshold,
-                                                   opts.state_threshold,
-                                                   AnyArcFilter<Arc>(),
-                                                   &odistance);
-      Prune(dfst, ofst, popts);
-      } else */ {
-      *ofst = DeterminizeFst<Arc>(ifst, nopts);
-      Prune(ofst, opts.weight_threshold, opts.state_threshold);
-    }
-  } else {
-    *ofst = DeterminizeFst<Arc>(ifst, nopts);
-  }
-  head_.resize(ofst->NumStates(), kNoStateId);
-}
-
-template <class Arc>
-void Disambiguator<Arc>::FindAmbiguities(const ExpandedFst<Arc> &fst) {
-  if (fst.Start() == kNoStateId) return;
-  candidates_.reset(new ArcIdMap(ArcIdCompare(head_)));
-  const auto start_pr = std::make_pair(fst.Start(), fst.Start());
-  coreachable_.insert(start_pr);
-  queue_.push_back(start_pr);
-  while (!queue_.empty()) {
-    const auto &pr = queue_.front();
-    const auto s1 = pr.first;
-    const auto s2 = pr.second;
-    queue_.pop_front();
-    FindAmbiguousPairs(fst, s1, s2);
-  }
-}
-
-template <class Arc>
-void Disambiguator<Arc>::FindAmbiguousPairs(const ExpandedFst<Arc> &fst,
-                                            StateId s1, StateId s2) {
-  if (fst.NumArcs(s2) > fst.NumArcs(s1)) FindAmbiguousPairs(fst, s2, s1);
-  SortedMatcher<Fst<Arc>> matcher(fst, MATCH_INPUT);
-  matcher.SetState(s2);
-  for (ArcIterator<Fst<Arc>> aiter(fst, s1); !aiter.Done(); aiter.Next()) {
-    const auto &arc1 = aiter.Value();
-    const ArcId a1(s1, aiter.Position());
-    if (matcher.Find(arc1.ilabel)) {
-      for (; !matcher.Done(); matcher.Next()) {
-        const auto &arc2 = matcher.Value();
-        // Continues on implicit epsilon match.
-        if (arc2.ilabel == kNoLabel) continue;
-        const ArcId a2(s2, matcher.Position());
-        // Actual transition is ambiguous.
-        if (s1 != s2 && arc1.nextstate == arc2.nextstate) {
-          InsertCandidate(s1, s2, a1, a2);
-        }
-        const auto spr = arc1.nextstate <= arc2.nextstate
-                             ? std::make_pair(arc1.nextstate, arc2.nextstate)
-                             : std::make_pair(arc2.nextstate, arc1.nextstate);
-        // Not already marked as coreachable?
-        if (coreachable_.insert(spr).second) {
-          // Only possible if state split by quantization issues.
-          if (spr.first != spr.second &&
-              head_[spr.first] == head_[spr.second]) {
-            if (!merge_) {
-              merge_.reset(new UnionFind<StateId>(fst.NumStates(), kNoStateId));
-              merge_->MakeAllSet(fst.NumStates());
-            }
-            merge_->Union(spr.first, spr.second);
-          } else {
-            queue_.push_back(spr);
-          }
-        }
-      }
-    }
-  }
-  // Super-final transition is ambiguous.
-  if (s1 != s2 && fst.Final(s1) != Weight::Zero() &&
-      fst.Final(s2) != Weight::Zero()) {
-    const ArcId a1(s1, -1);
-    const ArcId a2(s2, -1);
-    InsertCandidate(s1, s2, a1, a2);
-  }
-}
-
-template <class Arc>
-void Disambiguator<Arc>::MarkAmbiguities() {
-  if (!candidates_) return;
-  for (auto it = candidates_->begin(); it != candidates_->end(); ++it) {
-    const auto a = it->first;
-    const auto b = it->second;
-    // If b is not to be removed, then a is.
-    if (ambiguous_.count(b) == 0) ambiguous_.insert(a);
-  }
-  coreachable_.clear();
-  candidates_.reset();
-}
-
-template <class Arc>
-void Disambiguator<Arc>::RemoveSplits(MutableFst<Arc> *ofst) {
-  if (!merge_) return;
-  // Merges split states to remove spurious ambiguities.
-  for (StateIterator<MutableFst<Arc>> siter(*ofst); !siter.Done();
-       siter.Next()) {
-    for (MutableArcIterator<MutableFst<Arc>> aiter(ofst, siter.Value());
-         !aiter.Done(); aiter.Next()) {
-      auto arc = aiter.Value();
-      const auto nextstate = merge_->FindSet(arc.nextstate);
-      if (nextstate != arc.nextstate) {
-        arc.nextstate = nextstate;
-        aiter.SetValue(arc);
-      }
-    }
-  }
-  // Repeats search for actual ambiguities on modified FST.
-  coreachable_.clear();
-  merge_.reset();
-  candidates_.reset();
-  FindAmbiguities(*ofst);
-  if (merge_) {  // Shouldn't get here; sanity test.
-    FSTERROR() << "Disambiguate: Unable to remove spurious ambiguities";
-    error_ = true;
-    return;
-  }
-}
-
-template <class Arc>
-void Disambiguator<Arc>::RemoveAmbiguities(MutableFst<Arc> *ofst) {
-  if (ambiguous_.empty()) return;
-  // Adds dead state to redirect ambiguous transitions to be removed.
-  const auto dead = ofst->AddState();
-  for (auto it = ambiguous_.begin(); it != ambiguous_.end(); ++it) {
-    const auto pos = it->second;
-    if (pos >= 0) {  // Actual transition.
-      MutableArcIterator<MutableFst<Arc>> aiter(ofst, it->first);
-      aiter.Seek(pos);
-      auto arc = aiter.Value();
-      arc.nextstate = dead;
-      aiter.SetValue(arc);
-    } else {  // Super-final transition.
-      ofst->SetFinal(it->first, Weight::Zero());
-    }
-  }
-  Connect(ofst);
-  ambiguous_.clear();
-}
-
-}  // namespace internal
-
-// Disambiguates a weighted FST. This version writes the disambiguated FST to an
-// output MutableFst. The result will be an equivalent FST that has the
-// property that there are not two distinct paths from the initial state to a
-// final state with the same input labeling.
-//
-// The weights must be (weakly) left divisible (valid for Tropical and
-// LogWeight).
-//
-// Complexity:
-//
-//   Disambiguable: exponential (polynomial in the size of the output).
-//   Non-disambiguable: does not terminate.
-//
-// The disambiguable transducers include all automata and functional transducers
-// that are unweighted or that are acyclic or that are unambiguous.
-//
-// For more information, see:
-//
-// Mohri, M. and Riley, M. 2015. On the disambiguation of weighted automata.
-// In CIAA, pages 263-278.
-template <class Arc>
-void Disambiguate(
-    const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-    const DisambiguateOptions<Arc> &opts = DisambiguateOptions<Arc>()) {
-  internal::Disambiguator<Arc> disambiguator;
-  disambiguator.Disambiguate(ifst, ofst, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_DISAMBIGUATE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/edit-fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/edit-fst.h
deleted file mode 100644
index e7bdfd61..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/edit-fst.h
+++ /dev/null
@@ -1,702 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// An FST implementation that allows non-destructive edit operations on an
-// existing FST.
-//
-// The EditFst class enables non-destructive edit operations on a wrapped
-// ExpandedFst. The implementation uses copy-on-write semantics at the node
-// level: if a user has an underlying fst on which he or she wants to perform a
-// relatively small number of edits (read: mutations), then this implementation
-// will copy the edited node to an internal MutableFst and perform any edits in
-// situ on that copied node. This class supports all the methods of MutableFst
-// except for DeleteStates(const std::vector<StateId> &); thus, new nodes may
-// also be
-// added, and one may add transitions from existing nodes of the wrapped fst to
-// new nodes.
-//
-// N.B.: The documentation for Fst::Copy(true) says that its behavior is
-// undefined if invoked on an fst that has already been accessed.  This class
-// requires that the Fst implementation it wraps provides consistent, reliable
-// behavior when its Copy(true) method is invoked, where consistent means
-// the graph structure, graph properties and state numbering and do not change.
-// VectorFst and CompactFst, for example, are both well-behaved in this regard.
-
-#ifndef FST_EDIT_FST_H_
-#define FST_EDIT_FST_H_
-
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-
-
-namespace fst {
-namespace internal {
-
-// The EditFstData class is a container for all mutable data for EditFstImpl;
-// also, this class provides most of the actual implementation of what EditFst
-// does (that is, most of EditFstImpl's methods delegate to methods in this, the
-// EditFstData class).  Instances of this class are reference-counted and can be
-// shared between otherwise independent EditFstImpl instances. This scheme
-// allows EditFstImpl to implement the thread-safe, copy-on-write semantics
-// required by Fst::Copy(true).
-//
-// template parameters:
-//   A the type of arc to use
-//   WrappedFstT the type of fst wrapped by the EditFst instance that
-//     this EditFstData instance is backing
-//   MutableFstT the type of mutable fst to use internally for edited states;
-//     crucially, MutableFstT::Copy(false) *must* yield an fst that is
-//     thread-safe for reading (VectorFst, for example, has this property)
-template <typename Arc, typename WrappedFstT = ExpandedFst<Arc>,
-          typename MutableFstT = VectorFst<Arc>>
-class EditFstData {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  EditFstData() : num_new_states_(0) {}
-
-  EditFstData(const EditFstData &other)
-      : edits_(other.edits_),
-        external_to_internal_ids_(other.external_to_internal_ids_),
-        edited_final_weights_(other.edited_final_weights_),
-        num_new_states_(other.num_new_states_) {}
-
-  ~EditFstData() {}
-
-  static EditFstData<Arc, WrappedFstT, MutableFstT> *Read(
-      std::istream &strm, const FstReadOptions &opts);
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
-    // Serialize all private data members of this class.
-    FstWriteOptions edits_opts(opts);
-    edits_opts.write_header = true;  // Force writing contained header.
-    edits_.Write(strm, edits_opts);
-    WriteType(strm, external_to_internal_ids_);
-    WriteType(strm, edited_final_weights_);
-    WriteType(strm, num_new_states_);
-    if (!strm) {
-      LOG(ERROR) << "EditFstData::Write: Write failed: " << opts.source;
-      return false;
-    }
-    return true;
-  }
-
-  StateId NumNewStates() const { return num_new_states_; }
-
-  // accessor methods for the fst holding edited states
-  StateId EditedStart() const { return edits_.Start(); }
-
-  Weight Final(StateId s, const WrappedFstT *wrapped) const {
-    auto final_weight_it = GetFinalWeightIterator(s);
-    if (final_weight_it == NotInFinalWeightMap()) {
-      auto it = GetEditedIdMapIterator(s);
-      return it == NotInEditedMap() ? wrapped->Final(s)
-                                    : edits_.Final(it->second);
-    } else {
-      return final_weight_it->second;
-    }
-  }
-
-  size_t NumArcs(StateId s, const WrappedFstT *wrapped) const {
-    auto it = GetEditedIdMapIterator(s);
-    return it == NotInEditedMap() ? wrapped->NumArcs(s)
-                                  : edits_.NumArcs(it->second);
-  }
-
-  size_t NumInputEpsilons(StateId s, const WrappedFstT *wrapped) const {
-    auto it = GetEditedIdMapIterator(s);
-    return it == NotInEditedMap() ? wrapped->NumInputEpsilons(s)
-                                  : edits_.NumInputEpsilons(it->second);
-  }
-
-  size_t NumOutputEpsilons(StateId s, const WrappedFstT *wrapped) const {
-    auto it = GetEditedIdMapIterator(s);
-    return it == NotInEditedMap() ? wrapped->NumOutputEpsilons(s)
-                                  : edits_.NumOutputEpsilons(it->second);
-  }
-
-  void SetEditedProperties(uint64 props, uint64 mask) {
-    edits_.SetProperties(props, mask);
-  }
-
-  // Non-const MutableFst operations.
-
-  // Sets the start state for this FST.
-  void SetStart(StateId s) { edits_.SetStart(s); }
-
-  // Sets the final state for this FST.
-  Weight SetFinal(StateId s, Weight w, const WrappedFstT *wrapped) {
-    Weight old_weight = Final(s, wrapped);
-    auto it = GetEditedIdMapIterator(s);
-    // If we haven't already edited state s, don't add it to edited_ (which can
-    // be expensive if s has many transitions); just use the
-    // edited_final_weights_ map.
-    if (it == NotInEditedMap()) {
-      edited_final_weights_[s] = w;
-    } else {
-      edits_.SetFinal(GetEditableInternalId(s, wrapped), w);
-    }
-    return old_weight;
-  }
-
-  // Adds a new state to this FST, initially with no arcs.
-  StateId AddState(StateId curr_num_states) {
-    StateId internal_state_id = edits_.AddState();
-    StateId external_state_id = curr_num_states;
-    external_to_internal_ids_[external_state_id] = internal_state_id;
-    num_new_states_++;
-    return external_state_id;
-  }
-
-  // Adds the specified arc to the specified state of this FST.
-  const Arc *AddArc(StateId s, const Arc &arc, const WrappedFstT *wrapped) {
-    const auto internal_id = GetEditableInternalId(s, wrapped);
-    const auto num_arcs = edits_.NumArcs(internal_id);
-    ArcIterator<MutableFstT> arc_it(edits_, internal_id);
-    const Arc *prev_arc = nullptr;
-    if (num_arcs > 0) {
-      // grab the final arc associated with this state in edits_
-      arc_it.Seek(num_arcs - 1);
-      prev_arc = &(arc_it.Value());
-    }
-    edits_.AddArc(internal_id, arc);
-    return prev_arc;
-  }
-
-  void DeleteStates() {
-    edits_.DeleteStates();
-    num_new_states_ = 0;
-    external_to_internal_ids_.clear();
-    edited_final_weights_.clear();
-  }
-
-  // Removes all but the first n outgoing arcs of the specified state.
-  void DeleteArcs(StateId s, size_t n, const WrappedFstT *wrapped) {
-    edits_.DeleteArcs(GetEditableInternalId(s, wrapped), n);
-  }
-
-  // Removes all outgoing arcs from the specified state.
-  void DeleteArcs(StateId s, const WrappedFstT *wrapped) {
-    edits_.DeleteArcs(GetEditableInternalId(s, wrapped));
-  }
-
-  // End methods for non-const MutableFst operations.
-
-  // Provides information for the generic arc iterator.
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data,
-                       const WrappedFstT *wrapped) const {
-    auto id_map_it = GetEditedIdMapIterator(s);
-    if (id_map_it == NotInEditedMap()) {
-      VLOG(3) << "EditFstData::InitArcIterator: iterating on state " << s
-              << " of original fst";
-      wrapped->InitArcIterator(s, data);
-    } else {
-      VLOG(2) << "EditFstData::InitArcIterator: iterating on edited state " << s
-              << " (internal state id: " << id_map_it->second << ")";
-      edits_.InitArcIterator(id_map_it->second, data);
-    }
-  }
-
-  // Provides information for the generic mutable arc iterator.
-  void InitMutableArcIterator(StateId s, MutableArcIteratorData<Arc> *data,
-                              const WrappedFstT *wrapped) {
-    data->base = new MutableArcIterator<MutableFstT>(
-        &edits_, GetEditableInternalId(s, wrapped));
-  }
-
-  // Prints out the map from external to internal state id's (for debugging
-  // purposes).
-  void PrintMap() {
-    for (auto map_it = external_to_internal_ids_.begin();
-         map_it != NotInEditedMap(); ++map_it) {
-      LOG(INFO) << "(external,internal)=(" << map_it->first << ","
-                << map_it->second << ")";
-    }
-  }
-
- private:
-  // Returns the iterator of the map from external to internal state id's
-  // of edits_ for the specified external state id.
-  typename std::unordered_map<StateId, StateId>::const_iterator
-      GetEditedIdMapIterator(StateId s) const {
-    return external_to_internal_ids_.find(s);
-  }
-
-  typename std::unordered_map<StateId, StateId>::const_iterator
-      NotInEditedMap() const {
-    return external_to_internal_ids_.end();
-  }
-
-  typename std::unordered_map<StateId, Weight>::const_iterator
-      GetFinalWeightIterator(StateId s) const {
-    return edited_final_weights_.find(s);
-  }
-
-  typename std::unordered_map<StateId, Weight>::const_iterator
-      NotInFinalWeightMap() const {
-    return edited_final_weights_.end();
-  }
-
-  // Returns the internal state ID of the specified external ID if the state has
-  // already been made editable, or else copies the state from wrapped_ to
-  // edits_ and returns the state id of the newly editable state in edits_.
-  StateId GetEditableInternalId(StateId s, const WrappedFstT *wrapped) {
-    auto id_map_it = GetEditedIdMapIterator(s);
-    if (id_map_it == NotInEditedMap()) {
-      StateId new_internal_id = edits_.AddState();
-      VLOG(2) << "EditFstData::GetEditableInternalId: editing state " << s
-              << " of original fst; new internal state id:" << new_internal_id;
-      external_to_internal_ids_[s] = new_internal_id;
-      for (ArcIterator<Fst<Arc>> arc_iterator(*wrapped, s);
-           !arc_iterator.Done(); arc_iterator.Next()) {
-        edits_.AddArc(new_internal_id, arc_iterator.Value());
-      }
-      // Copies the final weight.
-      auto final_weight_it = GetFinalWeightIterator(s);
-      if (final_weight_it == NotInFinalWeightMap()) {
-        edits_.SetFinal(new_internal_id, wrapped->Final(s));
-      } else {
-        edits_.SetFinal(new_internal_id, final_weight_it->second);
-        edited_final_weights_.erase(s);
-      }
-      return new_internal_id;
-    } else {
-      return id_map_it->second;
-    }
-  }
-
-  // A mutable FST (by default, a VectorFst) to contain new states, and/or
-  // copies of states from a wrapped ExpandedFst that have been modified in
-  // some way.
-  MutableFstT edits_;
-  // A mapping from external state IDs to the internal IDs of states that
-  // appear in edits_.
-  std::unordered_map<StateId, StateId> external_to_internal_ids_;
-  // A mapping from external state IDs to final state weights assigned to
-  // those states.  The states in this map are *only* those whose final weight
-  // has been modified; if any other part of the state has been modified,
-  // the entire state is copied to edits_, and all modifications reside there.
-  std::unordered_map<StateId, Weight> edited_final_weights_;
-  // The number of new states added to this mutable fst impl, which is <= the
-  // number of states in edits_ (since edits_ contains both edited *and* new
-  // states).
-  StateId num_new_states_;
-};
-
-// EditFstData method implementations: just the Read method.
-template <typename A, typename WrappedFstT, typename MutableFstT>
-EditFstData<A, WrappedFstT, MutableFstT> *
-EditFstData<A, WrappedFstT, MutableFstT>::Read(std::istream &strm,
-                                               const FstReadOptions &opts) {
-  auto *data = new EditFstData<A, WrappedFstT, MutableFstT>();
-  // next read in MutabelFstT machine that stores edits
-  FstReadOptions edits_opts(opts);
-  // Contained header was written out, so read it in.
-  edits_opts.header = nullptr;
-
-  // Because our internal representation of edited states is a solid object
-  // of type MutableFstT (defaults to VectorFst<A>) and not a pointer,
-  // and because the static Read method allocates a new object on the heap,
-  // we need to call Read, check if there was a failure, use
-  // MutableFstT::operator= to assign the object (not the pointer) to the
-  // edits_ data member (which will increase the ref count by 1 on the impl)
-  // and, finally, delete the heap-allocated object.
-  std::unique_ptr<MutableFstT> edits(MutableFstT::Read(strm, edits_opts));
-  if (!edits) return nullptr;
-  data->edits_ = *edits;
-  edits.reset();
-  // Finally, reads in rest of private data members.
-  ReadType(strm, &data->external_to_internal_ids_);
-  ReadType(strm, &data->edited_final_weights_);
-  ReadType(strm, &data->num_new_states_);
-  if (!strm) {
-    LOG(ERROR) << "EditFst::Read: read failed: " << opts.source;
-    return nullptr;
-  }
-  return data;
-}
-
-// This class enables non-destructive edit operations on a wrapped ExpandedFst.
-// The implementation uses copy-on-write semantics at the node level: if a user
-// has an underlying fst on which he or she wants to perform a relatively small
-// number of edits (read: mutations), then this implementation will copy the
-// edited node to an internal MutableFst and perform any edits in situ on that
-// copied node. This class supports all the methods of MutableFst except for
-// DeleteStates(const std::vector<StateId> &); thus, new nodes may also be
-// added, and
-// one may add transitions from existing nodes of the wrapped fst to new nodes.
-//
-// template parameters:
-//   A the type of arc to use
-//   WrappedFstT the type of fst wrapped by the EditFst instance that
-//     this EditFstImpl instance is backing
-//   MutableFstT the type of mutable fst to use internally for edited states;
-//     crucially, MutableFstT::Copy(false) *must* yield an fst that is
-//     thread-safe for reading (VectorFst, for example, has this property)
-template <typename A, typename WrappedFstT = ExpandedFst<A>,
-          typename MutableFstT = VectorFst<A>>
-class EditFstImpl : public FstImpl<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-  using FstImpl<Arc>::WriteHeader;
-
-  // Constructs an editable FST implementation with no states. Effectively, this
-  // initially-empty fst will in every way mimic the behavior of a
-  // VectorFst---more precisely, a VectorFstImpl instance---but with slightly
-  // slower performance (by a constant factor), due to the fact that
-  // this class maintains a mapping between external state id's and
-  // their internal equivalents.
-  EditFstImpl() : wrapped_(new MutableFstT()) {
-    FstImpl<Arc>::SetType("edit");
-    InheritPropertiesFromWrapped();
-    data_ = std::make_shared<EditFstData<Arc, WrappedFstT, MutableFstT>>();
-  }
-
-  // Wraps the specified ExpandedFst. This constructor requires that the
-  // specified Fst is an ExpandedFst instance. This requirement is only enforced
-  // at runtime. (See below for the reason.)
-  //
-  // This library uses the pointer-to-implementation or "PIMPL" design pattern.
-  // In particular, to make it convenient to bind an implementation class to its
-  // interface, there are a pair of template "binder" classes, one for immutable
-  // and one for mutable fst's (ImplToFst and ImplToMutableFst, respectively).
-  // As it happens, the API for the ImplToMutableFst<I,F> class requires that
-  // the implementation class--the template parameter "I"--have a constructor
-  // taking a const Fst<A> reference.  Accordingly, the constructor here must
-  // perform a static_cast to the WrappedFstT type required by EditFst and
-  // therefore EditFstImpl.
-  explicit EditFstImpl(const Fst<Arc> &wrapped)
-      : wrapped_(static_cast<WrappedFstT *>(wrapped.Copy())) {
-    FstImpl<Arc>::SetType("edit");
-    data_ = std::make_shared<EditFstData<Arc, WrappedFstT, MutableFstT>>();
-    // have edits_ inherit all properties from wrapped_
-    data_->SetEditedProperties(wrapped_->Properties(kFstProperties, false),
-                               kFstProperties);
-    InheritPropertiesFromWrapped();
-  }
-
-  // A copy constructor for this implementation class, used to implement
-  // the Copy() method of the Fst interface.
-  EditFstImpl(const EditFstImpl &impl)
-      : FstImpl<Arc>(),
-        wrapped_(static_cast<WrappedFstT *>(impl.wrapped_->Copy(true))),
-        data_(impl.data_) {
-    SetProperties(impl.Properties());
-  }
-
-  // const Fst/ExpandedFst operations, declared in the Fst and ExpandedFst
-  // interfaces
-  StateId Start() const {
-    const auto edited_start = data_->EditedStart();
-    return edited_start == kNoStateId ? wrapped_->Start() : edited_start;
-  }
-
-  Weight Final(StateId s) const { return data_->Final(s, wrapped_.get()); }
-
-  size_t NumArcs(StateId s) const { return data_->NumArcs(s, wrapped_.get()); }
-
-  size_t NumInputEpsilons(StateId s) const {
-    return data_->NumInputEpsilons(s, wrapped_.get());
-  }
-
-  size_t NumOutputEpsilons(StateId s) const {
-    return data_->NumOutputEpsilons(s, wrapped_.get());
-  }
-
-  StateId NumStates() const {
-    return wrapped_->NumStates() + data_->NumNewStates();
-  }
-
-  static EditFstImpl<Arc, WrappedFstT, MutableFstT> *Read(
-      std::istream &strm, const FstReadOptions &opts);
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
-    FstHeader hdr;
-    hdr.SetStart(Start());
-    hdr.SetNumStates(NumStates());
-    FstWriteOptions header_opts(opts);
-    // Allows the contained FST to hold any symbols.
-    header_opts.write_isymbols = false;
-    header_opts.write_osymbols = false;
-    WriteHeader(strm, header_opts, kFileVersion, &hdr);
-    // First, serializes the wrapped FST to stream.
-    FstWriteOptions wrapped_opts(opts);
-    // Forcse writing the contained header.
-    wrapped_opts.write_header = true;
-    wrapped_->Write(strm, wrapped_opts);
-    data_->Write(strm, opts);
-    strm.flush();
-    if (!strm) {
-      LOG(ERROR) << "EditFst::Write: Write failed: " << opts.source;
-      return false;
-    }
-    return true;
-  }
-
-  // Sets the start state for this FST.
-  void SetStart(StateId s) {
-    MutateCheck();
-    data_->SetStart(s);
-    SetProperties(SetStartProperties(FstImpl<Arc>::Properties()));
-  }
-
-  // Sets the final state for this fst.
-  void SetFinal(StateId s, Weight weight) {
-    MutateCheck();
-    Weight old_weight = data_->SetFinal(s, weight, wrapped_.get());
-    SetProperties(
-        SetFinalProperties(FstImpl<Arc>::Properties(), old_weight, weight));
-  }
-
-  // Adds a new state to this fst, initially with no arcs.
-  StateId AddState() {
-    MutateCheck();
-    SetProperties(AddStateProperties(FstImpl<Arc>::Properties()));
-    return data_->AddState(NumStates());
-  }
-
-  // Adds the specified arc to the specified state of this fst.
-  void AddArc(StateId s, const Arc &arc) {
-    MutateCheck();
-    const auto *prev_arc = data_->AddArc(s, arc, wrapped_.get());
-    SetProperties(
-        AddArcProperties(FstImpl<Arc>::Properties(), s, arc, prev_arc));
-  }
-
-  void DeleteStates(const std::vector<StateId> &dstates) {
-    FSTERROR() << ": EditFstImpl::DeleteStates(const std::vector<StateId>&): "
-               << " not implemented";
-    SetProperties(kError, kError);
-  }
-
-  // Deletes all states in this fst.
-  void DeleteStates();
-
-  // Removes all but the first n outgoing arcs of the specified state.
-  void DeleteArcs(StateId s, size_t n) {
-    MutateCheck();
-    data_->DeleteArcs(s, n, wrapped_.get());
-    SetProperties(DeleteArcsProperties(FstImpl<Arc>::Properties()));
-  }
-
-  // Removes all outgoing arcs from the specified state.
-  void DeleteArcs(StateId s) {
-    MutateCheck();
-    data_->DeleteArcs(s, wrapped_.get());
-    SetProperties(DeleteArcsProperties(FstImpl<Arc>::Properties()));
-  }
-
-  void ReserveStates(StateId s) {}
-
-  void ReserveArcs(StateId s, size_t n) {}
-
-  // Ends non-const MutableFst operations.
-
-  // Provides information for the generic state iterator.
-  void InitStateIterator(StateIteratorData<Arc> *data) const {
-    data->base = nullptr;
-    data->nstates = NumStates();
-  }
-
-  // Provides information for the generic arc iterator.
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const {
-    data_->InitArcIterator(s, data, wrapped_.get());
-  }
-
-  // Provides information for the generic mutable arc iterator.
-  void InitMutableArcIterator(StateId s, MutableArcIteratorData<Arc> *data) {
-    MutateCheck();
-    data_->InitMutableArcIterator(s, data, wrapped_.get());
-  }
-
- private:
-  // Properties always true of this FST class.
-  static constexpr uint64 kStaticProperties = kExpanded | kMutable;
-  // Current file format version.
-  static constexpr int kFileVersion = 2;
-  // Minimum file format version supported
-  static constexpr int kMinFileVersion = 2;
-
-  // Causes this FST to inherit all the properties from its wrapped FST, except
-  // for the two properties that always apply to EditFst instances: kExpanded
-  // and kMutable.
-  void InheritPropertiesFromWrapped() {
-    SetProperties(wrapped_->Properties(kCopyProperties, false) |
-                  kStaticProperties);
-    SetInputSymbols(wrapped_->InputSymbols());
-    SetOutputSymbols(wrapped_->OutputSymbols());
-  }
-
-  // This method ensures that any operations that alter the mutable data
-  // portion of this EditFstImpl cause the data_ member to be copied when its
-  // reference count is greater than 1.  Note that this method is distinct from
-  // MutableFst::Mutate, which gets invoked whenever one of the basic mutation
-  // methods defined in MutableFst is invoked, such as SetInputSymbols.
-  // The MutateCheck here in EditFstImpl is invoked whenever one of the
-  // mutating methods specifically related to the types of edits provided
-  // by EditFst is performed, such as changing an arc of an existing state
-  // of the wrapped fst via a MutableArcIterator, or adding a new state via
-  // AddState().
-  void MutateCheck() {
-    if (!data_.unique()) {
-      data_ =
-          std::make_shared<EditFstData<Arc, WrappedFstT, MutableFstT>>(*data_);
-    }
-  }
-
-  // The FST that this FST wraps. The purpose of this class is to enable
-  // non-destructive edits on this wrapped FST.
-  std::unique_ptr<const WrappedFstT> wrapped_;
-  // The mutable data for this EditFst instance, with delegates for all the
-  // methods that can mutate data.
-  std::shared_ptr<EditFstData<Arc, WrappedFstT, MutableFstT>> data_;
-};
-
-template <typename Arc, typename WrappedFstT, typename MutableFstT>
-constexpr uint64 EditFstImpl<Arc, WrappedFstT, MutableFstT>::kStaticProperties;
-
-template <typename Arc, typename WrappedFstT, typename MutableFstT>
-constexpr int EditFstImpl<Arc, WrappedFstT, MutableFstT>::kFileVersion;
-
-template <typename Arc, typename WrappedFstT, typename MutableFstT>
-constexpr int EditFstImpl<Arc, WrappedFstT, MutableFstT>::kMinFileVersion;
-
-template <typename Arc, typename WrappedFstT, typename MutableFstT>
-inline void EditFstImpl<Arc, WrappedFstT, MutableFstT>::DeleteStates() {
-  data_->DeleteStates();
-  // we are deleting all states, so just forget about pointer to wrapped_
-  // and do what default constructor does: set wrapped_ to a new VectorFst
-  wrapped_.reset(new MutableFstT());
-  const auto new_props =
-      DeleteAllStatesProperties(FstImpl<Arc>::Properties(), kStaticProperties);
-  FstImpl<Arc>::SetProperties(new_props);
-}
-
-template <typename Arc, typename WrappedFstT, typename MutableFstT>
-EditFstImpl<Arc, WrappedFstT, MutableFstT> *
-EditFstImpl<Arc, WrappedFstT, MutableFstT>::Read(std::istream &strm,
-                                                 const FstReadOptions &opts) {
-  auto *impl = new EditFstImpl();
-  FstHeader hdr;
-  if (!impl->ReadHeader(strm, opts, kMinFileVersion, &hdr)) return nullptr;
-  impl->SetStart(hdr.Start());
-  // Reads in wrapped FST.
-  FstReadOptions wrapped_opts(opts);
-  // Contained header was written out, so reads it in too.
-  wrapped_opts.header = nullptr;
-  std::unique_ptr<Fst<Arc>> wrapped_fst(Fst<Arc>::Read(strm, wrapped_opts));
-  if (!wrapped_fst) return nullptr;
-  impl->wrapped_.reset(static_cast<WrappedFstT *>(wrapped_fst.release()));
-  impl->data_ = std::shared_ptr<EditFstData<Arc, WrappedFstT, MutableFstT>>(
-      EditFstData<Arc, WrappedFstT, MutableFstT>::Read(strm, opts));
-  if (!impl->data_) return nullptr;
-  return impl;
-}
-
-}  // namespace internal
-
-// Concrete, editable FST.  This class attaches interface to implementation.
-template <typename A, typename WrappedFstT = ExpandedFst<A>,
-          typename MutableFstT = VectorFst<A>>
-class EditFst : public ImplToMutableFst<
-                    internal::EditFstImpl<A, WrappedFstT, MutableFstT>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  using Impl = internal::EditFstImpl<Arc, WrappedFstT, MutableFstT>;
-
-  friend class MutableArcIterator<EditFst<Arc, WrappedFstT, MutableFstT>>;
-
-  EditFst() : ImplToMutableFst<Impl>(std::make_shared<Impl>()) {}
-
-  explicit EditFst(const Fst<Arc> &fst)
-      : ImplToMutableFst<Impl>(std::make_shared<Impl>(fst)) {}
-
-  explicit EditFst(const WrappedFstT &fst)
-      : ImplToMutableFst<Impl>(std::make_shared<Impl>(fst)) {}
-
-  // See Fst<>::Copy() for doc.
-  EditFst(const EditFst<Arc, WrappedFstT, MutableFstT> &fst, bool safe = false)
-      : ImplToMutableFst<Impl>(fst, safe) {}
-
-  ~EditFst() override {}
-
-  // Gets a copy of this EditFst. See Fst<>::Copy() for further doc.
-  EditFst<Arc, WrappedFstT, MutableFstT> *Copy(
-      bool safe = false) const override {
-    return new EditFst<Arc, WrappedFstT, MutableFstT>(*this, safe);
-  }
-
-  EditFst<Arc, WrappedFstT, MutableFstT> &operator=(
-      const EditFst<Arc, WrappedFstT, MutableFstT> &fst) {
-    SetImpl(fst.GetSharedImpl());
-    return *this;
-  }
-
-  EditFst<Arc, WrappedFstT, MutableFstT> &operator=(
-      const Fst<Arc> &fst) override {
-    SetImpl(std::make_shared<Impl>(fst));
-    return *this;
-  }
-
-  // Reads an EditFst from an input stream, returning nullptr on error.
-  static EditFst<Arc, WrappedFstT, MutableFstT> *Read(
-      std::istream &strm, const FstReadOptions &opts) {
-    auto *impl = Impl::Read(strm, opts);
-    return impl ? new EditFst<Arc>(std::shared_ptr<Impl>(impl)) : nullptr;
-  }
-
-  // Reads an EditFst from a file, returning nullptr on error. If the filename
-  // argument is an empty string, it reads from standard input.
-  static EditFst<Arc, WrappedFstT, MutableFstT> *Read(const string &filename) {
-    auto *impl = ImplToExpandedFst<Impl, MutableFst<Arc>>::Read(filename);
-    return impl ? new EditFst<Arc, WrappedFstT, MutableFstT>(
-                      std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return GetImpl()->Write(strm, opts);
-  }
-
-  bool Write(const string &filename) const override {
-    return Fst<Arc>::WriteFile(filename);
-  }
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const override {
-    GetImpl()->InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetImpl()->InitArcIterator(s, data);
-  }
-
-  void InitMutableArcIterator(StateId s,
-                              MutableArcIteratorData<A> *data) override {
-    GetMutableImpl()->InitMutableArcIterator(s, data);
-  }
-
- private:
-  explicit EditFst(std::shared_ptr<Impl> impl) : ImplToMutableFst<Impl>(impl) {}
-
-  using ImplToFst<Impl, MutableFst<Arc>>::GetImpl;
-  using ImplToFst<Impl, MutableFst<Arc>>::GetMutableImpl;
-  using ImplToFst<Impl, MutableFst<Arc>>::SetImpl;
-};
-
-}  // namespace fst
-
-#endif  // FST_EDIT_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/encode.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/encode.h
deleted file mode 100644
index f251bbfc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/encode.h
+++ /dev/null
@@ -1,556 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to encode and decode an FST.
-
-#ifndef FST_ENCODE_H_
-#define FST_ENCODE_H_
-
-#include <iostream>
-#include <memory>
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-#include <fstream>
-
-#include <fst/arc-map.h>
-#include <fst/rmfinalepsilon.h>
-
-
-namespace fst {
-
-enum EncodeType { ENCODE = 1, DECODE = 2 };
-
-static constexpr uint32 kEncodeLabels = 0x0001;
-static constexpr uint32 kEncodeWeights = 0x0002;
-static constexpr uint32 kEncodeFlags = 0x0003;
-
-namespace internal {
-
-static constexpr uint32 kEncodeHasISymbols = 0x0004;
-static constexpr uint32 kEncodeHasOSymbols = 0x0008;
-
-// Identifies stream data as an encode table (and its endianity)
-static const int32 kEncodeMagicNumber = 2129983209;
-
-// The following class encapsulates implementation details for the encoding and
-// decoding of label/weight tuples used for encoding and decoding of FSTs. The
-// EncodeTable is bidirectional. I.e, it stores both the Tuple of encode labels
-// and weights to a unique label, and the reverse.
-template <class Arc>
-class EncodeTable {
- public:
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-
-  // Encoded data consists of arc input/output labels and arc weight.
-  struct Tuple {
-    Tuple() {}
-
-    Tuple(Label ilabel_, Label olabel_, Weight weight_)
-        : ilabel(ilabel_), olabel(olabel_), weight(std::move(weight_)) {}
-
-    Tuple(const Tuple &tuple)
-        : ilabel(tuple.ilabel),
-          olabel(tuple.olabel),
-          weight(std::move(tuple.weight)) {}
-
-    Label ilabel;
-    Label olabel;
-    Weight weight;
-  };
-
-  // Comparison object for hashing EncodeTable Tuple(s).
-  class TupleEqual {
-   public:
-    bool operator()(const Tuple *x, const Tuple *y) const {
-      return (x->ilabel == y->ilabel && x->olabel == y->olabel &&
-              x->weight == y->weight);
-    }
-  };
-
-  // Hash function for EncodeTabe Tuples. Based on the encode flags
-  // we either hash the labels, weights or combination of them.
-  class TupleKey {
-   public:
-    TupleKey() : encode_flags_(kEncodeLabels | kEncodeWeights) {}
-
-    TupleKey(const TupleKey &key) : encode_flags_(key.encode_flags_) {}
-
-    explicit TupleKey(uint32 encode_flags) : encode_flags_(encode_flags) {}
-
-    size_t operator()(const Tuple *x) const {
-      size_t hash = x->ilabel;
-      static constexpr int lshift = 5;
-      static constexpr int rshift = CHAR_BIT * sizeof(size_t) - 5;
-      if (encode_flags_ & kEncodeLabels) {
-        hash = hash << lshift ^ hash >> rshift ^ x->olabel;
-      }
-      if (encode_flags_ & kEncodeWeights) {
-        hash = hash << lshift ^ hash >> rshift ^ x->weight.Hash();
-      }
-      return hash;
-    }
-
-   private:
-    int32 encode_flags_;
-  };
-
-  explicit EncodeTable(uint32 encode_flags)
-      : flags_(encode_flags), encode_hash_(1024, TupleKey(encode_flags)) {}
-
-  using EncodeHash = std::unordered_map<const Tuple *, Label, TupleKey,
-                                        TupleEqual>;
-
-  // Given an arc, encodes either input/output labels or input/costs or both.
-  Label Encode(const Arc &arc) {
-    std::unique_ptr<Tuple> tuple(
-        new Tuple(arc.ilabel, flags_ & kEncodeLabels ? arc.olabel : 0,
-                  flags_ & kEncodeWeights ? arc.weight : Weight::One()));
-    auto insert_result = encode_hash_.insert(
-        std::make_pair(tuple.get(), encode_tuples_.size() + 1));
-    if (insert_result.second) encode_tuples_.push_back(std::move(tuple));
-    return insert_result.first->second;
-  }
-
-  // Given an arc, looks up its encoded label or returns kNoLabel if not found.
-  Label GetLabel(const Arc &arc) const {
-    const Tuple tuple(arc.ilabel, flags_ & kEncodeLabels ? arc.olabel : 0,
-                      flags_ & kEncodeWeights ? arc.weight : Weight::One());
-    auto it = encode_hash_.find(&tuple);
-    return (it == encode_hash_.end()) ?  kNoLabel : it->second;
-  }
-
-  // Given an encoded arc label, decodes back to input/output labels and costs.
-  const Tuple *Decode(Label key) const {
-    if (key < 1 || key > encode_tuples_.size()) {
-      LOG(ERROR) << "EncodeTable::Decode: Unknown decode key: " << key;
-      return nullptr;
-    }
-    return encode_tuples_[key - 1].get();
-  }
-
-  size_t Size() const { return encode_tuples_.size(); }
-
-  bool Write(std::ostream &strm, const string &source) const;
-
-  static EncodeTable<Arc> *Read(std::istream &strm, const string &source);
-
-  uint32 Flags() const { return flags_ & kEncodeFlags; }
-
-  const SymbolTable *InputSymbols() const { return isymbols_.get(); }
-
-  const SymbolTable *OutputSymbols() const { return osymbols_.get(); }
-
-  void SetInputSymbols(const SymbolTable *syms) {
-    if (syms) {
-      isymbols_.reset(syms->Copy());
-      flags_ |= kEncodeHasISymbols;
-    } else {
-      isymbols_.reset();
-      flags_ &= ~kEncodeHasISymbols;
-    }
-  }
-
-  void SetOutputSymbols(const SymbolTable *syms) {
-    if (syms) {
-      osymbols_.reset(syms->Copy());
-      flags_ |= kEncodeHasOSymbols;
-    } else {
-      osymbols_.reset();
-      flags_ &= ~kEncodeHasOSymbols;
-    }
-  }
-
- private:
-  uint32 flags_;
-  std::vector<std::unique_ptr<Tuple>> encode_tuples_;
-  EncodeHash encode_hash_;
-  std::unique_ptr<SymbolTable> isymbols_;  // Pre-encoded input symbol table.
-  std::unique_ptr<SymbolTable> osymbols_;  // Pre-encoded output symbol table.
-
-  EncodeTable(const EncodeTable &) = delete;
-  EncodeTable &operator=(const EncodeTable &) = delete;
-};
-
-template <class Arc>
-bool EncodeTable<Arc>::Write(std::ostream &strm,
-                                  const string &source) const {
-  WriteType(strm, kEncodeMagicNumber);
-  WriteType(strm, flags_);
-  const int64 size = encode_tuples_.size();
-  WriteType(strm, size);
-  for (const auto &tuple : encode_tuples_) {
-    WriteType(strm, tuple->ilabel);
-    WriteType(strm, tuple->olabel);
-    tuple->weight.Write(strm);
-  }
-  if (flags_ & kEncodeHasISymbols) isymbols_->Write(strm);
-  if (flags_ & kEncodeHasOSymbols) osymbols_->Write(strm);
-  strm.flush();
-  if (!strm) {
-    LOG(ERROR) << "EncodeTable::Write: Write failed: " << source;
-    return false;
-  }
-  return true;
-}
-
-template <class Arc>
-EncodeTable<Arc> *EncodeTable<Arc>::Read(std::istream &strm,
-                                         const string &source) {
-  int32 magic_number = 0;
-  ReadType(strm, &magic_number);
-  if (magic_number != kEncodeMagicNumber) {
-    LOG(ERROR) << "EncodeTable::Read: Bad encode table header: " << source;
-    return nullptr;
-  }
-  uint32 flags;
-  ReadType(strm, &flags);
-  int64 size;
-  ReadType(strm, &size);
-  if (!strm) {
-    LOG(ERROR) << "EncodeTable::Read: Read failed: " << source;
-    return nullptr;
-  }
-  std::unique_ptr<EncodeTable<Arc>> table(new EncodeTable<Arc>(flags));
-  for (int64 i = 0; i < size; ++i) {
-    std::unique_ptr<Tuple> tuple(new Tuple());
-    ReadType(strm, &tuple->ilabel);
-    ReadType(strm, &tuple->olabel);
-    tuple->weight.Read(strm);
-    if (!strm) {
-      LOG(ERROR) << "EncodeTable::Read: Read failed: " << source;
-      return nullptr;
-    }
-    table->encode_tuples_.push_back(std::move(tuple));
-    table->encode_hash_[table->encode_tuples_.back().get()] =
-        table->encode_tuples_.size();
-  }
-  if (flags & kEncodeHasISymbols) {
-    table->isymbols_.reset(SymbolTable::Read(strm, source));
-  }
-  if (flags & kEncodeHasOSymbols) {
-    table->osymbols_.reset(SymbolTable::Read(strm, source));
-  }
-  return table.release();
-}
-
-}  // namespace internal
-
-// A mapper to encode/decode weighted transducers. Encoding of an FST is used
-// for performing classical determinization or minimization on a weighted
-// transducer viewing it as an unweighted acceptor over encoded labels.
-//
-// The mapper stores the encoding in a local hash table (EncodeTable). This
-// table is shared (and reference-counted) between the encoder and decoder.
-// A decoder has read-only access to the EncodeTable.
-//
-// The EncodeMapper allows on the fly encoding of the machine. As the
-// EncodeTable is generated the same table may by used to decode the machine
-// on the fly. For example in the following sequence of operations
-//
-//  Encode -> Determinize -> Decode
-//
-// we will use the encoding table generated during the encode step in the
-// decode, even though the encoding is not complete.
-template <class Arc>
-class EncodeMapper {
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-
- public:
-  EncodeMapper(uint32 flags, EncodeType type)
-      : flags_(flags),
-        type_(type),
-        table_(std::make_shared<internal::EncodeTable<Arc>>(flags)),
-        error_(false) {}
-
-  EncodeMapper(const EncodeMapper &mapper)
-      : flags_(mapper.flags_),
-        type_(mapper.type_),
-        table_(mapper.table_),
-        error_(false) {}
-
-  // Copy constructor but setting the type, typically to DECODE.
-  EncodeMapper(const EncodeMapper &mapper, EncodeType type)
-      : flags_(mapper.flags_),
-        type_(type),
-        table_(mapper.table_),
-        error_(mapper.error_) {}
-
-  Arc operator()(const Arc &arc);
-
-  MapFinalAction FinalAction() const {
-    return (type_ == ENCODE && (flags_ & kEncodeWeights))
-               ? MAP_REQUIRE_SUPERFINAL
-               : MAP_NO_SUPERFINAL;
-  }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_CLEAR_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 inprops) {
-    uint64 outprops = inprops;
-    if (error_) outprops |= kError;
-    uint64 mask = kFstProperties;
-    if (flags_ & kEncodeLabels) {
-      mask &= kILabelInvariantProperties & kOLabelInvariantProperties;
-    }
-    if (flags_ & kEncodeWeights) {
-      mask &= kILabelInvariantProperties & kWeightInvariantProperties &
-              (type_ == ENCODE ? kAddSuperFinalProperties
-                               : kRmSuperFinalProperties);
-    }
-    return outprops & mask;
-  }
-
-  uint32 Flags() const { return flags_; }
-
-  EncodeType Type() const { return type_; }
-
-  bool Write(std::ostream &strm, const string &source) const {
-    return table_->Write(strm, source);
-  }
-
-  bool Write(const string &filename) const {
-    std::ofstream strm(filename,
-                             std::ios_base::out | std::ios_base::binary);
-    if (!strm) {
-      LOG(ERROR) << "EncodeMap: Can't open file: " << filename;
-      return false;
-    }
-    return Write(strm, filename);
-  }
-
-  static EncodeMapper<Arc> *Read(std::istream &strm, const string &source,
-                               EncodeType type = ENCODE) {
-    auto *table = internal::EncodeTable<Arc>::Read(strm, source);
-    return table ? new EncodeMapper(table->Flags(), type, table) : nullptr;
-  }
-
-  static EncodeMapper<Arc> *Read(const string &filename,
-                                 EncodeType type = ENCODE) {
-    std::ifstream strm(filename,
-                            std::ios_base::in | std::ios_base::binary);
-    if (!strm) {
-      LOG(ERROR) << "EncodeMap: Can't open file: " << filename;
-      return nullptr;
-    }
-    return Read(strm, filename, type);
-  }
-
-  const SymbolTable *InputSymbols() const { return table_->InputSymbols(); }
-
-  const SymbolTable *OutputSymbols() const { return table_->OutputSymbols(); }
-
-  void SetInputSymbols(const SymbolTable *syms) {
-    table_->SetInputSymbols(syms);
-  }
-
-  void SetOutputSymbols(const SymbolTable *syms) {
-    table_->SetOutputSymbols(syms);
-  }
-
- private:
-  uint32 flags_;
-  EncodeType type_;
-  std::shared_ptr<internal::EncodeTable<Arc>> table_;
-  bool error_;
-
-  explicit EncodeMapper(uint32 flags, EncodeType type,
-                        internal::EncodeTable<Arc> *table)
-      : flags_(flags), type_(type), table_(table), error_(false) {}
-
-  EncodeMapper &operator=(const EncodeMapper &) = delete;
-};
-
-template <class Arc>
-Arc EncodeMapper<Arc>::operator()(const Arc &arc) {
-  if (type_ == ENCODE) {
-    if ((arc.nextstate == kNoStateId && !(flags_ & kEncodeWeights)) ||
-        (arc.nextstate == kNoStateId && (flags_ & kEncodeWeights) &&
-         arc.weight == Weight::Zero())) {
-      return arc;
-    } else {
-      const auto label = table_->Encode(arc);
-      return Arc(label, flags_ & kEncodeLabels ? label : arc.olabel,
-                 flags_ & kEncodeWeights ? Weight::One() : arc.weight,
-                 arc.nextstate);
-    }
-  } else {  // type_ == DECODE
-    if (arc.nextstate == kNoStateId) {
-      return arc;
-    } else {
-      if (arc.ilabel == 0) return arc;
-      if (flags_ & kEncodeLabels && arc.ilabel != arc.olabel) {
-        FSTERROR() << "EncodeMapper: Label-encoded arc has different "
-                      "input and output labels";
-        error_ = true;
-      }
-      if (flags_ & kEncodeWeights && arc.weight != Weight::One()) {
-        FSTERROR() << "EncodeMapper: Weight-encoded arc has non-trivial weight";
-        error_ = true;
-      }
-      const auto tuple = table_->Decode(arc.ilabel);
-      if (!tuple) {
-        FSTERROR() << "EncodeMapper: Decode failed";
-        error_ = true;
-        return Arc(kNoLabel, kNoLabel, Weight::NoWeight(), arc.nextstate);
-      } else {
-        return Arc(tuple->ilabel,
-                   flags_ & kEncodeLabels ? tuple->olabel : arc.olabel,
-                   flags_ & kEncodeWeights ? tuple->weight : arc.weight,
-                   arc.nextstate);
-      }
-    }
-  }
-}
-
-// Complexity: O(E + V).
-template <class Arc>
-inline void Encode(MutableFst<Arc> *fst, EncodeMapper<Arc> *mapper) {
-  mapper->SetInputSymbols(fst->InputSymbols());
-  mapper->SetOutputSymbols(fst->OutputSymbols());
-  ArcMap(fst, mapper);
-}
-
-template <class Arc>
-inline void Decode(MutableFst<Arc> *fst, const EncodeMapper<Arc> &mapper) {
-  ArcMap(fst, EncodeMapper<Arc>(mapper, DECODE));
-  RmFinalEpsilon(fst);
-  fst->SetInputSymbols(mapper.InputSymbols());
-  fst->SetOutputSymbols(mapper.OutputSymbols());
-}
-
-// On-the-fly encoding of an input FST.
-//
-// Complexity:
-//
-//   Construction: O(1)
-//   Traversal: O(e + v)
-//
-// where e is the number of arcs visited and v is the number of states visited.
-// Constant time and space to visit an input state or arc is assumed and
-// exclusive of caching.
-template <class Arc>
-class EncodeFst : public ArcMapFst<Arc, Arc, EncodeMapper<Arc>> {
- public:
-  using Mapper = EncodeMapper<Arc>;
-  using Impl = internal::ArcMapFstImpl<Arc, Arc, Mapper>;
-
-  EncodeFst(const Fst<Arc> &fst, Mapper *encoder)
-      : ArcMapFst<Arc, Arc, Mapper>(fst, encoder, ArcMapFstOptions()) {
-    encoder->SetInputSymbols(fst.InputSymbols());
-    encoder->SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  EncodeFst(const Fst<Arc> &fst, const Mapper &encoder)
-      : ArcMapFst<Arc, Arc, Mapper>(fst, encoder, ArcMapFstOptions()) {}
-
-  // See Fst<>::Copy() for doc.
-  EncodeFst(const EncodeFst<Arc> &fst, bool copy = false)
-      : ArcMapFst<Arc, Arc, Mapper>(fst, copy) {}
-
-  // Makes a copy of this EncodeFst. See Fst<>::Copy() for further doc.
-  EncodeFst<Arc> *Copy(bool safe = false) const override {
-    if (safe) {
-      FSTERROR() << "EncodeFst::Copy(true): Not allowed";
-      GetImpl()->SetProperties(kError, kError);
-    }
-    return new EncodeFst(*this);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-};
-
-// On-the-fly decoding of an input FST.
-//
-// Complexity:
-//
-//   Construction: O(1).
-//   Traversal: O(e + v)
-//
-// Constant time and space to visit an input state or arc is assumed and
-// exclusive of caching.
-template <class Arc>
-class DecodeFst : public ArcMapFst<Arc, Arc, EncodeMapper<Arc>> {
- public:
-  using Mapper = EncodeMapper<Arc>;
-  using Impl = internal::ArcMapFstImpl<Arc, Arc, Mapper>;
-  using ImplToFst<Impl>::GetImpl;
-
-  DecodeFst(const Fst<Arc> &fst, const Mapper &encoder)
-      : ArcMapFst<Arc, Arc, Mapper>(fst, Mapper(encoder, DECODE),
-                                    ArcMapFstOptions()) {
-    GetMutableImpl()->SetInputSymbols(encoder.InputSymbols());
-    GetMutableImpl()->SetOutputSymbols(encoder.OutputSymbols());
-  }
-
-  // See Fst<>::Copy() for doc.
-  DecodeFst(const DecodeFst<Arc> &fst, bool safe = false)
-      : ArcMapFst<Arc, Arc, Mapper>(fst, safe) {}
-
-  // Makes a copy of this DecodeFst. See Fst<>::Copy() for further doc.
-  DecodeFst<Arc> *Copy(bool safe = false) const override {
-    return new DecodeFst(*this, safe);
-  }
-
- private:
-  using ImplToFst<Impl>::GetMutableImpl;
-};
-
-// Specialization for EncodeFst.
-template <class Arc>
-class StateIterator<EncodeFst<Arc>>
-    : public StateIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>> {
- public:
-  explicit StateIterator(const EncodeFst<Arc> &fst)
-      : StateIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>>(fst) {}
-};
-
-// Specialization for EncodeFst.
-template <class Arc>
-class ArcIterator<EncodeFst<Arc>>
-    : public ArcIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>> {
- public:
-  ArcIterator(const EncodeFst<Arc> &fst, typename Arc::StateId s)
-      : ArcIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>>(fst, s) {}
-};
-
-// Specialization for DecodeFst.
-template <class Arc>
-class StateIterator<DecodeFst<Arc>>
-    : public StateIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>> {
- public:
-  explicit StateIterator(const DecodeFst<Arc> &fst)
-      : StateIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>>(fst) {}
-};
-
-// Specialization for DecodeFst.
-template <class Arc>
-class ArcIterator<DecodeFst<Arc>>
-    : public ArcIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>> {
- public:
-  ArcIterator(const DecodeFst<Arc> &fst, typename Arc::StateId s)
-      : ArcIterator<ArcMapFst<Arc, Arc, EncodeMapper<Arc>>>(fst, s) {}
-};
-
-// Useful aliases when using StdArc.
-
-using StdEncodeFst = EncodeFst<StdArc>;
-
-using StdDecodeFst = DecodeFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_ENCODE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/epsnormalize.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/epsnormalize.h
deleted file mode 100644
index 3496edda..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/epsnormalize.h
+++ /dev/null
@@ -1,58 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function that implements epsilon-normalization.
-
-#ifndef FST_EPSNORMALIZE_H_
-#define FST_EPSNORMALIZE_H_
-
-
-#include <fst/arc-map.h>
-#include <fst/factor-weight.h>
-#include <fst/invert.h>
-#include <fst/rmepsilon.h>
-
-
-namespace fst {
-
-enum EpsNormalizeType { EPS_NORM_INPUT, EPS_NORM_OUTPUT };
-
-// Returns an equivalent FST that is epsilon-normalized. An acceptor is
-// epsilon-normalized if it is epsilon-removed. A transducer is input
-// epsilon-normalized if additionally if on each path any epsilon input
-// label follows all non-epsilon input labels. Output epsilon-normalized
-// is defined similarly.
-//
-// For more information, see:
-//
-// Mohri, M. 2002. Generic epsilon-removal and input epsilon-normalization
-// algorithms for weighted transducers. International Journal of Computer
-// Science, 13(1): 129-143, 2002.
-template <class Arc>
-void EpsNormalize(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-                  EpsNormalizeType type = EPS_NORM_INPUT) {
-  EpsNormalize<Arc, GALLIC>(ifst, ofst, type);
-}
-
-// Same as above, expect allows specifying explicitely the gallic weight type.
-template <class Arc, GallicType G>
-void EpsNormalize(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-                  EpsNormalizeType type) {
-  VectorFst<GallicArc<Arc, G>> gfst;
-  if (type == EPS_NORM_INPUT) {
-    ArcMap(ifst, &gfst, ToGallicMapper<Arc, G>());
-  } else {  // type == EPS_NORM_OUTPUT
-    ArcMap(InvertFst<Arc>(ifst), &gfst, ToGallicMapper<Arc, G>());
-  }
-  RmEpsilon(&gfst);
-  FactorWeightFst<GallicArc<Arc, G>,
-                  GallicFactor<typename Arc::Label, typename Arc::Weight, G>>
-      fwfst(gfst);
-  ArcMap(fwfst, ofst, FromGallicMapper<Arc, G>());
-  ofst->SetOutputSymbols(ifst.OutputSymbols());
-  if (type == EPS_NORM_OUTPUT) Invert(ofst);
-}
-
-}  // namespace fst
-
-#endif  // FST_EPSNORMALIZE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/equal.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/equal.h
deleted file mode 100644
index 6a035c93..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/equal.h
+++ /dev/null
@@ -1,139 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function to test equality of two FSTs.
-
-#ifndef FST_EQUAL_H_
-#define FST_EQUAL_H_
-
-#include <fst/log.h>
-
-#include <fst/fst.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-constexpr uint32 kEqualFsts = 0x0001;
-constexpr uint32 kEqualFstTypes = 0x0002;
-constexpr uint32 kEqualCompatProperties = 0x0004;
-constexpr uint32 kEqualCompatSymbols = 0x0008;
-constexpr uint32 kEqualAll =
-    kEqualFsts | kEqualFstTypes | kEqualCompatProperties | kEqualCompatSymbols;
-
-// Tests if two Fsts have the same states and arcs in the same order (when
-// etype & kEqualFst).
-// Also optional checks equality of Fst types (etype & kEqualFstTypes) and
-// compatibility of stored properties (etype & kEqualCompatProperties) and
-// of symbol tables (etype & kEqualCompatSymbols).
-template <class Arc>
-bool Equal(const Fst<Arc> &fst1, const Fst<Arc> &fst2, float delta = kDelta,
-           uint32 etype = kEqualFsts) {
-  if ((etype & kEqualFstTypes) && (fst1.Type() != fst2.Type())) {
-    VLOG(1) << "Equal: Mismatched FST types (" << fst1.Type() << " != "
-            << fst2.Type() << ")";
-    return false;
-  }
-  if ((etype & kEqualCompatProperties) &&
-      !CompatProperties(fst1.Properties(kCopyProperties, false),
-                        fst2.Properties(kCopyProperties, false))) {
-    VLOG(1) << "Equal: Properties not compatible";
-    return false;
-  }
-  if (etype & kEqualCompatSymbols) {
-    if (!CompatSymbols(fst1.InputSymbols(), fst2.InputSymbols(), false)) {
-      VLOG(1) << "Equal: Input symbols not compatible";
-      return false;
-    }
-    if (!CompatSymbols(fst1.OutputSymbols(), fst2.OutputSymbols(), false)) {
-      VLOG(1) << "Equal: Output symbols not compatible";
-      return false;
-    }
-  }
-  if (!(etype & kEqualFsts)) return true;
-  if (fst1.Start() != fst2.Start()) {
-    VLOG(1) << "Equal: Mismatched start states (" << fst1.Start() << " != "
-            << fst2.Start() << ")";
-    return false;
-  }
-  StateIterator<Fst<Arc>> siter1(fst1);
-  StateIterator<Fst<Arc>> siter2(fst2);
-  while (!siter1.Done() || !siter2.Done()) {
-    if (siter1.Done() || siter2.Done()) {
-      VLOG(1) << "Equal: Mismatched number of states";
-      return false;
-    }
-    const auto s1 = siter1.Value();
-    const auto s2 = siter2.Value();
-    if (s1 != s2) {
-      VLOG(1) << "Equal: Mismatched states (" << s1 << "!= "
-              << s2 << ")";
-      return false;
-    }
-    const auto &final1 = fst1.Final(s1);
-    const auto &final2 = fst2.Final(s2);
-    if (!ApproxEqual(final1, final2, delta)) {
-      VLOG(1) << "Equal: Mismatched final weights at state " << s1
-              << " (" << final1 << " != " << final2 << ")";
-      return false;
-    }
-    ArcIterator<Fst<Arc>> aiter1(fst1, s1);
-    ArcIterator<Fst<Arc>> aiter2(fst2, s2);
-    for (auto a = 0; !aiter1.Done() || !aiter2.Done(); ++a) {
-      if (aiter1.Done() || aiter2.Done()) {
-        VLOG(1) << "Equal: Mismatched number of arcs at state " << s1;
-        return false;
-      }
-      const auto &arc1 = aiter1.Value();
-      const auto &arc2 = aiter2.Value();
-      if (arc1.ilabel != arc2.ilabel) {
-        VLOG(1) << "Equal: Mismatched arc input labels at state " << s1
-                << ", arc " << a << " (" << arc1.ilabel << " != "
-                << arc2.ilabel << ")";
-        return false;
-      } else if (arc1.olabel != arc2.olabel) {
-        VLOG(1) << "Equal: Mismatched arc output labels at state " << s1
-                << ", arc " << a << " (" << arc1.olabel << " != "
-                << arc2.olabel << ")";
-        return false;
-      } else if (!ApproxEqual(arc1.weight, arc2.weight, delta)) {
-        VLOG(1) << "Equal: Mismatched arc weights at state " << s1
-                << ", arc " << a << " (" << arc1.weight << " != "
-                << arc2.weight << ")";
-        return false;
-      } else if (arc1.nextstate != arc2.nextstate) {
-        VLOG(1) << "Equal: Mismatched next state at state " << s1
-                << ", arc " << a << " (" << arc1.nextstate << " != "
-                << arc2.nextstate << ")";
-        return false;
-      }
-      aiter1.Next();
-      aiter2.Next();
-    }
-    // Sanity checks: should never fail.
-    if (fst1.NumArcs(s1) != fst2.NumArcs(s2)) {
-      FSTERROR() << "Equal: Inconsistent arc counts at state " << s1
-                 << " (" << fst1.NumArcs(s1) << " != "
-                 << fst2.NumArcs(s2) << ")";
-      return false;
-    }
-    if (fst1.NumInputEpsilons(s1) != fst2.NumInputEpsilons(s2)) {
-      FSTERROR() << "Equal: Inconsistent input epsilon counts at state " << s1
-                 << " (" << fst1.NumInputEpsilons(s1) << " != "
-                 << fst2.NumInputEpsilons(s2) << ")";
-      return false;
-    }
-    if (fst1.NumOutputEpsilons(s1) != fst2.NumOutputEpsilons(s2)) {
-      FSTERROR() << "Equal: Inconsistent output epsilon counts at state " << s1
-                 << " (" << fst1.NumOutputEpsilons(s1) << " != "
-                 << fst2.NumOutputEpsilons(s2) << ")";
-    }
-    siter1.Next();
-    siter2.Next();
-  }
-  return true;
-}
-
-}  // namespace fst
-
-#endif  // FST_EQUAL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/equivalent.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/equivalent.h
deleted file mode 100644
index 7c7b7b36..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/equivalent.h
+++ /dev/null
@@ -1,230 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to determine the equivalence of two FSTs.
-
-#ifndef FST_EQUIVALENT_H_
-#define FST_EQUIVALENT_H_
-
-#include <algorithm>
-#include <deque>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-#include <fst/log.h>
-
-#include <fst/encode.h>
-#include <fst/push.h>
-#include <fst/union-find.h>
-#include <fst/vector-fst.h>
-
-
-namespace fst {
-namespace internal {
-
-// Traits-like struct holding utility functions/typedefs/constants for
-// the equivalence algorithm.
-//
-// Encoding device: in order to make the statesets of the two acceptors
-// disjoint, we map Arc::StateId on the type MappedId. The states of
-// the first acceptor are mapped on odd numbers (s -> 2s + 1), and
-// those of the second one on even numbers (s -> 2s + 2). The number 0
-// is reserved for an implicit (non-final) dead state (required for
-// the correct treatment of non-coaccessible states; kNoStateId is mapped to
-// kDeadState for both acceptors). The union-find algorithm operates on the
-// mapped IDs.
-template <class Arc>
-struct EquivalenceUtil {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using MappedId = StateId;  // ID for an equivalence class.
-
-  // MappedId for an implicit dead state.
-  static constexpr MappedId kDeadState = 0;
-
-  // MappedId for lookup failure.
-  static constexpr MappedId kInvalidId = -1;
-
-  // Maps state ID to the representative of the corresponding
-  // equivalence class. The parameter 'which_fst' takes the values 1
-  // and 2, identifying the input FST.
-  static MappedId MapState(StateId s, int32 which_fst) {
-    return (kNoStateId == s) ? kDeadState
-                             : (static_cast<MappedId>(s) << 1) + which_fst;
-  }
-
-  // Maps set ID to State ID.
-  static StateId UnMapState(MappedId id) {
-    return static_cast<StateId>((--id) >> 1);
-  }
-
-  // Convenience function: checks if state with MappedId s is final in
-  // acceptor fa.
-  static bool IsFinal(const Fst<Arc> &fa, MappedId s) {
-    return (kDeadState == s) ? false
-                             : (fa.Final(UnMapState(s)) != Weight::Zero());
-  }
-  // Convenience function: returns the representative of ID in sets,
-  // creating a new set if needed.
-  static MappedId FindSet(UnionFind<MappedId> *sets, MappedId id) {
-    const auto repr = sets->FindSet(id);
-    if (repr != kInvalidId) {
-      return repr;
-    } else {
-      sets->MakeSet(id);
-      return id;
-    }
-  }
-};
-
-template <class Arc>
-constexpr
-    typename EquivalenceUtil<Arc>::MappedId EquivalenceUtil<Arc>::kDeadState;
-
-template <class Arc>
-constexpr
-    typename EquivalenceUtil<Arc>::MappedId EquivalenceUtil<Arc>::kInvalidId;
-
-}  // namespace internal
-
-// Equivalence checking algorithm: determines if the two FSTs fst1 and fst2
-// are equivalent. The input FSTs must be deterministic input-side epsilon-free
-// acceptors, unweighted or with weights over a left semiring. Two acceptors are
-// considered equivalent if they accept exactly the same set of strings (with
-// the same weights).
-//
-// The algorithm (cf. Aho, Hopcroft and Ullman, "The Design and Analysis of
-// Computer Programs") successively constructs sets of states that can be
-// reached by the same prefixes, starting with a set containing the start states
-// of both acceptors. A disjoint tree forest (the union-find algorithm) is used
-// to represent the sets of states. The algorithm returns false if one of the
-// constructed sets contains both final and non-final states. Returns an
-// optional error value (useful when FLAGS_error_fatal = false).
-//
-// Complexity:
-//
-// Quasi-linear, i.e., O(n G(n)), where
-//
-//   n = |S1| + |S2| is the number of states in both acceptors
-//
-//   G(n) is a very slowly growing function that can be approximated
-//        by 4 by all practical purposes.
-template <class Arc>
-bool Equivalent(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-                double delta = kDelta, bool *error = nullptr) {
-  using Weight = typename Arc::Weight;
-  if (error) *error = false;
-  // Check that the symbol table are compatible.
-  if (!CompatSymbols(fst1.InputSymbols(), fst2.InputSymbols()) ||
-      !CompatSymbols(fst1.OutputSymbols(), fst2.OutputSymbols())) {
-    FSTERROR() << "Equivalent: Input/output symbol tables of 1st argument "
-               << "do not match input/output symbol tables of 2nd argument";
-    if (error) *error = true;
-    return false;
-  }
-  // Check properties first.
-  static constexpr auto props = kNoEpsilons | kIDeterministic | kAcceptor;
-  if (fst1.Properties(props, true) != props) {
-    FSTERROR() << "Equivalent: 1st argument not an"
-               << " epsilon-free deterministic acceptor";
-    if (error) *error = true;
-    return false;
-  }
-  if (fst2.Properties(props, true) != props) {
-    FSTERROR() << "Equivalent: 2nd argument not an"
-               << " epsilon-free deterministic acceptor";
-    if (error) *error = true;
-    return false;
-  }
-  if ((fst1.Properties(kUnweighted, true) != kUnweighted) ||
-      (fst2.Properties(kUnweighted, true) != kUnweighted)) {
-    VectorFst<Arc> efst1(fst1);
-    VectorFst<Arc> efst2(fst2);
-    Push(&efst1, REWEIGHT_TO_INITIAL, delta);
-    Push(&efst2, REWEIGHT_TO_INITIAL, delta);
-    ArcMap(&efst1, QuantizeMapper<Arc>(delta));
-    ArcMap(&efst2, QuantizeMapper<Arc>(delta));
-    EncodeMapper<Arc> mapper(kEncodeWeights | kEncodeLabels, ENCODE);
-    ArcMap(&efst1, &mapper);
-    ArcMap(&efst2, &mapper);
-    return Equivalent(efst1, efst2);
-  }
-  using Util = internal::EquivalenceUtil<Arc>;
-  using MappedId = typename Util::MappedId;
-  enum { FST1 = 1, FST2 = 2 };  // Required by Util::MapState(...)
-  auto s1 = Util::MapState(fst1.Start(), FST1);
-  auto s2 = Util::MapState(fst2.Start(), FST2);
-  // The union-find structure.
-  UnionFind<MappedId> eq_classes(1000, Util::kInvalidId);
-  // Initializes the union-find structure.
-  eq_classes.MakeSet(s1);
-  eq_classes.MakeSet(s2);
-  // Data structure for the (partial) acceptor transition function of fst1 and
-  // fst2: input labels mapped to pairs of MappedIds representing destination
-  // states of the corresponding arcs in fst1 and fst2, respectively.
-  using Label2StatePairMap =
-      std::unordered_map<typename Arc::Label, std::pair<MappedId, MappedId>>;
-  Label2StatePairMap arc_pairs;
-  // Pairs of MappedId's to be processed, organized in a queue.
-  std::deque<std::pair<MappedId, MappedId>> q;
-  bool ret = true;
-  // Returns early if the start states differ w.r.t. finality.
-  if (Util::IsFinal(fst1, s1) != Util::IsFinal(fst2, s2)) ret = false;
-  // Main loop: explores the two acceptors in a breadth-first manner, updating
-  // the equivalence relation on the statesets. Loop invariant: each block of
-  // the states contains either final states only or non-final states only.
-  for (q.push_back(std::make_pair(s1, s2)); ret && !q.empty(); q.pop_front()) {
-    s1 = q.front().first;
-    s2 = q.front().second;
-    // Representatives of the equivalence classes of s1/s2.
-    const auto rep1 = Util::FindSet(&eq_classes, s1);
-    const auto rep2 = Util::FindSet(&eq_classes, s2);
-    if (rep1 != rep2) {
-      eq_classes.Union(rep1, rep2);
-      arc_pairs.clear();
-      // Copies outgoing arcs starting at s1 into the hash-table.
-      if (Util::kDeadState != s1) {
-        ArcIterator<Fst<Arc>> arc_iter(fst1, Util::UnMapState(s1));
-        for (; !arc_iter.Done(); arc_iter.Next()) {
-          const auto &arc = arc_iter.Value();
-          // Zero-weight arcs are treated as if they did not exist.
-          if (arc.weight != Weight::Zero()) {
-            arc_pairs[arc.ilabel].first = Util::MapState(arc.nextstate, FST1);
-          }
-        }
-      }
-      // Copies outgoing arcs starting at s2 into the hashtable.
-      if (Util::kDeadState != s2) {
-        ArcIterator<Fst<Arc>> arc_iter(fst2, Util::UnMapState(s2));
-        for (; !arc_iter.Done(); arc_iter.Next()) {
-          const auto &arc = arc_iter.Value();
-          // Zero-weight arcs are treated as if they did not exist.
-          if (arc.weight != Weight::Zero()) {
-            arc_pairs[arc.ilabel].second = Util::MapState(arc.nextstate, FST2);
-          }
-        }
-      }
-      // Iterates through the hashtable and process pairs of target states.
-      for (const auto &arc_iter : arc_pairs) {
-        const auto &pair = arc_iter.second;
-        if (Util::IsFinal(fst1, pair.first) !=
-            Util::IsFinal(fst2, pair.second)) {
-          // Detected inconsistency: return false.
-          ret = false;
-          break;
-        }
-        q.push_back(pair);
-      }
-    }
-  }
-  if (fst1.Properties(kError, false) || fst2.Properties(kError, false)) {
-    if (error) *error = true;
-    return false;
-  }
-  return ret;
-}
-
-}  // namespace fst
-
-#endif  // FST_EQUIVALENT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/expanded-fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/expanded-fst.h
deleted file mode 100644
index 614ecef0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/expanded-fst.h
+++ /dev/null
@@ -1,183 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Generic FST augmented with state count-interface class definition.
-
-#ifndef FST_EXPANDED_FST_H_
-#define FST_EXPANDED_FST_H_
-
-#include <sys/types.h>
-#include <istream>
-#include <string>
-
-#include <fst/log.h>
-#include <fstream>
-
-#include <fst/fst.h>
-
-
-namespace fst {
-
-// A generic FST plus state count.
-template <class A>
-class ExpandedFst : public Fst<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  virtual StateId NumStates() const = 0;  // State count
-
-  // Get a copy of this ExpandedFst. See Fst<>::Copy() for further doc.
-  ExpandedFst<Arc> *Copy(bool safe = false) const override = 0;
-
-  // Read an ExpandedFst from an input stream; return NULL on error.
-  static ExpandedFst<Arc> *Read(std::istream &strm,
-                                const FstReadOptions &opts) {
-    FstReadOptions ropts(opts);
-    FstHeader hdr;
-    if (ropts.header) {
-      hdr = *opts.header;
-    } else {
-      if (!hdr.Read(strm, opts.source)) return nullptr;
-      ropts.header = &hdr;
-    }
-    if (!(hdr.Properties() & kExpanded)) {
-      LOG(ERROR) << "ExpandedFst::Read: Not an ExpandedFst: " << ropts.source;
-      return nullptr;
-    }
-    const auto reader =
-        FstRegister<Arc>::GetRegister()->GetReader(hdr.FstType());
-    if (!reader) {
-      LOG(ERROR) << "ExpandedFst::Read: Unknown FST type \"" << hdr.FstType()
-                 << "\" (arc type = \"" << A::Type() << "\"): " << ropts.source;
-      return nullptr;
-    }
-    auto *fst = reader(strm, ropts);
-    if (!fst) return nullptr;
-    return static_cast<ExpandedFst<Arc> *>(fst);
-  }
-
-  // Read an ExpandedFst from a file; return NULL on error.
-  // Empty filename reads from standard input.
-  static ExpandedFst<Arc> *Read(const string &filename) {
-    if (!filename.empty()) {
-      std::ifstream strm(filename,
-                              std::ios_base::in | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "ExpandedFst::Read: Can't open file: " << filename;
-        return nullptr;
-      }
-      return Read(strm, FstReadOptions(filename));
-    } else {
-      return Read(std::cin, FstReadOptions("standard input"));
-    }
-  }
-};
-
-namespace internal {
-
-//  ExpandedFst<A> case - abstract methods.
-template <class Arc>
-inline typename Arc::Weight Final(const ExpandedFst<Arc> &fst,
-                                  typename Arc::StateId s) {
-  return fst.Final(s);
-}
-
-template <class Arc>
-inline ssize_t NumArcs(const ExpandedFst<Arc> &fst, typename Arc::StateId s) {
-  return fst.NumArcs(s);
-}
-
-template <class Arc>
-inline ssize_t NumInputEpsilons(const ExpandedFst<Arc> &fst,
-                                typename Arc::StateId s) {
-  return fst.NumInputEpsilons(s);
-}
-
-template <class Arc>
-inline ssize_t NumOutputEpsilons(const ExpandedFst<Arc> &fst,
-                                 typename Arc::StateId s) {
-  return fst.NumOutputEpsilons(s);
-}
-
-}  // namespace internal
-
-// A useful alias when using StdArc.
-using StdExpandedFst = ExpandedFst<StdArc>;
-
-// This is a helper class template useful for attaching an ExpandedFst
-// interface to its implementation, handling reference counting. It
-// delegates to ImplToFst the handling of the Fst interface methods.
-template <class Impl, class FST = ExpandedFst<typename Impl::Arc>>
-class ImplToExpandedFst : public ImplToFst<Impl, FST> {
- public:
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using ImplToFst<Impl, FST>::operator=;
-
-  StateId NumStates() const override { return GetImpl()->NumStates(); }
-
- protected:
-  using ImplToFst<Impl, FST>::GetImpl;
-
-  explicit ImplToExpandedFst(std::shared_ptr<Impl> impl)
-      : ImplToFst<Impl, FST>(impl) {}
-
-  ImplToExpandedFst(const ImplToExpandedFst<Impl, FST> &fst)
-      : ImplToFst<Impl, FST>(fst) {}
-
-  ImplToExpandedFst(const ImplToExpandedFst<Impl, FST> &fst, bool safe)
-      : ImplToFst<Impl, FST>(fst, safe) {}
-
-  static Impl *Read(std::istream &strm, const FstReadOptions &opts) {
-    return Impl::Read(strm, opts);
-  }
-
-  // Read FST implementation from a file; return NULL on error.
-  // Empty filename reads from standard input.
-  static Impl *Read(const string &filename) {
-    if (!filename.empty()) {
-      std::ifstream strm(filename,
-                              std::ios_base::in | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "ExpandedFst::Read: Can't open file: " << filename;
-        return nullptr;
-      }
-      return Impl::Read(strm, FstReadOptions(filename));
-    } else {
-      return Impl::Read(std::cin, FstReadOptions("standard input"));
-    }
-  }
-};
-
-// Function to return the number of states in an FST, counting them
-// if necessary.
-template <class Arc>
-typename Arc::StateId CountStates(const Fst<Arc> &fst) {
-  if (fst.Properties(kExpanded, false)) {
-    const auto *efst = static_cast<const ExpandedFst<Arc> *>(&fst);
-    return efst->NumStates();
-  } else {
-    typename Arc::StateId nstates = 0;
-    for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-      ++nstates;
-    }
-    return nstates;
-  }
-}
-
-// Function to return the number of arcs in an FST.
-template <class Arc>
-typename Arc::StateId CountArcs(const Fst<Arc> &fst) {
-  size_t narcs = 0;
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    narcs += fst.NumArcs(siter.Value());
-  }
-  return narcs;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXPANDED_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/expectation-weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/expectation-weight.h
deleted file mode 100644
index e6395794..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/expectation-weight.h
+++ /dev/null
@@ -1,137 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Expectation semiring as described by Jason Eisner:
-// See: doi=10.1.1.22.9398
-// Multiplex semiring operations and identities:
-//    One: <One, Zero>
-//    Zero: <Zero, Zero>
-//    Plus: <a1, b1> + <a2, b2> = < (a1 + a2) , (b1 + b2) >
-//    Times: <a1, b1> * <a2, b2> = < (a1 * a2) , [(a1 * b2) + (a2 * b1)] >
-//    Division: Undefined (currently)
-//
-// Usually used to store the pair <probability, random_variable> so that
-// ShortestDistance[Fst<ArcTpl<ExpectationWeight<P, V>>>]
-//    == < PosteriorProbability, Expected_Value[V] >
-
-#ifndef FST_EXPECTATION_WEIGHT_H_
-#define FST_EXPECTATION_WEIGHT_H_
-
-#include <string>
-
-#include <fst/log.h>
-
-#include <fst/pair-weight.h>
-#include <fst/product-weight.h>
-
-
-namespace fst {
-
-// X1 is usually a probability weight like LogWeight.
-// X2 is usually a random variable or vector (see SignedLogWeight or
-// SparsePowerWeight).
-//
-// If X1 is distinct from X2, it is required that there is an external product
-// between X1 and X2 and if both semriring are commutative, or left or right
-// semirings, then result must have those properties.
-template <class X1, class X2>
-class ExpectationWeight : public PairWeight<X1, X2> {
- public:
-  using PairWeight<X1, X2>::Value1;
-  using PairWeight<X1, X2>::Value2;
-
-  using PairWeight<X1, X2>::Reverse;
-  using PairWeight<X1, X2>::Quantize;
-  using PairWeight<X1, X2>::Member;
-
-  using ReverseWeight =
-      ExpectationWeight<typename X1::ReverseWeight, typename X2::ReverseWeight>;
-
-  ExpectationWeight() : PairWeight<X1, X2>(Zero()) {}
-
-  ExpectationWeight(const ExpectationWeight &weight)
-      : PairWeight<X1, X2>(weight) {}
-
-  explicit ExpectationWeight(const PairWeight<X1, X2> &weight)
-      : PairWeight<X1, X2>(weight) {}
-
-  ExpectationWeight(const X1 &x1, const X2 &x2) : PairWeight<X1, X2>(x1, x2) {}
-
-  static const ExpectationWeight &Zero() {
-    static const ExpectationWeight zero(X1::Zero(), X2::Zero());
-    return zero;
-  }
-
-  static const ExpectationWeight &One() {
-    static const ExpectationWeight one(X1::One(), X2::Zero());
-    return one;
-  }
-
-  static const ExpectationWeight &NoWeight() {
-    static const ExpectationWeight no_weight(X1::NoWeight(), X2::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type =
-        new string("expectation_" + X1::Type() + "_" + X2::Type());
-    return *type;
-  }
-
-  PairWeight<X1, X2> Quantize(float delta = kDelta) const {
-    return ExpectationWeight(PairWeight<X1, X2>::Quantize());
-  }
-
-  ReverseWeight Reverse() const {
-    return ReverseWeight(PairWeight<X1, X2>::Reverse());
-  }
-
-  bool Member() const { return PairWeight<X1, X2>::Member(); }
-
-  static constexpr uint64 Properties() {
-    return X1::Properties() & X2::Properties() &
-           (kLeftSemiring | kRightSemiring | kCommutative | kIdempotent);
-  }
-};
-
-template <class X1, class X2>
-inline ExpectationWeight<X1, X2> Plus(const ExpectationWeight<X1, X2> &w1,
-                                      const ExpectationWeight<X1, X2> &w2) {
-  return ExpectationWeight<X1, X2>(Plus(w1.Value1(), w2.Value1()),
-                                   Plus(w1.Value2(), w2.Value2()));
-}
-
-template <class X1, class X2>
-inline ExpectationWeight<X1, X2> Times(const ExpectationWeight<X1, X2> &w1,
-                                       const ExpectationWeight<X1, X2> &w2) {
-  return ExpectationWeight<X1, X2>(
-      Times(w1.Value1(), w2.Value1()),
-      Plus(Times(w1.Value1(), w2.Value2()), Times(w1.Value2(), w2.Value1())));
-}
-
-template <class X1, class X2>
-inline ExpectationWeight<X1, X2> Divide(const ExpectationWeight<X1, X2> &w1,
-                                        const ExpectationWeight<X1, X2> &w2,
-                                        DivideType typ = DIVIDE_ANY) {
-  FSTERROR() << "ExpectationWeight::Divide: Not implemented";
-  return ExpectationWeight<X1, X2>::NoWeight();
-}
-
-// This function object generates weights by calling the underlying generators
-// for the template weight types, like all other pair weight types. This is
-// intended primarily for testing.
-template <class X1, class X2>
-class WeightGenerate<ExpectationWeight<X1, X2>>
-    : public WeightGenerate<PairWeight<X1, X2>> {
- public:
-  using Weight = ExpectationWeight<X1, X2>;
-  using Generate = WeightGenerate<PairWeight<X1, X2>>;
-
-  explicit WeightGenerate(bool allow_zero = true) : Generate(allow_zero) {}
-
-  Weight operator()() const { return Weight(Generate::operator()()); }
-};
-
-}  // namespace fst
-
-#endif  // FST_EXPECTATION_WEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/compress-script.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/compress-script.h
deleted file mode 100644
index d666598b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/compress-script.h
+++ /dev/null
@@ -1,54 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Declarations of 'scriptable' versions of compression operations, that is,
-// those that can be called with FstClass-type arguments.
-
-#ifndef FST_EXTENSIONS_COMPRESS_COMPRESS_SCRIPT_H_
-#define FST_EXTENSIONS_COMPRESS_COMPRESS_SCRIPT_H_
-
-#include <utility>
-#include <vector>
-
-#include <fst/extensions/compress/compress.h>
-
-#include <fst/log.h>
-#include <fst/util.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-typedef std::tuple<const FstClass &, const string &, const bool> CompressArgs;
-
-template <class Arc>
-void Compress(CompressArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  const string &filename = std::get<1>(*args);
-  const bool gzip = std::get<2>(*args);
-
-  if (!fst::Compress(fst, filename, gzip)) FSTERROR() << "Compress: failed";
-}
-
-void Compress(const FstClass &fst, const string &filename, const bool gzip);
-
-typedef std::tuple<const string &, MutableFstClass *, const bool>
-    DecompressArgs;
-
-template <class Arc>
-void Decompress(DecompressArgs *args) {
-  const string &filename = std::get<0>(*args);
-  MutableFst<Arc> *fst = std::get<1>(*args)->GetMutableFst<Arc>();
-  const bool gzip = std::get<2>(*args);
-
-  if (!fst::Decompress(filename, fst, gzip))
-    FSTERROR() << "Decompress: failed";
-}
-
-void Decompress(const string &filename, MutableFstClass *fst, const bool gzip);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_COMPRESS_COMPRESS_SCRIPT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/compress.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/compress.h
deleted file mode 100644
index 59d25628..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/compress.h
+++ /dev/null
@@ -1,904 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Compresses and decompresses unweighted FSTs.
-
-#ifndef FST_EXTENSIONS_COMPRESS_COMPRESS_H_
-#define FST_EXTENSIONS_COMPRESS_COMPRESS_H_
-
-#include <cstdio>
-
-#include <iostream>
-#include <queue>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/extensions/compress/elias.h>
-#include <fst/extensions/compress/gzfile.h>
-#include <fstream>
-
-#include <fst/encode.h>
-#include <fst/expanded-fst.h>
-#include <fst/fst.h>
-#include <fst/mutable-fst.h>
-#include <fst/statesort.h>
-
-namespace fst {
-
-// Identifies stream data as a vanilla compressed FST.
-static const int32 kCompressMagicNumber = 1858869554;
-// Identifies stream data as (probably) a Gzip file accidentally read from
-// a vanilla stream, without gzip support.
-static const int32 kGzipMagicNumber = 0x8b1f;
-// Selects the two most significant bytes.
-constexpr uint32 kGzipMask = 0xffffffff >> 16;
-
-namespace internal {
-
-// Expands a Lempel Ziv code and returns the set of code words. expanded_code[i]
-// is the i^th Lempel Ziv codeword.
-template <class Var, class Edge>
-bool ExpandLZCode(const std::vector<std::pair<Var, Edge>> &code,
-                  std::vector<std::vector<Edge>> *expanded_code) {
-  expanded_code->resize(code.size());
-  for (int i = 0; i < code.size(); ++i) {
-    if (code[i].first > i) {
-      LOG(ERROR) << "ExpandLZCode: Not a valid code";
-      return false;
-    }
-    if (code[i].first == 0) {
-      (*expanded_code)[i].resize(1, code[i].second);
-    } else {
-      (*expanded_code)[i].resize((*expanded_code)[code[i].first - 1].size() +
-                                 1);
-      std::copy((*expanded_code)[code[i].first - 1].begin(),
-                (*expanded_code)[code[i].first - 1].end(),
-                (*expanded_code)[i].begin());
-      (*expanded_code)[i][(*expanded_code)[code[i].first - 1].size()] =
-          code[i].second;
-    }
-  }
-  return true;
-}
-
-}  // namespace internal
-
-// Lempel Ziv on data structure Edge, with a less than operator
-// EdgeLessThan and an equals operator  EdgeEquals.
-// Edge has a value defaultedge which it never takes and
-// Edge is defined, it is initialized to defaultedge
-template <class Var, class Edge, class EdgeLessThan, class EdgeEquals>
-class LempelZiv {
- public:
-  LempelZiv() : dict_number_(0), default_edge_() {
-    root_.current_number = dict_number_++;
-    root_.current_edge = default_edge_;
-    decode_vector_.push_back(std::make_pair(0, default_edge_));
-  }
-  // Encodes a vector input into output
-  void BatchEncode(const std::vector<Edge> &input,
-                   std::vector<std::pair<Var, Edge>> *output);
-
-  // Decodes codedvector to output. Returns false if
-  // the index exceeds the size.
-  bool BatchDecode(const std::vector<std::pair<Var, Edge>> &input,
-                   std::vector<Edge> *output);
-
-  // Decodes a single dictionary element. Returns false
-  // if the index exceeds the size.
-  bool SingleDecode(const Var &index, Edge *output) {
-    if (index >= decode_vector_.size()) {
-      LOG(ERROR) << "LempelZiv::SingleDecode: "
-                 << "Index exceeded the dictionary size";
-      return false;
-    } else {
-      *output = decode_vector_[index].second;
-      return true;
-    }
-  }
-
-  ~LempelZiv() {
-    for (auto it = (root_.next_number).begin(); it != (root_.next_number).end();
-         ++it) {
-      CleanUp(it->second);
-    }
-  }
-  // Adds a single dictionary element while decoding
-  //  void AddDictElement(const std::pair<Var, Edge> &newdict) {
-  //    EdgeEquals InstEdgeEquals;
-  //  if (InstEdgeEquals(newdict.second, default_edge_) != 1)
-  //     decode_vector_.push_back(newdict);
-  //  }
-
- private:
-  // Node datastructure is used for encoding
-
-  struct Node {
-    Var current_number;
-    Edge current_edge;
-    std::map<Edge, Node *, EdgeLessThan> next_number;
-  };
-
-  void CleanUp(Node *temp) {
-    for (auto it = (temp->next_number).begin(); it != (temp->next_number).end();
-         ++it) {
-      CleanUp(it->second);
-    }
-    delete temp;
-  }
-  Node root_;
-  Var dict_number_;
-  // decode_vector_ is used for decoding
-  std::vector<std::pair<Var, Edge>> decode_vector_;
-  Edge default_edge_;
-};
-
-template <class Var, class Edge, class EdgeLessThan, class EdgeEquals>
-void LempelZiv<Var, Edge, EdgeLessThan, EdgeEquals>::BatchEncode(
-    const std::vector<Edge> &input, std::vector<std::pair<Var, Edge>> *output) {
-  for (typename std::vector<Edge>::const_iterator it = input.begin();
-       it != input.end(); ++it) {
-    Node *temp_node = &root_;
-    while (it != input.end()) {
-      auto next = (temp_node->next_number).find(*it);
-      if (next != (temp_node->next_number).end()) {
-        temp_node = next->second;
-        ++it;
-      } else {
-        break;
-      }
-    }
-    if (it == input.end() && temp_node->current_number != 0) {
-      output->push_back(
-          std::make_pair(temp_node->current_number, default_edge_));
-    } else if (it != input.end()) {
-      output->push_back(std::make_pair(temp_node->current_number, *it));
-      Node *new_node = new (Node);
-      new_node->current_number = dict_number_++;
-      new_node->current_edge = *it;
-      (temp_node->next_number)[*it] = new_node;
-    }
-    if (it == input.end()) break;
-  }
-}
-
-template <class Var, class Edge, class EdgeLessThan, class EdgeEquals>
-bool LempelZiv<Var, Edge, EdgeLessThan, EdgeEquals>::BatchDecode(
-    const std::vector<std::pair<Var, Edge>> &input, std::vector<Edge> *output) {
-  for (typename std::vector<std::pair<Var, Edge>>::const_iterator it =
-           input.begin();
-       it != input.end(); ++it) {
-    std::vector<Edge> temp_output;
-    EdgeEquals InstEdgeEquals;
-    if (InstEdgeEquals(it->second, default_edge_) != 1) {
-      decode_vector_.push_back(*it);
-      temp_output.push_back(it->second);
-    }
-    Var temp_integer = it->first;
-    if (temp_integer >= decode_vector_.size()) {
-      LOG(ERROR) << "LempelZiv::BatchDecode: "
-                 << "Index exceeded the dictionary size";
-      return false;
-    } else {
-      while (temp_integer != 0) {
-        temp_output.push_back(decode_vector_[temp_integer].second);
-        temp_integer = decode_vector_[temp_integer].first;
-      }
-      std::reverse(temp_output.begin(), temp_output.end());
-      output->insert(output->end(), temp_output.begin(), temp_output.end());
-    }
-  }
-  return true;
-}
-
-// The main Compressor class
-template <class Arc>
-class Compressor {
- public:
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Label Label;
-  typedef typename Arc::Weight Weight;
-
-  Compressor() {}
-
-  // Compresses fst into a boolean vector code. Returns true on sucesss.
-  bool Compress(const Fst<Arc> &fst, std::ostream &strm);
-
-  // Decompresses the boolean vector into Fst. Returns true on sucesss.
-  bool Decompress(std::istream &strm, const string &source,
-                  MutableFst<Arc> *fst);
-
-  // Finds the BFS order of a fst
-  void BfsOrder(const ExpandedFst<Arc> &fst, std::vector<StateId> *order);
-
-  // Preprocessing step to convert fst to a isomorphic fst
-  // Returns a preproccess fst and a dictionary
-  void Preprocess(const Fst<Arc> &fst, MutableFst<Arc> *preprocessedfst,
-                  EncodeMapper<Arc> *encoder);
-
-  // Performs Lempel Ziv and outputs a stream of integers
-  // and sends it to a stream
-  void EncodeProcessedFst(const ExpandedFst<Arc> &fst, std::ostream &strm);
-
-  // Decodes fst from the stream
-  void DecodeProcessedFst(const std::vector<StateId> &input,
-                          MutableFst<Arc> *fst, bool unweighted);
-
-  // Converts buffer_code_ to uint8 and writes to a stream.
-
-  // Writes the boolean file to the stream
-  void WriteToStream(std::ostream &strm);
-
-  // Writes the weights to the stream
-  void WriteWeight(const std::vector<Weight> &input, std::ostream &strm);
-
-  void ReadWeight(std::istream &strm, std::vector<Weight> *output);
-
-  // Same as fst::Decode without the line RmFinalEpsilon(fst)
-  void DecodeForCompress(MutableFst<Arc> *fst, const EncodeMapper<Arc> &mapper);
-
-  // Updates the buffer_code_
-  template <class CVar>
-  void WriteToBuffer(CVar input) {
-    std::vector<bool> current_code;
-    Elias<CVar>::DeltaEncode(input, &current_code);
-    if (!buffer_code_.empty()) {
-      buffer_code_.insert(buffer_code_.end(), current_code.begin(),
-                          current_code.end());
-    } else {
-      buffer_code_.assign(current_code.begin(), current_code.end());
-    }
-  }
-
- private:
-  struct LZLabel {
-    LZLabel() : label(0) {}
-    Label label;
-  };
-
-  struct LabelLessThan {
-    bool operator()(const LZLabel &labelone, const LZLabel &labeltwo) const {
-      return labelone.label < labeltwo.label;
-    }
-  };
-
-  struct LabelEquals {
-    bool operator()(const LZLabel &labelone, const LZLabel &labeltwo) const {
-      return labelone.label == labeltwo.label;
-    }
-  };
-
-  struct Transition {
-    Transition() : nextstate(0), label(0), weight(Weight::Zero()) {}
-
-    StateId nextstate;
-    Label label;
-    Weight weight;
-  };
-
-  struct TransitionLessThan {
-    bool operator()(const Transition &transition_one,
-                    const Transition &transition_two) const {
-      if (transition_one.nextstate == transition_two.nextstate)
-        return transition_one.label < transition_two.label;
-      else
-        return transition_one.nextstate < transition_two.nextstate;
-    }
-  } transition_less_than;
-
-  struct TransitionEquals {
-    bool operator()(const Transition &transition_one,
-                    const Transition &transition_two) const {
-      return transition_one.nextstate == transition_two.nextstate &&
-             transition_one.label == transition_two.label;
-    }
-  } transition_equals;
-
-  struct OldDictCompare {
-    bool operator()(const std::pair<StateId, Transition> &pair_one,
-                    const std::pair<StateId, Transition> &pair_two) const {
-      if ((pair_one.second).nextstate == (pair_two.second).nextstate)
-        return (pair_one.second).label < (pair_two.second).label;
-      else
-        return (pair_one.second).nextstate < (pair_two.second).nextstate;
-    }
-  } old_dict_compare;
-
-  std::vector<bool> buffer_code_;
-  std::vector<Weight> arc_weight_;
-  std::vector<Weight> final_weight_;
-};
-
-template <class Arc>
-inline void Compressor<Arc>::DecodeForCompress(
-    MutableFst<Arc> *fst, const EncodeMapper<Arc> &mapper) {
-  ArcMap(fst, EncodeMapper<Arc>(mapper, DECODE));
-  fst->SetInputSymbols(mapper.InputSymbols());
-  fst->SetOutputSymbols(mapper.OutputSymbols());
-}
-
-// Compressor::BfsOrder
-template <class Arc>
-void Compressor<Arc>::BfsOrder(const ExpandedFst<Arc> &fst,
-                               std::vector<StateId> *order) {
-  Arc arc;
-  StateId bfs_visit_number = 0;
-  std::queue<StateId> states_queue;
-  order->assign(fst.NumStates(), kNoStateId);
-  states_queue.push(fst.Start());
-  (*order)[fst.Start()] = bfs_visit_number++;
-  while (!states_queue.empty()) {
-    for (ArcIterator<Fst<Arc>> aiter(fst, states_queue.front()); !aiter.Done();
-         aiter.Next()) {
-      arc = aiter.Value();
-      StateId nextstate = arc.nextstate;
-      if ((*order)[nextstate] == kNoStateId) {
-        (*order)[nextstate] = bfs_visit_number++;
-        states_queue.push(nextstate);
-      }
-    }
-    states_queue.pop();
-  }
-
-  // If the FST is unconnected, then the following
-  // code finds them
-  while (bfs_visit_number < fst.NumStates()) {
-    int unseen_state = 0;
-    for (unseen_state = 0; unseen_state < fst.NumStates(); ++unseen_state) {
-      if ((*order)[unseen_state] == kNoStateId) break;
-    }
-    states_queue.push(unseen_state);
-    (*order)[unseen_state] = bfs_visit_number++;
-    while (!states_queue.empty()) {
-      for (ArcIterator<Fst<Arc>> aiter(fst, states_queue.front());
-           !aiter.Done(); aiter.Next()) {
-        arc = aiter.Value();
-        StateId nextstate = arc.nextstate;
-        if ((*order)[nextstate] == kNoStateId) {
-          (*order)[nextstate] = bfs_visit_number++;
-          states_queue.push(nextstate);
-        }
-      }
-      states_queue.pop();
-    }
-  }
-}
-
-template <class Arc>
-void Compressor<Arc>::Preprocess(const Fst<Arc> &fst,
-                                 MutableFst<Arc> *preprocessedfst,
-                                 EncodeMapper<Arc> *encoder) {
-  *preprocessedfst = fst;
-  if (!preprocessedfst->NumStates()) {
-    return;
-  }
-  // Relabels the edges and develops a dictionary
-  Encode(preprocessedfst, encoder);
-  std::vector<StateId> order;
-  // Finds the BFS sorting order of the fst
-  BfsOrder(*preprocessedfst, &order);
-  // Reorders the states according to the BFS order
-  StateSort(preprocessedfst, order);
-}
-
-template <class Arc>
-void Compressor<Arc>::EncodeProcessedFst(const ExpandedFst<Arc> &fst,
-                                         std::ostream &strm) {
-  std::vector<StateId> output;
-  LempelZiv<StateId, LZLabel, LabelLessThan, LabelEquals> dict_new;
-  LempelZiv<StateId, Transition, TransitionLessThan, TransitionEquals> dict_old;
-  std::vector<LZLabel> current_new_input;
-  std::vector<Transition> current_old_input;
-  std::vector<std::pair<StateId, LZLabel>> current_new_output;
-  std::vector<std::pair<StateId, Transition>> current_old_output;
-  std::vector<StateId> final_states;
-
-  StateId number_of_states = fst.NumStates();
-
-  StateId seen_states = 0;
-  // Adding the number of states
-  WriteToBuffer<StateId>(number_of_states);
-
-  for (StateId state = 0; state < number_of_states; ++state) {
-    current_new_input.clear();
-    current_old_input.clear();
-    current_new_output.clear();
-    current_old_output.clear();
-    if (state > seen_states) ++seen_states;
-
-    // Collecting the final states
-    if (fst.Final(state) != Weight::Zero()) {
-      final_states.push_back(state);
-      final_weight_.push_back(fst.Final(state));
-    }
-
-    // Reading the states
-    for (ArcIterator<Fst<Arc>> aiter(fst, state); !aiter.Done(); aiter.Next()) {
-      Arc arc = aiter.Value();
-      if (arc.nextstate > seen_states) {  // RILEY: > or >= ?
-        ++seen_states;
-        LZLabel temp_label;
-        temp_label.label = arc.ilabel;
-        arc_weight_.push_back(arc.weight);
-        current_new_input.push_back(temp_label);
-      } else {
-        Transition temp_transition;
-        temp_transition.nextstate = arc.nextstate;
-        temp_transition.label = arc.ilabel;
-        temp_transition.weight = arc.weight;
-        current_old_input.push_back(temp_transition);
-      }
-    }
-    // Adding new states
-    dict_new.BatchEncode(current_new_input, &current_new_output);
-    WriteToBuffer<StateId>(current_new_output.size());
-
-    for (auto it = current_new_output.begin(); it != current_new_output.end();
-         ++it) {
-      WriteToBuffer<StateId>(it->first);
-      WriteToBuffer<Label>((it->second).label);
-    }
-    // Adding old states by sorting and using difference coding
-    std::sort(current_old_input.begin(), current_old_input.end(),
-              transition_less_than);
-    for (auto it = current_old_input.begin(); it != current_old_input.end();
-         ++it) {
-      arc_weight_.push_back(it->weight);
-    }
-    dict_old.BatchEncode(current_old_input, &current_old_output);
-    std::vector<StateId> dict_old_temp;
-    std::vector<Transition> transition_old_temp;
-    for (auto it = current_old_output.begin(); it != current_old_output.end();
-         ++it) {
-      dict_old_temp.push_back(it->first);
-      transition_old_temp.push_back(it->second);
-    }
-    if (!transition_old_temp.empty()) {
-      if ((transition_old_temp.back()).nextstate == 0 &&
-          (transition_old_temp.back()).label == 0) {
-        transition_old_temp.pop_back();
-      }
-    }
-    std::sort(dict_old_temp.begin(), dict_old_temp.end());
-    std::sort(transition_old_temp.begin(), transition_old_temp.end(),
-              transition_less_than);
-
-    WriteToBuffer<StateId>(dict_old_temp.size());
-    if (dict_old_temp.size() != transition_old_temp.size())
-      WriteToBuffer<int>(1);
-    else
-      WriteToBuffer<int>(0);
-
-    StateId previous;
-    if (!dict_old_temp.empty()) {
-      WriteToBuffer<StateId>(dict_old_temp.front());
-      previous = dict_old_temp.front();
-    }
-    if (dict_old_temp.size() > 1) {
-      for (auto it = dict_old_temp.begin() + 1; it != dict_old_temp.end();
-           ++it) {
-        WriteToBuffer<StateId>(*it - previous);
-        previous = *it;
-      }
-    }
-    if (!transition_old_temp.empty()) {
-      WriteToBuffer<StateId>((transition_old_temp.front()).nextstate);
-      previous = (transition_old_temp.front()).nextstate;
-      WriteToBuffer<Label>((transition_old_temp.front()).label);
-    }
-    if (transition_old_temp.size() > 1) {
-      for (auto it = transition_old_temp.begin() + 1;
-           it != transition_old_temp.end(); ++it) {
-        WriteToBuffer<StateId>(it->nextstate - previous);
-        previous = it->nextstate;
-        WriteToBuffer<StateId>(it->label);
-      }
-    }
-  }
-  // Adding final states
-  WriteToBuffer<StateId>(final_states.size());
-  if (!final_states.empty()) {
-    for (auto it = final_states.begin(); it != final_states.end(); ++it) {
-      WriteToBuffer<StateId>(*it);
-    }
-  }
-  WriteToStream(strm);
-  uint8 unweighted = (fst.Properties(kUnweighted, true) == kUnweighted);
-  WriteType(strm, unweighted);
-  if (unweighted == 0) {
-    WriteWeight(arc_weight_, strm);
-    WriteWeight(final_weight_, strm);
-  }
-}
-
-template <class Arc>
-void Compressor<Arc>::DecodeProcessedFst(const std::vector<StateId> &input,
-                                         MutableFst<Arc> *fst,
-                                         bool unweighted) {
-  LempelZiv<StateId, LZLabel, LabelLessThan, LabelEquals> dict_new;
-  LempelZiv<StateId, Transition, TransitionLessThan, TransitionEquals> dict_old;
-  std::vector<std::pair<StateId, LZLabel>> current_new_input;
-  std::vector<std::pair<StateId, Transition>> current_old_input;
-  std::vector<LZLabel> current_new_output;
-  std::vector<Transition> current_old_output;
-  std::vector<std::pair<StateId, Transition>> actual_old_dict_numbers;
-  std::vector<Transition> actual_old_dict_transitions;
-  auto arc_weight_it = arc_weight_.begin();
-  Transition default_transition;
-  StateId seen_states = 1;
-
-  // Adding states.
-  const StateId num_states = input.front();
-  if (num_states > 0) {
-    const StateId start_state = fst->AddState();
-    fst->SetStart(start_state);
-    for (StateId state = 1; state < num_states; ++state) {
-      fst->AddState();
-    }
-  }
-
-  typename std::vector<StateId>::const_iterator main_it = input.begin();
-  ++main_it;
-
-  for (StateId current_state = 0; current_state < num_states; ++current_state) {
-    if (current_state >= seen_states) ++seen_states;
-    current_new_input.clear();
-    current_new_output.clear();
-    current_old_input.clear();
-    current_old_output.clear();
-    // New states
-    StateId current_number_new_elements = *main_it;
-    ++main_it;
-    for (StateId new_integer = 0; new_integer < current_number_new_elements;
-         ++new_integer) {
-      std::pair<StateId, LZLabel> temp_new_dict_element;
-      temp_new_dict_element.first = *main_it;
-      ++main_it;
-      LZLabel temp_label;
-      temp_label.label = *main_it;
-      ++main_it;
-      temp_new_dict_element.second = temp_label;
-      current_new_input.push_back(temp_new_dict_element);
-    }
-    dict_new.BatchDecode(current_new_input, &current_new_output);
-    for (auto it = current_new_output.begin(); it != current_new_output.end();
-         ++it) {
-      if (!unweighted) {
-        fst->AddArc(current_state,
-                    Arc(it->label, it->label, *arc_weight_it, seen_states++));
-        ++arc_weight_it;
-      } else {
-        fst->AddArc(current_state,
-                    Arc(it->label, it->label, Weight::One(), seen_states++));
-      }
-    }
-
-    // Old states dictionary
-    StateId current_number_old_elements = *main_it;
-    ++main_it;
-    StateId is_zero_removed = *main_it;
-    ++main_it;
-    StateId previous = 0;
-    actual_old_dict_numbers.clear();
-    for (StateId new_integer = 0; new_integer < current_number_old_elements;
-         ++new_integer) {
-      std::pair<StateId, Transition> pair_temp_transition;
-      if (new_integer == 0) {
-        pair_temp_transition.first = *main_it;
-        previous = *main_it;
-      } else {
-        pair_temp_transition.first = *main_it + previous;
-        previous = pair_temp_transition.first;
-      }
-      ++main_it;
-      Transition temp_test;
-      if (!dict_old.SingleDecode(pair_temp_transition.first, &temp_test)) {
-        FSTERROR() << "Compressor::Decode: failed";
-        fst->DeleteStates();
-        fst->SetProperties(kError, kError);
-        return;
-      }
-      pair_temp_transition.second = temp_test;
-      actual_old_dict_numbers.push_back(pair_temp_transition);
-    }
-
-    // Reordering the dictionary elements
-    std::sort(actual_old_dict_numbers.begin(), actual_old_dict_numbers.end(),
-              old_dict_compare);
-
-    // Transitions
-    previous = 0;
-    actual_old_dict_transitions.clear();
-
-    for (StateId new_integer = 0;
-         new_integer < current_number_old_elements - is_zero_removed;
-         ++new_integer) {
-      Transition temp_transition;
-      if (new_integer == 0) {
-        temp_transition.nextstate = *main_it;
-        previous = *main_it;
-      } else {
-        temp_transition.nextstate = *main_it + previous;
-        previous = temp_transition.nextstate;
-      }
-      ++main_it;
-      temp_transition.label = *main_it;
-      ++main_it;
-      actual_old_dict_transitions.push_back(temp_transition);
-    }
-
-    if (is_zero_removed == 1) {
-      actual_old_dict_transitions.push_back(default_transition);
-    }
-
-    auto trans_it = actual_old_dict_transitions.begin();
-    auto dict_it = actual_old_dict_numbers.begin();
-
-    while (trans_it != actual_old_dict_transitions.end() &&
-           dict_it != actual_old_dict_numbers.end()) {
-      if (dict_it->first == 0) {
-        ++dict_it;
-      } else {
-        std::pair<StateId, Transition> temp_pair;
-        if (transition_equals(*trans_it, default_transition) == 1) {
-          temp_pair.first = dict_it->first;
-          temp_pair.second = default_transition;
-          ++dict_it;
-        } else if (transition_less_than(dict_it->second, *trans_it) == 1) {
-          temp_pair.first = dict_it->first;
-          temp_pair.second = *trans_it;
-          ++dict_it;
-        } else {
-          temp_pair.first = 0;
-          temp_pair.second = *trans_it;
-        }
-        ++trans_it;
-        current_old_input.push_back(temp_pair);
-      }
-    }
-    while (trans_it != actual_old_dict_transitions.end()) {
-      std::pair<StateId, Transition> temp_pair;
-      temp_pair.first = 0;
-      temp_pair.second = *trans_it;
-      ++trans_it;
-      current_old_input.push_back(temp_pair);
-    }
-
-    // Adding old elements in the dictionary
-    if (!dict_old.BatchDecode(current_old_input, &current_old_output)) {
-      FSTERROR() << "Compressor::Decode: Failed";
-      fst->DeleteStates();
-      fst->SetProperties(kError, kError);
-      return;
-    }
-
-    for (auto it = current_old_output.begin(); it != current_old_output.end();
-         ++it) {
-      if (!unweighted) {
-        fst->AddArc(current_state,
-                    Arc(it->label, it->label, *arc_weight_it, it->nextstate));
-        ++arc_weight_it;
-      } else {
-        fst->AddArc(current_state,
-                    Arc(it->label, it->label, Weight::One(), it->nextstate));
-      }
-    }
-  }
-  // Adding the final states
-  StateId number_of_final_states = *main_it;
-  if (number_of_final_states > 0) {
-    ++main_it;
-    for (StateId temp_int = 0; temp_int < number_of_final_states; ++temp_int) {
-      if (!unweighted) {
-        fst->SetFinal(*main_it, final_weight_[temp_int]);
-      } else {
-        fst->SetFinal(*main_it, Weight(0));
-      }
-      ++main_it;
-    }
-  }
-}
-
-template <class Arc>
-void Compressor<Arc>::ReadWeight(std::istream &strm,
-                                 std::vector<Weight> *output) {
-  int64 size;
-  Weight weight;
-  ReadType(strm, &size);
-  for (int64 i = 0; i < size; ++i) {
-    weight.Read(strm);
-    output->push_back(weight);
-  }
-}
-
-template <class Arc>
-bool Compressor<Arc>::Decompress(std::istream &strm, const string &source,
-                                 MutableFst<Arc> *fst) {
-  fst->DeleteStates();
-  int32 magic_number = 0;
-  ReadType(strm, &magic_number);
-  if (magic_number != kCompressMagicNumber) {
-    LOG(ERROR) << "Decompress: Bad compressed Fst: " << source;
-    // If the most significant two bytes of the magic number match the
-    // gzip magic number, then we are probably reading a gzip file as an
-    // ordinary stream.
-    if ((magic_number & kGzipMask) == kGzipMagicNumber) {
-      LOG(ERROR) << "Decompress: Fst appears to be compressed with Gzip, but "
-                    "gzip decompression was not requested. Try with "
-                    "the --gzip flag"
-                    ".";
-    }
-    return false;
-  }
-  std::unique_ptr<EncodeMapper<Arc>> encoder(
-      EncodeMapper<Arc>::Read(strm, "Decoding", DECODE));
-  std::vector<bool> bool_code;
-  uint8 block;
-  uint8 msb = 128;
-  int64 data_size;
-  ReadType(strm, &data_size);
-  for (int64 i = 0; i < data_size; ++i) {
-    ReadType(strm, &block);
-    for (int j = 0; j < 8; ++j) {
-      uint8 temp = msb & block;
-      if (temp == 128)
-        bool_code.push_back(1);
-      else
-        bool_code.push_back(0);
-      block = block << 1;
-    }
-  }
-  std::vector<StateId> int_code;
-  Elias<StateId>::BatchDecode(bool_code, &int_code);
-  bool_code.clear();
-  uint8 unweighted;
-  ReadType(strm, &unweighted);
-  if (unweighted == 0) {
-    ReadWeight(strm, &arc_weight_);
-    ReadWeight(strm, &final_weight_);
-  }
-  DecodeProcessedFst(int_code, fst, unweighted);
-  DecodeForCompress(fst, *encoder);
-  return !fst->Properties(kError, false);
-}
-
-template <class Arc>
-void Compressor<Arc>::WriteWeight(const std::vector<Weight> &input,
-                                  std::ostream &strm) {
-  int64 size = input.size();
-  WriteType(strm, size);
-  for (typename std::vector<Weight>::const_iterator it = input.begin();
-       it != input.end(); ++it) {
-    it->Write(strm);
-  }
-}
-
-template <class Arc>
-void Compressor<Arc>::WriteToStream(std::ostream &strm) {
-  while (buffer_code_.size() % 8 != 0) buffer_code_.push_back(1);
-  int64 data_size = buffer_code_.size() / 8;
-  WriteType(strm, data_size);
-  std::vector<bool>::const_iterator it;
-  int64 i;
-  uint8 block;
-  for (it = buffer_code_.begin(), i = 0; it != buffer_code_.end(); ++it, ++i) {
-    if (i % 8 == 0) {
-      if (i > 0) WriteType(strm, block);
-      block = 0;
-    } else {
-      block = block << 1;
-    }
-    block |= *it;
-  }
-  WriteType(strm, block);
-}
-
-template <class Arc>
-bool Compressor<Arc>::Compress(const Fst<Arc> &fst, std::ostream &strm) {
-  VectorFst<Arc> processedfst;
-  EncodeMapper<Arc> encoder(kEncodeLabels, ENCODE);
-  Preprocess(fst, &processedfst, &encoder);
-  WriteType(strm, kCompressMagicNumber);
-  encoder.Write(strm, "encoder stream");
-  EncodeProcessedFst(processedfst, strm);
-  return true;
-}
-
-// Convenience functions that call the compressor and decompressor.
-
-template <class Arc>
-void Compress(const Fst<Arc> &fst, std::ostream &strm) {
-  Compressor<Arc> comp;
-  comp.Compress(fst, strm);
-}
-
-// Returns true on success.
-template <class Arc>
-bool Compress(const Fst<Arc> &fst, const string &file_name,
-              const bool gzip = false) {
-  if (gzip) {
-    if (file_name.empty()) {
-      std::stringstream strm;
-      Compress(fst, strm);
-      OGzFile gzfile(fileno(stdout));
-      gzfile.write(strm);
-      if (!gzfile) {
-        LOG(ERROR) << "Compress: Can't write to file: stdout";
-        return false;
-      }
-    } else {
-      std::stringstream strm;
-      Compress(fst, strm);
-      OGzFile gzfile(file_name);
-      if (!gzfile) {
-        LOG(ERROR) << "Compress: Can't open file: " << file_name;
-        return false;
-      }
-      gzfile.write(strm);
-      if (!gzfile) {
-        LOG(ERROR) << "Compress: Can't write to file: " << file_name;
-        return false;
-      }
-    }
-  } else if (file_name.empty()) {
-    Compress(fst, std::cout);
-  } else {
-    std::ofstream strm(file_name,
-                             std::ios_base::out | std::ios_base::binary);
-    if (!strm) {
-      LOG(ERROR) << "Compress: Can't open file: " << file_name;
-      return false;
-    }
-    Compress(fst, strm);
-  }
-  return true;
-}
-
-template <class Arc>
-void Decompress(std::istream &strm, const string &source,
-                MutableFst<Arc> *fst) {
-  Compressor<Arc> comp;
-  comp.Decompress(strm, source, fst);
-}
-
-// Returns true on success.
-template <class Arc>
-bool Decompress(const string &file_name, MutableFst<Arc> *fst,
-                const bool gzip = false) {
-  if (gzip) {
-    if (file_name.empty()) {
-      IGzFile gzfile(fileno(stdin));
-      Decompress(*gzfile.read(), "stdin", fst);
-      if (!gzfile) {
-        LOG(ERROR) << "Decompress: Can't read from file: stdin";
-        return false;
-      }
-    } else {
-      IGzFile gzfile(file_name);
-      if (!gzfile) {
-        LOG(ERROR) << "Decompress: Can't open file: " << file_name;
-        return false;
-      }
-      Decompress(*gzfile.read(), file_name, fst);
-      if (!gzfile) {
-        LOG(ERROR) << "Decompress: Can't read from file: " << file_name;
-        return false;
-      }
-    }
-  } else if (file_name.empty()) {
-    Decompress(std::cin, "stdin", fst);
-  } else {
-    std::ifstream strm(file_name,
-                            std::ios_base::in | std::ios_base::binary);
-    if (!strm) {
-      LOG(ERROR) << "Decompress: Can't open file: " << file_name;
-      return false;
-    }
-    Decompress(strm, file_name, fst);
-  }
-  return true;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_COMPRESS_COMPRESS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/elias.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/elias.h
deleted file mode 100644
index 6c576d33..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/elias.h
+++ /dev/null
@@ -1,97 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Compresses and decompresses unweighted FSTs.
-
-#ifndef FST_EXTENSIONS_COMPRESS_ELIAS_H_
-#define FST_EXTENSIONS_COMPRESS_ELIAS_H_
-
-#include <stack>
-#include <vector>
-
-#include <fst/compat.h>
-namespace fst {
-
-template <class Var>
-class Elias {
- public:
-  // Gamma encoding is a subroutine for Delta encoding
-  static void GammaEncode(const Var &input, std::vector<bool> *code);
-
-  // Elias Delta encoding for a single integer
-  static void DeltaEncode(const Var &input, std::vector<bool> *code);
-
-  // Batch decoding of a set of integers
-  static void BatchDecode(const std::vector<bool> &input,
-                          std::vector<Var> *output);
-};
-
-template <class Var>
-void Elias<Var>::GammaEncode(const Var &input, std::vector<bool> *code) {
-  Var input_copy = input;
-  std::stack<bool> reverse_code;
-  while (input_copy > 0) {
-    reverse_code.push(input_copy % 2);
-    input_copy = input_copy / 2;
-  }
-  for (Var auxvar = 0; auxvar < reverse_code.size() - 1; auxvar++)
-    code->push_back(0);
-  while (reverse_code.empty() != 1) {
-    code->push_back(reverse_code.top());
-    reverse_code.pop();
-  }
-}
-template <class Var>
-void Elias<Var>::DeltaEncode(const Var &input, std::vector<bool> *code) {
-  Var input_copy = input + 1;
-  std::stack<bool> reverse_remainder;
-  Var auxvar = 0;
-  while (input_copy != 0) {
-    reverse_remainder.push(input_copy % 2);
-    input_copy = input_copy / 2;
-    auxvar = auxvar + 1;
-  }
-  GammaEncode(auxvar, code);
-  reverse_remainder.pop();
-  while (reverse_remainder.empty() != 1) {
-    code->push_back(reverse_remainder.top());
-    reverse_remainder.pop();
-  }
-}
-
-template <class Var>
-void Elias<Var>::BatchDecode(const std::vector<bool> &input,
-                             std::vector<Var> *output) {
-  Var lead_zeros = 0;
-  Var remainder_bits = 0;
-  Var current_word = 1;
-  Var value = 1;
-  std::vector<bool>::const_iterator it = input.begin();
-  while (it != input.end()) {
-    lead_zeros = 0;
-    remainder_bits = 0;
-    current_word = 1;
-    value = 1;
-    while (*it != 1) {
-      it++;
-      lead_zeros++;
-    }
-    it++;
-    while (lead_zeros > 0) {
-      lead_zeros--;
-      current_word = 2 * current_word + *it;
-      it++;
-    }
-    current_word--;
-    while (current_word > 0) {
-      value = 2 * value + *it;
-      current_word--;
-      it++;
-    }
-    output->push_back(value - 1);
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_COMPRESS_ELIAS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/gzfile.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/gzfile.h
deleted file mode 100644
index 39c4f9bb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/gzfile.h
+++ /dev/null
@@ -1,126 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Resource handles for gzip files written to or read from stringstreams. These
-// are necessary to provide the compression routines with streams reading from
-// or writing to compressed files (or the UNIX standard streams), and are not
-// intended for general use.
-
-#ifndef FST_EXTENSIONS_COMPRESS_GZFILE_H_
-#define FST_EXTENSIONS_COMPRESS_GZFILE_H_
-
-#include <iostream>
-#include <memory>
-#include <sstream>
-#include <string>
-
-#include <fst/compat.h>
-#include <fst/fst.h>
-#include <zlib.h>
-
-using std::stringstream;
-using std::unique_ptr;
-
-namespace fst {
-
-// Gives the zlib gzFile type an OO-like interface. String inputs are all
-// C-style strings. The caller is responsible to get the file modes appropriate
-// for the IO methods being called, and for opening the file descriptors
-// if that constructor is used. The ! operator can be used to check for errors
-// after construction or read/writing.
-class GzFile {
- public:
-  GzFile(const char *filename, const char *mode)
-      : gzfile_(gzopen(filename, mode)), error_(check_handle()) {
-  }
-
-  // The caller is responsible to ensure the corresponding FD is open and has
-  // the needed modes ("r" for reading, "w" or "a" for writing).
-  explicit GzFile(const int fd, const char *mode)
-      : gzfile_(gzdopen(fd, mode)), error_(check_handle()), close_me_(false) {
-  }
-
-  // If the instance was constructed from an FD, flush the buffer; otherwise,
-  // close the file, which flushes the buffer as a side-effect.
-  ~GzFile() { close_me_ ? gzclose(gzfile_) : gzflush(gzfile_, Z_FINISH); }
-
-  inline bool operator!() const { return error_; }
-
-  // Returns false on EOF and sets error if short read does not reach an EOF.
-  int read(void *buf, unsigned int size) {
-    int bytes_read = gzread(gzfile_, buf, size);
-    if ((bytes_read < size) && !gzeof(gzfile_)) error_ = true;
-    return bytes_read;
-  }
-
-  // Sets error on short writes.
-  void write(const char *buf, unsigned int size) {
-    if (gzwrite(gzfile_, buf, size) != size) error_ = true;
-  }
-
- private:
-  // gzopen and gzdopen signal failure by returning null.
-  bool check_handle() { return gzfile_ == nullptr; }
-
-  gzFile gzfile_ = nullptr;
-  bool error_ = false;
-  bool close_me_ = false;
-};
-
-// Resource handle for writing stringstream to GzFile.
-class OGzFile {
- public:
-  explicit OGzFile(const string &filename) : OGzFile(filename.c_str()) {}
-
-  explicit OGzFile(const char *filename) : gz_(GzFile(filename, mode_)) {}
-
-  explicit OGzFile(const int fd) : gz_(GzFile(fd, mode_)) {}
-
-  inline bool operator!() const { return !gz_; }
-
-  void write(const stringstream &ssbuf) {
-    string sbuf = ssbuf.str();
-    gz_.write(sbuf.data(), sbuf.size());
-  }
-
- private:
-  GzFile gz_;
-  static constexpr auto &mode_ = "wb";
-};
-
-// Resource handle for reading stringstream from GzFile.
-class IGzFile {
- public:
-  explicit IGzFile(const string &filename) : IGzFile(filename.c_str()) {}
-
-  explicit IGzFile(const char *filename) : gz_(GzFile(filename, mode_)) {}
-
-  explicit IGzFile(const int fd) : gz_(GzFile(fd, mode_)) {}
-
-  inline bool operator!() const { return !gz_; }
-
-  // This is a great case for "move", but GCC 4 is missing the C+11 standard
-  // move constructor for stringstream, so a unique_ptr is the next best thing.
-  unique_ptr<stringstream> read() {
-    char buf[bufsize_];
-    unique_ptr<stringstream> sstrm(new stringstream);
-    // We always read at least once, and the result of the last read is always
-    // pushed onto the stringstream. We use the "write" member because << onto
-    // a stringstream stops at the null byte, which might be data!
-    int bytes_read;
-    while ((bytes_read = gz_.read(buf, bufsize_)) == bufsize_)
-      sstrm->write(buf, bufsize_);
-    sstrm->write(buf, bytes_read);
-    return sstrm;
-  }
-
- private:
-  GzFile gz_;
-  static constexpr auto &mode_ = "rb";
-  // This is the same size as the default internal buffer for zlib.
-  static const size_t bufsize_ = 8192;
-};
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_COMPRESS_GZFILE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/randmod.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/randmod.h
deleted file mode 100644
index de2b0281..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/compress/randmod.h
+++ /dev/null
@@ -1,101 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Generates a random FST according to a class-specific transition model.
-
-#ifndef FST_EXTENSIONS_COMPRESS_RANDMOD_H_
-#define FST_EXTENSIONS_COMPRESS_RANDMOD_H_
-
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/mutable-fst.h>
-
-namespace fst {
-
-template <class Arc, class G>
-class RandMod {
- public:
-  typedef typename Arc::StateId StateId;
-  typedef typename Arc::Label Label;
-  typedef typename Arc::Weight Weight;
-
-  // Generates random FST with 'nstates' with 'nclasses' in the probability
-  // generation model, and 'nlabels' in the alphabet. If 'trans' = true, then
-  // a transducer is generated; iff 'generate_' is non-null, the output is
-  // randomly weighted.
-  RandMod(StateId nstates, StateId nclasses, Label nlabels, bool trans,
-          const G *generate)
-      : nstates_(nstates),
-        nclasses_(nclasses),
-        nlabels_(nlabels),
-        trans_(trans),
-        generate_(generate) {
-    for (StateId s = 0; s < nstates; ++s) {
-      classes_.push_back(rand() % nclasses);  // NOLINT
-    }
-  }
-
-  // Generates a random FST according to a class-specific transition model
-  void Generate(StdMutableFst *fst) {
-    StateId start = rand() % nstates_;  // NOLINT
-    fst->DeleteStates();
-    for (StateId s = 0; s < nstates_; ++s) {
-      fst->AddState();
-      if (s == start) fst->SetStart(start);
-      for (StateId n = 0; n <= nstates_; ++n) {
-        Arc arc;
-        StateId d = n == nstates_ ? kNoStateId : n;
-        if (!RandArc(s, d, &arc)) continue;
-        if (d == kNoStateId) {  // A super-final transition?
-          fst->SetFinal(s, arc.weight);
-        } else {
-          fst->AddArc(s, arc);
-        }
-      }
-    }
-  }
-
- private:
-  // Generates a transition from s to d. If d == kNoStateId, a superfinal
-  // transition is generated. Returns false if no transition generated.
-  bool RandArc(StateId s, StateId d, Arc *arc) {
-    StateId sclass = classes_[s];
-    StateId dclass = d != kNoStateId ? classes_[d] : 0;
-
-    int r = sclass + dclass + 2;
-    if ((rand() % r) != 0)  // NOLINT
-      return false;
-
-    arc->nextstate = d;
-
-    Label ilabel = kNoLabel;
-    Label olabel = kNoLabel;
-    if (d != kNoStateId) {
-      ilabel = (dclass % nlabels_) + 1;
-      if (trans_)
-        olabel = (sclass % nlabels_) + 1;
-      else
-        olabel = ilabel;
-    }
-
-    Weight weight = Weight::One();
-    if (generate_) weight = (*generate_)();
-
-    arc->ilabel = ilabel;
-    arc->olabel = olabel;
-    arc->weight = weight;
-    return true;
-  }
-
-  StateId nstates_;
-  StateId nclasses_;
-  Label nlabels_;
-  bool trans_;
-  const G *generate_;
-  std::vector<StateId> classes_;
-};
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_COMPRESS_RANDMOD_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/compile-strings.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/compile-strings.h
deleted file mode 100644
index 96271c17..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/compile-strings.h
+++ /dev/null
@@ -1,260 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_FAR_COMPILE_STRINGS_H_
-#define FST_EXTENSIONS_FAR_COMPILE_STRINGS_H_
-
-#ifndef _MSC_VER
-#include <libgen.h>
-#else
-#include <fst/compat.h>
-#endif
-
-#include <fstream>
-#include <istream>
-#include <string>
-#include <vector>
-
-#include <fst/extensions/far/far.h>
-#include <fstream>
-#include <fst/string.h>
-
-namespace fst {
-
-// Constructs a reader that provides FSTs from a file (stream) either on a
-// line-by-line basis or on a per-stream basis. Note that the freshly
-// constructed reader is already set to the first input.
-//
-// Sample usage:
-//
-//   for (StringReader<Arc> reader(...); !reader.Done(); reader.Next()) {
-//     auto *fst = reader.GetVectorFst();
-//   }
-template <class Arc>
-class StringReader {
- public:
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-
-  enum EntryType { LINE = 1, FILE = 2 };
-
-  StringReader(std::istream &istrm, const string &source, EntryType entry_type,
-               StringTokenType token_type, bool allow_negative_labels,
-               const SymbolTable *syms = nullptr,
-               Label unknown_label = kNoStateId)
-      : nline_(0),
-        istrm_(istrm),
-        source_(source),
-        entry_type_(entry_type),
-        token_type_(token_type),
-        symbols_(syms),
-        done_(false),
-        compiler_(token_type, syms, unknown_label, allow_negative_labels) {
-    Next();  // Initialize the reader to the first input.
-  }
-
-  bool Done() { return done_; }
-
-  void Next() {
-    VLOG(1) << "Processing source " << source_ << " at line " << nline_;
-    if (!istrm_) {  // We're done if we have no more input.
-      done_ = true;
-      return;
-    }
-    if (entry_type_ == LINE) {
-      getline(istrm_, content_);
-      ++nline_;
-    } else {
-      content_.clear();
-      string line;
-      while (getline(istrm_, line)) {
-        ++nline_;
-        content_.append(line);
-        content_.append("\n");
-      }
-    }
-    if (!istrm_ && content_.empty())  // We're also done if we read off all the
-      done_ = true;                   // whitespace at the end of a file.
-  }
-
-  VectorFst<Arc> *GetVectorFst(bool keep_symbols = false) {
-    std::unique_ptr<VectorFst<Arc>> fst(new VectorFst<Arc>());
-    if (keep_symbols) {
-      fst->SetInputSymbols(symbols_);
-      fst->SetOutputSymbols(symbols_);
-    }
-    if (compiler_(content_, fst.get())) {
-      return fst.release();
-    } else {
-      return nullptr;
-    }
-  }
-
-  CompactStringFst<Arc> *GetCompactFst(bool keep_symbols = false) {
-    std::unique_ptr<CompactStringFst<Arc>> fst;
-    if (keep_symbols) {
-      VectorFst<Arc> tmp;
-      tmp.SetInputSymbols(symbols_);
-      tmp.SetOutputSymbols(symbols_);
-      fst.reset(new CompactStringFst<Arc>(tmp));
-    } else {
-      fst.reset(new CompactStringFst<Arc>());
-    }
-    if (compiler_(content_, fst.get())) {
-      return fst.release();
-    } else {
-      return nullptr;
-    }
-  }
-
- private:
-  size_t nline_;
-  std::istream &istrm_;
-  string source_;
-  EntryType entry_type_;
-  StringTokenType token_type_;
-  const SymbolTable *symbols_;
-  bool done_;
-  StringCompiler<Arc> compiler_;
-  string content_;  // The actual content of the input stream's next FST.
-
-  StringReader(const StringReader &) = delete;
-  StringReader &operator=(const StringReader &) = delete;
-};
-
-// Computes the minimal length required to encode each line number as a decimal
-// number.
-int KeySize(const char *filename);
-
-template <class Arc>
-void FarCompileStrings(const std::vector<string> &in_fnames,
-                       const string &out_fname, const string &fst_type,
-                       const FarType &far_type, int32 generate_keys,
-                       FarEntryType fet, FarTokenType tt,
-                       const string &symbols_fname,
-                       const string &unknown_symbol, bool keep_symbols,
-                       bool initial_symbols, bool allow_negative_labels,
-                       const string &key_prefix, const string &key_suffix) {
-  typename StringReader<Arc>::EntryType entry_type;
-  if (fet == FET_LINE) {
-    entry_type = StringReader<Arc>::LINE;
-  } else if (fet == FET_FILE) {
-    entry_type = StringReader<Arc>::FILE;
-  } else {
-    FSTERROR() << "FarCompileStrings: Unknown entry type";
-    return;
-  }
-  StringTokenType token_type;
-  if (tt == FTT_SYMBOL) {
-    token_type = StringTokenType::SYMBOL;
-  } else if (tt == FTT_BYTE) {
-    token_type = StringTokenType::BYTE;
-  } else if (tt == FTT_UTF8) {
-    token_type = StringTokenType::UTF8;
-  } else {
-    FSTERROR() << "FarCompileStrings: Unknown token type";
-    return;
-  }
-  bool compact;
-  if (fst_type.empty() || (fst_type == "vector")) {
-    compact = false;
-  } else if (fst_type == "compact") {
-    compact = true;
-  } else {
-    FSTERROR() << "FarCompileStrings: Unknown FST type: " << fst_type;
-    return;
-  }
-  std::unique_ptr<const SymbolTable> syms;
-  typename Arc::Label unknown_label = kNoLabel;
-  if (!symbols_fname.empty()) {
-    const SymbolTableTextOptions opts(allow_negative_labels);
-    syms.reset(SymbolTable::ReadText(symbols_fname, opts));
-    if (!syms) {
-      LOG(ERROR) << "FarCompileStrings: Error reading symbol table: "
-                 << symbols_fname;
-      return;
-    }
-    if (!unknown_symbol.empty()) {
-      unknown_label = syms->Find(unknown_symbol);
-      if (unknown_label == kNoLabel) {
-        FSTERROR() << "FarCompileStrings: Label \"" << unknown_label
-                   << "\" missing from symbol table: " << symbols_fname;
-        return;
-      }
-    }
-  }
-  std::unique_ptr<FarWriter<Arc>> far_writer(
-      FarWriter<Arc>::Create(out_fname, far_type));
-  if (!far_writer) return;
-  int n = 0;
-  for (const auto &in_fname : in_fnames) {
-    if (generate_keys == 0 && in_fname.empty()) {
-      FSTERROR() << "FarCompileStrings: Read from a file instead of stdin or"
-                 << " set the --generate_keys flags.";
-      return;
-    }
-    int key_size =
-        generate_keys ? generate_keys : (entry_type == StringReader<Arc>::FILE
-                                             ? 1 : KeySize(in_fname.c_str()));
-    std::ifstream fstrm;
-    if (!in_fname.empty()) {
-      fstrm.open(in_fname);
-      if (!fstrm) {
-        FSTERROR() << "FarCompileStrings: Can't open file: " << in_fname;
-        return;
-      }
-    }
-    std::istream &istrm = fstrm.is_open() ? fstrm : std::cin;
-    bool keep_syms = keep_symbols;
-    for (StringReader<Arc> reader(
-             istrm, in_fname.empty() ? "stdin" : in_fname, entry_type,
-             token_type, allow_negative_labels, syms.get(), unknown_label);
-         !reader.Done(); reader.Next()) {
-      ++n;
-      std::unique_ptr<const Fst<Arc>> fst;
-      if (compact) {
-        fst.reset(reader.GetCompactFst(keep_syms));
-      } else {
-        fst.reset(reader.GetVectorFst(keep_syms));
-      }
-      if (initial_symbols) keep_syms = false;
-      if (!fst) {
-        FSTERROR() << "FarCompileStrings: Compiling string number " << n
-                   << " in file " << in_fname << " failed with token_type = "
-                   << (tt == FTT_BYTE
-                           ? "byte"
-                           : (tt == FTT_UTF8
-                                  ? "utf8"
-                                  : (tt == FTT_SYMBOL ? "symbol" : "unknown")))
-                   << " and entry_type = "
-                   << (fet == FET_LINE
-                           ? "line"
-                           : (fet == FET_FILE ? "file" : "unknown"));
-        return;
-      }
-      std::ostringstream keybuf;
-      keybuf.width(key_size);
-      keybuf.fill('0');
-      keybuf << n;
-      string key;
-      if (generate_keys > 0) {
-        key = keybuf.str();
-      } else {
-        auto *filename = new char[in_fname.size() + 1];
-        strcpy(filename, in_fname.c_str());
-        key = basename(filename);
-        if (entry_type != StringReader<Arc>::FILE) {
-          key += "-";
-          key += keybuf.str();
-        }
-        delete[] filename;
-      }
-      far_writer->Add(key_prefix + key + key_suffix, *fst);
-    }
-    if (generate_keys == 0) n = 0;
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_COMPILE_STRINGS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/create.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/create.h
deleted file mode 100644
index 31c95896..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/create.h
+++ /dev/null
@@ -1,46 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Creates a finite-state archive from component FSTs.
-
-#ifndef FST_EXTENSIONS_FAR_CREATE_H_
-#define FST_EXTENSIONS_FAR_CREATE_H_
-
-#include <sstream>
-#include <string>
-#include <vector>
-
-#include <fst/extensions/far/far.h>
-
-namespace fst {
-
-template <class Arc>
-void FarCreate(const std::vector<string> &in_fnames, const string &out_fname,
-               const int32 generate_keys, const FarType &far_type,
-               const string &key_prefix, const string &key_suffix) {
-  std::unique_ptr<FarWriter<Arc>> far_writer(
-      FarWriter<Arc>::Create(out_fname, far_type));
-  if (!far_writer) return;
-  for (size_t i = 0; i < in_fnames.size(); ++i) {
-    std::unique_ptr<Fst<Arc>> ifst(Fst<Arc>::Read(in_fnames[i]));
-    if (!ifst) return;
-    string key;
-    if (generate_keys > 0) {
-      std::ostringstream keybuf;
-      keybuf.width(generate_keys);
-      keybuf.fill('0');
-      keybuf << i + 1;
-      key = keybuf.str();
-    } else {
-      auto *filename = new char[in_fnames[i].size() + 1];
-      strcpy(filename, in_fnames[i].c_str());
-      key = basename(filename);
-      delete[] filename;
-    }
-    far_writer->Add(key_prefix + key + key_suffix, *ifst);
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_CREATE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/equal.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/equal.h
deleted file mode 100644
index 04511d21..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/equal.h
+++ /dev/null
@@ -1,69 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_FAR_EQUAL_H_
-#define FST_EXTENSIONS_FAR_EQUAL_H_
-
-#include <memory>
-#include <string>
-
-#include <fst/extensions/far/far.h>
-#include <fst/equal.h>
-
-namespace fst {
-
-template <class Arc>
-bool FarEqual(const string &filename1, const string &filename2,
-              float delta = kDelta, const string &begin_key = string(),
-              const string &end_key = string()) {
-  std::unique_ptr<FarReader<Arc>> reader1(FarReader<Arc>::Open(filename1));
-  if (!reader1) {
-    LOG(ERROR) << "FarEqual: Could not open FAR file " << filename1;
-    return false;
-  }
-  std::unique_ptr<FarReader<Arc>> reader2(FarReader<Arc>::Open(filename2));
-  if (!reader2) {
-    LOG(ERROR) << "FarEqual: Could not open FAR file " << filename2;
-    return false;
-  }
-  if (!begin_key.empty()) {
-    bool find_begin1 = reader1->Find(begin_key);
-    bool find_begin2 = reader2->Find(begin_key);
-    if (!find_begin1 || !find_begin2) {
-      bool ret = !find_begin1 && !find_begin2;
-      if (!ret) {
-        LOG(ERROR) << "FarEqual: Key " << begin_key << " missing from "
-                   << (find_begin1 ? "second" : "first") << " archive";
-      }
-      return ret;
-    }
-  }
-  for (; !reader1->Done() && !reader2->Done();
-       reader1->Next(), reader2->Next()) {
-    const auto &key1 = reader1->GetKey();
-    const auto &key2 = reader2->GetKey();
-    if (!end_key.empty() && end_key < key1 && end_key < key2) {
-      return true;
-    }
-    if (key1 != key2) {
-      LOG(ERROR) << "FarEqual: Mismatched keys " << key1 << " and " << key2;
-      return false;
-    }
-    if (!Equal(*(reader1->GetFst()), *(reader2->GetFst()), delta)) {
-      LOG(ERROR) << "FarEqual: FSTs for key " << key1 << " are not equal";
-      return false;
-    }
-  }
-  if (!reader1->Done() || !reader2->Done()) {
-    LOG(ERROR) << "FarEqual: Key "
-               << (reader1->Done() ? reader2->GetKey() : reader1->GetKey())
-               << " missing from " << (reader2->Done() ? "first" : "second")
-               << " archive";
-    return false;
-  }
-  return true;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_EQUAL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/extract.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/extract.h
deleted file mode 100644
index 79107d60..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/extract.h
+++ /dev/null
@@ -1,117 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Extracts component FSTs from an finite-state archive.
-
-#ifndef FST_EXTENSIONS_FAR_EXTRACT_H_
-#define FST_EXTENSIONS_FAR_EXTRACT_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include <fst/extensions/far/far.h>
-
-namespace fst {
-
-template <class Arc>
-inline void FarWriteFst(const Fst<Arc> *fst, string key, string *okey,
-                        int *nrep, int32 generate_filenames, int i,
-                        const string &filename_prefix,
-                        const string &filename_suffix) {
-  if (key == *okey) {
-    ++*nrep;
-  } else {
-    *nrep = 0;
-  }
-  *okey = key;
-  string ofilename;
-  if (generate_filenames) {
-    std::ostringstream tmp;
-    tmp.width(generate_filenames);
-    tmp.fill('0');
-    tmp << i;
-    ofilename = tmp.str();
-  } else {
-    if (*nrep > 0) {
-      std::ostringstream tmp;
-      tmp << '.' << nrep;
-      key.append(tmp.str().data(), tmp.str().size());
-    }
-    ofilename = key;
-  }
-  fst->Write(filename_prefix + ofilename + filename_suffix);
-}
-
-template <class Arc>
-void FarExtract(const std::vector<string> &ifilenames, int32 generate_filenames,
-                const string &keys, const string &key_separator,
-                const string &range_delimiter, const string &filename_prefix,
-                const string &filename_suffix) {
-  std::unique_ptr<FarReader<Arc>> far_reader(
-      FarReader<Arc>::Open(ifilenames));
-  if (!far_reader) return;
-  string okey;
-  int nrep = 0;
-  std::vector<char *> key_vector;
-  // User has specified a set of FSTs to extract, where some of these may in
-  // fact be ranges.
-  if (!keys.empty()) {
-    auto *keys_cstr = new char[keys.size() + 1];
-    strcpy(keys_cstr, keys.c_str());
-    SplitToVector(keys_cstr, key_separator.c_str(), &key_vector, true);
-    int i = 0;
-    for (size_t k = 0; k < key_vector.size(); ++k, ++i) {
-      string key = key_vector[k];
-      auto *key_cstr = new char[key.size() + 1];
-      strcpy(key_cstr, key.c_str());
-      std::vector<char *> range_vector;
-      SplitToVector(key_cstr, range_delimiter.c_str(), &range_vector, false);
-      if (range_vector.size() == 1) {  // Not a range
-        if (!far_reader->Find(key)) {
-          LOG(ERROR) << "FarExtract: Cannot find key " << key;
-          return;
-        }
-        const auto *fst = far_reader->GetFst();
-        FarWriteFst(fst, key, &okey, &nrep, generate_filenames, i,
-                    filename_prefix, filename_suffix);
-      } else if (range_vector.size() == 2) {  // A legal range
-        string begin_key = range_vector[0];
-        string end_key = range_vector[1];
-        if (begin_key.empty() || end_key.empty()) {
-          LOG(ERROR) << "FarExtract: Illegal range specification " << key;
-          return;
-        }
-        if (!far_reader->Find(begin_key)) {
-          LOG(ERROR) << "FarExtract: Cannot find key " << begin_key;
-          return;
-        }
-        for (; !far_reader->Done(); far_reader->Next(), ++i) {
-          const auto &ikey = far_reader->GetKey();
-          if (end_key < ikey) break;
-          const auto *fst = far_reader->GetFst();
-          FarWriteFst(fst, ikey, &okey, &nrep, generate_filenames, i,
-                      filename_prefix, filename_suffix);
-        }
-      } else {
-        LOG(ERROR) << "FarExtract: Illegal range specification " << key;
-        return;
-      }
-      delete[] key_cstr;
-    }
-    delete[] keys_cstr;
-    return;
-  }
-  // Nothing specified, so just extracts everything.
-  for (size_t i = 1; !far_reader->Done(); far_reader->Next(), ++i) {
-    const auto &key = far_reader->GetKey();
-    const auto *fst = far_reader->GetFst();
-    FarWriteFst(fst, key, &okey, &nrep, generate_filenames, i, filename_prefix,
-                filename_suffix);
-  }
-  return;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_EXTRACT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/far-class.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/far-class.h
deleted file mode 100644
index d4fedd26..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/far-class.h
+++ /dev/null
@@ -1,270 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Scripting API support for FarReader and FarWriter.
-
-#ifndef FST_EXTENSIONS_FAR_FAR_CLASS_H_
-#define FST_EXTENSIONS_FAR_FAR_CLASS_H_
-
-#include <memory>
-#include <string>
-#include <vector>
-
-#include <fst/script/arg-packs.h>
-#include <fst/script/fstscript.h>
-#include <fst/extensions/far/far.h>
-
-
-namespace fst {
-namespace script {
-
-
-// FarReader API.
-
-// Virtual interface implemented by each concrete FarReaderImpl<A>.
-// See the FarReader interface in far.h for the exact semantics.
-class FarReaderImplBase {
- public:
-  virtual const string &ArcType() const = 0;
-  virtual bool Done() const = 0;
-  virtual bool Error() const = 0;
-  virtual const string &GetKey() const = 0;
-  virtual const FstClass *GetFstClass() const = 0;
-  virtual bool Find(const string &key) = 0;
-  virtual void Next() = 0;
-  virtual void Reset() = 0;
-  virtual FarType Type() const = 0;
-  virtual ~FarReaderImplBase() {}
-};
-
-// Templated implementation.
-template <class Arc>
-class FarReaderClassImpl : public FarReaderImplBase {
- public:
-  explicit FarReaderClassImpl(const string &filename)
-      : impl_(FarReader<Arc>::Open(filename)) {}
-
-  explicit FarReaderClassImpl(const std::vector<string> &filenames)
-      : impl_(FarReader<Arc>::Open(filenames)) {}
-
-  const string &ArcType() const final { return Arc::Type(); }
-
-  bool Done() const final { return impl_->Done(); }
-
-  bool Error() const final { return impl_->Error(); }
-
-  bool Find(const string &key) final { return impl_->Find(key); }
-
-  const FstClass *GetFstClass() const final {
-    fstc_.reset(new FstClass(*impl_->GetFst()));
-    return fstc_.get();
-  }
-
-  const string &GetKey() const final { return impl_->GetKey(); }
-
-  void Next() final { return impl_->Next(); }
-
-  void Reset() final { impl_->Reset(); }
-
-  FarType Type() const final { return impl_->Type(); }
-
-  const FarReader<Arc> *GetImpl() const { return impl_.get(); }
-
-  FarReader<Arc> *GetImpl() { return impl_.get(); }
-
- private:
-  std::unique_ptr<FarReader<Arc>> impl_;
-  mutable std::unique_ptr<FstClass> fstc_;
-};
-
-
-class FarReaderClass;
-
-using OpenFarReaderClassArgs1 =
-    WithReturnValue<FarReaderClass *, const string &>;
-
-using OpenFarReaderClassArgs2 =
-    WithReturnValue<FarReaderClass *, const std::vector<string> &>;
-
-// Untemplated user-facing class holding a templated pimpl.
-class FarReaderClass {
- public:
-  const string &ArcType() const { return impl_->ArcType(); }
-
-  bool Done() const { return impl_->Done(); }
-
-  // Returns True if the impl is null (i.e., due to read failure).
-  // Attempting to call any other function will result in null dereference.
-  bool Error() const { return (impl_) ? impl_->Error() : true; }
-
-  bool Find(const string &key) { return impl_->Find(key); }
-
-  const FstClass *GetFstClass() const { return impl_->GetFstClass(); }
-
-  const string &GetKey() const { return impl_->GetKey(); }
-
-  void Next() { impl_->Next(); }
-
-  void Reset() { impl_->Reset(); }
-
-  FarType Type() const { return impl_->Type(); }
-
-  template <class Arc>
-  const FarReader<Arc> *GetFarReader() const {
-    if (Arc::Type() != ArcType()) return nullptr;
-    const FarReaderClassImpl<Arc> *typed_impl =
-        static_cast<FarReaderClassImpl<Arc> *>(impl_.get());
-    return typed_impl->GetImpl();
-  }
-
-  template <class Arc>
-  FarReader<Arc> *GetFarReader() {
-    if (Arc::Type() != ArcType()) return nullptr;
-    FarReaderClassImpl<Arc> *typed_impl =
-        static_cast<FarReaderClassImpl<Arc> *>(impl_.get());
-    return typed_impl->GetImpl();
-  }
-
-  template <class Arc>
-  friend void OpenFarReaderClass(OpenFarReaderClassArgs1 *args);
-
-  template <class Arc>
-  friend void OpenFarReaderClass(OpenFarReaderClassArgs2 *args);
-
-  // Defined in the CC.
-
-  static FarReaderClass *Open(const string &filename);
-
-  static FarReaderClass *Open(const std::vector<string> &filenames);
-
- private:
-  template <class Arc>
-  explicit FarReaderClass(FarReaderClassImpl<Arc> *impl) : impl_(impl) {}
-
-  std::unique_ptr<FarReaderImplBase> impl_;
-};
-
-// These exist solely for registration purposes; users should call the
-// static method FarReaderClass::Open instead.
-
-template <class Arc>
-void OpenFarReaderClass(OpenFarReaderClassArgs1 *args) {
-  args->retval = new FarReaderClass(new FarReaderClassImpl<Arc>(args->args));
-}
-
-template <class Arc>
-void OpenFarReaderClass(OpenFarReaderClassArgs2 *args) {
-  args->retval = new FarReaderClass(new FarReaderClassImpl<Arc>(args->args));
-}
-
-// FarWriter API.
-
-// Virtual interface implemented by each concrete FarWriterImpl<A>.
-class FarWriterImplBase {
- public:
-  // Unlike the lower-level library, this returns a boolean to signal failure
-  // due to non-conformant arc types.
-  virtual bool Add(const string &key, const FstClass &fst) = 0;
-  virtual const string &ArcType() const = 0;
-  virtual bool Error() const = 0;
-  virtual FarType Type() const = 0;
-  virtual ~FarWriterImplBase() {}
-};
-
-
-// Templated implementation.
-template <class Arc>
-class FarWriterClassImpl : public FarWriterImplBase {
- public:
-  explicit FarWriterClassImpl(const string &filename,
-                              FarType type = FAR_DEFAULT)
-      : impl_(FarWriter<Arc>::Create(filename, type)) {}
-
-  bool Add(const string &key, const FstClass &fst) final {
-    if (ArcType() != fst.ArcType()) {
-      FSTERROR() << "Cannot write FST with " << fst.ArcType() << " arcs to "
-                 << "FAR with " << ArcType() << " arcs";
-      return false;
-    }
-    impl_->Add(key, *(fst.GetFst<Arc>()));
-    return true;
-  }
-
-  const string &ArcType() const final { return Arc::Type(); }
-
-  bool Error() const final { return impl_->Error(); }
-
-  FarType Type() const final { return impl_->Type(); }
-
-  const FarWriter<Arc> *GetImpl() const { return impl_.get(); }
-
-  FarWriter<Arc> *GetImpl() { return impl_.get(); }
-
- private:
-  std::unique_ptr<FarWriter<Arc>> impl_;
-};
-
-
-class FarWriterClass;
-
-using CreateFarWriterClassInnerArgs = std::pair<const string &, FarType>;
-
-using CreateFarWriterClassArgs =
-    WithReturnValue<FarWriterClass *, CreateFarWriterClassInnerArgs>;
-
-// Untemplated user-facing class holding a templated pimpl.
-class FarWriterClass {
- public:
-  static FarWriterClass *Create(const string &filename, const string &arc_type,
-                                FarType type = FAR_DEFAULT);
-
-  bool Add(const string &key, const FstClass &fst) {
-    return impl_->Add(key, fst);
-  }
-
-  // Returns True if the impl is null (i.e., due to construction failure).
-  // Attempting to call any other function will result in null dereference.
-  bool Error() const { return (impl_) ? impl_->Error() : true; }
-
-  const string &ArcType() const { return impl_->ArcType(); }
-
-  FarType Type() const { return impl_->Type(); }
-
-  template <class Arc>
-  const FarWriter<Arc> *GetFarWriter() const {
-    if (Arc::Type() != ArcType()) return nullptr;
-    const FarWriterClassImpl<Arc> *typed_impl =
-        static_cast<FarWriterClassImpl<Arc> *>(impl_.get());
-    return typed_impl->GetImpl();
-  }
-
-  template <class Arc>
-  FarWriter<Arc> *GetFarWriter() {
-    if (Arc::Type() != ArcType()) return nullptr;
-    FarWriterClassImpl<Arc> *typed_impl =
-        static_cast<FarWriterClassImpl<Arc> *>(impl_.get());
-    return typed_impl->GetImpl();
-  }
-
-  template <class Arc>
-  friend void CreateFarWriterClass(CreateFarWriterClassArgs *args);
-
- private:
-  template <class Arc>
-  explicit FarWriterClass(FarWriterClassImpl<Arc> *impl) : impl_(impl) {}
-
-  std::unique_ptr<FarWriterImplBase> impl_;
-};
-
-// This exists solely for registration purposes; users should call the
-// static method FarWriterClass::Create instead.
-template <class Arc>
-void CreateFarWriterClass(CreateFarWriterClassArgs *args) {
-  args->retval = new FarWriterClass(new FarWriterClassImpl<Arc>(
-      std::get<0>(args->args), std::get<1>(args->args)));
-}
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_FAR_CLASS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/far.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/far.h
deleted file mode 100644
index 806f126f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/far.h
+++ /dev/null
@@ -1,520 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Finite-State Transducer (FST) archive classes.
-
-#ifndef FST_EXTENSIONS_FAR_FAR_H_
-#define FST_EXTENSIONS_FAR_FAR_H_
-
-#include <iostream>
-#include <sstream>
-
-#include <fst/log.h>
-#include <fst/extensions/far/stlist.h>
-#include <fst/extensions/far/sttable.h>
-#include <fst/fst.h>
-#include <fst/vector-fst.h>
-#include <fstream>
-
-namespace fst {
-
-enum FarEntryType { FET_LINE, FET_FILE };
-
-enum FarTokenType { FTT_SYMBOL, FTT_BYTE, FTT_UTF8 };
-
-inline bool IsFst(const string &filename) {
-  std::ifstream strm(filename, std::ios_base::in | std::ios_base::binary);
-  if (!strm) return false;
-  return IsFstHeader(strm, filename);
-}
-
-// FST archive header class
-class FarHeader {
- public:
-  const string &ArcType() const { return arctype_; }
-
-  const string &FarType() const { return fartype_; }
-
-  bool Read(const string &filename) {
-    FstHeader fsthdr;
-    if (filename.empty()) {
-      // Header reading unsupported on stdin. Assumes STList and StdArc.
-      fartype_ = "stlist";
-      arctype_ = "standard";
-      return true;
-    } else if (IsSTTable(filename)) {  // Checks if STTable.
-      ReadSTTableHeader(filename, &fsthdr);
-      fartype_ = "sttable";
-      arctype_ = fsthdr.ArcType().empty() ? "unknown" : fsthdr.ArcType();
-      return true;
-    } else if (IsSTList(filename)) {  // Checks if STList.
-      ReadSTListHeader(filename, &fsthdr);
-      fartype_ = "stlist";
-      arctype_ = fsthdr.ArcType().empty() ? "unknown" : fsthdr.ArcType();
-      return true;
-    } else if (IsFst(filename)) {  // Checks if FST.
-      std::ifstream istrm(filename,
-                               std::ios_base::in | std::ios_base::binary);
-      fsthdr.Read(istrm, filename);
-      fartype_ = "fst";
-      arctype_ = fsthdr.ArcType().empty() ? "unknown" : fsthdr.ArcType();
-      return true;
-    }
-    return false;
-  }
-
- private:
-  string fartype_;
-  string arctype_;
-};
-
-enum FarType {
-  FAR_DEFAULT = 0,
-  FAR_STTABLE = 1,
-  FAR_STLIST = 2,
-  FAR_FST = 3,
-};
-
-// This class creates an archive of FSTs.
-template <class A>
-class FarWriter {
- public:
-  using Arc = A;
-
-  // Creates a new (empty) FST archive; returns null on error.
-  static FarWriter *Create(const string &filename, FarType type = FAR_DEFAULT);
-
-  // Adds an FST to the end of an archive. Keys must be non-empty and
-  // in lexicographic order. FSTs must have a suitable write method.
-  virtual void Add(const string &key, const Fst<Arc> &fst) = 0;
-
-  virtual FarType Type() const = 0;
-
-  virtual bool Error() const = 0;
-
-  virtual ~FarWriter() {}
-
- protected:
-  FarWriter() {}
-};
-
-// This class iterates through an existing archive of FSTs.
-template <class A>
-class FarReader {
- public:
-  using Arc = A;
-
-  // Opens an existing FST archive in a single file; returns null on error.
-  // Sets current position to the beginning of the achive.
-  static FarReader *Open(const string &filename);
-
-  static FarReader *Open(std::unique_ptr<std::istream> stream);
-  static FarReader *Open(std::vector<std::unique_ptr<std::istream>> streams);
-
-  // Opens an existing FST archive in multiple files; returns null on error.
-  // Sets current position to the beginning of the achive.
-  static FarReader *Open(const std::vector<string> &filenames);
-
-  // Resets current position to beginning of archive.
-  virtual void Reset() = 0;
-
-  // Sets current position to first entry >= key.  Returns true if a match.
-  virtual bool Find(const string &key) = 0;
-
-  // Current position at end of archive?
-  virtual bool Done() const = 0;
-
-  // Move current position to next FST.
-  virtual void Next() = 0;
-
-  // Returns key at the current position. This reference is invalidated if
-  // the current position in the archive is changed.
-  virtual const string &GetKey() const = 0;
-
-  // Returns pointer to FST at the current position. This is invalidated if
-  // the current position in the archive is changed.
-  virtual const Fst<Arc> *GetFst() const = 0;
-
-  virtual FarType Type() const = 0;
-
-  virtual bool Error() const = 0;
-
-  virtual ~FarReader() {}
-
- protected:
-  FarReader() {}
-};
-
-template <class Arc>
-class FstWriter {
- public:
-  void operator()(std::ostream &strm, const Fst<Arc> &fst) const {
-    fst.Write(strm, FstWriteOptions());
-  }
-};
-
-template <class A>
-class STTableFarWriter : public FarWriter<A> {
- public:
-  using Arc = A;
-
-  static STTableFarWriter *Create(const string &filename) {
-    auto *writer = STTableWriter<Fst<Arc>, FstWriter<Arc>>::Create(filename);
-    return new STTableFarWriter(writer);
-  }
-
-  void Add(const string &key, const Fst<Arc> &fst) final {
-    writer_->Add(key, fst);
-  }
-
-  FarType Type() const final { return FAR_STTABLE; }
-
-  bool Error() const final { return writer_->Error(); }
-
- private:
-  explicit STTableFarWriter(STTableWriter<Fst<Arc>, FstWriter<Arc>> *writer)
-      : writer_(writer) {}
-
-  std::unique_ptr<STTableWriter<Fst<Arc>, FstWriter<Arc>>> writer_;
-};
-
-template <class A>
-class STListFarWriter : public FarWriter<A> {
- public:
-  using Arc = A;
-
-  static STListFarWriter *Create(const string &filename) {
-    auto *writer = STListWriter<Fst<Arc>, FstWriter<Arc>>::Create(filename);
-    return new STListFarWriter(writer);
-  }
-
-  void Add(const string &key, const Fst<Arc> &fst) final {
-    writer_->Add(key, fst);
-  }
-
-  constexpr FarType Type() const final { return FAR_STLIST; }
-
-  bool Error() const final { return writer_->Error(); }
-
- private:
-  explicit STListFarWriter(STListWriter<Fst<Arc>, FstWriter<Arc>> *writer)
-      : writer_(writer) {}
-
-  std::unique_ptr<STListWriter<Fst<Arc>, FstWriter<Arc>>> writer_;
-};
-
-template <class A>
-class FstFarWriter : public FarWriter<A> {
- public:
-  using Arc = A;
-
-  explicit FstFarWriter(const string &filename)
-      : filename_(filename), error_(false), written_(false) {}
-
-  static FstFarWriter *Create(const string &filename) {
-    return new FstFarWriter(filename);
-  }
-
-  void Add(const string &key, const Fst<A> &fst) final {
-    if (written_) {
-      LOG(WARNING) << "FstFarWriter::Add: only one FST supported,"
-                   << " subsequent entries discarded.";
-    } else {
-      error_ = !fst.Write(filename_);
-      written_ = true;
-    }
-  }
-
-  constexpr FarType Type() const final { return FAR_FST; }
-
-  bool Error() const final { return error_; }
-
-  ~FstFarWriter() {}
-
- private:
-  string filename_;
-  bool error_;
-  bool written_;
-};
-
-template <class Arc>
-FarWriter<Arc> *FarWriter<Arc>::Create(const string &filename, FarType type) {
-  switch (type) {
-    case FAR_DEFAULT:
-      if (filename.empty()) return STListFarWriter<Arc>::Create(filename);
-    case FAR_STTABLE:
-      return STTableFarWriter<Arc>::Create(filename);
-    case FAR_STLIST:
-      return STListFarWriter<Arc>::Create(filename);
-    case FAR_FST:
-      return FstFarWriter<Arc>::Create(filename);
-    default:
-      LOG(ERROR) << "FarWriter::Create: Unknown FAR type";
-      return nullptr;
-  }
-}
-
-template <class Arc>
-class FstReader {
- public:
-  Fst<Arc> *operator()(std::istream &strm) const {
-    return Fst<Arc>::Read(strm, FstReadOptions());
-  }
-};
-
-template <class A>
-class STTableFarReader : public FarReader<A> {
- public:
-  using Arc = A;
-
-  static STTableFarReader *Open(const string &filename) {
-    auto *reader = STTableReader<Fst<Arc>, FstReader<Arc>>::Open(filename);
-    if (!reader || reader->Error()) {
-      if (reader) {
-        delete reader;
-      }
-
-      return nullptr;
-    }
-    return new STTableFarReader(reader);
-  }
-
-  static STTableFarReader *Open(std::unique_ptr<std::istream> stream) {
-    auto *reader = STTableReader<Fst<Arc>, FstReader<Arc>>::Open(std::move(stream));
-    if (!reader || reader->Error()) return nullptr;
-    return new STTableFarReader(reader);
-  }
-
-  static STTableFarReader *Open(std::vector<std::unique_ptr<std::istream>> streams) {
-    auto *reader = STTableReader<Fst<Arc>, FstReader<Arc>>::Open(std::move(streams));
-    if (!reader || reader->Error()) return nullptr;
-    return new STTableFarReader(reader);
-  }
-
-  static STTableFarReader *Open(const std::vector<string> &filenames) {
-    auto *reader = STTableReader<Fst<Arc>, FstReader<Arc>>::Open(filenames);
-    if (!reader || reader->Error()) return nullptr;
-    return new STTableFarReader(reader);
-  }
-
-  void Reset() final { reader_->Reset(); }
-
-  bool Find(const string &key) final { return reader_->Find(key); }
-
-  bool Done() const final { return reader_->Done(); }
-
-  void Next() final { return reader_->Next(); }
-
-  const string &GetKey() const final { return reader_->GetKey(); }
-
-  const Fst<Arc> *GetFst() const final { return reader_->GetEntry(); }
-
-  constexpr FarType Type() const final { return FAR_STTABLE; }
-
-  bool Error() const final { return reader_->Error(); }
-
- private:
-  explicit STTableFarReader(STTableReader<Fst<Arc>, FstReader<Arc>> *reader)
-      : reader_(reader) {}
-
-  std::unique_ptr<STTableReader<Fst<Arc>, FstReader<Arc>>> reader_;
-};
-
-template <class A>
-class STListFarReader : public FarReader<A> {
- public:
-  using Arc = A;
-
-  static STListFarReader *Open(const string &filename) {
-    auto *reader = STListReader<Fst<Arc>, FstReader<Arc>>::Open(filename);
-    if (!reader || reader->Error()) {
-      if (reader) {
-        delete reader;
-      }
-      return nullptr;
-    }
-    return new STListFarReader(reader);
-  }
-
-  static STListFarReader *Open(const std::vector<string> &filenames) {
-    auto *reader = STListReader<Fst<Arc>, FstReader<Arc>>::Open(filenames);
-    if (!reader || reader->Error()) {
-      if (reader) {
-        delete reader;
-      }
-      return nullptr;
-    }
-    return new STListFarReader(reader);
-  }
-
-  void Reset() final { reader_->Reset(); }
-
-  bool Find(const string &key) final { return reader_->Find(key); }
-
-  bool Done() const final { return reader_->Done(); }
-
-  void Next() final { return reader_->Next(); }
-
-  const string &GetKey() const final { return reader_->GetKey(); }
-
-  const Fst<Arc> *GetFst() const final { return reader_->GetEntry(); }
-
-  constexpr FarType Type() const final { return FAR_STLIST; }
-
-  bool Error() const final { return reader_->Error(); }
-
- private:
-  explicit STListFarReader(STListReader<Fst<Arc>, FstReader<Arc>> *reader)
-      : reader_(reader) {}
-
-  std::unique_ptr<STListReader<Fst<Arc>, FstReader<Arc>>> reader_;
-};
-
-template <class A>
-class FstFarReader : public FarReader<A> {
- public:
-  using Arc = A;
-
-  static FstFarReader *Open(const string &filename) {
-    std::vector<string> filenames;
-    filenames.push_back(filename);
-    return new FstFarReader<Arc>(filenames);
-  }
-
-  static FstFarReader *Open(const std::vector<string> &filenames) {
-    return new FstFarReader<Arc>(filenames);
-  }
-
-  explicit FstFarReader(const std::vector<string> &filenames)
-      : keys_(filenames), has_stdin_(false), pos_(0), error_(false) {
-    std::sort(keys_.begin(), keys_.end());
-    streams_.resize(keys_.size(), 0);
-    for (size_t i = 0; i < keys_.size(); ++i) {
-      if (keys_[i].empty()) {
-        if (!has_stdin_) {
-          streams_[i] = &std::cin;
-          // sources_[i] = "stdin";
-          has_stdin_ = true;
-        } else {
-          FSTERROR() << "FstFarReader::FstFarReader: standard input should "
-                        "only appear once in the input file list";
-          error_ = true;
-          return;
-        }
-      } else {
-        streams_[i] = new std::ifstream(
-            keys_[i], std::ios_base::in | std::ios_base::binary);
-      }
-    }
-    if (pos_ >= keys_.size()) return;
-    ReadFst();
-  }
-
-  void Reset() final {
-    if (has_stdin_) {
-      FSTERROR()
-          << "FstFarReader::Reset: Operation not supported on standard input";
-      error_ = true;
-      return;
-    }
-    pos_ = 0;
-    ReadFst();
-  }
-
-  bool Find(const string &key) final {
-    if (has_stdin_) {
-      FSTERROR()
-          << "FstFarReader::Find: Operation not supported on standard input";
-      error_ = true;
-      return false;
-    }
-    pos_ = 0;  // TODO
-    ReadFst();
-    return true;
-  }
-
-  bool Done() const final { return error_ || pos_ >= keys_.size(); }
-
-  void Next() final {
-    ++pos_;
-    ReadFst();
-  }
-
-  const string &GetKey() const final { return keys_[pos_]; }
-
-  const Fst<Arc> *GetFst() const final { return fst_.get(); }
-
-  constexpr FarType Type() const final { return FAR_FST; }
-
-  bool Error() const final { return error_; }
-
-  ~FstFarReader() {
-    for (size_t i = 0; i < keys_.size(); ++i) {
-      if (streams_[i] != &std::cin) {
-        delete streams_[i];
-      }
-    }
-  }
-
- private:
-  void ReadFst() {
-    fst_.reset();
-    if (pos_ >= keys_.size()) return;
-    streams_[pos_]->seekg(0);
-    fst_.reset(Fst<Arc>::Read(*streams_[pos_], FstReadOptions()));
-    if (!fst_) {
-      FSTERROR() << "FstFarReader: Error reading Fst from: " << keys_[pos_];
-      error_ = true;
-    }
-  }
-
-  std::vector<string> keys_;
-  std::vector<std::istream *> streams_;
-  bool has_stdin_;
-  size_t pos_;
-  mutable std::unique_ptr<Fst<Arc>> fst_;
-  mutable bool error_;
-};
-
-
-template <class Arc>
-FarReader<Arc> *FarReader<Arc>::Open(std::unique_ptr<std::istream> stream) {
-  // TODO(fangjun): Support other types
-    return STTableFarReader<Arc>::Open(std::move(stream));
-}
-
-template <class Arc>
-FarReader<Arc> *FarReader<Arc>::Open(std::vector<std::unique_ptr<std::istream>> streams) {
-  // TODO(fangjun): Support other types
-    return STTableFarReader<Arc>::Open(std::move(streams));
-}
-
-template <class Arc>
-FarReader<Arc> *FarReader<Arc>::Open(const string &filename) {
-  if (filename.empty())
-    return STListFarReader<Arc>::Open(filename);
-  else if (IsSTTable(filename))
-    return STTableFarReader<Arc>::Open(filename);
-  else if (IsSTList(filename))
-    return STListFarReader<Arc>::Open(filename);
-  else if (IsFst(filename))
-    return FstFarReader<Arc>::Open(filename);
-  return nullptr;
-}
-
-template <class Arc>
-FarReader<Arc> *FarReader<Arc>::Open(const std::vector<string> &filenames) {
-  if (!filenames.empty() && filenames[0].empty())
-    return STListFarReader<Arc>::Open(filenames);
-  else if (!filenames.empty() && IsSTTable(filenames[0]))
-    return STTableFarReader<Arc>::Open(filenames);
-  else if (!filenames.empty() && IsSTList(filenames[0]))
-    return STListFarReader<Arc>::Open(filenames);
-  else if (!filenames.empty() && IsFst(filenames[0]))
-    return FstFarReader<Arc>::Open(filenames);
-  return nullptr;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_FAR_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/farlib.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/farlib.h
deleted file mode 100644
index c9bb1710..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/farlib.h
+++ /dev/null
@@ -1,19 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// A finite-state archive (FAR) is used to store an indexable collection of
-// FSTs in a single file. Utilities are provided to create FARs from FSTs,
-// to iterate over FARs, and to extract specific FSTs from FARs.
-
-#ifndef FST_EXTENSIONS_FAR_FARLIB_H_
-#define FST_EXTENSIONS_FAR_FARLIB_H_
-
-#include <fst/extensions/far/compile-strings.h>
-#include <fst/extensions/far/create.h>
-#include <fst/extensions/far/extract.h>
-#include <fst/extensions/far/far.h>
-#include <fst/extensions/far/getters.h>
-#include <fst/extensions/far/info.h>
-#include <fst/extensions/far/print-strings.h>
-
-#endif  // FST_EXTENSIONS_FAR_FARLIB_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/farscript.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/farscript.h
deleted file mode 100644
index 4bd11a94..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/farscript.h
+++ /dev/null
@@ -1,269 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Convenience file for including all of the FAR operations, or registering
-// them for new arc types.
-
-#ifndef FST_EXTENSIONS_FAR_FARSCRIPT_H_
-#define FST_EXTENSIONS_FAR_FARSCRIPT_H_
-
-#include <string>
-#include <vector>
-
-#include <fst/types.h>
-#include <fst/extensions/far/compile-strings.h>
-#include <fst/extensions/far/create.h>
-#include <fst/extensions/far/equal.h>
-#include <fst/extensions/far/extract.h>
-#include <fst/extensions/far/far.h>
-#include <fst/extensions/far/far-class.h>
-#include <fst/extensions/far/info.h>
-#include <fst/extensions/far/isomorphic.h>
-#include <fst/extensions/far/print-strings.h>
-#include <fst/extensions/far/script-impl.h>
-#include <fst/script/arg-packs.h>
-
-namespace fst {
-namespace script {
-
-// Note: it is safe to pass these strings as references because this struct is
-// only used to pass them deeper in the call graph. Be sure you understand why
-// this is so before using this struct for anything else!
-struct FarCompileStringsArgs {
-  const std::vector<string> &in_fnames;
-  const string &out_fname;
-  const string &fst_type;
-  const FarType &far_type;
-  const int32 generate_keys;
-  const FarEntryType fet;
-  const FarTokenType tt;
-  const string &symbols_fname;
-  const string &unknown_symbol;
-  const bool keep_symbols;
-  const bool initial_symbols;
-  const bool allow_negative_labels;
-  const string &key_prefix;
-  const string &key_suffix;
-
-  FarCompileStringsArgs(const std::vector<string> &in_fnames,
-                        const string &out_fname, const string &fst_type,
-                        const FarType &far_type, int32 generate_keys,
-                        FarEntryType fet, FarTokenType tt,
-                        const string &symbols_fname,
-                        const string &unknown_symbol, bool keep_symbols,
-                        bool initial_symbols, bool allow_negative_labels,
-                        const string &key_prefix, const string &key_suffix)
-      : in_fnames(in_fnames),
-        out_fname(out_fname),
-        fst_type(fst_type),
-        far_type(far_type),
-        generate_keys(generate_keys),
-        fet(fet),
-        tt(tt),
-        symbols_fname(symbols_fname),
-        unknown_symbol(unknown_symbol),
-        keep_symbols(keep_symbols),
-        initial_symbols(initial_symbols),
-        allow_negative_labels(allow_negative_labels),
-        key_prefix(key_prefix),
-        key_suffix(key_suffix) {}
-};
-
-template <class Arc>
-void FarCompileStrings(FarCompileStringsArgs *args) {
-  FarCompileStrings<Arc>(
-      args->in_fnames, args->out_fname, args->fst_type, args->far_type,
-      args->generate_keys, args->fet, args->tt, args->symbols_fname,
-      args->unknown_symbol, args->keep_symbols, args->initial_symbols,
-      args->allow_negative_labels, args->key_prefix, args->key_suffix);
-}
-
-void FarCompileStrings(const std::vector<string> &in_fnames,
-                       const string &out_fname, const string &arc_type,
-                       const string &fst_type, const FarType &far_type,
-                       int32 generate_keys, FarEntryType fet, FarTokenType tt,
-                       const string &symbols_fname,
-                       const string &unknown_symbol, bool keep_symbols,
-                       bool initial_symbols, bool allow_negative_labels,
-                       const string &key_prefix, const string &key_suffix);
-
-// Note: it is safe to pass these strings as references because this struct is
-// only used to pass them deeper in the call graph. Be sure you understand why
-// this is so before using this struct for anything else!
-struct FarCreateArgs {
-  const std::vector<string> &in_fnames;
-  const string &out_fname;
-  const int32 generate_keys;
-  const FarType &far_type;
-  const string &key_prefix;
-  const string &key_suffix;
-
-  FarCreateArgs(const std::vector<string> &in_fnames, const string &out_fname,
-                const int32 generate_keys, const FarType &far_type,
-                const string &key_prefix, const string &key_suffix)
-      : in_fnames(in_fnames),
-        out_fname(out_fname),
-        generate_keys(generate_keys),
-        far_type(far_type),
-        key_prefix(key_prefix),
-        key_suffix(key_suffix) {}
-};
-
-template <class Arc>
-void FarCreate(FarCreateArgs *args) {
-  FarCreate<Arc>(args->in_fnames, args->out_fname, args->generate_keys,
-                 args->far_type, args->key_prefix, args->key_suffix);
-}
-
-void FarCreate(const std::vector<string> &in_fnames, const string &out_fname,
-               const string &arc_type, const int32 generate_keys,
-               const FarType &far_type, const string &key_prefix,
-               const string &key_suffix);
-
-using FarEqualInnerArgs = std::tuple<const string &, const string &, float,
-                                     const string &, const string &>;
-
-using FarEqualArgs = WithReturnValue<bool, FarEqualInnerArgs>;
-
-template <class Arc>
-void FarEqual(FarEqualArgs *args) {
-  args->retval = fst::FarEqual<Arc>(
-      std::get<0>(args->args), std::get<1>(args->args), std::get<2>(args->args),
-      std::get<3>(args->args), std::get<4>(args->args));
-}
-
-bool FarEqual(const string &filename1, const string &filename2,
-              const string &arc_type, float delta = kDelta,
-              const string &begin_key = string(),
-              const string &end_key = string());
-
-using FarExtractArgs =
-    std::tuple<const std::vector<string> &, int32, const string &,
-               const string &, const string &, const string &, const string &>;
-
-template <class Arc>
-void FarExtract(FarExtractArgs *args) {
-  fst::FarExtract<Arc>(std::get<0>(*args), std::get<1>(*args),
-                           std::get<2>(*args), std::get<3>(*args),
-                           std::get<4>(*args), std::get<5>(*args),
-                           std::get<6>(*args));
-}
-
-void FarExtract(const std::vector<string> &ifilenames, const string &arc_type,
-                int32 generate_filenames, const string &keys,
-                const string &key_separator, const string &range_delimiter,
-                const string &filename_prefix, const string &filename_suffix);
-
-using FarInfoArgs = std::tuple<const std::vector<string> &, const string &,
-                               const string &, const bool>;
-
-template <class Arc>
-void FarInfo(FarInfoArgs *args) {
-  fst::FarInfo<Arc>(std::get<0>(*args), std::get<1>(*args),
-                        std::get<2>(*args), std::get<3>(*args));
-}
-
-void FarInfo(const std::vector<string> &filenames, const string &arc_type,
-             const string &begin_key, const string &end_key,
-             const bool list_fsts);
-
-using GetFarInfoArgs = std::tuple<const std::vector<string> &, const string &,
-                                  const string &, const bool, FarInfoData *>;
-
-template <class Arc>
-void GetFarInfo(GetFarInfoArgs *args) {
-  fst::GetFarInfo<Arc>(std::get<0>(*args), std::get<1>(*args),
-                           std::get<2>(*args), std::get<3>(*args),
-                           std::get<4>(*args));
-}
-
-void GetFarInfo(const std::vector<string> &filenames, const string &arc_type,
-                const string &begin_key, const string &end_key,
-                const bool list_fsts, FarInfoData *);
-
-using FarIsomorphicInnerArgs = std::tuple<const string &, const string &, float,
-                                          const string &, const string &>;
-
-using FarIsomorphicArgs = WithReturnValue<bool, FarIsomorphicInnerArgs>;
-
-template <class Arc>
-void FarIsomorphic(FarIsomorphicArgs *args) {
-  args->retval = fst::FarIsomorphic<Arc>(
-      std::get<0>(args->args), std::get<1>(args->args), std::get<2>(args->args),
-      std::get<3>(args->args), std::get<4>(args->args));
-}
-
-bool FarIsomorphic(const string &filename1, const string &filename2,
-                   const string &arc_type, float delta = kDelta,
-                   const string &begin_key = string(),
-                   const string &end_key = string());
-
-struct FarPrintStringsArgs {
-  const std::vector<string> &ifilenames;
-  const FarEntryType entry_type;
-  const FarTokenType token_type;
-  const string &begin_key;
-  const string &end_key;
-  const bool print_key;
-  const bool print_weight;
-  const string &symbols_fname;
-  const bool initial_symbols;
-  const int32 generate_filenames;
-  const string &filename_prefix;
-  const string &filename_suffix;
-
-  FarPrintStringsArgs(const std::vector<string> &ifilenames,
-                      const FarEntryType entry_type,
-                      const FarTokenType token_type, const string &begin_key,
-                      const string &end_key, const bool print_key,
-                      const bool print_weight, const string &symbols_fname,
-                      const bool initial_symbols,
-                      const int32 generate_filenames,
-                      const string &filename_prefix,
-                      const string &filename_suffix)
-      : ifilenames(ifilenames),
-        entry_type(entry_type),
-        token_type(token_type),
-        begin_key(begin_key),
-        end_key(end_key),
-        print_key(print_key),
-        print_weight(print_weight),
-        symbols_fname(symbols_fname),
-        initial_symbols(initial_symbols),
-        generate_filenames(generate_filenames),
-        filename_prefix(filename_prefix),
-        filename_suffix(filename_suffix) {}
-};
-
-template <class Arc>
-void FarPrintStrings(FarPrintStringsArgs *args) {
-  fst::FarPrintStrings<Arc>(
-      args->ifilenames, args->entry_type, args->token_type, args->begin_key,
-      args->end_key, args->print_key, args->print_weight, args->symbols_fname,
-      args->initial_symbols, args->generate_filenames, args->filename_prefix,
-      args->filename_suffix);
-}
-
-void FarPrintStrings(const std::vector<string> &ifilenames,
-                     const string &arc_type, const FarEntryType entry_type,
-                     const FarTokenType token_type, const string &begin_key,
-                     const string &end_key, const bool print_key,
-                     const bool print_weight, const string &symbols_fname,
-                     const bool initial_symbols, const int32 generate_filenames,
-                     const string &filename_prefix,
-                     const string &filename_suffix);
-
-}  // namespace script
-}  // namespace fst
-
-#define REGISTER_FST_FAR_OPERATIONS(ArcType)                                 \
-  REGISTER_FST_OPERATION(FarCompileStrings, ArcType, FarCompileStringsArgs); \
-  REGISTER_FST_OPERATION(FarCreate, ArcType, FarCreateArgs);                 \
-  REGISTER_FST_OPERATION(FarEqual, ArcType, FarEqualArgs);                   \
-  REGISTER_FST_OPERATION(FarExtract, ArcType, FarExtractArgs);               \
-  REGISTER_FST_OPERATION(FarInfo, ArcType, FarInfoArgs);                     \
-  REGISTER_FST_OPERATION(FarIsomorphic, ArcType, FarIsomorphicArgs);         \
-  REGISTER_FST_OPERATION(FarPrintStrings, ArcType, FarPrintStringsArgs);     \
-  REGISTER_FST_OPERATION(GetFarInfo, ArcType, GetFarInfoArgs)
-
-#endif  // FST_EXTENSIONS_FAR_FARSCRIPT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/getters.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/getters.h
deleted file mode 100644
index 3dde4194..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/getters.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes and functions for registering and invoking FAR main
-// functions that support multiple and extensible arc types.
-
-#ifndef FST_EXTENSIONS_FAR_GETTERS_H_
-#define FST_EXTENSIONS_FAR_GETTERS_H_
-
-#include <fst/flags.h>
-#include <fst/extensions/far/far.h>
-
-namespace fst {
-namespace script {
-
-FarType GetFarType(const string &str);
-
-bool GetFarEntryType(const string &str, FarEntryType *entry_type);
-
-bool GetFarTokenType(const string &str, FarTokenType *token_type);
-
-void ExpandArgs(int argc, char **argv, int *argcp, char ***argvp);
-
-}  // namespace script
-
-string GetFarTypeString(FarType type);
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_GETTERS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/info.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/info.h
deleted file mode 100644
index 0391c1f4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/info.h
+++ /dev/null
@@ -1,147 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_FAR_INFO_H_
-#define FST_EXTENSIONS_FAR_INFO_H_
-
-#include <iomanip>
-#include <memory>
-#include <set>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fst/extensions/far/far.h>
-#include <fst/extensions/far/getters.h>
-
-namespace fst {
-
-template <class Arc>
-void AccumulateStatesAndArcs(const Fst<Arc> &fst, size_t *nstate, size_t *narc,
-                             size_t *nfinal) {
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done();
-       siter.Next(), ++(*nstate)) {
-    ArcIterator<Fst<Arc>> aiter(fst, siter.Value());
-    for (; !aiter.Done(); aiter.Next(), ++(*narc)) {
-    }
-    if (fst.Final(siter.Value()) != Arc::Weight::Zero()) ++(*nfinal);
-  }
-}
-
-struct KeyInfo {
-  string key;
-  string type;
-  size_t nstate = 0;
-  size_t narc = 0;
-  size_t nfinal = 0;
-};
-
-struct FarInfoData {
-  std::vector<KeyInfo> key_infos;
-  string far_type;
-  string arc_type;
-  size_t nfst = 0;
-  size_t nstate = 0;
-  size_t narc = 0;
-  size_t nfinal = 0;
-  std::set<string> fst_types;
-};
-
-template <class Arc>
-void GetFarInfo(const std::vector<string> &filenames, const string &begin_key,
-                const string &end_key, const bool list_fsts,
-                FarInfoData *far_info) {
-  *far_info = FarInfoData();
-  std::unique_ptr<FarReader<Arc>> reader(FarReader<Arc>::Open(filenames));
-  if (!reader) {
-    LOG(ERROR) << "GetFarInfo: failed to create far reader.";
-    return;
-  }
-  if (!begin_key.empty()) reader->Find(begin_key);
-
-  for (; !reader->Done(); reader->Next()) {
-    const auto &key = reader->GetKey();
-    if (!end_key.empty() && end_key < key) break;
-    ++far_info->nfst;
-    const auto *fst = reader->GetFst();
-    far_info->fst_types.insert(fst->Type());
-    if (list_fsts) {
-      KeyInfo info;
-      info.key = key;
-      info.type = fst->Type();
-      AccumulateStatesAndArcs(*fst, &info.nstate, &info.narc, &info.nfinal);
-      far_info->nstate += info.nstate;
-      far_info->narc += info.narc;
-      far_info->nfinal += info.nfinal;
-      far_info->key_infos.push_back(info);
-    } else {
-      AccumulateStatesAndArcs(*fst, &far_info->nstate, &far_info->narc,
-                              &far_info->nfinal);
-    }
-  }
-  far_info->far_type = GetFarTypeString(reader->Type());
-  far_info->arc_type = Arc::Type();
-}
-
-template <class Arc>
-void FarInfo(const std::vector<string> &filenames, const string &begin_key,
-             const string &end_key, const bool list_fsts) {
-  FarInfoData info;
-  GetFarInfo<Arc>(filenames, begin_key, end_key, list_fsts, &info);
-  if (!list_fsts) {
-    std::cout << std::left << std::setw(50) << "far type" << info.far_type
-              << std::endl;
-    std::cout << std::left << std::setw(50) << "arc type" << Arc::Type()
-              << std::endl;
-    std::cout << std::left << std::setw(50) << "fst type";
-    for (auto iter = info.fst_types.begin(); iter != info.fst_types.end();
-         ++iter) {
-      if (iter != info.fst_types.begin()) std::cout << ",";
-      std::cout << *iter;
-    }
-    std::cout << std::endl;
-    std::cout << std::left << std::setw(50) << "# of FSTs" << info.nfst
-              << std::endl;
-    std::cout << std::left << std::setw(50) << "total # of states"
-              << info.nstate << std::endl;
-    std::cout << std::left << std::setw(50) << "total # of arcs" << info.narc
-              << std::endl;
-    std::cout << std::left << std::setw(50) << "total # of final states"
-              << info.nfinal << std::endl;
-  } else {
-    // FIXME(kbg): Grok, then document this.
-    int wkey = 10;
-    int wtype = 10;
-    int wnstate = 14;
-    int wnarc = 12;
-    int wnfinal = 20;
-    for (const auto &key_info : info.key_infos) {
-      if (key_info.key.size() + 2 > wkey) wkey = key_info.key.size() + 2;
-      if (key_info.type.size() + 2 > wtype) wtype = key_info.type.size() + 2;
-      if (ceil(log10(key_info.nstate)) + 2 > wnstate) {
-        wnstate = ceil(log10(key_info.nstate)) + 2;
-      }
-      if (ceil(log10(key_info.narc)) + 2 > wnarc) {
-        wnarc = ceil(log10(key_info.narc)) + 2;
-      }
-      if (ceil(log10(key_info.nfinal)) + 2 > wnfinal) {
-        wnfinal = ceil(log10(key_info.nfinal)) + 2;
-      }
-    }
-    std::cout << std::left << std::setw(wkey) << "key" << std::setw(wtype)
-              << "type" << std::right << std::setw(wnstate) << "# of states"
-              << std::setw(wnarc) << "# of arcs" << std::setw(wnfinal)
-              << "# of final states" << std::endl;
-    for (const auto &key_info : info.key_infos) {
-      std::cout << std::left << std::setw(wkey) << key_info.key
-                << std::setw(wtype) << key_info.type << std::right
-                << std::setw(wnstate) << key_info.nstate << std::setw(wnarc)
-                << key_info.narc << std::setw(wnfinal) << key_info.nfinal
-                << std::endl;
-    }
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_INFO_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/isomorphic.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/isomorphic.h
deleted file mode 100644
index 1e6e9cb3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/isomorphic.h
+++ /dev/null
@@ -1,69 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_FAR_ISOMORPHIC_H_
-#define FST_EXTENSIONS_FAR_ISOMORPHIC_H_
-
-#include <memory>
-#include <string>
-
-#include <fst/extensions/far/far.h>
-#include <fst/isomorphic.h>
-
-namespace fst {
-
-template <class Arc>
-bool FarIsomorphic(const string &filename1, const string &filename2,
-                   float delta = kDelta, const string &begin_key = string(),
-                   const string &end_key = string()) {
-  std::unique_ptr<FarReader<Arc>> reader1(FarReader<Arc>::Open(filename1));
-  if (!reader1) {
-    LOG(ERROR) << "FarIsomorphic: Cannot open FAR file " << filename1;
-    return false;
-  }
-  std::unique_ptr<FarReader<Arc>> reader2(FarReader<Arc>::Open(filename2));
-  if (!reader2) {
-    LOG(ERROR) << "FarIsomorphic: Cannot open FAR file " << filename2;
-    return false;
-  }
-  if (!begin_key.empty()) {
-    bool find_begin1 = reader1->Find(begin_key);
-    bool find_begin2 = reader2->Find(begin_key);
-    if (!find_begin1 || !find_begin2) {
-      bool ret = !find_begin1 && !find_begin2;
-      if (!ret) {
-        VLOG(1) << "FarIsomorphic: Key " << begin_key << " missing from "
-                << (find_begin1 ? "second" : "first") << " archive.";
-      }
-      return ret;
-    }
-  }
-  for (; !reader1->Done() && !reader2->Done();
-       reader1->Next(), reader2->Next()) {
-    const auto &key1 = reader1->GetKey();
-    const auto &key2 = reader2->GetKey();
-    if (!end_key.empty() && end_key < key1 && end_key < key2) return true;
-    if (key1 != key2) {
-      LOG(ERROR) << "FarIsomorphic: Mismatched keys " << key1 << " and "
-                 << key2;
-      return false;
-    }
-    if (!Isomorphic(*(reader1->GetFst()), *(reader2->GetFst()), delta)) {
-      LOG(ERROR) << "FarIsomorphic: FSTs for key " << key1
-                 << " are not isomorphic";
-      return false;
-    }
-  }
-  if (!reader1->Done() || !reader2->Done()) {
-    LOG(ERROR) << "FarIsomorphic: Key "
-               << (reader1->Done() ? reader2->GetKey() : reader1->GetKey())
-               << " missing form " << (reader2->Done() ? "first" : "second")
-               << " archive";
-    return false;
-  }
-  return true;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_ISOMORPHIC_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/print-strings.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/print-strings.h
deleted file mode 100644
index dc428401..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/print-strings.h
+++ /dev/null
@@ -1,105 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Outputs as strings the string FSTs in a finite-state archive.
-
-#ifndef FST_EXTENSIONS_FAR_PRINT_STRINGS_H_
-#define FST_EXTENSIONS_FAR_PRINT_STRINGS_H_
-
-#include <iomanip>
-#include <string>
-#include <vector>
-
-#include <fst/flags.h>
-#include <fst/extensions/far/far.h>
-#include <fstream>
-#include <fst/shortest-distance.h>
-#include <fst/string.h>
-
-DECLARE_string(far_field_separator);
-
-namespace fst {
-
-template <class Arc>
-void FarPrintStrings(const std::vector<string> &ifilenames,
-                     FarEntryType entry_type, FarTokenType far_token_type,
-                     const string &begin_key, const string &end_key,
-                     bool print_key, bool print_weight,
-                     const string &symbols_fname, bool initial_symbols,
-                     int32 generate_filenames, const string &filename_prefix,
-                     const string &filename_suffix) {
-  StringTokenType token_type;
-  if (far_token_type == FTT_SYMBOL) {
-    token_type = StringTokenType::SYMBOL;
-  } else if (far_token_type == FTT_BYTE) {
-    token_type = StringTokenType::BYTE;
-  } else if (far_token_type == FTT_UTF8) {
-    token_type = StringTokenType::UTF8;
-  } else {
-    FSTERROR() << "FarPrintStrings: Unknown token type";
-    return;
-  }
-  std::unique_ptr<const SymbolTable> syms;
-  if (!symbols_fname.empty()) {
-    // TODO(kbg): Allow negative flag?
-    const SymbolTableTextOptions opts(true);
-    syms.reset(SymbolTable::ReadText(symbols_fname, opts));
-    if (!syms) {
-      LOG(ERROR) << "FarPrintStrings: Error reading symbol table "
-                 << symbols_fname;
-      return;
-    }
-  }
-  std::unique_ptr<FarReader<Arc>> far_reader(FarReader<Arc>::Open(ifilenames));
-  if (!far_reader) return;
-  if (!begin_key.empty()) far_reader->Find(begin_key);
-  string okey;
-  int nrep = 0;
-  for (int i = 1; !far_reader->Done(); far_reader->Next(), ++i) {
-    const auto &key = far_reader->GetKey();
-    if (!end_key.empty() && end_key < key) break;
-    if (okey == key) {
-      ++nrep;
-    } else {
-      nrep = 0;
-    }
-    okey = key;
-    const auto *fst = far_reader->GetFst();
-    if (i == 1 && initial_symbols && !syms && fst->InputSymbols())
-      syms.reset(fst->InputSymbols()->Copy());
-    string str;
-    VLOG(2) << "Handling key: " << key;
-    StringPrinter<Arc> string_printer(token_type,
-                                      syms ? syms.get() : fst->InputSymbols());
-    string_printer(*fst, &str);
-    if (entry_type == FET_LINE) {
-      if (print_key) std::cout << key << FLAGS_far_field_separator[0];
-      std::cout << str;
-      if (print_weight)
-        std::cout << FLAGS_far_field_separator[0] << ShortestDistance(*fst);
-      std::cout << std::endl;
-    } else if (entry_type == FET_FILE) {
-      std::stringstream sstrm;
-      if (generate_filenames) {
-        sstrm.fill('0');
-        sstrm << std::right << std::setw(generate_filenames) << i;
-      } else {
-        sstrm << key;
-        if (nrep > 0) sstrm << "." << nrep;
-      }
-      string filename;
-      filename = filename_prefix + sstrm.str() + filename_suffix;
-      std::ofstream ostrm(filename);
-      if (!ostrm) {
-        LOG(ERROR) << "FarPrintStrings: Can't open file: " << filename;
-        return;
-      }
-      ostrm << str;
-      if (token_type == StringTokenType::SYMBOL) ostrm << "\n";
-    }
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_PRINT_STRINGS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/script-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/script-impl.h
deleted file mode 100644
index a0586cc3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/script-impl.h
+++ /dev/null
@@ -1,23 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes and functions for registering and invoking Far main
-// functions that support multiple and extensible arc types.
-
-#ifndef FST_EXTENSIONS_FAR_SCRIPT_IMPL_H_
-#define FST_EXTENSIONS_FAR_SCRIPT_IMPL_H_
-
-#include <string>
-
-#include <fst/compat.h>
-namespace fst {
-namespace script {
-
-string LoadArcTypeFromFar(const string &far_fname);
-
-string LoadArcTypeFromFst(const string &fst_fname);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_SCRIPT_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/stlist.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/stlist.h
deleted file mode 100644
index fcb80d04..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/stlist.h
+++ /dev/null
@@ -1,267 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// A generic (string,type) list file format.
-//
-// This is a stripped-down version of STTable that does not support the Find()
-// operation but that does support reading/writting from standard in/out.
-
-#ifndef FST_EXTENSIONS_FAR_STLIST_H_
-#define FST_EXTENSIONS_FAR_STLIST_H_
-
-#include <algorithm>
-#include <functional>
-#include <iostream>
-#include <memory>
-#include <queue>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fstream>
-#include <fst/util.h>
-
-namespace fst {
-
-static constexpr int32 kSTListMagicNumber = 5656924;
-static constexpr int32 kSTListFileVersion = 1;
-
-// String-type list writing class for object of type T using a functor Writer.
-// The Writer functor must provide at least the following interface:
-//
-//   struct Writer {
-//     void operator()(std::ostream &, const T &) const;
-//   };
-template <class T, class Writer>
-class STListWriter {
- public:
-  explicit STListWriter(const string &filename)
-      : stream_(filename.empty() ? &std::cout : new std::ofstream(
-                                                    filename,
-                                                    std::ios_base::out |
-                                                        std::ios_base::binary)),
-        error_(false) {
-    WriteType(*stream_, kSTListMagicNumber);
-    WriteType(*stream_, kSTListFileVersion);
-    if (!stream_) {
-      FSTERROR() << "STListWriter::STListWriter: Error writing to file: "
-                 << filename;
-      error_ = true;
-    }
-  }
-
-  static STListWriter<T, Writer> *Create(const string &filename) {
-    return new STListWriter<T, Writer>(filename);
-  }
-
-  void Add(const string &key, const T &t) {
-    if (key == "") {
-      FSTERROR() << "STListWriter::Add: Key empty: " << key;
-      error_ = true;
-    } else if (key < last_key_) {
-      FSTERROR() << "STListWriter::Add: Key out of order: " << key;
-      error_ = true;
-    }
-    if (error_) return;
-    last_key_ = key;
-    WriteType(*stream_, key);
-    entry_writer_(*stream_, t);
-  }
-
-  bool Error() const { return error_; }
-
-  ~STListWriter() {
-    WriteType(*stream_, string());
-    if (stream_ != &std::cout) delete stream_;
-  }
-
- private:
-  Writer entry_writer_;
-  std::ostream *stream_;  // Output stream.
-  string last_key_;       // Last key.
-  bool error_;
-
-  STListWriter(const STListWriter &) = delete;
-  STListWriter &operator=(const STListWriter &) = delete;
-};
-
-// String-type list reading class for object of type T using a functor Reader.
-// Reader must provide at least the following interface:
-//
-//   struct Reader {
-//     T *operator()(std::istream &) const;
-//   };
-template <class T, class Reader>
-class STListReader {
- public:
-  explicit STListReader(const std::vector<string> &filenames)
-      : sources_(filenames), error_(false) {
-    streams_.resize(filenames.size(), 0);
-    bool has_stdin = false;
-    for (size_t i = 0; i < filenames.size(); ++i) {
-      if (filenames[i].empty()) {
-        if (!has_stdin) {
-          streams_[i] = &std::cin;
-          sources_[i] = "stdin";
-          has_stdin = true;
-        } else {
-          FSTERROR() << "STListReader::STListReader: Cannot read multiple "
-                     << "inputs from standard input";
-          error_ = true;
-          return;
-        }
-      } else {
-        streams_[i] = new std::ifstream(
-            filenames[i], std::ios_base::in | std::ios_base::binary);
-      }
-      int32 magic_number = 0;
-      ReadType(*streams_[i], &magic_number);
-      int32 file_version = 0;
-      ReadType(*streams_[i], &file_version);
-      if (magic_number != kSTListMagicNumber) {
-        FSTERROR() << "STListReader::STListReader: Wrong file type: "
-                   << filenames[i];
-        error_ = true;
-        return;
-      }
-      if (file_version != kSTListFileVersion) {
-        FSTERROR() << "STListReader::STListReader: Wrong file version: "
-                   << filenames[i];
-        error_ = true;
-        return;
-      }
-      string key;
-      ReadType(*streams_[i], &key);
-      if (!key.empty()) heap_.push(std::make_pair(key, i));
-      if (!*streams_[i]) {
-        FSTERROR() << "STListReader: Error reading file: " << sources_[i];
-        error_ = true;
-        return;
-      }
-    }
-    if (heap_.empty()) return;
-    const auto current = heap_.top().second;
-    entry_.reset(entry_reader_(*streams_[current]));
-    if (!entry_ || !*streams_[current]) {
-      FSTERROR() << "STListReader: Error reading entry for key "
-                 << heap_.top().first << ", file " << sources_[current];
-      error_ = true;
-    }
-  }
-
-  ~STListReader() {
-    for (auto &stream : streams_) {
-      if (stream != &std::cin) delete stream;
-    }
-  }
-
-  static STListReader<T, Reader> *Open(const string &filename) {
-    std::vector<string> filenames;
-    filenames.push_back(filename);
-    return new STListReader<T, Reader>(filenames);
-  }
-
-  static STListReader<T, Reader> *Open(const std::vector<string> &filenames) {
-    return new STListReader<T, Reader>(filenames);
-  }
-
-  void Reset() {
-    FSTERROR() << "STListReader::Reset: Operation not supported";
-    error_ = true;
-  }
-
-  bool Find(const string &key) {
-    FSTERROR() << "STListReader::Find: Operation not supported";
-    error_ = true;
-    return false;
-  }
-
-  bool Done() const { return error_ || heap_.empty(); }
-
-  void Next() {
-    if (error_) return;
-    auto current = heap_.top().second;
-    string key;
-    heap_.pop();
-    ReadType(*(streams_[current]), &key);
-    if (!*streams_[current]) {
-      FSTERROR() << "STListReader: Error reading file: " << sources_[current];
-      error_ = true;
-      return;
-    }
-    if (!key.empty()) heap_.push(std::make_pair(key, current));
-    if (!heap_.empty()) {
-      current = heap_.top().second;
-      entry_.reset(entry_reader_(*streams_[current]));
-      if (!entry_ || !*streams_[current]) {
-        FSTERROR() << "STListReader: Error reading entry for key: "
-                   << heap_.top().first << ", file: " << sources_[current];
-        error_ = true;
-      }
-    }
-  }
-
-  const string &GetKey() const { return heap_.top().first; }
-
-  const T *GetEntry() const { return entry_.get(); }
-
-  bool Error() const { return error_; }
-
- private:
-  Reader entry_reader_;                  // Read functor.
-  std::vector<std::istream *> streams_;  // Input streams.
-  std::vector<string> sources_;          // Corresponding filenames.
-  std::priority_queue<
-      std::pair<string, size_t>, std::vector<std::pair<string, size_t>>,
-      std::greater<std::pair<string, size_t>>> heap_;  // (Key, stream id) heap
-  mutable std::unique_ptr<T> entry_;  // The currently read entry.
-  bool error_;
-
-  STListReader(const STListReader &) = delete;
-  STListReader &operator=(const STListReader &) = delete;
-};
-
-// String-type list header reading function, templated on the entry header type.
-// The Header type must provide at least the following interface:
-//
-//  struct Header {
-//    void Read(std::istream &strm, const string &filename);
-//  };
-template <class Header>
-bool ReadSTListHeader(const string &filename, Header *header) {
-  if (filename.empty()) {
-    LOG(ERROR) << "ReadSTListHeader: Can't read header from standard input";
-    return false;
-  }
-  std::ifstream strm(filename, std::ios_base::in | std::ios_base::binary);
-  if (!strm) {
-    LOG(ERROR) << "ReadSTListHeader: Could not open file: " << filename;
-    return false;
-  }
-  int32 magic_number = 0;
-  ReadType(strm, &magic_number);
-  int32 file_version = 0;
-  ReadType(strm, &file_version);
-  if (magic_number != kSTListMagicNumber) {
-    LOG(ERROR) << "ReadSTListHeader: Wrong file type: " << filename;
-    return false;
-  }
-  if (file_version != kSTListFileVersion) {
-    LOG(ERROR) << "ReadSTListHeader: Wrong file version: " << filename;
-    return false;
-  }
-  string key;
-  ReadType(strm, &key);
-  header->Read(strm, filename + ":" + key);
-  if (!strm) {
-    LOG(ERROR) << "ReadSTListHeader: Error reading file: " << filename;
-    return false;
-  }
-  return true;
-}
-
-bool IsSTList(const string &filename);
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_STLIST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/sttable.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/sttable.h
deleted file mode 100644
index 927d33b8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/far/sttable.h
+++ /dev/null
@@ -1,414 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// A generic string-to-type table file format.
-//
-// This is not meant as a generalization of SSTable. This is more of a simple
-// replacement for SSTable in order to provide an open-source implementation
-// of the FAR format for the external version of the FST library.
-
-#ifndef FST_EXTENSIONS_FAR_STTABLE_H_
-#define FST_EXTENSIONS_FAR_STTABLE_H_
-
-#include <algorithm>
-#include <istream>
-#include <memory>
-
-#include <fstream>
-#include <fst/util.h>
-
-namespace fst {
-
-static constexpr int32 kSTTableMagicNumber = 2125656924;
-static constexpr int32 kSTTableFileVersion = 1;
-
-// String-type table writing class for an object of type T using a functor
-// Writer. The Writer functor must provide at least the following interface:
-//
-//   struct Writer {
-//     void operator()(std::ostream &, const T &) const;
-//   };
-template <class T, class Writer>
-class STTableWriter {
- public:
-  explicit STTableWriter(const string &filename)
-      : stream_(filename, std::ios_base::out | std::ios_base::binary),
-        error_(false) {
-    WriteType(stream_, kSTTableMagicNumber);
-    WriteType(stream_, kSTTableFileVersion);
-    if (stream_.fail()) {
-      FSTERROR() << "STTableWriter::STTableWriter: Error writing to file: "
-                 << filename;
-      error_ = true;
-    }
-  }
-
-  static STTableWriter<T, Writer> *Create(const string &filename) {
-    if (filename.empty()) {
-      LOG(ERROR) << "STTableWriter: Writing to standard out unsupported.";
-      return nullptr;
-    }
-    return new STTableWriter<T, Writer>(filename);
-  }
-
-  void Add(const string &key, const T &t) {
-    if (key == "") {
-      FSTERROR() << "STTableWriter::Add: Key empty: " << key;
-      error_ = true;
-    } else if (key < last_key_) {
-      FSTERROR() << "STTableWriter::Add: Key out of order: " << key;
-      error_ = true;
-    }
-    if (error_) return;
-    last_key_ = key;
-    positions_.push_back(stream_.tellp());
-    WriteType(stream_, key);
-    entry_writer_(stream_, t);
-  }
-
-  bool Error() const { return error_; }
-
-  ~STTableWriter() {
-    WriteType(stream_, positions_);
-    WriteType(stream_, static_cast<int64>(positions_.size()));
-  }
-
- private:
-  Writer entry_writer_;
-  std::ofstream stream_;
-  std::vector<int64> positions_;  // Position in file of each key-entry pair.
-  string last_key_;               // Last key.
-  bool error_;
-
-  STTableWriter(const STTableWriter &) = delete;
-  STTableWriter &operator=(const STTableWriter &) = delete;
-};
-
-// String-type table reading class for object of type T using a functor Reader.
-// Reader must provide at least the following interface:
-//
-//   struct Reader {
-//     T *operator()(std::istream &) const;
-//   };
-//
-template <class T, class Reader>
-class STTableReader {
- public:
-  explicit STTableReader(const std::vector<string> &filenames)
-      : sources_(filenames), error_(false) {
-    compare_.reset(new Compare(&keys_));
-    keys_.resize(filenames.size());
-    streams_.resize(filenames.size(), 0);
-    positions_.resize(filenames.size());
-    for (size_t i = 0; i < filenames.size(); ++i) {
-      streams_[i] = new std::ifstream(
-          filenames[i], std::ios_base::in | std::ios_base::binary);
-      int32 magic_number = 0;
-      ReadType(*streams_[i], &magic_number);
-      int32 file_version = 0;
-      ReadType(*streams_[i], &file_version);
-      if (magic_number != kSTTableMagicNumber) {
-        FSTERROR() << "STTableReader::STTableReader: Wrong file type: "
-                   << filenames[i];
-        error_ = true;
-        return;
-      }
-      if (file_version != kSTTableFileVersion) {
-        FSTERROR() << "STTableReader::STTableReader: Wrong file version: "
-                   << filenames[i];
-        error_ = true;
-        return;
-      }
-      int64 num_entries;
-      streams_[i]->seekg(-static_cast<int>(sizeof(int64)), std::ios_base::end);
-      ReadType(*streams_[i], &num_entries);
-      if (num_entries > 0) {
-        streams_[i]->seekg(-static_cast<int>(sizeof(int64)) * (num_entries + 1),
-                           std::ios_base::end);
-        positions_[i].resize(num_entries);
-        for (size_t j = 0; (j < num_entries) && (!streams_[i]->fail()); ++j) {
-          ReadType(*streams_[i], &(positions_[i][j]));
-        }
-        streams_[i]->seekg(positions_[i][0]);
-        if (streams_[i]->fail()) {
-          FSTERROR() << "STTableReader::STTableReader: Error reading file: "
-                     << filenames[i];
-          error_ = true;
-          return;
-        }
-      }
-    }
-    MakeHeap();
-  }
-
-  explicit STTableReader(std::vector<std::unique_ptr<std::istream>> streams)
-      : error_(false) {
-    sources_.resize(streams.size());
-    compare_.reset(new Compare(&keys_));
-    keys_.resize(streams.size());
-
-    streams_.reserve(streams.size());
-    for (auto& s : streams) {
-      streams_.push_back(s.release());
-    }
-
-    positions_.resize(streams.size());
-    for (size_t i = 0; i < streams.size(); ++i) {
-      if (streams_[i]->fail()) {
-        FSTERROR() << "STTableReader::STTableReader: Error reading file: "
-                 << i;
-        error_ = true;
-        return;
-      }
-      int32 magic_number = 0;
-      ReadType(*streams_[i], &magic_number);
-      int32 file_version = 0;
-      ReadType(*streams_[i], &file_version);
-      if (magic_number != kSTTableMagicNumber) {
-        FSTERROR() << "STTableReader::STTableReader: Wrong file type: "
-                   << i;
-        error_ = true;
-        return;
-      }
-      if (file_version != kSTTableFileVersion) {
-        FSTERROR() << "STTableReader::STTableReader: Wrong file version: "
-                   << i;
-        error_ = true;
-        return;
-      }
-      int64 num_entries;
-      streams_[i]->seekg(-static_cast<int>(sizeof(int64)), std::ios_base::end);
-      ReadType(*streams_[i], &num_entries);
-      if (num_entries > 0) {
-        streams_[i]->seekg(-static_cast<int>(sizeof(int64)) * (num_entries + 1),
-                           std::ios_base::end);
-        positions_[i].resize(num_entries);
-        for (size_t j = 0; (j < num_entries) && (!streams_[i]->fail()); ++j) {
-          ReadType(*streams_[i], &(positions_[i][j]));
-        }
-        streams_[i]->seekg(positions_[i][0]);
-        if (streams_[i]->fail()) {
-          FSTERROR() << "STTableReader::STTableReader: Error reading file: "
-                     << i;
-          error_ = true;
-          return;
-        }
-      }
-    }
-    MakeHeap();
-  }
-
-  ~STTableReader() {
-    for (auto &stream : streams_) delete stream;
-  }
-
-  static STTableReader<T, Reader> *Open(const string &filename) {
-    if (filename.empty()) {
-      LOG(ERROR) << "STTableReader: Operation not supported on standard input";
-      return nullptr;
-    }
-    std::vector<string> filenames;
-    filenames.push_back(filename);
-    return new STTableReader<T, Reader>(filenames);
-  }
-
-  static STTableReader<T, Reader> *Open(std::unique_ptr<std::istream> stream) {
-    std::vector<std::unique_ptr<std::istream>> streams;
-    streams.push_back(std::move(stream));
-    return new STTableReader<T, Reader>(std::move(streams));
-  }
-
-  static STTableReader<T, Reader> *Open(std::vector<std::unique_ptr<std::istream>> streams) {
-    return new STTableReader<T, Reader>(std::move(streams));
-  }
-
-  static STTableReader<T, Reader> *Open(const std::vector<string> &filenames) {
-    return new STTableReader<T, Reader>(filenames);
-  }
-
-  void Reset() {
-    if (error_) return;
-    for (size_t i = 0; i < streams_.size(); ++i)
-      streams_[i]->seekg(positions_[i].front());
-    MakeHeap();
-  }
-
-  bool Find(const string &key) {
-    if (error_) return false;
-    for (size_t i = 0; i < streams_.size(); ++i) LowerBound(i, key);
-    MakeHeap();
-    if (heap_.empty()) return false;
-    return keys_[current_] == key;
-  }
-
-  bool Done() const { return error_ || heap_.empty(); }
-
-  void Next() {
-    if (error_) return;
-    if (streams_[current_]->tellg() <= positions_[current_].back()) {
-      ReadType(*(streams_[current_]), &(keys_[current_]));
-      if (streams_[current_]->fail()) {
-        FSTERROR() << "STTableReader: Error reading file: "
-                   << sources_[current_];
-        error_ = true;
-        return;
-      }
-      std::push_heap(heap_.begin(), heap_.end(), *compare_);
-    } else {
-      heap_.pop_back();
-    }
-    if (!heap_.empty()) PopHeap();
-  }
-
-  const string &GetKey() const { return keys_[current_]; }
-
-  const T *GetEntry() const { return entry_.get(); }
-
-  bool Error() const { return error_; }
-
- private:
-  // Comparison functor used to compare stream IDs in the heap.
-  struct Compare {
-    explicit Compare(const std::vector<string> *keys) : keys(keys) {}
-
-    bool operator()(size_t i, size_t j) const {
-      return (*keys)[i] > (*keys)[j];
-    };
-
-   private:
-    const std::vector<string> *keys;
-  };
-
-  // Positions the stream at the position corresponding to the lower bound for
-  // the specified key.
-  void LowerBound(size_t id, const string &find_key) {
-    auto *strm = streams_[id];
-    const auto &positions = positions_[id];
-    if (positions.empty()) return;
-    size_t low = 0;
-    size_t high = positions.size() - 1;
-    while (low < high) {
-      size_t mid = (low + high) / 2;
-      strm->seekg(positions[mid]);
-      string key;
-      ReadType(*strm, &key);
-      if (key > find_key) {
-        high = mid;
-      } else if (key < find_key) {
-        low = mid + 1;
-      } else {
-        for (size_t i = mid; i > low; --i) {
-          strm->seekg(positions[i - 1]);
-          ReadType(*strm, &key);
-          if (key != find_key) {
-            strm->seekg(positions[i]);
-            return;
-          }
-        }
-        strm->seekg(positions[low]);
-        return;
-      }
-    }
-    strm->seekg(positions[low]);
-  }
-
-  // Adds all streams to the heap.
-  void MakeHeap() {
-    heap_.clear();
-    for (size_t i = 0; i < streams_.size(); ++i) {
-      if (positions_[i].empty()) continue;
-      ReadType(*streams_[i], &(keys_[i]));
-      if (streams_[i]->fail()) {
-        FSTERROR() << "STTableReader: Error reading file: " << sources_[i];
-        error_ = true;
-        return;
-      }
-      heap_.push_back(i);
-    }
-    if (heap_.empty()) return;
-    std::make_heap(heap_.begin(), heap_.end(), *compare_);
-    PopHeap();
-  }
-
-  // Positions the stream with the lowest key at the top of the heap, sets
-  // current_ to the ID of that stream, and reads the current entry from that
-  // stream.
-  void PopHeap() {
-    std::pop_heap(heap_.begin(), heap_.end(), *compare_);
-    current_ = heap_.back();
-    entry_.reset(entry_reader_(*streams_[current_]));
-    if (!entry_) error_ = true;
-    if (streams_[current_]->fail()) {
-      FSTERROR() << "STTableReader: Error reading entry for key: "
-                 << keys_[current_] << ", file: " << sources_[current_];
-      error_ = true;
-    }
-  }
-
-  Reader entry_reader_;
-  std::vector<std::istream *> streams_;        // Input streams.
-  std::vector<string> sources_;                // Corresponding file names.
-  std::vector<std::vector<int64>> positions_;  // Index of positions.
-  std::vector<string> keys_;  // Lowest unread key for each stream.
-  std::vector<int64> heap_;   // Heap containing ID of streams with unread keys.
-  int64 current_;             // ID of current stream to be read.
-  std::unique_ptr<Compare> compare_;  // Functor comparing stream IDs.
-  mutable std::unique_ptr<T> entry_;  // The currently read entry.
-  bool error_;
-};
-
-// String-type table header reading function template on the entry header type.
-// The Header type must provide at least the following interface:
-//
-//   struct Header {
-//     void Read(std::istream &istrm, const string &filename);
-//   };
-template <class Header>
-bool ReadSTTableHeader(const string &filename, Header *header) {
-  if (filename.empty()) {
-    LOG(ERROR) << "ReadSTTable: Can't read header from standard input";
-    return false;
-  }
-  std::ifstream strm(filename, std::ios_base::in | std::ios_base::binary);
-  if (!strm) {
-    LOG(ERROR) << "ReadSTTableHeader: Could not open file: " << filename;
-    return false;
-  }
-  int32 magic_number = 0;
-  ReadType(strm, &magic_number);
-  int32 file_version = 0;
-  ReadType(strm, &file_version);
-  if (magic_number != kSTTableMagicNumber) {
-    LOG(ERROR) << "ReadSTTableHeader: Wrong file type: " << filename;
-    return false;
-  }
-  if (file_version != kSTTableFileVersion) {
-    LOG(ERROR) << "ReadSTTableHeader: Wrong file version: " << filename;
-    return false;
-  }
-  int64 i = -1;
-  strm.seekg(-static_cast<int>(sizeof(int64)), std::ios_base::end);
-  ReadType(strm, &i);  // Reads number of entries
-  if (strm.fail()) {
-    LOG(ERROR) << "ReadSTTableHeader: Error reading file: " << filename;
-    return false;
-  }
-  if (i == 0) return true;  // No entry header to read.
-  strm.seekg(-2 * static_cast<int>(sizeof(int64)), std::ios_base::end);
-  ReadType(strm, &i);  // Reads position for last entry in file.
-  strm.seekg(i);
-  string key;
-  ReadType(strm, &key);
-  header->Read(strm, filename + ":" + key);
-  if (strm.fail()) {
-    LOG(ERROR) << "ReadSTTableHeader: Error reading file: " << filename;
-    return false;
-  }
-  return true;
-}
-
-bool IsSTTable(const string &filename);
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_FAR_STTABLE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linear-fst-data-builder.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linear-fst-data-builder.h
deleted file mode 100644
index a6ac7279..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linear-fst-data-builder.h
+++ /dev/null
@@ -1,1074 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_BUILDER_H_
-#define FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_BUILDER_H_
-
-#include <map>
-#include <queue>
-#include <set>
-#include <sstream>
-#include <stack>
-#include <string>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/log.h>
-#include <fst/fst.h>
-#include <fst/symbol-table.h>
-#include <fst/util.h>
-
-#include <fst/extensions/linear/linear-fst-data.h>
-
-namespace fst {
-
-// Forward declaration
-template <class A>
-class FeatureGroupBuilder;
-
-// For logging purposes
-inline string TranslateLabel(int64 label, const SymbolTable *syms);
-template <class Iterator>
-string JoinLabels(Iterator begin, Iterator end, const SymbolTable *syms);
-template <class Label>
-string JoinLabels(const std::vector<Label> &labels, const SymbolTable *syms);
-
-// Guesses the appropriate boundary label (start- or end-of-sentence)
-// for all labels equal to `boundary` and modifies the `sequence`
-// in-place. Returns the number of positions that are still uncertain.
-template <class A>
-typename A::Label GuessStartOrEnd(std::vector<typename A::Label> *sequence,
-                                  typename A::Label boundary);
-
-// Builds a `LinearFstData` object by adding words and feature
-// weights. A few conventions:
-//
-// - Input labels forms a dense non-empty range from 1 to `MaxInputLabel()`.
-// - Feature labels, output labels are > 0.
-// - Being a discriminative linear model, it only makes sense to use tropical
-// semirings.
-template <class A>
-class LinearFstDataBuilder {
- public:
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-
-  // Constructs a builder with associated symbol tables for diagonstic
-  // output. Each of these symbol tables may also be nullptr.
-  explicit LinearFstDataBuilder(const SymbolTable *isyms = nullptr,
-                                const SymbolTable *fsyms = nullptr,
-                                const SymbolTable *osyms = nullptr)
-      : error_(false),
-        max_future_size_(0),
-        max_input_label_(1),
-        isyms_(isyms),
-        fsyms_(fsyms),
-        osyms_(osyms) {}
-
-  // Tests whether the builder has encountered any error. No operation
-  // is valid if the builder is already at error state. All other
-  // public methods should check this before any actual operations.
-  bool Error() const { return error_; }
-
-  // Adds a word and its feature labels to the vocabulary; this
-  // version allows the word to have any output label. Returns true
-  // iff the word is added.
-  //
-  // This may fail if the word is added twice or if the feature labels
-  // are non-positive.
-  bool AddWord(Label word, const std::vector<Label> &features);
-
-  // Adds a word and its feature labels to the vocabulary; this
-  // version puts constraint on possible output labels the word can
-  // have. `possible_output` must not be empty. Returns true iff the
-  // word is added.
-  //
-  // In addition to the reasons above in the two-parameter version,
-  // this may also fail if `possible_output` is empty or any output
-  // label in it is non-positive.
-  bool AddWord(Label word, const std::vector<Label> &word_features,
-               const std::vector<Label> &possible_output);
-
-  // Creates a new feature group with specified future size (size of
-  // the look-ahead window), returns the group id to be used for
-  // adding actual feature weights or a negative number when called at
-  // error state.
-  //
-  // This does not fail unless called at error state.
-  int AddGroup(size_t future_size);
-
-  // Adds an instance of feature weight to the specified feature
-  // group. If some weight has already been added with the same
-  // feature, the product of the old and new weights are
-  // stored. Returns true iff the weight is added. A weight is not
-  // added when the context has ill-formed context involving start-,
-  // end-of-sentence marks.
-  //
-  // For two features to be within the same group, it must satisfy
-  // that (1) they have the same future size; (2) the two either have
-  // disjoint context or one is the back-off context of the
-  // other. Furthermore, for all features in a single group, there
-  // must be one and only one other context (not necessarily an active
-  // feature) that the feature immediately backs off to (i.e. there is
-  // no other context that is the back-off of the first and backs off
-  // to the second).
-  //
-  // Consider for example features with zero look-ahead of the form
-  // (input, OUTPUT).
-  //
-  // - The following two features can be put in the same group because
-  // their context is disjoint: (a a a, A A), (b, B B);
-  //
-  // - The following two features can be put in the same group because
-  // one is the back-off context of the other: (a a a, A A), (a a, A
-  // A);
-  //
-  // - The following two features can NOT be put in the same group
-  // because there is overlap but neither is the other's back-off: (a
-  // a a, A), (a a, A A);
-  //
-  // - Finally, the following three features cannot be in a same group
-  // because the first one can immediately back off to either of the
-  // rest: (a a a, A A), (a a, A A), (a a a, A).
-  //
-  // The easiest way to satisfy the constraints is to create a feature
-  // group for each feature template. However, better feature grouping
-  // may help improve speed.
-  //
-  // This may fail if any of input or output labels are non-positive,
-  // or if any call to `FeatureGroupBuilder<>::AddWeight()` fails.
-  bool AddWeight(size_t group, const std::vector<Label> &input,
-                 const std::vector<Label> &output, Weight weight);
-
-  // Returns a newly created `LinearFstData` object or nullptr in case
-  // of failure. The caller takes the ownership of the memory. No
-  // other methods shall be called after this --- this is enforced by
-  // putting the builder at error state, even when a
-  // `LinearFstData<>` object is successfully built.
-  //
-  // This may fail if the call to any `FeatureGroupBuilder<>::Dump()`
-  // fails.
-  LinearFstData<A> *Dump();
-
- private:
-  bool error_;
-  CompactSet<Label, kNoLabel> all_output_labels_;
-  std::map<Label, std::set<Label>> word_output_map_, word_feat_map_;
-  std::map<Label, std::set<size_t>> feat_groups_;
-  std::vector<std::unique_ptr<FeatureGroupBuilder<A>>> groups_;
-  size_t max_future_size_;
-  Label max_input_label_;
-  const SymbolTable *isyms_, *fsyms_, *osyms_;
-
-  LinearFstDataBuilder(const LinearFstDataBuilder &) = delete;
-  LinearFstDataBuilder &operator=(const LinearFstDataBuilder &) = delete;
-};
-
-// Builds a LinearFstData tailored for a LinearClassifierFst. The
-// major difference between an ordinary LinearFstData that works on
-// taggers and a LinearFstData that works on classifiers is that
-// feature groups are divided into sections by the prediction class
-// label. For a prediction label `pred` and a logical group id
-// `group`, the actual group id is `group * num_classes + pred -
-// 1`.
-//
-// This layout saves us from recording output labels in each single
-// FeatureGroup. Because there is no need for any delaying, stripping
-// the output allows features with different shapes but using the same
-// set of feature label mapping to reside in a single FeatureGroup.
-template <class A>
-class LinearClassifierFstDataBuilder {
- public:
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-
-  // Constructs a builder for a `num_classes`-class classifier,
-  // optinally with associated symbol tables for diagnostic
-  // output. The output labels (i.e. prediction) must be in the range
-  // of [1, num_classes].
-  explicit LinearClassifierFstDataBuilder(size_t num_classes,
-                                          const SymbolTable *isyms = nullptr,
-                                          const SymbolTable *fsyms = nullptr,
-                                          const SymbolTable *osyms = nullptr)
-      : error_(false),
-        num_classes_(num_classes),
-        num_groups_(0),
-        builder_(isyms, fsyms, osyms) {}
-
-  // Tests whether the builder has encountered any error. Similar to
-  // LinearFstDataBuilder<>::Error().
-  bool Error() const { return error_; }
-
-  // Same as LinearFstDataBuilder<>::AddWord().
-  bool AddWord(Label word, const std::vector<Label> &features);
-
-  // Adds a logical feature group. Similar to
-  // LinearFstDataBuilder<>::AddGroup(), with the exception that the
-  // returned group id is the logical group id. Also there is no need
-  // for "future" in a classifier.
-  int AddGroup();
-
-  // Adds an instance of feature weight to the specified logical
-  // feature group. Instead of a vector of output, only a single
-  // prediction is needed as the output.
-  //
-  // This may fail if `pred` is not in the range of [1, num_classes_].
-  bool AddWeight(size_t group, const std::vector<Label> &input, Label pred,
-                 Weight weight);
-
-  // Returns a newly created `LinearFstData` object or nullptr in case of
-  // failure.
-  LinearFstData<A> *Dump();
-
- private:
-  std::vector<Label> empty_;
-  bool error_;
-  size_t num_classes_, num_groups_;
-  LinearFstDataBuilder<A> builder_;
-};
-
-// Builds a single feature group. Usually used in
-// `LinearFstDataBuilder::AddWeight()`. See that method for the
-// constraints on grouping features.
-template <class A>
-class FeatureGroupBuilder {
- public:
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-
-  // Constructs a builder with the given future size. All features
-  // added to the group will have look-ahead windows of this size.
-  FeatureGroupBuilder(size_t future_size, const SymbolTable *fsyms,
-                      const SymbolTable *osyms)
-      : error_(false), future_size_(future_size), fsyms_(fsyms), osyms_(osyms) {
-    // This edge is special; see doc of class `FeatureGroup` on the
-    // details.
-    start_ = trie_.Insert(trie_.Root(), InputOutputLabel(kNoLabel, kNoLabel));
-  }
-
-  // Tests whether the builder has encountered any error. No operation
-  // is valid if the builder is already at error state. All other
-  // public methods should check this before any actual operations.
-  bool Error() const { return error_; }
-
-  // Adds a feature weight with the given context. Returns true iff
-  // the weight is added. A weight is not added if it has ill-formed
-  // context involving start-, end-of-sentence marks.
-  //
-  // Note: `input` is the sequence of input
-  // features, instead of input labels themselves. `input` must be at
-  // least as long as `future_size`; `output` may be empty, but
-  // usually should be non-empty because an empty output context is
-  // useless in discriminative modelling. All labels in both `input`
-  // and `output` must be > 0 (this is checked in
-  // `LinearFstDataBuilder::AddWeight()`). See
-  // LinearFstDataBuilder<>::AddWeight for more details.
-  //
-  // This may fail if the input is smaller than the look-ahead window.
-  bool AddWeight(const std::vector<Label> &input,
-                 const std::vector<Label> &output, Weight weight);
-
-  // Creates an actual FeatureGroup<> object. Connects back-off links;
-  // pre-accumulates weights from back-off features. Returns nullptr if
-  // there is any violation in unique immediate back-off
-  // constraints.
-  //
-  // Regardless of whether the call succeeds or not, the error flag is
-  // always set before this returns, to prevent repeated dumping.
-  //
-  // TODO(wuke): check overlapping top-level contexts (see
-  // `DumpOverlappingContext()` in tests).
-  FeatureGroup<A> *Dump(size_t max_future_size);
-
- private:
-  typedef typename FeatureGroup<A>::InputOutputLabel InputOutputLabel;
-  typedef typename FeatureGroup<A>::InputOutputLabelHash InputOutputLabelHash;
-  typedef typename FeatureGroup<A>::WeightBackLink WeightBackLink;
-  // Nested trie topology uses more memory but we can traverse a
-  // node's children easily, which is required in `BuildBackLinks()`.
-  typedef NestedTrieTopology<InputOutputLabel, InputOutputLabelHash> Topology;
-  typedef MutableTrie<InputOutputLabel, WeightBackLink, Topology> Trie;
-
-  // Finds the first node with an arc with `label` following the
-  // back-off chain of `parent`. Returns the node index or
-  // `kNoTrieNodeId` when not found. The number of hops is stored in
-  // `hop` when it is not `nullptr`.
-  //
-  // This does not fail.
-  int FindFirstMatch(InputOutputLabel label, int parent, int *hop) const;
-
-  // Links each node to its immediate back-off. root is linked to -1.
-  //
-  // This may fail when the unique immediate back-off constraint is
-  // violated.
-  void BuildBackLinks();
-
-  // Traces back on the back-chain for each node to multiply the
-  // weights from back-offs to the node itself.
-  //
-  // This does not fail.
-  void PreAccumulateWeights();
-
-  // Reconstruct the path from trie root to given node for logging.
-  bool TrieDfs(const Topology &topology, int cur, int target,
-               std::vector<InputOutputLabel> *path) const;
-  string TriePath(int node, const Topology &topology) const;
-
-  bool error_;
-  size_t future_size_;
-  Trie trie_;
-  int start_;
-  const SymbolTable *fsyms_, *osyms_;
-
-  FeatureGroupBuilder(const FeatureGroupBuilder &) = delete;
-  FeatureGroupBuilder &operator=(const FeatureGroupBuilder &) = delete;
-};
-
-//
-// Implementation of methods in `LinearFstDataBuilder`
-//
-template <class A>
-bool LinearFstDataBuilder<A>::AddWord(Label word,
-                                      const std::vector<Label> &features) {
-  if (error_) {
-    FSTERROR() << "Calling LinearFstDataBuilder<>::AddWord() at error state";
-    return false;
-  }
-  if (word == LinearFstData<A>::kStartOfSentence ||
-      word == LinearFstData<A>::kEndOfSentence) {
-    LOG(WARNING) << "Ignored: adding boundary label: "
-                 << TranslateLabel(word, isyms_)
-                 << "(start-of-sentence=" << LinearFstData<A>::kStartOfSentence
-                 << ", end-of-sentence=" << LinearFstData<A>::kEndOfSentence
-                 << ")";
-    return false;
-  }
-  if (word <= 0) {
-    error_ = true;
-    FSTERROR() << "Word label must be > 0; got " << word;
-    return false;
-  }
-  if (word > max_input_label_) max_input_label_ = word;
-  // Make sure the word hasn't been added before
-  if (word_feat_map_.find(word) != word_feat_map_.end()) {
-    error_ = true;
-    FSTERROR() << "Input word " << TranslateLabel(word, isyms_)
-               << " is added twice";
-    return false;
-  }
-  // Store features
-  std::set<Label> *feats = &word_feat_map_[word];
-  for (size_t i = 0; i < features.size(); ++i) {
-    Label feat = features[i];
-    if (feat <= 0) {
-      error_ = true;
-      FSTERROR() << "Feature label must be > 0; got " << feat;
-      return false;
-    }
-    feats->insert(feat);
-  }
-  return true;
-}
-
-template <class A>
-bool LinearFstDataBuilder<A>::AddWord(
-    Label word, const std::vector<Label> &word_features,
-    const std::vector<Label> &possible_output) {
-  if (error_) {
-    FSTERROR() << "Calling LinearFstDataBuilder<>::AddWord() at error state";
-    return false;
-  }
-  if (!AddWord(word, word_features)) return false;
-  // Store possible output constraint
-  if (possible_output.empty()) {
-    error_ = true;
-    FSTERROR() << "Empty possible output constraint; "
-               << "use the two-parameter version if no constraint is need.";
-    return false;
-  }
-  std::set<Label> *outputs = &word_output_map_[word];
-  for (size_t i = 0; i < possible_output.size(); ++i) {
-    Label output = possible_output[i];
-    if (output == LinearFstData<A>::kStartOfSentence ||
-        output == LinearFstData<A>::kEndOfSentence) {
-      LOG(WARNING) << "Ignored: word = " << TranslateLabel(word, isyms_)
-                   << ": adding boundary label as possible output: " << output
-                   << "(start-of-sentence="
-                   << LinearFstData<A>::kStartOfSentence
-                   << ", end-of-sentence=" << LinearFstData<A>::kEndOfSentence
-                   << ")";
-      continue;
-    }
-    if (output <= 0) {
-      error_ = true;
-      FSTERROR() << "Output label must be > 0; got " << output;
-      return false;
-    }
-    outputs->insert(output);
-    all_output_labels_.Insert(output);
-  }
-  return true;
-}
-
-template <class A>
-inline int LinearFstDataBuilder<A>::AddGroup(size_t future_size) {
-  if (error_) {
-    FSTERROR() << "Calling LinearFstDataBuilder<>::AddGroup() at error state";
-    return -1;
-  }
-  size_t ret = groups_.size();
-  groups_.emplace_back(new FeatureGroupBuilder<A>(future_size, fsyms_, osyms_));
-  if (future_size > max_future_size_) max_future_size_ = future_size;
-  return ret;
-}
-
-template <class A>
-bool LinearFstDataBuilder<A>::AddWeight(size_t group,
-                                        const std::vector<Label> &input,
-                                        const std::vector<Label> &output,
-                                        Weight weight) {
-  if (error_) {
-    FSTERROR() << "Calling LinearFstDataBuilder<>::AddWeight() at error state";
-    return false;
-  }
-  // Check well-formedness of boundary marks on the input.
-  {
-    bool start_in_middle = false, end_in_middle = false;
-    for (int i = 1; i < input.size(); ++i) {
-      if (input[i] == LinearFstData<A>::kStartOfSentence &&
-          input[i - 1] != LinearFstData<A>::kStartOfSentence)
-        start_in_middle = true;
-      if (input[i - 1] == LinearFstData<A>::kEndOfSentence &&
-          input[i] != LinearFstData<A>::kEndOfSentence)
-        end_in_middle = true;
-    }
-    if (start_in_middle) {
-      LOG(WARNING) << "Ignored: start-of-sentence in the middle of the input!";
-      LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-      LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-      return false;
-    }
-    if (end_in_middle) {
-      LOG(WARNING) << "Ignored: end-of-sentence in the middle of the input!";
-      LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-      LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-      return false;
-    }
-  }
-  // Check well-formedness of boundary marks on the output.
-  {
-    bool non_first_start = false, non_last_end = false;
-    for (int i = 1; i < output.size(); ++i) {
-      if (output[i] == LinearFstData<A>::kStartOfSentence)
-        non_first_start = true;
-      if (output[i - 1] == LinearFstData<A>::kEndOfSentence)
-        non_last_end = true;
-    }
-    if (non_first_start) {
-      LOG(WARNING) << "Ignored: start-of-sentence not appearing "
-                   << "as the first label in the output!";
-      LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-      LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-      return false;
-    }
-    if (non_last_end) {
-      LOG(WARNING) << "Ignored: end-of-sentence not appearing "
-                   << "as the last label in the output!";
-      LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-      LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-      return false;
-    }
-  }
-
-  for (size_t i = 0; i < input.size(); ++i) {
-    Label feat = input[i];
-    if (feat != LinearFstData<A>::kStartOfSentence &&
-        feat != LinearFstData<A>::kEndOfSentence && feat <= 0) {
-      error_ = true;
-      FSTERROR() << "Feature label must be > 0; got " << feat;
-      return false;
-    }
-    feat_groups_[feat].insert(group);
-  }
-  for (size_t i = 0; i < output.size(); ++i) {
-    Label label = output[i];
-    if (label != LinearFstData<A>::kStartOfSentence &&
-        label != LinearFstData<A>::kEndOfSentence && label <= 0) {
-      error_ = true;
-      FSTERROR() << "Output label must be > 0; got " << label;
-      return false;
-    }
-    if (label != LinearFstData<A>::kStartOfSentence &&
-        label != LinearFstData<A>::kEndOfSentence)
-      all_output_labels_.Insert(label);
-  }
-
-  // Everything looks good at this point (more checks on the way in
-  // the feature group). Add this feature weight.
-  bool added = groups_[group]->AddWeight(input, output, weight);
-  if (groups_[group]->Error()) {
-    error_ = true;
-    FSTERROR() << "FeatureGroupBuilder<>::AddWeight() failed";
-    return false;
-  }
-  return added;
-}
-
-template <class A>
-LinearFstData<A> *LinearFstDataBuilder<A>::Dump() {
-  if (error_) {
-    FSTERROR() << "Calling LinearFstDataBuilder<>::Dump() at error state";
-    return nullptr;
-  }
-
-  std::unique_ptr<LinearFstData<A>> data(new LinearFstData<A>());
-  data->max_future_size_ = max_future_size_;
-  data->max_input_label_ = max_input_label_;
-
-  // Feature groups; free builders after it's dumped.
-  data->groups_.resize(groups_.size());
-  for (int group = 0; group != groups_.size(); ++group) {
-    FeatureGroup<A> *new_group = groups_[group]->Dump(max_future_size_);
-    if (new_group == nullptr) {
-      error_ = true;
-      FSTERROR() << "Error in dumping group " << group;
-      return nullptr;
-    }
-    data->groups_[group].reset(new_group);
-    groups_[group].reset();
-    VLOG(1) << "Group " << group << ": " << new_group->Stats();
-  }
-
-  // Per-group feature mapping
-  data->group_feat_map_.Init(data->NumGroups(), max_input_label_ + 1);
-  for (Label word = 1; word <= max_input_label_; ++word) {
-    typename std::map<Label, std::set<Label>>::const_iterator it =
-        word_feat_map_.find(word);
-    if (it == word_feat_map_.end()) continue;
-    for (typename std::set<Label>::const_iterator oit = it->second.begin();
-         oit != it->second.end(); ++oit) {
-      Label feat = *oit;
-      typename std::map<Label, std::set<size_t>>::const_iterator jt =
-          feat_groups_.find(feat);
-      if (jt == feat_groups_.end()) continue;
-      for (std::set<size_t>::const_iterator git = jt->second.begin();
-           git != jt->second.end(); ++git) {
-        size_t group_id = *git;
-        if (!data->group_feat_map_.Set(group_id, word, feat)) {
-          error_ = true;
-          return nullptr;
-        }
-      }
-    }
-  }
-
-  // Possible output labels
-  {
-    std::vector<typename LinearFstData<A>::InputAttribute> *input_attribs =
-        &data->input_attribs_;
-    std::vector<Label> *output_pool = &data->output_pool_;
-    input_attribs->resize(max_input_label_ + 1);
-    for (Label word = 0; word <= max_input_label_; ++word) {
-      typename std::map<Label, std::set<Label>>::const_iterator it =
-          word_output_map_.find(word);
-      if (it == word_output_map_.end()) {
-        (*input_attribs)[word].output_begin = 0;
-        (*input_attribs)[word].output_length = 0;
-      } else {
-        (*input_attribs)[word].output_begin = output_pool->size();
-        (*input_attribs)[word].output_length = it->second.size();
-        for (typename std::set<Label>::const_iterator oit = it->second.begin();
-             oit != it->second.end(); ++oit) {
-          Label olabel = *oit;
-          output_pool->push_back(olabel);
-        }
-      }
-    }
-  }
-
-  for (typename CompactSet<Label, kNoLabel>::const_iterator it =
-           all_output_labels_.Begin();
-       it != all_output_labels_.End(); ++it)
-    data->output_set_.push_back(*it);
-
-  error_ = true;  // prevent future calls on this object
-  return data.release();
-}
-
-//
-// Implementation of methods in `LinearClassifierFstDataBuilder`
-//
-template <class A>
-inline bool LinearClassifierFstDataBuilder<A>::AddWord(
-    Label word, const std::vector<Label> &features) {
-  if (error_) {
-    FSTERROR() << "Calling LinearClassifierFstDataBuilder<>::AddWord() at "
-                  "error state";
-    return false;
-  }
-  bool added = builder_.AddWord(word, features);
-  if (builder_.Error()) error_ = true;
-  return added;
-}
-
-template <class A>
-inline int LinearClassifierFstDataBuilder<A>::AddGroup() {
-  if (error_) {
-    FSTERROR() << "Calling LinearClassifierFstDataBuilder<>::AddGroup() at "
-                  "error state";
-    return -1;
-  }
-  for (int i = 0; i < num_classes_; ++i) builder_.AddGroup(0);
-  if (builder_.Error()) {
-    error_ = true;
-    return -1;
-  }
-  return num_groups_++;
-}
-
-template <class A>
-inline bool LinearClassifierFstDataBuilder<A>::AddWeight(
-    size_t group, const std::vector<Label> &input, Label pred, Weight weight) {
-  if (error_) {
-    FSTERROR() << "Calling LinearClassifierFstDataBuilder<>::AddWeight() at "
-                  "error state";
-    return false;
-  }
-  if (pred <= 0 || pred > num_classes_) {
-    FSTERROR() << "Out-of-range prediction label: " << pred
-               << " (num classes = " << num_classes_ << ")";
-    error_ = true;
-    return false;
-  }
-  size_t real_group = group * num_classes_ + pred - 1;
-  bool added = builder_.AddWeight(real_group, input, empty_, weight);
-  if (builder_.Error()) error_ = true;
-  return added;
-}
-
-template <class A>
-inline LinearFstData<A> *LinearClassifierFstDataBuilder<A>::Dump() {
-  if (error_) {
-    FSTERROR()
-        << "Calling LinearClassifierFstDataBuilder<>::Dump() at error state";
-    return nullptr;
-  }
-  LinearFstData<A> *data = builder_.Dump();
-  error_ = true;
-  return data;
-}
-
-//
-// Implementation of methods in `FeatureGroupBuilder`
-//
-template <class A>
-bool FeatureGroupBuilder<A>::AddWeight(const std::vector<Label> &input,
-                                       const std::vector<Label> &output,
-                                       Weight weight) {
-  if (error_) {
-    FSTERROR() << "Calling FeatureGroupBuilder<>::AddWeight() at error state";
-    return false;
-  }
-
-  // `LinearFstDataBuilder<>::AddWeight()` ensures prefix/suffix
-  // properties for us. We can directly count.
-  int num_input_start = 0;
-  while (num_input_start < input.size() &&
-         input[num_input_start] == LinearFstData<A>::kStartOfSentence)
-    ++num_input_start;
-  int num_output_start = 0;
-  while (num_output_start < output.size() &&
-         output[num_output_start] == LinearFstData<A>::kStartOfSentence)
-    ++num_output_start;
-  int num_input_end = 0;
-  for (int i = input.size() - 1;
-       i >= 0 && input[i] == LinearFstData<A>::kEndOfSentence; --i)
-    ++num_input_end;
-  int num_output_end = 0;
-  for (int i = output.size() - 1;
-       i >= 0 && output[i] == LinearFstData<A>::kEndOfSentence; --i)
-    ++num_output_end;
-
-  DCHECK_LE(num_output_end, 1);
-
-  if (input.size() - num_input_start < future_size_) {
-    LOG(WARNING) << "Ignored: start-of-sentence in the future!";
-    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-    LOG(WARNING) << "\tOutput: " << JoinLabels(output, fsyms_);
-    return false;
-  }
-  if (num_input_start > 0 &&
-      input.size() - future_size_ - num_input_start <
-          output.size() - num_output_start) {
-    LOG(WARNING) << "Ignored: matching start-of-sentence with actual output!";
-    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-    LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-    return false;
-  }
-  if (num_output_start > 0 &&
-      input.size() - future_size_ - num_input_start >
-          output.size() - num_output_start) {
-    LOG(WARNING) << "Ignored: matching start-of-sentence with actual input!";
-    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-    LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-    return false;
-  }
-  // The following two require `num_output_end` <= 1.
-  if (num_input_end > future_size_ && num_input_end - future_size_ != 1) {
-    LOG(WARNING) << "Ignored: matching end-of-sentence with actual output!";
-    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-    LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-    return false;
-  }
-  if (num_output_end > 0 &&
-      ((input.size() == future_size_ && future_size_ != num_input_end) ||
-       (input.size() > future_size_ &&
-        num_input_end != future_size_ + num_output_end))) {
-    LOG(WARNING) << "Ignored: matching end-of-sentence with actual input!";
-    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-    LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-    return false;
-  }
-  // Check if the context has no other labels than boundary marks
-  // (such features are useless).
-  if (num_input_start + num_input_end == input.size() &&
-      num_output_start + num_output_end == output.size()) {
-    LOG(WARNING)
-        << "Ignored: feature context consisting of only boundary marks!";
-    LOG(WARNING) << "\tInput: " << JoinLabels(input, fsyms_);
-    LOG(WARNING) << "\tOutput: " << JoinLabels(output, osyms_);
-    return false;
-  }
-
-  // Start point for insertion in the trie. Insert at `start_` iff the
-  // beginning of the context is non-consumed start-of-sentence.
-  int cur = (num_input_start == 0 && num_output_start <= future_size_)
-                ? trie_.Root()
-                : start_;
-  // Skip all input start-of-sentence marks
-  size_t ipos = num_input_start;
-  // Skip to keep at most `future_size_` start-of-sentence marks
-  size_t opos =
-      num_output_start <= future_size_ ? 0 : num_output_start - future_size_;
-  // Skip `num_output_end` end-of-sentence marks on both input and output
-  size_t iend = !input.empty() ? input.size() - num_output_end : 0,
-         oend = output.size() - num_output_end;
-  // Further, when output is empty, keep at most `future_size_`
-  // end-of-sentence marks on input.
-  if (output.empty() && num_input_end > future_size_)
-    iend = input.size() - num_input_end + future_size_;
-
-  // Actual feature context is (input[ipos:iend], output[opos:oend]).
-
-  // Pad `kNoLabel` as don't cares on the shorter of actual `input`
-  // and `output`.
-  const size_t effective_input_size = iend - ipos,
-               effective_output_size = oend - opos;
-  if (effective_input_size > effective_output_size) {
-    for (size_t pad = effective_input_size - effective_output_size; pad != 0;
-         --pad, ++ipos)
-      cur = trie_.Insert(cur, InputOutputLabel(input[ipos], kNoLabel));
-  } else if (effective_input_size < effective_output_size) {
-    for (size_t pad = effective_output_size - effective_input_size; pad != 0;
-         --pad, ++opos)
-      cur = trie_.Insert(cur, InputOutputLabel(kNoLabel, output[opos]));
-  }
-  CHECK_EQ(iend - ipos, oend - opos);
-  for (; ipos != iend; ++ipos, ++opos)
-    cur = trie_.Insert(cur, InputOutputLabel(input[ipos], output[opos]));
-  // We only need to attach final weight when there is an output
-  // end-of-sentence. When there is only end-of-sentence on the input,
-  // they are all consumed as the end-of-sentence paddings from
-  // `LinearFstImpl<>::ShiftBuffer()`. `LinearFstImpl<>::Expand()`
-  // and `LinearFstImpl<>::MatchInput()` ensures no other
-  // transition takes place after consuming the padding.
-  if (num_output_end > 0 || (output.empty() && num_input_end > future_size_))
-    trie_[cur].final_weight = Times(trie_[cur].final_weight, weight);
-  else
-    trie_[cur].weight = Times(trie_[cur].weight, weight);
-
-  return true;
-}
-
-template <class A>
-FeatureGroup<A> *FeatureGroupBuilder<A>::Dump(size_t max_future_size) {
-  if (error_) {
-    FSTERROR() << "Calling FeatureGroupBuilder<>::PreAccumulateWeights() "
-               << "at error state";
-    return nullptr;
-  }
-
-  if (max_future_size < future_size_) {
-    error_ = true;
-    FSTERROR() << "max_future_size (= " << max_future_size
-               << ") is smaller the builder's future_size (= " << future_size_
-               << ")";
-    return nullptr;
-  }
-
-  BuildBackLinks();
-  if (error_) return nullptr;
-  PreAccumulateWeights();  // does not fail
-
-  FeatureGroup<A> *ret =
-      new FeatureGroup<A>(max_future_size - future_size_, start_);
-
-  // Walk around the trie to compute next states
-  ret->next_state_.resize(trie_.NumNodes());
-  const Topology &topology = trie_.TrieTopology();
-  for (int i = 0; i < topology.NumNodes(); ++i) {
-    int next = i;
-    while (next != topology.Root() && topology.ChildrenOf(next).empty() &&
-           trie_[next].final_weight ==
-               trie_[trie_[next].back_link].final_weight)
-      next = trie_[next].back_link;
-    ret->next_state_[i] = next;
-  }
-
-  // Copy the trie
-  typename FeatureGroup<A>::Trie store_trie(trie_);
-  ret->trie_.swap(store_trie);
-
-  // Put the builder at error state to prevent repeated call of `Dump()`.
-  error_ = true;
-  return ret;
-}
-
-template <class A>
-int FeatureGroupBuilder<A>::FindFirstMatch(InputOutputLabel label, int parent,
-                                           int *hop) const {
-  int hop_count = 0;
-  int ret = kNoTrieNodeId;
-  for (; parent >= 0; parent = trie_[parent].back_link, ++hop_count) {
-    int next = trie_.Find(parent, label);
-    if (next != kNoTrieNodeId) {
-      ret = next;
-      break;
-    }
-  }
-  if (hop != nullptr) *hop = hop_count;
-  return ret;
-}
-
-template <class A>
-void FeatureGroupBuilder<A>::BuildBackLinks() {
-  // Breadth first search from the root. In the case where we only
-  // have the input label, the immedate back-off is simply the longest
-  // suffix of the current node that is also in the trie. For a node
-  // reached from its parent with label L, we can simply walk through
-  // the parent's back-off chain to find the first state with an arc
-  // of the same label L. The uniqueness is always
-  // guanranteed. However, in the case with both input and output
-  // labels, it is possible to back off by removing first labels from
-  // either side, which in general causes non-uniqueness.
-
-  const Topology &topology = trie_.TrieTopology();
-  std::queue<int> q;  // all enqueued or visited nodes have known links
-
-  // Note: nodes have back link initialized to -1 in their
-  // constructor.
-  q.push(trie_.Root());
-  while (!error_ && !q.empty()) {
-    int parent = q.front();
-    q.pop();
-    // Find links for every child
-    const typename Topology::NextMap &children = topology.ChildrenOf(parent);
-    for (typename Topology::NextMap::const_iterator eit = children.begin();
-         eit != children.end(); ++eit) {
-      const std::pair<InputOutputLabel, int> &edge = *eit;
-      InputOutputLabel label = edge.first;
-      int child = edge.second;
-      if (label.input == kNoLabel || label.output == kNoLabel) {
-        // Label pairs from root to here all have one and only one
-        // `kNoLabel` on the same side; equivalent to the
-        // "longest-suffix" case.
-        trie_[child].back_link =
-            FindFirstMatch(label, trie_[parent].back_link, nullptr);
-      } else {
-        // Neither side is `kNoLabel` at this point, there are
-        // three possible ways to back-off: if the parent backs
-        // off to some context with only one side non-empty, the
-        // empty side may remain empty; or else an exact match of
-        // both sides is needed. Try to find all three possible
-        // backs and look for the closest one (in terms of hops
-        // along the parent's back-off chain).
-        int only_input_hop, only_output_hop, full_hop;
-        int only_input_link =
-                FindFirstMatch(InputOutputLabel(label.input, kNoLabel), parent,
-                               &only_input_hop),
-            only_output_link =
-                FindFirstMatch(InputOutputLabel(kNoLabel, label.output), parent,
-                               &only_output_hop),
-            full_link =
-                FindFirstMatch(label, trie_[parent].back_link, &full_hop);
-        if (only_input_link != -1 && only_output_link != -1) {
-          error_ = true;
-          FSTERROR() << "Branching back-off chain:\n"
-                     << "\tnode " << child << ": " << TriePath(child, topology)
-                     << "\n"
-                     << "\tcan back-off to node " << only_input_link << ": "
-                     << TriePath(only_input_link, topology) << "\n"
-                     << "\tcan back-off to node " << only_output_link << ": "
-                     << TriePath(only_output_link, topology);
-          return;
-        } else if (full_link != -1) {
-          ++full_hop;
-          if (full_hop <= only_input_hop && full_hop <= only_output_hop) {
-            trie_[child].back_link = full_link;
-          } else {
-            error_ = true;
-            int problem_link = only_input_link != kNoTrieNodeId
-                                   ? only_input_link
-                                   : only_output_link;
-            CHECK_NE(problem_link, kNoTrieNodeId);
-            FSTERROR() << "Branching back-off chain:\n"
-                       << "\tnode " << child << ": "
-                       << TriePath(child, topology) << "\n"
-                       << "\tcan back-off to node " << full_link << ": "
-                       << TriePath(full_link, topology) << "\n"
-                       << "tcan back-off to node " << problem_link << ": "
-                       << TriePath(problem_link, topology);
-            return;
-          }
-        } else {
-          trie_[child].back_link =
-              only_input_link != -1 ? only_input_link : only_output_link;
-        }
-      }
-      if (error_) break;
-      // Point to empty context (root) when no back-off can be found
-      if (trie_[child].back_link == -1) trie_[child].back_link = 0;
-      q.push(child);
-    }
-  }
-}
-
-template <class A>
-void FeatureGroupBuilder<A>::PreAccumulateWeights() {
-  std::vector<bool> visited(trie_.NumNodes(), false);
-  visited[trie_.Root()] = true;
-
-  for (size_t i = 0; i != trie_.NumNodes(); ++i) {
-    std::stack<int> back_offs;
-    for (int j = i; !visited[j]; j = trie_[j].back_link) back_offs.push(j);
-    while (!back_offs.empty()) {
-      int j = back_offs.top();
-      back_offs.pop();
-      WeightBackLink &node = trie_[j];
-      node.weight = Times(node.weight, trie_[node.back_link].weight);
-      node.final_weight =
-          Times(node.final_weight, trie_[node.back_link].final_weight);
-      visited[j] = true;
-    }
-  }
-}
-
-template <class A>
-bool FeatureGroupBuilder<A>::TrieDfs(
-    const Topology &topology, int cur, int target,
-    std::vector<InputOutputLabel> *path) const {
-  if (cur == target) return true;
-  const typename Topology::NextMap &children = topology.ChildrenOf(cur);
-  for (typename Topology::NextMap::const_iterator eit = children.begin();
-       eit != children.end(); ++eit) {
-    const std::pair<InputOutputLabel, int> &edge = *eit;
-    path->push_back(edge.first);
-    if (TrieDfs(topology, edge.second, target, path)) return true;
-    path->pop_back();
-  }
-  return false;
-}
-
-template <class A>
-string FeatureGroupBuilder<A>::TriePath(int node,
-                                        const Topology &topology) const {
-  std::vector<InputOutputLabel> labels;
-  TrieDfs(topology, topology.Root(), node, &labels);
-  bool first = true;
-  std::ostringstream strm;
-  for (typename std::vector<InputOutputLabel>::const_iterator it =
-           labels.begin();
-       it != labels.end(); ++it) {
-    InputOutputLabel i = *it;
-    if (first)
-      first = false;
-    else
-      strm << ", ";
-    strm << "(" << TranslateLabel(i.input, fsyms_) << ", "
-         << TranslateLabel(i.output, osyms_) << ")";
-  }
-  return strm.str();
-}
-
-inline string TranslateLabel(int64 label, const SymbolTable *syms) {
-  string ret;
-  if (syms != nullptr) ret += syms->Find(label);
-  if (ret.empty()) {
-    std::ostringstream strm;
-    strm << '<' << label << '>';
-    ret = strm.str();
-  }
-  return ret;
-}
-
-template <class Iterator>
-string JoinLabels(Iterator begin, Iterator end, const SymbolTable *syms) {
-  if (begin == end) return "<empty>";
-  std::ostringstream strm;
-  bool first = true;
-  for (Iterator it = begin; it != end; ++it) {
-    if (first)
-      first = false;
-    else
-      strm << '|';
-    strm << TranslateLabel(*it, syms);
-  }
-  return strm.str();
-}
-
-template <class Label>
-string JoinLabels(const std::vector<Label> &labels, const SymbolTable *syms) {
-  return JoinLabels(labels.begin(), labels.end(), syms);
-}
-
-template <class A>
-typename A::Label GuessStartOrEnd(std::vector<typename A::Label> *sequence,
-                                  typename A::Label boundary) {
-  const size_t length = sequence->size();
-  std::vector<bool> non_boundary_on_left(length, false),
-      non_boundary_on_right(length, false);
-  for (size_t i = 1; i < length; ++i) {
-    non_boundary_on_left[i] =
-        non_boundary_on_left[i - 1] || (*sequence)[i - 1] != boundary;
-    non_boundary_on_right[length - 1 - i] = non_boundary_on_right[length - i] ||
-                                            (*sequence)[length - i] != boundary;
-  }
-  int unresolved = 0;
-  for (size_t i = 0; i < length; ++i) {
-    if ((*sequence)[i] != boundary) continue;
-    const bool left = non_boundary_on_left[i], right = non_boundary_on_right[i];
-    if (left && right) {
-      // Boundary in the middle
-      LOG(WARNING) << "Boundary label in the middle of the sequence! position: "
-                   << i << "; boundary: " << boundary
-                   << "; sequence: " << JoinLabels(*sequence, nullptr);
-      LOG(WARNING)
-          << "This is an invalid sequence anyway so I will set it to start.";
-      (*sequence)[i] = LinearFstData<A>::kStartOfSentence;
-    } else if (left && !right) {
-      // Can only be end
-      (*sequence)[i] = LinearFstData<A>::kEndOfSentence;
-    } else if (!left && right) {
-      // Can only be start
-      (*sequence)[i] = LinearFstData<A>::kStartOfSentence;
-    } else {
-      // !left && !right; can't really tell
-      ++unresolved;
-    }
-  }
-  return unresolved;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_BUILDER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linear-fst-data.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linear-fst-data.h
deleted file mode 100644
index 3b39c29d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linear-fst-data.h
+++ /dev/null
@@ -1,526 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Data structures for storing and looking up the actual feature weights.
-
-#ifndef FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_H_
-#define FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_H_
-
-#include <memory>
-#include <numeric>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/fst.h>
-
-#include <fst/extensions/linear/trie.h>
-
-namespace fst {
-
-// Forward declarations
-template <class A>
-class LinearFstDataBuilder;
-template <class A>
-class FeatureGroup;
-
-// Immutable data storage of the feature weights in a linear
-// model. Produces state tuples that represent internal states of a
-// LinearTaggerFst. Object of this class can only be constructed via
-// either `LinearFstDataBuilder::Dump()` or `LinearFstData::Read()`
-// and usually used as refcount'd object shared across mutiple
-// `LinearTaggerFst` copies.
-//
-// TODO(wuke): more efficient trie implementation
-template <class A>
-class LinearFstData {
- public:
-  friend class LinearFstDataBuilder<A>;  // For builder access
-
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-
-  // Sentence boundary labels. Both of them are negative labels other
-  // than `kNoLabel`.
-  static const Label kStartOfSentence;
-  static const Label kEndOfSentence;
-
-  // Constructs empty data; for non-trivial ways of construction see
-  // `Read()` and `LinearFstDataBuilder`.
-  LinearFstData()
-      : max_future_size_(0), max_input_label_(1), input_attribs_(1) {}
-
-  // Appends the state tuple of the start state to `output`, where
-  // each tuple holds the node ids of a trie for each feature group.
-  void EncodeStartState(std::vector<Label> *output) const {
-    for (int i = 0; i < NumGroups(); ++i) output->push_back(GroupStartState(i));
-  }
-
-  // Takes a transition from the trie states stored in
-  // `(trie_state_begin, trie_state_end)` with input label `ilabel`
-  // and output label `olabel`; appends the destination state tuple to
-  // `next` and multiplies the weight of the transition onto
-  // `weight`. `next` should be the shifted input buffer of the caller
-  // in `LinearTaggerFstImpl` (i.e. of size `LinearTaggerFstImpl::delay_`;
-  // the last element is `ilabel`).
-  template <class Iterator>
-  void TakeTransition(Iterator buffer_end, Iterator trie_state_begin,
-                      Iterator trie_state_end, Label ilabel, Label olabel,
-                      std::vector<Label> *next, Weight *weight) const;
-
-  // Returns the final weight of the given trie state sequence.
-  template <class Iterator>
-  Weight FinalWeight(Iterator trie_state_begin, Iterator trie_state_end) const;
-
-  // Returns the start trie state of the given group.
-  Label GroupStartState(int group_id) const {
-    return groups_[group_id]->Start();
-  }
-
-  // Takes a transition only within the given group. Returns the
-  // destination trie state and multiplies the weight onto `weight`.
-  Label GroupTransition(int group_id, int trie_state, Label ilabel,
-                        Label olabel, Weight *weight) const;
-
-  // Returns the final weight of the given trie state in the given group.
-  Weight GroupFinalWeight(int group_id, int trie_state) const {
-    return groups_[group_id]->FinalWeight(trie_state);
-  }
-
-  Label MinInputLabel() const { return 1; }
-
-  Label MaxInputLabel() const { return max_input_label_; }
-
-  // Returns the maximum future size of all feature groups. Future is
-  // the look-ahead window of a feature, e.g. if a feature looks at
-  // the next 2 words after the current input, then the future size is
-  // 2. There is no look-ahead for output. Features inside a single
-  // `FeatureGroup` must have equal future size.
-  size_t MaxFutureSize() const { return max_future_size_; }
-
-  // Returns the number of feature groups
-  size_t NumGroups() const { return groups_.size(); }
-
-  // Returns the range of possible output labels for an input label.
-  std::pair<typename std::vector<Label>::const_iterator,
-            typename std::vector<Label>::const_iterator>
-  PossibleOutputLabels(Label word) const;
-
-  static LinearFstData<A> *Read(std::istream &strm);  // NOLINT
-  std::ostream &Write(std::ostream &strm) const;      // NOLINT
-
- private:
-  // Offsets in `output_pool_`
-  struct InputAttribute {
-    size_t output_begin, output_length;
-
-    std::istream &Read(std::istream &strm);         // NOLINT
-    std::ostream &Write(std::ostream &strm) const;  // NOLINT
-  };
-
-  // Mapping from input label to per-group feature label
-  class GroupFeatureMap;
-
-  // Translates the input label into input feature label of group
-  // `group`; returns `kNoLabel` when there is no feature for that
-  // group.
-  Label FindFeature(size_t group, Label word) const;
-
-  size_t max_future_size_;
-  Label max_input_label_;
-  std::vector<std::unique_ptr<const FeatureGroup<A>>> groups_;
-  std::vector<InputAttribute> input_attribs_;
-  std::vector<Label> output_pool_, output_set_;
-  GroupFeatureMap group_feat_map_;
-
-  LinearFstData(const LinearFstData &) = delete;
-  LinearFstData &operator=(const LinearFstData &) = delete;
-};
-
-template <class A>
-const typename A::Label LinearFstData<A>::kStartOfSentence = -3;
-template <class A>
-const typename A::Label LinearFstData<A>::kEndOfSentence = -2;
-
-template <class A>
-template <class Iterator>
-void LinearFstData<A>::TakeTransition(Iterator buffer_end,
-                                      Iterator trie_state_begin,
-                                      Iterator trie_state_end, Label ilabel,
-                                      Label olabel, std::vector<Label> *next,
-                                      Weight *weight) const {
-  DCHECK_EQ(trie_state_end - trie_state_begin, groups_.size());
-  DCHECK(ilabel > 0 || ilabel == kEndOfSentence);
-  DCHECK(olabel > 0 || olabel == kStartOfSentence);
-  size_t group_id = 0;
-  for (Iterator it = trie_state_begin; it != trie_state_end; ++it, ++group_id) {
-    size_t delay = groups_[group_id]->Delay();
-    // On the buffer, there may also be `kStartOfSentence` from the
-    // initial empty buffer.
-    Label real_ilabel = delay == 0 ? ilabel : *(buffer_end - delay);
-    next->push_back(
-        GroupTransition(group_id, *it, real_ilabel, olabel, weight));
-  }
-}
-
-template <class A>
-typename A::Label LinearFstData<A>::GroupTransition(int group_id,
-                                                    int trie_state,
-                                                    Label ilabel, Label olabel,
-                                                    Weight *weight) const {
-  Label group_ilabel = FindFeature(group_id, ilabel);
-  return groups_[group_id]->Walk(trie_state, group_ilabel, olabel, weight);
-}
-
-template <class A>
-template <class Iterator>
-inline typename A::Weight LinearFstData<A>::FinalWeight(
-    Iterator trie_state_begin, Iterator trie_state_end) const {
-  DCHECK_EQ(trie_state_end - trie_state_begin, groups_.size());
-  size_t group_id = 0;
-  Weight accum = Weight::One();
-  for (Iterator it = trie_state_begin; it != trie_state_end; ++it, ++group_id)
-    accum = Times(accum, GroupFinalWeight(group_id, *it));
-  return accum;
-}
-
-template <class A>
-inline std::pair<typename std::vector<typename A::Label>::const_iterator,
-                 typename std::vector<typename A::Label>::const_iterator>
-LinearFstData<A>::PossibleOutputLabels(Label word) const {
-  const InputAttribute &attrib = input_attribs_[word];
-  if (attrib.output_length == 0)
-    return std::make_pair(output_set_.begin(), output_set_.end());
-  else
-    return std::make_pair(
-        output_pool_.begin() + attrib.output_begin,
-        output_pool_.begin() + attrib.output_begin + attrib.output_length);
-}
-
-template <class A>
-inline LinearFstData<A> *LinearFstData<A>::Read(std::istream &strm) {  // NOLINT
-  std::unique_ptr<LinearFstData<A>> data(new LinearFstData<A>());
-  ReadType(strm, &(data->max_future_size_));
-  ReadType(strm, &(data->max_input_label_));
-  // Feature groups
-  size_t num_groups = 0;
-  ReadType(strm, &num_groups);
-  data->groups_.resize(num_groups);
-  for (size_t i = 0; i < num_groups; ++i)
-    data->groups_[i].reset(FeatureGroup<A>::Read(strm));
-  // Other data
-  ReadType(strm, &(data->input_attribs_));
-  ReadType(strm, &(data->output_pool_));
-  ReadType(strm, &(data->output_set_));
-  ReadType(strm, &(data->group_feat_map_));
-  if (strm) {
-    return data.release();
-  } else {
-    return nullptr;
-  }
-}
-
-template <class A>
-inline std::ostream &LinearFstData<A>::Write(
-    std::ostream &strm) const {  // NOLINT
-  WriteType(strm, max_future_size_);
-  WriteType(strm, max_input_label_);
-  // Feature groups
-  WriteType(strm, groups_.size());
-  for (size_t i = 0; i < groups_.size(); ++i) {
-    groups_[i]->Write(strm);
-  }
-  // Other data
-  WriteType(strm, input_attribs_);
-  WriteType(strm, output_pool_);
-  WriteType(strm, output_set_);
-  WriteType(strm, group_feat_map_);
-  return strm;
-}
-
-template <class A>
-typename A::Label LinearFstData<A>::FindFeature(size_t group,
-                                                Label word) const {
-  DCHECK(word > 0 || word == kStartOfSentence || word == kEndOfSentence);
-  if (word == kStartOfSentence || word == kEndOfSentence)
-    return word;
-  else
-    return group_feat_map_.Find(group, word);
-}
-
-template <class A>
-inline std::istream &LinearFstData<A>::InputAttribute::Read(
-    std::istream &strm) {  // NOLINT
-  ReadType(strm, &output_begin);
-  ReadType(strm, &output_length);
-  return strm;
-}
-
-template <class A>
-inline std::ostream &LinearFstData<A>::InputAttribute::Write(
-    std::ostream &strm) const {  // NOLINT
-  WriteType(strm, output_begin);
-  WriteType(strm, output_length);
-  return strm;
-}
-
-// Forward declaration
-template <class A>
-class FeatureGroupBuilder;
-
-// An immutable grouping of features with similar context shape. Like
-// `LinearFstData`, this can only be constructed via `Read()` or
-// via its builder.
-//
-// Internally it uses a trie to store all feature n-grams and their
-// weights. The label of a trie edge is a pair (feat, olabel) of
-// labels. They can be either positive (ordinary label), `kNoLabel`,
-// `kStartOfSentence`, or `kEndOfSentence`. `kNoLabel` usually means
-// matching anything, with one exception: from the root of the trie,
-// there is a special (kNoLabel, kNoLabel) that leads to the implicit
-// start-of-sentence state. This edge is never actually matched
-// (`FindFirstMatch()` ensures this).
-template <class A>
-class FeatureGroup {
- public:
-  friend class FeatureGroupBuilder<A>;  // for builder access
-
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-
-  int Start() const { return start_; }
-
-  // Finds destination node from `cur` by consuming `ilabel` and
-  // `olabel`. The transition weight is multiplied onto `weight`.
-  int Walk(int cur, Label ilabel, Label olabel, Weight *weight) const;
-
-  // Returns the final weight of the current trie state. Only valid if
-  // the state is already known to be part of a final state (see
-  // `LinearFstData<>::CanBeFinal()`).
-  Weight FinalWeight(int trie_state) const {
-    return trie_[trie_state].final_weight;
-  }
-
-  static FeatureGroup<A> *Read(std::istream &strm) {  // NOLINT
-    size_t delay;
-    ReadType(strm, &delay);
-    int start;
-    ReadType(strm, &start);
-    Trie trie;
-    ReadType(strm, &trie);
-    std::unique_ptr<FeatureGroup<A>> ret(new FeatureGroup<A>(delay, start));
-    ret->trie_.swap(trie);
-    ReadType(strm, &ret->next_state_);
-    if (strm) {
-      return ret.release();
-    } else {
-      return nullptr;
-    }
-  }
-
-  std::ostream &Write(std::ostream &strm) const {  // NOLINT
-    WriteType(strm, delay_);
-    WriteType(strm, start_);
-    WriteType(strm, trie_);
-    WriteType(strm, next_state_);
-    return strm;
-  }
-
-  size_t Delay() const { return delay_; }
-
-  string Stats() const;
-
- private:
-  // Label along the arcs on the trie. `kNoLabel` means anything
-  // (non-negative label) can match; both sides holding `kNoLabel`
-  // is not allow; otherwise the label is > 0 (enforced by
-  // `LinearFstDataBuilder::AddWeight()`).
-  struct InputOutputLabel;
-  struct InputOutputLabelHash;
-
-  // Data to be stored on the trie
-  struct WeightBackLink {
-    int back_link;
-    Weight weight, final_weight;
-
-    WeightBackLink()
-        : back_link(kNoTrieNodeId),
-          weight(Weight::One()),
-          final_weight(Weight::One()) {}
-
-    std::istream &Read(std::istream &strm) {  // NOLINT
-      ReadType(strm, &back_link);
-      ReadType(strm, &weight);
-      ReadType(strm, &final_weight);
-      return strm;
-    }
-
-    std::ostream &Write(std::ostream &strm) const {  // NOLINT
-      WriteType(strm, back_link);
-      WriteType(strm, weight);
-      WriteType(strm, final_weight);
-      return strm;
-    }
-  };
-
-  typedef FlatTrieTopology<InputOutputLabel, InputOutputLabelHash> Topology;
-  typedef MutableTrie<InputOutputLabel, WeightBackLink, Topology> Trie;
-
-  explicit FeatureGroup(size_t delay, int start)
-      : delay_(delay), start_(start) {}
-
-  // Finds the first node with an arc with `label` following the
-  // back-off chain of `parent`. Returns the node index or
-  // `kNoTrieNodeId` when not found.
-  int FindFirstMatch(InputOutputLabel label, int parent) const;
-
-  size_t delay_;
-  int start_;
-  Trie trie_;
-  // Where to go after hitting this state. When we reach a state with
-  // no child and with no additional final weight (i.e. its final
-  // weight is the same as its back-off), we can immediately go to its
-  // back-off state.
-  std::vector<int> next_state_;
-
-  FeatureGroup(const FeatureGroup &) = delete;
-  FeatureGroup &operator=(const FeatureGroup &) = delete;
-};
-
-template <class A>
-struct FeatureGroup<A>::InputOutputLabel {
-  Label input, output;
-
-  InputOutputLabel(Label i = kNoLabel, Label o = kNoLabel)
-      : input(i), output(o) {}
-
-  bool operator==(InputOutputLabel that) const {
-    return input == that.input && output == that.output;
-  }
-
-  std::istream &Read(std::istream &strm) {  // NOLINT
-    ReadType(strm, &input);
-    ReadType(strm, &output);
-    return strm;
-  }
-
-  std::ostream &Write(std::ostream &strm) const {  // NOLINT
-    WriteType(strm, input);
-    WriteType(strm, output);
-    return strm;
-  }
-};
-
-template <class A>
-struct FeatureGroup<A>::InputOutputLabelHash {
-  size_t operator()(InputOutputLabel label) const {
-    return static_cast<size_t>(label.input * 7853 + label.output);
-  }
-};
-
-template <class A>
-int FeatureGroup<A>::Walk(int cur, Label ilabel, Label olabel,
-                          Weight *weight) const {
-  // Note: user of this method need to ensure `ilabel` and `olabel`
-  // are valid (e.g. see DCHECKs in
-  // `LinearFstData<>::TakeTransition()` and
-  // `LinearFstData<>::FindFeature()`).
-  int next;
-  if (ilabel == LinearFstData<A>::kStartOfSentence) {
-    // An observed start-of-sentence only occurs in the beginning of
-    // the input, when this feature group is delayed (i.e. there is
-    // another feature group with a larger future size). The actual
-    // input hasn't arrived so stay at the start state.
-    DCHECK_EQ(cur, start_);
-    next = start_;
-  } else {
-    // First, try exact match
-    next = FindFirstMatch(InputOutputLabel(ilabel, olabel), cur);
-    // Then try with don't cares
-    if (next == kNoTrieNodeId)
-      next = FindFirstMatch(InputOutputLabel(ilabel, kNoLabel), cur);
-    if (next == kNoTrieNodeId)
-      next = FindFirstMatch(InputOutputLabel(kNoLabel, olabel), cur);
-    // All failed, go to empty context
-    if (next == kNoTrieNodeId) next = trie_.Root();
-    *weight = Times(*weight, trie_[next].weight);
-    next = next_state_[next];
-  }
-  return next;
-}
-
-template <class A>
-inline int FeatureGroup<A>::FindFirstMatch(InputOutputLabel label,
-                                           int parent) const {
-  if (label.input == kNoLabel && label.output == kNoLabel)
-    return kNoTrieNodeId;  // very important; see class doc.
-  for (; parent != kNoTrieNodeId; parent = trie_[parent].back_link) {
-    int next = trie_.Find(parent, label);
-    if (next != kNoTrieNodeId) return next;
-  }
-  return kNoTrieNodeId;
-}
-
-template <class A>
-inline string FeatureGroup<A>::Stats() const {
-  std::ostringstream strm;
-  int num_states = 2;
-  for (int i = 2; i < next_state_.size(); ++i)
-    num_states += i == next_state_[i];
-  strm << trie_.NumNodes() << " node(s); " << num_states << " state(s)";
-  return strm.str();
-}
-
-template <class A>
-class LinearFstData<A>::GroupFeatureMap {
- public:
-  GroupFeatureMap() {}
-
-  void Init(size_t num_groups, size_t num_words) {
-    num_groups_ = num_groups;
-    pool_.clear();
-    pool_.resize(num_groups * num_words, kNoLabel);
-  }
-
-  Label Find(size_t group_id, Label ilabel) const {
-    return pool_[IndexOf(group_id, ilabel)];
-  }
-
-  bool Set(size_t group_id, Label ilabel, Label feat) {
-    size_t i = IndexOf(group_id, ilabel);
-    if (pool_[i] != kNoLabel && pool_[i] != feat) {
-      FSTERROR() << "Feature group " << group_id
-                 << " already has feature for word " << ilabel;
-      return false;
-    }
-    pool_[i] = feat;
-    return true;
-  }
-
-  std::istream &Read(std::istream &strm) {  // NOLINT
-    ReadType(strm, &num_groups_);
-    ReadType(strm, &pool_);
-    return strm;
-  }
-
-  std::ostream &Write(std::ostream &strm) const {  // NOLINT
-    WriteType(strm, num_groups_);
-    WriteType(strm, pool_);
-    return strm;
-  }
-
- private:
-  size_t IndexOf(size_t group_id, Label ilabel) const {
-    return ilabel * num_groups_ + group_id;
-  }
-
-  size_t num_groups_;
-  // `pool_[ilabel * num_groups_ + group_id]` is the feature active
-  // for group `group_id` with input `ilabel`
-  std::vector<Label> pool_;
-};
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_LINEAR_LINEAR_FST_DATA_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linear-fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linear-fst.h
deleted file mode 100644
index a2c19e73..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linear-fst.h
+++ /dev/null
@@ -1,1147 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes for building, storing and representing log-linear models as FSTs.
-
-#ifndef FST_EXTENSIONS_LINEAR_LINEAR_FST_H_
-#define FST_EXTENSIONS_LINEAR_LINEAR_FST_H_
-
-#include <algorithm>
-#include <iostream>
-#include <memory>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/log.h>
-#include <fst/extensions/pdt/collection.h>
-#include <fst/bi-table.h>
-#include <fst/cache.h>
-#include <fstream>
-#include <fst/fst.h>
-#include <fst/matcher.h>
-#include <fst/symbol-table.h>
-
-#include <fst/extensions/linear/linear-fst-data.h>
-
-namespace fst {
-
-// Forward declaration of the specialized matcher for both
-// LinearTaggerFst and LinearClassifierFst.
-template <class F>
-class LinearFstMatcherTpl;
-
-namespace internal {
-
-// Implementation class for on-the-fly generated LinearTaggerFst with
-// special optimization in matching.
-template <class A>
-class LinearTaggerFstImpl : public CacheImpl<A> {
- public:
-  using FstImpl<A>::SetType;
-  using FstImpl<A>::SetProperties;
-  using FstImpl<A>::SetInputSymbols;
-  using FstImpl<A>::SetOutputSymbols;
-  using FstImpl<A>::WriteHeader;
-
-  using CacheBaseImpl<CacheState<A>>::PushArc;
-  using CacheBaseImpl<CacheState<A>>::HasArcs;
-  using CacheBaseImpl<CacheState<A>>::HasFinal;
-  using CacheBaseImpl<CacheState<A>>::HasStart;
-  using CacheBaseImpl<CacheState<A>>::SetArcs;
-  using CacheBaseImpl<CacheState<A>>::SetFinal;
-  using CacheBaseImpl<CacheState<A>>::SetStart;
-
-  typedef A Arc;
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-  typedef typename A::StateId StateId;
-  typedef typename Collection<StateId, Label>::SetIterator NGramIterator;
-
-  // Constructs an empty FST by default.
-  LinearTaggerFstImpl()
-      : CacheImpl<A>(CacheOptions()),
-        data_(std::make_shared<LinearFstData<A>>()),
-        delay_(0) {
-    SetType("linear-tagger");
-  }
-
-  // Constructs the FST with given data storage and symbol
-  // tables.
-  //
-  // TODO(wuke): when there is no constraint on output we can delay
-  // less than `data->MaxFutureSize` positions.
-  LinearTaggerFstImpl(const LinearFstData<Arc> *data, const SymbolTable *isyms,
-                      const SymbolTable *osyms, CacheOptions opts)
-      : CacheImpl<A>(opts), data_(data), delay_(data->MaxFutureSize()) {
-    SetType("linear-tagger");
-    SetProperties(kILabelSorted, kFstProperties);
-    SetInputSymbols(isyms);
-    SetOutputSymbols(osyms);
-    ReserveStubSpace();
-  }
-
-  // Copy by sharing the underlying data storage.
-  LinearTaggerFstImpl(const LinearTaggerFstImpl &impl)
-      : CacheImpl<A>(impl), data_(impl.data_), delay_(impl.delay_) {
-    SetType("linear-tagger");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-    ReserveStubSpace();
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      StateId start = FindStartState();
-      SetStart(start);
-    }
-    return CacheImpl<A>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) {
-      state_stub_.clear();
-      FillState(s, &state_stub_);
-      if (CanBeFinal(state_stub_))
-        SetFinal(s, data_->FinalWeight(InternalBegin(state_stub_),
-                                       InternalEnd(state_stub_)));
-      else
-        SetFinal(s, Weight::Zero());
-    }
-    return CacheImpl<A>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<A>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<A>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<A>::NumOutputEpsilons(s);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<A> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<A>::InitArcIterator(s, data);
-  }
-
-  // Computes the outgoing transitions from a state, creating new
-  // destination states as needed.
-  void Expand(StateId s);
-
-  // Appends to `arcs` all out-going arcs from state `s` that matches `label` as
-  // the input label.
-  void MatchInput(StateId s, Label ilabel, std::vector<Arc> *arcs);
-
-  static LinearTaggerFstImpl *Read(std::istream &strm,
-                                   const FstReadOptions &opts);
-
-  bool Write(std::ostream &strm,  // NOLINT
-             const FstWriteOptions &opts) const {
-    FstHeader header;
-    header.SetStart(kNoStateId);
-    WriteHeader(strm, opts, kFileVersion, &header);
-    data_->Write(strm);
-    if (!strm) {
-      LOG(ERROR) << "LinearTaggerFst::Write: Write failed: " << opts.source;
-      return false;
-    }
-    return true;
-  }
-
- private:
-  static const int kMinFileVersion;
-  static const int kFileVersion;
-
-  // A collection of functions to access parts of the state tuple. A
-  // state tuple is a vector of `Label`s with two parts:
-  //   [buffer] [internal].
-  //
-  // - [buffer] is a buffer of observed input labels with length
-  // `delay_`. `LinearFstData<A>::kStartOfSentence`
-  // (resp. `LinearFstData<A>::kEndOfSentence`) are used as
-  // paddings when the buffer has fewer than `delay_` elements, which
-  // can only appear as the prefix (resp. suffix) of the buffer.
-  //
-  // - [internal] is the internal state tuple for `LinearFstData`
-  typename std::vector<Label>::const_iterator BufferBegin(
-      const std::vector<Label> &state) const {
-    return state.begin();
-  }
-
-  typename std::vector<Label>::const_iterator BufferEnd(
-      const std::vector<Label> &state) const {
-    return state.begin() + delay_;
-  }
-
-  typename std::vector<Label>::const_iterator InternalBegin(
-      const std::vector<Label> &state) const {
-    return state.begin() + delay_;
-  }
-
-  typename std::vector<Label>::const_iterator InternalEnd(
-      const std::vector<Label> &state) const {
-    return state.end();
-  }
-
-  // The size of state tuples are fixed, reserve them in stubs
-  void ReserveStubSpace() {
-    state_stub_.reserve(delay_ + data_->NumGroups());
-    next_stub_.reserve(delay_ + data_->NumGroups());
-  }
-
-  // Computes the start state tuple and maps it to the start state id.
-  StateId FindStartState() {
-    // Empty buffer with start-of-sentence paddings
-    state_stub_.clear();
-    state_stub_.resize(delay_, LinearFstData<A>::kStartOfSentence);
-    // Append internal states
-    data_->EncodeStartState(&state_stub_);
-    return FindState(state_stub_);
-  }
-
-  // Tests whether the buffer in `(begin, end)` is empty.
-  bool IsEmptyBuffer(typename std::vector<Label>::const_iterator begin,
-                     typename std::vector<Label>::const_iterator end) const {
-    // The following is guanranteed by `ShiftBuffer()`:
-    // - buffer[i] == LinearFstData<A>::kEndOfSentence =>
-    //       buffer[i+x] == LinearFstData<A>::kEndOfSentence
-    // - buffer[i] == LinearFstData<A>::kStartOfSentence =>
-    //       buffer[i-x] == LinearFstData<A>::kStartOfSentence
-    return delay_ == 0 || *(end - 1) == LinearFstData<A>::kStartOfSentence ||
-           *begin == LinearFstData<A>::kEndOfSentence;
-  }
-
-  // Tests whether the given state tuple can be a final state. A state
-  // is final iff there is no observed input in the buffer.
-  bool CanBeFinal(const std::vector<Label> &state) {
-    return IsEmptyBuffer(BufferBegin(state), BufferEnd(state));
-  }
-
-  // Finds state corresponding to an n-gram. Creates new state if n-gram not
-  // found.
-  StateId FindState(const std::vector<Label> &ngram) {
-    StateId sparse = ngrams_.FindId(ngram, true);
-    StateId dense = condensed_.FindId(sparse, true);
-    return dense;
-  }
-
-  // Appends after `output` the state tuple corresponding to the state id. The
-  // state id must exist.
-  void FillState(StateId s, std::vector<Label> *output) {
-    s = condensed_.FindEntry(s);
-    for (NGramIterator it = ngrams_.FindSet(s); !it.Done(); it.Next()) {
-      Label label = it.Element();
-      output->push_back(label);
-    }
-  }
-
-  // Shifts the buffer in `state` by appending `ilabel` and popping
-  // the one in the front as the return value. `next_stub_` is a
-  // shifted buffer of size `delay_` where the first `delay_ - 1`
-  // elements are the last `delay_ - 1` elements in the buffer of
-  // `state`. The last (if any) element in `next_stub_` will be
-  // `ilabel` after the call returns.
-  Label ShiftBuffer(const std::vector<Label> &state, Label ilabel,
-                    std::vector<Label> *next_stub_);
-
-  // Builds an arc from state tuple `state` consuming `ilabel` and
-  // `olabel`. `next_stub_` is the buffer filled in `ShiftBuffer`.
-  Arc MakeArc(const std::vector<Label> &state, Label ilabel, Label olabel,
-              std::vector<Label> *next_stub_);
-
-  // Expands arcs from state `s`, equivalent to state tuple `state`,
-  // with input `ilabel`. `next_stub_` is the buffer filled in
-  // `ShiftBuffer`.
-  void ExpandArcs(StateId s, const std::vector<Label> &state, Label ilabel,
-                  std::vector<Label> *next_stub_);
-
-  // Appends arcs from state `s`, equivalent to state tuple `state`,
-  // with input `ilabel` to `arcs`. `next_stub_` is the buffer filled
-  // in `ShiftBuffer`.
-  void AppendArcs(StateId s, const std::vector<Label> &state, Label ilabel,
-                  std::vector<Label> *next_stub_, std::vector<Arc> *arcs);
-
-  std::shared_ptr<const LinearFstData<A>> data_;
-  size_t delay_;
-  // Mapping from internal state tuple to *non-consecutive* ids
-  Collection<StateId, Label> ngrams_;
-  // Mapping from non-consecutive id to actual state id
-  CompactHashBiTable<StateId, StateId, std::hash<StateId>> condensed_;
-  // Two frequently used vectors, reuse to avoid repeated heap
-  // allocation
-  std::vector<Label> state_stub_, next_stub_;
-
-  LinearTaggerFstImpl &operator=(const LinearTaggerFstImpl &) = delete;
-};
-
-template <class A>
-const int LinearTaggerFstImpl<A>::kMinFileVersion = 1;
-
-template <class A>
-const int LinearTaggerFstImpl<A>::kFileVersion = 1;
-
-template <class A>
-inline typename A::Label LinearTaggerFstImpl<A>::ShiftBuffer(
-    const std::vector<Label> &state, Label ilabel,
-    std::vector<Label> *next_stub_) {
-  DCHECK(ilabel > 0 || ilabel == LinearFstData<A>::kEndOfSentence);
-  if (delay_ == 0) {
-    DCHECK_GT(ilabel, 0);
-    return ilabel;
-  } else {
-    (*next_stub_)[BufferEnd(*next_stub_) - next_stub_->begin() - 1] = ilabel;
-    return *BufferBegin(state);
-  }
-}
-
-template <class A>
-inline A LinearTaggerFstImpl<A>::MakeArc(const std::vector<Label> &state,
-                                         Label ilabel, Label olabel,
-                                         std::vector<Label> *next_stub_) {
-  DCHECK(ilabel > 0 || ilabel == LinearFstData<A>::kEndOfSentence);
-  DCHECK(olabel > 0 || olabel == LinearFstData<A>::kStartOfSentence);
-  Weight weight(Weight::One());
-  data_->TakeTransition(BufferEnd(state), InternalBegin(state),
-                        InternalEnd(state), ilabel, olabel, next_stub_,
-                        &weight);
-  StateId nextstate = FindState(*next_stub_);
-  // Restore `next_stub_` to its size before the call
-  next_stub_->resize(delay_);
-  // In the actual arc, we use epsilons instead of boundaries.
-  return A(ilabel == LinearFstData<A>::kEndOfSentence ? 0 : ilabel,
-           olabel == LinearFstData<A>::kStartOfSentence ? 0 : olabel, weight,
-           nextstate);
-}
-
-template <class A>
-inline void LinearTaggerFstImpl<A>::ExpandArcs(StateId s,
-                                               const std::vector<Label> &state,
-                                               Label ilabel,
-                                               std::vector<Label> *next_stub_) {
-  // Input label to constrain the output with, observed `delay_` steps
-  // back. `ilabel` is the input label to be put on the arc, which
-  // fires features.
-  Label obs_ilabel = ShiftBuffer(state, ilabel, next_stub_);
-  if (obs_ilabel == LinearFstData<A>::kStartOfSentence) {
-    // This happens when input is shorter than `delay_`.
-    PushArc(s, MakeArc(state, ilabel, LinearFstData<A>::kStartOfSentence,
-                       next_stub_));
-  } else {
-    std::pair<typename std::vector<typename A::Label>::const_iterator,
-              typename std::vector<typename A::Label>::const_iterator> range =
-        data_->PossibleOutputLabels(obs_ilabel);
-    for (typename std::vector<typename A::Label>::const_iterator it =
-             range.first;
-         it != range.second; ++it)
-      PushArc(s, MakeArc(state, ilabel, *it, next_stub_));
-  }
-}
-
-// TODO(wuke): this has much in duplicate with `ExpandArcs()`
-template <class A>
-inline void LinearTaggerFstImpl<A>::AppendArcs(StateId /*s*/,
-                                               const std::vector<Label> &state,
-                                               Label ilabel,
-                                               std::vector<Label> *next_stub_,
-                                               std::vector<Arc> *arcs) {
-  // Input label to constrain the output with, observed `delay_` steps
-  // back. `ilabel` is the input label to be put on the arc, which
-  // fires features.
-  Label obs_ilabel = ShiftBuffer(state, ilabel, next_stub_);
-  if (obs_ilabel == LinearFstData<A>::kStartOfSentence) {
-    // This happens when input is shorter than `delay_`.
-    arcs->push_back(
-        MakeArc(state, ilabel, LinearFstData<A>::kStartOfSentence, next_stub_));
-  } else {
-    std::pair<typename std::vector<typename A::Label>::const_iterator,
-              typename std::vector<typename A::Label>::const_iterator> range =
-        data_->PossibleOutputLabels(obs_ilabel);
-    for (typename std::vector<typename A::Label>::const_iterator it =
-             range.first;
-         it != range.second; ++it)
-      arcs->push_back(MakeArc(state, ilabel, *it, next_stub_));
-  }
-}
-
-template <class A>
-void LinearTaggerFstImpl<A>::Expand(StateId s) {
-  VLOG(3) << "Expand " << s;
-  state_stub_.clear();
-  FillState(s, &state_stub_);
-
-  // Precompute the first `delay_ - 1` elements in the buffer of
-  // next states, which are identical for different input/output.
-  next_stub_.clear();
-  next_stub_.resize(delay_);
-  if (delay_ > 0)
-    std::copy(BufferBegin(state_stub_) + 1, BufferEnd(state_stub_),
-              next_stub_.begin());
-
-  // Epsilon transition for flushing out the next observed input
-  if (!IsEmptyBuffer(BufferBegin(state_stub_), BufferEnd(state_stub_)))
-    ExpandArcs(s, state_stub_, LinearFstData<A>::kEndOfSentence, &next_stub_);
-
-  // Non-epsilon input when we haven't flushed
-  if (delay_ == 0 ||
-      *(BufferEnd(state_stub_) - 1) != LinearFstData<A>::kEndOfSentence)
-    for (Label ilabel = data_->MinInputLabel();
-         ilabel <= data_->MaxInputLabel(); ++ilabel)
-      ExpandArcs(s, state_stub_, ilabel, &next_stub_);
-
-  SetArcs(s);
-}
-
-template <class A>
-void LinearTaggerFstImpl<A>::MatchInput(StateId s, Label ilabel,
-                                        std::vector<Arc> *arcs) {
-  state_stub_.clear();
-  FillState(s, &state_stub_);
-
-  // Precompute the first `delay_ - 1` elements in the buffer of
-  // next states, which are identical for different input/output.
-  next_stub_.clear();
-  next_stub_.resize(delay_);
-  if (delay_ > 0)
-    std::copy(BufferBegin(state_stub_) + 1, BufferEnd(state_stub_),
-              next_stub_.begin());
-
-  if (ilabel == 0) {
-    // Epsilon transition for flushing out the next observed input
-    if (!IsEmptyBuffer(BufferBegin(state_stub_), BufferEnd(state_stub_)))
-      AppendArcs(s, state_stub_, LinearFstData<A>::kEndOfSentence, &next_stub_,
-                 arcs);
-  } else {
-    // Non-epsilon input when we haven't flushed
-    if (delay_ == 0 ||
-        *(BufferEnd(state_stub_) - 1) != LinearFstData<A>::kEndOfSentence)
-      AppendArcs(s, state_stub_, ilabel, &next_stub_, arcs);
-  }
-}
-
-template <class A>
-inline LinearTaggerFstImpl<A> *LinearTaggerFstImpl<A>::Read(
-    std::istream &strm, const FstReadOptions &opts) {  // NOLINT
-  std::unique_ptr<LinearTaggerFstImpl<A>> impl(new LinearTaggerFstImpl<A>());
-  FstHeader header;
-  if (!impl->ReadHeader(strm, opts, kMinFileVersion, &header)) {
-    return nullptr;
-  }
-  impl->data_ = std::shared_ptr<LinearFstData<A>>(LinearFstData<A>::Read(strm));
-  if (!impl->data_) {
-    return nullptr;
-  }
-  impl->delay_ = impl->data_->MaxFutureSize();
-  impl->ReserveStubSpace();
-  return impl.release();
-}
-
-}  // namespace internal
-
-// This class attaches interface to implementation and handles
-// reference counting, delegating most methods to ImplToFst.
-template <class A>
-class LinearTaggerFst : public ImplToFst<internal::LinearTaggerFstImpl<A>> {
- public:
-  friend class ArcIterator<LinearTaggerFst<A>>;
-  friend class StateIterator<LinearTaggerFst<A>>;
-  friend class LinearFstMatcherTpl<LinearTaggerFst<A>>;
-
-  typedef A Arc;
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-  typedef typename A::StateId StateId;
-  typedef DefaultCacheStore<A> Store;
-  typedef typename Store::State State;
-  using Impl = internal::LinearTaggerFstImpl<A>;
-
-  LinearTaggerFst() : ImplToFst<Impl>(std::make_shared<Impl>()) {}
-
-  explicit LinearTaggerFst(LinearFstData<A> *data,
-                           const SymbolTable *isyms = nullptr,
-                           const SymbolTable *osyms = nullptr,
-                           CacheOptions opts = CacheOptions())
-      : ImplToFst<Impl>(std::make_shared<Impl>(data, isyms, osyms, opts)) {}
-
-  explicit LinearTaggerFst(const Fst<A> &fst)
-      : ImplToFst<Impl>(std::make_shared<Impl>()) {
-    LOG(FATAL) << "LinearTaggerFst: no constructor from arbitrary FST.";
-  }
-
-  // See Fst<>::Copy() for doc.
-  LinearTaggerFst(const LinearTaggerFst<A> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this LinearTaggerFst. See Fst<>::Copy() for further doc.
-  LinearTaggerFst<A> *Copy(bool safe = false) const override {
-    return new LinearTaggerFst<A>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<A> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<A> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
-  MatcherBase<A> *InitMatcher(MatchType match_type) const override {
-    return new LinearFstMatcherTpl<LinearTaggerFst<A>>(*this, match_type);
-  }
-
-  static LinearTaggerFst<A> *Read(const string &filename) {
-    if (!filename.empty()) {
-      std::ifstream strm(filename,
-                              std::ios_base::in | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "LinearTaggerFst::Read: Can't open file: " << filename;
-        return nullptr;
-      }
-      return Read(strm, FstReadOptions(filename));
-    } else {
-      return Read(std::cin, FstReadOptions("standard input"));
-    }
-  }
-
-  static LinearTaggerFst<A> *Read(std::istream &in,  // NOLINT
-                                  const FstReadOptions &opts) {
-    auto *impl = Impl::Read(in, opts);
-    return impl ? new LinearTaggerFst<A>(std::shared_ptr<Impl>(impl)) : nullptr;
-  }
-
-  bool Write(const string &filename) const override {
-    if (!filename.empty()) {
-      std::ofstream strm(filename,
-                               std::ios_base::out | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "LinearTaggerFst::Write: Can't open file: " << filename;
-        return false;
-      }
-      return Write(strm, FstWriteOptions(filename));
-    } else {
-      return Write(std::cout, FstWriteOptions("standard output"));
-    }
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return GetImpl()->Write(strm, opts);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  explicit LinearTaggerFst(std::shared_ptr<Impl> impl)
-      : ImplToFst<Impl>(impl) {}
-
-  void operator=(const LinearTaggerFst<A> &fst) = delete;
-};
-
-// Specialization for LinearTaggerFst.
-template <class Arc>
-class StateIterator<LinearTaggerFst<Arc>>
-    : public CacheStateIterator<LinearTaggerFst<Arc>> {
- public:
-  explicit StateIterator(const LinearTaggerFst<Arc> &fst)
-      : CacheStateIterator<LinearTaggerFst<Arc>>(fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for LinearTaggerFst.
-template <class Arc>
-class ArcIterator<LinearTaggerFst<Arc>>
-    : public CacheArcIterator<LinearTaggerFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const LinearTaggerFst<Arc> &fst, StateId s)
-      : CacheArcIterator<LinearTaggerFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc>
-inline void LinearTaggerFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<LinearTaggerFst<Arc>>(*this);
-}
-
-namespace internal {
-
-// Implementation class for on-the-fly generated LinearClassifierFst with
-// special optimization in matching.
-template <class A>
-class LinearClassifierFstImpl : public CacheImpl<A> {
- public:
-  using FstImpl<A>::SetType;
-  using FstImpl<A>::SetProperties;
-  using FstImpl<A>::SetInputSymbols;
-  using FstImpl<A>::SetOutputSymbols;
-  using FstImpl<A>::WriteHeader;
-
-  using CacheBaseImpl<CacheState<A>>::PushArc;
-  using CacheBaseImpl<CacheState<A>>::HasArcs;
-  using CacheBaseImpl<CacheState<A>>::HasFinal;
-  using CacheBaseImpl<CacheState<A>>::HasStart;
-  using CacheBaseImpl<CacheState<A>>::SetArcs;
-  using CacheBaseImpl<CacheState<A>>::SetFinal;
-  using CacheBaseImpl<CacheState<A>>::SetStart;
-
-  typedef A Arc;
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-  typedef typename A::StateId StateId;
-  typedef typename Collection<StateId, Label>::SetIterator NGramIterator;
-
-  // Constructs an empty FST by default.
-  LinearClassifierFstImpl()
-      : CacheImpl<A>(CacheOptions()),
-        data_(std::make_shared<LinearFstData<A>>()) {
-    SetType("linear-classifier");
-    num_classes_ = 0;
-    num_groups_ = 0;
-  }
-
-  // Constructs the FST with given data storage, number of classes and
-  // symbol tables.
-  LinearClassifierFstImpl(const LinearFstData<Arc> *data, size_t num_classes,
-                          const SymbolTable *isyms, const SymbolTable *osyms,
-                          CacheOptions opts)
-      : CacheImpl<A>(opts),
-        data_(data),
-        num_classes_(num_classes),
-        num_groups_(data_->NumGroups() / num_classes_) {
-    SetType("linear-classifier");
-    SetProperties(kILabelSorted, kFstProperties);
-    SetInputSymbols(isyms);
-    SetOutputSymbols(osyms);
-    ReserveStubSpace();
-  }
-
-  // Copy by sharing the underlying data storage.
-  LinearClassifierFstImpl(const LinearClassifierFstImpl &impl)
-      : CacheImpl<A>(impl),
-        data_(impl.data_),
-        num_classes_(impl.num_classes_),
-        num_groups_(impl.num_groups_) {
-    SetType("linear-classifier");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-    ReserveStubSpace();
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      StateId start = FindStartState();
-      SetStart(start);
-    }
-    return CacheImpl<A>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) {
-      state_stub_.clear();
-      FillState(s, &state_stub_);
-      SetFinal(s, FinalWeight(state_stub_));
-    }
-    return CacheImpl<A>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<A>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<A>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<A>::NumOutputEpsilons(s);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<A> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<A>::InitArcIterator(s, data);
-  }
-
-  // Computes the outgoing transitions from a state, creating new
-  // destination states as needed.
-  void Expand(StateId s);
-
-  // Appends to `arcs` all out-going arcs from state `s` that matches
-  // `label` as the input label.
-  void MatchInput(StateId s, Label ilabel, std::vector<Arc> *arcs);
-
-  static LinearClassifierFstImpl<A> *Read(std::istream &strm,
-                                          const FstReadOptions &opts);
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
-    FstHeader header;
-    header.SetStart(kNoStateId);
-    WriteHeader(strm, opts, kFileVersion, &header);
-    data_->Write(strm);
-    WriteType(strm, num_classes_);
-    if (!strm) {
-      LOG(ERROR) << "LinearClassifierFst::Write: Write failed: " << opts.source;
-      return false;
-    }
-    return true;
-  }
-
- private:
-  static const int kMinFileVersion;
-  static const int kFileVersion;
-
-  // A collection of functions to access parts of the state tuple. A
-  // state tuple is a vector of `Label`s with two parts:
-  //   [prediction] [internal].
-  //
-  // - [prediction] is a single label of the predicted class. A state
-  //   must have a positive class label, unless it is the start state.
-  //
-  // - [internal] is the internal state tuple for `LinearFstData` of
-  //   the given class; or kNoTrieNodeId's if in start state.
-  Label &Prediction(std::vector<Label> &state) { return state[0]; }  // NOLINT
-  Label Prediction(const std::vector<Label> &state) const { return state[0]; }
-
-  Label &InternalAt(std::vector<Label> &state, int index) {  // NOLINT
-    return state[index + 1];
-  }
-  Label InternalAt(const std::vector<Label> &state, int index) const {
-    return state[index + 1];
-  }
-
-  // The size of state tuples are fixed, reserve them in stubs
-  void ReserveStubSpace() {
-    size_t size = 1 + num_groups_;
-    state_stub_.reserve(size);
-    next_stub_.reserve(size);
-  }
-
-  // Computes the start state tuple and maps it to the start state id.
-  StateId FindStartState() {
-    // A start state tuple has no prediction
-    state_stub_.clear();
-    state_stub_.push_back(kNoLabel);
-    // For a start state, we don't yet know where we are in the tries.
-    for (size_t i = 0; i < num_groups_; ++i)
-      state_stub_.push_back(kNoTrieNodeId);
-    return FindState(state_stub_);
-  }
-
-  // Tests if the state tuple represents the start state.
-  bool IsStartState(const std::vector<Label> &state) const {
-    return state[0] == kNoLabel;
-  }
-
-  // Computes the actual group id in the data storage.
-  int GroupId(Label pred, int group) const {
-    return group * num_classes_ + pred - 1;
-  }
-
-  // Finds out the final weight of the given state. A state is final
-  // iff it is not the start.
-  Weight FinalWeight(const std::vector<Label> &state) const {
-    if (IsStartState(state)) {
-      return Weight::Zero();
-    }
-    Label pred = Prediction(state);
-    DCHECK_GT(pred, 0);
-    DCHECK_LE(pred, num_classes_);
-    Weight final_weight = Weight::One();
-    for (size_t group = 0; group < num_groups_; ++group) {
-      int group_id = GroupId(pred, group);
-      int trie_state = InternalAt(state, group);
-      final_weight =
-          Times(final_weight, data_->GroupFinalWeight(group_id, trie_state));
-    }
-    return final_weight;
-  }
-
-  // Finds state corresponding to an n-gram. Creates new state if n-gram not
-  // found.
-  StateId FindState(const std::vector<Label> &ngram) {
-    StateId sparse = ngrams_.FindId(ngram, true);
-    StateId dense = condensed_.FindId(sparse, true);
-    return dense;
-  }
-
-  // Appends after `output` the state tuple corresponding to the state id. The
-  // state id must exist.
-  void FillState(StateId s, std::vector<Label> *output) {
-    s = condensed_.FindEntry(s);
-    for (NGramIterator it = ngrams_.FindSet(s); !it.Done(); it.Next()) {
-      Label label = it.Element();
-      output->push_back(label);
-    }
-  }
-
-  std::shared_ptr<const LinearFstData<A>> data_;
-  // Division of groups in `data_`; num_classes_ * num_groups_ ==
-  // data_->NumGroups().
-  size_t num_classes_, num_groups_;
-  // Mapping from internal state tuple to *non-consecutive* ids
-  Collection<StateId, Label> ngrams_;
-  // Mapping from non-consecutive id to actual state id
-  CompactHashBiTable<StateId, StateId, std::hash<StateId>> condensed_;
-  // Two frequently used vectors, reuse to avoid repeated heap
-  // allocation
-  std::vector<Label> state_stub_, next_stub_;
-
-  void operator=(const LinearClassifierFstImpl<A> &) = delete;
-};
-
-template <class A>
-const int LinearClassifierFstImpl<A>::kMinFileVersion = 0;
-
-template <class A>
-const int LinearClassifierFstImpl<A>::kFileVersion = 0;
-
-template <class A>
-void LinearClassifierFstImpl<A>::Expand(StateId s) {
-  VLOG(3) << "Expand " << s;
-  state_stub_.clear();
-  FillState(s, &state_stub_);
-  next_stub_.clear();
-  next_stub_.resize(1 + num_groups_);
-
-  if (IsStartState(state_stub_)) {
-    // Make prediction
-    for (Label pred = 1; pred <= num_classes_; ++pred) {
-      Prediction(next_stub_) = pred;
-      for (int i = 0; i < num_groups_; ++i)
-        InternalAt(next_stub_, i) = data_->GroupStartState(GroupId(pred, i));
-      PushArc(s, A(0, pred, Weight::One(), FindState(next_stub_)));
-    }
-  } else {
-    Label pred = Prediction(state_stub_);
-    DCHECK_GT(pred, 0);
-    DCHECK_LE(pred, num_classes_);
-    for (Label ilabel = data_->MinInputLabel();
-         ilabel <= data_->MaxInputLabel(); ++ilabel) {
-      Prediction(next_stub_) = pred;
-      Weight weight = Weight::One();
-      for (int i = 0; i < num_groups_; ++i)
-        InternalAt(next_stub_, i) =
-            data_->GroupTransition(GroupId(pred, i), InternalAt(state_stub_, i),
-                                   ilabel, pred, &weight);
-      PushArc(s, A(ilabel, 0, weight, FindState(next_stub_)));
-    }
-  }
-
-  SetArcs(s);
-}
-
-template <class A>
-void LinearClassifierFstImpl<A>::MatchInput(StateId s, Label ilabel,
-                                            std::vector<Arc> *arcs) {
-  state_stub_.clear();
-  FillState(s, &state_stub_);
-  next_stub_.clear();
-  next_stub_.resize(1 + num_groups_);
-
-  if (IsStartState(state_stub_)) {
-    // Make prediction if `ilabel` is epsilon.
-    if (ilabel == 0) {
-      for (Label pred = 1; pred <= num_classes_; ++pred) {
-        Prediction(next_stub_) = pred;
-        for (int i = 0; i < num_groups_; ++i)
-          InternalAt(next_stub_, i) = data_->GroupStartState(GroupId(pred, i));
-        arcs->push_back(A(0, pred, Weight::One(), FindState(next_stub_)));
-      }
-    }
-  } else if (ilabel != 0) {
-    Label pred = Prediction(state_stub_);
-    Weight weight = Weight::One();
-    Prediction(next_stub_) = pred;
-    for (int i = 0; i < num_groups_; ++i)
-      InternalAt(next_stub_, i) = data_->GroupTransition(
-          GroupId(pred, i), InternalAt(state_stub_, i), ilabel, pred, &weight);
-    arcs->push_back(A(ilabel, 0, weight, FindState(next_stub_)));
-  }
-}
-
-template <class A>
-inline LinearClassifierFstImpl<A> *LinearClassifierFstImpl<A>::Read(
-    std::istream &strm, const FstReadOptions &opts) {
-  std::unique_ptr<LinearClassifierFstImpl<A>> impl(
-      new LinearClassifierFstImpl<A>());
-  FstHeader header;
-  if (!impl->ReadHeader(strm, opts, kMinFileVersion, &header)) {
-    return nullptr;
-  }
-  impl->data_ = std::shared_ptr<LinearFstData<A>>(LinearFstData<A>::Read(strm));
-  if (!impl->data_) {
-    return nullptr;
-  }
-  ReadType(strm, &impl->num_classes_);
-  if (!strm) {
-    return nullptr;
-  }
-  impl->num_groups_ = impl->data_->NumGroups() / impl->num_classes_;
-  if (impl->num_groups_ * impl->num_classes_ != impl->data_->NumGroups()) {
-    FSTERROR() << "Total number of feature groups is not a multiple of the "
-                  "number of classes: num groups = "
-               << impl->data_->NumGroups()
-               << ", num classes = " << impl->num_classes_;
-    return nullptr;
-  }
-  impl->ReserveStubSpace();
-  return impl.release();
-}
-
-}  // namespace internal
-
-// This class attaches interface to implementation and handles
-// reference counting, delegating most methods to ImplToFst.
-template <class A>
-class LinearClassifierFst
-    : public ImplToFst<internal::LinearClassifierFstImpl<A>> {
- public:
-  friend class ArcIterator<LinearClassifierFst<A>>;
-  friend class StateIterator<LinearClassifierFst<A>>;
-  friend class LinearFstMatcherTpl<LinearClassifierFst<A>>;
-
-  typedef A Arc;
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-  typedef typename A::StateId StateId;
-  typedef DefaultCacheStore<A> Store;
-  typedef typename Store::State State;
-  using Impl = internal::LinearClassifierFstImpl<A>;
-
-  LinearClassifierFst() : ImplToFst<Impl>(std::make_shared<Impl>()) {}
-
-  explicit LinearClassifierFst(LinearFstData<A> *data, size_t num_classes,
-                               const SymbolTable *isyms = nullptr,
-                               const SymbolTable *osyms = nullptr,
-                               CacheOptions opts = CacheOptions())
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(data, num_classes, isyms, osyms, opts)) {}
-
-  explicit LinearClassifierFst(const Fst<A> &fst)
-      : ImplToFst<Impl>(std::make_shared<Impl>()) {
-    LOG(FATAL) << "LinearClassifierFst: no constructor from arbitrary FST.";
-  }
-
-  // See Fst<>::Copy() for doc.
-  LinearClassifierFst(const LinearClassifierFst<A> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this LinearClassifierFst. See Fst<>::Copy() for further doc.
-  LinearClassifierFst<A> *Copy(bool safe = false) const override {
-    return new LinearClassifierFst<A>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<A> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<A> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
-  MatcherBase<A> *InitMatcher(MatchType match_type) const override {
-    return new LinearFstMatcherTpl<LinearClassifierFst<A>>(*this, match_type);
-  }
-
-  static LinearClassifierFst<A> *Read(const string &filename) {
-    if (!filename.empty()) {
-      std::ifstream strm(filename,
-                              std::ios_base::in | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "LinearClassifierFst::Read: Can't open file: "
-                   << filename;
-        return nullptr;
-      }
-      return Read(strm, FstReadOptions(filename));
-    } else {
-      return Read(std::cin, FstReadOptions("standard input"));
-    }
-  }
-
-  static LinearClassifierFst<A> *Read(std::istream &in,
-                                      const FstReadOptions &opts) {
-    auto *impl = Impl::Read(in, opts);
-    return impl ? new LinearClassifierFst<A>(std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  bool Write(const string &filename) const override {
-    if (!filename.empty()) {
-      std::ofstream strm(filename,
-                               std::ios_base::out | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "ProdLmFst::Write: Can't open file: " << filename;
-        return false;
-      }
-      return Write(strm, FstWriteOptions(filename));
-    } else {
-      return Write(std::cout, FstWriteOptions("standard output"));
-    }
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return GetImpl()->Write(strm, opts);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  explicit LinearClassifierFst(std::shared_ptr<Impl> impl)
-      : ImplToFst<Impl>(impl) {}
-
-  void operator=(const LinearClassifierFst<A> &fst) = delete;
-};
-
-// Specialization for LinearClassifierFst.
-template <class Arc>
-class StateIterator<LinearClassifierFst<Arc>>
-    : public CacheStateIterator<LinearClassifierFst<Arc>> {
- public:
-  explicit StateIterator(const LinearClassifierFst<Arc> &fst)
-      : CacheStateIterator<LinearClassifierFst<Arc>>(fst,
-                                                     fst.GetMutableImpl()) {}
-};
-
-// Specialization for LinearClassifierFst.
-template <class Arc>
-class ArcIterator<LinearClassifierFst<Arc>>
-    : public CacheArcIterator<LinearClassifierFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const LinearClassifierFst<Arc> &fst, StateId s)
-      : CacheArcIterator<LinearClassifierFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc>
-inline void LinearClassifierFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<LinearClassifierFst<Arc>>(*this);
-}
-
-// Specialized Matcher for LinearFsts. This matcher only supports
-// matching from the input side. This is intentional because comparing
-// the scores of different input sequences with the same output
-// sequence is meaningless in a discriminative model.
-template <class F>
-class LinearFstMatcherTpl : public MatcherBase<typename F::Arc> {
- public:
-  typedef typename F::Arc Arc;
-  typedef typename Arc::Label Label;
-  typedef typename Arc::Weight Weight;
-  typedef typename Arc::StateId StateId;
-  typedef F FST;
-
-  LinearFstMatcherTpl(const FST &fst, MatchType match_type)
-      : fst_(fst.Copy()),
-        match_type_(match_type),
-        s_(kNoStateId),
-        current_loop_(false),
-        loop_(kNoLabel, 0, Weight::One(), kNoStateId),
-        cur_arc_(0),
-        error_(false) {
-    switch (match_type_) {
-      case MATCH_INPUT:
-      case MATCH_OUTPUT:
-      case MATCH_NONE:
-        break;
-      default:
-        FSTERROR() << "LinearFstMatcherTpl: Bad match type";
-        match_type_ = MATCH_NONE;
-        error_ = true;
-    }
-  }
-
-  LinearFstMatcherTpl(const LinearFstMatcherTpl<F> &matcher, bool safe = false)
-      : fst_(matcher.fst_->Copy(safe)),
-        match_type_(matcher.match_type_),
-        s_(kNoStateId),
-        current_loop_(false),
-        loop_(matcher.loop_),
-        cur_arc_(0),
-        error_(matcher.error_) {}
-
-  LinearFstMatcherTpl<F> *Copy(bool safe = false) const override {
-    return new LinearFstMatcherTpl<F>(*this, safe);
-  }
-
-  MatchType Type(bool /*test*/) const override {
-    // `MATCH_INPUT` is the only valid type
-    return match_type_ == MATCH_INPUT ? match_type_ : MATCH_NONE;
-  }
-
-  void SetState(StateId s) final {
-    if (s_ == s) return;
-    s_ = s;
-    // `MATCH_INPUT` is the only valid type
-    if (match_type_ != MATCH_INPUT) {
-      FSTERROR() << "LinearFstMatcherTpl: Bad match type";
-      error_ = true;
-    }
-    loop_.nextstate = s;
-  }
-
-  bool Find(Label label) final {
-    if (error_) {
-      current_loop_ = false;
-      return false;
-    }
-    current_loop_ = label == 0;
-    if (label == kNoLabel) label = 0;
-    arcs_.clear();
-    cur_arc_ = 0;
-    fst_->GetMutableImpl()->MatchInput(s_, label, &arcs_);
-    return current_loop_ || !arcs_.empty();
-  }
-
-  bool Done() const final {
-    return !(current_loop_ || cur_arc_ < arcs_.size());
-  }
-
-  const Arc &Value() const final {
-    return current_loop_ ? loop_ : arcs_[cur_arc_];
-  }
-
-  void Next() final {
-    if (current_loop_)
-      current_loop_ = false;
-    else
-      ++cur_arc_;
-  }
-
-  ssize_t Priority(StateId s) final { return kRequirePriority; }
-
-  const FST &GetFst() const override { return *fst_; }
-
-  uint64 Properties(uint64 props) const override {
-    if (error_) props |= kError;
-    return props;
-  }
-
-  uint32 Flags() const override { return kRequireMatch; }
-
- private:
-  std::unique_ptr<const FST> fst_;
-  MatchType match_type_;  // Type of match to perform.
-  StateId s_;             // Current state.
-  bool current_loop_;     // Current arc is the implicit loop.
-  Arc loop_;              // For non-consuming symbols.
-  // All out-going arcs matching the label in last Find() call.
-  std::vector<Arc> arcs_;
-  size_t cur_arc_;  // Index to the arc that `Value()` should return.
-  bool error_;      // Error encountered.
-};
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_LINEAR_LINEAR_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linearscript.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linearscript.h
deleted file mode 100644
index 54106d20..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/linearscript.h
+++ /dev/null
@@ -1,391 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_LINEAR_LINEARSCRIPT_H_
-#define FST_EXTENSIONS_LINEAR_LINEARSCRIPT_H_
-
-#include <istream>
-#include <sstream>
-#include <string>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/extensions/linear/linear-fst-data-builder.h>
-#include <fst/extensions/linear/linear-fst.h>
-#include <fstream>
-#include <fst/symbol-table.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/script-impl.h>
-
-DECLARE_string(delimiter);
-DECLARE_string(empty_symbol);
-DECLARE_string(start_symbol);
-DECLARE_string(end_symbol);
-DECLARE_bool(classifier);
-
-namespace fst {
-namespace script {
-typedef std::tuple<const string &, const string &, const string &, char **, int,
-                   const string &, const string &, const string &,
-                   const string &>
-    LinearCompileArgs;
-
-bool ValidateDelimiter();
-bool ValidateEmptySymbol();
-
-// Returns the proper label given the symbol. For symbols other than
-// `FLAGS_start_symbol` or `FLAGS_end_symbol`, looks up the symbol
-// table to decide the label. Depending on whether
-// `FLAGS_start_symbol` and `FLAGS_end_symbol` are identical, it
-// either returns `kNoLabel` for later processing or decides the label
-// right away.
-template <class Arc>
-inline typename Arc::Label LookUp(const string &str, SymbolTable *syms) {
-  if (str == FLAGS_start_symbol)
-    return str == FLAGS_end_symbol ? kNoLabel
-                                   : LinearFstData<Arc>::kStartOfSentence;
-  else if (str == FLAGS_end_symbol)
-    return LinearFstData<Arc>::kEndOfSentence;
-  else
-    return syms->AddSymbol(str);
-}
-
-// Splits `str` with `delim` as the delimiter and stores the labels in
-// `output`.
-template <class Arc>
-void SplitAndPush(const string &str, const char delim, SymbolTable *syms,
-                  std::vector<typename Arc::Label> *output) {
-  if (str == FLAGS_empty_symbol) return;
-  std::istringstream strm(str);
-  string buf;
-  while (std::getline(strm, buf, delim))
-    output->push_back(LookUp<Arc>(buf, syms));
-}
-
-// Like `std::replace_copy` but returns the number of modifications
-template <class InputIterator, class OutputIterator, class T>
-size_t ReplaceCopy(InputIterator first, InputIterator last,
-                   OutputIterator result, const T &old_value,
-                   const T &new_value) {
-  size_t changes = 0;
-  while (first != last) {
-    if (*first == old_value) {
-      *result = new_value;
-      ++changes;
-    } else {
-      *result = *first;
-    }
-    ++first;
-    ++result;
-  }
-  return changes;
-}
-
-template <class Arc>
-bool GetVocabRecord(const string &vocab, std::istream &strm,  // NOLINT
-                    SymbolTable *isyms, SymbolTable *fsyms, SymbolTable *osyms,
-                    typename Arc::Label *word,
-                    std::vector<typename Arc::Label> *feature_labels,
-                    std::vector<typename Arc::Label> *possible_labels,
-                    size_t *num_line);
-
-template <class Arc>
-bool GetModelRecord(const string &model, std::istream &strm,  // NOLINT
-                    SymbolTable *fsyms, SymbolTable *osyms,
-                    std::vector<typename Arc::Label> *input_labels,
-                    std::vector<typename Arc::Label> *output_labels,
-                    typename Arc::Weight *weight, size_t *num_line);
-
-// Reads in vocabulary file. Each line is in the following format
-//
-//   word <whitespace> features [ <whitespace> possible output ]
-//
-// where features and possible output are `FLAGS_delimiter`-delimited lists of
-// tokens
-template <class Arc>
-void AddVocab(const string &vocab, SymbolTable *isyms, SymbolTable *fsyms,
-              SymbolTable *osyms, LinearFstDataBuilder<Arc> *builder) {
-  std::ifstream in(vocab);
-  if (!in) LOG(FATAL) << "Can't open file: " << vocab;
-  size_t num_line = 0, num_added = 0;
-  std::vector<string> fields;
-  std::vector<typename Arc::Label> feature_labels, possible_labels;
-  typename Arc::Label word;
-  while (GetVocabRecord<Arc>(vocab, in, isyms, fsyms, osyms, &word,
-                             &feature_labels, &possible_labels, &num_line)) {
-    if (word == kNoLabel) {
-      LOG(WARNING) << "Ignored: boundary word: " << fields[0];
-      continue;
-    }
-    if (possible_labels.empty())
-      num_added += builder->AddWord(word, feature_labels);
-    else
-      num_added += builder->AddWord(word, feature_labels, possible_labels);
-  }
-  VLOG(1) << "Read " << num_added << " words in " << num_line << " lines from "
-          << vocab;
-}
-
-template <class Arc>
-void AddVocab(const string &vocab, SymbolTable *isyms, SymbolTable *fsyms,
-              SymbolTable *osyms,
-              LinearClassifierFstDataBuilder<Arc> *builder) {
-  std::ifstream in(vocab);
-  if (!in) LOG(FATAL) << "Can't open file: " << vocab;
-  size_t num_line = 0, num_added = 0;
-  std::vector<string> fields;
-  std::vector<typename Arc::Label> feature_labels, possible_labels;
-  typename Arc::Label word;
-  while (GetVocabRecord<Arc>(vocab, in, isyms, fsyms, osyms, &word,
-                             &feature_labels, &possible_labels, &num_line)) {
-    if (!possible_labels.empty())
-      LOG(FATAL)
-          << "Classifier vocabulary should not have possible output constraint";
-    if (word == kNoLabel) {
-      LOG(WARNING) << "Ignored: boundary word: " << fields[0];
-      continue;
-    }
-    num_added += builder->AddWord(word, feature_labels);
-  }
-  VLOG(1) << "Read " << num_added << " words in " << num_line << " lines from "
-          << vocab;
-}
-
-// Reads in model file. The first line is an integer designating the
-// size of future window in the input sequences. After this, each line
-// is in the following format
-//
-//   input sequence <whitespace> output sequence <whitespace> weight
-//
-// input sequence is a `FLAGS_delimiter`-delimited sequence of feature
-// labels (see `AddVocab()`) . output sequence is a
-// `FLAGS_delimiter`-delimited sequence of output labels where the
-// last label is the output of the feature position before the history
-// boundary.
-template <class Arc>
-void AddModel(const string &model, SymbolTable *fsyms, SymbolTable *osyms,
-              LinearFstDataBuilder<Arc> *builder) {
-  std::ifstream in(model);
-  if (!in) LOG(FATAL) << "Can't open file: " << model;
-  string line;
-  std::getline(in, line);
-  if (!in) LOG(FATAL) << "Empty file: " << model;
-  size_t future_size;
-  {
-    std::istringstream strm(line);
-    strm >> future_size;
-    if (!strm) LOG(FATAL) << "Can't read future size: " << model;
-  }
-  size_t num_line = 1, num_added = 0;
-  const int group = builder->AddGroup(future_size);
-  VLOG(1) << "Group " << group << ": from " << model << "; future size is "
-          << future_size << ".";
-  // Add the rest of lines as a single feature group
-  std::vector<string> fields;
-  std::vector<typename Arc::Label> input_labels, output_labels;
-  typename Arc::Weight weight;
-  while (GetModelRecord<Arc>(model, in, fsyms, osyms, &input_labels,
-                             &output_labels, &weight, &num_line)) {
-    if (output_labels.empty())
-      LOG(FATAL) << "Empty output sequence in source " << model << ", line "
-                 << num_line;
-
-    const typename Arc::Label marks[] = {LinearFstData<Arc>::kStartOfSentence,
-                                         LinearFstData<Arc>::kEndOfSentence};
-
-    std::vector<typename Arc::Label> copy_input(input_labels.size()),
-        copy_output(output_labels.size());
-    for (int i = 0; i < 2; ++i) {
-      for (int j = 0; j < 2; ++j) {
-        size_t num_input_changes =
-            ReplaceCopy(input_labels.begin(), input_labels.end(),
-                        copy_input.begin(), kNoLabel, marks[i]);
-        size_t num_output_changes =
-            ReplaceCopy(output_labels.begin(), output_labels.end(),
-                        copy_output.begin(), kNoLabel, marks[j]);
-        if ((num_input_changes > 0 || i == 0) &&
-            (num_output_changes > 0 || j == 0))
-          num_added +=
-              builder->AddWeight(group, copy_input, copy_output, weight);
-      }
-    }
-  }
-  VLOG(1) << "Group " << group << ": read " << num_added << " weight(s) in "
-          << num_line << " lines.";
-}
-
-template <class Arc>
-void AddModel(const string &model, SymbolTable *fsyms, SymbolTable *osyms,
-              LinearClassifierFstDataBuilder<Arc> *builder) {
-  std::ifstream in(model);
-  if (!in) LOG(FATAL) << "Can't open file: " << model;
-  string line;
-  std::getline(in, line);
-  if (!in) LOG(FATAL) << "Empty file: " << model;
-  size_t future_size;
-  {
-    std::istringstream strm(line);
-    strm >> future_size;
-    if (!strm) LOG(FATAL) << "Can't read future size: " << model;
-  }
-  if (future_size != 0)
-    LOG(FATAL) << "Classifier model must have future size = 0; got "
-               << future_size << " from " << model;
-  size_t num_line = 1, num_added = 0;
-  const int group = builder->AddGroup();
-  VLOG(1) << "Group " << group << ": from " << model << "; future size is "
-          << future_size << ".";
-  // Add the rest of lines as a single feature group
-  std::vector<string> fields;
-  std::vector<typename Arc::Label> input_labels, output_labels;
-  typename Arc::Weight weight;
-  while (GetModelRecord<Arc>(model, in, fsyms, osyms, &input_labels,
-                             &output_labels, &weight, &num_line)) {
-    if (output_labels.size() != 1)
-      LOG(FATAL) << "Output not a single label in source " << model << ", line "
-                 << num_line;
-
-    const typename Arc::Label marks[] = {LinearFstData<Arc>::kStartOfSentence,
-                                         LinearFstData<Arc>::kEndOfSentence};
-
-    typename Arc::Label pred = output_labels[0];
-
-    std::vector<typename Arc::Label> copy_input(input_labels.size());
-    for (int i = 0; i < 2; ++i) {
-      size_t num_input_changes =
-          ReplaceCopy(input_labels.begin(), input_labels.end(),
-                      copy_input.begin(), kNoLabel, marks[i]);
-      if (num_input_changes > 0 || i == 0)
-        num_added += builder->AddWeight(group, copy_input, pred, weight);
-    }
-  }
-  VLOG(1) << "Group " << group << ": read " << num_added << " weight(s) in "
-          << num_line << " lines.";
-}
-
-void SplitByWhitespace(const string &str, std::vector<string> *out);
-int ScanNumClasses(char **models, int models_length);
-
-template <class Arc>
-void LinearCompileTpl(LinearCompileArgs *args) {
-  const string &epsilon_symbol = std::get<0>(*args);
-  const string &unknown_symbol = std::get<1>(*args);
-  const string &vocab = std::get<2>(*args);
-  char **models = std::get<3>(*args);
-  const int models_length = std::get<4>(*args);
-  const string &out = std::get<5>(*args);
-  const string &save_isymbols = std::get<6>(*args);
-  const string &save_fsymbols = std::get<7>(*args);
-  const string &save_osymbols = std::get<8>(*args);
-
-  SymbolTable isyms,  // input (e.g. word tokens)
-      osyms,          // output (e.g. tags)
-      fsyms;          // feature (e.g. word identity, suffix, etc.)
-  isyms.AddSymbol(epsilon_symbol);
-  osyms.AddSymbol(epsilon_symbol);
-  fsyms.AddSymbol(epsilon_symbol);
-  isyms.AddSymbol(unknown_symbol);
-
-  VLOG(1) << "start-of-sentence label is "
-          << LinearFstData<Arc>::kStartOfSentence;
-  VLOG(1) << "end-of-sentence label is " << LinearFstData<Arc>::kEndOfSentence;
-
-  if (FLAGS_classifier) {
-    int num_classes = ScanNumClasses(models, models_length);
-    LinearClassifierFstDataBuilder<Arc> builder(num_classes, &isyms, &fsyms,
-                                                &osyms);
-
-    AddVocab(vocab, &isyms, &fsyms, &osyms, &builder);
-    for (int i = 0; i < models_length; ++i)
-      AddModel(models[i], &fsyms, &osyms, &builder);
-
-    LinearClassifierFst<Arc> fst(builder.Dump(), num_classes, &isyms, &osyms);
-    fst.Write(out);
-  } else {
-    LinearFstDataBuilder<Arc> builder(&isyms, &fsyms, &osyms);
-
-    AddVocab(vocab, &isyms, &fsyms, &osyms, &builder);
-    for (int i = 0; i < models_length; ++i)
-      AddModel(models[i], &fsyms, &osyms, &builder);
-
-    LinearTaggerFst<Arc> fst(builder.Dump(), &isyms, &osyms);
-    fst.Write(out);
-  }
-
-  if (!save_isymbols.empty()) isyms.WriteText(save_isymbols);
-  if (!save_fsymbols.empty()) fsyms.WriteText(save_fsymbols);
-  if (!save_osymbols.empty()) osyms.WriteText(save_osymbols);
-}
-
-void LinearCompile(const string &arc_type, const string &epsilon_symbol,
-                   const string &unknown_symbol, const string &vocab,
-                   char **models, int models_len, const string &out,
-                   const string &save_isymbols, const string &save_fsymbols,
-                   const string &save_osymbols);
-
-template <class Arc>
-bool GetVocabRecord(const string &vocab, std::istream &strm,  // NOLINT
-                    SymbolTable *isyms, SymbolTable *fsyms, SymbolTable *osyms,
-                    typename Arc::Label *word,
-                    std::vector<typename Arc::Label> *feature_labels,
-                    std::vector<typename Arc::Label> *possible_labels,
-                    size_t *num_line) {
-  string line;
-  if (!std::getline(strm, line)) return false;
-  ++(*num_line);
-
-  std::vector<string> fields;
-  SplitByWhitespace(line, &fields);
-  if (fields.size() != 3)
-    LOG(FATAL) << "Wrong number of fields in source " << vocab << ", line "
-               << num_line;
-
-  feature_labels->clear();
-  possible_labels->clear();
-
-  *word = LookUp<Arc>(fields[0], isyms);
-
-  const char delim = FLAGS_delimiter[0];
-  SplitAndPush<Arc>(fields[1], delim, fsyms, feature_labels);
-  SplitAndPush<Arc>(fields[2], delim, osyms, possible_labels);
-
-  return true;
-}
-
-template <class Arc>
-bool GetModelRecord(const string &model, std::istream &strm,  // NOLINT
-                    SymbolTable *fsyms, SymbolTable *osyms,
-                    std::vector<typename Arc::Label> *input_labels,
-                    std::vector<typename Arc::Label> *output_labels,
-                    typename Arc::Weight *weight, size_t *num_line) {
-  string line;
-  if (!std::getline(strm, line)) return false;
-  ++(*num_line);
-
-  std::vector<string> fields;
-  SplitByWhitespace(line, &fields);
-  if (fields.size() != 3)
-    LOG(FATAL) << "Wrong number of fields in source " << model << ", line "
-               << num_line;
-
-  input_labels->clear();
-  output_labels->clear();
-
-  const char delim = FLAGS_delimiter[0];
-  SplitAndPush<Arc>(fields[0], delim, fsyms, input_labels);
-  SplitAndPush<Arc>(fields[1], delim, osyms, output_labels);
-
-  *weight = StrToWeight<typename Arc::Weight>(fields[2], model, *num_line);
-
-  GuessStartOrEnd<Arc>(input_labels, kNoLabel);
-  GuessStartOrEnd<Arc>(output_labels, kNoLabel);
-
-  return true;
-}
-}  // namespace script
-}  // namespace fst
-
-#define REGISTER_FST_LINEAR_OPERATIONS(Arc) \
-  REGISTER_FST_OPERATION(LinearCompileTpl, Arc, LinearCompileArgs);
-
-#endif  // FST_EXTENSIONS_LINEAR_LINEARSCRIPT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/loglinear-apply.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/loglinear-apply.h
deleted file mode 100644
index 1b5d2eaf..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/loglinear-apply.h
+++ /dev/null
@@ -1,77 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_LINEAR_LOGLINEAR_APPLY_H_
-#define FST_EXTENSIONS_LINEAR_LOGLINEAR_APPLY_H_
-
-#include <fst/compat.h>
-#include <fst/arc.h>
-#include <fst/arc-map.h>
-#include <fst/compose.h>
-#include <fst/determinize.h>
-#include <fst/float-weight.h>
-#include <fst/fst.h>
-#include <fst/minimize.h>
-#include <fst/mutable-fst.h>
-#include <fst/project.h>
-#include <fst/rmepsilon.h>
-#include <fst/vector-fst.h>
-
-namespace fst {
-
-// Applies a FST model as a discriminative model to weighted input
-// `ifst`. `A` is an arc type with tropical weight of all the
-// input/output FSTs.
-//
-// In general, consider `ifst` an unnormalized probability
-// distribution between its input X and output Y, P(X, Y); and `lfst`
-// a group of unnormalized probability distributions of all its output
-// Z for every input Y, Q(Z|Y). `normalize` controls whether Q is
-// normalized for every Y before chaining with P(X, Y). I.e., for a
-// path (X, Y, Z) in `ofst` (where Y is hidden),
-//
-// - When `normalize` is true, its weight is P(X, Y) Q(Z|Y) / sum_z Q(z|Y);
-// - When `normalize` is false, its weight is P(X, Y) Q(Z|Y).
-template <class A>
-void LogLinearApply(const Fst<A> &ifst, const Fst<A> &lfst, MutableFst<A> *ofst,
-                    bool normalize = true) {
-  LogLinearApply<A, LogArc>(ifst, lfst, ofst, normalize);
-}
-
-// This version gives finer control over the arc type (`B`) to be used
-// in normalization. `B` is an arc type with log weight (e.g. `LogArc`
-// or `Log64Arc`).
-template <class A, class B>
-void LogLinearApply(const Fst<A> &ifst, const Fst<A> &lfst, MutableFst<A> *ofst,
-                    bool normalize = true) {
-  if (normalize) {
-    VectorFst<A> unnormalized_ofst, rescored_ifsa;
-    Compose(ifst, lfst, &unnormalized_ofst);
-    {
-      VectorFst<A> tropical_ifsa(unnormalized_ofst);
-      Project(&tropical_ifsa, PROJECT_INPUT);
-      {
-        VectorFst<B> minimal_log_ifsa;
-        {
-          VectorFst<B> log_ifsa;
-          ArcMap(tropical_ifsa, &log_ifsa, WeightConvertMapper<A, B>());
-          RmEpsilon(&log_ifsa);
-          Determinize(log_ifsa, &minimal_log_ifsa);
-        }
-        Minimize(&minimal_log_ifsa);
-        ArcMap(&minimal_log_ifsa, InvertWeightMapper<B>());
-        ArcMap(minimal_log_ifsa, &tropical_ifsa, WeightConvertMapper<B, A>());
-      }
-      ArcSort(&tropical_ifsa, OLabelCompare<A>());
-      Compose(tropical_ifsa, ifst, &rescored_ifsa);
-    }
-    ArcSort(&rescored_ifsa, OLabelCompare<A>());
-    Compose(rescored_ifsa, unnormalized_ofst, ofst);
-  } else {
-    Compose(ifst, lfst, ofst);
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_LINEAR_LOGLINEAR_APPLY_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/trie.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/trie.h
deleted file mode 100644
index b5ddb3ad..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/linear/trie.h
+++ /dev/null
@@ -1,444 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_LINEAR_TRIE_H_
-#define FST_EXTENSIONS_LINEAR_TRIE_H_
-
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/util.h>
-
-namespace fst {
-
-const int kNoTrieNodeId = -1;
-
-// Forward declarations of all available trie topologies.
-template <class L, class H>
-class NestedTrieTopology;
-template <class L, class H>
-class FlatTrieTopology;
-
-// A pair of parent node id and label, part of a trie edge
-template <class L>
-struct ParentLabel {
-  int parent;
-  L label;
-
-  ParentLabel() {}
-  ParentLabel(int p, L l) : parent(p), label(l) {}
-
-  bool operator==(const ParentLabel &that) const {
-    return parent == that.parent && label == that.label;
-  }
-
-  std::istream &Read(std::istream &strm) {  // NOLINT
-    ReadType(strm, &parent);
-    ReadType(strm, &label);
-    return strm;
-  }
-
-  std::ostream &Write(std::ostream &strm) const {  // NOLINT
-    WriteType(strm, parent);
-    WriteType(strm, label);
-    return strm;
-  }
-};
-
-template <class L, class H>
-struct ParentLabelHash {
-  size_t operator()(const ParentLabel<L> &pl) const {
-    return static_cast<size_t>(pl.parent * 7853 + H()(pl.label));
-  }
-};
-
-// The trie topology in a nested tree of hash maps; allows efficient
-// iteration over children of a specific node.
-template <class L, class H>
-class NestedTrieTopology {
- public:
-  typedef L Label;
-  typedef H Hash;
-  typedef std::unordered_map<L, int, H> NextMap;
-
-  class const_iterator {
-   public:
-    typedef std::forward_iterator_tag iterator_category;
-    typedef std::pair<ParentLabel<L>, int> value_type;
-    typedef std::ptrdiff_t difference_type;
-    typedef const value_type *pointer;
-    typedef const value_type &reference;
-
-    friend class NestedTrieTopology<L, H>;
-
-    const_iterator() : ptr_(nullptr), cur_node_(kNoTrieNodeId), cur_edge_() {}
-
-    reference operator*() {
-      UpdateStub();
-      return stub_;
-    }
-    pointer operator->() {
-      UpdateStub();
-      return &stub_;
-    }
-
-    const_iterator &operator++();
-    const_iterator &operator++(int);  // NOLINT
-
-    bool operator==(const const_iterator &that) const {
-      return ptr_ == that.ptr_ && cur_node_ == that.cur_node_ &&
-             cur_edge_ == that.cur_edge_;
-    }
-    bool operator!=(const const_iterator &that) const {
-      return !(*this == that);
-    }
-
-   private:
-    const_iterator(const NestedTrieTopology *ptr, int cur_node)
-        : ptr_(ptr), cur_node_(cur_node) {
-      SetProperCurEdge();
-    }
-
-    void SetProperCurEdge() {
-      if (cur_node_ < ptr_->NumNodes())
-        cur_edge_ = ptr_->nodes_[cur_node_]->begin();
-      else
-        cur_edge_ = ptr_->nodes_[0]->begin();
-    }
-
-    void UpdateStub() {
-      stub_.first = ParentLabel<L>(cur_node_, cur_edge_->first);
-      stub_.second = cur_edge_->second;
-    }
-
-    const NestedTrieTopology *ptr_;
-    int cur_node_;
-    typename NextMap::const_iterator cur_edge_;
-    value_type stub_;
-  };
-
-  NestedTrieTopology();
-  NestedTrieTopology(const NestedTrieTopology &that);
-  ~NestedTrieTopology();
-  void swap(NestedTrieTopology &that);
-  NestedTrieTopology &operator=(const NestedTrieTopology &that);
-  bool operator==(const NestedTrieTopology &that) const;
-  bool operator!=(const NestedTrieTopology &that) const;
-
-  int Root() const { return 0; }
-  size_t NumNodes() const { return nodes_.size(); }
-  int Insert(int parent, const L &label);
-  int Find(int parent, const L &label) const;
-  const NextMap &ChildrenOf(int parent) const { return *nodes_[parent]; }
-
-  std::istream &Read(std::istream &strm);         // NOLINT
-  std::ostream &Write(std::ostream &strm) const;  // NOLINT
-
-  const_iterator begin() const { return const_iterator(this, 0); }
-  const_iterator end() const { return const_iterator(this, NumNodes()); }
-
- private:
-  std::vector<NextMap *>
-      nodes_;  // Use pointers to avoid copying the maps when the
-               // vector grows
-};
-
-template <class L, class H>
-NestedTrieTopology<L, H>::NestedTrieTopology() {
-  nodes_.push_back(new NextMap);
-}
-
-template <class L, class H>
-NestedTrieTopology<L, H>::NestedTrieTopology(const NestedTrieTopology &that) {
-  nodes_.reserve(that.nodes_.size());
-  for (size_t i = 0; i < that.nodes_.size(); ++i) {
-    NextMap *node = that.nodes_[i];
-    nodes_.push_back(new NextMap(*node));
-  }
-}
-
-template <class L, class H>
-NestedTrieTopology<L, H>::~NestedTrieTopology() {
-  for (size_t i = 0; i < nodes_.size(); ++i) {
-    NextMap *node = nodes_[i];
-    delete node;
-  }
-}
-
-// TODO(wuke): std::swap compatibility
-template <class L, class H>
-inline void NestedTrieTopology<L, H>::swap(NestedTrieTopology &that) {
-  nodes_.swap(that.nodes_);
-}
-
-template <class L, class H>
-inline NestedTrieTopology<L, H> &NestedTrieTopology<L, H>::operator=(
-    const NestedTrieTopology &that) {
-  NestedTrieTopology copy(that);
-  swap(copy);
-  return *this;
-}
-
-template <class L, class H>
-inline bool NestedTrieTopology<L, H>::operator==(
-    const NestedTrieTopology &that) const {
-  if (NumNodes() != that.NumNodes()) return false;
-  for (int i = 0; i < NumNodes(); ++i)
-    if (ChildrenOf(i) != that.ChildrenOf(i)) return false;
-  return true;
-}
-
-template <class L, class H>
-inline bool NestedTrieTopology<L, H>::operator!=(
-    const NestedTrieTopology &that) const {
-  return !(*this == that);
-}
-
-template <class L, class H>
-inline int NestedTrieTopology<L, H>::Insert(int parent, const L &label) {
-  int ret = Find(parent, label);
-  if (ret == kNoTrieNodeId) {
-    ret = NumNodes();
-    (*nodes_[parent])[label] = ret;
-    nodes_.push_back(new NextMap);
-  }
-  return ret;
-}
-
-template <class L, class H>
-inline int NestedTrieTopology<L, H>::Find(int parent, const L &label) const {
-  typename NextMap::const_iterator it = nodes_[parent]->find(label);
-  return it == nodes_[parent]->end() ? kNoTrieNodeId : it->second;
-}
-
-template <class L, class H>
-inline std::istream &NestedTrieTopology<L, H>::Read(
-    std::istream &strm) {  // NOLINT
-  NestedTrieTopology new_trie;
-  size_t num_nodes;
-  if (!ReadType(strm, &num_nodes)) return strm;
-  for (size_t i = 1; i < num_nodes; ++i) new_trie.nodes_.push_back(new NextMap);
-  for (size_t i = 0; i < num_nodes; ++i) ReadType(strm, new_trie.nodes_[i]);
-  if (strm) swap(new_trie);
-  return strm;
-}
-
-template <class L, class H>
-inline std::ostream &NestedTrieTopology<L, H>::Write(
-    std::ostream &strm) const {  // NOLINT
-  WriteType(strm, NumNodes());
-  for (size_t i = 0; i < NumNodes(); ++i) WriteType(strm, *nodes_[i]);
-  return strm;
-}
-
-template <class L, class H>
-inline typename NestedTrieTopology<L, H>::const_iterator
-    &NestedTrieTopology<L, H>::const_iterator::operator++() {
-  ++cur_edge_;
-  if (cur_edge_ == ptr_->nodes_[cur_node_]->end()) {
-    ++cur_node_;
-    while (cur_node_ < ptr_->NumNodes() && ptr_->nodes_[cur_node_]->empty())
-      ++cur_node_;
-    SetProperCurEdge();
-  }
-  return *this;
-}
-
-template <class L, class H>
-inline typename NestedTrieTopology<L, H>::const_iterator
-    &NestedTrieTopology<L, H>::const_iterator::operator++(int) {  // NOLINT
-  const_iterator save(*this);
-  ++(*this);
-  return save;
-}
-
-// The trie topology in a single hash map; only allows iteration over
-// all the edges in arbitrary order.
-template <class L, class H>
-class FlatTrieTopology {
- private:
-  typedef std::unordered_map<ParentLabel<L>, int, ParentLabelHash<L, H>>
-      NextMap;
-
- public:
-  // Iterator over edges as std::pair<ParentLabel<L>, int>
-  typedef typename NextMap::const_iterator const_iterator;
-  typedef L Label;
-  typedef H Hash;
-
-  FlatTrieTopology() {}
-  FlatTrieTopology(const FlatTrieTopology &that) : next_(that.next_) {}
-  template <class T>
-  explicit FlatTrieTopology(const T &that);
-
-  // TODO(wuke): std::swap compatibility
-  void swap(FlatTrieTopology &that) { next_.swap(that.next_); }
-
-  bool operator==(const FlatTrieTopology &that) const {
-    return next_ == that.next_;
-  }
-  bool operator!=(const FlatTrieTopology &that) const {
-    return !(*this == that);
-  }
-
-  int Root() const { return 0; }
-  size_t NumNodes() const { return next_.size() + 1; }
-  int Insert(int parent, const L &label);
-  int Find(int parent, const L &label) const;
-
-  std::istream &Read(std::istream &strm) {  // NOLINT
-    return ReadType(strm, &next_);
-  }
-  std::ostream &Write(std::ostream &strm) const {  // NOLINT
-    return WriteType(strm, next_);
-  }
-
-  const_iterator begin() const { return next_.begin(); }
-  const_iterator end() const { return next_.end(); }
-
- private:
-  NextMap next_;
-};
-
-template <class L, class H>
-template <class T>
-FlatTrieTopology<L, H>::FlatTrieTopology(const T &that)
-    : next_(that.begin(), that.end()) {}
-
-template <class L, class H>
-inline int FlatTrieTopology<L, H>::Insert(int parent, const L &label) {
-  int ret = Find(parent, label);
-  if (ret == kNoTrieNodeId) {
-    ret = NumNodes();
-    next_[ParentLabel<L>(parent, label)] = ret;
-  }
-  return ret;
-}
-
-template <class L, class H>
-inline int FlatTrieTopology<L, H>::Find(int parent, const L &label) const {
-  typename NextMap::const_iterator it =
-      next_.find(ParentLabel<L>(parent, label));
-  return it == next_.end() ? kNoTrieNodeId : it->second;
-}
-
-// A collection of implementations of the trie data structure. The key
-// is a sequence of type `L` which must be hashable. The value is of
-// `V` which must be default constructible and copyable. In addition,
-// a value object is stored for each node in the trie therefore
-// copying `V` should be cheap.
-//
-// One can access the store values with an integer node id, using the
-// [] operator. A valid node id can be obtained by the following ways:
-//
-// 1. Using the `Root()` method to get the node id of the root.
-//
-// 2. Iterating through 0 to `NumNodes() - 1`. The node ids are dense
-// so every integer in this range is a valid node id.
-//
-// 3. Using the node id returned from a successful `Insert()` or
-// `Find()` call.
-//
-// 4. Iterating over the trie edges with an `EdgeIterator` and using
-// the node ids returned from its `Parent()` and `Child()` methods.
-//
-// Below is an example of inserting keys into the trie:
-//
-//   const string words[] = {"hello", "health", "jello"};
-//   Trie<char, bool> dict;
-//   for (auto word : words) {
-//     int cur = dict.Root();
-//     for (char c : word) {
-//       cur = dict.Insert(cur, c);
-//     }
-//     dict[cur] = true;
-//   }
-//
-// And the following is an example of looking up the longest prefix of
-// a string using the trie constructed above:
-//
-//   string query = "healed";
-//   size_t prefix_length = 0;
-//   int cur = dict.Find(dict.Root(), query[prefix_length]);
-//   while (prefix_length < query.size() &&
-//     cur != Trie<char, bool>::kNoNodeId) {
-//     ++prefix_length;
-//     cur = dict.Find(cur, query[prefix_length]);
-//   }
-template <class L, class V, class T>
-class MutableTrie {
- public:
-  template <class LL, class VV, class TT>
-  friend class MutableTrie;
-
-  typedef L Label;
-  typedef V Value;
-  typedef T Topology;
-
-  // Constructs a trie with only the root node.
-  MutableTrie() {}
-
-  // Conversion from another trie of a possiblly different
-  // topology. The underlying topology must supported conversion.
-  template <class S>
-  explicit MutableTrie(const MutableTrie<L, V, S> &that)
-      : topology_(that.topology_), values_(that.values_) {}
-
-  // TODO(wuke): std::swap compatibility
-  void swap(MutableTrie &that) {
-    topology_.swap(that.topology_);
-    values_.swap(that.values_);
-  }
-
-  int Root() const { return topology_.Root(); }
-  size_t NumNodes() const { return topology_.NumNodes(); }
-
-  // Inserts an edge with given `label` at node `parent`. Returns the
-  // child node id. If the node already exists, returns the node id
-  // right away.
-  int Insert(int parent, const L &label) {
-    int ret = topology_.Insert(parent, label);
-    values_.resize(NumNodes());
-    return ret;
-  }
-
-  // Finds the node id of the node from `parent` via `label`. Returns
-  // `kNoTrieNodeId` when such a node does not exist.
-  int Find(int parent, const L &label) const {
-    return topology_.Find(parent, label);
-  }
-
-  const T &TrieTopology() const { return topology_; }
-
-  // Accesses the value stored for the given node.
-  V &operator[](int node_id) { return values_[node_id]; }
-  const V &operator[](int node_id) const { return values_[node_id]; }
-
-  // Comparison by content
-  bool operator==(const MutableTrie &that) const {
-    return topology_ == that.topology_ && values_ == that.values_;
-  }
-
-  bool operator!=(const MutableTrie &that) const { return !(*this == that); }
-
-  std::istream &Read(std::istream &strm) {  // NOLINT
-    ReadType(strm, &topology_);
-    ReadType(strm, &values_);
-    return strm;
-  }
-  std::ostream &Write(std::ostream &strm) const {  // NOLINT
-    WriteType(strm, topology_);
-    WriteType(strm, values_);
-    return strm;
-  }
-
- private:
-  T topology_;
-  std::vector<V> values_;
-};
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_LINEAR_TRIE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/compose.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/compose.h
deleted file mode 100644
index 47714e37..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/compose.h
+++ /dev/null
@@ -1,267 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Compose an MPDT and an FST.
-
-#ifndef FST_EXTENSIONS_MPDT_COMPOSE_H_
-#define FST_EXTENSIONS_MPDT_COMPOSE_H_
-
-#include <list>
-
-#include <fst/extensions/mpdt/mpdt.h>
-#include <fst/extensions/pdt/compose.h>
-#include <fst/compose.h>
-
-namespace fst {
-
-template <class Filter>
-class MPdtParenFilter {
- public:
-  using FST1 = typename Filter::FST1;
-  using FST2 = typename Filter::FST2;
-  using Arc = typename Filter::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Matcher1 = typename Filter::Matcher1;
-  using Matcher2 = typename Filter::Matcher2;
-
-  using StackId = StateId;
-  using ParenStack = internal::MPdtStack<StackId, Label>;
-  using FilterState1 = typename Filter::FilterState;
-  using FilterState2 = IntegerFilterState<StackId>;
-  using FilterState = PairFilterState<FilterState1, FilterState2>;
-
-  MPdtParenFilter(const FST1 &fst1, const FST2 &fst2,
-                  Matcher1 *matcher1 = nullptr, Matcher2 *matcher2 = nullptr,
-                  const std::vector<std::pair<Label, Label>> *parens = nullptr,
-                  const std::vector<Label> *assignments = nullptr,
-                  bool expand = false, bool keep_parens = true)
-      : filter_(fst1, fst2, matcher1, matcher2),
-        parens_(parens ? *parens : std::vector<std::pair<Label, Label>>()),
-        assignments_(assignments ? *assignments : std::vector<Label>()),
-        expand_(expand),
-        keep_parens_(keep_parens),
-        fs_(FilterState::NoState()),
-        stack_(parens_, assignments_),
-        paren_id_(-1) {
-    if (parens) {
-      for (const auto &pair : *parens) {
-        parens_.push_back(pair);
-        GetMatcher1()->AddOpenParen(pair.first);
-        GetMatcher2()->AddOpenParen(pair.first);
-        if (!expand_) {
-          GetMatcher1()->AddCloseParen(pair.second);
-          GetMatcher2()->AddCloseParen(pair.second);
-        }
-      }
-    }
-  }
-
-  MPdtParenFilter(const MPdtParenFilter &filter, bool safe = false)
-      : filter_(filter.filter_, safe),
-        parens_(filter.parens_),
-        expand_(filter.expand_),
-        keep_parens_(filter.keep_parens_),
-        fs_(FilterState::NoState()),
-        stack_(filter.parens_, filter.assignments_),
-        paren_id_(-1) {}
-
-  FilterState Start() const {
-    return FilterState(filter_.Start(), FilterState2(0));
-  }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    fs_ = fs;
-    filter_.SetState(s1, s2, fs_.GetState1());
-    if (!expand_) return;
-    const auto paren_id = stack_.Top(fs.GetState2().GetState());
-    if (paren_id != paren_id_) {
-      if (paren_id_ != -1) {
-        GetMatcher1()->RemoveCloseParen(parens_[paren_id_].second);
-        GetMatcher2()->RemoveCloseParen(parens_[paren_id_].second);
-      }
-      paren_id_ = paren_id;
-      if (paren_id_ != -1) {
-        GetMatcher1()->AddCloseParen(parens_[paren_id_].second);
-        GetMatcher2()->AddCloseParen(parens_[paren_id_].second);
-      }
-    }
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    const auto fs1 = filter_.FilterArc(arc1, arc2);
-    const auto &fs2 = fs_.GetState2();
-    if (fs1 == FilterState1::NoState()) return FilterState::NoState();
-    if (arc1->olabel == kNoLabel && arc2->ilabel) {  // arc2 parentheses.
-      if (keep_parens_) {
-        arc1->ilabel = arc2->ilabel;
-      } else if (arc2->ilabel) {
-        arc2->olabel = arc1->ilabel;
-      }
-      return FilterParen(arc2->ilabel, fs1, fs2);
-    } else if (arc2->ilabel == kNoLabel && arc1->olabel) {  // arc1 parentheses
-      if (keep_parens_) {
-        arc2->olabel = arc1->olabel;
-      } else {
-        arc1->ilabel = arc2->olabel;
-      }
-      return FilterParen(arc1->olabel, fs1, fs2);
-    } else {
-      return FilterState(fs1, fs2);
-    }
-  }
-
-  void FilterFinal(Weight *w1, Weight *w2) const {
-    if (fs_.GetState2().GetState() != 0) *w1 = Weight::Zero();
-    filter_.FilterFinal(w1, w2);
-  }
-
-  // Returns respective matchers; ownership stays with filter.
-
-  Matcher1 *GetMatcher1() { return filter_.GetMatcher1(); }
-
-  Matcher2 *GetMatcher2() { return filter_.GetMatcher2(); }
-
-  uint64 Properties(uint64 iprops) const {
-    const auto oprops = filter_.Properties(iprops);
-    return oprops & kILabelInvariantProperties & kOLabelInvariantProperties;
-  }
-
- private:
-  const FilterState FilterParen(Label label, const FilterState1 &fs1,
-                                const FilterState2 &fs2) const {
-    if (!expand_) return FilterState(fs1, fs2);
-    const auto stack_id = stack_.Find(fs2.GetState(), label);
-    if (stack_id < 0) {
-      return FilterState::NoState();
-    } else {
-      return FilterState(fs1, FilterState2(stack_id));
-    }
-  }
-
-  Filter filter_;
-  std::vector<std::pair<Label, Label>> parens_;
-  std::vector<Label> assignments_;
-  bool expand_;       // Expands to FST?
-  bool keep_parens_;  // Retains parentheses in output?
-  FilterState fs_;    // Current filter state.
-  mutable ParenStack stack_;
-  ssize_t paren_id_;
-};
-
-// Class to setup composition options for MPDT composition. Default is to take
-// the MPDT as the first composition argument.
-template <class Arc, bool left_pdt = true>
-class MPdtComposeFstOptions
-    : public ComposeFstOptions<Arc, ParenMatcher<Fst<Arc>>,
-                               MPdtParenFilter<AltSequenceComposeFilter<
-                                   ParenMatcher<Fst<Arc>> >> > {
- public:
-  using Label = typename Arc::Label;
-  using MPdtMatcher = ParenMatcher<Fst<Arc>>;
-  using MPdtFilter = MPdtParenFilter<AltSequenceComposeFilter<MPdtMatcher>>;
-
-  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::matcher1;
-  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::matcher2;
-  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::filter;
-
-  MPdtComposeFstOptions(const Fst<Arc> &ifst1,
-                        const std::vector<std::pair<Label, Label>> &parens,
-                        const std::vector<typename Arc::Label> &assignments,
-                        const Fst<Arc> &ifst2, bool expand = false,
-                        bool keep_parens = true) {
-    matcher1 = new MPdtMatcher(ifst1, MATCH_OUTPUT, kParenList);
-    matcher2 = new MPdtMatcher(ifst2, MATCH_INPUT, kParenLoop);
-    filter = new MPdtFilter(ifst1, ifst2, matcher1, matcher2, &parens,
-                            &assignments, expand, keep_parens);
-  }
-};
-
-// Class to setup composition options for PDT with FST composition.
-// Specialization is for the FST as the first composition argument.
-template <class Arc>
-class MPdtComposeFstOptions<Arc, false>
-    : public ComposeFstOptions<
-          Arc, ParenMatcher<Fst<Arc>>,
-          MPdtParenFilter<SequenceComposeFilter<ParenMatcher<Fst<Arc>> >> > {
- public:
-  using Label = typename Arc::Label;
-  using MPdtMatcher = ParenMatcher<Fst<Arc>>;
-  using MPdtFilter = MPdtParenFilter<SequenceComposeFilter<MPdtMatcher>>;
-
-  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::matcher1;
-  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::matcher2;
-  using ComposeFstOptions<Arc, MPdtMatcher, MPdtFilter>::filter;
-
-  MPdtComposeFstOptions(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-                        const std::vector<std::pair<Label, Label>> &parens,
-                        const std::vector<typename Arc::Label> &assignments,
-                        bool expand = false, bool keep_parens = true) {
-    matcher1 = new MPdtMatcher(ifst1, MATCH_OUTPUT, kParenLoop);
-    matcher2 = new MPdtMatcher(ifst2, MATCH_INPUT, kParenList);
-    filter = new MPdtFilter(ifst1, ifst2, matcher1, matcher2, &parens,
-                            &assignments, expand, keep_parens);
-  }
-};
-
-struct MPdtComposeOptions {
-  bool connect;                  // Connect output?
-  PdtComposeFilter filter_type;  // Which pre-defined filter to use.
-
-  explicit MPdtComposeOptions(bool connect = true,
-                              PdtComposeFilter filter_type = PAREN_FILTER)
-      : connect(connect), filter_type(filter_type) {}
-};
-
-// Composes multi-pushdown transducer (MPDT) encoded as an FST (1st arg) and an
-// FST (2nd arg) with the result also an MPDT encoded as an FST (3rd arg). In
-// theMPDTs, some transitions are labeled with open or close parentheses (and
-// associated with a stack). To be interpreted as an MPDT, the parents on each
-// stack must balance on a path (see MPdtExpand()). The open-close parenthesis
-// label pairs are passed using the parens arguments, and the stack assignments
-// are passed using the assignments argument.
-template <class Arc>
-void Compose(
-    const Fst<Arc> &ifst1,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens,
-    const std::vector<typename Arc::Label> &assignments, const Fst<Arc> &ifst2,
-    MutableFst<Arc> *ofst,
-    const MPdtComposeOptions &opts = MPdtComposeOptions()) {
-  bool expand = opts.filter_type != PAREN_FILTER;
-  bool keep_parens = opts.filter_type != EXPAND_FILTER;
-  MPdtComposeFstOptions<Arc, true> copts(ifst1, parens, assignments, ifst2,
-                                         expand, keep_parens);
-  copts.gc_limit = 0;
-  *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-  if (opts.connect) Connect(ofst);
-}
-
-// Composes an FST (1st arg) and a multi-pushdown transducer (MPDT) encoded as
-// an FST (2nd arg) with the result also an MPDT encoded as an FST (3rd arg).
-// In the MPDTs, some transitions are labeled with open or close parentheses
-// (and associated with a stack). To be interpreted as an MPDT, the parents on
-// each stack must balance on a path (see MPdtExpand()). The open-close
-// parenthesis label pairs are passed using the parens arguments, and the stack
-// assignments are passed using the assignments argument.
-template <class Arc>
-void Compose(
-    const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens,
-    const std::vector<typename Arc::Label> &assignments, MutableFst<Arc> *ofst,
-    const MPdtComposeOptions &opts = MPdtComposeOptions()) {
-  bool expand = opts.filter_type != PAREN_FILTER;
-  bool keep_parens = opts.filter_type != EXPAND_FILTER;
-  MPdtComposeFstOptions<Arc, false> copts(ifst1, ifst2, parens, assignments,
-                                          expand, keep_parens);
-  copts.gc_limit = 0;
-  *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-  if (opts.connect) Connect(ofst);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_MPDT_COMPOSE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/expand.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/expand.h
deleted file mode 100644
index 0e06cdce..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/expand.h
+++ /dev/null
@@ -1,335 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Expands an MPDT to an FST.
-
-#ifndef FST_EXTENSIONS_MPDT_EXPAND_H_
-#define FST_EXTENSIONS_MPDT_EXPAND_H_
-
-#include <vector>
-
-#include <fst/extensions/mpdt/mpdt.h>
-#include <fst/extensions/pdt/paren.h>
-#include <fst/cache.h>
-#include <fst/mutable-fst.h>
-#include <fst/queue.h>
-#include <fst/state-table.h>
-#include <fst/test-properties.h>
-
-namespace fst {
-
-template <class Arc>
-struct MPdtExpandFstOptions : public CacheOptions {
-  bool keep_parentheses;
-  internal::MPdtStack<typename Arc::StateId, typename Arc::Label> *stack;
-  PdtStateTable<typename Arc::StateId, typename Arc::StateId> *state_table;
-
-  MPdtExpandFstOptions(
-      const CacheOptions &opts = CacheOptions(), bool kp = false,
-      internal::MPdtStack<typename Arc::StateId, typename Arc::Label> *s =
-          nullptr,
-      PdtStateTable<typename Arc::StateId, typename Arc::StateId> *st = nullptr)
-      : CacheOptions(opts), keep_parentheses(kp), stack(s), state_table(st) {}
-};
-
-// Properties for an expanded PDT.
-inline uint64 MPdtExpandProperties(uint64 inprops) {
-  return inprops & (kAcceptor | kAcyclic | kInitialAcyclic | kUnweighted);
-}
-
-namespace internal {
-
-// Implementation class for ExpandFst
-template <class Arc>
-class MPdtExpandFstImpl : public CacheImpl<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using StackId = StateId;
-  using StateTuple = PdtStateTuple<StateId, StackId>;
-  using ParenStack = internal::MPdtStack<StateId, Label>;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::Properties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheBaseImpl<CacheState<Arc>>::PushArc;
-  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
-  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
-  using CacheBaseImpl<CacheState<Arc>>::HasStart;
-  using CacheBaseImpl<CacheState<Arc>>::SetArcs;
-  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
-  using CacheBaseImpl<CacheState<Arc>>::SetStart;
-
-  MPdtExpandFstImpl(const Fst<Arc> &fst,
-                    const std::vector<std::pair<Label, Label>> &parens,
-                    const std::vector<Label> &assignments,
-                    const MPdtExpandFstOptions<Arc> &opts)
-      : CacheImpl<Arc>(opts),
-        fst_(fst.Copy()),
-        stack_(opts.stack ? opts.stack : new ParenStack(parens, assignments)),
-        state_table_(opts.state_table ? opts.state_table
-                                      : new PdtStateTable<StateId, StackId>()),
-        own_stack_(!opts.stack),
-        own_state_table_(!opts.state_table),
-        keep_parentheses_(opts.keep_parentheses) {
-    SetType("expand");
-    const auto props = fst.Properties(kFstProperties, false);
-    SetProperties(MPdtExpandProperties(props), kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  MPdtExpandFstImpl(const MPdtExpandFstImpl &impl)
-      : CacheImpl<Arc>(impl),
-        fst_(impl.fst_->Copy(true)),
-        stack_(new ParenStack(*impl.stack_)),
-        state_table_(new PdtStateTable<StateId, StackId>()),
-        own_stack_(true),
-        own_state_table_(true),
-        keep_parentheses_(impl.keep_parentheses_) {
-    SetType("expand");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  ~MPdtExpandFstImpl() override {
-    if (own_stack_) delete stack_;
-    if (own_state_table_) delete state_table_;
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      const auto s = fst_->Start();
-      if (s == kNoStateId) return kNoStateId;
-      const StateTuple tuple(s, 0);
-      const auto start = state_table_->FindState(tuple);
-      SetStart(start);
-    }
-    return CacheImpl<Arc>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) {
-      const auto &tuple = state_table_->Tuple(s);
-      const auto weight = fst_->Final(tuple.state_id);
-      SetFinal(s,
-               (weight != Weight::Zero() && tuple.stack_id == 0)
-                   ? weight
-                   : Weight::Zero());
-    }
-    return CacheImpl<Arc>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) ExpandState(s);
-    return CacheImpl<Arc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) ExpandState(s);
-    return CacheImpl<Arc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) ExpandState(s);
-    return CacheImpl<Arc>::NumOutputEpsilons(s);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) ExpandState(s);
-    CacheImpl<Arc>::InitArcIterator(s, data);
-  }
-
-  // Computes the outgoing transitions from a state, creating new destination
-  // states as needed.
-  void ExpandState(StateId s) {
-    const auto tuple = state_table_->Tuple(s);
-    for (ArcIterator<Fst<Arc>> aiter(*fst_, tuple.state_id); !aiter.Done();
-         aiter.Next()) {
-      auto arc = aiter.Value();
-      const auto stack_id = stack_->Find(tuple.stack_id, arc.ilabel);
-      if (stack_id == -1) {
-        continue;  // Non-matching close parenthesis.
-      } else if ((stack_id != tuple.stack_id) && !keep_parentheses_) {
-        arc.ilabel = arc.olabel = 0;  // Stack push/pop.
-      }
-      const StateTuple ntuple(arc.nextstate, stack_id);
-      arc.nextstate = state_table_->FindState(ntuple);
-      PushArc(s, arc);
-    }
-    SetArcs(s);
-  }
-
-  const ParenStack &GetStack() const { return *stack_; }
-
-  const PdtStateTable<StateId, StackId> &GetStateTable() const {
-    return *state_table_;
-  }
-
- private:
-  std::unique_ptr<const Fst<Arc>> fst_;
-  ParenStack *stack_;
-  PdtStateTable<StateId, StackId> *state_table_;
-  const bool own_stack_;
-  const bool own_state_table_;
-  const bool keep_parentheses_;
-
-  MPdtExpandFstImpl &operator=(const MPdtExpandFstImpl &) = delete;
-};
-
-}  // namespace internal
-
-// Expands a multi-pushdown transducer (MPDT) encoded as an FST into an FST.
-// This version is a delayed FST. In the MPDT, some transitions are labeled with
-// open or close parentheses. To be interpreted as an MPDT, the parens for each
-// stack must balance on a path. The open-close parenthesis label
-// pairs are passed using the parens argument, and the assignment of those pairs
-// to stacks is passed using the assignments argument. Expansion enforces the
-// parenthesis constraints. The MPDT must be
-// expandable as an FST.
-//
-// This class attaches interface to implementation and handles
-// reference counting, delegating most methods to ImplToFst.
-template <class A>
-class MPdtExpandFst : public ImplToFst<internal::MPdtExpandFstImpl<A>> {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using StackId = StateId;
-  using ParenStack = internal::MPdtStack<StackId, Label>;
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::MPdtExpandFstImpl<Arc>;
-
-  friend class ArcIterator<MPdtExpandFst<Arc>>;
-  friend class StateIterator<MPdtExpandFst<Arc>>;
-
-  MPdtExpandFst(const Fst<Arc> &fst,
-                const std::vector<std::pair<Label, Label>> &parens,
-                const std::vector<Label> &assignments)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, parens, assignments,
-                                               MPdtExpandFstOptions<Arc>())) {}
-
-  MPdtExpandFst(const Fst<Arc> &fst,
-                const std::vector<std::pair<Label, Label>> &parens,
-                const std::vector<Label> &assignments,
-                const MPdtExpandFstOptions<Arc> &opts)
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, parens, assignments, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  MPdtExpandFst(const MPdtExpandFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this ExpandFst. See Fst<>::Copy() for further doc.
-  MPdtExpandFst<Arc> *Copy(bool safe = false) const override {
-    return new MPdtExpandFst<A>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
-  const ParenStack &GetStack() const { return GetImpl()->GetStack(); }
-
-  const PdtStateTable<StateId, StackId> &GetStateTable() const {
-    return GetImpl()->GetStateTable();
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  void operator=(const MPdtExpandFst &) = delete;
-};
-
-// Specialization for MPdtExpandFst.
-template <class Arc>
-class StateIterator<MPdtExpandFst<Arc>>
-    : public CacheStateIterator<MPdtExpandFst<Arc>> {
- public:
-  explicit StateIterator(const MPdtExpandFst<Arc> &fst)
-      : CacheStateIterator<MPdtExpandFst<Arc>>(fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for MPdtExpandFst.
-template <class Arc>
-class ArcIterator<MPdtExpandFst<Arc>>
-    : public CacheArcIterator<MPdtExpandFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const MPdtExpandFst<Arc> &fst, StateId s)
-      : CacheArcIterator<MPdtExpandFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->ExpandState(s);
-  }
-};
-
-template <class Arc>
-inline void MPdtExpandFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<MPdtExpandFst<Arc>>(*this);
-}
-
-struct MPdtExpandOptions {
-  bool connect;
-  bool keep_parentheses;
-
-  explicit MPdtExpandOptions(bool connect = true, bool keep_parentheses = false)
-      : connect(connect), keep_parentheses(keep_parentheses) {}
-};
-
-// Expands a multi-pushdown transducer (MPDT) encoded as an FST into an FST.
-// This version writes the expanded PDT to a mutable FST. In the MPDT, some
-// transitions are labeled with open or close parentheses. To be interpreted as
-// an MPDT, the parens for each stack must balance on a path. The open-close
-// parenthesis label pair sets are passed using the parens argument, and the
-// assignment of those pairs to stacks is passed using the assignments argument.
-// The expansion enforces the parenthesis constraints. The MPDT must be
-// expandable as an FST.
-template <class Arc>
-void Expand(const Fst<Arc> &ifst,
-            const std::vector<
-            std::pair<typename Arc::Label, typename Arc::Label>> &parens,
-            const std::vector<typename Arc::Label> &assignments,
-            MutableFst<Arc> *ofst, const MPdtExpandOptions &opts) {
-  MPdtExpandFstOptions<Arc> eopts;
-  eopts.gc_limit = 0;
-  eopts.keep_parentheses = opts.keep_parentheses;
-  *ofst = MPdtExpandFst<Arc>(ifst, parens, assignments, eopts);
-  if (opts.connect) Connect(ofst);
-}
-
-// Expands a multi-pushdown transducer (MPDT) encoded as an FST into an FST.
-// This version writes the expanded PDT to a mutable FST. In the MPDT, some
-// transitions are labeled with open or close parentheses. To be interpreted as
-// an MPDT, the parens for each stack must balance on a path. The open-close
-// parenthesis label pair sets are passed using the parens argument, and the
-// assignment of those pairs to stacks is passed using the assignments argument.
-// The expansion enforces the parenthesis constraints. The MPDT must be
-// expandable as an FST.
-template <class Arc>
-void Expand(const Fst<Arc> &ifst,
-            const std::vector<std::pair<typename Arc::Label,
-            typename Arc::Label>> &parens,
-            const std::vector<typename Arc::Label> &assignments,
-            MutableFst<Arc> *ofst, bool connect = true,
-            bool keep_parentheses = false) {
-  const MPdtExpandOptions opts(connect, keep_parentheses);
-  Expand(ifst, parens, assignments, ofst, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_MPDT_EXPAND_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/info.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/info.h
deleted file mode 100644
index 512fcfa8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/info.h
+++ /dev/null
@@ -1,190 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Prints information about an MPDT.
-
-#ifndef FST_EXTENSIONS_MPDT_INFO_H_
-#define FST_EXTENSIONS_MPDT_INFO_H_
-
-#include <unordered_map>
-#include <vector>
-
-#include <fst/extensions/mpdt/mpdt.h>
-#include <fst/fst.h>
-
-namespace fst {
-
-// Compute various information about MPDTs, helper class for mpdtinfo.cc.
-template <class Arc, typename Arc::Label nlevels = 2>
-class MPdtInfo {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  MPdtInfo(const Fst<Arc> &fst,
-           const std::vector<std::pair<Label, Label>> &parens,
-           const std::vector<Label> &assignments);
-
-  const string &FstType() const { return fst_type_; }
-
-  const string &ArcType() const { return Arc::Type(); }
-
-  int64 NumStates() const { return nstates_; }
-
-  int64 NumArcs() const { return narcs_; }
-
-  int64 NumLevels() const { return nlevels; }
-
-  int64 NumOpenParens(Label level) const { return nopen_parens_[level]; }
-
-  int64 NumCloseParens(Label level) const { return nclose_parens_[level]; }
-
-  int64 NumUniqueOpenParens(Label level) const {
-    return nuniq_open_parens_[level];
-  }
-
-  int64 NumUniqueCloseParens(Label level) const {
-    return nuniq_close_parens_[level];
-  }
-  int64 NumOpenParenStates(Label level) const {
-    return nopen_paren_states_[level];
-  }
-
-  int64 NumCloseParenStates(Label level) const {
-    return nclose_paren_states_[level];
-  }
-
-  void Print();
-
- private:
-  string fst_type_;
-  int64 nstates_;
-  int64 narcs_;
-  int64 nopen_parens_[nlevels];
-  int64 nclose_parens_[nlevels];
-  int64 nuniq_open_parens_[nlevels];
-  int64 nuniq_close_parens_[nlevels];
-  int64 nopen_paren_states_[nlevels];
-  int64 nclose_paren_states_[nlevels];
-  bool error_;
-};
-
-template <class Arc, typename Arc::Label nlevels>
-MPdtInfo<Arc, nlevels>::MPdtInfo(
-    const Fst<Arc> &fst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens,
-    const std::vector<typename Arc::Label> &assignments)
-    : fst_type_(fst.Type()), nstates_(0), narcs_(0), error_(false) {
-  std::unordered_map<Label, size_t> paren_map;
-  std::unordered_set<Label> paren_set;
-  std::unordered_map<Label, int> paren_levels;
-  std::unordered_set<StateId> open_paren_state_set;
-  std::unordered_set<StateId> close_paren_state_set;
-  if (parens.size() != assignments.size()) {
-    FSTERROR() << "MPdtInfo: Parens of different size from assignments";
-    error_ = true;
-    return;
-  }
-  for (Label i = 0; i < assignments.size(); ++i) {
-    // Assignments here start at 0, so assuming the human-readable version has
-    // them starting at 1, we should subtract 1 here.
-    Label level = assignments[i] - 1;
-    if (level < 0 || level >= nlevels) {
-      FSTERROR() << "MPdtInfo: Specified level " << level << " out of bounds";
-      error_ = true;
-      return;
-    }
-    const auto &pair = parens[i];
-    paren_levels[pair.first] = level;
-    paren_levels[pair.second] = level;
-    paren_map[pair.first] = i;
-    paren_map[pair.second] = i;
-  }
-  for (Label i = 0; i < nlevels; ++i) {
-    nopen_parens_[i] = 0;
-    nclose_parens_[i] = 0;
-    nuniq_open_parens_[i] = 0;
-    nuniq_close_parens_[i] = 0;
-    nopen_paren_states_[i] = 0;
-    nclose_paren_states_[i] = 0;
-  }
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    ++nstates_;
-    const auto s = siter.Value();
-    for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      ++narcs_;
-      const auto it = paren_map.find(arc.ilabel);
-      if (it != paren_map.end()) {
-        const auto open_paren = parens[it->second].first;
-        const auto close_paren = parens[it->second].second;
-        const auto level = paren_levels[arc.ilabel];
-        if (arc.ilabel == open_paren) {
-          ++nopen_parens_[level];
-          if (!paren_set.count(open_paren)) {
-            ++nuniq_open_parens_[level];
-            paren_set.insert(open_paren);
-          }
-          if (!open_paren_state_set.count(arc.nextstate)) {
-            ++nopen_paren_states_[level];
-            open_paren_state_set.insert(arc.nextstate);
-          }
-        } else {
-          ++nclose_parens_[level];
-          if (!paren_set.count(close_paren)) {
-            ++nuniq_close_parens_[level];
-            paren_set.insert(close_paren);
-          }
-          if (!close_paren_state_set.count(s)) {
-            ++nclose_paren_states_[level];
-            close_paren_state_set.insert(s);
-          }
-        }
-      }
-    }
-  }
-}
-
-template <class Arc, typename Arc::Label nlevels>
-void MPdtInfo<Arc, nlevels>::Print() {
-  const auto old = std::cout.setf(std::ios::left);
-  std::cout.width(50);
-  std::cout << "fst type" << FstType() << std::endl;
-  std::cout.width(50);
-  std::cout << "arc type" << ArcType() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of states" << NumStates() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of arcs" << NumArcs() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of levels" << NumLevels() << std::endl;
-  std::cout.width(50);
-  for (typename Arc::Label i = 0; i < nlevels; ++i) {
-    int level = i + 1;
-    std::cout << "# of open parentheses at levelel " << level << "\t"
-              << NumOpenParens(i) << std::endl;
-    std::cout.width(50);
-    std::cout << "# of close parentheses at levelel " << level << "\t"
-              << NumCloseParens(i) << std::endl;
-    std::cout.width(50);
-    std::cout << "# of unique open parentheses at levelel " << level << "\t"
-              << NumUniqueOpenParens(i) << std::endl;
-    std::cout.width(50);
-    std::cout << "# of unique close parentheses at levelel " << level << "\t"
-              << NumUniqueCloseParens(i) << std::endl;
-    std::cout.width(50);
-    std::cout << "# of open parenthesis dest. states at levelel " << level
-              << "\t" << NumOpenParenStates(i) << std::endl;
-    std::cout.width(50);
-    std::cout << "# of close parenthesis source states at levelel " << level
-              << "\t" << NumCloseParenStates(i) << std::endl;
-    std::cout.width(50);
-  }
-  std::cout.setf(old);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_MPDT_INFO_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/mpdt.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/mpdt.h
deleted file mode 100644
index 77ff7574..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/mpdt.h
+++ /dev/null
@@ -1,357 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Common classes for Multi Pushdown Transducer (MPDT) expansion/traversal.
-
-#ifndef FST_EXTENSIONS_MPDT_MPDT_H_
-#define FST_EXTENSIONS_MPDT_MPDT_H_
-
-#include <array>
-#include <functional>
-#include <map>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/extensions/pdt/pdt.h>
-
-namespace fst {
-
-enum MPdtType {
-  MPDT_READ_RESTRICT,   // Can only read from first empty stack
-  MPDT_WRITE_RESTRICT,  // Can only write to first empty stack
-  MPDT_NO_RESTRICT,     // No read-write restrictions
-};
-
-namespace internal {
-
-// PLEASE READ THIS CAREFULLY:
-//
-// When USEVECTOR is set, the stack configurations --- the statewise
-// representation of the StackId's for each substack --- is stored in a vector.
-// I would like to do this using an array for efficiency reasons, thus the
-// definition of StackConfig below. However, while this *works* in that tests
-// pass, etc. It causes a memory leak in the compose and expand tests, evidently
-// in the map[] that is being used to store the mapping between these
-// StackConfigs and the external StackId that the caller sees. There are no
-// memory leaks when I use a vector, only with this StackId. Why there should be
-// memory leaks given that I am not mallocing anything is a mystery. In case you
-// were wondering, clearing the map at the end does not help.
-
-template <typename StackId, typename Level, Level nlevels>
-struct StackConfig {
-  StackConfig() : array_() {}
-
-  StackConfig(const StackConfig<StackId, Level, nlevels> &config) {
-    array_ = config.array_;
-  }
-
-  StackId &operator[](const int index) { return array_[index]; }
-
-  const StackId &operator[](const int index) const { return array_[index]; }
-
-  StackConfig &operator=(const StackConfig<StackId, Level, nlevels> &config) {
-    if (this == &config) return *this;
-    array_ = config.array_;
-    return *this;
-  }
-
-  std::array<StackId, nlevels> array_;
-};
-
-template <typename StackId, typename Level, Level nlevels>
-class CompConfig {
- public:
-  using Config = StackConfig<StackId, Level, nlevels>;
-
-  bool operator()(const Config &x, const Config &y) const {
-    for (Level level = 0; level < nlevels; ++level) {
-      if (x.array_[level] < y.array_[level]) {
-        return true;
-      } else if (x.array_[level] > y.array_[level]) {
-        return false;
-      }
-    }
-    return false;
-  }
-};
-
-// Defines the KeyPair type used as the key to MPdtStack.paren_id_map_. The hash
-// function is provided as a separate struct to match templating syntax.
-template <typename Level>
-struct KeyPair {
-  Level level;
-  size_t underlying_id;
-
-  KeyPair(Level level, size_t id) : level(level), underlying_id(id) {}
-
-  inline bool operator==(const KeyPair<Level> &rhs) const {
-    return level == rhs.level && underlying_id == rhs.underlying_id;
-  }
-};
-
-template <typename Level>
-struct KeyPairHasher {
-  inline size_t operator()(const KeyPair<Level> &keypair) const {
-    return std::hash<Level>()(keypair.level) ^
-           (std::hash<size_t>()(keypair.underlying_id) << 1);
-  }
-};
-
-template <typename StackId, typename Level, Level nlevels = 2,
-          MPdtType restrict = MPDT_READ_RESTRICT>
-class MPdtStack {
- public:
-  using Label = Level;
-  using Config = StackConfig<StackId, Level, nlevels>;
-  using ConfigToStackId =
-      std::map<Config, StackId, CompConfig<StackId, Level, nlevels>>;
-
-  MPdtStack(const std::vector<std::pair<Label, Label>> &parens,
-            const std::vector<Level> &assignments);
-
-  MPdtStack(const MPdtStack &mstack);
-
-  ~MPdtStack() {
-    for (Level level = 0; level < nlevels; ++level) delete stacks_[level];
-  }
-
-  StackId Find(StackId stack_id, Label label);
-
-  // For now we do not implement Pop since this is needed only for
-  // ShortestPath().
-
-  // For Top we find the first non-empty config, and find the paren ID of that
-  // (or -1) if there is none, then map that to the external stack_id to return.
-  ssize_t Top(StackId stack_id) const {
-    if (stack_id == -1) return -1;
-    const auto config = InternalStackIds(stack_id);
-    Level level = 0;
-    StackId underlying_id = -1;
-    for (; level < nlevels; ++level) {
-      if (!Empty(config, level)) {
-        underlying_id = stacks_[level]->Top(config[level]);
-        break;
-      }
-    }
-    if (underlying_id == -1) return -1;
-    const auto it = paren_id_map_.find(KeyPair<Level>(level, underlying_id));
-    if (it == paren_id_map_.end()) return -1;  // NB: shouldn't happen.
-    return it->second;
-  }
-
-  ssize_t ParenId(Label label) const {
-    const auto it = paren_map_.find(label);
-    return it != paren_map_.end() ? it->second : -1;
-  }
-
-  // TODO(rws): For debugging purposes only: remove later.
-  string PrintConfig(const Config &config) const {
-    string result = "[";
-    for (Level i = 0; i < nlevels; ++i) {
-      char s[128];
-      snprintf(s, sizeof(s), "%d", config[i]);
-      result += string(s);
-      if (i < nlevels - 1) result += ", ";
-    }
-    result += "]";
-    return result;
-  }
-
-  bool Error() { return error_; }
-
-  // Each component stack has an internal stack ID for a given configuration and
-  // label.
-  // This function maps a configuration of those to the stack ID the caller
-  // sees.
-  inline StackId ExternalStackId(const Config &config) {
-    const auto it = config_to_stack_id_map_.find(config);
-    StackId result;
-    if (it == config_to_stack_id_map_.end()) {
-      result = next_stack_id_++;
-      config_to_stack_id_map_.insert(
-          std::pair<Config, StackId>(config, result));
-      stack_id_to_config_map_[result] = config;
-    } else {
-      result = it->second;
-    }
-    return result;
-  }
-
-  // Retrieves the internal stack ID for a corresponding external stack ID.
-  inline const Config InternalStackIds(StackId stack_id) const {
-    auto it = stack_id_to_config_map_.find(stack_id);
-    if (it == stack_id_to_config_map_.end()) {
-      it = stack_id_to_config_map_.find(-1);
-    }
-    return it->second;
-  }
-
-  inline bool Empty(const Config &config, Level level) const {
-    return config[level] <= 0;
-  }
-
-  inline bool AllEmpty(const Config &config) {
-    for (Level level = 0; level < nlevels; ++level) {
-      if (!Empty(config, level)) return false;
-    }
-    return true;
-  }
-
-  bool error_;
-  Label min_paren_;
-  Label max_paren_;
-  // Stores level of each paren.
-  std::unordered_map<Label, Label> paren_levels_;
-  std::vector<std::pair<Label, Label>> parens_;  // As in pdt.h.
-  std::unordered_map<Label, size_t> paren_map_;  // As in pdt.h.
-  // Maps between internal paren_id and external paren_id.
-  std::unordered_map<KeyPair<Level>, size_t, KeyPairHasher<Level>>
-      paren_id_map_;
-  // Maps between internal stack ids and external stack id.
-  ConfigToStackId config_to_stack_id_map_;
-  std::unordered_map<StackId, Config> stack_id_to_config_map_;
-  StackId next_stack_id_;
-  // Array of stacks.
-  PdtStack<StackId, Label> *stacks_[nlevels];
-};
-
-template <typename StackId, typename Level, Level nlevels, MPdtType restrict>
-MPdtStack<StackId, Level, nlevels, restrict>::MPdtStack(
-    const std::vector<std::pair<Level, Level>> &parens,  // NB: Label = Level.
-    const std::vector<Level> &assignments)
-    : error_(false),
-      min_paren_(kNoLabel),
-      max_paren_(kNoLabel),
-      parens_(parens),
-      next_stack_id_(1) {
-  using Label = Level;
-  if (parens.size() != assignments.size()) {
-    FSTERROR() << "MPdtStack: Parens of different size from assignments";
-    error_ = true;
-    return;
-  }
-  std::vector<std::pair<Label, Label>> vectors[nlevels];
-  for (Level i = 0; i < assignments.size(); ++i) {
-    // Assignments here start at 0, so assuming the human-readable version has
-    // them starting at 1, we should subtract 1 here
-    const auto level = assignments[i] - 1;
-    if (level < 0 || level >= nlevels) {
-      FSTERROR() << "MPdtStack: Specified level " << level << " out of bounds";
-      error_ = true;
-      return;
-    }
-    const auto &pair = parens[i];
-    vectors[level].push_back(pair);
-    paren_levels_[pair.first] = level;
-    paren_levels_[pair.second] = level;
-    paren_map_[pair.first] = i;
-    paren_map_[pair.second] = i;
-    const KeyPair<Level> key(level, vectors[level].size() - 1);
-    paren_id_map_[key] = i;
-    if (min_paren_ == kNoLabel || pair.first < min_paren_) {
-      min_paren_ = pair.first;
-    }
-    if (pair.second < min_paren_) min_paren_ = pair.second;
-    if (max_paren_ == kNoLabel || pair.first > max_paren_) {
-      max_paren_ = pair.first;
-    }
-    if (pair.second > max_paren_) max_paren_ = pair.second;
-  }
-  using Config = StackConfig<StackId, Level, nlevels>;
-  Config neg_one;
-  Config zero;
-  for (Level level = 0; level < nlevels; ++level) {
-    stacks_[level] = new PdtStack<StackId, Label>(vectors[level]);
-    neg_one[level] = -1;
-    zero[level] = 0;
-  }
-  config_to_stack_id_map_[neg_one] = -1;
-  config_to_stack_id_map_[zero] = 0;
-  stack_id_to_config_map_[-1] = neg_one;
-  stack_id_to_config_map_[0] = zero;
-}
-
-template <typename StackId, typename Level, Level nlevels, MPdtType restrict>
-MPdtStack<StackId, Level, nlevels, restrict>::MPdtStack(
-    const MPdtStack<StackId, Level, nlevels, restrict> &mstack)
-    : error_(mstack.error_),
-      min_paren_(mstack.min_paren_),
-      max_paren_(mstack.max_paren_),
-      parens_(mstack.parens_),
-      next_stack_id_(mstack.next_stack_id_) {
-  for (const auto &kv : mstack.paren_levels_) {
-    paren_levels_[kv.first] = kv.second;
-  }
-  for (const auto &paren : mstack.parens_) parens_.push_back(paren);
-  for (const auto &kv : mstack.paren_map_) {
-    paren_map_[kv.first] = kv.second;
-  }
-  for (const auto &kv : mstack.paren_id_map_) {
-    paren_id_map_[kv.first] = kv.second;
-  }
-  for (auto it = mstack.config_to_stack_id_map_.begin();
-       it != mstack.config_to_stack_id_map_.end(); ++it) {
-    config_to_stack_id_map_[it->first] = it->second;
-  }
-  for (const auto &kv : mstack.stack_id_to_config_map_) {
-    using Config = StackConfig<StackId, Level, nlevels>;
-    const Config config(kv.second);
-    stack_id_to_config_map_[kv.first] = config;
-  }
-  for (Level level = 0; level < nlevels; ++level)
-    stacks_[level] = mstack.stacks_[level];
-}
-
-template <typename StackId, typename Level, Level nlevels, MPdtType restrict>
-StackId MPdtStack<StackId, Level, nlevels, restrict>::Find(StackId stack_id,
-                                                           Level label) {
-  // Non-paren.
-  if (min_paren_ == kNoLabel || label < min_paren_ || label > max_paren_) {
-    return stack_id;
-  }
-  const auto it = paren_map_.find(label);
-  // Non-paren.
-  if (it == paren_map_.end()) return stack_id;
-  ssize_t paren_id = it->second;
-  // Gets the configuration associated with this stack_id.
-  const auto config = InternalStackIds(stack_id);
-  // Gets the level.
-  const auto level = paren_levels_.find(label)->second;
-  // If the label is an open paren we push:
-  //
-  // 1) if the restrict type is not MPDT_WRITE_RESTRICT, or
-  // 2) the restrict type is MPDT_WRITE_RESTRICT, and all the stacks above the
-  // level are empty.
-  if (label == parens_[paren_id].first) {  // Open paren.
-    if (restrict == MPDT_WRITE_RESTRICT) {
-      for (Level upper_level = 0; upper_level < level; ++upper_level) {
-        if (!Empty(config, upper_level)) return -1;  // Non-empty stack blocks.
-      }
-    }
-    // If the label is an close paren we pop:
-    //
-    // 1) if the restrict type is not MPDT_READ_RESTRICT, or
-    // 2) the restrict type is MPDT_READ_RESTRICT, and all the stacks above the
-    // level are empty.
-  } else if (restrict == MPDT_READ_RESTRICT) {
-    for (Level lower_level = 0; lower_level < level; ++lower_level) {
-      if (!Empty(config, lower_level)) return -1;  // Non-empty stack blocks.
-    }
-  }
-  const auto nid = stacks_[level]->Find(config[level], label);
-  // If the new ID is -1, that means that there is no valid transition at the
-  // level we want.
-  if (nid == -1) {
-    return -1;
-  } else {
-    using Config = StackConfig<StackId, Level, nlevels>;
-    Config nconfig(config);
-    nconfig[level] = nid;
-    return ExternalStackId(nconfig);
-  }
-}
-
-}  // namespace internal
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_MPDT_MPDT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/mpdtlib.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/mpdtlib.h
deleted file mode 100644
index bbf206f1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/mpdtlib.h
+++ /dev/null
@@ -1,18 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// This is an experimental multipush-down transducer (MPDT) library. An MPDT is
-// encoded as an FST, where some transitions are labeled with open or close
-// parentheses, each mated pair of which is associated to one stack. To be
-// interpreted as an MPDT, the parentheses within a stack must balance on a
-// path.
-
-#ifndef FST_EXTENSIONS_MPDT_MPDTLIB_H_
-#define FST_EXTENSIONS_MPDT_MPDTLIB_H_
-
-#include <fst/extensions/mpdt/compose.h>
-#include <fst/extensions/mpdt/expand.h>
-#include <fst/extensions/mpdt/mpdt.h>
-#include <fst/extensions/mpdt/reverse.h>
-
-#endif  // FST_EXTENSIONS_MPDT_MPDTLIB_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/mpdtscript.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/mpdtscript.h
deleted file mode 100644
index 66340992..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/mpdtscript.h
+++ /dev/null
@@ -1,156 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Convenience file for including all MPDT operations at once, and/or
-// registering them for new arc types.
-
-#ifndef FST_EXTENSIONS_MPDT_MPDTSCRIPT_H_
-#define FST_EXTENSIONS_MPDT_MPDTSCRIPT_H_
-
-#include <algorithm>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-#include <fst/compose.h>  // for ComposeOptions
-#include <fst/util.h>
-
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/shortest-path.h>
-
-#include <fst/extensions/mpdt/compose.h>
-#include <fst/extensions/mpdt/expand.h>
-#include <fst/extensions/mpdt/info.h>
-#include <fst/extensions/mpdt/reverse.h>
-
-#include <fst/extensions/pdt/pdtscript.h>  // For LabelClassPair,
-                                               // FstClassPair, and to detect
-                                               // any collisions.
-
-namespace fst {
-namespace script {
-
-using MPdtComposeArgs =
-    std::tuple<const FstClass &, const FstClass &,
-               const std::vector<LabelPair> &, const std::vector<int64> &,
-               MutableFstClass *, const MPdtComposeOptions &, bool>;
-
-template <class Arc>
-void MPdtCompose(MPdtComposeArgs *args) {
-  const Fst<Arc> &ifst1 = *(std::get<0>(*args).GetFst<Arc>());
-  const Fst<Arc> &ifst2 = *(std::get<1>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<4>(*args)->GetMutableFst<Arc>();
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<2>(*args).size());
-  std::copy(std::get<2>(*args).begin(), std::get<2>(*args).end(),
-            typed_parens.begin());
-  using Level = typename Arc::Label;
-  std::vector<Level> typed_assignments(std::get<3>(*args).size());
-  std::copy(std::get<3>(*args).begin(), std::get<3>(*args).end(),
-            typed_assignments.begin());
-  if (std::get<6>(*args)) {
-    Compose(ifst1, typed_parens, typed_assignments, ifst2, ofst,
-            std::get<5>(*args));
-  } else {
-    Compose(ifst1, ifst2, typed_parens, typed_assignments, ofst,
-            std::get<5>(*args));
-  }
-}
-
-void MPdtCompose(const FstClass &ifst1, const FstClass &ifst2,
-                 const std::vector<LabelPair> &parens,
-                 const std::vector<int64> &assignments, MutableFstClass *ofst,
-                 const MPdtComposeOptions &copts, bool left_pdt);
-
-using MPdtExpandArgs =
-    std::tuple<const FstClass &, const std::vector<LabelPair> &,
-               const std::vector<int64> &, MutableFstClass *,
-               const MPdtExpandOptions &>;
-
-template <class Arc>
-void MPdtExpand(MPdtExpandArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<3>(*args)->GetMutableFst<Arc>();
-  // In case Arc::Label is not the same as FstClass::Label, we make copies.
-  // Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_parens.begin());
-  using Level = typename Arc::Label;
-  std::vector<Level> typed_assignments(std::get<2>(*args).size());
-  std::copy(std::get<2>(*args).begin(), std::get<2>(*args).end(),
-            typed_assignments.begin());
-  Expand(fst, typed_parens, typed_assignments, ofst,
-         MPdtExpandOptions(std::get<4>(*args).connect,
-                           std::get<4>(*args).keep_parentheses));
-}
-
-void MPdtExpand(const FstClass &ifst, const std::vector<LabelPair> &parens,
-                const std::vector<int64> &assignments, MutableFstClass *ofst,
-                const MPdtExpandOptions &opts);
-
-using MPdtReverseArgs =
-    std::tuple<const FstClass &, const std::vector<LabelPair> &,
-               std::vector<int64> *, MutableFstClass *>;
-
-template <class Arc>
-void MPdtReverse(MPdtReverseArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<3>(*args)->GetMutableFst<Arc>();
-  // In case Arc::Label is not the same as FstClass::Label, we make copies.
-  // Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_parens.begin());
-  using Level = typename Arc::Label;
-  std::vector<Level> typed_assignments(std::get<2>(*args)->size());
-  std::copy(std::get<2>(*args)->begin(), std::get<2>(*args)->end(),
-            typed_assignments.begin());
-  Reverse(fst, typed_parens, &typed_assignments, ofst);
-  // Reassign stack assignments to input assignment vector.
-  std::copy(typed_assignments.begin(), typed_assignments.end(),
-            std::get<2>(*args)->begin());
-}
-
-void MPdtReverse(const FstClass &ifst, const std::vector<LabelPair> &parens,
-                 std::vector<int64> *assignments, MutableFstClass *ofst);
-
-using PrintMPdtInfoArgs =
-    std::tuple<const FstClass &, const std::vector<LabelPair> &,
-               const std::vector<int64> &>;
-
-template <class Arc>
-void PrintMPdtInfo(PrintMPdtInfoArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  // In case Arc::Label is not the same as FstClass::Label, we make copies.
-  // Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_parens.begin());
-  using Level = typename Arc::Label;
-  std::vector<Level> typed_assignments(std::get<2>(*args).size());
-  std::copy(std::get<2>(*args).begin(), std::get<2>(*args).end(),
-            typed_assignments.begin());
-  MPdtInfo<Arc> mpdtinfo(fst, typed_parens, typed_assignments);
-  mpdtinfo.Print();
-}
-
-void PrintMPdtInfo(const FstClass &ifst, const std::vector<LabelPair> &parens,
-                   const std::vector<int64> &assignments);
-
-}  // namespace script
-}  // namespace fst
-
-#define REGISTER_FST_MPDT_OPERATIONS(ArcType)                    \
-  REGISTER_FST_OPERATION(MPdtCompose, ArcType, MPdtComposeArgs); \
-  REGISTER_FST_OPERATION(MPdtExpand, ArcType, MPdtExpandArgs);   \
-  REGISTER_FST_OPERATION(MPdtReverse, ArcType, MPdtReverseArgs); \
-  REGISTER_FST_OPERATION(PrintMPdtInfo, ArcType, PrintMPdtInfoArgs)
-#endif  // FST_EXTENSIONS_MPDT_MPDTSCRIPT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/read_write_utils.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/read_write_utils.h
deleted file mode 100644
index c77d7488..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/read_write_utils.h
+++ /dev/null
@@ -1,86 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Definition of ReadLabelTriples based on ReadLabelPairs, like that in
-// nlp/fst/lib/util.h for pairs, and similarly for WriteLabelTriples.
-
-#ifndef FST_EXTENSIONS_MPDT_READ_WRITE_UTILS_H_
-#define FST_EXTENSIONS_MPDT_READ_WRITE_UTILS_H_
-
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fstream>
-#include <fst/test-properties.h>
-
-namespace fst {
-
-// Returns true on success.
-template <typename Label>
-bool ReadLabelTriples(const string &filename,
-                      std::vector<std::pair<Label, Label>> *pairs,
-                      std::vector<Label> *assignments,
-                      bool allow_negative = false) {
-  std::ifstream fstrm(filename);
-  if (!fstrm) {
-    LOG(ERROR) << "ReadIntTriples: Can't open file: " << filename;
-    return false;
-  }
-  static constexpr auto kLineLen = 8096;
-  char line[kLineLen];
-  size_t nline = 0;
-  pairs->clear();
-  while (fstrm.getline(line, kLineLen)) {
-    ++nline;
-    std::vector<char *> col;
-    SplitToVector(line, "\n\t ", &col, true);
-    // Empty line or comment?
-    if (col.empty() || col[0][0] == '\0' || col[0][0] == '#') continue;
-    if (col.size() != 3) {
-      LOG(ERROR) << "ReadLabelTriples: Bad number of columns, "
-                 << "file = " << filename << ", line = " << nline;
-      return false;
-    }
-    bool err;
-    const Label i1 = StrToInt64(col[0], filename, nline, allow_negative, &err);
-    if (err) return false;
-    const Label i2 = StrToInt64(col[1], filename, nline, allow_negative, &err);
-    if (err) return false;
-    using Level = Label;
-    const Level i3 = StrToInt64(col[2], filename, nline, allow_negative, &err);
-    if (err) return false;
-    pairs->push_back(std::make_pair(i1, i2));
-    assignments->push_back(i3);
-  }
-  return true;
-}
-
-// Returns true on success.
-template <typename Label>
-bool WriteLabelTriples(const string &filename,
-                       const std::vector<std::pair<Label, Label>> &pairs,
-                       const std::vector<Label> &assignments) {
-  if (pairs.size() != assignments.size()) {
-    LOG(ERROR) << "WriteLabelTriples: Pairs and assignments of different sizes";
-    return false;
-  }
-  std::ofstream fstrm(filename);
-  if (!fstrm) {
-    LOG(ERROR) << "WriteLabelTriples: Can't open file: " << filename;
-    return false;
-  }
-  for (size_t n = 0; n < pairs.size(); ++n)
-    fstrm << pairs[n].first << "\t" << pairs[n].second << "\t" << assignments[n]
-          << "\n";
-  if (!fstrm) {
-    LOG(ERROR) << "WriteLabelTriples: Write failed: "
-               << (filename.empty() ? "standard output" : filename);
-    return false;
-  }
-  return true;
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_MPDT_READ_WRITE_UTILS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/reverse.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/reverse.h
deleted file mode 100644
index 413dd3f0..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/mpdt/reverse.h
+++ /dev/null
@@ -1,54 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Reverses an MPDT.
-
-#ifndef FST_EXTENSIONS_MPDT_REVERSE_H_
-#define FST_EXTENSIONS_MPDT_REVERSE_H_
-
-#include <limits>
-#include <vector>
-
-#include <fst/mutable-fst.h>
-#include <fst/relabel.h>
-#include <fst/reverse.h>
-
-namespace fst {
-
-// Reverses a multi-stack pushdown transducer (MPDT) encoded as an FST.
-template <class Arc, class RevArc>
-void Reverse(
-    const Fst<Arc> &ifst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens,
-    std::vector<typename Arc::Label> *assignments, MutableFst<RevArc> *ofst) {
-  using Label = typename Arc::Label;
-  // Reverses FST component.
-  Reverse(ifst, ofst);
-  // Exchanges open and close parenthesis pairs.
-  std::vector<std::pair<Label, Label>> relabel_pairs;
-  relabel_pairs.reserve(2 * parens.size());
-  for (const auto &pair : parens) {
-    relabel_pairs.emplace_back(pair.first, pair.second);
-    relabel_pairs.emplace_back(pair.second, pair.first);
-  }
-  Relabel(ofst, relabel_pairs, relabel_pairs);
-  // Computes new bounds for the stack assignments.
-  Label max_level = -1;
-  Label min_level = std::numeric_limits<Label>::max();
-  for (const auto assignment : *assignments) {
-    if (assignment < min_level) {
-      min_level = assignment;
-    } else if (assignment > max_level) {
-      max_level = assignment;
-    }
-  }
-  // Actually reverses stack assignments.
-  for (auto &assignment : *assignments) {
-    assignment = (max_level - assignment) + min_level;
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_MPDT_REVERSE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/ngram/bitmap-index.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/ngram/bitmap-index.h
deleted file mode 100644
index 09ad6d7d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/ngram/bitmap-index.h
+++ /dev/null
@@ -1,168 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_NGRAM_BITMAP_INDEX_H_
-#define FST_EXTENSIONS_NGRAM_BITMAP_INDEX_H_
-
-#include <utility>
-#include <vector>
-
-#include <fst/compat.h>
-
-// This class is a bitstring storage class with an index that allows
-// seeking to the Nth set or clear bit in time O(Log(N)) where N is
-// the length of the bit vector.  In addition, it allows counting set or
-// clear bits over ranges in constant time.
-//
-// This is accomplished by maintaining an "secondary" index of limited
-// size in bits that maintains a running count of the number of bits set
-// in each block of bitmap data.  A block is defined as the number of
-// uint64 values that can fit in the secondary index before an overflow
-// occurs.
-//
-// To handle overflows, a "primary" index containing a running count of
-// bits set in each block is created using the type uint64.
-
-namespace fst {
-
-class BitmapIndex {
- public:
-  static size_t StorageSize(size_t size) {
-    return ((size + kStorageBlockMask) >> kStorageLogBitSize);
-  }
-
-  BitmapIndex() : bits_(nullptr), size_(0) {}
-
-  bool Get(size_t index) const {
-    return (bits_[index >> kStorageLogBitSize] &
-            (kOne << (index & kStorageBlockMask))) != 0;
-  }
-
-  static void Set(uint64* bits, size_t index) {
-    bits[index >> kStorageLogBitSize] |= (kOne << (index & kStorageBlockMask));
-  }
-
-  static void Clear(uint64* bits, size_t index) {
-    bits[index >> kStorageLogBitSize] &= ~(kOne << (index & kStorageBlockMask));
-  }
-
-  size_t Bits() const { return size_; }
-
-  size_t ArraySize() const { return StorageSize(size_); }
-
-  // Returns the number of one bits in the bitmap
-  size_t GetOnesCount() const {
-    return primary_index_[primary_index_size() - 1];
-  }
-
-  // Returns the number of one bits in positions 0 to limit - 1.
-  // REQUIRES: limit <= Bits()
-  size_t Rank1(size_t end) const;
-
-  // Returns the number of one bits in the range start to end - 1.
-  // REQUIRES: limit <= Bits()
-  size_t GetOnesCountInRange(size_t start, size_t end) const {
-    return Rank1(end) - Rank1(start);
-  }
-
-  // Returns the number of zero bits in positions 0 to limit - 1.
-  // REQUIRES: limit <= Bits()
-  size_t Rank0(size_t end) const { return end - Rank1(end); }
-
-  // Returns the number of zero bits in the range start to end - 1.
-  // REQUIRES: limit <= Bits()
-  size_t GetZeroesCountInRange(size_t start, size_t end) const {
-    return end - start - GetOnesCountInRange(start, end);
-  }
-
-  // Return true if any bit between begin inclusive and end exclusive
-  // is set.  0 <= begin <= end <= Bits() is required.
-  //
-  bool TestRange(size_t start, size_t end) const {
-    return Rank1(end) > Rank1(start);
-  }
-
-  // Returns the offset to the nth set bit (zero based)
-  // or Bits() if index >= number of ones
-  size_t Select1(size_t bit_index) const;
-
-  // Returns the offset to the nth clear bit (zero based)
-  // or Bits() if index > number of
-  size_t Select0(size_t bit_index) const;
-
-  // Returns the offset of the nth and nth+1 clear bit (zero based),
-  // equivalent to two calls to Select0, but more efficient.
-  std::pair<size_t, size_t> Select0s(size_t bit_index) const;
-
-  // Rebuilds from index for the associated Bitmap, should be called
-  // whenever changes have been made to the Bitmap or else behavior
-  // of the indexed bitmap methods will be undefined.
-  void BuildIndex(const uint64* bits, size_t size);
-
-  // the secondary index accumulates counts until it can possibly overflow
-  // this constant computes the number of uint64 units that can fit into
-  // units the size of uint16.
-  static const uint64 kOne = 1;
-  static const uint32 kStorageBitSize = 64;
-  static const uint32 kStorageLogBitSize = 6;
-  static const uint32 kSecondaryBlockSize =
-      ((1 << 16) - 1) >> kStorageLogBitSize;
-
- private:
-  static const uint32 kStorageBlockMask = kStorageBitSize - 1;
-
-  // returns, from the index, the count of ones up to array_index
-  size_t get_index_ones_count(size_t array_index) const;
-
-  // because the indexes, both primary and secondary, contain a running
-  // count of the population of one bits contained in [0,i), there is
-  // no reason to have an element in the zeroth position as this value would
-  // necessarily be zero.  (The bits are indexed in a zero based way.)  Thus
-  // we don't store the 0th element in either index.  Both of the following
-  // functions, if greater than 0, must be decremented by one before retreiving
-  // the value from the corresponding array.
-  // returns the 1 + the block that contains the bitindex in question
-  // the inverted version works the same but looks for zeros using an inverted
-  // view of the index
-  size_t find_primary_block(size_t bit_index) const;
-
-  size_t find_inverted_primary_block(size_t bit_index) const;
-
-  // similarly, the secondary index (which resets its count to zero at
-  // the end of every kSecondaryBlockSize entries) does not store the element
-  // at 0.  Note that the rem_bit_index parameter is the number of bits
-  // within the secondary block, after the bits accounted for by the primary
-  // block have been removed (i.e. the remaining bits)  And, because we
-  // reset to zero with each new block, there is no need to store those
-  // actual zeros.
-  // returns 1 + the secondary block that contains the bitindex in question
-  size_t find_secondary_block(size_t block, size_t rem_bit_index) const;
-
-  size_t find_inverted_secondary_block(size_t block,
-                                       size_t rem_bit_index) const;
-
-  // We create a primary index based upon the number of secondary index
-  // blocks.  The primary index uses fields wide enough to accomodate any
-  // index of the bitarray so cannot overflow
-  // The primary index is the actual running
-  // count of one bits set for all blocks (and, thus, all uint64s).
-  size_t primary_index_size() const {
-    return (ArraySize() + kSecondaryBlockSize - 1) / kSecondaryBlockSize;
-  }
-
-  const uint64* bits_;
-  size_t size_;
-
-  // The primary index contains the running popcount of all blocks
-  // which means the nth value contains the popcounts of
-  // [0,n*kSecondaryBlockSize], however, the 0th element is omitted.
-  std::vector<uint32> primary_index_;
-  // The secondary index contains the running popcount of the associated
-  // bitmap.  It is the same length (in units of uint16) as the
-  // bitmap's map is in units of uint64s.
-  std::vector<uint16> secondary_index_;
-};
-
-}  // end namespace fst
-
-#endif  // FST_EXTENSIONS_NGRAM_BITMAP_INDEX_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/ngram/ngram-fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/ngram/ngram-fst.h
deleted file mode 100644
index 532f458e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/ngram/ngram-fst.h
+++ /dev/null
@@ -1,1010 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// NgramFst implements a n-gram language model based upon the LOUDS data
-// structure.  Please refer to "Unary Data Structures for Language Models"
-// http://research.google.com/pubs/archive/37218.pdf
-
-#ifndef FST_EXTENSIONS_NGRAM_NGRAM_FST_H_
-#define FST_EXTENSIONS_NGRAM_NGRAM_FST_H_
-
-#include <stddef.h>
-#include <string.h>
-#include <algorithm>
-#include <iostream>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/log.h>
-#include <fstream>
-#include <fst/extensions/ngram/bitmap-index.h>
-#include <fst/fstlib.h>
-#include <fst/mapped-file.h>
-
-namespace fst {
-template <class A>
-class NGramFst;
-template <class A>
-class NGramFstMatcher;
-
-// Instance data containing mutable state for bookkeeping repeated access to
-// the same state.
-template <class A>
-struct NGramFstInst {
-  typedef typename A::Label Label;
-  typedef typename A::StateId StateId;
-  typedef typename A::Weight Weight;
-  StateId state_;
-  size_t num_futures_;
-  size_t offset_;
-  size_t node_;
-  StateId node_state_;
-  std::vector<Label> context_;
-  StateId context_state_;
-  NGramFstInst()
-      : state_(kNoStateId),
-        node_state_(kNoStateId),
-        context_state_(kNoStateId) {}
-};
-
-namespace internal {
-
-// Implementation class for LOUDS based NgramFst interface.
-template <class A>
-class NGramFstImpl : public FstImpl<A> {
-  using FstImpl<A>::SetInputSymbols;
-  using FstImpl<A>::SetOutputSymbols;
-  using FstImpl<A>::SetType;
-  using FstImpl<A>::WriteHeader;
-
-  friend class ArcIterator<NGramFst<A>>;
-  friend class NGramFstMatcher<A>;
-
- public:
-  using FstImpl<A>::InputSymbols;
-  using FstImpl<A>::SetProperties;
-  using FstImpl<A>::Properties;
-
-  typedef A Arc;
-  typedef typename A::Label Label;
-  typedef typename A::StateId StateId;
-  typedef typename A::Weight Weight;
-
-  NGramFstImpl() {
-    SetType("ngram");
-    SetInputSymbols(nullptr);
-    SetOutputSymbols(nullptr);
-    SetProperties(kStaticProperties);
-  }
-
-  NGramFstImpl(const Fst<A> &fst, std::vector<StateId> *order_out);
-
-  explicit NGramFstImpl(const Fst<A> &fst) : NGramFstImpl(fst, nullptr) {}
-
-  NGramFstImpl(const NGramFstImpl &other) {
-    FSTERROR() << "Copying NGramFst Impls is not supported, use safe = false.";
-    SetProperties(kError, kError);
-  }
-
-  ~NGramFstImpl() override {
-    if (owned_) {
-      delete[] data_;
-    }
-  }
-
-  static NGramFstImpl<A> *Read(std::istream &strm,  // NOLINT
-                               const FstReadOptions &opts) {
-    NGramFstImpl<A> *impl = new NGramFstImpl();
-    FstHeader hdr;
-    if (!impl->ReadHeader(strm, opts, kMinFileVersion, &hdr)) return 0;
-    uint64 num_states, num_futures, num_final;
-    const size_t offset =
-        sizeof(num_states) + sizeof(num_futures) + sizeof(num_final);
-    // Peek at num_states and num_futures to see how much more needs to be read.
-    strm.read(reinterpret_cast<char *>(&num_states), sizeof(num_states));
-    strm.read(reinterpret_cast<char *>(&num_futures), sizeof(num_futures));
-    strm.read(reinterpret_cast<char *>(&num_final), sizeof(num_final));
-    size_t size = Storage(num_states, num_futures, num_final);
-    MappedFile *data_region = MappedFile::Allocate(size);
-    char *data = reinterpret_cast<char *>(data_region->mutable_data());
-    // Copy num_states, num_futures and num_final back into data.
-    memcpy(data, reinterpret_cast<char *>(&num_states), sizeof(num_states));
-    memcpy(data + sizeof(num_states), reinterpret_cast<char *>(&num_futures),
-           sizeof(num_futures));
-    memcpy(data + sizeof(num_states) + sizeof(num_futures),
-           reinterpret_cast<char *>(&num_final), sizeof(num_final));
-    strm.read(data + offset, size - offset);
-    if (strm.fail()) {
-      delete impl;
-      return nullptr;
-    }
-    impl->Init(data, false, data_region);
-    return impl;
-  }
-
-  bool Write(std::ostream &strm,  // NOLINT
-             const FstWriteOptions &opts) const {
-    FstHeader hdr;
-    hdr.SetStart(Start());
-    hdr.SetNumStates(num_states_);
-    WriteHeader(strm, opts, kFileVersion, &hdr);
-    strm.write(data_, StorageSize());
-    return !strm.fail();
-  }
-
-  StateId Start() const { return start_; }
-
-  Weight Final(StateId state) const {
-    if (final_index_.Get(state)) {
-      return final_probs_[final_index_.Rank1(state)];
-    } else {
-      return Weight::Zero();
-    }
-  }
-
-  size_t NumArcs(StateId state, NGramFstInst<A> *inst = nullptr) const {
-    if (inst == nullptr) {
-      const std::pair<size_t, size_t> zeros =
-          (state == 0) ? select_root_ : future_index_.Select0s(state);
-      return zeros.second - zeros.first - 1;
-    }
-    SetInstFuture(state, inst);
-    return inst->num_futures_ + ((state == 0) ? 0 : 1);
-  }
-
-  size_t NumInputEpsilons(StateId state) const {
-    // State 0 has no parent, thus no backoff.
-    if (state == 0) return 0;
-    return 1;
-  }
-
-  size_t NumOutputEpsilons(StateId state) const {
-    return NumInputEpsilons(state);
-  }
-
-  StateId NumStates() const { return num_states_; }
-
-  void InitStateIterator(StateIteratorData<A> *data) const {
-    data->base = 0;
-    data->nstates = num_states_;
-  }
-
-  static size_t Storage(uint64 num_states, uint64 num_futures,
-                        uint64 num_final) {
-    uint64 b64;
-    Weight weight;
-    Label label;
-    size_t offset =
-        sizeof(num_states) + sizeof(num_futures) + sizeof(num_final);
-    offset +=
-        sizeof(b64) * (BitmapIndex::StorageSize(num_states * 2 + 1) +
-                       BitmapIndex::StorageSize(num_futures + num_states + 1) +
-                       BitmapIndex::StorageSize(num_states));
-    offset += (num_states + 1) * sizeof(label) + num_futures * sizeof(label);
-    // Pad for alignemnt, see
-    // http://en.wikipedia.org/wiki/Data_structure_alignment#Computing_padding
-    offset = (offset + sizeof(weight) - 1) & ~(sizeof(weight) - 1);
-    offset += (num_states + 1) * sizeof(weight) + num_final * sizeof(weight) +
-              (num_futures + 1) * sizeof(weight);
-    return offset;
-  }
-
-  void SetInstFuture(StateId state, NGramFstInst<A> *inst) const {
-    if (inst->state_ != state) {
-      inst->state_ = state;
-      const std::pair<size_t, size_t> zeros = future_index_.Select0s(state);
-      inst->num_futures_ = zeros.second - zeros.first - 1;
-      inst->offset_ = future_index_.Rank1(zeros.first + 1);
-    }
-  }
-
-  void SetInstNode(NGramFstInst<A> *inst) const {
-    if (inst->node_state_ != inst->state_) {
-      inst->node_state_ = inst->state_;
-      inst->node_ = context_index_.Select1(inst->state_);
-    }
-  }
-
-  void SetInstContext(NGramFstInst<A> *inst) const {
-    SetInstNode(inst);
-    if (inst->context_state_ != inst->state_) {
-      inst->context_state_ = inst->state_;
-      inst->context_.clear();
-      size_t node = inst->node_;
-      while (node != 0) {
-        inst->context_.push_back(context_words_[context_index_.Rank1(node)]);
-        node = context_index_.Select1(context_index_.Rank0(node) - 1);
-      }
-    }
-  }
-
-  // Access to the underlying representation
-  const char *GetData(size_t *data_size) const {
-    *data_size = StorageSize();
-    return data_;
-  }
-
-  void Init(const char *data, bool owned, MappedFile *file = nullptr);
-
-  const std::vector<Label> &GetContext(StateId s, NGramFstInst<A> *inst) const {
-    SetInstFuture(s, inst);
-    SetInstContext(inst);
-    return inst->context_;
-  }
-
-  size_t StorageSize() const {
-    return Storage(num_states_, num_futures_, num_final_);
-  }
-
-  void GetStates(const std::vector<Label> &context,
-                 std::vector<StateId> *states) const;
-
- private:
-  StateId Transition(const std::vector<Label> &context, Label future) const;
-
-  // Properties always true for this Fst class.
-  static const uint64 kStaticProperties =
-      kAcceptor | kIDeterministic | kODeterministic | kEpsilons | kIEpsilons |
-      kOEpsilons | kILabelSorted | kOLabelSorted | kWeighted | kCyclic |
-      kInitialAcyclic | kNotTopSorted | kAccessible | kCoAccessible |
-      kNotString | kExpanded;
-  // Current file format version.
-  static const int kFileVersion = 4;
-  // Minimum file format version supported.
-  static const int kMinFileVersion = 4;
-
-  std::unique_ptr<MappedFile> data_region_;
-  const char *data_ = nullptr;
-  bool owned_ = false;  // True if we own data_
-  StateId start_ = fst::kNoStateId;
-  uint64 num_states_ = 0;
-  uint64 num_futures_ = 0;
-  uint64 num_final_ = 0;
-  std::pair<size_t, size_t> select_root_;
-  const Label *root_children_ = nullptr;
-  // borrowed references
-  const uint64 *context_ = nullptr;
-  const uint64 *future_ = nullptr;
-  const uint64 *final_ = nullptr;
-  const Label *context_words_ = nullptr;
-  const Label *future_words_ = nullptr;
-  const Weight *backoff_ = nullptr;
-  const Weight *final_probs_ = nullptr;
-  const Weight *future_probs_ = nullptr;
-  BitmapIndex context_index_;
-  BitmapIndex future_index_;
-  BitmapIndex final_index_;
-};
-
-template <typename A>
-inline void NGramFstImpl<A>::GetStates(
-    const std::vector<Label> &context,
-    std::vector<typename A::StateId> *states) const {
-  states->clear();
-  states->push_back(0);
-  typename std::vector<Label>::const_reverse_iterator cit = context.rbegin();
-  const Label *children = root_children_;
-  size_t num_children = select_root_.second - 2;
-  const Label *loc = std::lower_bound(children, children + num_children, *cit);
-  if (loc == children + num_children || *loc != *cit) return;
-  size_t node = 2 + loc - children;
-  states->push_back(context_index_.Rank1(node));
-  if (context.size() == 1) return;
-  size_t node_rank = context_index_.Rank1(node);
-  std::pair<size_t, size_t> zeros =
-      node_rank == 0 ? select_root_ : context_index_.Select0s(node_rank);
-  size_t first_child = zeros.first + 1;
-  ++cit;
-  if (context_index_.Get(first_child) != false) {
-    size_t last_child = zeros.second - 1;
-    while (cit != context.rend()) {
-      children = context_words_ + context_index_.Rank1(first_child);
-      loc = std::lower_bound(children, children + last_child - first_child + 1,
-                             *cit);
-      if (loc == children + last_child - first_child + 1 || *loc != *cit) {
-        break;
-      }
-      ++cit;
-      node = first_child + loc - children;
-      states->push_back(context_index_.Rank1(node));
-      node_rank = context_index_.Rank1(node);
-      zeros =
-          node_rank == 0 ? select_root_ : context_index_.Select0s(node_rank);
-      first_child = zeros.first + 1;
-      if (context_index_.Get(first_child) == false) break;
-      last_child = zeros.second - 1;
-    }
-  }
-}
-
-}  // namespace internal
-
-/*****************************************************************************/
-template <class A>
-class NGramFst : public ImplToExpandedFst<internal::NGramFstImpl<A>> {
-  friend class ArcIterator<NGramFst<A>>;
-  friend class NGramFstMatcher<A>;
-
- public:
-  typedef A Arc;
-  typedef typename A::StateId StateId;
-  typedef typename A::Label Label;
-  typedef typename A::Weight Weight;
-  typedef internal::NGramFstImpl<A> Impl;
-
-  explicit NGramFst(const Fst<A> &dst)
-      : ImplToExpandedFst<Impl>(std::make_shared<Impl>(dst, nullptr)) {}
-
-  NGramFst(const Fst<A> &fst, std::vector<StateId> *order_out)
-      : ImplToExpandedFst<Impl>(std::make_shared<Impl>(fst, order_out)) {}
-
-  // Because the NGramFstImpl is a const stateless data structure, there
-  // is never a need to do anything beside copy the reference.
-  NGramFst(const NGramFst<A> &fst, bool safe = false)
-      : ImplToExpandedFst<Impl>(fst, false) {}
-
-  NGramFst() : ImplToExpandedFst<Impl>(std::make_shared<Impl>()) {}
-
-  // Non-standard constructor to initialize NGramFst directly from data.
-  NGramFst(const char *data, bool owned)
-      : ImplToExpandedFst<Impl>(std::make_shared<Impl>()) {
-    GetMutableImpl()->Init(data, owned, nullptr);
-  }
-
-  // Get method that gets the data associated with Init().
-  const char *GetData(size_t *data_size) const {
-    return GetImpl()->GetData(data_size);
-  }
-
-  const std::vector<Label> GetContext(StateId s) const {
-    return GetImpl()->GetContext(s, &inst_);
-  }
-
-  // Consumes as much as possible of context from right to left, returns the
-  // the states corresponding to the increasingly conditioned input sequence.
-  void GetStates(const std::vector<Label> &context,
-                 std::vector<StateId> *state) const {
-    return GetImpl()->GetStates(context, state);
-  }
-
-  size_t NumArcs(StateId s) const override {
-    return GetImpl()->NumArcs(s, &inst_);
-  }
-
-  NGramFst<A> *Copy(bool safe = false) const override {
-    return new NGramFst(*this, safe);
-  }
-
-  static NGramFst<A> *Read(std::istream &strm, const FstReadOptions &opts) {
-    Impl *impl = Impl::Read(strm, opts);
-    return impl ? new NGramFst<A>(std::shared_ptr<Impl>(impl)) : nullptr;
-  }
-
-  static NGramFst<A> *Read(const string &filename) {
-    if (!filename.empty()) {
-      std::ifstream strm(filename,
-                              std::ios_base::in | std::ios_base::binary);
-      if (!strm.good()) {
-        LOG(ERROR) << "NGramFst::Read: Can't open file: " << filename;
-        return nullptr;
-      }
-      return Read(strm, FstReadOptions(filename));
-    } else {
-      return Read(std::cin, FstReadOptions("standard input"));
-    }
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return GetImpl()->Write(strm, opts);
-  }
-
-  bool Write(const string &filename) const override {
-    return Fst<A>::WriteFile(filename);
-  }
-
-  inline void InitStateIterator(StateIteratorData<A> *data) const override {
-    GetImpl()->InitStateIterator(data);
-  }
-
-  inline void InitArcIterator(StateId s,
-                              ArcIteratorData<A> *data) const override;
-
-  MatcherBase<A> *InitMatcher(MatchType match_type) const override {
-    return new NGramFstMatcher<A>(*this, match_type);
-  }
-
-  size_t StorageSize() const { return GetImpl()->StorageSize(); }
-
-  static bool HasRequiredProps(const Fst<A> &fst) {
-    int64 props =
-        kAcceptor | kIDeterministic | kILabelSorted | kIEpsilons | kAccessible;
-    return fst.Properties(props, true) == props;
-  }
-
-  static bool HasRequiredStructure(const Fst<A> &fst) {
-    if (!HasRequiredProps(fst)) {
-      return false;
-    }
-    typename A::StateId unigram = fst.Start();
-    while (true) {  // Follows epsilon arc chain to find unigram state.
-      if (unigram == fst::kNoStateId) return false;  // No unigram state.
-      typename fst::ArcIterator<Fst<A>> aiter(fst, unigram);
-      if (aiter.Done() || aiter.Value().ilabel != 0) break;
-      unigram = aiter.Value().nextstate;
-      aiter.Next();
-    }
-    // Other requirement: all states other than unigram an epsilon arc.
-    for (fst::StateIterator<Fst<A>> siter(fst); !siter.Done();
-         siter.Next()) {
-      const typename A::StateId &state = siter.Value();
-      fst::ArcIterator<Fst<A>> aiter(fst, state);
-      if (state != unigram) {
-        if (aiter.Done()) return false;
-        if (aiter.Value().ilabel != 0) return false;
-        aiter.Next();
-        if (!aiter.Done() && aiter.Value().ilabel == 0) return false;
-      }
-    }
-    return true;
-  }
-
- private:
-  using ImplToExpandedFst<Impl, ExpandedFst<A>>::GetImpl;
-  using ImplToExpandedFst<Impl, ExpandedFst<A>>::GetMutableImpl;
-
-  explicit NGramFst(std::shared_ptr<Impl> impl)
-      : ImplToExpandedFst<Impl>(impl) {}
-
-  mutable NGramFstInst<A> inst_;
-};
-
-template <class A>
-inline void NGramFst<A>::InitArcIterator(StateId s,
-                                         ArcIteratorData<A> *data) const {
-  GetImpl()->SetInstFuture(s, &inst_);
-  GetImpl()->SetInstNode(&inst_);
-  data->base = new ArcIterator<NGramFst<A>>(*this, s);
-}
-
-namespace internal {
-
-template <typename A>
-NGramFstImpl<A>::NGramFstImpl(const Fst<A> &fst,
-                              std::vector<StateId> *order_out) {
-  typedef A Arc;
-  typedef typename Arc::Label Label;
-  typedef typename Arc::Weight Weight;
-  typedef typename Arc::StateId StateId;
-  SetType("ngram");
-  SetInputSymbols(fst.InputSymbols());
-  SetOutputSymbols(fst.OutputSymbols());
-  SetProperties(kStaticProperties);
-
-  // Check basic requirements for an OpenGrm language model Fst.
-  if (!NGramFst<A>::HasRequiredProps(fst)) {
-    FSTERROR() << "NGramFst only accepts OpenGrm language models as input";
-    SetProperties(kError, kError);
-    return;
-  }
-
-  int64 num_states = CountStates(fst);
-  Label *context = new Label[num_states];
-
-  // Find the unigram state by starting from the start state, following
-  // epsilons.
-  StateId unigram = fst.Start();
-  while (1) {
-    if (unigram == kNoStateId) {
-      FSTERROR() << "Could not identify unigram state";
-      SetProperties(kError, kError);
-      return;
-    }
-    ArcIterator<Fst<A>> aiter(fst, unigram);
-    if (aiter.Done()) {
-      LOG(WARNING) << "Unigram state " << unigram << " has no arcs.";
-      break;
-    }
-    if (aiter.Value().ilabel != 0) break;
-    unigram = aiter.Value().nextstate;
-  }
-
-  // Each state's context is determined by the subtree it is under from the
-  // unigram state.
-  std::queue<std::pair<StateId, Label>> label_queue;
-  std::vector<bool> visited(num_states);
-  // Force an epsilon link to the start state.
-  label_queue.push(std::make_pair(fst.Start(), 0));
-  for (ArcIterator<Fst<A>> aiter(fst, unigram); !aiter.Done(); aiter.Next()) {
-    label_queue.push(
-        std::make_pair(aiter.Value().nextstate, aiter.Value().ilabel));
-  }
-  // investigate states in breadth first fashion to assign context words.
-  while (!label_queue.empty()) {
-    std::pair<StateId, Label> &now = label_queue.front();
-    if (!visited[now.first]) {
-      context[now.first] = now.second;
-      visited[now.first] = true;
-      for (ArcIterator<Fst<A>> aiter(fst, now.first); !aiter.Done();
-           aiter.Next()) {
-        const Arc &arc = aiter.Value();
-        if (arc.ilabel != 0) {
-          label_queue.push(std::make_pair(arc.nextstate, now.second));
-        }
-      }
-    }
-    label_queue.pop();
-  }
-  visited.clear();
-
-  // The arc from the start state should be assigned an epsilon to put it
-  // in front of the all other labels (which makes Start state 1 after
-  // unigram which is state 0).
-  context[fst.Start()] = 0;
-
-  // Build the tree of contexts fst by reversing the epsilon arcs from fst.
-  VectorFst<Arc> context_fst;
-  uint64 num_final = 0;
-  for (int i = 0; i < num_states; ++i) {
-    if (fst.Final(i) != Weight::Zero()) {
-      ++num_final;
-    }
-    context_fst.SetFinal(context_fst.AddState(), fst.Final(i));
-  }
-  context_fst.SetStart(unigram);
-  context_fst.SetInputSymbols(fst.InputSymbols());
-  context_fst.SetOutputSymbols(fst.OutputSymbols());
-  int64 num_context_arcs = 0;
-  int64 num_futures = 0;
-  for (StateIterator<Fst<A>> siter(fst); !siter.Done(); siter.Next()) {
-    const StateId &state = siter.Value();
-    num_futures += fst.NumArcs(state) - fst.NumInputEpsilons(state);
-    ArcIterator<Fst<A>> aiter(fst, state);
-    if (!aiter.Done()) {
-      const Arc &arc = aiter.Value();
-      // this arc goes from state to arc.nextstate, so create an arc from
-      // arc.nextstate to state to reverse it.
-      if (arc.ilabel == 0) {
-        context_fst.AddArc(arc.nextstate, Arc(context[state], context[state],
-                                              arc.weight, state));
-        num_context_arcs++;
-      }
-    }
-  }
-  if (num_context_arcs != context_fst.NumStates() - 1) {
-    FSTERROR() << "Number of contexts arcs != number of states - 1";
-    SetProperties(kError, kError);
-    return;
-  }
-  if (context_fst.NumStates() != num_states) {
-    FSTERROR() << "Number of contexts != number of states";
-    SetProperties(kError, kError);
-    return;
-  }
-  int64 context_props =
-      context_fst.Properties(kIDeterministic | kILabelSorted, true);
-  if (!(context_props & kIDeterministic)) {
-    FSTERROR() << "Input Fst is not structured properly";
-    SetProperties(kError, kError);
-    return;
-  }
-  if (!(context_props & kILabelSorted)) {
-    ArcSort(&context_fst, ILabelCompare<Arc>());
-  }
-
-  delete[] context;
-
-  uint64 b64;
-  Weight weight;
-  Label label = kNoLabel;
-  const size_t storage = Storage(num_states, num_futures, num_final);
-  MappedFile *data_region = MappedFile::Allocate(storage);
-  char *data = reinterpret_cast<char *>(data_region->mutable_data());
-  memset(data, 0, storage);
-  size_t offset = 0;
-  memcpy(data + offset, reinterpret_cast<char *>(&num_states),
-         sizeof(num_states));
-  offset += sizeof(num_states);
-  memcpy(data + offset, reinterpret_cast<char *>(&num_futures),
-         sizeof(num_futures));
-  offset += sizeof(num_futures);
-  memcpy(data + offset, reinterpret_cast<char *>(&num_final),
-         sizeof(num_final));
-  offset += sizeof(num_final);
-  uint64 *context_bits = reinterpret_cast<uint64 *>(data + offset);
-  offset += BitmapIndex::StorageSize(num_states * 2 + 1) * sizeof(b64);
-  uint64 *future_bits = reinterpret_cast<uint64 *>(data + offset);
-  offset +=
-      BitmapIndex::StorageSize(num_futures + num_states + 1) * sizeof(b64);
-  uint64 *final_bits = reinterpret_cast<uint64 *>(data + offset);
-  offset += BitmapIndex::StorageSize(num_states) * sizeof(b64);
-  Label *context_words = reinterpret_cast<Label *>(data + offset);
-  offset += (num_states + 1) * sizeof(label);
-  Label *future_words = reinterpret_cast<Label *>(data + offset);
-  offset += num_futures * sizeof(label);
-  offset = (offset + sizeof(weight) - 1) & ~(sizeof(weight) - 1);
-  Weight *backoff = reinterpret_cast<Weight *>(data + offset);
-  offset += (num_states + 1) * sizeof(weight);
-  Weight *final_probs = reinterpret_cast<Weight *>(data + offset);
-  offset += num_final * sizeof(weight);
-  Weight *future_probs = reinterpret_cast<Weight *>(data + offset);
-  int64 context_arc = 0, future_arc = 0, context_bit = 0, future_bit = 0,
-        final_bit = 0;
-
-  // pseudo-root bits
-  BitmapIndex::Set(context_bits, context_bit++);
-  ++context_bit;
-  context_words[context_arc] = label;
-  backoff[context_arc] = Weight::Zero();
-  context_arc++;
-
-  ++future_bit;
-  if (order_out) {
-    order_out->clear();
-    order_out->resize(num_states);
-  }
-
-  std::queue<StateId> context_q;
-  context_q.push(context_fst.Start());
-  StateId state_number = 0;
-  while (!context_q.empty()) {
-    const StateId &state = context_q.front();
-    if (order_out) {
-      (*order_out)[state] = state_number;
-    }
-
-    const Weight final_weight = context_fst.Final(state);
-    if (final_weight != Weight::Zero()) {
-      BitmapIndex::Set(final_bits, state_number);
-      final_probs[final_bit] = final_weight;
-      ++final_bit;
-    }
-
-    for (ArcIterator<VectorFst<A>> aiter(context_fst, state); !aiter.Done();
-         aiter.Next()) {
-      const Arc &arc = aiter.Value();
-      context_words[context_arc] = arc.ilabel;
-      backoff[context_arc] = arc.weight;
-      ++context_arc;
-      BitmapIndex::Set(context_bits, context_bit++);
-      context_q.push(arc.nextstate);
-    }
-    ++context_bit;
-
-    for (ArcIterator<Fst<A>> aiter(fst, state); !aiter.Done(); aiter.Next()) {
-      const Arc &arc = aiter.Value();
-      if (arc.ilabel != 0) {
-        future_words[future_arc] = arc.ilabel;
-        future_probs[future_arc] = arc.weight;
-        ++future_arc;
-        BitmapIndex::Set(future_bits, future_bit++);
-      }
-    }
-    ++future_bit;
-    ++state_number;
-    context_q.pop();
-  }
-
-  if ((state_number != num_states) || (context_bit != num_states * 2 + 1) ||
-      (context_arc != num_states) || (future_arc != num_futures) ||
-      (future_bit != num_futures + num_states + 1) ||
-      (final_bit != num_final)) {
-    FSTERROR() << "Structure problems detected during construction";
-    SetProperties(kError, kError);
-    return;
-  }
-
-  Init(data, false, data_region);
-}
-
-template <typename A>
-inline void NGramFstImpl<A>::Init(const char *data, bool owned,
-                                  MappedFile *data_region) {
-  if (owned_) {
-    delete[] data_;
-  }
-  data_region_.reset(data_region);
-  owned_ = owned;
-  data_ = data;
-  size_t offset = 0;
-  num_states_ = *(reinterpret_cast<const uint64 *>(data_ + offset));
-  offset += sizeof(num_states_);
-  num_futures_ = *(reinterpret_cast<const uint64 *>(data_ + offset));
-  offset += sizeof(num_futures_);
-  num_final_ = *(reinterpret_cast<const uint64 *>(data_ + offset));
-  offset += sizeof(num_final_);
-  uint64 bits;
-  size_t context_bits = num_states_ * 2 + 1;
-  size_t future_bits = num_futures_ + num_states_ + 1;
-  context_ = reinterpret_cast<const uint64 *>(data_ + offset);
-  offset += BitmapIndex::StorageSize(context_bits) * sizeof(bits);
-  future_ = reinterpret_cast<const uint64 *>(data_ + offset);
-  offset += BitmapIndex::StorageSize(future_bits) * sizeof(bits);
-  final_ = reinterpret_cast<const uint64 *>(data_ + offset);
-  offset += BitmapIndex::StorageSize(num_states_) * sizeof(bits);
-  context_words_ = reinterpret_cast<const Label *>(data_ + offset);
-  offset += (num_states_ + 1) * sizeof(*context_words_);
-  future_words_ = reinterpret_cast<const Label *>(data_ + offset);
-  offset += num_futures_ * sizeof(*future_words_);
-  offset = (offset + sizeof(*backoff_) - 1) & ~(sizeof(*backoff_) - 1);
-  backoff_ = reinterpret_cast<const Weight *>(data_ + offset);
-  offset += (num_states_ + 1) * sizeof(*backoff_);
-  final_probs_ = reinterpret_cast<const Weight *>(data_ + offset);
-  offset += num_final_ * sizeof(*final_probs_);
-  future_probs_ = reinterpret_cast<const Weight *>(data_ + offset);
-
-  context_index_.BuildIndex(context_, context_bits);
-  future_index_.BuildIndex(future_, future_bits);
-  final_index_.BuildIndex(final_, num_states_);
-
-  select_root_ = context_index_.Select0s(0);
-  if (context_index_.Rank1(0) != 0 || select_root_.first != 1 ||
-      context_index_.Get(2) == false) {
-    FSTERROR() << "Malformed file";
-    SetProperties(kError, kError);
-    return;
-  }
-  root_children_ = context_words_ + context_index_.Rank1(2);
-  start_ = 1;
-}
-
-template <typename A>
-inline typename A::StateId NGramFstImpl<A>::Transition(
-    const std::vector<Label> &context, Label future) const {
-  const Label *children = root_children_;
-  size_t num_children = select_root_.second - 2;
-  const Label *loc =
-      std::lower_bound(children, children + num_children, future);
-  if (loc == children + num_children || *loc != future) {
-    return context_index_.Rank1(0);
-  }
-  size_t node = 2 + loc - children;
-  size_t node_rank = context_index_.Rank1(node);
-  std::pair<size_t, size_t> zeros =
-      (node_rank == 0) ? select_root_ : context_index_.Select0s(node_rank);
-  size_t first_child = zeros.first + 1;
-  if (context_index_.Get(first_child) == false) {
-    return context_index_.Rank1(node);
-  }
-  size_t last_child = zeros.second - 1;
-  for (int word = context.size() - 1; word >= 0; --word) {
-    children = context_words_ + context_index_.Rank1(first_child);
-    loc = std::lower_bound(children, children + last_child - first_child + 1,
-                           context[word]);
-    if (loc == children + last_child - first_child + 1 ||
-        *loc != context[word]) {
-      break;
-    }
-    node = first_child + loc - children;
-    node_rank = context_index_.Rank1(node);
-    zeros =
-        (node_rank == 0) ? select_root_ : context_index_.Select0s(node_rank);
-    first_child = zeros.first + 1;
-    if (context_index_.Get(first_child) == false) break;
-    last_child = zeros.second - 1;
-  }
-  return context_index_.Rank1(node);
-}
-
-}  // namespace internal
-
-/*****************************************************************************/
-template <class A>
-class NGramFstMatcher : public MatcherBase<A> {
- public:
-  typedef A Arc;
-  typedef typename A::Label Label;
-  typedef typename A::StateId StateId;
-  typedef typename A::Weight Weight;
-
-  NGramFstMatcher(const NGramFst<A> &fst, MatchType match_type)
-      : fst_(fst),
-        inst_(fst.inst_),
-        match_type_(match_type),
-        current_loop_(false),
-        loop_(kNoLabel, 0, A::Weight::One(), kNoStateId) {
-    if (match_type_ == MATCH_OUTPUT) {
-      std::swap(loop_.ilabel, loop_.olabel);
-    }
-  }
-
-  NGramFstMatcher(const NGramFstMatcher<A> &matcher, bool safe = false)
-      : fst_(matcher.fst_),
-        inst_(matcher.inst_),
-        match_type_(matcher.match_type_),
-        current_loop_(false),
-        loop_(kNoLabel, 0, A::Weight::One(), kNoStateId) {
-    if (match_type_ == MATCH_OUTPUT) {
-      std::swap(loop_.ilabel, loop_.olabel);
-    }
-  }
-
-  NGramFstMatcher<A> *Copy(bool safe = false) const override {
-    return new NGramFstMatcher<A>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override { return match_type_; }
-
-  const Fst<A> &GetFst() const override { return fst_; }
-
-  uint64 Properties(uint64 props) const override { return props; }
-
-  void SetState(StateId s) final {
-    fst_.GetImpl()->SetInstFuture(s, &inst_);
-    current_loop_ = false;
-  }
-
-  bool Find(Label label) final {
-    const Label nolabel = kNoLabel;
-    done_ = true;
-    if (label == 0 || label == nolabel) {
-      if (label == 0) {
-        current_loop_ = true;
-        loop_.nextstate = inst_.state_;
-      }
-      // The unigram state has no epsilon arc.
-      if (inst_.state_ != 0) {
-        arc_.ilabel = arc_.olabel = 0;
-        fst_.GetImpl()->SetInstNode(&inst_);
-        arc_.nextstate = fst_.GetImpl()->context_index_.Rank1(
-            fst_.GetImpl()->context_index_.Select1(
-                fst_.GetImpl()->context_index_.Rank0(inst_.node_) - 1));
-        arc_.weight = fst_.GetImpl()->backoff_[inst_.state_];
-        done_ = false;
-      }
-    } else {
-      current_loop_ = false;
-      const Label *start = fst_.GetImpl()->future_words_ + inst_.offset_;
-      const Label *end = start + inst_.num_futures_;
-      const Label *search = std::lower_bound(start, end, label);
-      if (search != end && *search == label) {
-        size_t state = search - start;
-        arc_.ilabel = arc_.olabel = label;
-        arc_.weight = fst_.GetImpl()->future_probs_[inst_.offset_ + state];
-        fst_.GetImpl()->SetInstContext(&inst_);
-        arc_.nextstate = fst_.GetImpl()->Transition(inst_.context_, label);
-        done_ = false;
-      }
-    }
-    return !Done();
-  }
-
-  bool Done() const final { return !current_loop_ && done_; }
-
-  const Arc &Value() const final { return (current_loop_) ? loop_ : arc_; }
-
-  void Next() final {
-    if (current_loop_) {
-      current_loop_ = false;
-    } else {
-      done_ = true;
-    }
-  }
-
-  ssize_t Priority(StateId s) final { return fst_.NumArcs(s); }
-
- private:
-  const NGramFst<A> &fst_;
-  NGramFstInst<A> inst_;
-  MatchType match_type_;  // Supplied by caller
-  bool done_;
-  Arc arc_;
-  bool current_loop_;  // Current arc is the implicit loop
-  Arc loop_;
-};
-
-/*****************************************************************************/
-// Specialization for NGramFst; see generic version in fst.h
-// for sample usage (but use the ProdLmFst type!). This version
-// should inline.
-template <class A>
-class StateIterator<NGramFst<A>> : public StateIteratorBase<A> {
- public:
-  typedef typename A::StateId StateId;
-
-  explicit StateIterator(const NGramFst<A> &fst)
-      : s_(0), num_states_(fst.NumStates()) {}
-
-  bool Done() const final { return s_ >= num_states_; }
-
-  StateId Value() const final { return s_; }
-
-  void Next() final { ++s_; }
-
-  void Reset() final { s_ = 0; }
-
- private:
-  StateId s_;
-  StateId num_states_;
-};
-
-/*****************************************************************************/
-template <class A>
-class ArcIterator<NGramFst<A>> : public ArcIteratorBase<A> {
- public:
-  typedef A Arc;
-  typedef typename A::Label Label;
-  typedef typename A::StateId StateId;
-  typedef typename A::Weight Weight;
-
-  ArcIterator(const NGramFst<A> &fst, StateId state)
-      : lazy_(~0), impl_(fst.GetImpl()), i_(0), flags_(kArcValueFlags) {
-    inst_ = fst.inst_;
-    impl_->SetInstFuture(state, &inst_);
-    impl_->SetInstNode(&inst_);
-  }
-
-  bool Done() const final {
-    return i_ >=
-           ((inst_.node_ == 0) ? inst_.num_futures_ : inst_.num_futures_ + 1);
-  }
-
-  const Arc &Value() const final {
-    bool eps = (inst_.node_ != 0 && i_ == 0);
-    StateId state = (inst_.node_ == 0) ? i_ : i_ - 1;
-    if (flags_ & lazy_ & (kArcILabelValue | kArcOLabelValue)) {
-      arc_.ilabel = arc_.olabel =
-          eps ? 0 : impl_->future_words_[inst_.offset_ + state];
-      lazy_ &= ~(kArcILabelValue | kArcOLabelValue);
-    }
-    if (flags_ & lazy_ & kArcNextStateValue) {
-      if (eps) {
-        arc_.nextstate =
-            impl_->context_index_.Rank1(impl_->context_index_.Select1(
-                impl_->context_index_.Rank0(inst_.node_) - 1));
-      } else {
-        if (lazy_ & kArcNextStateValue) {
-          impl_->SetInstContext(&inst_);  // first time only.
-        }
-        arc_.nextstate = impl_->Transition(
-            inst_.context_, impl_->future_words_[inst_.offset_ + state]);
-      }
-      lazy_ &= ~kArcNextStateValue;
-    }
-    if (flags_ & lazy_ & kArcWeightValue) {
-      arc_.weight = eps ? impl_->backoff_[inst_.state_]
-                        : impl_->future_probs_[inst_.offset_ + state];
-      lazy_ &= ~kArcWeightValue;
-    }
-    return arc_;
-  }
-
-  void Next() final {
-    ++i_;
-    lazy_ = ~0;
-  }
-
-  size_t Position() const final { return i_; }
-
-  void Reset() final {
-    i_ = 0;
-    lazy_ = ~0;
-  }
-
-  void Seek(size_t a) final {
-    if (i_ != a) {
-      i_ = a;
-      lazy_ = ~0;
-    }
-  }
-
-  uint32 Flags() const final { return flags_; }
-
-  void SetFlags(uint32 flags, uint32 mask) final {
-    flags_ &= ~mask;
-    flags_ |= (flags & kArcValueFlags);
-  }
-
- private:
-  mutable Arc arc_;
-  mutable uint32 lazy_;
-  const internal::NGramFstImpl<A> *impl_;  // Borrowed reference.
-  mutable NGramFstInst<A> inst_;
-
-  size_t i_;
-  uint32 flags_;
-};
-
-}  // namespace fst
-#endif  // FST_EXTENSIONS_NGRAM_NGRAM_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/ngram/nthbit.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/ngram/nthbit.h
deleted file mode 100644
index 8d28ae6c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/ngram/nthbit.h
+++ /dev/null
@@ -1,33 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_NGRAM_NTHBIT_H_
-#define FST_EXTENSIONS_NGRAM_NTHBIT_H_
-
-#include <fst/types.h>
-#include <fst/compat.h>
-
-extern uint32 nth_bit_bit_offset[];
-
-inline uint32 nth_bit(uint64 v, uint32 r) {
-  uint32 shift = 0;
-  uint32 c = __builtin_popcount(v & 0xffffffff);
-  uint32 mask = -(r > c);
-  r -= c & mask;
-  shift += (32 & mask);
-
-  c = __builtin_popcount((v >> shift) & 0xffff);
-  mask = -(r > c);
-  r -= c & mask;
-  shift += (16 & mask);
-
-  c = __builtin_popcount((v >> shift) & 0xff);
-  mask = -(r > c);
-  r -= c & mask;
-  shift += (8 & mask);
-
-  return shift +
-         ((nth_bit_bit_offset[(v >> shift) & 0xff] >> ((r - 1) << 2)) & 0xf);
-}
-
-#endif  // FST_EXTENSIONS_NGRAM_NTHBIT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/collection.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/collection.h
deleted file mode 100644
index ae34aba7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/collection.h
+++ /dev/null
@@ -1,107 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to store a collection of ordered (multi-)sets with elements of type T.
-
-#ifndef FST_EXTENSIONS_PDT_COLLECTION_H_
-#define FST_EXTENSIONS_PDT_COLLECTION_H_
-
-#include <functional>
-#include <vector>
-
-#include <fst/log.h>
-#include <fst/bi-table.h>
-
-namespace fst {
-
-// Stores a collection of non-empty, ordered (multi-)sets with elements of type
-// T. A default constructor, operator==, and an STL-style hash functor must be
-// defined on the elements. Provides signed integer ID (of type I) for each
-// unique set. The IDs are allocated starting from 0 in order.
-template <class I, class T>
-class Collection {
- public:
-  struct Node {  // Trie node.
-    I node_id;   // Root is kNoNodeId;
-    T element;
-
-    Node() : node_id(kNoNodeId), element(T()) {}
-
-    Node(I i, const T &t) : node_id(i), element(t) {}
-
-    bool operator==(const Node &n) const {
-      return n.node_id == node_id && n.element == element;
-    }
-  };
-
-  struct NodeHash {
-    size_t operator()(const Node &n) const {
-      static constexpr auto kPrime = 7853;
-      return n.node_id + hash_(n.element) * kPrime;
-    }
-  };
-
-  using NodeTable = CompactHashBiTable<I, Node, NodeHash>;
-
-  class SetIterator {
-   public:
-    SetIterator(I id, Node node, NodeTable *node_table)
-        : id_(id), node_(node), node_table_(node_table) {}
-
-    bool Done() const { return id_ == kNoNodeId; }
-
-    const T &Element() const { return node_.element; }
-
-    void Next() {
-      id_ = node_.node_id;
-      if (id_ != kNoNodeId) node_ = node_table_->FindEntry(id_);
-    }
-
-   private:
-    I id_;       // Iterator set node ID.
-    Node node_;  // Iterator set node.
-    NodeTable *node_table_;
-  };
-
-  Collection() {}
-
-  // Looks up integer ID from ordered multi-se, and if it doesn't exist and
-  // insert is true, then adds it. Otherwise returns -1.
-  I FindId(const std::vector<T> &set, bool insert = true) {
-    I node_id = kNoNodeId;
-    for (ssize_t i = set.size() - 1; i >= 0; --i) {
-      Node node(node_id, set[i]);
-      node_id = node_table_.FindId(node, insert);
-      if (node_id == -1) break;
-    }
-    return node_id;
-  }
-
-  // Finds ordered (multi-)set given integer ID. Returns set iterator to
-  // traverse result.
-  SetIterator FindSet(I id) {
-    if (id < 0 || id >= node_table_.Size()) {
-      return SetIterator(kNoNodeId, Node(kNoNodeId, T()), &node_table_);
-    } else {
-      return SetIterator(id, node_table_.FindEntry(id), &node_table_);
-    }
-  }
-
-  I Size() const { return node_table_.Size(); }
-
- private:
-  static constexpr I kNoNodeId = -1;
-  static const std::hash<T> hash_;
-
-  NodeTable node_table_;
-};
-
-template <class I, class T>
-constexpr I Collection<I, T>::kNoNodeId;
-
-template <class I, class T>
-const std::hash<T> Collection<I, T>::hash_ = {};
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_COLLECTION_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/compose.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/compose.h
deleted file mode 100644
index c1352b52..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/compose.h
+++ /dev/null
@@ -1,476 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Composes a PDT and an FST.
-
-#ifndef FST_EXTENSIONS_PDT_COMPOSE_H_
-#define FST_EXTENSIONS_PDT_COMPOSE_H_
-
-#include <list>
-
-#include <fst/extensions/pdt/pdt.h>
-#include <fst/compose.h>
-
-namespace fst {
-
-// Returns paren arcs for Find(kNoLabel).
-constexpr uint32 kParenList = 0x00000001;
-
-// Returns a kNolabel loop for Find(paren).
-constexpr uint32 kParenLoop = 0x00000002;
-
-// This class is a matcher that treats parens as multi-epsilon labels.
-// It is most efficient if the parens are in a range non-overlapping with
-// the non-paren labels.
-template <class F>
-class ParenMatcher {
- public:
-  using FST = F;
-  using M = SortedMatcher<FST>;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  ParenMatcher(const FST &fst, MatchType match_type,
-               uint32 flags = (kParenLoop | kParenList))
-      : matcher_(fst, match_type), match_type_(match_type), flags_(flags) {
-    if (match_type == MATCH_INPUT) {
-      loop_.ilabel = kNoLabel;
-      loop_.olabel = 0;
-    } else {
-      loop_.ilabel = 0;
-      loop_.olabel = kNoLabel;
-    }
-    loop_.weight = Weight::One();
-    loop_.nextstate = kNoStateId;
-  }
-
-  ParenMatcher(const ParenMatcher<FST> &matcher, bool safe = false)
-      : matcher_(matcher.matcher_, safe),
-        match_type_(matcher.match_type_),
-        flags_(matcher.flags_),
-        open_parens_(matcher.open_parens_),
-        close_parens_(matcher.close_parens_),
-        loop_(matcher.loop_) {
-    loop_.nextstate = kNoStateId;
-  }
-
-  ParenMatcher<FST> *Copy(bool safe = false) const {
-    return new ParenMatcher<FST>(*this, safe);
-  }
-
-  MatchType Type(bool test) const { return matcher_.Type(test); }
-
-  void SetState(StateId s) {
-    matcher_.SetState(s);
-    loop_.nextstate = s;
-  }
-
-  bool Find(Label match_label);
-
-  bool Done() const { return done_; }
-
-  const Arc &Value() const { return paren_loop_ ? loop_ : matcher_.Value(); }
-
-  void Next();
-
-  Weight Final(StateId s) { return matcher_.Final(s); }
-
-  ssize_t Priority(StateId s) { return matcher_.Priority(s); }
-
-  const FST &GetFst() const { return matcher_.GetFst(); }
-
-  uint64 Properties(uint64 props) const { return matcher_.Properties(props); }
-
-  uint32 Flags() const { return matcher_.Flags(); }
-
-  void AddOpenParen(Label label) {
-    if (label == 0) {
-      FSTERROR() << "ParenMatcher: Bad open paren label: 0";
-    } else {
-      open_parens_.Insert(label);
-    }
-  }
-
-  void AddCloseParen(Label label) {
-    if (label == 0) {
-      FSTERROR() << "ParenMatcher: Bad close paren label: 0";
-    } else {
-      close_parens_.Insert(label);
-    }
-  }
-
-  void RemoveOpenParen(Label label) {
-    if (label == 0) {
-      FSTERROR() << "ParenMatcher: Bad open paren label: 0";
-    } else {
-      open_parens_.Erase(label);
-    }
-  }
-
-  void RemoveCloseParen(Label label) {
-    if (label == 0) {
-      FSTERROR() << "ParenMatcher: Bad close paren label: 0";
-    } else {
-      close_parens_.Erase(label);
-    }
-  }
-
-  void ClearOpenParens() { open_parens_.Clear(); }
-
-  void ClearCloseParens() { close_parens_.Clear(); }
-
-  bool IsOpenParen(Label label) const { return open_parens_.Member(label); }
-
-  bool IsCloseParen(Label label) const { return close_parens_.Member(label); }
-
- private:
-  // Advances matcher to next open paren, returning true if it exists.
-  bool NextOpenParen();
-
-  // Advances matcher to next close paren, returning true if it exists.
-  bool NextCloseParen();
-
-  M matcher_;
-  MatchType match_type_;  // Type of match to perform.
-  uint32 flags_;
-  // Open paren label set.
-  CompactSet<Label, kNoLabel> open_parens_;
-  // Close paren label set.
-  CompactSet<Label, kNoLabel> close_parens_;
-  bool open_paren_list_;   // Matching open paren list?
-  bool close_paren_list_;  // Matching close paren list?
-  bool paren_loop_;        // Current arc is the implicit paren loop?
-  mutable Arc loop_;       // For non-consuming symbols.
-  bool done_;              // Matching done?
-
-  ParenMatcher &operator=(const ParenMatcher &) = delete;
-};
-
-template <class FST>
-inline bool ParenMatcher<FST>::Find(Label match_label) {
-  open_paren_list_ = false;
-  close_paren_list_ = false;
-  paren_loop_ = false;
-  done_ = false;
-  // Returns all parenthesis arcs.
-  if (match_label == kNoLabel && (flags_ & kParenList)) {
-    if (open_parens_.LowerBound() != kNoLabel) {
-      matcher_.LowerBound(open_parens_.LowerBound());
-      open_paren_list_ = NextOpenParen();
-      if (open_paren_list_) return true;
-    }
-    if (close_parens_.LowerBound() != kNoLabel) {
-      matcher_.LowerBound(close_parens_.LowerBound());
-      close_paren_list_ = NextCloseParen();
-      if (close_paren_list_) return true;
-    }
-  }
-  // Returns the implicit paren loop.
-  if (match_label > 0 && (flags_ & kParenLoop) &&
-      (IsOpenParen(match_label) || IsCloseParen(match_label))) {
-    paren_loop_ = true;
-    return true;
-  }
-  // Returns all other labels.
-  if (matcher_.Find(match_label)) return true;
-  done_ = true;
-  return false;
-}
-
-template <class FST>
-inline void ParenMatcher<FST>::Next() {
-  if (paren_loop_) {
-    paren_loop_ = false;
-    done_ = true;
-  } else if (open_paren_list_) {
-    matcher_.Next();
-    open_paren_list_ = NextOpenParen();
-    if (open_paren_list_) return;
-    if (close_parens_.LowerBound() != kNoLabel) {
-      matcher_.LowerBound(close_parens_.LowerBound());
-      close_paren_list_ = NextCloseParen();
-      if (close_paren_list_) return;
-    }
-    done_ = !matcher_.Find(kNoLabel);
-  } else if (close_paren_list_) {
-    matcher_.Next();
-    close_paren_list_ = NextCloseParen();
-    if (close_paren_list_) return;
-    done_ = !matcher_.Find(kNoLabel);
-  } else {
-    matcher_.Next();
-    done_ = matcher_.Done();
-  }
-}
-
-// Advances matcher to next open paren, returning true if it exists.
-template <class FST>
-inline bool ParenMatcher<FST>::NextOpenParen() {
-  for (; !matcher_.Done(); matcher_.Next()) {
-    Label label = match_type_ == MATCH_INPUT ? matcher_.Value().ilabel
-                                             : matcher_.Value().olabel;
-    if (label > open_parens_.UpperBound()) return false;
-    if (IsOpenParen(label)) return true;
-  }
-  return false;
-}
-
-// Advances matcher to next close paren, returning true if it exists.
-template <class FST>
-inline bool ParenMatcher<FST>::NextCloseParen() {
-  for (; !matcher_.Done(); matcher_.Next()) {
-    Label label = match_type_ == MATCH_INPUT ? matcher_.Value().ilabel
-                                             : matcher_.Value().olabel;
-    if (label > close_parens_.UpperBound()) return false;
-    if (IsCloseParen(label)) return true;
-  }
-  return false;
-}
-
-template <class Filter>
-class ParenFilter {
- public:
-  using FST1 = typename Filter::FST1;
-  using FST2 = typename Filter::FST2;
-  using Arc = typename Filter::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Matcher1 = typename Filter::Matcher1;
-  using Matcher2 = typename Filter::Matcher2;
-
-  using StackId = StateId;
-  using ParenStack = PdtStack<StackId, Label>;
-  using FilterState1 = typename Filter::FilterState;
-  using FilterState2 = IntegerFilterState<StackId>;
-  using FilterState = PairFilterState<FilterState1, FilterState2>;
-
-  ParenFilter(const FST1 &fst1, const FST2 &fst2, Matcher1 *matcher1 = nullptr,
-              Matcher2 *matcher2 = nullptr,
-              const std::vector<std::pair<Label, Label>> *parens = nullptr,
-              bool expand = false, bool keep_parens = true)
-      : filter_(fst1, fst2, matcher1, matcher2),
-        parens_(parens ? *parens : std::vector<std::pair<Label, Label>>()),
-        expand_(expand),
-        keep_parens_(keep_parens),
-        fs_(FilterState::NoState()),
-        stack_(parens_),
-        paren_id_(-1) {
-    if (parens) {
-      for (const auto &pair : *parens) {
-        parens_.push_back(pair);
-        GetMatcher1()->AddOpenParen(pair.first);
-        GetMatcher2()->AddOpenParen(pair.first);
-        if (!expand_) {
-          GetMatcher1()->AddCloseParen(pair.second);
-          GetMatcher2()->AddCloseParen(pair.second);
-        }
-      }
-    }
-  }
-
-  ParenFilter(const ParenFilter &filter, bool safe = false)
-      : filter_(filter.filter_, safe),
-        parens_(filter.parens_),
-        expand_(filter.expand_),
-        keep_parens_(filter.keep_parens_),
-        fs_(FilterState::NoState()),
-        stack_(filter.parens_),
-        paren_id_(-1) {}
-
-  FilterState Start() const {
-    return FilterState(filter_.Start(), FilterState2(0));
-  }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    fs_ = fs;
-    filter_.SetState(s1, s2, fs_.GetState1());
-    if (!expand_) return;
-    ssize_t paren_id = stack_.Top(fs.GetState2().GetState());
-    if (paren_id != paren_id_) {
-      if (paren_id_ != -1) {
-        GetMatcher1()->RemoveCloseParen(parens_[paren_id_].second);
-        GetMatcher2()->RemoveCloseParen(parens_[paren_id_].second);
-      }
-      paren_id_ = paren_id;
-      if (paren_id_ != -1) {
-        GetMatcher1()->AddCloseParen(parens_[paren_id_].second);
-        GetMatcher2()->AddCloseParen(parens_[paren_id_].second);
-      }
-    }
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    const auto fs1 = filter_.FilterArc(arc1, arc2);
-    const auto &fs2 = fs_.GetState2();
-    if (fs1 == FilterState1::NoState()) return FilterState::NoState();
-    if (arc1->olabel == kNoLabel && arc2->ilabel) {  // arc2 parentheses.
-      if (keep_parens_) {
-        arc1->ilabel = arc2->ilabel;
-      } else if (arc2->ilabel) {
-        arc2->olabel = arc1->ilabel;
-      }
-      return FilterParen(arc2->ilabel, fs1, fs2);
-    } else if (arc2->ilabel == kNoLabel && arc1->olabel) {  // arc1 parentheses.
-      if (keep_parens_) {
-        arc2->olabel = arc1->olabel;
-      } else {
-        arc1->ilabel = arc2->olabel;
-      }
-      return FilterParen(arc1->olabel, fs1, fs2);
-    } else {
-      return FilterState(fs1, fs2);
-    }
-  }
-
-  void FilterFinal(Weight *w1, Weight *w2) const {
-    if (fs_.GetState2().GetState() != 0) *w1 = Weight::Zero();
-    filter_.FilterFinal(w1, w2);
-  }
-
-  // Returns respective matchers; ownership stays with filter.
-
-  Matcher1 *GetMatcher1() { return filter_.GetMatcher1(); }
-
-  Matcher2 *GetMatcher2() { return filter_.GetMatcher2(); }
-
-  uint64 Properties(uint64 iprops) const {
-    return filter_.Properties(iprops) & kILabelInvariantProperties &
-           kOLabelInvariantProperties;
-  }
-
- private:
-  const FilterState FilterParen(Label label, const FilterState1 &fs1,
-                                const FilterState2 &fs2) const {
-    if (!expand_) return FilterState(fs1, fs2);
-    const auto stack_id = stack_.Find(fs2.GetState(), label);
-    if (stack_id < 0) {
-      return FilterState::NoState();
-    } else {
-      return FilterState(fs1, FilterState2(stack_id));
-    }
-  }
-
-  Filter filter_;
-  std::vector<std::pair<Label, Label>> parens_;
-  bool expand_;       // Expands to FST?
-  bool keep_parens_;  // Retains parentheses in output?
-  FilterState fs_;    // Current filter state.
-  mutable ParenStack stack_;
-  ssize_t paren_id_;
-};
-
-// Class to setup composition options for PDT composition. Default is to take
-// the PDT as the first composition argument.
-template <class Arc, bool left_pdt = true>
-class PdtComposeFstOptions
-    : public ComposeFstOptions<
-          Arc, ParenMatcher<Fst<Arc>>,
-          ParenFilter<AltSequenceComposeFilter<ParenMatcher<Fst<Arc>>>>> {
- public:
-  using Label = typename Arc::Label;
-  using PdtMatcher = ParenMatcher<Fst<Arc>>;
-  using PdtFilter = ParenFilter<AltSequenceComposeFilter<PdtMatcher>>;
-
-  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::matcher1;
-  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::matcher2;
-  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::filter;
-
-  PdtComposeFstOptions(const Fst<Arc> &ifst1,
-                       const std::vector<std::pair<Label, Label>> &parens,
-                       const Fst<Arc> &ifst2, bool expand = false,
-                       bool keep_parens = true) {
-    matcher1 = new PdtMatcher(ifst1, MATCH_OUTPUT, kParenList);
-    matcher2 = new PdtMatcher(ifst2, MATCH_INPUT, kParenLoop);
-    filter = new PdtFilter(ifst1, ifst2, matcher1, matcher2, &parens, expand,
-                           keep_parens);
-  }
-};
-
-// Class to setup composition options for PDT with FST composition.
-// Specialization is for the FST as the first composition argument.
-template <class Arc>
-class PdtComposeFstOptions<Arc, false>
-    : public ComposeFstOptions<
-          Arc, ParenMatcher<Fst<Arc>>,
-          ParenFilter<SequenceComposeFilter<ParenMatcher<Fst<Arc>>>>> {
- public:
-  using Label = typename Arc::Label;
-  using PdtMatcher = ParenMatcher<Fst<Arc>>;
-  using PdtFilter = ParenFilter<SequenceComposeFilter<PdtMatcher>>;
-
-  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::matcher1;
-  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::matcher2;
-  using ComposeFstOptions<Arc, PdtMatcher, PdtFilter>::filter;
-
-  PdtComposeFstOptions(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-                       const std::vector<std::pair<Label, Label>> &parens,
-                       bool expand = false, bool keep_parens = true) {
-    matcher1 = new PdtMatcher(ifst1, MATCH_OUTPUT, kParenLoop);
-    matcher2 = new PdtMatcher(ifst2, MATCH_INPUT, kParenList);
-    filter = new PdtFilter(ifst1, ifst2, matcher1, matcher2, &parens, expand,
-                           keep_parens);
-  }
-};
-
-enum PdtComposeFilter {
-  PAREN_FILTER,         // Bar-Hillel construction; keeps parentheses.
-  EXPAND_FILTER,        // Bar-Hillel + expansion; removes parentheses.
-  EXPAND_PAREN_FILTER,  // Bar-Hillel + expansion; keeps parentheses.
-};
-
-struct PdtComposeOptions {
-  bool connect;                  // Connect output?
-  PdtComposeFilter filter_type;  // Pre-defined filter to use.
-
-  explicit PdtComposeOptions(bool connect = true,
-                             PdtComposeFilter filter_type = PAREN_FILTER)
-      : connect(connect), filter_type(filter_type) {}
-};
-
-// Composes pushdown transducer (PDT) encoded as an FST (1st arg) and an FST
-// (2nd arg) with the result also a PDT encoded as an FST (3rd arg). In the
-// PDTs, some transitions are labeled with open or close parentheses. To be
-// interpreted as a PDT, the parens must balance on a path (see PdtExpand()).
-// The open-close parenthesis label pairs are passed using the parens argument.
-template <class Arc>
-void Compose(const Fst<Arc> &ifst1,
-             const std::vector<
-                 std::pair<typename Arc::Label, typename Arc::Label>> &parens,
-             const Fst<Arc> &ifst2, MutableFst<Arc> *ofst,
-             const PdtComposeOptions &opts = PdtComposeOptions()) {
-  bool expand = opts.filter_type != PAREN_FILTER;
-  bool keep_parens = opts.filter_type != EXPAND_FILTER;
-  PdtComposeFstOptions<Arc, true> copts(ifst1, parens, ifst2, expand,
-                                        keep_parens);
-  copts.gc_limit = 0;
-  *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-  if (opts.connect) Connect(ofst);
-}
-
-// Composes an FST (1st arg) and pushdown transducer (PDT) encoded as an FST
-// (2nd arg) with the result also a PDT encoded as an FST (3rd arg). In the
-// PDTs, some transitions are labeled with open or close parentheses. To be
-// interpreted as a PDT, the parens must balance on a path (see ExpandFst()).
-// The open-close parenthesis label pairs are passed using the parens argument.
-template <class Arc>
-void Compose(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-             const std::vector<
-                 std::pair<typename Arc::Label, typename Arc::Label>> &parens,
-             MutableFst<Arc> *ofst,
-             const PdtComposeOptions &opts = PdtComposeOptions()) {
-  bool expand = opts.filter_type != PAREN_FILTER;
-  bool keep_parens = opts.filter_type != EXPAND_FILTER;
-  PdtComposeFstOptions<Arc, false> copts(ifst1, ifst2, parens, expand,
-                                         keep_parens);
-  copts.gc_limit = 0;
-  *ofst = ComposeFst<Arc>(ifst1, ifst2, copts);
-  if (opts.connect) Connect(ofst);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_COMPOSE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/expand.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/expand.h
deleted file mode 100644
index eeee781e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/expand.h
+++ /dev/null
@@ -1,933 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Expands a PDT to an FST.
-
-#ifndef FST_EXTENSIONS_PDT_EXPAND_H_
-#define FST_EXTENSIONS_PDT_EXPAND_H_
-
-#include <forward_list>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/extensions/pdt/paren.h>
-#include <fst/extensions/pdt/pdt.h>
-#include <fst/extensions/pdt/reverse.h>
-#include <fst/extensions/pdt/shortest-path.h>
-#include <fst/cache.h>
-#include <fst/mutable-fst.h>
-#include <fst/queue.h>
-#include <fst/state-table.h>
-#include <fst/test-properties.h>
-
-namespace fst {
-
-template <class Arc>
-struct PdtExpandFstOptions : public CacheOptions {
-  bool keep_parentheses;
-  PdtStack<typename Arc::StateId, typename Arc::Label> *stack;
-  PdtStateTable<typename Arc::StateId, typename Arc::StateId> *state_table;
-
-  explicit PdtExpandFstOptions(
-      const CacheOptions &opts = CacheOptions(), bool keep_parentheses = false,
-      PdtStack<typename Arc::StateId, typename Arc::Label> *stack = nullptr,
-      PdtStateTable<typename Arc::StateId, typename Arc::StateId> *state_table =
-          nullptr)
-      : CacheOptions(opts),
-        keep_parentheses(keep_parentheses),
-        stack(stack),
-        state_table(state_table) {}
-};
-
-namespace internal {
-
-// Implementation class for PdtExpandFst.
-template <class Arc>
-class PdtExpandFstImpl : public CacheImpl<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using StackId = StateId;
-  using StateTuple = PdtStateTuple<StateId, StackId>;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::Properties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheBaseImpl<CacheState<Arc>>::PushArc;
-  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
-  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
-  using CacheBaseImpl<CacheState<Arc>>::HasStart;
-  using CacheBaseImpl<CacheState<Arc>>::SetArcs;
-  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
-  using CacheBaseImpl<CacheState<Arc>>::SetStart;
-
-  PdtExpandFstImpl(const Fst<Arc> &fst,
-                   const std::vector<std::pair<Label, Label>> &parens,
-                   const PdtExpandFstOptions<Arc> &opts)
-      : CacheImpl<Arc>(opts),
-        fst_(fst.Copy()),
-        stack_(opts.stack ? opts.stack : new PdtStack<StateId, Label>(parens)),
-        state_table_(opts.state_table ? opts.state_table
-                                      : new PdtStateTable<StateId, StackId>()),
-        own_stack_(opts.stack == 0),
-        own_state_table_(opts.state_table == 0),
-        keep_parentheses_(opts.keep_parentheses) {
-    SetType("expand");
-    const auto props = fst.Properties(kFstProperties, false);
-    SetProperties(PdtExpandProperties(props), kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  PdtExpandFstImpl(const PdtExpandFstImpl &impl)
-      : CacheImpl<Arc>(impl),
-        fst_(impl.fst_->Copy(true)),
-        stack_(new PdtStack<StateId, Label>(*impl.stack_)),
-        state_table_(new PdtStateTable<StateId, StackId>()),
-        own_stack_(true),
-        own_state_table_(true),
-        keep_parentheses_(impl.keep_parentheses_) {
-    SetType("expand");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  ~PdtExpandFstImpl() override {
-    if (own_stack_) delete stack_;
-    if (own_state_table_) delete state_table_;
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      const auto s = fst_->Start();
-      if (s == kNoStateId) return kNoStateId;
-      StateTuple tuple(s, 0);
-      const auto start = state_table_->FindState(tuple);
-      SetStart(start);
-    }
-    return CacheImpl<Arc>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) {
-      const auto &tuple = state_table_->Tuple(s);
-      const auto weight = fst_->Final(tuple.state_id);
-      if (weight != Weight::Zero() && tuple.stack_id == 0)
-        SetFinal(s, weight);
-      else
-        SetFinal(s, Weight::Zero());
-    }
-    return CacheImpl<Arc>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) ExpandState(s);
-    return CacheImpl<Arc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) ExpandState(s);
-    return CacheImpl<Arc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) ExpandState(s);
-    return CacheImpl<Arc>::NumOutputEpsilons(s);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) ExpandState(s);
-    CacheImpl<Arc>::InitArcIterator(s, data);
-  }
-
-  // Computes the outgoing transitions from a state, creating new destination
-  // states as needed.
-  void ExpandState(StateId s) {
-    StateTuple tuple = state_table_->Tuple(s);
-    for (ArcIterator<Fst<Arc>> aiter(*fst_, tuple.state_id); !aiter.Done();
-         aiter.Next()) {
-      auto arc = aiter.Value();
-      const auto stack_id = stack_->Find(tuple.stack_id, arc.ilabel);
-      if (stack_id == -1) {  // Non-matching close parenthesis.
-        continue;
-      } else if ((stack_id != tuple.stack_id) && !keep_parentheses_) {
-        // Stack push/pop.
-        arc.ilabel = 0;
-        arc.olabel = 0;
-      }
-      StateTuple ntuple(arc.nextstate, stack_id);
-      arc.nextstate = state_table_->FindState(ntuple);
-      PushArc(s, arc);
-    }
-    SetArcs(s);
-  }
-
-  const PdtStack<StackId, Label> &GetStack() const { return *stack_; }
-
-  const PdtStateTable<StateId, StackId> &GetStateTable() const {
-    return *state_table_;
-  }
-
- private:
-  // Properties for an expanded PDT.
-  inline uint64 PdtExpandProperties(uint64 inprops) {
-    return inprops & (kAcceptor | kAcyclic | kInitialAcyclic | kUnweighted);
-  }
-
-  std::unique_ptr<const Fst<Arc>> fst_;
-  PdtStack<StackId, Label> *stack_;
-  PdtStateTable<StateId, StackId> *state_table_;
-  bool own_stack_;
-  bool own_state_table_;
-  bool keep_parentheses_;
-};
-
-}  // namespace internal
-
-// Expands a pushdown transducer (PDT) encoded as an FST into an FST. This
-// version is a delayed FST. In the PDT, some transitions are labeled with open
-// or close parentheses. To be interpreted as a PDT, the parens must balance on
-// a path. The open-close parenthesis label pairs are passed using the parens
-// argument. The expansion enforces the parenthesis constraints. The PDT must be
-// expandable as an FST.
-//
-// This class attaches interface to implementation and handles reference
-// counting, delegating most methods to ImplToFst.
-template <class A>
-class PdtExpandFst : public ImplToFst<internal::PdtExpandFstImpl<A>> {
- public:
-  using Arc = A;
-
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using StackId = StateId;
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::PdtExpandFstImpl<Arc>;
-
-  friend class ArcIterator<PdtExpandFst<Arc>>;
-  friend class StateIterator<PdtExpandFst<Arc>>;
-
-  PdtExpandFst(const Fst<Arc> &fst,
-               const std::vector<std::pair<Label, Label>> &parens)
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, parens, PdtExpandFstOptions<A>())) {}
-
-  PdtExpandFst(const Fst<Arc> &fst,
-               const std::vector<std::pair<Label, Label>> &parens,
-               const PdtExpandFstOptions<Arc> &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, parens, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  PdtExpandFst(const PdtExpandFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Gets a copy of this ExpandFst. See Fst<>::Copy() for further doc.
-  PdtExpandFst<Arc> *Copy(bool safe = false) const override {
-    return new PdtExpandFst<Arc>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
-  const PdtStack<StackId, Label> &GetStack() const {
-    return GetImpl()->GetStack();
-  }
-
-  const PdtStateTable<StateId, StackId> &GetStateTable() const {
-    return GetImpl()->GetStateTable();
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  void operator=(const PdtExpandFst &) = delete;
-};
-
-// Specialization for PdtExpandFst.
-template <class Arc>
-class StateIterator<PdtExpandFst<Arc>>
-    : public CacheStateIterator<PdtExpandFst<Arc>> {
- public:
-  explicit StateIterator(const PdtExpandFst<Arc> &fst)
-      : CacheStateIterator<PdtExpandFst<Arc>>(fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for PdtExpandFst.
-template <class Arc>
-class ArcIterator<PdtExpandFst<Arc>>
-    : public CacheArcIterator<PdtExpandFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const PdtExpandFst<Arc> &fst, StateId s)
-      : CacheArcIterator<PdtExpandFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->ExpandState(s);
-  }
-};
-
-template <class Arc>
-inline void PdtExpandFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<PdtExpandFst<Arc>>(*this);
-}
-
-// PrunedExpand prunes the delayed expansion of a pushdown transducer (PDT)
-// encoded as an FST into an FST. In the PDT, some transitions are labeled with
-// open or close parentheses. To be interpreted as a PDT, the parens must
-// balance on a path. The open-close parenthesis label pairs are passed
-// using the parens argument. The expansion enforces the parenthesis
-// constraints.
-//
-// The algorithm works by visiting the delayed ExpandFst using a shortest-stack
-// first queue discipline and relies on the shortest-distance information
-// computed using a reverse shortest-path call to perform the pruning.
-//
-// The algorithm maintains the same state ordering between the ExpandFst being
-// visited (efst_) and the result of pruning written into the MutableFst (ofst_)
-// to improve readability.
-template <class Arc>
-class PdtPrunedExpand {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using StackId = StateId;
-  using Stack = PdtStack<StackId, Label>;
-  using StateTable = PdtStateTable<StateId, StackId>;
-  using SetIterator = typename internal::PdtBalanceData<Arc>::SetIterator;
-
-  // Constructor taking as input a PDT specified by by an input FST and a vector
-  // of parentheses. The keep_parentheses argument specifies whether parentheses
-  // are replaced by epsilons or not during the expansion. The cache options are
-  // passed to the underlying ExpandFst.
-  PdtPrunedExpand(const Fst<Arc> &ifst,
-                  const std::vector<std::pair<Label, Label>> &parens,
-                  bool keep_parentheses = false,
-                  const CacheOptions &opts = CacheOptions())
-      : ifst_(ifst.Copy()),
-        keep_parentheses_(keep_parentheses),
-        stack_(parens),
-        efst_(ifst, parens,
-              PdtExpandFstOptions<Arc>(opts, true, &stack_, &state_table_)),
-        queue_(state_table_, stack_, stack_length_, distance_, fdistance_),
-        error_(false) {
-    Reverse(*ifst_, parens, &rfst_);
-    VectorFst<Arc> path;
-    reverse_shortest_path_.reset(new PdtShortestPath<Arc, FifoQueue<StateId>>(
-        rfst_, parens,
-        PdtShortestPathOptions<Arc, FifoQueue<StateId>>(true, false)));
-    reverse_shortest_path_->ShortestPath(&path);
-    error_ = (path.Properties(kError, true) == kError);
-    balance_data_.reset(reverse_shortest_path_->GetBalanceData()->Reverse(
-        rfst_.NumStates(), 10, -1));
-    InitCloseParenMultimap(parens);
-  }
-
-  bool Error() const { return error_; }
-
-  // Expands and prunes the input PDT according to the provided weight
-  // threshold, wirting the result into an output mutable FST.
-  void Expand(MutableFst<Arc> *ofst, const Weight &threshold);
-
- private:
-  static constexpr uint8 kEnqueued = 0x01;
-  static constexpr uint8 kExpanded = 0x02;
-  static constexpr uint8 kSourceState = 0x04;
-
-  // Comparison functor used by the queue:
-  //
-  // 1. States corresponding to shortest stack first, and
-  // 2. for stacks of matching length, reverse lexicographic order is used, and
-  // 3. for states with the same stack, shortest-first order is used.
-  class StackCompare {
-   public:
-    StackCompare(const StateTable &state_table, const Stack &stack,
-                 const std::vector<StackId> &stack_length,
-                 const std::vector<Weight> &distance,
-                 const std::vector<Weight> &fdistance)
-        : state_table_(state_table),
-          stack_(stack),
-          stack_length_(stack_length),
-          distance_(distance),
-          fdistance_(fdistance) {}
-
-    bool operator()(StateId s1, StateId s2) const {
-      auto si1 = state_table_.Tuple(s1).stack_id;
-      auto si2 = state_table_.Tuple(s2).stack_id;
-      if (stack_length_[si1] < stack_length_[si2]) return true;
-      if (stack_length_[si1] > stack_length_[si2]) return false;
-      // If stack IDs are equal, use A*.
-      if (si1 == si2) {
-        return less_(Distance(s1), Distance(s2));
-      }
-      // If lengths are equal, uses reverse lexicographic order.
-      for (; si1 != si2; si1 = stack_.Pop(si1), si2 = stack_.Pop(si2)) {
-        if (stack_.Top(si1) < stack_.Top(si2)) return true;
-        if (stack_.Top(si1) > stack_.Top(si2)) return false;
-      }
-      return false;
-    }
-
-   private:
-    Weight Distance(StateId s) const {
-      return (s < distance_.size()) && (s < fdistance_.size())
-                 ? Times(distance_[s], fdistance_[s])
-                 : Weight::Zero();
-    }
-
-    const StateTable &state_table_;
-    const Stack &stack_;
-    const std::vector<StackId> &stack_length_;
-    const std::vector<Weight> &distance_;
-    const std::vector<Weight> &fdistance_;
-    const NaturalLess<Weight> less_;
-  };
-
-  class ShortestStackFirstQueue
-      : public ShortestFirstQueue<StateId, StackCompare> {
-   public:
-    ShortestStackFirstQueue(const PdtStateTable<StateId, StackId> &state_table,
-                            const Stack &stack,
-                            const std::vector<StackId> &stack_length,
-                            const std::vector<Weight> &distance,
-                            const std::vector<Weight> &fdistance)
-        : ShortestFirstQueue<StateId, StackCompare>(StackCompare(
-              state_table, stack, stack_length, distance, fdistance)) {}
-  };
-
-  void InitCloseParenMultimap(
-      const std::vector<std::pair<Label, Label>> &parens);
-
-  Weight DistanceToDest(StateId source, StateId dest) const;
-
-  uint8 Flags(StateId s) const;
-
-  void SetFlags(StateId s, uint8 flags, uint8 mask);
-
-  Weight Distance(StateId s) const;
-
-  void SetDistance(StateId s, Weight weight);
-
-  Weight FinalDistance(StateId s) const;
-
-  void SetFinalDistance(StateId s, Weight weight);
-
-  StateId SourceState(StateId s) const;
-
-  void SetSourceState(StateId s, StateId p);
-
-  void AddStateAndEnqueue(StateId s);
-
-  void Relax(StateId s, const Arc &arc, Weight weight);
-
-  bool PruneArc(StateId s, const Arc &arc);
-
-  void ProcStart();
-
-  void ProcFinal(StateId s);
-
-  bool ProcNonParen(StateId s, const Arc &arc, bool add_arc);
-
-  bool ProcOpenParen(StateId s, const Arc &arc, StackId si, StackId nsi);
-
-  bool ProcCloseParen(StateId s, const Arc &arc);
-
-  void ProcDestStates(StateId s, StackId si);
-
-  // Input PDT.
-  std::unique_ptr<Fst<Arc>> ifst_;
-  // Reversed PDT.
-  VectorFst<Arc> rfst_;
-  // Keep parentheses in ofst?
-  const bool keep_parentheses_;
-  // State table for efst_.
-  StateTable state_table_;
-  // Stack trie.
-  Stack stack_;
-  // Expanded PDT.
-  PdtExpandFst<Arc> efst_;
-  // Length of stack for given stack ID.
-  std::vector<StackId> stack_length_;
-  // Distance from initial state in efst_/ofst.
-  std::vector<Weight> distance_;
-  // Distance to final states in efst_/ofst.
-  std::vector<Weight> fdistance_;
-  // Queue used to visit efst_.
-  ShortestStackFirstQueue queue_;
-  // Construction time failure?
-  bool error_;
-  // Status flags for states in efst_/ofst.
-  std::vector<uint8> flags_;
-  // PDT source state for each expanded state.
-  std::vector<StateId> sources_;
-  // Shortest path for rfst_.
-  std::unique_ptr<PdtShortestPath<Arc, FifoQueue<StateId>>>
-      reverse_shortest_path_;
-  std::unique_ptr<internal::PdtBalanceData<Arc>> balance_data_;
-  // Maps open paren arcs to balancing close paren arcs.
-  typename PdtShortestPath<Arc, FifoQueue<StateId>>::CloseParenMultimap
-      close_paren_multimap_;
-  MutableFst<Arc> *ofst_;  // Output FST.
-  Weight limit_;           // Weight limit.
-
-  // Maps a state s in ifst (i.e., the source of a closed paranthesis matching
-  // the top of current_stack_id_ to final states in efst_.
-  std::unordered_map<StateId, Weight> dest_map_;
-  // Stack ID of the states currently at the top of the queue, i.e., the states
-  // currently being popped and processed.
-  StackId current_stack_id_;
-  ssize_t current_paren_id_;  // Paren ID at top of current stack.
-  ssize_t cached_stack_id_;
-  StateId cached_source_;
-  // The set of pairs of destination states and weights to final states for the
-  // source state cached_source_ and the paren ID cached_paren_id_; i.e., the
-  // set of source states of a closed parenthesis with paren ID cached_paren_id
-  // balancing an incoming open parenthesis with paren ID cached_paren_id_ in
-  // state cached_source_.
-  std::forward_list<std::pair<StateId, Weight>> cached_dest_list_;
-  NaturalLess<Weight> less_;
-};
-
-// Initializes close paren multimap, mapping pairs (s, paren_id) to all the arcs
-// out of s labeled with close parenthese for paren_id.
-template <class Arc>
-void PdtPrunedExpand<Arc>::InitCloseParenMultimap(
-    const std::vector<std::pair<Label, Label>> &parens) {
-  std::unordered_map<Label, Label> paren_map;
-  for (size_t i = 0; i < parens.size(); ++i) {
-    const auto &pair = parens[i];
-    paren_map[pair.first] = i;
-    paren_map[pair.second] = i;
-  }
-  for (StateIterator<Fst<Arc>> siter(*ifst_); !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    for (ArcIterator<Fst<Arc>> aiter(*ifst_, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto it = paren_map.find(arc.ilabel);
-      if (it == paren_map.end()) continue;
-      if (arc.ilabel == parens[it->second].second) {  // Close paren.
-        const internal::ParenState<Arc> key(it->second, s);
-        close_paren_multimap_.emplace(key, arc);
-      }
-    }
-  }
-}
-
-// Returns the weight of the shortest balanced path from source to dest
-// in ifst_; dest must be the source state of a close paren arc.
-template <class Arc>
-typename Arc::Weight PdtPrunedExpand<Arc>::DistanceToDest(StateId source,
-                                                          StateId dest) const {
-  using SearchState =
-      typename PdtShortestPath<Arc, FifoQueue<StateId>>::SearchState;
-  const SearchState ss(source + 1, dest + 1);
-  const auto distance =
-      reverse_shortest_path_->GetShortestPathData().Distance(ss);
-  VLOG(2) << "D(" << source << ", " << dest << ") =" << distance;
-  return distance;
-}
-
-// Returns the flags for state s in ofst_.
-template <class Arc>
-uint8 PdtPrunedExpand<Arc>::Flags(StateId s) const {
-  return s < flags_.size() ? flags_[s] : 0;
-}
-
-// Modifies the flags for state s in ofst_.
-template <class Arc>
-void PdtPrunedExpand<Arc>::SetFlags(StateId s, uint8 flags, uint8 mask) {
-  while (flags_.size() <= s) flags_.push_back(0);
-  flags_[s] &= ~mask;
-  flags_[s] |= flags & mask;
-}
-
-// Returns the shortest distance from the initial state to s in ofst_.
-template <class Arc>
-typename Arc::Weight PdtPrunedExpand<Arc>::Distance(StateId s) const {
-  return s < distance_.size() ? distance_[s] : Weight::Zero();
-}
-
-// Sets the shortest distance from the initial state to s in ofst_.
-template <class Arc>
-void PdtPrunedExpand<Arc>::SetDistance(StateId s, Weight weight) {
-  while (distance_.size() <= s) distance_.push_back(Weight::Zero());
-  distance_[s] = std::move(weight);
-}
-
-// Returns the shortest distance from s to the final states in ofst_.
-template <class Arc>
-typename Arc::Weight PdtPrunedExpand<Arc>::FinalDistance(StateId s) const {
-  return s < fdistance_.size() ? fdistance_[s] : Weight::Zero();
-}
-
-// Sets the shortest distance from s to the final states in ofst_.
-template <class Arc>
-void PdtPrunedExpand<Arc>::SetFinalDistance(StateId s, Weight weight) {
-  while (fdistance_.size() <= s) fdistance_.push_back(Weight::Zero());
-  fdistance_[s] = std::move(weight);
-}
-
-// Returns the PDT source state of state s in ofst_.
-template <class Arc>
-typename Arc::StateId PdtPrunedExpand<Arc>::SourceState(StateId s) const {
-  return s < sources_.size() ? sources_[s] : kNoStateId;
-}
-
-// Sets the PDT source state of state s in ofst_ to state p'in ifst_.
-template <class Arc>
-void PdtPrunedExpand<Arc>::SetSourceState(StateId s, StateId p) {
-  while (sources_.size() <= s) sources_.push_back(kNoStateId);
-  sources_[s] = p;
-}
-
-// Adds state s of efst_ to ofst_ and inserts it in the queue, modifying the
-// flags for s accordingly.
-template <class Arc>
-void PdtPrunedExpand<Arc>::AddStateAndEnqueue(StateId s) {
-  if (!(Flags(s) & (kEnqueued | kExpanded))) {
-    while (ofst_->NumStates() <= s) ofst_->AddState();
-    queue_.Enqueue(s);
-    SetFlags(s, kEnqueued, kEnqueued);
-  } else if (Flags(s) & kEnqueued) {
-    queue_.Update(s);
-  }
-  // TODO(allauzen): Check everything is fine when kExpanded?
-}
-
-// Relaxes arc out of state s in ofst_ as follows:
-//
-// 1. If the distance to s times the weight of arc is smaller than
-//   the currently stored distance for arc.nextstate, updates
-//   Distance(arc.nextstate) with a new estimate
-// 2. If fd is less than the currently stored distance from arc.nextstate to the
-// final state, updates with new estimate.
-template <class Arc>
-void PdtPrunedExpand<Arc>::Relax(StateId s, const Arc &arc, Weight fd) {
-  const auto nd = Times(Distance(s), arc.weight);
-  if (less_(nd, Distance(arc.nextstate))) {
-    SetDistance(arc.nextstate, nd);
-    SetSourceState(arc.nextstate, SourceState(s));
-  }
-  if (less_(fd, FinalDistance(arc.nextstate))) {
-    SetFinalDistance(arc.nextstate, fd);
-  }
-  VLOG(2) << "Relax: " << s << ", d[s] = " << Distance(s) << ", to "
-          << arc.nextstate << ", d[ns] = " << Distance(arc.nextstate)
-          << ", nd = " << nd;
-}
-
-// Returns whether the arc out of state s in efst needs pruned.
-template <class Arc>
-bool PdtPrunedExpand<Arc>::PruneArc(StateId s, const Arc &arc) {
-  VLOG(2) << "Prune ?";
-  auto fd = Weight::Zero();
-  if ((cached_source_ != SourceState(s)) ||
-      (cached_stack_id_ != current_stack_id_)) {
-    cached_source_ = SourceState(s);
-    cached_stack_id_ = current_stack_id_;
-    cached_dest_list_.clear();
-    if (cached_source_ != ifst_->Start()) {
-      for (auto set_iter =
-               balance_data_->Find(current_paren_id_, cached_source_);
-           !set_iter.Done(); set_iter.Next()) {
-        auto dest = set_iter.Element();
-        const auto it = dest_map_.find(dest);
-        cached_dest_list_.push_front(*it);
-      }
-    } else {
-      // TODO(allauzen): queue discipline should prevent this from ever
-      // happening.
-      // Replace by a check.
-      cached_dest_list_.push_front(
-          std::make_pair(rfst_.Start() - 1, Weight::One()));
-    }
-  }
-  for (auto it = cached_dest_list_.begin(); it != cached_dest_list_.end();
-       ++it) {
-    const auto d =
-        DistanceToDest(state_table_.Tuple(arc.nextstate).state_id, it->first);
-    fd = Plus(fd, Times(d, it->second));
-  }
-  Relax(s, arc, fd);
-  return less_(limit_, Times(Distance(s), Times(arc.weight, fd)));
-}
-
-// Adds start state of efst_ to ofst_, enqueues it, and initializes the distance
-// data structures.
-template <class Arc>
-void PdtPrunedExpand<Arc>::ProcStart() {
-  const auto s = efst_.Start();
-  AddStateAndEnqueue(s);
-  ofst_->SetStart(s);
-  SetSourceState(s, ifst_->Start());
-  current_stack_id_ = 0;
-  current_paren_id_ = -1;
-  stack_length_.push_back(0);
-  const auto r = rfst_.Start() - 1;
-  cached_source_ = ifst_->Start();
-  cached_stack_id_ = 0;
-  cached_dest_list_.push_front(std::make_pair(r, Weight::One()));
-  const PdtStateTuple<StateId, StackId> tuple(r, 0);
-  SetFinalDistance(state_table_.FindState(tuple), Weight::One());
-  SetDistance(s, Weight::One());
-  const auto d = DistanceToDest(ifst_->Start(), r);
-  SetFinalDistance(s, d);
-  VLOG(2) << d;
-}
-
-// Makes s final in ofst_ if shortest accepting path ending in s is below
-// threshold.
-template <class Arc>
-void PdtPrunedExpand<Arc>::ProcFinal(StateId s) {
-  const auto weight = efst_.Final(s);
-  if (weight == Weight::Zero()) return;
-  if (less_(limit_, Times(Distance(s), weight))) return;
-  ofst_->SetFinal(s, weight);
-}
-
-// Returns true when an arc (or meta-arc) leaving state s in efst_ is below the
-// threshold. When add_arc is true, arc is added to ofst_.
-template <class Arc>
-bool PdtPrunedExpand<Arc>::ProcNonParen(StateId s, const Arc &arc,
-                                        bool add_arc) {
-  VLOG(2) << "ProcNonParen: " << s << " to " << arc.nextstate << ", "
-          << arc.ilabel << ":" << arc.olabel << " / " << arc.weight
-          << ", add_arc = " << (add_arc ? "true" : "false");
-  if (PruneArc(s, arc)) return false;
-  if (add_arc) ofst_->AddArc(s, arc);
-  AddStateAndEnqueue(arc.nextstate);
-  return true;
-}
-
-// Processes an open paren arc leaving state s in ofst_. When the arc is labeled
-// with an open paren,
-//
-// 1. Considers each (shortest) balanced path starting in s by taking the arc
-// and ending by a close paren balancing the open paren of as a meta-arc,
-// processing and pruning each meta-arc as a non-paren arc, inserting its
-// destination to the queue;
-// 2. if at least one of these meta-arcs has not been pruned, adds the
-// destination of arc to ofst_ as a new source state for the stack ID nsi, and
-// inserts it in the queue.
-template <class Arc>
-bool PdtPrunedExpand<Arc>::ProcOpenParen(StateId s, const Arc &arc, StackId si,
-                                         StackId nsi) {
-  // Updates the stack length when needed.
-  while (stack_length_.size() <= nsi) stack_length_.push_back(-1);
-  if (stack_length_[nsi] == -1) stack_length_[nsi] = stack_length_[si] + 1;
-  const auto ns = arc.nextstate;
-  VLOG(2) << "Open paren: " << s << "(" << state_table_.Tuple(s).state_id
-          << ") to " << ns << "(" << state_table_.Tuple(ns).state_id << ")";
-  bool proc_arc = false;
-  auto fd = Weight::Zero();
-  const auto paren_id = stack_.ParenId(arc.ilabel);
-  std::forward_list<StateId> sources;
-  for (auto set_iter =
-           balance_data_->Find(paren_id, state_table_.Tuple(ns).state_id);
-       !set_iter.Done(); set_iter.Next()) {
-    sources.push_front(set_iter.Element());
-  }
-  for (const auto source : sources) {
-    VLOG(2) << "Close paren source: " << source;
-    const internal::ParenState<Arc> paren_state(paren_id, source);
-    for (auto it = close_paren_multimap_.find(paren_state);
-         it != close_paren_multimap_.end() && paren_state == it->first; ++it) {
-      auto meta_arc = it->second;
-      const PdtStateTuple<StateId, StackId> tuple(meta_arc.nextstate, si);
-      meta_arc.nextstate = state_table_.FindState(tuple);
-      const auto state_id = state_table_.Tuple(ns).state_id;
-      const auto d = DistanceToDest(state_id, source);
-      VLOG(2) << state_id << ", " << source;
-      VLOG(2) << "Meta arc weight = " << arc.weight << " Times " << d
-              << " Times " << meta_arc.weight;
-      meta_arc.weight = Times(arc.weight, Times(d, meta_arc.weight));
-      proc_arc |= ProcNonParen(s, meta_arc, false);
-      fd = Plus(
-          fd,
-          Times(Times(DistanceToDest(state_table_.Tuple(ns).state_id, source),
-                      it->second.weight),
-                FinalDistance(meta_arc.nextstate)));
-    }
-  }
-  if (proc_arc) {
-    VLOG(2) << "Proc open paren " << s << " to " << arc.nextstate;
-    ofst_->AddArc(
-        s, keep_parentheses_ ? arc : Arc(0, 0, arc.weight, arc.nextstate));
-    AddStateAndEnqueue(arc.nextstate);
-    const auto nd = Times(Distance(s), arc.weight);
-    if (less_(nd, Distance(arc.nextstate))) SetDistance(arc.nextstate, nd);
-    // FinalDistance not necessary for source state since pruning decided using
-    // meta-arcs above.  But this is a problem with A*, hence the following.
-    if (less_(fd, FinalDistance(arc.nextstate)))
-      SetFinalDistance(arc.nextstate, fd);
-    SetFlags(arc.nextstate, kSourceState, kSourceState);
-  }
-  return proc_arc;
-}
-
-// Checks that shortest path through close paren arc in efst_ is below
-// threshold, and if so, adds it to ofst_.
-template <class Arc>
-bool PdtPrunedExpand<Arc>::ProcCloseParen(StateId s, const Arc &arc) {
-  const auto weight =
-      Times(Distance(s), Times(arc.weight, FinalDistance(arc.nextstate)));
-  if (less_(limit_, weight)) return false;
-  ofst_->AddArc(s,
-                keep_parentheses_ ? arc : Arc(0, 0, arc.weight, arc.nextstate));
-  return true;
-}
-
-// When state s in ofst_ is a source state for stack ID si, identifies all the
-// corresponding possible destination states, that is, all the states in ifst_
-// that have an outgoing close paren arc balancing the incoming open paren taken
-// to get to s. For each such state t, computes the shortest distance from (t,
-// si) to the final states in ofst_. Stores this information in dest_map_.
-template <class Arc>
-void PdtPrunedExpand<Arc>::ProcDestStates(StateId s, StackId si) {
-  if (!(Flags(s) & kSourceState)) return;
-  if (si != current_stack_id_) {
-    dest_map_.clear();
-    current_stack_id_ = si;
-    current_paren_id_ = stack_.Top(current_stack_id_);
-    VLOG(2) << "StackID " << si << " dequeued for first time";
-  }
-  // TODO(allauzen): clean up source state business; rename current function to
-  // ProcSourceState.
-  SetSourceState(s, state_table_.Tuple(s).state_id);
-  const auto paren_id = stack_.Top(si);
-  for (auto set_iter =
-           balance_data_->Find(paren_id, state_table_.Tuple(s).state_id);
-       !set_iter.Done(); set_iter.Next()) {
-    const auto dest_state = set_iter.Element();
-    if (dest_map_.find(dest_state) != dest_map_.end()) continue;
-    auto dest_weight = Weight::Zero();
-    internal::ParenState<Arc> paren_state(paren_id, dest_state);
-    for (auto it = close_paren_multimap_.find(paren_state);
-         it != close_paren_multimap_.end() && paren_state == it->first; ++it) {
-      const auto &arc = it->second;
-      const PdtStateTuple<StateId, StackId> tuple(arc.nextstate,
-                                                  stack_.Pop(si));
-      dest_weight =
-          Plus(dest_weight,
-               Times(arc.weight, FinalDistance(state_table_.FindState(tuple))));
-    }
-    dest_map_[dest_state] = dest_weight;
-    VLOG(2) << "State " << dest_state << " is a dest state for stack ID " << si
-            << " with weight " << dest_weight;
-  }
-}
-
-// Expands and prunes the input PDT, writing the result in ofst.
-template <class Arc>
-void PdtPrunedExpand<Arc>::Expand(MutableFst<Arc> *ofst,
-                                  const typename Arc::Weight &threshold) {
-  ofst_ = ofst;
-  if (error_) {
-    ofst_->SetProperties(kError, kError);
-    return;
-  }
-  ofst_->DeleteStates();
-  ofst_->SetInputSymbols(ifst_->InputSymbols());
-  ofst_->SetOutputSymbols(ifst_->OutputSymbols());
-  limit_ = Times(DistanceToDest(ifst_->Start(), rfst_.Start() - 1), threshold);
-  flags_.clear();
-  ProcStart();
-  while (!queue_.Empty()) {
-    const auto s = queue_.Head();
-    queue_.Dequeue();
-    SetFlags(s, kExpanded, kExpanded | kEnqueued);
-    VLOG(2) << s << " dequeued!";
-    ProcFinal(s);
-    StackId stack_id = state_table_.Tuple(s).stack_id;
-    ProcDestStates(s, stack_id);
-    for (ArcIterator<PdtExpandFst<Arc>> aiter(efst_, s); !aiter.Done();
-         aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto nextstack_id = state_table_.Tuple(arc.nextstate).stack_id;
-      if (stack_id == nextstack_id) {
-        ProcNonParen(s, arc, true);
-      } else if (stack_id == stack_.Pop(nextstack_id)) {
-        ProcOpenParen(s, arc, stack_id, nextstack_id);
-      } else {
-        ProcCloseParen(s, arc);
-      }
-    }
-    VLOG(2) << "d[" << s << "] = " << Distance(s) << ", fd[" << s
-            << "] = " << FinalDistance(s);
-  }
-}
-
-// Expand functions.
-
-template <class Arc>
-struct PdtExpandOptions {
-  using Weight = typename Arc::Weight;
-
-  bool connect;
-  bool keep_parentheses;
-  Weight weight_threshold;
-
-  PdtExpandOptions(bool connect = true, bool keep_parentheses = false,
-                   Weight weight_threshold = Weight::Zero())
-      : connect(connect),
-        keep_parentheses(keep_parentheses),
-        weight_threshold(std::move(weight_threshold)) {}
-};
-
-// Expands a pushdown transducer (PDT) encoded as an FST into an FST. This
-// version writes the expanded PDT to a mutable FST. In the PDT, some
-// transitions are labeled with open or close parentheses. To be interpreted as
-// a PDT, the parens must balance on a path. The open-close parenthesis label
-// pairs are passed using the parens argument. Expansion enforces the
-// parenthesis constraints. The PDT must be expandable as an FST.
-template <class Arc>
-void Expand(
-    const Fst<Arc> &ifst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens,
-    MutableFst<Arc> *ofst, const PdtExpandOptions<Arc> &opts) {
-  PdtExpandFstOptions<Arc> eopts;
-  eopts.gc_limit = 0;
-  if (opts.weight_threshold == Arc::Weight::Zero()) {
-    eopts.keep_parentheses = opts.keep_parentheses;
-    *ofst = PdtExpandFst<Arc>(ifst, parens, eopts);
-  } else {
-    PdtPrunedExpand<Arc> pruned_expand(ifst, parens, opts.keep_parentheses);
-    pruned_expand.Expand(ofst, opts.weight_threshold);
-  }
-  if (opts.connect) Connect(ofst);
-}
-
-// Expands a pushdown transducer (PDT) encoded as an FST into an FST. This
-// version writes the expanded PDT result to a mutable FST. In the PDT, some
-// transitions are labeled with open or close parentheses. To be interpreted as
-// a PDT, the parens must balance on a path. The open-close parenthesis label
-// pairs are passed using the parents argument. Expansion enforces the
-// parenthesis constraints. The PDT must be expandable as an FST.
-template <class Arc>
-void Expand(const Fst<Arc> &ifst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-    &parens, MutableFst<Arc> *ofst, bool connect = true,
-    bool keep_parentheses = false) {
-  const PdtExpandOptions<Arc> opts(connect, keep_parentheses);
-  Expand(ifst, parens, ofst, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_EXPAND_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/getters.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/getters.h
deleted file mode 100644
index 69dd150d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/getters.h
+++ /dev/null
@@ -1,22 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_PDT_GETTERS_H_
-#define FST_EXTENSIONS_PDT_GETTERS_H_
-
-#include <string>
-
-#include <fst/extensions/pdt/compose.h>
-#include <fst/extensions/pdt/replace.h>
-
-namespace fst {
-namespace script {
-
-bool GetPdtComposeFilter(const string &str, PdtComposeFilter *cf);
-
-bool GetPdtParserType(const string &str, PdtParserType *pt);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_GETTERS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/info.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/info.h
deleted file mode 100644
index 3de54772..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/info.h
+++ /dev/null
@@ -1,152 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Prints information about a PDT.
-
-#ifndef FST_EXTENSIONS_PDT_INFO_H_
-#define FST_EXTENSIONS_PDT_INFO_H_
-
-#include <unordered_map>
-#include <unordered_set>
-#include <vector>
-
-#include <fst/extensions/pdt/pdt.h>
-#include <fst/fst.h>
-
-namespace fst {
-
-// Compute various information about PDTs.
-template <class Arc>
-class PdtInfo {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  PdtInfo(const Fst<Arc> &fst,
-          const std::vector<std::pair<Label, Label>> &parents);
-
-  const string &FstType() const { return fst_type_; }
-
-  const string &ArcType() const { return Arc::Type(); }
-
-  int64 NumStates() const { return nstates_; }
-
-  int64 NumArcs() const { return narcs_; }
-
-  int64 NumOpenParens() const { return nopen_parens_; }
-
-  int64 NumCloseParens() const { return nclose_parens_; }
-
-  int64 NumUniqueOpenParens() const { return nuniq_open_parens_; }
-
-  int64 NumUniqueCloseParens() const { return nuniq_close_parens_; }
-
-  int64 NumOpenParenStates() const { return nopen_paren_states_; }
-
-  int64 NumCloseParenStates() const { return nclose_paren_states_; }
-
- private:
-  string fst_type_;
-  int64 nstates_;
-  int64 narcs_;
-  int64 nopen_parens_;
-  int64 nclose_parens_;
-  int64 nuniq_open_parens_;
-  int64 nuniq_close_parens_;
-  int64 nopen_paren_states_;
-  int64 nclose_paren_states_;
-};
-
-template <class Arc>
-PdtInfo<Arc>::PdtInfo(
-    const Fst<Arc> &fst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens)
-    : fst_type_(fst.Type()),
-      nstates_(0),
-      narcs_(0),
-      nopen_parens_(0),
-      nclose_parens_(0),
-      nuniq_open_parens_(0),
-      nuniq_close_parens_(0),
-      nopen_paren_states_(0),
-      nclose_paren_states_(0) {
-  std::unordered_map<Label, size_t> paren_map;
-  std::unordered_set<Label> paren_set;
-  std::unordered_set<StateId> open_paren_state_set;
-  std::unordered_set<StateId> close_paren_state_set;
-  for (size_t i = 0; i < parens.size(); ++i) {
-    const auto &pair = parens[i];
-    paren_map[pair.first] = i;
-    paren_map[pair.second] = i;
-  }
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    ++nstates_;
-    const auto s = siter.Value();
-    for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      ++narcs_;
-      const auto it = paren_map.find(arc.ilabel);
-      if (it != paren_map.end()) {
-        const auto open_paren = parens[it->second].first;
-        const auto close_paren = parens[it->second].second;
-        if (arc.ilabel == open_paren) {
-          ++nopen_parens_;
-          if (!paren_set.count(open_paren)) {
-            ++nuniq_open_parens_;
-            paren_set.insert(open_paren);
-          }
-          if (!open_paren_state_set.count(arc.nextstate)) {
-            ++nopen_paren_states_;
-            open_paren_state_set.insert(arc.nextstate);
-          }
-        } else {
-          ++nclose_parens_;
-          if (!paren_set.count(close_paren)) {
-            ++nuniq_close_parens_;
-            paren_set.insert(close_paren);
-          }
-          if (!close_paren_state_set.count(s)) {
-            ++nclose_paren_states_;
-            close_paren_state_set.insert(s);
-          }
-        }
-      }
-    }
-  }
-}
-
-template <class Arc>
-void PrintPdtInfo(const PdtInfo<Arc> &info) {
-  const auto old = std::cout.setf(std::ios::left);
-  std::cout.width(50);
-  std::cout << "fst type" << info.FstType() << std::endl;
-  std::cout.width(50);
-  std::cout << "arc type" << info.ArcType() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of states" << info.NumStates() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of arcs" << info.NumArcs() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of open parentheses" << info.NumOpenParens() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of close parentheses" << info.NumCloseParens() << std::endl;
-  std::cout.width(50);
-  std::cout << "# of unique open parentheses" << info.NumUniqueOpenParens()
-            << std::endl;
-  std::cout.width(50);
-  std::cout << "# of unique close parentheses" << info.NumUniqueCloseParens()
-            << std::endl;
-  std::cout.width(50);
-  std::cout << "# of open parenthesis dest. states" << info.NumOpenParenStates()
-            << std::endl;
-  std::cout.width(50);
-  std::cout << "# of close parenthesis source states"
-            << info.NumCloseParenStates() << std::endl;
-  std::cout.setf(old);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_INFO_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/paren.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/paren.h
deleted file mode 100644
index 0347848d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/paren.h
+++ /dev/null
@@ -1,440 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Common classes for PDT parentheses.
-
-#ifndef FST_EXTENSIONS_PDT_PAREN_H_
-#define FST_EXTENSIONS_PDT_PAREN_H_
-
-#include <algorithm>
-#include <set>
-#include <unordered_map>
-#include <unordered_set>
-
-#include <fst/log.h>
-
-#include <fst/extensions/pdt/collection.h>
-#include <fst/extensions/pdt/pdt.h>
-#include <fst/dfs-visit.h>
-#include <fst/fst.h>
-
-
-namespace fst {
-namespace internal {
-
-// ParenState: Pair of an open (close) parenthesis and its destination (source)
-// state.
-
-template <class Arc>
-struct ParenState {
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-
-  Label paren_id;    // ID of open (close) paren.
-  StateId state_id;  // Destination (source) state of open (close) paren.
-
-  explicit ParenState(Label paren_id = kNoLabel, StateId state_id = kNoStateId)
-      : paren_id(paren_id), state_id(state_id) {}
-
-  bool operator==(const ParenState<Arc> &other) const {
-    if (&other == this) return true;
-    return other.paren_id == paren_id && other.state_id == state_id;
-  }
-
-  bool operator!=(const ParenState<Arc> &other) const {
-    return !(other == *this);
-  }
-
-  struct Hash {
-    size_t operator()(const ParenState<Arc> &pstate) const {
-      static constexpr auto prime = 7853;
-      return pstate.paren_id + pstate.state_id * prime;
-    }
-  };
-};
-
-// Creates an FST-style const iterator from an STL-style map.
-template <class Map>
-class MapIterator {
- public:
-  using StlIterator = typename Map::const_iterator;
-  using ValueType = typename Map::mapped_type;
-
-  MapIterator(const Map &map, StlIterator it)
-      : begin_(it), end_(map.end()), it_(it) {}
-
-  bool Done() const { return it_ == end_ || it_->first != begin_->first; }
-
-  ValueType Value() const { return it_->second; }
-
-  void Next() { ++it_; }
-
-  void Reset() { it_ = begin_; }
-
- private:
-  const StlIterator begin_;
-  const StlIterator end_;
-  StlIterator it_;
-};
-
-// PdtParenReachable: Provides various parenthesis reachability information.
-
-template <class Arc>
-class PdtParenReachable {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-
-  using State = ParenState<Arc>;
-  using StateHash = typename State::Hash;
-
-  // Maps from state ID to reachable paren IDs from (to) that state.
-  using ParenMultimap = std::unordered_multimap<StateId, Label>;
-
-  // Maps from paren ID and state ID to reachable state set ID.
-  using StateSetMap = std::unordered_map<State, ssize_t, StateHash>;
-
-  // Maps from paren ID and state ID to arcs exiting that state with that
-  // Label.
-  using ParenArcMultimap = std::unordered_map<State, Arc, StateHash>;
-
-  using ParenIterator = MapIterator<ParenMultimap>;
-
-  using ParenArcIterator = MapIterator<ParenArcMultimap>;
-
-  using SetIterator = typename Collection<ssize_t, StateId>::SetIterator;
-
-  // Computes close (open) parenthesis reachability information for a PDT with
-  // bounded stack.
-  PdtParenReachable(const Fst<Arc> &fst,
-                    const std::vector<std::pair<Label, Label>> &parens,
-                    bool close)
-      : fst_(fst), parens_(parens), close_(close), error_(false) {
-    paren_map_.reserve(2 * parens.size());
-    for (size_t i = 0; i < parens.size(); ++i) {
-      const auto &pair = parens[i];
-      paren_map_[pair.first] = i;
-      paren_map_[pair.second] = i;
-    }
-    if (close_) {
-      const auto start = fst.Start();
-      if (start == kNoStateId) return;
-      if (!DFSearch(start)) {
-        FSTERROR() << "PdtReachable: Underlying cyclicity not supported";
-        error_ = true;
-      }
-    } else {
-      FSTERROR() << "PdtParenReachable: Open paren info not implemented";
-      error_ = true;
-    }
-  }
-
-  bool Error() const { return error_; }
-
-  // Given a state ID, returns an iterator over paren IDs for close (open)
-  // parens reachable from that state along balanced paths.
-  ParenIterator FindParens(StateId s) const {
-    return ParenIterator(paren_multimap_, paren_multimap_.find(s));
-  }
-
-  // Given a paren ID and a state ID s, returns an iterator over states that can
-  // be reached along balanced paths from (to) s that have have close (open)
-  // parentheses matching the paren ID exiting (entering) those states.
-  SetIterator FindStates(Label paren_id, StateId s) const {
-    const State paren_state(paren_id, s);
-    const auto it = set_map_.find(paren_state);
-    if (it == set_map_.end()) {
-      return state_sets_.FindSet(-1);
-    } else {
-      return state_sets_.FindSet(it->second);
-    }
-  }
-
-  // Given a paren ID and a state ID s, return an iterator over arcs that exit
-  // (enter) s and are labeled with a close (open) parenthesis matching the
-  // paren ID.
-  ParenArcIterator FindParenArcs(Label paren_id, StateId s) const {
-    const State paren_state(paren_id, s);
-    return ParenArcIterator(paren_arc_multimap_,
-                            paren_arc_multimap_.find(paren_state));
-  }
-
- private:
-  // Returns false when cycle detected during DFS gathering paren and state set
-  // information.
-  bool DFSearch(StateId s);
-
-  // Unions state sets together gathered by the DFS.
-  void ComputeStateSet(StateId s);
-
-  // Gathers state set(s) from state.
-  void UpdateStateSet(StateId nextstate, std::set<Label> *paren_set,
-                      std::vector<std::set<StateId>> *state_sets) const;
-
-  const Fst<Arc> &fst_;
-  // Paren IDs to labels.
-  const std::vector<std::pair<Label, Label>> &parens_;
-  // Close/open paren info?
-  const bool close_;
-  // Labels to paren IDs.
-  std::unordered_map<Label, Label> paren_map_;
-  // Paren reachability.
-  ParenMultimap paren_multimap_;
-  // Paren arcs.
-  ParenArcMultimap paren_arc_multimap_;
-  // DFS states.
-  std::vector<uint8> state_color_;
-  // Reachable states to IDs.
-  mutable Collection<ssize_t, StateId> state_sets_;
-  // IDs to reachable states.
-  StateSetMap set_map_;
-  bool error_;
-
-  PdtParenReachable(const PdtParenReachable &) = delete;
-  PdtParenReachable &operator=(const PdtParenReachable &) = delete;
-};
-
-// Gathers paren and state set information.
-template <class Arc>
-bool PdtParenReachable<Arc>::DFSearch(StateId s) {
-  static constexpr uint8 kWhiteState = 0x01;  // Undiscovered.
-  static constexpr uint8 kGreyState = 0x02;   // Discovered & unfinished.
-  static constexpr uint8 kBlackState = 0x04;  // Finished.
-  if (s >= state_color_.size()) state_color_.resize(s + 1, kWhiteState);
-  if (state_color_[s] == kBlackState) return true;
-  if (state_color_[s] == kGreyState) return false;
-  state_color_[s] = kGreyState;
-  for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
-    const auto &arc = aiter.Value();
-    const auto it = paren_map_.find(arc.ilabel);
-    if (it != paren_map_.end()) {  // Paren?
-      const auto paren_id = it->second;
-      if (arc.ilabel == parens_[paren_id].first) {  // Open paren?
-        if (!DFSearch(arc.nextstate)) return false;
-        for (auto set_iter = FindStates(paren_id, arc.nextstate);
-             !set_iter.Done(); set_iter.Next()) {
-          for (auto paren_arc_iter =
-                   FindParenArcs(paren_id, set_iter.Element());
-               !paren_arc_iter.Done(); paren_arc_iter.Next()) {
-            const auto &cparc = paren_arc_iter.Value();
-            if (!DFSearch(cparc.nextstate)) return false;
-          }
-        }
-      }
-    } else if (!DFSearch(arc.nextstate)) {  // Non-paren.
-      return false;
-    }
-  }
-  ComputeStateSet(s);
-  state_color_[s] = kBlackState;
-  return true;
-}
-
-// Unions state sets.
-template <class Arc>
-void PdtParenReachable<Arc>::ComputeStateSet(StateId s) {
-  std::set<Label> paren_set;
-  std::vector<std::set<StateId>> state_sets(parens_.size());
-  for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
-    const auto &arc = aiter.Value();
-    const auto it = paren_map_.find(arc.ilabel);
-    if (it != paren_map_.end()) {  // Paren?
-      const auto paren_id = it->second;
-      if (arc.ilabel == parens_[paren_id].first) {  // Open paren?
-        for (auto set_iter = FindStates(paren_id, arc.nextstate);
-             !set_iter.Done(); set_iter.Next()) {
-          for (auto paren_arc_iter =
-                   FindParenArcs(paren_id, set_iter.Element());
-               !paren_arc_iter.Done(); paren_arc_iter.Next()) {
-            const auto &cparc = paren_arc_iter.Value();
-            UpdateStateSet(cparc.nextstate, &paren_set, &state_sets);
-          }
-        }
-      } else {  // Close paren.
-        paren_set.insert(paren_id);
-        state_sets[paren_id].insert(s);
-        const State paren_state(paren_id, s);
-        paren_arc_multimap_.insert(std::make_pair(paren_state, arc));
-      }
-    } else {  // Non-paren.
-      UpdateStateSet(arc.nextstate, &paren_set, &state_sets);
-    }
-  }
-  std::vector<StateId> state_set;
-  for (auto paren_iter = paren_set.begin(); paren_iter != paren_set.end();
-       ++paren_iter) {
-    state_set.clear();
-    const auto paren_id = *paren_iter;
-    paren_multimap_.insert(std::make_pair(s, paren_id));
-    for (auto state_iter = state_sets[paren_id].begin();
-         state_iter != state_sets[paren_id].end(); ++state_iter) {
-      state_set.push_back(*state_iter);
-    }
-    const State paren_state(paren_id, s);
-    set_map_[paren_state] = state_sets_.FindId(state_set);
-  }
-}
-
-// Gathers state sets.
-template <class Arc>
-void PdtParenReachable<Arc>::UpdateStateSet(
-    StateId nextstate, std::set<Label> *paren_set,
-    std::vector<std::set<StateId>> *state_sets) const {
-  for (auto paren_iter = FindParens(nextstate); !paren_iter.Done();
-       paren_iter.Next()) {
-    const auto paren_id = paren_iter.Value();
-    paren_set->insert(paren_id);
-    for (auto set_iter = FindStates(paren_id, nextstate); !set_iter.Done();
-         set_iter.Next()) {
-      (*state_sets)[paren_id].insert(set_iter.Element());
-    }
-  }
-}
-
-// Stores balancing parenthesis data for a PDT. Unlike PdtParenReachable above
-// this allows on-the-fly construction (e.g., in PdtShortestPath).
-template <class Arc>
-class PdtBalanceData {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-
-  using State = ParenState<Arc>;
-  using StateHash = typename State::Hash;
-
-  // Set for open parens.
-  using OpenParenSet = std::unordered_set<State, StateHash>;
-
-  // Maps from open paren destination state to parenthesis ID.
-  using OpenParenMap = std::unordered_multimap<StateId, Label>;
-
-  // Maps from open paren state to source states of matching close parens
-  using CloseParenMap = std::unordered_multimap<State, StateId, StateHash>;
-
-  // Maps from open paren state to close source set ID.
-  using CloseSourceMap = std::unordered_map<State, ssize_t, StateHash>;
-
-  using SetIterator = typename Collection<ssize_t, StateId>::SetIterator;
-
-  PdtBalanceData() {}
-
-  void Clear() {
-    open_paren_map_.clear();
-    close_paren_map_.clear();
-  }
-
-  // Adds an open parenthesis with destination state open_dest.
-  void OpenInsert(Label paren_id, StateId open_dest) {
-    const State key(paren_id, open_dest);
-    if (!open_paren_set_.count(key)) {
-      open_paren_set_.insert(key);
-      open_paren_map_.emplace(open_dest, paren_id);
-    }
-  }
-
-  // Adds a matching closing parenthesis with source state close_source
-  // balancing an open_parenthesis with destination state open_dest if
-  // OpenInsert() previously called.
-  void CloseInsert(Label paren_id, StateId open_dest, StateId close_source) {
-    const State key(paren_id, open_dest);
-    if (open_paren_set_.count(key)) {
-      close_paren_map_.emplace(key, close_source);
-    }
-  }
-
-  // Finds close paren source states matching an open parenthesis. The following
-  // methods are then used to iterate through those matching states. Should be
-  // called only after FinishInsert(open_dest).
-  SetIterator Find(Label paren_id, StateId open_dest) {
-    const State key(paren_id, open_dest);
-    const auto it = close_source_map_.find(key);
-    if (it == close_source_map_.end()) {
-      return close_source_sets_.FindSet(-1);
-    } else {
-      return close_source_sets_.FindSet(it->second);
-    }
-  }
-
-  // Called when all open and close parenthesis insertions (w.r.t. open
-  // parentheses entering state open_dest) are finished. Must be called before
-  // Find(open_dest).
-  void FinishInsert(StateId open_dest) {
-    std::vector<StateId> close_sources;
-    for (auto oit = open_paren_map_.find(open_dest);
-         oit != open_paren_map_.end() && oit->first == open_dest;) {
-      const auto paren_id = oit->second;
-      close_sources.clear();
-      const State key(paren_id, open_dest);
-      open_paren_set_.erase(open_paren_set_.find(key));
-      for (auto cit = close_paren_map_.find(key);
-           cit != close_paren_map_.end() && cit->first == key;) {
-        close_sources.push_back(cit->second);
-        close_paren_map_.erase(cit++);
-      }
-      std::sort(close_sources.begin(), close_sources.end());
-      auto unique_end = std::unique(close_sources.begin(), close_sources.end());
-      close_sources.resize(unique_end - close_sources.begin());
-      if (!close_sources.empty()) {
-        close_source_map_[key] = close_source_sets_.FindId(close_sources);
-      }
-      open_paren_map_.erase(oit++);
-    }
-  }
-
-  // Returns a new balance data object representing the reversed balance
-  // information.
-  PdtBalanceData<Arc> *Reverse(StateId num_states, StateId num_split,
-                               StateId state_id_shift) const;
-
- private:
-  // Open paren at destintation state?
-  OpenParenSet open_paren_set_;
-  // Open parens per state.
-  OpenParenMap open_paren_map_;
-  // Current open destination state.
-  State open_dest_;
-  // Current open paren/state.
-  typename OpenParenMap::const_iterator open_iter_;
-  // Close states to (open paren, state).
-  CloseParenMap close_paren_map_;
-  // (Paren, state) to set ID.
-  CloseSourceMap close_source_map_;
-  mutable Collection<ssize_t, StateId> close_source_sets_;
-};
-
-// Return a new balance data object representing the reversed balance
-// information.
-template <class Arc>
-PdtBalanceData<Arc> *PdtBalanceData<Arc>::Reverse(
-    StateId num_states, StateId num_split, StateId state_id_shift) const {
-  auto *bd = new PdtBalanceData<Arc>;
-  std::unordered_set<StateId> close_sources;
-  const auto split_size = num_states / num_split;
-  for (StateId i = 0; i < num_states; i += split_size) {
-    close_sources.clear();
-    for (auto it = close_source_map_.begin(); it != close_source_map_.end();
-         ++it) {
-      const auto &okey = it->first;
-      const auto open_dest = okey.state_id;
-      const auto paren_id = okey.paren_id;
-      for (auto set_iter = close_source_sets_.FindSet(it->second);
-           !set_iter.Done(); set_iter.Next()) {
-        const auto close_source = set_iter.Element();
-        if ((close_source < i) || (close_source >= i + split_size)) continue;
-        close_sources.insert(close_source + state_id_shift);
-        bd->OpenInsert(paren_id, close_source + state_id_shift);
-        bd->CloseInsert(paren_id, close_source + state_id_shift,
-                        open_dest + state_id_shift);
-      }
-    }
-    for (auto it = close_sources.begin(); it != close_sources.end(); ++it) {
-      bd->FinishInsert(*it);
-    }
-  }
-  return bd;
-}
-
-}  // namespace internal
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_PAREN_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/pdt.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/pdt.h
deleted file mode 100644
index d6c5946f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/pdt.h
+++ /dev/null
@@ -1,165 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Common classes for PDT expansion/traversal.
-
-#ifndef FST_EXTENSIONS_PDT_PDT_H_
-#define FST_EXTENSIONS_PDT_PDT_H_
-
-#include <map>
-#include <set>
-#include <unordered_map>
-
-#include <fst/compat.h>
-#include <fst/log.h>
-#include <fst/fst.h>
-#include <fst/state-table.h>
-
-namespace fst {
-
-// Provides bijection between parenthesis stacks and signed integral stack IDs.
-// Each stack ID is unique to each distinct stack. The open-close parenthesis
-// label pairs are passed using the parens argument.
-template <typename StackId, typename Label>
-class PdtStack {
- public:
-  // The stacks are stored in a tree. The nodes are stored in a vector. Each
-  // node represents the top of some stack and is identified by its position in
-  // the vector. Its' parent node represents the stack with the top popped and
-  // its children are stored in child_map_ and accessed by stack_id and label.
-  // The paren_id is
-  // the position in parens of the parenthesis for that node.
-  struct StackNode {
-    StackId parent_id;
-    size_t paren_id;
-
-    StackNode(StackId p, size_t i) : parent_id(p), paren_id(i) {}
-  };
-
-  explicit PdtStack(const std::vector<std::pair<Label, Label>> &parens)
-      : parens_(parens), min_paren_(kNoLabel), max_paren_(kNoLabel) {
-    for (size_t i = 0; i < parens.size(); ++i) {
-      const auto &pair = parens[i];
-      paren_map_[pair.first] = i;
-      paren_map_[pair.second] = i;
-      if (min_paren_ == kNoLabel || pair.first < min_paren_) {
-        min_paren_ = pair.first;
-      }
-      if (pair.second < min_paren_) min_paren_ = pair.second;
-      if (max_paren_ == kNoLabel || pair.first > max_paren_) {
-        max_paren_ = pair.first;
-      }
-      if (pair.second > max_paren_) max_paren_ = pair.second;
-    }
-    nodes_.push_back(StackNode(-1, -1));  // Tree root.
-  }
-
-  // Returns stack ID given the current stack ID (0 if empty) and label read.
-  // Pushes onto the stack if the label is an open parenthesis, returning the
-  // new stack ID. Pops the stack if the label is a close parenthesis that
-  // matches the top of the stack, returning the parent stack ID. Returns -1 if
-  // label is an unmatched close parenthesis. Otherwise, returns the current
-  // stack ID.
-  StackId Find(StackId stack_id, Label label) {
-    if (min_paren_ == kNoLabel || label < min_paren_ || label > max_paren_) {
-      return stack_id;  // Non-paren.
-    }
-    const auto it = paren_map_.find(label);
-    // Non-paren.
-    if (it == paren_map_.end()) return stack_id;
-    const auto paren_id = it->second;
-    // Open paren.
-    if (label == parens_[paren_id].first) {
-      auto &child_id = child_map_[std::make_pair(stack_id, label)];
-      if (child_id == 0) {  // Child not found; pushes label.
-        child_id = nodes_.size();
-        nodes_.push_back(StackNode(stack_id, paren_id));
-      }
-      return child_id;
-    }
-    const auto &node = nodes_[stack_id];
-    // Matching close paren.
-    if (paren_id == node.paren_id) return node.parent_id;
-    // Non-matching close paren.
-    return -1;
-  }
-
-  // Returns the stack ID obtained by popping the label at the top of the
-  // current stack ID.
-  StackId Pop(StackId stack_id) const { return nodes_[stack_id].parent_id; }
-
-  // Returns the paren ID at the top of the stack.
-  ssize_t Top(StackId stack_id) const { return nodes_[stack_id].paren_id; }
-
-  ssize_t ParenId(Label label) const {
-    const auto it = paren_map_.find(label);
-    if (it == paren_map_.end()) return -1;  // Non-paren.
-    return it->second;
-  }
-
- private:
-  struct ChildHash {
-    size_t operator()(const std::pair<StackId, Label> &pair) const {
-      static constexpr size_t prime = 7853;
-      return static_cast<size_t>(pair.first) +
-             static_cast<size_t>(pair.second) * prime;
-    }
-  };
-
-  std::vector<std::pair<Label, Label>> parens_;
-  std::vector<StackNode> nodes_;
-  std::unordered_map<Label, size_t> paren_map_;
-  // Child of stack node w.r.t label
-  std::unordered_map<std::pair<StackId, Label>, StackId, ChildHash> child_map_;
-  Label min_paren_;
-  Label max_paren_;
-};
-
-// State tuple for PDT expansion.
-template <typename S, typename K>
-struct PdtStateTuple {
-  using StateId = S;
-  using StackId = K;
-
-  StateId state_id;
-  StackId stack_id;
-
-  PdtStateTuple(StateId state_id = kNoStateId, StackId stack_id = -1)
-      : state_id(state_id), stack_id(stack_id) {}
-};
-
-// Equality of PDT state tuples.
-template <typename S, typename K>
-inline bool operator==(const PdtStateTuple<S, K> &x,
-                       const PdtStateTuple<S, K> &y) {
-  if (&x == &y) return true;
-  return x.state_id == y.state_id && x.stack_id == y.stack_id;
-}
-
-// Hash function object for PDT state tuples
-template <class T>
-class PdtStateHash {
- public:
-  size_t operator()(const T &tuple) const {
-    static constexpr auto prime = 7853;
-    return tuple.state_id + tuple.stack_id * prime;
-  }
-};
-
-// Tuple to PDT state bijection.
-template <class StateId, class StackId>
-class PdtStateTable : public CompactHashStateTable<
-                          PdtStateTuple<StateId, StackId>,
-                          PdtStateHash<PdtStateTuple<StateId, StackId>>> {
- public:
-  PdtStateTable() {}
-
-  PdtStateTable(const PdtStateTable &other) {}
-
- private:
-  PdtStateTable &operator=(const PdtStateTable &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_PDT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/pdtlib.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/pdtlib.h
deleted file mode 100644
index 41d98986..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/pdtlib.h
+++ /dev/null
@@ -1,19 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// This is an experimental push-down transducer (PDT) library. A PDT is
-// encoded as an FST, where some transitions are labeled with open or close
-// parentheses. To be interpreted as a PDT, the parentheses must balance on a
-// path.
-
-#ifndef FST_EXTENSIONS_PDT_PDTLIB_H_
-#define FST_EXTENSIONS_PDT_PDTLIB_H_
-
-#include <fst/extensions/pdt/compose.h>
-#include <fst/extensions/pdt/expand.h>
-#include <fst/extensions/pdt/pdt.h>
-#include <fst/extensions/pdt/replace.h>
-#include <fst/extensions/pdt/reverse.h>
-#include <fst/extensions/pdt/shortest-path.h>
-
-#endif  // FST_EXTENSIONS_PDT_PDTLIB_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/pdtscript.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/pdtscript.h
deleted file mode 100644
index bdf5b4ea..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/pdtscript.h
+++ /dev/null
@@ -1,244 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Convenience file for including all PDT operations at once, and/or
-// registering them for new arc types.
-
-#ifndef FST_EXTENSIONS_PDT_PDTSCRIPT_H_
-#define FST_EXTENSIONS_PDT_PDTSCRIPT_H_
-
-#include <algorithm>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-#include <fst/compose.h>  // for ComposeOptions
-#include <fst/util.h>
-
-#include <fst/script/arg-packs.h>
-#include <fst/script/fstscript.h>
-#include <fst/script/shortest-path.h>
-
-#include <fst/extensions/pdt/compose.h>
-#include <fst/extensions/pdt/expand.h>
-#include <fst/extensions/pdt/info.h>
-#include <fst/extensions/pdt/replace.h>
-#include <fst/extensions/pdt/reverse.h>
-#include <fst/extensions/pdt/shortest-path.h>
-
-namespace fst {
-namespace script {
-
-using PdtComposeArgs =
-    std::tuple<const FstClass &, const FstClass &,
-               const std::vector<LabelPair> &, MutableFstClass *,
-               const PdtComposeOptions &, bool>;
-
-template <class Arc>
-void PdtCompose(PdtComposeArgs *args) {
-  const Fst<Arc> &ifst1 = *(std::get<0>(*args).GetFst<Arc>());
-  const Fst<Arc> &ifst2 = *(std::get<1>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<3>(*args)->GetMutableFst<Arc>();
-  // In case Arc::Label is not the same as FstClass::Label, we make a
-  // copy. Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<2>(*args).size());
-  std::copy(std::get<2>(*args).begin(), std::get<2>(*args).end(),
-            typed_parens.begin());
-  if (std::get<5>(*args)) {
-    Compose(ifst1, typed_parens, ifst2, ofst, std::get<4>(*args));
-  } else {
-    Compose(ifst1, ifst2, typed_parens, ofst, std::get<4>(*args));
-  }
-}
-
-void PdtCompose(const FstClass &ifst1, const FstClass &ifst2,
-                const std::vector<LabelPair> &parens,
-                MutableFstClass *ofst, const PdtComposeOptions &opts,
-                bool left_pdt);
-
-struct PdtExpandOptions {
-  bool connect;
-  bool keep_parentheses;
-  const WeightClass &weight_threshold;
-
-  PdtExpandOptions(bool c, bool k, const WeightClass &w)
-      : connect(c), keep_parentheses(k), weight_threshold(w) {}
-};
-
-using PdtExpandArgs =
-    std::tuple<const FstClass &, const std::vector<LabelPair> &,
-               MutableFstClass *, const PdtExpandOptions &>;
-
-template <class Arc>
-void PdtExpand(PdtExpandArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
-  // In case Arc::Label is not the same as FstClass::Label, we make a
-  // copy. Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_parens.begin());
-  Expand(fst, typed_parens, ofst,
-         fst::PdtExpandOptions<Arc>(
-             std::get<3>(*args).connect, std::get<3>(*args).keep_parentheses,
-             *(std::get<3>(*args)
-                   .weight_threshold.GetWeight<typename Arc::Weight>())));
-}
-
-void PdtExpand(const FstClass &ifst, const std::vector<LabelPair> &parens,
-               MutableFstClass *ofst, const PdtExpandOptions &opts);
-
-void PdtExpand(const FstClass &ifst, const std::vector<LabelPair> &parens,
-               MutableFstClass *ofst, bool connect, bool keep_parentheses,
-               const WeightClass &weight_threshold);
-
-using PdtReplaceArgs =
-    std::tuple<const std::vector<LabelFstClassPair> &, MutableFstClass *,
-               std::vector<LabelPair> *, int64, PdtParserType, int64,
-               const string &, const string &>;
-
-template <class Arc>
-void PdtReplace(PdtReplaceArgs *args) {
-  const auto &untyped_pairs = std::get<0>(*args);
-  auto size = untyped_pairs.size();
-  std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>> typed_pairs(
-      size);
-  for (size_t i = 0; i < size; ++i) {
-    typed_pairs[i].first = untyped_pairs[i].first;
-    typed_pairs[i].second = untyped_pairs[i].second->GetFst<Arc>();
-  }
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens;
-  const PdtReplaceOptions<Arc> opts(std::get<3>(*args), std::get<4>(*args),
-                                    std::get<5>(*args), std::get<6>(*args),
-                                    std::get<7>(*args));
-  Replace(typed_pairs, ofst, &typed_parens, opts);
-  // Copies typed parens into arg3.
-  std::get<2>(*args)->resize(typed_parens.size());
-  std::copy(typed_parens.begin(), typed_parens.end(),
-            std::get<2>(*args)->begin());
-}
-
-void PdtReplace(const std::vector<LabelFstClassPair> &pairs,
-                MutableFstClass *ofst, std::vector<LabelPair> *parens,
-                int64 root, PdtParserType parser_type = PDT_LEFT_PARSER,
-                int64 start_paren_labels = kNoLabel,
-                const string &left_paren_prefix = "(_",
-                const string &right_paren_prefix = "_)");
-
-using PdtReverseArgs =
-    std::tuple<const FstClass &, const std::vector<LabelPair> &,
-               MutableFstClass *>;
-
-template <class Arc>
-void PdtReverse(PdtReverseArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
-  // In case Arc::Label is not the same as FstClass::Label, we make a
-  // copy. Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_parens.begin());
-  Reverse(fst, typed_parens, ofst);
-}
-
-void PdtReverse(const FstClass &ifst, const std::vector<LabelPair> &,
-                MutableFstClass *ofst);
-
-// PDT SHORTESTPATH
-
-struct PdtShortestPathOptions {
-  QueueType queue_type;
-  bool keep_parentheses;
-  bool path_gc;
-
-  PdtShortestPathOptions(QueueType qt = FIFO_QUEUE, bool kp = false,
-                         bool gc = true)
-      : queue_type(qt), keep_parentheses(kp), path_gc(gc) {}
-};
-
-using PdtShortestPathArgs =
-    std::tuple<const FstClass &, const std::vector<LabelPair> &,
-               MutableFstClass *, const PdtShortestPathOptions &>;
-
-template <class Arc>
-void PdtShortestPath(PdtShortestPathArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
-  const PdtShortestPathOptions &opts = std::get<3>(*args);
-  // In case Arc::Label is not the same as FstClass::Label, we make a
-  // copy. Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_parens.begin());
-  switch (opts.queue_type) {
-    default:
-      FSTERROR() << "Unknown queue type: " << opts.queue_type;
-    case FIFO_QUEUE: {
-      using Queue = FifoQueue<typename Arc::StateId>;
-      fst::PdtShortestPathOptions<Arc, Queue> spopts(opts.keep_parentheses,
-                                                         opts.path_gc);
-      ShortestPath(fst, typed_parens, ofst, spopts);
-      return;
-    }
-    case LIFO_QUEUE: {
-      using Queue = LifoQueue<typename Arc::StateId>;
-      fst::PdtShortestPathOptions<Arc, Queue> spopts(opts.keep_parentheses,
-                                                         opts.path_gc);
-      ShortestPath(fst, typed_parens, ofst, spopts);
-      return;
-    }
-    case STATE_ORDER_QUEUE: {
-      using Queue = StateOrderQueue<typename Arc::StateId>;
-      fst::PdtShortestPathOptions<Arc, Queue> spopts(opts.keep_parentheses,
-                                                         opts.path_gc);
-      ShortestPath(fst, typed_parens, ofst, spopts);
-      return;
-    }
-  }
-}
-
-void PdtShortestPath(const FstClass &ifst,
-    const std::vector<LabelPair> &parens, MutableFstClass *ofst,
-    const PdtShortestPathOptions &opts = PdtShortestPathOptions());
-
-// PRINT INFO
-
-using PrintPdtInfoArgs =
-    std::pair<const FstClass &, const std::vector<LabelPair> &>;
-
-template <class Arc>
-void PrintPdtInfo(PrintPdtInfoArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  // In case Arc::Label is not the same as FstClass::Label, we make a
-  // copy. Truncation may occur if FstClass::Label has more precision than
-  // Arc::Label.
-  std::vector<std::pair<typename Arc::Label, typename Arc::Label>> typed_parens(
-      std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_parens.begin());
-  PdtInfo<Arc> pdtinfo(fst, typed_parens);
-  PrintPdtInfo(pdtinfo);
-}
-
-void PrintPdtInfo(const FstClass &ifst, const std::vector<LabelPair> &parens);
-
-}  // namespace script
-}  // namespace fst
-
-#define REGISTER_FST_PDT_OPERATIONS(ArcType)                             \
-  REGISTER_FST_OPERATION(PdtCompose, ArcType, PdtComposeArgs);           \
-  REGISTER_FST_OPERATION(PdtExpand, ArcType, PdtExpandArgs);             \
-  REGISTER_FST_OPERATION(PdtReplace, ArcType, PdtReplaceArgs);           \
-  REGISTER_FST_OPERATION(PdtReverse, ArcType, PdtReverseArgs);           \
-  REGISTER_FST_OPERATION(PdtShortestPath, ArcType, PdtShortestPathArgs); \
-  REGISTER_FST_OPERATION(PrintPdtInfo, ArcType, PrintPdtInfoArgs)
-#endif  // FST_EXTENSIONS_PDT_PDTSCRIPT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/replace.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/replace.h
deleted file mode 100644
index 76341e8d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/replace.h
+++ /dev/null
@@ -1,827 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Recursively replaces FST arcs with other FSTs, returning a PDT.
-
-#ifndef FST_EXTENSIONS_PDT_REPLACE_H_
-#define FST_EXTENSIONS_PDT_REPLACE_H_
-
-#include <map>
-#include <memory>
-#include <set>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/replace.h>
-#include <fst/replace-util.h>
-#include <fst/symbol-table-ops.h>
-
-namespace fst {
-namespace internal {
-
-// Hash to paren IDs
-template <typename S>
-struct ReplaceParenHash {
-  size_t operator()(const std::pair<size_t, S> &paren) const {
-    static constexpr auto prime = 7853;
-    return paren.first + paren.second * prime;
-  }
-};
-
-}  // namespace internal
-
-// Parser types characterize the PDT construction method. When applied to a CFG,
-// each non-terminal is encoded as a DFA that accepts precisely the RHS's of
-// productions of that non-terminal. For parsing (rather than just recognition),
-// production numbers can used as outputs (placed as early as possible) in the
-// DFAs promoted to DFTs. For more information on the strongly regular
-// construction, see:
-//
-// Mohri, M., and Pereira, F. 1998. Dynamic compilation of weighted context-free
-// grammars. In Proc. ACL, pages 891-897.
-enum PdtParserType {
-  // Top-down construction. Applied to a simple LL(1) grammar (among others),
-  // gives a DPDA. If promoted to a DPDT, with outputs being production
-  // numbers, gives a leftmost derivation. Left recursive grammars are
-  // problematic in use.
-  PDT_LEFT_PARSER,
-
-  // Top-down construction. Similar to PDT_LEFT_PARSE except bounded-stack
-  // (expandable as an FST) result with regular or, more generally, strongly
-  // regular grammars. Epsilons may replace some parentheses, which may
-  // introduce some non-determinism.
-  PDT_LEFT_SR_PARSER,
-
-  /* TODO(riley):
-  // Bottom-up construction. Applied to a LR(0) grammar, gives a DPDA.
-  // If promoted to a DPDT, with outputs being the production nubmers,
-  // gives the reverse of a rightmost derivation.
-  PDT_RIGHT_PARSER,
-  */
-};
-
-template <class Arc>
-struct PdtReplaceOptions {
-  using Label = typename Arc::Label;
-
-  explicit PdtReplaceOptions(Label root,
-                             PdtParserType type = PDT_LEFT_PARSER,
-                             Label start_paren_labels = kNoLabel,
-                             string left_paren_prefix = "(_",
-                             string right_paren_prefix = ")_") :
-      root(root), type(type), start_paren_labels(start_paren_labels),
-      left_paren_prefix(std::move(left_paren_prefix)),
-      right_paren_prefix(std::move(right_paren_prefix)) {}
-
-  Label root;
-  PdtParserType type;
-  Label start_paren_labels;
-  const string left_paren_prefix;
-  const string right_paren_prefix;
-};
-
-// PdtParser: Base PDT parser class common to specific parsers.
-
-template <class Arc>
-class PdtParser {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using LabelFstPair = std::pair<Label, const Fst<Arc> *>;
-  using LabelPair = std::pair<Label, Label>;
-  using LabelStatePair = std::pair<Label, StateId>;
-  using StateWeightPair = std::pair<StateId, Weight>;
-  using ParenKey = std::pair<size_t, StateId>;
-  using ParenMap =
-      std::unordered_map<ParenKey, size_t, internal::ReplaceParenHash<StateId>>;
-
-  PdtParser(const std::vector<LabelFstPair> &fst_array,
-            const PdtReplaceOptions<Arc> &opts) :
-      root_(opts.root), start_paren_labels_(opts.start_paren_labels),
-      left_paren_prefix_(std::move(opts.left_paren_prefix)),
-      right_paren_prefix_(std::move(opts.right_paren_prefix)),
-      error_(false) {
-    for (size_t i = 0; i < fst_array.size(); ++i) {
-      if (!CompatSymbols(fst_array[0].second->InputSymbols(),
-                         fst_array[i].second->InputSymbols())) {
-        FSTERROR() << "PdtParser: Input symbol table of input FST " << i
-                   << " does not match input symbol table of 0th input FST";
-        error_ = true;
-      }
-      if (!CompatSymbols(fst_array[0].second->OutputSymbols(),
-                         fst_array[i].second->OutputSymbols())) {
-        FSTERROR() << "PdtParser: Output symbol table of input FST " << i
-                   << " does not match input symbol table of 0th input FST";
-        error_ = true;
-      }
-      fst_array_.emplace_back(fst_array[i].first, fst_array[i].second->Copy());
-      // Builds map from non-terminal label to FST ID.
-      label2id_[fst_array[i].first] = i;
-    }
-  }
-
-  virtual ~PdtParser() {
-    for (auto &pair : fst_array_) delete pair.second;
-  }
-
-  // Constructs the output PDT, dependent on the derived parser type.
-  virtual void GetParser(MutableFst<Arc> *ofst,
-                         std::vector<LabelPair> *parens) = 0;
-
- protected:
-  const std::vector<LabelFstPair> &FstArray() const { return fst_array_; }
-
-  Label Root() const { return root_; }
-
-  // Maps from non-terminal label to corresponding FST ID, or returns
-  // kNoStateId to signal lookup failure.
-  StateId Label2Id(Label l) const {
-    auto it = label2id_.find(l);
-    return it == label2id_.end() ? kNoStateId : it->second;
-  }
-
-  // Maps from output state to input FST label, state pair, or returns a
-  // (kNoLabel, kNoStateId) pair to signal lookup failure.
-  LabelStatePair GetLabelStatePair(StateId os) const {
-    if (os >= label_state_pairs_.size()) {
-      static const LabelStatePair no_pair(kNoLabel, kNoLabel);
-      return no_pair;
-    } else {
-      return label_state_pairs_[os];
-    }
-  }
-
-  // Maps to output state from input FST (label, state) pair, or returns
-  // kNoStateId to signal lookup failure.
-  StateId GetState(const LabelStatePair &lsp) const {
-    auto it = state_map_.find(lsp);
-    if (it == state_map_.end()) {
-      return kNoStateId;
-    } else {
-      return it->second;
-    }
-  }
-
-  // Builds single FST combining all referenced input FSTs, leaving in the
-  // non-termnals for now; also tabulates the PDT states that correspond to the
-  // start and final states of the input FSTs.
-  void CreateFst(MutableFst<Arc> *ofst, std::vector<StateId> *open_dest,
-                 std::vector<std::vector<StateWeightPair>> *close_src);
-
-  // Assigns parenthesis labels from total allocated paren IDs.
-  void AssignParenLabels(size_t total_nparens, std::vector<LabelPair> *parens) {
-    parens->clear();
-    for (size_t paren_id = 0; paren_id < total_nparens; ++paren_id) {
-      const auto open_paren = start_paren_labels_ + paren_id;
-      const auto close_paren = open_paren + total_nparens;
-      parens->emplace_back(open_paren, close_paren);
-    }
-  }
-
-  // Determines how non-terminal instances are assigned parentheses IDs.
-  virtual size_t AssignParenIds(const Fst<Arc> &ofst,
-                                ParenMap *paren_map) const = 0;
-
-  // Changes a non-terminal transition to an open parenthesis transition
-  // redirected to the PDT state specified in the open_dest argument, when
-  // indexed by the input FST ID for the non-terminal. Adds close parenthesis
-  // transitions (with specified weights) from the PDT states specified in the
-  // close_src argument, when indexed by the input FST ID for the non-terminal,
-  // to the former destination state of the non-terminal transition. The
-  // paren_map argument gives the parenthesis ID for a given non-terminal FST ID
-  // and destination state pair. The close_non_term_weight vector specifies
-  // non-terminals for which the non-terminal arc weight should be applied on
-  // the close parenthesis (multiplying the close_src weight above) rather than
-  // on the open parenthesis. If no paren ID is found, then an epsilon replaces
-  // the parenthesis that would carry the non-terminal arc weight and the other
-  // parenthesis is omitted (appropriate for the strongly-regular case).
-  void AddParensToFst(
-      const std::vector<LabelPair> &parens,
-      const ParenMap &paren_map,
-      const std::vector<StateId> &open_dest,
-      const std::vector<std::vector<StateWeightPair>> &close_src,
-      const std::vector<bool> &close_non_term_weight,
-      MutableFst<Arc> *ofst);
-
-  // Ensures that parentheses arcs are added to the symbol table.
-  void AddParensToSymbolTables(const std::vector<LabelPair> &parens,
-                               MutableFst<Arc> *ofst);
-
- private:
-  std::vector<LabelFstPair> fst_array_;
-  Label root_;
-  // Index to use for the first parenthesis.
-  Label start_paren_labels_;
-  const string left_paren_prefix_;
-  const string right_paren_prefix_;
-  // Maps from non-terminal label to FST ID.
-  std::unordered_map<Label, StateId> label2id_;
-  // Given an output state, specifies the input FST (label, state) pair.
-  std::vector<LabelStatePair> label_state_pairs_;
-  // Given an FST (label, state) pair, specifies the output FST state ID.
-  std::map<LabelStatePair, StateId> state_map_;
-  bool error_;
-};
-
-template <class Arc>
-void PdtParser<Arc>::CreateFst(
-    MutableFst<Arc> *ofst, std::vector<StateId> *open_dest,
-    std::vector<std::vector<StateWeightPair>> *close_src) {
-  ofst->DeleteStates();
-  if (error_) {
-    ofst->SetProperties(kError, kError);
-    return;
-  }
-  open_dest->resize(fst_array_.size(), kNoStateId);
-  close_src->resize(fst_array_.size());
-  // Queue of non-terminals to replace.
-  std::deque<Label> non_term_queue;
-  non_term_queue.push_back(root_);
-  // Has a non-terminal been enqueued?
-  std::vector<bool> enqueued(fst_array_.size(), false);
-  enqueued[label2id_[root_]] = true;
-  Label max_label = kNoLabel;
-  for (StateId soff = 0; !non_term_queue.empty(); soff = ofst->NumStates()) {
-    const auto label = non_term_queue.front();
-    non_term_queue.pop_front();
-    StateId fst_id = Label2Id(label);
-    const auto *ifst = fst_array_[fst_id].second;
-    for (StateIterator<Fst<Arc>> siter(*ifst); !siter.Done(); siter.Next()) {
-      const auto is = siter.Value();
-      const auto os = ofst->AddState();
-      const LabelStatePair lsp(label, is);
-      label_state_pairs_.push_back(lsp);
-      state_map_[lsp] = os;
-      if (is == ifst->Start()) {
-        (*open_dest)[fst_id] = os;
-        if (label == root_) ofst->SetStart(os);
-      }
-      if (ifst->Final(is) != Weight::Zero()) {
-        if (label == root_) ofst->SetFinal(os, ifst->Final(is));
-        (*close_src)[fst_id].emplace_back(os, ifst->Final(is));
-      }
-      for (ArcIterator<Fst<Arc>> aiter(*ifst, is); !aiter.Done();
-           aiter.Next()) {
-        auto arc = aiter.Value();
-        arc.nextstate += soff;
-        if (max_label == kNoLabel || arc.olabel > max_label)
-          max_label = arc.olabel;
-        const auto nfst_id = Label2Id(arc.olabel);
-        if (nfst_id != kNoStateId) {
-          if (fst_array_[nfst_id].second->Start() == kNoStateId) continue;
-          if (!enqueued[nfst_id]) {
-            non_term_queue.push_back(arc.olabel);
-            enqueued[nfst_id] = true;
-          }
-        }
-        ofst->AddArc(os, arc);
-      }
-    }
-  }
-  if (start_paren_labels_ == kNoLabel)
-    start_paren_labels_ = max_label + 1;
-}
-
-template <class Arc>
-void PdtParser<Arc>::AddParensToFst(
-    const std::vector<LabelPair> &parens,
-    const ParenMap &paren_map,
-    const std::vector<StateId> &open_dest,
-    const std::vector<std::vector<StateWeightPair>> &close_src,
-    const std::vector<bool> &close_non_term_weight,
-    MutableFst<Arc> *ofst) {
-  StateId dead_state = kNoStateId;
-  using MIter = MutableArcIterator<MutableFst<Arc>>;
-  for (StateIterator<Fst<Arc>> siter(*ofst); !siter.Done(); siter.Next()) {
-    StateId os = siter.Value();
-    std::unique_ptr<MIter> aiter(new MIter(ofst, os));
-    for (auto n = 0; !aiter->Done(); aiter->Next(), ++n) {
-      const auto arc = aiter->Value();  // A reference here may go stale.
-      StateId nfst_id = Label2Id(arc.olabel);
-      if (nfst_id != kNoStateId) {
-        // Gets parentheses.
-        const ParenKey paren_key(nfst_id, arc.nextstate);
-        auto it = paren_map.find(paren_key);
-        Label open_paren = 0;
-        Label close_paren = 0;
-        if (it != paren_map.end()) {
-          const auto paren_id = it->second;
-          open_paren = parens[paren_id].first;
-          close_paren = parens[paren_id].second;
-        }
-        // Sets open parenthesis.
-        if (open_paren != 0 || !close_non_term_weight[nfst_id]) {
-          const auto open_weight =
-              close_non_term_weight[nfst_id] ? Weight::One() : arc.weight;
-          const Arc sarc(open_paren, open_paren, open_weight,
-                         open_dest[nfst_id]);
-          aiter->SetValue(sarc);
-        } else {
-          if (dead_state == kNoStateId) {
-            dead_state = ofst->AddState();
-          }
-          const Arc sarc(0, 0, Weight::One(), dead_state);
-          aiter->SetValue(sarc);
-        }
-        // Adds close parentheses.
-        if (close_paren != 0 || close_non_term_weight[nfst_id]) {
-          for (size_t i = 0; i < close_src[nfst_id].size(); ++i) {
-            const auto &pair = close_src[nfst_id][i];
-            const auto close_weight = close_non_term_weight[nfst_id]
-                                          ? Times(arc.weight, pair.second)
-                                          : pair.second;
-            const Arc farc(close_paren, close_paren, close_weight,
-                           arc.nextstate);
-
-            ofst->AddArc(pair.first, farc);
-            if (os == pair.first) {  // Invalidated iterator.
-              aiter.reset(new MIter(ofst, os));
-              aiter->Seek(n);
-            }
-          }
-        }
-      }
-    }
-  }
-}
-
-template <class Arc>
-void PdtParser<Arc>::AddParensToSymbolTables(
-    const std::vector<LabelPair> &parens, MutableFst<Arc> *ofst) {
-  auto size = parens.size();
-  if (ofst->InputSymbols()) {
-    if (!AddAuxiliarySymbols(left_paren_prefix_, start_paren_labels_, size,
-                             ofst->MutableInputSymbols())) {
-      ofst->SetProperties(kError, kError);
-      return;
-    }
-    if (!AddAuxiliarySymbols(right_paren_prefix_, start_paren_labels_ + size,
-                             size, ofst->MutableInputSymbols())) {
-      ofst->SetProperties(kError, kError);
-      return;
-    }
-  }
-  if (ofst->OutputSymbols()) {
-    if (!AddAuxiliarySymbols(left_paren_prefix_, start_paren_labels_, size,
-                             ofst->MutableOutputSymbols())) {
-      ofst->SetProperties(kError, kError);
-      return;
-    }
-    if (!AddAuxiliarySymbols(right_paren_prefix_, start_paren_labels_ + size,
-                             size, ofst->MutableOutputSymbols())) {
-      ofst->SetProperties(kError, kError);
-      return;
-    }
-  }
-}
-
-// Builds a PDT by recursive replacement top-down, where the call and return are
-// encoded in the parentheses.
-template <class Arc>
-class PdtLeftParser final : public PdtParser<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using LabelFstPair = typename PdtParser<Arc>::LabelFstPair;
-  using LabelPair = typename PdtParser<Arc>::LabelPair;
-  using LabelStatePair = typename PdtParser<Arc>::LabelStatePair;
-  using StateWeightPair = typename PdtParser<Arc>::StateWeightPair;
-  using ParenKey = typename PdtParser<Arc>::ParenKey;
-  using ParenMap = typename PdtParser<Arc>::ParenMap;
-
-  using PdtParser<Arc>::AddParensToFst;
-  using PdtParser<Arc>::AddParensToSymbolTables;
-  using PdtParser<Arc>::AssignParenLabels;
-  using PdtParser<Arc>::CreateFst;
-  using PdtParser<Arc>::FstArray;
-  using PdtParser<Arc>::GetLabelStatePair;
-  using PdtParser<Arc>::GetState;
-  using PdtParser<Arc>::Label2Id;
-  using PdtParser<Arc>::Root;
-
-  PdtLeftParser(const std::vector<LabelFstPair> &fst_array,
-                const PdtReplaceOptions<Arc> &opts) :
-      PdtParser<Arc>(fst_array, opts) { }
-
-  void GetParser(MutableFst<Arc> *ofst,
-                 std::vector<LabelPair> *parens) override;
-
- protected:
-  // Assigns a unique parenthesis ID for each non-terminal, destination
-  // state pair.
-  size_t AssignParenIds(const Fst<Arc> &ofst,
-                        ParenMap *paren_map) const override;
-};
-
-template <class Arc>
-void PdtLeftParser<Arc>::GetParser(
-    MutableFst<Arc> *ofst,
-    std::vector<LabelPair> *parens) {
-  ofst->DeleteStates();
-  parens->clear();
-  const auto &fst_array = FstArray();
-  // Map that gives the paren ID for a (non-terminal, dest. state) pair
-  // (which can be unique).
-  ParenMap paren_map;
-  // Specifies the open parenthesis destination state for a given non-terminal.
-  // The source is the non-terminal instance source state.
-  std::vector<StateId> open_dest(fst_array.size(), kNoStateId);
-  // Specifies close parenthesis source states and weights for a given
-  // non-terminal. The destination is the non-terminal instance destination
-  // state.
-  std::vector<std::vector<StateWeightPair>> close_src(fst_array.size());
-  // Specifies non-terminals for which the non-terminal arc weight
-  // should be applied on the close parenthesis (multiplying the
-  // 'close_src' weight above) rather than on the open parenthesis.
-  std::vector<bool> close_non_term_weight(fst_array.size(), false);
-  CreateFst(ofst, &open_dest, &close_src);
-  auto total_nparens = AssignParenIds(*ofst, &paren_map);
-  AssignParenLabels(total_nparens, parens);
-  AddParensToFst(*parens, paren_map, open_dest, close_src,
-                 close_non_term_weight, ofst);
-  if (!fst_array.empty()) {
-    ofst->SetInputSymbols(fst_array[0].second->InputSymbols());
-    ofst->SetOutputSymbols(fst_array[0].second->OutputSymbols());
-  }
-  AddParensToSymbolTables(*parens, ofst);
-}
-
-template <class Arc>
-size_t PdtLeftParser<Arc>::AssignParenIds(
-    const Fst<Arc> &ofst,
-    ParenMap *paren_map) const {
-  // Number of distinct parenthesis pairs per FST.
-  std::vector<size_t> nparens(FstArray().size(), 0);
-  // Number of distinct parenthesis pairs overall.
-  size_t total_nparens = 0;
-  for (StateIterator<Fst<Arc>> siter(ofst); !siter.Done(); siter.Next()) {
-    const auto os = siter.Value();
-    for (ArcIterator<Fst<Arc>> aiter(ofst, os); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto nfst_id = Label2Id(arc.olabel);
-      if (nfst_id != kNoStateId) {
-        const ParenKey paren_key(nfst_id, arc.nextstate);
-        auto it = paren_map->find(paren_key);
-        if (it == paren_map->end()) {
-          // Assigns new paren ID for this (FST, dest state) pair.
-          (*paren_map)[paren_key] = nparens[nfst_id]++;
-          if (nparens[nfst_id] > total_nparens)
-            total_nparens = nparens[nfst_id];
-        }
-      }
-    }
-  }
-  return total_nparens;
-}
-
-// Similar to PdtLeftParser but:
-//
-// 1. Uses epsilons rather than parentheses labels for any non-terminal
-//    instances within a left- (right-) linear dependency SCC,
-// 2. Allocates a paren ID uniquely for each such dependency SCC (rather than
-//    non-terminal = dependency state) and destination state.
-template <class Arc>
-class PdtLeftSRParser final : public PdtParser<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using LabelFstPair = typename PdtParser<Arc>::LabelFstPair;
-  using LabelPair = typename PdtParser<Arc>::LabelPair;
-  using LabelStatePair = typename PdtParser<Arc>::LabelStatePair;
-  using StateWeightPair = typename PdtParser<Arc>::StateWeightPair;
-  using ParenKey = typename PdtParser<Arc>::ParenKey;
-  using ParenMap = typename PdtParser<Arc>::ParenMap;
-
-  using PdtParser<Arc>::AddParensToFst;
-  using PdtParser<Arc>::AddParensToSymbolTables;
-  using PdtParser<Arc>::AssignParenLabels;
-  using PdtParser<Arc>::CreateFst;
-  using PdtParser<Arc>::FstArray;
-  using PdtParser<Arc>::GetLabelStatePair;
-  using PdtParser<Arc>::GetState;
-  using PdtParser<Arc>::Label2Id;
-  using PdtParser<Arc>::Root;
-
-  PdtLeftSRParser(const std::vector<LabelFstPair> &fst_array,
-                  const PdtReplaceOptions<Arc> &opts) :
-      PdtParser<Arc>(fst_array, opts),
-      replace_util_(fst_array, ReplaceUtilOptions(opts.root)) { }
-
-  void GetParser(MutableFst<Arc> *ofst,
-                 std::vector<LabelPair> *parens) override;
-
- protected:
-  // Assigns a unique parenthesis ID for each non-terminal, destination state
-  // pair when the non-terminal refers to a non-linear FST. Otherwise, assigns
-  // a unique parenthesis ID for each dependency SCC, destination state pair if
-  // the non-terminal instance is between
-  // SCCs. Otherwise does nothing.
-  size_t AssignParenIds(const Fst<Arc> &ofst,
-                        ParenMap *paren_map) const override;
-
-  // Returns dependency SCC for given label.
-  size_t SCC(Label label) const { return replace_util_.SCC(label); }
-
-  // Is a given dependency SCC left-linear?
-  bool SCCLeftLinear(size_t scc_id) const {
-    const auto ll_props = kReplaceSCCLeftLinear | kReplaceSCCNonTrivial;
-    const auto scc_props = replace_util_.SCCProperties(scc_id);
-    return (scc_props & ll_props) == ll_props;
-  }
-
-  // Is a given dependency SCC right-linear?
-  bool SCCRightLinear(size_t scc_id) const {
-    const auto lr_props = kReplaceSCCRightLinear | kReplaceSCCNonTrivial;
-    const auto scc_props = replace_util_.SCCProperties(scc_id);
-    return (scc_props & lr_props) == lr_props;
-  }
-
-  // Components of left- (right-) linear dependency SCC; empty o.w.
-  const std::vector<size_t> &SCCComps(size_t scc_id) const {
-    if (scc_comps_.empty()) GetSCCComps();
-    return scc_comps_[scc_id];
-  }
-
-  // Returns the representative state of an SCC. For left-linear grammars, it
-  // is one of the initial states. For right-linear grammars, it is one of the
-  // non-terminal destination states; otherwise, it is kNoStateId.
-  StateId RepState(size_t scc_id) const {
-    if (SCCComps(scc_id).empty()) return kNoStateId;
-    const auto fst_id = SCCComps(scc_id).front();
-    const auto &fst_array = FstArray();
-    const auto label = fst_array[fst_id].first;
-    const auto *ifst = fst_array[fst_id].second;
-    if (SCCLeftLinear(scc_id)) {
-      const LabelStatePair lsp(label, ifst->Start());
-      return GetState(lsp);
-    } else {  // Right-linear.
-      const LabelStatePair lsp(label, *NonTermDests(fst_id).begin());
-      return GetState(lsp);
-    }
-    return kNoStateId;
-  }
-
- private:
-  // Merges initial (final) states of in a left- (right-) linear dependency SCC
-  // after dealing with the non-terminal arc and final weights.
-  void ProcSCCs(MutableFst<Arc> *ofst,
-                std::vector<StateId> *open_dest,
-                std::vector<std::vector<StateWeightPair>> *close_src,
-                std::vector<bool> *close_non_term_weight) const;
-
-  // Computes components of left- (right-) linear dependency SCC.
-  void GetSCCComps() const {
-    const std::vector<LabelFstPair> &fst_array = FstArray();
-    for (size_t i = 0; i < fst_array.size(); ++i) {
-      const auto label = fst_array[i].first;
-      const auto scc_id = SCC(label);
-      if (scc_comps_.size() <= scc_id) scc_comps_.resize(scc_id + 1);
-      if (SCCLeftLinear(scc_id) || SCCRightLinear(scc_id)) {
-        scc_comps_[scc_id].push_back(i);
-      }
-    }
-  }
-
-  const std::set<StateId> &NonTermDests(StateId fst_id) const {
-    if (non_term_dests_.empty()) GetNonTermDests();
-    return non_term_dests_[fst_id];
-  }
-
-  // Finds non-terminal destination states for right-linear FSTS, or does
-  // nothing if not found.
-  void GetNonTermDests() const;
-
-  // Dependency SCC info.
-  mutable ReplaceUtil<Arc> replace_util_;
-  // Components of left- (right-) linear dependency SCCs, or empty otherwise.
-  mutable std::vector<std::vector<size_t>> scc_comps_;
-  // States that have non-terminals entering them for each (right-linear) FST.
-  mutable std::vector<std::set<StateId>> non_term_dests_;
-};
-
-template <class Arc>
-void PdtLeftSRParser<Arc>::GetParser(
-    MutableFst<Arc> *ofst,
-    std::vector<LabelPair> *parens) {
-  ofst->DeleteStates();
-  parens->clear();
-  const auto &fst_array = FstArray();
-  // Map that gives the paren ID for a (non-terminal, dest. state) pair.
-  ParenMap paren_map;
-  // Specifies the open parenthesis destination state for a given non-terminal.
-  // The source is the non-terminal instance source state.
-  std::vector<StateId> open_dest(fst_array.size(), kNoStateId);
-  // Specifies close parenthesis source states and weights for a given
-  // non-terminal. The destination is the non-terminal instance destination
-  // state.
-  std::vector<std::vector<StateWeightPair>> close_src(fst_array.size());
-  // Specifies non-terminals for which the non-terminal arc weight should be
-  // applied on the close parenthesis (multiplying the close_src weight above)
-  // rather than on the open parenthesis.
-  std::vector<bool> close_non_term_weight(fst_array.size(), false);
-  CreateFst(ofst, &open_dest, &close_src);
-  ProcSCCs(ofst, &open_dest, &close_src, &close_non_term_weight);
-  const auto total_nparens = AssignParenIds(*ofst, &paren_map);
-  AssignParenLabels(total_nparens, parens);
-  AddParensToFst(*parens, paren_map, open_dest, close_src,
-                 close_non_term_weight, ofst);
-  if (!fst_array.empty()) {
-    ofst->SetInputSymbols(fst_array[0].second->InputSymbols());
-    ofst->SetOutputSymbols(fst_array[0].second->OutputSymbols());
-  }
-  AddParensToSymbolTables(*parens, ofst);
-  Connect(ofst);
-}
-
-template <class Arc>
-void PdtLeftSRParser<Arc>::ProcSCCs(
-    MutableFst<Arc> *ofst,
-    std::vector<StateId> *open_dest,
-    std::vector<std::vector<StateWeightPair>> *close_src,
-    std::vector<bool> *close_non_term_weight) const {
-  const auto &fst_array = FstArray();
-  for (StateIterator<Fst<Arc>> siter(*ofst); !siter.Done(); siter.Next()) {
-    const auto os = siter.Value();
-    const auto label = GetLabelStatePair(os).first;
-    const auto is = GetLabelStatePair(os).second;
-    const auto fst_id = Label2Id(label);
-    const auto scc_id = SCC(label);
-    const auto rs = RepState(scc_id);
-    const auto *ifst = fst_array[fst_id].second;
-    // SCC LEFT-LINEAR: puts non-terminal weights on close parentheses. Merges
-    // initial states into SCC representative state and updates open_dest.
-    if (SCCLeftLinear(scc_id)) {
-      (*close_non_term_weight)[fst_id] = true;
-      if (is == ifst->Start() && os != rs) {
-        for (ArcIterator<Fst<Arc>> aiter(*ofst, os); !aiter.Done();
-             aiter.Next()) {
-          const auto &arc = aiter.Value();
-          ofst->AddArc(rs, arc);
-        }
-        ofst->DeleteArcs(os);
-        if (os == ofst->Start())
-          ofst->SetStart(rs);
-        (*open_dest)[fst_id] = rs;
-      }
-    }
-    // SCC RIGHT-LINEAR: pushes back final weights onto non-terminals, if
-    // possible, or adds weighted epsilons to the SCC representative state.
-    // Merges final states into SCC representative state and updates close_src.
-    if (SCCRightLinear(scc_id)) {
-      for (MutableArcIterator<MutableFst<Arc>> aiter(ofst, os); !aiter.Done();
-           aiter.Next()) {
-        auto arc = aiter.Value();
-        const auto idest = GetLabelStatePair(arc.nextstate).second;
-        if (NonTermDests(fst_id).count(idest) > 0) {
-          if (ofst->Final(arc.nextstate) != Weight::Zero()) {
-            ofst->SetFinal(arc.nextstate, Weight::Zero());
-            ofst->SetFinal(rs, Weight::One());
-          }
-          arc.weight = Times(arc.weight, ifst->Final(idest));
-          arc.nextstate = rs;
-          aiter.SetValue(arc);
-        }
-      }
-      const auto final_weight = ifst->Final(is);
-      if (final_weight != Weight::Zero() &&
-          NonTermDests(fst_id).count(is) == 0) {
-        ofst->AddArc(os, Arc(0, 0, final_weight, rs));
-        if (ofst->Final(os) != Weight::Zero()) {
-          ofst->SetFinal(os, Weight::Zero());
-          ofst->SetFinal(rs, Weight::One());
-        }
-      }
-      if (is == ifst->Start()) {
-        (*close_src)[fst_id].clear();
-        (*close_src)[fst_id].emplace_back(rs, Weight::One());
-      }
-    }
-  }
-}
-
-template <class Arc>
-void PdtLeftSRParser<Arc>::GetNonTermDests() const {
-  const auto &fst_array = FstArray();
-  non_term_dests_.resize(fst_array.size());
-  for (size_t fst_id = 0; fst_id < fst_array.size(); ++fst_id) {
-    const auto label = fst_array[fst_id].first;
-    const auto scc_id = SCC(label);
-    if (SCCRightLinear(scc_id)) {
-      const auto *ifst = fst_array[fst_id].second;
-      for (StateIterator<Fst<Arc>> siter(*ifst); !siter.Done(); siter.Next()) {
-        const auto is = siter.Value();
-        for (ArcIterator<Fst<Arc>> aiter(*ifst, is); !aiter.Done();
-             aiter.Next()) {
-          const auto &arc = aiter.Value();
-          if (Label2Id(arc.olabel) != kNoStateId) {
-            non_term_dests_[fst_id].insert(arc.nextstate);
-          }
-        }
-      }
-    }
-  }
-}
-
-template <class Arc>
-size_t PdtLeftSRParser<Arc>::AssignParenIds(
-    const Fst<Arc> &ofst,
-    ParenMap *paren_map) const {
-  const auto &fst_array = FstArray();
-  // Number of distinct parenthesis pairs per FST.
-  std::vector<size_t> nparens(fst_array.size(), 0);
-  // Number of distinct parenthesis pairs overall.
-  size_t total_nparens = 0;
-  for (StateIterator<Fst<Arc>> siter(ofst); !siter.Done(); siter.Next()) {
-    const auto os = siter.Value();
-    const auto label = GetLabelStatePair(os).first;
-    const auto scc_id = SCC(label);
-    for (ArcIterator<Fst<Arc>> aiter(ofst, os); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto nfst_id = Label2Id(arc.olabel);
-      if (nfst_id != kNoStateId) {
-        size_t nscc_id = SCC(arc.olabel);
-        bool nscc_linear = !SCCComps(nscc_id).empty();
-        // Assigns a parenthesis ID for the non-terminal transition
-        // if the non-terminal belongs to a (left-/right-) linear dependency
-        // SCC or if the transition is in an FST from a different SCC
-        if (!nscc_linear || scc_id != nscc_id) {
-          // For (left-/right-) linear SCCs instead of using nfst_id, we
-          // will use its SCC prototype pfst_id for assigning distinct
-          // parenthesis IDs.
-          const auto pfst_id =
-              nscc_linear ? SCCComps(nscc_id).front() : nfst_id;
-          ParenKey paren_key(pfst_id, arc.nextstate);
-          const auto it = paren_map->find(paren_key);
-          if (it == paren_map->end()) {
-            // Assigns new paren ID for this (FST/SCC, dest. state) pair.
-            if (nscc_linear) {
-              // This is mapping we'll need, but we also store (harmlessly)
-              // for the prototype below so we can easily keep count per SCC.
-              const ParenKey nparen_key(nfst_id, arc.nextstate);
-              (*paren_map)[nparen_key] = nparens[pfst_id];
-            }
-            (*paren_map)[paren_key] = nparens[pfst_id]++;
-            if (nparens[pfst_id] > total_nparens) {
-              total_nparens = nparens[pfst_id];
-            }
-          }
-        }
-      }
-    }
-  }
-  return total_nparens;
-}
-
-// Builds a pushdown transducer (PDT) from an RTN specification. The result is
-// a PDT written to a mutable FST where some transitions are labeled with
-// open or close parentheses. To be interpreted as a PDT, the parens must
-// balance on a path (see PdtExpand()). The open/close parenthesis label pairs
-// are returned in the parens argument.
-template <class Arc>
-void Replace(
-    const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
-        &ifst_array,
-    MutableFst<Arc> *ofst,
-    std::vector<std::pair<typename Arc::Label, typename Arc::Label>> *parens,
-    const PdtReplaceOptions<Arc> &opts) {
-  switch (opts.type) {
-    case PDT_LEFT_PARSER:
-      {
-        PdtLeftParser<Arc> pr(ifst_array, opts);
-        pr.GetParser(ofst, parens);
-        return;
-      }
-    case PDT_LEFT_SR_PARSER:
-      {
-        PdtLeftSRParser<Arc> pr(ifst_array, opts);
-        pr.GetParser(ofst, parens);
-        return;
-      }
-    default:
-      FSTERROR() << "Replace: Unknown PDT parser type: " << opts.type;
-      ofst->DeleteStates();
-      ofst->SetProperties(kError, kError);
-      parens->clear();
-      return;
-  }
-}
-
-// Variant where the only user-controlled arguments is the root ID.
-template <class Arc>
-void Replace(
-    const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
-        &ifst_array,
-    MutableFst<Arc> *ofst,
-    std::vector<std::pair<typename Arc::Label, typename Arc::Label>> *parens,
-    typename Arc::Label root) {
-  PdtReplaceOptions<Arc> opts(root);
-  Replace(ifst_array, ofst, parens, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_REPLACE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/reverse.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/reverse.h
deleted file mode 100644
index 3bd44e21..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/reverse.h
+++ /dev/null
@@ -1,38 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Expands a PDT to an FST.
-
-#ifndef FST_EXTENSIONS_PDT_REVERSE_H_
-#define FST_EXTENSIONS_PDT_REVERSE_H_
-
-#include <vector>
-
-#include <fst/mutable-fst.h>
-#include <fst/relabel.h>
-#include <fst/reverse.h>
-
-namespace fst {
-
-// Reverses a pushdown transducer (PDT) encoded as an FST.
-template <class Arc, class RevArc>
-void Reverse(const Fst<Arc> &ifst,
-             const std::vector<
-                 std::pair<typename Arc::Label, typename Arc::Label>> &parens,
-             MutableFst<RevArc> *ofst) {
-  using Label = typename Arc::Label;
-  // Reverses FST component.
-  Reverse(ifst, ofst);
-  // Exchanges open and close parenthesis pairs.
-  std::vector<std::pair<Label, Label>> relabel_pairs;
-  relabel_pairs.reserve(2 * parens.size());
-  for (const auto &pair : parens) {
-    relabel_pairs.emplace_back(pair.first, pair.second);
-    relabel_pairs.emplace_back(pair.second, pair.first);
-  }
-  Relabel(ofst, relabel_pairs, relabel_pairs);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_REVERSE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/shortest-path.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/shortest-path.h
deleted file mode 100644
index 591c1ea4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/pdt/shortest-path.h
+++ /dev/null
@@ -1,715 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions to find shortest paths in a PDT.
-
-#ifndef FST_EXTENSIONS_PDT_SHORTEST_PATH_H_
-#define FST_EXTENSIONS_PDT_SHORTEST_PATH_H_
-
-#include <stack>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/extensions/pdt/paren.h>
-#include <fst/extensions/pdt/pdt.h>
-#include <fst/shortest-path.h>
-
-namespace fst {
-
-template <class Arc, class Queue>
-struct PdtShortestPathOptions {
-  bool keep_parentheses;
-  bool path_gc;
-
-  PdtShortestPathOptions(bool keep_parentheses = false, bool path_gc = true)
-      : keep_parentheses(keep_parentheses), path_gc(path_gc) {}
-};
-
-namespace internal {
-
-// Flags for shortest path data.
-
-constexpr uint8 kPdtInited = 0x01;
-constexpr uint8 kPdtFinal = 0x02;
-constexpr uint8 kPdtMarked = 0x04;
-
-// Stores shortest path tree info Distance(), Parent(), and ArcParent()
-// information keyed on two types:
-//
-// 1. SearchState: This is a usual node in a shortest path tree but:
-//    a. is w.r.t a PDT search state (a pair of a PDT state and a "start" state,
-//    either the PDT start state or the destination state of an open
-//    parenthesis).
-//    b. the Distance() is from this "start" state to the search state.
-//    c. Parent().state is kNoLabel for the "start" state.
-//
-// 2. ParenSpec: This connects shortest path trees depending on the the
-// parenthesis taken. Given the parenthesis spec:
-//    a. the Distance() is from the Parent() "start" state to the parenthesis
-//    destination state.
-//    b. The ArcParent() is the parenthesis arc.
-template <class Arc>
-class PdtShortestPathData {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  struct SearchState {
-    StateId state;  // PDT state.
-    StateId start;  // PDT paren "start" state.
-
-    SearchState(StateId s = kNoStateId, StateId t = kNoStateId)
-        : state(s), start(t) {}
-
-    bool operator==(const SearchState &other) const {
-      if (&other == this) return true;
-      return other.state == state && other.start == start;
-    }
-  };
-
-  // Specifies paren ID, source and dest "start" states of a paren. These are
-  // the "start" states of the respective sub-graphs.
-  struct ParenSpec {
-    ParenSpec(Label paren_id = kNoLabel, StateId src_start = kNoStateId,
-              StateId dest_start = kNoStateId)
-        : paren_id(paren_id), src_start(src_start), dest_start(dest_start) {}
-
-    Label paren_id;
-    StateId src_start;   // Sub-graph "start" state for paren source.
-    StateId dest_start;  // Sub-graph "start" state for paren dest.
-
-    bool operator==(const ParenSpec &other) const {
-      if (&other == this) return true;
-      return (other.paren_id == paren_id &&
-              other.src_start == other.src_start &&
-              other.dest_start == dest_start);
-    }
-  };
-
-  struct SearchData {
-    SearchData()
-        : distance(Weight::Zero()),
-          parent(kNoStateId, kNoStateId),
-          paren_id(kNoLabel),
-          flags(0) {}
-
-    Weight distance;     // Distance to this state from PDT "start" state.
-    SearchState parent;  // Parent state in shortest path tree.
-    int16 paren_id;      // If parent arc has paren, paren ID (or kNoLabel).
-    uint8 flags;         // First byte reserved for PdtShortestPathData use.
-  };
-
-  PdtShortestPathData(bool gc)
-      : gc_(gc), nstates_(0), ngc_(0), finished_(false) {}
-
-  ~PdtShortestPathData() {
-    VLOG(1) << "opm size: " << paren_map_.size();
-    VLOG(1) << "# of search states: " << nstates_;
-    if (gc_) VLOG(1) << "# of GC'd search states: " << ngc_;
-  }
-
-  void Clear() {
-    search_map_.clear();
-    search_multimap_.clear();
-    paren_map_.clear();
-    state_ = SearchState(kNoStateId, kNoStateId);
-    nstates_ = 0;
-    ngc_ = 0;
-  }
-
-  // TODO(kbg): Currently copying SearchState and passing a const reference to
-  // ParenSpec. Benchmark to confirm this is the right thing to do.
-
-  Weight Distance(SearchState s) const { return GetSearchData(s)->distance; }
-
-  Weight Distance(const ParenSpec &paren) const {
-    return GetSearchData(paren)->distance;
-  }
-
-  SearchState Parent(SearchState s) const { return GetSearchData(s)->parent; }
-
-  SearchState Parent(const ParenSpec &paren) const {
-    return GetSearchData(paren)->parent;
-  }
-
-  Label ParenId(SearchState s) const { return GetSearchData(s)->paren_id; }
-
-  uint8 Flags(SearchState s) const { return GetSearchData(s)->flags; }
-
-  void SetDistance(SearchState s, Weight weight) {
-    GetSearchData(s)->distance = std::move(weight);
-  }
-
-  void SetDistance(const ParenSpec &paren, Weight weight) {
-    GetSearchData(paren)->distance = std::move(weight);
-  }
-
-  void SetParent(SearchState s, SearchState p) { GetSearchData(s)->parent = p; }
-
-  void SetParent(const ParenSpec &paren, SearchState p) {
-    GetSearchData(paren)->parent = p;
-  }
-
-  void SetParenId(SearchState s, Label p) {
-    if (p >= 32768) {
-      FSTERROR() << "PdtShortestPathData: Paren ID does not fit in an int16";
-    }
-    GetSearchData(s)->paren_id = p;
-  }
-
-  void SetFlags(SearchState s, uint8 f, uint8 mask) {
-    auto *data = GetSearchData(s);
-    data->flags &= ~mask;
-    data->flags |= f & mask;
-  }
-
-  void GC(StateId s);
-
-  void Finish() { finished_ = true; }
-
- private:
-  // Hash for search state.
-  struct SearchStateHash {
-    size_t operator()(const SearchState &s) const {
-      static constexpr auto prime = 7853;
-      return s.state + s.start * prime;
-    }
-  };
-
-  // Hash for paren map.
-  struct ParenHash {
-    size_t operator()(const ParenSpec &paren) const {
-      static constexpr auto prime0 = 7853;
-      static constexpr auto prime1 = 7867;
-      return paren.paren_id + paren.src_start * prime0 +
-             paren.dest_start * prime1;
-    }
-  };
-
-  using SearchMap =
-      std::unordered_map<SearchState, SearchData, SearchStateHash>;
-
-  using SearchMultimap = std::unordered_multimap<StateId, StateId>;
-
-  // Hash map from paren spec to open paren data.
-  using ParenMap = std::unordered_map<ParenSpec, SearchData, ParenHash>;
-
-  SearchData *GetSearchData(SearchState s) const {
-    if (s == state_) return state_data_;
-    if (finished_) {
-      auto it = search_map_.find(s);
-      if (it == search_map_.end()) return &null_search_data_;
-      state_ = s;
-      return state_data_ = &(it->second);
-    } else {
-      state_ = s;
-      state_data_ = &search_map_[s];
-      if (!(state_data_->flags & kPdtInited)) {
-        ++nstates_;
-        if (gc_) search_multimap_.insert(std::make_pair(s.start, s.state));
-        state_data_->flags = kPdtInited;
-      }
-      return state_data_;
-    }
-  }
-
-  SearchData *GetSearchData(ParenSpec paren) const {
-    if (paren == paren_) return paren_data_;
-    if (finished_) {
-      auto it = paren_map_.find(paren);
-      if (it == paren_map_.end()) return &null_search_data_;
-      paren_ = paren;
-      return state_data_ = &(it->second);
-    } else {
-      paren_ = paren;
-      return paren_data_ = &paren_map_[paren];
-    }
-  }
-
-  mutable SearchMap search_map_;            // Maps from search state to data.
-  mutable SearchMultimap search_multimap_;  // Maps from "start" to subgraph.
-  mutable ParenMap paren_map_;              // Maps paren spec to search data.
-  mutable SearchState state_;               // Last state accessed.
-  mutable SearchData *state_data_;          // Last state data accessed.
-  mutable ParenSpec paren_;                 // Last paren spec accessed.
-  mutable SearchData *paren_data_;          // Last paren data accessed.
-  bool gc_;                                 // Allow GC?
-  mutable size_t nstates_;                  // Total number of search states.
-  size_t ngc_;                              // Number of GC'd search states.
-  mutable SearchData null_search_data_;     // Null search data.
-  bool finished_;                           // Read-only access when true.
-
-  PdtShortestPathData(const PdtShortestPathData &) = delete;
-  PdtShortestPathData &operator=(const PdtShortestPathData &) = delete;
-};
-
-// Deletes inaccessible search data from a given "start" (open paren dest)
-// state. Assumes "final" (close paren source or PDT final) states have
-// been flagged kPdtFinal.
-template <class Arc>
-void PdtShortestPathData<Arc>::GC(StateId start) {
-  if (!gc_) return;
-  std::vector<StateId> finals;
-  for (auto it = search_multimap_.find(start);
-       it != search_multimap_.end() && it->first == start; ++it) {
-    const SearchState s(it->second, start);
-    if (search_map_[s].flags & kPdtFinal) finals.push_back(s.state);
-  }
-  // Mark phase.
-  for (const auto state : finals) {
-    SearchState ss(state, start);
-    while (ss.state != kNoLabel) {
-      auto &sdata = search_map_[ss];
-      if (sdata.flags & kPdtMarked) break;
-      sdata.flags |= kPdtMarked;
-      const auto p = sdata.parent;
-      if (p.start != start && p.start != kNoLabel) {  // Entering sub-subgraph.
-        const ParenSpec paren(sdata.paren_id, ss.start, p.start);
-        ss = paren_map_[paren].parent;
-      } else {
-        ss = p;
-      }
-    }
-  }
-  // Sweep phase.
-  auto it = search_multimap_.find(start);
-  while (it != search_multimap_.end() && it->first == start) {
-    const SearchState s(it->second, start);
-    auto mit = search_map_.find(s);
-    const SearchData &data = mit->second;
-    if (!(data.flags & kPdtMarked)) {
-      search_map_.erase(mit);
-      ++ngc_;
-    }
-    search_multimap_.erase(it++);
-  }
-}
-
-}  // namespace internal
-
-// This computes the single source shortest (balanced) path (SSSP) through a
-// weighted PDT that has a bounded stack (i.e., is expandable as an FST). It is
-// a generalization of the classic SSSP graph algorithm that removes a state s
-// from a queue (defined by a user-provided queue type) and relaxes the
-// destination states of transitions leaving s. In this PDT version, states that
-// have entering open parentheses are treated as source states for a sub-graph
-// SSSP problem with the shortest path up to the open parenthesis being first
-// saved. When a close parenthesis is then encountered any balancing open
-// parenthesis is examined for this saved information and multiplied back. In
-// this way, each sub-graph is entered only once rather than repeatedly. If
-// every state in the input PDT has the property that there is a unique "start"
-// state for it with entering open parentheses, then this algorithm is quite
-// straightforward. In general, this will not be the case, so the algorithm
-// (implicitly) creates a new graph where each state is a pair of an original
-// state and a possible parenthesis "start" state for that state.
-template <class Arc, class Queue>
-class PdtShortestPath {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using SpData = internal::PdtShortestPathData<Arc>;
-  using SearchState = typename SpData::SearchState;
-  using ParenSpec = typename SpData::ParenSpec;
-  using CloseSourceIterator =
-      typename internal::PdtBalanceData<Arc>::SetIterator;
-
-  PdtShortestPath(const Fst<Arc> &ifst,
-                  const std::vector<std::pair<Label, Label>> &parens,
-                  const PdtShortestPathOptions<Arc, Queue> &opts)
-      : ifst_(ifst.Copy()),
-        parens_(parens),
-        keep_parens_(opts.keep_parentheses),
-        start_(ifst.Start()),
-        sp_data_(opts.path_gc),
-        error_(false) {
-    // TODO(kbg): Make this a compile-time static_assert once:
-    // 1) All weight properties are made constexpr for all weight types.
-    // 2) We have a pleasant way to "deregister" this oepration for non-path
-    //    semirings so an informative error message is produced. The best
-    //    solution will probably involve some kind of SFINAE magic.
-    if ((Weight::Properties() & (kPath | kRightSemiring)) !=
-        (kPath | kRightSemiring)) {
-      FSTERROR() << "PdtShortestPath: Weight needs to have the path"
-                 << " property and be right distributive: " << Weight::Type();
-      error_ = true;
-    }
-    for (Label i = 0; i < parens.size(); ++i) {
-      const auto &pair = parens[i];
-      paren_map_[pair.first] = i;
-      paren_map_[pair.second] = i;
-    }
-  }
-
-  ~PdtShortestPath() {
-    VLOG(1) << "# of input states: " << CountStates(*ifst_);
-    VLOG(1) << "# of enqueued: " << nenqueued_;
-    VLOG(1) << "cpmm size: " << close_paren_multimap_.size();
-  }
-
-  void ShortestPath(MutableFst<Arc> *ofst) {
-    Init(ofst);
-    GetDistance(start_);
-    GetPath();
-    sp_data_.Finish();
-    if (error_) ofst->SetProperties(kError, kError);
-  }
-
-  const internal::PdtShortestPathData<Arc> &GetShortestPathData() const {
-    return sp_data_;
-  }
-
-  internal::PdtBalanceData<Arc> *GetBalanceData() { return &balance_data_; }
-
- public:
-  // Hash multimap from close paren label to an paren arc.
-  using CloseParenMultimap =
-      std::unordered_multimap<internal::ParenState<Arc>, Arc,
-                              typename internal::ParenState<Arc>::Hash>;
-
-  const CloseParenMultimap &GetCloseParenMultimap() const {
-    return close_paren_multimap_;
-  }
-
- private:
-  void Init(MutableFst<Arc> *ofst);
-
-  void GetDistance(StateId start);
-
-  void ProcFinal(SearchState s);
-
-  void ProcArcs(SearchState s);
-
-  void ProcOpenParen(Label paren_id, SearchState s, StateId nexstate,
-                     const Weight &weight);
-
-  void ProcCloseParen(Label paren_id, SearchState s, const Weight &weight);
-
-  void ProcNonParen(SearchState s, StateId nextstate, const Weight &weight);
-
-  void Relax(SearchState s, SearchState t, StateId nextstate,
-             const Weight &weight, Label paren_id);
-
-  void Enqueue(SearchState d);
-
-  void GetPath();
-
-  Arc GetPathArc(SearchState s, SearchState p, Label paren_id, bool open);
-
-  std::unique_ptr<Fst<Arc>> ifst_;
-  MutableFst<Arc> *ofst_;
-  const std::vector<std::pair<Label, Label>> &parens_;
-  bool keep_parens_;
-  Queue *state_queue_;
-  StateId start_;
-  Weight fdistance_;
-  SearchState f_parent_;
-  SpData sp_data_;
-  std::unordered_map<Label, Label> paren_map_;
-  CloseParenMultimap close_paren_multimap_;
-  internal::PdtBalanceData<Arc> balance_data_;
-  ssize_t nenqueued_;
-  bool error_;
-
-  static constexpr uint8 kEnqueued = 0x10;
-  static constexpr uint8 kExpanded = 0x20;
-  static constexpr uint8 kFinished = 0x40;
-
-  static const Arc kNoArc;
-};
-
-template <class Arc, class Queue>
-void PdtShortestPath<Arc, Queue>::Init(MutableFst<Arc> *ofst) {
-  ofst_ = ofst;
-  ofst->DeleteStates();
-  ofst->SetInputSymbols(ifst_->InputSymbols());
-  ofst->SetOutputSymbols(ifst_->OutputSymbols());
-  if (ifst_->Start() == kNoStateId) return;
-  fdistance_ = Weight::Zero();
-  f_parent_ = SearchState(kNoStateId, kNoStateId);
-  sp_data_.Clear();
-  close_paren_multimap_.clear();
-  balance_data_.Clear();
-  nenqueued_ = 0;
-  // Finds open parens per destination state and close parens per source state.
-  for (StateIterator<Fst<Arc>> siter(*ifst_); !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    for (ArcIterator<Fst<Arc>> aiter(*ifst_, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto it = paren_map_.find(arc.ilabel);
-      if (it != paren_map_.end()) {  // Is a paren?
-        const auto paren_id = it->second;
-        if (arc.ilabel == parens_[paren_id].first) {  // Open paren.
-          balance_data_.OpenInsert(paren_id, arc.nextstate);
-        } else {  // Close paren.
-          const internal::ParenState<Arc> paren_state(paren_id, s);
-          close_paren_multimap_.emplace(paren_state, arc);
-        }
-      }
-    }
-  }
-}
-
-// Computes the shortest distance stored in a recursive way. Each sub-graph
-// (i.e., different paren "start" state) begins with weight One().
-template <class Arc, class Queue>
-void PdtShortestPath<Arc, Queue>::GetDistance(StateId start) {
-  if (start == kNoStateId) return;
-  Queue state_queue;
-  state_queue_ = &state_queue;
-  const SearchState q(start, start);
-  Enqueue(q);
-  sp_data_.SetDistance(q, Weight::One());
-  while (!state_queue_->Empty()) {
-    const auto state = state_queue_->Head();
-    state_queue_->Dequeue();
-    const SearchState s(state, start);
-    sp_data_.SetFlags(s, 0, kEnqueued);
-    ProcFinal(s);
-    ProcArcs(s);
-    sp_data_.SetFlags(s, kExpanded, kExpanded);
-  }
-  sp_data_.SetFlags(q, kFinished, kFinished);
-  balance_data_.FinishInsert(start);
-  sp_data_.GC(start);
-}
-
-// Updates best complete path.
-template <class Arc, class Queue>
-void PdtShortestPath<Arc, Queue>::ProcFinal(SearchState s) {
-  if (ifst_->Final(s.state) != Weight::Zero() && s.start == start_) {
-    const auto weight = Times(sp_data_.Distance(s), ifst_->Final(s.state));
-    if (fdistance_ != Plus(fdistance_, weight)) {
-      if (f_parent_.state != kNoStateId) {
-        sp_data_.SetFlags(f_parent_, 0, internal::kPdtFinal);
-      }
-      sp_data_.SetFlags(s, internal::kPdtFinal, internal::kPdtFinal);
-      fdistance_ = Plus(fdistance_, weight);
-      f_parent_ = s;
-    }
-  }
-}
-
-// Processes all arcs leaving the state s.
-template <class Arc, class Queue>
-void PdtShortestPath<Arc, Queue>::ProcArcs(SearchState s) {
-  for (ArcIterator<Fst<Arc>> aiter(*ifst_, s.state); !aiter.Done();
-       aiter.Next()) {
-    const auto &arc = aiter.Value();
-    const auto weight = Times(sp_data_.Distance(s), arc.weight);
-    const auto it = paren_map_.find(arc.ilabel);
-    if (it != paren_map_.end()) {  // Is a paren?
-      const auto paren_id = it->second;
-      if (arc.ilabel == parens_[paren_id].first) {
-        ProcOpenParen(paren_id, s, arc.nextstate, weight);
-      } else {
-        ProcCloseParen(paren_id, s, weight);
-      }
-    } else {
-      ProcNonParen(s, arc.nextstate, weight);
-    }
-  }
-}
-
-// Saves the shortest path info for reaching this parenthesis and starts a new
-// SSSP in the sub-graph pointed to by the parenthesis if previously unvisited.
-// Otherwise it finds any previously encountered closing parentheses and relaxes
-// them using the recursively stored shortest distance to them.
-template <class Arc, class Queue>
-inline void PdtShortestPath<Arc, Queue>::ProcOpenParen(Label paren_id,
-                                                       SearchState s,
-                                                       StateId nextstate,
-                                                       const Weight &weight) {
-  const SearchState d(nextstate, nextstate);
-  const ParenSpec paren(paren_id, s.start, d.start);
-  const auto pdist = sp_data_.Distance(paren);
-  if (pdist != Plus(pdist, weight)) {
-    sp_data_.SetDistance(paren, weight);
-    sp_data_.SetParent(paren, s);
-    const auto dist = sp_data_.Distance(d);
-    if (dist == Weight::Zero()) {
-      auto *state_queue = state_queue_;
-      GetDistance(d.start);
-      state_queue_ = state_queue;
-    } else if (!(sp_data_.Flags(d) & kFinished)) {
-      FSTERROR()
-          << "PdtShortestPath: open parenthesis recursion: not bounded stack";
-      error_ = true;
-    }
-    for (auto set_iter = balance_data_.Find(paren_id, nextstate);
-         !set_iter.Done(); set_iter.Next()) {
-      const SearchState cpstate(set_iter.Element(), d.start);
-      const internal::ParenState<Arc> paren_state(paren_id, cpstate.state);
-      for (auto cpit = close_paren_multimap_.find(paren_state);
-           cpit != close_paren_multimap_.end() && paren_state == cpit->first;
-           ++cpit) {
-        const auto &cparc = cpit->second;
-        const auto cpw =
-            Times(weight, Times(sp_data_.Distance(cpstate), cparc.weight));
-        Relax(cpstate, s, cparc.nextstate, cpw, paren_id);
-      }
-    }
-  }
-}
-
-// Saves the correspondence between each closing parenthesis and its balancing
-// open parenthesis info. Relaxes any close parenthesis destination state that
-// has a balancing previously encountered open parenthesis.
-template <class Arc, class Queue>
-inline void PdtShortestPath<Arc, Queue>::ProcCloseParen(Label paren_id,
-                                                        SearchState s,
-                                                        const Weight &weight) {
-  const internal::ParenState<Arc> paren_state(paren_id, s.start);
-  if (!(sp_data_.Flags(s) & kExpanded)) {
-    balance_data_.CloseInsert(paren_id, s.start, s.state);
-    sp_data_.SetFlags(s, internal::kPdtFinal, internal::kPdtFinal);
-  }
-}
-
-// Classical relaxation for non-parentheses.
-template <class Arc, class Queue>
-inline void PdtShortestPath<Arc, Queue>::ProcNonParen(SearchState s,
-                                                      StateId nextstate,
-                                                      const Weight &weight) {
-  Relax(s, s, nextstate, weight, kNoLabel);
-}
-
-// Classical relaxation on the search graph for an arc with destination state
-// nexstate from state s. State t is in the same sub-graph as nextstate (i.e.,
-// has the same paren "start").
-template <class Arc, class Queue>
-inline void PdtShortestPath<Arc, Queue>::Relax(SearchState s, SearchState t,
-                                               StateId nextstate,
-                                               const Weight &weight,
-                                               Label paren_id) {
-  const SearchState d(nextstate, t.start);
-  Weight dist = sp_data_.Distance(d);
-  if (dist != Plus(dist, weight)) {
-    sp_data_.SetParent(d, s);
-    sp_data_.SetParenId(d, paren_id);
-    sp_data_.SetDistance(d, Plus(dist, weight));
-    Enqueue(d);
-  }
-}
-
-template <class Arc, class Queue>
-inline void PdtShortestPath<Arc, Queue>::Enqueue(SearchState s) {
-  if (!(sp_data_.Flags(s) & kEnqueued)) {
-    state_queue_->Enqueue(s.state);
-    sp_data_.SetFlags(s, kEnqueued, kEnqueued);
-    ++nenqueued_;
-  } else {
-    state_queue_->Update(s.state);
-  }
-}
-
-// Follows parent pointers to find the shortest path. A stack is used since the
-// shortest distance is stored recursively.
-template <class Arc, class Queue>
-void PdtShortestPath<Arc, Queue>::GetPath() {
-  SearchState s = f_parent_;
-  SearchState d = SearchState(kNoStateId, kNoStateId);
-  StateId s_p = kNoStateId;
-  StateId d_p = kNoStateId;
-  auto arc = kNoArc;
-  Label paren_id = kNoLabel;
-  std::stack<ParenSpec> paren_stack;
-  while (s.state != kNoStateId) {
-    d_p = s_p;
-    s_p = ofst_->AddState();
-    if (d.state == kNoStateId) {
-      ofst_->SetFinal(s_p, ifst_->Final(f_parent_.state));
-    } else {
-      if (paren_id != kNoLabel) {                     // Paren?
-        if (arc.ilabel == parens_[paren_id].first) {  // Open paren?
-          paren_stack.pop();
-        } else {  // Close paren?
-          const ParenSpec paren(paren_id, d.start, s.start);
-          paren_stack.push(paren);
-        }
-        if (!keep_parens_) arc.ilabel = arc.olabel = 0;
-      }
-      arc.nextstate = d_p;
-      ofst_->AddArc(s_p, arc);
-    }
-    d = s;
-    s = sp_data_.Parent(d);
-    paren_id = sp_data_.ParenId(d);
-    if (s.state != kNoStateId) {
-      arc = GetPathArc(s, d, paren_id, false);
-    } else if (!paren_stack.empty()) {
-      const ParenSpec paren = paren_stack.top();
-      s = sp_data_.Parent(paren);
-      paren_id = paren.paren_id;
-      arc = GetPathArc(s, d, paren_id, true);
-    }
-  }
-  ofst_->SetStart(s_p);
-  ofst_->SetProperties(
-      ShortestPathProperties(ofst_->Properties(kFstProperties, false)),
-      kFstProperties);
-}
-
-// Finds transition with least weight between two states with label matching
-// paren_id and open/close paren type or a non-paren if kNoLabel.
-template <class Arc, class Queue>
-Arc PdtShortestPath<Arc, Queue>::GetPathArc(SearchState s, SearchState d,
-                                            Label paren_id, bool open_paren) {
-  auto path_arc = kNoArc;
-  for (ArcIterator<Fst<Arc>> aiter(*ifst_, s.state); !aiter.Done();
-       aiter.Next()) {
-    const auto &arc = aiter.Value();
-    if (arc.nextstate != d.state) continue;
-    Label arc_paren_id = kNoLabel;
-    const auto it = paren_map_.find(arc.ilabel);
-    if (it != paren_map_.end()) {
-      arc_paren_id = it->second;
-      bool arc_open_paren = (arc.ilabel == parens_[arc_paren_id].first);
-      if (arc_open_paren != open_paren) continue;
-    }
-    if (arc_paren_id != paren_id) continue;
-    if (arc.weight == Plus(arc.weight, path_arc.weight)) path_arc = arc;
-  }
-  if (path_arc.nextstate == kNoStateId) {
-    FSTERROR() << "PdtShortestPath::GetPathArc: Failed to find arc";
-    error_ = true;
-  }
-  return path_arc;
-}
-
-template <class Arc, class Queue>
-const Arc PdtShortestPath<Arc, Queue>::kNoArc = Arc(kNoLabel, kNoLabel,
-                                                    Weight::Zero(), kNoStateId);
-
-// Functional variants.
-
-template <class Arc, class Queue>
-void ShortestPath(
-    const Fst<Arc> &ifst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens,
-    MutableFst<Arc> *ofst, const PdtShortestPathOptions<Arc, Queue> &opts) {
-  PdtShortestPath<Arc, Queue> psp(ifst, parens, opts);
-  psp.ShortestPath(ofst);
-}
-
-template <class Arc>
-void ShortestPath(
-    const Fst<Arc> &ifst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &parens,
-    MutableFst<Arc> *ofst) {
-  using Q = FifoQueue<typename Arc::StateId>;
-  const PdtShortestPathOptions<Arc, Q> opts;
-  PdtShortestPath<Arc, Q> psp(ifst, parens, opts);
-  psp.ShortestPath(ofst);
-}
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_PDT_SHORTEST_PATH_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/special/phi-fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/special/phi-fst.h
deleted file mode 100644
index 4e7be957..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/special/phi-fst.h
+++ /dev/null
@@ -1,171 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_SPECIAL_PHI_FST_H_
-#define FST_EXTENSIONS_SPECIAL_PHI_FST_H_
-
-#include <memory>
-#include <string>
-
-#include <fst/const-fst.h>
-#include <fst/matcher-fst.h>
-#include <fst/matcher.h>
-
-DECLARE_int64(phi_fst_phi_label);
-DECLARE_bool(phi_fst_phi_loop);
-DECLARE_string(phi_fst_rewrite_mode);
-
-namespace fst {
-namespace internal {
-
-template <class Label>
-class PhiFstMatcherData {
- public:
-  PhiFstMatcherData(
-      Label phi_label = FLAGS_phi_fst_phi_label,
-      bool phi_loop = FLAGS_phi_fst_phi_loop,
-      MatcherRewriteMode rewrite_mode = RewriteMode(FLAGS_phi_fst_rewrite_mode))
-      : phi_label_(phi_label),
-        phi_loop_(phi_loop),
-        rewrite_mode_(rewrite_mode) {}
-
-  PhiFstMatcherData(const PhiFstMatcherData &data)
-      : phi_label_(data.phi_label_),
-        phi_loop_(data.phi_loop_),
-        rewrite_mode_(data.rewrite_mode_) {}
-
-  static PhiFstMatcherData<Label> *Read(std::istream &istrm,
-                                        const FstReadOptions &read) {
-    auto *data = new PhiFstMatcherData<Label>();
-    ReadType(istrm, &data->phi_label_);
-    ReadType(istrm, &data->phi_loop_);
-    int32 rewrite_mode;
-    ReadType(istrm, &rewrite_mode);
-    data->rewrite_mode_ = static_cast<MatcherRewriteMode>(rewrite_mode);
-    return data;
-  }
-
-  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
-    WriteType(ostrm, phi_label_);
-    WriteType(ostrm, phi_loop_);
-    WriteType(ostrm, static_cast<int32>(rewrite_mode_));
-    return !ostrm ? false : true;
-  }
-
-  Label PhiLabel() const { return phi_label_; }
-
-  bool PhiLoop() const { return phi_loop_; }
-
-  MatcherRewriteMode RewriteMode() const { return rewrite_mode_; }
-
- private:
-  static MatcherRewriteMode RewriteMode(const string &mode) {
-    if (mode == "auto") return MATCHER_REWRITE_AUTO;
-    if (mode == "always") return MATCHER_REWRITE_ALWAYS;
-    if (mode == "never") return MATCHER_REWRITE_NEVER;
-    LOG(WARNING) << "PhiFst: Unknown rewrite mode: " << mode << ". "
-                 << "Defaulting to auto.";
-    return MATCHER_REWRITE_AUTO;
-  }
-
-  Label phi_label_;
-  bool phi_loop_;
-  MatcherRewriteMode rewrite_mode_;
-};
-
-}  // namespace internal
-
-constexpr uint8 kPhiFstMatchInput = 0x01;   // Input matcher is PhiMatcher.
-constexpr uint8 kPhiFstMatchOutput = 0x02;  // Output matcher is PhiMatcher.
-
-template <class M, uint8 flags = kPhiFstMatchInput | kPhiFstMatchOutput>
-class PhiFstMatcher : public PhiMatcher<M> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename M::Arc;
-  using StateId = typename Arc::StateId;
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-  using MatcherData = internal::PhiFstMatcherData<Label>;
-
-  enum : uint8 { kFlags = flags };
-
-  PhiFstMatcher(const FST &fst, MatchType match_type,
-      std::shared_ptr<MatcherData> data = std::make_shared<MatcherData>())
-      : PhiMatcher<M>(fst, match_type,
-                      PhiLabel(match_type, data ? data->PhiLabel()
-                                                : MatcherData().PhiLabel()),
-                      data ? data->PhiLoop() : MatcherData().PhiLoop(),
-                      data ? data->RewriteMode() : MatcherData().RewriteMode()),
-        data_(data) {}
-
-  PhiFstMatcher(const PhiFstMatcher<M, flags> &matcher, bool safe = false)
-      : PhiMatcher<M>(matcher, false), data_(matcher.data_) {}
-
-  PhiFstMatcher<M, flags> *Copy(bool safe = false) const override {
-    return new PhiFstMatcher<M, flags>(*this, safe);
-  }
-
-  const MatcherData *GetData() const { return data_.get(); }
-
-  std::shared_ptr<MatcherData> GetSharedData() const { return data_; }
-
- private:
-  static Label PhiLabel(MatchType match_type, Label label) {
-    if (match_type == MATCH_INPUT && flags & kPhiFstMatchInput) return label;
-    if (match_type == MATCH_OUTPUT && flags & kPhiFstMatchOutput) return label;
-    return kNoLabel;
-  }
-
-  std::shared_ptr<MatcherData> data_;
-};
-
-extern const char phi_fst_type[];
-extern const char input_phi_fst_type[];
-extern const char output_phi_fst_type[];
-
-using StdPhiFst =
-    MatcherFst<ConstFst<StdArc>, PhiFstMatcher<SortedMatcher<ConstFst<StdArc>>>,
-               phi_fst_type>;
-
-using LogPhiFst =
-    MatcherFst<ConstFst<LogArc>, PhiFstMatcher<SortedMatcher<ConstFst<LogArc>>>,
-               phi_fst_type>;
-
-using Log64PhiFst = MatcherFst<ConstFst<Log64Arc>,
-                               PhiFstMatcher<SortedMatcher<ConstFst<Log64Arc>>>,
-                               input_phi_fst_type>;
-
-using StdInputPhiFst =
-    MatcherFst<ConstFst<StdArc>, PhiFstMatcher<SortedMatcher<ConstFst<StdArc>>,
-                                               kPhiFstMatchInput>,
-               input_phi_fst_type>;
-
-using LogInputPhiFst =
-    MatcherFst<ConstFst<LogArc>, PhiFstMatcher<SortedMatcher<ConstFst<LogArc>>,
-                                               kPhiFstMatchInput>,
-               input_phi_fst_type>;
-
-using Log64InputPhiFst = MatcherFst<
-    ConstFst<Log64Arc>,
-    PhiFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kPhiFstMatchInput>,
-    input_phi_fst_type>;
-
-using StdOutputPhiFst =
-    MatcherFst<ConstFst<StdArc>, PhiFstMatcher<SortedMatcher<ConstFst<StdArc>>,
-                                               kPhiFstMatchOutput>,
-               output_phi_fst_type>;
-
-using LogOutputPhiFst =
-    MatcherFst<ConstFst<LogArc>, PhiFstMatcher<SortedMatcher<ConstFst<LogArc>>,
-                                               kPhiFstMatchOutput>,
-               output_phi_fst_type>;
-
-using Log64OutputPhiFst = MatcherFst<
-    ConstFst<Log64Arc>,
-    PhiFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kPhiFstMatchOutput>,
-    output_phi_fst_type>;
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_SPECIAL_PHI_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/special/rho-fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/special/rho-fst.h
deleted file mode 100644
index 1f876637..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/special/rho-fst.h
+++ /dev/null
@@ -1,160 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_SPECIAL_RHO_FST_H_
-#define FST_EXTENSIONS_SPECIAL_RHO_FST_H_
-
-#include <memory>
-#include <string>
-
-#include <fst/const-fst.h>
-#include <fst/matcher-fst.h>
-#include <fst/matcher.h>
-
-DECLARE_int64(rho_fst_rho_label);
-DECLARE_string(rho_fst_rewrite_mode);
-
-namespace fst {
-namespace internal {
-
-template <class Label>
-class RhoFstMatcherData {
- public:
-  explicit RhoFstMatcherData(
-      Label rho_label = FLAGS_rho_fst_rho_label,
-      MatcherRewriteMode rewrite_mode = RewriteMode(FLAGS_rho_fst_rewrite_mode))
-      : rho_label_(rho_label), rewrite_mode_(rewrite_mode) {}
-
-  RhoFstMatcherData(const RhoFstMatcherData &data)
-      : rho_label_(data.rho_label_), rewrite_mode_(data.rewrite_mode_) {}
-
-  static RhoFstMatcherData<Label> *Read(std::istream &istrm,
-                                    const FstReadOptions &read) {
-    auto *data = new RhoFstMatcherData<Label>();
-    ReadType(istrm, &data->rho_label_);
-    int32 rewrite_mode;
-    ReadType(istrm, &rewrite_mode);
-    data->rewrite_mode_ = static_cast<MatcherRewriteMode>(rewrite_mode);
-    return data;
-  }
-
-  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
-    WriteType(ostrm, rho_label_);
-    WriteType(ostrm, static_cast<int32>(rewrite_mode_));
-    return !ostrm ? false : true;
-  }
-
-  Label RhoLabel() const { return rho_label_; }
-
-  MatcherRewriteMode RewriteMode() const { return rewrite_mode_; }
-
- private:
-  static MatcherRewriteMode RewriteMode(const string &mode) {
-    if (mode == "auto") return MATCHER_REWRITE_AUTO;
-    if (mode == "always") return MATCHER_REWRITE_ALWAYS;
-    if (mode == "never") return MATCHER_REWRITE_NEVER;
-    LOG(WARNING) << "RhoFst: Unknown rewrite mode: " << mode << ". "
-                 << "Defaulting to auto.";
-    return MATCHER_REWRITE_AUTO;
-  }
-
-  Label rho_label_;
-  MatcherRewriteMode rewrite_mode_;
-};
-
-}  // namespace internal
-
-constexpr uint8 kRhoFstMatchInput = 0x01;   // Input matcher is RhoMatcher.
-constexpr uint8 kRhoFstMatchOutput = 0x02;  // Output matcher is RhoMatcher.
-
-template <class M, uint8 flags = kRhoFstMatchInput | kRhoFstMatchOutput>
-class RhoFstMatcher : public RhoMatcher<M> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename M::Arc;
-  using StateId = typename Arc::StateId;
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-  using MatcherData = internal::RhoFstMatcherData<Label>;
-
-  enum : uint8 { kFlags = flags };
-
-  RhoFstMatcher(
-      const FST &fst, MatchType match_type,
-      std::shared_ptr<MatcherData> data = std::make_shared<MatcherData>())
-      : RhoMatcher<M>(fst, match_type,
-                      RhoLabel(match_type, data ? data->RhoLabel()
-                                                : MatcherData().RhoLabel()),
-                      data ? data->RewriteMode() : MatcherData().RewriteMode()),
-        data_(data) {}
-
-  RhoFstMatcher(const RhoFstMatcher<M, flags> &matcher, bool safe = false)
-      : RhoMatcher<M>(matcher, false), data_(matcher.data_) {}
-
-  RhoFstMatcher<M, flags> *Copy(bool safe = false) const override {
-    return new RhoFstMatcher<M, flags>(*this, safe);
-  }
-
-  const MatcherData *GetData() const { return data_.get(); }
-
-  std::shared_ptr<MatcherData> GetSharedData() const { return data_; }
-
- private:
-  static Label RhoLabel(MatchType match_type, Label label) {
-    if (match_type == MATCH_INPUT && flags & kRhoFstMatchInput) return label;
-    if (match_type == MATCH_OUTPUT && flags & kRhoFstMatchOutput) return label;
-    return kNoLabel;
-  }
-
-  std::shared_ptr<MatcherData> data_;
-};
-
-extern const char rho_fst_type[];
-extern const char input_rho_fst_type[];
-extern const char output_rho_fst_type[];
-
-using StdRhoFst =
-    MatcherFst<ConstFst<StdArc>, RhoFstMatcher<SortedMatcher<ConstFst<StdArc>>>,
-               rho_fst_type>;
-
-using LogRhoFst =
-    MatcherFst<ConstFst<LogArc>, RhoFstMatcher<SortedMatcher<ConstFst<LogArc>>>,
-               rho_fst_type>;
-
-using Log64RhoFst = MatcherFst<ConstFst<Log64Arc>,
-                               RhoFstMatcher<SortedMatcher<ConstFst<Log64Arc>>>,
-                               input_rho_fst_type>;
-
-using StdInputRhoFst =
-    MatcherFst<ConstFst<StdArc>, RhoFstMatcher<SortedMatcher<ConstFst<StdArc>>,
-                                               kRhoFstMatchInput>,
-               input_rho_fst_type>;
-
-using LogInputRhoFst =
-    MatcherFst<ConstFst<LogArc>, RhoFstMatcher<SortedMatcher<ConstFst<LogArc>>,
-                                               kRhoFstMatchInput>,
-               input_rho_fst_type>;
-
-using Log64InputRhoFst = MatcherFst<
-    ConstFst<Log64Arc>,
-    RhoFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kRhoFstMatchInput>,
-    input_rho_fst_type>;
-
-using StdOutputRhoFst =
-    MatcherFst<ConstFst<StdArc>, RhoFstMatcher<SortedMatcher<ConstFst<StdArc>>,
-                                               kRhoFstMatchOutput>,
-               output_rho_fst_type>;
-
-using LogOutputRhoFst =
-    MatcherFst<ConstFst<LogArc>, RhoFstMatcher<SortedMatcher<ConstFst<LogArc>>,
-                                               kRhoFstMatchOutput>,
-               output_rho_fst_type>;
-
-using Log64OutputRhoFst = MatcherFst<
-    ConstFst<Log64Arc>,
-    RhoFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kRhoFstMatchOutput>,
-    output_rho_fst_type>;
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_SPECIAL_RHO_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/special/sigma-fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/special/sigma-fst.h
deleted file mode 100644
index 9ab8fe4f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/extensions/special/sigma-fst.h
+++ /dev/null
@@ -1,163 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_EXTENSIONS_SPECIAL_SIGMA_FST_H_
-#define FST_EXTENSIONS_SPECIAL_SIGMA_FST_H_
-
-#include <memory>
-#include <string>
-
-#include <fst/const-fst.h>
-#include <fst/matcher-fst.h>
-#include <fst/matcher.h>
-
-DECLARE_int64(sigma_fst_sigma_label);
-DECLARE_string(sigma_fst_rewrite_mode);
-
-namespace fst {
-namespace internal {
-
-template <class Label>
-class SigmaFstMatcherData {
- public:
-  explicit SigmaFstMatcherData(Label sigma_label = FLAGS_sigma_fst_sigma_label,
-      MatcherRewriteMode rewrite_mode =
-                          RewriteMode(FLAGS_sigma_fst_rewrite_mode))
-      : sigma_label_(sigma_label), rewrite_mode_(rewrite_mode) {}
-
-  SigmaFstMatcherData(const SigmaFstMatcherData &data)
-      : sigma_label_(data.sigma_label_), rewrite_mode_(data.rewrite_mode_) {}
-
-  static SigmaFstMatcherData<Label> *Read(std::istream &istrm,
-                                      const FstReadOptions &read) {
-    auto *data = new SigmaFstMatcherData<Label>();
-    ReadType(istrm, &data->sigma_label_);
-    int32 rewrite_mode;
-    ReadType(istrm, &rewrite_mode);
-    data->rewrite_mode_ = static_cast<MatcherRewriteMode>(rewrite_mode);
-    return data;
-  }
-
-  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
-    WriteType(ostrm, sigma_label_);
-    WriteType(ostrm, static_cast<int32>(rewrite_mode_));
-    return !ostrm ? false : true;
-  }
-
-  Label SigmaLabel() const { return sigma_label_; }
-
-  MatcherRewriteMode RewriteMode() const { return rewrite_mode_; }
-
- private:
-  static MatcherRewriteMode RewriteMode(const string &mode) {
-    if (mode == "auto") return MATCHER_REWRITE_AUTO;
-    if (mode == "always") return MATCHER_REWRITE_ALWAYS;
-    if (mode == "never") return MATCHER_REWRITE_NEVER;
-    LOG(WARNING) << "SigmaFst: Unknown rewrite mode: " << mode << ". "
-                 << "Defaulting to auto.";
-    return MATCHER_REWRITE_AUTO;
-  }
-
-  Label sigma_label_;
-  MatcherRewriteMode rewrite_mode_;
-};
-
-}  // namespace internal
-
-constexpr uint8 kSigmaFstMatchInput = 0x01;   // Input matcher is SigmaMatcher.
-constexpr uint8 kSigmaFstMatchOutput = 0x02;  // Output matcher is SigmaMatcher.
-
-template <class M, uint8 flags = kSigmaFstMatchInput | kSigmaFstMatchOutput>
-class SigmaFstMatcher : public SigmaMatcher<M> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename M::Arc;
-  using StateId = typename Arc::StateId;
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-  using MatcherData = internal::SigmaFstMatcherData<Label>;
-
-  enum : uint8 { kFlags = flags };
-
-  SigmaFstMatcher(
-      const FST &fst, MatchType match_type,
-      std::shared_ptr<MatcherData> data = std::make_shared<MatcherData>())
-      : SigmaMatcher<M>(
-            fst, match_type,
-            SigmaLabel(match_type,
-                       data ? data->SigmaLabel() : MatcherData().SigmaLabel()),
-            data ? data->RewriteMode() : MatcherData().RewriteMode()),
-        data_(data) {}
-
-  SigmaFstMatcher(const SigmaFstMatcher<M, flags> &matcher, bool safe = false)
-      : SigmaMatcher<M>(matcher, false), data_(matcher.data_) {}
-
-  SigmaFstMatcher<M, flags> *Copy(bool safe = false) const override {
-    return new SigmaFstMatcher<M, flags>(*this, safe);
-  }
-
-  const MatcherData *GetData() const { return data_.get(); }
-
-  std::shared_ptr<MatcherData> GetSharedData() const { return data_; }
-
- private:
-  static Label SigmaLabel(MatchType match_type, Label label) {
-    if (match_type == MATCH_INPUT && flags & kSigmaFstMatchInput) return label;
-    if (match_type == MATCH_OUTPUT && flags & kSigmaFstMatchOutput)
-      return label;
-    return kNoLabel;
-  }
-
-  std::shared_ptr<MatcherData> data_;
-};
-
-extern const char sigma_fst_type[];
-extern const char input_sigma_fst_type[];
-extern const char output_sigma_fst_type[];
-
-using StdSigmaFst = MatcherFst<ConstFst<StdArc>,
-                               SigmaFstMatcher<SortedMatcher<ConstFst<StdArc>>>,
-                               sigma_fst_type>;
-
-using LogSigmaFst = MatcherFst<ConstFst<LogArc>,
-                               SigmaFstMatcher<SortedMatcher<ConstFst<LogArc>>>,
-                               sigma_fst_type>;
-
-using Log64SigmaFst =
-    MatcherFst<ConstFst<Log64Arc>,
-               SigmaFstMatcher<SortedMatcher<ConstFst<Log64Arc>>>,
-               input_sigma_fst_type>;
-
-using StdInputSigmaFst = MatcherFst<
-    ConstFst<StdArc>,
-    SigmaFstMatcher<SortedMatcher<ConstFst<StdArc>>, kSigmaFstMatchInput>,
-    input_sigma_fst_type>;
-
-using LogInputSigmaFst = MatcherFst<
-    ConstFst<LogArc>,
-    SigmaFstMatcher<SortedMatcher<ConstFst<LogArc>>, kSigmaFstMatchInput>,
-    input_sigma_fst_type>;
-
-using Log64InputSigmaFst = MatcherFst<
-    ConstFst<Log64Arc>,
-    SigmaFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kSigmaFstMatchInput>,
-    input_sigma_fst_type>;
-
-using StdOutputSigmaFst = MatcherFst<
-    ConstFst<StdArc>,
-    SigmaFstMatcher<SortedMatcher<ConstFst<StdArc>>, kSigmaFstMatchOutput>,
-    output_sigma_fst_type>;
-
-using LogOutputSigmaFst = MatcherFst<
-    ConstFst<LogArc>,
-    SigmaFstMatcher<SortedMatcher<ConstFst<LogArc>>, kSigmaFstMatchOutput>,
-    output_sigma_fst_type>;
-
-using Log64OutputSigmaFst = MatcherFst<
-    ConstFst<Log64Arc>,
-    SigmaFstMatcher<SortedMatcher<ConstFst<Log64Arc>>, kSigmaFstMatchOutput>,
-    output_sigma_fst_type>;
-
-}  // namespace fst
-
-#endif  // FST_EXTENSIONS_SPECIAL_SIGMA_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/factor-weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/factor-weight.h
deleted file mode 100644
index a9dde875..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/factor-weight.h
+++ /dev/null
@@ -1,496 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes to factor weights in an FST.
-
-#ifndef FST_FACTOR_WEIGHT_H_
-#define FST_FACTOR_WEIGHT_H_
-
-#include <algorithm>
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-constexpr uint32 kFactorFinalWeights = 0x00000001;
-constexpr uint32 kFactorArcWeights = 0x00000002;
-
-template <class Arc>
-struct FactorWeightOptions : CacheOptions {
-  using Label = typename Arc::Label;
-
-  float delta;
-  uint32 mode;         // Factor arc weights and/or final weights.
-  Label final_ilabel;  // Input label of arc when factoring final weights.
-  Label final_olabel;  // Output label of arc when factoring final weights.
-  bool increment_final_ilabel;  // When factoring final w' results in > 1 arcs
-  bool increment_final_olabel;  // at state, increment labels to make distinct?
-
-  explicit FactorWeightOptions(const CacheOptions &opts, float delta = kDelta,
-                               uint32 mode = kFactorArcWeights |
-                                             kFactorFinalWeights,
-                               Label final_ilabel = 0, Label final_olabel = 0,
-                               bool increment_final_ilabel = false,
-                               bool increment_final_olabel = false)
-      : CacheOptions(opts),
-        delta(delta),
-        mode(mode),
-        final_ilabel(final_ilabel),
-        final_olabel(final_olabel),
-        increment_final_ilabel(increment_final_ilabel),
-        increment_final_olabel(increment_final_olabel) {}
-
-  explicit FactorWeightOptions(float delta = kDelta,
-                               uint32 mode = kFactorArcWeights |
-                                             kFactorFinalWeights,
-                               Label final_ilabel = 0, Label final_olabel = 0,
-                               bool increment_final_ilabel = false,
-                               bool increment_final_olabel = false)
-      : delta(delta),
-        mode(mode),
-        final_ilabel(final_ilabel),
-        final_olabel(final_olabel),
-        increment_final_ilabel(increment_final_ilabel),
-        increment_final_olabel(increment_final_olabel) {}
-};
-
-// A factor iterator takes as argument a weight w and returns a sequence of
-// pairs of weights (xi, yi) such that the sum of the products xi times yi is
-// equal to w. If w is fully factored, the iterator should return nothing.
-//
-// template <class W>
-// class FactorIterator {
-//  public:
-//   explicit FactorIterator(W w);
-//
-//   bool Done() const;
-//
-//   void Next();
-//
-//   std::pair<W, W> Value() const;
-//
-//   void Reset();
-// }
-
-// Factors trivially.
-template <class W>
-class IdentityFactor {
- public:
-  explicit IdentityFactor(const W &weight) {}
-
-  bool Done() const { return true; }
-
-  void Next() {}
-
-  std::pair<W, W> Value() const { return std::make_pair(W::One(), W::One()); }
-
-  void Reset() {}
-};
-
-// Factors a StringWeight w as 'ab' where 'a' is a label.
-template <typename Label, StringType S = STRING_LEFT>
-class StringFactor {
- public:
-  explicit StringFactor(const StringWeight<Label, S> &weight)
-      : weight_(weight), done_(weight.Size() <= 1) {}
-
-  bool Done() const { return done_; }
-
-  void Next() { done_ = true; }
-
-  std::pair<StringWeight<Label, S>, StringWeight<Label, S>> Value() const {
-    using Weight = StringWeight<Label, S>;
-    typename Weight::Iterator siter(weight_);
-    Weight w1(siter.Value());
-    Weight w2;
-    for (siter.Next(); !siter.Done(); siter.Next()) w2.PushBack(siter.Value());
-    return std::make_pair(w1, w2);
-  }
-
-  void Reset() { done_ = weight_.Size() <= 1; }
-
- private:
-  const StringWeight<Label, S> weight_;
-  bool done_;
-};
-
-// Factor a GallicWeight using StringFactor.
-template <class Label, class W, GallicType G = GALLIC_LEFT>
-class GallicFactor {
- public:
-  using GW = GallicWeight<Label, W, G>;
-
-  explicit GallicFactor(const GW &weight)
-      : weight_(weight), done_(weight.Value1().Size() <= 1) {}
-
-  bool Done() const { return done_; }
-
-  void Next() { done_ = true; }
-
-  std::pair<GW, GW> Value() const {
-    StringFactor<Label, GallicStringType(G)> siter(weight_.Value1());
-    GW w1(siter.Value().first, weight_.Value2());
-    GW w2(siter.Value().second, W::One());
-    return std::make_pair(w1, w2);
-  }
-
-  void Reset() { done_ = weight_.Value1().Size() <= 1; }
-
- private:
-  const GW weight_;
-  bool done_;
-};
-
-// Specialization for the (general) GALLIC type GallicWeight.
-template <class Label, class W>
-class GallicFactor<Label, W, GALLIC> {
- public:
-  using GW = GallicWeight<Label, W, GALLIC>;
-  using GRW = GallicWeight<Label, W, GALLIC_RESTRICT>;
-
-  explicit GallicFactor(const GW &weight)
-      : iter_(weight),
-        done_(weight.Size() == 0 ||
-              (weight.Size() == 1 && weight.Back().Value1().Size() <= 1)) {}
-
-  bool Done() const { return done_ || iter_.Done(); }
-
-  void Next() { iter_.Next(); }
-
-  void Reset() { iter_.Reset(); }
-
-  std::pair<GW, GW> Value() const {
-    const auto weight = iter_.Value();
-    StringFactor<Label, GallicStringType(GALLIC_RESTRICT)> siter(
-        weight.Value1());
-    GRW w1(siter.Value().first, weight.Value2());
-    GRW w2(siter.Value().second, W::One());
-    return std::make_pair(GW(w1), GW(w2));
-  }
-
- private:
-  UnionWeightIterator<GRW, GallicUnionWeightOptions<Label, W>> iter_;
-  bool done_;
-};
-
-namespace internal {
-
-// Implementation class for FactorWeight
-template <class Arc, class FactorIterator>
-class FactorWeightFstImpl : public CacheImpl<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheBaseImpl<CacheState<Arc>>::PushArc;
-  using CacheBaseImpl<CacheState<Arc>>::HasStart;
-  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
-  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
-  using CacheBaseImpl<CacheState<Arc>>::SetArcs;
-  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
-  using CacheBaseImpl<CacheState<Arc>>::SetStart;
-
-  struct Element {
-    Element() {}
-
-    Element(StateId s, Weight weight_) : state(s), weight(std::move(weight_)) {}
-
-    StateId state;  // Input state ID.
-    Weight weight;  // Residual weight.
-  };
-
-  FactorWeightFstImpl(const Fst<Arc> &fst, const FactorWeightOptions<Arc> &opts)
-      : CacheImpl<Arc>(opts),
-        fst_(fst.Copy()),
-        delta_(opts.delta),
-        mode_(opts.mode),
-        final_ilabel_(opts.final_ilabel),
-        final_olabel_(opts.final_olabel),
-        increment_final_ilabel_(opts.increment_final_ilabel),
-        increment_final_olabel_(opts.increment_final_olabel) {
-    SetType("factor_weight");
-    const auto props = fst.Properties(kFstProperties, false);
-    SetProperties(FactorWeightProperties(props), kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-    if (mode_ == 0) {
-      LOG(WARNING) << "FactorWeightFst: Factor mode is set to 0; "
-                   << "factoring neither arc weights nor final weights";
-    }
-  }
-
-  FactorWeightFstImpl(const FactorWeightFstImpl<Arc, FactorIterator> &impl)
-      : CacheImpl<Arc>(impl),
-        fst_(impl.fst_->Copy(true)),
-        delta_(impl.delta_),
-        mode_(impl.mode_),
-        final_ilabel_(impl.final_ilabel_),
-        final_olabel_(impl.final_olabel_),
-        increment_final_ilabel_(impl.increment_final_ilabel_),
-        increment_final_olabel_(impl.increment_final_olabel_) {
-    SetType("factor_weight");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      const auto s = fst_->Start();
-      if (s == kNoStateId) return kNoStateId;
-      SetStart(FindState(Element(fst_->Start(), Weight::One())));
-    }
-    return CacheImpl<Arc>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) {
-      const auto &element = elements_[s];
-      // TODO(sorenj): fix so cast is unnecessary
-      const auto weight =
-          element.state == kNoStateId
-              ? element.weight
-              : (Weight)Times(element.weight, fst_->Final(element.state));
-      FactorIterator siter(weight);
-      if (!(mode_ & kFactorFinalWeights) || siter.Done()) {
-        SetFinal(s, weight);
-      } else {
-        SetFinal(s, Weight::Zero());
-      }
-    }
-    return CacheImpl<Arc>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumOutputEpsilons(s);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && fst_->Properties(kError, false)) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<Arc>::InitArcIterator(s, data);
-  }
-
-  // Finds state corresponding to an element, creating new state if element not
-  // found.
-  StateId FindState(const Element &element) {
-    if (!(mode_ & kFactorArcWeights) && element.weight == Weight::One() &&
-        element.state != kNoStateId) {
-      while (unfactored_.size() <= element.state)
-        unfactored_.push_back(kNoStateId);
-      if (unfactored_[element.state] == kNoStateId) {
-        unfactored_[element.state] = elements_.size();
-        elements_.push_back(element);
-      }
-      return unfactored_[element.state];
-    } else {
-      const auto insert_result =
-          element_map_.insert(std::make_pair(element, elements_.size()));
-      if (insert_result.second) {
-        elements_.push_back(element);
-      }
-      return insert_result.first->second;
-    }
-  }
-
-  // Computes the outgoing transitions from a state, creating new destination
-  // states as needed.
-  void Expand(StateId s) {
-    const auto element = elements_[s];
-    if (element.state != kNoStateId) {
-      for (ArcIterator<Fst<Arc>> ait(*fst_, element.state); !ait.Done();
-           ait.Next()) {
-        const auto &arc = ait.Value();
-        const auto weight = Times(element.weight, arc.weight);
-        FactorIterator fiter(weight);
-        if (!(mode_ & kFactorArcWeights) || fiter.Done()) {
-          const auto dest = FindState(Element(arc.nextstate, Weight::One()));
-          PushArc(s, Arc(arc.ilabel, arc.olabel, weight, dest));
-        } else {
-          for (; !fiter.Done(); fiter.Next()) {
-            const auto &pair = fiter.Value();
-            const auto dest =
-                FindState(Element(arc.nextstate, pair.second.Quantize(delta_)));
-            PushArc(s, Arc(arc.ilabel, arc.olabel, pair.first, dest));
-          }
-        }
-      }
-    }
-    if ((mode_ & kFactorFinalWeights) &&
-        ((element.state == kNoStateId) ||
-         (fst_->Final(element.state) != Weight::Zero()))) {
-      const auto weight =
-          element.state == kNoStateId
-              ? element.weight
-              : Times(element.weight, fst_->Final(element.state));
-      auto ilabel = final_ilabel_;
-      auto olabel = final_olabel_;
-      for (FactorIterator fiter(weight); !fiter.Done(); fiter.Next()) {
-        const auto &pair = fiter.Value();
-        const auto dest =
-            FindState(Element(kNoStateId, pair.second.Quantize(delta_)));
-        PushArc(s, Arc(ilabel, olabel, pair.first, dest));
-        if (increment_final_ilabel_) ++ilabel;
-        if (increment_final_olabel_) ++olabel;
-      }
-    }
-    SetArcs(s);
-  }
-
- private:
-  // Equality function for Elements, assume weights have been quantized.
-  class ElementEqual {
-   public:
-    bool operator()(const Element &x, const Element &y) const {
-      return x.state == y.state && x.weight == y.weight;
-    }
-  };
-
-  // Hash function for Elements to Fst states.
-  class ElementKey {
-   public:
-    size_t operator()(const Element &x) const {
-      static constexpr auto prime = 7853;
-      return static_cast<size_t>(x.state * prime + x.weight.Hash());
-    }
-  };
-
-  using ElementMap =
-      std::unordered_map<Element, StateId, ElementKey, ElementEqual>;
-
-  std::unique_ptr<const Fst<Arc>> fst_;
-  float delta_;
-  uint32 mode_;         // Factoring arc and/or final weights.
-  Label final_ilabel_;  // ilabel of arc created when factoring final weights.
-  Label final_olabel_;  // olabel of arc created when factoring final weights.
-  bool increment_final_ilabel_;    // When factoring final weights results in
-  bool increment_final_olabel_;    // mutiple arcs, increment labels?
-  std::vector<Element> elements_;  // mapping from FST state to Element.
-  ElementMap element_map_;         // mapping from Element to FST state.
-  // Mapping between old/new StateId for states that do not need to be factored
-  // when mode_ is 0 or kFactorFinalWeights.
-  std::vector<StateId> unfactored_;
-};
-
-}  // namespace internal
-
-// FactorWeightFst takes as template parameter a FactorIterator as defined
-// above. The result of weight factoring is a transducer equivalent to the
-// input whose path weights have been factored according to the FactorIterator.
-// States and transitions will be added as necessary. The algorithm is a
-// generalization to arbitrary weights of the second step of the input
-// epsilon-normalization algorithm.
-//
-// This class attaches interface to implementation and handles reference
-// counting, delegating most methods to ImplToFst.
-template <class A, class FactorIterator>
-class FactorWeightFst
-    : public ImplToFst<internal::FactorWeightFstImpl<A, FactorIterator>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::FactorWeightFstImpl<Arc, FactorIterator>;
-
-  friend class ArcIterator<FactorWeightFst<Arc, FactorIterator>>;
-  friend class StateIterator<FactorWeightFst<Arc, FactorIterator>>;
-
-  explicit FactorWeightFst(const Fst<Arc> &fst)
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, FactorWeightOptions<Arc>())) {}
-
-  FactorWeightFst(const Fst<Arc> &fst, const FactorWeightOptions<Arc> &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  FactorWeightFst(const FactorWeightFst<Arc, FactorIterator> &fst, bool copy)
-      : ImplToFst<Impl>(fst, copy) {}
-
-  // Get a copy of this FactorWeightFst. See Fst<>::Copy() for further doc.
-  FactorWeightFst<Arc, FactorIterator> *Copy(bool copy = false) const override {
-    return new FactorWeightFst<Arc, FactorIterator>(*this, copy);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  FactorWeightFst &operator=(const FactorWeightFst &) = delete;
-};
-
-// Specialization for FactorWeightFst.
-template <class Arc, class FactorIterator>
-class StateIterator<FactorWeightFst<Arc, FactorIterator>>
-    : public CacheStateIterator<FactorWeightFst<Arc, FactorIterator>> {
- public:
-  explicit StateIterator(const FactorWeightFst<Arc, FactorIterator> &fst)
-      : CacheStateIterator<FactorWeightFst<Arc, FactorIterator>>(
-            fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for FactorWeightFst.
-template <class Arc, class FactorIterator>
-class ArcIterator<FactorWeightFst<Arc, FactorIterator>>
-    : public CacheArcIterator<FactorWeightFst<Arc, FactorIterator>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const FactorWeightFst<Arc, FactorIterator> &fst, StateId s)
-      : CacheArcIterator<FactorWeightFst<Arc, FactorIterator>>(
-            fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc, class FactorIterator>
-inline void FactorWeightFst<Arc, FactorIterator>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<FactorWeightFst<Arc, FactorIterator>>(*this);
-}
-
-}  // namespace fst
-
-#endif  // FST_FACTOR_WEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/filter-state.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/filter-state.h
deleted file mode 100644
index 289b628e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/filter-state.h
+++ /dev/null
@@ -1,199 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes for storing filter state in various algorithms like composition.
-
-#ifndef FST_FILTER_STATE_H_
-#define FST_FILTER_STATE_H_
-
-#include <forward_list>
-#include <utility>
-
-#include <fst/fst-decl.h>  // For optional argument declarations
-#include <fst/fst.h>
-#include <fst/matcher.h>
-
-
-namespace fst {
-
-// The filter state interface represents the state of a (e.g., composition)
-// filter.
-//
-// class FilterState {
-//  public:
-//   // Required constructors.
-//
-//   FilterState();
-//
-//   FilterState(const FilterState &fs);
-//
-//   // An invalid filter state.
-//   static const FilterState NoState();
-//
-//   // Maps state to integer for hashing.
-//   size_t Hash() const;
-//
-//   // Equality of filter states.
-//   bool operator==(const FilterState &fs) const;
-//
-//   // Inequality of filter states.
-//   bool operator!=(const FilterState &fs) const;
-//
-//   // Assignment to filter states.
-//   FilterState &operator=(const FilterState& fs);
-// };
-
-// Filter state that is a signed integral type.
-template <typename T>
-class IntegerFilterState {
- public:
-  IntegerFilterState() : state_(kNoStateId) {}
-
-  explicit IntegerFilterState(T s) : state_(s) {}
-
-  static const IntegerFilterState NoState() { return IntegerFilterState(); }
-
-  size_t Hash() const { return static_cast<size_t>(state_); }
-
-  bool operator==(const IntegerFilterState &fs) const {
-    return state_ == fs.state_;
-  }
-
-  bool operator!=(const IntegerFilterState &fs) const {
-    return state_ != fs.state_;
-  }
-
-  T GetState() const { return state_; }
-
-  void SetState(T state) { state_ = state; }
-
- private:
-  T state_;
-};
-
-using CharFilterState = IntegerFilterState<signed char>;
-using ShortFilterState = IntegerFilterState<short>;  // NOLINT
-using IntFilterState = IntegerFilterState<int>;
-
-// Filter state that is a weight (class).
-template <class W>
-class WeightFilterState {
- public:
-  WeightFilterState() : weight_(W::Zero()) {}
-
-  explicit WeightFilterState(W weight) : weight_(std::move(weight)) {}
-
-  static const WeightFilterState NoState() { return WeightFilterState(); }
-
-  size_t Hash() const { return weight_.Hash(); }
-
-  bool operator==(const WeightFilterState &fs) const {
-    return weight_ == fs.weight_;
-  }
-
-  bool operator!=(const WeightFilterState &fs) const {
-    return weight_ != fs.weight_;
-  }
-
-  W GetWeight() const { return weight_; }
-
-  void SetWeight(W weight) { weight_ = std::move(weight); }
-
- private:
-  W weight_;
-};
-
-// Filter state is a list of signed integer types T. Order matters
-// for equality.
-template <typename T>
-class ListFilterState {
- public:
-  ListFilterState() {}
-
-  explicit ListFilterState(T s) { list_.push_front(s); }
-
-  static const ListFilterState NoState() { return ListFilterState(kNoStateId); }
-
-  size_t Hash() const {
-    size_t h = 0;
-    for (const auto &elem : list_) h ^= h << 1 ^ elem;
-    return h;
-  }
-
-  bool operator==(const ListFilterState &fs) const { return list_ == fs.list_; }
-
-  bool operator!=(const ListFilterState &fs) const { return list_ != fs.list_; }
-
-  const std::forward_list<T> &GetState() const { return list_; }
-
-  std::forward_list<T> *GetMutableState() { return &list_; }
-
-  void SetState(const std::forward_list<T> &state) { list_ = state; }
-
- private:
-  std::forward_list<T> list_;
-};
-
-// Filter state that is the combination of two filter states.
-template <class FS1, class FS2>
-class PairFilterState {
- public:
-  PairFilterState() : fs1_(FS1::NoState()), fs2_(FS2::NoState()) {}
-
-  PairFilterState(const FS1 &fs1, const FS2 &fs2) : fs1_(fs1), fs2_(fs2) {}
-
-  static const PairFilterState NoState() { return PairFilterState(); }
-
-  size_t Hash() const {
-    const auto h1 = fs1_.Hash();
-    static constexpr auto lshift = 5;
-    static constexpr auto rshift = CHAR_BIT * sizeof(size_t) - 5;
-    return h1 << lshift ^ h1 >> rshift ^ fs2_.Hash();
-  }
-
-  bool operator==(const PairFilterState &fs) const {
-    return fs1_ == fs.fs1_ && fs2_ == fs.fs2_;
-  }
-
-  bool operator!=(const PairFilterState &fs) const {
-    return fs1_ != fs.fs1_ || fs2_ != fs.fs2_;
-  }
-
-  const FS1 &GetState1() const { return fs1_; }
-
-  const FS2 &GetState2() const { return fs2_; }
-
-  void SetState(const FS1 &fs1, const FS2 &fs2) {
-    fs1_ = fs1;
-    fs2_ = fs2;
-  }
-
- private:
-  FS1 fs1_;
-  FS2 fs2_;
-};
-
-// Single non-blocking filter state.
-class TrivialFilterState {
- public:
-  explicit TrivialFilterState(bool state = false) : state_(state) {}
-
-  static const TrivialFilterState NoState() { return TrivialFilterState(); }
-
-  size_t Hash() const { return 0; }
-
-  bool operator==(const TrivialFilterState &fs) const {
-    return state_ == fs.state_;
-  }
-
-  bool operator!=(const TrivialFilterState &fs) const {
-    return state_ != fs.state_;
-  }
-
- private:
-  bool state_;
-};
-
-}  // namespace fst
-
-#endif  // FST_FILTER_STATE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/flags.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/flags.h
deleted file mode 100644
index 94c5bb34..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/flags.h
+++ /dev/null
@@ -1,243 +0,0 @@
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-// 
-// Google-style flag handling declarations and inline definitions.
-
-#ifndef FST_LIB_FLAGS_H_
-#define FST_LIB_FLAGS_H_
-
-#include <cstdlib>
-
-#include <iostream>
-#include <map>
-#include <set>
-#include <sstream>
-#include <string>
-
-#include <fst/types.h>
-#include <fst/lock.h>
-
-using std::string;
-
-// FLAGS USAGE:
-//
-// Definition example:
-//
-//    DEFINE_int32(length, 0, "length");
-//
-// This defines variable FLAGS_length, initialized to 0.
-//
-// Declaration example:
-//
-//    DECLARE_int32(length);
-//
-// SET_FLAGS() can be used to set flags from the command line
-// using, for example, '--length=2'.
-//
-// ShowUsage() can be used to print out command and flag usage.
-#if defined(_WIN32)
-#if defined(FST_BUILD_SHARED_LIBS)
-#define FST_EXPORT __declspec(dllexport)
-#define FST_IMPORT __declspec(dllimport)
-#else
-#define FST_EXPORT
-#define FST_IMPORT
-#endif
-#else  // WIN32
-#define FST_EXPORT __attribute__((visibility("default")))
-
-#define FST_IMPORT FST_EXPORT
-#endif  // WIN32
-
-#if defined(FST_BUILD_MAIN_LIB)
-#define FST_API FST_EXPORT
-#else
-#define FST_API FST_IMPORT
-#endif
-
-#define DECLARE_bool(name) FST_API extern bool FLAGS_ ## name
-#define DECLARE_string(name) FST_API extern string FLAGS_ ## name
-#define DECLARE_int32(name) FST_API extern int32 FLAGS_ ## name
-#define DECLARE_int64(name) FST_API extern int64 FLAGS_ ## name
-#define DECLARE_double(name) FST_API extern double FLAGS_ ## name
-
-template <typename T>
-struct FlagDescription {
-  FlagDescription(T *addr, const char *doc, const char *type,
-		  const char *file, const T val)
-      : address(addr),
-    doc_string(doc),
-    type_name(type),
-    file_name(file),
-    default_value(val) {}
-
-  T *address;
-  const char *doc_string;
-  const char *type_name;
-  const char *file_name;
-  const T default_value;
-};
-
-template <typename T>
-class FlagRegister {
- public:
-  static FlagRegister<T> *GetRegister() {
-    static auto reg = new FlagRegister<T>;
-    return reg;
-  }
-
-  const FlagDescription<T> &GetFlagDescription(const string &name) const {
-    fst::MutexLock l(&flag_lock_);
-    auto it = flag_table_.find(name);
-    return it != flag_table_.end() ? it->second : 0;
-  }
-
-  void SetDescription(const string &name,
-                      const FlagDescription<T> &desc) {
-    fst::MutexLock l(&flag_lock_);
-    flag_table_.insert(make_pair(name, desc));
-  }
-
-  bool SetFlag(const string &val, bool *address) const {
-    if (val == "true" || val == "1" || val.empty()) {
-      *address = true;
-      return true;
-    } else if (val == "false" || val == "0") {
-      *address = false;
-      return true;
-    }
-    else {
-      return false;
-    }
-  }
-
-  bool SetFlag(const string &val, string *address) const {
-    *address = val;
-    return true;
-  }
-
-  bool SetFlag(const string &val, int32 *address) const {
-    char *p = 0;
-    *address = strtol(val.c_str(), &p, 0);
-    return !val.empty() && *p == '\0';
-  }
-
-  bool SetFlag(const string &val, int64 *address) const {
-    char *p = 0;
-    *address = strtoll(val.c_str(), &p, 0);
-    return !val.empty() && *p == '\0';
-  }
-
-  bool SetFlag(const string &val, double *address) const {
-    char *p = 0;
-    *address = strtod(val.c_str(), &p);
-    return !val.empty() && *p == '\0';
-  }
-
-  bool SetFlag(const string &arg, const string &val) const {
-    for (typename std::map< string, FlagDescription<T> >::const_iterator it =
-           flag_table_.begin();
-         it != flag_table_.end();
-         ++it) {
-      const string &name = it->first;
-      const FlagDescription<T> &desc = it->second;
-      if (arg == name)
-        return SetFlag(val, desc.address);
-    }
-    return false;
-  }
-
-  void GetUsage(std::set<std::pair<string, string>> *usage_set) const {
-    for (auto it = flag_table_.begin(); it != flag_table_.end(); ++it) {
-      const string &name = it->first;
-      const FlagDescription<T> &desc = it->second;
-      string usage = "  --" + name;
-      usage += ": type = ";
-      usage += desc.type_name;
-      usage += ", default = ";
-      usage += GetDefault(desc.default_value) + "\n  ";
-      usage += desc.doc_string;
-      usage_set->insert(make_pair(desc.file_name, usage));
-    }
-  }
-
- private:
-  string GetDefault(bool default_value) const {
-    return default_value ? "true" : "false";
-  }
-
-  string GetDefault(const string &default_value) const {
-    return "\"" + default_value + "\"";
-  }
-
-  template <class V>
-  string GetDefault(const V &default_value) const {
-    std::ostringstream strm;
-    strm << default_value;
-    return strm.str();
-  }
-
-  mutable fst::Mutex flag_lock_;        // Multithreading lock.
-  std::map<string, FlagDescription<T>> flag_table_;
-};
-
-template <typename T>
-class FlagRegisterer {
- public:
-  FlagRegisterer(const string &name, const FlagDescription<T> &desc) {
-    auto registr = FlagRegister<T>::GetRegister();
-    registr->SetDescription(name, desc);
-  }
-
- private:
-  FlagRegisterer(const FlagRegisterer &) = delete;
-  FlagRegisterer &operator=(const FlagRegisterer &) = delete;
-};
-
-
-#define DEFINE_VAR(type, name, value, doc)                                \
-  FST_API type FLAGS_ ## name = value;                                            \
-  static FlagRegisterer<type>                                             \
-  name ## _flags_registerer(#name, FlagDescription<type>(&FLAGS_ ## name, \
-                                                         doc,             \
-                                                         #type,           \
-                                                         __FILE__,        \
-                                                         value))
-
-#define DEFINE_bool(name, value, doc) DEFINE_VAR(bool, name, value, doc)
-#define DEFINE_string(name, value, doc) \
-  DEFINE_VAR(string, name, value, doc)
-#define DEFINE_int32(name, value, doc) DEFINE_VAR(int32, name, value, doc)
-#define DEFINE_int64(name, value, doc) DEFINE_VAR(int64, name, value, doc)
-#define DEFINE_double(name, value, doc) DEFINE_VAR(double, name, value, doc)
-
-
-// Temporary directory.
-DECLARE_string(tmpdir);
-
-void SetFlags(const char *usage, int *argc, char ***argv, bool remove_flags,
-              const char *src = "");
-
-#define SET_FLAGS(usage, argc, argv, rmflags) \
-SetFlags(usage, argc, argv, rmflags, __FILE__)
-
-// Deprecated; for backward compatibility.
-inline void InitFst(const char *usage, int *argc, char ***argv, bool rmflags) {
-  return SetFlags(usage, argc, argv, rmflags);
-}
-
-void ShowUsage(bool long_usage = true);
-
-#endif  // FST_LIB_FLAGS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/float-weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/float-weight.h
deleted file mode 100644
index 43c59134..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/float-weight.h
+++ /dev/null
@@ -1,780 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Float weight set and associated semiring operation definitions.
-
-#ifndef FST_FLOAT_WEIGHT_H_
-#define FST_FLOAT_WEIGHT_H_
-
-#include <climits>
-#include <cmath>
-#include <cstdlib>
-#include <cstring>
-
-#include <algorithm>
-#include <limits>
-#include <sstream>
-#include <string>
-
-#include <fst/util.h>
-#include <fst/weight.h>
-
-
-namespace fst {
-
-// Numeric limits class.
-template <class T>
-class FloatLimits {
- public:
-  static constexpr T PosInfinity() {
-    return std::numeric_limits<T>::infinity();
-  }
-
-  static constexpr T NegInfinity() { return -PosInfinity(); }
-
-  static constexpr T NumberBad() { return std::numeric_limits<T>::quiet_NaN(); }
-};
-
-// Weight class to be templated on floating-points types.
-template <class T = float>
-class FloatWeightTpl {
- public:
-  using ValueType = T;
-
-  FloatWeightTpl() {}
-
-  FloatWeightTpl(T f) : value_(f) {}
-
-  FloatWeightTpl(const FloatWeightTpl<T> &weight) : value_(weight.value_) {}
-
-  FloatWeightTpl<T> &operator=(const FloatWeightTpl<T> &weight) {
-    value_ = weight.value_;
-    return *this;
-  }
-
-  std::istream &Read(std::istream &strm) { return ReadType(strm, &value_); }
-
-  std::ostream &Write(std::ostream &strm) const {
-    return WriteType(strm, value_);
-  }
-
-  size_t Hash() const {
-    size_t hash = 0;
-    // Avoid using union, which would be undefined behavior.
-    // Use memcpy, similar to bit_cast, but sizes may be different.
-    // This should be optimized into a single move instruction by
-    // any reasonable compiler.
-    std::memcpy(&hash, &value_, std::min(sizeof(hash), sizeof(value_)));
-    return hash;
-  }
-
-  const T &Value() const { return value_; }
-
- protected:
-  void SetValue(const T &f) { value_ = f; }
-
-  static constexpr const char *GetPrecisionString() {
-    return sizeof(T) == 4
-               ? ""
-               : sizeof(T) == 1
-                     ? "8"
-                     : sizeof(T) == 2 ? "16"
-                                      : sizeof(T) == 8 ? "64" : "unknown";
-  }
-
- private:
-  T value_;
-};
-
-// Single-precision float weight.
-using FloatWeight = FloatWeightTpl<float>;
-
-template <class T>
-inline bool operator==(const FloatWeightTpl<T> &w1,
-                       const FloatWeightTpl<T> &w2) {
-  // Volatile qualifier thwarts over-aggressive compiler optimizations that
-  // lead to problems esp. with NaturalLess().
-  volatile T v1 = w1.Value();
-  volatile T v2 = w2.Value();
-  return v1 == v2;
-}
-
-// These seemingly unnecessary overloads are actually needed to make
-// comparisons like FloatWeightTpl<float> == float compile.  If only the
-// templated version exists, the FloatWeightTpl<float>(float) conversion
-// won't be found.
-inline bool operator==(const FloatWeightTpl<double> &w1,
-                       const FloatWeightTpl<double> &w2) {
-  return operator==<double>(w1, w2);
-}
-
-inline bool operator==(const FloatWeightTpl<float> &w1,
-                       const FloatWeightTpl<float> &w2) {
-  return operator==<float>(w1, w2);
-}
-
-template <class T>
-inline bool operator!=(const FloatWeightTpl<T> &w1,
-                       const FloatWeightTpl<T> &w2) {
-  return !(w1 == w2);
-}
-
-inline bool operator!=(const FloatWeightTpl<double> &w1,
-                       const FloatWeightTpl<double> &w2) {
-  return operator!=<double>(w1, w2);
-}
-
-inline bool operator!=(const FloatWeightTpl<float> &w1,
-                       const FloatWeightTpl<float> &w2) {
-  return operator!=<float>(w1, w2);
-}
-
-template <class T>
-inline bool ApproxEqual(const FloatWeightTpl<T> &w1,
-                        const FloatWeightTpl<T> &w2, float delta = kDelta) {
-  return w1.Value() <= w2.Value() + delta && w2.Value() <= w1.Value() + delta;
-}
-
-template <class T>
-inline std::ostream &operator<<(std::ostream &strm,
-                                const FloatWeightTpl<T> &w) {
-  if (w.Value() == FloatLimits<T>::PosInfinity()) {
-    return strm << "Infinity";
-  } else if (w.Value() == FloatLimits<T>::NegInfinity()) {
-    return strm << "-Infinity";
-  } else if (w.Value() != w.Value()) {  // Fails for IEEE NaN.
-    return strm << "BadNumber";
-  } else {
-    return strm << w.Value();
-  }
-}
-
-template <class T>
-inline std::istream &operator>>(std::istream &strm, FloatWeightTpl<T> &w) {
-  string s;
-  strm >> s;
-  if (s == "Infinity") {
-    w = FloatWeightTpl<T>(FloatLimits<T>::PosInfinity());
-  } else if (s == "-Infinity") {
-    w = FloatWeightTpl<T>(FloatLimits<T>::NegInfinity());
-  } else {
-    char *p;
-    T f = strtod(s.c_str(), &p);
-    if (p < s.c_str() + s.size()) {
-      strm.clear(std::ios::badbit);
-    } else {
-      w = FloatWeightTpl<T>(f);
-    }
-  }
-  return strm;
-}
-
-// Tropical semiring: (min, +, inf, 0).
-template <class T>
-class TropicalWeightTpl : public FloatWeightTpl<T> {
- public:
-  using typename FloatWeightTpl<T>::ValueType;
-  using FloatWeightTpl<T>::Value;
-  using ReverseWeight = TropicalWeightTpl<T>;
-  using Limits = FloatLimits<T>;
-
-  constexpr TropicalWeightTpl() : FloatWeightTpl<T>() {}
-
-  constexpr TropicalWeightTpl(T f) : FloatWeightTpl<T>(f) {}
-
-  constexpr TropicalWeightTpl(const TropicalWeightTpl<T> &weight)
-      : FloatWeightTpl<T>(weight) {}
-
-  static const TropicalWeightTpl<T> &Zero() {
-    static const TropicalWeightTpl zero(Limits::PosInfinity());
-    return zero;
-  }
-
-  static const TropicalWeightTpl<T> &One() {
-    static const TropicalWeightTpl one(0.0F);
-    return one;
-  }
-
-  static const TropicalWeightTpl<T> &NoWeight() {
-    static const TropicalWeightTpl no_weight(Limits::NumberBad());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type =
-        new string(string("tropical") +
-                   FloatWeightTpl<T>::GetPrecisionString());
-    return *type;
-  }
-
-  bool Member() const {
-    // First part fails for IEEE NaN.
-    return Value() == Value() && Value() != Limits::NegInfinity();
-  }
-
-  TropicalWeightTpl<T> Quantize(float delta = kDelta) const {
-    if (!Member() || Value() == Limits::PosInfinity()) {
-      return *this;
-    } else {
-      return TropicalWeightTpl<T>(floor(Value() / delta + 0.5F) * delta);
-    }
-  }
-
-  TropicalWeightTpl<T> Reverse() const { return *this; }
-
-  static constexpr uint64 Properties() {
-    return kLeftSemiring | kRightSemiring | kCommutative | kPath | kIdempotent;
-  }
-};
-
-// Single precision tropical weight.
-using TropicalWeight = TropicalWeightTpl<float>;
-
-template <class T>
-inline TropicalWeightTpl<T> Plus(const TropicalWeightTpl<T> &w1,
-                                 const TropicalWeightTpl<T> &w2) {
-  if (!w1.Member() || !w2.Member()) return TropicalWeightTpl<T>::NoWeight();
-  return w1.Value() < w2.Value() ? w1 : w2;
-}
-
-// See comment at operator==(FloatWeightTpl<float>, FloatWeightTpl<float>)
-// for why these overloads are present.
-inline TropicalWeightTpl<float> Plus(const TropicalWeightTpl<float> &w1,
-                                     const TropicalWeightTpl<float> &w2) {
-  return Plus<float>(w1, w2);
-}
-
-inline TropicalWeightTpl<double> Plus(const TropicalWeightTpl<double> &w1,
-                                      const TropicalWeightTpl<double> &w2) {
-  return Plus<double>(w1, w2);
-}
-
-template <class T>
-inline TropicalWeightTpl<T> Times(const TropicalWeightTpl<T> &w1,
-                                  const TropicalWeightTpl<T> &w2) {
-  using Limits = FloatLimits<T>;
-  if (!w1.Member() || !w2.Member()) return TropicalWeightTpl<T>::NoWeight();
-  const T f1 = w1.Value();
-  const T f2 = w2.Value();
-  if (f1 == Limits::PosInfinity()) {
-    return w1;
-  } else if (f2 == Limits::PosInfinity()) {
-    return w2;
-  } else {
-    return TropicalWeightTpl<T>(f1 + f2);
-  }
-}
-
-inline TropicalWeightTpl<float> Times(const TropicalWeightTpl<float> &w1,
-                                      const TropicalWeightTpl<float> &w2) {
-  return Times<float>(w1, w2);
-}
-
-inline TropicalWeightTpl<double> Times(const TropicalWeightTpl<double> &w1,
-                                       const TropicalWeightTpl<double> &w2) {
-  return Times<double>(w1, w2);
-}
-
-template <class T>
-inline TropicalWeightTpl<T> Divide(const TropicalWeightTpl<T> &w1,
-                                   const TropicalWeightTpl<T> &w2,
-                                   DivideType typ = DIVIDE_ANY) {
-  using Limits = FloatLimits<T>;
-  if (!w1.Member() || !w2.Member()) return TropicalWeightTpl<T>::NoWeight();
-  const T f1 = w1.Value();
-  const T f2 = w2.Value();
-  if (f2 == Limits::PosInfinity()) {
-    return Limits::NumberBad();
-  } else if (f1 == Limits::PosInfinity()) {
-    return Limits::PosInfinity();
-  } else {
-    return TropicalWeightTpl<T>(f1 - f2);
-  }
-}
-
-inline TropicalWeightTpl<float> Divide(const TropicalWeightTpl<float> &w1,
-                                       const TropicalWeightTpl<float> &w2,
-                                       DivideType typ = DIVIDE_ANY) {
-  return Divide<float>(w1, w2, typ);
-}
-
-inline TropicalWeightTpl<double> Divide(const TropicalWeightTpl<double> &w1,
-                                        const TropicalWeightTpl<double> &w2,
-                                        DivideType typ = DIVIDE_ANY) {
-  return Divide<double>(w1, w2, typ);
-}
-
-template <class T>
-inline TropicalWeightTpl<T> Power(const TropicalWeightTpl<T> &weight,
-                                  T scalar) {
-  return TropicalWeightTpl<T>(weight.Value() * scalar);
-}
-
-// Log semiring: (log(e^-x + e^-y), +, inf, 0).
-template <class T>
-class LogWeightTpl : public FloatWeightTpl<T> {
- public:
-  using typename FloatWeightTpl<T>::ValueType;
-  using FloatWeightTpl<T>::Value;
-  using ReverseWeight = LogWeightTpl;
-  using Limits = FloatLimits<T>;
-
-  constexpr LogWeightTpl() : FloatWeightTpl<T>() {}
-
-  constexpr LogWeightTpl(T f) : FloatWeightTpl<T>(f) {}
-
-  constexpr LogWeightTpl(const LogWeightTpl<T> &weight)
-      : FloatWeightTpl<T>(weight) {}
-
-  static const LogWeightTpl &Zero() {
-    static const LogWeightTpl zero(Limits::PosInfinity());
-    return zero;
-  }
-
-  static const LogWeightTpl &One() {
-    static const LogWeightTpl one(0.0F);
-    return one;
-  }
-
-  static const LogWeightTpl &NoWeight() {
-    static const LogWeightTpl no_weight(Limits::NumberBad());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type =
-        new string(string("log") + FloatWeightTpl<T>::GetPrecisionString());
-    return *type;
-  }
-
-  bool Member() const {
-    // First part fails for IEEE NaN.
-    return Value() == Value() && Value() != Limits::NegInfinity();
-  }
-
-  LogWeightTpl<T> Quantize(float delta = kDelta) const {
-    if (!Member() || Value() == Limits::PosInfinity()) {
-      return *this;
-    } else {
-      return LogWeightTpl<T>(floor(Value() / delta + 0.5F) * delta);
-    }
-  }
-
-  LogWeightTpl<T> Reverse() const { return *this; }
-
-  static constexpr uint64 Properties() {
-    return kLeftSemiring | kRightSemiring | kCommutative;
-  }
-};
-
-// Single-precision log weight.
-using LogWeight = LogWeightTpl<float>;
-
-// Double-precision log weight.
-using Log64Weight = LogWeightTpl<double>;
-
-namespace internal {
-
-// -log(e^-x + e^-y) = x - LogPosExp(y - x), assuming x >= 0.0.
-inline double LogPosExp(double x) {
-  DCHECK(!(x < 0));  // NB: NaN values are allowed.
-  return log1p(exp(-x));
-}
-
-// -log(e^-x - e^-y) = x - LogNegExp(y - x), assuming x > 0.0.
-inline double LogNegExp(double x) {
-  DCHECK_GT(x, 0);
-  return log1p(-exp(-x));
-}
-
-// a +_log b = -log(e^-a + e^-b) = KahanLogSum(a, b, ...).
-// Kahan compensated summation provides an error bound that is
-// independent of the number of addends. Assumes b >= a;
-// c is the compensation.
-inline double KahanLogSum(double a, double b, double *c) {
-  DCHECK_GE(b, a);
-  double y = -LogPosExp(b - a) - *c;
-  double t = a + y;
-  *c = (t - a) - y;
-  return t;
-}
-
-// a -_log b = -log(e^-a - e^-b) = KahanLogDiff(a, b, ...).
-// Kahan compensated summation provides an error bound that is
-// independent of the number of addends. Assumes b > a;
-// c is the compensation.
-inline double KahanLogDiff(double a, double b, double *c) {
-  DCHECK_GT(b, a);
-  double y = -LogNegExp(b - a) - *c;
-  double t = a + y;
-  *c = (t - a) - y;
-  return t;
-}
-
-}  // namespace internal
-
-template <class T>
-inline LogWeightTpl<T> Plus(const LogWeightTpl<T> &w1,
-                            const LogWeightTpl<T> &w2) {
-  using Limits = FloatLimits<T>;
-  const T f1 = w1.Value();
-  const T f2 = w2.Value();
-  if (f1 == Limits::PosInfinity()) {
-    return w2;
-  } else if (f2 == Limits::PosInfinity()) {
-    return w1;
-  } else if (f1 > f2) {
-    return LogWeightTpl<T>(f2 - internal::LogPosExp(f1 - f2));
-  } else {
-    return LogWeightTpl<T>(f1 - internal::LogPosExp(f2 - f1));
-  }
-}
-
-inline LogWeightTpl<float> Plus(const LogWeightTpl<float> &w1,
-                                const LogWeightTpl<float> &w2) {
-  return Plus<float>(w1, w2);
-}
-
-inline LogWeightTpl<double> Plus(const LogWeightTpl<double> &w1,
-                                 const LogWeightTpl<double> &w2) {
-  return Plus<double>(w1, w2);
-}
-
-template <class T>
-inline LogWeightTpl<T> Times(const LogWeightTpl<T> &w1,
-                             const LogWeightTpl<T> &w2) {
-  using Limits = FloatLimits<T>;
-  if (!w1.Member() || !w2.Member()) return LogWeightTpl<T>::NoWeight();
-  const T f1 = w1.Value();
-  const T f2 = w2.Value();
-  if (f1 == Limits::PosInfinity()) {
-    return w1;
-  } else if (f2 == Limits::PosInfinity()) {
-    return w2;
-  } else {
-    return LogWeightTpl<T>(f1 + f2);
-  }
-}
-
-inline LogWeightTpl<float> Times(const LogWeightTpl<float> &w1,
-                                 const LogWeightTpl<float> &w2) {
-  return Times<float>(w1, w2);
-}
-
-inline LogWeightTpl<double> Times(const LogWeightTpl<double> &w1,
-                                  const LogWeightTpl<double> &w2) {
-  return Times<double>(w1, w2);
-}
-
-template <class T>
-inline LogWeightTpl<T> Divide(const LogWeightTpl<T> &w1,
-                              const LogWeightTpl<T> &w2,
-                              DivideType typ = DIVIDE_ANY) {
-  using Limits = FloatLimits<T>;
-  if (!w1.Member() || !w2.Member()) return LogWeightTpl<T>::NoWeight();
-  const T f1 = w1.Value();
-  const T f2 = w2.Value();
-  if (f2 == Limits::PosInfinity()) {
-    return Limits::NumberBad();
-  } else if (f1 == Limits::PosInfinity()) {
-    return Limits::PosInfinity();
-  } else {
-    return LogWeightTpl<T>(f1 - f2);
-  }
-}
-
-inline LogWeightTpl<float> Divide(const LogWeightTpl<float> &w1,
-                                  const LogWeightTpl<float> &w2,
-                                  DivideType typ = DIVIDE_ANY) {
-  return Divide<float>(w1, w2, typ);
-}
-
-inline LogWeightTpl<double> Divide(const LogWeightTpl<double> &w1,
-                                   const LogWeightTpl<double> &w2,
-                                   DivideType typ = DIVIDE_ANY) {
-  return Divide<double>(w1, w2, typ);
-}
-
-template <class T>
-inline LogWeightTpl<T> Power(const LogWeightTpl<T> &weight, T scalar) {
-  return LogWeightTpl<T>(weight.Value() * scalar);
-}
-
-// Specialization using the Kahan compensated summation.
-template <class T>
-class Adder<LogWeightTpl<T>> {
- public:
-  using Weight = LogWeightTpl<T>;
-
-  explicit Adder(Weight w = Weight::Zero())
-      : sum_(w.Value()),
-        c_(0.0) { }
-
-  Weight Add(const Weight &w) {
-    using Limits = FloatLimits<T>;
-    const T f = w.Value();
-    if (f == Limits::PosInfinity()) {
-      return Sum();
-    } else if (sum_ == Limits::PosInfinity()) {
-      sum_ = f;
-      c_ = 0.0;
-    } else if (f > sum_) {
-      sum_ = internal::KahanLogSum(sum_, f, &c_);
-    } else {
-      sum_ = internal::KahanLogSum(f, sum_, &c_);
-    }
-    return Sum();
-  }
-
-  Weight Sum() { return Weight(sum_); }
-
-  void Reset(Weight w = Weight::Zero()) {
-    sum_ = w.Value();
-    c_ = 0.0;
-  }
-
- private:
-  double sum_;
-  double c_;   // Kahan compensation.
-};
-
-// MinMax semiring: (min, max, inf, -inf).
-template <class T>
-class MinMaxWeightTpl : public FloatWeightTpl<T> {
- public:
-  using typename FloatWeightTpl<T>::ValueType;
-  using FloatWeightTpl<T>::Value;
-  using ReverseWeight = MinMaxWeightTpl<T>;
-  using Limits = FloatLimits<T>;
-
-  MinMaxWeightTpl() : FloatWeightTpl<T>() {}
-
-  MinMaxWeightTpl(T f) : FloatWeightTpl<T>(f) {}
-
-  MinMaxWeightTpl(const MinMaxWeightTpl<T> &weight)
-      : FloatWeightTpl<T>(weight) {}
-
-  static const MinMaxWeightTpl &Zero() {
-    static const MinMaxWeightTpl zero(Limits::PosInfinity());
-    return zero;
-  }
-
-  static const MinMaxWeightTpl &One() {
-    static const MinMaxWeightTpl one(Limits::NegInfinity());
-    return one;
-  }
-
-  static const MinMaxWeightTpl &NoWeight() {
-    static const MinMaxWeightTpl no_weight(Limits::NumberBad());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type =
-        new string(string("minmax") + FloatWeightTpl<T>::GetPrecisionString());
-    return *type;
-  }
-
-  // Fails for IEEE NaN.
-  bool Member() const { return Value() == Value(); }
-
-  MinMaxWeightTpl<T> Quantize(float delta = kDelta) const {
-    // If one of infinities, or a NaN.
-    if (!Member() ||
-        Value() == Limits::NegInfinity() || Value() == Limits::PosInfinity()) {
-      return *this;
-    } else {
-      return MinMaxWeightTpl<T>(floor(Value() / delta + 0.5F) * delta);
-    }
-  }
-
-  MinMaxWeightTpl<T> Reverse() const { return *this; }
-
-  static constexpr uint64 Properties() {
-    return kLeftSemiring | kRightSemiring | kCommutative | kIdempotent | kPath;
-  }
-};
-
-// Single-precision min-max weight.
-using MinMaxWeight = MinMaxWeightTpl<float>;
-
-// Min.
-template <class T>
-inline MinMaxWeightTpl<T> Plus(const MinMaxWeightTpl<T> &w1,
-                               const MinMaxWeightTpl<T> &w2) {
-  if (!w1.Member() || !w2.Member()) return MinMaxWeightTpl<T>::NoWeight();
-  return w1.Value() < w2.Value() ? w1 : w2;
-}
-
-inline MinMaxWeightTpl<float> Plus(const MinMaxWeightTpl<float> &w1,
-                                   const MinMaxWeightTpl<float> &w2) {
-  return Plus<float>(w1, w2);
-}
-
-inline MinMaxWeightTpl<double> Plus(const MinMaxWeightTpl<double> &w1,
-                                    const MinMaxWeightTpl<double> &w2) {
-  return Plus<double>(w1, w2);
-}
-
-// Max.
-template <class T>
-inline MinMaxWeightTpl<T> Times(const MinMaxWeightTpl<T> &w1,
-                                const MinMaxWeightTpl<T> &w2) {
-  if (!w1.Member() || !w2.Member()) return MinMaxWeightTpl<T>::NoWeight();
-  return w1.Value() >= w2.Value() ? w1 : w2;
-}
-
-inline MinMaxWeightTpl<float> Times(const MinMaxWeightTpl<float> &w1,
-                                    const MinMaxWeightTpl<float> &w2) {
-  return Times<float>(w1, w2);
-}
-
-inline MinMaxWeightTpl<double> Times(const MinMaxWeightTpl<double> &w1,
-                                     const MinMaxWeightTpl<double> &w2) {
-  return Times<double>(w1, w2);
-}
-
-// Defined only for special cases.
-template <class T>
-inline MinMaxWeightTpl<T> Divide(const MinMaxWeightTpl<T> &w1,
-                                 const MinMaxWeightTpl<T> &w2,
-                                 DivideType typ = DIVIDE_ANY) {
-  if (!w1.Member() || !w2.Member()) return MinMaxWeightTpl<T>::NoWeight();
-  // min(w1, x) = w2, w1 >= w2 => min(w1, x) = w2, x = w2.
-  return w1.Value() >= w2.Value() ? w1 : FloatLimits<T>::NumberBad();
-}
-
-inline MinMaxWeightTpl<float> Divide(const MinMaxWeightTpl<float> &w1,
-                                     const MinMaxWeightTpl<float> &w2,
-                                     DivideType typ = DIVIDE_ANY) {
-  return Divide<float>(w1, w2, typ);
-}
-
-inline MinMaxWeightTpl<double> Divide(const MinMaxWeightTpl<double> &w1,
-                                      const MinMaxWeightTpl<double> &w2,
-                                      DivideType typ = DIVIDE_ANY) {
-  return Divide<double>(w1, w2, typ);
-}
-
-// Converts to tropical.
-template <>
-struct WeightConvert<LogWeight, TropicalWeight> {
-  TropicalWeight operator()(const LogWeight &w) const { return w.Value(); }
-};
-
-template <>
-struct WeightConvert<Log64Weight, TropicalWeight> {
-  TropicalWeight operator()(const Log64Weight &w) const { return w.Value(); }
-};
-
-// Converts to log.
-template <>
-struct WeightConvert<TropicalWeight, LogWeight> {
-  LogWeight operator()(const TropicalWeight &w) const { return w.Value(); }
-};
-
-template <>
-struct WeightConvert<Log64Weight, LogWeight> {
-  LogWeight operator()(const Log64Weight &w) const { return w.Value(); }
-};
-
-// Converts to log64.
-template <>
-struct WeightConvert<TropicalWeight, Log64Weight> {
-  Log64Weight operator()(const TropicalWeight &w) const { return w.Value(); }
-};
-
-template <>
-struct WeightConvert<LogWeight, Log64Weight> {
-  Log64Weight operator()(const LogWeight &w) const { return w.Value(); }
-};
-
-// This function object returns random integers chosen from [0,
-// num_random_weights). The boolean 'allow_zero' determines whether Zero() and
-// zero divisors should be returned in the random weight generation. This is
-// intended primary for testing.
-template <class Weight>
-class FloatWeightGenerate {
- public:
-  explicit FloatWeightGenerate(
-      bool allow_zero = true,
-      const size_t num_random_weights = kNumRandomWeights)
-      : allow_zero_(allow_zero), num_random_weights_(num_random_weights) {}
-
-  Weight operator()() const {
-    const int n = rand() % (num_random_weights_ + allow_zero_);  // NOLINT
-    if (allow_zero_ && n == num_random_weights_) return Weight::Zero();
-    return Weight(n);
-  }
-
- private:
-  // Permits Zero() and zero divisors.
-  const bool allow_zero_;
-  // Number of alternative random weights.
-  const size_t num_random_weights_;
-};
-
-template <class T>
-class WeightGenerate<TropicalWeightTpl<T>>
-    : public FloatWeightGenerate<TropicalWeightTpl<T>> {
- public:
-  using Weight = TropicalWeightTpl<T>;
-  using Generate = FloatWeightGenerate<Weight>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t num_random_weights = kNumRandomWeights)
-      : Generate(allow_zero, num_random_weights) {}
-
-  Weight operator()() const { return Weight(Generate::operator()()); }
-};
-
-template <class T>
-class WeightGenerate<LogWeightTpl<T>>
-    : public FloatWeightGenerate<LogWeightTpl<T>> {
- public:
-  using Weight = LogWeightTpl<T>;
-  using Generate = FloatWeightGenerate<Weight>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t num_random_weights = kNumRandomWeights)
-      : Generate(allow_zero, num_random_weights) {}
-
-  Weight operator()() const { return Weight(Generate::operator()()); }
-};
-
-// This function object returns random integers chosen from [0,
-// num_random_weights). The boolean 'allow_zero' determines whether Zero() and
-// zero divisors should be returned in the random weight generation. This is
-// intended primary for testing.
-template <class T>
-class WeightGenerate<MinMaxWeightTpl<T>> {
- public:
-  using Weight = MinMaxWeightTpl<T>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t num_random_weights = kNumRandomWeights)
-      : allow_zero_(allow_zero), num_random_weights_(num_random_weights) {}
-
-  Weight operator()() const {
-    const int n = (rand() %  // NOLINT
-                   (2 * num_random_weights_ + allow_zero_)) -
-                  num_random_weights_;
-    if (allow_zero_ && n == num_random_weights_) {
-      return Weight::Zero();
-    } else if (n == -num_random_weights_) {
-      return Weight::One();
-    } else {
-      return Weight(n);
-    }
-  }
-
- private:
-  // Permits Zero() and zero divisors.
-  const bool allow_zero_;
-  // Number of alternative random weights.
-  const size_t num_random_weights_;
-};
-
-}  // namespace fst
-
-#endif  // FST_FLOAT_WEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/fst-decl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/fst-decl.h
deleted file mode 100644
index 0e18e155..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/fst-decl.h
+++ /dev/null
@@ -1,251 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// This file contains declarations of classes in the Fst template library.
-
-#ifndef FST_FST_DECL_H_
-#define FST_FST_DECL_H_
-
-#include <sys/types.h>
-#include <memory>  // for allocator<>
-
-#include <fst/types.h>
-
-namespace fst {
-
-// Symbol table and iterator.
-
-class SymbolTable;
-
-class SymbolTableIterator;
-
-// Weight templates and weights.
-
-template <class T>
-class FloatWeightTpl;
-
-template <class T>
-class TropicalWeightTpl;
-
-template <class T>
-class LogWeightTpl;
-
-template <class T>
-class MinMaxWeightTpl;
-
-using FloatWeight = FloatWeightTpl<float>;
-
-using TropicalWeight = TropicalWeightTpl<float>;
-
-using LogWeight = LogWeightTpl<float>;
-
-using MinMaxWeight = MinMaxWeightTpl<float>;
-
-// Arc templates and arcs.
-
-template <class Weight>
-struct ArcTpl;
-
-using StdArc = ArcTpl<TropicalWeight>;
-
-using LogArc = ArcTpl<LogWeight>;
-
-// Stores.
-
-template <class Element, class U>
-class DefaultCompactStore;
-
-template <class Arc>
-class DefaultCacheStore;
-
-// FST templates.
-
-template <class A, class ArcCompactor, class Unsigned = uint32,
-    class CompactStore = DefaultCompactStore<typename ArcCompactor::Element, Unsigned>,
-    class CacheStore = DefaultCacheStore<A>>
-class CompactFst;
-
-template <class Arc, class U = uint32>
-class ConstFst;
-
-template <class Arc, class Weight, class Matcher>
-class EditFst;
-
-template <class Arc>
-class ExpandedFst;
-
-template <class Arc>
-class Fst;
-
-template <class Arc>
-class MutableFst;
-
-template <class A, class Allocator = std::allocator<A>>
-class VectorState;
-
-template <class A, class State = VectorState<A>>
-class VectorFst;
-
-template <class Arc, class U = ssize_t>
-class DefaultReplaceStateTable;
-
-// On-the-fly operations.
-
-template <class Arc, class Compare>
-class ArcSortFst;
-
-template <class Arc>
-class ClosureFst;
-
-template <class A, class Store = DefaultCacheStore<A>>
-class ComposeFst;
-
-template <class Arc>
-class ConcatFst;
-
-template <class Arc>
-class DeterminizeFst;
-
-template <class Arc>
-class DifferenceFst;
-
-template <class Arc>
-class IntersectFst;
-
-template <class Arc>
-class InvertFst;
-
-template <class AArc, class BArc, class Mapper>
-class ArcMapFst;
-
-template <class Arc>
-class ProjectFst;
-
-template <class AArc, class BArc, class Selector>
-class RandGenFst;
-
-template <class Arc>
-class RelabelFst;
-
-template <class A, class StateTable = DefaultReplaceStateTable<A>,
-          class Store = DefaultCacheStore<A>>
-class ReplaceFst;
-
-template <class Arc>
-class RmEpsilonFst;
-
-template <class Arc>
-class UnionFst;
-
-// Heap.
-
-template <class T, class Compare>
-class Heap;
-
-// Compactors.
-
-template <class Arc>
-class AcceptorCompactor;
-
-template <class Arc>
-class StringCompactor;
-
-template <class Arc>
-class UnweightedAcceptorCompactor;
-
-template <class Arc>
-class UnweightedCompactor;
-
-template <class Arc>
-class WeightedStringCompactor;
-
-// Compact FSTs.
-
-template <class Arc, class U = uint32>
-using CompactStringFst = CompactFst<Arc, StringCompactor<Arc>, U>;
-
-template <class Arc, class U = uint32>
-using CompactWeightedStringFst =
-    CompactFst<Arc, WeightedStringCompactor<Arc>, U>;
-
-template <class Arc, class U = uint32>
-using CompactAcceptorFst = CompactFst<Arc, AcceptorCompactor<Arc>, U>;
-
-template <class Arc, class U = uint32>
-using CompactUnweightedFst = CompactFst<Arc, UnweightedCompactor<Arc>, U>;
-
-template <class Arc, class U = uint32>
-using CompactUnweightedAcceptorFst =
-    CompactFst<Arc, UnweightedAcceptorCompactor<Arc>, U>;
-
-// StdArc aliases for FSTs.
-
-using StdConstFst = ConstFst<StdArc>;
-using StdExpandedFst = ExpandedFst<StdArc>;
-using StdFst = Fst<StdArc>;
-using StdMutableFst = MutableFst<StdArc>;
-using StdVectorFst = VectorFst<StdArc>;
-
-// StdArc aliases for on-the-fly operations.
-
-template <class Compare>
-using StdArcSortFst = ArcSortFst<StdArc, Compare>;
-
-using StdClosureFst = ClosureFst<StdArc>;
-
-using StdComposeFst = ComposeFst<StdArc>;
-
-using StdConcatFst = ConcatFst<StdArc>;
-
-using StdDeterminizeFst = DeterminizeFst<StdArc>;
-
-using StdDifferenceFst = DifferenceFst<StdArc>;
-
-using StdIntersectFst = IntersectFst<StdArc>;
-
-using StdInvertFst = InvertFst<StdArc>;
-
-using StdProjectFst = ProjectFst<StdArc>;
-
-using StdRelabelFst = RelabelFst<StdArc>;
-
-using StdReplaceFst = ReplaceFst<StdArc>;
-
-using StdRmEpsilonFst = RmEpsilonFst<StdArc>;
-
-using StdUnionFst = UnionFst<StdArc>;
-
-// Filter states.
-
-template <class T>
-class IntegerFilterState;
-
-using CharFilterState = IntegerFilterState<signed char>;
-
-using ShortFilterState = IntegerFilterState<short>;  // NOLINT
-
-using IntFilterState = IntegerFilterState<int>;
-
-// Matchers and filters.
-
-template <class FST>
-class Matcher;
-
-template <class M1, class M2 = M1>
-class NullComposeFilter;
-
-template <class M1, class M2 = M1>
-class TrivialComposeFilter;
-
-template <class M1, class M2 = M1>
-class SequenceComposeFilter;
-
-template <class M1, class M2 = M1>
-class AltSequenceComposeFilter;
-
-template <class M1, class M2 = M1>
-class MatchComposeFilter;
-
-}  // namespace fst
-
-#endif  // FST_FST_DECL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/fst.h
deleted file mode 100644
index 1e8c5211..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/fst.h
+++ /dev/null
@@ -1,960 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// FST abstract base class definition, state and arc iterator interface, and
-// suggested base implementation.
-
-#ifndef FST_FST_H_
-#define FST_FST_H_
-
-#include <sys/types.h>
-
-#include <cmath>
-#include <cstddef>
-
-#include <iostream>
-#include <memory>
-#include <sstream>
-#include <string>
-#include <utility>
-
-#include <fst/compat.h>
-#include <fst/types.h>
-#include <fst/flags.h>
-#include <fst/log.h>
-#include <fstream>
-
-#include <fst/arc.h>
-#include <fst/memory.h>
-#include <fst/properties.h>
-#include <fst/register.h>
-#include <fst/symbol-table.h>
-#include <fst/util.h>
-
-
-DECLARE_bool(fst_align);
-
-namespace fst {
-
-bool IsFstHeader(std::istream &, const string &);
-
-class FstHeader;
-template <class Arc>
-
-struct StateIteratorData;
-template <class Arc>
-
-struct ArcIteratorData;
-
-template <class Arc>
-class MatcherBase;
-
-struct FstReadOptions {
-  // FileReadMode(s) are advisory, there are many conditions than prevent a
-  // file from being mapped, READ mode will be selected in these cases with
-  // a warning indicating why it was chosen.
-  enum FileReadMode { READ, MAP };
-
-  string source;                // Where you're reading from.
-  const FstHeader *header;      // Pointer to FST header; if non-zero, use
-                                // this info (don't read a stream header).
-  const SymbolTable *isymbols;  // Pointer to input symbols; if non-zero, use
-                                // this info (read and skip stream isymbols)
-  const SymbolTable *osymbols;  // Pointer to output symbols; if non-zero, use
-                                // this info (read and skip stream osymbols)
-  FileReadMode mode;            // Read or map files (advisory, if possible)
-  bool read_isymbols;           // Read isymbols, if any (default: true).
-  bool read_osymbols;           // Read osymbols, if any (default: true).
-
-  explicit FstReadOptions(const string &source = "<unspecified>",
-                          const FstHeader *header = nullptr,
-                          const SymbolTable *isymbols = nullptr,
-                          const SymbolTable *osymbols = nullptr);
-
-  explicit FstReadOptions(const string &source, const SymbolTable *isymbols,
-                          const SymbolTable *osymbols = nullptr);
-
-  // Helper function to convert strings FileReadModes into their enum value.
-  static FileReadMode ReadMode(const string &mode);
-
-  // Outputs a debug string for the FstReadOptions object.
-  string DebugString() const;
-};
-
-struct FstWriteOptions {
-  string source;        // Where you're writing to.
-  bool write_header;    // Write the header?
-  bool write_isymbols;  // Write input symbols?
-  bool write_osymbols;  // Write output symbols?
-  bool align;           // Write data aligned (may fail on pipes)?
-  bool stream_write;    // Avoid seek operations in writing.
-
-  explicit FstWriteOptions(const string &source = "<unspecifed>",
-                           bool write_header = true, bool write_isymbols = true,
-                           bool write_osymbols = true,
-                           bool align = FLAGS_fst_align,
-                           bool stream_write = false)
-      : source(source),
-        write_header(write_header),
-        write_isymbols(write_isymbols),
-        write_osymbols(write_osymbols),
-        align(align),
-        stream_write(stream_write) {}
-};
-
-// Header class.
-//
-// This is the recommended file header representation.
-
-class FstHeader {
- public:
-  enum {
-    HAS_ISYMBOLS = 0x1,  // Has input symbol table.
-    HAS_OSYMBOLS = 0x2,  // Has output symbol table.
-    IS_ALIGNED = 0x4,    // Memory-aligned (where appropriate).
-  } Flags;
-
-  FstHeader() : version_(0), flags_(0), properties_(0), start_(-1),
-      numstates_(0), numarcs_(0) {}
-
-  const string &FstType() const { return fsttype_; }
-
-  const string &ArcType() const { return arctype_; }
-
-  int32 Version() const { return version_; }
-
-  int32 GetFlags() const { return flags_; }
-
-  uint64 Properties() const { return properties_; }
-
-  int64 Start() const { return start_; }
-
-  int64 NumStates() const { return numstates_; }
-
-  int64 NumArcs() const { return numarcs_; }
-
-  void SetFstType(const string &type) { fsttype_ = type; }
-
-  void SetArcType(const string &type) { arctype_ = type; }
-
-  void SetVersion(int32 version) { version_ = version; }
-
-  void SetFlags(int32 flags) { flags_ = flags; }
-
-  void SetProperties(uint64 properties) { properties_ = properties; }
-
-  void SetStart(int64 start) { start_ = start; }
-
-  void SetNumStates(int64 numstates) { numstates_ = numstates; }
-
-  void SetNumArcs(int64 numarcs) { numarcs_ = numarcs; }
-
-  bool Read(std::istream &strm, const string &source,
-            bool rewind = false);
-
-  bool Write(std::ostream &strm, const string &source) const;
-
-  // Outputs a debug string for the FstHeader object.
-  string DebugString() const;
-
- private:
-  string fsttype_;     // E.g. "vector".
-  string arctype_;     // E.g. "standard".
-  int32 version_;      // Type version number.
-  int32 flags_;        // File format bits.
-  uint64 properties_;  // FST property bits.
-  int64 start_;        // Start state.
-  int64 numstates_;    // # of states.
-  int64 numarcs_;      // # of arcs.
-};
-
-// Specifies matcher action.
-enum MatchType {
-  MATCH_INPUT = 1,   // Match input label.
-  MATCH_OUTPUT = 2,  // Match output label.
-  MATCH_BOTH = 3,    // Match input or output label.
-  MATCH_NONE = 4,    // Match nothing.
-  MATCH_UNKNOWN = 5
-};  // Otherwise, match type unknown.
-
-constexpr int kNoStateId = -1;  // Not a valid state ID.
-constexpr int kNoLabel = -1;    // Not a valid label.
-
-// A generic FST, templated on the arc definition, with common-demoninator
-// methods (use StateIterator and ArcIterator to iterate over its states and
-// arcs).
-template <class A>
-class Fst {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  virtual ~Fst() {}
-
-  // Initial state.
-  virtual StateId Start() const = 0;
-
-  // State's final weight.
-  virtual Weight Final(StateId) const = 0;
-
-  // State's arc count.
-  virtual size_t NumArcs(StateId) const = 0;
-
-  // State's input epsilon count.
-  virtual size_t NumInputEpsilons(StateId) const = 0;
-
-  // State's output epsilon count.
-  virtual size_t NumOutputEpsilons(StateId) const = 0;
-
-  // Property bits. If test = false, return stored properties bits for mask
-  // (some possibly unknown); if test = true, return property bits for mask
-  // (computing o.w. unknown).
-  virtual uint64 Properties(uint64 mask, bool test) const = 0;
-
-  // FST type name.
-  virtual const string &Type() const = 0;
-
-  // Gets a copy of this Fst. The copying behaves as follows:
-  //
-  // (1) The copying is constant time if safe = false or if safe = true
-  // and is on an otherwise unaccessed FST.
-  //
-  // (2) If safe = true, the copy is thread-safe in that the original
-  // and copy can be safely accessed (but not necessarily mutated) by
-  // separate threads. For some FST types, 'Copy(true)' should only be
-  // called on an FST that has not otherwise been accessed. Behavior is
-  // otherwise undefined.
-  //
-  // (3) If a MutableFst is copied and then mutated, then the original is
-  // unmodified and vice versa (often by a copy-on-write on the initial
-  // mutation, which may not be constant time).
-  virtual Fst<Arc> *Copy(bool safe = false) const = 0;
-
-  // Reads an FST from an input stream; returns nullptr on error.
-  static Fst<Arc> *Read(std::istream &strm, const FstReadOptions &opts) {
-    FstReadOptions ropts(opts);
-    FstHeader hdr;
-    if (ropts.header) {
-      hdr = *opts.header;
-    } else {
-      if (!hdr.Read(strm, opts.source)) return nullptr;
-      ropts.header = &hdr;
-    }
-    const auto &fst_type = hdr.FstType();
-    const auto reader = FstRegister<Arc>::GetRegister()->GetReader(fst_type);
-    if (!reader) {
-      LOG(ERROR) << "Fst::Read: Unknown FST type " << fst_type
-                 << " (arc type = " << Arc::Type() << "): " << ropts.source;
-      return nullptr;
-    }
-    return reader(strm, ropts);
-  }
-
-  // Reads an FST from a file; returns nullptr on error. An empty filename
-  // results in reading from standard input.
-  static Fst<Arc> *Read(const string &filename) {
-    if (!filename.empty()) {
-      std::ifstream strm(filename,
-                              std::ios_base::in | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "Fst::Read: Can't open file: " << filename;
-        return nullptr;
-      }
-      return Read(strm, FstReadOptions(filename));
-    } else {
-      return Read(std::cin, FstReadOptions("standard input"));
-    }
-  }
-
-  // Writes an FST to an output stream; returns false on error.
-  virtual bool Write(std::ostream &strm, const FstWriteOptions &opts) const {
-    LOG(ERROR) << "Fst::Write: No write stream method for " << Type()
-               << " FST type";
-    return false;
-  }
-
-  // Writes an FST to a file; returns false on error; an empty filename
-  // results in writing to standard output.
-  virtual bool Write(const string &filename) const {
-    LOG(ERROR) << "Fst::Write: No write filename method for " << Type()
-               << " FST type";
-    return false;
-  }
-
-  // Returns input label symbol table; return nullptr if not specified.
-  virtual const SymbolTable *InputSymbols() const = 0;
-
-  // Return output label symbol table; return nullptr if not specified.
-  virtual const SymbolTable *OutputSymbols() const = 0;
-
-  // For generic state iterator construction (not normally called directly by
-  // users).
-  virtual void InitStateIterator(StateIteratorData<Arc> *data) const = 0;
-
-  // For generic arc iterator construction (not normally called directly by
-  // users).
-  virtual void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const = 0;
-
-  // For generic matcher construction (not normally called directly by users).
-  virtual MatcherBase<Arc> *InitMatcher(MatchType match_type) const;
-
- protected:
-  bool WriteFile(const string &filename) const {
-    if (!filename.empty()) {
-      std::ofstream strm(filename,
-                               std::ios_base::out | std::ios_base::binary);
-      if (!strm) {
-        LOG(ERROR) << "Fst::Write: Can't open file: " << filename;
-        return false;
-      }
-      bool val = Write(strm, FstWriteOptions(filename));
-      if (!val) LOG(ERROR) << "Fst::Write failed: " << filename;
-      return val;
-    } else {
-      return Write(std::cout, FstWriteOptions("standard output"));
-    }
-  }
-};
-
-// A useful alias when using StdArc.
-using StdFst = Fst<StdArc>;
-
-// State and arc iterator definitions.
-//
-// State iterator interface templated on the Arc definition; used for
-// StateIterator specializations returned by the InitStateIterator FST method.
-template <class Arc>
-class StateIteratorBase {
- public:
-  using StateId = typename Arc::StateId;
-
-  virtual ~StateIteratorBase() {}
-
-  // End of iterator?
-  virtual bool Done() const = 0;
-  // Returns current state (when !Done()).
-  virtual StateId Value() const = 0;
-  // Advances to next state (when !Done()).
-  virtual void Next() = 0;
-  // Resets to initial condition.
-  virtual void Reset() = 0;
-};
-
-// StateIterator initialization data.
-
-template <class Arc>
-struct StateIteratorData {
-  using StateId = typename Arc::StateId;
-
-  // Specialized iterator if non-zero.
-  StateIteratorBase<Arc> *base;
-  // Otherwise, the total number of states.
-  StateId nstates;
-
-  StateIteratorData() : base(nullptr), nstates(0) {}
-
-  StateIteratorData(const StateIteratorData &) = delete;
-  StateIteratorData &operator=(const StateIteratorData &) = delete;
-};
-
-// Generic state iterator, templated on the FST definition (a wrapper
-// around a pointer to a specific one). Here is a typical use:
-//
-//   for (StateIterator<StdFst> siter(fst);
-//        !siter.Done();
-//        siter.Next()) {
-//     StateId s = siter.Value();
-//     ...
-//   }
-template <class FST>
-class StateIterator {
- public:
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-
-  explicit StateIterator(const FST &fst) : s_(0) {
-    fst.InitStateIterator(&data_);
-  }
-
-  ~StateIterator() { delete data_.base; }
-
-  bool Done() const {
-    return data_.base ? data_.base->Done() : s_ >= data_.nstates;
-  }
-
-  StateId Value() const { return data_.base ? data_.base->Value() : s_; }
-
-  void Next() {
-    if (data_.base) {
-      data_.base->Next();
-    } else {
-      ++s_;
-    }
-  }
-
-  void Reset() {
-    if (data_.base) {
-      data_.base->Reset();
-    } else {
-      s_ = 0;
-    }
-  }
-
- private:
-  StateIteratorData<Arc> data_;
-  StateId s_;
-};
-
-// Flags to control the behavior on an arc iterator.
-static constexpr uint32 kArcILabelValue =
-    0x0001;  // Value() gives valid ilabel.
-static constexpr uint32 kArcOLabelValue = 0x0002;  //  "       "     " olabel.
-static constexpr uint32 kArcWeightValue = 0x0004;  //  "       "     " weight.
-static constexpr uint32 kArcNextStateValue =
-    0x0008;                                    //  "       "     " nextstate.
-static constexpr uint32 kArcNoCache = 0x0010;  // No need to cache arcs.
-
-static constexpr uint32 kArcValueFlags =
-    kArcILabelValue | kArcOLabelValue | kArcWeightValue | kArcNextStateValue;
-
-static constexpr uint32 kArcFlags = kArcValueFlags | kArcNoCache;
-
-// Arc iterator interface, templated on the arc definition; used for arc
-// iterator specializations that are returned by the InitArcIterator FST method.
-template <class Arc>
-class ArcIteratorBase {
- public:
-  using StateId = typename Arc::StateId;
-
-  virtual ~ArcIteratorBase() {}
-
-  // End of iterator?
-  virtual bool Done() const = 0;
-  // Returns current arc (when !Done()).
-  virtual const Arc &Value() const = 0;
-  // Advances to next arc (when !Done()).
-  virtual void Next() = 0;
-  // Returns current position.
-  virtual size_t Position() const = 0;
-  // Returns to initial condition.
-  virtual void Reset() = 0;
-  // Advances to arbitrary arc by position.
-  virtual void Seek(size_t) = 0;
-  // Returns current behavorial flags
-  virtual uint32 Flags() const = 0;
-  // Sets behavorial flags.
-  virtual void SetFlags(uint32, uint32) = 0;
-};
-
-// ArcIterator initialization data.
-template <class Arc>
-struct ArcIteratorData {
-  ArcIteratorData()
-      : base(nullptr), arcs(nullptr), narcs(0), ref_count(nullptr) {}
-
-  ArcIteratorData(const ArcIteratorData &) = delete;
-
-  ArcIteratorData &operator=(const ArcIteratorData &) = delete;
-
-  ArcIteratorBase<Arc> *base;  // Specialized iterator if non-zero.
-  const Arc *arcs;             // O.w. arcs pointer
-  size_t narcs;                // ... and arc count.
-  int *ref_count;              // ... and reference count if non-zero.
-};
-
-// Generic arc iterator, templated on the FST definition (a wrapper around a
-// pointer to a specific one). Here is a typical use:
-//
-//   for (ArcIterator<StdFst> aiter(fst, s);
-//        !aiter.Done();
-//         aiter.Next()) {
-//     StdArc &arc = aiter.Value();
-//     ...
-//   }
-template <class FST>
-class ArcIterator {
- public:
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const FST &fst, StateId s) : i_(0) {
-    fst.InitArcIterator(s, &data_);
-  }
-
-  explicit ArcIterator(const ArcIteratorData<Arc> &data) : data_(data), i_(0) {
-    if (data_.ref_count) ++(*data_.ref_count);
-  }
-
-  ~ArcIterator() {
-    if (data_.base) {
-      delete data_.base;
-    } else if (data_.ref_count) {
-      --(*data_.ref_count);
-    }
-  }
-
-  bool Done() const {
-    return data_.base ? data_.base->Done() : i_ >= data_.narcs;
-  }
-
-  const Arc &Value() const {
-    return data_.base ? data_.base->Value() : data_.arcs[i_];
-  }
-
-  void Next() {
-    if (data_.base) {
-      data_.base->Next();
-    } else {
-      ++i_;
-    }
-  }
-
-  void Reset() {
-    if (data_.base) {
-      data_.base->Reset();
-    } else {
-      i_ = 0;
-    }
-  }
-
-  void Seek(size_t a) {
-    if (data_.base) {
-      data_.base->Seek(a);
-    } else {
-      i_ = a;
-    }
-  }
-
-  size_t Position() const { return data_.base ? data_.base->Position() : i_; }
-
-  uint32 Flags() const {
-    if (data_.base) {
-      return data_.base->Flags();
-    } else {
-      return kArcValueFlags;
-    }
-  }
-
-  void SetFlags(uint32 flags, uint32 mask) {
-    if (data_.base) data_.base->SetFlags(flags, mask);
-  }
-
- private:
-  ArcIteratorData<Arc> data_;
-  size_t i_;
-};
-
-}  // namespace fst
-
-// ArcIterator placement operator new and destroy function; new needs to be in
-// the global namespace.
-
-template <class FST>
-void *operator new(size_t size,
-                   fst::MemoryPool<fst::ArcIterator<FST>> *pool) {
-  return pool->Allocate();
-}
-
-namespace fst {
-
-template <class FST>
-void Destroy(ArcIterator<FST> *aiter, MemoryPool<ArcIterator<FST>> *pool) {
-  if (aiter) {
-    aiter->~ArcIterator<FST>();
-    pool->Free(aiter);
-  }
-}
-
-// Matcher definitions.
-
-template <class Arc>
-MatcherBase<Arc> *Fst<Arc>::InitMatcher(MatchType match_type) const {
-  return nullptr;  // One should just use the default matcher.
-}
-
-// FST accessors, useful in high-performance applications.
-
-namespace internal {
-
-// General case, requires non-abstract, 'final' methods. Use for inlining.
-
-template <class F>
-inline typename F::Arc::Weight Final(const F &fst, typename F::Arc::StateId s) {
-  return fst.F::Final(s);
-}
-
-template <class F>
-inline ssize_t NumArcs(const F &fst, typename F::Arc::StateId s) {
-  return fst.F::NumArcs(s);
-}
-
-template <class F>
-inline ssize_t NumInputEpsilons(const F &fst, typename F::Arc::StateId s) {
-  return fst.F::NumInputEpsilons(s);
-}
-
-template <class F>
-inline ssize_t NumOutputEpsilons(const F &fst, typename F::Arc::StateId s) {
-  return fst.F::NumOutputEpsilons(s);
-}
-
-// Fst<Arc> case, abstract methods.
-
-template <class Arc>
-inline typename Arc::Weight Final(const Fst<Arc> &fst,
-                                  typename Arc::StateId s) {
-  return fst.Final(s);
-}
-
-template <class Arc>
-inline size_t NumArcs(const Fst<Arc> &fst, typename Arc::StateId s) {
-  return fst.NumArcs(s);
-}
-
-template <class Arc>
-inline size_t NumInputEpsilons(const Fst<Arc> &fst, typename Arc::StateId s) {
-  return fst.NumInputEpsilons(s);
-}
-
-template <class Arc>
-inline size_t NumOutputEpsilons(const Fst<Arc> &fst, typename Arc::StateId s) {
-  return fst.NumOutputEpsilons(s);
-}
-
-// FST implementation base.
-//
-// This is the recommended FST implementation base class. It will handle
-// reference counts, property bits, type information and symbols.
-//
-// Users are discouraged, but not prohibited, from subclassing this outside the
-// FST library.
-template <class Arc>
-class FstImpl {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  FstImpl() : properties_(0), type_("null") {}
-
-  FstImpl(const FstImpl<Arc> &impl)
-      : properties_(impl.properties_),
-        type_(impl.type_),
-        isymbols_(impl.isymbols_ ? impl.isymbols_->Copy() : nullptr),
-        osymbols_(impl.osymbols_ ? impl.osymbols_->Copy() : nullptr) {}
-
-  virtual ~FstImpl() {}
-
-  const string &Type() const { return type_; }
-
-  void SetType(const string &type) { type_ = type; }
-
-  virtual uint64 Properties() const { return properties_; }
-
-  virtual uint64 Properties(uint64 mask) const { return properties_ & mask; }
-
-  void SetProperties(uint64 props) {
-    properties_ &= kError;  // kError can't be cleared.
-    properties_ |= props;
-  }
-
-  void SetProperties(uint64 props, uint64 mask) {
-    properties_ &= ~mask | kError;  // kError can't be cleared.
-    properties_ |= props & mask;
-  }
-
-  // Allows (only) setting error bit on const FST implementations.
-  void SetProperties(uint64 props, uint64 mask) const {
-    if (mask != kError) {
-      FSTERROR() << "FstImpl::SetProperties() const: Can only set kError";
-    }
-    properties_ |= kError;
-  }
-
-  const SymbolTable *InputSymbols() const { return isymbols_.get(); }
-
-  const SymbolTable *OutputSymbols() const { return osymbols_.get(); }
-
-  SymbolTable *InputSymbols() { return isymbols_.get(); }
-
-  SymbolTable *OutputSymbols() { return osymbols_.get(); }
-
-  void SetInputSymbols(const SymbolTable *isyms) {
-    isymbols_.reset(isyms ? isyms->Copy() : nullptr);
-  }
-
-  void SetOutputSymbols(const SymbolTable *osyms) {
-    osymbols_.reset(osyms ? osyms->Copy() : nullptr);
-  }
-
-  // Reads header and symbols from input stream, initializes FST, and returns
-  // the header. If opts.header is non-null, skips reading and uses the option
-  // value instead. If opts.[io]symbols is non-null, reads in (if present), but
-  // uses the option value.
-  bool ReadHeader(std::istream &strm, const FstReadOptions &opts,
-                  int min_version, FstHeader *hdr);
-
-  // Writes header and symbols to output stream. If opts.header is false, skips
-  // writing header. If opts.[io]symbols is false, skips writing those symbols.
-  // This method is needed for implementations that implement Write methods.
-  void WriteHeader(std::ostream &strm, const FstWriteOptions &opts,
-                   int version, FstHeader *hdr) const {
-    if (opts.write_header) {
-      hdr->SetFstType(type_);
-      hdr->SetArcType(Arc::Type());
-      hdr->SetVersion(version);
-      hdr->SetProperties(properties_);
-      int32 file_flags = 0;
-      if (isymbols_ && opts.write_isymbols) {
-        file_flags |= FstHeader::HAS_ISYMBOLS;
-      }
-      if (osymbols_ && opts.write_osymbols) {
-        file_flags |= FstHeader::HAS_OSYMBOLS;
-      }
-      if (opts.align) file_flags |= FstHeader::IS_ALIGNED;
-      hdr->SetFlags(file_flags);
-      hdr->Write(strm, opts.source);
-    }
-    if (isymbols_ && opts.write_isymbols) isymbols_->Write(strm);
-    if (osymbols_ && opts.write_osymbols) osymbols_->Write(strm);
-  }
-
-  // Writes out header and symbols to output stream. If opts.header is false,
-  // skips writing header. If opts.[io]symbols is false, skips writing those
-  // symbols. `type` is the FST type being written. This method is used in the
-  // cross-type serialization methods Fst::WriteFst.
-  static void WriteFstHeader(const Fst<Arc> &fst, std::ostream &strm,
-                             const FstWriteOptions &opts, int version,
-                             const string &type, uint64 properties,
-                             FstHeader *hdr) {
-    if (opts.write_header) {
-      hdr->SetFstType(type);
-      hdr->SetArcType(Arc::Type());
-      hdr->SetVersion(version);
-      hdr->SetProperties(properties);
-      int32 file_flags = 0;
-      if (fst.InputSymbols() && opts.write_isymbols) {
-        file_flags |= FstHeader::HAS_ISYMBOLS;
-      }
-      if (fst.OutputSymbols() && opts.write_osymbols) {
-        file_flags |= FstHeader::HAS_OSYMBOLS;
-      }
-      if (opts.align) file_flags |= FstHeader::IS_ALIGNED;
-      hdr->SetFlags(file_flags);
-      hdr->Write(strm, opts.source);
-    }
-    if (fst.InputSymbols() && opts.write_isymbols) {
-      fst.InputSymbols()->Write(strm);
-    }
-    if (fst.OutputSymbols() && opts.write_osymbols) {
-      fst.OutputSymbols()->Write(strm);
-    }
-  }
-
-  // In serialization routines where the header cannot be written until after
-  // the machine has been serialized, this routine can be called to seek to the
-  // beginning of the file an rewrite the header with updated fields. It
-  // repositions the file pointer back at the end of the file. Returns true on
-  // success, false on failure.
-  static bool UpdateFstHeader(const Fst<Arc> &fst, std::ostream &strm,
-                              const FstWriteOptions &opts, int version,
-                              const string &type, uint64 properties,
-                              FstHeader *hdr, size_t header_offset) {
-    strm.seekp(header_offset);
-    if (!strm) {
-      LOG(ERROR) << "Fst::UpdateFstHeader: Write failed: " << opts.source;
-      return false;
-    }
-    WriteFstHeader(fst, strm, opts, version, type, properties, hdr);
-    if (!strm) {
-      LOG(ERROR) << "Fst::UpdateFstHeader: Write failed: " << opts.source;
-      return false;
-    }
-    strm.seekp(0, std::ios_base::end);
-    if (!strm) {
-      LOG(ERROR) << "Fst::UpdateFstHeader: Write failed: " << opts.source;
-      return false;
-    }
-    return true;
-  }
-
- protected:
-  mutable uint64 properties_;  // Property bits.
-
- private:
-  string type_;  // Unique name of FST class.
-  std::unique_ptr<SymbolTable> isymbols_;
-  std::unique_ptr<SymbolTable> osymbols_;
-};
-
-template <class Arc>
-bool FstImpl<Arc>::ReadHeader(std::istream &strm, const FstReadOptions &opts,
-                              int min_version, FstHeader *hdr) {
-  if (opts.header) {
-    *hdr = *opts.header;
-  } else if (!hdr->Read(strm, opts.source)) {
-    return false;
-  }
-  if (FLAGS_v >= 2) {
-    LOG(INFO) << "FstImpl::ReadHeader: source: " << opts.source
-              << ", fst_type: " << hdr->FstType()
-              << ", arc_type: " << Arc::Type()
-              << ", version: " << hdr->Version()
-              << ", flags: " << hdr->GetFlags();
-  }
-  if (hdr->FstType() != type_) {
-    LOG(ERROR) << "FstImpl::ReadHeader: FST not of type " << type_
-               << ": " << opts.source;
-    return false;
-  }
-  if (hdr->ArcType() != Arc::Type()) {
-    LOG(ERROR) << "FstImpl::ReadHeader: Arc not of type " << Arc::Type()
-               << ": " << opts.source;
-    return false;
-  }
-  if (hdr->Version() < min_version) {
-    LOG(ERROR) << "FstImpl::ReadHeader: Obsolete " << type_
-               << " FST version: " << opts.source;
-    return false;
-  }
-  properties_ = hdr->Properties();
-  if (hdr->GetFlags() & FstHeader::HAS_ISYMBOLS) {
-    isymbols_.reset(SymbolTable::Read(strm, opts.source));
-  }
-  // Deletes input symbol table.
-  if (!opts.read_isymbols) SetInputSymbols(nullptr);
-  if (hdr->GetFlags() & FstHeader::HAS_OSYMBOLS) {
-    osymbols_.reset(SymbolTable::Read(strm, opts.source));
-  }
-  // Deletes output symbol table.
-  if (!opts.read_osymbols) SetOutputSymbols(nullptr);
-  if (opts.isymbols) {
-    isymbols_.reset(opts.isymbols->Copy());
-  }
-  if (opts.osymbols) {
-    osymbols_.reset(opts.osymbols->Copy());
-  }
-  return true;
-}
-
-}  // namespace internal
-
-template <class Arc>
-uint64 TestProperties(const Fst<Arc> &fst, uint64 mask, uint64 *known);
-
-// This is a helper class template useful for attaching an FST interface to
-// its implementation, handling reference counting.
-template <class Impl, class FST = Fst<typename Impl::Arc>>
-class ImplToFst : public FST {
- public:
-  using Arc = typename Impl::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using FST::operator=;
-
-  StateId Start() const override { return impl_->Start(); }
-
-  Weight Final(StateId s) const override { return impl_->Final(s); }
-
-  size_t NumArcs(StateId s) const override { return impl_->NumArcs(s); }
-
-  size_t NumInputEpsilons(StateId s) const override {
-    return impl_->NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) const override {
-    return impl_->NumOutputEpsilons(s);
-  }
-
-  uint64 Properties(uint64 mask, bool test) const override {
-    if (test) {
-      uint64 knownprops, testprops = TestProperties(*this, mask, &knownprops);
-      impl_->SetProperties(testprops, knownprops);
-      return testprops & mask;
-    } else {
-      return impl_->Properties(mask);
-    }
-  }
-
-  const string &Type() const override { return impl_->Type(); }
-
-  const SymbolTable *InputSymbols() const override {
-    return impl_->InputSymbols();
-  }
-
-  const SymbolTable *OutputSymbols() const override {
-    return impl_->OutputSymbols();
-  }
-
- protected:
-  explicit ImplToFst(std::shared_ptr<Impl> impl) : impl_(std::move(impl)) {}
-
-  // This constructor presumes there is a copy constructor for the
-  // implementation.
-  ImplToFst(const ImplToFst<Impl, FST> &fst, bool safe) {
-    if (safe) {
-      impl_ = std::make_shared<Impl>(*(fst.impl_));
-    } else {
-      impl_ = fst.impl_;
-    }
-  }
-
-  // Returns raw pointers to the shared object.
-  const Impl *GetImpl() const { return impl_.get(); }
-
-  Impl *GetMutableImpl() const { return impl_.get(); }
-
-  // Returns a ref-counted smart poiner to the implementation.
-  std::shared_ptr<Impl> GetSharedImpl() const { return impl_; }
-
-  bool Unique() const { return impl_.unique(); }
-
-  void SetImpl(std::shared_ptr<Impl> impl) { impl_ = impl; }
-
- private:
-  template <class IFST, class OFST>
-  friend void Cast(const IFST &ifst, OFST *ofst);
-
-  std::shared_ptr<Impl> impl_;
-};
-
-// Converts FSTs by casting their implementations, where this makes sense
-// (which excludes implementations with weight-dependent virtual methods).
-// Must be a friend of the FST classes involved (currently the concrete FSTs:
-// ConstFst, CompactFst, and VectorFst). This can only be safely used for arc
-// types that have identical storage characteristics. As with an FST
-// copy constructor and Copy() method, this is a constant time operation
-// (but subject to copy-on-write if it is a MutableFst and modified).
-template <class IFST, class OFST>
-void Cast(const IFST &ifst, OFST *ofst) {
-  using OImpl = typename OFST::Impl;
-  ofst->impl_ = std::shared_ptr<OImpl>(ifst.impl_,
-      reinterpret_cast<OImpl *>(ifst.impl_.get()));
-}
-
-// FST serialization.
-
-template <class Arc>
-void FstToString(const Fst<Arc> &fst, string *result) {
-  std::ostringstream ostrm;
-  fst.Write(ostrm, FstWriteOptions("FstToString"));
-  *result = ostrm.str();
-}
-
-template <class Arc>
-void FstToString(const Fst<Arc> &fst, string *result,
-                 const FstWriteOptions &options) {
-  std::ostringstream ostrm;
-  fst.Write(ostrm, options);
-  *result = ostrm.str();
-}
-
-template <class Arc>
-Fst<Arc> *StringToFst(const string &s) {
-  std::istringstream istrm(s);
-  return Fst<Arc>::Read(istrm, FstReadOptions("StringToFst"));
-}
-
-}  // namespace fst
-
-#endif  // FST_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/fstlib.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/fstlib.h
deleted file mode 100644
index a99cff47..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/fstlib.h
+++ /dev/null
@@ -1,130 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// This is a library for constructing, combining, optimizing, and searching
-// "weighted finite-state transducers" (FSTs). Weighted finite-state transducers
-// are automata where each transition has an input label, an output label, and a
-// weight. The more familiar finite-state acceptor is represented as a
-// transducer with each transition's input and output the same. Finite-state
-// acceptors are used to represent sets of strings (specifically, "regular" or
-// "rational sets"); finite-state transducers are used to represent binary
-// relations between pairs of strings (specifically, "rational transductions").
-// The weights can be used to represent the cost of taking a particular
-// transition.
-//
-// In this library, transducers are templated on the Arc (transition)
-// definition, which allows changing the label, weight, and state ID sets.
-// Labels and state IDs are restricted to signed integral types but the weight
-// can be an arbitrary type whose members satisfy certain algebraic ("semiring")
-// properties.
-//
-// This convenience file includes all other FST header files.
-
-#ifndef FST_FSTLIB_H_
-#define FST_FSTLIB_H_
-
-
-// Abstract FST classes.
-#include <fst/expanded-fst.h>
-#include <fst/fst.h>
-#include <fst/mutable-fst.h>
-
-// Concrete FST classes.
-#include <fst/compact-fst.h>
-#include <fst/const-fst.h>
-#include <fst/edit-fst.h>
-#include <fst/vector-fst.h>
-
-// FST algorithms and delayed FST classes.
-#include <fst/arc-map.h>
-#include <fst/arcsort.h>
-#include <fst/closure.h>
-#include <fst/compose.h>
-#include <fst/concat.h>
-#include <fst/connect.h>
-#include <fst/determinize.h>
-#include <fst/difference.h>
-#include <fst/disambiguate.h>
-#include <fst/encode.h>
-#include <fst/epsnormalize.h>
-#include <fst/equal.h>
-#include <fst/equivalent.h>
-#include <fst/factor-weight.h>
-#include <fst/intersect.h>
-#include <fst/invert.h>
-#include <fst/isomorphic.h>
-#include <fst/map.h>
-#include <fst/minimize.h>
-#include <fst/project.h>
-#include <fst/prune.h>
-#include <fst/push.h>
-#include <fst/randequivalent.h>
-#include <fst/randgen.h>
-#include <fst/rational.h>
-#include <fst/relabel.h>
-#include <fst/replace.h>
-#include <fst/replace-util.h>
-#include <fst/reverse.h>
-#include <fst/reweight.h>
-#include <fst/rmepsilon.h>
-#include <fst/rmfinalepsilon.h>
-#include <fst/shortest-distance.h>
-#include <fst/shortest-path.h>
-#include <fst/state-map.h>
-#include <fst/statesort.h>
-#include <fst/synchronize.h>
-#include <fst/topsort.h>
-#include <fst/union.h>
-#include <fst/verify.h>
-#include <fst/visit.h>
-
-// Weights.
-#include <fst/expectation-weight.h>
-#include <fst/float-weight.h>
-#include <fst/lexicographic-weight.h>
-#include <fst/pair-weight.h>
-#include <fst/power-weight.h>
-#include <fst/product-weight.h>
-#include <fst/signed-log-weight.h>
-#include <fst/sparse-power-weight.h>
-#include <fst/sparse-tuple-weight.h>
-#include <fst/string-weight.h>
-#include <fst/tuple-weight.h>
-#include <fst/weight.h>
-
-// Auxiliary classes for composition.
-#include <fst/compose-filter.h>
-#include <fst/lookahead-filter.h>
-#include <fst/lookahead-matcher.h>
-#include <fst/matcher-fst.h>
-#include <fst/matcher.h>
-#include <fst/state-table.h>
-
-// Data structures.
-#include <fst/heap.h>
-#include <fst/interval-set.h>
-#include <fst/queue.h>
-#include <fst/union-find.h>
-
-// Miscellaneous.
-#include <fst/accumulator.h>
-#include <fst/add-on.h>
-#include <fst/arc.h>
-#include <fst/arcfilter.h>
-#include <fst/cache.h>
-#include <fst/complement.h>
-#include <fst/dfs-visit.h>
-#include <fst/generic-register.h>
-#include <fst/label-reachable.h>
-#include <fst/partition.h>
-#include <fst/properties.h>
-#include <fst/register.h>
-#include <fst/state-reachable.h>
-#include <fst/string.h>
-#include <fst/symbol-table.h>
-#include <fst/symbol-table-ops.h>
-#include <fst/test-properties.h>
-#include <fst/util.h>
-
-
-#endif  // FST_FSTLIB_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/generic-register.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/generic-register.h
deleted file mode 100644
index ad243396..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/generic-register.h
+++ /dev/null
@@ -1,128 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_GENERIC_REGISTER_H_
-#define FST_GENERIC_REGISTER_H_
-
-#ifndef FST_NO_DYNAMIC_LINKING
-#ifndef _MSC_VER
-#include <dlfcn.h>
-#endif // !_MSC_VER
-#include <fst/compat.h>
-#endif
-#include <map>
-#include <string>
-
-#include <fst/types.h>
-#include <fst/log.h>
-
-// Generic class representing a globally-stored correspondence between
-// objects of KeyType and EntryType.
-//
-// KeyType must:
-//
-// * be such as can be stored as a key in a std::map<>.
-// * be concatenable with a const char* with the + operator
-//   (or you must subclass and redefine LoadEntryFromSharedObject)
-//
-// EntryType must be default constructible.
-//
-// The third template parameter should be the type of a subclass of this class
-// (think CRTP). This is to allow GetRegister() to instantiate and return an
-// object of the appropriate type.
-
-namespace fst {
-
-template <class KeyType, class EntryType, class RegisterType>
-class GenericRegister {
- public:
-  using Key = KeyType;
-  using Entry = EntryType;
-
-  static RegisterType *GetRegister() {
-    static auto reg = new RegisterType;
-    return reg;
-  }
-
-  void SetEntry(const KeyType &key, const EntryType &entry) {
-    MutexLock l(&register_lock_);
-    register_table_.insert(std::make_pair(key, entry));
-  }
-
-  EntryType GetEntry(const KeyType &key) const {
-    const auto *entry = LookupEntry(key);
-    if (entry) {
-      return *entry;
-    } else {
-      return LoadEntryFromSharedObject(key);
-    }
-  }
-
-  virtual ~GenericRegister() {}
-
- protected:
-  // Override this if you want to be able to load missing definitions from
-  // shared object files.
-  virtual EntryType LoadEntryFromSharedObject(const KeyType &key) const {
-#ifdef FST_NO_DYNAMIC_LINKING
-    return EntryType();
-#else
-    const auto so_filename = ConvertKeyToSoFilename(key);
-    void *handle = dlopen(so_filename.c_str(), RTLD_LAZY);
-    if (handle == nullptr) {
-      LOG(ERROR) << "GenericRegister::GetEntry: " << dlerror();
-      return EntryType();
-    }
-#ifdef RUN_MODULE_INITIALIZERS
-    RUN_MODULE_INITIALIZERS();
-#endif
-    // We assume that the DSO constructs a static object in its global scope
-    // that does the registration. Thus we need only load it, not call any
-    // methods.
-    const auto *entry = this->LookupEntry(key);
-    if (entry == nullptr) {
-      LOG(ERROR) << "GenericRegister::GetEntry: "
-                 << "lookup failed in shared object: " << so_filename;
-      return EntryType();
-    }
-    return *entry;
-#endif  // FST_NO_DYNAMIC_LINKING
-  }
-
-  // Override this to define how to turn a key into an SO filename.
-  virtual string ConvertKeyToSoFilename(const KeyType &key) const = 0;
-
-  virtual const EntryType *LookupEntry(const KeyType &key) const {
-    MutexLock l(&register_lock_);
-    const auto it = register_table_.find(key);
-    if (it != register_table_.end()) {
-      return &it->second;
-    } else {
-      return nullptr;
-    }
-  }
-
- private:
-  mutable Mutex register_lock_;
-  std::map<KeyType, EntryType> register_table_;
-};
-
-// Generic register-er class capable of creating new register entries in the
-// given RegisterType template parameter. This type must define types Key and
-// Entry, and have appropriate static GetRegister() and instance SetEntry()
-// functions. An easy way to accomplish this is to have RegisterType be the
-// type of a subclass of GenericRegister.
-template <class RegisterType>
-class GenericRegisterer {
- public:
-  using Key = typename RegisterType::Key;
-  using Entry = typename RegisterType::Entry;
-
-  GenericRegisterer(Key key, Entry entry) {
-    RegisterType::GetRegister()->SetEntry(key, entry);
-  }
-};
-
-}  // namespace fst
-
-#endif  // FST_GENERIC_REGISTER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/heap.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/heap.h
deleted file mode 100644
index 2a4e6356..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/heap.h
+++ /dev/null
@@ -1,166 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Implementation of a heap as in STL, but allows tracking positions in heap
-// using a key. The key can be used to do an in-place update of values in the
-// heap.
-
-#ifndef FST_HEAP_H_
-#define FST_HEAP_H_
-
-#include <utility>
-#include <vector>
-
-#include <fst/compat.h>
-namespace fst {
-
-// A templated heap implementation that supports in-place update of values.
-//
-// The templated heap implementation is a little different from the STL
-// priority_queue and the *_heap operations in STL. This heap supports
-// indexing of values in the heap via an associated key.
-//
-// Each value is internally associated with a key which is returned to the
-// calling functions on heap insert. This key can be used to later update
-// the specific value in the heap.
-//
-// T: the element type of the hash. It can be POD, Data or a pointer to Data.
-// Compare: comparison functor for determining min-heapness.
-template <class T, class Compare>
-class Heap {
- public:
-  using Value = T;
-
-  static constexpr int kNoKey = -1;
-
-  // Initializes with a specific comparator.
-  explicit Heap(Compare comp = Compare()) : comp_(comp), size_(0) {}
-
-  // Inserts a value into the heap.
-  int Insert(const Value &value) {
-    if (size_ < values_.size()) {
-      values_[size_] = value;
-      pos_[key_[size_]] = size_;
-    } else {
-      values_.push_back(value);
-      pos_.push_back(size_);
-      key_.push_back(size_);
-    }
-    ++size_;
-    return Insert(value, size_ - 1);
-  }
-
-  // Updates a value at position given by the key. The pos_ array is first
-  // indexed by the key. The position gives the position in the heap array.
-  // Once we have the position we can then use the standard heap operations
-  // to calculate the parent and child positions.
-  void Update(int key, const Value &value) {
-    const auto i = pos_[key];
-    const bool is_better = comp_(value, values_[Parent(i)]);
-    values_[i] = value;
-    if (is_better) {
-      Insert(value, i);
-    } else {
-      Heapify(i);
-    }
-  }
-
-  // Returns the least value.
-  Value Pop() {
-    Value top = values_.front();
-    Swap(0, size_-1);
-    size_--;
-    Heapify(0);
-    return top;
-  }
-
-  // Returns the least value w.r.t.  the comparison function from the
-  // heap.
-  const Value &Top() const { return values_.front(); }
-
-  // Returns the element for the given key.
-  const Value &Get(int key) const { return values_[pos_[key]]; }
-
-  // Checks if the heap is empty.
-  bool Empty() const { return size_ == 0; }
-
-  void Clear() { size_ = 0; }
-
-  int Size() const { return size_; }
-
-  void Reserve(int size) {
-    values_.reserve(size);
-    pos_.reserve(size);
-    key_.reserve(size);
-  }
-
- private:
-  // The following private routines are used in a supportive role
-  // for managing the heap and keeping the heap properties.
-
-  // Computes left child of parent.
-  static int Left(int i) {
-    return 2 * (i + 1) - 1;  // 0 -> 1, 1 -> 3
-  }
-
-  // Computes right child of parent.
-  static int Right(int i) {
-    return 2 * (i + 1);  // 0 -> 2, 1 -> 4
-  }
-
-  // Given a child computes parent.
-  static int Parent(int i) {
-    return (i - 1) / 2;  // 0 -> 0, 1 -> 0, 2 -> 0,  3 -> 1,  4 -> 1, ...
-  }
-
-  // Swaps a child and parent. Use to move element up/down tree. Note the use of
-  // a little trick here. When we swap we need to swap:
-  //
-  // - the value
-  // - the associated keys
-  // - the position of the value in the heap
-  void Swap(int j, int k) {
-    const auto tkey = key_[j];
-    pos_[key_[j] = key_[k]] = j;
-    pos_[key_[k] = tkey] = k;
-    using std::swap;
-    swap(values_[j], values_[k]);
-  }
-
-  // Heapifies the subtree rooted at index i.
-  void Heapify(int i) {
-    const auto l = Left(i);
-    const auto r = Right(i);
-    auto largest = (l < size_ && comp_(values_[l], values_[i])) ? l : i;
-    if (r < size_ && comp_(values_[r], values_[largest])) largest = r;
-    if (largest != i) {
-      Swap(i, largest);
-      Heapify(largest);
-    }
-  }
-
-  // Inserts (updates) element at subtree rooted at index i.
-  int Insert(const Value &value, int i) {
-    int p;
-    while (i > 0 && !comp_(values_[p = Parent(i)], value)) {
-      Swap(i, p);
-      i = p;
-    }
-    return key_[i];
-  }
-
- private:
-  const Compare comp_;
-
-  std::vector<int> pos_;
-  std::vector<int> key_;
-  std::vector<Value> values_;
-  int size_;
-};
-
-template <class T, class Compare>
-constexpr int Heap<T, Compare>::kNoKey;
-
-}  // namespace fst
-
-#endif  // FST_HEAP_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/icu.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/icu.h
deleted file mode 100644
index 4cd91ba9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/icu.h
+++ /dev/null
@@ -1,108 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// This library implements an unrestricted Thompson/Pike UTF-8 parser and
-// serializer.  UTF-8 is a restricted subset of this byte stream encoding. See
-// http://en.wikipedia.org/wiki/UTF-8 for a good description of the encoding
-// details.
-
-#ifndef FST_ICU_H_
-#define FST_ICU_H_
-
-#include <sstream>
-#include <vector>
-
-#include <fst/log.h>
-
-namespace fst {
-
-// This function writes UTF-8 codepoints into a vector of Labels, truncating if
-// necessary. It is possible to use this sensibly with as little as 16 bits of
-// Label precision (i.e., when all characters are within the Basic Multilingual
-// Plane). With 21 bits, one can encode all UTF-8 codepoints, including those
-// from the various Astral Planes. Naturally, it is safe to use this with larger
-// Labels (e.g., 64 bits).
-template <class Label>
-bool UTF8StringToLabels(const string &str, std::vector<Label> *labels) {
-  const auto *data = str.data();
-  const auto length = str.size();
-  for (size_t i = 0; i < length;) {
-    int c = data[i++] & 0xff;
-    if ((c & 0x80) == 0) {
-      labels->push_back(c);
-    } else {
-      if ((c & 0xc0) == 0x80) {
-        LOG(ERROR) << "UTF8StringToLabels: Continuation byte as lead byte";
-        return false;
-      }
-      int count =
-          (c >= 0xc0) + (c >= 0xe0) + (c >= 0xf0) + (c >= 0xf8) + (c >= 0xfc);
-      int32 code = c & ((1 << (6 - count)) - 1);
-      while (count != 0) {
-        if (i == length) {
-          LOG(ERROR) << "UTF8StringToLabels: Truncated UTF-8 byte sequence";
-          return false;
-        }
-        char cb = data[i++];
-        if ((cb & 0xc0) != 0x80) {
-          LOG(ERROR) << "UTF8StringToLabels: Missing/invalid continuation byte";
-          return false;
-        }
-        code = (code << 6) | (cb & 0x3f);
-        count--;
-      }
-      if (code < 0) {
-        // Should be unreachable.
-        LOG(ERROR) << "UTF8StringToLabels: Invalid character found: " << c;
-        return false;
-      }
-      labels->push_back(code);
-    }
-  }
-  return true;
-}
-
-template <class Label>
-bool LabelsToUTF8String(const std::vector<Label> &labels, string *str) {
-  std::ostringstream ostr;
-  for (size_t i = 0; i < labels.size(); ++i) {
-    int32 code = labels[i];
-    if (code < 0) {
-      LOG(ERROR) << "LabelsToUTF8String: Invalid character found: " << code;
-      return false;
-    } else if (code < 0x80) {
-      ostr << static_cast<char>(code);
-    } else if (code < 0x800) {
-      ostr << static_cast<char>((code >> 6) | 0xc0);
-      ostr << static_cast<char>((code & 0x3f) | 0x80);
-    } else if (code < 0x10000) {
-      ostr << static_cast<char>((code >> 12) | 0xe0);
-      ostr << static_cast<char>(((code >> 6) & 0x3f) | 0x80);
-      ostr << static_cast<char>((code & 0x3f) | 0x80);
-    } else if (code < 0x200000) {
-      ostr << static_cast<char>((code >> 18) | 0xf0);
-      ostr << static_cast<char>(((code >> 12) & 0x3f) | 0x80);
-      ostr << static_cast<char>(((code >> 6) & 0x3f) | 0x80);
-      ostr << static_cast<char>((code & 0x3f) | 0x80);
-    } else if (code < 0x4000000) {
-      ostr << static_cast<char>((code >> 24) | 0xf8);
-      ostr << static_cast<char>(((code >> 18) & 0x3f) | 0x80);
-      ostr << static_cast<char>(((code >> 12) & 0x3f) | 0x80);
-      ostr << static_cast<char>(((code >> 6) & 0x3f) | 0x80);
-      ostr << static_cast<char>((code & 0x3f) | 0x80);
-    } else {
-      ostr << static_cast<char>((code >> 30) | 0xfc);
-      ostr << static_cast<char>(((code >> 24) & 0x3f) | 0x80);
-      ostr << static_cast<char>(((code >> 18) & 0x3f) | 0x80);
-      ostr << static_cast<char>(((code >> 12) & 0x3f) | 0x80);
-      ostr << static_cast<char>(((code >> 6) & 0x3f) | 0x80);
-      ostr << static_cast<char>((code & 0x3f) | 0x80);
-    }
-  }
-  *str = ostr.str();
-  return true;
-}
-
-}  // namespace fst
-
-#endif  // FST_ICU_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/intersect.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/intersect.h
deleted file mode 100644
index b07ff455..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/intersect.h
+++ /dev/null
@@ -1,153 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to compute the intersection of two FSAs.
-
-#ifndef FST_INTERSECT_H_
-#define FST_INTERSECT_H_
-
-#include <algorithm>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/compose.h>
-
-
-namespace fst {
-
-using IntersectOptions = ComposeOptions;
-
-template <class Arc, class M = Matcher<Fst<Arc>>,
-          class Filter = SequenceComposeFilter<M>,
-          class StateTable =
-              GenericComposeStateTable<Arc, typename Filter::FilterState>>
-struct IntersectFstOptions
-    : public ComposeFstOptions<Arc, M, Filter, StateTable> {
-  IntersectFstOptions() {}
-
-  explicit IntersectFstOptions(const CacheOptions &opts, M *matcher1 = nullptr,
-                               M *matcher2 = nullptr, Filter *filter = nullptr,
-                               StateTable *state_table = nullptr)
-      : ComposeFstOptions<Arc, M, Filter, StateTable>(opts, matcher1, matcher2,
-                                                      filter, state_table) {}
-};
-
-// Computes the intersection (Hadamard product) of two FSAs. This version is a
-// delayed FST. Only strings that are in both automata are retained in the
-// result.
-//
-// The two arguments must be acceptors. One of the arguments must be
-// label-sorted.
-//
-// Complexity: same as ComposeFst.
-//
-// Caveats: same as ComposeFst.
-template <class A>
-class IntersectFst : public ComposeFst<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using ComposeFst<A>::CreateBase;
-  using ComposeFst<A>::CreateBase1;
-  using ComposeFst<A>::Properties;
-
-  IntersectFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-               const CacheOptions &opts = CacheOptions())
-      : ComposeFst<Arc>(CreateBase(fst1, fst2, opts)) {
-    const bool acceptors =
-        fst1.Properties(kAcceptor, true) && fst2.Properties(kAcceptor, true);
-    if (!acceptors) {
-      FSTERROR() << "IntersectFst: Input FSTs are not acceptors";
-      GetMutableImpl()->SetProperties(kError);
-    }
-  }
-
-  template <class M, class Filter, class StateTable>
-  IntersectFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-               const IntersectFstOptions<Arc, M, Filter, StateTable> &opts)
-      : ComposeFst<Arc>(CreateBase1(fst1, fst2, opts)) {
-    const bool acceptors =
-        fst1.Properties(kAcceptor, true) && fst2.Properties(kAcceptor, true);
-    if (!acceptors) {
-      FSTERROR() << "IntersectFst: input FSTs are not acceptors";
-      GetMutableImpl()->SetProperties(kError);
-    }
-  }
-
-  // See Fst<>::Copy() for doc.
-  IntersectFst(const IntersectFst<Arc> &fst, bool safe = false)
-      : ComposeFst<Arc>(fst, safe) {}
-
-  // Get a copy of this IntersectFst. See Fst<>::Copy() for further doc.
-  IntersectFst<Arc> *Copy(bool safe = false) const override {
-    return new IntersectFst<Arc>(*this, safe);
-  }
-
- private:
-  using ImplToFst<internal::ComposeFstImplBase<A>>::GetImpl;
-  using ImplToFst<internal::ComposeFstImplBase<A>>::GetMutableImpl;
-};
-
-// Specialization for IntersectFst.
-template <class Arc>
-class StateIterator<IntersectFst<Arc>> : public StateIterator<ComposeFst<Arc>> {
- public:
-  explicit StateIterator(const IntersectFst<Arc> &fst)
-      : StateIterator<ComposeFst<Arc>>(fst) {}
-};
-
-// Specialization for IntersectFst.
-template <class Arc>
-class ArcIterator<IntersectFst<Arc>> : public ArcIterator<ComposeFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const IntersectFst<Arc> &fst, StateId s)
-      : ArcIterator<ComposeFst<Arc>>(fst, s) {}
-};
-
-// Useful alias when using StdArc.
-using StdIntersectFst = IntersectFst<StdArc>;
-
-// Computes the intersection (Hadamard product) of two FSAs. This version
-// writes the intersection to an output MurableFst. Only strings that are in
-// both automata are retained in the result.
-//
-// The two arguments must be acceptors. One of the arguments must be
-// label-sorted.
-//
-// Complexity: same as Compose.
-//
-// Caveats: same as Compose.
-template <class Arc>
-void Intersect(const Fst<Arc> &ifst1, const Fst<Arc> &ifst2,
-               MutableFst<Arc> *ofst,
-               const IntersectOptions &opts = IntersectOptions()) {
-  using M = Matcher<Fst<Arc>>;
-  if (opts.filter_type == AUTO_FILTER) {
-    CacheOptions nopts;
-    nopts.gc_limit = 0;  // Cache only the last state for fastest copy.
-    *ofst = IntersectFst<Arc>(ifst1, ifst2, nopts);
-  } else if (opts.filter_type == SEQUENCE_FILTER) {
-    IntersectFstOptions<Arc> iopts;
-    iopts.gc_limit = 0;  // Cache only the last state for fastest copy.
-    *ofst = IntersectFst<Arc>(ifst1, ifst2, iopts);
-  } else if (opts.filter_type == ALT_SEQUENCE_FILTER) {
-    IntersectFstOptions<Arc, M, AltSequenceComposeFilter<M>> iopts;
-    iopts.gc_limit = 0;  // Cache only the last state for fastest copy.
-    *ofst = IntersectFst<Arc>(ifst1, ifst2, iopts);
-  } else if (opts.filter_type == MATCH_FILTER) {
-    IntersectFstOptions<Arc, M, MatchComposeFilter<M>> iopts;
-    iopts.gc_limit = 0;  // Cache only the last state for fastest copy.
-    *ofst = IntersectFst<Arc>(ifst1, ifst2, iopts);
-  }
-  if (opts.connect) Connect(ofst);
-}
-
-}  // namespace fst
-
-#endif  // FST_INTERSECT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/interval-set.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/interval-set.h
deleted file mode 100644
index 0942ea00..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/interval-set.h
+++ /dev/null
@@ -1,398 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to represent and operate on sets of intervals.
-
-#ifndef FST_INTERVAL_SET_H_
-#define FST_INTERVAL_SET_H_
-
-#include <algorithm>
-#include <iostream>
-#include <vector>
-
-
-#include <fst/util.h>
-
-
-namespace fst {
-
-// Half-open integral interval [a, b) of signed integers of type T.
-template <class T>
-struct IntInterval {
-  T begin;
-  T end;
-
-  IntInterval() : begin(-1), end(-1) {}
-
-  IntInterval(T begin, T end) : begin(begin), end(end) {}
-
-  bool operator<(const IntInterval<T> &i) const {
-    return begin < i.begin || (begin == i.begin && end > i.end);
-  }
-
-  bool operator==(const IntInterval<T> &i) const {
-    return begin == i.begin && end == i.end;
-  }
-
-  bool operator!=(const IntInterval<T> &i) const {
-    return begin != i.begin || end != i.end;
-  }
-
-  std::istream &Read(std::istream &strm) {
-    T n;
-    ReadType(strm, &n);
-    begin = n;
-    ReadType(strm, &n);
-    end = n;
-    return strm;
-  }
-
-  std::ostream &Write(std::ostream &strm) const {
-    T n = begin;
-    WriteType(strm, n);
-    n = end;
-    WriteType(strm, n);
-    return strm;
-  }
-};
-
-// Stores IntIntervals<T> in a vector. In addition, keeps the count of points in
-// all intervals.
-template <class T>
-class VectorIntervalStore {
- public:
-  using Interval = IntInterval<T>;
-  using Iterator = typename std::vector<Interval>::const_iterator;
-
-  VectorIntervalStore() : count_(-1) {}
-
-  std::vector<Interval> *MutableIntervals() { return &intervals_; }
-
-  const Interval *Intervals() const { return intervals_.data(); }
-
-  T Size() const { return intervals_.size(); }
-
-  T Count() const { return count_; }
-
-  void SetCount(T count) { count_ = count; }
-
-  void Clear() {
-    intervals_.clear();
-    count_ = 0;
-  }
-
-  Iterator begin() const { return intervals_.begin(); }
-
-  Iterator end() const { return intervals_.end(); }
-
-  std::istream &Read(std::istream &strm) {
-    ReadType(strm, &intervals_);
-    return ReadType(strm, &count_);
-  }
-
-  std::ostream &Write(std::ostream &strm) const {
-    WriteType(strm, intervals_);
-    return WriteType(strm, count_);
-  }
-
- private:
-  std::vector<Interval> intervals_;
-  T count_;
-};
-
-// Stores and operates on a set of half-open integral intervals [a, b)
-// of signed integers of type T.
-template <class T, class Store = VectorIntervalStore<T>>
-class IntervalSet {
- public:
-  using Interval = IntInterval<T>;
-
-  template <class... A>
-  explicit IntervalSet(A... args) : intervals_(args...) {}
-
-  // Returns the interval set as a vector.
-  std::vector<Interval> *MutableIntervals() {
-    return intervals_.MutableIntervals();
-  }
-
-  // Returns a pointer to an array of Size() elements.
-  const Interval *Intervals() const { return intervals_.Intervals(); }
-
-  bool Empty() const { return Size() == 0; }
-
-  T Size() const { return intervals_.Size(); }
-
-  // Number of points in the intervals (undefined if not normalized).
-  T Count() const { return intervals_.Count(); }
-
-  void Clear() { intervals_.Clear(); }
-
-  // Adds an interval set to the set. The result may not be normalized.
-  void Union(const IntervalSet<T, Store> &iset) {
-    intervals_.MutableIntervals()->insert(intervals_.MutableIntervals()->end(),
-                                          iset.intervals_.begin(),
-                                          iset.intervals_.end());
-  }
-
-  // Requires intervals be normalized.
-  bool Member(T value) const {
-    const Interval interval(value, value);
-    auto lb = std::lower_bound(intervals_.begin(), intervals_.end(), interval);
-    if (lb == intervals_.begin()) return false;
-    return (--lb)->end > value;
-  }
-
-  // Requires intervals be normalized.
-  bool operator==(const IntervalSet<T, Store> &iset) const {
-    return Size() == iset.Size() &&
-           std::equal(intervals_.begin(), intervals_.end(),
-                      iset.intervals_.begin());
-  }
-
-  // Requires intervals be normalized.
-  bool operator!=(const IntervalSet<T, Store> &iset) const {
-    return Size() != iset.Size() ||
-           !std::equal(intervals_.begin(), intervals_.end(),
-                       iset.intervals_.begin());
-  }
-
-  bool Singleton() const {
-    return Size() == 1 &&
-           intervals_.begin()->begin + 1 == intervals_.begin()->end;
-  }
-
-  // Sorts, collapses overlapping and adjacent interals, and sets count.
-  void Normalize();
-
-  // Intersects an interval set with the set. Requires intervals be normalized.
-  // The result is normalized.
-  void Intersect(const IntervalSet<T, Store> &iset,
-                 IntervalSet<T, Store> *oset) const;
-
-  // Complements the set w.r.t [0, maxval). Requires intervals be normalized.
-  // The result is normalized.
-  void Complement(T maxval, IntervalSet<T, Store> *oset) const;
-
-  // Subtract an interval set from the set. Requires intervals be normalized.
-  // The result is normalized.
-  void Difference(const IntervalSet<T, Store> &iset,
-                  IntervalSet<T, Store> *oset) const;
-
-  // Determines if an interval set overlaps with the set. Requires intervals be
-  // normalized.
-  bool Overlaps(const IntervalSet<T, Store> &iset) const;
-
-  // Determines if an interval set overlaps with the set but neither is
-  // contained in the other. Requires intervals be normalized.
-  bool StrictlyOverlaps(const IntervalSet<T, Store> &iset) const;
-
-  // Determines if an interval set is contained within the set. Requires
-  // intervals be normalized.
-  bool Contains(const IntervalSet<T, Store> &iset) const;
-
-  std::istream &Read(std::istream &strm) { return intervals_.Read(strm); }
-
-  std::ostream &Write(std::ostream &strm) const {
-    return intervals_.Write(strm);
-  }
-
-  typename Store::Iterator begin() const { return intervals_.begin(); }
-
-  typename Store::Iterator end() const { return intervals_.end(); }
-
- private:
-  Store intervals_;
-};
-
-// Sorts, collapses overlapping and adjacent intervals, and sets count.
-template <typename T, class Store>
-void IntervalSet<T, Store>::Normalize() {
-  auto &intervals = *intervals_.MutableIntervals();
-  std::sort(intervals.begin(), intervals.end());
-  T count = 0;
-  T size = 0;
-  for (T i = 0; i < intervals.size(); ++i) {
-    auto &inti = intervals[i];
-    if (inti.begin == inti.end) continue;
-    for (T j = i + 1; j < intervals.size(); ++j) {
-      auto &intj = intervals[j];
-      if (intj.begin > inti.end) break;
-      if (intj.end > inti.end) inti.end = intj.end;
-      ++i;
-    }
-    count += inti.end - inti.begin;
-    intervals[size++] = inti;
-  }
-  intervals.resize(size);
-  intervals_.SetCount(count);
-}
-
-// Intersects an interval set with the set. Requires intervals be normalized.
-// The result is normalized.
-template <typename T, class Store>
-void IntervalSet<T, Store>::Intersect(const IntervalSet<T, Store> &iset,
-                                      IntervalSet<T, Store> *oset) const {
-  auto *ointervals = oset->MutableIntervals();
-  auto it1 = intervals_.begin();
-  auto it2 = iset.intervals_.begin();
-  ointervals->clear();
-  T count = 0;
-  while (it1 != intervals_.end() && it2 != iset.intervals_.end()) {
-    if (it1->end <= it2->begin) {
-      ++it1;
-    } else if (it2->end <= it1->begin) {
-      ++it2;
-    } else {
-      ointervals->emplace_back(std::max(it1->begin, it2->begin),
-                               std::min(it1->end, it2->end));
-      count += ointervals->back().end - ointervals->back().begin;
-      if (it1->end < it2->end) {
-        ++it1;
-      } else {
-        ++it2;
-      }
-    }
-  }
-  oset->intervals_.SetCount(count);
-}
-
-// Complements the set w.r.t [0, maxval). Requires intervals be normalized.
-// The result is normalized.
-template <typename T, class Store>
-void IntervalSet<T, Store>::Complement(T maxval,
-                                       IntervalSet<T, Store> *oset) const {
-  auto *ointervals = oset->MutableIntervals();
-  ointervals->clear();
-  T count = 0;
-  Interval interval;
-  interval.begin = 0;
-  for (auto it = intervals_.begin(); it != intervals_.end(); ++it) {
-    interval.end = std::min(it->begin, maxval);
-    if ((interval.begin) < (interval.end)) {
-      ointervals->push_back(interval);
-      count += interval.end - interval.begin;
-    }
-    interval.begin = it->end;
-  }
-  interval.end = maxval;
-  if ((interval.begin) < (interval.end)) {
-    ointervals->push_back(interval);
-    count += interval.end - interval.begin;
-  }
-  oset->intervals_.SetCount(count);
-}
-
-// Subtract an interval set from the set. Requires intervals be normalized.
-// The result is normalized.
-template <typename T, class Store>
-void IntervalSet<T, Store>::Difference(const IntervalSet<T, Store> &iset,
-                                       IntervalSet<T, Store> *oset) const {
-  if (Empty()) {
-    oset->MutableIntervals()->clear();
-    oset->intervals_.SetCount(0);
-  } else {
-    IntervalSet<T, Store> cset;
-    iset.Complement(intervals_.Intervals()[intervals_.Size() - 1].end, &cset);
-    Intersect(cset, oset);
-  }
-}
-
-// Determines if an interval set overlaps with the set. Requires intervals be
-// normalized.
-template <typename T, class Store>
-bool IntervalSet<T, Store>::Overlaps(const IntervalSet<T, Store> &iset) const {
-  auto it1 = intervals_.begin();
-  auto it2 = iset.intervals_.begin();
-  while (it1 != intervals_.end() && it2 != iset.intervals_.end()) {
-    if (it1->end <= it2->begin) {
-      ++it1;
-    } else if (it2->end <= it1->begin) {
-      ++it2;
-    } else {
-      return true;
-    }
-  }
-  return false;
-}
-
-// Determines if an interval set overlaps with the set but neither is contained
-// in the other. Requires intervals be normalized.
-template <typename T, class Store>
-bool IntervalSet<T, Store>::StrictlyOverlaps(
-    const IntervalSet<T, Store> &iset) const {
-  auto it1 = intervals_.begin();
-  auto it2 = iset.intervals_.begin();
-  bool only1 = false;    // Point in intervals_ but not intervals.
-  bool only2 = false;    // Point in intervals but not intervals_.
-  bool overlap = false;  // Point in both intervals_ and intervals.
-  while (it1 != intervals_.end() && it2 != iset.intervals_.end()) {
-    if (it1->end <= it2->begin) {  // no overlap - it1 first
-      only1 = true;
-      ++it1;
-    } else if (it2->end <= it1->begin) {  // no overlap - it2 first
-      only2 = true;
-      ++it2;
-    } else if (it2->begin == it1->begin && it2->end == it1->end) {  // equals
-      overlap = true;
-      ++it1;
-      ++it2;
-    } else if (it2->begin <= it1->begin && it2->end >= it1->end) {  // 1 c 2
-      only2 = true;
-      overlap = true;
-      ++it1;
-    } else if (it1->begin <= it2->begin && it1->end >= it2->end) {  // 2 c 1
-      only1 = true;
-      overlap = true;
-      ++it2;
-    } else {  // Strict overlap.
-      only1 = true;
-      only2 = true;
-      overlap = true;
-    }
-    if (only1 == true && only2 == true && overlap == true) return true;
-  }
-  if (it1 != intervals_.end()) only1 = true;
-  if (it2 != iset.intervals_.end()) only2 = true;
-  return only1 == true && only2 == true && overlap == true;
-}
-
-// Determines if an interval set is contained within the set. Requires intervals
-// be normalized.
-template <typename T, class Store>
-bool IntervalSet<T, Store>::Contains(const IntervalSet<T, Store> &iset) const {
-  if (iset.Count() > Count()) return false;
-  auto it1 = intervals_.begin();
-  auto it2 = iset.intervals_.begin();
-  while (it1 != intervals_.end() && it2 != iset.intervals_.end()) {
-    if ((it1->end) <= (it2->begin)) {  // No overlap; it1 first.
-      ++it1;
-    } else if ((it2->begin) < (it1->begin) ||
-               (it2->end) > (it1->end)) {  // No C.
-      return false;
-    } else if (it2->end == it1->end) {
-      ++it1;
-      ++it2;
-    } else {
-      ++it2;
-    }
-  }
-  return it2 == iset.intervals_.end();
-}
-
-template <typename T, class Store>
-std::ostream &operator<<(std::ostream &strm, const IntervalSet<T, Store> &s) {
-  strm << "{";
-  for (T i = 0; i < s.Size(); ++i) {
-    if (i > 0) {
-      strm << ",";
-    }
-    const auto &interval = s.Intervals()[i];
-    strm << "[" << interval.begin << "," << interval.end << ")";
-  }
-  strm << "}";
-  return strm;
-}
-
-}  // namespace fst
-
-#endif  // FST_INTERVAL_SET_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/invert.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/invert.h
deleted file mode 100644
index b37bd116..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/invert.h
+++ /dev/null
@@ -1,119 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to invert an FST.
-
-#ifndef FST_INVERT_H_
-#define FST_INVERT_H_
-
-#include <fst/arc-map.h>
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-// Mapper to implement inversion of an arc.
-template <class A>
-struct InvertMapper {
-  using FromArc = A;
-  using ToArc = A;
-
-  InvertMapper() {}
-
-  ToArc operator()(const FromArc &arc) {
-    return ToArc(arc.olabel, arc.ilabel, arc.weight, arc.nextstate);
-  }
-
-  MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  MapSymbolsAction InputSymbolsAction() const { return MAP_CLEAR_SYMBOLS; }
-
-  MapSymbolsAction OutputSymbolsAction() const { return MAP_CLEAR_SYMBOLS; }
-
-  uint64 Properties(uint64 props) { return InvertProperties(props); }
-};
-
-// Inverts the transduction corresponding to an FST by exchanging the
-// FST's input and output labels. This version modifies its input.
-//
-// Complexity:
-//
-//   Time: O(V + E)
-//   Space: O(1)
-//
-// where V is the number of states and E is the number of arcs.
-template <class Arc>
-inline void Invert(MutableFst<Arc> *fst) {
-  std::unique_ptr<SymbolTable> input(
-      fst->InputSymbols() ? fst->InputSymbols()->Copy() : nullptr);
-  std::unique_ptr<SymbolTable> output(
-      fst->OutputSymbols() ? fst->OutputSymbols()->Copy() : nullptr);
-  ArcMap(fst, InvertMapper<Arc>());
-  fst->SetInputSymbols(output.get());
-  fst->SetOutputSymbols(input.get());
-}
-
-// Inverts the transduction corresponding to an FST by exchanging the
-// FST's input and output labels. This version is a delayed FST.
-//
-// Complexity:
-//
-//   Time: O(v + e)
-//   Space: O(1)
-//
-// where v is the number of states visited and e is the number of arcs visited.
-// Constant time and to visit an input state or arc is assumed and exclusive of
-// caching.
-template <class A>
-class InvertFst : public ArcMapFst<A, A, InvertMapper<A>> {
- public:
-  using Arc = A;
-
-  using Mapper = InvertMapper<Arc>;
-  using Impl = internal::ArcMapFstImpl<A, A, InvertMapper<A>>;
-
-  explicit InvertFst(const Fst<Arc> &fst)
-      : ArcMapFst<Arc, Arc, Mapper>(fst, Mapper()) {
-    GetMutableImpl()->SetOutputSymbols(fst.InputSymbols());
-    GetMutableImpl()->SetInputSymbols(fst.OutputSymbols());
-  }
-
-  // See Fst<>::Copy() for doc.
-  InvertFst(const InvertFst<Arc> &fst, bool safe = false)
-      : ArcMapFst<Arc, Arc, Mapper>(fst, safe) {}
-
-  // Get a copy of this InvertFst. See Fst<>::Copy() for further doc.
-  InvertFst<Arc> *Copy(bool safe = false) const override {
-    return new InvertFst(*this, safe);
-  }
-
- private:
-  using ImplToFst<Impl>::GetMutableImpl;
-};
-
-// Specialization for InvertFst.
-template <class Arc>
-class StateIterator<InvertFst<Arc>>
-    : public StateIterator<ArcMapFst<Arc, Arc, InvertMapper<Arc>>> {
- public:
-  explicit StateIterator(const InvertFst<Arc> &fst)
-      : StateIterator<ArcMapFst<Arc, Arc, InvertMapper<Arc>>>(fst) {}
-};
-
-// Specialization for InvertFst.
-template <class Arc>
-class ArcIterator<InvertFst<Arc>>
-    : public ArcIterator<ArcMapFst<Arc, Arc, InvertMapper<Arc>>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const InvertFst<Arc> &fst, StateId s)
-      : ArcIterator<ArcMapFst<Arc, Arc, InvertMapper<Arc>>>(fst, s) {}
-};
-
-// Useful alias when using StdArc.
-using StdInvertFst = InvertFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_INVERT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/isomorphic.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/isomorphic.h
deleted file mode 100644
index f14625a2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/isomorphic.h
+++ /dev/null
@@ -1,178 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function to test two FSTs are isomorphic, i.e., they are equal up to a state
-// and arc re-ordering. FSTs should be deterministic when viewed as
-// unweighted automata.
-
-#ifndef FST_ISOMORPHIC_H_
-#define FST_ISOMORPHIC_H_
-
-#include <algorithm>
-#include <list>
-#include <type_traits>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/fst.h>
-
-
-namespace fst {
-namespace internal {
-
-// Orders weights for equality checking.
-template <class Weight>
-bool WeightCompare(Weight w1, Weight w2, float delta, bool *error) {
- if ((Weight::Properties() & kIdempotent) == kIdempotent) {
-  return NaturalLess<Weight>()(w1, w2);
- } else { // (Weight::Properties() & kIdempotent) != kIdempotent
-  // No natural order; use hash
-  const auto q1 = w1.Quantize(delta);
-  const auto q2 = w2.Quantize(delta);
-  auto n1 = q1.Hash();
-  auto n2 = q2.Hash();
-  // Hash not unique; very unlikely to happen.
-  if (n1 == n2 && q1 != q2) {
-    VLOG(1) << "Isomorphic: Weight hash collision";
-    *error = true;
-  }
-  return n1 < n2;
- }
-}
-
-template <class Arc>
-class Isomorphism {
-  using StateId = typename Arc::StateId;
-
- public:
-  Isomorphism(const Fst<Arc> &fst1, const Fst<Arc> &fst2, float delta)
-      : fst1_(fst1.Copy()),
-        fst2_(fst2.Copy()),
-        delta_(delta),
-        error_(false),
-        comp_(delta, &error_) {}
-
-  // Checks if input FSTs are isomorphic.
-  bool IsIsomorphic() {
-    if (fst1_->Start() == kNoStateId && fst2_->Start() == kNoStateId) {
-      return true;
-    }
-    if (fst1_->Start() == kNoStateId || fst2_->Start() == kNoStateId) {
-      return false;
-    }
-    PairState(fst1_->Start(), fst2_->Start());
-    while (!queue_.empty()) {
-      const auto &pr = queue_.front();
-      if (!IsIsomorphicState(pr.first, pr.second)) return false;
-      queue_.pop_front();
-    }
-    return true;
-  }
-
-  bool Error() const { return error_; }
-
- private:
-  // Orders arcs for equality checking.
-  class ArcCompare {
-   public:
-    ArcCompare(float delta, bool *error) : delta_(delta), error_(error) {}
-
-    bool operator()(const Arc &arc1, const Arc &arc2) const {
-      if (arc1.ilabel < arc2.ilabel) return true;
-      if (arc1.ilabel > arc2.ilabel) return false;
-      if (arc1.olabel < arc2.olabel) return true;
-      if (arc1.olabel > arc2.olabel) return false;
-      return WeightCompare(arc1.weight, arc2.weight, delta_, error_);
-    }
-
-   private:
-    float delta_;
-    bool *error_;
-  };
-
-  // Maintains state correspondences and queue.
-  bool PairState(StateId s1, StateId s2) {
-    if (state_pairs_.size() <= s1) state_pairs_.resize(s1 + 1, kNoStateId);
-    if (state_pairs_[s1] == s2) {
-      return true;  // already seen this pair
-    } else if (state_pairs_[s1] != kNoStateId) {
-      return false;  // s1 already paired with another s2
-    }
-    state_pairs_[s1] = s2;
-    queue_.push_back(std::make_pair(s1, s2));
-    return true;
-  }
-
-  // Checks if state pair is isomorphic
-  bool IsIsomorphicState(StateId s1, StateId s2);
-
-  std::unique_ptr<Fst<Arc>> fst1_;
-  std::unique_ptr<Fst<Arc>> fst2_;
-  float delta_;                          // Weight equality delta.
-  std::vector<Arc> arcs1_;               // For sorting arcs on FST1.
-  std::vector<Arc> arcs2_;               // For sorting arcs on FST2.
-  std::vector<StateId> state_pairs_;     // Maintains state correspondences.
-  std::list<std::pair<StateId, StateId>> queue_;  // Queue of state pairs.
-  bool error_;                           // Error flag.
-  ArcCompare comp_;
-};
-
-template <class Arc>
-bool Isomorphism<Arc>::IsIsomorphicState(StateId s1, StateId s2) {
-  if (!ApproxEqual(fst1_->Final(s1), fst2_->Final(s2), delta_)) return false;
-  auto narcs1 = fst1_->NumArcs(s1);
-  auto narcs2 = fst2_->NumArcs(s2);
-  if (narcs1 != narcs2) return false;
-  ArcIterator<Fst<Arc>> aiter1(*fst1_, s1);
-  ArcIterator<Fst<Arc>> aiter2(*fst2_, s2);
-  arcs1_.clear();
-  arcs1_.reserve(narcs1);
-  arcs2_.clear();
-  arcs2_.reserve(narcs2);
-  for (; !aiter1.Done(); aiter1.Next(), aiter2.Next()) {
-    arcs1_.push_back(aiter1.Value());
-    arcs2_.push_back(aiter2.Value());
-  }
-  std::sort(arcs1_.begin(), arcs1_.end(), comp_);
-  std::sort(arcs2_.begin(), arcs2_.end(), comp_);
-  for (size_t i = 0; i < arcs1_.size(); ++i) {
-    const auto &arc1 = arcs1_[i];
-    const auto &arc2 = arcs2_[i];
-    if (arc1.ilabel != arc2.ilabel) return false;
-    if (arc1.olabel != arc2.olabel) return false;
-    if (!ApproxEqual(arc1.weight, arc2.weight, delta_)) return false;
-    if (!PairState(arc1.nextstate, arc2.nextstate)) return false;
-    if (i > 0) {  // Checks for non-determinism.
-      const auto &arc0 = arcs1_[i - 1];
-      if (arc1.ilabel == arc0.ilabel && arc1.olabel == arc0.olabel &&
-          ApproxEqual(arc1.weight, arc0.weight, delta_)) {
-        VLOG(1) << "Isomorphic: Non-determinism as an unweighted automaton";
-        error_ = true;
-        return false;
-      }
-    }
-  }
-  return true;
-}
-
-}  // namespace internal
-
-// Tests if two FSTs have the same states and arcs up to a reordering.
-// Inputs should be non-deterministic when viewed as unweighted automata.
-template <class Arc>
-bool Isomorphic(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-                float delta = kDelta) {
-  internal::Isomorphism<Arc> iso(fst1, fst2, delta);
-  bool result = iso.IsIsomorphic();
-  if (iso.Error()) {
-    FSTERROR() << "Isomorphic: Cannot determine if inputs are isomorphic";
-    return false;
-  } else {
-    return result;
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_ISOMORPHIC_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/label-reachable.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/label-reachable.h
deleted file mode 100644
index fca6cc27..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/label-reachable.h
+++ /dev/null
@@ -1,511 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to determine if a non-epsilon label can be read as the first
-// non-epsilon symbol along some path from a given state.
-
-#ifndef FST_LABEL_REACHABLE_H_
-#define FST_LABEL_REACHABLE_H_
-
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/accumulator.h>
-#include <fst/arcsort.h>
-#include <fst/interval-set.h>
-#include <fst/state-reachable.h>
-#include <fst/util.h>
-#include <fst/vector-fst.h>
-
-
-namespace fst {
-
-// Stores shareable data for label reachable class copies.
-template <typename Label>
-class LabelReachableData {
- public:
-  using LabelIntervalSet = IntervalSet<Label>;
-  using Interval = typename LabelIntervalSet::Interval;
-
-  explicit LabelReachableData(bool reach_input, bool keep_relabel_data = true)
-      : reach_input_(reach_input),
-        keep_relabel_data_(keep_relabel_data),
-        have_relabel_data_(true),
-        final_label_(kNoLabel) {}
-
-  ~LabelReachableData() {}
-
-  bool ReachInput() const { return reach_input_; }
-
-  std::vector<LabelIntervalSet> *MutableIntervalSets() {
-    return &interval_sets_;
-  }
-
-  const LabelIntervalSet &GetIntervalSet(int s) const {
-    return interval_sets_[s];
-  }
-
-  int NumIntervalSets() const { return interval_sets_.size(); }
-
-  std::unordered_map<Label, Label> *Label2Index() {
-    if (!have_relabel_data_) {
-      FSTERROR() << "LabelReachableData: No relabeling data";
-    }
-    return &label2index_;
-  }
-
-  void SetFinalLabel(Label final_label) { final_label_ = final_label; }
-
-  Label FinalLabel() const { return final_label_; }
-
-  static LabelReachableData<Label> *Read(std::istream &istrm,
-                                         const FstReadOptions &opts) {
-    auto *data = new LabelReachableData<Label>();
-    ReadType(istrm, &data->reach_input_);
-    ReadType(istrm, &data->keep_relabel_data_);
-    data->have_relabel_data_ = data->keep_relabel_data_;
-    if (data->keep_relabel_data_) ReadType(istrm, &data->label2index_);
-    ReadType(istrm, &data->final_label_);
-    ReadType(istrm, &data->interval_sets_);
-    return data;
-  }
-
-  bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const {
-    WriteType(ostrm, reach_input_);
-    WriteType(ostrm, keep_relabel_data_);
-    if (keep_relabel_data_) WriteType(ostrm, label2index_);
-    WriteType(ostrm, FinalLabel());
-    WriteType(ostrm, interval_sets_);
-    return true;
-  }
-
- private:
-  LabelReachableData() {}
-
-  bool reach_input_;                              // Input labels considered?
-  bool keep_relabel_data_;                        // Save label2index_ to file?
-  bool have_relabel_data_;                        // Using label2index_?
-  Label final_label_;                             // Final label.
-  std::unordered_map<Label, Label> label2index_;  // Finds index for a label.
-  std::vector<LabelIntervalSet> interval_sets_;   // Interval sets per state.
-};
-
-// Tests reachability of labels from a given state. If reach_input is true, then
-// input labels are considered, o.w. output labels are considered. To test for
-// reachability from a state s, first do SetState(s), then a label l can be
-// reached from state s of FST f iff Reach(r) is true where r = Relabel(l). The
-// relabeling is required to ensure a compact representation of the reachable
-// labels.
-
-// The whole FST can be relabeled instead with Relabel(&f, reach_input) so that
-// the test Reach(r) applies directly to the labels of the transformed FST f.
-// The relabeled FST will also be sorted appropriately for composition.
-//
-// Reachablity of a final state from state s (via an epsilon path) can be
-// tested with ReachFinal().
-//
-// Reachability can also be tested on the set of labels specified by an arc
-// iterator, useful for FST composition. In particular, Reach(aiter, ...) is
-// true if labels on the input (output) side of the transitions of the arc
-// iterator, when iter_input is true (false), can be reached from the state s.
-// The iterator labels must have already been relabeled.
-//
-// With the arc iterator test of reachability, the begin position, end position
-// and accumulated arc weight of the matches can be returned. The optional
-// template argument controls how reachable arc weights are accumulated. The
-// default uses semiring Plus(). Alternative ones can be used to distribute the
-// weights in composition in various ways.
-template <class Arc, class Accumulator = DefaultAccumulator<Arc>,
-          class D = LabelReachableData<typename Arc::Label>>
-class LabelReachable {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using Data = D;
-
-  using LabelIntervalSet = typename Data::LabelIntervalSet;
-
-  using Interval = typename LabelIntervalSet::Interval;
-
-  LabelReachable(const Fst<Arc> &fst, bool reach_input,
-                 Accumulator *accumulator = nullptr,
-                 bool keep_relabel_data = true)
-      : fst_(new VectorFst<Arc>(fst)),
-        s_(kNoStateId),
-        data_(std::make_shared<Data>(reach_input, keep_relabel_data)),
-        accumulator_(accumulator ? accumulator : new Accumulator()),
-        ncalls_(0),
-        nintervals_(0),
-        reach_fst_input_(false),
-        error_(false) {
-    const auto ins = fst_->NumStates();
-    TransformFst();
-    FindIntervals(ins);
-    fst_.reset();
-  }
-
-  explicit LabelReachable(std::shared_ptr<Data> data,
-                          Accumulator *accumulator = nullptr)
-      : s_(kNoStateId),
-        data_(std::move(data)),
-        accumulator_(accumulator ? accumulator : new Accumulator()),
-        ncalls_(0),
-        nintervals_(0),
-        reach_fst_input_(false),
-        error_(false) {}
-
-  LabelReachable(const LabelReachable<Arc, Accumulator, Data> &reachable,
-                 bool safe = false)
-      : s_(kNoStateId),
-        data_(reachable.data_),
-        accumulator_(new Accumulator(*reachable.accumulator_, safe)),
-        ncalls_(0),
-        nintervals_(0),
-        reach_fst_input_(reachable.reach_fst_input_),
-        error_(reachable.error_) {}
-
-  ~LabelReachable() {
-    if (ncalls_ > 0) {
-      VLOG(2) << "# of calls: " << ncalls_;
-      VLOG(2) << "# of intervals/call: " << (nintervals_ / ncalls_);
-    }
-  }
-
-  // Relabels w.r.t labels that give compact label sets.
-  Label Relabel(Label label) {
-    if (label == 0 || error_) return label;
-    auto &label2index = *data_->Label2Index();
-    auto &relabel = label2index[label];
-    if (!relabel) relabel = label2index.size() + 1;  // Adds new label.
-    return relabel;
-  }
-
-  // Relabels FST w.r.t to labels that give compact label sets.
-  void Relabel(MutableFst<Arc> *fst, bool relabel_input) {
-    for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
-         siter.Next()) {
-      for (MutableArcIterator<MutableFst<Arc>> aiter(fst, siter.Value());
-           !aiter.Done(); aiter.Next()) {
-        auto arc = aiter.Value();
-        if (relabel_input) {
-          arc.ilabel = Relabel(arc.ilabel);
-        } else {
-          arc.olabel = Relabel(arc.olabel);
-        }
-        aiter.SetValue(arc);
-      }
-    }
-    if (relabel_input) {
-      ArcSort(fst, ILabelCompare<Arc>());
-      fst->SetInputSymbols(nullptr);
-    } else {
-      ArcSort(fst, OLabelCompare<Arc>());
-      fst->SetOutputSymbols(nullptr);
-    }
-  }
-
-  // Returns relabeling pairs (cf. relabel.h::Relabel()). If avoid_collisions is
-  // true, extra pairs are added to ensure no collisions when relabeling
-  // automata that have labels unseen here.
-  void RelabelPairs(std::vector<std::pair<Label, Label>> *pairs,
-                    bool avoid_collisions = false) {
-    pairs->clear();
-    const auto &label2index = *data_->Label2Index();
-    // Maps labels to their new values in [1, label2index().size()].
-    for (auto it = label2index.begin(); it != label2index.end(); ++it) {
-      if (it->second != data_->FinalLabel()) {
-        pairs->push_back(std::make_pair(it->first, it->second));
-      }
-    }
-    if (avoid_collisions) {
-      // Ensures any label in [1, label2index().size()] is mapped either
-      // by the above step or to label2index() + 1 (to avoid collisions).
-      for (size_t i = 1; i <= label2index.size(); ++i) {
-        const auto it = label2index.find(i);
-        if (it == label2index.end() || it->second == data_->FinalLabel()) {
-          pairs->push_back(std::make_pair(i, label2index.size() + 1));
-        }
-      }
-    }
-  }
-
-  // Set current state. Optionally set state associated
-  // with arc iterator to be passed to Reach.
-  void SetState(StateId s, StateId aiter_s = kNoStateId) {
-    s_ = s;
-    if (aiter_s != kNoStateId) {
-      accumulator_->SetState(aiter_s);
-      if (accumulator_->Error()) error_ = true;
-    }
-  }
-
-  // Can reach this label from current state?
-  // Original labels must be transformed by the Relabel methods above.
-  bool Reach(Label label) const {
-    if (label == 0 || error_) return false;
-    return data_->GetIntervalSet(s_).Member(label);
-  }
-
-  // Can reach final state (via epsilon transitions) from this state?
-  bool ReachFinal() const {
-    if (error_) return false;
-    return data_->GetIntervalSet(s_).Member(data_->FinalLabel());
-  }
-
-  // Initialize with secondary FST to be used with Reach(Iterator,...).
-  // If reach_input = true, then arc input labels are considered in
-  // Reach(aiter, ...), o.w. output labels are considered. If copy is true, then
-  // the FST is a copy of the FST used in the previous call to this method
-  // (useful to avoid unnecessary updates).
-  template <class FST>
-  void ReachInit(const FST &fst, bool reach_input, bool copy = false) {
-    reach_fst_input_ = reach_input;
-    if (!fst.Properties(reach_fst_input_ ? kILabelSorted : kOLabelSorted,
-                        true)) {
-      FSTERROR() << "LabelReachable::ReachInit: Fst is not sorted";
-      error_ = true;
-    }
-    accumulator_->Init(fst, copy);
-    if (accumulator_->Error()) error_ = true;
-  }
-
-  // Can reach any arc iterator label between iterator positions
-  // aiter_begin and aiter_end?
-  // Arc iterator labels must be transformed by the Relabel methods
-  // above. If compute_weight is true, user may call ReachWeight().
-  template <class Iterator>
-  bool Reach(Iterator *aiter, ssize_t aiter_begin, ssize_t aiter_end,
-             bool compute_weight) {
-    if (error_) return false;
-    const auto &interval_set = data_->GetIntervalSet(s_);
-    ++ncalls_;
-    nintervals_ += interval_set.Size();
-    reach_begin_ = -1;
-    reach_end_ = -1;
-    reach_weight_ = Weight::Zero();
-    const auto flags = aiter->Flags();  // Save flags to restore them on exit.
-    aiter->SetFlags(kArcNoCache, kArcNoCache);  // Makes caching optional.
-    aiter->Seek(aiter_begin);
-    if (2 * (aiter_end - aiter_begin) < interval_set.Size()) {
-      // Checks each arc against intervals, setting arc iterator flags to only
-      // compute the ilabel or olabel values, since they are the only values
-      // required for most of the arcs processed.
-      aiter->SetFlags(reach_fst_input_ ? kArcILabelValue : kArcOLabelValue,
-                      kArcValueFlags);
-      Label reach_label = kNoLabel;
-      for (auto aiter_pos = aiter_begin; aiter_pos < aiter_end;
-           aiter->Next(), ++aiter_pos) {
-        const auto &arc = aiter->Value();
-        const auto label = reach_fst_input_ ? arc.ilabel : arc.olabel;
-        if (label == reach_label || Reach(label)) {
-          reach_label = label;
-          if (reach_begin_ < 0) reach_begin_ = aiter_pos;
-          reach_end_ = aiter_pos + 1;
-          if (compute_weight) {
-            if (!(aiter->Flags() & kArcWeightValue)) {
-              // If arc.weight wasn't computed by the call to aiter->Value()
-              // above, we need to call aiter->Value() again after having set
-              // the arc iterator flags to compute the arc weight value.
-              aiter->SetFlags(kArcWeightValue, kArcValueFlags);
-              const auto &arcb = aiter->Value();
-              // Call the accumulator.
-              reach_weight_ = accumulator_->Sum(reach_weight_, arcb.weight);
-              // Only ilabel or olabel required to process the following arcs.
-              aiter->SetFlags(
-                  reach_fst_input_ ? kArcILabelValue : kArcOLabelValue,
-                  kArcValueFlags);
-            } else {
-              // Calls the accumulator.
-              reach_weight_ = accumulator_->Sum(reach_weight_, arc.weight);
-            }
-          }
-        }
-      }
-    } else {
-      // Checks each interval against arcs.
-      auto begin_low = aiter_begin;
-      auto end_low = aiter_begin;
-      for (const auto &interval : interval_set) {
-        begin_low = LowerBound(aiter, end_low, aiter_end, interval.begin);
-        end_low = LowerBound(aiter, begin_low, aiter_end, interval.end);
-        if (end_low - begin_low > 0) {
-          if (reach_begin_ < 0) reach_begin_ = begin_low;
-          reach_end_ = end_low;
-          if (compute_weight) {
-            aiter->SetFlags(kArcWeightValue, kArcValueFlags);
-            reach_weight_ =
-                accumulator_->Sum(reach_weight_, aiter, begin_low, end_low);
-          }
-        }
-      }
-    }
-    aiter->SetFlags(flags, kArcFlags);  // Restores original flag values.
-    return reach_begin_ >= 0;
-  }
-
-  // Returns iterator position of first matching arc.
-  ssize_t ReachBegin() const { return reach_begin_; }
-
-  // Returns iterator position one past last matching arc.
-  ssize_t ReachEnd() const { return reach_end_; }
-
-  // Return the sum of the weights for matching arcs. Valid only if
-  // compute_weight was true in Reach() call.
-  Weight ReachWeight() const { return reach_weight_; }
-
-  // Access to the relabeling map. Excludes epsilon (0) label but
-  // includes kNoLabel that is used internally for super-final
-  // transitons.
-  const std::unordered_map<Label, Label> &Label2Index() const {
-    return *data_->Label2Index();
-  }
-
-  const Data *GetData() const { return data_.get(); }
-
-  std::shared_ptr<Data> GetSharedData() const { return data_; }
-
-  bool Error() const { return error_ || accumulator_->Error(); }
-
- private:
-  // Redirects labeled arcs (input or output labels determined by ReachInput())
-  // to new label-specific final states. Each original final state is
-  // redirected via a transition labeled with kNoLabel to a new
-  // kNoLabel-specific final state. Creates super-initial state for all states
-  // with zero in-degree.
-  void TransformFst() {
-    auto ins = fst_->NumStates();
-    auto ons = ins;
-    std::vector<ssize_t> indeg(ins, 0);
-    // Redirects labeled arcs to new final states.
-    for (StateId s = 0; s < ins; ++s) {
-      for (MutableArcIterator<VectorFst<Arc>> aiter(fst_.get(), s);
-           !aiter.Done(); aiter.Next()) {
-        auto arc = aiter.Value();
-        const auto label = data_->ReachInput() ? arc.ilabel : arc.olabel;
-        if (label) {
-          auto insert_result = label2state_.insert(std::make_pair(label, ons));
-          if (insert_result.second) {
-            indeg.push_back(0);
-            ++ons;
-          }
-          arc.nextstate = label2state_[label];
-          aiter.SetValue(arc);
-        }
-        ++indeg[arc.nextstate];  // Finds in-degrees for next step.
-      }
-      // Redirects final weights to new final state.
-      const auto final_weight = fst_->Final(s);
-      if (final_weight != Weight::Zero()) {
-        auto insert_result = label2state_.insert(std::make_pair(kNoLabel, ons));
-        if (insert_result.second) {
-          indeg.push_back(0);
-          ++ons;
-        }
-        Arc arc(kNoLabel, kNoLabel, final_weight, label2state_[kNoLabel]);
-        fst_->AddArc(s, arc);
-        ++indeg[arc.nextstate];  // Finds in-degrees for next step.
-        fst_->SetFinal(s, Weight::Zero());
-      }
-    }
-    // Adds new final states to the FST.
-    while (fst_->NumStates() < ons) {
-      StateId s = fst_->AddState();
-      fst_->SetFinal(s, Weight::One());
-    }
-    // Creates a super-initial state for all states with zero in-degree.
-    const auto start = fst_->AddState();
-    fst_->SetStart(start);
-    for (StateId s = 0; s < start; ++s) {
-      if (indeg[s] == 0) {
-        Arc arc(0, 0, Weight::One(), s);
-        fst_->AddArc(start, arc);
-      }
-    }
-  }
-
-  void FindIntervals(StateId ins) {
-    StateReachable<Arc, Label, LabelIntervalSet> state_reachable(*fst_);
-    if (state_reachable.Error()) {
-      error_ = true;
-      return;
-    }
-    auto &state2index = state_reachable.State2Index();
-    auto &interval_sets = *data_->MutableIntervalSets();
-    interval_sets = state_reachable.IntervalSets();
-    interval_sets.resize(ins);
-    auto &label2index = *data_->Label2Index();
-    for (const auto &kv : label2state_) {
-      Label i = state2index[kv.second];
-      label2index[kv.first] = i;
-      if (kv.first == kNoLabel) data_->SetFinalLabel(i);
-    }
-    label2state_.clear();
-    double nintervals = 0;
-    ssize_t non_intervals = 0;
-    for (StateId s = 0; s < ins; ++s) {
-      nintervals += interval_sets[s].Size();
-      if (interval_sets[s].Size() > 1) {
-        ++non_intervals;
-        VLOG(3) << "state: " << s
-                << " # of intervals: " << interval_sets[s].Size();
-      }
-    }
-    VLOG(2) << "# of states: " << ins;
-    VLOG(2) << "# of intervals: " << nintervals;
-    VLOG(2) << "# of intervals/state: " << nintervals / ins;
-    VLOG(2) << "# of non-interval states: " << non_intervals;
-  }
-
-  template <class Iterator>
-  ssize_t LowerBound(Iterator *aiter, ssize_t aiter_begin, ssize_t aiter_end,
-                     Label match_label) const {
-    // Only needs to compute the ilabel or olabel of arcs when performing the
-    // binary search.
-    aiter->SetFlags(reach_fst_input_ ? kArcILabelValue : kArcOLabelValue,
-                    kArcValueFlags);
-    ssize_t low = aiter_begin;
-    ssize_t high = aiter_end;
-    while (low < high) {
-      const ssize_t mid = low + (high - low) / 2;
-      aiter->Seek(mid);
-      auto label =
-          reach_fst_input_ ? aiter->Value().ilabel : aiter->Value().olabel;
-      if (label < match_label) {
-        low = mid + 1;
-      } else {
-        high = mid;
-      }
-    }
-    aiter->Seek(low);
-    aiter->SetFlags(kArcValueFlags, kArcValueFlags);
-    return low;
-  }
-
-  std::unique_ptr<VectorFst<Arc>> fst_;
-  // Current state
-  StateId s_;
-  // Finds final state for a label
-  std::unordered_map<Label, StateId> label2state_;
-  // Iterator position of first match.
-  ssize_t reach_begin_;
-  // Iterator position after last match.
-  ssize_t reach_end_;
-  // Gives weight sum of arc iterator arcs with reachable labels.
-  Weight reach_weight_;
-  // Shareable data between copies.
-  std::shared_ptr<Data> data_;
-  // Sums arc weights.
-  std::unique_ptr<Accumulator> accumulator_;
-  double ncalls_;
-  double nintervals_;
-  bool reach_fst_input_;
-  bool error_;
-};
-
-}  // namespace fst
-
-#endif  // FST_LABEL_REACHABLE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lexicographic-weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lexicographic-weight.h
deleted file mode 100644
index af3c6db6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lexicographic-weight.h
+++ /dev/null
@@ -1,173 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Lexicographic weight set and associated semiring operation definitions.
-//
-// A lexicographic weight is a sequence of weights, each of which must have the
-// path property and Times() must be (strongly) cancellative
-// (for all a,b,c != Zero(): Times(c, a) = Times(c, b) => a = b,
-// Times(a, c) = Times(b, c) => a = b).
-// The + operation on two weights a and b is the lexicographically
-// prior of a and b.
-
-#ifndef FST_LEXICOGRAPHIC_WEIGHT_H_
-#define FST_LEXICOGRAPHIC_WEIGHT_H_
-
-#include <cstdlib>
-
-#include <string>
-
-#include <fst/log.h>
-
-#include <fst/pair-weight.h>
-#include <fst/weight.h>
-
-
-namespace fst {
-
-template <class W1, class W2>
-class LexicographicWeight : public PairWeight<W1, W2> {
- public:
-  using ReverseWeight = LexicographicWeight<typename W1::ReverseWeight,
-                                            typename W2::ReverseWeight>;
-
-  using PairWeight<W1, W2>::Value1;
-  using PairWeight<W1, W2>::Value2;
-  using PairWeight<W1, W2>::SetValue1;
-  using PairWeight<W1, W2>::SetValue2;
-  using PairWeight<W1, W2>::Zero;
-  using PairWeight<W1, W2>::One;
-  using PairWeight<W1, W2>::NoWeight;
-  using PairWeight<W1, W2>::Quantize;
-  using PairWeight<W1, W2>::Reverse;
-
-  LexicographicWeight() {}
-
-  explicit LexicographicWeight(const PairWeight<W1, W2> &w)
-      : PairWeight<W1, W2>(w) {}
-
-  LexicographicWeight(W1 w1, W2 w2) : PairWeight<W1, W2>(w1, w2) {
-    if ((W1::Properties() & kPath) != kPath) {
-      FSTERROR() << "LexicographicWeight must "
-                 << "have the path property: " << W1::Type();
-      SetValue1(W1::NoWeight());
-    }
-    if ((W2::Properties() & kPath) != kPath) {
-      FSTERROR() << "LexicographicWeight must "
-                 << "have the path property: " << W2::Type();
-      SetValue2(W2::NoWeight());
-    }
-  }
-
-  static const LexicographicWeight &Zero() {
-    static const LexicographicWeight zero(PairWeight<W1, W2>::Zero());
-    return zero;
-  }
-
-  static const LexicographicWeight &One() {
-    static const LexicographicWeight one(PairWeight<W1, W2>::One());
-    return one;
-  }
-
-  static const LexicographicWeight &NoWeight() {
-    static const LexicographicWeight no_weight(PairWeight<W1, W2>::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type =
-        new string(W1::Type() + "_LT_" + W2::Type());
-    return *type;
-  }
-
-  bool Member() const {
-    if (!Value1().Member() || !Value2().Member()) return false;
-    // Lexicographic weights cannot mix zeroes and non-zeroes.
-    if (Value1() == W1::Zero() && Value2() == W2::Zero()) return true;
-    if (Value1() != W1::Zero() && Value2() != W2::Zero()) return true;
-    return false;
-  }
-
-  LexicographicWeight Quantize(float delta = kDelta) const {
-    return LexicographicWeight(PairWeight<W1, W2>::Quantize());
-  }
-
-  ReverseWeight Reverse() const {
-    return ReverseWeight(PairWeight<W1, W2>::Reverse());
-  }
-
-  static constexpr uint64 Properties() {
-    return W1::Properties() & W2::Properties() &
-           (kLeftSemiring | kRightSemiring | kPath | kIdempotent |
-            kCommutative);
-  }
-};
-
-template <class W1, class W2>
-inline LexicographicWeight<W1, W2> Plus(const LexicographicWeight<W1, W2> &w,
-                                        const LexicographicWeight<W1, W2> &v) {
-  if (!w.Member() || !v.Member()) {
-    return LexicographicWeight<W1, W2>::NoWeight();
-  }
-  NaturalLess<W1> less1;
-  NaturalLess<W2> less2;
-  if (less1(w.Value1(), v.Value1())) return w;
-  if (less1(v.Value1(), w.Value1())) return v;
-  if (less2(w.Value2(), v.Value2())) return w;
-  if (less2(v.Value2(), w.Value2())) return v;
-  return w;
-}
-
-template <class W1, class W2>
-inline LexicographicWeight<W1, W2> Times(const LexicographicWeight<W1, W2> &w,
-                                         const LexicographicWeight<W1, W2> &v) {
-  return LexicographicWeight<W1, W2>(Times(w.Value1(), v.Value1()),
-                                     Times(w.Value2(), v.Value2()));
-}
-
-template <class W1, class W2>
-inline LexicographicWeight<W1, W2> Divide(const LexicographicWeight<W1, W2> &w,
-                                          const LexicographicWeight<W1, W2> &v,
-                                          DivideType typ = DIVIDE_ANY) {
-  return LexicographicWeight<W1, W2>(Divide(w.Value1(), v.Value1(), typ),
-                                     Divide(w.Value2(), v.Value2(), typ));
-}
-
-// This function object generates weights by calling the underlying generators
-// for the templated weight types, like all other pair weight types. However,
-// for lexicographic weights, we cannot generate zeroes for the two subweights
-// separately: weights are members iff both members are zero or both members
-// are non-zero. This is intended primarily for testing.
-template <class W1, class W2>
-class WeightGenerate<LexicographicWeight<W1, W2>> {
- public:
-  using Weight = LexicographicWeight<W1, W1>;
-  using Generate1 = WeightGenerate<W1>;
-  using Generate2 = WeightGenerate<W2>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t num_random_weights = kNumRandomWeights)
-      : generator1_(false, num_random_weights),
-        generator2_(false, num_random_weights), allow_zero_(allow_zero),
-        num_random_weights_(num_random_weights) {}
-
-  Weight operator()() const {
-    if (allow_zero_) {
-      const int n = rand() % (num_random_weights_ + 1);  // NOLINT
-      if (n == num_random_weights_) return Weight(W1::Zero(), W2::Zero());
-    }
-    return Weight(generator1_(), generator2_());
-  }
-
- private:
-  const Generate1 generator1_;
-  const Generate2 generator2_;
-  // Permits Zero() and zero divisors.
-  const bool allow_zero_;
-  // The number of alternative random weights.
-  const size_t num_random_weights_;
-};
-
-}  // namespace fst
-
-#endif  // FST_LEXICOGRAPHIC_WEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lock.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lock.h
deleted file mode 100644
index 388d42f5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lock.h
+++ /dev/null
@@ -1,62 +0,0 @@
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Google-compatibility locking declarations and inline definitions.
-
-#ifndef FST_LIB_LOCK_H_
-#define FST_LIB_LOCK_H_
-
-#include <mutex>
-
-namespace fst {
-
-using namespace std;
-
-class Mutex {
- public:
-  Mutex() {}
-
-  inline void Lock() { mu_.lock(); }
-
-  inline void Unlock() { mu_.unlock(); }
-
- private:
-  std::mutex mu_;
-
-  Mutex(const Mutex &) = delete;
-  Mutex &operator=(const Mutex &) = delete;
-};
-
-class MutexLock {
- public:
-  explicit MutexLock(Mutex *mu) : mu_(mu) { mu_->Lock(); }
-
-  ~MutexLock() { mu_->Unlock(); }
-
- private:
-  Mutex *mu_;
-
-  MutexLock(const MutexLock &) = delete;
-  MutexLock &operator=(const MutexLock &) = delete;
-};
-
-// Currently, we don't use a separate reader lock.
-// TODO(kbg): Implement this with std::shared_mutex once C++17 becomes widely
-// available.
-using ReaderMutexLock = MutexLock;
-
-}  // namespace fst
-
-#endif  // FST_LIB_LOCK_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/log.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/log.h
deleted file mode 100644
index 0a8d6d73..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/log.h
+++ /dev/null
@@ -1,82 +0,0 @@
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Google-style logging declarations and inline definitions.
-
-#ifndef FST_LIB_LOG_H_
-#define FST_LIB_LOG_H_
-
-#include <cassert>
-#include <iostream>
-#include <string>
-
-#include <fst/types.h>
-#include <fst/flags.h>
-
-using std::string;
-
-DECLARE_int32(v);
-
-class LogMessage {
- public:
-  LogMessage(const string &type) : fatal_(type == "FATAL") {
-    std::cerr << type << ": ";
-  }
-  ~LogMessage() {
-    std::cerr << std::endl;
-    if(fatal_)
-      exit(1);
-  }
-  std::ostream &stream() { return std::cerr; }
-
- private:
-  bool fatal_;
-};
-
-#define LOG(type) LogMessage(#type).stream()
-#define VLOG(level) if ((level) <= FLAGS_v) LOG(INFO)
-
-// Checks
-inline void FstCheck(bool x, const char* expr,
-                const char *file, int line) {
-  if (!x) {
-    LOG(FATAL) << "Check failed: \"" << expr
-               << "\" file: " << file
-               << " line: " << line;
-  }
-}
-
-#define CHECK(x) FstCheck(static_cast<bool>(x), #x, __FILE__, __LINE__)
-#define CHECK_EQ(x, y) CHECK((x) == (y))
-#define CHECK_LT(x, y) CHECK((x) < (y))
-#define CHECK_GT(x, y) CHECK((x) > (y))
-#define CHECK_LE(x, y) CHECK((x) <= (y))
-#define CHECK_GE(x, y) CHECK((x) >= (y))
-#define CHECK_NE(x, y) CHECK((x) != (y))
-
-// Debug checks
-#define DCHECK(x) assert(x)
-#define DCHECK_EQ(x, y) DCHECK((x) == (y))
-#define DCHECK_LT(x, y) DCHECK((x) < (y))
-#define DCHECK_GT(x, y) DCHECK((x) > (y))
-#define DCHECK_LE(x, y) DCHECK((x) <= (y))
-#define DCHECK_GE(x, y) DCHECK((x) >= (y))
-#define DCHECK_NE(x, y) DCHECK((x) != (y))
-
-
-// Ports
-#define ATTRIBUTE_DEPRECATED __attribute__((deprecated))
-
-#endif  // FST_LIB_LOG_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lookahead-filter.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lookahead-filter.h
deleted file mode 100644
index df3fc6c9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lookahead-filter.h
+++ /dev/null
@@ -1,623 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Composition filters to support lookahead matchers, useful for improving
-// composition efficiency with certain inputs.
-
-#ifndef FST_LOOKAHEAD_FILTER_H_
-#define FST_LOOKAHEAD_FILTER_H_
-
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/filter-state.h>
-#include <fst/fst.h>
-#include <fst/lookahead-matcher.h>
-
-
-namespace fst {
-
-// Identifies and verifies the capabilities of the matcher to be used for
-// lookahead with the composition filters below. This version is passed two
-// matchers.
-template <class Matcher1, class Matcher2>
-MatchType LookAheadMatchType(const Matcher1 &m1, const Matcher2 &m2) {
-  const auto type1 = m1.Type(false);
-  const auto type2 = m2.Type(false);
-  if (type1 == MATCH_OUTPUT && m1.Flags() & kOutputLookAheadMatcher) {
-    return MATCH_OUTPUT;
-  } else if (type2 == MATCH_INPUT && m2.Flags() & kInputLookAheadMatcher) {
-    return MATCH_INPUT;
-  } else if (m1.Flags() & kOutputLookAheadMatcher &&
-             m1.Type(true) == MATCH_OUTPUT) {
-    return MATCH_OUTPUT;
-  } else if (m2.Flags() & kInputLookAheadMatcher &&
-             m2.Type(true) == MATCH_INPUT) {
-    return MATCH_INPUT;
-  } else {
-    return MATCH_NONE;
-  }
-}
-
-// Identifies and verifies the capabilities of the matcher to be used for
-// lookahead with the composition filters below. This version uses the FST's
-// default matchers.
-template <class Arc>
-MatchType LookAheadMatchType(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {
-  LookAheadMatcher<Fst<Arc>> matcher1(fst1, MATCH_OUTPUT);
-  LookAheadMatcher<Fst<Arc>> matcher2(fst2, MATCH_INPUT);
-  return LookAheadMatchType(matcher1, matcher2);
-}
-
-// LookAheadSelector is a helper class for selecting among possibly distinct
-// FST and matcher types without using a common base class. This lets us avoid
-// virtual function calls. It stores and returns the appropriate FSTs and
-// matcher for lookahead. It is templated on the matcher types. General case
-// has no methods.
-template <class Matcher1, class Matcher2, MatchType MT>
-class LookAheadSelector {};
-
-// Stores and returns the appropriate FST and matcher for lookahead. Specialized
-// for two matchers of same type with the (match) type argument determining
-// which is used for lookahead.
-template <class Matcher, MatchType MT>
-class LookAheadSelector<Matcher, Matcher, MT> {
- public:
-  using FST = typename Matcher::FST;
-
-  LookAheadSelector(Matcher *lmatcher1, Matcher *lmatcher2, MatchType type)
-      : lmatcher1_(lmatcher1->Copy()),
-        lmatcher2_(lmatcher2->Copy()),
-        type_(type) {}
-
-  LookAheadSelector(const LookAheadSelector<Matcher, Matcher, MT> &selector)
-      : lmatcher1_(selector.lmatcher1_->Copy()),
-        lmatcher2_(selector.lmatcher2_->Copy()),
-        type_(selector.type_) {}
-
-  const FST &GetFst() const {
-    return type_ == MATCH_OUTPUT ? lmatcher2_->GetFst() : lmatcher1_->GetFst();
-  }
-
-  Matcher *GetMatcher() const {
-    return type_ == MATCH_OUTPUT ? lmatcher1_.get() : lmatcher2_.get();
-  }
-
- private:
-  std::unique_ptr<Matcher> lmatcher1_;
-  std::unique_ptr<Matcher> lmatcher2_;
-  MatchType type_;
-};
-
-// Stores and returns the appropriate FST and matcher for lookahead.
-// Specialized for lookahead on input labels.
-template <class Matcher1, class Matcher2>
-class LookAheadSelector<Matcher1, Matcher2, MATCH_INPUT> {
- public:
-  using FST1 = typename Matcher1::FST;
-
-  LookAheadSelector(Matcher1 *lmatcher1, Matcher2 *lmatcher2, MatchType)
-      : fst_(lmatcher1->GetFst().Copy()), lmatcher_(lmatcher2->Copy()) {}
-
-  LookAheadSelector(
-      const LookAheadSelector<Matcher1, Matcher2, MATCH_INPUT> &selector)
-      : fst_(selector.fst_->Copy()), lmatcher_(selector.lmatcher_->Copy()) {}
-
-  const FST1 &GetFst() const { return *fst_; }
-
-  Matcher2 *GetMatcher() const { return lmatcher_.get(); }
-
- private:
-  std::unique_ptr<const FST1> fst_;
-  std::unique_ptr<Matcher2> lmatcher_;
-};
-
-// Stores and returns the appropriate FST and matcher for lookahead.
-// Specialized for lookahead on output labels.
-template <class Matcher1, class Matcher2>
-class LookAheadSelector<Matcher1, Matcher2, MATCH_OUTPUT> {
- public:
-  using FST2 = typename Matcher2::FST;
-
-  LookAheadSelector(Matcher1 *lmatcher1, Matcher2 *lmatcher2, MatchType)
-      : fst_(lmatcher2->GetFst().Copy()), lmatcher_(lmatcher1->Copy()) {}
-
-  LookAheadSelector(
-      const LookAheadSelector<Matcher1, Matcher2, MATCH_OUTPUT> &selector)
-      : fst_(selector.fst_->Copy()), lmatcher_(selector.lmatcher_->Copy()) {}
-
-  const FST2 &GetFst() const { return *fst_; }
-
-  Matcher1 *GetMatcher() const { return lmatcher_.get(); }
-
- private:
-  std::unique_ptr<const FST2> fst_;
-  std::unique_ptr<Matcher1> lmatcher_;
-};
-
-// This filter uses a lookahead matcher in FilterArc(arc1, arc2) to examine the
-// future of the composition state (arc1.nextstate, arc2.nextstate), blocking
-// moving forward when its determined to be
-// non-coaccessible. It is templated on an underlying filter, typically the
-// epsilon filter. Which matcher is the lookahead matcher is determined by the
-// template argument MT unless it is MATCH_BOTH. In that case, both matcher
-// arguments must be lookahead matchers of the same type and one will be
-// selected by LookAheadMatchType() based on their capability.
-template <class Filter, class M1 = LookAheadMatcher<typename Filter::FST1>,
-          class M2 = M1, MatchType MT = MATCH_BOTH>
-class LookAheadComposeFilter {
- public:
-  using Arc = typename Filter::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FST1 = typename Filter::FST1;
-  using FST2 = typename Filter::FST2;
-  using Matcher1 = typename Filter::Matcher1;
-  using Matcher2 = typename Filter::Matcher2;
-  using FilterState = typename Filter::FilterState;
-
-  LookAheadComposeFilter(const FST1 &fst1, const FST2 &fst2, M1 *matcher1,
-                         M2 *matcher2)
-      : filter_(fst1, fst2, matcher1, matcher2),
-        lookahead_type_(MT == MATCH_BOTH
-                            ? LookAheadMatchType(*filter_.GetMatcher1(),
-                                                 *filter_.GetMatcher2())
-                            : MT),
-        selector_(filter_.GetMatcher1(), filter_.GetMatcher2(),
-                  lookahead_type_),
-        flags_(lookahead_type_ == MATCH_OUTPUT
-                   ? filter_.GetMatcher1()->Flags()
-                   : filter_.GetMatcher2()->Flags()) {
-    if (lookahead_type_ == MATCH_NONE) {
-      FSTERROR() << "LookAheadComposeFilter: 1st argument cannot "
-                 << "match/look-ahead on output labels and 2nd argument "
-                 << "cannot match/look-ahead on input labels";
-    }
-    selector_.GetMatcher()->InitLookAheadFst(selector_.GetFst());
-  }
-
-  LookAheadComposeFilter(
-      const LookAheadComposeFilter<Filter, M1, M2, MT> &filter,
-      bool safe = false)
-      : filter_(filter.filter_, safe),
-        lookahead_type_(filter.lookahead_type_),
-        selector_(filter_.GetMatcher1(), filter_.GetMatcher2(),
-                  lookahead_type_),
-        flags_(filter.flags_) {
-    selector_.GetMatcher()->InitLookAheadFst(selector_.GetFst(), true);
-  }
-
-  FilterState Start() const { return filter_.Start(); }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    filter_.SetState(s1, s2, fs);
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    lookahead_arc_ = false;
-    const FilterState &fs = filter_.FilterArc(arc1, arc2);
-    if (fs == FilterState::NoState()) return FilterState::NoState();
-    return LookAheadOutput() ? LookAheadFilterArc(arc1, arc2, fs)
-                             : LookAheadFilterArc(arc2, arc1, fs);
-  }
-
-  void FilterFinal(Weight *weight1, Weight *weight2) const {
-    filter_.FilterFinal(weight1, weight2);
-  }
-
-  // Returns matchers; ownership stays with filter.
-
-  Matcher1 *GetMatcher1() { return filter_.GetMatcher1(); }
-
-  Matcher2 *GetMatcher2() { return filter_.GetMatcher2(); }
-
-  const LookAheadSelector<Matcher1, Matcher2, MT> &Selector() const {
-    return selector_;
-  }
-
-  uint64 Properties(uint64 inprops) const {
-    auto outprops = filter_.Properties(inprops);
-    if (lookahead_type_ == MATCH_NONE) outprops |= kError;
-    return outprops;
-  }
-
-  uint32 LookAheadFlags() const { return flags_; }
-
-  bool LookAheadArc() const { return lookahead_arc_; }
-
-  bool LookAheadOutput() const {
-    if (MT == MATCH_OUTPUT) {
-      return true;
-    } else if (MT == MATCH_INPUT) {
-      return false;
-    } else if (lookahead_type_ == MATCH_OUTPUT) {
-      return true;
-    } else {
-      return false;
-    }
-  }
-
- private:
-  FilterState LookAheadFilterArc(Arc *arca, Arc *arcb,
-                                 const FilterState &fs) const {
-    auto &labela = LookAheadOutput() ? arca->olabel : arca->ilabel;
-    if (labela != 0 && !(flags_ & kLookAheadNonEpsilons)) return fs;
-    if (labela == 0 && !(flags_ & kLookAheadEpsilons)) return fs;
-    lookahead_arc_ = true;
-    selector_.GetMatcher()->SetState(arca->nextstate);
-    return selector_.GetMatcher()->LookAheadFst(selector_.GetFst(),
-                                                arcb->nextstate)
-               ? fs
-               : FilterState::NoState();
-  }
-
-  Filter filter_;             // Underlying filter.
-  MatchType lookahead_type_;  // Lookahead match type.
-  LookAheadSelector<Matcher1, Matcher2, MT> selector_;
-  uint32 flags_;                // Lookahead flags.
-  mutable bool lookahead_arc_;  // Look-ahead performed at last FilterArc()?
-
-  LookAheadComposeFilter &operator=(const LookAheadComposeFilter &) = delete;
-};
-
-// This filter adds weight-pushing to a lookahead composition filter using the
-// LookAheadWeight() method of matcher argument. It is templated on an
-// underlying lookahead filter, typically the basic lookahead filter.
-// Weight-pushing in composition brings weights forward as much as possible
-// based on the lookahead information.
-template <class Filter, class M1 = LookAheadMatcher<typename Filter::FST1>,
-          class M2 = M1, MatchType MT = MATCH_BOTH>
-class PushWeightsComposeFilter {
- public:
-  using Arc = typename Filter::Arc;
-  using StateId = typename Filter::StateId;
-  using Weight = typename Filter::Weight;
-
-  using FST1 = typename Filter::FST1;
-  using FST2 = typename Filter::FST2;
-  using Matcher1 = typename Filter::Matcher1;
-  using Matcher2 = typename Filter::Matcher2;
-
-  using FilterState1 = typename Filter::FilterState;
-  using FilterState2 = WeightFilterState<Weight>;
-  using FilterState = PairFilterState<FilterState1, FilterState2>;
-
-  PushWeightsComposeFilter(const FST1 &fst1, const FST2 &fst2, M1 *matcher1,
-                           M2 *matcher2)
-      : filter_(fst1, fst2, matcher1, matcher2), fs_(FilterState::NoState()) {}
-
-  PushWeightsComposeFilter(
-      const PushWeightsComposeFilter<Filter, M1, M2, MT> &filter,
-      bool safe = false)
-      : filter_(filter.filter_, safe), fs_(FilterState::NoState()) {}
-
-  FilterState Start() const {
-    return FilterState(filter_.Start(), FilterState2(Weight::One()));
-  }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    fs_ = fs;
-    filter_.SetState(s1, s2, fs.GetState1());
-  }
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    const auto &fs1 = filter_.FilterArc(arc1, arc2);
-    if (fs1 == FilterState1::NoState()) return FilterState::NoState();
-    if (!(LookAheadFlags() & kLookAheadWeight)) {
-      return FilterState(fs1, FilterState2(Weight::One()));
-    }
-    const auto &lweight = filter_.LookAheadArc()
-                              ? Selector().GetMatcher()->LookAheadWeight()
-                              : Weight::One();
-    const auto &fs2 = fs_.GetState2();
-    const auto &fweight = fs2.GetWeight();
-    // Disallows Zero() weight futures.
-    if (lweight == Weight::Zero()) return FilterState::NoState();
-    arc2->weight = Divide(Times(arc2->weight, lweight), fweight);
-    return FilterState(fs1, FilterState2(lweight.Quantize()));
-  }
-
-  void FilterFinal(Weight *weight1, Weight *weight2) const {
-    filter_.FilterFinal(weight1, weight2);
-    if (!(LookAheadFlags() & kLookAheadWeight) || *weight1 == Weight::Zero()) {
-      return;
-    }
-    const auto &fs2 = fs_.GetState2();
-    const auto &fweight = fs2.GetWeight();
-    *weight1 = Divide(*weight1, fweight);
-  }
-
-  // Returns matchers; ownership states with filter.
-
-  Matcher1 *GetMatcher1() { return filter_.GetMatcher1(); }
-
-  Matcher2 *GetMatcher2() { return filter_.GetMatcher2(); }
-
-  const LookAheadSelector<Matcher1, Matcher2, MT> &Selector() const {
-    return filter_.Selector();
-  }
-
-  uint32 LookAheadFlags() const { return filter_.LookAheadFlags(); }
-
-  bool LookAheadArc() const { return filter_.LookAheadArc(); }
-
-  bool LookAheadOutput() const { return filter_.LookAheadOutput(); }
-
-  uint64 Properties(uint64 props) const {
-    return filter_.Properties(props) & kWeightInvariantProperties;
-  }
-
- private:
-  Filter filter_;   // Underlying filter.
-  FilterState fs_;  // Current filter state.
-
-  PushWeightsComposeFilter &operator=(const PushWeightsComposeFilter &) =
-      delete;
-};
-
-// This filter adds label-pushing to a lookahead composition filter using the
-// LookAheadPrefix() method of the matcher argument. It is templated on an
-// underlying filter, typically the basic lookahead or weight-pushing lookahead
-// filter. Label-pushing in composition matches labels as early as possible
-// based on the lookahead information.
-template <class Filter, class M1 = LookAheadMatcher<typename Filter::FST1>,
-          class M2 = M1, MatchType MT = MATCH_BOTH>
-class PushLabelsComposeFilter {
- public:
-  using Arc = typename Filter::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FST1 = typename Filter::FST1;
-  using FST2 = typename Filter::FST2;
-  using Matcher1 = MultiEpsMatcher<typename Filter::Matcher1>;
-  using Matcher2 = MultiEpsMatcher<typename Filter::Matcher2>;
-  using FilterState1 = typename Filter::FilterState;
-  using FilterState2 = IntegerFilterState<Label>;
-  using FilterState = PairFilterState<FilterState1, FilterState2>;
-
-  PushLabelsComposeFilter(const FST1 &fst1, const FST2 &fst2, M1 *matcher1,
-                          M2 *matcher2)
-      : filter_(fst1, fst2, matcher1, matcher2),
-        fs_(FilterState::NoState()),
-        fst1_(filter_.GetMatcher1()->GetFst()),
-        fst2_(filter_.GetMatcher2()->GetFst()),
-        matcher1_(fst1_, MATCH_OUTPUT,
-                  filter_.LookAheadOutput() ? kMultiEpsList : kMultiEpsLoop,
-                  filter_.GetMatcher1(), false),
-        matcher2_(fst2_, MATCH_INPUT,
-                  filter_.LookAheadOutput() ? kMultiEpsLoop : kMultiEpsList,
-                  filter_.GetMatcher2(), false) {}
-
-  PushLabelsComposeFilter(
-      const PushLabelsComposeFilter<Filter, M1, M2, MT> &filter,
-      bool safe = false)
-      : filter_(filter.filter_, safe),
-        fs_(FilterState::NoState()),
-        fst1_(filter_.GetMatcher1()->GetFst()),
-        fst2_(filter_.GetMatcher2()->GetFst()),
-        matcher1_(fst1_, MATCH_OUTPUT,
-                  filter_.LookAheadOutput() ? kMultiEpsList : kMultiEpsLoop,
-                  filter_.GetMatcher1(), false),
-        matcher2_(fst2_, MATCH_INPUT,
-                  filter_.LookAheadOutput() ? kMultiEpsLoop : kMultiEpsList,
-                  filter_.GetMatcher2(), false) {}
-
-  FilterState Start() const {
-    return FilterState(filter_.Start(), FilterState2(kNoLabel));
-  }
-
-  void SetState(StateId s1, StateId s2, const FilterState &fs) {
-    fs_ = fs;
-    filter_.SetState(s1, s2, fs.GetState1());
-    if (!(LookAheadFlags() & kLookAheadPrefix)) return;
-    narcsa_ = LookAheadOutput() ? internal::NumArcs(fst1_, s1)
-                                : internal::NumArcs(fst2_, s2);
-    const auto &fs2 = fs_.GetState2();
-    const auto &flabel = fs2.GetState();
-    GetMatcher1()->ClearMultiEpsLabels();
-    GetMatcher2()->ClearMultiEpsLabels();
-    if (flabel != kNoLabel) {                   // Have a lookahead label?
-      GetMatcher1()->AddMultiEpsLabel(flabel);  // Yes, make it a multi-epsilon
-      GetMatcher2()->AddMultiEpsLabel(flabel);  // label so that it matches the
-    }                                           // implicit epsilon arc to be
-  }                                             // modified below when pushing.
-
-  FilterState FilterArc(Arc *arc1, Arc *arc2) const {
-    if (!(LookAheadFlags() & kLookAheadPrefix)) {
-      return FilterState(filter_.FilterArc(arc1, arc2), FilterState2(kNoLabel));
-    }
-    const auto &fs2 = fs_.GetState2();
-    const auto &flabel = fs2.GetState();
-    if (flabel != kNoLabel) {  // Have a lookahead label?
-      return LookAheadOutput() ? PushedLabelFilterArc(arc1, arc2, flabel)
-                               : PushedLabelFilterArc(arc2, arc1, flabel);
-    }
-    const auto &fs1 = filter_.FilterArc(arc1, arc2);
-    if (fs1 == FilterState1::NoState()) return FilterState::NoState();
-    if (!filter_.LookAheadArc())
-      return FilterState(fs1, FilterState2(kNoLabel));
-    return LookAheadOutput() ? PushLabelFilterArc(arc1, arc2, fs1)
-                             : PushLabelFilterArc(arc2, arc1, fs1);
-  }
-
-  void FilterFinal(Weight *weight1, Weight *weight2) const {
-    filter_.FilterFinal(weight1, weight2);
-    if (!(LookAheadFlags() & kLookAheadPrefix) || *weight1 == Weight::Zero()) {
-      return;
-    }
-    const auto &fs2 = fs_.GetState2();
-    const auto &flabel = fs2.GetState();
-    if (flabel != kNoLabel) *weight1 = Weight::Zero();
-  }
-
-  // Returns matchers; ownership states with filter.
-
-  Matcher1 *GetMatcher1() { return &matcher1_; }
-
-  Matcher2 *GetMatcher2() { return &matcher2_; }
-
-  uint64 Properties(uint64 iprops) const {
-    const auto oprops = filter_.Properties(iprops);
-    if (LookAheadOutput()) {
-      return oprops & kOLabelInvariantProperties;
-    } else {
-      return oprops & kILabelInvariantProperties;
-    }
-  }
-
- private:
-  const LookAheadSelector<typename Filter::Matcher1, typename Filter::Matcher2,
-                          MT>
-      &Selector() const {
-    return filter_.Selector();
-  }
-
-  // Consumes an already pushed label.
-  FilterState PushedLabelFilterArc(Arc *arca, Arc *arcb, Label flabel) const {
-    auto &labela = LookAheadOutput() ? arca->olabel : arca->ilabel;
-    const auto &labelb = LookAheadOutput() ? arcb->ilabel : arcb->olabel;
-    if (labelb != kNoLabel) {
-      return FilterState::NoState();  // Blocks non-(multi-)epsilon label
-    } else if (labela == flabel) {
-      labela = 0;  // Converts match to multi-epsilon to epsilon.
-      return Start();
-    } else if (labela == 0) {
-      if (narcsa_ == 1) return fs_;  // Takes epsilon, keeping state with label.
-      Selector().GetMatcher()->SetState(arca->nextstate);
-      if (Selector().GetMatcher()->LookAheadLabel(flabel)) {
-        return fs_;  // Takes epsilon, keeping state with label.
-      } else {
-        return FilterState::NoState();  // Blocks non-coaccessible path.
-      }
-    } else {
-      return FilterState::NoState();  // Blocks mismatch to multi-epsilon label.
-    }
-  }
-
-  // Pushes a label forward when possible.
-  FilterState PushLabelFilterArc(Arc *arca, Arc *arcb,
-                                 const FilterState1 &fs1) const {
-    auto &labela = LookAheadOutput() ? arca->olabel : arca->ilabel;
-    const auto &labelb = LookAheadOutput() ? arcb->olabel : arcb->ilabel;
-    if (labelb != 0) {  // No place to push.
-      return FilterState(fs1, FilterState2(kNoLabel));
-    }
-    if (labela != 0 &&  // Wrong lookahead prefix type?
-        LookAheadFlags() & kLookAheadNonEpsilonPrefix) {
-      return FilterState(fs1, FilterState2(kNoLabel));
-    }
-    Arc larc(kNoLabel, kNoLabel, Weight::Zero(), kNoStateId);
-    if (Selector().GetMatcher()->LookAheadPrefix(&larc)) {  // Have prefix arc?
-      labela = LookAheadOutput() ? larc.ilabel : larc.olabel;
-      arcb->ilabel = larc.ilabel;  // Goes forward on that arc,
-      arcb->olabel = larc.olabel;  // thus pushing the label.
-      arcb->weight = Times(arcb->weight, larc.weight);
-      arcb->nextstate = larc.nextstate;
-      return FilterState(fs1, FilterState2(labela));
-    } else {
-      return FilterState(fs1, FilterState2(kNoLabel));
-    }
-  }
-
-  uint32 LookAheadFlags() const { return filter_.LookAheadFlags(); }
-
-  bool LookAheadArc() const { return filter_.LookAheadArc(); }
-
-  bool LookAheadOutput() const { return filter_.LookAheadOutput(); }
-
-  Filter filter_;   // Underlying filter.
-  FilterState fs_;  // Current filter state.
-  const FST1 &fst1_;
-  const FST2 &fst2_;
-  Matcher1 matcher1_;  // Multi-epsilon matcher for fst1_.
-  Matcher2 matcher2_;  // Multi-epsilon matcher for fst2_.
-  ssize_t narcsa_;     // Number of arcs leaving look-ahead match FST.
-
-  PushLabelsComposeFilter &operator=(const PushLabelsComposeFilter &) = delete;
-};
-
-// Convenience class for setting up composition with a default lookahead matcher
-// and filter.
-template <class Arc, MatchType type>
-class DefaultLookAhead {
- public:
-  using M = Matcher<Fst<Arc>>;
-  using ComposeFilter = SequenceComposeFilter<M>;
-  using FstMatcher = M;
-};
-
-// Specializes for MATCH_INPUT to allow lookahead.
-template <class Arc>
-class DefaultLookAhead<Arc, MATCH_INPUT> {
- public:
-  using M = LookAheadMatcher<Fst<Arc>>;
-  using SF = SequenceComposeFilter<M>;
-  using ComposeFilter = LookAheadComposeFilter<SF, M>;
-  using FstMatcher = M;
-};
-
-// Specializes for MATCH_OUTPUT to allow lookahead.
-template <class Arc>
-class DefaultLookAhead<Arc, MATCH_OUTPUT> {
- public:
-  using M = LookAheadMatcher<Fst<Arc>>;
-  using SF = AltSequenceComposeFilter<M>;
-  using ComposeFilter = LookAheadComposeFilter<SF, M>;
-  using FstMatcher = M;
-};
-
-// Specializes for StdArc to allow weight and label pushing.
-template <>
-class DefaultLookAhead<StdArc, MATCH_INPUT> {
- public:
-  using M = LookAheadMatcher<Fst<StdArc>>;
-  using SF = SequenceComposeFilter<M>;
-  using LF = LookAheadComposeFilter<SF, M>;
-  using WF = PushWeightsComposeFilter<LF, M>;
-  using ComposeFilter = PushLabelsComposeFilter<WF, M>;
-  using FstMatcher = M;
-};
-
-// Specializes for StdArc to allow weight and label pushing.
-template <>
-class DefaultLookAhead<StdArc, MATCH_OUTPUT> {
- public:
-  using M = LookAheadMatcher<Fst<StdArc>>;
-  using SF = AltSequenceComposeFilter<M>;
-  using LF = LookAheadComposeFilter<SF, M>;
-  using WF = PushWeightsComposeFilter<LF, M>;
-  using ComposeFilter = PushLabelsComposeFilter<WF, M>;
-  using FstMatcher = M;
-};
-
-// Specializes for LogArc to allow weight and label pushing.
-template <>
-class DefaultLookAhead<LogArc, MATCH_INPUT> {
- public:
-  using M = LookAheadMatcher<Fst<LogArc>>;
-  using SF = SequenceComposeFilter<M>;
-  using LF = LookAheadComposeFilter<SF, M>;
-  using WF = PushWeightsComposeFilter<LF, M>;
-  using ComposeFilter = PushLabelsComposeFilter<WF, M>;
-  using FstMatcher = M;
-};
-
-// Specializes for LogArc to allow weight and label pushing.
-template <>
-class DefaultLookAhead<LogArc, MATCH_OUTPUT> {
- public:
-  using M = LookAheadMatcher<Fst<LogArc>>;
-  using SF = AltSequenceComposeFilter<M>;
-  using LF = LookAheadComposeFilter<SF, M>;
-  using WF = PushWeightsComposeFilter<LF, M>;
-  using ComposeFilter = PushLabelsComposeFilter<WF, M>;
-  using FstMatcher = M;
-};
-
-}  // namespace fst
-
-#endif  // FST_LOOKAHEAD_FILTER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lookahead-matcher.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lookahead-matcher.h
deleted file mode 100644
index b95fbb23..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/lookahead-matcher.h
+++ /dev/null
@@ -1,786 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes to add lookahead to FST matchers, useful for improving composition
-// efficiency with certain inputs.
-
-#ifndef FST_LOOKAHEAD_MATCHER_H_
-#define FST_LOOKAHEAD_MATCHER_H_
-
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include <fst/flags.h>
-#include <fst/log.h>
-
-#include <fst/add-on.h>
-#include <fst/const-fst.h>
-#include <fst/fst.h>
-#include <fst/label-reachable.h>
-#include <fst/matcher.h>
-
-
-DECLARE_string(save_relabel_ipairs);
-DECLARE_string(save_relabel_opairs);
-
-namespace fst {
-
-// Lookahead matches extend the matcher interface with following additional
-// methods:
-//
-// template <class FST>
-// class LookAheadMatcher {
-//  public:
-//   using Arc = typename FST::Arc;
-//   using Label = typename Arc::Label;
-//   using StateId = typename Arc::StateId;
-//   using Weight = typename Arc::Weight;
-//
-//  // Required constructors.
-//  LookAheadMatcher(const FST &fst, MatchType match_type);
-//   // If safe=true, the copy is thread-safe (except the lookahead FST is
-//   // preserved). See Fst<>::Cop() for further doc.
-//  LookAheadMatcher(const LookAheadMatcher &matcher, bool safe = false);
-//
-//  // Below are methods for looking ahead for a match to a label and more
-//  // generally, to a rational set. Each returns false if there is definitely
-//  // not a match and returns true if there possibly is a match.
-//
-//  // Optionally pre-specifies the lookahead FST that will be passed to
-//  // LookAheadFst() for possible precomputation. If copy is true, then the FST
-//  // argument is a copy of the FST used in the previous call to this method
-//  // (to avoid unnecessary updates).
-//  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) override;
-//
-//  // Are there paths from a state in the lookahead FST that can be read from
-//  // the curent matcher state?
-//  bool LookAheadFst(const Fst<Arc> &fst, StateId s) override;
-//
-//  // Can the label be read from the current matcher state after possibly
-//  // following epsilon transitions?
-//  bool LookAheadLabel(Label label) const override;
-//
-//  // The following methods allow looking ahead for an arbitrary rational set
-//  // of strings, specified by an FST and a state from which to begin the
-//  // matching. If the lookahead FST is a transducer, this looks on the side
-//  // different from the matcher's match_type (cf. composition).
-//  // Is there is a single non-epsilon arc found in the lookahead FST that
-//  // begins the path (after possibly following any epsilons) in the last call
-//  // to LookAheadFst? If so, return true and copy it to the arc argument;
-//  // otherwise, return false. Non-trivial implementations are useful for
-//  // label-pushing in composition.
-//  bool LookAheadPrefix(Arc *arc) override;
-//
-//  // Gives an estimate of the combined weight of the paths in the lookahead
-//  // and matcher FSTs for the last call to LookAheadFst. Non-trivial
-//  // implementations are useful for weight-pushing in composition.
-//  Weight LookAheadWeight() const override;
-// };
-
-// Look-ahead flags.
-// Matcher is a lookahead matcher when match_type is MATCH_INPUT.
-constexpr uint32 kInputLookAheadMatcher = 0x00000010;
-
-// Matcher is a lookahead matcher when match_type is MATCH_OUTPUT.
-constexpr uint32 kOutputLookAheadMatcher = 0x00000020;
-
-// Is a non-trivial implementation of LookAheadWeight() method defined and
-// if so, should it be used?
-constexpr uint32 kLookAheadWeight = 0x00000040;
-
-// Is a non-trivial implementation of LookAheadPrefix() method defined and
-// if so, should it be used?
-constexpr uint32 kLookAheadPrefix = 0x00000080;
-
-// Look-ahead of matcher FST non-epsilon arcs?
-constexpr uint32 kLookAheadNonEpsilons = 0x00000100;
-
-// Look-ahead of matcher FST epsilon arcs?
-constexpr uint32 kLookAheadEpsilons = 0x00000200;
-
-// Ignore epsilon paths for the lookahead prefix? This gives correct results in
-// composition only with an appropriate composition filter since it depends on
-// the filter blocking the ignored paths.
-constexpr uint32 kLookAheadNonEpsilonPrefix = 0x00000400;
-
-// For LabelLookAheadMatcher, save relabeling data to file?
-constexpr uint32 kLookAheadKeepRelabelData = 0x00000800;
-
-// Flags used for lookahead matchers.
-constexpr uint32 kLookAheadFlags = 0x00000ff0;
-
-// LookAhead Matcher interface, templated on the Arc definition; used
-// for lookahead matcher specializations that are returned by the
-// InitMatcher() Fst method.
-template <class Arc>
-class LookAheadMatcherBase : public MatcherBase<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  virtual void InitLookAheadFst(const Fst<Arc> &, bool copy = false) = 0;
-  virtual bool LookAheadFst(const Fst<Arc> &, StateId) = 0;
-  virtual bool LookAheadLabel(Label) const = 0;
-
-  // Suggested concrete implementation of lookahead methods.
-
-  bool LookAheadPrefix(Arc *arc) const {
-    if (prefix_arc_.nextstate != kNoStateId) {
-      *arc = prefix_arc_;
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  Weight LookAheadWeight() const { return weight_; }
-
- protected:
-  // Concrete implementations for lookahead helper methods.
-
-  void ClearLookAheadWeight() { weight_ = Weight::One(); }
-
-  void SetLookAheadWeight(Weight weight) { weight_ = std::move(weight); }
-
-  void ClearLookAheadPrefix() { prefix_arc_.nextstate = kNoStateId; }
-
-  void SetLookAheadPrefix(Arc arc) { prefix_arc_ = std::move(arc); }
-
- private:
-  Arc prefix_arc_;
-  Weight weight_;
-};
-
-// Doesn't actually lookahead, just declares that the future looks good.
-template <class M>
-class TrivialLookAheadMatcher
-    : public LookAheadMatcherBase<typename M::FST::Arc> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  TrivialLookAheadMatcher(const FST &fst, MatchType match_type)
-      : matcher_(fst, match_type) {}
-
-  TrivialLookAheadMatcher(const TrivialLookAheadMatcher<M> &lmatcher,
-                          bool safe = false)
-      : matcher_(lmatcher.matcher_, safe) {}
-
-  TrivialLookAheadMatcher<M> *Copy(bool safe = false) const override {
-    return new TrivialLookAheadMatcher<M>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override { return matcher_.Type(test); }
-
-  void SetState(StateId s) final { return matcher_.SetState(s); }
-
-  bool Find(Label label) final { return matcher_.Find(label); }
-
-  bool Done() const final { return matcher_.Done(); }
-
-  const Arc &Value() const final { return matcher_.Value(); }
-
-  void Next() final { matcher_.Next(); }
-
-  Weight Final(StateId s) const final { return matcher_.Final(s); }
-
-  ssize_t Priority(StateId s) final { return matcher_.Priority(s); }
-
-  const FST &GetFst() const override { return matcher_.GetFst(); }
-
-  uint64 Properties(uint64 props) const override {
-    return matcher_.Properties(props);
-  }
-
-  uint32 Flags() const override {
-    return matcher_.Flags() | kInputLookAheadMatcher | kOutputLookAheadMatcher;
-  }
-
-  // Lookahead methods (all trivial).
-
-  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) override {}
-
-  bool LookAheadFst(const Fst<Arc> &, StateId) final { return true; }
-
-  bool LookAheadLabel(Label) const final { return true; }
-
-  bool LookAheadPrefix(Arc *) const { return false; }
-
-  Weight LookAheadWeight() const { return Weight::One(); }
-
- private:
-  M matcher_;
-};
-
-// Look-ahead of one transition. Template argument flags accepts flags to
-// control behavior.
-template <class M,
-          uint32 flags = kLookAheadNonEpsilons | kLookAheadEpsilons |
-                         kLookAheadWeight | kLookAheadPrefix>
-class ArcLookAheadMatcher : public LookAheadMatcherBase<typename M::FST::Arc> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using MatcherData = NullAddOn;
-
-  using LookAheadMatcherBase<Arc>::ClearLookAheadWeight;
-  using LookAheadMatcherBase<Arc>::LookAheadWeight;
-  using LookAheadMatcherBase<Arc>::SetLookAheadWeight;
-  using LookAheadMatcherBase<Arc>::ClearLookAheadPrefix;
-  using LookAheadMatcherBase<Arc>::LookAheadPrefix;
-  using LookAheadMatcherBase<Arc>::SetLookAheadPrefix;
-
-  enum : uint32 { kFlags = flags };
-
-  ArcLookAheadMatcher(
-      const FST &fst, MatchType match_type,
-      std::shared_ptr<MatcherData> data = std::shared_ptr<MatcherData>())
-      : matcher_(fst, match_type),
-        fst_(matcher_.GetFst()),
-        lfst_(nullptr),
-        state_(kNoStateId) {}
-
-  ArcLookAheadMatcher(const ArcLookAheadMatcher<M, flags> &lmatcher,
-                      bool safe = false)
-      : matcher_(lmatcher.matcher_, safe),
-        fst_(matcher_.GetFst()),
-        lfst_(lmatcher.lfst_),
-        state_(kNoStateId) {}
-
-  // General matcher methods.
-  ArcLookAheadMatcher<M, flags> *Copy(bool safe = false) const override {
-    return new ArcLookAheadMatcher<M, flags>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override { return matcher_.Type(test); }
-
-  void SetState(StateId s) final {
-    state_ = s;
-    matcher_.SetState(s);
-  }
-
-  bool Find(Label label) final { return matcher_.Find(label); }
-
-  bool Done() const final { return matcher_.Done(); }
-
-  const Arc &Value() const final { return matcher_.Value(); }
-
-  void Next() final { matcher_.Next(); }
-
-  Weight Final(StateId s) const final { return matcher_.Final(s); }
-
-  ssize_t Priority(StateId s) final { return matcher_.Priority(s); }
-
-  const FST &GetFst() const override { return fst_; }
-
-  uint64 Properties(uint64 props) const override {
-    return matcher_.Properties(props);
-  }
-
-  uint32 Flags() const override {
-    return matcher_.Flags() | kInputLookAheadMatcher | kOutputLookAheadMatcher |
-           kFlags;
-  }
-
-  const MatcherData *GetData() const { return nullptr; }
-
-  std::shared_ptr<MatcherData> GetSharedData() const {
-    return std::shared_ptr<MatcherData>();
-  }
-
-  // Look-ahead methods.
-
-  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) override {
-    lfst_ = &fst;
-  }
-
-  // Checks if there is a matching (possibly super-final) transition
-  // at (state_, s).
-  bool LookAheadFst(const Fst<Arc> &, StateId) final;
-
-  bool LookAheadLabel(Label label) const final { return matcher_.Find(label); }
-
- private:
-  mutable M matcher_;
-  const FST &fst_;        // Matcher FST.
-  const Fst<Arc> *lfst_;  // Look-ahead FST.
-  StateId state_;         // Matcher state.
-};
-
-template <class M, uint32 flags>
-bool ArcLookAheadMatcher<M, flags>::LookAheadFst(const Fst<Arc> &fst,
-                                                 StateId s) {
-  if (&fst != lfst_) InitLookAheadFst(fst);
-  bool result = false;
-  ssize_t nprefix = 0;
-  if (kFlags & kLookAheadWeight) ClearLookAheadWeight();
-  if (kFlags & kLookAheadPrefix) ClearLookAheadPrefix();
-  if (fst_.Final(state_) != Weight::Zero() &&
-      lfst_->Final(s) != Weight::Zero()) {
-    if (!(kFlags & (kLookAheadWeight | kLookAheadPrefix))) return true;
-    ++nprefix;
-    if (kFlags & kLookAheadWeight) {
-      SetLookAheadWeight(
-          Plus(LookAheadWeight(), Times(fst_.Final(state_), lfst_->Final(s))));
-    }
-    result = true;
-  }
-  if (matcher_.Find(kNoLabel)) {
-    if (!(kFlags & (kLookAheadWeight | kLookAheadPrefix))) return true;
-    ++nprefix;
-    if (kFlags & kLookAheadWeight) {
-      for (; !matcher_.Done(); matcher_.Next()) {
-        SetLookAheadWeight(Plus(LookAheadWeight(), matcher_.Value().weight));
-      }
-    }
-    result = true;
-  }
-  for (ArcIterator<Fst<Arc>> aiter(*lfst_, s); !aiter.Done(); aiter.Next()) {
-    const auto &arc = aiter.Value();
-    Label label = kNoLabel;
-    switch (matcher_.Type(false)) {
-      case MATCH_INPUT:
-        label = arc.olabel;
-        break;
-      case MATCH_OUTPUT:
-        label = arc.ilabel;
-        break;
-      default:
-        FSTERROR() << "ArcLookAheadMatcher::LookAheadFst: Bad match type";
-        return true;
-    }
-    if (label == 0) {
-      if (!(kFlags & (kLookAheadWeight | kLookAheadPrefix))) return true;
-      if (!(kFlags & kLookAheadNonEpsilonPrefix)) ++nprefix;
-      if (kFlags & kLookAheadWeight) {
-        SetLookAheadWeight(Plus(LookAheadWeight(), arc.weight));
-      }
-      result = true;
-    } else if (matcher_.Find(label)) {
-      if (!(kFlags & (kLookAheadWeight | kLookAheadPrefix))) return true;
-      for (; !matcher_.Done(); matcher_.Next()) {
-        ++nprefix;
-        if (kFlags & kLookAheadWeight) {
-          SetLookAheadWeight(Plus(LookAheadWeight(),
-                                  Times(arc.weight, matcher_.Value().weight)));
-        }
-        if ((kFlags & kLookAheadPrefix) && nprefix == 1)
-          SetLookAheadPrefix(arc);
-      }
-      result = true;
-    }
-  }
-  if (kFlags & kLookAheadPrefix) {
-    if (nprefix == 1) {
-      ClearLookAheadWeight();  // Avoids double counting.
-    } else {
-      ClearLookAheadPrefix();
-    }
-  }
-  return result;
-}
-
-// Template argument flags accepts flags to control behavior. It must include
-// precisely one of kInputLookAheadMatcher or kOutputLookAheadMatcher.
-template <class M,
-          uint32 flags = kLookAheadEpsilons | kLookAheadWeight |
-                         kLookAheadPrefix | kLookAheadNonEpsilonPrefix |
-                         kLookAheadKeepRelabelData,
-          class Accumulator = DefaultAccumulator<typename M::Arc>,
-          class Reachable = LabelReachable<typename M::Arc, Accumulator>>
-class LabelLookAheadMatcher
-    : public LookAheadMatcherBase<typename M::FST::Arc> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using MatcherData = typename Reachable::Data;
-
-  using LookAheadMatcherBase<Arc>::ClearLookAheadWeight;
-  using LookAheadMatcherBase<Arc>::LookAheadWeight;
-  using LookAheadMatcherBase<Arc>::SetLookAheadWeight;
-  using LookAheadMatcherBase<Arc>::ClearLookAheadPrefix;
-  using LookAheadMatcherBase<Arc>::LookAheadPrefix;
-  using LookAheadMatcherBase<Arc>::SetLookAheadPrefix;
-
-  enum : uint32 { kFlags = flags };
-
-  LabelLookAheadMatcher(
-      const FST &fst, MatchType match_type,
-      std::shared_ptr<MatcherData> data = std::shared_ptr<MatcherData>(),
-      Accumulator *accumulator = nullptr)
-      : matcher_(fst, match_type),
-        lfst_(nullptr),
-        state_(kNoStateId),
-        error_(false) {
-    if (!(kFlags & (kInputLookAheadMatcher | kOutputLookAheadMatcher))) {
-      FSTERROR() << "LabelLookaheadMatcher: Bad matcher flags: " << kFlags;
-      error_ = true;
-    }
-    const bool reach_input = match_type == MATCH_INPUT;
-    if (data) {
-      if (reach_input == data->ReachInput()) {
-        label_reachable_.reset(new Reachable(data, accumulator));
-      }
-    } else if ((reach_input && (kFlags & kInputLookAheadMatcher)) ||
-               (!reach_input && (kFlags & kOutputLookAheadMatcher))) {
-      label_reachable_.reset(new Reachable(fst, reach_input, accumulator,
-                                           kFlags & kLookAheadKeepRelabelData));
-    }
-  }
-
-  LabelLookAheadMatcher(
-      const LabelLookAheadMatcher<M, flags, Accumulator, Reachable> &lmatcher,
-      bool safe = false)
-      : matcher_(lmatcher.matcher_, safe),
-        lfst_(lmatcher.lfst_),
-        label_reachable_(lmatcher.label_reachable_
-                             ? new Reachable(*lmatcher.label_reachable_, safe)
-                             : nullptr),
-        state_(kNoStateId),
-        error_(lmatcher.error_) {}
-
-  LabelLookAheadMatcher<M, flags, Accumulator, Reachable> *Copy(
-      bool safe = false) const override {
-    return new LabelLookAheadMatcher<M, flags, Accumulator, Reachable>(*this,
-                                                                       safe);
-  }
-
-  MatchType Type(bool test) const override { return matcher_.Type(test); }
-
-  void SetState(StateId s) final {
-    if (state_ == s) return;
-    state_ = s;
-    match_set_state_ = false;
-    reach_set_state_ = false;
-  }
-
-  bool Find(Label label) final {
-    if (!match_set_state_) {
-      matcher_.SetState(state_);
-      match_set_state_ = true;
-    }
-    return matcher_.Find(label);
-  }
-
-  bool Done() const final { return matcher_.Done(); }
-
-  const Arc &Value() const final { return matcher_.Value(); }
-
-  void Next() final { matcher_.Next(); }
-
-  Weight Final(StateId s) const final { return matcher_.Final(s); }
-
-  ssize_t Priority(StateId s) final { return matcher_.Priority(s); }
-
-  const FST &GetFst() const override { return matcher_.GetFst(); }
-
-  uint64 Properties(uint64 inprops) const override {
-    auto outprops = matcher_.Properties(inprops);
-    if (error_ || (label_reachable_ && label_reachable_->Error())) {
-      outprops |= kError;
-    }
-    return outprops;
-  }
-
-  uint32 Flags() const override {
-    if (label_reachable_ && label_reachable_->GetData()->ReachInput()) {
-      return matcher_.Flags() | kFlags | kInputLookAheadMatcher;
-    } else if (label_reachable_ && !label_reachable_->GetData()->ReachInput()) {
-      return matcher_.Flags() | kFlags | kOutputLookAheadMatcher;
-    } else {
-      return matcher_.Flags();
-    }
-  }
-
-  const MatcherData *GetData() const {
-    return label_reachable_ ? label_reachable_->GetData() : nullptr;
-  };
-
-  std::shared_ptr<MatcherData> GetSharedData() const {
-    return label_reachable_ ? label_reachable_->GetSharedData()
-                            : std::shared_ptr<MatcherData>();
-  }
-  // Checks if there is a matching (possibly super-final) transition at
-  // (state_, s).
-  template <class LFST>
-  bool LookAheadFst(const LFST &fst, StateId s);
-
-  // Required to make class concrete.
-  bool LookAheadFst(const Fst<Arc> &fst, StateId s) final {
-    return LookAheadFst<Fst<Arc>>(fst, s);
-  }
-
-  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) override {
-    lfst_ = &fst;
-    if (label_reachable_) {
-      const bool reach_input = Type(false) == MATCH_OUTPUT;
-      label_reachable_->ReachInit(fst, reach_input, copy);
-    }
-  }
-
-  template <class LFST>
-  void InitLookAheadFst(const LFST &fst, bool copy = false) {
-    lfst_ = static_cast<const Fst<Arc> *>(&fst);
-    if (label_reachable_) {
-      const bool reach_input = Type(false) == MATCH_OUTPUT;
-      label_reachable_->ReachInit(fst, reach_input, copy);
-    }
-  }
-
-  bool LookAheadLabel(Label label) const final {
-    if (label == 0) return true;
-    if (label_reachable_) {
-      if (!reach_set_state_) {
-        label_reachable_->SetState(state_);
-        reach_set_state_ = true;
-      }
-      return label_reachable_->Reach(label);
-    } else {
-      return true;
-    }
-  }
-
- private:
-  mutable M matcher_;
-  const Fst<Arc> *lfst_;                        // Look-ahead FST.
-  std::unique_ptr<Reachable> label_reachable_;  // Label reachability info.
-  StateId state_;                               // Matcher state.
-  bool match_set_state_;                        // matcher_.SetState called?
-  mutable bool reach_set_state_;                // reachable_.SetState called?
-  bool error_;                                  // Error encountered?
-};
-
-template <class M, uint32 flags, class Accumulator, class Reachable>
-template <class LFST>
-inline bool LabelLookAheadMatcher<M, flags, Accumulator,
-                                  Reachable>::LookAheadFst(const LFST &fst,
-                                                           StateId s) {
-  if (static_cast<const Fst<Arc> *>(&fst) != lfst_) InitLookAheadFst(fst);
-  ClearLookAheadWeight();
-  ClearLookAheadPrefix();
-  if (!label_reachable_) return true;
-  label_reachable_->SetState(state_, s);
-  reach_set_state_ = true;
-  bool compute_weight = kFlags & kLookAheadWeight;
-  bool compute_prefix = kFlags & kLookAheadPrefix;
-  ArcIterator<LFST> aiter(fst, s);
-  aiter.SetFlags(kArcNoCache, kArcNoCache);  // Makes caching optional.
-  const bool reach_arc = label_reachable_->Reach(
-      &aiter, 0, internal::NumArcs(*lfst_, s), compute_weight);
-  const auto lfinal = internal::Final(*lfst_, s);
-  const bool reach_final =
-      lfinal != Weight::Zero() && label_reachable_->ReachFinal();
-  if (reach_arc) {
-    const auto begin = label_reachable_->ReachBegin();
-    const auto end = label_reachable_->ReachEnd();
-    if (compute_prefix && end - begin == 1 && !reach_final) {
-      aiter.Seek(begin);
-      SetLookAheadPrefix(aiter.Value());
-      compute_weight = false;
-    } else if (compute_weight) {
-      SetLookAheadWeight(label_reachable_->ReachWeight());
-    }
-  }
-  if (reach_final && compute_weight) {
-    SetLookAheadWeight(reach_arc ? Plus(LookAheadWeight(), lfinal) : lfinal);
-  }
-  return reach_arc || reach_final;
-}
-
-// Label-lookahead relabeling class.
-template <class Arc, class Data = LabelReachableData<typename Arc::Label>>
-class LabelLookAheadRelabeler {
- public:
-  using Label = typename Arc::Label;
-  using Reachable = LabelReachable<Arc, DefaultAccumulator<Arc>, Data>;
-
-  // Relabels matcher FST (initialization function object).
-  template <typename Impl>
-  explicit LabelLookAheadRelabeler(std::shared_ptr<Impl> *impl);
-
-  // Relabels arbitrary FST. Class LFST should be a label-lookahead FST.
-  template <class LFST>
-  static void Relabel(MutableFst<Arc> *fst, const LFST &mfst,
-                      bool relabel_input) {
-    const auto *data = mfst.GetAddOn();
-    Reachable reachable(data->First() ? data->SharedFirst()
-                                      : data->SharedSecond());
-    reachable.Relabel(fst, relabel_input);
-  }
-
-  // Returns relabeling pairs (cf. relabel.h::Relabel()). Class LFST should be a
-  // label-lookahead FST. If avoid_collisions is true, extra pairs are added to
-  // ensure no collisions when relabeling automata that have labels unseen here.
-  template <class LFST>
-  static void RelabelPairs(const LFST &mfst,
-                           std::vector<std::pair<Label, Label>> *pairs,
-                           bool avoid_collisions = false) {
-    const auto *data = mfst.GetAddOn();
-    Reachable reachable(data->First() ? data->SharedFirst()
-                                      : data->SharedSecond());
-    reachable.RelabelPairs(pairs, avoid_collisions);
-  }
-};
-
-template <class Arc, class Data>
-template <typename Impl>
-inline LabelLookAheadRelabeler<Arc, Data>::LabelLookAheadRelabeler(
-    std::shared_ptr<Impl> *impl) {
-  Fst<Arc> &fst = (*impl)->GetFst();
-  auto data = (*impl)->GetSharedAddOn();
-  const auto name = (*impl)->Type();
-  const bool is_mutable = fst.Properties(kMutable, false);
-  std::unique_ptr<MutableFst<Arc>> mfst;
-  if (is_mutable) {
-    mfst.reset(static_cast<MutableFst<Arc> *>(&fst));
-  } else {
-    mfst.reset(new VectorFst<Arc>(fst));
-  }
-  if (data->First()) {  // reach_input.
-    Reachable reachable(data->SharedFirst());
-    reachable.Relabel(mfst.get(), true);
-    if (!FLAGS_save_relabel_ipairs.empty()) {
-      std::vector<std::pair<Label, Label>> pairs;
-      reachable.RelabelPairs(&pairs, true);
-      WriteLabelPairs(FLAGS_save_relabel_ipairs, pairs);
-    }
-  } else {
-    Reachable reachable(data->SharedSecond());
-    reachable.Relabel(mfst.get(), false);
-    if (!FLAGS_save_relabel_opairs.empty()) {
-      std::vector<std::pair<Label, Label>> pairs;
-      reachable.RelabelPairs(&pairs, true);
-      WriteLabelPairs(FLAGS_save_relabel_opairs, pairs);
-    }
-  }
-  if (!is_mutable) {
-    *impl = std::make_shared<Impl>(*mfst, name);
-    (*impl)->SetAddOn(data);
-  }
-}
-
-// Generic lookahead matcher, templated on the FST definition (a wrapper around
-// a pointer to specific one).
-template <class F>
-class LookAheadMatcher {
- public:
-  using FST = F;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using LBase = LookAheadMatcherBase<Arc>;
-
-  LookAheadMatcher(const FST &fst, MatchType match_type)
-      : base_(fst.InitMatcher(match_type)) {
-    if (!base_) base_.reset(new SortedMatcher<FST>(fst, match_type));
-    lookahead_ = false;
-  }
-
-  // Takes ownership of base.
-  explicit LookAheadMatcher(MatcherBase<Arc> *base)
-      : base_(base), lookahead_(false) {}
-
-  LookAheadMatcher(const LookAheadMatcher<FST> &matcher, bool safe = false)
-      : base_(matcher.base_->Copy(safe)) {
-    lookahead_ = matcher.lookahead_;
-  }
-
-  LookAheadMatcher<FST> *Copy(bool safe = false) const {
-    return new LookAheadMatcher<FST>(*this, safe);
-  }
-
-  MatchType Type(bool test) const { return base_->Type(test); }
-
-  void SetState(StateId s) { base_->SetState(s); }
-
-  bool Find(Label label) { return base_->Find(label); }
-
-  bool Done() const { return base_->Done(); }
-
-  const Arc &Value() const { return base_->Value(); }
-
-  void Next() { base_->Next(); }
-
-  Weight Final(StateId s) const { return base_->Final(s); }
-
-  ssize_t Priority(StateId s) { return base_->Priority(s); }
-
-  const FST &GetFst() const {
-    return static_cast<const FST &>(base_->GetFst());
-  }
-
-  uint64 Properties(uint64 props) const { return base_->Properties(props); }
-
-  uint32 Flags() const { return base_->Flags(); }
-
-  bool LookAheadLabel(Label label) const {
-    if (LookAheadCheck()) {
-      return static_cast<LBase *>(base_.get())->LookAheadLabel(label);
-    } else {
-      return true;
-    }
-  }
-
-  bool LookAheadFst(const Fst<Arc> &fst, StateId s) {
-    if (LookAheadCheck()) {
-      return static_cast<LBase *>(base_.get())->LookAheadFst(fst, s);
-    } else {
-      return true;
-    }
-  }
-
-  Weight LookAheadWeight() const {
-    if (LookAheadCheck()) {
-      return static_cast<LBase *>(base_.get())->LookAheadWeight();
-    } else {
-      return Weight::One();
-    }
-  }
-
-  bool LookAheadPrefix(Arc *arc) const {
-    if (LookAheadCheck()) {
-      return static_cast<LBase *>(base_.get())->LookAheadPrefix(arc);
-    } else {
-      return false;
-    }
-  }
-
-  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) {
-    if (LookAheadCheck()) {
-      static_cast<LBase *>(base_.get())->InitLookAheadFst(fst, copy);
-    }
-  }
-
- private:
-  bool LookAheadCheck() const {
-    if (!lookahead_) {
-      lookahead_ =
-          base_->Flags() & (kInputLookAheadMatcher | kOutputLookAheadMatcher);
-      if (!lookahead_) {
-        FSTERROR() << "LookAheadMatcher: No look-ahead matcher defined";
-      }
-    }
-    return lookahead_;
-  }
-
-  std::unique_ptr<MatcherBase<Arc>> base_;
-  mutable bool lookahead_;
-
-  LookAheadMatcher &operator=(const LookAheadMatcher &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_LOOKAHEAD_MATCHER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/map.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/map.h
deleted file mode 100644
index 5d663ae8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/map.h
+++ /dev/null
@@ -1,110 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Compatibility file for old-style Map() functions and MapFst class that have
-// been renamed to ArcMap (cf. StateMap).
-
-#ifndef FST_MAP_H_
-#define FST_MAP_H_
-
-
-#include <fst/arc-map.h>
-
-
-namespace fst {
-
-template <class A, class C>
-void Map(MutableFst<A> *fst, C *mapper) {
-  ArcMap(fst, mapper);
-}
-
-template <class A, class C>
-void Map(MutableFst<A> *fst, C mapper) {
-  ArcMap(fst, mapper);
-}
-
-template <class A, class B, class C>
-void Map(const Fst<A> &ifst, MutableFst<B> *ofst, C *mapper) {
-  ArcMap(ifst, ofst, mapper);
-}
-
-template <class A, class B, class C>
-void Map(const Fst<A> &ifst, MutableFst<B> *ofst, C mapper) {
-  ArcMap(ifst, ofst, mapper);
-}
-
-using MapFstOptions = ArcMapFstOptions;
-
-template <class A, class B, class C>
-class MapFst : public ArcMapFst<A, B, C> {
- public:
-  using FromArc = A;
-  using ToArc = B;
-
-  using StateId = typename ToArc::StateId;
-  using Weight = typename ToArc::Weight;
-
-  using State = CacheState<B>;
-
-  MapFst(const Fst<A> &fst, const C &mapper, const MapFstOptions &opts)
-      : ArcMapFst<A, B, C>(fst, mapper, opts) {}
-
-  MapFst(const Fst<A> &fst, C *mapper, const MapFstOptions &opts)
-      : ArcMapFst<A, B, C>(fst, mapper, opts) {}
-
-  MapFst(const Fst<A> &fst, const C &mapper)
-      : ArcMapFst<A, B, C>(fst, mapper) {}
-
-  MapFst(const Fst<A> &fst, C *mapper) : ArcMapFst<A, B, C>(fst, mapper) {}
-
-  // See Fst<>::Copy() for doc.
-  MapFst(const MapFst<A, B, C> &fst, bool safe = false)
-      : ArcMapFst<A, B, C>(fst, safe) {}
-
-  // Get a copy of this MapFst. See Fst<>::Copy() for further doc.
-  MapFst<A, B, C> *Copy(bool safe = false) const override {
-    return new MapFst(*this, safe);
-  }
-};
-
-// Specialization for MapFst.
-template <class A, class B, class C>
-class StateIterator<MapFst<A, B, C>>
-    : public StateIterator<ArcMapFst<A, B, C>> {
- public:
-  explicit StateIterator(const ArcMapFst<A, B, C> &fst)
-      : StateIterator<ArcMapFst<A, B, C>>(fst) {}
-};
-
-// Specialization for MapFst.
-template <class A, class B, class C>
-class ArcIterator<MapFst<A, B, C>> : public ArcIterator<ArcMapFst<A, B, C>> {
- public:
-  ArcIterator(const ArcMapFst<A, B, C> &fst, typename A::StateId s)
-      : ArcIterator<ArcMapFst<A, B, C>>(fst, s) {}
-};
-
-// For backwards compatibility only; use IdentityArcMapper otherwise.
-template <class A>
-struct IdentityMapper {
-  using FromArc = A;
-  using ToArc = A;
-
-  ToArc operator()(const FromArc &arc) const { return arc; }
-
-  constexpr MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const { return props; }
-};
-
-}  // namespace fst
-
-#endif  // FST_MAP_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/mapped-file.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/mapped-file.h
deleted file mode 100644
index 0471ff80..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/mapped-file.h
+++ /dev/null
@@ -1,75 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_MAPPED_FILE_H_
-#define FST_MAPPED_FILE_H_
-
-#include <cstddef>
-#include <istream>
-#include <string>
-
-#include <fst/compat.h>
-#include <fst/flags.h>
-
-namespace fst {
-
-// A memory region is a simple abstraction for allocated memory or data from
-// memory-mapped files. If mmap is null, then data represents an owned region
-// of size bytes. Otherwise, mmap and size refer to the mapping and data is a
-// casted pointer to a region contained within [mmap, mmap + size). If size is
-// 0, then mmap and data refer to a block of memory managed externally by some
-// other allocator. The offset is used when allocating memory to providing
-// padding for alignment.
-struct MemoryRegion {
-  void *data;
-  void *mmap;
-  size_t size;
-  int offset;
-};
-
-class MappedFile {
- public:
-  ~MappedFile();
-
-  void *mutable_data() const { return region_.data; }
-
-  const void *data() const { return region_.data; }
-
-  // Returns a MappedFile object that contains the contents of the input stream
-  // strm starting from the current file position with size bytes. The memorymap
-  // bool is advisory, and Map will default to allocating and reading. The
-  // source argument needs to contain the filename that was used to open the
-  // input stream.
-  static MappedFile *Map(std::istream *istrm, bool memorymap,
-                         const string &source, size_t size);
-
-  // Creates a MappedFile object with a new[]'ed block of memory of size. The
-  // align argument can be used to specify a desired block alignment.
-  // This is RECOMMENDED FOR INTERNAL USE ONLY as it may change in future
-  // releases.
-  static MappedFile *Allocate(size_t size, int align = kArchAlignment);
-
-  // Creates a MappedFile object pointing to a borrowed reference to data. This
-  // block of memory is not owned by the MappedFile object and will not be
-  // freed. This is RECOMMENDED FOR INTERNAL USE ONLY, may change in future
-  // releases.
-  static MappedFile *Borrow(void *data);
-
-  // Alignment required for mapping structures in bytes. Regions of memory that
-  // are not aligned upon a 128-bit boundary are read from the file instead.
-  // This is consistent with the alignment boundary set in ConstFst and
-  // CompactFst.
-  static constexpr int kArchAlignment = 16;
-
-  static constexpr size_t kMaxReadChunk = 256 * 1024 * 1024;  // 256 MB.
-
- private:
-  explicit MappedFile(const MemoryRegion &region);
-
-  MemoryRegion region_;
-  MappedFile(const MappedFile &) = delete;
-  MappedFile &operator=(const MappedFile &) = delete;
-};
-}  // namespace fst
-
-#endif  // FST_MAPPED_FILE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/matcher-fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/matcher-fst.h
deleted file mode 100644
index 7aad31eb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/matcher-fst.h
+++ /dev/null
@@ -1,347 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to add a matcher to an FST.
-
-#ifndef FST_MATCHER_FST_H_
-#define FST_MATCHER_FST_H_
-
-#include <memory>
-#include <string>
-
-#include <fst/add-on.h>
-#include <fst/const-fst.h>
-#include <fst/lookahead-matcher.h>
-
-
-namespace fst {
-
-// Writeable matchers have the same interface as Matchers (as defined in
-// matcher.h) along with the following additional methods:
-//
-// template <class F>
-// class Matcher {
-//  public:
-//   using FST = F;
-//   ...
-//   using MatcherData = ...;   // Initialization data.
-//
-//   // Constructor with additional argument for external initialization data;
-//   // matcher increments its reference count on construction and decrements
-//   // the reference count, and deletes once the reference count has reached
-//   // zero.
-//   Matcher(const FST &fst, MatchType type, MatcherData *data);
-//
-//   // Returns pointer to initialization data that can be passed to a Matcher
-//   // constructor.
-//   MatcherData *GetData() const;
-// };
-
-// The matcher initialization data class must also provide the following
-// interface:
-//
-// class MatcherData {
-// public:
-//   // Required copy constructor.
-//   MatcherData(const MatcherData &);
-//
-//   // Required I/O methods.
-//   static MatcherData *Read(std::istream &istrm, const FstReadOptions &opts);
-//   bool Write(std::ostream &ostrm, const FstWriteOptions &opts) const;
-// };
-
-// Trivial (no-op) MatcherFst initializer functor.
-template <class M>
-class NullMatcherFstInit {
- public:
-  using MatcherData = typename M::MatcherData;
-  using Data = AddOnPair<MatcherData, MatcherData>;
-  using Impl = internal::AddOnImpl<typename M::FST, Data>;
-
-  explicit NullMatcherFstInit(std::shared_ptr<Impl> *) {}
-};
-
-// Class adding a matcher to an FST type. Creates a new FST whose name is given
-// by N. An optional functor Init can be used to initialize the FST. The Data
-// template parameter allows the user to select the type of the add-on.
-template <
-    class F, class M, const char *Name, class Init = NullMatcherFstInit<M>,
-    class Data = AddOnPair<typename M::MatcherData, typename M::MatcherData>>
-class MatcherFst : public ImplToExpandedFst<internal::AddOnImpl<F, Data>> {
- public:
-  using FST = F;
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-
-  using FstMatcher = M;
-  using MatcherData = typename FstMatcher::MatcherData;
-
-  using Impl = internal::AddOnImpl<FST, Data>;
-  using D = Data;
-
-  friend class StateIterator<MatcherFst<FST, FstMatcher, Name, Init, Data>>;
-  friend class ArcIterator<MatcherFst<FST, FstMatcher, Name, Init, Data>>;
-
-  MatcherFst() : ImplToExpandedFst<Impl>(std::make_shared<Impl>(FST(), Name)) {}
-
-  explicit MatcherFst(const FST &fst, std::shared_ptr<Data> data = nullptr)
-      : ImplToExpandedFst<Impl>(data ? CreateImpl(fst, Name, data)
-                                     : CreateDataAndImpl(fst, Name)) {}
-
-  explicit MatcherFst(const Fst<Arc> &fst)
-      : ImplToExpandedFst<Impl>(CreateDataAndImpl(fst, Name)) {}
-
-  // See Fst<>::Copy() for doc.
-  MatcherFst(const MatcherFst<FST, FstMatcher, Name, Init, Data> &fst,
-             bool safe = false)
-      : ImplToExpandedFst<Impl>(fst, safe) {}
-
-  // Get a copy of this MatcherFst. See Fst<>::Copy() for further doc.
-  MatcherFst<FST, FstMatcher, Name, Init, Data> *Copy(
-      bool safe = false) const override {
-    return new MatcherFst<FST, FstMatcher, Name, Init, Data>(*this, safe);
-  }
-
-  // Read a MatcherFst from an input stream; return nullptr on error
-  static MatcherFst<FST, M, Name, Init, Data> *Read(
-      std::istream &strm, const FstReadOptions &opts) {
-    auto *impl = Impl::Read(strm, opts);
-    return impl ? new MatcherFst<FST, FstMatcher, Name, Init, Data>(
-                      std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  // Read a MatcherFst from a file; return nullptr on error
-  // Empty filename reads from standard input
-  static MatcherFst<FST, FstMatcher, Name, Init, Data> *Read(
-      const string &filename) {
-    auto *impl = ImplToExpandedFst<Impl>::Read(filename);
-    return impl ? new MatcherFst<FST, FstMatcher, Name, Init, Data>(
-                      std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return GetImpl()->Write(strm, opts);
-  }
-
-  bool Write(const string &filename) const override {
-    return Fst<Arc>::WriteFile(filename);
-  }
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const override {
-    return GetImpl()->InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    return GetImpl()->InitArcIterator(s, data);
-  }
-
-  FstMatcher *InitMatcher(MatchType match_type) const override {
-    return new FstMatcher(GetFst(), match_type, GetSharedData(match_type));
-  }
-
-  const FST &GetFst() const { return GetImpl()->GetFst(); }
-
-  const Data *GetAddOn() const { return GetImpl()->GetAddOn(); }
-
-  std::shared_ptr<Data> GetSharedAddOn() const {
-    return GetImpl()->GetSharedAddOn();
-  }
-
-  const MatcherData *GetData(MatchType match_type) const {
-    const auto *data = GetAddOn();
-    return match_type == MATCH_INPUT ? data->First() : data->Second();
-  }
-
-  std::shared_ptr<MatcherData> GetSharedData(MatchType match_type) const {
-    const auto *data = GetAddOn();
-    return match_type == MATCH_INPUT ? data->SharedFirst()
-                                     : data->SharedSecond();
-  }
-
- protected:
-  using ImplToFst<Impl, ExpandedFst<Arc>>::GetImpl;
-
-  static std::shared_ptr<Impl> CreateDataAndImpl(const FST &fst,
-                                                 const string &name) {
-    FstMatcher imatcher(fst, MATCH_INPUT);
-    FstMatcher omatcher(fst, MATCH_OUTPUT);
-    return CreateImpl(fst, name,
-                      std::make_shared<Data>(imatcher.GetSharedData(),
-                                             omatcher.GetSharedData()));
-  }
-
-  static std::shared_ptr<Impl> CreateDataAndImpl(const Fst<Arc> &fst,
-                                                 const string &name) {
-    FST result(fst);
-    return CreateDataAndImpl(result, name);
-  }
-
-  static std::shared_ptr<Impl> CreateImpl(const FST &fst, const string &name,
-                                          std::shared_ptr<Data> data) {
-    auto impl = std::make_shared<Impl>(fst, name);
-    impl->SetAddOn(data);
-    Init init(&impl);
-    return impl;
-  }
-
-  explicit MatcherFst(std::shared_ptr<Impl> impl)
-      : ImplToExpandedFst<Impl>(impl) {}
-
- private:
-  MatcherFst &operator=(const MatcherFst &) = delete;
-};
-
-// Specialization for MatcherFst.
-template <class FST, class M, const char *Name, class Init>
-class StateIterator<MatcherFst<FST, M, Name, Init>>
-    : public StateIterator<FST> {
- public:
-  explicit StateIterator(const MatcherFst<FST, M, Name, Init> &fst)
-      : StateIterator<FST>(fst.GetImpl()->GetFst()) {}
-};
-
-// Specialization for MatcherFst.
-template <class FST, class M, const char *Name, class Init>
-class ArcIterator<MatcherFst<FST, M, Name, Init>> : public ArcIterator<FST> {
- public:
-  using StateId = typename FST::Arc::StateId;
-
-  ArcIterator(const MatcherFst<FST, M, Name, Init> &fst,
-              typename FST::Arc::StateId s)
-      : ArcIterator<FST>(fst.GetImpl()->GetFst(), s) {}
-};
-
-// Specialization for MatcherFst.
-template <class F, class M, const char *Name, class Init>
-class Matcher<MatcherFst<F, M, Name, Init>> {
- public:
-  using FST = MatcherFst<F, M, Name, Init>;
-  using Arc = typename F::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-
-  Matcher(const FST &fst, MatchType match_type)
-      : matcher_(fst.InitMatcher(match_type)) {}
-
-  Matcher(const Matcher<FST> &matcher) : matcher_(matcher.matcher_->Copy()) {}
-
-  Matcher<FST> *Copy() const { return new Matcher<FST>(*this); }
-
-  MatchType Type(bool test) const { return matcher_->Type(test); }
-
-  void SetState(StateId s) { matcher_->SetState(s); }
-
-  bool Find(Label label) { return matcher_->Find(label); }
-
-  bool Done() const { return matcher_->Done(); }
-
-  const Arc &Value() const { return matcher_->Value(); }
-
-  void Next() { matcher_->Next(); }
-
-  uint64 Properties(uint64 props) const { return matcher_->Properties(props); }
-
-  uint32 Flags() const { return matcher_->Flags(); }
-
- private:
-  std::unique_ptr<M> matcher_;
-};
-
-// Specialization for MatcherFst.
-template <class F, class M, const char *Name, class Init>
-class LookAheadMatcher<MatcherFst<F, M, Name, Init>> {
- public:
-  using FST = MatcherFst<F, M, Name, Init>;
-  using Arc = typename F::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  LookAheadMatcher(const FST &fst, MatchType match_type)
-      : matcher_(fst.InitMatcher(match_type)) {}
-
-  LookAheadMatcher(const LookAheadMatcher<FST> &matcher, bool safe = false)
-      : matcher_(matcher.matcher_->Copy(safe)) {}
-
-  // General matcher methods.
-  LookAheadMatcher<FST> *Copy(bool safe = false) const {
-    return new LookAheadMatcher<FST>(*this, safe);
-  }
-
-  MatchType Type(bool test) const { return matcher_->Type(test); }
-
-  void SetState(StateId s) { matcher_->SetState(s); }
-
-  bool Find(Label label) { return matcher_->Find(label); }
-
-  bool Done() const { return matcher_->Done(); }
-
-  const Arc &Value() const { return matcher_->Value(); }
-
-  void Next() { matcher_->Next(); }
-
-  const FST &GetFst() const { return matcher_->GetFst(); }
-
-  uint64 Properties(uint64 props) const { return matcher_->Properties(props); }
-
-  uint32 Flags() const { return matcher_->Flags(); }
-
-  bool LookAheadLabel(Label label) const {
-    return matcher_->LookAheadLabel(label);
-  }
-
-  bool LookAheadFst(const Fst<Arc> &fst, StateId s) {
-    return matcher_->LookAheadFst(fst, s);
-  }
-
-  Weight LookAheadWeight() const { return matcher_->LookAheadWeight(); }
-
-  bool LookAheadPrefix(Arc *arc) const {
-    return matcher_->LookAheadPrefix(arc);
-  }
-
-  void InitLookAheadFst(const Fst<Arc> &fst, bool copy = false) {
-    matcher_->InitLookAheadFst(fst, copy);
-  }
-
- private:
-  std::unique_ptr<M> matcher_;
-};
-
-// Useful aliases when using StdArc.
-
-extern const char arc_lookahead_fst_type[];
-
-using StdArcLookAheadFst =
-    MatcherFst<ConstFst<StdArc>,
-               ArcLookAheadMatcher<SortedMatcher<ConstFst<StdArc>>>,
-               arc_lookahead_fst_type>;
-
-extern const char ilabel_lookahead_fst_type[];
-extern const char olabel_lookahead_fst_type[];
-
-constexpr auto ilabel_lookahead_flags =
-    kInputLookAheadMatcher | kLookAheadWeight | kLookAheadPrefix |
-    kLookAheadEpsilons | kLookAheadNonEpsilonPrefix;
-
-constexpr auto olabel_lookahead_flags =
-    kOutputLookAheadMatcher | kLookAheadWeight | kLookAheadPrefix |
-    kLookAheadEpsilons | kLookAheadNonEpsilonPrefix;
-
-using StdILabelLookAheadFst = MatcherFst<
-    ConstFst<StdArc>,
-    LabelLookAheadMatcher<SortedMatcher<ConstFst<StdArc>>,
-                          ilabel_lookahead_flags, FastLogAccumulator<StdArc>>,
-    ilabel_lookahead_fst_type, LabelLookAheadRelabeler<StdArc>>;
-
-using StdOLabelLookAheadFst = MatcherFst<
-    ConstFst<StdArc>,
-    LabelLookAheadMatcher<SortedMatcher<ConstFst<StdArc>>,
-                          olabel_lookahead_flags, FastLogAccumulator<StdArc>>,
-    olabel_lookahead_fst_type, LabelLookAheadRelabeler<StdArc>>;
-
-}  // namespace fst
-
-#endif  // FST_MATCHER_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/matcher.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/matcher.h
deleted file mode 100644
index 975036ac..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/matcher.h
+++ /dev/null
@@ -1,1458 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes to allow matching labels leaving FST states.
-
-#ifndef FST_MATCHER_H_
-#define FST_MATCHER_H_
-
-#include <algorithm>
-#include <unordered_map>
-#include <utility>
-
-#include <fst/log.h>
-
-#include <fst/mutable-fst.h>  // for all internal FST accessors.
-
-
-namespace fst {
-
-// Matchers find and iterate through requested labels at FST states. In the
-// simplest form, these are just some associative map or search keyed on labels.
-// More generally, they may implement matching special labels that represent
-// sets of labels such as sigma (all), rho (rest), or phi (fail). The Matcher
-// interface is:
-//
-// template <class F>
-// class Matcher {
-//  public:
-//   using FST = F;
-//   using Arc = typename FST::Arc;
-//   using Label = typename Arc::Label;
-//   using StateId = typename Arc::StateId;
-//   using Weight = typename Arc::Weight;
-//
-//   // Required constructors.
-//
-//   Matcher(const FST &fst, MatchType type);
-//   Matcher(const Matcher &matcher, bool safe = false);
-//
-//   // Standard copy method.
-//   Matcher<FST> *Copy(bool safe = false) const override;
-//
-//   // Returns the match type that can be provided (depending on compatibility
-//   of the input FST). It is either the requested match type, MATCH_NONE, or
-//   MATCH_UNKNOWN. If test is false, a costly testing is avoided, but
-//   MATCH_UNKNOWN may be returned. If test is true, a definite answer is
-//   returned, but may involve more costly computation (e.g., visiting the FST).
-//   MatchType Type(bool test) const override;
-//
-//   // Specifies the current state.
-//   void SetState(StateId s) final;
-//
-//   // Finds matches to a label at the current state, returning true if a match
-//   // found. kNoLabel matches any non-consuming transitions, e.g., epsilon
-//   // transitions, which do not require a matching symbol.
-//   bool Find(Label label) final;
-//
-//   // Iterator methods. Note that initially and after SetState() these have
-//   undefined behavior until Find() is called.
-//
-//   bool Done() const final;
-//
-//   const Arc &Value() const final;
-//
-//   void Next() final;
-//
-//   // Returns final weight of a state.
-//   Weight Final(StateId) const final;
-//
-//   // Indicates preference for being the side used for matching in
-//   // composition. If the value is kRequirePriority, then it is
-//   // mandatory that it be used. Calling this method without passing the
-//   // current state of the matcher invalidates the state of the matcher.
-//   ssize_t Priority(StateId s) final;
-//
-//   // This specifies the known FST properties as viewed from this matcher. It
-//   // takes as argument the input FST's known properties.
-//   uint64 Properties(uint64 props) const override;
-//
-//   // Returns matcher flags.
-//   uint32 Flags() const override;
-//
-//   // Returns matcher FST.
-//   const FST &GetFst() const override;
-// };
-
-// Basic matcher flags.
-
-// Matcher needs to be used as the matching side in composition for
-// at least one state (has kRequirePriority).
-constexpr uint32 kRequireMatch = 0x00000001;
-
-// Flags used for basic matchers (see also lookahead.h).
-constexpr uint32 kMatcherFlags = kRequireMatch;
-
-// Matcher priority that is mandatory.
-constexpr ssize_t kRequirePriority = -1;
-
-// Matcher interface, templated on the Arc definition; used for matcher
-// specializations that are returned by the InitMatcher FST method.
-template <class A>
-class MatcherBase {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  virtual ~MatcherBase() {}
-
-  // Virtual interface.
-
-  virtual MatcherBase<Arc> *Copy(bool safe = false) const = 0;
-  virtual MatchType Type(bool) const = 0;
-  virtual void SetState(StateId) = 0;
-  virtual bool Find(Label) = 0;
-  virtual bool Done() const = 0;
-  virtual const Arc &Value() const = 0;
-  virtual void Next() = 0;
-  virtual const Fst<Arc> &GetFst() const = 0;
-  virtual uint64 Properties(uint64) const = 0;
-
-  // Trivial implementations that can be used by derived classes. Full
-  // devirtualization is expected for any derived class marked final.
-  virtual uint32 Flags() const { return 0; }
-
-  virtual Weight Final(StateId s) const { return internal::Final(GetFst(), s); }
-
-  virtual ssize_t Priority(StateId s) { return internal::NumArcs(GetFst(), s); }
-};
-
-// A matcher that expects sorted labels on the side to be matched.
-// If match_type == MATCH_INPUT, epsilons match the implicit self-loop
-// Arc(kNoLabel, 0, Weight::One(), current_state) as well as any
-// actual epsilon transitions. If match_type == MATCH_OUTPUT, then
-// Arc(0, kNoLabel, Weight::One(), current_state) is instead matched.
-template <class F>
-class SortedMatcher : public MatcherBase<typename F::Arc> {
- public:
-  using FST = F;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using MatcherBase<Arc>::Flags;
-  using MatcherBase<Arc>::Properties;
-
-  // Labels >= binary_label will be searched for by binary search;
-  // o.w. linear search is used.
-  SortedMatcher(const FST &fst, MatchType match_type, Label binary_label = 1)
-      : fst_(fst.Copy()),
-        state_(kNoStateId),
-        aiter_(nullptr),
-        match_type_(match_type),
-        binary_label_(binary_label),
-        match_label_(kNoLabel),
-        narcs_(0),
-        loop_(kNoLabel, 0, Weight::One(), kNoStateId),
-        error_(false),
-        aiter_pool_(1) {
-    switch (match_type_) {
-      case MATCH_INPUT:
-      case MATCH_NONE:
-        break;
-      case MATCH_OUTPUT:
-        std::swap(loop_.ilabel, loop_.olabel);
-        break;
-      default:
-        FSTERROR() << "SortedMatcher: Bad match type";
-        match_type_ = MATCH_NONE;
-        error_ = true;
-    }
-  }
-
-  SortedMatcher(const SortedMatcher<FST> &matcher, bool safe = false)
-      : fst_(matcher.fst_->Copy(safe)),
-        state_(kNoStateId),
-        aiter_(nullptr),
-        match_type_(matcher.match_type_),
-        binary_label_(matcher.binary_label_),
-        match_label_(kNoLabel),
-        narcs_(0),
-        loop_(matcher.loop_),
-        error_(matcher.error_),
-        aiter_pool_(1) {}
-
-  ~SortedMatcher() override { Destroy(aiter_, &aiter_pool_); }
-
-  SortedMatcher<FST> *Copy(bool safe = false) const override {
-    return new SortedMatcher<FST>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override {
-    if (match_type_ == MATCH_NONE) return match_type_;
-    const auto true_prop =
-        match_type_ == MATCH_INPUT ? kILabelSorted : kOLabelSorted;
-    const auto false_prop =
-        match_type_ == MATCH_INPUT ? kNotILabelSorted : kNotOLabelSorted;
-    const auto props = fst_->Properties(true_prop | false_prop, test);
-    if (props & true_prop) {
-      return match_type_;
-    } else if (props & false_prop) {
-      return MATCH_NONE;
-    } else {
-      return MATCH_UNKNOWN;
-    }
-  }
-
-  void SetState(StateId s) final {
-    if (state_ == s) return;
-    state_ = s;
-    if (match_type_ == MATCH_NONE) {
-      FSTERROR() << "SortedMatcher: Bad match type";
-      error_ = true;
-    }
-    Destroy(aiter_, &aiter_pool_);
-    aiter_ = new (&aiter_pool_) ArcIterator<FST>(*fst_, s);
-    aiter_->SetFlags(kArcNoCache, kArcNoCache);
-    narcs_ = internal::NumArcs(*fst_, s);
-    loop_.nextstate = s;
-  }
-
-  bool Find(Label match_label) final {
-    exact_match_ = true;
-    if (error_) {
-      current_loop_ = false;
-      match_label_ = kNoLabel;
-      return false;
-    }
-    current_loop_ = match_label == 0;
-    match_label_ = match_label == kNoLabel ? 0 : match_label;
-    if (Search()) {
-      return true;
-    } else {
-      return current_loop_;
-    }
-  }
-
-  // Positions matcher to the first position where inserting match_label would
-  // maintain the sort order.
-  void LowerBound(Label label) {
-    exact_match_ = false;
-    current_loop_ = false;
-    if (error_) {
-      match_label_ = kNoLabel;
-      return;
-    }
-    match_label_ = label;
-    Search();
-  }
-
-  // After Find(), returns false if no more exact matches.
-  // After LowerBound(), returns false if no more arcs.
-  bool Done() const final {
-    if (current_loop_) return false;
-    if (aiter_->Done()) return true;
-    if (!exact_match_) return false;
-    aiter_->SetFlags(
-        match_type_ == MATCH_INPUT ? kArcILabelValue : kArcOLabelValue,
-        kArcValueFlags);
-    const auto label = match_type_ == MATCH_INPUT ? aiter_->Value().ilabel
-                                                  : aiter_->Value().olabel;
-    return label != match_label_;
-  }
-
-  const Arc &Value() const final {
-    if (current_loop_) return loop_;
-    aiter_->SetFlags(kArcValueFlags, kArcValueFlags);
-    return aiter_->Value();
-  }
-
-  void Next() final {
-    if (current_loop_) {
-      current_loop_ = false;
-    } else {
-      aiter_->Next();
-    }
-  }
-
-  Weight Final(StateId s) const final {
-    return MatcherBase<Arc>::Final(s);
-  }
-
-  ssize_t Priority(StateId s) final {
-    return MatcherBase<Arc>::Priority(s);
-  }
-
-  const FST &GetFst() const override { return *fst_; }
-
-  uint64 Properties(uint64 inprops) const override {
-    return inprops | (error_ ? kError : 0);
-  }
-
-  size_t Position() const { return aiter_ ? aiter_->Position() : 0; }
-
- private:
-  Label GetLabel() const {
-    const auto &arc = aiter_->Value();
-    return match_type_ == MATCH_INPUT ? arc.ilabel : arc.olabel;
-  }
-
-  bool BinarySearch();
-  bool LinearSearch();
-  bool Search();
-
-  std::unique_ptr<const FST> fst_;
-  StateId state_;            // Matcher state.
-  ArcIterator<FST> *aiter_;  // Iterator for current state.
-  MatchType match_type_;     // Type of match to perform.
-  Label binary_label_;       // Least label for binary search.
-  Label match_label_;        // Current label to be matched.
-  size_t narcs_;             // Current state arc count.
-  Arc loop_;                 // For non-consuming symbols.
-  bool current_loop_;        // Current arc is the implicit loop.
-  bool exact_match_;         // Exact match or lower bound?
-  bool error_;               // Error encountered?
-  MemoryPool<ArcIterator<FST>> aiter_pool_;  // Pool of arc iterators.
-};
-
-// Returns true iff match to match_label_. The arc iterator is positioned at the
-// lower bound, that is, the first element greater than or equal to
-// match_label_, or the end if all elements are less than match_label_.
-template <class FST>
-inline bool SortedMatcher<FST>::BinarySearch() {
-  size_t low = 0;
-  size_t high = narcs_;
-  while (low < high) {
-    const size_t mid = low + (high - low) / 2;
-    aiter_->Seek(mid);
-    if (GetLabel() < match_label_) {
-      low = mid + 1;
-    } else {
-      high = mid;
-    }
-  }
-
-  aiter_->Seek(low);
-  return low < narcs_ && GetLabel() == match_label_;
-}
-
-// Returns true iff match to match_label_, positioning arc iterator at lower
-// bound.
-template <class FST>
-inline bool SortedMatcher<FST>::LinearSearch() {
-  for (aiter_->Reset(); !aiter_->Done(); aiter_->Next()) {
-    const auto label = GetLabel();
-    if (label == match_label_) return true;
-    if (label > match_label_) break;
-  }
-  return false;
-}
-
-// Returns true iff match to match_label_, positioning arc iterator at lower
-// bound.
-template <class FST>
-inline bool SortedMatcher<FST>::Search() {
-  aiter_->SetFlags(match_type_ == MATCH_INPUT ?
-                   kArcILabelValue : kArcOLabelValue,
-                   kArcValueFlags);
-  if (match_label_ >= binary_label_) {
-    return BinarySearch();
-  } else {
-    return LinearSearch();
-  }
-}
-
-// A matcher that stores labels in a per-state hash table populated upon the
-// first visit to that state. Sorting is not required. Treatment of
-// epsilons are the same as with SortedMatcher.
-template <class F>
-class HashMatcher : public MatcherBase<typename F::Arc> {
- public:
-  using FST = F;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using MatcherBase<Arc>::Flags;
-  using MatcherBase<Arc>::Final;
-  using MatcherBase<Arc>::Priority;
-
-  HashMatcher(const FST &fst, MatchType match_type)
-      : fst_(fst.Copy()),
-        state_(kNoStateId),
-        match_type_(match_type),
-        loop_(kNoLabel, 0, Weight::One(), kNoStateId) {
-    switch (match_type_) {
-      case MATCH_INPUT:
-      case MATCH_NONE:
-        break;
-      case MATCH_OUTPUT:
-        std::swap(loop_.ilabel, loop_.olabel);
-        break;
-      default:
-        FSTERROR() << "SortedMatcher: Bad match type";
-        match_type_ = MATCH_NONE;
-        error_ = true;
-    }
-  }
-
-  HashMatcher(const HashMatcher<FST> &matcher, bool safe = false)
-      : fst_(matcher.fst_->Copy(safe)),
-        state_(kNoStateId),
-        match_type_(matcher.match_type_),
-        loop_(matcher.loop_),
-        error_(matcher.error_) {}
-
-  HashMatcher<FST> *Copy(bool safe = false) const override {
-    return new HashMatcher<FST>(*this, safe);
-  }
-
-  // The argument is ignored as there are no relevant properties to test.
-  MatchType Type(bool test) const override { return match_type_; }
-
-  void SetState(StateId s) final;
-
-  bool Find(Label label) final {
-    current_loop_ = label == 0;
-    if (label == 0) {
-      Search(label);
-      return true;
-    }
-    if (label == kNoLabel) label = 0;
-    return Search(label);
-  }
-
-  bool Done() const final {
-    if (current_loop_) return false;
-    return label_it_ == label_end_;
-  }
-
-  const Arc &Value() const final {
-    if (current_loop_) return loop_;
-    aiter_->Seek(label_it_->second);
-    return aiter_->Value();
-  }
-
-  void Next() final {
-    if (current_loop_) {
-      current_loop_ = false;
-    } else {
-      ++label_it_;
-    }
-  }
-
-  const FST &GetFst() const override { return *fst_; }
-
-  uint64 Properties(uint64 inprops) const override {
-    return inprops | (error_ ? kError : 0);
-  }
-
- private:
-  bool Search(Label match_label);
-
-  using LabelTable = std::unordered_multimap<Label, size_t>;
-  using StateTable = std::unordered_map<StateId, LabelTable>;
-
-  std::unique_ptr<const FST> fst_;
-  StateId state_;  // Matcher state.
-  MatchType match_type_;
-  Arc loop_;           // The implicit loop itself.
-  bool current_loop_;  // Is the current arc is the implicit loop?
-  bool error_;         // Error encountered?
-  std::unique_ptr<ArcIterator<FST>> aiter_;
-  StateTable state_table_;   // Table from states to label table.
-  LabelTable *label_table_;  // Pointer to current state's label table.
-  typename LabelTable::iterator label_it_;   // Position for label.
-  typename LabelTable::iterator label_end_;  // Position for last label + 1.
-};
-
-template <class FST>
-void HashMatcher<FST>::SetState(typename FST::Arc::StateId s) {
-  if (state_ == s) return;
-  // Resets everything for the state.
-  state_ = s;
-  loop_.nextstate = state_;
-  aiter_.reset(new ArcIterator<FST>(*fst_, state_));
-  if (match_type_ == MATCH_NONE) {
-    FSTERROR() << "HashMatcher: Bad match type";
-    error_ = false;
-  }
-  // Attempts to insert a new label table; if it already exists,
-  // no additional work is done and we simply return.
-  auto it_and_success = state_table_.emplace(state_, LabelTable());
-  if (!it_and_success.second) return;
-  // Otherwise, populate this new table.
-  // Sets instance's pointer to the label table for this state.
-  label_table_ = &(it_and_success.first->second);
-  // Populates the label table.
-  label_table_->reserve(internal::NumArcs(*fst_, state_));
-  const auto aiter_flags =
-      (match_type_ == MATCH_INPUT ? kArcILabelValue : kArcOLabelValue) |
-      kArcNoCache;
-  aiter_->SetFlags(aiter_flags, kArcFlags);
-  for (; !aiter_->Done(); aiter_->Next()) {
-    const auto label = (match_type_ == MATCH_INPUT) ? aiter_->Value().ilabel
-                                                    : aiter_->Value().olabel;
-    label_table_->emplace(label, aiter_->Position());
-  }
-  aiter_->SetFlags(kArcValueFlags, kArcValueFlags);
-}
-
-template <class FST>
-inline bool HashMatcher<FST>::Search(typename FST::Arc::Label match_label) {
-  auto range = label_table_->equal_range(match_label);
-  if (range.first == range.second) return false;
-  label_it_ = range.first;
-  label_end_ = range.second;
-  aiter_->Seek(label_it_->second);
-  return true;
-}
-
-// Specifies whether we rewrite both the input and output sides during matching.
-enum MatcherRewriteMode {
-  MATCHER_REWRITE_AUTO = 0,  // Rewrites both sides iff acceptor.
-  MATCHER_REWRITE_ALWAYS,
-  MATCHER_REWRITE_NEVER
-};
-
-// For any requested label that doesn't match at a state, this matcher
-// considers the *unique* transition that matches the label 'phi_label'
-// (phi = 'fail'), and recursively looks for a match at its
-// destination.  When 'phi_loop' is true, if no match is found but a
-// phi self-loop is found, then the phi transition found is returned
-// with the phi_label rewritten as the requested label (both sides if
-// an acceptor, or if 'rewrite_both' is true and both input and output
-// labels of the found transition are 'phi_label').  If 'phi_label' is
-// kNoLabel, this special matching is not done.  PhiMatcher is
-// templated itself on a matcher, which is used to perform the
-// underlying matching.  By default, the underlying matcher is
-// constructed by PhiMatcher. The user can instead pass in this
-// object; in that case, PhiMatcher takes its ownership.
-// Phi non-determinism not supported. No non-consuming symbols other
-// than epsilon supported with the underlying template argument matcher.
-template <class M>
-class PhiMatcher : public MatcherBase<typename M::Arc> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  PhiMatcher(const FST &fst, MatchType match_type, Label phi_label = kNoLabel,
-             bool phi_loop = true,
-             MatcherRewriteMode rewrite_mode = MATCHER_REWRITE_AUTO,
-             M *matcher = nullptr)
-      : matcher_(matcher ? matcher : new M(fst, match_type)),
-        match_type_(match_type),
-        phi_label_(phi_label),
-        state_(kNoStateId),
-        phi_loop_(phi_loop),
-        error_(false) {
-    if (match_type == MATCH_BOTH) {
-      FSTERROR() << "PhiMatcher: Bad match type";
-      match_type_ = MATCH_NONE;
-      error_ = true;
-    }
-    if (rewrite_mode == MATCHER_REWRITE_AUTO) {
-      rewrite_both_ = fst.Properties(kAcceptor, true);
-    } else if (rewrite_mode == MATCHER_REWRITE_ALWAYS) {
-      rewrite_both_ = true;
-    } else {
-      rewrite_both_ = false;
-    }
-  }
-
-  PhiMatcher(const PhiMatcher<M> &matcher, bool safe = false)
-      : matcher_(new M(*matcher.matcher_, safe)),
-        match_type_(matcher.match_type_),
-        phi_label_(matcher.phi_label_),
-        rewrite_both_(matcher.rewrite_both_),
-        state_(kNoStateId),
-        phi_loop_(matcher.phi_loop_),
-        error_(matcher.error_) {}
-
-  PhiMatcher<M> *Copy(bool safe = false) const override {
-    return new PhiMatcher<M>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override { return matcher_->Type(test); }
-
-  void SetState(StateId s) final {
-    if (state_ == s) return;
-    matcher_->SetState(s);
-    state_ = s;
-    has_phi_ = phi_label_ != kNoLabel;
-  }
-
-  bool Find(Label match_label) final;
-
-  bool Done() const final { return matcher_->Done(); }
-
-  const Arc &Value() const final {
-    if ((phi_match_ == kNoLabel) && (phi_weight_ == Weight::One())) {
-      return matcher_->Value();
-    } else if (phi_match_ == 0) {  // Virtual epsilon loop.
-      phi_arc_ = Arc(kNoLabel, 0, Weight::One(), state_);
-      if (match_type_ == MATCH_OUTPUT) {
-        std::swap(phi_arc_.ilabel, phi_arc_.olabel);
-      }
-      return phi_arc_;
-    } else {
-      phi_arc_ = matcher_->Value();
-      phi_arc_.weight = Times(phi_weight_, phi_arc_.weight);
-      if (phi_match_ != kNoLabel) {  // Phi loop match.
-        if (rewrite_both_) {
-          if (phi_arc_.ilabel == phi_label_) phi_arc_.ilabel = phi_match_;
-          if (phi_arc_.olabel == phi_label_) phi_arc_.olabel = phi_match_;
-        } else if (match_type_ == MATCH_INPUT) {
-          phi_arc_.ilabel = phi_match_;
-        } else {
-          phi_arc_.olabel = phi_match_;
-        }
-      }
-      return phi_arc_;
-    }
-  }
-
-  void Next() final { matcher_->Next(); }
-
-  Weight Final(StateId s) const final {
-    auto weight = matcher_->Final(s);
-    if (phi_label_ == kNoLabel || weight != Weight::Zero()) {
-      return weight;
-    }
-    weight = Weight::One();
-    matcher_->SetState(s);
-    while (matcher_->Final(s) == Weight::Zero()) {
-      if (!matcher_->Find(phi_label_ == 0 ? -1 : phi_label_)) break;
-      weight = Times(weight, matcher_->Value().weight);
-      if (s == matcher_->Value().nextstate) {
-        return Weight::Zero();  // Does not follow phi self-loops.
-      }
-      s = matcher_->Value().nextstate;
-      matcher_->SetState(s);
-    }
-    weight = Times(weight, matcher_->Final(s));
-    return weight;
-  }
-
-  ssize_t Priority(StateId s) final {
-    if (phi_label_ != kNoLabel) {
-      matcher_->SetState(s);
-      const bool has_phi = matcher_->Find(phi_label_ == 0 ? -1 : phi_label_);
-      return has_phi ? kRequirePriority : matcher_->Priority(s);
-    } else {
-      return matcher_->Priority(s);
-    }
-  }
-
-  const FST &GetFst() const override { return matcher_->GetFst(); }
-
-  uint64 Properties(uint64 props) const override;
-
-  uint32 Flags() const override {
-    if (phi_label_ == kNoLabel || match_type_ == MATCH_NONE) {
-      return matcher_->Flags();
-    }
-    return matcher_->Flags() | kRequireMatch;
-  }
-
-  Label PhiLabel() const { return phi_label_; }
-
- private:
-  mutable std::unique_ptr<M> matcher_;
-  MatchType match_type_;  // Type of match requested.
-  Label phi_label_;       // Label that represents the phi transition.
-  bool rewrite_both_;     // Rewrite both sides when both are phi_label_?
-  bool has_phi_;          // Are there possibly phis at the current state?
-  Label phi_match_;       // Current label that matches phi loop.
-  mutable Arc phi_arc_;   // Arc to return.
-  StateId state_;         // Matcher state.
-  Weight phi_weight_;     // Product of the weights of phi transitions taken.
-  bool phi_loop_;         // When true, phi self-loop are allowed and treated
-                          // as rho (required for Aho-Corasick).
-  bool error_;            // Error encountered?
-
-  PhiMatcher &operator=(const PhiMatcher &) = delete;
-};
-
-template <class M>
-inline bool PhiMatcher<M>::Find(Label label) {
-  if (label == phi_label_ && phi_label_ != kNoLabel && phi_label_ != 0) {
-    FSTERROR() << "PhiMatcher::Find: bad label (phi): " << phi_label_;
-    error_ = true;
-    return false;
-  }
-  matcher_->SetState(state_);
-  phi_match_ = kNoLabel;
-  phi_weight_ = Weight::One();
-  // If phi_label_ == 0, there are no more true epsilon arcs.
-  if (phi_label_ == 0) {
-    if (label == kNoLabel) {
-      return false;
-    }
-    if (label == 0) {  // but a virtual epsilon loop needs to be returned.
-      if (!matcher_->Find(kNoLabel)) {
-        return matcher_->Find(0);
-      } else {
-        phi_match_ = 0;
-        return true;
-      }
-    }
-  }
-  if (!has_phi_ || label == 0 || label == kNoLabel) {
-    return matcher_->Find(label);
-  }
-  auto s = state_;
-  while (!matcher_->Find(label)) {
-    // Look for phi transition (if phi_label_ == 0, we need to look
-    // for -1 to avoid getting the virtual self-loop)
-    if (!matcher_->Find(phi_label_ == 0 ? -1 : phi_label_)) return false;
-    if (phi_loop_ && matcher_->Value().nextstate == s) {
-      phi_match_ = label;
-      return true;
-    }
-    phi_weight_ = Times(phi_weight_, matcher_->Value().weight);
-    s = matcher_->Value().nextstate;
-    matcher_->Next();
-    if (!matcher_->Done()) {
-      FSTERROR() << "PhiMatcher: Phi non-determinism not supported";
-      error_ = true;
-    }
-    matcher_->SetState(s);
-  }
-  return true;
-}
-
-template <class M>
-inline uint64 PhiMatcher<M>::Properties(uint64 inprops) const {
-  auto outprops = matcher_->Properties(inprops);
-  if (error_) outprops |= kError;
-  if (match_type_ == MATCH_NONE) {
-    return outprops;
-  } else if (match_type_ == MATCH_INPUT) {
-    if (phi_label_ == 0) {
-      outprops &= ~kEpsilons | ~kIEpsilons | ~kOEpsilons;
-      outprops |= kNoEpsilons | kNoIEpsilons;
-    }
-    if (rewrite_both_) {
-      return outprops &
-             ~(kODeterministic | kNonODeterministic | kString | kILabelSorted |
-               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
-    } else {
-      return outprops &
-             ~(kODeterministic | kAcceptor | kString | kILabelSorted |
-               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
-    }
-  } else if (match_type_ == MATCH_OUTPUT) {
-    if (phi_label_ == 0) {
-      outprops &= ~kEpsilons | ~kIEpsilons | ~kOEpsilons;
-      outprops |= kNoEpsilons | kNoOEpsilons;
-    }
-    if (rewrite_both_) {
-      return outprops &
-             ~(kIDeterministic | kNonIDeterministic | kString | kILabelSorted |
-               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
-    } else {
-      return outprops &
-             ~(kIDeterministic | kAcceptor | kString | kILabelSorted |
-               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
-    }
-  } else {
-    // Shouldn't ever get here.
-    FSTERROR() << "PhiMatcher: Bad match type: " << match_type_;
-    return 0;
-  }
-}
-
-// For any requested label that doesn't match at a state, this matcher
-// considers all transitions that match the label 'rho_label' (rho =
-// 'rest').  Each such rho transition found is returned with the
-// rho_label rewritten as the requested label (both sides if an
-// acceptor, or if 'rewrite_both' is true and both input and output
-// labels of the found transition are 'rho_label').  If 'rho_label' is
-// kNoLabel, this special matching is not done.  RhoMatcher is
-// templated itself on a matcher, which is used to perform the
-// underlying matching.  By default, the underlying matcher is
-// constructed by RhoMatcher.  The user can instead pass in this
-// object; in that case, RhoMatcher takes its ownership.
-// No non-consuming symbols other than epsilon supported with
-// the underlying template argument matcher.
-template <class M>
-class RhoMatcher : public MatcherBase<typename M::Arc> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  RhoMatcher(const FST &fst, MatchType match_type, Label rho_label = kNoLabel,
-             MatcherRewriteMode rewrite_mode = MATCHER_REWRITE_AUTO,
-             M *matcher = nullptr)
-      : matcher_(matcher ? matcher : new M(fst, match_type)),
-        match_type_(match_type),
-        rho_label_(rho_label),
-        error_(false),
-        state_(kNoStateId) {
-    if (match_type == MATCH_BOTH) {
-      FSTERROR() << "RhoMatcher: Bad match type";
-      match_type_ = MATCH_NONE;
-      error_ = true;
-    }
-    if (rho_label == 0) {
-      FSTERROR() << "RhoMatcher: 0 cannot be used as rho_label";
-      rho_label_ = kNoLabel;
-      error_ = true;
-    }
-    if (rewrite_mode == MATCHER_REWRITE_AUTO) {
-      rewrite_both_ = fst.Properties(kAcceptor, true);
-    } else if (rewrite_mode == MATCHER_REWRITE_ALWAYS) {
-      rewrite_both_ = true;
-    } else {
-      rewrite_both_ = false;
-    }
-  }
-
-  RhoMatcher(const RhoMatcher<M> &matcher, bool safe = false)
-      : matcher_(new M(*matcher.matcher_, safe)),
-        match_type_(matcher.match_type_),
-        rho_label_(matcher.rho_label_),
-        rewrite_both_(matcher.rewrite_both_),
-        error_(matcher.error_),
-        state_(kNoStateId) {}
-
-  RhoMatcher<M> *Copy(bool safe = false) const override {
-    return new RhoMatcher<M>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override { return matcher_->Type(test); }
-
-  void SetState(StateId s) final {
-    if (state_ == s) return;
-    state_ = s;
-    matcher_->SetState(s);
-    has_rho_ = rho_label_ != kNoLabel;
-  }
-
-  bool Find(Label label) final {
-    if (label == rho_label_ && rho_label_ != kNoLabel) {
-      FSTERROR() << "RhoMatcher::Find: bad label (rho)";
-      error_ = true;
-      return false;
-    }
-    if (matcher_->Find(label)) {
-      rho_match_ = kNoLabel;
-      return true;
-    } else if (has_rho_ && label != 0 && label != kNoLabel &&
-               (has_rho_ = matcher_->Find(rho_label_))) {
-      rho_match_ = label;
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  bool Done() const final { return matcher_->Done(); }
-
-  const Arc &Value() const final {
-    if (rho_match_ == kNoLabel) {
-      return matcher_->Value();
-    } else {
-      rho_arc_ = matcher_->Value();
-      if (rewrite_both_) {
-        if (rho_arc_.ilabel == rho_label_) rho_arc_.ilabel = rho_match_;
-        if (rho_arc_.olabel == rho_label_) rho_arc_.olabel = rho_match_;
-      } else if (match_type_ == MATCH_INPUT) {
-        rho_arc_.ilabel = rho_match_;
-      } else {
-        rho_arc_.olabel = rho_match_;
-      }
-      return rho_arc_;
-    }
-  }
-
-  void Next() final { matcher_->Next(); }
-
-  Weight Final(StateId s) const final { return matcher_->Final(s); }
-
-  ssize_t Priority(StateId s) final {
-    state_ = s;
-    matcher_->SetState(s);
-    has_rho_ = matcher_->Find(rho_label_);
-    if (has_rho_) {
-      return kRequirePriority;
-    } else {
-      return matcher_->Priority(s);
-    }
-  }
-
-  const FST &GetFst() const override { return matcher_->GetFst(); }
-
-  uint64 Properties(uint64 props) const override;
-
-  uint32 Flags() const override {
-    if (rho_label_ == kNoLabel || match_type_ == MATCH_NONE) {
-      return matcher_->Flags();
-    }
-    return matcher_->Flags() | kRequireMatch;
-  }
-
-  Label RhoLabel() const { return rho_label_; }
-
- private:
-  std::unique_ptr<M> matcher_;
-  MatchType match_type_;  // Type of match requested.
-  Label rho_label_;       // Label that represents the rho transition
-  bool rewrite_both_;     // Rewrite both sides when both are rho_label_?
-  bool has_rho_;          // Are there possibly rhos at the current state?
-  Label rho_match_;       // Current label that matches rho transition.
-  mutable Arc rho_arc_;   // Arc to return when rho match.
-  bool error_;            // Error encountered?
-  StateId state_;         // Matcher state.
-};
-
-template <class M>
-inline uint64 RhoMatcher<M>::Properties(uint64 inprops) const {
-  auto outprops = matcher_->Properties(inprops);
-  if (error_) outprops |= kError;
-  if (match_type_ == MATCH_NONE) {
-    return outprops;
-  } else if (match_type_ == MATCH_INPUT) {
-    if (rewrite_both_) {
-      return outprops &
-             ~(kODeterministic | kNonODeterministic | kString | kILabelSorted |
-               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
-    } else {
-      return outprops &
-             ~(kODeterministic | kAcceptor | kString | kILabelSorted |
-               kNotILabelSorted);
-    }
-  } else if (match_type_ == MATCH_OUTPUT) {
-    if (rewrite_both_) {
-      return outprops &
-             ~(kIDeterministic | kNonIDeterministic | kString | kILabelSorted |
-               kNotILabelSorted | kOLabelSorted | kNotOLabelSorted);
-    } else {
-      return outprops &
-             ~(kIDeterministic | kAcceptor | kString | kOLabelSorted |
-               kNotOLabelSorted);
-    }
-  } else {
-    // Shouldn't ever get here.
-    FSTERROR() << "RhoMatcher: Bad match type: " << match_type_;
-    return 0;
-  }
-}
-
-// For any requested label, this matcher considers all transitions
-// that match the label 'sigma_label' (sigma = "any"), and this in
-// additions to transitions with the requested label.  Each such sigma
-// transition found is returned with the sigma_label rewritten as the
-// requested label (both sides if an acceptor, or if 'rewrite_both' is
-// true and both input and output labels of the found transition are
-// 'sigma_label').  If 'sigma_label' is kNoLabel, this special
-// matching is not done.  SigmaMatcher is templated itself on a
-// matcher, which is used to perform the underlying matching.  By
-// default, the underlying matcher is constructed by SigmaMatcher.
-// The user can instead pass in this object; in that case,
-// SigmaMatcher takes its ownership.  No non-consuming symbols other
-// than epsilon supported with the underlying template argument matcher.
-template <class M>
-class SigmaMatcher : public MatcherBase<typename M::Arc> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  SigmaMatcher(const FST &fst, MatchType match_type,
-               Label sigma_label = kNoLabel,
-               MatcherRewriteMode rewrite_mode = MATCHER_REWRITE_AUTO,
-               M *matcher = nullptr)
-      : matcher_(matcher ? matcher : new M(fst, match_type)),
-        match_type_(match_type),
-        sigma_label_(sigma_label),
-        error_(false),
-        state_(kNoStateId) {
-    if (match_type == MATCH_BOTH) {
-      FSTERROR() << "SigmaMatcher: Bad match type";
-      match_type_ = MATCH_NONE;
-      error_ = true;
-    }
-    if (sigma_label == 0) {
-      FSTERROR() << "SigmaMatcher: 0 cannot be used as sigma_label";
-      sigma_label_ = kNoLabel;
-      error_ = true;
-    }
-    if (rewrite_mode == MATCHER_REWRITE_AUTO) {
-      rewrite_both_ = fst.Properties(kAcceptor, true);
-    } else if (rewrite_mode == MATCHER_REWRITE_ALWAYS) {
-      rewrite_both_ = true;
-    } else {
-      rewrite_both_ = false;
-    }
-  }
-
-  SigmaMatcher(const SigmaMatcher<M> &matcher, bool safe = false)
-      : matcher_(new M(*matcher.matcher_, safe)),
-        match_type_(matcher.match_type_),
-        sigma_label_(matcher.sigma_label_),
-        rewrite_both_(matcher.rewrite_both_),
-        error_(matcher.error_),
-        state_(kNoStateId) {}
-
-  SigmaMatcher<M> *Copy(bool safe = false) const override {
-    return new SigmaMatcher<M>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override { return matcher_->Type(test); }
-
-  void SetState(StateId s) final {
-    if (state_ == s) return;
-    state_ = s;
-    matcher_->SetState(s);
-    has_sigma_ =
-        (sigma_label_ != kNoLabel) ? matcher_->Find(sigma_label_) : false;
-  }
-
-  bool Find(Label match_label) final {
-    match_label_ = match_label;
-    if (match_label == sigma_label_ && sigma_label_ != kNoLabel) {
-      FSTERROR() << "SigmaMatcher::Find: bad label (sigma)";
-      error_ = true;
-      return false;
-    }
-    if (matcher_->Find(match_label)) {
-      sigma_match_ = kNoLabel;
-      return true;
-    } else if (has_sigma_ && match_label != 0 && match_label != kNoLabel &&
-               matcher_->Find(sigma_label_)) {
-      sigma_match_ = match_label;
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  bool Done() const final { return matcher_->Done(); }
-
-  const Arc &Value() const final {
-    if (sigma_match_ == kNoLabel) {
-      return matcher_->Value();
-    } else {
-      sigma_arc_ = matcher_->Value();
-      if (rewrite_both_) {
-        if (sigma_arc_.ilabel == sigma_label_) sigma_arc_.ilabel = sigma_match_;
-        if (sigma_arc_.olabel == sigma_label_) sigma_arc_.olabel = sigma_match_;
-      } else if (match_type_ == MATCH_INPUT) {
-        sigma_arc_.ilabel = sigma_match_;
-      } else {
-        sigma_arc_.olabel = sigma_match_;
-      }
-      return sigma_arc_;
-    }
-  }
-
-  void Next() final {
-    matcher_->Next();
-    if (matcher_->Done() && has_sigma_ && (sigma_match_ == kNoLabel) &&
-        (match_label_ > 0)) {
-      matcher_->Find(sigma_label_);
-      sigma_match_ = match_label_;
-    }
-  }
-
-  Weight Final(StateId s) const final { return matcher_->Final(s); }
-
-  ssize_t Priority(StateId s) final {
-    if (sigma_label_ != kNoLabel) {
-      SetState(s);
-      return has_sigma_ ? kRequirePriority : matcher_->Priority(s);
-    } else {
-      return matcher_->Priority(s);
-    }
-  }
-
-  const FST &GetFst() const override { return matcher_->GetFst(); }
-
-  uint64 Properties(uint64 props) const override;
-
-  uint32 Flags() const override {
-    if (sigma_label_ == kNoLabel || match_type_ == MATCH_NONE) {
-      return matcher_->Flags();
-    }
-    return matcher_->Flags() | kRequireMatch;
-  }
-
-  Label SigmaLabel() const { return sigma_label_; }
-
- private:
-  std::unique_ptr<M> matcher_;
-  MatchType match_type_;   // Type of match requested.
-  Label sigma_label_;      // Label that represents the sigma transition.
-  bool rewrite_both_;      // Rewrite both sides when both are sigma_label_?
-  bool has_sigma_;         // Are there sigmas at the current state?
-  Label sigma_match_;      // Current label that matches sigma transition.
-  mutable Arc sigma_arc_;  // Arc to return when sigma match.
-  Label match_label_;      // Label being matched.
-  bool error_;             // Error encountered?
-  StateId state_;          // Matcher state.
-};
-
-template <class M>
-inline uint64 SigmaMatcher<M>::Properties(uint64 inprops) const {
-  auto outprops = matcher_->Properties(inprops);
-  if (error_) outprops |= kError;
-  if (match_type_ == MATCH_NONE) {
-    return outprops;
-  } else if (rewrite_both_) {
-    return outprops &
-           ~(kIDeterministic | kNonIDeterministic | kODeterministic |
-             kNonODeterministic | kILabelSorted | kNotILabelSorted |
-             kOLabelSorted | kNotOLabelSorted | kString);
-  } else if (match_type_ == MATCH_INPUT) {
-    return outprops &
-           ~(kIDeterministic | kNonIDeterministic | kODeterministic |
-             kNonODeterministic | kILabelSorted | kNotILabelSorted | kString |
-             kAcceptor);
-  } else if (match_type_ == MATCH_OUTPUT) {
-    return outprops &
-           ~(kIDeterministic | kNonIDeterministic | kODeterministic |
-             kNonODeterministic | kOLabelSorted | kNotOLabelSorted | kString |
-             kAcceptor);
-  } else {
-    // Shouldn't ever get here.
-    FSTERROR() << "SigmaMatcher: Bad match type: " << match_type_;
-    return 0;
-  }
-}
-
-// Flags for MultiEpsMatcher.
-
-// Return multi-epsilon arcs for Find(kNoLabel).
-const uint32 kMultiEpsList = 0x00000001;
-
-// Return a kNolabel loop for Find(multi_eps).
-const uint32 kMultiEpsLoop = 0x00000002;
-
-// MultiEpsMatcher: allows treating multiple non-0 labels as
-// non-consuming labels in addition to 0 that is always
-// non-consuming. Precise behavior controlled by 'flags' argument. By
-// default, the underlying matcher is constructed by
-// MultiEpsMatcher. The user can instead pass in this object; in that
-// case, MultiEpsMatcher takes its ownership iff 'own_matcher' is
-// true.
-template <class M>
-class MultiEpsMatcher {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  MultiEpsMatcher(const FST &fst, MatchType match_type,
-                  uint32 flags = (kMultiEpsLoop | kMultiEpsList),
-                  M *matcher = nullptr, bool own_matcher = true)
-      : matcher_(matcher ? matcher : new M(fst, match_type)),
-        flags_(flags),
-        own_matcher_(matcher ? own_matcher : true) {
-    if (match_type == MATCH_INPUT) {
-      loop_.ilabel = kNoLabel;
-      loop_.olabel = 0;
-    } else {
-      loop_.ilabel = 0;
-      loop_.olabel = kNoLabel;
-    }
-    loop_.weight = Weight::One();
-    loop_.nextstate = kNoStateId;
-  }
-
-  MultiEpsMatcher(const MultiEpsMatcher<M> &matcher, bool safe = false)
-      : matcher_(new M(*matcher.matcher_, safe)),
-        flags_(matcher.flags_),
-        own_matcher_(true),
-        multi_eps_labels_(matcher.multi_eps_labels_),
-        loop_(matcher.loop_) {
-    loop_.nextstate = kNoStateId;
-  }
-
-  ~MultiEpsMatcher() {
-    if (own_matcher_) delete matcher_;
-  }
-
-  MultiEpsMatcher<M> *Copy(bool safe = false) const {
-    return new MultiEpsMatcher<M>(*this, safe);
-  }
-
-  MatchType Type(bool test) const { return matcher_->Type(test); }
-
-  void SetState(StateId state) {
-    matcher_->SetState(state);
-    loop_.nextstate = state;
-  }
-
-  bool Find(Label label);
-
-  bool Done() const { return done_; }
-
-  const Arc &Value() const { return current_loop_ ? loop_ : matcher_->Value(); }
-
-  void Next() {
-    if (!current_loop_) {
-      matcher_->Next();
-      done_ = matcher_->Done();
-      if (done_ && multi_eps_iter_ != multi_eps_labels_.End()) {
-        ++multi_eps_iter_;
-        while ((multi_eps_iter_ != multi_eps_labels_.End()) &&
-               !matcher_->Find(*multi_eps_iter_)) {
-          ++multi_eps_iter_;
-        }
-        if (multi_eps_iter_ != multi_eps_labels_.End()) {
-          done_ = false;
-        } else {
-          done_ = !matcher_->Find(kNoLabel);
-        }
-      }
-    } else {
-      done_ = true;
-    }
-  }
-
-  const FST &GetFst() const { return matcher_->GetFst(); }
-
-  uint64 Properties(uint64 props) const { return matcher_->Properties(props); }
-
-  const M *GetMatcher() const { return matcher_; }
-
-  Weight Final(StateId s) const { return matcher_->Final(s); }
-
-  uint32 Flags() const { return matcher_->Flags(); }
-
-  ssize_t Priority(StateId s) { return matcher_->Priority(s); }
-
-  void AddMultiEpsLabel(Label label) {
-    if (label == 0) {
-      FSTERROR() << "MultiEpsMatcher: Bad multi-eps label: 0";
-    } else {
-      multi_eps_labels_.Insert(label);
-    }
-  }
-
-  void RemoveMultiEpsLabel(Label label) {
-    if (label == 0) {
-      FSTERROR() << "MultiEpsMatcher: Bad multi-eps label: 0";
-    } else {
-      multi_eps_labels_.Erase(label);
-    }
-  }
-
-  void ClearMultiEpsLabels() { multi_eps_labels_.Clear(); }
-
- private:
-  M *matcher_;
-  uint32 flags_;
-  bool own_matcher_;  // Does this class delete the matcher?
-
-  // Multi-eps label set.
-  CompactSet<Label, kNoLabel> multi_eps_labels_;
-  typename CompactSet<Label, kNoLabel>::const_iterator multi_eps_iter_;
-
-  bool current_loop_;  // Current arc is the implicit loop?
-  mutable Arc loop_;   // For non-consuming symbols.
-  bool done_;          // Matching done?
-
-  MultiEpsMatcher &operator=(const MultiEpsMatcher &) = delete;
-};
-
-template <class M>
-inline bool MultiEpsMatcher<M>::Find(Label label) {
-  multi_eps_iter_ = multi_eps_labels_.End();
-  current_loop_ = false;
-  bool ret;
-  if (label == 0) {
-    ret = matcher_->Find(0);
-  } else if (label == kNoLabel) {
-    if (flags_ & kMultiEpsList) {
-      // Returns all non-consuming arcs (including epsilon).
-      multi_eps_iter_ = multi_eps_labels_.Begin();
-      while ((multi_eps_iter_ != multi_eps_labels_.End()) &&
-             !matcher_->Find(*multi_eps_iter_)) {
-        ++multi_eps_iter_;
-      }
-      if (multi_eps_iter_ != multi_eps_labels_.End()) {
-        ret = true;
-      } else {
-        ret = matcher_->Find(kNoLabel);
-      }
-    } else {
-      // Returns all epsilon arcs.
-      ret = matcher_->Find(kNoLabel);
-    }
-  } else if ((flags_ & kMultiEpsLoop) &&
-             multi_eps_labels_.Find(label) != multi_eps_labels_.End()) {
-    // Returns implicit loop.
-    current_loop_ = true;
-    ret = true;
-  } else {
-    ret = matcher_->Find(label);
-  }
-  done_ = !ret;
-  return ret;
-}
-
-// This class discards any implicit matches (e.g., the implicit epsilon
-// self-loops in the SortedMatcher). Matchers are most often used in
-// composition/intersection where the implicit matches are needed
-// e.g. for epsilon processing. However, if a matcher is simply being
-// used to look-up explicit label matches, this class saves the user
-// from having to check for and discard the unwanted implicit matches
-// themselves.
-template <class M>
-class ExplicitMatcher : public MatcherBase<typename M::Arc> {
- public:
-  using FST = typename M::FST;
-  using Arc = typename FST::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  ExplicitMatcher(const FST &fst, MatchType match_type, M *matcher = nullptr)
-      : matcher_(matcher ? matcher : new M(fst, match_type)),
-        match_type_(match_type),
-        error_(false) {}
-
-  ExplicitMatcher(const ExplicitMatcher<M> &matcher, bool safe = false)
-      : matcher_(new M(*matcher.matcher_, safe)),
-        match_type_(matcher.match_type_),
-        error_(matcher.error_) {}
-
-  ExplicitMatcher<M> *Copy(bool safe = false) const override {
-    return new ExplicitMatcher<M>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override { return matcher_->Type(test); }
-
-  void SetState(StateId s) final { matcher_->SetState(s); }
-
-  bool Find(Label label) final {
-    matcher_->Find(label);
-    CheckArc();
-    return !Done();
-  }
-
-  bool Done() const final { return matcher_->Done(); }
-
-  const Arc &Value() const final { return matcher_->Value(); }
-
-  void Next() final {
-    matcher_->Next();
-    CheckArc();
-  }
-
-  Weight Final(StateId s) const final { return matcher_->Final(s); }
-
-  ssize_t Priority(StateId s) final { return matcher_->Priority(s); }
-
-  const FST &GetFst() const final { return matcher_->GetFst(); }
-
-  uint64 Properties(uint64 inprops) const override {
-    return matcher_->Properties(inprops);
-  }
-
-  const M *GetMatcher() const { return matcher_.get(); }
-
-  uint32 Flags() const override { return matcher_->Flags(); }
-
- private:
-  // Checks current arc if available and explicit. If not available, stops. If
-  // not explicit, checks next ones.
-  void CheckArc() {
-    for (; !matcher_->Done(); matcher_->Next()) {
-      const auto label = match_type_ == MATCH_INPUT ? matcher_->Value().ilabel
-                                                    : matcher_->Value().olabel;
-      if (label != kNoLabel) return;
-    }
-  }
-
-  std::unique_ptr<M> matcher_;
-  MatchType match_type_;  // Type of match requested.
-  bool error_;            // Error encountered?
-};
-
-// Generic matcher, templated on the FST definition.
-//
-// Here is a typical use:
-//
-//   Matcher<StdFst> matcher(fst, MATCH_INPUT);
-//   matcher.SetState(state);
-//   if (matcher.Find(label))
-//     for (; !matcher.Done(); matcher.Next()) {
-//       auto &arc = matcher.Value();
-//       ...
-//     }
-template <class F>
-class Matcher {
- public:
-  using FST = F;
-  using Arc = typename F::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  Matcher(const FST &fst, MatchType match_type) {
-    base_.reset(fst.InitMatcher(match_type));
-    if (!base_) base_.reset(new SortedMatcher<FST>(fst, match_type));
-  }
-
-  Matcher(const Matcher<FST> &matcher, bool safe = false) {
-    base_.reset(matcher.base_->Copy(safe));
-  }
-
-  // Takes ownership of the provided matcher.
-  explicit Matcher(MatcherBase<Arc> *base_matcher) {
-    base_.reset(base_matcher);
-  }
-
-  Matcher<FST> *Copy(bool safe = false) const {
-    return new Matcher<FST>(*this, safe);
-  }
-
-  MatchType Type(bool test) const { return base_->Type(test); }
-
-  void SetState(StateId s) { base_->SetState(s); }
-
-  bool Find(Label label) { return base_->Find(label); }
-
-  bool Done() const { return base_->Done(); }
-
-  const Arc &Value() const { return base_->Value(); }
-
-  void Next() { base_->Next(); }
-
-  const FST &GetFst() const {
-    return static_cast<const FST &>(base_->GetFst());
-  }
-
-  uint64 Properties(uint64 props) const { return base_->Properties(props); }
-
-  Weight Final(StateId s) const { return base_->Final(s); }
-
-  uint32 Flags() const { return base_->Flags() & kMatcherFlags; }
-
-  ssize_t Priority(StateId s) { return base_->Priority(s); }
-
- private:
-  std::unique_ptr<MatcherBase<Arc>> base_;
-};
-
-}  // namespace fst
-
-#endif  // FST_MATCHER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/memory.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/memory.h
deleted file mode 100644
index 612fdcf2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/memory.h
+++ /dev/null
@@ -1,419 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// FST memory utilities.
-
-#ifndef FST_MEMORY_H_
-#define FST_MEMORY_H_
-
-#include <list>
-#include <memory>
-#include <utility>
-#include <vector>
-
-#include <fst/types.h>
-#include <fst/log.h>
-#include <fstream>
-
-namespace fst {
-
-// Default block allocation size.
-constexpr int kAllocSize = 64;
-
-// Minimum number of allocations per block.
-constexpr int kAllocFit = 4;
-
-// Base class for MemoryArena that allows (e.g.) MemoryArenaCollection to
-// easily manipulate collections of variously sized arenas.
-class MemoryArenaBase {
- public:
-  virtual ~MemoryArenaBase() {}
-  virtual size_t Size() const = 0;
-};
-
-// Allocates 'size' unintialized memory chunks of size sizeof(T) from underlying
-// blocks of (at least) size 'block_size * sizeof(T)'. All blocks are freed when
-// this class is deleted. Result of allocate() will be aligned to sizeof(T).
-template <typename T>
-class MemoryArena : public MemoryArenaBase {
- public:
-  explicit MemoryArena(size_t block_size = kAllocSize)
-      : block_size_(block_size * sizeof(T)), block_pos_(0) {
-    blocks_.emplace_front(new char[block_size_]);
-  }
-
-  void *Allocate(size_t size) {
-    const auto byte_size = size * sizeof(T);
-    if (byte_size * kAllocFit > block_size_) {
-      // Large block; adds new large block.
-      auto *ptr = new char[byte_size];
-      blocks_.emplace_back(ptr);
-      return ptr;
-    }
-    if (block_pos_ + byte_size > block_size_) {
-      // Doesn't fit; adds new standard block.
-      auto *ptr = new char[block_size_];
-      block_pos_ = 0;
-      blocks_.emplace_front(ptr);
-    }
-    // Fits; uses current block.
-    auto *ptr = blocks_.front().get() + block_pos_;
-    block_pos_ += byte_size;
-    return ptr;
-  }
-
-  size_t Size() const override { return sizeof(T); }
-
- private:
-  size_t block_size_;  // Default block size in bytes.
-  size_t block_pos_;   // Current position in block in bytes.
-  std::list<std::unique_ptr<char[]>> blocks_;  // List of allocated blocks.
-};
-
-// Base class for MemoryPool that allows (e.g.) MemoryPoolCollection to easily
-// manipulate collections of variously sized pools.
-class MemoryPoolBase {
- public:
-  virtual ~MemoryPoolBase() {}
-  virtual size_t Size() const = 0;
-};
-
-// Allocates and frees initially uninitialized memory chunks of size sizeof(T).
-// Keeps an internal list of freed chunks that are reused (as is) on the next
-// allocation if available. Chunks are constructed in blocks of size
-// 'pool_size'. All memory is freed when the class is deleted. The result of
-// Allocate() will be suitably memory-aligned.
-//
-// Combined with placement operator new and destroy fucntions for the T class,
-// this can be used to improve allocation efficiency. See nlp/fst/lib/visit.h
-// (global new) and nlp/fst/lib/dfs-visit.h (class new) for examples.
-template <typename T>
-class MemoryPool : public MemoryPoolBase {
- public:
-  struct Link {
-    char buf[sizeof(T)];
-    Link *next;
-  };
-
-  // 'pool_size' specifies the size of the initial pool and how it is extended.
-  explicit MemoryPool(size_t pool_size = kAllocSize)
-      : mem_arena_(pool_size), free_list_(nullptr) {}
-
-  void *Allocate() {
-    if (free_list_ == nullptr) {
-      auto *link = static_cast<Link *>(mem_arena_.Allocate(1));
-      link->next = nullptr;
-      return link;
-    } else {
-      auto *link = free_list_;
-      free_list_ = link->next;
-      return link;
-    }
-  }
-
-  void Free(void *ptr) {
-    if (ptr) {
-      auto *link = static_cast<Link *>(ptr);
-      link->next = free_list_;
-      free_list_ = link;
-    }
-  }
-
-  size_t Size() const override { return sizeof(T); }
-
- private:
-  MemoryArena<Link> mem_arena_;
-  Link *free_list_;
-
-  MemoryPool(const MemoryPool &) = delete;
-  MemoryPool &operator=(const MemoryPool &) = delete;
-};
-
-// Stores a collection of memory arenas.
-class MemoryArenaCollection {
- public:
-  // 'block_size' specifies the block size of the arenas.
-  explicit MemoryArenaCollection(size_t block_size = kAllocSize)
-      : block_size_(block_size), ref_count_(1) {}
-
-  template <typename T>
-  MemoryArena<T> *Arena() {
-    if (sizeof(T) >= arenas_.size()) arenas_.resize(sizeof(T) + 1);
-    MemoryArenaBase *arena = arenas_[sizeof(T)].get();
-    if (arena == nullptr) {
-      arena = new MemoryArena<T>(block_size_);
-      arenas_[sizeof(T)].reset(arena);
-    }
-    return static_cast<MemoryArena<T> *>(arena);
-  }
-
-  size_t BlockSize() const { return block_size_; }
-
-  size_t RefCount() const { return ref_count_; }
-
-  size_t IncrRefCount() { return ++ref_count_; }
-
-  size_t DecrRefCount() { return --ref_count_; }
-
- private:
-  size_t block_size_;
-  size_t ref_count_;
-  std::vector<std::unique_ptr<MemoryArenaBase>> arenas_;
-};
-
-// Stores a collection of memory pools
-class MemoryPoolCollection {
- public:
-  // 'pool_size' specifies the size of initial pool and how it is extended.
-  explicit MemoryPoolCollection(size_t pool_size = kAllocSize)
-      : pool_size_(pool_size), ref_count_(1) {}
-
-  template <typename T>
-  MemoryPool<T> *Pool() {
-    if (sizeof(T) >= pools_.size()) pools_.resize(sizeof(T) + 1);
-    MemoryPoolBase *pool = pools_[sizeof(T)].get();
-    if (pool == nullptr) {
-      pool = new MemoryPool<T>(pool_size_);
-      pools_[sizeof(T)].reset(pool);
-    }
-    return static_cast<MemoryPool<T> *>(pool);
-  }
-
-  size_t PoolSize() const { return pool_size_; }
-
-  size_t RefCount() const { return ref_count_; }
-
-  size_t IncrRefCount() { return ++ref_count_; }
-
-  size_t DecrRefCount() { return --ref_count_; }
-
- private:
-  size_t pool_size_;
-  size_t ref_count_;
-  std::vector<std::unique_ptr<MemoryPoolBase>> pools_;
-};
-
-// STL allocator using memory arenas. Memory is allocated from underlying
-// blocks of size 'block_size * sizeof(T)'. Memory is freed only when all
-// objects using this allocator are destroyed and there is otherwise no reuse
-// (unlike PoolAllocator).
-//
-// This allocator has object-local state so it should not be used with splicing
-// or swapping operations between objects created with different allocators nor
-// should it be used if copies must be thread-safe. The result of allocate()
-// will be suitably memory-aligned.
-template <typename T>
-class BlockAllocator {
- public:
-  using Allocator = std::allocator<T>;
-  using size_type = typename Allocator::size_type;
-  using difference_type = typename Allocator::difference_type;
-  using pointer = typename Allocator::pointer;
-  using const_pointer = typename Allocator::const_pointer;
-  using reference = typename Allocator::reference;
-  using const_reference = typename Allocator::const_reference;
-  using value_type = typename Allocator::value_type;
-
-  template <typename U>
-  struct rebind {
-    using other = BlockAllocator<U>;
-  };
-
-  explicit BlockAllocator(size_t block_size = kAllocSize)
-      : arenas_(new MemoryArenaCollection(block_size)) {}
-
-  BlockAllocator(const BlockAllocator<T> &arena_alloc)
-      : arenas_(arena_alloc.Arenas()) {
-    Arenas()->IncrRefCount();
-  }
-
-  template <typename U>
-  explicit BlockAllocator(const BlockAllocator<U> &arena_alloc)
-      : arenas_(arena_alloc.Arenas()) {
-    Arenas()->IncrRefCount();
-  }
-
-  ~BlockAllocator() {
-    if (Arenas()->DecrRefCount() == 0) delete Arenas();
-  }
-
-  pointer address(reference ref) const { return Allocator().address(ref); }
-
-  const_pointer address(const_reference ref) const {
-    return Allocator().address(ref);
-  }
-
-  size_type max_size() const { return Allocator().max_size(); }
-
-  template <class U, class... Args>
-  void construct(U *p, Args &&... args) {
-    Allocator().construct(p, std::forward<Args>(args)...);
-  }
-
-  void destroy(pointer p) { Allocator().destroy(p); }
-
-  pointer allocate(size_type n, const void *hint = nullptr) {
-    if (n * kAllocFit <= kAllocSize) {
-      return static_cast<pointer>(Arena()->Allocate(n));
-    } else {
-      return Allocator().allocate(n, hint);
-    }
-  }
-
-  void deallocate(pointer p, size_type n) {
-    if (n * kAllocFit > kAllocSize) Allocator().deallocate(p, n);
-  }
-
-  MemoryArenaCollection *Arenas() const { return arenas_; }
-
- private:
-  MemoryArena<T> *Arena() { return arenas_->Arena<T>(); }
-
-  MemoryArenaCollection *arenas_;
-
-  BlockAllocator<T> operator=(const BlockAllocator<T> &);
-};
-
-template <typename T, typename U>
-bool operator==(const BlockAllocator<T> &alloc1,
-                const BlockAllocator<U> &alloc2) {
-  return false;
-}
-
-template <typename T, typename U>
-bool operator!=(const BlockAllocator<T> &alloc1,
-                const BlockAllocator<U> &alloc2) {
-  return true;
-}
-
-// STL allocator using memory pools. Memory is allocated from underlying
-// blocks of size 'block_size * sizeof(T)'. Keeps an internal list of freed
-// chunks thare are reused on the next allocation.
-//
-// This allocator has object-local state so it should not be used with splicing
-// or swapping operations between objects created with different allocators nor
-// should it be used if copies must be thread-safe. The result of allocate()
-// will be suitably memory-aligned.
-template <typename T>
-class PoolAllocator {
- public:
-  using Allocator = std::allocator<T>;
-  using size_type = typename Allocator::size_type;
-  using difference_type = typename Allocator::difference_type;
-  using pointer = typename Allocator::pointer;
-  using const_pointer = typename Allocator::const_pointer;
-  using reference = typename Allocator::reference;
-  using const_reference = typename Allocator::const_reference;
-  using value_type = typename Allocator::value_type;
-
-  template <typename U>
-  struct rebind {
-    using other = PoolAllocator<U>;
-  };
-
-  explicit PoolAllocator(size_t pool_size = kAllocSize)
-      : pools_(new MemoryPoolCollection(pool_size)) {}
-
-  PoolAllocator(const PoolAllocator<T> &pool_alloc)
-      : pools_(pool_alloc.Pools()) {
-    Pools()->IncrRefCount();
-  }
-
-  template <typename U>
-  explicit PoolAllocator(const PoolAllocator<U> &pool_alloc)
-      : pools_(pool_alloc.Pools()) {
-    Pools()->IncrRefCount();
-  }
-
-  ~PoolAllocator() {
-    if (Pools()->DecrRefCount() == 0) delete Pools();
-  }
-
-  pointer address(reference ref) const { return Allocator().address(ref); }
-
-  const_pointer address(const_reference ref) const {
-    return Allocator().address(ref);
-  }
-
-  size_type max_size() const { return Allocator().max_size(); }
-
-  template <class U, class... Args>
-  void construct(U *p, Args &&... args) {
-    Allocator().construct(p, std::forward<Args>(args)...);
-  }
-
-  void destroy(pointer p) { Allocator().destroy(p); }
-
-  pointer allocate(size_type n, const void *hint = nullptr) {
-    if (n == 1) {
-      return static_cast<pointer>(Pool<1>()->Allocate());
-    } else if (n == 2) {
-      return static_cast<pointer>(Pool<2>()->Allocate());
-    } else if (n <= 4) {
-      return static_cast<pointer>(Pool<4>()->Allocate());
-    } else if (n <= 8) {
-      return static_cast<pointer>(Pool<8>()->Allocate());
-    } else if (n <= 16) {
-      return static_cast<pointer>(Pool<16>()->Allocate());
-    } else if (n <= 32) {
-      return static_cast<pointer>(Pool<32>()->Allocate());
-    } else if (n <= 64) {
-      return static_cast<pointer>(Pool<64>()->Allocate());
-    } else {
-      return Allocator().allocate(n, hint);
-    }
-  }
-
-  void deallocate(pointer p, size_type n) {
-    if (n == 1) {
-      Pool<1>()->Free(p);
-    } else if (n == 2) {
-      Pool<2>()->Free(p);
-    } else if (n <= 4) {
-      Pool<4>()->Free(p);
-    } else if (n <= 8) {
-      Pool<8>()->Free(p);
-    } else if (n <= 16) {
-      Pool<16>()->Free(p);
-    } else if (n <= 32) {
-      Pool<32>()->Free(p);
-    } else if (n <= 64) {
-      Pool<64>()->Free(p);
-    } else {
-      Allocator().deallocate(p, n);
-    }
-  }
-
-  MemoryPoolCollection *Pools() const { return pools_; }
-
- private:
-  template <int n>
-  struct TN {
-    T buf[n];
-  };
-
-  template <int n>
-  MemoryPool<TN<n>> *Pool() {
-    return pools_->Pool<TN<n>>();
-  }
-
-  MemoryPoolCollection *pools_;
-
-  PoolAllocator<T> operator=(const PoolAllocator<T> &);
-};
-
-template <typename T, typename U>
-bool operator==(const PoolAllocator<T> &alloc1,
-                const PoolAllocator<U> &alloc2) {
-  return false;
-}
-
-template <typename T, typename U>
-bool operator!=(const PoolAllocator<T> &alloc1,
-                const PoolAllocator<U> &alloc2) {
-  return true;
-}
-
-}  // namespace fst
-
-#endif  // FST_MEMORY_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/minimize.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/minimize.h
deleted file mode 100644
index d68c43a7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/minimize.h
+++ /dev/null
@@ -1,567 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to minimize an FST.
-
-#ifndef FST_MINIMIZE_H_
-#define FST_MINIMIZE_H_
-
-#include <cmath>
-
-#include <algorithm>
-#include <map>
-#include <queue>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arcsort.h>
-#include <fst/connect.h>
-#include <fst/dfs-visit.h>
-#include <fst/encode.h>
-#include <fst/factor-weight.h>
-#include <fst/fst.h>
-#include <fst/mutable-fst.h>
-#include <fst/partition.h>
-#include <fst/push.h>
-#include <fst/queue.h>
-#include <fst/reverse.h>
-#include <fst/state-map.h>
-
-
-namespace fst {
-namespace internal {
-
-// Comparator for creating partition.
-template <class Arc>
-class StateComparator {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  StateComparator(const Fst<Arc> &fst, const Partition<StateId> &partition)
-      : fst_(fst), partition_(partition) {}
-
-  // Compares state x with state y based on sort criteria.
-  bool operator()(const StateId x, const StateId y) const {
-    // Checks for final state equivalence.
-    const auto xfinal = fst_.Final(x).Hash();
-    const auto yfinal = fst_.Final(y).Hash();
-    if (xfinal < yfinal) {
-      return true;
-    } else if (xfinal > yfinal) {
-      return false;
-    }
-    // Checks for number of arcs.
-    if (fst_.NumArcs(x) < fst_.NumArcs(y)) return true;
-    if (fst_.NumArcs(x) > fst_.NumArcs(y)) return false;
-    // If the number of arcs are equal, checks for arc match.
-    for (ArcIterator<Fst<Arc>> aiter1(fst_, x), aiter2(fst_, y);
-         !aiter1.Done() && !aiter2.Done(); aiter1.Next(), aiter2.Next()) {
-      const auto &arc1 = aiter1.Value();
-      const auto &arc2 = aiter2.Value();
-      if (arc1.ilabel < arc2.ilabel) return true;
-      if (arc1.ilabel > arc2.ilabel) return false;
-      if (partition_.ClassId(arc1.nextstate) <
-          partition_.ClassId(arc2.nextstate))
-        return true;
-      if (partition_.ClassId(arc1.nextstate) >
-          partition_.ClassId(arc2.nextstate))
-        return false;
-    }
-    return false;
-  }
-
- private:
-  const Fst<Arc> &fst_;
-  const Partition<StateId> &partition_;
-};
-
-// Computes equivalence classes for cyclic unweighted acceptors. For cyclic
-// minimization we use the classic Hopcroft minimization algorithm, which has
-// complexity O(E log V) where E is the number of arcs and V is the number of
-// states.
-//
-// For more information, see:
-//
-//  Hopcroft, J. 1971. An n Log n algorithm for minimizing states in a finite
-//  automaton. Ms, Stanford University.
-//
-// Note: the original presentation of the paper was for a finite automaton (==
-// deterministic, unweighted acceptor), but we also apply it to the
-// nondeterministic case, where it is also applicable as long as the semiring is
-// idempotent (if the semiring is not idempotent, there are some complexities
-// in keeping track of the weight when there are multiple arcs to states that
-// will be merged, and we don't deal with this).
-template <class Arc, class Queue>
-class CyclicMinimizer {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using ClassId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using RevArc = ReverseArc<Arc>;
-
-  explicit CyclicMinimizer(const ExpandedFst<Arc> &fst) {
-    Initialize(fst);
-    Compute(fst);
-  }
-
-  const Partition<StateId> &GetPartition() const { return P_; }
-
- private:
-  // StateILabelHasher is a hashing object that computes a hash-function
-  // of an FST state that depends only on the set of ilabels on arcs leaving
-  // the state [note: it assumes that the arcs are ilabel-sorted].
-  // In order to work correctly for non-deterministic automata, multiple
-  // instances of the same ilabel count the same as a single instance.
-  class StateILabelHasher {
-   public:
-    explicit StateILabelHasher(const Fst<Arc> &fst) : fst_(fst) {}
-
-    using Label = typename Arc::Label;
-    using StateId = typename Arc::StateId;
-
-    size_t operator()(const StateId s) {
-      const size_t p1 = 7603;
-      const size_t p2 = 433024223;
-      size_t result = p2;
-      size_t current_ilabel = kNoLabel;
-      for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
-        Label this_ilabel = aiter.Value().ilabel;
-        if (this_ilabel != current_ilabel) {  // Ignores repeats.
-          result = p1 * result + this_ilabel;
-          current_ilabel = this_ilabel;
-        }
-      }
-      return result;
-    }
-
-   private:
-    const Fst<Arc> &fst_;
-  };
-
-  class ArcIterCompare {
-   public:
-    explicit ArcIterCompare(const Partition<StateId> &partition)
-        : partition_(partition) {}
-
-    ArcIterCompare(const ArcIterCompare &comp) : partition_(comp.partition_) {}
-
-    // Compares two iterators based on their input labels.
-    bool operator()(const ArcIterator<Fst<RevArc>> *x,
-                    const ArcIterator<Fst<RevArc>> *y) const {
-      const auto &xarc = x->Value();
-      const auto &yarc = y->Value();
-      return xarc.ilabel > yarc.ilabel;
-    }
-
-   private:
-    const Partition<StateId> &partition_;
-  };
-
-  using ArcIterQueue =
-      std::priority_queue<ArcIterator<Fst<RevArc>> *,
-                          std::vector<ArcIterator<Fst<RevArc>> *>,
-                          ArcIterCompare>;
-
- private:
-  // Prepartitions the space into equivalence classes. We ensure that final and
-  // non-final states always go into different equivalence classes, and we use
-  // class StateILabelHasher to make sure that most of the time, states with
-  // different sets of ilabels on arcs leaving them, go to different partitions.
-  // Note: for the O(n) guarantees we don't rely on the goodness of this
-  // hashing function---it just provides a bonus speedup.
-  void PrePartition(const ExpandedFst<Arc> &fst) {
-    VLOG(5) << "PrePartition";
-    StateId next_class = 0;
-    auto num_states = fst.NumStates();
-    // Allocates a temporary vector to store the initial class mappings, so that
-    // we can allocate the classes all at once.
-    std::vector<StateId> state_to_initial_class(num_states);
-    {
-      // We maintain two maps from hash-value to class---one for final states
-      // (final-prob == One()) and one for non-final states
-      // (final-prob == Zero()). We are processing unweighted acceptors, so the
-      // are the only two possible values.
-      using HashToClassMap = std::unordered_map<size_t, StateId>;
-      HashToClassMap hash_to_class_nonfinal;
-      HashToClassMap hash_to_class_final;
-      StateILabelHasher hasher(fst);
-      for (StateId s = 0; s < num_states; ++s) {
-        size_t hash = hasher(s);
-        HashToClassMap &this_map =
-            (fst.Final(s) != Weight::Zero() ? hash_to_class_final
-                                            : hash_to_class_nonfinal);
-        // Avoids two map lookups by using 'insert' instead of 'find'.
-        auto p = this_map.insert(std::make_pair(hash, next_class));
-        state_to_initial_class[s] = p.second ? next_class++ : p.first->second;
-      }
-      // Lets the unordered_maps go out of scope before we allocate the classes,
-      // to reduce the maximum amount of memory used.
-    }
-    P_.AllocateClasses(next_class);
-    for (StateId s = 0; s < num_states; ++s) {
-      P_.Add(s, state_to_initial_class[s]);
-    }
-    for (StateId c = 0; c < next_class; ++c) L_.Enqueue(c);
-    VLOG(5) << "Initial Partition: " << P_.NumClasses();
-  }
-
-  // Creates inverse transition Tr_ = rev(fst), loops over states in FST and
-  // splits on final, creating two blocks in the partition corresponding to
-  // final, non-final.
-  void Initialize(const ExpandedFst<Arc> &fst) {
-    // Constructs Tr.
-    Reverse(fst, &Tr_);
-    ILabelCompare<RevArc> ilabel_comp;
-    ArcSort(&Tr_, ilabel_comp);
-    // Tells the partition how many elements to allocate. The first state in
-    // Tr_ is super-final state.
-    P_.Initialize(Tr_.NumStates() - 1);
-    // Prepares initial partition.
-    PrePartition(fst);
-    // Allocates arc iterator queue.
-    ArcIterCompare comp(P_);
-    aiter_queue_.reset(new ArcIterQueue(comp));
-  }
-  // Partitions all classes with destination C.
-  void Split(ClassId C) {
-    // Prepares priority queue: opens arc iterator for each state in C, and
-    // inserts into priority queue.
-    for (PartitionIterator<StateId> siter(P_, C); !siter.Done(); siter.Next()) {
-      StateId s = siter.Value();
-      if (Tr_.NumArcs(s + 1)) {
-        aiter_queue_->push(new ArcIterator<Fst<RevArc>>(Tr_, s + 1));
-      }
-    }
-    // Now pops arc iterator from queue, splits entering equivalence class, and
-    // re-inserts updated iterator into queue.
-    Label prev_label = -1;
-    while (!aiter_queue_->empty()) {
-      std::unique_ptr<ArcIterator<Fst<RevArc>>> aiter(aiter_queue_->top());
-      aiter_queue_->pop();
-      if (aiter->Done()) continue;
-      const auto &arc = aiter->Value();
-      auto from_state = aiter->Value().nextstate - 1;
-      auto from_label = arc.ilabel;
-      if (prev_label != from_label) P_.FinalizeSplit(&L_);
-      auto from_class = P_.ClassId(from_state);
-      if (P_.ClassSize(from_class) > 1) P_.SplitOn(from_state);
-      prev_label = from_label;
-      aiter->Next();
-      if (!aiter->Done()) aiter_queue_->push(aiter.release());
-    }
-    P_.FinalizeSplit(&L_);
-  }
-
-  // Main loop for Hopcroft minimization.
-  void Compute(const Fst<Arc> &fst) {
-    // Processes active classes (FIFO, or FILO).
-    while (!L_.Empty()) {
-      const auto C = L_.Head();
-      L_.Dequeue();
-      Split(C);  // Splits on C, all labels in C.
-    }
-  }
-
- private:
-  // Partioning of states into equivalence classes.
-  Partition<StateId> P_;
-  // Set of active classes to be processed in partition P.
-  Queue L_;
-  // Reverses transition function.
-  VectorFst<RevArc> Tr_;
-  // Priority queue of open arc iterators for all states in the splitter
-  // equivalence class.
-  std::unique_ptr<ArcIterQueue> aiter_queue_;
-};
-
-// Computes equivalence classes for acyclic FST.
-//
-// Complexity:
-//
-//   O(E)
-//
-// where E is the number of arcs.
-//
-// For more information, see:
-//
-// Revuz, D. 1992. Minimization of acyclic deterministic automata in linear
-// time. Theoretical Computer Science 92(1): 181-189.
-template <class Arc>
-class AcyclicMinimizer {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using ClassId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit AcyclicMinimizer(const ExpandedFst<Arc> &fst) {
-    Initialize(fst);
-    Refine(fst);
-  }
-
-  const Partition<StateId> &GetPartition() { return partition_; }
-
- private:
-  // DFS visitor to compute the height (distance) to final state.
-  class HeightVisitor {
-   public:
-    HeightVisitor() : max_height_(0), num_states_(0) {}
-
-    // Invoked before DFS visit.
-    void InitVisit(const Fst<Arc> &fst) {}
-
-    // Invoked when state is discovered (2nd arg is DFS tree root).
-    bool InitState(StateId s, StateId root) {
-      // Extends height array and initialize height (distance) to 0.
-      for (StateId i = height_.size(); i <= s; ++i) height_.push_back(-1);
-      if (s >= num_states_) num_states_ = s + 1;
-      return true;
-    }
-
-    // Invoked when tree arc examined (to undiscovered state).
-    bool TreeArc(StateId s, const Arc &arc) { return true; }
-
-    // Invoked when back arc examined (to unfinished state).
-    bool BackArc(StateId s, const Arc &arc) { return true; }
-
-    // Invoked when forward or cross arc examined (to finished state).
-    bool ForwardOrCrossArc(StateId s, const Arc &arc) {
-      if (height_[arc.nextstate] + 1 > height_[s]) {
-        height_[s] = height_[arc.nextstate] + 1;
-      }
-      return true;
-    }
-
-    // Invoked when state finished (parent is kNoStateId for tree root).
-    void FinishState(StateId s, StateId parent, const Arc *parent_arc) {
-      if (height_[s] == -1) height_[s] = 0;
-      const auto h = height_[s] + 1;
-      if (parent >= 0) {
-        if (h > height_[parent]) height_[parent] = h;
-        if (h > max_height_) max_height_ = h;
-      }
-    }
-
-    // Invoked after DFS visit.
-    void FinishVisit() {}
-
-    size_t max_height() const { return max_height_; }
-
-    const std::vector<StateId> &height() const { return height_; }
-
-    size_t num_states() const { return num_states_; }
-
-   private:
-    std::vector<StateId> height_;
-    size_t max_height_;
-    size_t num_states_;
-  };
-
- private:
-  // Cluster states according to height (distance to final state)
-  void Initialize(const Fst<Arc> &fst) {
-    // Computes height (distance to final state).
-    HeightVisitor hvisitor;
-    DfsVisit(fst, &hvisitor);
-    // Creates initial partition based on height.
-    partition_.Initialize(hvisitor.num_states());
-    partition_.AllocateClasses(hvisitor.max_height() + 1);
-    const auto &hstates = hvisitor.height();
-    for (StateId s = 0; s < hstates.size(); ++s) partition_.Add(s, hstates[s]);
-  }
-
-  // Refines states based on arc sort (out degree, arc equivalence).
-  void Refine(const Fst<Arc> &fst) {
-    using EquivalenceMap = std::map<StateId, StateId, StateComparator<Arc>>;
-    StateComparator<Arc> comp(fst, partition_);
-    // Starts with tail (height = 0).
-    auto height = partition_.NumClasses();
-    for (StateId h = 0; h < height; ++h) {
-      EquivalenceMap equiv_classes(comp);
-      // Sorts states within equivalence class.
-      PartitionIterator<StateId> siter(partition_, h);
-      equiv_classes[siter.Value()] = h;
-      for (siter.Next(); !siter.Done(); siter.Next()) {
-        auto insert_result =
-            equiv_classes.insert(std::make_pair(siter.Value(), kNoStateId));
-        if (insert_result.second) {
-          insert_result.first->second = partition_.AddClass();
-        }
-      }
-      // Creates refined partition.
-      for (siter.Reset(); !siter.Done();) {
-        const auto s = siter.Value();
-        const auto old_class = partition_.ClassId(s);
-        const auto new_class = equiv_classes[s];
-        // A move operation can invalidate the iterator, so we first update
-        // the iterator to the next element before we move the current element
-        // out of the list.
-        siter.Next();
-        if (old_class != new_class) partition_.Move(s, new_class);
-      }
-    }
-  }
-
- private:
-  Partition<StateId> partition_;
-};
-
-// Given a partition and a Mutable FST, merges states of Fst in place (i.e.,
-// destructively). Merging works by taking the first state in a class of the
-// partition to be the representative state for the class. Each arc is then
-// reconnected to this state. All states in the class are merged by adding
-// their arcs to the representative state.
-template <class Arc>
-void MergeStates(const Partition<typename Arc::StateId> &partition,
-                 MutableFst<Arc> *fst) {
-  using StateId = typename Arc::StateId;
-  std::vector<StateId> state_map(partition.NumClasses());
-  for (StateId i = 0; i < partition.NumClasses(); ++i) {
-    PartitionIterator<StateId> siter(partition, i);
-    state_map[i] = siter.Value();  // First state in partition.
-  }
-  // Relabels destination states.
-  for (StateId c = 0; c < partition.NumClasses(); ++c) {
-    for (PartitionIterator<StateId> siter(partition, c); !siter.Done();
-         siter.Next()) {
-      const auto s = siter.Value();
-      for (MutableArcIterator<MutableFst<Arc>> aiter(fst, s); !aiter.Done();
-           aiter.Next()) {
-        auto arc = aiter.Value();
-        arc.nextstate = state_map[partition.ClassId(arc.nextstate)];
-        if (s == state_map[c]) {  // For the first state, just sets destination.
-          aiter.SetValue(arc);
-        } else {
-          fst->AddArc(state_map[c], arc);
-        }
-      }
-    }
-  }
-  fst->SetStart(state_map[partition.ClassId(fst->Start())]);
-  Connect(fst);
-}
-
-template <class Arc>
-void AcceptorMinimize(MutableFst<Arc> *fst,
-                      bool allow_acyclic_minimization = true) {
-  if (!(fst->Properties(kAcceptor | kUnweighted, true) ==
-        (kAcceptor | kUnweighted))) {
-    FSTERROR() << "FST is not an unweighted acceptor";
-    fst->SetProperties(kError, kError);
-    return;
-  }
-  // Connects FST before minimization, handles disconnected states.
-  Connect(fst);
-  if (fst->NumStates() == 0) return;
-  if (allow_acyclic_minimization && fst->Properties(kAcyclic, true)) {
-    // Acyclic minimization (Revuz).
-    VLOG(2) << "Acyclic minimization";
-    ArcSort(fst, ILabelCompare<Arc>());
-    AcyclicMinimizer<Arc> minimizer(*fst);
-    MergeStates(minimizer.GetPartition(), fst);
-  } else {
-    // Either the FST has cycles, or it's generated from non-deterministic input
-    // (which the Revuz algorithm can't handle), so use the cyclic minimization
-    // algorithm of Hopcroft.
-    VLOG(2) << "Cyclic minimization";
-    CyclicMinimizer<Arc, LifoQueue<typename Arc::StateId>> minimizer(*fst);
-    MergeStates(minimizer.GetPartition(), fst);
-  }
-  // Merges in appropriate semiring
-  ArcUniqueMapper<Arc> mapper(*fst);
-  StateMap(fst, mapper);
-}
-
-}  // namespace internal
-
-// In place minimization of deterministic weighted automata and transducers,
-// and also non-deterministic ones if they use an idempotent semiring.
-// For transducers, if the 'sfst' argument is not null, the algorithm
-// produces a compact factorization of the minimal transducer.
-//
-// In the acyclic deterministic case, we use an algorithm from Revuz that is
-// linear in the number of arcs (edges) in the machine.
-//
-// In the cyclic or non-deterministic case, we use the classical Hopcroft
-// minimization (which was presented for the deterministic case but which
-// also works for non-deterministic FSTs); this has complexity O(e log v).
-//
-template <class Arc>
-void Minimize(MutableFst<Arc> *fst, MutableFst<Arc> *sfst = nullptr,
-              float delta = kDelta, bool allow_nondet = false) {
-  using Weight = typename Arc::Weight;
-  const auto props = fst->Properties(
-      kAcceptor | kIDeterministic | kWeighted | kUnweighted, true);
-  bool allow_acyclic_minimization;
-  if (props & kIDeterministic) {
-    allow_acyclic_minimization = true;
-  } else {
-    // Our approach to minimization of non-deterministic FSTs will only work in
-    // idempotent semirings---for non-deterministic inputs, a state could have
-    // multiple transitions to states that will get merged, and we'd have to
-    // sum their weights. The algorithm doesn't handle that.
-    if (!(Weight::Properties() & kIdempotent)) {
-      fst->SetProperties(kError, kError);
-      FSTERROR() << "Cannot minimize a non-deterministic FST over a "
-                    "non-idempotent semiring";
-      return;
-    } else if (!allow_nondet) {
-      fst->SetProperties(kError, kError);
-      FSTERROR() << "Refusing to minimize a non-deterministic FST with "
-                 << "allow_nondet = false";
-      return;
-    }
-    // The Revuz algorithm won't work for nondeterministic inputs, so if the
-    // input is nondeterministic, we'll have to pass a bool saying not to use
-    // that algorithm. We check at this level rather than in AcceptorMinimize(),
-    // because it's possible that the FST at this level could be deterministic,
-    // but a harmless type of non-determinism could be introduced by Encode()
-    // (thanks to kEncodeWeights, if the FST has epsilons and has a final
-    // weight with weights equal to some epsilon arc.)
-    allow_acyclic_minimization = false;
-  }
-  if (!(props & kAcceptor)) {  // Weighted transducer.
-    VectorFst<GallicArc<Arc, GALLIC_LEFT>> gfst;
-    ArcMap(*fst, &gfst, ToGallicMapper<Arc, GALLIC_LEFT>());
-    fst->DeleteStates();
-    gfst.SetProperties(kAcceptor, kAcceptor);
-    Push(&gfst, REWEIGHT_TO_INITIAL, delta);
-    ArcMap(&gfst, QuantizeMapper<GallicArc<Arc, GALLIC_LEFT>>(delta));
-    EncodeMapper<GallicArc<Arc, GALLIC_LEFT>> encoder(
-        kEncodeLabels | kEncodeWeights, ENCODE);
-    Encode(&gfst, &encoder);
-    internal::AcceptorMinimize(&gfst, allow_acyclic_minimization);
-    Decode(&gfst, encoder);
-    if (!sfst) {
-      FactorWeightFst<GallicArc<Arc, GALLIC_LEFT>,
-                      GallicFactor<typename Arc::Label, Weight, GALLIC_LEFT>>
-          fwfst(gfst);
-      std::unique_ptr<SymbolTable> osyms(
-          fst->OutputSymbols() ? fst->OutputSymbols()->Copy() : nullptr);
-      ArcMap(fwfst, fst, FromGallicMapper<Arc, GALLIC_LEFT>());
-      fst->SetOutputSymbols(osyms.get());
-    } else {
-      sfst->SetOutputSymbols(fst->OutputSymbols());
-      GallicToNewSymbolsMapper<Arc, GALLIC_LEFT> mapper(sfst);
-      ArcMap(gfst, fst, &mapper);
-      fst->SetOutputSymbols(sfst->InputSymbols());
-    }
-  } else if (props & kWeighted) {  // Weighted acceptor.
-    Push(fst, REWEIGHT_TO_INITIAL, delta);
-    ArcMap(fst, QuantizeMapper<Arc>(delta));
-    EncodeMapper<Arc> encoder(kEncodeLabels | kEncodeWeights, ENCODE);
-    Encode(fst, &encoder);
-    internal::AcceptorMinimize(fst, allow_acyclic_minimization);
-    Decode(fst, encoder);
-  } else {  // Unweighted acceptor.
-    internal::AcceptorMinimize(fst, allow_acyclic_minimization);
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_MINIMIZE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/mutable-fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/mutable-fst.h
deleted file mode 100644
index cf15da9a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/mutable-fst.h
+++ /dev/null
@@ -1,389 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Expanded FST augmented with mutators; interface class definition and
-// mutable arc iterator interface.
-
-#ifndef FST_MUTABLE_FST_H_
-#define FST_MUTABLE_FST_H_
-
-#include <stddef.h>
-#include <sys/types.h>
-
-#include <istream>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-#include <fstream>
-
-#include <fst/expanded-fst.h>
-
-
-namespace fst {
-
-template <class Arc>
-struct MutableArcIteratorData;
-
-// Abstract interface for an expanded FST which also supports mutation
-// operations. To modify arcs, use MutableArcIterator.
-template <class A>
-class MutableFst : public ExpandedFst<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  virtual MutableFst<Arc> &operator=(const Fst<Arc> &fst) = 0;
-
-  MutableFst<Arc> &operator=(const MutableFst<Arc> &fst) {
-    return operator=(static_cast<const Fst<Arc> &>(fst));
-  }
-
-  // Sets the initial state.
-  virtual void SetStart(StateId) = 0;
-
-  // Sets a state's final weight.
-  virtual void SetFinal(StateId, Weight) = 0;
-
-  // Sets property bits w.r.t. mask.
-  virtual void SetProperties(uint64 props, uint64 mask) = 0;
-
-  // Adds a state and returns its ID.
-  virtual StateId AddState() = 0;
-
-  // Adds an arc to state.
-  virtual void AddArc(StateId, const Arc &arc) = 0;
-
-  // Deletes some states, preserving original StateId ordering.
-  virtual void DeleteStates(const std::vector<StateId> &) = 0;
-
-  // Delete all states.
-  virtual void DeleteStates() = 0;
-
-  // Delete some arcs at a given state.
-  virtual void DeleteArcs(StateId, size_t n) = 0;
-
-  // Delete all arcs at a given state.
-  virtual void DeleteArcs(StateId) = 0;
-
-  // Optional, best effort only.
-  virtual void ReserveStates(StateId n) {}
-
-  // Optional, best effort only.
-  virtual void ReserveArcs(StateId s, size_t n) {}
-
-  // Returns input label symbol table or nullptr if not specified.
-  const SymbolTable *InputSymbols() const override = 0;
-
-  // Returns output label symbol table or nullptr if not specified.
-  const SymbolTable *OutputSymbols() const override = 0;
-
-  // Returns input label symbol table or nullptr if not specified.
-  virtual SymbolTable *MutableInputSymbols() = 0;
-
-  // Returns output label symbol table or nullptr if not specified.
-  virtual SymbolTable *MutableOutputSymbols() = 0;
-
-  // Sets input label symbol table; pass nullptr to delete table.
-  virtual void SetInputSymbols(const SymbolTable *isyms) = 0;
-
-  // Sets output label symbol table; pass nullptr to delete table.
-  virtual void SetOutputSymbols(const SymbolTable *osyms) = 0;
-
-  // Gets a copy of this MutableFst. See Fst<>::Copy() for further doc.
-  MutableFst<A> *Copy(bool safe = false) const override = 0;
-
-  // Reads a MutableFst from an input stream, returning nullptr on error.
-  static MutableFst<Arc> *Read(std::istream &strm, const FstReadOptions &opts) {
-    FstReadOptions ropts(opts);
-    FstHeader hdr;
-    if (ropts.header) {
-      hdr = *opts.header;
-    } else {
-      if (!hdr.Read(strm, opts.source)) return nullptr;
-      ropts.header = &hdr;
-    }
-    if (!(hdr.Properties() & kMutable)) {
-      LOG(ERROR) << "MutableFst::Read: Not a MutableFst: " << ropts.source;
-      return nullptr;
-    }
-    const auto &fst_type = hdr.FstType();
-    const auto reader = FstRegister<Arc>::GetRegister()->GetReader(fst_type);
-    if (!reader) {
-      LOG(ERROR) << "MutableFst::Read: Unknown FST type \"" << fst_type
-                 << "\" (arc type = \"" << A::Type() << "\"): " << ropts.source;
-      return nullptr;
-    }
-    auto *fst = reader(strm, ropts);
-    if (!fst) return nullptr;
-    return static_cast<MutableFst<Arc> *>(fst);
-  }
-
-  // Reads a MutableFst from a file; returns nullptr on error. An empty
-  // filename results in reading from standard input. If convert is true,
-  // convert to a mutable FST subclass (given by convert_type) in the case
-  // that the input FST is non-mutable.
-  static MutableFst<Arc> *Read(const string &filename, bool convert = false,
-                               const string &convert_type = "vector") {
-    if (convert == false) {
-      if (!filename.empty()) {
-        std::ifstream strm(filename,
-                                std::ios_base::in | std::ios_base::binary);
-        if (!strm) {
-          LOG(ERROR) << "MutableFst::Read: Can't open file: " << filename;
-          return nullptr;
-        }
-        return Read(strm, FstReadOptions(filename));
-      } else {
-        return Read(std::cin, FstReadOptions("standard input"));
-      }
-    } else {  // Converts to 'convert_type' if not mutable.
-      std::unique_ptr<Fst<Arc>> ifst(Fst<Arc>::Read(filename));
-      if (!ifst) return nullptr;
-      if (ifst->Properties(kMutable, false)) {
-        return static_cast<MutableFst<Arc> *>(ifst.release());
-      } else {
-        std::unique_ptr<Fst<Arc>> ofst(Convert(*ifst, convert_type));
-        ifst.reset();
-        if (!ofst) return nullptr;
-        if (!ofst->Properties(kMutable, false)) {
-          LOG(ERROR) << "MutableFst: Bad convert type: " << convert_type;
-        }
-        return static_cast<MutableFst<Arc> *>(ofst.release());
-      }
-    }
-  }
-
-  // For generic mutuble arc iterator construction; not normally called
-  // directly by users.
-  virtual void InitMutableArcIterator(StateId s,
-                                      MutableArcIteratorData<Arc> *data) = 0;
-};
-
-// Mutable arc iterator interface, templated on the Arc definition. This is
-// used by mutable arc iterator specializations that are returned by the
-// InitMutableArcIterator MutableFst method.
-template <class Arc>
-class MutableArcIteratorBase : public ArcIteratorBase<Arc> {
- public:
-  // Sets current arc.
-  virtual void SetValue(const Arc &) = 0;
-};
-
-template <class Arc>
-struct MutableArcIteratorData {
-  MutableArcIteratorBase<Arc> *base;  // Specific iterator.
-};
-
-// Generic mutable arc iterator, templated on the FST definition; a wrapper
-// around a pointer to a more specific one.
-//
-// Here is a typical use:
-//
-//   for (MutableArcIterator<StdFst> aiter(&fst, s);
-//        !aiter.Done();
-//         aiter.Next()) {
-//     StdArc arc = aiter.Value();
-//     arc.ilabel = 7;
-//     aiter.SetValue(arc);
-//     ...
-//   }
-//
-// This version requires function calls.
-template <class FST>
-class MutableArcIterator {
- public:
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-
-  MutableArcIterator(FST *fst, StateId s) {
-    fst->InitMutableArcIterator(s, &data_);
-  }
-
-  ~MutableArcIterator() { delete data_.base; }
-
-  bool Done() const { return data_.base->Done(); }
-
-  const Arc &Value() const { return data_.base->Value(); }
-
-  void Next() { data_.base->Next(); }
-
-  size_t Position() const { return data_.base->Position(); }
-
-  void Reset() { data_.base->Reset(); }
-
-  void Seek(size_t a) { data_.base->Seek(a); }
-
-  void SetValue(const Arc &arc) { data_.base->SetValue(arc); }
-
-  uint32 Flags() const { return data_.base->Flags(); }
-
-  void SetFlags(uint32 flags, uint32 mask) {
-    return data_.base->SetFlags(flags, mask);
-  }
-
- private:
-  MutableArcIteratorData<Arc> data_;
-
-  MutableArcIterator(const MutableArcIterator &) = delete;
-  MutableArcIterator &operator=(const MutableArcIterator &) = delete;
-};
-
-namespace internal {
-
-// MutableFst<A> case: abstract methods.
-template <class Arc>
-inline typename Arc::Weight Final(const MutableFst<Arc> &fst,
-                                  typename Arc::StateId s) {
-  return fst.Final(s);
-}
-
-template <class Arc>
-inline ssize_t NumArcs(const MutableFst<Arc> &fst, typename Arc::StateId s) {
-  return fst.NumArcs(s);
-}
-
-template <class Arc>
-inline ssize_t NumInputEpsilons(const MutableFst<Arc> &fst,
-                                typename Arc::StateId s) {
-  return fst.NumInputEpsilons(s);
-}
-
-template <class Arc>
-inline ssize_t NumOutputEpsilons(const MutableFst<Arc> &fst,
-                                 typename Arc::StateId s) {
-  return fst.NumOutputEpsilons(s);
-}
-
-}  // namespace internal
-
-// A useful alias when using StdArc.
-using StdMutableFst = MutableFst<StdArc>;
-
-// This is a helper class template useful for attaching a MutableFst interface
-// to its implementation, handling reference counting and COW semantics.
-template <class Impl, class FST = MutableFst<typename Impl::Arc>>
-class ImplToMutableFst : public ImplToExpandedFst<Impl, FST> {
- public:
-  using Arc = typename Impl::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using ImplToExpandedFst<Impl, FST>::operator=;
-
-  void SetStart(StateId s) override {
-    MutateCheck();
-    GetMutableImpl()->SetStart(s);
-  }
-
-  void SetFinal(StateId s, Weight weight) override {
-    MutateCheck();
-    GetMutableImpl()->SetFinal(s, std::move(weight));
-  }
-
-  void SetProperties(uint64 props, uint64 mask) override {
-    // Can skip mutate check if extrinsic properties don't change,
-    // since it is then safe to update all (shallow) copies
-    const auto exprops = kExtrinsicProperties & mask;
-    if (GetImpl()->Properties(exprops) != (props & exprops)) MutateCheck();
-    GetMutableImpl()->SetProperties(props, mask);
-  }
-
-  StateId AddState() override {
-    MutateCheck();
-    return GetMutableImpl()->AddState();
-  }
-
-  void AddArc(StateId s, const Arc &arc) override {
-    MutateCheck();
-    GetMutableImpl()->AddArc(s, arc);
-  }
-
-  void DeleteStates(const std::vector<StateId> &dstates) override {
-    MutateCheck();
-    GetMutableImpl()->DeleteStates(dstates);
-  }
-
-  void DeleteStates() override {
-    if (!Unique()) {
-      const auto *isymbols = GetImpl()->InputSymbols();
-      const auto *osymbols = GetImpl()->OutputSymbols();
-      SetImpl(std::make_shared<Impl>());
-      GetMutableImpl()->SetInputSymbols(isymbols);
-      GetMutableImpl()->SetOutputSymbols(osymbols);
-    } else {
-      GetMutableImpl()->DeleteStates();
-    }
-  }
-
-  void DeleteArcs(StateId s, size_t n) override {
-    MutateCheck();
-    GetMutableImpl()->DeleteArcs(s, n);
-  }
-
-  void DeleteArcs(StateId s) override {
-    MutateCheck();
-    GetMutableImpl()->DeleteArcs(s);
-  }
-
-  void ReserveStates(StateId s) override {
-    MutateCheck();
-    GetMutableImpl()->ReserveStates(s);
-  }
-
-  void ReserveArcs(StateId s, size_t n) override {
-    MutateCheck();
-    GetMutableImpl()->ReserveArcs(s, n);
-  }
-
-  const SymbolTable *InputSymbols() const override {
-    return GetImpl()->InputSymbols();
-  }
-
-  const SymbolTable *OutputSymbols() const override {
-    return GetImpl()->OutputSymbols();
-  }
-
-  SymbolTable *MutableInputSymbols() override {
-    MutateCheck();
-    return GetMutableImpl()->InputSymbols();
-  }
-
-  SymbolTable *MutableOutputSymbols() override {
-    MutateCheck();
-    return GetMutableImpl()->OutputSymbols();
-  }
-
-  void SetInputSymbols(const SymbolTable *isyms) override {
-    MutateCheck();
-    GetMutableImpl()->SetInputSymbols(isyms);
-  }
-
-  void SetOutputSymbols(const SymbolTable *osyms) override {
-    MutateCheck();
-    GetMutableImpl()->SetOutputSymbols(osyms);
-  }
-
- protected:
-  using ImplToExpandedFst<Impl, FST>::GetImpl;
-  using ImplToExpandedFst<Impl, FST>::GetMutableImpl;
-  using ImplToExpandedFst<Impl, FST>::Unique;
-  using ImplToExpandedFst<Impl, FST>::SetImpl;
-  using ImplToExpandedFst<Impl, FST>::InputSymbols;
-
-  explicit ImplToMutableFst(std::shared_ptr<Impl> impl)
-      : ImplToExpandedFst<Impl, FST>(impl) {}
-
-  ImplToMutableFst(const ImplToMutableFst<Impl, FST> &fst, bool safe)
-      : ImplToExpandedFst<Impl, FST>(fst, safe) {}
-
-  void MutateCheck() {
-    if (!Unique()) SetImpl(std::make_shared<Impl>(*this));
-  }
-};
-
-}  // namespace fst
-
-#endif  // FST_MUTABLE_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/pair-weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/pair-weight.h
deleted file mode 100644
index 05ad15a4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/pair-weight.h
+++ /dev/null
@@ -1,164 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Pair weight templated base class for weight classes that contain two weights
-// (e.g. Product, Lexicographic).
-
-#ifndef FST_PAIR_WEIGHT_H_
-#define FST_PAIR_WEIGHT_H_
-
-#include <climits>
-#include <stack>
-#include <string>
-#include <utility>
-
-#include <fst/flags.h>
-#include <fst/log.h>
-
-#include <fst/weight.h>
-
-
-namespace fst {
-
-template <class W1, class W2>
-class PairWeight {
- public:
-  using ReverseWeight =
-      PairWeight<typename W1::ReverseWeight, typename W2::ReverseWeight>;
-
-  PairWeight() {}
-
-  PairWeight(const PairWeight &weight)
-      : value1_(weight.value1_), value2_(weight.value2_) {}
-
-  PairWeight(W1 w1, W2 w2) : value1_(std::move(w1)), value2_(std::move(w2)) {}
-
-  static const PairWeight<W1, W2> &Zero() {
-    static const PairWeight zero(W1::Zero(), W2::Zero());
-    return zero;
-  }
-
-  static const PairWeight<W1, W2> &One() {
-    static const PairWeight one(W1::One(), W2::One());
-    return one;
-  }
-
-  static const PairWeight<W1, W2> &NoWeight() {
-    static const PairWeight no_weight(W1::NoWeight(), W2::NoWeight());
-    return no_weight;
-  }
-
-  std::istream &Read(std::istream &strm) {
-    value1_.Read(strm);
-    return value2_.Read(strm);
-  }
-
-  std::ostream &Write(std::ostream &strm) const {
-    value1_.Write(strm);
-    return value2_.Write(strm);
-  }
-
-  PairWeight<W1, W2> &operator=(const PairWeight<W1, W2> &weight) {
-    value1_ = weight.Value1();
-    value2_ = weight.Value2();
-    return *this;
-  }
-
-  bool Member() const { return value1_.Member() && value2_.Member(); }
-
-  size_t Hash() const {
-    const auto h1 = value1_.Hash();
-    const auto h2 = value2_.Hash();
-    static constexpr int lshift = 5;
-    static constexpr int rshift = CHAR_BIT * sizeof(size_t) - 5;
-    return h1 << lshift ^ h1 >> rshift ^ h2;
-  }
-
-  PairWeight<W1, W2> Quantize(float delta = kDelta) const {
-    return PairWeight<W1, W2>(value1_.Quantize(delta), value2_.Quantize(delta));
-  }
-
-  ReverseWeight Reverse() const {
-    return ReverseWeight(value1_.Reverse(), value2_.Reverse());
-  }
-
-  const W1 &Value1() const { return value1_; }
-
-  const W2 &Value2() const { return value2_; }
-
-  void SetValue1(const W1 &weight) { value1_ = weight; }
-
-  void SetValue2(const W2 &weight) { value2_ = weight; }
-
- private:
-  W1 value1_;
-  W2 value2_;
-};
-
-template <class W1, class W2>
-inline bool operator==(const PairWeight<W1, W2> &w1,
-                       const PairWeight<W1, W2> &w2) {
-  return w1.Value1() == w2.Value1() && w1.Value2() == w2.Value2();
-}
-
-template <class W1, class W2>
-inline bool operator!=(const PairWeight<W1, W2> &w1,
-                       const PairWeight<W1, W2> &w2) {
-  return w1.Value1() != w2.Value1() || w1.Value2() != w2.Value2();
-}
-
-template <class W1, class W2>
-inline bool ApproxEqual(const PairWeight<W1, W2> &w1,
-                        const PairWeight<W1, W2> &w2, float delta = kDelta) {
-  return ApproxEqual(w1.Value1(), w2.Value1(), delta) &&
-         ApproxEqual(w1.Value2(), w2.Value2(), delta);
-}
-
-template <class W1, class W2>
-inline std::ostream &operator<<(std::ostream &strm,
-                                const PairWeight<W1, W2> &weight) {
-  CompositeWeightWriter writer(strm);
-  writer.WriteBegin();
-  writer.WriteElement(weight.Value1());
-  writer.WriteElement(weight.Value2());
-  writer.WriteEnd();
-  return strm;
-}
-
-template <class W1, class W2>
-inline std::istream &operator>>(std::istream &strm,
-                                PairWeight<W1, W2> &weight) {
-  CompositeWeightReader reader(strm);
-  reader.ReadBegin();
-  W1 w1;
-  reader.ReadElement(&w1);
-  weight.SetValue1(w1);
-  W2 w2;
-  reader.ReadElement(&w2, true);
-  weight.SetValue2(w2);
-  reader.ReadEnd();
-  return strm;
-}
-
-// This function object returns weights by calling the underlying generators
-// and forming a pair. This is intended primarily for testing.
-template <class W1, class W2>
-class WeightGenerate<PairWeight<W1, W2>> {
- public:
-  using Weight = PairWeight<W1, W2>;
-  using Generate1 = WeightGenerate<W1>;
-  using Generate2 = WeightGenerate<W2>;
-
-  explicit WeightGenerate(bool allow_zero = true)
-      : generate1_(allow_zero), generate2_(allow_zero) {}
-
-  Weight operator()() const { return Weight(generate1_(), generate2_()); }
-
- private:
-  Generate1 generate1_;
-  Generate2 generate2_;
-};
-
-}  // namespace fst
-
-#endif  // FST_PAIR_WEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/partition.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/partition.h
deleted file mode 100644
index 5dbbe46a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/partition.h
+++ /dev/null
@@ -1,305 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to create a partition of states.
-
-#ifndef FST_PARTITION_H_
-#define FST_PARTITION_H_
-
-#include <algorithm>
-#include <vector>
-
-
-#include <fst/queue.h>
-
-
-namespace fst {
-namespace internal {
-
-template <typename T>
-class PartitionIterator;
-
-// Defines a partitioning of elements, used to represent equivalence classes
-// for FST operations like minimization. T must be a signed integer type.
-//
-// The elements are numbered from 0 to num_elements - 1.
-// Initialize(num_elements) sets up the class for a given number of elements.
-// We maintain a partition of these elements into classes. The classes are also
-// numbered from zero; you can add a class with AddClass(), or add them in bulk
-// with AllocateClasses(num_classes). Initially the elements are not assigned
-// to any class; you set up the initial mapping from elements to classes by
-// calling Add(element_id, class_id). You can also move an element to a
-// different class by calling Move(element_id, class_id).
-//
-// We also support a rather specialized interface that allows you to efficiently
-// split classes in the Hopcroft minimization algorithm. This maintains a
-// binary partition of each class.  Let's call these, rather arbitrarily, the
-// 'yes' subset and the 'no' subset of each class, and assume that by default,
-// each element of a class is in its 'no' subset. When one calls
-// SplitOn(element_id), element_id is moved to the 'yes' subset of its class.
-// (If it was already in the 'yes' set, it just stays there). The aim is to
-// enable (later) splitting the class in two in time no greater than the time
-// already spent calling SplitOn() for that class. We keep a list of the classes
-// which have nonempty 'yes' sets, as visited_classes_. When one calls
-// FinalizeSplit(Queue *l), for each class in visited_classes_ whose 'yes'
-// and 'no' sets are both nonempty, it will create a new class consisting of
-// the smaller of the two subsets (and this class will be added to the queue),
-// and the old class will now be the larger of the two subsets. This call also
-// resets all the yes/no partitions so that everything is in the 'no' subsets.
-//
-// One cannot use the Move() function if SplitOn() has been called without
-// a subsequent call to FinalizeSplit()
-template <typename T>
-class Partition {
- public:
-  Partition() {}
-
-  explicit Partition(T num_elements) { Initialize(num_elements); }
-
-  // Creates an empty partition for num_elements. This means that the elements
-  // are not assigned to a class (i.e class_index = -1); you should set up the
-  // number of classes using AllocateClasses() or AddClass(), and allocate each
-  // element to a class by calling Add(element, class_id).
-  void Initialize(size_t num_elements) {
-    elements_.resize(num_elements);
-    classes_.reserve(num_elements);
-    classes_.clear();
-    yes_counter_ = 1;
-  }
-
-  // Adds a class; returns new number of classes.
-  T AddClass() {
-    auto num_classes = classes_.size();
-    classes_.resize(num_classes + 1);
-    return num_classes;
-  }
-
-  // Adds 'num_classes' new (empty) classes.
-  void AllocateClasses(T num_classes) {
-    classes_.resize(classes_.size() + num_classes);
-  }
-
-  // Adds element_id to class_id. element_id should already have been allocated
-  // by calling Initialize(num_elements)---or the constructor taking
-  // num_elements---with num_elements > element_id. element_id must not
-  // currently be a member of any class; once elements have been added to a
-  // class, use the Move() method to move them from one class to another.
-  void Add(T element_id, T class_id) {
-    auto &this_element = elements_[element_id];
-    auto &this_class = classes_[class_id];
-    ++this_class.size;
-    // Adds the element to the 'no' subset of the class.
-    auto no_head = this_class.no_head;
-    if (no_head >= 0) elements_[no_head].prev_element = element_id;
-    this_class.no_head = element_id;
-    this_element.class_id = class_id;
-    // Adds to the 'no' subset of the class.
-    this_element.yes = 0;
-    this_element.next_element = no_head;
-    this_element.prev_element = -1;
-  }
-
-  // Moves element_id from 'no' subset of its current class to 'no' subset of
-  // class class_id. This may not work correctly if you have called SplitOn()
-  // [for any element] and haven't subsequently called FinalizeSplit().
-  void Move(T element_id, T class_id) {
-    auto elements = &(elements_[0]);
-    auto &element = elements[element_id];
-    auto &old_class = classes_[element.class_id];
-    --old_class.size;
-    // Excises the element from the 'no' list of its old class, where it is
-    // assumed to be.
-    if (element.prev_element >= 0) {
-      elements[element.prev_element].next_element = element.next_element;
-    } else {
-      old_class.no_head = element.next_element;
-    }
-    if (element.next_element >= 0) {
-      elements[element.next_element].prev_element = element.prev_element;
-    }
-    // Adds to new class.
-    Add(element_id, class_id);
-  }
-
-  // Moves element_id to the 'yes' subset of its class if it was in the 'no'
-  // subset, and marks the class as having been visited.
-  void SplitOn(T element_id) {
-    auto elements = &(elements_[0]);
-    auto &element = elements[element_id];
-    if (element.yes == yes_counter_) {
-      return;  // Already in the 'yes' set; nothing to do.
-    }
-    auto class_id = element.class_id;
-    auto &this_class = classes_[class_id];
-    // Excises the element from the 'no' list of its class.
-    if (element.prev_element >= 0) {
-      elements[element.prev_element].next_element = element.next_element;
-    } else {
-      this_class.no_head = element.next_element;
-    }
-    if (element.next_element >= 0) {
-      elements[element.next_element].prev_element = element.prev_element;
-    }
-    // Adds the element to the 'yes' list.
-    if (this_class.yes_head >= 0) {
-      elements[this_class.yes_head].prev_element = element_id;
-    } else {
-      visited_classes_.push_back(class_id);
-    }
-    element.yes = yes_counter_;
-    element.next_element = this_class.yes_head;
-    element.prev_element = -1;
-    this_class.yes_head = element_id;
-    this_class.yes_size++;
-  }
-
-  // This should be called after one has possibly called SplitOn for one or more
-  // elements, thus moving those elements to the 'yes' subset for their class.
-  // For each class that has a nontrivial split (i.e., it's not the case that
-  // all members are in the 'yes' or 'no' subset), this function creates a new
-  // class containing the smaller of the two subsets of elements, leaving the
-  // larger group of elements in the old class. The identifier of the new class
-  // will be added to the queue provided as the pointer L. This method then
-  // moves all elements to the 'no' subset of their class.
-  template <class Queue>
-  void FinalizeSplit(Queue *queue) {
-    for (const auto &visited_class : visited_classes_) {
-      const auto new_class = SplitRefine(visited_class);
-      if (new_class != -1 && queue) queue->Enqueue(new_class);
-    }
-    visited_classes_.clear();
-    // Incrementation sets all the 'yes' members of the elements to false.
-    ++yes_counter_;
-  }
-
-  const T ClassId(T element_id) const { return elements_[element_id].class_id; }
-
-  const size_t ClassSize(T class_id) const { return classes_[class_id].size; }
-
-  const T NumClasses() const { return classes_.size(); }
-
- private:
-  friend class PartitionIterator<T>;
-
-  // Information about a given element.
-  struct Element {
-    T class_id;      // Class ID of this element.
-    T yes;           // This is to be interpreted as a bool, true if it's in the
-                     // 'yes' set of this class. The interpretation as bool is
-                     // (yes == yes_counter_ ? true : false).
-    T next_element;  // Next element in the 'no' list or 'yes' list of this
-                     // class, whichever of the two we belong to (think of
-                     // this as the 'next' in a doubly-linked list, although
-                     // it is an index into the elements array). Negative
-                     // values corresponds to null.
-    T prev_element;  // Previous element in the 'no' or 'yes' doubly linked
-                     // list. Negative values corresponds to null.
-  };
-
-  // Information about a given class.
-  struct Class {
-    Class() : size(0), yes_size(0), no_head(-1), yes_head(-1) {}
-    T size;      // Total number of elements in this class ('no' plus 'yes'
-                 // subsets).
-    T yes_size;  // Total number of elements of 'yes' subset of this class.
-    T no_head;   // Index of head element of doubly-linked list in 'no' subset.
-                 // Everything is in the 'no' subset until you call SplitOn().
-                 // -1 means no element.
-    T yes_head;  // Index of head element of doubly-linked list in 'yes' subset.
-                 // -1 means no element.
-  };
-
-  // This method, called from FinalizeSplit(), checks whether a class has to
-  // be split (a class will be split only if its 'yes' and 'no' subsets are
-  // both nonempty, but one can assume that since this function was called, the
-  // 'yes' subset is nonempty). It splits by taking the smaller subset and
-  // making it a new class, and leaving the larger subset of elements in the
-  // 'no' subset of the old class. It returns the new class if created, or -1
-  // if none was created.
-  T SplitRefine(T class_id) {
-    auto yes_size = classes_[class_id].yes_size;
-    auto size = classes_[class_id].size;
-    auto no_size = size - yes_size;
-    if (no_size == 0) {
-      // All members are in the 'yes' subset, so we don't have to create a new
-      // class, just move them all to the 'no' subset.
-      classes_[class_id].no_head = classes_[class_id].yes_head;
-      classes_[class_id].yes_head = -1;
-      classes_[class_id].yes_size = 0;
-      return -1;
-    } else {
-      auto new_class_id = classes_.size();
-      classes_.resize(classes_.size() + 1);
-      auto &old_class = classes_[class_id];
-      auto &new_class = classes_[new_class_id];
-      // The new_class will have the values from the constructor.
-      if (no_size < yes_size) {
-        // Moves the 'no' subset to new class ('no' subset).
-        new_class.no_head = old_class.no_head;
-        new_class.size = no_size;
-        // And makes the 'yes' subset of the old class ('no' subset).
-        old_class.no_head = old_class.yes_head;
-        old_class.yes_head = -1;
-        old_class.size = yes_size;
-        old_class.yes_size = 0;
-      } else {
-        // Moves the 'yes' subset to the new class (to the 'no' subset)
-        new_class.size = yes_size;
-        new_class.no_head = old_class.yes_head;
-        // Retains only the 'no' subset in the old class.
-        old_class.size = no_size;
-        old_class.yes_size = 0;
-        old_class.yes_head = -1;
-      }
-      auto elements = &(elements_[0]);
-      // Updates the 'class_id' of all the elements we moved.
-      for (auto e = new_class.no_head; e >= 0; e = elements[e].next_element) {
-        elements[e].class_id = new_class_id;
-      }
-      return new_class_id;
-    }
-  }
-
-  // elements_[i] contains all info about the i'th element.
-  std::vector<Element> elements_;
-  // classes_[i] contains all info about the i'th class.
-  std::vector<Class> classes_;
-  // Set of visited classes to be used in split refine.
-  std::vector<T> visited_classes_;
-  // yes_counter_ is used in interpreting the 'yes' members of class Element.
-  // If element.yes == yes_counter_, we interpret that element as being in the
-  // 'yes' subset of its class. This allows us to, in effect, set all those
-  // bools to false at a stroke by incrementing yes_counter_.
-  T yes_counter_;
-};
-
-// Iterates over members of the 'no' subset of a class in a partition. (When
-// this is used, everything is in the 'no' subset).
-template <typename T>
-class PartitionIterator {
- public:
-  using Element = typename Partition<T>::Element;
-
-  PartitionIterator(const Partition<T> &partition, T class_id)
-      : partition_(partition),
-        element_id_(partition_.classes_[class_id].no_head),
-        class_id_(class_id) {}
-
-  bool Done() { return element_id_ < 0; }
-
-  const T Value() { return element_id_; }
-
-  void Next() { element_id_ = partition_.elements_[element_id_].next_element; }
-
-  void Reset() { element_id_ = partition_.classes_[class_id_].no_head; }
-
- private:
-  const Partition<T> &partition_;
-  T element_id_;
-  T class_id_;
-};
-
-}  // namespace internal
-}  // namespace fst
-
-#endif  // FST_PARTITION_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/power-weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/power-weight.h
deleted file mode 100644
index f2f3cbdb..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/power-weight.h
+++ /dev/null
@@ -1,168 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Cartesian power weight semiring operation definitions.
-
-#ifndef FST_POWER_WEIGHT_H_
-#define FST_POWER_WEIGHT_H_
-
-#include <string>
-
-#include <fst/tuple-weight.h>
-#include <fst/weight.h>
-
-
-namespace fst {
-
-// Cartesian power semiring: W ^ n
-//
-// Forms:
-//  - a left semimodule when W is a left semiring,
-//  - a right semimodule when W is a right semiring,
-//  - a bisemimodule when W is a semiring,
-//    the free semimodule of rank n over W
-// The Times operation is overloaded to provide the left and right scalar
-// products.
-template <class W, size_t n>
-class PowerWeight : public TupleWeight<W, n> {
- public:
-  using ReverseWeight = PowerWeight<typename W::ReverseWeight, n>;
-
-  PowerWeight() {}
-
-  explicit PowerWeight(const TupleWeight<W, n> &weight)
-      : TupleWeight<W, n>(weight) {}
-
-  template <class Iterator>
-  PowerWeight(Iterator begin, Iterator end) : TupleWeight<W, n>(begin, end) {}
-
-  // Initialize component `index` to `weight`; initialize all other components
-  // to `default_weight`
-  PowerWeight(size_t index, const W &weight,
-              const W &default_weight = W::Zero())
-      : TupleWeight<W, n>(index, weight, default_weight) {}
-
-  static const PowerWeight &Zero() {
-    static const PowerWeight zero(TupleWeight<W, n>::Zero());
-    return zero;
-  }
-
-  static const PowerWeight &One() {
-    static const PowerWeight one(TupleWeight<W, n>::One());
-    return one;
-  }
-
-  static const PowerWeight &NoWeight() {
-    static const PowerWeight no_weight(TupleWeight<W, n>::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type =
-        new string(W::Type() + "_^" + std::to_string(n));
-    return *type;
-  }
-
-  static constexpr uint64 Properties() {
-    return W::Properties() &
-           (kLeftSemiring | kRightSemiring | kCommutative | kIdempotent);
-  }
-
-  PowerWeight Quantize(float delta = kDelta) const {
-    return PowerWeight(TupleWeight<W, n>::Quantize(delta));
-  }
-
-  ReverseWeight Reverse() const {
-    return ReverseWeight(TupleWeight<W, n>::Reverse());
-  }
-};
-
-// Semiring plus operation.
-template <class W, size_t n>
-inline PowerWeight<W, n> Plus(const PowerWeight<W, n> &w1,
-                              const PowerWeight<W, n> &w2) {
-  PowerWeight<W, n> result;
-  for (size_t i = 0; i < n; ++i) {
-    result.SetValue(i, Plus(w1.Value(i), w2.Value(i)));
-  }
-  return result;
-}
-
-// Semiring times operation.
-template <class W, size_t n>
-inline PowerWeight<W, n> Times(const PowerWeight<W, n> &w1,
-                               const PowerWeight<W, n> &w2) {
-  PowerWeight<W, n> result;
-  for (size_t i = 0; i < n; ++i) {
-    result.SetValue(i, Times(w1.Value(i), w2.Value(i)));
-  }
-  return result;
-}
-
-// Semiring divide operation.
-template <class W, size_t n>
-inline PowerWeight<W, n> Divide(const PowerWeight<W, n> &w1,
-                                const PowerWeight<W, n> &w2,
-                                DivideType type = DIVIDE_ANY) {
-  PowerWeight<W, n> result;
-  for (size_t i = 0; i < n; ++i) {
-    result.SetValue(i, Divide(w1.Value(i), w2.Value(i), type));
-  }
-  return result;
-}
-
-// Semimodule left scalar product.
-template <class W, size_t n>
-inline PowerWeight<W, n> Times(const W &scalar,
-                               const PowerWeight<W, n> &weight) {
-  PowerWeight<W, n> result;
-  for (size_t i = 0; i < n; ++i) {
-    result.SetValue(i, Times(scalar, weight.Value(i)));
-  }
-  return result;
-}
-
-// Semimodule right scalar product.
-template <class W, size_t n>
-inline PowerWeight<W, n> Times(const PowerWeight<W, n> &weight,
-                               const W &scalar) {
-  PowerWeight<W, n> result;
-  for (size_t i = 0; i < n; ++i) {
-    result.SetValue(i, Times(weight.Value(i), scalar));
-  }
-  return result;
-}
-
-// Semimodule dot product.
-template <class W, size_t n>
-inline W DotProduct(const PowerWeight<W, n> &w1, const PowerWeight<W, n> &w2) {
-  W result(W::Zero());
-  for (size_t i = 0; i < n; ++i) {
-    result = Plus(result, Times(w1.Value(i), w2.Value(i)));
-  }
-  return result;
-}
-
-// This function object generates weights over the Cartesian power of rank
-// n over the underlying weight. This is intended primarily for testing.
-template <class W, size_t n>
-class WeightGenerate<PowerWeight<W, n>> {
- public:
-  using Weight = PowerWeight<W, n>;
-  using Generate = WeightGenerate<W>;
-
-  explicit WeightGenerate(bool allow_zero = true) : generate_(allow_zero) {}
-
-  Weight operator()() const {
-    Weight result;
-    for (size_t i = 0; i < n; ++i) result.SetValue(i, generate_());
-    return result;
-  }
-
- private:
-  Generate generate_;
-};
-
-}  // namespace fst
-
-#endif  // FST_POWER_WEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/product-weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/product-weight.h
deleted file mode 100644
index 56a18be1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/product-weight.h
+++ /dev/null
@@ -1,107 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Product weight set and associated semiring operation definitions.
-
-#ifndef FST_PRODUCT_WEIGHT_H_
-#define FST_PRODUCT_WEIGHT_H_
-
-#include <string>
-#include <utility>
-
-#include <fst/pair-weight.h>
-#include <fst/weight.h>
-
-
-namespace fst {
-
-// Product semiring: W1 * W2.
-template <class W1, class W2>
-class ProductWeight : public PairWeight<W1, W2> {
- public:
-  using ReverseWeight =
-      ProductWeight<typename W1::ReverseWeight, typename W2::ReverseWeight>;
-
-  ProductWeight() {}
-
-  explicit ProductWeight(const PairWeight<W1, W2> &weight)
-      : PairWeight<W1, W2>(weight) {}
-
-  ProductWeight(W1 w1, W2 w2)
-      : PairWeight<W1, W2>(std::move(w1), std::move(w2)) {}
-
-  static const ProductWeight &Zero() {
-    static const ProductWeight zero(PairWeight<W1, W2>::Zero());
-    return zero;
-  }
-
-  static const ProductWeight &One() {
-    static const ProductWeight one(PairWeight<W1, W2>::One());
-    return one;
-  }
-
-  static const ProductWeight &NoWeight() {
-    static const ProductWeight no_weight(PairWeight<W1, W2>::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type =
-        new string(W1::Type() + "_X_" + W2::Type());
-    return *type;
-  }
-
-  static constexpr uint64 Properties() {
-    return W1::Properties() & W2::Properties() &
-           (kLeftSemiring | kRightSemiring | kCommutative | kIdempotent);
-  }
-
-  ProductWeight Quantize(float delta = kDelta) const {
-    return ProductWeight(PairWeight<W1, W2>::Quantize(delta));
-  }
-
-  ReverseWeight Reverse() const {
-    return ReverseWeight(PairWeight<W1, W2>::Reverse());
-  }
-};
-
-template <class W1, class W2>
-inline ProductWeight<W1, W2> Plus(const ProductWeight<W1, W2> &w1,
-                                  const ProductWeight<W1, W2> &w2) {
-  return ProductWeight<W1, W2>(Plus(w1.Value1(), w2.Value1()),
-                               Plus(w1.Value2(), w2.Value2()));
-}
-
-template <class W1, class W2>
-inline ProductWeight<W1, W2> Times(const ProductWeight<W1, W2> &w1,
-                                   const ProductWeight<W1, W2> &w2) {
-  return ProductWeight<W1, W2>(Times(w1.Value1(), w2.Value1()),
-                               Times(w1.Value2(), w2.Value2()));
-}
-
-template <class W1, class W2>
-inline ProductWeight<W1, W2> Divide(const ProductWeight<W1, W2> &w1,
-                                    const ProductWeight<W1, W2> &w2,
-                                    DivideType typ = DIVIDE_ANY) {
-  return ProductWeight<W1, W2>(Divide(w1.Value1(), w2.Value1(), typ),
-                               Divide(w1.Value2(), w2.Value2(), typ));
-}
-
-// This function object generates weights by calling the underlying generators
-// for the template weight types, like all other pair weight types. This is
-// intended primarily for testing.
-template <class W1, class W2>
-class WeightGenerate<ProductWeight<W1, W2>> :
-    public WeightGenerate<PairWeight<W1, W2>> {
- public:
-  using Weight = ProductWeight<W1, W2>;
-  using Generate = WeightGenerate<PairWeight<W1, W2>>;
-
-  explicit WeightGenerate(bool allow_zero = true) : Generate(allow_zero) {}
-
-  Weight operator()() const { return Weight(Generate::operator()()); }
-};
-
-}  // namespace fst
-
-#endif  // FST_PRODUCT_WEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/project.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/project.h
deleted file mode 100644
index 68c73869..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/project.h
+++ /dev/null
@@ -1,138 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to project an FST on to its domain or range.
-
-#ifndef FST_PROJECT_H_
-#define FST_PROJECT_H_
-
-#include <fst/arc-map.h>
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-// This specifies whether to project on input or output.
-enum ProjectType { PROJECT_INPUT = 1, PROJECT_OUTPUT = 2 };
-
-// Mapper to implement projection per arc.
-template <class A>
-class ProjectMapper {
- public:
-  using FromArc = A;
-  using ToArc = A;
-
-  explicit ProjectMapper(ProjectType project_type)
-      : project_type_(project_type) {}
-
-  ToArc operator()(const FromArc &arc) {
-    const auto label = project_type_ == PROJECT_INPUT ? arc.ilabel : arc.olabel;
-    return ToArc(label, label, arc.weight, arc.nextstate);
-  }
-
-  MapFinalAction FinalAction() const { return MAP_NO_SUPERFINAL; }
-
-  MapSymbolsAction InputSymbolsAction() const {
-    return project_type_ == PROJECT_INPUT ? MAP_COPY_SYMBOLS
-                                          : MAP_CLEAR_SYMBOLS;
-  }
-
-  MapSymbolsAction OutputSymbolsAction() const {
-    return project_type_ == PROJECT_OUTPUT ? MAP_COPY_SYMBOLS
-                                           : MAP_CLEAR_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) {
-    return ProjectProperties(props, project_type_ == PROJECT_INPUT);
-  }
-
- private:
-  ProjectType project_type_;
-};
-
-// Projects an FST onto its domain or range by either copying each arcs' input
-// label to the output label or vice versa. This version modifies its input.
-//
-// Complexity:
-//
-//   Time: O(V + E)
-//   Space: O(1)
-//
-// where V is the number of states and E is the number of arcs.
-template <class Arc>
-inline void Project(MutableFst<Arc> *fst, ProjectType project_type) {
-  ArcMap(fst, ProjectMapper<Arc>(project_type));
-  if (project_type == PROJECT_INPUT) fst->SetOutputSymbols(fst->InputSymbols());
-  if (project_type == PROJECT_OUTPUT) {
-    fst->SetInputSymbols(fst->OutputSymbols());
-  }
-}
-
-// Projects an FST onto its domain or range by either copying each arc's input
-// label to the output label or vice versa. This version is a delayed FST.
-//
-// Complexity:
-//
-//   Time: O(v + e)
-//   Space: O(1)
-//
-// where v is the number of states visited and e is the number of arcs visited.
-// Constant time and to visit an input state or arc is assumed and exclusive of
-// caching.
-template <class A>
-class ProjectFst : public ArcMapFst<A, A, ProjectMapper<A>> {
- public:
-  using FromArc = A;
-  using ToArc = A;
-
-  using Impl = internal::ArcMapFstImpl<A, A, ProjectMapper<A>>;
-
-  ProjectFst(const Fst<A> &fst, ProjectType project_type)
-      : ArcMapFst<A, A, ProjectMapper<A>>(fst, ProjectMapper<A>(project_type)) {
-    if (project_type == PROJECT_INPUT) {
-      GetMutableImpl()->SetOutputSymbols(fst.InputSymbols());
-    }
-    if (project_type == PROJECT_OUTPUT) {
-      GetMutableImpl()->SetInputSymbols(fst.OutputSymbols());
-    }
-  }
-
-  // See Fst<>::Copy() for doc.
-  ProjectFst(const ProjectFst<A> &fst, bool safe = false)
-      : ArcMapFst<A, A, ProjectMapper<A>>(fst, safe) {}
-
-  // Gets a copy of this ProjectFst. See Fst<>::Copy() for further doc.
-  ProjectFst<A> *Copy(bool safe = false) const override {
-    return new ProjectFst(*this, safe);
-  }
-
- private:
-  using ImplToFst<Impl>::GetMutableImpl;
-};
-
-// Specialization for ProjectFst.
-template <class A>
-class StateIterator<ProjectFst<A>>
-    : public StateIterator<ArcMapFst<A, A, ProjectMapper<A>>> {
- public:
-  explicit StateIterator(const ProjectFst<A> &fst)
-      : StateIterator<ArcMapFst<A, A, ProjectMapper<A>>>(fst) {}
-};
-
-// Specialization for ProjectFst.
-template <class A>
-class ArcIterator<ProjectFst<A>>
-    : public ArcIterator<ArcMapFst<A, A, ProjectMapper<A>>> {
- public:
-  using StateId = typename A::StateId;
-
-  ArcIterator(const ProjectFst<A> &fst, StateId s)
-      : ArcIterator<ArcMapFst<A, A, ProjectMapper<A>>>(fst, s) {}
-};
-
-// Useful alias when using StdArc.
-using StdProjectFst = ProjectFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_PROJECT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/properties.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/properties.h
deleted file mode 100644
index 1f39fedc..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/properties.h
+++ /dev/null
@@ -1,467 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// FST property bits.
-
-#ifndef FST_PROPERTIES_H_
-#define FST_PROPERTIES_H_
-
-#include <sys/types.h>
-#include <vector>
-
-#include <fst/compat.h>
-
-namespace fst {
-
-// The property bits here assert facts about an FST. If individual bits are
-// added, then the composite properties below, the property functions and
-// property names in properties.cc, and TestProperties() in test-properties.h
-// should be updated.
-
-// BINARY PROPERTIES
-//
-// For each property below, there is a single bit. If it is set, the property is
-// true. If it is not set, the property is false.
-
-// The Fst is an ExpandedFst.
-constexpr uint64 kExpanded = 0x0000000000000001ULL;
-
-// The Fst is a MutableFst.
-constexpr uint64 kMutable = 0x0000000000000002ULL;
-
-// An error was detected while constructing/using the FST.
-constexpr uint64 kError = 0x0000000000000004ULL;
-
-// TRINARY PROPERTIES
-//
-// For each of these properties below there is a pair of property bits, one
-// positive and one negative. If the positive bit is set, the property is true.
-// If the negative bit is set, the property is false. If neither is set, the
-// property has unknown value. Both should never be simultaneously set. The
-// individual positive and negative bit pairs should be adjacent with the
-// positive bit at an odd and lower position.
-
-// ilabel == olabel for each arc.
-constexpr uint64 kAcceptor = 0x0000000000010000ULL;
-// ilabel != olabel for some arc.
-constexpr uint64 kNotAcceptor = 0x0000000000020000ULL;
-
-// ilabels unique leaving each state.
-constexpr uint64 kIDeterministic = 0x0000000000040000ULL;
-// ilabels not unique leaving some state.
-constexpr uint64 kNonIDeterministic = 0x0000000000080000ULL;
-
-// olabels unique leaving each state.
-constexpr uint64 kODeterministic = 0x0000000000100000ULL;
-// olabels not unique leaving some state.
-constexpr uint64 kNonODeterministic = 0x0000000000200000ULL;
-
-// FST has input/output epsilons.
-constexpr uint64 kEpsilons = 0x0000000000400000ULL;
-// FST has no input/output epsilons.
-constexpr uint64 kNoEpsilons = 0x0000000000800000ULL;
-
-// FST has input epsilons.
-constexpr uint64 kIEpsilons = 0x0000000001000000ULL;
-// FST has no input epsilons.
-constexpr uint64 kNoIEpsilons = 0x0000000002000000ULL;
-
-// FST has output epsilons.
-constexpr uint64 kOEpsilons = 0x0000000004000000ULL;
-// FST has no output epsilons.
-constexpr uint64 kNoOEpsilons = 0x0000000008000000ULL;
-
-// ilabels sorted wrt < for each state.
-constexpr uint64 kILabelSorted = 0x0000000010000000ULL;
-// ilabels not sorted wrt < for some state.
-constexpr uint64 kNotILabelSorted = 0x0000000020000000ULL;
-
-// olabels sorted wrt < for each state.
-constexpr uint64 kOLabelSorted = 0x0000000040000000ULL;
-// olabels not sorted wrt < for some state.
-constexpr uint64 kNotOLabelSorted = 0x0000000080000000ULL;
-
-// Non-trivial arc or final weights.
-constexpr uint64 kWeighted = 0x0000000100000000ULL;
-// Only trivial arc and final weights.
-constexpr uint64 kUnweighted = 0x0000000200000000ULL;
-
-// FST has cycles.
-constexpr uint64 kCyclic = 0x0000000400000000ULL;
-// FST has no cycles.
-constexpr uint64 kAcyclic = 0x0000000800000000ULL;
-
-// FST has cycles containing the initial state.
-constexpr uint64 kInitialCyclic = 0x0000001000000000ULL;
-// FST has no cycles containing the initial state.
-constexpr uint64 kInitialAcyclic = 0x0000002000000000ULL;
-
-// FST is topologically sorted.
-constexpr uint64 kTopSorted = 0x0000004000000000ULL;
-// FST is not topologically sorted.
-constexpr uint64 kNotTopSorted = 0x0000008000000000ULL;
-
-// All states reachable from the initial state.
-constexpr uint64 kAccessible = 0x0000010000000000ULL;
-// Not all states reachable from the initial state.
-constexpr uint64 kNotAccessible = 0x0000020000000000ULL;
-
-// All states can reach a final state.
-constexpr uint64 kCoAccessible = 0x0000040000000000ULL;
-// Not all states can reach a final state.
-constexpr uint64 kNotCoAccessible = 0x0000080000000000ULL;
-
-// If NumStates() > 0, then state 0 is initial, state NumStates() - 1 is final,
-// there is a transition from each non-final state i to state i + 1, and there
-// are no other transitions.
-constexpr uint64 kString = 0x0000100000000000ULL;
-
-// Not a string FST.
-constexpr uint64 kNotString = 0x0000200000000000ULL;
-
-// FST has least one weighted cycle.
-constexpr uint64 kWeightedCycles = 0x0000400000000000ULL;
-
-// Only unweighted cycles.
-constexpr uint64 kUnweightedCycles = 0x0000800000000000ULL;
-
-// COMPOSITE PROPERTIES
-
-// Properties of an empty machine.
-constexpr uint64 kNullProperties =
-    kAcceptor | kIDeterministic | kODeterministic | kNoEpsilons | kNoIEpsilons |
-    kNoOEpsilons | kILabelSorted | kOLabelSorted | kUnweighted | kAcyclic |
-    kInitialAcyclic | kTopSorted | kAccessible | kCoAccessible | kString |
-    kUnweightedCycles;
-
-// Properties that are preserved when an FST is copied.
-constexpr uint64 kCopyProperties =
-    kError | kAcceptor | kNotAcceptor | kIDeterministic | kNonIDeterministic |
-    kODeterministic | kNonODeterministic | kEpsilons | kNoEpsilons |
-    kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons | kILabelSorted |
-    kNotILabelSorted | kOLabelSorted | kNotOLabelSorted | kWeighted |
-    kUnweighted | kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic |
-    kTopSorted | kNotTopSorted | kAccessible | kNotAccessible | kCoAccessible |
-    kNotCoAccessible | kString | kNotString | kWeightedCycles |
-    kUnweightedCycles;
-
-// Properties that are intrinsic to the FST.
-constexpr uint64 kIntrinsicProperties =
-    kExpanded | kMutable | kAcceptor | kNotAcceptor | kIDeterministic |
-    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
-    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
-    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
-    kWeighted | kUnweighted | kCyclic | kAcyclic | kInitialCyclic |
-    kInitialAcyclic | kTopSorted | kNotTopSorted | kAccessible |
-    kNotAccessible | kCoAccessible | kNotCoAccessible | kString | kNotString |
-    kWeightedCycles | kUnweightedCycles;
-
-// Properties that are (potentially) extrinsic to the FST.
-constexpr uint64 kExtrinsicProperties = kError;
-
-// Properties that are preserved when an FST start state is set.
-constexpr uint64 kSetStartProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
-    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
-    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
-    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
-    kWeighted | kUnweighted | kCyclic | kAcyclic | kTopSorted | kNotTopSorted |
-    kCoAccessible | kNotCoAccessible | kWeightedCycles | kUnweightedCycles;
-
-// Properties that are preserved when an FST final weight is set.
-constexpr uint64 kSetFinalProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
-    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
-    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
-    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
-    kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic | kTopSorted |
-    kNotTopSorted | kAccessible | kNotAccessible | kWeightedCycles |
-    kUnweightedCycles;
-
-// Properties that are preserved when an FST state is added.
-constexpr uint64 kAddStateProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
-    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
-    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
-    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
-    kWeighted | kUnweighted | kCyclic | kAcyclic | kInitialCyclic |
-    kInitialAcyclic | kTopSorted | kNotTopSorted | kNotAccessible |
-    kNotCoAccessible | kNotString | kWeightedCycles | kUnweightedCycles;
-
-// Properties that are preserved when an FST arc is added.
-constexpr uint64 kAddArcProperties =
-    kExpanded | kMutable | kError | kNotAcceptor | kNonIDeterministic |
-    kNonODeterministic | kEpsilons | kIEpsilons | kOEpsilons |
-    kNotILabelSorted | kNotOLabelSorted | kWeighted | kCyclic | kInitialCyclic |
-    kNotTopSorted | kAccessible | kCoAccessible | kWeightedCycles;
-
-// Properties that are preserved when an FST arc is set.
-constexpr uint64 kSetArcProperties = kExpanded | kMutable | kError;
-
-// Properties that are preserved when FST states are deleted.
-constexpr uint64 kDeleteStatesProperties =
-    kExpanded | kMutable | kError | kAcceptor | kIDeterministic |
-    kODeterministic | kNoEpsilons | kNoIEpsilons | kNoOEpsilons |
-    kILabelSorted | kOLabelSorted | kUnweighted | kAcyclic | kInitialAcyclic |
-    kTopSorted | kUnweightedCycles;
-
-// Properties that are preserved when FST arcs are deleted.
-constexpr uint64 kDeleteArcsProperties =
-    kExpanded | kMutable | kError | kAcceptor | kIDeterministic |
-    kODeterministic | kNoEpsilons | kNoIEpsilons | kNoOEpsilons |
-    kILabelSorted | kOLabelSorted | kUnweighted | kAcyclic | kInitialAcyclic |
-    kTopSorted | kNotAccessible | kNotCoAccessible | kUnweightedCycles;
-
-// Properties that are preserved when an FST's states are reordered.
-constexpr uint64 kStateSortProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
-    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
-    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
-    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
-    kWeighted | kUnweighted | kCyclic | kAcyclic | kInitialCyclic |
-    kInitialAcyclic | kAccessible | kNotAccessible | kCoAccessible |
-    kNotCoAccessible | kWeightedCycles | kUnweightedCycles;
-
-// Properties that are preserved when an FST's arcs are reordered.
-constexpr uint64 kArcSortProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
-    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
-    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
-    kWeighted | kUnweighted | kCyclic | kAcyclic | kInitialCyclic |
-    kInitialAcyclic | kTopSorted | kNotTopSorted | kAccessible |
-    kNotAccessible | kCoAccessible | kNotCoAccessible | kString | kNotString |
-    kWeightedCycles | kUnweightedCycles;
-
-// Properties that are preserved when an FST's input labels are changed.
-constexpr uint64 kILabelInvariantProperties =
-    kExpanded | kMutable | kError | kODeterministic | kNonODeterministic |
-    kOEpsilons | kNoOEpsilons | kOLabelSorted | kNotOLabelSorted | kWeighted |
-    kUnweighted | kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic |
-    kTopSorted | kNotTopSorted | kAccessible | kNotAccessible | kCoAccessible |
-    kNotCoAccessible | kString | kNotString | kWeightedCycles |
-    kUnweightedCycles;
-
-// Properties that are preserved when an FST's output labels are changed.
-constexpr uint64 kOLabelInvariantProperties =
-    kExpanded | kMutable | kError | kIDeterministic | kNonIDeterministic |
-    kIEpsilons | kNoIEpsilons | kILabelSorted | kNotILabelSorted | kWeighted |
-    kUnweighted | kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic |
-    kTopSorted | kNotTopSorted | kAccessible | kNotAccessible | kCoAccessible |
-    kNotCoAccessible | kString | kNotString | kWeightedCycles |
-    kUnweightedCycles;
-
-// Properties that are preserved when an FST's weights are changed. This
-// assumes that the set of states that are non-final is not changed.
-constexpr uint64 kWeightInvariantProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
-    kNonIDeterministic | kODeterministic | kNonODeterministic | kEpsilons |
-    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons | kNoOEpsilons |
-    kILabelSorted | kNotILabelSorted | kOLabelSorted | kNotOLabelSorted |
-    kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic | kTopSorted |
-    kNotTopSorted | kAccessible | kNotAccessible | kCoAccessible |
-    kNotCoAccessible | kString | kNotString;
-
-// Properties that are preserved when a superfinal state is added and an FST's
-// final weights are directed to it via new transitions.
-constexpr uint64 kAddSuperFinalProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor |
-    kNonIDeterministic | kNonODeterministic | kEpsilons | kIEpsilons |
-    kOEpsilons | kNotILabelSorted | kNotOLabelSorted | kWeighted | kUnweighted |
-    kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic | kNotTopSorted |
-    kNotAccessible | kCoAccessible | kNotCoAccessible | kNotString |
-    kWeightedCycles | kUnweightedCycles;
-
-// Properties that are preserved when a superfinal state is removed and the
-// epsilon transitions directed to it are made final weights.
-constexpr uint64 kRmSuperFinalProperties =
-    kExpanded | kMutable | kError | kAcceptor | kNotAcceptor | kIDeterministic |
-    kODeterministic | kNoEpsilons | kNoIEpsilons | kNoOEpsilons |
-    kILabelSorted | kOLabelSorted | kWeighted | kUnweighted | kCyclic |
-    kAcyclic | kInitialCyclic | kInitialAcyclic | kTopSorted | kAccessible |
-    kCoAccessible | kNotCoAccessible | kString | kWeightedCycles |
-    kUnweightedCycles;
-
-// All binary properties.
-constexpr uint64 kBinaryProperties = 0x0000000000000007ULL;
-
-// All trinary properties.
-constexpr uint64 kTrinaryProperties = 0x0000ffffffff0000ULL;
-
-// COMPUTED PROPERTIES
-
-// 1st bit of trinary properties.
-constexpr uint64 kPosTrinaryProperties = kTrinaryProperties &
-    0x5555555555555555ULL;
-
-// 2nd bit of trinary properties.
-constexpr uint64 kNegTrinaryProperties = kTrinaryProperties &
-    0xaaaaaaaaaaaaaaaaULL;
-
-// All properties.
-constexpr uint64 kFstProperties = kBinaryProperties | kTrinaryProperties;
-
-// PROPERTY FUNCTIONS and STRING NAMES (defined in properties.cc).
-
-// Below are functions for getting property bit vectors when executing
-// mutation operations.
-inline uint64 SetStartProperties(uint64 inprops);
-
-template <typename Weight>
-uint64 SetFinalProperties(uint64 inprops, Weight old_weight, Weight new_weight);
-
-inline uint64 AddStateProperties(uint64 inprops);
-
-template <typename A>
-uint64 AddArcProperties(uint64 inprops, typename A::StateId s, const A &arc,
-                        const A *prev_arc);
-
-inline uint64 DeleteStatesProperties(uint64 inprops);
-
-inline uint64 DeleteAllStatesProperties(uint64 inprops, uint64 staticProps);
-
-inline uint64 DeleteArcsProperties(uint64 inprops);
-
-uint64 ClosureProperties(uint64 inprops, bool star, bool delayed = false);
-
-uint64 ComplementProperties(uint64 inprops);
-
-uint64 ComposeProperties(uint64 inprops1, uint64 inprops2);
-
-uint64 ConcatProperties(uint64 inprops1, uint64 inprops2, bool delayed = false);
-
-uint64 DeterminizeProperties(uint64 inprops, bool has_subsequential_label,
-                             bool distinct_psubsequential_labels);
-
-uint64 FactorWeightProperties(uint64 inprops);
-
-uint64 InvertProperties(uint64 inprops);
-
-uint64 ProjectProperties(uint64 inprops, bool project_input);
-
-uint64 RandGenProperties(uint64 inprops, bool weighted);
-
-uint64 RelabelProperties(uint64 inprops);
-
-uint64 ReplaceProperties(const std::vector<uint64> &inprops, ssize_t root,
-                         bool epsilon_on_call, bool epsilon_on_return,
-                         bool out_epsilon_on_call, bool out_epsilon_on_return,
-                         bool replace_transducer, bool no_empty_fst,
-                         bool all_ilabel_sorted, bool all_olabel_sorted,
-                         bool all_negative_or_dense);
-
-uint64 ReverseProperties(uint64 inprops, bool has_superinitial);
-
-uint64 ReweightProperties(uint64 inprops);
-
-uint64 RmEpsilonProperties(uint64 inprops, bool delayed = false);
-
-uint64 ShortestPathProperties(uint64 props, bool tree = false);
-
-uint64 SynchronizeProperties(uint64 inprops);
-
-uint64 UnionProperties(uint64 inprops1, uint64 inprops2, bool delayed = false);
-
-// Definitions of inlined functions.
-
-uint64 SetStartProperties(uint64 inprops) {
-  auto outprops = inprops & kSetStartProperties;
-  if (inprops & kAcyclic) {
-    outprops |= kInitialAcyclic;
-  }
-  return outprops;
-}
-
-uint64 AddStateProperties(uint64 inprops) {
-  return inprops & kAddStateProperties;
-}
-
-uint64 DeleteStatesProperties(uint64 inprops) {
-  return inprops & kDeleteStatesProperties;
-}
-
-uint64 DeleteAllStatesProperties(uint64 inprops, uint64 staticprops) {
-  const auto outprops = inprops & kError;
-  return outprops | kNullProperties | staticprops;
-}
-
-uint64 DeleteArcsProperties(uint64 inprops) {
-  return inprops & kDeleteArcsProperties;
-}
-
-// Definitions of template functions.
-
-template <typename Weight>
-uint64 SetFinalProperties(uint64 inprops, Weight old_weight,
-                          Weight new_weight) {
-  auto outprops = inprops;
-  if (old_weight != Weight::Zero() && old_weight != Weight::One()) {
-    outprops &= ~kWeighted;
-  }
-  if (new_weight != Weight::Zero() && new_weight != Weight::One()) {
-    outprops |= kWeighted;
-    outprops &= ~kUnweighted;
-  }
-  outprops &= kSetFinalProperties | kWeighted | kUnweighted;
-  return outprops;
-}
-
-/// Gets the properties for the MutableFst::AddArc method.
-///
-/// \param inprops  the current properties of the FST
-/// \param s        the ID of the state to which an arc is being added.
-/// \param arc      the arc being added to the state with the specified ID
-/// \param prev_arc the previously-added (or "last") arc of state s, or nullptr
-//                  if s currently has no arcs.
-template <typename Arc>
-uint64 AddArcProperties(uint64 inprops, typename Arc::StateId s,
-                        const Arc &arc, const Arc *prev_arc) {
-  using Weight = typename Arc::Weight;
-  auto outprops = inprops;
-  if (arc.ilabel != arc.olabel) {
-    outprops |= kNotAcceptor;
-    outprops &= ~kAcceptor;
-  }
-  if (arc.ilabel == 0) {
-    outprops |= kIEpsilons;
-    outprops &= ~kNoIEpsilons;
-    if (arc.olabel == 0) {
-      outprops |= kEpsilons;
-      outprops &= ~kNoEpsilons;
-    }
-  }
-  if (arc.olabel == 0) {
-    outprops |= kOEpsilons;
-    outprops &= ~kNoOEpsilons;
-  }
-  if (prev_arc) {
-    if (prev_arc->ilabel > arc.ilabel) {
-      outprops |= kNotILabelSorted;
-      outprops &= ~kILabelSorted;
-    }
-    if (prev_arc->olabel > arc.olabel) {
-      outprops |= kNotOLabelSorted;
-      outprops &= ~kOLabelSorted;
-    }
-  }
-  if (arc.weight != Weight::Zero() && arc.weight != Weight::One()) {
-    outprops |= kWeighted;
-    outprops &= ~kUnweighted;
-  }
-  if (arc.nextstate <= s) {
-    outprops |= kNotTopSorted;
-    outprops &= ~kTopSorted;
-  }
-  outprops &= kAddArcProperties | kAcceptor | kNoEpsilons | kNoIEpsilons |
-              kNoOEpsilons | kILabelSorted | kOLabelSorted | kUnweighted |
-              kTopSorted;
-  if (outprops & kTopSorted) {
-    outprops |= kAcyclic | kInitialAcyclic;
-  }
-  return outprops;
-}
-
-FST_API extern const char *PropertyNames[];
-
-}  // namespace fst
-
-#endif  // FST_PROPERTIES_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/prune.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/prune.h
deleted file mode 100644
index b50e1f97..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/prune.h
+++ /dev/null
@@ -1,325 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions implementing pruning.
-
-#ifndef FST_PRUNE_H_
-#define FST_PRUNE_H_
-
-#include <type_traits>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arcfilter.h>
-#include <fst/heap.h>
-#include <fst/shortest-distance.h>
-
-
-namespace fst {
-namespace internal {
-
-template <class StateId, class Weight>
-class PruneCompare {
- public:
-  PruneCompare(const std::vector<Weight> &idistance,
-               const std::vector<Weight> &fdistance)
-      : idistance_(idistance), fdistance_(fdistance) {}
-
-  bool operator()(const StateId x, const StateId y) const {
-    const auto wx = Times(IDistance(x), FDistance(x));
-    const auto wy = Times(IDistance(y), FDistance(y));
-    return less_(wx, wy);
-  }
-
- private:
-  Weight IDistance(const StateId s) const {
-    return s < idistance_.size() ? idistance_[s] : Weight::Zero();
-  }
-
-  Weight FDistance(const StateId s) const {
-    return s < fdistance_.size() ? fdistance_[s] : Weight::Zero();
-  }
-
-  const std::vector<Weight> &idistance_;
-  const std::vector<Weight> &fdistance_;
-  NaturalLess<Weight> less_;
-};
-
-}  // namespace internal
-
-template <class Arc, class ArcFilter>
-struct PruneOptions {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  PruneOptions(const Weight &weight_threshold, StateId state_threshold,
-               ArcFilter filter, std::vector<Weight> *distance = nullptr,
-               float delta = kDelta, bool threshold_initial = false)
-      : weight_threshold(std::move(weight_threshold)),
-        state_threshold(state_threshold),
-        filter(std::move(filter)),
-        distance(distance),
-        delta(delta),
-        threshold_initial(threshold_initial) {}
-
-  // Pruning weight threshold.
-  Weight weight_threshold;
-  // Pruning state threshold.
-  StateId state_threshold;
-  // Arc filter.
-  ArcFilter filter;
-  // If non-zero, passes in pre-computed shortest distance to final states.
-  const std::vector<Weight> *distance;
-  // Determines the degree of convergence required when computing shortest
-  // distances.
-  float delta;
-  // Determines if the shortest path weight is left (true) or right
-  // (false) multiplied by the threshold to get the limit for
-  // keeping a state or arc (matters if the semiring is not
-  // commutative).
-  bool threshold_initial;
-};
-
-// Pruning algorithm: this version modifies its input and it takes an options
-// class as an argument. After pruning the FST contains states and arcs that
-// belong to a successful path in the FST whose weight is no more than the
-// weight of the shortest path Times() the provided weight threshold. When the
-// state threshold is not kNoStateId, the output FST is further restricted to
-// have no more than the number of states in opts.state_threshold. Weights must
-// have the path property. The weight of any cycle needs to be bounded; i.e.,
-//
-//   Plus(weight, Weight::One()) == Weight::One()
-template <class Arc, class ArcFilter>
-void Prune(MutableFst<Arc> *fst, const PruneOptions<Arc, ArcFilter> &opts) {
- if ((Arc::Weight::Properties() & kPath) == kPath) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using StateHeap = Heap<StateId, internal::PruneCompare<StateId, Weight>>;
-  auto ns = fst->NumStates();
-  if (ns < 1) return;
-  std::vector<Weight> idistance(ns, Weight::Zero());
-  std::vector<Weight> tmp;
-  if (!opts.distance) {
-    tmp.reserve(ns);
-    ShortestDistance(*fst, &tmp, true, opts.delta);
-  }
-  const auto *fdistance = opts.distance ? opts.distance : &tmp;
-  if ((opts.state_threshold == 0) || (fdistance->size() <= fst->Start()) ||
-      ((*fdistance)[fst->Start()] == Weight::Zero())) {
-    fst->DeleteStates();
-    return;
-  }
-  internal::PruneCompare<StateId, Weight> compare(idistance, *fdistance);
-  StateHeap heap(compare);
-  std::vector<bool> visited(ns, false);
-  std::vector<size_t> enqueued(ns, StateHeap::kNoKey);
-  std::vector<StateId> dead;
-  dead.push_back(fst->AddState());
-  NaturalLess<Weight> less;
-  auto s = fst->Start();
-  const auto limit = opts.threshold_initial ?
-      Times(opts.weight_threshold, (*fdistance)[s]) :
-      Times((*fdistance)[s], opts.weight_threshold);
-  StateId num_visited = 0;
-
-  if (!less(limit, (*fdistance)[s])) {
-    idistance[s] = Weight::One();
-    enqueued[s] = heap.Insert(s);
-    ++num_visited;
-  }
-  while (!heap.Empty()) {
-    s = heap.Top();
-    heap.Pop();
-    enqueued[s] = StateHeap::kNoKey;
-    visited[s] = true;
-    if (less(limit, Times(idistance[s], fst->Final(s)))) {
-      fst->SetFinal(s, Weight::Zero());
-    }
-    for (MutableArcIterator<MutableFst<Arc>> aiter(fst, s); !aiter.Done();
-         aiter.Next()) {
-      auto arc = aiter.Value();  // Copy intended.
-      if (!opts.filter(arc)) continue;
-      const auto weight = Times(Times(idistance[s], arc.weight),
-                                arc.nextstate < fdistance->size() ?
-                                (*fdistance)[arc.nextstate] : Weight::Zero());
-      if (less(limit, weight)) {
-        arc.nextstate = dead[0];
-        aiter.SetValue(arc);
-        continue;
-      }
-      if (less(Times(idistance[s], arc.weight), idistance[arc.nextstate])) {
-        idistance[arc.nextstate] = Times(idistance[s], arc.weight);
-      }
-      if (visited[arc.nextstate]) continue;
-      if ((opts.state_threshold != kNoStateId) &&
-          (num_visited >= opts.state_threshold)) {
-        continue;
-      }
-      if (enqueued[arc.nextstate] == StateHeap::kNoKey) {
-        enqueued[arc.nextstate] = heap.Insert(arc.nextstate);
-        ++num_visited;
-      } else {
-        heap.Update(enqueued[arc.nextstate], arc.nextstate);
-      }
-    }
-  }
-  for (StateId i = 0; i < visited.size(); ++i) {
-    if (!visited[i]) dead.push_back(i);
-  }
-  fst->DeleteStates(dead);
- } else {  // (Arc::Weight::Properties() & kPath) != kPath
-  FSTERROR() << "Prune: Weight needs to have the path property: "
-             << Arc::Weight::Type();
-  fst->SetProperties(kError, kError);
- }
-}
-
-// Pruning algorithm: this version modifies its input and takes the
-// pruning threshold as an argument. It deletes states and arcs in the
-// FST that do not belong to a successful path whose weight is more
-// than the weight of the shortest path Times() the provided weight
-// threshold. When the state threshold is not kNoStateId, the output
-// FST is further restricted to have no more than the number of states
-// in opts.state_threshold. Weights must have the path property. The
-// weight of any cycle needs to be bounded; i.e.,
-//
-//   Plus(weight, Weight::One()) == Weight::One()
-template <class Arc>
-void Prune(MutableFst<Arc> *fst, typename Arc::Weight weight_threshold,
-           typename Arc::StateId state_threshold = kNoStateId,
-           double delta = kDelta) {
-  const PruneOptions<Arc, AnyArcFilter<Arc>> opts(
-      weight_threshold, state_threshold, AnyArcFilter<Arc>(), nullptr, delta);
-  Prune(fst, opts);
-}
-
-// Pruning algorithm: this version writes the pruned input FST to an
-// output MutableFst and it takes an options class as an argument. The
-// output FST contains states and arcs that belong to a successful
-// path in the input FST whose weight is more than the weight of the
-// shortest path Times() the provided weight threshold. When the state
-// threshold is not kNoStateId, the output FST is further restricted
-// to have no more than the number of states in
-// opts.state_threshold. Weights have the path property.  The weight
-// of any cycle needs to be bounded; i.e.,
-//
-//   Plus(weight, Weight::One()) == Weight::One()
-template <class Arc, class ArcFilter>
-void Prune(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-           const PruneOptions<Arc, ArcFilter> &opts) {
- if ((Arc::Weight::Properties() & kPath) == kPath) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using StateHeap = Heap<StateId, internal::PruneCompare<StateId, Weight>>;
-  ofst->DeleteStates();
-  ofst->SetInputSymbols(ifst.InputSymbols());
-  ofst->SetOutputSymbols(ifst.OutputSymbols());
-  if (ifst.Start() == kNoStateId) return;
-  NaturalLess<Weight> less;
-  if (less(opts.weight_threshold, Weight::One()) ||
-      (opts.state_threshold == 0)) {
-    return;
-  }
-  std::vector<Weight> idistance;
-  std::vector<Weight> tmp;
-  if (!opts.distance) ShortestDistance(ifst, &tmp, true, opts.delta);
-  const auto *fdistance = opts.distance ? opts.distance : &tmp;
-  if ((fdistance->size() <= ifst.Start()) ||
-      ((*fdistance)[ifst.Start()] == Weight::Zero())) {
-    return;
-  }
-  internal::PruneCompare<StateId, Weight> compare(idistance, *fdistance);
-  StateHeap heap(compare);
-  std::vector<StateId> copy;
-  std::vector<size_t> enqueued;
-  std::vector<bool> visited;
-  auto s = ifst.Start();
-  const auto limit = opts.threshold_initial ?
-      Times(opts.weight_threshold, (*fdistance)[s]) :
-      Times((*fdistance)[s], opts.weight_threshold);
-  while (copy.size() <= s) copy.push_back(kNoStateId);
-  copy[s] = ofst->AddState();
-  ofst->SetStart(copy[s]);
-  while (idistance.size() <= s) idistance.push_back(Weight::Zero());
-  idistance[s] = Weight::One();
-  while (enqueued.size() <= s) {
-    enqueued.push_back(StateHeap::kNoKey);
-    visited.push_back(false);
-  }
-  enqueued[s] = heap.Insert(s);
-  while (!heap.Empty()) {
-    s = heap.Top();
-    heap.Pop();
-    enqueued[s] = StateHeap::kNoKey;
-    visited[s] = true;
-    if (!less(limit, Times(idistance[s], ifst.Final(s)))) {
-      ofst->SetFinal(copy[s], ifst.Final(s));
-    }
-    for (ArcIterator<Fst<Arc>> aiter(ifst, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      if (!opts.filter(arc)) continue;
-      const auto weight = Times(Times(idistance[s], arc.weight),
-                                arc.nextstate < fdistance->size() ?
-                                (*fdistance)[arc.nextstate] : Weight::Zero());
-      if (less(limit, weight)) continue;
-      if ((opts.state_threshold != kNoStateId) &&
-          (ofst->NumStates() >= opts.state_threshold)) {
-        continue;
-      }
-      while (idistance.size() <= arc.nextstate) {
-        idistance.push_back(Weight::Zero());
-      }
-      if (less(Times(idistance[s], arc.weight), idistance[arc.nextstate])) {
-        idistance[arc.nextstate] = Times(idistance[s], arc.weight);
-      }
-      while (copy.size() <= arc.nextstate) copy.push_back(kNoStateId);
-      if (copy[arc.nextstate] == kNoStateId) {
-        copy[arc.nextstate] = ofst->AddState();
-      }
-      ofst->AddArc(copy[s], Arc(arc.ilabel, arc.olabel, arc.weight,
-                                copy[arc.nextstate]));
-      while (enqueued.size() <= arc.nextstate) {
-        enqueued.push_back(StateHeap::kNoKey);
-        visited.push_back(false);
-      }
-      if (visited[arc.nextstate]) continue;
-      if (enqueued[arc.nextstate] == StateHeap::kNoKey) {
-        enqueued[arc.nextstate] = heap.Insert(arc.nextstate);
-      } else {
-        heap.Update(enqueued[arc.nextstate], arc.nextstate);
-      }
-    }
-  }
- } else {  // (Arc::Weight::Properties() & kPath) != kPath
-  FSTERROR() << "Prune: Weight needs to have the path property: "
-             << Arc::Weight::Type();
-  ofst->SetProperties(kError, kError);
- }
-}
-
-// Pruning algorithm: this version writes the pruned input FST to an
-// output MutableFst and simply takes the pruning threshold as an
-// argument. The output FST contains states and arcs that belong to a
-// successful path in the input FST whose weight is no more than the
-// weight of the shortest path Times() the provided weight
-// threshold. When the state threshold is not kNoStateId, the output
-// FST is further restricted to have no more than the number of states
-// in opts.state_threshold. Weights must have the path property. The
-// weight of any cycle needs to be bounded; i.e.,
-//
-// Plus(weight, Weight::One()) = Weight::One();
-template <class Arc>
-void Prune(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-           typename Arc::Weight weight_threshold,
-           typename Arc::StateId state_threshold = kNoStateId,
-           float delta = kDelta) {
-  const PruneOptions<Arc, AnyArcFilter<Arc>> opts(
-      weight_threshold, state_threshold, AnyArcFilter<Arc>(), nullptr, delta);
-  Prune(ifst, ofst, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_PRUNE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/push.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/push.h
deleted file mode 100644
index 1f772739..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/push.h
+++ /dev/null
@@ -1,155 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to reweight/push an FST, and utility functions to weigh and reweight
-// an FST.
-
-#ifndef FST_PUSH_H_
-#define FST_PUSH_H_
-
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arc-map.h>
-#include <fst/factor-weight.h>
-#include <fst/fst.h>
-#include <fst/reweight.h>
-#include <fst/shortest-distance.h>
-
-
-namespace fst {
-
-// Computes the total weight (sum of the weights of all accepting paths) from
-// the output of ShortestDistance, using the shortest distance from the final
-// state when reverse is true and from the initial state otherwise.
-template <class Arc>
-typename Arc::Weight ComputeTotalWeight(
-    const Fst<Arc> &fst, const std::vector<typename Arc::Weight> &distance,
-    bool reverse) {
-  if (reverse) {
-    return fst.Start() < distance.size() ? distance[fst.Start()]
-                                         : Arc::Weight::Zero();
-  }
-  auto sum = Arc::Weight::Zero();
-  for (typename Arc::StateId s = 0; s < distance.size(); ++s) {
-    sum = Plus(sum, Times(distance[s], fst.Final(s)));
-  }
-  return sum;
-}
-
-// Divides the weight of every accepting path by a fixed weight. This weight
-// is also divided at the final state if at_final is true and at the initial
-// state otherwise.
-template <class Arc>
-void RemoveWeight(MutableFst<Arc> *fst, const typename Arc::Weight &weight,
-                  bool at_final) {
-  using Weight = typename Arc::Weight;
-  if ((weight == Weight::One()) || (weight == Weight::Zero())) return;
-  if (at_final) {
-    for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
-         siter.Next()) {
-      fst->SetFinal(siter.Value(),
-                    Divide(fst->Final(siter.Value()), weight, DIVIDE_RIGHT));
-    }
-  } else {
-    const auto start = fst->Start();
-    for (MutableArcIterator<MutableFst<Arc>> aiter(fst, start); !aiter.Done();
-         aiter.Next()) {
-      auto arc = aiter.Value();
-      arc.weight = Divide(arc.weight, weight, DIVIDE_LEFT);
-      aiter.SetValue(arc);
-    }
-    fst->SetFinal(start, Divide(fst->Final(start), weight, DIVIDE_LEFT));
-  }
-}
-
-// Pushes the weights in FST in the direction defined by TYPE. If
-// pushing towards the initial state, the sum of the weight of the
-// outgoing transitions and final weight at a non-initial state is
-// equal to One() in the resulting machine. If pushing towards the
-// final state, the same property holds on the reverse machine.
-//
-// Weight needs to be left distributive when pushing towards the
-// initial state and right distributive when pushing towards the final
-// states.
-template <class Arc>
-void Push(MutableFst<Arc> *fst, ReweightType type, float delta = kDelta,
-          bool remove_total_weight = false) {
-  using Weight = typename Arc::Weight;
-  std::vector<Weight> distance;
-  ShortestDistance(*fst, &distance, type == REWEIGHT_TO_INITIAL, delta);
-  auto total_weight = Weight::One();
-  if (remove_total_weight) {
-    total_weight =
-        ComputeTotalWeight(*fst, distance, type == REWEIGHT_TO_INITIAL);
-  }
-  Reweight(fst, distance, type);
-  if (remove_total_weight) {
-    RemoveWeight(fst, total_weight, type == REWEIGHT_TO_FINAL);
-  }
-}
-
-constexpr uint32 kPushWeights = 0x0001;
-constexpr uint32 kPushLabels = 0x0002;
-constexpr uint32 kPushRemoveTotalWeight = 0x0004;
-constexpr uint32 kPushRemoveCommonAffix = 0x0008;
-
-// Pushes the weights and/or labels of the input FST into the output
-// mutable FST by pushing weights and/or labels (as determined by the
-// ptype argument) towards the initial state or final states (as
-// determined by the rtype template parameter). The weight type must
-// be left distributive when pushing weights towards the initial state, and
-// right distribution when pushing weights towards the final states.
-template <class Arc, ReweightType rtype>
-void Push(const Fst<Arc> &ifst, MutableFst<Arc> *ofst, uint32 ptype,
-          float delta = kDelta) {
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-  if ((ptype & (kPushWeights | kPushLabels)) == kPushWeights) {
-    *ofst = ifst;
-    Push(ofst, rtype, delta, ptype & kPushRemoveTotalWeight);
-  } else if (ptype & kPushLabels) {
-    const auto gtype =
-        rtype == REWEIGHT_TO_INITIAL ? GALLIC_LEFT : GALLIC_RIGHT;
-    using GallicWeight = typename GallicArc<Arc, gtype>::Weight;
-    std::vector<GallicWeight> gdistance;
-    VectorFst<GallicArc<Arc, gtype>> gfst;
-    ArcMap(ifst, &gfst, ToGallicMapper<Arc, gtype>());
-    if (ptype & kPushWeights) {
-      ShortestDistance(gfst, &gdistance, rtype == REWEIGHT_TO_INITIAL, delta);
-    } else {
-      ArcMapFst<Arc, Arc, RmWeightMapper<Arc>> uwfst(ifst,
-                                                      RmWeightMapper<Arc>());
-      ArcMapFst<Arc, GallicArc<Arc, gtype>, ToGallicMapper<Arc, gtype>> guwfst(
-          uwfst, ToGallicMapper<Arc, gtype>());
-      ShortestDistance(guwfst, &gdistance, rtype == REWEIGHT_TO_INITIAL, delta);
-    }
-    auto total_weight = GallicWeight::One();
-    if (ptype & (kPushRemoveTotalWeight | kPushRemoveCommonAffix)) {
-      total_weight =
-          ComputeTotalWeight(gfst, gdistance, rtype == REWEIGHT_TO_INITIAL);
-      total_weight = GallicWeight(
-          ptype & kPushRemoveCommonAffix
-              ? total_weight.Value1()
-              : StringWeight<Label, GallicStringType(gtype)>::One(),
-          ptype & kPushRemoveTotalWeight ? total_weight.Value2()
-                                         : Weight::One());
-    }
-    Reweight(&gfst, gdistance, rtype);
-    if (ptype & (kPushRemoveTotalWeight | kPushRemoveCommonAffix)) {
-      RemoveWeight(&gfst, total_weight, rtype == REWEIGHT_TO_FINAL);
-    }
-    FactorWeightFst<GallicArc<Arc, gtype>, GallicFactor<Label, Weight, gtype>>
-        fwfst(gfst);
-    ArcMap(fwfst, ofst, FromGallicMapper<Arc, gtype>());
-    ofst->SetOutputSymbols(ifst.OutputSymbols());
-  } else {
-    LOG(WARNING) << "Push: pushing type is set to 0, so not pushing";
-    *ofst = ifst;
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_PUSH_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/queue.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/queue.h
deleted file mode 100644
index bc74a64e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/queue.h
+++ /dev/null
@@ -1,858 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes for various FST state queues with a unified interface.
-
-#ifndef FST_QUEUE_H_
-#define FST_QUEUE_H_
-
-#include <deque>
-#include <memory>
-#include <type_traits>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arcfilter.h>
-#include <fst/connect.h>
-#include <fst/heap.h>
-#include <fst/topsort.h>
-
-
-namespace fst {
-
-// The Queue interface is:
-//
-// template <class S>
-// class Queue {
-//  public:
-//   using StateId = S;
-//
-//   // Constructor: may need args (e.g., FST, comparator) for some queues.
-//   Queue(...) override;
-//
-//   // Returns the head of the queue.
-//   StateId Head() const override;
-//
-//   // Inserts a state.
-//   void Enqueue(StateId s) override;
-//
-//   // Removes the head of the queue.
-//   void Dequeue() override;
-//
-//   // Updates ordering of state s when weight changes, if necessary.
-//   void Update(StateId s) override;
-//
-//   // Is the queue empty?
-//   bool Empty() const override;
-//
-//   // Removes all states from the queue.
-//   void Clear() override;
-// };
-
-// State queue types.
-enum QueueType {
-  TRIVIAL_QUEUE = 0,         // Single state queue.
-  FIFO_QUEUE = 1,            // First-in, first-out queue.
-  LIFO_QUEUE = 2,            // Last-in, first-out queue.
-  SHORTEST_FIRST_QUEUE = 3,  // Shortest-first queue.
-  TOP_ORDER_QUEUE = 4,       // Topologically-ordered queue.
-  STATE_ORDER_QUEUE = 5,     // State ID-ordered queue.
-  SCC_QUEUE = 6,             // Component graph top-ordered meta-queue.
-  AUTO_QUEUE = 7,            // Auto-selected queue.
-  OTHER_QUEUE = 8
-};
-
-// QueueBase, templated on the StateId, is a virtual base class shared by all
-// queues considered by AutoQueue.
-template <class S>
-class QueueBase {
- public:
-  using StateId = S;
-
-  virtual ~QueueBase() {}
-
-  // Concrete implementation.
-
-  explicit QueueBase(QueueType type) : queue_type_(type), error_(false) {}
-
-  void SetError(bool error) { error_ = error; }
-
-  bool Error() const { return error_; }
-
-  QueueType Type() const { return queue_type_; }
-
-  // Virtual interface.
-
-  virtual StateId Head() const = 0;
-  virtual void Enqueue(StateId) = 0;
-  virtual void Dequeue() = 0;
-  virtual void Update(StateId) = 0;
-  virtual bool Empty() const = 0;
-  virtual void Clear() = 0;
-
- private:
-  QueueType queue_type_;
-  bool error_;
-};
-
-// Trivial queue discipline; one may enqueue at most one state at a time. It
-// can be used for strongly connected components with only one state and no
-// self-loops.
-template <class S>
-class TrivialQueue final : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  TrivialQueue() : QueueBase<StateId>(TRIVIAL_QUEUE), front_(kNoStateId) {}
-
-  virtual ~TrivialQueue() = default;
-
-  StateId Head() const override { return front_; }
-
-  void Enqueue(StateId s) override { front_ = s; }
-
-  void Dequeue() override { front_ = kNoStateId; }
-
-  void Update(StateId) override {}
-
-  bool Empty() const override { return front_ == kNoStateId; }
-
-  void Clear() override { front_ = kNoStateId; }
-
- private:
-  StateId front_;
-};
-
-// First-in, first-out queue discipline.
-//
-// This is not a final class.
-template <class S>
-class FifoQueue : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  FifoQueue() : QueueBase<StateId>(FIFO_QUEUE) {}
-
-  virtual ~FifoQueue() = default;
-
-  StateId Head() const override { return queue_.back(); }
-
-  void Enqueue(StateId s) override { queue_.push_front(s); }
-
-  void Dequeue() override { queue_.pop_back(); }
-
-  void Update(StateId) override {}
-
-  bool Empty() const override { return queue_.empty(); }
-
-  void Clear() override { queue_.clear(); }
-
- private:
-  std::deque<StateId> queue_;
-};
-
-// Last-in, first-out queue discipline.
-template <class S>
-class LifoQueue final : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  LifoQueue() : QueueBase<StateId>(LIFO_QUEUE) {}
-
-  virtual ~LifoQueue() = default;
-
-  StateId Head() const override { return queue_.front(); }
-
-  void Enqueue(StateId s) override { queue_.push_front(s); }
-
-  void Dequeue() override { queue_.pop_front(); }
-
-  void Update(StateId) override {}
-
-  bool Empty() const override { return queue_.empty(); }
-
-  void Clear() override { queue_.clear(); }
-
- private:
-  std::deque<StateId> queue_;
-};
-
-// Shortest-first queue discipline, templated on the StateId and as well as a
-// comparison functor used to compare two StateIds. If a (single) state's order
-// changes, it can be reordered in the queue with a call to Update(). If update
-// is false, call to Update() does not reorder the queue.
-//
-// This is not a final class.
-template <typename S, typename Compare, bool update = true>
-class ShortestFirstQueue : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  static constexpr StateId kNoKey = -1;
-
-  explicit ShortestFirstQueue(Compare comp)
-      : QueueBase<StateId>(SHORTEST_FIRST_QUEUE), heap_(comp) {}
-
-  StateId Head() const override { return heap_.Top(); }
-
-  void Enqueue(StateId s) override {
-    if (update) {
-      for (StateId i = key_.size(); i <= s; ++i) key_.push_back(kNoKey);
-      key_[s] = heap_.Insert(s);
-    } else {
-      heap_.Insert(s);
-    }
-  }
-
-  void Dequeue() override {
-    if (update) {
-      key_[heap_.Pop()] = kNoKey;
-    } else {
-      heap_.Pop();
-    }
-  }
-
-  void Update(StateId s) override {
-    if (!update) return;
-    if (s >= key_.size() || key_[s] == kNoKey) {
-      Enqueue(s);
-    } else {
-      heap_.Update(key_[s], s);
-    }
-  }
-
-  bool Empty() const override { return heap_.Empty(); }
-
-  void Clear() override {
-    heap_.Clear();
-    if (update) key_.clear();
-  }
-
- private:
-  Heap<StateId, Compare> heap_;
-  std::vector<ssize_t> key_;
-};
-
-template <typename StateId, typename Compare, bool update>
-constexpr StateId ShortestFirstQueue<StateId, Compare, update>::kNoKey;
-
-namespace internal {
-
-// Given a vector that maps from states to weights, and a comparison functor
-// for weights, this class defines a comparison function object between states.
-template <typename StateId, typename Less>
-class StateWeightCompare {
- public:
-  using Weight = typename Less::Weight;
-
-  StateWeightCompare(const std::vector<Weight> &weights, const Less &less)
-      : weights_(weights), less_(less) {}
-
-  bool operator()(const StateId s1, const StateId s2) const {
-    return less_(weights_[s1], weights_[s2]);
-  }
-
- private:
-  // Borrowed references.
-  const std::vector<Weight> &weights_;
-  const Less &less_;
-};
-
-}  // namespace internal
-
-// Shortest-first queue discipline, templated on the StateId and Weight, is
-// specialized to use the weight's natural order for the comparison function.
-template <typename S, typename Weight>
-class NaturalShortestFirstQueue final
-    : public ShortestFirstQueue<
-          S, internal::StateWeightCompare<S, NaturalLess<Weight>>> {
- public:
-  using StateId = S;
-  using Compare = internal::StateWeightCompare<StateId, NaturalLess<Weight>>;
-
-  explicit NaturalShortestFirstQueue(const std::vector<Weight> &distance)
-      : ShortestFirstQueue<StateId, Compare>(Compare(distance, less_)) {}
-
-  virtual ~NaturalShortestFirstQueue() = default;
-
- private:
-  // This is non-static because the constructor for non-idempotent weights will
-  // result in a an error.
-  const NaturalLess<Weight> less_{};
-};
-
-// Topological-order queue discipline, templated on the StateId. States are
-// ordered in the queue topologically. The FST must be acyclic.
-template <class S>
-class TopOrderQueue final : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  // This constructor computes the topological order. It accepts an arc filter
-  // to limit the transitions considered in that computation (e.g., only the
-  // epsilon graph).
-  template <class Arc, class ArcFilter>
-  TopOrderQueue(const Fst<Arc> &fst, ArcFilter filter)
-      : QueueBase<StateId>(TOP_ORDER_QUEUE),
-        front_(0),
-        back_(kNoStateId),
-        order_(0),
-        state_(0) {
-    bool acyclic;
-    TopOrderVisitor<Arc> top_order_visitor(&order_, &acyclic);
-    DfsVisit(fst, &top_order_visitor, filter);
-    if (!acyclic) {
-      FSTERROR() << "TopOrderQueue: FST is not acyclic";
-      QueueBase<S>::SetError(true);
-    }
-    state_.resize(order_.size(), kNoStateId);
-  }
-
-  // This constructor is passed the pre-computed topological order.
-  explicit TopOrderQueue(const std::vector<StateId> &order)
-      : QueueBase<StateId>(TOP_ORDER_QUEUE),
-        front_(0),
-        back_(kNoStateId),
-        order_(order),
-        state_(order.size(), kNoStateId) {}
-
-  virtual ~TopOrderQueue() = default;
-
-  StateId Head() const override { return state_[front_]; }
-
-  void Enqueue(StateId s) override {
-    if (front_ > back_) {
-      front_ = back_ = order_[s];
-    } else if (order_[s] > back_) {
-      back_ = order_[s];
-    } else if (order_[s] < front_) {
-      front_ = order_[s];
-    }
-    state_[order_[s]] = s;
-  }
-
-  void Dequeue() override {
-    state_[front_] = kNoStateId;
-    while ((front_ <= back_) && (state_[front_] == kNoStateId)) ++front_;
-  }
-
-  void Update(StateId) override {}
-
-  bool Empty() const override { return front_ > back_; }
-
-  void Clear() override {
-    for (StateId s = front_; s <= back_; ++s) state_[s] = kNoStateId;
-    back_ = kNoStateId;
-    front_ = 0;
-  }
-
- private:
-  StateId front_;
-  StateId back_;
-  std::vector<StateId> order_;
-  std::vector<StateId> state_;
-};
-
-// State order queue discipline, templated on the StateId. States are ordered in
-// the queue by state ID.
-template <class S>
-class StateOrderQueue final : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  StateOrderQueue()
-      : QueueBase<StateId>(STATE_ORDER_QUEUE), front_(0), back_(kNoStateId) {}
-
-  virtual ~StateOrderQueue() = default;
-
-  StateId Head() const override { return front_; }
-
-  void Enqueue(StateId s) override {
-    if (front_ > back_) {
-      front_ = back_ = s;
-    } else if (s > back_) {
-      back_ = s;
-    } else if (s < front_) {
-      front_ = s;
-    }
-    while (enqueued_.size() <= s) enqueued_.push_back(false);
-    enqueued_[s] = true;
-  }
-
-  void Dequeue() override {
-    enqueued_[front_] = false;
-    while ((front_ <= back_) && (enqueued_[front_] == false)) ++front_;
-  }
-
-  void Update(StateId) override {}
-
-  bool Empty() const override { return front_ > back_; }
-
-  void Clear() override {
-    for (StateId i = front_; i <= back_; ++i) enqueued_[i] = false;
-    front_ = 0;
-    back_ = kNoStateId;
-  }
-
- private:
-  StateId front_;
-  StateId back_;
-  std::vector<bool> enqueued_;
-};
-
-// SCC topological-order meta-queue discipline, templated on the StateId and a
-// queue used inside each SCC. It visits the SCCs of an FST in topological
-// order. Its constructor is passed the queues to to use within an SCC.
-template <class S, class Queue>
-class SccQueue final : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  // Constructor takes a vector specifying the SCC number per state and a
-  // vector giving the queue to use per SCC number.
-  SccQueue(const std::vector<StateId> &scc,
-           std::vector<std::unique_ptr<Queue>> *queue)
-      : QueueBase<StateId>(SCC_QUEUE),
-        queue_(queue),
-        scc_(scc),
-        front_(0),
-        back_(kNoStateId) {}
-
-  virtual ~SccQueue() = default;
-
-  StateId Head() const override {
-    while ((front_ <= back_) &&
-           (((*queue_)[front_] && (*queue_)[front_]->Empty()) ||
-            (((*queue_)[front_] == nullptr) &&
-             ((front_ >= trivial_queue_.size()) ||
-              (trivial_queue_[front_] == kNoStateId))))) {
-      ++front_;
-    }
-    if ((*queue_)[front_]) {
-      return (*queue_)[front_]->Head();
-    } else {
-      return trivial_queue_[front_];
-    }
-  }
-
-  void Enqueue(StateId s) override {
-    if (front_ > back_) {
-      front_ = back_ = scc_[s];
-    } else if (scc_[s] > back_) {
-      back_ = scc_[s];
-    } else if (scc_[s] < front_) {
-      front_ = scc_[s];
-    }
-    if ((*queue_)[scc_[s]]) {
-      (*queue_)[scc_[s]]->Enqueue(s);
-    } else {
-      while (trivial_queue_.size() <= scc_[s]) {
-        trivial_queue_.push_back(kNoStateId);
-      }
-      trivial_queue_[scc_[s]] = s;
-    }
-  }
-
-  void Dequeue() override {
-    if ((*queue_)[front_]) {
-      (*queue_)[front_]->Dequeue();
-    } else if (front_ < trivial_queue_.size()) {
-      trivial_queue_[front_] = kNoStateId;
-    }
-  }
-
-  void Update(StateId s) override {
-    if ((*queue_)[scc_[s]]) (*queue_)[scc_[s]]->Update(s);
-  }
-
-  bool Empty() const override {
-    // Queue SCC number back_ is not empty unless back_ == front_.
-    if (front_ < back_) {
-      return false;
-    } else if (front_ > back_) {
-      return true;
-    } else if ((*queue_)[front_]) {
-      return (*queue_)[front_]->Empty();
-    } else {
-      return (front_ >= trivial_queue_.size()) ||
-             (trivial_queue_[front_] == kNoStateId);
-    }
-  }
-
-  void Clear() override {
-    for (StateId i = front_; i <= back_; ++i) {
-      if ((*queue_)[i]) {
-        (*queue_)[i]->Clear();
-      } else if (i < trivial_queue_.size()) {
-        trivial_queue_[i] = kNoStateId;
-      }
-    }
-    front_ = 0;
-    back_ = kNoStateId;
-  }
-
- private:
-  std::vector<std::unique_ptr<Queue>> *queue_;
-  const std::vector<StateId> &scc_;
-  mutable StateId front_;
-  StateId back_;
-  std::vector<StateId> trivial_queue_;
-};
-
-// Automatic queue discipline. It selects a queue discipline for a given FST
-// based on its properties.
-template <class S>
-class AutoQueue final : public QueueBase<S> {
- public:
-  using StateId = S;
-
-  // This constructor takes a state distance vector that, if non-null and if
-  // the Weight type has the path property, will entertain the shortest-first
-  // queue using the natural order w.r.t to the distance.
-  template <class Arc, class ArcFilter>
-  AutoQueue(const Fst<Arc> &fst,
-            const std::vector<typename Arc::Weight> *distance, ArcFilter filter)
-      : QueueBase<StateId>(AUTO_QUEUE) {
-    using Weight = typename Arc::Weight;
-    // TrivialLess is never instantiated since the construction of Less is
-    // guarded by Properties() & kPath.  It is only here to avoid instantiating
-    // NaturalLess for non-path weights.
-    struct TrivialLess {
-      using Weight = typename Arc::Weight;
-      bool operator()(const Weight &, const Weight &) const { return false; }
-    };
-    using Less =
-        typename std::conditional<(Weight::Properties() & kPath) == kPath,
-                                  NaturalLess<Weight>, TrivialLess>::type;
-    using Compare = internal::StateWeightCompare<StateId, Less>;
-    // First checks if the FST is known to have these properties.
-    const auto props =
-        fst.Properties(kAcyclic | kCyclic | kTopSorted | kUnweighted, false);
-    if ((props & kTopSorted) || fst.Start() == kNoStateId) {
-      queue_.reset(new StateOrderQueue<StateId>());
-      VLOG(2) << "AutoQueue: using state-order discipline";
-    } else if (props & kAcyclic) {
-      queue_.reset(new TopOrderQueue<StateId>(fst, filter));
-      VLOG(2) << "AutoQueue: using top-order discipline";
-    } else if ((props & kUnweighted) && (Weight::Properties() & kIdempotent)) {
-      queue_.reset(new LifoQueue<StateId>());
-      VLOG(2) << "AutoQueue: using LIFO discipline";
-    } else {
-      uint64 properties;
-      // Decomposes into strongly-connected components.
-      SccVisitor<Arc> scc_visitor(&scc_, nullptr, nullptr, &properties);
-      DfsVisit(fst, &scc_visitor, filter);
-      auto nscc = *std::max_element(scc_.begin(), scc_.end()) + 1;
-      std::vector<QueueType> queue_types(nscc);
-      std::unique_ptr<Less> less;
-      std::unique_ptr<Compare> comp;
-      if (distance && (Weight::Properties() & kPath)) {
-        less.reset(new Less);
-        comp.reset(new Compare(*distance, *less));
-      }
-      // Finds the queue type to use per SCC.
-      bool unweighted;
-      bool all_trivial;
-      SccQueueType(fst, scc_, &queue_types, filter, less.get(), &all_trivial,
-                   &unweighted);
-      // If unweighted and semiring is idempotent, uses LIFO queue.
-      if (unweighted) {
-        queue_.reset(new LifoQueue<StateId>());
-        VLOG(2) << "AutoQueue: using LIFO discipline";
-        return;
-      }
-      // If all the SCC are trivial, the FST is acyclic and the scc number gives
-      // the topological order.
-      if (all_trivial) {
-        queue_.reset(new TopOrderQueue<StateId>(scc_));
-        VLOG(2) << "AutoQueue: using top-order discipline";
-        return;
-      }
-      VLOG(2) << "AutoQueue: using SCC meta-discipline";
-      queues_.resize(nscc);
-      for (StateId i = 0; i < nscc; ++i) {
-        switch (queue_types[i]) {
-          case TRIVIAL_QUEUE:
-            queues_[i].reset();
-            VLOG(3) << "AutoQueue: SCC #" << i << ": using trivial discipline";
-            break;
-          case SHORTEST_FIRST_QUEUE:
-            queues_[i].reset(
-                new ShortestFirstQueue<StateId, Compare, false>(*comp));
-            VLOG(3) << "AutoQueue: SCC #" << i
-                    << ": using shortest-first discipline";
-            break;
-          case LIFO_QUEUE:
-            queues_[i].reset(new LifoQueue<StateId>());
-            VLOG(3) << "AutoQueue: SCC #" << i << ": using LIFO discipline";
-            break;
-          case FIFO_QUEUE:
-          default:
-            queues_[i].reset(new FifoQueue<StateId>());
-            VLOG(3) << "AutoQueue: SCC #" << i << ": using FIFO discipine";
-            break;
-        }
-      }
-      queue_.reset(new SccQueue<StateId, QueueBase<StateId>>(scc_, &queues_));
-    }
-  }
-
-  virtual ~AutoQueue() = default;
-
-  StateId Head() const override { return queue_->Head(); }
-
-  void Enqueue(StateId s) override { queue_->Enqueue(s); }
-
-  void Dequeue() override { queue_->Dequeue(); }
-
-  void Update(StateId s) override { queue_->Update(s); }
-
-  bool Empty() const override { return queue_->Empty(); }
-
-  void Clear() override { queue_->Clear(); }
-
- private:
-  template <class Arc, class ArcFilter, class Less>
-  static void SccQueueType(const Fst<Arc> &fst, const std::vector<StateId> &scc,
-                           std::vector<QueueType> *queue_types,
-                           ArcFilter filter, Less *less, bool *all_trivial,
-                           bool *unweighted);
-
-  std::unique_ptr<QueueBase<StateId>> queue_;
-  std::vector<std::unique_ptr<QueueBase<StateId>>> queues_;
-  std::vector<StateId> scc_;
-};
-
-// Examines the states in an FST's strongly connected components and determines
-// which type of queue to use per SCC. Stores result as a vector of QueueTypes
-// which is assumed to have length equal to the number of SCCs. An arc filter
-// is used to limit the transitions considered (e.g., only the epsilon graph).
-// The argument all_trivial is set to true if every queue is the trivial queue.
-// The argument unweighted is set to true if the semiring is idempotent and all
-// the arc weights are equal to Zero() or One().
-template <class StateId>
-template <class Arc, class ArcFilter, class Less>
-void AutoQueue<StateId>::SccQueueType(const Fst<Arc> &fst,
-                                      const std::vector<StateId> &scc,
-                                      std::vector<QueueType> *queue_type,
-                                      ArcFilter filter, Less *less,
-                                      bool *all_trivial, bool *unweighted) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  *all_trivial = true;
-  *unweighted = true;
-  for (StateId i = 0; i < queue_type->size(); ++i) {
-    (*queue_type)[i] = TRIVIAL_QUEUE;
-  }
-  for (StateIterator<Fst<Arc>> sit(fst); !sit.Done(); sit.Next()) {
-    const auto state = sit.Value();
-    for (ArcIterator<Fst<Arc>> ait(fst, state); !ait.Done(); ait.Next()) {
-      const auto &arc = ait.Value();
-      if (!filter(arc)) continue;
-      if (scc[state] == scc[arc.nextstate]) {
-        auto &type = (*queue_type)[scc[state]];
-        if (!less || ((*less)(arc.weight, Weight::One()))) {
-          type = FIFO_QUEUE;
-        } else if ((type == TRIVIAL_QUEUE) || (type == LIFO_QUEUE)) {
-          if (!(Weight::Properties() & kIdempotent) ||
-              (arc.weight != Weight::Zero() && arc.weight != Weight::One())) {
-            type = SHORTEST_FIRST_QUEUE;
-          } else {
-            type = LIFO_QUEUE;
-          }
-        }
-        if (type != TRIVIAL_QUEUE) *all_trivial = false;
-      }
-      if (!(Weight::Properties() & kIdempotent) ||
-          (arc.weight != Weight::Zero() && arc.weight != Weight::One())) {
-        *unweighted = false;
-      }
-    }
-  }
-}
-
-// An A* estimate is a function object that maps from a state ID to a an
-// estimate of the shortest distance to the final states.
-
-// The trivial A* estimate is always One().
-template <typename StateId, typename Weight>
-struct TrivialAStarEstimate {
-  Weight operator()(StateId) const { return Weight::One(); }
-};
-
-// Given a vector that maps from states to weights representing the shortest
-// distance from the initial state, a comparison function object between
-// weights, and an estimate of the shortest distance to the final states, this
-// class defines a comparison function object between states.
-template <typename S, typename Less, typename Estimate>
-class AStarWeightCompare {
- public:
-  using StateId = S;
-  using Weight = typename Less::Weight;
-
-  AStarWeightCompare(const std::vector<Weight> &weights, const Less &less,
-                     const Estimate &estimate)
-      : weights_(weights), less_(less), estimate_(estimate) {}
-
-  bool operator()(const StateId s1, const StateId s2) const {
-    const auto w1 = Times(weights_[s1], estimate_(s1));
-    const auto w2 = Times(weights_[s2], estimate_(s2));
-    return less_(w1, w2);
-  }
-
- private:
-  // Borrowed references.
-  const std::vector<Weight> &weights_;
-  const Less &less_;
-  const Estimate &estimate_;
-};
-
-// A* queue discipline templated on StateId, Weight, and Estimate.
-template <typename S, typename Weight, typename Estimate>
-class NaturalAStarQueue final
-    : public ShortestFirstQueue<
-          S, AStarWeightCompare<S, NaturalLess<Weight>, Estimate>> {
- public:
-  using StateId = S;
-  using Compare = AStarWeightCompare<StateId, NaturalLess<Weight>, Estimate>;
-
-  NaturalAStarQueue(const std::vector<Weight> &distance,
-                    const Estimate &estimate)
-      : ShortestFirstQueue<StateId, Compare>(
-            Compare(distance, less_, estimate)) {}
-
-  virtual ~NaturalAStarQueue() = default;
-
- private:
-  // This is non-static because the constructor for non-idempotent weights will
-  // result in a an error.
-  const NaturalLess<Weight> less_{};
-};
-
-// A state equivalence class is a function object that maps from a state ID to
-// an equivalence class (state) ID. The trivial equivalence class maps a state
-// ID to itself.
-template <typename StateId>
-struct TrivialStateEquivClass {
-  StateId operator()(StateId s) const { return s; }
-};
-
-// Distance-based pruning queue discipline: Enqueues a state only when its
-// shortest distance (so far), as specified by distance, is less than (as
-// specified by comp) the shortest distance Times() the threshold to any state
-// in the same equivalence class, as specified by the functor class_func. The
-// underlying queue discipline is specified by queue. The ownership of queue is
-// given to this class.
-//
-// This is not a final class.
-template <typename Queue, typename Less, typename ClassFnc>
-class PruneQueue : public QueueBase<typename Queue::StateId> {
- public:
-  using StateId = typename Queue::StateId;
-  using Weight = typename Less::Weight;
-
-  PruneQueue(const std::vector<Weight> &distance, Queue *queue,
-             const Less &less, const ClassFnc &class_fnc, Weight threshold)
-      : QueueBase<StateId>(OTHER_QUEUE),
-        distance_(distance),
-        queue_(queue),
-        less_(less),
-        class_fnc_(class_fnc),
-        threshold_(std::move(threshold)) {}
-
-  virtual ~PruneQueue() = default;
-
-  StateId Head() const override { return queue_->Head(); }
-
-  void Enqueue(StateId s) override {
-    const auto c = class_fnc_(s);
-    if (c >= class_distance_.size()) {
-      class_distance_.resize(c + 1, Weight::Zero());
-    }
-    if (less_(distance_[s], class_distance_[c])) {
-      class_distance_[c] = distance_[s];
-    }
-    // Enqueues only if below threshold limit.
-    const auto limit = Times(class_distance_[c], threshold_);
-    if (less_(distance_[s], limit)) queue_->Enqueue(s);
-  }
-
-  void Dequeue() override { queue_->Dequeue(); }
-
-  void Update(StateId s) override {
-    const auto c = class_fnc_(s);
-    if (less_(distance_[s], class_distance_[c])) {
-      class_distance_[c] = distance_[s];
-    }
-    queue_->Update(s);
-  }
-
-  bool Empty() const override { return queue_->Empty(); }
-
-  void Clear() override { queue_->Clear(); }
-
- private:
-  const std::vector<Weight> &distance_;  // Shortest distance to state.
-  std::unique_ptr<Queue> queue_;
-  const Less &less_;                    // Borrowed reference.
-  const ClassFnc &class_fnc_;           // Equivalence class functor.
-  Weight threshold_;                    // Pruning weight threshold.
-  std::vector<Weight> class_distance_;  // Shortest distance to class.
-};
-
-// Pruning queue discipline (see above) using the weight's natural order for the
-// comparison function. The ownership of the queue argument is given to this
-// class.
-template <typename Queue, typename Weight, typename ClassFnc>
-class NaturalPruneQueue final
-    : public PruneQueue<Queue, NaturalLess<Weight>, ClassFnc> {
- public:
-  using StateId = typename Queue::StateId;
-
-  NaturalPruneQueue(const std::vector<Weight> &distance, Queue *queue,
-                    const ClassFnc &class_fnc, Weight threshold)
-      : PruneQueue<Queue, NaturalLess<Weight>, ClassFnc>(
-            distance, queue, NaturalLess<Weight>(), class_fnc, threshold) {}
-
-  virtual ~NaturalPruneQueue() = default;
-};
-
-// Filter-based pruning queue discipline: enqueues a state only if allowed by
-// the filter, specified by the state filter functor argument. The underlying
-// queue discipline is specified by the queue argument. The ownership of the
-// queue is given to this class.
-template <typename Queue, typename Filter>
-class FilterQueue final : public QueueBase<typename Queue::StateId> {
- public:
-  using StateId = typename Queue::StateId;
-
-  FilterQueue(Queue *queue, const Filter &filter)
-      : QueueBase<StateId>(OTHER_QUEUE), queue_(queue), filter_(filter) {}
-
-  virtual ~FilterQueue() = default;
-
-  StateId Head() const override { return queue_->Head(); }
-
-  // Enqueues only if allowed by state filter.
-  void Enqueue(StateId s) override {
-    if (filter_(s)) queue_->Enqueue(s);
-  }
-
-  void Dequeue() override { queue_->Dequeue(); }
-
-  void Update(StateId s) override {}
-
-  bool Empty() const override { return queue_->Empty(); }
-
-  void Clear() override { queue_->Clear(); }
-
- private:
-  std::unique_ptr<Queue> queue_;
-  const Filter &filter_;
-};
-
-}  // namespace fst
-
-#endif  // FST_QUEUE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/randequivalent.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/randequivalent.h
deleted file mode 100644
index 73108b46..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/randequivalent.h
+++ /dev/null
@@ -1,114 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Tests if two FSTS are equivalent by checking if random strings from one FST
-// are transduced the same by both FSTs.
-
-#ifndef FST_RANDEQUIVALENT_H_
-#define FST_RANDEQUIVALENT_H_
-
-#include <fst/log.h>
-
-#include <fst/arcsort.h>
-#include <fst/compose.h>
-#include <fst/project.h>
-#include <fst/randgen.h>
-#include <fst/shortest-distance.h>
-#include <fst/vector-fst.h>
-
-
-namespace fst {
-
-// Test if two FSTs are stochastically equivalent by randomly generating
-// random paths through the FSTs.
-//
-// For each randomly generated path, the algorithm computes for each
-// of the two FSTs the sum of the weights of all the successful paths
-// sharing the same input and output labels as the considered randomly
-// generated path and checks that these two values are within a user-specified
-// delta. Returns optional error value (when FLAGS_error_fatal = false).
-template <class Arc, class ArcSelector>
-bool RandEquivalent(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-                    int32 num_paths, float delta,
-                    const RandGenOptions<ArcSelector> &opts,
-                    bool *error = nullptr) {
-  using Weight = typename Arc::Weight;
-  if (error) *error = false;
-  // Checks that the symbol table are compatible.
-  if (!CompatSymbols(fst1.InputSymbols(), fst2.InputSymbols()) ||
-      !CompatSymbols(fst1.OutputSymbols(), fst2.OutputSymbols())) {
-    FSTERROR() << "RandEquivalent: Input/output symbol tables of 1st "
-               << "argument do not match input/output symbol tables of 2nd "
-               << "argument";
-    if (error) *error = true;
-    return false;
-  }
-  static const ILabelCompare<Arc> icomp;
-  static const OLabelCompare<Arc> ocomp;
-  VectorFst<Arc> sfst1(fst1);
-  VectorFst<Arc> sfst2(fst2);
-  Connect(&sfst1);
-  Connect(&sfst2);
-  ArcSort(&sfst1, icomp);
-  ArcSort(&sfst2, icomp);
-  bool result = true;
-  for (int32 n = 0; n < num_paths; ++n) {
-    VectorFst<Arc> path;
-    const auto &fst = rand() % 2 ? sfst1 : sfst2;  // NOLINT
-    RandGen(fst, &path, opts);
-    VectorFst<Arc> ipath(path);
-    VectorFst<Arc> opath(path);
-    Project(&ipath, PROJECT_INPUT);
-    Project(&opath, PROJECT_OUTPUT);
-    VectorFst<Arc> cfst1, pfst1;
-    Compose(ipath, sfst1, &cfst1);
-    ArcSort(&cfst1, ocomp);
-    Compose(cfst1, opath, &pfst1);
-    // Gives up if there are epsilon cycles in a non-idempotent semiring.
-    if (!(Weight::Properties() & kIdempotent) &&
-        pfst1.Properties(kCyclic, true)) {
-      continue;
-    }
-    const auto sum1 = ShortestDistance(pfst1);
-    VectorFst<Arc> cfst2;
-    Compose(ipath, sfst2, &cfst2);
-    ArcSort(&cfst2, ocomp);
-    VectorFst<Arc> pfst2;
-    Compose(cfst2, opath, &pfst2);
-    // Gives up if there are epsilon cycles in a non-idempotent semiring.
-    if (!(Weight::Properties() & kIdempotent) &&
-        pfst2.Properties(kCyclic, true)) {
-      continue;
-    }
-    const auto sum2 = ShortestDistance(pfst2);
-    if (!ApproxEqual(sum1, sum2, delta)) {
-      VLOG(1) << "Sum1 = " << sum1;
-      VLOG(1) << "Sum2 = " << sum2;
-      result = false;
-      break;
-    }
-  }
-  if (fst1.Properties(kError, false) || fst2.Properties(kError, false)) {
-    if (error) *error = true;
-    return false;
-  }
-  return result;
-}
-
-// Tests if two FSTs are equivalent by randomly generating a nnum_paths paths
-// (no longer than the path_length) using a user-specified seed, optionally
-// indicating an error setting an optional error argument to true.
-template <class Arc>
-bool RandEquivalent(const Fst<Arc> &fst1, const Fst<Arc> &fst2, int32 num_paths,
-                    float delta = kDelta, time_t seed = time(nullptr),
-                    int32 max_length = std::numeric_limits<int32>::max(),
-                    bool *error = nullptr) {
-  const UniformArcSelector<Arc> uniform_selector(seed);
-  const RandGenOptions<UniformArcSelector<Arc>> opts(uniform_selector,
-                                                     max_length);
-  return RandEquivalent(fst1, fst2, num_paths, delta, opts, error);
-}
-
-}  // namespace fst
-
-#endif  // FST_RANDEQUIVALENT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/randgen.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/randgen.h
deleted file mode 100644
index d312ef17..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/randgen.h
+++ /dev/null
@@ -1,750 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes and functions to generate random paths through an FST.
-
-#ifndef FST_RANDGEN_H_
-#define FST_RANDGEN_H_
-
-#include <math.h>
-#include <stddef.h>
-#include <limits>
-#include <map>
-#include <memory>
-#include <random>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/accumulator.h>
-#include <fst/cache.h>
-#include <fst/dfs-visit.h>
-#include <fst/float-weight.h>
-#include <fst/fst-decl.h>
-#include <fst/fst.h>
-#include <fst/mutable-fst.h>
-#include <fst/properties.h>
-#include <fst/util.h>
-#include <fst/weight.h>
-
-namespace fst {
-
-// The RandGenFst class is roughly similar to ArcMapFst in that it takes two
-// template parameters denoting the input and output arc types. However, it also
-// takes an additional template parameter which specifies a sampler object which
-// samples (with replacement) arcs from an FST state. The sampler in turn takes
-// a template parameter for a selector object which actually chooses the arc.
-//
-// Arc selector functors are used to select a random transition given an FST
-// state s, returning a number N such that 0 <= N <= NumArcs(s). If N is
-// NumArcs(s), then the final weight is selected; otherwise the N-th arc is
-// selected. It is assumed these are not applied to any state which is neither
-// final nor has any arcs leaving it.
-
-// Randomly selects a transition using the uniform distribution. This class is
-// not thread-safe.
-template <class Arc>
-class UniformArcSelector {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // Constructs a selector with a non-deterministic seed.
-  UniformArcSelector() : rand_(std::random_device()()) {}
-  // Constructs a selector with a given seed.
-  explicit UniformArcSelector(uint64 seed) : rand_(seed) {}
-
-  size_t operator()(const Fst<Arc> &fst, StateId s) const {
-    const auto n = fst.NumArcs(s) + (fst.Final(s) != Weight::Zero());
-    return static_cast<size_t>(
-        std::uniform_int_distribution<>(0, n - 1)(rand_));
-  }
-
- private:
-  mutable std::mt19937_64 rand_;
-};
-
-// Randomly selects a transition w.r.t. the weights treated as negative log
-// probabilities after normalizing for the total weight leaving the state. Zero
-// transitions are disregarded. It assumed that Arc::Weight::Value() accesses
-// the floating point representation of the weight. This class is not
-// thread-safe.
-template <class Arc>
-class LogProbArcSelector {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // Constructs a selector with a non-deterministic seed.
-  LogProbArcSelector() : rand_(std::random_device()()) {}
-  // Constructs a selector with a given seed.
-  explicit LogProbArcSelector(uint64 seed) : rand_(seed) {}
-
-  size_t operator()(const Fst<Arc> &fst, StateId s) const {
-    // Finds total weight leaving state.
-    auto sum = Log64Weight::Zero();
-    ArcIterator<Fst<Arc>> aiter(fst, s);
-    for (; !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      sum = Plus(sum, to_log_weight_(arc.weight));
-    }
-    sum = Plus(sum, to_log_weight_(fst.Final(s)));
-    const double threshold =
-        std::uniform_real_distribution<>(0, exp(-sum.Value()))(rand_);
-    auto p = Log64Weight::Zero();
-    size_t n = 0;
-    for (aiter.Reset(); !aiter.Done(); aiter.Next(), ++n) {
-      p = Plus(p, to_log_weight_(aiter.Value().weight));
-      if (exp(-p.Value()) > threshold) return n;
-    }
-    return n;
-  }
-
- private:
-  mutable std::mt19937_64 rand_;
-  WeightConvert<Weight, Log64Weight> to_log_weight_;
-};
-
-// Useful alias when using StdArc.
-using StdArcSelector = LogProbArcSelector<StdArc>;
-
-// Same as LogProbArcSelector but use CacheLogAccumulator to cache the weight
-// accumulation computations. This class is not thread-safe.
-template <class Arc>
-class FastLogProbArcSelector : public LogProbArcSelector<Arc> {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using LogProbArcSelector<Arc>::operator();
-
-  // Constructs a selector with a non-deterministic seed.
-  FastLogProbArcSelector() : seed_(std::random_device()()), rand_(seed_) {}
-  // Constructs a selector with a given seed.
-  explicit FastLogProbArcSelector(uint64 seed) : seed_(seed), rand_(seed_) {}
-
-  size_t operator()(const Fst<Arc> &fst, StateId s,
-                    CacheLogAccumulator<Arc> *accumulator) const {
-    accumulator->SetState(s);
-    ArcIterator<Fst<Arc>> aiter(fst, s);
-    // Finds total weight leaving state.
-    const double sum = to_log_weight_(accumulator->Sum(fst.Final(s), &aiter, 0,
-                                                       fst.NumArcs(s)))
-                           .Value();
-    const double r = -log(std::uniform_real_distribution<>(0, 1)(rand_));
-    Weight w = from_log_weight_(r + sum);
-    aiter.Reset();
-    return accumulator->LowerBound(w, &aiter);
-  }
-
-  uint64 Seed() const { return seed_; }
-
- private:
-  const uint64 seed_;
-  mutable std::mt19937_64 rand_;
-  WeightConvert<Weight, Log64Weight> to_log_weight_;
-  WeightConvert<Log64Weight, Weight> from_log_weight_;
-};
-
-// Random path state info maintained by RandGenFst and passed to samplers.
-template <typename Arc>
-struct RandState {
-  using StateId = typename Arc::StateId;
-
-  StateId state_id;  // Current input FST state.
-  size_t nsamples;   // Number of samples to be sampled at this state.
-  size_t length;     // Length of path to this random state.
-  size_t select;     // Previous sample arc selection.
-  const RandState<Arc> *parent;  // Previous random state on this path.
-
-  explicit RandState(StateId state_id, size_t nsamples = 0, size_t length = 0,
-                     size_t select = 0, const RandState<Arc> *parent = nullptr)
-      : state_id(state_id),
-        nsamples(nsamples),
-        length(length),
-        select(select),
-        parent(parent) {}
-
-  RandState() : RandState(kNoStateId) {}
-};
-
-// This class, given an arc selector, samples, with replacement, multiple random
-// transitions from an FST's state. This is a generic version with a
-// straightforward use of the arc selector. Specializations may be defined for
-// arc selectors for greater efficiency or special behavior.
-template <class Arc, class Selector>
-class ArcSampler {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // The max_length argument may be interpreted (or ignored) by a selector as
-  // it chooses. This generic version interprets this literally.
-  ArcSampler(const Fst<Arc> &fst, const Selector &selector,
-             int32 max_length = std::numeric_limits<int32>::max())
-      : fst_(fst), selector_(selector), max_length_(max_length) {}
-
-  // Allow updating FST argument; pass only if changed.
-  ArcSampler(const ArcSampler<Arc, Selector> &sampler,
-             const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? *fst : sampler.fst_),
-        selector_(sampler.selector_),
-        max_length_(sampler.max_length_) {
-    Reset();
-  }
-
-  // Samples a fixed number of samples from the given state. The length argument
-  // specifies the length of the path to the state. Returns true if the samples
-  // were collected. No samples may be collected if either there are no
-  // transitions leaving the state and the state is non-final, or if the path
-  // length has been exceeded. Iterator members are provided to read the samples
-  // in the order in which they were collected.
-  bool Sample(const RandState<Arc> &rstate) {
-    sample_map_.clear();
-    if ((fst_.NumArcs(rstate.state_id) == 0 &&
-         fst_.Final(rstate.state_id) == Weight::Zero()) ||
-        rstate.length == max_length_) {
-      Reset();
-      return false;
-    }
-    for (size_t i = 0; i < rstate.nsamples; ++i) {
-      ++sample_map_[selector_(fst_, rstate.state_id)];
-    }
-    Reset();
-    return true;
-  }
-
-  // More samples?
-  bool Done() const { return sample_iter_ == sample_map_.end(); }
-
-  // Gets the next sample.
-  void Next() { ++sample_iter_; }
-
-  std::pair<size_t, size_t> Value() const { return *sample_iter_; }
-
-  void Reset() { sample_iter_ = sample_map_.begin(); }
-
-  bool Error() const { return false; }
-
- private:
-  const Fst<Arc> &fst_;
-  const Selector &selector_;
-  const int32 max_length_;
-
-  // Stores (N, K) as described for Value().
-  std::map<size_t, size_t> sample_map_;
-  std::map<size_t, size_t>::const_iterator sample_iter_;
-
-  ArcSampler<Arc, Selector> &operator=(const ArcSampler &) = delete;
-};
-
-// Samples one sample of num_to_sample dimensions from a multinomial
-// distribution parameterized by a vector of probabilities. The result
-// container should be pre-initialized (e.g., an empty map or a zeroed vector
-// sized the same as the vector of probabilities.
-// probs.size()).
-template <class Result, class RNG>
-void OneMultinomialSample(const std::vector<double> &probs,
-                          size_t num_to_sample, Result *result, RNG *rng) {
-  // Left-over probability mass.
-  double norm = 0;
-  for (double p : probs) norm += p;
-  // Left-over number of samples needed.
-  for (size_t i = 0; i < probs.size(); ++i) {
-    size_t num_sampled = 0;
-    if (probs[i] > 0) {
-      std::binomial_distribution<> d(num_to_sample, probs[i] / norm);
-      num_sampled = d(*rng);
-    }
-    if (num_sampled != 0) (*result)[i] = num_sampled;
-    norm -= probs[i];
-    num_to_sample -= num_sampled;
-  }
-}
-
-// Specialization for FastLogProbArcSelector.
-template <class Arc>
-class ArcSampler<Arc, FastLogProbArcSelector<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Accumulator = CacheLogAccumulator<Arc>;
-  using Selector = FastLogProbArcSelector<Arc>;
-
-  ArcSampler(const Fst<Arc> &fst, const Selector &selector,
-             int32 max_length = std::numeric_limits<int32>::max())
-      : fst_(fst),
-        selector_(selector),
-        max_length_(max_length),
-        accumulator_(new Accumulator()) {
-    accumulator_->Init(fst);
-    rng_.seed(selector_.Seed());
-  }
-
-  ArcSampler(const ArcSampler<Arc, Selector> &sampler,
-             const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? *fst : sampler.fst_),
-        selector_(sampler.selector_),
-        max_length_(sampler.max_length_) {
-    if (fst) {
-      accumulator_.reset(new Accumulator());
-      accumulator_->Init(*fst);
-    } else {  // Shallow copy.
-      accumulator_.reset(new Accumulator(*sampler.accumulator_));
-    }
-  }
-
-  bool Sample(const RandState<Arc> &rstate) {
-    sample_map_.clear();
-    if ((fst_.NumArcs(rstate.state_id) == 0 &&
-         fst_.Final(rstate.state_id) == Weight::Zero()) ||
-        rstate.length == max_length_) {
-      Reset();
-      return false;
-    }
-    if (fst_.NumArcs(rstate.state_id) + 1 < rstate.nsamples) {
-      MultinomialSample(rstate);
-      Reset();
-      return true;
-    }
-    for (size_t i = 0; i < rstate.nsamples; ++i) {
-      ++sample_map_[selector_(fst_, rstate.state_id, accumulator_.get())];
-    }
-    Reset();
-    return true;
-  }
-
-  bool Done() const { return sample_iter_ == sample_map_.end(); }
-
-  void Next() { ++sample_iter_; }
-
-  std::pair<size_t, size_t> Value() const { return *sample_iter_; }
-
-  void Reset() { sample_iter_ = sample_map_.begin(); }
-
-  bool Error() const { return accumulator_->Error(); }
-
- private:
-  using RNG = std::mt19937;
-
-  // Sample according to the multinomial distribution of rstate.nsamples draws
-  // from p_.
-  void MultinomialSample(const RandState<Arc> &rstate) {
-    p_.clear();
-    for (ArcIterator<Fst<Arc>> aiter(fst_, rstate.state_id); !aiter.Done();
-         aiter.Next()) {
-      p_.push_back(exp(-to_log_weight_(aiter.Value().weight).Value()));
-    }
-    if (fst_.Final(rstate.state_id) != Weight::Zero()) {
-      p_.push_back(exp(-to_log_weight_(fst_.Final(rstate.state_id)).Value()));
-    }
-    if (rstate.nsamples < std::numeric_limits<RNG::result_type>::max()) {
-      OneMultinomialSample(p_, rstate.nsamples, &sample_map_, &rng_);
-    } else {
-      for (size_t i = 0; i < p_.size(); ++i) {
-        sample_map_[i] = ceil(p_[i] * rstate.nsamples);
-      }
-    }
-  }
-
-  const Fst<Arc> &fst_;
-  const Selector &selector_;
-  const int32 max_length_;
-
-  // Stores (N, K) for Value().
-  std::map<size_t, size_t> sample_map_;
-  std::map<size_t, size_t>::const_iterator sample_iter_;
-
-  std::unique_ptr<Accumulator> accumulator_;
-  RNG rng_;                // Random number generator.
-  std::vector<double> p_;  // Multinomial parameters.
-  WeightConvert<Weight, Log64Weight> to_log_weight_;
-};
-
-// Options for random path generation with RandGenFst. The template argument is
-// a sampler, typically the class ArcSampler. Ownership of the sampler is taken
-// by RandGenFst.
-template <class Sampler>
-struct RandGenFstOptions : public CacheOptions {
-  Sampler *sampler;          // How to sample transitions at a state.
-  int32 npath;               // Number of paths to generate.
-  bool weighted;             // Is the output tree weighted by path count, or
-                             // is it just an unweighted DAG?
-  bool remove_total_weight;  // Remove total weight when output is weighted.
-
-  RandGenFstOptions(const CacheOptions &opts, Sampler *sampler, int32 npath = 1,
-                    bool weighted = true, bool remove_total_weight = false)
-      : CacheOptions(opts),
-        sampler(sampler),
-        npath(npath),
-        weighted(weighted),
-        remove_total_weight(remove_total_weight) {}
-};
-
-namespace internal {
-
-// Implementation of RandGenFst.
-template <class FromArc, class ToArc, class Sampler>
-class RandGenFstImpl : public CacheImpl<ToArc> {
- public:
-  using FstImpl<ToArc>::SetType;
-  using FstImpl<ToArc>::SetProperties;
-  using FstImpl<ToArc>::SetInputSymbols;
-  using FstImpl<ToArc>::SetOutputSymbols;
-
-  using CacheBaseImpl<CacheState<ToArc>>::PushArc;
-  using CacheBaseImpl<CacheState<ToArc>>::HasArcs;
-  using CacheBaseImpl<CacheState<ToArc>>::HasFinal;
-  using CacheBaseImpl<CacheState<ToArc>>::HasStart;
-  using CacheBaseImpl<CacheState<ToArc>>::SetArcs;
-  using CacheBaseImpl<CacheState<ToArc>>::SetFinal;
-  using CacheBaseImpl<CacheState<ToArc>>::SetStart;
-
-  using Label = typename FromArc::Label;
-  using StateId = typename FromArc::StateId;
-  using FromWeight = typename FromArc::Weight;
-
-  using ToWeight = typename ToArc::Weight;
-
-  RandGenFstImpl(const Fst<FromArc> &fst,
-                 const RandGenFstOptions<Sampler> &opts)
-      : CacheImpl<ToArc>(opts),
-        fst_(fst.Copy()),
-        sampler_(opts.sampler),
-        npath_(opts.npath),
-        weighted_(opts.weighted),
-        remove_total_weight_(opts.remove_total_weight),
-        superfinal_(kNoLabel) {
-    SetType("randgen");
-    SetProperties(
-        RandGenProperties(fst.Properties(kFstProperties, false), weighted_),
-        kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  RandGenFstImpl(const RandGenFstImpl &impl)
-      : CacheImpl<ToArc>(impl),
-        fst_(impl.fst_->Copy(true)),
-        sampler_(new Sampler(*impl.sampler_, fst_.get())),
-        npath_(impl.npath_),
-        weighted_(impl.weighted_),
-        superfinal_(kNoLabel) {
-    SetType("randgen");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      const auto s = fst_->Start();
-      if (s == kNoStateId) return kNoStateId;
-      SetStart(state_table_.size());
-      state_table_.emplace_back(
-          new RandState<FromArc>(s, npath_, 0, 0, nullptr));
-    }
-    return CacheImpl<ToArc>::Start();
-  }
-
-  ToWeight Final(StateId s) {
-    if (!HasFinal(s)) Expand(s);
-    return CacheImpl<ToArc>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<ToArc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<ToArc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<ToArc>::NumOutputEpsilons(s);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) &&
-        (fst_->Properties(kError, false) || sampler_->Error())) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<ToArc>::Properties(mask);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<ToArc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<ToArc>::InitArcIterator(s, data);
-  }
-
-  // Computes the outgoing transitions from a state, creating new destination
-  // states as needed.
-  void Expand(StateId s) {
-    if (s == superfinal_) {
-      SetFinal(s, ToWeight::One());
-      SetArcs(s);
-      return;
-    }
-    SetFinal(s, ToWeight::Zero());
-    const auto &rstate = *state_table_[s];
-    sampler_->Sample(rstate);
-    ArcIterator<Fst<FromArc>> aiter(*fst_, rstate.state_id);
-    const auto narcs = fst_->NumArcs(rstate.state_id);
-    for (; !sampler_->Done(); sampler_->Next()) {
-      const auto &sample_pair = sampler_->Value();
-      const auto pos = sample_pair.first;
-      const auto count = sample_pair.second;
-      double prob = static_cast<double>(count) / rstate.nsamples;
-      if (pos < narcs) {  // Regular transition.
-        aiter.Seek(sample_pair.first);
-        const auto &aarc = aiter.Value();
-        const auto weight =
-            weighted_ ? to_weight_(Log64Weight(-log(prob))) : ToWeight::One();
-        const ToArc barc(aarc.ilabel, aarc.olabel, weight, state_table_.size());
-        PushArc(s, barc);
-        auto *nrstate = new RandState<FromArc>(aarc.nextstate, count,
-                                               rstate.length + 1, pos, &rstate);
-        state_table_.emplace_back(nrstate);
-      } else {  // Super-final transition.
-        if (weighted_) {
-          const auto weight =
-              remove_total_weight_
-                  ? to_weight_(Log64Weight(-log(prob)))
-                  : to_weight_(Log64Weight(-log(prob * npath_)));
-          SetFinal(s, weight);
-        } else {
-          if (superfinal_ == kNoLabel) {
-            superfinal_ = state_table_.size();
-            state_table_.emplace_back(
-                new RandState<FromArc>(kNoStateId, 0, 0, 0, nullptr));
-          }
-          for (size_t n = 0; n < count; ++n) {
-            const ToArc barc(0, 0, ToWeight::One(), superfinal_);
-            PushArc(s, barc);
-          }
-        }
-      }
-    }
-    SetArcs(s);
-  }
-
- private:
-  const std::unique_ptr<Fst<FromArc>> fst_;
-  std::unique_ptr<Sampler> sampler_;
-  const int32 npath_;
-  std::vector<std::unique_ptr<RandState<FromArc>>> state_table_;
-  const bool weighted_;
-  bool remove_total_weight_;
-  StateId superfinal_;
-  WeightConvert<Log64Weight, ToWeight> to_weight_;
-};
-
-}  // namespace internal
-
-// FST class to randomly generate paths through an FST, with details controlled
-// by RandGenOptionsFst. Output format is a tree weighted by the path count.
-template <class FromArc, class ToArc, class Sampler>
-class RandGenFst
-    : public ImplToFst<internal::RandGenFstImpl<FromArc, ToArc, Sampler>> {
- public:
-  using Label = typename FromArc::Label;
-  using StateId = typename FromArc::StateId;
-  using Weight = typename FromArc::Weight;
-
-  using Store = DefaultCacheStore<FromArc>;
-  using State = typename Store::State;
-
-  using Impl = internal::RandGenFstImpl<FromArc, ToArc, Sampler>;
-
-  friend class ArcIterator<RandGenFst<FromArc, ToArc, Sampler>>;
-  friend class StateIterator<RandGenFst<FromArc, ToArc, Sampler>>;
-
-  RandGenFst(const Fst<FromArc> &fst, const RandGenFstOptions<Sampler> &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  RandGenFst(const RandGenFst<FromArc, ToArc, Sampler> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this RandGenFst. See Fst<>::Copy() for further doc.
-  RandGenFst<FromArc, ToArc, Sampler> *Copy(bool safe = false) const override {
-    return new RandGenFst<FromArc, ToArc, Sampler>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<ToArc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<ToArc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  RandGenFst &operator=(const RandGenFst &) = delete;
-};
-
-// Specialization for RandGenFst.
-template <class FromArc, class ToArc, class Sampler>
-class StateIterator<RandGenFst<FromArc, ToArc, Sampler>>
-    : public CacheStateIterator<RandGenFst<FromArc, ToArc, Sampler>> {
- public:
-  explicit StateIterator(const RandGenFst<FromArc, ToArc, Sampler> &fst)
-      : CacheStateIterator<RandGenFst<FromArc, ToArc, Sampler>>(
-            fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for RandGenFst.
-template <class FromArc, class ToArc, class Sampler>
-class ArcIterator<RandGenFst<FromArc, ToArc, Sampler>>
-    : public CacheArcIterator<RandGenFst<FromArc, ToArc, Sampler>> {
- public:
-  using StateId = typename FromArc::StateId;
-
-  ArcIterator(const RandGenFst<FromArc, ToArc, Sampler> &fst, StateId s)
-      : CacheArcIterator<RandGenFst<FromArc, ToArc, Sampler>>(
-            fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class FromArc, class ToArc, class Sampler>
-inline void RandGenFst<FromArc, ToArc, Sampler>::InitStateIterator(
-    StateIteratorData<ToArc> *data) const {
-  data->base = new StateIterator<RandGenFst<FromArc, ToArc, Sampler>>(*this);
-}
-
-// Options for random path generation.
-template <class Selector>
-struct RandGenOptions {
-  const Selector &selector;  // How an arc is selected at a state.
-  int32 max_length;          // Maximum path length.
-  int32 npath;               // Number of paths to generate.
-  bool weighted;             // Is the output tree weighted by path count, or
-                             // is it just an unweighted DAG?
-  bool remove_total_weight;  // Remove total weight when output is weighted?
-
-  explicit RandGenOptions(const Selector &selector,
-                          int32 max_length = std::numeric_limits<int32>::max(),
-                          int32 npath = 1, bool weighted = false,
-                          bool remove_total_weight = false)
-      : selector(selector),
-        max_length(max_length),
-        npath(npath),
-        weighted(weighted),
-        remove_total_weight(remove_total_weight) {}
-};
-
-namespace internal {
-
-template <class FromArc, class ToArc>
-class RandGenVisitor {
- public:
-  using StateId = typename FromArc::StateId;
-  using Weight = typename FromArc::Weight;
-
-  explicit RandGenVisitor(MutableFst<ToArc> *ofst) : ofst_(ofst) {}
-
-  void InitVisit(const Fst<FromArc> &ifst) {
-    ifst_ = &ifst;
-    ofst_->DeleteStates();
-    ofst_->SetInputSymbols(ifst.InputSymbols());
-    ofst_->SetOutputSymbols(ifst.OutputSymbols());
-    if (ifst.Properties(kError, false)) ofst_->SetProperties(kError, kError);
-    path_.clear();
-  }
-
-  constexpr bool InitState(StateId, StateId) const { return true; }
-
-  bool TreeArc(StateId, const ToArc &arc) {
-    if (ifst_->Final(arc.nextstate) == Weight::Zero()) {
-      path_.push_back(arc);
-    } else {
-      OutputPath();
-    }
-    return true;
-  }
-
-  bool BackArc(StateId, const FromArc &) {
-    FSTERROR() << "RandGenVisitor: cyclic input";
-    ofst_->SetProperties(kError, kError);
-    return false;
-  }
-
-  bool ForwardOrCrossArc(StateId, const FromArc &) {
-    OutputPath();
-    return true;
-  }
-
-  void FinishState(StateId s, StateId p, const FromArc *) {
-    if (p != kNoStateId && ifst_->Final(s) == Weight::Zero()) path_.pop_back();
-  }
-
-  void FinishVisit() {}
-
- private:
-  void OutputPath() {
-    if (ofst_->Start() == kNoStateId) {
-      const auto start = ofst_->AddState();
-      ofst_->SetStart(start);
-    }
-    auto src = ofst_->Start();
-    for (size_t i = 0; i < path_.size(); ++i) {
-      const auto dest = ofst_->AddState();
-      const ToArc arc(path_[i].ilabel, path_[i].olabel, Weight::One(), dest);
-      ofst_->AddArc(src, arc);
-      src = dest;
-    }
-    ofst_->SetFinal(src, Weight::One());
-  }
-
-  const Fst<FromArc> *ifst_;
-  MutableFst<ToArc> *ofst_;
-  std::vector<ToArc> path_;
-
-  RandGenVisitor(const RandGenVisitor &) = delete;
-  RandGenVisitor &operator=(const RandGenVisitor &) = delete;
-};
-
-}  // namespace internal
-
-// Randomly generate paths through an FST; details controlled by
-// RandGenOptions.
-template <class FromArc, class ToArc, class Selector>
-void RandGen(const Fst<FromArc> &ifst, MutableFst<ToArc> *ofst,
-             const RandGenOptions<Selector> &opts) {
-  using State = typename ToArc::StateId;
-  using Weight = typename ToArc::Weight;
-  using Sampler = ArcSampler<FromArc, Selector>;
-  auto *sampler = new Sampler(ifst, opts.selector, opts.max_length);
-  RandGenFstOptions<Sampler> fopts(CacheOptions(true, 0), sampler, opts.npath,
-                                   opts.weighted, opts.remove_total_weight);
-  RandGenFst<FromArc, ToArc, Sampler> rfst(ifst, fopts);
-  if (opts.weighted) {
-    *ofst = rfst;
-  } else {
-    internal::RandGenVisitor<FromArc, ToArc> rand_visitor(ofst);
-    DfsVisit(rfst, &rand_visitor);
-  }
-}
-
-// Randomly generate a path through an FST with the uniform distribution
-// over the transitions.
-template <class FromArc, class ToArc>
-void RandGen(const Fst<FromArc> &ifst, MutableFst<ToArc> *ofst) {
-  const UniformArcSelector<FromArc> uniform_selector;
-  RandGenOptions<UniformArcSelector<ToArc>> opts(uniform_selector);
-  RandGen(ifst, ofst, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_RANDGEN_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/rational.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/rational.h
deleted file mode 100644
index 5a9e6024..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/rational.h
+++ /dev/null
@@ -1,307 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// An FST implementation and base interface for delayed unions, concatenations,
-// and closures.
-
-#ifndef FST_RATIONAL_H_
-#define FST_RATIONAL_H_
-
-#include <algorithm>
-#include <string>
-#include <vector>
-
-#include <fst/mutable-fst.h>
-#include <fst/replace.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-using RationalFstOptions = CacheOptions;
-
-// This specifies whether to add the empty string.
-enum ClosureType {
-  CLOSURE_STAR = 0,  // Add the empty string.
-  CLOSURE_PLUS = 1   // Don't add the empty string.
-};
-
-template <class Arc>
-class RationalFst;
-
-template <class Arc>
-void Union(RationalFst<Arc> *fst1, const Fst<Arc> &fst2);
-
-template <class Arc>
-void Concat(RationalFst<Arc> *fst1, const Fst<Arc> &fst2);
-
-template <class Arc>
-void Concat(const Fst<Arc> &fst1, RationalFst<Arc> *fst2);
-
-template <class Arc>
-void Closure(RationalFst<Arc> *fst, ClosureType closure_type);
-
-namespace internal {
-
-// Implementation class for delayed unions, concatenations and closures.
-template <class A>
-class RationalFstImpl : public FstImpl<A> {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::WriteHeader;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  explicit RationalFstImpl(const RationalFstOptions &opts)
-      : nonterminals_(0), replace_options_(opts, 0) {
-    SetType("rational");
-    fst_tuples_.push_back(std::make_pair(0, nullptr));
-  }
-
-  RationalFstImpl(const RationalFstImpl<Arc> &impl)
-      : rfst_(impl.rfst_),
-        nonterminals_(impl.nonterminals_),
-        replace_(impl.replace_ ? impl.replace_->Copy(true) : nullptr),
-        replace_options_(impl.replace_options_) {
-    SetType("rational");
-    fst_tuples_.reserve(impl.fst_tuples_.size());
-    for (const auto &pair : impl.fst_tuples_) {
-      fst_tuples_.emplace_back(pair.first,
-                               pair.second ? pair.second->Copy(true) : nullptr);
-    }
-  }
-
-  ~RationalFstImpl() override {
-    for (auto &tuple : fst_tuples_) delete tuple.second;
-  }
-
-  StateId Start() { return Replace()->Start(); }
-
-  Weight Final(StateId s) { return Replace()->Final(s); }
-
-  size_t NumArcs(StateId s) { return Replace()->NumArcs(s); }
-
-  size_t NumInputEpsilons(StateId s) { return Replace()->NumInputEpsilons(s); }
-
-  size_t NumOutputEpsilons(StateId s) {
-    return Replace()->NumOutputEpsilons(s);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && Replace()->Properties(kError, false)) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  // Implementation of UnionFst(fst1, fst2).
-  void InitUnion(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {
-    replace_.reset();
-    const auto props1 = fst1.Properties(kFstProperties, false);
-    const auto props2 = fst2.Properties(kFstProperties, false);
-    SetInputSymbols(fst1.InputSymbols());
-    SetOutputSymbols(fst1.OutputSymbols());
-    rfst_.AddState();
-    rfst_.AddState();
-    rfst_.SetStart(0);
-    rfst_.SetFinal(1, Weight::One());
-    rfst_.SetInputSymbols(fst1.InputSymbols());
-    rfst_.SetOutputSymbols(fst1.OutputSymbols());
-    nonterminals_ = 2;
-    rfst_.AddArc(0, Arc(0, -1, Weight::One(), 1));
-    rfst_.AddArc(0, Arc(0, -2, Weight::One(), 1));
-    fst_tuples_.push_back(std::make_pair(-1, fst1.Copy()));
-    fst_tuples_.push_back(std::make_pair(-2, fst2.Copy()));
-    SetProperties(UnionProperties(props1, props2, true), kCopyProperties);
-  }
-
-  // Implementation of ConcatFst(fst1, fst2).
-  void InitConcat(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {
-    replace_.reset();
-    const auto props1 = fst1.Properties(kFstProperties, false);
-    const auto props2 = fst2.Properties(kFstProperties, false);
-    SetInputSymbols(fst1.InputSymbols());
-    SetOutputSymbols(fst1.OutputSymbols());
-    rfst_.AddState();
-    rfst_.AddState();
-    rfst_.AddState();
-    rfst_.SetStart(0);
-    rfst_.SetFinal(2, Weight::One());
-    rfst_.SetInputSymbols(fst1.InputSymbols());
-    rfst_.SetOutputSymbols(fst1.OutputSymbols());
-    nonterminals_ = 2;
-    rfst_.AddArc(0, Arc(0, -1, Weight::One(), 1));
-    rfst_.AddArc(1, Arc(0, -2, Weight::One(), 2));
-    fst_tuples_.push_back(std::make_pair(-1, fst1.Copy()));
-    fst_tuples_.push_back(std::make_pair(-2, fst2.Copy()));
-    SetProperties(ConcatProperties(props1, props2, true), kCopyProperties);
-  }
-
-  // Implementation of ClosureFst(fst, closure_type).
-  void InitClosure(const Fst<Arc> &fst, ClosureType closure_type) {
-    replace_.reset();
-    const auto props = fst.Properties(kFstProperties, false);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-    if (closure_type == CLOSURE_STAR) {
-      rfst_.AddState();
-      rfst_.SetStart(0);
-      rfst_.SetFinal(0, Weight::One());
-      rfst_.AddArc(0, Arc(0, -1, Weight::One(), 0));
-    } else {
-      rfst_.AddState();
-      rfst_.AddState();
-      rfst_.SetStart(0);
-      rfst_.SetFinal(1, Weight::One());
-      rfst_.AddArc(0, Arc(0, -1, Weight::One(), 1));
-      rfst_.AddArc(1, Arc(0, 0, Weight::One(), 0));
-    }
-    rfst_.SetInputSymbols(fst.InputSymbols());
-    rfst_.SetOutputSymbols(fst.OutputSymbols());
-    fst_tuples_.push_back(std::make_pair(-1, fst.Copy()));
-    nonterminals_ = 1;
-    SetProperties(ClosureProperties(props, closure_type == CLOSURE_STAR, true),
-                  kCopyProperties);
-  }
-
-  // Implementation of Union(Fst &, RationalFst *).
-  void AddUnion(const Fst<Arc> &fst) {
-    replace_.reset();
-    const auto props1 = FstImpl<A>::Properties();
-    const auto props2 = fst.Properties(kFstProperties, false);
-    VectorFst<Arc> afst;
-    afst.AddState();
-    afst.AddState();
-    afst.SetStart(0);
-    afst.SetFinal(1, Weight::One());
-    ++nonterminals_;
-    afst.AddArc(0, Arc(0, -nonterminals_, Weight::One(), 1));
-    Union(&rfst_, afst);
-    fst_tuples_.push_back(std::make_pair(-nonterminals_, fst.Copy()));
-    SetProperties(UnionProperties(props1, props2, true), kCopyProperties);
-  }
-
-  // Implementation of Concat(Fst &, RationalFst *).
-  void AddConcat(const Fst<Arc> &fst, bool append) {
-    replace_.reset();
-    const auto props1 = FstImpl<A>::Properties();
-    const auto props2 = fst.Properties(kFstProperties, false);
-    VectorFst<Arc> afst;
-    afst.AddState();
-    afst.AddState();
-    afst.SetStart(0);
-    afst.SetFinal(1, Weight::One());
-    ++nonterminals_;
-    afst.AddArc(0, Arc(0, -nonterminals_, Weight::One(), 1));
-    if (append) {
-      Concat(&rfst_, afst);
-    } else {
-      Concat(afst, &rfst_);
-    }
-    fst_tuples_.push_back(std::make_pair(-nonterminals_, fst.Copy()));
-    SetProperties(ConcatProperties(props1, props2, true), kCopyProperties);
-  }
-
-  // Implementation of Closure(RationalFst *, closure_type).
-  void AddClosure(ClosureType closure_type) {
-    replace_.reset();
-    const auto props = FstImpl<A>::Properties();
-    Closure(&rfst_, closure_type);
-    SetProperties(ClosureProperties(props, closure_type == CLOSURE_STAR, true),
-                  kCopyProperties);
-  }
-
-  // Returns the underlying ReplaceFst, preserving ownership of the underlying
-  // object.
-  ReplaceFst<Arc> *Replace() const {
-    if (!replace_) {
-      fst_tuples_[0].second = rfst_.Copy();
-      replace_.reset(new ReplaceFst<Arc>(fst_tuples_, replace_options_));
-    }
-    return replace_.get();
-  }
-
- private:
-  // Rational topology machine, using negative non-terminals.
-  VectorFst<Arc> rfst_;
-  // Number of nonterminals used.
-  Label nonterminals_;
-  // Contains the nonterminals and their corresponding FSTs.
-  mutable std::vector<std::pair<Label, const Fst<Arc> *>> fst_tuples_;
-  // Underlying ReplaceFst.
-  mutable std::unique_ptr<ReplaceFst<Arc>> replace_;
-  const ReplaceFstOptions<Arc> replace_options_;
-};
-
-}  // namespace internal
-
-// Parent class for the delayed rational operations (union, concatenation, and
-// closure). This class attaches interface to implementation and handles
-// reference counting, delegating most methods to ImplToFst.
-template <class A>
-class RationalFst : public ImplToFst<internal::RationalFstImpl<A>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  using Impl = internal::RationalFstImpl<Arc>;
-
-  friend class StateIterator<RationalFst<Arc>>;
-  friend class ArcIterator<RationalFst<Arc>>;
-  friend void Union<>(RationalFst<Arc> *fst1, const Fst<Arc> &fst2);
-  friend void Concat<>(RationalFst<Arc> *fst1, const Fst<Arc> &fst2);
-  friend void Concat<>(const Fst<Arc> &fst1, RationalFst<Arc> *fst2);
-  friend void Closure<>(RationalFst<Arc> *fst, ClosureType closure_type);
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const override {
-    GetImpl()->Replace()->InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetImpl()->Replace()->InitArcIterator(s, data);
-  }
-
- protected:
-  using ImplToFst<Impl>::GetImpl;
-
-  explicit RationalFst(const RationalFstOptions &opts = RationalFstOptions())
-      : ImplToFst<Impl>(std::make_shared<Impl>(opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  RationalFst(const RationalFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
- private:
-  RationalFst &operator=(const RationalFst &) = delete;
-};
-
-// Specialization for RationalFst.
-template <class Arc>
-class StateIterator<RationalFst<Arc>> : public StateIterator<ReplaceFst<Arc>> {
- public:
-  explicit StateIterator(const RationalFst<Arc> &fst)
-      : StateIterator<ReplaceFst<Arc>>(*(fst.GetImpl()->Replace())) {}
-};
-
-// Specialization for RationalFst.
-template <class Arc>
-class ArcIterator<RationalFst<Arc>> : public CacheArcIterator<ReplaceFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const RationalFst<Arc> &fst, StateId s)
-      : ArcIterator<ReplaceFst<Arc>>(*(fst.GetImpl()->Replace()), s) {}
-};
-
-}  // namespace fst
-
-#endif  // FST_RATIONAL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/register.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/register.h
deleted file mode 100644
index 2d1a6ea7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/register.h
+++ /dev/null
@@ -1,115 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes for registering derived FST for generic reading.
-
-#ifndef FST_REGISTER_H_
-#define FST_REGISTER_H_
-
-#include <string>
-#include <type_traits>
-
-
-#include <fst/compat.h>
-#include <fst/generic-register.h>
-#include <fst/util.h>
-
-
-#include <fst/types.h>
-#include <fst/log.h>
-
-namespace fst {
-
-template <class Arc>
-class Fst;
-
-struct FstReadOptions;
-
-// This class represents a single entry in a FstRegister
-template <class Arc>
-struct FstRegisterEntry {
-  using Reader = Fst<Arc> *(*)(std::istream &istrm, const FstReadOptions &opts);
-  using Converter = Fst<Arc> *(*)(const Fst<Arc> &fst);
-
-  Reader reader;
-  Converter converter;
-
-  explicit FstRegisterEntry(Reader reader = nullptr,
-                            Converter converter = nullptr)
-      : reader(reader), converter(converter) {}
-};
-
-// This class maintains the correspondence between a string describing
-// an FST type, and its reader and converter.
-template <class Arc>
-class FstRegister
-    : public GenericRegister<string, FstRegisterEntry<Arc>, FstRegister<Arc>> {
- public:
-  using Reader = typename FstRegisterEntry<Arc>::Reader;
-  using Converter = typename FstRegisterEntry<Arc>::Converter;
-
-  const Reader GetReader(const string &type) const {
-    return this->GetEntry(type).reader;
-  }
-
-  const Converter GetConverter(const string &type) const {
-    return this->GetEntry(type).converter;
-  }
-
- protected:
-  string ConvertKeyToSoFilename(const string &key) const override {
-    string legal_type(key);
-    ConvertToLegalCSymbol(&legal_type);
-    return legal_type + "-fst.so";
-  }
-};
-
-// This class registers an FST type for generic reading and creating.
-// The type must have a default constructor and a copy constructor from
-// Fst<Arc>.
-template <class FST>
-class FstRegisterer : public GenericRegisterer<FstRegister<typename FST::Arc>> {
- public:
-  using Arc = typename FST::Arc;
-  using Entry = typename FstRegister<Arc>::Entry;
-  using Reader = typename FstRegister<Arc>::Reader;
-
-  FstRegisterer()
-      : GenericRegisterer<FstRegister<typename FST::Arc>>(FST().Type(),
-                                                          BuildEntry()) {}
-
- private:
-  static Fst<Arc> *ReadGeneric(
-      std::istream &strm, const FstReadOptions &opts) {
-    static_assert(std::is_base_of<Fst<Arc>, FST>::value,
-                  "FST class does not inherit from Fst<Arc>");
-    return FST::Read(strm, opts);
-  }
-
-  static Entry BuildEntry() {
-    return Entry(&ReadGeneric, &FstRegisterer<FST>::Convert);
-  }
-
-  static Fst<Arc> *Convert(const Fst<Arc> &fst) { return new FST(fst); }
-};
-
-// Convenience macro to generate static FstRegisterer instance.
-#define REGISTER_FST(FST, Arc) \
-  static fst::FstRegisterer<FST<Arc>> FST##_##Arc##_registerer
-
-// Converts an FST to the specified type.
-template <class Arc>
-Fst<Arc> *Convert(const Fst<Arc> &fst, const string &fst_type) {
-  auto *reg = FstRegister<Arc>::GetRegister();
-  const auto converter = reg->GetConverter(fst_type);
-  if (!converter) {
-    FSTERROR() << "Fst::Convert: Unknown FST type " << fst_type << " (arc type "
-               << Arc::Type() << ")";
-    return nullptr;
-  }
-  return converter(fst);
-}
-
-}  // namespace fst
-
-#endif  // FST_REGISTER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/relabel.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/relabel.h
deleted file mode 100644
index 3d73184e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/relabel.h
+++ /dev/null
@@ -1,478 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to relabel an FST (either on input or output).
-
-#ifndef FST_RELABEL_H_
-#define FST_RELABEL_H_
-
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/test-properties.h>
-
-
-#include <unordered_map>
-
-namespace fst {
-
-// Relabels either the input labels or output labels. The old to
-// new labels are specified using a vector of std::pair<Label, Label>.
-// Any label associations not specified are assumed to be identity
-// mapping. The destination labels must be valid labels (e.g., not kNoLabel).
-template <class Arc>
-void Relabel(
-    MutableFst<Arc> *fst,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &ipairs,
-    const std::vector<std::pair<typename Arc::Label, typename Arc::Label>>
-        &opairs) {
-  using Label = typename Arc::Label;
-  const auto props = fst->Properties(kFstProperties, false);
-  // Constructs label-to-label maps.
-  std::unordered_map<Label, Label> input_map;
-  for (auto &ipair : ipairs) input_map[ipair.first] = ipair.second;
-  std::unordered_map<Label, Label> output_map;
-  for (auto &opair : opairs) output_map[opair.first] = opair.second;
-  for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
-       siter.Next()) {
-    for (MutableArcIterator<MutableFst<Arc>> aiter(fst, siter.Value());
-         !aiter.Done(); aiter.Next()) {
-      auto arc = aiter.Value();
-      // Relabels input.
-      auto it = input_map.find(arc.ilabel);
-      if (it != input_map.end()) {
-        if (it->second == kNoLabel) {
-          FSTERROR() << "Input symbol ID " << arc.ilabel
-                     << " missing from target vocabulary";
-          fst->SetProperties(kError, kError);
-          return;
-        }
-        arc.ilabel = it->second;
-      }
-      // Relabels output.
-      it = output_map.find(arc.olabel);
-      if (it != output_map.end()) {
-        if (it->second == kNoLabel) {
-          FSTERROR() << "Output symbol id " << arc.olabel
-                     << " missing from target vocabulary";
-          fst->SetProperties(kError, kError);
-          return;
-        }
-        arc.olabel = it->second;
-      }
-      aiter.SetValue(arc);
-    }
-  }
-  fst->SetProperties(RelabelProperties(props), kFstProperties);
-}
-
-// Relabels either the input labels or output labels. The old to
-// new labels are specified using pairs of old and new symbol tables.
-// The tables must contain (at least) all labels on the appropriate side of the
-// FST. If the 'unknown_i(o)symbol' is non-empty, it is used to label any
-// missing symbol in new_i(o)symbols table.
-template <class Arc>
-void Relabel(MutableFst<Arc> *fst,
-             const SymbolTable *old_isymbols, const SymbolTable *new_isymbols,
-             const string &unknown_isymbol, bool attach_new_isymbols,
-             const SymbolTable *old_osymbols, const SymbolTable *new_osymbols,
-             const string &unknown_osymbol, bool attach_new_osymbols) {
-  using Label = typename Arc::Label;
-  // Constructs vectors of input-side label pairs.
-  std::vector<std::pair<Label, Label>> ipairs;
-  if (old_isymbols && new_isymbols) {
-    size_t num_missing_syms = 0;
-    Label unknown_ilabel = kNoLabel;
-    if (!unknown_isymbol.empty()) {
-      unknown_ilabel = new_isymbols->Find(unknown_isymbol);
-      if (unknown_ilabel == kNoLabel) {
-        VLOG(1) << "Input symbol '" << unknown_isymbol
-                << "' missing from target symbol table";
-        ++num_missing_syms;
-      }
-    }
-
-    for (SymbolTableIterator siter(*old_isymbols); !siter.Done();
-         siter.Next()) {
-      const auto old_index = siter.Value();
-      const auto symbol = siter.Symbol();
-      auto new_index = new_isymbols->Find(siter.Symbol());
-      if (new_index == kNoLabel) {
-        if (unknown_ilabel != kNoLabel) {
-          new_index = unknown_ilabel;
-        } else {
-          VLOG(1) << "Input symbol ID " << old_index << " symbol '" << symbol
-                  << "' missing from target symbol table";
-          ++num_missing_syms;
-        }
-      }
-      ipairs.push_back(std::make_pair(old_index, new_index));
-    }
-    if (num_missing_syms > 0) {
-      LOG(WARNING) << "Target symbol table missing: " << num_missing_syms
-                   << " input symbols";
-    }
-    if (attach_new_isymbols) fst->SetInputSymbols(new_isymbols);
-  }
-  // Constructs vectors of output-side label pairs.
-  std::vector<std::pair<Label, Label>> opairs;
-  if (old_osymbols && new_osymbols) {
-    size_t num_missing_syms = 0;
-    Label unknown_olabel = kNoLabel;
-    if (!unknown_osymbol.empty()) {
-      unknown_olabel = new_osymbols->Find(unknown_osymbol);
-      if (unknown_olabel == kNoLabel) {
-        VLOG(1) << "Output symbol '" << unknown_osymbol
-                << "' missing from target symbol table";
-        ++num_missing_syms;
-      }
-    }
-
-    for (SymbolTableIterator siter(*old_osymbols); !siter.Done();
-         siter.Next()) {
-      const auto old_index = siter.Value();
-      const auto symbol = siter.Symbol();
-      auto new_index = new_osymbols->Find(siter.Symbol());
-      if (new_index == kNoLabel) {
-        if (unknown_olabel != kNoLabel) {
-          new_index = unknown_olabel;
-        } else {
-          VLOG(1) << "Output symbol ID " << old_index << " symbol '" << symbol
-                  << "' missing from target symbol table";
-          ++num_missing_syms;
-        }
-      }
-      opairs.push_back(std::make_pair(old_index, new_index));
-    }
-    if (num_missing_syms > 0) {
-      LOG(WARNING) << "Target symbol table missing: " << num_missing_syms
-                   << " output symbols";
-    }
-    if (attach_new_osymbols) fst->SetOutputSymbols(new_osymbols);
-  }
-  // Calls relabel using vector of relabel pairs.
-  Relabel(fst, ipairs, opairs);
-}
-
-// Same as previous but no special allowance for unknown symbols. Kept
-// for backward compat.
-template <class Arc>
-void Relabel(MutableFst<Arc> *fst, const SymbolTable *old_isymbols,
-             const SymbolTable *new_isymbols, bool attach_new_isymbols,
-             const SymbolTable *old_osymbols, const SymbolTable *new_osymbols,
-             bool attach_new_osymbols) {
-  Relabel(fst,
-          old_isymbols, new_isymbols, "" /* no unknown isymbol */,
-          attach_new_isymbols,
-          old_osymbols, new_osymbols, "" /* no unknown ioymbol */,
-          attach_new_osymbols);
-}
-
-
-// Relabels either the input labels or output labels. The old to
-// new labels are specified using symbol tables. Any label associations not
-// specified are assumed to be identity mapping.
-template <class Arc>
-void Relabel(MutableFst<Arc> *fst, const SymbolTable *new_isymbols,
-             const SymbolTable *new_osymbols) {
-  Relabel(fst, fst->InputSymbols(), new_isymbols, true, fst->OutputSymbols(),
-          new_osymbols, true);
-}
-
-using RelabelFstOptions = CacheOptions;
-
-template <class Arc>
-class RelabelFst;
-
-namespace internal {
-
-// Relabels an FST from one symbol set to another. Relabeling can either be on
-// input or output space. RelabelFst implements a delayed version of the
-// relabel. Arcs are relabeled on the fly and not cached; i.e., each request is
-// recomputed.
-template <class Arc>
-class RelabelFstImpl : public CacheImpl<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::WriteHeader;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheImpl<Arc>::PushArc;
-  using CacheImpl<Arc>::HasArcs;
-  using CacheImpl<Arc>::HasFinal;
-  using CacheImpl<Arc>::HasStart;
-  using CacheImpl<Arc>::SetArcs;
-  using CacheImpl<Arc>::SetFinal;
-  using CacheImpl<Arc>::SetStart;
-
-  friend class StateIterator<RelabelFst<Arc>>;
-
-  RelabelFstImpl(const Fst<Arc> &fst,
-                 const std::vector<std::pair<Label, Label>> &ipairs,
-                 const std::vector<std::pair<Label, Label>> &opairs,
-                 const RelabelFstOptions &opts)
-      : CacheImpl<Arc>(opts),
-        fst_(fst.Copy()),
-        relabel_input_(false),
-        relabel_output_(false) {
-    SetProperties(RelabelProperties(fst.Properties(kCopyProperties, false)));
-    SetType("relabel");
-    // Creates input label map.
-    if (!ipairs.empty()) {
-      for (auto &ipair : ipairs) input_map_[ipair.first] = ipair.second;
-      relabel_input_ = true;
-    }
-    // Creates output label map.
-    if (!opairs.empty()) {
-      for (auto &opair : opairs) output_map_[opair.first] = opair.second;
-      relabel_output_ = true;
-    }
-  }
-
-  RelabelFstImpl(const Fst<Arc> &fst, const SymbolTable *old_isymbols,
-                 const SymbolTable *new_isymbols,
-                 const SymbolTable *old_osymbols,
-                 const SymbolTable *new_osymbols, const RelabelFstOptions &opts)
-      : CacheImpl<Arc>(opts),
-        fst_(fst.Copy()),
-        relabel_input_(false),
-        relabel_output_(false) {
-    SetType("relabel");
-    SetProperties(RelabelProperties(fst.Properties(kCopyProperties, false)));
-    SetInputSymbols(old_isymbols);
-    SetOutputSymbols(old_osymbols);
-    if (old_isymbols && new_isymbols &&
-        old_isymbols->LabeledCheckSum() != new_isymbols->LabeledCheckSum()) {
-      for (SymbolTableIterator siter(*old_isymbols); !siter.Done();
-           siter.Next()) {
-        input_map_[siter.Value()] = new_isymbols->Find(siter.Symbol());
-      }
-      SetInputSymbols(new_isymbols);
-      relabel_input_ = true;
-    }
-    if (old_osymbols && new_osymbols &&
-        old_osymbols->LabeledCheckSum() != new_osymbols->LabeledCheckSum()) {
-      for (SymbolTableIterator siter(*old_osymbols); !siter.Done();
-           siter.Next()) {
-        output_map_[siter.Value()] = new_osymbols->Find(siter.Symbol());
-      }
-      SetOutputSymbols(new_osymbols);
-      relabel_output_ = true;
-    }
-  }
-
-  RelabelFstImpl(const RelabelFstImpl<Arc> &impl)
-      : CacheImpl<Arc>(impl),
-        fst_(impl.fst_->Copy(true)),
-        input_map_(impl.input_map_),
-        output_map_(impl.output_map_),
-        relabel_input_(impl.relabel_input_),
-        relabel_output_(impl.relabel_output_) {
-    SetType("relabel");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  StateId Start() {
-    if (!HasStart()) SetStart(fst_->Start());
-    return CacheImpl<Arc>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) SetFinal(s, fst_->Final(s));
-    return CacheImpl<Arc>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumOutputEpsilons(s);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && fst_->Properties(kError, false)) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<Arc>::InitArcIterator(s, data);
-  }
-
-  void Expand(StateId s) {
-    for (ArcIterator<Fst<Arc>> aiter(*fst_, s); !aiter.Done(); aiter.Next()) {
-      auto arc = aiter.Value();
-      if (relabel_input_) {
-        auto it = input_map_.find(arc.ilabel);
-        if (it != input_map_.end()) arc.ilabel = it->second;
-      }
-      if (relabel_output_) {
-        auto it = output_map_.find(arc.olabel);
-        if (it != output_map_.end()) {
-          arc.olabel = it->second;
-        }
-      }
-      PushArc(s, arc);
-    }
-    SetArcs(s);
-  }
-
- private:
-  std::unique_ptr<const Fst<Arc>> fst_;
-
-  std::unordered_map<Label, Label> input_map_;
-  std::unordered_map<Label, Label> output_map_;
-  bool relabel_input_;
-  bool relabel_output_;
-};
-
-}  // namespace internal
-
-// This class attaches interface to implementation and handles
-// reference counting, delegating most methods to ImplToFst.
-template <class A>
-class RelabelFst : public ImplToFst<internal::RelabelFstImpl<A>> {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::RelabelFstImpl<Arc>;
-
-  friend class ArcIterator<RelabelFst<A>>;
-  friend class StateIterator<RelabelFst<A>>;
-
-  RelabelFst(const Fst<Arc> &fst,
-             const std::vector<std::pair<Label, Label>> &ipairs,
-             const std::vector<std::pair<Label, Label>> &opairs,
-             const RelabelFstOptions &opts = RelabelFstOptions())
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, ipairs, opairs, opts)) {}
-
-  RelabelFst(const Fst<Arc> &fst, const SymbolTable *new_isymbols,
-             const SymbolTable *new_osymbols,
-             const RelabelFstOptions &opts = RelabelFstOptions())
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, fst.InputSymbols(), new_isymbols,
-                                   fst.OutputSymbols(), new_osymbols, opts)) {}
-
-  RelabelFst(const Fst<Arc> &fst, const SymbolTable *old_isymbols,
-             const SymbolTable *new_isymbols, const SymbolTable *old_osymbols,
-             const SymbolTable *new_osymbols,
-             const RelabelFstOptions &opts = RelabelFstOptions())
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, old_isymbols, new_isymbols,
-                                               old_osymbols, new_osymbols,
-                                               opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  RelabelFst(const RelabelFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Gets a copy of this RelabelFst. See Fst<>::Copy() for further doc.
-  RelabelFst<Arc> *Copy(bool safe = false) const override {
-    return new RelabelFst<Arc>(*this, safe);
-  }
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    return GetMutableImpl()->InitArcIterator(s, data);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  RelabelFst &operator=(const RelabelFst &) = delete;
-};
-
-// Specialization for RelabelFst.
-template <class Arc>
-class StateIterator<RelabelFst<Arc>> : public StateIteratorBase<Arc> {
- public:
-  using StateId = typename Arc::StateId;
-
-  explicit StateIterator(const RelabelFst<Arc> &fst)
-      : impl_(fst.GetImpl()), siter_(*impl_->fst_), s_(0) {}
-
-  bool Done() const final { return siter_.Done(); }
-
-  StateId Value() const final { return s_; }
-
-  void Next() final {
-    if (!siter_.Done()) {
-      ++s_;
-      siter_.Next();
-    }
-  }
-
-  void Reset() final {
-    s_ = 0;
-    siter_.Reset();
-  }
-
- private:
-  const internal::RelabelFstImpl<Arc>* impl_;
-  StateIterator<Fst<Arc>> siter_;
-  StateId s_;
-
-  StateIterator(const StateIterator &) = delete;
-  StateIterator &operator=(const StateIterator &) = delete;
-};
-
-// Specialization for RelabelFst.
-template <class Arc>
-class ArcIterator<RelabelFst<Arc>> : public CacheArcIterator<RelabelFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const RelabelFst<Arc> &fst, StateId s)
-      : CacheArcIterator<RelabelFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc>
-inline void RelabelFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<RelabelFst<Arc>>(*this);
-}
-
-// Useful alias when using StdArc.
-using StdRelabelFst = RelabelFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_RELABEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/replace-util.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/replace-util.h
deleted file mode 100644
index f4e15621..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/replace-util.h
+++ /dev/null
@@ -1,629 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Utility classes for the recursive replacement of FSTs (RTNs).
-
-#ifndef FST_REPLACE_UTIL_H_
-#define FST_REPLACE_UTIL_H_
-
-#include <map>
-#include <unordered_map>
-#include <unordered_set>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/connect.h>
-#include <fst/mutable-fst.h>
-#include <fst/topsort.h>
-#include <fst/vector-fst.h>
-
-
-namespace fst {
-
-// This specifies what labels to output on the call or return arc. Note that
-// REPLACE_LABEL_INPUT and REPLACE_LABEL_OUTPUT will produce transducers when
-// applied to acceptors.
-enum ReplaceLabelType {
-  // Epsilon labels on both input and output.
-  REPLACE_LABEL_NEITHER = 1,
-  // Non-epsilon labels on input and epsilon on output.
-  REPLACE_LABEL_INPUT = 2,
-  // Epsilon on input and non-epsilon on output.
-  REPLACE_LABEL_OUTPUT = 3,
-  // Non-epsilon labels on both input and output.
-  REPLACE_LABEL_BOTH = 4
-};
-
-// By default ReplaceUtil will copy the input label of the replace arc.
-// The call_label_type and return_label_type options specify how to manage
-// the labels of the call arc and the return arc of the replace FST
-struct ReplaceUtilOptions {
-  int64 root;                          // Root rule for expansion.
-  ReplaceLabelType call_label_type;    // How to label call arc.
-  ReplaceLabelType return_label_type;  // How to label return arc.
-  int64 return_label;                  // Label to put on return arc.
-
-  explicit ReplaceUtilOptions(
-      int64 root = kNoLabel,
-      ReplaceLabelType call_label_type = REPLACE_LABEL_INPUT,
-      ReplaceLabelType return_label_type = REPLACE_LABEL_NEITHER,
-      int64 return_label = 0)
-      : root(root),
-        call_label_type(call_label_type),
-        return_label_type(return_label_type),
-        return_label(return_label) {}
-
-  // For backwards compatibility.
-  ReplaceUtilOptions(int64 root, bool epsilon_replace_arc)
-      : ReplaceUtilOptions(root,
-                           epsilon_replace_arc ? REPLACE_LABEL_NEITHER
-                                               : REPLACE_LABEL_INPUT) {}
-};
-
-// Every non-terminal on a path appears as the first label on that path in every
-// FST associated with a given SCC of the replace dependency graph. This would
-// be true if the SCC were formed from left-linear grammar rules.
-constexpr uint8 kReplaceSCCLeftLinear = 0x01;
-// Every non-terminal on a path appears as the final label on that path in every
-// FST associated with a given SCC of the replace dependency graph. This would
-// be true if the SCC were formed from right-linear grammar rules.
-constexpr uint8 kReplaceSCCRightLinear = 0x02;
-// The SCC in the replace dependency graph has more than one state or a
-// self-loop.
-constexpr uint8 kReplaceSCCNonTrivial = 0x04;
-
-// Defined in replace.h.
-template <class Arc>
-void Replace(
-    const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>> &,
-    MutableFst<Arc> *, const ReplaceUtilOptions &);
-
-// Utility class for the recursive replacement of FSTs (RTNs). The user provides
-// a set of label/FST pairs at construction. These are used by methods for
-// testing cyclic dependencies and connectedness and doing RTN connection and
-// specific FST replacement by label or for various optimization properties. The
-// modified results can be obtained with the GetFstPairs() or
-// GetMutableFstPairs() methods.
-template <class Arc>
-class ReplaceUtil {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstPair = std::pair<Label, const Fst<Arc> *>;
-  using MutableFstPair = std::pair<Label, MutableFst<Arc> *>;
-  using NonTerminalHash = std::unordered_map<Label, Label>;
-
-  // Constructs from mutable FSTs; FST ownership is given to ReplaceUtil.
-  ReplaceUtil(const std::vector<MutableFstPair> &fst_pairs,
-              const ReplaceUtilOptions &opts);
-
-  // Constructs from FSTs; FST ownership is retained by caller.
-  ReplaceUtil(const std::vector<FstPair> &fst_pairs,
-              const ReplaceUtilOptions &opts);
-
-  // Constructs from ReplaceFst internals; FST ownership is retained by caller.
-  ReplaceUtil(const std::vector<std::unique_ptr<const Fst<Arc>>> &fst_array,
-              const NonTerminalHash &nonterminal_hash,
-              const ReplaceUtilOptions &opts);
-
-  ~ReplaceUtil() {
-    for (Label i = 0; i < fst_array_.size(); ++i) delete fst_array_[i];
-  }
-
-  // True if the non-terminal dependencies are cyclic. Cyclic dependencies will
-  // result in an unexpandable FST.
-  bool CyclicDependencies() const {
-    GetDependencies(false);
-    return depprops_ & kCyclic;
-  }
-
-  // Returns the strongly-connected component ID in the dependency graph of the
-  // replace FSTS.
-  StateId SCC(Label label) const {
-    GetDependencies(false);
-    const auto it = nonterminal_hash_.find(label);
-    if (it == nonterminal_hash_.end()) return kNoStateId;
-    return depscc_[it->second];
-  }
-
-  // Returns properties for the strongly-connected component in the dependency
-  // graph of the replace FSTs. If the SCC is kReplaceSCCLeftLinear or
-  // kReplaceSCCRightLinear, that SCC can be represented as finite-state despite
-  // any cyclic dependencies, but not by the usual replacement operation (see
-  // fst/extensions/pdt/replace.h).
-  uint8 SCCProperties(StateId scc_id) {
-    GetSCCProperties();
-    return depsccprops_[scc_id];
-  }
-
-  // Returns true if no useless FSTs, states or transitions are present in the
-  // RTN.
-  bool Connected() const {
-    GetDependencies(false);
-    uint64 props = kAccessible | kCoAccessible;
-    for (Label i = 0; i < fst_array_.size(); ++i) {
-      if (!fst_array_[i]) continue;
-      if (fst_array_[i]->Properties(props, true) != props || !depaccess_[i]) {
-        return false;
-      }
-    }
-    return true;
-  }
-
-  // Removes useless FSTs, states and transitions from the RTN.
-  void Connect();
-
-  // Replaces FSTs specified by labels, unless there are cyclic dependencies.
-  void ReplaceLabels(const std::vector<Label> &labels);
-
-  // Replaces FSTs that have at most nstates states, narcs arcs and nnonterm
-  // non-terminals (updating in reverse dependency order), unless there are
-  // cyclic dependencies.
-  void ReplaceBySize(size_t nstates, size_t narcs, size_t nnonterms);
-
-  // Replaces singleton FSTS, unless there are cyclic dependencies.
-  void ReplaceTrivial() { ReplaceBySize(2, 1, 1); }
-
-  // Replaces non-terminals that have at most ninstances instances (updating in
-  // dependency order), unless there are cyclic dependencies.
-  void ReplaceByInstances(size_t ninstances);
-
-  // Replaces non-terminals that have only one instance, unless there are cyclic
-  // dependencies.
-  void ReplaceUnique() { ReplaceByInstances(1); }
-
-  // Returns label/FST pairs, retaining FST ownership.
-  void GetFstPairs(std::vector<FstPair> *fst_pairs);
-
-  // Returns label/mutable FST pairs, giving FST ownership over to the caller.
-  void GetMutableFstPairs(std::vector<MutableFstPair> *mutable_fst_pairs);
-
- private:
-  // FST statistics.
-  struct ReplaceStats {
-    StateId nstates;  // Number of states.
-    StateId nfinal;   // Number of final states.
-    size_t narcs;     // Number of arcs.
-    Label nnonterms;  // Number of non-terminals in FST.
-    size_t nref;      // Number of non-terminal instances referring to this FST.
-    // Number of times that ith FST references this FST
-    std::map<Label, size_t> inref;
-    // Number of times that this FST references the ith FST
-    std::map<Label, size_t> outref;
-
-    ReplaceStats() : nstates(0), nfinal(0), narcs(0), nnonterms(0), nref(0) {}
-  };
-
-  // Checks that Mutable FSTs exists, creating them if necessary.
-  void CheckMutableFsts();
-
-  // Computes the dependency graph for the RTN, computing dependency statistics
-  // if stats is true.
-  void GetDependencies(bool stats) const;
-
-  void ClearDependencies() const {
-    depfst_.DeleteStates();
-    stats_.clear();
-    depprops_ = 0;
-    depsccprops_.clear();
-    have_stats_ = false;
-  }
-
-  // Gets topological order of dependencies, returning false with cyclic input.
-  bool GetTopOrder(const Fst<Arc> &fst, std::vector<Label> *toporder) const;
-
-  // Updates statistics to reflect the replacement of the jth FST.
-  void UpdateStats(Label j);
-
-  // Computes the properties for the strongly-connected component in the
-  // dependency graph of the replace FSTs.
-  void GetSCCProperties() const;
-
-  Label root_label_;                                  // Root non-terminal.
-  Label root_fst_;                                    // Root FST ID.
-  ReplaceLabelType call_label_type_;                  // See Replace().
-  ReplaceLabelType return_label_type_;                // See Replace().
-  int64 return_label_;                                // See Replace().
-  std::vector<const Fst<Arc> *> fst_array_;           // FST per ID.
-  std::vector<MutableFst<Arc> *> mutable_fst_array_;  // Mutable FST per ID.
-  std::vector<Label> nonterminal_array_;              // FST ID to non-terminal.
-  NonTerminalHash nonterminal_hash_;                  // Non-terminal to FST ID.
-  mutable VectorFst<Arc> depfst_;                     // FST ID dependencies.
-  mutable std::vector<StateId> depscc_;               // FST SCC ID.
-  mutable std::vector<bool> depaccess_;               // FST ID accessibility.
-  mutable uint64 depprops_;                           // Dependency FST props.
-  mutable bool have_stats_;                  // Have dependency statistics?
-  mutable std::vector<ReplaceStats> stats_;  // Per-FST statistics.
-  mutable std::vector<uint8> depsccprops_;   // SCC properties.
-  ReplaceUtil(const ReplaceUtil &) = delete;
-  ReplaceUtil &operator=(const ReplaceUtil &) = delete;
-};
-
-template <class Arc>
-ReplaceUtil<Arc>::ReplaceUtil(const std::vector<MutableFstPair> &fst_pairs,
-                              const ReplaceUtilOptions &opts)
-    : root_label_(opts.root),
-      call_label_type_(opts.call_label_type),
-      return_label_type_(opts.return_label_type),
-      return_label_(opts.return_label),
-      depprops_(0),
-      have_stats_(false) {
-  fst_array_.push_back(nullptr);
-  mutable_fst_array_.push_back(nullptr);
-  nonterminal_array_.push_back(kNoLabel);
-  for (const auto &fst_pair : fst_pairs) {
-    const auto label = fst_pair.first;
-    auto *fst = fst_pair.second;
-    nonterminal_hash_[label] = fst_array_.size();
-    nonterminal_array_.push_back(label);
-    fst_array_.push_back(fst);
-    mutable_fst_array_.push_back(fst);
-  }
-  root_fst_ = nonterminal_hash_[root_label_];
-  if (!root_fst_) {
-    FSTERROR() << "ReplaceUtil: No root FST for label: " << root_label_;
-  }
-}
-
-template <class Arc>
-ReplaceUtil<Arc>::ReplaceUtil(const std::vector<FstPair> &fst_pairs,
-                              const ReplaceUtilOptions &opts)
-    : root_label_(opts.root),
-      call_label_type_(opts.call_label_type),
-      return_label_type_(opts.return_label_type),
-      return_label_(opts.return_label),
-      depprops_(0),
-      have_stats_(false) {
-  fst_array_.push_back(nullptr);
-  nonterminal_array_.push_back(kNoLabel);
-  for (const auto &fst_pair : fst_pairs) {
-    const auto label = fst_pair.first;
-    const auto *fst = fst_pair.second;
-    nonterminal_hash_[label] = fst_array_.size();
-    nonterminal_array_.push_back(label);
-    fst_array_.push_back(fst->Copy());
-  }
-  root_fst_ = nonterminal_hash_[root_label_];
-  if (!root_fst_) {
-    FSTERROR() << "ReplaceUtil: No root FST for label: " << root_label_;
-  }
-}
-
-template <class Arc>
-ReplaceUtil<Arc>::ReplaceUtil(
-    const std::vector<std::unique_ptr<const Fst<Arc>>> &fst_array,
-    const NonTerminalHash &nonterminal_hash, const ReplaceUtilOptions &opts)
-    : root_fst_(opts.root),
-      call_label_type_(opts.call_label_type),
-      return_label_type_(opts.return_label_type),
-      return_label_(opts.return_label),
-      nonterminal_array_(fst_array.size()),
-      nonterminal_hash_(nonterminal_hash),
-      depprops_(0),
-      have_stats_(false) {
-  fst_array_.push_back(nullptr);
-  for (size_t i = 1; i < fst_array.size(); ++i) {
-    fst_array_.push_back(fst_array[i]->Copy());
-  }
-  for (auto it = nonterminal_hash.begin(); it != nonterminal_hash.end(); ++it) {
-    nonterminal_array_[it->second] = it->first;
-  }
-  root_label_ = nonterminal_array_[root_fst_];
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::GetDependencies(bool stats) const {
-  if (depfst_.NumStates() > 0) {
-    if (stats && !have_stats_) {
-      ClearDependencies();
-    } else {
-      return;
-    }
-  }
-  have_stats_ = stats;
-  if (have_stats_) stats_.reserve(fst_array_.size());
-  for (Label i = 0; i < fst_array_.size(); ++i) {
-    depfst_.AddState();
-    depfst_.SetFinal(i, Weight::One());
-    if (have_stats_) stats_.push_back(ReplaceStats());
-  }
-  depfst_.SetStart(root_fst_);
-  // An arc from each state (representing the FST) to the state representing the
-  // FST being replaced
-  for (Label i = 0; i < fst_array_.size(); ++i) {
-    const auto *ifst = fst_array_[i];
-    if (!ifst) continue;
-    for (StateIterator<Fst<Arc>> siter(*ifst); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      if (have_stats_) {
-        ++stats_[i].nstates;
-        if (ifst->Final(s) != Weight::Zero()) ++stats_[i].nfinal;
-      }
-      for (ArcIterator<Fst<Arc>> aiter(*ifst, s); !aiter.Done();
-           aiter.Next()) {
-        if (have_stats_) ++stats_[i].narcs;
-        const auto &arc = aiter.Value();
-        auto it = nonterminal_hash_.find(arc.olabel);
-        if (it != nonterminal_hash_.end()) {
-          const auto j = it->second;
-          depfst_.AddArc(i, Arc(arc.olabel, arc.olabel, Weight::One(), j));
-          if (have_stats_) {
-            ++stats_[i].nnonterms;
-            ++stats_[j].nref;
-            ++stats_[j].inref[i];
-            ++stats_[i].outref[j];
-          }
-        }
-      }
-    }
-  }
-  // Computes accessibility info.
-  SccVisitor<Arc> scc_visitor(&depscc_, &depaccess_, nullptr, &depprops_);
-  DfsVisit(depfst_, &scc_visitor);
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::UpdateStats(Label j) {
-  if (!have_stats_) {
-    FSTERROR() << "ReplaceUtil::UpdateStats: Stats not available";
-    return;
-  }
-  if (j == root_fst_) return;  // Can't replace root.
-  for (auto in = stats_[j].inref.begin(); in != stats_[j].inref.end(); ++in) {
-    const auto i = in->first;
-    const auto ni = in->second;
-    stats_[i].nstates += stats_[j].nstates * ni;
-    stats_[i].narcs += (stats_[j].narcs + 1) * ni;
-    stats_[i].nnonterms += (stats_[j].nnonterms - 1) * ni;
-    stats_[i].outref.erase(j);
-    for (auto out = stats_[j].outref.begin(); out != stats_[j].outref.end();
-         ++out) {
-      const auto k = out->first;
-      const auto nk = out->second;
-      stats_[i].outref[k] += ni * nk;
-    }
-  }
-  for (auto out = stats_[j].outref.begin(); out != stats_[j].outref.end();
-       ++out) {
-    const auto k = out->first;
-    const auto nk = out->second;
-    stats_[k].nref -= nk;
-    stats_[k].inref.erase(j);
-    for (auto in = stats_[j].inref.begin(); in != stats_[j].inref.end(); ++in) {
-      const auto i = in->first;
-      const auto ni = in->second;
-      stats_[k].inref[i] += ni * nk;
-      stats_[k].nref += ni * nk;
-    }
-  }
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::CheckMutableFsts() {
-  if (mutable_fst_array_.empty()) {
-    for (Label i = 0; i < fst_array_.size(); ++i) {
-      if (!fst_array_[i]) {
-        mutable_fst_array_.push_back(nullptr);
-      } else {
-        mutable_fst_array_.push_back(new VectorFst<Arc>(*fst_array_[i]));
-        delete fst_array_[i];
-        fst_array_[i] = mutable_fst_array_[i];
-      }
-    }
-  }
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::Connect() {
-  CheckMutableFsts();
-  static constexpr auto props = kAccessible | kCoAccessible;
-  for (auto *mutable_fst : mutable_fst_array_) {
-    if (!mutable_fst) continue;
-    if (mutable_fst->Properties(props, false) != props) {
-      fst::Connect(mutable_fst);
-    }
-  }
-  GetDependencies(false);
-  for (Label i = 0; i < mutable_fst_array_.size(); ++i) {
-    auto *fst = mutable_fst_array_[i];
-    if (fst && !depaccess_[i]) {
-      delete fst;
-      fst_array_[i] = nullptr;
-      mutable_fst_array_[i] = nullptr;
-    }
-  }
-  ClearDependencies();
-}
-
-template <class Arc>
-bool ReplaceUtil<Arc>::GetTopOrder(const Fst<Arc> &fst,
-                                   std::vector<Label> *toporder) const {
-  // Finds topological order of dependencies.
-  std::vector<StateId> order;
-  bool acyclic = false;
-  TopOrderVisitor<Arc> top_order_visitor(&order, &acyclic);
-  DfsVisit(fst, &top_order_visitor);
-  if (!acyclic) {
-    LOG(WARNING) << "ReplaceUtil::GetTopOrder: Cyclical label dependencies";
-    return false;
-  }
-  toporder->resize(order.size());
-  for (Label i = 0; i < order.size(); ++i) (*toporder)[order[i]] = i;
-  return true;
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::ReplaceLabels(const std::vector<Label> &labels) {
-  CheckMutableFsts();
-  std::unordered_set<Label> label_set;
-  for (const auto label : labels) {
-    // Can't replace root.
-    if (label != root_label_) label_set.insert(label);
-  }
-  // Finds FST dependencies restricted to the labels requested.
-  GetDependencies(false);
-  VectorFst<Arc> pfst(depfst_);
-  for (StateId i = 0; i < pfst.NumStates(); ++i) {
-    std::vector<Arc> arcs;
-    for (ArcIterator<VectorFst<Arc>> aiter(pfst, i); !aiter.Done();
-         aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto label = nonterminal_array_[arc.nextstate];
-      if (label_set.count(label) > 0) arcs.push_back(arc);
-    }
-    pfst.DeleteArcs(i);
-    for (const auto &arc : arcs) pfst.AddArc(i, arc);
-  }
-  std::vector<Label> toporder;
-  if (!GetTopOrder(pfst, &toporder)) {
-    ClearDependencies();
-    return;
-  }
-  // Visits FSTs in reverse topological order of dependencies and performs
-  // replacements.
-  for (Label o = toporder.size() - 1; o >= 0; --o) {
-    std::vector<FstPair> fst_pairs;
-    auto s = toporder[o];
-    for (ArcIterator<VectorFst<Arc>> aiter(pfst, s); !aiter.Done();
-         aiter.Next()) {
-      const auto &arc = aiter.Value();
-      const auto label = nonterminal_array_[arc.nextstate];
-      const auto *fst = fst_array_[arc.nextstate];
-      fst_pairs.push_back(std::make_pair(label, fst));
-    }
-    if (fst_pairs.empty()) continue;
-    const auto label = nonterminal_array_[s];
-    const auto *fst = fst_array_[s];
-    fst_pairs.push_back(std::make_pair(label, fst));
-    const ReplaceUtilOptions opts(label, call_label_type_, return_label_type_,
-                                  return_label_);
-    Replace(fst_pairs, mutable_fst_array_[s], opts);
-  }
-  ClearDependencies();
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::ReplaceBySize(size_t nstates, size_t narcs,
-                                     size_t nnonterms) {
-  std::vector<Label> labels;
-  GetDependencies(true);
-  std::vector<Label> toporder;
-  if (!GetTopOrder(depfst_, &toporder)) {
-    ClearDependencies();
-    return;
-  }
-  for (Label o = toporder.size() - 1; o >= 0; --o) {
-    const auto j = toporder[o];
-    if (stats_[j].nstates <= nstates && stats_[j].narcs <= narcs &&
-        stats_[j].nnonterms <= nnonterms) {
-      labels.push_back(nonterminal_array_[j]);
-      UpdateStats(j);
-    }
-  }
-  ReplaceLabels(labels);
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::ReplaceByInstances(size_t ninstances) {
-  std::vector<Label> labels;
-  GetDependencies(true);
-  std::vector<Label> toporder;
-  if (!GetTopOrder(depfst_, &toporder)) {
-    ClearDependencies();
-    return;
-  }
-  for (Label o = 0; o < toporder.size(); ++o) {
-    const auto j = toporder[o];
-    if (stats_[j].nref <= ninstances) {
-      labels.push_back(nonterminal_array_[j]);
-      UpdateStats(j);
-    }
-  }
-  ReplaceLabels(labels);
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::GetFstPairs(std::vector<FstPair> *fst_pairs) {
-  CheckMutableFsts();
-  fst_pairs->clear();
-  for (Label i = 0; i < fst_array_.size(); ++i) {
-    const auto label = nonterminal_array_[i];
-    const auto *fst = fst_array_[i];
-    if (!fst) continue;
-    fst_pairs->push_back(std::make_pair(label, fst));
-  }
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::GetMutableFstPairs(
-    std::vector<MutableFstPair> *mutable_fst_pairs) {
-  CheckMutableFsts();
-  mutable_fst_pairs->clear();
-  for (Label i = 0; i < mutable_fst_array_.size(); ++i) {
-    const auto label = nonterminal_array_[i];
-    const auto *fst = mutable_fst_array_[i];
-    if (!fst) continue;
-    mutable_fst_pairs->push_back(std::make_pair(label, fst->Copy()));
-  }
-}
-
-template <class Arc>
-void ReplaceUtil<Arc>::GetSCCProperties() const {
-  if (!depsccprops_.empty()) return;
-  GetDependencies(false);
-  if (depscc_.empty()) return;
-  for (StateId scc = 0; scc < depscc_.size(); ++scc) {
-    depsccprops_.push_back(kReplaceSCCLeftLinear | kReplaceSCCRightLinear);
-  }
-  if (!(depprops_ & kCyclic)) return;  // No cyclic dependencies.
-  // Checks for self-loops in the dependency graph.
-  for (StateId scc = 0; scc < depscc_.size(); ++scc) {
-    for (ArcIterator<Fst<Arc> > aiter(depfst_, scc);
-         !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      if (arc.nextstate == scc) {  // SCC has a self loop.
-        depsccprops_[scc] |= kReplaceSCCNonTrivial;
-      }
-    }
-  }
-  std::vector<bool> depscc_visited(depscc_.size(), false);
-  for (Label i = 0; i < fst_array_.size(); ++i) {
-    const auto *fst = fst_array_[i];
-    if (!fst) continue;
-    const auto depscc = depscc_[i];
-    if (depscc_visited[depscc]) {  // SCC has more than one state.
-      depsccprops_[depscc] |= kReplaceSCCNonTrivial;
-    }
-    depscc_visited[depscc] = true;
-    std::vector<StateId> fstscc;  // SCCs of the current FST.
-    uint64 fstprops;
-    SccVisitor<Arc> scc_visitor(&fstscc, nullptr, nullptr, &fstprops);
-    DfsVisit(*fst, &scc_visitor);
-    for (StateIterator<Fst<Arc>> siter(*fst); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      for (ArcIterator<Fst<Arc>> aiter(*fst, s); !aiter.Done(); aiter.Next()) {
-        const auto &arc = aiter.Value();
-        auto it = nonterminal_hash_.find(arc.olabel);
-        if (it == nonterminal_hash_.end() || depscc_[it->second] != depscc) {
-          continue;  // Skips if a terminal or a non-terminal not in SCC.
-        }
-        const bool arc_in_cycle = fstscc[s] == fstscc[arc.nextstate];
-        // Left linear iff all non-terminals are initial.
-        if (s != fst->Start() || arc_in_cycle) {
-          depsccprops_[depscc] &= ~kReplaceSCCLeftLinear;
-        }
-        // Right linear iff all non-terminals are final.
-        if (fst->Final(arc.nextstate) == Weight::Zero() || arc_in_cycle) {
-          depsccprops_[depscc] &= ~kReplaceSCCRightLinear;
-        }
-      }
-    }
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_REPLACE_UTIL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/replace.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/replace.h
deleted file mode 100644
index 54a65627..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/replace.h
+++ /dev/null
@@ -1,1470 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes for the recursive replacement of FSTs.
-
-#ifndef FST_REPLACE_H_
-#define FST_REPLACE_H_
-
-#include <set>
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/expanded-fst.h>
-#include <fst/fst-decl.h>  // For optional argument declarations.
-#include <fst/fst.h>
-#include <fst/matcher.h>
-#include <fst/replace-util.h>
-#include <fst/state-table.h>
-#include <fst/test-properties.h>
-
-namespace fst {
-
-// Replace state tables have the form:
-//
-// template <class Arc, class P>
-// class ReplaceStateTable {
-//  public:
-//   using Label = typename Arc::Label Label;
-//   using StateId = typename Arc::StateId;
-//
-//   using PrefixId = P;
-//   using StateTuple = ReplaceStateTuple<StateId, PrefixId>;
-//   using StackPrefix = ReplaceStackPrefix<Label, StateId>;
-//
-//   // Required constructor.
-//   ReplaceStateTable(
-//       const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_list,
-//       Label root);
-//
-//   // Required copy constructor that does not copy state.
-//   ReplaceStateTable(const ReplaceStateTable<Arc, PrefixId> &table);
-//
-//   // Looks up state ID by tuple, adding it if it doesn't exist.
-//   StateId FindState(const StateTuple &tuple);
-//
-//   // Looks up state tuple by ID.
-//   const StateTuple &Tuple(StateId id) const;
-//
-//   // Lookus up prefix ID by stack prefix, adding it if it doesn't exist.
-//   PrefixId FindPrefixId(const StackPrefix &stack_prefix);
-//
-//  // Looks up stack prefix by ID.
-//  const StackPrefix &GetStackPrefix(PrefixId id) const;
-// };
-
-// Tuple that uniquely defines a state in replace.
-template <class S, class P>
-struct ReplaceStateTuple {
-  using StateId = S;
-  using PrefixId = P;
-
-  ReplaceStateTuple(PrefixId prefix_id = -1, StateId fst_id = kNoStateId,
-                    StateId fst_state = kNoStateId)
-      : prefix_id(prefix_id), fst_id(fst_id), fst_state(fst_state) {}
-
-  PrefixId prefix_id;  // Index in prefix table.
-  StateId fst_id;      // Current FST being walked.
-  StateId fst_state;   // Current state in FST being walked (not to be
-                       // confused with the thse StateId of the combined FST).
-};
-
-// Equality of replace state tuples.
-template <class StateId, class PrefixId>
-inline bool operator==(const ReplaceStateTuple<StateId, PrefixId> &x,
-                       const ReplaceStateTuple<StateId, PrefixId> &y) {
-  return x.prefix_id == y.prefix_id && x.fst_id == y.fst_id &&
-         x.fst_state == y.fst_state;
-}
-
-// Functor returning true for tuples corresponding to states in the root FST.
-template <class StateId, class PrefixId>
-class ReplaceRootSelector {
- public:
-  bool operator()(const ReplaceStateTuple<StateId, PrefixId> &tuple) const {
-    return tuple.prefix_id == 0;
-  }
-};
-
-// Functor for fingerprinting replace state tuples.
-template <class StateId, class PrefixId>
-class ReplaceFingerprint {
- public:
-  explicit ReplaceFingerprint(const std::vector<uint64> *size_array)
-      : size_array_(size_array) {}
-
-  uint64 operator()(const ReplaceStateTuple<StateId, PrefixId> &tuple) const {
-    return tuple.prefix_id * size_array_->back() +
-           size_array_->at(tuple.fst_id - 1) + tuple.fst_state;
-  }
-
- private:
-  const std::vector<uint64> *size_array_;
-};
-
-// Useful when the fst_state uniquely define the tuple.
-template <class StateId, class PrefixId>
-class ReplaceFstStateFingerprint {
- public:
-  uint64 operator()(const ReplaceStateTuple<StateId, PrefixId> &tuple) const {
-    return tuple.fst_state;
-  }
-};
-
-// A generic hash function for replace state tuples.
-template <typename S, typename P>
-class ReplaceHash {
- public:
-  size_t operator()(const ReplaceStateTuple<S, P>& t) const {
-    static constexpr auto prime0 = 7853;
-    static constexpr auto prime1 = 7867;
-    return t.prefix_id + t.fst_id * prime0 + t.fst_state * prime1;
-  }
-};
-
-// Container for stack prefix.
-template <class Label, class StateId>
-class ReplaceStackPrefix {
- public:
-  struct PrefixTuple {
-    PrefixTuple(Label fst_id = kNoLabel, StateId nextstate = kNoStateId)
-        : fst_id(fst_id), nextstate(nextstate) {}
-
-    Label fst_id;
-    StateId nextstate;
-  };
-
-  ReplaceStackPrefix() {}
-
-  ReplaceStackPrefix(const ReplaceStackPrefix &other)
-      : prefix_(other.prefix_) {}
-
-  void Push(StateId fst_id, StateId nextstate) {
-    prefix_.push_back(PrefixTuple(fst_id, nextstate));
-  }
-
-  void Pop() { prefix_.pop_back(); }
-
-  const PrefixTuple &Top() const { return prefix_[prefix_.size() - 1]; }
-
-  size_t Depth() const { return prefix_.size(); }
-
- public:
-  std::vector<PrefixTuple> prefix_;
-};
-
-// Equality stack prefix classes.
-template <class Label, class StateId>
-inline bool operator==(const ReplaceStackPrefix<Label, StateId> &x,
-                       const ReplaceStackPrefix<Label, StateId> &y) {
-  if (x.prefix_.size() != y.prefix_.size()) return false;
-  for (size_t i = 0; i < x.prefix_.size(); ++i) {
-    if (x.prefix_[i].fst_id != y.prefix_[i].fst_id ||
-        x.prefix_[i].nextstate != y.prefix_[i].nextstate) {
-      return false;
-    }
-  }
-  return true;
-}
-
-// Hash function for stack prefix to prefix id.
-template <class Label, class StateId>
-class ReplaceStackPrefixHash {
- public:
-  size_t operator()(const ReplaceStackPrefix<Label, StateId> &prefix) const {
-    size_t sum = 0;
-    for (const auto &pair : prefix.prefix_) {
-      static constexpr auto prime = 7863;
-      sum += pair.fst_id + pair.nextstate * prime;
-    }
-    return sum;
-  }
-};
-
-// Replace state tables.
-
-// A two-level state table for replace. Warning: calls CountStates to compute
-// the number of states of each component FST.
-template <class Arc, class P = ssize_t>
-class VectorHashReplaceStateTable {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-
-  using PrefixId = P;
-
-  using StateTuple = ReplaceStateTuple<StateId, PrefixId>;
-  using StateTable =
-      VectorHashStateTable<ReplaceStateTuple<StateId, PrefixId>,
-                           ReplaceRootSelector<StateId, PrefixId>,
-                           ReplaceFstStateFingerprint<StateId, PrefixId>,
-                           ReplaceFingerprint<StateId, PrefixId>>;
-  using StackPrefix = ReplaceStackPrefix<Label, StateId>;
-  using StackPrefixTable =
-      CompactHashBiTable<PrefixId, StackPrefix,
-                         ReplaceStackPrefixHash<Label, StateId>>;
-
-  VectorHashReplaceStateTable(
-      const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_list,
-      Label root)
-      : root_size_(0) {
-    size_array_.push_back(0);
-    for (const auto &fst_pair : fst_list) {
-      if (fst_pair.first == root) {
-        root_size_ = CountStates(*(fst_pair.second));
-        size_array_.push_back(size_array_.back());
-      } else {
-        size_array_.push_back(size_array_.back() +
-                              CountStates(*(fst_pair.second)));
-      }
-    }
-    state_table_.reset(
-        new StateTable(new ReplaceRootSelector<StateId, PrefixId>,
-                       new ReplaceFstStateFingerprint<StateId, PrefixId>,
-                       new ReplaceFingerprint<StateId, PrefixId>(&size_array_),
-                       root_size_, root_size_ + size_array_.back()));
-  }
-
-  VectorHashReplaceStateTable(
-      const VectorHashReplaceStateTable<Arc, PrefixId> &table)
-      : root_size_(table.root_size_),
-        size_array_(table.size_array_),
-        prefix_table_(table.prefix_table_) {
-    state_table_.reset(
-        new StateTable(new ReplaceRootSelector<StateId, PrefixId>,
-                       new ReplaceFstStateFingerprint<StateId, PrefixId>,
-                       new ReplaceFingerprint<StateId, PrefixId>(&size_array_),
-                       root_size_, root_size_ + size_array_.back()));
-  }
-
-  StateId FindState(const StateTuple &tuple) {
-    return state_table_->FindState(tuple);
-  }
-
-  const StateTuple &Tuple(StateId id) const { return state_table_->Tuple(id); }
-
-  PrefixId FindPrefixId(const StackPrefix &prefix) {
-    return prefix_table_.FindId(prefix);
-  }
-
-  const StackPrefix& GetStackPrefix(PrefixId id) const {
-    return prefix_table_.FindEntry(id);
-  }
-
- private:
-  StateId root_size_;
-  std::vector<uint64> size_array_;
-  std::unique_ptr<StateTable> state_table_;
-  StackPrefixTable prefix_table_;
-};
-
-// Default replace state table.
-template <class Arc, class P /* = size_t */>
-class DefaultReplaceStateTable
-    : public CompactHashStateTable<ReplaceStateTuple<typename Arc::StateId, P>,
-                                   ReplaceHash<typename Arc::StateId, P>> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-
-  using PrefixId = P;
-  using StateTuple = ReplaceStateTuple<StateId, PrefixId>;
-  using StateTable =
-      CompactHashStateTable<StateTuple, ReplaceHash<StateId, PrefixId>>;
-  using StackPrefix = ReplaceStackPrefix<Label, StateId>;
-  using StackPrefixTable =
-      CompactHashBiTable<PrefixId, StackPrefix,
-                         ReplaceStackPrefixHash<Label, StateId>>;
-
-  using StateTable::FindState;
-  using StateTable::Tuple;
-
-  DefaultReplaceStateTable(
-      const std::vector<std::pair<Label, const Fst<Arc> *>> &, Label) {}
-
-  DefaultReplaceStateTable(const DefaultReplaceStateTable<Arc, PrefixId> &table)
-      : StateTable(), prefix_table_(table.prefix_table_) {}
-
-  PrefixId FindPrefixId(const StackPrefix &prefix) {
-    return prefix_table_.FindId(prefix);
-  }
-
-  const StackPrefix &GetStackPrefix(PrefixId id) const {
-    return prefix_table_.FindEntry(id);
-  }
-
- private:
-  StackPrefixTable prefix_table_;
-};
-
-// By default ReplaceFst will copy the input label of the replace arc.
-// The call_label_type and return_label_type options specify how to manage
-// the labels of the call arc and the return arc of the replace FST
-template <class Arc, class StateTable = DefaultReplaceStateTable<Arc>,
-          class CacheStore = DefaultCacheStore<Arc>>
-struct ReplaceFstOptions : CacheImplOptions<CacheStore> {
-  using Label = typename Arc::Label;
-
-  // Index of root rule for expansion.
-  Label root;
-  // How to label call arc.
-  ReplaceLabelType call_label_type = REPLACE_LABEL_INPUT;
-  // How to label return arc.
-  ReplaceLabelType return_label_type = REPLACE_LABEL_NEITHER;
-  // Specifies output label to put on call arc; if kNoLabel, use existing label
-  // on call arc. Otherwise, use this field as the output label.
-  Label call_output_label = kNoLabel;
-  // Specifies label to put on return arc.
-  Label return_label = 0;
-  // Take ownership of input FSTs?
-  bool take_ownership = false;
-  // Pointer to optional pre-constructed state table.
-  StateTable *state_table = nullptr;
-
-  explicit ReplaceFstOptions(const CacheImplOptions<CacheStore> &opts,
-                             Label root = kNoLabel)
-      : CacheImplOptions<CacheStore>(opts), root(root) {}
-
-  explicit ReplaceFstOptions(const CacheOptions &opts, Label root = kNoLabel)
-      : CacheImplOptions<CacheStore>(opts), root(root) {}
-
-  // FIXME(kbg): There are too many constructors here. Come up with a consistent
-  // position for call_output_label (probably the very end) so that it is
-  // possible to express all the remaining constructors with a single
-  // default-argument constructor. Also move clients off of the "backwards
-  // compatibility" constructor, for good.
-
-  explicit ReplaceFstOptions(Label root) : root(root) {}
-
-  explicit ReplaceFstOptions(Label root, ReplaceLabelType call_label_type,
-                             ReplaceLabelType return_label_type,
-                             Label return_label)
-      : root(root),
-        call_label_type(call_label_type),
-        return_label_type(return_label_type),
-        return_label(return_label) {}
-
-  explicit ReplaceFstOptions(Label root, ReplaceLabelType call_label_type,
-                             ReplaceLabelType return_label_type,
-                             Label call_output_label, Label return_label)
-      : root(root),
-        call_label_type(call_label_type),
-        return_label_type(return_label_type),
-        call_output_label(call_output_label),
-        return_label(return_label) {}
-
-  explicit ReplaceFstOptions(const ReplaceUtilOptions &opts)
-      : ReplaceFstOptions(opts.root, opts.call_label_type,
-                          opts.return_label_type, opts.return_label) {}
-
-  ReplaceFstOptions() : root(kNoLabel) {}
-
-  // For backwards compatibility.
-  ReplaceFstOptions(int64 root, bool epsilon_replace_arc)
-      : root(root),
-        call_label_type(epsilon_replace_arc ? REPLACE_LABEL_NEITHER
-                                            : REPLACE_LABEL_INPUT),
-        call_output_label(epsilon_replace_arc ? 0 : kNoLabel) {}
-};
-
-// Forward declaration.
-template <class Arc, class StateTable, class CacheStore>
-class ReplaceFstMatcher;
-
-template <class Arc>
-using FstList = std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>;
-
-// Returns true if label type on arc results in epsilon input label.
-inline bool EpsilonOnInput(ReplaceLabelType label_type) {
-  return label_type == REPLACE_LABEL_NEITHER ||
-         label_type == REPLACE_LABEL_OUTPUT;
-}
-
-// Returns true if label type on arc results in epsilon input label.
-inline bool EpsilonOnOutput(ReplaceLabelType label_type) {
-  return label_type == REPLACE_LABEL_NEITHER ||
-         label_type == REPLACE_LABEL_INPUT;
-}
-
-// Returns true if for either the call or return arc ilabel != olabel.
-template <class Label>
-bool ReplaceTransducer(ReplaceLabelType call_label_type,
-                       ReplaceLabelType return_label_type,
-                       Label call_output_label) {
-  return call_label_type == REPLACE_LABEL_INPUT ||
-         call_label_type == REPLACE_LABEL_OUTPUT ||
-         (call_label_type == REPLACE_LABEL_BOTH &&
-          call_output_label != kNoLabel) ||
-         return_label_type == REPLACE_LABEL_INPUT ||
-         return_label_type == REPLACE_LABEL_OUTPUT;
-}
-
-template <class Arc>
-uint64 ReplaceFstProperties(typename Arc::Label root_label,
-                            const FstList<Arc> &fst_list,
-                            ReplaceLabelType call_label_type,
-                            ReplaceLabelType return_label_type,
-                            typename Arc::Label call_output_label,
-                            bool *sorted_and_non_empty) {
-  using Label = typename Arc::Label;
-  std::vector<uint64> inprops;
-  bool all_ilabel_sorted = true;
-  bool all_olabel_sorted = true;
-  bool all_non_empty = true;
-  // All nonterminals are negative?
-  bool all_negative = true;
-  // All nonterminals are positive and form a dense range containing 1?
-  bool dense_range = true;
-  Label root_fst_idx = 0;
-  for (Label i = 0; i < fst_list.size(); ++i) {
-    const auto label = fst_list[i].first;
-    if (label >= 0) all_negative = false;
-    if (label > fst_list.size() || label <= 0) dense_range = false;
-    if (label == root_label) root_fst_idx = i;
-    const auto *fst = fst_list[i].second;
-    if (fst->Start() == kNoStateId) all_non_empty = false;
-    if (!fst->Properties(kILabelSorted, false)) all_ilabel_sorted = false;
-    if (!fst->Properties(kOLabelSorted, false)) all_olabel_sorted = false;
-    inprops.push_back(fst->Properties(kCopyProperties, false));
-  }
-  const auto props = ReplaceProperties(
-      inprops, root_fst_idx, EpsilonOnInput(call_label_type),
-      EpsilonOnInput(return_label_type), EpsilonOnOutput(call_label_type),
-      EpsilonOnOutput(return_label_type),
-      ReplaceTransducer(call_label_type, return_label_type, call_output_label),
-      all_non_empty, all_ilabel_sorted, all_olabel_sorted,
-      all_negative || dense_range);
-  const bool sorted = props & (kILabelSorted | kOLabelSorted);
-  *sorted_and_non_empty = all_non_empty && sorted;
-  return props;
-}
-
-namespace internal {
-
-// The replace implementation class supports a dynamic expansion of a recursive
-// transition network represented as label/FST pairs with dynamic replacable
-// arcs.
-template <class Arc, class StateTable, class CacheStore>
-class ReplaceFstImpl
-    : public CacheBaseImpl<typename CacheStore::State, CacheStore> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using State = typename CacheStore::State;
-  using CacheImpl = CacheBaseImpl<State, CacheStore>;
-  using PrefixId = typename StateTable::PrefixId;
-  using StateTuple = ReplaceStateTuple<StateId, PrefixId>;
-  using StackPrefix = ReplaceStackPrefix<Label, StateId>;
-  using NonTerminalHash = std::unordered_map<Label, Label>;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::WriteHeader;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-  using FstImpl<Arc>::InputSymbols;
-  using FstImpl<Arc>::OutputSymbols;
-
-  using CacheImpl::PushArc;
-  using CacheImpl::HasArcs;
-  using CacheImpl::HasFinal;
-  using CacheImpl::HasStart;
-  using CacheImpl::SetArcs;
-  using CacheImpl::SetFinal;
-  using CacheImpl::SetStart;
-
-  friend class ReplaceFstMatcher<Arc, StateTable, CacheStore>;
-
-  ReplaceFstImpl(const FstList<Arc> &fst_list,
-                 const ReplaceFstOptions<Arc, StateTable, CacheStore> &opts)
-      : CacheImpl(opts),
-        call_label_type_(opts.call_label_type),
-        return_label_type_(opts.return_label_type),
-        call_output_label_(opts.call_output_label),
-        return_label_(opts.return_label),
-        state_table_(opts.state_table ? opts.state_table
-                                      : new StateTable(fst_list, opts.root)) {
-    SetType("replace");
-    // If the label is epsilon, then all replace label options are equivalent,
-    // so we set the label types to NEITHER for simplicity.
-    if (call_output_label_ == 0) call_label_type_ = REPLACE_LABEL_NEITHER;
-    if (return_label_ == 0) return_label_type_ = REPLACE_LABEL_NEITHER;
-    if (!fst_list.empty()) {
-      SetInputSymbols(fst_list[0].second->InputSymbols());
-      SetOutputSymbols(fst_list[0].second->OutputSymbols());
-    }
-    fst_array_.push_back(nullptr);
-    for (Label i = 0; i < fst_list.size(); ++i) {
-      const auto label = fst_list[i].first;
-      const auto *fst = fst_list[i].second;
-      nonterminal_hash_[label] = fst_array_.size();
-      nonterminal_set_.insert(label);
-      fst_array_.emplace_back(opts.take_ownership ? fst : fst->Copy());
-      if (i) {
-        if (!CompatSymbols(InputSymbols(), fst->InputSymbols())) {
-          FSTERROR() << "ReplaceFstImpl: Input symbols of FST " << i
-                     << " do not match input symbols of base FST (0th FST)";
-          SetProperties(kError, kError);
-        }
-        if (!CompatSymbols(OutputSymbols(), fst->OutputSymbols())) {
-          FSTERROR() << "ReplaceFstImpl: Output symbols of FST " << i
-                     << " do not match output symbols of base FST (0th FST)";
-          SetProperties(kError, kError);
-        }
-      }
-    }
-    const auto nonterminal = nonterminal_hash_[opts.root];
-    if ((nonterminal == 0) && (fst_array_.size() > 1)) {
-      FSTERROR() << "ReplaceFstImpl: No FST corresponding to root label "
-                 << opts.root << " in the input tuple vector";
-      SetProperties(kError, kError);
-    }
-    root_ = (nonterminal > 0) ? nonterminal : 1;
-    bool all_non_empty_and_sorted = false;
-    SetProperties(ReplaceFstProperties(opts.root, fst_list, call_label_type_,
-                                       return_label_type_, call_output_label_,
-                                       &all_non_empty_and_sorted));
-    // Enables optional caching as long as sorted and all non-empty.
-    always_cache_ = !all_non_empty_and_sorted;
-    VLOG(2) << "ReplaceFstImpl::ReplaceFstImpl: always_cache = "
-            << (always_cache_ ? "true" : "false");
-  }
-
-  ReplaceFstImpl(const ReplaceFstImpl &impl)
-      : CacheImpl(impl),
-        call_label_type_(impl.call_label_type_),
-        return_label_type_(impl.return_label_type_),
-        call_output_label_(impl.call_output_label_),
-        return_label_(impl.return_label_),
-        always_cache_(impl.always_cache_),
-        state_table_(new StateTable(*(impl.state_table_))),
-        nonterminal_set_(impl.nonterminal_set_),
-        nonterminal_hash_(impl.nonterminal_hash_),
-        root_(impl.root_) {
-    SetType("replace");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-    fst_array_.reserve(impl.fst_array_.size());
-    fst_array_.emplace_back(nullptr);
-    for (Label i = 1; i < impl.fst_array_.size(); ++i) {
-      fst_array_.emplace_back(impl.fst_array_[i]->Copy(true));
-    }
-  }
-
-  // Computes the dependency graph of the replace class and returns
-  // true if the dependencies are cyclic. Cyclic dependencies will result
-  // in an un-expandable FST.
-  bool CyclicDependencies() const {
-    const ReplaceUtilOptions opts(root_);
-    ReplaceUtil<Arc> replace_util(fst_array_, nonterminal_hash_, opts);
-    return replace_util.CyclicDependencies();
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      if (fst_array_.size() == 1) {
-        SetStart(kNoStateId);
-        return kNoStateId;
-      } else {
-        const auto fst_start = fst_array_[root_]->Start();
-        if (fst_start == kNoStateId) return kNoStateId;
-        const auto prefix = GetPrefixId(StackPrefix());
-        const auto start =
-            state_table_->FindState(StateTuple(prefix, root_, fst_start));
-        SetStart(start);
-        return start;
-      }
-    } else {
-      return CacheImpl::Start();
-    }
-  }
-
-  Weight Final(StateId s) {
-    if (HasFinal(s)) return CacheImpl::Final(s);
-    const auto &tuple = state_table_->Tuple(s);
-    auto weight = Weight::Zero();
-    if (tuple.prefix_id == 0) {
-      const auto fst_state = tuple.fst_state;
-      weight = fst_array_[tuple.fst_id]->Final(fst_state);
-    }
-    if (always_cache_ || HasArcs(s)) SetFinal(s, weight);
-    return weight;
-  }
-
-  size_t NumArcs(StateId s) {
-    if (HasArcs(s)) {
-      return CacheImpl::NumArcs(s);
-    } else if (always_cache_) {  // If always caching, expands and caches state.
-      Expand(s);
-      return CacheImpl::NumArcs(s);
-    } else {  // Otherwise computes the number of arcs without expanding.
-      const auto tuple = state_table_->Tuple(s);
-      if (tuple.fst_state == kNoStateId) return 0;
-      auto num_arcs = fst_array_[tuple.fst_id]->NumArcs(tuple.fst_state);
-      if (ComputeFinalArc(tuple, nullptr)) ++num_arcs;
-      return num_arcs;
-    }
-  }
-
-  // Returns whether a given label is a non-terminal.
-  bool IsNonTerminal(Label label) const {
-    if (label < *nonterminal_set_.begin() ||
-        label > *nonterminal_set_.rbegin()) {
-      return false;
-    } else {
-      return nonterminal_hash_.count(label);
-    }
-    // TODO(allauzen): be smarter and take advantage of all_dense or
-    // all_negative. Also use this in ComputeArc. This would require changes to
-    // Replace so that recursing into an empty FST lead to a non co-accessible
-    // state instead of deleting the arc as done currently. The current use
-    // correct, since labels are sorted if all_non_empty is true.
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (HasArcs(s)) {
-      return CacheImpl::NumInputEpsilons(s);
-    } else if (always_cache_ || !Properties(kILabelSorted)) {
-      // If always caching or if the number of input epsilons is too expensive
-      // to compute without caching (i.e., not ilabel-sorted), then expands and
-      // caches state.
-      Expand(s);
-      return CacheImpl::NumInputEpsilons(s);
-    } else {
-      // Otherwise, computes the number of input epsilons without caching.
-      const auto tuple = state_table_->Tuple(s);
-      if (tuple.fst_state == kNoStateId) return 0;
-      size_t num = 0;
-      if (!EpsilonOnInput(call_label_type_)) {
-        // If EpsilonOnInput(c) is false, all input epsilon arcs
-        // are also input epsilons arcs in the underlying machine.
-        num = fst_array_[tuple.fst_id]->NumInputEpsilons(tuple.fst_state);
-      } else {
-        // Otherwise, one need to consider that all non-terminal arcs
-        // in the underlying machine also become input epsilon arc.
-        ArcIterator<Fst<Arc>> aiter(*fst_array_[tuple.fst_id], tuple.fst_state);
-        for (; !aiter.Done() && ((aiter.Value().ilabel == 0) ||
-                                 IsNonTerminal(aiter.Value().olabel));
-             aiter.Next()) {
-          ++num;
-        }
-      }
-      if (EpsilonOnInput(return_label_type_) &&
-          ComputeFinalArc(tuple, nullptr)) {
-        ++num;
-      }
-      return num;
-    }
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (HasArcs(s)) {
-      return CacheImpl::NumOutputEpsilons(s);
-    } else if (always_cache_ || !Properties(kOLabelSorted)) {
-      // If always caching or if the number of output epsilons is too expensive
-      // to compute without caching (i.e., not olabel-sorted), then expands and
-      // caches state.
-      Expand(s);
-      return CacheImpl::NumOutputEpsilons(s);
-    } else {
-      // Otherwise, computes the number of output epsilons without caching.
-      const auto tuple = state_table_->Tuple(s);
-      if (tuple.fst_state == kNoStateId) return 0;
-      size_t num = 0;
-      if (!EpsilonOnOutput(call_label_type_)) {
-        // If EpsilonOnOutput(c) is false, all output epsilon arcs are also
-        // output epsilons arcs in the underlying machine.
-        num = fst_array_[tuple.fst_id]->NumOutputEpsilons(tuple.fst_state);
-      } else {
-        // Otherwise, one need to consider that all non-terminal arcs in the
-        // underlying machine also become output epsilon arc.
-        ArcIterator<Fst<Arc>> aiter(*fst_array_[tuple.fst_id], tuple.fst_state);
-        for (; !aiter.Done() && ((aiter.Value().olabel == 0) ||
-                                 IsNonTerminal(aiter.Value().olabel));
-             aiter.Next()) {
-          ++num;
-        }
-      }
-      if (EpsilonOnOutput(return_label_type_) &&
-          ComputeFinalArc(tuple, nullptr)) {
-        ++num;
-      }
-      return num;
-    }
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if (mask & kError) {
-      for (Label i = 1; i < fst_array_.size(); ++i) {
-        if (fst_array_[i]->Properties(kError, false)) {
-          SetProperties(kError, kError);
-        }
-      }
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  // Returns the base arc iterator, and if arcs have not been computed yet,
-  // extends and recurses for new arcs.
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl::InitArcIterator(s, data);
-    // TODO(allauzen): Set behaviour of generic iterator.
-    // Warning: ArcIterator<ReplaceFst<A>>::InitCache() relies on current
-    // behaviour.
-  }
-
-  // Extends current state (walk arcs one level deep).
-  void Expand(StateId s) {
-    const auto tuple = state_table_->Tuple(s);
-    if (tuple.fst_state == kNoStateId) {  // Local FST is empty.
-      SetArcs(s);
-      return;
-    }
-    ArcIterator<Fst<Arc>> aiter(*fst_array_[tuple.fst_id], tuple.fst_state);
-    Arc arc;
-    // Creates a final arc when needed.
-    if (ComputeFinalArc(tuple, &arc)) PushArc(s, arc);
-    // Expands all arcs leaving the state.
-    for (; !aiter.Done(); aiter.Next()) {
-      if (ComputeArc(tuple, aiter.Value(), &arc)) PushArc(s, arc);
-    }
-    SetArcs(s);
-  }
-
-  void Expand(StateId s, const StateTuple &tuple,
-              const ArcIteratorData<Arc> &data) {
-    if (tuple.fst_state == kNoStateId) {  // Local FST is empty.
-      SetArcs(s);
-      return;
-    }
-    ArcIterator<Fst<Arc>> aiter(data);
-    Arc arc;
-    // Creates a final arc when needed.
-    if (ComputeFinalArc(tuple, &arc)) AddArc(s, arc);
-    // Expands all arcs leaving the state.
-    for (; !aiter.Done(); aiter.Next()) {
-      if (ComputeArc(tuple, aiter.Value(), &arc)) AddArc(s, arc);
-    }
-    SetArcs(s);
-  }
-
-  // If acpp is null, only returns true if a final arcp is required, but does
-  // not actually compute it.
-  bool ComputeFinalArc(const StateTuple &tuple, Arc *arcp,
-                       uint32 flags = kArcValueFlags) {
-    const auto fst_state = tuple.fst_state;
-    if (fst_state == kNoStateId) return false;
-    // If state is final, pops the stack.
-    if (fst_array_[tuple.fst_id]->Final(fst_state) != Weight::Zero() &&
-        tuple.prefix_id) {
-      if (arcp) {
-        arcp->ilabel = (EpsilonOnInput(return_label_type_)) ? 0 : return_label_;
-        arcp->olabel =
-            (EpsilonOnOutput(return_label_type_)) ? 0 : return_label_;
-        if (flags & kArcNextStateValue) {
-          const auto &stack = state_table_->GetStackPrefix(tuple.prefix_id);
-          const auto prefix_id = PopPrefix(stack);
-          const auto &top = stack.Top();
-          arcp->nextstate = state_table_->FindState(
-              StateTuple(prefix_id, top.fst_id, top.nextstate));
-        }
-        if (flags & kArcWeightValue) {
-          arcp->weight = fst_array_[tuple.fst_id]->Final(fst_state);
-        }
-      }
-      return true;
-    } else {
-      return false;
-    }
-  }
-
-  // Computes an arc in the FST corresponding to one in the underlying machine.
-  // Returns false if the underlying arc corresponds to no arc in the resulting
-  // FST.
-  bool ComputeArc(const StateTuple &tuple, const Arc &arc, Arc *arcp,
-                  uint32 flags = kArcValueFlags) {
-    if (!EpsilonOnInput(call_label_type_) &&
-        (flags == (flags & (kArcILabelValue | kArcWeightValue)))) {
-      *arcp = arc;
-      return true;
-    }
-    if (arc.olabel == 0 || arc.olabel < *nonterminal_set_.begin() ||
-        arc.olabel > *nonterminal_set_.rbegin()) {  // Expands local FST.
-      const auto nextstate =
-          flags & kArcNextStateValue
-              ? state_table_->FindState(
-                    StateTuple(tuple.prefix_id, tuple.fst_id, arc.nextstate))
-              : kNoStateId;
-      *arcp = Arc(arc.ilabel, arc.olabel, arc.weight, nextstate);
-    } else {
-      // Checks for non-terminal.
-      const auto it = nonterminal_hash_.find(arc.olabel);
-      if (it != nonterminal_hash_.end()) {  // Recurses into non-terminal.
-        const auto nonterminal = it->second;
-        const auto nt_prefix =
-            PushPrefix(state_table_->GetStackPrefix(tuple.prefix_id),
-                       tuple.fst_id, arc.nextstate);
-        // If the start state is valid, replace; othewise, the arc is implicitly
-        // deleted.
-        const auto nt_start = fst_array_[nonterminal]->Start();
-        if (nt_start != kNoStateId) {
-          const auto nt_nextstate = flags & kArcNextStateValue
-                                        ? state_table_->FindState(StateTuple(
-                                              nt_prefix, nonterminal, nt_start))
-                                        : kNoStateId;
-          const auto ilabel =
-              (EpsilonOnInput(call_label_type_)) ? 0 : arc.ilabel;
-          const auto olabel =
-              (EpsilonOnOutput(call_label_type_))
-                  ? 0
-                  : ((call_output_label_ == kNoLabel) ? arc.olabel
-                                                      : call_output_label_);
-          *arcp = Arc(ilabel, olabel, arc.weight, nt_nextstate);
-        } else {
-          return false;
-        }
-      } else {
-        const auto nextstate =
-            flags & kArcNextStateValue
-                ? state_table_->FindState(
-                      StateTuple(tuple.prefix_id, tuple.fst_id, arc.nextstate))
-                : kNoStateId;
-        *arcp = Arc(arc.ilabel, arc.olabel, arc.weight, nextstate);
-      }
-    }
-    return true;
-  }
-
-  // Returns the arc iterator flags supported by this FST.
-  uint32 ArcIteratorFlags() const {
-    uint32 flags = kArcValueFlags;
-    if (!always_cache_) flags |= kArcNoCache;
-    return flags;
-  }
-
-  StateTable *GetStateTable() const { return state_table_.get(); }
-
-  const Fst<Arc> *GetFst(Label fst_id) const {
-    return fst_array_[fst_id].get();
-  }
-
-  Label GetFstId(Label nonterminal) const {
-    const auto it = nonterminal_hash_.find(nonterminal);
-    if (it == nonterminal_hash_.end()) {
-      FSTERROR() << "ReplaceFstImpl::GetFstId: Nonterminal not found: "
-                 << nonterminal;
-    }
-    return it->second;
-  }
-
-  // Returns true if label type on call arc results in epsilon input label.
-  bool EpsilonOnCallInput() { return EpsilonOnInput(call_label_type_); }
-
- private:
-  // The unique index into stack prefix table.
-  PrefixId GetPrefixId(const StackPrefix &prefix) {
-    return state_table_->FindPrefixId(prefix);
-  }
-
-  // The prefix ID after a stack pop.
-  PrefixId PopPrefix(StackPrefix prefix) {
-    prefix.Pop();
-    return GetPrefixId(prefix);
-  }
-
-  // The prefix ID after a stack push.
-  PrefixId PushPrefix(StackPrefix prefix, Label fst_id, StateId nextstate) {
-    prefix.Push(fst_id, nextstate);
-    return GetPrefixId(prefix);
-  }
-
-  // Runtime options
-  ReplaceLabelType call_label_type_;    // How to label call arc.
-  ReplaceLabelType return_label_type_;  // How to label return arc.
-  int64 call_output_label_;  // Specifies output label to put on call arc
-  int64 return_label_;       // Specifies label to put on return arc.
-  bool always_cache_;        // Disable optional caching of arc iterator?
-
-  // State table.
-  std::unique_ptr<StateTable> state_table_;
-
-  // Replace components.
-  std::set<Label> nonterminal_set_;
-  NonTerminalHash nonterminal_hash_;
-  std::vector<std::unique_ptr<const Fst<Arc>>> fst_array_;
-  Label root_;
-};
-
-}  // namespace internal
-
-//
-// ReplaceFst supports dynamic replacement of arcs in one FST with another FST.
-// This replacement is recursive. ReplaceFst can be used to support a variety of
-// delayed constructions such as recursive
-// transition networks, union, or closure. It is constructed with an array of
-// FST(s). One FST represents the root (or topology) machine. The root FST
-// refers to other FSTs by recursively replacing arcs labeled as non-terminals
-// with the matching non-terminal FST. Currently the ReplaceFst uses the output
-// symbols of the arcs to determine whether the arc is a non-terminal arc or
-// not. A non-terminal can be any label that is not a non-zero terminal label in
-// the output alphabet.
-//
-// Note that the constructor uses a vector of pairs. These correspond to the
-// tuple of non-terminal Label and corresponding FST. For example to implement
-// the closure operation we need 2 FSTs. The first root FST is a single
-// self-loop arc on the start state.
-//
-// The ReplaceFst class supports an optionally caching arc iterator.
-//
-// The ReplaceFst needs to be built such that it is known to be ilabel- or
-// olabel-sorted (see usage below).
-//
-// Observe that Matcher<Fst<A>> will use the optionally caching arc iterator
-// when available (the FST is ilabel-sorted and matching on the input, or the
-// FST is olabel -orted and matching on the output).  In order to obtain the
-// most efficient behaviour, it is recommended to set call_label_type to
-// REPLACE_LABEL_INPUT or REPLACE_LABEL_BOTH and return_label_type to
-// REPLACE_LABEL_OUTPUT or REPLACE_LABEL_NEITHER. This means that the call arc
-// does not have epsilon on the input side and the return arc has epsilon on the
-// input side) and matching on the input side.
-//
-// This class attaches interface to implementation and handles reference
-// counting, delegating most methods to ImplToFst.
-template <class A, class T /* = DefaultReplaceStateTable<A> */,
-          class CacheStore /* = DefaultCacheStore<A> */>
-class ReplaceFst
-    : public ImplToFst<internal::ReplaceFstImpl<A, T, CacheStore>> {
- public:
-  using Arc = A;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using StateTable = T;
-  using Store = CacheStore;
-  using State = typename CacheStore::State;
-  using Impl = internal::ReplaceFstImpl<Arc, StateTable, CacheStore>;
-  using CacheImpl = internal::CacheBaseImpl<State, CacheStore>;
-
-  using ImplToFst<Impl>::Properties;
-
-  friend class ArcIterator<ReplaceFst<Arc, StateTable, CacheStore>>;
-  friend class StateIterator<ReplaceFst<Arc, StateTable, CacheStore>>;
-  friend class ReplaceFstMatcher<Arc, StateTable, CacheStore>;
-
-  ReplaceFst(const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_array,
-             Label root)
-      : ImplToFst<Impl>(std::make_shared<Impl>(
-            fst_array, ReplaceFstOptions<Arc, StateTable, CacheStore>(root))) {}
-
-  ReplaceFst(const std::vector<std::pair<Label, const Fst<Arc> *>> &fst_array,
-             const ReplaceFstOptions<Arc, StateTable, CacheStore> &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst_array, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  ReplaceFst(const ReplaceFst<Arc, StateTable, CacheStore> &fst,
-             bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this ReplaceFst. See Fst<>::Copy() for further doc.
-  ReplaceFst<Arc, StateTable, CacheStore> *Copy(
-      bool safe = false) const override {
-    return new ReplaceFst<Arc, StateTable, CacheStore>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
-  MatcherBase<Arc> *InitMatcher(MatchType match_type) const override {
-    if ((GetImpl()->ArcIteratorFlags() & kArcNoCache) &&
-        ((match_type == MATCH_INPUT && Properties(kILabelSorted, false)) ||
-         (match_type == MATCH_OUTPUT && Properties(kOLabelSorted, false)))) {
-      return new ReplaceFstMatcher<Arc, StateTable, CacheStore>(*this,
-                                                                match_type);
-    } else {
-      VLOG(2) << "Not using replace matcher";
-      return nullptr;
-    }
-  }
-
-  bool CyclicDependencies() const { return GetImpl()->CyclicDependencies(); }
-
-  const StateTable &GetStateTable() const {
-    return *GetImpl()->GetStateTable();
-  }
-
-  const Fst<Arc> &GetFst(Label nonterminal) const {
-    return *GetImpl()->GetFst(GetImpl()->GetFstId(nonterminal));
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  ReplaceFst &operator=(const ReplaceFst &) = delete;
-};
-
-// Specialization for ReplaceFst.
-template <class Arc, class StateTable, class CacheStore>
-class StateIterator<ReplaceFst<Arc, StateTable, CacheStore>>
-    : public CacheStateIterator<ReplaceFst<Arc, StateTable, CacheStore>> {
- public:
-  explicit StateIterator(const ReplaceFst<Arc, StateTable, CacheStore> &fst)
-      : CacheStateIterator<ReplaceFst<Arc, StateTable, CacheStore>>(
-            fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for ReplaceFst, implementing optional caching. It is be used
-// as follows:
-//
-//   ReplaceFst<A> replace;
-//   ArcIterator<ReplaceFst<A>> aiter(replace, s);
-//   // Note: ArcIterator< Fst<A>> is always a caching arc iterator.
-//   aiter.SetFlags(kArcNoCache, kArcNoCache);
-//   // Uses the arc iterator, no arc will be cached, no state will be expanded.
-//   // Arc flags can be used to decide which component of the arc need to be
-//   computed.
-//   aiter.SetFlags(kArcILabelValue, kArcValueFlags);
-//   // Wants the ilabel for this arc.
-//   aiter.Value();  // Does not compute the destination state.
-//   aiter.Next();
-//   aiter.SetFlags(kArcNextStateValue, kArcNextStateValue);
-//   // Wants the ilabel and next state for this arc.
-//   aiter.Value();  // Does compute the destination state and inserts it
-//                   // in the replace state table.
-//   // No additional arcs have been cached at this point.
-template <class Arc, class StateTable, class CacheStore>
-class ArcIterator<ReplaceFst<Arc, StateTable, CacheStore>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  using StateTuple = typename StateTable::StateTuple;
-
-  ArcIterator(const ReplaceFst<Arc, StateTable, CacheStore> &fst, StateId s)
-      : fst_(fst),
-        s_(s),
-        pos_(0),
-        offset_(0),
-        flags_(kArcValueFlags),
-        arcs_(nullptr),
-        data_flags_(0),
-        final_flags_(0) {
-    cache_data_.ref_count = nullptr;
-    local_data_.ref_count = nullptr;
-    // If FST does not support optional caching, forces caching.
-    if (!(fst_.GetImpl()->ArcIteratorFlags() & kArcNoCache) &&
-        !(fst_.GetImpl()->HasArcs(s_))) {
-      fst_.GetMutableImpl()->Expand(s_);
-    }
-    // If state is already cached, use cached arcs array.
-    if (fst_.GetImpl()->HasArcs(s_)) {
-      (fst_.GetImpl())
-          ->internal::template CacheBaseImpl<
-              typename CacheStore::State,
-              CacheStore>::InitArcIterator(s_, &cache_data_);
-      num_arcs_ = cache_data_.narcs;
-      arcs_ = cache_data_.arcs;      // arcs_ is a pointer to the cached arcs.
-      data_flags_ = kArcValueFlags;  // All the arc member values are valid.
-    } else {  // Otherwise delay decision until Value() is called.
-      tuple_ = fst_.GetImpl()->GetStateTable()->Tuple(s_);
-      if (tuple_.fst_state == kNoStateId) {
-        num_arcs_ = 0;
-      } else {
-        // The decision to cache or not to cache has been defered until Value()
-        // or
-        // SetFlags() is called. However, the arc iterator is set up now to be
-        // ready for non-caching in order to keep the Value() method simple and
-        // efficient.
-        const auto *rfst = fst_.GetImpl()->GetFst(tuple_.fst_id);
-        rfst->InitArcIterator(tuple_.fst_state, &local_data_);
-        // arcs_ is a pointer to the arcs in the underlying machine.
-        arcs_ = local_data_.arcs;
-        // Computes the final arc (but not its destination state) if a final arc
-        // is required.
-        bool has_final_arc = fst_.GetMutableImpl()->ComputeFinalArc(
-            tuple_, &final_arc_, kArcValueFlags & ~kArcNextStateValue);
-        // Sets the arc value flags that hold for final_arc_.
-        final_flags_ = kArcValueFlags & ~kArcNextStateValue;
-        // Computes the number of arcs.
-        num_arcs_ = local_data_.narcs;
-        if (has_final_arc) ++num_arcs_;
-        // Sets the offset between the underlying arc positions and the
-        // positions
-        // in the arc iterator.
-        offset_ = num_arcs_ - local_data_.narcs;
-        // Defers the decision to cache or not until Value() or SetFlags() is
-        // called.
-        data_flags_ = 0;
-      }
-    }
-  }
-
-  ~ArcIterator() {
-    if (cache_data_.ref_count) --(*cache_data_.ref_count);
-    if (local_data_.ref_count) --(*local_data_.ref_count);
-  }
-
-  void ExpandAndCache() const  {
-    // TODO(allauzen): revisit this.
-    // fst_.GetImpl()->Expand(s_, tuple_, local_data_);
-    // (fst_.GetImpl())->CacheImpl<A>*>::InitArcIterator(s_,
-    //                                               &cache_data_);
-    //
-    fst_.InitArcIterator(s_, &cache_data_);  // Expand and cache state.
-    arcs_ = cache_data_.arcs;      // arcs_ is a pointer to the cached arcs.
-    data_flags_ = kArcValueFlags;  // All the arc member values are valid.
-    offset_ = 0;                   // No offset.
-  }
-
-  void Init() {
-    if (flags_ & kArcNoCache) {  // If caching is disabled
-      // arcs_ is a pointer to the arcs in the underlying machine.
-      arcs_ = local_data_.arcs;
-      // Sets the arcs value flags that hold for arcs_.
-      data_flags_ = kArcWeightValue;
-      if (!fst_.GetMutableImpl()->EpsilonOnCallInput()) {
-        data_flags_ |= kArcILabelValue;
-      }
-      // Sets the offset between the underlying arc positions and the positions
-      // in the arc iterator.
-      offset_ = num_arcs_ - local_data_.narcs;
-    } else {
-      ExpandAndCache();
-    }
-  }
-
-  bool Done() const { return pos_ >= num_arcs_; }
-
-  const Arc &Value() const {
-    // If data_flags_ is 0, non-caching was not requested.
-    if (!data_flags_) {
-      // TODO(allauzen): Revisit this.
-      if (flags_ & kArcNoCache) {
-        // Should never happen.
-        FSTERROR() << "ReplaceFst: Inconsistent arc iterator flags";
-      }
-      ExpandAndCache();
-    }
-    if (pos_ - offset_ >= 0) {  // The requested arc is not the final arc.
-      const auto &arc = arcs_[pos_ - offset_];
-      if ((data_flags_ & flags_) == (flags_ & kArcValueFlags)) {
-        // If the value flags match the recquired value flags then returns the
-        // arc.
-        return arc;
-      } else {
-        // Otherwise, compute the corresponding arc on-the-fly.
-        fst_.GetMutableImpl()->ComputeArc(tuple_, arc, &arc_,
-                                          flags_ & kArcValueFlags);
-        return arc_;
-      }
-    } else {  // The requested arc is the final arc.
-      if ((final_flags_ & flags_) != (flags_ & kArcValueFlags)) {
-        // If the arc value flags that hold for the final arc do not match the
-        // requested value flags, then
-        // final_arc_ needs to be updated.
-        fst_.GetMutableImpl()->ComputeFinalArc(tuple_, &final_arc_,
-                                               flags_ & kArcValueFlags);
-        final_flags_ = flags_ & kArcValueFlags;
-      }
-      return final_arc_;
-    }
-  }
-
-  void Next() { ++pos_; }
-
-  size_t Position() const { return pos_; }
-
-  void Reset() { pos_ = 0; }
-
-  void Seek(size_t pos) { pos_ = pos; }
-
-  uint32 Flags() const { return flags_; }
-
-  void SetFlags(uint32 flags, uint32 mask) {
-    // Updates the flags taking into account what flags are supported
-    // by the FST.
-    flags_ &= ~mask;
-    flags_ |= (flags & fst_.GetImpl()->ArcIteratorFlags());
-    // If non-caching is not requested (and caching has not already been
-    // performed), then flush data_flags_ to request caching during the next
-    // call to Value().
-    if (!(flags_ & kArcNoCache) && data_flags_ != kArcValueFlags) {
-      if (!fst_.GetImpl()->HasArcs(s_)) data_flags_ = 0;
-    }
-    // If data_flags_ has been flushed but non-caching is requested before
-    // calling Value(), then set up the iterator for non-caching.
-    if ((flags & kArcNoCache) && (!data_flags_)) Init();
-  }
-
- private:
-  const ReplaceFst<Arc, StateTable, CacheStore> &fst_;  // Reference to the FST.
-  StateId s_;                                           // State in the FST.
-  mutable StateTuple tuple_;  // Tuple corresponding to state_.
-
-  ssize_t pos_;             // Current position.
-  mutable ssize_t offset_;  // Offset between position in iterator and in arcs_.
-  ssize_t num_arcs_;        // Number of arcs at state_.
-  uint32 flags_;            // Behavorial flags for the arc iterator
-  mutable Arc arc_;         // Memory to temporarily store computed arcs.
-
-  mutable ArcIteratorData<Arc> cache_data_;  // Arc iterator data in cache.
-  mutable ArcIteratorData<Arc> local_data_;  // Arc iterator data in local FST.
-
-  mutable const Arc *arcs_;     // Array of arcs.
-  mutable uint32 data_flags_;   // Arc value flags valid for data in arcs_.
-  mutable Arc final_arc_;       // Final arc (when required).
-  mutable uint32 final_flags_;  // Arc value flags valid for final_arc_.
-
-  ArcIterator(const ArcIterator &) = delete;
-  ArcIterator &operator=(const ArcIterator &) = delete;
-};
-
-template <class Arc, class StateTable, class CacheStore>
-class ReplaceFstMatcher : public MatcherBase<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FST = ReplaceFst<Arc, StateTable, CacheStore>;
-  using LocalMatcher = MultiEpsMatcher<Matcher<Fst<Arc>>>;
-
-  using StateTuple = typename StateTable::StateTuple;
-
-  ReplaceFstMatcher(const ReplaceFst<Arc, StateTable, CacheStore> &fst,
-                    MatchType match_type)
-      : fst_(fst),
-        impl_(fst_.GetMutableImpl()),
-        s_(fst::kNoStateId),
-        match_type_(match_type),
-        current_loop_(false),
-        final_arc_(false),
-        loop_(kNoLabel, 0, Weight::One(), kNoStateId) {
-    if (match_type_ == fst::MATCH_OUTPUT) {
-      std::swap(loop_.ilabel, loop_.olabel);
-    }
-    InitMatchers();
-  }
-
-  ReplaceFstMatcher(
-      const ReplaceFstMatcher<Arc, StateTable, CacheStore> &matcher,
-      bool safe = false)
-      : fst_(matcher.fst_),
-        impl_(fst_.GetMutableImpl()),
-        s_(fst::kNoStateId),
-        match_type_(matcher.match_type_),
-        current_loop_(false),
-        final_arc_(false),
-        loop_(fst::kNoLabel, 0, Weight::One(), fst::kNoStateId) {
-    if (match_type_ == fst::MATCH_OUTPUT) {
-      std::swap(loop_.ilabel, loop_.olabel);
-    }
-    InitMatchers();
-  }
-
-  // Creates a local matcher for each component FST in the RTN. LocalMatcher is
-  // a multi-epsilon wrapper matcher. MultiEpsilonMatcher is used to match each
-  // non-terminal arc, since these non-terminal
-  // turn into epsilons on recursion.
-  void InitMatchers() {
-    const auto &fst_array = impl_->fst_array_;
-    matcher_.resize(fst_array.size());
-    for (Label i = 0; i < fst_array.size(); ++i) {
-      if (fst_array[i]) {
-        matcher_[i].reset(
-            new LocalMatcher(*fst_array[i], match_type_, kMultiEpsList));
-        auto it = impl_->nonterminal_set_.begin();
-        for (; it != impl_->nonterminal_set_.end(); ++it) {
-          matcher_[i]->AddMultiEpsLabel(*it);
-        }
-      }
-    }
-  }
-
-  ReplaceFstMatcher<Arc, StateTable, CacheStore> *Copy(
-      bool safe = false) const override {
-    return new ReplaceFstMatcher<Arc, StateTable, CacheStore>(*this, safe);
-  }
-
-  MatchType Type(bool test) const override {
-    if (match_type_ == MATCH_NONE) return match_type_;
-    const auto true_prop =
-        match_type_ == MATCH_INPUT ? kILabelSorted : kOLabelSorted;
-    const auto false_prop =
-        match_type_ == MATCH_INPUT ? kNotILabelSorted : kNotOLabelSorted;
-    const auto props = fst_.Properties(true_prop | false_prop, test);
-    if (props & true_prop) {
-      return match_type_;
-    } else if (props & false_prop) {
-      return MATCH_NONE;
-    } else {
-      return MATCH_UNKNOWN;
-    }
-  }
-
-  const Fst<Arc> &GetFst() const override { return fst_; }
-
-  uint64 Properties(uint64 props) const override { return props; }
-
-  // Sets the state from which our matching happens.
-  void SetState(StateId s) final {
-    if (s_ == s) return;
-    s_ = s;
-    tuple_ = impl_->GetStateTable()->Tuple(s_);
-    if (tuple_.fst_state == kNoStateId) {
-      done_ = true;
-      return;
-    }
-    // Gets current matcher, used for non-epsilon matching.
-    current_matcher_ = matcher_[tuple_.fst_id].get();
-    current_matcher_->SetState(tuple_.fst_state);
-    loop_.nextstate = s_;
-    final_arc_ = false;
-  }
-
-  // Searches for label from previous set state. If label == 0, first
-  // hallucinate an epsilon loop; otherwise use the underlying matcher to
-  // search for the label or epsilons. Note since the ReplaceFst recursion
-  // on non-terminal arcs causes epsilon transitions to be created we use
-  // MultiEpsilonMatcher to search for possible matches of non-terminals. If the
-  // component FST
-  // reaches a final state we also need to add the exiting final arc.
-  bool Find(Label label) final {
-    bool found = false;
-    label_ = label;
-    if (label_ == 0 || label_ == kNoLabel) {
-      // Computes loop directly, avoiding Replace::ComputeArc.
-      if (label_ == 0) {
-        current_loop_ = true;
-        found = true;
-      }
-      // Searches for matching multi-epsilons.
-      final_arc_ = impl_->ComputeFinalArc(tuple_, nullptr);
-      found = current_matcher_->Find(kNoLabel) || final_arc_ || found;
-    } else {
-      // Searches on a sub machine directly using sub machine matcher.
-      found = current_matcher_->Find(label_);
-    }
-    return found;
-  }
-
-  bool Done() const final {
-    return !current_loop_ && !final_arc_ && current_matcher_->Done();
-  }
-
-  const Arc &Value() const final {
-    if (current_loop_) return loop_;
-    if (final_arc_) {
-      impl_->ComputeFinalArc(tuple_, &arc_);
-      return arc_;
-    }
-    const auto &component_arc = current_matcher_->Value();
-    impl_->ComputeArc(tuple_, component_arc, &arc_);
-    return arc_;
-  }
-
-  void Next() final {
-    if (current_loop_) {
-      current_loop_ = false;
-      return;
-    }
-    if (final_arc_) {
-      final_arc_ = false;
-      return;
-    }
-    current_matcher_->Next();
-  }
-
-  ssize_t Priority(StateId s) final { return fst_.NumArcs(s); }
-
- private:
-  const ReplaceFst<Arc, StateTable, CacheStore> &fst_;
-  internal::ReplaceFstImpl<Arc, StateTable, CacheStore> *impl_;
-  LocalMatcher *current_matcher_;
-  std::vector<std::unique_ptr<LocalMatcher>> matcher_;
-  StateId s_;             // Current state.
-  Label label_;           // Current label.
-  MatchType match_type_;  // Supplied by caller.
-  mutable bool done_;
-  mutable bool current_loop_;  // Current arc is the implicit loop.
-  mutable bool final_arc_;     // Current arc for exiting recursion.
-  mutable StateTuple tuple_;   // Tuple corresponding to state_.
-  mutable Arc arc_;
-  Arc loop_;
-
-  ReplaceFstMatcher &operator=(const ReplaceFstMatcher &) = delete;
-};
-
-template <class Arc, class StateTable, class CacheStore>
-inline void ReplaceFst<Arc, StateTable, CacheStore>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base =
-      new StateIterator<ReplaceFst<Arc, StateTable, CacheStore>>(*this);
-}
-
-using StdReplaceFst = ReplaceFst<StdArc>;
-
-// Recursively replaces arcs in the root FSTs with other FSTs.
-// This version writes the result of replacement to an output MutableFst.
-//
-// Replace supports replacement of arcs in one Fst with another FST. This
-// replacement is recursive. Replace takes an array of FST(s). One FST
-// represents the root (or topology) machine. The root FST refers to other FSTs
-// by recursively replacing arcs labeled as non-terminals with the matching
-// non-terminal FST. Currently Replace uses the output symbols of the arcs to
-// determine whether the arc is a non-terminal arc or not. A non-terminal can be
-// any label that is not a non-zero terminal label in the output alphabet.
-//
-// Note that input argument is a vector of pairs. These correspond to the tuple
-// of non-terminal Label and corresponding FST.
-template <class Arc>
-void Replace(const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
-                 &ifst_array,
-             MutableFst<Arc> *ofst,
-             ReplaceFstOptions<Arc> opts = ReplaceFstOptions<Arc>()) {
-  opts.gc = true;
-  opts.gc_limit = 0;  // Caches only the last state for fastest copy.
-  *ofst = ReplaceFst<Arc>(ifst_array, opts);
-}
-
-template <class Arc>
-void Replace(const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
-                 &ifst_array,
-             MutableFst<Arc> *ofst, const ReplaceUtilOptions &opts) {
-  Replace(ifst_array, ofst, ReplaceFstOptions<Arc>(opts));
-}
-
-// For backwards compatibility.
-template <class Arc>
-void Replace(const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
-                 &ifst_array,
-             MutableFst<Arc> *ofst, typename Arc::Label root,
-             bool epsilon_on_replace) {
-  Replace(ifst_array, ofst, ReplaceFstOptions<Arc>(root, epsilon_on_replace));
-}
-
-template <class Arc>
-void Replace(const std::vector<std::pair<typename Arc::Label, const Fst<Arc> *>>
-                 &ifst_array,
-             MutableFst<Arc> *ofst, typename Arc::Label root) {
-  Replace(ifst_array, ofst, ReplaceFstOptions<Arc>(root));
-}
-
-}  // namespace fst
-
-#endif  // FST_REPLACE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/reverse.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/reverse.h
deleted file mode 100644
index 7c7c89db..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/reverse.h
+++ /dev/null
@@ -1,116 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to sort arcs in an FST.
-
-#ifndef FST_REVERSE_H_
-#define FST_REVERSE_H_
-
-#include <algorithm>
-#include <vector>
-
-#include <fst/cache.h>
-
-
-namespace fst {
-
-// Reverses an FST. The reversed result is written to an output mutable FST.
-// If A transduces string x to y with weight a, then the reverse of A
-// transduces the reverse of x to the reverse of y with weight a.Reverse().
-//
-// Typically, a = a.Reverse() and an arc is its own reverse (e.g., for
-// TropicalWeight or LogWeight). In general, e.g., when the weights only form a
-// left or right semiring, the output arc type must match the input arc type
-// except having the reversed Weight type.
-//
-// When require_superinitial is false, a superinitial state is not created in
-// the reversed FST iff the input FST has exactly one final state (which becomes
-// the initial state of the reversed FST) with a final weight of semiring One,
-// or if it does not belong to any cycle. When require_superinitial is true, a
-// superinitial state is always created.
-template <class FromArc, class ToArc>
-void Reverse(const Fst<FromArc> &ifst, MutableFst<ToArc> *ofst,
-             bool require_superinitial = true) {
-  using StateId = typename FromArc::StateId;
-  using FromWeight = typename FromArc::Weight;
-  using ToWeight = typename ToArc::Weight;
-  ofst->DeleteStates();
-  ofst->SetInputSymbols(ifst.InputSymbols());
-  ofst->SetOutputSymbols(ifst.OutputSymbols());
-  if (ifst.Properties(kExpanded, false)) {
-    ofst->ReserveStates(CountStates(ifst) + 1);
-  }
-  StateId istart = ifst.Start();
-  StateId ostart = kNoStateId;
-  StateId offset = 0;
-  uint64 dfs_iprops = 0;
-  uint64 dfs_oprops = 0;
-  if (!require_superinitial) {
-    for (StateIterator<Fst<FromArc>> siter(ifst); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      if (ifst.Final(s) == FromWeight::Zero()) continue;
-      if (ostart != kNoStateId) {
-        ostart = kNoStateId;
-        break;
-      } else {
-        ostart = s;
-      }
-    }
-    if (ostart != kNoStateId && ifst.Final(ostart) != FromWeight::One()) {
-      std::vector<StateId> scc;
-      SccVisitor<FromArc> scc_visitor(&scc, nullptr, nullptr, &dfs_iprops);
-      DfsVisit(ifst, &scc_visitor);
-      if (count(scc.begin(), scc.end(), scc[ostart]) > 1) {
-        ostart = kNoStateId;
-      } else {
-        for (ArcIterator<Fst<FromArc>> aiter(ifst, ostart); !aiter.Done();
-             aiter.Next()) {
-          if (aiter.Value().nextstate == ostart) {
-            ostart = kNoStateId;
-            break;
-          }
-        }
-      }
-      if (ostart != kNoStateId) dfs_oprops = kInitialAcyclic;
-    }
-  }
-  if (ostart == kNoStateId) {  // Super-initial requested or needed.
-    ostart = ofst->AddState();
-    offset = 1;
-  }
-  for (StateIterator<Fst<FromArc>> siter(ifst); !siter.Done(); siter.Next()) {
-    const auto is = siter.Value();
-    const auto os = is + offset;
-    while (ofst->NumStates() <= os) ofst->AddState();
-    if (is == istart) ofst->SetFinal(os, ToWeight::One());
-    const auto weight = ifst.Final(is);
-    if ((weight != FromWeight::Zero()) && (offset == 1)) {
-      const ToArc oarc(0, 0, weight.Reverse(), os);
-      ofst->AddArc(0, oarc);
-    }
-    for (ArcIterator<Fst<FromArc>> aiter(ifst, is); !aiter.Done();
-         aiter.Next()) {
-      const auto &iarc = aiter.Value();
-      const auto nos = iarc.nextstate + offset;
-      auto weight = iarc.weight.Reverse();
-      if (!offset && (nos == ostart)) {
-        weight = Times(ifst.Final(ostart).Reverse(), weight);
-      }
-      const ToArc oarc(iarc.ilabel, iarc.olabel, weight, os);
-      while (ofst->NumStates() <= nos) ofst->AddState();
-      ofst->AddArc(nos, oarc);
-    }
-  }
-  ofst->SetStart(ostart);
-  if (offset == 0 && ostart == istart) {
-    ofst->SetFinal(ostart, ifst.Final(ostart).Reverse());
-  }
-  const auto iprops = ifst.Properties(kCopyProperties, false) | dfs_iprops;
-  const auto oprops = ofst->Properties(kFstProperties, false) | dfs_oprops;
-  ofst->SetProperties(ReverseProperties(iprops, offset == 1) | oprops,
-                      kFstProperties);
-}
-
-}  // namespace fst
-
-#endif  // FST_REVERSE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/reweight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/reweight.h
deleted file mode 100644
index 64e68cb7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/reweight.h
+++ /dev/null
@@ -1,127 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function to reweight an FST.
-
-#ifndef FST_REWEIGHT_H_
-#define FST_REWEIGHT_H_
-
-#include <vector>
-#include <fst/log.h>
-
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-enum ReweightType { REWEIGHT_TO_INITIAL, REWEIGHT_TO_FINAL };
-
-// Reweights an FST according to a vector of potentials in a given direction.
-// The weight must be left distributive when reweighting towards the initial
-// state and right distributive when reweighting towards the final states.
-//
-// An arc of weight w, with an origin state of potential p and destination state
-// of potential q, is reweighted by p^-1 \otimes (w \otimes q) when reweighting
-// torwards the initial state, and by (p \otimes w) \otimes q^-1 when
-// reweighting towards the final states.
-template <class Arc>
-void Reweight(MutableFst<Arc> *fst,
-              const std::vector<typename Arc::Weight> &potential,
-              ReweightType type) {
-  using Weight = typename Arc::Weight;
-  if (fst->NumStates() == 0) return;
-  // TODO(kbg): Make this a compile-time static_assert once we have a pleasant
-  // way to "deregister" this operation for non-distributive semirings so an
-  // informative error message is produced.
-  if (type == REWEIGHT_TO_FINAL && !(Weight::Properties() & kRightSemiring)) {
-    FSTERROR() << "Reweight: Reweighting to the final states requires "
-               << "Weight to be right distributive: " << Weight::Type();
-    fst->SetProperties(kError, kError);
-    return;
-  }
-  // TODO(kbg): Make this a compile-time static_assert once we have a pleasant
-  // way to "deregister" this operation for non-distributive semirings so an
-  // informative error message is produced.
-  if (type == REWEIGHT_TO_INITIAL && !(Weight::Properties() & kLeftSemiring)) {
-    FSTERROR() << "Reweight: Reweighting to the initial state requires "
-               << "Weight to be left distributive: " << Weight::Type();
-    fst->SetProperties(kError, kError);
-    return;
-  }
-  StateIterator<MutableFst<Arc>> siter(*fst);
-  for (; !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    if (s == potential.size()) break;
-    const auto &weight = potential[s];
-    if (weight != Weight::Zero()) {
-      for (MutableArcIterator<MutableFst<Arc>> aiter(fst, s); !aiter.Done();
-           aiter.Next()) {
-        auto arc = aiter.Value();
-        if (arc.nextstate >= potential.size()) continue;
-        const auto &nextweight = potential[arc.nextstate];
-        if (nextweight == Weight::Zero()) continue;
-        if (type == REWEIGHT_TO_INITIAL) {
-          arc.weight =
-              Divide(Times(arc.weight, nextweight), weight, DIVIDE_LEFT);
-        }
-        if (type == REWEIGHT_TO_FINAL) {
-          arc.weight =
-              Divide(Times(weight, arc.weight), nextweight, DIVIDE_RIGHT);
-        }
-        aiter.SetValue(arc);
-      }
-      if (type == REWEIGHT_TO_INITIAL) {
-        fst->SetFinal(s, Divide(fst->Final(s), weight, DIVIDE_LEFT));
-      }
-    }
-    if (type == REWEIGHT_TO_FINAL) {
-      fst->SetFinal(s, Times(weight, fst->Final(s)));
-    }
-  }
-  // This handles elements past the end of the potentials array.
-  for (; !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    if (type == REWEIGHT_TO_FINAL) {
-      fst->SetFinal(s, Times(Weight::Zero(), fst->Final(s)));
-    }
-  }
-  const auto startweight = fst->Start() < potential.size()
-                               ? potential[fst->Start()]
-                               : Weight::Zero();
-  if ((startweight != Weight::One()) && (startweight != Weight::Zero())) {
-    if (fst->Properties(kInitialAcyclic, true) & kInitialAcyclic) {
-      const auto s = fst->Start();
-      for (MutableArcIterator<MutableFst<Arc>> aiter(fst, s); !aiter.Done();
-           aiter.Next()) {
-        auto arc = aiter.Value();
-        if (type == REWEIGHT_TO_INITIAL) {
-          arc.weight = Times(startweight, arc.weight);
-        } else {
-          arc.weight = Times(Divide(Weight::One(), startweight, DIVIDE_RIGHT),
-                             arc.weight);
-        }
-        aiter.SetValue(arc);
-      }
-      if (type == REWEIGHT_TO_INITIAL) {
-        fst->SetFinal(s, Times(startweight, fst->Final(s)));
-      } else {
-        fst->SetFinal(s, Times(Divide(Weight::One(), startweight, DIVIDE_RIGHT),
-                               fst->Final(s)));
-      }
-    } else {
-      const auto s = fst->AddState();
-      const auto weight =
-          (type == REWEIGHT_TO_INITIAL)
-              ? startweight
-              : Divide(Weight::One(), startweight, DIVIDE_RIGHT);
-      fst->AddArc(s, Arc(0, 0, weight, fst->Start()));
-      fst->SetStart(s);
-    }
-  }
-  fst->SetProperties(ReweightProperties(fst->Properties(kFstProperties, false)),
-                     kFstProperties);
-}
-
-}  // namespace fst
-
-#endif  // FST_REWEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/rmepsilon.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/rmepsilon.h
deleted file mode 100644
index ee17b418..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/rmepsilon.h
+++ /dev/null
@@ -1,548 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes that implemement epsilon-removal.
-
-#ifndef FST_RMEPSILON_H_
-#define FST_RMEPSILON_H_
-
-#include <forward_list>
-#include <stack>
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arcfilter.h>
-#include <fst/cache.h>
-#include <fst/connect.h>
-#include <fst/factor-weight.h>
-#include <fst/invert.h>
-#include <fst/prune.h>
-#include <fst/queue.h>
-#include <fst/shortest-distance.h>
-#include <fst/topsort.h>
-
-
-namespace fst {
-
-template <class Arc, class Queue>
-class RmEpsilonOptions
-    : public ShortestDistanceOptions<Arc, Queue, EpsilonArcFilter<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  bool connect;             // Connect output
-  Weight weight_threshold;  // Pruning weight threshold.
-  StateId state_threshold;  // Pruning state threshold.
-
-  explicit RmEpsilonOptions(Queue *queue, float delta = kDelta,
-                            bool connect = true,
-                            Weight weight_threshold = Weight::Zero(),
-                            StateId state_threshold = kNoStateId)
-      : ShortestDistanceOptions<Arc, Queue, EpsilonArcFilter<Arc>>(
-            queue, EpsilonArcFilter<Arc>(), kNoStateId, delta),
-        connect(connect),
-        weight_threshold(std::move(weight_threshold)),
-        state_threshold(state_threshold) {}
-};
-
-namespace internal {
-
-// Computation state of the epsilon-removal algorithm.
-template <class Arc, class Queue>
-class RmEpsilonState {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  RmEpsilonState(const Fst<Arc> &fst, std::vector<Weight> *distance,
-                 const RmEpsilonOptions<Arc, Queue> &opts)
-      : fst_(fst),
-        distance_(distance),
-        sd_state_(fst_, distance, opts, true),
-        expand_id_(0) {}
-
-  void Expand(StateId s);
-
-  std::vector<Arc> &Arcs() { return arcs_; }
-
-  const Weight &Final() const { return final_; }
-
-  bool Error() const { return sd_state_.Error(); }
-
- private:
-  struct Element {
-    Label ilabel;
-    Label olabel;
-    StateId nextstate;
-
-    Element() {}
-
-    Element(Label ilabel, Label olabel, StateId nexstate)
-        : ilabel(ilabel), olabel(olabel), nextstate(nexstate) {}
-  };
-
-  struct ElementHash {
-   public:
-    size_t operator()(const Element &element) const {
-      static constexpr size_t prime0 = 7853;
-      static constexpr size_t prime1 = 7867;
-      return static_cast<size_t>(element.nextstate) +
-             static_cast<size_t>(element.ilabel) * prime0 +
-             static_cast<size_t>(element.olabel) * prime1;
-    }
-  };
-
-  class ElementEqual {
-   public:
-    bool operator()(const Element &e1, const Element &e2) const {
-      return (e1.ilabel == e2.ilabel) && (e1.olabel == e2.olabel) &&
-             (e1.nextstate == e2.nextstate);
-    }
-  };
-
-  using ElementMap = std::unordered_map<Element, std::pair<StateId, size_t>,
-                                        ElementHash, ElementEqual>;
-
-  const Fst<Arc> &fst_;
-  // Distance from state being expanded in epsilon-closure.
-  std::vector<Weight> *distance_;
-  // Shortest distance algorithm computation state.
-  internal::ShortestDistanceState<Arc, Queue, EpsilonArcFilter<Arc>> sd_state_;
-  // Maps an element to a pair corresponding to a position in the arcs vector
-  // of the state being expanded. The element corresopnds to the position in
-  // the arcs_ vector if p.first is equal to the state being expanded.
-  ElementMap element_map_;
-  EpsilonArcFilter<Arc> eps_filter_;
-  std::stack<StateId> eps_queue_;  // Queue used to visit the epsilon-closure.
-  std::vector<bool> visited_;      // True if the state has been visited.
-  std::forward_list<StateId> visited_states_;  // List of visited states.
-  std::vector<Arc> arcs_;                      // Arcs of state being expanded.
-  Weight final_;       // Final weight of state being expanded.
-  StateId expand_id_;  // Unique ID for each call to Expand
-
-  RmEpsilonState(const RmEpsilonState &) = delete;
-  RmEpsilonState &operator=(const RmEpsilonState &) = delete;
-};
-
-template <class Arc, class Queue>
-void RmEpsilonState<Arc, Queue>::Expand(typename Arc::StateId source) {
-  final_ = Weight::Zero();
-  arcs_.clear();
-  sd_state_.ShortestDistance(source);
-  if (sd_state_.Error()) return;
-  eps_queue_.push(source);
-  while (!eps_queue_.empty()) {
-    const auto state = eps_queue_.top();
-    eps_queue_.pop();
-    while (visited_.size() <= state) visited_.push_back(false);
-    if (visited_[state]) continue;
-    visited_[state] = true;
-    visited_states_.push_front(state);
-    for (ArcIterator<Fst<Arc>> aiter(fst_, state); !aiter.Done();
-         aiter.Next()) {
-      auto arc = aiter.Value();
-      arc.weight = Times((*distance_)[state], arc.weight);
-      if (eps_filter_(arc)) {
-        while (visited_.size() <= arc.nextstate) visited_.push_back(false);
-        if (!visited_[arc.nextstate]) eps_queue_.push(arc.nextstate);
-      } else {
-        const Element element(arc.ilabel, arc.olabel, arc.nextstate);
-        auto insert_result = element_map_.insert(
-            std::make_pair(element, std::make_pair(expand_id_, arcs_.size())));
-        if (insert_result.second) {
-          arcs_.push_back(arc);
-        } else {
-          if (insert_result.first->second.first == expand_id_) {
-            auto &weight = arcs_[insert_result.first->second.second].weight;
-            weight = Plus(weight, arc.weight);
-          } else {
-            insert_result.first->second.first = expand_id_;
-            insert_result.first->second.second = arcs_.size();
-            arcs_.push_back(arc);
-          }
-        }
-      }
-    }
-    final_ = Plus(final_, Times((*distance_)[state], fst_.Final(state)));
-  }
-  while (!visited_states_.empty()) {
-    visited_[visited_states_.front()] = false;
-    visited_states_.pop_front();
-  }
-  ++expand_id_;
-}
-
-}  // namespace internal
-
-// Removes epsilon-transitions (when both the input and output label are an
-// epsilon) from a transducer. The result will be an equivalent FST that has no
-// such epsilon transitions. This version modifies its input. It allows fine
-// control via the options argument; see below for a simpler interface.
-//
-// The distance vector will be used to hold the shortest distances during the
-// epsilon-closure computation. The state queue discipline and convergence delta
-// are taken in the options argument.
-template <class Arc, class Queue>
-void RmEpsilon(MutableFst<Arc> *fst,
-               std::vector<typename Arc::Weight> *distance,
-               const RmEpsilonOptions<Arc, Queue> &opts) {
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  if (fst->Start() == kNoStateId) return;
-  // noneps_in[s] will be set to true iff s admits a non-epsilon incoming
-  // transition or is the start state.
-  std::vector<bool> noneps_in(fst->NumStates(), false);
-  noneps_in[fst->Start()] = true;
-  for (size_t i = 0; i < fst->NumStates(); ++i) {
-    for (ArcIterator<Fst<Arc>> aiter(*fst, i); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      if (arc.ilabel != 0 || arc.olabel != 0) {
-        noneps_in[arc.nextstate] = true;
-      }
-    }
-  }
-  // States sorted in topological order when (acyclic) or generic topological
-  // order (cyclic).
-  std::vector<StateId> states;
-  states.reserve(fst->NumStates());
-  if (fst->Properties(kTopSorted, false) & kTopSorted) {
-    for (size_t i = 0; i < fst->NumStates(); i++) states.push_back(i);
-  } else if (fst->Properties(kAcyclic, false) & kAcyclic) {
-    std::vector<StateId> order;
-    bool acyclic;
-    TopOrderVisitor<Arc> top_order_visitor(&order, &acyclic);
-    DfsVisit(*fst, &top_order_visitor, EpsilonArcFilter<Arc>());
-    // Sanity check: should be acyclic if property bit is set.
-    if (!acyclic) {
-      FSTERROR() << "RmEpsilon: Inconsistent acyclic property bit";
-      fst->SetProperties(kError, kError);
-      return;
-    }
-    states.resize(order.size());
-    for (StateId i = 0; i < order.size(); i++) states[order[i]] = i;
-  } else {
-    uint64 props;
-    std::vector<StateId> scc;
-    SccVisitor<Arc> scc_visitor(&scc, nullptr, nullptr, &props);
-    DfsVisit(*fst, &scc_visitor, EpsilonArcFilter<Arc>());
-    std::vector<StateId> first(scc.size(), kNoStateId);
-    std::vector<StateId> next(scc.size(), kNoStateId);
-    for (StateId i = 0; i < scc.size(); i++) {
-      if (first[scc[i]] != kNoStateId) next[i] = first[scc[i]];
-      first[scc[i]] = i;
-    }
-    for (StateId i = 0; i < first.size(); i++) {
-      for (auto j = first[i]; j != kNoStateId; j = next[j]) {
-        states.push_back(j);
-      }
-    }
-  }
-  internal::RmEpsilonState<Arc, Queue> rmeps_state(*fst, distance, opts);
-  while (!states.empty()) {
-    const auto state = states.back();
-    states.pop_back();
-    if (!noneps_in[state] &&
-        (opts.connect || opts.weight_threshold != Weight::Zero() ||
-         opts.state_threshold != kNoStateId)) {
-      continue;
-    }
-    rmeps_state.Expand(state);
-    fst->SetFinal(state, rmeps_state.Final());
-    fst->DeleteArcs(state);
-    auto &arcs = rmeps_state.Arcs();
-    fst->ReserveArcs(state, arcs.size());
-    while (!arcs.empty()) {
-      fst->AddArc(state, arcs.back());
-      arcs.pop_back();
-    }
-  }
-  if (opts.connect || opts.weight_threshold != Weight::Zero() ||
-      opts.state_threshold != kNoStateId) {
-    for (size_t s = 0; s < fst->NumStates(); ++s) {
-      if (!noneps_in[s]) fst->DeleteArcs(s);
-    }
-  }
-  if (rmeps_state.Error()) fst->SetProperties(kError, kError);
-  fst->SetProperties(
-      RmEpsilonProperties(fst->Properties(kFstProperties, false)),
-      kFstProperties);
-  if (opts.weight_threshold != Weight::Zero() ||
-      opts.state_threshold != kNoStateId) {
-    Prune(fst, opts.weight_threshold, opts.state_threshold);
-  }
-  if (opts.connect && opts.weight_threshold == Weight::Zero() &&
-      opts.state_threshold == kNoStateId) {
-    Connect(fst);
-  }
-}
-
-// Removes epsilon-transitions (when both the input and output label
-// are an epsilon) from a transducer. The result will be an equivalent
-// FST that has no such epsilon transitions. This version modifies its
-// input. It has a simplified interface; see above for a version that
-// allows finer control.
-//
-// Complexity:
-//
-// - Time:
-//
-//   Unweighted: O(v^2 + ve).
-//   Acyclic: O(v^2 + V e).
-//   Tropical semiring: O(v^2 log V + ve).
-//   General: exponential.
-//
-// - Space: O(vE)
-//
-// where v is the number of states visited and e is the number of arcs visited.
-//
-// For more information, see:
-//
-// Mohri, M. 2002. Generic epsilon-removal and input epsilon-normalization
-// algorithms for weighted transducers. International Journal of Computer
-// Science 13(1): 129-143.
-template <class Arc>
-void RmEpsilon(MutableFst<Arc> *fst, bool connect = true,
-               typename Arc::Weight weight_threshold = Arc::Weight::Zero(),
-               typename Arc::StateId state_threshold = kNoStateId,
-               float delta = kDelta) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  std::vector<Weight> distance;
-  AutoQueue<StateId> state_queue(*fst, &distance, EpsilonArcFilter<Arc>());
-  RmEpsilonOptions<Arc, AutoQueue<StateId>> opts(
-      &state_queue, delta, connect, weight_threshold, state_threshold);
-  RmEpsilon(fst, &distance, opts);
-}
-
-struct RmEpsilonFstOptions : CacheOptions {
-  float delta;
-
-  explicit RmEpsilonFstOptions(const CacheOptions &opts, float delta = kDelta)
-      : CacheOptions(opts), delta(delta) {}
-
-  explicit RmEpsilonFstOptions(float delta = kDelta) : delta(delta) {}
-};
-
-namespace internal {
-
-// Implementation of delayed RmEpsilonFst.
-template <class Arc>
-class RmEpsilonFstImpl : public CacheImpl<Arc> {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-
-  using FstImpl<Arc>::Properties;
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
-  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
-  using CacheBaseImpl<CacheState<Arc>>::HasStart;
-  using CacheBaseImpl<CacheState<Arc>>::PushArc;
-  using CacheBaseImpl<CacheState<Arc>>::SetArcs;
-  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
-  using CacheBaseImpl<CacheState<Arc>>::SetStart;
-
-  RmEpsilonFstImpl(const Fst<Arc> &fst, const RmEpsilonFstOptions &opts)
-      : CacheImpl<Arc>(opts),
-        fst_(fst.Copy()),
-        delta_(opts.delta),
-        rmeps_state_(
-            *fst_, &distance_,
-            RmEpsilonOptions<Arc, FifoQueue<StateId>>(&queue_, delta_, false)) {
-    SetType("rmepsilon");
-    SetProperties(
-        RmEpsilonProperties(fst.Properties(kFstProperties, false), true),
-        kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  RmEpsilonFstImpl(const RmEpsilonFstImpl &impl)
-      : CacheImpl<Arc>(impl),
-        fst_(impl.fst_->Copy(true)),
-        delta_(impl.delta_),
-        rmeps_state_(
-            *fst_, &distance_,
-            RmEpsilonOptions<Arc, FifoQueue<StateId>>(&queue_, delta_, false)) {
-    SetType("rmepsilon");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  StateId Start() {
-    if (!HasStart()) SetStart(fst_->Start());
-    return CacheImpl<Arc>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) Expand(s);
-    return CacheImpl<Arc>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumOutputEpsilons(s);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found and returns other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) &&
-        (fst_->Properties(kError, false) || rmeps_state_.Error())) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<Arc>::InitArcIterator(s, data);
-  }
-
-  void Expand(StateId s) {
-    rmeps_state_.Expand(s);
-    SetFinal(s, rmeps_state_.Final());
-    auto &arcs = rmeps_state_.Arcs();
-    while (!arcs.empty()) {
-      PushArc(s, arcs.back());
-      arcs.pop_back();
-    }
-    SetArcs(s);
-  }
-
- private:
-  std::unique_ptr<const Fst<Arc>> fst_;
-  float delta_;
-  std::vector<Weight> distance_;
-  FifoQueue<StateId> queue_;
-  internal::RmEpsilonState<Arc, FifoQueue<StateId>> rmeps_state_;
-};
-
-}  // namespace internal
-
-// Removes epsilon-transitions (when both the input and output label are an
-// epsilon) from a transducer. The result will be an equivalent FST that has no
-// such epsilon transitions. This version is a
-// delayed FST.
-//
-// Complexity:
-//
-// - Time:
-//   Unweighted: O(v^2 + ve).
-//   General: exponential.
-//
-// - Space: O(vE)
-//
-// where v is the number of states visited and e is the number of arcs visited.
-// Constant time to visit an input state or arc is assumed and exclusive of
-// caching.
-//
-// For more information, see:
-//
-// Mohri, M. 2002. Generic epsilon-removal and input epsilon-normalization
-// algorithms for weighted transducers. International Journal of Computer
-// Science 13(1): 129-143.
-//
-// This class attaches interface to implementation and handles
-// reference counting, delegating most methods to ImplToFst.
-template <class A>
-class RmEpsilonFst : public ImplToFst<internal::RmEpsilonFstImpl<A>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::RmEpsilonFstImpl<Arc>;
-
-  friend class ArcIterator<RmEpsilonFst<Arc>>;
-  friend class StateIterator<RmEpsilonFst<Arc>>;
-
-  explicit RmEpsilonFst(const Fst<Arc> &fst)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, RmEpsilonFstOptions())) {}
-
-  RmEpsilonFst(const Fst<A> &fst, const RmEpsilonFstOptions &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  RmEpsilonFst(const RmEpsilonFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this RmEpsilonFst. See Fst<>::Copy() for further doc.
-  RmEpsilonFst<Arc> *Copy(bool safe = false) const override {
-    return new RmEpsilonFst<Arc>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  RmEpsilonFst &operator=(const RmEpsilonFst &) = delete;
-};
-
-// Specialization for RmEpsilonFst.
-template <class Arc>
-class StateIterator<RmEpsilonFst<Arc>>
-    : public CacheStateIterator<RmEpsilonFst<Arc>> {
- public:
-  explicit StateIterator(const RmEpsilonFst<Arc> &fst)
-      : CacheStateIterator<RmEpsilonFst<Arc>>(fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for RmEpsilonFst.
-template <class Arc>
-class ArcIterator<RmEpsilonFst<Arc>>
-    : public CacheArcIterator<RmEpsilonFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const RmEpsilonFst<Arc> &fst, StateId s)
-      : CacheArcIterator<RmEpsilonFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc>
-inline void RmEpsilonFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<RmEpsilonFst<Arc>>(*this);
-}
-
-// Useful alias when using StdArc.
-using StdRmEpsilonFst = RmEpsilonFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_RMEPSILON_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/rmfinalepsilon.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/rmfinalepsilon.h
deleted file mode 100644
index 87e3a714..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/rmfinalepsilon.h
+++ /dev/null
@@ -1,80 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function to remove of final states that have epsilon-only input arcs.
-
-#ifndef FST_RMFINALEPSILON_H_
-#define FST_RMFINALEPSILON_H_
-
-#include <unordered_set>
-#include <vector>
-
-#include <fst/connect.h>
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-// Removes final states that have epsilon-only input arcs.
-template <class Arc>
-void RmFinalEpsilon(MutableFst<Arc> *fst) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  // Determines the coaccesibility of states.
-  std::vector<bool> access;
-  std::vector<bool> coaccess;
-  uint64 props = 0;
-  SccVisitor<Arc> scc_visitor(nullptr, &access, &coaccess, &props);
-  DfsVisit(*fst, &scc_visitor);
-  // Finds potential list of removable final states. These are final states that
-  // have no outgoing transitions or final states that have a non-coaccessible
-  // future.
-  std::unordered_set<StateId> finals;
-  for (StateIterator<Fst<Arc>> siter(*fst); !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    if (fst->Final(s) != Weight::Zero()) {
-      bool future_coaccess = false;
-      for (ArcIterator<Fst<Arc>> aiter(*fst, s); !aiter.Done(); aiter.Next()) {
-        const auto &arc = aiter.Value();
-        if (coaccess[arc.nextstate]) {
-          future_coaccess = true;
-          break;
-        }
-      }
-      if (!future_coaccess) finals.insert(s);
-    }
-  }
-  // Moves the final weight.
-  std::vector<Arc> arcs;
-  for (StateIterator<Fst<Arc>> siter(*fst); !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    auto weight = fst->Final(s);
-    arcs.clear();
-    for (ArcIterator<Fst<Arc>> aiter(*fst, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      // Next state is in the list of finals.
-      if (finals.find(arc.nextstate) != finals.end()) {
-        // Sums up all epsilon arcs.
-        if (arc.ilabel == 0 && arc.olabel == 0) {
-          weight = Plus(Times(fst->Final(arc.nextstate), arc.weight), weight);
-        } else {
-          arcs.push_back(arc);
-        }
-      } else {
-        arcs.push_back(arc);
-      }
-    }
-    // If some arcs (epsilon arcs) were deleted, delete all arcs and add back
-    // only the non-epsilon arcs.
-    if (arcs.size() < fst->NumArcs(s)) {
-      fst->DeleteArcs(s);
-      fst->SetFinal(s, weight);
-      for (const auto &arc : arcs) fst->AddArc(s, arc);
-    }
-  }
-  Connect(fst);
-}
-
-}  // namespace fst
-
-#endif  // FST_RMFINALEPSILON_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arc-class.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arc-class.h
deleted file mode 100644
index 551266d7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arc-class.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_ARC_CLASS_H_
-#define FST_SCRIPT_ARC_CLASS_H_
-
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-// A struct representing an arc while ignoring arc type. It is passed as an
-// argument to AddArc.
-
-struct ArcClass {
-  template <class Arc>
-  explicit ArcClass(const Arc &arc)
-      : ilabel(arc.ilabel), olabel(arc.olabel), weight(arc.weight),
-        nextstate(arc.nextstate) {}
-
-  ArcClass(int64 ilabel, int64 olabel, const WeightClass &weight,
-           int64 nextstate)
-      : ilabel(ilabel), olabel(olabel), weight(weight), nextstate(nextstate) {}
-
-  template <class Arc>
-  Arc GetArc() const {
-    return Arc(ilabel, olabel, *(weight.GetWeight<typename Arc::Weight>()),
-               nextstate);
-  }
-
-  int64 ilabel;
-  int64 olabel;
-  WeightClass weight;
-  int64 nextstate;
-};
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_ARC_CLASS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arciterator-class.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arciterator-class.h
deleted file mode 100644
index 8e4ca4f8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arciterator-class.h
+++ /dev/null
@@ -1,212 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_ARCITERATOR_CLASS_H_
-#define FST_SCRIPT_ARCITERATOR_CLASS_H_
-
-#include <memory>
-#include <utility>
-
-#include <fst/fstlib.h>
-#include <fst/script/fst-class.h>
-
-// Scripting API support for ArcIterator.
-//
-// A call to Value() causes the underlying arc to be used to construct the
-// associated ArcClass.
-
-namespace fst {
-namespace script {
-
-// Non-mutable arc iterators.
-
-// Virtual interface implemented by each concrete ArcIteratorImpl<F>.
-class ArcIteratorImplBase {
- public:
-  virtual bool Done() const = 0;
-  virtual uint32 Flags() const = 0;
-  virtual void Next() = 0;
-  virtual size_t Position() const = 0;
-  virtual void Reset() = 0;
-  virtual void Seek(size_t a) = 0;
-  virtual void SetFlags(uint32 flags, uint32 mask) = 0;
-  virtual ArcClass Value() const = 0;
-  virtual ~ArcIteratorImplBase() {}
-};
-
-// Templated implementation.
-template <class Arc>
-class ArcIteratorClassImpl : public ArcIteratorImplBase {
- public:
-  explicit ArcIteratorClassImpl(const Fst<Arc> &fst, int64 s)
-      : aiter_(fst, s) {}
-
-  bool Done() const final { return aiter_.Done(); }
-
-  uint32 Flags() const final { return aiter_.Flags(); }
-
-  void Next() final { aiter_.Next(); }
-
-  size_t Position() const final { return aiter_.Position(); }
-
-  void Reset() final { aiter_.Reset(); }
-
-  void Seek(size_t a) final { aiter_.Seek(a); }
-
-  void SetFlags(uint32 flags, uint32 mask) final {
-    aiter_.SetFlags(flags, mask);
-  }
-
-  // This is returned by value because it has not yet been constructed, and
-  // is likely to participate in return-value optimization.
-  ArcClass Value() const final { return ArcClass(aiter_.Value()); }
-
-  ~ArcIteratorClassImpl() final {}
-
- private:
-  ArcIterator<Fst<Arc>> aiter_;
-};
-
-class ArcIteratorClass;
-
-using InitArcIteratorClassArgs =
-    std::tuple<const FstClass &, int64, ArcIteratorClass *>;
-
-// Untemplated user-facing class holding a templated pimpl.
-class ArcIteratorClass {
- public:
-  ArcIteratorClass(const FstClass &fst, int64 s);
-
-  template <class Arc>
-  ArcIteratorClass(const Fst<Arc> &fst, int64 s)
-      : impl_(new ArcIteratorClassImpl<Arc>(fst, s)) {}
-
-  bool Done() const { return impl_->Done(); }
-
-  uint32 Flags() const { return impl_->Flags(); }
-
-  void Next() { impl_->Next(); }
-
-  size_t Position() const { return impl_->Position(); }
-
-  void Reset() { impl_->Reset(); }
-
-  void Seek(size_t a) { impl_->Seek(a); }
-
-  void SetFlags(uint32 flags, uint32 mask) { impl_->SetFlags(flags, mask); }
-
-  ArcClass Value() const { return impl_->Value(); }
-
-  template <class Arc>
-  friend void InitArcIteratorClass(InitArcIteratorClassArgs *args);
-
- private:
-  std::unique_ptr<ArcIteratorImplBase> impl_;
-};
-
-template <class Arc>
-void InitArcIteratorClass(InitArcIteratorClassArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  std::get<2>(*args)->impl_.reset(
-      new ArcIteratorClassImpl<Arc>(fst, std::get<1>(*args)));
-}
-
-// Mutable arc iterators.
-
-// Virtual interface implemented by each concrete MutableArcIteratorImpl<F>.
-class MutableArcIteratorImplBase : public ArcIteratorImplBase {
- public:
-  virtual void SetValue(const ArcClass &) = 0;
-
-  ~MutableArcIteratorImplBase() override {}
-};
-
-// Templated implementation.
-template <class Arc>
-class MutableArcIteratorClassImpl
-    : public MutableArcIteratorImplBase {
- public:
-  explicit MutableArcIteratorClassImpl(MutableFst<Arc> *fst, int64 s)
-      : aiter_(fst, s) {}
-
-  bool Done() const final { return aiter_.Done(); }
-
-  uint32 Flags() const final { return aiter_.Flags(); }
-
-  void Next() final { aiter_.Next(); }
-
-  size_t Position() const final { return aiter_.Position(); }
-
-  void Reset() final { aiter_.Reset(); }
-
-  void Seek(size_t a) final { aiter_.Seek(a); }
-
-  void SetFlags(uint32 flags, uint32 mask) final {
-    aiter_.SetFlags(flags, mask);
-  }
-
-  void SetValue(const Arc &arc) { aiter_.SetValue(arc); }
-
-  void SetValue(const ArcClass &ac) final { aiter_.SetValue(ac.GetArc<Arc>()); }
-
-  // This is returned by value because it has not yet been constructed, and
-  // is likely to participate in return-value optimization.
-  ArcClass Value() const final { return ArcClass(aiter_.Value()); }
-
-  ~MutableArcIteratorClassImpl() override {}
-
- private:
-  MutableArcIterator<MutableFst<Arc>> aiter_;
-};
-
-class MutableArcIteratorClass;
-
-using InitMutableArcIteratorClassArgs =
-    std::tuple<MutableFstClass *, int64, MutableArcIteratorClass *>;
-
-// Untemplated user-facing class holding a templated pimpl.
-class MutableArcIteratorClass {
- public:
-  MutableArcIteratorClass(MutableFstClass *fst, int64 s);
-
-  template <class Arc>
-  MutableArcIteratorClass(MutableFst<Arc> *fst, int64 s)
-      : impl_(new MutableArcIteratorClassImpl<Arc>(fst, s)) {}
-
-  bool Done() const { return impl_->Done(); }
-
-  uint32 Flags() const { return impl_->Flags(); }
-
-  void Next() { impl_->Next(); }
-
-  size_t Position() const { return impl_->Position(); }
-
-  void Reset() { impl_->Reset(); }
-
-  void Seek(size_t a) { impl_->Seek(a); }
-
-  void SetFlags(uint32 flags, uint32 mask) { impl_->SetFlags(flags, mask); }
-
-  void SetValue(const ArcClass &ac) { impl_->SetValue(ac); }
-
-  ArcClass Value() const { return impl_->Value(); }
-
-  template <class Arc>
-  friend void InitMutableArcIteratorClass(
-      InitMutableArcIteratorClassArgs *args);
-
- private:
-  std::unique_ptr<MutableArcIteratorImplBase> impl_;
-};
-
-template <class Arc>
-void InitMutableArcIteratorClass(InitMutableArcIteratorClassArgs *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  std::get<2>(*args)->impl_.reset(
-      new MutableArcIteratorClassImpl<Arc>(fst, std::get<1>(*args)));
-}
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_ARCITERATOR_CLASS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arcsort.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arcsort.h
deleted file mode 100644
index 3e56fe5c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arcsort.h
+++ /dev/null
@@ -1,44 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_ARCSORT_H_
-#define FST_SCRIPT_ARCSORT_H_
-
-#include <utility>
-
-#include <fst/arcsort.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-enum ArcSortType {
-  ILABEL_SORT,
-  OLABEL_SORT
-};
-
-using ArcSortArgs = std::pair<MutableFstClass *, ArcSortType>;
-
-template <class Arc>
-void ArcSort(ArcSortArgs *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  switch (std::get<1>(*args)) {
-    case ILABEL_SORT: {
-      const ILabelCompare<Arc> icomp;
-      ArcSort(fst, icomp);
-      return;
-    }
-    case OLABEL_SORT: {
-      const OLabelCompare<Arc> ocomp;
-      ArcSort(fst, ocomp);
-      return;
-    }
-  }
-}
-
-void ArcSort(MutableFstClass *ofst, ArcSortType);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_ARCSORT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arg-packs.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arg-packs.h
deleted file mode 100644
index 93cd4a05..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/arg-packs.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// std::pair and std::tuple are used for the arguments of FstClass operations.
-//
-// If a function with a return value is required, use the WithReturnValue
-// template as follows:
-//
-// WithReturnValue<bool, std::tuple<...>>
-
-#ifndef FST_SCRIPT_ARG_PACKS_H_
-#define FST_SCRIPT_ARG_PACKS_H_
-
-#include <type_traits>
-
-namespace fst {
-namespace script {
-
-// Tack this on to an existing type to add a return value. The syntax for
-// accessing the args is then slightly more stilted, as you must do an extra
-// member access (since the args are stored as a member of this class).
-
-template <class Retval, class ArgTuple>
-struct WithReturnValue {
-  // Avoid reference-to-reference if ArgTuple is a reference.
-  using Args = typename std::remove_reference<ArgTuple>::type;
-
-  Retval retval;
-  const Args &args;
-
-  explicit WithReturnValue(const Args &args) : args(args) {}
-};
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_ARG_PACKS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/closure.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/closure.h
deleted file mode 100644
index 7c68604a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/closure.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_CLOSURE_H_
-#define FST_SCRIPT_CLOSURE_H_
-
-#include <utility>
-
-#include <fst/closure.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using ClosureArgs = std::pair<MutableFstClass *, const ClosureType>;
-
-template <class Arc>
-void Closure(ClosureArgs *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  Closure(fst, std::get<1>(*args));
-}
-
-void Closure(MutableFstClass *ofst, ClosureType closure_type);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_CLOSURE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/compile-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/compile-impl.h
deleted file mode 100644
index ba8aebf8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/compile-impl.h
+++ /dev/null
@@ -1,217 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to to compile a binary FST from textual input.
-
-#ifndef FST_SCRIPT_COMPILE_IMPL_H_
-#define FST_SCRIPT_COMPILE_IMPL_H_
-
-#include <iostream>
-#include <memory>
-#include <sstream>
-#include <string>
-#include <vector>
-
-#include <fst/fst.h>
-#include <fst/util.h>
-#include <fst/vector-fst.h>
-#include <unordered_map>
-
-DECLARE_string(fst_field_separator);
-
-namespace fst {
-
-// Compile a binary Fst from textual input, helper class for fstcompile.cc
-// WARNING: Stand-alone use of this class not recommended, most code should
-// read/write using the binary format which is much more efficient.
-template <class Arc>
-class FstCompiler {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  // WARNING: use of negative labels not recommended as it may cause conflicts.
-  // If add_symbols_ is true, then the symbols will be dynamically added to the
-  // symbol tables. This is only useful if you set the (i/o)keep flag to attach
-  // the final symbol table, or use the accessors. (The input symbol tables are
-  // const and therefore not changed.)
-  FstCompiler(std::istream &istrm, const string &source,  // NOLINT
-              const SymbolTable *isyms, const SymbolTable *osyms,
-              const SymbolTable *ssyms, bool accep, bool ikeep,
-              bool okeep, bool nkeep, bool allow_negative_labels = false) {
-    std::unique_ptr<SymbolTable> misyms(isyms ? isyms->Copy() : nullptr);
-    std::unique_ptr<SymbolTable> mosyms(osyms ? osyms->Copy() : nullptr);
-    std::unique_ptr<SymbolTable> mssyms(ssyms ? ssyms->Copy() : nullptr);
-    Init(istrm, source, misyms.get(), mosyms.get(), mssyms.get(), accep,
-         ikeep, okeep, nkeep, allow_negative_labels, false);
-  }
-
-  FstCompiler(std::istream &istrm, const string &source,  // NOLINT
-              SymbolTable *isyms, SymbolTable *osyms, SymbolTable *ssyms,
-              bool accep, bool ikeep, bool okeep, bool nkeep,
-              bool allow_negative_labels, bool add_symbols) {
-    Init(istrm, source, isyms, osyms, ssyms, accep, ikeep, okeep, nkeep,
-         allow_negative_labels, add_symbols);
-  }
-
-  void Init(std::istream &istrm, const string &source,  // NOLINT
-            SymbolTable *isyms, SymbolTable *osyms, SymbolTable *ssyms,
-            bool accep, bool ikeep, bool okeep, bool nkeep,
-            bool allow_negative_labels, bool add_symbols) {
-    nline_ = 0;
-    source_ = source;
-    isyms_ = isyms;
-    osyms_ = osyms;
-    ssyms_ = ssyms;
-    nstates_ = 0;
-    keep_state_numbering_ = nkeep;
-    allow_negative_labels_ = allow_negative_labels;
-    add_symbols_ = add_symbols;
-    bool start_state_populated = false;
-    char line[kLineLen];
-    const string separator = FLAGS_fst_field_separator + "\n";
-    while (istrm.getline(line, kLineLen)) {
-      ++nline_;
-      std::vector<char *> col;
-      SplitToVector(line, separator.c_str(), &col, true);
-      if (col.empty() || col[0][0] == '\0')
-        continue;
-      if (col.size() > 5 || (col.size() > 4 && accep) ||
-          (col.size() == 3 && !accep)) {
-        FSTERROR() << "FstCompiler: Bad number of columns, source = " << source_
-                   << ", line = " << nline_;
-        fst_.SetProperties(kError, kError);
-        return;
-      }
-      StateId s = StrToStateId(col[0]);
-      while (s >= fst_.NumStates()) fst_.AddState();
-      if (!start_state_populated) {
-        fst_.SetStart(s);
-        start_state_populated = true;
-      }
-
-      Arc arc;
-      StateId d = s;
-      switch (col.size()) {
-        case 1:
-          fst_.SetFinal(s, Weight::One());
-          break;
-        case 2:
-          fst_.SetFinal(s, StrToWeight(col[1], true));
-          break;
-        case 3:
-          arc.nextstate = d = StrToStateId(col[1]);
-          arc.ilabel = StrToILabel(col[2]);
-          arc.olabel = arc.ilabel;
-          arc.weight = Weight::One();
-          fst_.AddArc(s, arc);
-          break;
-        case 4:
-          arc.nextstate = d = StrToStateId(col[1]);
-          arc.ilabel = StrToILabel(col[2]);
-          if (accep) {
-            arc.olabel = arc.ilabel;
-            arc.weight = StrToWeight(col[3], true);
-          } else {
-            arc.olabel = StrToOLabel(col[3]);
-            arc.weight = Weight::One();
-          }
-          fst_.AddArc(s, arc);
-          break;
-        case 5:
-          arc.nextstate = d = StrToStateId(col[1]);
-          arc.ilabel = StrToILabel(col[2]);
-          arc.olabel = StrToOLabel(col[3]);
-          arc.weight = StrToWeight(col[4], true);
-          fst_.AddArc(s, arc);
-      }
-      while (d >= fst_.NumStates()) fst_.AddState();
-    }
-    if (ikeep) fst_.SetInputSymbols(isyms);
-    if (okeep) fst_.SetOutputSymbols(osyms);
-  }
-
-  const VectorFst<Arc> &Fst() const { return fst_; }
-
- private:
-  // Maximum line length in text file.
-  static constexpr int kLineLen = 8096;
-
-  StateId StrToId(const char *s, SymbolTable *syms, const char *name,
-                  bool allow_negative = false) const {
-    StateId n = 0;
-    if (syms) {
-      n = (add_symbols_) ? syms->AddSymbol(s) : syms->Find(s);
-      if (n == -1 || (!allow_negative && n < 0)) {
-        FSTERROR() << "FstCompiler: Symbol \"" << s
-                   << "\" is not mapped to any integer " << name
-                   << ", symbol table = " << syms->Name()
-                   << ", source = " << source_ << ", line = " << nline_;
-        fst_.SetProperties(kError, kError);
-      }
-    } else {
-      char *p;
-      n = strtoll(s, &p, 10);
-      if (p < s + strlen(s) || (!allow_negative && n < 0)) {
-        FSTERROR() << "FstCompiler: Bad " << name << " integer = \"" << s
-                   << "\", source = " << source_ << ", line = " << nline_;
-        fst_.SetProperties(kError, kError);
-      }
-    }
-    return n;
-  }
-
-  StateId StrToStateId(const char *s) {
-    StateId n = StrToId(s, ssyms_, "state ID");
-    if (keep_state_numbering_) return n;
-    // Remaps state IDs to make dense set.
-    const auto it = states_.find(n);
-    if (it == states_.end()) {
-      states_[n] = nstates_;
-      return nstates_++;
-    } else {
-      return it->second;
-    }
-  }
-
-  StateId StrToILabel(const char *s) const {
-    return StrToId(s, isyms_, "arc ilabel", allow_negative_labels_);
-  }
-
-  StateId StrToOLabel(const char *s) const {
-    return StrToId(s, osyms_, "arc olabel", allow_negative_labels_);
-  }
-
-  Weight StrToWeight(const char *s, bool allow_zero) const {
-    Weight w;
-    std::istringstream strm(s);
-    strm >> w;
-    if (!strm || (!allow_zero && w == Weight::Zero())) {
-      FSTERROR() << "FstCompiler: Bad weight = \"" << s
-                 << "\", source = " << source_ << ", line = " << nline_;
-      fst_.SetProperties(kError, kError);
-      w = Weight::NoWeight();
-    }
-    return w;
-  }
-
-  mutable VectorFst<Arc> fst_;
-  size_t nline_;
-  string source_;       // Text FST source name.
-  SymbolTable *isyms_;  // ilabel symbol table (not owned).
-  SymbolTable *osyms_;  // olabel symbol table (not owned).
-  SymbolTable *ssyms_;  // slabel symbol table (not owned).
-  std::unordered_map<StateId, StateId> states_;  // State ID map.
-  StateId nstates_;                              // Number of seen states.
-  bool keep_state_numbering_;
-  bool allow_negative_labels_;  // Not recommended; may cause conflicts.
-  bool add_symbols_;            // Add to symbol tables on-the fly.
-
-  FstCompiler(const FstCompiler &) = delete;
-  FstCompiler &operator=(const FstCompiler &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_SCRIPT_COMPILE_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/compile.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/compile.h
deleted file mode 100644
index c82ed477..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/compile.h
+++ /dev/null
@@ -1,98 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_COMPILE_H_
-#define FST_SCRIPT_COMPILE_H_
-
-#include <istream>
-#include <memory>
-
-#include <fst/script/arg-packs.h>
-#include <fst/script/compile-impl.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-// This operation exists in two forms. 1 is a void operation which writes the
-// compiled machine to disk; 2 returns an FstClass. I/O should normally be done
-// using the binary format for efficiency, so users are STRONGLY ENCOURAGED to
-// use 1 or to construct FSTs using the C++ FST mutation operations.
-
-// Note: it is safe to pass these strings as references because
-// this struct is only used to pass them deeper in the call graph.
-// Be sure you understand why this is so before using this struct
-// for anything else!
-struct CompileFstInnerArgs {
-  std::istream &istrm;
-  const string &source;
-  const string &fst_type;
-  const fst::SymbolTable *isyms;
-  const fst::SymbolTable *osyms;
-  const fst::SymbolTable *ssyms;
-  const bool accep;
-  const bool ikeep;
-  const bool okeep;
-  const bool nkeep;
-  const bool allow_negative_labels;
-
-  CompileFstInnerArgs(std::istream &istrm, const string &source,
-                      const string &fst_type, const fst::SymbolTable *isyms,
-                      const fst::SymbolTable *osyms,
-                      const fst::SymbolTable *ssyms, bool accep, bool ikeep,
-                      bool okeep, bool nkeep,
-                      bool allow_negative_labels = false)
-      : istrm(istrm),
-        source(source),
-        fst_type(fst_type),
-        isyms(isyms),
-        osyms(osyms),
-        ssyms(ssyms),
-        accep(accep),
-        ikeep(ikeep),
-        okeep(okeep),
-        nkeep(nkeep),
-        allow_negative_labels(allow_negative_labels) {}
-};
-
-using CompileFstArgs = WithReturnValue<FstClass *, CompileFstInnerArgs>;
-
-template <class Arc>
-void CompileFstInternal(CompileFstArgs *args) {
-  using fst::Convert;
-  using fst::Fst;
-  using fst::FstCompiler;
-  FstCompiler<Arc> fstcompiler(
-      args->args.istrm, args->args.source, args->args.isyms, args->args.osyms,
-      args->args.ssyms, args->args.accep, args->args.ikeep, args->args.okeep,
-      args->args.nkeep, args->args.allow_negative_labels);
-  const Fst<Arc> *fst = &fstcompiler.Fst();
-  std::unique_ptr<const Fst<Arc>> owned_fst;
-  if (args->args.fst_type != "vector") {
-    owned_fst.reset(Convert<Arc>(*fst, args->args.fst_type));
-    if (!owned_fst) {
-      FSTERROR() << "Failed to convert FST to desired type: "
-                 << args->args.fst_type;
-    }
-    fst = owned_fst.get();
-  }
-  args->retval = fst ? new FstClass(*fst) : nullptr;
-}
-
-void CompileFst(std::istream &istrm, const string &source, const string &dest,
-                const string &fst_type, const string &arc_type,
-                const SymbolTable *isyms, const SymbolTable *osyms,
-                const SymbolTable *ssyms, bool accep, bool ikeep, bool okeep,
-                bool nkeep, bool allow_negative_labels);
-
-FstClass *CompileFstInternal(std::istream &istrm, const string &source,
-                             const string &fst_type, const string &arc_type,
-                             const SymbolTable *isyms, const SymbolTable *osyms,
-                             const SymbolTable *ssyms, bool accep, bool ikeep,
-                             bool okeep, bool nkeep,
-                             bool allow_negative_labels);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_COMPILE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/compose.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/compose.h
deleted file mode 100644
index a1735803..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/compose.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_COMPOSE_H_
-#define FST_SCRIPT_COMPOSE_H_
-
-#include <tuple>
-
-#include <fst/compose.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using ComposeArgs = std::tuple<const FstClass &, const FstClass &,
-                               MutableFstClass *, const ComposeOptions &>;
-
-template <class Arc>
-void Compose(ComposeArgs *args) {
-  const Fst<Arc> &ifst1 = *(std::get<0>(*args).GetFst<Arc>());
-  const Fst<Arc> &ifst2 = *(std::get<1>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
-  const auto &opts = std::get<3>(*args);
-  Compose(ifst1, ifst2, ofst, opts);
-}
-
-void Compose(const FstClass &ifst1, const FstClass &ifst2,
-             MutableFstClass *ofst,
-             const ComposeOptions &opts = ComposeOptions());
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_COMPOSE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/concat.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/concat.h
deleted file mode 100644
index 4bf8dc61..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/concat.h
+++ /dev/null
@@ -1,40 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_CONCAT_H_
-#define FST_SCRIPT_CONCAT_H_
-
-#include <utility>
-
-#include <fst/concat.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using ConcatArgs1 = std::pair<MutableFstClass *, const FstClass &>;
-
-template <class Arc>
-void Concat(ConcatArgs1 *args) {
-  MutableFst<Arc> *ofst = std::get<0>(*args)->GetMutableFst<Arc>();
-  const Fst<Arc> &ifst = *(std::get<1>(*args).GetFst<Arc>());
-  Concat(ofst, ifst);
-}
-
-using ConcatArgs2 = std::pair<const FstClass &, MutableFstClass *>;
-
-template <class Arc>
-void Concat(ConcatArgs2 *args) {
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  Concat(ifst, ofst);
-}
-
-void Concat(MutableFstClass *ofst, const FstClass &ifst);
-
-void Concat(const FstClass &ifst, MutableFstClass *ofst);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_CONCAT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/connect.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/connect.h
deleted file mode 100644
index 030102ac..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/connect.h
+++ /dev/null
@@ -1,23 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_CONNECT_H_
-#define FST_SCRIPT_CONNECT_H_
-
-#include <fst/connect.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-template <class Arc>
-void Connect(MutableFstClass *fst) {
-  Connect(fst->GetMutableFst<Arc>());
-}
-
-void Connect(MutableFstClass *fst);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_CONNECT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/convert.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/convert.h
deleted file mode 100644
index 1a6eeaa3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/convert.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_CONVERT_H_
-#define FST_SCRIPT_CONVERT_H_
-
-#include <memory>
-#include <string>
-#include <utility>
-
-#include <fst/register.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using ConvertInnerArgs = std::pair<const FstClass &, const string &>;
-
-using ConvertArgs = WithReturnValue<FstClass *, ConvertInnerArgs>;
-
-template <class Arc>
-void Convert(ConvertArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(args->args).GetFst<Arc>());
-  const string &new_type = std::get<1>(args->args);
-  std::unique_ptr<Fst<Arc>> result(Convert(fst, new_type));
-  args->retval = result ? new FstClass(*result) : nullptr;
-}
-
-FstClass *Convert(const FstClass &fst, const string &new_type);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_CONVERT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/decode.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/decode.h
deleted file mode 100644
index 09f25391..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/decode.h
+++ /dev/null
@@ -1,49 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_DECODE_H_
-#define FST_SCRIPT_DECODE_H_
-
-#include <memory>
-#include <string>
-#include <utility>
-
-#include <fst/encode.h>
-#include <fst/script/encodemapper-class.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using DecodeArgs1 = std::pair<MutableFstClass *, const string &>;
-
-template <class Arc>
-void Decode(DecodeArgs1 *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  std::unique_ptr<EncodeMapper<Arc>> decoder(
-      EncodeMapper<Arc>::Read(std::get<1>(*args), DECODE));
-  if (!decoder) {
-    fst->SetProperties(kError, kError);
-    return;
-  }
-  Decode(fst, *decoder);
-}
-
-using DecodeArgs2 = std::pair<MutableFstClass *, const EncodeMapperClass &>;
-
-template <class Arc>
-void Decode(DecodeArgs2 *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  const EncodeMapper<Arc> &encoder =
-      *(std::get<1>(*args).GetEncodeMapper<Arc>());
-  Decode(fst, encoder);
-}
-
-void Decode(MutableFstClass *fst, const string &coder_fname);
-
-void Decode(MutableFstClass *fst, const EncodeMapperClass &encoder);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_DECODE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/determinize.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/determinize.h
deleted file mode 100644
index 383a8fe4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/determinize.h
+++ /dev/null
@@ -1,59 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_DETERMINIZE_H_
-#define FST_SCRIPT_DETERMINIZE_H_
-
-#include <tuple>
-
-#include <fst/determinize.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-struct DeterminizeOptions {
-  const float delta;
-  const WeightClass &weight_threshold;
-  const int64 state_threshold;
-  const int64 subsequential_label;
-  const DeterminizeType det_type;
-  const bool increment_subsequential_label;
-
-  DeterminizeOptions(float delta, const WeightClass &weight_threshold,
-                     int64 state_threshold = kNoStateId,
-                     int64 subsequential_label = 0,
-                     DeterminizeType det_type = DETERMINIZE_FUNCTIONAL,
-                     bool increment_subsequential_label = false)
-      : delta(delta),
-        weight_threshold(weight_threshold),
-        state_threshold(state_threshold),
-        subsequential_label(subsequential_label),
-        det_type(det_type),
-        increment_subsequential_label(increment_subsequential_label) {}
-};
-
-using DeterminizeArgs = std::tuple<const FstClass &, MutableFstClass *,
-                                   const DeterminizeOptions &>;
-
-template <class Arc>
-void Determinize(DeterminizeArgs *args) {
-  using Weight = typename Arc::Weight;
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  const auto &opts = std::get<2>(*args);
-  const auto weight_threshold = *(opts.weight_threshold.GetWeight<Weight>());
-  const fst::DeterminizeOptions<Arc> detargs(opts.delta, weight_threshold,
-      opts.state_threshold, opts.subsequential_label, opts.det_type,
-      opts.increment_subsequential_label);
-  Determinize(ifst, ofst, detargs);
-}
-
-void Determinize(const FstClass &ifst, MutableFstClass *ofst,
-                 const DeterminizeOptions &opts);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_DETERMINIZE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/difference.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/difference.h
deleted file mode 100644
index 7af6200a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/difference.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_DIFFERENCE_H_
-#define FST_SCRIPT_DIFFERENCE_H_
-
-#include <tuple>
-
-#include <fst/difference.h>
-#include <fst/script/compose.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using DifferenceArgs = std::tuple<const FstClass &, const FstClass &,
-                                  MutableFstClass *, const ComposeOptions &>;
-
-template <class Arc>
-void Difference(DifferenceArgs *args) {
-  const Fst<Arc> &ifst1 = *(std::get<0>(*args).GetFst<Arc>());
-  const Fst<Arc> &ifst2 = *(std::get<1>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
-  const auto &opts = std::get<3>(*args);
-  Difference(ifst1, ifst2, ofst, opts);
-}
-
-void Difference(const FstClass &ifst1, const FstClass &ifst2,
-                MutableFstClass *ofst,
-                const ComposeOptions &opts = ComposeOptions());
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_DIFFERENCE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/disambiguate.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/disambiguate.h
deleted file mode 100644
index acc1fba2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/disambiguate.h
+++ /dev/null
@@ -1,54 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_DISAMBIGUATE_H_
-#define FST_SCRIPT_DISAMBIGUATE_H_
-
-#include <tuple>
-#include <utility>
-
-#include <fst/disambiguate.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-struct DisambiguateOptions {
-  const float delta;
-  const WeightClass &weight_threshold;
-  const int64 state_threshold;
-  const int64 subsequential_label;
-
-  DisambiguateOptions(float delta, const WeightClass &weight_threshold,
-                      int64 state_threshold = kNoStateId,
-                      int64 subsequential_label = 0)
-      : delta(delta),
-        weight_threshold(weight_threshold),
-        state_threshold(state_threshold),
-        subsequential_label(subsequential_label) {}
-};
-
-using DisambiguateArgs = std::tuple<const FstClass &, MutableFstClass *,
-                                     const DisambiguateOptions &>;
-
-template <class Arc>
-void Disambiguate(DisambiguateArgs *args) {
-  using Weight = typename Arc::Weight;
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  const auto &opts = std::get<2>(*args);
-  const auto weight_threshold = *(opts.weight_threshold.GetWeight<Weight>());
-  const fst::DisambiguateOptions<Arc> disargs(opts.delta, weight_threshold,
-                                                  opts.state_threshold,
-                                                  opts.subsequential_label);
-  Disambiguate(ifst, ofst, disargs);
-}
-
-void Disambiguate(const FstClass &ifst, MutableFstClass *ofst,
-                  const DisambiguateOptions &opts);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_DISAMBIGUATE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/draw-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/draw-impl.h
deleted file mode 100644
index f204b2e6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/draw-impl.h
+++ /dev/null
@@ -1,227 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to draw a binary FST by producing a text file in dot format, a helper
-// class to fstdraw.cc.
-
-#ifndef FST_SCRIPT_DRAW_IMPL_H_
-#define FST_SCRIPT_DRAW_IMPL_H_
-
-#include <ostream>
-#include <sstream>
-#include <string>
-
-#include <fst/fst.h>
-#include <fst/util.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-
-// Print a binary FST in GraphViz textual format (helper class for fstdraw.cc).
-// WARNING: Stand-alone use not recommend.
-template <class Arc>
-class FstDrawer {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  FstDrawer(const Fst<Arc> &fst, const SymbolTable *isyms,
-            const SymbolTable *osyms, const SymbolTable *ssyms, bool accep,
-            const string &title, float width, float height, bool portrait,
-            bool vertical, float ranksep, float nodesep, int fontsize,
-            int precision, const string &float_format, bool show_weight_one)
-      : fst_(fst),
-        isyms_(isyms),
-        osyms_(osyms),
-        ssyms_(ssyms),
-        accep_(accep && fst.Properties(kAcceptor, true)),
-        ostrm_(nullptr),
-        title_(title),
-        width_(width),
-        height_(height),
-        portrait_(portrait),
-        vertical_(vertical),
-        ranksep_(ranksep),
-        nodesep_(nodesep),
-        fontsize_(fontsize),
-        precision_(precision),
-        float_format_(float_format),
-        show_weight_one_(show_weight_one) {}
-
-  // Draws FST to an output buffer.
-  void Draw(std::ostream *strm, const string &dest) {
-    ostrm_ = strm;
-    SetStreamState(ostrm_);
-    dest_ = dest;
-    StateId start = fst_.Start();
-    if (start == kNoStateId) return;
-    PrintString("digraph FST {\n");
-    if (vertical_) {
-      PrintString("rankdir = BT;\n");
-    } else {
-      PrintString("rankdir = LR;\n");
-    }
-    PrintString("size = \"");
-    Print(width_);
-    PrintString(",");
-    Print(height_);
-    PrintString("\";\n");
-    if (!dest_.empty()) PrintString("label = \"" + title_ + "\";\n");
-    PrintString("center = 1;\n");
-    if (portrait_) {
-      PrintString("orientation = Portrait;\n");
-    } else {
-      PrintString("orientation = Landscape;\n");
-    }
-    PrintString("ranksep = \"");
-    Print(ranksep_);
-    PrintString("\";\n");
-    PrintString("nodesep = \"");
-    Print(nodesep_);
-    PrintString("\";\n");
-    // Initial state first.
-    DrawState(start);
-    for (StateIterator<Fst<Arc>> siter(fst_); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      if (s != start) DrawState(s);
-    }
-    PrintString("}\n");
-  }
-
- private:
-  void SetStreamState(std::ostream* strm) const {
-    strm->precision(precision_);
-    if (float_format_ == "e")
-        strm->setf(std::ios_base::scientific, std::ios_base::floatfield);
-    if (float_format_ == "f")
-        strm->setf(std::ios_base::fixed, std::ios_base::floatfield);
-    // O.w. defaults to "g" per standard lib.
-  }
-
-  void PrintString(const string &str) const { *ostrm_ << str; }
-
-  // Escapes backslash and double quote if these occur in the string. Dot will
-  // not deal gracefully with these if they are not escaped.
-  static string Escape(const string &str) {
-    string ns;
-    for (char c : str) {
-      if (c == '\\' || c == '"') ns.push_back('\\');
-      ns.push_back(c);
-    }
-    return ns;
-  }
-
-  void PrintId(StateId id, const SymbolTable *syms, const char *name) const {
-    if (syms) {
-      auto symbol = syms->Find(id);
-      if (symbol.empty()) {
-        FSTERROR() << "FstDrawer: Integer " << id
-                   << " is not mapped to any textual symbol"
-                   << ", symbol table = " << syms->Name()
-                   << ", destination = " << dest_;
-        symbol = "?";
-      }
-      PrintString(Escape(symbol));
-    } else {
-      PrintString(std::to_string(id));
-    }
-  }
-
-  void PrintStateId(StateId s) const { PrintId(s, ssyms_, "state ID"); }
-
-  void PrintILabel(Label label) const {
-    PrintId(label, isyms_, "arc input label");
-  }
-
-  void PrintOLabel(Label label) const {
-    PrintId(label, osyms_, "arc output label");
-  }
-
-  void PrintWeight(Weight w) const {
-    // Weight may have double quote characters in it, so escape it.
-    PrintString(Escape(ToString(w)));
-  }
-
-  template <class T>
-  void Print(T t) const { *ostrm_ << t; }
-
-  template <class T>
-  string ToString(T t) const {
-    std::stringstream ss;
-    SetStreamState(&ss);
-    ss << t;
-    return ss.str();
-  }
-
-  void DrawState(StateId s) const {
-    Print(s);
-    PrintString(" [label = \"");
-    PrintStateId(s);
-    const auto weight = fst_.Final(s);
-    if (weight != Weight::Zero()) {
-      if (show_weight_one_ || (weight != Weight::One())) {
-        PrintString("/");
-        PrintWeight(weight);
-      }
-      PrintString("\", shape = doublecircle,");
-    } else {
-      PrintString("\", shape = circle,");
-    }
-    if (s == fst_.Start()) {
-      PrintString(" style = bold,");
-    } else {
-      PrintString(" style = solid,");
-    }
-    PrintString(" fontsize = ");
-    Print(fontsize_);
-    PrintString("]\n");
-    for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      PrintString("\t");
-      Print(s);
-      PrintString(" -> ");
-      Print(arc.nextstate);
-      PrintString(" [label = \"");
-      PrintILabel(arc.ilabel);
-      if (!accep_) {
-        PrintString(":");
-        PrintOLabel(arc.olabel);
-      }
-      if (show_weight_one_ || (arc.weight != Weight::One())) {
-        PrintString("/");
-        PrintWeight(arc.weight);
-      }
-      PrintString("\", fontsize = ");
-      Print(fontsize_);
-      PrintString("];\n");
-    }
-  }
-
-  const Fst<Arc> &fst_;
-  const SymbolTable *isyms_;  // ilabel symbol table.
-  const SymbolTable *osyms_;  // olabel symbol table.
-  const SymbolTable *ssyms_;  // slabel symbol table.
-  bool accep_;                // Print as acceptor when possible.
-  std::ostream *ostrm_;       // Drawn FST destination.
-  string dest_;               // Drawn FST destination name.
-
-  string title_;
-  float width_;
-  float height_;
-  bool portrait_;
-  bool vertical_;
-  float ranksep_;
-  float nodesep_;
-  int fontsize_;
-  int precision_;
-  string float_format_;
-  bool show_weight_one_;
-
-  FstDrawer(const FstDrawer &) = delete;
-  FstDrawer &operator=(const FstDrawer &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_SCRIPT_DRAW_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/draw.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/draw.h
deleted file mode 100644
index cb37df1e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/draw.h
+++ /dev/null
@@ -1,85 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_DRAW_H_
-#define FST_SCRIPT_DRAW_H_
-
-#include <ostream>
-
-#include <fst/script/draw-impl.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-// Note: it is safe to pass these strings as references because
-// this struct is only used to pass them deeper in the call graph.
-// Be sure you understand why this is so before using this struct
-// for anything else!
-struct FstDrawerArgs {
-  const FstClass &fst;
-  const SymbolTable *isyms;
-  const SymbolTable *osyms;
-  const SymbolTable *ssyms;
-  const bool accep;
-  const string &title;
-  const float width;
-  const float height;
-  const bool portrait;
-  const bool vertical;
-  const float ranksep;
-  const float nodesep;
-  const int fontsize;
-  const int precision;
-  const string &float_format;  // NOLINT
-  const bool show_weight_one;
-  std::ostream *ostrm;
-  const string &dest;
-
-  FstDrawerArgs(const FstClass &fst, const SymbolTable *isyms,
-                const SymbolTable *osyms, const SymbolTable *ssyms, bool accep,
-                const string &title, float width, float height, bool portrait,
-                bool vertical, float ranksep, float nodesep, int fontsize,
-                int precision, const string &float_format,
-                bool show_weight_one, std::ostream *ostrm,  const string &dest)
-      : fst(fst),
-        isyms(isyms),
-        osyms(osyms),
-        ssyms(ssyms),
-        accep(accep),
-        title(title),
-        width(width),
-        height(height),
-        portrait(portrait),
-        vertical(vertical),
-        ranksep(ranksep),
-        nodesep(nodesep),
-        fontsize(fontsize),
-        precision(precision),
-        float_format(float_format),
-        show_weight_one(show_weight_one),
-        ostrm(ostrm),
-        dest(dest) {}
-};
-
-template <class Arc>
-void DrawFst(FstDrawerArgs *args) {
-  const Fst<Arc> &fst = *(args->fst.GetFst<Arc>());
-  FstDrawer<Arc> fstdrawer(fst, args->isyms, args->osyms, args->ssyms,
-      args->accep, args->title, args->width, args->height, args->portrait,
-      args->vertical, args->ranksep, args->nodesep, args->fontsize,
-      args->precision, args->float_format, args->show_weight_one);
-  fstdrawer.Draw(args->ostrm, args->dest);
-}
-
-void DrawFst(const FstClass &fst, const SymbolTable *isyms,
-             const SymbolTable *osyms, const SymbolTable *ssyms, bool accep,
-             const string &title, float width, float height, bool portrait,
-             bool vertical, float ranksep, float nodesep, int fontsize,
-             int precision, const string &float_format, bool show_weight_one,
-             std::ostream *ostrm, const string &dest);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_DRAW_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/encode.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/encode.h
deleted file mode 100644
index 6a869680..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/encode.h
+++ /dev/null
@@ -1,51 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_ENCODE_H_
-#define FST_SCRIPT_ENCODE_H_
-
-#include <memory>
-#include <string>
-#include <tuple>
-#include <utility>
-
-#include <fst/encode.h>
-#include <fst/script/encodemapper-class.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using EncodeArgs1 = std::tuple<MutableFstClass *, uint32, bool, const string &>;
-
-template <class Arc>
-void Encode(EncodeArgs1 *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  const string &coder_fname = std::get<3>(*args);
-  // If true, reuse encode from disk. If false, make a new encoder and just use
-  // the filename argument as the destination state.
-  std::unique_ptr<EncodeMapper<Arc>> encoder(
-      std::get<2>(*args) ? EncodeMapper<Arc>::Read(coder_fname, ENCODE)
-                         : new EncodeMapper<Arc>(std::get<1>(*args), ENCODE));
-  Encode(fst, encoder.get());
-  if (!std::get<2>(*args)) encoder->Write(coder_fname);
-}
-
-using EncodeArgs2 = std::pair<MutableFstClass *, EncodeMapperClass *>;
-
-template <class Arc>
-void Encode(EncodeArgs2 *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  EncodeMapper<Arc> *encoder = std::get<1>(*args)->GetEncodeMapper<Arc>();
-  Encode(fst, encoder);
-}
-
-void Encode(MutableFstClass *fst, uint32 flags, bool reuse_encoder,
-            const string &coder_fname);
-
-void Encode(MutableFstClass *fst, EncodeMapperClass *encoder);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_ENCODE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/encodemapper-class.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/encodemapper-class.h
deleted file mode 100644
index b62824f1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/encodemapper-class.h
+++ /dev/null
@@ -1,169 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_ENCODEMAPPER_CLASS_H_
-#define FST_SCRIPT_ENCODEMAPPER_CLASS_H_
-
-#include <memory>
-#include <string>
-#include <iostream>
-
-#include <fst/fstlib.h>
-#include <fst/script/arc-class.h>
-#include <fst/script/fst-class.h>
-
-// Scripting API support for EncodeMapper.
-
-namespace fst {
-namespace script {
-
-// Virtual interface implemented by each concrete EncodeMapperClassImpl<A>.
-class EncodeMapperImplBase {
- public:
-  // Returns an encoded ArcClass.
-  virtual ArcClass operator()(const ArcClass &a) = 0;
-  virtual const string &ArcType() const = 0;
-  virtual uint32 Flags() const = 0;
-  virtual uint64 Properties(uint64 inprops) = 0;
-  virtual EncodeType Type() const = 0;
-  virtual const SymbolTable *InputSymbols() const = 0;
-  virtual const SymbolTable *OutputSymbols() const = 0;
-  virtual void SetInputSymbols(const SymbolTable *syms) = 0;
-  virtual void SetOutputSymbols(const SymbolTable *syms) = 0;
-  virtual const string &WeightType() const = 0;
-  virtual ~EncodeMapperImplBase() {}
-};
-
-// Templated implementation.
-template <class Arc>
-class EncodeMapperClassImpl : public EncodeMapperImplBase {
- public:
-  EncodeMapperClassImpl(uint32 flags, EncodeType type)
-      : encoder_(flags, type) {}
-
-  ArcClass operator()(const ArcClass &a) final;
-
-  const string &ArcType() const final { return Arc::Type(); }
-
-  uint32 Flags() const final { return encoder_.Flags(); }
-
-  uint64 Properties(uint64 inprops) final {
-    return encoder_.Properties(inprops);
-  }
-
-  EncodeType Type() const final { return encoder_.Type(); }
-
-  const SymbolTable *InputSymbols() const final {
-    return encoder_.InputSymbols();
-  }
-
-  const SymbolTable *OutputSymbols() const final {
-    return encoder_.OutputSymbols();
-  }
-
-  void SetInputSymbols(const SymbolTable *syms) final {
-    encoder_.SetInputSymbols(syms);
-  }
-
-  void SetOutputSymbols(const SymbolTable *syms) final {
-    encoder_.SetOutputSymbols(syms);
-  }
-
-  const string &WeightType() const final { return Arc::Weight::Type(); }
-
-  ~EncodeMapperClassImpl() override {}
-
-  EncodeMapper<Arc> *GetImpl() const { return &encoder_; }
-
-  EncodeMapper<Arc> *GetImpl() { return &encoder_; }
-
- private:
-  EncodeMapper<Arc> encoder_;
-};
-
-// This is returned by value because it is very likely to undergo return-value
-// optimization.
-template <class Arc>
-inline ArcClass EncodeMapperClassImpl<Arc>::operator()(const ArcClass &a) {
-  Arc arc(a.ilabel, a.olabel, *(a.weight.GetWeight<typename Arc::Weight>()),
-          a.nextstate);
-  return ArcClass(encoder_(arc));
-}
-
-class EncodeMapperClass;
-
-using InitEncodeMapperClassArgs =
-    std::tuple<uint32, EncodeType, EncodeMapperClass *>;
-
-class EncodeMapperClass {
- public:
-  EncodeMapperClass(const string &arc_type, uint32 flags, EncodeType type);
-
-  template <class Arc>
-  EncodeMapperClass(uint32 flags, EncodeType type)
-      : impl_(new EncodeMapperClassImpl<Arc>(flags, type)) {}
-
-  ArcClass operator()(const ArcClass &arc) { return (*impl_)(arc); }
-
-  const string &ArcType() const { return impl_->ArcType(); }
-
-  uint32 Flags() const { return impl_->Flags(); }
-
-  uint64 Properties(uint64 inprops) { return impl_->Properties(inprops); }
-
-  EncodeType Type() const { return impl_->Type(); }
-
-  const SymbolTable *InputSymbols() const { return impl_->InputSymbols(); }
-
-  const SymbolTable *OutputSymbols() const { return impl_->OutputSymbols(); }
-
-  void SetInputSymbols(const SymbolTable *syms) {
-    impl_->SetInputSymbols(syms);
-  }
-
-  void SetOutputSymbols(const SymbolTable *syms) {
-    impl_->SetOutputSymbols(syms);
-  }
-
-  const string &WeightType() const { return impl_->WeightType(); }
-
-  template <class Arc>
-  friend void InitEncodeMapperClass(InitEncodeMapperClassArgs *args);
-
-  // Naturally, this exists in non-const and const forms. Encoding arcs or FSTs
-  // mutates the underlying encoder; decoding them does not.
-
-  template <class Arc>
-  EncodeMapper<Arc> *GetEncodeMapper() {
-    if (Arc::Type() != ArcType()) {
-      return nullptr;
-    } else {
-      auto *typed_impl = static_cast<EncodeMapperClassImpl<Arc> *>(impl_.get());
-      return typed_impl->GetImpl();
-    }
-  }
-
-  template <class Arc>
-  const EncodeMapper<Arc> *GetEncodeMapper() const {
-    if (Arc::Type() != ArcType()) {
-      return nullptr;
-    } else {
-      auto *typed_impl = static_cast<EncodeMapperClassImpl<Arc> *>(impl_.get());
-      return typed_impl->GetImpl();
-    }
-  }
-
- private:
-  std::unique_ptr<EncodeMapperImplBase> impl_;
-};
-
-template <class Arc>
-void InitEncodeMapperClass(InitEncodeMapperClassArgs *args) {
-  std::get<2>(*args)->impl_.reset(
-      new EncodeMapperClassImpl<Arc>(std::get<0>(*args), std::get<1>(*args)));
-}
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_ENCODEMAPPER_CLASS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/epsnormalize.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/epsnormalize.h
deleted file mode 100644
index b55fefae..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/epsnormalize.h
+++ /dev/null
@@ -1,31 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_EPSNORMALIZE_H_
-#define FST_SCRIPT_EPSNORMALIZE_H_
-
-#include <tuple>
-
-#include <fst/epsnormalize.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using EpsNormalizeArgs = std::tuple<const FstClass &, MutableFstClass *,
-                                    EpsNormalizeType>;
-
-template <class Arc>
-void EpsNormalize(EpsNormalizeArgs *args) {
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  EpsNormalize(ifst, ofst, std::get<2>(*args));
-}
-
-void EpsNormalize(const FstClass &ifst, MutableFstClass *ofst,
-                  EpsNormalizeType norm_type = EPS_NORM_INPUT);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_EPSNORMALIZE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/equal.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/equal.h
deleted file mode 100644
index 79ea9aa4..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/equal.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_EQUAL_H_
-#define FST_SCRIPT_EQUAL_H_
-
-#include <tuple>
-
-#include <fst/equal.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using EqualInnerArgs = std::tuple<const FstClass &, const FstClass &, float>;
-
-using EqualArgs = WithReturnValue<bool, EqualInnerArgs>;
-
-template <class Arc>
-void Equal(EqualArgs *args) {
-  const Fst<Arc> &fst1 = *(std::get<0>(args->args).GetFst<Arc>());
-  const Fst<Arc> &fst2 = *(std::get<1>(args->args).GetFst<Arc>());
-  args->retval = Equal(fst1, fst2, std::get<2>(args->args));
-}
-
-bool Equal(const FstClass &fst1, const FstClass &fst2, float delta = kDelta);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_EQUAL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/equivalent.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/equivalent.h
deleted file mode 100644
index 7cdff45e..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/equivalent.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_EQUIVALENT_H_
-#define FST_SCRIPT_EQUIVALENT_H_
-
-#include <tuple>
-
-#include <fst/equivalent.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using EquivalentInnerArgs = std::tuple<const FstClass &, const FstClass &,
-                                       float>;
-
-using EquivalentArgs = WithReturnValue<bool, EquivalentInnerArgs>;
-
-template <class Arc>
-void Equivalent(EquivalentArgs *args) {
-  const Fst<Arc> &fst1 = *(std::get<0>(args->args).GetFst<Arc>());
-  const Fst<Arc> &fst2 = *(std::get<1>(args->args).GetFst<Arc>());
-  args->retval = Equivalent(fst1, fst2, std::get<2>(args->args));
-}
-
-bool Equivalent(const FstClass &fst1, const FstClass &fst2,
-                float delta = kDelta);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_EQUIVALENT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/fst-class.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/fst-class.h
deleted file mode 100644
index 6378d0a3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/fst-class.h
+++ /dev/null
@@ -1,533 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_FST_CLASS_H_
-#define FST_SCRIPT_FST_CLASS_H_
-
-#include <algorithm>
-#include <limits>
-#include <string>
-#include <type_traits>
-
-#include <fst/expanded-fst.h>
-#include <fst/fst.h>
-#include <fst/mutable-fst.h>
-#include <fst/vector-fst.h>
-#include <fst/script/arc-class.h>
-#include <fst/script/weight-class.h>
-
-// Classes to support "boxing" all existing types of FST arcs in a single
-// FstClass which hides the arc types. This allows clients to load
-// and work with FSTs without knowing the arc type. These classes are only
-// recommended for use in high-level scripting applications. Most users should
-// use the lower-level templated versions corresponding to these classes.
-
-namespace fst {
-namespace script {
-
-// Abstract base class defining the set of functionalities implemented in all
-// impls and passed through by all bases. Below FstClassBase the class
-// hierarchy bifurcates; FstClassImplBase serves as the base class for all
-// implementations (of which FstClassImpl is currently the only one) and
-// FstClass serves as the base class for all interfaces.
-
-class FstClassBase {
- public:
-  virtual const string &ArcType() const = 0;
-  virtual WeightClass Final(int64) const = 0;
-  virtual const string &FstType() const = 0;
-  virtual const SymbolTable *InputSymbols() const = 0;
-  virtual size_t NumArcs(int64) const = 0;
-  virtual size_t NumInputEpsilons(int64) const = 0;
-  virtual size_t NumOutputEpsilons(int64) const = 0;
-  virtual const SymbolTable *OutputSymbols() const = 0;
-  virtual uint64 Properties(uint64, bool) const = 0;
-  virtual int64 Start() const = 0;
-  virtual const string &WeightType() const = 0;
-  virtual bool ValidStateId(int64) const = 0;
-  virtual bool Write(const string &fname) const = 0;
-  virtual bool Write(std::ostream &, const FstWriteOptions &) const = 0;
-  virtual ~FstClassBase() {}
-};
-
-// Adds all the MutableFst methods.
-class FstClassImplBase : public FstClassBase {
- public:
-  virtual bool AddArc(int64, const ArcClass &) = 0;
-  virtual int64 AddState() = 0;
-  virtual FstClassImplBase *Copy() = 0;
-  virtual bool DeleteArcs(int64, size_t) = 0;
-  virtual bool DeleteArcs(int64) = 0;
-  virtual bool DeleteStates(const std::vector<int64> &) = 0;
-  virtual void DeleteStates() = 0;
-  virtual SymbolTable *MutableInputSymbols() = 0;
-  virtual SymbolTable *MutableOutputSymbols() = 0;
-  virtual int64 NumStates() const = 0;
-  virtual bool ReserveArcs(int64, size_t) = 0;
-  virtual void ReserveStates(int64) = 0;
-  virtual void SetInputSymbols(SymbolTable *) = 0;
-  virtual bool SetFinal(int64, const WeightClass &) = 0;
-  virtual void SetOutputSymbols(SymbolTable *) = 0;
-  virtual void SetProperties(uint64, uint64) = 0;
-  virtual bool SetStart(int64) = 0;
-  ~FstClassImplBase() override {}
-};
-
-// Containiner class wrapping an Fst<Arc>, hiding its arc type. Whether this
-// Fst<Arc> pointer refers to a special kind of FST (e.g. a MutableFst) is
-// known by the type of interface class that owns the pointer to this
-// container.
-
-template <class Arc>
-class FstClassImpl : public FstClassImplBase {
- public:
-  explicit FstClassImpl(Fst<Arc> *impl, bool should_own = false)
-      : impl_(should_own ? impl : impl->Copy()) {}
-
-  explicit FstClassImpl(const Fst<Arc> &impl) : impl_(impl.Copy()) {}
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  bool AddArc(int64 s, const ArcClass &ac) final {
-    if (!ValidStateId(s)) return false;
-    // Note that we do not check that the destination state is valid, so users
-    // can add arcs before they add the corresponding states. Verify can be
-    // used to determine whether any arc has a nonexisting destination.
-    Arc arc(ac.ilabel, ac.olabel, *ac.weight.GetWeight<typename Arc::Weight>(),
-            ac.nextstate);
-    static_cast<MutableFst<Arc> *>(impl_.get())->AddArc(s, arc);
-    return true;
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  int64 AddState() final {
-    return static_cast<MutableFst<Arc> *>(impl_.get())->AddState();
-  }
-
-  const string &ArcType() const final { return Arc::Type(); }
-
-  FstClassImpl *Copy() final { return new FstClassImpl<Arc>(impl_.get()); }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  bool DeleteArcs(int64 s, size_t n) final {
-    if (!ValidStateId(s)) return false;
-    static_cast<MutableFst<Arc> *>(impl_.get())->DeleteArcs(s, n);
-    return true;
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  bool DeleteArcs(int64 s) final {
-    if (!ValidStateId(s)) return false;
-    static_cast<MutableFst<Arc> *>(impl_.get())->DeleteArcs(s);
-    return true;
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  bool DeleteStates(const std::vector<int64> &dstates) final {
-    for (const auto &state : dstates)
-      if (!ValidStateId(state)) return false;
-    // Warning: calling this method with any integers beyond the precision of
-    // the underlying FST will result in truncation.
-    std::vector<typename Arc::StateId> typed_dstates(dstates.size());
-    std::copy(dstates.begin(), dstates.end(), typed_dstates.begin());
-    static_cast<MutableFst<Arc> *>(impl_.get())->DeleteStates(typed_dstates);
-    return true;
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  void DeleteStates() final {
-    static_cast<MutableFst<Arc> *>(impl_.get())->DeleteStates();
-  }
-
-  WeightClass Final(int64 s) const final {
-    if (!ValidStateId(s)) return WeightClass::NoWeight(WeightType());
-    WeightClass w(impl_->Final(s));
-    return w;
-  }
-
-  const string &FstType() const final { return impl_->Type(); }
-
-  const SymbolTable *InputSymbols() const final {
-    return impl_->InputSymbols();
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  SymbolTable *MutableInputSymbols() final {
-    return static_cast<MutableFst<Arc> *>(impl_.get())->MutableInputSymbols();
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  SymbolTable *MutableOutputSymbols() final {
-    return static_cast<MutableFst<Arc> *>(impl_.get())->MutableOutputSymbols();
-  }
-
-  // Signals failure by returning size_t max.
-  size_t NumArcs(int64 s) const final {
-    return ValidStateId(s) ? impl_->NumArcs(s)
-                           : std::numeric_limits<size_t>::max();
-  }
-
-  // Signals failure by returning size_t max.
-  size_t NumInputEpsilons(int64 s) const final {
-    return ValidStateId(s) ? impl_->NumInputEpsilons(s)
-                           : std::numeric_limits<size_t>::max();
-  }
-
-  // Signals failure by returning size_t max.
-  size_t NumOutputEpsilons(int64 s) const final {
-    return ValidStateId(s) ? impl_->NumOutputEpsilons(s)
-                           : std::numeric_limits<size_t>::max();
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  int64 NumStates() const final {
-    return static_cast<MutableFst<Arc> *>(impl_.get())->NumStates();
-  }
-
-  uint64 Properties(uint64 mask, bool test) const final {
-    return impl_->Properties(mask, test);
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  bool ReserveArcs(int64 s, size_t n) final {
-    if (!ValidStateId(s)) return false;
-    static_cast<MutableFst<Arc> *>(impl_.get())->ReserveArcs(s, n);
-    return true;
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  void ReserveStates(int64 s) final {
-    static_cast<MutableFst<Arc> *>(impl_.get())->ReserveStates(s);
-  }
-
-  const SymbolTable *OutputSymbols() const final {
-    return impl_->OutputSymbols();
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  void SetInputSymbols(SymbolTable *isyms) final {
-    static_cast<MutableFst<Arc> *>(impl_.get())->SetInputSymbols(isyms);
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  bool SetFinal(int64 s, const WeightClass &weight) final {
-    if (!ValidStateId(s)) return false;
-    static_cast<MutableFst<Arc> *>(impl_.get())
-        ->SetFinal(s, *weight.GetWeight<typename Arc::Weight>());
-    return true;
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  void SetOutputSymbols(SymbolTable *osyms) final {
-    static_cast<MutableFst<Arc> *>(impl_.get())->SetOutputSymbols(osyms);
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  void SetProperties(uint64 props, uint64 mask) final {
-    static_cast<MutableFst<Arc> *>(impl_.get())->SetProperties(props, mask);
-  }
-
-  // Warning: calling this method casts the FST to a mutable FST.
-  bool SetStart(int64 s) final {
-    if (!ValidStateId(s)) return false;
-    static_cast<MutableFst<Arc> *>(impl_.get())->SetStart(s);
-    return true;
-  }
-
-  int64 Start() const final { return impl_->Start(); }
-
-  bool ValidStateId(int64 s) const final {
-    // This cowardly refuses to count states if the FST is not yet expanded.
-    if (!Properties(kExpanded, true)) {
-      FSTERROR() << "Cannot get number of states for unexpanded FST";
-      return false;
-    }
-    // If the FST is already expanded, CountStates calls NumStates.
-    if (s < 0 || s >= CountStates(*impl_)) {
-      FSTERROR() << "State ID " << s << " not valid";
-      return false;
-    }
-    return true;
-  }
-
-  const string &WeightType() const final { return Arc::Weight::Type(); }
-
-  bool Write(const string &fname) const final { return impl_->Write(fname); }
-
-  bool Write(std::ostream &ostr, const FstWriteOptions &opts) const final {
-    return impl_->Write(ostr, opts);
-  }
-
-  ~FstClassImpl() override {}
-
-  Fst<Arc> *GetImpl() const { return impl_.get(); }
-
- private:
-  std::unique_ptr<Fst<Arc>> impl_;
-};
-
-// BASE CLASS DEFINITIONS
-
-class MutableFstClass;
-
-class FstClass : public FstClassBase {
- public:
-  FstClass() : impl_(nullptr) {}
-
-  template <class Arc>
-  explicit FstClass(const Fst<Arc> &fst) : impl_(new FstClassImpl<Arc>(fst)) {}
-
-  FstClass(const FstClass &other)
-      : impl_(other.impl_ == nullptr ? nullptr : other.impl_->Copy()) {}
-
-  FstClass &operator=(const FstClass &other) {
-    impl_.reset(other.impl_ == nullptr ? nullptr : other.impl_->Copy());
-    return *this;
-  }
-
-  WeightClass Final(int64 s) const final { return impl_->Final(s); }
-
-  const string &ArcType() const final { return impl_->ArcType(); }
-
-  const string &FstType() const final { return impl_->FstType(); }
-
-  const SymbolTable *InputSymbols() const final {
-    return impl_->InputSymbols();
-  }
-
-  size_t NumArcs(int64 s) const final { return impl_->NumArcs(s); }
-
-  size_t NumInputEpsilons(int64 s) const final {
-    return impl_->NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(int64 s) const final {
-    return impl_->NumOutputEpsilons(s);
-  }
-
-  const SymbolTable *OutputSymbols() const final {
-    return impl_->OutputSymbols();
-  }
-
-  uint64 Properties(uint64 mask, bool test) const final {
-    // Special handling for FSTs with a null impl.
-    if (!impl_) return kError & mask;
-    return impl_->Properties(mask, test);
-  }
-
-  static FstClass *Read(const string &fname);
-
-  static FstClass *Read(std::istream &istr, const string &source);
-
-  static FstClass *ReadFromString(const string &fst_string);
-
-  int64 Start() const final { return impl_->Start(); }
-
-  bool ValidStateId(int64 s) const final { return impl_->ValidStateId(s); }
-
-  const string &WeightType() const final { return impl_->WeightType(); }
-
-  // Helper that logs an ERROR if the weight type of an FST and a WeightClass
-  // don't match.
-
-  bool WeightTypesMatch(const WeightClass &weight, const string &op_name) const;
-
-  bool Write(const string &fname) const final { return impl_->Write(fname); }
-
-  bool Write(std::ostream &ostr, const FstWriteOptions &opts) const final {
-    return impl_->Write(ostr, opts);
-  }
-
-  const string WriteToString() const;
-
-  ~FstClass() override {}
-
-  // These methods are required by IO registration.
-
-  template <class Arc>
-  static FstClassImplBase *Convert(const FstClass &other) {
-    FSTERROR() << "Doesn't make sense to convert any class to type FstClass";
-    return nullptr;
-  }
-
-  template <class Arc>
-  static FstClassImplBase *Create() {
-    FSTERROR() << "Doesn't make sense to create an FstClass with a "
-               << "particular arc type";
-    return nullptr;
-  }
-
-  template <class Arc>
-  const Fst<Arc> *GetFst() const {
-    if (Arc::Type() != ArcType()) {
-      return nullptr;
-    } else {
-      FstClassImpl<Arc> *typed_impl =
-          static_cast<FstClassImpl<Arc> *>(impl_.get());
-      return typed_impl->GetImpl();
-    }
-  }
-
-  template <class Arc>
-  static FstClass *Read(std::istream &stream, const FstReadOptions &opts) {
-    if (!opts.header) {
-      LOG(ERROR) << "FstClass::Read: Options header not specified";
-      return nullptr;
-    }
-    const FstHeader &hdr = *opts.header;
-    if (hdr.Properties() & kMutable) {
-      return ReadTypedFst<MutableFstClass, MutableFst<Arc>>(stream, opts);
-    } else {
-      return ReadTypedFst<FstClass, Fst<Arc>>(stream, opts);
-    }
-  }
-
- protected:
-  explicit FstClass(FstClassImplBase *impl) : impl_(impl) {}
-
-  const FstClassImplBase *GetImpl() const { return impl_.get(); }
-
-  FstClassImplBase *GetImpl() { return impl_.get(); }
-
-  // Generic template method for reading an arc-templated FST of type
-  // UnderlyingT, and returning it wrapped as FstClassT, with appropriat
-  // error checking. Called from arc-templated Read() static methods.
-  template <class FstClassT, class UnderlyingT>
-  static FstClassT *ReadTypedFst(std::istream &stream,
-                                 const FstReadOptions &opts) {
-    std::unique_ptr<UnderlyingT> u(UnderlyingT::Read(stream, opts));
-    return u ? new FstClassT(*u) : nullptr;
-  }
-
- private:
-  std::unique_ptr<FstClassImplBase> impl_;
-};
-
-// Specific types of FstClass with special properties
-
-class MutableFstClass : public FstClass {
- public:
-  bool AddArc(int64 s, const ArcClass &ac) {
-    if (!WeightTypesMatch(ac.weight, "AddArc")) return false;
-    return GetImpl()->AddArc(s, ac);
-  }
-
-  int64 AddState() { return GetImpl()->AddState(); }
-
-  bool DeleteArcs(int64 s, size_t n) { return GetImpl()->DeleteArcs(s, n); }
-
-  bool DeleteArcs(int64 s) { return GetImpl()->DeleteArcs(s); }
-
-  bool DeleteStates(const std::vector<int64> &dstates) {
-    return GetImpl()->DeleteStates(dstates);
-  }
-
-  void DeleteStates() { GetImpl()->DeleteStates(); }
-
-  SymbolTable *MutableInputSymbols() {
-    return GetImpl()->MutableInputSymbols();
-  }
-
-  SymbolTable *MutableOutputSymbols() {
-    return GetImpl()->MutableOutputSymbols();
-  }
-
-  int64 NumStates() const { return GetImpl()->NumStates(); }
-
-  bool ReserveArcs(int64 s, size_t n) { return GetImpl()->ReserveArcs(s, n); }
-
-  void ReserveStates(int64 s) { GetImpl()->ReserveStates(s); }
-
-  static MutableFstClass *Read(const string &fname, bool convert = false);
-
-  void SetInputSymbols(SymbolTable *isyms) {
-    GetImpl()->SetInputSymbols(isyms);
-  }
-
-  bool SetFinal(int64 s, const WeightClass &weight) {
-    if (!WeightTypesMatch(weight, "SetFinal")) return false;
-    return GetImpl()->SetFinal(s, weight);
-  }
-
-  void SetOutputSymbols(SymbolTable *osyms) {
-    GetImpl()->SetOutputSymbols(osyms);
-  }
-
-  void SetProperties(uint64 props, uint64 mask) {
-    GetImpl()->SetProperties(props, mask);
-  }
-
-  bool SetStart(int64 s) { return GetImpl()->SetStart(s); }
-
-  template <class Arc>
-  explicit MutableFstClass(const MutableFst<Arc> &fst) : FstClass(fst) {}
-
-  // These methods are required by IO registration.
-
-  template <class Arc>
-  static FstClassImplBase *Convert(const FstClass &other) {
-    FSTERROR() << "Doesn't make sense to convert any class to type "
-               << "MutableFstClass";
-    return nullptr;
-  }
-
-  template <class Arc>
-  static FstClassImplBase *Create() {
-    FSTERROR() << "Doesn't make sense to create a MutableFstClass with a "
-               << "particular arc type";
-    return nullptr;
-  }
-
-  template <class Arc>
-  MutableFst<Arc> *GetMutableFst() {
-    Fst<Arc> *fst = const_cast<Fst<Arc> *>(this->GetFst<Arc>());
-    MutableFst<Arc> *mfst = static_cast<MutableFst<Arc> *>(fst);
-    return mfst;
-  }
-
-  template <class Arc>
-  static MutableFstClass *Read(std::istream &stream,
-                               const FstReadOptions &opts) {
-    std::unique_ptr<MutableFst<Arc>> mfst(MutableFst<Arc>::Read(stream, opts));
-    return mfst ? new MutableFstClass(*mfst) : nullptr;
-  }
-
- protected:
-  explicit MutableFstClass(FstClassImplBase *impl) : FstClass(impl) {}
-};
-
-class VectorFstClass : public MutableFstClass {
- public:
-  explicit VectorFstClass(FstClassImplBase *impl) : MutableFstClass(impl) {}
-
-  explicit VectorFstClass(const FstClass &other);
-
-  explicit VectorFstClass(const string &arc_type);
-
-  static VectorFstClass *Read(const string &fname);
-
-  template <class Arc>
-  static VectorFstClass *Read(std::istream &stream,
-                              const FstReadOptions &opts) {
-    std::unique_ptr<VectorFst<Arc>> mfst(VectorFst<Arc>::Read(stream, opts));
-    return mfst ? new VectorFstClass(*mfst) : nullptr;
-  }
-
-  template <class Arc>
-  explicit VectorFstClass(const VectorFst<Arc> &fst) : MutableFstClass(fst) {}
-
-  template <class Arc>
-  static FstClassImplBase *Convert(const FstClass &other) {
-    return new FstClassImpl<Arc>(new VectorFst<Arc>(*other.GetFst<Arc>()),
-                                 true);
-  }
-
-  template <class Arc>
-  static FstClassImplBase *Create() {
-    return new FstClassImpl<Arc>(new VectorFst<Arc>(), true);
-  }
-};
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_FST_CLASS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/fstscript-decl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/fstscript-decl.h
deleted file mode 100644
index 294d0159..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/fstscript-decl.h
+++ /dev/null
@@ -1,32 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Forward declarations for the FST and FST script classes.
-
-#ifndef FST_SCRIPT_FSTSCRIPT_DECL_H_
-#define FST_SCRIPT_FSTSCRIPT_DECL_H_
-
-#include <fst/fst-decl.h>
-
-namespace fst {
-namespace script {
-
-class ArcClass;
-
-class ArcIteratorClass;
-class MutableArcIteratorClass;
-
-class EncodeMapperClass;
-
-class FstClass;
-class MutableFstClass;
-class VectorFstClass;
-
-class StateIteratorClass;
-
-class WeightClass;
-
-}  // namespace script
-}  // namespace fst;
-
-#endif  // FST_SCRIPT_FSTSCRIPT_DECL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/fstscript.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/fstscript.h
deleted file mode 100644
index 45f16175..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/fstscript.h
+++ /dev/null
@@ -1,155 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// The FST script interface permits users to interact with FSTs without knowing
-// their arc type. It does this by mapping compile-time polymorphism (in the
-// form of a arc-templated FST types) onto a shared virtual interface. It also
-// supports arc extension via a DSO interface. Due to the overhead of virtual
-// dispatch and registered function lookups, the script API is somewhat slower
-// then library API provided by types like StdVectorFst, but has the advantage
-// that it is designed not to crash (and to provide useful debugging
-// information) upon common user errors like passing invalid indices or
-// attempting comparison of incompatible FSTs. It is used both by the FST
-// binaries and the Python extension.
-//
-// This header includes all of the FST script functionality.
-
-#ifndef FST_SCRIPT_FSTSCRIPT_H_
-#define FST_SCRIPT_FSTSCRIPT_H_
-
-// Major classes
-#include <fst/script/arciterator-class.h>
-#include <fst/script/encodemapper-class.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/stateiterator-class.h>
-#include <fst/script/text-io.h>
-#include <fst/script/weight-class.h>
-
-// Flag-to-enum parsers.
-#include <fst/script/getters.h>
-// Templates like Operation<> and Apply<>.
-#include <fst/script/script-impl.h>
-
-// Operations.
-#include <fst/script/arcsort.h>
-#include <fst/script/closure.h>
-#include <fst/script/compile.h>
-#include <fst/script/compose.h>
-#include <fst/script/concat.h>
-#include <fst/script/connect.h>
-#include <fst/script/convert.h>
-#include <fst/script/decode.h>
-#include <fst/script/determinize.h>
-#include <fst/script/difference.h>
-#include <fst/script/disambiguate.h>
-#include <fst/script/draw.h>
-#include <fst/script/encode.h>
-#include <fst/script/epsnormalize.h>
-#include <fst/script/equal.h>
-#include <fst/script/equivalent.h>
-#include <fst/script/info.h>
-#include <fst/script/intersect.h>
-#include <fst/script/invert.h>
-#include <fst/script/isomorphic.h>
-#include <fst/script/map.h>
-#include <fst/script/minimize.h>
-#include <fst/script/print.h>
-#include <fst/script/project.h>
-#include <fst/script/prune.h>
-#include <fst/script/push.h>
-#include <fst/script/randequivalent.h>
-#include <fst/script/randgen.h>
-#include <fst/script/relabel.h>
-#include <fst/script/replace.h>
-#include <fst/script/reverse.h>
-#include <fst/script/reweight.h>
-#include <fst/script/rmepsilon.h>
-#include <fst/script/shortest-distance.h>
-#include <fst/script/shortest-path.h>
-#include <fst/script/synchronize.h>
-#include <fst/script/topsort.h>
-#include <fst/script/union.h>
-#include <fst/script/verify.h>
-
-// This class is necessary because registering each of the operations
-// separately overfills the stack, as there's so many of them.
-namespace fst {
-namespace script {
-
-template <class Arc>
-class AllFstOperationsRegisterer {
- public:
-  AllFstOperationsRegisterer() {
-    RegisterBatch1();
-    RegisterBatch2();
-  }
-
- private:
-  void RegisterBatch1() {
-    REGISTER_FST_OPERATION(ArcSort, Arc, ArcSortArgs);
-    REGISTER_FST_OPERATION(Closure, Arc, ClosureArgs);
-    REGISTER_FST_OPERATION(CompileFstInternal, Arc, CompileFstArgs);
-    REGISTER_FST_OPERATION(Compose, Arc, ComposeArgs);
-    REGISTER_FST_OPERATION(Concat, Arc, ConcatArgs1);
-    REGISTER_FST_OPERATION(Concat, Arc, ConcatArgs2);
-    REGISTER_FST_OPERATION(Connect, Arc, MutableFstClass);
-    REGISTER_FST_OPERATION(Convert, Arc, ConvertArgs);
-    REGISTER_FST_OPERATION(Decode, Arc, DecodeArgs1);
-    REGISTER_FST_OPERATION(Decode, Arc, DecodeArgs2);
-    REGISTER_FST_OPERATION(Determinize, Arc, DeterminizeArgs);
-    REGISTER_FST_OPERATION(Difference, Arc, DifferenceArgs);
-    REGISTER_FST_OPERATION(Disambiguate, Arc, DisambiguateArgs);
-    REGISTER_FST_OPERATION(DrawFst, Arc, FstDrawerArgs);
-    REGISTER_FST_OPERATION(Encode, Arc, EncodeArgs1);
-    REGISTER_FST_OPERATION(Encode, Arc, EncodeArgs2);
-    REGISTER_FST_OPERATION(EpsNormalize, Arc, EpsNormalizeArgs);
-    REGISTER_FST_OPERATION(Equal, Arc, EqualArgs);
-    REGISTER_FST_OPERATION(Equivalent, Arc, EquivalentArgs);
-    REGISTER_FST_OPERATION(PrintFstInfo, Arc, InfoArgs);
-    REGISTER_FST_OPERATION(GetFstInfo, Arc, GetInfoArgs);
-    REGISTER_FST_OPERATION(InitArcIteratorClass, Arc,
-                           InitArcIteratorClassArgs);
-    REGISTER_FST_OPERATION(InitEncodeMapperClass, Arc,
-                           InitEncodeMapperClassArgs);
-    REGISTER_FST_OPERATION(InitMutableArcIteratorClass, Arc,
-                           InitMutableArcIteratorClassArgs);
-    REGISTER_FST_OPERATION(InitStateIteratorClass, Arc,
-                           InitStateIteratorClassArgs);
-  }
-
-  void RegisterBatch2() {
-    REGISTER_FST_OPERATION(Intersect, Arc, IntersectArgs);
-    REGISTER_FST_OPERATION(Invert, Arc, MutableFstClass);
-    REGISTER_FST_OPERATION(Map, Arc, MapArgs);
-    REGISTER_FST_OPERATION(Minimize, Arc, MinimizeArgs);
-    REGISTER_FST_OPERATION(PrintFst, Arc, FstPrinterArgs);
-    REGISTER_FST_OPERATION(Project, Arc, ProjectArgs);
-    REGISTER_FST_OPERATION(Prune, Arc, PruneArgs1);
-    REGISTER_FST_OPERATION(Prune, Arc, PruneArgs2);
-    REGISTER_FST_OPERATION(Push, Arc, PushArgs1);
-    REGISTER_FST_OPERATION(Push, Arc, PushArgs2);
-    REGISTER_FST_OPERATION(RandEquivalent, Arc, RandEquivalentArgs);
-    REGISTER_FST_OPERATION(RandGen, Arc, RandGenArgs);
-    REGISTER_FST_OPERATION(Relabel, Arc, RelabelArgs1);
-    REGISTER_FST_OPERATION(Relabel, Arc, RelabelArgs2);
-    REGISTER_FST_OPERATION(Replace, Arc, ReplaceArgs);
-    REGISTER_FST_OPERATION(Reverse, Arc, ReverseArgs);
-    REGISTER_FST_OPERATION(Reweight, Arc, ReweightArgs);
-    REGISTER_FST_OPERATION(RmEpsilon, Arc, RmEpsilonArgs);
-    REGISTER_FST_OPERATION(ShortestDistance, Arc, ShortestDistanceArgs1);
-    REGISTER_FST_OPERATION(ShortestDistance, Arc, ShortestDistanceArgs2);
-    REGISTER_FST_OPERATION(ShortestPath, Arc, ShortestPathArgs);
-    REGISTER_FST_OPERATION(Synchronize, Arc, SynchronizeArgs);
-    REGISTER_FST_OPERATION(TopSort, Arc, TopSortArgs);
-    REGISTER_FST_OPERATION(Union, Arc, UnionArgs);
-    REGISTER_FST_OPERATION(Verify, Arc, VerifyArgs);
-  }
-};
-
-}  // namespace script
-}  // namespace fst
-
-#define REGISTER_FST_OPERATIONS(Arc) \
-  AllFstOperationsRegisterer<Arc> register_all_fst_operations##Arc;
-
-#endif  // FST_SCRIPT_FSTSCRIPT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/getters.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/getters.h
deleted file mode 100644
index 5cd727e8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/getters.h
+++ /dev/null
@@ -1,76 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Getters for converting command-line arguments into the appropriate enums
-// or bitmasks, with the simplest ones defined as inline.
-
-#ifndef FST_SCRIPT_GETTERS_H_
-#define FST_SCRIPT_GETTERS_H_
-
-#include <string>
-
-#include <fst/compose.h>          // For ComposeFilter.
-#include <fst/determinize.h>      // For DeterminizeType.
-#include <fst/encode.h>           // For kEncodeLabels (etc.).
-#include <fst/epsnormalize.h>     // For EpsNormalizeType.
-#include <fst/project.h>          // For ProjectType.
-#include <fst/push.h>             // For kPushWeights (etc.).
-#include <fst/queue.h>            // For QueueType.
-#include <fst/rational.h>         // For ClosureType.
-#include <fst/script/arcsort.h>       // For ArcSortType.
-#include <fst/script/map.h>           // For MapType.
-#include <fst/script/script-impl.h>   // For RandArcSelection.
-
-#include <fst/log.h>
-
-namespace fst {
-namespace script {
-
-bool GetArcSortType(const string &str, ArcSortType *sort_type);
-
-inline ClosureType GetClosureType(bool closure_plus) {
-  return closure_plus ? CLOSURE_PLUS : CLOSURE_STAR;
-}
-
-bool GetComposeFilter(const string &str, ComposeFilter *compose_filter);
-
-bool GetDeterminizeType(const string &str, DeterminizeType *det_type);
-
-inline uint32 GetEncodeFlags(bool encode_labels, bool encode_weights) {
-  return (encode_labels ? kEncodeLabels : 0) |
-         (encode_weights ? kEncodeWeights : 0);
-}
-
-inline EpsNormalizeType GetEpsNormalizeType(bool eps_norm_output) {
-  return eps_norm_output ? EPS_NORM_OUTPUT : EPS_NORM_INPUT;
-}
-
-bool GetMapType(const string &str, MapType *map_type);
-
-inline ProjectType GetProjectType(bool project_output) {
-  return project_output ? PROJECT_OUTPUT : PROJECT_INPUT;
-}
-
-inline uint32 GetPushFlags(bool push_weights, bool push_labels,
-                           bool remove_total_weight, bool remove_common_affix) {
-  return ((push_weights ? kPushWeights : 0) |
-          (push_labels ? kPushLabels : 0) |
-          (remove_total_weight ? kPushRemoveTotalWeight : 0) |
-          (remove_common_affix ? kPushRemoveCommonAffix : 0));
-}
-
-bool GetQueueType(const string &str, QueueType *queue_type);
-
-bool GetRandArcSelection(const string &str, RandArcSelection *ras);
-
-bool GetReplaceLabelType(const string &str, bool epsilon_on_replace,
-                         ReplaceLabelType *rlt);
-
-inline ReweightType GetReweightType(bool to_final) {
-  return to_final ? REWEIGHT_TO_FINAL : REWEIGHT_TO_INITIAL;
-}
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_GETTERS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/info-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/info-impl.h
deleted file mode 100644
index e8956498..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/info-impl.h
+++ /dev/null
@@ -1,314 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to compute various information about FSTs, a helper class for
-// fstinfo.cc.
-
-#ifndef FST_SCRIPT_INFO_IMPL_H_
-#define FST_SCRIPT_INFO_IMPL_H_
-
-#include <map>
-#include <string>
-#include <vector>
-
-#include <fst/connect.h>
-#include <fst/dfs-visit.h>
-#include <fst/fst.h>
-#include <fst/lookahead-matcher.h>
-#include <fst/matcher.h>
-#include <fst/queue.h>
-#include <fst/test-properties.h>
-#include <fst/verify.h>
-#include <fst/visit.h>
-
-namespace fst {
-
-// Compute various information about FSTs, helper class for fstinfo.cc.
-// WARNING: Stand-alone use of this class is not recommended, most code
-// should call directly the relevant library functions: Fst<Arc>::NumStates,
-// Fst<Arc>::NumArcs, TestProperties, etc.
-class FstInfo {
- public:
-  FstInfo() {}
-
-  // When info_type is "short" (or "auto" and not an ExpandedFst) then only
-  // minimal info is computed and can be requested.
-  template <typename Arc>
-  FstInfo(const Fst<Arc> &fst, bool test_properties,
-          const string &arc_filter_type = "any",
-          const string &info_type = "auto", bool verify = true)
-      : fst_type_(fst.Type()),
-        input_symbols_(fst.InputSymbols() ? fst.InputSymbols()->Name()
-                                          : "none"),
-        output_symbols_(fst.OutputSymbols() ? fst.OutputSymbols()->Name()
-                                            : "none"),
-        nstates_(0),
-        narcs_(0),
-        start_(kNoStateId),
-        nfinal_(0),
-        nepsilons_(0),
-        niepsilons_(0),
-        noepsilons_(0),
-        ilabel_mult_(0.0),
-        olabel_mult_(0.0),
-        naccess_(0),
-        ncoaccess_(0),
-        nconnect_(0),
-        ncc_(0),
-        nscc_(0),
-        input_match_type_(MATCH_NONE),
-        output_match_type_(MATCH_NONE),
-        input_lookahead_(false),
-        output_lookahead_(false),
-        properties_(0),
-        arc_filter_type_(arc_filter_type),
-        long_info_(true),
-        arc_type_(Arc::Type()) {
-    using Label = typename Arc::Label;
-    using StateId = typename Arc::StateId;
-    using Weight = typename Arc::Weight;
-    if (info_type == "long") {
-      long_info_ = true;
-    } else if (info_type == "short") {
-      long_info_ = false;
-    } else if (info_type == "auto") {
-      long_info_ = fst.Properties(kExpanded, false);
-    } else {
-      FSTERROR() << "Bad info type: " << info_type;
-      return;
-    }
-    if (!long_info_) return;
-    // If the FST is not sane, we return.
-    if (verify && !Verify(fst)) {
-      FSTERROR() << "FstInfo: Verify: FST not well-formed";
-      return;
-    }
-    start_ = fst.Start();
-    properties_ = fst.Properties(kFstProperties, test_properties);
-    for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-      ++nstates_;
-      const auto s = siter.Value();
-      if (fst.Final(s) != Weight::Zero()) ++nfinal_;
-      std::map<Label, size_t> ilabel_count;
-      std::map<Label, size_t> olabel_count;
-      for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-        const auto &arc = aiter.Value();
-        ++narcs_;
-        if (arc.ilabel == 0 && arc.olabel == 0) ++nepsilons_;
-        if (arc.ilabel == 0) ++niepsilons_;
-        if (arc.olabel == 0) ++noepsilons_;
-        ++ilabel_count[arc.ilabel];
-        ++olabel_count[arc.olabel];
-      }
-      for (auto it = ilabel_count.begin(); it != ilabel_count.end(); ++it) {
-        ilabel_mult_ += it->second * it->second;
-      }
-      for (auto it = olabel_count.begin(); it != olabel_count.end(); ++it) {
-        olabel_mult_ += it->second * it->second;
-      }
-    }
-    if (narcs_ > 0) {
-      ilabel_mult_ /= narcs_;
-      olabel_mult_ /= narcs_;
-    }
-    {
-      std::vector<StateId> cc;
-      CcVisitor<Arc> cc_visitor(&cc);
-      FifoQueue<StateId> fifo_queue;
-      if (arc_filter_type == "any") {
-        Visit(fst, &cc_visitor, &fifo_queue);
-      } else if (arc_filter_type == "epsilon") {
-        Visit(fst, &cc_visitor, &fifo_queue, EpsilonArcFilter<Arc>());
-      } else if (arc_filter_type == "iepsilon") {
-        Visit(fst, &cc_visitor, &fifo_queue, InputEpsilonArcFilter<Arc>());
-      } else if (arc_filter_type == "oepsilon") {
-        Visit(fst, &cc_visitor, &fifo_queue, OutputEpsilonArcFilter<Arc>());
-      } else {
-        FSTERROR() << "Bad arc filter type: " << arc_filter_type;
-        return;
-      }
-      for (StateId s = 0; s < cc.size(); ++s) {
-        if (cc[s] >= ncc_) ncc_ = cc[s] + 1;
-      }
-    }
-    {
-      std::vector<StateId> scc;
-      std::vector<bool> access, coaccess;
-      uint64 props = 0;
-      SccVisitor<Arc> scc_visitor(&scc, &access, &coaccess, &props);
-      if (arc_filter_type == "any") {
-        DfsVisit(fst, &scc_visitor);
-      } else if (arc_filter_type == "epsilon") {
-        DfsVisit(fst, &scc_visitor, EpsilonArcFilter<Arc>());
-      } else if (arc_filter_type == "iepsilon") {
-        DfsVisit(fst, &scc_visitor, InputEpsilonArcFilter<Arc>());
-      } else if (arc_filter_type == "oepsilon") {
-        DfsVisit(fst, &scc_visitor, OutputEpsilonArcFilter<Arc>());
-      } else {
-        FSTERROR() << "Bad arc filter type: " << arc_filter_type;
-        return;
-      }
-      for (StateId s = 0; s < scc.size(); ++s) {
-        if (access[s]) ++naccess_;
-        if (coaccess[s]) ++ncoaccess_;
-        if (access[s] && coaccess[s]) ++nconnect_;
-        if (scc[s] >= nscc_) nscc_ = scc[s] + 1;
-      }
-    }
-    LookAheadMatcher<Fst<Arc>> imatcher(fst, MATCH_INPUT);
-    input_match_type_ = imatcher.Type(test_properties);
-    input_lookahead_ = imatcher.Flags() & kInputLookAheadMatcher;
-    LookAheadMatcher<Fst<Arc>> omatcher(fst, MATCH_OUTPUT);
-    output_match_type_ = omatcher.Type(test_properties);
-    output_lookahead_ = omatcher.Flags() & kOutputLookAheadMatcher;
-  }
-
-  // Short info.
-
-  const string &FstType() const { return fst_type_; }
-
-  const string &ArcType() const { return arc_type_; }
-
-  const string &InputSymbols() const { return input_symbols_; }
-
-  const string &OutputSymbols() const { return output_symbols_; }
-
-  bool LongInfo() const { return long_info_; }
-
-  const string &ArcFilterType() const { return arc_filter_type_; }
-
-  // Long info.
-
-  MatchType InputMatchType() const {
-    CheckLong();
-    return input_match_type_;
-  }
-
-  MatchType OutputMatchType() const {
-    CheckLong();
-    return output_match_type_;
-  }
-
-  bool InputLookAhead() const {
-    CheckLong();
-    return input_lookahead_;
-  }
-
-  bool OutputLookAhead() const {
-    CheckLong();
-    return output_lookahead_;
-  }
-
-  int64 NumStates() const {
-    CheckLong();
-    return nstates_;
-  }
-
-  size_t NumArcs() const {
-    CheckLong();
-    return narcs_;
-  }
-
-  int64 Start() const {
-    CheckLong();
-    return start_;
-  }
-
-  size_t NumFinal() const {
-    CheckLong();
-    return nfinal_;
-  }
-
-  size_t NumEpsilons() const {
-    CheckLong();
-    return nepsilons_;
-  }
-
-  size_t NumInputEpsilons() const {
-    CheckLong();
-    return niepsilons_;
-  }
-
-  size_t NumOutputEpsilons() const {
-    CheckLong();
-    return noepsilons_;
-  }
-
-  double InputLabelMultiplicity() const {
-    CheckLong();
-    return ilabel_mult_;
-  }
-
-  double OutputLabelMultiplicity() const {
-    CheckLong();
-    return olabel_mult_;
-  }
-
-  size_t NumAccessible() const {
-    CheckLong();
-    return naccess_;
-  }
-
-  size_t NumCoAccessible() const {
-    CheckLong();
-    return ncoaccess_;
-  }
-
-  size_t NumConnected() const {
-    CheckLong();
-    return nconnect_;
-  }
-
-  size_t NumCc() const {
-    CheckLong();
-    return ncc_;
-  }
-
-  size_t NumScc() const {
-    CheckLong();
-    return nscc_;
-  }
-
-  uint64 Properties() const {
-    CheckLong();
-    return properties_;
-  }
-
- private:
-  void CheckLong() const {
-    if (!long_info_)
-      FSTERROR() << "FstInfo: Method only available with long info signature";
-  }
-
-  string fst_type_;
-  string input_symbols_;
-  string output_symbols_;
-  int64 nstates_;
-  size_t narcs_;
-  int64 start_;
-  size_t nfinal_;
-  size_t nepsilons_;
-  size_t niepsilons_;
-  size_t noepsilons_;
-  double ilabel_mult_;
-  double olabel_mult_;
-  size_t naccess_;
-  size_t ncoaccess_;
-  size_t nconnect_;
-  size_t ncc_;
-  size_t nscc_;
-  MatchType input_match_type_;
-  MatchType output_match_type_;
-  bool input_lookahead_;
-  bool output_lookahead_;
-  uint64 properties_;
-  string arc_filter_type_;
-  bool long_info_;
-  string arc_type_;
-};
-
-void PrintFstInfoImpl(const FstInfo &fstinfo, bool pipe = false);
-
-}  // namespace fst
-
-#endif  // FST_SCRIPT_INFO_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/info.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/info.h
deleted file mode 100644
index 039d06d8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/info.h
+++ /dev/null
@@ -1,50 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_INFO_H_
-#define FST_SCRIPT_INFO_H_
-
-#include <string>
-#include <tuple>
-
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/info-impl.h>
-
-namespace fst {
-namespace script {
-
-using InfoArgs = std::tuple<const FstClass &, bool, const string &,
-                            const string &, bool, bool>;
-
-template <class Arc>
-void PrintFstInfo(InfoArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  const FstInfo fstinfo(fst, std::get<1>(*args), std::get<2>(*args),
-                        std::get<3>(*args), std::get<4>(*args));
-  PrintFstInfoImpl(fstinfo, std::get<5>(*args));
-  if (std::get<5>(*args)) fst.Write("");
-}
-
-void PrintFstInfo(const FstClass &f, bool test_properties,
-                  const string &arc_filter, const string &info_type, bool pipe,
-                  bool verify);
-
-using GetInfoArgs = std::tuple<const FstClass &, bool, const string &,
-                               const string &, bool, FstInfo *>;
-
-template <class Arc>
-void GetFstInfo(GetInfoArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  *(std::get<5>(*args)) = FstInfo(fst, std::get<1>(*args), std::get<2>(*args),
-                                  std::get<3>(*args), std::get<4>(*args));
-}
-
-void GetFstInfo(const FstClass &fst, bool test_properties,
-                const string &arc_filter, const string &info_type, bool verify,
-                FstInfo *info);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_INFO_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/intersect.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/intersect.h
deleted file mode 100644
index 229bd56f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/intersect.h
+++ /dev/null
@@ -1,35 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_INTERSECT_H_
-#define FST_SCRIPT_INTERSECT_H_
-
-#include <tuple>
-
-#include <fst/intersect.h>
-#include <fst/script/compose.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using IntersectArgs = std::tuple<const FstClass &, const FstClass &,
-                                 MutableFstClass *, const ComposeOptions &>;
-
-template <class Arc>
-void Intersect(IntersectArgs *args) {
-  const Fst<Arc> &ifst1 = *(std::get<0>(*args).GetFst<Arc>());
-  const Fst<Arc> &ifst2 = *(std::get<1>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<2>(*args)->GetMutableFst<Arc>();
-  const auto &opts = std::get<3>(*args);
-  Intersect(ifst1, ifst2, ofst, opts);
-}
-
-void Intersect(const FstClass &ifst, const FstClass &ifst2,
-               MutableFstClass *ofst,
-               const ComposeOptions &opts = ComposeOptions());
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_INTERSECT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/invert.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/invert.h
deleted file mode 100644
index 5bc31317..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/invert.h
+++ /dev/null
@@ -1,23 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_INVERT_H_
-#define FST_SCRIPT_INVERT_H_
-
-#include <fst/invert.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-template <class Arc>
-void Invert(MutableFstClass *fst) {
-  Invert(fst->GetMutableFst<Arc>());
-}
-
-void Invert(MutableFstClass *fst);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_INVERT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/isomorphic.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/isomorphic.h
deleted file mode 100644
index 94ea77f9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/isomorphic.h
+++ /dev/null
@@ -1,34 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_ISOMORPHIC_H_
-#define FST_SCRIPT_ISOMORPHIC_H_
-
-#include <tuple>
-
-#include <fst/isomorphic.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using IsomorphicInnerArgs = std::tuple<const FstClass &, const FstClass &,
-                                       float>;
-
-using IsomorphicArgs = WithReturnValue<bool, IsomorphicInnerArgs>;
-
-template <class Arc>
-void Isomorphic(IsomorphicArgs *args) {
-  const Fst<Arc> &fst1 = *(std::get<0>(args->args).GetFst<Arc>());
-  const Fst<Arc> &fst2 = *(std::get<1>(args->args).GetFst<Arc>());
-  args->retval = Isomorphic(fst1, fst2, std::get<2>(args->args));
-}
-
-bool Isomorphic(const FstClass &fst1, const FstClass &fst2,
-                float delta = kDelta);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_ISOMORPHIC_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/map.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/map.h
deleted file mode 100644
index bef7797b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/map.h
+++ /dev/null
@@ -1,150 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_MAP_H_
-#define FST_SCRIPT_MAP_H_
-
-#include <memory>
-#include <tuple>
-
-#include <fst/arc-map.h>
-#include <fst/state-map.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-template <class M>
-Fst<typename M::ToArc> *ArcMap(const Fst<typename M::FromArc> &fst,
-                               const M &mapper) {
-  using ToArc = typename M::ToArc;
-  auto *ofst = new VectorFst<ToArc>;
-  ArcMap(fst, ofst, mapper);
-  return ofst;
-}
-
-template <class M>
-Fst<typename M::ToArc> *StateMap(const Fst<typename M::FromArc> &fst,
-                                 const M &mapper) {
-  using ToArc = typename M::ToArc;
-  auto *ofst = new VectorFst<ToArc>;
-  StateMap(fst, ofst, mapper);
-  return ofst;
-}
-
-enum MapType {
-  ARC_SUM_MAPPER,
-  ARC_UNIQUE_MAPPER,
-  IDENTITY_MAPPER,
-  INPUT_EPSILON_MAPPER,
-  INVERT_MAPPER,
-  OUTPUT_EPSILON_MAPPER,
-  PLUS_MAPPER,
-  QUANTIZE_MAPPER,
-  RMWEIGHT_MAPPER,
-  SUPERFINAL_MAPPER,
-  TIMES_MAPPER,
-  TO_LOG_MAPPER,
-  TO_LOG64_MAPPER,
-  TO_STD_MAPPER
-};
-
-using MapInnerArgs = std::tuple<const FstClass &, MapType, float,
-                                const WeightClass &>;
-
-using MapArgs = WithReturnValue<FstClass *, MapInnerArgs>;
-
-template <class Arc>
-void Map(MapArgs *args) {
-  using Weight = typename Arc::Weight;
-  const Fst<Arc> &ifst = *(std::get<0>(args->args).GetFst<Arc>());
-  const auto map_type = std::get<1>(args->args);
-  const auto weight = *(std::get<3>(args->args).GetWeight<Weight>());
-  switch (map_type) {
-    case ARC_SUM_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(StateMap(ifst, ArcSumMapper<Arc>(ifst)));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case ARC_UNIQUE_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(StateMap(ifst,
-                                              ArcUniqueMapper<Arc>(ifst)));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case IDENTITY_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, IdentityArcMapper<Arc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case INPUT_EPSILON_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, InputEpsilonMapper<Arc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case INVERT_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, InvertWeightMapper<Arc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case OUTPUT_EPSILON_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, OutputEpsilonMapper<Arc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case PLUS_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, PlusMapper<Arc>(weight)));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case QUANTIZE_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, QuantizeMapper<Arc>(
-          std::get<2>(args->args))));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case RMWEIGHT_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, RmWeightMapper<Arc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case SUPERFINAL_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, SuperFinalMapper<Arc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case TIMES_MAPPER: {
-      std::unique_ptr<Fst<Arc>> ofst(ArcMap(ifst, TimesMapper<Arc>(weight)));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case TO_LOG_MAPPER: {
-      std::unique_ptr<Fst<LogArc>> ofst(ArcMap(ifst,
-          WeightConvertMapper<Arc, LogArc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case TO_LOG64_MAPPER: {
-      std::unique_ptr<Fst<Log64Arc>> ofst(ArcMap(ifst,
-          WeightConvertMapper<Arc, Log64Arc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-    case TO_STD_MAPPER: {
-      std::unique_ptr<Fst<StdArc>> ofst(ArcMap(ifst,
-          WeightConvertMapper<Arc, StdArc>()));
-      args->retval = new FstClass(*ofst);
-      return;
-    }
-  }
-}
-
-FstClass *Map(const FstClass &ifst, MapType map_type, float delta,
-              const WeightClass &weight);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_MAP_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/minimize.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/minimize.h
deleted file mode 100644
index 8b852f68..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/minimize.h
+++ /dev/null
@@ -1,33 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_MINIMIZE_H_
-#define FST_SCRIPT_MINIMIZE_H_
-
-#include <tuple>
-
-#include <fst/minimize.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using MinimizeArgs = std::tuple<MutableFstClass *, MutableFstClass *, float,
-                                bool>;
-
-template <class Arc>
-void Minimize(MinimizeArgs *args) {
-  MutableFst<Arc> *ofst1 = std::get<0>(*args)->GetMutableFst<Arc>();
-  MutableFst<Arc> *ofst2 = (std::get<1>(*args) ?
-                            std::get<1>(*args)->GetMutableFst<Arc>() :
-                            nullptr);
-  Minimize(ofst1, ofst2, std::get<2>(*args), std::get<3>(*args));
-}
-
-void Minimize(MutableFstClass *ofst1, MutableFstClass *ofst2 = nullptr,
-              float delta = kDelta, bool allow_nondet = false);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_MINIMIZE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/print-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/print-impl.h
deleted file mode 100644
index 539c6d8f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/print-impl.h
+++ /dev/null
@@ -1,132 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Stand-alone class to print out binary FSTs in the AT&T format, a helper
-// class for fstprint.cc.
-
-#ifndef FST_SCRIPT_PRINT_IMPL_H_
-#define FST_SCRIPT_PRINT_IMPL_H_
-
-#include <ostream>
-#include <sstream>
-#include <string>
-
-#include <fst/fstlib.h>
-#include <fst/util.h>
-
-DECLARE_string(fst_field_separator);
-
-namespace fst {
-
-// Print a binary FST in textual format (helper class for fstprint.cc).
-// WARNING: Stand-alone use of this class not recommended, most code should
-// read/write using the binary format which is much more efficient.
-template <class Arc>
-class FstPrinter {
- public:
-  using StateId = typename Arc::StateId;
-  using Label = typename Arc::Label;
-  using Weight = typename Arc::Weight;
-
-  FstPrinter(const Fst<Arc> &fst, const SymbolTable *isyms,
-             const SymbolTable *osyms, const SymbolTable *ssyms, bool accep,
-             bool show_weight_one, const string &field_separator,
-             const string &missing_symbol = "")
-      : fst_(fst),
-        isyms_(isyms),
-        osyms_(osyms),
-        ssyms_(ssyms),
-        accep_(accep && fst.Properties(kAcceptor, true)),
-        ostrm_(nullptr),
-        show_weight_one_(show_weight_one),
-        sep_(field_separator),
-        missing_symbol_(missing_symbol) {}
-
-  // Prints FST to an output stream.
-  void Print(std::ostream *ostrm, const string &dest) {
-    ostrm_ = ostrm;
-    dest_ = dest;
-    const auto start = fst_.Start();
-    if (start == kNoStateId) return;
-    // Initial state first.
-    PrintState(start);
-    for (StateIterator<Fst<Arc>> siter(fst_); !siter.Done(); siter.Next()) {
-      const auto s = siter.Value();
-      if (s != start) PrintState(s);
-    }
-  }
-
- private:
-  void PrintId(StateId id, const SymbolTable *syms, const char *name) const {
-    if (syms) {
-      string symbol = syms->Find(id);
-      if (symbol.empty()) {
-        if (missing_symbol_.empty()) {
-          FSTERROR() << "FstPrinter: Integer " << id
-                     << " is not mapped to any textual symbol"
-                     << ", symbol table = " << syms->Name()
-                     << ", destination = " << dest_;
-          symbol = "?";
-        } else {
-          symbol = missing_symbol_;
-        }
-      }
-      *ostrm_ << symbol;
-    } else {
-      *ostrm_ << id;
-    }
-  }
-
-  void PrintStateId(StateId s) const { PrintId(s, ssyms_, "state ID"); }
-
-  void PrintILabel(Label l) const { PrintId(l, isyms_, "arc input label"); }
-
-  void PrintOLabel(Label l) const { PrintId(l, osyms_, "arc output label"); }
-
-  void PrintState(StateId s) const {
-    bool output = false;
-    for (ArcIterator<Fst<Arc>> aiter(fst_, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      PrintStateId(s);
-      *ostrm_ << sep_;
-      PrintStateId(arc.nextstate);
-      *ostrm_ << sep_;
-      PrintILabel(arc.ilabel);
-      if (!accep_) {
-        *ostrm_ << sep_;
-        PrintOLabel(arc.olabel);
-      }
-      if (show_weight_one_ || arc.weight != Weight::One())
-        *ostrm_ << sep_ << arc.weight;
-      *ostrm_ << "\n";
-      output = true;
-    }
-    const auto weight = fst_.Final(s);
-    if (weight != Weight::Zero() || !output) {
-      PrintStateId(s);
-      if (show_weight_one_ || weight != Weight::One()) {
-        *ostrm_ << sep_ << weight;
-      }
-      *ostrm_ << "\n";
-    }
-  }
-
-  const Fst<Arc> &fst_;
-  const SymbolTable *isyms_;  // ilabel symbol table.
-  const SymbolTable *osyms_;  // olabel symbol table.
-  const SymbolTable *ssyms_;  // slabel symbol table.
-  bool accep_;                // Print as acceptor when possible?
-  std::ostream *ostrm_;       // Text FST destination.
-  string dest_;               // Text FST destination name.
-  bool show_weight_one_;      // Print weights equal to Weight::One()?
-  string sep_;                // Separator character between fields.
-  string missing_symbol_;     // Symbol to print when lookup fails (default
-                              // "" means raise error).
-                              //
-  FstPrinter(const FstPrinter &) = delete;
-  FstPrinter &operator=(const FstPrinter &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_SCRIPT_PRINT_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/print.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/print.h
deleted file mode 100644
index 687606b3..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/print.h
+++ /dev/null
@@ -1,79 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_PRINT_H_
-#define FST_SCRIPT_PRINT_H_
-
-#include <ostream>
-
-#include <fst/flags.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/print-impl.h>
-
-DECLARE_string(fst_field_separator);
-
-namespace fst {
-namespace script {
-
-// Note: it is safe to pass these strings as references because
-// this struct is only used to pass them deeper in the call graph.
-// Be sure you understand why this is so before using this struct
-// for anything else!
-struct FstPrinterArgs {
-  const FstClass &fst;
-  const SymbolTable *isyms;
-  const SymbolTable *osyms;
-  const SymbolTable *ssyms;
-  const bool accept;
-  const bool show_weight_one;
-  std::ostream *ostrm;
-  const string &dest;
-  const string &sep;  // NOLINT
-  const string &missing_symbol;
-
-  FstPrinterArgs(const FstClass &fst, const SymbolTable *isyms,
-                 const SymbolTable *osyms, const SymbolTable *ssyms,
-                 bool accept, bool show_weight_one, std::ostream *ostrm,
-                 const string &dest, const string &sep,
-                 const string &missing_sym = "")
-      : fst(fst),
-        isyms(isyms),
-        osyms(osyms),
-        ssyms(ssyms),
-        accept(accept),
-        show_weight_one(show_weight_one),
-        ostrm(ostrm),
-        dest(dest),
-        sep(sep),
-        missing_symbol(missing_sym) {}
-};
-
-template <class Arc>
-void PrintFst(FstPrinterArgs *args) {
-  const Fst<Arc> &fst = *(args->fst.GetFst<Arc>());
-  FstPrinter<Arc> fstprinter(fst, args->isyms, args->osyms, args->ssyms,
-                             args->accept, args->show_weight_one, args->sep,
-                             args->missing_symbol);
-  fstprinter.Print(args->ostrm, args->dest);
-}
-
-void PrintFst(const FstClass &fst, std::ostream &ostrm, const string &dest,
-              const SymbolTable *isyms, const SymbolTable *osyms,
-              const SymbolTable *ssyms, bool accept, bool show_weight_one,
-              const string &missing_sym = "");
-
-// The same, but with more sensible defaults.
-template <class Arc>
-void PrintFst(const Fst<Arc> &fst, std::ostream &ostrm, const string &dest = "",
-              const SymbolTable *isyms = nullptr,
-              const SymbolTable *osyms = nullptr,
-              const SymbolTable *ssyms = nullptr) {
-  const string sep = FLAGS_fst_field_separator.substr(0, 1);
-  FstPrinter<Arc> fstprinter(fst, isyms, osyms, ssyms, true, true, sep);
-  fstprinter.Print(&ostrm, dest);
-}
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_PRINT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/project.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/project.h
deleted file mode 100644
index 13edeb1d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/project.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_PROJECT_H_
-#define FST_SCRIPT_PROJECT_H_
-
-#include <utility>
-
-#include <fst/project.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using ProjectArgs = std::pair<MutableFstClass *, ProjectType>;
-
-template <class Arc>
-void Project(ProjectArgs *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  Project(fst, std::get<1>(*args));
-}
-
-void Project(MutableFstClass *fst, ProjectType project_type);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_PROJECT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/prune.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/prune.h
deleted file mode 100644
index ed10b540..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/prune.h
+++ /dev/null
@@ -1,51 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_PRUNE_H_
-#define FST_SCRIPT_PRUNE_H_
-
-#include <tuple>
-#include <utility>
-
-#include <fst/prune.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-using PruneArgs1 = std::tuple<const FstClass &, MutableFstClass *,
-                              const WeightClass &, int64, float>;
-
-template <class Arc>
-void Prune(PruneArgs1 *args) {
-  using Weight = typename Arc::Weight;
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  const auto weight_threshold = *(std::get<2>(*args).GetWeight<Weight>());
-  Prune(ifst, ofst, weight_threshold, std::get<3>(*args), std::get<4>(*args));
-}
-
-using PruneArgs2 = std::tuple<MutableFstClass *, const WeightClass &, int64,
-                               float>;
-
-template <class Arc>
-void Prune(PruneArgs2 *args) {
-  using Weight = typename Arc::Weight;
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  const auto weight_threshold = *(std::get<1>(*args).GetWeight<Weight>());
-  Prune(fst, weight_threshold, std::get<2>(*args), std::get<3>(*args));
-}
-
-void Prune(const FstClass &ifst, MutableFstClass *ofst,
-           const WeightClass &weight_threshold,
-           int64 state_threshold = kNoStateId,
-           float delta = kDelta);
-
-void Prune(MutableFstClass *fst, const WeightClass &weight_threshold,
-           int64 state_threshold = kNoStateId, float delta = kDelta);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_PRUNE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/push.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/push.h
deleted file mode 100644
index 018cd8f8..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/push.h
+++ /dev/null
@@ -1,53 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_PUSH_H_
-#define FST_SCRIPT_PUSH_H_
-
-#include <tuple>
-
-#include <fst/push.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using PushArgs1 = std::tuple<MutableFstClass *, ReweightType, float, bool>;
-
-template <class Arc>
-void Push(PushArgs1 *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  Push(fst, std::get<1>(*args), std::get<2>(*args), std::get<3>(*args));
-}
-
-using PushArgs2 = std::tuple<const FstClass &, MutableFstClass *, uint32,
-                             ReweightType, float>;
-
-template <class Arc>
-void Push(PushArgs2 *args) {
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  switch (std::get<3>(*args)) {
-    case REWEIGHT_TO_FINAL: {
-      Push<Arc, REWEIGHT_TO_FINAL>(ifst, ofst, std::get<2>(*args),
-                                   std::get<4>(*args));
-      return;
-    }
-    case REWEIGHT_TO_INITIAL: {
-      Push<Arc, REWEIGHT_TO_INITIAL>(ifst, ofst, std::get<2>(*args),
-                                     std::get<4>(*args));
-      return;
-    }
-  }
-}
-
-void Push(MutableFstClass *fst, ReweightType rew_type, float delta = kDelta,
-          bool remove_total_weight = false);
-
-void Push(const FstClass &ifst, MutableFstClass *ofst, uint32 flags,
-          ReweightType rew_type, float delta = kDelta);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_PUSH_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/randequivalent.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/randequivalent.h
deleted file mode 100644
index 945f8a06..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/randequivalent.h
+++ /dev/null
@@ -1,67 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_RANDEQUIVALENT_H_
-#define FST_SCRIPT_RANDEQUIVALENT_H_
-
-#include <climits>
-#include <ctime>
-
-#include <tuple>
-
-#include <fst/randequivalent.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/script-impl.h>
-
-namespace fst {
-namespace script {
-
-using RandEquivalentInnerArgs = std::tuple<const FstClass &, const FstClass &,
-    int32, float, time_t, const RandGenOptions<RandArcSelection> &>;
-
-using RandEquivalentArgs = WithReturnValue<bool, RandEquivalentInnerArgs>;
-
-template <class Arc>
-void RandEquivalent(RandEquivalentArgs *args) {
-  const Fst<Arc> &fst1 = *(std::get<0>(args->args).GetFst<Arc>());
-  const Fst<Arc> &fst2 = *(std::get<1>(args->args).GetFst<Arc>());
-  const auto seed = std::get<4>(args->args);
-  const auto &opts = std::get<5>(args->args);
-  switch (opts.selector) {
-    case UNIFORM_ARC_SELECTOR: {
-      const UniformArcSelector<Arc> selector(seed);
-      const RandGenOptions<UniformArcSelector<Arc>> ropts(selector,
-                                                          opts.max_length);
-      args->retval = RandEquivalent(fst1, fst2, std::get<2>(args->args),
-                                    std::get<3>(args->args), ropts);
-      return;
-    }
-    case FAST_LOG_PROB_ARC_SELECTOR: {
-      const FastLogProbArcSelector<Arc> selector(seed);
-      const RandGenOptions<FastLogProbArcSelector<Arc>> ropts(selector,
-                                                              opts.max_length);
-      args->retval = RandEquivalent(fst1, fst2, std::get<2>(args->args),
-                                    std::get<3>(args->args), ropts);
-      return;
-    }
-    case LOG_PROB_ARC_SELECTOR: {
-      const LogProbArcSelector<Arc> selector(seed);
-      const RandGenOptions<LogProbArcSelector<Arc>> ropts(selector,
-                                                          opts.max_length);
-      args->retval = RandEquivalent(fst1, fst2, std::get<2>(args->args),
-                                    std::get<3>(args->args), ropts);
-      return;
-    }
-  }
-}
-
-bool RandEquivalent(const FstClass &fst1, const FstClass &fst2, int32 npath = 1,
-    float delta = kDelta, time_t seed = time(nullptr),
-    const RandGenOptions<RandArcSelection> &opts =
-        RandGenOptions<RandArcSelection>(UNIFORM_ARC_SELECTOR));
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_RANDEQUIVALENT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/randgen.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/randgen.h
deleted file mode 100644
index 5ce79d01..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/randgen.h
+++ /dev/null
@@ -1,63 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_RANDGEN_H_
-#define FST_SCRIPT_RANDGEN_H_
-
-#include <ctime>
-
-#include <tuple>
-
-#include <fst/randgen.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/script-impl.h>
-
-namespace fst {
-namespace script {
-
-using RandGenArgs = std::tuple<const FstClass &, MutableFstClass *, time_t,
-                               const RandGenOptions<RandArcSelection> &>;
-
-template <class Arc>
-void RandGen(RandGenArgs *args) {
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  const time_t seed = std::get<2>(*args);
-  const auto &opts = std::get<3>(*args);
-  switch (opts.selector) {
-    case UNIFORM_ARC_SELECTOR: {
-      const UniformArcSelector<Arc> selector(seed);
-      const RandGenOptions<UniformArcSelector<Arc>> ropts(
-          selector, opts.max_length, opts.npath, opts.weighted,
-          opts.remove_total_weight);
-      RandGen(ifst, ofst, ropts);
-      return;
-    }
-    case FAST_LOG_PROB_ARC_SELECTOR: {
-      const FastLogProbArcSelector<Arc> selector(seed);
-      const RandGenOptions<FastLogProbArcSelector<Arc>> ropts(
-          selector, opts.max_length, opts.npath, opts.weighted,
-          opts.remove_total_weight);
-      RandGen(ifst, ofst, ropts);
-      return;
-    }
-    case LOG_PROB_ARC_SELECTOR: {
-      const LogProbArcSelector<Arc> selector(seed);
-      const RandGenOptions<LogProbArcSelector<Arc>> ropts(
-          selector, opts.max_length, opts.npath, opts.weighted,
-          opts.remove_total_weight);
-      RandGen(ifst, ofst, ropts);
-      return;
-    }
-  }
-}
-
-void RandGen(const FstClass &ifst, MutableFstClass *ofst,
-             time_t seed = time(nullptr),
-             const RandGenOptions<RandArcSelection> &opts =
-                 RandGenOptions<RandArcSelection>(UNIFORM_ARC_SELECTOR));
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_RANDGEN_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/register.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/register.h
deleted file mode 100644
index 1be2cde7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/register.h
+++ /dev/null
@@ -1,114 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_REGISTER_H_
-#define FST_SCRIPT_REGISTER_H_
-
-#include <istream>
-#include <string>
-
-#include <fst/generic-register.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/weight-class.h>
-
-// Holds methods and classes responsible for maintaining
-// the register for FstClass arc types.
-
-namespace fst {
-namespace script {
-
-// Registers for reading and converting various kinds of FST classes.
-
-// This class definition is to avoid a nested class definition inside the
-// IORegistration struct.
-
-template <class Reader, class Creator, class Converter>
-struct FstClassRegEntry {
-  Reader reader;
-  Creator creator;
-  Converter converter;
-
-  FstClassRegEntry(Reader r, Creator cr, Converter co)
-      : reader(r), creator(cr), converter(co) {}
-
-  FstClassRegEntry()
-      : reader(NullReader), creator(NullCreator), converter(NullConverter) {}
-
-  // Null-returning reader, creator, and converter, used when registry lookup
-  // fails.
-
-  template <class FstClassType>
-  static FstClassType *NullReader(std::istream &strm,
-                                  const FstReadOptions &opts) {
-    return nullptr;
-  }
-
-  static FstClassImplBase *NullCreator() { return nullptr; }
-
-  static FstClassImplBase *NullConverter(const FstClass &other) {
-    return nullptr;
-  }
-};
-
-template <class Reader, class Creator, class Converter>
-class FstClassIORegister
-    : public GenericRegister<string,
-                             FstClassRegEntry<Reader, Creator, Converter>,
-                             FstClassIORegister<Reader, Creator, Converter>> {
- public:
-  Reader GetReader(const string &arc_type) const {
-    return this->GetEntry(arc_type).reader;
-  }
-
-  Creator GetCreator(const string &arc_type) const {
-    return this->GetEntry(arc_type).creator;
-  }
-
-  Converter GetConverter(const string &arc_type) const {
-    return this->GetEntry(arc_type).converter;
-  }
-
- protected:
-  string ConvertKeyToSoFilename(const string &key) const final {
-    string legal_type(key);
-    ConvertToLegalCSymbol(&legal_type);
-    return legal_type + "-arc.so";
-  }
-};
-
-// Struct containing everything needed to register a particular type
-// of FST class (e.g., a plain FstClass, or a MutableFstClass, etc.).
-template <class FstClassType>
-struct IORegistration {
-  using Reader = FstClassType *(*)(std::istream &stream,
-                                   const FstReadOptions &opts);
-
-  using Creator = FstClassImplBase *(*)();
-
-  using Converter = FstClassImplBase *(*)(const FstClass &other);
-
-  using Entry = FstClassRegEntry<Reader, Creator, Converter>;
-
-  // FST class Register.
-  using Register = FstClassIORegister<Reader, Creator, Converter>;
-
-  // FST class Register-er.
-  using Registerer =
-      GenericRegisterer<FstClassIORegister<Reader, Creator, Converter>>;
-};
-
-#define REGISTER_FST_CLASS(Class, Arc)                                   \
-  static IORegistration<Class>::Registerer Class##_##Arc##_registerer(   \
-      Arc::Type(),                                                       \
-      IORegistration<Class>::Entry(Class::Read<Arc>, Class::Create<Arc>, \
-                                   Class::Convert<Arc>))
-
-#define REGISTER_FST_CLASSES(Arc)           \
-  REGISTER_FST_CLASS(FstClass, Arc);        \
-  REGISTER_FST_CLASS(MutableFstClass, Arc); \
-  REGISTER_FST_CLASS(VectorFstClass, Arc);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_REGISTER_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/relabel.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/relabel.h
deleted file mode 100644
index 74443490..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/relabel.h
+++ /dev/null
@@ -1,64 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_RELABEL_H_
-#define FST_SCRIPT_RELABEL_H_
-
-#include <algorithm>
-#include <tuple>
-#include <utility>
-#include <vector>
-
-#include <fst/relabel.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using RelabelArgs1 = std::tuple<MutableFstClass *, const SymbolTable *,
-                                const SymbolTable *, const string &, bool,
-                                const SymbolTable *, const SymbolTable *,
-                                const string &, bool>;
-
-template <class Arc>
-void Relabel(RelabelArgs1 *args) {
-  MutableFst<Arc> *ofst = std::get<0>(*args)->GetMutableFst<Arc>();
-  Relabel(ofst, std::get<1>(*args), std::get<2>(*args), std::get<3>(*args),
-          std::get<4>(*args), std::get<5>(*args), std::get<6>(*args),
-          std::get<7>(*args), std::get<8>(*args));
-}
-
-using LabelPair = std::pair<int64, int64>;
-
-using RelabelArgs2 = std::tuple<MutableFstClass *,
-                                const std::vector<LabelPair> &,
-                                const std::vector<LabelPair> &>;
-
-template <class Arc>
-void Relabel(RelabelArgs2 *args) {
-  MutableFst<Arc> *ofst = std::get<0>(*args)->GetMutableFst<Arc>();
-  using LabelPair = std::pair<typename Arc::Label, typename Arc::Label>;
-  // In case the MutableFstClass::Label is not the same as Arc::Label,
-  // make a copy.
-  std::vector<LabelPair> typed_ipairs(std::get<1>(*args).size());
-  std::copy(std::get<1>(*args).begin(), std::get<1>(*args).end(),
-            typed_ipairs.begin());
-  std::vector<LabelPair> typed_opairs(std::get<2>(*args).size());
-  std::copy(std::get<2>(*args).begin(), std::get<2>(*args).end(),
-            typed_opairs.begin());
-  Relabel(ofst, typed_ipairs, typed_opairs);
-}
-
-void Relabel(MutableFstClass *ofst,
-             const SymbolTable *old_isymbols, const SymbolTable *new_isymbols,
-             const string &unknown_isymbol,  bool attach_new_isymbols,
-             const SymbolTable *old_osymbols, const SymbolTable *new_osymbols,
-             const string &unknown_osymbol, bool attach_new_osymbols);
-
-void Relabel(MutableFstClass *ofst, const std::vector<LabelPair> &ipairs,
-             const std::vector<LabelPair> &opairs);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_RELABEL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/replace.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/replace.h
deleted file mode 100644
index 926ece24..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/replace.h
+++ /dev/null
@@ -1,72 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_REPLACE_H_
-#define FST_SCRIPT_REPLACE_H_
-
-#include <tuple>
-#include <utility>
-#include <vector>
-
-#include <fst/replace.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-struct ReplaceOptions {
-  const int64 root;                          // Root rule for expansion.
-  const ReplaceLabelType call_label_type;    // How to label call arc.
-  const ReplaceLabelType return_label_type;  // How to label return arc.
-  const int64 return_label;                  // Specifies return arc label.
-
-  explicit ReplaceOptions(int64 root,
-      ReplaceLabelType call_label_type = REPLACE_LABEL_INPUT,
-      ReplaceLabelType return_label_type = REPLACE_LABEL_NEITHER,
-      int64 return_label = 0)
-      : root(root),
-        call_label_type(call_label_type),
-        return_label_type(return_label_type),
-        return_label(return_label) {}
-};
-
-using LabelFstClassPair = std::pair<int64, const FstClass *>;
-
-using ReplaceArgs = std::tuple<const std::vector<LabelFstClassPair> &,
-                               MutableFstClass *, const ReplaceOptions &>;
-
-template <class Arc>
-void Replace(ReplaceArgs *args) {
-  using LabelFstPair = std::pair<typename Arc::Label, const Fst<Arc> *>;
-  // Now that we know the arc type, we construct a vector of
-  // std::pair<real label, real fst> that the real Replace will use.
-  const auto &untyped_pairs = std::get<0>(*args);
-  std::vector<LabelFstPair> typed_pairs;
-  typed_pairs.reserve(untyped_pairs.size());
-  for (const auto &untyped_pair : untyped_pairs) {
-    typed_pairs.emplace_back(untyped_pair.first,  // Converts label.
-                             untyped_pair.second->GetFst<Arc>());
-  }
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  const auto &opts = std::get<2>(*args);
-  ReplaceFstOptions<Arc> typed_opts(opts.root, opts.call_label_type,
-                                    opts.return_label_type, opts.return_label);
-  ReplaceFst<Arc> rfst(typed_pairs, typed_opts);
-  // Checks for cyclic dependencies before attempting expansion.
-  if (rfst.CyclicDependencies()) {
-    FSTERROR() << "Replace: Cyclic dependencies detected; cannot expand";
-    ofst->SetProperties(kError, kError);
-    return;
-  }
-  typed_opts.gc = true;     // Caching options to speed up batch copy.
-  typed_opts.gc_limit = 0;
-  *ofst = rfst;
-}
-
-void Replace(const std::vector<LabelFstClassPair> &pairs,
-             MutableFstClass *ofst, const ReplaceOptions &opts);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_REPLACE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/reverse.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/reverse.h
deleted file mode 100644
index badd96b5..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/reverse.h
+++ /dev/null
@@ -1,30 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_REVERSE_H_
-#define FST_SCRIPT_REVERSE_H_
-
-#include <tuple>
-
-#include <fst/reverse.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using ReverseArgs = std::tuple<const FstClass &, MutableFstClass *, bool>;
-
-template <class Arc>
-void Reverse(ReverseArgs *args) {
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  Reverse(ifst, ofst, std::get<2>(*args));
-}
-
-void Reverse(const FstClass &ifst, MutableFstClass *ofst,
-             bool require_superinitial = true);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_REVERSE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/reweight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/reweight.h
deleted file mode 100644
index 3893ad85..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/reweight.h
+++ /dev/null
@@ -1,37 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_REWEIGHT_H_
-#define FST_SCRIPT_REWEIGHT_H_
-
-#include <tuple>
-#include <vector>
-
-#include <fst/reweight.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/script-impl.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-using ReweightArgs = std::tuple<MutableFstClass *,
-                                const std::vector<WeightClass> &, ReweightType>;
-
-template <class Arc>
-void Reweight(ReweightArgs *args) {
-  using Weight = typename Arc::Weight;
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  const std::vector<WeightClass> &potentials = std::get<1>(*args);
-  std::vector<Weight> typed_potentials;
-  internal::CopyWeights(potentials, &typed_potentials);
-  Reweight(fst, typed_potentials, std::get<2>(*args));
-}
-
-void Reweight(MutableFstClass *fst, const std::vector<WeightClass> &potentials,
-              ReweightType reweight_type);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_REWEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/rmepsilon.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/rmepsilon.h
deleted file mode 100644
index c878f96a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/rmepsilon.h
+++ /dev/null
@@ -1,114 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_RMEPSILON_H_
-#define FST_SCRIPT_RMEPSILON_H_
-
-#include <utility>
-#include <vector>
-
-#include <fst/queue.h>
-#include <fst/rmepsilon.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/shortest-distance.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-
-struct RmEpsilonOptions : public ShortestDistanceOptions {
-  const bool connect;
-  const WeightClass &weight_threshold;
-  const int64 state_threshold;
-
-  RmEpsilonOptions(QueueType queue_type, bool connect,
-                   const WeightClass &weight_threshold,
-                   int64 state_threshold = kNoStateId,
-                   float delta = kDelta)
-      : ShortestDistanceOptions(queue_type, EPSILON_ARC_FILTER, kNoStateId,
-                                delta),
-        connect(connect),
-        weight_threshold(weight_threshold),
-        state_threshold(state_threshold) {}
-};
-
-namespace internal {
-
-// Code to implement switching on queue types.
-
-template <class Arc, class Queue>
-void RmEpsilon(MutableFst<Arc> *fst,
-               std::vector<typename Arc::Weight> *distance,
-               const RmEpsilonOptions &opts,
-               Queue *queue) {
-  using Weight = typename Arc::Weight;
-  const fst::RmEpsilonOptions<Arc, Queue> ropts(queue, opts.delta,
-      opts.connect, *opts.weight_threshold.GetWeight<Weight>(),
-      opts.state_threshold);
-  RmEpsilon(fst, distance, ropts);
-}
-
-template <class Arc>
-void RmEpsilon(MutableFst<Arc> *fst,
-               const RmEpsilonOptions &opts) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  std::vector<Weight> distance;
-  switch (opts.queue_type) {
-    case AUTO_QUEUE: {
-      AutoQueue<StateId> queue(*fst, &distance, EpsilonArcFilter<Arc>());
-      RmEpsilon(fst, &distance, opts, &queue);
-      return;
-    }
-    case FIFO_QUEUE: {
-      FifoQueue<StateId> queue;
-      RmEpsilon(fst, &distance, opts, &queue);
-      return;
-    }
-    case LIFO_QUEUE: {
-      LifoQueue<StateId> queue;
-      RmEpsilon(fst, &distance, opts, &queue);
-      return;
-    }
-    case SHORTEST_FIRST_QUEUE: {
-      NaturalShortestFirstQueue<StateId, Weight> queue(distance);
-      RmEpsilon(fst, &distance, opts, &queue);
-      return;
-    }
-    case STATE_ORDER_QUEUE: {
-      StateOrderQueue<StateId> queue;
-      RmEpsilon(fst, &distance, opts, &queue);
-      return;
-    }
-    case TOP_ORDER_QUEUE: {
-      TopOrderQueue<StateId> queue(*fst, EpsilonArcFilter<Arc>());
-      internal::RmEpsilon(fst, &distance, opts, &queue);
-      return;
-    }
-    default: {
-      FSTERROR() << "RmEpsilon: Unknown queue type: "
-                 << opts.queue_type;
-      fst->SetProperties(kError, kError);
-      return;
-    }
-  }
-}
-
-}  // namespace internal
-
-using RmEpsilonArgs = std::pair<MutableFstClass *, const RmEpsilonOptions &>;
-
-template <class Arc>
-void RmEpsilon(RmEpsilonArgs *args) {
-  MutableFst<Arc> *fst = std::get<0>(*args)->GetMutableFst<Arc>();
-  const auto &opts = std::get<1>(*args);
-  internal::RmEpsilon(fst, opts);
-}
-
-void RmEpsilon(MutableFstClass *fst, const RmEpsilonOptions &opts);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_RMEPSILON_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/script-impl.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/script-impl.h
deleted file mode 100644
index 33c2853a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/script-impl.h
+++ /dev/null
@@ -1,211 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// This file defines the registration mechanism for new operations.
-// These operations are designed to enable scripts to work with FST classes
-// at a high level.
-//
-// If you have a new arc type and want these operations to work with FSTs
-// with that arc type, see below for the registration steps
-// you must take.
-//
-// These methods are only recommended for use in high-level scripting
-// applications. Most users should use the lower-level templated versions
-// corresponding to these.
-//
-// If you have a new arc type you'd like these operations to work with,
-// use the REGISTER_FST_OPERATIONS macro defined in fstscript.h.
-//
-// If you have a custom operation you'd like to define, you need four
-// components. In the following, assume you want to create a new operation
-// with the signature
-//
-//    void Foo(const FstClass &ifst, MutableFstClass *ofst);
-//
-//  You need:
-//
-//  1) A way to bundle the args that your new Foo operation will take, as
-//     a single struct. The template structs in arg-packs.h provide a handy
-//     way to do this. In Foo's case, that might look like this:
-//
-//       using FooArgs = std::pair<const FstClass &, MutableFstClass *>;
-//
-//     Note: this package of args is going to be passed by non-const pointer.
-//
-//  2) A function template that is able to perform Foo, given the args and
-//     arc type. Yours might look like this:
-//
-//       template<class Arc>
-//       void Foo(FooArgs *args) {
-//          // Pulls out the actual, arc-templated FSTs.
-//          const Fst<Arc> &ifst = std::get<0>(*args).GetFst<Arc>();
-//          MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-//          // Actually perform Foo on ifst and ofst.
-//       }
-//
-//  3) a client-facing function for your operation. This would look like
-//     the following:
-//
-//     void Foo(const FstClass &ifst, MutableFstClass *ofst) {
-//       // Check that the arc types of the FSTs match
-//       if (!ArcTypesMatch(ifst, *ofst, "Foo")) return;
-//       // package the args
-//       FooArgs args(ifst, ofst);
-//       // Finally, call the operation
-//       Apply<Operation<FooArgs>>("Foo", ifst->ArcType(), &args);
-//     }
-//
-//  The Apply<> function template takes care of the link between 2 and 3,
-//  provided you also have:
-//
-//  4) A registration for your new operation, on the arc types you care about.
-//     This can be provided easily by the REGISTER_FST_OPERATION macro in
-//     operations.h:
-//
-//       REGISTER_FST_OPERATION(Foo, StdArc, FooArgs);
-//       REGISTER_FST_OPERATION(Foo, MyArc, FooArgs);
-//       // .. etc
-//
-//
-//  That's it! Now when you call Foo(const FstClass &, MutableFstClass *),
-//  it dispatches (in #3) via the Apply<> function to the correct
-//  instantiation of the template function in #2.
-//
-
-#ifndef FST_SCRIPT_SCRIPT_IMPL_H_
-#define FST_SCRIPT_SCRIPT_IMPL_H_
-
-// This file contains general-purpose templates which are used in the
-// implementation of the operations.
-
-#include <string>
-#include <utility>
-
-#include <fst/generic-register.h>
-#include <fst/script/fst-class.h>
-
-#include <fst/log.h>
-
-namespace fst {
-namespace script {
-
-enum RandArcSelection {
-  UNIFORM_ARC_SELECTOR,
-  LOG_PROB_ARC_SELECTOR,
-  FAST_LOG_PROB_ARC_SELECTOR
-};
-
-// A generic register for operations with various kinds of signatures.
-// Needed since every function signature requires a new registration class.
-// The std::pair<string, string> is understood to be the operation name and arc
-// type; subclasses (or typedefs) need only provide the operation signature.
-
-template <class OperationSignature>
-class GenericOperationRegister
-    : public GenericRegister<std::pair<string, string>, OperationSignature,
-                             GenericOperationRegister<OperationSignature>> {
- public:
-  void RegisterOperation(const string &operation_name, const string &arc_type,
-                         OperationSignature op) {
-    this->SetEntry(std::make_pair(operation_name, arc_type), op);
-  }
-
-  OperationSignature GetOperation(const string &operation_name,
-                                  const string &arc_type) {
-    return this->GetEntry(std::make_pair(operation_name, arc_type));
-  }
-
- protected:
-  string ConvertKeyToSoFilename(
-      const std::pair<string, string> &key) const final {
-    // Uses the old-style FST for now.
-    string legal_type(key.second);  // The arc type.
-    ConvertToLegalCSymbol(&legal_type);
-    return legal_type + "-arc.so";
-  }
-};
-
-// Operation package: everything you need to register a new type of operation.
-// The ArgPack should be the type that's passed into each wrapped function;
-// for instance, it might be a struct containing all the args. It's always
-// passed by pointer, so const members should be used to enforce constness where
-// it's needed. Return values should be implemented as a member of ArgPack as
-// well.
-
-template <class Args>
-struct Operation {
-  using ArgPack = Args;
-
-  using OpType = void (*)(ArgPack *args);
-
-  // The register (hash) type.
-  using Register = GenericOperationRegister<OpType>;
-
-  // The register-er type
-  using Registerer = GenericRegisterer<Register>;
-};
-
-// Macro for registering new types of operations.
-
-#define REGISTER_FST_OPERATION(Op, Arc, ArgPack)               \
-  static fst::script::Operation<ArgPack>::Registerer       \
-      arc_dispatched_operation_##ArgPack##Op##Arc##_registerer \
-      (std::make_pair(#Op, Arc::Type()), Op<Arc>)
-
-// Template function to apply an operation by name.
-
-template <class OpReg>
-void Apply(const string &op_name, const string &arc_type,
-           typename OpReg::ArgPack *args) {
-  const auto op = OpReg::Register::GetRegister()->GetOperation(op_name,
-                                                               arc_type);
-  if (!op) {
-    FSTERROR() << "No operation found for " << op_name << " on "
-               << "arc type " << arc_type;
-    return;
-  }
-  op(args);
-}
-
-namespace internal {
-
-// Helper that logs to ERROR if the arc types of m and n don't match,
-// assuming that both m and n implement .ArcType(). The op_name argument is
-// used to construct the error message.
-template <class M, class N>
-bool ArcTypesMatch(const M &m, const N &n, const string &op_name) {
-  if (m.ArcType() != n.ArcType()) {
-    FSTERROR() << "Arguments with non-matching arc types passed to "
-               << op_name << ":\t" << m.ArcType() << " and " << n.ArcType();
-    return false;
-  }
-  return true;
-}
-
-// From untyped to typed weights.
-template <class Weight>
-void CopyWeights(const std::vector<WeightClass> &weights,
-                 std::vector<Weight> *typed_weights) {
-  typed_weights->clear();
-  typed_weights->reserve(weights.size());
-  for (const auto &weight : weights) {
-    typed_weights->push_back(*weight.GetWeight<Weight>());
-  }
-}
-
-// From typed to untyped weights.
-template <class Weight>
-void CopyWeights(const std::vector<Weight> &typed_weights,
-                 std::vector<WeightClass> *weights) {
-  weights->clear();
-  weights->reserve(typed_weights.size());
-  for (const auto &typed_weight : typed_weights) {
-    weights->emplace_back(typed_weight);
-  }
-}
-
-}  // namespace internal
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_SCRIPT_IMPL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/shortest-distance.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/shortest-distance.h
deleted file mode 100644
index 98dd1907..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/shortest-distance.h
+++ /dev/null
@@ -1,217 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_SHORTEST_DISTANCE_H_
-#define FST_SCRIPT_SHORTEST_DISTANCE_H_
-
-#include <tuple>
-#include <vector>
-
-#include <fst/queue.h>
-#include <fst/shortest-distance.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/prune.h>
-#include <fst/script/script-impl.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-enum ArcFilterType {
-  ANY_ARC_FILTER,
-  EPSILON_ARC_FILTER,
-  INPUT_EPSILON_ARC_FILTER,
-  OUTPUT_EPSILON_ARC_FILTER
-};
-
-struct ShortestDistanceOptions {
-  const QueueType queue_type;
-  const ArcFilterType arc_filter_type;
-  const int64 source;
-  const float delta;
-
-  ShortestDistanceOptions(QueueType queue_type, ArcFilterType arc_filter_type,
-                          int64 source, float delta)
-      : queue_type(queue_type),
-        arc_filter_type(arc_filter_type),
-        source(source),
-        delta(delta) {}
-};
-
-namespace internal {
-
-// Code to implement switching on queue and arc filter types.
-
-template <class Arc, class Queue, class ArcFilter>
-struct QueueConstructor {
-  using Weight = typename Arc::Weight;
-
-  static Queue *Construct(const Fst<Arc> &, const std::vector<Weight> *) {
-    return new Queue();
-  }
-};
-
-// Specializations to support queues with different constructors.
-
-template <class Arc, class ArcFilter>
-struct QueueConstructor<Arc, AutoQueue<typename Arc::StateId>, ArcFilter> {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  //  template<class Arc, class ArcFilter>
-  static AutoQueue<StateId> *Construct(const Fst<Arc> &fst,
-                                      const std::vector<Weight> *distance) {
-    return new AutoQueue<StateId>(fst, distance, ArcFilter());
-  }
-};
-
-template <class Arc, class ArcFilter>
-struct QueueConstructor<Arc,
-                        NaturalShortestFirstQueue<typename Arc::StateId,
-                                                  typename Arc::Weight>,
-                        ArcFilter> {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  static NaturalShortestFirstQueue<StateId, Weight> *Construct(
-      const Fst<Arc> &, const std::vector<Weight> *distance) {
-    return new NaturalShortestFirstQueue<StateId, Weight>(*distance);
-  }
-};
-
-template <class Arc, class ArcFilter>
-struct QueueConstructor<Arc, TopOrderQueue<typename Arc::StateId>, ArcFilter> {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  static TopOrderQueue<StateId> *Construct(const Fst<Arc> &fst,
-                                           const std::vector<Weight> *) {
-    return new TopOrderQueue<StateId>(fst, ArcFilter());
-  }
-};
-
-template <class Arc, class Queue, class ArcFilter>
-void ShortestDistance(const Fst<Arc> &fst,
-                      std::vector<typename Arc::Weight> *distance,
-                      const ShortestDistanceOptions &opts) {
-  std::unique_ptr<Queue> queue(
-      QueueConstructor<Arc, Queue, ArcFilter>::Construct(fst, distance));
-  const fst::ShortestDistanceOptions<Arc, Queue, ArcFilter> sopts(
-      queue.get(), ArcFilter(), opts.source, opts.delta);
-  ShortestDistance(fst, distance, sopts);
-}
-
-template <class Arc, class Queue>
-void ShortestDistance(const Fst<Arc> &fst,
-                      std::vector<typename Arc::Weight> *distance,
-                      const ShortestDistanceOptions &opts) {
-  switch (opts.arc_filter_type) {
-    case ANY_ARC_FILTER: {
-      ShortestDistance<Arc, Queue, AnyArcFilter<Arc>>(fst, distance, opts);
-      return;
-    }
-    case EPSILON_ARC_FILTER: {
-      ShortestDistance<Arc, Queue, EpsilonArcFilter<Arc>>(fst, distance, opts);
-      return;
-    }
-    case INPUT_EPSILON_ARC_FILTER: {
-      ShortestDistance<Arc, Queue, InputEpsilonArcFilter<Arc>>(fst,
-          distance, opts);
-      return;
-    }
-    case OUTPUT_EPSILON_ARC_FILTER: {
-      ShortestDistance<Arc, Queue, OutputEpsilonArcFilter<Arc>>(fst,
-          distance, opts);
-      return;
-    }
-    default: {
-      FSTERROR() << "ShortestDistance: Unknown arc filter type: "
-                 << opts.arc_filter_type;
-      distance->clear();
-      distance->resize(1, Arc::Weight::NoWeight());
-      return;
-    }
-  }
-}
-
-}  // namespace internal
-
-using ShortestDistanceArgs1 = std::tuple<const FstClass &,
-                                        std::vector<WeightClass> *,
-                                        const ShortestDistanceOptions &>;
-
-template <class Arc>
-void ShortestDistance(ShortestDistanceArgs1 *args) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  const auto &opts = std::get<2>(*args);
-  std::vector<Weight> typed_distance;
-  switch (opts.queue_type) {
-    case AUTO_QUEUE: {
-      internal::ShortestDistance<Arc, AutoQueue<StateId>>(fst, &typed_distance,
-                                                          opts);
-      break;
-    }
-    case FIFO_QUEUE: {
-      internal::ShortestDistance<Arc, FifoQueue<StateId>>(fst, &typed_distance,
-                                                          opts);
-      break;
-    }
-    case LIFO_QUEUE: {
-      internal::ShortestDistance<Arc, LifoQueue<StateId>>(fst, &typed_distance,
-                                                          opts);
-      break;
-    }
-    case SHORTEST_FIRST_QUEUE: {
-      internal::ShortestDistance<Arc,
-          NaturalShortestFirstQueue<StateId, Weight>>(fst, &typed_distance,
-                                                      opts);
-      break;
-    }
-    case STATE_ORDER_QUEUE: {
-      internal::ShortestDistance<Arc, StateOrderQueue<StateId>>(fst,
-                                                                &typed_distance,
-                                                                opts);
-      break;
-    }
-    case TOP_ORDER_QUEUE: {
-      internal::ShortestDistance<Arc, TopOrderQueue<StateId>>(fst,
-                                                              &typed_distance,
-                                                              opts);
-      break;
-    }
-    default: {
-      FSTERROR() << "ShortestDistance: Unknown queue type: "
-                 << opts.queue_type;
-      typed_distance.clear();
-      typed_distance.resize(1, Arc::Weight::NoWeight());
-      break;
-    }
-  }
-  internal::CopyWeights(typed_distance, std::get<1>(*args));
-}
-
-using ShortestDistanceArgs2 =
-    std::tuple<const FstClass &, std::vector<WeightClass> *, bool, double>;
-
-template <class Arc>
-void ShortestDistance(ShortestDistanceArgs2 *args) {
-  using Weight = typename Arc::Weight;
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  std::vector<Weight> typed_distance;
-  ShortestDistance(fst, &typed_distance, std::get<2>(*args),
-                   std::get<3>(*args));
-  internal::CopyWeights(typed_distance, std::get<1>(*args));
-}
-
-void ShortestDistance(const FstClass &fst, std::vector<WeightClass> *distance,
-                      const ShortestDistanceOptions &opts);
-
-void ShortestDistance(const FstClass &ifst, std::vector<WeightClass> *distance,
-                      bool reverse = false, double delta = fst::kDelta);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_SHORTEST_DISTANCE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/shortest-path.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/shortest-path.h
deleted file mode 100644
index 86bc88da..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/shortest-path.h
+++ /dev/null
@@ -1,116 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_SHORTEST_PATH_H_
-#define FST_SCRIPT_SHORTEST_PATH_H_
-
-#include <memory>
-#include <vector>
-
-#include <fst/shortest-path.h>
-#include <fst/script/fst-class.h>
-#include <fst/script/shortest-distance.h>
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-// Slightly simplified interface: `has_distance` and `first_path` are disabled.
-
-struct ShortestPathOptions : public ShortestDistanceOptions {
-  const int32 nshortest;
-  const bool unique;
-  const WeightClass &weight_threshold;
-  const int64 state_threshold;
-
-  ShortestPathOptions(QueueType queue_type, int32 nshortest, bool unique,
-                      float delta, const WeightClass &weight_threshold,
-                      int64 state_threshold = kNoStateId)
-      : ShortestDistanceOptions(queue_type, ANY_ARC_FILTER, kNoStateId, delta),
-        nshortest(nshortest),
-        unique(unique),
-        weight_threshold(weight_threshold),
-        state_threshold(state_threshold) {}
-};
-
-namespace internal {
-
-// Code to implement switching on queue types.
-
-template <class Arc, class Queue>
-void ShortestPath(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-                  std::vector<typename Arc::Weight> *distance,
-                  const ShortestPathOptions &opts) {
-  using ArcFilter = AnyArcFilter<Arc>;
-  using Weight = typename Arc::Weight;
-  const std::unique_ptr<Queue> queue(
-      QueueConstructor<Arc, Queue, ArcFilter>::Construct(ifst, distance));
-  const fst::ShortestPathOptions<Arc, Queue, ArcFilter> sopts(
-      queue.get(), ArcFilter(), opts.nshortest, opts.unique,
-      /* has_distance=*/false, opts.delta, /* first_path=*/false,
-      *opts.weight_threshold.GetWeight<Weight>(), opts.state_threshold);
-  ShortestPath(ifst, ofst, distance, sopts);
-}
-
-template <class Arc>
-void ShortestPath(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-                  const ShortestPathOptions &opts) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  std::vector<Weight> distance;
-  switch (opts.queue_type) {
-    case AUTO_QUEUE: {
-      ShortestPath<Arc, AutoQueue<StateId>>(ifst, ofst, &distance, opts);
-      return;
-    }
-    case FIFO_QUEUE: {
-      ShortestPath<Arc, FifoQueue<StateId>>(ifst, ofst, &distance, opts);
-      return;
-    }
-    case LIFO_QUEUE: {
-      ShortestPath<Arc, LifoQueue<StateId>>(ifst, ofst, &distance, opts);
-      return;
-    }
-    case SHORTEST_FIRST_QUEUE: {
-      ShortestPath<Arc, NaturalShortestFirstQueue<StateId, Weight>>(ifst, ofst,
-                                                                    &distance,
-                                                                    opts);
-      return;
-    }
-    case STATE_ORDER_QUEUE: {
-      ShortestPath<Arc, StateOrderQueue<StateId>>(ifst, ofst, &distance, opts);
-      return;
-    }
-    case TOP_ORDER_QUEUE: {
-      ShortestPath<Arc, TopOrderQueue<StateId>>(ifst, ofst, &distance, opts);
-      return;
-    }
-    default: {
-      FSTERROR() << "ShortestPath: Unknown queue type: "
-                 << opts.queue_type;
-      ofst->SetProperties(kError, kError);
-      return;
-    }
-  }
-}
-
-}  // namespace internal
-
-using ShortestPathArgs = std::tuple<const FstClass &, MutableFstClass *,
-                                    const ShortestPathOptions &>;
-
-template <class Arc>
-void ShortestPath(ShortestPathArgs *args) {
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  const ShortestPathOptions &opts = std::get<2>(*args);
-  internal::ShortestPath(ifst, ofst, opts);
-}
-
-void ShortestPath(const FstClass &ifst, MutableFstClass *ofst,
-                  const ShortestPathOptions &opts);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_SHORTEST_PATH_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/stateiterator-class.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/stateiterator-class.h
deleted file mode 100644
index f6fddfe6..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/stateiterator-class.h
+++ /dev/null
@@ -1,85 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_STATEITERATOR_CLASS_H_
-#define FST_SCRIPT_STATEITERATOR_CLASS_H_
-
-#include <memory>
-
-#include <fst/fstlib.h>
-#include <fst/script/fst-class.h>
-
-// Scripting API support for StateIterator.
-
-namespace fst {
-namespace script {
-
-// Virtual interface implemented by each concrete StateIteratorImpl<F>.
-class StateIteratorImplBase {
- public:
-  virtual bool Done() const = 0;
-  virtual int64 Value() const = 0;
-  virtual void Next() = 0;
-  virtual void Reset() = 0;
-  virtual ~StateIteratorImplBase() {}
-};
-
-// Templated implementation.
-template <class Arc>
-class StateIteratorClassImpl : public StateIteratorImplBase {
- public:
-  explicit StateIteratorClassImpl(const Fst<Arc> &fst) : siter_(fst) {}
-
-  bool Done() const final { return siter_.Done(); }
-
-  int64 Value() const final { return siter_.Value(); }
-
-  void Next() final { siter_.Next(); }
-
-  void Reset() final { siter_.Reset(); }
-
-  ~StateIteratorClassImpl() override {}
-
- private:
-  StateIterator<Fst<Arc>> siter_;
-};
-
-class StateIteratorClass;
-
-using InitStateIteratorClassArgs =
-    std::pair<const FstClass &, StateIteratorClass *>;
-
-// Untemplated user-facing class holding a templated pimpl.
-class StateIteratorClass {
- public:
-  explicit StateIteratorClass(const FstClass &fst);
-
-  template <class Arc>
-  explicit StateIteratorClass(const Fst<Arc> &fst)
-      : impl_(new StateIteratorClassImpl<Arc>(fst)) {}
-
-  bool Done() const { return impl_->Done(); }
-
-  int64 Value() const { return impl_->Value(); }
-
-  void Next() { impl_->Next(); }
-
-  void Reset() { impl_->Reset(); }
-
-  template <class Arc>
-  friend void InitStateIteratorClass(InitStateIteratorClassArgs *args);
-
- private:
-  std::unique_ptr<StateIteratorImplBase> impl_;
-};
-
-template <class Arc>
-void InitStateIteratorClass(InitStateIteratorClassArgs *args) {
-  const Fst<Arc> &fst = *(std::get<0>(*args).GetFst<Arc>());
-  std::get<1>(*args)->impl_.reset(new StateIteratorClassImpl<Arc>(fst));
-}
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_STATEITERATOR_CLASS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/synchronize.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/synchronize.h
deleted file mode 100644
index 01df151a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/synchronize.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_SYNCHRONIZE_H_
-#define FST_SCRIPT_SYNCHRONIZE_H_
-
-#include <utility>
-
-#include <fst/synchronize.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using SynchronizeArgs = std::pair<const FstClass &, MutableFstClass *>;
-
-template <class Arc>
-void Synchronize(SynchronizeArgs *args) {
-  const Fst<Arc> &ifst = *(std::get<0>(*args).GetFst<Arc>());
-  MutableFst<Arc> *ofst = std::get<1>(*args)->GetMutableFst<Arc>();
-  Synchronize(ifst, ofst);
-}
-
-void Synchronize(const FstClass &ifst, MutableFstClass *ofst);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_SYNCHRONIZE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/text-io.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/text-io.h
deleted file mode 100644
index 464bf885..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/text-io.h
+++ /dev/null
@@ -1,28 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Utilities for reading and writing textual strings representing states,
-// labels, and weights and files specifying label-label pairs and potentials
-// (state-weight pairs).
-
-#ifndef FST_SCRIPT_TEXT_IO_H__
-#define FST_SCRIPT_TEXT_IO_H__
-
-#include <string>
-#include <vector>
-
-#include <fst/script/weight-class.h>
-
-namespace fst {
-namespace script {
-
-bool ReadPotentials(const string &weight_type, const string &filename,
-                    std::vector<WeightClass> *potentials);
-
-bool WritePotentials(const string &filename,
-                     const std::vector<WeightClass> &potentials);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_TEXT_IO_H__
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/topsort.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/topsort.h
deleted file mode 100644
index fb6738d7..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/topsort.h
+++ /dev/null
@@ -1,26 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_TOPSORT_H_
-#define FST_SCRIPT_TOPSORT_H_
-
-#include <fst/topsort.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using TopSortArgs = WithReturnValue<bool, MutableFstClass *>;
-
-template <class Arc>
-void TopSort(TopSortArgs *args) {
-  args->retval = TopSort(args->args->GetMutableFst<Arc>());
-}
-
-bool TopSort(MutableFstClass *fst);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_TOPSORT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/union.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/union.h
deleted file mode 100644
index 9493e2b1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/union.h
+++ /dev/null
@@ -1,29 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_UNION_H_
-#define FST_SCRIPT_UNION_H_
-
-#include <utility>
-
-#include <fst/union.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using UnionArgs = std::pair<MutableFstClass *, const FstClass &>;
-
-template <class Arc>
-void Union(UnionArgs *args) {
-  MutableFst<Arc> *fst1 = std::get<0>(*args)->GetMutableFst<Arc>();
-  const Fst<Arc> &fst2 = *(std::get<1>(*args).GetFst<Arc>());
-  Union(fst1, fst2);
-}
-
-void Union(MutableFstClass *fst1, const FstClass &fst2);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_UNION_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/verify.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/verify.h
deleted file mode 100644
index 52f58641..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/verify.h
+++ /dev/null
@@ -1,27 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SCRIPT_VERIFY_H_
-#define FST_SCRIPT_VERIFY_H_
-
-#include <fst/verify.h>
-#include <fst/script/arg-packs.h>
-#include <fst/script/fst-class.h>
-
-namespace fst {
-namespace script {
-
-using VerifyArgs = WithReturnValue<bool, const FstClass &>;
-
-template <class Arc>
-void Verify(VerifyArgs *args) {
-  const Fst<Arc> &fst = *(args->args.GetFst<Arc>());
-  args->retval = Verify(fst);
-}
-
-bool Verify(const FstClass &fst);
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_VERIFY_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/weight-class.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/weight-class.h
deleted file mode 100644
index f1fc2b49..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/script/weight-class.h
+++ /dev/null
@@ -1,230 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Represents a generic weight in an FST; that is, represents a specific type
-// of weight underneath while hiding that type from a client.
-
-#ifndef FST_SCRIPT_WEIGHT_CLASS_H_
-#define FST_SCRIPT_WEIGHT_CLASS_H_
-
-#include <memory>
-#include <ostream>
-#include <string>
-
-#include <fst/arc.h>
-#include <fst/generic-register.h>
-#include <fst/util.h>
-#include <fst/weight.h>
-
-namespace fst {
-namespace script {
-
-class WeightImplBase {
- public:
-  virtual WeightImplBase *Copy() const = 0;
-  virtual void Print(std::ostream *o) const = 0;
-  virtual const string &Type() const = 0;
-  virtual string ToString() const = 0;
-  virtual bool operator==(const WeightImplBase &other) const = 0;
-  virtual bool operator!=(const WeightImplBase &other) const = 0;
-  virtual WeightImplBase &PlusEq(const WeightImplBase &other) = 0;
-  virtual WeightImplBase &TimesEq(const WeightImplBase &other) = 0;
-  virtual WeightImplBase &DivideEq(const WeightImplBase &other) = 0;
-  virtual WeightImplBase &PowerEq(size_t n) = 0;
-  virtual ~WeightImplBase() {}
-};
-
-template <class W>
-class WeightClassImpl : public WeightImplBase {
- public:
-  explicit WeightClassImpl(const W &weight) : weight_(weight) {}
-
-  WeightClassImpl<W> *Copy() const final {
-    return new WeightClassImpl<W>(weight_);
-  }
-
-  const string &Type() const final { return W::Type(); }
-
-  void Print(std::ostream *ostrm) const final { *ostrm << weight_; }
-
-  string ToString() const final {
-    string str;
-    WeightToStr(weight_, &str);
-    return str;
-  }
-
-  bool operator==(const WeightImplBase &other) const final {
-    const auto *typed_other = static_cast<const WeightClassImpl<W> *>(&other);
-    return weight_ == typed_other->weight_;
-  }
-
-  bool operator!=(const WeightImplBase &other) const final {
-    return !(*this == other);
-  }
-
-  WeightClassImpl<W> &PlusEq(const WeightImplBase &other) final {
-    const auto *typed_other = static_cast<const WeightClassImpl<W> *>(&other);
-    weight_ = Plus(weight_, typed_other->weight_);
-    return *this;
-  }
-
-  WeightClassImpl<W> &TimesEq(const WeightImplBase &other) final {
-    const auto *typed_other = static_cast<const WeightClassImpl<W> *>(&other);
-    weight_ = Times(weight_, typed_other->weight_);
-    return *this;
-  }
-
-  WeightClassImpl<W> &DivideEq(const WeightImplBase &other) final {
-    const auto *typed_other = static_cast<const WeightClassImpl<W> *>(&other);
-    weight_ = Divide(weight_, typed_other->weight_);
-    return *this;
-  }
-
-  WeightClassImpl<W> &PowerEq(size_t n) final {
-    weight_ = Power(weight_, n);
-    return *this;
-  }
-
-  W *GetImpl() { return &weight_; }
-
- private:
-  W weight_;
-};
-
-
-class WeightClass {
- public:
-  WeightClass() = default;
-
-  template <class W>
-  explicit WeightClass(const W &weight)
-      : impl_(new WeightClassImpl<W>(weight)) {}
-
-  template <class W>
-  explicit WeightClass(const WeightClassImpl<W> &impl)
-      : impl_(new WeightClassImpl<W>(impl)) {}
-
-  WeightClass(const string &weight_type, const string &weight_str);
-
-  WeightClass(const WeightClass &other)
-      : impl_(other.impl_ ? other.impl_->Copy() : nullptr) {}
-
-  WeightClass &operator=(const WeightClass &other) {
-    impl_.reset(other.impl_ ? other.impl_->Copy() : nullptr);
-    return *this;
-  }
-
-  static constexpr const char *__ZERO__ = "__ZERO__";  // NOLINT
-
-  static WeightClass Zero(const string &weight_type);
-
-  static constexpr const char *__ONE__ = "__ONE__";  // NOLINT
-
-  static WeightClass One(const string &weight_type);
-
-  static constexpr const char *__NOWEIGHT__ = "__NOWEIGHT__";  // NOLINT
-
-  static WeightClass NoWeight(const string &weight_type);
-
-  template <class W>
-  const W *GetWeight() const {
-    if (W::Type() != impl_->Type()) {
-       return nullptr;
-    } else {
-      auto *typed_impl = static_cast<WeightClassImpl<W> *>(impl_.get());
-      return typed_impl->GetImpl();
-    }
-  }
-
-  string ToString() const { return (impl_) ? impl_->ToString() : "none"; }
-
-  const string &Type() const {
-    if (impl_) return impl_->Type();
-    static const string *const no_type = new string("none");
-    return *no_type;
-  }
-
-  bool WeightTypesMatch(const WeightClass &other, const string &op_name) const;
-
-  friend bool operator==(const WeightClass &lhs, const WeightClass &rhs);
-
-  friend WeightClass Plus(const WeightClass &lhs, const WeightClass &rhs);
-
-  friend WeightClass Times(const WeightClass &lhs, const WeightClass &rhs);
-
-  friend WeightClass Divide(const WeightClass &lhs, const WeightClass &rhs);
-
-  friend WeightClass Power(const WeightClass &w, size_t n);
-
- private:
-  const WeightImplBase *GetImpl() const { return impl_.get(); }
-
-  WeightImplBase *GetImpl() { return impl_.get(); }
-
-  std::unique_ptr<WeightImplBase> impl_;
-
-  friend std::ostream &operator<<(std::ostream &o, const WeightClass &c);
-};
-
-bool operator==(const WeightClass &lhs, const WeightClass &rhs);
-
-bool operator!=(const WeightClass &lhs, const WeightClass &rhs);
-
-WeightClass Plus(const WeightClass &lhs, const WeightClass &rhs);
-
-WeightClass Times(const WeightClass &lhs, const WeightClass &rhs);
-
-WeightClass Divide(const WeightClass &lhs, const WeightClass &rhs);
-
-WeightClass Power(const WeightClass &w, size_t n);
-
-std::ostream &operator<<(std::ostream &o, const WeightClass &c);
-
-// Registration for generic weight types.
-
-using StrToWeightImplBaseT = WeightImplBase *(*)(const string &str,
-                                                 const string &src,
-                                                 size_t nline);
-
-template <class W>
-WeightImplBase *StrToWeightImplBase(const string &str, const string &src,
-                                    size_t nline) {
-  if (str == WeightClass::__ZERO__)
-    return new WeightClassImpl<W>(W::Zero());
-  else if (str == WeightClass::__ONE__)
-    return new WeightClassImpl<W>(W::One());
-  else if (str == WeightClass::__NOWEIGHT__)
-    return new WeightClassImpl<W>(W::NoWeight());
-  return new WeightClassImpl<W>(StrToWeight<W>(str, src, nline));
-}
-
-class WeightClassRegister : public GenericRegister<string, StrToWeightImplBaseT,
-                                                   WeightClassRegister> {
- protected:
-  string ConvertKeyToSoFilename(const string &key) const final {
-    string legal_type(key);
-    ConvertToLegalCSymbol(&legal_type);
-    return legal_type + ".so";
-  }
-};
-
-using WeightClassRegisterer = GenericRegisterer<WeightClassRegister>;
-
-// Internal version; needs to be called by wrapper in order for macro args to
-// expand.
-#define REGISTER_FST_WEIGHT__(Weight, line)                \
-  static WeightClassRegisterer weight_registerer##_##line( \
-      Weight::Type(), StrToWeightImplBase<Weight>)
-
-// This layer is where __FILE__ and __LINE__ are expanded.
-#define REGISTER_FST_WEIGHT_EXPANDER(Weight, line) \
-  REGISTER_FST_WEIGHT__(Weight, line)
-
-// Macro for registering new weight types. Clients call this.
-#define REGISTER_FST_WEIGHT(Weight) \
-  REGISTER_FST_WEIGHT_EXPANDER(Weight, __LINE__)
-
-}  // namespace script
-}  // namespace fst
-
-#endif  // FST_SCRIPT_WEIGHT_CLASS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/set-weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/set-weight.h
deleted file mode 100644
index f106e45b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/set-weight.h
+++ /dev/null
@@ -1,592 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Weights consisting of sets (of integral Labels) and
-// associated semiring operation definitions using intersect
-// and union.
-
-#ifndef FST_SET_WEIGHT_H_
-#define FST_SET_WEIGHT_H_
-
-#include <cstdlib>
-
-#include <algorithm>
-#include <list>
-#include <string>
-#include <vector>
-
-#include <fst/union-weight.h>
-#include <fst/weight.h>
-
-
-namespace fst {
-
-constexpr int kSetEmpty = 0;         // Label for the empty set.
-constexpr int kSetUniv = -1;         // Label for the universal set.
-constexpr int kSetBad = -2;          // Label for a non-set.
-constexpr char kSetSeparator = '_';  // Label separator in sets.
-
-// Determines whether to use (intersect, union) or (union, intersect)
-// as (+, *) for the semiring. SET_INTERSECT_UNION_RESTRICTED is a
-// restricted version of (intersect, union) that requires summed
-// arguments to be equal (or an error is signalled), useful for
-// algorithms that require a unique labelled path weight.  SET_BOOLEAN
-// treats all non-Zero() elements as equivalent (with Zero() ==
-// UnivSet()), useful for algorithms that don't really depend on the
-// detailed sets.
-enum SetType { SET_INTERSECT_UNION = 0,
-               SET_UNION_INTERSECT = 1,
-               SET_INTERSECT_UNION_RESTRICT = 2,
-               SET_BOOLEAN = 3 };
-
-template <class>
-class SetWeightIterator;
-
-// Set semiring of integral labels.
-template <typename Label_, SetType S = SET_INTERSECT_UNION>
-class SetWeight {
- public:
-  using Label = Label_;
-  using ReverseWeight = SetWeight<Label, S>;
-  using Iterator = SetWeightIterator<SetWeight>;
-  friend class SetWeightIterator<SetWeight>;
-
-  SetWeight() {}
-
-  // Input should be positive, sorted and unique.
-  template <typename Iterator>
-  SetWeight(const Iterator &begin, const Iterator &end) {
-    for (auto iter = begin; iter != end; ++iter) PushBack(*iter);
-  }
-
-  // Input should be positive. (Non-positive value has
-  // special internal meaning w.r.t. integral constants above.)
-  explicit SetWeight(Label label) { PushBack(label); }
-
-  static const SetWeight &Zero() {
-    return S == SET_UNION_INTERSECT ? EmptySet() : UnivSet();
-  }
-
-  static const SetWeight &One() {
-    return S == SET_UNION_INTERSECT ? UnivSet() : EmptySet();
-  }
-
-  static const SetWeight &NoWeight() {
-    static const auto *const no_weight = new SetWeight(Label(kSetBad));
-    return *no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string(
-        S == SET_UNION_INTERSECT
-        ? "union_intersect_set"
-        : (S == SET_INTERSECT_UNION
-           ? "intersect_union_set"
-           : (S == SET_INTERSECT_UNION_RESTRICT
-              ? "restricted_set_intersect_union"
-              : "boolean_set")));
-    return *type;
-  }
-
-  bool Member() const;
-
-  std::istream &Read(std::istream &strm);
-
-  std::ostream &Write(std::ostream &strm) const;
-
-  size_t Hash() const;
-
-  SetWeight Quantize(float delta = kDelta) const { return *this; }
-
-  ReverseWeight Reverse() const;
-
-  static constexpr uint64 Properties() {
-    return kIdempotent | kLeftSemiring | kRightSemiring | kCommutative;
-  }
-
-  // These operations combined with the SetWeightIterator
-  // provide the access and mutation of the set internal elements.
-
-  // The empty set.
-  static const SetWeight &EmptySet() {
-    static const auto *const empty = new SetWeight(Label(kSetEmpty));
-    return *empty;
-  }
-
-  // The univeral set.
-  static const SetWeight &UnivSet() {
-    static const auto *const univ = new SetWeight(Label(kSetUniv));
-    return *univ;
-  }
-
-  // Clear existing SetWeight.
-  void Clear() {
-    first_ = kSetEmpty;
-    rest_.clear();
-  }
-
-  size_t Size() const { return first_ == kSetEmpty ? 0 : rest_.size() + 1; }
-
-  Label Back() {
-    if (rest_.empty()) {
-      return first_;
-    } else {
-      return rest_.back();
-    }
-  }
-
-  // Caller must add in sort order and be unique (or error signalled).
-  // Input should also be positive. Non-positive value (for the first
-  // push) has special internal meaning w.r.t. integral constants above.
-  void PushBack(Label label) {
-    if (first_ == kSetEmpty) {
-      first_ = label;
-    } else {
-      if (label <= Back() || label <= 0) {
-        FSTERROR() << "SetWeight: labels must be positive, added"
-                   << " in sort order and be unique.";
-        rest_.push_back(Label(kSetBad));
-      }
-      rest_.push_back(label);
-    }
-  }
-
- private:
-  Label first_ = kSetEmpty;  // First label in set (kSetEmpty if empty).
-  std::list<Label> rest_;    // Remaining labels in set.
-};
-
-// Traverses set in forward direction.
-template <class SetWeight_>
-class SetWeightIterator {
- public:
-  using Weight = SetWeight_;
-  using Label = typename Weight::Label;
-
-  explicit SetWeightIterator(const Weight &w)
-      : first_(w.first_), rest_(w.rest_), init_(true), iter_(rest_.begin()) {}
-
-  bool Done() const {
-    if (init_) {
-      return first_ == kSetEmpty;
-    } else {
-      return iter_ == rest_.end();
-    }
-  }
-
-  const Label &Value() const { return init_ ? first_ : *iter_; }
-
-  void Next() {
-    if (init_) {
-      init_ = false;
-    } else {
-      ++iter_;
-    }
-  }
-
-  void Reset() {
-    init_ = true;
-    iter_ = rest_.begin();
-  }
-
- private:
-  const Label &first_;
-  const decltype(Weight::rest_) &rest_;
-  bool init_;  // In the initialized state?
-  typename std::remove_reference<decltype(Weight::rest_)>::type::const_iterator iter_;
-};
-
-
-// SetWeight member functions follow that require SetWeightIterator
-
-template <typename Label, SetType S>
-inline std::istream &SetWeight<Label, S>::Read(std::istream &strm) {
-  Clear();
-  int32 size;
-  ReadType(strm, &size);
-  for (int32 i = 0; i < size; ++i) {
-    Label label;
-    ReadType(strm, &label);
-    PushBack(label);
-  }
-  return strm;
-}
-
-template <typename Label, SetType S>
-inline std::ostream &SetWeight<Label, S>::Write(std::ostream &strm) const {
-  const int32 size = Size();
-  WriteType(strm, size);
-  for (Iterator iter(*this); !iter.Done(); iter.Next()) {
-    WriteType(strm, iter.Value());
-  }
-  return strm;
-}
-
-template <typename Label, SetType S>
-inline bool SetWeight<Label, S>::Member() const {
-  Iterator iter(*this);
-  return iter.Value() != Label(kSetBad);
-}
-
-template <typename Label, SetType S>
-inline typename SetWeight<Label, S>::ReverseWeight
-SetWeight<Label, S>::Reverse() const {
-  return *this;
-}
-
-template <typename Label, SetType S>
-inline size_t SetWeight<Label, S>::Hash() const {
-  using Weight = SetWeight<Label, S>;
-  if (S == SET_BOOLEAN) {
-    return *this == Weight::Zero() ? 0 : 1;
-  } else {
-    size_t h = 0;
-    for (Iterator iter(*this); !iter.Done(); iter.Next()) {
-      h ^= h << 1 ^ iter.Value();
-    }
-    return h;
-  }
-}
-
-// Default ==
-template <typename Label, SetType S>
-inline bool operator==(const SetWeight<Label, S> &w1,
-                       const SetWeight<Label, S> &w2) {
-  if (w1.Size() != w2.Size()) return false;
-  using Iterator = typename SetWeight<Label, S>::Iterator;
-  Iterator iter1(w1);
-  Iterator iter2(w2);
-  for (; !iter1.Done(); iter1.Next(), iter2.Next()) {
-    if (iter1.Value() != iter2.Value()) return false;
-  }
-  return true;
-}
-
-// Boolean ==
-template <typename Label>
-inline bool operator==(const SetWeight<Label, SET_BOOLEAN> &w1,
-                       const SetWeight<Label, SET_BOOLEAN> &w2) {
-  // x == kSetEmpty if x \nin {kUnivSet, kSetBad}
-  if (!w1.Member() || !w2.Member()) return false;
-  using Iterator = typename SetWeight<Label, SET_BOOLEAN>::Iterator;
-  Iterator iter1(w1);
-  Iterator iter2(w2);
-  Label label1 = iter1.Done() ? kSetEmpty : iter1.Value();
-  Label label2 = iter2.Done() ? kSetEmpty : iter2.Value();
-  if (label1 == kSetUniv) return label2 == kSetUniv;
-  if (label2 == kSetUniv) return label1 == kSetUniv;
-  return true;
-}
-
-template <typename Label, SetType S>
-inline bool operator!=(const SetWeight<Label, S> &w1,
-                       const SetWeight<Label, S> &w2) {
-  return !(w1 == w2);
-}
-
-template <typename Label, SetType S>
-inline bool ApproxEqual(const SetWeight<Label, S> &w1,
-                        const SetWeight<Label, S> &w2,
-                        float delta = kDelta) {
-  return w1 == w2;
-}
-
-template <typename Label, SetType S>
-inline std::ostream &operator<<(std::ostream &strm,
-                                const SetWeight<Label, S> &weight) {
-  typename SetWeight<Label, S>::Iterator iter(weight);
-  if (iter.Done()) {
-    return strm << "EmptySet";
-  } else if (iter.Value() == Label(kSetUniv)) {
-    return strm << "UnivSet";
-  } else if (iter.Value() == Label(kSetBad)) {
-    return strm << "BadSet";
-  } else {
-    for (size_t i = 0; !iter.Done(); ++i, iter.Next()) {
-      if (i > 0) strm << kSetSeparator;
-      strm << iter.Value();
-    }
-  }
-  return strm;
-}
-
-template <typename Label, SetType S>
-inline std::istream &operator>>(std::istream &strm,
-                                SetWeight<Label, S> &weight) {
-  string str;
-  strm >> str;
-  using Weight = SetWeight<Label, S>;
-  if (str == "EmptySet") {
-    weight = Weight(Label(kSetEmpty));
-  } else if (str == "UnivSet") {
-    weight = Weight(Label(kSetUniv));
-  } else {
-    weight.Clear();
-    char *p = nullptr;
-    for (const char *cs = str.c_str(); !p || *p != '\0'; cs = p + 1) {
-      const Label label = strtoll(cs, &p, 10);
-      if (p == cs || (*p != 0 && *p != kSetSeparator)) {
-        strm.clear(std::ios::badbit);
-        break;
-      }
-      weight.PushBack(label);
-    }
-  }
-  return strm;
-}
-
-template <typename Label, SetType S>
-inline SetWeight<Label, S> Union(
-    const SetWeight<Label, S> &w1,
-    const SetWeight<Label, S> &w2) {
-  using Weight = SetWeight<Label, S>;
-  using Iterator = typename SetWeight<Label, S>::Iterator;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::EmptySet()) return w2;
-  if (w2 == Weight::EmptySet()) return w1;
-  if (w1 == Weight::UnivSet()) return w1;
-  if (w2 == Weight::UnivSet()) return w2;
-  Iterator it1(w1);
-  Iterator it2(w2);
-  Weight result;
-  while (!it1.Done() && !it2.Done()) {
-    const auto v1 = it1.Value();
-    const auto v2 = it2.Value();
-    if (v1 < v2) {
-      result.PushBack(v1);
-      it1.Next();
-    } else if (v1 > v2) {
-      result.PushBack(v2);
-      it2.Next();
-    } else {
-      result.PushBack(v1);
-      it1.Next();
-      it2.Next();
-    }
-  }
-  for (; !it1.Done(); it1.Next()) result.PushBack(it1.Value());
-  for (; !it2.Done(); it2.Next()) result.PushBack(it2.Value());
-  return result;
-}
-
-template <typename Label, SetType S>
-inline SetWeight<Label, S> Intersect(
-    const SetWeight<Label, S> &w1,
-    const SetWeight<Label, S> &w2) {
-  using Weight = SetWeight<Label, S>;
-  using Iterator = typename SetWeight<Label, S>::Iterator;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::EmptySet()) return w1;
-  if (w2 == Weight::EmptySet()) return w2;
-  if (w1 == Weight::UnivSet()) return w2;
-  if (w2 == Weight::UnivSet()) return w1;
-  Iterator it1(w1);
-  Iterator it2(w2);
-  Weight result;
-  while (!it1.Done() && !it2.Done()) {
-    const auto v1 = it1.Value();
-    const auto v2 = it2.Value();
-    if (v1 < v2) {
-      it1.Next();
-    } else if (v1 > v2) {
-      it2.Next();
-    } else {
-      result.PushBack(v1);
-      it1.Next();
-      it2.Next();
-    }
-  }
-  return result;
-}
-
-template <typename Label, SetType S>
-inline SetWeight<Label, S> Difference(
-    const SetWeight<Label, S> &w1,
-    const SetWeight<Label, S> &w2) {
-  using Weight = SetWeight<Label, S>;
-  using Iterator = typename SetWeight<Label, S>::Iterator;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::EmptySet()) return w1;
-  if (w2 == Weight::EmptySet()) return w1;
-  if (w2 == Weight::UnivSet()) return Weight::EmptySet();
-  Iterator it1(w1);
-  Iterator it2(w2);
-  Weight result;
-  while (!it1.Done() && !it2.Done()) {
-    const auto v1 = it1.Value();
-    const auto v2 = it2.Value();
-    if (v1 < v2) {
-      result.PushBack(v1);
-      it1.Next();
-    } else if (v1 > v2) {
-      it2.Next();
-    } else {
-      it1.Next();
-      it2.Next();
-    }
-  }
-  for (; !it1.Done(); it1.Next()) result.PushBack(it1.Value());
-  return result;
-}
-
-// Default: Plus = Intersect.
-template <typename Label, SetType S>
-inline SetWeight<Label, S> Plus(
-    const SetWeight<Label, S> &w1,
-    const SetWeight<Label, S> &w2) {
-  return Intersect(w1, w2);
-}
-
-// Plus = Union.
-template <typename Label>
-inline SetWeight<Label, SET_UNION_INTERSECT> Plus(
-    const SetWeight<Label, SET_UNION_INTERSECT> &w1,
-    const SetWeight<Label, SET_UNION_INTERSECT> &w2) {
-  return Union(w1, w2);
-}
-
-// Plus = Set equality is required (for non-Zero() input). The
-// restriction is useful (e.g., in determinization) to ensure the input
-// has a unique labelled path weight.
-template <typename Label>
-inline SetWeight<Label, SET_INTERSECT_UNION_RESTRICT> Plus(
-    const SetWeight<Label, SET_INTERSECT_UNION_RESTRICT> &w1,
-    const SetWeight<Label, SET_INTERSECT_UNION_RESTRICT> &w2) {
-  using Weight = SetWeight<Label, SET_INTERSECT_UNION_RESTRICT>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::Zero()) return w2;
-  if (w2 == Weight::Zero()) return w1;
-  if (w1 != w2) {
-    FSTERROR() << "SetWeight::Plus: Unequal arguments "
-               << "(non-unique labelled path weights?)"
-               << " w1 = " << w1 << " w2 = " << w2;
-    return Weight::NoWeight();
-  }
-  return w1;
-}
-
-// Plus = Or.
-template <typename Label>
-inline SetWeight<Label, SET_BOOLEAN> Plus(
-    const SetWeight<Label, SET_BOOLEAN> &w1,
-    const SetWeight<Label, SET_BOOLEAN> &w2) {
-  using Weight = SetWeight<Label, SET_BOOLEAN>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::One()) return w1;
-  if (w2 == Weight::One()) return w2;
-  return Weight::Zero();
-}
-
-// Default: Times = Union.
-template <typename Label, SetType S>
-inline SetWeight<Label, S> Times(
-    const SetWeight<Label, S> &w1,
-    const SetWeight<Label, S> &w2) {
-  return Union(w1, w2);
-}
-
-// Times = Intersect.
-template <typename Label>
-inline SetWeight<Label, SET_UNION_INTERSECT> Times(
-    const SetWeight<Label, SET_UNION_INTERSECT> &w1,
-    const SetWeight<Label, SET_UNION_INTERSECT> &w2) {
-  return Intersect(w1, w2);
-}
-
-// Times = And.
-template <typename Label>
-inline SetWeight<Label, SET_BOOLEAN> Times(
-    const SetWeight<Label, SET_BOOLEAN> &w1,
-    const SetWeight<Label, SET_BOOLEAN> &w2) {
-  using Weight = SetWeight<Label, SET_BOOLEAN>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::One()) return w2;
-  return w1;
-}
-
-// Divide = Difference.
-template <typename Label, SetType S>
-inline SetWeight<Label, S> Divide(const SetWeight<Label, S> &w1,
-                                  const SetWeight<Label, S> &w2,
-                                  DivideType divide_type = DIVIDE_ANY) {
-  return Difference(w1, w2);
-}
-
-// Divide = dividend (or the universal set if the
-// dividend == divisor).
-template <typename Label>
-inline SetWeight<Label, SET_UNION_INTERSECT> Divide(
-    const SetWeight<Label, SET_UNION_INTERSECT> &w1,
-    const SetWeight<Label, SET_UNION_INTERSECT> &w2,
-    DivideType divide_type = DIVIDE_ANY) {
-  using Weight = SetWeight<Label, SET_UNION_INTERSECT>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == w2) return Weight::UnivSet();
-  return w1;
-}
-
-// Divide = Or Not.
-template <typename Label>
-inline SetWeight<Label, SET_BOOLEAN> Divide(
-    const SetWeight<Label, SET_BOOLEAN> &w1,
-    const SetWeight<Label, SET_BOOLEAN> &w2,
-    DivideType divide_type = DIVIDE_ANY) {
-  using Weight = SetWeight<Label, SET_BOOLEAN>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::One()) return w1;
-  if (w2 == Weight::Zero()) return Weight::One();
-  return Weight::Zero();
-}
-
-// Converts between different set types.
-template <typename Label, SetType S1, SetType S2>
-struct WeightConvert<SetWeight<Label, S1>, SetWeight<Label, S2>> {
-  SetWeight<Label, S2> operator()(const SetWeight<Label, S1> &w1) const {
-    using Iterator = SetWeightIterator<SetWeight<Label, S1>>;
-    SetWeight<Label, S2> w2;
-    for (Iterator iter(w1); !iter.Done(); iter.Next())
-      w2.PushBack(iter.Value());
-    return w2;
-  }
-};
-
-// This function object generates SetWeights that are random integer sets
-// from {1, ... , alphabet_size}^{0, max_set_length} U { Zero }. This is
-// intended primarily for testing.
-template <class Label, SetType S>
-class WeightGenerate<SetWeight<Label, S>> {
- public:
-  using Weight = SetWeight<Label, S>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t alphabet_size = kNumRandomWeights,
-                          size_t max_set_length = kNumRandomWeights)
-      : allow_zero_(allow_zero),
-        alphabet_size_(alphabet_size),
-        max_set_length_(max_set_length) {}
-
-  Weight operator()() const {
-    const size_t n = rand() % (max_set_length_ + allow_zero_);  // NOLINT
-    if (allow_zero_ && n == max_set_length_) return Weight::Zero();
-    std::vector<Label> labels;
-    for (size_t i = 0; i < n; ++i) {
-      labels.push_back(rand() % alphabet_size_ + 1);  // NOLINT
-    }
-    std::sort(labels.begin(), labels.end());
-    const auto labels_end = std::unique(labels.begin(), labels.end());
-    labels.resize(labels_end - labels.begin());
-    return Weight(labels.begin(), labels.end());
-  }
-
- private:
-  // Permits Zero() and zero divisors.
-  const bool allow_zero_;
-  // Alphabet size for random weights.
-  const size_t alphabet_size_;
-  // Number of alternative random weights.
-  const size_t max_set_length_;
-};
-
-}  // namespace fst
-
-#endif  // FST_SET_WEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/shortest-distance.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/shortest-distance.h
deleted file mode 100644
index 3bdb3753..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/shortest-distance.h
+++ /dev/null
@@ -1,336 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to find shortest distance in an FST.
-
-#ifndef FST_SHORTEST_DISTANCE_H_
-#define FST_SHORTEST_DISTANCE_H_
-
-#include <deque>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/arcfilter.h>
-#include <fst/cache.h>
-#include <fst/queue.h>
-#include <fst/reverse.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-template <class Arc, class Queue, class ArcFilter>
-struct ShortestDistanceOptions {
-  using StateId = typename Arc::StateId;
-
-  Queue *state_queue;    // Queue discipline used; owned by caller.
-  ArcFilter arc_filter;  // Arc filter (e.g., limit to only epsilon graph).
-  StateId source;        // If kNoStateId, use the FST's initial state.
-  float delta;           // Determines the degree of convergence required
-  bool first_path;       // For a semiring with the path property (o.w.
-                         // undefined), compute the shortest-distances along
-                         // along the first path to a final state found
-                         // by the algorithm. That path is the shortest-path
-                         // only if the FST has a unique final state (or all
-                         // the final states have the same final weight), the
-                         // queue discipline is shortest-first and all the
-                         // weights in the FST are between One() and Zero()
-                         // according to NaturalLess.
-
-  ShortestDistanceOptions(Queue *state_queue, ArcFilter arc_filter,
-                          StateId source = kNoStateId, float delta = kDelta)
-      : state_queue(state_queue),
-        arc_filter(arc_filter),
-        source(source),
-        delta(delta),
-        first_path(false) {}
-};
-
-namespace internal {
-
-// Computation state of the shortest-distance algorithm. Reusable information
-// is maintained across calls to member function ShortestDistance(source) when
-// retain is true for improved efficiency when calling multiple times from
-// different source states (e.g., in epsilon removal). Contrary to the usual
-// conventions, fst may not be freed before this class. Vector distance
-// should not be modified by the user between these calls. The Error() method
-// returns true iff an error was encountered.
-template <class Arc, class Queue, class ArcFilter>
-class ShortestDistanceState {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  ShortestDistanceState(
-      const Fst<Arc> &fst, std::vector<Weight> *distance,
-      const ShortestDistanceOptions<Arc, Queue, ArcFilter> &opts, bool retain)
-      : fst_(fst),
-        distance_(distance),
-        state_queue_(opts.state_queue),
-        arc_filter_(opts.arc_filter),
-        delta_(opts.delta),
-        first_path_(opts.first_path),
-        retain_(retain),
-        source_id_(0),
-        error_(false) {
-    distance_->clear();
-  }
-
-  void ShortestDistance(StateId source);
-
-  bool Error() const { return error_; }
-
- private:
-  const Fst<Arc> &fst_;
-  std::vector<Weight> *distance_;
-  Queue *state_queue_;
-  ArcFilter arc_filter_;
-  const float delta_;
-  const bool first_path_;
-  const bool retain_;  // Retain and reuse information across calls.
-
-  std::vector<Adder<Weight>> adder_;   // Sums distance_ accurately.
-  std::vector<Adder<Weight>> radder_;  // Relaxation distance.
-  std::vector<bool> enqueued_;     // Is state enqueued?
-  std::vector<StateId> sources_;   // Source ID for ith state in distance_,
-                                   // (r)adder_, and enqueued_ if retained.
-  StateId source_id_;              // Unique ID characterizing each call.
-  bool error_;
-};
-
-// Compute the shortest distance; if source is kNoStateId, uses the initial
-// state of the FST.
-template <class Arc, class Queue, class ArcFilter>
-void ShortestDistanceState<Arc, Queue, ArcFilter>::ShortestDistance(
-    StateId source) {
-  if (fst_.Start() == kNoStateId) {
-    if (fst_.Properties(kError, false)) error_ = true;
-    return;
-  }
-  if (!(Weight::Properties() & kRightSemiring)) {
-    FSTERROR() << "ShortestDistance: Weight needs to be right distributive: "
-               << Weight::Type();
-    error_ = true;
-    return;
-  }
-  if (first_path_ && !(Weight::Properties() & kPath)) {
-    FSTERROR() << "ShortestDistance: The first_path option is disallowed when "
-               << "Weight does not have the path property: " << Weight::Type();
-    error_ = true;
-    return;
-  }
-  state_queue_->Clear();
-  if (!retain_) {
-    distance_->clear();
-    adder_.clear();
-    radder_.clear();
-    enqueued_.clear();
-  }
-  if (source == kNoStateId) source = fst_.Start();
-  while (distance_->size() <= source) {
-    distance_->push_back(Weight::Zero());
-    adder_.push_back(Adder<Weight>());
-    radder_.push_back(Adder<Weight>());
-    enqueued_.push_back(false);
-  }
-  if (retain_) {
-    while (sources_.size() <= source) sources_.push_back(kNoStateId);
-    sources_[source] = source_id_;
-  }
-  (*distance_)[source] = Weight::One();
-  adder_[source].Reset(Weight::One());
-  radder_[source].Reset(Weight::One());
-  enqueued_[source] = true;
-  state_queue_->Enqueue(source);
-  while (!state_queue_->Empty()) {
-    const auto state = state_queue_->Head();
-    state_queue_->Dequeue();
-    while (distance_->size() <= state) {
-      distance_->push_back(Weight::Zero());
-      adder_.push_back(Adder<Weight>());
-      radder_.push_back(Adder<Weight>());
-      enqueued_.push_back(false);
-    }
-    if (first_path_ && (fst_.Final(state) != Weight::Zero())) break;
-    enqueued_[state] = false;
-    const auto r = radder_[state].Sum();
-    radder_[state].Reset();
-    for (ArcIterator<Fst<Arc>> aiter(fst_, state); !aiter.Done();
-         aiter.Next()) {
-      const auto &arc = aiter.Value();
-      if (!arc_filter_(arc)) continue;
-      while (distance_->size() <= arc.nextstate) {
-        distance_->push_back(Weight::Zero());
-        adder_.push_back(Adder<Weight>());
-        radder_.push_back(Adder<Weight>());
-        enqueued_.push_back(false);
-      }
-      if (retain_) {
-        while (sources_.size() <= arc.nextstate) sources_.push_back(kNoStateId);
-        if (sources_[arc.nextstate] != source_id_) {
-          (*distance_)[arc.nextstate] = Weight::Zero();
-          adder_[arc.nextstate].Reset();
-          radder_[arc.nextstate].Reset();
-          enqueued_[arc.nextstate] = false;
-          sources_[arc.nextstate] = source_id_;
-        }
-      }
-      auto &nd = (*distance_)[arc.nextstate];
-      auto &na = adder_[arc.nextstate];
-      auto &nr = radder_[arc.nextstate];
-      auto weight = Times(r, arc.weight);
-      if (!ApproxEqual(nd, Plus(nd, weight), delta_)) {
-        nd = na.Add(weight);
-        nr.Add(weight);
-        if (!nd.Member() || !nr.Sum().Member()) {
-          error_ = true;
-          return;
-        }
-        if (!enqueued_[arc.nextstate]) {
-          state_queue_->Enqueue(arc.nextstate);
-          enqueued_[arc.nextstate] = true;
-        } else {
-          state_queue_->Update(arc.nextstate);
-        }
-      }
-    }
-  }
-  ++source_id_;
-  if (fst_.Properties(kError, false)) error_ = true;
-}
-
-}  // namespace internal
-
-// Shortest-distance algorithm: this version allows fine control
-// via the options argument. See below for a simpler interface.
-//
-// This computes the shortest distance from the opts.source state to each
-// visited state S and stores the value in the distance vector. An
-// nvisited state S has distance Zero(), which will be stored in the
-// distance vector if S is less than the maximum visited state. The state
-// queue discipline, arc filter, and convergence delta are taken in the
-// options argument. The distance vector will contain a unique element for
-// which Member() is false if an error was encountered.
-//
-// The weights must must be right distributive and k-closed (i.e., 1 +
-// x + x^2 + ... + x^(k +1) = 1 + x + x^2 + ... + x^k).
-//
-// Complexity:
-//
-// Depends on properties of the semiring and the queue discipline.
-//
-// For more information, see:
-//
-// Mohri, M. 2002. Semiring framework and algorithms for shortest-distance
-// problems, Journal of Automata, Languages and
-// Combinatorics 7(3): 321-350, 2002.
-template <class Arc, class Queue, class ArcFilter>
-void ShortestDistance(
-    const Fst<Arc> &fst, std::vector<typename Arc::Weight> *distance,
-    const ShortestDistanceOptions<Arc, Queue, ArcFilter> &opts) {
-  internal::ShortestDistanceState<Arc, Queue, ArcFilter> sd_state(fst, distance,
-                                                                  opts, false);
-  sd_state.ShortestDistance(opts.source);
-  if (sd_state.Error()) {
-    distance->clear();
-    distance->resize(1, Arc::Weight::NoWeight());
-  }
-}
-
-// Shortest-distance algorithm: simplified interface. See above for a version
-// that permits finer control.
-//
-// If reverse is false, this computes the shortest distance from the initial
-// state to each state S and stores the value in the distance vector. If
-// reverse is true, this computes the shortest distance from each state to the
-// final states. An unvisited state S has distance Zero(), which will be stored
-// in the distance vector if S is less than the maximum visited state. The
-// state queue discipline is automatically-selected. The distance vector will
-// contain a unique element for which Member() is false if an error was
-// encountered.
-//
-// The weights must must be right (left) distributive if reverse is false (true)
-// and k-closed (i.e., 1 + x + x^2 + ... + x^(k +1) = 1 + x + x^2 + ... + x^k).
-//
-// Arc weights must satisfy the property that the sum of the weights of one or
-// more paths from some state S to T is never Zero(). In particular, arc weights
-// are never Zero().
-//
-// Complexity:
-//
-// Depends on properties of the semiring and the queue discipline.
-//
-// For more information, see:
-//
-// Mohri, M. 2002. Semiring framework and algorithms for
-// shortest-distance problems, Journal of Automata, Languages and
-// Combinatorics 7(3): 321-350, 2002.
-template <class Arc>
-void ShortestDistance(const Fst<Arc> &fst,
-                      std::vector<typename Arc::Weight> *distance,
-                      bool reverse = false, float delta = kDelta) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  if (!reverse) {
-    AnyArcFilter<Arc> arc_filter;
-    AutoQueue<StateId> state_queue(fst, distance, arc_filter);
-    const ShortestDistanceOptions<Arc, AutoQueue<StateId>, AnyArcFilter<Arc>>
-        opts(&state_queue, arc_filter, kNoStateId, delta);
-    ShortestDistance(fst, distance, opts);
-  } else {
-    using ReverseArc = ReverseArc<Arc>;
-    using ReverseWeight = typename ReverseArc::Weight;
-    AnyArcFilter<ReverseArc> rarc_filter;
-    VectorFst<ReverseArc> rfst;
-    Reverse(fst, &rfst);
-    std::vector<ReverseWeight> rdistance;
-    AutoQueue<StateId> state_queue(rfst, &rdistance, rarc_filter);
-    const ShortestDistanceOptions<ReverseArc, AutoQueue<StateId>,
-                                  AnyArcFilter<ReverseArc>>
-        ropts(&state_queue, rarc_filter, kNoStateId, delta);
-    ShortestDistance(rfst, &rdistance, ropts);
-    distance->clear();
-    if (rdistance.size() == 1 && !rdistance[0].Member()) {
-      distance->resize(1, Arc::Weight::NoWeight());
-      return;
-    }
-    while (distance->size() < rdistance.size() - 1) {
-      distance->push_back(rdistance[distance->size() + 1].Reverse());
-    }
-  }
-}
-
-// Return the sum of the weight of all successful paths in an FST, i.e., the
-// shortest-distance from the initial state to the final states. Returns a
-// weight such that Member() is false if an error was encountered.
-template <class Arc>
-typename Arc::Weight ShortestDistance(const Fst<Arc> &fst,
-                                      float delta = kDelta) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  std::vector<Weight> distance;
-  if (Weight::Properties() & kRightSemiring) {
-    ShortestDistance(fst, &distance, false, delta);
-    if (distance.size() == 1 && !distance[0].Member()) {
-      return Arc::Weight::NoWeight();
-    }
-    Adder<Weight> adder;  // maintains cumulative sum accurately
-    for (StateId state = 0; state < distance.size(); ++state) {
-      adder.Add(Times(distance[state], fst.Final(state)));
-    }
-    return adder.Sum();
-  } else {
-    ShortestDistance(fst, &distance, true, delta);
-    const auto state = fst.Start();
-    if (distance.size() == 1 && !distance[0].Member()) {
-      return Arc::Weight::NoWeight();
-    }
-    return state != kNoStateId && state < distance.size() ? distance[state]
-                                                          : Weight::Zero();
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_SHORTEST_DISTANCE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/shortest-path.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/shortest-path.h
deleted file mode 100644
index 624f86b1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/shortest-path.h
+++ /dev/null
@@ -1,515 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions to find shortest paths in an FST.
-
-#ifndef FST_SHORTEST_PATH_H_
-#define FST_SHORTEST_PATH_H_
-
-#include <functional>
-#include <type_traits>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/cache.h>
-#include <fst/determinize.h>
-#include <fst/queue.h>
-#include <fst/shortest-distance.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-template <class Arc, class Queue, class ArcFilter>
-struct ShortestPathOptions
-    : public ShortestDistanceOptions<Arc, Queue, ArcFilter> {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  int32 nshortest;    // Returns n-shortest paths.
-  bool unique;        // Only returns paths with distinct input strings.
-  bool has_distance;  // Distance vector already contains the
-                      // shortest distance from the initial state.
-  bool first_path;    // Single shortest path stops after finding the first
-                      // path to a final state; that path is the shortest path
-                      // only when using the ShortestFirstQueue and
-                      // only when all the weights in the FST are between
-                      // One() and Zero() according to NaturalLess.
-  Weight weight_threshold;  // Pruning weight threshold.
-  StateId state_threshold;  // Pruning state threshold.
-
-  ShortestPathOptions(Queue *queue, ArcFilter filter, int32 nshortest = 1,
-                      bool unique = false, bool has_distance = false,
-                      float delta = kDelta, bool first_path = false,
-                      Weight weight_threshold = Weight::Zero(),
-                      StateId state_threshold = kNoStateId)
-      : ShortestDistanceOptions<Arc, Queue, ArcFilter>(queue, filter,
-                                                       kNoStateId, delta),
-        nshortest(nshortest),
-        unique(unique),
-        has_distance(has_distance),
-        first_path(first_path),
-        weight_threshold(std::move(weight_threshold)),
-        state_threshold(state_threshold) {}
-};
-
-namespace internal {
-
-constexpr size_t kNoArc = -1;
-
-// Helper function for SingleShortestPath building the shortest path as a left-
-// to-right machine backwards from the best final state. It takes the input
-// FST passed to SingleShortestPath and the parent vector and f_parent returned
-// by that function, and builds the result into the provided output mutable FS
-// This is not normally called by users; see ShortestPath instead.
-template <class Arc>
-void SingleShortestPathBacktrace(
-    const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-    const std::vector<std::pair<typename Arc::StateId, size_t>> &parent,
-    typename Arc::StateId f_parent) {
-  using StateId = typename Arc::StateId;
-  ofst->DeleteStates();
-  ofst->SetInputSymbols(ifst.InputSymbols());
-  ofst->SetOutputSymbols(ifst.OutputSymbols());
-  StateId s_p = kNoStateId;
-  StateId d_p = kNoStateId;
-  for (StateId state = f_parent, d = kNoStateId; state != kNoStateId;
-       d = state, state = parent[state].first) {
-    d_p = s_p;
-    s_p = ofst->AddState();
-    if (d == kNoStateId) {
-      ofst->SetFinal(s_p, ifst.Final(f_parent));
-    } else {
-      ArcIterator<Fst<Arc>> aiter(ifst, state);
-      aiter.Seek(parent[d].second);
-      auto arc = aiter.Value();
-      arc.nextstate = d_p;
-      ofst->AddArc(s_p, arc);
-    }
-  }
-  ofst->SetStart(s_p);
-  if (ifst.Properties(kError, false)) ofst->SetProperties(kError, kError);
-  ofst->SetProperties(
-      ShortestPathProperties(ofst->Properties(kFstProperties, false), true),
-      kFstProperties);
-}
-
-// Helper function for SingleShortestPath building a tree of shortest paths to
-// every final state in the input FST. It takes the input FST and parent values
-// computed by SingleShortestPath and builds into the output mutable FST the
-// subtree of ifst that consists only of the best paths to all final states.
-// This is not normally called by users; see ShortestPath instead.
-template <class Arc>
-void SingleShortestTree(
-    const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-    const std::vector<std::pair<typename Arc::StateId, size_t>> &parent) {
-  ofst->DeleteStates();
-  ofst->SetInputSymbols(ifst.InputSymbols());
-  ofst->SetOutputSymbols(ifst.OutputSymbols());
-  ofst->SetStart(ifst.Start());
-  for (StateIterator<Fst<Arc>> siter(ifst); !siter.Done(); siter.Next()) {
-    ofst->AddState();
-    ofst->SetFinal(siter.Value(), ifst.Final(siter.Value()));
-  }
-  for (const auto &pair : parent) {
-    if (pair.first != kNoStateId && pair.second != kNoArc) {
-      ArcIterator<Fst<Arc>> aiter(ifst, pair.first);
-      aiter.Seek(pair.second);
-      ofst->AddArc(pair.first, aiter.Value());
-    }
-  }
-  if (ifst.Properties(kError, false)) ofst->SetProperties(kError, kError);
-  ofst->SetProperties(
-      ShortestPathProperties(ofst->Properties(kFstProperties, false), true),
-      kFstProperties);
-}
-
-// Shortest-path algorithm. It builds the output mutable FST so that it contains
-// the shortest path in the input FST; distance returns the shortest distances
-// from the source state to each state in the input FST, and the options struct
-// is
-// used to specify options such as the queue discipline, the arc filter and
-// delta. The super_final option is an output parameter indicating the final
-// state, and the parent argument is used for the storage of the backtrace path
-// for each state 1 to n, (i.e., the best previous state and the arc that
-// transition to state n.) The shortest path is the lowest weight path w.r.t.
-// the natural semiring order. The weights need to be right distributive and
-// have the path (kPath) property. False is returned if an error is encountered.
-//
-// This is not normally called by users; see ShortestPath instead (with n = 1).
-template <class Arc, class Queue, class ArcFilter>
-bool SingleShortestPath(
-    const Fst<Arc> &ifst, std::vector<typename Arc::Weight> *distance,
-    const ShortestPathOptions<Arc, Queue, ArcFilter> &opts,
-    typename Arc::StateId *f_parent,
-    std::vector<std::pair<typename Arc::StateId, size_t>> *parent) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  
-  if ((Weight::Properties() & (kPath | kRightSemiring)) !=
-    (kPath | kRightSemiring)) {
-    FSTERROR() << "SingleShortestPath: Weight needs to have the path"
-      << " property and be right distributive: " << Weight::Type();
-    return false;
-  }
-
-  parent->clear();
-  *f_parent = kNoStateId;
-  if (ifst.Start() == kNoStateId) return true;
-  std::vector<bool> enqueued;
-  auto state_queue = opts.state_queue;
-  const auto source = (opts.source == kNoStateId) ? ifst.Start() : opts.source;
-  bool final_seen = false;
-  auto f_distance = Weight::Zero();
-  distance->clear();
-  state_queue->Clear();
-  while (distance->size() < source) {
-    distance->push_back(Weight::Zero());
-    enqueued.push_back(false);
-    parent->push_back(std::make_pair(kNoStateId, kNoArc));
-  }
-  distance->push_back(Weight::One());
-  parent->push_back(std::make_pair(kNoStateId, kNoArc));
-  state_queue->Enqueue(source);
-  enqueued.push_back(true);
-  while (!state_queue->Empty()) {
-    const auto s = state_queue->Head();
-    state_queue->Dequeue();
-    enqueued[s] = false;
-    const auto sd = (*distance)[s];
-    // If we are using a shortest queue, no other path is going to be shorter
-    // than f_distance at this point.
-    if (opts.first_path && final_seen && f_distance == Plus(f_distance, sd)) {
-      break;
-    }
-    if (ifst.Final(s) != Weight::Zero()) {
-      const auto plus = Plus(f_distance, Times(sd, ifst.Final(s)));
-      if (f_distance != plus) {
-        f_distance = plus;
-        *f_parent = s;
-      }
-      if (!f_distance.Member()) return false;
-      final_seen = true;
-    }
-    for (ArcIterator<Fst<Arc>> aiter(ifst, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      while (distance->size() <= arc.nextstate) {
-        distance->push_back(Weight::Zero());
-        enqueued.push_back(false);
-        parent->push_back(std::make_pair(kNoStateId, kNoArc));
-      }
-      auto &nd = (*distance)[arc.nextstate];
-      const auto weight = Times(sd, arc.weight);
-      if (nd != Plus(nd, weight)) {
-        nd = Plus(nd, weight);
-        if (!nd.Member()) return false;
-        (*parent)[arc.nextstate] = std::make_pair(s, aiter.Position());
-        if (!enqueued[arc.nextstate]) {
-          state_queue->Enqueue(arc.nextstate);
-          enqueued[arc.nextstate] = true;
-        } else {
-          state_queue->Update(arc.nextstate);
-        }
-      }
-    }
-  }
-  return true;
-}
-
-template <class StateId, class Weight>
-class ShortestPathCompare {
- public:
-  ShortestPathCompare(const std::vector<std::pair<StateId, Weight>> &pairs,
-                      const std::vector<Weight> &distance, StateId superfinal,
-                      float delta)
-      : pairs_(pairs),
-        distance_(distance),
-        superfinal_(superfinal),
-        delta_(delta) {}
-
-  bool operator()(const StateId x, const StateId y) const {
-    const auto &px = pairs_[x];
-    const auto &py = pairs_[y];
-    const auto wx = Times(PWeight(px.first), px.second);
-    const auto wy = Times(PWeight(py.first), py.second);
-    // Penalize complete paths to ensure correct results with inexact weights.
-    // This forms a strict weak order so long as ApproxEqual(a, b) =>
-    // ApproxEqual(a, c) for all c s.t. less_(a, c) && less_(c, b).
-    if (px.first == superfinal_ && py.first != superfinal_) {
-      return less_(wy, wx) || ApproxEqual(wx, wy, delta_);
-    } else if (py.first == superfinal_ && px.first != superfinal_) {
-      return less_(wy, wx) && !ApproxEqual(wx, wy, delta_);
-    } else {
-      return less_(wy, wx);
-    }
-  }
-
- private:
-  Weight PWeight(StateId state) const {
-    return (state == superfinal_)
-               ? Weight::One()
-               : (state < distance_.size()) ? distance_[state] : Weight::Zero();
-  }
-
-  const std::vector<std::pair<StateId, Weight>> &pairs_;
-  const std::vector<Weight> &distance_;
-  const StateId superfinal_;
-  const float delta_;
-  NaturalLess<Weight> less_;
-};
-
-// N-Shortest-path algorithm: implements the core n-shortest path algorithm.
-// The output is built reversed. See below for versions with more options and
-// *not reversed*.
-//
-// The output mutable FST contains the REVERSE of n'shortest paths in the input
-// FST; distance must contain the shortest distance from each state to a final
-// state in the input FST; delta is the convergence delta.
-//
-// The n-shortest paths are the n-lowest weight paths w.r.t. the natural
-// semiring order. The single path that can be read from the ith of at most n
-// transitions leaving the initial state of the the input FST is the ith
-// shortest path. Disregarding the initial state and initial transitions, the
-// n-shortest paths, in fact, form a tree rooted at the single final state.
-//
-// The weights need to be left and right distributive (kSemiring) and have the
-// path (kPath) property.
-//
-// Arc weights must satisfy the property that the sum of the weights of one or
-// more paths from some state S to T is never Zero(). In particular, arc weights
-// are never Zero().
-//
-// For more information, see:
-//
-// Mohri, M, and Riley, M. 2002. An efficient algorithm for the n-best-strings
-// problem. In Proc. ICSLP.
-//
-// The algorithm relies on the shortest-distance algorithm. There are some
-// issues with the pseudo-code as written in the paper (viz., line 11).
-//
-// IMPLEMENTATION NOTE: The input FST can be a delayed FST and and at any state
-// in its expansion the values of distance vector need only be defined at that
-// time for the states that are known to exist.
-template <class Arc, class RevArc>
-void NShortestPath(const Fst<RevArc> &ifst, MutableFst<Arc> *ofst,
-                   const std::vector<typename Arc::Weight> &distance,
-                   int32 nshortest, float delta = kDelta,
-                   typename Arc::Weight weight_threshold = Arc::Weight::Zero(),
-                   typename Arc::StateId state_threshold = kNoStateId) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using Pair = std::pair<StateId, Weight>;
-
-  if (nshortest <= 0) return;
-  ofst->DeleteStates();
-  ofst->SetInputSymbols(ifst.InputSymbols());
-  ofst->SetOutputSymbols(ifst.OutputSymbols());
-
-  if ((Weight::Properties() & (kPath | kRightSemiring)) !=
-    (kPath | kRightSemiring)) {
-    FSTERROR() << "NShortestPath: Weight needs to have the path"
-      << " property and be right distributive: " << Weight::Type();
-    ofst->SetProperties(kError, kError);
-    return;
-  }
-  // Each state in ofst corresponds to a path with weight w from the initial
-  // state of ifst to a state s in ifst, that can be characterized by a pair
-  // (s, w). The vector pairs maps each state in ofst to the corresponding
-  // pair maps states in ofst to the corresponding pair (s, w).
-  std::vector<Pair> pairs;
-  // The supefinal state is denoted by kNoStateId. The distance from the
-  // superfinal state to the final state is semiring One, so
-  // `distance[kNoStateId]` is not needed.
-  const ShortestPathCompare<StateId, Weight> compare(pairs, distance,
-                                                     kNoStateId, delta);
-  const NaturalLess<Weight> less;
-  if (ifst.Start() == kNoStateId || distance.size() <= ifst.Start() ||
-      distance[ifst.Start()] == Weight::Zero() ||
-      less(weight_threshold, Weight::One()) || state_threshold == 0) {
-    if (ifst.Properties(kError, false)) ofst->SetProperties(kError, kError);
-    return;
-  }
-  ofst->SetStart(ofst->AddState());
-  const auto final_state = ofst->AddState();
-  ofst->SetFinal(final_state, Weight::One());
-  while (pairs.size() <= final_state) {
-    pairs.push_back(std::make_pair(kNoStateId, Weight::Zero()));
-  }
-  pairs[final_state] = std::make_pair(ifst.Start(), Weight::One());
-  std::vector<StateId> heap;
-  heap.push_back(final_state);
-  const auto limit = Times(distance[ifst.Start()], weight_threshold);
-  // r[s + 1], s state in fst, is the number of states in ofst which
-  // corresponding pair contains s, i.e., it is number of paths computed so far
-  // to s. Valid for s == kNoStateId (the superfinal state).
-  std::vector<int> r;
-  while (!heap.empty()) {
-    std::pop_heap(heap.begin(), heap.end(), compare);
-    const auto state = heap.back();
-    const auto p = pairs[state];
-    heap.pop_back();
-    const auto d =
-        (p.first == kNoStateId)
-            ? Weight::One()
-            : (p.first < distance.size()) ? distance[p.first] : Weight::Zero();
-    if (less(limit, Times(d, p.second)) ||
-        (state_threshold != kNoStateId &&
-         ofst->NumStates() >= state_threshold)) {
-      continue;
-    }
-    while (r.size() <= p.first + 1) r.push_back(0);
-    ++r[p.first + 1];
-    if (p.first == kNoStateId) {
-      ofst->AddArc(ofst->Start(), Arc(0, 0, Weight::One(), state));
-    }
-    if ((p.first == kNoStateId) && (r[p.first + 1] == nshortest)) break;
-    if (r[p.first + 1] > nshortest) continue;
-    if (p.first == kNoStateId) continue;
-    for (ArcIterator<Fst<RevArc>> aiter(ifst, p.first); !aiter.Done();
-         aiter.Next()) {
-      const auto &rarc = aiter.Value();
-      Arc arc(rarc.ilabel, rarc.olabel, rarc.weight.Reverse(), rarc.nextstate);
-      const auto weight = Times(p.second, arc.weight);
-      const auto next = ofst->AddState();
-      pairs.push_back(std::make_pair(arc.nextstate, weight));
-      arc.nextstate = state;
-      ofst->AddArc(next, arc);
-      heap.push_back(next);
-      std::push_heap(heap.begin(), heap.end(), compare);
-    }
-    const auto final_weight = ifst.Final(p.first).Reverse();
-    if (final_weight != Weight::Zero()) {
-      const auto weight = Times(p.second, final_weight);
-      const auto next = ofst->AddState();
-      pairs.push_back(std::make_pair(kNoStateId, weight));
-      ofst->AddArc(next, Arc(0, 0, final_weight, state));
-      heap.push_back(next);
-      std::push_heap(heap.begin(), heap.end(), compare);
-    }
-  }
-  Connect(ofst);
-  if (ifst.Properties(kError, false)) ofst->SetProperties(kError, kError);
-  ofst->SetProperties(
-      ShortestPathProperties(ofst->Properties(kFstProperties, false)),
-      kFstProperties);
-}
-
-}  // namespace internal
-
-// N-Shortest-path algorithm: this version allows finer control via the options
-// argument. See below for a simpler interface. The output mutable FST contains
-// the n-shortest paths in the input FST; the distance argument is used to
-// return the shortest distances from the source state to each state in the
-// input FST, and the options struct is used to specify the number of paths to
-// return, whether they need to have distinct input strings, the queue
-// discipline, the arc filter and the convergence delta.
-//
-// The n-shortest paths are the n-lowest weight paths w.r.t. the natural
-// semiring order. The single path that can be read from the ith of at most n
-// transitions leaving the initial state of the output FST is the ith shortest
-// path.
-// Disregarding the initial state and initial transitions, The n-shortest paths,
-// in fact, form a tree rooted at the single final state.
-//
-// The weights need to be right distributive and have the path (kPath) property.
-// They need to be left distributive as well for nshortest > 1.
-//
-// For more information, see:
-//
-// Mohri, M, and Riley, M. 2002. An efficient algorithm for the n-best-strings
-// problem. In Proc. ICSLP.
-//
-// The algorithm relies on the shortest-distance algorithm. There are some
-// issues with the pseudo-code as written in the paper (viz., line 11).
-template <class Arc, class Queue, class ArcFilter>
-void ShortestPath(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-                  std::vector<typename Arc::Weight> *distance,
-                  const ShortestPathOptions<Arc, Queue, ArcFilter> &opts) {
- if ((Arc::Weight::Properties() & (kPath | kSemiring)) == (kPath | kSemiring)) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using RevArc = ReverseArc<Arc>;
-  if (opts.nshortest == 1) {
-    std::vector<std::pair<StateId, size_t>> parent;
-    StateId f_parent;
-    if (internal::SingleShortestPath(ifst, distance, opts, &f_parent,
-                                     &parent)) {
-      internal::SingleShortestPathBacktrace(ifst, ofst, parent, f_parent);
-    } else {
-      ofst->SetProperties(kError, kError);
-    }
-    return;
-  }
-  if (opts.nshortest <= 0) return;
-  if (!opts.has_distance) {
-    ShortestDistance(ifst, distance, opts);
-    if (distance->size() == 1 && !(*distance)[0].Member()) {
-      ofst->SetProperties(kError, kError);
-      return;
-    }
-  }
-  // Algorithm works on the reverse of 'fst'; 'distance' is the distance to the
-  // final state in 'rfst', 'ofst' is built as the reverse of the tree of
-  // n-shortest path in 'rfst'.
-  VectorFst<RevArc> rfst;
-  Reverse(ifst, &rfst);
-  auto d = Weight::Zero();
-  for (ArcIterator<VectorFst<RevArc>> aiter(rfst, 0); !aiter.Done();
-       aiter.Next()) {
-    const auto &arc = aiter.Value();
-    const auto state = arc.nextstate - 1;
-    if (state < distance->size()) {
-      d = Plus(d, Times(arc.weight.Reverse(), (*distance)[state]));
-    }
-  }
-  // TODO(kbg): Avoid this expensive vector operation.
-  distance->insert(distance->begin(), d);
-  if (!opts.unique) {
-    internal::NShortestPath(rfst, ofst, *distance, opts.nshortest, opts.delta,
-                            opts.weight_threshold, opts.state_threshold);
-  } else {
-    std::vector<Weight> ddistance;
-    DeterminizeFstOptions<RevArc> dopts(opts.delta);
-    DeterminizeFst<RevArc> dfst(rfst, distance, &ddistance, dopts);
-    internal::NShortestPath(dfst, ofst, ddistance, opts.nshortest, opts.delta,
-                            opts.weight_threshold, opts.state_threshold);
-  }
-  // TODO(kbg): Avoid this expensive vector operation.
-  distance->erase(distance->begin());
- } else {  // (Arc::Weight::Properties() & (kPath | kSemiring)) != (kPath | kSemiring)
-  FSTERROR() << "ShortestPath: Weight needs to have the "
-             << "path property and be distributive: " << Arc::Weight::Type();
-  ofst->SetProperties(kError, kError);
- }
-}
-
-// Shortest-path algorithm: simplified interface. See above for a version that
-// allows finer control. The output mutable FST contains the n-shortest paths
-// in the input FST. The queue discipline is automatically selected. When unique
-// is true, only paths with distinct input label sequences are returned.
-//
-// The n-shortest paths are the n-lowest weight paths w.r.t. the natural
-// semiring order. The single path that can be read from the ith of at most n
-// transitions leaving the initial state of the ouput FST is the ith best path.
-// The weights need to be right distributive and have the path (kPath) property.
-template <class Arc>
-void ShortestPath(const Fst<Arc> &ifst, MutableFst<Arc> *ofst,
-                  int32 nshortest = 1, bool unique = false,
-                  bool first_path = false,
-                  typename Arc::Weight weight_threshold = Arc::Weight::Zero(),
-                  typename Arc::StateId state_threshold = kNoStateId) {
-  using StateId = typename Arc::StateId;
-  std::vector<typename Arc::Weight> distance;
-  AnyArcFilter<Arc> arc_filter;
-  AutoQueue<StateId> state_queue(ifst, &distance, arc_filter);
-  const ShortestPathOptions<Arc, AutoQueue<StateId>, AnyArcFilter<Arc>> opts(
-      &state_queue, arc_filter, nshortest, unique, false, kDelta, first_path,
-      weight_threshold, state_threshold);
-  ShortestPath(ifst, ofst, &distance, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_SHORTEST_PATH_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/signed-log-weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/signed-log-weight.h
deleted file mode 100644
index 63bd7852..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/signed-log-weight.h
+++ /dev/null
@@ -1,436 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// LogWeight along with sign information that represents the value X in the
-// linear domain as <sign(X), -ln(|X|)>
-//
-// The sign is a TropicalWeight:
-//  positive, TropicalWeight.Value() > 0.0, recommended value 1.0
-//  negative, TropicalWeight.Value() <= 0.0, recommended value -1.0
-
-#ifndef FST_SIGNED_LOG_WEIGHT_H_
-#define FST_SIGNED_LOG_WEIGHT_H_
-
-#include <cstdlib>
-
-#include <fst/float-weight.h>
-#include <fst/pair-weight.h>
-#include <fst/product-weight.h>
-
-
-namespace fst {
-template <class T>
-class SignedLogWeightTpl : public PairWeight<TropicalWeight, LogWeightTpl<T>> {
- public:
-  using X1 = TropicalWeight;
-  using X2 = LogWeightTpl<T>;
-  using ReverseWeight = SignedLogWeightTpl;
-
-  using PairWeight<X1, X2>::Value1;
-  using PairWeight<X1, X2>::Value2;
-
-  SignedLogWeightTpl() : PairWeight<X1, X2>() {}
-
-  SignedLogWeightTpl(const SignedLogWeightTpl &w) : PairWeight<X1, X2>(w) {}
-
-  explicit SignedLogWeightTpl(const PairWeight<X1, X2> &w)
-      : PairWeight<X1, X2>(w) {}
-
-  SignedLogWeightTpl(const X1 &x1, const X2 &x2) : PairWeight<X1, X2>(x1, x2) {}
-
-  static const SignedLogWeightTpl &Zero() {
-    static const SignedLogWeightTpl zero(X1(1.0), X2::Zero());
-    return zero;
-  }
-
-  static const SignedLogWeightTpl &One() {
-    static const SignedLogWeightTpl one(X1(1.0), X2::One());
-    return one;
-  }
-
-  static const SignedLogWeightTpl &NoWeight() {
-    static const SignedLogWeightTpl no_weight(X1(1.0), X2::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type =
-        new string("signed_log_" + X1::Type() + "_" + X2::Type());
-    return *type;
-  }
-
-  SignedLogWeightTpl Quantize(float delta = kDelta) const {
-    return SignedLogWeightTpl(PairWeight<X1, X2>::Quantize(delta));
-  }
-
-  ReverseWeight Reverse() const {
-    return SignedLogWeightTpl(PairWeight<X1, X2>::Reverse());
-  }
-
-  bool Member() const { return PairWeight<X1, X2>::Member(); }
-
-  // Neither idempotent nor path.
-  static constexpr uint64 Properties() {
-    return kLeftSemiring | kRightSemiring | kCommutative;
-  }
-
-  size_t Hash() const {
-    size_t h1;
-    if (Value2() == X2::Zero() || Value1().Value() > 0.0) {
-      h1 = TropicalWeight(1.0).Hash();
-    } else {
-      h1 = TropicalWeight(-1.0).Hash();
-    }
-    size_t h2 = Value2().Hash();
-    static constexpr int lshift = 5;
-    static constexpr int rshift = CHAR_BIT * sizeof(size_t) - 5;
-    return h1 << lshift ^ h1 >> rshift ^ h2;
-  }
-};
-
-template <class T>
-inline SignedLogWeightTpl<T> Plus(const SignedLogWeightTpl<T> &w1,
-                                  const SignedLogWeightTpl<T> &w2) {
-  using X1 = TropicalWeight;
-  using X2 = LogWeightTpl<T>;
-  if (!w1.Member() || !w2.Member()) return SignedLogWeightTpl<T>::NoWeight();
-  const auto s1 = w1.Value1().Value() > 0.0;
-  const auto s2 = w2.Value1().Value() > 0.0;
-  const bool equal = (s1 == s2);
-  const auto f1 = w1.Value2().Value();
-  const auto f2 = w2.Value2().Value();
-  if (f1 == FloatLimits<T>::PosInfinity()) {
-    return w2;
-  } else if (f2 == FloatLimits<T>::PosInfinity()) {
-    return w1;
-  } else if (f1 == f2) {
-    if (equal) {
-      return SignedLogWeightTpl<T>(X1(w1.Value1()), X2(f2 - log(2.0F)));
-    } else {
-      return SignedLogWeightTpl<T>::Zero();
-    }
-  } else if (f1 > f2) {
-    if (equal) {
-      return SignedLogWeightTpl<T>(X1(w1.Value1()),
-                                   X2(f2 - log(1.0F + exp(f2 - f1))));
-    } else {
-      return SignedLogWeightTpl<T>(X1(w2.Value1()),
-                                   X2((f2 - log(1.0F - exp(f2 - f1)))));
-    }
-  } else {
-    if (equal) {
-      return SignedLogWeightTpl<T>(X1(w2.Value1()),
-                                   X2((f1 - log(1.0F + exp(f1 - f2)))));
-    } else {
-      return SignedLogWeightTpl<T>(X1(w1.Value1()),
-                                   X2((f1 - log(1.0F - exp(f1 - f2)))));
-    }
-  }
-}
-
-template <class T>
-inline SignedLogWeightTpl<T> Minus(const SignedLogWeightTpl<T> &w1,
-                                   const SignedLogWeightTpl<T> &w2) {
-  SignedLogWeightTpl<T> minus_w2(-w2.Value1().Value(), w2.Value2());
-  return Plus(w1, minus_w2);
-}
-
-template <class T>
-inline SignedLogWeightTpl<T> Times(const SignedLogWeightTpl<T> &w1,
-                                   const SignedLogWeightTpl<T> &w2) {
-  using X2 = LogWeightTpl<T>;
-  if (!w1.Member() || !w2.Member()) return SignedLogWeightTpl<T>::NoWeight();
-  const auto s1 = w1.Value1().Value() > 0.0;
-  const auto s2 = w2.Value1().Value() > 0.0;
-  const auto f1 = w1.Value2().Value();
-  const auto f2 = w2.Value2().Value();
-  if (s1 == s2) {
-    return SignedLogWeightTpl<T>(TropicalWeight(1.0), X2(f1 + f2));
-  } else {
-    return SignedLogWeightTpl<T>(TropicalWeight(-1.0), X2(f1 + f2));
-  }
-}
-
-template <class T>
-inline SignedLogWeightTpl<T> Divide(const SignedLogWeightTpl<T> &w1,
-                                    const SignedLogWeightTpl<T> &w2,
-                                    DivideType typ = DIVIDE_ANY) {
-  using X2 = LogWeightTpl<T>;
-  if (!w1.Member() || !w2.Member()) return SignedLogWeightTpl<T>::NoWeight();
-  const auto s1 = w1.Value1().Value() > 0.0;
-  const auto s2 = w2.Value1().Value() > 0.0;
-  const auto f1 = w1.Value2().Value();
-  const auto f2 = w2.Value2().Value();
-  if (f2 == FloatLimits<T>::PosInfinity()) {
-    return SignedLogWeightTpl<T>(TropicalWeight(1.0),
-                                 X2(FloatLimits<T>::NumberBad()));
-  } else if (f1 == FloatLimits<T>::PosInfinity()) {
-    return SignedLogWeightTpl<T>(TropicalWeight(1.0),
-                                 X2(FloatLimits<T>::PosInfinity()));
-  } else if (s1 == s2) {
-    return SignedLogWeightTpl<T>(TropicalWeight(1.0), X2(f1 - f2));
-  } else {
-    return SignedLogWeightTpl<T>(TropicalWeight(-1.0), X2(f1 - f2));
-  }
-}
-
-template <class T>
-inline bool ApproxEqual(const SignedLogWeightTpl<T> &w1,
-                        const SignedLogWeightTpl<T> &w2, float delta = kDelta) {
-  const auto s1 = w1.Value1().Value() > 0.0;
-  const auto s2 = w2.Value1().Value() > 0.0;
-  if (s1 == s2) {
-    return ApproxEqual(w1.Value2(), w2.Value2(), delta);
-  } else {
-    return w1.Value2() == LogWeightTpl<T>::Zero() &&
-           w2.Value2() == LogWeightTpl<T>::Zero();
-  }
-}
-
-template <class T>
-inline bool operator==(const SignedLogWeightTpl<T> &w1,
-                       const SignedLogWeightTpl<T> &w2) {
-  const auto s1 = w1.Value1().Value() > 0.0;
-  const auto s2 = w2.Value1().Value() > 0.0;
-  if (s1 == s2) {
-    return w1.Value2() == w2.Value2();
-  } else {
-    return (w1.Value2() == LogWeightTpl<T>::Zero()) &&
-           (w2.Value2() == LogWeightTpl<T>::Zero());
-  }
-}
-
-// Single-precision signed-log weight.
-using SignedLogWeight = SignedLogWeightTpl<float>;
-
-// Double-precision signed-log weight.
-using SignedLog64Weight = SignedLogWeightTpl<double>;
-
-template <class W1, class W2>
-bool SignedLogConvertCheck(W1 weight) {
-  if (weight.Value1().Value() < 0.0) {
-    FSTERROR() << "WeightConvert: Can't convert weight " << weight
-               << " from " << W1::Type() << " to " << W2::Type();
-    return false;
-  }
-  return true;
-}
-
-// Specialization using the Kahan compensated summation
-template <class T>
-class Adder<SignedLogWeightTpl<T>> {
- public:
-  using Weight = SignedLogWeightTpl<T>;
-  using X1 = TropicalWeight;
-  using X2 = LogWeightTpl<T>;
-
-  explicit Adder(Weight w = Weight::Zero())
-     : ssum_(w.Value1().Value() > 0.0),
-        sum_(w.Value2().Value()),
-        c_(0.0) { }
-
-  Weight Add(const Weight &w) {
-    const auto sw = w.Value1().Value() > 0.0;
-    const auto f = w.Value2().Value();
-    const bool equal = (ssum_ == sw);
-
-    if (!Sum().Member() || f == FloatLimits<T>::PosInfinity()) {
-      return Sum();
-    } else if (!w.Member() || sum_ == FloatLimits<T>::PosInfinity()) {
-      sum_ = f;
-      ssum_ = sw;
-      c_ = 0.0;
-    } else if (f == sum_) {
-      if (equal) {
-        sum_ = internal::KahanLogSum(sum_, f, &c_);
-      } else {
-        sum_ = FloatLimits<T>::PosInfinity();
-        ssum_ = true;
-        c_ = 0.0;
-      }
-    } else if (f > sum_) {
-      if (equal) {
-        sum_ = internal::KahanLogSum(sum_, f, &c_);
-      } else {
-        sum_ = internal::KahanLogDiff(sum_, f, &c_);
-      }
-    } else {
-      if (equal) {
-        sum_ = internal::KahanLogSum(f, sum_, &c_);
-      } else {
-        sum_ = internal::KahanLogDiff(f, sum_, &c_);
-        ssum_ = sw;
-      }
-    }
-    return Sum();
-  }
-
-  Weight Sum() { return Weight(X1(ssum_ ? 1.0 : -1.0), X2(sum_)); }
-
-  void Reset(Weight w = Weight::Zero()) {
-    ssum_ = w.Value1().Value() > 0.0;
-    sum_ = w.Value2().Value();
-    c_ = 0.0;
-  }
-
- private:
-  bool ssum_;   // true iff sign of sum is positive
-  double sum_;  // unsigned sum
-  double c_;    // Kahan compensation
-};
-
-// Converts to tropical.
-template <>
-struct WeightConvert<SignedLogWeight, TropicalWeight> {
-  TropicalWeight operator()(const SignedLogWeight &weight) const {
-    if (!SignedLogConvertCheck<SignedLogWeight, TropicalWeight>(weight)) {
-      return TropicalWeight::NoWeight();
-    }
-    return TropicalWeight(weight.Value2().Value());
-  }
-};
-
-template <>
-struct WeightConvert<SignedLog64Weight, TropicalWeight> {
-  TropicalWeight operator()(const SignedLog64Weight &weight) const {
-    if (!SignedLogConvertCheck<SignedLog64Weight, TropicalWeight>(weight)) {
-      return TropicalWeight::NoWeight();
-    }
-    return TropicalWeight(weight.Value2().Value());
-  }
-};
-
-// Converts to log.
-template <>
-struct WeightConvert<SignedLogWeight, LogWeight> {
-  LogWeight operator()(const SignedLogWeight &weight) const {
-    if (!SignedLogConvertCheck<SignedLogWeight, LogWeight>(weight)) {
-      return LogWeight::NoWeight();
-    }
-    return LogWeight(weight.Value2().Value());
-  }
-};
-
-template <>
-struct WeightConvert<SignedLog64Weight, LogWeight> {
-  LogWeight operator()(const SignedLog64Weight &weight) const {
-    if (!SignedLogConvertCheck<SignedLog64Weight, LogWeight>(weight)) {
-      return LogWeight::NoWeight();
-    }
-    return LogWeight(weight.Value2().Value());
-  }
-};
-
-// Converts to log64.
-template <>
-struct WeightConvert<SignedLogWeight, Log64Weight> {
-  Log64Weight operator()(const SignedLogWeight &weight) const {
-    if (!SignedLogConvertCheck<SignedLogWeight, Log64Weight>(weight)) {
-      return Log64Weight::NoWeight();
-    }
-    return Log64Weight(weight.Value2().Value());
-  }
-};
-
-template <>
-struct WeightConvert<SignedLog64Weight, Log64Weight> {
-  Log64Weight operator()(const SignedLog64Weight &weight) const {
-    if (!SignedLogConvertCheck<SignedLog64Weight, Log64Weight>(weight)) {
-      return Log64Weight::NoWeight();
-    }
-    return Log64Weight(weight.Value2().Value());
-  }
-};
-
-// Converts to signed log.
-template <>
-struct WeightConvert<TropicalWeight, SignedLogWeight> {
-  SignedLogWeight operator()(const TropicalWeight &weight) const {
-    return SignedLogWeight(1.0, weight.Value());
-  }
-};
-
-template <>
-struct WeightConvert<LogWeight, SignedLogWeight> {
-  SignedLogWeight operator()(const LogWeight &weight) const {
-    return SignedLogWeight(1.0, weight.Value());
-  }
-};
-
-template <>
-struct WeightConvert<Log64Weight, SignedLogWeight> {
-  SignedLogWeight operator()(const Log64Weight &weight) const {
-    return SignedLogWeight(1.0, weight.Value());
-  }
-};
-
-template <>
-struct WeightConvert<SignedLog64Weight, SignedLogWeight> {
-  SignedLogWeight operator()(const SignedLog64Weight &weight) const {
-    return SignedLogWeight(weight.Value1(), weight.Value2().Value());
-  }
-};
-
-// Converts to signed log64.
-template <>
-struct WeightConvert<TropicalWeight, SignedLog64Weight> {
-  SignedLog64Weight operator()(const TropicalWeight &weight) const {
-    return SignedLog64Weight(1.0, weight.Value());
-  }
-};
-
-template <>
-struct WeightConvert<LogWeight, SignedLog64Weight> {
-  SignedLog64Weight operator()(const LogWeight &weight) const {
-    return SignedLog64Weight(1.0, weight.Value());
-  }
-};
-
-template <>
-struct WeightConvert<Log64Weight, SignedLog64Weight> {
-  SignedLog64Weight operator()(const Log64Weight &weight) const {
-    return SignedLog64Weight(1.0, weight.Value());
-  }
-};
-
-template <>
-struct WeightConvert<SignedLogWeight, SignedLog64Weight> {
-  SignedLog64Weight operator()(const SignedLogWeight &weight) const {
-    return SignedLog64Weight(weight.Value1(), weight.Value2().Value());
-  }
-};
-
-// This function object returns SignedLogWeightTpl<T>'s that are random integers
-// chosen from [0, num_random_weights) times a random sign. This is intended
-// primarily for testing.
-template <class T>
-class WeightGenerate<SignedLogWeightTpl<T>> {
- public:
-  using Weight = SignedLogWeightTpl<T>;
-  using X1 = typename Weight::X1;
-  using X2 = typename Weight::X2;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t num_random_weights = kNumRandomWeights)
-    : allow_zero_(allow_zero), num_random_weights_(num_random_weights) {}
-
-  Weight operator()() const {
-    static const X1 negative_one(-1.0);
-    static const X1 positive_one(+1.0);
-    const int m = rand() % 2;                                    // NOLINT
-    const int n = rand() % (num_random_weights_ + allow_zero_);  // NOLINT
-    return Weight((m == 0) ? negative_one : positive_one,
-                  (allow_zero_ && n == num_random_weights_) ?
-                   X2::Zero() : X2(n));
-  }
-
- private:
-  // Permits Zero() and zero divisors.
-  const bool allow_zero_;
-  // Number of alternative random weights.
-  const size_t num_random_weights_;
-};
-
-}  // namespace fst
-
-#endif  // FST_SIGNED_LOG_WEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/sparse-power-weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/sparse-power-weight.h
deleted file mode 100644
index a9911187..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/sparse-power-weight.h
+++ /dev/null
@@ -1,235 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Cartesian power weight semiring operation definitions, using
-// SparseTupleWeight as underlying representation.
-
-#ifndef FST_SPARSE_POWER_WEIGHT_H_
-#define FST_SPARSE_POWER_WEIGHT_H_
-
-#include <climits>
-#include <string>
-
-#include <fst/sparse-tuple-weight.h>
-#include <fst/weight.h>
-
-
-namespace fst {
-
-// Below SparseTupleWeight*Mapper are used in conjunction with
-// SparseTupleWeightMap to compute the respective semiring operations
-template <class W, class K>
-struct SparseTupleWeightPlusMapper {
-  W Map(const K &k, const W &v1, const W &v2) const { return Plus(v1, v2); }
-};
-
-template <class W, class K>
-struct SparseTupleWeightTimesMapper {
-  W Map(const K &k, const W &v1, const W &v2) const { return Times(v1, v2); }
-};
-
-template <class W, class K>
-struct SparseTupleWeightDivideMapper {
-  const DivideType type;
-
-  explicit SparseTupleWeightDivideMapper(DivideType type_) : type(type_) {}
-
-  W Map(const K &k, const W &v1, const W &v2) const {
-    return Divide(v1, v2, type);
-  }
-};
-
-template <class W, class K>
-struct SparseTupleWeightApproxMapper {
-  const float delta;
-
-  explicit SparseTupleWeightApproxMapper(float delta_ = kDelta)
-      : delta(delta_) {}
-
-  W Map(const K &k, const W &v1, const W &v2) const {
-    return ApproxEqual(v1, v2, delta) ? W::One() : W::Zero();
-  }
-};
-
-// Sparse cartesian power semiring: W ^ n
-//
-// Forms:
-//
-//  - a left semimodule when W is a left semiring,
-//  - a right semimodule when W is a right semiring,
-//  - a bisemimodule when W is a semiring,
-//    the free semimodule of rank n over W
-//
-// The Times operation is overloaded to provide the left and right scalar
-// products.
-//
-// K is the key value type. kNoKey (-1) is reserved for internal use
-template <class W, class K = int>
-class SparsePowerWeight : public SparseTupleWeight<W, K> {
- public:
-  using ReverseWeight = SparsePowerWeight<typename W::ReverseWeight, K>;
-
-  SparsePowerWeight() {}
-
-  explicit SparsePowerWeight(const SparseTupleWeight<W, K> &weight)
-      : SparseTupleWeight<W, K>(weight) {}
-
-  template <class Iterator>
-  SparsePowerWeight(Iterator begin, Iterator end)
-      : SparseTupleWeight<W, K>(begin, end) {}
-
-  // Initialize component `key` to `weight`, with `default_weight` for all
-  // other components.
-  SparsePowerWeight(const K &key, const W &weight,
-                    const W &default_weight = W::Zero())
-      : SparseTupleWeight<W, K>(key, weight, default_weight) {}
-
-  static const SparsePowerWeight &Zero() {
-    static const SparsePowerWeight zero(SparseTupleWeight<W, K>::Zero());
-    return zero;
-  }
-
-  static const SparsePowerWeight &One() {
-    static const SparsePowerWeight one(SparseTupleWeight<W, K>::One());
-    return one;
-  }
-
-  static const SparsePowerWeight &NoWeight() {
-    static const SparsePowerWeight no_weight(
-        SparseTupleWeight<W, K>::NoWeight());
-    return no_weight;
-  }
-
-  // Overide this: Overwrite the Type method to reflect the key type if using
-  // a non-default key type.
-  static const string &Type() {
-    static const string *const type = [] {
-      string type = W::Type() + "_^n";
-      if (sizeof(K) != sizeof(uint32)) {
-        type += "_" + std::to_string(CHAR_BIT * sizeof(K));
-      }
-      return new string(type);
-    }();
-    return *type;
-  }
-
-  static constexpr uint64 Properties() {
-    return W::Properties() &
-           (kLeftSemiring | kRightSemiring | kCommutative | kIdempotent);
-  }
-
-  SparsePowerWeight Quantize(float delta = kDelta) const {
-    return SparsePowerWeight(SparseTupleWeight<W, K>::Quantize(delta));
-  }
-
-  ReverseWeight Reverse() const {
-    return ReverseWeight(SparseTupleWeight<W, K>::Reverse());
-  }
-};
-
-// Semimodule plus operation.
-template <class W, class K>
-inline SparsePowerWeight<W, K> Plus(const SparsePowerWeight<W, K> &w1,
-                                    const SparsePowerWeight<W, K> &w2) {
-  SparsePowerWeight<W, K> result;
-  SparseTupleWeightPlusMapper<W, K> operator_mapper;
-  SparseTupleWeightMap(&result, w1, w2, operator_mapper);
-  return result;
-}
-
-// Semimodule times operation.
-template <class W, class K>
-inline SparsePowerWeight<W, K> Times(const SparsePowerWeight<W, K> &w1,
-                                     const SparsePowerWeight<W, K> &w2) {
-  SparsePowerWeight<W, K> result;
-  SparseTupleWeightTimesMapper<W, K> operator_mapper;
-  SparseTupleWeightMap(&result, w1, w2, operator_mapper);
-  return result;
-}
-
-// Semimodule divide operation.
-template <class W, class K>
-inline SparsePowerWeight<W, K> Divide(const SparsePowerWeight<W, K> &w1,
-                                      const SparsePowerWeight<W, K> &w2,
-                                      DivideType type = DIVIDE_ANY) {
-  SparsePowerWeight<W, K> result;
-  SparseTupleWeightDivideMapper<W, K> operator_mapper(type);
-  SparseTupleWeightMap(&result, w1, w2, operator_mapper);
-  return result;
-}
-
-// Semimodule dot product operation.
-template <class W, class K>
-inline const W &DotProduct(const SparsePowerWeight<W, K> &w1,
-                           const SparsePowerWeight<W, K> &w2) {
-  const SparsePowerWeight<W, K> product = Times(w1, w2);
-  W result(W::Zero());
-  for (SparseTupleWeightIterator<W, K> it(product); !it.Done(); it.Next()) {
-    result = Plus(result, it.Value().second);
-  }
-  return result;
-}
-
-template <class W, class K>
-inline bool ApproxEqual(const SparsePowerWeight<W, K> &w1,
-                        const SparsePowerWeight<W, K> &w2,
-                        float delta = kDelta) {
-  SparseTupleWeight<W, K> result;
-  SparseTupleWeightApproxMapper<W, K> operator_mapper(kDelta);
-  SparseTupleWeightMap(&result, w1, w2, operator_mapper);
-  return result == SparsePowerWeight<W, K>::One();
-}
-
-template <class W, class K>
-inline SparsePowerWeight<W, K> Times(const W &k,
-                                     const SparsePowerWeight<W, K> &w2) {
-  const SparseTupleWeight<W, K> t2(k);
-  const SparsePowerWeight<W, K> w1(t2);
-  return Times(w1, w2);
-}
-
-template <class W, class K>
-inline SparsePowerWeight<W, K> Times(const SparsePowerWeight<W, K> &w1,
-                                     const W &k) {
-  const SparseTupleWeight<W, K> t2(k);
-  const SparsePowerWeight<W, K> w2(t2);
-  return Times(w1, w2);
-}
-
-template <class W, class K>
-inline SparsePowerWeight<W, K> Divide(const SparsePowerWeight<W, K> &w1,
-                                      const W &k,
-                                      DivideType divide_type = DIVIDE_ANY) {
-  const SparseTupleWeight<W, K> t2(k);
-  const SparsePowerWeight<W, K> w2(t2);
-  return Divide(w1, w2, divide_type);
-}
-
-// This function object generates weights over the Cartesian power of rank
-// n over the underlying weight. This is intended primarily for testing.
-template <class W, class K>
-class WeightGenerate<SparsePowerWeight<W, K>> {
- public:
-  using Weight = SparsePowerWeight<W, K>;
-  using Generate = WeightGenerate<W>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t sparse_power_rank = 3)
-      : generate_(allow_zero), sparse_power_rank_(sparse_power_rank) {}
-
-  Weight operator()() const {
-    Weight weight;
-    for (size_t i = 1; i <= sparse_power_rank_; ++i) {
-      weight.PushBack(i, generate_(), true);
-    }
-    return weight;
-  }
-
- private:
-  const Generate generate_;
-  const size_t sparse_power_rank_;
-};
-
-}  // namespace fst
-
-#endif  // FST_SPARSE_POWER_WEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/sparse-tuple-weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/sparse-tuple-weight.h
deleted file mode 100644
index c1711c9c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/sparse-tuple-weight.h
+++ /dev/null
@@ -1,411 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Sparse version of tuple-weight, based on tuple-weight.h.
-// Internally stores sparse key, value pairs in linked list. The default value
-// element is the assumed value of unset keys. Internal singleton
-// implementation that stores first key, value pair as a initialized member
-// variable to avoid unnecessary allocation on heap. Use
-// SparseTupleWeightIterator to iterate through the key,value pairs. Note:
-// this does NOT iterate through the default value.
-//
-// Sparse tuple weight set operation definitions.
-
-#ifndef FST_SPARSE_TUPLE_WEIGHT_H_
-#define FST_SPARSE_TUPLE_WEIGHT_H_
-
-#include <algorithm>
-#include <list>
-#include <stack>
-#include <string>
-#include <unordered_map>
-#include <utility>
-
-
-#include <fst/weight.h>
-
-
-namespace fst {
-
-template <class W, class K>
-class SparseTupleWeightIterator;
-
-// Arbitrary dimension tuple weight, stored as a sorted linked-list.
-// W is any weight class, and K is the key value type. kNoKey (-1) is reserved
-// for internal use.
-template <class W, class K = int>
-class SparseTupleWeight {
- public:
-  using ReverseWeight = SparseTupleWeight<typename W::ReverseWeight, K>;
-
-  using Iterator = SparseTupleWeightIterator<W, K>;
-  using Pair = std::pair<K, W>;
-  using Weight = W;
-  using Index = K;
-
-  constexpr static K kNoKey = -1;
-
-  SparseTupleWeight() { Init(); }
-
-  template <class Iterator>
-  SparseTupleWeight(Iterator begin, Iterator end) {
-    Init();
-    // Assumes input iterator is sorted.
-    for (auto it = begin; it != end; ++it) PushBack(*it);
-  }
-
-  // Initialize component `key` to `weight`, with `default_weight` for all
-  // other components.
-  SparseTupleWeight(const K &key, const W &weight, const W &default_weight)
-      : default_(default_weight),
-        first_(weight == default_weight ? kNoKey : key, weight) {}
-
-  explicit SparseTupleWeight(const W &weight) { Init(weight); }
-
-  SparseTupleWeight(const SparseTupleWeight &weight) {
-    Init(weight.DefaultValue());
-    SetDefaultValue(weight.DefaultValue());
-    for (Iterator it(weight); !it.Done(); it.Next()) {
-      PushBack(it.Value());
-    }
-  }
-
-  SparseTupleWeight(SparseTupleWeight &&weight)
-    // Don't move the default, so weight.default_ is still valid.
-    : default_(weight.default_), first_(std::move(weight.first_)),
-      rest_(std::move(weight.rest_)) {
-    // move leaves the source in a valid but unspecified state.
-    // Make sure the source weight is empty.
-    weight.first_.first = kNoKey;
-    weight.rest_.clear();
-  }
-
-  static const SparseTupleWeight &Zero() {
-    static const SparseTupleWeight zero(W::Zero());
-    return zero;
-  }
-
-  static const SparseTupleWeight &One() {
-    static const SparseTupleWeight one(W::One());
-    return one;
-  }
-
-  static const SparseTupleWeight &NoWeight() {
-    static const SparseTupleWeight no_weight(W::NoWeight());
-    return no_weight;
-  }
-
-  std::istream &Read(std::istream &strm) {
-    ReadType(strm, &default_);
-    ReadType(strm, &first_);
-    return ReadType(strm, &rest_);
-  }
-
-  std::ostream &Write(std::ostream &strm) const {
-    WriteType(strm, default_);
-    WriteType(strm, first_);
-    return WriteType(strm, rest_);
-  }
-
-  SparseTupleWeight &operator=(const SparseTupleWeight &weight) {
-    if (this == &weight) return *this;  // Checks for identity.
-    Init(weight.DefaultValue());
-    for (Iterator it(weight); !it.Done(); it.Next()) {
-      PushBack(it.Value());
-    }
-    return *this;
-  }
-
-  SparseTupleWeight &operator=(SparseTupleWeight &&weight) {
-    if (this == &weight) return *this;  // Checks for identity.
-    default_ = weight.default_;
-    std::swap(first_, weight.first_);
-    std::swap(rest_, weight.rest_);
-    return *this;
-  }
-
-  bool Member() const {
-    if (!DefaultValue().Member()) return false;
-    for (Iterator it(*this); !it.Done(); it.Next()) {
-      if (!it.Value().second.Member()) return false;
-    }
-    return true;
-  }
-
-  // Assumes H() function exists for the hash of the key value.
-  size_t Hash() const {
-    size_t h = 0;
-    static const std::hash<K> H;
-    for (Iterator it(*this); !it.Done(); it.Next()) {
-      h = 5 * h + H(it.Value().first);
-      h = 13 * h + it.Value().second.Hash();
-    }
-    return h;
-  }
-
-  SparseTupleWeight Quantize(float delta = kDelta) const {
-    SparseTupleWeight weight;
-    for (Iterator it(*this); !it.Done(); it.Next()) {
-      weight.PushBack(it.Value().first, it.Value().second.Quantize(delta));
-    }
-    return weight;
-  }
-
-  ReverseWeight Reverse() const {
-    SparseTupleWeight weight;
-    for (Iterator it(*this); !it.Done(); it.Next()) {
-      weight.PushBack(it.Value().first, it.Value().second.Reverse());
-    }
-    return ReverseWeight(weight);
-  }
-
-  void Init(const W &default_value = W::Zero()) {
-    first_.first = kNoKey;
-    // Initialized to the reserved key value.
-    default_ = default_value;
-    rest_.clear();
-  }
-
-  size_t Size() const {
-    if (first_.first == kNoKey) {
-      return 0;
-    } else {
-      return rest_.size() + 1;
-    }
-  }
-
-  inline void PushBack(const K &key, const W &weight,
-                       bool default_value_check = true) {
-    PushBack(std::make_pair(key, weight), default_value_check);
-  }
-
-  inline void PushBack(const Pair &pair, bool default_value_check = true) {
-    if (default_value_check && pair.second == default_) return;
-    if (first_.first == kNoKey) {
-      first_ = pair;
-    } else {
-      rest_.push_back(pair);
-    }
-  }
-
-  // Returns the `key`-th component, or the default value if not set.
-  const W &Value(const K &key) const {
-    // TODO(rybach): Consider binary search.
-    Iterator iter(*this);
-    for (; !iter.Done() && iter.Value().first < key; iter.Next()) continue;
-    return !iter.Done() && iter.Value().first == key ? iter.Value().second
-                                                     : DefaultValue();
-  }
-
-  void SetValue(const K &key, const W &w) {
-    if (w == DefaultValue()) {
-      ClearValue(key);
-    } else {
-      SetValueToNonDefault(key, w);
-    }
-  }
-
-  void SetDefaultValue(const W &value) { default_ = value; }
-
-  const W &DefaultValue() const { return default_; }
-
- private:
-  void SetValueToNonDefault(const K &key, const W &w) {
-    // Don't use SparseTupleWeightIterator, since that's const.
-    if (first_.first == kNoKey) {
-      first_ = Pair(key, w);
-    } else if (key < first_.first) {
-      rest_.push_front(first_);
-      first_ = Pair(key, w);
-    } else if (key == first_.first) {
-      first_.second = w;
-    } else {
-      const auto i =
-          std::find_if(rest_.begin(), rest_.end(),
-                       [key](const Pair &p) { return p.first >= key; });
-      if (i != rest_.end() && i->first == key) {
-        i->second = w;
-      } else {
-        rest_.insert(i, Pair(key, w));
-      }
-    }
-  }
-
-  // Removes the weight value for `key`, having the effect of setting
-  // it to `DefaultValue()`.
-  void ClearValue(const K &key) {
-    if (key == first_.first) {
-      if (!rest_.empty()) {
-        first_ = rest_.front();
-        rest_.pop_front();
-      } else {
-        first_.first = kNoKey;
-      }
-    } else if (key > first_.first) {
-      const auto i =
-          std::find_if(rest_.begin(), rest_.end(),
-                       [key](const Pair &p) { return p.first >= key; });
-      if (i != rest_.end() && i->first == key) {
-        rest_.erase(i);
-      }
-    }
-  }
-
-  // Assumed default value of uninitialized keys, by default W::Zero().
-  W default_;
-
-  // Key values pairs are first stored in first_, then fill rest_ this way we
-  // can avoid dynamic allocation in the common case where the weight is a
-  // single key/value pair.
-  Pair first_;
-  std::list<Pair> rest_;
-
-  friend class SparseTupleWeightIterator<W, K>;
-};
-
-// Declare storage for kNoKey since it is passed by reference.
-template <class W, class K>
-constexpr K SparseTupleWeight<W, K>::kNoKey;
-
-template <class W, class K>
-class SparseTupleWeightIterator {
- public:
-  using Pair = typename SparseTupleWeight<W, K>::Pair;
-  using const_iterator = typename std::list<Pair>::const_iterator;
-  using iterator = typename std::list<Pair>::iterator;
-
-  explicit SparseTupleWeightIterator(const SparseTupleWeight<W, K> &weight)
-      : first_(weight.first_),
-        rest_(weight.rest_),
-        init_(true),
-        iter_(rest_.begin()) {}
-
-  bool Done() const {
-    if (init_) {
-      return first_.first == SparseTupleWeight<W, K>::kNoKey;
-    } else {
-      return iter_ == rest_.end();
-    }
-  }
-
-  const Pair &Value() const { return init_ ? first_ : *iter_; }
-
-  void Next() {
-    if (init_) {
-      init_ = false;
-    } else {
-      ++iter_;
-    }
-  }
-
-  void Reset() {
-    init_ = true;
-    iter_ = rest_.begin();
-  }
-
- private:
-  const Pair &first_;
-  const std::list<Pair> &rest_;
-  bool init_;  // In the initialized state?
-  const_iterator iter_;
-};
-
-template <class W, class K, class M>
-inline void SparseTupleWeightMap(SparseTupleWeight<W, K> *result,
-                                 const SparseTupleWeight<W, K> &w1,
-                                 const SparseTupleWeight<W, K> &w2,
-                                 const M &operator_mapper) {
-  SparseTupleWeightIterator<W, K> w1_it(w1);
-  SparseTupleWeightIterator<W, K> w2_it(w2);
-  const auto &v1_def = w1.DefaultValue();
-  const auto &v2_def = w2.DefaultValue();
-  result->SetDefaultValue(operator_mapper.Map(0, v1_def, v2_def));
-  while (!w1_it.Done() || !w2_it.Done()) {
-    const auto &k1 = (w1_it.Done()) ? w2_it.Value().first : w1_it.Value().first;
-    const auto &k2 = (w2_it.Done()) ? w1_it.Value().first : w2_it.Value().first;
-    const auto &v1 = (w1_it.Done()) ? v1_def : w1_it.Value().second;
-    const auto &v2 = (w2_it.Done()) ? v2_def : w2_it.Value().second;
-    if (k1 == k2) {
-      result->PushBack(k1, operator_mapper.Map(k1, v1, v2));
-      if (!w1_it.Done()) w1_it.Next();
-      if (!w2_it.Done()) w2_it.Next();
-    } else if (k1 < k2) {
-      result->PushBack(k1, operator_mapper.Map(k1, v1, v2_def));
-      w1_it.Next();
-    } else {
-      result->PushBack(k2, operator_mapper.Map(k2, v1_def, v2));
-      w2_it.Next();
-    }
-  }
-}
-
-template <class W, class K>
-inline bool operator==(const SparseTupleWeight<W, K> &w1,
-                       const SparseTupleWeight<W, K> &w2) {
-  const auto &v1_def = w1.DefaultValue();
-  const auto &v2_def = w2.DefaultValue();
-  if (v1_def != v2_def) return false;
-  SparseTupleWeightIterator<W, K> w1_it(w1);
-  SparseTupleWeightIterator<W, K> w2_it(w2);
-  while (!w1_it.Done() || !w2_it.Done()) {
-    const auto &k1 = (w1_it.Done()) ? w2_it.Value().first : w1_it.Value().first;
-    const auto &k2 = (w2_it.Done()) ? w1_it.Value().first : w2_it.Value().first;
-    const auto &v1 = (w1_it.Done()) ? v1_def : w1_it.Value().second;
-    const auto &v2 = (w2_it.Done()) ? v2_def : w2_it.Value().second;
-    if (k1 == k2) {
-      if (v1 != v2) return false;
-      if (!w1_it.Done()) w1_it.Next();
-      if (!w2_it.Done()) w2_it.Next();
-    } else if (k1 < k2) {
-      if (v1 != v2_def) return false;
-      w1_it.Next();
-    } else {
-      if (v1_def != v2) return false;
-      w2_it.Next();
-    }
-  }
-  return true;
-}
-
-template <class W, class K>
-inline bool operator!=(const SparseTupleWeight<W, K> &w1,
-                       const SparseTupleWeight<W, K> &w2) {
-  return !(w1 == w2);
-}
-
-template <class W, class K>
-inline std::ostream &operator<<(std::ostream &strm,
-                                const SparseTupleWeight<W, K> &weight) {
-  CompositeWeightWriter writer(strm);
-  writer.WriteBegin();
-  writer.WriteElement(weight.DefaultValue());
-  for (SparseTupleWeightIterator<W, K> it(weight); !it.Done(); it.Next()) {
-    writer.WriteElement(it.Value().first);
-    writer.WriteElement(it.Value().second);
-  }
-  writer.WriteEnd();
-  return strm;
-}
-
-template <class W, class K>
-inline std::istream &operator>>(std::istream &strm,
-                                SparseTupleWeight<W, K> &weight) {
-  CompositeWeightReader reader(strm);
-  reader.ReadBegin();
-  W def;
-  bool more = reader.ReadElement(&def);
-  weight.Init(def);
-  while (more) {
-    K key;
-    reader.ReadElement(&key);
-    W v;
-    more = reader.ReadElement(&v);
-    weight.PushBack(key, v);
-  }
-  reader.ReadEnd();
-  return strm;
-}
-
-}  // namespace fst
-
-#endif  // FST_SPARSE_TUPLE_WEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/state-map.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/state-map.h
deleted file mode 100644
index 065f1c0b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/state-map.h
+++ /dev/null
@@ -1,613 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to map over/transform states e.g., sort transitions.
-//
-// Consider using when operation does not change the number of states.
-
-#ifndef FST_STATE_MAP_H_
-#define FST_STATE_MAP_H_
-
-#include <algorithm>
-#include <string>
-#include <utility>
-
-#include <fst/log.h>
-
-#include <fst/arc-map.h>
-#include <fst/cache.h>
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-// StateMapper Interface. The class determines how states are mapped; useful for
-// implementing operations that do not change the number of states.
-//
-// class StateMapper {
-//  public:
-//   using FromArc = A;
-//   using ToArc = B;
-//
-//   // Typical constructor.
-//   StateMapper(const Fst<A> &fst);
-//
-//   // Required copy constructor that allows updating FST argument;
-//   // pass only if relevant and changed.
-//   StateMapper(const StateMapper &mapper, const Fst<A> *fst = 0);
-//
-//   // Specifies initial state of result.
-//   B::StateId Start() const;
-//   // Specifies state's final weight in result.
-//   B::Weight Final(B::StateId state) const;
-//
-//   // These methods iterate through a state's arcs in result.
-//
-//   // Specifies state to iterate over.
-//   void SetState(B::StateId state);
-//
-//   // End of arcs?
-//   bool Done() const;
-//
-//   // Current arc.
-//   const B &Value() const;
-//
-//   // Advances to next arc (when !Done)
-//   void Next();
-//
-//   // Specifies input symbol table action the mapper requires (see above).
-//   MapSymbolsAction InputSymbolsAction() const;
-//
-//   // Specifies output symbol table action the mapper requires (see above).
-//   MapSymbolsAction OutputSymbolsAction() const;
-//
-//   // This specifies the known properties of an FST mapped by this
-//   // mapper. It takes as argument the input FST's known properties.
-//   uint64 Properties(uint64 props) const;
-// };
-//
-// We include a various state map versions below. One dimension of variation is
-// whether the mapping mutates its input, writes to a new result FST, or is an
-// on-the-fly Fst. Another dimension is how we pass the mapper. We allow passing
-// the mapper by pointer for cases that we need to change the state of the
-// user's mapper. We also include map versions that pass the mapper by value or
-// const reference when this suffices.
-
-// Maps an arc type A using a mapper function object C, passed by pointer. This
-// version modifies the input FST.
-template <class A, class C>
-void StateMap(MutableFst<A> *fst, C *mapper) {
-  if (mapper->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    fst->SetInputSymbols(nullptr);
-  }
-  if (mapper->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    fst->SetOutputSymbols(nullptr);
-  }
-  if (fst->Start() == kNoStateId) return;
-  const auto props = fst->Properties(kFstProperties, false);
-  fst->SetStart(mapper->Start());
-  for (StateIterator<Fst<A>> siter(*fst); !siter.Done(); siter.Next()) {
-    const auto state = siter.Value();
-    mapper->SetState(state);
-    fst->DeleteArcs(state);
-    for (; !mapper->Done(); mapper->Next()) {
-      fst->AddArc(state, mapper->Value());
-    }
-    fst->SetFinal(state, mapper->Final(state));
-  }
-  fst->SetProperties(mapper->Properties(props), kFstProperties);
-}
-
-// Maps an arc type A using a mapper function object C, passed by value.
-// This version modifies the input FST.
-template <class A, class C>
-void StateMap(MutableFst<A> *fst, C mapper) {
-  StateMap(fst, &mapper);
-}
-
-// Maps an arc type A to an arc type B using mapper functor C, passed by
-// pointer. This version writes to an output FST.
-template <class A, class B, class C>
-void StateMap(const Fst<A> &ifst, MutableFst<B> *ofst, C *mapper) {
-  ofst->DeleteStates();
-  if (mapper->InputSymbolsAction() == MAP_COPY_SYMBOLS) {
-    ofst->SetInputSymbols(ifst.InputSymbols());
-  } else if (mapper->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    ofst->SetInputSymbols(nullptr);
-  }
-  if (mapper->OutputSymbolsAction() == MAP_COPY_SYMBOLS) {
-    ofst->SetOutputSymbols(ifst.OutputSymbols());
-  } else if (mapper->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-    ofst->SetOutputSymbols(nullptr);
-  }
-  const auto iprops = ifst.Properties(kCopyProperties, false);
-  if (ifst.Start() == kNoStateId) {
-    if (iprops & kError) ofst->SetProperties(kError, kError);
-    return;
-  }
-  // Adds all states.
-  if (ifst.Properties(kExpanded, false)) ofst->ReserveStates(CountStates(ifst));
-  for (StateIterator<Fst<A>> siter(ifst); !siter.Done(); siter.Next()) {
-    ofst->AddState();
-  }
-  ofst->SetStart(mapper->Start());
-  for (StateIterator<Fst<A>> siter(ifst); !siter.Done(); siter.Next()) {
-    const auto state = siter.Value();
-    mapper->SetState(state);
-    for (; !mapper->Done(); mapper->Next()) {
-      ofst->AddArc(state, mapper->Value());
-    }
-    ofst->SetFinal(state, mapper->Final(state));
-  }
-  const auto oprops = ofst->Properties(kFstProperties, false);
-  ofst->SetProperties(mapper->Properties(iprops) | oprops, kFstProperties);
-}
-
-// Maps an arc type A to an arc type B using mapper functor object C, passed by
-// value. This version writes to an output FST.
-template <class A, class B, class C>
-void StateMap(const Fst<A> &ifst, MutableFst<B> *ofst, C mapper) {
-  StateMap(ifst, ofst, &mapper);
-}
-
-using StateMapFstOptions = CacheOptions;
-
-template <class A, class B, class C>
-class StateMapFst;
-
-// Facade around StateIteratorBase<A> inheriting from StateIteratorBase<B>.
-template <class A, class B>
-class StateMapStateIteratorBase : public StateIteratorBase<B> {
- public:
-  using Arc = B;
-  using StateId = typename Arc::StateId;
-
-  explicit StateMapStateIteratorBase(StateIteratorBase<A> *base)
-      : base_(base) {}
-
-  bool Done() const final { return base_->Done(); }
-
-  StateId Value() const final { return base_->Value(); }
-
-  void Next() final { base_->Next(); }
-
-  void Reset() final { base_->Reset(); }
-
- private:
-  std::unique_ptr<StateIteratorBase<A>> base_;
-
-  StateMapStateIteratorBase() = delete;
-};
-
-namespace internal {
-
-// Implementation of delayed StateMapFst.
-template <class A, class B, class C>
-class StateMapFstImpl : public CacheImpl<B> {
- public:
-  using Arc = B;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<B>::SetType;
-  using FstImpl<B>::SetProperties;
-  using FstImpl<B>::SetInputSymbols;
-  using FstImpl<B>::SetOutputSymbols;
-
-  using CacheImpl<B>::PushArc;
-  using CacheImpl<B>::HasArcs;
-  using CacheImpl<B>::HasFinal;
-  using CacheImpl<B>::HasStart;
-  using CacheImpl<B>::SetArcs;
-  using CacheImpl<B>::SetFinal;
-  using CacheImpl<B>::SetStart;
-
-  friend class StateIterator<StateMapFst<A, B, C>>;
-
-  StateMapFstImpl(const Fst<A> &fst, const C &mapper,
-                  const StateMapFstOptions &opts)
-      : CacheImpl<B>(opts),
-        fst_(fst.Copy()),
-        mapper_(new C(mapper, fst_.get())),
-        own_mapper_(true) {
-    Init();
-  }
-
-  StateMapFstImpl(const Fst<A> &fst, C *mapper, const StateMapFstOptions &opts)
-      : CacheImpl<B>(opts),
-        fst_(fst.Copy()),
-        mapper_(mapper),
-        own_mapper_(false) {
-    Init();
-  }
-
-  StateMapFstImpl(const StateMapFstImpl<A, B, C> &impl)
-      : CacheImpl<B>(impl),
-        fst_(impl.fst_->Copy(true)),
-        mapper_(new C(*impl.mapper_, fst_.get())),
-        own_mapper_(true) {
-    Init();
-  }
-
-  ~StateMapFstImpl() override {
-    if (own_mapper_) delete mapper_;
-  }
-
-  StateId Start() {
-    if (!HasStart()) SetStart(mapper_->Start());
-    return CacheImpl<B>::Start();
-  }
-
-  Weight Final(StateId state) {
-    if (!HasFinal(state)) SetFinal(state, mapper_->Final(state));
-    return CacheImpl<B>::Final(state);
-  }
-
-  size_t NumArcs(StateId state) {
-    if (!HasArcs(state)) Expand(state);
-    return CacheImpl<B>::NumArcs(state);
-  }
-
-  size_t NumInputEpsilons(StateId state) {
-    if (!HasArcs(state)) Expand(state);
-    return CacheImpl<B>::NumInputEpsilons(state);
-  }
-
-  size_t NumOutputEpsilons(StateId state) {
-    if (!HasArcs(state)) Expand(state);
-    return CacheImpl<B>::NumOutputEpsilons(state);
-  }
-
-  void InitStateIterator(StateIteratorData<B> *datb) const {
-    StateIteratorData<A> data;
-    fst_->InitStateIterator(&data);
-    datb->base = data.base ? new StateMapStateIteratorBase<A, B>(data.base)
-        : nullptr;
-    datb->nstates = data.nstates;
-  }
-
-  void InitArcIterator(StateId state, ArcIteratorData<B> *data) {
-    if (!HasArcs(state)) Expand(state);
-    CacheImpl<B>::InitArcIterator(state, data);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && (fst_->Properties(kError, false) ||
-                            (mapper_->Properties(0) & kError))) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  void Expand(StateId state) {
-    // Adds exiting arcs.
-    for (mapper_->SetState(state); !mapper_->Done(); mapper_->Next()) {
-      PushArc(state, mapper_->Value());
-    }
-    SetArcs(state);
-  }
-
-  const Fst<A> *GetFst() const { return fst_.get(); }
-
- private:
-  void Init() {
-    SetType("statemap");
-    if (mapper_->InputSymbolsAction() == MAP_COPY_SYMBOLS) {
-      SetInputSymbols(fst_->InputSymbols());
-    } else if (mapper_->InputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-      SetInputSymbols(nullptr);
-    }
-    if (mapper_->OutputSymbolsAction() == MAP_COPY_SYMBOLS) {
-      SetOutputSymbols(fst_->OutputSymbols());
-    } else if (mapper_->OutputSymbolsAction() == MAP_CLEAR_SYMBOLS) {
-      SetOutputSymbols(nullptr);
-    }
-    const auto props = fst_->Properties(kCopyProperties, false);
-    SetProperties(mapper_->Properties(props));
-  }
-
-  std::unique_ptr<const Fst<A>> fst_;
-  C *mapper_;
-  bool own_mapper_;
-};
-
-}  // namespace internal
-
-// Maps an arc type A to an arc type B using Mapper function object
-// C. This version is a delayed FST.
-template <class A, class B, class C>
-class StateMapFst : public ImplToFst<internal::StateMapFstImpl<A, B, C>> {
- public:
-  friend class ArcIterator<StateMapFst<A, B, C>>;
-
-  using Arc = B;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::StateMapFstImpl<A, B, C>;
-
-  StateMapFst(const Fst<A> &fst, const C &mapper,
-              const StateMapFstOptions &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, mapper, opts)) {}
-
-  StateMapFst(const Fst<A> &fst, C *mapper, const StateMapFstOptions &opts)
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, mapper, opts)) {}
-
-  StateMapFst(const Fst<A> &fst, const C &mapper)
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, mapper, StateMapFstOptions())) {}
-
-  StateMapFst(const Fst<A> &fst, C *mapper)
-      : ImplToFst<Impl>(
-            std::make_shared<Impl>(fst, mapper, StateMapFstOptions())) {}
-
-  // See Fst<>::Copy() for doc.
-  StateMapFst(const StateMapFst<A, B, C> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Get a copy of this StateMapFst. See Fst<>::Copy() for further doc.
-  StateMapFst<A, B, C> *Copy(bool safe = false) const override {
-    return new StateMapFst<A, B, C>(*this, safe);
-  }
-
-  void InitStateIterator(StateIteratorData<B> *data) const override {
-    GetImpl()->InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId state, ArcIteratorData<B> *data) const override {
-    GetMutableImpl()->InitArcIterator(state, data);
-  }
-
- protected:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
- private:
-  StateMapFst &operator=(const StateMapFst &) = delete;
-};
-
-// Specialization for StateMapFst.
-template <class A, class B, class C>
-class ArcIterator<StateMapFst<A, B, C>>
-    : public CacheArcIterator<StateMapFst<A, B, C>> {
- public:
-  using StateId = typename A::StateId;
-
-  ArcIterator(const StateMapFst<A, B, C> &fst, StateId state)
-      : CacheArcIterator<StateMapFst<A, B, C>>(fst.GetMutableImpl(), state) {
-    if (!fst.GetImpl()->HasArcs(state)) fst.GetMutableImpl()->Expand(state);
-  }
-};
-
-// Utility mappers.
-
-// Mapper that returns its input.
-template <class Arc>
-class IdentityStateMapper {
- public:
-  using FromArc = Arc;
-  using ToArc = Arc;
-
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit IdentityStateMapper(const Fst<Arc> &fst) : fst_(fst) {}
-
-  // Allows updating FST argument; pass only if changed.
-  IdentityStateMapper(const IdentityStateMapper<Arc> &mapper,
-                      const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? *fst : mapper.fst_) {}
-
-  StateId Start() const { return fst_.Start(); }
-
-  Weight Final(StateId state) const { return fst_.Final(state); }
-
-  void SetState(StateId state) {
-    aiter_.reset(new ArcIterator<Fst<Arc>>(fst_, state));
-  }
-
-  bool Done() const { return aiter_->Done(); }
-
-  const Arc &Value() const { return aiter_->Value(); }
-
-  void Next() { aiter_->Next(); }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const { return props; }
-
- private:
-  const Fst<Arc> &fst_;
-  std::unique_ptr<ArcIterator<Fst<Arc>>> aiter_;
-};
-
-template <class Arc>
-class ArcSumMapper {
- public:
-  using FromArc = Arc;
-  using ToArc = Arc;
-
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit ArcSumMapper(const Fst<Arc> &fst) : fst_(fst), i_(0) {}
-
-  // Allows updating FST argument; pass only if changed.
-  ArcSumMapper(const ArcSumMapper<Arc> &mapper, const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? *fst : mapper.fst_), i_(0) {}
-
-  StateId Start() const { return fst_.Start(); }
-
-  Weight Final(StateId state) const { return fst_.Final(state); }
-
-  void SetState(StateId state) {
-    i_ = 0;
-    arcs_.clear();
-    arcs_.reserve(fst_.NumArcs(state));
-    for (ArcIterator<Fst<Arc>> aiter(fst_, state); !aiter.Done();
-         aiter.Next()) {
-      arcs_.push_back(aiter.Value());
-    }
-    // First sorts the exiting arcs by input label, output label and destination
-    // state and then sums weights of arcs with the same input label, output
-    // label, and destination state.
-    std::sort(arcs_.begin(), arcs_.end(), comp_);
-    size_t narcs = 0;
-    for (const auto &arc : arcs_) {
-      if (narcs > 0 && equal_(arc, arcs_[narcs - 1])) {
-        arcs_[narcs - 1].weight = Plus(arcs_[narcs - 1].weight, arc.weight);
-      } else {
-        arcs_[narcs] = arc;
-        ++narcs;
-      }
-    }
-    arcs_.resize(narcs);
-  }
-
-  bool Done() const { return i_ >= arcs_.size(); }
-
-  const Arc &Value() const { return arcs_[i_]; }
-
-  void Next() { ++i_; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return props & kArcSortProperties & kDeleteArcsProperties &
-           kWeightInvariantProperties;
-  }
-
- private:
-  struct Compare {
-    bool operator()(const Arc &x, const Arc &y) const {
-      if (x.ilabel < y.ilabel) return true;
-      if (x.ilabel > y.ilabel) return false;
-      if (x.olabel < y.olabel) return true;
-      if (x.olabel > y.olabel) return false;
-      if (x.nextstate < y.nextstate) return true;
-      if (x.nextstate > y.nextstate) return false;
-      return false;
-    }
-  };
-
-  struct Equal {
-    bool operator()(const Arc &x, const Arc &y) const {
-      return (x.ilabel == y.ilabel && x.olabel == y.olabel &&
-              x.nextstate == y.nextstate);
-    }
-  };
-
-  const Fst<Arc> &fst_;
-  Compare comp_;
-  Equal equal_;
-  std::vector<Arc> arcs_;
-  ssize_t i_;  // Current arc position.
-
-  ArcSumMapper &operator=(const ArcSumMapper &) = delete;
-};
-
-template <class Arc>
-class ArcUniqueMapper {
- public:
-  using FromArc = Arc;
-  using ToArc = Arc;
-
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit ArcUniqueMapper(const Fst<Arc> &fst) : fst_(fst), i_(0) {}
-
-  // Allows updating FST argument; pass only if changed.
-  ArcUniqueMapper(const ArcUniqueMapper<Arc> &mapper,
-                  const Fst<Arc> *fst = nullptr)
-      : fst_(fst ? *fst : mapper.fst_), i_(0) {}
-
-  StateId Start() const { return fst_.Start(); }
-
-  Weight Final(StateId state) const { return fst_.Final(state); }
-
-  void SetState(StateId state) {
-    i_ = 0;
-    arcs_.clear();
-    arcs_.reserve(fst_.NumArcs(state));
-    for (ArcIterator<Fst<Arc>> aiter(fst_, state); !aiter.Done();
-         aiter.Next()) {
-      arcs_.push_back(aiter.Value());
-    }
-    // First sorts the exiting arcs by input label, output label and destination
-    // state and then uniques identical arcs.
-    std::sort(arcs_.begin(), arcs_.end(), comp_);
-    arcs_.erase(std::unique(arcs_.begin(), arcs_.end(), equal_), arcs_.end());
-  }
-
-  bool Done() const { return i_ >= arcs_.size(); }
-
-  const Arc &Value() const { return arcs_[i_]; }
-
-  void Next() { ++i_; }
-
-  constexpr MapSymbolsAction InputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  constexpr MapSymbolsAction OutputSymbolsAction() const {
-    return MAP_COPY_SYMBOLS;
-  }
-
-  uint64 Properties(uint64 props) const {
-    return props & kArcSortProperties & kDeleteArcsProperties;
-  }
-
- private:
-  struct Compare {
-    bool operator()(const Arc &x, const Arc &y) const {
-      if (x.ilabel < y.ilabel) return true;
-      if (x.ilabel > y.ilabel) return false;
-      if (x.olabel < y.olabel) return true;
-      if (x.olabel > y.olabel) return false;
-      if (x.nextstate < y.nextstate) return true;
-      if (x.nextstate > y.nextstate) return false;
-      return false;
-    }
-  };
-
-  struct Equal {
-    bool operator()(const Arc &x, const Arc &y) const {
-      return (x.ilabel == y.ilabel && x.olabel == y.olabel &&
-              x.nextstate == y.nextstate && x.weight == y.weight);
-    }
-  };
-
-  const Fst<Arc> &fst_;
-  Compare comp_;
-  Equal equal_;
-  std::vector<Arc> arcs_;
-  size_t i_;  // Current arc position.
-
-  ArcUniqueMapper &operator=(const ArcUniqueMapper &) = delete;
-};
-
-// Useful aliases when using StdArc.
-
-using StdArcSumMapper = ArcSumMapper<StdArc>;
-
-using StdArcUniqueMapper = ArcUniqueMapper<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_STATE_MAP_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/state-reachable.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/state-reachable.h
deleted file mode 100644
index 36b5559a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/state-reachable.h
+++ /dev/null
@@ -1,224 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Class to determine whether a given (final) state can be reached from some
-// other given state.
-
-#ifndef FST_STATE_REACHABLE_H_
-#define FST_STATE_REACHABLE_H_
-
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/connect.h>
-#include <fst/dfs-visit.h>
-#include <fst/fst.h>
-#include <fst/interval-set.h>
-#include <fst/vector-fst.h>
-
-
-namespace fst {
-
-// Computes the (final) states reachable from a given state in an FST. After
-// this visitor has been called, a final state f can be reached from a state
-// s iff (*isets)[s].Member(state2index[f]) is true, where (*isets[s]) is a
-// set of half-open inteval of final state indices and state2index[f] maps from
-// a final state to its index. If state2index is empty, it is filled-in with
-// suitable indices. If it is non-empty, those indices are used; in this case,
-// the final states must have out-degree 0.
-template <class Arc, class I = typename Arc::StateId, class S = IntervalSet<I>>
-class IntervalReachVisitor {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Index = I;
-  using ISet = S;
-  using Interval = typename ISet::Interval;
-
-  IntervalReachVisitor(const Fst<Arc> &fst, std::vector<S> *isets,
-                       std::vector<Index> *state2index)
-      : fst_(fst),
-        isets_(isets),
-        state2index_(state2index),
-        index_(state2index->empty() ? 1 : -1),
-        error_(false) {
-    isets_->clear();
-  }
-
-  void InitVisit(const Fst<Arc> &) { error_ = false; }
-
-  bool InitState(StateId s, StateId r) {
-    while (isets_->size() <= s) isets_->push_back(S());
-    while (state2index_->size() <= s) state2index_->push_back(-1);
-    if (fst_.Final(s) != Weight::Zero()) {
-      // Create tree interval.
-      auto *intervals = (*isets_)[s].MutableIntervals();
-      if (index_ < 0) {  // Uses state2index_ map to set index.
-        if (fst_.NumArcs(s) > 0) {
-          FSTERROR() << "IntervalReachVisitor: state2index map must be empty "
-                     << "for this FST";
-          error_ = true;
-          return false;
-        }
-        const auto index = (*state2index_)[s];
-        if (index < 0) {
-          FSTERROR() << "IntervalReachVisitor: state2index map incomplete";
-          error_ = true;
-          return false;
-        }
-        intervals->push_back(Interval(index, index + 1));
-      } else {  // Use pre-order index.
-        intervals->push_back(Interval(index_, index_ + 1));
-        (*state2index_)[s] = index_++;
-      }
-    }
-    return true;
-  }
-
-  constexpr bool TreeArc(StateId, const Arc &) const { return true; }
-
-  bool BackArc(StateId s, const Arc &arc) {
-    FSTERROR() << "IntervalReachVisitor: Cyclic input";
-    error_ = true;
-    return false;
-  }
-
-  bool ForwardOrCrossArc(StateId s, const Arc &arc) {
-    // Non-tree interval.
-    (*isets_)[s].Union((*isets_)[arc.nextstate]);
-    return true;
-  }
-
-  void FinishState(StateId s, StateId p, const Arc *) {
-    if (index_ >= 0 && fst_.Final(s) != Weight::Zero()) {
-      auto *intervals = (*isets_)[s].MutableIntervals();
-      (*intervals)[0].end = index_;  // Updates tree interval end.
-    }
-    (*isets_)[s].Normalize();
-    if (p != kNoStateId) {
-      (*isets_)[p].Union((*isets_)[s]);  // Propagates intervals to parent.
-    }
-  }
-
-  void FinishVisit() {}
-
-  bool Error() const { return error_; }
-
- private:
-  const Fst<Arc> &fst_;
-  std::vector<ISet> *isets_;
-  std::vector<Index> *state2index_;
-  Index index_;
-  bool error_;
-};
-
-// Tests reachability of final states from a given state. To test for
-// reachability from a state s, first do SetState(s). Then a final state f can
-// be reached from state s of FST iff Reach(f) is true. The input can be cyclic,
-// but no cycle may contain a final state.
-template <class Arc, class I = typename Arc::StateId, class S = IntervalSet<I>>
-class StateReachable {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Index = I;
-  using ISet = S;
-  using Interval = typename ISet::Interval;
-
-  explicit StateReachable(const Fst<Arc> &fst) : error_(false) {
-    if (fst.Properties(kAcyclic, true)) {
-      AcyclicStateReachable(fst);
-    } else {
-      CyclicStateReachable(fst);
-    }
-  }
-
-  explicit StateReachable(const StateReachable<Arc> &reachable) {
-    FSTERROR() << "Copy constructor for state reachable class "
-               << "not implemented.";
-    error_ = true;
-  }
-
-  // Sets current state.
-  void SetState(StateId s) { s_ = s; }
-
-  // Can reach this final state from current state?
-  bool Reach(StateId s) {
-    if (s >= state2index_.size()) return false;
-    const auto i = state2index_[s];
-    if (i < 0) {
-      FSTERROR() << "StateReachable: State non-final: " << s;
-      error_ = true;
-      return false;
-    }
-    return isets_[s_].Member(i);
-  }
-
-  // Access to the state-to-index mapping. Unassigned states have index -1.
-  std::vector<Index> &State2Index() { return state2index_; }
-
-  // Access to the interval sets. These specify the reachability to the final
-  // states as intervals of the final state indices.
-  const std::vector<ISet> &IntervalSets() { return isets_; }
-
-  bool Error() const { return error_; }
-
- private:
-  void AcyclicStateReachable(const Fst<Arc> &fst) {
-    IntervalReachVisitor<Arc, StateId, ISet> reach_visitor(fst, &isets_,
-                                                           &state2index_);
-    DfsVisit(fst, &reach_visitor);
-    if (reach_visitor.Error()) error_ = true;
-  }
-
-  void CyclicStateReachable(const Fst<Arc> &fst) {
-    // Finds state reachability on the acyclic condensation FST.
-    VectorFst<Arc> cfst;
-    std::vector<StateId> scc;
-    Condense(fst, &cfst, &scc);
-    StateReachable reachable(cfst);
-    if (reachable.Error()) {
-      error_ = true;
-      return;
-    }
-    // Gets the number of states per SCC.
-    std::vector<size_t> nscc;
-    for (StateId s = 0; s < scc.size(); ++s) {
-      const auto c = scc[s];
-      while (c >= nscc.size()) nscc.push_back(0);
-      ++nscc[c];
-    }
-    // Constructs the interval sets and state index mapping for the original
-    // FST from the condensation FST.
-    state2index_.resize(scc.size(), -1);
-    isets_.resize(scc.size());
-    for (StateId s = 0; s < scc.size(); ++s) {
-      const auto c = scc[s];
-      isets_[s] = reachable.IntervalSets()[c];
-      state2index_[s] = reachable.State2Index()[c];
-      // Checks that each final state in an input FST is not contained in a
-      // cycle (i.e., not in a non-trivial SCC).
-      if (cfst.Final(c) != Weight::Zero() && nscc[c] > 1) {
-        FSTERROR() << "StateReachable: Final state contained in a cycle";
-        error_ = true;
-        return;
-      }
-    }
-  }
-
-  StateId s_;                       // Current state.
-  std::vector<ISet> isets_;         // Interval sets per state.
-  std::vector<Index> state2index_;  // Finds index for a final state.
-  bool error_;
-
-  StateReachable &operator=(const StateReachable &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_STATE_REACHABLE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/state-table.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/state-table.h
deleted file mode 100644
index eec4f9cd..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/state-table.h
+++ /dev/null
@@ -1,494 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes for representing the mapping between state tuples and state IDs.
-
-#ifndef FST_STATE_TABLE_H_
-#define FST_STATE_TABLE_H_
-
-#include <deque>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/bi-table.h>
-#include <fst/expanded-fst.h>
-#include <fst/filter-state.h>
-
-
-namespace fst {
-
-// State tables determine the bijective mapping between state tuples (e.g., in
-// composition, triples of two FST states and a composition filter state) and
-// their corresponding state IDs. They are classes, templated on state tuples,
-// with the following interface:
-//
-// template <class T>
-// class StateTable {
-//  public:
-//   using StateTuple = T;
-//
-//   // Required constructors.
-//   StateTable();
-//
-//   StateTable(const StateTable &);
-//
-//   // Looks up state ID by tuple. If it doesn't exist, then add it.
-//   StateId FindState(const StateTuple &tuple);
-//
-//   // Looks up state tuple by state ID.
-//   const StateTuple<StateId> &Tuple(StateId s) const;
-//
-//   // # of stored tuples.
-//   StateId Size() const;
-// };
-//
-// A state tuple has the form:
-//
-// template <class S>
-// struct StateTuple {
-//   using StateId = S;
-//
-//   // Required constructors.
-//
-//   StateTuple();
-//
-//   StateTuple(const StateTuple &tuple);
-// };
-
-// An implementation using a hash map for the tuple to state ID mapping. The
-// state tuple T must support operator==.
-template <class T, class H>
-class HashStateTable : public HashBiTable<typename T::StateId, T, H> {
- public:
-  using StateTuple = T;
-  using StateId = typename StateTuple::StateId;
-
-  using HashBiTable<StateId, StateTuple, H>::FindId;
-  using HashBiTable<StateId, StateTuple, H>::FindEntry;
-  using HashBiTable<StateId, StateTuple, H>::Size;
-
-  HashStateTable() : HashBiTable<StateId, StateTuple, H>() {}
-
-  explicit HashStateTable(size_t table_size)
-      : HashBiTable<StateId, StateTuple, H>(table_size) {}
-
-  StateId FindState(const StateTuple &tuple) { return FindId(tuple); }
-
-  const StateTuple &Tuple(StateId s) const { return FindEntry(s); }
-};
-
-// An implementation using a hash map for the tuple to state ID mapping. The
-// state tuple T must support operator==.
-template <class T, class H>
-class CompactHashStateTable
-    : public CompactHashBiTable<typename T::StateId, T, H> {
- public:
-  using StateTuple = T;
-  using StateId = typename StateTuple::StateId;
-
-  using CompactHashBiTable<StateId, StateTuple, H>::FindId;
-  using CompactHashBiTable<StateId, StateTuple, H>::FindEntry;
-  using CompactHashBiTable<StateId, StateTuple, H>::Size;
-
-  CompactHashStateTable() : CompactHashBiTable<StateId, StateTuple, H>() {}
-
-  explicit CompactHashStateTable(size_t table_size)
-      : CompactHashBiTable<StateId, StateTuple, H>(table_size) {}
-
-  StateId FindState(const StateTuple &tuple) { return FindId(tuple); }
-
-  const StateTuple &Tuple(StateId s) const { return FindEntry(s); }
-};
-
-// An implementation using a vector for the tuple to state mapping. It is
-// passed a fingerprint functor that should fingerprint tuples uniquely to an
-// integer that can used as a vector index. Normally, VectorStateTable
-// constructs the fingerprint functor. Alternately, the user can pass this
-// objert, in which case the table takes ownership.
-template <class T, class FP>
-class VectorStateTable : public VectorBiTable<typename T::StateId, T, FP> {
- public:
-  using StateTuple = T;
-  using StateId = typename StateTuple::StateId;
-
-  using VectorBiTable<StateId, StateTuple, FP>::FindId;
-  using VectorBiTable<StateId, StateTuple, FP>::FindEntry;
-  using VectorBiTable<StateId, StateTuple, FP>::Size;
-  using VectorBiTable<StateId, StateTuple, FP>::Fingerprint;
-
-  explicit VectorStateTable(FP *fingerprint = nullptr, size_t table_size = 0)
-      : VectorBiTable<StateId, StateTuple, FP>(fingerprint, table_size) {}
-
-  StateId FindState(const StateTuple &tuple) { return FindId(tuple); }
-
-  const StateTuple &Tuple(StateId s) const { return FindEntry(s); }
-};
-
-// An implementation using a vector and a compact hash table. The selection
-// functor returns true for tuples to be hashed in the vector. The fingerprint
-// functor should fingerprint tuples uniquely to an integer that can be used as
-// a vector index. A hash functor is used when hashing tuples into the compact
-// hash table.
-template <class T, class Select, class FP, class H>
-class VectorHashStateTable
-    : public VectorHashBiTable<typename T::StateId, T, Select, FP, H> {
- public:
-  using StateTuple = T;
-  using StateId = typename StateTuple::StateId;
-
-  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::FindId;
-  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::FindEntry;
-  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::Size;
-  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::Selector;
-  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::Fingerprint;
-  using VectorHashBiTable<StateId, StateTuple, Select, FP, H>::Hash;
-
-  VectorHashStateTable(Select *select, FP *fingerprint, H *hash,
-                       size_t vector_size = 0, size_t tuple_size = 0)
-      : VectorHashBiTable<StateId, StateTuple, Select, FP, H>(
-            select, fingerprint, hash, vector_size, tuple_size) {}
-
-  StateId FindState(const StateTuple &tuple) { return FindId(tuple); }
-
-  const StateTuple &Tuple(StateId s) const { return FindEntry(s); }
-};
-
-// An implementation using a hash map to map from tuples to state IDs. This
-// version permits erasing of states. The state tuple's default constructor
-// must produce a tuple that will never be seen and the table must suppor
-// operator==.
-template <class T, class H>
-class ErasableStateTable : public ErasableBiTable<typename T::StateId, T, H> {
- public:
-  using StateTuple = T;
-  using StateId = typename StateTuple::StateId;
-
-  using ErasableBiTable<StateId, StateTuple, H>::FindId;
-  using ErasableBiTable<StateId, StateTuple, H>::FindEntry;
-  using ErasableBiTable<StateId, StateTuple, H>::Size;
-  using ErasableBiTable<StateId, StateTuple, H>::Erase;
-
-  ErasableStateTable() : ErasableBiTable<StateId, StateTuple, H>() {}
-
-  StateId FindState(const StateTuple &tuple) { return FindId(tuple); }
-
-  const StateTuple &Tuple(StateId s) const { return FindEntry(s); }
-};
-
-// The composition state table has the form:
-//
-// template <class Arc, class FilterState>
-// class ComposeStateTable {
-//  public:
-//   using StateId = typename Arc::StateId;
-//
-//   // Required constructors.
-//
-//   ComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2);
-//   ComposeStateTable(const ComposeStateTable<Arc, FilterState> &table);
-//
-//   // Looks up a state ID by tuple, adding it if doesn't exist.
-//   StateId FindState(const StateTuple &tuple);
-//
-//   // Looks up a tuple by state ID.
-//   const ComposeStateTuple<StateId> &Tuple(StateId s) const;
-//
-//   // The number of of stored tuples.
-//   StateId Size() const;
-//
-//   // Return true if error was encountered.
-//   bool Error() const;
-// };
-//
-// The following interface is used to represent the composition state.
-//
-// template <class S, class FS>
-// class CompositionStateTuple {
-//  public:
-//   using StateId = typename StateId;
-//   using FS = FilterState;
-//
-//   // Required constructors.
-//   StateTuple();
-//   StateTuple(StateId s1, StateId s2, const FilterState &fs);
-//
-//   StateId StateId1() const;
-//   StateId StateId2() const;
-//
-//   FilterState GetFilterState() const;
-//
-//   std::pair<StateId, StateId> StatePair() const;
-//
-//   size_t Hash() const;
-//
-//   friend bool operator==(const StateTuple& x, const StateTuple &y);
-// }
-//
-template <typename S, typename FS>
-class DefaultComposeStateTuple {
- public:
-  using StateId = S;
-  using FilterState = FS;
-
-  DefaultComposeStateTuple()
-      : state_pair_(kNoStateId, kNoStateId), fs_(FilterState::NoState()) {}
-
-  DefaultComposeStateTuple(StateId s1, StateId s2, const FilterState &fs)
-      : state_pair_(s1, s2), fs_(fs) {}
-
-  StateId StateId1() const { return state_pair_.first; }
-
-  StateId StateId2() const { return state_pair_.second; }
-
-  FilterState GetFilterState() const { return fs_; }
-
-  const std::pair<StateId, StateId> &StatePair() const { return state_pair_; }
-
-  friend bool operator==(const DefaultComposeStateTuple &x,
-                         const DefaultComposeStateTuple &y) {
-    return (&x == &y) || (x.state_pair_ == y.state_pair_ && x.fs_ == y.fs_);
-  }
-
-  size_t Hash() const {
-    return static_cast<size_t>(StateId1()) +
-           static_cast<size_t>(StateId2()) * 7853u +
-           GetFilterState().Hash() * 7867u;
-  }
-
- private:
-  std::pair<StateId, StateId> state_pair_;
-  FilterState fs_;  // State of composition filter.
-};
-
-// Specialization for TrivialFilterState that does not explicitely store the
-// filter state since it is always the unique non-blocking state.
-template <typename S>
-class DefaultComposeStateTuple<S, TrivialFilterState> {
- public:
-  using StateId = S;
-  using FilterState = TrivialFilterState;
-
-  DefaultComposeStateTuple()
-      : state_pair_(kNoStateId, kNoStateId) {}
-
-  DefaultComposeStateTuple(StateId s1, StateId s2, const FilterState &)
-      : state_pair_(s1, s2) {}
-
-  StateId StateId1() const { return state_pair_.first; }
-
-  StateId StateId2() const { return state_pair_.second; }
-
-  FilterState GetFilterState() const { return FilterState(true); }
-
-  const std::pair<StateId, StateId> &StatePair() const { return state_pair_; }
-
-  friend bool operator==(const DefaultComposeStateTuple &x,
-                         const DefaultComposeStateTuple &y) {
-    return (&x == &y) || (x.state_pair_ == y.state_pair_);
-  }
-
-  size_t Hash() const { return StateId1() + StateId2() * 7853; }
-
- private:
-  std::pair<StateId, StateId> state_pair_;
-};
-
-// Hashing of composition state tuples.
-template <typename T>
-class ComposeHash {
- public:
-  size_t operator()(const T &t) const { return t.Hash(); }
-};
-
-// A HashStateTable over composition tuples.
-template <typename Arc, typename FilterState,
-          typename StateTuple =
-              DefaultComposeStateTuple<typename Arc::StateId, FilterState>,
-          typename StateTable =
-              CompactHashStateTable<StateTuple, ComposeHash<StateTuple>>>
-class GenericComposeStateTable : public StateTable {
- public:
-  using StateId = typename Arc::StateId;
-
-  GenericComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {}
-
-  GenericComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-                           size_t table_size)
-      : StateTable(table_size) {}
-
-  constexpr bool Error() const { return false; }
-
- private:
-  GenericComposeStateTable &operator=(const GenericComposeStateTable &table) =
-      delete;
-};
-
-//  Fingerprint for general composition tuples.
-template <typename StateTuple>
-class ComposeFingerprint {
- public:
-  using StateId = typename StateTuple::StateId;
-
-  // Required but suboptimal constructor.
-  ComposeFingerprint() : mult1_(8192), mult2_(8192) {
-    LOG(WARNING) << "TupleFingerprint: # of FST states should be provided.";
-  }
-
-  // Constructor is provided the sizes of the input FSTs.
-  ComposeFingerprint(StateId nstates1, StateId nstates2)
-      : mult1_(nstates1), mult2_(nstates1 * nstates2) {}
-
-  size_t operator()(const StateTuple &tuple) {
-    return tuple.StateId1() + tuple.StateId2() * mult1_ +
-           tuple.GetFilterState().Hash() * mult2_;
-  }
-
- private:
-  const ssize_t mult1_;
-  const ssize_t mult2_;
-};
-
-// Useful when the first composition state determines the tuple.
-template <typename StateTuple>
-class ComposeState1Fingerprint {
- public:
-  size_t operator()(const StateTuple &tuple) { return tuple.StateId1(); }
-};
-
-// Useful when the second composition state determines the tuple.
-template <typename StateTuple>
-class ComposeState2Fingerprint {
- public:
-  size_t operator()(const StateTuple &tuple) { return tuple.StateId2(); }
-};
-
-// A VectorStateTable over composition tuples. This can be used when the
-// product of number of states in FST1 and FST2 (and the composition filter
-// state hash) is manageable. If the FSTs are not expanded FSTs, they will
-// first have their states counted.
-template <typename Arc, typename StateTuple>
-class ProductComposeStateTable
-    : public VectorStateTable<StateTuple, ComposeFingerprint<StateTuple>> {
- public:
-  using StateId = typename Arc::StateId;
-  using StateTable =
-      VectorStateTable<StateTuple, ComposeFingerprint<StateTuple>>;
-
-  ProductComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-                           size_t table_size = 0)
-      : StateTable(new ComposeFingerprint<StateTuple>(CountStates(fst1),
-                                                      CountStates(fst2)),
-                   table_size) {}
-
-  ProductComposeStateTable(
-      const ProductComposeStateTable<Arc, StateTuple> &table)
-      : StateTable(new ComposeFingerprint<StateTuple>(table.Fingerprint())) {}
-
-  constexpr bool Error() const { return false; }
-
- private:
-  ProductComposeStateTable &operator=(const ProductComposeStateTable &table) =
-      delete;
-};
-
-// A vector-backed table over composition tuples which can be used when the
-// first FST is a string (i.e., satisfies kStringProperties) and the second is
-// deterministic and epsilon-free. It should be used with a composition filter
-// that creates at most one filter state per tuple under these conditions (e.g.,
-// SequenceComposeFilter or MatchComposeFilter).
-template <typename Arc, typename StateTuple>
-class StringDetComposeStateTable
-    : public VectorStateTable<StateTuple,
-                              ComposeState1Fingerprint<StateTuple>> {
- public:
-  using StateId = typename Arc::StateId;
-  using StateTable =
-      VectorStateTable<StateTuple, ComposeState1Fingerprint<StateTuple>>;
-
-  StringDetComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2)
-      : error_(false) {
-    static constexpr auto props2 = kIDeterministic | kNoIEpsilons;
-    if (fst1.Properties(kString, true) != kString) {
-      FSTERROR() << "StringDetComposeStateTable: 1st FST is not a string";
-      error_ = true;
-    } else if (fst2.Properties(props2, true) != props2) {
-      FSTERROR() << "StringDetComposeStateTable: 2nd FST is not deterministic "
-                    "and epsilon-free";
-      error_ = true;
-    }
-  }
-
-  StringDetComposeStateTable(
-      const StringDetComposeStateTable<Arc, StateTuple> &table)
-      : StateTable(table), error_(table.error_) {}
-
-  bool Error() const { return error_; }
-
- private:
-  bool error_;
-
-  StringDetComposeStateTable &operator=(const StringDetComposeStateTable &) =
-      delete;
-};
-
-// A vector-backed table over composition tuples which can be used when the
-// first FST is deterministic and epsilon-free and the second is a string (i.e.,
-// satisfies kString). It should be used with a composition filter that creates
-// at most one filter state per tuple under these conditions (e.g.,
-// SequenceComposeFilter or MatchComposeFilter).
-template <typename Arc, typename StateTuple>
-class DetStringComposeStateTable
-    : public VectorStateTable<StateTuple,
-                              ComposeState2Fingerprint<StateTuple>> {
- public:
-  using StateId = typename Arc::StateId;
-  using StateTable =
-      VectorStateTable<StateTuple, ComposeState2Fingerprint<StateTuple>>;
-
-  DetStringComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2)
-      : error_(false) {
-    static constexpr auto props = kODeterministic | kNoOEpsilons;
-    if (fst1.Properties(props, true) != props) {
-      FSTERROR() << "StringDetComposeStateTable: 1st FST is not "
-                 << "input-deterministic and epsilon-free";
-      error_ = true;
-    } else if (fst2.Properties(kString, true) != kString) {
-      FSTERROR() << "DetStringComposeStateTable: 2nd FST is not a string";
-      error_ = true;
-    }
-  }
-
-  DetStringComposeStateTable(
-      const DetStringComposeStateTable<Arc, StateTuple> &table)
-      : StateTable(table), error_(table.error_) {}
-
-  bool Error() const { return error_; }
-
- private:
-  bool error_;
-
-  DetStringComposeStateTable &operator=(const DetStringComposeStateTable &) =
-      delete;
-};
-
-// An erasable table over composition tuples. The Erase(StateId) method can be
-// called if the user either is sure that composition will never return to that
-// tuple or doesn't care that if it does, it is assigned a new state ID.
-template <typename Arc, typename StateTuple>
-class ErasableComposeStateTable
-    : public ErasableStateTable<StateTuple, ComposeHash<StateTuple>> {
- public:
-  ErasableComposeStateTable(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {}
-
-  constexpr bool Error() const { return false; }
-
- private:
-  ErasableComposeStateTable &operator=(const ErasableComposeStateTable &table) =
-      delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_STATE_TABLE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/statesort.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/statesort.h
deleted file mode 100644
index 346c7d32..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/statesort.h
+++ /dev/null
@@ -1,74 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function to sort states of an FST.
-
-#ifndef FST_STATESORT_H_
-#define FST_STATESORT_H_
-
-#include <algorithm>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-// Sorts the input states of an FST. order[i] gives the the state ID after
-// sorting that corresponds to the state ID i before sorting; it must
-// therefore be a permutation of the input FST's states ID sequence.
-template <class Arc>
-void StateSort(MutableFst<Arc> *fst,
-               const std::vector<typename Arc::StateId> &order) {
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  if (order.size() != fst->NumStates()) {
-    FSTERROR() << "StateSort: Bad order vector size: " << order.size();
-    fst->SetProperties(kError, kError);
-    return;
-  }
-  if (fst->Start() == kNoStateId) return;
-  const auto props = fst->Properties(kStateSortProperties, false);
-  std::vector<bool> done(order.size(), false);
-  std::vector<Arc> arcsa;
-  std::vector<Arc> arcsb;
-  fst->SetStart(order[fst->Start()]);
-  for (StateIterator<MutableFst<Arc>> siter(*fst); !siter.Done();
-       siter.Next()) {
-    auto s1 = siter.Value();
-    StateId s2;
-    if (done[s1]) continue;
-    auto final1 = fst->Final(s1);
-    auto final2 = Weight::Zero();
-    arcsa.clear();
-    for (ArcIterator<MutableFst<Arc>> aiter(*fst, s1); !aiter.Done();
-         aiter.Next()) {
-      arcsa.push_back(aiter.Value());
-    }
-    for (; !done[s1]; s1 = s2, final1 = final2, std::swap(arcsa, arcsb)) {
-      s2 = order[s1];
-      if (!done[s2]) {
-        final2 = fst->Final(s2);
-        arcsb.clear();
-        for (ArcIterator<MutableFst<Arc>> aiter(*fst, s2); !aiter.Done();
-             aiter.Next()) {
-          arcsb.push_back(aiter.Value());
-        }
-      }
-      fst->SetFinal(s2, final1);
-      fst->DeleteArcs(s2);
-      for (auto arc : arcsa) {  // Copy intended.
-        arc.nextstate = order[arc.nextstate];
-        fst->AddArc(s2, arc);
-      }
-      done[s1] = true;
-    }
-  }
-  fst->SetProperties(props, kFstProperties);
-}
-
-}  // namespace fst
-
-#endif  // FST_STATESORT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/string-weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/string-weight.h
deleted file mode 100644
index fb3e70f1..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/string-weight.h
+++ /dev/null
@@ -1,807 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// String weight set and associated semiring operation definitions.
-
-#ifndef FST_STRING_WEIGHT_H_
-#define FST_STRING_WEIGHT_H_
-
-#include <cstdlib>
-
-#include <list>
-#include <string>
-#include <vector>
-
-#include <fst/product-weight.h>
-#include <fst/union-weight.h>
-#include <fst/weight.h>
-
-
-namespace fst {
-
-constexpr int kStringInfinity = -1;     // Label for the infinite string.
-constexpr int kStringBad = -2;          // Label for a non-string.
-constexpr char kStringSeparator = '_';  // Label separator in strings.
-
-// Determines whether to use left or right string semiring.  Includes a
-// 'restricted' version that signals an error if proper prefixes/suffixes
-// would otherwise be returned by Plus, useful with various
-// algorithms that require functional transducer input with the
-// string semirings.
-enum StringType { STRING_LEFT = 0, STRING_RIGHT = 1, STRING_RESTRICT = 2 };
-
-constexpr StringType ReverseStringType(StringType s) {
-  return s == STRING_LEFT ? STRING_RIGHT
-                          : (s == STRING_RIGHT ? STRING_LEFT : STRING_RESTRICT);
-}
-
-template <class>
-class StringWeightIterator;
-template <class>
-class StringWeightReverseIterator;
-
-// String semiring: (longest_common_prefix/suffix, ., Infinity, Epsilon)
-template <typename Label_, StringType S = STRING_LEFT>
-class StringWeight {
- public:
-  using Label = Label_;
-  using ReverseWeight = StringWeight<Label, ReverseStringType(S)>;
-  using Iterator = StringWeightIterator<StringWeight>;
-  using ReverseIterator = StringWeightReverseIterator<StringWeight>;
-
-  friend class StringWeightIterator<StringWeight>;
-  friend class StringWeightReverseIterator<StringWeight>;
-
-  StringWeight() {}
-
-  template <typename Iterator>
-  StringWeight(const Iterator &begin, const Iterator &end) {
-    for (auto iter = begin; iter != end; ++iter) PushBack(*iter);
-  }
-
-  explicit StringWeight(Label label) { PushBack(label); }
-
-  static const StringWeight &Zero() {
-    static const auto *const zero = new StringWeight(Label(kStringInfinity));
-    return *zero;
-  }
-
-  static const StringWeight &One() {
-    static const auto *const one = new StringWeight();
-    return *one;
-  }
-
-  static const StringWeight &NoWeight() {
-    static const auto *const no_weight = new StringWeight(Label(kStringBad));
-    return *no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string(
-        S == STRING_LEFT
-            ? "left_string"
-            : (S == STRING_RIGHT ? "right_string" : "restricted_string"));
-    return *type;
-  }
-
-  bool Member() const;
-
-  std::istream &Read(std::istream &strm);
-
-  std::ostream &Write(std::ostream &strm) const;
-
-  size_t Hash() const;
-
-  StringWeight Quantize(float delta = kDelta) const { return *this; }
-
-  ReverseWeight Reverse() const;
-
-  static constexpr uint64 Properties() {
-    return kIdempotent |
-           (S == STRING_LEFT ? kLeftSemiring
-                             : (S == STRING_RIGHT
-                                    ? kRightSemiring
-                                    : /* S == STRING_RESTRICT */ kLeftSemiring |
-                                          kRightSemiring));
-  }
-
-  // These operations combined with the StringWeightIterator and
-  // StringWeightReverseIterator provide the access and mutation of the string
-  // internal elements.
-
-  // Clear existing StringWeight.
-  void Clear() {
-    first_ = 0;
-    rest_.clear();
-  }
-
-  size_t Size() const { return first_ ? rest_.size() + 1 : 0; }
-
-  void PushFront(Label label) {
-    if (first_) rest_.push_front(first_);
-    first_ = label;
-  }
-
-  void PushBack(Label label) {
-    if (!first_) {
-      first_ = label;
-    } else {
-      rest_.push_back(label);
-    }
-  }
-
- private:
-  Label first_ = 0;        // First label in string (0 if empty).
-  std::list<Label> rest_;  // Remaining labels in string.
-};
-
-// Traverses string in forward direction.
-template <class StringWeight_>
-class StringWeightIterator {
- public:
-  using Weight = StringWeight_;
-  using Label = typename Weight::Label;
-
-  explicit StringWeightIterator(const Weight &w)
-      : first_(w.first_), rest_(w.rest_), init_(true), iter_(rest_.begin()) {}
-
-  bool Done() const {
-    if (init_) {
-      return first_ == 0;
-    } else {
-      return iter_ == rest_.end();
-    }
-  }
-
-  const Label &Value() const { return init_ ? first_ : *iter_; }
-
-  void Next() {
-    if (init_) {
-      init_ = false;
-    } else {
-      ++iter_;
-    }
-  }
-
-  void Reset() {
-    init_ = true;
-    iter_ = rest_.begin();
-  }
-
- private:
-  const Label &first_;
-  const decltype(Weight::rest_) &rest_;
-  bool init_;  // In the initialized state?
-  typename std::remove_reference<decltype (Weight::rest_)>::type::const_iterator iter_;
-};
-
-// Traverses string in backward direction.
-template <class StringWeight_>
-class StringWeightReverseIterator {
- public:
-  using Weight = StringWeight_;
-  using Label = typename Weight::Label;
-
-  explicit StringWeightReverseIterator(const Weight &w)
-      : first_(w.first_),
-        rest_(w.rest_),
-        fin_(first_ == Label()),
-        iter_(rest_.rbegin()) {}
-
-  bool Done() const { return fin_; }
-
-  const Label &Value() const { return iter_ == rest_.rend() ? first_ : *iter_; }
-
-  void Next() {
-    if (iter_ == rest_.rend()) {
-      fin_ = true;
-    } else {
-      ++iter_;
-    }
-  }
-
-  void Reset() {
-    fin_ = false;
-    iter_ = rest_.rbegin();
-  }
-
- private:
-  const Label &first_;
-  const decltype(Weight::rest_) &rest_;
-  bool fin_;  // In the final state?
-  typename std::remove_reference<decltype (Weight::rest_)>::type::const_reverse_iterator iter_;
-};
-
-// StringWeight member functions follow that require
-// StringWeightIterator or StringWeightReverseIterator.
-
-template <typename Label, StringType S>
-inline std::istream &StringWeight<Label, S>::Read(std::istream &strm) {
-  Clear();
-  int32 size;
-  ReadType(strm, &size);
-  for (int32 i = 0; i < size; ++i) {
-    Label label;
-    ReadType(strm, &label);
-    PushBack(label);
-  }
-  return strm;
-}
-
-template <typename Label, StringType S>
-inline std::ostream &StringWeight<Label, S>::Write(std::ostream &strm) const {
-  const int32 size = Size();
-  WriteType(strm, size);
-  for (Iterator iter(*this); !iter.Done(); iter.Next()) {
-    WriteType(strm, iter.Value());
-  }
-  return strm;
-}
-
-template <typename Label, StringType S>
-inline bool StringWeight<Label, S>::Member() const {
-  Iterator iter(*this);
-  return iter.Value() != Label(kStringBad);
-}
-
-template <typename Label, StringType S>
-inline typename StringWeight<Label, S>::ReverseWeight
-StringWeight<Label, S>::Reverse() const {
-  ReverseWeight rweight;
-  for (Iterator iter(*this); !iter.Done(); iter.Next()) {
-    rweight.PushFront(iter.Value());
-  }
-  return rweight;
-}
-
-template <typename Label, StringType S>
-inline size_t StringWeight<Label, S>::Hash() const {
-  size_t h = 0;
-  for (Iterator iter(*this); !iter.Done(); iter.Next()) {
-    h ^= h << 1 ^ iter.Value();
-  }
-  return h;
-}
-
-template <typename Label, StringType S>
-inline bool operator==(const StringWeight<Label, S> &w1,
-                       const StringWeight<Label, S> &w2) {
-  if (w1.Size() != w2.Size()) return false;
-  using Iterator = typename StringWeight<Label, S>::Iterator;
-  Iterator iter1(w1);
-  Iterator iter2(w2);
-  for (; !iter1.Done(); iter1.Next(), iter2.Next()) {
-    if (iter1.Value() != iter2.Value()) return false;
-  }
-  return true;
-}
-
-template <typename Label, StringType S>
-inline bool operator!=(const StringWeight<Label, S> &w1,
-                       const StringWeight<Label, S> &w2) {
-  return !(w1 == w2);
-}
-
-template <typename Label, StringType S>
-inline bool ApproxEqual(const StringWeight<Label, S> &w1,
-                        const StringWeight<Label, S> &w2,
-                        float delta = kDelta) {
-  return w1 == w2;
-}
-
-template <typename Label, StringType S>
-inline std::ostream &operator<<(std::ostream &strm,
-                                const StringWeight<Label, S> &weight) {
-  typename StringWeight<Label, S>::Iterator iter(weight);
-  if (iter.Done()) {
-    return strm << "Epsilon";
-  } else if (iter.Value() == Label(kStringInfinity)) {
-    return strm << "Infinity";
-  } else if (iter.Value() == Label(kStringBad)) {
-    return strm << "BadString";
-  } else {
-    for (size_t i = 0; !iter.Done(); ++i, iter.Next()) {
-      if (i > 0) strm << kStringSeparator;
-      strm << iter.Value();
-    }
-  }
-  return strm;
-}
-
-template <typename Label, StringType S>
-inline std::istream &operator>>(std::istream &strm,
-                                StringWeight<Label, S> &weight) {
-  string str;
-  strm >> str;
-  using Weight = StringWeight<Label, S>;
-  if (str == "Infinity") {
-    weight = Weight::Zero();
-  } else if (str == "Epsilon") {
-    weight = Weight::One();
-  } else {
-    weight.Clear();
-    char *p = nullptr;
-    for (const char *cs = str.c_str(); !p || *p != '\0'; cs = p + 1) {
-      const Label label = strtoll(cs, &p, 10);
-      if (p == cs || (*p != 0 && *p != kStringSeparator)) {
-        strm.clear(std::ios::badbit);
-        break;
-      }
-      weight.PushBack(label);
-    }
-  }
-  return strm;
-}
-
-// Default is for the restricted string semiring. String equality is required
-// (for non-Zero() input). The restriction is used (e.g., in determinization)
-// to ensure the input is functional.
-template <typename Label, StringType S>
-inline StringWeight<Label, S> Plus(const StringWeight<Label, S> &w1,
-                                   const StringWeight<Label, S> &w2) {
-  using Weight = StringWeight<Label, S>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::Zero()) return w2;
-  if (w2 == Weight::Zero()) return w1;
-  if (w1 != w2) {
-    FSTERROR() << "StringWeight::Plus: Unequal arguments "
-               << "(non-functional FST?)"
-               << " w1 = " << w1 << " w2 = " << w2;
-    return Weight::NoWeight();
-  }
-  return w1;
-}
-
-// Longest common prefix for left string semiring.
-template <typename Label>
-inline StringWeight<Label, STRING_LEFT> Plus(
-    const StringWeight<Label, STRING_LEFT> &w1,
-    const StringWeight<Label, STRING_LEFT> &w2) {
-  using Weight = StringWeight<Label, STRING_LEFT>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::Zero()) return w2;
-  if (w2 == Weight::Zero()) return w1;
-  Weight sum;
-  typename Weight::Iterator iter1(w1);
-  typename Weight::Iterator iter2(w2);
-  for (; !iter1.Done() && !iter2.Done() && iter1.Value() == iter2.Value();
-       iter1.Next(), iter2.Next()) {
-    sum.PushBack(iter1.Value());
-  }
-  return sum;
-}
-
-// Longest common suffix for right string semiring.
-template <typename Label>
-inline StringWeight<Label, STRING_RIGHT> Plus(
-    const StringWeight<Label, STRING_RIGHT> &w1,
-    const StringWeight<Label, STRING_RIGHT> &w2) {
-  using Weight = StringWeight<Label, STRING_RIGHT>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::Zero()) return w2;
-  if (w2 == Weight::Zero()) return w1;
-  Weight sum;
-  typename Weight::ReverseIterator iter1(w1);
-  typename Weight::ReverseIterator iter2(w2);
-  for (; !iter1.Done() && !iter2.Done() && iter1.Value() == iter2.Value();
-       iter1.Next(), iter2.Next()) {
-    sum.PushFront(iter1.Value());
-  }
-  return sum;
-}
-
-template <typename Label, StringType S>
-inline StringWeight<Label, S> Times(const StringWeight<Label, S> &w1,
-                                    const StringWeight<Label, S> &w2) {
-  using Weight = StringWeight<Label, S>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w1 == Weight::Zero() || w2 == Weight::Zero()) return Weight::Zero();
-  Weight product(w1);
-  for (typename Weight::Iterator iter(w2); !iter.Done(); iter.Next()) {
-    product.PushBack(iter.Value());
-  }
-  return product;
-}
-
-// Left division in a left string semiring.
-template <typename Label, StringType S>
-inline StringWeight<Label, S> DivideLeft(const StringWeight<Label, S> &w1,
-                                         const StringWeight<Label, S> &w2) {
-  using Weight = StringWeight<Label, S>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w2 == Weight::Zero()) {
-    return Weight(Label(kStringBad));
-  } else if (w1 == Weight::Zero()) {
-    return Weight::Zero();
-  }
-  Weight result;
-  typename Weight::Iterator iter(w1);
-  size_t i = 0;
-  for (; !iter.Done() && i < w2.Size(); iter.Next(), ++i) {
-  }
-  for (; !iter.Done(); iter.Next()) result.PushBack(iter.Value());
-  return result;
-}
-
-// Right division in a right string semiring.
-template <typename Label, StringType S>
-inline StringWeight<Label, S> DivideRight(const StringWeight<Label, S> &w1,
-                                          const StringWeight<Label, S> &w2) {
-  using Weight = StringWeight<Label, S>;
-  if (!w1.Member() || !w2.Member()) return Weight::NoWeight();
-  if (w2 == Weight::Zero()) {
-    return Weight(Label(kStringBad));
-  } else if (w1 == Weight::Zero()) {
-    return Weight::Zero();
-  }
-  Weight result;
-  typename Weight::ReverseIterator iter(w1);
-  size_t i = 0;
-  for (; !iter.Done() && i < w2.Size(); iter.Next(), ++i) {
-  }
-  for (; !iter.Done(); iter.Next()) result.PushFront(iter.Value());
-  return result;
-}
-
-// Default is the restricted string semiring.
-template <typename Label, StringType S>
-inline StringWeight<Label, S> Divide(const StringWeight<Label, S> &w1,
-                                     const StringWeight<Label, S> &w2,
-                                     DivideType divide_type) {
-  using Weight = StringWeight<Label, S>;
-  if (divide_type == DIVIDE_LEFT) {
-    return DivideLeft(w1, w2);
-  } else if (divide_type == DIVIDE_RIGHT) {
-    return DivideRight(w1, w2);
-  } else {
-    FSTERROR() << "StringWeight::Divide: "
-               << "Only explicit left or right division is defined "
-               << "for the " << Weight::Type() << " semiring";
-    return Weight::NoWeight();
-  }
-}
-
-// Left division in the left string semiring.
-template <typename Label>
-inline StringWeight<Label, STRING_LEFT> Divide(
-    const StringWeight<Label, STRING_LEFT> &w1,
-    const StringWeight<Label, STRING_LEFT> &w2, DivideType divide_type) {
-  if (divide_type != DIVIDE_LEFT) {
-    FSTERROR() << "StringWeight::Divide: Only left division is defined "
-               << "for the left string semiring";
-    return StringWeight<Label, STRING_LEFT>::NoWeight();
-  }
-  return DivideLeft(w1, w2);
-}
-
-// Right division in the right string semiring.
-template <typename Label>
-inline StringWeight<Label, STRING_RIGHT> Divide(
-    const StringWeight<Label, STRING_RIGHT> &w1,
-    const StringWeight<Label, STRING_RIGHT> &w2, DivideType divide_type) {
-  if (divide_type != DIVIDE_RIGHT) {
-    FSTERROR() << "StringWeight::Divide: Only right division is defined "
-               << "for the right string semiring";
-    return StringWeight<Label, STRING_RIGHT>::NoWeight();
-  }
-  return DivideRight(w1, w2);
-}
-
-// This function object generates StringWeights that are random integer strings
-// from {1, ... , alphabet_size)^{0, max_string_length} U { Zero }. This is
-// intended primarily for testing.
-template <class Label, StringType S>
-class WeightGenerate<StringWeight<Label, S>> {
- public:
-  using Weight = StringWeight<Label, S>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t alphabet_size = kNumRandomWeights,
-                          size_t max_string_length = kNumRandomWeights)
-      : allow_zero_(allow_zero),
-        alphabet_size_(alphabet_size),
-        max_string_length_(max_string_length) {}
-
-  Weight operator()() const {
-    size_t n = rand() % (max_string_length_ + allow_zero_);  // NOLINT
-    if (allow_zero_ && n == max_string_length_) return Weight::Zero();
-    std::vector<Label> labels;
-    labels.reserve(n);
-    for (size_t i = 0; i < n; ++i) {
-      labels.push_back(rand() % alphabet_size_ + 1);  // NOLINT
-    }
-    return Weight(labels.begin(), labels.end());
-  }
-
- private:
-  // Permits Zero() and zero divisors.
-  const bool allow_zero_;
-  // Alphabet size for random weights.
-  const size_t alphabet_size_;
-  // Number of alternative random weights.
-  const size_t max_string_length_;
-};
-
-// Determines whether to use left, right, or (general) gallic semiring. Includes
-// a restricted version that signals an error if proper string prefixes or
-// suffixes would otherwise be returned by string Plus. This is useful with
-// algorithms that require functional transducer input. Also includes min
-// version that changes the Plus to keep only the lowest W weight string.
-enum GallicType {
-  GALLIC_LEFT = 0,
-  GALLIC_RIGHT = 1,
-  GALLIC_RESTRICT = 2,
-  GALLIC_MIN = 3,
-  GALLIC = 4
-};
-
-constexpr StringType GallicStringType(GallicType g) {
-  return g == GALLIC_LEFT
-             ? STRING_LEFT
-             : (g == GALLIC_RIGHT ? STRING_RIGHT : STRING_RESTRICT);
-}
-
-constexpr GallicType ReverseGallicType(GallicType g) {
-  return g == GALLIC_LEFT
-             ? GALLIC_RIGHT
-             : (g == GALLIC_RIGHT
-                    ? GALLIC_LEFT
-                    : (g == GALLIC_RESTRICT
-                           ? GALLIC_RESTRICT
-                           : (g == GALLIC_MIN ? GALLIC_MIN : GALLIC)));
-}
-
-// Product of string weight and an arbitraryy weight.
-template <class Label, class W, GallicType G = GALLIC_LEFT>
-struct GallicWeight
-    : public ProductWeight<StringWeight<Label, GallicStringType(G)>, W> {
-  using ReverseWeight =
-      GallicWeight<Label, typename W::ReverseWeight, ReverseGallicType(G)>;
-  using SW = StringWeight<Label, GallicStringType(G)>;
-
-  using ProductWeight<SW, W>::Properties;
-
-  GallicWeight() {}
-
-  GallicWeight(SW w1, W w2) : ProductWeight<SW, W>(w1, w2) {}
-
-  explicit GallicWeight(const string &s, int *nread = nullptr)
-      : ProductWeight<SW, W>(s, nread) {}
-
-  explicit GallicWeight(const ProductWeight<SW, W> &w)
-      : ProductWeight<SW, W>(w) {}
-
-  static const GallicWeight &Zero() {
-    static const GallicWeight zero(ProductWeight<SW, W>::Zero());
-    return zero;
-  }
-
-  static const GallicWeight &One() {
-    static const GallicWeight one(ProductWeight<SW, W>::One());
-    return one;
-  }
-
-  static const GallicWeight &NoWeight() {
-    static const GallicWeight no_weight(ProductWeight<SW, W>::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string(
-        G == GALLIC_LEFT
-            ? "left_gallic"
-            : (G == GALLIC_RIGHT
-                   ? "right_gallic"
-                   : (G == GALLIC_RESTRICT
-                          ? "restricted_gallic"
-                          : (G == GALLIC_MIN ? "min_gallic" : "gallic"))));
-    return *type;
-  }
-
-  GallicWeight Quantize(float delta = kDelta) const {
-    return GallicWeight(ProductWeight<SW, W>::Quantize(delta));
-  }
-
-  ReverseWeight Reverse() const {
-    return ReverseWeight(ProductWeight<SW, W>::Reverse());
-  }
-};
-
-// Default plus.
-template <class Label, class W, GallicType G>
-inline GallicWeight<Label, W, G> Plus(const GallicWeight<Label, W, G> &w,
-                                      const GallicWeight<Label, W, G> &v) {
-  return GallicWeight<Label, W, G>(Plus(w.Value1(), v.Value1()),
-                                   Plus(w.Value2(), v.Value2()));
-}
-
-// Min gallic plus.
-template <class Label, class W>
-inline GallicWeight<Label, W, GALLIC_MIN> Plus(
-    const GallicWeight<Label, W, GALLIC_MIN> &w1,
-    const GallicWeight<Label, W, GALLIC_MIN> &w2) {
-  static const NaturalLess<W> less;
-  return less(w1.Value2(), w2.Value2()) ? w1 : w2;
-}
-
-template <class Label, class W, GallicType G>
-inline GallicWeight<Label, W, G> Times(const GallicWeight<Label, W, G> &w,
-                                       const GallicWeight<Label, W, G> &v) {
-  return GallicWeight<Label, W, G>(Times(w.Value1(), v.Value1()),
-                                   Times(w.Value2(), v.Value2()));
-}
-
-template <class Label, class W, GallicType G>
-inline GallicWeight<Label, W, G> Divide(const GallicWeight<Label, W, G> &w,
-                                        const GallicWeight<Label, W, G> &v,
-                                        DivideType divide_type = DIVIDE_ANY) {
-  return GallicWeight<Label, W, G>(Divide(w.Value1(), v.Value1(), divide_type),
-                                   Divide(w.Value2(), v.Value2(), divide_type));
-}
-
-// This function object generates gallic weights by calling an underlying
-// product weight generator. This is intended primarily for testing.
-template <class Label, class W, GallicType G>
-class WeightGenerate<GallicWeight<Label, W, G>>
-    : public WeightGenerate<
-          ProductWeight<StringWeight<Label, GallicStringType(G)>, W>> {
- public:
-  using Weight = GallicWeight<Label, W, G>;
-  using Generate = WeightGenerate<
-      ProductWeight<StringWeight<Label, GallicStringType(G)>, W>>;
-
-  explicit WeightGenerate(bool allow_zero = true) : generate_(allow_zero) {}
-
-  Weight operator()() const { return Weight(generate_()); }
-
- private:
-  const Generate generate_;
-};
-
-// Union weight options for (general) GALLIC type.
-template <class Label, class W>
-struct GallicUnionWeightOptions {
-  using ReverseOptions = GallicUnionWeightOptions<Label, W>;
-  using GW = GallicWeight<Label, W, GALLIC_RESTRICT>;
-  using SW = StringWeight<Label, GallicStringType(GALLIC_RESTRICT)>;
-  using SI = StringWeightIterator<SW>;
-
-  // Military order.
-  struct Compare {
-    bool operator()(const GW &w1, const GW &w2) const {
-      const SW &s1 = w1.Value1();
-      const SW &s2 = w2.Value1();
-      if (s1.Size() < s2.Size()) return true;
-      if (s1.Size() > s2.Size()) return false;
-      SI iter1(s1);
-      SI iter2(s2);
-      while (!iter1.Done()) {
-        const auto l1 = iter1.Value();
-        const auto l2 = iter2.Value();
-        if (l1 < l2) return true;
-        if (l1 > l2) return false;
-        iter1.Next();
-        iter2.Next();
-      }
-      return false;
-    }
-  };
-
-  // Adds W weights when string part equal.
-  struct Merge {
-    GW operator()(const GW &w1, const GW &w2) const {
-      return GW(w1.Value1(), Plus(w1.Value2(), w2.Value2()));
-    }
-  };
-};
-
-// Specialization for the (general) GALLIC type.
-template <class Label, class W>
-struct GallicWeight<Label, W, GALLIC>
-    : public UnionWeight<GallicWeight<Label, W, GALLIC_RESTRICT>,
-                         GallicUnionWeightOptions<Label, W>> {
-  using GW = GallicWeight<Label, W, GALLIC_RESTRICT>;
-  using SW = StringWeight<Label, GallicStringType(GALLIC_RESTRICT)>;
-  using SI = StringWeightIterator<SW>;
-  using UW = UnionWeight<GW, GallicUnionWeightOptions<Label, W>>;
-  using UI = UnionWeightIterator<GW, GallicUnionWeightOptions<Label, W>>;
-  using ReverseWeight = GallicWeight<Label, W, GALLIC>;
-
-  using UW::Properties;
-
-  GallicWeight() {}
-
-  // Copy constructor.
-  GallicWeight(const UW &weight) : UW(weight) {}  // NOLINT
-
-  // Singleton constructors: create a GALLIC weight containing a single
-  // GALLIC_RESTRICT weight. Takes as argument (1) a GALLIC_RESTRICT weight or
-  // (2) the two components of a GALLIC_RESTRICT weight.
-  explicit GallicWeight(const GW &weight) : UW(weight) {}
-
-  GallicWeight(SW w1, W w2) : UW(GW(w1, w2)) {}
-
-  explicit GallicWeight(const string &str, int *nread = nullptr)
-      : UW(str, nread) {}
-
-  static const GallicWeight<Label, W, GALLIC> &Zero() {
-    static const GallicWeight<Label, W, GALLIC> zero(UW::Zero());
-    return zero;
-  }
-
-  static const GallicWeight<Label, W, GALLIC> &One() {
-    static const GallicWeight<Label, W, GALLIC> one(UW::One());
-    return one;
-  }
-
-  static const GallicWeight<Label, W, GALLIC> &NoWeight() {
-    static const GallicWeight<Label, W, GALLIC> no_weight(UW::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string("gallic");
-    return *type;
-  }
-
-  GallicWeight<Label, W, GALLIC> Quantize(float delta = kDelta) const {
-    return UW::Quantize(delta);
-  }
-
-  ReverseWeight Reverse() const { return UW::Reverse(); }
-};
-
-// (General) gallic plus.
-template <class Label, class W>
-inline GallicWeight<Label, W, GALLIC> Plus(
-    const GallicWeight<Label, W, GALLIC> &w1,
-    const GallicWeight<Label, W, GALLIC> &w2) {
-  using GW = GallicWeight<Label, W, GALLIC_RESTRICT>;
-  using UW = UnionWeight<GW, GallicUnionWeightOptions<Label, W>>;
-  return Plus(static_cast<UW>(w1), static_cast<UW>(w2));
-}
-
-// (General) gallic times.
-template <class Label, class W>
-inline GallicWeight<Label, W, GALLIC> Times(
-    const GallicWeight<Label, W, GALLIC> &w1,
-    const GallicWeight<Label, W, GALLIC> &w2) {
-  using GW = GallicWeight<Label, W, GALLIC_RESTRICT>;
-  using UW = UnionWeight<GW, GallicUnionWeightOptions<Label, W>>;
-  return Times(static_cast<UW>(w1), static_cast<UW>(w2));
-}
-
-// (General) gallic divide.
-template <class Label, class W>
-inline GallicWeight<Label, W, GALLIC> Divide(
-    const GallicWeight<Label, W, GALLIC> &w1,
-    const GallicWeight<Label, W, GALLIC> &w2,
-    DivideType divide_type = DIVIDE_ANY) {
-  using GW = GallicWeight<Label, W, GALLIC_RESTRICT>;
-  using UW = UnionWeight<GW, GallicUnionWeightOptions<Label, W>>;
-  return Divide(static_cast<UW>(w1), static_cast<UW>(w2), divide_type);
-}
-
-// This function object generates gallic weights by calling an underlying
-// union weight generator. This is intended primarily for testing.
-template <class Label, class W>
-class WeightGenerate<GallicWeight<Label, W, GALLIC>>
-    : public WeightGenerate<UnionWeight<GallicWeight<Label, W, GALLIC_RESTRICT>,
-                                        GallicUnionWeightOptions<Label, W>>> {
- public:
-  using Weight = GallicWeight<Label, W, GALLIC>;
-  using Generate =
-      WeightGenerate<UnionWeight<GallicWeight<Label, W, GALLIC_RESTRICT>,
-                                 GallicUnionWeightOptions<Label, W>>>;
-
-  explicit WeightGenerate(bool allow_zero = true) : generate_(allow_zero) {}
-
-  Weight operator()() const { return Weight(generate_()); }
-
- private:
-  const Generate generate_;
-};
-
-}  // namespace fst
-
-#endif  // FST_STRING_WEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/string.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/string.h
deleted file mode 100644
index 9bc8e43a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/string.h
+++ /dev/null
@@ -1,270 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Utilities to convert strings into FSTs.
-
-#ifndef FST_STRING_H_
-#define FST_STRING_H_
-
-#include <memory>
-#include <sstream>
-#include <string>
-#include <vector>
-
-#include <fst/flags.h>
-#include <fst/log.h>
-
-#include <fst/compact-fst.h>
-#include <fst/icu.h>
-#include <fst/mutable-fst.h>
-#include <fst/util.h>
-
-
-DECLARE_string(fst_field_separator);
-
-namespace fst {
-
-// This will eventually replace StringCompiler<Arc>::TokenType and
-// StringPrinter<Arc>::TokenType.
-enum StringTokenType { SYMBOL = 1, BYTE = 2, UTF8 = 3 };
-
-namespace internal {
-
-template <class Label>
-bool ConvertSymbolToLabel(const char *str, const SymbolTable *syms,
-                          Label unknown_label, bool allow_negative,
-                          Label *output) {
-  int64 n;
-  if (syms) {
-    n = syms->Find(str);
-    if ((n == -1) && (unknown_label != kNoLabel)) n = unknown_label;
-    if (n == -1 || (!allow_negative && n < 0)) {
-      VLOG(1) << "ConvertSymbolToLabel: Symbol \"" << str
-              << "\" is not mapped to any integer label, symbol table = "
-              << syms->Name();
-      return false;
-    }
-  } else {
-    char *p;
-    n = strtoll(str, &p, 10);
-    if (p < str + strlen(str) || (!allow_negative && n < 0)) {
-      VLOG(1) << "ConvertSymbolToLabel: Bad label integer "
-              << "= \"" << str << "\"";
-      return false;
-    }
-  }
-  *output = n;
-  return true;
-}
-
-template <class Label>
-bool ConvertStringToLabels(const string &str, StringTokenType token_type,
-                           const SymbolTable *syms, Label unknown_label,
-                           bool allow_negative, std::vector<Label> *labels) {
-  labels->clear();
-  if (token_type == StringTokenType::BYTE) {
-    for (const char c : str) labels->push_back(c);
-  } else if (token_type == StringTokenType::UTF8) {
-    return UTF8StringToLabels(str, labels);
-  } else {
-    std::unique_ptr<char[]> c_str(new char[str.size() + 1]);
-    str.copy(c_str.get(), str.size());
-    c_str[str.size()] = 0;
-    std::vector<char *> vec;
-    const string separator = "\n" + FLAGS_fst_field_separator;
-    SplitToVector(c_str.get(), separator.c_str(), &vec, true);
-    for (const char *c : vec) {
-      Label label;
-      if (!ConvertSymbolToLabel(c, syms, unknown_label, allow_negative,
-                                &label)) {
-        return false;
-      }
-      labels->push_back(label);
-    }
-  }
-  return true;
-}
-
-}  // namespace internal
-
-// Functor for compiling a string in an FST.
-template <class Arc>
-class StringCompiler {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit StringCompiler(StringTokenType token_type,
-                          const SymbolTable *syms = nullptr,
-                          Label unknown_label = kNoLabel,
-                          bool allow_negative = false)
-      : token_type_(token_type),
-        syms_(syms),
-        unknown_label_(unknown_label),
-        allow_negative_(allow_negative) {}
-
-  // Compiles string into an FST.
-  template <class FST>
-  bool operator()(const string &str, FST *fst) const {
-    std::vector<Label> labels;
-    if (!internal::ConvertStringToLabels(str, token_type_, syms_,
-                                         unknown_label_, allow_negative_,
-                                         &labels)) {
-      return false;
-    }
-    Compile(labels, fst);
-    return true;
-  }
-
-  template <class FST>
-  bool operator()(const string &str, FST *fst, Weight weight) const {
-    std::vector<Label> labels;
-    if (!internal::ConvertStringToLabels(str, token_type_, syms_,
-                                         unknown_label_, allow_negative_,
-                                         &labels)) {
-      return false;
-    }
-    Compile(labels, fst, std::move(weight));
-    return true;
-  }
-
- private:
-  void Compile(const std::vector<Label> &labels, MutableFst<Arc> *fst,
-               Weight weight = Weight::One()) const {
-    fst->DeleteStates();
-    while (fst->NumStates() <= labels.size()) fst->AddState();
-    for (StateId i = 0; i < labels.size(); ++i) {
-      fst->AddArc(i, Arc(labels[i], labels[i], Weight::One(), i + 1));
-    }
-    fst->SetStart(0);
-    fst->SetFinal(labels.size(), std::move(weight));
-  }
-
-  template <class Unsigned>
-  void Compile(const std::vector<Label> &labels,
-               CompactStringFst<Arc, Unsigned> *fst) const {
-    fst->SetCompactElements(labels.begin(), labels.end());
-  }
-
-  template <class Unsigned>
-  void Compile(const std::vector<Label> &labels,
-               CompactWeightedStringFst<Arc, Unsigned> *fst,
-               const Weight &weight = Weight::One()) const {
-    std::vector<std::pair<Label, Weight>> compacts;
-    compacts.reserve(labels.size() + 1);
-    for (StateId i = 0; i < static_cast<StateId>(labels.size()) - 1; ++i) {
-      compacts.emplace_back(labels[i], Weight::One());
-    }
-    compacts.emplace_back(!labels.empty() ? labels.back() : kNoLabel, weight);
-    fst->SetCompactElements(compacts.begin(), compacts.end());
-  }
-
-  const StringTokenType token_type_;
-  const SymbolTable *syms_;    // Symbol table (used when token type is symbol).
-  const Label unknown_label_;  // Label for token missing from symbol table.
-  const bool allow_negative_;  // Negative labels allowed?
-
-  StringCompiler(const StringCompiler &) = delete;
-  StringCompiler &operator=(const StringCompiler &) = delete;
-};
-
-// Functor for printing a string FST as a string.
-template <class Arc>
-class StringPrinter {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  explicit StringPrinter(StringTokenType token_type,
-                         const SymbolTable *syms = nullptr)
-      : token_type_(token_type), syms_(syms) {}
-
-  // Converts the FST into a string.
-  bool operator()(const Fst<Arc> &fst, string *result) {
-    if (!FstToLabels(fst)) {
-      VLOG(1) << "StringPrinter::operator(): FST is not a string";
-      return false;
-    }
-    result->clear();
-    if (token_type_ == StringTokenType::SYMBOL) {
-      std::stringstream sstrm;
-      for (size_t i = 0; i < labels_.size(); ++i) {
-        if (i) sstrm << *(FLAGS_fst_field_separator.rbegin());
-        if (!PrintLabel(labels_[i], sstrm)) return false;
-      }
-      *result = sstrm.str();
-    } else if (token_type_ == StringTokenType::BYTE) {
-      result->reserve(labels_.size());
-      for (size_t i = 0; i < labels_.size(); ++i) result->push_back(labels_[i]);
-    } else if (token_type_ == StringTokenType::UTF8) {
-      return LabelsToUTF8String(labels_, result);
-    } else {
-      VLOG(1) << "StringPrinter::operator(): Unknown token type: "
-              << token_type_;
-      return false;
-    }
-    return true;
-  }
-
- private:
-  bool FstToLabels(const Fst<Arc> &fst) {
-    labels_.clear();
-    auto s = fst.Start();
-    if (s == kNoStateId) {
-      VLOG(2) << "StringPrinter::FstToLabels: Invalid starting state for "
-              << "string FST";
-      return false;
-    }
-    while (fst.Final(s) == Weight::Zero()) {
-      ArcIterator<Fst<Arc>> aiter(fst, s);
-      if (aiter.Done()) {
-        VLOG(2) << "StringPrinter::FstToLabels: String FST traversal does "
-                << "not reach final state";
-        return false;
-      }
-      const auto &arc = aiter.Value();
-      labels_.push_back(arc.olabel);
-      s = arc.nextstate;
-      if (s == kNoStateId) {
-        VLOG(2) << "StringPrinter::FstToLabels: Transition to invalid state";
-        return false;
-      }
-      aiter.Next();
-      if (!aiter.Done()) {
-        VLOG(2) << "StringPrinter::FstToLabels: State with multiple "
-                << "outgoing arcs found";
-        return false;
-      }
-    }
-    return true;
-  }
-
-  bool PrintLabel(Label label, std::ostream &ostrm) {
-    if (syms_) {
-      const auto symbol = syms_->Find(label);
-      if (symbol == "") {
-        VLOG(2) << "StringPrinter::PrintLabel: Integer " << label << " is not "
-                << "mapped to any textual symbol, symbol table = "
-                << syms_->Name();
-        return false;
-      }
-      ostrm << symbol;
-    } else {
-      ostrm << label;
-    }
-    return true;
-  }
-
-  const StringTokenType token_type_;
-  const SymbolTable *syms_;    // Symbol table (used when token type is symbol).
-  std::vector<Label> labels_;  // Input FST labels.
-
-  StringPrinter(const StringPrinter &) = delete;
-  StringPrinter &operator=(const StringPrinter &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_STRING_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/symbol-table-ops.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/symbol-table-ops.h
deleted file mode 100644
index 38f9d819..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/symbol-table-ops.h
+++ /dev/null
@@ -1,76 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-
-#ifndef FST_SYMBOL_TABLE_OPS_H_
-#define FST_SYMBOL_TABLE_OPS_H_
-
-#include <string>
-#include <unordered_set>
-#include <vector>
-
-
-#include <fst/fst.h>
-#include <fst/symbol-table.h>
-
-
-namespace fst {
-
-// Returns a minimal symbol table containing only symbols referenced by the
-// passed fst.  Symbols preserve their original numbering, so fst does not
-// require relabeling.
-template <class Arc>
-SymbolTable *PruneSymbolTable(const Fst<Arc> &fst, const SymbolTable &syms,
-                              bool input) {
-  std::unordered_set<typename Arc::Label> seen;
-  seen.insert(0);  // Always keep epsilon.
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    for (ArcIterator<Fst<Arc>> aiter(fst, siter.Value()); !aiter.Done();
-         aiter.Next()) {
-      const auto sym = (input) ? aiter.Value().ilabel : aiter.Value().olabel;
-      seen.insert(sym);
-    }
-  }
-  auto *pruned = new SymbolTable(syms.Name() + "_pruned");
-  for (SymbolTableIterator stiter(syms); !stiter.Done(); stiter.Next()) {
-    const auto label = stiter.Value();
-    if (seen.count(label)) pruned->AddSymbol(stiter.Symbol(), label);
-  }
-  return pruned;
-}
-
-// Relabels a symbol table to make it a contiguous mapping.
-SymbolTable *CompactSymbolTable(const SymbolTable &syms);
-
-// Merges two SymbolTables, all symbols from left will be merged into right
-// with the same ids.  Symbols in right that have conflicting ids with those
-// in left will be assigned to value assigned from the left SymbolTable.
-// The returned symbol table will never modify symbol assignments from the left
-// side, but may do so on the right.  If right_relabel_output is non-NULL, it
-// will be assigned true if the symbols from the right table needed to be
-// reassigned.
-// A potential use case is to Compose two Fst's that have different symbol
-// tables.  You can reconcile them in the following way:
-//   Fst<Arc> a, b;
-//   bool relabel;
-//   std::unique_ptr<SymbolTable> bnew(MergeSymbolTable(a.OutputSymbols(),
-//                                     b.InputSymbols(), &relabel);
-//   if (relabel) {
-//     Relabel(b, bnew.get(), nullptr);
-//   }
-//   b.SetInputSymbols(bnew);
-SymbolTable *MergeSymbolTable(const SymbolTable &left, const SymbolTable &right,
-                              bool *right_relabel_output = nullptr);
-
-// Read the symbol table from any Fst::Read()able file, without loading the
-// corresponding Fst.  Returns nullptr if the Fst does not contain a symbol
-// table or the symbol table cannot be read.
-SymbolTable *FstReadSymbols(const string &filename, bool input);
-
-// Adds a contiguous range of symbols to a symbol table using a simple prefix
-// for the string, returning false if the inserted symbol string clashes with
-// any currently present.
-bool AddAuxiliarySymbols(const string &prefix, int64 start_label,
-                         int64 nlabels, SymbolTable *syms);
-
-}  // namespace fst
-#endif  // FST_SYMBOL_TABLE_OPS_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/symbol-table.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/symbol-table.h
deleted file mode 100644
index fed5806f..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/symbol-table.h
+++ /dev/null
@@ -1,466 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Classes to provide symbol-to-integer and integer-to-symbol mappings.
-
-#ifndef FST_SYMBOL_TABLE_H_
-#define FST_SYMBOL_TABLE_H_
-
-#include <cstring>
-#include <functional>
-#include <ios>
-#include <iostream>
-#include <memory>
-#include <sstream>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/flags.h>
-#include <fst/log.h>
-#include <fstream>
-#include <map>
-
-DECLARE_bool(fst_compat_symbols);
-
-namespace fst {
-
-// WARNING: Reading via symbol table read options should
-//          not be used. This is a temporary work around for
-//          reading symbol ranges of previously stored symbol sets.
-struct SymbolTableReadOptions {
-  SymbolTableReadOptions() {}
-
-  SymbolTableReadOptions(
-      std::vector<std::pair<int64, int64>> string_hash_ranges,
-      const string &source)
-      : string_hash_ranges(std::move(string_hash_ranges)), source(source) {}
-
-  std::vector<std::pair<int64, int64>> string_hash_ranges;
-  string source;
-};
-
-struct SymbolTableTextOptions {
-  explicit SymbolTableTextOptions(bool allow_negative_labels = false);
-
-  bool allow_negative_labels;
-  string fst_field_separator;
-};
-
-namespace internal {
-
-// List of symbols with a dense hash for looking up symbol index.
-// Hash uses linear probe, rehashes at 0.75% occupancy, avg 6 bytes overhead
-// per entry.  Rehash in place from symbol list.
-//
-// Symbols are stored as c strings to avoid adding memory overhead, but the
-// performance penalty for this is high because rehash must call strlen on
-// every symbol.  AddSymbol can be another 2x faster if symbol lengths were
-// stored.
-class DenseSymbolMap {
- public:
-  DenseSymbolMap();
-
-  DenseSymbolMap(const DenseSymbolMap &x);
-
-  ~DenseSymbolMap();
-
-  std::pair<int64, bool> InsertOrFind(const string &key);
-
-  int64 Find(const string &key) const;
-
-  const size_t size() const { return symbols_.size(); }
-
-  const string GetSymbol(size_t idx) const {
-    return string(symbols_[idx], strlen(symbols_[idx]));
-  }
-
-  void RemoveSymbol(size_t idx);
-
- private:
-  // num_buckets must be power of 2.
-  void Rehash(size_t num_buckets);
-
-  const char* NewSymbol(const string &sym);
-
-  int64 empty_;
-  std::vector<const char *> symbols_;
-  std::hash<string> str_hash_;
-  std::vector<int64> buckets_;
-  uint64 hash_mask_;
-};
-
-class SymbolTableImpl {
- public:
-  explicit SymbolTableImpl(const string &name)
-      : name_(name),
-        available_key_(0),
-        dense_key_limit_(0),
-        check_sum_finalized_(false) {}
-
-  SymbolTableImpl(const SymbolTableImpl &impl)
-      : name_(impl.name_),
-        available_key_(impl.available_key_),
-        dense_key_limit_(impl.dense_key_limit_),
-        symbols_(impl.symbols_),
-        idx_key_(impl.idx_key_),
-        key_map_(impl.key_map_),
-        check_sum_finalized_(false) {}
-
-  int64 AddSymbol(const string &symbol, int64 key);
-
-  int64 AddSymbol(const string &symbol) {
-    return AddSymbol(symbol, available_key_);
-  }
-
-  // Removes the symbol with the given key. The removal is costly
-  // (O(NumSymbols)) and may reduce the efficiency of Find() because of a
-  // potentially reduced size of the dense key interval.
-  void RemoveSymbol(int64 key);
-
-  static SymbolTableImpl *ReadText(
-      std::istream &strm, const string &name,
-      const SymbolTableTextOptions &opts = SymbolTableTextOptions());
-
-  static SymbolTableImpl* Read(std::istream &strm,
-                               const SymbolTableReadOptions &opts);
-
-  bool Write(std::ostream &strm) const;
-
-  // Return the string associated with the key. If the key is out of
-  // range (<0, >max), return an empty string.
-  string Find(int64 key) const {
-    int64 idx = key;
-    if (key < 0 || key >= dense_key_limit_) {
-      const auto it = key_map_.find(key);
-      if (it == key_map_.end()) return "";
-      idx = it->second;
-    }
-    if (idx < 0 || idx >= symbols_.size()) return "";
-    return symbols_.GetSymbol(idx);
-  }
-
-  // Returns the key associated with the symbol. If the symbol
-  // does not exists, return SymbolTable::kNoSymbol.
-  int64 Find(const string &symbol) const {
-    int64 idx = symbols_.Find(symbol);
-    if (idx == -1 || idx < dense_key_limit_) return idx;
-    return idx_key_[idx - dense_key_limit_];
-  }
-
-  bool Member(int64 key) const { return !Find(key).empty(); }
-
-  bool Member(const string &symbol) const { return Find(symbol) != -1; }
-
-  int64 GetNthKey(ssize_t pos) const {
-    if (pos < 0 || pos >= symbols_.size()) return -1;
-    if (pos < dense_key_limit_) return pos;
-    return Find(symbols_.GetSymbol(pos));
-  }
-
-  const string &Name() const { return name_; }
-
-  void SetName(const string &new_name) { name_ = new_name; }
-
-  string CheckSum() const {
-    MaybeRecomputeCheckSum();
-    return check_sum_string_;
-  }
-
-  string LabeledCheckSum() const {
-    MaybeRecomputeCheckSum();
-    return labeled_check_sum_string_;
-  }
-
-  int64 AvailableKey() const { return available_key_; }
-
-  size_t NumSymbols() const { return symbols_.size(); }
-
- private:
-  // Recomputes the checksums (both of them) if we've had changes since the last
-  // computation (i.e., if check_sum_finalized_ is false).
-  // Takes ~2.5 microseconds (dbg) or ~230 nanoseconds (opt) on a 2.67GHz Xeon
-  // if the checksum is up-to-date (requiring no recomputation).
-  void MaybeRecomputeCheckSum() const;
-
-  string name_;
-  int64 available_key_;
-  int64 dense_key_limit_;
-
-  DenseSymbolMap symbols_;
-  // Maps index to key for index >= dense_key_limit:
-  //   key = idx_key_[index - dense_key_limit]
-  std::vector<int64> idx_key_;
-  // Maps key to index for key >= dense_key_limit_.
-  //  index = key_map_[key]
-  map<int64, int64> key_map_;
-
-  mutable bool check_sum_finalized_;
-  mutable string check_sum_string_;
-  mutable string labeled_check_sum_string_;
-  mutable Mutex check_sum_mutex_;
-};
-
-}  // namespace internal
-
-// Symbol (string) to integer (and reverse) mapping.
-//
-// The SymbolTable implements the mappings of labels to strings and reverse.
-// SymbolTables are used to describe the alphabet of the input and output
-// labels for arcs in a Finite State Transducer.
-//
-// SymbolTables are reference-counted and can therefore be shared across
-// multiple machines. For example a language model grammar G, with a
-// SymbolTable for the words in the language model can share this symbol
-// table with the lexical representation L o G.
-class SymbolTable {
- public:
-  static constexpr int64 kNoSymbol = -1;
-
-  // Constructs symbol table with an unspecified name.
-  SymbolTable() :
-      impl_(std::make_shared<internal::SymbolTableImpl>("<unspecified>")) {}
-
-  // Constructs symbol table with a unique name.
-  explicit SymbolTable(const string &name)
-      : impl_(std::make_shared<internal::SymbolTableImpl>(name)) {}
-
-  virtual ~SymbolTable() {}
-
-  // Reads a text representation of the symbol table from an istream. Pass a
-  // name to give the resulting SymbolTable.
-  static SymbolTable* ReadText(std::istream &strm,
-      const string &name,
-      const SymbolTableTextOptions &opts = SymbolTableTextOptions()) {
-    auto *impl = internal::SymbolTableImpl::ReadText(strm, name, opts);
-    if (!impl) {
-      return nullptr;
-    } else {
-      return new SymbolTable(impl);
-    }
-  }
-
-  // Reads a text representation of the symbol table.
-  static SymbolTable *ReadText(const string &filename,
-      const SymbolTableTextOptions &opts = SymbolTableTextOptions()) {
-    std::ifstream strm(filename, std::ios_base::in);
-    if (!strm.good()) {
-      LOG(ERROR) << "SymbolTable::ReadText: Can't open file " << filename;
-      return nullptr;
-    }
-    return ReadText(strm, filename, opts);
-  }
-
-  // WARNING: Reading via symbol table read options should not be used. This is
-  // a temporary work-around.
-  static SymbolTable* Read(std::istream &strm,
-                           const SymbolTableReadOptions &opts) {
-    auto *impl = internal::SymbolTableImpl::Read(strm, opts);
-    return (impl) ? new SymbolTable(impl) : nullptr;
-  }
-
-  // Reads a binary dump of the symbol table from a stream.
-  static SymbolTable *Read(std::istream &strm,
-                           const string &source) {
-    SymbolTableReadOptions opts;
-    opts.source = source;
-    return Read(strm, opts);
-  }
-
-  // Reads a binary dump of the symbol table.
-  static SymbolTable *Read(const string& filename) {
-    std::ifstream strm(filename,
-                            std::ios_base::in | std::ios_base::binary);
-    if (!strm.good()) {
-      LOG(ERROR) << "SymbolTable::Read: Can't open file " << filename;
-      return nullptr;
-    }
-    return Read(strm, filename);
-  }
-
-  //--------------------------------------------------------
-  // Derivable Interface (final)
-  //--------------------------------------------------------
-  // Creates a reference counted copy.
-  virtual SymbolTable *Copy() const { return new SymbolTable(*this); }
-
-  // Adds a symbol with given key to table. A symbol table also keeps track of
-  // the last available key (highest key value in the symbol table).
-  virtual int64 AddSymbol(const string &symbol, int64 key) {
-    MutateCheck();
-    return impl_->AddSymbol(symbol, key);
-  }
-
-  // Adds a symbol to the table. The associated value key is automatically
-  // assigned by the symbol table.
-  virtual int64 AddSymbol(const string &symbol) {
-    MutateCheck();
-    return impl_->AddSymbol(symbol);
-  }
-
-  // Adds another symbol table to this table. All key values will be offset
-  // by the current available key (highest key value in the symbol table).
-  // Note string symbols with the same key value with still have the same
-  // key value after the symbol table has been merged, but a different
-  // value. Adding symbol tables do not result in changes in the base table.
-  virtual void AddTable(const SymbolTable &table);
-
-  virtual void RemoveSymbol(int64 key) {
-    MutateCheck();
-    return impl_->RemoveSymbol(key);
-  }
-
-  // Returns the name of the symbol table.
-  virtual const string &Name() const { return impl_->Name(); }
-
-  // Sets the name of the symbol table.
-  virtual void SetName(const string &new_name) {
-    MutateCheck();
-    impl_->SetName(new_name);
-  }
-
-  // Return the label-agnostic MD5 check-sum for this table.  All new symbols
-  // added to the table will result in an updated checksum. Deprecated.
-  virtual string CheckSum() const { return impl_->CheckSum(); }
-
-  // Same as CheckSum(), but returns an label-dependent version.
-  virtual string LabeledCheckSum() const { return impl_->LabeledCheckSum(); }
-
-  virtual bool Write(std::ostream &strm) const { return impl_->Write(strm); }
-
-  bool Write(const string &filename) const {
-    std::ofstream strm(filename,
-                             std::ios_base::out | std::ios_base::binary);
-    if (!strm.good()) {
-      LOG(ERROR) << "SymbolTable::Write: Can't open file " << filename;
-      return false;
-    }
-    return Write(strm);
-  }
-
-  // Dump a text representation of the symbol table via a stream.
-  virtual bool WriteText(std::ostream &strm,
-      const SymbolTableTextOptions &opts = SymbolTableTextOptions()) const;
-
-  // Dump an text representation of the symbol table.
-  bool WriteText(const string &filename) const {
-    std::ofstream strm(filename);
-    if (!strm.good()) {
-      LOG(ERROR) << "SymbolTable::WriteText: Can't open file " << filename;
-      return false;
-    }
-    return WriteText(strm);
-  }
-
-  // Returns the string associated with the key. If the key is out of
-  // range (<0, >max), returns an empty string.
-  virtual string Find(int64 key) const { return impl_->Find(key); }
-
-  // Returns the key associated with the symbol. If the symbol does not exist,
-  // SymbolTable::kNoSymbol is returned.
-  virtual int64 Find(const string &symbol) const { return impl_->Find(symbol); }
-
-  // Returns the key associated with the symbol. If the symbol does not exist,
-  // SymbolTable::kNoSymbol is returned.
-  virtual int64 Find(const char *symbol) const { return impl_->Find(symbol); }
-
-  virtual bool Member(int64 key) const { return impl_->Member(key); }
-
-  virtual bool Member(const string &symbol) const {
-    return impl_->Member(symbol);
-  }
-
-  // Returns the current available key (i.e., highest key + 1) in the symbol
-  // table.
-  virtual int64 AvailableKey() const { return impl_->AvailableKey(); }
-
-  // Returns the current number of symbols in table (not necessarily equal to
-  // AvailableKey()).
-  virtual size_t NumSymbols() const { return impl_->NumSymbols(); }
-
-  virtual int64 GetNthKey(ssize_t pos) const { return impl_->GetNthKey(pos); }
-
- private:
-  explicit SymbolTable(internal::SymbolTableImpl *impl) : impl_(impl) {}
-
-  void MutateCheck() {
-    if (!impl_.unique()) impl_.reset(new internal::SymbolTableImpl(*impl_));
-  }
-
-  const internal::SymbolTableImpl *Impl() const { return impl_.get(); }
-
- private:
-  std::shared_ptr<internal::SymbolTableImpl> impl_;
-};
-
-// Iterator class for symbols in a symbol table.
-class SymbolTableIterator {
- public:
-  explicit SymbolTableIterator(const SymbolTable &table)
-      : table_(table),
-        pos_(0),
-        nsymbols_(table.NumSymbols()),
-        key_(table.GetNthKey(0)) {}
-
-  ~SymbolTableIterator() {}
-
-  // Returns whether iterator is done.
-  bool Done() const { return (pos_ == nsymbols_); }
-
-  // Return the key of the current symbol.
-  int64 Value() const { return key_; }
-
-  // Return the string of the current symbol.
-  string Symbol() const { return table_.Find(key_); }
-
-  // Advances iterator.
-  void Next() {
-    ++pos_;
-    if (pos_ < nsymbols_) key_ = table_.GetNthKey(pos_);
-  }
-
-  // Resets iterator.
-  void Reset() {
-    pos_ = 0;
-    key_ = table_.GetNthKey(0);
-  }
-
- private:
-  const SymbolTable &table_;
-  ssize_t pos_;
-  size_t nsymbols_;
-  int64 key_;
-};
-
-// Relabels a symbol table as specified by the input vector of pairs
-// (old label, new label). The new symbol table only retains symbols
-// for which a relabeling is explicitly specified.
-//
-// TODO(allauzen): consider adding options to allow for some form of implicit
-// identity relabeling.
-template <class Label>
-SymbolTable *RelabelSymbolTable(const SymbolTable *table,
-    const std::vector<std::pair<Label, Label>> &pairs) {
-  auto new_table = new SymbolTable(table->Name().empty() ?
-      string() : (string("relabeled_") + table->Name()));
-  for (const auto &pair : pairs) {
-    new_table->AddSymbol(table->Find(pair.first), pair.second);
-  }
-  return new_table;
-}
-
-// Returns true if the two symbol tables have equal checksums. Passing in
-// nullptr for either table always returns true.
-bool CompatSymbols(const SymbolTable *syms1, const SymbolTable *syms2,
-                   bool warning = true);
-
-// Symbol Table serialization.
-
-void SymbolTableToString(const SymbolTable *table, string *result);
-
-SymbolTable *StringToSymbolTable(const string &str);
-
-}  // namespace fst
-
-#endif  // FST_SYMBOL_TABLE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/synchronize.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/synchronize.h
deleted file mode 100644
index b2fc3df9..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/synchronize.h
+++ /dev/null
@@ -1,414 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Synchronize an FST with bounded delay.
-
-#ifndef FST_SYNCHRONIZE_H_
-#define FST_SYNCHRONIZE_H_
-
-#include <algorithm>
-#include <string>
-#include <unordered_map>
-#include <unordered_set>
-#include <utility>
-#include <vector>
-
-#include <fst/cache.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-using SynchronizeFstOptions = CacheOptions;
-
-namespace internal {
-
-// Implementation class for SynchronizeFst.
-// TODO(kbg,sorenj): Refactor to guarantee thread-safety.
-
-template <class Arc>
-class SynchronizeFstImpl : public CacheImpl<Arc> {
- public:
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-
-  using CacheBaseImpl<CacheState<Arc>>::PushArc;
-  using CacheBaseImpl<CacheState<Arc>>::HasArcs;
-  using CacheBaseImpl<CacheState<Arc>>::HasFinal;
-  using CacheBaseImpl<CacheState<Arc>>::HasStart;
-  using CacheBaseImpl<CacheState<Arc>>::SetArcs;
-  using CacheBaseImpl<CacheState<Arc>>::SetFinal;
-  using CacheBaseImpl<CacheState<Arc>>::SetStart;
-
-  using String = basic_string<Label>;
-
-  struct Element {
-    Element() {}
-
-    Element(StateId state_, const String *i, const String *o)
-        : state(state_), istring(i), ostring(o) {}
-
-    StateId state;          // Input state ID.
-    const String *istring;  // Residual input labels.
-    const String *ostring;  // Residual output labels.
-    // Residual strings are represented by const pointers to
-    // basic_string<Label> and are stored in a hash_set. The pointed
-    // memory is owned by the hash_set string_set_.
-  };
-
-  SynchronizeFstImpl(const Fst<Arc> &fst, const SynchronizeFstOptions &opts)
-      : CacheImpl<Arc>(opts), fst_(fst.Copy()) {
-    SetType("synchronize");
-    const auto props = fst.Properties(kFstProperties, false);
-    SetProperties(SynchronizeProperties(props), kCopyProperties);
-    SetInputSymbols(fst.InputSymbols());
-    SetOutputSymbols(fst.OutputSymbols());
-  }
-
-  SynchronizeFstImpl(const SynchronizeFstImpl &impl)
-      : CacheImpl<Arc>(impl), fst_(impl.fst_->Copy(true)) {
-    SetType("synchronize");
-    SetProperties(impl.Properties(), kCopyProperties);
-    SetInputSymbols(impl.InputSymbols());
-    SetOutputSymbols(impl.OutputSymbols());
-  }
-
-  ~SynchronizeFstImpl() override {
-    for (const auto *ptr : string_set_) delete ptr;
-  }
-
-  StateId Start() {
-    if (!HasStart()) {
-      auto start = fst_->Start();
-      if (start == kNoStateId) return kNoStateId;
-      const auto *empty = FindString(new String());
-      start = FindState(Element(fst_->Start(), empty, empty));
-      SetStart(start);
-    }
-    return CacheImpl<Arc>::Start();
-  }
-
-  Weight Final(StateId s) {
-    if (!HasFinal(s)) {
-      const auto &element = elements_[s];
-      const auto weight = element.state == kNoStateId
-                              ? Weight::One()
-                              : fst_->Final(element.state);
-      if ((weight != Weight::Zero()) && (element.istring)->empty() &&
-          (element.ostring)->empty()) {
-        SetFinal(s, weight);
-      } else {
-        SetFinal(s, Weight::Zero());
-      }
-    }
-    return CacheImpl<Arc>::Final(s);
-  }
-
-  size_t NumArcs(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumArcs(s);
-  }
-
-  size_t NumInputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumInputEpsilons(s);
-  }
-
-  size_t NumOutputEpsilons(StateId s) {
-    if (!HasArcs(s)) Expand(s);
-    return CacheImpl<Arc>::NumOutputEpsilons(s);
-  }
-
-  uint64 Properties() const override { return Properties(kFstProperties); }
-
-  // Sets error if found, returning other FST impl properties.
-  uint64 Properties(uint64 mask) const override {
-    if ((mask & kError) && fst_->Properties(kError, false)) {
-      SetProperties(kError, kError);
-    }
-    return FstImpl<Arc>::Properties(mask);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) {
-    if (!HasArcs(s)) Expand(s);
-    CacheImpl<Arc>::InitArcIterator(s, data);
-  }
-
-  // Returns the first character of the string obtained by concatenating the
-  // string and the label.
-  Label Car(const String *str, Label label = 0) const {
-    if (!str->empty()) {
-      return (*str)[0];
-    } else {
-      return label;
-    }
-  }
-
-  // Computes the residual string obtained by removing the first
-  // character in the concatenation of the string and the label.
-  const String *Cdr(const String *str, Label label = 0) {
-    auto *r = new String();
-    for (size_t i = 1; i < str->size(); ++i) r->push_back((*str)[i]);
-    if (label && !(str->empty())) r->push_back(label);
-    return FindString(r);
-  }
-
-  // Computes the concatenation of the string and the label.
-  const String *Concat(const String *str, Label label = 0) {
-    auto *r = new String();
-    for (size_t i = 0; i < str->size(); ++i) r->push_back((*str)[i]);
-    if (label) r->push_back(label);
-    return FindString(r);
-  }
-
-  // Tests if the concatenation of the string and label is empty.
-  bool Empty(const String *str, Label label = 0) const {
-    if (str->empty()) {
-      return label == 0;
-    } else {
-      return false;
-    }
-  }
-
-  // Finds the string pointed by s in the hash set. Transfers the pointer
-  // ownership to the hash set.
-  const String *FindString(const String *str) {
-    const auto insert_result = string_set_.insert(str);
-    if (!insert_result.second) {
-      delete str;
-    }
-    return *insert_result.first;
-  }
-
-  // Finds state corresponding to an element. Creates new state if element
-  // is not found.
-  StateId FindState(const Element &element) {
-    const auto insert_result =
-        element_map_.insert(std::make_pair(element, elements_.size()));
-    if (insert_result.second) {
-      elements_.push_back(element);
-    }
-    return insert_result.first->second;
-  }
-
-  // Computes the outgoing transitions from a state, creating new destination
-  // states as needed.
-  void Expand(StateId s) {
-    const auto element = elements_[s];
-    if (element.state != kNoStateId) {
-      for (ArcIterator<Fst<Arc>> aiter(*fst_, element.state); !aiter.Done();
-           aiter.Next()) {
-        const auto &arc = aiter.Value();
-        if (!Empty(element.istring, arc.ilabel) &&
-            !Empty(element.ostring, arc.olabel)) {
-          const auto *istring = Cdr(element.istring, arc.ilabel);
-          const auto *ostring = Cdr(element.ostring, arc.olabel);
-          PushArc(s, Arc(Car(element.istring, arc.ilabel),
-                         Car(element.ostring, arc.olabel), arc.weight,
-                         FindState(Element(arc.nextstate, istring, ostring))));
-        } else {
-          const auto *istring = Concat(element.istring, arc.ilabel);
-          const auto *ostring = Concat(element.ostring, arc.olabel);
-          PushArc(s, Arc(0, 0, arc.weight,
-                         FindState(Element(arc.nextstate, istring, ostring))));
-        }
-      }
-    }
-    const auto weight = element.state == kNoStateId
-                            ? Weight::One()
-                            : fst_->Final(element.state);
-    if ((weight != Weight::Zero()) &&
-        ((element.istring)->size() + (element.ostring)->size() > 0)) {
-      const auto *istring = Cdr(element.istring);
-      const auto *ostring = Cdr(element.ostring);
-      PushArc(s, Arc(Car(element.istring), Car(element.ostring), weight,
-                     FindState(Element(kNoStateId, istring, ostring))));
-    }
-    SetArcs(s);
-  }
-
- private:
-  // Equality function for Elements; assumes strings have been hashed.
-  class ElementEqual {
-   public:
-    bool operator()(const Element &x, const Element &y) const {
-      return x.state == y.state && x.istring == y.istring &&
-             x.ostring == y.ostring;
-    }
-  };
-
-  // Hash function for Elements to FST states.
-  class ElementKey {
-   public:
-    size_t operator()(const Element &x) const {
-      size_t key = x.state;
-      key = (key << 1) ^ (x.istring)->size();
-      for (size_t i = 0; i < (x.istring)->size(); ++i) {
-        key = (key << 1) ^ (*x.istring)[i];
-      }
-      key = (key << 1) ^ (x.ostring)->size();
-      for (size_t i = 0; i < (x.ostring)->size(); ++i) {
-        key = (key << 1) ^ (*x.ostring)[i];
-      }
-      return key;
-    }
-  };
-
-  // Equality function for strings.
-  class StringEqual {
-   public:
-    bool operator()(const String *const &x, const String *const &y) const {
-      if (x->size() != y->size()) return false;
-      for (size_t i = 0; i < x->size(); ++i) {
-        if ((*x)[i] != (*y)[i]) return false;
-      }
-      return true;
-    }
-  };
-
-  // Hash function for set of strings
-  class StringKey {
-   public:
-    size_t operator()(const String *const &x) const {
-      size_t key = x->size();
-      for (size_t i = 0; i < x->size(); ++i) key = (key << 1) ^ (*x)[i];
-      return key;
-    }
-  };
-
-  using ElementMap =
-      std::unordered_map<Element, StateId, ElementKey, ElementEqual>;
-  using StringSet = std::unordered_set<const String *, StringKey, StringEqual>;
-
-  std::unique_ptr<const Fst<Arc>> fst_;
-  std::vector<Element> elements_;  // Maps FST state to Elements.
-  ElementMap element_map_;         // Maps Elements to FST state.
-  StringSet string_set_;
-};
-
-}  // namespace internal
-
-// Synchronizes a transducer. This version is a delayed FST. The result is an
-// equivalent FST that has the property that during the traversal of a path,
-// the delay is either zero or strictly increasing, where the delay is the
-// difference between the number of non-epsilon output labels and input labels
-// along the path.
-//
-// For the algorithm to terminate, the input transducer must have bounded
-// delay, i.e., the delay of every cycle must be zero.
-//
-// Complexity:
-//
-// - A has bounded delay: exponential.
-// - A does not have bounded delay: does not terminate.
-//
-// For more information, see:
-//
-// Mohri, M. 2003. Edit-distance of weighted automata: General definitions and
-// algorithms. International Journal of Computer Science 14(6): 957-982.
-//
-// This class attaches interface to implementation and handles reference
-// counting, delegating most methods to ImplToFst.
-template <class A>
-class SynchronizeFst : public ImplToFst<internal::SynchronizeFstImpl<A>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using Store = DefaultCacheStore<Arc>;
-  using State = typename Store::State;
-  using Impl = internal::SynchronizeFstImpl<A>;
-
-  friend class ArcIterator<SynchronizeFst<A>>;
-  friend class StateIterator<SynchronizeFst<A>>;
-
-  explicit SynchronizeFst(
-      const Fst<A> &fst,
-      const SynchronizeFstOptions &opts = SynchronizeFstOptions())
-      : ImplToFst<Impl>(std::make_shared<Impl>(fst, opts)) {}
-
-  // See Fst<>::Copy() for doc.
-  SynchronizeFst(const SynchronizeFst<Arc> &fst, bool safe = false)
-      : ImplToFst<Impl>(fst, safe) {}
-
-  // Gets a copy of this SynchronizeFst. See Fst<>::Copy() for further doc.
-  SynchronizeFst<Arc> *Copy(bool safe = false) const override {
-    return new SynchronizeFst<Arc>(*this, safe);
-  }
-
-  inline void InitStateIterator(StateIteratorData<Arc> *data) const override;
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetMutableImpl()->InitArcIterator(s, data);
-  }
-
- private:
-  using ImplToFst<Impl>::GetImpl;
-  using ImplToFst<Impl>::GetMutableImpl;
-
-  SynchronizeFst &operator=(const SynchronizeFst &) = delete;
-};
-
-// Specialization for SynchronizeFst.
-template <class Arc>
-class StateIterator<SynchronizeFst<Arc>>
-    : public CacheStateIterator<SynchronizeFst<Arc>> {
- public:
-  explicit StateIterator(const SynchronizeFst<Arc> &fst)
-      : CacheStateIterator<SynchronizeFst<Arc>>(fst, fst.GetMutableImpl()) {}
-};
-
-// Specialization for SynchronizeFst.
-template <class Arc>
-class ArcIterator<SynchronizeFst<Arc>>
-    : public CacheArcIterator<SynchronizeFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const SynchronizeFst<Arc> &fst, StateId s)
-      : CacheArcIterator<SynchronizeFst<Arc>>(fst.GetMutableImpl(), s) {
-    if (!fst.GetImpl()->HasArcs(s)) fst.GetMutableImpl()->Expand(s);
-  }
-};
-
-template <class Arc>
-inline void SynchronizeFst<Arc>::InitStateIterator(
-    StateIteratorData<Arc> *data) const {
-  data->base = new StateIterator<SynchronizeFst<Arc>>(*this);
-}
-
-// Synchronizes a transducer. This version writes the synchronized result to a
-// MutableFst. The result will be an equivalent FST that has the property that
-// during the traversal of a path, the delay is either zero or strictly
-// increasing, where the delay is the difference between the number of
-// non-epsilon output labels and input labels along the path.
-//
-// For the algorithm to terminate, the input transducer must have bounded
-// delay, i.e., the delay of every cycle must be zero.
-//
-// Complexity:
-//
-// - A has bounded delay: exponential.
-// - A does not have bounded delay: does not terminate.
-//
-// For more information, see:
-//
-// Mohri, M. 2003. Edit-distance of weighted automata: General definitions and
-// algorithms. International Journal of Computer Science 14(6): 957-982.
-template <class Arc>
-void Synchronize(const Fst<Arc> &ifst, MutableFst<Arc> *ofst) {
-  // Caches only the last state for fastest copy.
-  const SynchronizeFstOptions opts(FLAGS_fst_default_cache_gc, 0);
-  *ofst = SynchronizeFst<Arc>(ifst, opts);
-}
-
-}  // namespace fst
-
-#endif  // FST_SYNCHRONIZE_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/test-properties.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/test-properties.h
deleted file mode 100644
index 677ed01b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/test-properties.h
+++ /dev/null
@@ -1,246 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions to manipulate and test property bits.
-
-#ifndef FST_TEST_PROPERTIES_H_
-#define FST_TEST_PROPERTIES_H_
-
-#include <unordered_set>
-
-#include <fst/flags.h>
-#include <fst/log.h>
-
-#include <fst/connect.h>
-#include <fst/dfs-visit.h>
-
-
-DECLARE_bool(fst_verify_properties);
-
-namespace fst {
-// namespace internal {
-
-// For a binary property, the bit is always returned set. For a trinary (i.e.,
-// two-bit) property, both bits are returned set iff either corresponding input
-// bit is set.
-inline uint64 KnownProperties(uint64 props) {
-  return kBinaryProperties | (props & kTrinaryProperties) |
-         ((props & kPosTrinaryProperties) << 1) |
-         ((props & kNegTrinaryProperties) >> 1);
-}
-
-// Tests compatibility between two sets of properties.
-inline bool CompatProperties(uint64 props1, uint64 props2) {
-  const auto known_props1 = KnownProperties(props1);
-  const auto known_props2 = KnownProperties(props2);
-  const auto known_props = known_props1 & known_props2;
-  const auto incompat_props = (props1 & known_props) ^ (props2 & known_props);
-  if (incompat_props) {
-    uint64 prop = 1;
-    for (int i = 0; i < 64; ++i, prop <<= 1) {
-      if (prop & incompat_props) {
-        LOG(ERROR) << "CompatProperties: Mismatch: " << PropertyNames[i]
-                   << ": props1 = " << (props1 & prop ? "true" : "false")
-                   << ", props2 = " << (props2 & prop ? "true" : "false");
-      }
-    }
-    return false;
-  } else {
-    return true;
-  }
-}
-
-// Computes FST property values defined in properties.h. The value of each
-// property indicated in the mask will be determined and returned (these will
-// never be unknown here). In the course of determining the properties
-// specifically requested in the mask, certain other properties may be
-// determined (those with little additional expense) and their values will be
-// returned as well. The complete set of known properties (whether true or
-// false) determined by this operation will be assigned to the the value pointed
-// to by KNOWN. If 'use_stored' is true, pre-computed FST properties may be used
-// when possible. 'mask & required_mask' is used to determine whether the stored
-// propertoes can be used. This routine is seldom called directly; instead it is
-// used to implement fst.Properties(mask, true).
-template <class Arc>
-uint64 ComputeProperties(const Fst<Arc> &fst, uint64 mask, uint64 *known,
-                         bool use_stored) {
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  const auto fst_props = fst.Properties(kFstProperties, false);  // FST-stored.
-  // Check stored FST properties first if allowed.
-  if (use_stored) {
-    const auto known_props = KnownProperties(fst_props);
-    // If FST contains required info, return it.
-    if ((known_props & mask) == mask) {
-      if (known) *known = known_props;
-      return fst_props;
-    }
-  }
-  // Computes (trinary) properties explicitly.
-  // Initialize with binary properties (already known).
-  uint64 comp_props = fst_props & kBinaryProperties;
-  // Computes these trinary properties with a DFS. We compute only those that
-  // need a DFS here, since we otherwise would like to avoid a DFS since its
-  // stack could grow large.
-  uint64 dfs_props = kCyclic | kAcyclic | kInitialCyclic | kInitialAcyclic |
-                     kAccessible | kNotAccessible | kCoAccessible |
-                     kNotCoAccessible;
-  std::vector<StateId> scc;
-  if (mask & (dfs_props | kWeightedCycles | kUnweightedCycles)) {
-    SccVisitor<Arc> scc_visitor(&scc, nullptr, nullptr, &comp_props);
-    DfsVisit(fst, &scc_visitor);
-  }
-  // Computes any remaining trinary properties via a state and arcs iterations
-  if (mask & ~(kBinaryProperties | dfs_props)) {
-    comp_props |= kAcceptor | kNoEpsilons | kNoIEpsilons | kNoOEpsilons |
-                  kILabelSorted | kOLabelSorted | kUnweighted | kTopSorted |
-                  kString;
-    if (mask & (kIDeterministic | kNonIDeterministic)) {
-      comp_props |= kIDeterministic;
-    }
-    if (mask & (kODeterministic | kNonODeterministic)) {
-      comp_props |= kODeterministic;
-    }
-    if (mask & (dfs_props | kWeightedCycles | kUnweightedCycles)) {
-      comp_props |= kUnweightedCycles;
-    }
-    std::unique_ptr<std::unordered_set<Label>> ilabels;
-    std::unique_ptr<std::unordered_set<Label>> olabels;
-    StateId nfinal = 0;
-    for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-      StateId s = siter.Value();
-      Arc prev_arc;
-      // Creates these only if we need to.
-      if (mask & (kIDeterministic | kNonIDeterministic)) {
-        ilabels.reset(new std::unordered_set<Label>());
-      }
-      if (mask & (kODeterministic | kNonODeterministic)) {
-        olabels.reset(new std::unordered_set<Label>());
-      }
-      bool first_arc = true;
-      for (ArcIterator<Fst<Arc>> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-        const auto &arc = aiter.Value();
-        if (ilabels && ilabels->find(arc.ilabel) != ilabels->end()) {
-          comp_props |= kNonIDeterministic;
-          comp_props &= ~kIDeterministic;
-        }
-        if (olabels && olabels->find(arc.olabel) != olabels->end()) {
-          comp_props |= kNonODeterministic;
-          comp_props &= ~kODeterministic;
-        }
-        if (arc.ilabel != arc.olabel) {
-          comp_props |= kNotAcceptor;
-          comp_props &= ~kAcceptor;
-        }
-        if (arc.ilabel == 0 && arc.olabel == 0) {
-          comp_props |= kEpsilons;
-          comp_props &= ~kNoEpsilons;
-        }
-        if (arc.ilabel == 0) {
-          comp_props |= kIEpsilons;
-          comp_props &= ~kNoIEpsilons;
-        }
-        if (arc.olabel == 0) {
-          comp_props |= kOEpsilons;
-          comp_props &= ~kNoOEpsilons;
-        }
-        if (!first_arc) {
-          if (arc.ilabel < prev_arc.ilabel) {
-            comp_props |= kNotILabelSorted;
-            comp_props &= ~kILabelSorted;
-          }
-          if (arc.olabel < prev_arc.olabel) {
-            comp_props |= kNotOLabelSorted;
-            comp_props &= ~kOLabelSorted;
-          }
-        }
-        if (arc.weight != Weight::One() && arc.weight != Weight::Zero()) {
-          comp_props |= kWeighted;
-          comp_props &= ~kUnweighted;
-          if ((comp_props & kUnweightedCycles) &&
-              scc[s] == scc[arc.nextstate]) {
-            comp_props |= kWeightedCycles;
-            comp_props &= ~kUnweightedCycles;
-          }
-        }
-        if (arc.nextstate <= s) {
-          comp_props |= kNotTopSorted;
-          comp_props &= ~kTopSorted;
-        }
-        if (arc.nextstate != s + 1) {
-          comp_props |= kNotString;
-          comp_props &= ~kString;
-        }
-        prev_arc = arc;
-        first_arc = false;
-        if (ilabels) ilabels->insert(arc.ilabel);
-        if (olabels) olabels->insert(arc.olabel);
-      }
-
-      if (nfinal > 0) {  // Final state not last.
-        comp_props |= kNotString;
-        comp_props &= ~kString;
-      }
-      const auto final_weight = fst.Final(s);
-      if (final_weight != Weight::Zero()) {  // Final state.
-        if (final_weight != Weight::One()) {
-          comp_props |= kWeighted;
-          comp_props &= ~kUnweighted;
-        }
-        ++nfinal;
-      } else {  // Non-final state.
-        if (fst.NumArcs(s) != 1) {
-          comp_props |= kNotString;
-          comp_props &= ~kString;
-        }
-      }
-    }
-    if (fst.Start() != kNoStateId && fst.Start() != 0) {
-      comp_props |= kNotString;
-      comp_props &= ~kString;
-    }
-  }
-  if (known) *known = KnownProperties(comp_props);
-  return comp_props;
-}
-
-// This is a wrapper around ComputeProperties that will cause a fatal error if
-// the stored properties and the computed properties are incompatible when
-// FLAGS_fst_verify_properties is true. This routine is seldom called directly;
-// instead it is used to implement fst.Properties(mask, true).
-template <class Arc>
-uint64 TestProperties(const Fst<Arc> &fst, uint64 mask, uint64 *known) {
-  if (FLAGS_fst_verify_properties) {
-    const auto stored_props = fst.Properties(kFstProperties, false);
-    const auto computed_props = ComputeProperties(fst, mask, known, false);
-    if (!CompatProperties(stored_props, computed_props)) {
-      FSTERROR() << "TestProperties: stored FST properties incorrect"
-                 << " (stored: props1, computed: props2)";
-    }
-    return computed_props;
-  } else {
-    return ComputeProperties(fst, mask, known, true);
-  }
-}
-
-// If all the properties of 'fst' corresponding to 'check_mask' are known,
-// returns the stored properties. Otherwise, the properties corresponding to
-// both 'check_mask' and 'test_mask' are computed. This is used to check for
-// newly-added properties that might not be set in old binary files.
-template <class Arc>
-uint64 CheckProperties(const Fst<Arc> &fst, uint64 check_mask,
-                       uint64 test_mask) {
-  auto props = fst.Properties(kFstProperties, false);
-  if (FLAGS_fst_verify_properties) {
-    props = TestProperties(fst, check_mask | test_mask, nullptr);
-  } else if ((KnownProperties(props) & check_mask) != check_mask) {
-    props = ComputeProperties(fst, check_mask | test_mask, nullptr, false);
-  }
-  return props & (check_mask | test_mask);
-}
-
-//}  // namespace internal
-}  // namespace fst
-
-#endif  // FST_TEST_PROPERTIES_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/topsort.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/topsort.h
deleted file mode 100644
index cae3e154..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/topsort.h
+++ /dev/null
@@ -1,95 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Topological sort of FSTs.
-
-#ifndef FST_TOPSORT_H_
-#define FST_TOPSORT_H_
-
-#include <memory>
-#include <vector>
-
-
-#include <fst/dfs-visit.h>
-#include <fst/fst.h>
-#include <fst/statesort.h>
-
-
-namespace fst {
-
-// DFS visitor class to return topological ordering.
-template <class Arc>
-class TopOrderVisitor {
- public:
-  using StateId = typename Arc::StateId;
-
-  // If acyclic, order[i] gives the topological position of StateId i;
-  // otherwise it is unchanged. acyclic_ will be true iff the FST has no
-  // cycles. The caller retains ownership of the state order vector.
-  TopOrderVisitor(std::vector<StateId> *order, bool *acyclic)
-      : order_(order), acyclic_(acyclic) {}
-
-  void InitVisit(const Fst<Arc> &fst) {
-    finish_.reset(new std::vector<StateId>());
-    *acyclic_ = true;
-  }
-
-  constexpr bool InitState(StateId, StateId) const { return true; }
-
-  constexpr bool TreeArc(StateId, const Arc &) const { return true; }
-
-  bool BackArc(StateId, const Arc &) { return (*acyclic_ = false); }
-
-  constexpr bool ForwardOrCrossArc(StateId, const Arc &) const { return true; }
-
-  void FinishState(StateId s, StateId, const Arc *) { finish_->push_back(s); }
-
-  void FinishVisit() {
-    if (*acyclic_) {
-      order_->clear();
-      for (StateId s = 0; s < finish_->size(); ++s) {
-        order_->push_back(kNoStateId);
-      }
-      for (StateId s = 0; s < finish_->size(); ++s) {
-        (*order_)[(*finish_)[finish_->size() - s - 1]] = s;
-      }
-    }
-    finish_.reset();
-  }
-
- private:
-  std::vector<StateId> *order_;
-  bool *acyclic_;
-  // States in finish-time order.
-  std::unique_ptr<std::vector<StateId>> finish_;
-};
-
-// Topologically sorts its input if acyclic, modifying it. Otherwise, the input
-// is unchanged. When sorted, all transitions are from lower to higher state
-// IDs.
-//
-// Complexity:
-//
-//   Time:  O(V + E)
-//   Space: O(V + E)
-//
-// where V is the number of states and E is the number of arcs.
-template <class Arc>
-bool TopSort(MutableFst<Arc> *fst) {
-  std::vector<typename Arc::StateId> order;
-  bool acyclic;
-  TopOrderVisitor<Arc> top_order_visitor(&order, &acyclic);
-  DfsVisit(*fst, &top_order_visitor);
-  if (acyclic) {
-    StateSort(fst, order);
-    fst->SetProperties(kAcyclic | kInitialAcyclic | kTopSorted,
-                       kAcyclic | kInitialAcyclic | kTopSorted);
-  } else {
-    fst->SetProperties(kCyclic | kNotTopSorted, kCyclic | kNotTopSorted);
-  }
-  return acyclic;
-}
-
-}  // namespace fst
-
-#endif  // FST_TOPSORT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/tuple-weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/tuple-weight.h
deleted file mode 100644
index 97680e0c..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/tuple-weight.h
+++ /dev/null
@@ -1,170 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Tuple weight set operation definitions.
-
-#ifndef FST_TUPLE_WEIGHT_H_
-#define FST_TUPLE_WEIGHT_H_
-
-#include <algorithm>
-#include <array>
-#include <functional>
-#include <string>
-#include <vector>
-
-#include <fst/flags.h>
-#include <fst/log.h>
-
-#include <fst/weight.h>
-
-
-namespace fst {
-
-// n-tuple weight, element of the n-th Cartesian power of W.
-template <class W, size_t n>
-class TupleWeight {
- public:
-  using ReverseWeight = TupleWeight<typename W::ReverseWeight, n>;
-
-  using Weight = W;
-  using Index = size_t;
-
-  TupleWeight(const TupleWeight &other) { values_ = other.values_; }
-
-  TupleWeight<W, n> &operator=(const TupleWeight<W, n> &other) {
-    values_ = other.values_;
-    return *this;
-  }
-
-  template <class Iterator>
-  TupleWeight(Iterator begin, Iterator end) {
-    std::copy(begin, end, values_.begin());
-  }
-
-  explicit TupleWeight(const W &weight = W::Zero()) { values_.fill(weight); }
-
-  // Initialize component `index` to `weight`; initialize all other components
-  // to `default_weight`
-  TupleWeight(Index index, const W &weight, const W &default_weight)
-      : TupleWeight(default_weight) {
-    values_[index] = weight;
-  }
-
-  static const TupleWeight<W, n> &Zero() {
-    static const TupleWeight<W, n> zero(W::Zero());
-    return zero;
-  }
-
-  static const TupleWeight<W, n> &One() {
-    static const TupleWeight<W, n> one(W::One());
-    return one;
-  }
-
-  static const TupleWeight<W, n> &NoWeight() {
-    static const TupleWeight<W, n> no_weight(W::NoWeight());
-    return no_weight;
-  }
-
-  constexpr static size_t Length() { return n; }
-
-  std::istream &Read(std::istream &istrm) {
-    for (size_t i = 0; i < n; ++i) values_[i].Read(istrm);
-    return istrm;
-  }
-
-  std::ostream &Write(std::ostream &ostrm) const {
-    for (size_t i = 0; i < n; ++i) values_[i].Write(ostrm);
-    return ostrm;
-  }
-
-  bool Member() const {
-    return std::all_of(values_.begin(), values_.end(),
-                       std::mem_fn(&W::Member));
-  }
-
-  size_t Hash() const {
-    uint64 hash = 0;
-    for (size_t i = 0; i < n; ++i) hash = 5 * hash + values_[i].Hash();
-    return size_t(hash);
-  }
-
-  TupleWeight<W, n> Quantize(float delta = kDelta) const {
-    TupleWeight<W, n> weight;
-    for (size_t i = 0; i < n; ++i) {
-      weight.values_[i] = values_[i].Quantize(delta);
-    }
-    return weight;
-  }
-
-  ReverseWeight Reverse() const {
-    TupleWeight<W, n> w;
-    for (size_t i = 0; i < n; ++i) w.values_[i] = values_[i].Reverse();
-    return w;
-  }
-
-  const W &Value(size_t i) const { return values_[i]; }
-
-  void SetValue(size_t i, const W &w) { values_[i] = w; }
-
- private:
-  std::array<W, n> values_;
-};
-
-template <class W, size_t n>
-inline bool operator==(const TupleWeight<W, n> &w1,
-                       const TupleWeight<W, n> &w2) {
-  for (size_t i = 0; i < n; ++i) {
-    if (w1.Value(i) != w2.Value(i)) return false;
-  }
-  return true;
-}
-
-template <class W, size_t n>
-inline bool operator!=(const TupleWeight<W, n> &w1,
-                       const TupleWeight<W, n> &w2) {
-  for (size_t i = 0; i < n; ++i) {
-    if (w1.Value(i) != w2.Value(i)) return true;
-  }
-  return false;
-}
-
-template <class W, size_t n>
-inline bool ApproxEqual(const TupleWeight<W, n> &w1,
-                        const TupleWeight<W, n> &w2, float delta = kDelta) {
-  for (size_t i = 0; i < n; ++i) {
-    if (!ApproxEqual(w1.Value(i), w2.Value(i), delta)) return false;
-  }
-  return true;
-}
-
-template <class W, size_t n>
-inline std::ostream &operator<<(std::ostream &strm,
-                                const TupleWeight<W, n> &w) {
-  CompositeWeightWriter writer(strm);
-  writer.WriteBegin();
-  for (size_t i = 0; i < n; ++i) writer.WriteElement(w.Value(i));
-  writer.WriteEnd();
-  return strm;
-}
-
-template <class W, size_t n>
-inline std::istream &operator>>(std::istream &strm, TupleWeight<W, n> &w) {
-  CompositeWeightReader reader(strm);
-  reader.ReadBegin();
-  W v;
-  // Reads first n-1 elements.
-  static_assert(n > 0, "Size must be positive.");
-  for (size_t i = 0; i < n - 1; ++i) {
-    reader.ReadElement(&v);
-    w.SetValue(i, v);
-  }
-  // Reads n-th element.
-  reader.ReadElement(&v, true);
-  w.SetValue(n - 1, v);
-  reader.ReadEnd();
-  return strm;
-}
-
-}  // namespace fst
-
-#endif  // FST_TUPLE_WEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/types.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/types.h
deleted file mode 100644
index 9c0b7998..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/types.h
+++ /dev/null
@@ -1,41 +0,0 @@
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Various type definitions (mostly for Google compatibility).
-
-#include <cstddef>       // For std::ptrdiff_t.
-#include <cstdlib>       // for ssize_t.
-#include <cstdint>       // for ?int*_t.
-
-#ifndef FST_LIB_TYPES_H_
-#define FST_LIB_TYPES_H_
-
-using int8 = int8_t;
-using int16 = int16_t;
-using int32 = int32_t;
-using int64 = int64_t;
-
-using uint8 = uint8_t;
-using uint16 = uint16_t;
-using uint32 = uint32_t;
-using uint64 = uint64_t;
-
-#ifdef _MSC_VER
-// Not really Windows-specific: they should have used ptrdiff_t in the first
-// place. But on Windows there has never been ssize_t.
-using ssize_t = std::ptrdiff_t;
-#endif  // _MSC_VER
-
-#endif  // FST_LIB_TYPES_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/union-find.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/union-find.h
deleted file mode 100644
index b5c7f10b..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/union-find.h
+++ /dev/null
@@ -1,84 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Union-find algorithm for dense sets of non-negative integers, implemented
-// using disjoint tree forests with rank heuristics and path compression.
-
-#ifndef FST_UNION_FIND_H_
-#define FST_UNION_FIND_H_
-
-#include <stack>
-#include <vector>
-
-namespace fst {
-
-// Union-Find algorithm for dense sets of non-negative integers.
-template <class T>
-class UnionFind {
- public:
-  // Creates a disjoint set forest for the range [0; max); 'fail' is a value
-  // indicating that an element hasn't been initialized using MakeSet(...).
-  // The upper bound of the range can be reset (increased) using MakeSet(...).
-  UnionFind(T max, T fail) : parent_(max, fail), rank_(max), fail_(fail) {}
-
-  // Finds the representative of the set 'item' belongs to, performing path
-  // compression if necessary.
-  T FindSet(T item) {
-    if (item >= parent_.size() || item == fail_ || parent_[item] == fail_) {
-      return fail_;
-    }
-    auto *p = &parent_[item];
-    for (; *p != item; item = *p, p = &parent_[item]) exec_stack_.push(p);
-    for (; !exec_stack_.empty(); exec_stack_.pop()) *exec_stack_.top() = *p;
-    return *p;
-  }
-
-  // Creates the (destructive) union of the sets x and y belong to.
-  void Union(T x, T y) { Link(FindSet(x), FindSet(y)); }
-
-  // Initialization of an element: creates a singleton set containing 'item'.
-  // The range [0; max) is reset if item >= max.
-  T MakeSet(T item) {
-    if (item >= parent_.size()) {
-      // New value in parent_ should be initialized to fail_.
-      const auto nitem = item > 0 ? 2 * item : 2;
-      parent_.resize(nitem, fail_);
-      rank_.resize(nitem);
-    }
-    parent_[item] = item;
-    return item;
-  }
-
-  // Initialization of all elements starting from 0 to max - 1 to distinct sets.
-  void MakeAllSet(T max) {
-    parent_.resize(max);
-    for (T item = 0; item < max; ++item) parent_[item] = item;
-  }
-
- private:
-  // Links trees rooted in 'x' and 'y'.
-  void Link(T x, T y) {
-    if (x == y) return;
-    if (rank_[x] > rank_[y]) {
-      parent_[y] = x;
-    } else {
-      parent_[x] = y;
-      if (rank_[x] == rank_[y]) {
-        ++rank_[y];
-      }
-    }
-  }
-
-  UnionFind(const UnionFind &) = delete;
-
-  UnionFind &operator=(const UnionFind &) = delete;
-
-  std::vector<T> parent_;       // Parent nodes.
-  std::vector<int> rank_;       // Rank of an element = min. depth in tree.
-  T fail_;                      // Value indicating lookup failure.
-  std::stack<T *> exec_stack_;  // Used for path compression.
-};
-
-}  // namespace fst
-
-#endif  // FST_UNION_FIND_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/union-weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/union-weight.h
deleted file mode 100644
index bb2dea96..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/union-weight.h
+++ /dev/null
@@ -1,505 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Union weight set and associated semiring operation definitions.
-//
-// TODO(riley): add in normalizer functor
-
-#ifndef FST_UNION_WEIGHT_H_
-#define FST_UNION_WEIGHT_H_
-
-#include <cstdlib>
-
-#include <iostream>
-#include <list>
-#include <sstream>
-#include <string>
-#include <utility>
-
-#include <fst/weight.h>
-
-
-namespace fst {
-
-// Example UnionWeightOptions for UnionWeight template below. The Merge
-// operation is used to collapse elements of the set and the Compare function
-// to efficiently implement the merge. In the simplest case, merge would just
-// apply with equality of set elements so the result is a set (and not a
-// multiset). More generally, this can be used to maintain the multiplicity or
-// other such weight associated with the set elements (cf. Gallic weights).
-
-// template <class W>
-// struct UnionWeightOptions {
-//   // Comparison function C is a total order on W that is monotonic w.r.t. to
-//   // Times: for all a, b,c != Zero(): C(a, b) => C(ca, cb) and is
-//   // anti-monotonic w.r.rt to Divide: C(a, b) => C(c/b, c/a).
-//   //
-//   // For all a, b: only one of C(a, b), C(b, a) or a ~ b must true where
-//   // ~ is an equivalence relation on W. Also we require a ~ b iff
-//   // a.Reverse() ~ b.Reverse().
-//   using Compare = NaturalLess<W>;
-//
-//   // How to combine two weights if a ~ b as above. For all a, b: a ~ b =>
-//   // merge(a, b) ~ a, Merge must define a semiring endomorphism from the
-//   // unmerged weight sets to the merged weight sets.
-//   struct Merge {
-//     W operator()(const W &w1, const W &w2) const { return w1; }
-//   };
-//
-//   // For ReverseWeight.
-//   using ReverseOptions = UnionWeightOptions<ReverseWeight>;
-// };
-
-template <class W, class O>
-class UnionWeight;
-
-template <class W, class O>
-class UnionWeightIterator;
-
-template <class W, class O>
-class UnionWeightReverseIterator;
-
-template <class W, class O>
-bool operator==(const UnionWeight<W, O> &, const UnionWeight<W, O> &);
-
-// Semiring that uses Times() and One() from W and union and the empty set
-// for Plus() and Zero(), respectively. Template argument O specifies the union
-// weight options as above.
-template <class W, class O>
-class UnionWeight {
- public:
-  using Weight = W;
-  using Compare = typename O::Compare;
-  using Merge = typename O::Merge;
-
-  using ReverseWeight =
-      UnionWeight<typename W::ReverseWeight, typename O::ReverseOptions>;
-
-  friend class UnionWeightIterator<W, O>;
-  friend class UnionWeightReverseIterator<W, O>;
-  friend bool operator==
-      <>(const UnionWeight<W, O> &, const UnionWeight<W, O> &);
-
-  // Sets represented as first_ weight + rest_ weights. Uses first_ as
-  // NoWeight() to indicate the union weight Zero() ask the empty set. Uses
-  // rest_ containing NoWeight() to indicate the union weight NoWeight().
-  UnionWeight() : first_(W::NoWeight()) {}
-
-  explicit UnionWeight(W weight) : first_(weight) {
-    if (weight == W::NoWeight()) rest_.push_back(weight);
-  }
-
-  static const UnionWeight<W, O> &Zero() {
-    static const UnionWeight<W, O> zero(W::NoWeight());
-    return zero;
-  }
-
-  static const UnionWeight<W, O> &One() {
-    static const UnionWeight<W, O> one(W::One());
-    return one;
-  }
-
-  static const UnionWeight<W, O> &NoWeight() {
-    static const UnionWeight<W, O> no_weight(W::Zero(), W::NoWeight());
-    return no_weight;
-  }
-
-  static const string &Type() {
-    static const string *const type = new string(W::Type() + "_union");
-    return *type;
-  }
-
-  static constexpr uint64 Properties() {
-    return W::Properties() &
-           (kLeftSemiring | kRightSemiring | kCommutative | kIdempotent);
-  }
-
-  bool Member() const;
-
-  std::istream &Read(std::istream &strm);
-
-  std::ostream &Write(std::ostream &strm) const;
-
-  size_t Hash() const;
-
-  UnionWeight<W, O> Quantize(float delta = kDelta) const;
-
-  ReverseWeight Reverse() const;
-
-  // These operations combined with the UnionWeightIterator and
-  // UnionWeightReverseIterator provide the access and mutation of the union
-  // weight internal elements.
-
-  // Common initializer among constructors; clears existing UnionWeight.
-  void Clear() {
-    first_ = W::NoWeight();
-    rest_.clear();
-  }
-
-  size_t Size() const { return first_.Member() ? rest_.size() + 1 : 0; }
-
-  const W &Back() const { return rest_.empty() ? first_ : rest_.back(); }
-
-  // When srt is true, assumes elements added sorted w.r.t Compare and merging
-  // of weights performed as needed. Otherwise, just ensures first_ is the
-  // least element wrt Compare.
-  void PushBack(W weight, bool srt);
-
-  // Sorts the elements of the set. Assumes that first_, if present, is the
-  // least element.
-  void Sort() { rest_.sort(comp_); }
-
- private:
-  W &Back() {
-    if (rest_.empty()) {
-      return first_;
-    } else {
-      return rest_.back();
-    }
-  }
-
-  UnionWeight(W w1, W w2) : first_(std::move(w1)), rest_(1, std::move(w2)) {}
-
-  W first_;            // First weight in set.
-  std::list<W> rest_;  // Remaining weights in set.
-  Compare comp_;
-  Merge merge_;
-};
-
-template <class W, class O>
-void UnionWeight<W, O>::PushBack(W weight, bool srt) {
-  if (!weight.Member()) {
-    rest_.push_back(std::move(weight));
-  } else if (!first_.Member()) {
-    first_ = std::move(weight);
-  } else if (srt) {
-    auto &back = Back();
-    if (comp_(back, weight)) {
-      rest_.push_back(std::move(weight));
-    } else {
-      back = merge_(back, std::move(weight));
-    }
-  } else {
-    if (comp_(first_, weight)) {
-      rest_.push_back(std::move(weight));
-    } else {
-      rest_.push_back(first_);
-      first_ = std::move(weight);
-    }
-  }
-}
-
-// Traverses union weight in the forward direction.
-template <class W, class O>
-class UnionWeightIterator {
- public:
-  explicit UnionWeightIterator(const UnionWeight<W, O> &weight)
-      : first_(weight.first_),
-        rest_(weight.rest_),
-        init_(true),
-        it_(rest_.begin()) {}
-
-  bool Done() const { return init_ ? !first_.Member() : it_ == rest_.end(); }
-
-  const W &Value() const { return init_ ? first_ : *it_; }
-
-  void Next() {
-    if (init_) {
-      init_ = false;
-    } else {
-      ++it_;
-    }
-  }
-
-  void Reset() {
-    init_ = true;
-    it_ = rest_.begin();
-  }
-
- private:
-  const W &first_;
-  const std::list<W> &rest_;
-  bool init_;  // in the initialized state?
-  typename std::list<W>::const_iterator it_;
-};
-
-// Traverses union weight in backward direction.
-template <typename L, class O>
-class UnionWeightReverseIterator {
- public:
-  explicit UnionWeightReverseIterator(const UnionWeight<L, O> &weight)
-      : first_(weight.first_),
-        rest_(weight.rest_),
-        fin_(!first_.Member()),
-        it_(rest_.rbegin()) {}
-
-  bool Done() const { return fin_; }
-
-  const L &Value() const { return it_ == rest_.rend() ? first_ : *it_; }
-
-  void Next() {
-    if (it_ == rest_.rend()) {
-      fin_ = true;
-    } else {
-      ++it_;
-    }
-  }
-
-  void Reset() {
-    fin_ = !first_.Member();
-    it_ = rest_.rbegin();
-  }
-
- private:
-  const L &first_;
-  const std::list<L> &rest_;
-  bool fin_;  // in the final state?
-  typename std::list<L>::const_reverse_iterator it_;
-};
-
-// UnionWeight member functions follow that require UnionWeightIterator.
-template <class W, class O>
-inline std::istream &UnionWeight<W, O>::Read(std::istream &istrm) {
-  Clear();
-  int32 size;
-  ReadType(istrm, &size);
-  for (int i = 0; i < size; ++i) {
-    W weight;
-    ReadType(istrm, &weight);
-    PushBack(weight, true);
-  }
-  return istrm;
-}
-
-template <class W, class O>
-inline std::ostream &UnionWeight<W, O>::Write(std::ostream &ostrm) const {
-  const int32 size = Size();
-  WriteType(ostrm, size);
-  for (UnionWeightIterator<W, O> it(*this); !it.Done(); it.Next()) {
-    WriteType(ostrm, it.Value());
-  }
-  return ostrm;
-}
-
-template <class W, class O>
-inline bool UnionWeight<W, O>::Member() const {
-  if (Size() <= 1) return true;
-  for (UnionWeightIterator<W, O> it(*this); !it.Done(); it.Next()) {
-    if (!it.Value().Member()) return false;
-  }
-  return true;
-}
-
-template <class W, class O>
-inline UnionWeight<W, O> UnionWeight<W, O>::Quantize(float delta) const {
-  UnionWeight<W, O> weight;
-  for (UnionWeightIterator<W, O> it(*this); !it.Done(); it.Next()) {
-    weight.PushBack(it.Value().Quantize(delta), true);
-  }
-  return weight;
-}
-
-template <class W, class O>
-inline typename UnionWeight<W, O>::ReverseWeight UnionWeight<W, O>::Reverse()
-    const {
-  ReverseWeight weight;
-  for (UnionWeightIterator<W, O> it(*this); !it.Done(); it.Next()) {
-    weight.PushBack(it.Value().Reverse(), false);
-  }
-  weight.Sort();
-  return weight;
-}
-
-template <class W, class O>
-inline size_t UnionWeight<W, O>::Hash() const {
-  size_t h = 0;
-  static constexpr int lshift = 5;
-  static constexpr int rshift = CHAR_BIT * sizeof(size_t) - lshift;
-  for (UnionWeightIterator<W, O> it(*this); !it.Done(); it.Next()) {
-    h = h << lshift ^ h >> rshift ^ it.Value().Hash();
-  }
-  return h;
-}
-
-// Requires union weight has been canonicalized.
-template <class W, class O>
-inline bool operator==(const UnionWeight<W, O> &w1,
-                       const UnionWeight<W, O> &w2) {
-  if (w1.Size() != w2.Size()) return false;
-  UnionWeightIterator<W, O> it1(w1);
-  UnionWeightIterator<W, O> it2(w2);
-  for (; !it1.Done(); it1.Next(), it2.Next()) {
-    if (it1.Value() != it2.Value()) return false;
-  }
-  return true;
-}
-
-// Requires union weight has been canonicalized.
-template <class W, class O>
-inline bool operator!=(const UnionWeight<W, O> &w1,
-                       const UnionWeight<W, O> &w2) {
-  return !(w1 == w2);
-}
-
-// Requires union weight has been canonicalized.
-template <class W, class O>
-inline bool ApproxEqual(const UnionWeight<W, O> &w1,
-                        const UnionWeight<W, O> &w2, float delta = kDelta) {
-  if (w1.Size() != w2.Size()) return false;
-  UnionWeightIterator<W, O> it1(w1);
-  UnionWeightIterator<W, O> it2(w2);
-  for (; !it1.Done(); it1.Next(), it2.Next()) {
-    if (!ApproxEqual(it1.Value(), it2.Value(), delta)) return false;
-  }
-  return true;
-}
-
-template <class W, class O>
-inline std::ostream &operator<<(std::ostream &ostrm,
-                                const UnionWeight<W, O> &weight) {
-  UnionWeightIterator<W, O> it(weight);
-  if (it.Done()) {
-    return ostrm << "EmptySet";
-  } else if (!weight.Member()) {
-    return ostrm << "BadSet";
-  } else {
-    CompositeWeightWriter writer(ostrm);
-    writer.WriteBegin();
-    for (; !it.Done(); it.Next()) writer.WriteElement(it.Value());
-    writer.WriteEnd();
-  }
-  return ostrm;
-}
-
-template <class W, class O>
-inline std::istream &operator>>(std::istream &istrm,
-                                UnionWeight<W, O> &weight) {
-  string s;
-  istrm >> s;
-  if (s == "EmptySet") {
-    weight = UnionWeight<W, O>::Zero();
-  } else if (s == "BadSet") {
-    weight = UnionWeight<W, O>::NoWeight();
-  } else {
-    weight = UnionWeight<W, O>::Zero();
-    std::istringstream sstrm(s);
-    CompositeWeightReader reader(sstrm);
-    reader.ReadBegin();
-    bool more = true;
-    while (more) {
-      W v;
-      more = reader.ReadElement(&v);
-      weight.PushBack(v, true);
-    }
-    reader.ReadEnd();
-  }
-  return istrm;
-}
-
-template <class W, class O>
-inline UnionWeight<W, O> Plus(const UnionWeight<W, O> &w1,
-                              const UnionWeight<W, O> &w2) {
-  if (!w1.Member() || !w2.Member()) return UnionWeight<W, O>::NoWeight();
-  if (w1 == UnionWeight<W, O>::Zero()) return w2;
-  if (w2 == UnionWeight<W, O>::Zero()) return w1;
-  UnionWeightIterator<W, O> it1(w1);
-  UnionWeightIterator<W, O> it2(w2);
-  UnionWeight<W, O> sum;
-  typename O::Compare comp;
-  while (!it1.Done() && !it2.Done()) {
-    const auto v1 = it1.Value();
-    const auto v2 = it2.Value();
-    if (comp(v1, v2)) {
-      sum.PushBack(v1, true);
-      it1.Next();
-    } else {
-      sum.PushBack(v2, true);
-      it2.Next();
-    }
-  }
-  for (; !it1.Done(); it1.Next()) sum.PushBack(it1.Value(), true);
-  for (; !it2.Done(); it2.Next()) sum.PushBack(it2.Value(), true);
-  return sum;
-}
-
-template <class W, class O>
-inline UnionWeight<W, O> Times(const UnionWeight<W, O> &w1,
-                               const UnionWeight<W, O> &w2) {
-  if (!w1.Member() || !w2.Member()) return UnionWeight<W, O>::NoWeight();
-  if (w1 == UnionWeight<W, O>::Zero() || w2 == UnionWeight<W, O>::Zero()) {
-    return UnionWeight<W, O>::Zero();
-  }
-  UnionWeightIterator<W, O> it1(w1);
-  UnionWeightIterator<W, O> it2(w2);
-  UnionWeight<W, O> prod1;
-  for (; !it1.Done(); it1.Next()) {
-    UnionWeight<W, O> prod2;
-    for (; !it2.Done(); it2.Next()) {
-      prod2.PushBack(Times(it1.Value(), it2.Value()), true);
-    }
-    prod1 = Plus(prod1, prod2);
-    it2.Reset();
-  }
-  return prod1;
-}
-
-template <class W, class O>
-inline UnionWeight<W, O> Divide(const UnionWeight<W, O> &w1,
-                                const UnionWeight<W, O> &w2, DivideType typ) {
-  if (!w1.Member() || !w2.Member()) return UnionWeight<W, O>::NoWeight();
-  if (w1 == UnionWeight<W, O>::Zero() || w2 == UnionWeight<W, O>::Zero()) {
-    return UnionWeight<W, O>::Zero();
-  }
-  UnionWeightIterator<W, O> it1(w1);
-  UnionWeightReverseIterator<W, O> it2(w2);
-  UnionWeight<W, O> quot;
-  if (w1.Size() == 1) {
-    for (; !it2.Done(); it2.Next()) {
-      quot.PushBack(Divide(it1.Value(), it2.Value(), typ), true);
-    }
-  } else if (w2.Size() == 1) {
-    for (; !it1.Done(); it1.Next()) {
-      quot.PushBack(Divide(it1.Value(), it2.Value(), typ), true);
-    }
-  } else {
-    quot = UnionWeight<W, O>::NoWeight();
-  }
-  return quot;
-}
-
-// This function object generates weights over the union of weights for the
-// underlying generators for the template weight types. This is intended
-// primarily for testing.
-template <class W, class O>
-class WeightGenerate<UnionWeight<W, O>> {
- public:
-  using Weight = UnionWeight<W, O>;
-  using Generate = WeightGenerate<W>;
-
-  explicit WeightGenerate(bool allow_zero = true,
-                          size_t num_random_weights = kNumRandomWeights)
-      : generate_(false), allow_zero_(allow_zero),
-        num_random_weights_(num_random_weights) {}
-
-  Weight operator()() const {
-    const int n = rand() % (num_random_weights_ + 1);  // NOLINT
-    if (allow_zero_ && n == num_random_weights_) {
-      return Weight::Zero();
-    } else if (n % 2 == 0) {
-      return Weight(generate_());
-    } else {
-      return Plus(Weight(generate_()), Weight(generate_()));
-    }
-  }
-
- private:
-  Generate generate_;
-  // Permits Zero() and zero divisors.
-  bool allow_zero_;
-  // The number of alternative random weights.
-  const size_t num_random_weights_;
-};
-
-}  // namespace fst
-
-#endif  // FST_UNION_WEIGHT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/union.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/union.h
deleted file mode 100644
index 5f56c17d..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/union.h
+++ /dev/null
@@ -1,158 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Functions and classes to compute the union of two FSTs.
-
-#ifndef FST_UNION_H_
-#define FST_UNION_H_
-
-#include <algorithm>
-#include <vector>
-
-#include <fst/mutable-fst.h>
-#include <fst/rational.h>
-
-
-namespace fst {
-
-// Computes the union (sum) of two FSTs. This version writes the union to an
-// output MutableFst. If A transduces string x to y with weight a and B
-// transduces string w to v with weight b, then their union transduces x to y
-// with weight a and w to v with weight b.
-//
-// Complexity:
-//
-//   Time: (V_2 + E_2)
-//   Space: O(V_2 + E_2)
-//
-// where Vi is the number of states, and Ei is the number of arcs, in the ith
-// FST.
-template <class Arc>
-void Union(MutableFst<Arc> *fst1, const Fst<Arc> &fst2) {
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  // Checks for symbol table compatibility.
-  if (!CompatSymbols(fst1->InputSymbols(), fst2.InputSymbols()) ||
-      !CompatSymbols(fst1->OutputSymbols(), fst2.OutputSymbols())) {
-    FSTERROR() << "Union: Input/output symbol tables of 1st argument "
-               << "do not match input/output symbol tables of 2nd argument";
-    fst1->SetProperties(kError, kError);
-    return;
-  }
-  const auto numstates1 = fst1->NumStates();
-  const bool initial_acyclic1 = fst1->Properties(kInitialAcyclic, true);
-  const auto props1 = fst1->Properties(kFstProperties, false);
-  const auto props2 = fst2.Properties(kFstProperties, false);
-  const auto start2 = fst2.Start();
-  if (start2 == kNoStateId) {
-    if (props2 & kError) fst1->SetProperties(kError, kError);
-    return;
-  }
-  if (fst2.Properties(kExpanded, false)) {
-    fst1->ReserveStates(numstates1 + CountStates(fst2) +
-                        (initial_acyclic1 ? 0 : 1));
-  }
-  for (StateIterator<Fst<Arc>> siter(fst2); !siter.Done(); siter.Next()) {
-    const auto s1 = fst1->AddState();
-    const auto s2 = siter.Value();
-    fst1->SetFinal(s1, fst2.Final(s2));
-    fst1->ReserveArcs(s1, fst2.NumArcs(s2));
-    for (ArcIterator<Fst<Arc>> aiter(fst2, s2); !aiter.Done(); aiter.Next()) {
-      auto arc = aiter.Value();  // Copy intended.
-      arc.nextstate += numstates1;
-      fst1->AddArc(s1, arc);
-    }
-  }
-  const auto start1 = fst1->Start();
-  if (start1 == kNoStateId) {
-    fst1->SetStart(start2);
-    fst1->SetProperties(props2, kCopyProperties);
-    return;
-  }
-  if (initial_acyclic1) {
-    fst1->AddArc(start1, Arc(0, 0, Weight::One(), start2 + numstates1));
-  } else {
-    const auto nstart1 = fst1->AddState();
-    fst1->SetStart(nstart1);
-    fst1->AddArc(nstart1, Arc(0, 0, Weight::One(), start1));
-    fst1->AddArc(nstart1, Arc(0, 0, Weight::One(), start2 + numstates1));
-  }
-  fst1->SetProperties(UnionProperties(props1, props2), kFstProperties);
-}
-
-// Computes the union of two FSTs, modifying the RationalFst argument.
-template <class Arc>
-void Union(RationalFst<Arc> *fst1, const Fst<Arc> &fst2) {
-  fst1->GetMutableImpl()->AddUnion(fst2);
-}
-
-using UnionFstOptions = RationalFstOptions;
-
-// Computes the union (sum) of two FSTs. This version is a delayed FST. If A
-// transduces string x to y with weight a and B transduces string w to v with
-// weight b, then their union transduces x to y with weight a and w to v with
-// weight b.
-//
-// Complexity:
-//
-//   Time: O(v_1 + e_1 + v_2 + e_2)
-//   Space: O(v_1 + v_2)
-//
-// where vi is the number of states visited, and ei is the number of arcs
-// visited, in the ith FST. Constant time and space to visit an input state or
-// arc is assumed and exclusive of caching.
-template <class A>
-class UnionFst : public RationalFst<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  UnionFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2) {
-    GetMutableImpl()->InitUnion(fst1, fst2);
-  }
-
-  UnionFst(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
-           const UnionFstOptions &opts)
-      : RationalFst<Arc>(opts) {
-    GetMutableImpl()->InitUnion(fst1, fst2);
-  }
-
-  // See Fst<>::Copy() for doc.
-  UnionFst(const UnionFst<Arc> &fst, bool safe = false)
-      : RationalFst<Arc>(fst, safe) {}
-
-  // Gets a copy of this UnionFst. See Fst<>::Copy() for further doc.
-  UnionFst<Arc> *Copy(bool safe = false) const override {
-    return new UnionFst<Arc>(*this, safe);
-  }
-
- private:
-  using ImplToFst<internal::RationalFstImpl<Arc>>::GetImpl;
-  using ImplToFst<internal::RationalFstImpl<Arc>>::GetMutableImpl;
-};
-
-// Specialization for UnionFst.
-template <class Arc>
-class StateIterator<UnionFst<Arc>> : public StateIterator<RationalFst<Arc>> {
- public:
-  explicit StateIterator(const UnionFst<Arc> &fst)
-      : StateIterator<RationalFst<Arc>>(fst) {}
-};
-
-// Specialization for UnionFst.
-template <class Arc>
-class ArcIterator<UnionFst<Arc>> : public ArcIterator<RationalFst<Arc>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const UnionFst<Arc> &fst, StateId s)
-      : ArcIterator<RationalFst<Arc>>(fst, s) {}
-};
-
-using StdUnionFst = UnionFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_UNION_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/util.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/util.h
deleted file mode 100644
index bfbe11a2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/util.h
+++ /dev/null
@@ -1,399 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// FST utility inline definitions.
-
-#ifndef FST_UTIL_H_
-#define FST_UTIL_H_
-
-#include <iostream>
-#include <iterator>
-#include <list>
-#include <map>
-#include <set>
-#include <sstream>
-#include <string>
-#include <type_traits>
-#include <unordered_map>
-#include <unordered_set>
-#include <utility>
-#include <vector>
-
-#include <fst/compat.h>
-#include <fst/types.h>
-#include <fst/log.h>
-#include <fstream>
-
-#include <fst/flags.h>
-
-
-// Utility for error handling.
-
-DECLARE_bool(fst_error_fatal);
-
-#define FSTERROR() \
-  (FLAGS_fst_error_fatal ? LOG(FATAL) : LOG(ERROR))
-
-namespace fst {
-
-// Utility for type I/O.
-
-// Reads types from an input stream.
-
-// Generic case.
-template <class T,
-    typename std::enable_if<std::is_class<T>::value, T>::type* = nullptr>
-inline std::istream &ReadType(std::istream &strm, T *t) {
-  return t->Read(strm);
-}
-
-// Numeric (boolean, integral, floating-point) case.
-template <class T,
-    typename std::enable_if<std::is_arithmetic<T>::value, T>::type* = nullptr>
-inline std::istream &ReadType(std::istream &strm, T *t) {
-  return strm.read(reinterpret_cast<char *>(t), sizeof(T)); \
-}
-
-// String case.
-inline std::istream &ReadType(std::istream &strm, string *s) {  // NOLINT
-  s->clear();
-  int32 ns = 0;
-  strm.read(reinterpret_cast<char *>(&ns), sizeof(ns));
-  for (int32 i = 0; i < ns; ++i) {
-    char c;
-    strm.read(&c, 1);
-    *s += c;
-  }
-  return strm;
-}
-
-// Pair case.
-template <typename S, typename T>
-inline std::istream &ReadType(std::istream &strm, std::pair<S, T> *p) {
-  ReadType(strm, &p->first);
-  ReadType(strm, &p->second);
-  return strm;
-}
-
-template <typename S, typename T>
-inline std::istream &ReadType(std::istream &strm, std::pair<const S, T> *p) {
-  ReadType(strm, const_cast<S *>(&p->first));
-  ReadType(strm, &p->second);
-  return strm;
-}
-
-namespace internal {
-template <class C, class ReserveFn>
-std::istream &ReadContainerType(std::istream &strm, C *c, ReserveFn reserve) {
-  c->clear();
-  int64 n = 0;
-  ReadType(strm, &n);
-  reserve(c, n);
-  auto insert = std::inserter(*c, c->begin());
-  for (int64 i = 0; i < n; ++i) {
-    typename C::value_type value;
-    ReadType(strm, &value);
-    *insert = value;
-  }
-  return strm;
-}
-}  // namespace internal
-
-template <typename T, typename A>
-std::istream &ReadType(std::istream &strm, std::vector<T,A> *c) {
-  return internal::ReadContainerType(
-      strm, c, [](decltype(c) v, int n) { v->reserve(n); });
-}
-
-template <typename T, typename A>
-std::istream &ReadType(std::istream &strm, std::list<T,A> *c) {
-  return internal::ReadContainerType(strm, c, [](decltype(c) v, int n) {});
-}
-
-template <typename T, typename L, typename A>
-std::istream &ReadType(std::istream &strm, std::set<T,L,A> *c) {
-  return internal::ReadContainerType(strm, c, [](decltype(c) v, int n) {});
-}
-
-template <typename K, typename V, typename L, typename A>
-std::istream &ReadType(std::istream &strm, std::map<K,V,L,A> *c) {
-  return internal::ReadContainerType(strm, c, [](decltype(c) v, int n) {});
-}
-
-template <typename T, typename H, typename E, typename A>
-std::istream &ReadType(std::istream &strm, std::unordered_set<T,H,E,A> *c) {
-  return internal::ReadContainerType(
-      strm, c, [](decltype(c) v, int n) { v->reserve(n); });
-}
-
-template <typename K, typename V, typename H, typename E, typename A>
-std::istream &ReadType(std::istream &strm, std::unordered_map<K,V,H,E,A> *c) {
-  return internal::ReadContainerType(
-      strm, c, [](decltype(c) v, int n) { v->reserve(n); });
-}
-
-// Writes types to an output stream.
-
-// Generic case.
-template <class T,
-    typename std::enable_if<std::is_class<T>::value, T>::type* = nullptr>
-inline std::ostream &WriteType(std::ostream &strm, const T t) {
-  t.Write(strm);
-  return strm;
-}
-
-// Numeric (boolean, integral, floating-point) case.
-template <class T,
-    typename std::enable_if<std::is_arithmetic<T>::value, T>::type* = nullptr>
-inline std::ostream &WriteType(std::ostream &strm, const T t) {
-  return strm.write(reinterpret_cast<const char *>(&t), sizeof(T));
-}
-
-// String case.
-inline std::ostream &WriteType(std::ostream &strm, const string &s) {  // NOLINT
-  int32 ns = s.size();
-  strm.write(reinterpret_cast<const char *>(&ns), sizeof(ns));
-  return strm.write(s.data(), ns);
-}
-
-// Pair case.
-template <typename S, typename T>
-inline std::ostream &WriteType(std::ostream &strm,
-                               const std::pair<S, T> &p) {  // NOLINT
-  WriteType(strm, p.first);
-  WriteType(strm, p.second);
-  return strm;
-}
-
-namespace internal {
-template <class C>
-std::ostream &WriteContainer(std::ostream &strm, const C &c) {
-  const int64 n = c.size();
-  WriteType(strm, n);
-  for (const auto &e : c) {
-    WriteType(strm, e);
-  }
-  return strm;
-}
-}  // namespace internal
-
-template <typename T, typename A>
-std::ostream &WriteType(std::ostream &strm, const std::vector<T,A> &c) {
-  return internal::WriteContainer(strm, c);
-}
-
-template <typename T, typename A>
-std::ostream &WriteType(std::ostream &strm, const std::list<T,A> &c) {
-  return internal::WriteContainer(strm, c);
-}
-
-template <typename T, typename L, typename A>
-std::ostream &WriteType(std::ostream &strm, const std::set<T,L,A> &c) {
-  return internal::WriteContainer(strm, c);
-}
-
-template <typename K, typename V, typename L, typename A>
-std::ostream &WriteType(std::ostream &strm, const std::map<K,V,L,A> &c) {
-  return internal::WriteContainer(strm, c);
-}
-
-template <typename T, typename H, typename E, typename A>
-std::ostream &WriteType(std::ostream &strm, const std::unordered_set<T,H,E,A> &c) {
-  return internal::WriteContainer(strm, c);
-}
-
-template <typename K, typename V, typename H, typename E, typename A>
-std::ostream &WriteType(std::ostream &strm, const std::unordered_map<K,V,H,E,A> &c) {
-  return internal::WriteContainer(strm, c);
-}
-
-// Utilities for converting between int64 or Weight and string.
-
-int64 StrToInt64(const string &s, const string &src, size_t nline,
-                 bool allow_negative, bool *error = nullptr);
-
-template <typename Weight>
-Weight StrToWeight(const string &s, const string &src, size_t nline) {
-  Weight w;
-  std::istringstream strm(s);
-  strm >> w;
-  if (!strm) {
-    FSTERROR() << "StrToWeight: Bad weight = \"" << s << "\", source = " << src
-               << ", line = " << nline;
-    return Weight::NoWeight();
-  }
-  return w;
-}
-
-template <typename Weight>
-void WeightToStr(Weight w, string *s) {
-  std::ostringstream strm;
-  strm.precision(9);
-  strm << w;
-  s->append(strm.str().data(), strm.str().size());
-}
-
-// Utilities for reading/writing integer pairs (typically labels)
-
-// Modifies line using a vector of pointers to a buffer beginning with line.
-void SplitToVector(char *line, const char *delim, std::vector<char *> *vec,
-                   bool omit_empty_strings);
-
-template <typename I>
-bool ReadIntPairs(const string &filename, std::vector<std::pair<I, I>> *pairs,
-                  bool allow_negative = false) {
-  std::ifstream strm(filename, std::ios_base::in);
-  if (!strm) {
-    LOG(ERROR) << "ReadIntPairs: Can't open file: " << filename;
-    return false;
-  }
-  const int kLineLen = 8096;
-  char line[kLineLen];
-  size_t nline = 0;
-  pairs->clear();
-  while (strm.getline(line, kLineLen)) {
-    ++nline;
-    std::vector<char *> col;
-    SplitToVector(line, "\n\t ", &col, true);
-    // empty line or comment?
-    if (col.empty() || col[0][0] == '\0' || col[0][0] == '#') continue;
-    if (col.size() != 2) {
-      LOG(ERROR) << "ReadIntPairs: Bad number of columns, "
-                 << "file = " << filename << ", line = " << nline;
-      return false;
-    }
-    bool err;
-    I i1 = StrToInt64(col[0], filename, nline, allow_negative, &err);
-    if (err) return false;
-    I i2 = StrToInt64(col[1], filename, nline, allow_negative, &err);
-    if (err) return false;
-    pairs->push_back(std::make_pair(i1, i2));
-  }
-  return true;
-}
-
-template <typename I>
-bool WriteIntPairs(const string &filename,
-                   const std::vector<std::pair<I, I>> &pairs) {
-  std::ostream *strm = &std::cout;
-  if (!filename.empty()) {
-    strm = new std::ofstream(filename);
-    if (!*strm) {
-      LOG(ERROR) << "WriteIntPairs: Can't open file: " << filename;
-      return false;
-    }
-  }
-  for (ssize_t n = 0; n < pairs.size(); ++n) {
-    *strm << pairs[n].first << "\t" << pairs[n].second << "\n";
-  }
-  if (!*strm) {
-    LOG(ERROR) << "WriteIntPairs: Write failed: "
-               << (filename.empty() ? "standard output" : filename);
-    return false;
-  }
-  if (strm != &std::cout) delete strm;
-  return true;
-}
-
-// Utilities for reading/writing label pairs.
-
-template <typename Label>
-bool ReadLabelPairs(const string &filename,
-                    std::vector<std::pair<Label, Label>> *pairs,
-                    bool allow_negative = false) {
-  return ReadIntPairs(filename, pairs, allow_negative);
-}
-
-template <typename Label>
-bool WriteLabelPairs(const string &filename,
-                     const std::vector<std::pair<Label, Label>> &pairs) {
-  return WriteIntPairs(filename, pairs);
-}
-
-// Utilities for converting a type name to a legal C symbol.
-
-void ConvertToLegalCSymbol(string *s);
-
-// Utilities for stream I/O.
-
-bool AlignInput(std::istream &strm);
-bool AlignOutput(std::ostream &strm);
-
-// An associative container for which testing membership is faster than an STL
-// set if members are restricted to an interval that excludes most non-members.
-// A Key must have ==, !=, and < operators defined. Element NoKey should be a
-// key that marks an uninitialized key and is otherwise unused. Find() returns
-// an STL const_iterator to the match found, otherwise it equals End().
-template <class Key, Key NoKey>
-class CompactSet {
- public:
-  using const_iterator = typename std::set<Key>::const_iterator;
-
-  CompactSet() : min_key_(NoKey), max_key_(NoKey) {}
-
-  CompactSet(const CompactSet<Key, NoKey> &compact_set)
-      : set_(compact_set.set_),
-        min_key_(compact_set.min_key_),
-        max_key_(compact_set.max_key_) {}
-
-  void Insert(Key key) {
-    set_.insert(key);
-    if (min_key_ == NoKey || key < min_key_) min_key_ = key;
-    if (max_key_ == NoKey || max_key_ < key) max_key_ = key;
-  }
-
-  void Erase(Key key) {
-    set_.erase(key);
-    if (set_.empty()) {
-      min_key_ = max_key_ = NoKey;
-    } else if (key == min_key_) {
-      ++min_key_;
-    } else if (key == max_key_) {
-      --max_key_;
-    }
-  }
-
-  void Clear() {
-    set_.clear();
-    min_key_ = max_key_ = NoKey;
-  }
-
-  const_iterator Find(Key key) const {
-    if (min_key_ == NoKey || key < min_key_ || max_key_ < key) {
-      return set_.end();
-    } else {
-      return set_.find(key);
-    }
-  }
-
-  bool Member(Key key) const {
-    if (min_key_ == NoKey || key < min_key_ || max_key_ < key) {
-      return false;  // out of range
-    } else if (min_key_ != NoKey && max_key_ + 1 == min_key_ + set_.size()) {
-      return true;  // dense range
-    } else {
-      return set_.count(key);
-    }
-  }
-
-  const_iterator Begin() const { return set_.begin(); }
-
-  const_iterator End() const { return set_.end(); }
-
-  // All stored keys are greater than or equal to this value.
-  Key LowerBound() const { return min_key_; }
-
-  // All stored keys are less than or equal to this value.
-  Key UpperBound() const { return max_key_; }
-
- private:
-  std::set<Key> set_;
-  Key min_key_;
-  Key max_key_;
-
-  void operator=(const CompactSet &) = delete;
-};
-
-}  // namespace fst
-
-#endif  // FST_UTIL_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/vector-fst.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/vector-fst.h
deleted file mode 100644
index 0288d9d2..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/vector-fst.h
+++ /dev/null
@@ -1,717 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Simple concrete, mutable FST whose states and arcs are stored in STL vectors.
-
-#ifndef FST_VECTOR_FST_H_
-#define FST_VECTOR_FST_H_
-
-#include <string>
-#include <utility>
-#include <vector>
-
-#include <fst/log.h>
-
-#include <fst/fst-decl.h>  // For optional argument declarations
-#include <fst/mutable-fst.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-template <class A, class S>
-class VectorFst;
-
-template <class F, class G>
-void Cast(const F &, G *);
-
-// Arcs (of type A) implemented by an STL vector per state. M specifies Arc
-// allocator (default declared in fst-decl.h).
-template <class A, class M /* = std::allocator<A> */>
-class VectorState {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  using ArcAllocator = M;
-  using StateAllocator =
-      typename ArcAllocator::template rebind<VectorState<Arc, M>>::other;
-
-  // Provide STL allocator for arcs.
-  explicit VectorState(const ArcAllocator &alloc)
-      : final_(Weight::Zero()), niepsilons_(0), noepsilons_(0), arcs_(alloc) {}
-
-  VectorState(const VectorState<A, M> &state, const ArcAllocator &alloc)
-      : final_(state.Final()),
-        niepsilons_(state.NumInputEpsilons()),
-        noepsilons_(state.NumOutputEpsilons()),
-        arcs_(state.arcs_.begin(), state.arcs_.end(), alloc) {}
-
-  void Reset() {
-    final_ = Weight::Zero();
-    niepsilons_ = 0;
-    noepsilons_ = 0;
-    arcs_.clear();
-  }
-
-  Weight Final() const { return final_; }
-
-  size_t NumInputEpsilons() const { return niepsilons_; }
-
-  size_t NumOutputEpsilons() const { return noepsilons_; }
-
-  size_t NumArcs() const { return arcs_.size(); }
-
-  const Arc &GetArc(size_t n) const { return arcs_[n]; }
-
-  const Arc *Arcs() const { return !arcs_.empty() ? &arcs_[0] : nullptr; }
-
-  Arc *MutableArcs() { return !arcs_.empty() ? &arcs_[0] : nullptr; }
-
-  void ReserveArcs(size_t n) { arcs_.reserve(n); }
-
-  void SetFinal(Weight weight) { final_ = std::move(weight); }
-
-  void SetNumInputEpsilons(size_t n) { niepsilons_ = n; }
-
-  void SetNumOutputEpsilons(size_t n) { noepsilons_ = n; }
-
-  void AddArc(const Arc &arc) {
-    if (arc.ilabel == 0) ++niepsilons_;
-    if (arc.olabel == 0) ++noepsilons_;
-    arcs_.push_back(arc);
-  }
-
-  void SetArc(const Arc &arc, size_t n) {
-    if (arcs_[n].ilabel == 0) --niepsilons_;
-    if (arcs_[n].olabel == 0) --noepsilons_;
-    if (arc.ilabel == 0) ++niepsilons_;
-    if (arc.olabel == 0) ++noepsilons_;
-    arcs_[n] = arc;
-  }
-
-  void DeleteArcs() {
-    niepsilons_ = 0;
-    noepsilons_ = 0;
-    arcs_.clear();
-  }
-
-  void DeleteArcs(size_t n) {
-    for (size_t i = 0; i < n; ++i) {
-      if (arcs_.back().ilabel == 0) --niepsilons_;
-      if (arcs_.back().olabel == 0) --noepsilons_;
-      arcs_.pop_back();
-    }
-  }
-
-  // For state class allocation.
-  void *operator new(size_t size, StateAllocator *alloc) {
-    return alloc->allocate(1);
-  }
-
-  // For state destruction and memory freeing.
-  static void Destroy(VectorState<A, M> *state, StateAllocator *alloc) {
-    if (state) {
-      state->~VectorState<A, M>();
-      alloc->deallocate(state, 1);
-    }
-  }
-
- private:
-  Weight final_;                       // Final weight.
-  size_t niepsilons_;                  // # of input epsilons
-  size_t noepsilons_;                  // # of output epsilons
-  std::vector<A, ArcAllocator> arcs_;  // Arc container.
-};
-
-namespace internal {
-
-// States are implemented by STL vectors, templated on the
-// State definition. This does not manage the Fst properties.
-template <class S>
-class VectorFstBaseImpl : public FstImpl<typename S::Arc> {
- public:
-  using State = S;
-  using Arc = typename State::Arc;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  VectorFstBaseImpl() : start_(kNoStateId) {}
-
-  ~VectorFstBaseImpl() override {
-    for (StateId s = 0; s < states_.size(); ++s) {
-      State::Destroy(states_[s], &state_alloc_);
-    }
-  }
-
-  StateId Start() const { return start_; }
-
-  Weight Final(StateId state) const { return states_[state]->Final(); }
-
-  StateId NumStates() const { return states_.size(); }
-
-  size_t NumArcs(StateId state) const { return states_[state]->NumArcs(); }
-
-  size_t NumInputEpsilons(StateId state) const {
-    return GetState(state)->NumInputEpsilons();
-  }
-
-  size_t NumOutputEpsilons(StateId state) const {
-    return GetState(state)->NumOutputEpsilons();
-  }
-
-  void SetStart(StateId state) { start_ = state; }
-
-  void SetFinal(StateId state, Weight weight) {
-    states_[state]->SetFinal(std::move(weight));
-  }
-
-  StateId AddState() {
-    states_.push_back(new (&state_alloc_) State(arc_alloc_));
-    return states_.size() - 1;
-  }
-
-  StateId AddState(State *state) {
-    states_.push_back(state);
-    return states_.size() - 1;
-  }
-
-  void AddArc(StateId state, const Arc &arc) { states_[state]->AddArc(arc); }
-
-  void DeleteStates(const std::vector<StateId> &dstates) {
-    std::vector<StateId> newid(states_.size(), 0);
-    for (StateId i = 0; i < dstates.size(); ++i) newid[dstates[i]] = kNoStateId;
-    StateId nstates = 0;
-    for (StateId state = 0; state < states_.size(); ++state) {
-      if (newid[state] != kNoStateId) {
-        newid[state] = nstates;
-        if (state != nstates) states_[nstates] = states_[state];
-        ++nstates;
-      } else {
-        State::Destroy(states_[state], &state_alloc_);
-      }
-    }
-    states_.resize(nstates);
-    for (StateId state = 0; state < states_.size(); ++state) {
-      auto *arcs = states_[state]->MutableArcs();
-      size_t narcs = 0;
-      auto nieps = states_[state]->NumInputEpsilons();
-      auto noeps = states_[state]->NumOutputEpsilons();
-      for (size_t i = 0; i < states_[state]->NumArcs(); ++i) {
-        const auto t = newid[arcs[i].nextstate];
-        if (t != kNoStateId) {
-          arcs[i].nextstate = t;
-          if (i != narcs) arcs[narcs] = arcs[i];
-          ++narcs;
-        } else {
-          if (arcs[i].ilabel == 0) --nieps;
-          if (arcs[i].olabel == 0) --noeps;
-        }
-      }
-      states_[state]->DeleteArcs(states_[state]->NumArcs() - narcs);
-      states_[state]->SetNumInputEpsilons(nieps);
-      states_[state]->SetNumOutputEpsilons(noeps);
-    }
-    if (Start() != kNoStateId) SetStart(newid[Start()]);
-  }
-
-  void DeleteStates() {
-    for (StateId state = 0; state < states_.size(); ++state) {
-      State::Destroy(states_[state], &state_alloc_);
-    }
-    states_.clear();
-    SetStart(kNoStateId);
-  }
-
-  void DeleteArcs(StateId state, size_t n) { states_[state]->DeleteArcs(n); }
-
-  void DeleteArcs(StateId state) { states_[state]->DeleteArcs(); }
-
-  State *GetState(StateId state) { return states_[state]; }
-
-  const State *GetState(StateId state) const { return states_[state]; }
-
-  void SetState(StateId state, State *vstate) { states_[state] = vstate; }
-
-  void ReserveStates(StateId n) { states_.reserve(n); }
-
-  void ReserveArcs(StateId state, size_t n) { states_[state]->ReserveArcs(n); }
-
-  // Provide information needed for generic state iterator.
-  void InitStateIterator(StateIteratorData<Arc> *data) const {
-    data->base = nullptr;
-    data->nstates = states_.size();
-  }
-
-  // Provide information needed for generic arc iterator.
-  void InitArcIterator(StateId state, ArcIteratorData<Arc> *data) const {
-    data->base = nullptr;
-    data->narcs = states_[state]->NumArcs();
-    data->arcs = states_[state]->Arcs();
-    data->ref_count = nullptr;
-  }
-
- private:
-  std::vector<State *> states_;                 // States represenation.
-  StateId start_;                               // Initial state.
-  typename State::StateAllocator state_alloc_;  // For state allocation.
-  typename State::ArcAllocator arc_alloc_;      // For arc allocation.
-
-  VectorFstBaseImpl(const VectorFstBaseImpl &) = delete;
-  VectorFstBaseImpl &operator=(const VectorFstBaseImpl &) = delete;
-};
-
-// This is a VectorFstBaseImpl container that holds VectorStates and manages FST
-// properties.
-template <class S>
-class VectorFstImpl : public VectorFstBaseImpl<S> {
- public:
-  using State = S;
-  using Arc = typename State::Arc;
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  using FstImpl<Arc>::SetInputSymbols;
-  using FstImpl<Arc>::SetOutputSymbols;
-  using FstImpl<Arc>::SetType;
-  using FstImpl<Arc>::SetProperties;
-  using FstImpl<Arc>::Properties;
-
-  using VectorFstBaseImpl<S>::Start;
-  using VectorFstBaseImpl<S>::NumStates;
-  using VectorFstBaseImpl<S>::GetState;
-  using VectorFstBaseImpl<S>::ReserveArcs;
-
-  friend class MutableArcIterator<VectorFst<Arc, S>>;
-
-  using BaseImpl = VectorFstBaseImpl<S>;
-
-  VectorFstImpl() {
-    SetType("vector");
-    SetProperties(kNullProperties | kStaticProperties);
-  }
-
-  explicit VectorFstImpl(const Fst<Arc> &fst);
-
-  static VectorFstImpl<S> *Read(std::istream &strm, const FstReadOptions &opts);
-
-  void SetStart(StateId state) {
-    BaseImpl::SetStart(state);
-    SetProperties(SetStartProperties(Properties()));
-  }
-
-  void SetFinal(StateId state, Weight weight) {
-    const auto old_weight = BaseImpl::Final(state);
-    BaseImpl::SetFinal(state, std::move(weight));
-    SetProperties(SetFinalProperties(Properties(), old_weight, weight));
-  }
-
-  StateId AddState() {
-    const auto state = BaseImpl::AddState();
-    SetProperties(AddStateProperties(Properties()));
-    return state;
-  }
-
-  void AddArc(StateId state, const Arc &arc) {
-    auto *vstate = GetState(state);
-    const auto *parc = vstate->NumArcs() == 0
-                           ? nullptr
-                           : &(vstate->GetArc(vstate->NumArcs() - 1));
-    SetProperties(AddArcProperties(Properties(), state, arc, parc));
-    BaseImpl::AddArc(state, arc);
-  }
-
-  void DeleteStates(const std::vector<StateId> &dstates) {
-    BaseImpl::DeleteStates(dstates);
-    SetProperties(DeleteStatesProperties(Properties()));
-  }
-
-  void DeleteStates() {
-    BaseImpl::DeleteStates();
-    SetProperties(DeleteAllStatesProperties(Properties(), kStaticProperties));
-  }
-
-  void DeleteArcs(StateId state, size_t n) {
-    BaseImpl::DeleteArcs(state, n);
-    SetProperties(DeleteArcsProperties(Properties()));
-  }
-
-  void DeleteArcs(StateId state) {
-    BaseImpl::DeleteArcs(state);
-    SetProperties(DeleteArcsProperties(Properties()));
-  }
-
-  // Properties always true of this FST class
-  static constexpr uint64 kStaticProperties = kExpanded | kMutable;
-
- private:
-  // Minimum file format version supported.
-  static constexpr int kMinFileVersion = 2;
-};
-
-template <class S>
-constexpr uint64 VectorFstImpl<S>::kStaticProperties;
-
-template <class S>
-constexpr int VectorFstImpl<S>::kMinFileVersion;
-
-template <class S>
-VectorFstImpl<S>::VectorFstImpl(const Fst<Arc> &fst) {
-  SetType("vector");
-  SetInputSymbols(fst.InputSymbols());
-  SetOutputSymbols(fst.OutputSymbols());
-  BaseImpl::SetStart(fst.Start());
-  if (fst.Properties(kExpanded, false)) {
-    BaseImpl::ReserveStates(CountStates(fst));
-  }
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    const auto state = siter.Value();
-    BaseImpl::AddState();
-    BaseImpl::SetFinal(state, fst.Final(state));
-    ReserveArcs(state, fst.NumArcs(state));
-    for (ArcIterator<Fst<Arc>> aiter(fst, state); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      BaseImpl::AddArc(state, arc);
-    }
-  }
-  SetProperties(fst.Properties(kCopyProperties, false) | kStaticProperties);
-}
-
-template <class S>
-VectorFstImpl<S> *VectorFstImpl<S>::Read(std::istream &strm,
-                                         const FstReadOptions &opts) {
-  std::unique_ptr<VectorFstImpl<S>> impl(new VectorFstImpl());
-  FstHeader hdr;
-  if (!impl->ReadHeader(strm, opts, kMinFileVersion, &hdr)) return nullptr;
-  impl->BaseImpl::SetStart(hdr.Start());
-  if (hdr.NumStates() != kNoStateId) impl->ReserveStates(hdr.NumStates());
-  StateId state = 0;
-  for (; hdr.NumStates() == kNoStateId || state < hdr.NumStates(); ++state) {
-    Weight weight;
-    if (!weight.Read(strm)) break;
-    impl->BaseImpl::AddState();
-    auto *vstate = impl->GetState(state);
-    vstate->SetFinal(weight);
-    int64 narcs;
-    ReadType(strm, &narcs);
-    if (!strm) {
-      LOG(ERROR) << "VectorFst::Read: Read failed: " << opts.source;
-      return nullptr;
-    }
-    impl->ReserveArcs(state, narcs);
-    for (int64 i = 0; i < narcs; ++i) {
-      Arc arc;
-      ReadType(strm, &arc.ilabel);
-      ReadType(strm, &arc.olabel);
-      arc.weight.Read(strm);
-      ReadType(strm, &arc.nextstate);
-      if (!strm) {
-        LOG(ERROR) << "VectorFst::Read: Read failed: " << opts.source;
-        return nullptr;
-      }
-      impl->BaseImpl::AddArc(state, arc);
-    }
-  }
-  if (hdr.NumStates() != kNoStateId && state != hdr.NumStates()) {
-    LOG(ERROR) << "VectorFst::Read: Unexpected end of file: " << opts.source;
-    return nullptr;
-  }
-  return impl.release();
-}
-
-}  // namespace internal
-
-// Simple concrete, mutable FST. This class attaches interface to implementation
-// and handles reference counting, delegating most methods to ImplToMutableFst.
-// Also supports ReserveStates and ReserveArcs methods (cf. STL vector methods).
-// The second optional template argument gives the State definition.
-template <class A, class S /* = VectorState<A> */>
-class VectorFst : public ImplToMutableFst<internal::VectorFstImpl<S>> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  using State = S;
-  using Impl = internal::VectorFstImpl<State>;
-
-  friend class StateIterator<VectorFst<Arc, State>>;
-  friend class ArcIterator<VectorFst<Arc, State>>;
-  friend class MutableArcIterator<VectorFst<A, S>>;
-
-  template <class F, class G>
-  friend void Cast(const F &, G *);
-
-  VectorFst() : ImplToMutableFst<Impl>(std::make_shared<Impl>()) {}
-
-  explicit VectorFst(const Fst<Arc> &fst)
-      : ImplToMutableFst<Impl>(std::make_shared<Impl>(fst)) {}
-
-  VectorFst(const VectorFst<Arc, State> &fst, bool safe = false)
-      : ImplToMutableFst<Impl>(fst) {}
-
-  // Get a copy of this VectorFst. See Fst<>::Copy() for further doc.
-  VectorFst<Arc, State> *Copy(bool safe = false) const override {
-    return new VectorFst<Arc, State>(*this, safe);
-  }
-
-  VectorFst<Arc, State> &operator=(const VectorFst<Arc, State> &fst) {
-    SetImpl(fst.GetSharedImpl());
-    return *this;
-  }
-
-  VectorFst<Arc, State> &operator=(const Fst<Arc> &fst) override {
-    if (this != &fst) SetImpl(std::make_shared<Impl>(fst));
-    return *this;
-  }
-
-  // Reads a VectorFst from an input stream, returning nullptr on error.
-  static VectorFst<Arc, State> *Read(std::istream &strm,
-                                     const FstReadOptions &opts) {
-    auto *impl = Impl::Read(strm, opts);
-    return impl ? new VectorFst<Arc, State>(std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  // Read a VectorFst from a file, returning nullptr on error; empty filename
-  // reads from standard input.
-  static VectorFst<Arc, State> *Read(const string &filename) {
-    auto *impl = ImplToExpandedFst<Impl, MutableFst<Arc>>::Read(filename);
-    return impl ? new VectorFst<Arc, State>(std::shared_ptr<Impl>(impl))
-                : nullptr;
-  }
-
-  bool Write(std::ostream &strm, const FstWriteOptions &opts) const override {
-    return WriteFst(*this, strm, opts);
-  }
-
-  bool Write(const string &filename) const override {
-    return Fst<Arc>::WriteFile(filename);
-  }
-
-  template <class FST>
-  static bool WriteFst(const FST &fst, std::ostream &strm,
-                       const FstWriteOptions &opts);
-
-  void InitStateIterator(StateIteratorData<Arc> *data) const override {
-    GetImpl()->InitStateIterator(data);
-  }
-
-  void InitArcIterator(StateId s, ArcIteratorData<Arc> *data) const override {
-    GetImpl()->InitArcIterator(s, data);
-  }
-
-  inline void InitMutableArcIterator(StateId s,
-                                     MutableArcIteratorData<Arc> *) override;
-
-  using ImplToMutableFst<Impl, MutableFst<Arc>>::ReserveArcs;
-  using ImplToMutableFst<Impl, MutableFst<Arc>>::ReserveStates;
-
- private:
-  using ImplToMutableFst<Impl, MutableFst<Arc>>::GetImpl;
-  using ImplToMutableFst<Impl, MutableFst<Arc>>::MutateCheck;
-  using ImplToMutableFst<Impl, MutableFst<Arc>>::SetImpl;
-
-  explicit VectorFst(std::shared_ptr<Impl> impl)
-      : ImplToMutableFst<Impl>(impl) {}
-};
-
-// Writes FST to file in Vector format, potentially with a pass over the machine
-// before writing to compute number of states.
-template <class Arc, class State>
-template <class FST>
-bool VectorFst<Arc, State>::WriteFst(const FST &fst, std::ostream &strm,
-                                     const FstWriteOptions &opts) {
-  static constexpr int file_version = 2;
-  bool update_header = true;
-  FstHeader hdr;
-  hdr.SetStart(fst.Start());
-  hdr.SetNumStates(kNoStateId);
-  size_t start_offset = 0;
-  if (fst.Properties(kExpanded, false) || opts.stream_write ||
-      (start_offset = strm.tellp()) != -1) {
-    hdr.SetNumStates(CountStates(fst));
-    update_header = false;
-  }
-  const auto properties =
-      fst.Properties(kCopyProperties, false) | Impl::kStaticProperties;
-  internal::FstImpl<Arc>::WriteFstHeader(fst, strm, opts, file_version,
-                                         "vector", properties, &hdr);
-  StateId num_states = 0;
-  for (StateIterator<FST> siter(fst); !siter.Done(); siter.Next()) {
-    const auto s = siter.Value();
-    fst.Final(s).Write(strm);
-    const int64 narcs = fst.NumArcs(s);
-    WriteType(strm, narcs);
-    for (ArcIterator<FST> aiter(fst, s); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      WriteType(strm, arc.ilabel);
-      WriteType(strm, arc.olabel);
-      arc.weight.Write(strm);
-      WriteType(strm, arc.nextstate);
-    }
-    ++num_states;
-  }
-  strm.flush();
-  if (!strm) {
-    LOG(ERROR) << "VectorFst::Write: Write failed: " << opts.source;
-    return false;
-  }
-  if (update_header) {
-    hdr.SetNumStates(num_states);
-    return internal::FstImpl<Arc>::UpdateFstHeader(
-        fst, strm, opts, file_version, "vector", properties, &hdr,
-        start_offset);
-  } else {
-    if (num_states != hdr.NumStates()) {
-      LOG(ERROR) << "Inconsistent number of states observed during write";
-      return false;
-    }
-  }
-  return true;
-}
-
-// Specialization for VectorFst; see generic version in fst.h for sample usage
-// (but use the VectorFst type instead). This version should inline.
-template <class Arc, class State>
-class StateIterator<VectorFst<Arc, State>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  explicit StateIterator(const VectorFst<Arc, State> &fst)
-      : nstates_(fst.GetImpl()->NumStates()), s_(0) {}
-
-  bool Done() const { return s_ >= nstates_; }
-
-  StateId Value() const { return s_; }
-
-  void Next() { ++s_; }
-
-  void Reset() { s_ = 0; }
-
- private:
-  const StateId nstates_;
-  StateId s_;
-};
-
-// Specialization for VectorFst; see generic version in fst.h for sample usage
-// (but use the VectorFst type instead). This version should inline.
-template <class Arc, class State>
-class ArcIterator<VectorFst<Arc, State>> {
- public:
-  using StateId = typename Arc::StateId;
-
-  ArcIterator(const VectorFst<Arc, State> &fst, StateId s)
-      : arcs_(fst.GetImpl()->GetState(s)->Arcs()),
-        narcs_(fst.GetImpl()->GetState(s)->NumArcs()),
-        i_(0) {}
-
-  bool Done() const { return i_ >= narcs_; }
-
-  const Arc &Value() const { return arcs_[i_]; }
-
-  void Next() { ++i_; }
-
-  void Reset() { i_ = 0; }
-
-  void Seek(size_t a) { i_ = a; }
-
-  size_t Position() const { return i_; }
-
-  constexpr uint32 Flags() const { return kArcValueFlags; }
-
-  void SetFlags(uint32, uint32) {}
-
- private:
-  const Arc *arcs_;
-  size_t narcs_;
-  size_t i_;
-};
-
-// Specialization for VectorFst; see generic version in mutable-fst.h for sample
-// usage (but use the VectorFst type instead). This version should inline.
-template <class Arc, class State>
-class MutableArcIterator<VectorFst<Arc, State>>
-    : public MutableArcIteratorBase<Arc> {
- public:
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-
-  MutableArcIterator(VectorFst<Arc, State> *fst, StateId s) : i_(0) {
-    fst->MutateCheck();
-    state_ = fst->GetMutableImpl()->GetState(s);
-    properties_ = &fst->GetImpl()->properties_;
-  }
-
-  bool Done() const final { return i_ >= state_->NumArcs(); }
-
-  const Arc &Value() const final { return state_->GetArc(i_); }
-
-  void Next() final { ++i_; }
-
-  size_t Position() const final { return i_; }
-
-  void Reset() final { i_ = 0; }
-
-  void Seek(size_t a) final { i_ = a; }
-
-  void SetValue(const Arc &arc) final {
-    const auto &oarc = state_->GetArc(i_);
-    if (oarc.ilabel != oarc.olabel) *properties_ &= ~kNotAcceptor;
-    if (oarc.ilabel == 0) {
-      *properties_ &= ~kIEpsilons;
-      if (oarc.olabel == 0) *properties_ &= ~kEpsilons;
-    }
-    if (oarc.olabel == 0) *properties_ &= ~kOEpsilons;
-    if (oarc.weight != Weight::Zero() && oarc.weight != Weight::One()) {
-      *properties_ &= ~kWeighted;
-    }
-    state_->SetArc(arc, i_);
-    if (arc.ilabel != arc.olabel) {
-      *properties_ |= kNotAcceptor;
-      *properties_ &= ~kAcceptor;
-    }
-    if (arc.ilabel == 0) {
-      *properties_ |= kIEpsilons;
-      *properties_ &= ~kNoIEpsilons;
-      if (arc.olabel == 0) {
-        *properties_ |= kEpsilons;
-        *properties_ &= ~kNoEpsilons;
-      }
-    }
-    if (arc.olabel == 0) {
-      *properties_ |= kOEpsilons;
-      *properties_ &= ~kNoOEpsilons;
-    }
-    if (arc.weight != Weight::Zero() && arc.weight != Weight::One()) {
-      *properties_ |= kWeighted;
-      *properties_ &= ~kUnweighted;
-    }
-    *properties_ &= kSetArcProperties | kAcceptor | kNotAcceptor | kEpsilons |
-                    kNoEpsilons | kIEpsilons | kNoIEpsilons | kOEpsilons |
-                    kNoOEpsilons | kWeighted | kUnweighted;
-  }
-
-  uint32 Flags() const final { return kArcValueFlags; }
-
-  void SetFlags(uint32, uint32) final {}
-
- private:
-  State *state_;
-  uint64 *properties_;
-  size_t i_;
-};
-
-// Provides information needed for the generic mutable arc iterator.
-template <class Arc, class State>
-inline void VectorFst<Arc, State>::InitMutableArcIterator(
-    StateId s, MutableArcIteratorData<Arc> *data) {
-  data->base = new MutableArcIterator<VectorFst<Arc, State>>(this, s);
-}
-
-// A useful alias when using StdArc.
-using StdVectorFst = VectorFst<StdArc>;
-
-}  // namespace fst
-
-#endif  // FST_VECTOR_FST_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/verify.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/verify.h
deleted file mode 100644
index 69837441..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/verify.h
+++ /dev/null
@@ -1,102 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Function to verify an FST's contents.
-
-#ifndef FST_VERIFY_H_
-#define FST_VERIFY_H_
-
-#include <fst/log.h>
-
-#include <fst/fst.h>
-#include <fst/test-properties.h>
-
-
-namespace fst {
-
-// Verifies that an Fst's contents are sane.
-template <class Arc>
-bool Verify(const Fst<Arc> &fst, bool allow_negative_labels = false) {
-  using Label = typename Arc::Label;
-  using StateId = typename Arc::StateId;
-  using Weight = typename Arc::Weight;
-  const auto start = fst.Start();
-  const auto *isyms = fst.InputSymbols();
-  const auto *osyms = fst.OutputSymbols();
-  // Count states
-  StateId ns = 0;
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) ++ns;
-  if (start == kNoStateId && ns > 0) {
-    LOG(ERROR) << "Verify: FST start state ID not set";
-    return false;
-  } else if (start >= ns) {
-    LOG(ERROR) << "Verify: FST start state ID exceeds number of states";
-    return false;
-  }
-  for (StateIterator<Fst<Arc>> siter(fst); !siter.Done(); siter.Next()) {
-    auto state = siter.Value();
-    size_t na = 0;
-    for (ArcIterator<Fst<Arc>> aiter(fst, state); !aiter.Done(); aiter.Next()) {
-      const auto &arc = aiter.Value();
-      if (!allow_negative_labels && arc.ilabel < 0) {
-        LOG(ERROR) << "Verify: FST input label ID of arc at position " << na
-                   << " of state " << state << " is negative";
-        return false;
-      } else if (isyms && isyms->Find(arc.ilabel) == "") {
-        LOG(ERROR) << "Verify: FST input label ID " << arc.ilabel
-                   << " of arc at position " << na << " of state " << state
-                   << " is missing from input symbol table \"" << isyms->Name()
-                   << "\"";
-        return false;
-      } else if (!allow_negative_labels && arc.olabel < 0) {
-        LOG(ERROR) << "Verify: FST output label ID of arc at position " << na
-                   << " of state " << state << " is negative";
-        return false;
-      } else if (osyms && osyms->Find(arc.olabel) == "") {
-        LOG(ERROR) << "Verify: FST output label ID " << arc.olabel
-                   << " of arc at position " << na << " of state " << state
-                   << " is missing from output symbol table \"" << osyms->Name()
-                   << "\"";
-        return false;
-      } else if (!arc.weight.Member()) {
-        LOG(ERROR) << "Verify: FST weight of arc at position " << na
-                   << " of state " << state << " is invalid";
-        return false;
-      } else if (arc.nextstate < 0) {
-        LOG(ERROR) << "Verify: FST destination state ID of arc at position "
-                   << na << " of state " << state << " is negative";
-        return false;
-      } else if (arc.nextstate >= ns) {
-        LOG(ERROR) << "Verify: FST destination state ID of arc at position "
-                   << na << " of state " << state
-                   << " exceeds number of states";
-        return false;
-      }
-      ++na;
-    }
-    if (!fst.Final(state).Member()) {
-      LOG(ERROR) << "Verify: FST final weight of state " << state
-                 << " is invalid";
-      return false;
-    }
-  }
-  const auto fst_props = fst.Properties(kFstProperties, false);
-  if (fst_props & kError) {
-    LOG(ERROR) << "Verify: FST error property is set";
-    return false;
-  }
-  uint64 known_props;
-  uint64 test_props =
-      ComputeProperties(fst, kFstProperties, &known_props, false);
-  if (!CompatProperties(fst_props, test_props)) {
-    LOG(ERROR) << "Verify: Stored FST properties incorrect "
-               << "(props1 = stored props, props2 = tested)";
-    return false;
-  } else {
-    return true;
-  }
-}
-
-}  // namespace fst
-
-#endif  // FST_VERIFY_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/visit.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/visit.h
deleted file mode 100644
index c6658047..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/visit.h
+++ /dev/null
@@ -1,321 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// Queue-dependent visitation of finite-state transducers. See also dfs-visit.h.
-
-#ifndef FST_VISIT_H_
-#define FST_VISIT_H_
-
-
-#include <fst/arcfilter.h>
-#include <fst/mutable-fst.h>
-
-
-namespace fst {
-
-// Visitor Interface: class determining actions taken during a visit. If any of
-// the boolean member functions return false, the visit is aborted by first
-// calling FinishState() on all unfinished (grey) states and then calling
-// FinishVisit().
-//
-// Note this is more general than the visitor interface in dfs-visit.h but lacks
-// some DFS-specific behavior.
-//
-// template <class Arc>
-// class Visitor {
-//  public:
-//   using StateId = typename Arc::StateId;
-//
-//   Visitor(T *return_data);
-//
-//   // Invoked before visit.
-//   void InitVisit(const Fst<Arc> &fst);
-//
-//   // Invoked when state discovered (2nd arg is visitation root).
-//   bool InitState(StateId s, StateId root);
-//
-//   // Invoked when arc to white/undiscovered state examined.
-//   bool WhiteArc(StateId s, const Arc &arc);
-//
-//   // Invoked when arc to grey/unfinished state examined.
-//   bool GreyArc(StateId s, const Arc &arc);
-//
-//   // Invoked when arc to black/finished state examined.
-//   bool BlackArc(StateId s, const Arc &arc);
-//
-//   // Invoked when state finished.
-//   void FinishState(StateId s);
-//
-//   // Invoked after visit.
-//   void FinishVisit();
-// };
-
-// Performs queue-dependent visitation. Visitor class argument determines
-// actions and contains any return data. ArcFilter determines arcs that are
-// considered. If 'access_only' is true, performs visitation only to states
-// accessible from the initial state.
-template <class FST, class Visitor, class Queue, class ArcFilter>
-void Visit(const FST &fst, Visitor *visitor, Queue *queue, ArcFilter filter,
-           bool access_only = false) {
-  using Arc = typename FST::Arc;
-  using StateId = typename Arc::StateId;
-  visitor->InitVisit(fst);
-  const auto start = fst.Start();
-  if (start == kNoStateId) {
-    visitor->FinishVisit();
-    return;
-  }
-  // An FST's state's visit color.
-  static constexpr uint8 kWhiteState = 0x01;  // Undiscovered.
-  static constexpr uint8 kGreyState = 0x02;   // Discovered & unfinished.
-  static constexpr uint8 kBlackState = 0x04;  // Finished.
-  // We destroy an iterator as soon as possible and mark it so.
-  static constexpr uint8 kArcIterDone = 0x08;
-  std::vector<uint8> state_status;
-  std::vector<ArcIterator<FST> *> arc_iterator;
-  MemoryPool<ArcIterator<FST>> aiter_pool;
-  StateId nstates = start + 1;  // Number of known states in general case.
-  bool expanded = false;
-  if (fst.Properties(kExpanded, false)) {  // Tests if expanded, then uses
-    nstates = CountStates(fst);            // ExpandedFst::NumStates().
-    expanded = true;
-  }
-  state_status.resize(nstates, kWhiteState);
-  arc_iterator.resize(nstates);
-  StateIterator<Fst<Arc>> siter(fst);
-  // Continues visit while true.
-  bool visit = true;
-  // Iterates over trees in visit forest.
-  for (auto root = start; visit && root < nstates;) {
-    visit = visitor->InitState(root, root);
-    state_status[root] = kGreyState;
-    queue->Enqueue(root);
-    while (!queue->Empty()) {
-      auto state = queue->Head();
-      if (state >= state_status.size()) {
-        nstates = state + 1;
-        state_status.resize(nstates, kWhiteState);
-        arc_iterator.resize(nstates);
-      }
-      // Creates arc iterator if needed.
-      if (!arc_iterator[state] && !(state_status[state] & kArcIterDone) &&
-          visit) {
-        arc_iterator[state] = new (&aiter_pool) ArcIterator<FST>(fst, state);
-      }
-      // Deletes arc iterator if done.
-      auto *aiter = arc_iterator[state];
-      if ((aiter && aiter->Done()) || !visit) {
-        Destroy(aiter, &aiter_pool);
-        arc_iterator[state] = nullptr;
-        state_status[state] |= kArcIterDone;
-      }
-      // Dequeues state and marks black if done.
-      if (state_status[state] & kArcIterDone) {
-        queue->Dequeue();
-        visitor->FinishState(state);
-        state_status[state] = kBlackState;
-        continue;
-      }
-      const auto &arc = aiter->Value();
-      if (arc.nextstate >= state_status.size()) {
-        nstates = arc.nextstate + 1;
-        state_status.resize(nstates, kWhiteState);
-        arc_iterator.resize(nstates);
-      }
-      // Visits respective arc types.
-      if (filter(arc)) {
-        // Enqueues destination state and marks grey if white.
-        if (state_status[arc.nextstate] == kWhiteState) {
-          visit = visitor->WhiteArc(state, arc);
-          if (!visit) continue;
-          visit = visitor->InitState(arc.nextstate, root);
-          state_status[arc.nextstate] = kGreyState;
-          queue->Enqueue(arc.nextstate);
-        } else if (state_status[arc.nextstate] == kBlackState) {
-          visit = visitor->BlackArc(state, arc);
-        } else {
-          visit = visitor->GreyArc(state, arc);
-        }
-      }
-      aiter->Next();
-      // Destroys an iterator ASAP for efficiency.
-      if (aiter->Done()) {
-        Destroy(aiter, &aiter_pool);
-        arc_iterator[state] = nullptr;
-        state_status[state] |= kArcIterDone;
-      }
-    }
-    if (access_only) break;
-    // Finds next tree root.
-    for (root = (root == start) ? 0 : root + 1;
-         root < nstates && state_status[root] != kWhiteState; ++root) {
-    }
-    // Check for a state beyond the largest known state.
-    if (!expanded && root == nstates) {
-      for (; !siter.Done(); siter.Next()) {
-        if (siter.Value() == nstates) {
-          ++nstates;
-          state_status.push_back(kWhiteState);
-          arc_iterator.push_back(nullptr);
-          break;
-        }
-      }
-    }
-  }
-  visitor->FinishVisit();
-}
-
-template <class Arc, class Visitor, class Queue>
-inline void Visit(const Fst<Arc> &fst, Visitor *visitor, Queue *queue) {
-  Visit(fst, visitor, queue, AnyArcFilter<Arc>());
-}
-
-// Copies input FST to mutable FST following queue order.
-template <class A>
-class CopyVisitor {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  explicit CopyVisitor(MutableFst<Arc> *ofst) : ifst_(nullptr), ofst_(ofst) {}
-
-  void InitVisit(const Fst<A> &ifst) {
-    ifst_ = &ifst;
-    ofst_->DeleteStates();
-    ofst_->SetStart(ifst_->Start());
-  }
-
-  bool InitState(StateId state, StateId) {
-    while (ofst_->NumStates() <= state) ofst_->AddState();
-    return true;
-  }
-
-  bool WhiteArc(StateId state, const Arc &arc) {
-    ofst_->AddArc(state, arc);
-    return true;
-  }
-
-  bool GreyArc(StateId state, const Arc &arc) {
-    ofst_->AddArc(state, arc);
-    return true;
-  }
-
-  bool BlackArc(StateId state, const Arc &arc) {
-    ofst_->AddArc(state, arc);
-    return true;
-  }
-
-  void FinishState(StateId state) {
-    ofst_->SetFinal(state, ifst_->Final(state));
-  }
-
-  void FinishVisit() {}
-
- private:
-  const Fst<Arc> *ifst_;
-  MutableFst<Arc> *ofst_;
-};
-
-// Visits input FST up to a state limit following queue order.
-template <class A>
-class PartialVisitor {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  explicit PartialVisitor(StateId maxvisit)
-      : fst_(nullptr), maxvisit_(maxvisit) {}
-
-  void InitVisit(const Fst<A> &ifst) {
-    fst_ = &ifst;
-    ninit_ = 0;
-    nfinish_ = 0;
-  }
-
-  bool InitState(StateId state, StateId root) {
-    ++ninit_;
-    return ninit_ <= maxvisit_;
-  }
-
-  bool WhiteArc(StateId state, const Arc &arc) { return true; }
-
-  bool GreyArc(StateId state, const Arc &arc) { return true; }
-
-  bool BlackArc(StateId state, const Arc &arc) { return true; }
-
-  void FinishState(StateId state) {
-    fst_->Final(state);  // Visits super-final arc.
-    ++nfinish_;
-  }
-
-  void FinishVisit() {}
-
-  StateId NumInitialized() { return ninit_; }
-
-  StateId NumFinished() { return nfinish_; }
-
- private:
-  const Fst<Arc> *fst_;
-  StateId maxvisit_;
-  StateId ninit_;
-  StateId nfinish_;
-};
-
-// Copies input FST to mutable FST up to a state limit following queue order.
-template <class A>
-class PartialCopyVisitor : public CopyVisitor<A> {
- public:
-  using Arc = A;
-  using StateId = typename Arc::StateId;
-
-  using CopyVisitor<A>::WhiteArc;
-
-  PartialCopyVisitor(MutableFst<Arc> *ofst, StateId maxvisit,
-                     bool copy_grey = true, bool copy_black = true)
-      : CopyVisitor<A>(ofst), maxvisit_(maxvisit),
-        copy_grey_(copy_grey), copy_black_(copy_black) {}
-
-  void InitVisit(const Fst<A> &ifst) {
-    CopyVisitor<A>::InitVisit(ifst);
-    ninit_ = 0;
-    nfinish_ = 0;
-  }
-
-  bool InitState(StateId state, StateId root) {
-    CopyVisitor<A>::InitState(state, root);
-    ++ninit_;
-    return ninit_ <= maxvisit_;
-  }
-
-  bool GreyArc(StateId state, const Arc &arc) {
-    if (copy_grey_) return CopyVisitor<A>::GreyArc(state, arc);
-    return true;
-  }
-
-  bool BlackArc(StateId state, const Arc &arc) {
-    if (copy_black_) return CopyVisitor<A>::BlackArc(state, arc);
-    return true;
-  }
-
-  void FinishState(StateId state) {
-    CopyVisitor<A>::FinishState(state);
-    ++nfinish_;
-  }
-
-  void FinishVisit() {}
-
-  StateId NumInitialized() { return ninit_; }
-
-  StateId NumFinished() { return nfinish_; }
-
- private:
-  StateId maxvisit_;
-  StateId ninit_;
-  StateId nfinish_;
-  const bool copy_grey_;
-  const bool copy_black_;
-};
-
-}  // namespace fst
-
-#endif  // FST_VISIT_H_
diff --git a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/weight.h b/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/weight.h
deleted file mode 100644
index a72dde0a..00000000
--- a/projects/llm_framework/include/sherpa/sherpa-onnx/openfst-src/fst/weight.h
+++ /dev/null
@@ -1,358 +0,0 @@
-// See www.openfst.org for extensive documentation on this weighted
-// finite-state transducer library.
-//
-// General weight set and associated semiring operation definitions.
-
-#ifndef FST_WEIGHT_H_
-#define FST_WEIGHT_H_
-
-#include <cctype>
-#include <cmath>
-#include <iostream>
-#include <sstream>
-#include <utility>
-
-#include <fst/compat.h>
-#include <fst/log.h>
-
-#include <fst/util.h>
-
-
-DECLARE_string(fst_weight_parentheses);
-DECLARE_string(fst_weight_separator);
-
-namespace fst {
-
-// A semiring is specified by two binary operations Plus and Times and two
-// designated elements Zero and One with the following properties:
-//
-//   Plus: associative, commutative, and has Zero as its identity.
-//
-//   Times: associative and has identity One, distributes w.r.t. Plus, and
-//     has Zero as an annihilator:
-//          Times(Zero(), a) == Times(a, Zero()) = Zero().
-//
-// A left semiring distributes on the left; a right semiring is similarly
-// defined.
-//
-// A Weight class must have binary functions Plus and Times and static member
-// functions Zero() and One() and these must form (at least) a left or right
-// semiring.
-//
-// In addition, the following should be defined for a Weight:
-//
-//   Member: predicate on set membership.
-//
-//   NoWeight: static member function that returns an element that is
-//      not a set member; used to signal an error.
-//
-//   >>: reads textual representation of a weight.
-//
-//   <<: prints textual representation of a weight.
-//
-//   Read(istream &istrm): reads binary representation of a weight.
-//
-//   Write(ostream &ostrm): writes binary representation of a weight.
-//
-//   Hash: maps weight to size_t.
-//
-//   ApproxEqual: approximate equality (for inexact weights)
-//
-//   Quantize: quantizes w.r.t delta (for inexact weights)
-//
-//   Divide: for all a, b, c s.t. Times(a, b) == c
-//
-//     --> b' = Divide(c, a, DIVIDE_LEFT) if a left semiring, b'.Member()
-//      and Times(a, b') == c
-//     --> a' = Divide(c, b, DIVIDE_RIGHT) if a right semiring, a'.Member()
-//      and Times(a', b) == c
-//     --> b' = Divide(c, a) = Divide(c, a, DIVIDE_ANY) =
-//      Divide(c, a, DIVIDE_LEFT) = Divide(c, a, DIVIDE_RIGHT) if a
-//      commutative semiring, b'.Member() and Times(a, b') = Times(b', a) = c
-//
-//   ReverseWeight: the type of the corresponding reverse weight.
-//
-//     Typically the same type as Weight for a (both left and right) semiring.
-//     For the left string semiring, it is the right string semiring.
-//
-//   Reverse: a mapping from Weight to ReverseWeight s.t.
-//
-//     --> Reverse(Reverse(a)) = a
-//     --> Reverse(Plus(a, b)) = Plus(Reverse(a), Reverse(b))
-//     --> Reverse(Times(a, b)) = Times(Reverse(b), Reverse(a))
-//     Typically the identity mapping in a (both left and right) semiring.
-//     In the left string semiring, it maps to the reverse string in the right
-//     string semiring.
-//
-//   Properties: specifies additional properties that hold:
-//      LeftSemiring: indicates weights form a left semiring.
-//      RightSemiring: indicates weights form a right semiring.
-//      Commutative: for all a,b: Times(a,b) == Times(b,a)
-//      Idempotent: for all a: Plus(a, a) == a.
-//      Path: for all a, b: Plus(a, b) == a or Plus(a, b) == b.
-
-// CONSTANT DEFINITIONS
-
-// A representable float near .001.
-constexpr float kDelta = 1.0F / 1024.0F;
-
-// For all a, b, c: Times(c, Plus(a, b)) = Plus(Times(c, a), Times(c, b)).
-constexpr uint64 kLeftSemiring = 0x0000000000000001ULL;
-
-// For all a, b, c: Times(Plus(a, b), c) = Plus(Times(a, c), Times(b, c)).
-constexpr uint64 kRightSemiring = 0x0000000000000002ULL;
-
-constexpr uint64 kSemiring = kLeftSemiring | kRightSemiring;
-
-// For all a, b: Times(a, b) = Times(b, a).
-constexpr uint64 kCommutative = 0x0000000000000004ULL;
-
-// For all a: Plus(a, a) = a.
-constexpr uint64 kIdempotent = 0x0000000000000008ULL;
-
-// For all a, b: Plus(a, b) = a or Plus(a, b) = b.
-constexpr uint64 kPath = 0x0000000000000010ULL;
-
-// For random weight generation: default number of distinct weights.
-// This is also used for a few other weight generation defaults.
-constexpr size_t kNumRandomWeights = 5;
-
-// Determines direction of division.
-enum DivideType {
-  DIVIDE_LEFT,   // left division
-  DIVIDE_RIGHT,  // right division
-  DIVIDE_ANY
-};  // division in a commutative semiring
-
-// NATURAL ORDER
-//
-// By definition:
-//
-//                 a <= b iff a + b = a
-//
-// The natural order is a negative partial order iff the semiring is
-// idempotent. It is trivially monotonic for plus. It is left
-// (resp. right) monotonic for times iff the semiring is left
-// (resp. right) distributive. It is a total order iff the semiring
-// has the path property.
-//
-// For more information, see:
-//
-// Mohri, M. 2002. Semiring framework and algorithms for shortest-distance
-// problems, Journal of Automata, Languages and
-// Combinatorics 7(3): 321-350, 2002.
-//
-// We define the strict version of this order below.
-
-template <class W>
-class NaturalLess {
- public:
-  using Weight = W;
-
-  NaturalLess() {
-    // TODO(kbg): Make this a compile-time static_assert once we have a pleasant
-    // way to "deregister" this operation for non-path semirings so an
-    // informative error message is produced.
-    if (!(W::Properties() & kIdempotent)) {
-      FSTERROR() << "NaturalLess: Weight type is not idempotent: " << W::Type();
-    }
-  }
-
-  bool operator()(const W &w1, const W &w2) const {
-    return (Plus(w1, w2) == w1) && w1 != w2;
-  }
-};
-
-// Power is the iterated product for arbitrary semirings such that
-// Power(w, 0) is One() for the semiring, and Power(w, n) =
-// Times(Power(w, n-1), w).
-
-template <class Weight>
-Weight Power(Weight w, size_t n) {
-  auto result = Weight::One();
-  for (size_t i = 0; i < n; ++i) result = Times(result, w);
-  return result;
-}
-
-// Simple default adder class. Specializations might be more complex.
-template <class Weight>
-class Adder {
- public:
-  explicit Adder(Weight w = Weight::Zero()) : sum_(w) { }
-
-  Weight Add(const Weight &w) {
-    sum_ = Plus(sum_, w);
-    return sum_;
-  }
-
-  Weight Sum() { return sum_; }
-
-  void Reset(Weight w = Weight::Zero()) { sum_ = w; }
-
- private:
-  Weight sum_;
-};
-
-// General weight converter: raises error.
-template <class W1, class W2>
-struct WeightConvert {
-  W2 operator()(W1 w1) const {
-    FSTERROR() << "WeightConvert: Can't convert weight from \"" << W1::Type()
-               << "\" to \"" << W2::Type();
-    return W2::NoWeight();
-  }
-};
-
-// Specialized weight converter to self.
-template <class W>
-struct WeightConvert<W, W> {
-  W operator()(W weight) const { return weight; }
-};
-
-// General random weight generator: raises error.
-template <class W>
-struct WeightGenerate {
-  W operator()() const {
-    FSTERROR() << "WeightGenerate: No random generator for " << W::Type();
-    return W::NoWeight();
-  }
-};
-
-namespace internal {
-
-class CompositeWeightIO {
- public:
-  CompositeWeightIO();
-  CompositeWeightIO(char separator, std::pair<char, char> parentheses);
-
-  std::pair<char, char> parentheses() const {
-    return {open_paren_, close_paren_};
-  }
-  char separator() const { return separator_; }
-
-  bool error() const { return error_; }
-
- protected:
-  const char separator_;
-  const char open_paren_;
-  const char close_paren_;
-
- private:
-  bool error_;
-};
-
-}  // namespace internal
-
-// Helper class for writing textual composite weights.
-class CompositeWeightWriter : public internal::CompositeWeightIO {
- public:
-  // Uses configuration from flags (FLAGS_fst_weight_separator,
-  // FLAGS_fst_weight_parentheses).
-  explicit CompositeWeightWriter(std::ostream &ostrm);
-
-  // parentheses defines the opening and closing parenthesis characters.
-  // Set parentheses = {0, 0} to disable writing parenthesis.
-  CompositeWeightWriter(std::ostream &ostrm, char separator,
-                        std::pair<char, char> parentheses);
-
-  CompositeWeightWriter(const CompositeWeightWriter &) = delete;
-  CompositeWeightWriter &operator=(const CompositeWeightWriter &) = delete;
-
-  // Writes open parenthesis to a stream if option selected.
-  void WriteBegin();
-
-  // Writes element to a stream.
-  template <class T>
-  void WriteElement(const T &comp) {
-    if (i_++ > 0) ostrm_ << separator_;
-    ostrm_ << comp;
-  }
-
-  // Writes close parenthesis to a stream if option selected.
-  void WriteEnd();
-
- private:
-  std::ostream &ostrm_;
-  int i_ = 0;  // Element position.
-};
-
-// Helper class for reading textual composite weights. Elements are separated by
-// a separator character. There must be at least one element per textual
-// representation.  Parentheses characters should be set if the composite
-// weights themselves contain composite weights to ensure proper parsing.
-class CompositeWeightReader : public internal::CompositeWeightIO {
- public:
-  // Uses configuration from flags (FLAGS_fst_weight_separator,
-  // FLAGS_fst_weight_parentheses).
-  explicit CompositeWeightReader(std::istream &istrm);
-
-  // parentheses defines the opening and closing parenthesis characters.
-  // Set parentheses = {0, 0} to disable reading parenthesis.
-  CompositeWeightReader(std::istream &istrm, char separator,
-                        std::pair<char, char> parentheses);
-
-  CompositeWeightReader(const CompositeWeightReader &) = delete;
-  CompositeWeightReader &operator=(const CompositeWeightReader &) = delete;
-
-  // Reads open parenthesis from a stream if option selected.
-  void ReadBegin();
-
-  // Reads element from a stream. The second argument, when true, indicates that
-  // this will be the last element (allowing more forgiving formatting of the
-  // last element). Returns false when last element is read.
-  template <class T>
-  bool ReadElement(T *comp, bool last = false);
-
-  // Finalizes reading.
-  void ReadEnd();
-
- private:
-  std::istream &istrm_;  // Input stream.
-  int c_ = 0;            // Last character read, or EOF.
-  int depth_ = 0;        // Weight parentheses depth.
-};
-
-template <class T>
-inline bool CompositeWeightReader::ReadElement(T *comp, bool last) {
-  string s;
-  const bool has_parens = open_paren_ != 0;
-  while ((c_ != std::istream::traits_type::eof()) && !std::isspace(c_) &&
-         (c_ != separator_ || depth_ > 1 || last) &&
-         (c_ != close_paren_ || depth_ != 1)) {
-    s += c_;
-    // If parentheses encountered before separator, they must be matched.
-    if (has_parens && c_ == open_paren_) {
-      ++depth_;
-    } else if (has_parens && c_ == close_paren_) {
-      // Failure on unmatched parentheses.
-      if (depth_ == 0) {
-        FSTERROR() << "CompositeWeightReader: Unmatched close paren: "
-                   << "Is the fst_weight_parentheses flag set correctly?";
-        istrm_.clear(std::ios::badbit);
-        return false;
-      }
-      --depth_;
-    }
-    c_ = istrm_.get();
-  }
-  if (s.empty()) {
-    FSTERROR() << "CompositeWeightReader: Empty element: "
-               << "Is the fst_weight_parentheses flag set correctly?";
-    istrm_.clear(std::ios::badbit);
-    return false;
-  }
-  std::istringstream istrm(s);
-  istrm >> *comp;
-  // Skips separator/close parenthesis.
-  if (c_ != std::istream::traits_type::eof() && !std::isspace(c_)) {
-    c_ = istrm_.get();
-  }
-  const bool is_eof = c_ == std::istream::traits_type::eof();
-  // Clears fail bit if just EOF.
-  if (is_eof && !istrm_.bad()) istrm_.clear(std::ios::eofbit);
-  return !is_eof && !std::isspace(c_);
-}
-
-}  // namespace fst
-
-#endif  // FST_WEIGHT_H_
diff --git a/projects/llm_framework/include/utils/checker.h b/projects/llm_framework/include/utils/checker.h
deleted file mode 100644
index e8ee8faa..00000000
--- a/projects/llm_framework/include/utils/checker.h
+++ /dev/null
@@ -1,32 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2023 Axera Semiconductor (Ningbo) Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor (Ningbo) Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor (Ningbo) Co., Ltd.
- *
- **************************************************************************************************/
-
-#ifndef AX_CHECKER_H
-#define AX_CHECKER_H
-
-#include "utils/logger.h"
-
-#define CHECK_PTR(p)                     \
-    do {                                 \
-        if (!p) {                        \
-            ALOGE("%s nil pointer\n", #p);      \
-            return -1; \
-        }                                \
-    } while (0)
-
-#define CHECK_INITED(p)                     \
-    do {                                    \
-        if (!p->HasInit()) {                \
-            ALOGE("%s has not init\n", #p); \
-            return -1;    \
-        }                                   \
-    } while (0)
-
-#endif //AX_CHECKER_H
diff --git a/projects/llm_framework/include/utils/io.hpp b/projects/llm_framework/include/utils/io.hpp
deleted file mode 100644
index e1aacd69..00000000
--- a/projects/llm_framework/include/utils/io.hpp
+++ /dev/null
@@ -1,601 +0,0 @@
-/*
- * Licensed to the Apache Software Foundation (ASF) under one
- * or more contributor license agreements.  See the NOTICE file
- * distributed with this work for additional information
- * regarding copyright ownership.  The ASF licenses this file
- * to you under the Apache License, Version 2.0 (the
- * License); you may not use this file except in compliance
- * with the License.  You may obtain a copy of the License at
- *
- *   http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing,
- * software distributed under the License is distributed on an
- * AS IS BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
- * KIND, either express or implied.  See the License for the
- * specific language governing permissions and limitations
- * under the License.
- */
-
-#pragma once
-
-#include <cstdio>
-#include <cstring>
-#include <vector>
-#include <array>
-#include <string>
-#include <fstream>
-#include <cstdint>
-#include <cstdlib>
-
-#include "utils/checker.h"
-#include "ax_sys_api.h"
-#include "ax_engine_type.h"
-
-
-#define IO_CMM_ALIGN_SIZE 128
-
-namespace utils {
-    typedef enum {
-        IO_BUFFER_STRATEGY_DEFAULT,
-        IO_BUFFER_STRATEGY_CACHED
-    } IO_BUFFER_STRATEGY_T;
-
-    static inline AX_S32 query_model_input_size(const AX_ENGINE_IO_INFO_T* io_info, std::array<int, 2> &input_size,
-                                                AX_IMG_FORMAT_E &eDtype) {
-        int height = 0;
-        int width = 0;
-        int size = 0;
-        int channel = 0;
-        int data_type_size = 0;
-        auto& input = io_info->pInputs[0];
-
-        switch (input.eLayout) {
-            case AX_ENGINE_TENSOR_LAYOUT_NHWC:
-                height = input.pShape[1];
-                width = input.pShape[2];
-                channel = input.pShape[3];
-                size = input.nSize;
-                break;
-            case AX_ENGINE_TENSOR_LAYOUT_NCHW:
-                channel = input.pShape[1];
-                height = input.pShape[2];
-                width = input.pShape[3];
-                size = input.nSize;
-                break;
-            default: // NHWC
-                height = input.pShape[1];
-                width = input.pShape[2];
-                channel = input.pShape[3];
-                size = input.nSize;
-                break;
-        }
-
-        switch (input.eDataType) {
-            case AX_ENGINE_DT_UINT8:
-            case AX_ENGINE_DT_SINT8:
-                data_type_size = 1;
-                break;
-            case AX_ENGINE_DT_UINT16:
-            case AX_ENGINE_DT_SINT16:
-                data_type_size = 2;
-                break;
-            case AX_ENGINE_DT_FLOAT32:
-                data_type_size = 4;
-                break;
-            case AX_ENGINE_DT_SINT32:
-            case AX_ENGINE_DT_UINT32:
-                data_type_size = 4;
-                break;
-            case AX_ENGINE_DT_FLOAT64:
-                data_type_size = 8;
-                break;
-            default:
-                data_type_size = 1;
-                break;
-        }
-
-        if (channel == 0 || height == 0 || width == 0 || size == 0 || data_type_size == 0) {
-            return -1;
-        }
-
-        if (input.pExtraMeta) {
-            switch (input.pExtraMeta->eColorSpace) {
-                case AX_ENGINE_CS_BGR:
-                    input_size[0] = height;
-                    input_size[1] = width;
-                    eDtype = AX_FORMAT_BGR888;
-                    break;
-                case AX_ENGINE_CS_RGB:
-                    input_size[0] = height;
-                    input_size[1] = width;
-                    eDtype = AX_FORMAT_RGB888;
-                    break;
-                case AX_ENGINE_CS_NV12:
-                    input_size[0] = height * 2 / 3;
-                    input_size[1] = width;
-                    eDtype = AX_FORMAT_YUV420_SEMIPLANAR;
-                    break;
-                case AX_ENGINE_CS_NV21:
-                    input_size[0] = height * 2 / 3;
-                    input_size[1] = width;
-                    eDtype = AX_FORMAT_YUV420_SEMIPLANAR_VU;
-                    break;
-                default: // AX_ENGINE_CS_NV12
-                    input_size[0] = height * 2 / 3;
-                    input_size[1] = width;
-                    eDtype = AX_FORMAT_YUV420_SEMIPLANAR;
-                    break;
-            }
-        }
-        else {
-            input_size[0] = height * 2 / 3;
-            input_size[1] = width;
-            eDtype = AX_FORMAT_YUV420_SEMIPLANAR;
-        }
-
-        ALOGD("eLayout:%d, eDataType:%d, channel:%d, height:%d, width:%d, size:%d, data_type_size:%d",
-              input.eLayout, input.eDataType, channel, input_size[0], input_size[1], size, data_type_size);
-
-        return 0;
-    }
-
-    static inline void brief_io_info(std::string strModel, const AX_ENGINE_IO_INFO_T* io_info) {
-        auto describe_shape_type = [](AX_ENGINE_TENSOR_LAYOUT_T type) -> const char* {
-            switch (type) {
-                case AX_ENGINE_TENSOR_LAYOUT_NHWC:
-                    return "NHWC";
-                case AX_ENGINE_TENSOR_LAYOUT_NCHW:
-                    return "NCHW";
-                default:
-                    return "unknown";
-            }
-        };
-        auto describe_data_type = [](AX_ENGINE_DATA_TYPE_T type) -> const char* {
-            switch (type) {
-                case AX_ENGINE_DT_UINT8:
-                    return "uint8";
-                case AX_ENGINE_DT_UINT16:
-                    return "uint16";
-                case AX_ENGINE_DT_FLOAT32:
-                    return "float32";
-                case AX_ENGINE_DT_SINT16:
-                    return "sint16";
-                case AX_ENGINE_DT_SINT8:
-                    return "sint8";
-                case AX_ENGINE_DT_SINT32:
-                    return "sint32";
-                case AX_ENGINE_DT_UINT32:
-                    return "uint32";
-                case AX_ENGINE_DT_FLOAT64:
-                    return "float64";
-                case AX_ENGINE_DT_UINT10_PACKED:
-                    return "uint10_packed";
-                case AX_ENGINE_DT_UINT12_PACKED:
-                    return "uint12_packed";
-                case AX_ENGINE_DT_UINT14_PACKED:
-                    return "uint14_packed";
-                case AX_ENGINE_DT_UINT16_PACKED:
-                    return "uint16_packed";
-                default:
-                    return "unknown";
-            }
-        };
-        auto describe_memory_type = [](AX_ENGINE_MEMORY_TYPE_T type) -> const char* {
-            switch (type) {
-                case AX_ENGINE_MT_PHYSICAL:
-                    return "Physical";
-                case AX_ENGINE_MT_VIRTUAL:
-                    return "Virtual";
-                default:
-                    return "unknown";
-            }
-        };
-        auto describe_color_space = [](AX_ENGINE_COLOR_SPACE_T cs) -> const char* {
-            switch (cs) {
-                case AX_ENGINE_CS_FEATUREMAP:
-                    return "FeatureMap";
-                case AX_ENGINE_CS_BGR:
-                    return "BGR";
-                case AX_ENGINE_CS_RGB:
-                    return "RGB";
-                case AX_ENGINE_CS_RGBA:
-                    return "RGBA";
-                case AX_ENGINE_CS_GRAY:
-                    return "GRAY";
-                case AX_ENGINE_CS_NV12:
-                    return "NV12";
-                case AX_ENGINE_CS_NV21:
-                    return "NV21";
-                case AX_ENGINE_CS_YUV444:
-                    return "YUV444";
-                case AX_ENGINE_CS_RAW8:
-                    return "RAW8";
-                case AX_ENGINE_CS_RAW10:
-                    return "RAW10";
-                case AX_ENGINE_CS_RAW12:
-                    return "RAW12";
-                case AX_ENGINE_CS_RAW14:
-                    return "RAW14";
-                case AX_ENGINE_CS_RAW16:
-                    return "RAW16";
-                default:
-                    return "unknown";
-            }
-        };
-        printf("Model Name: %s\n", strModel.c_str());
-        printf("Max Batch Size %d\n", io_info->nMaxBatchSize);
-        printf("Support Dynamic Batch? %s\n", io_info->bDynamicBatchSize == AX_TRUE ? "Yes" : "No");
-
-        for (uint32_t i = 0; i < io_info->nInputSize; ++i) {
-            auto& input = io_info->pInputs[i];
-            printf("Input[%d]: %s\n", i, input.pName);
-            printf("    Shape [");
-            for (uint32_t j = 0; j < input.nShapeSize; ++j) {
-                printf("%d", (int)input.pShape[j]);
-                if (j + 1 < input.nShapeSize) printf(", ");
-            }
-            printf("] %s %s %s %s\n", describe_shape_type(input.eLayout), describe_data_type(input.eDataType),
-                   input.pExtraMeta ? describe_color_space(input.pExtraMeta->eColorSpace) : "",
-                   input.nQuantizationValue > 0 ? ("Q=" + std::to_string(input.nQuantizationValue)).c_str() : "");
-            printf("    Memory %s\n", describe_memory_type(input.eMemoryType));
-            printf("    Size %u\n", input.nSize);
-        }
-        for (uint32_t i = 0; i < io_info->nOutputSize; ++i) {
-            auto& output = io_info->pOutputs[i];
-            printf("Output[%d]: %s\n", i, output.pName);
-            printf("    Shape [");
-            for (uint32_t j = 0; j < output.nShapeSize; ++j) {
-                printf("%d", (int)output.pShape[j]);
-                if (j + 1 < output.nShapeSize) printf(", ");
-            }
-            printf("] %s %s %s\n", describe_shape_type(output.eLayout), describe_data_type(output.eDataType),
-                   output.nQuantizationValue > 0 ? ("Q=" + std::to_string(output.nQuantizationValue)).c_str() : "");
-            printf("    Memory %s\n", describe_memory_type(output.eMemoryType));
-            printf("    Size %u\n", output.nSize);
-        }
-    }
-
-    static inline AX_S32 alloc_engine_buffer(const std::string& token, const std::string& appendix, size_t index, const AX_ENGINE_IOMETA_T* pMeta, AX_ENGINE_IO_BUFFER_T* pBuf, IO_BUFFER_STRATEGY_T eStrategy = IO_BUFFER_STRATEGY_DEFAULT) {
-        AX_S32 ret = -1;
-        if (eStrategy != IO_BUFFER_STRATEGY_DEFAULT && eStrategy != IO_BUFFER_STRATEGY_CACHED) {
-            fprintf(stderr, "strategy %d not supported\n", (int)eStrategy);
-            return -1;
-        }
-        memset(pBuf, 0, sizeof(AX_ENGINE_IO_BUFFER_T));
-        pBuf->nSize = pMeta->nSize;
-
-        const std::string token_name = "skel_" + token + appendix + std::to_string(index);
-
-        if (eStrategy == IO_BUFFER_STRATEGY_CACHED) {
-            ret = AX_SYS_MemAllocCached((AX_U64*)&pBuf->phyAddr, &pBuf->pVirAddr, pBuf->nSize, IO_CMM_ALIGN_SIZE, (const AX_S8*)token_name.c_str());
-        }
-        else {
-            ret = AX_SYS_MemAlloc((AX_U64*)&pBuf->phyAddr, &pBuf->pVirAddr, pBuf->nSize, IO_CMM_ALIGN_SIZE, (const AX_S8*)token_name.c_str());
-        }
-
-        return ret;
-    }
-
-    static inline AX_S32 free_engine_buffer(AX_ENGINE_IO_BUFFER_T* pBuf) {
-        if (pBuf->phyAddr == 0) {
-            delete[] reinterpret_cast<uint8_t*>(pBuf->pVirAddr);
-        }
-        else {
-            AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
-        }
-        pBuf->phyAddr = 0;
-        pBuf->pVirAddr = nullptr;
-
-        return 0;
-    }
-
-    static inline void free_io_index(AX_ENGINE_IO_BUFFER_T* io_buf, size_t index) {
-        AX_ENGINE_IO_BUFFER_T* pBuf = io_buf + index;
-        free_engine_buffer(pBuf);
-    }
-
-    static inline void free_io(AX_ENGINE_IO_T &io) {
-        for (size_t j = 0; j < io.nInputSize; ++j)
-        {
-            AX_ENGINE_IO_BUFFER_T *pBuf = io.pInputs + j;
-            AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
-        }
-        for (size_t j = 0; j < io.nOutputSize; ++j)
-        {
-            AX_ENGINE_IO_BUFFER_T *pBuf = io.pOutputs + j;
-            AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
-        }
-        delete[] io.pInputs;
-        delete[] io.pOutputs;
-    }
-
-    static inline void free_io(AX_ENGINE_IO_T &io, std::vector<std::vector<AX_ENGINE_IO_BUFFER_T>> &vecOutputBuffer) {
-        if (io.pInputs) {
-            delete[] io.pInputs;
-            io.pInputs = nullptr;
-        }
-
-        if (io.pOutputs) {
-            for (size_t index = 0; index < vecOutputBuffer.size(); ++index) {
-                AX_ENGINE_IO_BUFFER_T *pOutputs = &vecOutputBuffer[index][0];
-                for (size_t j = 0; j < io.nOutputSize; ++j) {
-                    free_io_index(pOutputs, j);
-                }
-            }
-
-            delete[] io.pOutputs;
-            io.pOutputs = nullptr;
-        }
-    }
-
-    static inline int prepare_io(const std::string& token, const AX_ENGINE_IO_INFO_T* info, AX_ENGINE_IO_T &io, IO_BUFFER_STRATEGY_T strategy) {
-        auto ret = 0;
-
-        memset(&io, 0, sizeof(io));
-
-        io.pInputs = new AX_ENGINE_IO_BUFFER_T[info->nInputSize];
-
-        if (!io.pInputs) {
-            goto EXIT;
-        }
-
-        memset(io.pInputs, 0x00, sizeof(AX_ENGINE_IO_BUFFER_T) * info->nInputSize);
-        io.nInputSize = info->nInputSize;
-        for (AX_U32 i = 0; i < info->nInputSize; ++i)
-        {
-            auto meta = info->pInputs[i];
-            auto buffer = &io.pInputs[i];
-            ret = alloc_engine_buffer(token, "_input_", i, &meta, buffer, strategy);
-            if (ret != 0)
-            {
-                free_io_index(io.pInputs, i);
-                return ret;
-            }
-        }
-
-        io.pOutputs = new AX_ENGINE_IO_BUFFER_T[info->nOutputSize];
-
-        if (!io.pOutputs) {
-            goto EXIT;
-        }
-
-        memset(io.pOutputs, 0x00, sizeof(AX_ENGINE_IO_BUFFER_T) * info->nOutputSize);
-        io.nOutputSize = info->nOutputSize;
-
-        for (size_t i = 0; i < info->nOutputSize; ++i) {
-            auto meta = info->pOutputs[i];
-            auto buffer = &io.pOutputs[i];
-            ret = alloc_engine_buffer(token, "_output_", i, &meta, buffer, strategy);
-            if (ret != 0) {
-                goto EXIT;
-            }
-        }
-
-        EXIT:
-            if (ret != 0) {
-                free_io(io);
-                return -1;
-            }
-
-        return 0;
-    }
-
-    static inline int prepare_io(const std::string& token,
-                                 const AX_ENGINE_IO_INFO_T* info, AX_ENGINE_IO_T &io,
-                                 std::vector<AX_ENGINE_IO_BUFFER_T> &vecOutputBuffer,
-                                 const IO_BUFFER_STRATEGY_T &strategy) {
-        AX_S32 ret = 0;
-        memset(&io, 0, sizeof(io));
-
-        std::vector<AX_ENGINE_IO_BUFFER_T> outputBuffer;
-
-        if (1 != info->nInputSize) {
-            fprintf(stderr, "[ERR]: Only single input was accepted(got %u).\n", info->nInputSize);
-            return -1;
-        }
-
-        io.pInputs = new AX_ENGINE_IO_BUFFER_T[info->nInputSize];
-
-        if (!io.pInputs) {
-            goto EXIT;
-        }
-
-        memset(io.pInputs, 0x00, sizeof(AX_ENGINE_IO_BUFFER_T) * info->nInputSize);
-        io.nInputSize = info->nInputSize;
-
-        for (AX_U32 i = 0; i < info->nInputSize; ++i)
-        {
-            auto meta = info->pInputs[i];
-            auto buffer = &io.pInputs[i];
-            ret = alloc_engine_buffer(token, "_input_", i, &meta, buffer, strategy);
-            if (ret != 0)
-            {
-                free_io_index(io.pInputs, i);
-                return ret;
-            }
-        }
-
-        io.pOutputs = new AX_ENGINE_IO_BUFFER_T[info->nOutputSize];
-
-        if (!io.pOutputs) {
-            goto EXIT;
-        }
-
-        for (size_t i = 0; i < info->nOutputSize; ++i) {
-            auto meta = info->pOutputs[i];
-            auto buffer = &io.pOutputs[i];
-            ret = alloc_engine_buffer(token, "_output_", i, &meta, buffer, strategy);
-
-            if (ret != 0) {
-                goto EXIT;
-            }
-
-            vecOutputBuffer.push_back(*buffer);
-        }
-
-        memset(io.pOutputs, 0x00, sizeof(AX_ENGINE_IO_BUFFER_T) * info->nOutputSize);
-        io.nOutputSize = info->nOutputSize;
-
-        for (size_t i = 0; i < info->nOutputSize; ++i) {
-            auto buffer = &io.pOutputs[i];
-            *buffer = vecOutputBuffer[i];
-        }
-
-        EXIT:
-        if (ret != 0) {
-            free_io(io);
-            return -1;
-        }
-
-        return 0;
-    }
-
-    static inline AX_S32 push_io_output(const AX_ENGINE_IO_INFO_T* info,
-                                        AX_ENGINE_IO_T& io,
-                                        std::vector<AX_ENGINE_IO_BUFFER_T> &outputBuffer) {
-        for (size_t i = 0; i < info->nOutputSize; ++i) {
-            auto buffer = &io.pOutputs[i];
-            *buffer = outputBuffer[i];
-        }
-
-        return 0;
-    }
-
-    static inline AX_S32 push_io_input(void* input, int index, AX_ENGINE_IO_T& io) {
-        // img ranks_depth ranks_feat ranks_bev, n_points
-        AX_ENGINE_IO_BUFFER_T* pImg = &io.pInputs[index];
-        memcpy(pImg->pVirAddr, input, pImg->nSize);
-        return 0;
-    }
-
-    static inline AX_S32 cache_io_flush(const AX_ENGINE_IO_BUFFER_T *io_buf) {
-        if (io_buf->phyAddr != 0) {
-            AX_SYS_MflushCache(io_buf->phyAddr, io_buf->pVirAddr, io_buf->nSize);
-        }
-
-        return 0;
-    }
-
-    static inline AX_S32 push_io_output(void* output,
-                                        int index,
-                                        AX_ENGINE_IO_T& io) {
-        AX_ENGINE_IO_BUFFER_T* pImg = &io.pOutputs[index];
-        cache_io_flush(pImg);
-        memcpy(output, pImg->pVirAddr, pImg->nSize);
-        return 0;
-    }
-
-    static inline AX_S32 cpu_copy(AX_U64 nPhyAddrSrc, AX_U64 nPhyAddrDst, AX_U32 nLen) {
-        if (nPhyAddrSrc != 0 && nPhyAddrDst != 0 && nLen > 0) {
-            AX_VOID* pSrcVirAddr = AX_SYS_MmapCache(nPhyAddrSrc, nLen);
-            AX_VOID* pDstVirAddr = AX_SYS_MmapCache(nPhyAddrDst, nLen);
-
-            memcpy((AX_VOID*)pDstVirAddr, (AX_VOID*)pSrcVirAddr, nLen);
-
-            AX_SYS_Munmap(pSrcVirAddr, nLen);
-            AX_SYS_Munmap(pDstVirAddr, nLen);
-
-            return 0;
-        }
-
-        return -1;
-    }
-
-    static inline AX_S32 inc_io_ref_cnt(const AX_VIDEO_FRAME_T &stFrame) {
-        if (stFrame.u32BlkId[0] > 0) {
-            AX_POOL_IncreaseRefCnt(stFrame.u32BlkId[0]);
-        }
-        if (stFrame.u32BlkId[1] > 0) {
-            AX_POOL_IncreaseRefCnt(stFrame.u32BlkId[1]);
-        }
-        if (stFrame.u32BlkId[2] > 0) {
-            AX_POOL_IncreaseRefCnt(stFrame.u32BlkId[2]);
-        }
-
-        return 0;
-    }
-
-    static inline AX_S32 dec_io_ref_cnt(const AX_VIDEO_FRAME_T &stFrame) {
-        if (stFrame.u32BlkId[0] > 0) {
-            AX_POOL_DecreaseRefCnt(stFrame.u32BlkId[0]);
-        }
-        if (stFrame.u32BlkId[1] > 0) {
-            AX_POOL_DecreaseRefCnt(stFrame.u32BlkId[1]);
-        }
-        if (stFrame.u32BlkId[2] > 0) {
-            AX_POOL_DecreaseRefCnt(stFrame.u32BlkId[2]);
-        }
-
-        return 0;
-    }
-
-    static inline bool read_file(const char* path, std::vector<char>& data) {
-        std::fstream fs(path, std::ios::in | std::ios::binary);
-
-        if (!fs.is_open()) {
-            return false;
-        }
-
-        fs.seekg(std::ios::end);
-        auto fs_end = fs.tellg();
-        fs.seekg(std::ios::beg);
-        auto fs_beg = fs.tellg();
-
-        auto file_size = static_cast<size_t>(fs_end - fs_beg);
-        auto vector_size = data.size();
-
-        data.reserve(vector_size + file_size);
-        data.insert(data.end(), std::istreambuf_iterator<char>(fs), std::istreambuf_iterator<char>());
-
-        fs.close();
-
-        return true;
-    }
-
-    static inline bool read_file(const char* path, AX_VOID **pModelBufferVirAddr,
-                                 AX_U64 &u64ModelBufferPhyAddr, AX_U32 &nModelBufferSize) {
-        std::fstream fs(path, std::ios::in | std::ios::binary);
-
-        if (!fs.is_open()) {
-            return false;
-        }
-
-        fs.seekg(0, std::ios::end);
-        int file_size = fs.tellg();
-        fs.seekg(0, std::ios::beg);
-
-        nModelBufferSize = (AX_U32)file_size;
-
-        AX_SYS_MemAlloc(&u64ModelBufferPhyAddr, pModelBufferVirAddr, nModelBufferSize, 0x100, (AX_S8 *)"SKEL-CV");
-
-        if (!pModelBufferVirAddr || (u64ModelBufferPhyAddr == 0)) {
-            return false;
-        }
-
-        fs.read((AX_CHAR *)*pModelBufferVirAddr, nModelBufferSize);
-
-        fs.close();
-
-        return true;
-    }
-
-    static inline void dequant(float** pptrOutput, const AX_ENGINE_IOMETA_T& ptrIoInfo, const AX_ENGINE_IO_BUFFER_T& ioBuf, float zp, float scale)
-    {
-        if (ptrIoInfo.eDataType == AX_ENGINE_DT_FLOAT32)
-        {
-            *pptrOutput = (float*)ioBuf.pVirAddr;
-            return;
-        }
-
-        *pptrOutput = (float*)malloc(ptrIoInfo.nSize * sizeof(float));
-        uint8_t *pBuf = (uint8_t*)ioBuf.pVirAddr;
-        float* pOutput = *pptrOutput;
-        // float inv_scale = 1.0f / scale;
-        for (int i = 0; i < ptrIoInfo.nSize; i++)
-        {
-            pOutput[i] = ((float)pBuf[i] - zp) * scale;
-        }
-    }
-}
-
diff --git a/projects/llm_framework/include/utils/logger.h b/projects/llm_framework/include/utils/logger.h
deleted file mode 100644
index a3f119dc..00000000
--- a/projects/llm_framework/include/utils/logger.h
+++ /dev/null
@@ -1,82 +0,0 @@
-/**************************************************************************************************
- *
- * Copyright (c) 2019-2023 Axera Semiconductor (Ningbo) Co., Ltd. All Rights Reserved.
- *
- * This source file is the property of Axera Semiconductor (Ningbo) Co., Ltd. and
- * may not be copied or distributed in any isomorphic form without the prior
- * written consent of Axera Semiconductor (Ningbo) Co., Ltd.
- *
- **************************************************************************************************/
-
-#ifndef SKEL_LOGGER_H
-#define SKEL_LOGGER_H
-
-#include "ax_global_type.h"
-#include "ax_sys_log.h"
-
-#include <stdio.h>
-
-#ifdef __cplusplus
-extern "C"
-{
-#endif
-
-//#define SKEL_LOG_TAG "SKEL"
-//
-//#define ALOGE(fmt, ...) AX_LOG_ERR_EX(SKEL_LOG_TAG, AX_ID_SKEL, fmt, ##__VA_ARGS__)
-//#define ALOGW(fmt, ...) AX_LOG_WARN_EX(SKEL_LOG_TAG, AX_ID_SKEL, fmt, ##__VA_ARGS__)
-//#define ALOGI(fmt, ...) AX_LOG_INFO_EX(SKEL_LOG_TAG, AX_ID_SKEL, fmt, ##__VA_ARGS__)
-//#define ALOGD(fmt, ...) AX_LOG_DBG_EX(SKEL_LOG_TAG, AX_ID_SKEL, fmt, ##__VA_ARGS__)
-//#define ALOGN(fmt, ...) AX_LOG_NOTICE_EX(SKEL_LOG_TAG, AX_ID_SKEL, fmt, ##__VA_ARGS__)
-
-
-typedef enum {
-    SKEL_LOG_MIN         = -1,
-    SKEL_LOG_EMERGENCY   = 0,
-    SKEL_LOG_ALERT       = 1,
-    SKEL_LOG_CRITICAL    = 2,
-    SKEL_LOG_ERROR       = 3,
-    SKEL_LOG_WARN        = 4,
-    SKEL_LOG_NOTICE      = 5,
-    SKEL_LOG_INFO        = 6,
-    SKEL_LOG_DEBUG       = 7,
-    SKEL_LOG_MAX
-} SKEL_LOG_LEVEL_E;
-
-static SKEL_LOG_LEVEL_E log_level = SKEL_LOG_DEBUG;
-
-#if 1
-#define MACRO_BLACK "\033[1;30;30m"
-#define MACRO_RED "\033[1;30;31m"
-#define MACRO_GREEN "\033[1;30;32m"
-#define MACRO_YELLOW "\033[1;30;33m"
-#define MACRO_BLUE "\033[1;30;34m"
-#define MACRO_PURPLE "\033[1;30;35m"
-#define MACRO_WHITE "\033[1;30;37m"
-#define MACRO_END "\033[0m"
-#else
-#define MACRO_BLACK
-#define MACRO_RED
-#define MACRO_GREEN
-#define MACRO_YELLOW
-#define MACRO_BLUE
-#define MACRO_PURPLE
-#define MACRO_WHITE
-#define MACRO_END
-#endif
-
-#define ALOGE(fmt, ...) printf(MACRO_RED "[E][%32s][%4d]: " fmt MACRO_END "\n", __FUNCTION__, __LINE__, ##__VA_ARGS__)
-#define ALOGW(fmt, ...) if (log_level >= SKEL_LOG_WARN) \
-    printf(MACRO_YELLOW "[W][%32s][%4d]: " fmt MACRO_END "\n", __FUNCTION__, __LINE__, ##__VA_ARGS__)
-#define ALOGI(fmt, ...) if (log_level >= SKEL_LOG_INFO) \
-    printf(MACRO_GREEN "[I][%32s][%4d]: " fmt MACRO_END "\n", __FUNCTION__, __LINE__, ##__VA_ARGS__)
-#define ALOGD(fmt, ...) if (log_level >= SKEL_LOG_DEBUG) \
-    printf(MACRO_WHITE "[D][%32s][%4d]: " fmt MACRO_END "\n", __FUNCTION__, __LINE__, ##__VA_ARGS__)
-#define ALOGN(fmt, ...) if (log_level >= SKEL_LOG_NOTICE) \
-    printf(MACRO_PURPLE "[N][%32s][%4d]: " fmt MACRO_END "\n", __FUNCTION__, __LINE__, ##__VA_ARGS__)
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif //SKEL_LOGGER_H
diff --git a/projects/llm_framework/main/SConstruct b/projects/llm_framework/main/SConstruct
index 0c82732f..ed4cbc5b 100644
--- a/projects/llm_framework/main/SConstruct
+++ b/projects/llm_framework/main/SConstruct
@@ -27,7 +27,8 @@ STATIC_FILES += [AFile('../static_lib/sherpa/ncnn/libsherpa-ncnn-core.so'),
                  AFile('../static_lib/wetext/libglog.so.0'),
                  AFile('../static_lib/wetext/libfst.so.16'),
                  AFile('../static_lib/libonnxruntime.so.1'),
-                AFile('../static_lib/libonnxruntime.so.1.14.0')
+                 AFile('../static_lib/libonnxruntime.so.1.14.0'),
+                 AFile('../static_lib/libzmq.so.5')
                  ]
 
 env['COMPONENTS'].append({'target':'static_file-1.0',
diff --git a/projects/llm_framework/main_asr/SConstruct b/projects/llm_framework/main_asr/SConstruct
index 344162bb..d9ce42c1 100644
--- a/projects/llm_framework/main_asr/SConstruct
+++ b/projects/llm_framework/main_asr/SConstruct
@@ -20,7 +20,7 @@ LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-
 DEFINITIONS += ['-std=c++17']
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 
-INCLUDE += [ADir('../include/sherpa'), ADir('../include/sherpa/sherpa-ncnn')]
+INCLUDE += [ADir('../static_lib/include/sherpa'), ADir('../static_lib/include/sherpa/sherpa-ncnn')]
 LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/ncnn')]
 REQUIREMENTS += ['ncnn', 'sherpa-ncnn-core']
 
diff --git a/projects/llm_framework/main_audio/SConstruct b/projects/llm_framework/main_audio/SConstruct
index 5491339b..2fa5dbbb 100644
--- a/projects/llm_framework/main_audio/SConstruct
+++ b/projects/llm_framework/main_audio/SConstruct
@@ -20,7 +20,6 @@ STATIC_FILES = []
 
 DEFINITIONS += ['-std=c++17']
 LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
-# INCLUDE += [ADir('../include'), ADir('../include/cppzmq'), ADir('../include/libzmq')]
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 REQUIREMENTS += ['ax_sys', 'ax_audio', 'ax_audio_3a', 'tinyalsa', 'fdk-aac', 'opus', 'samplerate']
 if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
diff --git a/projects/llm_framework/main_camera/SConstruct b/projects/llm_framework/main_camera/SConstruct
index a68311d9..14bc4405 100644
--- a/projects/llm_framework/main_camera/SConstruct
+++ b/projects/llm_framework/main_camera/SConstruct
@@ -20,17 +20,8 @@ STATIC_FILES = []
 DEFINITIONS += ['-O2']
 DEFINITIONS += ['-std=c++17']
 LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
-# INCLUDE += [ADir('../include'), ADir('../include/cppzmq'), ADir('../include/libzmq')]
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 
-# REQUIREMENTS += ['Backward_cpp']
-# DYNAMIC_LIB += [ AFile('../static_lib/libdw.so.1'),
-#                 AFile('../static_lib/libelf.so.1'),
-#                 AFile('../static_lib/libz.so.1'),
-#                 AFile('../static_lib/liblzma.so.5'),
-#                 AFile('../static_lib/libbz2.so.1.0')]
-# DEFINITIONS += ["-DENABLE_BACKWARD"]
-
 INCLUDE.append(os.path.join(env["MSP_PATH"], 'third-party/live/out/arm64/glibc/include/liveMedia'))
 INCLUDE.append(os.path.join(env["MSP_PATH"], 'third-party/live/out/arm64/glibc/include/groupsock'))
 INCLUDE.append(os.path.join(env["MSP_PATH"], 'third-party/live/out/arm64/glibc/include/UsageEnvironment'))
@@ -47,7 +38,7 @@ if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
     REQUIREMENTS += ['ax_audio_3a','samplerate', 'ax_fdk', 'fdk-aac', 'ax_awb', 'ax_af', 'ax_nt_stream', 'ax_nt_ctrl']
     REQUIREMENTS += ['liveMedia', 'UsageEnvironment', 'groupsock', 'BasicUsageEnvironment']
 
-INCLUDE += [ADir('../include/opencv4')]
+INCLUDE += [ADir('../static_lib/include/opencv4')]
 static_file = []
 static_file += [AFile('../static_lib/libopencv-4.6-aarch64-none/lib/libopencv_core.a')]
 static_file += [AFile('../static_lib/libopencv-4.6-aarch64-none/lib/libittnotify.a')]
diff --git a/projects/llm_framework/main_depth_anything/SConstruct b/projects/llm_framework/main_depth_anything/SConstruct
index 66caf78a..d4354487 100644
--- a/projects/llm_framework/main_depth_anything/SConstruct
+++ b/projects/llm_framework/main_depth_anything/SConstruct
@@ -21,7 +21,7 @@ LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 REQUIREMENTS += ['ax_engine', 'ax_interpreter', 'ax_sys']
 
-INCLUDE += [ADir('../include'), ADir('../include/opencv4')]
+INCLUDE += [ADir('../static_lib/include'), ADir('../static_lib/include/opencv4')]
 
 static_file = Glob('../static_lib/module-llm/libabsl_*')
 static_file = Glob('../static_lib/libopencv-4.6-aarch64-none/lib/lib*')
diff --git a/projects/llm_framework/main_kws/SConstruct b/projects/llm_framework/main_kws/SConstruct
index b417fc2b..fc175abb 100644
--- a/projects/llm_framework/main_kws/SConstruct
+++ b/projects/llm_framework/main_kws/SConstruct
@@ -22,10 +22,10 @@ DEFINITIONS += ['-std=c++17']
 LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 
-INCLUDE += [ADir('../include/sherpa'),
-            ADir('../include/sherpa/sherpa-onnx'),
-            ADir('../include/sherpa/sherpa-onnx/onnxruntime-src'),
-            ADir('../include/sherpa/sherpa-onnx/openfst-src')
+INCLUDE += [ADir('../static_lib/include/sherpa'),
+            ADir('../static_lib/include/sherpa/sherpa-onnx'),
+            ADir('../static_lib/include/sherpa/sherpa-onnx/onnxruntime-src'),
+            ADir('../static_lib/include/sherpa/sherpa-onnx/openfst-src')
             ]
 
 LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/onnx')]
diff --git a/projects/llm_framework/main_llm/SConstruct b/projects/llm_framework/main_llm/SConstruct
index 1b12643b..da100ad7 100644
--- a/projects/llm_framework/main_llm/SConstruct
+++ b/projects/llm_framework/main_llm/SConstruct
@@ -34,10 +34,10 @@ LINK_SEARCH_PATH += [ADir('../static_lib')]
 
 
 INCLUDE += [ADir('src/runner'), ADir('src/runner/utils')]
-INCLUDE += [ADir('../include/sentencepiece'),
-            ADir('../include/protobuf-lite'),
-            ADir('../include/abseil-cpp'),
-            ADir('../include/re2')]
+INCLUDE += [ADir('../static_lib/include/sentencepiece'),
+            ADir('../static_lib/include/protobuf-lite'),
+            ADir('../static_lib/include/abseil-cpp'),
+            ADir('../static_lib/include/re2')]
 
 static_file = Glob('../static_lib/module-llm/libabsl_*')
 static_file += [AFile('../static_lib/module-llm/libre2.a'), AFile('../static_lib/module-llm/libsentencepiece.a'), AFile('../static_lib/module-llm/libsentencepiece_train.a')]
diff --git a/projects/llm_framework/main_llm/src/main.cpp b/projects/llm_framework/main_llm/src/main.cpp
index a1cb066c..71ad2f7f 100644
--- a/projects/llm_framework/main_llm/src/main.cpp
+++ b/projects/llm_framework/main_llm/src/main.cpp
@@ -64,6 +64,7 @@ class llm_task {
     std::vector<unsigned short> prompt_data;
     std::vector<int> tokens_ids, tokens_diff;
     std::vector<std::vector<unsigned short>> k_caches, v_caches;
+    std::string kvcache_path;
     int precompute_len = 0;
     std::vector<int> _token_ids;
     task_callback_t out_callback_;
@@ -103,6 +104,23 @@ class llm_task {
         return false;
     }
 
+    void prepare_kvcache_folder(const std::string &kvcache_path)
+    {
+        try {
+            if (!std::filesystem::exists(kvcache_path)) {
+                std::filesystem::create_directories(kvcache_path);
+            }
+
+            if (std::filesystem::exists(kvcache_path) && std::filesystem::is_directory(kvcache_path)) {
+                for (const auto &entry : std::filesystem::directory_iterator(kvcache_path)) {
+                    std::filesystem::remove_all(entry.path());
+                }
+            }
+        } catch (const std::exception &e) {
+            ALOGI("prepare_kvcache_folder: skip clear/create due to error: %s", e.what());
+        }
+    }
+
     int load_model(const nlohmann::json &config_body)
     {
         if (parse_config(config_body)) {
@@ -237,7 +255,9 @@ class llm_task {
 
             if (lLaMa_ctx_) {
                 lLaMa_ctx_->SetSystemPrompt(mode_config_.system_prompt, _token_ids);
-                std::string kvcache_path = "/tmp/.llm/";
+                if (!kvcache_path.empty()) {
+                    prepare_kvcache_folder(kvcache_path);
+                }
                 if (!kvcache_path.empty() && kvcache_path != "") {
                     if (lLaMa_ctx_->load_kvcache(kvcache_path, mode_config_.axmodel_num, k_caches, v_caches,
                                                  mode_config_.system_prompt, precompute_len)) {
diff --git a/projects/llm_framework/main_melotts/SConstruct b/projects/llm_framework/main_melotts/SConstruct
index 1701cb76..0ea5d61f 100644
--- a/projects/llm_framework/main_melotts/SConstruct
+++ b/projects/llm_framework/main_melotts/SConstruct
@@ -23,8 +23,8 @@ LINK_SEARCH_PATH += [ADir('../static_lib')]
 REQUIREMENTS += ['ax_engine', 'ax_interpreter', 'ax_sys']
 REQUIREMENTS += ['samplerate']
 
-INCLUDE += [ADir('../include')]
-INCLUDE += [ADir('src/runner'), ADir('../include/onnxruntime/core/session')]
+INCLUDE += [ADir('../static_lib/include')]
+INCLUDE += [ADir('src/runner'), ADir('../static_lib/include/onnxruntime/core/session')]
 LINK_SEARCH_PATH += [ADir('../static_lib/wetext')]
 REQUIREMENTS += ['glog', 'fst']
 
diff --git a/projects/llm_framework/main_skel/SConstruct b/projects/llm_framework/main_skel/SConstruct
index 55e508b6..a768f334 100644
--- a/projects/llm_framework/main_skel/SConstruct
+++ b/projects/llm_framework/main_skel/SConstruct
@@ -26,7 +26,7 @@ REQUIREMENTS += ['ax_engine', 'ax_interpreter', 'ax_sys']
 if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
     REQUIREMENTS += ['ax_skel', 'ax_engine', 'ax_venc', 'ax_vdec', 'ax_ivps', 'ax_proton']
 
-INCLUDE += [ADir('../include'), ADir('src/runner'), ADir('../include/opencv4')]
+INCLUDE += [ADir('../static_lib/include'), ADir('src/runner'), ADir('../static_lib/include/opencv4')]
 
 static_file = Glob('../static_lib/module-llm/libabsl_*')
 static_file = Glob('../static_lib/libopencv-4.6-aarch64-none/lib/lib*')
diff --git a/projects/llm_framework/main_vad/SConstruct b/projects/llm_framework/main_vad/SConstruct
index bbc289d5..3c1b4b05 100644
--- a/projects/llm_framework/main_vad/SConstruct
+++ b/projects/llm_framework/main_vad/SConstruct
@@ -20,7 +20,7 @@ DEFINITIONS += ['-std=c++17']
 LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 
-INCLUDE += [ADir('../include/sherpa')]
+INCLUDE += [ADir('../static_lib/include/sherpa')]
 
 LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/onnx')]
 LDFLAGS += ['-l:libsherpa-onnx-core.a']
diff --git a/projects/llm_framework/main_vlm/SConstruct b/projects/llm_framework/main_vlm/SConstruct
index bba83f8f..b5f429f3 100644
--- a/projects/llm_framework/main_vlm/SConstruct
+++ b/projects/llm_framework/main_vlm/SConstruct
@@ -27,11 +27,11 @@ LINK_SEARCH_PATH += [ADir('../static_lib')]
 
 
 INCLUDE += [ADir('src/runner'), ADir('src/runner/utils')]
-INCLUDE += [ADir('../include/sentencepiece'),
-            ADir('../include/protobuf-lite'),
-            ADir('../include/abseil-cpp'),
-            ADir('../include/opencv4'),
-            ADir('../include/re2')]
+INCLUDE += [ADir('../static_lib/include/sentencepiece'),
+            ADir('../static_lib/include/protobuf-lite'),
+            ADir('../static_lib/include/abseil-cpp'),
+            ADir('../static_lib/include/opencv4'),
+            ADir('../static_lib/include/re2')]
 
 static_file = Glob('../static_lib/module-llm/libabsl_*')
 static_file += [AFile('../static_lib/module-llm/libre2.a'), AFile('../static_lib/module-llm/libsentencepiece.a'), AFile('../static_lib/module-llm/libsentencepiece_train.a')]
diff --git a/projects/llm_framework/main_vlm/models/mode_internvl3-1B-448-ax650.json b/projects/llm_framework/main_vlm/models/mode_internvl3-1B-448-ax650.json
new file mode 100644
index 00000000..4da7a27e
--- /dev/null
+++ b/projects/llm_framework/main_vlm/models/mode_internvl3-1B-448-ax650.json
@@ -0,0 +1,39 @@
+{
+    "mode": "internvl3-1B-448-ax650",
+    "type": "vlm",
+    "homepage": "https://huggingface.co/AXERA-TECH/InternVL3-1B",
+    "capabilities": [
+        "text_generation",
+        "chat"
+    ],
+    "input_type": [
+        "vlm.chat_completion",
+        "vlm.chat_completion.stream"
+    ],
+    "output_type": [
+        "vlm.utf-8",
+        "vlm.utf-8.stream"
+    ],
+    "mode_param": {
+        "tokenizer_type": 2,
+        "url_tokenizer_model": "http://localhost:8080",
+        "filename_tokens_embed": "model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel": "qwen2_post.axmodel",
+        "template_filename_axmodel": "qwen2_p128_l%d_together.axmodel",
+        "filename_image_encoder_axmodel": "internvl3_1b_vit.axmodel",
+        "b_use_topk": false,
+        "b_bos": false,
+        "b_eos": false,
+        "axmodel_num": 24,
+        "tokens_embed_num": 151674,
+        "img_token_id": 151667,
+        "tokens_embed_size": 896,
+        "b_use_mmap_load_embed": true,
+        "b_dynamic_load_axmodel_layer": false,
+        "precompute_len": 2048,
+        "cmm_size": 1919044,
+        "ext_scripts": [
+            "tokenizer_internvl3-1B-448-ax650.py"
+        ]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/scripts/tokenizer_internvl3-1B-448-ax650.py b/projects/llm_framework/main_vlm/scripts/tokenizer_internvl3-1B-448-ax650.py
new file mode 100644
index 00000000..cf6a3dc4
--- /dev/null
+++ b/projects/llm_framework/main_vlm/scripts/tokenizer_internvl3-1B-448-ax650.py
@@ -0,0 +1,241 @@
+from transformers import AutoTokenizer
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+import uuid
+
+tokenizers = {}
+
+class Tokenizer_Http():
+    def __init__(self, model_id):
+        self.tokenizer = AutoTokenizer.from_pretrained(model_id)
+        self.messages = [
+            {"role": "system", "content": "你是由上海人工智能实验室联合商汤科技开发的书生多模态大模型，英文名叫InternVL, 是一个有用无害的人工智能助手。"},
+        ]
+        self.token_ids = []
+        self.token_ids_cache = []
+
+    def encode(self, prompt, last_reply=None):
+        if last_reply is not None:
+            self.messages.append({"role": "assistant", "content": last_reply})
+            text = self.tokenizer.apply_chat_template(
+                self.messages,
+                tokenize=False,
+                add_generation_prompt=True
+            )
+            self.token_ids = self.tokenizer.encode(text)[:-3]
+        self.messages.append({"role": "user", "content": prompt})
+
+        text = self.tokenizer.apply_chat_template(
+            self.messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        token_ids = self.tokenizer.encode(text)
+        diff = token_ids[len(self.token_ids):]
+        self.token_ids = token_ids
+        return token_ids, diff
+
+    def encode_with_image(self, question: str, num_of_images: int, imgsz: int, last_reply=None):
+        if last_reply is not None:
+            self.messages.append({"role": "assistant", "content": last_reply})
+
+        # 根据图片尺寸设定 context_len
+        if imgsz == 448:
+            context_len = 256
+        elif imgsz == 224:
+            context_len = 64
+        else:
+            print(f"Unsupported imgsz: {imgsz}")
+            return None, None
+
+        # 拼接带图片的用户输入
+        question_with_images = question
+        if num_of_images > 0:
+            for _ in range(num_of_images):
+                question_with_images += "\n<img>" + "<IMG_CONTEXT>" * context_len + "</img>\n"
+
+        self.messages.append({"role": "user", "content": question_with_images})
+
+        text = self.tokenizer.apply_chat_template(
+            self.messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        token_ids = self.tokenizer.encode(text)
+        diff = token_ids[len(self.token_ids):]
+        self.token_ids = token_ids
+        return token_ids, diff
+
+    def decode(self, token_ids):
+        self.token_ids_cache += token_ids
+        text = self.tokenizer.decode(self.token_ids_cache)
+        if "\ufffd" in text:
+            print("Text 中包含非法字符")
+            return ""
+        else:
+            self.token_ids_cache.clear()
+            return text
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+    @property
+    def bos_token(self):
+        return self.tokenizer.bos_token
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+    @property
+    def img_start_token(self):
+        return self.tokenizer.encode("<img>")[0]
+    @property
+    def img_context_token(self):
+        return self.tokenizer.encode("<IMG_CONTEXT>")[0]
+
+    def reset(self, system_prompt=None):
+        if system_prompt is None:
+            system_prompt = args.content
+        self.messages = [
+            {"role": "system", "content": system_prompt},
+        ]
+        text = self.tokenizer.apply_chat_template(
+            self.messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        token_ids = self.tokenizer.encode(text)[:-3]
+        self.token_ids = token_ids
+        print(self.decode(token_ids))
+        return token_ids
+
+
+class Request(BaseHTTPRequestHandler):
+    timeout = 5
+    server_version = 'Apache'
+
+    def do_GET(self):
+        print(self.path)
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.end_headers()
+        if '/get_uid' in self.path:
+            new_uid = str(uuid.uuid4())
+            print("新 uid:", new_uid)
+            tokenizers[new_uid] = Tokenizer_Http(args.model_id)
+            msg = json.dumps({'uid': new_uid})
+        elif '/bos_id' in self.path:
+            uid = self.get_query_param("uid")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                msg = json.dumps({'bos_id': instance.bos_id if instance.bos_id is not None else -1})
+        elif '/eos_id' in self.path:
+            uid = self.get_query_param("uid")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                msg = json.dumps({'eos_id': instance.eos_id if instance.eos_id is not None else -1})
+        elif '/img_start_token' in self.path:
+            uid = self.get_query_param("uid")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                msg = json.dumps({'img_start_token': instance.img_start_token})
+
+        elif '/img_context_token' in self.path:
+            uid = self.get_query_param("uid")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                msg = json.dumps({'img_context_token': instance.img_context_token})
+        else:
+            msg = json.dumps({'error': 'Invalid GET endpoint'})
+        print(msg)
+        self.wfile.write(msg.encode())
+
+    def do_POST(self):
+        content_length = int(self.headers.get('content-length', 0))
+        data = self.rfile.read(content_length).decode()
+        req = json.loads(data)
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.end_headers()
+
+        if '/encode' in self.path:
+            uid = req.get('uid')
+            prompt = req.get('text')
+            last_reply = req.get('last_reply')
+            b_img_prompt = False
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if 'img_prompt' in req:
+                b_img_prompt = req['img_prompt']
+            if b_img_prompt:
+                num_img = req['num_img']
+                imgsz = req['imgsz']
+
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                if b_img_prompt:
+                    token_ids, diff = instance.encode_with_image(prompt, num_img, imgsz, last_reply)
+                else:
+                    token_ids, diff = instance.encode(prompt, last_reply)
+                msg = json.dumps({'token_ids': token_ids, 'diff': diff})
+
+        elif '/decode' in self.path:
+            uid = req.get('uid')
+            token_ids = req.get('token_ids')
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                text = instance.decode(token_ids)
+                msg = json.dumps({'text': text})
+
+        elif '/reset' in self.path:
+            uid = req.get("uid")
+            system_prompt = req.get("system_prompt")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                if system_prompt is not None:
+                    print("system_prompt:", system_prompt)
+                    token_ids = instance.reset(system_prompt)
+                else:
+                    token_ids = instance.reset()
+                msg = json.dumps({'token_ids': token_ids})
+
+        else:
+            msg = json.dumps({'error': 'Invalid POST endpoint'})
+
+        self.wfile.write(msg.encode())
+
+    def get_query_param(self, key):
+        from urllib.parse import urlparse, parse_qs
+        query = urlparse(self.path).query
+        params = parse_qs(query)
+        values = params.get(key)
+        return values[0] if values else None
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", type=str, default="localhost")
+    parser.add_argument("--port", type=int, default=8080)
+    parser.add_argument('--model_id', type=str, default='internvl3_tokenizer')
+    parser.add_argument('--content', type=str, default='你是由上海人工智能实验室联合商汤科技开发的书生多模态大模型，英文名叫InternVL, 是一个有用无害的人工智能助手。')
+    args = parser.parse_args()
+
+    host = (args.host, args.port)
+    print("http://%s:%s" % host)
+    server = HTTPServer(host, Request)
+    server.serve_forever()
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/main.cpp b/projects/llm_framework/main_vlm/src/main.cpp
index 6460cff1..36f278a7 100644
--- a/projects/llm_framework/main_vlm/src/main.cpp
+++ b/projects/llm_framework/main_vlm/src/main.cpp
@@ -65,6 +65,7 @@ class llm_task {
     std::string last_reply;
     std::vector<int> tokens_ids, tokens_diff;
     std::vector<std::vector<unsigned short>> k_caches, v_caches;
+    std::string kvcache_path;
     int precompute_len = 0;
     std::vector<int> _token_ids;
     task_callback_t out_callback_;
@@ -103,6 +104,23 @@ class llm_task {
         return false;
     }
 
+    void prepare_kvcache_folder(const std::string &kvcache_path)
+    {
+        try {
+            if (!std::filesystem::exists(kvcache_path)) {
+                std::filesystem::create_directories(kvcache_path);
+            }
+
+            if (std::filesystem::exists(kvcache_path) && std::filesystem::is_directory(kvcache_path)) {
+                for (const auto &entry : std::filesystem::directory_iterator(kvcache_path)) {
+                    std::filesystem::remove_all(entry.path());
+                }
+            }
+        } catch (const std::exception &e) {
+            ALOGI("prepare_kvcache_folder: skip clear/create due to error: %s", e.what());
+        }
+    }
+
     int load_model(const nlohmann::json &config_body)
     {
         if (parse_config(config_body)) {
@@ -247,7 +265,9 @@ class llm_task {
 
             if (lLaMa_ctx_) {
                 lLaMa_ctx_->SetSystemPrompt(mode_config_.system_prompt, _token_ids);
-                std::string kvcache_path = "/tmp/.vlm/";
+                if (!kvcache_path.empty()) {
+                    prepare_kvcache_folder(kvcache_path);
+                }
                 if (!kvcache_path.empty() && kvcache_path != "") {
                     if (lLaMa_ctx_->load_kvcache(kvcache_path, mode_config_.axmodel_num, k_caches, v_caches,
                                                  mode_config_.system_prompt, precompute_len)) {
@@ -353,7 +373,7 @@ class llm_task {
                     return;
                 }
 
-                if (image_data_.empty()) {
+                if (images_data.empty()) {
                     lLaMa_ctx_->Encode(prompt_data_, prompt_complete(msg), last_reply, tokens_ids, tokens_diff);
                     if (auto ret = lLaMa_ctx_->SetKVCache(k_caches, v_caches, precompute_len, tokens_diff.size());
                         ret != 0) {
@@ -376,17 +396,15 @@ class llm_task {
                     }
                     if (mats.empty()) return;
                     images_data.clear();
-                    cv::Mat src = cv::imdecode(image_data_, cv::IMREAD_COLOR);
-                    if (src.empty()) return;
-                    image_data_.clear();
-                    std::vector<unsigned short> img_embed;
-                    if (auto ret = lLaMa_ctx_->Encode(src, img_embed); ret != 0) {
+                    lLaMa_ctx_->ClearImgsEmbed();
+                    std::vector<std::vector<unsigned short>> all_embeds;
+                    if (auto ret = lLaMa_ctx_->Encode(mats, all_embeds); ret != 0) {
                         ALOGE("lLaMaCtx.Encode failed");
                         return;
                     }
                     mats.clear();
                     if (auto ret =
-                            lLaMa_ctx_->Encode(img_embed, prompt_data_, prompt_complete(msg), tokens_ids, tokens_diff);
+                            lLaMa_ctx_->Encode(all_embeds, prompt_data_, prompt_complete(msg), tokens_ids, tokens_diff);
                         ret != 0) {
                         ALOGE("lLaMaCtx.Encode failed");
                         return;
diff --git a/projects/llm_framework/main_whisper/SConstruct b/projects/llm_framework/main_whisper/SConstruct
index 6395e923..bc4d84b5 100644
--- a/projects/llm_framework/main_whisper/SConstruct
+++ b/projects/llm_framework/main_whisper/SConstruct
@@ -23,8 +23,8 @@ LINK_SEARCH_PATH += [ADir('../static_lib')]
 REQUIREMENTS += ['ax_engine', 'ax_interpreter', 'ax_sys']
 # REQUIREMENTS += ['onnxruntime', 'samplerate']
 
-INCLUDE += [ADir('../include')]
-INCLUDE += [ADir('src/runner'), ADir('../include/onnxruntime/core/session')]
+INCLUDE += [ADir('../static_lib/include')]
+INCLUDE += [ADir('src/runner'), ADir('../static_lib/include/onnxruntime/core/session')]
 INCLUDE += [ADir('src/runner/opencc/include/opencc')]
 
 LINK_SEARCH_PATH += [ADir('../static_lib/opencc/lib')]
diff --git a/projects/llm_framework/main_yolo/SConstruct b/projects/llm_framework/main_yolo/SConstruct
index 54502a54..55137bbb 100644
--- a/projects/llm_framework/main_yolo/SConstruct
+++ b/projects/llm_framework/main_yolo/SConstruct
@@ -21,7 +21,7 @@ LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 REQUIREMENTS += ['ax_engine', 'ax_interpreter', 'ax_sys']
 
-INCLUDE += [ADir('../include'), ADir('../include/opencv4')]
+INCLUDE += [ADir('../static_lib/include'), ADir('../static_lib/include/opencv4')]
 
 static_file = Glob('../static_lib/module-llm/libabsl_*')
 static_file = Glob('../static_lib/libopencv-4.6-aarch64-none/lib/lib*')

From 1df8ab9d7b4d035be9671dd9a8a4b650feb7a2c9 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 4 Sep 2025 14:52:54 +0800
Subject: [PATCH 37/79] [update] update model list

---
 .../mode_qwen2.5-HA-0.5B-ctx-ax630c.json      |  56 ++---
 .../mode_qwen2.5-HA-0.5B-ctx-ax650.json       |  46 ++++
 .../tokenizer_qwen2.5-HA-0.5B-ctx-ax650.py    | 203 ++++++++++++++++++
 ...json => mode_internvl3-1B-448-ax630c.json} |  19 +-
 .../models/mode_internvl3-1B-448-ax650.json   |  13 +-
 ...y => tokenizer_internvl3-1B-448-ax630c.py} |   0
 projects/llm_framework/main_vlm/src/main.cpp  |   4 +-
 .../llm_framework/main_vlm/src/runner/LLM.hpp |  10 +-
 projects/llm_framework/tools/llm_pack.py      |   3 +
 9 files changed, 313 insertions(+), 41 deletions(-)
 create mode 100644 projects/llm_framework/main_llm/models/mode_qwen2.5-HA-0.5B-ctx-ax650.json
 create mode 100644 projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-HA-0.5B-ctx-ax650.py
 rename projects/llm_framework/main_vlm/models/{mode_internvl3-1B-ax630c.json => mode_internvl3-1B-448-ax630c.json} (64%)
 rename projects/llm_framework/main_vlm/scripts/{tokenizer_internvl3-1B-ax630c.py => tokenizer_internvl3-1B-448-ax630c.py} (100%)

diff --git a/projects/llm_framework/main_llm/models/mode_qwen2.5-HA-0.5B-ctx-ax630c.json b/projects/llm_framework/main_llm/models/mode_qwen2.5-HA-0.5B-ctx-ax630c.json
index 1481d454..83d7cb72 100644
--- a/projects/llm_framework/main_llm/models/mode_qwen2.5-HA-0.5B-ctx-ax630c.json
+++ b/projects/llm_framework/main_llm/models/mode_qwen2.5-HA-0.5B-ctx-ax630c.json
@@ -1,38 +1,46 @@
 {
-    "mode":"qwen2.5-HA-0.5B-ctx-ax630c",
-    "type":"llm",
-    "homepage":"https://huggingface.co/yunyu1258/qwen2.5-0.5b-ha",
-    "compile_flage":"pulsar2 llm_build --input_path Qwen/qwen2.5-0.5b-ha --output_path Qwen/qwen2.5-0.5B-p1024-ha-ax630c --hidden_state_type bf16 --prefill_len 128 --kv_cache_len 1280 --last_kv_cache_len 128 --last_kv_cache_len 512 --last_kv_cache_len 1024 --chip AX620E --parallel 24",
-    "pulsar_version":"4.1-patch1-c37957c7",
-    "capabilities":[
+    "mode": "qwen2.5-HA-0.5B-ctx-ax630c",
+    "type": "llm",
+    "homepage": "https://huggingface.co/yunyu1258/qwen2.5-0.5b-ha",
+    "compile_flage": "pulsar2 llm_build --input_path Qwen/qwen2.5-0.5b-ha --output_path Qwen/qwen2.5-0.5B-p1024-ha-ax630c --hidden_state_type bf16 --prefill_len 128 --kv_cache_len 1280 --last_kv_cache_len 128 --last_kv_cache_len 512 --last_kv_cache_len 1024 --chip AX620E --parallel 24",
+    "pulsar_version": "4.1-patch1-c37957c7",
+    "capabilities": [
         "text_generation",
         "chat"
     ],
-    "input_type":[
+    "input_type": [
         "llm.utf-8",
         "llm.utf-8.stream",
         "llm.chat_completion",
         "llm.chat_completion.stream"
     ],
-    "output_type":[
+    "output_type": [
         "llm.utf-8",
         "llm.utf-8.stream"
     ],
-    "mode_param":{
-        "tokenizer_type":2,
-        "url_tokenizer_model":"http://localhost:8080",
-        "filename_tokens_embed":"model.embed_tokens.weight.bfloat16.bin",
-        "filename_post_axmodel":"qwen2_post.axmodel",
-        "template_filename_axmodel":"qwen2_p128_l%d_together.axmodel",
-        "b_use_topk":false,
-        "b_bos":false,
-        "b_eos":false,
-        "axmodel_num":24,
-        "tokens_embed_num":151936,
-        "tokens_embed_size":896,
-        "b_use_mmap_load_embed":true,
-        "b_dynamic_load_axmodel_layer":false,
-        "precompute_len":1202,
-        "ext_scripts":["tokenizer_qwen2.5-HA-0.5B-ctx-ax630c.py"]
+    "mode_param": {
+        "tokenizer_type": 2,
+        "url_tokenizer_model": "http://localhost:8080",
+        "filename_tokens_embed": "model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel": "qwen2_post.axmodel",
+        "template_filename_axmodel": "qwen2_p128_l%d_together.axmodel",
+        "enable_temperature": true,
+        "temperature": 0.7,
+        "enable_top_p_sampling": false,
+        "top_p": 0.9,
+        "enable_top_k_sampling": true,
+        "top_k": 40,
+        "enable_repetition_penalty": false,
+        "repetition_penalty": 1.1,
+        "penalty_window": 50,
+        "axmodel_num": 24,
+        "tokens_embed_num": 151936,
+        "tokens_embed_size": 896,
+        "b_use_mmap_load_embed": true,
+        "precompute_len": 1024,
+        "cmm_size": 1840108,
+        "ext_scripts": [
+            "tokenizer_qwen2.5-HA-0.5B-ctx-ax630c.py"
+        ]
     }
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_llm/models/mode_qwen2.5-HA-0.5B-ctx-ax650.json b/projects/llm_framework/main_llm/models/mode_qwen2.5-HA-0.5B-ctx-ax650.json
new file mode 100644
index 00000000..c54669f7
--- /dev/null
+++ b/projects/llm_framework/main_llm/models/mode_qwen2.5-HA-0.5B-ctx-ax650.json
@@ -0,0 +1,46 @@
+{
+    "mode": "qwen2.5-HA-0.5B-ctx-ax650",
+    "type": "llm",
+    "homepage": "https://huggingface.co/yunyu1258/qwen2.5-0.5b-ha",
+    "compile_flage": "pulsar2 llm_build --input_path Qwen/qwen2.5-0.5b-ha --output_path Qwen/qwen2.5-0.5B-p1024-ha-ax650 --hidden_state_type bf16 --prefill_len 128 --kv_cache_len 1280 --last_kv_cache_len 128 --last_kv_cache_len 512 --last_kv_cache_len 1024 --chip AX650 --parallel 24",
+    "pulsar_version": "4.1-patch1-c37957c7",
+    "capabilities": [
+        "text_generation",
+        "chat"
+    ],
+    "input_type": [
+        "llm.utf-8",
+        "llm.utf-8.stream",
+        "llm.chat_completion",
+        "llm.chat_completion.stream"
+    ],
+    "output_type": [
+        "llm.utf-8",
+        "llm.utf-8.stream"
+    ],
+    "mode_param": {
+        "tokenizer_type": 2,
+        "url_tokenizer_model": "http://localhost:8080",
+        "filename_tokens_embed": "model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel": "qwen2_post.axmodel",
+        "template_filename_axmodel": "qwen2_p128_l%d_together.axmodel",
+        "enable_temperature": true,
+        "temperature": 0.7,
+        "enable_top_p_sampling": false,
+        "top_p": 0.9,
+        "enable_top_k_sampling": true,
+        "top_k": 40,
+        "enable_repetition_penalty": false,
+        "repetition_penalty": 1.1,
+        "penalty_window": 50,
+        "axmodel_num": 24,
+        "tokens_embed_num": 151936,
+        "tokens_embed_size": 896,
+        "b_use_mmap_load_embed": true,
+        "precompute_len": 1024,
+        "cmm_size": 730656,
+        "ext_scripts": [
+            "tokenizer_qwen2.5-HA-0.5B-ctx-ax650.py"
+        ]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-HA-0.5B-ctx-ax650.py b/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-HA-0.5B-ctx-ax650.py
new file mode 100644
index 00000000..4a63e88a
--- /dev/null
+++ b/projects/llm_framework/main_llm/scripts/tokenizer_qwen2.5-HA-0.5B-ctx-ax650.py
@@ -0,0 +1,203 @@
+from transformers import AutoTokenizer, PreTrainedTokenizerFast
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+import uuid
+
+# 全局字典：存储 uid 到 Tokenizer_Http 实例的映射
+tokenizers = {}
+
+class Tokenizer_Http():
+    def __init__(self, model_id):
+        self.tokenizer = AutoTokenizer.from_pretrained(model_id)
+        self.messages = [
+            {"role": "system", "content": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant."},
+        ]
+        self.token_ids = []
+        
+        self.token_ids_cache = []
+
+    def encode(self, prompt, last_reply=None):
+        if last_reply is not None:
+            self.messages.append({"role": "assistant", "content": last_reply})
+            text = self.tokenizer.apply_chat_template(
+                self.messages,
+                tokenize=False,
+                add_generation_prompt=True
+            )
+            # print("生成的文本:\n============\n", text, "============\n")
+            self.token_ids = self.tokenizer.encode(text)[:-3]
+        self.messages.append({"role": "user", "content": prompt})
+        
+        text = self.tokenizer.apply_chat_template(
+            self.messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        print("生成的文本:\n============\n", text, "============\n")
+        token_ids = self.tokenizer.encode(text)
+        # 找出新增部分
+        diff = token_ids[len(self.token_ids):]
+        self.token_ids = token_ids
+        print(self.decode(diff))
+        return token_ids, diff
+
+    def decode(self, token_ids):
+        self.token_ids_cache += token_ids
+        text = self.tokenizer.decode(self.token_ids_cache)
+        if "\ufffd" in text:
+            print("text 中包含非法字符")
+            return ""
+        else:
+            self.token_ids_cache.clear()
+            return text
+        
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+    
+    @property
+    def bos_token(self):
+        return self.tokenizer.bos_token
+
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+    
+    def reset(self, system_prompt=None):
+        if system_prompt is None:
+            system_prompt = args.content
+        self.messages = [
+            {"role": "system", "content": system_prompt},
+        ]
+        text = self.tokenizer.apply_chat_template(
+            self.messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        token_ids = self.tokenizer.encode(text)[:-3]
+        self.token_ids = token_ids
+        print(self.decode(token_ids))
+        return token_ids
+        
+
+class Request(BaseHTTPRequestHandler):
+    timeout = 5
+    server_version = 'Apache'
+
+    def do_GET(self):
+        print("GET 请求路径:", self.path)
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.end_headers()
+
+        # 新增接口：获取 uid
+        if '/get_uid' in self.path:
+            new_uid = str(uuid.uuid4())
+            print("新 uid:", new_uid)
+            # 为该 uid 创建一个新的 Tokenizer_Http 实例
+            tokenizers[new_uid] = Tokenizer_Http(args.model_id)
+            msg = json.dumps({'uid': new_uid})
+        elif '/bos_id' in self.path:
+            # 获取 uid 参数（例如 ?uid=xxx）
+            uid = self.get_query_param("uid")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                bos_id = instance.bos_id
+                msg = json.dumps({'bos_id': bos_id if bos_id is not None else -1})
+        elif '/eos_id' in self.path:
+            uid = self.get_query_param("uid")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                eos_id = instance.eos_id
+                msg = json.dumps({'eos_id': eos_id if eos_id is not None else -1})
+        else:
+            msg = json.dumps({'error': 'Invalid GET endpoint'})
+
+        print("响应消息:", msg)
+        self.wfile.write(msg.encode())
+
+    def do_POST(self):
+        content_length = int(self.headers.get('content-length', 0))
+        data = self.rfile.read(content_length).decode()
+        print("POST 请求路径:", self.path)
+        print("接收到的数据:", data)
+        req = json.loads(data)
+
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.end_headers()
+
+        if '/encode' in self.path:
+            # 请求数据中必须包含 uid, text, 和可选的 last_reply
+            uid = req.get('uid')
+            prompt = req.get('text')
+            last_reply = req.get('last_reply')
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                token_ids, diff = instance.encode(prompt, last_reply)
+                msg = json.dumps({'token_ids': token_ids, 'diff': diff})
+        elif '/decode' in self.path:
+            uid = req.get('uid')
+            token_ids = req.get('token_ids')
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                text = instance.decode(token_ids)
+                msg = json.dumps({'text': text})
+        elif '/reset' in self.path:
+            uid = req.get("uid")
+            system_prompt = req.get("system_prompt")
+            instance: Tokenizer_Http = tokenizers.get(uid)
+            if instance is None:
+                msg = json.dumps({'error': 'Invalid uid'})
+            else:
+                if system_prompt is not None:
+                    print("system_prompt:", system_prompt)
+                    token_ids = instance.reset(system_prompt)
+                    msg = json.dumps({'token_ids': token_ids})
+                else:
+                    token_ids = instance.reset()
+                    msg = json.dumps({'token_ids': token_ids})
+        else:
+            msg = json.dumps({'error': 'Invalid POST endpoint'})
+
+        print("响应消息:", msg)
+        self.wfile.write(msg.encode())
+
+    def get_query_param(self, key):
+        """
+        辅助函数：从 GET 请求的 URL 中获取查询参数的值
+        例如：/bos_id?uid=xxx
+        """
+        from urllib.parse import urlparse, parse_qs
+        query = urlparse(self.path).query
+        params = parse_qs(query)
+        values = params.get(key)
+        return values[0] if values else None
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--host', type=str, default='0.0.0.0')
+    parser.add_argument('--port', type=int, default=12345)
+    parser.add_argument('--model_id', type=str, default='qwen3_1.7B_tokenizer')
+    parser.add_argument('--content', type=str, default='You are Qwen, created by Alibaba Cloud. You are a helpful assistant.')
+
+    args = parser.parse_args()
+
+    host = (args.host, args.port)
+    print('Server running at http://%s:%s' % host)
+    server = HTTPServer(host, Request)
+    server.serve_forever()
diff --git a/projects/llm_framework/main_vlm/models/mode_internvl3-1B-ax630c.json b/projects/llm_framework/main_vlm/models/mode_internvl3-1B-448-ax630c.json
similarity index 64%
rename from projects/llm_framework/main_vlm/models/mode_internvl3-1B-ax630c.json
rename to projects/llm_framework/main_vlm/models/mode_internvl3-1B-448-ax630c.json
index 50848e20..84d7611d 100644
--- a/projects/llm_framework/main_vlm/models/mode_internvl3-1B-ax630c.json
+++ b/projects/llm_framework/main_vlm/models/mode_internvl3-1B-448-ax630c.json
@@ -1,5 +1,5 @@
 {
-    "mode": "internvl3-1B-ax630c",
+    "mode": "internvl3-1B-448-ax630c",
     "type": "vlm",
     "homepage": "https://huggingface.co/AXERA-TECH/InternVL3-1B",
     "capabilities": [
@@ -20,18 +20,25 @@
         "filename_tokens_embed": "model.embed_tokens.weight.bfloat16.bin",
         "filename_post_axmodel": "qwen2_post.axmodel",
         "template_filename_axmodel": "qwen2_p128_l%d_together.axmodel",
-        "filename_image_encoder_axmodedl": "internvl3_1b_vit.axmodel",
-        "b_bos": false,
-        "b_eos": false,
+        "filename_image_encoder_axmodel": "internvl3_1b_vit.axmodel",
+        "enable_temperature": true,
+        "temperature": 0.7,
+        "enable_top_p_sampling": false,
+        "top_p": 0.9,
+        "enable_top_k_sampling": true,
+        "top_k": 40,
+        "enable_repetition_penalty": false,
+        "repetition_penalty": 1.1,
+        "penalty_window": 50,
         "axmodel_num": 24,
         "tokens_embed_num": 151674,
         "img_token_id": 151667,
         "tokens_embed_size": 896,
         "b_use_mmap_load_embed": true,
-        "b_dynamic_load_axmodel_layer": false,
         "precompute_len": 1024,
+        "cmm_size": 1674992,
         "ext_scripts": [
-            "tokenizer_internvl3-1B-ax630c.py"
+            "tokenizer_internvl3-1B-448-ax630c.py"
         ]
     }
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/models/mode_internvl3-1B-448-ax650.json b/projects/llm_framework/main_vlm/models/mode_internvl3-1B-448-ax650.json
index 4da7a27e..91519756 100644
--- a/projects/llm_framework/main_vlm/models/mode_internvl3-1B-448-ax650.json
+++ b/projects/llm_framework/main_vlm/models/mode_internvl3-1B-448-ax650.json
@@ -21,15 +21,20 @@
         "filename_post_axmodel": "qwen2_post.axmodel",
         "template_filename_axmodel": "qwen2_p128_l%d_together.axmodel",
         "filename_image_encoder_axmodel": "internvl3_1b_vit.axmodel",
-        "b_use_topk": false,
-        "b_bos": false,
-        "b_eos": false,
+        "enable_temperature": true,
+        "temperature": 0.7,
+        "enable_top_p_sampling": false,
+        "top_p": 0.9,
+        "enable_top_k_sampling": true,
+        "top_k": 40,
+        "enable_repetition_penalty": false,
+        "repetition_penalty": 1.1,
+        "penalty_window": 50,
         "axmodel_num": 24,
         "tokens_embed_num": 151674,
         "img_token_id": 151667,
         "tokens_embed_size": 896,
         "b_use_mmap_load_embed": true,
-        "b_dynamic_load_axmodel_layer": false,
         "precompute_len": 2048,
         "cmm_size": 1919044,
         "ext_scripts": [
diff --git a/projects/llm_framework/main_vlm/scripts/tokenizer_internvl3-1B-ax630c.py b/projects/llm_framework/main_vlm/scripts/tokenizer_internvl3-1B-448-ax630c.py
similarity index 100%
rename from projects/llm_framework/main_vlm/scripts/tokenizer_internvl3-1B-ax630c.py
rename to projects/llm_framework/main_vlm/scripts/tokenizer_internvl3-1B-448-ax630c.py
diff --git a/projects/llm_framework/main_vlm/src/main.cpp b/projects/llm_framework/main_vlm/src/main.cpp
index 36f278a7..926976d2 100644
--- a/projects/llm_framework/main_vlm/src/main.cpp
+++ b/projects/llm_framework/main_vlm/src/main.cpp
@@ -154,7 +154,7 @@ class llm_task {
             CONFIG_AUTO_SET(file_body["mode_param"], url_tokenizer_model);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_tokens_embed);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_post_axmodel);
-            CONFIG_AUTO_SET(file_body["mode_param"], filename_vpm_resampler_axmodel);
+            CONFIG_AUTO_SET(file_body["mode_param"], filename_vpm_resampler_axmodedl);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_image_encoder_axmodel);
             CONFIG_AUTO_SET(file_body["mode_param"], template_filename_axmodel);
             CONFIG_AUTO_SET(file_body["mode_param"], b_use_topk);
@@ -238,7 +238,7 @@ class llm_task {
             mode_config_.filename_tokens_embed          = base_model + mode_config_.filename_tokens_embed;
             mode_config_.filename_post_axmodel          = base_model + mode_config_.filename_post_axmodel;
             mode_config_.template_filename_axmodel      = base_model + mode_config_.template_filename_axmodel;
-            mode_config_.filename_vpm_resampler_axmodel = base_model + mode_config_.filename_vpm_resampler_axmodel;
+            mode_config_.filename_vpm_resampler_axmodedl = base_model + mode_config_.filename_vpm_resampler_axmodedl;
             mode_config_.filename_image_encoder_axmodel = base_model + mode_config_.filename_image_encoder_axmodel;
             mode_config_.runing_callback = [this](int *p_token, int n_token, const char *p_str, float token_per_sec,
                                                   void *reserve) {
diff --git a/projects/llm_framework/main_vlm/src/runner/LLM.hpp b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
index ff232473..ecb2196d 100644
--- a/projects/llm_framework/main_vlm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
@@ -26,7 +26,7 @@ struct LLMAttrType {
     std::string filename_post_axmodel          = "tinyllama-int8/tinyllama_post.axmodel";
     std::string filename_image_encoder_axmodel = "minicpmv/vpm_resampler_version0_fp16.axmodel";
     std::string filename_vpm_encoder_axmodel   = "minicpmv/vpm_resampler_version0_fp16.axmodel";
-    std::string filename_vpm_resampler_axmodel = "minicpmv/vpm_resampler_version0_fp16.axmodel";
+    std::string filename_vpm_resampler_axmodedl = "minicpmv/vpm_resampler_version0_fp16.axmodel";
 
     int image_encoder_width  = 448;
     int image_encoder_height = 448;
@@ -190,18 +190,18 @@ class LLM {
                 return false;
             }
 
-            ret = vpm_resampler.init(attr.filename_vpm_resampler_axmodel.c_str(), false);
+            ret = vpm_resampler.init(attr.filename_vpm_resampler_axmodedl.c_str(), false);
             if (ret != 0) {
-                ALOGE("init vpm axmodel(%s) failed", attr.filename_vpm_resampler_axmodel.c_str());
+                ALOGE("init vpm axmodel(%s) failed", attr.filename_vpm_resampler_axmodedl.c_str());
                 return false;
             }
 
             _attr.vpm_height = vpm_encoder.get_input(0).vShape[1];
             _attr.vpm_width  = vpm_encoder.get_input(0).vShape[2];
         } else {
-            ret = vpm_resampler.init(attr.filename_vpm_resampler_axmodel.c_str(), false);
+            ret = vpm_resampler.init(attr.filename_vpm_resampler_axmodedl.c_str(), false);
             if (ret != 0) {
-                ALOGE("init vpm axmodel(%s) failed", attr.filename_vpm_resampler_axmodel.c_str());
+                ALOGE("init vpm axmodel(%s) failed", attr.filename_vpm_resampler_axmodedl.c_str());
                 return false;
             }
             _attr.vpm_height = vpm_resampler.get_input(0).vShape[1];
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index fc0097d2..5f8ab577 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -453,8 +453,11 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-model-llama3.2-1B-p256-ax630c':[create_data_deb,'llm-model-llama3.2-1B-p256-ax630c', '0.4', src_folder, revision],
         'llm-model-openbuddy-llama3.2-1B-ax630c':[create_data_deb,'llm-model-openbuddy-llama3.2-1B-ax630c', data_version, src_folder, revision],
         # InternVL model
+        ## AX630C
         'llm-model-internvl2.5-1B-ax630c':[create_data_deb,'llm-model-internvl2.5-1B-ax630c', '0.4', src_folder, revision],
         'llm-model-internvl2.5-1B-364-ax630c':[create_data_deb,'llm-model-internvl2.5-1B-364-ax630c', '0.4', src_folder, revision],
+        'llm-model-internvl3-1B-448-ax630c':[create_data_deb,'llm-model-internvl3-1B-448-ax630c', '0.6', src_folder, revision],
+        ## AX650
         'llm-model-internvl3-1B-448-ax650':[create_data_deb,'llm-model-internvl3-1B-448-ax650', '0.6', src_folder, revision],
         # DeepSeek model
         ## AX630C

From e628093302bd7409719696cc673a5eefe9273044 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 4 Sep 2025 18:39:47 +0800
Subject: [PATCH 38/79] [update] update asr kws llm vlm vad whisper melotts
 version

---
 projects/llm_framework/main_asr/SConstruct    |   2 +-
 projects/llm_framework/main_kws/SConstruct    |   2 +-
 projects/llm_framework/main_llm/SConstruct    |   2 +-
 projects/llm_framework/main_llm/src/main.cpp  |   2 +-
 .../llm_framework/main_llm/src/runner/LLM.hpp |   4 +-
 .../llm_framework/main_melotts/SConstruct     |   2 +-
 projects/llm_framework/main_vad/SConstruct    |   2 +-
 projects/llm_framework/main_vlm/SConstruct    |   2 +-
 projects/llm_framework/main_vlm/src/main.cpp  |   3 +-
 .../llm_framework/main_vlm/src/runner/LLM.hpp | 113 ++++++++++--------
 .../llm_framework/main_whisper/SConstruct     |   2 +-
 projects/llm_framework/tools/llm_pack.py      |  14 +--
 12 files changed, 79 insertions(+), 71 deletions(-)

diff --git a/projects/llm_framework/main_asr/SConstruct b/projects/llm_framework/main_asr/SConstruct
index d9ce42c1..cbc98e40 100644
--- a/projects/llm_framework/main_asr/SConstruct
+++ b/projects/llm_framework/main_asr/SConstruct
@@ -26,7 +26,7 @@ REQUIREMENTS += ['ncnn', 'sherpa-ncnn-core']
 
 STATIC_FILES += Glob('mode_*.json')
 
-env['COMPONENTS'].append({'target':'llm_asr-1.6',
+env['COMPONENTS'].append({'target':'llm_asr-1.7',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_kws/SConstruct b/projects/llm_framework/main_kws/SConstruct
index fc175abb..b46a1eb1 100644
--- a/projects/llm_framework/main_kws/SConstruct
+++ b/projects/llm_framework/main_kws/SConstruct
@@ -57,7 +57,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_kws-1.8',
+env['COMPONENTS'].append({'target':'llm_kws-1.9',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_llm/SConstruct b/projects/llm_framework/main_llm/SConstruct
index da100ad7..b6a3d191 100644
--- a/projects/llm_framework/main_llm/SConstruct
+++ b/projects/llm_framework/main_llm/SConstruct
@@ -66,7 +66,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_llm-1.8',
+env['COMPONENTS'].append({'target':'llm_llm-1.9',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_llm/src/main.cpp b/projects/llm_framework/main_llm/src/main.cpp
index 71ad2f7f..a845299a 100644
--- a/projects/llm_framework/main_llm/src/main.cpp
+++ b/projects/llm_framework/main_llm/src/main.cpp
@@ -224,7 +224,7 @@ class llm_task {
                 if (!process_field(mode_config_.filename_tokenizer_model, "filename_tokenizer_model") &&
                     !process_field(mode_config_.url_tokenizer_model, "url_tokenizer_model")) {
                     mode_config_.filename_tokenizer_model = base_model + mode_config_.filename_tokenizer_model;
-                    SLOGE("filename_tokenizer_model: %s", mode_config_.filename_tokenizer_model.c_str());
+                    SLOGI("filename_tokenizer_model: %s", mode_config_.filename_tokenizer_model.c_str());
                 }
             }
             mode_config_.filename_tokens_embed     = base_model + mode_config_.filename_tokens_embed;
diff --git a/projects/llm_framework/main_llm/src/runner/LLM.hpp b/projects/llm_framework/main_llm/src/runner/LLM.hpp
index 5639fb89..9de1f1fe 100644
--- a/projects/llm_framework/main_llm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_llm/src/runner/LLM.hpp
@@ -55,8 +55,8 @@ struct LLMAttrType {
     bool enable_top_p_sampling = false;
     float top_p                = 0.7f;
 
-    bool enable_top_k_sampling = false;
-    int top_k                  = 50;
+    bool enable_top_k_sampling = true;
+    int top_k                  = 10;
 
     bool enable_repetition_penalty = false;
     float repetition_penalty       = 1.2f;
diff --git a/projects/llm_framework/main_melotts/SConstruct b/projects/llm_framework/main_melotts/SConstruct
index 0ea5d61f..5466d8e1 100644
--- a/projects/llm_framework/main_melotts/SConstruct
+++ b/projects/llm_framework/main_melotts/SConstruct
@@ -32,7 +32,7 @@ REQUIREMENTS += ['onnxruntime']
 
 STATIC_FILES += Glob('models/mode_*.json')
 
-env['COMPONENTS'].append({'target':'llm_melotts-1.8',
+env['COMPONENTS'].append({'target':'llm_melotts-1.9',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_vad/SConstruct b/projects/llm_framework/main_vad/SConstruct
index 3c1b4b05..2d802067 100644
--- a/projects/llm_framework/main_vad/SConstruct
+++ b/projects/llm_framework/main_vad/SConstruct
@@ -29,7 +29,7 @@ REQUIREMENTS += ['onnxruntime']
 
 STATIC_FILES += Glob('mode_*.json')
 
-env['COMPONENTS'].append({'target':'llm_vad-1.7',
+env['COMPONENTS'].append({'target':'llm_vad-1.8',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_vlm/SConstruct b/projects/llm_framework/main_vlm/SConstruct
index b5f429f3..2bb327ba 100644
--- a/projects/llm_framework/main_vlm/SConstruct
+++ b/projects/llm_framework/main_vlm/SConstruct
@@ -73,7 +73,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_vlm-1.8',
+env['COMPONENTS'].append({'target':'llm_vlm-1.9',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_vlm/src/main.cpp b/projects/llm_framework/main_vlm/src/main.cpp
index 926976d2..7b5d5ead 100644
--- a/projects/llm_framework/main_vlm/src/main.cpp
+++ b/projects/llm_framework/main_vlm/src/main.cpp
@@ -157,7 +157,6 @@ class llm_task {
             CONFIG_AUTO_SET(file_body["mode_param"], filename_vpm_resampler_axmodedl);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_image_encoder_axmodel);
             CONFIG_AUTO_SET(file_body["mode_param"], template_filename_axmodel);
-            CONFIG_AUTO_SET(file_body["mode_param"], b_use_topk);
             CONFIG_AUTO_SET(file_body["mode_param"], b_vpm_two_stage);
             CONFIG_AUTO_SET(file_body["mode_param"], b_bos);
             CONFIG_AUTO_SET(file_body["mode_param"], b_eos);
@@ -232,7 +231,7 @@ class llm_task {
                 if (!process_field(mode_config_.filename_tokenizer_model, "filename_tokenizer_model") &&
                     !process_field(mode_config_.url_tokenizer_model, "url_tokenizer_model")) {
                     mode_config_.filename_tokenizer_model = base_model + mode_config_.filename_tokenizer_model;
-                    SLOGE("filename_tokenizer_model: %s", mode_config_.filename_tokenizer_model.c_str());
+                    SLOGI("filename_tokenizer_model: %s", mode_config_.filename_tokenizer_model.c_str());
                 }
             }
             mode_config_.filename_tokens_embed          = base_model + mode_config_.filename_tokens_embed;
diff --git a/projects/llm_framework/main_vlm/src/runner/LLM.hpp b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
index ecb2196d..0caf4641 100644
--- a/projects/llm_framework/main_vlm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
@@ -20,48 +20,56 @@ typedef std::function<void(int *, int, const char *, float, void *)> LLMRuningCa
 
 struct LLMAttrType {
     std::string system_prompt;
+
     std::string template_filename_axmodel = "tinyllama-int8/tinyllama_l%d.axmodel";
+    std::string post_config_path          = "post_config.json";
     int axmodel_num                       = 22;
 
-    std::string filename_post_axmodel          = "tinyllama-int8/tinyllama_post.axmodel";
-    std::string filename_image_encoder_axmodel = "minicpmv/vpm_resampler_version0_fp16.axmodel";
-    std::string filename_vpm_encoder_axmodel   = "minicpmv/vpm_resampler_version0_fp16.axmodel";
+    std::string filename_image_encoder_axmodel  = "minicpmv/vpm_resampler_version0_fp16.axmodel";
+    std::string filename_vpm_encoder_axmodel    = "minicpmv/vpm_resampler_version0_fp16.axmodel";
     std::string filename_vpm_resampler_axmodedl = "minicpmv/vpm_resampler_version0_fp16.axmodel";
 
-    int image_encoder_width  = 448;
-    int image_encoder_height = 448;
-    int vpm_width            = 280;
-    int vpm_height           = 280;
-    bool b_vpm_two_stage     = false;
+    int image_encoder_width       = 448;
+    int image_encoder_height      = 448;
+    int vpm_width                 = 280;
+    int vpm_height                = 280;
+    bool b_vpm_two_stage          = false;
+    int IMAGE_CONTEXT_TOKEN       = 151667;
+    int IMAGE_START_TOKEN         = 151665;
+    int IMAGE_ENCODER_INPUT_NCHW  = -1;
+    int IMAGE_ENCODER_OUTPUT_BF16 = -1;
 
     int prefill_token_num     = 96;
     int prefill_max_token_num = 512;
-    std::vector<int> prefill_max_kv_cache_num_grp;
-    int precompute_len = 0;
-    int prefill_grpid  = -1;
+
+    std::string filename_post_axmodel = "tinyllama-int8/tinyllama_post.axmodel";
 
     TokenizerType tokenizer_type         = TKT_LLaMa;
     std::string filename_tokenizer_model = "tokenizer.model";
     std::string url_tokenizer_model;
-    bool b_bos = true, b_eos = false;
+    bool b_bos                        = true;
+    bool b_eos                        = false;
     std::string filename_tokens_embed = "tinyllama.model.embed_tokens.weight.bfloat16.bin";
     int tokens_embed_num              = 32000;
     int img_token_id                  = 151667;
     int tokens_embed_size             = 2048;
 
     int max_token_len = 127;
-
     int kv_cache_num  = 1024;
     int kv_cache_size = 256;
 
+    int precompute_len = 0;
+    std::vector<int> prefill_max_kv_cache_num_grp;
+    int prefill_grpid = -1;
+
     bool enable_temperature = false;
     float temperature       = 0.7f;
 
     bool enable_top_p_sampling = false;
     float top_p                = 0.7f;
 
-    bool enable_top_k_sampling = false;
-    int top_k                  = 50;
+    bool enable_top_k_sampling = true;
+    int top_k                  = 10;
 
     bool enable_repetition_penalty = false;
     float repetition_penalty       = 1.2f;
@@ -69,20 +77,10 @@ struct LLMAttrType {
 
     bool b_use_mmap_load_embed        = false;
     bool b_dynamic_load_axmodel_layer = false;
+    bool b_use_mmap_load_layer        = true;
 
-    bool b_use_mmap_load_layer = true;
-
-    bool b_use_topk              = false;
-    std::string post_config_path = "post_config.json";
-
-    // bool b_live_print = true;
     LLMRuningCallback runing_callback = nullptr;
     void *reserve                     = nullptr;
-
-    int IMAGE_CONTEXT_TOKEN       = 151667;
-    int IMAGE_START_TOKEN         = 151665;
-    int IMAGE_ENCODER_INPUT_NCHW  = -1;
-    int IMAGE_ENCODER_OUTPUT_BF16 = -1;
 };
 
 class LLM {
@@ -142,7 +140,6 @@ class LLM {
             return false;
         }
         update_cqdm(&cqdm, 1, "count", "embed_selector init ok");
-
         llama_layers.resize(attr.axmodel_num);
 
         char axmodel_path[1024];
@@ -241,13 +238,34 @@ class LLM {
 
             _attr.prefill_token_num = llama_layers[0].layer.get_input(prefill_grpid, "indices").vShape[1];
             ALOGI("prefill_token_num : %d", _attr.prefill_token_num);
-
             ALOGI("vpm_height : %d,vpm_width : %d", _attr.vpm_height, _attr.vpm_width);
         }
         if (attr.b_dynamic_load_axmodel_layer) {
             auto &layer = llama_layers[0];
             layer.layer.deinit();
         }
+        nlohmann::json dynamic_config;
+
+        dynamic_config["enable_temperature"] = _attr.enable_temperature;
+        dynamic_config["temperature"]        = _attr.temperature;
+
+        dynamic_config["enable_repetition_penalty"] = _attr.enable_repetition_penalty;
+        dynamic_config["repetition_penalty"]        = _attr.repetition_penalty;
+        dynamic_config["penalty_window"]            = _attr.penalty_window;
+
+        dynamic_config["enable_top_p_sampling"] = _attr.enable_top_p_sampling;
+        dynamic_config["top_p"]                 = _attr.top_p;
+
+        dynamic_config["enable_top_k_sampling"] = _attr.enable_top_k_sampling;
+        dynamic_config["top_k"]                 = _attr.top_k;
+
+        if (!postprocess.load_config(attr.post_config_path)) {
+            ALOGW("load postprocess config(%s) failed", attr.post_config_path.c_str());
+        }
+
+        if (!postprocess.load_config(dynamic_config)) {
+            ALOGW("load postprocess config(%s) failed", dynamic_config.dump(4).c_str());
+        }
 
         // Reset();
         ALOGI("LLM init ok");
@@ -483,19 +501,15 @@ class LLM {
             auto &input = llama_post.get_input("input");
             memcpy(input.pVirAddr, embed.data(), embed.size() * sizeof(unsigned short));
             llama_post.inference();
+
             int max_index;
-            if (_attr.b_use_topk) {
-                AX_SYS_MinvalidateCache(llama_post.get_output("indices").phyAddr,
-                                        llama_post.get_output("indices").pVirAddr,
-                                        llama_post.get_output("indices").nSize);
-                max_index = *(int *)llama_post.get_output("indices").pVirAddr;
-            } else {
-                auto &output_post = llama_post.get_output("output");
-                AX_SYS_MinvalidateCache(output_post.phyAddr, output_post.pVirAddr, output_post.nSize);
-                unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
-                float max_val            = -MAXFLOAT;
-                max_index = post_process(postprocess, post_out, _attr.tokens_embed_num, token_ids, &max_val);
-            }
+
+            auto &output_post = llama_post.get_output("output");
+            AX_SYS_MinvalidateCache(output_post.phyAddr, output_post.pVirAddr, output_post.nSize);
+            unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
+            float max_val            = -MAXFLOAT;
+            max_index                = post_process(postprocess, post_out, _attr.tokens_embed_num, token_ids, &max_val);
+
             next_token = max_index;
 
             token_ids.push_back(max_index);
@@ -574,18 +588,13 @@ class LLM {
                 memcpy(input.pVirAddr, embed.data(), embed.size() * sizeof(unsigned short));
                 llama_post.inference();
                 int max_index;
-                if (_attr.b_use_topk) {
-                    AX_SYS_MinvalidateCache(llama_post.get_output("indices").phyAddr,
-                                            llama_post.get_output("indices").pVirAddr,
-                                            llama_post.get_output("indices").nSize);
-                    max_index = *(int *)llama_post.get_output("indices").pVirAddr;
-                } else {
-                    auto &output_post = llama_post.get_output("output");
-                    AX_SYS_MinvalidateCache(output_post.phyAddr, output_post.pVirAddr, output_post.nSize);
-                    unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
-                    float max_val            = -MAXFLOAT;
-                    max_index = post_process(postprocess, post_out, _attr.tokens_embed_num, token_ids, &max_val);
-                }
+
+                auto &output_post = llama_post.get_output("output");
+                AX_SYS_MinvalidateCache(output_post.phyAddr, output_post.pVirAddr, output_post.nSize);
+                unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
+                float max_val            = -MAXFLOAT;
+                max_index = post_process(postprocess, post_out, _attr.tokens_embed_num, token_ids, &max_val);
+
                 next_token = max_index;
 
                 if (tokenizer->isEnd(max_index)) {
diff --git a/projects/llm_framework/main_whisper/SConstruct b/projects/llm_framework/main_whisper/SConstruct
index bc4d84b5..6b7d54b0 100644
--- a/projects/llm_framework/main_whisper/SConstruct
+++ b/projects/llm_framework/main_whisper/SConstruct
@@ -33,7 +33,7 @@ LDFLAGS += ['-l:libopencc.a', '-l:libmarisa.a']
 
 STATIC_FILES += Glob('models/mode_*.json')
 
-env['COMPONENTS'].append({'target':'llm_whisper-1.7',
+env['COMPONENTS'].append({'target':'llm_whisper-1.8',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 5f8ab577..aa1ecb25 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -356,18 +356,18 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'lib-llm':[create_lib_deb,'lib-llm', '1.8', src_folder, revision],
         'llm-sys':[create_bin_deb,'llm-sys', '1.6', src_folder, revision],
         'llm-audio':[create_bin_deb,'llm-audio', '1.6', src_folder, revision],
-        'llm-kws':[create_bin_deb,'llm-kws', '1.8', src_folder, revision],
-        'llm-asr':[create_bin_deb,'llm-asr', '1.6', src_folder, revision],
-        'llm-llm':[create_bin_deb,'llm-llm', '1.8', src_folder, revision],
+        'llm-kws':[create_bin_deb,'llm-kws', '1.9', src_folder, revision],
+        'llm-asr':[create_bin_deb,'llm-asr', '1.7', src_folder, revision],
+        'llm-llm':[create_bin_deb,'llm-llm', '1.9', src_folder, revision],
         'llm-tts':[create_bin_deb,'llm-tts', '1.6', src_folder, revision],
-        'llm-melotts':[create_bin_deb,'llm-melotts', '1.8', src_folder, revision],
+        'llm-melotts':[create_bin_deb,'llm-melotts', '1.9', src_folder, revision],
         'llm-camera':[create_bin_deb,'llm-camera', '1.9', src_folder, revision, 'lib-llm'],
-        'llm-vlm':[create_bin_deb,'llm-vlm', '1.8', src_folder, revision],
+        'llm-vlm':[create_bin_deb,'llm-vlm', '1.9', src_folder, revision],
         'llm-yolo':[create_bin_deb,'llm-yolo', '1.9', src_folder, revision],
         'llm-skel':[create_bin_deb,'llm-skel', version, src_folder, revision],
         'llm-depth-anything':[create_bin_deb,'llm-depth-anything', '1.7', src_folder, revision],
-        'llm-vad':[create_bin_deb,'llm-vad', '1.7', src_folder, revision],
-        'llm-whisper':[create_bin_deb,'llm-whisper', '1.7', src_folder, revision],
+        'llm-vad':[create_bin_deb,'llm-vad', '1.8', src_folder, revision],
+        'llm-whisper':[create_bin_deb,'llm-whisper', '1.8', src_folder, revision],
         'llm-openai-api':[create_bin_deb,'llm-openai-api', '1.7', src_folder, revision],
         # keyword spotting Audio file
         'llm-model-audio-en-us':[create_data_deb,'llm-model-audio-en-us', data_version, src_folder, revision],

From a7d82afcecefc5dfcd1f649ab4830816112110d0 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Mon, 8 Sep 2025 18:05:59 +0800
Subject: [PATCH 39/79] [fix] fix alsa audio cap

---
 projects/llm_framework/main_audio/src/alsa_audio.c | 4 +++-
 projects/llm_framework/main_audio/src/main.cpp     | 5 +++++
 2 files changed, 8 insertions(+), 1 deletion(-)

diff --git a/projects/llm_framework/main_audio/src/alsa_audio.c b/projects/llm_framework/main_audio/src/alsa_audio.c
index 73aca0f4..944497a9 100644
--- a/projects/llm_framework/main_audio/src/alsa_audio.c
+++ b/projects/llm_framework/main_audio/src/alsa_audio.c
@@ -11,6 +11,7 @@ static int gcapLoopExit = 0;
 void alsa_cap_start(unsigned int card, unsigned int device, float Volume, int channel, int rate, int bit,
                     AUDIOCallback callback)
 {
+    gcapLoopExit = 0;
     struct pcm_config config;
     unsigned int pcm_open_flags;
     struct pcm *pcm;
@@ -23,7 +24,7 @@ void alsa_cap_start(unsigned int card, unsigned int device, float Volume, int ch
     memset(&config, 0, sizeof(config));
     config.channels          = channel;
     config.rate              = 48000;  // TODO: 部分USB MIC仅支持48k，暂时固定采集为48k
-    config.period_size       = 512;
+    config.period_size       = 120;
     config.period_count      = 4;
     config.format            = PCM_FORMAT_S16_LE;
     config.start_threshold   = 0;
@@ -123,6 +124,7 @@ void alsa_cap_start(unsigned int card, unsigned int device, float Volume, int ch
     }
     free(buffer);
     pcm_close(pcm);
+    printf("Total frames captured: %u\n", total_frames_read);
 }
 
 void alsa_close_cap()
diff --git a/projects/llm_framework/main_audio/src/main.cpp b/projects/llm_framework/main_audio/src/main.cpp
index fdd25063..a892e3b2 100644
--- a/projects/llm_framework/main_audio/src/main.cpp
+++ b/projects/llm_framework/main_audio/src/main.cpp
@@ -105,7 +105,12 @@ class llm_audio : public StackFlow {
 
     void _cap_stop()
     {
+#if (defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)) && !defined(CONFIG_AXCL_ENABLED)
         ax_close_cap();
+
+#else
+        alsa_close_cap();
+#endif
         if (audio_cap_thread_) {
             audio_cap_thread_->join();
             audio_cap_thread_.reset();

From bb482368e09717203f57d878a5f5af4a003da0b0 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Mon, 15 Sep 2025 17:29:57 +0800
Subject: [PATCH 40/79] [update] add cosyvioce2

---
 .../llm_framework/main_cosy_voice/Kconfig     |     0
 .../llm_framework/main_cosy_voice/SConstruct  |    84 +
 .../models/mode_CosyVoice2-0.5B-ax650.json    |    44 +
 .../multilingual_zh_ja_yue_char_del.tiktoken  | 58836 ++++++++++++++++
 .../scripts/tokenizer/tokenizer.py            |   279 +
 .../tokenizer_CosyVoice2-0.5B-ax650.py        |   122 +
 .../main_cosy_voice/src/main.cpp              |   885 +
 .../main_cosy_voice/src/runner/LLM.hpp        |   705 +
 .../src/runner/LLMEmbedSelector.hpp           |   110 +
 .../src/runner/LLMPostprocess.hpp             |   363 +
 .../main_cosy_voice/src/runner/Token2wav.hpp  |   635 +
 .../src/runner/Tokenizer/QwenTokenizer.cpp    |   127 +
 .../src/runner/Tokenizer/QwenTokenizer.hpp    |    34 +
 .../src/runner/Tokenizer/Tokenizer.cpp        |   912 +
 .../src/runner/Tokenizer/Tokenizer.hpp        |    80 +
 .../src/runner/Tokenizer/base64.h             |    54 +
 .../src/runner/Tokenizer/tiktoken.h           |   327 +
 .../src/runner/Tokenizer/unordered_dense.h    |  2240 +
 .../ax_model_runner/ax_model_runner.hpp       |   157 +
 .../ax_model_runner/ax_model_runner_ax650.cpp |   454 +
 .../ax_model_runner/ax_model_runner_ax650.hpp |    22 +
 .../src/runner/utils/ax_cmm_utils.hpp         |    43 +
 .../src/runner/utils/bfloat16.hpp             |    71 +
 .../src/runner/utils/concat_3d.h              |    96 +
 .../main_cosy_voice/src/runner/utils/cqdm.cpp |    48 +
 .../main_cosy_voice/src/runner/utils/cqdm.h   |    28 +
 .../src/runner/utils/http_utils.hpp           |   102 +
 .../src/runner/utils/httplib.h                |  9483 +++
 .../main_cosy_voice/src/runner/utils/json.hpp | 24766 +++++++
 .../src/runner/utils/memory_utils.cpp         |    65 +
 .../src/runner/utils/memory_utils.hpp         |    74 +
 .../src/runner/utils/sample_log.h             |    65 +
 .../src/runner/utils/sampling.hpp             |   200 +
 .../src/runner/utils/slice_3d.h               |    44 +
 .../src/runner/utils/slice_3d.tpp             |   253 +
 .../src/runner/utils/string_utility.hpp       |   217 +
 .../src/runner/utils/timer.hpp                |    57 +
 .../src/runner/utils/utils.hpp                |    97 +
 .../main_cosy_voice/src/runner/utils/wav.hpp  |   160 +
 39 files changed, 102339 insertions(+)
 create mode 100644 projects/llm_framework/main_cosy_voice/Kconfig
 create mode 100644 projects/llm_framework/main_cosy_voice/SConstruct
 create mode 100644 projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json
 create mode 100755 projects/llm_framework/main_cosy_voice/scripts/tokenizer/assets/multilingual_zh_ja_yue_char_del.tiktoken
 create mode 100755 projects/llm_framework/main_cosy_voice/scripts/tokenizer/tokenizer.py
 create mode 100755 projects/llm_framework/main_cosy_voice/scripts/tokenizer_CosyVoice2-0.5B-ax650.py
 create mode 100644 projects/llm_framework/main_cosy_voice/src/main.cpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/LLMEmbedSelector.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/LLMPostprocess.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/QwenTokenizer.cpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/QwenTokenizer.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/base64.h
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/tiktoken.h
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/unordered_dense.h
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.cpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/ax_cmm_utils.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/bfloat16.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/concat_3d.h
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/cqdm.cpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/cqdm.h
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/http_utils.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/httplib.h
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/json.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/memory_utils.cpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/memory_utils.hpp
 create mode 100755 projects/llm_framework/main_cosy_voice/src/runner/utils/sample_log.h
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/sampling.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/slice_3d.h
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/slice_3d.tpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/string_utility.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/timer.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/utils.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/utils/wav.hpp

diff --git a/projects/llm_framework/main_cosy_voice/Kconfig b/projects/llm_framework/main_cosy_voice/Kconfig
new file mode 100644
index 00000000..e69de29b
diff --git a/projects/llm_framework/main_cosy_voice/SConstruct b/projects/llm_framework/main_cosy_voice/SConstruct
new file mode 100644
index 00000000..f82f28aa
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/SConstruct
@@ -0,0 +1,84 @@
+import os
+
+Import('env')
+with open(env['PROJECT_TOOL_S']) as f:
+    exec(f.read())
+
+
+SRCS = append_srcs_dir(ADir('src'))
+INCLUDE = [ADir('include'), ADir('.')]
+PRIVATE_INCLUDE = [ADir('../include')]
+REQUIREMENTS = ['pthread','atomic','rt', 'dl', 'm', 'ax_msp', 'utilities', 'eventpp', 'StackFlow']
+STATIC_LIB = []
+DYNAMIC_LIB = []
+DEFINITIONS = []
+DEFINITIONS_PRIVATE = []
+LDFLAGS = []
+LINK_SEARCH_PATH = []
+STATIC_FILES = []
+
+python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-llm-python-venv_v1.7.tar.gz", 'm5stack_llm-llm-python-venv_v1.7.tar.gz')
+
+# REQUIREMENTS += ['Backward_cpp']
+# DYNAMIC_LIB += [ AFile('../static_lib/libdw.so.1'),
+#                 AFile('../static_lib/libelf.so.1'),
+#                 AFile('../static_lib/libz.so.1'),
+#                 AFile('../static_lib/liblzma.so.5'),
+#                 AFile('../static_lib/libbz2.so.1.0')]
+# DEFINITIONS += ["-DENABLE_BACKWARD"]
+DEFINITIONS += ['-O2']
+DEFINITIONS += ['-std=c++17']
+LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
+REQUIREMENTS += ['ax_engine', 'ax_interpreter', 'ax_sys', 'utilities']
+LINK_SEARCH_PATH += [ADir('../static_lib')]
+
+
+INCLUDE += [ADir('src/runner'), ADir('src/runner/utils')]
+INCLUDE += [ADir('../static_lib/include/sentencepiece'),
+            ADir('../static_lib/include/protobuf-lite'),
+            ADir('../static_lib/include/abseil-cpp'),
+            ADir('../static_lib/include/re2')]
+
+static_file = Glob('../static_lib/module-llm/libabsl_*')
+static_file += [AFile('../static_lib/module-llm/libre2.a'), AFile('../static_lib/module-llm/libsentencepiece.a'), AFile('../static_lib/module-llm/libsentencepiece_train.a')]
+STATIC_LIB += static_file * 4
+
+# STATIC_FILES += [os.path.join(python_venv, 'llm')]
+# STATIC_FILES += Glob('scripts/tokenizer_*.py')
+STATIC_FILES += Glob('scripts/tokenizer*')
+STATIC_FILES += Glob('models/mode_*.json')
+# STATIC_FILES += [AFile('scripts/llm-llm_tokenizer_auto.py')]
+
+IGNORE_FILES = []
+IGNORE_FILES += ['llm']
+
+import json
+if not os.path.exists('../dist'):
+    os.makedirs('../dist')
+ignore = {'ignore':[]}
+try:
+    with open('../dist/fileignore', 'a+') as f:
+        f.seek(0)
+        ignore = json.load(f)
+except:
+    pass
+ignore['ignore'] += IGNORE_FILES
+ignore['ignore'] = list(set(ignore['ignore']))
+with open('../dist/fileignore', 'w') as f:
+    json.dump(ignore, f, indent=4)
+
+env['COMPONENTS'].append({'target':'llm_cosy_voice-1.8',
+                          'SRCS':SRCS,
+                          'INCLUDE':INCLUDE,
+                          'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
+                          'REQUIREMENTS':REQUIREMENTS,
+                          'STATIC_LIB':STATIC_LIB,
+                          'DYNAMIC_LIB':DYNAMIC_LIB,
+                          'DEFINITIONS':DEFINITIONS,
+                          'DEFINITIONS_PRIVATE':DEFINITIONS_PRIVATE,
+                          'LDFLAGS':LDFLAGS,
+                          'LINK_SEARCH_PATH':LINK_SEARCH_PATH,
+                          'STATIC_FILES':STATIC_FILES,
+                          'REGISTER':'project'
+                          })
+
diff --git a/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json b/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json
new file mode 100644
index 00000000..2955af82
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json
@@ -0,0 +1,44 @@
+{
+    "mode": "CosyVoice2-0.5B-ax650",
+    "type": "tts",
+    "homepage": "https://huggingface.co/FunAudioLLM/CosyVoice2-0.5B",
+    "capabilities": [
+        "tts",
+        "clone_voice"
+    ],
+    "input_type": [
+        "tts.utf-8"
+    ],
+    "output_type": [
+        "tts.wav",
+        "sys.play.0_1"
+    ],
+    "mode_param": {
+        "tokenizer_type": 2,
+        "filename_tokenizer_model": "http://localhost:8080",
+        "filename_tokens_embed": "model.embed_tokens.weight.bfloat16.bin",
+        "filename_llm_embed": "llm.llm_embedding.float16.bin",
+        "filename_speech_embed": "llm.speech_embedding.float16.bin",
+        "filename_post_axmodel": "qwen2_post.axmodel",
+        "filename_decoder_axmodel": "llm_decoder.axmodel",
+        "template_filename_axmodel": "qwen2_p128_l%d_together.axmodel",
+        "token2wav_axmodel_dir": "",
+        "prompt_files": "zh_man1",
+        "b_use_topk": false,
+        "b_bos": false,
+        "b_eos": false,
+        "axmodel_num": 24,
+        "tokens_embed_num": 151936,
+        "tokens_embed_size": 896,
+        "llm_embed_num": 2,
+        "llm_embed_size": 896,
+        "speech_embed_num": 6564,
+        "speech_embed_size": 896,
+        "n_timesteps": 10,
+        "b_use_mmap_load_embed": true,
+        "b_dynamic_load_axmodel_layer": false,
+        "ext_scripts": [
+            "tokenizer_cosyvoice2-0.5B-ax650.py"
+        ]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/scripts/tokenizer/assets/multilingual_zh_ja_yue_char_del.tiktoken b/projects/llm_framework/main_cosy_voice/scripts/tokenizer/assets/multilingual_zh_ja_yue_char_del.tiktoken
new file mode 100755
index 00000000..ac47fe68
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/scripts/tokenizer/assets/multilingual_zh_ja_yue_char_del.tiktoken
@@ -0,0 +1,58836 @@
+IQ== 0
+Ig== 1
+Iw== 2
+JA== 3
+JQ== 4
+Jg== 5
+Jw== 6
+KA== 7
+KQ== 8
+Kg== 9
+Kw== 10
+LA== 11
+LQ== 12
+Lg== 13
+Lw== 14
+MA== 15
+MQ== 16
+Mg== 17
+Mw== 18
+NA== 19
+NQ== 20
+Ng== 21
+Nw== 22
+OA== 23
+OQ== 24
+Og== 25
+Ow== 26
+PA== 27
+PQ== 28
+Pg== 29
+Pw== 30
+QA== 31
+QQ== 32
+Qg== 33
+Qw== 34
+RA== 35
+RQ== 36
+Rg== 37
+Rw== 38
+SA== 39
+SQ== 40
+Sg== 41
+Sw== 42
+TA== 43
+TQ== 44
+Tg== 45
+Tw== 46
+UA== 47
+UQ== 48
+Ug== 49
+Uw== 50
+VA== 51
+VQ== 52
+Vg== 53
+Vw== 54
+WA== 55
+WQ== 56
+Wg== 57
+Ww== 58
+XA== 59
+XQ== 60
+Xg== 61
+Xw== 62
+YA== 63
+YQ== 64
+Yg== 65
+Yw== 66
+ZA== 67
+ZQ== 68
+Zg== 69
+Zw== 70
+aA== 71
+aQ== 72
+ag== 73
+aw== 74
+bA== 75
+bQ== 76
+bg== 77
+bw== 78
+cA== 79
+cQ== 80
+cg== 81
+cw== 82
+dA== 83
+dQ== 84
+dg== 85
+dw== 86
+eA== 87
+eQ== 88
+eg== 89
+ew== 90
+fA== 91
+fQ== 92
+fg== 93
+oQ== 94
+og== 95
+ow== 96
+pA== 97
+pQ== 98
+pg== 99
+pw== 100
+qA== 101
+qQ== 102
+qg== 103
+qw== 104
+rA== 105
+rg== 106
+rw== 107
+sA== 108
+sQ== 109
+sg== 110
+sw== 111
+tA== 112
+tQ== 113
+tg== 114
+tw== 115
+uA== 116
+uQ== 117
+ug== 118
+uw== 119
+vA== 120
+vQ== 121
+vg== 122
+vw== 123
+wA== 124
+wQ== 125
+wg== 126
+ww== 127
+xA== 128
+xQ== 129
+xg== 130
+xw== 131
+yA== 132
+yQ== 133
+yg== 134
+yw== 135
+zA== 136
+zQ== 137
+zg== 138
+zw== 139
+0A== 140
+0Q== 141
+0g== 142
+0w== 143
+1A== 144
+1Q== 145
+1g== 146
+1w== 147
+2A== 148
+2Q== 149
+2g== 150
+2w== 151
+3A== 152
+3Q== 153
+3g== 154
+3w== 155
+4A== 156
+4Q== 157
+4g== 158
+4w== 159
+5A== 160
+5Q== 161
+5g== 162
+5w== 163
+6A== 164
+6Q== 165
+6g== 166
+6w== 167
+7A== 168
+7Q== 169
+7g== 170
+7w== 171
+8A== 172
+8Q== 173
+8g== 174
+8w== 175
+9A== 176
+9Q== 177
+9g== 178
+9w== 179
++A== 180
++Q== 181
++g== 182
++w== 183
+/A== 184
+/Q== 185
+/g== 186
+/w== 187
+AA== 188
+AQ== 189
+Ag== 190
+Aw== 191
+BA== 192
+BQ== 193
+Bg== 194
+Bw== 195
+CA== 196
+CQ== 197
+Cg== 198
+Cw== 199
+DA== 200
+DQ== 201
+Dg== 202
+Dw== 203
+EA== 204
+EQ== 205
+Eg== 206
+Ew== 207
+FA== 208
+FQ== 209
+Fg== 210
+Fw== 211
+GA== 212
+GQ== 213
+Gg== 214
+Gw== 215
+HA== 216
+HQ== 217
+Hg== 218
+Hw== 219
+IA== 220
+fw== 221
+gA== 222
+gQ== 223
+gg== 224
+gw== 225
+hA== 226
+hQ== 227
+hg== 228
+hw== 229
+iA== 230
+iQ== 231
+ig== 232
+iw== 233
+jA== 234
+jQ== 235
+jg== 236
+jw== 237
+kA== 238
+kQ== 239
+kg== 240
+kw== 241
+lA== 242
+lQ== 243
+lg== 244
+lw== 245
+mA== 246
+mQ== 247
+mg== 248
+mw== 249
+nA== 250
+nQ== 251
+ng== 252
+nw== 253
+oA== 254
+rQ== 255
+IHQ= 256
+IGE= 257
+IHRo 258
+aW4= 259
+ZXI= 260
+IHc= 261
+IHM= 262
+b3U= 263
+IHRoZQ== 264
+cmU= 265
+b24= 266
+YXQ= 267
+ZW4= 268
+IGM= 269
+aXQ= 270
+aXM= 271
+IGI= 272
+bmQ= 273
+IGQ= 274
+IG0= 275
+IGg= 276
+IG8= 277
+aW5n 278
+ZXM= 279
+IHA= 280
+IHRv 281
+YW4= 282
+IGY= 283
+b3I= 284
+bGw= 285
+IEk= 286
+IGw= 287
+IHk= 288
+YXI= 289
+IGc= 290
+IHlvdQ== 291
+ZWQ= 292
+IGFuZA== 293
+IGlu 294
+IG9m 295
+YXM= 296
+IG4= 297
+b20= 298
+aWM= 299
+IHRoYXQ= 300
+dXM= 301
+ZXQ= 302
+dmU= 303
+YWw= 304
+b3c= 305
+bGU= 306
+IGlz 307
+IGU= 308
+IGl0 309
+b3Q= 310
+J3M= 311
+IGJl 312
+aW9u 313
+IFQ= 314
+IHdo 315
+IEE= 316
+ZW50 317
+IFM= 318
+IHJl 319
+YXk= 320
+IHdl 321
+IG9u 322
+ZXJl 323
+IGhh 324
+dXQ= 325
+YWM= 326
+aWQ= 327
+aWc= 328
+b3M= 329
+a2U= 330
+dmVy 331
+aW0= 332
+INA= 333
+IFRo 334
+YW0= 335
+YWxs 336
+IGZvcg== 337
+ZWw= 338
+Y2g= 339
+cm8= 340
+IHRoaXM= 341
+IHN0 342
+IFc= 343
+IHU= 344
+YWQ= 345
+b3V0 346
+aXI= 347
+bGQ= 348
+Y3Q= 349
+IGs= 350
+aWY= 351
+IGdv 352
+Li4= 353
+0L4= 354
+aXRo 355
+bHk= 356
+aHQ= 357
+cXU= 358
+IC0= 359
+IGRv 360
+IGo= 361
+IGhhdmU= 362
+IEI= 363
+IGFu 364
+IHdpdGg= 365
+IGFyZQ== 366
+IHI= 367
+IGRl 368
+IHNl 369
+IHNv 370
+IHY= 371
+c3Q= 372
+aWxs 373
+dXI= 374
+IGxp 375
+IE0= 376
+ZXN0 377
+b2Q= 378
+YWxseQ== 379
+J3Q= 380
+dXN0 381
+IGFz 382
+IEM= 383
+Y2U= 384
+IG1l 385
+0LA= 386
+0LU= 387
+aWw= 388
+IEg= 389
+IHdhcw== 390
+dGVy 391
+dGg= 392
+IGNhbg== 393
+YW50 394
+IGNvbQ== 395
+b3Vy 396
+aWdodA== 397
+IFk= 398
+YXRpb24= 399
+IEFuZA== 400
+b2w= 401
+IHNo 402
+0YI= 403
+b3A= 404
+c2U= 405
+IG5vdA== 406
+IFNv 407
+IG5l 408
+dW4= 409
+IGFi 410
+IGxpa2U= 411
+IGF0 412
+IEQ= 413
+aWU= 414
+IGhl 415
+IGNvbg== 416
+IGNo 417
+b3Jl 418
+IGFs 419
+IG9y 420
+IHF1 421
+IE8= 422
+b21l 423
+cmE= 424
+dWw= 425
+IE4= 426
+cHA= 427
+IHlvdXI= 428
+b3VsZA== 429
+IFA= 430
+IGZy 431
+Z2U= 432
+ZXJz 433
+J3Jl 434
+0Lg= 435
+IHRoZXk= 436
+IHdoYXQ= 437
+dXNl 438
+IGFsbA== 439
+IFRoZQ== 440
+IEw= 441
+ZXNz 442
+ZW0= 443
+IGtu 444
+IGp1c3Q= 445
+YXJ0 446
+IHBybw== 447
+dmVyeQ== 448
+dW0= 449
+IGxv 450
+IOw= 451
+IG15 452
+b2s= 453
+IGV4 454
+YWI= 455
+IHRoZXJl 456
+IGJ1dA== 457
+IGtub3c= 458
+IHN1 459
+IEc= 460
+0YE= 461
+IEU= 462
+IG1h 463
+0L7Q 464
+IGVu 465
+IGFib3V0 466
+IEl0 467
+aXN0 468
+IHdvcg== 469
+cmk= 470
+aW5k 471
+IG9uZQ== 472
+YXRl 473
+YW5k 474
+aW5r 475
+IGxl 476
+b3J0 477
+J20= 478
+IEY= 479
+aWNo 480
+0YA= 481
+aWRl 482
+IGdldA== 483
+IG91dA== 484
+Li4u 485
+IHdpbGw= 486
+44E= 487
+aXZl 488
+0L0= 489
+IGZyb20= 490
+YWlu 491
+IFdl 492
+IHVw 493
+cGU= 494
+cmVz 495
+Y2E= 496
+IFI= 497
+IGlm 498
+IHBs 499
+IGRvbg== 500
+YWNr 501
+IDE= 502
+ICI= 503
+IHRy 504
+IHVz 505
+IFdo 506
+aXR5 507
+IEo= 508
+IFlvdQ== 509
+IGhlcmU= 510
+aGVy 511
+IHNvbWU= 512
+b3Vn 513
+YWs= 514
+YXJk 515
+IGdvaW5n 516
+IHVu 517
+bWVudA== 518
+IHRoaW5r 519
+IHBl 520
+ZW5k 521
+ICg= 522
+Y2F1c2U= 523
+IHRpbQ== 524
+YXN0 525
+w6k= 526
+IG91cg== 527
+IHdhbnQ= 528
+YW1l 529
+aWVz 530
+IOs= 531
+dWQ= 532
+aW5l 533
+IHJlYWxseQ== 534
+IHRl 535
+IHNlZQ== 536
+Y2k= 537
+IGJ5 538
+c28= 539
+dXJl 540
+b3Nl 541
+IFs= 542
+YXJl 543
+IG1vcmU= 544
+YWg= 545
+b25l 546
+Y2s= 547
+b3BsZQ== 548
+0LDQ 549
+IHRoZW4= 550
+IHRoaW5n 551
+IHRoZW0= 552
+dmVu 553
+b3VuZA== 554
+b3N0 555
+b25n 556
+ZWN0 557
+IHJpZ2h0 558
+YWc= 559
+IGludA== 560
+IHBlb3BsZQ== 561
+IHdoZW4= 562
+b3Vz 563
+cGw= 564
+IHRpbWU= 565
+IGlt 566
+IHdobw== 567
+IDI= 568
+YXA= 569
+IGJlY2F1c2U= 570
+aGluZw== 571
+IG5v 572
+aWNl 573
+IGxvb2s= 574
+IGhhcw== 575
+IHdvdWxk 576
+IGhvdw== 577
+YWN0 578
+IGZl 579
+bnQ= 580
+b3VnaA== 581
+IHBy 582
+IEJ1dA== 583
+IHNheQ== 584
+0YM= 585
+IG5vdw== 586
+IG1hbg== 587
+IHZlcnk= 588
+IHdvcms= 589
+aXo= 590
+IEs= 591
+aXY= 592
+aXR0 593
+IGFy 594
+ZXA= 595
+IGNs 596
+IHdoaWNo 597
+IGNv 598
+YW5z 599
+J3Zl 600
+IHNh 601
+ZmY= 602
+J2xs 603
+IGFueQ== 604
+IGFjdA== 605
+IHll 606
+YmVy 607
+YWNo 608
+YWdl 609
+cGVy 610
+IGFsc28= 611
+ZmVy 612
+IHRoZXNl 613
+IGFk 614
+0LXQ 615
+dGhlcg== 616
+YWNl 617
+aWNr 618
+YWtl 619
+cmVhdA== 620
+aXJl 621
+dWU= 622
+IGFn 623
+IFU= 624
+dWNo 625
+aW9ucw== 626
+cnk= 627
+MDA= 628
+bmE= 629
+IGRpZA== 630
+IHF1ZQ== 631
+IGhhZA== 632
+IGV2ZXJ5 633
+IEhl 634
+IGxh 635
+IHdheQ== 636
+IHNw 637
+Ymxl 638
+IFRoaXM= 639
+YXNz 640
+IHRoZWly 641
+aXRl 642
+IG5lZWQ= 643
+IHBhcnQ= 644
+IHdlcmU= 645
+IGJhY2s= 646
+aXA= 647
+b3du 648
+b21ldA== 649
+YmU= 650
+YXNl 651
+IG1ha2U= 652
+aXJzdA== 653
+aWE= 654
+ZW5jZQ== 655
+YW5n 656
+YW5r 657
+IGdvdA== 658
+IHByZQ== 659
+IGNvbnQ= 660
+IG90aGVy 661
+cHQ= 662
+IFRoYXQ= 663
+b2c= 664
+IGdvb2Q= 665
+IGludG8= 666
+YWxr 667
+IGJlZW4= 668
+IGFt 669
+IG92ZXI= 670
+dWFsbHk= 671
+IOI= 672
+7J0= 673
+IHVuZA== 674
+aGU= 675
+d2F5 676
+IGdy 677
+0Yw= 678
+IGRpZg== 679
+IHBlcg== 680
+0Y8= 681
+IElu 682
+IHR3 683
+b25k 684
+YXJz 685
+aW50 686
+b3Jt 687
+IGxvdA== 688
+IHdoZXJl 689
+IMM= 690
+IFY= 691
+IHNvbWV0 692
+0Ls= 693
+ZW5z 694
+IGd1 695
+IGFj 696
+dWc= 697
+0Ys= 698
+xLE= 699
+IGZpcnN0 700
+cmVl 701
+IGhpcw== 702
+aXR0bGU= 703
+IGltcA== 704
+IG1v 705
+YXY= 706
+IGxpdHRsZQ== 707
+IFdoYXQ= 708
+IG11Y2g= 709
+IHo= 710
+IOo= 711
+YWJsZQ== 712
+INC/ 713
+IHBv 714
+IGNvbXA= 715
+bmU= 716
+IGRpcw== 717
+IGxldA== 718
+YW5jZQ== 719
+IGhlcg== 720
+IHRoaW5ncw== 721
+IHN0YXJ0 722
+dWx0 723
+IGFwcA== 724
+IHJlcw== 725
+IGZv 726
+IGNvdWxk 727
+IGludGVy 728
+IHRob3Nl 729
+IGRlcw== 730
+IHdlbGw= 731
+IHR3bw== 732
+IGtpbmQ= 733
+eHQ= 734
+cmVzcw== 735
+ZWx5 736
+w6Q= 737
+IGJy 738
+IHRocg== 739
+INCy 740
+IGk= 741
+aXNo 742
+IGRpZmZlcg== 743
+IHJv 744
+IFN0 745
+IHNvbWV0aGluZw== 746
+IHRha2U= 747
+IGJv 748
+eXM= 749
+IHNoZQ== 750
+IHRhbGs= 751
+bG8= 752
+0Yc= 753
+IGV2ZW4= 754
+0Lo= 755
+44A= 756
+INC9 757
+IGJ1 758
+IElm 759
+IGRvd24= 760
+IENo 761
+YWRl 762
+YXRpb25z 763
+IHVzZQ== 764
+b3Jk 765
+IG9mZg== 766
+IGFjdHVhbGx5 767
+IHNwZQ== 768
+ZHU= 769
+YXRlZA== 770
+YXRlcg== 771
+b3Nz 772
+bmluZw== 773
+w7w= 774
+IGRvZXM= 775
+INGB 776
+IG5ldw== 777
+IGJldA== 778
+dmVs 779
+Y2Vzcw== 780
+cGxl 781
+IGhhcHA= 782
+dGluZw== 783
+b25uYQ== 784
+IGVz 785
+IGRheQ== 786
+IG9ubHk= 787
+aWdu 788
+a2F5 789
+c2Vs 790
+ZW50cw== 791
+b3VudA== 792
+aWxk 793
+aWxl 794
+IHNj 795
+IGhpbQ== 796
+IGFnYWlu 797
+dmluZw== 798
+IGdvbm5h 799
+IGNvbW0= 800
+IGhlbA== 801
+b3RoZXI= 802
+IGtl 803
+aWNhbA== 804
+IDM= 805
+IGVs 806
+IHRocm91Z2g= 807
+IGNvbWU= 808
+YXJr 809
+ZGF5 810
+aWVy 811
+w7M= 812
+IHRoYW4= 813
+IFRoZXk= 814
+IG1heQ== 815
+IHNlcg== 816
+7ZU= 817
+IGNhbGw= 818
+IGRpZmZlcmVudA== 819
+IHNob3VsZA== 820
+IFRoZXJl 821
+YXJ5 822
+IE5vdw== 823
+44I= 824
+dGhpbmc= 825
+d2U= 826
+b3J5 827
+ZnRlcg== 828
+IHB1dA== 829
+b3Jz 830
+aWFs 831
+64s= 832
+IHVuZGVy 833
+IGluYw== 834
+IFll 835
+dWI= 836
+Zm9ybQ== 837
+IHZpZGU= 838
+4Lg= 839
+dmVycw== 840
+IGZlZWw= 841
+w6E= 842
+b2R5 843
+ZnQ= 844
+Zm9yZQ== 845
+IGVt 846
+Z2V0 847
+IHNhaWQ= 848
+aXRpb24= 849
+IHJlYw== 850
+aW91cw== 851
+YXRjaA== 852
+IHRyeQ== 853
+IGhlbHA= 854
+IHNob3c= 855
+0LQ= 856
+IGJpdA== 857
+dWxs 858
+0LI= 859
+0YLQvg== 860
+Z3I= 861
+IHBsYXk= 862
+aWZl 863
+YWls 864
+IFllYWg= 865
+IHF1ZXN0 866
+IG1hbnk= 867
+IHBlcnM= 868
+IGdyZWF0 869
+w60= 870
+IGVzdA== 871
+bmc= 872
+IOKZ 873
+dHk= 874
+bGE= 875
+IE9o 876
+INc= 877
+4K4= 878
+IEJl 879
+YWR5 880
+IG1vc3Q= 881
+Y3Rpb24= 882
+IE5v 883
+IGRvaW5n 884
+IGJlaW5n 885
+IHRvbw== 886
+Y2Vz 887
+IGJs 888
+LiI= 889
+IHJlbQ== 890
+aXNz 891
+b25z 892
+Pj4= 893
+cnU= 894
+d24= 895
+b250 896
+aWI= 897
+ZWxs 898
+IHNt 899
+b3Ro 900
+dWFs 901
+ID4+ 902
+IHBo 903
+bGVz 904
+b2M= 905
+ZnVs 906
+IHNlYw== 907
+aXNl 908
+IGFkZA== 909
+aWdo 910
+ZXJ0 911
+IHNhbWU= 912
+4oA= 913
+IG1lYW4= 914
+IGZpbmQ= 915
+ZWs= 916
+IGVuZA== 917
+LS0= 918
+0Lw= 919
+IHN0aWxs 920
+YXo= 921
+ICc= 922
+IG1pbg== 923
+IHllYXJz 924
+dXJu 925
+IGFyb3VuZA== 926
+c2VsZg== 927
+IHdy 928
+YnM= 929
+b3VnaHQ= 930
+IOKZqg== 931
+IGZs 932
+YW5nZQ== 933
+IGFmdGVy 934
+IHBvaW50 935
+bWVy 936
+dmVk 937
+IGxvbmc= 938
+b3k= 939
+5Lg= 940
+IGNy 941
+d2F5cw== 942
+IHN5 943
+IHRyYQ== 944
+IDIw 945
+YXZl 946
+IGNoZQ== 947
+IGVudA== 948
+IGJlZm9yZQ== 949
+cGg= 950
+IGF0dA== 951
+aWFu 952
+aWx5 953
+IHBlcnNvbg== 954
+IGJpZw== 955
+IHNjaA== 956
+IHJlYWw= 957
+IG5leHQ= 958
+IGxvdmU= 959
+IHZpZGVv 960
+IExldA== 961
+IGZpbg== 962
+IG1haw== 963
+aWJsZQ== 964
+IHRvZGF5 965
+ZXJt 966
+IEFs 967
+b3dlcg== 968
+YW5u 969
+aXg= 970
+IHBhcg== 971
+IHN0dWQ= 972
+w7Y= 973
+IGltcG9ydA== 974
+dGU= 975
+IGdpdmU= 976
+dmVz 977
+IGRpZQ== 978
+IGRlYw== 979
+IHRlbGw= 980
+INC6 981
+0YHRgg== 982
+IHdoeQ== 983
+aWNhbGx5 984
+aWN0 985
+cmVk 986
+IGJhcw== 987
+IHN1cmU= 988
+IGJlbA== 989
+YXRpbmc= 990
+IHRhaw== 991
+IHNldA== 992
+IGxpZmU= 993
+IGRpZG4= 994
+2Kc= 995
+b2I= 996
+dW5k 997
+YXRo 998
+IG9w 999
+INC+ 1000
+YWl0 1001
+IHdvcmxk 1002
+IHN1cHA= 1003
+aW8= 1004
+IGNvdXI= 1005
+INC4 1006
+d2FyZA== 1007
+0LXQvQ== 1008
+IGFsd2F5cw== 1009
+dXA= 1010
+IGhhbmQ= 1011
+IEhvdw== 1012
+Y2lhbA== 1013
+IGNvbnM= 1014
+INE= 1015
+IGluZA== 1016
+IDQ= 1017
+IEFz 1018
+IGZ1bg== 1019
+amVjdA== 1020
+IGltcG9ydGFudA== 1021
+IHN1cg== 1022
+ZXc= 1023
+YXRlcw== 1024
+IDU= 1025
+IGRp 1026
+IG1hZGU= 1027
+IGlucw== 1028
+IGFzaw== 1029
+IGV0 1030
+IG51bQ== 1031
+IGNhcg== 1032
+IE9rYXk= 1033
+IHNpbQ== 1034
+aWs= 1035
+IGxhc3Q= 1036
+IEdv 1037
+IG11cw== 1038
+IHJlbA== 1039
+dWxhcg== 1040
+tOw= 1041
+IFdlbGw= 1042
+cGVjdA== 1043
+IFRoYW5r 1044
+IHRocmVl 1045
+w6M= 1046
+44M= 1047
+IGludg== 1048
+IGdlbg== 1049
+bGlj 1050
+IGhhcHBlbg== 1051
+64o= 1052
+aWVu 1053
+ZXZlcg== 1054
+0L7Qsg== 1055
+IHN0cg== 1056
+IEFsbA== 1057
+IGluc3Q= 1058
+IOKA 1059
+IGRlZg== 1060
+IHNs 1061
+IG1pZ2h0 1062
+dW5n 1063
+IHllYXI= 1064
+IG93bg== 1065
+IGtlZXA= 1066
+Ym9keQ== 1067
+ZGVy 1068
+INGC 1069
+INC0 1070
+IGFub3RoZXI= 1071
+IG1vZA== 1072
+IGV2 1073
+IGd1eXM= 1074
+IGFibGU= 1075
+w6Nv 1076
+cXVl 1077
+aWRlbnQ= 1078
+IFllcw== 1079
+IGl0cw== 1080
+IHBsYWNl 1081
+IHByb2R1 1082
+YXJu 1083
+INC8 1084
+IHJlcA== 1085
+IGV4cGVy 1086
+IGZhbQ== 1087
+aXRpZXM= 1088
+aWZpYw== 1089
+IGhpZ2g= 1090
+aWVk 1091
+b29s 1092
+aWV3 1093
+0LXRgg== 1094
+cmVu 1095
+IGRvbmU= 1096
+IC4uLg== 1097
+64qU 1098
+c3RlbQ== 1099
+IFNl 1100
+IGJldHRlcg== 1101
+Y29tZQ== 1102
+IGRlbA== 1103
+IHR5 1104
+IHVt 1105
+IGhv 1106
+IEFu 1107
+IG1vbg== 1108
+aW5ncw== 1109
+IHNr 1110
+IG9i 1111
+Y29t 1112
+YmxlbQ== 1113
+b3Bl 1114
+c3RhbmQ= 1115
+J2Q= 1116
+bWVudHM= 1117
+IGVsZQ== 1118
+IElz 1119
+IGRh 1120
+IHJlZw== 1121
+bGVhc2U= 1122
+aWtl 1123
+YWxz 1124
+aXpl 1125
+6rA= 1126
+IGNhcmU= 1127
+IG5ldmVy 1128
+7J20 1129
+ZXNl 1130
+IG1ldA== 1131
+b2xvZw== 1132
+IFdoZW4= 1133
+dWNr 1134
+0LXRgA== 1135
+IMOp 1136
+IGRhdA== 1137
+w6c= 1138
+IGV4YW0= 1139
+aWxpdHk= 1140
+IGRldA== 1141
+Y3Jp 1142
+IHVzZWQ= 1143
+IERv 1144
+IHRyYW5z 1145
+ZWc= 1146
+dGVu 1147
+0Y4= 1148
+Y3Vz 1149
+IHNlY29uZA== 1150
+IGJlc3Q= 1151
+IGhhcmQ= 1152
+IGlkZQ== 1153
+IHByb2JsZW0= 1154
+6rM= 1155
+IFVu 1156
+0YU= 1157
+IM4= 1158
+IHdhdGNo 1159
+IFNo 1160
+YXR0ZXI= 1161
+IHByZXQ= 1162
+IGRlcg== 1163
+IGNvdXJzZQ== 1164
+xZ8= 1165
+YXRpdmU= 1166
+aWNz 1167
+IHF1ZXN0aW9u 1168
+dXRl 1169
+7Jc= 1170
+IEZvcg== 1171
+YXRoZXI= 1172
+IGNvbA== 1173
+aWVuZA== 1174
+IO0= 1175
+IFo= 1176
+IGRvZXNu 1177
+YXJjaA== 1178
+IGludGVyZXN0 1179
+IHBvbA== 1180
+IGNvcg== 1181
+aWVuY2U= 1182
+IHByZXM= 1183
+IGVhY2g= 1184
+IHN5c3RlbQ== 1185
+IGZhY3Q= 1186
+aWVs 1187
+YWJseQ== 1188
+IGVy 1189
+IHJ1bg== 1190
+IOyd 1191
+IHRvcA== 1192
+bmVy 1193
+IHRob3VnaHQ= 1194
+IGVhcw== 1195
+aWVudA== 1196
+IGNyZQ== 1197
+0Yg= 1198
+IGNvbW11bg== 1199
+eWU= 1200
+cmVhZHk= 1201
+bGxvdw== 1202
+IGV2ZXJ5dGhpbmc= 1203
+b21t 1204
+IG1lZA== 1205
+mpQ= 1206
+IGNvdW50 1207
+aXRz 1208
+IGNvbXBs 1209
+aGlw 1210
+2YQ= 1211
+b29r 1212
+IHRvZ2V0 1213
+IHRvZ2V0aGVy 1214
+YW1w 1215
+IGdhbWU= 1216
+IGFscmVhZHk= 1217
+0LDQuw== 1218
+IGNhbGxlZA== 1219
+YWxl 1220
+xYI= 1221
+IE15 1222
+IHVuZGVyc3RhbmQ= 1223
+IGRy 1224
+IG1vbQ== 1225
+aXRlZA== 1226
+0L7Quw== 1227
+IHVzaW5n 1228
+enk= 1229
+IG51bWJlcg== 1230
+44CB 1231
+Y2Vk 1232
+IGNsZQ== 1233
+0L3Qvg== 1234
+64uk 1235
+aW5jZQ== 1236
+IGxvb2tpbmc= 1237
+IHByZXR0eQ== 1238
+IHByb2I= 1239
+IFNoZQ== 1240
+IHZl 1241
+IGdldHRpbmc= 1242
+IHdlZWs= 1243
+IGVmZg== 1244
+dWZm 1245
+YWly 1246
+dWVz 1247
+ZXJu 1248
+IFE= 1249
+b3Vw 1250
+ZW50aW9u 1251
+IHNpZGU= 1252
+0L7QvA== 1253
+IGZvcm0= 1254
+IGJ1cw== 1255
+IGFzcw== 1256
+IGVk 1257
+YXNvbg== 1258
+d2Vlbg== 1259
+4oCm 1260
+IHR1cm4= 1261
+IGN1cg== 1262
+IGNvbGw= 1263
+IGRpcmU= 1264
+IEdvZA== 1265
+IDEw 1266
+IGVxdQ== 1267
+INCx 1268
+IG9wZW4= 1269
+IHN1Y2g= 1270
+aXJk 1271
+0LDQug== 1272
+IGVhcg== 1273
+xJk= 1274
+Z2Fu 1275
+IHBhcnRpYw== 1276
+IGZyaWVuZA== 1277
+IGV4cA== 1278
+IGV4dA== 1279
+IGhvbWU= 1280
+IHdhdGVy 1281
+IE9u 1282
+0YLRjA== 1283
+b3Jr 1284
+INC/0YA= 1285
+IG1vdmU= 1286
+bmVzcw== 1287
+ZW5zZQ== 1288
+aG8= 1289
+IGNoYXI= 1290
+Y28= 1291
+aW5z 1292
+IGJvdGg= 1293
+IDE5 1294
+IGdyYQ== 1295
+IGJldHdlZW4= 1296
+4bs= 1297
+IOyV 1298
+YXNo 1299
+IFJl 1300
+YWk= 1301
+YWx0aA== 1302
+dXJlcw== 1303
+ZW1iZXI= 1304
+IGF2 1305
+IHZlcg== 1306
+w6o= 1307
+b25leQ== 1308
+IHRoYW5r 1309
+IG1heWJl 1310
+dWM= 1311
+aW1l 1312
+6rOg 1313
+IGF3YXk= 1314
+IG5hbWU= 1315
+b3VzZQ== 1316
+IGFjYw== 1317
+IG11c2lj 1318
+IGNoYW5nZQ== 1319
+IHBhc3M= 1320
+Z2Vy 1321
+IGJ1aWxk 1322
+IHZhbA== 1323
+aW5lc3M= 1324
+YW55 1325
+IGZldw== 1326
+tOs= 1327
+dGE= 1328
+IGxpc3Q= 1329
+w6U= 1330
+IG9sZA== 1331
+IOye 1332
+IHNvcnQ= 1333
+IG1lbQ== 1334
+IGNh 1335
+Y2VwdA== 1336
+IGdlbmVy 1337
+IHllYWg= 1338
+IHdoaWxl 1339
+IGFueXRoaW5n 1340
+cmlj 1341
+Z3JhbQ== 1342
+IGVpbg== 1343
+Y3k= 1344
+dXJpbmc= 1345
+IERl 1346
+IHBvd2Vy 1347
+IGNvbWluZw== 1348
+IHdvcmQ= 1349
+IC0t 1350
+IGJlbGll 1351
+IGZvdW5k 1352
+dG8= 1353
+0L8= 1354
+IG1lYW5z 1355
+IGluZm9ybQ== 1356
+INg= 1357
+INGH 1358
+IHNtYWxs 1359
+MDAw 1360
+IGNhbWU= 1361
+IO2V 1362
+d2g= 1363
+IHdvcmtpbmc= 1364
+IGV4YW1wbGU= 1365
+IHBvcw== 1366
+IGRlcA== 1367
+6rI= 1368
+5Lo= 1369
+b3Rl 1370
+IGRlbQ== 1371
+7Kc= 1372
+dHM= 1373
+IHZhcg== 1374
+YXV0 1375
+IHRyaQ== 1376
+Y2hu 1377
+IGhlYWQ= 1378
+IHdob2xl 1379
+15k= 1380
+emU= 1381
+IHRyeWluZw== 1382
+IHRlbQ== 1383
+IGNvdQ== 1384
+ZXRz 1385
+IDY= 1386
+IGZpbA== 1387
+dmVsb3A= 1388
+IGNhc2U= 1389
+4K8= 1390
+IHByb2JhYmx5 1391
+IG9rYXk= 1392
+IHBsYW4= 1393
+IHNpdA== 1394
+IHNjaG9vbA== 1395
+IFRoZW4= 1396
+uOs= 1397
+bWU= 1398
+IHByb2Nlc3M= 1399
+IGZhcg== 1400
+IHJlYWQ= 1401
+IHBvc3M= 1402
+IGJyZQ== 1403
+IHNvbA== 1404
+aWNodA== 1405
+IHN1cHBvcnQ= 1406
+IFRv 1407
+ZXJ0YWlu 1408
+IHN0YXJ0ZWQ= 1409
+IGNhcA== 1410
+IGxlZnQ= 1411
+IGRhdGE= 1412
+IHRpbWVz 1413
+0LXQuw== 1414
+IHdhbnRlZA== 1415
+0LDQvQ== 1416
+IHRhbGtpbmc= 1417
+IGlzdA== 1418
+IGhhdmluZw== 1419
+dW1w 1420
+IGNvbnRpbg== 1421
+IHN1Yg== 1422
+INC3 1423
+cHI= 1424
+64uI 1425
+aW5h 1426
+xbw= 1427
+IGNyZWF0 1428
+b2Rl 1429
+15U= 1430
+5pg= 1431
+ISE= 1432
+IHRlcm0= 1433
+aXNt 1434
+0L7QtA== 1435
+IEJlY2F1c2U= 1436
+IHdlbnQ= 1437
+aWRlcg== 1438
+IHByb3Y= 1439
+IGNoaWxk 1440
+IGRlbg== 1441
+IGxpZ2h0 1442
+YnI= 1443
+s9C+ 1444
+b2g= 1445
+IGJvb2s= 1446
+INk= 1447
+dXRpb24= 1448
+IEp1c3Q= 1449
+ZW5l 1450
+IGZvdXI= 1451
+IHZpcw== 1452
+6rCA 1453
+IGhvcGU= 1454
+IG1ha2luZw== 1455
+IExl 1456
+7JU= 1457
+IG9wcA== 1458
+YXU= 1459
+IG1vbmV5 1460
+IHByb2dyYW0= 1461
+w6g= 1462
+IHN0YW5k 1463
+SU4= 1464
+IHNpZ24= 1465
+IGxlYXJu 1466
+w6A= 1467
+IERvbg== 1468
+IHRlYW0= 1469
+INC90LA= 1470
+bHVk 1471
+IHJlc3Q= 1472
+aWNlcw== 1473
+5pw= 1474
+INGA 1475
+IGF1dA== 1476
+IGxlYWQ= 1477
+YXRpb25hbA== 1478
+ZGU= 1479
+Z3k= 1480
+IG5pY2U= 1481
+IGRhcw== 1482
+IGRpc3Q= 1483
+IGh1bQ== 1484
+IE9uZQ== 1485
+5og= 1486
+IGNvbWVz 1487
+IGpv 1488
+IGNlbnQ= 1489
+IGV4cGw= 1490
+IG1hcms= 1491
+cmVlbg== 1492
+bGVk 1493
+Z2lu 1494
+7JqU 1495
+IGxldmVs 1496
+IGNvbmY= 1497
+dXNo 1498
+IGRldmVsb3A= 1499
+IHRlc3Q= 1500
+ZW5n 1501
+dmlvdXM= 1502
+YXR1cmU= 1503
+0LXQvA== 1504
+cmV0 1505
+IGpl 1506
+IHN0dWZm 1507
+IGNsYXNz 1508
+b3dz 1509
+IOq3 1510
+IHNp 1511
+IGxlcw== 1512
+cm9w 1513
+55o= 1514
+IHBvcg== 1515
+IHdhcg== 1516
+7JeQ 1517
+IGV2ZXJ5b25l 1518
+IGdl 1519
+IGNoZWNr 1520
+b3R0 1521
+IHNpbmc= 1522
+IGFydA== 1523
+IGZvbGxvdw== 1524
+IDIwMQ== 1525
+IEZy 1526
+YWlz 1527
+7JY= 1528
+zrE= 1529
+5bA= 1530
+IMOg 1531
+aW1lcw== 1532
+IHJldA== 1533
+IGNoYW5n 1534
+IHB1Yg== 1535
+IGluZg== 1536
+IHRlY2hu 1537
+YWRh 1538
+aXZlcw== 1539
+IGJlaA== 1540
+IGxvb2tz 1541
+44CC 1542
+0Lc= 1543
+IFdoeQ== 1544
+IGVub3VnaA== 1545
+IGJyYQ== 1546
+aXRjaA== 1547
+5Ls= 1548
+IGFkdg== 1549
+0LE= 1550
+IHdpdGhvdXQ= 1551
+d2Vy 1552
+bWVyaWM= 1553
+ZGVu 1554
+IGNvbXBsZXQ= 1555
+IGlkZWE= 1556
+dGVycw== 1557
+b2Nr 1558
+IGRlZmlu 1559
+IGV2ZXI= 1560
+IGds 1561
+IG9uY2U= 1562
+IGJyaW5n 1563
+IHNheWluZw== 1564
+IGFucw== 1565
+IGhlYXI= 1566
+bmVjdA== 1567
+IGxlc3M= 1568
+Z28= 1569
+cmVhbQ== 1570
+YWRv 1571
+7J4= 1572
+IG1pbmQ= 1573
+ZW50ZQ== 1574
+IGZ1bGw= 1575
+IGJhZA== 1576
+IHdvbQ== 1577
+IHNvbWVvbmU= 1578
+IGR1 1579
+IHdvbg== 1580
+IGNvbnRybw== 1581
+b3J0dW4= 1582
+IGhlYWx0aA== 1583
+IGNobw== 1584
+IEFy 1585
+IGNvbmM= 1586
+IGluZm9ybWF0aW9u 1587
+IHN0b3A= 1588
+YXR0 1589
+YXRlbHk= 1590
+5L0= 1591
+IGdyb3Vw 1592
+INGD 1593
+IHF1aXRl 1594
+IHJlc3A= 1595
+RVI= 1596
+dWdodA== 1597
+6rg= 1598
+bWFu 1599
+aXplZA== 1600
+IEJy 1601
+IHJlbWVtYmVy 1602
+IGZhbWlseQ== 1603
+IGJ1c2luZXNz 1604
+YXc= 1605
+IHNwZWM= 1606
+IGF1 1607
+IE9y 1608
+xIU= 1609
+IHNlZW4= 1610
+IGxhcg== 1611
+IDc= 1612
+Z2c= 1613
+YmVycw== 1614
+IGRyYQ== 1615
+IG1vbnRo 1616
+IHNheXM= 1617
+IGlzcw== 1618
+IGxpdmU= 1619
+IGxpbmU= 1620
+IG1vbWVudA== 1621
+IGV4Yw== 1622
+ZWxz 1623
+IHNvdW5k 1624
+IGNvb2w= 1625
+IGxvYw== 1626
+IGNlcnRhaW4= 1627
+IGRyaQ== 1628
+0L7Rgg== 1629
+YW1lcw== 1630
+IG11c3Q= 1631
+bnk= 1632
+0LjRgg== 1633
+IGtpZA== 1634
+IGluY2x1ZA== 1635
+7J2E 1636
+YXRvcg== 1637
+xJ8= 1638
+aGE= 1639
+YXJlZA== 1640
+IHNlZW0= 1641
+0Lk= 1642
+7IQ= 1643
+IGVsc2U= 1644
+IOyg 1645
+aXJs 1646
+IDg= 1647
+IHZv 1648
+IHF1ZXN0aW9ucw== 1649
+aW5lcw== 1650
+ZWU= 1651
+w7xy 1652
+IEFtZXJpYw== 1653
+IHN0b3J5 1654
+IHNlcnY= 1655
+dmVybg== 1656
+YWdlcw== 1657
+bGFuZA== 1658
+IOKAkw== 1659
+ZXJh 1660
+IENhbg== 1661
+IHBvcA== 1662
+ZXRoZXI= 1663
+IG5h 1664
+IG9yZGVy 1665
+IG1ha2Vz 1666
+IHNpbmNl 1667
+Y29u 1668
+Y3Rvcg== 1669
+IHRob3VnaA== 1670
+IHByb2R1Y3Q= 1671
+0LvQuA== 1672
+IGxlZw== 1673
+IG1lZXQ= 1674
+YWxm 1675
+0YHRjw== 1676
+dW5jaA== 1677
+aXRlcg== 1678
+b3Zl 1679
+15XX 1680
+aWV0 1681
+0LDQvA== 1682
+aXRhbA== 1683
+IHN1cGVy 1684
+bGluZw== 1685
+IHBheQ== 1686
+IHBhcmE= 1687
+IGpvYg== 1688
+IEhlcmU= 1689
+IHN3 1690
+a3M= 1691
+cHRpb24= 1692
+bWE= 1693
+IGJlbGlldmU= 1694
+rOs= 1695
+IHdhaXQ= 1696
+0L7QuQ== 1697
+IHVudA== 1698
+IHF1aWNr 1699
+aHI= 1700
+INGN 1701
+IFBybw== 1702
+IG1lbg== 1703
+4Lk= 1704
+IGRheXM= 1705
+IGdvZXM= 1706
+IHNwZWFr 1707
+IEF0 1708
+ZW1lbnQ= 1709
+IG1pc3M= 1710
+IGF3 1711
+IGRlc2lnbg== 1712
+IHByb2plY3Q= 1713
+0L7RgA== 1714
+aWo= 1715
+YW50cw== 1716
+YXRz 1717
+IENocg== 1718
+IDk= 1719
+IGN1dA== 1720
+IHJlcXU= 1721
+INC90LU= 1722
+IE5vdA== 1723
+YXN0ZXI= 1724
+IG1pbGw= 1725
+IHBhcnRpY3VsYXI= 1726
+IHBpZQ== 1727
+IHN0dWRlbnRz 1728
+IGZpdmU= 1729
+b3Vu 1730
+IE5l 1731
+IGdp 1732
+IHBhcw== 1733
+IGZyZWU= 1734
+IFNw 1735
+bGljaA== 1736
+IHByb2Y= 1737
+IGVuZw== 1738
+IHByb3Q= 1739
+IExpa2U= 1740
+b3NlZA== 1741
+IGNvbm5lY3Q= 1742
+YXBw 1743
+IOun 1744
+aXRpbmc= 1745
+IGJsbw== 1746
+IGxvcw== 1747
+aXN0cw== 1748
+IGV4cGVyaWVuY2U= 1749
+cmVudA== 1750
+IHN0YXk= 1751
+IGZvb2Q= 1752
+dG9u 1753
+cnVjdA== 1754
+IGhpc3Q= 1755
+dmlldw== 1756
+aW5pbmc= 1757
+bW9zdA== 1758
+aXZlcnM= 1759
+Ym8= 1760
+IFRy 1761
+Z2Vu 1762
+IHBsZWFzZQ== 1763
+IGNvbW11bml0eQ== 1764
+IGNl 1765
+QU4= 1766
+bm8= 1767
+IGJvZHk= 1768
+IGhvdXI= 1769
+IHZlcnM= 1770
+4bo= 1771
+Y2Vy 1772
+IOqw 1773
+IHJlYXNvbg== 1774
+IFJpZ2h0 1775
+IGxhdGVy 1776
+z4Q= 1777
+IGhvdXNl 1778
+IFg= 1779
+0L7QvQ== 1780
+IHN0YXRl 1781
+Zmlj 1782
+5aQ= 1783
+xZs= 1784
+aWVsZA== 1785
+IHByaQ== 1786
+IHBhc3Q= 1787
+IHdhbGs= 1788
+b2xvZ3k= 1789
+ZXJpbmc= 1790
+YW5uYQ== 1791
+IHRlcg== 1792
+IGhvbGQ= 1793
+IG9yZ2Fu 1794
+YmVu 1795
+zr8= 1796
+w7Nu 1797
+IGVmZmVjdA== 1798
+IHlvdXJzZWxm 1799
+IHBsdXM= 1800
+YWo= 1801
+YW5kbw== 1802
+dXJhbA== 1803
+IHJvb20= 1804
+bGVjdA== 1805
+6rKM 1806
+PyI= 1807
+c2lkZQ== 1808
+IGJlY29tZQ== 1809
+0YY= 1810
+IMI= 1811
+b29k 1812
+IGNvbnN0 1813
+IG5pZ2h0 1814
+dXRlcw== 1815
+0LY= 1816
+IGJyZWFr 1817
+IHBhaW4= 1818
+IHN0ZXA= 1819
+aXJlZA== 1820
+IG5vdGhpbmc= 1821
+IHVudGls 1822
+0ZY= 1823
+0LDQsg== 1824
+2Yo= 1825
+IGR1cmluZw== 1826
+7KeA 1827
+bGVzcw== 1828
+b2xs 1829
+0L3Riw== 1830
+zrk= 1831
+ZmVjdA== 1832
+aXZlcg== 1833
+j4Q= 1834
+aXRoZXI= 1835
+eWluZw== 1836
+IGJlZ2lu 1837
+15nX 1838
+aXZpZA== 1839
+IMOn 1840
+IHNhbA== 1841
+IHRh 1842
+IHBvdA== 1843
+ICQ= 1844
+IG1hcg== 1845
+IGNsZWFy 1846
+IGZhY2U= 1847
+IGdyb3c= 1848
+ICo= 1849
+IGluc2lkZQ== 1850
+IGZyaWVuZHM= 1851
+IGxlYXZl 1852
+ZW5u 1853
+IGVhc3k= 1854
+IGFyZWE= 1855
+YWxpdHk= 1856
+b3Vk 1857
+IGVhdA== 1858
+2YY= 1859
+IHB1cg== 1860
+b3Ju 1861
+IHNhdw== 1862
+IGFuc3dlcg== 1863
+IGZyb250 1864
+IGJlYXV0 1865
+vOs= 1866
+IG1hdHRlcg== 1867
+IHNvbg== 1868
+IE5ldw== 1869
+IHJlc3VsdA== 1870
+aWRlcw== 1871
+Y2hl 1872
+IGZ1dA== 1873
+cHM= 1874
+IGZvY3Vz 1875
+IGludGVyZXN0aW5n 1876
+5aU= 1877
+IGFw 1878
+Ii4= 1879
+IGNyZWF0ZQ== 1880
+0L7RgQ== 1881
+IHByZXNz 1882
+cm9zcw== 1883
+IHBpY2s= 1884
+bGluZQ== 1885
+IHRvb2s= 1886
+IE1heQ== 1887
+cm93 1888
+IGljaA== 1889
+mOs= 1890
+IHJlZg== 1891
+IG1vcg== 1892
+cmFjdA== 1893
+YXJlbnQ= 1894
+QVI= 1895
+IGV4YWN0 1896
+IHNwYWNl 1897
+d29yaw== 1898
+0L3QuA== 1899
+IGJpcg== 1900
+IGRldg== 1901
+0LM= 1902
+IHRvbGQ= 1903
+IHB1YmxpYw== 1904
+Y2lhbGx5 1905
+IHZpZXc= 1906
+IEhleQ== 1907
+bWVk 1908
+bGxv 1909
+Y2M= 1910
+IGZhYw== 1911
+IGNvdXBsZQ== 1912
+IGhlYXJ0 1913
+bGVy 1914
+IHJlYWR5 1915
+IGFsbW9zdA== 1916
+YXJpbmc= 1917
+IGhhbGY= 1918
+IE1l 1919
+YXZvcg== 1920
+aXF1ZQ== 1921
+IGNoYXJhYw== 1922
+IHByYWN0 1923
+T04= 1924
+YW5l 1925
+IGls 1926
+0L3QsA== 1927
+IHZp 1928
+bGlzaA== 1929
+aGVhZA== 1930
+IGxlYXN0 1931
+IGJhc2ljYWxseQ== 1932
+YXNlZA== 1933
+cmlnaHQ= 1934
+IHlldA== 1935
+IHRha2luZw== 1936
+IGNvdW50cnk= 1937
+IHdpbg== 1938
+IGlzbg== 1939
+IHBvc3NpYmxl 1940
+IGNhbQ== 1941
+IGluY3Jl 1942
+IHBhdA== 1943
+IHdhbm5h 1944
+IGNvbnNpZGVy 1945
+IGFicw== 1946
+IHdpdGhpbg== 1947
+IGh1bWFu 1948
+IHRoaW5raW5n 1949
+IG9o 1950
+oZw= 1951
+IHF1aQ== 1952
+YXNlcw== 1953
+IDA= 1954
+aXRlbHk= 1955
+IGtpbGw= 1956
+IG1pbA== 1957
+IGludmVzdA== 1958
+aXN0ZXI= 1959
+IHN1Yw== 1960
+aW9uYWw= 1961
+ZWxm 1962
+IHdoZXRoZXI= 1963
+IGNvbnRyb2w= 1964
+IGFnYWluc3Q= 1965
+b3Rz 1966
+64uI64uk 1967
+aW9y 1968
+IHByZXNlbnQ= 1969
+INin 1970
+IHdhdGNoaW5n 1971
+dWJl 1972
+ZXJ2 1973
+IG5pY2h0 1974
+IGdvdmVybg== 1975
+IFRoZXNl 1976
+IDo= 1977
+dWl0 1978
+dWdo 1979
+IHdvcmtz 1980
+b28= 1981
+IHdpcg== 1982
+IGFpcg== 1983
+IFRl 1984
+0LDQtw== 1985
+aXNpb24= 1986
+d2hlcmU= 1987
+IHRvdA== 1988
+am95 1989
+7Is= 1990
+IHZvbA== 1991
+INC1 1992
+IGNsb3Nl 1993
+IEFk 1994
+0Yk= 1995
+aW5lZA== 1996
+IHVuYQ== 1997
+IOq3uOs= 1998
+sOs= 1999
+b3JyeQ== 2000
+IGJybw== 2001
+IGZpbG0= 2002
+aWZ0 2003
+MjA= 2004
+IHR5cGU= 2005
+IGhhcHBlbmVk 2006
+IEFt 2007
+IGdpcmw= 2008
+IEFyZQ== 2009
+d2FyZHM= 2010
+IHBvdXI= 2011
+IGNvbG9y 2012
+ZWx0 2013
+0LDRgQ== 2014
+IHNlbnNl 2015
+bGV4 2016
+IFdpdGg= 2017
+dXNz 2018
+cmli 2019
+IHJlc2U= 2020
+IG5vcm0= 2021
+IGZ1dHVyZQ== 2022
+IGRlYWw= 2023
+ZW5kaW5n 2024
+ZXk= 2025
+IHg= 2026
+ZXJv 2027
+IENs 2028
+dWs= 2029
+IHdoYXRldmVy 2030
+c2VsdmVz 2031
+IHlvdW5n 2032
+7Io= 2033
+IE1hcg== 2034
+IENocmlzdA== 2035
+IGd1ZXNz 2036
+IHBlcmZvcm0= 2037
+IGVuZXI= 2038
+cm9u 2039
+IGhpdA== 2040
+IHdvbmQ= 2041
+IGRpcmVjdA== 2042
+IEV2ZXJ5 2043
+IG9mdGVu 2044
+IGZh 2045
+IGFsb25n 2046
+IGNsaWNr 2047
+IExvb2s= 2048
+IHNpdHU= 2049
+IGhhcHB5 2050
+ZWFk 2051
+IGFnbw== 2052
+IGVuYw== 2053
+IG15c2VsZg== 2054
+IGNvdmVy 2055
+0L7QsQ== 2056
+IG1pZA== 2057
+IGNvc3Q= 2058
+IHRlbg== 2059
+IFNjaA== 2060
+IGV4cGVjdA== 2061
+IHdhc24= 2062
+IHN0cm9uZw== 2063
+aWZ1bA== 2064
+IG9wcG9ydHVu 2065
+aW5hbA== 2066
+eWxl 2067
+IHNoYXJl 2068
+IHRydWU= 2069
+IGFwcHJv 2070
+IGNoYWxs 2071
+IG1pbnV0ZXM= 2072
+IGNoYW5u 2073
+IOuC 2074
+zrU= 2075
+bGk= 2076
+IG1lc3M= 2077
+b3JpZXM= 2078
+cGVjaWFsbHk= 2079
+IHdyb25n 2080
+IHllcw== 2081
+IOyX 2082
+aXJvbg== 2083
+IGFsbG93 2084
+IHN1YnM= 2085
+IGZvcmU= 2086
+IGZpZ2h0 2087
+IHNvY2lhbA== 2088
+IGNyYQ== 2089
+YW5h 2090
+IGFmZg== 2091
+IGVzcw== 2092
+IHdheXM= 2093
+IHNob3J0 2094
+IGZhbGw= 2095
+IGxhdw== 2096
+IFdobw== 2097
+IGVuam95 2098
+IGNhbA== 2099
+IGFjY2Vzcw== 2100
+ZmU= 2101
+IG5vbg== 2102
+IGFjcm9zcw== 2103
+ZXJ5 2104
+dmlvdXNseQ== 2105
+IEV4 2106
+aWRlZA== 2107
+IGxpbms= 2108
+IFBy 2109
+IHRlcm1z 2110
+YWNlcw== 2111
+IGxhbmQ= 2112
+YXppbmc= 2113
+IDE1 2114
+IG11bHQ= 2115
+IHNwZWNpYWw= 2116
+5YA= 2117
+aXZpbmc= 2118
+7J2A 2119
+IHR5cA== 2120
+IHN0ZQ== 2121
+IMQ= 2122
+IGZvcndhcmQ= 2123
+5Y8= 2124
+IGZyZQ== 2125
+IHJlc2VhcmNo 2126
+4K+N 2127
+0LDRgg== 2128
+IG1haW4= 2129
+IHJlY29yZA== 2130
+IGh1 2131
+IGRlZmluaXRlbHk= 2132
+IGVpdGhlcg== 2133
+IGxpc3Rlbg== 2134
+IGtleQ== 2135
+IG1hcmtldA== 2136
+INGH0YLQvg== 2137
+aXphdGlvbg== 2138
+IHZpZGVvcw== 2139
+IGd1eQ== 2140
+IGZpZw== 2141
+IHN0cmE= 2142
+IFBs 2143
+dWxseQ== 2144
+YW1vcw== 2145
+IG1lbnRpb24= 2146
+IHNvbmc= 2147
+IGludGVybg== 2148
+cmFs 2149
+dXJz 2150
+IGhvbg== 2151
+IHZhbHVl 2152
+IGJhcg== 2153
+Y2xl 2154
+0L7Qtg== 2155
+xIc= 2156
+nOs= 2157
+IHp1 2158
+0LjQvA== 2159
+IHNpbmdsZQ== 2160
+IGF1Y2g= 2161
+Y3Vzcw== 2162
+IGdldHM= 2163
+IHNvbWV0aW1lcw== 2164
+5b4= 2165
+YW1i 2166
+bW0= 2167
+Y2luZw== 2168
+IHBlcmZlY3Q= 2169
+IEJs 2170
+b3V0aA== 2171
+7KA= 2172
+IHNjaQ== 2173
+cGFy 2174
+IHJlZA== 2175
+IHBvc3Q= 2176
+IG1vdA== 2177
+IGVsZWN0 2178
+IEV1 2179
+aXRpdmU= 2180
+IFNvbWU= 2181
+IGRlc2NyaQ== 2182
+IGN1cnJlbnQ= 2183
+w6lz 2184
+IHRyZQ== 2185
+IEVu 2186
+IG1pdA== 2187
+RU4= 2188
+iOs= 2189
+aXVt 2190
+IGhlYXJk 2191
+IHNpbXBsZQ== 2192
+bGFy 2193
+IGV2ZXJ5Ym9keQ== 2194
+aWxhcg== 2195
+IG5lZWRz 2196
+IGRpZmZpYw== 2197
+IEdvb2Q= 2198
+dW1lbnQ= 2199
+Y2VudA== 2200
+IG9wZXI= 2201
+0LDRgtGM 2202
+ZXR5 2203
+IGJsYWNr 2204
+IGdpdmVu 2205
+b25lcw== 2206
+IHdlbA== 2207
+6YA= 2208
+IOyVhA== 2209
+IDMw 2210
+QVQ= 2211
+IHN0YXQ= 2212
+b3VjaA== 2213
+IE1y 2214
+0LDRgA== 2215
+IHNobw== 2216
+IGNvbmQ= 2217
+15Q= 2218
+bXk= 2219
+IGNoaWxkcmVu 2220
+IGV1 2221
+0LXQtA== 2222
+7JWE 2223
+dGVybg== 2224
+IHVo 2225
+IGhhcg== 2226
+IHByb20= 2227
+IHB1bGw= 2228
+cmV3 2229
+IGNvbXBhbnk= 2230
+IGJlYXV0aWZ1bA== 2231
+dXN0b20= 2232
+7ZWY 2233
+0LrQuA== 2234
+IHN0cmU= 2235
+IGFtYXppbmc= 2236
+cmllcw== 2237
+IHN1Y2Nlc3M= 2238
+IG1hY2g= 2239
+bm90 2240
+IGRpc2N1c3M= 2241
+IG5hdA== 2242
+pqw= 2243
+IHVuZQ== 2244
+IGRpZmZpY3VsdA== 2245
+IHJpcw== 2246
+zr0= 2247
+IGNhbXA= 2248
+IGJ1eQ== 2249
+IG1hZw== 2250
+cG8= 2251
+IFlvdXI= 2252
+IGJlaGluZA== 2253
+aWNh 2254
+xLFu 2255
+IE9L 2256
+IGxhbmc= 2257
+IHdvbWVu 2258
+IGVudg== 2259
+IHJlY2U= 2260
+IGNoYW5uZWw= 2261
+aWFsbHk= 2262
+dWxl 2263
+IDEy 2264
+dGhlcnM= 2265
+IGJvdHQ= 2266
+IHJlcG9ydA== 2267
+ZW50bHk= 2268
+ZnVsbHk= 2269
+VGhl 2270
+IHNlbnQ= 2271
+IGV2ZW50 2272
+IGVuZXJneQ== 2273
+bHQ= 2274
+IHdvcmRz 2275
+YXJy 2276
+ZGxl 2277
+IGFoZWFk 2278
+YXJkcw== 2279
+2LE= 2280
+IHRvb2w= 2281
+Y29ub20= 2282
+0LXRgQ== 2283
+IGV4YWN0bHk= 2284
+IGZhdm9y 2285
+IGxvdw== 2286
+IHByb3Blcg== 2287
+IOyeiA== 2288
+ICE= 2289
+IHJlbGF0aW9ucw== 2290
+IG1hcw== 2291
+IGtpZHM= 2292
+IGVudGlyZQ== 2293
+dWRl 2294
+2YU= 2295
+IFdoZXJl 2296
+IG9uZXM= 2297
+IGNpdHk= 2298
+b2x1dA== 2299
+IHNpeA== 2300
+YWJpbGl0eQ== 2301
+w7Zy 2302
+aWxp 2303
+IEVz 2304
+IGhhcHBlbnM= 2305
+YWlucw== 2306
+IG1vZGVs 2307
+IHBpY3Q= 2308
+IGVzcGVjaWFsbHk= 2309
+IDEwMA== 2310
+a3Q= 2311
+IHNvb24= 2312
+Ynk= 2313
+cm9kdQ== 2314
+IGFubg== 2315
+IHN1YnNjcmk= 2316
+IFF1 2317
+IGF2YWls 2318
+aW1lbnQ= 2319
+IHZvYw== 2320
+a2E= 2321
+IDIwMA== 2322
+YXBlcg== 2323
+IEluZA== 2324
+IOyn 2325
+aG9y 2326
+jbA= 2327
+am9y 2328
+0LjQuw== 2329
+IHNxdQ== 2330
+QVU= 2331
+YXJuaW5n 2332
+INCz 2333
+SVM= 2334
+INC7 2335
+0LXQuQ== 2336
+eWVz 2337
+5YU= 2338
+INCS 2339
+IG9yaWc= 2340
+0L7Qs9C+ 2341
+IGFza2Vk 2342
+aWx0 2343
+0L7Qsw== 2344
+IGNvbnRpbnVl 2345
+IOyY 2346
+cmFt 2347
+IG90aGVycw== 2348
+RVM= 2349
+b2hu 2350
+IGxheQ== 2351
+IGJhc2Vk 2352
+IHB1 2353
+IGFwcGU= 2354
+IGxpbQ== 2355
+IHByb3A= 2356
+gOs= 2357
+bWlu 2358
+IGhvdA== 2359
+IExh 2360
+IGZhc3Q= 2361
+IHByb3RlY3Q= 2362
+IGFtb3VudA== 2363
+IGFxdQ== 2364
+IGZ1bmQ= 2365
+IGN1c3RvbQ== 2366
+IGN1bHQ= 2367
+IGhhbmRz 2368
+IGhhdmVu 2369
+IGF1ZA== 2370
+IG91dHNpZGU= 2371
+IEFmdGVy 2372
+YXBz 2373
+IGFuaW0= 2374
+cGxveQ== 2375
+IGhhdA== 2376
+IEZpcnN0 2377
+IHRyZWF0 2378
+IGVw 2379
+IG1hdGVy 2380
+IGJ1aWxkaW5n 2381
+IOuw 2382
+5ZA= 2383
+7ISc 2384
+emE= 2385
+dWdodGVy 2386
+IFBl 2387
+bmV5 2388
+ZXRlcg== 2389
+YXRpYw== 2390
+IGVkdWM= 2391
+6riw 2392
+IG1vdg== 2393
+k6Q= 2394
+YW1h 2395
+cmF0aW9u 2396
+IHNu 2397
+2Yg= 2398
+IHN1bQ== 2399
+IHBob3Q= 2400
+INCd 2401
+IC4= 2402
+IGZpbmlzaA== 2403
+aXR0aW5n 2404
+5a4= 2405
+IGxhcmdl 2406
+IOyW 2407
+IHdoaXRl 2408
+YXJh 2409
+IG1haXM= 2410
+IEhp 2411
+IGRhbQ== 2412
+INin2YQ= 2413
+IGJveA== 2414
+IEhlbGxv 2415
+IHNsZQ== 2416
+IG9wdA== 2417
+cmllZA== 2418
+pbw= 2419
+IGFjdGl2 2420
+IG7Do28= 2421
+IENvbQ== 2422
+IHBsYXlpbmc= 2423
+VGg= 2424
+IGF2YWlsYWJsZQ== 2425
+IHBvcnQ= 2426
+5Yg= 2427
+IEFo 2428
+IGxhcw== 2429
+IGVhcmx5 2430
+IHdvbmRlcg== 2431
+sbA= 2432
+IDE4 2433
+Y3Vs 2434
+IGZ1bmN0aW9u 2435
+IG1vcm5pbmc= 2436
+bGxl 2437
+aWVudHM= 2438
+dXg= 2439
+IGNpcg== 2440
+aXRpb25z 2441
+IGRlZXA= 2442
+IHBvbGl0 2443
+eW9y 2444
+bXA= 2445
+YWtpbmc= 2446
+jOs= 2447
+IE1hbg== 2448
+IG1pbGxpb24= 2449
+IC8= 2450
+IGluZGl2aWQ= 2451
+IHBhbg== 2452
+IGdvdmVybm1lbnQ= 2453
+IHdyaXRl 2454
+IFRvZA== 2455
+YW1lbnQ= 2456
+IM8= 2457
+IHdpbmQ= 2458
+IEVuZw== 2459
+Y2hlbg== 2460
+V2g= 2461
+7Jw= 2462
+IGlkZW50 2463
+dmVudA== 2464
+dXJjaA== 2465
+IGh5 2466
+IHlh 2467
+IHRyYWQ= 2468
+IHJlbGF0aW9uc2hpcA== 2469
+w7o= 2470
+IGRvdQ== 2471
+T1I= 2472
+IHN3ZQ== 2473
+IG5lZw== 2474
+aW5hdGlvbg== 2475
+IHRleHQ= 2476
+aXBw 2477
+IGZpbmU= 2478
+w6Fz 2479
+IERy 2480
+IENvbWU= 2481
+IG1vbnRocw== 2482
+LCI= 2483
+0LXQvdC4 2484
+IGhvdXJz 2485
+IHBvZA== 2486
+aXJ0 2487
+IGludm9s 2488
+IGNvbGxlY3Q= 2489
+IGF1Zg== 2490
+IHBh 2491
+IGhpc3Rvcnk= 2492
+bWI= 2493
+aWZ5 2494
+ID8= 2495
+IGJlbG93 2496
+YXN1cmU= 2497
+YWJ5 2498
+IGxhbmd1 2499
+IGFudA== 2500
+IGNvbWI= 2501
+YXRv 2502
+IGV4aXN0 2503
+IOuL 2504
+IHRha2Vz 2505
+IGNoYXJhY3Rlcg== 2506
+YWZm 2507
+IGZpZWxk 2508
+IGVjb25vbQ== 2509
+aWVm 2510
+IHBpZWNl 2511
+5Zw= 2512
+IHJlYWNo 2513
+IOqy 2514
+b255 2515
+IG1hdGVyaWFs 2516
+IGRpZw== 2517
+IHBoeXM= 2518
+IGltcHJv 2519
+IHNpbWlsYXI= 2520
+SUM= 2521
+IG5ldA== 2522
+eW4= 2523
+IHBvc2l0aW9u 2524
+w58= 2525
+IGJlbmU= 2526
+cmVhZA== 2527
+IGxlYXJuaW5n 2528
+dW1l 2529
+IGNsZWFu 2530
+0YLQvtGA 2531
+IGNvb2s= 2532
+IHNlZW1z 2533
+IG9s 2534
+IFVT 2535
+IEplcw== 2536
+IOCu 2537
+ZW50aWFs 2538
+aXZlcnNpdHk= 2539
+YWN5 2540
+INGP 2541
+b2x1dGVseQ== 2542
+cmVjdA== 2543
+IFBsZWFzZQ== 2544
+IHJlcHJlcw== 2545
+IHRvdWNo 2546
+bWVu 2547
+INCw 2548
+acOzbg== 2549
+IFRoYW5rcw== 2550
+IGFuZw== 2551
+IG1ham9y 2552
+IGl0c2VsZg== 2553
+aWxscw== 2554
+Iiw= 2555
+aWFucw== 2556
+IHNjcmVlbg== 2557
+IGhvcg== 2558
+IGtub3du 2559
+IGVudmlyb24= 2560
+IGZpbmFs 2561
+IGZpZ3VyZQ== 2562
+IFR3 2563
+IGV5ZXM= 2564
+IGltYWc= 2565
+IHNlZWluZw== 2566
+IGhhaXI= 2567
+cmVt 2568
+IGFwcGxpYw== 2569
+ZW5kcw== 2570
+cHV0 2571
+IG5ld3M= 2572
+IGNvbXBsZXRlbHk= 2573
+dWdocw== 2574
+IGtuZXc= 2575
+aWZpZWQ= 2576
+IEpl 2577
+IERpZA== 2578
+IHNpdHVhdGlvbg== 2579
+IGZsbw== 2580
+bXM= 2581
+IHBob25l 2582
+IGJhbGw= 2583
+ZG8= 2584
+IHBhcmVudA== 2585
+IHNvcnJ5 2586
+dXJ5 2587
+0LjQvQ== 2588
+aXBz 2589
+0LDQtA== 2590
+IGluc3RlYWQ= 2591
+IGh1Z2U= 2592
+IHR1 2593
+IOOB 2594
+IEdy 2595
+IGRldGFpbA== 2596
+INCf 2597
+IGluZGl2aWR1YWw= 2598
+IGZpcmU= 2599
+IGNsb3M= 2600
+IHdlcg== 2601
+dW5l 2602
+IHJ1bm5pbmc= 2603
+IGNvbnZlcnM= 2604
+IHJlY29tbQ== 2605
+IGNvbW8= 2606
+IHNvbWVib2R5 2607
+IEpvaG4= 2608
+IOydtA== 2609
+IE91cg== 2610
+cGxlcw== 2611
+IFBo 2612
+IGFuYWw= 2613
+IDUw 2614
+IG9mZmVy 2615
+IDw= 2616
+aXRpb25hbA== 2617
+Z2VzdA== 2618
+IHZvdXM= 2619
+bGV0 2620
+aWN5 2621
+IGZlZWxpbmc= 2622
+TEU= 2623
+cm9z 2624
+IHRoaXJk 2625
+0L7Qug== 2626
+IHNlcmllcw== 2627
+IEFueQ== 2628
+aXNlZA== 2629
+b2xk 2630
+IGRyYXc= 2631
+IHNlcnZpY2U= 2632
+IGNhbm5vdA== 2633
+YmFs 2634
+IGxpdmluZw== 2635
+xLFt 2636
+IGRpZmZlcmVuY2U= 2637
+IG9wcG9ydHVuaXR5 2638
+IG5lYXI= 2639
+b3J0aA== 2640
+a2Vu 2641
+IGxvY2Fs 2642
+2Ko= 2643
+IENvbg== 2644
+IG9iamVjdA== 2645
+IGRhc3M= 2646
+kNc= 2647
+IHF1aWNrbHk= 2648
+cmFwaA== 2649
+IGlzc3Vlcw== 2650
+IEFtZXJpY2Fu 2651
+IHByZXA= 2652
+ZW5jZXM= 2653
+IHByb2Zlc3M= 2654
+bGxpbmc= 2655
+b2Y= 2656
+IGZvb3Q= 2657
+YnJl 2658
+IHVzdWFsbHk= 2659
+IGdlbmVyYWw= 2660
+ZGE= 2661
+YW5jZXM= 2662
+IGRlc3Q= 2663
+IG9jYw== 2664
+IG1lbWJlcnM= 2665
+IGRhbnM= 2666
+IGVxdWFs 2667
+enQ= 2668
+IGJlY29t 2669
+IG1vdmluZw== 2670
+IHNwZWNpZmlj 2671
+w61h 2672
+IGZ1cg== 2673
+IG5lY2Vzcw== 2674
+IGNvbW1vbg== 2675
+IGF0dGFjaw== 2676
+INGN0YLQvg== 2677
+IFRvZGF5 2678
+IHVucw== 2679
+IEd1 2680
+aW9k 2681
+IGFjY291bnQ= 2682
+IGdyYW5k 2683
+IHNlbGY= 2684
+IEVs 2685
+IHRhc3Q= 2686
+IGNvbnRlbnQ= 2687
+IGN1 2688
+hOs= 2689
+IE1heWJl 2690
+IEplc3Vz 2691
+b3Jlcw== 2692
+cG9ydA== 2693
+qbQ= 2694
+IGdpdmVz 2695
+IG5vcm1hbA== 2696
+0YDRgw== 2697
+IGltcGFjdA== 2698
+w6Ry 2699
+IGRpZXM= 2700
+IGxhYg== 2701
+c2g= 2702
+aW9z 2703
+IFByZXM= 2704
+IFVuZA== 2705
+IE9m 2706
+IGZpbmFsbHk= 2707
+IGRvbGw= 2708
+IHZvY8Oq 2709
+cGx5 2710
+IEFn 2711
+IHRha2Vu 2712
+IGdyb3VuZA== 2713
+Zm9ydA== 2714
+IGdhdmU= 2715
+IEluc3Q= 2716
+IGxvc3Q= 2717
+IHdvcmtlZA== 2718
+IGxpdGVy 2719
+IGlzc3Vl 2720
+IGluZHVzdA== 2721
+IHJldHVybg== 2722
+IGhhcHBlbmluZw== 2723
+IHdhbnRz 2724
+0LjQsg== 2725
+IHByb2JsZW1z 2726
+IENhcg== 2727
+nbw= 2728
+IEFsc28= 2729
+IHNpemU= 2730
+IG9idmlvdXNseQ== 2731
+IFN1 2732
+IFNj 2733
+IHJlY29tbWVuZA== 2734
+b3VyY2Vz 2735
+YXN0aWM= 2736
+Li4uLg== 2737
+IG1p 2738
+bGllcg== 2739
+IEV2ZW4= 2740
+Y2lh 2741
+IGh1cg== 2742
+dmE= 2743
+IG1hc3M= 2744
+IHdvdWxkbg== 2745
+dW50 2746
+Y2tz 2747
+IGZlbHQ= 2748
+b3Nw 2749
+bGlnaHQ= 2750
+0L7Qu9GM 2751
+bmll 2752
+IGJvdHRvbQ== 2753
+INCx0Ys= 2754
+b3JlZA== 2755
+aXNvbg== 2756
+IGdyYWQ= 2757
+IHVtYQ== 2758
+IHZh 2759
+IOyC 2760
+cmVzc2lvbg== 2761
+dWxhdGlvbg== 2762
+SUQ= 2763
+aWRlbmNl 2764
+IGJ1cg== 2765
+IGdvbmU= 2766
+bHU= 2767
+7Ja07A== 2768
+IHJlZHU= 2769
+IGph 2770
+7J2Y 2771
+aXRh 2772
+IHNvZnQ= 2773
+IMOnYQ== 2774
+aWNv 2775
+ZXJhbA== 2776
+w7E= 2777
+YWY= 2778
+IHBvaW50cw== 2779
+Z3U= 2780
+IGTDqQ== 2781
+YXB0 2782
+YXg= 2783
+IEFscmlnaHQ= 2784
+IGNhbWVyYQ== 2785
+IGFjaA== 2786
+INC/0L4= 2787
+IHNldmVy 2788
+NTA= 2789
+IHNpZQ== 2790
+z4E= 2791
+IG1hbA== 2792
+IGNvbXB1dA== 2793
+IG1pZGRsZQ== 2794
+IGNvdWxkbg== 2795
+bWluZw== 2796
+IOyL 2797
+IEhpcw== 2798
+IGdhbWVz 2799
+IGludHJvZHU= 2800
+IGNlbGw= 2801
+cG9y 2802
+IHNsZWVw 2803
+IOuz 2804
+aWRpbmc= 2805
+IG91 2806
+IGRlZw== 2807
+IGRyaW5r 2808
+IGVudmlyb25tZW50 2809
+IFVuaXRlZA== 2810
+IHRhbGtlZA== 2811
+IGNob29zZQ== 2812
+IGpvdXI= 2813
+ZWdl 2814
+IE1pbg== 2815
+IGludGU= 2816
+IHJhdGhlcg== 2817
+IG9mZmlj 2818
+0LrQsA== 2819
+YWNoaW5n 2820
+IG1lbnRpb25lZA== 2821
+IGZpbGw= 2822
+IHRyYWNr 2823
+IG5pZQ== 2824
+IHV0 2825
+INCy0Ys= 2826
+aWJpbGl0eQ== 2827
+IHZhYw== 2828
+IHJhZA== 2829
+IHBhY2s= 2830
+IHNlbmQ= 2831
+IERhcw== 2832
+IEFi 2833
+IGVuZ2luZQ== 2834
+IGNvbXBldA== 2835
+w7Q= 2836
+INCy0YE= 2837
+IGRvb3I= 2838
+IGxvbmdlcg== 2839
+IGxhbmd1YWdl 2840
+IGV4dHJh 2841
+cGxheQ== 2842
+IHdlYnM= 2843
+dW1i 2844
+cm9vbQ== 2845
+55w= 2846
+IGJlZ2lubmluZw== 2847
+IHJlZmVy 2848
+QU0= 2849
+bmVu 2850
+aWdoZXI= 2851
+ZmFjZQ== 2852
+ZXJj 2853
+IGZvcmdldA== 2854
+IGNvbW1lbnQ= 2855
+0LXQug== 2856
+0LvRjw== 2857
+cm9y 2858
+xbxl 2859
+IEdl 2860
+IGRhcms= 2861
+IGFueW9uZQ== 2862
+YW50ZQ== 2863
+Z2Vz 2864
+7Iq1 2865
+0ZE= 2866
+YmVk 2867
+amU= 2868
+cnVjdHVyZQ== 2869
+IHByaW0= 2870
+aWRh 2871
+6KY= 2872
+IG1peA== 2873
+IHN0YXJ0aW5n 2874
+IOydtOs= 2875
+IHByb3ZpZGU= 2876
+YWN0aW9u 2877
+IG1vdGhlcg== 2878
+IHBlcmlvZA== 2879
+IHN0aWNr 2880
+IFlvdVQ= 2881
+IHRlY2hub2xvZ3k= 2882
+6rk= 2883
+IGJlZA== 2884
+IGdpdmluZw== 2885
+IGV4cGxhaW4= 2886
+emVu 2887
+aW1hdGU= 2888
+IHJlcHJlc2VudA== 2889
+bG9hZA== 2890
+IEhvd2V2ZXI= 2891
+IGxpdmVz 2892
+dXRo 2893
+aXJpdA== 2894
+b2du 2895
+IGxpaw== 2896
+IHJlc3BvbnM= 2897
+IHByaXY= 2898
+IHRvbQ== 2899
+w6fDo28= 2900
+aWFt 2901
+IGV4Y2l0ZWQ= 2902
+IGNhcmQ= 2903
+Z3JvdW5k 2904
+INeU 2905
+IHNlbnM= 2906
+IHRlYWNo 2907
+aWRv 2908
+aG9k 2909
+IGVwaXM= 2910
+IHdlbGNvbWU= 2911
+IHdhbGw= 2912
+5Lk= 2913
+IGNoYW5jZQ== 2914
+aGVu 2915
+INCh 2916
+IMSR 2917
+IHNpbXBseQ== 2918
+INGC0LDQug== 2919
+cmluZw== 2920
+amE= 2921
+Ym9vaw== 2922
+IHNldmVyYWw= 2923
+c3Rl 2924
+IGNyZWF0ZWQ= 2925
+INC+0YI= 2926
+IHB1c2g= 2927
+PT0= 2928
+IGhpZ2hlcg== 2929
+dWY= 2930
+b3VyY2U= 2931
+b2tl 2932
+IG9ubGluZQ== 2933
+IHJlbGU= 2934
+IHRvbg== 2935
+ZW5zaXZl 2936
+IGZhdm9yaXRl 2937
+0YPQtA== 2938
+IGxvb2tlZA== 2939
+IHZvbg== 2940
+4oCU 2941
+IGbDvHI= 2942
+IGJ1dHRvbg== 2943
+IGJpbGw= 2944
+IGNoYW5nZXM= 2945
+ISI= 2946
+IHNsb3c= 2947
+YWJsZXM= 2948
+IGRlYXRo 2949
+YW5kcw== 2950
+YXRlZw== 2951
+IHRoZW1zZWx2ZXM= 2952
+IGNvcA== 2953
+IHBlcnNvbmFs 2954
+dWdoaW5n 2955
+IDEx 2956
+Z2Fy 2957
+YWRlcw== 2958
+IG5lZWRlZA== 2959
+IHN0dWR5 2960
+YWdlZA== 2961
+0YHRgtCy 2962
+aW5v 2963
+IGRpc2M= 2964
+a2k= 2965
+IGFkZHJlc3M= 2966
+16g= 2967
+aXR0ZW4= 2968
+ZXNvbWU= 2969
+INC2 2970
+pOs= 2971
+dXJh 2972
+IG11 2973
+IGNvbnRpbnU= 2974
+Zm9y 2975
+IG1hdGNo 2976
+IHN0cmFpZ2h0 2977
+kOs= 2978
+bmVycw== 2979
+IGRvZw== 2980
+IGRlYg== 2981
+IENP 2982
+IG9z 2983
+Z2Vk 2984
+Y2FtZQ== 2985
+IGNvcnJlY3Q= 2986
+ZXR0ZQ== 2987
+IFNlZQ== 2988
+IGluY2x1ZGluZw== 2989
+IEV1cm8= 2990
+ZXN0ZXI= 2991
+IGp1bXA= 2992
+IFdoaWNo 2993
+INC60LDQug== 2994
+c29u 2995
+eWE= 2996
+SU5H 2997
+IGVpbmU= 2998
+b3No 2999
+ZW5jeQ== 3000
+IG1lZGlh 3001
+IHN1YnNjcmliZQ== 3002
+6YI= 3003
+IHByaW4= 3004
+IGhhYg== 3005
+IFBlcg== 3006
+IFdhcw== 3007
+IHBhZ2U= 3008
+aXRvcg== 3009
+IHRvd2FyZHM= 3010
+IHRyaWVk 3011
+ZW5nZQ== 3012
+YXJ0bWVudA== 3013
+IHZhcmk= 3014
+IHBhcGVy 3015
+IHBpY3R1cmU= 3016
+IHZlcnNpb24= 3017
+IGJyb3VnaHQ= 3018
+d2FyZQ== 3019
+IFN0YXRlcw== 3020
+IHNpY2g= 3021
+bGVkZ2U= 3022
+IHBlcmNlbnQ= 3023
+IGdvZA== 3024
+ZWM= 3025
+IENvbW0= 3026
+IGRlY2lkZWQ= 3027
+IHNlbGVjdA== 3028
+7ZWc 3029
+KS4= 3030
+dXJpdHk= 3031
+IGZ1cnRoZXI= 3032
+IGNvbW1lbnRz 3033
+bGVtZW50 3034
+IGRyZWFt 3035
+IGNlbnRlcg== 3036
+bWk= 3037
+IGNhcw== 3038
+IHdvbWFu 3039
+IHJvYWQ= 3040
+IGZhaWw= 3041
+IGJlY2FtZQ== 3042
+bHVz 3043
+aWxpdGllcw== 3044
+IENv 3045
+IG1hbmFnZQ== 3046
+IHJlY29nbg== 3047
+IGFjdGlvbg== 3048
+IGJlbmVm 3049
+IGVhcmxpZXI= 3050
+15w= 3051
+IHNwZWVk 3052
+IG1lbnQ= 3053
+IHNvY2k= 3054
+IHNob290 3055
+dWk= 3056
+IMOk 3057
+IGFwcGx5 3058
+dm8= 3059
+eGlt 3060
+IGNhdXNl 3061
+IHN1cnBy 3062
+IGhhYmVu 3063
+REk= 3064
+IGZhdGhlcg== 3065
+IE5leHQ= 3066
+IFlvdVR1YmU= 3067
+IGNvZGU= 3068
+IHJvbGU= 3069
+Z3Jlc3M= 3070
+IGdyZWVu 3071
+ZXR0 3072
+IGJ1aWx0 3073
+IGZsb3c= 3074
+IGJhc2U= 3075
+IHRyYWluaW5n 3076
+IHJvdW5k 3077
+IFdpbGw= 3078
+IHBhdGg= 3079
+IFJv 3080
+IGludGVyZXN0ZWQ= 3081
+7Ja0 3082
+IHJlc3BlY3Q= 3083
+IGNoYW5nZWQ= 3084
+aXNzaW9u 3085
+IHN0dWRlbnQ= 3086
+b2dyYXBo 3087
+IGFwcHJvYWNo 3088
+IHNob3dz 3089
+IHRhcg== 3090
+IGNyaXQ= 3091
+IGdsbw== 3092
+7Iq164uI64uk 3093
+IGRlYWQ= 3094
+IFByZXNpZGVudA== 3095
+IHRob3Vz 3096
+IGJhbA== 3097
+c3Rlcg== 3098
+ZXg= 3099
+IGFic29sdXRlbHk= 3100
+IG1pYw== 3101
+IHByYWN0aWNl 3102
+IHF1YWxpdHk= 3103
+IGxvd2Vy 3104
+b2dsZQ== 3105
+IHNlcGFy 3106
+YmFsbA== 3107
+bWVkaQ== 3108
+IHJldmlldw== 3109
+IEFwcA== 3110
+IG9r 3111
+4oCL 3112
+IGV4cGVyaWVu 3113
+IGNvbmNlcm4= 3114
+ZW50aWFsbHk= 3115
+bW9yZQ== 3116
+IEpv 3117
+YXBhbg== 3118
+IEljaA== 3119
+aXN0aWM= 3120
+IGZhaXI= 3121
+IHdlYnNpdGU= 3122
+aXJlcw== 3123
+IEJ5 3124
+IHRyYXZlbA== 3125
+IHJpc2s= 3126
+IG1pcg== 3127
+IGJvYXJk 3128
+IHNlbg== 3129
+IHBhcmVudHM= 3130
+IFdvdw== 3131
+IGZlZWQ= 3132
+IHNhdmU= 3133
+IHNlcmlvdXM= 3134
+IGluaXQ= 3135
+RUw= 3136
+dW5kcmVk 3137
+QVM= 3138
+IHZhbg== 3139
+b3Jyb3c= 3140
+IHdvcnRo 3141
+IHNlYXJjaA== 3142
+IDE2 3143
+IHBhcnRz 3144
+0YHRgtGM 3145
+IGNvbXBhbg== 3146
+IG1vdmll 3147
+IG1ldGhvZA== 3148
+IGlsbA== 3149
+IHdpc2g= 3150
+ZHk= 3151
+IGl0ZW0= 3152
+IG1pbnVz 3153
+YW5nZXI= 3154
+IHZvaWNl 3155
+IHNraW4= 3156
+IGFyZWFz 3157
+IGVpZ2h0 3158
+IG9icw== 3159
+ICw= 3160
+0LDQuQ== 3161
+IG9pbA== 3162
+IGN5 3163
+IGJhYnk= 3164
+c3k= 3165
+IGVtcGxveQ== 3166
+IEtl 3167
+IHBsYWNlcw== 3168
+IGZpeA== 3169
+IGVzdMOh 3170
+aXZlZA== 3171
+IGxvdHM= 3172
+IHNlYXNvbg== 3173
+dW5r 3174
+YWx0 3175
+IHRhYmxl 3176
+INCi 3177
+w6I= 3178
+IGF0dGVudGlvbg== 3179
+IEhlcg== 3180
+IGFnZQ== 3181
+IHByYQ== 3182
+YmFjaw== 3183
+Y2ls 3184
+IG5ldHdvcms= 3185
+cml0 3186
+IGRvYw== 3187
+IGFyZW4= 3188
+aWdlbg== 3189
+IOuE 3190
+2K8= 3191
+ZW5kZXI= 3192
+IHRvdGFs 3193
+IHByaWNl 3194
+IGNyYXp5 3195
+7Jo= 3196
+aXF1 3197
+dGhvdWdo 3198
+WW91 3199
+2Yc= 3200
+z4U= 3201
+IHNhdA== 3202
+IGJp 3203
+IERpZQ== 3204
+IHNoYQ== 3205
+IHRoYW5rcw== 3206
+dWg= 3207
+IHN0YWdl 3208
+0LDQtg== 3209
+IEZs 3210
+IGxlYXY= 3211
+IGJveQ== 3212
+IGFm 3213
+w7Zu 3214
+IEdldA== 3215
+IGFjY2VwdA== 3216
+IGVudGVy 3217
+IHR1cg== 3218
+IHNpxJk= 3219
+IGhvbmVzdA== 3220
+44CM 3221
+IHNhbQ== 3222
+IHJlcGw= 3223
+Z2luZw== 3224
+IGRldmVsb3BtZW50 3225
+IEFjdA== 3226
+b3Jh 3227
+44CN 3228
+5L4= 3229
+IGtub3dz 3230
+IGltYWdl 3231
+IExvcmQ= 3232
+0LjRgtGM 3233
+IHdlZWtz 3234
+IHNleA== 3235
+lOs= 3236
+IGh1bmRyZWQ= 3237
+IHNvdW5kcw== 3238
+IGxlYXJuZWQ= 3239
+IGJ1ZA== 3240
+INGB0YI= 3241
+IGluY3JlZA== 3242
+4pk= 3243
+IG5vcw== 3244
+IGRyb3A= 3245
+IGJlbg== 3246
+INCY 3247
+IHNhZmU= 3248
+YXRh 3249
+IGZ1Y2s= 3250
+c29jaQ== 3251
+IGRhbg== 3252
+IGNyb3Nz 3253
+MTA= 3254
+bW8= 3255
+dmVydA== 3256
+IDE3 3257
+emll 3258
+5ZU= 3259
+IGRvbQ== 3260
+IEJv 3261
+IHNldHRpbmc= 3262
+IGludm9sdmVk 3263
+YXJpbHk= 3264
+IHNpbmQ= 3265
+IHN1cw== 3266
+IHdvcnJ5 3267
+ZXRo 3268
+6rmM 3269
+IHN1bg== 3270
+IGhpZXI= 3271
+IGNlcnRhaW5seQ== 3272
+b3Vs 3273
+b3J0cw== 3274
+IEVy 3275
+IFVt 3276
+IGNhdXM= 3277
+IG5hdHVyYWw= 3278
+IMO8 3279
+IGNyeQ== 3280
+IFNlYw== 3281
+IHNvbQ== 3282
+5rI= 3283
+IGVkdWNhdGlvbg== 3284
+0LDQtdGC 3285
+IG11bHRpcA== 3286
+IGFsb25l 3287
+IGV5ZQ== 3288
+IHJhdGU= 3289
+IEV1cm9wZQ== 3290
+6L8= 3291
+bW9u 3292
+IGZpdA== 3293
+aXppbmc= 3294
+cHBlZA== 3295
+IHByZXNzdXJl 3296
+dGhl 3297
+0LjRgQ== 3298
+aXRlcw== 3299
+IEFm 3300
+cmVjaQ== 3301
+YXR0bGU= 3302
+IHNlcnZpY2Vz 3303
+IEdvb2dsZQ== 3304
+6YE= 3305
+IGNhc2Vz 3306
+IGRyaXZl 3307
+IGNoYWxsZW5n 3308
+dXo= 3309
+IE1v 3310
+7Jy86w== 3311
+dmFs 3312
+IGZvbA== 3313
+IOyi 3314
+ZmZpYw== 3315
+IHJh 3316
+IHNpbg== 3317
+IGJsdWU= 3318
+IGFmZmVjdA== 3319
+IG1pcw== 3320
+IHNob3Q= 3321
+INC+0LE= 3322
+YXNpbmc= 3323
+IHNpZ25pZmlj 3324
+IENoZQ== 3325
+IOqz 3326
+IHBvc2l0aXZl 3327
+7KM= 3328
+IHdpZQ== 3329
+IDQw 3330
+b3JkaW5n 3331
+IEZyb20= 3332
+6rU= 3333
+IGJyYW5k 3334
+IHRydXN0 3335
+IHBsZQ== 3336
+IGNvbW11bmlj 3337
+IHdlaWdodA== 3338
+IGFza2luZw== 3339
+IHRheA== 3340
+IEphcGFu 3341
+IO2VmA== 3342
+b3Bz 3343
+z4I= 3344
+IHB1dHRpbmc= 3345
+IHJvbGw= 3346
+IEFtZXJpY2E= 3347
+cmVn 3348
+ntc= 3349
+YXR1cmVz 3350
+ZW5zaW9u 3351
+IFNvbWV0 3352
+IG9yaWdpbmFs 3353
+cGluZw== 3354
+IMWf 3355
+IHByb2R1Y3Rz 3356
+IGNvbnRhY3Q= 3357
+b2x1dGlvbg== 3358
+IGdvYWw= 3359
+IHBvdw== 3360
+IHBlcmZvcm1hbmNl 3361
+IGJsb29k 3362
+YXRvcnM= 3363
+IE1pY2g= 3364
+IHRlbXBlcg== 3365
+IERhbg== 3366
+IHN1Z2c= 3367
+0YLQuA== 3368
+IGltbQ== 3369
+IG9mZmljZQ== 3370
+IGFycmk= 3371
+IGNvbWZvcnQ= 3372
+INCU 3373
+IHN1Z2dlc3Q= 3374
+IHBsYXQ= 3375
+gpg= 3376
+MTk= 3377
+IG9t 3378
+IHNldmVu 3379
+IENlbnQ= 3380
+aWxsZQ== 3381
+IGNvbmNlcHQ= 3382
+IGJhZw== 3383
+w7xu 3384
+aXZlbHk= 3385
+IGRpdg== 3386
+bW9z 3387
+5ok= 3388
+IGZlZWxz 3389
+IGly 3390
+YWtlcw== 3391
+bGV5 3392
+IHBhcnRpY2lw 3393
+INCa 3394
+Zmw= 3395
+anVzdA== 3396
+IHNpbA== 3397
+IFBh 3398
+QUw= 3399
+IGdvdHRh 3400
+IGZhbg== 3401
+IGNoYWxsZW5nZQ== 3402
+IGNvbXBhbmllcw== 3403
+IFBlb3BsZQ== 3404
+PC8= 3405
+0L7Qtw== 3406
+IHBlbg== 3407
+aXNpbmc= 3408
+IGF1cw== 3409
+ZW1pYw== 3410
+YW1lbnRl 3411
+IG1lZXRpbmc= 3412
+IHZpc2l0 3413
+IHN1cHBvc2Vk 3414
+IE9uY2U= 3415
+0LTQsA== 3416
+b3JsZA== 3417
+MzA= 3418
+VVM= 3419
+IHZpb2w= 3420
+IG5vdGljZQ== 3421
+INCQ 3422
+aGFu 3423
+cGVk 3424
+7Jg= 3425
+aGg= 3426
+IHRyb3U= 3427
+IG1pbnV0ZQ== 3428
+IFBhcg== 3429
+cmF5 3430
+IHRpdA== 3431
+IHVwZA== 3432
+IGJsb2Nr 3433
+IGR1ZQ== 3434
+YXVy 3435
+IGZvcmNl 3436
+IGNvdW4= 3437
+IOKAlA== 3438
+IHR5cGVz 3439
+66c= 3440
+IGxhdGU= 3441
+IGltcHJvdmU= 3442
+IOyI 3443
+IGF2ZQ== 3444
+dWxlcw== 3445
+Y2w= 3446
+YW1lZA== 3447
+IGF3ZXNvbWU= 3448
+IE9r 3449
+IHZvdA== 3450
+IG1hY2hpbmU= 3451
+IGZvbGxvd2luZw== 3452
+IG1lYXN1cmU= 3453
+YWNpw7Nu 3454
+dWVs 3455
+Y2hhbg== 3456
+IGFiaWxpdHk= 3457
+IHRvdXQ= 3458
+IGlkZWFz 3459
+IGluY3JlYXNl 3460
+IGVucw== 3461
+INGF 3462
+IOuq 3463
+IGplc3Q= 3464
+INCc 3465
+IHRydXRo 3466
+aHk= 3467
+IHNwZW5k 3468
+IHNjaWVuY2U= 3469
+ZXRl 3470
+IDE0 3471
+IGVwaXNvZGU= 3472
+IGFsZw== 3473
+ZW5kZWQ= 3474
+YXJp 3475
+bGxh 3476
+IGZpc2g= 3477
+IHRocm93 3478
+bWl0 3479
+5bk= 3480
+IGNpcmM= 3481
+IENhbA== 3482
+IHRvdXI= 3483
+IGRpcmVjdGlvbg== 3484
+IG5vY2g= 3485
+0LXQsg== 3486
+w6lu 3487
+IGNvdW50cmllcw== 3488
+IGluZHVzdHJ5 3489
+aW55 3490
+aWNsZQ== 3491
+IGZlZXQ= 3492
+SXQ= 3493
+IGxlYWRlcnM= 3494
+ZXR6dA== 3495
+IHN0YWZm 3496
+55Q= 3497
+IHB1cnA= 3498
+aXRv 3499
+PyE= 3500
+IEph 3501
+IHN0b3Jl 3502
+ZXRpYw== 3503
+IENoaW5h 3504
+IOuQ 3505
+IFVuaXZlcnNpdHk= 3506
+ICM= 3507
+IGRlY2lzaW9u 3508
+IGFjaGll 3509
+IGFjdHVhbA== 3510
+dWx5 3511
+IHNlY3Rpb24= 3512
+IHJlc3VsdHM= 3513
+IHN0YXI= 3514
+IG1pc3Q= 3515
+aWJseQ== 3516
+IGRhZA== 3517
+IG51bWJlcnM= 3518
+b21i 3519
+6Ko= 3520
+IFNwZQ== 3521
+IG1lcg== 3522
+IDI1 3523
+IGF1dG9t 3524
+IGNvbGQ= 3525
+2Kg= 3526
+hJw= 3527
+YWdlcg== 3528
+IFRW 3529
+IFNpZQ== 3530
+IEhhdmU= 3531
+IMW8ZQ== 3532
+dWdn 3533
+YWluZWQ= 3534
+IHVwb24= 3535
+IGxvZw== 3536
+IGNvbXBsZXRl 3537
+IGJyYWlu 3538
+YWdpbmc= 3539
+IE11cw== 3540
+b3Zlcg== 3541
+IGVhc2llcg== 3542
+IGludGVncg== 3543
+IG3DoXM= 3544
+IHR1cm5lZA== 3545
+IHN0cmk= 3546
+aXZhbA== 3547
+IGhlYXY= 3548
+IFRI 3549
+IHdyaXRpbmc= 3550
+0YDQsA== 3551
+IGNsYQ== 3552
+ZGluZw== 3553
+IHRlbGxpbmc= 3554
+0LjQtA== 3555
+aWNhdGVk 3556
+YWNodA== 3557
+aGFwcw== 3558
+IFN0ZQ== 3559
+IHJlc291cmNlcw== 3560
+IGRhbm4= 3561
+IHBhcnR5 3562
+IM+E 3563
+IHNhZg== 3564
+aXNlcw== 3565
+dHJl 3566
+b2ludA== 3567
+IGtub3dsZWRnZQ== 3568
+IGFueW1vcmU= 3569
+IGZseQ== 3570
+IG1haW50 3571
+0LjQug== 3572
+5ZE= 3573
+IHNlbGw= 3574
+bGF1Z2hz 3575
+IFlvcms= 3576
+IGJpZW4= 3577
+IG9k 3578
+IGVhc2lseQ== 3579
+IHJhbmdl 3580
+IG9wdGlvbg== 3581
+2Lk= 3582
+IGFwcHJlY2k= 3583
+b2Ny 3584
+IGRldGVybQ== 3585
+0YQ= 3586
+IG1lYW5pbmc= 3587
+IHNpdGU= 3588
+IGRpc2Nv 3589
+dmVyYWdl 3590
+IGxvc2U= 3591
+IGluc3RhbGw= 3592
+IGVtb3Q= 3593
+YW50bHk= 3594
+w6R0 3595
+IHRhbWI= 3596
+IFdhcg== 3597
+IEhv 3598
+IEdlbg== 3599
+ZW15 3600
+0LXQtw== 3601
+IFBvbA== 3602
+IG1lc3NhZ2U= 3603
+IG5vdGU= 3604
+jIA= 3605
+IGhldA== 3606
+IGltbWVkaQ== 3607
+IGF2bw== 3608
+IGJvb2tz 3609
+IGJlY29tZXM= 3610
+cmVzaA== 3611
+w6hz 3612
+YXNvbnM= 3613
+IGhpbXNlbGY= 3614
+dXRz 3615
+IGp1 3616
+IGF3YXJl 3617
+IHJlcXVpcmU= 3618
+IHN5c3RlbXM= 3619
+IEhhcg== 3620
+IGFtb25n 3621
+IGhvbQ== 3622
+IGJyZWF0 3623
+IHdlaXJk 3624
+IOu2 3625
+zrs= 3626
+2Kk= 3627
+aWZm 3628
+b3Jpbmc= 3629
+IHBsYXRmb3Jt 3630
+IFRha2U= 3631
+IGhlbHBz 3632
+dXRpb25z 3633
+IGZvcmc= 3634
+IGx1Y2s= 3635
+IEVuZ2xpc2g= 3636
+IHdlYg== 3637
+IG5lZ2F0aXZl 3638
+IHR1dA== 3639
+IGFib3Zl 3640
+bmd0aA== 3641
+IOqxsA== 3642
+IHN0b3JpZXM= 3643
+IGxvYWQ= 3644
+IGJhY2tncm91bmQ= 3645
+IHN3aXRjaA== 3646
+Z2E= 3647
+IHByaW5jaQ== 3648
+IGZpbmFu 3649
+IHZhcmlvdXM= 3650
+IGzDoA== 3651
+IGtpbmRz 3652
+YWluaW5n 3653
+IG5hdHVyZQ== 3654
+INCe 3655
+Y3o= 3656
+IHByYXk= 3657
+IGdhcg== 3658
+aXJt 3659
+ICY= 3660
+IOyD 3661
+bnM= 3662
+IFJlcA== 3663
+IEZl 3664
+IHJldg== 3665
+cmFuZA== 3666
+IGxpa2VseQ== 3667
+IHVuZGVyc3RhbmRpbmc= 3668
+xLFy 3669
+IGZhbA== 3670
+IDEz 3671
+0YbQuA== 3672
+IHN1ZA== 3673
+IGJyb3RoZXI= 3674
+IHBsYW50 3675
+IHRocm91Z2hvdXQ= 3676
+d2lzZQ== 3677
+cHJl 3678
+IGN1bHR1cmU= 3679
+INmF 3680
+IHdvbmRlcmZ1bA== 3681
+IGFo 3682
+cHBlcg== 3683
+IHNvbGQ= 3684
+IHN0YXJ0cw== 3685
+IHdyaXR0ZW4= 3686
+zq8= 3687
+bmk= 3688
+INeU1w== 3689
+IERhdg== 3690
+IHVsdA== 3691
+IGFybQ== 3692
+IHJvY2s= 3693
+IHdlYXI= 3694
+642w 3695
+YW5v 3696
+cmFn 3697
+IHNxdWFyZQ== 3698
+0LDQvdC4 3699
+Y2FzdA== 3700
+bGVicg== 3701
+IGxpdGVyYWxseQ== 3702
+IHBsYXllZA== 3703
+IGhlYXQ= 3704
+b25zZQ== 3705
+cmljdA== 3706
+IGluc3A= 3707
+aWRz 3708
+IHBvcHVsYXI= 3709
+64+E 3710
+IGNhdGNo 3711
+IG1vdW50 3712
+IGp1ZA== 3713
+V2hhdA== 3714
+0LXQsQ== 3715
+UkE= 3716
+YXVk 3717
+0LrQvg== 3718
+IHN1cmZhY2U= 3719
+IGNvbnY= 3720
+IHBpZWNlcw== 3721
+T2g= 3722
+5oA= 3723
+IHN0eWxl 3724
+cHBpbmc= 3725
+IHJlYWRpbmc= 3726
+IGNvbnZlcnNhdGlvbg== 3727
+0L7Qvw== 3728
+IEFnYWlu 3729
+IGJhbms= 3730
+dGltZQ== 3731
+0YPRgg== 3732
+ZXJ2ZQ== 3733
+IEdyZWF0 3734
+IGNhcHQ= 3735
+0LDQsQ== 3736
+YXlz 3737
+IEZpbg== 3738
+aWZpY2F0aW9u 3739
+IMOkcg== 3740
+0LDRjg== 3741
+IGVnZw== 3742
+IFdlbA== 3743
+IHRhcmdldA== 3744
+dWxh 3745
+Y2hlcw== 3746
+YW5p 3747
+T08= 3748
+aWNpb3Vz 3749
+bm93 3750
+z4M= 3751
+Ym9hcmQ= 3752
+IGdlbnRl 3753
+IGRybw== 3754
+IEV0 3755
+IGRpbg== 3756
+IGNvcw== 3757
+IGF1dGhvcg== 3758
+2LM= 3759
+IG9jaA== 3760
+IGVtYWls 3761
+IHNwaXJpdA== 3762
+IHNpdHRpbmc= 3763
+bWFz 3764
+IHN0cmVuZ3Ro 3765
+IGJpZ2dlcg== 3766
+IFdhaXQ= 3767
+IG1hdA== 3768
+IHBvbGljZQ== 3769
+cmVzc2Vk 3770
+IHdhaXRpbmc= 3771
+aXNoaW5n 3772
+IGRvbGxhcnM= 3773
+aG9vZA== 3774
+c3M= 3775
+IGltYWdpbmU= 3776
+aW5p 3777
+IG1lcw== 3778
+IGRpc2U= 3779
+aWRnZQ== 3780
+YWJvcg== 3781
+IHBldA== 3782
+IGhvcA== 3783
+IEtpbmc= 3784
+IGNvbXB1dGVy 3785
+IGdvbGQ= 3786
+IG51 3787
+IGZpbmc= 3788
+KSw= 3789
+IHNlY3VyaXR5 3790
+cnVjdGlvbg== 3791
+IHNvbHV0aW9u 3792
+ZXh0 3793
+IHBhdHRlcg== 3794
+aWNrZW4= 3795
+dXJlZA== 3796
+IHN0YW5kYXJk 3797
+7Iuc 3798
+IGRvdWJsZQ== 3799
+zrc= 3800
+IHdpZmU= 3801
+aXNh 3802
+IGRpcmVjdGx5 3803
+YWNlZA== 3804
+IGJ1bmNo 3805
+IMK/ 3806
+0LDQu9GM 3807
+IHJlZ2FyZA== 3808
+IHN3ZWV0 3809
+IHVuaXF1ZQ== 3810
+IOKZqw== 3811
+IHRyYWlu 3812
+IEdlcm0= 3813
+zqw= 3814
+UkU= 3815
+IGJlaGF2 3816
+IHByZWQ= 3817
+7IM= 3818
+c2V0 3819
+IGRlc2NyaXB0aW9u 3820
+w6ll 3821
+IGNhdA== 3822
+5ZM= 3823
+IGNvbGxlZ2U= 3824
+7Js= 3825
+IGFwcGxpY2F0aW9u 3826
+IFNlbg== 3827
+YXNr 3828
+IGNyZWQ= 3829
+dWJsaWM= 3830
+IG11bHRpcGxl 3831
+IG5p 3832
+IHByZXNpZGVudA== 3833
+IGFkZGVk 3834
+IHJvYg== 3835
+IGFxdWk= 3836
+IGhvc3A= 3837
+IHRvb2xz 3838
+IGd1bg== 3839
+IGJhc2lj 3840
+IGxpbmVz 3841
+IHN0cnVjdHVyZQ== 3842
+IFJ1c3M= 3843
+IHRvdGFsbHk= 3844
+IGJpZ2dlc3Q= 3845
+IGVlbg== 3846
+IGFyZw== 3847
+INec 3848
+IHBhcms= 3849
+IERlcw== 3850
+IGNlbGVicg== 3851
+IGZhaXQ= 3852
+0LXQvdGM 3853
+IHN1ZmY= 3854
+IHJlZ3VsYXI= 3855
+qOs= 3856
+IG1pbmU= 3857
+IEtvcmU= 3858
+IHByZXZpb3Vz 3859
+IHBp 3860
+IHNlZw== 3861
+IHBvbGljeQ== 3862
+INC60L4= 3863
+IFRydW1w 3864
+IHZhY2M= 3865
+w7N3 3866
+IFN5 3867
+0LjRhw== 3868
+aXR0ZXI= 3869
+IHBvbGl0aWNhbA== 3870
+cmFz 3871
+IGFscw== 3872
+0LXQu9GM 3873
+IHNoYXBl 3874
+YW56 3875
+IG9udG8= 3876
+IGFyY2g= 3877
+IGFtYg== 3878
+YWdyYW0= 3879
+IFNt 3880
+Y3Rpb25z 3881
+IGpvaW4= 3882
+Ym9y 3883
+5Zs= 3884
+IGZyYW1l 3885
+oIc= 3886
+IGNob2ljZQ== 3887
+4K+B 3888
+0YPRjg== 3889
+IENvcg== 3890
+IFN3 3891
+SVQ= 3892
+IHRlbmQ= 3893
+IEVhcg== 3894
+IHRvcg== 3895
+IGV2ZW50cw== 3896
+IGNsYWlt 3897
+IERh 3898
+IE1hcms= 3899
+IGdyb3Vwcw== 3900
+IGVhdGluZw== 3901
+IFdvcmxk 3902
+IHJlY2VudGx5 3903
+IHRhc3Rl 3904
+IHN1cnY= 3905
+4KQ= 3906
+IHNraWxscw== 3907
+INC40Lc= 3908
+aXR0ZWQ= 3909
+IHNob3A= 3910
+7J207A== 3911
+IGVzdGFi 3912
+IOuCmA== 3913
+IHNlY29uZHM= 3914
+IFRob3Nl 3915
+IEVudA== 3916
+IOyE 3917
+ZXJzb24= 3918
+IHRvd24= 3919
+IGNhbmQ= 3920
+IG9wdGlvbnM= 3921
+IGluZw== 3922
+VklE 3923
+IGVuY291cg== 3924
+IHLDqQ== 3925
+4pmq 3926
+IGVudHJl 3927
+IG1vdmVtZW50 3928
+IEJlbg== 3929
+IGJpcnRo 3930
+IHdoZQ== 3931
+IGhhbmc= 3932
+IEVt 3933
+aWdl 3934
+cm9sbA== 3935
+IHVuZg== 3936
+7II= 3937
+IHJpZA== 3938
+IHNwcmVhZA== 3939
+IGhvc3Q= 3940
+YWxk 3941
+IEVk 3942
+IGNvbnN1bQ== 3943
+VU4= 3944
+IG9waW4= 3945
+aXRhcg== 3946
+IE1lZA== 3947
+IHN1YmplY3Q= 3948
+IHBhbA== 3949
+IGNhcnJ5 3950
+IGFncmVl 3951
+IFdoaWxl 3952
+IGNhcmVlcg== 3953
+IHNjaWVudA== 3954
+IHN1ZGRlbg== 3955
+IGZpbGU= 3956
+emk= 3957
+IGV4Y2VwdA== 3958
+6bo= 3959
+IHBvdGVudGlhbA== 3960
+IEFub3RoZXI= 3961
+IGNvbXBsZXg= 3962
+IFNpbQ== 3963
+ZW5kbw== 3964
+IHJhaXM= 3965
+IHBoeXNpY2Fs 3966
+IGRhdGU= 3967
+YWtlcg== 3968
+IENvbA== 3969
+IHBvd2VyZnVs 3970
+IG1lbWJlcg== 3971
+cmFw 3972
+IHNwb3Q= 3973
+IHNvdXJjZQ== 3974
+IGZlbQ== 3975
+w6lt 3976
+IGVtcA== 3977
+amk= 3978
+aWV0eQ== 3979
+IGluZmx1 3980
+IGRyeQ== 3981
+IGxvY2s= 3982
+IHplcm8= 3983
+IFVo 3984
+IHJvdXQ= 3985
+IHBvcnF1ZQ== 3986
+IDI0 3987
+IHRhbA== 3988
+IGZvbGtz 3989
+IGxhdW5jaA== 3990
+IGNvbXBvbg== 3991
+IFdlbGNvbWU= 3992
+IGthbm4= 3993
+w6Ru 3994
+INGN0YI= 3995
+ZWVz 3996
+INmI 3997
+IGFueXdheQ== 3998
+IGF1ZGllbmNl 3999
+IHNsaWdodA== 4000
+b25h 4001
+IHVy 4002
+IHJlbGln 4003
+IGV4dHJlbQ== 4004
+xLF6 4005
+IE1h 4006
+zrw= 4007
+IMO2 4008
+IGFsbG93cw== 4009
+IGZhdA== 4010
+IEZhY2U= 4011
+IG5hdGlvbmFs 4012
+IGludGVydmlldw== 4013
+IE1j 4014
+w6l0 4015
+IGN1dGU= 4016
+ZWxh 4017
+IHNlY3JldA== 4018
+IFdlc3Q= 4019
+IERlcA== 4020
+IGV4ZXJj 4021
+IGhpc3Rvcg== 4022
+IHByaW9y 4023
+IDYw 4024
+YXZh 4025
+YWNoZXI= 4026
+eW9uZA== 4027
+IEhh 4028
+IGVzdGU= 4029
+aW5hcnk= 4030
+IE5vcnRo 4031
+b25zdA== 4032
+IHNtYXJ0 4033
+YW1z 4034
+0LDQu9C4 4035
+IGRhcg== 4036
+ZXJlZA== 4037
+IGZ1bm55 4038
+IE9i 4039
+IEJsYWNr 4040
+IHJlbGF0ZWQ= 4041
+IEJ1 4042
+IHNvbWV3aGVyZQ== 4043
+IFJlbQ== 4044
+bmVz 4045
+bWVudGU= 4046
+IFJlYWxseQ== 4047
+IGNyZWF0aW5n 4048
+IGZhbWls 4049
+IHNvY2lldHk= 4050
+IGdlbA== 4051
+IHRyYW5zZm9ybQ== 4052
+xIM= 4053
+IGluY2x1ZGU= 4054
+IGhvbA== 4055
+bGlrZQ== 4056
+a28= 4057
+YWlycw== 4058
+INC/0L7QtA== 4059
+IHBlcnNwZWN0 4060
+IGJlcw== 4061
+IHBhcnRpY3VsYXJseQ== 4062
+IHNob3dpbmc= 4063
+IFBhcnQ= 4064
+IHF1YWw= 4065
+bG9jaw== 4066
+IHJlYWxpdHk= 4067
+aG9sZA== 4068
+aWN0aW9u 4069
+b29u 4070
+IHZpcg== 4071
+aXRhcnk= 4072
+IGRydWc= 4073
+IGZlYXR1cmU= 4074
+IHJlYXNvbnM= 4075
+INep 4076
+IHdyb3Rl 4077
+IGZhbnQ= 4078
+IGJhbmQ= 4079
+2YM= 4080
+ZW5h 4081
+a2V5 4082
+IGVhcnRo 4083
+ZG9t 4084
+IGZlYXR1cmVz 4085
+IGZsb29y 4086
+IHNwZWFraW5n 4087
+IHRpcA== 4088
+IEF1c3Q= 4089
+IHN0b2Nr 4090
+IGNodXJjaA== 4091
+IHJhYw== 4092
+7Jy866Gc 4093
+4LiZ 4094
+a3k= 4095
+IHJlc3BvbnNl 4096
+24w= 4097
+dWxhdGlvbnM= 4098
+IHNsaWRl 4099
+IGdyYWR1 4100
+Y2lvdXM= 4101
+IG1lYW50 4102
+ID09 4103
+INeQ1w== 4104
+44U= 4105
+IGtpbmRh 4106
+IHNjZW5l 4107
+IG11aXQ= 4108
+IOqwgA== 4109
+cmFzdA== 4110
+cmVzdA== 4111
+IHBsYXllcnM= 4112
+d2E= 4113
+IGJyb2Fk 4114
+IHRvbW9ycm93 4115
+b2NvbA== 4116
+INGB0LI= 4117
+IEJhcg== 4118
+xLFr 4119
+IHNlYQ== 4120
+IHJlbW92ZQ== 4121
+IHJlbWluZA== 4122
+0L7QvNGD 4123
+IFNpbmNl 4124
+IGF2ZWM= 4125
+Y2VsbA== 4126
+0LjRhQ== 4127
+IGRvY3VtZW50 4128
+IOq3uOuf 4129
+IG5laWdo 4130
+YmVhdA== 4131
+IHDDpQ== 4132
+IGFzcGVjdA== 4133
+IGRlZA== 4134
+bGlzaGVk 4135
+aWxz 4136
+IG91cnNlbHZlcw== 4137
+dWNl 4138
+IGhleQ== 4139
+INC/0YDQvg== 4140
+ZW50eQ== 4141
+IGFzc29jaQ== 4142
+YWRvcw== 4143
+dW1iZXI= 4144
+IF0= 4145
+bm92 4146
+IOyZ 4147
+0YPRhw== 4148
+IGNvbmRpdGlvbg== 4149
+64qU642w 4150
+IHZhbHVlcw== 4151
+IHNjZW4= 4152
+bWluaXN0 4153
+IGNhc3Q= 4154
+IGdyb3dpbmc= 4155
+IHVzZXI= 4156
+IHJlc3BvbmQ= 4157
+bGlt 4158
+w6ly 4159
+eW0= 4160
+b3Nlcw== 4161
+c3ljaA== 4162
+INGA0LDQtw== 4163
+IGFwcGVhcg== 4164
+IHByb2dyZXNz 4165
+ZW5ndGg= 4166
+IGphaw== 4167
+IERpcw== 4168
+IHBhdGllbnRz 4169
+IFNlcg== 4170
+IGdhcw== 4171
+w6hyZQ== 4172
+7Ja07JqU 4173
+IHJlY2k= 4174
+7J24 4175
+IHNjYQ== 4176
+ZXBlbmQ= 4177
+0YHQug== 4178
+0LDQvw== 4179
+IGJhdHRlcg== 4180
+IHZlaA== 4181
+8J8= 4182
+IGFjY29t 4183
+IGJlYXQ= 4184
+IHBhaW50 4185
+IGNvbnRyaWI= 4186
+IHNhZA== 4187
+xrA= 4188
+YWxlcw== 4189
+IHRyZWU= 4190
+YmE= 4191
+IGJvcm4= 4192
+aWNlZA== 4193
+4K6V 4194
+YmFuZA== 4195
+IG1lY2hhbg== 4196
+IERldA== 4197
+IGNhcGl0YWw= 4198
+IGRlbGl2ZXI= 4199
+IGZlYXI= 4200
+npg= 4201
+IFNvdXRo 4202
+IGJvdWdodA== 4203
+IHN0cmVzcw== 4204
+IHZvcg== 4205
+Pz8= 4206
+aWg= 4207
+7JW8 4208
+IGVyYQ== 4209
+7J206w== 4210
+0LDRjw== 4211
+aXNpb25z 4212
+aXZpdHk= 4213
+IGhlbHBlZA== 4214
+IGFzc2lzdA== 4215
+IHBsYXllcg== 4216
+cmFu 4217
+IGltbWVkaWF0ZWx5 4218
+IG1vdmVk 4219
+Y2ll 4220
+6rE= 4221
+IGFubm91bg== 4222
+5b8= 4223
+7J6Q 4224
+IHByb2R1Y3Rpb24= 4225
+IHN1bW1lcg== 4226
+IHR1bg== 4227
+IHByb2dyYW1z 4228
+R0g= 4229
+YWxpbmc= 4230
+aXJh 4231
+ZWxlc3M= 4232
+Lik= 4233
+IGF2ZXJhZ2U= 4234
+IGdsYXNz 4235
+b21hbg== 4236
+aWZpY2FsbHk= 4237
+IOuLpA== 4238
+IENvbmc= 4239
+IFZlcg== 4240
+IHRyaWNr 4241
+IGJlZ2Fu 4242
+IHZpbGw= 4243
+6rGw 4244
+aG93 4245
+5q0= 4246
+IHRpbGw= 4247
+IDkw 4248
+YmVydA== 4249
+IOq4 4250
+IHRlbXBlcmF0dXJl 4251
+w7I= 4252
+4LmI 4253
+IGdyYXBo 4254
+IOq3uA== 4255
+IHJvdA== 4256
+IG1vYg== 4257
+QVk= 4258
+YWVs 4259
+IHJlcGU= 4260
+IGRldmljZQ== 4261
+IDE5OQ== 4262
+IHRlbGU= 4263
+IGtlcHQ= 4264
+cGE= 4265
+5pY= 4266
+dmVyc2U= 4267
+IHN0cmVhbQ== 4268
+0LXRhw== 4269
+ZXNzaW9u 4270
+IHN0cnVnZw== 4271
+eno= 4272
+IGRlZ3JlZQ== 4273
+IGhlbHBpbmc= 4274
+IHNtZWxs 4275
+IHBlcmhhcHM= 4276
+cHJv 4277
+IGNvbnRleHQ= 4278
+IGlr 4279
+INC/0LXRgA== 4280
+IGNhbGN1bA== 4281
+YmluZw== 4282
+IHJlYWxpemU= 4283
+bGFt 4284
+IENoYXI= 4285
+eXQ= 4286
+IOydtOw= 4287
+IGRhbmdlcg== 4288
+IElt 4289
+YWE= 4290
+IGxvdmVk 4291
+IHB1cnBvc2U= 4292
+IGZpbmlzaGVk 4293
+IHBlYWNl 4294
+IG90 4295
+IGdsb2JhbA== 4296
+z4A= 4297
+IGFiZXI= 4298
+log= 4299
+IGNoYXJhY3RlcnM= 4300
+IG51cg== 4301
+IGRhbWFnZQ== 4302
+IGVtZXI= 4303
+IHByZWM= 4304
+IFdpcg== 4305
+IGluc3RpdA== 4306
+kdc= 4307
+IGFsbG93ZWQ= 4308
+Ym9u 4309
+IHRvZA== 4310
+0LXQs9C+ 4311
+IGpldHp0 4312
+IG1lZGlj 4313
+IHNtYWxsZXI= 4314
+Y2VlZA== 4315
+IGxldmVscw== 4316
+IGludGVsbA== 4317
+V2U= 4318
+IHNlbQ== 4319
+IGN1cnJlbnRseQ== 4320
+IG1vZGVybg== 4321
+IGNvbnRyYWN0 4322
+IGRldGFpbHM= 4323
+b3J0dW5hdGVseQ== 4324
+T1M= 4325
+IHN0YXRlcw== 4326
+IGFkanVzdA== 4327
+YW50YWdl 4328
+ZXo= 4329
+IFZlcnk= 4330
+IHNjYWxl 4331
+IHJlbGVhc2U= 4332
+IGZheg== 4333
+IGlj 4334
+aXR1ZGU= 4335
+QUM= 4336
+IFBhdA== 4337
+aWRlbg== 4338
+rZA= 4339
+IHByZWZlcg== 4340
+b2xvZ2ljYWw= 4341
+IEZhY2Vib29r 4342
+IOqwmQ== 4343
+IC4u 4344
+IE1ha2U= 4345
+INC60L7RgtC+0YA= 4346
+IERhdmlk 4347
+IEFmcmlj 4348
+IG1vZGU= 4349
+IENpdHk= 4350
+IHNoYWxs 4351
+INGE 4352
+aW1pbg== 4353
+INC30LA= 4354
+cm9t 4355
+dWE= 4356
+IGJleW9uZA== 4357
+IGRpc3RyaWI= 4358
+0LrRgw== 4359
+IERvZXM= 4360
+IHZpY3Q= 4361
+cmF0ZQ== 4362
+IHZhaQ== 4363
+IHN1Y2Nlc3NmdWw= 4364
+IGhvdXM= 4365
+YWhh 4366
+ZXN0cw== 4367
+IEVzdA== 4368
+IGRpc2NvdmVy 4369
+IHRoZXJlZm9yZQ== 4370
+Y2hh 4371
+IGN1cA== 4372
+IHBvcHVsYXRpb24= 4373
+IEls 4374
+c2M= 4375
+IHNwZW50 4376
+cmVs 4377
+IHVzZWZ1bA== 4378
+IHRhYg== 4379
+5p0= 4380
+IMU= 4381
+IOygnA== 4382
+IGNvbnNl 4383
+IHF1YW50 4384
+YXlh 4385
+IGJvbg== 4386
+IENoaW4= 4387
+IOqygw== 4388
+b3VuZHM= 4389
+0LXRiA== 4390
+ZWxsZQ== 4391
+IGljZQ== 4392
+MjE= 4393
+IGtpY2s= 4394
+IHN0ZXBz 4395
+IHRvbmlnaHQ= 4396
+0L3Ri9C5 4397
+cmVuY2g= 4398
+Lic= 4399
+IGdyYWI= 4400
+IGltcGxlbWVudA== 4401
+IOyImA== 4402
+IG1pc3Npb24= 4403
+IGNsZWFybHk= 4404
+IGFwcHJlY2lhdGU= 4405
+6IA= 4406
+IGZyZXNo 4407
+YXJt 4408
+IFR3bw== 4409
+IGV4ZWM= 4410
+IHByb2plY3Rz 4411
+IGNvbW11bml0aWVz 4412
+cmlibGU= 4413
+IHJlZ2lvbg== 4414
+IGZyZXF1 4415
+cm95 4416
+IGhvd2V2ZXI= 4417
+IHBhcnRuZXJz 4418
+YW5j 4419
+IG1pbmlt 4420
+IGxhdA== 4421
+IGZhbWlsaWVz 4422
+IGV2aWRlbmNl 4423
+IHB1bg== 4424
+cmFmdA== 4425
+IGxvc3M= 4426
+IG1hcA== 4427
+IGFueWJvZHk= 4428
+IGNoYW5naW5n 4429
+IHJ1bGVz 4430
+IG9yZ2FuaXphdGlvbg== 4431
+IGVzc2VudGlhbGx5 4432
+IFJlZA== 4433
+IGVsZW1lbnQ= 4434
+5pc= 4435
+IHZpcnQ= 4436
+cmF0 4437
+IHByaW50 4438
+YW5kZXI= 4439
+YXJlbg== 4440
+ZW1vcw== 4441
+zr/PhQ== 4442
+IGNvbmRpdGlvbnM= 4443
+YWJl 4444
+IGRhbmNl 4445
+0LjRgA== 4446
+IGRvcw== 4447
+0L7Rhw== 4448
+IFF1ZQ== 4449
+IHdhbGtpbmc= 4450
+IHRybw== 4451
+IGlk 4452
+IGFkZGl0aW9uYWw= 4453
+IGZ1bGx5 4454
+IGZhbnM= 4455
+IGFkZGl0aW9u 4456
+IGxpa2Vk 4457
+IMO8YmVy 4458
+IGJvdw== 4459
+ZGk= 4460
+IG1hc3Rlcg== 4461
+b2Zm 4462
+KTo= 4463
+bWJlcg== 4464
+IOus 4465
+5a8= 4466
+bGF1c2U= 4467
+IG9kZXI= 4468
+IHNhZmV0eQ== 4469
+IHJlYWN0 4470
+4K6/ 4471
+YnQ= 4472
+IGRpc2FwcA== 4473
+IGdpcmxz 4474
+U3Q= 4475
+IEFuZw== 4476
+IGZhaXRo 4477
+IHR1cm5z 4478
+IHRpZ2h0 4479
+IG1vdXRo 4480
+YW1p 4481
+emVy 4482
+IHdlYXA= 4483
+INCx0YPQtA== 4484
+IGhvc3BpdGFs 4485
+cmFpZA== 4486
+IG1pY3Jv 4487
+IFN0YXRl 4488
+IE1vc3Q= 4489
+YWdu 4490
+IGRlY2lkZQ== 4491
+IHBhdGllbnQ= 4492
+IGNvcm5lcg== 4493
+IGRpZWQ= 4494
+Tm8= 4495
+IFN0dWQ= 4496
+cmVuZA== 4497
+ZW1wdA== 4498
+IGxpZQ== 4499
+IGxpZg== 4500
+IEJlZm9yZQ== 4501
+dMOz 4502
+IFN1cGVy 4503
+IGJlbGw= 4504
+NjA= 4505
+IHByaXZhdGU= 4506
+IFBhdWw= 4507
+IGdpYg== 4508
+IGFncmU= 4509
+tOyEnA== 4510
+IHNpZw== 4511
+IGludmVzdGln 4512
+0Y/Rgg== 4513
+ZW5pbmc= 4514
+IGRpc3RhbmNl 4515
+IHdhcm0= 4516
+IGRpZ2l0YWw= 4517
+aW5lcg== 4518
+IHBhbmQ= 4519
+IENPVklE 4520
+0LPQvg== 4521
+Z24= 4522
+IHJhY2U= 4523
+IHByb3Vk 4524
+IHRlYWNoaW5n 4525
+INGC0L4= 4526
+7J6l 4527
+IEFsbGFo 4528
+SW4= 4529
+IHdvb2Q= 4530
+IGNvbG9ycw== 4531
+IHdpcmQ= 4532
+dWo= 4533
+aWRhZA== 4534
+IGN1c3RvbWVycw== 4535
+IGNvbm5lY3RlZA== 4536
+IGxheWVy 4537
+IGFjaGlldmU= 4538
+IHBlcnNwZWN0aXZl 4539
+IENvbGw= 4540
+2YI= 4541
+IGNsb3Vk 4542
+ISEh 4543
+IGVuZGVk 4544
+oIfqsow= 4545
+IG1hbmFnZW1lbnQ= 4546
+IHJpY2g= 4547
+IHN1YnN0 4548
+IHJlbW8= 4549
+IHNlcnZl 4550
+IHJlc2lzdA== 4551
+IHRob3VnaHRz 4552
+IGdyb3d0aA== 4553
+aWxpYXI= 4554
+IHJpZ2h0cw== 4555
+IGNoYXJnZQ== 4556
+IGNvbnNpc3Q= 4557
+IHdlcmRlbg== 4558
+IGVtYg== 4559
+YW5kb20= 4560
+IGh1cnQ= 4561
+IGthbg== 4562
+aWFz 4563
+0LvQvg== 4564
+IHNoaXQ= 4565
+IGJlZw== 4566
+IHJlY2VpdmVk 4567
+aXRhdGlvbg== 4568
+IG1lYXQ= 4569
+IGlzc28= 4570
+ZmZlZQ== 4571
+IGZhbW91cw== 4572
+IGNvbWZvcnRhYmxl 4573
+SUw= 4574
+IEJ5ZQ== 4575
+b3RoZXM= 4576
+IG1lZGljYWw= 4577
+IGVuam95ZWQ= 4578
+IGhlYWx0aHk= 4579
+IHd5 4580
+Y2llcw== 4581
+IGVmZm9ydA== 4582
+IGRvY3Rvcg== 4583
+IG1pbGl0YXJ5 4584
+TEFV 4585
+IGdybw== 4586
+IGJhdHRsZQ== 4587
+IGZlZA== 4588
+IGNhcGFj 4589
+IGFmcmFpZA== 4590
+aXZpbA== 4591
+INCy0YHQtQ== 4592
+IGxlbmd0aA== 4593
+eXNpcw== 4594
+IGJlaQ== 4595
+pO0= 4596
+IG9yZ2FuaXo= 4597
+b3Jn 4598
+aW5j 4599
+IGludGVyYWN0 4600
+IENoaW5lc2U= 4601
+IGFjY29yZGluZw== 4602
+IGluY3JlZGlibGU= 4603
+IGtpbGxlZA== 4604
+IGRhdWdodGVy 4605
+IM+A 4606
+0YvQsg== 4607
+IHNjaG9vbHM= 4608
+IMKr 4609
+bGxlcg== 4610
+IHNob3VsZG4= 4611
+bmFs 4612
+IGNyaXM= 4613
+IGNoaWNrZW4= 4614
+IGZhc3Rlcg== 4615
+IGV4dHJlbWVseQ== 4616
+IG9wcG9z 4617
+IG5vdXM= 4618
+ICs= 4619
+cmlh 4620
+IGZpbmFuY2lhbA== 4621
+IGV4Y2l0aW5n 4622
+IGpvdXJuZXk= 4623
+15nXnQ== 4624
+oOs= 4625
+IGRpc3BsYXk= 4626
+IG1lbW9yeQ== 4627
+IGhlYXZ5 4628
+0L3QtQ== 4629
+IHBhc3NlZA== 4630
+0YDQuA== 4631
+aWxlcw== 4632
+IHBzeWNo 4633
+IHNwZWNpZmljYWxseQ== 4634
+IGVuZ2FnZQ== 4635
+IGxlZA== 4636
+b3JnZQ== 4637
+IERlbQ== 4638
+b3JkZXI= 4639
+IDgw 4640
+IGNyZWFt 4641
+ZXN0ZXJkYXk= 4642
+IGVkZ2U= 4643
+INC/0L7Quw== 4644
+IGJ1bGw= 4645
+IGluZGlj 4646
+IGt0w7M= 4647
+IGhvcGVmdWxseQ== 4648
+dW1lbnRz 4649
+YWdlbg== 4650
+0L3QvtCz0L4= 4651
+IGhhdGU= 4652
+Y2h0 4653
+ODA= 4654
+IGVmZmlj 4655
+IOyngA== 4656
+IGludGVybmV0 4657
+IGJ1ZGdldA== 4658
+IHByb3BlcnR5 4659
+aWRheQ== 4660
+IOya 4661
+INC80L7Qtg== 4662
+b2xh 4663
+IHNob3dlZA== 4664
+IE1vbg== 4665
+IHRob3VzYW5k 4666
+QVA= 4667
+IHBvb3I= 4668
+dXNlZA== 4669
+IEphY2s= 4670
+IHPDpQ== 4671
+g70= 4672
+IGVzYw== 4673
+IHNvZnR3YXJl 4674
+IHF1YXI= 4675
+INio 4676
+IG5lY2Vzc2FyaWx5 4677
+b21lbg== 4678
+aXk= 4679
+IGV2ZW50dWFsbHk= 4680
+aXNoZWQ= 4681
+IGJyaWdodA== 4682
+RUQ= 4683
+IHNwbA== 4684
+IGRlbWFuZA== 4685
+IHRocmVhdA== 4686
+IHNpcg== 4687
+IHJlbGVhc2Vk 4688
+Y2tldA== 4689
+IOKAqw== 4690
+IHJlcXVpcmVk 4691
+IHZvdGU= 4692
+7Lk= 4693
+4K6k 4694
+IGRldmVsb3BlZA== 4695
+IOyCrA== 4696
+YXRvcnk= 4697
+IGRpcg== 4698
+Y2FwZQ== 4699
+IHNsaWdodGx5 4700
+w6w= 4701
+4LmJ 4702
+cmVldA== 4703
+IGRpc2Vhc2U= 4704
+IGNvdXJ0 4705
+IGl0ZW1z 4706
+IEVhcnRo 4707
+0YHRgtC4 4708
+0LbQtQ== 4709
+7LI= 4710
+IGNoYWxsZW5nZXM= 4711
+IEJyaXQ= 4712
+IGRlc2lnbmVk 4713
+MTI= 4714
+IGhlYXJpbmc= 4715
+IGxpc3RlbmluZw== 4716
+em8= 4717
+INGB0Ls= 4718
+IHBlcm8= 4719
+IHdlYXJpbmc= 4720
+cGxpYw== 4721
+IGNoZW0= 4722
+IGJhbGFuY2U= 4723
+IGJh 4724
+IHJlY2VpdmU= 4725
+aW1h 4726
+IHNpZ25pZmljYW50 4727
+INC80Ys= 4728
+YW5jaA== 4729
+IENy 4730
+IENvdW4= 4731
+6riI 4732
+IGpvYnM= 4733
+IG9mZmljaWFs 4734
+IHBlcm0= 4735
+b21z 4736
+IG9wcG9ydHVuaXRpZXM= 4737
+IG92ZXJhbGw= 4738
+IGh1cw== 4739
+b2Rlcw== 4740
+IG5hdGlvbg== 4741
+IFJlZw== 4742
+IG9yZA== 4743
+IHJlc3RhdXI= 4744
+IOyG 4745
+IG1lbA== 4746
+dmlu 4747
+IHdlbm4= 4748
+IGvDtm4= 4749
+5oM= 4750
+IG9waW5pb24= 4751
+6Kw= 4752
+IFNvbWV0aW1lcw== 4753
+54I= 4754
+0YnQtQ== 4755
+YXNj 4756
+T1U= 4757
+IDIwMjA= 4758
+IGRlbGljaW91cw== 4759
+aWdlcg== 4760
+IOyViA== 4761
+b2xl 4762
+IGhhbmRsZQ== 4763
+IGNpdA== 4764
+IO2VnA== 4765
+IGbDtnI= 4766
+b290aA== 4767
+IG5lY2Vzc2FyeQ== 4768
+IGluZGVwZW5k 4769
+5oQ= 4770
+aXN0ZW4= 4771
+aGFt 4772
+IMOpdA== 4773
+IG11bHRp 4774
+z4w= 4775
+Pyk= 4776
+IGNhbXB1cw== 4777
+IHRvcGlj 4778
+IHJhaW4= 4779
+IHBhbmVs 4780
+IFNhbQ== 4781
+IGxhcmdlcg== 4782
+YXVkaWVuY2U= 4783
+IHBhaWQ= 4784
+IGVjb25vbWlj 4785
+b2x0 4786
+IHN0cmVldA== 4787
+IENvbnQ= 4788
+IGRyaXZpbmc= 4789
+IOyggA== 4790
+IGhheQ== 4791
+IHByb2Zlc3Npb25hbA== 4792
+IEludGVybg== 4793
+5bg= 4794
+IGlucHV0 4795
+IGNhdGVn 4796
+IGNybw== 4797
+IGxs 4798
+RVQ= 4799
+0YvQuQ== 4800
+Kio= 4801
+IFpl 4802
+QkxF 4803
+IOyk 4804
+cmVlcw== 4805
+INCv 4806
+ZWRl 4807
+aWVydA== 4808
+IGZvbGQ= 4809
+IGR1cg== 4810
+IE5hdGlvbmFs 4811
+IOyWtOs= 4812
+YW5jZWQ= 4813
+IGZhaXJl 4814
+dXRlZA== 4815
+IGtpbmc= 4816
+IHdpbGQ= 4817
+b2k= 4818
+dXBiZWF0 4819
+IHByZXZlbnQ= 4820
+aXVz 4821
+IMOo 4822
+IHdpZGU= 4823
+IHJpbmc= 4824
+IHRpdGxl 4825
+IHN0YW5kaW5n 4826
+IGFsdGhvdWdo 4827
+IGhp 4828
+IHNhdWNl 4829
+IHNpZGVz 4830
+IGFuaW1hbHM= 4831
+aWxpbmc= 4832
+YXRpdmVz 4833
+7JeQ7ISc 4834
+IE92ZXI= 4835
+IGRlc3A= 4836
+IGNvbnNpZGVyZWQ= 4837
+YXJpZXM= 4838
+aWVycw== 4839
+IGVpbmVu 4840
+IHNpc3Rlcg== 4841
+IOuV 4842
+IFN1cmU= 4843
+cmllbmQ= 4844
+YWlnbg== 4845
+IHNob3du 4846
+IHNhYw== 4847
+IHNvbnQ= 4848
+IGNlbnR1cnk= 4849
+IHRpZW4= 4850
+IM66 4851
+IFNU 4852
+IG9sZGVy 4853
+aWVt 4854
+IHRydWx5 4855
+IFNp 4856
+IHdpbmRvdw== 4857
+aXF1ZXM= 4858
+YXJpbw== 4859
+IGxvY2F0aW9u 4860
+zro= 4861
+IOyc 4862
+dmk= 4863
+YWd1ZQ== 4864
+IFNvcnJ5 4865
+IGRpc3A= 4866
+IGhlbGw= 4867
+IMOJ 4868
+IHRyYWRl 4869
+IGNyaXRpY2Fs 4870
+IOqx 4871
+IG5hbWVk 4872
+IHByZXBhcmVk 4873
+IEhvdXNl 4874
+YWx1 4875
+IHRvdWdo 4876
+IHRyaXA= 4877
+IHNhbmQ= 4878
+Y2Vs 4879
+w7x6 4880
+IFB1dA== 4881
+IGFwYXJ0 4882
+aXNm 4883
+dmlz 4884
+IGxpYnI= 4885
+YXZlbg== 4886
+IHZpZQ== 4887
+IGVmZmVjdGl2ZQ== 4888
+4Liy 4889
+IG1hZ24= 4890
+IG11aXRv 4891
+IOq1 4892
+aGFs 4893
+IGxpbWl0 4894
+IG5pbmU= 4895
+IHdpbGxpbmc= 4896
+xLHFnw== 4897
+c3A= 4898
+0LXQsw== 4899
+aGk= 4900
+IGFsdA== 4901
+IEphbg== 4902
+IG9yaWdpbg== 4903
+IFVz 4904
+IGVsZW1lbnRz 4905
+IHVzZXM= 4906
+IGhlbHBmdWw= 4907
+IGZsYXQ= 4908
+IGZhbWlsaWFy 4909
+IFBhcms= 4910
+IGNvcmU= 4911
+IGNsb3Nlcg== 4912
+IGFjdGl2ZQ== 4913
+IGFkbWluaXN0 4914
+Q0U= 4915
+0L3Ri9C1 4916
+54Q= 4917
+IHJlbGF0aXZl 4918
+IG1lbnRhbA== 4919
+IHJhbmRvbQ== 4920
+IHBhcnRuZXI= 4921
+IHV0aWw= 4922
+cGhvbmU= 4923
+IHJ1bGU= 4924
+d3c= 4925
+IOyglQ== 4926
+IHNjaG9u 4927
+IGNvZmZlZQ== 4928
+SEE= 4929
+IGNvbm5lY3Rpb24= 4930
+IHVuaXQ= 4931
+bGF1Z2hpbmc= 4932
+bG9n 4933
+IGFwcGw= 4934
+0LvQsA== 4935
+dXNpYw== 4936
+IEJyYQ== 4937
+IGFueXdoZXJl 4938
+QVVESQ== 4939
+IHNlcGFyYXRl 4940
+Ym94 4941
+IGRpdmlk 4942
+IHRlc3Rpbmc= 4943
+IHNpY2s= 4944
+IHdlcmVu 4945
+INec1w== 4946
+IGFkdmFudGFnZQ== 4947
+IHRyYW5zZmVy 4948
+Jy4= 4949
+IOu5 4950
+IGZpbmRpbmc= 4951
+0L3QvtC5 4952
+IOyiiw== 4953
+IGZvcnQ= 4954
+IGVjb25vbXk= 4955
+IGxhY2s= 4956
+IGxlYXZpbmc= 4957
+IGRpbQ== 4958
+5Y4= 4959
+IFJlcw== 4960
+2K0= 4961
+IGRpc2N1c3Npb24= 4962
+0LXQvw== 4963
+IGdlcw== 4964
+ZHVjdA== 4965
+IGNoYWlu 4966
+IHVzZXJz 4967
+ZWNo 4968
+xYJh 4969
+IGRpc2g= 4970
+IGNhcmVmdWw= 4971
+IHRlYWNoZXI= 4972
+IG9wdGlt 4973
+IGZsdQ== 4974
+YXRpY2FsbHk= 4975
+IHJlZmxlY3Q= 4976
+IHRyZWF0bWVudA== 4977
+ZWVk 4978
+acSZ 4979
+w7k= 4980
+4K6+ 4981
+IGVxdWlw 4982
+IHBsYW5uaW5n 4983
+IHNvbHZl 4984
+IFRvbQ== 4985
+IGF2b2lk 4986
+IHBvdQ== 4987
+IGdyZWF0ZXI= 4988
+bGlu 4989
+T0w= 4990
+IEx1 4991
+IE1vcmU= 4992
+IGF0dHJhY3Q= 4993
+w6pu 4994
+dW5h 4995
+IHBob3Rv 4996
+ZXJhdGlvbg== 4997
+IHBsYW5ldA== 4998
+IGNvcHk= 4999
+IHZpc3VhbA== 5000
+aXJpbmc= 5001
+IGludGVybmF0aW9uYWw= 5002
+IGxhdWdoaW5n 5003
+IHRoaWNr 5004
+IGhvbGRpbmc= 5005
+IGJyaW5naW5n 5006
+IGxldHRlcg== 5007
+IGJ1cm4= 5008
+IGVmZmVjdHM= 5009
+aXTDqQ== 5010
+b3Vycw== 5011
+T1Q= 5012
+w6ptZQ== 5013
+IFNjaG9vbA== 5014
+15XXqg== 5015
+cm9wcmk= 5016
+bGln 5017
+zrHOuQ== 5018
+IGFkdWx0 5019
+IHN1Z2Fy 5020
+IHJpZGU= 5021
+IGhpZ2hsaWdodA== 5022
+IG5vYm9keQ== 5023
+IDIx 5024
+IGNoYXQ= 5025
+INC/0YDQuA== 5026
+IGlubm92 5027
+dW5nZW4= 5028
+IGF0dGFjaA== 5029
+ZWRvbQ== 5030
+5Yo= 5031
+eWw= 5032
+IGxlZ2Fs 5033
+IHJpY2U= 5034
+IGNvbGxhYm9y 5035
+a2luZw== 5036
+ZG93bg== 5037
+5pk= 5038
+IGlo 5039
+IEFj 5040
+b3VzbHk= 5041
+IHJhcA== 5042
+IHNvbGlk 5043
+IGdlbmVyYWxseQ== 5044
+IHBhdHRlcm4= 5045
+YWxp 5046
+4Lit 5047
+IHRyYW5zbA== 5048
+aW50ZXI= 5049
+YXVsdA== 5050
+IOuo 5051
+IGV4cHJlc3M= 5052
+IGV4YW1wbGVz 5053
+IGNob3Nl 5054
+IHRlbGxz 5055
+w61z 5056
+YWludA== 5057
+IFRlbGw= 5058
+IE1pY2hhZWw= 5059
+5qg= 5060
+IE51bWJlcg== 5061
+IHRhcA== 5062
+IGV4cGVyaW1lbnQ= 5063
+IGJlbmVmaXQ= 5064
+IOyw 5065
+IHNlcXU= 5066
+IGV4cGVuc2l2ZQ== 5067
+IGdlbmVyYXRpb24= 5068
+IE1hbnk= 5069
+IGFkZGluZw== 5070
+IGtpbA== 5071
+IGNhbXBhaWdu 5072
+IEFudA== 5073
+cmF3 5074
+b21tZW4= 5075
+IHNvdWw= 5076
+am8= 5077
+IEFjdHVhbGx5 5078
+YW1t 5079
+6rKg 5080
+IG1heGlt 5081
+IHNhbHQ= 5082
+IGNydQ== 5083
+IGNhbGxpbmc= 5084
+IGJhc2lz 5085
+YmFu 5086
+IGtlZXBpbmc= 5087
+IE1vcg== 5088
+ZWRz 5089
+7IY= 5090
+IHRvZG8= 5091
+0LDQvNC4 5092
+0L3Rjw== 5093
+IGxpdmVk 5094
+IER1 5095
+Zm9yY2U= 5096
+ZmVyZW5jZQ== 5097
+YWxh 5098
+IG9jY3Vy 5099
+c2s= 5100
+IHJlY2VudA== 5101
+IGNhcnM= 5102
+IHRyYWRpdGlvbmFs 5103
+ZW50bGU= 5104
+sog= 5105
+IGhlbGQ= 5106
+IG5hY2g= 5107
+IENlbnRlcg== 5108
+ZXJlbg== 5109
+IGJpbg== 5110
+2YE= 5111
+IGNvbW1l 5112
+IHJldmU= 5113
+IOyYpA== 5114
+IGV4cGVjdGVk 5115
+YWJpbA== 5116
+IGZvY3VzZWQ= 5117
+b3Y= 5118
+IGlQ 5119
+b3JpYWw= 5120
+aXJv 5121
+IGV0Yw== 5122
+YW1pbmc= 5123
+IFNvbg== 5124
+IHllc3RlcmRheQ== 5125
+IHN0cmF0ZQ== 5126
+INGG 5127
+IOuP 5128
+cGVz 5129
+IGFjdGl2aXR5 5130
+IGFkdmljZQ== 5131
+IG9wZW5pbmc= 5132
+Zmlu 5133
+IHJlbGE= 5134
+6ZY= 5135
+IGluc3RhbmNl 5136
+IEV2ZXJ5b25l 5137
+Ymw= 5138
+cGVu 5139
+IHZpc2lvbg== 5140
+IEFsZXg= 5141
+aWZvcm4= 5142
+IHRpY2s= 5143
+SGU= 5144
+IHN0cmF0ZWd5 5145
+IGtvbQ== 5146
+UEU= 5147
+IEds 5148
+IGVsZWN0cmlj 5149
+MTU= 5150
+IGRhaWx5 5151
+IGh1c2JhbmQ= 5152
+IHN0YXRpb24= 5153
+IGFuYWx5c2lz 5154
+eW5hbQ== 5155
+IGF0dGVtcHQ= 5156
+IGJpbGxpb24= 5157
+dmFudA== 5158
+IGZvcnRo 5159
+IG1hdGg= 5160
+YWx5 5161
+IGJlaGF2aW9y 5162
+IE1hcw== 5163
+a2Fu 5164
+IERheQ== 5165
+IGJsZXNz 5166
+IGd1dA== 5167
+IEhpZ2g= 5168
+b3g= 5169
+IGRyZXNz 5170
+IGplZA== 5171
+6K8= 5172
+5ZY= 5173
+IGV4cGVyaWVuY2Vz 5174
+aXN0YQ== 5175
+IGZpZ2h0aW5n 5176
+5bc= 5177
+INGB0Lo= 5178
+IG1vc3RseQ== 5179
+YXVzZQ== 5180
+IHBpY3R1cmVz 5181
+0LXQvdGC 5182
+IG1hZA== 5183
+IG1vZGVscw== 5184
+0YjQtQ== 5185
+IENvdW50 5186
+xYQ= 5187
+xYJv 5188
+ZXB0 5189
+T00= 5190
+IEFO 5191
+IHRyb3VibGU= 5192
+NDA= 5193
+IGJpcmQ= 5194
+dWxhdGU= 5195
+IG11cg== 5196
+IHByb2R1Y2U= 5197
+IG1hcnJpZWQ= 5198
+Yml0 5199
+IHRoZW9yeQ== 5200
+7Zg= 5201
+IGxlYWRlcg== 5202
+IExhc3Q= 5203
+QUE= 5204
+6LU= 5205
+IGltYWdlcw== 5206
+IGV4cGFuZA== 5207
+IFBvcg== 5208
+IHB1cmNo 5209
+IFNhbg== 5210
+IENocmlzdG1hcw== 5211
+IEF1c3RyYWw= 5212
+IHdpZA== 5213
+IE1pc3M= 5214
+IGtub3dpbmc= 5215
+IHpl 5216
+c2hpcA== 5217
+a3U= 5218
+0YXQvtC0 5219
+IEluc3RhZ3JhbQ== 5220
+IEluZGlh 5221
+IGVzdGE= 5222
+IENhbGlmb3Ju 5223
+IDcw 5224
+IGRyYWc= 5225
+IGJydXNo 5226
+IG5hbWVz 5227
+QW5k 5228
+IHlv 5229
+aWxsYQ== 5230
+IHNjaGVk 5231
+IGRlc3Ryb3k= 5232
+eWVhcg== 5233
+IHZhbW9z 5234
+INmE 5235
+w6dh 5236
+IGZvcmdvdA== 5237
+0LjQtQ== 5238
+IHJhaXNl 5239
+cmVtZQ== 5240
+7ZW0 5241
+IEdpdmU= 5242
+IGNvbnRhaW4= 5243
+cmFi 5244
+IGdpZnQ= 5245
+INGB0L8= 5246
+IHJlcXVlc3Q= 5247
+IHNodXQ= 5248
+IGRlZ3JlZXM= 5249
+IGJlbmVmaXRz 5250
+0YvQtQ== 5251
+IHN0dWRpZXM= 5252
+IGVuZHM= 5253
+IGV2ZXJ5d2hlcmU= 5254
+IGhlcm8= 5255
+b3Bo 5256
+ZXJyeQ== 5257
+IG1hdGVyaWFscw== 5258
+ZW5lZA== 5259
+TkE= 5260
+5Y0= 5261
+IG11eQ== 5262
+IHdvcnNl 5263
+IE1hZA== 5264
+IGRlY2lzaW9ucw== 5265
+aW9uZQ== 5266
+IGZvcmVpZ24= 5267
+bGF1Z2h0ZXI= 5268
+aWJlcg== 5269
+0LXQvdC40Y8= 5270
+44WL 5271
+IHJlYWxpemVk 5272
+IGlnbg== 5273
+IHdlYWs= 5274
+IM68 5275
+IHNjYXJlZA== 5276
+IGFzc3Vt 5277
+QUs= 5278
+778= 5279
+77+9 5280
+IGNvdmVyZWQ= 5281
+IFNhdA== 5282
+INC+0L0= 5283
+IGluZGl2aWR1YWxz 5284
+IGNvbXBhcmVk 5285
+MTE= 5286
+IEFkZA== 5287
+aWNsZXM= 5288
+IGNlcnQ= 5289
+cmFy 5290
+IGJyaWVm 5291
+IGFjdGl2aXRpZXM= 5292
+IGZhYg== 5293
+YmFy 5294
+IGFzdA== 5295
+IE90aGVy 5296
+IGNsYXNzZXM= 5297
+IG9n 5298
+IG1pc3Npbmc= 5299
+6Z0= 5300
+d2Vycw== 5301
+16k= 5302
+IGludHJvZHVjZQ== 5303
+IGVxdWF0aW9u 5304
+IG5vbQ== 5305
+IHBhaW50aW5n 5306
+dXNoaW5n 5307
+IEFQ 5308
+IGVuY291cmFnZQ== 5309
+IHNoaXA= 5310
+aXR0ZWU= 5311
+aXZlcnNl 5312
+b3Rh 5313
+bmFt 5314
+IGV4ZXJjaXNl 5315
+INCt 5316
+IG5hcw== 5317
+IHRob3VzYW5kcw== 5318
+IENhbGlmb3JuaWE= 5319
+IHNlcw== 5320
+IHJvdw== 5321
+nog= 5322
+IHBhbmRlbWlj 5323
+IHNraWxs 5324
+YmVs 5325
+IGRpcmVjdG9y 5326
+IG1pbGs= 5327
+IG51dA== 5328
+IG1vdGlvbg== 5329
+IGNsb3NlZA== 5330
+6Kg= 5331
+IGNyZWRpdA== 5332
+YWhy 5333
+IGNoZWVzZQ== 5334
+IGFsdGVybg== 5335
+aW1hdGVseQ== 5336
+IHN1c3Q= 5337
+IFRyYQ== 5338
+IGdsYWQ= 5339
+IGhpZ2hseQ== 5340
+IHdh 5341
+IHJlZHVjZQ== 5342
+IGJsZQ== 5343
+YWRvcg== 5344
+aW5hdGVk 5345
+aW9uZXM= 5346
+Y2llbnQ= 5347
+IGRlcGVuZGluZw== 5348
+IHNoYXJpbmc= 5349
+IGNhdWdodA== 5350
+cmFlbA== 5351
+IG1laHI= 5352
+IHBhc3Npb24= 5353
+55s= 5354
+IHJ1 5355
+IGZhcm0= 5356
+VEk= 5357
+YXZlcw== 5358
+IFJvYg== 5359
+IEJybw== 5360
+IG1vdGl2 5361
+cmV0Y2g= 5362
+cnVwdA== 5363
+IEJpZw== 5364
+IGFsbGU= 5365
+IGV0dA== 5366
+dWJz 5367
+IEphcGFuZXNl 5368
+IEhhbGw= 5369
+0LjQu9C4 5370
+QVVESUJMRQ== 5371
+56w= 5372
+IGNlbGxz 5373
+aWth 5374
+ZWxpbmU= 5375
+aWxlcg== 5376
+IOyj 5377
+IHNreQ== 5378
+SU5BVURJQkxF 5379
+ZW5kZQ== 5380
+YXB0ZXI= 5381
+IHBpbg== 5382
+IGdhdGhlcg== 5383
+aG9s 5384
+bGVjdGlvbg== 5385
+IHN5bg== 5386
+IHBsdWc= 5387
+cm91bmQ= 5388
+IHVuaXZlcnNpdHk= 5389
+aGli 5390
+IGZhbnRhc3RpYw== 5391
+a24= 5392
+IGhvbGU= 5393
+IFJlbWVtYmVy 5394
+aW5jdA== 5395
+YWtz 5396
+Q0g= 5397
+IGJyb2tlbg== 5398
+IHN0cmF0ZWc= 5399
+IGFsaXZl 5400
+IHRhbms= 5401
+IGNhcnQ= 5402
+cmF0ZWQ= 5403
+cmll 5404
+IFN0ZXA= 5405
+IEV2ZXJ5dGhpbmc= 5406
+IGJvdW5k 5407
+IHNvYnJl 5408
+IGN1c3RvbWVy 5409
+oYw= 5410
+dXJn 5411
+IEJpbGw= 5412
+TGE= 5413
+d2hhdA== 5414
+IHJlYWN0aW9u 5415
+IHNlc3Npb24= 5416
+IHBsYW5z 5417
+IOydtOugh+qyjA== 5418
+IGRvd25sb2Fk 5419
+7Jk= 5420
+dWVy 5421
+IGNhYg== 5422
+IGluc3Ry 5423
+aWZ5aW5n 5424
+IE5pY2U= 5425
+IHRlYW1z 5426
+xLFs 5427
+IGdvYWxz 5428
+aXNjaA== 5429
+IHRyYW5zcG9ydA== 5430
+IGFuaW1hbA== 5431
+IGNvc3Rz 5432
+IGNhbGxz 5433
+IHNlaHI= 5434
+7Ig= 5435
+cmlhbg== 5436
+IGRpYWw= 5437
+IHdlYXRoZXI= 5438
+4LmA 5439
+INCy0L7Rgg== 5440
+IFBsYXk= 5441
+IHNoYXJlZA== 5442
+IHNtb290aA== 5443
+YWJh 5444
+IGxlYXZlcw== 5445
+4K6p 5446
+IGNvbmNlbnQ= 5447
+IHNoaWZ0 5448
+IOuQmA== 5449
+IEdvdmVybg== 5450
+IGRlbW9uc3Q= 5451
+IGJ1dHRlcg== 5452
+IOyXrA== 5453
+IHNhdGlzZg== 5454
+iOus 5455
+IHJlY29nbml6ZQ== 5456
+IEZyZW5jaA== 5457
+IHZvbHVtZQ== 5458
+w6RuZA== 5459
+0YPQvA== 5460
+IOynhA== 5461
+IEtlZXA= 5462
+b3dh 5463
+aXBwZWQ= 5464
+0YHRgtGA 5465
+IGRldGVjdA== 5466
+IM+D 5467
+IGxpZnQ= 5468
+IGNsb3RoZXM= 5469
+IFN0b3A= 5470
+w7U= 5471
+bWV0 5472
+IGNsaW4= 5473
+IGFycg== 5474
+ZnJpZW5k 5475
+IHN0dWNr 5476
+WWU= 5477
+aGFuZA== 5478
+dW1h 5479
+IHNjcmk= 5480
+IGZ1Y2tpbmc= 5481
+Y3RvcnM= 5482
+16o= 5483
+IGpvaW5pbmc= 5484
+IGNldHRl 5485
+INij 5486
+IFdoaXRl 5487
+IGlocg== 5488
+zq0= 5489
+IGluY2x1ZGVk 5490
+ZXNzbw== 5491
+IGFjYWQ= 5492
+YnVt 5493
+IHNhYg== 5494
+INC00LvRjw== 5495
+dWZhY3Q= 5496
+IFJlcHVibGlj 5497
+cmlt 5498
+IHllbGxvdw== 5499
+IGxpbWl0ZWQ= 5500
+VEVS 5501
+IFR5 5502
+IG5vdGVz 5503
+dmVzdA== 5504
+0LjQtw== 5505
+YWxlZA== 5506
+IHBoYXNl 5507
+YW5kYQ== 5508
+IE1vbQ== 5509
+Ukk= 5510
+IGltbWVy 5511
+bWFs 5512
+IGluag== 5513
+IHlhbmc= 5514
+dWRpYmxl 5515
+0LDQsw== 5516
+IHNldHQ= 5517
+IG1hZ2lj 5518
+IGVuc3VyZQ== 5519
+IHNwcmluZw== 5520
+IHNob2Nr 5521
+IHdoZWVs 5522
+0L7Qs9C00LA= 5523
+IGNhbmNlcg== 5524
+IHJvb3Q= 5525
+0JA= 5526
+Z2VuY3k= 5527
+IOuN 5528
+aWk= 5529
+IG91dHB1dA== 5530
+IGNvbW1pdA== 5531
+IHdvcmtlcnM= 5532
+7JWE7JqU 5533
+INGB0LDQvA== 5534
+dmV5 5535
+IHBldQ== 5536
+IGNpdmls 5537
+aXNj 5538
+IGJyaW5ncw== 5539
+0YDQsNCy 5540
+YW5pYQ== 5541
+xIE= 5542
+Y3JhZnQ= 5543
+bWJvbA== 5544
+IGludGVsbGln 5545
+Ymk= 5546
+YWNpbmc= 5547
+eW91 5548
+IGJlY29taW5n 5549
+IERlcg== 5550
+ZW1h 5551
+IGluZ3JlZA== 5552
+IGNvbW1hbmQ= 5553
+IHVwZGF0ZQ== 5554
+IHByZW0= 5555
+IG9wZW5lZA== 5556
+hKQ= 5557
+0LXQvdC40LU= 5558
+IGdhcmQ= 5559
+IHN0YXRlbWVudA== 5560
+IHNjcmV3 5561
+IHByb3Rl 5562
+IGNhcmRz 5563
+IHRhc2s= 5564
+IGV2ZW5pbmc= 5565
+IHN0aXRjaA== 5566
+aW5lbg== 5567
+IEJlcg== 5568
+bWFyaw== 5569
+IERhZA== 5570
+INC10YHRgtGM 5571
+INee1w== 5572
+7JeI 5573
+IGJhbg== 5574
+IGNsaW0= 5575
+IGZyZWVkb20= 5576
+IG5vcm1hbGx5 5577
+0LXRgdGM 5578
+5aY= 5579
+IHByb3ZpZGVk 5580
+IOyekA== 5581
+IOyVhOuLiA== 5582
+IEtpbQ== 5583
+aWVkZXI= 5584
+7J2M 5585
+IGNpdGl6 5586
+IGJpa2U= 5587
+IGJhaw== 5588
+IG5vaXNl 5589
+IGNsaW1hdGU= 5590
+aXplcw== 5591
+IGluY3JlYXNpbmc= 5592
+IFRIRQ== 5593
+IGxpcXU= 5594
+IHBlcnNvbmFsbHk= 5595
+ZWY= 5596
+cmVzcA== 5597
+IGxlZ3M= 5598
+aW5kZXI= 5599
+IHBlZA== 5600
+IOunjg== 5601
+IGRlcGVuZA== 5602
+IHZhcmlldHk= 5603
+IElzcmFlbA== 5604
+IHdhc2g= 5605
+5YY= 5606
+IHF1aWV0 5607
+IEphbWVz 5608
+IEpldw== 5609
+IGZvcmV2ZXI= 5610
+IEludA== 5611
+IGNvdW50ZXI= 5612
+dXJhbmNl 5613
+IEFueXdheQ== 5614
+Y2FyZQ== 5615
+IE9ubHk= 5616
+Y2nDs24= 5617
+YWRp 5618
+IEV2 5619
+64uI6rmM 5620
+IM6x 5621
+IHNsb3dseQ== 5622
+INC+0LQ= 5623
+IG5vdGljZWQ= 5624
+aWVyZW4= 5625
+IGZlbGw= 5626
+INCR 5627
+IG3Dqm1l 5628
+IHdoZW5ldmVy 5629
+ISk= 5630
+IEh5 5631
+5bw= 5632
+b3Jkcw== 5633
+dXNpb24= 5634
+IFN0YXI= 5635
+IO2Y 5636
+IE1hYw== 5637
+aXZlbg== 5638
+IOyLnA== 5639
+IOyXhg== 5640
+IFR1cg== 5641
+IGdlcg== 5642
+cmlz 5643
+IHZleg== 5644
+INC70Y4= 5645
+IHZlcnN1cw== 5646
+2KfY 5647
+b2NvbGF0ZQ== 5648
+IHBsYW5l 5649
+IHpv 5650
+IHN1aXQ= 5651
+VGhpcw== 5652
+IG5lcnY= 5653
+IEFjYw== 5654
+0YPQtg== 5655
+7IKs 5656
+bmg= 5657
+ZW1l 5658
+IGF1c3M= 5659
+IG1lYXM= 5660
+IHRyw6hz 5661
+z4k= 5662
+0YHQu9C4 5663
+IEFydA== 5664
+IFNlY29uZA== 5665
+0L7Qu9GM0LrQvg== 5666
+Y2hv 5667
+aXRlY3Q= 5668
+0LXRgdGC 5669
+IGJvc3M= 5670
+IGluY29tZQ== 5671
+oKQ= 5672
+IHNoYWQ= 5673
+IGFwcHJvcHJp 5674
+IE1hbA== 5675
+b3B0 5676
+IGFydGlzdA== 5677
+IHBsYXlz 5678
+b3RoZXJz 5679
+IEludGVy 5680
+IHZpcnVz 5681
+IGh1bmc= 5682
+IGNvbnN0YW50 5683
+IHNjcmlwdA== 5684
+IHNub3c= 5685
+dWxm 5686
+a2V0 5687
+IGRldmljZXM= 5688
+IG1ldGFs 5689
+aWdodHM= 5690
+7IS4 5691
+IHNhbGVz 5692
+IHZlZ2V0 5693
+IGNvbGxlY3Rpb24= 5694
+IHZpYQ== 5695
+a2Vy 5696
+IGdvdHRlbg== 5697
+T1c= 5698
+acOpbg== 5699
+IGFjY3Vy 5700
+IHdhdmU= 5701
+dWx0eQ== 5702
+IEFpcg== 5703
+IGxlYWRpbmc= 5704
+aWNpbmc= 5705
+IGNlbnRyYWw= 5706
+IENocmlzdGlhbg== 5707
+ZnI= 5708
+IEFsdGhvdWdo 5709
+IHNvbmdz 5710
+IGZpZg== 5711
+0L3Ri9GF 5712
+IGJlbG9uZw== 5713
+b3NzaWJsZQ== 5714
+7LA= 5715
+IHBob3Rvcw== 5716
+aXNs 5717
+IHJlbGF4 5718
+c2E= 5719
+VVNJQw== 5720
+6rc= 5721
+IG1hbnVmYWN0 5722
+IFR3aXR0ZXI= 5723
+IGRhbmdlcm91cw== 5724
+IGh5ZA== 5725
+bGVhcg== 5726
+aWFudA== 5727
+IOKApg== 5728
+IHN1ZGRlbmx5 5729
+IGxhdWdo 5730
+IGFuZ2xl 5731
+IEdvdA== 5732
+IHdvcnJpZWQ= 5733
+0L7QtQ== 5734
+IHBhcA== 5735
+IE1hcnQ= 5736
+ZW5v 5737
+IGJhdHRlcnk= 5738
+INC/0L7RgQ== 5739
+IGxpZ2h0cw== 5740
+IGFybXM= 5741
+IEFicw== 5742
+bWVz 5743
+4oCT 5744
+dXNldW0= 5745
+IHRlYQ== 5746
+IE1pYw== 5747
+IGZvcm1lcg== 5748
+b2dyYXBoeQ== 5749
+IGFwcGxpY2F0aW9ucw== 5750
+IERpcmU= 5751
+IGZlZWRiYWNr 5752
+aXRjaGVu 5753
+eW9ydW0= 5754
+dWVk 5755
+aWd0 5756
+xrDhuw== 5757
+b3NpdGlvbg== 5758
+IERlbA== 5759
+IO2VmOs= 5760
+IEJhY2s= 5761
+YWRz 5762
+IHByaW1l 5763
+7KO8 5764
+7KOg 5765
+15E= 5766
+IG11dA== 5767
+XS4= 5768
+INCX 5769
+bG9j 5770
+a2lu 5771
+IGV4cGVydA== 5772
+IGFscmlnaHQ= 5773
+dW5ncw== 5774
+IHN1cHBseQ== 5775
+IGxlYWRlcnNoaXA= 5776
+IEZyYQ== 5777
+IHR5cGljYWxseQ== 5778
+IHNlbA== 5779
+IHRyZWVz 5780
+IDIy 5781
+aGFy 5782
+IHdvcnN0 5783
+IGJ1c3k= 5784
+YW50bw== 5785
+IFVw 5786
+IEJhcw== 5787
+IHByZXNlbnRhdGlvbg== 5788
+IHN0cmFuZ2U= 5789
+IHRoaW4= 5790
+0YLQtQ== 5791
+IHZlaGljbGU= 5792
+INC00L4= 5793
+Y2VsbGVudA== 5794
+NzA= 5795
+IHRpcmVk 5796
+IGNyaXNpcw== 5797
+IHRpbnk= 5798
+YXN5 5799
+IHJhbg== 5800
+6Yc= 5801
+IGZvcmNlcw== 5802
+INC+0Yc= 5803
+IGlkZW50aWZ5 5804
+IGFzc2Vzcw== 5805
+0LjRgtC1 5806
+U0U= 5807
+IGNyZWF0aXZl 5808
+558= 5809
+IGRlcGFydG1lbnQ= 5810
+IGluaXRpYWw= 5811
+IERhbQ== 5812
+YWt0 5813
+dmVyZQ== 5814
+IGluZmVjdA== 5815
+IHB1bXA= 5816
+4bqh 5817
+IHZpZWw= 5818
+IHJhcmU= 5819
+IGRvdA== 5820
+YXNoaW9u 5821
+ZW1wbA== 5822
+IGZsZXg= 5823
+IGtvbg== 5824
+IHRydWNr 5825
+IGxlY3Q= 5826
+IHBsYXN0aWM= 5827
+bGF3 5828
+IGxpa2Vz 5829
+IHJvdWdo 5830
+IE1BVA== 5831
+7Z6I 5832
+IGNvbW1lcg== 5833
+IGFzc2U= 5834
+IGNha2U= 5835
+IGFjdGlvbnM= 5836
+IGFkbQ== 5837
+IG90aGVyd2lzZQ== 5838
+IEhlYWx0aA== 5839
+IGNvbGxl 5840
+4LmA4Lg= 5841
+IHJ1Yg== 5842
+5pQ= 5843
+IHNjcg== 5844
+IHp1bQ== 5845
+IEhpbQ== 5846
+IGNoYW1w 5847
+IGNvbmNlcm5lZA== 5848
+IDUwMA== 5849
+IHBsYXRl 5850
+IE91dA== 5851
+IGRvbmM= 5852
+IGVxdWlwbWVudA== 5853
+IHRhdWdodA== 5854
+bGxlZA== 5855
+IO2Z 5856
+aXZh 5857
+IG1vdG9y 5858
+wrs= 5859
+IGd1aWRl 5860
+5Yk= 5861
+IHN0b3BwZWQ= 5862
+IHJhdA== 5863
+IGxhYm9y 5864
+IGFpbQ== 5865
+IHByZXBhcmU= 5866
+INGI 5867
+IHNob290aW5n 5868
+YW5uZWQ= 5869
+Y3JpcHQ= 5870
+IGVuZW15 5871
+IGRlcGVuZHM= 5872
+IG5hdg== 5873
+IGJlcg== 5874
+IGxhbmRz 5875
+IHVuaXZlcnM= 5876
+aXU= 5877
+IGZhY3Rvcg== 5878
+b2tpbmc= 5879
+IGNhcmJvbg== 5880
+YnV0 5881
+IExvdmU= 5882
+ZWxk 5883
+IM61 5884
+IGdh 5885
+IMOpcw== 5886
+IGJyZWFk 5887
+IHZvbHQ= 5888
+7Yo= 5889
+IHdhc3Rl 5890
+IGtlZXBz 5891
+IHN0b3I= 5892
+IGhvbm9y 5893
+IHVubGVzcw== 5894
+IGNvbHVt 5895
+IOuMgA== 5896
+IHBsYW50cw== 5897
+WWVhaA== 5898
+IGluY2x1ZGVz 5899
+IG94 5900
+IHBldXQ= 5901
+66eM 5902
+7IOB 5903
+aXN0cnk= 5904
+4Lix 5905
+IERlcGFydG1lbnQ= 5906
+YW50YQ== 5907
+IGZpbmdlcg== 5908
+IHN0cmV0Y2g= 5909
+IHN5bWJvbA== 5910
+IG5laWdoYm9y 5911
+5qw= 5912
+6rCE 5913
+fn4= 5914
+INGC0Ys= 5915
+IEFiZXI= 5916
+a2Vz 5917
+IG1hc3NpdmU= 5918
+IENI 5919
+IFNhbA== 5920
+16A= 5921
+IGR5bmFt 5922
+YWNoZQ== 5923
+IFByZQ== 5924
+IG1vbml0b3I= 5925
+ZW50ZWQ= 5926
+RU8= 5927
+IHJhaXNlZA== 5928
+aXN0aWNz 5929
+2qk= 5930
+IHZvdQ== 5931
+aXRlbg== 5932
+obA= 5933
+IGJ1c2luZXNzZXM= 5934
+IGVhcm4= 5935
+IG1vYmlsZQ== 5936
+aWRhZGU= 5937
+IGhhYmU= 5938
+eXI= 5939
+bGljdA== 5940
+IGNvbmR1Y3Q= 5941
+IGZlZGVyYWw= 5942
+IHdv 5943
+YnU= 5944
+IG5vbmU= 5945
+IHRlYWNoZXJz 5946
+INin2YTY 5947
+aWRlbnRz 5948
+2KfZhA== 5949
+IHRyZW5k 5950
+0LXQtg== 5951
+IGFsYnVt 5952
+IG1pY2g= 5953
+YmFzZWQ= 5954
+4Li1 5955
+IHRyYW5zaXRpb24= 5956
+INC90L4= 5957
+w7Vlcw== 5958
+aG9zdA== 5959
+ZWR5 5960
+IFByb2Y= 5961
+cGFu 5962
+aWpu 5963
+IGNhcGFjaXR5 5964
+dW5kbw== 5965
+INeR1w== 5966
+IGJyZWF0aA== 5967
+INC80LXQvQ== 5968
+IG3DvA== 5969
+7Zk= 5970
+IEF1dA== 5971
+aGluZ3Rvbg== 5972
+IG5vcg== 5973
+IGdhaW4= 5974
+cG9pbnQ= 5975
+WWVz 5976
+INiq 5977
+IE5h 5978
+w6Vy 5979
+IGnDpw== 5980
+IE1hcnk= 5981
+IHNwaW4= 5982
+IGFudGk= 5983
+IHNvbWVob3c= 5984
+IGxhd3M= 5985
+IG1vbWVudHM= 5986
+IGdyZQ== 5987
+IG1vdmVz 5988
+IFdvdWxk 5989
+IHByZWRpY3Q= 5990
+IHZyYQ== 5991
+IDIwMTk= 5992
+toQ= 5993
+IGZ1bmRhbWVudA== 5994
+MjU= 5995
+IHB1cmU= 5996
+IHdvdw== 5997
+IGlzbGFuZA== 5998
+IGludmVzdG1lbnQ= 5999
+IGJhdGg= 6000
+IFlh 6001
+IGhhcmRlcg== 6002
+IHRpcHM= 6003
+5Zc= 6004
+IGVsZWN0cm9u 6005
+IEJvYg== 6006
+IGJvbmQ= 6007
+b2RpZXM= 6008
+IEF1Zw== 6009
+IGdpYnQ= 6010
+IGNoYWly 6011
+IHR3aWNl 6012
+d29vZA== 6013
+IGNsYXI= 6014
+IG1hc2s= 6015
+IGhvbmVzdGx5 6016
+IDIwMTg= 6017
+dGllcw== 6018
+Jyw= 6019
+IHBlbnM= 6020
+IHN1cnByaXNlZA== 6021
+IGNvbW11bmljYXRpb24= 6022
+IHNwcg== 6023
+IHdob3Nl 6024
+IHN0YXJz 6025
+15DX 6026
+IOKAiw== 6027
+IHByb3Blcmx5 6028
+IGdyZXc= 6029
+b3Npbmc= 6030
+IGRpdmVycw== 6031
+QUQ= 6032
+IGVtcHQ= 6033
+IGV4cHJlc3Npb24= 6034
+4bq/ 6035
+IFBhbA== 6036
+IGp1c3RpY2U= 6037
+IHBhaXI= 6038
+d28= 6039
+IHNlYXQ= 6040
+b3J0ZXI= 6041
+IGxpbmtz 6042
+IE1lcg== 6043
+IHJlbmQ= 6044
+0L3QvtC1 6045
+dXBpZA== 6046
+IEhlbA== 6047
+IE1hcmNo 6048
+IExv 6049
+0YHRjA== 6050
+IGhhc24= 6051
+IGV2YWx1 6052
+aWxvcw== 6053
+IGZ1bmRpbmc= 6054
+IHZlbg== 6055
+dWFu 6056
+IE1hc3Rlcg== 6057
+IE9s 6058
+IEZyZQ== 6059
+IHlhcA== 6060
+IFNpcg== 6061
+c2No 6062
+IG1pc3Rha2U= 6063
+YW1hbg== 6064
+IGRpbm5lcg== 6065
+IFdhc2hpbmd0b24= 6066
+IG9yZ2FuaXphdGlvbnM= 6067
+INC20LU= 6068
+YXZpbmc= 6069
+IHbDrQ== 6070
+IGJpcnRoZGF5 6071
+IGJlYXI= 6072
+INmB 6073
+IGFmZm9yZA== 6074
+IHJldmVu 6075
+IHJlbGF0aW9uc2hpcHM= 6076
+cm91Z2g= 6077
+IFRpbWU= 6078
+IHRhZw== 6079
+IFN1bg== 6080
+dWFyeQ== 6081
+IFBv 6082
+Y2Fy 6083
+YWJpbGl0aWVz 6084
+IHByaXNvbg== 6085
+IGxpYw== 6086
+7KCV 6087
+aWRkZW4= 6088
+IHNwZWNpZXM= 6089
+6bs= 6090
+IGZpcm0= 6091
+IHNjb3Jl 6092
+IGRpdA== 6093
+IHNwZWN0 6094
+IHBlbA== 6095
+IGNvbXBsaWNhdGVk 6096
+IHJhbms= 6097
+IG9wcG9zaXRl 6098
+IHBpY2tlZA== 6099
+INC60L7QvQ== 6100
+ZWxlcg== 6101
+IG1pZw== 6102
+IFNs 6103
+IE5ldA== 6104
+IG5lY2s= 6105
+IEZyYW5jZQ== 6106
+IHRlY2huaWNhbA== 6107
+4Lih 6108
+IG1pbGVz 6109
+IHByaW1hcnk= 6110
+IHNlaW4= 6111
+c2Vz 6112
+IGxhdWdocw== 6113
+YnJh 6114
+xZtjaQ== 6115
+cmlhZ2U= 6116
+IG5pYw== 6117
+ZXRlcnM= 6118
+IMOq 6119
+b2xvZ2llcw== 6120
+IElT 6121
+cmFk 6122
+dWRv 6123
+xLFuZA== 6124
+bWFy 6125
+IGV4Y2g= 6126
+IGNvbXBldGl0aW9u 6127
+IGF1c3Np 6128
+IFNlcnY= 6129
+IHJlbnQ= 6130
+IGNob2NvbGF0ZQ== 6131
+IHdpZWRlcg== 6132
+IG5lYXJseQ== 6133
+IHNwZWVjaA== 6134
+IHVuYw== 6135
+IHBhcmFt 6136
+IEJyaXRpc2g= 6137
+IHJlbWFpbg== 6138
+4LiB 6139
+dXJ0 6140
+INi5 6141
+IGNyYWNr 6142
+YWlscw== 6143
+IHByb21pc2U= 6144
+IHBheWluZw== 6145
+acOf 6146
+IGFkYXB0 6147
+0LDQu9Cw 6148
+IG1vdmllcw== 6149
+IHdpcmU= 6150
+n6w= 6151
+IHRlcnJpYmxl 6152
+IHPDsw== 6153
+IHBlcmZlY3RseQ== 6154
+b3JkaW4= 6155
+IGrDoQ== 6156
+IGltcG9zc2libGU= 6157
+IFRocmVl 6158
+IG5o 6159
+IHR1cm5pbmc= 6160
+cnVt 6161
+IEJlbA== 6162
+aWdn 6163
+IHJlc3BvbnNpYmxl 6164
+0LjQuQ== 6165
+IGluY3JlZGlibHk= 6166
+d2k= 6167
+aWFubw== 6168
+IGh1bWFucw== 6169
+IMOH 6170
+IHNldHRpbmdz 6171
+IGpveQ== 6172
+b290 6173
+IGRlYWxpbmc= 6174
+aWxsZWQ= 6175
+IHN1cnJvdW5k 6176
+IGZvbGxvd2Vk 6177
+IHBvc3NpYmx5 6178
+IGluaXRp 6179
+c3Rlbg== 6180
+IHByb3M= 6181
+IGNhbmRpZA== 6182
+IGFzc2lnbg== 6183
+IHZpb2xlbmNl 6184
+V2VsbA== 6185
+IHJpc2U= 6186
+UFM= 6187
+IHRhbWLDqW0= 6188
+IOuTpA== 6189
+aWFuY2U= 6190
+eWFu 6191
+IGF1ZGlv 6192
+IEJldA== 6193
+IEFtZXJpY2Fucw== 6194
+IEFzcw== 6195
+aXNjaGVu 6196
+7J6F 6197
+IHVsdGltYXRlbHk= 6198
+IHBvbGlj 6199
+IG1ham9yaXR5 6200
+IEZpbmFsbHk= 6201
+ZXJhcA== 6202
+IGd1YXJk 6203
+IE1BVFQ= 6204
+IGJyb3du 6205
+0LzQuA== 6206
+IGNoYQ== 6207
+IEhvbHk= 6208
+IG5lcnZvdXM= 6209
+aXBwaW5n 6210
+xJlk 6211
+IFNh 6212
+k5zr 6213
+toA= 6214
+bGll 6215
+IG51Yw== 6216
+IEFwcg== 6217
+6Zs= 6218
+IEtvcmVh 6219
+ZWdv 6220
+IENhbmFkYQ== 6221
+IGvDtm5uZW4= 6222
+IGNvbXBhcg== 6223
+IGdhbno= 6224
+IE1haXM= 6225
+IHRoZW1l 6226
+IGtp 6227
+IGRyYXdpbmc= 6228
+YXpvbg== 6229
+IE9mZg== 6230
+dHQ= 6231
+IFdpbmQ= 6232
+IHRvZG9z 6233
+IG9idmlvdXM= 6234
+0L3QsNGP 6235
+SU0= 6236
+INCg 6237
+d2VsbA== 6238
+IGJsb3c= 6239
+IGhvb2s= 6240
+IGNpcmNsZQ== 6241
+IOuztA== 6242
+IGFyY2hpdGVjdA== 6243
+IEty 6244
+IGPDsw== 6245
+IHByb3RlY3Rpb24= 6246
+ZWdh 6247
+5Yc= 6248
+IHdhdGNoZWQ= 6249
+IGFuc3dlcnM= 6250
+IGRpZXQ= 6251
+aXZv 6252
+IHBvd2Rlcg== 6253
+IHlvdXJz 6254
+IGhpZ2hlc3Q= 6255
+RkY= 6256
+5bo= 6257
+IGJveXM= 6258
+w7Z5bGU= 6259
+IGx1bmNo 6260
+IElJ 6261
+IHNldHM= 6262
+IG1vbGU= 6263
+24E= 6264
+IHdpbnRlcg== 6265
+IGx1Y2t5 6266
+IHJlc3BvbnNpYmlsaXR5 6267
+IHNpZ25hbA== 6268
+IHdvbmRlcmluZw== 6269
+IGF4 6270
+IGNvb2tpbmc= 6271
+0L7QstC+0YA= 6272
+bGVn 6273
+INC/0L7Rgg== 6274
+IHN1cnByaXNl 6275
+IGRlbW9jcg== 6276
+IGxvb3A= 6277
+IGphZw== 6278
+IGN1cmlvdXM= 6279
+IG1hcmtldGluZw== 6280
+0J0= 6281
+YXJvbg== 6282
+IEFwcGxl 6283
+IHZpcnR1YWw= 6284
+IDE5OA== 6285
+bm9vbg== 6286
+IE1ldA== 6287
+0L7RgdGC0L4= 6288
+0L7QsdGL 6289
+aXR1 6290
+IEF3 6291
+IGJ1eWluZw== 6292
+IHJlc3RhdXJhbnQ= 6293
+IEJ1ZA== 6294
+IGRvdWJ0 6295
+IGdyYW50 6296
+IHZlcmQ= 6297
+IGNhc2g= 6298
+IGZhY3VsdHk= 6299
+VGhhdA== 6300
+IEVpbg== 6301
+IHdlZA== 6302
+aXRuZXNz 6303
+IE1hZw== 6304
+bmVs 6305
+IG5hcnI= 6306
+IGFjY2lkZW50 6307
+IG1lZGl1bQ== 6308
+ZW1lbnRz 6309
+IGNyb3c= 6310
+bmlnaHQ= 6311
+7J28 6312
+IGxpYnJhcnk= 6313
+0LDRjtGC 6314
+IHRhbWJpw6lu 6315
+IHJlZmVyZW5jZQ== 6316
+IGZvdXJ0aA== 6317
+aG91c2U= 6318
+dmVudGlvbg== 6319
+IGZpbGxlZA== 6320
+IENvdXI= 6321
+aWJy 6322
+IG5n 6323
+IGRldmVsb3Bpbmc= 6324
+IHByb3ZpZGVz 6325
+IHBvbGw= 6326
+IHRyYWZmaWM= 6327
+YXJlbnRseQ== 6328
+4K6f 6329
+IGZvcm1z 6330
+IGNsaWVudA== 6331
+IGdlbnRsZQ== 6332
+IG11c3M= 6333
+IENvbmdyZXNz 6334
+IEluZGlhbg== 6335
+Y2Vhbg== 6336
+IHBpbA== 6337
+IGN6eQ== 6338
+c3Rvb2Q= 6339
+dXR5 6340
+IG7DpA== 6341
+IHNwZW5kaW5n 6342
+IGNvbnN0cnVjdGlvbg== 6343
+aW5hdWRpYmxl 6344
+IOuniA== 6345
+iOustA== 6346
+IOyDnQ== 6347
+b21h 6348
+b3Nlbg== 6349
+YWdv 6350
+IGxhcmdlc3Q= 6351
+44WL44WL 6352
+IHVuaXZlcnNl 6353
+YmVz 6354
+b3Nh 6355
+INC10LPQvg== 6356
+IGR1ZGU= 6357
+IE1BUg== 6358
+IGluZGVlZA== 6359
+zrXOuQ== 6360
+IG1hbmFnZWQ= 6361
+IFNob3VsZA== 6362
+U28= 6363
+IGFwcGxpZWQ= 6364
+IGZhaXJseQ== 6365
+IERlbg== 6366
+IGFuYWx5 6367
+IGNvbnN0YW50bHk= 6368
+0YHQvw== 6369
+SG93 6370
+IFNheQ== 6371
+ZW5jaWVz 6372
+IFBD 6373
+IGVnZ3M= 6374
+4K6w 6375
+IGV0aA== 6376
+IEVudMOjbw== 6377
+aW5hcg== 6378
+aW90 6379
+IGN6 6380
+IEV1cm9wZWFu 6381
+IEFN 6382
+IGPDoQ== 6383
+IHJhZGlv 6384
+p4w= 6385
+IGhpZGU= 6386
+IFN0YXJ0 6387
+IGNsdWI= 6388
+IEhvcGU= 6389
+IGVmZm9ydHM= 6390
+bHVzaW9u 6391
+IGNpdGllcw== 6392
+aG9uZQ== 6393
+IHJlYWNoZWQ= 6394
+IGd1aWQ= 6395
+cm9pZA== 6396
+IGhhcm0= 6397
+IGN1dHRpbmc= 6398
+IGJ1bA== 6399
+MTg= 6400
+aWVzdA== 6401
+IE1leA== 6402
+IGlyb24= 6403
+IGFmdGVybm9vbg== 6404
+IGhhbGw= 6405
+IHByenk= 6406
+IGdvc2g= 6407
+IGluZmx1ZW5jZQ== 6408
+INCy0LjQtA== 6409
+IGluY3JlYXNlZA== 6410
+IE1pbmlzdGVy 6411
+IGRpc2Np 6412
+IFBldGVy 6413
+IHZlcnQ= 6414
+IG1lbnU= 6415
+IHNlbGxpbmc= 6416
+dXJhbGx5 6417
+IHF1b3Rl 6418
+IMKh 6419
+IGNvbnRpbnVlcw== 6420
+bXByZQ== 6421
+IMWfZXk= 6422
+aXR1dGlvbg== 6423
+INC90LDRgQ== 6424
+Y2xlcw== 6425
+IEdlcm1hbg== 6426
+Y3p5 6427
+INCj 6428
+QmU= 6429
+IGtpdGNoZW4= 6430
+IFRyeQ== 6431
+aXBl 6432
+IGljb24= 6433
+YXJw 6434
+IHByb3ZpZGluZw== 6435
+IFRyYW5z 6436
+IHRlY2huaXF1ZQ== 6437
+IGjDpHI= 6438
+IGluZnJhc3Q= 6439
+IHN1c3A= 6440
+w7xjaw== 6441
+aWNpcA== 6442
+INCV 6443
+IGNpbg== 6444
+7Ja06w== 6445
+IHByeg== 6446
+IGNvbXBvbmVudA== 6447
+IGJ5ZQ== 6448
+IEJpYmxl 6449
+aXplcg== 6450
+Q2g= 6451
+IHNvbHV0aW9ucw== 6452
+IGFjY29tcGw= 6453
+IDIwMTY= 6454
+SUU= 6455
+IFRh 6456
+IGFzc3VtZQ== 6457
+IGxpcXVpZA== 6458
+IOuouQ== 6459
+IHF1YXJ0ZXI= 6460
+IGZlbWFsZQ== 6461
+IFRoaW5r 6462
+IHN0YXR1cw== 6463
+aXR1dGU= 6464
+IGNvYWNo 6465
+IHJlaW4= 6466
+IGNvbWJpbmF0aW9u 6467
+6Lc= 6468
+IFRlcg== 6469
+IG9iamVjdHM= 6470
+IGRpc3RyaWN0 6471
+IG1ha2V1cA== 6472
+IG11cmRlcg== 6473
+d2Fz 6474
+ZmVu 6475
+IGJvd2w= 6476
+IHB1Ymxpc2hlZA== 6477
+IHNwb3J0cw== 6478
+IGlkZW50aXR5 6479
+IHNlZW1lZA== 6480
+IGFjdGluZw== 6481
+0LvRjg== 6482
+cml4 6483
+IHVwbG9hZA== 6484
+IGhhc3Q= 6485
+IGJvYXQ= 6486
+IE1vZA== 6487
+cmlv 6488
+ID0= 6489
+IGN5Y2xl 6490
+r7g= 6491
+IGxvdWQ= 6492
+dXN0ZWQ= 6493
+Y29taW5n 6494
+IDIwMTc= 6495
+IG9udA== 6496
+IGxlZ2lzbA== 6497
+IHN0cnVjdA== 6498
+IFNvbWV0aGluZw== 6499
+IGNvbmZsaWN0 6500
+IHVwcGVy 6501
+IG1hbmFnZXI= 6502
+IG1vcnQ= 6503
+IGZyYQ== 6504
+IMSw 6505
+IE1pa2U= 6506
+IFdvcms= 6507
+IG7Dsw== 6508
+cGhlcmU= 6509
+IOyCrOs= 6510
+IExhbmQ= 6511
+IGZpbHRlcg== 6512
+IHByb21vdA== 6513
+5rA= 6514
+lbw= 6515
+IHJlY29yZGluZw== 6516
+150= 6517
+IGFzc29jaWF0ZWQ= 6518
+IGZ1ZWw= 6519
+dW5kZXI= 6520
+IGVsZWN0aW9u 6521
+IGVtcGxveWVlcw== 6522
+IENvbXA= 6523
+0YDRg9Cz 6524
+IFdv 6525
+cm9s 6526
+IHNhdmVk 6527
+IEhvbg== 6528
+IFZp 6529
+YWNh 6530
+cHJldA== 6531
+IHdldA== 6532
+IHN0dXBpZA== 6533
+IGxhZA== 6534
+IGZlc3Q= 6535
+IHdha2U= 6536
+INC40L0= 6537
+IGdyZWF0ZXN0 6538
+IEppbQ== 6539
+IHNlcmlvdXNseQ== 6540
+IOy5 6541
+IGZlZWxpbmdz 6542
+IDMwMA== 6543
+aWF0aW9u 6544
+IGJlYXV0eQ== 6545
+IOyemA== 6546
+IHNhbg== 6547
+k6A= 6548
+IC0o 6549
+IGNvbnNjaW91cw== 6550
+INC00LXQuw== 6551
+Ynll 6552
+55k= 6553
+TWFu 6554
+IGxldHM= 6555
+IHNob2Vz 6556
+eWQ= 6557
+IGRpc2FwcGU= 6558
+IENvdW50eQ== 6559
+IFNjb3R0 6560
+IGJ1dHQ= 6561
+IGFxdcOt 6562
+IGNvbmZpZw== 6563
+cmVzcG9uZA== 6564
+TEFVR0g= 6565
+qeuLiOuLpA== 6566
+IGRpdmlkZWQ= 6567
+IGFjcXU= 6568
+IHpvbmU= 6569
+IGtvbW0= 6570
+YcOnw6Nv 6571
+7Kec 6572
+Y3V0 6573
+IDIz 6574
+IG1heGltdW0= 6575
+cm9n 6576
+IHJ1bnM= 6577
+IGNvbXBvbmVudHM= 6578
+IGFycml2ZWQ= 6579
+IGNvbmZpZGVudA== 6580
+0YDQvtCy 6581
+IGhlaWdodA== 6582
+IHByb2NlZA== 6583
+RU0= 6584
+INCt0YLQvg== 6585
+IE1lbg== 6586
+IHRhbGtz 6587
+IGNvbmZpZGVuY2U= 6588
+IENocmlz 6589
+IGxlYWRz 6590
+IG5vc2U= 6591
+ZmFsbA== 6592
+YmI= 6593
+IE5vdGhpbmc= 6594
+aXNlcg== 6595
+IGluZGVwZW5kZW50 6596
+IG1pbm9y 6597
+IHN5bQ== 6598
+bGVu 6599
+Y2llbmNl 6600
+IGZhc2hpb24= 6601
+IHNleHVhbA== 6602
+IGJ1bg== 6603
+aGVyZQ== 6604
+IHNvaWw= 6605
+IGRpZXNl 6606
+IHNoYXA= 6607
+IGVtcHR5 6608
+IGpvdXJuYWw= 6609
+YWdvbg== 6610
+IFRoZWly 6611
+IHdlZWtlbmQ= 6612
+w610 6613
+IGVycm9y 6614
+IG5hcg== 6615
+w7g= 6616
+6Kk= 6617
+YW5jeQ== 6618
+IOyVig== 6619
+IGZvcmVzdA== 6620
+IGhhY2Vy 6621
+IG1pc3NlZA== 6622
+IGV2aWw= 6623
+IHN0b3JhZ2U= 6624
+IHNpbmdpbmc= 6625
+aW5oYQ== 6626
+IGtub2Nr 6627
+IGltcHJlc3M= 6628
+INC+0YfQtdC90Yw= 6629
+IEdvbGQ= 6630
+IFN1cg== 6631
+IFBvcnQ= 6632
+IExvbmQ= 6633
+IGZhemVy 6634
+b3R5 6635
+b3Rv 6636
+IGFueA== 6637
+IFdpbGxpYW0= 6638
+IGV4aXN0aW5n 6639
+cGxhY2U= 6640
+IENE 6641
+zrM= 6642
+IENvbGxlZ2U= 6643
+bG9y 6644
+IEVhc3Q= 6645
+c2Vu 6646
+ZmFjaA== 6647
+b2Z0 6648
+IGV4cGVyaWVuY2Vk 6649
+IGxvdmVz 6650
+aW1t 6651
+IHBvbHk= 6652
+IGVzc2U= 6653
+7KQ= 6654
+IEdyYW5k 6655
+6Kc= 6656
+Y2hlcg== 6657
+IHZpY3RpbQ== 6658
+IEdlcw== 6659
+0LvRjA== 6660
+dmlzaW9u 6661
+IHRhbGw= 6662
+IGxlbnM= 6663
+INC30L3QsA== 6664
+IEJvdGg= 6665
+IOyy 6666
+IHN1c3RhaW4= 6667
+IGFyZ3VtZW50 6668
+IGZhY3RvcnM= 6669
+IGF1dG9tYXRpY2FsbHk= 6670
+IGZydWl0 6671
+IGxpYmVy 6672
+IGFsZQ== 6673
+IFByZXNz 6674
+IEJh 6675
+INCz0L4= 6676
+IGh1bmRyZWRz 6677
+dGhhdA== 6678
+IFJpY2g= 6679
+IHJlY2lwZQ== 6680
+IElU 6681
+6Ic= 6682
+4bql 6683
+IGRlc2NyaWJl 6684
+IGRyaXZlcg== 6685
+IE9jdA== 6686
+IE1hdA== 6687
+0LTQtQ== 6688
+IG1lYWw= 6689
+IGxhdGVzdA== 6690
+IHRoZXJhcA== 6691
+IGNvbXBhcmU= 6692
+IEFtYXpvbg== 6693
+IOyigA== 6694
+IFJ1c3NpYQ== 6695
+IHN0cmluZw== 6696
+IGth 6697
+IENvbW11bg== 6698
+IGRpYQ== 6699
+SXM= 6700
+IG1pbGxpb25z 6701
+IGNvcnBvcg== 6702
+IGNvcnJlc3BvbmQ= 6703
+IGZpeGVk 6704
+IEpvZQ== 6705
+2Y4= 6706
+IHZpZXdz 6707
+IHJpdmVy 6708
+IHN0dWRpbw== 6709
+aWdnZXI= 6710
+IGZsYXZvcg== 6711
+IHByZXNlbmNl 6712
+IHVuaXRz 6713
+IHNhdmluZw== 6714
+YXZvdXI= 6715
+IHBlc3Nv 6716
+b3JpdGg= 6717
+IGhlcnM= 6718
+IE5hdA== 6719
+YXNpb24= 6720
+IEZyYW5r 6721
+0L7RiA== 6722
+xYJ5 6723
+7YQ= 6724
+IGVpbmVt 6725
+IGZ1bmN0aW9ucw== 6726
+dW1hbg== 6727
+IG5vcnRo 6728
+IOyghA== 6729
+IGhvcnNl 6730
+dmlk 6731
+IHBsZWFzdXJl 6732
+0LDRiA== 6733
+w6llcw== 6734
+aW5kYQ== 6735
+IHRhaWw= 6736
+IGV4cGxvcmU= 6737
+U1Q= 6738
+IGNvbW1lcmNpYWw= 6739
+IER1cmluZw== 6740
+YXJs 6741
+XTo= 6742
+Zml0 6743
+IHJhdGVz 6744
+5rM= 6745
+TVVTSUM= 6746
+IGhvdXNpbmc= 6747
+IGVpbmVy 6748
+IHNpdHVhdGlvbnM= 6749
+5os= 6750
+IGRlY3Jl 6751
+IGFwcHJvcHJpYXRl 6752
+0LXQvdC90L4= 6753
+JS4= 6754
+IGJhYw== 6755
+IHdhdA== 6756
+ZW5zaXR5 6757
+w6Ro 6758
+a25vd24= 6759
+aXR6 6760
+IGVtb3Rpb25hbA== 6761
+ZXJ2YXRpb24= 6762
+IGJsaW5k 6763
+MTY= 6764
+7YM= 6765
+IGpvaW5lZA== 6766
+IGxvY2F0ZWQ= 6767
+INGB0Lw= 6768
+YWRhcw== 6769
+YmVyZw== 6770
+IGRlc3M= 6771
+IGRlYXI= 6772
+ZWRlbg== 6773
+Y29z 6774
+IGFkb3B0 6775
+MTAw 6776
+b3dl 6777
+IENoZWNr 6778
+aXNtbw== 6779
+IHNpbXBs 6780
+IGFuZ3J5 6781
+INC80LXQvdGP 6782
+IENhbQ== 6783
+IHBhZA== 6784
+IGF0dGVuZA== 6785
+IHNhbXBsZQ== 6786
+IOyb 6787
+IElO 6788
+dWxvdXM= 6789
+IFNhcg== 6790
+IFNob3c= 6791
+IGluZnJhc3RydWN0dXJl 6792
+IEF1Z3VzdA== 6793
+IGxlc3Nvbg== 6794
+IG5pZXQ= 6795
+5o4= 6796
+IGZvaQ== 6797
+IGJyb2tl 6798
+dHI= 6799
+55U= 6800
+IDQ1 6801
+IGdldw== 6802
+0YPQvw== 6803
+YXRp 6804
+IG1haW50YWlu 6805
+IGFydGlzdHM= 6806
+aW5nZXI= 6807
+ZXJ2ZWQ= 6808
+SUE= 6809
+IGVxdWFscw== 6810
+IG9wZXJhdGlvbg== 6811
+aWxseQ== 6812
+IOuCtA== 6813
+IGNyb3dk 6814
+IGludGVybmFs 6815
+IHRlc3Rz 6816
+IFJvY2s= 6817
+IENvbnM= 6818
+IOuEiOustA== 6819
+d2Fy 6820
+IHNvdQ== 6821
+IGNoYXJ0 6822
+IEp1bmU= 6823
+IEFwcmls 6824
+Z2VudA== 6825
+IHZlbnQ= 6826
+IHF1YW5k 6827
+IEtvcmVhbg== 6828
+aW1v 6829
+54k= 6830
+aWRlcnM= 6831
+IG1vdW50YWlu 6832
+0YHRgtCw0LI= 6833
+aWpr 6834
+IGRpc2NvdmVyZWQ= 6835
+IFN1bmQ= 6836
+IFNpbA== 6837
+IHNvbG8= 6838
+wrQ= 6839
+IHNjaG9s 6840
+IEVhY2g= 6841
+57U= 6842
+IGJhcmU= 6843
+IO2M 6844
+IHbDrWRl 6845
+IGluZ3JlZGllbnRz 6846
+IEl0cw== 6847
+nbzqs6A= 6848
+IOyK 6849
+z40= 6850
+IExlZQ== 6851
+IHNjYXJ5 6852
+IHByaW5jaXA= 6853
+IHNwaXJpdHVhbA== 6854
+7IU= 6855
+IEhvbGQ= 6856
+IGRlZmluZQ== 6857
+IExlcw== 6858
+IE5vcg== 6859
+IEVuZA== 6860
+IGJsb2c= 6861
+IEdyZWVu 6862
+0LDQtdGC0YHRjw== 6863
+cGFydA== 6864
+ZWxlcw== 6865
+IFVuZGVy 6866
+IHBhcnRl 6867
+IDM1 6868
+IHNlY3Rvcg== 6869
+IFNlcHQ= 6870
+IGF1dGg= 6871
+4K6u 6872
+b21pbg== 6873
+IGNsaWVudHM= 6874
+IGNp 6875
+IEZyaWRheQ== 6876
+ZXJhcw== 6877
+IHR3ZQ== 6878
+dWxhdGVk 6879
+IGN1bHR1cmFs 6880
+INGB0LLQvg== 6881
+IOuNlA== 6882
+IMO6 6883
+IHBhcmNl 6884
+4K6y 6885
+IHRyYWRpdGlvbg== 6886
+IGp1ZGdl 6887
+IEdlbmVyYWw= 6888
+IGRldGVybWluZQ== 6889
+IElzbg== 6890
+IFBM 6891
+bmVhdGg= 6892
+IG1hdHRlcnM= 6893
+7ZW07A== 6894
+IV0= 6895
+0LDRhQ== 6896
+IHBvb2w= 6897
+IHZhcmlhYmxl 6898
+IHZhY2NpbmU= 6899
+IGNhdXNlZA== 6900
+IHdlc3Q= 6901
+IFllcA== 6902
+ZmFzdA== 6903
+IHBoaWxvcw== 6904
+aG9yYQ== 6905
+IGNvbnRpbnVlZA== 6906
+IHVuZm9ydHVuYXRlbHk= 6907
+5pU= 6908
+IGZsaWdodA== 6909
+IHdyYXA= 6910
+IGh1aA== 6911
+IEFic29sdXRlbHk= 6912
+IHBpbms= 6913
+IHJlbWFpbnM= 6914
+IG7DqQ== 6915
+IGZsZQ== 6916
+IFNvbA== 6917
+IGxvc2luZw== 6918
+IGFsZ29yaXRo 6919
+IHJlcXVpcmVz 6920
+IGZvdW5kYXRpb24= 6921
+IEJ1cg== 6922
+IHByb2Zlc3Npb24= 6923
+IE1pZA== 6924
+IOutkA== 6925
+Y2Fu 6926
+IE1pbA== 6927
+IHlvdW5nZXI= 6928
+IGFwcGVhcnM= 6929
+dGVybQ== 6930
+7ZWY6rOg 6931
+YWNsZQ== 6932
+IExvbmRvbg== 6933
+IGVuZ2luZWVyaW5n 6934
+4Lii 6935
+IGFkdmVudA== 6936
+7IS47JqU 6937
+IOq4sA== 6938
+IE1hag== 6939
+0YDQtdC8 6940
+aW5ndQ== 6941
+IFVL 6942
+dXJv 6943
+c3Bl 6944
+IHRlbnQ= 6945
+IHJlcG9ydGVk 6946
+IEFM 6947
+SGV5 6948
+IOunkA== 6949
+IGRlbnQ= 6950
+IEF1c3RyYWxpYQ== 6951
+IEphbnVhcnk= 6952
+s7Q= 6953
+YWd1ZXM= 6954
+YXJzaA== 6955
+cmln 6956
+IHRpZW5l 6957
+4Lij 6958
+zq4= 6959
+IG1hY2hlbg== 6960
+dW50ZQ== 6961
+0YPRgQ== 6962
+IGVsZWN0cg== 6963
+IHR1dG9yaWFs 6964
+IHBsYWNlZA== 6965
+IOydtOqxsA== 6966
+IENvdW5jaWw= 6967
+7ZaI 6968
+sOumrA== 6969
+YWhyZW4= 6970
+IOq3uOuemA== 6971
+IHByb3Zl 6972
+Zm9s 6973
+IHF1ZXI= 6974
+IGNoZWFw 6975
+IEZhdGhlcg== 6976
+IFBvd2Vy 6977
+k5w= 6978
+IHB1cnM= 6979
+IGVzcA== 6980
+IEJyZQ== 6981
+6riw6w== 6982
+b21hcw== 6983
+0LjQu9GM 6984
+IGdlaHQ= 6985
+b3N0ZXI= 6986
+6rO8 6987
+IGZpbGVz 6988
+INCn 6989
+YmVsbA== 6990
+IHdob20= 6991
+IOuY 6992
+IGV4Y2VsbGVudA== 6993
+IGRhdGFi 6994
+IGfDtg== 6995
+IOynhOynnA== 6996
+IGJlbGllZg== 6997
+amV0 6998
+IGphY2s= 6999
+IHN3aW0= 7000
+cmlhbA== 7001
+dW1pbg== 7002
+YXVj 7003
+IHNvbGw= 7004
+IGVzc2VudGlhbA== 7005
+7ZWY64qU 7006
+IGV2b2w= 7007
+Y2hhZnQ= 7008
+YWluZQ== 7009
+dGhsZXQ= 7010
+IGluY29y 7011
+IHJlcG9ydHM= 7012
+IGRlZmluaXRpb24= 7013
+a2Vs 7014
+IGNpcmN1bQ== 7015
+IHByb2R1Y2Vk 7016
+INeb 7017
+YW50aWM= 7018
+bmV0 7019
+IGF3YXJk 7020
+IGR1cmNo 7021
+IHRyYW5zcA== 7022
+IG1hbGU= 7023
+pqzr 7024
+IG1vb24= 7025
+IEdlb3JnZQ== 7026
+IGZseWluZw== 7027
+acOz 7028
+IHNvdXJjZXM= 7029
+IHBsZW50eQ== 7030
+IERlbW9jcg== 7031
+Uk8= 7032
+IDAw 7033
+IHNlY3VyZQ== 7034
+IEJpcg== 7035
+cmFpbg== 7036
+IHp1cg== 7037
+IGVmZmljaWVudA== 7038
+IHJlcGVhdA== 7039
+IG1ldGhvZHM= 7040
+IGNhbG0= 7041
+IGRpc2N1c3NlZA== 7042
+IOyeiOuKlA== 7043
+IHNlcnZlcg== 7044
+YW5pZQ== 7045
+IEluc3RlYWQ= 7046
+IGlkZWFs 7047
+IGNvbnZlbg== 7048
+IGhvcGluZw== 7049
+IFRvcg== 7050
+IGRlcHRo 7051
+IGhlYXZlbg== 7052
+RU5DRQ== 7053
+IGhhYml0 7054
+Z3JhZA== 7055
+IGZsYWc= 7056
+IGluZQ== 7057
+IGto 7058
+IExJ 7059
+IGZhY2luZw== 7060
+IEFV 7061
+IFRpbQ== 7062
+IGdlbQ== 7063
+IEp1bA== 7064
+IGVsYQ== 7065
+aXp6YQ== 7066
+IGZlbGxvdw== 7067
+IHF1ZWw= 7068
+IHNwb2tl 7069
+IGNpdGl6ZW5z 7070
+dWdl 7071
+IHBhZ2Vz 7072
+IGZhc2M= 7073
+IHJlbGlnaW91cw== 7074
+YXRlbg== 7075
+IGNoYXB0ZXI= 7076
+IFZhbA== 7077
+IGNvbnN1bHQ= 7078
+IE1pbGw= 7079
+Z2w= 7080
+b3Blcg== 7081
+IGluZmlu 7082
+IG1hcnJpYWdl 7083
+IG1lZGljaW5l 7084
+INC00LI= 7085
+IGRvZ3M= 7086
+IGluc3RydW1lbnQ= 7087
+IEV4YWN0 7088
+w6Fu 7089
+IDIwMjE= 7090
+IGZlcg== 7091
+IHdlYWx0aA== 7092
+IGdyYWRl 7093
+0YvRhQ== 7094
+IGNyaW1l 7095
+IHRocmVhZA== 7096
+IGVzc2E= 7097
+IHdpbmU= 7098
+Y29ob2w= 7099
+cGhh 7100
+4LiH 7101
+b2d1ZQ== 7102
+IGluc3VyYW5jZQ== 7103
+YXJyYXRvcg== 7104
+IFNlcHRlbWJlcg== 7105
+IHZpZA== 7106
+IFNwaXJpdA== 7107
+IGdlc3Q= 7108
+IFJ1c3NpYW4= 7109
+IHByb3BlcnRpZXM= 7110
+IGFydGljbGU= 7111
+IHVuZGVybmVhdGg= 7112
+eWVy 7113
+IGpvaW50 7114
+IHJlbGF0aXZlbHk= 7115
+IGluY2g= 7116
+IGRlc3BpdGU= 7117
+IEdyZWU= 7118
+IGNsYXNzaWM= 7119
+IHN1cHBvcnRpbmc= 7120
+IGluc3RydWN0 7121
+bHVzaXZl 7122
+IGRpYWdu 7123
+5oo= 7124
+IGFkbWluaXN0cmF0aW9u 7125
+0LDQsdC+0YI= 7126
+IE9wZW4= 7127
+INC/0L7Qug== 7128
+IGRvbGxhcg== 7129
+IGNvbnNlcXU= 7130
+b2Jlcg== 7131
+IEdlcm1hbnk= 7132
+IHRlcnI= 7133
+IFFV 7134
+INCT 7135
+574= 7136
+IHN0cm9uZ2Vy 7137
+yZk= 7138
+INmK 7139
+IGlQaG9uZQ== 7140
+IGZhYnJpYw== 7141
+w7xo 7142
+IGVuZW0= 7143
+5q8= 7144
+IHN1YnQ= 7145
+RUU= 7146
+b25kZQ== 7147
+IGNyZXc= 7148
+IHJlbW92ZWQ= 7149
+IGxhZHk= 7150
+IHBvdGVudGlhbGx5 7151
+INCd0L4= 7152
+eWFs 7153
+IHN5bXB0 7154
+IGFybXk= 7155
+IGludHJvZHVjZWQ= 7156
+dGVz 7157
+IGFzcGVjdHM= 7158
+MTQ= 7159
+IExvdQ== 7160
+ICk= 7161
+IGRlcGxveQ== 7162
+cGV0 7163
+IGhhbg== 7164
+IFdhdGNo 7165
+IHdlYXBvbnM= 7166
+IHBoZW4= 7167
+IHJlZ2lzdGVy 7168
+IGVpbmZhY2g= 7169
+IHNwb3J0 7170
+IGJyaWRnZQ== 7171
+IGlubmVy 7172
+IG1pbmltdW0= 7173
+IHdpdG5lc3M= 7174
+IGVzbw== 7175
+IHZpbGxhZ2U= 7176
+IG93bmVy 7177
+pqzqs6A= 7178
+IHNjcmVhbQ== 7179
+aWxlZA== 7180
+IHBpdGNo 7181
+YnJ1 7182
+IGFkdmFuY2U= 7183
+IHN1cHBvc2U= 7184
+IEF0dA== 7185
+0LXRgtGB0Y8= 7186
+IGRpZmZlcmVuY2Vz 7187
+YWtlZA== 7188
+IGludGVycHJldA== 7189
+w6Y= 7190
+aWVuZG8= 7191
+IGFic29s 7192
+INCx0YPQtNC10YI= 7193
+IOuy 7194
+IHRyaWFs 7195
+IHRoaW5rcw== 7196
+bHlpbmc= 7197
+Y2VwdGlvbg== 7198
+IEFmcmljYW4= 7199
+IGNoZW1pY2Fs 7200
+IHRhcGU= 7201
+IGNvbnZlcnNhdGlvbnM= 7202
+IGRpc3RyaWJ1dGlvbg== 7203
+dGk= 7204
+IEFJ 7205
+IGZsYXNo 7206
+IHVuZGVyc3Rvb2Q= 7207
+IEdvdmVybm1lbnQ= 7208
+IT8= 7209
+IFNr 7210
+6rGw6w== 7211
+cmllcg== 7212
+VFM= 7213
+IEFjY29yZGluZw== 7214
+0Y7Rgg== 7215
+IHNwb25z 7216
+0YLQvtCx0Ys= 7217
+IHZhbHU= 7218
+ZXJlbQ== 7219
+aWNodGln 7220
+IHJlc2lzdGFuY2U= 7221
+IEdhbA== 7222
+Z2VyeQ== 7223
+IGJlZ2lucw== 7224
+IGFkdmFuY2Vk 7225
+IHJlbGV2YW50 7226
+IHBvbGl0aWNz 7227
+IEZhbQ== 7228
+IMOnb2s= 7229
+IE5ldmVy 7230
+aWxsaW5n 7231
+IGZvb3RiYWxs 7232
+0LjQuA== 7233
+IElE 7234
+IEFmcmljYQ== 7235
+IGZpbmdlcnM= 7236
+INCx0L7Qu9GM 7237
+IMOh 7238
+IGNsaXA= 7239
+IExhdA== 7240
+IOyngOq4iA== 7241
+ZXNzZQ== 7242
+IHZvb3I= 7243
+IGFzaWRl 7244
+5p4= 7245
+IHRvd2FyZA== 7246
+IGJhdA== 7247
+IHZhbGlk 7248
+IE1lbnM= 7249
+IGNvbXBsZXRlZA== 7250
+xLHEnw== 7251
+IHBvZGNhc3Q= 7252
+IEJvbg== 7253
+25I= 7254
+IEp1bHk= 7255
+aWxh 7256
+IHBhY2thZ2U= 7257
+IHB1bGxlZA== 7258
+Y2hhcg== 7259
+IE1lbA== 7260
+b2lz 7261
+IHNvdXRo 7262
+IOuU 7263
+IGltcG9ydGFuY2U= 7264
+IHB1c2hpbmc= 7265
+IGlzb2w= 7266
+IHN0YW5kcw== 7267
+Y2lsbA== 7268
+5Lw= 7269
+IPCf 7270
+b3Jp 7271
+6rCB 7272
+IGhvbWVz 7273
+IGNvbmNlcm5z 7274
+IGJpeg== 7275
+5b0= 7276
+Ymll 7277
+IGJpcw== 7278
+IGdlYXI= 7279
+IE1T 7280
+IGh1bg== 7281
+IE1hdHQ= 7282
+4bqj 7283
+c2V5 7284
+IFNlY3JldA== 7285
+IG9kZA== 7286
+IE1heA== 7287
+b2xseQ== 7288
+Zm9yZA== 7289
+IFNI 7290
+IHJlcGxhY2U= 7291
+IG5hdmln 7292
+IGluaQ== 7293
+0LjRjw== 7294
+IGdpYW50 7295
+IG1hbmQ= 7296
+IEhhcHA= 7297
+VElPTg== 7298
+Z3Vu 7299
+aWFtbw== 7300
+7J6F64uI64uk 7301
+IGdhcA== 7302
+IMOqdHJl 7303
+IGNsYXNzcm9vbQ== 7304
+IGh5cA== 7305
+YWtp 7306
+6K4= 7307
+aXN0ZXJz 7308
+YWNrcw== 7309
+INGB0L4= 7310
+IGJ1Zw== 7311
+IGdyYXY= 7312
+YW1pbg== 7313
+IGV2ZXJ5ZGF5 7314
+IOyhsA== 7315
+IGdhcmRlbg== 7316
+Y2VtYmVy 7317
+IGVzdG8= 7318
+2Kw= 7319
+n7A= 7320
+5YE= 7321
+IHJvbQ== 7322
+IOygnOqwgA== 7323
+IGZhbGxpbmc= 7324
+IGZhdWx0 7325
+ZWxseQ== 7326
+IGNoZXN0 7327
+INC70Lg= 7328
+IHBvdGF0bw== 7329
+IGJ1aWxkaW5ncw== 7330
+IG9wZXJhdGluZw== 7331
+IHBhcmU= 7332
+d3I= 7333
+RG9u 7334
+IEZvdXI= 7335
+IHZ1bA== 7336
+IGzDoQ== 7337
+IGZydXN0 7338
+IERhbm4= 7339
+b2xlcw== 7340
+bnlh 7341
+IOy2 7342
+INGA0LDRgQ== 7343
+15s= 7344
+IGHDrQ== 7345
+d29yZA== 7346
+IHdlYXBvbg== 7347
+IG9idA== 7348
+IEZhbGw= 7349
+IFN0ZXZl 7350
+IG1peGVk 7351
+IHBvZGU= 7352
+IEFT 7353
+IExlZw== 7354
+IGRlc2M= 7355
+IHNwbGl0 7356
+IGVtZXJnZW5jeQ== 7357
+IFNpbmc= 7358
+IHByb2ZpdA== 7359
+IHR5cGljYWw= 7360
+IERvbmM= 7361
+IGFubm91bmNl 7362
+IFRleA== 7363
+IHNhY3I= 7364
+dGVybmFs 7365
+IGNvbW1pdHRlZQ== 7366
+aWdv 7367
+IGRpYW0= 7368
+cGhhcw== 7369
+IGRlZmU= 7370
+IFByb2Zlc3M= 7371
+IGRlY2w= 7372
+0YPRgA== 7373
+MjI= 7374
+b2xm 7375
+IE1vbmQ= 7376
+dXk= 7377
+IGF5 7378
+IGxlbQ== 7379
+IGxvdmVseQ== 7380
+IENvdWxk 7381
+IGd1YXI= 7382
+SEg= 7383
+IGNhcmVmdWxseQ== 7384
+IExpc3Rlbg== 7385
+INC60YA= 7386
+IHlvdXRo 7387
+IFRoZXJlZm9yZQ== 7388
+IGRyZWFtcw== 7389
+IEplZmY= 7390
+P10= 7391
+IOuI 7392
+REE= 7393
+IGJvZGllcw== 7394
+YXV4 7395
+IHRlY2huaXF1ZXM= 7396
+IG1lY2hhbmlzbQ== 7397
+15M= 7398
+INC+0L3QuA== 7399
+IGRlc2lyZQ== 7400
+w64= 7401
+IFZv 7402
+cXVlcw== 7403
+INGD0LbQtQ== 7404
+IFdob2E= 7405
+IEdhbWU= 7406
+IGhhbA== 7407
+YW5pc2g= 7408
+IHByYWN0aWNlcw== 7409
+NTAw 7410
+IHNvcnRz 7411
+dXBz 7412
+YXRlZnVs 7413
+IGhlcnNlbGY= 7414
+IGd1aXRhcg== 7415
+IHByb3Bvcw== 7416
+IHNpdGVz 7417
+IGJlYWNo 7418
+INei 7419
+0L3Rgw== 7420
+IGRyYW0= 7421
+IE5vdmU= 7422
+VkU= 7423
+cmFudA== 7424
+IHBsb3Q= 7425
+IOyXrOq4sA== 7426
+IENh 7427
+IGVzdGFibGlzaGVk 7428
+IDIwMTU= 7429
+IGluc3BpcmVk 7430
+IGFubm91bmNlZA== 7431
+INGC0YA= 7432
+IDI2 7433
+IHZveQ== 7434
+IHRlY2g= 7435
+7KCB 7436
+IHByb2Nlc3Nlcw== 7437
+b250bw== 7438
+IFBhbg== 7439
+IHJhcGlk 7440
+aXN0YW4= 7441
+IDE5Nw== 7442
+IHJlbGlnaW9u 7443
+IDI4 7444
+IHNtaWxl 7445
+IGJhYg== 7446
+INqp 7447
+IFZpcg== 7448
+IHNjaGVkdWxl 7449
+IGV4ZWN1dA== 7450
+IHByb24= 7451
+0Y0= 7452
+INCd0YM= 7453
+bXVzaWM= 7454
+7JuQ 7455
+IGdhbg== 7456
+7Iug 7457
+IGRlZmF1bHQ= 7458
+IGJlbQ== 7459
+2Yk= 7460
+IGZvcmNlZA== 7461
+IE9idmlvdXNseQ== 7462
+IHN0b25l 7463
+IHRpZQ== 7464
+IGRyaW5raW5n 7465
+IHNlcnZlZA== 7466
+Q2F1c2U= 7467
+IGNvbmZlcmVuY2U= 7468
+IEV4YWN0bHk= 7469
+oJw= 7470
+7JmA 7471
+IFJh 7472
+IGZha2U= 7473
+IGRpZmY= 7474
+IGNoYWxsZW5naW5n 7475
+IOykkQ== 7476
+z4c= 7477
+IGludGVsbGlnZW5jZQ== 7478
+cmV0ZQ== 7479
+IHN0dWR5aW5n 7480
+IGFwcG9pbnQ= 7481
+IHRhbg== 7482
+INC40Lw= 7483
+IGN1cnZl 7484
+IFRlYW0= 7485
+IEF6 7486
+INC30LQ= 7487
+IE11c2lj 7488
+ZmllbGQ= 7489
+aXJhdGlvbg== 7490
+IGZhaWxlZA== 7491
+IG5vdmVs 7492
+IGRpZmZlcmVudGx5 7493
+IGVzY2FwZQ== 7494
+IFlv 7495
+IE9jdG9iZXI= 7496
+xLF5b3I= 7497
+IGRlc2NyaWJlZA== 7498
+IGNvbnZlcnQ= 7499
+YWNlbWVudA== 7500
+IGhvdGVs 7501
+aXNhdGlvbg== 7502
+IHN1aXM= 7503
+IHdhbGtlZA== 7504
+MjAw 7505
+IG5laWdoYm9yaG9vZA== 7506
+aXNw 7507
+IExvcw== 7508
+IGhpZGRlbg== 7509
+IDI3 7510
+0LvQtQ== 7511
+IHBocg== 7512
+IElzbGFuZA== 7513
+IFN0cmVldA== 7514
+ZW5kYQ== 7515
+aGlwcw== 7516
+b3N1cmU= 7517
+IGRlZmluZWQ= 7518
+4Lin 7519
+IHZpZGE= 7520
+IGxhYmVs 7521
+IEV2ZXJ5Ym9keQ== 7522
+IGpva2U= 7523
+aWFv 7524
+2KfZhg== 7525
+IGF0aGxldA== 7526
+Li4uIg== 7527
+IEZpcmU= 7528
+RG8= 7529
+IGRlZmVuc2U= 7530
+IGVudGVydGFpbg== 7531
+w6F0 7532
+IHBvbGljaWVz 7533
+IGFsY29ob2w= 7534
+IEVuZ2luZQ== 7535
+IGdhbA== 7536
+IEp1ZA== 7537
+IHZvbHVudGU= 7538
+aWNrcw== 7539
+ZXRh 7540
+YWd0 7541
+INeV 7542
+IG3Dtg== 7543
+MTM= 7544
+IGVuY291bg== 7545
+IGVo 7546
+IG9yYW5nZQ== 7547
+IGFic29y 7548
+IHNwYWNlcw== 7549
+IE5vdmVtYmVy 7550
+6rWs 7551
+aWF0 7552
+IHRhbQ== 7553
+Y2tub3c= 7554
+IHN0b3Jt 7555
+IERpcmVjdG9y 7556
+IHByZWdu 7557
+IOydvA== 7558
+INC+0L8= 7559
+IHJlc291cmNl 7560
+IGJhcmQ= 7561
+bmV3 7562
+IERlY2VtYmVy 7563
+dWl0cw== 7564
+IHdlaWw= 7565
+IGNvbnN0cnVjdA== 7566
+c2k= 7567
+bmlj 7568
+IGZsb3Vy 7569
+IHJlc3RyaWN0 7570
+w7x0 7571
+IGVudGlyZWx5 7572
+IGJyZWFraW5n 7573
+ZW50bGljaA== 7574
+IHR3ZW50eQ== 7575
+IGNhdXNlcw== 7576
+IGVsZXY= 7577
+IFNwcg== 7578
+IEludGVybmV0 7579
+IGtpc3M= 7580
+IG9wZXJhdGlvbnM= 7581
+c3p5 7582
+IOuK 7583
+IHNjaWVudGlzdHM= 7584
+IGdyb3du 7585
+IG93bmVycw== 7586
+b3V0cw== 7587
+IGNvdXJzZXM= 7588
+IHVzdWFs 7589
+IGlubg== 7590
+IHRyYW5zbQ== 7591
+w7Fv 7592
+IG51ZXN0 7593
+0LrQvtCy 7594
+IGNhdGVnb3J5 7595
+IExpZmU= 7596
+IFBsdXM= 7597
+IGF0bW9z 7598
+d2hpbGU= 7599
+IHJlY29yZHM= 7600
+IGRlxJ8= 7601
+64uk6rOg 7602
+IOyCrOue 7603
+IHJlcXVpcmVtZW50cw== 7604
+aW5u 7605
+IGltbWln 7606
+IGRlZXBlcg== 7607
+57Q= 7608
+IGFwcHM= 7609
+IGNvbGxlYWd1ZXM= 7610
+xbx5 7611
+IG9mZmVycw== 7612
+IHTDoQ== 7613
+IGNvbHVtbg== 7614
+bGF1ZA== 7615
+SVI= 7616
+IE1z 7617
+IGV4Y2hhbmdl 7618
+bGFz 7619
+IExhdw== 7620
+IEpvbg== 7621
+aXNzZQ== 7622
+cm9nZW4= 7623
+IG1vaQ== 7624
+15c= 7625
+IHNlbmRpbmc= 7626
+IGhlbGxv 7627
+0LXQtQ== 7628
+xZvEhw== 7629
+IHN1Y2NlZWQ= 7630
+IHN1ZmZlcmluZw== 7631
+IGFkdmVydA== 7632
+IOyjvA== 7633
+IHJlY28= 7634
+xLFuxLE= 7635
+INC60L7QvA== 7636
+YWxsZXk= 7637
+IGZhaWx1cmU= 7638
+aWVq 7639
+IOuVjA== 7640
+IGRydWdz 7641
+IGN1YW5kbw== 7642
+IOyWtOuW 7643
+IEFib3V0 7644
+IHF1YW5kbw== 7645
+OTA= 7646
+IEZlZA== 7647
+MTc= 7648
+U2g= 7649
+aW5obw== 7650
+IFN1bmRheQ== 7651
+IFBoaWw= 7652
+IGFjYWRlbWlj 7653
+IEluYw== 7654
+IG1haW50ZW4= 7655
+IHJld2FyZA== 7656
+ZXJk 7657
+IGNvbW1pdHRlZA== 7658
+7Iqk 7659
+0LPRgA== 7660
+IHN0YW5kYXJkcw== 7661
+IGthbA== 7662
+IGludGVudGlvbg== 7663
+IFpo 7664
+IGFja25vdw== 7665
+5L8= 7666
+ID09PQ== 7667
+b2d5 7668
+5ac= 7669
+IGZpbG1z 7670
+aXNr 7671
+IHRlZXRo 7672
+IHN0cnVnZ2xl 7673
+cmQ= 7674
+dWVu 7675
+IGRpc3M= 7676
+IERhcg== 7677
+YW15 7678
+IGVuZW1pZXM= 7679
+IHZlbG9j 7680
+IENhbGw= 7681
+dW1icw== 7682
+0LjRgtC10LvRjA== 7683
+IG9jZWFu 7684
+w6lk 7685
+7Jqw 7686
+IHRyZW0= 7687
+aWVudG8= 7688
+0LXRiNGM 7689
+ZmZpY2llbnQ= 7690
+IGJvdHRsZQ== 7691
+IGluc3RpdHV0aW9u 7692
+ZXN0eQ== 7693
+IEhhbg== 7694
+aGFi 7695
+64qY 7696
+IGFycmVzdA== 7697
+IGxldHRlcnM= 7698
+b3VuY2U= 7699
+7Yw= 7700
+QW4= 7701
+IGNyZWF0ZXM= 7702
+IGNsb2Nr 7703
+IGRlYnQ= 7704
+IGFuY2llbnQ= 7705
+aWZpY2F0aW9ucw== 7706
+Z2k= 7707
+QnV0 7708
+IFR1 7709
+a2w= 7710
+IGJvcmRlcg== 7711
+IG9vaw== 7712
+IEJheQ== 7713
+ZXN0YQ== 7714
+IOuztOw= 7715
+IHdyYQ== 7716
+cHJlbmU= 7717
+IOqyjA== 7718
+YW5nbGU= 7719
+IGJlbGlldmVk 7720
+aWVuY3k= 7721
+YWth 7722
+IGNyaXRpYw== 7723
+IGJvbWI= 7724
+IGhhbQ== 7725
+INCb 7726
+6rWt 7727
+IEd1eXM= 7728
+cm9zb2Z0 7729
+IGNyaW0= 7730
+ZXRjaA== 7731
+QVJS 7732
+IHNpZ2h0 7733
+0LjQvdCw 7734
+IGFpbg== 7735
+4buR 7736
+aXNjaGU= 7737
+IGF1eA== 7738
+IG51bWVy 7739
+IHN1cnZpdmU= 7740
+QWxs 7741
+QkM= 7742
+IHN6 7743
+n6zr 7744
+IGphbQ== 7745
+IENvdXJ0 7746
+IGFsbGVz 7747
+IHRyaWdnZXI= 7748
+0J4= 7749
+IGZvcm1hdA== 7750
+IGRlY2FkZXM= 7751
+IGNlcw== 7752
+IHNpZ25z 7753
+IHJvYm90 7754
+IENodXJjaA== 7755
+IGF6 7756
+IHNvdXA= 7757
+IFRleGFz 7758
+dXRlbg== 7759
+INGH0YLQvtCx0Ys= 7760
+IG5laWdoYg== 7761
+lteU 7762
+IGNvbW11bmljYXRl 7763
+xaE= 7764
+IGVsaW1pbg== 7765
+IGZyZXF1ZW5jeQ== 7766
+aGVybg== 7767
+aWRvcw== 7768
+IGVtcGhhcw== 7769
+IG1lc3NhZ2Vz 7770
+IGdlbmRlcg== 7771
+IFdlbm4= 7772
+INCy0L4= 7773
+IHByaWNlcw== 7774
+b2xv 7775
+INC/0L7QvQ== 7776
+d2luZw== 7777
+IEZpbA== 7778
+0LDQtdC8 7779
+IEN1cg== 7780
+IGZhbHNl 7781
+IGZpZWxkcw== 7782
+IHPDqQ== 7783
+MjQ= 7784
+IG1hYw== 7785
+dcWf 7786
+IGxheWVycw== 7787
+IGFkdm9j 7788
+d2Fu 7789
+IGthcg== 7790
+IMWe 7791
+IGRlY29y 7792
+IHdhbGxz 7793
+b2U= 7794
+aXNzaW9ucw== 7795
+IHJlc29s 7796
+16I= 7797
+IENhcm9s 7798
+IFZpZGU= 7799
+bGVlcA== 7800
+IFlPVQ== 7801
+IGZsaXA= 7802
+IHN1cmdlcnk= 7803
+IGNob3A= 7804
+VVI= 7805
+Liw= 7806
+IGFnZW5jeQ== 7807
+IHdhbnRpbmc= 7808
+IHNvbGFy 7809
+IGhvcml6 7810
+IEFkYW0= 7811
+IHN0YXlpbmc= 7812
+b2xpYw== 7813
+IGdyYXRlZnVs 7814
+IHJlbWFyaw== 7815
+IHRlY2hub2xvZ2llcw== 7816
+IHByb3RlaW4= 7817
+0LTQtdC7 7818
+IE1vbnQ= 7819
+IHNob3VsZGVy 7820
+IHph 7821
+cmV5 7822
+IE9vaA== 7823
+IHN0eQ== 7824
+aWNhcg== 7825
+0L7RgtGA 7826
+IHJvdXRl 7827
+IFR1cm4= 7828
+IGJvbQ== 7829
+IGRlYmF0ZQ== 7830
+IHBvc3NpYmlsaXR5 7831
+IO2VtOw= 7832
+YXBh 7833
+IGludmVudA== 7834
+w7xybGljaA== 7835
+IHByb2ZpbGU= 7836
+IHNlbmlvcg== 7837
+cHB5 7838
+dmFz 7839
+IG11bmRv 7840
+YXRldmVy 7841
+IGFwcGFyZW50bHk= 7842
+ZW5lcg== 7843
+15A= 7844
+560= 7845
+IHByZWNpcw== 7846
+IGFsaWdu 7847
+IGtuaWZl 7848
+IFJvYmVydA== 7849
+5Ys= 7850
+IGZvb2w= 7851
+IGludml0ZQ== 7852
+dXNpbmc= 7853
+IGNpcmN1bXN0 7854
+IGNhcHR1cmU= 7855
+IGRvdWdo 7856
+IFNhbmQ= 7857
+IHNldQ== 7858
+IE5ld3M= 7859
+IGJpdGU= 7860
+IG5ldXQ= 7861
+d2lkZQ== 7862
+IGxlY3R1cmU= 7863
+IOuYkA== 7864
+IG9yaWdpbmFsbHk= 7865
+IGNob2ljZXM= 7866
+IEdhcg== 7867
+IHZlcnNl 7868
+IGxpdA== 7869
+IDE5Ng== 7870
+7ZWg 7871
+IG1lYXN1cmVz 7872
+w6fDtWVz 7873
+d2F0ZXI= 7874
+cml2ZQ== 7875
+IHppam4= 7876
+7YE= 7877
+IEJ1cw== 7878
+IGhlYg== 7879
+0LXRhQ== 7880
+IEthcg== 7881
+IE7Do28= 7882
+IGtpbGxpbmc= 7883
+4K6q 7884
+IG1pcnJvcg== 7885
+bW9k 7886
+IG1vbA== 7887
+IGNyZWF0aW9u 7888
+IGVzdGlt 7889
+IGF0bW9zcGhlcmU= 7890
+IGdhbQ== 7891
+IHRhYmxlcw== 7892
+aXNp 7893
+IExpdHRsZQ== 7894
+IHRhcw== 7895
+IEVsZQ== 7896
+w6ls 7897
+IHNjZW5lcw== 7898
+IHRvbmU= 7899
+IGFmZmVjdGVk 7900
+IEFVREk= 7901
+IEJyb3du 7902
+SWY= 7903
+INmH 7904
+IERhbmllbA== 7905
+cXVlcg== 7906
+Y2hp 7907
+7ZWY6w== 7908
+IG1pc3Rha2Vz 7909
+IHNsYQ== 7910
+IGVudHI= 7911
+INC10YHQu9C4 7912
+IHNob3V0 7913
+IHBvcnRpb24= 7914
+0Zc= 7915
+IHByZXZpb3VzbHk= 7916
+4buZ 7917
+INC/0YDQtdC0 7918
+0L7RgdGM 7919
+IGhlYWRz 7920
+544= 7921
+5a0= 7922
+IGdyYXNz 7923
+4Liw 7924
+Y3JpYmU= 7925
+IHF1w6k= 7926
+IFNwYW5pc2g= 7927
+IG9mZmVyZWQ= 7928
+INCx0YvQu9C+ 7929
+IENsb3Vk 7930
+IHZlY3Rvcg== 7931
+IEh1aA== 7932
+IGthZA== 7933
+aWZ0cw== 7934
+IM69 7935
+IGh1bmdyeQ== 7936
+0KE= 7937
+IHBhcmFsbA== 7938
+QU5E 7939
+IHbDrWRlbw== 7940
+aXp6 7941
+IG9jY3Vw 7942
+IO2U 7943
+IHNlZWs= 7944
+aGVz 7945
+IGRvb3Jz 7946
+IGhvdXNlcw== 7947
+IGNvbnNpZGVyaW5n 7948
+IGdyYWR1YXRl 7949
+IGZ1bGY= 7950
+6KM= 7951
+IGV4dHJlbWU= 7952
+IGZsb3dlcnM= 7953
+aXRhdGU= 7954
+IFByaQ== 7955
+IGZ1bmRhbWVudGFs 7956
+0YfQsNGB 7957
+IHRleHR1cmU= 7958
+jZg= 7959
+IEFORA== 7960
+4K6x 7961
+IFRlbQ== 7962
+IG5hZGE= 7963
+7KeE 7964
+IGNlbGVicmF0ZQ== 7965
+dW1z 7966
+IHBpbGw= 7967
+INC40LvQuA== 7968
+Z29pbmc= 7969
+IGhpcA== 7970
+IHN1cHBvcnRlZA== 7971
+IHBlcm1hbg== 7972
+IGFncmVlbWVudA== 7973
+IHR5bQ== 7974
+IOuR 7975
+k6TsnbQ= 7976
+IHB1cmNoYXNl 7977
+7ZQ= 7978
+IFBsYW4= 7979
+ZWdlbg== 7980
+IHJlY292ZXI= 7981
+UFU= 7982
+IE1pY3Jvc29mdA== 7983
+ZHVj 7984
+IGhvbGVz 7985
+IGRyb3BwZWQ= 7986
+IHBpZw== 7987
+IGVuZGluZw== 7988
+IGF0dGFja3M= 7989
+YmVj 7990
+IHJlbg== 7991
+IHJhcHA= 7992
+IOyasOumrA== 7993
+IHRlcnJvcg== 7994
+INeZ 7995
+IGVkaXQ= 7996
+IGFv 7997
+Ljwv 7998
+IDIwMDA= 7999
+IFVuaW9u 8000
+IHNjaWVudGlmaWM= 8001
+IHB1bmNo 8002
+b3J0aW9u 8003
+IHB1dHM= 8004
+IE1vbmRheQ== 8005
+IEplcg== 8006
+RUM= 8007
+IG1hdHJpeA== 8008
+IGluc3RpdHV0aW9ucw== 8009
+IG1vbnQ= 8010
+IGV4aGli 8011
+IHNwZWFrZXI= 8012
+IG1ldGVycw== 8013
+Ll0= 8014
+IHNlcnZpbmc= 8015
+IGRhdGFiYXNl 8016
+IExBVQ== 8017
+IGRhbW4= 8018
+IHBvZGVy 8019
+ISEhIQ== 8020
+IO2WiA== 8021
+IEFVRElFTkNF 8022
+IGp1bg== 8023
+IEFD 8024
+IEl0YWw= 8025
+c2Vj 8026
+IFlvdW5n 8027
+cnVjaw== 8028
+b3V2ZQ== 8029
+4LiE 8030
+54g= 8031
+IOunjOs= 8032
+YWRpbmc= 8033
+dXJhdGlvbg== 8034
+IFBT 8035
+0Jo= 8036
+IFVuZg== 8037
+6IE= 8038
+b3JpYQ== 8039
+IG1hbmlm 8040
+IHNlbnRlbmNl 8041
+IHNpZ25lZA== 8042
+QlM= 8043
+IHByb29m 8044
+IE11c2xpbQ== 8045
+IG51Y2xlYXI= 8046
+INCz0L7QstC+0YA= 8047
+IHdvbGw= 8048
+IGZhdm91cg== 8049
+IFdI 8050
+IHZ1bG5lcg== 8051
+IGNsb3NlbHk= 8052
+IGluZGV4 8053
+0YLQtdGA 8054
+YWNoZWw= 8055
+IGNhcGFibGU= 8056
+IEJlcw== 8057
+IGNyb2No 8058
+ZWt0 8059
+IHNoZWV0 8060
+IHNlZXM= 8061
+IG5hdHVyYWxseQ== 8062
+IEVuZ2xhbmQ= 8063
+IHBhcnRpY2lwYXRl 8064
+IGV4aXN0cw== 8065
+IHNoYXJw 8066
+cHk= 8067
+IGJyZWFrZmFzdA== 8068
+Ym93 8069
+IHR3aXN0 8070
+56c= 8071
+aW5hdGluZw== 8072
+b3Rp 8073
+IEZvdW5k 8074
+IGRldXg= 8075
+IHNlbGVjdGVk 8076
+7KCE 8077
+b3Npcw== 8078
+IHByZXNlbnRlZA== 8079
+IGxpbmVhcg== 8080
+IOq0 8081
+IGt1bg== 8082
+w7RuZw== 8083
+IGLEmWQ= 8084
+IHRlbXBvcg== 8085
+IGNhYmxl 8086
+INC/0YDQvtGB0YLQvg== 8087
+0LrQtQ== 8088
+INGC0LDQvA== 8089
+IHdpbm5pbmc= 8090
+mOuPhA== 8091
+IDIwMTQ= 8092
+IOyXrOs= 8093
+IFVO 8094
+IENsaWNr 8095
+IHByZXBhcg== 8096
+IFRP 8097
+IHN1YQ== 8098
+IEhhbQ== 8099
+IGzDpA== 8100
+IGFic29sdXRl 8101
+IGVuZ2FnZWQ= 8102
+IEhtbQ== 8103
+IGRhc2g= 8104
+VEE= 8105
+w7Fvcw== 8106
+IHNwbw== 8107
+KV0= 8108
+IHRlc3RlZA== 8109
+IGJsYW5r 8110
+IHJlamVjdA== 8111
+IGFzc2lt 8112
+IHJlYXI= 8113
+IFN0cg== 8114
+IGNyYXNo 8115
+INC90LDRiA== 8116
+0LjRgtGB0Y8= 8117
+IGNvbG9u 8118
+IFVudA== 8119
+IENl 8120
+IGFjaWQ= 8121
+6Zc= 8122
+IGtpdA== 8123
+aWJpbGl0aWVz 8124
+dXRv 8125
+IHZhbHVhYmxl 8126
+bGlzdA== 8127
+IHBhcnRpZXM= 8128
+IE1t 8129
+IGNvbG91cg== 8130
+IGNoYW0= 8131
+IHN0ZWVs 8132
+IEltcA== 8133
+IGZ1bmRz 8134
+IEROQQ== 8135
+IEtlbg== 8136
+aW5kZQ== 8137
+7ZW07ISc 8138
+IEhhcHB5 8139
+IFVzZQ== 8140
+IExpZ2h0 8141
+IGxpcA== 8142
+IGF1dGhvcml0eQ== 8143
+IExvbmc= 8144
+IElyYW4= 8145
+IGVsbA== 8146
+IGNvb3JkaW4= 8147
+IHN1Ym0= 8148
+IHJlY29yZGVk 8149
+0YPRiA== 8150
+IGRlbHRh 8151
+IHJlZm9ybQ== 8152
+IFN0aWxs 8153
+IG9wcG9u 8154
+IGFsbG93aW5n 8155
+IHBhdHRlcm5z 8156
+IGxldHRpbmc= 8157
+IHNsZWVwaW5n 8158
+T2theQ== 8159
+IHBpenph 8160
+IMWb 8161
+INC00L7Quw== 8162
+IHRhbGVudA== 8163
+ZW5zaW9ucw== 8164
+IGVudmlyb25tZW50YWw= 8165
+IHByb2Zlc3Nvcg== 8166
+IHNob3Rz 8167
+IGNvbnRhaW5z 8168
+dWdhcg== 8169
+eW8= 8170
+j5k= 8171
+IHNlcXVlbmNl 8172
+zrnOsQ== 8173
+YWRlcg== 8174
+6aA= 8175
+0LDRhw== 8176
+2YbYpw== 8177
+IElr 8178
+IHRvdXM= 8179
+dXJpZXM= 8180
+IHBvdW5kcw== 8181
+IGV4dGVybmFs 8182
+aW1lbnRz 8183
+IHZyYWltZW50 8184
+7Iuk 8185
+IGhhcHBpbmVzcw== 8186
+IHByemU= 8187
+ZXN0aWM= 8188
+IGVzdGFibGlzaA== 8189
+IEZsb3I= 8190
+IHJpZw== 8191
+IGhvbmV5 8192
+IHB1bA== 8193
+IHN5bXB0b21z 8194
+IGJyb3dz 8195
+0LXQu9C4 8196
+IM+Ezr8= 8197
+IHNoaXJ0 8198
+IFRlY2hu 8199
+IFByb2dyYW0= 8200
+0LXQvNGD 8201
+IHVwc2V0 8202
+IGd1ZXN0 8203
+YnVyZw== 8204
+IHVubGlrZQ== 8205
+IHNvbWV3aGF0 8206
+IGhhbmdpbmc= 8207
+YWU= 8208
+IHJ1bQ== 8209
+IHBob3RvZ3JhcGg= 8210
+IExp 8211
+IHN0YWJsZQ== 8212
+IHZvbHRhZ2U= 8213
+IEVsbA== 8214
+IGVudHJlcHJlbmU= 8215
+dXNlcw== 8216
+YXNzZW4= 8217
+rLg= 8218
+IOunjuydtA== 8219
+IGdob3N0 8220
+IHNhZ2Vu 8221
+IGNvbWJhdA== 8222
+IGfDtnI= 8223
+IENhcA== 8224
+IHPDo28= 8225
+IEthdA== 8226
+IGZvcm1h 8227
+IHN1bW0= 8228
+IG1hcmNo 8229
+IHZhc3Q= 8230
+w7xr 8231
+IGNvbW1pdG1lbnQ= 8232
+aW1vcw== 8233
+TGV0 8234
+IGRlZGljYXRlZA== 8235
+aXN0ZQ== 8236
+bGF5 8237
+IHRvcGljcw== 8238
+IG1hY2hpbmVz 8239
+IFBhcmlz 8240
+IOydtOufsA== 8241
+IG1pbmk= 8242
+IG1hcmtldHM= 8243
+IGtv 8244
+zrQ= 8245
+dmlsbGU= 8246
+IGdvb2RuZXNz 8247
+IGZyYW1ld29yaw== 8248
+dWx0dXJl 8249
+IGJhc2tldA== 8250
+ZXNzYQ== 8251
+0LDRhtC4 8252
+dXN0ZXI= 8253
+IOq5 8254
+IGV4dGVudA== 8255
+IE1lbnNjaGVu 8256
+IGNvbnNpc3RlbnQ= 8257
+IGF1dG8= 8258
+cmlw 8259
+IG1lcmU= 8260
+4K+I 8261
+0ZQ= 8262
+IGVsbGU= 8263
+jIDr 8264
+b2tlbg== 8265
+IHB1bGxpbmc= 8266
+IGNvdw== 8267
+b3V0aGVybg== 8268
+IG1lZXRpbmdz 8269
+IGNhZGE= 8270
+0L3Ri9C8 8271
+aWVudGU= 8272
+IGJhc3Q= 8273
+YW5pbmc= 8274
+IGZvY3VzaW5n 8275
+cm9hZA== 8276
+IHJvb2Y= 8277
+IFByb2Zlc3Nvcg== 8278
+IFNQ 8279
+0YDQsNC3 8280
+IG5vb2Q= 8281
+IDQwMA== 8282
+IOydtOygnA== 8283
+7J6I 8284
+IE1vdW50 8285
+0LXQudGH0LDRgQ== 8286
+INeQ 8287
+V2h5 8288
+154= 8289
+xLFuZGE= 8290
+IHBvc2l0aW9ucw== 8291
+w6htZQ== 8292
+548= 8293
+INC00YDRg9Cz 8294
+aXlvcg== 8295
+IHBhc3Npbmc= 8296
+IGFzc2VtYg== 8297
+IHNtb2tl 8298
+IHRpbA== 8299
+IG11c2V1bQ== 8300
+0JQ= 8301
+IFBlcnNvbg== 8302
+0L3QuNC8 8303
+bGVpY2g= 8304
+IGludGVudA== 8305
+IHNxdWU= 8306
+IGNyYWZ0 8307
+7IiY 8308
+b3JzdW4= 8309
+IDE1MA== 8310
+IGJyb3RoZXJz 8311
+dm9y 8312
+IFNwZWFrZXI= 8313
+aWNpYW5z 8314
+IG9mZmljZXI= 8315
+IGnDp2lu 8316
+INGC0LXQsQ== 8317
+IHNjcmF0Y2g= 8318
+IGdlbmVyYXRl 8319
+eWk= 8320
+IGVtb3Rpb25z 8321
+YXVz 8322
+7LmY 8323
+NDU= 8324
+IExpbms= 8325
+IFJlYWw= 8326
+IGF0ZQ== 8327
+INC90LDQtA== 8328
+IG5hdGl2ZQ== 8329
+4buH 8330
+xLF5 8331
+IGVub3Jt 8332
+IGJsb2Nrcw== 8333
+IGZhY2Vz 8334
+YWNj 8335
+aXZlbmVzcw== 8336
+IGluY2hlcw== 8337
+dWlz 8338
+aGVpdA== 8339
+IHN0cmVldHM= 8340
+IHByb2JhYmlsaXR5 8341
+YXNp 8342
+IGltcGw= 8343
+IOCk 8344
+dXJkYXk= 8345
+IGZhdXQ= 8346
+b215 8347
+IHBpcA== 8348
+IGlsbHVzdA== 8349
+4K6v 8350
+IEp1bg== 8351
+IGx5aW5n 8352
+OTk= 8353
+IG1lbW9yaWVz 8354
+IHByYWN0aWNhbA== 8355
+aWFuYQ== 8356
+b25jZXM= 8357
+IHZpZXdlcnM= 8358
+IFRob21hcw== 8359
+5ow= 8360
+IEdpcmw= 8361
+IFdoZXRoZXI= 8362
+IGlubm92YXRpb24= 8363
+IGRpc2FwcG9pbnQ= 8364
+TXk= 8365
+IHdpbm5lcg== 8366
+IGln 8367
+IHJhdGlv 8368
+IEJsdWU= 8369
+IFN1Yg== 8370
+IGRvY3VtZW50cw== 8371
+IGZvcm11bGE= 8372
+IOup 8373
+0Yo= 8374
+IGFwcGVhcmVk 8375
+dmFy 8376
+YW5kb24= 8377
+IHNwcmF5 8378
+bWFr 8379
+IFFVRVM= 8380
+S0U= 8381
+IHdlZGRpbmc= 8382
+UmU= 8383
+0LDRgtGM0YHRjw== 8384
+IHVubw== 8385
+IGdhbGw= 8386
+7YSw 8387
+Y2lv 8388
+Y2Vycw== 8389
+INC80L3QtQ== 8390
+IHBlcHBlcg== 8391
+IEZlYnJ1 8392
+IGFsdGVybmF0aXZl 8393
+IGZ1 8394
+IEJhc2ljYWxseQ== 8395
+IFNtaXRo 8396
+IGdhdGU= 8397
+IFRhbQ== 8398
+IFdoYXRldmVy 8399
+IGFwcHJveGlt 8400
+IGNvbmNlcnQ= 8401
+IGp1aWNl 8402
+IEVzcGVjaWFsbHk= 8403
+IGR5bmFtaWM= 8404
+UXU= 8405
+b25kZXI= 8406
+aXZlcnk= 8407
+IGJhbmc= 8408
+IHJ1bA== 8409
+IFBhcnR5 8410
+IHNjaG9sYXJz 8411
+IGNyeWluZw== 8412
+asSF 8413
+0KI= 8414
+IFFVRVNUSU9O 8415
+cmlk 8416
+IGFjY3VyYXRl 8417
+w6dv 8418
+IENvb2w= 8419
+Y29pbg== 8420
+IOyDgQ== 8421
+IEZv 8422
+IHByw7M= 8423
+IFJvbWFu 8424
+INCf0YA= 8425
+IGNoZWNraW5n 8426
+Pyc= 8427
+IGF0dGFjaGVk 8428
+IElzbGFt 8429
+IGV4cGVydHM= 8430
+16c= 8431
+IENvbnN0 8432
+0YDQsNC9 8433
+IHNoYWRvdw== 8434
+IGRlbGF5 8435
+0JI= 8436
+IG9yaWVudA== 8437
+64I= 8438
+ZWxsZW4= 8439
+IGFzw60= 8440
+0LrQuNC5 8441
+IGhpc3RvcmljYWw= 8442
+IHVuY29t 8443
+b21w 8444
+aG0= 8445
+IGJpbA== 8446
+IHBsYW5uZWQ= 8447
+IFVuZm9ydHVuYXRlbHk= 8448
+IFdpbmRvd3M= 8449
+2LQ= 8450
+IGVuY291bnRlcg== 8451
+IOyDneqwgQ== 8452
+IHJlZ2FyZGluZw== 8453
+YXJyYXNz 8454
+IHJlY292ZXJ5 8455
+IEh1cg== 8456
+IEVtcA== 8457
+IHPDrQ== 8458
+7ZWY6rKM 8459
+IGRlZmVuZA== 8460
+IGNldA== 8461
+YXNzZQ== 8462
+64uo 8463
+b2tlcw== 8464
+IHJlbW90ZQ== 8465
+INiz 8466
+IGFydHM= 8467
+aXNjbw== 8468
+YXVjb3Vw 8469
+IE1leGljbw== 8470
+INC/0L7QvA== 8471
+IGNob3Nlbg== 8472
+ZW1hdA== 8473
+b2Rpbmc= 8474
+IGZsb3dlcg== 8475
+c3RhbmRpbmc= 8476
+IEFzc29jaQ== 8477
+dW1teQ== 8478
+SUxM 8479
+IGNhbWVyYXM= 8480
+IEFyYWI= 8481
+IFN1bQ== 8482
+IHRlZ28= 8483
+IGNyaW1pbmFs 8484
+aWZvcm0= 8485
+IHN0YWNr 8486
+7ISx 8487
+IERvbmFsZA== 8488
+IE9sZA== 8489
+IGR1c3Q= 8490
+IEpvc2U= 8491
+IGhlbQ== 8492
+IGluY3JlYXNlcw== 8493
+b3N0YQ== 8494
+IGR5aW5n 8495
+IFJpdmVy 8496
+IG1vaXN0 8497
+0YLQvtCy 8498
+YXJlcw== 8499
+IGRpc2NpcGw= 8500
+cmFpdA== 8501
+IEhhcw== 8502
+eWdlbg== 8503
+IFRyZQ== 8504
+IOu0 8505
+IGxhbmd1YWdlcw== 8506
+IEhlbg== 8507
+IDM2 8508
+IERpc25leQ== 8509
+aW50cw== 8510
+IGFsZ28= 8511
+IGZvb2Rz 8512
+IHNldHVw 8513
+bGFu 8514
+IGVmZmVjdGl2ZWx5 8515
+IHdoZXJldmVy 8516
+IHVudGVy 8517
+Zm9ybWF0aW9u 8518
+IGhpdHM= 8519
+IHByaW5jaXBsZQ== 8520
+IHRhc3Rlcw== 8521
+p4g= 8522
+IHRyZWF0ZWQ= 8523
+IHJlc29sdXRpb24= 8524
+IHByaXZpbGU= 8525
+IElQ 8526
+67A= 8527
+IHRlcnJpdA== 8528
+IHBvd2Vycw== 8529
+IO2D 8530
+IFZpY3Q= 8531
+IGJvdGhlcg== 8532
+IENoYWly 8533
+IG11c2NsZQ== 8534
+IHNhbGU= 8535
+IGRlY2VudA== 8536
+IGNvdXA= 8537
+IFNxdQ== 8538
+IGNvYXN0 8539
+IHJvZA== 8540
+IEZyYW5j 8541
+IGJhdGhyb29t 8542
+IHNob3BwaW5n 8543
+INC80L7QttC10YI= 8544
+IGnFnw== 8545
+IFN0YXk= 8546
+Z3JhZGU= 8547
+IGZvcm1lZA== 8548
+IGJhxZ8= 8549
+IGJyaWxs 8550
+am91cg== 8551
+7ZY= 8552
+d2ll 8553
+aWNhdGU= 8554
+IOKAi+KAiw== 8555
+IE5vcm0= 8556
+4KU= 8557
+IG1haW5seQ== 8558
+IFNwYWNl 8559
+IHRyZW1lbmQ= 8560
+aXRp 8561
+4K61 8562
+VVQ= 8563
+TXVzaWM= 8564
+IEZlYnJ1YXJ5 8565
+IGNvbnRyYXN0 8566
+ZXN0aW5n 8567
+IM60 8568
+aW5naW5n 8569
+INmG 8570
+c3Nlbg== 8571
+IEhvbWU= 8572
+IHNoZWxs 8573
+IEhheQ== 8574
+IGFsbGVy 8575
+IEFw 8576
+IFdlc3Rlcm4= 8577
+IFdvcmQ= 8578
+IFBMQVk= 8579
+IOuF 8580
+IEFxdQ== 8581
+IGVudHJ5 8582
+IGxhdW5jaGVk 8583
+IE1lbQ== 8584
+IFBvdXI= 8585
+IHp3ZQ== 8586
+IFNvbWVvbmU= 8587
+aW5nZQ== 8588
+IFByb2I= 8589
+bWJsZQ== 8590
+IFJlbA== 8591
+dXJ1 8592
+IHJoeQ== 8593
+IGdpZw== 8594
+IGVuZ2FnZW1lbnQ= 8595
+w7zFnw== 8596
+IG9mZmVyaW5n 8597
+d2hlbA== 8598
+IGFjdG9y 8599
+QVBQ 8600
+d2VzdA== 8601
+IFJveQ== 8602
+IHJldHVybmVk 8603
+IHNpbHZlcg== 8604
+cmF0aW5n 8605
+IGVzdGFy 8606
+IHNrZQ== 8607
+IHRp 8608
+aWNhdGlvbg== 8609
+IGFubm95 8610
+IGRlZXBseQ== 8611
+7Jqp 8612
+IG5hdMO8cmxpY2g= 8613
+RUxM 8614
+IENhdGg= 8615
+IHJhaWw= 8616
+0L3QvtCy 8617
+IHByYXllcg== 8618
+Y29s 8619
+R0I= 8620
+INCi0LDQug== 8621
+IGdsYQ== 8622
+IFdhdGVy 8623
+0Y/RgtGM 8624
+IE5vbg== 8625
+w7R0 8626
+YWdlcnM= 8627
+IGh1Zw== 8628
+IGRvY3RvcnM= 8629
+YW5jaW5n 8630
+IFRhbGs= 8631
+emluZw== 8632
+IGhhZG4= 8633
+IGx1aQ== 8634
+IGF0w6k= 8635
+IOq3uOumrOqzoA== 8636
+6rmM7KeA 8637
+aWNp 8638
+IGluY29ycG9y 8639
+IERp 8640
+emls 8641
+YW55YQ== 8642
+qoU= 8643
+IMK7 8644
+MzU= 8645
+IGJlZXI= 8646
+IGJlYXVjb3Vw 8647
+IE1D 8648
+IGVhcnM= 8649
+b2dlbg== 8650
+IFF1ZXN0 8651
+ZWRh 8652
+IFNhdHVyZGF5 8653
+IGZhbGxz 8654
+c3Rvbg== 8655
+Ymxlcw== 8656
+IHRodXM= 8657
+IOuEpA== 8658
+4LmE 8659
+IHRoZXJt 8660
+IGRpdmVyc2l0eQ== 8661
+IHNveQ== 8662
+YXp1 8663
+aW1w 8664
+IHRlbGV2aXNpb24= 8665
+INep15w= 8666
+IHd1cg== 8667
+IGVkZ2Vz 8668
+IGxlc3NvbnM= 8669
+IEF1ZA== 8670
+dm9pcg== 8671
+YW1lbnRv 8672
+IGV4cGxhaW5lZA== 8673
+INC+0L3QsA== 8674
+IHRlbXBz 8675
+z44= 8676
+VGhleQ== 8677
+IHN1cnByaXNpbmc= 8678
+0LDQvdC40Y8= 8679
+IERyYWc= 8680
+IENsZQ== 8681
+IG5hbQ== 8682
+INC70Y7QtA== 8683
+IGhhcmR3YXJl 8684
+IHRodW1icw== 8685
+IM66zrHOuQ== 8686
+IFRvcA== 8687
+IMOl 8688
+6Zk= 8689
+15XXqA== 8690
+IOq3uOuemOyEnA== 8691
+IEJ1ZGQ= 8692
+dGhlcm4= 8693
+IGludGVyZXN0cw== 8694
+2LA= 8695
+IGRldmVsb3BlcnM= 8696
+IGhpdHRpbmc= 8697
+IG9wcG9zZWQ= 8698
+IGhlYXJ0cw== 8699
+IEFuZHJvaWQ= 8700
+IEhhbmQ= 8701
+IHJlcHJlc2VudHM= 8702
+Z2xpY2g= 8703
+7Yq4 8704
+IDMy 8705
+IGRvbWlu 8706
+IEFubg== 8707
+IMOpdMOp 8708
+IHpvb20= 8709
+IGt0w7NyZQ== 8710
+IGFkdWx0cw== 8711
+IG9yZGVyZWQ= 8712
+IHBpY2tpbmc= 8713
+IEhvbmc= 8714
+IGZpbG1pbmc= 8715
+IHNlZWQ= 8716
+IEFU 8717
+IGNhbGN1bGF0ZQ== 8718
+INC60L7Qs9C00LA= 8719
+IE9z 8720
+aWNpdA== 8721
+IHJlbWFpbmluZw== 8722
+IHNlZ3U= 8723
+w7s= 8724
+IOyYpOuKmA== 8725
+IGFycml2ZQ== 8726
+IGNvbmdy 8727
+IGdyYW5kZQ== 8728
+IGhlYWx0aGNhcmU= 8729
+INC80L7QttC90L4= 8730
+U0E= 8731
+ZXN0ZQ== 8732
+IGF3YXJlbmVzcw== 8733
+IHNxdWFyZWQ= 8734
+eHR1cmU= 8735
+IEJlaW5n 8736
+IHNvbGRpZXJz 8737
+0YPQsQ== 8738
+IHJldm9sdXRpb24= 8739
+IHRyYWluZWQ= 8740
+ZW5kZW4= 8741
+6LA= 8742
+IGRhbmNpbmc= 8743
+IGluc3RhbGxlZA== 8744
+cHJpc2U= 8745
+IHZldGVy 8746
+IG1lbm9z 8747
+bmVsbA== 8748
+IEJyb3RoZXI= 8749
+IG51bg== 8750
+IGltcG9ydGFudGx5 8751
+YWxsZWQ= 8752
+aWHFgg== 8753
+YWJsZWQ= 8754
+IFN5c3RlbQ== 8755
+IFZvbA== 8756
+IGVsZA== 8757
+IGVtb3Rpb24= 8758
+aWNhbg== 8759
+IEJhbms= 8760
+aWtlcw== 8761
+IHZsb2c= 8762
+INCy0L7Qtw== 8763
+IHB1ZWRl 8764
+7Jik 8765
+IHRlZW4= 8766
+IHNldmVyZQ== 8767
+JSw= 8768
+IGNsZWFuaW5n 8769
+esSF 8770
+l5A= 8771
+IFRocm91Z2g= 8772
+IFNldA== 8773
+RVA= 8774
+Ij8= 8775
+IE1vdGhlcg== 8776
+IGZpZ3VyZWQ= 8777
+IG11ZA== 8778
+INGW 8779
+IE9mZmljZQ== 8780
+IHJhdw== 8781
+IGRlc3Ryb3llZA== 8782
+ZW50YQ== 8783
+IGFnZ3Jlc3M= 8784
+INC+0YE= 8785
+IOuqqOs= 8786
+w6TDpA== 8787
+IEFS 8788
+IGNvcnJlY3RseQ== 8789
+IHN0aXI= 8790
+IGV4dHJhY3Q= 8791
+IHZlaGljbGVz 8792
+IFJ1bg== 8793
+INCy0YDQtdC8 8794
+IHBhcmFsbGVs 8795
+IGxhZw== 8796
+anU= 8797
+IGRhcmU= 8798
+IE1vdA== 8799
+b25v 8800
+IGJlaW5ncw== 8801
+IHN0cm8= 8802
+IGV4Y3VzZQ== 8803
+IGFscGhh 8804
+IGFza3M= 8805
+IHBvY2tldA== 8806
+Li4uPw== 8807
+IGtpdGE= 8808
+w7xt 8809
+IGFwcGVhcmFuY2U= 8810
+b3JkYW4= 8811
+IGluc2VydA== 8812
+INC90LDRhw== 8813
+m2k= 8814
+IHRlbXBv 8815
+IGZhY2lsaXR5 8816
+IHZpc2libGU= 8817
+5ZI= 8818
+IFNjaWVuY2U= 8819
+dXJvcw== 8820
+INmB2Yo= 8821
+IFZhbg== 8822
+IHRlbnNpb24= 8823
+IO2VoA== 8824
+IGRlbGl2ZXJ5 8825
+IHN0aW0= 8826
+IHN1cnZleQ== 8827
+IEdyYQ== 8828
+IGJvbA== 8829
+5qA= 8830
+IHdlaXRlcg== 8831
+w59lbg== 8832
+IHByb2NlZWQ= 8833
+IGltcHJlc3NpdmU= 8834
+IFZvYw== 8835
+aW91c2x5 8836
+INC00LA= 8837
+aGFsZQ== 8838
+b2No 8839
+IGdsdWU= 8840
+cGhldA== 8841
+Y29udA== 8842
+IGZpdHM= 8843
+IGJveGVz 8844
+IGNvbnRyb2xz 8845
+IENoaWxk 8846
+IHNjZW5hcmlv 8847
+IHRyb3A= 8848
+IHByb2Nlc3Npbmc= 8849
+INGC0L7Qu9GM0LrQvg== 8850
+IGJpcmRz 8851
+IENoaWM= 8852
+INC90LDQvw== 8853
+IDIwMTM= 8854
+IG3DvHNzZW4= 8855
+IEphZw== 8856
+IHPEhQ== 8857
+IHBlcmNl 8858
+cmVo 8859
+IEZvcmU= 8860
+IGNvbmZ1c2Vk 8861
+YWlyZQ== 8862
+IGFjY29tcGxpc2g= 8863
+IGNhc2E= 8864
+Y2xvY2s= 8865
+IGluZmx1ZW4= 8866
+IFJP 8867
+IGJvbmU= 8868
+aWNpYW4= 8869
+IFND 8870
+IHN0cmF0ZWdpZXM= 8871
+Z2g= 8872
+0LTRgw== 8873
+IGl0dQ== 8874
+IHBlcnNvbmFsaXR5 8875
+IGJhcmR6bw== 8876
+IGFjY2VwdGVk 8877
+IHN0b20= 8878
+aWV2 8879
+IEhpc3Q= 8880
+IEF1cw== 8881
+IOuwlOs= 8882
+QVRPUg== 8883
+b2ly 8884
+IG1hZ2F6 8885
+IGV4cGxhbg== 8886
+IGNvcm4= 8887
+IGlscw== 8888
+IGNpcmN1aXQ= 8889
+IGdheQ== 8890
+aG9w 8891
+IGVxdWl2YWw= 8892
+IGRpZXNlcg== 8893
+ZXJ2ZXM= 8894
+Y29tZXM= 8895
+a2xpY2g= 8896
+IOuVjOs= 8897
+YWJldA== 8898
+IGV4aGE= 8899
+IG1hbm5lcg== 8900
+IOKZquKZqg== 8901
+w6lj 8902
+w6Rs 8903
+IGNvbmZpcm0= 8904
+IGVudGVyZWQ= 8905
+ZW1wbG8= 8906
+IEZhcg== 8907
+IG/DuQ== 8908
+ZXNzaW9ucw== 8909
+IG51cnM= 8910
+IGVudMOjbw== 8911
+IGFiYW5kb24= 8912
+bGlmZQ== 8913
+IHdpcw== 8914
+TmFycmF0b3I= 8915
+IOyWtA== 8916
+VGhlcmU= 8917
+IFJhbQ== 8918
+YXN0ZQ== 8919
+IGF0dHJpYg== 8920
+IEF5 8921
+IG1lc21v 8922
+IM69zrE= 8923
+6as= 8924
+ZW5zZXM= 8925
+IGNyb3A= 8926
+INC30LTQtdGB0Yw= 8927
+IFVudGls 8928
+c3RlaW4= 8929
+IG92ZW4= 8930
+IHN1c3BlY3Q= 8931
+aGV0 8932
+IHB1aXM= 8933
+IGNhcnJpZWQ= 8934
+w6ln 8935
+IERldg== 8936
+ZW1z 8937
+cmVlbnM= 8938
+YmVycnk= 8939
+IHRlbXBs 8940
+IEJpdA== 8941
+IHZhcmlhYmxlcw== 8942
+IG92ZXJ3aGVs 8943
+zrzOtQ== 8944
+IGluaXRpYWxseQ== 8945
+7JWY 8946
+b3RoaW5n 8947
+0LXRgtGM 8948
+IEhpbGw= 8949
+IGRlcGFydA== 8950
+IG15c3Q= 8951
+YXp6 8952
+IGZsdWlk 8953
+IERD 8954
+IGNsaW5pY2Fs 8955
+IFJ5YW4= 8956
+IEZsb3JpZGE= 8957
+IFRhaw== 8958
+IGFueGlldHk= 8959
+YnJv 8960
+IGNpcmN1bXN0YW5jZXM= 8961
+INmD 8962
+IGV4aXN0ZW5jZQ== 8963
+IHRvbmc= 8964
+IDIwMTI= 8965
+IFNlY3JldGFyeQ== 8966
+IHNwaWN5 8967
+IFso 8968
+IFdpdGhvdXQ= 8969
+IGZhY3Rz 8970
+IHRvbnM= 8971
+QXBw 8972
+IFN0YW5k 8973
+IGxpZXM= 8974
+IEFE 8975
+d2lu 8976
+z4TOtQ== 8977
+YXBwbGF1c2U= 8978
+SVA= 8979
+c3Rh 8980
+IFN1cA== 8981
+cGhvbmVz 8982
+npE= 8983
+cGll 8984
+IFBvdA== 8985
+IE5P 8986
+INee 8987
+INCU0LA= 8988
+aWNhcw== 8989
+IEly 8990
+IHB1c2hlZA== 8991
+IHVuY2xl 8992
+INmF2YY= 8993
+IGxvbg== 8994
+IHByaW5jaXBsZXM= 8995
+IEludGVybmF0aW9uYWw= 8996
+IMOW 8997
+xb4= 8998
+IHNheWE= 8999
+IOqzoA== 9000
+IHJpYg== 9001
+IHBhc3Rl 9002
+IHdhcm5pbmc= 9003
+IG11c2ljYWw= 9004
+IGFncmVlZA== 9005
+0L7RgNC8 9006
+IGdhcmxpYw== 9007
+IG94eWdlbg== 9008
+7JiI 9009
+QWw= 9010
+IOunng== 9011
+ZWxpbmVz 9012
+TEFVU0U= 9013
+Z3lwdA== 9014
+R0U= 9015
+Y2tlcg== 9016
+dHU= 9017
+IHNoZWw= 9018
+IHN0YXllZA== 9019
+INCz0L7QtA== 9020
+IGxhcHQ= 9021
+IE1hcnRpbg== 9022
+IGludml0ZWQ= 9023
+IGNvbmZpcg== 9024
+IGVtYmFycmFzcw== 9025
+YWNpb25lcw== 9026
+IENhbXA= 9027
+IGhvbGRz 9028
+YXh5 9029
+IGRpdmU= 9030
+dWNrbGVz 9031
+IGJvb3N0 9032
+IHfDvHI= 9033
+c3RhbA== 9034
+INGA0LDQsdC+0YI= 9035
+IGTDqWM= 9036
+IG9mZmljZXJz 9037
+IOyVhOs= 9038
+b2xvZ2lzdA== 9039
+157X 9040
+IHNlZWRz 9041
+IGJ1ZmY= 9042
+IHVwZGF0ZXM= 9043
+ZGVk 9044
+IGZyaWVuZGx5 9045
+IGNvdW5jaWw= 9046
+IFByb2JhYmx5 9047
+IHBpYW5v 9048
+IHJlZHVjZWQ= 9049
+z4TOsQ== 9050
+IGF1dGhlbnQ= 9051
+IGV4cGxvcw== 9052
+cGFzcw== 9053
+IEhpdA== 9054
+anVk 9055
+IE5hdg== 9056
+b21p 9057
+IGNvbW1pc3Npb24= 9058
+IGd5bQ== 9059
+0J8= 9060
+IHBvbg== 9061
+0YDQvtGB 9062
+IGludGVyZmFjZQ== 9063
+IHN0cnVjdHVyZXM= 9064
+IEplbg== 9065
+IHlvaw== 9066
+IG1ldQ== 9067
+7KeA66eM 9068
+bmVk 9069
+IFdpZQ== 9070
+IGlkZW50aWZpZWQ= 9071
+IGNoYW5uZWxz 9072
+xLFuYQ== 9073
+IHBoaWxvc29w 9074
+a2VpdA== 9075
+IGJpdHM= 9076
+ZW50ZXM= 9077
+IGZyYWc= 9078
+IEtpbmQ= 9079
+IGRvY2g= 9080
+IHNuZQ== 9081
+aW5kaW5n 9082
+IEpld2lzaA== 9083
+0L7RgNC+0Yg= 9084
+IGZ1ZQ== 9085
+IO2P 9086
+IG3EsQ== 9087
+IGtlaW5l 9088
+IGxvY2F0aW9ucw== 9089
+IG1ldGVy 9090
+IGJlZWY= 9091
+IG1hbmlw 9092
+IHNvbm8= 9093
+enpsZQ== 9094
+57Y= 9095
+IHBlcw== 9096
+IGhvcnJpYmxl 9097
+IFNu 9098
+IGZhY3Rvcnk= 9099
+IGZpZnRo 9100
+IGNvb2tlZA== 9101
+IG1vb2Q= 9102
+IHZlbG9jaXR5 9103
+IG9ibGln 9104
+IGNvbm5lY3Rpb25z 9105
+xJ9pbQ== 9106
+IOqztQ== 9107
+IGRvbWFpbg== 9108
+IGFwcGx5aW5n 9109
+IHJpZGlj 9110
+IGNlbA== 9111
+IGNoaWxkaG9vZA== 9112
+IFRlc3Q= 9113
+cmF0dWxhdGlvbnM= 9114
+IFZpcmdpbg== 9115
+IENFTw== 9116
+INC/0Ls= 9117
+IGFsZ29yaXRobQ== 9118
+IGludGVyYWN0aW9u 9119
+YWdh 9120
+IGtpZGRpbmc= 9121
+IHRvbWF0bw== 9122
+IGNvbnRpbnVpbmc= 9123
+bGFk 9124
+c3RyZWFt 9125
+0L7QttC1 9126
+IOyYgQ== 9127
+0LXQu9C+0LI= 9128
+QkE= 9129
+IG5hcA== 9130
+IE5vYm9keQ== 9131
+IHRodW1i 9132
+IE9O 9133
+IHJ1c2g= 9134
+RFI= 9135
+IHN0cmlrZQ== 9136
+IGV2b2x1dGlvbg== 9137
+aWNoZQ== 9138
+IOy7 9139
+IOq3uOufsA== 9140
+2KfYqg== 9141
+IGFr 9142
+IHdpbmRvd3M= 9143
+IGV4Y2Vzcw== 9144
+IGNvbmNsdWQ= 9145
+IGVwaXNvZGVz 9146
+IHN0cnVnZ2xpbmc= 9147
+IERhdA== 9148
+nbzr 9149
+IGtleXM= 9150
+IGtsZQ== 9151
+IHZlZ2V0YWJsZXM= 9152
+eXN0ZW0= 9153
+w6puY2lh 9154
+cmljaw== 9155
+IHJldmVudWU= 9156
+IEhhdw== 9157
+IGxhbg== 9158
+YW50ZXM= 9159
+aW5peg== 9160
+0LjRgdGC 9161
+IHN1cA== 9162
+qbTshJw= 9163
+IG1vbWVudG8= 9164
+aXN0bw== 9165
+IEVyaWM= 9166
+aW9ycw== 9167
+YmFq 9168
+IGludHJvZHVjdGlvbg== 9169
+aXJ0eQ== 9170
+IGRlY2s= 9171
+cmVhbA== 9172
+IE1hcmlv 9173
+IGxvdmluZw== 9174
+4LiU 9175
+IHN1cHBvcnRz 9176
+0LjRh9C10YE= 9177
+IGluY2lkZW50 9178
+dXRjaA== 9179
+dXY= 9180
+IGJvb20= 9181
+0LXRgNGM 9182
+INC90YPQtg== 9183
+IGNvbWJpbmVk 9184
+IExpbg== 9185
+MjM= 9186
+b3JhdGlvbg== 9187
+bnRl 9188
+IHNvcg== 9189
+IGRpcnR5 9190
+aWZlcg== 9191
+IEFQSQ== 9192
+IGNvbGxhYm9yYXRpb24= 9193
+aWFibGU= 9194
+IHByaW9yaXR5 9195
+IEFsZQ== 9196
+IFByaW4= 9197
+IEV4Yw== 9198
+IHZhaXM= 9199
+IGdyYW4= 9200
+IHN0b29k 9201
+IHJlY3J1 9202
+IE11cg== 9203
+ZXNpcw== 9204
+YXNw 9205
+IGxvY2tlZA== 9206
+IFBlcm8= 9207
+IEhhcnJ5 9208
+IHR1ZG8= 9209
+IFRlbg== 9210
+2LU= 9211
+Zm9yY2VtZW50 9212
+KSk= 9213
+b2xp 9214
+IOyduA== 9215
+IHN1cHBs 9216
+IGNyb2NoZXQ= 9217
+IHBoZW5vbWVu 9218
+bG9z 9219
+YXRoYW4= 9220
+IFN1cHA= 9221
+IGVtYnI= 9222
+IGJlaw== 9223
+IFplaXQ= 9224
+Z2VuZA== 9225
+IHJvb21z 9226
+qr0= 9227
+VkVS 9228
+bnljaA== 9229
+IGRvbnQ= 9230
+IGNhYmlu 9231
+IGFjY291bnRz 9232
+IEVhc3Rlcg== 9233
+15XXnA== 9234
+IGZhY2lsaXRpZXM= 9235
+YmVpdA== 9236
+IGxpbmtlZA== 9237
+IEdlcg== 9238
+IHByb2dyYW1taW5n 9239
+b3RpYw== 9240
+IGRyYW1h 9241
+IDI5 9242
+IO2B 9243
+IGluc3RydWN0aW9ucw== 9244
+IGltcG9ydGFudGU= 9245
+IHdhdmVz 9246
+IGFpZA== 9247
+Q0s= 9248
+6rKg7Iq164uI64uk 9249
+IE1pcg== 9250
+IHRpZA== 9251
+IEhvdA== 9252
+IGFycmFuZ2U= 9253
+IEJhYnk= 9254
+IHRhY2s= 9255
+INGJ 9256
+7Z0= 9257
+IHZlcnRpY2Fs 9258
+IGhlZWw= 9259
+IEN1dA== 9260
+IG5hcnJvdw== 9261
+IEFyaQ== 9262
+IGtuZWU= 9263
+IEJyYXppbA== 9264
+IEZpdmU= 9265
+IHBvc3RlZA== 9266
+VUQ= 9267
+IHJvbGxpbmc= 9268
+zrg= 9269
+IGNsYWltcw== 9270
+IElucw== 9271
+T0s= 9272
+dWlu 9273
+IEluc3RpdHV0ZQ== 9274
+IGludGVuc2U= 9275
+aWFy 9276
+IE5pY2s= 9277
+IHNlbGVjdGlvbg== 9278
+IGxlZ2VuZA== 9279
+IHVuaWZvcm0= 9280
+w7pu 9281
+IHN0dWRpZWQ= 9282
+INCl 9283
+IOyVjA== 9284
+Z2Vycw== 9285
+IGRvdw== 9286
+IENT 9287
+IGFnZW50 9288
+IEF1Zg== 9289
+IGpvZw== 9290
+IGFpcmNyYWZ0 9291
+64uY 9292
+IHZpdA== 9293
+dWxz 9294
+IHNlZ21lbnQ= 9295
+IG9yZGVycw== 9296
+IENsYXNz 9297
+IGFwb2xvZw== 9298
+IHBsYXRmb3Jtcw== 9299
+IG15dGg= 9300
+0LDQttC1 9301
+IEJvb2s= 9302
+IHNlbnNpdGl2ZQ== 9303
+INC/0L7Qu9GD0Yc= 9304
+IGRhbWl0 9305
+IENhcHQ= 9306
+c29sZQ== 9307
+IGFyY2hpdGVjdHVyZQ== 9308
+IFdpbA== 9309
+IGluaGVy 9310
+Y2Fw 9311
+IEJveQ== 9312
+IGJ1cm5pbmc= 9313
+IFB1YmxpYw== 9314
+IGJlaGFsZg== 9315
+IOychA== 9316
+IHRoZXJhcHk= 9317
+dWJzY3JpYmU= 9318
+IGludm9sdmU= 9319
+IGV4cG9zZWQ= 9320
+acWf 9321
+w6p0cmU= 9322
+IHRvaWw= 9323
+IHNpbms= 9324
+cGly 9325
+5YM= 9326
+SUk= 9327
+IGFnZW5jaWVz 9328
+IHE= 9329
+IERvd24= 9330
+YXVm 9331
+IOunmw== 9332
+IHByb2M= 9333
+b2tlZA== 9334
+IHN0b3Jlcw== 9335
+cG93ZXI= 9336
+IFRoaW5ncw== 9337
+IGFjY2Vzc2libGU= 9338
+IHRlxbw= 9339
+IEVkdWM= 9340
+IHNwZWFrZXJz 9341
+IFNhcmFo 9342
+lJQ= 9343
+IGRpdmVyc2U= 9344
+7J6W 9345
+IFVsdA== 9346
+w6B5 9347
+IENoaWNhZ28= 9348
+U2hl 9349
+YXRoeQ== 9350
+IGVuYWJsZQ== 9351
+IHRyYWRpbmc= 9352
+IG11c2NsZXM= 9353
+5ps= 9354
+IENhcmU= 9355
+IFVy 9356
+IFNjb3Q= 9357
+IHBocmFzZQ== 9358
+RU5U 9359
+IOqyvQ== 9360
+IEphYw== 9361
+cGFjaw== 9362
+IGRldGVybWluZWQ= 9363
+w7xuZA== 9364
+IG5lZ290aQ== 9365
+IHZpZMOp 9366
+IHJveg== 9367
+IFN1cw== 9368
+IHJpZGluZw== 9369
+aG1lbg== 9370
+IERlZg== 9371
+IENyZQ== 9372
+IFdhbGw= 9373
+aWdhbg== 9374
+IHNlbXByZQ== 9375
+0ZbQtA== 9376
+IGRyaXZlbg== 9377
+IGZvb3RhZ2U= 9378
+IGZvbmQ= 9379
+IFdheQ== 9380
+w6Rt 9381
+IE9iYW1h 9382
+IFNlcnZpY2U= 9383
+IDc1 9384
+IERhcms= 9385
+IOq3vOs= 9386
+IENhdA== 9387
+2Lc= 9388
+6Yw= 9389
+IGp1Zw== 9390
+IGV0d2Fz 9391
+IGJyZWF0aGluZw== 9392
+4buD 9393
+IFdlYg== 9394
+IGZvaXM= 9395
+IGxpZ2h0aW5n 9396
+IERB 9397
+IG9ic3Q= 9398
+IGxldXI= 9399
+IEVneXB0 9400
+IEFybXk= 9401
+aWNpZGU= 9402
+0LDRgtC4 9403
+IOuLpOs= 9404
+IGFwYXJ0bWVudA== 9405
+IGNoaWVm 9406
+IFdlZA== 9407
+IG5ldHdvcmtz 9408
+IGJhdHQ= 9409
+5rg= 9410
+IEx1Yw== 9411
+IG5pY2VseQ== 9412
+IHZlcmI= 9413
+4Li0 9414
+7LY= 9415
+b3NpdA== 9416
+IHJldmVhbGVk 9417
+IHRhdA== 9418
+IHRpZWQ= 9419
+4buB 9420
+IGFuaW1hdGlvbg== 9421
+IHJvbGVz 9422
+7Iqk7Q== 9423
+IHZlcnNpb25z 9424
+0YfQuNGC 9425
+IHRhc2tz 9426
+r7w= 9427
+IHJlc2M= 9428
+c2hl 9429
+IGxvb3Nl 9430
+IGPhuw== 9431
+IGNvaXNh 9432
+IGFsZXJ0 9433
+IG5pbg== 9434
+IFNBTQ== 9435
+IHRyYWJhag== 9436
+aXJ1cw== 9437
+VEg= 9438
+xqE= 9439
+b2dldGhlcg== 9440
+IFRhaQ== 9441
+IGZpZ3VyZXM= 9442
+INeQ16o= 9443
+IGNyZWVw 9444
+IGludmVzdGlnYXRpb24= 9445
+IHJlY29tbWVuZGVk 9446
+IEFr 9447
+IHJlc2lkZW50cw== 9448
+0YHRgtCy0L4= 9449
+c2VjdA== 9450
+0LDQvdC40LU= 9451
+IG1pbmRz 9452
+dWluZw== 9453
+5bE= 9454
+b3dpbmc= 9455
+IG5vZw== 9456
+IHJheg== 9457
+2KfYsQ== 9458
+IHF1b3Q= 9459
+INC40YU= 9460
+IHNlZA== 9461
+IGFwcGxhdWQ= 9462
+IGNvdmVyYWdl 9463
+dm9s 9464
+IFJlYw== 9465
+xJs= 9466
+INCy0YHRkQ== 9467
+IGV4cGVjdGluZw== 9468
+IG9wZXJhdGU= 9469
+IGNvbnZlcg== 9470
+IFN1Y2g= 9471
+IFJhZA== 9472
+IFByaW1l 9473
+IHB1cnBsZQ== 9474
+IDIwMTA= 9475
+IOyViOs= 9476
+IGV4ZW0= 9477
+IGNvbXBhcmlzb24= 9478
+IGxhbmRzY2FwZQ== 9479
+IG5laXRoZXI= 9480
+IEVo 9481
+64U= 9482
+IHN0b21hY2g= 9483
+IGNhc28= 9484
+w6Ju 9485
+IHBlcmNlbnRhZ2U= 9486
+d2ljaA== 9487
+aXRhbg== 9488
+IGts 9489
+IGV4cGFucw== 9490
+INin2YTZhQ== 9491
+IG9jY2FzaW9u 9492
+cmV0cw== 9493
+aWduaW5n 9494
+IGtpbG9tZXQ= 9495
+IGd1c3Q= 9496
+Y3pl 9497
+IHVyYmFu 9498
+IGFncmlj 9499
+IGFzc2lzdGFuY2U= 9500
+IHN1cmY= 9501
+aW1ldGVy 9502
+IHBldGl0 9503
+IGFzc2Vzc21lbnQ= 9504
+IG1hbnVhbA== 9505
+IGltcHJvdmVk 9506
+YnN0 9507
+IHBpbG90 9508
+IE1hcnM= 9509
+IHZpZWxl 9510
+IENvbmdyYXR1bGF0aW9ucw== 9511
+IGFyZ3Vl 9512
+IHdpcmtsaWNo 9513
+IGNsaWNraW5n 9514
+UklT 9515
+IGxvZ28= 9516
+IG91dGNvbWU= 9517
+IENlbnRyYWw= 9518
+IEpp 9519
+IGdhbWluZw== 9520
+IGNvbnNlcnY= 9521
+IHVsdGltYXRl 9522
+IFZl 9523
+IFdhbA== 9524
+YXJv 9525
+c3Rhcg== 9526
+IGNvbnN1bWVy 9527
+IHRyYXZlbGluZw== 9528
+aW1lcg== 9529
+IDEwMDA= 9530
+0L3QuNC6 9531
+IHByaW5jaXBhbA== 9532
+IHNha2U= 9533
+0ZbQsg== 9534
+IG1vdXNl 9535
+YXJpb3M= 9536
+IHJlbGF0aW9u 9537
+IG1vcmFs 9538
+IHRoZXRh 9539
+d3k= 9540
+IGthbQ== 9541
+IGVpZw== 9542
+IGdvbGRlbg== 9543
+16Q= 9544
+IGFtcGw= 9545
+IHZ1 9546
+c3Ry 9547
+cm9ycw== 9548
+IHdoZXJlYXM= 9549
+aXphcg== 9550
+IGFkbWluaXN0cg== 9551
+IG7Ds3M= 9552
+IFByZXQ= 9553
+IEFjYWQ= 9554
+YW5naW5n 9555
+YmFnZQ== 9556
+w6l0YWl0 9557
+dXJp 9558
+IGhlYWxpbmc= 9559
+IHRpcG8= 9560
+IG1hcnJ5 9561
+0YPQsg== 9562
+IGVzdGF0ZQ== 9563
+dXU= 9564
+7JQ= 9565
+IEJlc3Q= 9566
+IHN1ZmZlcg== 9567
+IDE5NA== 9568
+IGJhY3Rlcg== 9569
+INCS0L7Rgg== 9570
+IE9t 9571
+IGR6 9572
+6LY= 9573
+7KY= 9574
+IG9sZHU= 9575
+IHBoeXNpY2FsbHk= 9576
+IExvdWlz 9577
+ZXRpbWU= 9578
+Y2FzZQ== 9579
+IHBpZXI= 9580
+7KCc 9581
+dmFu 9582
+IGFzc2V0cw== 9583
+IOuB 9584
+dmV0 9585
+0LjQsQ== 9586
+IHByb21vdGU= 9587
+IGNvbmdyYXQ= 9588
+dWVzZGF5 9589
+IGR1dHk= 9590
+IFZpZGVv 9591
+2K4= 9592
+IEpvaG5zb24= 9593
+a3Rpb24= 9594
+IFZvY8Oq 9595
+44CL 9596
+IGFp 9597
+IGFubnVhbA== 9598
+IEpvc2g= 9599
+aXR0ZQ== 9600
+IEpP 9601
+IHNsaWRlcw== 9602
+IGFuYw== 9603
+uYQ= 9604
+dGVlbg== 9605
+IGNhcnJ5aW5n 9606
+bHltcA== 9607
+ZWRpbmc= 9608
+IGZybw== 9609
+IGFkbWl0 9610
+cmVy 9611
+IG9mZmljaWFscw== 9612
+cHRpb25z 9613
+Z2Fs 9614
+IGhldXRl 9615
+IHZvaWNlcw== 9616
+IGJhbGxz 9617
+IGd1ZXN0cw== 9618
+YW5uZXI= 9619
+44CK 9620
+aXNoZXI= 9621
+IE1S 9622
+IFJpY2hhcmQ= 9623
+IHJvdWdobHk= 9624
+bMSx 9625
+IHZpY3Rvcnk= 9626
+IGFsZ3Vu 9627
+IE1ycw== 9628
+xZtjaWU= 9629
+IFVr 9630
+IGV5 9631
+IFdhcnM= 9632
+IGJyYW5jaA== 9633
+YXN0eQ== 9634
+IFByaW5jZQ== 9635
+0LXQutGC 9636
+IHJlY29nbml6ZWQ= 9637
+IG11Y2hv 9638
+IExlYXZl 9639
+Y29ubmVjdA== 9640
+IHNwZWxs 9641
+IHRvdWNoZWQ= 9642
+IGFnZW5kYQ== 9643
+6L4= 9644
+YXJpYQ== 9645
+IEtvbmc= 9646
+b2dh 9647
+IHBhcmFtZXRlcnM= 9648
+64uk6w== 9649
+IGluc3RhbnQ= 9650
+IHJlZ3Vs 9651
+Q29u 9652
+IGVkaXRvcg== 9653
+IERpc3Q= 9654
+IHVua25vd24= 9655
+IHB1bmlzaA== 9656
+IGV4cGVjdGF0aW9ucw== 9657
+IGNyeXB0 9658
+IGRpdmlkZQ== 9659
+YWtlbg== 9660
+IE1lc3M= 9661
+IGh5cGVy 9662
+IFByb2plY3Q= 9663
+aWtp 9664
+IGFnb3Jh 9665
+IGFidXNl 9666
+IGNhdXNpbmc= 9667
+IGNvbnZpbg== 9668
+IExB 9669
+IGNvbmNlbnRyYXRpb24= 9670
+IGJyZWFrcw== 9671
+dXJlcg== 9672
+IGNvbmNyZXRl 9673
+IGZvcm1hbA== 9674
+IGJldGE= 9675
+aXRvcnM= 9676
+IENoYW1w 9677
+IGhlYWRpbmc= 9678
+IEJsbw== 9679
+IHByZW5k 9680
+IFNlbmF0ZQ== 9681
+IGFkdmVudHVyZQ== 9682
+b3Nv 9683
+IG9wZW5z 9684
+IFBMQVlJTkc= 9685
+IFNV 9686
+dXJlbg== 9687
+aWt0 9688
+INC70Y7QsQ== 9689
+IEZvbGxvdw== 9690
+IEJpZGVu 9691
+ZWxu 9692
+IFNreQ== 9693
+ZXRpbmc= 9694
+IEV4dA== 9695
+0L3Rg9GO 9696
+IOyZnA== 9697
+IHNocg== 9698
+ZWxsYQ== 9699
+IERpdg== 9700
+IHRyYW5zZm9ybWF0aW9u 9701
+IGhvdXNlaG9sZA== 9702
+ZXRyeQ== 9703
+6KE= 9704
+IERlc3A= 9705
+IGNvdXJhZ2U= 9706
+IHBhcmtpbmc= 9707
+IGV0dMOk 9708
+Y2Fs 9709
+bHlu 9710
+IGxhaWQ= 9711
+IHRyaWVz 9712
+aXJ0cw== 9713
+aWdh 9714
+IHJlY2FsbA== 9715
+aWZpZXI= 9716
+z4HOsQ== 9717
+IGFhbg== 9718
+IGJ1dHRvbnM= 9719
+IHJlYWNoaW5n 9720
+IOq3vOuNsA== 9721
+IHNwYXJr 9722
+IFNvY2lhbA== 9723
+INC10YnQtQ== 9724
+IGNhbmFs 9725
+IGNyaXRlcg== 9726
+IGt0w7NyeQ== 9727
+IHRlbmVtb3M= 9728
+gqw= 9729
+INC90LXRgg== 9730
+IHR1YmU= 9731
+YWNsZXM= 9732
+0LjRiA== 9733
+IGRlxJ9pbA== 9734
+IHN0YW1w 9735
+IGluZmw= 9736
+IGFob3Jh 9737
+IHRyYWls 9738
+IG1peHR1cmU= 9739
+IFJvbGw= 9740
+IHJvdXRpbmU= 9741
+IGNvdW50eQ== 9742
+IGVuam95aW5n 9743
+0L3QvtGB0YLRjA== 9744
+ZXJlcw== 9745
+IHB1cnBvc2Vz 9746
+IFNhbnRh 9747
+IGJyZWFzdA== 9748
+w6RuZw== 9749
+IHdyaXRlcg== 9750
+5Yw= 9751
+0YDQvg== 9752
+IG5lbQ== 9753
+aWNvcw== 9754
+0LDRgdGC 9755
+IGRldGFpbGVk 9756
+IHJldmVyc2U= 9757
+IFJlYWR5 9758
+IGRpc3RyYWN0 9759
+IEFsb3Jz 9760
+dXR0ZXI= 9761
+IGRlc2VydmU= 9762
+IFJvbg== 9763
+0L3QvtC8 9764
+IG9ic2Vydg== 9765
+IGxvZ2lj 9766
+IFB5 9767
+IEtldmlu 9768
+pbQ= 9769
+2YrZhg== 9770
+IHNrYQ== 9771
+IHRhY3Q= 9772
+IGhvbGlkYXk= 9773
+IGJ1bXA= 9774
+INC80L7Qsw== 9775
+IGRlaXg= 9776
+7YU= 9777
+IHdvcnNoaXA= 9778
+Q2w= 9779
+IHN1Y2s= 9780
+INGB0LXQsQ== 9781
+IGFwcGxhdXNl 9782
+IEVw 9783
+INC80L4= 9784
+IHBhdGNo 9785
+4bqt 9786
+IGxhZGllcw== 9787
+IGJyb2FkY2FzdA== 9788
+IGlsbGVn 9789
+IG5hcnJhdGl2ZQ== 9790
+b3NzYQ== 9791
+QVJSQVRPUg== 9792
+IHNhbmc= 9793
+IG1vdmVtZW50cw== 9794
+IHBhcnRuZXJzaGlw 9795
+IG9yZ2FuaXplZA== 9796
+IG5vZGU= 9797
+ZXN0eWxl 9798
+IE1lZw== 9799
+IGluZHVzdHJpYWw= 9800
+IGdvbA== 9801
+IGJvcmluZw== 9802
+IGN1dHM= 9803
+IHJlY29u 9804
+YXNh 9805
+IGltcHJlc3Npb24= 9806
+7Jq0 9807
+Z2ll 9808
+TUE= 9809
+hrU= 9810
+IGVkaXRpbmc= 9811
+cm9udA== 9812
+IGZvbGxvd3M= 9813
+IEl0YWxpYW4= 9814
+0YDQvtC0 9815
+IOqwmeydgA== 9816
+IOuwqQ== 9817
+IHBhcnRpY2xlcw== 9818
+IEJvYXJk 9819
+15nXqg== 9820
+anVu 9821
+cm9uaWM= 9822
+IGVq 9823
+IM+Ezrc= 9824
+15XXkw== 9825
+Y2lvbg== 9826
+aXR0eQ== 9827
+IFR1ZXNkYXk= 9828
+dW1lcw== 9829
+IFByb3Q= 9830
+ZWRlcg== 9831
+IHBlc3NvYXM= 9832
+INC90L7Qsg== 9833
+IHNraXA= 9834
+IG9iamVjdGl2ZQ== 9835
+w61hcw== 9836
+IGRlc2s= 9837
+IExvb2tz 9838
+dW5kZW4= 9839
+IHByaW1hcmlseQ== 9840
+aW1lbnRv 9841
+IHJlcG9ydGluZw== 9842
+IGhhY2U= 9843
+IGNoZWNrZWQ= 9844
+6Zg= 9845
+IOuztOs= 9846
+IHNtZWxscw== 9847
+IGFjdG9ycw== 9848
+IEFzaWE= 9849
+aWzDoA== 9850
+IHJlY2VpdmluZw== 9851
+IHRheGVz 9852
+IGdyYWNl 9853
+IGNvbXBldGl0aXZl 9854
+IGRpdmlzaW9u 9855
+IGVzcGVy 9856
+IHdoZWVscw== 9857
+IGtvbW10 9858
+IHRyZW1lbmRvdXM= 9859
+IGVzcGU= 9860
+Li4uKQ== 9861
+IOyehQ== 9862
+IGxpc3RlZA== 9863
+w6RsbA== 9864
+IHVudXM= 9865
+IEhvbGx5 9866
+IGd1aWRhbmNl 9867
+IGN1Yg== 9868
+IGludGVsbGVjdA== 9869
+INCx0YvQuw== 9870
+IHJlZ2FyZGxlc3M= 9871
+IFN0YW4= 9872
+IGNvbmNsdXNpb24= 9873
+YWNhxJ8= 9874
+IGxvbA== 9875
+IEJhdA== 9876
+IG1hbmlmZXN0 9877
+IENoaWVm 9878
+IHNoYW1l 9879
+IG91dGNvbWVz 9880
+IG1haWw= 9881
+IGt1cg== 9882
+zrnOug== 9883
+ZXR6 9884
+IHByZXBhcmluZw== 9885
+Mjc= 9886
+IFF1ZWVu 9887
+4K6z 9888
+IOu5hA== 9889
+IHRpc3M= 9890
+IGNvbnNjaW91c25lc3M= 9891
+IHBhbnRz 9892
+IG1lbHQ= 9893
+dWNodA== 9894
+aW5o 9895
+7JuM 9896
+IHZvdHJl 9897
+IG1vZHVsZQ== 9898
+b3d5 9899
+IG1vbnN0ZXI= 9900
+IOuG 9901
+IGVsZWN0cm9uaWM= 9902
+IGNlbnRyZQ== 9903
+IHN0b3Bz 9904
+IHRvdQ== 9905
+IOut 9906
+IGxhbWI= 9907
+IGNvbnNlcXVlbmNlcw== 9908
+IHN0cmF3 9909
+IGltcGVy 9910
+IGV4dGVuZA== 9911
+IGFuc3dlcmVk 9912
+IE1haA== 9913
+IExBVVJB 9914
+aWZ0aW5n 9915
+dWF0ZQ== 9916
+IFVTQg== 9917
+IEFuZHJldw== 9918
+IEZyZWQ= 9919
+IERF 9920
+IEdlb3Jn 9921
+57s= 9922
+w6xuaA== 9923
+IGRyYXdu 9924
+IGxpcHM= 9925
+Ymly 9926
+IG1heW9y 9927
+aW1p 9928
+IGVuY29yZQ== 9929
+Zm9ydGFibGU= 9930
+dXJzZGF5 9931
+IEZvcm0= 9932
+IGJsYW1l 9933
+IHNob3dlcg== 9934
+IGNvbnRhaW5lcg== 9935
+c3RlcnM= 9936
+dWRlcw== 9937
+IFRheQ== 9938
+4Lil 9939
+IOyYiA== 9940
+IHZvbQ== 9941
+IGJhc3M= 9942
+IExhYg== 9943
+aXNzYQ== 9944
+IGRpbWVuc2lvbg== 9945
+IGV4ZWN1dGl2ZQ== 9946
+IFJvbQ== 9947
+6rKM7JqU 9948
+IERvY3Rvcg== 9949
+IGRlbGl2ZXJlZA== 9950
+IGdhbmc= 9951
+IGNlcg== 9952
+IHBpdA== 9953
+ZWxp 9954
+IGV4dHJhb3Jk 9955
+amFy 9956
+IGRlcml2 9957
+IGlsbG5lc3M= 9958
+IGd1bnM= 9959
+IDIwMTE= 9960
+IGFpcnBvcnQ= 9961
+0JU= 9962
+IGF0dGl0dWRl 9963
+IGdyYXQ= 9964
+IFdy 9965
+IE5BUlJBVE9S 9966
+IOyalA== 9967
+IHJlbmV3 9968
+IGNvc2E= 9969
+IGNvbnRyb2xsZWQ= 9970
+b21teQ== 9971
+b25kcw== 9972
+IGVzZQ== 9973
+w6RjaA== 9974
+IHZlbmQ= 9975
+ZGFt 9976
+IGFyZ3U= 9977
+IGFjY2VsZXI= 9978
+IG5haWw= 9979
+aWVuZQ== 9980
+7IOd 9981
+IGVuY29udA== 9982
+ZXNlYXJjaA== 9983
+6aE= 9984
+IGdvb2Rz 9985
+IGZpc2hpbmc= 9986
+QVBQTEFVU0U= 9987
+IE5BUw== 9988
+ZWN0aW9u 9989
+IHRlbXBsZQ== 9990
+bGljaGU= 9991
+IGtleWJvYXJk 9992
+IGRlc2Rl 9993
+IGVkdWNhdGlvbmFs 9994
+IE5pZ2h0 9995
+MzM= 9996
+IGJyZWF0aGU= 9997
+bGljaGVu 9998
+dGht 9999
+acOocmU= 10000
+4Lia 10001
+bGFyxLE= 10002
+IGFsaQ== 10003
+IGNvbXBvcw== 10004
+IHNlbnNvcg== 10005
+IOu2gOs= 10006
+IG5ld3Nw 10007
+IEJ1bmQ= 10008
+IE1p 10009
+IHBlcmZvcm1pbmc= 10010
+IGRydW0= 10011
+QkU= 10012
+IHBvcms= 10013
+IGNvYWw= 10014
+ZW5nZXI= 10015
+IHJhbQ== 10016
+IOuyiA== 10017
+0LjRgNC+0LI= 10018
+IFBvcA== 10019
+IHBob25lcw== 10020
+IGZhY2ls 10021
+IHRyYWNrcw== 10022
+b250ZQ== 10023
+IG9yZ2FuaWM= 10024
+IGRpYWxvZ3Vl 10025
+IEhhdmluZw== 10026
+IFBvc3Q= 10027
+IHBheW1lbnQ= 10028
+IGFycmF5 10029
+IGludGVuZGVk 10030
+w7pz 10031
+IGJhcnM= 10032
+IHJldmlld3M= 10033
+bGFuZHM= 10034
+IGtpbmdkb20= 10035
+IHN0YWdlcw== 10036
+IG1vdW50YWlucw== 10037
+IGR1bg== 10038
+IGRlY2ly 10039
+xI0= 10040
+IGJhbmtz 10041
+IHRocm93aW5n 10042
+IOuquw== 10043
+IGFuZ2Vy 10044
+INGB0LXQudGH0LDRgQ== 10045
+IGRpc3R1cg== 10046
+IGh1bWFuaXR5 10047
+IGVsZXM= 10048
+IHNob3VsZGVycw== 10049
+IFBlcmZlY3Q= 10050
+IGZhbmN5 10051
+IGJyaWxsaWFudA== 10052
+IGluc3BpcmF0aW9u 10053
+aG1t 10054
+IGxpZA== 10055
+VUw= 10056
+IG3DpQ== 10057
+aW5kaQ== 10058
+6Ig= 10059
+IHNoaWVsZA== 10060
+IOyYpOs= 10061
+Q1Q= 10062
+YWdpbmU= 10063
+dWJlcg== 10064
+IEJS 10065
+IHF1ZXN0bw== 10066
+INC30LDQug== 10067
+IEtub3c= 10068
+IHRhbmc= 10069
+7ZWp64uI64uk 10070
+IGJhcmVseQ== 10071
+IFNF 10072
+IG1hcmdpbg== 10073
+cmVp 10074
+0LDRgtC10LvRjA== 10075
+IGNvbnRy 10076
+IHbDoA== 10077
+IGxlZ2l0 10078
+0Jg= 10079
+a2lucw== 10080
+0YDQtdC0 10081
+IEFzaA== 10082
+IGFkdmlz 10083
+IEdyZWVr 10084
+0YPQug== 10085
+IHNoYWtl 10086
+aWRhZGVz 10087
+0LDRgdGM 10088
+IGNvbnZlbnRpb24= 10089
+IGNvbnRlc3Q= 10090
+TVM= 10091
+IFllYXI= 10092
+IHJlcHJlc2VudGF0aW9u 10093
+aW5kZW4= 10094
+ZW5kYXI= 10095
+IHByb3N0 10096
+IEh1bWFu 10097
+IEN5 10098
+YW5nZWQ= 10099
+UEE= 10100
+IGF4aXM= 10101
+IHRoZW9yZQ== 10102
+YXR6 10103
+IO2VmOqzoA== 10104
+IGVscw== 10105
+IFJlc2VhcmNo 10106
+IGJlbmVmaWM= 10107
+IGRlbnNpdHk= 10108
+aW5kbw== 10109
+7Jy8 10110
+aW1kaQ== 10111
+IHJlc2VhcmNoZXJz 10112
+6rGw65Og 10113
+aWdocw== 10114
+ZGFu 10115
+IGRpY2U= 10116
+IG1hYXI= 10117
+IHN1Ym1pdA== 10118
+IGR1bWI= 10119
+IGJpag== 10120
+YXdheQ== 10121
+IFBhc3M= 10122
+IGV4dGVuc2lvbg== 10123
+IGNydXNo 10124
+IGNvdmVyaW5n 10125
+ZWRp 10126
+Ym9ybg== 10127
+aW5hdGlvbnM= 10128
+INGB0LTQtdC7 10129
+0LLQtdGA 10130
+IE90aGVyd2lzZQ== 10131
+aXN0YW50 10132
+0LDQudGC0LU= 10133
+IHRhbnRv 10134
+IHBlcmZvcm1lZA== 10135
+INC30LDQvw== 10136
+YWxv 10137
+IEZvdW5kYXRpb24= 10138
+IHByb3RvY29s 10139
+IFpv 10140
+bWF5 10141
+IGhhY2s= 10142
+IGJ1ZGR5 10143
+bWFkZQ== 10144
+IGFkcw== 10145
+IGZhc2NpbmF0aW5n 10146
+IGVxdWl2YWxlbnQ= 10147
+Z2Vs 10148
+IGFyYw== 10149
+INGH0LXQu9C+0LI= 10150
+IHByb3Bvc2Vk 10151
+IG5vdHJl 10152
+YW5nZXM= 10153
+IGNvdW5zZWw= 10154
+YWxsYQ== 10155
+IDMx 10156
+d2VldA== 10157
+yJk= 10158
+IGVsZWN0cmljaXR5 10159
+IHRveA== 10160
+xYJhZA== 10161
+IOy0 10162
+IGRpZmZpY3VsdHk= 10163
+oNeZ 10164
+bmVzZGF5 10165
+0LjRgdGM 10166
+IGFsbGVn 10167
+IEdP 10168
+IHF1aXQ= 10169
+IEhlcnI= 10170
+IGVzdMOhbg== 10171
+IGdpcmxmcmllbmQ= 10172
+IHRlbmc= 10173
+aWZpY2lhbA== 10174
+IEphbQ== 10175
+IGNhbmNlbA== 10176
+IGZyZXF1ZW50bHk= 10177
+SVY= 10178
+IGNsb3Npbmc= 10179
+IGRlY2FkZQ== 10180
+IHJlcHJlc2VudGVk 10181
+IENhbmFk 10182
+INC60L7RgtC+0YDRi9C1 10183
+IGVzdGFtb3M= 10184
+IFRodXJzZGF5 10185
+IEdh 10186
+IExpdmU= 10187
+bGVt 10188
+YmJsZQ== 10189
+U09O 10190
+IDIwMDg= 10191
+IGRpY2g= 10192
+IEF3ZXNvbWU= 10193
+IGNvbmNlcHRz 10194
+UEVBSw== 10195
+IGxpdGVyYXR1cmU= 10196
+IE9seW1w 10197
+0LvQsNC0 10198
+IG5vc3Q= 10199
+dml0 10200
+IEVudGVy 10201
+b3JkZXJz 10202
+aWNraW5n 10203
+bmllag== 10204
+IGV1Y2g= 10205
+IFRob3VnaA== 10206
+IGJhZ3M= 10207
+IGxpbWl0cw== 10208
+IHN0YWtl 10209
+g6U= 10210
+IG9j 10211
+IFZpcw== 10212
+IDEyMA== 10213
+IG51ZQ== 10214
+IGNvbmNl 10215
+IGRpc2Fn 10216
+56g= 10217
+IGFudGljaXA= 10218
+oIg= 10219
+c2w= 10220
+IHZvdGluZw== 10221
+IGV4cG9zdXJl 10222
+IENvbW11bml0eQ== 10223
+IEp1c3RpY2U= 10224
+b3JuZXk= 10225
+c3p5c3Q= 10226
+IGZyaWVk 10227
+7Iuc6w== 10228
+IFdpbg== 10229
+IEA= 10230
+IEhvcGVmdWxseQ== 10231
+ZXN6 10232
+IG1vbmRl 10233
+IGNvbWJpbmU= 10234
+Z21lbnQ= 10235
+IHJlY29tbWVuZGF0aW9ucw== 10236
+IHByZWduYW50 10237
+7Iud 10238
+cmFm 10239
+IGx1 10240
+ZG9vcg== 10241
+0LDQt9GL0LI= 10242
+dWVnbw== 10243
+IGltcHJvdmVtZW50 10244
+IHRyaW0= 10245
+IGVpZ2Vu 10246
+IGFwcHJveGltYXRlbHk= 10247
+INCy0LDQvA== 10248
+YXdh 10249
+INGB0L7QsQ== 10250
+IGNvcm9u 10251
+IG9uZ29pbmc= 10252
+IGhlcw== 10253
+IGluanVyeQ== 10254
+IGZyYW5r 10255
+IGthZGFy 10256
+cmVuY3k= 10257
+IENvbG9y 10258
+IEdydQ== 10259
+IGRpcA== 10260
+0YDRiw== 10261
+IHRlYXJz 10262
+Z3Q= 10263
+IFBE 10264
+IHBhdXNl 10265
+b3Nj 10266
+IHVzdGVk 10267
+IFdvbw== 10268
+IHdpxJk= 10269
+IGRlbm4= 10270
+IFBldA== 10271
+IG92ZXJjb21l 10272
+IOuCtOqwgA== 10273
+IE1vdmU= 10274
+IGxpY2Vuc2U= 10275
+IHJlcGVhdGVk 10276
+4K+H 10277
+IGNhdGVnb3JpZXM= 10278
+IG5vb2RsZXM= 10279
+IGZsb29k 10280
+IE1hc3M= 10281
+IG51dHM= 10282
+IEplc3M= 10283
+IElo 10284
+IGNoYW5jZXM= 10285
+kJg= 10286
+IGRvbmRl 10287
+SUc= 10288
+IGFuZGVyZQ== 10289
+IGJvbmVz 10290
+7J6R 10291
+IGVmZmljaWVuY3k= 10292
+IG1vZGVy 10293
+cm9hdA== 10294
+IOydtOqyjA== 10295
+aWxsZXI= 10296
+IG9tZWdh 10297
+INC/0L7Qsg== 10298
+IEdyb3Vw 10299
+IHByb2R1Y2luZw== 10300
+YW1v 10301
+IHBhcnRpY2lwYW50cw== 10302
+dXBw 10303
+aWZpY2U= 10304
+IGZvcnR1bg== 10305
+aWV0bmFt 10306
+YWNhaw== 10307
+IEtv 10308
+bWnFnw== 10309
+IGphaWw= 10310
+IEpvbmVz 10311
+xZtteQ== 10312
+IERldXRz 10313
+IGJyaWVmbHk= 10314
+IFRhbA== 10315
+IFBlcmhhcHM= 10316
+IFJ1Yg== 10317
+IEtu 10318
+64uk64qU 10319
+csOp 10320
+IHZvY8Oqcw== 10321
+IENoYXJsZXM= 10322
+0LXRgtC1 10323
+cmllcnM= 10324
+IGhlYWw= 10325
+YW50ZWU= 10326
+IGRlbW9jcmFjeQ== 10327
+IGxvYW4= 10328
+IGNoZWY= 10329
+0Y/QvA== 10330
+IHVuY29tZm9ydGFibGU= 10331
+IGV0ZXJu 10332
+YXBwaW5n 10333
+IHJlcGFpcg== 10334
+cm90 10335
+IFRhcg== 10336
+IGNvdmVycw== 10337
+b21pbmc= 10338
+IEV0aA== 10339
+IM6t 10340
+0YfQvdC+ 10341
+IGFmdGVyd2FyZHM= 10342
+INCy0LXRgA== 10343
+IGRhaGE= 10344
+IGtuZWVz 10345
+IG9yZGluYXJ5 10346
+w7xs 10347
+Z2Fz 10348
+IHRpY2tldA== 10349
+IOyggOuKlA== 10350
+IOyeiOyKteuLiOuLpA== 10351
+Y2h0ZQ== 10352
+TXI= 10353
+IHNpc3Q= 10354
+aHVp 10355
+6re46w== 10356
+7Jes 10357
+IHZhcnk= 10358
+IG1lbW9y 10359
+IGNvbnRyb2xsZXI= 10360
+IGLEmWR6aWU= 10361
+IG1pbmlzdGVy 10362
+15I= 10363
+Zmxvdw== 10364
+QUg= 10365
+IHRvd2Vy 10366
+55A= 10367
+IHNjYXI= 10368
+IFBlbg== 10369
+IHBhw61z 10370
+15g= 10371
+7J246w== 10372
+IGVuZXJn 10373
+IHN3b3Jk 10374
+IHBhcGVycw== 10375
+0LjQu9Cw 10376
+IFdlZG5lc2RheQ== 10377
+IEZvcmNl 10378
+IGV4dHJhb3JkaW5hcnk= 10379
+IExha2U= 10380
+IOqwgOs= 10381
+IEJlYXV0 10382
+IHJlYXNvbmFibGU= 10383
+IGNvbnRyaWJ1dGU= 10384
+IHBsZWFzZWQ= 10385
+IHVwZGF0ZWQ= 10386
+IHBpw7k= 10387
+ZWxv 10388
+IHNpZ25pZmljYW50bHk= 10389
+IGJvdA== 10390
+IGdlbmVyYXRpb25z 10391
+IHByb3RlY3RlZA== 10392
+IGhpZGluZw== 10393
+IElsbA== 10394
+IG5ldXRyYWw= 10395
+XSw= 10396
+z4TOvw== 10397
+IHRvbmd1ZQ== 10398
+VGhhbms= 10399
+IOqzhA== 10400
+IHBheXM= 10401
+zq/OvQ== 10402
+IGFwcGxl 10403
+MDE= 10404
+ZXJr 10405
+aWVyYQ== 10406
+IGplZw== 10407
+IFN1YnNjcmliZQ== 10408
+IHRoZWF0ZXI= 10409
+IHN0cm9uZ2x5 10410
+IOyGjA== 10411
+INC/0YDQsNCy 10412
+dWNreQ== 10413
+IEppbg== 10414
+a3dhcmQ= 10415
+6rG0 10416
+IG9wcG9uZW50 10417
+IFNP 10418
+IGhvbHk= 10419
+IGZpbGxpbmc= 10420
+Ol0= 10421
+IGhpag== 10422
+0Jw= 10423
+IGJpc3M= 10424
+IGJsZW5k 10425
+IGltcGxpYw== 10426
+IOy9 10427
+bGxlaWNodA== 10428
+2YrYqQ== 10429
+YXNhbnQ= 10430
+ZXJ0ZQ== 10431
+IFNhbWU= 10432
+IGludGVyaW9y 10433
+U2U= 10434
+IGJlbmNo 10435
+IHBvY28= 10436
+IG1hcmtz 10437
+IHdpbnM= 10438
+IM6z 10439
+IGRpc3RpbmN0 10440
+IEFzaWFu 10441
+IG1vbGVj 10442
+IEphY2tzb24= 10443
+IGVhc3Q= 10444
+IHBoeXNpY3M= 10445
+aW1hbA== 10446
+IHBlYWs= 10447
+YXJpYW4= 10448
+ZXBz 10449
+IG5lYXQ= 10450
+INCy0LDRgQ== 10451
+dXJuaW5n 10452
+IHN5bnRo 10453
+IHJldmVhbA== 10454
+xbo= 10455
+Z29u 10456
+bmlz 10457
+YXRpdg== 10458
+IExhcw== 10459
+IHB5 10460
+IE1hamVzdHk= 10461
+IFZhbGxleQ== 10462
+IGVuZg== 10463
+IGdlbnM= 10464
+IHJvb3Rz 10465
+ZXpl 10466
+YmV0 10467
+IGFjdHM= 10468
+6Zo= 10469
+6JA= 10470
+IHBoaWxvc29waHk= 10471
+IG1hdGNoZXM= 10472
+nWk= 10473
+IGp1xbw= 10474
+IGRlc3Blcg== 10475
+IEVkdWNhdGlvbg== 10476
+IHNwb3Rz 10477
+IHJlZ2lvbnM= 10478
+QXI= 10479
+IE5hbQ== 10480
+ZWVu 10481
+IGRpYWdyYW0= 10482
+IHJlbHk= 10483
+IHRlbnM= 10484
+IGRhdGluZw== 10485
+IGNvYXQ= 10486
+IEhvcg== 10487
+IGFja25vd2xlZGdl 10488
+IFByZXR0eQ== 10489
+INC/0L7Qvw== 10490
+IHZvaXI= 10491
+IGZhdm91cml0ZQ== 10492
+IG1vxbw= 10493
+IGtt 10494
+IERP 10495
+IGZlcnQ= 10496
+IOuPhA== 10497
+IFBhYw== 10498
+IGZvbnQ= 10499
+IGZpbmRz 10500
+IEl0YWx5 10501
+INC60L7Quw== 10502
+IGNvbXBhc3M= 10503
+67M= 10504
+bGlhbWVudA== 10505
+IG5vdGlvbg== 10506
+IGluamVjdA== 10507
+IHdpc2RvbQ== 10508
+IMOc 10509
+IE1vb24= 10510
+IEJ1c2luZXNz 10511
+cmljcw== 10512
+IFlvdXQ= 10513
+IGZvcmdpdmU= 10514
+IGZpbmFuY2U= 10515
+aWxv 10516
+2KM= 10517
+YWhs 10518
+IGRlbW8= 10519
+IGNsaW1i 10520
+IGV4cG9ydA== 10521
+5aA= 10522
+IHN1Y2Nlc3NmdWxseQ== 10523
+IEZlcg== 10524
+cGVjdGVk 10525
+ZGVt 10526
+IHJldGlyZQ== 10527
+IGxhcHRvcA== 10528
+IHNwaXI= 10529
+IEFzc29jaWF0aW9u 10530
+INCz0Ls= 10531
+IFNlbA== 10532
+IO2VnOs= 10533
+IGVtcGxveWVl 10534
+IG1vbHQ= 10535
+Ukw= 10536
+0K8= 10537
+IGNvbnRyYQ== 10538
+IHVn 10539
+IEJhbGw= 10540
+IEphdmE= 10541
+w6lyaWU= 10542
+IHByb2NlZHVyZQ== 10543
+IGdyaWQ= 10544
+IOuKkOs= 10545
+IGJlbHQ= 10546
+INGN0YLQvtCz0L4= 10547
+dXJk 10548
+IGNvbXByZWg= 10549
+IGRldmVsb3Blcg== 10550
+INGN0YLQvtC8 10551
+5Zg= 10552
+Y3I= 10553
+IOuT 10554
+IHNwb2tlbg== 10555
+cmVuY2U= 10556
+IHRlcm1pbg== 10557
+IGFnZ3Jlc3NpdmU= 10558
+IGJpc3NjaGVu 10559
+IGhhc3Rh 10560
+IEJyaWFu 10561
+IENvbW1pc3Npb24= 10562
+IFl1 10563
+IHByb21pc2Vk 10564
+IGVxdWl0eQ== 10565
+aWtv 10566
+dmVydHk= 10567
+IHJlcGxhY2Vk 10568
+IEhlbHA= 10569
+IHBvc2U= 10570
+IE1pZGRsZQ== 10571
+IGtpbQ== 10572
+IG1laW4= 10573
+IENvdW5jaWxs 10574
+INCS0YE= 10575
+b3Jv 10576
+IEJlcm4= 10577
+IGJleg== 10578
+IGFuYWx5dA== 10579
+YW5nZW4= 10580
+IOyLtg== 10581
+IEdsbw== 10582
+IHF1YWQ= 10583
+0YLQsA== 10584
+IHNwZWFrcw== 10585
+7JiI7JqU 10586
+IOyXrOufrOs= 10587
+ZnJlZQ== 10588
+0L3Rlg== 10589
+cmljaA== 10590
+IOuvuA== 10591
+IERpZXM= 10592
+YWJi 10593
+pbg= 10594
+IGRlcHJlc3Npb24= 10595
+IHJldGFpbA== 10596
+hOuTpA== 10597
+IFZvdXM= 10598
+IExhdGlu 10599
+4bk= 10600
+IOyii+yVhA== 10601
+IHRvcnQ= 10602
+IGNvbXB1dGVycw== 10603
+IHNlYXJjaGluZw== 10604
+IHR1Yg== 10605
+YXRlbGw= 10606
+IG1lcmM= 10607
+IGdsYXNzZXM= 10608
+cGVyc29u 10609
+IGRpc2hlcw== 10610
+IGd1YXJhbnRlZQ== 10611
+IG1lZw== 10612
+c20= 10613
+IFdhbGs= 10614
+7Jy866m0 10615
+IGZvbGRlcg== 10616
+IE1pdA== 10617
+IHRpbWluZw== 10618
+IGFic3Q= 10619
+IExvZw== 10620
+IGFwcHJvdmVk 10621
+IFVTQQ== 10622
+0LLQtdGC 10623
+IHdpc2U= 10624
+ZXNzZWQ= 10625
+IGRvdWI= 10626
+IHJlc2lkZW50 10627
+IGdlbmVyYXRlZA== 10628
+IHN0YXlz 10629
+IGV4cGxhbmF0aW9u 10630
+IHBvaXNvbg== 10631
+YXRyZQ== 10632
+IGluc2FuZQ== 10633
+IHJlZmVycmVk 10634
+YWlyZXM= 10635
+IFRSQQ== 10636
+IHNlaQ== 10637
+IGlubm9j 10638
+QWg= 10639
+IG1hbnQ= 10640
+aHVz 10641
+IG91dGVy 10642
+Z2Vi 10643
+b2ljZQ== 10644
+IGRpc2N1c3Npbmc= 10645
+IGNvbnZlbmllbnQ= 10646
+X18= 10647
+IGF2b2ly 10648
+IHNoYXBlcw== 10649
+IGdyYXk= 10650
+IGRlbnRybw== 10651
+IG1hY2h0 10652
+IDE5NQ== 10653
+2Y8= 10654
+IGFkZHM= 10655
+dXRpbmc= 10656
+IGNhcGFiaWxpdGllcw== 10657
+IHNlY3Rpb25z 10658
+IHR1bmU= 10659
+IENhdXNl 10660
+YXJkZQ== 10661
+INGB0LrQsNC3 10662
+YXZpcnVz 10663
+IFJF 10664
+IHR1bmVk 10665
+IGxlYWY= 10666
+dGVyaW9y 10667
+IENhcHRhaW4= 10668
+INis 10669
+IGNob29zaW5n 10670
+aGlu 10671
+Z2dpbmc= 10672
+dmlldA== 10673
+IHJlZ3JldA== 10674
+MjY= 10675
+b25kZXJu 10676
+IGJvbnVz 10677
+IFJheQ== 10678
+QXM= 10679
+IHRvcm4= 10680
+IEhpZXI= 10681
+IEVV 10682
+IHJpc2tz 10683
+IGFtYQ== 10684
+IFlldA== 10685
+IGNoYXJhY3RlcmlzdGljcw== 10686
+IOqwkA== 10687
+IFNlbmF0b3I= 10688
+IFZhbW9z 10689
+IHJvc2U= 10690
+IGNvcnBvcmF0ZQ== 10691
+Z2hhbg== 10692
+IGNlbnRlcnM= 10693
+c3RhaXJz 10694
+IG5pdA== 10695
+IHVudXN1YWw= 10696
+IFRvbnk= 10697
+IEdS 10698
+IFdpbGQ= 10699
+IFNpbWlsYXI= 10700
+IHRvZGFz 10701
+IGhvcml6b250 10702
+bWVs 10703
+IHN0cmljdA== 10704
+IGN1YWw= 10705
+IHdyaXQ= 10706
+IGV4dGVuZGVk 10707
+IO2VmOuKlA== 10708
+IHJlbGllZg== 10709
+IG9uaW9u 10710
+IGJhYmllcw== 10711
+IGRpZmVy 10712
+IGludGVncmF0ZWQ= 10713
+w7x6aWs= 10714
+ZXBpbmc= 10715
+LS0tLQ== 10716
+IG1lbnM= 10717
+IHN0cmF0ZWdpYw== 10718
+ZmluaXRlbHk= 10719
+IGVpZ2VudGxpY2g= 10720
+V2hv 10721
+IHs= 10722
+IFRyaQ== 10723
+IHBvaW50ZWQ= 10724
+8J0= 10725
+bmFtZW50 10726
+0LXRhg== 10727
+IHByaWRl 10728
+IFJlcHVibGljYW4= 10729
+IHNhbXBsZXM= 10730
+IGRvbWVzdGlj 10731
+TFk= 10732
+dmV6 10733
+IHdlYmluYXI= 10734
+2KfZhQ== 10735
+IGVuaA== 10736
+IHN1Z2dlc3RlZA== 10737
+IG1laW5l 10738
+IHB1ZWQ= 10739
+b3Jlbg== 10740
+cmly 10741
+IGhlYXZpbHk= 10742
+IGluc3RydWN0aW9u 10743
+IG1pY3JvcGhvbmU= 10744
+IGlndWFs 10745
+IElyYQ== 10746
+IHZ1bG5lcmFibGU= 10747
+IFZpcmdpbmlh 10748
+IGNvbnRpbnVvdXM= 10749
+IHBvdmVydHk= 10750
+IGJsYWRl 10751
+IHJlbGF0ZQ== 10752
+IGNhcmE= 10753
+IEdvaW5n 10754
+IHJlZ2lvbmFs 10755
+IEZ1Y2s= 10756
+IHRvdw== 10757
+IE11c2V1bQ== 10758
+cmFudHM= 10759
+INCx0LXQtw== 10760
+bGFpbQ== 10761
+IGNoYW1waW9u 10762
+dGxl 10763
+w61u 10764
+ZW5jaWE= 10765
+IGRpZXNlbQ== 10766
+IERpZw== 10767
+bWF0ZXM= 10768
+IGludmVzdGluZw== 10769
+IEpvcmRhbg== 10770
+IGludGVncmF0aW9u 10771
+IO2O 10772
+4Lir 10773
+ZW5zdXM= 10774
+IEFyY2g= 10775
+IHBlbmNpbA== 10776
+0LDQu9GM0L3Qvg== 10777
+aXNzZW4= 10778
+IEth 10779
+IHJvY2tz 10780
+IHJhdGluZw== 10781
+IHJlZnVnZQ== 10782
+IGFwcg== 10783
+ZXRlZA== 10784
+IGFzc2lzdGFudA== 10785
+IG1lYW5pbmdmdWw= 10786
+IHBlcm1hbmVudA== 10787
+IGhpbGw= 10788
+IHdzenlzdA== 10789
+IHdvdW5k 10790
+IEF0bA== 10791
+IGxha2U= 10792
+IEZvcnQ= 10793
+IHJlZHVjdGlvbg== 10794
+IHZpdg== 10795
+IHNvdXI= 10796
+IGVjb3M= 10797
+IGhheg== 10798
+IHN0ZWFs 10799
+IG15c3Rlcg== 10800
+INCa0LDQug== 10801
+INGN0YLQuA== 10802
+IFZpZXRuYW0= 10803
+IGFudGVz 10804
+IGNvbm5lY3Rpbmc= 10805
+IERhdmU= 10806
+IGLDtnlsZQ== 10807
+IENhc3Q= 10808
+TGU= 10809
+IGN1bA== 10810
+IGdlbnJl 10811
+66eQ 10812
+IGNvbXBsYWlu 10813
+IGh1cnJ5 10814
+YXJ0ZQ== 10815
+Z3JlZw== 10816
+IG1vbml0b3Jpbmc= 10817
+IGRlc2VydA== 10818
+INGB0L7Qsg== 10819
+ZWxpbmc= 10820
+IFN1cHJlbWU= 10821
+IGdpYmk= 10822
+IGxhcmc= 10823
+IG5hdGlvbnM= 10824
+IFRvaw== 10825
+IG5lZWRsZQ== 10826
+5rU= 10827
+IGFzbGVlcA== 10828
+IGNvbXVu 10829
+IEpld3M= 10830
+IGFjaGlldmVk 10831
+IGV4aXQ= 10832
+IGRpc2Vhc2Vz 10833
+bGluZXM= 10834
+cmllbmRz 10835
+IHJlY3Q= 10836
+IHNjYW4= 10837
+IGh1cnRz 10838
+esSZ 10839
+IExvb2tpbmc= 10840
+7ZI= 10841
+dWx0dXJhbA== 10842
+4buT 10843
+aW5lbnQ= 10844
+IHB1ZXM= 10845
+IGNoZWVyaW5n 10846
+p4A= 10847
+YWdnZXI= 10848
+IGFkYQ== 10849
+TGF1Z2h0ZXI= 10850
+IFdvbWVu 10851
+6Ks= 10852
+IG9jY3VycmVk 10853
+IHNlYXRz 10854
+IGVtcG93ZXI= 10855
+dW51 10856
+ZWxsaW5n 10857
+QkVS 10858
+ZW5zaW9uYWw= 10859
+IGNvbnNvbGU= 10860
+YXNoaW5n 10861
+IGVpbm1hbA== 10862
+ZmFyZQ== 10863
+IOuPvA== 10864
+IHNlc3Npb25z 10865
+2ZA= 10866
+IHJpZGljdWxvdXM= 10867
+w61hbg== 10868
+IEhlbnJ5 10869
+IEhvbA== 10870
+IGNvbGxlY3RlZA== 10871
+IGRpc2N1c3Npb25z 10872
+RGU= 10873
+IGRpc2FiaWxpdHk= 10874
+IO2b 10875
+IHN1YnNjcmliZXJz 10876
+IGlyZ2VuZA== 10877
+IGZlbA== 10878
+IGRpcmVjdGlvbnM= 10879
+IG1hbnVmYWN0dXJpbmc= 10880
+IFJvZA== 10881
+IOyWmA== 10882
+4LiX 10883
+IGNyaXRlcmlh 10884
+IG1vbGQ= 10885
+IGVudGVyaW5n 10886
+cmlq 10887
+aXNlbg== 10888
+IFBhcmE= 10889
+aWV2ZQ== 10890
+IGNoYXJnZWQ= 10891
+IGpvdQ== 10892
+IGNhdHM= 10893
+0LvQtdC0 10894
+YWRheXM= 10895
+0LDQvdC+0LI= 10896
+asSZ 10897
+dmF0aW9u 10898
+IGFzdHJvbg== 10899
+aXRhbHM= 10900
+IEJyYW5k 10901
+IEthbg== 10902
+IHBsYWlu 10903
+IGFuZGVyZW4= 10904
+YW5kZQ== 10905
+0Y/Qtw== 10906
+IHRvbGVy 10907
+xYJlbQ== 10908
+IHByw6k= 10909
+0LzQvtGC0YA= 10910
+YWdlbWVudA== 10911
+dWN0 10912
+Y2jDqQ== 10913
+IEVuZXI= 10914
+YWrEhQ== 10915
+IO2VtOs= 10916
+IHN0YQ== 10917
+IHJpbmdz 10918
+IHRvaWxldA== 10919
+IENyYQ== 10920
+IGV4cGVyaWVuY2luZw== 10921
+IHNsaXA= 10922
+IHNhbmR3aWNo 10923
+IFVzaW5n 10924
+IHNwZWN0cnVt 10925
+IFJvcw== 10926
+YXBzZQ== 10927
+IEpheQ== 10928
+0LzRgw== 10929
+RXg= 10930
+IHJlY29nbml0aW9u 10931
+IERpZG4= 10932
+dWRh 10933
+YWpl 10934
+ZXN0bHk= 10935
+IGZlbWlu 10936
+aXR1cmU= 10937
+0YDQsNGC 10938
+IGhpcmU= 10939
+IG5vd2hlcmU= 10940
+4bqn 10941
+IHdpbmc= 10942
+IHNhdg== 10943
+IFNlY3VyaXR5 10944
+IHJ1cmFs 10945
+IEZ1bg== 10946
+YXllcg== 10947
+IGFjY3Vz 10948
+IG1t 10949
+IEpvc2VwaA== 10950
+IHNjcmVlbnM= 10951
+IGJvcnJvdw== 10952
+IHN3aW5n 10953
+IDQ4 10954
+IHRvdWNoaW5n 10955
+dGhpcw== 10956
+aW50ZW5kbw== 10957
+6YM= 10958
+0KA= 10959
+IFNjb3RsYW5k 10960
+IEphc29u 10961
+IFZlbg== 10962
+IGV4Y2VwdGlvbg== 10963
+IG5lYXJieQ== 10964
+IGJyb3dzZXI= 10965
+YW5nZXJz 10966
+IFNpbg== 10967
+z4DOvw== 10968
+b3NwZWw= 10969
+IHd1cmRl 10970
+IGRydW5r 10971
+7Zo= 10972
+7IaN 10973
+IOyKpO0= 10974
+IExpZQ== 10975
+b2Nv 10976
+IExlYWd1ZQ== 10977
+IGlnbm9yZQ== 10978
+IDop 10979
+IGxhbmRpbmc= 10980
+INi52YQ= 10981
+IFRhZw== 10982
+Mjg= 10983
+IGRyYWZ0 10984
+IGFlcg== 10985
+IOq3uOuDpQ== 10986
+IHBlbnNl 10987
+INC00LDQttC1 10988
+IGJlZHJvb20= 10989
+IG5hag== 10990
+7KeA6rOg 10991
+aWdlbm91cw== 10992
+IGRlYWxz 10993
+ZWxsbw== 10994
+IHBvc2l0 10995
+6rs= 10996
+IHZpc2l0ZWQ= 10997
+aWZpZXM= 10998
+IHByZW1p 10999
+IGNhbnQ= 11000
+IFJpY2s= 11001
+IHJhaXNpbmc= 11002
+IHBlcm1pc3Npb24= 11003
+IHB1Ymw= 11004
+dW5jaQ== 11005
+IGJlbmQ= 11006
+IGNoYW1waW9ucw== 11007
+ZGll 11008
+IGF3ZnVs 11009
+IGp1bXBpbmc= 11010
+IGxsZWc= 11011
+IHN1c3RhaW5hYmxl 11012
+IFRvdA== 11013
+IGNhbmR5 11014
+IHNhdGlzZmllZA== 11015
+IHBpcGU= 11016
+IGNvY2s= 11017
+2LY= 11018
+c3RvbmU= 11019
+IG1vbWVudHVt 11020
+INCd0LA= 11021
+IGFsb3Jz 11022
+IHJldHVybnM= 11023
+YW1tZW4= 11024
+564= 11025
+0YvQvA== 11026
+YXdu 11027
+b3R0ZWQ= 11028
+IHdvbGxlbg== 11029
+aWN0ZWQ= 11030
+IGNhbmRpZGF0ZXM= 11031
+IExhZHk= 11032
+IHlpZWxk 11033
+IG1haW50ZW5hbmNl 11034
+ZmZlY3Q= 11035
+IGV4cGFuc2lvbg== 11036
+IExFRA== 11037
+IGRhcmtuZXNz 11038
+IG91dGZpdA== 11039
+7JWI 11040
+INC40YHQvw== 11041
+cnVwdGlvbg== 11042
+IGVuZ2FnaW5n 11043
+IGluc2lnaHQ= 11044
+IEFsd2F5cw== 11045
+IGdlZg== 11046
+cmFr 11047
+IHBpeA== 11048
+IHF1YW50aXR5 11049
+IGluaw== 11050
+IEtpbmdkb20= 11051
+IGNvcnQ= 11052
+IGdvdmVybm1lbnRz 11053
+IHByb3Rlc3Q= 11054
+cG9vbg== 11055
+INGC0L7Qs9C+ 11056
+dWNoZW4= 11057
+cXVhbGl0eQ== 11058
+IFBvcnF1ZQ== 11059
+IENsdWI= 11060
+IHJpdA== 11061
+IGFydGljbGVz 11062
+Qkk= 11063
+aWdpYmxl 11064
+IGRpc2FzdGVy 11065
+0LjQsw== 11066
+INC90LjQug== 11067
+2YfYpw== 11068
+66W8 11069
+YXJldA== 11070
+IHVuYWJsZQ== 11071
+IMOu 11072
+IGVyc3Q= 11073
+INeg 11074
+dmFyZA== 11075
+IGFubm95aW5n 11076
+IEtpcg== 11077
+0LXRgNC2 11078
+ZW5uaXM= 11079
+IHVuY2VydGFpbg== 11080
+MzY= 11081
+w7Zz 11082
+IGVjb3N5c3RlbQ== 11083
+emVk 11084
+asOg 11085
+c3Vu 11086
+7Ja07ISc 11087
+IMW8ZWJ5 11088
+IG1hcHM= 11089
+64KY 11090
+IEp1c3Rpbg== 11091
+IHRyYXNo 11092
+IGVub3Jtb3Vz 11093
+IHN0YXRlZA== 11094
+IGJyYW5kcw== 11095
+IHlvdXQ= 11096
+INGH0LXQu9C+0LLQtdC6 11097
+IE1hdHRo 11098
+IHRyYW5zcG9ydGF0aW9u 11099
+IGxlZ2lzbGF0aW9u 11100
+IHByb3ZpZGVycw== 11101
+INit 11102
+IG1hZ2F6aW5l 11103
+IHNlaGVu 11104
+IERlc3BpdGU= 11105
+IHBhc3Nlcw== 11106
+IGFsdGVy 11107
+YWRhbg== 11108
+IGZhcm1lcnM= 11109
+IGNvbmZpcm1lZA== 11110
+IGVzYQ== 11111
+aXRvcw== 11112
+IHJvYWRz 11113
+VklT 11114
+IHdvcmtlcg== 11115
+IGRlc2lnbnM= 11116
+IFNvdmlldA== 11117
+YnJpZA== 11118
+IHByYWN0aWNpbmc= 11119
+IOu2gA== 11120
+IFNlYQ== 11121
+INC/0YDQvtC0 11122
+IGNoaWxs 11123
+IGxlbW9u 11124
+7JeQ64qU 11125
+IGZsZXhpYmxl 11126
+IEV4Y3VzZQ== 11127
+IHRlcnJpdG9yeQ== 11128
+IGx1eA== 11129
+IGxpZmV0aW1l 11130
+IGRpc3Rpbmd1 11131
+IFRpbWVz 11132
+IGdyb3Nz 11133
+ZW56 11134
+IHNjcm9sbA== 11135
+bcSxxZ8= 11136
+Y2lw 11137
+o7w= 11138
+RFA= 11139
+IHB1Ymxpc2g= 11140
+IGViZW4= 11141
+IHJlZ2lzdA== 11142
+IGVkaXRpb24= 11143
+IExF 11144
+IGNoYXJnaW5n 11145
+dXRhdGlvbg== 11146
+eXJpY3M= 11147
+aWRhcw== 11148
+IM6/ 11149
+INC60L7RgA== 11150
+IFRvbg== 11151
+IHdob2V2ZXI= 11152
+IEZveA== 11153
+6rGw65Og7JqU 11154
+IGZvdWdodA== 11155
+IGRyaWxs 11156
+IEFmZ2hhbg== 11157
+fiE= 11158
+IFRvbw== 11159
+IHNlY29uZGFyeQ== 11160
+csOk 11161
+IEhlYWQ= 11162
+aW5uZW4= 11163
+IHlhcm4= 11164
+INC90LDQvA== 11165
+IHdpZHRo 11166
+IGVuZ2luZWVy 11167
+acSF 11168
+IHdpbmdz 11169
+IOuVjOusuA== 11170
+IHRyYXVtYQ== 11171
+IHJlcHJvZHU= 11172
+IGNoaXA= 11173
+IHBhc3Npb25hdGU= 11174
+IGF3a3dhcmQ= 11175
+IO2K 11176
+0LDQttC0 11177
+IEJpdGNvaW4= 11178
+IGtow7RuZw== 11179
+IHLDsw== 11180
+cmVjdGlvbg== 11181
+INCz0LTQtQ== 11182
+IFVzdWFsbHk= 11183
+IGltcGxlbWVudGF0aW9u 11184
+IGdhbWVwbGF5 11185
+IG15c3Rlcnk= 11186
+INC+0Lo= 11187
+IGHDsW9z 11188
+YW5keQ== 11189
+0LjQvNC4 11190
+IHByaXZhY3k= 11191
+YWNv 11192
+IGR1bXA= 11193
+IFBheQ== 11194
+IGRpcGw= 11195
+IGZ1cm4= 11196
+IHNoaXBz 11197
+TEE= 11198
+INGF0L7RgNC+0Yg= 11199
+IGVj 11200
+IGRyb3Bz 11201
+Y2hs 11202
+MzI= 11203
+IG9ic2VydmU= 11204
+IERldmVsb3A= 11205
+TcO8emlr 11206
+YW5uZWw= 11207
+b3dhxIc= 11208
+IGZhY2Vk 11209
+w6Fs 11210
+IHZpY3RpbXM= 11211
+IGdpZnRz 11212
+IGJvb3Q= 11213
+w59l 11214
+cm9k 11215
+IDIwMDk= 11216
+w7ZydA== 11217
+IHVuaXZlcnNhbA== 11218
+IG5vdXZl 11219
+IGJveWZyaWVuZA== 11220
+IGNldGVyYQ== 11221
+0YHRgtCw 11222
+J1M= 11223
+IG5pdmU= 11224
+IGNydWNpYWw= 11225
+IHN1cnZl 11226
+IGNvaW4= 11227
+IHNvbmRlcm4= 11228
+IHNoYWRl 11229
+IGx1Z2Fy 11230
+IHN1cmVseQ== 11231
+IG1heA== 11232
+IGltcHJvdmluZw== 11233
+IHdlbg== 11234
+INeR 11235
+IOyWtOw= 11236
+IGVuZm9yY2VtZW50 11237
+aWJs 11238
+IGxpdg== 11239
+bGVyaQ== 11240
+IG1lam9y 11241
+IENhcm9saW5h 11242
+IHZhcw== 11243
+IGNvbXByb20= 11244
+IGRpcnQ= 11245
+IHVwZ3JhZGU= 11246
+IEJlbGw= 11247
+IHJlc3RhdXJhbnRz 11248
+IHRyYXA= 11249
+IHRlYXM= 11250
+YmxpYw== 11251
+IEdyZWc= 11252
+c2Fu 11253
+IG93 11254
+dWVzdA== 11255
+IHByb3Bvc2Fs 11256
+IFJldA== 11257
+ZnJvbnQ= 11258
+IHBhc3NhZ2U= 11259
+IHN1cnJvdW5kaW5n 11260
+IMO6bHQ= 11261
+IHVwY29taW5n 11262
+IGhvcnJvcg== 11263
+IGNsb3RoaW5n 11264
+IOyVvQ== 11265
+IGRpbA== 11266
+cm9tZQ== 11267
+IElk 11268
+IFJvYWQ= 11269
+INGN0YLQvtGC 11270
+Y2hhaW4= 11271
+INCx0YvRgtGM 11272
+IE9mZmlj 11273
+INCd0LU= 11274
+IGluc2Fu 11275
+IGRlY3JlYXNl 11276
+INGF0L7Rgg== 11277
+YnVpbGQ= 11278
+IERyYWdvbg== 11279
+5YI= 11280
+IGludmVzdG9ycw== 11281
+YW50aQ== 11282
+IHNhY3JpZmljZQ== 11283
+IHRyb29wcw== 11284
+IEJhZA== 11285
+IHBhc3N3b3Jk 11286
+IGNvbnN0cmE= 11287
+4LiV 11288
+IMOHYQ== 11289
+YWRvdw== 11290
+dGhyb3VnaA== 11291
+0YbQsA== 11292
+Q2Fu 11293
+IGNhbmRpZGF0ZQ== 11294
+IGFudGli 11295
+7JeF 11296
+IHRhc3R5 11297
+2YjZhg== 11298
+IEluZg== 11299
+IEJhbmc= 11300
+acOfdA== 11301
+aW5pdHk= 11302
+ZmF0aGVy 11303
+IGNvbnRyb3ZlcnM= 11304
+IFBhaw== 11305
+aWx0eQ== 11306
+6rWs6w== 11307
+IGxpZ2h0ZXI= 11308
+IGZhbGxlbg== 11309
+IHp1cw== 11310
+IEd1YXJk 11311
+IGNvdHQ= 11312
+IEZyZWU= 11313
+IGluaXRpYXRpdmU= 11314
+YWxvdXM= 11315
+IG5vdGlmaWNhdGlvbg== 11316
+IE1lZGlj 11317
+IENvbW1pdHRlZQ== 11318
+7Jew 11319
+IFdvb2Q= 11320
+IG11c2g= 11321
+dWx1bQ== 11322
+6LI= 11323
+YWhhaA== 11324
+IHN1ZmZpY2llbnQ= 11325
+IHNpbmdlcg== 11326
+0LrQvtC5 11327
+QUxJ 11328
+w6R0dA== 11329
+IFBS 11330
+IExhcg== 11331
+Y3VsZXM= 11332
+aWVtcG8= 11333
+IHVuZXg= 11334
+IGludGVncmFs 11335
+IHRyYW5zbWlzc2lvbg== 11336
+IGljaQ== 11337
+0YPRhQ== 11338
+Z2lj 11339
+IE5pbnRlbmRv 11340
+IENvcA== 11341
+IFRydXN0 11342
+ZW5hcw== 11343
+IGFiaWxpdGllcw== 11344
+IGNoaXBz 11345
+cGF0 11346
+IGFuY2hl 11347
+0LvQtdC9 11348
+IGFwcHJvYWNoZXM= 11349
+IHRob3I= 11350
+IHNpc3RlcnM= 11351
+IGRyaXZlcnM= 11352
+IGFsbGE= 11353
+MDM= 11354
+IHJ1YmJlcg== 11355
+IG7DpQ== 11356
+QUNL 11357
+IGRpc2FwcGVhcg== 11358
+6rCc 11359
+IGNvbXBlbnM= 11360
+IHZpYnI= 11361
+R08= 11362
+IHNpemVz 11363
+IHRyYWNraW5n 11364
+7ZmU 11365
+IOyEuA== 11366
+IGltcGFjdHM= 11367
+aWJpbA== 11368
+ZmlzaA== 11369
+QlI= 11370
+IGFycm93 11371
+IGxhcmdlbHk= 11372
+YW5ueQ== 11373
+IGxhd3llcg== 11374
+am91cnM= 11375
+2ro= 11376
+dmlh 11377
+IGRlbGxh 11378
+IG1hdGhlbWF0 11379
+IE1pbmU= 11380
+IEtvbGw= 11381
+2LI= 11382
+IENyb3Nz 11383
+IDY1 11384
+IGdyYWM= 11385
+IGludm9sdmVz 11386
+IGRlbGlnaHQ= 11387
+IEhvbGx5d29vZA== 11388
+IGltbWVkaWF0ZQ== 11389
+b25pYw== 11390
+IGxhZG8= 11391
+IGJ1bGxldA== 11392
+IE5vdGU= 11393
+IHVubG9jaw== 11394
+IGRpc2NvdW50 11395
+IHJpc2luZw== 11396
+cHJlc3M= 11397
+IHBhY2U= 11398
+IHNob3J0ZXI= 11399
+IHRlbmVy 11400
+Z2Vvbg== 11401
+IG1hbmFnaW5n 11402
+IGNlcmU= 11403
+Q2hy 11404
+V2hlbg== 11405
+YWNoZW4= 11406
+IOyT 11407
+IEh1bg== 11408
+IG9mdA== 11409
+IDI1MA== 11410
+aWVydW5n 11411
+IHN0YWJpbA== 11412
+IENvbm5lY3Q= 11413
+IHlhbmk= 11414
+IGRvd250 11415
+zrzOsQ== 11416
+IHZvY2Fs 11417
+zr3OsQ== 11418
+IGxlYW4= 11419
+IHZpZMOpbw== 11420
+IEZhbWlseQ== 11421
+cmVzZW50 11422
+IGFtb3VudHM= 11423
+7KeB 11424
+Y2xhc3M= 11425
+IHZpYg== 11426
+IEF2 11427
+YXJzZQ== 11428
+IGdlbnRsZW1lbg== 11429
+IHNlZWtpbmc= 11430
+IHVuaW9u 11431
+IHJlZ3VsYXJseQ== 11432
+5o8= 11433
+IEphaHI= 11434
+IEZvb2Q= 11435
+IFByb2JsZW0= 11436
+IGFydGlmaWNpYWw= 11437
+IFNpeA== 11438
+IGltcHJlc3NlZA== 11439
+IHRvb3Ro 11440
+IEto 11441
+IHlhcmQ= 11442
+IO2VtA== 11443
+IG93bmVk 11444
+IOuPmQ== 11445
+7LKt 11446
+IHRvZGE= 11447
+IHBvcnRmb2w= 11448
+IOuCqA== 11449
+b3JnZW91cw== 11450
+IGRhdGVz 11451
+0L7Qu9GM0Lc= 11452
+0LXRh9C90L4= 11453
+IGNvbmZpZ3VyYXRpb24= 11454
+IHJlcXVpcmVtZW50 11455
+IGJlbGx5 11456
+IHBhaW5mdWw= 11457
+IGRlbW9uc3RyYXRl 11458
+IGdsZWljaA== 11459
+IHZpc2l0aW5n 11460
+IENvbmY= 11461
+IGRhbA== 11462
+2ZE= 11463
+IGFtZW5k 11464
+IEZ1cg== 11465
+IHZpdGFs 11466
+4buL 11467
+IG1hdGU= 11468
+IE91 11469
+IGxlZ2FjeQ== 11470
+dXN0aW5n 11471
+IGFjY29tbW9k 11472
+IHF1b2k= 11473
+YXVlbg== 11474
+IGxpZmVzdHlsZQ== 11475
+Q0M= 11476
+w6TDpG4= 11477
+YXJ0ZW4= 11478
+IG1pbmhh 11479
+csOz 11480
+IOuqqA== 11481
+IGZvcm1hdGlvbg== 11482
+IHRyYWlsZXI= 11483
+cGVyb3I= 11484
+INCz0YA= 11485
+IHVk 11486
+enU= 11487
+IGtvbW1lbg== 11488
+IGNhdmU= 11489
+IENvdW5jaWxsb3I= 11490
+IHRocm93bg== 11491
+IHRyaWNrcw== 11492
+TEFVR0hURVI= 11493
+IEFjYWRlbXk= 11494
+cm93ZA== 11495
+oZ0= 11496
+7KCA 11497
+IEltYWdpbmU= 11498
+IGluZm9ybWVk 11499
+cm9waA== 11500
+IGxpZw== 11501
+IHNrdWxs 11502
+YWJldGg= 11503
+IGZ1bmN0aW9uYWw= 11504
+ZXJlaw== 11505
+T24= 11506
+6aY= 11507
+IGFuY2VzdA== 11508
+IHNhZmVseQ== 11509
+IEhU 11510
+64u5 11511
+IGRhdg== 11512
+IGRyaXZlcw== 11513
+QW1lcmlj 11514
+IHRpcmU= 11515
+IHNhaXM= 11516
+w6FyaQ== 11517
+YXZvcnM= 11518
+IGNvcnJlc3BvbmRpbmc= 11519
+IHByw6lz 11520
+Y2hlc3Q= 11521
+IGJhY3Rlcmlh 11522
+IGluZmVjdGlvbg== 11523
+dXNhbA== 11524
+IGF2ZXo= 11525
+IGJhc2tldGJhbGw= 11526
+IHN1cHBsaWVz 11527
+IGV4cGVydGlzZQ== 11528
+oKU= 11529
+ZmE= 11530
+IHRpZW1wbw== 11531
+IFdhbnQ= 11532
+IHNpbGx5 11533
+IHVwcA== 11534
+IGVsZWN0ZWQ= 11535
+IGZpcmVk 11536
+INiv 11537
+IHVuaXZlcnNpdGllcw== 11538
+YWxsZQ== 11539
+IGphY2tldA== 11540
+wrA= 11541
+IHRyYXY= 11542
+bHM= 11543
+IGRlZmVhdA== 11544
+IGNvZ24= 11545
+IGVxdWF0aW9ucw== 11546
+dWtp 11547
+IFNoZXI= 11548
+IHRoaXJ0eQ== 11549
+IHN0cmVhbWluZw== 11550
+b3Ryb3M= 11551
+IFByb2R1 11552
+bmVq 11553
+IGRlc2lnbmVy 11554
+IOuKkOuC 11555
+IHBhaW50ZWQ= 11556
+cmFpbmU= 11557
+bWFpbA== 11558
+IHZlc3M= 11559
+IHJoeXRobQ== 11560
+bGVzaA== 11561
+IDk5 11562
+IGFpbmRh 11563
+Y2hpZWQ= 11564
+IMOpdGFpdA== 11565
+IGFmZmVjdHM= 11566
+6aM= 11567
+IEZpbmQ= 11568
+IMOpbA== 11569
+IHBvdGF0b2Vz 11570
+IHBhZw== 11571
+INC/0LDRgA== 11572
+YXJ0cw== 11573
+IE5hY2g= 11574
+IDMz 11575
+IEhhcmQ= 11576
+IElyYXE= 11577
+IG9waW5pb25z 11578
+d2l0aA== 11579
+ZXJtYW4= 11580
+w70= 11581
+6K0= 11582
+IFNQRUFL 11583
+rLw= 11584
+IHN0YWJpbGl0eQ== 11585
+IEhF 11586
+IGNhcHR1cmVk 11587
+IGJ1Y2tz 11588
+IG1hc2tz 11589
+IGNvbXBldGU= 11590
+IGZvcmdvdHRlbg== 11591
+0LvRjtGH 11592
+c2VxdQ== 11593
+IOyEoA== 11594
+aWxsaW9u 11595
+IGdyYXBoaWNz 11596
+IGh1Yg== 11597
+IOyXsA== 11598
+ZW1wb3I= 11599
+IGNyb3du 11600
+IHdpZGVy 11601
+IG9jY3Vycw== 11602
+RFM= 11603
+5oE= 11604
+IEJhdHRsZQ== 11605
+IGR1YWw= 11606
+IDYwMA== 11607
+YXRoZXJz 11608
+4LmB 11609
+IHNldHRsZQ== 11610
+IGF2YWl0 11611
+IGRvaXM= 11612
+0LrQuNGF 11613
+YWRvcmVz 11614
+IMOz 11615
+bmVnbw== 11616
+IEdlb3JnaWE= 11617
+IFJvZw== 11618
+IGRpdm9y 11619
+IFNvbmc= 11620
+IFNwZWNpYWw= 11621
+IG11bg== 11622
+IHByZXRlbmQ= 11623
+TUFO 11624
+IHZpb2xlbnQ= 11625
+IGJlc2lkZXM= 11626
+dnk= 11627
+IE5heg== 11628
+Mjk= 11629
+IHN3ZWF0 11630
+IHp3 11631
+IEh1 11632
+IEJ1aWxk 11633
+IGhvcm0= 11634
+IENhcmQ= 11635
+IOycoA== 11636
+IHJlY29tbWVuZGF0aW9u 11637
+Y2FsbGVk 11638
+c3RpY2s= 11639
+IFBvbGljZQ== 11640
+IGNvbnN1bWVycw== 11641
+IGdyb2Nlcg== 11642
+IHN0dW4= 11643
+INCS0Ys= 11644
+0KM= 11645
+IERhdGE= 11646
+IHN1YnN0YW50 11647
+aXJlY3Q= 11648
+4LI= 11649
+INCy0Lc= 11650
+IEFybQ== 11651
+IHNlbWVzdGVy 11652
+IEJyYWQ= 11653
+IG91cnM= 11654
+INC60L7RgtC+0YDRi9C5 11655
+p2E= 11656
+IGdyYW1z 11657
+IGV4ZXJjaXNlcw== 11658
+NzU= 11659
+IHN3ZWFy 11660
+IGluY2VudA== 11661
+z4HOvw== 11662
+IGlsbGVnYWw= 11663
+IERhbW4= 11664
+IG7Dug== 11665
+IG5lY2Vz 11666
+IGN1eg== 11667
+aWNvbg== 11668
+IGhvcnM= 11669
+IENvbW8= 11670
+IOuRkA== 11671
+IG92ZXJzZQ== 11672
+IGhhcnZlc3Q= 11673
+IHRocmV3 11674
+INC/0L7RgtC+0LzRgw== 11675
+15nXlA== 11676
+IG90cm8= 11677
+INC/0LXRgNCy 11678
+IHNjb3Bl 11679
+IGdsb3J5 11680
+IE1pY2hpZ2Fu 11681
+IGFzc3VtaW5n 11682
+INGD0LQ= 11683
+IGJvbGQ= 11684
+Z3Vl 11685
+bW90aGVy 11686
+IHdhcmVu 11687
+INil 11688
+IEthbQ== 11689
+aXNwaWVs 11690
+IHRvdWpvdXJz 11691
+IGNvbnN0aXR1dGlvbg== 11692
+IH4= 11693
+IGZyYW5rbHk= 11694
+b2xlbg== 11695
+b25zY2lvdXM= 11696
+IHfDvHJkZQ== 11697
+dGhvbg== 11698
+IE9G 11699
+7J6Q6w== 11700
+dW5kYQ== 11701
+INC/0L7RgA== 11702
+IGVtcGxveW1lbnQ= 11703
+0ZHRgg== 11704
+IHN0ZWFt 11705
+IERJ 11706
+IHByb2Zlc3Npb25hbHM= 11707
+IGVuZ2luZWVycw== 11708
+IFhpYQ== 11709
+56s= 11710
+7JiB 11711
+IER1bg== 11712
+IGJpdGNo 11713
+IEZvcmQ= 11714
+c2VjdGlvbg== 11715
+IHZpY2U= 11716
+IExhdGVy 11717
+b3N0b24= 11718
+0Y3Rgg== 11719
+16Y= 11720
+IEF6dXJl 11721
+cGxpbmc= 11722
+IDE4MA== 11723
+IENyZWF0 11724
+SVNIQQ== 11725
+IGJ1ZW5v 11726
+7Z2s 11727
+cnVw 11728
+bGVycw== 11729
+IFlhbmc= 11730
+IEhB 11731
+YmF0 11732
+IENhdGhvbGlj 11733
+IGFjY2VudA== 11734
+IG1peGluZw== 11735
+Y2tldHM= 11736
+IGVuaGFuY2U= 11737
+w7xocg== 11738
+w6pz 11739
+IO2W 11740
+IHN3aW1taW5n 11741
+IGPhu6dh 11742
+IEVsaXo= 11743
+IGltbXVuZQ== 11744
+INCx0L7Quw== 11745
+IGZhcmU= 11746
+IEdhYg== 11747
+16E= 11748
+IHNhdGVsbA== 11749
+IEFueXRoaW5n 11750
+IGFzc2V0 11751
+IHNjaGVkdWw= 11752
+IHJhZGljYWw= 11753
+IHp3ZWk= 11754
+IE1F 11755
+cmVsYXRlZA== 11756
+IHNlcGFyYXRlZA== 11757
+IExpYnI= 11758
+IGdyaXA= 11759
+IOCuqg== 11760
+IGJlYW5z 11761
+IE9w 11762
+7IaM 11763
+IHBvdW5k 11764
+IGVudHJhbmNl 11765
+z4Y= 11766
+IE5pZQ== 11767
+IFJlcHVibGljYW5z 11768
+IGF0b20= 11769
+IHBlcnNvbmFz 11770
+IEFsaQ== 11771
+w6Rocg== 11772
+IGRyYW1hdGlj 11773
+IEZpbmU= 11774
+IHJlbWluZHM= 11775
+6Jk= 11776
+IGTDqWrDoA== 11777
+IGFmZm9yZGFibGU= 11778
+IGJyYW4= 11779
+aWVybw== 11780
+YWxhcg== 11781
+Y3U= 11782
+IFw= 11783
+IG1vxbxl 11784
+IGJsYXN0 11785
+IHJlY3k= 11786
+ZmlyZQ== 11787
+IGxsZQ== 11788
+INCy0YDQtdC80Y8= 11789
+IFdX 11790
+IHZz 11791
+IER1ZGU= 11792
+IFJvbWU= 11793
+IGdyZWV0 11794
+IEhldA== 11795
+Y2lhcw== 11796
+IOuLuQ== 11797
+bGVzc2x5 11798
+IHByZW1pdW0= 11799
+IGV4cGVyaW1lbnRz 11800
+YXRhcg== 11801
+w6lyaQ== 11802
+IG9mZmljaWFsbHk= 11803
+IGZlZQ== 11804
+4LmH 11805
+INGH0LXQvA== 11806
+cmVh 11807
+IHRveQ== 11808
+T1A= 11809
+IFRheWxvcg== 11810
+IE1jQw== 11811
+aWxleQ== 11812
+IEJhaw== 11813
+IGFsdW0= 11814
+IFVudGVy 11815
+IG1hZ2ljYWw= 11816
+IHRyYWJhbA== 11817
+IHZvdGVz 11818
+aXRhZ2U= 11819
+IG1lY2hhbmljYWw= 11820
+aG4= 11821
+INC40L3RgtC10YA= 11822
+IGhpbnQ= 11823
+IGF1dGhvcml0aWVz 11824
+IE5BU0E= 11825
+aXZlcnNhcnk= 11826
+INC/0L7Rhw== 11827
+cmFj 11828
+IFNQRUFLRVI= 11829
+w7Z0 11830
+IGZyYW1lcw== 11831
+IGdvb2RieWU= 11832
+IGNoZXI= 11833
+aHU= 11834
+IG5ldXI= 11835
+IE1hY2g= 11836
+IEhlbGw= 11837
+IGZlc3RpdmFs 11838
+64WE 11839
+dXRh 11840
+IG11c2hyb29t 11841
+IHRhbnQ= 11842
+IHRhdHRv 11843
+IGRlbGV0ZQ== 11844
+IGRpeg== 11845
+IHbDpA== 11846
+IHNldmVudA== 11847
+IFF1aWNr 11848
+IGJha2luZw== 11849
+IGFzc2VtYmx5 11850
+R28= 11851
+IERyZWFt 11852
+IExhZA== 11853
+w6J5 11854
+YWdz 11855
+IGdyYXZpdHk= 11856
+IOynkQ== 11857
+ZW1wbG95 11858
+IGRpZXNlcw== 11859
+IGRpc2NvdmVyeQ== 11860
+0YHRgtCy0LA= 11861
+IGhlYmJlbg== 11862
+IGdlcmFkZQ== 11863
+IERS 11864
+ICcn 11865
+IHRlY2huaWNhbGx5 11866
+INCf0L4= 11867
+IHByaXZpbGVnZQ== 11868
+IEV2ZXI= 11869
+IFNlcnZpY2Vz 11870
+dXJhbg== 11871
+IGNvbnN1bXB0aW9u 11872
+IFJldg== 11873
+IFNoYWxs 11874
+YXNzZXI= 11875
+toDthLA= 11876
+IHJhY2lhbA== 11877
+IFlvdXR1YmU= 11878
+IFByYQ== 11879
+0YHRgtCy0LXQvQ== 11880
+Y2Vr 11881
+5rQ= 11882
+YXNoYQ== 11883
+INuB 11884
+kZw= 11885
+YWhu 11886
+SUNL 11887
+IGRyaW5rcw== 11888
+IGNhcmI= 11889
+IDY0 11890
+IE1tbQ== 11891
+IGVsZWN0cmljYWw= 11892
+IGZydWl0cw== 11893
+IEhE 11894
+w7Fh 11895
+IERlZmluaXRlbHk= 11896
+IOuwmw== 11897
+IGZhaXM= 11898
+cmF0aW9ucw== 11899
+IGNvZQ== 11900
+YWh1 11901
+IEZhaXI= 11902
+IGVhdGVu 11903
+IGZpcg== 11904
+IEF1 11905
+0YPQvQ== 11906
+dWxhdGluZw== 11907
+aW5nbHk= 11908
+IHZhY2NpbmVz 11909
+IGRyYWdvbg== 11910
+IHBvaW50aW5n 11911
+IHBlbG8= 11912
+b3J0ZXJz 11913
+IHdvcmtvdXQ= 11914
+0LjQvNC10YA= 11915
+bW9uZA== 11916
+IE5vcGU= 11917
+INeW15Q= 11918
+INmC 11919
+IGFkb3B0ZWQ= 11920
+YnVs 11921
+IHNhbnM= 11922
+IHBvc3NpYmlsaXRpZXM= 11923
+IHBlbmQ= 11924
+IHphbWFu 11925
+aG91 11926
+IHNoYXJlcw== 11927
+IGNhbGVuZGFy 11928
+IHBlcnNvbmE= 11929
+IHNlYWw= 11930
+IGdlbmU= 11931
+IHN0b3JlZA== 11932
+INC/0L7Qtw== 11933
+IGx5cmljcw== 11934
+IGluc3RydW1lbnRz 11935
+IE1B 11936
+gOydtA== 11937
+IGNsb3Vkcw== 11938
+aG90 11939
+4bqv 11940
+IOqwmeyVhOyalA== 11941
+IEVtcGlyZQ== 11942
+IGJpbw== 11943
+d2luZA== 11944
+aWVnbw== 11945
+IEV1cm9w 11946
+ZWRnZQ== 11947
+IGJhY2t3YXJkcw== 11948
+IOyngOs= 11949
+IHF1ZWVu 11950
+IHNoaW5l 11951
+IMOnxLFr 11952
+IGNhZA== 11953
+IE9k 11954
+IOyCrOuejA== 11955
+IGJ1YmJsZQ== 11956
+w7Rp 11957
+emVz 11958
+IHJlYWN0aW9ucw== 11959
+IGp1ZGdtZW50 11960
+IERlbW9jcmF0cw== 11961
+IGNvc2Fz 11962
+YXNoZWQ= 11963
+INC00L7Qu9C2 11964
+xZtuaWU= 11965
+6rQ= 11966
+IGV4ZW1wbGU= 11967
+TVA= 11968
+4buv 11969
+IFZlcnM= 11970
+IHJlc2ls 11971
+IG3DoQ== 11972
+xYRzdA== 11973
+YmVsaWV2 11974
+IFZvcg== 11975
+IHNjaGVtZQ== 11976
+b25kYQ== 11977
+IHBvZGVtb3M= 11978
+IGNoYXJnZXM= 11979
+IGRlc3RpbmF0aW9u 11980
+IEt5 11981
+IFNT 11982
+IHNpbGVuY2U= 11983
+IGVtYmVk 11984
+bmF0 11985
+4bubaQ== 11986
+QU5U 11987
+Z2dlZA== 11988
+IHJlZHVjaW5n 11989
+IHVnbHk= 11990
+IG1pbQ== 11991
+0YPQtNCw 11992
+MzQ= 11993
+IGNvcmQ= 11994
+INGC0L7QttC1 11995
+IExpc2E= 11996
+IMO2bg== 11997
+IENocmlzdGlhbnM= 11998
+dW1ibGVz 11999
+b2xvZ2lzdHM= 12000
+YXph 12001
+IHRlbmRz 12002
+IENvb2s= 12003
+IGdlc2FndA== 12004
+IO2VmOuCmA== 12005
+IFRlcw== 12006
+ZXJlbW9ueQ== 12007
+INC90YPQttC90L4= 12008
+IE1BUklTSEE= 12009
+IGVucm9sbA== 12010
+IENyeQ== 12011
+RVNT 12012
+IFNhZA== 12013
+IGltcGxlbWVudGVk 12014
+IGTDrWE= 12015
+w5w= 12016
+IHBpc3Q= 12017
+RHI= 12018
+IHNhYmU= 12019
+IFNvY2k= 12020
+w6RyZQ== 12021
+INC60YLQvg== 12022
+IEZyYW5jaXNjbw== 12023
+IOyepQ== 12024
+IGNyZWF0dXJlcw== 12025
+YXdz 12026
+IGVhcm5lZA== 12027
+IGNoZWFwZXI= 12028
+IGRsYQ== 12029
+IHdhcm4= 12030
+c2NoZQ== 12031
+IGJsYWg= 12032
+IG51dHI= 12033
+6Lw= 12034
+IGdvcmdlb3Vz 12035
+IEFuZ2VsZXM= 12036
+IGdlbWFjaHQ= 12037
+IGhvbWVsZXNz 12038
+b2dyYXBoaWM= 12039
+IFRhaXdhbg== 12040
+IFNvbQ== 12041
+IEhhZA== 12042
+YWN0aW9ucw== 12043
+IHBvc3Rz 12044
+IG91dHJh 12045
+IE1lYW4= 12046
+a2Fy 12047
+IGNvdXM= 12048
+IGJyYWNr 12049
+0LjRgtGM0YHRjw== 12050
+IGJlbGlldmVz 12051
+IHN1aWNpZGU= 12052
+IGVxdWFsbHk= 12053
+IGNhcmVz 12054
+0L7QttC90L4= 12055
+IHN0ZW0= 12056
+IE11Y2g= 12057
+IHByb2R1Y2Vy 12058
+15XXkA== 12059
+IHByb3RlY3Rpbmc= 12060
+IFRSQVZJUw== 12061
+IGludGVydmlld3M= 12062
+IGFsaWVu 12063
+IEFzaw== 12064
+IHNvbGU= 12065
+Q08= 12066
+IFN1ZA== 12067
+IHN1cnZpdg== 12068
+IHNrZXRjaA== 12069
+IHfFgmE= 12070
+IGNvbG9j 12071
+IGFwb2xvZ2l6ZQ== 12072
+d2VpZ2h0 12073
+IDU1 12074
+ID4= 12075
+IGhlcm9lcw== 12076
+IEJvc3Rvbg== 12077
+IGRlcGVuZGVudA== 12078
+IG1vdGl2YXRpb24= 12079
+ZmxpeA== 12080
+IHNlYW0= 12081
+0LrQuNC1 12082
+IGRyYWlu 12083
+b2RlZA== 12084
+IGd1aWx0eQ== 12085
+IEplbm4= 12086
+aW5nZW4= 12087
+IGdyYW50ZWQ= 12088
+IEtlbGx5 12089
+IFNhdg== 12090
+IFVuY2xl 12091
+IEhvbmVzdGx5 12092
+RUxJ 12093
+IG5hdmlnYXRl 12094
+IGJsZXNzZWQ= 12095
+Y29yZQ== 12096
+IGVhcm5pbmc= 12097
+IHNpZ25hbHM= 12098
+IGRpc2s= 12099
+aWFscw== 12100
+IGFnZXM= 12101
+5oU= 12102
+IHBhcnRpY2xl 12103
+INGH0LXRgA== 12104
+IGNhbm4= 12105
+IHRpZXI= 12106
+IHN0YXRlbWVudHM= 12107
+6rOg7JqU 12108
+IOuVjOusuOyXkA== 12109
+IENobw== 12110
+IHBvbGFy 12111
+YW7Dpw== 12112
+IEtlbm4= 12113
+IE5p 12114
+IEZpZ2h0 12115
+b3JnYW4= 12116
+6ZU= 12117
+IENoYQ== 12118
+IFPDrQ== 12119
+IHNsaWM= 12120
+IGNlcnRpZmlj 12121
+IHRlbXBsYXRl 12122
+IEZlZGVyYWw= 12123
+IGNvbnNpZGVyYXRpb24= 12124
+IGV4cGxv 12125
+IE1haW4= 12126
+IE5F 12127
+IGFsb25nc2lkZQ== 12128
+IGRyZXNzZWQ= 12129
+IFBvaW50 12130
+IGVudmlyb25tZW50cw== 12131
+IHByw7N4aW0= 12132
+IGRhYXI= 12133
+IHByb21wdA== 12134
+IHB1cnN1ZQ== 12135
+IGVudGVydGFpbm1lbnQ= 12136
+IHRocm9hdA== 12137
+IHByb2JsZW1h 12138
+IG1hcnQ= 12139
+7Lw= 12140
+IHByb3ZpZGVy 12141
+2Iw= 12142
+INeX 12143
+aW50ZQ== 12144
+bWFraW5n 12145
+IHN0cm9rZQ== 12146
+IHRpc3N1ZQ== 12147
+VW4= 12148
+IHByZWNpb3Vz 12149
+IEFydHM= 12150
+aW5raW5n 12151
+INCe0L0= 12152
+INC40YE= 12153
+bmFo 12154
+INCV0YHQu9C4 12155
+IGNvcm5lcnM= 12156
+IHRyaWNreQ== 12157
+aW5jaA== 12158
+bGlqaw== 12159
+IHByZXNzaW5n 12160
+bGV2ZWw= 12161
+QU5H 12162
+IHJhZGlhdGlvbg== 12163
+7ISg 12164
+IGNvbmZyb250 12165
+IHZldA== 12166
+IHJlcHJlc2VudGF0aXZl 12167
+IHByb3BhZw== 12168
+IGNyYXA= 12169
+IERlYw== 12170
+IHJhbXA= 12171
+0LXQv9C10YDRjA== 12172
+dcOpcw== 12173
+ZXNzZW4= 12174
+Y3JpcHRpb24= 12175
+IGJpbGxz 12176
+IE1hdHRoZXc= 12177
+IGFuaW1l 12178
+4bqldA== 12179
+IGxvd2VzdA== 12180
+aGFz 12181
+c2NyZWVu 12182
+b2dyYXA= 12183
+0LDQu9C+ 12184
+aW50b24= 12185
+IEphaA== 12186
+aXTDoA== 12187
+IGtheQ== 12188
+IHJvdGF0aW9u 12189
+IFdlcmU= 12190
+YWJlaQ== 12191
+IHRyaWFscw== 12192
+IGxldmVy 12193
+aWdodHk= 12194
+IHNwb29u 12195
+IGh1bnQ= 12196
+Y2xpbmc= 12197
+IGRpc20= 12198
+INCx0L7Qu9GM0Yg= 12199
+IGFzc2F1bHQ= 12200
+IO2YlQ== 12201
+IHdlZWtseQ== 12202
+IG1pc21v 12203
+IGdlbmV0aWM= 12204
+dWxwdA== 12205
+IFN0dWRlbnQ= 12206
+IHJlYWxpc3RpYw== 12207
+IGF1dGhlbnRpYw== 12208
+YXN0YQ== 12209
+IGFycmVzdGVk 12210
+IGd1aWRlbGluZXM= 12211
+INec15A= 12212
+INC00LDQsg== 12213
+IENvbWluZw== 12214
+ZsO8cg== 12215
+IHJlcXVlc3Rz 12216
+g5A= 12217
+IGFuYWx5emU= 12218
+IGludGVyZXNz 12219
+IGhhbHQ= 12220
+IE9wZXI= 12221
+b25vbQ== 12222
+IGR1Y2s= 12223
+IHdpdGhk 12224
+c2Vy 12225
+IM+M 12226
+IEhpc3Rvcnk= 12227
+IHlvdXR1YmU= 12228
+IHNhYmVy 12229
+d2Fsaw== 12230
+Zm9udA== 12231
+IG92ZXJ2aWV3 12232
+Mzk= 12233
+w7x5 12234
+ZXR0aQ== 12235
+IGZyb3plbg== 12236
+IGZsZXNo 12237
+xJ9p 12238
+IFBN 12239
+IOyZgA== 12240
+6aI= 12241
+0YbQuNC4 12242
+IOq4sOs= 12243
+7YGs 12244
+IHByb3Nl 12245
+b29vbw== 12246
+cmF0ZXM= 12247
+V1M= 12248
+IGF1dG9tYXRpYw== 12249
+IGNvbGxlY3Rpbmc= 12250
+xZE= 12251
+IG5laWdoYm9ycw== 12252
+wrsu 12253
+IEV4cGw= 12254
+IGNpcmN1bA== 12255
+Y292ZXI= 12256
+d2Vn 12257
+IHN0aWNrcw== 12258
+IGVsbGVy 12259
+IHd3dw== 12260
+IGRvcm0= 12261
+IEV4cGVy 12262
+IHN0YXRpc3RpY3M= 12263
+IGVtYWlscw== 12264
+IGdyYXZl 12265
+aW1peg== 12266
+SFM= 12267
+IHVpdA== 12268
+LCc= 12269
+IGxhc2Vy 12270
+6Ik= 12271
+INGC0LXQvA== 12272
+0YvRiA== 12273
+0YnRkQ== 12274
+IGdlbmF1 12275
+IHRpZW5lbg== 12276
+IG1lZGl0YXRpb24= 12277
+IE9yZ2Fu 12278
+IGVzdGltYXRl 12279
+IOustOw= 12280
+bGV0cw== 12281
+IG7DoHk= 12282
+IG1pbmRzZXQ= 12283
+IHJlc29u 12284
+IG3DqXM= 12285
+IG51bWVyb3Vz 12286
+IHZpZWxsZWljaHQ= 12287
+IFRoaXJk 12288
+dW91cw== 12289
+IERlYWQ= 12290
+0LDQvdC0 12291
+SE4= 12292
+IHJhY2luZw== 12293
+IGFnZW50cw== 12294
+IFV0 12295
+IHRlYXI= 12296
+IEhQ 12297
+IGNoZW1pc3RyeQ== 12298
+IHN1cnZpdmFs 12299
+IGNvbnZpbmNlZA== 12300
+IDs= 12301
+IHJlZ3VsYXRpb25z 12302
+IEVT 12303
+MzAw 12304
+IGVuc2U= 12305
+IOy1 12306
+IGRpY3Q= 12307
+R0E= 12308
+IGFow60= 12309
+IHRlag== 12310
+INC+0YHRgg== 12311
+IEVsZWN0 12312
+IGludGVsbGVjdHVhbA== 12313
+IGJpYXM= 12314
+IGJ1cmRlbg== 12315
+IOyWtOuWuw== 12316
+IGNoZWVy 12317
+IHNvcGg= 12318
+IHBvcnRmb2xpbw== 12319
+dWJh 12320
+IGVzdG9z 12321
+VFY= 12322
+Rm9y 12323
+IGFzaA== 12324
+IGtvbW1lcg== 12325
+IGNvbGxlY3RpdmU= 12326
+IHdyZXN0 12327
+IEpldHp0 12328
+IFdhdA== 12329
+cmVpY2g= 12330
+IHByaW1lcg== 12331
+YWN0aXZl 12332
+IG1pZQ== 12333
+aWNrZWQ= 12334
+IGh1bnRpbmc= 12335
+IHRlc3RpbQ== 12336
+IGNvbXBhc3Npb24= 12337
+INix 12338
+IGJydXQ= 12339
+IHNhbGFk 12340
+0L7QsdGJ0LU= 12341
+IHNvbHZpbmc= 12342
+IGZsb2F0aW5n 12343
+57c= 12344
+IGF0dHJhY3RpdmU= 12345
+2YjZhA== 12346
+IHBlcmQ= 12347
+aWZmZXI= 12348
+IHNjdWxwdA== 12349
+aGho 12350
+IFdlZWs= 12351
+IGVudGh1cw== 12352
+IG5hZA== 12353
+IG1lcmNo 12354
+IO2ZlQ== 12355
+IG1pbGU= 12356
+IM64 12357
+IOuCmOs= 12358
+Mzg= 12359
+IGNoYWlucw== 12360
+IEFsbW9zdA== 12361
+IHRpY2tldHM= 12362
+cmlu 12363
+IEND 12364
+IGRpc3RyaWJ1dGVk 12365
+YWJldGVz 12366
+IHRlbXBlcmF0dXJlcw== 12367
+IGdhaW5lZA== 12368
+IGZsZXhpYmlsaXR5 12369
+IHNjcmVhbWluZw== 12370
+IGFicm9hZA== 12371
+dW5v 12372
+IGVudHJlcHJlbmV1cnM= 12373
+IE5ldHdvcms= 12374
+IENhbmFkaWFu 12375
+IHByZXY= 12376
+IHPDtg== 12377
+INGC0LXQsdGP 12378
+IFBva2U= 12379
+IFBvZA== 12380
+IFR1cmtleQ== 12381
+IGFic3RyYWN0 12382
+IHNuYWtl 12383
+IEFteQ== 12384
+IOuKkOuCjA== 12385
+IGJyYXZl 12386
+IOyeiOyWtOyalA== 12387
+IEthbA== 12388
+IDIwMDc= 12389
+w6FyaW8= 12390
+IG1hcmtlZA== 12391
+Z2luZXM= 12392
+IGFsbG9j 12393
+T05H 12394
+IHNjaWVudGlzdA== 12395
+IGVzY2E= 12396
+IHJhY2lzbQ== 12397
+15HX 12398
+IFNhbXM= 12399
+IFBlbm4= 12400
+IGxvYWRz 12401
+IOCuqA== 12402
+w7xiZXI= 12403
+TWU= 12404
+aXjDsg== 12405
+IHBlcsOy 12406
+YW5uZQ== 12407
+IGV4cHJlc3NlZA== 12408
+0LzQtdGA 12409
+IG1vZXQ= 12410
+IHJldHVybmluZw== 12411
+bmlh 12412
+IGV4cG9u 12413
+UHJv 12414
+IGxveWFs 12415
+TUw= 12416
+IGxhbXA= 12417
+IHNoeQ== 12418
+IGNvbXBvc2l0aW9u 12419
+IEx5 12420
+IG1hZ25ldGlj 12421
+IHByZW1pZXI= 12422
+IG1lYXN1cmVk 12423
+IHN1bW1hcnk= 12424
+IGF0dGFja2Vk 12425
+IGZpbmlzaGluZw== 12426
+0Jc= 12427
+56U= 12428
+IHNpdHM= 12429
+IGh5ZHJvZ2Vu 12430
+IG1haQ== 12431
+IERldXRzY2g= 12432
+YXPEsQ== 12433
+IG9idGFpbg== 12434
+dmll 12435
+IHNvaXQ= 12436
+IOuwlA== 12437
+IGxhbmU= 12438
+IGNvbnNlZ3U= 12439
+0LLQvg== 12440
+IGVhc2U= 12441
+YWtpbg== 12442
+IEZh 12443
+IHVudHVr 12444
+IGJ1cnN0 12445
+IGN1bQ== 12446
+YWzEsW0= 12447
+w7pibGlj 12448
+aWRp 12449
+IFJveWFs 12450
+IEtvbg== 12451
+IGNvbW1vbmx5 12452
+IHJlbW92aW5n 12453
+IGp1cg== 12454
+aWxpYg== 12455
+IGFuY2g= 12456
+7ZaJ 12457
+xrDhu6M= 12458
+INCc0Ys= 12459
+IEFudGg= 12460
+IFPDpQ== 12461
+IGludGVycnVwdA== 12462
+IHN0ZXJl 12463
+IE9T 12464
+b255bQ== 12465
+dGVyeQ== 12466
+IE1hcmlh 12467
+6rKD 12468
+IGV4cGxvcmluZw== 12469
+IHRyYW5zcGFyZW50 12470
+IGZhdGU= 12471
+IEp1bmc= 12472
+IGdydXA= 12473
+IGRhcmtlcg== 12474
+IERvdWc= 12475
+IG1hbmU= 12476
+4bqhaQ== 12477
+ZHJp 12478
+bG9vaw== 12479
+IERlc2lnbg== 12480
+IHR1dGFq 12481
+IGhvcml6b250YWw= 12482
+cmVvbg== 12483
+b3J0ZQ== 12484
+IENvcnJlY3Q= 12485
+IFN0ZXZlbg== 12486
+IHZpbmU= 12487
+MDI= 12488
+acSH 12489
+IHNpZW1wcmU= 12490
+IEtleQ== 12491
+IEdhbWVz 12492
+IG5hYXI= 12493
+IHNob2NrZWQ= 12494
+ZWx2ZQ== 12495
+IFJvc2U= 12496
+7Ius 12497
+IHN0b3BwaW5n 12498
+b2hs 12499
+IE1peA== 12500
+IHN1ZmZlcmVk 12501
+IHNpZ21h 12502
+IHdlYWtuZXNz 12503
+IE93 12504
+4Li14LmI 12505
+SUY= 12506
+IOCuhQ== 12507
+YWRlZA== 12508
+IE5ldGZsaXg= 12509
+YW5lcw== 12510
+IHJlbWFpbmVk 12511
+aXJ5 12512
+IHJpcA== 12513
+ZWxsdA== 12514
+IHNpbGVudA== 12515
+IHByb3Zlbg== 12516
+IHRveGlj 12517
+IGFsdW1pbg== 12518
+IG11bHRpcGw= 12519
+YWxhbmQ= 12520
+IDM0 12521
+MDY= 12522
+IEJydQ== 12523
+IOygleunkA== 12524
+SnVzdA== 12525
+Ym95 12526
+IHNob2U= 12527
+IGNyZWF0dXJl 12528
+IGhlYWRlZA== 12529
+INC+0YLQug== 12530
+5rE= 12531
+IGVzc2VuY2U= 12532
+IHJlbWFya2FibGU= 12533
+IG7Dum1lcg== 12534
+IGRyZXc= 12535
+IHB1enpsZQ== 12536
+IExpYnJhcnk= 12537
+IEZ1 12538
+YXNoZXM= 12539
+a2s= 12540
+IElzdA== 12541
+prA= 12542
+IEJyeQ== 12543
+IGNlcmVtb255 12544
+IOCujg== 12545
+IGNyaQ== 12546
+ZXF1 12547
+IHByaXpl 12548
+IGRpbWVuc2lvbnM= 12549
+b2dyYW0= 12550
+IGxlYXRoZXI= 12551
+IHBvcHVsYXRpb25z 12552
+dXVt 12553
+IHZlZ2Fu 12554
+0Y/QtA== 12555
+IGPDs21v 12556
+5YQ= 12557
+IHN0cmlw 12558
+5aM= 12559
+IHZhY2F0aW9u 12560
+hZU= 12561
+IG1lYWxz 12562
+aWxpcHA= 12563
+IGVudHM= 12564
+YXJhbQ== 12565
+cmljaHQ= 12566
+IGdyYWlu 12567
+IFNwYWlu 12568
+IGNoZWVr 12569
+IEFmZg== 12570
+SU9O 12571
+IEJyaW5n 12572
+IDM4 12573
+aWVsZW4= 12574
+dWx1 12575
+INCx0L7Qu9GM0YjQtQ== 12576
+IGFubm91bmNlbWVudA== 12577
+INGC0YPRgg== 12578
+IFByb3BoZXQ= 12579
+YXJkbw== 12580
+Mzc= 12581
+IHdva2U= 12582
+IHRyYW5zbGF0aW9u 12583
+IE5PVA== 12584
+IENM 12585
+IGTDvMWf 12586
+0YbRlg== 12587
+YWNlcg== 12588
+IExvYw== 12589
+IHBlcmNlcHRpb24= 12590
+Tk8= 12591
+IGRpZXNlbg== 12592
+TG9vaw== 12593
+aGVhcnQ= 12594
+YXZlZA== 12595
+IGJvdW5kYXJ5 12596
+IGZsb3dz 12597
+0ZHQvA== 12598
+IGFyZ3VtZW50cw== 12599
+IGVsZWN0aW9ucw== 12600
+xLFz 12601
+IGhlY2s= 12602
+IHN1aXRhYmxl 12603
+IGZpYmVy 12604
+IFN0cmE= 12605
+eHk= 12606
+IEh1bQ== 12607
+IG1vbnRobHk= 12608
+dXBlcg== 12609
+IGdvbGY= 12610
+IGxhdGVseQ== 12611
+IEdhcmQ= 12612
+IFJlbg== 12613
+IEFzdA== 12614
+IEZhbnQ= 12615
+0LDRgdGB 12616
+IG9ic2Vy 12617
+66Gc 12618
+IGVhc2llc3Q= 12619
+jZTr 12620
+IHdlYnNpdGVz 12621
+cG9s 12622
+IGNvY29u 12623
+IOCuhw== 12624
+IFZlZw== 12625
+IHdhbGtz 12626
+IGludHJv 12627
+IGRpcmVjdGVk 12628
+IEFubmE= 12629
+IOuTpOyWtA== 12630
+IEVhc3Rlcm4= 12631
+IFNhaW50 12632
+IEJvdw== 12633
+IHJvYXN0 12634
+IFVSTA== 12635
+IGplZGVu 12636
+dXJhcw== 12637
+YWph 12638
+IHNlbWk= 12639
+IHJhcGlkbHk= 12640
+IHRhcmdldHM= 12641
+IENvbnRyb2w= 12642
+IGJhaA== 12643
+IHJlZmxlY3Rpb24= 12644
+IGNyZWF0aXZpdHk= 12645
+aG9sZGVycw== 12646
+IOyYrOs= 12647
+IGFtb25nc3Q= 12648
+IGZlZWRpbmc= 12649
+0Y3RgtC+0LzRgw== 12650
+INCy0LjQtNC1 12651
+IOunjOuTpA== 12652
+IFNtYXJ0 12653
+IHJlbGlhYmxl 12654
+IHZlemVz 12655
+INeo 12656
+Y2h1Y2tsZXM= 12657
+YXppb25l 12658
+IFdpbGxpYW1z 12659
+IGHDpw== 12660
+IHNsZWU= 12661
+0LXRiQ== 12662
+IHRpbWVsaW5l 12663
+IHRob3JvdWdo 12664
+4buN 12665
+IE90 12666
+4bqhbg== 12667
+IGltYWdpbmF0aW9u 12668
+IG1lY2hhbmljcw== 12669
+cmlzdA== 12670
+IGNsYWltZWQ= 12671
+z4TOtw== 12672
+w6p0ZQ== 12673
+IEh1cnJ5 12674
+IGlQYWQ= 12675
+IGNvbnN0cnU= 12676
+IENsYQ== 12677
+IEFscw== 12678
+dXR6 12679
+IGN1bHR1cmVz 12680
+IOyWtOuWu+qyjA== 12681
+IGJlbG9uZ3M= 12682
+IHllcg== 12683
+IERvZXNu 12684
+IGdlb21ldA== 12685
+IGJpZA== 12686
+IGZvYW0= 12687
+IGhvYg== 12688
+IEJyaXRhaW4= 12689
+IHN1YnN0YW5jZQ== 12690
+IGFubml2ZXJzYXJ5 12691
+IOuEiA== 12692
+IG5vdGVk 12693
+IGdvdmVybm9y 12694
+IHN0b2Nrcw== 12695
+MzE= 12696
+IGRpeWU= 12697
+7Iqk6w== 12698
+IHJlYg== 12699
+emVs 12700
+IG11bHRpcGx5 12701
+IG9wZXJhdG9y 12702
+hKTsmpQ= 12703
+IHdhdGVycw== 12704
+IGTDpHI= 12705
+IHVuc2Vy 12706
+IEVsaXphYmV0aA== 12707
+IGluY3JlYXNpbmdseQ== 12708
+IEdybw== 12709
+IGVuZ2luZXM= 12710
+aXJz 12711
+2Ks= 12712
+IHRyZWFzdXJl 12713
+UEM= 12714
+aW5jdGlvbg== 12715
+aXJp 12716
+IGFjY3Vt 12717
+IHZhcmlhdGlvbg== 12718
+IHBvbQ== 12719
+IHRpdGxlcw== 12720
+IEZlc3Q= 12721
+w7Nz 12722
+IGVsZGVy 12723
+bnlt 12724
+cnVu 12725
+0Y/Qsg== 12726
+IGlubm92YXRpdmU= 12727
+IG5vbWJyZQ== 12728
+IGNvaW5j 12729
+IGZyYW5jaA== 12730
+IGVudG9uY2Vz 12731
+IG5pY2h0cw== 12732
+IGV4Y2x1c2l2ZQ== 12733
+IENoZWVycw== 12734
+IEJp 12735
+dWpl 12736
+IHBvaw== 12737
+IFByZW0= 12738
+IHJvY2tldA== 12739
+RUxJUEU= 12740
+IGhvc3BpdGFscw== 12741
+cml1bQ== 12742
+IGp1c3Rl 12743
+IGhhbW1lcg== 12744
+IHF1YW50dW0= 12745
+IHJlc3BvbnNlcw== 12746
+bGx5 12747
+ZW5kaQ== 12748
+IGFjdGl2ZWx5 12749
+IGZyaWRnZQ== 12750
+aWF0ZQ== 12751
+bG9uZw== 12752
+IHF1ZW0= 12753
+IGRlYXRocw== 12754
+IHN1cGVyaW9y 12755
+Y2tlbg== 12756
+7J207JeQ 12757
+a3RvcA== 12758
+IGdhdGhlcmVk 12759
+o6g= 12760
+IGRhenU= 12761
+IHJlY2lwZXM= 12762
+IGJ1eno= 12763
+Y2Vu 12764
+IGFueXRpbWU= 12765
+b25zZW5zZQ== 12766
+IGNpcmNsZXM= 12767
+IHNvbHZlZA== 12768
+IOyLoA== 12769
+IGNvcm9uYXZpcnVz 12770
+IEx1a2U= 12771
+IGJ1YmI= 12772
+IGNvbnRlbXBvcg== 12773
+cnp5 12774
+IEphbmU= 12775
+INC00L7QvA== 12776
+IHNjcmV3cw== 12777
+IGh5YnJpZA== 12778
+IGNhc3VhbA== 12779
+IHNlbGJzdA== 12780
+YmVpbmc= 12781
+IMSQ 12782
+IENvbHVtYg== 12783
+INGF0L7Rhw== 12784
+IGJ1Y2tldA== 12785
+IGV2YWx1YXRl 12786
+IGlkb2w= 12787
+IHJlcHV0YXRpb24= 12788
+IOyGjOs= 12789
+2YjYsQ== 12790
+IGhlY2hv 12791
+IHBvZW0= 12792
+IHN1YmplY3Rz 12793
+cGxhbnQ= 12794
+IEJlaA== 12795
+IFNwZWFraW5n 12796
+IGJhdHRlcmllcw== 12797
+IGZvbGxvd2Vycw== 12798
+w7Zs 12799
+IGdlbnRseQ== 12800
+IHNpeHQ= 12801
+IHBhcmFtZXRlcg== 12802
+IGlra2U= 12803
+IFRvdXI= 12804
+IERK 12805
+b3R0ZQ== 12806
+IEphaHJlbg== 12807
+IHByZXBhcmF0aW9u 12808
+INC00YPQvA== 12809
+IDgwMA== 12810
+Y29w 12811
+aWtpbmc= 12812
+IOusuA== 12813
+INC90YM= 12814
+INC70LXRgg== 12815
+IElkZQ== 12816
+IOyhsOq4iA== 12817
+IGxhdWdodGVy 12818
+IG1vbGVjdWxlcw== 12819
+IFJlc3Q= 12820
+IG9ic2VydmVk 12821
+ZHppZQ== 12822
+IGFkdmVydGlzaW5n 12823
+ZXJ0bw== 12824
+IG1vaW5z 12825
+IE1JVA== 12826
+IGV4Y2l0 12827
+IHR1bQ== 12828
+IHR5bA== 12829
+IGludmVzdGVk 12830
+IHBoYXJt 12831
+IHVuZXhwZWN0ZWQ= 12832
+IHBoaQ== 12833
+b3R5cGU= 12834
+d2Vpc2U= 12835
+IGdlw6c= 12836
+am91cmQ= 12837
+IGhvcnNlcw== 12838
+bsSF 12839
+PSI= 12840
+IFNN 12841
+IGZpYg== 12842
+IGNsaXBz 12843
+IHJlZ2ltZQ== 12844
+IHJvdGF0ZQ== 12845
+cm91 12846
+bmlr 12847
+IGFybW9y 12848
+8J+Y 12849
+0LXRgNCw 12850
+IE9jaA== 12851
+IHJpY2h0aWc= 12852
+w7x6ZWw= 12853
+YW5lb3VzbHk= 12854
+bWVr 12855
+IFhpYW8= 12856
+IGV4aXN0ZWQ= 12857
+d29ydGg= 12858
+IG5hdWdodA== 12859
+IGhlacOfdA== 12860
+IEJhbA== 12861
+IHJlc2lk 12862
+aXZvdA== 12863
+b21hdGlj 12864
+IGhpcmVk 12865
+IGdyYWR1YWxseQ== 12866
+IG9uaW9ucw== 12867
+IGNvbXBhdA== 12868
+IGludGlt 12869
+IGpldw== 12870
+IGNvbnRyaWJ1dGlvbg== 12871
+IElyZQ== 12872
+YWNqaQ== 12873
+IHNsaWNl 12874
+IGltbXVu 12875
+IFJ1cw== 12876
+IGdyb3dz 12877
+IFNpbWlsYXJseQ== 12878
+IGhhcmRlc3Q= 12879
+IHN0cnVjaw== 12880
+IG1lYXN1cmVtZW50 12881
+Li4uXQ== 12882
+dGhleQ== 12883
+IOyggOs= 12884
+IHNuZWFr 12885
+IGFwcGxpZXM= 12886
+INC90LXQvA== 12887
+5pM= 12888
+15HXqA== 12889
+INCn0YLQvg== 12890
+IG91dHJv 12891
+IGlubm9jZW50 12892
+IG1vZw== 12893
+IFNhbXN1bmc= 12894
+IG1lcmN5 12895
+IGhhbmRsaW5n 12896
+IGludGVydmVudGlvbg== 12897
+aWRheXM= 12898
+Z290 12899
+IGN1cnJpYw== 12900
+IGJvdW5kYXJpZXM= 12901
+IGNvbmZ1c2luZw== 12902
+nbzripQ= 12903
+5oc= 12904
+IHN0aXRjaGVz 12905
+w612ZWw= 12906
+IHR1bm5lbA== 12907
+aXTDpA== 12908
+IGdvc3Q= 12909
+aW15 12910
+IGN6YXM= 12911
+IG3DqQ== 12912
+IGNhdGFs 12913
+IFNpbW9u 12914
+IExJQU0= 12915
+bWlj 12916
+INCk 12917
+IGV5ZWw= 12918
+aXNhcw== 12919
+IENQVQ== 12920
+IERvdQ== 12921
+IG7DpGNo 12922
+IGluZmluaXR5 12923
+IHJpZg== 12924
+IFBlYWNl 12925
+IEN1 12926
+IG1pbmltYWw= 12927
+IGxpc3RlbmVk 12928
+IHBvbGU= 12929
+aGFsYg== 12930
+IGxvYWRlZA== 12931
+IHN0ZWFkeQ== 12932
+IEJlc2lkZXM= 12933
+w6pt 12934
+IGxhcA== 12935
+IGNvb3A= 12936
+IGZyaWVuZHNoaXA= 12937
+d29ybGQ= 12938
+IGdlaA== 12939
+IHR5bGtv 12940
+IExhdXJh 12941
+IHN1cnJvdW5kZWQ= 12942
+IEV2ZW50 12943
+IGNoYXA= 12944
+IFdvbmRlcg== 12945
+YnJlYWs= 12946
+IGRyb3Zl 12947
+IGJyb2FkZXI= 12948
+IGNoaQ== 12949
+Rmk= 12950
+IGdlaGVu 12951
+IHdlc3Rlcm4= 12952
+IGludGVsbGlnZW50 12953
+IHBlcnNpc3Q= 12954
+IGZvdW5kZWQ= 12955
+IGhpc3Rvcmlj 12956
+IGZyw6U= 12957
+Y2tzw6U= 12958
+IGhhbmR5 12959
+IHN5bXA= 12960
+IHJvd3M= 12961
+IG51dHJp 12962
+YnVy 12963
+IExlb24= 12964
+IHNpc3RlbWE= 12965
+IGV4dGVuc2l2ZQ== 12966
+INGD0LI= 12967
+7Y8= 12968
+IG5pZ2h0cw== 12969
+IGPDoWM= 12970
+IGNvdW50aW5n 12971
+IE11c3Q= 12972
+YWxsb3c= 12973
+0LXRgdGB 12974
+TW9t 12975
+INC90LDQtNC+ 12976
+IGJhcnJlbA== 12977
+QVJE 12978
+IGluc3RhbGxhdGlvbg== 12979
+IGluc2VjdA== 12980
+IOuFuOs= 12981
+dWrEhQ== 12982
+IMSRaQ== 12983
+IHBhY2tlZA== 12984
+IGZpY3Rpb24= 12985
+Tm93 12986
+IFlheQ== 12987
+IHBlcnQ= 12988
+cm9ucw== 12989
+dW5kZQ== 12990
+YWNoZXM= 12991
+IHN0eWxlcw== 12992
+IGFwcsOocw== 12993
+b2t1 12994
+IFZpY2U= 12995
+xLFuxLF6 12996
+Y29tbQ== 12997
+IGFzc2lnbmVk 12998
+IGludGVyYWN0aW9ucw== 12999
+IGFjYWI= 13000
+RkVMSVBF 13001
+IHJlc2N1ZQ== 13002
+IGluZHVzdHJpZXM= 13003
+IEFuZHk= 13004
+IHByYWlzZQ== 13005
+IGZsYW1l 13006
+IHNuYWNr 13007
+7YI= 13008
+54E= 13009
+IHN3bw== 13010
+cmVuZGVy 13011
+IGJvYXJkcw== 13012
+INGC0L7QvA== 13013
+ZW5uZQ== 13014
+IHBhc3Rh 13015
+IGRldmls 13016
+IEZlbA== 13017
+IGhhdHRl 13018
+IGNvbGxlZw== 13019
+ZWg= 13020
+7Ls= 13021
+IHByb2R1Y3RpdmU= 13022
+Zm9yd2FyZA== 13023
+0LjQvw== 13024
+IHNtYXJ0cGhvbmU= 13025
+IGludmlz 13026
+IGJ1bQ== 13027
+IHdob2E= 13028
+7J6E 13029
+IG9ja3PDpQ== 13030
+IExhbmc= 13031
+IFN5cmlh 13032
+IHNlc2k= 13033
+zq/OsQ== 13034
+IGFwcHJvdmFs 13035
+NDg= 13036
+INC+0LTQuNC9 13037
+IOuW 13038
+IEhhcnI= 13039
+IEFkbWluaXN0 13040
+INek 13041
+IERlYW4= 13042
+Zmk= 13043
+IGNpdGl6ZW4= 13044
+IHNoYXJr 13045
+MDU= 13046
+IGJvaWw= 13047
+IGluZGljYXRl 13048
+5aE= 13049
+QXJl 13050
+IGxheW91dA== 13051
+IHJlZnI= 13052
+IFBhY2lmaWM= 13053
+QUFBQQ== 13054
+IEF1c3RyYWxpYW4= 13055
+Z3Jlc3Npb24= 13056
+Vm9pY2U= 13057
+0LDQu9GB0Y8= 13058
+IHNoZWx0ZXI= 13059
+VG8= 13060
+YXVwdA== 13061
+IGV2YWx1YXRpb24= 13062
+YXBvcg== 13063
+IGN1cnJlbmN5 13064
+INC80L3QvtCz0L4= 13065
+aWdvcw== 13066
+IG9jdA== 13067
+IHJveWFs 13068
+6LM= 13069
+YXNpbA== 13070
+IENoaWxkcmVu 13071
+IHJpZW4= 13072
+IOuTnOs= 13073
+IGJhcnJpZXI= 13074
+IGVqZW1wbG8= 13075
+IGVr 13076
+TkQ= 13077
+ZXNw 13078
+0LXQvdCw 13079
+IHBpYw== 13080
+IGtpbGxlcg== 13081
+IGludGVncmF0ZQ== 13082
+IGZld2Vy 13083
+IGRpc2FiaWxpdGllcw== 13084
+IC4uLi4= 13085
+IHRyaWFuZ2xl 13086
+IGZlZXM= 13087
+IHdpZGVseQ== 13088
+ZW1p 13089
+IG92ZXJ3aGVsbWluZw== 13090
+IHpvbWI= 13091
+IGJlcmU= 13092
+IGhvb2Q= 13093
+IEF5ZQ== 13094
+IEhhcnZhcmQ= 13095
+ZXY= 13096
+IM+Ezr/PhQ== 13097
+IGN1cHM= 13098
+IEF1Y2g= 13099
+em9uYQ== 13100
+IDE5OTA= 13101
+IHdlacOf 13102
+IGNydW5jaA== 13103
+5qU= 13104
+INC30LDQsg== 13105
+IG1lYXN1cmluZw== 13106
+IHN0YXRpb25z 13107
+IFN0ZXBoZW4= 13108
+IHNob3J0bHk= 13109
+IHNpZ25pbmc= 13110
+IGNvbWVkeQ== 13111
+b21v 13112
+IHN1Z2dlc3Rpb25z 13113
+IHNpZ25hdHVyZQ== 13114
+INC/0YDQuNCy 13115
+IGRpc29yZGVy 13116
+YXNrYQ== 13117
+IHdvcmxkcw== 13118
+IHByZWNpc2VseQ== 13119
+bm9ybQ== 13120
+cmF2 13121
+IENpdmls 13122
+SW50ZXI= 13123
+IENlcnRhaW4= 13124
+IGluanVyZWQ= 13125
+IHN1Z2dlc3Rz 13126
+IEdvbGRlbg== 13127
+IGN5YmVy 13128
+INi0 13129
+IHRlbXBvcmFyeQ== 13130
+IGNvb3Blcg== 13131
+IHZvdGVk 13132
+IG91Z2h0 13133
+4bqleQ== 13134
+eHVhbA== 13135
+IHBhbmVscw== 13136
+IDk1 13137
+IGhhbmRzb21l 13138
+INC/0YDQvtCy 13139
+IHBlcm1pdA== 13140
+IGtlaW4= 13141
+IGJhZGx5 13142
+IG5vdGlmaWNhdGlvbnM= 13143
+aXph 13144
+IE5vdGljZQ== 13145
+IGluY2x1c2l2ZQ== 13146
+IGFuc3dlcmluZw== 13147
+IO2X 13148
+dWxk 13149
+7YWM 13150
+IG5vd2FkYXlz 13151
+IDM3 13152
+IGJvbHQ= 13153
+IHN0YXRpYw== 13154
+IEhvcA== 13155
+IGF2YW50 13156
+YWpv 13157
+IOunm+yeiA== 13158
+IGZpZnR5 13159
+IEZpbmFs 13160
+IHNjb3Jlcw== 13161
+IFRhcA== 13162
+IGN5bA== 13163
+IGNvbnZpbmNl 13164
+IGFueXdheXM= 13165
+b2Rh 13166
+IOyVvA== 13167
+IHNlcnZlcw== 13168
+INGC0LDQutC+0Lk= 13169
+IFpvb20= 13170
+IHNhdmluZ3M= 13171
+dWxv 13172
+IHNvdXRoZXJu 13173
+dmlld2Vy 13174
+IGhvamU= 13175
+IHNlamE= 13176
+IHJlcHJlc2VudGluZw== 13177
+iOuNmA== 13178
+bGlr 13179
+IFNvbWVib2R5 13180
+IGJlYXN0 13181
+IHN0aWNraW5n 13182
+IGluc2lzdA== 13183
+IHRhbGVudGVk 13184
+IGV4cGxhaW5pbmc= 13185
+IGF0dG9ybmV5 13186
+IHN0YWlycw== 13187
+IERvZw== 13188
+7Ys= 13189
+IGNpZw== 13190
+IHNoYXBlZA== 13191
+IHNvbnM= 13192
+z4HOuQ== 13193
+dXR0 13194
+IOyU 13195
+IHBhcmFk 13196
+7J24642w 13197
+IGhvcm4= 13198
+IEpvdXI= 13199
+YW5ubw== 13200
+IHdvcmxkd2lkZQ== 13201
+IHBhcnRpY2lwYXRpb24= 13202
+poQ= 13203
+IG3Ds3c= 13204
+IGJ1cm5lZA== 13205
+IHdyaXRlcnM= 13206
+YWxsYWg= 13207
+IEZ1bmQ= 13208
+IGNsZXZlcg== 13209
+IExldXRl 13210
+Ymlu 13211
+IGJlYXRpbmc= 13212
+Zm9vdA== 13213
+IOybkA== 13214
+IFN0dWRpbw== 13215
+IHZhZw== 13216
+YmV5 13217
+cnpl 13218
+IG9wcG9zaXRpb24= 13219
+INC20LjQtw== 13220
+d2hv 13221
+IOqxtA== 13222
+IHRyYWNl 13223
+INC00LXQvdGM 13224
+IGVwaWQ= 13225
+IGdlc2No 13226
+IE5hcg== 13227
+IEJF 13228
+0YPQuQ== 13229
+IFNpZ24= 13230
+ZWRseQ== 13231
+IGNsYXk= 13232
+IGluc3RhbnRseQ== 13233
+IGdhdGhlcmluZw== 13234
+IEdhbGF4eQ== 13235
+IGJvcmVk 13236
+IEJ1ZGRo 13237
+Y8Op 13238
+IG1hbQ== 13239
+IHNsb3Bl 13240
+IOuLpOydjA== 13241
+IHNjaMO2bg== 13242
+IHBpcg== 13243
+Z2Vm 13244
+YW1lcg== 13245
+IGjDtg== 13246
+IGNvbGxlYWd1ZQ== 13247
+IHByZXNlbnRz 13248
+YWRpdW0= 13249
+IOCutQ== 13250
+IGZhbGFy 13251
+YmVlcA== 13252
+IGRyaWVk 13253
+aXNtcw== 13254
+IHJvcGU= 13255
+IHdvcmtzaG9w 13256
+IGVzdHVk 13257
+IGJhbmRz 13258
+IHRoZW1lcw== 13259
+2YrYsQ== 13260
+IHJlbWluZGVy 13261
+0YLRgw== 13262
+IEJo 13263
+IGNvY29udXQ= 13264
+INGB0YLQvg== 13265
+IENoYW5uZWw= 13266
+IGltbWlncmF0aW9u 13267
+w6Rz 13268
+Li4uLi4= 13269
+c3RvcA== 13270
+INC60LDRgA== 13271
+IGNvaW5z 13272
+INGH0LDRgQ== 13273
+IGRlc3RydWN0aW9u 13274
+bGluZWQ= 13275
+IGJhcnJpZXJz 13276
+YW50aW5l 13277
+IHByaW50ZWQ= 13278
+IGNvbmdyYXR1bGF0aW9ucw== 13279
+IEhlYXJ0 13280
+IGlucXU= 13281
+dGhh 13282
+IGhhcmRseQ== 13283
+IEF2ZW4= 13284
+IHRpbmhh 13285
+IFNvbnk= 13286
+IE5G 13287
+IGdyYWR1YXRlcw== 13288
+IHNxdWVlemU= 13289
+ZXJlbXk= 13290
+z4TOuQ== 13291
+IGVwaWM= 13292
+IEp1 13293
+IG9sbQ== 13294
+IExhdWdodGVy 13295
+IGJlbGllZnM= 13296
+IENydQ== 13297
+IFRydWU= 13298
+IFNvdWw= 13299
+b3dlZW4= 13300
+IHJvbWFudGlj 13301
+INC30LI= 13302
+IGFub3M= 13303
+IFl1cA== 13304
+ZGlt 13305
+IGluZmVy 13306
+INC30LDQvA== 13307
+IHNvYw== 13308
+dWth 13309
+IHByZWNpc2U= 13310
+IGRyb3BwaW5n 13311
+IGNsdWU= 13312
+IGVycm9ycw== 13313
+Y2hhcmdl 13314
+IFB1 13315
+b21ldGVy 13316
+IGxhbWJkYQ== 13317
+YWNpb25hbA== 13318
+IERvbmc= 13319
+IGNoYW1iZXI= 13320
+IHRoYW5rZnVs 13321
+IE51 13322
+IEhhd2Fp 13323
+IGluZm8= 13324
+IGFjdGl2YXRl 13325
+IFF1YWw= 13326
+IHF1ZWQ= 13327
+0YPQu9GM 13328
+IGNsb3Ro 13329
+IHdpY2h0aWc= 13330
+NTU= 13331
+IG90cmE= 13332
+b2dyYXBoZXI= 13333
+IGN1cmlvcw== 13334
+IDE5ODA= 13335
+IGVtcHJlcw== 13336
+ZGVzcw== 13337
+ZXVy 13338
+IGNsdXN0ZXI= 13339
+YXJ0ZXI= 13340
+b2JpbGU= 13341
+IFlhbg== 13342
+IEFkdg== 13343
+IGRpc2NpcGxpbmU= 13344
+IOygleuPhA== 13345
+IFBsYWNl 13346
+IFNlbGVjdA== 13347
+VEU= 13348
+INCx0YvQu9Cw 13349
+IHdoaXM= 13350
+IGJheQ== 13351
+IERvcg== 13352
+ZW5jaW5n 13353
+IHJlcGV0 13354
+IGZpY2Fy 13355
+cGFk 13356
+IGZvZw== 13357
+dXlvcg== 13358
+IHNuYXA= 13359
+aWJ0 13360
+IHNvYmll 13361
+IGFwcG9pbnRtZW50 13362
+IFJ5 13363
+IGNlaWxpbmc= 13364
+b3Vyc2U= 13365
+IHdyaXRlcw== 13366
+IEFmZ2hhbmlzdGFu 13367
+IG1vcw== 13368
+YXpl 13369
+IHBlbmFs 13370
+IGNyeXN0YWw= 13371
+SUNF 13372
+6rCQ 13373
+6Z8= 13374
+IFRlc2xh 13375
+IHRoZW9yaWVz 13376
+IGFwcGVhbA== 13377
+IG5ld3NwYXBlcg== 13378
+IGNvb2tpZXM= 13379
+5qk= 13380
+INin2YTZhA== 13381
+IG1hag== 13382
+IEdldHRpbmc= 13383
+a29tbWVu 13384
+IEhlYXZlbg== 13385
+ZWxscw== 13386
+IGRpdmluZQ== 13387
+xKs= 13388
+IGFrdA== 13389
+IGhvcGVz 13390
+IENoZW4= 13391
+d2VnZW4= 13392
+Kioq 13393
+IEZyYWdl 13394
+INC90Lg= 13395
+4Li5 13396
+bWluaXN0ZXI= 13397
+bmVzb3Rh 13398
+d2hpY2g= 13399
+IGV4cGxpY2l0 13400
+IHZlcmRhZA== 13401
+IGdyYWR1YXRlZA== 13402
+IFBoaWxpcHA= 13403
+UUw= 13404
+IE1J 13405
+IGRldm90 13406
+IGN1cmU= 13407
+IGNsb3Nlc3Q= 13408
+IMOE 13409
+IHNleHk= 13410
+IERlYXRo 13411
+b2tv 13412
+dWd1 13413
+IEFubmU= 13414
+aXRhcmlhbg== 13415
+ZXNh 13416
+0LXQs9C+0LQ= 13417
+IER1cg== 13418
+IDAwMA== 13419
+emVpdA== 13420
+IHRvdXJuYW1lbnQ= 13421
+IG1lbGhvcg== 13422
+4Liq 13423
+IGluZHU= 13424
+IGZsYXc= 13425
+IHdhcnM= 13426
+IE1pbmQ= 13427
+IElyb24= 13428
+0YLQsNC6 13429
+IFZS 13430
+IHNpeg== 13431
+IFNvdXRoZXJu 13432
+IOq3uOufrOs= 13433
+IGF3YWs= 13434
+IOyVng== 13435
+IGN1YmU= 13436
+YmVsaWV2YWJsZQ== 13437
+aWZhbGw= 13438
+ZGlz 13439
+IGFiYW5kb25lZA== 13440
+bWluZA== 13441
+IHBhcmw= 13442
+IGNsYXNzaWNhbA== 13443
+6Is= 13444
+4buZdA== 13445
+IEF1dG8= 13446
+IEJvcg== 13447
+56k= 13448
+NDAw 13449
+IFNvY2lldHk= 13450
+IHN1YnRsZQ== 13451
+IG1pc3Npb25z 13452
+IHJlbWVtYmVyZWQ= 13453
+IEVpdGhlcg== 13454
+IGRhZsO8cg== 13455
+T1JE 13456
+IGludGVuc2l0eQ== 13457
+RVNJTg== 13458
+IEN1cA== 13459
+IHJhcmVseQ== 13460
+IHRveXM= 13461
+IENoYXJsaWU= 13462
+4buf 13463
+IGdsYXViZQ== 13464
+IHJvdW5kcw== 13465
+VElO 13466
+IGNhcGFiaWxpdHk= 13467
+IGRlcml2YXRpdmU= 13468
+IHJlZmVycmluZw== 13469
+IGTDpQ== 13470
+IFRBTEk= 13471
+IGNvdHRvbg== 13472
+IGNvbmZlcg== 13473
+IGNvbHVtbnM= 13474
+IGxpYmVyYWw= 13475
+IG51bmNh 13476
+IM68zrU= 13477
+IGluZG8= 13478
+aWJlbg== 13479
+IEJlaXNwaWVs 13480
+IOq3uOughw== 13481
+INGD0Yc= 13482
+IGhveQ== 13483
+IGZyeQ== 13484
+IFNjb3R0aXNo 13485
+6Io= 13486
+IGNpdg== 13487
+IGNvbnNlcnZhdGl2ZQ== 13488
+IGFpcnBs 13489
+IHNhcg== 13490
+cnVz 13491
+IGludmVzdG1lbnRz 13492
+IGluZmluaXRl 13493
+IOCulQ== 13494
+IFRBTElFU0lO 13495
+IEdhcnk= 13496
+dWVsbA== 13497
+INCw0Lo= 13498
+IENpcg== 13499
+IHJpdHVhbA== 13500
+ID4+Pg== 13501
+IHRlbXB0 13502
+IFRlY2g= 13503
+IFBva2Vtb24= 13504
+IGltcHJvdmVtZW50cw== 13505
+IHNwYXJl 13506
+IHRyYW5zbGF0ZQ== 13507
+IHNvbnJh 13508
+IEZpbG0= 13509
+d29ydA== 13510
+INC80Lg= 13511
+IHBlcmlvZHM= 13512
+IGplYWxvdXM= 13513
+IHRpcg== 13514
+TUk= 13515
+IGNvbmR1Y3RlZA== 13516
+IOyViOuFlQ== 13517
+MDk= 13518
+IFBvbGl0 13519
+IFdoZXJlYXM= 13520
+IG1vaXN0dXJl 13521
+IHNpbnM= 13522
+IGthcA== 13523
+INGN0Lo= 13524
+IGJlbmlt 13525
+IGVsaW1pbmF0ZQ== 13526
+IGF0aGxldGVz 13527
+IE1hbmFnZXI= 13528
+IGZlYXR1cmVk 13529
+YXBvcmU= 13530
+IOuwnA== 13531
+IHBlcmY= 13532
+IFRodXM= 13533
+IGRlYnV0 13534
+0L7QsdGA 13535
+IHNlw7E= 13536
+IG15c3RlcmlvdXM= 13537
+d29yZHM= 13538
+lOqwgA== 13539
+IGNoZWNrcw== 13540
+IHZvbHVudGVlcg== 13541
+IHdhc2hpbmc= 13542
+IE1hcnZlbA== 13543
+IEFC 13544
+aXNzb3Jz 13545
+ISc= 13546
+IEZ1bGw= 13547
+eWVvbg== 13548
+IHdlaWdo 13549
+IEpPSE4= 13550
+IHZvcw== 13551
+IHByb2NlZHVyZXM= 13552
+IGFkZHJlc3NlZA== 13553
+IEJlcmxpbg== 13554
+cHV0ZXI= 13555
+IEJhbg== 13556
+IG1lZGljYXRpb24= 13557
+IGRyb25l 13558
+INGD0LE= 13559
+IEplYW4= 13560
+IGNhcHM= 13561
+IGRpc2FwcG9pbnRlZA== 13562
+IHdvcmU= 13563
+IOq1rQ== 13564
+IG9yZ2FuaXpl 13565
+IEhhbGxvd2Vlbg== 13566
+IGZhbnRhc3k= 13567
+eWFyZA== 13568
+IG5vc290cm9z 13569
+IGp1bXBlZA== 13570
+IHBob3RvZ3JhcGh5 13571
+IE5hbWU= 13572
+cmVj 13573
+QUI= 13574
+IGJsZXNzaW5n 13575
+IFNodXQ= 13576
+IGJpdHRlcg== 13577
+cG9w 13578
+IGRlaQ== 13579
+IGZ1bGZpbGw= 13580
+IGRlbmdhbg== 13581
+IGJlbG8= 13582
+IE1lYW53aGlsZQ== 13583
+IGRlcG9pcw== 13584
+IGRpYWJldGVz 13585
+IGJ1bmQ= 13586
+IFplYWxhbmQ= 13587
+IGRpZ2VzdA== 13588
+IHRpcmVz 13589
+IGRvZA== 13590
+YWduZQ== 13591
+4bq/dA== 13592
+IHBlZWw= 13593
+INC30LDQsQ== 13594
+IG5vZGVz 13595
+IHRyZW5kcw== 13596
+IFN3aXRjaA== 13597
+IEF3YXJk 13598
+IE9yaWc= 13599
+IEhhbA== 13600
+IGVzdGFz 13601
+IDM2MA== 13602
+IHNpbXVsdA== 13603
+IGNvbWlj 13604
+IG3DoA== 13605
+IGJhbGFuY2Vk 13606
+IFByaW5jZXNz 13607
+IGtpbG9tZXRlcnM= 13608
+4bup 13609
+IHBhcnRpcg== 13610
+7KSR 13611
+c29mdA== 13612
+IFZpZXc= 13613
+IGJpb2xvZ2ljYWw= 13614
+aW5zdA== 13615
+NDQ= 13616
+IG1hbmVyYQ== 13617
+IGNvbXByZWhlbnNpdmU= 13618
+IFNhYg== 13619
+IGNyaW1lcw== 13620
+eWVycw== 13621
+IENvbXBhbnk= 13622
+IFBob3Q= 13623
+IHBvdWNv 13624
+aWFj 13625
+IGJlaW0= 13626
+aW5hdGU= 13627
+IHN1YnNlcXU= 13628
+IE1heW9y 13629
+IGNlbnR1cmllcw== 13630
+w6hyZXM= 13631
+7J6W7JWE7JqU 13632
+IOq3uOufvA== 13633
+IEZyYXU= 13634
+IE9I 13635
+IOuBnQ== 13636
+IE5haA== 13637
+IFNlcmllcw== 13638
+IG92ZXJuaWdodA== 13639
+7ZKI 13640
+IOKAog== 13641
+IHRyYXZl 13642
+YXR0ZXJlZA== 13643
+IHdhcnJp 13644
+IEdydW5k 13645
+IEluZG9uZXM= 13646
+IHNjcmE= 13647
+b2J5 13648
+IEJyb29r 13649
+IGN1cnM= 13650
+IOu4 13651
+IGV4cGxhaW5z 13652
+cmFtYXRpYw== 13653
+IHBhcnRpY2lwYXRpbmc= 13654
+IG1pbnV0 13655
+IGNvbnRyYWN0cw== 13656
+IGdlZ2Vu 13657
+IGRpc2FwcGVhcmVk 13658
+IFNO 13659
+IHJvYnVzdA== 13660
+YXBo 13661
+IHNocmlt 13662
+IGRldmFzdA== 13663
+Y29wZQ== 13664
+IG1lZXRz 13665
+IHBlYWNlZnVs 13666
+bWF0ZQ== 13667
+IHdlbGQ= 13668
+INeq 13669
+ZG9u 13670
+0YPRgtGM 13671
+IHJlZ2lzdGVyZWQ= 13672
+IE5paw== 13673
+amlu 13674
+IGNhdg== 13675
+IGVjaHQ= 13676
+aW94 13677
+IGZsb3dpbmc= 13678
+0L3QvtGB0YLQuA== 13679
+IHRvZQ== 13680
+IGVudGl0eQ== 13681
+0L7QstCw 13682
+Zml0cw== 13683
+IFBhdHJpY2s= 13684
+0YLRgA== 13685
+IGxldmVyYWdl 13686
+IGNvcnJlbA== 13687
+aWFo 13688
+IHN0cmluZ3M= 13689
+aXN0aW5jdA== 13690
+IGd1ZQ== 13691
+YXJjaHk= 13692
+IHRlbmdv 13693
+xLFtxLF6 13694
+IG9yYml0 13695
+INC10YnRkQ== 13696
+Y2FrZQ== 13697
+INec15Q= 13698
+IE1pbm5lc290YQ== 13699
+IGJyYWtl 13700
+b3dpZQ== 13701
+IGNyYXc= 13702
+6riw66W8 13703
+IHByb2dyYW1tZQ== 13704
+INGB0LvRg9GH 13705
+aWVuY2Vz 13706
+IE91aQ== 13707
+IFBlcnM= 13708
+aW1pZW50bw== 13709
+IEludmVzdA== 13710
+IHNsb3dlcg== 13711
+IEJldGg= 13712
+IG51cnNl 13713
+IFNwcmluZw== 13714
+U3A= 13715
+IHVuZW1wbG95 13716
+0LTQuA== 13717
+IGdlbml1cw== 13718
+IEFhcm9u 13719
+IOq3uOufrA== 13720
+IGVp 13721
+IHRhbmtz 13722
+IGF1am91cmQ= 13723
+IGNvbXBsZXhpdHk= 13724
+INGA0LXRiA== 13725
+IG9sZGVzdA== 13726
+IGxldHo= 13727
+IHBoZW5vbWVub24= 13728
+cHJpbnQ= 13729
+IEJ1bmRlcw== 13730
+aXRhdA== 13731
+6ruY 13732
+IDQy 13733
+IFdp 13734
+IGluY29t 13735
+IGdlaw== 13736
+IGVtYnJhY2U= 13737
+IHRpZXM= 13738
+b3V0ZQ== 13739
+IGRvc2U= 13740
+IEZyaWVuZHM= 13741
+0YvRgg== 13742
+0LXQs9C+0LTQvdGP 13743
+IG9yZw== 13744
+hOuhnA== 13745
+w7Nn 13746
+IGV4Y2VlZA== 13747
+IGdvZHM= 13748
+IOqxsOyYiOyalA== 13749
+IHNvY2lldA== 13750
+IFVuaXZlcnM= 13751
+aXTDpHQ= 13752
+IHdvcmRlbg== 13753
+IHNtb2tpbmc= 13754
+IGludGVucw== 13755
+YWJ1bA== 13756
+ZW1pYQ== 13757
+6JE= 13758
+NDc= 13759
+Zmx5 13760
+IDIwMDY= 13761
+IFNlcmlvdXNseQ== 13762
+IHByemV6 13763
+5rw= 13764
+Y3Jl 13765
+IG5hbg== 13766
+IG1vZGVz 13767
+0L7QstCw0YLRjA== 13768
+IEhhbmc= 13769
+ZW1lbg== 13770
+IGJlbmVmaWNpYWw= 13771
+IHZvdGVycw== 13772
+IEJyb2Fk 13773
+IGJlbnQ= 13774
+V293 13775
+IG11bA== 13776
+IFVD 13777
+IGRhbWFnZWQ= 13778
+IFVrcmFpbmU= 13779
+IHdpcGU= 13780
+IHN0b25lcw== 13781
+IG1hbmFnZXJz 13782
+IHJhYg== 13783
+0YHRgtGA0L4= 13784
+bGF0 13785
+IGRlY2U= 13786
+IGdyYXBoaWM= 13787
+IGZvc3M= 13788
+IGRpc2FncmVl 13789
+IEFtZW4= 13790
+IHNlY3JldHM= 13791
+aG9sZQ== 13792
+aW5rbGU= 13793
+IGZvcnR1bmF0ZQ== 13794
+IOyx 13795
+7JyE 13796
+IGhhYml0cw== 13797
+IGJ1cmllZA== 13798
+IGhpbg== 13799
+IHZpcnR1YWxseQ== 13800
+b2xhcw== 13801
+IFJQ 13802
+IFRhYg== 13803
+bG93 13804
+IHNhY3JpZmlj 13805
+IGVzdGltYXRlZA== 13806
+b2xu 13807
+2Ys= 13808
+Y3Vy 13809
+IEZlZWw= 13810
+IGNhc3RsZQ== 13811
+IHVzZWxlc3M= 13812
+IGRpc2c= 13813
+IEphY29i 13814
+IGdhYW4= 13815
+IHVwc2lkZQ== 13816
+IHBhcmVjZQ== 13817
+IHNoaXBwaW5n 13818
+IENS 13819
+IGRpc3J1cHQ= 13820
+YWN0ZXI= 13821
+VU5E 13822
+ZnU= 13823
+IFBpY2s= 13824
+IENoYXJs 13825
+IEJ1bGw= 13826
+IGVudGVycHJpc2U= 13827
+IHB1bmlzaG1lbnQ= 13828
+YWNraW5n 13829
+IGZyYWN0aW9u 13830
+IHRhYmxldA== 13831
+IGNob3Jk 13832
+IHNpbWlsYXJseQ== 13833
+IFRvcm9udG8= 13834
+IGNvdXJ0cw== 13835
+xJ9s 13836
+ZXN6Y3pl 13837
+IHByb25vdW4= 13838
+IFNpc3Rlcg== 13839
+IE1Q 13840
+IGdyZWF0bHk= 13841
+IERhbms= 13842
+aWNvcA== 13843
+IGdhcmJhZ2U= 13844
+IHJlc29sdmU= 13845
+IFNhZg== 13846
+IEd1bg== 13847
+IGNvbXBvdW5k 13848
+IOuwsA== 13849
+IE11c2lr 13850
+4pmr 13851
+IGNoYW9z 13852
+IFdoZW5ldmVy 13853
+IGV1cm9z 13854
+IG9yY2hlc3Q= 13855
+IHJlZnJpZ2Vy 13856
+YWxhbg== 13857
+4Li3 13858
+IEFtYXppbmc= 13859
+IHB1ZA== 13860
+YWdhbg== 13861
+IGplc3pjemU= 13862
+aXN5 13863
+IGFjY3VyYWN5 13864
+IEFtYQ== 13865
+aXNvZGU= 13866
+64yA 13867
+IGludGVycHJldGF0aW9u 13868
+IExpYmVy 13869
+5rc= 13870
+Y2Ft 13871
+IGV2b2x2ZWQ= 13872
+IEtheQ== 13873
+0YbRiw== 13874
+IGNyZWF0b3I= 13875
+aXRhcw== 13876
+IGFsYXJt 13877
+IGNlbGVicmF0aW9u 13878
+emVudA== 13879
+IGZ1bmNpb24= 13880
+IG92 13881
+dW1ibGluZw== 13882
+ICU= 13883
+4LiI 13884
+IHJlc3RyaWN0aW9ucw== 13885
+INC90LDQsg== 13886
+IEtpbmRlcg== 13887
+IGJhbmFuYQ== 13888
+0YzRjw== 13889
+IGRpYW1ldGVy 13890
+IG5vcnRoZXJu 13891
+dXJlcnM= 13892
+IFBhcw== 13893
+IHdvcmtmb3JjZQ== 13894
+IGp1bmc= 13895
+IGd1YXJhbnRl 13896
+IGVxdWlsaWI= 13897
+IHN1aXRl 13898
+IGV1cm8= 13899
+IGRlbGliZXI= 13900
+U3Rl 13901
+IGRvd250b3du 13902
+IGNoaW4= 13903
+IGNvZGVz 13904
+ZWRpYQ== 13905
+IHNoZWVw 13906
+cmVzaG9sZA== 13907
+d25pZQ== 13908
+w7Ni 13909
+IHVuZGVybHlpbmc= 13910
+bGlh 13911
+amVy 13912
+z4DPjA== 13913
+550= 13914
+dGhyb3A= 13915
+IHphcA== 13916
+IHZhY3V1bQ== 13917
+IEhhYg== 13918
+IHdyYXBwZWQ= 13919
+7KI= 13920
+IGludmVudG9yeQ== 13921
+0LzQsA== 13922
+IGNvb3Jk 13923
+IHBsYXRlcw== 13924
+IHN5bW0= 13925
+VGU= 13926
+IHfFgmHFm25pZQ== 13927
+IHJlYWNoZXM= 13928
+IGxvbmVseQ== 13929
+U2NyaXB0 13930
+bGVl 13931
+ZXNzZXI= 13932
+IOqxuA== 13933
+IEdlc2No 13934
+IE1vdmluZw== 13935
+IHLDqXA= 13936
+IFZpbGw= 13937
+IFJhY2hlbA== 13938
+IHRlbW9z 13939
+T05F 13940
+IHN0cmFpbg== 13941
+IGFuZ2Vs 13942
+IGbDpQ== 13943
+VHI= 13944
+IGFjaG8= 13945
+IGhpZ2hsaWdodHM= 13946
+IFdlcg== 13947
+IENhcmw= 13948
+IGJsdXI= 13949
+IHJlZ2FyZHM= 13950
+wrc= 13951
+0LjQu9GB0Y8= 13952
+IHJlY3Jl 13953
+IFlhbmk= 13954
+VUNL 13955
+oLg= 13956
+IGVsZWN0cm9ucw== 13957
+IFNwaWVs 13958
+IHZlZA== 13959
+2r4= 13960
+IGJlYW0= 13961
+IGlkaW90 13962
+65Ok 13963
+0L3QsNGH 13964
+aWRk 13965
+IHNraQ== 13966
+aXRhdGl2ZQ== 13967
+IGh5cG90aGVz 13968
+ZW50ZXI= 13969
+IOyVhOuLiOs= 13970
+IGlocmU= 13971
+IHByZXZpZXc= 13972
+YW5nZWw= 13973
+IGRlbW9u 13974
+IGR1cw== 13975
+IGRpYw== 13976
+IEtvbQ== 13977
+TEVZ 13978
+Li4uIQ== 13979
+IHNpZWh0 13980
+IFNvbmlj 13981
+IHRlbmhv 13982
+YW5hcw== 13983
+IGRpZ2l0 13984
+IE1hYXI= 13985
+IHVuZGVyZ3JhZA== 13986
+b3VuY2Vy 13987
+dWZmeQ== 13988
+IGNvbnZlcnNpb24= 13989
+IGRpc2Nvbm5lY3Q= 13990
+IGVjaG8= 13991
+b21lcg== 13992
+IGN1cnJpY3VsdW0= 13993
+IHBlcmNow6k= 13994
+IHdhbmQ= 13995
+Li4/ 13996
+IHJvbGxlZA== 13997
+IGVudHJlcHJlbmV1cg== 13998
+IHRoZW9yZXQ= 13999
+INGJ0L4= 14000
+IGluc2lnaHRz 14001
+IHp1c2FtbWVu 14002
+b2lu 14003
+cmV0dA== 14004
+cHJvZHU= 14005
+IHZpc2l0b3Jz 14006
+ZW91cw== 14007
+IGdyYW5kbW90aGVy 14008
+IGh1bW9y 14009
+INC90LjRhQ== 14010
+emVuaWE= 14011
+aW5zb24= 14012
+IHJlc2V0 14013
+IGJhc2ViYWxs 14014
+IG1hdGNoaW5n 14015
+64uk6rCA 14016
+IHB1bnRv 14017
+7KE= 14018
+IHJlZGU= 14019
+IGFkZHJlc3Npbmc= 14020
+IGZvcmVjYXN0 14021
+IEJvbA== 14022
+IGNvbG9yZWQ= 14023
+IGRvY3VtZW50YXRpb24= 14024
+IGV4cGVjdGF0aW9u 14025
+IE5vcnRoZXJu 14026
+IGNyZW8= 14027
+IOCumg== 14028
+Zm9u 14029
+IHVuc2VyZQ== 14030
+VU0= 14031
+IGNvcGllcw== 14032
+IGV4cGFuZGVk 14033
+IHZldGVyYW5z 14034
+IEFsbQ== 14035
+INCy0L7QvtCx0YnQtQ== 14036
+IHBzeWNob2xvZ2ljYWw= 14037
+IG5vc3Nv 14038
+IHBheW1lbnRz 14039
+aW1ldGVycw== 14040
+IC0tPg== 14041
+IEplbm5pZmVy 14042
+IHZvbHVudGVlcnM= 14043
+b3NzZQ== 14044
+b3Jpb3Vz 14045
+INCx0YvQu9C4 14046
+6II= 14047
+IEVzcw== 14048
+d3M= 14049
+IEJD 14050
+IElD 14051
+V29tYW4= 14052
+IHZvbnQ= 14053
+IGV0aG5pYw== 14054
+RU5O 14055
+0LjQvNC+ 14056
+IGxvYg== 14057
+IG91aQ== 14058
+Y3M= 14059
+IHJlaGU= 14060
+IOyggQ== 14061
+IGNoaWNr 14062
+w7pzaWNh 14063
+IGtvbnQ= 14064
+IERpc3RyaWN0 14065
+IHBpbGU= 14066
+INCw0LI= 14067
+0LXQudGB0YLQsg== 14068
+IMKj 14069
+IGlzc3VlZA== 14070
+INC60L7QvNC/ 14071
+IHByb3NwZXI= 14072
+IHByb2ZvdW5k 14073
+IERlYXI= 14074
+IGZ1bmRlZA== 14075
+IGJpc2E= 14076
+npjr 14077
+158= 14078
+IOydmA== 14079
+IHR3ZWx2ZQ== 14080
+IENoYW1waW9ucw== 14081
+0YHQuw== 14082
+IDIwMDU= 14083
+cG0= 14084
+IG9uZGU= 14085
+IGRpZmbDqQ== 14086
+IENoYWxs 14087
+IGRpZmZpY3VsdGllcw== 14088
+IGdhcmFnZQ== 14089
+IGTDoQ== 14090
+w7xuaw== 14091
+IOusvA== 14092
+IHRyYW4= 14093
+IHN1Ym1pdHRlZA== 14094
+enc= 14095
+2YjYpw== 14096
+IGFyaw== 14097
+IOyEsQ== 14098
+IGdyb2Nlcnk= 14099
+0L7QvdCw 14100
+aWVyZQ== 14101
+IGFlc3Q= 14102
+IGV4aGliaXRpb24= 14103
+IHLDqXM= 14104
+IGNvbnNpc3RlbmN5 14105
+IGNvb2tpZQ== 14106
+0L3QtdC5 14107
+IHJlcGxhY2VtZW50 14108
+IFNlbQ== 14109
+IOyCrOyaqQ== 14110
+ODAw 14111
+IGdlbmVz 14112
+IHRyYW5zYWN0aW9u 14113
+IEVM 14114
+IGR1cmFudGU= 14115
+aWJsZXM= 14116
+IEVhdA== 14117
+dGFpbA== 14118
+aXNzYW5jZQ== 14119
+IHRvc3M= 14120
+IHN1cnZpdmVk 14121
+IG9mZmljZXM= 14122
+IHN1cHBvcnRpdmU= 14123
+V2hlcmU= 14124
+IHRvdXRlcw== 14125
+IOuniQ== 14126
+IGpva2Vz 14127
+aWVyb24= 14128
+YXBlcnM= 14129
+IG1hdHVyZQ== 14130
+IE1hcnNo 14131
+IHNpZG8= 14132
+a2luZA== 14133
+IHJlYWxtZW50ZQ== 14134
+IENoZWY= 14135
+IHF1ZWxxdWU= 14136
+IGp1ZGdlcw== 14137
+ZWZ0 14138
+RVJT 14139
+IGpldA== 14140
+IHBlcnNvbnM= 14141
+6Ls= 14142
+aXphdGlvbnM= 14143
+cmlr 14144
+IHNob3Bz 14145
+IFd5 14146
+IGVsZWc= 14147
+cXXDqA== 14148
+cXVvaQ== 14149
+IGp1Z2E= 14150
+IO2VnOuyiA== 14151
+IFF1ZXN0aW9u 14152
+IEdsb2JhbA== 14153
+IOyVveqwhA== 14154
+IFN0YXRpb24= 14155
+IE9oaW8= 14156
+IHN0aWNreQ== 14157
+IHN0cmVzc2Vk 14158
+IGfDvG4= 14159
+IO2d 14160
+0YHRgtGD0L8= 14161
+IFBoRA== 14162
+aW1tZXI= 14163
+IG1lbnRvcg== 14164
+IGludmVudGVk 14165
+IHJldW4= 14166
+IGluZXZpdA== 14167
+IHBvbMOtdA== 14168
+IGV4ZWN1dGU= 14169
+IFN0b3J5 14170
+IG91dHN0YW5kaW5n 14171
+IGd1ZXI= 14172
+IFJhaW4= 14173
+IGNob3Nlcw== 14174
+IFRpdA== 14175
+INGB0LXRgA== 14176
+IFNpbmdhcG9yZQ== 14177
+IE5vbmU= 14178
+IGNocm9uaWM= 14179
+sOuNsA== 14180
+IGVnbw== 14181
+RVNU 14182
+IFdhbmc= 14183
+IE5BVA== 14184
+IGF1Zw== 14185
+IGRlc2t0b3A= 14186
+IGV0ZXJuYWw= 14187
+IOyCrOyLpA== 14188
+IENvbnN0aXR1dGlvbg== 14189
+7IKs6w== 14190
+15nXnA== 14191
+cHJlcw== 14192
+INCi0Ys= 14193
+IGludGVyZg== 14194
+IGxpc3Rz 14195
+IGZpZ2h0cw== 14196
+ZnRlbg== 14197
+IElvd2E= 14198
+IG1vdGl2YXRlZA== 14199
+IEhvc3A= 14200
+IGVsc2V3aGVyZQ== 14201
+IHBhdGhz 14202
+IGluc3RhbmNlcw== 14203
+Qmw= 14204
+cmFuZ2U= 14205
+4bux 14206
+IFNpdA== 14207
+bWFuYQ== 14208
+IOyLnOyekQ== 14209
+IG3DrG5o 14210
+YW5zYXM= 14211
+IHNuYQ== 14212
+IHBoaWxvc29waA== 14213
+IHBhc3Nl 14214
+xrDhu51p 14215
+YWto 14216
+ZW50YWw= 14217
+IGlobg== 14218
+cnVjdG9y 14219
+INCy0LDRiA== 14220
+IGdlbmVyb3Vz 14221
+IHBpdm90 14222
+0L/QvtC7 14223
+IGphbWFpcw== 14224
+IGNvbWVudA== 14225
+IExldw== 14226
+b2R6aQ== 14227
+IFhib3g= 14228
+INCy0L7QtA== 14229
+IGNvbnNlbnQ= 14230
+ieyepQ== 14231
+IGRpc3Bhcg== 14232
+bGFzcw== 14233
+IEdvdmVybm9y 14234
+QmVpZmFsbA== 14235
+IOqwnA== 14236
+IGJlbG92ZWQ= 14237
+16DXlQ== 14238
+c2VsbA== 14239
+IGhvbm9yZWQ= 14240
+bGVo 14241
+IHfDpHJl 14242
+dW50aW5n 14243
+IGZyYXVk 14244
+IFJBTQ== 14245
+6rG4 14246
+IGtpbGxz 14247
+IGVjb25vbWljcw== 14248
+MDQ= 14249
+0L/QtdGA 14250
+IGNvaXNhcw== 14251
+INC40LPRgA== 14252
+w61t 14253
+IG3DtmNodGU= 14254
+IOy1nA== 14255
+IHN0aW11bA== 14256
+IGZhc3Rlc3Q= 14257
+bHY= 14258
+IGfDqW4= 14259
+IFNvdW5kcw== 14260
+IDE5NzA= 14261
+IGhvbWV3b3Jr 14262
+c3BlYWtpbmc= 14263
+IGVuY291cmFnaW5n 14264
+IHF1ZXJ5 14265
+IHJldmVycw== 14266
+cHJvZml0 14267
+IGR5 14268
+IOyekQ== 14269
+64qU642w7JqU 14270
+IHNvYXA= 14271
+IEdhbGw= 14272
+IENO 14273
+IEFucw== 14274
+IGZpYw== 14275
+YW5rcw== 14276
+IGRlc3NlcnQ= 14277
+IOyggO2drA== 14278
+IE1ha2luZw== 14279
+IGNvbWXDpw== 14280
+6rOE 14281
+IGFzc29jaWF0aW9u 14282
+RGFk 14283
+aGVl 14284
+IGhvZ3k= 14285
+IGFwcm8= 14286
+IGludmlzaWJsZQ== 14287
+QW1lcmljYW4= 14288
+7Y4= 14289
+IHZpYmU= 14290
+IGVtaXNzaW9ucw== 14291
+IGFkdm9jYXRl 14292
+IGtpY2tlZA== 14293
+IHZlbA== 14294
+IHN1bW1hcg== 14295
+IGZyZWFraW5n 14296
+Y2hyb24= 14297
+IHBpbmNo 14298
+IHdzenlzdGs= 14299
+aXNjYWw= 14300
+IHByb3ZlZA== 14301
+IG1pbmRmdWw= 14302
+IHTDpA== 14303
+IG5vaXNlcw== 14304
+IGlzb2xhdGVk 14305
+IGNyb3NzZWQ= 14306
+IOqwlQ== 14307
+IHZvaWzDoA== 14308
+IGNob3Jl 14309
+IFJB 14310
+Q29t 14311
+IHJlbGF4ZWQ= 14312
+YXRybw== 14313
+IHByZXZlbnRpb24= 14314
+Vm9pY2VvdmVy 14315
+T0Q= 14316
+IENvdmlk 14317
+IHNlcGFyYXRpb24= 14318
+IC1b 14319
+0LjRh9C10LPQvg== 14320
+IFNE 14321
+YmxlZXA= 14322
+IGluZGVwZW5kZW5jZQ== 14323
+IHBhcnRpYWw= 14324
+IGFsZ29yaXRobXM= 14325
+IEFueW9uZQ== 14326
+IGFzc29jaWF0ZQ== 14327
+aHVt 14328
+aWN1bGFy 14329
+IGLhuqFu 14330
+IGJhdHRsZXM= 14331
+R29vZA== 14332
+QXBwbGF1c2U= 14333
+IGJhc3RhbnRl 14334
+IGFkdmFudA== 14335
+IFN3ZWV0 14336
+IHJlZnVzZWQ= 14337
+INGC0LXQsdC1 14338
+cGxldA== 14339
+IGVuY291cmFnZWQ= 14340
+IG1pcmFjbGU= 14341
+IEJ1bg== 14342
+IFZhcg== 14343
+cmltaW5hdGlvbg== 14344
+ZWxlY3Q= 14345
+IE11bHQ= 14346
+IGRlbGl2ZXJpbmc= 14347
+ZWluZw== 14348
+IGNt 14349
+bmVobWVu 14350
+IExpbmU= 14351
+IOunjA== 14352
+ZW5jZWQ= 14353
+IFNvdW5k 14354
+IENvbnRpbg== 14355
+aWpk 14356
+VU5H 14357
+a2xl 14358
+IHRocmVzaG9sZA== 14359
+IGNvbXBhY3Q= 14360
+YWR0 14361
+IHRvZXM= 14362
+IFB1cg== 14363
+b3duZWQ= 14364
+bWVudGVk 14365
+IGRlc2lnbmluZw== 14366
+IHZhY2NpbmF0ZWQ= 14367
+IGV4aGF1c3Q= 14368
+IGJhc2ljcw== 14369
+IGNvbnNpc3Rz 14370
+IEd1eQ== 14371
+YWN6eQ== 14372
+IG3DrQ== 14373
+d29u 14374
+IDg1 14375
+5oI= 14376
+IG11bQ== 14377
+IGlnbm9y 14378
+IHByaW50aW5n 14379
+YWN1bGFy 14380
+cG93 14381
+IGV4cGFuZGluZw== 14382
+IGdpcg== 14383
+IENhYg== 14384
+7Zi4 14385
+0YLRjNGB0Y8= 14386
+IOyXrOufrOu2hA== 14387
+IGFuZ2xlcw== 14388
+IHRlcm1pbmFs 14389
+IFdvbg== 14390
+IEludGVyZXN0aW5n 14391
+IGNyb3NzaW5n 14392
+IGJvbmRz 14393
+IHB1ZWRlbg== 14394
+IG9yYg== 14395
+bGFyxLFu 14396
+IGNyZWVweQ== 14397
+IG51dHJpdGlvbg== 14398
+IGFsbGllcw== 14399
+IHdpcmVsZXNz 14400
+IGRlc2lyZWQ= 14401
+IGNvbXB1dGU= 14402
+IEFyaXpvbmE= 14403
+IEJlYXV0aWZ1bA== 14404
+IHByb2R1Y2Vz 14405
+IG51ZXN0cm8= 14406
+dGVk 14407
+IGVsaWdpYmxl 14408
+INGB0L7Qtw== 14409
+aWNpYWw= 14410
+IEhlcm8= 14411
+IGNvbnN1bWU= 14412
+IHJvYm90cw== 14413
+IHB1cmNoYXNlZA== 14414
+Y2Npw7Nu 14415
+IGl6 14416
+xrDhu6Nj 14417
+zq/Ovc6xzrk= 14418
+INij2YY= 14419
+IHNoYWRvd3M= 14420
+IE1lZGlh 14421
+IHByaW5jZXNz 14422
+IGtsYXI= 14423
+IHdvb2Rlbg== 14424
+IHVzYXI= 14425
+IGfDvHplbA== 14426
+IHNsb3Q= 14427
+cmFkZQ== 14428
+IOuS 14429
+IGhhcm1vbg== 14430
+IGluZ3JlZGllbnQ= 14431
+b3JzaGlw 14432
+ZWtp 14433
+IGdyYW5kZmF0aGVy 14434
+IGV4Y2l0ZW1lbnQ= 14435
+IHBvbGl0aWNpYW5z 14436
+Li4h 14437
+IG91dHM= 14438
+IHNlcGFyYXRlbHk= 14439
+INGP0Lo= 14440
+IFdlbHQ= 14441
+IFBvdw== 14442
+amFu 14443
+IG9yaWVudGF0aW9u 14444
+TEM= 14445
+YWdlbQ== 14446
+24zaug== 14447
+IGJyYW5jaGVz 14448
+YWRlbg== 14449
+cmVudGU= 14450
+IElocg== 14451
+YXNt 14452
+IGVzdMOjbw== 14453
+IE5pYw== 14454
+IHNsYXZl 14455
+IGNvbXByZXNz 14456
+Y3Jvd2Q= 14457
+IGNsaW1iaW5n 14458
+IE1hbmFnZW1lbnQ= 14459
+IEJhaA== 14460
+IHBhbmlj 14461
+IGtvcg== 14462
+IGNvb2xpbmc= 14463
+IGJpbmQ= 14464
+INC30LDQtA== 14465
+IHJhY2s= 14466
+IGVudGl0 14467
+IHNlbmRz 14468
+IHlvdXJzZWx2ZXM= 14469
+ZGVz 14470
+IE11c2xpbXM= 14471
+IO2a 14472
+aXNtYQ== 14473
+Y3ljbGU= 14474
+dW5rdA== 14475
+IENvcmU= 14476
+IGluanVyaWVz 14477
+IGlkZW50aWNhbA== 14478
+0LrQsNGP 14479
+IERldXRzY2hsYW5k 14480
+INC10LU= 14481
+aXNhbg== 14482
+IHRydWM= 14483
+bGV0b24= 14484
+IGJhY2t1cA== 14485
+IHVsdHJh 14486
+IGFidW5k 14487
+aWxsZXVycw== 14488
+IGJ5xYJv 14489
+b3J0ZWQ= 14490
+IGVhcnRocXU= 14491
+INC60Ls= 14492
+IG9ic2VydmF0aW9u 14493
+IG1haW50ZW5hbnQ= 14494
+ZWxlbg== 14495
+IHNldHRsZWQ= 14496
+IHBlbGE= 14497
+IEVjb25vbQ== 14498
+INU= 14499
+IHN0ZWVyaW5n 14500
+IEFMTA== 14501
+IENoZXI= 14502
+IHBhdGllbmNl 14503
+IFNub3c= 14504
+IGJvcg== 14505
+IHdvcnRoeQ== 14506
+IGPDoWk= 14507
+INen 14508
+IM66zrE= 14509
+ZG9n 14510
+IEthcmVu 14511
+aWxsZXM= 14512
+zrI= 14513
+IGFncmljdWx0dXJl 14514
+15XXnw== 14515
+IFNlYW4= 14516
+IHNlbnNvcnM= 14517
+7ZW06w== 14518
+YWdo 14519
+IHB1YmxpY2x5 14520
+IHBldXg= 14521
+IEFsZXhhbmRlcg== 14522
+IHByaW9yaXQ= 14523
+IGxhenk= 14524
+YXJkb24= 14525
+YXR0ZXJpbmc= 14526
+IGNvc3R1bWU= 14527
+2LPYqg== 14528
+IHVudw== 14529
+0Js= 14530
+IHRoaWNrbmVzcw== 14531
+cXVpdG8= 14532
+Z3VudA== 14533
+aXN0YXM= 14534
+bmV5cw== 14535
+IOuQmOqyjA== 14536
+IEJyYXNpbA== 14537
+IHRva2Vu 14538
+IGFmZmlsaQ== 14539
+bG9u 14540
+IGbDpXI= 14541
+IEJlYWNo 14542
+IHdpdGNo 14543
+IFNldmVu 14544
+IHBhbnQ= 14545
+zrvOuw== 14546
+IGNhcHRhaW4= 14547
+5Z0= 14548
+IHZldXQ= 14549
+IHBvdXZvaXI= 14550
+YWN6 14551
+IEJhcmI= 14552
+IHV0aWxpdHk= 14553
+IGNvbnRlbXBvcmFyeQ== 14554
+IG9idGFpbmVk 14555
+IHBhaW50aW5ncw== 14556
+ZWFy 14557
+IHBlYW4= 14558
+IE9n 14559
+IGN1c3Q= 14560
+0LvQtdC8 14561
+gpjr 14562
+IElzc28= 14563
+IGFjb250ZQ== 14564
+IFRlbGU= 14565
+IEFzc2lzdGFudA== 14566
+w4k= 14567
+7ZaI7Iq164uI64uk 14568
+IGNvdW50cw== 14569
+IGJ1Y2s= 14570
+IERlZXA= 14571
+IHRhY2tsZQ== 14572
+IGhhcnNo 14573
+IGRlY2lkZXM= 14574
+LuKAiw== 14575
+IEFuZ2Vs 14576
+IGxheWluZw== 14577
+IGNhbG9yaWVz 14578
+IGNvbnRyb2xsaW5n 14579
+IGFkdmFudGFnZXM= 14580
+INGN0YLQvtC5 14581
+IGFwcHJvYWNoaW5n 14582
+IHRocmVhdHM= 14583
+YWthbg== 14584
+ZW1hdGlj 14585
+bWFubg== 14586
+6rO1 14587
+bXVtYmxlcw== 14588
+YWNpw7M= 14589
+IG1haW50YWluaW5n 14590
+IGZvdW5kZXI= 14591
+bGFo 14592
+ZmlnaHQ= 14593
+IGFkbWl0dGVk 14594
+4oCmLg== 14595
+lYw= 14596
+YWJvbA== 14597
+IHVzYWdl 14598
+IG5vbnNlbnNl 14599
+IFBhbGVzdA== 14600
+IGNvbnRyZQ== 14601
+IERlbW9jcmF0aWM= 14602
+IEVS 14603
+amVrdA== 14604
+IGFyYml0 14605
+INCz0L7Quw== 14606
+IE1pY2hlbGxl 14607
+aWNoZXI= 14608
+ZXNo 14609
+IFBobw== 14610
+0LrQvtC8 14611
+NDk= 14612
+IEVuZXJneQ== 14613
+zr/PjQ== 14614
+IGNlbnRz 14615
+IHJlZmVycw== 14616
+IGdvc3BlbA== 14617
+IFNoYQ== 14618
+IFNoYXJl 14619
+15nXoA== 14620
+IGNsaW5pYw== 14621
+IOuEow== 14622
+IGVxdWFsaXR5 14623
+dWdz 14624
+IHNoZWQ= 14625
+IHBsYW5lcw== 14626
+IHRvdXRl 14627
+cmVjaw== 14628
+IHN0cmFuZA== 14629
+IGJpb2xvZ3k= 14630
+IGxlYWd1ZQ== 14631
+IFBvaw== 14632
+IG7Dum1lcm8= 14633
+IENvYXN0 14634
+IGNvbnNpc3RlbnRseQ== 14635
+IG51Y2xl 14636
+T09PTw== 14637
+IG9iamV0 14638
+IGNob3I= 14639
+IGdpbmdlcg== 14640
+IGRhYmVp 14641
+IGNvb3BlcmF0aW9u 14642
+4K+NLg== 14643
+bnRlbg== 14644
+56Q= 14645
+bMOg 14646
+7JaR 14647
+cmFkbw== 14648
+IHBhc3NpdmU= 14649
+IGdsb3Zlcw== 14650
+IHVuZGVyZ3JvdW5k 14651
+IGxvZ2ljYWw= 14652
+IGtldA== 14653
+IGZ1bmN0aW9uYWxpdHk= 14654
+uOumrA== 14655
+IHBvcnRhbA== 14656
+ZWxsZXI= 14657
+15nXqA== 14658
+IFRlZA== 14659
+IEdyZQ== 14660
+kJw= 14661
+IHBlcnNvbm5lbA== 14662
+IGVtZXJnaW5n 14663
+IEbDvHI= 14664
+IG1lYW50aW1l 14665
+dXNhbGVt 14666
+IENsZWFy 14667
+IHRyYXBwZWQ= 14668
+IOyasA== 14669
+IGRpc3Bs 14670
+IG1ldHRyZQ== 14671
+IG11bmljaXA= 14672
+IHdpdGhkcmF3 14673
+IHNwYXQ= 14674
+dW5lcw== 14675
+IGFjY2Vzc2liaWxpdHk= 14676
+IGFwYXJl 14677
+IHByb3NwZWN0 14678
+INC90LDQtw== 14679
+IGNvcHBlcg== 14680
+IFBSTw== 14681
+z4XPhA== 14682
+IGF0dGFja2luZw== 14683
+IFZpbg== 14684
+IFN0b25l 14685
+IGludmVzdGlnYXRl 14686
+c3R5bGU= 14687
+IM67 14688
+66Gd 14689
+66eI 14690
+IGluc3BlY3Q= 14691
+IGxpdmVy 14692
+0LDQu9C40YHRjA== 14693
+IHNlcmE= 14694
+aGFsdGVu 14695
+ZW1hbg== 14696
+IG1pbmlzdHJ5 14697
+Jyc= 14698
+IGRvdHM= 14699
+44WL44WL44WL44WL 14700
+0YPRgdGC 14701
+IEphaw== 14702
+QUtF 14703
+IGdhcHM= 14704
+dWNrZXI= 14705
+INC40L3RgtC10YDQtdGB 14706
+IEVtaWx5 14707
+IGludGVydmFs 14708
+IHRlbmRlcg== 14709
+IFRlY2hub2xvZ3k= 14710
+Z2FtZQ== 14711
+IHRyaWI= 14712
+2YTYpw== 14713
+IERldmVsb3BtZW50 14714
+2YXYpw== 14715
+IHdyaXN0 14716
+IGZpcmVz 14717
+IHRhcmdldGVk 14718
+7KCQ 14719
+IHNvZA== 14720
+7ZqM 14721
+IG9sZHXEnw== 14722
+IHNlYXNvbnM= 14723
+dmVudGlvbnM= 14724
+INC90LXQs9C+ 14725
+IHNvbWV0aW1l 14726
+0LvQuNCy 14727
+bsOp 14728
+IHTDug== 14729
+IERldXM= 14730
+IGV4ZWN1dGlvbg== 14731
+w6Fw 14732
+IENoYW5nZQ== 14733
+IEluZGVlZA== 14734
+IHJlZ3VsYXRpb24= 14735
+IEh1bmc= 14736
+w6lpcw== 14737
+IHdpc2hlcw== 14738
+IGpheno= 14739
+IHN0cnVjdHVyYWw= 14740
+IGJsb3dpbmc= 14741
+IGJ5xIc= 14742
+IHRoZXJtYWw= 14743
+cGhhbnQ= 14744
+0YDRg9C3 14745
+0LDQvdGC 14746
+IFB1bGw= 14747
+IGNvbmZ1c2lvbg== 14748
+0L3Ri9C80Lg= 14749
+IHNjZW5hcmlvcw== 14750
+7KCB7Jy866Gc 14751
+INC00LXRgg== 14752
+IHRhdHRvbw== 14753
+IGF1dHJl 14754
+IGhlYXRpbmc= 14755
+IHRyZWF0aW5n 14756
+INC/0L7QvdC40Lw= 14757
+IGV4Y2x1cw== 14758
+IExPTA== 14759
+d2Vhcg== 14760
+YWdsZQ== 14761
+IHp1csO8Y2s= 14762
+IHJhdGlvbmFs 14763
+c3U= 14764
+IGRldGVy 14765
+IE5hdGl2ZQ== 14766
+4K6V4K6z 14767
+YWNoZWQ= 14768
+IOOD 14769
+IEVudG9uY2Vz 14770
+IGhvcmE= 14771
+7J207JeQ7JqU 14772
+IGxpdGU= 14773
+w6s= 14774
+IHNpeHRo 14775
+INCx0L7Qu9C10LU= 14776
+YWN0b3I= 14777
+IHBzeWNob2xvZ3k= 14778
+IGRlbWFuZHM= 14779
+IHBlZXI= 14780
+IG5ld2x5 14781
+IFdXRQ== 14782
+RG9uYWxk 14783
+IEJveA== 14784
+IHBpbmU= 14785
+IGxvYWRpbmc= 14786
+IE5pY28= 14787
+IHPFgg== 14788
+b21tZQ== 14789
+QVJU 14790
+IHJlY3J1aXQ= 14791
+IGJ1Z3M= 14792
+YXJlbnRz 14793
+INC/0YDQvtCx 14794
+IEluc2lkZQ== 14795
+aXBwZXI= 14796
+ZHJhbWF0aWM= 14797
+IHBsYW5ldHM= 14798
+b3JkZQ== 14799
+IHlvZ2E= 14800
+Y2hpbGQ= 14801
+IE1hcmll 14802
+IEJM 14803
+IGZpbG1lZA== 14804
+IHJlZnJlc2g= 14805
+IHRvbWF0b2Vz 14806
+IGZldA== 14807
+UXXDqQ== 14808
+ICEh 14809
+IOuCtOs= 14810
+cmluZQ== 14811
+IGludGVyYWN0aXZl 14812
+c2Fs 14813
+YW5uYWg= 14814
+cGV6 14815
+IHVuZGVyc3RhbmRz 14816
+IFRva3lv 14817
+IGxpYnJhcmllcw== 14818
+IHJlYWRlcg== 14819
+kZA= 14820
+b3o= 14821
+IEVuZGU= 14822
+IEZsbw== 14823
+IG1pbGQ= 14824
+IHBvZXRyeQ== 14825
+INC20LjQsg== 14826
+IGJlaGF2ZQ== 14827
+IGRvZW4= 14828
+IFN1c2Fu 14829
+cGFnZQ== 14830
+cmFoYW0= 14831
+IGNvbW11bmljYXRpb25z 14832
+IHR1bmluZw== 14833
+IHBhYw== 14834
+IGFueGlvdXM= 14835
+SU8= 14836
+TWFyaw== 14837
+IGhpw6c= 14838
+Ym9va3M= 14839
+IHBpc3M= 14840
+IGVuYWJsZWQ= 14841
+YWNoZWxvcg== 14842
+IEZPUg== 14843
+IMOpYw== 14844
+IFRS 14845
+aWxzdA== 14846
+aGF0 14847
+IOydjA== 14848
+IHR5Y2g= 14849
+IGphcg== 14850
+IGJ1aWxkcw== 14851
+IEFyZ2VudA== 14852
+IGludGVybWVkaQ== 14853
+IGxvdQ== 14854
+IGFyYQ== 14855
+IGFzc2lnbm1lbnQ= 14856
+IGNhYmluZXQ= 14857
+IHJldGlyZW1lbnQ= 14858
+IGRpc2FibGVk 14859
+cmljYQ== 14860
+IGF3YXJkcw== 14861
+IGJvb3Rz 14862
+IGFja25vd2xlZA== 14863
+IHRoeQ== 14864
+IOq1rA== 14865
+IHN5bmQ= 14866
+0L3QuNC5 14867
+aWx0b24= 14868
+IHByb2Js 14869
+IEZhbA== 14870
+IHZlcmRhZGU= 14871
+IDcwMA== 14872
+IExlYXJuaW5n 14873
+b2N1cw== 14874
+IHBhbGFjZQ== 14875
+Tm90 14876
+dGFpbg== 14877
+Y20= 14878
+IG1hZ25ldA== 14879
+aW5jb2xu 14880
+IGZpZ3VyaW5n 14881
+IEx5bg== 14882
+IEJvc3M= 14883
+IFZP 14884
+IGRpYWdub3Npcw== 14885
+IGVxdWlwcGVk 14886
+d2F0Y2g= 14887
+aW5vcw== 14888
+YWRlcnM= 14889
+IHNoZWxm 14890
+IG9yZ2FuaXM= 14891
+IG5vZA== 14892
+IGvEsXo= 14893
+cHBlcnM= 14894
+IHJlc3RvcmU= 14895
+IGFydGlj 14896
+IFZvaWNl 14897
+xLF5b3J1bQ== 14898
+6rKp 14899
+IHNwcmVhZGluZw== 14900
+IGhpcHM= 14901
+IHdhcmQ= 14902
+dXJlYXU= 14903
+IGludGVyc2VjdGlvbg== 14904
+NjY= 14905
+IDM5 14906
+57M= 14907
+IHdhaXRlZA== 14908
+7LQ= 14909
+aGhoaA== 14910
+IGR5cw== 14911
+IEVO 14912
+IGJhdGNo 14913
+IGNhZg== 14914
+IG1hcmtlcg== 14915
+b3JhYmxl 14916
+w7NyaWE= 14917
+IHN0ZXBwZWQ= 14918
+IGNlbGVicmF0aW5n 14919
+0LDQvdCw 14920
+IHdvcm4= 14921
+IEZvbA== 14922
+IHBsYQ== 14923
+IGF0dGVtcHRz 14924
+IHR3ZWV0 14925
+IHJ1c3Q= 14926
+Z2VuY2U= 14927
+7Ya1 14928
+IHJldmVs 14929
+IHJlY2VwdA== 14930
+ZW5lc3M= 14931
+ICgo 14932
+IeKAiw== 14933
+IOyGkA== 14934
+IGluZmx1ZW5jZWQ= 14935
+0LjQtg== 14936
+INC60L7QvdC10YfQvdC+ 14937
+IGNvbGxlZ2Vz 14938
+aW9uaQ== 14939
+IHNhZw== 14940
+QW5u 14941
+b2xhcg== 14942
+IGV4cHJlc3Npb25z 14943
+IHN1aXRz 14944
+IG93bmVyc2hpcA== 14945
+ZWxhbmQ= 14946
+cGllY2U= 14947
+IGRlc3B1w6lz 14948
+IHRlbA== 14949
+IGluc3VsdA== 14950
+IOq1ieyepQ== 14951
+IFNtYWxs 14952
+IEZS 14953
+b2th 14954
+YmVycmllcw== 14955
+IEFudG9u 14956
+0LXQu9GP 14957
+0Y/RgQ== 14958
+IHZhbHZl 14959
+YWN0cw== 14960
+IHdvb2Rz 14961
+4K6j 14962
+IGN1bHRpdg== 14963
+IGbDoQ== 14964
+IGNoZWVycw== 14965
+IGFzc3VtcHRpb24= 14966
+IGZpdG5lc3M= 14967
+w61jdWw= 14968
+IHBvZHI= 14969
+IHdlaXQ= 14970
+IEhpbmQ= 14971
+IGRpZ24= 14972
+INC30L0= 14973
+IHNxdWFk 14974
+IGRlc3Rybw== 14975
+Y2VyZQ== 14976
+c2hpcnQ= 14977
+aW1tdA== 14978
+ZW5nZXJz 14979
+IHPDpA== 14980
+a8WCYWQ= 14981
+IMiZ 14982
+IG9jY2Fz 14983
+IOykhA== 14984
+IHByb2Nlc3Nvcg== 14985
+IERN 14986
+IERhZGR5 14987
+IHNvb25lcg== 14988
+IHN0cmFpZ2h0Zm9yd2FyZA== 14989
+IGRlcGFydG1lbnRz 14990
+IENocm9tZQ== 14991
+IHdvcmtwbGFjZQ== 14992
+IFB5dGhvbg== 14993
+IG1lbmc= 14994
+IERBTg== 14995
+IEljZQ== 14996
+IOuIiA== 14997
+IEdp 14998
+IGhpcmluZw== 14999
+IGxhbmRlZA== 15000
+IGRlbW9jcmF0aWM= 15001
+aWVkeg== 15002
+IHNldg== 15003
+aWNpYQ== 15004
+IGVzcGVjaWFs 15005
+IE5vdXM= 15006
+IGjDpHQ= 15007
+IGJvdQ== 15008
+cGVydA== 15009
+aWVzeg== 15010
+IHZpbA== 15011
+xZtsaQ== 15012
+IMOubg== 15013
+IGxvc3Nlcw== 15014
+IHRvYXN0 15015
+IHJlYWxt 15016
+IEF1c3Rpbg== 15017
+IEluZm9ybWF0aW9u 15018
+IHJlc3VtZQ== 15019
+IGNoYXNl 15020
+IHNhbGFyeQ== 15021
+IOu2hA== 15022
+0LvQuNGH 15023
+INGB0LvQtdC0 15024
+IEZ1cnRoZXI= 15025
+IGNhcmluZw== 15026
+IHZpZw== 15027
+IHZhbG9y 15028
+INGH0LA= 15029
+IGFuYWx5dGljcw== 15030
+IGdsb2Jl 15031
+IE1BTg== 15032
+IG5lbA== 15033
+7J207JW8 15034
+n7w= 15035
+IG95 15036
+7ZWY7IS47JqU 15037
+amVu 15038
+IHRyb3VibGVz 15039
+YWhhaGE= 15040
+IGNodXJjaGVz 15041
+dWV0 15042
+IG1lYXN1cmVtZW50cw== 15043
+Ymls 15044
+7L0= 15045
+aWZ1bGx5 15046
+0LjQvdGD 15047
+IFdpbHNvbg== 15048
+prQ= 15049
+IO2MjA== 15050
+IOywqA== 15051
+IHDDumJsaWM= 15052
+IEplcnVzYWxlbQ== 15053
+IG5haWxz 15054
+IHNwaW5l 15055
+IGhlbW9z 15056
+IHpu 15057
+cXVpcw== 15058
+IExlYmVu 15059
+IHJlZmVyZW5jZXM= 15060
+SVRI 15061
+aXBlcg== 15062
+INGB0LXQsdGP 15063
+7IE= 15064
+IFdh 15065
+c3RhdGU= 15066
+p50= 15067
+IEdlbmVy 15068
+IGFjdHJlc3M= 15069
+IEVuam95 15070
+4LmD 15071
+INeS 15072
+IGluZmVjdGVk 15073
+IHNoYWtpbmc= 15074
+IG5pY2s= 15075
+4Li4 15076
+IGZvdA== 15077
+IGFjY29tcGxpc2hlZA== 15078
+dWtl 15079
+IHNoZWV0cw== 15080
+IGZlbmNl 15081
+IG51cnNpbmc= 15082
+IGludHJvZHVjaW5n 15083
+IGZlYXQ= 15084
+T25l 15085
+VE8= 15086
+IGNsdWJz 15087
+IEJydWNl 15088
+b25nZQ== 15089
+Y2hhbmdl 15090
+IEJhdG1hbg== 15091
+IE9mZmljZXI= 15092
+IGh5ZHJv 15093
+IHN1cHBsZW1lbnQ= 15094
+IGNlbGE= 15095
+IGxvbmdlc3Q= 15096
+IGNvbXBldGluZw== 15097
+IGNvbmhl 15098
+Z2l2aW5n 15099
+IGJyYWlucw== 15100
+IGxvYW5z 15101
+IHdhZ2U= 15102
+IENsaW50b24= 15103
+IHPEgw== 15104
+YW5lb3Vz 15105
+IGxvcmQ= 15106
+0YDRg9C2 15107
+IHF1aXo= 15108
+IHN0aWZm 15109
+IExHQg== 15110
+c3o= 15111
+TUU= 15112
+bWFyZQ== 15113
+dGhlcmU= 15114
+IG7DpHI= 15115
+IE1hbmQ= 15116
+bGFzdA== 15117
+IGRhZw== 15118
+IGhhbGZ3YXk= 15119
+IEJhbmQ= 15120
+IOuLpOyLnA== 15121
+IEFyZW4= 15122
+IGlsZQ== 15123
+UE4= 15124
+ZW50bw== 15125
+IGFsZ3Vt 15126
+IHNvY2Nlcg== 15127
+IGJsb2NrZWQ= 15128
+IEpvbmF0aGFu 15129
+IHNldw== 15130
+IFRlc3RhbWVudA== 15131
+IHZhbGU= 15132
+IGJlaGF2aQ== 15133
+IGNvbm5h 15134
+SUNI 15135
+IGF1ZGllbmNlcw== 15136
+bWw= 15137
+YW1tYWQ= 15138
+IOyCtOw= 15139
+SUdI 15140
+IHJhY2Vz 15141
+ZW1lZA== 15142
+IG3hu5l0 15143
+w68= 15144
+IG92ZXJz 15145
+IGRlY2xhcmVk 15146
+IHNhbmE= 15147
+IFVuYQ== 15148
+INGA0LU= 15149
+dWNrcw== 15150
+IHBhaXJz 15151
+IGFuZ2U= 15152
+TmU= 15153
+IHVwcw== 15154
+YXZ5 15155
+w7hy 15156
+cmVlaw== 15157
+IGJlaGF2aW9ycw== 15158
+IHJlZmxlY3RlZA== 15159
+IHByaW9yaXRpZXM= 15160
+IGNvbmR1 15161
+IHJldHJlYXQ= 15162
+IGV4cGVuc2Vz 15163
+IOu0kA== 15164
+IHRyaXBsZQ== 15165
+IOq1ieyepe2eiA== 15166
+w6RsdA== 15167
+IGluZGlnZW5vdXM= 15168
+IG1pbmluZw== 15169
+IGFjY2VwdGFibGU= 15170
+IHJ1aW4= 15171
+Q0E= 15172
+dWluZQ== 15173
+IHBpcGVsaW5l 15174
+Y3RpYw== 15175
+w6p0 15176
+INCy0YHQtdCz0L4= 15177
+IGJvdW4= 15178
+IERpZ2l0YWw= 15179
+IEJvb20= 15180
+0YbQtQ== 15181
+INC70YPRhw== 15182
+IGFzYw== 15183
+jIDroZw= 15184
+IEdvb2RieWU= 15185
+IHJlbmRlcg== 15186
+ZW5leg== 15187
+YXJyZQ== 15188
+IFRIQVQ= 15189
+Ym91cg== 15190
+aWNpw7Nu 15191
+RXZlcnk= 15192
+IHdpcmVz 15193
+IFBhcmxpYW1lbnQ= 15194
+bnVuZw== 15195
+YXRldXI= 15196
+IFNhdmU= 15197
+IFBoeXM= 15198
+IGFtb3I= 15199
+IEV2ZQ== 15200
+IGZyaWdodA== 15201
+IGdhbW1h 15202
+IG1pY3Jvcw== 15203
+bWl0dA== 15204
+IENvZGU= 15205
+IEJleQ== 15206
+cGxlZA== 15207
+INC40YHQv9C+0LvRjNC3 15208
+55c= 15209
+7IOJ 15210
+IG1vbmV0 15211
+IEphaHJl 15212
+IGx1eHVyeQ== 15213
+IGRlYWY= 15214
+IGJldHJheQ== 15215
+IOqysA== 15216
+0LjQutC4 15217
+IGRlZmVhdGVk 15218
+IHVuZGVydA== 15219
+IHdlZw== 15220
+IGNvb2xlcg== 15221
+aWFtaQ== 15222
+IEplc3NpY2E= 15223
+IEpveQ== 15224
+IHNvcGhpc3RpYw== 15225
+0LXQvdC40Lg= 15226
+8J2Y 15227
+IGNoaWxp 15228
+IFR5cGU= 15229
+IHByb3RlaW5z 15230
+IHByZXNlbnRpbmc= 15231
+YWxpYQ== 15232
+7Jq4 15233
+IE1ham9y 15234
+IG1vbGVjdWxl 15235
+dW1lcg== 15236
+IGNvbGxhcHNl 15237
+IEFueXdheXM= 15238
+IE1vdW50YWlu 15239
+YW50ZWQ= 15240
+44CQ 15241
+INCy0LjQtNC10L4= 15242
+QXVk 15243
+IGNvbnF1 15244
+IHZvbGw= 15245
+IGtuaXQ= 15246
+IG1lbWJy 15247
+IE1hcmtldA== 15248
+IGRhcmk= 15249
+IGNhbGN1bGF0ZWQ= 15250
+0LPQuA== 15251
+IHNocmltcA== 15252
+IE11 15253
+INC/0YDQvtGC 15254
+IOyYgeyDgQ== 15255
+IHByb2R1Y3Rpdml0eQ== 15256
+IGNvZ25pdGl2ZQ== 15257
+IEhlYg== 15258
+aWN0aW9ucw== 15259
+6rK9 15260
+IGNyw6k= 15261
+ZsO2cg== 15262
+IHByYXlpbmc= 15263
+YXNoaQ== 15264
+IFRpaw== 15265
+w7Ny 15266
+d2Vu 15267
+0YzRjg== 15268
+aXhv 15269
+ICgi 15270
+INGC0LXQuw== 15271
+IOyWtOuWpA== 15272
+INC/0LXRgNC10LQ= 15273
+IERyaXZl 15274
+44CR 15275
+IEVxdQ== 15276
+IGVxdWlsaWJyaXVt 15277
+IGRlc2NyaWJlcw== 15278
+0L3QtdC1 15279
+NDI= 15280
+IEN1cnJlbnQ= 15281
+eXk= 15282
+IGFic29yYg== 15283
+IHNvbGRpZXI= 15284
+ZGVycw== 15285
+IHRlc3RpbW9ueQ== 15286
+IGRlY2xpbmU= 15287
+nOuhnA== 15288
+Z2FnZQ== 15289
+IGluc3BpcmU= 15290
+bGFwcGluZw== 15291
+IHNwaW5uaW5n 15292
+IHNsYXZlcnk= 15293
+IGZhY2lhbA== 15294
+IHRyYWRpdGlvbnM= 15295
+w6FyaW9z 15296
+IEhvc3BpdGFs 15297
+IG5lc3Q= 15298
+IOuIhA== 15299
+IHRvaQ== 15300
+IGZlYXJz 15301
+7IWo 15302
+IE11aA== 15303
+IGdyYWR1YXRpb24= 15304
+IGltcGFjdGVk 15305
+IGF1bnQ= 15306
+IExldHM= 15307
+IGFsdW1pbnVt 15308
+IGRvbWluYW50 15309
+IERhdmlz 15310
+IE5hdnk= 15311
+IGNvbXB0 15312
+b3BsZXM= 15313
+IGVzdGF2YQ== 15314
+6KU= 15315
+IHNjYWw= 15316
+IHByZXNlcnZl 15317
+IE9wcA== 15318
+IHByYWN0aWNhbGx5 15319
+IG1hZ25pdHVkZQ== 15320
+IGZpdHRpbmc= 15321
+IGNvb3JkaW5hdGU= 15322
+IGZ1cm5pdHVyZQ== 15323
+IEZhbWls 15324
+IGV4cGxvc2lvbg== 15325
+IGRvY3VtZW50YXJ5 15326
+IFNjcmlwdA== 15327
+IHBvcnRyYXk= 15328
+bWF0 15329
+IHNjaGVkdWxlZA== 15330
+IGR5bmFtaWNz 15331
+cGh5 15332
+YWt5 15333
+IFVJ 15334
+Q2hl 15335
+IGNvbnRpbnVvdXNseQ== 15336
+IFByb3Y= 15337
+0YPQtw== 15338
+cmFo 15339
+IGdlcm5l 15340
+cHJvb2Y= 15341
+IHNlY3JldGFyeQ== 15342
+IFBhdHJlb24= 15343
+c2NyZWFt 15344
+IEtpZHM= 15345
+4buTaQ== 15346
+IGtn 15347
+IHVuY2VydGFpbnR5 15348
+INC60LDQttC0 15349
+IG1pdGln 15350
+IHJlYWRz 15351
+IFJ1 15352
+IHByaWVzdA== 15353
+INC90LXQtA== 15354
+IGxpbWl0YXRpb25z 15355
+IGZsb2F0 15356
+NjAw 15357
+IFRveQ== 15358
+IEppbW15 15359
+IG9mZmVuc2l2ZQ== 15360
+ZW5p 15361
+IFhp 15362
+IGV5ZWJy 15363
+IFR1cms= 15364
+IGFjY2lkZW50YWxseQ== 15365
+IG9obmU= 15366
+IFNhdWQ= 15367
+OTU= 15368
+IER1dGNo 15369
+0LDQvdGB 15370
+IFNlYXR0bGU= 15371
+IOuTsQ== 15372
+Y2hlY2s= 15373
+a8SZ 15374
+IGNvbnRyaWJ1dGlvbnM= 15375
+IGJlc2lkZQ== 15376
+IHF1aW5kaQ== 15377
+IGZsZXc= 15378
+2LDYpw== 15379
+IExP 15380
+IHdhaXN0 15381
+IEVW 15382
+IGhvbGlkYXlz 15383
+am9u 15384
+IG1pc3VuZGVy 15385
+0Y/QvQ== 15386
+IGJvdXQ= 15387
+IGRpbWlu 15388
+4bq9 15389
+w7Ns 15390
+IEdyYWNl 15391
+IGlucHV0cw== 15392
+IGRlbnk= 15393
+IGZvcm1pbmc= 15394
+IEJpbGQ= 15395
+IGFkZXF1 15396
+IGZvbGs= 15397
+IHJlamVjdGVk 15398
+c2VtYg== 15399
+IGZydXN0cmF0ZWQ= 15400
+b3Blbg== 15401
+IEJldHRlcg== 15402
+aWxvbg== 15403
+IHRvd2Vs 15404
+IGRpZmZlcmVudGlhbA== 15405
+IHNhY3JlZA== 15406
+IHNhaWw= 15407
+ZW50aW1lcw== 15408
+IGdlbnRsZW1hbg== 15409
+IGljb25pYw== 15410
+IGNvbXBhcmluZw== 15411
+IHNhZ3Q= 15412
+IHRleHRz 15413
+IGdyYW5kbWE= 15414
+IHJvbGxz 15415
+IGNvbnRlbnRz 15416
+0L7RgdGB 15417
+IHN1c3BlbnNpb24= 15418
+cm9pdA== 15419
+prw= 15420
+IGFzc2V6 15421
+IGRvcnQ= 15422
+IE1hdGg= 15423
+IFZpY3Rvcg== 15424
+IEphdmFTY3JpcHQ= 15425
+IGVuaGFu 15426
+xZk= 15427
+IEJ1c2g= 15428
+IHByb21vdGlvbg== 15429
+IGtpbg== 15430
+IG1vbnN0ZXJz 15431
+IENvbG9yYWRv 15432
+IM6y 15433
+7ZW07JqU 15434
+aWZmZXJlbnQ= 15435
+IG5ha2Vk 15436
+IHByb2Q= 15437
+ZXRpY3M= 15438
+IFdvbWFu 15439
+IHRyZWF0bWVudHM= 15440
+IGVzdG95 15441
+dsOp 15442
+IGxpZnRpbmc= 15443
+IHlhcHQ= 15444
+IFJvYmVy 15445
+IOy5nA== 15446
+IHN1YnN0aXR1dGU= 15447
+YWt1 15448
+cmlkZ2U= 15449
+IOqxsOs= 15450
+IHJlc3BvbmRlZA== 15451
+IGLDqQ== 15452
+IEVuZ2luZWVy 15453
+IHRyYW5zZmVycmVk 15454
+67I= 15455
+IGhhYmVy 15456
+b29w 15457
+IFdF 15458
+IHZlc3Q= 15459
+IGZvcnR5 15460
+IERT 15461
+IDIwMDQ= 15462
+IGNvYWNoaW5n 15463
+bm9t 15464
+IEJhYg== 15465
+IG5vc3Nh 15466
+IEpha2U= 15467
+IGd5 15468
+IGRlbGVn 15469
+IOyeoA== 15470
+INC60YDQsNGB 15471
+IHN0YW5kcG9pbnQ= 15472
+IGRpc2Fk 15473
+IGFydHdvcms= 15474
+QWQ= 15475
+aWxsbw== 15476
+IMSRxrDhu6Nj 15477
+IFByb20= 15478
+IExpYg== 15479
+IGNyaXRpY2lzbQ== 15480
+IGNvbnRhY3Rz 15481
+0YDQsNC8 15482
+IGFjaGlldmVtZW50 15483
+0JTQsA== 15484
+IGRpc3NvbA== 15485
+IFZlZ2Fz 15486
+IHN0cmVhbXM= 15487
+IEtlbnQ= 15488
+INi52YTZiQ== 15489
+IHJhZGl1cw== 15490
+IHN1Y2tz 15491
+IEFjaA== 15492
+IGZp 15493
+b3VzdA== 15494
+INC70Y7QtNC4 15495
+IHBhbGV0dGU= 15496
+IEhheg== 15497
+IEFudGhvbnk= 15498
+IHRlbWE= 15499
+IENvcw== 15500
+IHNhZmVy 15501
+zrHPgg== 15502
+IGNvbnRyYWQ= 15503
+IG1haW9y 15504
+IGluZmxhdGlvbg== 15505
+IFNpbHZlcg== 15506
+IGF0dGVuZGluZw== 15507
+7ZWc7YWM 15508
+YXJ0bw== 15509
+IGFwcGxhdWRpbmc= 15510
+IGNvbXB1dGluZw== 15511
+IEhhdA== 15512
+5rs= 15513
+a25vdw== 15514
+bWFrZXJz 15515
+IGNvbm9j 15516
+IGVkdWNhdGVk 15517
+IG1vZGlmaWVk 15518
+IGluY2x1c2lvbg== 15519
+bWVudGFs 15520
+npA= 15521
+aXNpYQ== 15522
+IM+Azr/PhQ== 15523
+IGF1bg== 15524
+IElyZWxhbmQ= 15525
+IGvDtg== 15526
+IGNvbXBsaWFuY2U= 15527
+IGluc3BpcmluZw== 15528
+0LjRgtC10LvRjNC90L4= 15529
+IGRpc3Bvcw== 15530
+7LCo 15531
+IHdpcA== 15532
+cmljYWw= 15533
+cmF3ZA== 15534
+IHRyZXM= 15535
+IG1vYmls 15536
+b2x1dGlvbnM= 15537
+Qk8= 15538
+IGJvdW5jZQ== 15539
+IGFzc3VtZWQ= 15540
+IE1lZGljYWw= 15541
+IGZpc2NhbA== 15542
+IG5nxrDhu51p 15543
+aXRpb25hbGx5 15544
+IHN0b2xlbg== 15545
+IEJN 15546
+IG1lY2hhbmlzbXM= 15547
+zrXOrw== 15548
+IHF1YWxpZmllZA== 15549
+IOyekOs= 15550
+dWdodGVycw== 15551
+IEhJVg== 15552
+IExvdHM= 15553
+IHNlcnZlcnM= 15554
+IGNhcnI= 15555
+IFRvZ2V0aGVy 15556
+IGF0dHJhY3RlZA== 15557
+IGty 15558
+dGh1cg== 15559
+aW5pbg== 15560
+IEhhbGY= 15561
+yJs= 15562
+IFBhcA== 15563
+IHJlbWluZGVk 15564
+QUxM 15565
+IGhlbG1ldA== 15566
+IGJvdHRsZXM= 15567
+IHByb2Zlc3NvcnM= 15568
+IHNlaW5l 15569
+xYLEhQ== 15570
+IOqxsOyVvA== 15571
+INei15w= 15572
+ZnVu 15573
+IEJpcmQ= 15574
+IGZpZ2h0ZXI= 15575
+IOuUsOs= 15576
+IFRvb2w= 15577
+IHRpbg== 15578
+aW5vaXM= 15579
+67aE 15580
+15nXnw== 15581
+IENBUg== 15582
+aXJzdHk= 15583
+IG91dGRvb3I= 15584
+IE5T 15585
+44WO 15586
+ZmZlbg== 15587
+IGx1ZA== 15588
+SGVsbG8= 15589
+IHJvbGxlcg== 15590
+aWVsZQ== 15591
+IFBvbGFuZA== 15592
+IGFwYQ== 15593
+ZXhw 15594
+IGNlcnRpZmljYXRl 15595
+IFRvd24= 15596
+0LDRjtGC0YHRjw== 15597
+aWxkZQ== 15598
+IGRldGVybWlu 15599
+UFI= 15600
+IGZyZWV6ZQ== 15601
+IG1haW5zdHJlYW0= 15602
+IG9iamVjdGl2ZXM= 15603
+Ymxv 15604
+IHRha2ll 15605
+IOuwlOuhnA== 15606
+ZWxldA== 15607
+IElW 15608
+IEZhc3Q= 15609
+IGRlcmU= 15610
+ZW1w 15611
+IERyYQ== 15612
+IOyeiOyXiA== 15613
+IGRpc2NyaW1pbmF0aW9u 15614
+IM61zq/Ovc6xzrk= 15615
+bmVjZXNz 15616
+5q4= 15617
+xLHEn8Sx 15618
+IHBvc3Rpbmc= 15619
+d2nFm2NpZQ== 15620
+IGx1Yg== 15621
+IG9saXZl 15622
+IHJpbQ== 15623
+IG1vZGVsaW5n 15624
+IGHDsW8= 15625
+IFBha2lzdGFu 15626
+IG92ZXJs 15627
+IGluZmxhbQ== 15628
+TkU= 15629
+7JeQ6rKM 15630
+IGF0dGVuZGVk 15631
+IGRlYWx0 15632
+IEFsdA== 15633
+IExpbmNvbG4= 15634
+IGF3YWtl 15635
+IGZpbHRlcnM= 15636
+IFdpdGhpbg== 15637
+Y3p5d2nFm2NpZQ== 15638
+IHPDuw== 15639
+IEpvaG5ueQ== 15640
+IGludGVncml0eQ== 15641
+IGlzb2xhdGlvbg== 15642
+IEVhc3k= 15643
+INC/0YDQuNC9 15644
+IEFsaWNl 15645
+IHNtaWxpbmc= 15646
+ZW5peA== 15647
+LC4uLg== 15648
+zrY= 15649
+IGJlZ3Vu 15650
+IGpld2Vs 15651
+IGNvbnZlbnRpb25hbA== 15652
+IHN0YXRpc3Q= 15653
+IGhhbmRlZA== 15654
+IGlycmU= 15655
+IHByb2hpYg== 15656
+IHNhdGVsbGl0ZQ== 15657
+IEluZHVzdA== 15658
+IHRyYWdlZA== 15659
+IHRyYXZh 15660
+IGlobQ== 15661
+IGNydWVs 15662
+IEFnb3Jh 15663
+IERvYw== 15664
+IHpvbmVz 15665
+IG1hbGw= 15666
+IHRyYXk= 15667
+15XXoA== 15668
+IGlycml0 15669
+IGthbnM= 15670
+IEJlYXQ= 15671
+dWRnZQ== 15672
+aWVsbGU= 15673
+IHRydXN0ZWQ= 15674
+IGJpa2Vz 15675
+INGD0L8= 15676
+IE1lbWJlcg== 15677
+d2ljaw== 15678
+IGNyZWF0b3Jz 15679
+IGhlcml0YWdl 15680
+aW5kaXN0aW5jdA== 15681
+IHJlc3Vy 15682
+ZW5uZW4= 15683
+Q29tZQ== 15684
+IGZpcmluZw== 15685
+IEJ1ZW5v 15686
+INCi0L4= 15687
+aWthbg== 15688
+ZXR0ZXM= 15689
+IGtlcw== 15690
+IHRyaXBz 15691
+IGRpdm9yY2U= 15692
+IEts 15693
+IGNvbnNvbA== 15694
+a2VlcA== 15695
+6riw6rCA 15696
+IFJlcG9ydA== 15697
+IGhvc3Rpbmc= 15698
+IGRpYW1vbmQ= 15699
+IGNvbXBsaWM= 15700
+IGhlbGljb3A= 15701
+IGRlcHVpcw== 15702
+ZHM= 15703
+IENoYW4= 15704
+0Y/Quw== 15705
+IHNjaXNzb3Jz 15706
+aWxhdGlvbg== 15707
+IHByb3BvcnRpb24= 15708
+RVJF 15709
+INmI2KfZhA== 15710
+aW50YQ== 15711
+IG11Y2hhcw== 15712
+dWF0aW9u 15713
+aXRpcw== 15714
+0Y/RiQ== 15715
+IG5paW4= 15716
+IGVtcGhhc2l6ZQ== 15717
+dWVsYQ== 15718
+IHByb2R1Y2Vycw== 15719
+IHJ6ZQ== 15720
+w6RuZGVy 15721
+RVRI 15722
+5ro= 15723
+IGNvbnN0aXR1 15724
+IHBlcmZvcm1hbmNlcw== 15725
+aXN0bGU= 15726
+Z292 15727
+IExpdGVy 15728
+IGluY29ycG9yYXRl 15729
+IGVkdWNhdGU= 15730
+IE5pbg== 15731
+7Kq9 15732
+2YfZhQ== 15733
+ZWxlcmF0aW9u 15734
+15XXkQ== 15735
+IHlhxZ8= 15736
+b3JvdXM= 15737
+IENhcw== 15738
+IGdyYW50cw== 15739
+64ql 15740
+YW1lbA== 15741
+IOq3uOugh+qyjA== 15742
+IEVzdGU= 15743
+0YXQvtC00LjRgg== 15744
+INC/0L7RgdC70LU= 15745
+IGdlbnQ= 15746
+IGZvY3VzZXM= 15747
+YWxpdGllcw== 15748
+IFJo 15749
+67O0 15750
+IERhbmNl 15751
+cnI= 15752
+IGFtZXI= 15753
+IHV0aWxpemU= 15754
+IGzDrQ== 15755
+IEFtb25n 15756
+IHByZWduYW5jeQ== 15757
+IGxvb3Bz 15758
+0LDQu9C+0YHRjA== 15759
+IE1vaA== 15760
+IGNhdGNoaW5n 15761
+IGdsb2I= 15762
+IGFqdWQ= 15763
+IFs/ 15764
+IEFuYWw= 15765
+bG9va2luZw== 15766
+IHN1cmZhY2Vz 15767
+IHByb2dyZXNzaXZl 15768
+IHZpcmFs 15769
+MDg= 15770
+zr4= 15771
+S0E= 15772
+IMW8eQ== 15773
+IHBpY2tz 15774
+YW5ub24= 15775
+IGJ1bGs= 15776
+IFJvc3M= 15777
+IGRlc2NyaWJpbmc= 15778
+IEdlbA== 15779
+IGxvY2FsbHk= 15780
+IGVuZGxlc3M= 15781
+IG1hc3NhZ2U= 15782
+IGNsZWFuZWQ= 15783
+IHRyYXZlbGVk 15784
+0LXQvdGL 15785
+IHNlbnRpbWVudA== 15786
+aWdtYQ== 15787
+IE5hcw== 15788
+IGNoZW1pY2Fscw== 15789
+IHJpZ2h0ZW91cw== 15790
+IE1hZ2lj 15791
+IHJlbGF0ZXM= 15792
+IHRydWNrcw== 15793
+IDE5NjA= 15794
+IGFwcGV0 15795
+IHNuYWNrcw== 15796
+IFN1bW1lcg== 15797
+IHnDvHo= 15798
+IHByaXM= 15799
+IE1leGljYW4= 15800
+IHRyYW5zcGFyZW4= 15801
+IG1pbm9yaXR5 15802
+IHZlcnRl 15803
+IGxhc3Nlbg== 15804
+NDY= 15805
+0LvQtdC6 15806
+w6lw 15807
+INGE0LjQu9GM 15808
+IGl5aQ== 15809
+IHNwYW4= 15810
+7ZWY7KeA 15811
+IGluZGljYXRlZA== 15812
+cXVhcg== 15813
+IHNjaG9sYXJzaGlw 15814
+IExHQlQ= 15815
+IGhpc3RvcmljYWxseQ== 15816
+w7PFgg== 15817
+IG1pbmlzdA== 15818
+IHBlbmV0 15819
+IFJhcA== 15820
+IGNvbnNlcnZhdGlvbg== 15821
+IEhvbmV5 15822
+IEJlaQ== 15823
+aWRlbA== 15824
+IHJlc3BvbnNpYmlsaXRpZXM= 15825
+IG1lc3N5 15826
+IEV4Y2VwdA== 15827
+T1JF 15828
+IGluaXRpYXRpdmVz 15829
+IGp1bmlvcg== 15830
+IGRlc2lnbmVycw== 15831
+IGV4cGxvcmF0aW9u 15832
+IHNwb25zb3I= 15833
+IG1vYmlsaXR5 15834
+IGludGVn 15835
+bGFuZG8= 15836
+IGJhcms= 15837
+IGluZGljYXRlcw== 15838
+4LY= 15839
+IGVtcGxveWVy 15840
+IGNvdXNpbg== 15841
+IGJvaWxpbmc= 15842
+IGNocm9t 15843
+IMOnYWw= 15844
+IHBlcnBldA== 15845
+IGNvbnRhaW5lZA== 15846
+IHBhcmtz 15847
+0Ks= 15848
+IEVuZ2luZWVyaW5n 15849
+UGxlYXNl 15850
+IFN0YXJ0aW5n 15851
+aGVybw== 15852
+IGxhd3llcnM= 15853
+IHpk 15854
+IGZyYW5jaGlzZQ== 15855
+cmFnZQ== 15856
+IGludHVpdA== 15857
+IEdM 15858
+cmVhY2g= 15859
+IEVsbGU= 15860
+IG5oxrA= 15861
+IE5vcmQ= 15862
+IGJlYW4= 15863
+MDc= 15864
+IHBsZWFzYW50 15865
+dmlyb24= 15866
+IGdyYWRpZW50 15867
+enVz 15868
+IEVN 15869
+IGVzc2F5 15870
+7JeQ7JqU 15871
+4bq/bg== 15872
+bnU= 15873
+4bur 15874
+IMOJcw== 15875
+IGRlbm9taW4= 15876
+IEdpcmxz 15877
+IHBlcnNvbm5lcw== 15878
+INin2YTYow== 15879
+YmlsZA== 15880
+IFN0YXQ= 15881
+IGNvbXBsaW1lbnQ= 15882
+IEthdGU= 15883
+IG9wdGltYWw= 15884
+IGhpZA== 15885
+2K/Zig== 15886
+IHF1aWNrZXI= 15887
+d2FsbA== 15888
+RW4= 15889
+SU5F 15890
+Pz8/ 15891
+7LK0 15892
+IEFjdGlvbg== 15893
+5Z8= 15894
+IHBlbmFsdHk= 15895
+IEtheg== 15896
+Jz8= 15897
+IGNyaWVk 15898
+IGNhbnZhcw== 15899
+ZnRl 15900
+IGV4Y2x1ZA== 15901
+uOuhnA== 15902
+IGVtcGhhc2lz 15903
+IGVuenk= 15904
+IEhvdQ== 15905
+IG92ZXJzZWFz 15906
+w61hbW9z 15907
+w7ZnbGljaA== 15908
+IGhlYWRwaG9uZXM= 15909
+Y24= 15910
+IEFnZQ== 15911
+IGFrYW4= 15912
+IGNoYXJhY3RlcmlzdGlj 15913
+7ZWY66m0 15914
+Z2V0cw== 15915
+IOu2iA== 15916
+IHJpdmFs 15917
+IGJvcmRlcnM= 15918
+ZW1lbnRl 15919
+ZW3DoXM= 15920
+IHlvbA== 15921
+IGNvbXBl 15922
+ZW5kZXJz 15923
+xLFuZGFu 15924
+IG3DtmdsaWNo 15925
+IGJ1YmJsZXM= 15926
+bmF0dXJhbA== 15927
+IGFybWVk 15928
+IGVsYWJvcg== 15929
+IOydtOuyiA== 15930
+IHdhc2hlZA== 15931
+zr/Phc68zrU= 15932
+IGZsYXZvcnM= 15933
+IGV4aXN0ZQ== 15934
+IHByZXN0 15935
+IFRoZW1h 15936
+0L7Qv9GA0L7RgQ== 15937
+ZXJvbg== 15938
+VUU= 15939
+ZXJp 15940
+IGNvbmNlcg== 15941
+IGFpeMOy 15942
+IHByb3RlY3RpdmU= 15943
+INC30L3QsNGO 15944
+IOuCoA== 15945
+IElJSQ== 15946
+IG1lZXI= 15947
+IFNob3A= 15948
+bGxp 15949
+IE9yZGVy 15950
+IE1Z 15951
+IEdob3N0 15952
+YWRlbA== 15953
+IHN0b2xl 15954
+IHJlbGVhc2luZw== 15955
+IENvbW1lbnQ= 15956
+IHRyYWlucw== 15957
+66qF 15958
+IHdpc3Nlbg== 15959
+ZW5zZWQ= 15960
+IGRlc2NlbmQ= 15961
+IGZpZXI= 15962
+IHJhZGk= 15963
+IHBlcnN1 15964
+56I= 15965
+INC80L0= 15966
+IERlc3Q= 15967
+IHdvcnJpZXM= 15968
+aXRldA== 15969
+YmFz 15970
+IHN0YWI= 15971
+bmFtZQ== 15972
+b3JpYw== 15973
+IENsb3Nl 15974
+IGFsdW1uaQ== 15975
+IFNlbGY= 15976
+ZmZl 15977
+aXRhdGluZw== 15978
+YXRoZXJpbmU= 15979
+IFJpZ2h0cw== 15980
+IGVsbG9z 15981
+IHdhcnJhbnQ= 15982
+IG5lcnZl 15983
+IHZlZ2V0YWJsZQ== 15984
+IFRlaWw= 15985
+IOqwmeydtA== 15986
+Ulk= 15987
+IHN1c3RhaW5hYmlsaXR5 15988
+IHN0ZWh0 15989
+IGJyaWQ= 15990
+YWRhxZ8= 15991
+IHR2 15992
+IGR1cmF0aW9u 15993
+IHBlc3NvYQ== 15994
+IG1ldHJpY3M= 15995
+IGFkYW0= 15996
+Y2Fz 15997
+0LDRgNC4 15998
+IGV2aWRlbnQ= 15999
+IGRpc3BsYXllZA== 16000
+2KfYpg== 16001
+IHJlY2s= 16002
+IEJ1ZGRoYQ== 16003
+IGRlbGU= 16004
+IERpZWdv 16005
+b3NwaA== 16006
+IGJsYQ== 16007
+IE1paw== 16008
+dWxhdG9y 16009
+IDIwMDE= 16010
+IHByb21vdGluZw== 16011
+eWNo 16012
+IEVY 16013
+IGxhc3RseQ== 16014
+IG91dGxpbmU= 16015
+IHNwaXJpdHM= 16016
+IHZldXg= 16017
+IHN1YnRyYWN0 16018
+IMWfaW1kaQ== 16019
+IHBpbnM= 16020
+IGJ1cmdlcg== 16021
+IG1vbHRv 16022
+IGhhYsOtYQ== 16023
+IOuwmA== 16024
+aWd1 16025
+ZXJzdA== 16026
+IG5lbg== 16027
+IGJhY29u 16028
+aXRpb3Vz 16029
+IGNhcnJpZXM= 16030
+IHByb21pc2Vz 16031
+bmRl 16032
+IExlZnQ= 16033
+IExpbQ== 16034
+5qM= 16035
+IDQ0 16036
+IGNhcmVlcnM= 16037
+IOyjvOs= 16038
+IHNwZWVkcw== 16039
+cXXDqQ== 16040
+bWFk 16041
+bWFya2V0 16042
+aXNtZQ== 16043
+IDIwMDM= 16044
+IHJlY2Vzcw== 16045
+IEpVRA== 16046
+IHJhY2lzdA== 16047
+IFNjaGw= 16048
+IHBhcmxlcg== 16049
+IG90cm9z 16050
+aXNoZXM= 16051
+IGNvbnZlcnRlZA== 16052
+YWFhYQ== 16053
+0LDQvdC40Lg= 16054
+IEFyaw== 16055
+IENoYW5jZQ== 16056
+IGVsZW1lbnRhcnk= 16057
+zrXOvQ== 16058
+aW5rcw== 16059
+SW50ZXJ2aWV3ZXI= 16060
+IGZyZWVseQ== 16061
+YWxhaA== 16062
+IOuLpOuluA== 16063
+IHJlcXVlc3RlZA== 16064
+IHRvcnF1ZQ== 16065
+bm/Fm2Np 16066
+b3VyZWQ= 16067
+IFN0YWZm 16068
+IHN0YWlu 16069
+IEFsYW4= 16070
+IHZlcmU= 16071
+IFdpbnRlcg== 16072
+IGRlZmVjdA== 16073
+aWVkeQ== 16074
+IGJlYXRz 16075
+IGjDoQ== 16076
+dW1u 16077
+b29ucw== 16078
+aXR1ZGVz 16079
+IHNlaXQ= 16080
+b2x5 16081
+IHJlc2Vydg== 16082
+IGV4dHI= 16083
+IHBoeXNpY2lhbg== 16084
+dmlzb3I= 16085
+IGhhbmRmdWw= 16086
+IE5hdGlvbnM= 16087
+IOyii+ydgA== 16088
+dWNjZXNz 16089
+IHVwc3RhaXJz 16090
+IFNxdWFyZQ== 16091
+IGhlaW4= 16092
+IFNlYXNvbg== 16093
+b2xpcw== 16094
+IHByaW5jZQ== 16095
+IGRlZmVuc2l2ZQ== 16096
+570= 16097
+INC80LXRgdGC 16098
+0ZbQuQ== 16099
+INin2YY= 16100
+dW1ibGU= 16101
+6rmM7JqU 16102
+IGFzc2Fzcw== 16103
+IGNpcmN1bGFy 16104
+IHF1YWxpdGllcw== 16105
+IGhtbQ== 16106
+IGJsb3du 16107
+IExpeg== 16108
+IEt1cg== 16109
+IFNB 16110
+IGZpbmRpbmdz 16111
+IGNvbG91cnM= 16112
+IGRlbGxl 16113
+IElS 16114
+IEF0aA== 16115
+IER1Yg== 16116
+IE94 16117
+INiu 16118
+IHBvY2tldHM= 16119
+IGdyaWxs 16120
+IHN3aXRjaGluZw== 16121
+IHByZWZlcnJlZA== 16122
+IFdhbGVz 16123
+IGV4ZW1wbG8= 16124
+IGNob3BwZWQ= 16125
+IHZhY2NpbmF0aW9u 16126
+IG5ldXJv 16127
+IHNwZWNpZnk= 16128
+aXZvcw== 16129
+IHNlcsOh 16130
+IHppZQ== 16131
+IOCurg== 16132
+IHJlc3VsdGluZw== 16133
+IFVnaA== 16134
+IG1lc3NlZA== 16135
+Q0Q= 16136
+IHBhYXI= 16137
+IGNvbWVy 16138
+IGNvdWNo 16139
+IEZlc3RpdmFs 16140
+IDQ5 16141
+dm91cw== 16142
+emVucw== 16143
+IEtlbm5lZHk= 16144
+IFRz 16145
+IOuztOyX 16146
+IGRlbW9uc3RyYXRpb24= 16147
+IHVudG8= 16148
+IGZydXN0cmF0aW5n 16149
+IGxhYm9yYXRvcnk= 16150
+IGVneQ== 16151
+IGJlYXV0aWZ1bGx5 16152
+IOyerOs= 16153
+IGFsZ3U= 16154
+IMO2eWxl 16155
+IFBI 16156
+IGZvcnR1bmU= 16157
+IGNsZWFuZXI= 16158
+IFJvYmlu 16159
+IHNhdXM= 16160
+IEdlbGQ= 16161
+IGthdA== 16162
+b2Jz 16163
+IG9sdXI= 16164
+IG1hdHQ= 16165
+IHF1ZXN0YQ== 16166
+IHN1Z2dlc3Rpb24= 16167
+ZW5jZXI= 16168
+0L7RgdGC 16169
+IHJhZGFy 16170
+IOyeoQ== 16171
+aXNoYQ== 16172
+4K6o 16173
+amVz 16174
+IHZlZWw= 16175
+7IKw 16176
+IGF1dGhvcnM= 16177
+44CO 16178
+cGxhbg== 16179
+IGNvbGxhYm9yYXRpdmU= 16180
+IGluc3RpbmN0 16181
+IGZhcm1pbmc= 16182
+YXVnZQ== 16183
+RWR1 16184
+IG1lbWJlcnNoaXA= 16185
+IHNpbXVsdGFuZW91c2x5 16186
+IGJha2U= 16187
+IGvDpA== 16188
+IGxlY3R1cmVz 16189
+0YfQtdGB 16190
+IHByZW5kcmU= 16191
+IGNvbGxhcHM= 16192
+IFNheWE= 16193
+IEZ1dA== 16194
+IHlvZw== 16195
+IFJhdGhlcg== 16196
+2LHZig== 16197
+IGNhbXBz 16198
+0L7Qu9C+0LQ= 16199
+IHNpbXVsYXRpb24= 16200
+IE1haw== 16201
+TGF1Z2hz 16202
+IGdyZXk= 16203
+IHNlbnRlbmNlcw== 16204
+eWVu 16205
+IFVubGVzcw== 16206
+SmU= 16207
+IFNhdGFu 16208
+INGC0LDQutC20LU= 16209
+IE5B 16210
+IGJyb24= 16211
+ID9d 16212
+IHNvdWxz 16213
+IGxpZ2h0bmluZw== 16214
+IGltYWdpbmVk 16215
+IGN6eWxp 16216
+cHNpbG9u 16217
+ZXR0YQ== 16218
+IGJlbGlldmluZw== 16219
+IHN0cm9uZ2VzdA== 16220
+IENPTg== 16221
+IHF1ZWxxdWVz 16222
+IGltbWlncmFudHM= 16223
+IHdhbGxldA== 16224
+IEplcnNleQ== 16225
+IGltcGxpY2F0aW9ucw== 16226
+IGZvcmI= 16227
+44CP 16228
+IHVuYmVsaWV2YWJsZQ== 16229
+2KfYoQ== 16230
+IG9wZXJhdGlvbmFs 16231
+w7xz 16232
+IEdN 16233
+IOq3uOufsOuNsA== 16234
+IGdyYWNpYXM= 16235
+IGVudGVuZA== 16236
+IFJlZ2FyZA== 16237
+cm9i 16238
+INGC0LXRhQ== 16239
+6I8= 16240
+IFJldm9sdXRpb24= 16241
+IHdhYXI= 16242
+IEJpeg== 16243
+dGhlbGVzcw== 16244
+IHNwb25zb3JlZA== 16245
+cXVpZXI= 16246
+IOydvOs= 16247
+IHRlaw== 16248
+IOuQoA== 16249
+aWdrZWl0 16250
+IEx1Y2s= 16251
+IENlcnRhaW5seQ== 16252
+IHRvbGw= 16253
+INC90LjRh9C10LPQvg== 16254
+IE1vbmV5 16255
+INGB0YLQvtGA 16256
+IERvdWJsZQ== 16257
+IFdvbGY= 16258
+IGNodW5r 16259
+zqzOvQ== 16260
+aXTDqXM= 16261
+b25pbmc= 16262
+TWFy 16263
+IGdyYW5kZXM= 16264
+IGNvbGxlY3Rpb25z 16265
+IEV1cm9wYQ== 16266
+INCw0YA= 16267
+IOKAi+KAi+KAiw== 16268
+IOq3uOufrOuptA== 16269
+INC+0LHRig== 16270
+IOyLnOqwhA== 16271
+IEN1c3RvbQ== 16272
+IOyymA== 16273
+0ZbQu9GM 16274
+IGluZGl2aWR1YWxseQ== 16275
+7Zc= 16276
+IGRvemVu 16277
+IG93ZQ== 16278
+IFZpY3Rvcmlh 16279
+IGJlZXQ= 16280
+dXJi 16281
+IGFuYWxvZw== 16282
+acOnw6Nv 16283
+gpw= 16284
+c29ldmVy 16285
+IG1vZG8= 16286
+IHN1YnNjcmliZWQ= 16287
+7J6s 16288
+IGVudGl0aWVz 16289
+IGNsb3NldA== 16290
+IHJlc3BvbmRpbmc= 16291
+IHByaW50ZXI= 16292
+IFN0ZXBoYW4= 16293
+IGJ5xYI= 16294
+IERvbQ== 16295
+IEZlcm4= 16296
+IFBpZXI= 16297
+IHdpxJlj 16298
+IGhlbmNl 16299
+IG1vZHVsZXM= 16300
+IOuUsQ== 16301
+IERhbm55 16302
+INGB0LXQsdC1 16303
+IHZhZA== 16304
+IOyXhA== 16305
+IHNvdXM= 16306
+IHNwaGVyZQ== 16307
+Qlk= 16308
+IFBlZA== 16309
+aWduZWQ= 16310
+IHdoZWF0 16311
+IHVuZGVycw== 16312
+IGV2b2x2ZQ== 16313
+IGRlY2xhcg== 16314
+IGxpZ2h0bHk= 16315
+IGlkZW50aWZ5aW5n 16316
+IGxlZ2VuZGFyeQ== 16317
+IGdlbnVpbmU= 16318
+IGdyaW5k 16319
+IFVuZQ== 16320
+Z2ViZW4= 16321
+IGJpY3k= 16322
+IGp1bXBz 16323
+IHByb3ZpbmNl 16324
+emnEmQ== 16325
+INeQ16DXmQ== 16326
+IGhvYw== 16327
+INCx0Ls= 16328
+IEdyYWQ= 16329
+IHJldmVuZ2U= 16330
+INin2YTYqg== 16331
+b29o 16332
+0LDRhtC40Lg= 16333
+IGVsZWN0cm8= 16334
+IOuQkA== 16335
+IGZhbHM= 16336
+cmllbA== 16337
+b2tlcg== 16338
+IEV4Y2VsbGVudA== 16339
+IE1vcmdhbg== 16340
+IGJyaWNr 16341
+IHN1YnN0YW50aWFs 16342
+IHBvbGx1dGlvbg== 16343
+IFTDvHI= 16344
+IEV2ZXQ= 16345
+IGx1bmc= 16346
+15nXqQ== 16347
+b21tZXM= 16348
+IHJlYWxpemluZw== 16349
+IGh1bWJsZQ== 16350
+IExvY2s= 16351
+IGJvZA== 16352
+IOyWuA== 16353
+IHBlZXJz 16354
+dXp6 16355
+IGVtYmVkZGVk 16356
+IGNsYXJv 16357
+IGFnZ3JlZw== 16358
+IGVtcGxveWVycw== 16359
+IFJhag== 16360
+IFlp 16361
+IGpldQ== 16362
+YXRlcnM= 16363
+IHN0cmlrZXM= 16364
+bm9z 16365
+YXV0cmVz 16366
+ZHI= 16367
+b3BoZXI= 16368
+IEFwcGFyZW50bHk= 16369
+7ZiE 16370
+IGluZmFudA== 16371
+2KfYqA== 16372
+0YLRiw== 16373
+7Zs= 16374
+2q8= 16375
+IHJlZGVz 16376
+YWNhxJ/EsW0= 16377
+IERBVklE 16378
+IENoaWNrZW4= 16379
+IHBlcnNwZWN0aXZlcw== 16380
+IHZpZXdlcg== 16381
+IHNoYXI= 16382
+INC/0YDQvtC40Lc= 16383
+bGlndA== 16384
+ZXJvcw== 16385
+aXRhYmxl 16386
+0LjQu9C+0YHRjA== 16387
+IGRpZsOt 16388
+tOuNsA== 16389
+IHJldGlyZWQ= 16390
+IHRoYXRz 16391
+emVuaWU= 16392
+YmVpdGVu 16393
+IG15Y2tldA== 16394
+IFJhYg== 16395
+IGluZmxhbW0= 16396
+7LCu 16397
+IGR1bQ== 16398
+IGRhZGR5 16399
+IGltbWVycw== 16400
+IHBsYXlsaXN0 16401
+4K+G 16402
+IHRyYXVt 16403
+IHJlZnVzZQ== 16404
+c3RlcA== 16405
+4K6a 16406
+Y3Vw 16407
+IHBvcHM= 16408
+cmltaW4= 16409
+YXnEsW0= 16410
+IGFsZA== 16411
+IHVubmVjZXNz 16412
+IGRhaA== 16413
+IElyaXNo 16414
+IGNvbXBy 16415
+bGHFnw== 16416
+VFA= 16417
+IHRyYW5zbGF0ZWQ= 16418
+U2M= 16419
+Y2XEn2lt 16420
+tJA= 16421
+IGRyZWk= 16422
+INC70Y7QtNC10Lk= 16423
+IHF1aWVybw== 16424
+IGhlbGU= 16425
+emxpY2g= 16426
+IGFwcGxlcw== 16427
+IGRpc3RyaWN0cw== 16428
+IGNyZWRpdHM= 16429
+IGFzcA== 16430
+IOuLqA== 16431
+b3JhbA== 16432
+IHN0ZXBwaW5n 16433
+IFZh 16434
+IGdhaW5z 16435
+NjU= 16436
+IG51ZXN0cmE= 16437
+ZWRheQ== 16438
+YXNzYWRvcg== 16439
+IExpbmQ= 16440
+IGNyb3Bz 16441
+Y2llbmRv 16442
+aWd1ZQ== 16443
+IGJhbmE= 16444
+QW0= 16445
+IHBlbnQ= 16446
+IGFkZGljdGlvbg== 16447
+IHBhY2thZ2luZw== 16448
+w6Rk 16449
+qqg= 16450
+IHBlcnF1w6g= 16451
+IGNhbXBhaWducw== 16452
+IHN0ZWVw 16453
+IG5ldWU= 16454
+IGVtYmFycmFzc2Vk 16455
+IGRpc3RpbmN0aW9u 16456
+aXR6ZXI= 16457
+IHJlZ2lzdHJhdGlvbg== 16458
+IGxsYW0= 16459
+IEFsbWlnaHR5 16460
+bGllc3Q= 16461
+IHV6 16462
+bmFr 16463
+57o= 16464
+IHRlcmF6 16465
+aWFtZW50ZQ== 16466
+IHRyYW5zYWN0aW9ucw== 16467
+IGPDtHQ= 16468
+IHN3aXRjaGVk 16469
+IGNvbWJv 16470
+IHByYXllcnM= 16471
+IGludGVybnNoaXA= 16472
+IGFkZHJlc3Nlcw== 16473
+IGNoYXJpdHk= 16474
+IFdPTw== 16475
+IGJhaXQ= 16476
+IO+/vQ== 16477
+IGZpY2E= 16478
+IFR5bGVy 16479
+YXJ1 16480
+IGF0b21z 16481
+IExldmVs 16482
+INC/0L7RgtC+0Lw= 16483
+IGZhbWU= 16484
+dWxr 16485
+IHRlYWNoZXM= 16486
+IHJlYnVpbGQ= 16487
+0LXQtNGM 16488
+IEluZG9uZXNpYQ== 16489
+dXNoaQ== 16490
+IFNob3J0 16491
+IGVuc3VyaW5n 16492
+ZnM= 16493
+ZWxl 16494
+IG1hcmdpbmFs 16495
+IGNvbmNsdWRl 16496
+YW10 16497
+IHZlcmlmeQ== 16498
+IE1jRG9uYWxk 16499
+IHNrYWw= 16500
+IHJlY29uc3Q= 16501
+IE1hbm4= 16502
+IGJhc2VtZW50 16503
+IHRyYW5zZm9ybWVk 16504
+IG9jY2FzaW9uYWxseQ== 16505
+em9uZQ== 16506
+IERhbnM= 16507
+INC60LDQutC+0Lk= 16508
+IGRpYWdub3NlZA== 16509
+IM+EzrE= 16510
+IGNvbW1hbmRz 16511
+IHByZXNpZGVudGlhbA== 16512
+IGFiYg== 16513
+IGJyYWNrZXQ= 16514
+IExlbQ== 16515
+w6VuZw== 16516
+IGZhdm9yaXRlcw== 16517
+IHJldm9s 16518
+IO2KuQ== 16519
+IGhhcmFzcw== 16520
+6YU= 16521
+IGNsZWFucw== 16522
+c3TDpG5k 16523
+IGtub2NrZWQ= 16524
+IHBlb3BsZXM= 16525
+IG11c2ljaWFucw== 16526
+IG11dHVhbA== 16527
+IENvbGQ= 16528
+ODg= 16529
+emVq 16530
+YXRpZQ== 16531
+IEhvbm9y 16532
+IG9ic2Vzc2Vk 16533
+IE1VU0lD 16534
+IEJyZWFr 16535
+w7puZw== 16536
+IG1vZGlmeQ== 16537
+IHPDtnlsZQ== 16538
+INee15Q= 16539
+IE9ubGluZQ== 16540
+Zm8= 16541
+IE1pbGxlcg== 16542
+IGxpa2luZw== 16543
+IGluaGFi 16544
+IGdyYXRpdHVkZQ== 16545
+IEpvdXJuYWw= 16546
+YXJuZXNz 16547
+Sm9obg== 16548
+IEdpdA== 16549
+IHNpbmNlcmU= 16550
+IFNjaQ== 16551
+IEVsaQ== 16552
+IHN5bWJvbHM= 16553
+IG1hbnVhbGx5 16554
+zrXPgg== 16555
+INCy0ZbQtA== 16556
+IEZhdA== 16557
+IGxhYmVscw== 16558
+IHNvcGhpc3RpY2F0ZWQ= 16559
+dW1wcw== 16560
+IHJlbGVhc2Vz 16561
+IDQ3 16562
+IE9N 16563
+6rCA6w== 16564
+IEJpZW4= 16565
+IFJlZg== 16566
+IFN0YQ== 16567
+IEVnZw== 16568
+IGluZGljYXRvcg== 16569
+cHNvbg== 16570
+IG5hc8SxbA== 16571
+UmlnaHQ= 16572
+IGNvbnZleQ== 16573
+IGtub3Q= 16574
+IGNvbm5lY3Rz 16575
+dWxhcw== 16576
+IHByZWNlZA== 16577
+IGluZXF1YWxpdHk= 16578
+YW1pZW50bw== 16579
+IHJlcGx5 16580
+T1k= 16581
+IGRpc21pc3M= 16582
+IOuQnA== 16583
+INGF0L7RgNC+0YjQvg== 16584
+IG3DqWQ= 16585
+IHJhbmRvbWx5 16586
+IE9udA== 16587
+dWFyZA== 16588
+IHB1bGxz 16589
+INGC0LXQv9C10YDRjA== 16590
+IE5lZWQ= 16591
+IFNvZnQ= 16592
+IHN0cmVuZ3Rocw== 16593
+IGdvZWQ= 16594
+dW1lbg== 16595
+IO2OuA== 16596
+INC00L7QsQ== 16597
+IGNsYXJpdHk= 16598
+IEFp 16599
+IGJhbGxvb24= 16600
+IFBhbmQ= 16601
+IOyVhOuL 16602
+IHNoaW55 16603
+IHNtYWxsZXN0 16604
+b25pYQ== 16605
+aGlsbA== 16606
+b3Rpbmc= 16607
+IGVpbmc= 16608
+IG1lcmVseQ== 16609
+IHNldXM= 16610
+INC90LXQvw== 16611
+IO2GtQ== 16612
+IGd1aWRlcw== 16613
+IHNwZWNpYWxpc3Q= 16614
+IHN0ZWFr 16615
+IG1pZ3JhdGlvbg== 16616
+cXVlbGU= 16617
+IHJ1aW5lZA== 16618
+IHB1cHA= 16619
+IGtlbmQ= 16620
+YW5nYW4= 16621
+IHBhbG0= 16622
+IHVuZmFpcg== 16623
+IHpt 16624
+IERW 16625
+Y2hlc3Rlcg== 16626
+0LjRjg== 16627
+IG9vaA== 16628
+ZXJn 16629
+QVRI 16630
+sKk= 16631
+cmlzb24= 16632
+IGludm9sdmluZw== 16633
+IHBhcnRseQ== 16634
+YW7Dp2Fpcw== 16635
+IHZvdw== 16636
+IHByb21pbmVudA== 16637
+IGNyeXN0 16638
+aWJh 16639
+IGRlc2VydmVz 16640
+IG92ZXJ0 16641
+IHNlbnNpdA== 16642
+IFdoZQ== 16643
+IHRpZ2h0ZW4= 16644
+IGludGltaWQ= 16645
+IGFsaW1lbnQ= 16646
+d2lsbA== 16647
+IHN0cmVuZ3RoZW4= 16648
+IFRhbg== 16649
+b25p 16650
+IE11bg== 16651
+IHByb3Bo 16652
+IHJlaGVhcnM= 16653
+IEtsZQ== 16654
+IHZlY2Vz 16655
+IHdvbmRlcmVk 16656
+b2tp 16657
+IHNlbnNlcw== 16658
+tOyL 16659
+xrDhu5s= 16660
+IMiZaQ== 16661
+IG11Y2hvcw== 16662
+IHdhdGNoZXM= 16663
+b3J0dW5hdGU= 16664
+IEp1YW4= 16665
+7J6W7JWE 16666
+0YDQtQ== 16667
+ZWk= 16668
+aW9uZW4= 16669
+IGV4cGVyaW1lbnRhbA== 16670
+IGRhdWdodGVycw== 16671
+4Lib 16672
+IG1lbnRhbGx5 16673
+YmVjY2E= 16674
+YXdhcmU= 16675
+7ISd 16676
+IHdoYXRzb2V2ZXI= 16677
+IGVuYWJsZXM= 16678
+IExvdw== 16679
+b2lk 16680
+4LiK 16681
+w7Nk 16682
+2Lo= 16683
+IGNvbnN0cnVjdGVk 16684
+IExhZGllcw== 16685
+IGFjY3VzZWQ= 16686
+INCw0L0= 16687
+RGFu 16688
+IHNwYXdu 16689
+IGNvbnRhaW5lcnM= 16690
+IGFydGlzdGlj 16691
+xLFw 16692
+IGRpc2Ns 16693
+IGF1dHJlcw== 16694
+aW5hcw== 16695
+IE5hdGlvbg== 16696
+IG5hZw== 16697
+YmVhbg== 16698
+d2hl 16699
+nOuPhA== 16700
+IFNlb3Vs 16701
+IO2PrA== 16702
+IE5pY2g= 16703
+IGNvbXBsZW1lbnQ= 16704
+IGludGVydmVu 16705
+IE1vZGVs 16706
+IE9yYW5nZQ== 16707
+bmFtb24= 16708
+IGNhbGN1bGF0aW9u 16709
+c2Vl 16710
+IHVzdGVkZXM= 16711
+IGxlYg== 16712
+IGRvY3Q= 16713
+0ZbQvQ== 16714
+IGZvc3Rlcg== 16715
+IGVsYXN0aWM= 16716
+IEFoaA== 16717
+IGFjZQ== 16718
+IFBpbms= 16719
+IEplZw== 16720
+IGRlZXI= 16721
+c2lz 16722
+IGpha28= 16723
+IEVtbWE= 16724
+0YHRgtCy0LXQvdC90L4= 16725
+IHBvcnRyYWl0 16726
+IG1ha2Vy 16727
+IGF1bWVudA== 16728
+0YDQvtCx 16729
+IGFpcnBsYW5l 16730
+IHRyYW5zcGFyZW5jeQ== 16731
+IGFkanVzdG1lbnQ= 16732
+IENEQw== 16733
+w6dvbg== 16734
+IHVwbG9hZGVk 16735
+INC00LXQudGB0YLQsg== 16736
+INCz0L7RgtC+0LI= 16737
+IGl0ZXI= 16738
+IGN1cnNl 16739
+w7Ru 16740
+bWVyY2U= 16741
+YXJhbg== 16742
+IGxlYWs= 16743
+IGFic2VuY2U= 16744
+0YHQutC40Lk= 16745
+IHJlYWRlcnM= 16746
+YWxlcg== 16747
+IGJlbmVhdGg= 16748
+YW5nbw== 16749
+aGV0aWM= 16750
+IGZpbm5z 16751
+IHBvb3A= 16752
+IGR1cGxpYw== 16753
+SGk= 16754
+aWdz 16755
+b2xvZ2ljYWxseQ== 16756
+b3Bw 16757
+IGRpemVy 16758
+IEFsbGVu 16759
+IGdsaQ== 16760
+IGFjY2VsZXJhdGlvbg== 16761
+IHZpdGFtaW4= 16762
+dsOk 16763
+IEFjY2Vzcw== 16764
+4K6Z 16765
+csOhcw== 16766
+IGFwcHJlY2lhdGVk 16767
+IG5haA== 16768
+IHBvc3Rlcg== 16769
+IHRhbGU= 16770
+IGhpZ2hsaWdodGVk 16771
+xbxlbGk= 16772
+IGJsb2NrY2hhaW4= 16773
+IG1pY3Jvdw== 16774
+IGNpbmVtYQ== 16775
+IENoYW5n 16776
+IFNlYXJjaA== 16777
+dXN0ZXJz 16778
+IFplcm8= 16779
+IERpdmlzaW9u 16780
+0YDQsNGB 16781
+IHNjYXJl 16782
+IGplbGx5 16783
+IEFkbWluaXN0cmF0aW9u 16784
+U08= 16785
+IGxpbmVk 16786
+IOqwhA== 16787
+IGdlYmVu 16788
+IHNvZGE= 16789
+IHdpbm5lcnM= 16790
+s7w= 16791
+2ZI= 16792
+IEFtYg== 16793
+5ZQ= 16794
+IHBlZw== 16795
+NDM= 16796
+IHJhdXM= 16797
+IHJld2FyZHM= 16798
+IGluY2x1cw== 16799
+IGhpZ2h3YXk= 16800
+IGhhaA== 16801
+IG11bHRpcGxpZWQ= 16802
+IHPhur0= 16803
+IGRpc2NpcGxlcw== 16804
+IG5pbmc= 16805
+IGRyZXNzaW5n 16806
+IGF0dHJpYnV0ZXM= 16807
+IE1vc2M= 16808
+IEdyZWVjZQ== 16809
+IHNlaw== 16810
+IExlYXJu 16811
+IGp1cw== 16812
+cmVuZHJl 16813
+IHBlcnNvbm5l 16814
+cGxldGU= 16815
+IHBsYWNpbmc= 16816
+IGx1ZWdv 16817
+aWxsYW5jZQ== 16818
+INC+0LHRiQ== 16819
+IHByb3Zpc2lvbg== 16820
+IGxpb24= 16821
+dHJh 16822
+Ym9hcmRz 16823
+IGJlaGF2aW91cg== 16824
+aGV5 16825
+IHN1YnNjcmlwdGlvbg== 16826
+IHByb3RhZ29u 16827
+IHZhcmE= 16828
+IMWfdQ== 16829
+IGhhaGE= 16830
+IHRlYXNwb29u 16831
+5p8= 16832
+YXZvaXI= 16833
+IGNyeXB0bw== 16834
+INGB0YLQsNGA 16835
+IFN0b3Jl 16836
+YWJz 16837
+IFN0dWRlbnRz 16838
+IGxhdW5k 16839
+aW50bw== 16840
+IGFwcHJvYWNoZWQ= 16841
+sJw= 16842
+0YPRjtGJ 16843
+IExhYm9y 16844
+b3Rlcw== 16845
+aWF0cmlj 16846
+IGdyb8Of 16847
+dXRpdmU= 16848
+INC40LQ= 16849
+IEdpYg== 16850
+IHBsYWNlbWVudA== 16851
+IGRpZsOtY2ls 16852
+IGZyb2c= 16853
+INCy0YHQtdGF 16854
+IEpy 16855
+YXplZA== 16856
+0YPRiQ== 16857
+IOq8 16858
+ZnJhbWU= 16859
+0LDQtdGI0Yw= 16860
+IGxvY2tkb3du 16861
+IG1lZGk= 16862
+INeU157X 16863
+0LXQvdC40Lk= 16864
+ZW1hbGU= 16865
+7KKF 16866
+YXRlcmFs 16867
+IGRpc3RhbnQ= 16868
+IGJlYXJz 16869
+IGpvdXJuYWxpc3Q= 16870
+IE1hcnNoYWxs 16871
+IElobmVu 16872
+dWV0b290aA== 16873
+YmFn 16874
+IMSRw6M= 16875
+IEhpZ2huZXNz 16876
+IOywjQ== 16877
+0LjQutCw 16878
+IFd1 16879
+IEZyYW4= 16880
+IHBlbmc= 16881
+IGZvbg== 16882
+IGh5cG90aGVzaXM= 16883
+INGA0YM= 16884
+IGx5 16885
+15o= 16886
+7JuU 16887
+IFJhZGlv 16888
+4Lie 16889
+RGF2 16890
+IGVtYmFycmFzc2luZw== 16891
+IOyeiOyWtA== 16892
+IGNhc3Rpbmc= 16893
+IGNhZ2U= 16894
+IFBzeWNo 16895
+IOydvOuLqA== 16896
+IMW+ 16897
+aW1i 16898
+IGRpcmVjdG9ycw== 16899
+U0g= 16900
+IM+EzrfOvQ== 16901
+4buBdQ== 16902
+IGtvbnXFnw== 16903
+IG9wdGlvbmFs 16904
+cXVhcnRlcnM= 16905
+aWtlcg== 16906
+IFNhbnQ= 16907
+IHZlcnNlcw== 16908
+67aA 16909
+IG9sYXI= 16910
+IM+H 16911
+IM6zzrnOsQ== 16912
+IEltbQ== 16913
+IGNvbnRyb3ZlcnNpYWw= 16914
+IGVyc3Rlbg== 16915
+IHJlY2lw 16916
+IENocmlzdGlhbml0eQ== 16917
+IOq0nA== 16918
+b3Jkb24= 16919
+15XXqQ== 16920
+IHNsYXNo 16921
+IFBm 16922
+0YPQtNGM 16923
+15XXnQ== 16924
+IFBlcnJ5 16925
+IG1hbXk= 16926
+IGJhY2tncm91bmRz 16927
+IOCujuCuqQ== 16928
+IHBlbmRhbnQ= 16929
+IENvbHVtYmlh 16930
+IGludmVyc2U= 16931
+INGH0LXRgNC10Lc= 16932
+IHN2 16933
+IGRpZ2dpbmc= 16934
+NDE= 16935
+Y2hlbQ== 16936
+IG5hdmlnYXRpb24= 16937
+IFNoaW4= 16938
+IEZyb250 16939
+UEQ= 16940
+IGJlYXJpbmc= 16941
+IFdhc3Nlcg== 16942
+IHdheA== 16943
+IENIUklT 16944
+Y2hpbmc= 16945
+IHByZXNzZWQ= 16946
+RWw= 16947
+IERhbA== 16948
+b25zaW4= 16949
+IGJpbmRpbmc= 16950
+0YHQutC+0Lk= 16951
+cG9vbnM= 16952
+IG1vY2s= 16953
+YXJlc3Q= 16954
+0LrRgNCw 16955
+TU0= 16956
+IGNvcnJ1cHQ= 16957
+c3Rvcm0= 16958
+IHJlZnJlcw== 16959
+IENvYWNo 16960
+bGzDpA== 16961
+IFRISVM= 16962
+IHBhcmFn 16963
+IOyTsA== 16964
+cG9vbA== 16965
+IGJpbGxpb25z 16966
+IOq5gA== 16967
+Z3JvdXA= 16968
+IHdlbGNvbWluZw== 16969
+Y2VsbGVuY2U= 16970
+IER1a2U= 16971
+6ri0 16972
+IHByaW1lcmE= 16973
+7KC4 16974
+IHBvbmQ= 16975
+IHN0YXR1ZQ== 16976
+IOq1rOs= 16977
+IGhhdGNo 16978
+IGluc3RydW1lbnRhbA== 16979
+IHJlc2lkZW50aWFs 16980
+7Luk 16981
+IGFjY2VwdGluZw== 16982
+b3NoaQ== 16983
+ZGF0ZQ== 16984
+IOyUqA== 16985
+IHBsYW50ZWQ= 16986
+IGpva2luZw== 16987
+IOyEnA== 16988
+IGhhdGVk 16989
+INGA0LDRgdGB0Lo= 16990
+IHNsZXB0 16991
+IHBhY2thZ2Vz 16992
+IGlzbGFuZHM= 16993
+ZXNlbg== 16994
+xJ/EsQ== 16995
+IGRpYWdvbg== 16996
+IE9zYw== 16997
+IG1lc2g= 16998
+IHNjYWxlcw== 16999
+YXJpdHk= 17000
+IERlZmVuc2U= 17001
+IExld2lz 17002
+INGB0LXQs9C+0LTQvdGP 17003
+IGZsaWVz 17004
+dWluZWx5 17005
+IENvbnNpZGVy 17006
+IHN0YXJr 17007
+aGV3 17008
+IEFzw60= 17009
+s7Tr 17010
+IHByb3Bvc2U= 17011
+IO2VmOuptA== 17012
+b2Rv 17013
+IE5vcm1hbGx5 17014
+IGhlZWZ0 17015
+IEhhcnJpcw== 17016
+Z3Jv 17017
+IEJsb29k 17018
+YmFzZQ== 17019
+IGlPUw== 17020
+IHRvdWNoZXM= 17021
+IGluc3Bpcg== 17022
+INeT 17023
+IGJpbmFyeQ== 17024
+IOy2lA== 17025
+IHNlcmlhbA== 17026
+IGlvbg== 17027
+IHVuZW1wbG95bWVudA== 17028
+IG9kZHM= 17029
+IEZhYg== 17030
+IEZCSQ== 17031
+QlJVTg== 17032
+IHdlaWdodHM= 17033
+zr3Ovw== 17034
+YXRpbGU= 17035
+IG51cnNlcw== 17036
+IGludm9sdmVtZW50 17037
+IO2UvA== 17038
+IGdvdmVybmFuY2U= 17039
+IOKCrA== 17040
+0YDRg9C/ 17041
+aWVycmE= 17042
+7ZiV 17043
+IEplcnJ5 17044
+IGJlYXJk 17045
+IHNhbHZhdGlvbg== 17046
+IEFsb25n 17047
+Z2VudGxl 17048
+IEtp 17049
+Ym9s 17050
+IFBsYXQ= 17051
+IGhhc2h0 17052
+IHdhcmU= 17053
+IHBhcnRpZQ== 17054
+eWN6 17055
+IGludHI= 17056
+Rmlo 17057
+bmVudA== 17058
+IGNoZWF0 17059
+aWxlbg== 17060
+IOuv 17061
+b3JpZQ== 17062
+IGbDoWNpbA== 17063
+ZXRyaWM= 17064
+IGFmZmVjdGluZw== 17065
+dW5jaWF0aW9u 17066
+IGFmZmFpcnM= 17067
+IGJlZQ== 17068
+IHZpZXdpbmc= 17069
+IG9yYW5n 17070
+IExhbg== 17071
+INCh0YI= 17072
+IE1lcw== 17073
+g4E= 17074
+ZXJpZQ== 17075
+IGVzcGE= 17076
+IGludGVycHJl 17077
+IHBvc3Nlc3M= 17078
+IHB1cmVseQ== 17079
+cml0bw== 17080
+Zm91bmQ= 17081
+YXNtYQ== 17082
+7KCB7J24 17083
+IGV4YW1pbmU= 17084
+INGD0Lw= 17085
+IGJlc2No 17086
+IFRvbW9ycm93 17087
+IEJsb2Nr 17088
+IHZhcmlhbnQ= 17089
+IHByZWZlcmVuY2U= 17090
+IGNvYWNoZXM= 17091
+IG1lZGljYXRpb25z 17092
+IO2YhA== 17093
+IGVtcGlyZQ== 17094
+64Sk 17095
+IElsbGlub2lz 17096
+IGNyaXNweQ== 17097
+IHRow6w= 17098
+IGJlZXM= 17099
+Nzc= 17100
+IGdsb3c= 17101
+6Lo= 17102
+IFN0dWRpZXM= 17103
+IENoYWxsZW5nZQ== 17104
+IHVubGlrZWx5 17105
+0Kc= 17106
+xLF5b3JzdW4= 17107
+RElF 17108
+IG1pbmltaXpl 17109
+aXphcmQ= 17110
+IMO6bg== 17111
+IGVuY29udHJhcg== 17112
+IEtpbGw= 17113
+5bs= 17114
+IHZhbmlsbGE= 17115
+IEdyYW50 17116
+IEdU 17117
+c2Vh 17118
+IHNvdWdodA== 17119
+0LLQvtC0 17120
+IG7DpG0= 17121
+IEF1bnQ= 17122
+T1dO 17123
+IHB1bXBraW4= 17124
+c3RlbGxlbg== 17125
+IHJhZw== 17126
+0LXQs9C00LA= 17127
+IHN0b3J5dA== 17128
+IGZvcnVt 17129
+IGVzdGFiYQ== 17130
+dWNoZQ== 17131
+IGNvbmdyZXNz 17132
+IFJleQ== 17133
+IGRyYW1hdGljYWxseQ== 17134
+IFNwb3J0 17135
+IFllbGxvdw== 17136
+IOqzhOyGjQ== 17137
+IGRpc2d1c3Rpbmc= 17138
+IFJlY2VudA== 17139
+IGFjcXVpcmVk 17140
+IGNhYmxlcw== 17141
+ZGlu 17142
+IHZpc3Rv 17143
+IGNvbW11bmljYXRpbmc= 17144
+0YHRgtCw0LLQu9GP 17145
+0LXRgdGC0L4= 17146
+IHLDqWc= 17147
+IHNvY2tz 17148
+IHByb2Nlcw== 17149
+YmVjYXVzZQ== 17150
+IHV0dGVy 17151
+IGNvbG9jYXI= 17152
+IG5ld2VzdA== 17153
+IGdyYW1t 17154
+IHNoaWZ0aW5n 17155
+IGNhcnJpZXI= 17156
+INGB0LrQvtGA 17157
+IFNjaHc= 17158
+IGV4ZWN1dGVk 17159
+IG1haW50YWluZWQ= 17160
+IM+G 17161
+IE1vc2Vz 17162
+IGRpc3Nl 17163
+IGhvcnI= 17164
+44Cc 17165
+IHJhbGx5 17166
+IGFsbGVt 17167
+IEV2ZW50dWFsbHk= 17168
+IGRpeW9y 17169
+bHZhbmlh 17170
+IHNjaG5lbGw= 17171
+IOqzvA== 17172
+IOunpA== 17173
+IHN0cnVnZ2xlcw== 17174
+bGF0ZQ== 17175
+IGNsYXJpZnk= 17176
+w6ltZW50 17177
+IG11bHRpcGxpYw== 17178
+0LjQsdC+ 17179
+IGpvdXJu 17180
+IGZyYWdy 17181
+IHN1cnByaXNpbmdseQ== 17182
+IGRlc3BlcmF0ZQ== 17183
+NTI= 17184
+IHN1bA== 17185
+IFJlYWQ= 17186
+IEZyaWVk 17187
+IG1vbmQ= 17188
+d29v 17189
+IG9yZ2FuaXppbmc= 17190
+IFNvb24= 17191
+INCy0L7Qv9GA0L7RgQ== 17192
+IE51cg== 17193
+INCX0LQ= 17194
+IHNwaWRlcg== 17195
+0LXRgdGP 17196
+IHR1dG9yaWFscw== 17197
+IG51dHJpZW50cw== 17198
+b3Jlcg== 17199
+IGNvZWZmaWNpZW50 17200
+IGFycmFuZ2VtZW50 17201
+IHByaWNpbmc= 17202
+bmFu 17203
+eXU= 17204
+Qkw= 17205
+IHRyaWJl 17206
+IEhvd2FyZA== 17207
+dW5rcw== 17208
+IG5ld2Vy 17209
+IHByb3Zpbg== 17210
+IHByZWRpY3Rpb24= 17211
+aG9z 17212
+IG9sc3Vu 17213
+IEFyb3VuZA== 17214
+IHZpZXI= 17215
+INGB0YLQvtGA0L7QvQ== 17216
+IHZhbGxleQ== 17217
+IEVsYQ== 17218
+aWZp 17219
+IGdhbGF4eQ== 17220
+IHRyYW5xdQ== 17221
+IGFkdmVycw== 17222
+IFRlbXBsZQ== 17223
+aWZmcw== 17224
+aWdlbmNl 17225
+IGvDtm5udGU= 17226
+IMSRw7M= 17227
+RGlk 17228
+IHBob3RvZ3JhcGhz 17229
+IEFXUw== 17230
+0YbQuNGP 17231
+IGd1YXJkcw== 17232
+IGFwcG9pbnRlZA== 17233
+IEdpbA== 17234
+INC80L7QvA== 17235
+IGNvZA== 17236
+IFVubGlrZQ== 17237
+IGV2ZW5seQ== 17238
+aXNjb25zaW4= 17239
+IGVzdG91 17240
+IG1uaWU= 17241
+IEV4ZWM= 17242
+IE1W 17243
+IEVpbmU= 17244
+IFJvZ2Vy 17245
+IEZhYw== 17246
+IExpc3Q= 17247
+IGZ1ZXI= 17248
+0LDQtdGC0LU= 17249
+b21lZA== 17250
+IGF0dHJhY3Rpb24= 17251
+IHRlcnJhaW4= 17252
+IERyb3A= 17253
+IGNvcnBvcmF0aW9ucw== 17254
+IHNjaWVuY2Vz 17255
+IHRocm9uZQ== 17256
+IGFq 17257
+IFJvdA== 17258
+IHN1cHBvcnRlcnM= 17259
+IEJlcmU= 17260
+SGVyZQ== 17261
+IGRpZmVyZW50ZXM= 17262
+IHNpZ25pZmljYW5jZQ== 17263
+z4POtw== 17264
+IGNsYW1w 17265
+IOuMgOs= 17266
+IGZhYnVsb3Vz 17267
+cmV6 17268
+IGFzc3VtcHRpb25z 17269
+dXRoZXI= 17270
+d2lk 17271
+cG90 17272
+IHlhbg== 17273
+dWxpbg== 17274
+0YDRi9Cy 17275
+IFNsb3c= 17276
+IFBlbm5zeQ== 17277
+IO2VtOyEnA== 17278
+IG1laW8= 17279
+IHdlYWx0aHk= 17280
+IEVpZ2h0 17281
+IHB1bHNl 17282
+IGZyaWN0aW9u 17283
+aWRpdHk= 17284
+IEhvbGw= 17285
+aXlvcnVt 17286
+IHNvdW5kZWQ= 17287
+IENhcnI= 17288
+IGZvcms= 17289
+4pg= 17290
+IFBB 17291
+IGNvbnNwaXI= 17292
+IGNvZGluZw== 17293
+cnQ= 17294
+IFR5cA== 17295
+IOyWkQ== 17296
+INC/0L7Qsw== 17297
+IG1pc2Vy 17298
+INGB0LzQvtGC0YA= 17299
+IFN3ZWRlbg== 17300
+IG9sYXJhaw== 17301
+IFpoYW5n 17302
+IENoaQ== 17303
+IFRpdGFu 17304
+IHNjcmVlbmluZw== 17305
+IFNwaWRlcg== 17306
+IMWeaW1kaQ== 17307
+IG9ic3RhY2xlcw== 17308
+bGFyYQ== 17309
+IGNoYWxsZW5nZWQ= 17310
+cHNl 17311
+VE9O 17312
+4bul 17313
+IFBp 17314
+IGxhZ2k= 17315
+aWV1cnM= 17316
+IGh1cnRpbmc= 17317
+IG5lZ2xlY3Q= 17318
+IGdlbmVyYXRpbmc= 17319
+IHlvdW5nZXN0 17320
+IGF1ZGl0 17321
+INGA0LXQtw== 17322
+z4HOrA== 17323
+IGRvbmF0ZQ== 17324
+IFBERg== 17325
+IHZpc2l0cw== 17326
+IGNydWlzZQ== 17327
+UFA= 17328
+YXNlcg== 17329
+IHdzcA== 17330
+YmFja3M= 17331
+aXZhbHM= 17332
+IGRldmU= 17333
+IHByb3BvcnQ= 17334
+IGNhdGg= 17335
+IEVmZmVjdA== 17336
+IHdpbmRz 17337
+IOyZlA== 17338
+IGNoYXJ0cw== 17339
+IHNhbWE= 17340
+IGF1dG9tYXRpb24= 17341
+INC/0L7QutCw 17342
+IG9sYW4= 17343
+IGJvYXRz 17344
+IGNhZmU= 17345
+IGRlbmllZA== 17346
+IE1hbWE= 17347
+IGJsb2NraW5n 17348
+IFRob3I= 17349
+IHBoZW5vbWVuYWw= 17350
+IHN0YWtlaG9sZGVycw== 17351
+IHVub3M= 17352
+0YPQtdGC 17353
+IEFicmFoYW0= 17354
+IGRldGVjdGlvbg== 17355
+IGp1cmlz 17356
+IHBvd2VyZWQ= 17357
+emlhbA== 17358
+IHdlbGZhcmU= 17359
+IHVwZ3JhZA== 17360
+IG1vxbxuYQ== 17361
+IENhc2U= 17362
+Y3VsYXI= 17363
+lOydtA== 17364
+IEd1ZXNz 17365
+IGN5Y2xlcw== 17366
+cm9jaw== 17367
+dW1p 17368
+IGVsaXRl 17369
+IHF1w6g= 17370
+0YLQvtC8 17371
+IHNob3Jl 17372
+Z3VudGE= 17373
+IGt1 17374
+IGZhaXRoZnVs 17375
+IEplcmVteQ== 17376
+YWlk 17377
+4Lc= 17378
+dWdhbA== 17379
+IFZlbA== 17380
+IHZyYWk= 17381
+c3RlbGw= 17382
+qLg= 17383
+IGtvbA== 17384
+6L0= 17385
+IHF1YW50bw== 17386
+INC30LDRgA== 17387
+IDIwMDI= 17388
+ZXN5 17389
+IHJlc2VydmU= 17390
+INC80L7QvNC10L3Rgg== 17391
+IGRlcGxveWVk 17392
+IGRlZmluaW5n 17393
+IHNhdQ== 17394
+IGdhYXQ= 17395
+Iik= 17396
+IHRyYW5zbWl0 17397
+IHB1Ymxpc2hpbmc= 17398
+IHJhbmtpbmc= 17399
+IG9mZmVuc2U= 17400
+IDQ2 17401
+cGlu 17402
+IFRha2luZw== 17403
+IGVudGl0bGVk 17404
+IGdlbnVpbmVseQ== 17405
+IHZhcmlhdGlvbnM= 17406
+IGZpbmRl 17407
+IHRhdQ== 17408
+IHVuZm9ydHVuYXRl 17409
+IFJhaA== 17410
+cG9ydHM= 17411
+IGPF 17412
+IG1vbmtleQ== 17413
+IGJyYWM= 17414
+d2Vp 17415
+bHVuZw== 17416
+IGFydGlm 17417
+IHN5cnVw 17418
+INCU0LDQsg== 17419
+IGxpZnRlZA== 17420
+IGNoZXo= 17421
+IEFkdmVudA== 17422
+IFN0b2Nr 17423
+IGRvbA== 17424
+0LzQtdC9 17425
+0LjRiNGM 17426
+IHlu 17427
+Z2lv 17428
+ZGV0 17429
+IGRlc3Nl 17430
+IGdyaQ== 17431
+IENoYWlybWFu 17432
+54U= 17433
+IGN1ZW50YQ== 17434
+YW5pbQ== 17435
+IGNyYWI= 17436
+IGVzY2Fs 17437
+IHByZW1pw6hyZQ== 17438
+IEdlZg== 17439
+IGRpbmluZw== 17440
+IHNldmVudGg= 17441
+IGNoYXNpbmc= 17442
+IFRvd2Vy 17443
+IGJydXRhbA== 17444
+IGZ1bmRhbWVudGFsbHk= 17445
+0LvQtdC90LjRjw== 17446
+c3RhZ2U= 17447
+IGFjcXVpcw== 17448
+IGN5bGluZGVy 17449
+IGNvbW1hbmRlcg== 17450
+bWVt 17451
+IFVW 17452
+aGFwcHk= 17453
+IGVwc2lsb24= 17454
+IGludml0YXRpb24= 17455
+IGZhcm1lcg== 17456
+Y2hhaXI= 17457
+IGRlc3Rpbnk= 17458
+IHNvdmVyZQ== 17459
+IEhlYnJldw== 17460
+IHNlcnZhbnQ= 17461
+IGJldw== 17462
+IGdhc3Q= 17463
+dXRpZXM= 17464
+IGFkbWluaXN0cmF0aXZl 17465
+IENvbW1hbmQ= 17466
+w6l0YQ== 17467
+IG5pdHJvZ2Vu 17468
+6re8 17469
+IGFiaQ== 17470
+IHZpbGxhaW4= 17471
+IGJsYW5rZXQ= 17472
+IFNlbmQ= 17473
+IGJlYXRlbg== 17474
+soQ= 17475
+IHZvbHVudA== 17476
+IHNjaG9sYXI= 17477
+IEVtcGVyb3I= 17478
+IDQz 17479
+dmFibGU= 17480
+IER1cw== 17481
+IEdV 17482
+IHRhcmdldGluZw== 17483
+d3d3 17484
+IGFtZW5kbWVudA== 17485
+7IaM6w== 17486
+IHRpbmc= 17487
+IG5hc3R5 17488
+IGdhdWdl 17489
+INGA0L7QtA== 17490
+IEhhbnM= 17491
+WW91cg== 17492
+zrHOvQ== 17493
+IHByb2pldA== 17494
+IEhhd2FpaQ== 17495
+IHN1c3BpY2lvdXM= 17496
+IHNjaHc= 17497
+IHJlbW92YWw= 17498
+IGludHJpZw== 17499
+IE1V 17500
+IHBvbnRv 17501
+4KS+ 17502
+INC+0LHRgNCw0Lc= 17503
+IGd1ZXNzaW5n 17504
+cGFjZQ== 17505
+IG1vdGhlcnM= 17506
+IG1pbGxpbWV0ZXI= 17507
+0LvQtdC90LjQtQ== 17508
+IGF2YWlsYWJpbGl0eQ== 17509
+aWN6 17510
+IGZyYWN0 17511
+IGJhc2Vz 17512
+a20= 17513
+IEJUUw== 17514
+IEZpZWxk 17515
+IGR6aWU= 17516
+IHNlZ3VuZG8= 17517
+IOuCmOuKlA== 17518
+IGxlZ2l0aW1hdGU= 17519
+aW1hcw== 17520
+INCy0L0= 17521
+IGNvcnJ1cHRpb24= 17522
+IHNtYXNo 17523
+IFZhbGVudA== 17524
+IGFsaWduZWQ= 17525
+IFBlbm5zeWx2YW5pYQ== 17526
+IGdhYg== 17527
+IEV1bg== 17528
+ZW50aA== 17529
+IE1vcm5pbmc= 17530
+IGNhbmRsZQ== 17531
+IGJhY2twYWNr 17532
+IElzbGFtaWM= 17533
+YcOnw7Vlcw== 17534
+IGVuY3J5 17535
+IG11c2hyb29tcw== 17536
+7YyM 17537
+ZGl0 17538
+IHRyYW5zaXQ= 17539
+IFdpc2NvbnNpbg== 17540
+IHBhcnRpY2lwYXRlZA== 17541
+IElscw== 17542
+IHVuZm9sZA== 17543
+toDr 17544
+IHByb2ZpdHM= 17545
+IHdhcm1pbmc= 17546
+IEdhbmc= 17547
+IG5ldHdvcmtpbmc= 17548
+IG1lZ2E= 17549
+IHRob3JvdWdobHk= 17550
+bGVtZW50cw== 17551
+IEht 17552
+IGRlY2lkaW5n 17553
+IGVtb3Rpb25hbGx5 17554
+IGV4aGF1c3RlZA== 17555
+INCf0L7Rgg== 17556
+Y2lkbw== 17557
+IEhUTUw= 17558
+IGNvcHlyaWdodA== 17559
+IG1lbG9keQ== 17560
+eWlt 17561
+IGFuZGVycw== 17562
+b3Nob3A= 17563
+IOuzvA== 17564
+IGF0aGxldGU= 17565
+IEdF 17566
+IGZyZXF1ZW50 17567
+IGRlc2lyZXM= 17568
+IG5lZWRpbmc= 17569
+IFl1bg== 17570
+IHJpZmxl 17571
+IGxvdmVy 17572
+J1Q= 17573
+IGRlbnNl 17574
+IHTDo28= 17575
+IG5vdGlmaWVk 17576
+IGlkaQ== 17577
+7Jet 17578
+7YY= 17579
+IGludGVyYWN0aW5n 17580
+IHJhcHBvcnQ= 17581
+0LXRgNC4 17582
+c2tp 17583
+IGJlc3Nlcg== 17584
+IG1hbnVmYWN0dXJlcg== 17585
+IEt5bGU= 17586
+IGFjY291bnRhYmxl 17587
+IFNhaw== 17588
+IFBpbA== 17589
+IERvbWlu 17590
+IHByZXN1bQ== 17591
+INCS0YHQtQ== 17592
+IHZpbmVnYXI= 17593
+IGd1YXJhbnRlZWQ= 17594
+IGhhbmRsZWQ= 17595
+Y2F0 17596
+IGNpdmlsaXphdGlvbg== 17597
+IGFjY29tcA== 17598
+IFZN 17599
+w6ltb24= 17600
+IGRlemU= 17601
+IGdyYWRlcw== 17602
+IHNvbGx0ZQ== 17603
+IHN0YXJpbmc= 17604
+15DXqg== 17605
+YXJudA== 17606
+IGhvcml6b24= 17607
+IHRyYXZhaWw= 17608
+aG91cg== 17609
+IEVE 17610
+IERhaw== 17611
+IG55 17612
+IGNvbnZl 17613
+IENoYW0= 17614
+IGZpcm1z 17615
+IExpdQ== 17616
+INGB0YLRgNCw0L0= 17617
+IGxpYmVydA== 17618
+IGxlbnNlcw== 17619
+IGludGFrZQ== 17620
+INCy0YvQsQ== 17621
+IG1lbnNlbg== 17622
+aGVs 17623
+IHByYWN0aXRpb24= 17624
+IDM1MA== 17625
+Rk8= 17626
+IGJlZHM= 17627
+IGFuY2VzdG9ycw== 17628
+IOyXhOyyrQ== 17629
+IGRpc3R1cmI= 17630
+IExhc3RseQ== 17631
+IFN1cHBvcnQ= 17632
+4Li14LmJ 17633
+IENvcm9uYQ== 17634
+IGVudGh1c2k= 17635
+INCy0L7Qt9C8 17636
+IOyCrOuejOs= 17637
+IDUy 17638
+YmlyZA== 17639
+IHJlZHVjZXM= 17640
+IOyeiOydhA== 17641
+IEdlbmU= 17642
+6rWQ 17643
+xJlw 17644
+IMOcYmVy 17645
+IGNvbmNlcm5pbmc= 17646
+dXNlcg== 17647
+IGNvbmNlbnRyYXRl 17648
+IFdIQVQ= 17649
+aXNob3A= 17650
+b255bW91cw== 17651
+bm9sZA== 17652
+IHN1Z2dlc3Rpbmc= 17653
+qbA= 17654
+IEZpc2g= 17655
+Li4uLi4uLi4= 17656
+IHZlc3NlbA== 17657
+IHRyYWJham8= 17658
+IE9jZWFu 17659
+eWc= 17660
+IHRvd25z 17661
+ZGVs 17662
+IHRlcnJpZnlpbmc= 17663
+IMOnYWzEscWf 17664
+IHNpbm8= 17665
+IGVhdHM= 17666
+IGdleg== 17667
+IGdlbWU= 17668
+IOyZhA== 17669
+IGNvbXBhcnQ= 17670
+IGltcGxlbWVudGluZw== 17671
+IFBvdHRlcg== 17672
+IEdlcm1hbnM= 17673
+IGfFgg== 17674
+IHRlbm5pcw== 17675
+IGNhcnBldA== 17676
+YXVlcg== 17677
+IFNhdWRp 17678
+eWVvbmc= 17679
+IGN1cnJ5 17680
+IEZvcmVzdA== 17681
+0YvQuw== 17682
+IGZpZnRlZW4= 17683
+IGJvbHRz 17684
+IHtc 17685
+rLQ= 17686
+IHNldHRsZW1lbnQ= 17687
+IGxhbmdl 17688
+IGJhbQ== 17689
+R2V0 17690
+7ZWZ 17691
+IHN3YXA= 17692
+IEtoYW4= 17693
+IGNvbW1lbmNl 17694
+IHF1YXJhbnRpbmU= 17695
+IHNjb3JlZA== 17696
+55Y= 17697
+IDE5NTA= 17698
+IHRoaWNrZXI= 17699
+IHPDu3I= 17700
+IExhcnJ5 17701
+IGFsbGV6 17702
+7Iuc64qU 17703
+IGfDvA== 17704
+IHNwZWN0YWN1bGFy 17705
+Ly8= 17706
+Ym90aA== 17707
+IHN0YXRz 17708
+IE5hbmN5 17709
+IGJ1bnU= 17710
+IGNydXN0 17711
+IGFjdGl2YXRlZA== 17712
+IOq3uOue 17713
+b3V0aGU= 17714
+IHBvcnRz 17715
+IG5ldXJhbA== 17716
+IGphdw== 17717
+IG9ic2VydmF0aW9ucw== 17718
+IHZvaXQ= 17719
+YWJhbg== 17720
+4bqjaQ== 17721
+pqzrpbw= 17722
+b21lcw== 17723
+4K+L 17724
+cXVp 17725
+IGtpbmRuZXNz 17726
+0JE= 17727
+IDQx 17728
+IG1vZGVyYXRl 17729
+IGFuZ2Vscw== 17730
+IFRhbWI= 17731
+w6h0 17732
+IGNobG9y 17733
+IEJpbGx5 17734
+7LKY6w== 17735
+YWNvbg== 17736
+IHNlbGVjdGluZw== 17737
+IERlbHRh 17738
+IG51bGw= 17739
+ZGVubHk= 17740
+IGNpdWQ= 17741
+IHRlbmRlbmN5 17742
+IGJyZWFrZG93bg== 17743
+IG1pbnQ= 17744
+0YTQvtGA0Lw= 17745
+b3JwaA== 17746
+IGRhd24= 17747
+c3By 17748
+IFdJTEw= 17749
+w6RjaGxpY2g= 17750
+IHB1cHB5 17751
+NzAw 17752
+IOCupA== 17753
+IGZhaWxz 17754
+IENvbmM= 17755
+IHJlbGF0aXZlcw== 17756
+IGludml0aW5n 17757
+IGF1dG9ub20= 17758
+IGNvbXBvc2Vk 17759
+IHVuaXR5 17760
+IGRlY2lz 17761
+IGFjY2Vzc29yaWVz 17762
+IENhc3M= 17763
+IGJpc3Q= 17764
+IFRpcA== 17765
+7Ke4 17766
+IHB1bnQ= 17767
+IHLDoXA= 17768
+QU5L 17769
+ZXhpc3Q= 17770
+IGNvbXBhdGlibGU= 17771
+IG5lcg== 17772
+INC10LzRgw== 17773
+IGFwbGlj 17774
+IGJhcHQ= 17775
+IGZhaWxpbmc= 17776
+IFRhbWFt 17777
+IG9zY2lsbA== 17778
+IGxldHp0ZW4= 17779
+IHJlcGVhdGVkbHk= 17780
+IGp1bmdsZQ== 17781
+IFB1c2g= 17782
+aGFp 17783
+IM63 17784
+IGRlYWRseQ== 17785
+0Y/Qtg== 17786
+d2nEhQ== 17787
+IENvbW1vbg== 17788
+IM6V 17789
+IHNrYXRl 17790
+VEM= 17791
+IE1pbmk= 17792
+IGhvYmJ5 17793
+4bqnbg== 17794
+IHJvdXRlcw== 17795
+IGFtaWdvcw== 17796
+IGNvbmp1bg== 17797
+IHBhcnRuZXJzaGlwcw== 17798
+IG5vdm8= 17799
+IGF2ZXI= 17800
+IHBvdXZleg== 17801
+YnJpZGdl 17802
+IHByZW9j 17803
+aGlt 17804
+IHR1cmI= 17805
+IHNvYg== 17806
+IFNuYXA= 17807
+IOywuA== 17808
+bWludXRl 17809
+IHRyYWplY3Q= 17810
+dWrEmQ== 17811
+IGVhZ2Vy 17812
+IHJlZ3VsYXRvcnk= 17813
+IGJhbmtpbmc= 17814
+Ymxpbmc= 17815
+0YjRjA== 17816
+YcW8 17817
+IGJpemFycmU= 17818
+aXRhdGVk 17819
+ZGlyZQ== 17820
+IHRocmVhdGVuZWQ= 17821
+IHNoaW5pbmc= 17822
+IG5lc3Nl 17823
+IGNvcnBz 17824
+INGB0YM= 17825
+IHRlbGVz 17826
+IHRlbXA= 17827
+dGVt 17828
+INC60LDQvQ== 17829
+IGZldmVy 17830
+TmV3 17831
+IGhlYXZpZXI= 17832
+IFNhaA== 17833
+YnVk 17834
+IG91dHJvcw== 17835
+IOywvg== 17836
+IOuqhQ== 17837
+YXJyaW5n 17838
+IOq0nOywrg== 17839
+IE5hcA== 17840
+IHNlbWlu 17841
+IFRoYW4= 17842
+aWZz 17843
+IGRlc2Vu 17844
+INGC0LDQutC+0LU= 17845
+IGxvc2Vz 17846
+IEJhbHQ= 17847
+a29u 17848
+INC90LDQv9GA 17849
+IHZvaXM= 17850
+IE1vc2Nvdw== 17851
+IGNoYWlycw== 17852
+aGlz 17853
+IHJlZnVnZWVz 17854
+a2c= 17855
+IGtvbGU= 17856
+jag= 17857
+0LDRgdC40LHQvg== 17858
+pr0= 17859
+IFVuaXZlcnNl 17860
+IERpcmVjdA== 17861
+IGNoZWF0aW5n 17862
+IENpbg== 17863
+IHBhdHJp 17864
+IGFkdmlzZQ== 17865
+IE5ldGhlcg== 17866
+IHByaW1laXJv 17867
+IG1lbnRpb25pbmc= 17868
+bnV0 17869
+NTY= 17870
+YXLEsQ== 17871
+IHBldGl0ZQ== 17872
+YmxlZA== 17873
+IHBlbnNhcg== 17874
+aWNpbw== 17875
+SU5E 17876
+IHZldGVyYW4= 17877
+IGxhZGRlcg== 17878
+IGNvbnNlcXVlbmNl 17879
+0L7QttCw0Ls= 17880
+IEJ1cm4= 17881
+IHJ1Zw== 17882
+IE1hZGU= 17883
+IGdpdA== 17884
+Ii4uLg== 17885
+IGNvbXBldGl0b3Jz 17886
+IHByemVk 17887
+IGFwcGFyZW50 17888
+IEFyZ2VudGluYQ== 17889
+IFdvcmtpbmc= 17890
+IGNvbGxhYm9yYXRl 17891
+d29tYW4= 17892
+IHJldGFpbg== 17893
+IGxldXJz 17894
+IGRhc2hib2FyZA== 17895
+15nXkw== 17896
+IEVhcmx5 17897
+Qk0= 17898
+INC10ZE= 17899
+0L7Qu9C+0LM= 17900
+IHNhdGlzZnlpbmc= 17901
+IG9mdGVudGltZXM= 17902
+IG1hcHBpbmc= 17903
+w7xua8O8 17904
+YXJ0aA== 17905
+Zm9sZA== 17906
+IGxhdW5jaGluZw== 17907
+IGF1cmE= 17908
+IHByZWNpc2lvbg== 17909
+d29ya3M= 17910
+R29k 17911
+IHN0cmFw 17912
+IEltcGVy 17913
+IHJpdmVycw== 17914
+IHw= 17915
+IGN1ZXI= 17916
+cmVnb24= 17917
+IGFycml2YWw= 17918
+0LrQsNGF 17919
+IE1pYW1p 17920
+0LDQvdGL 17921
+IHN1cnZpdm9ycw== 17922
+IFNlbmlvcg== 17923
+RGF2aWQ= 17924
+IGVzdGFkbw== 17925
+IHNlY3RvcnM= 17926
+IHBvcHBpbmc= 17927
+IGNoaW0= 17928
+YXnEsQ== 17929
+IGt1bm5lbg== 17930
+IGdhbGxlcnk= 17931
+IHN1bmxpZ2h0 17932
+ZXNlaGVu 17933
+IHllbGxpbmc= 17934
+IE1laW4= 17935
+IFBob2VuaXg= 17936
+IG1hbm8= 17937
+IGhpc3Rvcmlh 17938
+IG9jY3VycmluZw== 17939
+7Lg= 17940
+0LDQtNC4 17941
+IGluc3RpdHV0aW9uYWw= 17942
+IFR1dA== 17943
+57I= 17944
+IHNsYXZlcw== 17945
+IGZvcmdpdmVuZXNz 17946
+IHR3aW4= 17947
+IEh5dW4= 17948
+0L3RjA== 17949
+IEtvbW0= 17950
+YW5kcmE= 17951
+c2hvdA== 17952
+c3PDpA== 17953
+INGG0LU= 17954
+YXR0YQ== 17955
+IGV4cGVuc2U= 17956
+IEdQVQ== 17957
+IFBhc3Q= 17958
+cmlibHk= 17959
+IOutkOyVvA== 17960
+INCz0L7QtNCw 17961
+IHJlc3Bpcg== 17962
+IFF1ZWVucw== 17963
+aG9wcw== 17964
+IHPDqXJpZQ== 17965
+IHByZWY= 17966
+IGNvbWVk 17967
+IHBsdXQ= 17968
+IE92ZXJhbGw= 17969
+IGN1c2g= 17970
+IHJpbmdpbmc= 17971
+IGluY29ycmVjdA== 17972
+INGB0YLRgA== 17973
+IGdlb21ldHJ5 17974
+IGFkdmVydGlz 17975
+INCo 17976
+IHJldmlld2Vk 17977
+IGRvemVucw== 17978
+IGRldGVybWluYXRpb24= 17979
+IFBoaWxs 17980
+IGNvbnRyaWJ1dGVk 17981
+IENpdA== 17982
+IHBhc3NlbmdlcnM= 17983
+IGPDtHTDqQ== 17984
+IHJldmVy 17985
+IHRlY2hub2xvZ2ljYWw= 17986
+IGFsbGVu 17987
+IHJhaW5pbmc= 17988
+YXZp 17989
+IHNhbHR5 17990
+IHR5cGluZw== 17991
+INGC0LU= 17992
+IHRpbHQ= 17993
+IOy5mA== 17994
+INC+0YA= 17995
+INC/0YDRj9C8 17996
+IHJvdQ== 17997
+IGFyZW5h 17998
+YXJhdA== 17999
+SEhISA== 18000
+IG1hbnVmYWN0dXJlcnM= 18001
+IEVkd2FyZA== 18002
+IHR1Y2s= 18003
+IGJsb3dz 18004
+aW5nbw== 18005
+IE1hcmM= 18006
+7JWE7ISc 18007
+TWljaA== 18008
+IENsZWFu 18009
+6LQ= 18010
+ZXN0bw== 18011
+IFBhY2s= 18012
+IHNoYWZ0 18013
+QlJVTk8= 18014
+IGF2ZW4= 18015
+dXVy 18016
+0YHQutC+0LvRjNC60L4= 18017
+6rSA 18018
+IGF1dG9tYXRlZA== 18019
+IHZlbnR1cmU= 18020
+IHN1cnZlaWxsYW5jZQ== 18021
+IEdyb3c= 18022
+IEVtZXI= 18023
+INC00L7RgA== 18024
+IGludmVzdG9y 18025
+IFlvaw== 18026
+IGxhdHRlcg== 18027
+IE5J 18028
+IGZ1bmN0aW9uaW5n 18029
+IEhhbWlsdG9u 18030
+IDUx 18031
+IG11cmRlcmVk 18032
+IGFuY2hvcg== 18033
+IGN1Yw== 18034
+IFNDUA== 18035
+IE1hZGFt 18036
+IGNvbnN0cmFpbnRz 18037
+IGJhcm4= 18038
+YW5rZW4= 18039
+IOunjuydgA== 18040
+IE1vdG9y 18041
+IERvaW5n 18042
+IGFtZW4= 18043
+ZXR0cw== 18044
+IGluc3RydWN0b3I= 18045
+ZWd0 18046
+YWtv 18047
+IHBvc3R1cmU= 18048
+aXZpYQ== 18049
+IFBvbGlzaA== 18050
+INC00LLQsA== 18051
+IGNvbG9yZnVs 18052
+IGVsYm93 18053
+IHBhcmxl 18054
+IHBhc3Nlcg== 18055
+IGNvbmRlbQ== 18056
+b3J0YWw= 18057
+IGZlcnRpbA== 18058
+2KfYrw== 18059
+IENvbG9tYg== 18060
+IGFsaWdubWVudA== 18061
+IGFzdHJvbmF1dA== 18062
+IE11dA== 18063
+IHNhbG1vbg== 18064
+IHN0cnVjdHVyZWQ= 18065
+nteo 18066
+IGNsaWNrcw== 18067
+IG1pZWo= 18068
+IFJvdW5k 18069
+IHJhaW5ib3c= 18070
+IFZB 18071
+7KeI 18072
+b3R6 18073
+LDwv 18074
+IE5pY29sZQ== 18075
+bGlzaGluZw== 18076
+IHdoaWxzdA== 18077
+IHJlcHVibGlj 18078
+IHRhbWFt 18079
+dmVydGVk 18080
+IHJlY29nbml6aW5n 18081
+INCz0LvQsNCy 18082
+IGR1Yg== 18083
+IEpvcw== 18084
+ZmFsbHM= 18085
+aWNoaQ== 18086
+IGN6xJk= 18087
+INCm 18088
+IE1pdGNo 18089
+Q1I= 18090
+Y2xpY2s= 18091
+IHN0dW5uaW5n 18092
+IEp1bGlh 18093
+bWVycw== 18094
+IFBvbHk= 18095
+IGRlc3Nh 18096
+IGludMOp 18097
+IOqzoOs= 18098
+IGRvxJ8= 18099
+IGRpdmVy 18100
+IHN0cmlraW5n 18101
+YXBob3I= 18102
+IGFwZW5hcw== 18103
+b3VzZXM= 18104
+IHRyYWdlZHk= 18105
+IEZhbg== 18106
+IFR1cmtpc2g= 18107
+IHByb3BoZXQ= 18108
+IGRpc3RhbmNpbmc= 18109
+IEhlbQ== 18110
+IGNhcnRvb24= 18111
+S2U= 18112
+YW50aW5n 18113
+IENsYXJr 18114
+578= 18115
+IGRhdm9u 18116
+IO2F 18117
+IHl1bW15 18118
+IGNvbXByb21pc2U= 18119
+IHN0YXJ0dXA= 18120
+cml0dA== 18121
+IGNlcnRpZmllZA== 18122
+IHBpbGxvdw== 18123
+YmVyZQ== 18124
+7KSA 18125
+IHNlZ3Vpcg== 18126
+IHN0YWRpdW0= 18127
+YXRpdm8= 18128
+IHNpbXBsZXI= 18129
+s7g= 18130
+IHZpc2E= 18131
+IHBhdGh3YXk= 18132
+IG51ZXZv 18133
+IHJheQ== 18134
+6Zw= 18135
+w7bDnw== 18136
+INC30LDQvQ== 18137
+IGNlbGVicml0eQ== 18138
+0LfQsA== 18139
+IGVpbmVz 18140
+IEdpdmVu 18141
+IEFyYQ== 18142
+IEpvYg== 18143
+IHlhaw== 18144
+IEFyYmVpdA== 18145
+cmVzc2luZw== 18146
+w6FuZA== 18147
+IGdyYWJiZWQ= 18148
+cGVuZA== 18149
+IHNpbmU= 18150
+aXJr 18151
+INCe0YI= 18152
+IEZsZQ== 18153
+aWNoZW4= 18154
+56Y= 18155
+IE5laWw= 18156
+IHJlcGVhdGluZw== 18157
+IGRyYXdpbmdz 18158
+cmlzZQ== 18159
+IGdsaXR0ZXI= 18160
+Zml2ZQ== 18161
+IHN1cnQ= 18162
+IHNpY2hlcg== 18163
+IGFkanVzdG1lbnRz 18164
+aXBwaQ== 18165
+Y2tl 18166
+IHJlcHJlc2VudGF0aXZlcw== 18167
+IG1pZHN0 18168
+IHNwb2ls 18169
+bWV5ZQ== 18170
+IHRhZ3M= 18171
+IHllcA== 18172
+IFN0ZXBoYW5pZQ== 18173
+IGdlcmU= 18174
+IFJ1ZA== 18175
+54s= 18176
+IGdyb3M= 18177
+IHF1ZXVl 18178
+IGFjY29yZA== 18179
+IG9yZ2FuaXNhdGlvbg== 18180
+ZW5keQ== 18181
+IFRleHQ= 18182
+w7x5b3I= 18183
+IMOt 18184
+IGNvbmNsdXM= 18185
+IOykgOs= 18186
+IGFtcA== 18187
+IExlc3M= 18188
+IOuQmOuKlA== 18189
+Y2Fubw== 18190
+IFBpeA== 18191
+YXBlZA== 18192
+IGRhcmF1Zg== 18193
+dW8= 18194
+eW50aA== 18195
+YWJlbA== 18196
+IERvbmU= 18197
+IGRpY2s= 18198
+YXRob24= 18199
+IGhpbGFy 18200
+YWNjbw== 18201
+IOyGjQ== 18202
+IE9yZWdvbg== 18203
+IFdlaWw= 18204
+IG1hdGhlbWF0aWNz 18205
+IGFsbQ== 18206
+IHBpeGVscw== 18207
+IGZyw6Vu 18208
+0LHQvg== 18209
+RkM= 18210
+0L3Rjg== 18211
+aGVpbQ== 18212
+Z29z 18213
+IEZvcmdldA== 18214
+ZmVuZA== 18215
+IFZvaWzDoA== 18216
+IEdyZWV0 18217
+IM6xz4XPhA== 18218
+IHJlY3Vy 18219
+NTE= 18220
+IOyeiOqzoA== 18221
+QXQ= 18222
+IHlhcmRz 18223
+0LjRgtC4 18224
+IG9mZnNldA== 18225
+cm9sbGluZw== 18226
+INCf0L7RgQ== 18227
+IGVubGlnaHQ= 18228
+IFBhZA== 18229
+bGltaXRlZA== 18230
+0LjQu9GM0L3Qvg== 18231
+IFNhcmE= 18232
+INGB0LTQtdC70LDRgtGM 18233
+bWFydA== 18234
+IEp1bXA= 18235
+IGFkb3JhYmxl 18236
+b3JzZQ== 18237
+Y2hlZXJpbmc= 18238
+IGVtcGF0aHk= 18239
+IFRvbmlnaHQ= 18240
+b3Jw 18241
+IEh1bnRlcg== 18242
+UG9pbnQ= 18243
+0LPQsA== 18244
+IHBhc3Nlbmdlcg== 18245
+IEtuaWdodA== 18246
+IHNlZW1pbmdseQ== 18247
+aHVo 18248
+IHRoZWF0cmU= 18249
+IHRvbWI= 18250
+IGRlcHJlc3NlZA== 18251
+IHN1bW1vbg== 18252
+IHNhdGlzZmFjdGlvbg== 18253
+ZG9vcnM= 18254
+IEhvdXN0b24= 18255
+0LDRjtGJ 18256
+IFJpbw== 18257
+0LPQu9GP 18258
+IGFycmFuZ2Vk 18259
+IGhhbmRsZXM= 18260
+IHRyaWxsaW9u 18261
+IG5pZ2h0bWFyZQ== 18262
+IFF1YW5kbw== 18263
+IG9sZQ== 18264
+IEd1aWRl 18265
+b29v 18266
+IGJpbGU= 18267
+IGVtcGV6 18268
+IDcy 18269
+Y3JpYmVk 18270
+IHByb2dyZXNzaW9u 18271
+IExpbnV4 18272
+66as 18273
+IOyymOydjA== 18274
+IGZvc3NpbA== 18275
+IHF1ZXJv 18276
+7Iah 18277
+YXRpdmE= 18278
+IHB1eno= 18279
+IFp1cw== 18280
+IHRocmlsbGVk 18281
+IENC 18282
+IG1pbmVy 18283
+0YDQsNGJ 18284
+IFNBUg== 18285
+IE5vcw== 18286
+INCz0L7RgNC+0LQ= 18287
+IGNhbWI= 18288
+INGC0LA= 18289
+IHJlc3VsdGVk 18290
+IERpY2s= 18291
+b3VuZw== 18292
+IGNvbWljcw== 18293
+IGFic29sdXQ= 18294
+c3Rhbg== 18295
+ZGltZW5zaW9uYWw= 18296
+IHRlbnNl 18297
+bXVz 18298
+IEludGVsbA== 18299
+INGN0YLRgw== 18300
+IHBoYXNlcw== 18301
+IHZvbHRh 18302
+IHbDo28= 18303
+Ym91bmQ= 18304
+IEFuZGVyc29u 18305
+IGN1cmlvc2l0eQ== 18306
+IHBvbnQ= 18307
+IGRlbW9uc3RyYXRlZA== 18308
+b2xpbmU= 18309
+IFNwZWVk 18310
+IG1hbWE= 18311
+IHNob2NraW5n 18312
+IGtpZWR5 18313
+IGVhcnRocXVha2U= 18314
+IGltcGxpZXM= 18315
+IGVudGVycw== 18316
+noA= 18317
+IGVsZXZhdG9y 18318
+IGRlbGlnaHRlZA== 18319
+IE1pdHQ= 18320
+IEJhc2Vk 18321
+IERvbA== 18322
+IGtlbg== 18323
+IHdvcnJ5aW5n 18324
+IGZpbGVk 18325
+YWlsYW5k 18326
+INC80LXRgg== 18327
+IG1hc2M= 18328
+IM6R 18329
+IEp1bGll 18330
+IGRpbWVuc2lvbmFs 18331
+aHVtYW4= 18332
+VG9r 18333
+w78= 18334
+IHVuc3Q= 18335
+IHNldWxl 18336
+IGVtYmFy 18337
+IO2VqeuLiOuLpA== 18338
+YWNpb24= 18339
+IOyJ 18340
+IOu2gOu2hA== 18341
+IGhlYXRlZA== 18342
+4oCm4oCm 18343
+IiE= 18344
+IHJlYWxpc2U= 18345
+0LXRgtGL 18346
+aWVuaWE= 18347
+aWV6 18348
+IGbDvGg= 18349
+IEVzc2U= 18350
+IHBz 18351
+IGTDsw== 18352
+YXN0ZXJz 18353
+IG9ucw== 18354
+UE0= 18355
+IHJldHJv 18356
+bWFrZXI= 18357
+d2hlbg== 18358
+IGVsbGE= 18359
+IExpdmluZw== 18360
+IExhbQ== 18361
+IHRyb25n 18362
+IGFwcHJvdmU= 18363
+IM64zrE= 18364
+IHN1bmc= 18365
+0LXQvdC40Y4= 18366
+IFJlbW92ZQ== 18367
+w6huZQ== 18368
+aXJlbg== 18369
+IHN0cmFuZ2Vy 18370
+0LjQvdGL 18371
+IHbDpg== 18372
+YWZ0ZXI= 18373
+b3R0bw== 18374
+lOuhnA== 18375
+IEFob3Jh 18376
+bWlsbA== 18377
+SVNI 18378
+IGdyYWR1YXRpbmc= 18379
+a3Rl 18380
+IHJlbm92 18381
+IHByb2Nlc3NlZA== 18382
+a2V5cw== 18383
+0LXQutC+ 18384
+IGVucmljaA== 18385
+IMWfZWs= 18386
+IGluc2Vj 18387
+IE5hbg== 18388
+Y2FrZXM= 18389
+IGlsbHVzaW9u 18390
+mOulvA== 18391
+IGFpcmw= 18392
+aW1z 18393
+IGFudGVu 18394
+4buvbmc= 18395
+c24= 18396
+IHByZWNpc2E= 18397
+6riw7J6Q 18398
+INin2YTYuQ== 18399
+IGZvcmVtb3N0 18400
+IHBhcmFncmFwaA== 18401
+YXZhaXM= 18402
+INCy0L7RgQ== 18403
+IG1hbnM= 18404
+w61maWM= 18405
+Ym90 18406
+INi52YY= 18407
+IGJyb3Ro 18408
+IGFsdGVybmF0ZQ== 18409
+IENoYXB0ZXI= 18410
+IHZlY3RvcnM= 18411
+ZXNhcg== 18412
+IGluZGljYXRpb24= 18413
+IE5laW4= 18414
+toE= 18415
+IGplYW5z 18416
+WUU= 18417
+Y29uZA== 18418
+IHVuaXRlZA== 18419
+YWJp 18420
+IFNlcmdl 18421
+IHBhcnRpYWxseQ== 18422
+IG1hY3Jv 18423
+IGV0aGljYWw= 18424
+cnVpdA== 18425
+IHNoaWZ0ZWQ= 18426
+IGNhYmU= 18427
+IG1hdGhlbWF0aWNhbA== 18428
+IHJ1ZGU= 18429
+15nXldeq 18430
+IE1lcmM= 18431
+IGdhbnpl 18432
+aWNpb24= 18433
+IHVuY29uc2Npb3Vz 18434
+IGJ1cm50 18435
+INGA0LXQsQ== 18436
+7Yq46w== 18437
+IGNoYXJt 18438
+YW5kYWw= 18439
+7LKc 18440
+b3RoeQ== 18441
+IEhhZGk= 18442
+IGFwcHJlY2lhdGlvbg== 18443
+RU5E 18444
+IHLDqWFs 18445
+toTrk6Q= 18446
+IE5hZw== 18447
+oKTqs6A= 18448
+IExhdXJlbg== 18449
+IHbhu5tp 18450
+IEJyaWRnZQ== 18451
+IFVtbQ== 18452
+IFdlZw== 18453
+IGNoYXF1ZQ== 18454
+IFNvcGg= 18455
+IGdkemll 18456
+7ZGc 18457
+IHN0ZXI= 18458
+IEJsYQ== 18459
+IHJlZmxlY3Rz 18460
+IGJlbmNobWFyaw== 18461
+0LLQsNGC 18462
+YW1pbmU= 18463
+IGFuaA== 18464
+IGNvbnRpbmVudA== 18465
+IEZEQQ== 18466
+7KGw 18467
+IMOqdGVz 18468
+15nXkA== 18469
+IGJsb29keQ== 18470
+IE5pbmU= 18471
+aWVsdA== 18472
+ZW1hbmQ= 18473
+IOuztOqzoA== 18474
+IHRpZGFr 18475
+IFNjaWVudA== 18476
+cGxleA== 18477
+b3N0ZW4= 18478
+IGFuaW1hdGVk 18479
+YXNzYQ== 18480
+IGRlcml2ZWQ= 18481
+INC40YHRgtC+0YA= 18482
+IE1pZw== 18483
+7IWY 18484
+IHJvcw== 18485
+cGx1cw== 18486
+b3NhdXI= 18487
+IF4= 18488
+IGludGVuc2l2ZQ== 18489
+IGdsb2JhbGx5 18490
+IGRpZmVyZW4= 18491
+7J206rOg 18492
+xIVk 18493
+IGTDqXM= 18494
+IHByZXNlbnRhdGlvbnM= 18495
+IENybw== 18496
+IGVzc2Vz 18497
+IEJldHdlZW4= 18498
+UGE= 18499
+IG5hdw== 18500
+4Lit4LiH 18501
+IGJyZWVk 18502
+aWNodGU= 18503
+INCe0L3QuA== 18504
+IEJ1aWxkaW5n 18505
+IGNvbmZvcm0= 18506
+TU8= 18507
+INCW 18508
+IEtpZA== 18509
+bmFz 18510
+IER1ZQ== 18511
+csOpcw== 18512
+IGRpb3g= 18513
+IEJpbg== 18514
+IHRheGk= 18515
+IHNhcA== 18516
+IEh1Yg== 18517
+IGNlbnRlcmVk 18518
+IHN1cmdl 18519
+IGF2b25z 18520
+IGxlYXJudA== 18521
+IFlhbQ== 18522
+IERpZXNl 18523
+0L3QuNC60Lg= 18524
+IEJlaWo= 18525
+V2lsbA== 18526
+IGF0dGVtcHRlZA== 18527
+IGdyaWVm 18528
+w7Nq 18529
+IGtpZG5leQ== 18530
+IG9wcG9uZW50cw== 18531
+IG5vbWU= 18532
+NTc= 18533
+0Y/RgtC90L4= 18534
+IG1pZG5pZ2h0 18535
+QW5ub3VuY2Vy 18536
+YWNpdHk= 18537
+b25lZA== 18538
+IHB1ZWRlcw== 18539
+IHByb2JsZW1hdGlj 18540
+IGNvcHM= 18541
+IFBldGU= 18542
+cmludA== 18543
+dW50ZWQ= 18544
+IGJpcA== 18545
+5qI= 18546
+IMOA 18547
+IGNlbnM= 18548
+YXRpdmVseQ== 18549
+IHVyZ2VudA== 18550
+IHN0cnVnZ2xlZA== 18551
+YWNodXM= 18552
+IG1pY3Jvd2F2ZQ== 18553
+IFNpZGU= 18554
+IERlbm4= 18555
+INGP0LI= 18556
+IHVyZ2U= 18557
+IGZvcmNpbmc= 18558
+d2FuZw== 18559
+INC60L7RgtC+0YDQsNGP 18560
+IG1hbW0= 18561
+IPCfjg== 18562
+IHRyaWJlcw== 18563
+IFNoYWRvdw== 18564
+IFNhbmc= 18565
+IEhpdGxlcg== 18566
+IGx1bg== 18567
+IHNjZW50 18568
+7KeR 18569
+IG92ZXJ3aGVsbWVk 18570
+IGJvbWJz 18571
+IGNyaW1pbg== 18572
+IGNvbnNvbGlk 18573
+IG1vbGVjdWxhcg== 18574
+15XXpw== 18575
+bm9y 18576
+IHBlcmNlaXZlZA== 18577
+IHbDqQ== 18578
+IGFsdG9nZXRoZXI= 18579
+IG9ydGg= 18580
+IHZlbQ== 18581
+IHp3YXI= 18582
+aXpv 18583
+xas= 18584
+IG1lbHRlZA== 18585
+b3JkZW4= 18586
+IENoYXJsb3R0ZQ== 18587
+IEV4Y2Vs 18588
+YXJ0YQ== 18589
+7Jyg 18590
+IEdldw== 18591
+IHJvbWFuY2U= 18592
+ZXJlbW9z 18593
+IGNvbG9uaWFs 18594
+IHRyYWRpdGlvbmFsbHk= 18595
+IHF1YW4= 18596
+aG9v 18597
+IGNoYW1waW9uc2hpcA== 18598
+IGFyYml0cg== 18599
+7IWU 18600
+INC80LjQvQ== 18601
+IHNlbGZpc2g= 18602
+IGJsZXc= 18603
+cnlpbmc= 18604
+IG9wZXJhdG9ycw== 18605
+IGp1cmlzZA== 18606
+j4U= 18607
+dWl0aW9u 18608
+IEVD 18609
+IEFueWJvZHk= 18610
+dmF0ZQ== 18611
+aWV0aWVz 18612
+IGFuYWx5c3Q= 18613
+tOyXkA== 18614
+INCy0YHQtdCz0LTQsA== 18615
+w6dlaw== 18616
+IEt1bg== 18617
+IGFnaW5n 18618
+1aE= 18619
+0YDQsNGE 18620
+IE1vbWVudA== 18621
+IEh1YQ== 18622
+6IM= 18623
+dGhlbg== 18624
+0LXQu9Cw 18625
+ZXN0b25l 18626
+IGVuZGU= 18627
+IGF3YXJkZWQ= 18628
+IG7DpGNoc3Rlbg== 18629
+IFNwb3Q= 18630
+IE5lZw== 18631
+IGZhaXJ5 18632
+IENvdmVy 18633
+IGRlcG9zaXQ= 18634
+IHN0cmVzc2Z1bA== 18635
+IGp1bms= 18636
+IG1ldGFib2w= 18637
+SmE= 18638
+IOq3gA== 18639
+IHVuZGVyZ3JhZHVhdGU= 18640
+IGNhbmNlbGw= 18641
+IGNvbnNlbnN1cw== 18642
+IG9zbw== 18643
+4bq3 18644
+xJ9lcg== 18645
+cmFkYQ== 18646
+IFBhbGFjZQ== 18647
+IHBlZGFs 18648
+IGV4YWdnZXI= 18649
+IGJlaGF2aW9yYWw= 18650
+cGxheWVy 18651
+bGxlcw== 18652
+IGNvbm5lY3Rvcg== 18653
+IHNrZXB0 18654
+jZTrnbzqs6A= 18655
+IG1pdHQ= 18656
+IEhhaGE= 18657
+IHBlcXVl 18658
+IEdvdHQ= 18659
+ZmFuZw== 18660
+4LA= 18661
+am9z 18662
+IGtpY2tpbmc= 18663
+IG1vdW50ZWQ= 18664
+IHJlcGxhY2luZw== 18665
+dm9z 18666
+IHF1aWV0bHk= 18667
+IG1pbGl0 18668
+IG93bnM= 18669
+IG5pdmVhdQ== 18670
+IGF1cg== 18671
+IEJ1eQ== 18672
+IHByZWRpY3RlZA== 18673
+IGNvd3M= 18674
+IHBvbmVy 18675
+IERyaQ== 18676
+IHJlbWFya3M= 18677
+IHJlcG9ydGVy 18678
+IGFya2FkYcWf 18679
+0LXRgdGC0Lg= 18680
+IHNhdmVz 18681
+IMOnb2M= 18682
+IG1ldGFwaG9y 18683
+IEtlbA== 18684
+c3RhdGlvbg== 18685
+c2VtYmx5 18686
+IGFkdmlzb3I= 18687
+IHdvcmtzaG9wcw== 18688
+IGFjY291bnRpbmc= 18689
+IHRvaw== 18690
+bmllcg== 18691
+aW5uZXI= 18692
+IGJ1cmFkYQ== 18693
+IEJC 18694
+IE9seW1waWM= 18695
+IFByYWN0 18696
+Q2hyaXN0 18697
+INGB0Y4= 18698
+IGthcw== 18699
+IHZpZXdlZA== 18700
+IG1hcmtlcnM= 18701
+IGZvdG8= 18702
+Z2V0aWM= 18703
+IEx1Y2Fz 18704
+IHBhZHM= 18705
+IEpvaA== 18706
+IENEVQ== 18707
+YWZmZW4= 18708
+YXJlbQ== 18709
+IEJlY2s= 18710
+IEdvc2g= 18711
+c2hpdA== 18712
+IE1hdGVy 18713
+YWJ1bGFyeQ== 18714
+IFJvb20= 18715
+bGxlbg== 18716
+IEZvbGxvd2luZw== 18717
+IGRvaXQ= 18718
+YmFsbHM= 18719
+aXhh 18720
+IGdyb3VuZHM= 18721
+IOyeiOuKlOuNsA== 18722
+TFM= 18723
+IHdpbGRsaWZl 18724
+IFNRTA== 18725
+IHNoaWZ0cw== 18726
+Qm9vaw== 18727
+IGhvc3RlZA== 18728
+bGxvcg== 18729
+IHNuYXBz 18730
+IGJlc29pbg== 18731
+INep15Q= 18732
+IHBlYW51dA== 18733
+w6RmdA== 18734
+uaA= 18735
+xZts 18736
+QXVkaWVuY2U= 18737
+IEJhcmJhcmE= 18738
+IGFkb3B0aW9u 18739
+IHdvbGY= 18740
+INC+0YHQvdC+0LI= 18741
+YXJkYQ== 18742
+IGV4cG9zZQ== 18743
+IOym 18744
+amFz 18745
+xJM= 18746
+IGNvdW50bGVzcw== 18747
+IOyngQ== 18748
+aGVhbHRo 18749
+dWVudA== 18750
+aXNv 18751
+b3Rpb24= 18752
+IGh1bmdlcg== 18753
+IG1vaXM= 18754
+b2Zmcw== 18755
+IGNsYWltaW5n 18756
+IM6a 18757
+IEJlbGc= 18758
+INC90LDQuQ== 18759
+6riw64+E 18760
+IHVucHJl 18761
+IGdlZA== 18762
+IElv 18763
+INC/0L7RgdC80L7RgtGA 18764
+IGNvxZs= 18765
+IE5hcnJhdG9y 18766
+IMOHb2s= 18767
+7Zmp 18768
+4Lit4Lii 18769
+Y2lwbA== 18770
+IHRpbWVy 18771
+IGRlZmlj 18772
+YXZpbg== 18773
+IGNhdGVnb3I= 18774
+IHRocm93cw== 18775
+IOuCnA== 18776
+INC/0L7RgdC70LXQtA== 18777
+IFRoYWk= 18778
+IG1hc2N1bA== 18779
+IGJla29tbWVu 18780
+IGludGVybmF0aW9u 18781
+dWxzZQ== 18782
+IGF5ZQ== 18783
+IHBvaQ== 18784
+IHBpeGVs 18785
+Q2hyaXM= 18786
+IHN0b3Zl 18787
+zr/OuQ== 18788
+IGdlbmVyYXRvcg== 18789
+IOy7rOs= 18790
+IGFjYWRlbQ== 18791
+IHByYWN0aWNlZA== 18792
+IGFxdWVzdA== 18793
+IGNvbnRyaWJ1dGluZw== 18794
+IEln 18795
+IOG7nw== 18796
+IGNvbnRhaW5pbmc= 18797
+IHdyZXN0bGluZw== 18798
+INGH0LXQs9C+ 18799
+aGF1cHQ= 18800
+IGVzc2Fz 18801
+dmVsb3Bl 18802
+IGV4Y2VwdGlvbmFs 18803
+WVU= 18804
+IEFwcGxhdXNl 18805
+cmljYW5l 18806
+IGNvbnZlbmllbmNl 18807
+INC00LXQu9Cw0YLRjA== 18808
+0LjQu9C40YHRjA== 18809
+IEVudmlyb24= 18810
+ODU= 18811
+IGPDog== 18812
+IOyViOuFle2VmOyEuOyalA== 18813
+IE1P 18814
+IFBvcGU= 18815
+IHNhaA== 18816
+b2Jp 18817
+IG1hc3RlcnM= 18818
+YWluZXM= 18819
+IGJsZXNzaW5ncw== 18820
+IG9iZXk= 18821
+IGZsdXg= 18822
+IGJyb3c= 18823
+IOyLpA== 18824
+IHBvcHVsYXJpdHk= 18825
+IExhbWI= 18826
+emV1Zw== 18827
+7JmU 18828
+j4TroZ0= 18829
+aXR1YXRpb24= 18830
+IGFjY29tcGFu 18831
+IGRpYWxvZw== 18832
+IEphbWll 18833
+IHNld2luZw== 18834
+IGJsZWVkaW5n 18835
+IGJhaWw= 18836
+IHRocmVhZHM= 18837
+b2RnZQ== 18838
+IFNoYW5n 18839
+IGRlcGxveW1lbnQ= 18840
+Y2hlZA== 18841
+IHNhdGlzZnk= 18842
+IGxheg== 18843
+IG1pc3NpbGU= 18844
+IExpbmtlZA== 18845
+IG1ha2Vycw== 18846
+Y2l1bQ== 18847
+ZnJl 18848
+IOuovA== 18849
+IOustOs= 18850
+IEVkZ2U= 18851
+IHNvY2lldGllcw== 18852
+IGFndWE= 18853
+IHN5bmNocm9u 18854
+oaA= 18855
+dW5mdA== 18856
+IHVubQ== 18857
+IHRyaWFuZw== 18858
+IGluanVzdA== 18859
+dG9w 18860
+IG9yYWw= 18861
+a29y 18862
+IO2VqA== 18863
+bGRpZ3Q= 18864
+Y2XEnw== 18865
+cXVldA== 18866
+IExlbw== 18867
+IHNhdm9pcg== 18868
+IGVhc3Rlcm4= 18869
+aWV1 18870
+IGV4cGVk 18871
+INCh0L8= 18872
+IHVubmVjZXNzYXJ5 18873
+IFBlcmZvcm0= 18874
+IE1pbmc= 18875
+INGA0LDQsg== 18876
+IGludGVudGlvbnM= 18877
+IGNvbXByZXNzaW9u 18878
+IFNhYw== 18879
+zr/Ouw== 18880
+YXJzb24= 18881
+IHRyb3V2ZQ== 18882
+IE11aGFtbWFk 18883
+INCy0YvRgQ== 18884
+IGZpbml0ZQ== 18885
+INC90LDRhdC+0LQ= 18886
+dWdh 18887
+0YDQsNC30YM= 18888
+IGNlbGVicmF0ZWQ= 18889
+IGNvbmZlc3M= 18890
+IHNxdWFyZXM= 18891
+IEdvcmRvbg== 18892
+IOuCmOyY 18893
+IHN5bmRyb21l 18894
+IGNvbXBsZXRpb24= 18895
+IGJhY2tpbmc= 18896
+IGRhcmY= 18897
+IFF1cmFu 18898
+IGludGVybWVkaWF0ZQ== 18899
+IGtlcg== 18900
+IGTDvA== 18901
+aGVzaXZl 18902
+IGFjY291bnRhYmlsaXR5 18903
+IFJlYmVjY2E= 18904
+IFNsZWVw 18905
+IGRpZmbDqXJlbnQ= 18906
+b2xz 18907
+IFJpY2U= 18908
+IOuzuA== 18909
+IGFudGliaW90 18910
+z4TOrA== 18911
+cno= 18912
+YW1ibGluZw== 18913
+IHNlbnNpdGl2aXR5 18914
+IGNocm9u 18915
+YWxsYXM= 18916
+NjQ= 18917
+IGZsZWV0 18918
+IG9wdGltaXN0aWM= 18919
+0YHQutC+0LPQvg== 18920
+IGphZGk= 18921
+YWlsbGV1cnM= 18922
+IEVub3VnaA== 18923
+IHNlbmlu 18924
+IHBhY2tz 18925
+Ym4= 18926
+IEFyZWE= 18927
+IFRybw== 18928
+qOumrA== 18929
+0LDRlA== 18930
+IFRob20= 18931
+IGhhcm1vbnk= 18932
+0L3QuNC60LA= 18933
+IHNvbWVkYXk= 18934
+SVNF 18935
+IEJyb2Fkd2F5 18936
+bGFyZXM= 18937
+ZXJuZXNz 18938
+4LmE4Lih 18939
+IFRlbm4= 18940
+IE5BVE8= 18941
+IG1pbnV0b3M= 18942
+IEthbnNhcw== 18943
+IE1vbmc= 18944
+IGNvbXB0ZQ== 18945
+iqQ= 18946
+IOyXrQ== 18947
+IHN1cGVyaGVybw== 18948
+IEdhcmRlbg== 18949
+IE1vcw== 18950
+IGF0dGFjaG1lbnQ= 18951
+IGJ1c3Q= 18952
+4K+K 18953
+IFRoYWlsYW5k 18954
+c3RhdA== 18955
+IHNwaWNl 18956
+IExlYg== 18957
+IGxlYXA= 18958
+emVjaA== 18959
+R0w= 18960
+IHZlcmw= 18961
+IGZpeGluZw== 18962
+IOuztOuptA== 18963
+IHBvcm4= 18964
+IGLDvHk= 18965
+INmF2Kc= 18966
+IFZpcnQ= 18967
+IFRvbW15 18968
+IGNhcmdv 18969
+IE9saGE= 18970
+IHJva3U= 18971
+2YPZhg== 18972
+IGJha2Vk 18973
+IHRhY3RpY3M= 18974
+IG1hcmtldHBsYWNl 18975
+IGt0w7NyYQ== 18976
+YXJsbw== 18977
+IHN3aXRjaGVz 18978
+IGNhY2hl 18979
+IEhS 18980
+IEdhbg== 18981
+IEdQUw== 18982
+IGR1YXM= 18983
+aGVyZXM= 18984
+0LXRgNGI 18985
+dHJhY2s= 18986
+IGx1bmdz 18987
+U3RhdGlvbg== 18988
+aWdnbGVz 18989
+IGNhbXBpbmc= 18990
+IGNvbXBsZXRpbmc= 18991
+YW1hcw== 18992
+IGN5Y2w= 18993
+IHByb3RvdHlwZQ== 18994
+IEp1ZGdl 18995
+b3R5cGVz 18996
+IGluZmVjdGlvbnM= 18997
+oKTr 18998
+0LXRgNCz 18999
+b2Jh 19000
+IEJvZA== 19001
+IFNlY29uZGx5 19002
+IGFwb3N0 19003
+IHNvZ2Fy 19004
+IHJlYXNz 19005
+aWVr 19006
+IGFzaGFtZWQ= 19007
+IGN1cnZlcw== 19008
+INCy0LDQtg== 19009
+IGVuc2VtYmxl 19010
+YXR1cg== 19011
+IHBob3RvZ3JhcGhlcg== 19012
+IGVpZ2h0aA== 19013
+IHdhc3RlZA== 19014
+IGRhbXA= 19015
+INC80LDQuw== 19016
+YXJlbmE= 19017
+IGludGVybmFsbHk= 19018
+IGhlZWxz 19019
+IFNhbHQ= 19020
+IGJsaXI= 19021
+iOuCmA== 19022
+IGNvbnRyYXJ5 19023
+IHByaW1h 19024
+IG9zcw== 19025
+IHJhYmJpdA== 19026
+IGF1dG9y 19027
+IGJyb2FkbHk= 19028
+w61zdA== 19029
+IGJhY2tz 19030
+7ZSE 19031
+ZXRv 19032
+IGp1cnk= 19033
+6LE= 19034
+IHByb3N0dQ== 19035
+IGJhcmE= 19036
+IHBhcmxpYW1lbnQ= 19037
+b3JpZW50 19038
+0LjQu9Cw0YHRjA== 19039
+IGluZGlyZWN0 19040
+w6Ft 19041
+IMOlcg== 19042
+IHRyYWl0cw== 19043
+IGTDrWFz 19044
+2YTZhQ== 19045
+IENU 19046
+YWx5c3Q= 19047
+IGxpdmVzdA== 19048
+IGtvcw== 19049
+TWF5 19050
+IEppbmc= 19051
+IGpvdXJuYWxpc3Rz 19052
+0YfQuNC6 19053
+YXJtcw== 19054
+IOqwkOyCrA== 19055
+INC40LzQtQ== 19056
+IMOpZ2Fs 19057
+IE5ld3Rvbg== 19058
+IHJlY292ZXJlZA== 19059
+IGJyYXVjaGVu 19060
+IEJyb24= 19061
+0LDQvdC+ 19062
+IHBhbGU= 19063
+cHJpc2Vz 19064
+IGhvcmFz 19065
+Y2h0cw== 19066
+w7/Dvw== 19067
+YWtlcnM= 19068
+IEFsYXNrYQ== 19069
+emllag== 19070
+IHNjb29w 19071
+7J206rCA 19072
+Y29y 19073
+w6lsw6k= 19074
+IHN1cmc= 19075
+IHZpZW5l 19076
+IEtyaXN0 19077
+NTQ= 19078
+IGJhbm5lZA== 19079
+IHNtb290aGx5 19080
+IHRyZWF0cw== 19081
+IHByb25vdW5jZQ== 19082
+IGZsdXNo 19083
+IGNhbWJp 19084
+IG11c2ljaWFu 19085
+IEFzaGxleQ== 19086
+IFNQRA== 19087
+IEJvYmJ5 19088
+IGdsb3Nz 19089
+cmVzcGVjdA== 19090
+IHJldmlld2luZw== 19091
+IGdlbmVyaWM= 19092
+xrDhu5tj 19093
+YXRzw6RjaGxpY2g= 19094
+IGhlYWx0aGllcg== 19095
+dWJlcnM= 19096
+INC00LDQvQ== 19097
+IE1lZGljYXJl 19098
+NTM= 19099
+IGNvbXBsYWludHM= 19100
+amFj 19101
+IGFncmljdWx0dXJhbA== 19102
+U3Bl 19103
+IEpvbmc= 19104
+IGRpb3hpZGU= 19105
+6rKo 19106
+ZWxpams= 19107
+IFNoaXQ= 19108
+YWludHM= 19109
+IElhbg== 19110
+IFNpbXBseQ== 19111
+IFN0cmU= 19112
+IEdEUA== 19113
+NTk= 19114
+YXN6 19115
+IEthdGll 19116
+INCx0YA= 19117
+IHBlZWs= 19118
+b3d5Y2g= 19119
+IHJlc29ydA== 19120
+IHJlc2lkZW5jZQ== 19121
+IHNwaWNlcw== 19122
+Y2nDsw== 19123
+IGplZGVy 19124
+IGVtbw== 19125
+YXJpdW0= 19126
+IHB1ZmY= 19127
+66eJ 19128
+0YPQu9GM0YI= 19129
+IG1ldGE= 19130
+IOyghOs= 19131
+IG9wdGltaXphdGlvbg== 19132
+Z2FuZw== 19133
+IO2VhA== 19134
+IGVmZmljaWVudGx5 19135
+IHZpc3VhbGx5 19136
+IGZyb3N0 19137
+IEFydGh1cg== 19138
+IMW8 19139
+IGFjaGlldmluZw== 19140
+IHJvdGF0aW5n 19141
+IGxpbmluZw== 19142
+IG9jY3VwaWVk 19143
+bWVudGF0aW9u 19144
+IHN0cmV0Y2hpbmc= 19145
+IHN0YWxs 19146
+b3N0aWM= 19147
+IFNldmVy 19148
+IGdsdWM= 19149
+IHLDs8W8 19150
+IG91dHJlYWNo 19151
+c3RyYQ== 19152
+aWtlbg== 19153
+IOyWmOq4sA== 19154
+IEpvaW4= 19155
+IGltcGU= 19156
+IGNvbXBlbnNhdGlvbg== 19157
+IFRhdA== 19158
+IENhcmxvcw== 19159
+w7xocnQ= 19160
+IEZyYW5jaXM= 19161
+Y2pp 19162
+eWVhaA== 19163
+IG1lbWJyYW5l 19164
+IGV4aGFsZQ== 19165
+IHJlbGk= 19166
+IE9S 19167
+IHJlZnJpZ2VyYXRvcg== 19168
+IFZlbmV6 19169
+TGlrZQ== 19170
+IHJhaXNlcw== 19171
+b3R0bGU= 19172
+YXR1cmE= 19173
+IHJ1bGVy 19174
+IHdlZXI= 19175
+IGd1aWRlZA== 19176
+IE1hZ24= 19177
+IENvcnBvcg== 19178
+jZQ= 19179
+IGF0dHJpYnV0ZQ== 19180
+IFdvYWg= 19181
+IGFycm93cw== 19182
+IGF3YWl0 19183
+IFByaW0= 19184
+IGRpZ25pdHk= 19185
+IE9udGFyaW8= 19186
+aXNjaGVy 19187
+IOyLnQ== 19188
+aW1lbg== 19189
+b3V2ZXI= 19190
+QVNT 19191
+4buHbg== 19192
+b3B5 19193
+YWNodXNldHRz 19194
+IGVsZGVybHk= 19195
+RkE= 19196
+IERhaWx5 19197
+c2hpbmU= 19198
+IDU2 19199
+6KI= 19200
+aWVybm8= 19201
+IHNraWxsZWQ= 19202
+IGdyb8OfZQ== 19203
+IE9haw== 19204
+aWdnbGU= 19205
+0LXQu9C10Lk= 19206
+IGJpcmF6 19207
+IGFyZ3Vpbmc= 19208
+INC/0L7RjdGC0L7QvNGD 19209
+IGRyaWZ0 19210
+IGhhcm5lc3M= 19211
+IGRlaXhhcg== 19212
+YXV0cmU= 19213
+IFNlZWluZw== 19214
+IGNhcGl0YWxpc20= 19215
+IEVsZA== 19216
+emlvbmU= 19217
+IEJleW9uZA== 19218
+IHBlcmZlY3Rpb24= 19219
+IGhvZQ== 19220
+IGRlY2xhcmU= 19221
+0LDQu9Cw0YHRjA== 19222
+IHBva2U= 19223
+INeh 19224
+IGZpZ2h0ZXJz 19225
+6rKg64uk 19226
+0L7RgNC+0LI= 19227
+IGFjY29yZGluZ2x5 19228
+IElzYQ== 19229
+IG9wdGltaXpl 19230
+IE1pbmlzdHJ5 19231
+IHNhZ2U= 19232
+7Iuc66m0 19233
+IGJlbmk= 19234
+IGRvbmF0aW9u 19235
+IGNsZWFyZWQ= 19236
+IEx1Y2tpbHk= 19237
+IGhhcm1mdWw= 19238
+tey7pA== 19239
+IGNlbWVudA== 19240
+0L/QuNGB 19241
+IGRlZGk= 19242
+IENyYWln 19243
+IGRlbW9ucw== 19244
+IGN1c3RvbWl6ZQ== 19245
+IGlnbm9yZWQ= 19246
+IFRpYW4= 19247
+IGhvcGVk 19248
+IEJ1cmVhdQ== 19249
+IHJp 19250
+IFlhaA== 19251
+IHNvY2tldA== 19252
+IGZlYXR1cmluZw== 19253
+IHBhcmY= 19254
+IFRF 19255
+IFRlYWNoZXI= 19256
+IGNhdGFsb2c= 19257
+6rCA7KeA6rOg 19258
+IFNlaXRl 19259
+IGNvbmU= 19260
+IFBhbGVzdGlu 19261
+IGdld29vbg== 19262
+IGdhaW5pbmc= 19263
+INii 19264
+IGNhdGFzdA== 19265
+IG5laWdoYm91cg== 19266
+SVNU 19267
+IHN0ZWFsaW5n 19268
+IHRyb2lz 19269
+IGludGVuZA== 19270
+IFNob290 19271
+IHBpb25l 19272
+IEludGVs 19273
+IExJTg== 19274
+IGJyaWdodGVy 19275
+IFllc3RlcmRheQ== 19276
+IHNvdw== 19277
+c2lu 19278
+b2Rz 19279
+IGV0aGljcw== 19280
+IGludGVydmlld2Vk 19281
+cmVsbA== 19282
+IHJlZnJlc2hpbmc= 19283
+c8Ol 19284
+IGFic3VyZA== 19285
+IHBob3NwaA== 19286
+Zmls 19287
+IHN0ZWhlbg== 19288
+dmFscw== 19289
+IGNhcmVk 19290
+IGRlbGw= 19291
+Ym9uZQ== 19292
+IGhvY2g= 19293
+IHB1cA== 19294
+IGlv 19295
+IGFjb250ZWNl 19296
+ZWxsZXM= 19297
+IFNwbA== 19298
+aWdp 19299
+IHTDpG4= 19300
+IGVsZXBoYW50 19301
+IGdhdGVz 19302
+IHNsaWNlcw== 19303
+IHByYW5r 19304
+b2tyYXQ= 19305
+IGhpbGFyaW91cw== 19306
+IFNpZA== 19307
+IOuSpA== 19308
+IGVzc2VyZQ== 19309
+IHRlbGVwaG9uZQ== 19310
+aW5hbGx5 19311
+cmF0b3I= 19312
+IGhlbGljb3B0ZXI= 19313
+IGnFn3Rl 19314
+IGdpZA== 19315
+IHRvdXJpc3Q= 19316
+IGNvbmZsaWN0cw== 19317
+0LDRgtCw 19318
+IHTDqQ== 19319
+IGFzc2VydA== 19320
+IGxhdW5kcnk= 19321
+IEJvbQ== 19322
+IHNwZWNpYWxpemVk 19323
+IE1vZGVybg== 19324
+b2dyYWY= 19325
+IGFubw== 19326
+IHJldHJpZQ== 19327
+IFB1dGlu 19328
+IEhBUg== 19329
+INC80LDRiA== 19330
+IM6xz4DPjA== 19331
+IHR1dHRp 19332
+INCy0YLQvtGA 19333
+7Ja1 19334
+IEJ1bA== 19335
+64uk66m0 19336
+xYJl 19337
+YXJpbg== 19338
+IHRoZXJhcGlzdA== 19339
+IGfDpXI= 19340
+IEN6eQ== 19341
+cHBl 19342
+bWly 19343
+IFRlcm0= 19344
+IEJlYXI= 19345
+bGFjZQ== 19346
+IE1vcmVvdmVy 19347
+IERpc2M= 19348
+IO2DgA== 19349
+IHRpdGxlZA== 19350
+IHN0cmlwcw== 19351
+IEZhaHI= 19352
+IFJpbmc= 19353
+cmFuZG8= 19354
+YWZh 19355
+IHNob3J0cw== 19356
+IHRydW5r 19357
+IHNlbnRpZG8= 19358
+z4nOvQ== 19359
+IGFjcmVz 19360
+IG92ZXJk 19361
+IE9seW1waWNz 19362
+IE1lcmNp 19363
+IOuCmOyYpA== 19364
+IGdlcm0= 19365
+YW1tZWQ= 19366
+IHByZWd1bnQ= 19367
+IE51dA== 19368
+IDwv 19369
+IHRyYXZlbHM= 19370
+IHZvY2FidWxhcnk= 19371
+ZXRlbg== 19372
+b2Rlcg== 19373
+IGNvbnN1bWluZw== 19374
+d3JpdGluZw== 19375
+IGFwcGVhcmluZw== 19376
+IGFkanVzdGVk 19377
+c2Vt 19378
+IGZyZW50ZQ== 19379
+IG1heGltaXpl 19380
+IHp3aXNjaGVu 19381
+IHphbQ== 19382
+Y29uc2Npb3Vz 19383
+emVr 19384
+aGFv 19385
+7LKY65+8 19386
+IEVwaXNvZGU= 19387
+IHZpc2liaWxpdHk= 19388
+IG1pam4= 19389
+IHZpZWxlbg== 19390
+IEJyb3RoZXJz 19391
+15nXkQ== 19392
+IHbDpGxkaWd0 19393
+IGNydXNoZWQ= 19394
+dWZlbg== 19395
+YWN0aWM= 19396
+IEJlZA== 19397
+IEZB 19398
+aXNzaXBwaQ== 19399
+IHJlbW90 19400
+IHBldHM= 19401
+IHRodW5kZXI= 19402
+IE1hbQ== 19403
+7JW17Luk 19404
+cGFyZW50cw== 19405
+IGLEsQ== 19406
+IHN1cnRvdXQ= 19407
+IHNlZ21lbnRz 19408
+IG5laG1lbg== 19409
+IHV0aWxpeg== 19410
+IFJ1Ynk= 19411
+IHLhu5Np 19412
+IGhhcHBpbHk= 19413
+IGJ1c2g= 19414
+dWx0YW4= 19415
+2Lg= 19416
+IEhpbA== 19417
+IGxhd24= 19418
+IGV5ZWJyb3dz 19419
+bWV6 19420
+IFN5ZA== 19421
+cmVw 19422
+aW5m 19423
+IG92ZXJoZWFk 19424
+Y3puaWU= 19425
+IG94aWQ= 19426
+IFdvbA== 19427
+IGRlc3Ryb3lpbmc= 19428
+IEFkZGl0aW9uYWxseQ== 19429
+dW1ibGVk 19430
+ZGVw 19431
+IGRlcG9z 19432
+IGNvbW1vZA== 19433
+IGNha2Vz 19434
+IHRhbGVudHM= 19435
+IHBvdXJxdW9p 19436
+IGNvbnRlbXBs 19437
+bmVscw== 19438
+0L7RiQ== 19439
+IEFyYWJpYw== 19440
+IE1hcnlsYW5k 19441
+b3dv 19442
+IFBsYQ== 19443
+xJ9sdW0= 19444
+IHByb3BoZQ== 19445
+IFJlcHJlc2VudA== 19446
+b3BvbA== 19447
+YWNjb3Jk 19448
+IE1lYW5pbmc= 19449
+IGpvaW50cw== 19450
+IGJyYWtlcw== 19451
+Y2t0 19452
+IDE5OTk= 19453
+IHB1YmxpY2F0aW9u 19454
+IFJldmlldw== 19455
+0L7QudC0 19456
+IG5pY2hl 19457
+IHNpZ25pZmljYQ== 19458
+IGRlYnI= 19459
+IG92ZXJsYXA= 19460
+IGRlbWFuZGluZw== 19461
+IFPDsw== 19462
+IHN1YnNlcXVlbnQ= 19463
+IHF1b3Rlcw== 19464
+IEN1cnJlbnRseQ== 19465
+IHByZXZlbnRpbmc= 19466
+IDEzMA== 19467
+IENlbA== 19468
+b25u 19469
+d25pZcW8 19470
+7JW9 19471
+INC60LDQutC40LU= 19472
+QUNI 19473
+IGd1bQ== 19474
+IElzcmFlbGk= 19475
+7Jy864uI6rmM 19476
+5ag= 19477
+cnVrdA== 19478
+IGNsYXBwaW5n 19479
+IE1hc3NhY2h1c2V0dHM= 19480
+IHJlc2lsaWVuY2U= 19481
+IHN1YnNjcmliaW5n 19482
+IGpld2Vscnk= 19483
+Z2VicmE= 19484
+IGNvcnJlY3Rpb24= 19485
+Ym9v 19486
+2KY= 19487
+bGlv 19488
+c2Ft 19489
+IGVudmVsb3Bl 19490
+a2Fs 19491
+IEZhcm0= 19492
+IGNhdHRsZQ== 19493
+IGJyYXM= 19494
+IHJlcGVudA== 19495
+IHRvbmVz 19496
+b3Npb24= 19497
+cGVjdGlvbg== 19498
+IGRlbmVu 19499
+yJtp 19500
+IE1hcmc= 19501
+IGFjcXVpcmU= 19502
+aWJsaW5ncw== 19503
+IGFzcGly 19504
+IHNpemVk 19505
+IGFsYw== 19506
+IHZpYnJhdGlvbg== 19507
+dGls 19508
+ZW1pbg== 19509
+IGNvcnJlbGF0aW9u 19510
+IHNpbmd1bGFy 19511
+INC/0L7Rj9Cy 19512
+cmVr 19513
+IGNoYXB0ZXJz 19514
+bWJyZQ== 19515
+IGF1ZGl0aW9u 19516
+w6dhcw== 19517
+IHZhbXA= 19518
+IHRlcw== 19519
+INGA0LDQt9Cy 19520
+IHJlc3BlY3RlZA== 19521
+Y2lu 19522
+IGZ1Y2tpbg== 19523
+IMO8YmVyaGF1cHQ= 19524
+INC/0L7QsQ== 19525
+IGFsaWtl 19526
+tog= 19527
+cm9iaQ== 19528
+w650 19529
+IFRvdWNo 19530
+YW56YQ== 19531
+IGZpcm1seQ== 19532
+IEdyZWV0aW5ncw== 19533
+c2NhbGU= 19534
+ZGFk 19535
+0LDQutGC0Lg= 19536
+IGJhY2t5YXJk 19537
+0L7QttC0 19538
+R3I= 19539
+IFNURQ== 19540
+0L7RgNGC 19541
+IGjDpHR0ZQ== 19542
+IEZpcnN0bHk= 19543
+IE9mdGVu 19544
+YXN1cmVz 19545
+IGRyYXdz 19546
+cmVkaXQ= 19547
+QVRF 19548
+UGU= 19549
+Q1A= 19550
+IGNvbXBlbGxpbmc= 19551
+IHN1YnNpZA== 19552
+IG5laWdoYm9yaG9vZHM= 19553
+IGRpcGxvbQ== 19554
+IGVudGVuZGVy 19555
+cGVyaW5n 19556
+YXVn 19557
+Y2hhdA== 19558
+0J3Rgw== 19559
+IERvbGw= 19560
+IOygkA== 19561
+IGhvc2U= 19562
+bmFy 19563
+IHJld2FyZGluZw== 19564
+IFNvbGQ= 19565
+IHRha2k= 19566
+IGJsYWRlcw== 19567
+IEthdGg= 19568
+IGpvZ28= 19569
+IHNlbnNhdGlvbg== 19570
+dWFuYQ== 19571
+cGVs 19572
+IFJlY2VudGx5 19573
+IHBvbHltZXI= 19574
+IFVQ 19575
+LS0t 19576
+IGhvdmVy 19577
+IHJ1bGVk 19578
+INeU15DX 19579
+IGFmZmVjdGlvbg== 19580
+IMSR4buD 19581
+IGJyZWU= 19582
+IExheQ== 19583
+IFlvbmc= 19584
+IHJlY2VpdmVy 19585
+nOulvA== 19586
+IGRpc3Nv 19587
+IFFpbmc= 19588
+IMOpdg== 19589
+IG3DunNpY2E= 19590
+IGFlc3RoZXRpYw== 19591
+IEJyZWF0 19592
+IFRB 19593
+IGFjY3VyYXRlbHk= 19594
+P+KAiw== 19595
+IHdhZ2Vz 19596
+cmF3ZMSZ 19597
+IHN3YWxsb3c= 19598
+IGNvbXBsYWludA== 19599
+IGxpZWQ= 19600
+YmVjdWU= 19601
+IHJlbGF4aW5n 19602
+IFBva8OpbW9u 19603
+IHRlY24= 19604
+YmFuZw== 19605
+s7Ts 19606
+IHF1aWVu 19607
+0L3QvtC80YM= 19608
+IGhhYml0YXQ= 19609
+Li4uLi4u 19610
+YWJsaW5n 19611
+INGC0LDQutC40LU= 19612
+IGJlc29uZA== 19613
+IGVtcGxveWVk 19614
+IGFycml2ZXM= 19615
+IHZlc3NlbHM= 19616
+IEF4 19617
+IGRpc3BsYXlz 19618
+MTUw 19619
+b2xvZ2ll 19620
+IOyXkA== 19621
+IGNsbw== 19622
+INC00L7Qsg== 19623
+INCe0LQ= 19624
+IHZ1ZWw= 19625
+d2VuZA== 19626
+IHNsaXBw 19627
+dXJw 19628
+IExvdA== 19629
+IGJ1bGxldHM= 19630
+IHJhZ2U= 19631
+IHNraXJ0 19632
+aWVudGVz 19633
+IG5o4buvbmc= 19634
+IE5hdHVyYWw= 19635
+IGhpbmQ= 19636
+IHdvcmtsb2Fk 19637
+bXU= 19638
+7YOc 19639
+IHN1bnNldA== 19640
+0LLQvtC7 19641
+cGl0 19642
+IEFTSA== 19643
+IOu2hOuTpA== 19644
+IGRvd25zdGFpcnM= 19645
+6a0= 19646
+IGNvdW50ZWQ= 19647
+IG5heg== 19648
+15XXpA== 19649
+IFBoaWxpcHBpbmVz 19650
+IDExMA== 19651
+IFBhcmtlcg== 19652
+IGdpdHU= 19653
+IGludGVyZXM= 19654
+IHVtYnJl 19655
+IE5hdHVyZQ== 19656
+IGplcg== 19657
+ZW5vcw== 19658
+IHBhbmVsaXN0cw== 19659
+IGNvYXRpbmc= 19660
+IGNoZXJyeQ== 19661
+IFBlbnQ= 19662
+IE1pc3Q= 19663
+cmVnYXRpb24= 19664
+IHZpbmQ= 19665
+IENvcnBz 19666
+IE1pc3Npb24= 19667
+IG5vYmxl 19668
+IGZvbmN0aW9u 19669
+IHdhcnJpb3I= 19670
+IHByb3Rlc3Rz 19671
+b3VyaQ== 19672
+IGNvbnN0aXR1dGlvbmFs 19673
+xYJhbQ== 19674
+IGVtZXJnZWQ= 19675
+IGR5ZQ== 19676
+IFRyeWluZw== 19677
+aWdt 19678
+w6lxdQ== 19679
+TE8= 19680
+IFZlcm0= 19681
+ZXJ2aW5n 19682
+IFRJTQ== 19683
+IENp 19684
+IGZyZWV6ZXI= 19685
+IGdydXBv 19686
+IFNwb3J0cw== 19687
+INC/0YDQvtCz 19688
+INmE2Kc= 19689
+b3RoZXJhcA== 19690
+aWZmYW55 19691
+Ymlhbg== 19692
+IHJhbmtlZA== 19693
+IHByb3Bvc2Fscw== 19694
+IMSRw6J5 19695
+IGZyZWV6aW5n 19696
+IGluc2VjdHM= 19697
+dmls 19698
+IGNvbXBvc3Q= 19699
+IHNlbWFuYQ== 19700
+IGRpc3Rpbmd1aXNo 19701
+IGZhY2lsaXRhdGU= 19702
+IHBsdXNpZXVycw== 19703
+IHZlcmc= 19704
+IGFsZ3Vucw== 19705
+IFRpa1Rvaw== 19706
+IEV4cHJlc3M= 19707
+0LzQtdC90YI= 19708
+U1U= 19709
+IGludGltYXRl 19710
+IEF1dGhvcg== 19711
+IHdpdG5lc3Nlcw== 19712
+IGthbGF1 19713
+IGFyZ3VlZA== 19714
+IGF2b2lkaW5n 19715
+Y3RpdmU= 19716
+IHB1cnN1aW5n 19717
+IHN5bGw= 19718
+w6F2ZWw= 19719
+IEF0bGFudGE= 19720
+IFV0YWg= 19721
+IFRpbGw= 19722
+IGVyZg== 19723
+IDIwMjI= 19724
+w6R0ZXI= 19725
+IGZ1bmVyYWw= 19726
+IEZsYXNo 19727
+IEF0bGFudGlj 19728
+IGdlbGU= 19729
+7KaI 19730
+IG1vcnRnYWdl 19731
+IOuEmA== 19732
+bGljaHQ= 19733
+IGFtYml0aW91cw== 19734
+IEJlaWppbmc= 19735
+IGRpdmluZw== 19736
+IHVuYm94 19737
+aWxsYXM= 19738
+IG90cmFz 19739
+IGV2YWM= 19740
+IG1hcmluZQ== 19741
+INGB0L7Qt9C0 19742
+IENyZWF0ZQ== 19743
+IGdq 19744
+IGZyZXF1ZW5jaWVz 19745
+aW5ndG9u 19746
+IFJvbWFucw== 19747
+IGFpbWluZw== 19748
+IEJ1ZmY= 19749
+IGVtcGVyb3I= 19750
+IE1vaQ== 19751
+IHByb21pc2luZw== 19752
+IGFsZ3VtYQ== 19753
+IHBhc2E= 19754
+IGRpc29yZGVycw== 19755
+U0k= 19756
+IHN1Y2NlZWRlZA== 19757
+IGN1ZXJwbw== 19758
+IHNvZGl1bQ== 19759
+IHN0dWI= 19760
+aGVpcm8= 19761
+IGRlbGF5ZWQ= 19762
+ZXRlcmE= 19763
+dHc= 19764
+IHN5bmM= 19765
+aGQ= 19766
+IHRvdXJpc3Rz 19767
+IHN5c3Q= 19768
+IG3DqXQ= 19769
+IHF1YWxpZnk= 19770
+IE90aGVycw== 19771
+bGxlcnM= 19772
+0LDRgtC10LvRjNC90L4= 19773
+INCe0L3QsA== 19774
+IHBlcmNlaXZl 19775
+IOqygA== 19776
+IOqwgOyepQ== 19777
+INC40YHQug== 19778
+IE1hdHRlcg== 19779
+IEJsdWV0b290aA== 19780
+IHBlYXJs 19781
+IGFyaXNl 19782
+IG1vbnVtZW50 19783
+INC40LzQtdC90L3Qvg== 19784
+YWdp 19785
+2YTZig== 19786
+IHJobw== 19787
+IHNtYXJ0ZXI= 19788
+IGNvbmo= 19789
+0L7QutCw 19790
+IGtlZW4= 19791
+IFRyZWF0 19792
+0LrQu9GO0Yc= 19793
+IHBhY2tldA== 19794
+ZWxzaXVz 19795
+IEFsYWI= 19796
+0LjQvdC4 19797
+IHBzaQ== 19798
+IGVuam95YWJsZQ== 19799
+IEVsbGVu 19800
+INCy0Lw= 19801
+IGVsaW1pbmF0ZWQ= 19802
+IFJvdw== 19803
+IHpvbWJpZQ== 19804
+IEt1 19805
+IHBocmFzZXM= 19806
+IGdyZW4= 19807
+dXRlcg== 19808
+IGRpcmVrdA== 19809
+15Y= 19810
+ZW5lbg== 19811
+dXNh 19812
+INGB0LvQvtCy 19813
+xLA= 19814
+IEdo 19815
+IGNvcnJpZA== 19816
+IHF1ZWVy 19817
+IExpbmRh 19818
+IG9uYQ== 19819
+IG9ibGlnYXRpb24= 19820
+ZGFy 19821
+INi1 19822
+ZW1tZW50 19823
+YWNpZXM= 19824
+IHNjcmV3ZWQ= 19825
+IG5haw== 19826
+IGF5dWQ= 19827
+w6Fy 19828
+bGV6 19829
+IGRyb3du 19830
+IE1lZGljaW5l 19831
+IGxhYnM= 19832
+IGp1c3F1 19833
+IEdvbm5h 19834
+IHRlcnJvcmlzdA== 19835
+cXVlc3Q= 19836
+IGZhcnRoZXI= 19837
+IHJlcGxpZWQ= 19838
+IFNX 19839
+IE1pc3Npc3NpcHBp 19840
+aXNobmE= 19841
+IGhvbGRlcg== 19842
+IHJlaWdu 19843
+IGFjY2VwdGFuY2U= 19844
+IHVs 19845
+tow= 19846
+IEhvdGVs 19847
+IENvb3Blcg== 19848
+dGFu 19849
+IEdyYWI= 19850
+IHZhcG9y 19851
+IGFjdGVk 19852
+IEthbmc= 19853
+ZmFu 19854
+IOydtOyDgQ== 19855
+dXRldA== 19856
+IHdvcmR0 19857
+IGZhcm1z 19858
+ZGF0 19859
+IGNvdXBsZXM= 19860
+IGJlYWRz 19861
+aWVudG9z 19862
+VGhlbg== 19863
+b3NpdHk= 19864
+IFN0YW5mb3Jk 19865
+Li0= 19866
+V2FpdA== 19867
+IGRhdGFz 19868
+b2lyZQ== 19869
+IGhhc2h0YWc= 19870
+aW1tZQ== 19871
+IGVuY291bnRlcmVk 19872
+IHNob3V0aW5n 19873
+IHJlc2lzdGFudA== 19874
+IFNldW5n 19875
+IHRyYWdpYw== 19876
+IERyYXc= 19877
+LCw= 19878
+IHNob3djYXNl 19879
+IEFG 19880
+IFN0cmk= 19881
+IGJhY2tlZA== 19882
+INGD0LM= 19883
+INCx0YPQtNGD0YI= 19884
+IENvbGU= 19885
+ZXVycw== 19886
+KD8p 19887
+IGVzY2FwZWQ= 19888
+QVNU 19889
+IEFzc2VtYmx5 19890
+IHN0aWNrZXI= 19891
+IG1pZXV4 19892
+IGVudGVydGFpbmluZw== 19893
+IERPTg== 19894
+IEFtZW5k 19895
+IEthcmw= 19896
+IGluaGli 19897
+c3N0 19898
+aWVn 19899
+fn5+ 19900
+IGhvb2tlZA== 19901
+IGxpdGVyYWw= 19902
+IHN1bm55 19903
+c3RlcHM= 19904
+IOuwnOs= 19905
+IE1hcmluZQ== 19906
+IHN1ZQ== 19907
+IHByaXNvbmVycw== 19908
+IEVi 19909
+NTg= 19910
+IGRydW1z 19911
+IGd1aWx0 19912
+YWxn 19913
+IGhhcHBpZXI= 19914
+IENN 19915
+IOyVhOuLiOyVvA== 19916
+INCf0LXRgA== 19917
+0YPQu9GP 19918
+IGtleXdvcmQ= 19919
+IFBhcmNl 19920
+IEZvcmVpZ24= 19921
+IEFtYW5kYQ== 19922
+IOuqqQ== 19923
+cGxlc3M= 19924
+iKw= 19925
+w7Ntbw== 19926
+IHF1YWxxdWVy 19927
+7J2065286rOg 19928
+IGNvbnNwaXJhY3k= 19929
+IHN0cmF3YmVycnk= 19930
+IGhhdHRlbg== 19931
+RXM= 19932
+IHNwb3M= 19933
+IHZpbGxhZ2Vz 19934
+IGxldg== 19935
+INGB0YDQtdC0 19936
+IHdha2luZw== 19937
+IGNhbGN1bGF0aW9ucw== 19938
+INmF2Lk= 19939
+IHBvdXJpbmc= 19940
+IGxlYmlo 19941
+IHBvbGlzaA== 19942
+IFRvdXQ= 19943
+IGZ1bmt0aW9u 19944
+0LzQvg== 19945
+IFRp 19946
+IHdhc3Rpbmc= 19947
+aXN0aWNhbGx5 19948
+IG1hbmlwdWxhdGU= 19949
+IHNpbXBsaWZ5 19950
+IHRlYW1tYXRlcw== 19951
+INCx0L4= 19952
+IGNvbnRhbQ== 19953
+IFF1aXRl 19954
+IGt1cno= 19955
+IENhbmQ= 19956
+dHlwZQ== 19957
+b3V0aGVhc3Q= 19958
+IGZpbmFuY2lhbGx5 19959
+0L7Qu9C9 19960
+ZWxzb24= 19961
+IGZvcmVoZWFk 19962
+dWFnZQ== 19963
+bmF1ZGlibGU= 19964
+IEJlaGluZA== 19965
+IG5lZ290aWF0aW9ucw== 19966
+IOuniOydjA== 19967
+IGFsdGVybmF0aXZlcw== 19968
+cmFuaw== 19969
+aG9sZGVy 19970
+IGhlYWxlZA== 19971
+0YLQvtGH 19972
+IFNwZWM= 19973
+IGV4aGliaXQ= 19974
+IHNoYWxsb3c= 19975
+IGdvYg== 19976
+IOuc 19977
+IGZydXN0cmF0aW9u 19978
+w61v 19979
+IG1lbHRpbmc= 19980
+IFN0b3Jt 19981
+IHBhdGVudA== 19982
+IEJhcmNlbA== 19983
+IHBlZGVzdA== 19984
+2YjZhQ== 19985
+IHRhaQ== 19986
+IE1vZGU= 19987
+IHdpbA== 19988
+IOuqqOultA== 19989
+IMOpZ2FsZW1lbnQ= 19990
+INeQ15c= 19991
+YXlhbg== 19992
+IGFtYXplZA== 19993
+7KeA64qU 19994
+IGhhY2llbmRv 19995
+IOydtOyVvA== 19996
+zrvOsQ== 19997
+4LiC 19998
+0LXRgtCw 19999
+IGV4YW1z 20000
+IHRyYXZlbGxpbmc= 20001
+UHJlc3M= 20002
+0LjRgNGD 20003
+IGJhc2VsaW5l 20004
+IGJ1c2Vz 20005
+IHJlaW5mb3I= 20006
+dmVuYW50 20007
+IFRydXRo 20008
+nb0= 20009
+b2Jl 20010
+IHllbGw= 20011
+IHNhdXNhZ2U= 20012
+VEY= 20013
+IEV2aWw= 20014
+IG1laW5lcg== 20015
+15nXpw== 20016
+IGhvcGVmdWw= 20017
+IHLDs3duaWXFvA== 20018
+IFBlcsOy 20019
+dHdv 20020
+bmRlcg== 20021
+INC80LjRgA== 20022
+IGNvbnNjaWVuY2U= 20023
+IFdhcnJlbg== 20024
+aWNreQ== 20025
+IGFpbWVk 20026
+IGfDtnJh 20027
+WFQ= 20028
+IHB5cmFt 20029
+UmVk 20030
+YXR1 20031
+IEVzdGE= 20032
+IGVhcm5pbmdz 20033
+IGhhdHM= 20034
+IFN0YWR0 20035
+aWNrZXQ= 20036
+cG9pbnRz 20037
+aW5hbmRlcg== 20038
+IG1vdG9yY3ljbGU= 20039
+IOuPjA== 20040
+IO2VtOyVvA== 20041
+a29t 20042
+IERpbmc= 20043
+5pI= 20044
+IHJlY3Vycw== 20045
+IGVzdGltYXRlcw== 20046
+IGRlcm5p 20047
+IHZlcnNjaA== 20048
+IE1JQw== 20049
+0LjQstCw0YLRjA== 20050
+INC/0YDQvtGI 20051
+IGRvc3Q= 20052
+INCy0YHRgtGA 20053
+IHdpZWw= 20054
+IHNpYmxpbmdz 20055
+INC00LXQsg== 20056
+IGVhcmxpZXN0 20057
+IGZhdGlndWU= 20058
+IG5oaQ== 20059
+IGd1c3Rh 20060
+IGJvbm5l 20061
+ZnJvbQ== 20062
+IEplbm55 20063
+IHN1cHBvc2VkbHk= 20064
+aW50YWdl 20065
+IGNvdW50aWVz 20066
+IHVucmU= 20067
+IHBsYW50aW5n 20068
+IEdyYWM= 20069
+IEdlbmVzaXM= 20070
+IEFscGhh 20071
+eXN6 20072
+IHRpbGU= 20073
+IOqyveyasA== 20074
+INeZ16k= 20075
+cXVlbA== 20076
+IGRpc3RyaWJ1dGU= 20077
+ZGVm 20078
+w6lyYWw= 20079
+IGNsdXRjaA== 20080
+YWRlbHBo 20081
+IFBsYXlTdGF0aW9u 20082
+hLg= 20083
+IHNq 20084
+YnJlYWtpbmc= 20085
+IOuQmOs= 20086
+IEN1YmE= 20087
+IFJ1c3NpYW5z 20088
+IE1BUks= 20089
+IHBlcnNl 20090
+IHJlc3RyaWN0ZWQ= 20091
+aWdlcw== 20092
+IFRyYXZlbA== 20093
+IGVsZWN0cm9uaWNz 20094
+IHF1YXJ0ZXJz 20095
+IEtlaXRo 20096
+c2l6ZWQ= 20097
+IGRlYWRsaW5l 20098
+YXJlbnRo 20099
+IHbDrWRlb3M= 20100
+IHByb3RvY29scw== 20101
+YW1tZW50 20102
+IFRyYWluaW5n 20103
+IMOi 20104
+IHNlcXVlbA== 20105
+0L3QsNC6 20106
+IGtlaW5lbg== 20107
+IG1hdHRyZXNz 20108
+bHVkaW5n 20109
+IGNsYXNzaWZpZWQ= 20110
+IHJlYWN0b3I= 20111
+IEtvbnQ= 20112
+IHBhc3Nhcg== 20113
+IGhvbm91cg== 20114
+b3JpZw== 20115
+SU5B 20116
+IE5hdGhhbg== 20117
+0LLQsA== 20118
+INGB0LrQsNC30LDRgtGM 20119
+dMSxcg== 20120
+IGV4Y2x1c2l2ZWx5 20121
+IHNoYWRlcw== 20122
+INC/0YDQvtGG 20123
+IG9jY2FzaW9ucw== 20124
+aWph 20125
+Zmln 20126
+IHR1cw== 20127
+IHJlbWVt 20128
+IENocmlzdG9waGVy 20129
+IHNsaW1l 20130
+IGFsZ3VuYQ== 20131
+IEZvcnR1bmF0ZWx5 20132
+IGxvcnM= 20133
+dm9sbA== 20134
+YXZlcg== 20135
+IG91dGxldA== 20136
+IExpbmtlZElu 20137
+IEV4ZWN1dGl2ZQ== 20138
+IG9yZ2Fucw== 20139
+IEJlZ2lu 20140
+IO2ZlA== 20141
+IHRyYW5zcGxhbnQ= 20142
+cmFnZW4= 20143
+Vk8= 20144
+IEbDtnI= 20145
+INio2KfZhA== 20146
+IEFuZHJl 20147
+aXNpbmU= 20148
+IGxhc3Rz 20149
+IGhpc3TDs3JpYQ== 20150
+IGx1eg== 20151
+IGNvbGxhcg== 20152
+IGtpZG5h 20153
+IG9wdGljYWw= 20154
+aW92 20155
+IHRvYg== 20156
+IGV4dGVyaW9y 20157
+IG1ldHJpYw== 20158
+aWV1cg== 20159
+IHRyb2xs 20160
+INGA0L7Qtw== 20161
+IHTDtA== 20162
+IOyYiOyB 20163
+IEdlc2V0eg== 20164
+INC10LQ= 20165
+IGRlbm9taW5hdG9y 20166
+7LM= 20167
+IGxldHQ= 20168
+IGdyw7bDnw== 20169
+IEx1dGhlcg== 20170
+IHJlc3Rl 20171
+IHJlc2VtYg== 20172
+IHBlcm1ldA== 20173
+a3Np 20174
+IGZpc2hlcg== 20175
+IFZvbg== 20176
+7ZS8 20177
+IM+Dz4TOvw== 20178
+IGxvY2tz 20179
+IHNob290cw== 20180
+IGthbXU= 20181
+IEtlcg== 20182
+IE9icw== 20183
+IGJpbGk= 20184
+IOuwsQ== 20185
+IHRvcnR1cmU= 20186
+YXNzeQ== 20187
+INC40LM= 20188
+IGxhc3Rpbmc= 20189
+IHRpZW5lcw== 20190
+IHJlY2VpdmVz 20191
+IE9zY2Fy 20192
+IHJlbWVtYmVyaW5n 20193
+IHByb2JsZW1hcw== 20194
+IGlh 20195
+IG1lbW9yYWJsZQ== 20196
+IGpvdXJz 20197
+IGZhw6dvbg== 20198
+YW1pYw== 20199
+IOu0pA== 20200
+YXRpcXVl 20201
+IOutlOqwgA== 20202
+IHppcA== 20203
+aGFsdA== 20204
+IPCfmA== 20205
+IGZyaWVz 20206
+IGZpbmRlbg== 20207
+Z3Jh 20208
+0YDRg9C0 20209
+aW1wb3J0 20210
+IOuLrOs= 20211
+IGlraQ== 20212
+IGNvbXBsYWluaW5n 20213
+IGZhemVuZG8= 20214
+IGdvb2dsZQ== 20215
+IHRhYnM= 20216
+IOuTpOyWtOw= 20217
+dWdv 20218
+aWVydG8= 20219
+YXVmZW4= 20220
+IOuovOyggA== 20221
+IHNrdWxsZQ== 20222
+IHN1aXY= 20223
+IHNweQ== 20224
+IEthaQ== 20225
+IG1hcnRpYWw= 20226
+IG9uZGVy 20227
+YXRpbGl0eQ== 20228
+IGlyZ2VuZHdpZQ== 20229
+IGNsYXA= 20230
+aW50ZWxs 20231
+IGluc3RhbGxpbmc= 20232
+IHVuaXF1 20233
+IENlbnRyZQ== 20234
+YXN0cw== 20235
+dWFy 20236
+IHJldmlz 20237
+IHRocmVhdGVuaW5n 20238
+cmFpcw== 20239
+IGN1aWQ= 20240
+c2th 20241
+IHJlc29sdmVk 20242
+IHJpZGVz 20243
+IGZhaWx1cmVz 20244
+IHNlbWI= 20245
+IG1hbGVz 20246
+VUZG 20247
+IHRyw6pz 20248
+YXBwZWQ= 20249
+IG5ld3NwYXBlcnM= 20250
+cmlldA== 20251
+IGFwcGxhdWRz 20252
+0JM= 20253
+IE5D 20254
+IGhldGVy 20255
+IGhhemFyZA== 20256
+IHJ5 20257
+IHN0cmljdGx5 20258
+IDU0 20259
+IOuTpOyWtOqwgA== 20260
+IHNwb250 20261
+IHRhdHPDpGNobGljaA== 20262
+IOunkOyU 20263
+bGF1Yg== 20264
+IGFic29yYmVk 20265
+YWNhxJ/EsXo= 20266
+IG9udQ== 20267
+INCQ0L0= 20268
+IGV4cGxpY2l0bHk= 20269
+IOyerA== 20270
+IEZ1dHVyZQ== 20271
+YWNodGVu 20272
+w6Bv 20273
+eW9u 20274
+IHNlcmlh 20275
+IEhlcnJlbg== 20276
+Y2Vq 20277
+IEFsYmVydA== 20278
+7J2064qU 20279
+ZWN0b3I= 20280
+IHBhY2tpbmc= 20281
+IHZpcnR1ZQ== 20282
+IHZlbmly 20283
+REQ= 20284
+IHlheg== 20285
+IGxvZ3M= 20286
+IFBob3Rvc2hvcA== 20287
+IHNpZA== 20288
+bGluZ3M= 20289
+IHJlbW90ZWx5 20290
+IERpZmZlcmVudA== 20291
+IG9wZXJhdGVk 20292
+bGlnaHRz 20293
+IGRpc2NyaW1pbg== 20294
+aXN0YW5jZQ== 20295
+IEdSRQ== 20296
+IHBsYWM= 20297
+IHNoaXJ0cw== 20298
+IGp1c3RpZnk= 20299
+IHRyYWJhbGhv 20300
+dXRpbA== 20301
+dm9j 20302
+IHF1YXJ0 20303
+IM6k 20304
+U0M= 20305
+IFNS 20306
+IC0i 20307
+IGhlc2l0YXRl 20308
+IHBhaw== 20309
+Z3Vh 20310
+Sm8= 20311
+IHNvdXZlbnQ= 20312
+IEFuZ2VsYQ== 20313
+ZXNzZWU= 20314
+YWRlbHBoaWE= 20315
+YXJrcw== 20316
+IHdlZWQ= 20317
+IGthbm5zdA== 20318
+IOq3uOufrOuLiOq5jA== 20319
+IHBsdXTDtHQ= 20320
+IENvbW1hbmRlcg== 20321
+IHN1bW1hcml6ZQ== 20322
+4K+A 20323
+IDk4 20324
+IGRldmVsb3BtZW50cw== 20325
+IENvc3Q= 20326
+IHRoZW9yZXRpY2Fs 20327
+IG9yZQ== 20328
+IG1ldGFsbA== 20329
+zr/Phc69 20330
+ZmFocg== 20331
+0JrQkA== 20332
+IGNodWNr 20333
+IGFkYXB0ZWQ= 20334
+IE9rbGFo 20335
+IE5ldGhlcmxhbmRz 20336
+IHBvZXQ= 20337
+c3Rv 20338
+a2F0 20339
+IHdlYXJz 20340
+568= 20341
+IOyWtOuUlA== 20342
+IEVzdG8= 20343
+IGxhdWdoZWQ= 20344
+IGRvbm5lcg== 20345
+IOuNsA== 20346
+IOybkOs= 20347
+b2N1cg== 20348
+IEtpY2s= 20349
+IERldHJvaXQ= 20350
+IGJpY3ljbGU= 20351
+IGxhY2tpbmc= 20352
+cGhhYmV0 20353
+IEtlbmQ= 20354
+QXNz 20355
+IHJldmVhbHM= 20356
+IM6g 20357
+IE5vYWg= 20358
+pqzripQ= 20359
+IHNlbGxz 20360
+IEFsYWJhbWE= 20361
+IHRlcnJpZmlj 20362
+IEVsZW1lbnQ= 20363
+IO2G 20364
+IHR1cmJv 20365
+IEhvbQ== 20366
+IHRoZW9yZW0= 20367
+IGFkdmVudHVyZXM= 20368
+IHB1cmNoYXNpbmc= 20369
+IFTDoQ== 20370
+INC80LDRgg== 20371
+IHZlbW9z 20372
+IGR1dGllcw== 20373
+IHdlbmln 20374
+IGJvb3Ro 20375
+IGVudHJhcg== 20376
+VkE= 20377
+IGdlYXJz 20378
+IEphZQ== 20379
+w6hu 20380
+IGNhbGNpdW0= 20381
+IFJvYmVydHM= 20382
+INC/0YDQvtCx0LvQtdC8 20383
+IHJpYmJvbg== 20384
+INC90LDQt9GL0LI= 20385
+IGxhdg== 20386
+IGludGVydmVudGlvbnM= 20387
+IFVsdHJh 20388
+IG5hbWVseQ== 20389
+IGFkZXF1YXRl 20390
+IHJlY2Fw 20391
+IGRvY2s= 20392
+ZnRpbmc= 20393
+IHZvaQ== 20394
+IGNvbnN1bHRhdGlvbg== 20395
+INGB0LXQvA== 20396
+IHBvZGVt 20397
+IHBvc3Nlc3Npb24= 20398
+IGNsdWVz 20399
+IFJ1c3NlbGw= 20400
+IHJlbmV3YWJsZQ== 20401
+INGD0Lc= 20402
+aW5mb3JtYXRpb24= 20403
+aWdnZXJz 20404
+V2l0aA== 20405
+d25v 20406
+IGVsYWJvcmF0ZQ== 20407
+Y3RvcmFs 20408
+IERvdw== 20409
+IHJhbWVu 20410
+4buV 20411
+IGVyc3Rl 20412
+IFplbA== 20413
+IHF1YXNp 20414
+INC90LDQug== 20415
+IFN0YXJz 20416
+IHRyaWJhbA== 20417
+IHNlYXRlZA== 20418
+IHdvbA== 20419
+IGNob2w= 20420
+w6Rtw6Q= 20421
+IG91dGJyZWFr 20422
+IGNyZXM= 20423
+IHVuc2VyZXI= 20424
+IO2RnA== 20425
+IHVuZGVyd2F0ZXI= 20426
+IGFzc3VyZQ== 20427
+T09E 20428
+IG5hcHJhd2TEmQ== 20429
+IGVzdGFibGlzaG1lbnQ= 20430
+IGluY29u 20431
+IGRpZmVyZW50ZQ== 20432
+IGV4Y3Vz 20433
+IERpbQ== 20434
+0L7RhQ== 20435
+IExpbmc= 20436
+cm9sb2c= 20437
+IG91dGRvb3Jz 20438
+bmFq 20439
+IGVwaWRlbWlj 20440
+IHVudGVycw== 20441
+IDMwMDA= 20442
+IEdhYnJpZWw= 20443
+IOyXhuuKlA== 20444
+IGVuY2w= 20445
+IE9kZXI= 20446
+IEZvb3Q= 20447
+cGFz 20448
+IFp1aw== 20449
+IHdvcmtmbG93 20450
+IHVucA== 20451
+IGFsbGlhbmNl 20452
+ZW5zY2hhZnQ= 20453
+IHlvZ3VydA== 20454
+0LjQvdC1 20455
+IGVydQ== 20456
+IGZpeg== 20457
+IGHFnw== 20458
+IGFwcmVuZA== 20459
+IGN1YWxxdWllcg== 20460
+IGNhcnJvdHM= 20461
+xLFuxLFu 20462
+YWZvb2Q= 20463
+IGZsb29ycw== 20464
+IGtleXdvcmRz 20465
+IHNwb3R0ZWQ= 20466
+IGRyYW5r 20467
+IHBhcmFz 20468
+IMO6bHRpbW8= 20469
+IGhhYmxhcg== 20470
+IHByb3NlY3V0 20471
+7JeQ64+E 20472
+IMOpcA== 20473
+IHN0aWNrZXJz 20474
+IHB1c2hlcw== 20475
+a2g= 20476
+IHJlc3RhcnQ= 20477
+IFRodW5kZXI= 20478
+4budaQ== 20479
+IG11aXRh 20480
+IGZveA== 20481
+YXJkZcWf 20482
+IFphY2g= 20483
+IE1pbmVjcmFmdA== 20484
+57g= 20485
+ID09PT0= 20486
+IGfDtnJl 20487
+IHN0YW5jZQ== 20488
+aWd1bmc= 20489
+2Y7ZkQ== 20490
+a8Ok 20491
+IHRlYWNoaW5ncw== 20492
+6YY= 20493
+IGRlY2F5 20494
+IHJpYw== 20495
+b21lbmE= 20496
+INCy0YHQtdC8 20497
+Y2h0ZW4= 20498
+IFZlcnQ= 20499
+IO2VnOq1rQ== 20500
+rLTr 20501
+IGNvYw== 20502
+Oik= 20503
+a2VpdGVu 20504
+IEJB 20505
+ZXRoZWxlc3M= 20506
+IGhlYWRxdWFydGVycw== 20507
+IHNwaWtl 20508
+IEJhc2U= 20509
+IDEwMQ== 20510
+IGNvb3JkaW5hdGVz 20511
+IHRhcmQ= 20512
+IGJvaWxlZA== 20513
+IE1vbnN0ZXI= 20514
+IG5vdGVib29r 20515
+IOq0gA== 20516
+IFdha2U= 20517
+IFNldHRpbmc= 20518
+7J207Jc= 20519
+IFN5ZG5leQ== 20520
+IEZpbm4= 20521
+IGxvYmJ5 20522
+IHNlbmlvcnM= 20523
+0L3QuNGF 20524
+YXZhbg== 20525
+IEpF 20526
+IHRyYWZm 20527
+dGhpbms= 20528
+IHNsYXA= 20529
+IENhc3RsZQ== 20530
+qW5n 20531
+IGFsZ3Vub3M= 20532
+IFBlcnNvbmFsbHk= 20533
+IE1hbGU= 20534
+7Yuw 20535
+IEdlbmVyYWxseQ== 20536
+IFBlbA== 20537
+IGRpYXM= 20538
+IGV2b2x2aW5n 20539
+aXRvbA== 20540
+0LLQvtGA 20541
+IHBsZWlu 20542
+IGZsaWdodHM= 20543
+IGVsZXZlbg== 20544
+b3dlag== 20545
+4buRbmc= 20546
+IGFrdQ== 20547
+IGdsYW5jZQ== 20548
+IGNvbm5lY3Rpdml0eQ== 20549
+IGJhbGQ= 20550
+0YvRhw== 20551
+IGludGVzdA== 20552
+w6Fn 20553
+IEdSw5w= 20554
+aWJsaWNhbA== 20555
+IFBhcGE= 20556
+IHBpdHk= 20557
+IGZhaW50 20558
+IHd1cmRlbg== 20559
+IGxlZ2FsbHk= 20560
+IHByZXk= 20561
+IFNjaWVuY2Vz 20562
+INC/0YDQvtGB 20563
+IHRyYWluZXI= 20564
+IHByb2Jsw6htZQ== 20565
+IGtpbG8= 20566
+0LrQvtCz0L4= 20567
+IGJyaWRnZXM= 20568
+ODk= 20569
+IGxhc3RlZA== 20570
+IGVsZWdhbnQ= 20571
+Ym93cw== 20572
+IHBhbGFi 20573
+IGRpcmVjdG9yeQ== 20574
+IGJ1bGI= 20575
+cGVvcGxl 20576
+SVg= 20577
+IGdlYg== 20578
+IDY2 20579
+IFRlbm5lc3NlZQ== 20580
+YWhsZW4= 20581
+aWV2YWw= 20582
+IGNhdXQ= 20583
+IERhbWVu 20584
+cGxv 20585
+aWFuZQ== 20586
+0LDQu9C1 20587
+YXR0YW4= 20588
+INin2YTYsw== 20589
+IHJpc2t5 20590
+IHNsZWV2ZQ== 20591
+IGluY2lkZW50cw== 20592
+IOuwlQ== 20593
+Q28= 20594
+IGFwcGxpY2FibGU= 20595
+IGltcGVyaWFs 20596
+IFBoaWxpcA== 20597
+IFllYQ== 20598
+0LXRgNC+ 20599
+INC/0L7QutCw0Lc= 20600
+w7xuZQ== 20601
+7JiA 20602
+SHVi 20603
+dG9y 20604
+IHNpZ3U= 20605
+Y2VuZA== 20606
+IHBvbGl0aWNhbGx5 20607
+IOyCtA== 20608
+IHBhcnM= 20609
+IG91dg== 20610
+IHByaW1laXJh 20611
+IFNoYWg= 20612
+IHNhdHVy 20613
+IGNvbWJ1c3Q= 20614
+IHByb21vdGVk 20615
+7KO86w== 20616
+IHRlbXBsYXRlcw== 20617
+IOuLrA== 20618
+IGhhdWw= 20619
+INGC0LXRgA== 20620
+IHNsaWRpbmc= 20621
+Y2VkZW50ZWQ= 20622
+Y2hpbGRyZW4= 20623
+TVI= 20624
+IFdlaQ== 20625
+IGLDtnI= 20626
+IHByw7N4aW1v 20627
+YXLDrWE= 20628
+IHNhbXBsaW5n 20629
+0LXQu9C10L0= 20630
+ZXNp 20631
+IERhbmllbGxl 20632
+IE9rbGFob21h 20633
+6IU= 20634
+0LXRgdC/ 20635
+IERWRA== 20636
+INCy0YvQvw== 20637
+cm91cw== 20638
+Y29ucw== 20639
+IGVuaGFuY2Vk 20640
+IHBhc3Rvcg== 20641
+IFN1ZGRlbmx5 20642
+ZmFy 20643
+UEVS 20644
+IE5n 20645
+MTAwMA== 20646
+IGNoZXc= 20647
+IHJ1bW9ycw== 20648
+IEFuYQ== 20649
+IGFubsOpZXM= 20650
+INGD0YHRgg== 20651
+IFBoaWxhZGVscGhpYQ== 20652
+0LXQttC00YM= 20653
+IGVmZmVjdGl2ZW5lc3M= 20654
+w6l0w6k= 20655
+IGRpbmc= 20656
+IHJlbGlnaW9ucw== 20657
+IGFnZWQ= 20658
+emllxIc= 20659
+IFJpYw== 20660
+IEthcA== 20661
+IFBhZ2U= 20662
+IHPDvA== 20663
+IG7DpG1saWNo 20664
+IG1hbmtpbmQ= 20665
+IHJlc3Rpbmc= 20666
+IGluZmx1ZW5jZXM= 20667
+IFNjaHVs 20668
+INC90LXQsg== 20669
+IG1hbmE= 20670
+IGNvbnN1bWVk 20671
+IFBvbQ== 20672
+IGNvbnNlZ3Vpcg== 20673
+IFRoYW5rc2dpdmluZw== 20674
+IEhpbmR1 20675
+bGFpcw== 20676
+IHRocml2ZQ== 20677
+IGNvbnRvdXI= 20678
+0LDRhtC40Y8= 20679
+IGZhbGFuZG8= 20680
+IErDoQ== 20681
+emFu 20682
+0LjRgtGD 20683
+aXBoZXI= 20684
+amFtaW4= 20685
+IEhhbGxv 20686
+IDE2MA== 20687
+INC+0YHQvtCx 20688
+IG1ldGU= 20689
+IOyVjOs= 20690
+IEJhcmNlbG9uYQ== 20691
+bGV0dGVy 20692
+INCd0LXRgg== 20693
+5Zk= 20694
+IGFkZW3DoXM= 20695
+IGNvb3JkaW5hdGlvbg== 20696
+dW50cw== 20697
+IHNsb3A= 20698
+INC/0YDQuNC0 20699
+7KeA66eJ 20700
+IHF1ZXN0aW9uaW5n 20701
+IGRpZXNlbA== 20702
+IGRlag== 20703
+IGFmZmlybQ== 20704
+jZTrnbzqs6DsmpQ= 20705
+aWVubmU= 20706
+IGNyYW5r 20707
+IHByZWRpY3Rpb25z 20708
+IHBoeXNp 20709
+Y2hzZWw= 20710
+IGNvbWJpbmF0aW9ucw== 20711
+IGV4Y2VsbGVuY2U= 20712
+4bud 20713
+d2lkdGg= 20714
+d2VlZA== 20715
+hOulvA== 20716
+hOuniA== 20717
+IGFsdG8= 20718
+IGRhaXJ5 20719
+IE5vcm1hbA== 20720
+cHBlbg== 20721
+IG9iZW4= 20722
+IGRldmFzdGF0aW5n 20723
+IHBveg== 20724
+IEh1cw== 20725
+bWF6 20726
+IHdhcm5lZA== 20727
+IGRlbms= 20728
+IEF1c3M= 20729
+IHRyYWRlcw== 20730
+aGVsbA== 20731
+IHByaW1lcm8= 20732
+IG1pYQ== 20733
+0LLQsNGA 20734
+2KjZig== 20735
+IGtpY2tz 20736
+IGHEnw== 20737
+IE3DvA== 20738
+IGx1Yw== 20739
+0LXQvdC40LXQvA== 20740
+IFN0YW5kYXJk 20741
+cmljZQ== 20742
+IEN1Yg== 20743
+IGdvdQ== 20744
+IEpvw6Nv 20745
+0YPRgdC6 20746
+IGVucXU= 20747
+o4w= 20748
+Z2V3 20749
+IO2BsA== 20750
+b3dhbmlh 20751
+aWFuaQ== 20752
+IGZha3Q= 20753
+0Y/QvdC4 20754
+IGJlZg== 20755
+IHRodW1ibmE= 20756
+IGNldXg= 20757
+YXBwbGU= 20758
+TkVO 20759
+IGdhZA== 20760
+YXBvbg== 20761
+IEZhbnRhc3RpYw== 20762
+IGNvbmNlbnRyYXRlZA== 20763
+Z2lybA== 20764
+bGVuZQ== 20765
+INCU0LvRjw== 20766
+IMOpdGE= 20767
+YWFu 20768
+IG91dHRh 20769
+IG5hcmM= 20770
+IEJvZHk= 20771
+YnJ1c2g= 20772
+IGxlZ2lzbGF0aXZl 20773
+IE1lZ2Fu 20774
+IG1pc3Rha2Vu 20775
+IE1pc3NvdXJp 20776
+IGxhYmVsZWQ= 20777
+0LvRj9C10YLRgdGP 20778
+IHJlYWxpc2Vk 20779
+eW9yc3Vu 20780
+IFNhZmV0eQ== 20781
+IGFjY2VsZXJhdGU= 20782
+IHNhbmN0aW9ucw== 20783
+IHBlZQ== 20784
+IGp1ZWdv 20785
+IHBlcHBlcnM= 20786
+IHdhbA== 20787
+6riJ 20788
+ZWxsb3c= 20789
+INC20LXQvQ== 20790
+IGNpbmNv 20791
+INGB0LjRgdGC 20792
+Y292ZXJ5 20793
+IGdyYW0= 20794
+IMOpcG8= 20795
+IEJNVw== 20796
+aXZvbA== 20797
+IENoZW0= 20798
+dXNlbWVudA== 20799
+IFN1cHBvc2U= 20800
+IOqwgOyngOqzoA== 20801
+IG1pbGxlbm4= 20802
+IFR1bg== 20803
+IG1lZGFs 20804
+IGhhY2lh 20805
+IHN0aW11bHVz 20806
+IGJyaWdodG5lc3M= 20807
+YWllbnQ= 20808
+IEhhbmRz 20809
+aW5ldA== 20810
+IGNvYWxpdGlvbg== 20811
+IHJpc2Vz 20812
+cmluYQ== 20813
+IHNjb290 20814
+IGRlZmVuZGluZw== 20815
+IGludmVycw== 20816
+IGhpbGxz 20817
+IGZ1bGZpbGxlZA== 20818
+bGxpZQ== 20819
+IGFkb2xlcw== 20820
+IENoYXNl 20821
+IEpK 20822
+IG5ldWVu 20823
+IFRydQ== 20824
+IGluaGVyaXQ= 20825
+IHNpeHR5 20826
+IEV4cA== 20827
+IENsYXk= 20828
+0L7RgdC+0LE= 20829
+YXJuYQ== 20830
+IEltcGVyaWFs 20831
+INGN0YLQsA== 20832
+IHNvY2lhbGx5 20833
+YXR5 20834
+b2R5bmFt 20835
+IHJpYnM= 20836
+b21pYw== 20837
+IFRvbA== 20838
+0L7Qu9C2 20839
+IDE5OTg= 20840
+IGZyYW0= 20841
+IHJhbmtz 20842
+INCx0YPQtNGD 20843
+IENvbG9u 20844
+SHo= 20845
+IGFjY29tbW9kYXRl 20846
+IGV4cGxvZGU= 20847
+7YSw6w== 20848
+SEFFTA== 20849
+IEhhcnQ= 20850
+INC20LjQt9C90Lg= 20851
+5qE= 20852
+IGRlbGljYXRl 20853
+oNeX 20854
+IHRvZnU= 20855
+IGFjaGlldmVtZW50cw== 20856
+IFNvcg== 20857
+IGFncmVlbWVudHM= 20858
+IDU3 20859
+IHRhbXA= 20860
+IGZyYW7Dp2Fpcw== 20861
+IGhlcmJz 20862
+Y29ybg== 20863
+IGtvbms= 20864
+QU5B 20865
+IFFp 20866
+IHByw7Nw 20867
+IHRpZ2Vy 20868
+IOuRmA== 20869
+xINt 20870
+IGFwcHJlbnQ= 20871
+YWhhbg== 20872
+IHJ1bGluZw== 20873
+IHRzcA== 20874
+IHR3aXR0ZXI= 20875
+IHRlZW5hZ2Vy 20876
+YnVz 20877
+IO2S 20878
+IEFtZW5kbWVudA== 20879
+IHRhcHBpbmc= 20880
+IEFQSXM= 20881
+IG1hdGNoZWQ= 20882
+66m0 20883
+V0E= 20884
+IEJlYXV0eQ== 20885
+IGluZXZpdGFibGU= 20886
+IGdhc2Vz 20887
+INm+ 20888
+aGlnaA== 20889
+IE9wdA== 20890
+IHByZWRvbWlu 20891
+z4HPjA== 20892
+IHR1YmVz 20893
+IOyVoA== 20894
+IEFh 20895
+b21ldG93bg== 20896
+IElN 20897
+IGRlc2Fy 20898
+w6RyZW4= 20899
+INC80LDRgQ== 20900
+IE3DtmdsaWNo 20901
+IHJlbnRhbA== 20902
+IO2VqOq7mA== 20903
+IERpYW5h 20904
+IGF1dGlzbQ== 20905
+IFB1ZXJ0bw== 20906
+xLFsZA== 20907
+IGZhbGFu 20908
+IGRyZWFtaW5n 20909
+IGd1dGU= 20910
+INC60LDQvA== 20911
+IHdyZWNr 20912
+IHN0b3J5dGVsbGluZw== 20913
+IExlZ2VuZA== 20914
+IFVrcmFpbg== 20915
+INC/0YDQvtC40YE= 20916
+IFNL 20917
+IO2WiQ== 20918
+IMWbd2k= 20919
+IEJlbGlldmU= 20920
+IG1vc3RyYXI= 20921
+IFRvZGQ= 20922
+IE5pZ2Vy 20923
+aWN0aW5n 20924
+aGFyZA== 20925
+Oi8v 20926
+aXJhYmxl 20927
+aWdhdGlvbg== 20928
+IE1lbWJlcnM= 20929
+IOygnO2SiA== 20930
+IGRpc2NvdXI= 20931
+n70= 20932
+cmlrYQ== 20933
+IERO 20934
+IEZpZg== 20935
+IENhcGl0YWw= 20936
+0YDQvtC8 20937
+IFNhbnM= 20938
+eXVu 20939
+IHBpbG90cw== 20940
+IHRyYXQ= 20941
+IG55dA== 20942
+IOuvvA== 20943
+IGV4cG9uZW50aWFs 20944
+IGVtZXJnZQ== 20945
+IHRyYWplY3Rvcnk= 20946
+INC/0L7Rh9C10LzRgw== 20947
+IHNlYWxlZA== 20948
+YXR0aQ== 20949
+IHdpZGVz 20950
+INC+0LPRgA== 20951
+aWFuY2Vz 20952
+IHdpdG5lc3NlZA== 20953
+T3I= 20954
+b3Np 20955
+IEpvZWw= 20956
+b25hbA== 20957
+IEludGU= 20958
+Y2VkZXM= 20959
+IEdvdHRh 20960
+YW5pdW0= 20961
+IGZlbWFsZXM= 20962
+IExlYmVucw== 20963
+IG1vaXN0dXI= 20964
+IFNpbXBsZQ== 20965
+IERvY2g= 20966
+YXLDoQ== 20967
+IGdlc2VoZW4= 20968
+VVNU 20969
+xqFp 20970
+IGNsYXNzaWZpY2F0aW9u 20971
+IGRpYWdvbmFs 20972
+IHBlcm1ldHQ= 20973
+Y29tcA== 20974
+INin2YTYrQ== 20975
+IE1hbGF5cw== 20976
+IGdlaMO2cnQ= 20977
+IHBvcHBlZA== 20978
+IGNvbnRhY3RlZA== 20979
+INeb15w= 20980
+IDE0MA== 20981
+IGFkYXB0YXRpb24= 20982
+IG1hbnVz 20983
+IHR1cmtleQ== 20984
+IHByZWFjaA== 20985
+YnJpZ2h0 20986
+IGRvd25z 20987
+IHVucHJlY2VkZW50ZWQ= 20988
+IG1pZ2h0eQ== 20989
+IGNhdGVy 20990
+aXR0aQ== 20991
+Z3M= 20992
+IERlcHV0eQ== 20993
+d3JpdGU= 20994
+IEJsZXNz 20995
+w6Fj 20996
+IHN1bW1pdA== 20997
+IOuPvOyalA== 20998
+IHRob3VnaHRmdWw= 20999
+IHNocmVk 21000
+c2luZ2luZw== 21001
+INC70YPRh9GI0LU= 21002
+IHllbg== 21003
+IHZpYnJhbnQ= 21004
+IFdhbHRlcg== 21005
+IGhvc3Rz 21006
+IGFtYnVs 21007
+IGludmFzaW9u 21008
+b2dhbg== 21009
+IHJlYXNvbmluZw== 21010
+IHN1Y2M= 21011
+0LvQtdC60YI= 21012
+IGZhbGE= 21013
+IGtpbmdz 21014
+IGdvaW4= 21015
+IGNhbGli 21016
+IEdSw5xORU4= 21017
+b3Rlcg== 21018
+IGVpbno= 21019
+IGluc3VsaW4= 21020
+iqg= 21021
+IHNjYWxpbmc= 21022
+IENvcm4= 21023
+aHlk 21024
+IG1hdHRl 21025
+UEw= 21026
+IGFsaWVucw== 21027
+IFNlZw== 21028
+ZXN0aQ== 21029
+YXN0aWNz 21030
+IHdhcm1lcg== 21031
+IGluZ2Vu 21032
+IE1M 21033
+IHJvZGU= 21034
+IEV5ZQ== 21035
+YmVpdHM= 21036
+IEJhcm4= 21037
+wrss 21038
+IENodWNr 21039
+IHByb2ZpdGFibGU= 21040
+dWd1ZXNl 21041
+IEFyYWJpYQ== 21042
+IGNvY28= 21043
+IHB1ZWRv 21044
+IGluZmxhbW1hdGlvbg== 21045
+Y2xpcA== 21046
+IHRhYmxlc3Bvb25z 21047
+IOygkQ== 21048
+IFN3ZWQ= 21049
+IGFuYXQ= 21050
+7Iig 21051
+IGFycmli 21052
+IGRhbmNlcg== 21053
+IENhcnRlcg== 21054
+IG1hZ25pZmlj 21055
+c3RvcmU= 21056
+IGZhZGU= 21057
+IGFjY29tcGFueQ== 21058
+IHdhaHI= 21059
+IHllYXN0 21060
+IG1pbmVyYWw= 21061
+IGxlZ2lzbGF0dXJl 21062
+aXJvcw== 21063
+IGNyb3dkZWQ= 21064
+0YDQsNGI 21065
+b2NhZG8= 21066
+7Ja07JW8 21067
+IO2bhA== 21068
+IEJhcnJ5 21069
+bWFzdGVy 21070
+IG5pY2tuYW1l 21071
+ICIuLi4= 21072
+IFJz 21073
+IE1vb3Jl 21074
+IHZlbnVl 21075
+INCx0YM= 21076
+bGlob29k 21077
+IEFnZW5jeQ== 21078
+0LvQvtCy 21079
+IGthaA== 21080
+IOyGjOumrA== 21081
+IG1hcnNo 21082
+IGluY29ycG9yYXRlZA== 21083
+YW50d29ydA== 21084
+IGtpbWNoaQ== 21085
+IHdvbw== 21086
+IGRpc3RyYWN0ZWQ= 21087
+ZXJpZXM= 21088
+IGluZm9ybWFjacOzbg== 21089
+IENob29zZQ== 21090
+IEphZGk= 21091
+IGFuYWxvZ3k= 21092
+c2F5 21093
+dWZmbGU= 21094
+Ym9r 21095
+IGFjaWRz 21096
+IGFjcXVpc2l0aW9u 21097
+IHZhcmlhbnRz 21098
+IHBhc3NpZXJ0 21099
+7J2064KY 21100
+cnVjdGl2ZQ== 21101
+YnJpZw== 21102
+IOOAjA== 21103
+ZXBoZXI= 21104
+IHBI 21105
+dXRsaWNo 21106
+IHJlbGll 21107
+dWl0ZQ== 21108
+IHJlY2VwdGlvbg== 21109
+IGNvaA== 21110
+IFByZXA= 21111
+IGFudGljaXBhdGU= 21112
+a2Vl 21113
+IGRlc2lnbmF0ZWQ= 21114
+0Y/RgtC4 21115
+IEtvcg== 21116
+IEFuaW0= 21117
+w7xobA== 21118
+IFdoaXQ= 21119
+IHVuY292ZXI= 21120
+IE1heWE= 21121
+INGC0L7Qs9C00LA= 21122
+sJU= 21123
+dXRlbmFudA== 21124
+IOyWvOs= 21125
+IGZvcmVzdHM= 21126
+IG1lbWU= 21127
+IGRpc3Rpbmd1aXNoZWQ= 21128
+IE1hcng= 21129
+IExpb24= 21130
+IHNlcnZhbnRz 21131
+IERpYW0= 21132
+IFBvbGljeQ== 21133
+jbw= 21134
+IHRyaWdnZXJlZA== 21135
+YWJpbGly 21136
+IOydkQ== 21137
+IG5lZ290aWF0ZQ== 21138
+IGZleg== 21139
+IGVydw== 21140
+IHZhcmllcw== 21141
+IGplbWFuZA== 21142
+IGRpc2NoYXJnZQ== 21143
+0YHRj9GH 21144
+IFBBUg== 21145
+IEFmZmFpcnM= 21146
+IHZvdGVy 21147
+IGF0ZW4= 21148
+IGNyb2lz 21149
+b2JpbA== 21150
+IE9vcHM= 21151
+IEFyYw== 21152
+IEhlYXRoZXI= 21153
+YW5rYQ== 21154
+IHNpbXBsZXM= 21155
+zr/OvQ== 21156
+Ij4= 21157
+IGNob3Jkcw== 21158
+IFNhbmRlcnM= 21159
+IOu2hOs= 21160
+QmVu 21161
+IGRhcsO8YmVy 21162
+aWxpYW5z 21163
+IG9yZGVyaW5n 21164
+IE1hbmg= 21165
+IGtpbG9ncmFt 21166
+IGthcsWf 21167
+IGdyYXNw 21168
+IGdob3N0cw== 21169
+YWxlbg== 21170
+IEplZGk= 21171
+INCx0LvQuA== 21172
+IGRvd25sb2FkZWQ= 21173
+IGNvbmR1Y3Rpbmc= 21174
+IEhhaw== 21175
+IHJlc2VhcmNoZXI= 21176
+aWxhbg== 21177
+Z29vZA== 21178
+IEhhbm5haA== 21179
+IGTDvMWfw7xu 21180
+IE1lc3NpYWg= 21181
+dWl0eQ== 21182
+aW9uYQ== 21183
+IHByb2JhYmxl 21184
+IFlF 21185
+IGluZGVwZW5kZW50bHk= 21186
+IGJ1ZmZlcg== 21187
+YnVybg== 21188
+b3VyZA== 21189
+IE1jSw== 21190
+IGxpbmd1 21191
+dWplbXk= 21192
+0LXRgNGC 21193
+IGludHVpdGl2ZQ== 21194
+IGNyYWNrcw== 21195
+YXBwcm9wcmk= 21196
+bnR5 21197
+IGdlZW4= 21198
+IGxlbmQ= 21199
+IGNlcnRpZmljYXRpb24= 21200
+SURT 21201
+dW50ZXI= 21202
+cGVlcw== 21203
+IHRydW1w 21204
+IGJhbmtydXB0 21205
+IGZlYXM= 21206
+6Jc= 21207
+IGR1xbw= 21208
+IHZpcnVzZXM= 21209
+IDU4 21210
+Z29k 21211
+INC20LXQuw== 21212
+IHN0YWxr 21213
+SW5k 21214
+YWNoaQ== 21215
+IENG 21216
+IENvbmQ= 21217
+IHNhbmN0 21218
+IGNvbnRlbg== 21219
+IGZyZWVk 21220
+IFJU 21221
+IG1lbnRvcnM= 21222
+7KGx 21223
+IHBvcnRhYmxl 21224
+IFBhdWxv 21225
+cmFuZQ== 21226
+SEFIQQ== 21227
+IFNlY3Rpb24= 21228
+54Y= 21229
+aHl1bg== 21230
+IM6tz4c= 21231
+IFB1Yg== 21232
+IEluZGVwZW5k 21233
+IGNvbXBvdW5kcw== 21234
+INGB0Ys= 21235
+IG1lc3NhZ2luZw== 21236
+IGRlZGljYXRpb24= 21237
+IG5vdGljaW5n 21238
+IGRldm90ZWQ= 21239
+0Y7RgtGB0Y8= 21240
+IHNuYWtlcw== 21241
+IGJhdHRsZWZpZWxk 21242
+cGVycw== 21243
+IGRlbGE= 21244
+OTI= 21245
+IGhhaQ== 21246
+aWxsw6Q= 21247
+w6lyZXI= 21248
+ZXZlcnk= 21249
+IHJlc3BvbnNpdmU= 21250
+15nXlQ== 21251
+b3Bm 21252
+6Yk= 21253
+irg= 21254
+QmVjYXVzZQ== 21255
+IHRvdXJpc20= 21256
+IOq3uOqyjA== 21257
+15XXpg== 21258
+IGNhbnM= 21259
+c3TDvHQ= 21260
+IGRvbm5l 21261
+IERpb3M= 21262
+IFViZXI= 21263
+YWN0b3J5 21264
+IG9yaWVudGVk 21265
+IEhlcm0= 21266
+IHBhdHJvbg== 21267
+dXJm 21268
+YmVp 21269
+IHByb2dyYW1h 21270
+IE9oaA== 21271
+Z2VuZXI= 21272
+IGZpc3Q= 21273
+IFdlbmR5 21274
+IGFuZGE= 21275
+IGd1ZXNzZWQ= 21276
+IGZyZWFr 21277
+IEtpbmdz 21278
+Y2hvb2w= 21279
+IG9mZmxpbmU= 21280
+IEluZGlhbmE= 21281
+IEFsbGlhbmNl 21282
+IDUz 21283
+IHBhcnRpY3Vs 21284
+IEZvY3Vz 21285
+IGluaGFiaXQ= 21286
+IOqwmeydgOuNsA== 21287
+IE1jRw== 21288
+b3dza2k= 21289
+IOydtOqxtA== 21290
+IHBhxYRzdA== 21291
+0L7QvdC4 21292
+aXR0YQ== 21293
+IGNvbmZpcm1hdGlvbg== 21294
+IEJyb29rbHlu 21295
+IG5vb2RsZQ== 21296
+ZnVuZA== 21297
+aXR1ZA== 21298
+IGdyYW5kcGFyZW50cw== 21299
+IGJhcmJlY3Vl 21300
+zrXOuc+C 21301
+IOE= 21302
+IGJhbGxvdA== 21303
+IFZldGVy 21304
+IHBpcGVz 21305
+aWdpb3Vz 21306
+IEdyYXBo 21307
+ZXN0ZWQ= 21308
+IOu4jOs= 21309
+IEtF 21310
+IGVpbnM= 21311
+IGhhdHJlZA== 21312
+IGRhbmc= 21313
+ZWVlZQ== 21314
+IGFyY2hhZQ== 21315
+IEplc3Nl 21316
+IGRldGVjdGVk 21317
+IHNlbmk= 21318
+YnVyZ2g= 21319
+IGRpc3BsYWNlbWVudA== 21320
+IGRvcA== 21321
+IGNvbmRpdGlvbmluZw== 21322
+INC90LXRgdC60L7Qu9GM0LrQvg== 21323
+IGRpc3R1cmJpbmc= 21324
+UEg= 21325
+IHRoaW5uZXI= 21326
+IHdvdW5kZWQ= 21327
+IEN1YW5kbw== 21328
+IGN1c2hpb24= 21329
+IHdoaXRlcw== 21330
+IHByZWZlcmVuY2Vz 21331
+IOykgOu5hA== 21332
+IGthxbw= 21333
+IEdhdGU= 21334
+IFBhdGg= 21335
+ZGxlcw== 21336
+4LiE4Lij 21337
+aW1vcmU= 21338
+IOuztOyXrA== 21339
+IGRpc2NpcGxpbmVz 21340
+4buP 21341
+IG1lc21h 21342
+IOyDiOs= 21343
+IOyLrA== 21344
+IGdpbmc= 21345
+IHVtYnJlbGxh 21346
+SUdIVA== 21347
+IHBlbnNpb24= 21348
+IGNvbWJpbmluZw== 21349
+U1M= 21350
+IHJlY3RhbmdsZQ== 21351
+4buHdA== 21352
+IHByb3hpbQ== 21353
+IENvdw== 21354
+uIw= 21355
+IGludGVudGlvbmFs 21356
+IGRlY2lk 21357
+INGB0LrQsNC2 21358
+IFVtYQ== 21359
+aWFzbQ== 21360
+YnV6 21361
+IGRlYnJpcw== 21362
+IGNhc3M= 21363
+IFByb3A= 21364
+aXNrYQ== 21365
+66Cl 21366
+ZXN0ZXJvbA== 21367
+dXNzaWFu 21368
+7J20656R 21369
+IHVubGltaXRlZA== 21370
+IGFkbWlyZQ== 21371
+IHRpZ2h0bHk= 21372
+IGdlbm9tZQ== 21373
+IEp1bmlvcg== 21374
+dmVuaXI= 21375
+Z3Vz 21376
+IGPEgw== 21377
+IFZsYWQ= 21378
+IO2C 21379
+IHJlbGF0aXY= 21380
+aW5jaQ== 21381
+IGF1bnF1ZQ== 21382
+IEJveXM= 21383
+0YbQuNC+0L0= 21384
+IFN3aXNz 21385
+IHBoeXNpY2lhbnM= 21386
+IO2PiQ== 21387
+IFBFVA== 21388
+IHdvdW5kcw== 21389
+YWJvdXQ= 21390
+w6Bp 21391
+b256 21392
+dXJpdGllcw== 21393
+INGD0LLQuNC0 21394
+IG1lbnRhbGl0eQ== 21395
+IHZhcmlhbmNl 21396
+IHNlZ3VuZGE= 21397
+IHZvbGNhbm8= 21398
+YWxpZQ== 21399
+4KWH 21400
+IHRpbGVz 21401
+IFRlcnJ5 21402
+INin2YTZhNmH 21403
+IGNhbm9u 21404
+IHNjYXR0ZXJlZA== 21405
+cHRvbg== 21406
+IGRlZmluaXRpb25z 21407
+IGFsZ2VicmE= 21408
+b3Rlbg== 21409
+YWJsbw== 21410
+aWp1YW5h 21411
+IHdyYXBwaW5n 21412
+IHNlc2FtZQ== 21413
+INC90LDRh9C40L3QsA== 21414
+IEFsZg== 21415
+INCg0L7RgdGB 21416
+b3Jubw== 21417
+IGFua2xl 21418
+IHNwZWNpYWx0eQ== 21419
+IGF0dGVtcHRpbmc= 21420
+aWxpYXRpb24= 21421
+IDE5MjA= 21422
+IHBoZW5vbWVuYQ== 21423
+IFByb2R1Y3Q= 21424
+IEJ1Y2s= 21425
+IEF3dw== 21426
+c2Vlbg== 21427
+IHZvaWQ= 21428
+IEZyYW5rbGlu 21429
+IGFkdm9jYWN5 21430
+IFNlcA== 21431
+IGNvb2xlc3Q= 21432
+INGB0YDQsNC30YM= 21433
+IFF1YW5k 21434
+IDkwMA== 21435
+IFRyYWQ= 21436
+ZGllcw== 21437
+IGhhc2g= 21438
+IHBvdHM= 21439
+IHNhZGx5 21440
+IHZpYWJsZQ== 21441
+IFRpZ2Vy 21442
+IE9ORQ== 21443
+IG5ldXJvbnM= 21444
+b3dhbmll 21445
+xJc= 21446
+IFNoYXI= 21447
+IExhbmRlcw== 21448
+IGNvbmZlcmVuY2Vz 21449
+IGNyZWRlbnRpYWw= 21450
+IGxpbWU= 21451
+aW5lZQ== 21452
+eGl0 21453
+cGF5 21454
+IGluY29ucw== 21455
+ID4+Og== 21456
+IO2emOs= 21457
+IGxlc3Nlcg== 21458
+IHNwaWxs 21459
+IHByZW1pc2U= 21460
+IDM2NQ== 21461
+IEhvc3Q= 21462
+IHRvbWFy 21463
+15DXnA== 21464
+67KI 21465
+IFdoYXRz 21466
+IGxpZ2h0d2VpZ2h0 21467
+IE1hcA== 21468
+Zmlh 21469
+ZWxsc2NoYWZ0 21470
+IHZlbmRvcnM= 21471
+dWVzdG8= 21472
+IE1pc3Rlcg== 21473
+INCf0YDQuA== 21474
+aG1h 21475
+IGludGVudGlvbmFsbHk= 21476
+IFRhbmc= 21477
+IGlkZW50aWZpY2F0aW9u 21478
+IGV0Y2V0ZXJh 21479
+IE5lZQ== 21480
+INGC0YDQuA== 21481
+6re4 21482
+IGNyeXB0b2N1cg== 21483
+IGluaGFsZQ== 21484
+IGFkZGljdA== 21485
+IG1hdQ== 21486
+INGC0LDQutCw0Y8= 21487
+IOuyhA== 21488
+IGNvbXByYXI= 21489
+aWVkemllxIc= 21490
+INC+0YLQvdC+ 21491
+IGJlZ2lubmVy 21492
+INC80YPQtg== 21493
+IG9ic2M= 21494
+IGxpbWl0aW5n 21495
+YXNjdWxhcg== 21496
+IGluc3BlY3Rpb24= 21497
+YWNp 21498
+IHJlam8= 21499
+TXVz 21500
+IHphdGVu 21501
+IHN6Y3o= 21502
+IE1hZHJpZA== 21503
+IHZhcmlldGllcw== 21504
+IGVzdMOg 21505
+IFNoYWtlcw== 21506
+IGtpdHM= 21507
+IGFkbWluaXN0ZXI= 21508
+IGxhdmE= 21509
+IGfDpQ== 21510
+16rXmQ== 21511
+IFdheW5l 21512
+IGluc3RhZ3JhbQ== 21513
+IHJhdGVk 21514
+cGFwZXI= 21515
+IGJpbGQ= 21516
+IHByZXRlbmRpbmc= 21517
+IG9ic2VydmluZw== 21518
+INGB0LDQvNC+0Lw= 21519
+IHRyb3I= 21520
+IG9yZ2FuaXNtcw== 21521
+IGZhbHRh 21522
+IGhvbWV0b3du 21523
+57E= 21524
+IO2L 21525
+IGNoZWc= 21526
+IOyh 21527
+IGNvbW1h 21528
+aXPDqQ== 21529
+IGxpa2VsaWhvb2Q= 21530
+YXZvcmVk 21531
+IGdlbGRp 21532
+0L3QuNC60L7Qsg== 21533
+IG1lZGlv 21534
+IGpha2ll 21535
+IEp1cA== 21536
+IGdyZWVuaG91c2U= 21537
+IHNwaXQ= 21538
+0LrQvtC1 21539
+INC60LDQtg== 21540
+IEdyYW0= 21541
+IENvbmZlcmVuY2U= 21542
+IGRlZmljaXQ= 21543
+c8Sxbg== 21544
+aW5zZQ== 21545
+dcSf 21546
+IHJpY2h0 21547
+IGNvaW5jaWRlbmNl 21548
+IGV1cm9w 21549
+IGJ1dHRlcmZseQ== 21550
+cHJlYWQ= 21551
+IOyWvA== 21552
+IHdhdmVs 21553
+IEluZmlu 21554
+IFBsYW5ldA== 21555
+IHNlbGZpZQ== 21556
+aWVudHJhcw== 21557
+IGFycm9n 21558
+b3Nlcg== 21559
+aWRhbA== 21560
+oNeX16DXlQ== 21561
+w7x0w7xu 21562
+IGZyZXNobWFu 21563
+IE1hY2hpbmU= 21564
+z4PPhA== 21565
+IERpYQ== 21566
+7J2064uk 21567
+bmVh 21568
+IGxpc3Rpbmc= 21569
+IGNvbmZpZ3VyZQ== 21570
+dXRvcg== 21571
+VXA= 21572
+dHNjaGFmdA== 21573
+cmnDqHJl 21574
+IHVwd2FyZHM= 21575
+INGF0L7Rh9GD 21576
+IHN3ZWVw 21577
+QnI= 21578
+IGV4cHJlc3Npbmc= 21579
+IHVuaGFwcHk= 21580
+IG1hbmRhdG9yeQ== 21581
+Z2VuZGVy 21582
+IEHDrQ== 21583
+IGluZGljYXRvcnM= 21584
+IG9pbHM= 21585
+bm90ZQ== 21586
+IHNlZ3Vy 21587
+0L7QttC10YI= 21588
+eW5hc3R5 21589
+IGRpc3RhbmNlcw== 21590
+IG1lcmdl 21591
+QkVSVA== 21592
+IHN1cnJlbmRlcg== 21593
+IGJ1YXQ= 21594
+IEF3YXJkcw== 21595
+IHNlw7Fvcg== 21596
+b2RveA== 21597
+IGZsYXZvdXI= 21598
+IGFiZG9t 21599
+IGNvbmZpZ3Vy 21600
+ODY= 21601
+IERJWQ== 21602
+IHJpZ2lk 21603
+sJg= 21604
+IGNvcnBvcmF0aW9u 21605
+IGdyb29t 21606
+amF3 21607
+IE5lYXI= 21608
+0LjQu9C+ 21609
+IG9wZXJh 21610
+IElubm92 21611
+0LjRgNCw 21612
+k7E= 21613
+IHNwZWNpZmllZA== 21614
+IGNvc20= 21615
+IEZyZWVkb20= 21616
+IGNsb3du 21617
+IE5lbQ== 21618
+INCy0L7Quw== 21619
+0ZHQvQ== 21620
+IGNoYXJnZXI= 21621
+4LmB4Lil 21622
+IGluZmx1ZW50aWFs 21623
+w6RzaWRlbnQ= 21624
+6aQ= 21625
+IOyEoOs= 21626
+IHZvbHVtZXM= 21627
+5pA= 21628
+IG91dHJhcw== 21629
+IFR3aXRjaA== 21630
+IGZvdW5kaW5n 21631
+IGF3aGlsZQ== 21632
+IGNvaWw= 21633
+6rCZ 21634
+IGPhuqM= 21635
+IFRocm93 21636
+IEhlbmNl 21637
+b21tdA== 21638
+IEJlbmphbWlu 21639
+0LPQu9GP0LQ= 21640
+VGltZQ== 21641
+b2JpYw== 21642
+IG1vdXI= 21643
+IGRyZWFk 21644
+IEzDoA== 21645
+IENoaWxl 21646
+IHByZXZhbA== 21647
+IHZhaW4= 21648
+IGFydMSxaw== 21649
+IHByZXNlcnZlZA== 21650
+INC+0YLQtA== 21651
+IHdhcmVob3VzZQ== 21652
+IGJlc3Rl 21653
+IFNldmVyYWw= 21654
+IFNpdHVhdGlvbg== 21655
+IGNhcmRib2FyZA== 21656
+VG9k 21657
+ZXJuYQ== 21658
+IGdhcmFudA== 21659
+IGdlc3R1cmU= 21660
+IGhlbg== 21661
+IHNwZWxsaW5n 21662
+b3NleHVhbA== 21663
+IGFubmU= 21664
+IG1pY2U= 21665
+IE1laW5l 21666
+Y2FyZA== 21667
+IHJlYmVsbA== 21668
+IGNlcnRv 21669
+IOycoOs= 21670
+IHZlcnNjaGllZA== 21671
+IEJvcw== 21672
+IGludmVudGlvbg== 21673
+IHRyemU= 21674
+IG1hbmnDqHJl 21675
+IENoYWQ= 21676
+IHNwcmU= 21677
+IG9yZ2FuaXNhdGlvbnM= 21678
+IHBvb3JseQ== 21679
+IGFudGVyaW9y 21680
+IHN0YWly 21681
+0LrRgA== 21682
+IGF0b21pYw== 21683
+IHN5bXBhdGg= 21684
+IGNvbnRpbnVhbGx5 21685
+IGtsZWluZQ== 21686
+w6h0ZQ== 21687
+0LjRiQ== 21688
+zr/Pgg== 21689
+cGV1dA== 21690
+IHJlcG9zaXQ= 21691
+IGVudHJh 21692
+RW0= 21693
+IGZpbmFuY2luZw== 21694
+INC80L3QvtCz 21695
+IHRoZXNpcw== 21696
+IENvbXB1dGVy 21697
+ZWF1 21698
+IFRyZWU= 21699
+IGJyaWRl 21700
+b25zaWV1cg== 21701
+c2hpcmU= 21702
+d2lj 21703
+REU= 21704
+IOyImOs= 21705
+IGFjb20= 21706
+IFBP 21707
+ZXJzY2g= 21708
+INC/0L7QvNC+0Yk= 21709
+IEFybWVu 21710
+IOyjvQ== 21711
+IHpvcg== 21712
+IHByaW50cw== 21713
+IERhc3M= 21714
+IGR1cmFibGU= 21715
+IFRyYW5zcG9ydA== 21716
+7J6Q6rCA 21717
+INC70LXQsw== 21718
+IGTDqXQ= 21719
+w7RsZQ== 21720
+YW1vdXM= 21721
+WU4= 21722
+IGNsaWZm 21723
+IGdyYW1tYXI= 21724
+INCf0L7RjdGC0L7QvNGD 21725
+IGzDoG0= 21726
+ZXNjaA== 21727
+IG1pc2VyYWJsZQ== 21728
+IHZvbHRz 21729
+IENhZA== 21730
+dWthbg== 21731
+0YLQuNCy 21732
+cnVzdA== 21733
+IOyYrOudvA== 21734
+IHZlcms= 21735
+IGNoaWNrZW5z 21736
+IFlvbw== 21737
+IG91dGZpdHM= 21738
+Y29kZQ== 21739
+IGhpZXJhcmNoeQ== 21740
+bmV0ZXM= 21741
+IGNvdW50ZXJwYXJ0 21742
+IHTDtGk= 21743
+IHRlZA== 21744
+IEJhcnQ= 21745
+IOudvA== 21746
+IEdlbmF1 21747
+IGluY29taW5n 21748
+IEFCQw== 21749
+cmlxdWU= 21750
+INC+0YLQvw== 21751
+cXVhbA== 21752
+IGluY2VudGl2ZQ== 21753
+IGlocmVu 21754
+16DXmQ== 21755
+bG9l 21756
+IDE5MzA= 21757
+IGJhcmc= 21758
+IGRpY3Rpb24= 21759
+IMO2bmNl 21760
+SU5T 21761
+IHJlaA== 21762
+aXNpYWo= 21763
+bW91dGg= 21764
+IHNjb3Jpbmc= 21765
+bMSxaw== 21766
+IOyVhOyjvA== 21767
+T1JJQQ== 21768
+IEVzdGFkb3M= 21769
+IGNvbXBhbmlvbg== 21770
+IGFzc2VtYmxl 21771
+IHB1bmlzaGVk 21772
+IGl0YWw= 21773
+IHByZXZlbnRz 21774
+aXN0ZXM= 21775
+IEtlbnR1Y2t5 21776
+IGxvY2F0ZQ== 21777
+IGZhc3Rpbmc= 21778
+g4A= 21779
+IFNlYg== 21780
+IENyb3du 21781
+b3BpYQ== 21782
+IHdoaXA= 21783
+dXN6 21784
+0LrQsNC80Lg= 21785
+IGRhdGFiYXNlcw== 21786
+IHByb3NlYw== 21787
+IDE5OTc= 21788
+IOyCtOynnQ== 21789
+IFNvbGFy 21790
+IFB1ZXM= 21791
+IFplbg== 21792
+b2xsbw== 21793
+IEd1cnU= 21794
+IHNxdWVleg== 21795
+INCX0LA= 21796
+IMSN 21797
+Y2VwdGlvbnM= 21798
+Y2Nh 21799
+aXphYmxl 21800
+bWFuZA== 21801
+IGJyZWFrdGhyb3VnaA== 21802
+IHRhYmxlc3Bvb24= 21803
+IFNFQw== 21804
+aWto 21805
+IFPDo28= 21806
+INC/0LvQvg== 21807
+YW1lbg== 21808
+IHByYWM= 21809
+IGRhcmxpbmc= 21810
+IHRhbGxlcg== 21811
+IHJlbmRlcmluZw== 21812
+IOyasOumrOqwgA== 21813
+IM+EzrfPgg== 21814
+IG3Dow== 21815
+IGVzb3M= 21816
+dWVyZG8= 21817
+INGB0YfQuNGC 21818
+YWxsZXI= 21819
+7JeI7Ja07JqU 21820
+IG1pbGxvbmVz 21821
+bGVyaW4= 21822
+IHBlZ2Fy 21823
+b25uZQ== 21824
+IGVucm9sbG1lbnQ= 21825
+IGxpZWd0 21826
+IGJvYQ== 21827
+d2nEmQ== 21828
+YnNw 21829
+IGN5Y2xpbmc= 21830
+IEJlcm5pZQ== 21831
+IDE5ODk= 21832
+INC00LDQu9GM 21833
+IERha290YQ== 21834
+INGB0LLRj9C3 21835
+IENQ 21836
+IHN0YXJl 21837
+7YKk 21838
+IHByb3NwZXJpdHk= 21839
+IGFycmFuZ2VtZW50cw== 21840
+IGFycml2aW5n 21841
+bcOk 21842
+IGtheWFr 21843
+aXB0 21844
+IHBhcmRvbg== 21845
+IHJlbGF0 21846
+IHZlcnN0ZQ== 21847
+IEZpZw== 21848
+IGZvaWw= 21849
+IFRhbGtpbmc= 21850
+cGVhcmU= 21851
+IG5vaQ== 21852
+INC/0YDQuNGI 21853
+IGhvY2tleQ== 21854
+IGFkbw== 21855
+IE9VVA== 21856
+Njc= 21857
+IGhvcm1vbmVz 21858
+IEF2ZW51ZQ== 21859
+IFN1cGVybWFu 21860
+IHByZXNjcmlwdGlvbg== 21861
+dWJlcm5ldGVz 21862
+Q0w= 21863
+b3RpdmU= 21864
+TklT 21865
+aWVuZW4= 21866
+IHNhZG5lc3M= 21867
+IFZpdA== 21868
+VHk= 21869
+IHN0YXJ0ZXI= 21870
+IGJlZGU= 21871
+IGZvdW5kYXRpb25z 21872
+IHNvcmU= 21873
+0YnQtdGB0YLQsg== 21874
+7Jqw6w== 21875
+INGH0YPQsg== 21876
+bGluaw== 21877
+IG1hbmV1 21878
+d29ya2luZw== 21879
+w6Bu 21880
+IEF0dGFjaw== 21881
+IENhcnQ= 21882
+dmVpcw== 21883
+IFJlc3A= 21884
+ZW5zaW5n 21885
+IOyii+yVhOyalA== 21886
+IGVzY3VjaA== 21887
+IFJOQQ== 21888
+grQ= 21889
+IGFkb3A= 21890
+IGJlbmRpbmc= 21891
+2LnYrw== 21892
+IG1hbmFnZXM= 21893
+dXNw 21894
+IHRhcnQ= 21895
+IHJvdXRlcg== 21896
+Qm8= 21897
+IGVzdGFibGlzaGluZw== 21898
+IGJhbGFuY2luZw== 21899
+IGF0aGxldGlj 21900
+IFNsbw== 21901
+IGZpbGxz 21902
+INC90LDQsQ== 21903
+INC00LDQuw== 21904
+IHBvc3Nv 21905
+IFZpZWxlbg== 21906
+IGNyaXRpY3M= 21907
+IGxhd3N1aXQ= 21908
+IElzYWFj 21909
+INGE0LjQu9GM0Lw= 21910
+IHRyYXM= 21911
+IHByYXc= 21912
+IENyYXp5 21913
+IG5ldQ== 21914
+IGt1bGw= 21915
+IHR1bW9y 21916
+IEFQUA== 21917
+Z2F0ZQ== 21918
+IEFSRQ== 21919
+OTg= 21920
+IFN0ZWFt 21921
+IGZ1Y2tlZA== 21922
+bGFnZQ== 21923
+IOKZrA== 21924
+IE1E 21925
+Znk= 21926
+IHNoZWxscw== 21927
+IFNlZW1z 21928
+aXplcnM= 21929
+IHJhbmdlcw== 21930
+IEFudG9uaW8= 21931
+QVRJT04= 21932
+IEJhYmE= 21933
+IOyDiQ== 21934
+a3Vu 21935
+IHByYXllZA== 21936
+0YDRjw== 21937
+INC/0YDQvtGC0LjQsg== 21938
+IHNlYXM= 21939
+YnVyeQ== 21940
+INeU16k= 21941
+IHRyYWl0 21942
+IERlcGVuZGluZw== 21943
+IGRyZQ== 21944
+IGvDtm5udA== 21945
+0YbRgw== 21946
+IGxpcHN0aWNr 21947
+ZWV6 21948
+INC/0YDQuNC80LXRgA== 21949
+IGFzc2lnbm1lbnRz 21950
+Qm9i 21951
+IG1ldGFscw== 21952
+IHNwZWNpYWxseQ== 21953
+IOyYiOs= 21954
+IMWh 21955
+IHZpc3Rh 21956
+IM6s 21957
+IHR3aW5z 21958
+IG5vdGFibGU= 21959
+IFNhdQ== 21960
+IGTDqXZlbG9w 21961
+IMOnZWs= 21962
+IHBvbHlub20= 21963
+YXZhbQ== 21964
+IHRhbWLDqQ== 21965
+0L7QvdC+0Lw= 21966
+IHBsYXNtYQ== 21967
+IGVmZWN0 21968
+IGzDpG5n 21969
+IGNhc2k= 21970
+0YHQsA== 21971
+xLFtxLE= 21972
+k6TsnYA= 21973
+IGxhYm91cg== 21974
+b3NzZW4= 21975
+IFB1bg== 21976
+cmlm 21977
+IGRvc2Vz 21978
+IG9wZXJhdGVz 21979
+0LjQu9C70Lg= 21980
+IGphYXI= 21981
+c3Rhdw== 21982
+IOyCrOuekQ== 21983
+IGF0bQ== 21984
+IHByb3RlY3Rz 21985
+IGltcGVk 21986
+SE8= 21987
+IGNpbWE= 21988
+IHRvY2g= 21989
+YWJpcw== 21990
+IHNlbmRv 21991
+bGF1cw== 21992
+IGN1cmw= 21993
+IE51bQ== 21994
+IHNwb25zb3Jz 21995
+IGTDqWJ1dA== 21996
+IEFsZXhh 21997
+IELDvHI= 21998
+IEFtZXI= 21999
+IGNvcGU= 22000
+INC40LfQsg== 22001
+amFs 22002
+IDE5OTU= 22003
+YXBhdA== 22004
+cmVzc2U= 22005
+IFByaXpl 22006
+IENsYWlyZQ== 22007
+IEJyYW5kb24= 22008
+IHdzenlzdGtv 22009
+IHZhbHVlZA== 22010
+4LiZ4Liw 22011
+IHNlY3Q= 22012
+IHNlY3JldGx5 22013
+IGRpYW1vbmRz 22014
+IEV2YW4= 22015
+IFJQRw== 22016
+iOuPhA== 22017
+IFVuaXZlcnNhbA== 22018
+IGRvdWJ0cw== 22019
+IFBpbg== 22020
+d2nEhXo= 22021
+mqk= 22022
+IGFsYm8= 22023
+IGJyYXVjaHQ= 22024
+QVVM 22025
+IE1vYmlsZQ== 22026
+Z3JhZGVz 22027
+IHNjaGVt 22028
+d2h5 22029
+IE5pY2h0 22030
+cGk= 22031
+Z2xl 22032
+IGNob3J1cw== 22033
+IGdseQ== 22034
+IHJlaW5mb3JjZQ== 22035
+IG11ZmY= 22036
+IFNoZW4= 22037
+IEhvbGE= 22038
+0YPQsw== 22039
+dmlkZW1tZW50 22040
+dmlhbA== 22041
+YWNpb3Vz 22042
+bGFpbWVk 22043
+IFJpY28= 22044
+IHZlZ2c= 22045
+IGlsbHVzdHJhdGlvbg== 22046
+IEJ1dHRlcg== 22047
+b3dhZA== 22048
+IGV1eA== 22049
+IGVuZmFudHM= 22050
+IExlYWRlcg== 22051
+IFZpbGxhZ2U= 22052
+ZXRpY2FsbHk= 22053
+2YbZig== 22054
+IHN0ZXc= 22055
+IHN1cnByaXNlcw== 22056
+IGN1ZQ== 22057
+IEdyYW5kbWE= 22058
+IENlbHNpdXM= 22059
+IFJpY2h0 22060
+ZW5j 22061
+IHBldGl0aW9u 22062
+IGhlcmI= 22063
+IHdpY2tlZA== 22064
+IHNjaGxl 22065
+b2NhbHk= 22066
+IHRyYW5zZg== 22067
+IHRva2Vucw== 22068
+IEdyYXk= 22069
+IEJCQw== 22070
+SUs= 22071
+IDE1MDA= 22072
+em4= 22073
+IE5ldg== 22074
+IGtveQ== 22075
+IHphcg== 22076
+IGJ1bGxzaGl0 22077
+IENvbG9tYmlh 22078
+dWxhdGl2ZQ== 22079
+IHdpZGVzcHJlYWQ= 22080
+eWVjdA== 22081
+a2l0 22082
+IGVtcHJlc2E= 22083
+IG5vdXI= 22084
+IGJ1cm5z 22085
+YXRpbg== 22086
+YWlyZWQ= 22087
+IHJldm9sdXRpb25hcnk= 22088
+INCz0L7QtNGD 22089
+IExvZ2Fu 22090
+IDE5OTY= 22091
+IEdyYWhhbQ== 22092
+cmVi 22093
+IE5IUw== 22094
+IGNvc3R1bWVz 22095
+IG5hd2V0 22096
+IGxvdmVycw== 22097
+IEx1Y3k= 22098
+IEluZGlnZW5vdXM= 22099
+7ZWY6riw 22100
+IGltbXVuaXR5 22101
+pbTr 22102
+dWl0bw== 22103
+IGV4Y2Vzc2l2ZQ== 22104
+IGRvbmF0aW9ucw== 22105
+INeU16g= 22106
+IOyyqw== 22107
+IGRyeWluZw== 22108
+bWVsb24= 22109
+IHN1cnZleXM= 22110
+IOustOyKqA== 22111
+YWFh 22112
+IHByb2Jl 22113
+YW5jaWFs 22114
+IGxvdWRlcg== 22115
+IGhvdGVscw== 22116
+w7zEnw== 22117
+YWduZXI= 22118
+IG9yaWdpbnM= 22119
+IOuniOyngOuniQ== 22120
+ICoq 22121
+IHN0cmFuZ2Vycw== 22122
+IEhhdXM= 22123
+Y29tZWQ= 22124
+IGFudGhyb3A= 22125
+IHVzbw== 22126
+IOyVhOyngQ== 22127
+IFl1YW4= 22128
+IO2VhOyalA== 22129
+cGxlcg== 22130
+cmVzc2l2ZQ== 22131
+IHNwcmF3 22132
+IFN0ZXc= 22133
+IDE5OTQ= 22134
+IGVsZGVycw== 22135
+IG1laW5lbg== 22136
+IGp1bnQ= 22137
+IGFjb3VzdA== 22138
+IFdvaG4= 22139
+IGJhbmFuYXM= 22140
+IHByb2plY3Rpb24= 22141
+IFN0aWNr 22142
+bGVndA== 22143
+c3BlZWQ= 22144
+IGPFqW5n 22145
+IFdvcnQ= 22146
+IEJhbHRpbW9yZQ== 22147
+INGG0LXQuw== 22148
+IGR1bm5v 22149
+Pyw= 22150
+IExvY2Fs 22151
+b3N0bw== 22152
+0K0= 22153
+0L7QtNCw 22154
+IFBvcnR1Z3Vlc2U= 22155
+IHRoZWlycw== 22156
+IGTDqW0= 22157
+IGRyYXVm 22158
+IEJ1ZGRoaXN0 22159
+ZXJ0YQ== 22160
+R2U= 22161
+IGNhcnJvdA== 22162
+IFdvbmRlcmZ1bA== 22163
+IHNvYWs= 22164
+IGNoYWlybWFu 22165
+Z2dp 22166
+SUNB 22167
+ZnJpZWQ= 22168
+IGZsaWNr 22169
+IFRocm91Z2hvdXQ= 22170
+IOyasOs= 22171
+IGNvdWdo 22172
+IGZsdWZmeQ== 22173
+c2Nob29s 22174
+IHJpcHBlZA== 22175
+LS0tLS0tLS0= 22176
+IFp1a3VuZnQ= 22177
+INC90LXQsQ== 22178
+IHN0bw== 22179
+IEJP 22180
+cGVudA== 22181
+IExhd3JlbmNl 22182
+z4nPgg== 22183
+c3RpY2tz 22184
+IEVpbnM= 22185
+INGA0Ys= 22186
+IFN0cm9uZw== 22187
+IGNhcmFtZWw= 22188
+IHNwaXRl 22189
+YXphcg== 22190
+IGNyaXRpY2FsbHk= 22191
+IG9icmE= 22192
+b3dpdHo= 22193
+IFpvbmU= 22194
+INGA0LXQug== 22195
+IHN1Zw== 22196
+YXJkZWQ= 22197
+IGfDrA== 22198
+ZmZlbnRsaWNo 22199
+YW5jaGU= 22200
+2J8= 22201
+YXN0aWNhbGx5 22202
+7J286w== 22203
+0LvQsNCy 22204
+IHNpbXBsZXN0 22205
+IEZyaWVuZA== 22206
+IHF1ZWxsbw== 22207
+IGFtYml0aW9u 22208
+IGFiYmlhbW8= 22209
+INGE0L7RgNC8 22210
+IEVzc2E= 22211
+IGVkdWNhdG9ycw== 22212
+IHN0YXRpc3RpY2Fs 22213
+IGNoYW5nZXI= 22214
+IGF0YXU= 22215
+w6l0YWlz 22216
+IFNoYWtlc3BlYXJl 22217
+65CY 22218
+IHRyaWdnZXJz 22219
+IHJlYWxpeg== 22220
+IGNlbHVp 22221
+d2hlZWw= 22222
+IGxveWFsdHk= 22223
+IHNjcmVhbXM= 22224
+a2Vocg== 22225
+IE1lZ2E= 22226
+ZWFzdA== 22227
+IHRvcHM= 22228
+IFRvdGFsbHk= 22229
+b3VudGFpbg== 22230
+bG9yZA== 22231
+IHZpb2xhdGlvbg== 22232
+IEdB 22233
+IG5pY2Vy 22234
+IEZyZXNo 22235
+IE1lbGlzc2E= 22236
+ZnVuY3Rpb24= 22237
+IHJhcGU= 22238
+IGV4Y2VwdGlvbnM= 22239
+IHNpbGljb24= 22240
+IGxpYmVydHk= 22241
+IGhvdXNlaG9sZHM= 22242
+IENB 22243
+INCe0LE= 22244
+IGxpYg== 22245
+now= 22246
+Y2lmaWM= 22247
+IHRyb3BpY2Fs 22248
+IGludmVzdGlnYXRpbmc= 22249
+SEQ= 22250
+IGFkYXB0ZXI= 22251
+IFBpdHQ= 22252
+YW5jaWE= 22253
+IFNoZWxs 22254
+ZnJpZW5kbHk= 22255
+IGNvbmNsdXNpb25z 22256
+IHR1cnRsZQ== 22257
+IGRlY29tcA== 22258
+IGFuaW1hdGlvbnM= 22259
+INGB0LXQug== 22260
+aW5zaQ== 22261
+IHJldGVudGlvbg== 22262
+a2ll 22263
+IGluamVjdGlvbg== 22264
+IE1hZGlzb24= 22265
+7LCw 22266
+IHZpZW50 22267
+IHZhcmllZA== 22268
+IHZpb2xpbg== 22269
+IEJpbA== 22270
+IGx1Y2tpbHk= 22271
+IGh0dA== 22272
+bMOk 22273
+IHJhbmNo 22274
+IHPDs2xv 22275
+7JWF 22276
+IERlcmVr 22277
+IFNjcmlwdHVyZQ== 22278
+0L7RgNCw 22279
+IGNsYXNzcm9vbXM= 22280
+YXZpbA== 22281
+Zm9ybWVk 22282
+IGJlZm9yZWhhbmQ= 22283
+IEdlbQ== 22284
+cHJlY2g= 22285
+IGxpbg== 22286
+IGdyZWVucw== 22287
+0YbQtdCy 22288
+IE1lcmNlZGVz 22289
+IGRyb3VnaHQ= 22290
+Z2FzcHM= 22291
+IGFib3J0aW9u 22292
+IHRlcnJpYmx5 22293
+IHNwb3PDs2I= 22294
+IHNlY3VyZWQ= 22295
+IGF0csOhcw== 22296
+IHdhdmVsZW5ndGg= 22297
+IGdyYWlucw== 22298
+ZWN0aXZl 22299
+IHNwYWNlY3JhZnQ= 22300
+IHRvdXJz 22301
+IHByb2Zlcw== 22302
+IHN1cmdlb24= 22303
+IFBpZQ== 22304
+IGlkZWFsbHk= 22305
+YXJuZXI= 22306
+VVA= 22307
+b3BhcmQ= 22308
+c2Nl 22309
+IGltbWVuc2U= 22310
+IE9ydA== 22311
+cm9sbGVy 22312
+IERhbGxhcw== 22313
+IE5pY2hvbGFz 22314
+IHN1bGY= 22315
+IFRveW90YQ== 22316
+IHF1YW50aXRpZXM= 22317
+Y2VhbnM= 22318
+IGN1aQ== 22319
+YW7Dp2E= 22320
+IENBTg== 22321
+aXR6ZXJsYW5k 22322
+IHpvdQ== 22323
+IEN5YmVy 22324
+bGVnZW4= 22325
+IEluaXQ= 22326
+ZWR1 22327
+IGFwZXJ0 22328
+IGFkamFj 22329
+b3V2 22330
+cnM= 22331
+IGNhYmJhZ2U= 22332
+IHdoZWVsY2hhaXI= 22333
+aW55bA== 22334
+IER5bmFt 22335
+IOyVhOuLiOudvA== 22336
+IGxpbmc= 22337
+aGw= 22338
+INC80L7Qs9GD 22339
+IGNyaXNw 22340
+IG1pag== 22341
+IGR1Zw== 22342
+bmlu 22343
+IGJsb3Nz 22344
+IGJlbG9uZ2luZw== 22345
+IGxvdWRseQ== 22346
+IG1pbmVyYWxz 22347
+IGNvbmNsdWRlZA== 22348
+IHNlYXJjaGVk 22349
+OTY= 22350
+IE1lZXQ= 22351
+IFNFTw== 22352
+INCh0Lo= 22353
+IEhvYg== 22354
+b3R0YQ== 22355
+IHByb3BhZ2FuZGE= 22356
+IGNpbm5hbW9u 22357
+IGh1bnRlcg== 22358
+IGdlbWVpbnM= 22359
+IHNjdWxwdHVyZQ== 22360
+dWxzaW9u 22361
+IHbDpGw= 22362
+IG1hZ2F6aW5lcw== 22363
+IGNvbnRyb3ZlcnN5 22364
+IHNlcXVlbmNlcw== 22365
+IO2ajA== 22366
+IGRlbGV0ZWQ= 22367
+kOuPhA== 22368
+IHZhcnlpbmc= 22369
+IG1vdW50aW5n 22370
+IGFmZmFpcg== 22371
+IHBhdGh3YXlz 22372
+5qY= 22373
+IGRpZ28= 22374
+INC00L7Qug== 22375
+QWxleA== 22376
+IHRvYmFjY28= 22377
+IENW 22378
+IGJvdGhlcmVk 22379
+IGFtYmllbnQ= 22380
+aW5reQ== 22381
+IFNM 22382
+IGhhdGVz 22383
+IGplxbxlbGk= 22384
+IGNvbmdyZWc= 22385
+IGVsYXM= 22386
+IGRldXRz 22387
+IFN0dWRpb3M= 22388
+Y2jEmQ== 22389
+IGRvY3VtZW50ZWQ= 22390
+IENydXo= 22391
+IExlbg== 22392
+IERvdWdsYXM= 22393
+IFBvcnR1Z2Fs 22394
+ZW50aQ== 22395
+IHNwb3VzZQ== 22396
+IGFuYWx5cw== 22397
+YXZpYQ== 22398
+IGVkaXRlZA== 22399
+IGzhuqFp 22400
+YnVpbHQ= 22401
+IHZpbGxl 22402
+YWRvcmE= 22403
+IGJyYWNlbGV0 22404
+IHN1c2hp 22405
+IHBt 22406
+IHRyYWlscw== 22407
+IGx1Zw== 22408
+IMO2dmVy 22409
+IHNvcnJvdw== 22410
+IGNvbG9ueQ== 22411
+YWRveA== 22412
+IHNlcmll 22413
+YW55YWs= 22414
+INi3 22415
+IEd1bGY= 22416
+IFBW 22417
+IFNhbXVlbA== 22418
+IEtpdA== 22419
+IFJhbA== 22420
+b250aW4= 22421
+ZXhwbA== 22422
+IGVudHJpZXM= 22423
+IGFjdGl2aXN0cw== 22424
+UHM= 22425
+IHNhbnQ= 22426
+INGC0L7Rhw== 22427
+IEJydW5v 22428
+a2VsZXk= 22429
+IHR1dHRv 22430
+6ZQ= 22431
+IHZpbnRhZ2U= 22432
+IHRlcnJpZmllZA== 22433
+INC/0L7RhQ== 22434
+dXNpdmU= 22435
+b3dlcnM= 22436
+0LDQudGC 22437
+64+Z 22438
+IHR3aXN0ZWQ= 22439
+IFRob3VnaHQ= 22440
+IHRhaA== 22441
+IHNocmluaw== 22442
+IHNoZWVy 22443
+bGl0 22444
+IGRhbGFt 22445
+IGRpYg== 22446
+IHZhcmQ= 22447
+b3dhbmU= 22448
+IGRvYnI= 22449
+IFJlbmE= 22450
+INGB0LLQvtGO 22451
+IHBhw61zZXM= 22452
+IEVyYQ== 22453
+IEJVVA== 22454
+c2lnaHM= 22455
+IOq3uOqxsA== 22456
+IGdyb8OfZW4= 22457
+IOu5qOumrA== 22458
+IG5lcnZlcw== 22459
+IGNvbnN0aXQ= 22460
+IHByZW9jdXA= 22461
+IEdheQ== 22462
+IFh1 22463
+a2VlcGVy 22464
+aGV1cmU= 22465
+Li4p 22466
+IENhbG0= 22467
+IFVuaWRvcw== 22468
+IOydtOqygw== 22469
+IEFxdWk= 22470
+IOygnOydvA== 22471
+ZMSxcg== 22472
+7KaY 22473
+eW91cg== 22474
+INGN0YLQuNC8 22475
+MjAyMA== 22476
+IHJ1bmQ= 22477
+IEhP 22478
+IENhdGhlcmluZQ== 22479
+aWVsaQ== 22480
+IGZ1c2lvbg== 22481
+IGlkZW9sb2d5 22482
+IGZvcmFt 22483
+c2hhcGVk 22484
+IO2bhOs= 22485
+IHd0 22486
+IHJldHI= 22487
+IHByw6lj 22488
+IOqwkQ== 22489
+IG9wZW5seQ== 22490
+dml0eQ== 22491
+6rWs7JqU 22492
+IG9ic3RhY2xl 22493
+IGJvbw== 22494
+IHNlaW5lcg== 22495
+aWNvcm4= 22496
+IGVpZ2VubGlqaw== 22497
+IGhlYWRlcg== 22498
+YXJlbW9z 22499
+IHNvZnRlcg== 22500
+INCf0L7QtA== 22501
+IHByZWp1ZA== 22502
+IGRlZmluZXM= 22503
+aWVydGU= 22504
+IGJsZW5kaW5n 22505
+IGJlbGlldmVycw== 22506
+IFdvY2hlbg== 22507
+INC90LjQutCw0Lo= 22508
+INCa0L7Qs9C00LA= 22509
+IFR5cGljYWxseQ== 22510
+IO2BrA== 22511
+Y2lvcw== 22512
+IG1pc3NpbGVz 22513
+IHNwb25nZQ== 22514
+IEtpdGNoZW4= 22515
+IHRyZW4= 22516
+bmluZ2Vu 22517
+IHNjcmFw 22518
+IHNlcmFpdA== 22519
+tOyg 22520
+57k= 22521
+IOuwmOs= 22522
+IHJlc3RvcmVk 22523
+IHByenlrxYJhZA== 22524
+IEt1YmVybmV0ZXM= 22525
+IHNhaXQ= 22526
+IHV3 22527
+IGVuYWJsaW5n 22528
+IHRyYXZlcnM= 22529
+YW1wcw== 22530
+IE9NRw== 22531
+ZW5zb3I= 22532
+IHpvc3Rh 22533
+IHByb25vdW5jZWQ= 22534
+QW5n 22535
+bm9ybWFs 22536
+IGVjb25vbWllcw== 22537
+dGlu 22538
+IENoYW1waW9u 22539
+aXplbg== 22540
+IGFyYmVpdGVu 22541
+IEdvc3BlbA== 22542
+IFp1 22543
+bmdh 22544
+IGxpdGVyYWN5 22545
+IE1hbnM= 22546
+IGNpcmN1bGF0aW9u 22547
+IGFkYXA= 22548
+IFRvdGFs 22549
+IG1lcmVrYQ== 22550
+IG9sYWNhaw== 22551
+0YHRgtCw0YLQuA== 22552
+SmFjaw== 22553
+IG11bmQ= 22554
+IHRoaWVm 22555
+Ymllcw== 22556
+IOqygQ== 22557
+YXF1ZQ== 22558
+INqp24w= 22559
+IFNjYXI= 22560
+5bI= 22561
+IGFib2w= 22562
+IGRldm90ZQ== 22563
+IDAx 22564
+IHNpdHRlbg== 22565
+IFZpc3VhbA== 22566
+d2Vlaw== 22567
+c29tZQ== 22568
+aW5ndA== 22569
+IGpvdXJuYWxpc20= 22570
+IEhpcg== 22571
+IEJhY2hlbG9y 22572
+aW5lcnk= 22573
+w5xORA== 22574
+IGNvbG9yaW5n 22575
+IENyaXN0 22576
+IGNlbGVicml0aWVz 22577
+INGH0LjRgQ== 22578
+IENyaXQ= 22579
+IGRpZmZlcmVudGlhdGU= 22580
+INCc0L3QtQ== 22581
+ZWxpbQ== 22582
+IHNlYWZvb2Q= 22583
+IGFsZ3VtYXM= 22584
+b3RoZXJhcHk= 22585
+IGdsYXVi 22586
+IGFyYml0cmFyeQ== 22587
+Z2Vucw== 22588
+INCx0YPQtNC10Lw= 22589
+IHRhdg== 22590
+IGNyZWFteQ== 22591
+IENvdW50cnk= 22592
+YcOx 22593
+0LzQtdGC 22594
+IGhpbnRlcg== 22595
+IG1pc20= 22596
+IGlsbHVzdHJhdGU= 22597
+w5xORE5JUw== 22598
+IGRlY3JlYXNpbmc= 22599
+IHdlbmlnZXI= 22600
+QUtJ 22601
+aXhvbg== 22602
+INC90LXQuQ== 22603
+IGZhdHRv 22604
+IG5lcmQ= 22605
+56A= 22606
+IGJpdHRl 22607
+UGVy 22608
+IHRhbmU= 22609
+IGfDtno= 22610
+IGZvcnRl 22611
+IEV5 22612
+INC90LDQstC10YA= 22613
+IFdvcmRQcmVzcw== 22614
+IE1pcw== 22615
+xa8= 22616
+esOkaA== 22617
+IGludMOpcmVzcw== 22618
+b3NhdXJz 22619
+IEZhbGxz 22620
+IG5lc3Nh 22621
+OTc= 22622
+IG11c2V1bXM= 22623
+IGNvcnJlc3BvbmRz 22624
+IHNpbmdz 22625
+Zm91cg== 22626
+IGVkZXI= 22627
+IENvbW11bmlzdA== 22628
+b2E= 22629
+bmVr 22630
+IFdITw== 22631
+IGNvcnBv 22632
+IG1lc3Npbmc= 22633
+z4TOsc65 22634
+IGJydXNoZXM= 22635
+IGJpc2M= 22636
+IEFyYmVpdHM= 22637
+IFRheA== 22638
+IHNlbGU= 22639
+IGZsYWdz 22640
+b3VwZQ== 22641
+IGFudGljaXBhdGVk 22642
+IE5hZA== 22643
+IHBvdXJlZA== 22644
+IG1s 22645
+IGxsYW1h 22646
+IHZpc3VhbGl6ZQ== 22647
+IGxpc3RlbmVycw== 22648
+2YTZgw== 22649
+YWx0ZW4= 22650
+TWljaGFlbA== 22651
+IGNvc8Os 22652
+1aHV 22653
+b3B1cw== 22654
+IO2VtOyjvA== 22655
+IGhpa2U= 22656
+IEF0dG9ybmV5 22657
+IEhpbGxhcnk= 22658
+dWRlZA== 22659
+IO2VmOyngOunjA== 22660
+IGRvdmU= 22661
+IHN0b3Jtcw== 22662
+0LDQutGB 22663
+IGRvY3RyaW5l 22664
+IGhleA== 22665
+aWtz 22666
+bm/Fm8SH 22667
+IHNjcmlwdHM= 22668
+IM60zrXOvQ== 22669
+INGN0YLQuNGF 22670
+INCG 22671
+YWJlcg== 22672
+IFZhcw== 22673
+IGNlbnRpbWV0ZXJz 22674
+157XlA== 22675
+0L3QuNCx 22676
+IHJpZGVycw== 22677
+IFRyaWI= 22678
+IHRha8W8ZQ== 22679
+IG5vdW4= 22680
+IGljb25z 22681
+IHNvbGVseQ== 22682
+bWluZGVk 22683
+IGRpc3Bvbg== 22684
+IFN3aXR6ZXJsYW5k 22685
+IGNsdXN0ZXJz 22686
+IHF1ZWRh 22687
+YWlsaW5n 22688
+IG1hbmdh 22689
+IDY4 22690
+hIg= 22691
+IHRldA== 22692
+Z2lucw== 22693
+aGF1cw== 22694
+IE9Q 22695
+b3RlZA== 22696
+IG5vdXZlYXU= 22697
+QUxMWQ== 22698
+2YjYrw== 22699
+w7Ju 22700
+IG1vcnRhbGl0eQ== 22701
+IEdpdEh1Yg== 22702
+ZHJvcA== 22703
+IGRpc2d1 22704
+IHJlY29t 22705
+IGxvY2Fscw== 22706
+IGhvbWVtYWRl 22707
+YW1iYQ== 22708
+IHByb251bmNpYXRpb24= 22709
+IGFscGhhYmV0 22710
+0LDQvdGM 22711
+b3dhbnk= 22712
+aXJhcw== 22713
+aWRlbmN5 22714
+T01F 22715
+INGA0LDRgdGB 22716
+YXJhaw== 22717
+dmlhbWVudGU= 22718
+IG5vbnByb2ZpdA== 22719
+IFlvdVR1YmVy 22720
+IHBhcmVudGg= 22721
+IEJvbw== 22722
+dmF0 22723
+IFN0aXI= 22724
+IHByZWNpcA== 22725
+IGFudHM= 22726
+IGFsbHk= 22727
+IE1hb3Jp 22728
+IOuMgO2VnA== 22729
+b2dlbmU= 22730
+IExhYm91cg== 22731
+YXJldHRl 22732
+IHJlY3ljbGluZw== 22733
+ZW5zYQ== 22734
+IHB1cnN1aXQ= 22735
+IHNhaw== 22736
+INCX0LTQtdGB0Yw= 22737
+IHRvbGVyYW5jZQ== 22738
+IHNhYXQ= 22739
+IGNsaWNrZWQ= 22740
+4pml 22741
+IGZhY2Vib29r 22742
+IEludG8= 22743
+IGluY2VudGl2ZXM= 22744
+6riw64qU 22745
+IERlbm5pcw== 22746
+IFdpaw== 22747
+Z2VzY2g= 22748
+4LmA4Lib 22749
+IM+AzrE= 22750
+IFdob28= 22751
+IHJvdW5kZWQ= 22752
+IGRvcGU= 22753
+IGNhcHR1cmluZw== 22754
+IFdhcnJp 22755
+IGNpdmlsaWFu 22756
+IGNoYXJtaW5n 22757
+IGVzYXM= 22758
+IHN1c3RhaW5lZA== 22759
+IGxlYW5pbmc= 22760
+IGFidW5kYW5jZQ== 22761
+w61saWE= 22762
+0LDQu9GM0L3Ri9C5 22763
+IHBo4bqjaQ== 22764
+YWNqYQ== 22765
+IOqwmeyVhA== 22766
+YWN0aXY= 22767
+4Liy4Lii 22768
+IDk3 22769
+INC80L7QuQ== 22770
+Y3Jv 22771
+IEphY2tpZQ== 22772
+aXR0ZWVz 22773
+YnJhY2h0 22774
+dWxlbnQ= 22775
+IOygnOs= 22776
+IHBsdWdpbg== 22777
+dmFudGFnZQ== 22778
+cGFydHk= 22779
+IHN1YXM= 22780
+IGFudGU= 22781
+0YPQuw== 22782
+0J3QkA== 22783
+IM+Dz4U= 22784
+IG1ldGg= 22785
+IGVudGh1c2lhc20= 22786
+0Y/RgtGB0Y8= 22787
+7ZmU6w== 22788
+IHN5bnRoZXRpYw== 22789
+IHNlYXNvbmluZw== 22790
+IExvc3Q= 22791
+b25vbXk= 22792
+IFNwYXJr 22793
+IGJ1cmU= 22794
+IGFzc3VyZWQ= 22795
+IGltYWdpbg== 22796
+IGNhcnJv 22797
+U2hh 22798
+xIV0 22799
+0L3Rg9GC0Yw= 22800
+w6F0aWNh 22801
+VFk= 22802
+IGtlcm4= 22803
+IEJyYXppbGlhbg== 22804
+w7A= 22805
+IHN1c3BlbmRlZA== 22806
+IENhcmli 22807
+IGJpemlt 22808
+IE9saXZlcg== 22809
+VG9t 22810
+INC/0LvQsNC9 22811
+IG5vcGU= 22812
+b21ldGhpbmc= 22813
+IGJlaWRlbg== 22814
+0YbQtdC9 22815
+IGZsdWN0 22816
+IM68zr/PhQ== 22817
+IGZhdGhlcnM= 22818
+IEJsYWtl 22819
+IHVwd2FyZA== 22820
+IERhc2g= 22821
+IExpbA== 22822
+IOyImOuPhA== 22823
+IHJldmVsYXRpb24= 22824
+IGVsZXZhdGVk 22825
+IEppYW5n 22826
+TEVE 22827
+IFRob21wc29u 22828
+INC80L7Qs9GD0YI= 22829
+0YHRgtGA0YM= 22830
+aWZpZXJz 22831
+IGNvbWViYWNr 22832
+IGJ1eWVycw== 22833
+6rKw 22834
+IFNhbGVz 22835
+0LjRh9C1 22836
+Y2lvbmVz 22837
+IHdoaXN0bGU= 22838
+IGR1bGw= 22839
+TEVY 22840
+IO2VmOqyoOyKteuLiOuLpA== 22841
+IGNyaW1pbmFscw== 22842
+IGRlc2NlbnQ= 22843
+aXBwbGU= 22844
+bWFzxLE= 22845
+IGZvb2xpc2g= 22846
+INC00YPQvNCw0Y4= 22847
+dGFy 22848
+IG1hbmdv 22849
+IGNob3Jlb2dyYXBoeQ== 22850
+TWF0dA== 22851
+IHRlcnJpdG9y 22852
+IGFjYWJh 22853
+IEVpbnN0ZWlu 22854
+IElCTQ== 22855
+IE1ldGFs 22856
+IENyeXN0YWw= 22857
+IHJhaA== 22858
+IGZvdWw= 22859
+IElzbGFuZHM= 22860
+IGludGFjdA== 22861
+IFJhaWw= 22862
+Ljo= 22863
+IGFjw6E= 22864
+INC/0YDQvtC/ 22865
+0LXRgNC1 22866
+IFdyaXRl 22867
+aGVoZQ== 22868
+IEZP 22869
+IM+Dz4TOtw== 22870
+IGRvaW4= 22871
+aGVsZA== 22872
+IGFwcHJvcHJpYXRlbHk= 22873
+IGRlbGliZXJhdGVseQ== 22874
+IGFyY2hpdmU= 22875
+IGdpdmVhd2F5 22876
+IGZpbmFsZQ== 22877
+0LvQsNGB 22878
+0LXQvdC+ 22879
+xqFu 22880
+b2dv 22881
+IEF1ZGllbmNl 22882
+44Wg 22883
+IHN1YnVy 22884
+IGhlYWRhY2hl 22885
+0LDQvdC90Y8= 22886
+IFdpdGNo 22887
+IFN3ZWRpc2g= 22888
+IEJJ 22889
+IGVyYXNl 22890
+IGtoaQ== 22891
+IGNvbW1lbnRhcnk= 22892
+IFN1bHRhbg== 22893
+7YOd 22894
+IExlYmFu 22895
+IOuztOyL 22896
+IFBhbQ== 22897
+cGVrdA== 22898
+bW9udGg= 22899
+IGdyb3VuZGVk 22900
+6r4= 22901
+IMWfZWtpbGRl 22902
+MjUw 22903
+IFNDSA== 22904
+aW9zbw== 22905
+IGluYXVn 22906
+aGVpbWVy 22907
+IHJlZmxlY3Rpbmc= 22908
+IFJ1dGg= 22909
+IE9pbA== 22910
+IHRyb3V2ZXI= 22911
+dWVw 22912
+Li5d 22913
+IOyeiOs= 22914
+IG9saGE= 22915
+IHJlYXNvbmFibHk= 22916
+IGdsaXRjaA== 22917
+VUI= 22918
+IEdyYW4= 22919
+IGFkYWxhaA== 22920
+IGxlbnQ= 22921
+2LHYpw== 22922
+IHRyYWN0aW9u 22923
+IGFkanVzdGluZw== 22924
+tKQ= 22925
+0L3QuNCx0YPQtNGM 22926
+INC00L7Qvw== 22927
+IHN0cmV0Y2hlZA== 22928
+IG9ydA== 22929
+IGNvc2luZQ== 22930
+dmlvbA== 22931
+IOyF 22932
+Y2ly 22933
+IGJhc3RhcmQ= 22934
+INGF0L7QtA== 22935
+IHF1aWVy 22936
+IHByZXNzdXJlcw== 22937
+IEFuaA== 22938
+IGVsbGVz 22939
+INC00YDRg9C3 22940
+INC80L7QttC10YLQtQ== 22941
+IGNo4bs= 22942
+IE3DqQ== 22943
+w7Zr 22944
+4bqndQ== 22945
+7KCI 22946
+emlu 22947
+IGNhdXRpb24= 22948
+aWJhbg== 22949
+IGp1ZGdpbmc= 22950
+0YPRjtGC 22951
+IGJhag== 22952
+INCh0LXQudGH0LDRgQ== 22953
+IFBvb3I= 22954
+IE5hemk= 22955
+IHVwYmVhdA== 22956
+eWFuZw== 22957
+IHdlZWtlbmRz 22958
+IEVzc2VudGlhbGx5 22959
+IG9sdXlvcg== 22960
+IHNwYXRpYWw= 22961
+YWNrZXI= 22962
+IHNlbGxlcg== 22963
+INeQ15XXqg== 22964
+kdec 22965
+IHZpdmlk 22966
+IEJvbmQ= 22967
+6raM 22968
+aXNrdA== 22969
+IGdvYXQ= 22970
+ZHJpdmVy 22971
+IG11Zw== 22972
+aWN0aW9uYWw= 22973
+IGFsbHQ= 22974
+IEluaXRp 22975
+IFJhbmQ= 22976
+IGZpbmlzaGVz 22977
+IOqwiA== 22978
+IHZpdGFt 22979
+IHRlZW5hZ2Vycw== 22980
+IE1vcnJpcw== 22981
+7KSE 22982
+IE9yaQ== 22983
+aXlh 22984
+IG15w7Zz 22985
+U3RlcA== 22986
+IEtyZQ== 22987
+IGRpbm9zYXVy 22988
+IOuqhw== 22989
+YWZmZQ== 22990
+IOuQqeuLiOuLpA== 22991
+IHplZw== 22992
+IE1hbmhhdHRhbg== 22993
+IHN1amV0 22994
+dWVsbGU= 22995
+c3RvZmY= 22996
+IGTDvHI= 22997
+IHN1Ym1hcg== 22998
+ZXNlcw== 22999
+IGFxdWVsZQ== 23000
+IG5vdQ== 23001
+IEZhaXRo 23002
+dHo= 23003
+INGC0L7QvNGD 23004
+YWNldXQ= 23005
+bGllcnM= 23006
+IGJhbmR3aWR0aA== 23007
+xrDhu50= 23008
+IHJlc3BlY3RpdmU= 23009
+IEF2ZQ== 23010
+IHNwcmVhZHNoZQ== 23011
+IFNlbnQ= 23012
+aWNhbWVudGU= 23013
+IGluZnJh 23014
+IGxlYXJuZXJz 23015
+IOCuiQ== 23016
+YWlhaA== 23017
+cmVuYWw= 23018
+IG11c3RhcmQ= 23019
+IGhhYnQ= 23020
+54M= 23021
+IFF1w6k= 23022
+IGFuYWx5emluZw== 23023
+IHNvbGlj 23024
+INeU15XXkA== 23025
+IGNhdXNh 23026
+IHdlbGNvbWVk 23027
+IFN1Y2Nlc3M= 23028
+IGZhY2lsZQ== 23029
+INCf0L7RgtC+0LzRgw== 23030
+c2NoZWlu 23031
+IGZldGNo 23032
+IHN0cmF0 23033
+INGB0YLQvtC40YI= 23034
+7JeQ7ISc64qU 23035
+INGB0L/QvtGB0L7QsQ== 23036
+bWFt 23037
+IHNlcsOtYQ== 23038
+bmFtZW50cw== 23039
+d3JpdGVy 23040
+IGNvbnN1bHRpbmc= 23041
+7ZiA 23042
+IEJlcmtlbGV5 23043
+ZXU= 23044
+YXNpdmU= 23045
+VVU= 23046
+IEFuYWx5dA== 23047
+IHN1Ym1pc3Npb24= 23048
+IG1hZ25pZmljZW50 23049
+ZW56YQ== 23050
+IGVjb24= 23051
+IHByb2ZpbGVz 23052
+IGluY2Fy 23053
+QWI= 23054
+IE51bg== 23055
+IGhpYw== 23056
+c2NyZWFtaW5n 23057
+IHJlc2lsaWVudA== 23058
+Z3J1bmQ= 23059
+IGNvbmN1cg== 23060
+IGJlcmVpdHM= 23061
+TEQ= 23062
+IG51cnQ= 23063
+7Ik= 23064
+IGZlYXN0 23065
+IGVuY3VlbnQ= 23066
+IE1pY2hlbA== 23067
+IHN1cHJlbQ== 23068
+Il0= 23069
+IGZlZWRz 23070
+IEtvbGxlZ2Vu 23071
+aXNzZXI= 23072
+IEZlbmc= 23073
+IFdlbg== 23074
+bXVu 23075
+IHRlbsOtYQ== 23076
+IFdyZXN0 23077
+IOyYpOuKmOydgA== 23078
+IHN0ZWFk 23079
+IHJlc3RvcmF0aW9u 23080
+IGRvbmF0ZWQ= 23081
+IGRlbHM= 23082
+IGNlbnN1cw== 23083
+IGRlc3BlcmF0ZWx5 23084
+d29ydGh5 23085
+SEU= 23086
+IFNwYQ== 23087
+IEJyeWFu 23088
+IGhq 23089
+IFJhdw== 23090
+7JWE6w== 23091
+IENhbWVyYQ== 23092
+IHppZW4= 23093
+IHN0eWw= 23094
+IFRX 23095
+IENoZWVzZQ== 23096
+Ym9ybmU= 23097
+IG9ibA== 23098
+IEFscmVhZHk= 23099
+IHVuc3RhYmxl 23100
+IGZsYW1lcw== 23101
+cG9zdA== 23102
+SGE= 23103
+cm9tYWdu 23104
+IOyXhOuniA== 23105
+ZGVzdA== 23106
+IGtvbGVq 23107
+IHRlbXBvcmFyaWx5 23108
+IGRldGVybWluaW5n 23109
+IEdsYXNz 23110
+0YDQvtC9 23111
+b2xhbg== 23112
+IGRvbWluYXRlZA== 23113
+X19fXw== 23114
+INmH2LDYpw== 23115
+IERhbmE= 23116
+IGRpbmhlaXJv 23117
+YXF1 23118
+66+8 23119
+IMOgcw== 23120
+IEpvZXk= 23121
+IEdyaWZm 23122
+IGF0dGFpbg== 23123
+IHRyYW5zaXRpb25z 23124
+IExpdGVyYWxseQ== 23125
+0LXQvdC0 23126
+IEhhdmVu 23127
+IGdyYWJiaW5n 23128
+IGNyeXN0YWxz 23129
+IEZvdXJ0aA== 23130
+IGNhbmRsZXM= 23131
+INGB0LvRg9GH0LA= 23132
+cmljbw== 23133
+IDUwMDA= 23134
+ZXR0bw== 23135
+IHVuZG8= 23136
+IGt0bw== 23137
+IGRpdmVydA== 23138
+IGNoaXI= 23139
+IHBlcnNlYw== 23140
+IGhpa2luZw== 23141
+IGFubm91bmNlbWVudHM= 23142
+0LfRiw== 23143
+IGF1Yw== 23144
+IHN5c3RlbWlj 23145
+IFJN 23146
+z4POsQ== 23147
+INCU0LY= 23148
+IHlhcg== 23149
+IFdhcmQ= 23150
+IHBpc3NlZA== 23151
+IGNhcm4= 23152
+IGF1dG9ub21vdXM= 23153
+44WO44WO 23154
+c292ZXI= 23155
+IHJlZmxleA== 23156
+IGdhcmRlbnM= 23157
+IGRhdGVk 23158
+7LE= 23159
+YW1pxJk= 23160
+IGNvbnRpbnVpdHk= 23161
+IGNpdGl6ZW5zaGlw 23162
+IHNjaHdlcg== 23163
+IHphaw== 23164
+dGFibGU= 23165
+INGB0Yc= 23166
+IM+DzrU= 23167
+IGdlbmVyYXRlcw== 23168
+6rWs64KY 23169
+w7Zo 23170
+w7Nt 23171
+YWxhbQ== 23172
+IEpVRFk= 23173
+IEJ1Zw== 23174
+IGRyb25lcw== 23175
+IMOhZ3Vh 23176
+YWNha3M= 23177
+5po= 23178
+INCa0L7QvQ== 23179
+15bXlA== 23180
+IHN0cml2ZQ== 23181
+IEFsdGVybg== 23182
+IG5lYXJlc3Q= 23183
+IHByb3llY3Q= 23184
+dGVyYQ== 23185
+IEFTSExFWQ== 23186
+IHdvcm0= 23187
+IHJlcGxheQ== 23188
+IHRhcmE= 23189
+IEluZGlhbnM= 23190
+aWNhaWQ= 23191
+IOyInA== 23192
+IGFwcGVhbGluZw== 23193
+IFdlcw== 23194
+IG1lbnRpb25z 23195
+INC00LXQu9C1 23196
+IGt3 23197
+IGZyYWdpbGU= 23198
+aXN6 23199
+a8Ozdw== 23200
+aGFuZw== 23201
+Y29sb3I= 23202
+IHByZXNpZGVudGU= 23203
+ODc= 23204
+0LXRhA== 23205
+INC00L7QsdCw0LI= 23206
+IE5lbHNvbg== 23207
+w6FmaWM= 23208
+IE1JQ0hBRUw= 23209
+IG1lY2hhbmlj 23210
+IG1ldHJlcw== 23211
+IG9jenl3acWbY2ll 23212
+IENpbmQ= 23213
+IG9nc8Ol 23214
+IGxhbmRzY2E= 23215
+QUNF 23216
+IGhlYWRsaW5lcw== 23217
+IGNhdGFseXN0 23218
+IENhdGNo 23219
+aW5rbGVz 23220
+IHBpbGxz 23221
+b3Jkbw== 23222
+IGltbWlncmFudA== 23223
+IGV4YW1pbmF0aW9u 23224
+IGFjY2lkZW50cw== 23225
+esSFZA== 23226
+IHF1aWVyZQ== 23227
+IG5lbGxh 23228
+IDY3 23229
+IHBhc3Nh 23230
+IHN1cGVyZmlj 23231
+aXN0b3I= 23232
+IG5vdg== 23233
+64u1 23234
+IG1hbmRhdGU= 23235
+aXNvbnM= 23236
+IFZpcnR1YWw= 23237
+IHNlbGJlcg== 23238
+IGNvdW5zZWxpbmc= 23239
+IE5CQQ== 23240
+IHNlcHQ= 23241
+IGJlbGlldmVy 23242
+IG1hcnZlbA== 23243
+IEludGVncg== 23244
+INC80ZY= 23245
+IG9ycGg= 23246
+IGJhY2t3YXJk 23247
+IEdlbmVyYXRpb24= 23248
+IFBpY3Q= 23249
+INGC0L7Rgg== 23250
+IHRhcGk= 23251
+cHJvY2hlbg== 23252
+IGhhbGx3YXk= 23253
+aHRl 23254
+INuB25I= 23255
+IFp1bQ== 23256
+YWNobWVudA== 23257
+aXF1ZXI= 23258
+Zm9sZw== 23259
+IEVkZGll 23260
+IEtpbA== 23261
+IHdlbGxuZXNz 23262
+c3RvY2s= 23263
+IGthw6c= 23264
+IHRlcnJvcmlzbQ== 23265
+IHBvaW50ZXI= 23266
+T2Y= 23267
+aGVyaWM= 23268
+IFVsdGltYXRlbHk= 23269
+IG1lc2Vz 23270
+IFRyYWRl 23271
+IHBpbnQ= 23272
+IHR1aXRpb24= 23273
+IGRpc2FncmU= 23274
+IOqyjOyehA== 23275
+IG1hbnVzY3JpcHQ= 23276
+IHJvb21t 23277
+IG91dHB1dHM= 23278
+0LXRhtC4 23279
+IHJpZXM= 23280
+IHNhbHVk 23281
+b3R6ZGVt 23282
+IG1hc3Nlcw== 23283
+IGJ5xYJh 23284
+IGNsZWFyaW5n 23285
+IGRpc2NvdXJzZQ== 23286
+YXRzb24= 23287
+IGZvbGRlZA== 23288
+IEphcg== 23289
+2YTZiQ== 23290
+OTAw 23291
+INGD0YHQvw== 23292
+IHByb3BoZWN5 23293
+IGludGVyZmVyZQ== 23294
+0LjRhdC+0LQ= 23295
+4LmM 23296
+IHRocmk= 23297
+INee16k= 23298
+IGxhesSxbQ== 23299
+IDE5OTI= 23300
+IGZ1dHVybw== 23301
+IGxvY2tpbmc= 23302
+IGVtYmFyZ28= 23303
+IE5laXRoZXI= 23304
+aXZhbWVudGU= 23305
+IG3DpXN0ZQ== 23306
+IG1paw== 23307
+IGNvbGxlY3Rvcg== 23308
+0LXQutC+0YLQvtGA 23309
+IEdhbmQ= 23310
+IHNlbnRpcg== 23311
+IE1pZ2h0 23312
+IGdhbnplbg== 23313
+VUM= 23314
+IHJlbGF0aW5n 23315
+U0Q= 23316
+IG1vc3F1aXRv 23317
+R1I= 23318
+IGhvbGxvdw== 23319
+4piF 23320
+IFdhbGtlcg== 23321
+IGFmZmlsaWF0ZQ== 23322
+IGR1cGxpY2F0ZQ== 23323
+0L3QtdC8 23324
+IGdyYXBl 23325
+IE9yZ2FuaXphdGlvbg== 23326
+IHN5bnQ= 23327
+Sm9l 23328
+IGdlZw== 23329
+IHJldmVhbGluZw== 23330
+IEV0aGFu 23331
+b3V0ZXI= 23332
+IHlheQ== 23333
+0LvQsNGA 23334
+IHJlcG9ydGVkbHk= 23335
+IGlocmVy 23336
+IHJlY29nbmlzZQ== 23337
+IGJ1bXBlcg== 23338
+IFJhbmR5 23339
+IFZlbnVz 23340
+dGxlcw== 23341
+IGFwcGV0aXRl 23342
+IGdsdWNvc2U= 23343
+IGNob2R6aQ== 23344
+IEZ1cnRoZXJtb3Jl 23345
+dGly 23346
+IGNvbnRh 23347
+IGludHVpdGlvbg== 23348
+IGFsdGl0dWRl 23349
+IGNodW5rcw== 23350
+IEpvc2h1YQ== 23351
+xLHEn8SxbQ== 23352
+cnlsaWM= 23353
+bGVhbnM= 23354
+IO2UvOs= 23355
+TEw= 23356
+UXVl 23357
+IGdvcg== 23358
+INC30L3QsNGH0LjRgg== 23359
+IHBvZW1z 23360
+IGV4Y2Vs 23361
+IGV4cGxvcmVk 23362
+IHBvcHVs 23363
+IGluY2x1c28= 23364
+c3TDpA== 23365
+IEdhdmlu 23366
+YWxsaW5n 23367
+IM+Ezr/OvQ== 23368
+6ak= 23369
+YXJiZWl0 23370
+IEdhcw== 23371
+IGdsb3Jpb3Vz 23372
+cmllYmVu 23373
+IHNwYW0= 23374
+IGluZG9vcg== 23375
+IHRocnVzdA== 23376
+IEFsZA== 23377
+IFByaW9y 23378
+IG9uYm9hcmQ= 23379
+b2Nh 23380
+QVNI 23381
+o6A= 23382
+IENocmlzdGluZQ== 23383
+IGRyYXdlcg== 23384
+IG5vb24= 23385
+IOyemOs= 23386
+IHBlcm1hbmVudGx5 23387
+INC90LDQv9GA0LjQvNC10YA= 23388
+IHBvZGNhc3Rz 23389
+ZXJhcGV1dA== 23390
+cHJpdA== 23391
+IHN0YWlubGVzcw== 23392
+INqp25I= 23393
+IGZhbWlsaWE= 23394
+INGA0LDQt9GA 23395
+dW50bw== 23396
+INGB0YLQvtC7 23397
+IGjDpA== 23398
+IEhhaQ== 23399
+IFBC 23400
+aXpvbg== 23401
+IGtvbm50ZQ== 23402
+IGLDvHnDvGs= 23403
+IHV0aWxpemFy 23404
+2oY= 23405
+IGFxdWVzdGE= 23406
+IG1peGVy 23407
+dWRlbnQ= 23408
+0LvQtdC60YE= 23409
+xYJ1 23410
+INGB0LjRgdGC0LXQvA== 23411
+INC90L7RgNC8 23412
+IGZhdGFs 23413
+IGNvbnNpZGVyYXRpb25z 23414
+IHZhbGlkYXRpb24= 23415
+IG9saQ== 23416
+IGthcmRlxZ8= 23417
+IEdMT1JJQQ== 23418
+IHBhbGw= 23419
+0LXRgdGC0LU= 23420
+IHJlY3Rhbmc= 23421
+IG1lZGlldmFs 23422
+YWxsYWhp 23423
+YXN0aQ== 23424
+IFN5cmlhbg== 23425
+IHNoZWFy 23426
+IGRlYnVn 23427
+IE1haQ== 23428
+IGtub2NraW5n 23429
+IExleA== 23430
+YXJkYW4= 23431
+cm92 23432
+IG1lbW9yaWFs 23433
+b29reQ== 23434
+IHN0dWZmZWQ= 23435
+IHBhc3PDqQ== 23436
+IHdpZw== 23437
+gqA= 23438
+IHByw7N4aW1h 23439
+IDE5OTE= 23440
+INC80LXQttC00YM= 23441
+IG51ZXN0cm9z 23442
+IEJlYXN0 23443
+IHNtbw== 23444
+YXRjaGVk 23445
+b2xvZ2lh 23446
+INC80L7QtA== 23447
+IGdlZQ== 23448
+IGNvbmNlcHR1YWw= 23449
+IMO0 23450
+IGRlY3JlYXNlcw== 23451
+IHF1ZXJpZXM= 23452
+0L7Qu9GM0Yg= 23453
+IEFwYXJ0 23454
+IGV4ZW1wbA== 23455
+IGZsZWQ= 23456
+IE9GRg== 23457
+Z2dhaw== 23458
+IGJlYWQ= 23459
+aGly 23460
+bGllcw== 23461
+IENsZWFybHk= 23462
+xLFsYXI= 23463
+IGNoZXNz 23464
+IHdoaWNoZXZlcg== 23465
+IDk2 23466
+4bqx 23467
+IHJlc3BlY3Rz 23468
+INC80L7RgA== 23469
+IG9yZ2FuaXNt 23470
+IGdyYW5kcGE= 23471
+IFZpZQ== 23472
+IGZsb29kaW5n 23473
+IHVwZ3JhZGVk 23474
+0ZHRgA== 23475
+IGNoZWVrcw== 23476
+IGNvbnF1ZXI= 23477
+IHN0dWJib3Ju 23478
+IHB1enpsZXM= 23479
+IGF1Y3Rpb24= 23480
+IHJlbHlpbmc= 23481
+IFBST0Y= 23482
+IEVzcGVy 23483
+INCc0KM= 23484
+IGh5cGU= 23485
+IHBvc3NpYmls 23486
+IGltcHJpc29u 23487
+IEVybg== 23488
+7JeI7Iq164uI64uk 23489
+IGVudmll 23490
+IHJlc3VycmVjdGlvbg== 23491
+IHNwZXI= 23492
+IFZlbmV6dWVsYQ== 23493
+c29t 23494
+IOyeoOq5 23495
+IG5vdXZlbGxl 23496
+IGNsb3Nlcw== 23497
+IDE5NDA= 23498
+IHF1YQ== 23499
+IEphcmVk 23500
+IFBpcg== 23501
+IGluZGU= 23502
+IHNjcnVi 23503
+dWt1 23504
+IHJlcXVpcmluZw== 23505
+INCy0LDQvNC4 23506
+IGNvbnNpZGVyYWJsZQ== 23507
+aWxpYQ== 23508
+IGlubmU= 23509
+IG1laW5lbQ== 23510
+IGhhcmRzaGlw 23511
+IHRyYXBz 23512
+cm9j 23513
+IOyEpOs= 23514
+IHJlc2VhcmNoaW5n 23515
+IE1hcmdhcmV0 23516
+IHBlbm55 23517
+IGLEsXJhaw== 23518
+0ZHQuw== 23519
+IHdvb2w= 23520
+IHJoZXQ= 23521
+IGZsYXR0ZW4= 23522
+54c= 23523
+4LmA4Lij 23524
+IHBpZWQ= 23525
+IENoYXA= 23526
+IHVuZGVybQ== 23527
+IGZyZXQ= 23528
+IGNyYXNoZWQ= 23529
+IEZyYXVlbg== 23530
+2LDZhw== 23531
+aXZhbg== 23532
+IGxpdGVyYXJ5 23533
+bGF0ZWdv 23534
+IHNww6R0ZXI= 23535
+IHNpbWlsYXJpdGllcw== 23536
+4oY= 23537
+IENvcm9u 23538
+IENyZWVr 23539
+IGJvc3Nlcw== 23540
+IGFjY29tcGFuaWVk 23541
+IGRlYmF0ZXM= 23542
+IGFzc2VtYmxlZA== 23543
+IMOB 23544
+IFZhaQ== 23545
+IHRyYWN0 23546
+IHNpbXBsZW1lbnQ= 23547
+IEFyaW4= 23548
+IHZ1bG5lcmFiaWxpdHk= 23549
+IGhvcm1vbmU= 23550
+SUVM 23551
+T09L 23552
+IHJlbGF5 23553
+IEFuZHJlYQ== 23554
+cmls 23555
+IG5lY2Vzc2l0eQ== 23556
+YWNldXRpY2Fs 23557
+0Y7RiQ== 23558
+b3VzaW5n 23559
+bmFobWVu 23560
+IGZvb3RwcmludA== 23561
+bWFw 23562
+IFRpZXI= 23563
+YW5ueWE= 23564
+aW50ZW5k 23565
+5aI= 23566
+IGRlY29yYXRl 23567
+IHpvbWJpZXM= 23568
+IEh5ZA== 23569
+IFN1eg== 23570
+IGNhbXB1c2Vz 23571
+IEVtYg== 23572
+IHRocm90dGxl 23573
+IGFkbWlu 23574
+IG9wb3J0dW4= 23575
+IG1pcnJvcnM= 23576
+IGlkZW50aXRpZXM= 23577
+IENsaW4= 23578
+IOu5hOs= 23579
+4bmj 23580
+IE90dA== 23581
+IGJsdWVz 23582
+IGltcHJlc3Npb25z 23583
+LSw= 23584
+IHZhZ3Vl 23585
+YWZl 23586
+IGluZmVyaW9y 23587
+ZXJhbGQ= 23588
+IG1lZGljaW5lcw== 23589
+IHByZWd1bnRh 23590
+b3NlbHk= 23591
+IHTDqWzDqQ== 23592
+IE1vbnRo 23593
+IExlYWRlcnM= 23594
+IEVneXB0aWFu 23595
+IHJhdGlvbg== 23596
+a2Vycw== 23597
+aGVpdHM= 23598
+IHJlY2h0 23599
+UGxheQ== 23600
+IGVn 23601
+IHBvbGxz 23602
+IFdPT0RS 23603
+IHNsb3Rz 23604
+amFt 23605
+Qm90aA== 23606
+IFJhdA== 23607
+0YDQsNC2 23608
+IEJyaWdodA== 23609
+4buRaQ== 23610
+dXJpb3Vz 23611
+IHNpbmdlcnM= 23612
+IGxvZ2lu 23613
+IHTDqm0= 23614
+bGF0aW9u 23615
+IE11bQ== 23616
+xrDhu51uZw== 23617
+IEVkaXRvcg== 23618
+IGlubm92YXRpb25z 23619
+aGF2ZQ== 23620
+IFNlaw== 23621
+IHdlYWtlcg== 23622
+IEdvYg== 23623
+QWZ0ZXI= 23624
+tOyngA== 23625
+IOusuOygnA== 23626
+IGRpc2FkdmFudGFnZQ== 23627
+IGdhemU= 23628
+IE1hY2s= 23629
+z4HOrw== 23630
+IEtpc3M= 23631
+IEhvbG8= 23632
+IEJpcnRo 23633
+aXpp 23634
+YmFi 23635
+7Iuc6rOg 23636
+0LTQtdGA0LY= 23637
+IHNxdWF0 23638
+0LrRg9GB 23639
+dW5p 23640
+IENvbW1l 23641
+IFdPT0RSVUZG 23642
+IENoYW1waW9uc2hpcA== 23643
+IHdlbGNoZQ== 23644
+IFlvdXRo 23645
+emVt 23646
+IG9kcG93 23647
+IHBlcnNpc3RlbnQ= 23648
+cnV0 23649
+7JSp 23650
+7Zal 23651
+bGFpcg== 23652
+aWt1 23653
+IHZlbmRvcg== 23654
+IGNow7puZw== 23655
+IGZpbmFuY2k= 23656
+IG92ZXJseQ== 23657
+w6J1 23658
+IGdsdXRlbg== 23659
+IDE4MDA= 23660
+IGRpdmlzaW9ucw== 23661
+IGNpdWRhZA== 23662
+IG9iZWQ= 23663
+IHdhcnVt 23664
+IGVoZXI= 23665
+IGVsaW0= 23666
+INCS0L4= 23667
+IHBldXZlbnQ= 23668
+IFdhbm5h 23669
+IGF0dGVuZGFuY2U= 23670
+IGFzc2Vzc21lbnRz 23671
+IEJvZw== 23672
+IGltYWdlcnk= 23673
+IGNvbGxlY3RpdmVseQ== 23674
+IGluZm9ybWFs 23675
+IFNjaHdl 23676
+IGRldXRsaWNo 23677
+IENoZWw= 23678
+IFBF 23679
+b3dlZA== 23680
+IGJhbm5lcg== 23681
+IHNoZWx2ZXM= 23682
+IFJldHVybg== 23683
+TEFVR0hT 23684
+IGNvbmdyYXR1bGF0ZQ== 23685
+IE5vcndheQ== 23686
+IGR3ZWxs 23687
+IENhcmliYmVhbg== 23688
+IG5vcm1z 23689
+IEFuaW1hbA== 23690
+IFZhbGVudGluZQ== 23691
+IGV4dGVuZGluZw== 23692
+IFZvdQ== 23693
+b3Jy 23694
+IENoZW5n 23695
+wqE= 23696
+INC00L7RgNC+0LM= 23697
+IHZlZw== 23698
+IGjDpQ== 23699
+IFhpbg== 23700
+IOy5tOs= 23701
+ZW1ldA== 23702
+IGh5cG90aA== 23703
+IGludGVyZXNzYW50ZQ== 23704
+cmljZXM= 23705
+SVo= 23706
+IFVTRA== 23707
+IHJ1bm5lcg== 23708
+IEJhZw== 23709
+IOq9 23710
+IGNvbWXDp2Fy 23711
+IHBpZ3M= 23712
+IHdlYWtuZXNzZXM= 23713
+UGg= 23714
+IFZpb2w= 23715
+IGRyYWdnaW5n 23716
+IEFxdcOt 23717
+IENTUw== 23718
+IG1pbGxpbWV0ZXJz 23719
+IGVzdMOhcw== 23720
+IGFjdXRl 23721
+IGRlamFy 23722
+acSf 23723
+b2JyYQ== 23724
+TG92ZQ== 23725
+IHNpbGs= 23726
+KioqKg== 23727
+IGpvaW5z 23728
+IHByb2w= 23729
+IOqwkOyCrO2VqeuLiOuLpA== 23730
+2K3Yrw== 23731
+YWdoZXR0aQ== 23732
+w6RubmVy 23733
+IHN0cmFuZw== 23734
+IGRvdWJsZWQ= 23735
+IGRlc2NyaXB0aW9ucw== 23736
+IHN0ZWxsZW4= 23737
+IHBhcnRp 23738
+soTr 23739
+IMO2xJ8= 23740
+aWdoaW5n 23741
+IGFuZ3VsYXI= 23742
+IG5hdHV1cg== 23743
+IFNoZWw= 23744
+xrDGoQ== 23745
+IHJheXM= 23746
+IHNlcGVy 23747
+c3RhcnQ= 23748
+dmlzZWQ= 23749
+IHJ1c2hlZA== 23750
+IGludGVybmF0aW9uYWxseQ== 23751
+IG5pdmVs 23752
+IGJveGluZw== 23753
+ZmFsbGVu 23754
+4buRYw== 23755
+IHNlaW5lbg== 23756
+cGxpY2l0eQ== 23757
+IGNhcmJvaA== 23758
+IFRyYXZpcw== 23759
+dXNv 23760
+IFBoYXNl 23761
+IGFjdGl2YXRpb24= 23762
+IG9waW8= 23763
+t6g= 23764
+IGRlY3JlYXNlZA== 23765
+Q2Fy 23766
+IGJ1bmRsZQ== 23767
+IGV4cGVuZA== 23768
+b3JtYWw= 23769
+IGFkamFjZW50 23770
+IG1lZQ== 23771
+INC+0YDQsw== 23772
+IHRyYW5zY3JpcHQ= 23773
+IExhbmd1YWdl 23774
+R1M= 23775
+IHNldWw= 23776
+w6BuaA== 23777
+IG55YQ== 23778
+bmluZ3M= 23779
+IOyLnOs= 23780
+IOuUsOudvA== 23781
+IEFncg== 23782
+w61k 23783
+IGFieQ== 23784
+IE5lbw== 23785
+xLF5b3J1eg== 23786
+IFRoaW5raW5n 23787
+YWltZQ== 23788
+IHZpdGU= 23789
+IHRyYXbDqXM= 23790
+INeR16I= 23791
+INC80LXQtA== 23792
+T3Vy 23793
+aG9vdA== 23794
+IGxpbmVy 23795
+IFBpenph 23796
+IGh5Zw== 23797
+ZmxpZXM= 23798
+IENvbnRpbnVl 23799
+IGRlbnRhbA== 23800
+IFRpYg== 23801
+IHJlZ3VsYXRl 23802
+bGllw58= 23803
+QUxL 23804
+IFRhZQ== 23805
+6ri4 23806
+IEJyZXhpdA== 23807
+IEd1dA== 23808
+IG9jY3VwYXRpb24= 23809
+IHpyb2Jp 23810
+w6Jt 23811
+IHdoaXNr 23812
+IGthbnNrZQ== 23813
+b21vbg== 23814
+cm9iZQ== 23815
+IHdhcmZhcmU= 23816
+IHRo4buD 23817
+IGpha2k= 23818
+IHN0cm9rZXM= 23819
+IHBlYXM= 23820
+IERhbWl0 23821
+SEFO 23822
+IGludGVyZmVyZW5jZQ== 23823
+INC80LjQvdGD0YI= 23824
+TkVS 23825
+b3V0aW5n 23826
+IHRleHR1cmVz 23827
+n4k= 23828
+b3dp 23829
+IO2VmQ== 23830
+IGRlbnM= 23831
+IHByb3RhZ29uaXN0 23832
+w6Rubg== 23833
+IGdvZGRlc3M= 23834
+IHdvbGx0ZQ== 23835
+aWpv 23836
+IFdvY2hl 23837
+IFZQTg== 23838
+c3Rvcnk= 23839
+IGtpbmRlcmc= 23840
+IGZ1bm5lbA== 23841
+IGRpc3RyZXNz 23842
+0L3QvtGB0YLRjNGO 23843
+IG5vaXN5 23844
+INC/0YDQvtC00L7Qu9C2 23845
+IGRhcmFu 23846
+IGVuenltZQ== 23847
+0LvQvtC2 23848
+IG11dGU= 23849
+IGR3YXI= 23850
+INin2LM= 23851
+IGtvbXBs 23852
+IG1lcml0 23853
+IGZvc3Nl 23854
+IERyaW5r 23855
+IGZvcmE= 23856
+IHdvaGw= 23857
+IGJyZWV6ZQ== 23858
+IHNhbml0 23859
+IGRyaW4= 23860
+IOydtOqxsOuKlA== 23861
+IDYy 23862
+IOywqOs= 23863
+YWJ5dGVz 23864
+IGRlZWRz 23865
+INC5 23866
+acOobWU= 23867
+aWdnbGluZw== 23868
+ICIn 23869
+INGH0LDRgdGC0Yw= 23870
+IEFuc3dlcg== 23871
+IGV2YW5nZWw= 23872
+IDEwODA= 23873
+IFZpc2l0 23874
+aWNpZW50 23875
+IHJlbGlhYmlsaXR5 23876
+0Y7RgdGM 23877
+IEVhcmxpZXI= 23878
+IGZpZA== 23879
+IHNsZWV2ZXM= 23880
+aXlvcnN1bg== 23881
+IGJpYg== 23882
+IEFjY291bnQ= 23883
+0Y/Qu9C4 23884
+Y2lwbGluYXJ5 23885
+emFz 23886
+INCx0LXRgA== 23887
+IG5lY2tsYWNl 23888
+IGJsZW5kZXI= 23889
+IFBoaWxsaXBz 23890
+ZXRp 23891
+IEp1cGl0ZXI= 23892
+IHByb3ZvYw== 23893
+IFllYXJz 23894
+ZW50cmU= 23895
+YWNpbw== 23896
+IGvDvA== 23897
+IGFudGVubmE= 23898
+IG5vdmVscw== 23899
+IGZhcnQ= 23900
+IFN1Z2Fy 23901
+IEp1ZHk= 23902
+IGNvbGxhcHNlZA== 23903
+57A= 23904
+cml0aXM= 23905
+IOyDge2ZqQ== 23906
+0JfQqw== 23907
+IFZlcmY= 23908
+cmFuZWFu 23909
+ZXJldW0= 23910
+IFRhcmdldA== 23911
+IDg4 23912
+INCY0Lc= 23913
+aWRlbw== 23914
+IHJlZ3Jlc3Npb24= 23915
+7Lac 23916
+IG3Ds3dp 23917
+IHN0dWRpb3M= 23918
+aWVucw== 23919
+aXBo 23920
+IGZyeWluZw== 23921
+IGZhc2NpbmF0ZWQ= 23922
+IFdhaA== 23923
+YnVja3M= 23924
+bWF5YQ== 23925
+IFNhdHVybg== 23926
+IE1vbW15 23927
+IHJhdGluZ3M= 23928
+IGF1dHVtbg== 23929
+xrDGoW5n 23930
+IGxvc2Vy 23931
+IGNlbnRybw== 23932
+w6lyaWV1cg== 23933
+IEZvbGQ= 23934
+IHN1cGVydmlzb3I= 23935
+IE5vYmVs 23936
+IHVuZGVyZXN0 23937
+b2JpYQ== 23938
+INCy0YHRjw== 23939
+IHZlcnc= 23940
+IGZ1ZWxz 23941
+IGFydGlmYWN0cw== 23942
+IOu2mQ== 23943
+IEF1dG9t 23944
+25Q= 23945
+15XXoQ== 23946
+IGlobmVu 23947
+IDU5 23948
+b3VuZGluZw== 23949
+0LXRgNGL 23950
+aW5hcnM= 23951
+Y2hhbnQ= 23952
+IGFkZGljdGVk 23953
+IGV4cGxvc2l2ZQ== 23954
+IGRpc3BlcnM= 23955
+4paI 23956
+YXhpcw== 23957
+QVJZ 23958
+IGx1bQ== 23959
+INGD0YHQuw== 23960
+INiM 23961
+IHJ1cGVlcw== 23962
+IFBlYXJs 23963
+Y2FtcA== 23964
+dHY= 23965
+b3lh 23966
+IGNvbmNsdWRlcw== 23967
+IGNvbGxpc2lvbg== 23968
+IGJ1eWVy 23969
+IHBsYXlncm91bmQ= 23970
+IHNwcmluZ3M= 23971
+IGZlbWluaW5l 23972
+IFJhcw== 23973
+IGluY2FyY2Vy 23974
+7ZeY 23975
+IGRpYWxlY3Q= 23976
+IGNsb3N1cmU= 23977
+IGNoYXR0aW5n 23978
+IGJhYmU= 23979
+IHNwb3RsaWdodA== 23980
+IG5vdGF0aW9u 23981
+U3Rhcg== 23982
+acOjbw== 23983
+IHTDqnRl 23984
+IHRpZGU= 23985
+IGp1bnRv 23986
+IHNlbmF0b3I= 23987
+0KU= 23988
+IGV4Y3VzZXM= 23989
+IGJsaW5r 23990
+IGFkbWlzc2lvbg== 23991
+IExpbHk= 23992
+0YvQvNC4 23993
+IGFtaWdv 23994
+IGx1c3Q= 23995
+64us 23996
+IGFtaW5v 23997
+IGNvbnN1bHRhbnQ= 23998
+IEVsZWN0cmlj 23999
+IOuFuOuemA== 24000
+dWphaA== 24001
+IHNob290ZXI= 24002
+aWNodGVu 24003
+IFVrcmFpbmlhbg== 24004
+IGFpbXM= 24005
+IEVudGVydGFpbg== 24006
+IG1pcmFjbGVz 24007
+IHplaWdlbg== 24008
+IGxhbQ== 24009
+IHJlc3M= 24010
+IEppbGw= 24011
+eWxhbg== 24012
+IHJvb2s= 24013
+IGhheWE= 24014
+IHBhc3Nwb3J0 24015
+YWRhdGE= 24016
+IGp1aWN5 24017
+Y29uZg== 24018
+0LvQtdC5 24019
+IFN6 24020
+IGludGVyY2VwdA== 24021
+IFRlYW1z 24022
+IG1ha2Vu 24023
+aXJyZWw= 24024
+IExJS0U= 24025
+4bqteQ== 24026
+6rWw 24027
+IHNob3J0YWdl 24028
+IHBhcmFkaWdt 24029
+IHBhcGVs 24030
+IGFzdGVybw== 24031
+IHNvbGxlbg== 24032
+IE1pY2tleQ== 24033
+IE9ybGVhbnM= 24034
+IGNob2xlc3Rlcm9s 24035
+IGdvb3Nl 24036
+0YbQuNGO 24037
+IEZM 24038
+INCz0L7Qu9C+0LI= 24039
+IHRyaWJ1dGU= 24040
+IEdhbQ== 24041
+IMOpdmlkZW1tZW50 24042
+0Y/RhQ== 24043
+IGluYXBwcm9wcmk= 24044
+dWhhbg== 24045
+IG9yZ2FuaXphdGlvbmFs 24046
+YWlsZWQ= 24047
+IGVuZHVyZQ== 24048
+IDc2 24049
+IHNob3RndW4= 24050
+IGxpdnJl 24051
+IHN1aXRlZA== 24052
+IHdhcm10aA== 24053
+IFNJTQ== 24054
+IGVudmlzaW9u 24055
+IGRlZ3JhZA== 24056
+w65uZQ== 24057
+TGF1Z2hpbmc= 24058
+IFdob2V2ZXI= 24059
+IEJ1ZGRoaXNt 24060
+IHNwcmlua2xl 24061
+Y2XEn2l6 24062
+IHJ1aW5z 24063
+IHN0YXJjaA== 24064
+IEhlcno= 24065
+IGluanVzdGljZQ== 24066
+IGh1bWlkaXR5 24067
+0L7QttCw0LvRg9C5 24068
+IE9iamVjdA== 24069
+IElnbg== 24070
+IEV4YW0= 24071
+aWdlcnM= 24072
+IHRob3U= 24073
+IFNveQ== 24074
+aXZhcw== 24075
+IHBvbGVz 24076
+bWF0aA== 24077
+INCy0L3QuNC8 24078
+SU5HSU5H 24079
+ZWRyYWw= 24080
+IGV4cGxvcg== 24081
+IHJvYXN0ZWQ= 24082
+IGNyYXds 24083
+IGNvZmY= 24084
+IGFub20= 24085
+IHdpag== 24086
+IGltcHJvdmVz 24087
+IHRyZWF0eQ== 24088
+IGRpc2NvdmVyaW5n 24089
+IHN0YXR1dGU= 24090
+IG1lcmNhZG8= 24091
+INGB0LjQuw== 24092
+IGludGVs 24093
+IENoYW5jZWxsb3I= 24094
+IE1lZGljYWlk 24095
+dWdp 24096
+IHZlcmJhbA== 24097
+IGTDtm4= 24098
+IHNjcmlwdHVyZQ== 24099
+IGl0ZXJhdGlvbg== 24100
+ZWtz 24101
+IE94Zm9yZA== 24102
+IHfDpGg= 24103
+IFZhZA== 24104
+IEFL 24105
+IOyVhOydtOs= 24106
+IGlldHM= 24107
+IG5lZWRsZXM= 24108
+2YPZhQ== 24109
+IHBhc2Fkbw== 24110
+IGFsYnVtcw== 24111
+IHllYQ== 24112
+ZXR6ZW4= 24113
+hOuPhA== 24114
+IGRldGVybWluZXM= 24115
+IHRoZWU= 24116
+IFBsYXlpbmc= 24117
+w6RydA== 24118
+INem 24119
+Y2xlZA== 24120
+IGRvd253YXJk 24121
+YWxvbmU= 24122
+IHNvbHU= 24123
+IHBhcnRpdGlvbg== 24124
+IHd6 24125
+ZGQ= 24126
+IHBlc3NvYWw= 24127
+IGZhY3Rvcmllcw== 24128
+IGJsZWlidA== 24129
+4Lih4Liy 24130
+YWxzYQ== 24131
+IE5GTA== 24132
+IGZ1ZXJh 24133
+IHJlc2VydmVk 24134
+IEVhcm4= 24135
+IGhlbHQ= 24136
+IHNob3J0Y3V0 24137
+IGNvbnZpbmNpbmc= 24138
+c3BhY2U= 24139
+IGVuZm9yY2U= 24140
+IGNvcmVz 24141
+IGVmdGVy 24142
+IHJlY2Vzc2lvbg== 24143
+eGljbw== 24144
+IHByb3Bvc2l0aW9u 24145
+YXJpYW5z 24146
+cm9wb2w= 24147
+IOuqsOs= 24148
+IM6c 24149
+IOyalOymmA== 24150
+IGFjdGl2aXN0 24151
+IGNvbnZpY3Rpb24= 24152
+IHphYg== 24153
+IGNhbmNlbGVk 24154
+0YLQvtGH0L3Qvg== 24155
+IM6u 24156
+bml0ZQ== 24157
+IGZ1bmRyYQ== 24158
+YnV6emVy 24159
+0LXQu9C+ 24160
+aWNhdGlvbnM= 24161
+IHpvbmE= 24162
+IHRlZW5z 24163
+IG1ldGhvZG9sb2d5 24164
+IOykkeyalA== 24165
+dGhhbg== 24166
+IFVs 24167
+IEdyZXk= 24168
+IGhvZw== 24169
+SU5L 24170
+IFN1bmc= 24171
+IENsYXVk 24172
+IENOTg== 24173
+IGRlbGl2ZXJz 24174
+YWxpbg== 24175
+IEFkb2Jl 24176
+b3RoZQ== 24177
+IERlc3dlZ2Vu 24178
+4Liz 24179
+IHdlcmRl 24180
+IGdyZWFzZQ== 24181
+IHVwZ3JhZGVz 24182
+IEZpbmxhbmQ= 24183
+YWNjZXB0 24184
+IGludGVycm9n 24185
+YmVl 24186
+IHByZWRl 24187
+IE5lcA== 24188
+IENhbWJyaWRnZQ== 24189
+IGdyYXBocw== 24190
+IGhhdW50ZWQ= 24191
+0YHQtdC8 24192
+5qc= 24193
+U29tZQ== 24194
+IE1hbGw= 24195
+IHJlaGVhcnNhbA== 24196
+IFVyYmFu 24197
+IExhZw== 24198
+IG5pbQ== 24199
+6rCV 24200
+IHBvc2l0aW9uZWQ= 24201
+IGF2b2lkZWQ= 24202
+RU1B 24203
+IGxsZWdhcg== 24204
+IHLDoXBpZG8= 24205
+IGdvdXZlcm4= 24206
+IGhpbmc= 24207
+IGRlYWxlcg== 24208
+IHJlZm9ybXM= 24209
+IGZhdHR5 24210
+0LrQvtC7 24211
+IEFjZQ== 24212
+IG5lcA== 24213
+IOyyrQ== 24214
+IGNvbXB1dGF0aW9u 24215
+IFN0cmVhbQ== 24216
+Ym91cm5l 24217
+dHVy 24218
+UG9y 24219
+IHNsZWVweQ== 24220
+IGJhbmdldA== 24221
+IHdlaWdocw== 24222
+IGJsZWliZW4= 24223
+IEdyZW4= 24224
+IHVuaW9ucw== 24225
+IOq1kA== 24226
+IGFwcmVuZGVy 24227
+dWl0YXI= 24228
+IEplc3Q= 24229
+dW1pbmc= 24230
+IFBsYXllcg== 24231
+IEV4dHJlbQ== 24232
+IGludGVnZXI= 24233
+0LDRh9C1 24234
+IGNvbmNlcnRz 24235
+15XXmw== 24236
+IHRyb2NoxJk= 24237
+IFJlcGU= 24238
+4LmC 24239
+xbxlbg== 24240
+IHNvdW5kaW5n 24241
+IGFub255bW91cw== 24242
+IGV4Y2E= 24243
+IElyYW5pYW4= 24244
+IGVuZXJnZXRpYw== 24245
+IHdpdmVz 24246
+INGG0LLQtdGC 24247
+IGFpcw== 24248
+IHN1ZGFo 24249
+IHVuZGVyd2Vhcg== 24250
+IGNydW5jaHk= 24251
+IFBhaW4= 24252
+IGdlcsOnZWs= 24253
+cmVkaWN0 24254
+IG1pc21h 24255
+0ZbRgg== 24256
+IHN1cnZpdmluZw== 24257
+zq3Pgg== 24258
+IHBhcnRpY2lwYW50 24259
+IEhlc3Nlbg== 24260
+w6FyaWFz 24261
+IHN1YndheQ== 24262
+aXN0w6Q= 24263
+IGNvcmFs 24264
+IG1hcmlqdWFuYQ== 24265
+IE1lbW9yaWFs 24266
+0YjQuNC5 24267
+cml6 24268
+IHNhdGVsbGl0ZXM= 24269
+IGxlYXNl 24270
+IENhbWVyb24= 24271
+dW1waA== 24272
+IGNsYXNzbWF0ZXM= 24273
+w6Row6Ru 24274
+0YHRgtCy0LU= 24275
+IGh1ZQ== 24276
+k6TsnYQ= 24277
+IHByb3BvcnRpb25hbA== 24278
+IG5vc3M= 24279
+IGxhcHM= 24280
+csOl 24281
+IGJpdGNvaW4= 24282
+0JfQq9Ca0JA= 24283
+IOy2qQ== 24284
+INmE2YQ= 24285
+IE1vcnQ= 24286
+IEVzcA== 24287
+YXJub3M= 24288
+INGB0LrQsNC30LDQuw== 24289
+IMOkbmQ= 24290
+15nXmded 24291
+IEdlYg== 24292
+Z2VoZW4= 24293
+SW5hdWRpYmxl 24294
+Ym9yb3VnaA== 24295
+0YTRhA== 24296
+IGZlbGxvd3NoaXA= 24297
+IFBhcGVy 24298
+IGN1cnZlZA== 24299
+IEdFT1I= 24300
+IGNhbGN1bGF0b3I= 24301
+IENhdGFs 24302
+IHbDoG8= 24303
+IGJ5cGFzcw== 24304
+0LvQtdGC 24305
+4LM= 24306
+dHJhbnM= 24307
+cmVuY2llcw== 24308
+7KGM 24309
+aWdlbnQ= 24310
+IHRhc3RlZA== 24311
+IG9jZWFucw== 24312
+dWZ0 24313
+ZXJ2aWNl 24314
+INCc0KPQl9Cr0JrQkA== 24315
+IENsYXNzaWM= 24316
+IHJlc3BlY3RpdmVseQ== 24317
+fik= 24318
+w650cmU= 24319
+IE5hc2g= 24320
+IHppdA== 24321
+IOybgw== 24322
+IOuGkg== 24323
+cXVvdGU= 24324
+IFVucw== 24325
+IHRhYw== 24326
+IHByb3Zlcw== 24327
+IFBvcnRsYW5k 24328
+Ymx5 24329
+IGVyZQ== 24330
+7LaU 24331
+IMOpcG9jYQ== 24332
+INGC0YvRgdGP0Yc= 24333
+NzY= 24334
+IGhhZGU= 24335
+IEZybw== 24336
+IHBvbMOtdGljYQ== 24337
+dGFn 24338
+IO2VrQ== 24339
+IHNjaMO2 24340
+YXJldHQ= 24341
+IHByb3Zpc2lvbnM= 24342
+IG1vdG9ycw== 24343
+IGltYWdpbmc= 24344
+IGRvaw== 24345
+dWxvdXNseQ== 24346
+IG1laWxsZQ== 24347
+65A= 24348
+IElTTw== 24349
+IFNURU0= 24350
+IEJvd2w= 24351
+IHRvd2Vycw== 24352
+IEVl 24353
+IFBlcmZvcm1hbmNl 24354
+IGxvaW4= 24355
+Y3Vzc2lvbg== 24356
+IGNvYXN0YWw= 24357
+aWFsZQ== 24358
+Y29tcGFzcw== 24359
+IHNwZWxscw== 24360
+IGRpc2FwcG9pbnRpbmc= 24361
+IOuyiOynuA== 24362
+RUVS 24363
+IHZlcnNhdGlsZQ== 24364
+YXN1cnk= 24365
+IGVuZmlu 24366
+IGRvd25zaWRl 24367
+IGd1aWRpbmc= 24368
+INin2YTZgg== 24369
+IG5pbmV0eQ== 24370
+Y2hhcmdlZA== 24371
+IEZhbnM= 24372
+IHBoaWxvc29waGljYWw= 24373
+IGdhcm4= 24374
+IG3DpW5nYQ== 24375
+IHdpbGxpbmduZXNz 24376
+IHBvcnRpb25z 24377
+YWJlbg== 24378
+IO8= 24379
+wr8= 24380
+cmF1bA== 24381
+IHNwcmludA== 24382
+aWZlbg== 24383
+xLF5bGE= 24384
+INC60YPQvw== 24385
+IGVuc3VpdGU= 24386
+IENhcGl0b2w= 24387
+IDYz 24388
+INCz0L7QstC+0YDQuNGC 24389
+IGFwcG9pbnRtZW50cw== 24390
+b21pYXN0 24391
+IGNhcmVn 24392
+IHB1Ymxpc2hlcg== 24393
+IGhlcmF1cw== 24394
+IM61zq8= 24395
+IFZT 24396
+IHNhY3JpZmljZXM= 24397
+dGhpcmQ= 24398
+IGh1bWFuaXRhcmlhbg== 24399
+IOuCtOw= 24400
+aW1vbg== 24401
+IGluZXF1 24402
+IHpvYg== 24403
+IGNvbWZvcnRhYmx5 24404
+IERpbmdl 24405
+IGNhbmNlbGxlZA== 24406
+IFBTQUtJ 24407
+IFJvYmluc29u 24408
+IGZpbnM= 24409
+KT8= 24410
+IEhpc3Rvcg== 24411
+INGH0LXQu9C+0LLQtdC60LA= 24412
+IHRic3A= 24413
+dGV4dA== 24414
+a2lt 24415
+IHVwZGF0aW5n 24416
+IGdlbGQ= 24417
+ZmVsZA== 24418
+j7w= 24419
+IG3DpA== 24420
+IGNhZsOp 24421
+1oA= 24422
+IFNyaQ== 24423
+IFJlZ2lvbg== 24424
+IEhhaGFoYQ== 24425
+IGZpbmFuY2Vz 24426
+INin2YTYtA== 24427
+IGJ1bms= 24428
+cnVr 24429
+aGFmdA== 24430
+IGxhdGVyYWw= 24431
+IGV4dGVuc2lvbnM= 24432
+IOyVhOydtA== 24433
+IGRlZmluaXRl 24434
+IFpoYW8= 24435
+IEx1aXM= 24436
+c3R5 24437
+IGNhc29z 24438
+IEtsaW0= 24439
+IDE5OTM= 24440
+IHJlYWxpemF0aW9u 24441
+IGhpc3Rvcmlhbg== 24442
+IGNyYWNrZWQ= 24443
+64K0 24444
+IHN5c3TDqG1l 24445
+IENJQQ== 24446
+INGC0LLQvg== 24447
+b3NwaGVyaWM= 24448
+IGZsZWU= 24449
+IHLhuqV0 24450
+IFJlZ2FyZGxlc3M= 24451
+IHJlbHVjdA== 24452
+IHRpbWVseQ== 24453
+IEp1bGlhbg== 24454
+R00= 24455
+6ZI= 24456
+YWR1cmE= 24457
+IGRyZXNzZXM= 24458
+IOuUlA== 24459
+IG5vbWluYXRlZA== 24460
+IGFkdm9jYXRlcw== 24461
+eW1waA== 24462
+IHJlY29yZGluZ3M= 24463
+IGRldmlhdGlvbg== 24464
+IHByaW9yaXRpemU= 24465
+IHNwaXJhbA== 24466
+IFlPVVI= 24467
+IHRyYW5zcG9zZQ== 24468
+YW1wb28= 24469
+IOybkOuemA== 24470
+IFZpc2lvbg== 24471
+IHBvbGl0ZQ== 24472
+IGhhbWI= 24473
+IFBhdGllbnQ= 24474
+7YGs6w== 24475
+IHNpYQ== 24476
+IOqzsw== 24477
+IMW+ZQ== 24478
+IHN1cGVybWFya2V0 24479
+67k= 24480
+IFNpZXJyYQ== 24481
+IGdyaWxsZWQ= 24482
+IFVwb24= 24483
+IGFic2VudA== 24484
+IG1lYw== 24485
+IEFwb2xsbw== 24486
+IHB1bms= 24487
+IFBhxYRzdA== 24488
+INGB0LLQvtC5 24489
+IOqxsOq4sA== 24490
+R2lybA== 24491
+IHNraW5ueQ== 24492
+IFByZW1pZXI= 24493
+IHRlcnJpdG9yaWVz 24494
+IGxpYWJpbGl0eQ== 24495
+IGplcms= 24496
+cmF0aWM= 24497
+IGRhbmNlcnM= 24498
+INGD0YDQvtCy 24499
+IOq0gOs= 24500
+b25seQ== 24501
+IFN0dQ== 24502
+IHNrZWxldG9u 24503
+IOutkOs= 24504
+INC30LDQutC+0L0= 24505
+xLFrdA== 24506
+IE1JS0U= 24507
+IGzDtg== 24508
+bWll 24509
+IHJlaXRlcg== 24510
+IEtvbGxlZw== 24511
+IEFkYW1z 24512
+bGljaGVy 24513
+IMOnb2N1aw== 24514
+0Y/Qsw== 24515
+IGJsdXNo 24516
+IHN1bnNoaW5l 24517
+IGV6 24518
+IERldmls 24519
+IOq4uA== 24520
+YWRk 24521
+IGxpY2Vuc2Vk 24522
+IHZpbnls 24523
+IEN6ZWNo 24524
+aW1hZw== 24525
+IGNyYWNraW5n 24526
+IOy6 24527
+IHVkYWg= 24528
+IHNvbW1lcw== 24529
+IOyWvOq1 24530
+d2HEhw== 24531
+IGZyZXM= 24532
+IFdhbG1hcnQ= 24533
+INCi0LXQv9C10YDRjA== 24534
+YXRpc2Y= 24535
+Q0k= 24536
+bGFuZw== 24537
+IGRpZmZ1c2lvbg== 24538
+IHNvbW9z 24539
+IE1ha2Vz 24540
+IFJpY2t5 24541
+IG11Y2hh 24542
+7ZWo 24543
+IGhvcnNlcG93ZXI= 24544
+YXNpYQ== 24545
+IGZpYmVycw== 24546
+IGVybQ== 24547
+0YHQutC40LU= 24548
+IGplc3Rl 24549
+IGZpcmVmaWdodA== 24550
+IGN1aXNpbmU= 24551
+IGJlc29uZGVycw== 24552
+ZGln 24553
+IOyihQ== 24554
+INGD0LY= 24555
+IHRyYWNpbmc= 24556
+IGNlcnRhaW5z 24557
+IEFwcGx5 24558
+0YvQstCw0YLRjA== 24559
+54w= 24560
+IGJydQ== 24561
+IFlFUw== 24562
+IEJhaQ== 24563
+IERpdA== 24564
+IEJpcw== 24565
+IHVubGU= 24566
+0YHRgtCw0YLQvtGH0L3Qvg== 24567
+IEF3YWs= 24568
+Li4i 24569
+IDEyNQ== 24570
+IHJvb3RlZA== 24571
+IGNhdXRpb3Vz 24572
+Y29uc3Q= 24573
+IG9yY2hlc3RyYQ== 24574
+INCy0L3Rg9GC 24575
+IHF1ZWxxdQ== 24576
+INC+0YLQstC10YI= 24577
+IE1ldGhvZA== 24578
+7Lmc 24579
+IM68zrHPgg== 24580
+bMO8 24581
+IOyVhOq5jA== 24582
+IG5hbWluZw== 24583
+Q2hhcg== 24584
+IFNpY2hlcg== 24585
+IHByaXZpbGVnZWQ= 24586
+IEZseQ== 24587
+4bqtdA== 24588
+IGFkdmFuY2Vz 24589
+IFplbGRh 24590
+IGFuZHJh 24591
+IGdyaW5kaW5n 24592
+IEVkaXRpb24= 24593
+cGY= 24594
+IHdhcnJpb3Jz 24595
+IGhlZGdl 24596
+IHVuc2VyZW4= 24597
+INGB0Y7QtNCw 24598
+ZWxpbmVzcw== 24599
+IHBlcnNvbmFsaXRpZXM= 24600
+IGbDtg== 24601
+J00= 24602
+INGC0L7Rh9C90L4= 24603
+IHNoaXBwZWQ= 24604
+IG1ldGVvcg== 24605
+IHN1cnJvdW5kaW5ncw== 24606
+IEZpbGw= 24607
+dWVzdGE= 24608
+IFBlcnNvbmFs 24609
+IEFsbGU= 24610
+T1JU 24611
+IFNjaGU= 24612
+Vkk= 24613
+IGNvbXBhcmFibGU= 24614
+ZGFtbg== 24615
+IGRpdGNo 24616
+WUFO 24617
+aXNtdXM= 24618
+IHBpY2t1cA== 24619
+IGRhaw== 24620
+IEVQ 24621
+YmVzdA== 24622
+IFN1ZQ== 24623
+w6RsbHQ= 24624
+IHBvcGNvcm4= 24625
+IGZvbGRpbmc= 24626
+aG9tZQ== 24627
+0LjQstCw0LXRgg== 24628
+IGFubm90 24629
+Y2h1Y2s= 24630
+IGZpZXJjZQ== 24631
+IGRhbWFnaW5n 24632
+IGZsb3A= 24633
+IHBhc2Fy 24634
+IHJlZWY= 24635
+INGB0LLQvtC10Lk= 24636
+IHpvbw== 24637
+b3ZlcnM= 24638
+amV0cw== 24639
+IHByw6hz 24640
+IFNpbGljb24= 24641
+dGVvaw== 24642
+IFNldGg= 24643
+YXRhbWVudGU= 24644
+IHRyYW5zbWl0dGVk 24645
+IHJlcGxpY2F0ZQ== 24646
+IHNsaW0= 24647
+IENyZWFt 24648
+IHNpZGV3YWxr 24649
+7IiY6w== 24650
+INC20LjQt9C90Yw= 24651
+IE1vbmljYQ== 24652
+IGNvcGllZA== 24653
+IFRlcnJh 24654
+aXN0ZW50 24655
+INC+0L3Qvg== 24656
+IHdoYWxl 24657
+IFdJVEg= 24658
+0LvRg9GI 24659
+IEVlbg== 24660
+INGB0LLQvtC4 24661
+IG9yZGlu 24662
+IHBsdXJhbA== 24663
+IHNwb2tlcw== 24664
+IGRpc3B1dGU= 24665
+IHNlbnNpYmxl 24666
+IHByZWFjaGluZw== 24667
+IGt0w7Nyenk= 24668
+cHRlZA== 24669
+YXZpZXI= 24670
+IHBpc3RvbA== 24671
+IFRhcGk= 24672
+IMWC 24673
+ZmZmZg== 24674
+IGFjcnlsaWM= 24675
+IGlnbm9yYW5jZQ== 24676
+IFppZWw= 24677
+cmFucw== 24678
+IHdlbGRpbmc= 24679
+bWlk 24680
+INC30LDQvdC40Lw= 24681
+IGxhbmVz 24682
+IG1pbmVz 24683
+IG1vbXM= 24684
+15XXlw== 24685
+IENoYW1iZXI= 24686
+dGllcg== 24687
+IG1vZGVzdA== 24688
+IOyXrOq4sOyEnA== 24689
+IHVuYXM= 24690
+IHdyZW5jaA== 24691
+aGFuZGVk 24692
+IHNhdHVyYXRlZA== 24693
+IEZhbmc= 24694
+IENvbW1pc3Npb25lcg== 24695
+4KSw 24696
+INeW 24697
+IExvdWlzaWFuYQ== 24698
+IE1hc2s= 24699
+IGN1YmVz 24700
+7JSo 24701
+IHZpZMOpb3M= 24702
+IG7DpWdvbg== 24703
+IHJpZGVy 24704
+IOy2nA== 24705
+IHPDs24= 24706
+IExhdGlubw== 24707
+YmFuaw== 24708
+7ZW07KO8 24709
+IEJyZW5k 24710
+IHNleHVhbGl0eQ== 24711
+Li4uLA== 24712
+IGZvcmdldHRpbmc= 24713
+INuM 24714
+IEF2ZW5nZXJz 24715
+IEJvbmpvdXI= 24716
+Y2Vzc29y 24717
+0LrRgNCw0Zc= 24718
+Y2VuY2U= 24719
+IGdlb2dyYXBo 24720
+Y3Vsbw== 24721
+0L7RgdGC0Yw= 24722
+IHN3ZWF0aW5n 24723
+7YOA 24724
+IHN5bW1ldHJ5 24725
+dHPDpQ== 24726
+IGphbg== 24727
+IEZlcnI= 24728
+IGFtYmFzc2Fkb3I= 24729
+emnEmWs= 24730
+IG11c3Vu 24731
+INGD0YI= 24732
+IExH 24733
+aXNzZW50 24734
+Y29tbXVu 24735
+IGNvdXJz 24736
+IGRldmVsb3Bz 24737
+IGJyb256ZQ== 24738
+IHN1YnN0YW5jZXM= 24739
+ZHJpdmVu 24740
+7KO87IS47JqU 24741
+IGFvcw== 24742
+IFBST0ZFU1M= 24743
+aGFsZg== 24744
+IHNvcnRlZA== 24745
+IEJvbWI= 24746
+0LvQsNCz 24747
+IE1hbGF5c2lh 24748
+IENocmlzdGluYQ== 24749
+IHRlYW1tYXRl 24750
+RlQ= 24751
+IGvEsQ== 24752
+aGVhcnRlZA== 24753
+Kys= 24754
+b2dlbmlj 24755
+IGJlbGxz 24756
+IE91YWlz 24757
+IHNwZWNpYWxpc3Rz 24758
+0LHRiw== 24759
+ZGVwdGg= 24760
+bGFzc2Vz 24761
+Z2llcw== 24762
+IENvZmZlZQ== 24763
+IG1hcmtpbmc= 24764
+IGZvbGw= 24765
+dWxp 24766
+IGFkaGVzaXZl 24767
+IEJvdA== 24768
+IFB1bmt0 24769
+ZXll 24770
+IEJ1Yg== 24771
+ZWxvbmc= 24772
+INC/0YDQuNC6 24773
+IGRvbm9y 24774
+ODQ= 24775
+IGVuZm9y 24776
+IGNhdGNoZXM= 24777
+IGJyaWNrcw== 24778
+IGtuaXR0aW5n 24779
+IEtub3dpbmc= 24780
+b2tz 24781
+SFk= 24782
+cmlkZQ== 24783
+IEZhbnRhc3k= 24784
+aW1hbg== 24785
+IHBzZQ== 24786
+IOyYqA== 24787
+INCy0LQ= 24788
+IHJlc3RyYQ== 24789
+IGV2YWx1YXRlZA== 24790
+0YDQtdCy 24791
+IGZvcnR1bmF0ZWx5 24792
+IGNoZWdhcg== 24793
+2LHYqA== 24794
+IGRvbWFpbnM= 24795
+aWJp 24796
+YXJyeQ== 24797
+IHNodXR0ZXI= 24798
+IGZpY291 24799
+TWlrZQ== 24800
+IGluY2x1 24801
+IGRvbm9ycw== 24802
+IGFwbA== 24803
+IExvd2Vy 24804
+IGltcG9ydGVk 24805
+IGFjYWRlbXk= 24806
+IGZpbmFscw== 24807
+IGRpc2FwcGVhcnM= 24808
+2YrYpw== 24809
+IGFkbWluaXN0cmF0b3I= 24810
+anM= 24811
+IGN1dHRlcg== 24812
+IHJhbmdpbmc= 24813
+w7ZycGVy 24814
+IGNvbnN0cmFpbnQ= 24815
+IFRhYmxl 24816
+IFNoYW4= 24817
+dmlj 24818
+IEZpeA== 24819
+IFN3aWZ0 24820
+b3VuY2Vz 24821
+IFdhcnVt 24822
+IGxldHR1Y2U= 24823
+YXBwZWxsZQ== 24824
+IHNoYXZl 24825
+IGLDoXM= 24826
+IDc3 24827
+IE9vbw== 24828
+YW8= 24829
+IE1jTQ== 24830
+IERyZXc= 24831
+IGx1bXA= 24832
+IGxhc2hlcw== 24833
+c2NoZWlubGljaA== 24834
+UmVw 24835
+aW5pcw== 24836
+IENldHRl 24837
+IGNvbXBvc2l0ZQ== 24838
+ZW1ldGVyeQ== 24839
+IHNvcnRl 24840
+IEZpbmFuY2lhbA== 24841
+0L7QvdC1 24842
+cm9uZXM= 24843
+IFZveQ== 24844
+IHTDqWM= 24845
+oLk= 24846
+IE5pbmph 24847
+IENvcmlu 24848
+0LXQvdC90Y8= 24849
+7J207JeI 24850
+IG5pY2g= 24851
+IGRldGVjdGl2ZQ== 24852
+4oCmIg== 24853
+z4POtQ== 24854
+nbzrj4Q= 24855
+IOuzgA== 24856
+IOu4lOs= 24857
+IHByb3Bl 24858
+IFdyaWdodA== 24859
+INeU16o= 24860
+IFNoaQ== 24861
+IGludmVzdGlnYXRpb25z 24862
+IFBvd2VyUG9pbnQ= 24863
+IENodQ== 24864
+IOyYpO0= 24865
+IOyZhOyghA== 24866
+IEZyYWdlbg== 24867
+dW5uaW5n 24868
+IHBvdXJyYWl0 24869
+IHRleHRib29r 24870
+0LzRiw== 24871
+IGZhaHJlbg== 24872
+INGC0L7RgA== 24873
+IGxha2Vz 24874
+w7xuZGU= 24875
+SW50 24876
+IE1ldHJv 24877
+IG1hbnNpb24= 24878
+INCw0LE= 24879
+IFpob3U= 24880
+IGNvcnJpZG9y 24881
+IGVzY29s 24882
+IGluZGljYXRpbmc= 24883
+aWHFgmE= 24884
+IG1vbW15 24885
+IGFyY2hpdmVz 24886
+IGZvdW5kZXJz 24887
+ZW5naW5l 24888
+IERpZXU= 24889
+IHNpY2tuZXNz 24890
+IOuztOuLiOq5jA== 24891
+IGFyYg== 24892
+IG5lZA== 24893
+IENob3A= 24894
+IGNvdmlk 24895
+IHNsYW0= 24896
+IHB1YmxpY2F0aW9ucw== 24897
+REM= 24898
+IHNwZW5kcw== 24899
+5r4= 24900
+IHJlZnVnZWU= 24901
+IGRpbGU= 24902
+INeQ15Y= 24903
+aWZpY2Fy 24904
+IFNhY2g= 24905
+R3U= 24906
+IHJlbG9hZA== 24907
+Pz8/Pw== 24908
+IGplxZtsaQ== 24909
+INGB0L7RgdGC0L4= 24910
+IHNpbXBsaWNpdHk= 24911
+IGJ1bGx5aW5n 24912
+INC80L7Quw== 24913
+IHJlYWxpZGFk 24914
+IHVuY2xlYXI= 24915
+YXBwYQ== 24916
+bGV2YW50 24917
+IElTSVM= 24918
+IFdhdHNvbg== 24919
+IGRlaW4= 24920
+IE1pY3Jv 24921
+7ZWc6w== 24922
+w7xn 24923
+IGRldmFt 24924
+IHR3ZWV0ZWQ= 24925
+IHVuZGVyc3RhbmRhYmxl 24926
+YXRhbg== 24927
+IHZlcnNh 24928
+IHByZWNh 24929
+IHbhu4E= 24930
+IENvcHk= 24931
+IE9yYWNsZQ== 24932
+IG1pbmRmdWxuZXNz 24933
+IGRpc2NyZXQ= 24934
+ZXJuZW4= 24935
+IFBsZQ== 24936
+SGF2ZQ== 24937
+IGlzb2xhdGU= 24938
+IGRldQ== 24939
+IHNldmVudHk= 24940
+IEhpbGxz 24941
+IGFyY2FkZQ== 24942
+INGB0L/QtdGG0Lg= 24943
+IHNpZ3VpZW50ZQ== 24944
+IELDnE5ETklT 24945
+bGlnYQ== 24946
+INCy0YHRgtGA0LXRhw== 24947
+w7Rt 24948
+IHR3ZWV0cw== 24949
+IHNjaGF1ZW4= 24950
+IGNyaXRpcXVl 24951
+IPCfjrU= 24952
+IHN0YXR0 24953
+INGB0LDQvNC+0LU= 24954
+w6JuY2lh 24955
+IHN1cGVybmF0dXJhbA== 24956
+IHBsdWdnZWQ= 24957
+Rmw= 24958
+eW7EsQ== 24959
+IFRhbWJpw6lu 24960
+IGVuY291cmFnZW1lbnQ= 24961
+IFNlcnZlcg== 24962
+64Kc 24963
+dXBh 24964
+IGFzdG9u 24965
+IGhlYXJz 24966
+0YDQsNGF 24967
+IHNjaGU= 24968
+IHJhdHM= 24969
+IHJlY3VwZXI= 24970
+IHVudGVu 24971
+IEZpZ2h0aW5n 24972
+IGFjYWRlbWljcw== 24973
+IFPDvA== 24974
+0YHQutC40YU= 24975
+IHBhaXJlZA== 24976
+gOydhA== 24977
+IMOhcmVh 24978
+IHN3ZWV0bmVzcw== 24979
+IGRlZmVy 24980
+IG11aXRhcw== 24981
+IEF1ZGlv 24982
+IGxvY2tlcg== 24983
+2YrYrw== 24984
+INGB0YLQsNCy 24985
+IGJ1ZW5h 24986
+QU5T 24987
+IGRldGVjdG9y 24988
+YXZv 24989
+YmVr 24990
+IM6xzr0= 24991
+7Y64 24992
+IGRyYWdnZWQ= 24993
+INC00L7Qu9C20LXQvQ== 24994
+w5Y= 24995
+2LHYqQ== 24996
+7J207KeA 24997
+IGNlbGxl 24998
+Y2tpbmc= 24999
+INin2YTYrA== 25000
+IENhbnZhcw== 25001
+IGVzcGHDsQ== 25002
+IGdsaW1w 25003
+IHNwcmVhZHM= 25004
+b25nbw== 25005
+IE1hc29u 25006
+IEluZw== 25007
+IOqwgOuKpQ== 25008
+z4TOuc66 25009
+IHNlY3VsYXI= 25010
+IGJhdGVy 25011
+IGlucXVpcnk= 25012
+IGVuZXJnaWVz 25013
+IG1hbnVmYWN0dXJlZA== 25014
+IHZlZ2V0YXJpYW4= 25015
+IHBpbmVhcHBsZQ== 25016
+0Y/RgtCw 25017
+IHByYWN0aXRpb25lcnM= 25018
+MjAwMA== 25019
+IO2VtOyalA== 25020
+IOyXrOufrOu2hOuTpA== 25021
+IOu2iOs= 25022
+IEplZmZlcnNvbg== 25023
+IEpvYW4= 25024
+IHRyYW0= 25025
+Y2htYWw= 25026
+IEhhaXQ= 25027
+4bmH 25028
+IHVucmVhbA== 25029
+IHN5bWJvbGlj 25030
+IHN0ZWFsdGg= 25031
+IHNwbGFzaA== 25032
+IEVudGVydGFpbm1lbnQ= 25033
+IG1ldGFsbGlj 25034
+PyIu 25035
+YXJvdW5k 25036
+IGRlc3BhaXI= 25037
+IE5ldmFkYQ== 25038
+IEZpbmFuY2U= 25039
+IGtyaWU= 25040
+IEx1eA== 25041
+IFNtYXNo 25042
+a2VlcGluZw== 25043
+INC30LDQsw== 25044
+IG5hcmNpc3M= 25045
+IGR6aXNpYWo= 25046
+IHRvbGVyYXRl 25047
+b2FyZA== 25048
+IGxpbmtpbmc= 25049
+IEVjb25vbWlj 25050
+IOy8 25051
+IG1vcnBo 25052
+IE5haw== 25053
+IEJha2Vy 25054
+YXRvbg== 25055
+cmluZ3M= 25056
+IFBlbmc= 25057
+IEFpcnBvcnQ= 25058
+7ZWY64uk 25059
+p4E= 25060
+cHJpbnRz 25061
+IGhhZGk= 25062
+IGVtcGly 25063
+IExpdmVz 25064
+YW5uZXJz 25065
+INC90LjQvA== 25066
+IFBST0ZFU1NPUg== 25067
+IHBvc2l0aXZlbHk= 25068
+YW50b20= 25069
+IGJhZGdl 25070
+a2VsdA== 25071
+IGludGVyZmVy 25072
+IGZ1bGZpbGxpbmc= 25073
+IHZpc3VhbGl6YXRpb24= 25074
+IFByaWNl 25075
+77+977+9 25076
+IHNjZW5lcnk= 25077
+IHByb25l 25078
+IHdpemFyZA== 25079
+IGJhbnlhaw== 25080
+dmVyYg== 25081
+c2t5 25082
+IHdpc2hlZA== 25083
+IHJhaWx3YXk= 25084
+IMO8emVy 25085
+IGFsZ3VpZW4= 25086
+IEFX 25087
+INC60L7Qu9C40YfQtQ== 25088
+IHJlYWN0aW5n 25089
+IEJ1Y2g= 25090
+4Li2 25091
+IGFudGg= 25092
+IHNpaA== 25093
+IGh1c3Q= 25094
+IFNjcmVlbg== 25095
+aWxhbnQ= 25096
+YWhv 25097
+IGZyYWdyYW5jZQ== 25098
+IGVsZXZhdGlvbg== 25099
+IE1lZGl0ZXI= 25100
+IOu/ 25101
+IMOpcXU= 25102
+IHdyYXBz 25103
+IGluZXJ0 25104
+IHJlY3JlYXRl 25105
+0LvQsNGC 25106
+IGJvbGVo 25107
+IGhhcmFzc21lbnQ= 25108
+dW5reQ== 25109
+IGdsaW1wc2U= 25110
+cmVnaWVydW5n 25111
+IGZ1dHVy 25112
+IHJlcG9zaXRvcnk= 25113
+IGVuZ3Jh 25114
+IHRyYWZmaWNraW5n 25115
+YXNzaXM= 25116
+IFRyZWs= 25117
+IOuyjA== 25118
+IOuniOs= 25119
+IEthYg== 25120
+YW5pdQ== 25121
+Z2l2ZQ== 25122
+IGRpbm9zYXVycw== 25123
+IGZlYXRoZXI= 25124
+IGF0dGl0dWRlcw== 25125
+IHBsdW0= 25126
+IFJT 25127
+IEFuZmFuZw== 25128
+aWxsZXJ5 25129
+IOyKpA== 25130
+TVk= 25131
+IHRyemViYQ== 25132
+IHNraWVz 25133
+IEFq 25134
+dXJhYmxl 25135
+Q1U= 25136
+IFNoYW5l 25137
+IGRlcGFydHVyZQ== 25138
+IFRPTg== 25139
+aWV0ZW4= 25140
+cmF0cw== 25141
+aXN1 25142
+IGJvcmQ= 25143
+IGludGVyZXN0aW5nbHk= 25144
+b3VnaGluZw== 25145
+IHJ1c2hpbmc= 25146
+IHZvbGF0aWxpdHk= 25147
+IHB5dA== 25148
+IGZvcm1hdHM= 25149
+INC30LDRgg== 25150
+IOq8rQ== 25151
+IHdoYXRub3Q= 25152
+IGNvbXBvcnQ= 25153
+c3c= 25154
+b3JlYW4= 25155
+IFJlbGF4 25156
+IGNsYW4= 25157
+IEFI 25158
+IHBldw== 25159
+IGRpY3Rpb25hcnk= 25160
+VGFrZQ== 25161
+c2hpcnRz 25162
+IEh1Z2g= 25163
+INi52YTZig== 25164
+IFBpYw== 25165
+IGVucm9sbGVk 25166
+IGplZG5haw== 25167
+IG9mZmVyaW5ncw== 25168
+IGNvcmF6 25169
+TGlmZQ== 25170
+ICEhIQ== 25171
+IGNsZXI= 25172
+IFZpZGVvcw== 25173
+IFJvZHJpZw== 25174
+IElkZW50 25175
+IFBvcw== 25176
+IFN0YWdl 25177
+IFJhY2U= 25178
+IGVuYWN0 25179
+IEd5 25180
+IEhpc3Bhbg== 25181
+IGRlZmVuY2U= 25182
+IENhbXBiZWxs 25183
+bWF0aWM= 25184
+IHJlbGV2 25185
+IHBlYWNo 25186
+hLjsmpQ= 25187
+IHBhcmFkaXNl 25188
+IGNlcmVtb24= 25189
+IGFubm95ZWQ= 25190
+bGF4 25191
+IGV4cGxvaXQ= 25192
+IGNsYXVzZQ== 25193
+ZWtlcg== 25194
+IEJsb29t 25195
+bmFudA== 25196
+YXRldXJz 25197
+IGhlaWdodHM= 25198
+RXZlbg== 25199
+0YHQvtC9 25200
+IG91dHJhZ2U= 25201
+IFZpZXRuYW1lc2U= 25202
+VFI= 25203
+IGVlcg== 25204
+IGNhbm5vbg== 25205
+IENvbWI= 25206
+kOunjA== 25207
+IOqyg+uPhA== 25208
+IGFjY29tcGxpc2htZW50cw== 25209
+IEFuYWx5dGljcw== 25210
+IHNoYXBpbmc= 25211
+cmVpYmVu 25212
+IGJhY2hlbG9y 25213
+IGZpbmdlcnQ= 25214
+YWNrZWQ= 25215
+IHB5cmFtaWQ= 25216
+IFN0ZXdhcnQ= 25217
+w6FzdA== 25218
+IHN1cnZpdm9y 25219
+IGR1Y3Q= 25220
+IGRlYWxlcnM= 25221
+2LnZhQ== 25222
+0LvQuNC9 25223
+IGVkZQ== 25224
+15XXog== 25225
+INmD2KfZhg== 25226
+IM+Ezrk= 25227
+IGNob29zZXM= 25228
+IE93bg== 25229
+0LPQvtGC0L7Qsg== 25230
+aGlyZQ== 25231
+0LDQu9GM0L3Ri9C1 25232
+INCb0Y4= 25233
+INC+0YHRgtCw0LI= 25234
+dGVjaA== 25235
+IGRyb2l0 25236
+IHN1YmplY3RpdmU= 25237
+ZW5lcw== 25238
+IGRpdmlz 25239
+YXZleg== 25240
+IG1hbmV1dmVy 25241
+4LmE4LiU 25242
+YWRlY2U= 25243
+IEVucw== 25244
+YWNpYWw= 25245
+IFByb3RlY3Rpb24= 25246
+lrQ= 25247
+IGZvcm1hbGx5 25248
+IHd5ZA== 25249
+aW5ndcOpbQ== 25250
+IHppZW0= 25251
+IHJlY3J1aXRpbmc= 25252
+15nXmg== 25253
+bmVt 25254
+IGZvcmJpZGRlbg== 25255
+IEJhcHQ= 25256
+15DXoNeZ 25257
+IHN1YnNldA== 25258
+IE1hZ2F6 25259
+bmVtZW50 25260
+IGFxdWVsYQ== 25261
+cmFnb24= 25262
+IGNvbW1pdHRlZXM= 25263
+IMOpdGFpZW50 25264
+dWRp 25265
+IERhd24= 25266
+IGJvcmU= 25267
+IGNvbXBvc2Vy 25268
+IHdpxJljZWo= 25269
+YW5nYQ== 25270
+IGRpc2xpa2U= 25271
+IERheXM= 25272
+IHBhcmFs 25273
+IG1pZW50cmFz 25274
+IGhlYXZlbnM= 25275
+aGVpZA== 25276
+IHRyYWRlcnM= 25277
+b25jZQ== 25278
+IG1hc2NhcmE= 25279
+IM+Az4HOvw== 25280
+IHdoaXNwZXI= 25281
+IE11c2s= 25282
+IEZhbWlsaWU= 25283
+QWxsYWg= 25284
+IE9saXZpYQ== 25285
+IFByb3M= 25286
+IG9saWth 25287
+aWxpbQ== 25288
+IHLDqXBvbmQ= 25289
+IFBldGVycw== 25290
+IGJpdGVz 25291
+IHZpYw== 25292
+IE5Z 25293
+ZW1wdGlvbg== 25294
+IDQ1MA== 25295
+IHZpc3VhbHM= 25296
+IGxpZXU= 25297
+w7xja2Vu 25298
+IFN0ZWVs 25299
+IEdQ 25300
+d2FpdA== 25301
+IG5vdGljZWFibGU= 25302
+dWNoYQ== 25303
+IHJlaGFiaWw= 25304
+IHJlamVjdGlvbg== 25305
+INGB0LvQtdC00YPRjtGJ 25306
+IHNsaWRlcg== 25307
+IHJlZ2FyZGVk 25308
+IGdyYXZpdA== 25309
+IFJlc2VydmU= 25310
+Y291bnQ= 25311
+IGJyZWVkaW5n 25312
+IGxvbmdl 25313
+YWxlYg== 25314
+IGtuaWdodA== 25315
+INCy0L7QuQ== 25316
+IHByw6lzZW50 25317
+gpjsmpQ= 25318
+IFNwZWNpZmljYWxseQ== 25319
+IHBvc2Vz 25320
+IHZldXJl 25321
+b2theQ== 25322
+ZW1hcw== 25323
+IG1hasSF 25324
+IHdlYmluYXJz 25325
+IGNhbm5hYmlz 25326
+IGRhbWFscw== 25327
+IE5vcnRod2VzdA== 25328
+IHBhZGE= 25329
+IGNyb3dkcw== 25330
+IGZ1dHVyZXM= 25331
+IMOkbg== 25332
+IGNpdmlsaWFucw== 25333
+IFNhY2hlbg== 25334
+5o0= 25335
+IHRyYWNlcw== 25336
+IOuoueqzoA== 25337
+UVU= 25338
+IElG 25339
+YW7EsW4= 25340
+7IK0 25341
+IGJpYmxpY2Fs 25342
+IFZlZA== 25343
+IHN0b3Jpbmc= 25344
+0YDQsNCy0LvRjw== 25345
+IG5hc3Q= 25346
+IGTDtg== 25347
+0YDQvtC/ 25348
+ZWxpYQ== 25349
+IHNpZGV3YXlz 25350
+IFVuZGVyc3RhbmQ= 25351
+IFF1cg== 25352
+IHBlcnBlbmQ= 25353
+IE1pbGxpb25lbg== 25354
+IHdhdGVybWVsb24= 25355
+IERpdmluZQ== 25356
+dWx0dXI= 25357
+YWJvcmQ= 25358
+IHN1Y2Nlc3Nlcw== 25359
+IGhvbWJyZQ== 25360
+IGNhcnA= 25361
+IHN1c2NlcHQ= 25362
+dW5na2lu 25363
+IGtpag== 25364
+dWx1cw== 25365
+2KfYrA== 25366
+IG5vdGNo 25367
+IHBvbHlub21pYWw= 25368
+5ak= 25369
+IMO6bmljbw== 25370
+IHRlbGVzY29wZQ== 25371
+IHBvbGl0aXF1ZQ== 25372
+a2llbQ== 25373
+IM6tzr3OsQ== 25374
+IGFnZ3JlZ2F0ZQ== 25375
+IEdlb2Zm 25376
+IHRyaWw= 25377
+IEdSQQ== 25378
+IHN1YnNjcmliZXI= 25379
+aW1ldA== 25380
+INC00L7Qu9C70LDRgA== 25381
+b3Bpbmc= 25382
+IHRoZXJhcGV1dA== 25383
+IENhbmNlcg== 25384
+IHBhcmFkZQ== 25385
+IGlycmln 25386
+4pmq4pmq 25387
+IGNsZWFyZXI= 25388
+IGJvZw== 25389
+IE1hdXI= 25390
+4Liy4LiH 25391
+IFNoYW5naGFp 25392
+YWNodGU= 25393
+IEtvbA== 25394
+ZWx1amFo 25395
+IGhhdg== 25396
+IENyaW1l 25397
+c2Vr 25398
+IOuhnA== 25399
+aWVubmE= 25400
+IEdvcg== 25401
+6Js= 25402
+INC/0L7RgtGA 25403
+INC60LDQttC10YLRgdGP 25404
+IExpZnQ= 25405
+IFNvcnQ= 25406
+IFBzYWw= 25407
+IHBpbmc= 25408
+k50= 25409
+cGhpcw== 25410
+IEZVQ0s= 25411
+IFN5bg== 25412
+IGJhbWJvbw== 25413
+rOyYgQ== 25414
+Y3V0cw== 25415
+IG1tbQ== 25416
+IGZ1bmt0aW9uaWVydA== 25417
+IF8= 25418
+w61jaW8= 25419
+U3RvcA== 25420
+IGltYWdpbmFyeQ== 25421
+IG5vdGFtbWVudA== 25422
+IEluaXRpYXRpdmU= 25423
+IEt1cnQ= 25424
+IGxvb3Nlbg== 25425
+IGJ1c2Nhcg== 25426
+IHplbGY= 25427
+IHByb3Bz 25428
+IG1vZXRlbg== 25429
+IG1pbGxp 25430
+IGhhbGxz 25431
+IE1hdGNo 25432
+IGJyYWNrZXRz 25433
+IENvdQ== 25434
+INCc0LDRgA== 25435
+SVNB 25436
+IGNpZ2FyZXR0ZQ== 25437
+IGNvbXBldGl0aW9ucw== 25438
+IE1JTg== 25439
+IGJlaMO2 25440
+dm9vcg== 25441
+IHVzdA== 25442
+IFpp 25443
+IE9jYw== 25444
+dWxhdGVz 25445
+IGJhbGxvb25z 25446
+IHByb250bw== 25447
+IE1peQ== 25448
+IEZpbGU= 25449
+INC60LvQsNGB0YE= 25450
+0L3Rg9C7 25451
+IGNlcmVhbA== 25452
+IGluY3JlbWVudA== 25453
+IHJlZmluZWQ= 25454
+cHJpc2luZw== 25455
+IFJG 25456
+IHJlc3BlY3RmdWw= 25457
+IGxvb3Q= 25458
+YXNrZXQ= 25459
+IGRlaXhh 25460
+aW5nbGU= 25461
+IGZ1bmNpb25h 25462
+IFJldmVs 25463
+IHNvYmVy 25464
+IHBlcmZvcm1z 25465
+IEdlbnRsZQ== 25466
+IHJlY2lwaWVudA== 25467
+IEhhdXNl 25468
+IOuD 25469
+RnJvbQ== 25470
+IG1pbmlzdGVycw== 25471
+IHBhcmFkb3g= 25472
+IHRhc3Rpbmc= 25473
+INeU15c= 25474
+IHJldXNl 25475
+IExhbmU= 25476
+INGB0L7QstC10YDRiA== 25477
+IHJlbWVtYmVycw== 25478
+IGZlbWluaXN0 25479
+IGNvbW1pdG1lbnRz 25480
+IHByb2plY3RlZA== 25481
+IGdheg== 25482
+aXlvcnV6 25483
+IG9ibGlnYXRpb25z 25484
+Um8= 25485
+emFy 25486
+IGNodw== 25487
+IEpBTQ== 25488
+IGLEmWTEhQ== 25489
+YXNwYmVycnk= 25490
+INC80LXRgdGC0L4= 25491
+67KV 25492
+IHJlZ3VsYXRlZA== 25493
+IHdpY2h0 25494
+IFRyZXZvcg== 25495
+IHNlY29uZGx5 25496
+IElocmU= 25497
+ZWxzaA== 25498
+IHJlcG9ydGVycw== 25499
+0YLQvtGA0LA= 25500
+b3lv 25501
+R0k= 25502
+IGludGVyY29ubmVjdA== 25503
+T1NI 25504
+IGJyYXNz 25505
+IGlnbm9yaW5n 25506
+aW5mZWN0 25507
+IHByb2pla3Q= 25508
+b3JldA== 25509
+z4TOsc69 25510
+INGC0LjQvw== 25511
+IG11dHRh 25512
+IHVuYm94aW5n 25513
+hLA= 25514
+IGFkdmlzZWQ= 25515
+IERlbnZlcg== 25516
+IHNldmVyZWx5 25517
+IE1obQ== 25518
+IGZsaXBwZWQ= 25519
+IHBpZW4= 25520
+IGtvbW11bg== 25521
+IEZSRQ== 25522
+IOCuh+CusA== 25523
+YWludGVk 25524
+IGtuaXZlcw== 25525
+IGhhYmw= 25526
+IGdld29yZGVu 25527
+YXJldHRlcw== 25528
+Q1M= 25529
+INC80LDQu9C10L3RjA== 25530
+IGdhbGF4 25531
+IG5pbmV0ZQ== 25532
+6rGw64KY 25533
+IHNpcw== 25534
+IGFkdmlzb3J5 25535
+IGRyaWxsaW5n 25536
+IFdvdWxkbg== 25537
+w7xuZg== 25538
+Z2VzdGVsbHQ= 25539
+IEhlbGVu 25540
+INee15A= 25541
+YXBvbGlz 25542
+IHJ6ZWN6eQ== 25543
+IHRlcnJh 25544
+IGhlcA== 25545
+IGFsZ8O6bg== 25546
+aWtr 25547
+IGFzdHJvbm9t 25548
+IFN0YXJidWNrcw== 25549
+a8SF 25550
+IHBhdHJvbA== 25551
+IOy9lA== 25552
+IGdvbg== 25553
+IOOAkA== 25554
+IHNvbnN0 25555
+IGVuY291bnRlcnM= 25556
+IHJldHJvdQ== 25557
+IHNoYXJrcw== 25558
+IGRvcg== 25559
+IFJldmVy 25560
+IGV2YXBvcg== 25561
+IHJlc2Vydm9pcg== 25562
+IGFsbGVnZWQ= 25563
+dWxlcg== 25564
+IHZlcm0= 25565
+IGNvbW1lcmNl 25566
+IGZpdHRlZA== 25567
+Z2Vt 25568
+IHRhY3RpY2Fs 25569
+IGxpdGg= 25570
+aGFk 25571
+IGNhcmJvaHlk 25572
+IGxlbmd0aHM= 25573
+zrnOvw== 25574
+IGRlbW9ncmFwaGlj 25575
+Um9i 25576
+IFNraW4= 25577
+Y2NvbGk= 25578
+IHNpbXBsaWZpZWQ= 25579
+IHJlYWRpbHk= 25580
+IEN1bQ== 25581
+YWRlc2g= 25582
+IETDpQ== 25583
+dXNzdA== 25584
+aWduZQ== 25585
+ZXRvbg== 25586
+IG1lbm9y 25587
+cWk= 25588
+T09N 25589
+4Lit4LiZ 25590
+IHBzeWNoaWF0 25591
+IGVpZ2h0eQ== 25592
+INC80LjQu9C70Lg= 25593
+IFRvYg== 25594
+ZWRv 25595
+IMSR4bq/bg== 25596
+IGNpcmN1aXRz 25597
+IExBVUdI 25598
+aWNpc20= 25599
+ZW1vcg== 25600
+IHJlZ2VuZXI= 25601
+ZWdyZWU= 25602
+IGJ1cmVhdWM= 25603
+IEFsYmVy 25604
+IFdvcg== 25605
+IHJlc2lu 25606
+IGJ5xYJ5 25607
+IElH 25608
+4K+NLA== 25609
+IDc4 25610
+IHdlZWRz 25611
+IE15dGg= 25612
+OTM= 25613
+5r8= 25614
+IOuCmOyZlA== 25615
+w6l2 25616
+4b0= 25617
+w7ZyZW4= 25618
+w6dhcg== 25619
+IFBBVUw= 25620
+IGRpc2FkdmFudA== 25621
+IHBvc2l0aW9uaW5n 25622
+IGNvY2t0YWls 25623
+IGFncmVlcw== 25624
+bm4= 25625
+IFNhbGx5 25626
+TXM= 25627
+IGluaGVyZW50 25628
+IG1vbmV0YXJ5 25629
+IG5hdHVy 25630
+IE5o 25631
+IEltcG9ydA== 25632
+IGxlYmVu 25633
+IHdp 25634
+dXNzeQ== 25635
+IG9iZXM= 25636
+IHdhbmRlcmluZw== 25637
+IOyLoOs= 25638
+xIVkYQ== 25639
+ZXRjaHVw 25640
+IGRpc3Bvc2Fs 25641
+IEpB 25642
+IENlcg== 25643
+emlsbGE= 25644
+IHZpcmdpbg== 25645
+IFNsaWRl 25646
+YW5kZWw= 25647
+IHJpZ2h0ZW91c25lc3M= 25648
+IM6j 25649
+IGlkZWlh 25650
+0LjRgNC+0LLQsNGC0Yw= 25651
+16jXkA== 25652
+Q29tbWVudA== 25653
+IHByZWxpbQ== 25654
+IFZhbGU= 25655
+IOyngOuCnA== 25656
+IFZhbmM= 25657
+T01BTg== 25658
+INC/0ZbQtA== 25659
+IHl1bQ== 25660
+c3RyZQ== 25661
+Y2Vt 25662
+IHBvY3o= 25663
+IGZyYWdtZW50 25664
+INGB0LvRg9GH0LDQtQ== 25665
+IHVuZGVyZ28= 25666
+IEhhbms= 25667
+Y2Vrcw== 25668
+IEZQUw== 25669
+IG9jdXI= 25670
+IGRldGVyaW9y 25671
+IGVtcHJlc2Fz 25672
+UGF1bA== 25673
+ICkpKQ== 25674
+INCy0YDQtdC80LXQvdC4 25675
+IHNjb2xk 25676
+15nXog== 25677
+IHN1c3BlY3RlZA== 25678
+IGFjY2Vzc2luZw== 25679
+IHN1YnN0aXQ= 25680
+IGhpc3RvcmlhbnM= 25681
+INC00LXQu9C+ 25682
+IHNvY2llZA== 25683
+cm9uZQ== 25684
+IHJlZGVu 25685
+IGV4dGVuZHM= 25686
+ZXBoZXJk 25687
+IGJhbGNvbg== 25688
+IFNvbG8= 25689
+IHBvbGl0aWNpYW4= 25690
+0L7Qu9GM0L3Qvg== 25691
+IGlyZ2VuZHc= 25692
+IHRyYXVtYXRpYw== 25693
+IHJhcHBlcg== 25694
+IFJPQkVSVA== 25695
+UmVhbGx5 25696
+IGxpbmV1cA== 25697
+QVNF 25698
+IGNvbnRyYWN0b3I= 25699
+IENvcnBvcmF0aW9u 25700
+Z29y 25701
+IFRvZG8= 25702
+0YHRgtGA0L7QuQ== 25703
+RkJF 25704
+IG5ld3NsZXR0ZXI= 25705
+IGtvxYQ= 25706
+YWx0aWVz 25707
+INC/0YDQuNGH 25708
+IEhlYXZ5 25709
+IHN3b3Jkcw== 25710
+IG1hbmlwdWxhdGlvbg== 25711
+IGZ1bms= 25712
+IHbDpXI= 25713
+IFRhbGliYW4= 25714
+IOuwpQ== 25715
+IGFjbmU= 25716
+w7xyw7w= 25717
+IGRlc3dlZ2Vu 25718
+IER1c3Q= 25719
+IHNpbGlj 25720
+IGhvb2tz 25721
+IGJsaWo= 25722
+IHBldGl0cw== 25723
+IGZpbG1l 25724
+IEJlcmVpY2g= 25725
+IFNhaWQ= 25726
+IGltcG9zZWQ= 25727
+IGRpYXJ5 25728
+INCz0L7RgA== 25729
+IEdhdGVz 25730
+IGFsdGE= 25731
+IGNoY2lh 25732
+cGxlYXNhbnQ= 25733
+IOuwnQ== 25734
+IG1vxbxlbXk= 25735
+IEF1c3RyaWE= 25736
+IGJyb2tlcg== 25737
+IHN1Y2tlZA== 25738
+IGNvbXBhcnRtZW50 25739
+IGNsb25l 25740
+INeU16I= 25741
+IERhbmtl 25742
+IG5vY2htYWw= 25743
+0LXQt9C0 25744
+IGFkcmVuYWw= 25745
+IGtsZWluZW4= 25746
+IHN1YnNlcXVlbnRseQ== 25747
+IGRlY2VudHJhbA== 25748
+IGdlbmV0aWNz 25749
+IOq0kQ== 25750
+IG1vbml0b3Jz 25751
+IEFwcGxpYw== 25752
+IFJlcG9ydGVy 25753
+d2VydA== 25754
+IHdpZW0= 25755
+IE1vdmVtZW50 25756
+IGludGVydmlld2luZw== 25757
+IGhhaXJz 25758
+IHB1w7I= 25759
+IENoZWxzZWE= 25760
+IGNvaGVy 25761
+IGNvdA== 25762
+IHphcw== 25763
+IHBhdGNoZXM= 25764
+IGxhaA== 25765
+0YPQvdC6 25766
+IFJlYWdhbg== 25767
+IE1hcmNv 25768
+Y2l0eQ== 25769
+IGRlZmVuZGVy 25770
+IGRlY29yYXRpb24= 25771
+aWpp 25772
+IGxpdHRlcg== 25773
+0Kg= 25774
+IGplZ28= 25775
+UkVX 25776
+IFBpaw== 25777
+IEhlZQ== 25778
+IEl2 25779
+INC40LTQtQ== 25780
+IFRoZWF0ZXI= 25781
+INGH0LDRgdGC0L4= 25782
+IHN3ZWF0ZXI= 25783
+IGhpZ2hsaWdodGluZw== 25784
+IGFpbnNp 25785
+IGRpcGxvbWF0aWM= 25786
+IE5ldmVydGhlbGVzcw== 25787
+5bM= 25788
+QVNPTg== 25789
+IHDDumJsaWNv 25790
+IGZlcm0= 25791
+cmVhdGVk 25792
+Y29k 25793
+IOusvOs= 25794
+IG1pc3Rlcg== 25795
+IFZhbmNvdXZlcg== 25796
+IHJlY29nbml6ZXM= 25797
+ZWNk 25798
+IGNvbXBsaWNhdGlvbnM= 25799
+ZW5jaWFs 25800
+IOqwgOyngA== 25801
+IFVsdGltYXRl 25802
+IHZhaWc= 25803
+IE1lcnJ5 25804
+15XXkg== 25805
+IE1hcmN1cw== 25806
+b3dlZ28= 25807
+IG1lbnRl 25808
+U20= 25809
+IGFqYQ== 25810
+IFRhbw== 25811
+IGp1ZGljaWFs 25812
+IGVudHJlcHJlbmV1cnNoaXA= 25813
+INC90LXQvNC90L7Qs9C+ 25814
+IHBpcw== 25815
+IGVyZw== 25816
+IGNocmlzdA== 25817
+IEN1cnQ= 25818
+INGA0LDRgdC/ 25819
+zrvOtQ== 25820
+ZW5zY2g= 25821
+w61yZQ== 25822
+IGZvY2Fs 25823
+IERpYW1vbmQ= 25824
+YXbDrWE= 25825
+IGhhbm5v 25826
+IFNxdWFk 25827
+IGFzc29jaWF0aW9ucw== 25828
+IENyZWF0aXZl 25829
+IG1lc3Nlbmdlcg== 25830
+IGJlZ2dpbmc= 25831
+IGRlY2ltYWw= 25832
+IGTEscWf 25833
+IG1ldGFkYXRh 25834
+c2Vscw== 25835
+IMSwxZ8= 25836
+4buvYQ== 25837
+IGRpZmZpY2lsZQ== 25838
+ZMSx 25839
+IHNsYXVnaHRlcg== 25840
+IFZlcmc= 25841
+INeS150= 25842
+IFRlYQ== 25843
+YXNzZXM= 25844
+T2s= 25845
+IHN5bnRoZXM= 25846
+b3RpYXRpb24= 25847
+IHBhaW50ZXI= 25848
+IGVsYm93cw== 25849
+IGFyY2hpdGVjdHVyYWw= 25850
+INGA0LDQtA== 25851
+IGdsb3I= 25852
+aW1hZ2U= 25853
+YW1wYQ== 25854
+Y3VsaWFy 25855
+oKg= 25856
+IHRldmU= 25857
+IFN0ZWxsZQ== 25858
+IEJhbQ== 25859
+IOy0iA== 25860
+YXNpcw== 25861
+aXBlZGlh 25862
+IEdJ 25863
+IEFjdGl2ZQ== 25864
+YXpp 25865
+IEx1Y2t5 25866
+7ZWp 25867
+INC/0YDQuNGF0L7QtA== 25868
+IHJ1bndheQ== 25869
+IGF1dGhlbnRpY2F0aW9u 25870
+IHBvc2libGU= 25871
+IHN1cHBsZW1lbnRz 25872
+IHN1cmdpY2Fs 25873
+R2Vu 25874
+IGZlYXNpYmxl 25875
+RE8= 25876
+IG91dGxvb2s= 25877
+IGludGVydmFscw== 25878
+IGFuZWNk 25879
+w6BuZw== 25880
+IHN0cmFwcw== 25881
+IFNodQ== 25882
+dWRk 25883
+aXNzZW5zY2hhZnQ= 25884
+IHBvcnRl 25885
+IGNvbW1pdHRpbmc= 25886
+IGFsbGV5 25887
+IGNvdmVuYW50 25888
+IFBlZHJv 25889
+bGVzc25lc3M= 25890
+IFNvbGlk 25891
+IE1vbGx5 25892
+INC90LXQutC+0YLQvtGA 25893
+IGNvb3BlcmF0ZQ== 25894
+b2xsZW4= 25895
+IHR1bmE= 25896
+IGtpbmRlcmdhcnRlbg== 25897
+IFNpeg== 25898
+IGR1xbxv 25899
+IE1CQQ== 25900
+IEdFT1JHRQ== 25901
+IEZpc2hlcg== 25902
+IENhZXNhcg== 25903
+INC60YDQsNGB0LjQsg== 25904
+IERlbGhp 25905
+enlt 25906
+IGV4cGxpY2Fy 25907
+6rCA7KeA 25908
+dW5z 25909
+Z3Jvdw== 25910
+INC/0YDQuNGB 25911
+IDg2 25912
+IHN0YXRpbmc= 25913
+IG1hc3Nh 25914
+Y2h0ZXI= 25915
+IOy7rOufrA== 25916
+IGRlcHV0eQ== 25917
+U00= 25918
+bm9j 25919
+IGdlb2dyYXBoeQ== 25920
+IEVudGVycHJpc2U= 25921
+IENhbnQ= 25922
+w7Z6 25923
+IHVucGFjaw== 25924
+IO2ZlOs= 25925
+IHNlYXJjaGVz 25926
+IHByZXNpZGVuY3k= 25927
+IHRyaXZpYWw= 25928
+IHBpZ2U= 25929
+b3VidA== 25930
+7LyA7J20 25931
+IGJ1ZGdldHM= 25932
+IHVi 25933
+IHBuZQ== 25934
+IFlhbGU= 25935
+IMWfw7Z5bGU= 25936
+cmVndWxhcg== 25937
+IGltcGVyZmVjdA== 25938
+QVJB 25939
+IGZhbcOtbGlh 25940
+dXJt 25941
+IEFkdmVudHVyZQ== 25942
+Y2lz 25943
+ZW1hcms= 25944
+IG5lZ28= 25945
+IGluYXBwcm9wcmlhdGU= 25946
+INC/0YDQuNC3 25947
+INGA0L7Quw== 25948
+IGRyZWFtZWQ= 25949
+QnJ5 25950
+IHNodXR0bGU= 25951
+IHBpbGxhcnM= 25952
+IGJpaw== 25953
+aW51bQ== 25954
+INGD0YE= 25955
+IE5lYnI= 25956
+IHBlcnBlbmRpY3VsYXI= 25957
+IGJvb2tlZA== 25958
+YmVyeQ== 25959
+IHZpa3Q= 25960
+YmVhcg== 25961
+ZXN1cw== 25962
+INCy0L7Qt9C80L7QttC90L4= 25963
+qLk= 25964
+IHByZXN1bWFibHk= 25965
+IE1lbXBoaXM= 25966
+IGFtYnVsYW5jZQ== 25967
+15XXnteo 25968
+IHRodW1ibmFpbA== 25969
+IG1vZGlmaWNhdGlvbg== 25970
+IGludGVycHJldGVk 25971
+IHByb21v 25972
+IM66zqw= 25973
+IM61z4A= 25974
+IGFjb3VzdGlj 25975
+IERC 25976
+IG5vbmV0aGVsZXNz 25977
+b3VsZQ== 25978
+IHBlcXU= 25979
+IGtub2I= 25980
+IOuPjOyVhA== 25981
+IHB1cmNoYXNlcw== 25982
+IMOHw7xua8O8 25983
+IGRpdmlkaW5n 25984
+cGVyZm9ybQ== 25985
+cmFjdGlvbg== 25986
+aGVhbHRoeQ== 25987
+IFRpdGxl 25988
+IHVr 25989
+IGNlcmNh 25990
+IGFyZ3VhYmx5 25991
+IGZhbGU= 25992
+67O1 25993
+IGdhbWVycw== 25994
+IHV0aWxpemluZw== 25995
+IG9mZmVuZGVk 25996
+IHRhdmE= 25997
+YWzEsQ== 25998
+IG1lZGlhbg== 25999
+IGluZmVjdGlvdXM= 26000
+IEFubmll 26001
+IHNtYXJ0cGhvbmVz 26002
+IHBhcm9sZQ== 26003
+IEVwaWM= 26004
+enph 26005
+IHVuaWZpZWQ= 26006
+IOq3uOuVjA== 26007
+IGN1cnRhaW4= 26008
+IMSD 26009
+IHNleHVhbGx5 26010
+IHVuc2VyZW0= 26011
+IENvbnZlbnRpb24= 26012
+IGFsbGVnZWRseQ== 26013
+WWE= 26014
+IEhvbw== 26015
+ZW5tZW50 26016
+7ZuE 26017
+IGdpZ2FudGlj 26018
+IG5vdGluZw== 26019
+IHJlYm8= 26020
+IEphbWE= 26021
+IEFseg== 26022
+IGJvcnJvd2Vk 26023
+7Lmo 26024
+IHBlcmlwaGVy 26025
+0L7RgtCw 26026
+IEdC 26027
+IEdlYXI= 26028
+IGVjb25vbWljYWxseQ== 26029
+IHRlbGVmb24= 26030
+IHF1ZXJlbW9z 26031
+INC00LDQu9GM0YjQtQ== 26032
+IHJhcw== 26033
+IFRlYWNo 26034
+aWNpb3M= 26035
+YXRvcw== 26036
+IHBsZWRnZQ== 26037
+YmF1 26038
+IEhpbXNlbGY= 26039
+TGluaw== 26040
+IGVzcGVybw== 26041
+IGNocm9tb3M= 26042
+IFBFUg== 26043
+IGVybGU= 26044
+IHBvZGl1bQ== 26045
+w6dvcw== 26046
+IG5pZXU= 26047
+IGZlbg== 26048
+IEdPRA== 26049
+IENob2NvbGF0ZQ== 26050
+d2Vyaw== 26051
+IHThu6s= 26052
+IHN1cHByZXNz 26053
+zrvOtw== 26054
+IDI0MA== 26055
+IHNpdMOk 26056
+IGhvbmVzdHk= 26057
+IEJpbw== 26058
+IEJhcmQ= 26059
+INC+0LHRidC10Lw= 26060
+INC80YPQtw== 26061
+IG1hcmJsZQ== 26062
+INGG0LXQvdGC 26063
+IHByb2N1cmU= 26064
+IHJvdG9y 26065
+YmVybg== 26066
+IHR1aA== 26067
+IGhlYWRzZXQ= 26068
+YXRlbQ== 26069
+IHdhcnJhbnR5 26070
+4K60 26071
+IGZpbGluZw== 26072
+zrnOrA== 26073
+IGNvbXByZW5kcmU= 26074
+IGltcHVsc2U= 26075
+IHNhbHY= 26076
+d3JpdHRlbg== 26077
+IGluc3RpdHV0ZQ== 26078
+S2lt 26079
+IExHQlRR 26080
+ZmljaWVudGU= 26081
+SGlz 26082
+IM6xz4XPhM+M 26083
+IHRlZW5hZ2U= 26084
+b3J1cw== 26085
+INGA0LDQt9Cx 26086
+U2Vl 26087
+IENvbnNlcnY= 26088
+4buBbg== 26089
+ZnVsbmVzcw== 26090
+IHN0cmF3YmVycmllcw== 26091
+IEFidQ== 26092
+0LjQvtC9 26093
+IG9sbGE= 26094
+Tk9JU0U= 26095
+IEVtcGxveQ== 26096
+IHdpcGVk 26097
+dXJnZXI= 26098
+IG1vZGlmaWNhdGlvbnM= 26099
+IO2VmOyngA== 26100
+IGZvb3RzdGVwcw== 26101
+IGhvbm9ycw== 26102
+IGFkdWw= 26103
+IGZsaXBwaW5n 26104
+IEhV 26105
+Wlk= 26106
+IGludGVncmF0aW5n 26107
+2KjYsQ== 26108
+dWxsYQ== 26109
+IG5hdHV1cmxpams= 26110
+IO2XiA== 26111
+IEV0aGVyZXVt 26112
+2YrZhA== 26113
+d2Vk 26114
+IHBlYWtz 26115
+IEtlcw== 26116
+IGJsb29t 26117
+IGNyYXNoaW5n 26118
+IDkxMQ== 26119
+INC+0YLQu9C40Yc= 26120
+IGNvbnRyb2xsZXJz 26121
+IERvZA== 26122
+INCy0LzQtdGB0YLQtQ== 26123
+IHNvcnRpcg== 26124
+IFN0cmFpZ2h0 26125
+IEdyYWNpYXM= 26126
+IGdyb292ZQ== 26127
+IHRvZ2c= 26128
+IOyLtuydgA== 26129
+w6lybw== 26130
+IG91dHdhcmQ= 26131
+IFdB 26132
+IFJvY2t5 26133
+IHNjYW0= 26134
+IGhheWF0 26135
+aWdudHk= 26136
+4oQ= 26137
+cGxpbmdz 26138
+IGFudGliaW90aWNz 26139
+IG5ldmVydGhlbGVzcw== 26140
+amFuZw== 26141
+Y29tbWVyY2U= 26142
+IHNwb2lsZXI= 26143
+IGdsb3Zl 26144
+IGNoYXR0ZXI= 26145
+IEJZ 26146
+fj8= 26147
+IO2YuA== 26148
+IGRlbW9s 26149
+d2VjaHNlbA== 26150
+aW1pcg== 26151
+IHJhaWQ= 26152
+0LXRgNGF 26153
+7J6Q6riw 26154
+ZW5m 26155
+IGNvbW1lbnRlZA== 26156
+IG9wdGltaXplZA== 26157
+IGNvbnZpY3RlZA== 26158
+IGJhdHM= 26159
+IFNC 26160
+IEF1cg== 26161
+IFRvbmc= 26162
+IGltcGxpY2l0 26163
+IEphbmV0 26164
+IHJlYWc= 26165
+IEFkdmFuY2Vk 26166
+IGltcG9zZQ== 26167
+16nXlA== 26168
+IHNjaGVtZXM= 26169
+b3VnaGVy 26170
+YWJvbGlj 26171
+IOqxsOyjoA== 26172
+IHNsb3dpbmc= 26173
+IHd0ZWR5 26174
+IGRlc3RydWN0aXZl 26175
+INC+0L/RgNC10LQ= 26176
+IGxhbmRtYXJr 26177
+IOuPiA== 26178
+IFdhbGtpbmc= 26179
+4bq5 26180
+IHRpamQ= 26181
+IEtO 26182
+IFF1YW50 26183
+7Jik6w== 26184
+INC60YDRgw== 26185
+IHBlcmRlcg== 26186
+IG5vdmU= 26187
+w6RuZGU= 26188
+Ymlh 26189
+IGN1c3RvZHk= 26190
+IGJpb2Q= 26191
+IGRpcmVjdGluZw== 26192
+Li4u4oCL 26193
+IHJlbG9j 26194
+IGRlbWFuZGU= 26195
+IG/En2x1bQ== 26196
+INC+0LTQvdCw 26197
+IE1pbGs= 26198
+IEtyYQ== 26199
+IEhvbmRh 26200
+IHB1ZQ== 26201
+IGVsZWt0 26202
+IGJlZ2lubmVycw== 26203
+IHNwZWFy 26204
+w61uaA== 26205
+IEx1ZnQ= 26206
+IG5pZw== 26207
+IFNjaG9vbHM= 26208
+IGZvcnVtcw== 26209
+IFFpbg== 26210
+cHBv 26211
+IHphZw== 26212
+INCu 26213
+IHRvb3RocA== 26214
+IFN0eWxl 26215
+7LSI 26216
+IHB1bmN0 26217
+IHJlcHM= 26218
+IEFseQ== 26219
+IGFtZW5kbWVudHM= 26220
+IMO2eg== 26221
+IGRpZ2l0cw== 26222
+dXJhaQ== 26223
+IGNoYW90aWM= 26224
+IE1hc3RlcnM= 26225
+ZW9u 26226
+IENhc2g= 26227
+IEN1eg== 26228
+IGJlZGV1dGV0 26229
+IHNjYW5uaW5n 26230
+INC20LQ= 26231
+0L3QtdGC 26232
+IGNlcnRhaW50eQ== 26233
+amVr 26234
+IGRpam8= 26235
+IENsaW1hdGU= 26236
+IHJpbnNl 26237
+IGtyaWo= 26238
+dmVsYW5k 26239
+IHNvdW5kdHJhY2s= 26240
+IFNhZmU= 26241
+IE5vdmE= 26242
+OTQ= 26243
+IGF0aGU= 26244
+IFZlcmI= 26245
+b2xlcg== 26246
+7J207KOg 26247
+IHZpbg== 26248
+IHJlc3BpcmF0b3J5 26249
+IFN0dWR5 26250
+IENBTQ== 26251
+IGF2b2NhZG8= 26252
+IFpoZW4= 26253
+IGxhdGVuY3k= 26254
+IGZlYXRoZXJz 26255
+IGNvbnRhcg== 26256
+INCy0LXRiQ== 26257
+IGZhcms= 26258
+IGJsZW5kZWQ= 26259
+IGV4cGxvZGVk 26260
+IFhY 26261
+IEJlbmlt 26262
+IGFsZ3XDqW0= 26263
+aXN0b2lyZQ== 26264
+IGNvbmZpZGVudGlhbA== 26265
+IG1hc3Q= 26266
+IOy/ 26267
+Z2Vo 26268
+IGRpc3Jlc3BlY3Q= 26269
+IFN5c3RlbXM= 26270
+xrBh 26271
+RWQ= 26272
+IHd5cw== 26273
+IGV4b3RpYw== 26274
+IGdsb3dpbmc= 26275
+w7luZw== 26276
+b3VuZ2U= 26277
+6IQ= 26278
+0LDQvdC40Lc= 26279
+IHBhbGF2 26280
+IFN3b3Jk 26281
+IGdpbQ== 26282
+IENyb3c= 26283
+IHBvdGVudA== 26284
+YmlzaA== 26285
+IGFidXNlZA== 26286
+IEplZA== 26287
+IGdhbWJsaW5n 26288
+IFNwZWN0 26289
+IGludmVzdGlnYXRvcnM= 26290
+IHJhdHQ= 26291
+IGRvYg== 26292
+IERFUw== 26293
+aG9n 26294
+INC+0YLQutGA0Ys= 26295
+7YyF 26296
+INC00LXQvdGM0LPQuA== 26297
+IO2YuQ== 26298
+IOuouOumrA== 26299
+IHNhdHVyYXRpb24= 26300
+IGluaGVyaXRlZA== 26301
+IElubm92YXRpb24= 26302
+7JeI642Y 26303
+IHRhbmdpYmxl 26304
+IGRlcHJp 26305
+aGVk 26306
+INC/0L7QvNC+0LM= 26307
+IHNsaWNlZA== 26308
+4KWN 26309
+IHRo4bq/ 26310
+xaU= 26311
+Njg= 26312
+IGNvcm9uYQ== 26313
+IGdpZnRlZA== 26314
+IHNvaXI= 26315
+IGh1bWlsaXR5 26316
+IOydtOqxuA== 26317
+IGZsYXdz 26318
+INC/0YDQsNC60YLQuA== 26319
+IGthbGQ= 26320
+d2HFvA== 26321
+eXc= 26322
+aXJ0ZWVu 26323
+IGNyb2NoZXRz 26324
+pqzqsIA= 26325
+IOyghOyXkA== 26326
+IGRlc2U= 26327
+INC80LDQsw== 26328
+IGR6aWHFgg== 26329
+IGzDqWc= 26330
+Y2hhbmdpbmc= 26331
+IGxsZXY= 26332
+xYRzaw== 26333
+IDE5ODQ= 26334
+b3Jucw== 26335
+IFdlbHNo 26336
+IHBoYXJtYWNldXRpY2Fs 26337
+IHB1bXBpbmc= 26338
+IFNoYXc= 26339
+cHVuaw== 26340
+IHZhdWx0 26341
+IGtpbmV0aWM= 26342
+IGh1cnJpY2FuZQ== 26343
+IEluY2x1ZGluZw== 26344
+4bupYw== 26345
+IEdyYW5kcGE= 26346
+YW5zaGlw 26347
+INCy0YvRhdC+0LQ= 26348
+0L3QvtC2 26349
+nKA= 26350
+dXR0YQ== 26351
+IOqygeuLiOuLpA== 26352
+IGJheg== 26353
+INC/0L7RiA== 26354
+IHBlY3VsaWFy 26355
+ennEhw== 26356
+IEVsbGll 26357
+IGxlYXJucw== 26358
+IEtyaXNobmE= 26359
+IGNvbnNlY3V0 26360
+IGVtcGF0aA== 26361
+IERpbg== 26362
+IHRyYWRlZA== 26363
+IEJvcmlz 26364
+dWdnYWdl 26365
+b2xsYQ== 26366
+INC90LDQt9Cy 26367
+IGV0ZXJuaXR5 26368
+INCy0L8= 26369
+w6htZXM= 26370
+IGdyYXBw 26371
+YsOp 26372
+INC/0YDQtdC00YHRgtCw0LI= 26373
+IEZD 26374
+jeuLiOuLpA== 26375
+ZXZlbg== 26376
+IE5lYnJhc2th 26377
+b3J0dW5l 26378
+IGthcmVuYQ== 26379
+IEFnZW50 26380
+IHN0aW5n 26381
+IFBJ 26382
+IG11bmljaXBhbA== 26383
+cG93ZXJlZA== 26384
+IGNvbnNlZ3Vl 26385
+IE1hbmNoZXN0ZXI= 26386
+IHJhaW55 26387
+IGJsaQ== 26388
+IGtvc3Q= 26389
+IGhhbHRlbg== 26390
+IEFoaGg= 26391
+aW5zdWxh 26392
+ZXJ0aW5n 26393
+INin2YTZgQ== 26394
+IHJlbGFjaW9u 26395
+IGtvbWVu 26396
+IGRvbWU= 26397
+IHByaWVzdHM= 26398
+IEludHJvZHU= 26399
+cm9waGU= 26400
+c2hvcmU= 26401
+dmVsdA== 26402
+Y2xpcHNl 26403
+INGA0YPRgQ== 26404
+15nXoQ== 26405
+IHNhYmVtb3M= 26406
+IEhvbGxhbmQ= 26407
+b2dp 26408
+YW5raQ== 26409
+IE1hdHM= 26410
+IHNtb2tlZA== 26411
+dWxsaWU= 26412
+IGV1cm9wZQ== 26413
+INC00LXQudGB0YLQstC40YLQtdC70YzQvdC+ 26414
+IGJhcmR6aWVq 26415
+IHRyYW5zZm9ybWluZw== 26416
+IEV6 26417
+b3BhdGg= 26418
+IOyWuOuLiA== 26419
+INGB0YLQsNC9 26420
+4bqxbmc= 26421
+4Lix4LmJ 26422
+IE91Y2g= 26423
+IGNsZWFyYW5jZQ== 26424
+dXN0YWlu 26425
+IHNvbGlkYXJpdHk= 26426
+IHByb3Zpbmc= 26427
+INCY0L0= 26428
+INGB0Yo= 26429
+IHByb2xvbmc= 26430
+0LDQtNC90L4= 26431
+IHNvcw== 26432
+IERlYWw= 26433
+IDE3MA== 26434
+bW9ucw== 26435
+INC30LXQvA== 26436
+IGxvZ2dlZA== 26437
+IGxpZmVsb25n 26438
+IHNlbnNvcnk= 26439
+IGJlaG9sZA== 26440
+IEZBUg== 26441
+w6h0ZW1lbnQ= 26442
+IEZlZGVyYXRpb24= 26443
+IGRvZGdl 26444
+IFNoaXI= 26445
+IGRyYWdvbnM= 26446
+IEFyY3RpYw== 26447
+xIXFvA== 26448
+xY0= 26449
+wro= 26450
+IGRlbmtl 26451
+IHBvZHLDrWE= 26452
+Y29sZQ== 26453
+0YPQu9GM0YLQsNGC 26454
+IHN5c3RlbWF0aWM= 26455
+0LDQvNCw 26456
+Y2hvcw== 26457
+IGNsaW5pY3M= 26458
+IEJT 26459
+IHRhbGVz 26460
+dXNpb25z 26461
+IO2IrA== 26462
+IHByZXNlcnZhdGlvbg== 26463
+IGxvcmU= 26464
+IFByb3Rlc3Q= 26465
+4bub 26466
+IGFja25vd2xlZGdlZA== 26467
+IElzYWlhaA== 26468
+IOuVjOuKlA== 26469
+INeY 26470
+IGNvbXBldGl0b3I= 26471
+IGFkdmFuY2luZw== 26472
+emlw 26473
+IHRlbnRo 26474
+IExhdXJl 26475
+IGhpbnRz 26476
+IGV4ZXJjaXNpbmc= 26477
+npzr 26478
+IEludGVsbGlnZW5jZQ== 26479
+dWF0ZWQ= 26480
+T1VU 26481
+b3BlZA== 26482
+IGF1dG9ub215 26483
+IGJyYW5kaW5n 26484
+IE1lZGl0ZXJyYW5lYW4= 26485
+0ZbQug== 26486
+IHNjcmV3ZHJpdmVy 26487
+IHN1cHJl 26488
+IHN0YXA= 26489
+IGp1cmlzZGljdGlvbg== 26490
+IFNldHRpbmdz 26491
+IGZvcmVmcm9udA== 26492
+IEZlbWFsZQ== 26493
+Y29tZm9ydA== 26494
+IG11bHRpcGxpY2F0aW9u 26495
+IE11cnJheQ== 26496
+IGJvYg== 26497
+IFRhcw== 26498
+IHRhaHU= 26499
+IG9udW4= 26500
+ZXR0ZXI= 26501
+IHByb3BoZXRz 26502
+bGFn 26503
+IHJldmVudWVz 26504
+IHByw6E= 26505
+IHVwbG9hZGluZw== 26506
+IG1hY2hpbmVyeQ== 26507
+YXNjYWw= 26508
+IEVzdMOh 26509
+IEdvdGg= 26510
+IEJhbGQ= 26511
+IFNhdw== 26512
+IHN0cmlwZXM= 26513
+7KCR 26514
+IHBvd2lu 26515
+IGhvc3RpbGU= 26516
+IGRhcnVt 26517
+IHByZXZlbnRlZA== 26518
+0L7QttCw0LvRg9C50YHRgtCw 26519
+IGFsZ3VuYXM= 26520
+IGhvcGVsZXNz 26521
+IHpuYWo= 26522
+IHJlYWRpbmdz 26523
+IGNyYXZpbmc= 26524
+dGF0 26525
+IFBpZw== 26526
+IGxpYXI= 26527
+IG11bHRpcGxheWVy 26528
+IGRhbGU= 26529
+IENvdXJzZQ== 26530
+7YG8 26531
+IEtpdGE= 26532
+IGN1c3RvbXM= 26533
+IHJlc3BvbmRz 26534
+ZW5kcmE= 26535
+IG1ldHJv 26536
+0YHQvtC7 26537
+IG1pdGlnYXRl 26538
+IG9wcHJlc3Npb24= 26539
+cXVpbmhv 26540
+IGFtbW8= 26541
+IGVuZmVy 26542
+IHBvbnk= 26543
+IG91bmNlcw== 26544
+sJQ= 26545
+IOyImOqwgA== 26546
+IGRpY2hv 26547
+IERlYg== 26548
+IHdvbmRlcnM= 26549
+IFJvb3Nl 26550
+IHByaXplcw== 26551
+IEFMRVg= 26552
+IHRoYW5rZnVsbHk= 26553
+IHRpc3N1ZXM= 26554
+INGA0LDQstC90L4= 26555
+IEx1bmE= 26556
+aW50ZWxsaWdpYmxl 26557
+IOyZuA== 26558
+6rCR 26559
+IEhlYXQ= 26560
+INGB0LjQtA== 26561
+IFF1aQ== 26562
+IGlvbnM= 26563
+IGFjY29tbW9kYXRpb24= 26564
+IEthcnQ= 26565
+aWVuc3Q= 26566
+IHRhcmRl 26567
+IHNvYWtlZA== 26568
+IENhc2V5 26569
+IOy0nQ== 26570
+INGA0YPQsQ== 26571
+IGRpZmZlcmVudGk= 26572
+IGxlZnRvdmVy 26573
+IGV4Y2hhbmdlcw== 26574
+c2Vjb25k 26575
+IGZpcnN0bHk= 26576
+IGJ1aWxkZXI= 26577
+cmllbg== 26578
+IGR3 26579
+IGJvdW5jaW5n 26580
+Pzwv 26581
+IOuMgO2VtOyEnA== 26582
+INGB0LU= 26583
+IE1pbGVz 26584
+aWVuaWU= 26585
+INC/0L7QtNC/0LjRgQ== 26586
+IOustA== 26587
+IGFyaXNlcw== 26588
+IHN1YmNvbnNjaW91cw== 26589
+IFNhbmR5 26590
+IGxvdHRlcnk= 26591
+4oCR 26592
+YW1pbGlhcg== 26593
+IGNvb3JkaW5hdG9y 26594
+6Iw= 26595
+IGV4dHJhb3JkaW4= 26596
+IFJvbmFsZA== 26597
+IE1PTg== 26598
+Z3JlZW4= 26599
+IG1hbnVmYWN0dXJl 26600
+IFJlY29yZA== 26601
+IE1hcmtldGluZw== 26602
+0LjRhg== 26603
+IGNyZWRlbnRpYWxz 26604
+IHVwcmlnaHQ= 26605
+IEhlcml0YWdl 26606
+IGfDtnJk 26607
+ZXhwZW5zaXZl 26608
+4bqtbg== 26609
+IOyxhA== 26610
+IG91dGxpbmVk 26611
+IE9vb2g= 26612
+b3JpZW50ZWQ= 26613
+IHdpcmVk 26614
+IG91dGxldHM= 26615
+IGh1Z2VseQ== 26616
+IO2WiOuKlOuNsA== 26617
+0LDRgNGC 26618
+IGxvZ2lzdGljcw== 26619
+IHNlYXNvbmFs 26620
+IGRlYmU= 26621
+IHRoZW9y 26622
+IHBpcmF0ZQ== 26623
+YXBweQ== 26624
+IGtub3Rz 26625
+IGZlbW1l 26626
+IFNvZnR3YXJl 26627
+Z2VuZGU= 26628
+0YLQsNC60Lg= 26629
+IHRlbXBsZXM= 26630
+IGxpbWl0YXRpb24= 26631
+IGFtcGxpdHVkZQ== 26632
+IGhhY2Vu 26633
+IGF1ZGk= 26634
+IOuWqA== 26635
+IFdhaGw= 26636
+IG5paA== 26637
+IGFtcGxpZmllcg== 26638
+YXJpdXM= 26639
+aXphZG8= 26640
+YWNoYQ== 26641
+IGt1bGxhbg== 26642
+IFR3aW4= 26643
+IEZvcmNlcw== 26644
+IGFicmly 26645
+IEVQQQ== 26646
+IEFoYQ== 26647
+IOq3uOuemOuPhA== 26648
+IGJpb20= 26649
+INCi0LDQvA== 26650
+IHNhaWxpbmc= 26651
+IEpva2Vy 26652
+Rmlyc3Q= 26653
+fl0= 26654
+b3JzY2g= 26655
+IHbDpnJl 26656
+IGJlZXRqZQ== 26657
+IFNwYcOf 26658
+cG9saXQ= 26659
+IHR1cmJ1bA== 26660
+IOyggO2drOqwgA== 26661
+IGNpYw== 26662
+IERyYWtl 26663
+IEJSSQ== 26664
+aXphw6fDo28= 26665
+IOyeiOuLpA== 26666
+IEx5bm4= 26667
+IHRyYW5zZ2VuZGVy 26668
+IHJlc2lnbg== 26669
+IGNoYXJ0ZXI= 26670
+IEpI 26671
+IEhvbG1lcw== 26672
+IExpcA== 26673
+ZGFz 26674
+IHBlZGlhdHJpYw== 26675
+IG1lbW9yaXpl 26676
+IGV2YWx1YXRpbmc= 26677
+IPCfkA== 26678
+Y2Fr 26679
+IGNvbmp1bmN0aW9u 26680
+IHJlc2VydmVz 26681
+IHNoYW1wb28= 26682
+IGp1ZGdlZA== 26683
+IHdpZHo= 26684
+VklO 26685
+IGFib2FyZA== 26686
+YXJpcw== 26687
+IFJvaA== 26688
+IGNvb2xlZA== 26689
+0YHRgtC1 26690
+Y2Vw 26691
+cm9zdA== 26692
+aG90cw== 26693
+IE1lbGJvdXJuZQ== 26694
+0L7Rh9GM 26695
+IHZlbnRpbA== 26696
+0LjQvdC+0LI= 26697
+IG1vdGlvbnM= 26698
+7JeI64qU642w 26699
+0LzQtdGA0LjQug== 26700
+IENoYXQ= 26701
+IGdvdXZlcm5lbWVudA== 26702
+IEtpdm9s 26703
+IEtpdm9sb3dpdHo= 26704
+IG7Ds2k= 26705
+INC60YPQtNCw 26706
+IGh5ZHJhdWw= 26707
+IEJlcmc= 26708
+eWx1bQ== 26709
+IFByw6RzaWRlbnQ= 26710
+cm9weQ== 26711
+IHNlbWlj 26712
+0Y/QtdGC 26713
+IENhcGU= 26714
+IGNhbmU= 26715
+IGJyaW5nZW4= 26716
+IHdpcmluZw== 26717
+dW55YQ== 26718
+IHJlcGF5 26719
+qqk= 26720
+IHdvbnQ= 26721
+w6FudA== 26722
+IGdvdmVy 26723
+IExpYmVydHk= 26724
+IGVsZWN0cm9tYWdu 26725
+IFNpbmdo 26726
+INCz0YDRg9C/ 26727
+0LPQvtCy 26728
+iOustOs= 26729
+IFJ1bGU= 26730
+IHVuZGVyd2F5 26731
+IEZyZWRlcg== 26732
+IHR1cmJpbmU= 26733
+aXNoaQ== 26734
+IGbDrXM= 26735
+IEN1bHR1cmU= 26736
+YWNyZQ== 26737
+IHdhbmRlcg== 26738
+IGd1ZXJyYQ== 26739
+IHPDtnk= 26740
+IEp1cg== 26741
+YXdheXM= 26742
+IHNjaHdpZXI= 26743
+Z3VhcmQ= 26744
+IEFiZA== 26745
+dWN0aW9u 26746
+IGFya2FkYcWfbGFy 26747
+IEhhbWI= 26748
+Py4= 26749
+c2l6ZQ== 26750
+IE9ydGg= 26751
+IHN3YXk= 26752
+IM6U 26753
+IGFic29ycHRpb24= 26754
+aW5lZXM= 26755
+IHBhdHJvbnM= 26756
+IGJlYWNoZXM= 26757
+R0c= 26758
+IGNvbnRhbWlu 26759
+aW50ZW5kZW50 26760
+INC90YDQsNCy 26761
+INC00LXRgNC2 26762
+IHF1aWx0 26763
+IGV2b2x1dGlvbmFyeQ== 26764
+7J206528 26765
+YXppb25p 26766
+IGVya2w= 26767
+IEJ1dGxlcg== 26768
+IGRvbw== 26769
+IG5lZ290aWF0aW9u 26770
+ZW5kdW0= 26771
+IHRlcm1pbm9sb2d5 26772
+IGt1bA== 26773
+IFVudGVybmVobWVu 26774
+w6lyaWM= 26775
+eGk= 26776
+YmFk 26777
+INC00L7Qu9C20L3Riw== 26778
+IE1pdGNoZWxs 26779
+dGhyZWU= 26780
+IHN1YnN0cmF0ZQ== 26781
+IEluaGFsZQ== 26782
+IEFncmlj 26783
+dW5nZQ== 26784
+INC30YA= 26785
+IGFkdmVyc2U= 26786
+IOyggOuPhA== 26787
+IHBpbGxhcg== 26788
+IE1pbnV0ZW4= 26789
+IE1hdGU= 26790
+IFBsYXR6 26791
+IGhlbHBsZXNz 26792
+IGFsYXI= 26793
+IGZyZW5jaA== 26794
+IGFsbG9jYXRpb24= 26795
+IHN0ZW1z 26796
+IG1hcmF0aG9u 26797
+IEhBUkY= 26798
+aXphY2nDs24= 26799
+SmVzcw== 26800
+INC30L3QsNGH 26801
+IGRlY2xhcmF0aW9u 26802
+RUVSSU5H 26803
+c3RlcmRhbQ== 26804
+YXNzaXVt 26805
+IHNlaXo= 26806
+IHByZXNpZGVudHM= 26807
+dGFrZQ== 26808
+IHdpbGRlcm5lc3M= 26809
+IGNvc21pYw== 26810
+IOuqqOuRkA== 26811
+c3Rybw== 26812
+IHBvd2llZHo= 26813
+IE1hZ2F6aW5l 26814
+IFZJ 26815
+INC00LXRgA== 26816
+IHfDvHJkZW4= 26817
+IHRhYmxldHM= 26818
+IHBpZXJ3cw== 26819
+IG1vcnRhbA== 26820
+IHN1cHBsaWVk 26821
+IE7Ds3M= 26822
+IFByb3Blcg== 26823
+INC60LDQttC00YvQuQ== 26824
+b2zDs2c= 26825
+67Cp 26826
+IG1pc2Nvbg== 26827
+IHByb3hpbWl0eQ== 26828
+IEFsbGVz 26829
+INCz0LvQsNC3 26830
+IGxhbWU= 26831
+IHZpYmVz 26832
+IGRlZW1lZA== 26833
+IHVyaW5l 26834
+IHJlbWluZGluZw== 26835
+IGNpcmN1bXN0YW5jZQ== 26836
+65Ok7J20 26837
+IGxhcHRvcHM= 26838
+wrI= 26839
+7ZW07JW8 26840
+IE9tZWdh 26841
+Tlk= 26842
+IHB1bXBz 26843
+IHJhaWxz 26844
+IHN1cnBhc3M= 26845
+IEJyb3M= 26846
+IG5hdGlvbmFsbHk= 26847
+IGdld2VzZW4= 26848
+s7Tri6Q= 26849
+b3NoaW5n 26850
+6rCI 26851
+IGNyaWFu 26852
+IOyCrOuejOydtA== 26853
+Y2F1c3Q= 26854
+0YbQuNC/ 26855
+IE9iZXI= 26856
+IERBWQ== 26857
+IENhbm9u 26858
+enVuZw== 26859
+IOqwlg== 26860
+INCw0LLRgtC+0Lw= 26861
+IGRpdm9yY2Vk 26862
+15nXpA== 26863
+z4HOtQ== 26864
+Y2VsYW5k 26865
+Y2llcg== 26866
+0YDQtdC3 26867
+VG9kYXk= 26868
+IG9yYml0YWw= 26869
+IHN0cmV0 26870
+IHNhdHU= 26871
+IO2BrOs= 26872
+em9z 26873
+IFNjbw== 26874
+zrzOrQ== 26875
+IEd1YXJkaWFu 26876
+aW50ZXJlc3Q= 26877
+IFZFUg== 26878
+w7xuZGVu 26879
+INGF0L7RgtC10Ls= 26880
+dGl0 26881
+Qnk= 26882
+IGFubGF0 26883
+U2hvdw== 26884
+IG9pbHk= 26885
+IGxlZ2VuZHM= 26886
+IHNwZWN1bGF0aW9u 26887
+IFdpc2g= 26888
+IG1vbms= 26889
+R0FO 26890
+IGjhu40= 26891
+IGRhbmdlcnM= 26892
+IEJlbmU= 26893
+aXF1ZW1lbnQ= 26894
+IOuCmOyZgA== 26895
+INCw0LQ= 26896
+IGRpc2NyZXRl 26897
+w4c= 26898
+IGNvbmRpdGlvbmFs 26899
+IEdpbGw= 26900
+dWF0ZXM= 26901
+INGB0L7QstGB0LXQvA== 26902
+IHNjcmVlbnNob3Q= 26903
+Y2Fkbw== 26904
+IOuqqOuToA== 26905
+IGZpbmdlcnRpcHM= 26906
+IE1BQw== 26907
+IGR1ZGVz 26908
+Y29zdA== 26909
+IGJ1bXBz 26910
+b25kbw== 26911
+IGRhdG9z 26912
+IGJlZXBz 26913
+IFByb24= 26914
+IEtoYWw= 26915
+emVnbw== 26916
+IEFiYnk= 26917
+VWg= 26918
+WW8= 26919
+IFRlbA== 26920
+IM68zq0= 26921
+S0k= 26922
+IHN0cmVzc2Vz 26923
+IHNwcmVhZHNoZWV0 26924
+IE5PVw== 26925
+REI= 26926
+IGxpYmVyYXRpb24= 26927
+IHByZWRpY3RhYmxl 26928
+IFF1ZXN0aW9ucw== 26929
+IHNwYWNpbmc= 26930
+IGluaGFiaXRhbnRz 26931
+IHp3acSFeg== 26932
+IFNBUA== 26933
+IGx1Z2dhZ2U= 26934
+IGhpcHA= 26935
+6JY= 26936
+IHRhbmdlbnQ= 26937
+IHbDpQ== 26938
+0LDQu9GM0L3QvtC5 26939
+c2VoZW4= 26940
+IHByb2Nlc3NvcnM= 26941
+IGZpbmRldA== 26942
+IGNhcnRyaWRnZQ== 26943
+IGFkbWluaXN0cmF0b3Jz 26944
+IOyWtOya 26945
+IHN1cHJlbWU= 26946
+IEFudGk= 26947
+IO2UhOuhnA== 26948
+IGluZm9ybWF0aXZl 26949
+IGtvbXQ= 26950
+15nXmA== 26951
+QXNzaXN0YW50 26952
+IGxpc3Rh 26953
+w7ZsbA== 26954
+IGRpc3RpbmN0aXZl 26955
+IEh1ZA== 26956
+IHNhbG9u 26957
+bcOqbWU= 26958
+IE1vdGlvbg== 26959
+IHNldWxlbWVudA== 26960
+IE1lbnNjaA== 26961
+IHB1bXBlZA== 26962
+w7xoZXI= 26963
+aWJv 26964
+IHdhxbw= 26965
+IHF1YW50aXRhdGl2ZQ== 26966
+2b4= 26967
+IOuqqOyKtQ== 26968
+IHBvdWNo 26969
+IFRoZWF0cmU= 26970
+YWhp 26971
+IHNwaW5hY2g= 26972
+IHJlYWxpdGllcw== 26973
+IGxleQ== 26974
+IE1hcnRoYQ== 26975
+IHJlY2hlcg== 26976
+ZWNoZXM= 26977
+IHBlcmlvZGlj 26978
+b2NpZGU= 26979
+IEluY3JlZA== 26980
+IHRo4bqleQ== 26981
+b3Rvbg== 26982
+IEVzbw== 26983
+IGfDqW7DqXJhbA== 26984
+aWxpZ2h0 26985
+IGltYWdpbmluZw== 26986
+aGVh 26987
+ZXRpY2Fs 26988
+4but 26989
+IERlbW9rcmF0 26990
+IGVuam8= 26991
+IGFkanVzdGFibGU= 26992
+IHJhaW5z 26993
+aWV3YcW8 26994
+IGp1c3RlbWVudA== 26995
+IGp1c3RpZmllZA== 26996
+IFNoYWtl 26997
+dml2 26998
+7IKs66W8 26999
+IG1ldHQ= 27000
+IEVudmlyb25tZW50YWw= 27001
+IHNvbGFtZW50ZQ== 27002
+IGludGVyc2VjdA== 27003
+IDE5ODg= 27004
+IHNpbXVsYXRl 27005
+SkE= 27006
+INC30LDRgQ== 27007
+IGNvbnRpbmc= 27008
+IFRlaw== 27009
+IHRvcmNo 27010
+INC00YDRg9Cz0L7QuQ== 27011
+IGluc2NyZQ== 27012
+IG1vZGVsbw== 27013
+IEdlZw== 27014
+IERlbW9jcmF0 27015
+0LrQsg== 27016
+IEJ1ZGR5 27017
+IHJlZHVuZA== 27018
+IGNyYWZ0cw== 27019
+IEhpag== 27020
+IGp1ZQ== 27021
+IEtpcms= 27022
+IGthYg== 27023
+4buj 27024
+IGFlc3RoZXQ= 27025
+IEpPTg== 27026
+IHN1cGVyY29t 27027
+INGB0LjRgtGD 27028
+IM+Mz4TOuQ== 27029
+2YXZhg== 27030
+IEVWRVI= 27031
+7JWY7Ja0 27032
+b2l0 27033
+IENsZXZlbGFuZA== 27034
+IHNpeHRlZW4= 27035
+IHdhdGVyZmFsbA== 27036
+77g= 27037
+aW5mbA== 27038
+IGNvdW5zZWxvcg== 27039
+IFB1bms= 27040
+IHNwcmVjaGVu 27041
+ZXhj 27042
+IFNraWxscw== 27043
+cm96 27044
+YWRhbWVudGU= 27045
+IHBhbmNha2Vz 27046
+6riw66Gc 27047
+IHBsYW5r 27048
+IHNvdmVyZWlnbnR5 27049
+IGZ1aQ== 27050
+INC90LXQvtCx 27051
+IFdpaQ== 27052
+IFNjaG9s 27053
+4oCO 27054
+IFNwZWFr 27055
+Y2lsaWF0aW9u 27056
+IHRoaWdo 27057
+IOqxsOydmA== 27058
+IGpvdA== 27059
+IOy0rOyYgQ== 27060
+INmF24zaug== 27061
+IENDUA== 27062
+INC/0L7RgdGC 27063
+IG9ic2VydmVy 27064
+w6Fi 27065
+IHN0aWdtYQ== 27066
+IHByb3ByaWV0 27067
+IGNpZGFkZQ== 27068
+IGJhxZ9rYQ== 27069
+2LnYqQ== 27070
+a3Jl 27071
+IHBvd2llZHppZcSH 27072
+IGNlYXNl 27073
+IHNraW5z 27074
+IHZlZ2dpZXM= 27075
+IG9wcG9zaW5n 27076
+b3BvbHk= 27077
+IEp1Zw== 27078
+IFlvb24= 27079
+IFVuaXQ= 27080
+IDE5ODY= 27081
+IGtvbnM= 27082
+IGRpYWdub3N0aWM= 27083
+IGVtcG93ZXJlZA== 27084
+IHRobw== 27085
+IGNlbg== 27086
+w6lyYXRpb24= 27087
+INGX 27088
+IHBoeXNpYw== 27089
+IFByYWN0aWNl 27090
+IFNvdXRoZWFzdA== 27091
+IEVzcGE= 27092
+IEdlb3I= 27093
+cm9wb3J0aW9u 27094
+IHNwZWNz 27095
+IGFkYXB0aXZl 27096
+IFVuaXR5 27097
+IFdvcmtz 27098
+dWdlbg== 27099
+IE1vbnRhbmE= 27100
+VGhhbmtz 27101
+IHdoaXBwZWQ= 27102
+IGR1bmdlb24= 27103
+IHZpdGFtaW5z 27104
+U1A= 27105
+IHNjYW5kYWw= 27106
+IGRpbmVybw== 27107
+b3Zh 27108
+IGVtYnJv 27109
+IEVhZ2xl 27110
+IHRoZW9sb2d5 27111
+IFZhbmVzc2E= 27112
+IEFJRFM= 27113
+65Cc 27114
+IGZyZWVs 27115
+IEFsemhlaW1lcg== 27116
+IMWa 27117
+SGVy 27118
+IHRvcm5hZG8= 27119
+YWdlbnM= 27120
+IOyeiOyWtOyEnA== 27121
+IFRyYW5zZm9ybQ== 27122
+IHByb2Nlc3Nv 27123
+IG1pbGxpc2U= 27124
+IHByb2Zlc3Npb25hbGx5 27125
+IG1lbWI= 27126
+b2NhdGlvbg== 27127
+IHN0eWxpbmc= 27128
+INC+0LHRj9C3 27129
+IE9wZXJhdGlvbg== 27130
+IHd5Z2w= 27131
+IFJhbg== 27132
+IEtpbg== 27133
+4buxYw== 27134
+IEJBUg== 27135
+IHBhcGVyd29yaw== 27136
+IHR1bGU= 27137
+IHF1ZXJpYQ== 27138
+IGNvbXBseQ== 27139
+IEhhaXI= 27140
+15nXmw== 27141
+INC/0YDQvtGB0YI= 27142
+IG11dGF0aW9u 27143
+IHJlcHLDqXM= 27144
+IG9jdG9wdXM= 27145
+IGltcG9ydGFudGVz 27146
+IGRlc2VydmVk 27147
+ZXRy 27148
+IGRpc2FzdGVycw== 27149
+bMSxbmRh 27150
+aXF1w6k= 27151
+IERlc2hhbGI= 27152
+c29v 27153
+b3NzaXA= 27154
+IHJlbGlldmVk 27155
+IENvbGxpbnM= 27156
+IHdhdGVycHJvb2Y= 27157
+IFl1aw== 27158
+IGNvcHlpbmc= 27159
+IGLDvHTDvG4= 27160
+IEhldXRl 27161
+IEVudHJl 27162
+IHJlc2lkdWFs 27163
+IGNvbG9uaWVz 27164
+IMOpbm9ybQ== 27165
+IEVyaW4= 27166
+IHN0YW4= 27167
+IHRyZW1lbmRvdXNseQ== 27168
+IGNhcHR1cmVz 27169
+IFNhaQ== 27170
+w6JjZQ== 27171
+IG1pYcWC 27172
+IDg3 27173
+IGxvZ2dpbmc= 27174
+IGluc2VydGVk 27175
+IGluaGVyZW50bHk= 27176
+7J2R 27177
+bGF2ZQ== 27178
+0L3QuNGH 27179
+IGZlbW1lcw== 27180
+IGTDqXA= 27181
+dWtz 27182
+YWNpYQ== 27183
+IFdhZGU= 27184
+IGppag== 27185
+IFZpbmNlbnQ= 27186
+IEljZWxhbmQ= 27187
+aGVt 27188
+IGFwb2xvZ3k= 27189
+IFBlZw== 27190
+IGdsdWVk 27191
+IGNvbXBhbmlvbnM= 27192
+IExpdmVy 27193
+IGNyaXRpY2l6ZWQ= 27194
+bGVhZGluZw== 27195
+IHPDpGdh 27196
+IHNxdWlk 27197
+IG5hcnJhdGl2ZXM= 27198
+IHRha2E= 27199
+bmV6 27200
+d2VpdA== 27201
+IHRyaXBvZA== 27202
+IGV4cGxpYw== 27203
+IHNwaW5hbA== 27204
+IGFwcHJveGltYXRpb24= 27205
+IHBhZ2Fy 27206
+IENhbHZpbg== 27207
+INCy0LXQtNGM 27208
+IGxhYw== 27209
+IHByb2FjdGl2ZQ== 27210
+IFRyYWlu 27211
+b3Jm 27212
+IHN0ZW4= 27213
+IGdyYXBlcw== 27214
+IG1ldXM= 27215
+IGF1dG9tYXQ= 27216
+IGJpYXNlZA== 27217
+IGNoYcOubmU= 27218
+Y29hbA== 27219
+IHJlbmNvbnQ= 27220
+IEt1bQ== 27221
+IGZlc3RpdmFscw== 27222
+IHN0YXJ0dXBz 27223
+IGFrYQ== 27224
+IGN5bGluZA== 27225
+c25h 27226
+Q1JJ 27227
+IHJlc3VsdGFkbw== 27228
+IG1pbGVzdG9uZQ== 27229
+IM+F 27230
+IHRlbGVwb3J0 27231
+enljaA== 27232
+NjI= 27233
+IEZlYXI= 27234
+IG51Y2xldXM= 27235
+IHNoaW5lcw== 27236
+aG92 27237
+IFBhcnRuZXJz 27238
+IEthcw== 27239
+IG5hZGll 27240
+IGFsZXJ0cw== 27241
+IEJJTEw= 27242
+c3Ryb25n 27243
+IE5hdGU= 27244
+IERlbm1hcms= 27245
+IENhdg== 27246
+T1NU 27247
+aMOkbHQ= 27248
+IOyVhOuLjA== 27249
+YW55b24= 27250
+IGVuY291cmFnZXM= 27251
+INC/0L7RgdGC0LDQsg== 27252
+IEh1YW5n 27253
+U1RB 27254
+IHBhaW50cw== 27255
+IHNjaGVkdWxlcw== 27256
+IGNoZWF0ZWQ= 27257
+IGFwcHJveA== 27258
+IO+3 27259
+IMK7Lg== 27260
+IHNtaWxlcw== 27261
+aXN1cmU= 27262
+IG5lcmVk 27263
+YXJkZW4= 27264
+IGN1cnQ= 27265
+IOuM 27266
+IFJvdGg= 27267
+IHB1aXNxdWU= 27268
+IEdFVA== 27269
+IFZlZ2V0 27270
+IHByb2R1eg== 27271
+IEJlbGdpdW0= 27272
+IENhbXB1cw== 27273
+16jXmded 27274
+aWN1dA== 27275
+INGB0L3QuNC8 27276
+IHLDqXVzcw== 27277
+IHNsaXBwZXJ5 27278
+IEV3 27279
+xbM= 27280
+IExlZ2VuZHM= 27281
+IFRpZmZhbnk= 27282
+0LDQu9C40Lc= 27283
+INC/0LXRgNC10LI= 27284
+INC+0LPRgNC+0Lw= 27285
+IGNyb3M= 27286
+IENF 27287
+QnU= 27288
+IGVuc3VyZXM= 27289
+IGdyYW5kY2hpbGRyZW4= 27290
+IGFjdWVyZG8= 27291
+IHByaXNvbmVy 27292
+IHRoaXJzdHk= 27293
+YmFuZQ== 27294
+IOu5oA== 27295
+IMO6bHRpbWE= 27296
+IExhdW5jaA== 27297
+bml0eQ== 27298
+IGNvbWJ1c3Rpb24= 27299
+IHVuaWNvcm4= 27300
+IGZhbWlsbGU= 27301
+IGxvd2VyaW5n 27302
+IFlpbmc= 27303
+YnVpbGRpbmc= 27304
+IGR1bw== 27305
+IE3DqXhpY28= 27306
+YXN0aWFu 27307
+IOuoueydhA== 27308
+IFJhbHBo 27309
+IHJld3JpdGU= 27310
+IGdsYW0= 27311
+aWZpcXVl 27312
+RXI= 27313
+IFJ1bm5pbmc= 27314
+0L7QvdC+0LI= 27315
+IG1lYW5pbmdz 27316
+IGNoZXd5 27317
+IExlc2xpZQ== 27318
+IGZpbmVzdA== 27319
+IGhhaGFoYQ== 27320
+IFNURVA= 27321
+IGxvbmVsaW5lc3M= 27322
+cmlhbnM= 27323
+IHF1ZXN0aW9uZWQ= 27324
+IGVzcXVl 27325
+IHNpbmtpbmc= 27326
+IHBlc28= 27327
+IFdyb25n 27328
+YXNtaW5l 27329
+IGRlZmluaXRpdmU= 27330
+IGJ1eXM= 27331
+IGNydWM= 27332
+Y29vbA== 27333
+IOugiA== 27334
+IHDDsw== 27335
+IHV0aWxpemVk 27336
+IHdvcnRod2hpbGU= 27337
+IER5bGFu 27338
+RVNF 27339
+IHZlcnRleA== 27340
+dMSx 27341
+IEZpcg== 27342
+IHphdw== 27343
+IEdlZA== 27344
+INCd0LDQvw== 27345
+ZHo= 27346
+IGN1cnNvcg== 27347
+IHN3aXBl 27348
+IGluZXZpdGFibHk= 27349
+IHBvc3RlcnM= 27350
+IGluY2xpbmVk 27351
+IGdyZWV0aW5n 27352
+IGRpc2FwcG9pbnRtZW50 27353
+IHJlbGHDp8Ojbw== 27354
+VFQ= 27355
+IHJhYmI= 27356
+IE1haW5l 27357
+IGFuYWx5emVk 27358
+RkU= 27359
+INCf0L7Quw== 27360
+IFNhbmRyYQ== 27361
+IHBsYWd1ZQ== 27362
+QVJF 27363
+IHbDpHI= 27364
+IFZpdg== 27365
+dW1lZA== 27366
+aGFuZG8= 27367
+aG91ZXR0ZQ== 27368
+IEJhaWxleQ== 27369
+eXNvbg== 27370
+IHNlbXVh 27371
+IGhhcmRjb3Jl 27372
+4oKs 27373
+0ZbQvA== 27374
+w6lyYQ== 27375
+T1RI 27376
+IGZvcmVpZ25lcnM= 27377
+IFBhbGVzdGluaWFu 27378
+IHByb3ByaW8= 27379
+0LDQvdC40Lk= 27380
+IG15dGhz 27381
+V0g= 27382
+IG5pbnRo 27383
+IENyZWF0b3I= 27384
+0LvQvtC8 27385
+IEZsaXA= 27386
+IGVtYW4= 27387
+IGtpxZ8= 27388
+emllaA== 27389
+IEVhcm5lc3Q= 27390
+c3lzdGVt 27391
+luyXkA== 27392
+IGFybWllcw== 27393
+IE91dHNpZGU= 27394
+IGhhcnVz 27395
+0L7QtNCw0YA= 27396
+IHZpc2l0b3I= 27397
+IHN0cmVuZ3RoZW5pbmc= 27398
+IDky 27399
+dmlv 27400
+IOumrA== 27401
+IGdyZWVkeQ== 27402
+IHBvcXVpdG8= 27403
+dWRlcg== 27404
+IEtvcGY= 27405
+IOuLpOydjOyXkA== 27406
+IHNlaXM= 27407
+w6F0aWNv 27408
+IHRydXN0aW5n 27409
+w61w 27410
+IEVtbQ== 27411
+bGVlbg== 27412
+INin2YTZhg== 27413
+IHJlY3J1aXRtZW50 27414
+IEZpbGlw 27415
+INmD2YQ= 27416
+Q2xpbnQ= 27417
+INCy0LXRgQ== 27418
+YXVmdA== 27419
+IGRvbWluYXRl 27420
+IHJlc3Rv 27421
+IGtyYQ== 27422
+w6Fp 27423
+IENhaXQ= 27424
+cm93cw== 27425
+IGNvdW50cnlzaWRl 27426
+IDE5NDU= 27427
+0LDRhtC40Y4= 27428
+INC00Lg= 27429
+IGtlcm5lbA== 27430
+bG92 27431
+IGNhbGN1bGF0aW5n 27432
+2K/Ypw== 27433
+IFdhbHQ= 27434
+IGVtcG93ZXJpbmc= 27435
+IGNoYXNzaXM= 27436
+bGluZWFy 27437
+0LPRgw== 27438
+IG5vdmE= 27439
+IHV5 27440
+IDY5 27441
+IGVuY29tcGFzcw== 27442
+dHJs 27443
+IGNvbXB1dGF0aW9uYWw= 27444
+IHdvcm1z 27445
+IG5oaeG7gXU= 27446
+IGFzdHJvbmF1dHM= 27447
+IHZlcw== 27448
+IHN5dHU= 27449
+IGRlbWFuZGVk 27450
+IGNz 27451
+IE1vbA== 27452
+IGA= 27453
+IGNoYW50 27454
+IHRoZXJlYnk= 27455
+IHBlbmlz 27456
+IGVtb2M= 27457
+d3lu 27458
+0YPQttC1 27459
+IHRyZWFk 27460
+w7NsZQ== 27461
+IGRlZXBlc3Q= 27462
+IG1hY2hl 27463
+IFZlbnQ= 27464
+IEFtc3RlcmRhbQ== 27465
+IHJlYmVs 27466
+IDYx 27467
+INCy0LrRg9GB 27468
+dWZmcw== 27469
+IGRvxJ9ydQ== 27470
+IE5hcG9sZQ== 27471
+zq7Pgw== 27472
+IHdvcmtvdXRz 27473
+IEdsYWQ= 27474
+0L3QtdGB 27475
+IHRlbnNpb25z 27476
+IFNoaWZ0 27477
+IEd1ZXI= 27478
+7YyQ 27479
+IOy5nOq1rA== 27480
+0JY= 27481
+IGltcGxhbnQ= 27482
+w6p1 27483
+6riA 27484
+IGF1dGhvcml6ZWQ= 27485
+Q0VS 27486
+IFJW 27487
+IGhpbA== 27488
+bGV2 27489
+Y2ltZW50bw== 27490
+IFVGTw== 27491
+7IOI 27492
+d29y 27493
+IGRhbmNlcw== 27494
+IFBpeGVs 27495
+IHRyb3R6ZGVt 27496
+IG9idGVu 27497
+IEFsZnJlZA== 27498
+IGNvc3RseQ== 27499
+IFN0YW5sZXk= 27500
+IHRlcnJvcmlzdHM= 27501
+IFdpZA== 27502
+heuLiOuLpA== 27503
+IGxlaWNodA== 27504
+7J207Iqk 27505
+IGRvYnJ6ZQ== 27506
+IGhlc2l0 27507
+IGVyesOkaA== 27508
+IGVpbmlnZQ== 27509
+IGhlYnQ= 27510
+0YHQtQ== 27511
+IHVucHJlZGljdA== 27512
+Q8OzbW8= 27513
+cmVtb3M= 27514
+IFRoYW5rZnVsbHk= 27515
+IHB1cnNl 27516
+Y2hz 27517
+YW5jZXI= 27518
+dWxvcw== 27519
+c3R1ZA== 27520
+IG5ldXJvbG9n 27521
+IEFuY2llbnQ= 27522
+T3V0 27523
+YXdzemU= 27524
+IG9wcG9zZQ== 27525
+IGFudGlib2RpZXM= 27526
+IFNvbWVob3c= 27527
+cm9wb2xpdGFu 27528
+a3Rvcg== 27529
+INGB0YLQvtGA0L7QvdGL 27530
+IHJvY2tldHM= 27531
+IGRpc2FibGU= 27532
+IGNhdGFzdHJvcGg= 27533
+tOye 27534
+IGN5bg== 27535
+INC00YDRg9C30YzRjw== 27536
+IGluc3RydWN0b3Jz 27537
+ZW1hYWw= 27538
+IGV0d2E= 27539
+IHl1YW4= 27540
+IEdyb3VuZA== 27541
+IHByZW1pZXJl 27542
+0YfQuNCy 27543
+IHNhaW50 27544
+eWJh 27545
+IGtvaw== 27546
+IGNvbnRyYWN0b3Jz 27547
+IOqwgQ== 27548
+INeQ15w= 27549
+IGhlYWRsaW5l 27550
+IGNvbXBsZXRhbWVudGU= 27551
+IGluZXhwZW5zaXZl 27552
+IHZpdQ== 27553
+IEdyYW5kZQ== 27554
+IGJsZWVk 27555
+66y8 27556
+IDcz 27557
+IHRvZGF2w61h 27558
+IFJ1c2g= 27559
+IEVsZGVy 27560
+6rCA64qU 27561
+IFJvdQ== 27562
+INC20LXQvdGJ 27563
+IE1pcmE= 27564
+IGRlaW5l 27565
+IGthcm1h 27566
+IHVtbQ== 27567
+IGVudHNjaGU= 27568
+IEhvbG9jYXVzdA== 27569
+IGRpc2NvdmVyaWVz 27570
+YW1lbnRz 27571
+IHJhaXNvbg== 27572
+IGJ1cmdlcnM= 27573
+QmFjaw== 27574
+IGdkeQ== 27575
+IEFH 27576
+IERhdw== 27577
+7JWg 27578
+aGVhZGVk 27579
+IENsYXI= 27580
+SW5zdA== 27581
+IExpZXV0ZW5hbnQ= 27582
+IEFmRA== 27583
+IENlcw== 27584
+IHBlcnNvbmFsaXplZA== 27585
+IGludGVyZmFjZXM= 27586
+4LiI4Liw 27587
+INGA0LXQtg== 27588
+IHN1aWM= 27589
+IHN0YXJ2aW5n 27590
+IG94aWRl 27591
+IGRlY29yYXRlZA== 27592
+IERV 27593
+IOyYiOyBmA== 27594
+IHF1bw== 27595
+IGRpc3RvcnRpb24= 27596
+IOuoueyWtOs= 27597
+IHN0YWtlcw== 27598
+IHN5bnRheA== 27599
+IGJp4bq/dA== 27600
+dGh5 27601
+aWNpZQ== 27602
+IGJyYXNpbGU= 27603
+aXNpcw== 27604
+UkM= 27605
+IHNob29r 27606
+IGRlcHRocw== 27607
+IENvc3Rh 27608
+IHZvY2Fscw== 27609
+IGNvYXN0ZXI= 27610
+IGZhbG91 27611
+ZXR0bGU= 27612
+IGtlbm5lbg== 27613
+IGRlcml2ZQ== 27614
+IGFpZHM= 27615
+INCd0LjQug== 27616
+IGVudHdpYw== 27617
+IHZlcnRpY2FsbHk= 27618
+IM0= 27619
+IFNVVg== 27620
+IGZpcmV3b3Jrcw== 27621
+IHNwZWNpZmljcw== 27622
+IGluc2lzdGVk 27623
+IGRlc2hhbGI= 27624
+IEdvbno= 27625
+bG92ZQ== 27626
+IE1pbGl0YXJ5 27627
+IFBpZXJyZQ== 27628
+IOKI 27629
+IFdob3Nl 27630
+IHBlcmZ1bWU= 27631
+IM+AzrU= 27632
+IGxvd2VyZWQ= 27633
+IGNyb3NzZXM= 27634
+IHRyYW5zbGF0ZXM= 27635
+IGFycmliYQ== 27636
+w61kbw== 27637
+IExldg== 27638
+IENpYW8= 27639
+IHNjaG9sYXJzaGlwcw== 27640
+IGdlc3R1cmVz 27641
+INGA0LXQt9GD0LvRjNGC0LDRgg== 27642
+IHF1ZXN0w6Nv 27643
+IENvbG9uZWw= 27644
+IEJvdHQ= 27645
+2LHZgQ== 27646
+TklORw== 27647
+IFdhdGNoaW5n 27648
+IFB1cnBsZQ== 27649
+0YHRgtGA0LDQvQ== 27650
+IGV4ZWN1dGl2ZXM= 27651
+IEtyaXM= 27652
+b3JuZXlz 27653
+0LXQvdC90YvQuQ== 27654
+IGNvYXRlZA== 27655
+xKk= 27656
+IHBhcmtlZA== 27657
+INGB0LLQtdGC 27658
+ISEhISE= 27659
+IEZsb3lk 27660
+xLFzxLE= 27661
+emnEhw== 27662
+IG1vdGl2YXRl 27663
+IEVsb24= 27664
+bGVhbg== 27665
+hpM= 27666
+IGlw 27667
+IG5pxbw= 27668
+IEV4cGVyaWVuY2U= 27669
+IFRpbmE= 27670
+IEtvbGxlZ2U= 27671
+IEFtYmFzc2Fkb3I= 27672
+aW55YQ== 27673
+IHRoZWZ0 27674
+IGhldXJlcw== 27675
+IE15c3Q= 27676
+IG1haXNvbg== 27677
+bGVi 27678
+IGJvd2xz 27679
+IELDvHJnZXI= 27680
+IFJvb3NldmVsdA== 27681
+UlA= 27682
+6rCA7JqU 27683
+IERlbGljaW91cw== 27684
+ZXJkaW5ncw== 27685
+IEFzc29jaWF0ZQ== 27686
+b3Vzc2U= 27687
+IENvcnQ= 27688
+IFJlcGVhdA== 27689
+IEdsb3J5 27690
+IGNvbnRhZw== 27691
+4LmA4Lil 27692
+IFBhcmFk 27693
+IEtlcnJ5 27694
+IOq/ 27695
+IFdhdmU= 27696
+IGdhdGV3YXk= 27697
+IeOAjQ== 27698
+IHRyYW5zY2VuZA== 27699
+IGRhbWFnZXM= 27700
+IHRhaWxz 27701
+IGdyYXZpdGF0aW9uYWw= 27702
+IFNoaWVsZA== 27703
+IHByaW1pdGl2ZQ== 27704
+IGNhcnJpZXJz 27705
+IEh1YXdlaQ== 27706
+2YLYrw== 27707
+IGZlbGl6 27708
+IE1pYQ== 27709
+INC/0YDRj9C80L4= 27710
+INC/0YDQvtC40YHRhdC+0LTQuNGC 27711
+IE11cnBoeQ== 27712
+IEFjdGl2 27713
+IGRpc2NvbWZvcnQ= 27714
+15HXlA== 27715
+IEtlbGw= 27716
+IENlbnR1cnk= 27717
+IHNwYWdoZXR0aQ== 27718
+IER1cmNo 27719
+IGNpZXJ0bw== 27720
+IEVtcHJlc3M= 27721
+IGd1dHM= 27722
+bmVn 27723
+INC00L7RgdGC0LDRgtC+0YfQvdC+ 27724
+IHZvbHVudGFyeQ== 27725
+IHNxdWlycmVs 27726
+IE1heg== 27727
+tOyLrA== 27728
+INCy0Lg= 27729
+INGC0LDQutC40YU= 27730
+IFNoYXJvbg== 27731
+IGVudGh1c2lhc3RpYw== 27732
+aXJlbWVudA== 27733
+IO2emOuTpA== 27734
+IHBvdHJ6ZQ== 27735
+IGluaXRpYXRlZA== 27736
+IMWbcm9k 27737
+IOydtOumhA== 27738
+IHJlbWFrZQ== 27739
+IGN1bG1pbg== 27740
+IGNvbmZ1c2U= 27741
+bWl5b3I= 27742
+dXJhcg== 27743
+Q1RPUg== 27744
+IGJ1bm55 27745
+ZWxw 27746
+IHZhbXBpcmU= 27747
+IGlsbHVtaW4= 27748
+IEhlbmQ= 27749
+INC60LDRh9C1 27750
+IFNhbHY= 27751
+INC60LDQvdCw0Ls= 27752
+IHBvcnRh 27753
+IGFzc2hvbGU= 27754
+IHN1cHBvcnRlcg== 27755
+IHNrZXB0aWNhbA== 27756
+IGtuZWFk 27757
+IOyYrA== 27758
+ZXph 27759
+IHF1w6o= 27760
+IERI 27761
+IHJvZHo= 27762
+b3duZXJz 27763
+IHBsb3Rz 27764
+IGRlbGF5cw== 27765
+IGJlbG9uZ2Vk 27766
+IGFoaA== 27767
+IGNhcnZlZA== 27768
+IHJpc2Vu 27769
+IG9yZGVu 27770
+cGhvbnk= 27771
+aXNzeQ== 27772
+ISEhISEhISE= 27773
+IG9sZHXEn3VudQ== 27774
+IHJvc2Vz 27775
+IGludHJpbnM= 27776
+IEFuZ3N0 27777
+IGZpbmFsZW1lbnQ= 27778
+7Ked 27779
+U09VTkQ= 27780
+IGluZHVs 27781
+sIw= 27782
+INeV15Q= 27783
+Y2h5 27784
+0LDQutGB0LjQvA== 27785
+IG5nZ2Fr 27786
+IGxpeg== 27787
+IGVsZWN0b3JhbA== 27788
+IFNoYXdu 27789
+cmljaWE= 27790
+IGFyc2Vu 27791
+IFBlcA== 27792
+IDIwMzA= 27793
+IHRyb3BoeQ== 27794
+IHNtb290aGVy 27795
+IGVycmU= 27796
+IGNyYXNoZXM= 27797
+IHNjaG5l 27798
+IGFzaQ== 27799
+IE1hw58= 27800
+0YPQu9C4 27801
+0YfQtdGB0LrQuA== 27802
+aWV2ZXM= 27803
+UkVBTQ== 27804
+IHN0aXJyaW5n 27805
+dXN0YQ== 27806
+IGludmVy 27807
+c2lnaHQ= 27808
+b3JkdQ== 27809
+b29y 27810
+IMSDbg== 27811
+IHBlcm1pdHRlZA== 27812
+0YDRjA== 27813
+IGNoYWxr 27814
+IHRhdHRvb3M= 27815
+IFJlbGF0aW9ucw== 27816
+IEhveQ== 27817
+a3NhbQ== 27818
+IGRlbnRpc3Q= 27819
+IOuvuOq1rQ== 27820
+IHNvZmE= 27821
+INGU 27822
+IGZvcm1l 27823
+2YLYqQ== 27824
+IOuyoA== 27825
+IGVtYnJhY2Vk 27826
+bWls 27827
+IHN1bmdsYXNzZXM= 27828
+IOqwlA== 27829
+IHNlYW1sZXNz 27830
+IGJlZXA= 27831
+w6RjaHN0 27832
+IHN3ZWV0cw== 27833
+IHNlbWFpbmU= 27834
+IGlycmVsZXZhbnQ= 27835
+IGRlc2Vudm9s 27836
+z4HPiQ== 27837
+INC/0YDQvtC40LfQstC+0LQ= 27838
+YW5ncw== 27839
+IGFyb21h 27840
+IHBvb2xz 27841
+IGdp4bud 27842
+IFVn 27843
+IGNsaW1iZWQ= 27844
+IHRyZW5kaW5n 27845
+IHNlcGVydGk= 27846
+IEJhcnI= 27847
+IHDFgg== 27848
+IE9yaWdpbmFsbHk= 27849
+INqv 27850
+dXR0bw== 27851
+irjr 27852
+INC60L7RgtC+0YDRi9GF 27853
+INC30LDRhQ== 27854
+IGVpZ2VuZW4= 27855
+IG11cmRlcmVy 27856
+ZXJuYW1l 27857
+xZ4= 27858
+IGFubm91bmNpbmc= 27859
+IFBsYXRmb3Jt 27860
+IGV4cGxhbmF0aW9ucw== 27861
+IHByZXNlbnRl 27862
+IE5hc8SxbA== 27863
+IG9ycGhhbg== 27864
+IEZvcnRuaXRl 27865
+cm9zcGVjdA== 27866
+ZXJlZGl0aA== 27867
+IOyXhuyWtA== 27868
+IE5JSA== 27869
+d2FnZW4= 27870
+IHJlbWVk 27871
+p4Dr 27872
+bW9udA== 27873
+IEplZmZyZXk= 27874
+cHJvbQ== 27875
+IGbDvG5m 27876
+INC90LDQt9Cw0LQ= 27877
+IGN1Y3VtYmVy 27878
+IFN1bW1pdA== 27879
+p6Q= 27880
+0J3QkNCv 27881
+IEpldA== 27882
+IGNhbWJpbw== 27883
+0YPQudGC0LU= 27884
+IGN1Ymlj 27885
+IGRpc3Byb3BvcnRpb24= 27886
+ZXJleg== 27887
+IG1hZG5lc3M= 27888
+IHRpbnQ= 27889
+IGZ1ZXJvbg== 27890
+IGt5 27891
+IGJpcGFydA== 27892
+U2Ft 27893
+IOu9 27894
+IHJpdg== 27895
+IFRhbms= 27896
+IOuGkw== 27897
+IHJlbmRlcmVk 27898
+xZtsxJk= 27899
+Y29uZHM= 27900
+IGRpc3J1cHRpb24= 27901
+IGluY29udmVu 27902
+IHF1aXNlcg== 27903
+IGRlbmlhbA== 27904
+IGdhbGF4aWVz 27905
+IHNvdmVyZWlnbg== 27906
+IHBvbHNr 27907
+z4HPjg== 27908
+IG1leA== 27909
+IGNhcmFjdGVy 27910
+IExlZ28= 27911
+YW5kZW4= 27912
+Lici 27913
+IO2UjOs= 27914
+IGNvbXByZXNzb3I= 27915
+IE1vdmll 27916
+IGFwcGxpY2FudHM= 27917
+emllaGVu 27918
+IHZlZ2V0YXRpb24= 27919
+IGJlbGxl 27920
+IEdPT0Q= 27921
+IEJhdQ== 27922
+IHJlc2VudA== 27923
+c2V4 27924
+YW1lbnRvcw== 27925
+INeU15bXlA== 27926
+IG92ZXJsb2Fk 27927
+IHNpbGljb25l 27928
+0LXRgdGC0L3Qvg== 27929
+IGRlbmtlbg== 27930
+IGRlZmluaXQ= 27931
+IFdhc24= 27932
+IGFsdGVyZWQ= 27933
+IFNvbw== 27934
+IFdpbmc= 27935
+aW5kcmU= 27936
+IE5QQw== 27937
+z4HOrQ== 27938
+IFR3ZW50eQ== 27939
+IExpZWJl 27940
+IGhvbWVsZXNzbmVzcw== 27941
+b3VsZGVy 27942
+INCY0YLQsNC6 27943
+0YHQutCw0Y8= 27944
+IGN1YXRybw== 27945
+IEhhcnZleQ== 27946
+IHBoaWxhbg== 27947
+IEJlZXQ= 27948
+IHBvbGljaW5n 27949
+IEFsZXhhbmQ= 27950
+INC80L7Qu9C+0LQ= 27951
+IG3DvHM= 27952
+IGhpem8= 27953
+67O064uk 27954
+INC/0L7Qt9Cy0L7Quw== 27955
+INC/0YvRgg== 27956
+0L7Rh9C10LzRgw== 27957
+IO2DnA== 27958
+IGNyeXB0b2N1cnJlbmN5 27959
+IGxvcm8= 27960
+IHN1bW1hdGlvbg== 27961
+IGJha2FsxLFt 27962
+IG5ldXJvcw== 27963
+2KU= 27964
+INC80L7QttC10Lw= 27965
+IMO8c3Q= 27966
+IHByZWxpbWluYXJ5 27967
+IGhvcm5z 27968
+IFRJ 27969
+2YPZhA== 27970
+WU8= 27971
+IGhpbmdl 27972
+IHJlcGFpcnM= 27973
+IGJvbmRpbmc= 27974
+IGJpemU= 27975
+INGI0YI= 27976
+IG1vdGl2ZQ== 27977
+IE5pZ2VyaWE= 27978
+MTIw 27979
+YmxvY2s= 27980
+IGF2aWF0aW9u 27981
+IEtvbW11bg== 27982
+INC+0LrQsNC3 27983
+IHRlbmhh 27984
+IGVkdWNhdGluZw== 27985
+IHN0YWF0 27986
+INGB0LrQvtC70YzQutC+ 27987
+IGZyaWdodGVuZWQ= 27988
+IHNlZWtz 27989
+0YDRg9GI 27990
+cXVlbnQ= 27991
+IE5vdQ== 27992
+IHByYXQ= 27993
+IFNob3Q= 27994
+V29yaw== 27995
+a2FyYW5n 27996
+IExpZ2h0bmluZw== 27997
+bm9sZHM= 27998
+cm9sbGVk 27999
+Z2xhc3M= 28000
+IGNyZWRpYmlsaXR5 28001
+SVRZ 28002
+IGF0bW9zcGhlcmlj 28003
+IGhhdmlh 28004
+w6RuZGVybg== 28005
+Y2hlZXJz 28006
+VGhlc2U= 28007
+IENlbGw= 28008
+IG1hZ25lcw== 28009
+IEJyYXZv 28010
+c2Vhc29u 28011
+IMWfZXlsZXI= 28012
+8J+O 28013
+d2hpdGU= 28014
+IE1C 28015
+IHN0YWNrZWQ= 28016
+IDc0 28017
+INC00LDQstCw0Lk= 28018
+IHBhdmU= 28019
+INC+0YU= 28020
+IGRhdGFzZXQ= 28021
+IHJldG91cg== 28022
+IG1hdHVyaXR5 28023
+IHF1YXNl 28024
+IDkz 28025
+IFN5bQ== 28026
+IGJyaWVmaW5n 28027
+IGN1bHR1cmFsbHk= 28028
+IOy3qA== 28029
+aW5oYXM= 28030
+IG1hZGFt 28031
+IGFqdWRhcg== 28032
+IFRpYmV0 28033
+IGxlYWtz 28034
+Y2lsZQ== 28035
+IHRoZWF0ZXJz 28036
+7Jio 28037
+NzI= 28038
+IFdhc2g= 28039
+IFF1YWxpdHk= 28040
+IEl2YW4= 28041
+IEJlbnQ= 28042
+aWdhdG9y 28043
+IEdlc2NoaWNodGU= 28044
+IHJlYWN0aXZl 28045
+IDE5MDA= 28046
+IGNvbnRyYWRpY3Q= 28047
+IHppZW1saWNo 28048
+IGNvaG9ydA== 28049
+4bun 28050
+IHBlc3RpYw== 28051
+IG9yYXo= 28052
+IHRlbGxlbWVudA== 28053
+6b4= 28054
+IE5vd2FkYXlz 28055
+Y3Jldw== 28056
+U3RldmU= 28057
+IGZpY3Rpb25hbA== 28058
+IGlsaw== 28059
+IGdhc29saW5l 28060
+emFt 28061
+IHBhbmNha2U= 28062
+w6huY2lh 28063
+IG11aXRvcw== 28064
+IGJ1cnk= 28065
+IGtvcA== 28066
+IElR 28067
+IHJlc2VydmF0aW9u 28068
+IFVwZGF0ZQ== 28069
+IGplag== 28070
+IEV5ZXM= 28071
+IHZpdmU= 28072
+IGNoY2U= 28073
+IEluaQ== 28074
+cmVzcG9ucw== 28075
+IHJlZmxlY3RpdmU= 28076
+IFdhbg== 28077
+0ZbQtw== 28078
+IGVuY2E= 28079
+IGVtYm9k 28080
+IEJ1cmdlcg== 28081
+IGFjYWRlbWlh 28082
+IENpcmM= 28083
+INC/0YDQtdC6 28084
+IGFubGFt 28085
+IHBoaWxhbnRocm9w 28086
+IEJhxZ8= 28087
+IEF1ZGk= 28088
+IHZvc3Q= 28089
+IHJlcGVy 28090
+UGV0ZXI= 28091
+IGNvbnNvbGVz 28092
+IHNjcnV0 28093
+IFR1cm5lcg== 28094
+INCx0YvQsg== 28095
+SUlJ 28096
+IEZsaWdodA== 28097
+4LiW 28098
+IFJhdmVu 28099
+IGNvcnJvcw== 28100
+ZmVybg== 28101
+IHByb3Zh 28102
+IFNldg== 28103
+IHJlY2lwcm8= 28104
+IDE5ODU= 28105
+IG51ZXZh 28106
+IGRhYg== 28107
+44CB44CM 28108
+IG1leg== 28109
+IFN0YXJr 28110
+cHBpbmdz 28111
+0L7RgdGC0Lg= 28112
+7Kad 28113
+IGZyYW1pbmc= 28114
+INCg0LDQtw== 28115
+IHBvc3Rw 28116
+IFNoYW5ub24= 28117
+INC60YPRgA== 28118
+IGpha2J5 28119
+aWVubmVudA== 28120
+IE1hcHM= 28121
+IFJldmVsYXRpb24= 28122
+INGB0YLQsNC7 28123
+7Jq0642w 28124
+IGRldmFudA== 28125
+IEdpdmluZw== 28126
+IFdBUw== 28127
+INC60L7Qs9C+ 28128
+IHJlbWE= 28129
+IFJD 28130
+bsOt 28131
+IHNsaXBwZWQ= 28132
+IFJhbXM= 28133
+IHdlZXQ= 28134
+IG1hc2N1bGluZQ== 28135
+IEVj 28136
+IHJlb3A= 28137
+IFBsYW50 28138
+IE1BWQ== 28139
+IHNwaWtlcw== 28140
+IG5venpsZQ== 28141
+IFdpa2lwZWRpYQ== 28142
+IENvaA== 28143
+SVNTQQ== 28144
+Y2hsb3NzZW4= 28145
+7KeA66W8 28146
+IOuvuOs= 28147
+IE5lZGVy 28148
+Sm9zaA== 28149
+INCg0L7RgdGB0LjQuA== 28150
+IDE5ODc= 28151
+IFRoZW9yeQ== 28152
+ZWtr 28153
+IHV0YW4= 28154
+INC00L7QvNCw 28155
+Y2h1 28156
+INGB0LE= 28157
+IGFwcm92ZQ== 28158
+VkVO 28159
+dWVwcmludA== 28160
+IDg0 28161
+Q29y 28162
+IHJpY2hlcg== 28163
+IHNhbmR3aWNoZXM= 28164
+YXRzdQ== 28165
+0YjQuNGF 28166
+IGxhdHQ= 28167
+fn5+fg== 28168
+ZnJpZW5kcw== 28169
+IGRlcm5pw6hyZQ== 28170
+IHN0ZXJlbw== 28171
+INGN0LrRgdC/ 28172
+IHByb3RlY3Rpb25z 28173
+IGhhdXQ= 28174
+RXZlcnlvbmU= 28175
+IGVudGVycHJpc2Vz 28176
+IE1vc3RseQ== 28177
+IFNwb3RpZnk= 28178
+IFNleA== 28179
+IHVuZw== 28180
+jOulvA== 28181
+IGFjdGl2aXNt 28182
+Y3RpY2E= 28183
+b3JpZ2luYWw= 28184
+INC/0YDQvtCz0YDQsNC8 28185
+IGJyb2Njb2xp 28186
+4KY= 28187
+0L7Qs9GA0LDRhA== 28188
+IHNla2FyYW5n 28189
+IGNyYWZ0aW5n 28190
+INCx0LDQvQ== 28191
+IFJheg== 28192
+IG5haXZl 28193
+IHNjcm9sbGluZw== 28194
+IG51bWVyaWNhbA== 28195
+IHNjaGVkdWxpbmc= 28196
+IGFwYXJ0bWVudHM= 28197
+540= 28198
+IHN0cmV0Y2hlcw== 28199
+YWNleQ== 28200
+IEhFUg== 28201
+IHppbmM= 28202
+IGRhcm4= 28203
+IGPDqWw= 28204
+IHdhcmRyb2Jl 28205
+IHJlZGlyZWN0 28206
+IGp1bQ== 28207
+IFN0cmFuZ2U= 28208
+IG7DoG8= 28209
+IGV4cGVyaW1lbnRpbmc= 28210
+w6lyw6k= 28211
+IHZvdWxleg== 28212
+IGdlYmU= 28213
+IEthbm4= 28214
+IMSR4buZ 28215
+IE1heGlt 28216
+IEvDtm4= 28217
+IEdsYXM= 28218
+IHBvbGlzaGVk 28219
+IG51bWE= 28220
+SWNo 28221
+IHJpdHVhbHM= 28222
+IFNJ 28223
+0LjRgtC10LvQuA== 28224
+IGluZmlsdA== 28225
+IHNjYXJm 28226
+b3BoeQ== 28227
+IHlpbmU= 28228
+IGNpdmlj 28229
+IE1lbmc= 28230
+w6RuZ2U= 28231
+1aU= 28232
+aGlzdG9pcmU= 28233
+IE9rZQ== 28234
+IOyYhg== 28235
+IHNvbGx0ZW4= 28236
+IDgy 28237
+IHByZXNjcmliZWQ= 28238
+IER1YmFp 28239
+IEVsdGVybg== 28240
+IG5hdGlvbndpZGU= 28241
+IHNrYXRpbmc= 28242
+aWFyeQ== 28243
+IHJld2FyZGVk 28244
+IG1vcmFsaXR5 28245
+IE1hZ2dpZQ== 28246
+IE9oaGg= 28247
+IEZhaHJlbg== 28248
+b2x2ZWQ= 28249
+IGRldXhpw6htZQ== 28250
+dGVjaG4= 28251
+cm9sZQ== 28252
+IGxlaWRlcg== 28253
+IEpBWQ== 28254
+INC40L3RhNC+0YDQvA== 28255
+IGNhZmZl 28256
+cmVpY2hlbg== 28257
+IGthcnQ= 28258
+IEN1dGU= 28259
+ZmZlY3RpdmU= 28260
+IGJ1bGx5 28261
+YWdhcg== 28262
+IGNvbW1vZGl0eQ== 28263
+IG9icmln 28264
+T1VS 28265
+IHVucGxlYXNhbnQ= 28266
+bm94 28267
+SnVs 28268
+b2xpdGg= 28269
+0YLQvtGP0Yk= 28270
+IEJlbGxh 28271
+IGRvbGxz 28272
+IEhvZmY= 28273
+IGFkdmlzb3Jz 28274
+IHRyYW5zZmVycw== 28275
+IEdva3U= 28276
+IDEyMDA= 28277
+aW5ob3M= 28278
+UGFs 28279
+IOuYkQ== 28280
+IHJlcHQ= 28281
+IGFjY29tcGxpc2htZW50 28282
+IHdlYXZl 28283
+IG92ZXJzaWdodA== 28284
+IHVuaGVhbHRoeQ== 28285
+IGZpbHQ= 28286
+IHB1ZGRpbmc= 28287
+IE1pZ3VlbA== 28288
+IGNodWNrbGVz 28289
+dmVyc2lvbg== 28290
+IGNvbmZlc3Npb24= 28291
+dmFsdWU= 28292
+IHRyaXVtcGg= 28293
+IHNhaXI= 28294
+IOuFuA== 28295
+IGFydGU= 28296
+IE1hdGVyaWFs 28297
+dXRp 28298
+IGxpcXVvcg== 28299
+IEJheWVybg== 28300
+IE1haWw= 28301
+IO2WpQ== 28302
+0YHQutC+0Lw= 28303
+IGNoZWFwZXN0 28304
+INGH0LDRgdGC0Lg= 28305
+IEpvYnM= 28306
+IENhbnlvbg== 28307
+aGFybWE= 28308
+YWxleQ== 28309
+YW5kcm8= 28310
+IGFwcGVhcmFuY2Vz 28311
+cHJvZg== 28312
+INC+0Lc= 28313
+bGFnZW4= 28314
+IC8v 28315
+INC70LjRiNGM 28316
+IHJlY292ZXJpbmc= 28317
+0LTQtg== 28318
+cHN5 28319
+IHN3aWZ0 28320
+IFNwaW4= 28321
+IHNlaW5lbQ== 28322
+IGRvbHBo 28323
+ZsO8aHI= 28324
+w6J0 28325
+IGFsdGlqZA== 28326
+IE1hcnR5 28327
+IEhvY2g= 28328
+IHByZWRhdG9ycw== 28329
+IHZvcmhlcg== 28330
+INCU0LDQstCw0Lk= 28331
+IGZyYWdtZW50cw== 28332
+IHBhc3RyeQ== 28333
+IGNvbW1lbg== 28334
+IFNhbmE= 28335
+IOqxtOuNsA== 28336
+dXNzZW4= 28337
+IHRlbGE= 28338
+IE5pbmE= 28339
+bGVr 28340
+IGNyaWVz 28341
+IHRoaWdocw== 28342
+IEZsZXg= 28343
+IEJ1eno= 28344
+44Q= 28345
+VXM= 28346
+IHBhc28= 28347
+IGRlY2xpbmVk 28348
+IE55 28349
+YmFsYW5jZQ== 28350
+IG1hc2E= 28351
+IGpvcw== 28352
+INCh0L/QsNGB0LjQsdC+ 28353
+YWNodQ== 28354
+bG91ZA== 28355
+IHBlbmE= 28356
+IFdhbGQ= 28357
+IGVsaW1pbmF0aW9u 28358
+INCy0LXRgdGM 28359
+b3JhZ2U= 28360
+IG1pc3VuZGVyc3RhbmRpbmc= 28361
+IGVuZG9yc2U= 28362
+IG9nw7NsZQ== 28363
+IGdyZWVk 28364
+IGtsZWlu 28365
+15zXlA== 28366
+UkVZ 28367
+IEVhdGluZw== 28368
+IHNlbWluYXI= 28369
+IEJpcnRoZGF5 28370
+IHF1ZWxsZQ== 28371
+IE11bHRp 28372
+IHRpcmFy 28373
+IHBlcmNo 28374
+IGxhdm9y 28375
+IEppYQ== 28376
+IG11dGF0aW9ucw== 28377
+IGNpZ2FyZXR0ZXM= 28378
+2YjYrA== 28379
+IGNvdXNpbnM= 28380
+IGNhcHN1bGU= 28381
+IGhvcnJpZmlj 28382
+IHN0dXI= 28383
+IHplaWd0 28384
+bnV0cw== 28385
+IG1lYW53aGlsZQ== 28386
+IENvbGlu 28387
+IGdvYmllcm5v 28388
+IGd3 28389
+IHVoaA== 28390
+IEpFUg== 28391
+c3BlY2lmaWM= 28392
+IGFsbGVnYXRpb25z 28393
+IOupiw== 28394
+IEVsbGE= 28395
+b29rZWQ= 28396
+IEZpdA== 28397
+YWZmbGU= 28398
+IEFwcsOocw== 28399
+IER1Y2s= 28400
+IGNlbGx1bGFy 28401
+Y8Ozdw== 28402
+INGH0YPQstGB0YLQsg== 28403
+Z2Vub21tZW4= 28404
+7Iqk7Yq4 28405
+IGxhaW4= 28406
+aXNvbA== 28407
+IGhvbGRlcnM= 28408
+IGJvb3N0ZXI= 28409
+IFNhc2hh 28410
+0YvQstCw0LXRgg== 28411
+gbw= 28412
+IHNlcGFyYXRpbmc= 28413
+IHJlaW5mb3JjZW1lbnQ= 28414
+INC+0LTQvdC+0Lk= 28415
+7JeG 28416
+SURF 28417
+IE9wdGlvbg== 28418
+cGhvbg== 28419
+IHBsYWlz 28420
+IENhbWI= 28421
+IO2ZmA== 28422
+IHVuY29tbW9u 28423
+Ijo= 28424
+bWl5b3J1bQ== 28425
+bW9p 28426
+YWNqZQ== 28427
+0LDQttGD 28428
+1bY= 28429
+IGdlbXM= 28430
+w7xsZXI= 28431
+b29scw== 28432
+IGVuenltZXM= 28433
+IGtpZG5hcHBlZA== 28434
+IGtldGNodXA= 28435
+dGFsaw== 28436
+IHphY2g= 28437
+IHdhc2hlcg== 28438
+44CC44CC 28439
+IEFyY2hpdGVjdA== 28440
+dmVudWU= 28441
+IFBsYW5uaW5n 28442
+IFNhdmlvcg== 28443
+INCz0YDRg9C/0L8= 28444
+7Yq8 28445
+YXJ5YQ== 28446
+IHByb2Nlc28= 28447
+IGxpbWJz 28448
+IHJlYWxpemVz 28449
+aWFuZGVy 28450
+RlM= 28451
+YWpp 28452
+IHVuaXRl 28453
+IOydmOs= 28454
+IHBvc3PDrXZlbA== 28455
+cmFpdHM= 28456
+IEFncmU= 28457
+24zaqQ== 28458
+7ISc64+E 28459
+INCy0LXQuw== 28460
+INC80LXRgdGP 28461
+YW5vcg== 28462
+UGF0 28463
+IGRlcm5pZXI= 28464
+z4PPhM61 28465
+INC60LDQutCw0Y8= 28466
+IGzDpHNzdA== 28467
+IE1laA== 28468
+IG5naA== 28469
+IGFtYXRldXI= 28470
+RmU= 28471
+IOq2gQ== 28472
+IHNpdHVhY2nDs24= 28473
+IHNlZGFu 28474
+IGNsZWFuc2luZw== 28475
+bGFzdGluZw== 28476
+IGNvbW11bmlzdA== 28477
+QU5F 28478
+IGlycmVndWxhcg== 28479
+IHNvdXQ= 28480
+IENhcm5leQ== 28481
+IGFsbGVtYWFs 28482
+IG11Y2jDrXM= 28483
+IGxpYnJv 28484
+0K3RgtC+ 28485
+INCw0L8= 28486
+IGNvbnRpbnVhdGlvbg== 28487
+IExvcg== 28488
+PyIs 28489
+cXVpbg== 28490
+IGNoYXJhY3Rlcml6ZWQ= 28491
+YWplcw== 28492
+IHNpZ2h0cw== 28493
+INGP0LfRiw== 28494
+IFVoaA== 28495
+YmlydGg= 28496
+ZG9uZw== 28497
+IGhhYmxhbmRv 28498
+IHN5bXB0b20= 28499
+IGNhcGFjaXRvcg== 28500
+IHRyYW5zcG9ydGVk 28501
+IGlnbm9yYW50 28502
+INC90LjQutC+0LPQtNCw 28503
+IGRyaXA= 28504
+IEV2YQ== 28505
+IGFkamVjdA== 28506
+IG1hc3NpdmVseQ== 28507
+IEV0aGk= 28508
+IENpcmNsZQ== 28509
+IHJhaW5mYWxs 28510
+IE1vdXNl 28511
+IHJlZnVuZA== 28512
+IFp3 28513
+YXNzZW1i 28514
+IDIyMA== 28515
+IE9yZA== 28516
+IHZlaW5z 28517
+IEdpYW50 28518
+IG3Do2U= 28519
+IHZhcA== 28520
+IG1pc3Nlcw== 28521
+zr/Phc+C 28522
+TW8= 28523
+IEVudHdpY2s= 28524
+SU5U 28525
+2YbYqg== 28526
+IHRoZW9yZXRpY2FsbHk= 28527
+IHRlYXJpbmc= 28528
+IHRyb3VibGVk 28529
+cHJlbQ== 28530
+IHJlcGV0aXRpdmU= 28531
+IOKW 28532
+IGhlYXZlbmx5 28533
+IEFtYmVy 28534
+INC/0L7Qu9C+0LY= 28535
+IO2VtOyk 28536
+IHZvd2Vs 28537
+YW5raW5n 28538
+IFdpcnRzY2hhZnQ= 28539
+IGlycg== 28540
+IGNvenk= 28541
+IHVuZmFtaWxpYXI= 28542
+IFBvcnM= 28543
+IOunnuyVhA== 28544
+IFRpbW90aHk= 28545
+0YHQvtC70Y7Rgg== 28546
+cGV4 28547
+IFZJUw== 28548
+KSg= 28549
+IHN1cGVyc3Q= 28550
+IGltcHJvdg== 28551
+IEJlbmc= 28552
+IGRpc2Nvbm5lY3RlZA== 28553
+IGFwdA== 28554
+0YDQtdC9 28555
+IEV4dHJh 28556
+INCx0LXQuw== 28557
+c2hvcA== 28558
+ZGluZ3M= 28559
+IENvbm5lY3RpY3V0 28560
+7LCs 28561
+IEdD 28562
+YmVo 28563
+SmVyZW15 28564
+IEJhdHQ= 28565
+YXRoYQ== 28566
+IFp1c2FtbWVu 28567
+c2NyZWFtcw== 28568
+IGdyYXM= 28569
+YWZmdA== 28570
+IEluaXRpYWxseQ== 28571
+IEJyZXR0 28572
+IHNwZWNpZmljYXRpb25z 28573
+IHNlYXdlZWQ= 28574
+IG9hdGg= 28575
+IGZvdW50YWlu 28576
+INC60L7RgtC+0YDQvtC5 28577
+IFN0ZWlu 28578
+IENvcmludGg= 28579
+IGNvbmp1Zw== 28580
+IGNvbXBlbnNhdGU= 28581
+IOuKkOuCjOydtA== 28582
+IG9uemU= 28583
+IHNraW5jYXJl 28584
+QnJpYW4= 28585
+IHNlcnZpcg== 28586
+fX0= 28587
+IFZpaw== 28588
+IHVuaW50 28589
+IHN1cHBsaWVycw== 28590
+IGJhbGNvbnk= 28591
+IGVuZXJnaWE= 28592
+b21ldHJpYw== 28593
+0LfRjw== 28594
+IHNpZ2g= 28595
+IFRPTQ== 28596
+IFB1cmU= 28597
+eXR0 28598
+0YvRgQ== 28599
+IFJhaW5ib3c= 28600
+IFBpdHRz 28601
+15nXng== 28602
+IHN0YXR1ZXM= 28603
+aGVhZHM= 28604
+IGNvdXBsZWQ= 28605
+IGhlcmQ= 28606
+IGV4Y2x1ZGVk 28607
+IGdpbHQ= 28608
+INGO 28609
+IHN3b2pl 28610
+IFN2ZXI= 28611
+NjM= 28612
+aXNzYW50 28613
+IGTDvHJmZW4= 28614
+oIjr 28615
+IGtpc3Npbmc= 28616
+b29m 28617
+IGN1cnNlZA== 28618
+IHNob3dlcnM= 28619
+IHN3aW5naW5n 28620
+IHJlcHJvZHVjZQ== 28621
+IHPDpHR0 28622
+ZWxjb21l 28623
+IGZ1bmRhbWVudGFscw== 28624
+IGFsbW9uZA== 28625
+IHDDqQ== 28626
+IHdlbGxiZWluZw== 28627
+IGh1bnRlcnM= 28628
+U2Vj 28629
+k5zrprQ= 28630
+IGVtaXNzaW9u 28631
+IHBzeWNob2xvZ2lzdA== 28632
+IGJldHJheWVk 28633
+IFJleW5vbGRz 28634
+TEVT 28635
+IHBvbGxpbmc= 28636
+IG5lZ2F0aXZlbHk= 28637
+IGNvbWJpbmVz 28638
+15zXkA== 28639
+0LDRgNCw 28640
+zrvOu86s 28641
+IFR1cm5z 28642
+T1RU 28643
+INeU15k= 28644
+YWlzb24= 28645
+IGFpcmxpbmU= 28646
+IHJlc3RyaWN0aW9u 28647
+d2Fs 28648
+IGF1cmFpdA== 28649
+IExlYmFub24= 28650
+IE1PUg== 28651
+IG1vbmtleXM= 28652
+w6luZXI= 28653
+0ZbRlw== 28654
+IG1vdGhlcmY= 28655
+INmH2LDZhw== 28656
+IGZldQ== 28657
+w7xocmVu 28658
+IGh5Z2llbmU= 28659
+ZW50ZWVu 28660
+RGVz 28661
+IGRpc3NpcA== 28662
+RXN0 28663
+IHNhaW50cw== 28664
+IHBvdGFzc2l1bQ== 28665
+IHJlY2tvbg== 28666
+Q2xpbnR1cw== 28667
+IG1hbmlmZXN0YXRpb24= 28668
+IEFwcHJv 28669
+IEluc3BlY3Q= 28670
+IHZlbnRpbGF0aW9u 28671
+IGhlbG0= 28672
+IGthcmE= 28673
+4Liy4LiZ 28674
+IGZhdm9yYWJsZQ== 28675
+IOyViuyVmA== 28676
+IEhpc3Bhbmlj 28677
+4Lic 28678
+INeU15s= 28679
+IHZhbGlkYXRl 28680
+IFJlc2lkZW50 28681
+IGNvbWVueg== 28682
+YmVpdGVy 28683
+ZXJlcg== 28684
+IGRhZG8= 28685
+YXRjaGluZw== 28686
+bWV0cm9z 28687
+IEhpbg== 28688
+IER1bQ== 28689
+IGhhesSxcg== 28690
+IE5hdGFsaWU= 28691
+IGVuY3J5cHRpb24= 28692
+0L7Rh9C60LA= 28693
+bW1h 28694
+aG91c2Vz 28695
+IGFuYWx5dGljYWw= 28696
+IERhbmc= 28697
+Zmlyc3Q= 28698
+IEVuYw== 28699
+Y2FuZG8= 28700
+IGx1ZHpp 28701
+d2FydA== 28702
+IHN0YXRpc3RpYw== 28703
+IOyCsA== 28704
+IGNvbW1lbnRpbmc= 28705
+IGNvb3JkaW5hdGVk 28706
+IEh5cGVy 28707
+5Zo= 28708
+IEJlcnQ= 28709
+IEhpcA== 28710
+a2Vt 28711
+w7xuw7w= 28712
+IHphbA== 28713
+IO2VmOuKlOuNsA== 28714
+IFJvYm90 28715
+cmF3bg== 28716
+IHJoZXRvcmlj 28717
+dWxsYWg= 28718
+IERpZXQ= 28719
+IHRha2ljaA== 28720
+IHBvc3Nlc3NlZA== 28721
+k5zripQ= 28722
+IHdha2Vz 28723
+IFJhZg== 28724
+TWFydA== 28725
+IGVjYw== 28726
+IEZN 28727
+IGRpZmlj 28728
+IEFsbGV6 28729
+IGN1cmVk 28730
+IFF1YWQ= 28731
+IGJlbGU= 28732
+IGpvdXJuYWxz 28733
+IHRhZA== 28734
+IHNvY2lhbGVz 28735
+IHdoYXRz 28736
+IEJhc3M= 28737
+IGplc3RlbQ== 28738
+IFNhZGx5 28739
+IFNvdXJjZQ== 28740
+IMO8w6c= 28741
+YWx0dW5n 28742
+aWVydGVu 28743
+IGp1bGxpZQ== 28744
+aWZh 28745
+INCa0L7RgA== 28746
+IERvb3I= 28747
+INCd0LDQtA== 28748
+INC30LTQvtGA0L7Qsg== 28749
+IHJ1bW9y 28750
+IHBpZXM= 28751
+INC/0LXRgNC1 28752
+INC+0YLQsg== 28753
+0LXQvdC90YvQtQ== 28754
+SG9zdA== 28755
+IFNvcGhpZQ== 28756
+YW50ZW4= 28757
+QW55 28758
+IEF1Zmc= 28759
+IEhEUg== 28760
+IFJvY2tldA== 28761
+cmVzc28= 28762
+IHZlcmRl 28763
+IHByw6lzaWRlbnQ= 28764
+IGluZG9vcnM= 28765
+IHN0YWdnZXI= 28766
+IHN0YXRv 28767
+IERpYWw= 28768
+IGJ1enppbmc= 28769
+ZW1lcg== 28770
+INCS0YHRkQ== 28771
+INC00LXRgNC10LI= 28772
+IHBvdXY= 28773
+IHN0cmFuZHM= 28774
+IOqyg+ydtA== 28775
+IFBhcmw= 28776
+0L7QutC+0Lk= 28777
+IHNpcA== 28778
+ICgq 28779
+w6RuZ3Q= 28780
+IGRlYmVy 28781
+IEFpbg== 28782
+IGRyYXN0aWNhbGx5 28783
+IFNsb3dseQ== 28784
+IEJyaWc= 28785
+IFRvcmFo 28786
+IGFjaGU= 28787
+ID8/Pw== 28788
+IERvYg== 28789
+a2FubnQ= 28790
+TWFyeQ== 28791
+IHN0YW0= 28792
+IERlbW9u 28793
+cGxh 28794
+IEZyZXVuZA== 28795
+IEJlbm4= 28796
+IGhpZ2hz 28797
+INqp2LE= 28798
+IFByZXBhcmU= 28799
+IHByb3h5 28800
+IGNhbXBv 28801
+IEF1Z2Vu 28802
+o6jr 28803
+IENobG9l 28804
+aWN1bGFybHk= 28805
+eW91bmc= 28806
+qZTr 28807
+IHNjcmF0Y2hpbmc= 28808
+IGdsYWM= 28809
+IGdlbWVpbnNhbQ== 28810
+YW5hbA== 28811
+YWNha3PEsW4= 28812
+IEZvcnVt 28813
+ZW5uaWFs 28814
+IFJlc291cmNlcw== 28815
+IG1laXN0ZW4= 28816
+IEZlbGw= 28817
+IHVuYW5pbQ== 28818
+IFRC 28819
+IFNlbGJzdA== 28820
+5oY= 28821
+IGludGltaWRhdGluZw== 28822
+IEdlZsO8aGw= 28823
+IOy9lOuhnA== 28824
+aWRvcg== 28825
+aWNpb25lcw== 28826
+YXJzYQ== 28827
+XS4u 28828
+YXpv 28829
+IGtlbmRp 28830
+IFRhZ2U= 28831
+dGVybWlu 28832
+IFByb3plbnQ= 28833
+TWF5YmU= 28834
+bMOp 28835
+IHF1ZXN0aQ== 28836
+IG1lbWVz 28837
+IGNvcnJl 28838
+IFZJUA== 28839
+IEdhbGxlcnk= 28840
+IHVyZ2VuY3k= 28841
+IG5vY2hl 28842
+IGtpbmRseQ== 28843
+IE1lcmVkaXRo 28844
+IHbhuq15 28845
+INin2YTYqA== 28846
+IEVzdGFkbw== 28847
+enVn 28848
+b3F1ZQ== 28849
+IG9iZXNpdHk= 28850
+T2Zm 28851
+IEV1cm9wZWFucw== 28852
+w7Zk 28853
+7Lm06w== 28854
+IGhvb3A= 28855
+IGVuam95cw== 28856
+IENoaXA= 28857
+cGF0aWVudA== 28858
+IG1pY3Jvc2NvcGU= 28859
+IGxlZ2l0aW0= 28860
+INGP0LLQu9GP0LXRgtGB0Y8= 28861
+z4POuQ== 28862
+YXJnZW50 28863
+IHNoYW0= 28864
+IGxpY2Vuc2luZw== 28865
+b2xpYQ== 28866
+U29ycnk= 28867
+cmFtYQ== 28868
+IGFjY2VsZXJhdGVk 28869
+IHd5bQ== 28870
+IGZhaXJuZXNz 28871
+IFJlYWRpbmc= 28872
+IHNsYWNr 28873
+IERvaw== 28874
+emnEmWt1asSZ 28875
+IHJ1YmJpbmc= 28876
+0LDRgtGD 28877
+IGFsbG9jYXRlZA== 28878
+anVuZw== 28879
+IHBhaW5z 28880
+IHdpbmRpbmc= 28881
+IGdlbGl5b3I= 28882
+IENV 28883
+bW90 28884
+Y29jaw== 28885
+IFBvc2l0aW9u 28886
+YnJvcw== 28887
+IGxpdmVzdHJlYW0= 28888
+IEJyYWlu 28889
+7LCp 28890
+IHByemVr 28891
+IEVp 28892
+IENvY28= 28893
+0LHQsA== 28894
+IHNob3ZlbA== 28895
+ZWE= 28896
+IGNob2NvbA== 28897
+IHJlYmVsbGlvbg== 28898
+IHNob3dj 28899
+IEhhbG8= 28900
+IGRpdmlkZW5k 28901
+bWlzc2lvbg== 28902
+IHVzYW5kbw== 28903
+IFsi 28904
+IGZhbGVp 28905
+QmxhY2s= 28906
+IFN1cmVseQ== 28907
+IMW7 28908
+IHBoaWxvc29waGVy 28909
+IG92ZXJoZQ== 28910
+IEJvcm4= 28911
+IG9iamV0aXZv 28912
+IDEyOA== 28913
+c2NoZWlk 28914
+IE5hemlz 28915
+IHNvbGNoZQ== 28916
+bGlmdA== 28917
+Y2VkZQ== 28918
+YWRvcnM= 28919
+IG1hcnNobQ== 28920
+IExPUkQ= 28921
+lOydtO2BrA== 28922
+IG93bmluZw== 28923
+Q29udA== 28924
+IGxhbmRzY2FwZXM= 28925
+IGxlbmRpbmc= 28926
+IEF1dGhvcml0eQ== 28927
+0L7QstC+0Lk= 28928
+b3F1 28929
+IFNlcw== 28930
+IEZlcnJhcmk= 28931
+IHJlc3BvbnNhYmls 28932
+IHbDoXJpb3M= 28933
+IGRlbGlj 28934
+IGVtYmFyaw== 28935
+IGVtYnJvaWRlcg== 28936
+IGZyYW1ld29ya3M= 28937
+IHNpbW1lcg== 28938
+IG5hY2lvbmFs 28939
+IHJlbWFpbmRlcg== 28940
+IFZpZWxsZWljaHQ= 28941
+IHF1aWVyZXM= 28942
+7JeU 28943
+IHRlc3Rvc3Rlcg== 28944
+aWhlbg== 28945
+IE96 28946
+w6hsZQ== 28947
+IHBvcnRyYXllZA== 28948
+zrrOtQ== 28949
+IFBvbGl0aWs= 28950
+IGFwZXJ0dXJl 28951
+IGJsYW5k 28952
+aW5kdXN0 28953
+INC+0LHRgNCw0YI= 28954
+IFRob3Vz 28955
+QmF5 28956
+IGRhbmRv 28957
+IHNoZXI= 28958
+IGFkbWlzc2lvbnM= 28959
+IENyZXc= 28960
+INGW0L0= 28961
+U0lOR0lORw== 28962
+IG91bmNl 28963
+IGl5 28964
+IGJhc2ls 28965
+IG92ZXJ0aW1l 28966
+IHRocmVhdGVu 28967
+IHBhcnRuZXJlZA== 28968
+IENhbm4= 28969
+YXZhbmE= 28970
+INC30L3QsNC10YLQtQ== 28971
+INC+0YLRgQ== 28972
+IFR1ZG8= 28973
+7L2U 28974
+IOuGgOs= 28975
+ZmVs 28976
+IHJlYXJy 28977
+IGlud2FyZA== 28978
+IFJvZ2Vycw== 28979
+4LmD4Lir 28980
+IHR3ZWFr 28981
+IGRyeWVy 28982
+Y2Vzc2lvbg== 28983
+IHJpZ29yb3Vz 28984
+IERhYXI= 28985
+b21pY3M= 28986
+IGZhdHM= 28987
+dmFk 28988
+IHppcHBlcg== 28989
+YWNjZXB0YWJsZQ== 28990
+IGRlbW9uc3RyYXRpbmc= 28991
+IFl1bQ== 28992
+IGJlYXU= 28993
+IHJvc3Rlcg== 28994
+IHByZWRvbWluYW50bHk= 28995
+0LXRgNGD 28996
+bmluZ2Fy 28997
+IHRyaWFuZ2xlcw== 28998
+IHRleHRpbmc= 28999
+IGJlcnJpZXM= 29000
+IOyCrOynhA== 29001
+YWRkZXI= 29002
+IGZhaXRlcw== 29003
+IEltYWdl 29004
+bGVyZQ== 29005
+IGJvdW5kcw== 29006
+IExhdXI= 29007
+IOyVhOustOs= 29008
+IG1pbw== 29009
+IHVzYQ== 29010
+INiw 29011
+IHRvZW4= 29012
+IEphbmc= 29013
+xb5l 29014
+Y2hvZA== 29015
+YW5hbg== 29016
+INC+0LHRgNCw0LfQvtC8 29017
+IHBlcnNldmVy 29018
+IFN3ZQ== 29019
+IGF1Z21lbnQ= 29020
+dWdnbGluZw== 29021
+acOocmVtZW50 29022
+aXN0bGVz 29023
+YWNqxJk= 29024
+OTE= 29025
+IG1haA== 29026
+IEtJUg== 29027
+RGll 29028
+IGRvd25oaWxs 29029
+IDE5Njg= 29030
+0L7RgNC+0YjQvg== 29031
+b2dyYXBoaWNz 29032
+IHTDpHNzw6Q= 29033
+6rKg7KOg 29034
+INC70LjRhw== 29035
+QVVESU8= 29036
+INC/0LvQvtGF 29037
+IHByb3Bvc2luZw== 29038
+IHRlbXB0ZWQ= 29039
+IGNvbnZlcnRpbmc= 29040
+IExlaHI= 29041
+IHBlcnNvbmU= 29042
+IEZlZWxpbmc= 29043
+7Ja07KO8 29044
+b21icmVz 29045
+INec15k= 29046
+IGd1cnU= 29047
+IGRlbWVudA== 29048
+0L3QuNC3 29049
+0LjRgtC10LvQtdC5 29050
+IGNvbXBhw7E= 29051
+IHJlZG8= 29052
+IGNvbmR1Y3Rvcg== 29053
+bWlh 29054
+IGlkb2xz 29055
+IE11bA== 29056
+IGluZXg= 29057
+IHTDpG3DpA== 29058
+IGltcGFjdGluZw== 29059
+IGRheWxpZ2h0 29060
+Z2ls 29061
+IGhlbGZlbg== 29062
+IGVudHNwcmVjaA== 29063
+IHdpxJlrcw== 29064
+IHNjcmlwdHVyZXM= 29065
+IGRpc21pc3NlZA== 29066
+IFBvZGNhc3Q= 29067
+2YXYsQ== 29068
+IGFubnVhbGx5 29069
+IHVzYWJsZQ== 29070
+IGxpYnJl 29071
+0L7Qt9C8 29072
+IHJ1YmJpc2g= 29073
+IGNvbnRpbnVhcg== 29074
+IGh1bWlsaQ== 29075
+IHNwZWVjaGVz 29076
+0YDQsNGH 29077
+YmFyZA== 29078
+NzE= 29079
+Pjw= 29080
+b2xvZ8OtYQ== 29081
+d2VhbHRo 29082
+IG1lZGl0YXRl 29083
+k6TsnZg= 29084
+IENyYWZ0 29085
+cml2 29086
+IEFnYWluc3Q= 29087
+IGNlcmFtaWM= 29088
+ZXNww6hyZQ== 29089
+IGNvbXBldGVudA== 29090
+IEhvcGtpbnM= 29091
+IGtpbG9z 29092
+IGdyYXZlbA== 29093
+IHBpc3Rvbg== 29094
+IGZyaWVuZHNoaXBz 29095
+IGVzY3Jl 29096
+IHZveg== 29097
+IEdlc2VsbHNjaGFmdA== 29098
+IHVudGVyc3TDvHQ= 29099
+IG11ag== 29100
+IHdhcm5pbmdz 29101
+cG9z 29102
+IFByb2Zlc3Npb25hbA== 29103
+d3N6eQ== 29104
+b2RsZQ== 29105
+YmFuZHM= 29106
+IHRlYW13b3Jr 29107
+c3RlbGx1bmc= 29108
+IGR4 29109
+IGF0dG9ybmV5cw== 29110
+IHdlaXRlcmU= 29111
+44WL44WL44WL 29112
+IE9yaWdpbmFs 29113
+15nXlw== 29114
+IGJyb2FkY2FzdGluZw== 29115
+INC/0LXRgNCy0YvQuQ== 29116
+dWNoaQ== 29117
+IGhldXJl 29118
+IGdyYWJz 29119
+IFdPUg== 29120
+IFBsYWlk 29121
+TWlu 29122
+IHBheg== 29123
+IFB1aXM= 29124
+dW11 29125
+aXRhdGVz 29126
+IGNvYXRz 29127
+IGJ1ZW4= 29128
+IGhlaXI= 29129
+IHBuZXVt 29130
+16nXqA== 29131
+ZW5zZXI= 29132
+IEpVREdF 29133
+IGJsb25kZQ== 29134
+4bmb 29135
+IGdhaw== 29136
+IHPEsWs= 29137
+IHF1b3RlZA== 29138
+IGVxdWlwbw== 29139
+IHdpc2hpbmc= 29140
+w61jaWE= 29141
+IHZlcmJz 29142
+IENhbmFkaWFucw== 29143
+IGdvdmVybmluZw== 29144
+IEV2YW5z 29145
+RXVybw== 29146
+IGdlbnJlcw== 29147
+IHVudGVyc2NoaWVk 29148
+IEJlY2t5 29149
+s7zqsozsmpQ= 29150
+IGVpbmdl 29151
+IFJhaXNl 29152
+b2xhbmQ= 29153
+IFN0cmF0ZWc= 29154
+IGVyZXM= 29155
+IFZldGVyYW5z 29156
+IGJyZWFrb3V0 29157
+IHNhbnTDqQ== 29158
+IGFkZWw= 29159
+IGludmVzdGlnYXRlZA== 29160
+IHBldXI= 29161
+IGFnaWxl 29162
+IHJhaWxyb2Fk 29163
+YW5za2E= 29164
+INC10Lk= 29165
+IGV4cG9z 29166
+YXRvcmllcw== 29167
+IENvbnRlbnQ= 29168
+IHRydXRocw== 29169
+IFRyYWls 29170
+IGd1YQ== 29171
+IHBvcmVz 29172
+IHdyaXRpbmdz 29173
+IFVocg== 29174
+IFRoYXRz 29175
+IGljaW5n 29176
+T0M= 29177
+IFByb2R1Y3Rpb24= 29178
+IGNhcm5l 29179
+SVNT 29180
+IG5pbmd1w6lt 29181
+bm9u 29182
+IHZpY2lvdXM= 29183
+15XXlA== 29184
+IHJlY29ubmVjdA== 29185
+IGNlbnRyZXM= 29186
+IEtlbQ== 29187
+IGNyZWFzZQ== 29188
+IOydtOuvuA== 29189
+0LDQudGC0LXRgdGM 29190
+INCx0L7RgA== 29191
+IEhhecSxcg== 29192
+INGB0YPQtA== 29193
+IMO6bmljYQ== 29194
+b3dhxYI= 29195
+IGFkaGVy 29196
+aHVh 29197
+Wlo= 29198
+IHByZWNpc28= 29199
+IGN1cnJlbnRz 29200
+IHNlYXNvbmVk 29201
+IElvVA== 29202
+IEJpc2hvcA== 29203
+c3RlZA== 29204
+IEJlcm5hcmQ= 29205
+7KSY 29206
+IEdsZW5u 29207
+IGt0w7NyeW0= 29208
+4Li34LmI 29209
+IGFzdHJvbG9n 29210
+IEtvdA== 29211
+IHBhcmZvaXM= 29212
+IGZvcndhcmRz 29213
+IFdpxJk= 29214
+IM6Y 29215
+IG5hbm8= 29216
+c3Vi 29217
+IEJyaWxs 29218
+IGdyaXQ= 29219
+IGNpdGVk 29220
+Z2Fkbw== 29221
+IG1lbHRz 29222
+IGZvcmPDqQ== 29223
+4paI4paI 29224
+IGJham8= 29225
+IGRpc2NyZXRpb24= 29226
+sLA= 29227
+YXRpdml0eQ== 29228
+IHNpdHVhdGVk 29229
+0YnQtdC1 29230
+INC/0YDQuNC90YbQuNC/ 29231
+YW1heg== 29232
+IGFxdWFyaXVt 29233
+IGRpc3NvbHZl 29234
+IEdvZHM= 29235
+U3VwZXI= 29236
+IGFtaWQ= 29237
+ems= 29238
+YW1wZg== 29239
+IGhlbGE= 29240
+JyE= 29241
+IGRldmVsb3BtZW50YWw= 29242
+IERpc2U= 29243
+INGA0LDQsdC+0YLQsNC10YI= 29244
+IHNuYXBzaG90 29245
+1bg= 29246
+IFl1ZQ== 29247
+IEh1bGs= 29248
+IERvb20= 29249
+IEZlbGl4 29250
+IHLDqWY= 29251
+TWFsZQ== 29252
+cGhhbnRz 29253
+RU5T 29254
+IE1lY2hhbg== 29255
+IEdvbGY= 29256
+IGdlbmVyb3NpdHk= 29257
+w6R0emU= 29258
+IHVubG9ja2Vk 29259
+7YOB 29260
+b2NhbHlwc2U= 29261
+QWxyaWdodA== 29262
+IOqwnOs= 29263
+INeQ15HXnA== 29264
+IEtlZXBpbmc= 29265
+IGNvbGxhYm9yYXRpbmc= 29266
+Y2hpZWY= 29267
+IEZlcm5hbmRv 29268
+IGNoZWZz 29269
+IO2UvOu2gA== 29270
+IHNraXBwZWQ= 29271
+IHBlcnNvbm4= 29272
+IGF4ZQ== 29273
+Y2hleg== 29274
+IGV4dHJhY3Rpb24= 29275
+IEFW 29276
+IEdpYmJz 29277
+IO2c 29278
+IHPEsQ== 29279
+SUFN 29280
+Vmlldw== 29281
+IEdSQU5U 29282
+IOuquA== 29283
+IHZlcmlmaWNhdGlvbg== 29284
+IGRlcGljdGVk 29285
+IE1veg== 29286
+b3V4 29287
+IHR1bA== 29288
+IHNjYW5uZXI= 29289
+IGNvbWVkaWFu 29290
+IFZvbGtz 29291
+IEpFRkY= 29292
+p4Q= 29293
+IGRpc3RyYWN0aW9u 29294
+csOh 29295
+IElOVEVS 29296
+IHNpbmNlcg== 29297
+INee16o= 29298
+INep16A= 29299
+IGNvbnN0cnVjdGl2ZQ== 29300
+YXJm 29301
+IOuIhOs= 29302
+IGVjbw== 29303
+cmFtb3M= 29304
+IHJlbmV3ZWQ= 29305
+aW5lbWVudA== 29306
+IFVi 29307
+IFBlcHBlcg== 29308
+7KeA6rCA 29309
+IERhcndpbg== 29310
+IG1lcmNoYW5k 29311
+IHbDoXJpYXM= 29312
+w6hjZQ== 29313
+Tkc= 29314
+IOychO2VtOyEnA== 29315
+INCw0LrRgtC40LI= 29316
+IFVudGVycw== 29317
+2LnZhA== 29318
+IGludHJpYw== 29319
+b21tYQ== 29320
+aWV2aW5n 29321
+IENhcm9saW5l 29322
+IFBSRVM= 29323
+IHBlcmZvcm1lcg== 29324
+IGF1dG91cg== 29325
+IHV0dGVybHk= 29326
+IHN5bnRoZXNpcw== 29327
+IGxlc2JpYW4= 29328
+IHJldHJpZXZl 29329
+IG1hbmVpcmE= 29330
+IGltcGFpcg== 29331
+IG1lbnRvcmluZw== 29332
+IFNvdWxz 29333
+IEdvUHJv 29334
+0YDQsNGC0Yw= 29335
+IGNvc2U= 29336
+IFNTRA== 29337
+SVJF 29338
+IHVwZnJvbnQ= 29339
+IEF1bg== 29340
+IGdhbWVy 29341
+IGxpdHQ= 29342
+IGFnZ3Jlc3Npb24= 29343
+IExpa2V3aXNl 29344
+IEJldHR5 29345
+IERhcnQ= 29346
+IERMQw== 29347
+aXNobWVudA== 29348
+7J6l7J2E 29349
+Y3JlYW0= 29350
+IEJhYnlsb24= 29351
+IG51Zw== 29352
+YnJhcg== 29353
+IGF5bsSx 29354
+YW1pbHk= 29355
+YmlrZQ== 29356
+YWhhaGFoYQ== 29357
+bG95ZA== 29358
+IG1pcmE= 29359
+IHBlcm1l 29360
+IEdhbWluZw== 29361
+IGZpcm13YXJl 29362
+TWE= 29363
+IGFzc2lzdGVk 29364
+YXRpY3M= 29365
+IOyVnuycvOuhnA== 29366
+IE1lbnRhbA== 29367
+bmllanM= 29368
+IEl6 29369
+b3fEhQ== 29370
+IHRvdWdoZXI= 29371
+IGRlZWQ= 29372
+IHN0eWxpc2g= 29373
+IFRvb2xz 29374
+IEhhbXA= 29375
+IHN1bnNjcmVlbg== 29376
+IGFydGljdWxhdGU= 29377
+aXll 29378
+0LjRhA== 29379
+IFNwcmVhZA== 29380
+IEhBVkU= 29381
+IHN3aXJs 29382
+IHNwb25zb3Jpbmc= 29383
+aW92YXNjdWxhcg== 29384
+bWVzaQ== 29385
+IHJlbGF4YXRpb24= 29386
+INGB0LLQvtC40YU= 29387
+IG1hcmdpbnM= 29388
+IHNhxJ8= 29389
+IFByaWRl 29390
+IM+Ezr/Phc+C 29391
+0LjRhtC4 29392
+ZW5jaQ== 29393
+RG9lcw== 29394
+IGNvcnBzZQ== 29395
+IGVuZHVyYW5jZQ== 29396
+IO2emA== 29397
+7Lm0 29398
+IGhhaXJjdXQ= 29399
+IGludGVycnVwdGVk 29400
+IHdpbmR5 29401
+IENhbGVi 29402
+z4HPhw== 29403
+IFBvdXJxdW9p 29404
+IGhvbGlzdGlj 29405
+dWNsZWFy 29406
+IFdob2xl 29407
+QWN0 29408
+IGdhbGxvbg== 29409
+Y2FkZQ== 29410
+IFJlZ2lvbmFs 29411
+cm9hZHM= 29412
+IFNjaG5l 29413
+w6FuZw== 29414
+INC40LfQvNC10L0= 29415
+IG1lbnVz 29416
+IHNwbGl0dGluZw== 29417
+IHByaWNlZA== 29418
+IM6T 29419
+IHVzZXJuYW1l 29420
+INCe0Yc= 29421
+IGNvbXByZXNzZWQ= 29422
+eWlu 29423
+IGd1YXJkaWFu 29424
+IGdvb2Y= 29425
+IGNoZWNrbGlzdA== 29426
+IGludGVyY2hhbmdl 29427
+IGV4cGVkaXRpb24= 29428
+IGV4dGVybg== 29429
+IGluZnJhcmVk 29430
+ZW5nbw== 29431
+IGRlbnlpbmc= 29432
+IHBhY2tldHM= 29433
+b25lbnQ= 29434
+QkI= 29435
+IEluY3Jl 29436
+IHNpbmk= 29437
+w59lcg== 29438
+w6hn 29439
+bWFhbA== 29440
+Z2VuZXJhdGlvbg== 29441
+IG1pbm9yaXRpZXM= 29442
+IGxsZXZhcg== 29443
+IG5vbWluYXRpb24= 29444
+IGNvbnNpZA== 29445
+INec16I= 29446
+bXXFnw== 29447
+IEVzYw== 29448
+IG51bWVyYXRvcg== 29449
+IGthaWs= 29450
+IGt0w7NyeWNo 29451
+aWVzZW4= 29452
+IHbDqg== 29453
+IFVTUw== 29454
+IFByaXZhdGU= 29455
+INC+0LTQvdC+ 29456
+IGFsw6lt 29457
+w610dWxv 29458
+IGxpbWI= 29459
+IGZvcmdpdmVu 29460
+IGRpc2Nsb3N1cmU= 29461
+z4TOrw== 29462
+IG5pbmfDum4= 29463
+IHRoZXJhcGV1dGlj 29464
+IG5lZ290aWF0aW5n 29465
+IE5pa2U= 29466
+ZW5zZWZ1bA== 29467
+IGluY2Fw 29468
+IGZsYWdzaGlw 29469
+dG93bg== 29470
+4og= 29471
+IM+Azr/Ouw== 29472
+IHdvbHZlcw== 29473
+IHZpb2xhdGlvbnM= 29474
+IEFybm9sZA== 29475
+IGludGVydmVuZQ== 29476
+IGhlYXRlcg== 29477
+IHJlY3Vyc29z 29478
+IG1haWQ= 29479
+6rK8 29480
+INC00LDQstCw0LnRgtC1 29481
+IENlbGVicg== 29482
+IGNhcGU= 29483
+IFN0eQ== 29484
+YWluZW4= 29485
+c2l0ZQ== 29486
+Ymlq 29487
+INC/0L7Qu9GM0Lc= 29488
+IGZyYW1lZA== 29489
+IHB1Ymxpc2hlcnM= 29490
+INGH0YPRgtGM 29491
+IHRlbXB0YXRpb24= 29492
+IGNlcnRlemE= 29493
+IGV4ZW1wdA== 29494
+7Iq5 29495
+c2VsbGluZw== 29496
+IFRhc2s= 29497
+aG9vbg== 29498
+IENvYw== 29499
+IFBhcmtz 29500
+IHJlcGV0aXRpb24= 29501
+INGC0YPQtNCw 29502
+IGVuc2w= 29503
+IGRlxJ9pxZ8= 29504
+IE9ybGFuZG8= 29505
+IE1haW50ZW4= 29506
+b2N1bWVudA== 29507
+IEhD 29508
+IHNjb290ZXI= 29509
+INC90LDQv9C40YE= 29510
+IHRpZ2h0ZXI= 29511
+IHRlYXNl 29512
+IHJlbW92ZXM= 29513
+IGtpamtlbg== 29514
+INGB0YPRidC10YHRgtCy 29515
+IHRow6k= 29516
+INCy0YvQs9C70Y/QtA== 29517
+IHJlbGlldmU= 29518
+IG1pdMOk 29519
+IHN0YXRpb25hcnk= 29520
+w7ZmZg== 29521
+cGFibGU= 29522
+IGFydGVy 29523
+IGTDqWY= 29524
+cmF0aXZl 29525
+IGNvbmVjdA== 29526
+IHNhZGRsZQ== 29527
+IERpYW5l 29528
+IGNvbW1lbW9y 29529
+ZmVuZGlt 29530
+U8Ot 29531
+IO2BtOs= 29532
+IG1hbmdl 29533
+YXR0ZQ== 29534
+IGFycm9nYW50 29535
+IHJvYm90aWM= 29536
+IGdpw6A= 29537
+IG5laWdoYm91cmhvb2Q= 29538
+aXNzb24= 29539
+INC00LLQuNC2 29540
+IFJJ 29541
+IE5vcm1hbg== 29542
+YnJhbmQ= 29543
+YW1hdGlvbg== 29544
+IHJhem9y 29545
+IG11cmRlcnM= 29546
+INGC0YM= 29547
+IHdzenlzdGtpbQ== 29548
+IHV0aWxpdGllcw== 29549
+IG1pY3Jvc2NvcA== 29550
+6r8= 29551
+IGRhcXVp 29552
+b2xsYXI= 29553
+INCU0LDQstCw0LnRgtC1 29554
+IGFubsOpZQ== 29555
+IGtpbG9tZXRyZXM= 29556
+IGhvbW9zZXh1YWw= 29557
+IGFyY2hpdGVjdHM= 29558
+IG5peWU= 29559
+TEVS 29560
+IG1pY3JvcGhvbmVz 29561
+IFN0dW5kZW4= 29562
+IGNvbnNlY3V0aXZl 29563
+aWVuZGE= 29564
+dsOkbmQ= 29565
+REVS 29566
+IGxpZnRz 29567
+IE1lYXQ= 29568
+IHNhdmV6 29569
+7ZaI642Y 29570
+TWVu 29571
+IGRpc21hbnQ= 29572
+6rGw66W8 29573
+IGluc3VsYXRpb24= 29574
+IHNjYWxs 29575
+IHNwb29reQ== 29576
+IHBhcmM= 29577
+IGJhbGxldA== 29578
+IFdoYXRzQXBw 29579
+IGZyYW5j 29580
+IGRlbGliZXJhdGU= 29581
+IO2FjA== 29582
+IG1hcnM= 29583
+IFp1cg== 29584
+UHI= 29585
+ZGlzY2lwbGluYXJ5 29586
+IG9ic2Vzc2lvbg== 29587
+0LzQtQ== 29588
+IG1hcmNoaW5n 29589
+IEVtZXJnZW5jeQ== 29590
+aWd1b3Vz 29591
+IHN6eQ== 29592
+IExhbmRz 29593
+IGJvYXJkaW5n 29594
+INC/0L7Rh9GC0Lg= 29595
+IGVudnk= 29596
+IGNvbXBhc3Npb25hdGU= 29597
+IG1lcmNp 29598
+IGRlc2lyYWJsZQ== 29599
+ZGFsZQ== 29600
+IGNhbsSxbQ== 29601
+IEFudGFy 29602
+dGVtcHM= 29603
+IGNvbmZpZ3VyZWQ= 29604
+IENvbXBhcmVk 29605
+bmVo 29606
+aWNhdGluZw== 29607
+IG5pY2tlbA== 29608
+2YjZgg== 29609
+2YPZiNmG 29610
+b3Blcw== 29611
+IGZvcm11bGFz 29612
+INCV0YHRgtGM 29613
+IHBvYmw= 29614
+IFBK 29615
+IEx1ZA== 29616
+IEJyaWQ= 29617
+IEhvZw== 29618
+IEJyaXM= 29619
+SmVu 29620
+IHNoYWRpbmc= 29621
+IFlhcw== 29622
+IGRpc3R1cmJlZA== 29623
+IHJlY29tbWVuZGluZw== 29624
+IGPDqQ== 29625
+IEhPVw== 29626
+7JeI7Ja0 29627
+IHJldmVyc2Vk 29628
+IEludGVyZXN0aW5nbHk= 29629
+aW94aWQ= 29630
+IOyYpOy8gOydtA== 29631
+4bq/dQ== 29632
+eHg= 29633
+IG91YWlz 29634
+IFlvdVR1YmVycw== 29635
+IFJvc2E= 29636
+IEhhdXB0 29637
+amFkaQ== 29638
+IHZsb2dz 29639
+IGN1bHR1cmE= 29640
+IExlYWRlcnNoaXA= 29641
+IEhlcA== 29642
+IGlsbHVt 29643
+tOuPmQ== 29644
+IGN1c3RvbWl6ZWQ= 29645
+IG1hcmNh 29646
+IHF1YXRybw== 29647
+INC90LDQsw== 29648
+IFNwYWNlWA== 29649
+IEVpZ2Vu 29650
+YXN0aW5n 29651
+IG9sZHXEn3U= 29652
+IGZvcnRz 29653
+cmltZW50 29654
+aWVuY2lh 29655
+IHRlbmly 29656
+cm9mZmVu 29657
+IDE5Nzk= 29658
+IGNpZQ== 29659
+IOuQmOqzoA== 29660
+IGVzY3Jp 29661
+z4zPgg== 29662
+7Y+s 29663
+dXp6eQ== 29664
+Q29uZw== 29665
+7J247J20 29666
+R3JlYXQ= 29667
+c2ls 29668
+w6ljaA== 29669
+IG11bHRpYw== 29670
+IERpc2s= 29671
+spU= 29672
+IGZhemxh 29673
+IGxldmFudA== 29674
+IGFiYWpv 29675
+dXJyeQ== 29676
+c3RydQ== 29677
+IOuoueuKlA== 29678
+IGFjY2Vzc29yeQ== 29679
+INC00LLQuNCz 29680
+IFJpZA== 29681
+MjAxOQ== 29682
+IGRvd25zdHJlYW0= 29683
+IGtheg== 29684
+dXRhbg== 29685
+IGNoYXJjb2Fs 29686
+IGFmZWN0 29687
+d3U= 29688
+IGNvbnRleHRz 29689
+IGZlYXJlZA== 29690
+IOyEpA== 29691
+IGhpc3Rvcmllcw== 29692
+IGZhcw== 29693
+ZW5zaWJsZQ== 29694
+IGNvY29h 29695
+aWxsYXI= 29696
+Z2VvbnM= 29697
+IHNwaXJpdHVhbGl0eQ== 29698
+IFBldw== 29699
+IHBoYXJtYWN5 29700
+IHBhc3Npb25z 29701
+IGJvcw== 29702
+IGFsbMOh 29703
+IHRocml2aW5n 29704
+IFJlYWN0 29705
+IG9jY3VweQ== 29706
+IHdpdGhkcmF3YWw= 29707
+IGFsbG93YW5jZQ== 29708
+IEZyYWt0aW9u 29709
+IGJ1ZGRpZXM= 29710
+IGlkbGU= 29711
+IGRpc3NvbHZlZA== 29712
+IHByZXZhbGVudA== 29713
+IG1pbGl0YXI= 29714
+IHNlbnNpbmc= 29715
+IHBvamF3 29716
+IGFuY29yYQ== 29717
+IGFidW5kYW50 29718
+IGhhaXJzdA== 29719
+IHR3ZWU= 29720
+IG7DpGNoc3Rl 29721
+IE3DtmdsaWNoa2VpdA== 29722
+IGhvbw== 29723
+dWZmaWNpZW50 29724
+IGZhbnRhc3Q= 29725
+IGVkaWJsZQ== 29726
+IOuWqOyWtOw= 29727
+7JuD 29728
+IHZlaW4= 29729
+dWNjaQ== 29730
+IGRldm90aW9u 29731
+IGNvbmNlYWxlcg== 29732
+aW5jb21l 29733
+IHJlY3ljbGVk 29734
+IOyKpO2DgA== 29735
+IHBvbnRvcw== 29736
+IGRlc3N1cw== 29737
+IHbDqXJpdA== 29738
+IHJlZmxlY3Rpb25z 29739
+IEFB 29740
+IHRha2Vhd2F5 29741
+YmFyZQ== 29742
+IENvbnRhY3Q= 29743
+ZWls 29744
+IEhlYXI= 29745
+IG1pcmFj 29746
+IEdlcmlsaW0= 29747
+INGB0LDQvNGL0Lk= 29748
+IHZpdm8= 29749
+IGtpbG9ncmFtcw== 29750
+IENyaW0= 29751
+w7t0 29752
+Nzg= 29753
+IHNpbmNlcmVseQ== 29754
+cmF6 29755
+IOuztQ== 29756
+IGFycml2 29757
+IGNvbmNlcHRpb24= 29758
+IFBlcnNpYW4= 29759
+IHNqw6Rs 29760
+IHN0YXJyaW5n 29761
+IOyVhOustA== 29762
+IEZvcmV2ZXI= 29763
+0LXRgdGC0Yw= 29764
+IHZlaWw= 29765
+IHN1YnRpdA== 29766
+b2RrYQ== 29767
+INC+0YLQvdC+0Yg= 29768
+IGNvb2tz 29769
+0LXQvdGP 29770
+S2F5 29771
+IG5pw7Fvcw== 29772
+IFBob25l 29773
+IHN0aXRjaGluZw== 29774
+IGZpbmdlcnByaW50 29775
+zrvOrA== 29776
+IGRlZGljYXRl 29777
+IExvYg== 29778
+IGJsYWNrcw== 29779
+IEJsZQ== 29780
+Ym91dA== 29781
+IMSRYW5n 29782
+IGVrcw== 29783
+IHNxdWFzaA== 29784
+IEvDvA== 29785
+b2Rp 29786
+IG7GsOG7m2M= 29787
+IHZveWFnZQ== 29788
+IHBsYXlmdWw= 29789
+INil2YTZiQ== 29790
+YW5pYw== 29791
+IGNvbmRlbW4= 29792
+IELDtnlsZQ== 29793
+IFBvbGl6ZQ== 29794
+IGF5dWRh 29795
+IHBhbQ== 29796
+4LmE4Lib 29797
+IEthdGh5 29798
+0LXQtNC40L0= 29799
+0L3QvtCy0LA= 29800
+IGJyaWc= 29801
+ZWdlcg== 29802
+IGVhZ2xl 29803
+IHZpc2lvbnM= 29804
+IO2VreyDgQ== 29805
+IHNoaXR0eQ== 29806
+IGhvdHQ= 29807
+IEJyaXR0 29808
+dXRvcnM= 29809
+RU5URQ== 29810
+IHBob24= 29811
+IEJpbmc= 29812
+INC/0L7QtNC00LXRgNC2 29813
+c3ByaW5n 29814
+ZXR0ZW4= 29815
+IHBpbGdy 29816
+IGVkaXlvcg== 29817
+0LXQvdGC0Ys= 29818
+YWdnaW8= 29819
+IGp1bA== 29820
+IGNvbXByZW5k 29821
+dGVpbA== 29822
+INiy 29823
+IHBlcmZvcm1lcnM= 29824
+IGluZmFtb3Vz 29825
+IE1L 29826
+56o= 29827
+b3RsZQ== 29828
+ZWZm 29829
+IEhhc2g= 29830
+IGNvd2FyZA== 29831
+IEJSQQ== 29832
+IERE 29833
+IGNvbWlkYQ== 29834
+IHBsYXRh 29835
+IGZsYXA= 29836
+IE1laHI= 29837
+cmlidXRpb24= 29838
+IFllbWVu 29839
+IG15c3Rlcmllcw== 29840
+IMSweWk= 29841
+IHN0ZWxs 29842
+IGV5ZWxpbmVy 29843
+IGRlbGVz 29844
+IG5haWxlZA== 29845
+IGlsbG5lc3Nlcw== 29846
+IHN0YWNrcw== 29847
+IHRyYWJhamFy 29848
+Zmxvd2Vy 29849
+Y2l1 29850
+IGNydWRl 29851
+IHN1YnN0YW50aWFsbHk= 29852
+IGhvbWVt 29853
+IG5lcGhldw== 29854
+IHN0YW1wcw== 29855
+IGNhcmJz 29856
+0YzRgtC1 29857
+bW9vdGg= 29858
+IHR1bm5lbHM= 29859
+YWNpZQ== 29860
+IFNlw7E= 29861
+IEhlcmE= 29862
+IOyVhOuLiOyXkOyalA== 29863
+IFd5b21pbmc= 29864
+IEhETUk= 29865
+IExpcw== 29866
+dWNpw7Nu 29867
+IHN0ZWVy 29868
+0L7Rjg== 29869
+0LjRgtCw 29870
+TlQ= 29871
+IOyWvOq1tA== 29872
+IHBhbG1z 29873
+IG5lb24= 29874
+0L7QstCw0L3QuNGP 29875
+IGZpbHRlcmluZw== 29876
+IGpvdWVy 29877
+IEjDtg== 29878
+INC90LXRgQ== 29879
+6rKg7Ja07JqU 29880
+IDgx 29881
+IHN0b3J5bGluZQ== 29882
+IHByemVw 29883
+IHRoYW5raW5n 29884
+IEJvZWluZw== 29885
+IHNvZnRseQ== 29886
+amVt 29887
+0LDQu9GM0L3Ri9GF 29888
+IGZsYXNobGlnaHQ= 29889
+INC/0YM= 29890
+IFdPTUFO 29891
+4bqvYw== 29892
+w61jaA== 29893
+IGx1eHVyaW91cw== 29894
+IHfDvG4= 29895
+IGltcGFjdGZ1bA== 29896
+IGNvbnNvbg== 29897
+cmV1 29898
+aXJyaW5n 29899
+aWZ0ZXI= 29900
+IGNvbnN0aXR1ZW50cw== 29901
+IDk0 29902
+IFRvdQ== 29903
+Z29t 29904
+IOyDneqwgeydhA== 29905
+IHN0ZXJlb3R5cGVz 29906
+IG1vxbxsaQ== 29907
+gqg= 29908
+IHBlbmNpbHM= 29909
+INGB0LvQvtC2 29910
+IGlocmVt 29911
+IEJlc2No 29912
+IEtvaA== 29913
+IEVudHNjaGVpZA== 29914
+IGxlaw== 29915
+IGbDtnJz 29916
+IHRvdGFsbWVudGU= 29917
+IGxpdmVseQ== 29918
+IGVudHJvcHk= 29919
+IGRpc2Nlcm4= 29920
+INCX0L3QsA== 29921
+IGRvdg== 29922
+IG15dGhvbG9neQ== 29923
+YXBhbmVzZQ== 29924
+IGFwcHJveGltYXRl 29925
+0LDRgtC40LI= 29926
+aWZpYWJsZQ== 29927
+IFNlbw== 29928
+tOyLrO2eiA== 29929
+IOyYtw== 29930
+IHRlbXBvcmFs 29931
+IGlU 29932
+IGVzdGF0 29933
+0LrQuNC8 29934
+IHNwcmluaw== 29935
+IGdydW5k 29936
+IGluZmFudHJ5 29937
+IHNjaGFmZmVu 29938
+IGFuaw== 29939
+cmlhZ2Vz 29940
+IFllb24= 29941
+IE1vcm9j 29942
+IGludmFzaXZl 29943
+gZQ= 29944
+IHBhcmVudGluZw== 29945
+IFJpcw== 29946
+aWJpbGU= 29947
+IG1vZHM= 29948
+INC/0YDQvtCy0LXRgA== 29949
+IFRoaW5n 29950
+IFdoZXJldmVy 29951
+IGFja25vd2xlZGdpbmc= 29952
+IHBhd24= 29953
+dW1tZXI= 29954
+b3Ji 29955
+Njk= 29956
+IHJldHJvdXZl 29957
+IHJlbGllcw== 29958
+IEhpZ2h3YXk= 29959
+IGF3ZQ== 29960
+aXRhaXJl 29961
+IGFwcGxpY2FudA== 29962
+IGFpc2xl 29963
+d29ybQ== 29964
+IHBheWxvYWQ= 29965
+IGNhcnJl 29966
+IEJhY2g= 29967
+IOy5nOq1rOs= 29968
+0L3QuNC1 29969
+IGl0w61z 29970
+b25uYWlzZQ== 29971
+c29s 29972
+YWxnaWE= 29973
+IHJvY2tpbmc= 29974
+IGJlc3Rlbg== 29975
+cml0ZXM= 29976
+Xl4= 29977
+0LjQvdC+0Lk= 29978
+IGJhaXhv 29979
+IOq4sOyWtQ== 29980
+0L7RgtGA0Lg= 29981
+c2lt 29982
+IGluY2Fybg== 29983
+64uk7J2M 29984
+IGxpY2s= 29985
+c2lkZWQ= 29986
+IDcx 29987
+Zm9yZGVy 29988
+IHJlc29uYW5jZQ== 29989
+IHRlZ2Vu 29990
+IG1ldGFwaA== 29991
+b3dzZXI= 29992
+INeQ16DXl9eg15U= 29993
+P+OAjQ== 29994
+IHNwaWVsZW4= 29995
+IHZvbGxleQ== 29996
+lOydtO2BrOyXhQ== 29997
+bG9va2Vk 29998
+IHNlbnRlbmNlZA== 29999
+IG11bHRpcGx5aW5n 30000
+IGlkZWFscw== 30001
+IHdhaHJzY2hlaW5saWNo 30002
+IGRlcG9zaXRz 30003
+YmlsaXI= 30004
+IGVmZmV0 30005
+aWxsb24= 30006
+iOunjA== 30007
+IHRlc3RpbW9u 30008
+IHphd3N6ZQ== 30009
+INC/0YDQvtGG0LXRgdGB 30010
+IExhdg== 30011
+IHRyYXZhaWxsZXI= 30012
+IGxhaXNzZQ== 30013
+IE1vdW50YWlucw== 30014
+INGA0L7QsQ== 30015
+IGV4YW1pbmVk 30016
+aXR1cw== 30017
+V2Fz 30018
+0LvRiw== 30019
+IGF0dHJpYnV0ZWQ= 30020
+IOyKuQ== 30021
+IEJhcm9u 30022
+IGdlcA== 30023
+IGF0dGVudA== 30024
+IENvbGxlY3Rpb24= 30025
+IHRoZWF0 30026
+IENhaQ== 30027
+IHdlbGxz 30028
+IGh1bWFubw== 30029
+IEhhc3Q= 30030
+INGF0L7RgtGP 30031
+Y3phcw== 30032
+IHBlcm1pdHM= 30033
+IGxlZ2c= 30034
+IGVwbw== 30035
+IEZlbg== 30036
+IHRoaQ== 30037
+IEZvaQ== 30038
+IMOpbGVjdA== 30039
+IDgz 30040
+IG92ZXJ0aA== 30041
+IHRlbmFudA== 30042
+TmV4dA== 30043
+IHByYWlzZWQ= 30044
+c2VjdXJpdHk= 30045
+IEltcGFjdA== 30046
+IHZvdWNo 30047
+IG5lZ8Oz 30048
+IHVudmU= 30049
+IGNyaXRpY2l6ZQ== 30050
+IEtlbnlh 30051
+IHRhY3RpYw== 30052
+IGxvZ3I= 30053
+IHBvaXM= 30054
+IHBhcGE= 30055
+c3BlYWtz 30056
+8J+R 30057
+aXNwZXJz 30058
+IHN1cnBsdXM= 30059
+IGNvbGRlcg== 30060
+cGxldHM= 30061
+IFZpZW5uYQ== 30062
+IExlYWQ= 30063
+IGFlcmlhbA== 30064
+IFRhaA== 30065
+0LXQvdGC0L7Qsg== 30066
+IEdyZWVrcw== 30067
+Q2Ft 30068
+IG3DoXhpbQ== 30069
+IGt1aW4= 30070
+Y2hpbw== 30071
+IGRlbW9uc3RyYXRlcw== 30072
+YW5vcw== 30073
+IENlcnQ= 30074
+INGN0L0= 30075
+IGJsb2dz 30076
+IOyEnOyauA== 30077
+IGJlYW1z 30078
+0LjQutC+0LI= 30079
+IHByb21wdGVk 30080
+IGZyaWdodGVuaW5n 30081
+IFBvcnNjaGU= 30082
+bGFyxLFuxLE= 30083
+IGNoaWxsaW5n 30084
+aXNwaGVyZQ== 30085
+IGZsYXNoaW5n 30086
+IEthcmQ= 30087
+YnJlYWQ= 30088
+IGV4aA== 30089
+IHR5Y2tlcg== 30090
+IGVjb2xvZ2ljYWw= 30091
+IE1hZQ== 30092
+INee15DXldeT 30093
+IOuCmOuPhA== 30094
+0LvQvtC9 30095
+eXNz 30096
+IHBlcmd1bnQ= 30097
+IHByaXg= 30098
+aXp6YXJk 30099
+IGNhbmNlcnM= 30100
+IDkx 30101
+c3VzcA== 30102
+IEl0ZW0= 30103
+xZ9h 30104
+IHBlc3Q= 30105
+IHRha8SF 30106
+IGx5bXBo 30107
+IFBhdHJp 30108
+ZmlsbA== 30109
+IHJlY29ubmE= 30110
+IG9wdGltaXNt 30111
+IG1pbWlj 30112
+IOyynA== 30113
+IE1hZGFtZQ== 30114
+b2N5 30115
+bGluaW5n 30116
+ZXJtZQ== 30117
+IGZvbGRlcnM= 30118
+IGN6xYI= 30119
+dWNoYXI= 30120
+IGN1cnNv 30121
+IGJyZWFjaA== 30122
+0L3QuNGC0Yw= 30123
+IHBhbWnEmQ== 30124
+IGVsaWc= 30125
+IGF1dG9w 30126
+Rmxvdw== 30127
+IHByb2dyYW1tZWQ= 30128
+IFByb2Nlc3M= 30129
+IGZpZ3Vy 30130
+IFNG 30131
+IEVsZXM= 30132
+IHByb2dyYW1tZXM= 30133
+IGRpenp5 30134
+7Iuc6rCE 30135
+INC70LjQsdC+ 30136
+IHNuaWZm 30137
+IFNlYmFzdGlhbg== 30138
+IEh5ZQ== 30139
+IDQwMDA= 30140
+IHBlcm1pdGU= 30141
+INC30LDRiQ== 30142
+IGd1aXQ= 30143
+IERhaXM= 30144
+IGFjY29yZGFuY2U= 30145
+IG1vZHVsYXI= 30146
+b2dlbmVvdXM= 30147
+IHBvdXF1aW5obw== 30148
+IGFydGlsbGVyeQ== 30149
+IGx1YnJpYw== 30150
+IHZvbGNhbg== 30151
+IE5I 30152
+8J+k 30153
+IGRlYW4= 30154
+Umg= 30155
+IG1pbmlzdHJl 30156
+IEludg== 30157
+IEJ1bGdhcg== 30158
+IERhdGVu 30159
+6I4= 30160
+SW0= 30161
+IG9yaWdpbmF0ZWQ= 30162
+IE5peG9u 30163
+aW50ZWdy 30164
+IGxhY2tz 30165
+IE5hY2h0 30166
+7Ja064KY 30167
+Y2FtZXJh 30168
+IHJhZGlzaA== 30169
+a2l5ZQ== 30170
+IGFuZ2Vz 30171
+IHByw6lm 30172
+anVr 30173
+IEJlZQ== 30174
+IEJV 30175
+INCy0L7RgdC/ 30176
+IEJU 30177
+w6ptZXM= 30178
+IFN0w7xjaw== 30179
+IEluaw== 30180
+IFNlcmdlYW50 30181
+IE11bHRpcA== 30182
+IGhpw6diaXI= 30183
+INCh0LDQvA== 30184
+IETDqQ== 30185
+b2xwaA== 30186
+7Ja4 30187
+IGltcGF0 30188
+IOyViuqzoA== 30189
+INGC0LDQutC+0LPQvg== 30190
+INC90LDQstC10YDQvdC+0LU= 30191
+IHVucHJlZGljdGFibGU= 30192
+IG1lbmQ= 30193
+IOyXhuyWtOyalA== 30194
+IGpha2llxZs= 30195
+IGFubmk= 30196
+IGRvbm7DqQ== 30197
+IEtpcnN0eQ== 30198
+IHJlY3Rhbmd1bGFy 30199
+IGVtcGV6YXI= 30200
+IEV4Y2hhbmdl 30201
+6rCU 30202
+IMOpY29ub20= 30203
+ZWxpbg== 30204
+cmVpYnQ= 30205
+INeU16Q= 30206
+IGNlbWV0ZXJ5 30207
+IGVzcGHDsW9s 30208
+b2xpbg== 30209
+0LvRjtC0 30210
+IGdyw6JjZQ== 30211
+YWxsZW4= 30212
+IFBoaWxvcw== 30213
+IEVyc3Q= 30214
+IOyDiA== 30215
+IFZpZA== 30216
+R2l2ZQ== 30217
+T0g= 30218
+zrzOvw== 30219
+IFBhcmU= 30220
+IG1ldGFib2xpc20= 30221
+IG1hcGxl 30222
+IGF4bGU= 30223
+IER5 30224
+IGtvbW1l 30225
+z47OvQ== 30226
+IGdyZWF0bmVzcw== 30227
+IHZlcmlmaWVk 30228
+IHNww6k= 30229
+IEZhaHJlbmhlaXQ= 30230
+IEJyZW4= 30231
+IENvbmZlZGVy 30232
+IGhpc3RvaXJl 30233
+IGVsaW1pbmF0aW5n 30234
+IEFkZGluZw== 30235
+IEFiaQ== 30236
+IGhvc3BpdGFsaXR5 30237
+dGlt 30238
+IGJvbml0bw== 30239
+IHBhcnRlcw== 30240
+INC00YDRg9Cz0LjRhQ== 30241
+IFNoYXk= 30242
+IFNlZA== 30243
+IHJlZ3JldHM= 30244
+0Y/QvNC4 30245
+IHRlbmFudHM= 30246
+IFBUUw== 30247
+IGRldmk= 30248
+IExhdGU= 30249
+dWV6 30250
+IHPDtnls 30251
+IOyerOuwjA== 30252
+IHRvZ2dsZQ== 30253
+IG1hc2tpbmc= 30254
+0LDQu9GM0L3QvtCz0L4= 30255
+IHBlcnPDtm4= 30256
+IGFtZXJpY2Fu 30257
+Zmlr 30258
+IFJHQg== 30259
+ZW5zb24= 30260
+IEtB 30261
+d3d3dw== 30262
+INGA0LXQsw== 30263
+bWV0aWNz 30264
+IGVkdWNhdG9y 30265
+cGFyaw== 30266
+0LXQu9GM0LfRjw== 30267
+YXJ1cw== 30268
+0YDQtdGC 30269
+IGZlaXRv 30270
+IGNob2ly 30271
+IGxhcmdv 30272
+IGVlbnM= 30273
+IHdhdHRz 30274
+IFNpbmdsZQ== 30275
+IHN1c2NlcHRpYmxl 30276
+aWNlcg== 30277
+INCy0LrQu9GO0Yc= 30278
+IHB1cw== 30279
+7ZmY 30280
+RW5n 30281
+IGZhbnRhcw== 30282
+IHNwZWNpZmljYXRpb24= 30283
+IGNvbmZyb250ZWQ= 30284
+IENvbHVtYnVz 30285
+0LjQstC10YI= 30286
+YXLEsW0= 30287
+IGNhZmZlaW5l 30288
+bXVuaXRpb24= 30289
+IG1pZ3JhbnRz 30290
+bGlkZQ== 30291
+aXRhdGlvbnM= 30292
+IEdlbWU= 30293
+4bqr 30294
+IHBsYW5uZXI= 30295
+IHN0aW11bGF0ZQ== 30296
+IGFwcm94aW0= 30297
+Y2V1 30298
+IE5vbQ== 30299
+IHZvZw== 30300
+INGA0LDRgdGC 30301
+IGVuc2XDsQ== 30302
+IHNlbGxlcnM= 30303
+IGd1dGVu 30304
+emQ= 30305
+Q2Fs 30306
+IGRlc2NyaXB0 30307
+IHJlY29uY2lsaWF0aW9u 30308
+emluaG8= 30309
+4bmHYQ== 30310
+YWN5ag== 30311
+IENPTA== 30312
+c2F3 30313
+IO2ZleyduA== 30314
+IHZhcml0 30315
+IHBhcnRuZXJpbmc= 30316
+IGRldGVudGlvbg== 30317
+IGJvbWJpbmc= 30318
+Y2xhcHBpbmc= 30319
+aWVuY2llcw== 30320
+b25kdQ== 30321
+QU1F 30322
+IOqwmeyKteuLiOuLpA== 30323
+Y8OtYQ== 30324
+INC/0L7RgdGC0L4= 30325
+IEFTTVI= 30326
+IGhvbWVwYWdl 30327
+IHNpw6g= 30328
+YW50aGE= 30329
+IFBvbGw= 30330
+IGlnZW4= 30331
+Y3ljaA== 30332
+IOqwkeyekOq4sA== 30333
+IGNvbnNpZGVyYWJseQ== 30334
+IEFyaXN0 30335
+IHdpdGhzdGFuZA== 30336
+IHF1YWxpdGF0aXZl 30337
+IEtyYWZ0 30338
+INGN0LvQtdC60YI= 30339
+IEJlYWQ= 30340
+0LXQutGC0LjQsg== 30341
+IGNydXNoaW5n 30342
+7LOQ 30343
+IG5hdnk= 30344
+2Yjaug== 30345
+c2hv 30346
+IG9haw== 30347
+aXBwZXJz 30348
+IHNvaWxz 30349
+IHBpZ21lbnQ= 30350
+IGV2aXRhcg== 30351
+IGZ1c2U= 30352
+IERhbGU= 30353
+OiI= 30354
+IGNvbXBsw6h0ZW1lbnQ= 30355
+IGtlbA== 30356
+4LmG 30357
+IHF1YXRyZQ== 30358
+IFVN 30359
+IOunkOs= 30360
+w61y 30361
+IGxlaXN1cmU= 30362
+IEhvdXNpbmc= 30363
+IGZvbGRz 30364
+ZXN0aW9u 30365
+QVJT 30366
+IG1hc2g= 30367
+dXJwb3Nl 30368
+IGFjY3VtdWxhdGVk 30369
+IFN0dWZm 30370
+IHRhcGVz 30371
+INGB0LjQu9GM0L3Qvg== 30372
+IExPVkU= 30373
+IDE5ODI= 30374
+IHNjYXJz 30375
+IGNhcGl0YWxpc3Q= 30376
+IE5lZA== 30377
+IHNvZnRlbg== 30378
+IG5vdGFibHk= 30379
+IGZvcmPDqW1lbnQ= 30380
+IFJhdW0= 30381
+INC90LXQvtCx0YXQvtC0 30382
+IHRyYWRlbWFyaw== 30383
+IGZlcnRpZw== 30384
+ID8h 30385
+IHJlaW5mb3JjZWQ= 30386
+IHJlY2hhcmdl 30387
+IFB1dHRpbmc= 30388
+IHZpbGxhaW5z 30389
+IGhhbmRpYw== 30390
+IGFkdmVydGlzZW1lbnQ= 30391
+2KrZig== 30392
+INGB0YPQvA== 30393
+IFJpbGV5 30394
+15XXkdc= 30395
+T3M= 30396
+2KfYsg== 30397
+Qm95 30398
+IHNxdWlzaA== 30399
+b2NrZXQ= 30400
+IHRlc3RpZnk= 30401
+INec157X 30402
+INC80LDRgdGB 30403
+bWFudWVs 30404
+IEFya2Fuc2Fz 30405
+aWZmZQ== 30406
+IGFuYWx5c3Rz 30407
+IERlYWY= 30408
+IGrDsw== 30409
+IGdyb2Nlcmllcw== 30410
+IFdoZWVs 30411
+INGA0LjRgQ== 30412
+IGPDsm4= 30413
+IENvYg== 30414
+IHByaXNvbnM= 30415
+w6h2ZQ== 30416
+IENhYmluZXQ= 30417
+IHBvc2Vk 30418
+IGd1ZXJyZQ== 30419
+IExsb3lk 30420
+IGNsZXJr 30421
+IGNyaXNlcw== 30422
+IFNobw== 30423
+IE9yZQ== 30424
+IEZvb3RiYWxs 30425
+IEFkdmlz 30426
+IFpoZW5n 30427
+6I0= 30428
+IEFNWQ== 30429
+IHVuZm9y 30430
+IG1vbmFzdGVy 30431
+IGNvbXBpbGU= 30432
+IGltbW9ydGFs 30433
+YXRhYmxl 30434
+IHBhcmFubw== 30435
+IHRpdmVy 30436
+IFN0ZXBo 30437
+IEZ1w58= 30438
+IGRpc2NvbnRpbg== 30439
+IHJpcGU= 30440
+IGhhY2tpbmc= 30441
+IHNpZW5kbw== 30442
+IHNlZ3Vybw== 30443
+YWx0cmVz 30444
+IGFuZGVyZXM= 30445
+IOumrOs= 30446
+IGV4cG9ydHM= 30447
+IHRhYmlp 30448
+IOq4sOuLpOs= 30449
+IGJvdGhlcmluZw== 30450
+IHBpY2tsZQ== 30451
+IEJSSUFO 30452
+IGFsdGFy 30453
+INC/0YDQuNCx 30454
+IHRyYW5zZmVycmluZw== 30455
+IFZvcnM= 30456
+INmH2Yg= 30457
+IFph 30458
+IEZyYW5jZXM= 30459
+IGJyb3dzZQ== 30460
+ZW1pdA== 30461
+IGNoZXdpbmc= 30462
+IEZyZWRkeQ== 30463
+IGVkaXRvcnM= 30464
+w6RsbGU= 30465
+IO2MgA== 30466
+IFNxdWU= 30467
+IEN1bHR1cmFs 30468
+YXdr 30469
+IFNhY2hl 30470
+IENhcmJvbg== 30471
+4bqvdA== 30472
+Rkw= 30473
+IE5HTw== 30474
+cGXFgg== 30475
+IFNvdQ== 30476
+IGh2b3I= 30477
+dW5pbnRlbGxpZ2libGU= 30478
+IOuylQ== 30479
+IMKw 30480
+aWlu 30481
+INei150= 30482
+IGRlcnJpw6hyZQ== 30483
+IGN6eW0= 30484
+IEFwb3N0 30485
+IHJlZ2FyZGVy 30486
+IGFncmFkZQ== 30487
+IENhbmR5 30488
+IG1hcmU= 30489
+IGludHJvZHVjZXM= 30490
+YmlyZHM= 30491
+IHVuaXF1ZWx5 30492
+IG11aw== 30493
+IGNvb2tlcg== 30494
+IGNyZXdz 30495
+IGplaXRv 30496
+RVJU 30497
+toTr 30498
+bmlzc2U= 30499
+IGVm 30500
+IGNhcnRl 30501
+IFlhaw== 30502
+IFBBVA== 30503
+0LjQvdC+ 30504
+Ym9ra2k= 30505
+IG1hdGVz 30506
+IGRpc3RpbnQ= 30507
+IOy9lOuhnOuCmA== 30508
+IHnEsWw= 30509
+IM66zqzOvQ== 30510
+IGNvbmZpZ3VyYXRpb25z 30511
+ZW5nYQ== 30512
+cmVjaHQ= 30513
+SGFwcHk= 30514
+aW52ZXN0 30515
+IHJlY29uc3RydWN0 30516
+INGN0YLQvtC80YM= 30517
+IG1vc3F1ZQ== 30518
+cmF1bQ== 30519
+IHZveWV6 30520
+IE5CQw== 30521
+IOyekOyLoA== 30522
+IHN0dXJkeQ== 30523
+INC60LDQvw== 30524
+IGFuc2No 30525
+YWxpZA== 30526
+IG1hc2lo 30527
+IFJFUA== 30528
+IOy9lOs= 30529
+IGRlZHVjdA== 30530
+IHNhbGly 30531
+d3VyZg== 30532
+aWxvdA== 30533
+IE11dHRlcg== 30534
+b2xkcw== 30535
+IEZFTUE= 30536
+IEJpYg== 30537
+IG5laWdoYm9yaW5n 30538
+IGJsaXNz 30539
+IO2YvA== 30540
+0LvQuNGB0Yw= 30541
+INGC0YDQtdCx 30542
+IGdyZW5hZGU= 30543
+IGVnYWw= 30544
+IGZpbmVseQ== 30545
+IHBldGFscw== 30546
+IGtlZXI= 30547
+IGNoeWJh 30548
+IHNraXBwaW5n 30549
+IHRoaXJ0ZWVu 30550
+IGdyYXZ5 30551
+IFNBVA== 30552
+NjE= 30553
+INC90L7Qsw== 30554
+IG1pbnM= 30555
+SVRF 30556
+IHNvemlhbA== 30557
+7ZWY66m07ISc 30558
+cnVrdHVy 30559
+INCy0L7Qt9C80L7Qtg== 30560
+INC+0L/Rj9GC0Yw= 30561
+IGFydGg= 30562
+IEN1YmFu 30563
+IHRyZWFzdXJlcw== 30564
+IGZlcnRpbGl6ZXI= 30565
+IGF3YWtlbmluZw== 30566
+IOuwseyLoA== 30567
+IHJhbGw= 30568
+IGRlcGljdA== 30569
+IFBhYmxv 30570
+IG5pbmV0ZWVu 30571
+IHdhdHQ= 30572
+IGVudGlyZXR5 30573
+S1M= 30574
+IFdvb2Rz 30575
+U2No 30576
+INqp2Yg= 30577
+IERyeQ== 30578
+dXZl 30579
+IHJlY29uc3RydWN0aW9u 30580
+IGFuYXRvbXk= 30581
+iOulvA== 30582
+IGJhYmE= 30583
+IGxpc3RlbmVy 30584
+IHNoYXJwZW4= 30585
+IFBlcnU= 30586
+INCy0YvQtw== 30587
+IHJlY3JlYXRpb24= 30588
+IGluaXRpYXRl 30589
+IGNhbG9y 30590
+IE5hag== 30591
+Z2Vl 30592
+IEZlZWxz 30593
+IFNuYXBjaGF0 30594
+IFRldA== 30595
+IE5lc3Q= 30596
+IERhZg== 30597
+IEZpbmlzaA== 30598
+INGC0LDQutC40Lw= 30599
+w7pj 30600
+aXplbnM= 30601
+IHNwaW5z 30602
+IGVtYnJ5 30603
+IHBhc3NhZ2Vz 30604
+IGNpZW50 30605
+IGp1c3RpZmljYXRpb24= 30606
+IG9sbWF6 30607
+IGZsb29kZWQ= 30608
+IGVtb2pp 30609
+IGVtYnJhY2luZw== 30610
+IGRpc2NhcmQ= 30611
+IEJhc2lj 30612
+YWdvZw== 30613
+IOychO2VtA== 30614
+IGFzeWx1bQ== 30615
+ZXJpbg== 30616
+IGZpbQ== 30617
+IG5pbmph 30618
+IGF1dG9tYXRl 30619
+IGFsbGVyZ2lj 30620
+w7/Dv8O/w78= 30621
+YW1hbQ== 30622
+INC80LDRgA== 30623
+IE9p 30624
+w6R1cw== 30625
+IGluZHVjdA== 30626
+IEJFTg== 30627
+IHrFgg== 30628
+IGthxbxkeQ== 30629
+IEFNUA== 30630
+bsSb 30631
+U3VyZQ== 30632
+IHF1aWw= 30633
+IGVzcGVj 30634
+cm9r 30635
+QlNDUkk= 30636
+IGxpZWJl 30637
+cHVz 30638
+YWNoc2Vu 30639
+IGNyaWNrZXQ= 30640
+64qQ 30641
+IEZyYW1l 30642
+ZWtrw7xy 30643
+YXJi 30644
+IHDFmQ== 30645
+0LjRgdGB 30646
+IHplZ2dlbg== 30647
+IGRvdWJsZXM= 30648
+IERyZQ== 30649
+dGVzdA== 30650
+aW5zcA== 30651
+Ym95cw== 30652
+IG3Do28= 30653
+IFZlcnNl 30654
+IG11c2N1bGFy 30655
+IE1BTEU= 30656
+IGR1bHU= 30657
+IG9jY2FzaW9uYWw= 30658
+TG8= 30659
+Y29ub21pYw== 30660
+IHZhaw== 30661
+IHJlbWVkeQ== 30662
+IOKZquKZquKZqg== 30663
+dmVt 30664
+IMO2bmVt 30665
+IGthcsWfxLE= 30666
+IFNoYXJw 30667
+aHVy 30668
+IOuwqeuylQ== 30669
+IGdyYW5kc29u 30670
+IGFrdGl2 30671
+IFRocm9uZXM= 30672
+IOyViOyXkA== 30673
+IHRvdHM= 30674
+IHN1YmQ= 30675
+IFBhdWxh 30676
+IGdyYXZlcw== 30677
+IEJyZW50 30678
+INC90LjQutGC0L4= 30679
+IHPDtno= 30680
+IGNyZWM= 30681
+IFZsYWRpbWly 30682
+INC/0L7QuQ== 30683
+ICIt 30684
+IHBzeQ== 30685
+YXRyaQ== 30686
+aWRhbg== 30687
+IGHDum4= 30688
+IHN0YW5kYXJkaXplZA== 30689
+7LmY6w== 30690
+INC60YDQvtCy 30691
+IFpodQ== 30692
+c29tZXRoaW5n 30693
+IDc1MA== 30694
+IG11amVyZXM= 30695
+IGFpdA== 30696
+YWd1 30697
+IGNvcnJlY3RlZA== 30698
+aWtrYQ== 30699
+ZWxlZA== 30700
+IENhcmVlcg== 30701
+b3d5bQ== 30702
+IHJvb21tYXRl 30703
+IGRlc2NlbmRhbnRz 30704
+IE5hcG9sZW9u 30705
+INCU0L4= 30706
+7ZaI7Ja07JqU 30707
+IGJ1bnVu 30708
+IE1pY2hh 30709
+IGRlc2NvYg== 30710
+UEk= 30711
+IHBhbGFicmE= 30712
+IHRyYWNrZWQ= 30713
+IGRlcGVuZGVuY2U= 30714
+IEJhcmFjaw== 30715
+IGZlcnRpbGl0eQ== 30716
+IFNvdXRod2VzdA== 30717
+IGluY29tcGxldGU= 30718
+IGNvbXVuaWM= 30719
+IGNvbXByaXM= 30720
+IFJlc3RhdXI= 30721
+IGFjcm9u 30722
+zrrOsQ== 30723
+IGFwcHJlbnRpY2Vz 30724
+IG11c3N0 30725
+IEFicg== 30726
+IHBlbnRydQ== 30727
+IENvbnNvcnQ= 30728
+IEF2ZWM= 30729
+IGR1bXBsaW5ncw== 30730
+TFI= 30731
+IHdzenlzdGtpZQ== 30732
+IHN3YW1w 30733
+0L3QtdCy 30734
+dWdnbGU= 30735
+IHdhdGVyY29sb3I= 30736
+IHByb3Rvbg== 30737
+IEVzcGHDsWE= 30738
+b2NraW5n 30739
+0L7QstCw0Ls= 30740
+IHRha2lt 30741
+VmVyeQ== 30742
+IGRlbWVudGlh 30743
+IMWfZXlp 30744
+SmFj 30745
+IE1hY0Jvb2s= 30746
+IExpdg== 30747
+ZmZpY2llbnRz 30748
+IEh1bnQ= 30749
+IG92ZXJsYXk= 30750
+IFNreXBl 30751
+cHVua3Q= 30752
+IGNvbmZpbmVk 30753
+IEFkcmlhbg== 30754
+2LHZgw== 30755
+IEplZXA= 30756
+IGVucXVhbnRv 30757
+IGFuZXN0 30758
+0L7RgtCy0LXRgg== 30759
+INC80LXQvdGM 30760
+IGlycmlnYXRpb24= 30761
+4buRbg== 30762
+IGVpZ2h0ZWVu 30763
+IFBvbg== 30764
+IHJlc2N1ZWQ= 30765
+IDE5ODM= 30766
+csO8 30767
+amFl 30768
+IEplb25n 30769
+IGFtYXppbmdseQ== 30770
+IEZEUA== 30771
+IGJhY2tzdGFnZQ== 30772
+Y3Vl 30773
+IM+Dz4TOt869 30774
+INin2YTYtQ== 30775
+IGxpdmVzdG9jaw== 30776
+IFdhcm5lcg== 30777
+IG1ham9ycw== 30778
+IGNvb3BlcmF0aXZl 30779
+IEJyYWR5 30780
+cmFpbmVk 30781
+cmllYg== 30782
+INeR157X 30783
+INC00L7QstC+0LvRjNC90L4= 30784
+IEZF 30785
+IGxlYWtlZA== 30786
+IE1lcmN1cnk= 30787
+IHBlcnN1YWRl 30788
+IHRyYW5zZm9ybWVy 30789
+IE5vcndlZw== 30790
+IOyXrOufrA== 30791
+IHpyb2JpxIc= 30792
+IGNhcmRpb3Zhc2N1bGFy 30793
+IENyYXNo 30794
+IGdvc3NpcA== 30795
+0LDRgdGC0Yw= 30796
+IOyqvQ== 30797
+IHN3ZXB0 30798
+IEhvcm4= 30799
+IEF0w6k= 30800
+IGJ1a2Fu 30801
+IEthdw== 30802
+S1k= 30803
+IFN0b3JpZXM= 30804
+R2FyeQ== 30805
+IGdhcmRlbmluZw== 30806
+IFF1aWNrbHk= 30807
+IEZhbGNvbg== 30808
+IG92YXQ= 30809
+Y8Sx 30810
+IENvbXBsZXQ= 30811
+IERhdGU= 30812
+INC/0YDQuNC8 30813
+IGzDpHVmdA== 30814
+IEF1ZHJleQ== 30815
+IFdlbnQ= 30816
+IHBlbMOtY3Vs 30817
+IGNhcnJpYWdl 30818
+IHVuYWNjZXB0YWJsZQ== 30819
+bnltaQ== 30820
+INGB0LvRi9GI 30821
+IHRlcnJl 30822
+dWVsbGVtZW50 30823
+RUVFRQ== 30824
+IHBoYXJtYWM= 30825
+aMO1ZXM= 30826
+IHppY2g= 30827
+IG1pZ3JhdGU= 30828
+IEZyeQ== 30829
+w7FhbmE= 30830
+IE11aXRv 30831
+RU9WRVI= 30832
+IGZvcnRyZXNz 30833
+IENvbXBhbg== 30834
+IEpTT04= 30835
+b3JkbnVuZw== 30836
+IHdhcnRv 30837
+IHVuZ2Vm 30838
+7IWU7ISc 30839
+INGA0L7Qug== 30840
+IHBhZGRsZQ== 30841
+SmFyZWQ= 30842
+IHN1Ym1pdHRpbmc= 30843
+IGxhdGNo 30844
+IGZ1Zw== 30845
+INC60L7RgQ== 30846
+IEVm 30847
+IGxhdW5jaGVz 30848
+IGZ0 30849
+b3RlY2hu 30850
+IHRyYXZlbGxlZA== 30851
+2KfZgQ== 30852
+IHByb2No 30853
+IGRlZGlt 30854
+ODM= 30855
+IHJlYm91bmQ= 30856
+IExV 30857
+cGF0aA== 30858
+INGB0L/RgNCw0LI= 30859
+IMO2bA== 30860
+IO2CpA== 30861
+IHByaXZhdA== 30862
+IHRyYWN0b3I= 30863
+IEF0dGVudGlvbg== 30864
+U2Vy 30865
+IGNvc2Vz 30866
+w6FyaWE= 30867
+cGFs 30868
+IOydgA== 30869
+IHN1Y2Nlc3Nvcg== 30870
+IGNvbm5lY3RvcnM= 30871
+INGD0YHRgtCw0L3QvtCy 30872
+IGdlbm9jaWRl 30873
+IHN1ZmZpY2llbnRseQ== 30874
+IEFpeMOy 30875
+IHN0YWJpbGl6ZQ== 30876
+IGNvbmdlc3Q= 30877
+IGNhcnZpbmc= 30878
+IHpvc3Q= 30879
+INCx0YvRgdGC0YDQvg== 30880
+IHNob3J0ZXN0 30881
+IGxpdmVs 30882
+IDg5 30883
+IGVyaw== 30884
+IHBvcnRyYWl0cw== 30885
+4KWA 30886
+6Jg= 30887
+Ym9hdA== 30888
+bGxhaA== 30889
+QU5D 30890
+IGVtcGlyaWNhbA== 30891
+IEVjaG8= 30892
+IE5lZGVybGFuZA== 30893
+TmV0 30894
+IGN1aWRhZG8= 30895
+IFJvbWE= 30896
+IGNhbGY= 30897
+IGdpYW50cw== 30898
+IEV4cGxvcmVy 30899
+IENvbGxlY3Q= 30900
+YWxpdGlvbg== 30901
+IERlc3Rpbnk= 30902
+IGF1c2dl 30903
+IEVkdQ== 30904
+IENsbw== 30905
+IGVhcnJpbmdz 30906
+IFRyYWNr 30907
+IFJPUw== 30908
+IEJlbGxl 30909
+IHB1ZWRh 30910
+IGRheXRpbWU= 30911
+IHN1cHBsaWVy 30912
+IFNW 30913
+IEV4aGFsZQ== 30914
+IGdhbGVyYQ== 30915
+Y291cnNl 30916
+IGNlbnRpbWV0ZXI= 30917
+IEJhc3Q= 30918
+bXVk 30919
+IHNhbmdhdA== 30920
+IFBoeXNpY2Fs 30921
+IHByaXZhdGVseQ== 30922
+IHRyYXRh 30923
+bHlubg== 30924
+aWxsaQ== 30925
+IOuplOydtO2BrOyXhQ== 30926
+IGNyeXN0YWxs 30927
+IHBvZHM= 30928
+4bqjbg== 30929
+aW5hdG9y 30930
+IFJlY29yZHM= 30931
+xJ9pbWl6 30932
+aXNzZW1lbnQ= 30933
+aGFyZQ== 30934
+aGFkb3c= 30935
+IERL 30936
+IOyVjOqzoA== 30937
+IHd5bg== 30938
+IHJlcXVlc3Rpbmc= 30939
+IERvbm5h 30940
+IOyXtOyLrO2eiA== 30941
+aW5lYQ== 30942
+IGV4ZXJ0 30943
+IER1bmNhbg== 30944
+INCy0LXRhw== 30945
+IEhhaA== 30946
+4KSC 30947
+IExpZg== 30948
+IEZpbmRpbmc= 30949
+IE5vdg== 30950
+INC30L3QsNC6 30951
+INC+0YQ= 30952
+IFF1w6g= 30953
+IHF1YXJ0ZXJiYWNr 30954
+INGE0LDQug== 30955
+IGJpcGFydGlzYW4= 30956
+xJ9pbg== 30957
+IG7DqWNlc3M= 30958
+IHJlZmVyZW5kdW0= 30959
+IGNvbXBpbGVy 30960
+IHByb2JhYmls 30961
+0LXQtNC4 30962
+IHRyYWRlcg== 30963
+IFJ1bQ== 30964
+Z2VtZQ== 30965
+IGRpbw== 30966
+IGLEmWR6aWVteQ== 30967
+IM+Azqw= 30968
+6r64 30969
+15XXmA== 30970
+IOCklQ== 30971
+INCx0LvQsNCz 30972
+IHNjYWxw 30973
+IFBhdXNl 30974
+IGNhcHRpb24= 30975
+IGVuZGFuZ2Vy 30976
+IGVubGFy 30977
+IHJvdHRlbg== 30978
+IHdhaA== 30979
+IGR6aQ== 30980
+IEluc3RhbGw= 30981
+QXk= 30982
+IGNyZWFy 30983
+0LXQvdGC0LA= 30984
+IHdlaWdoaW5n 30985
+IGJ1dHRlcmZsaWVz 30986
+IEdhc3Q= 30987
+aG9ybg== 30988
+d2Fyeg== 30989
+SUNFT1ZFUg== 30990
+INC90LDQudGC0Lg= 30991
+IGNvZWZmaWNpZW50cw== 30992
+IFNwZW5jZXI= 30993
+IEhpZ2hlcg== 30994
+IGNvd29yaw== 30995
+INC60L7RgtC+0YDQvtC1 30996
+IG1vbml0 30997
+IGR5c2Z1bmN0aW9u 30998
+INGB0YLQsNC90L7Qsg== 30999
+IHRvdXJuYW1lbnRz 31000
+IG95c3Rlcg== 31001
+Qk4= 31002
+IHRydWQ= 31003
+c2xvdw== 31004
+IFBlbm55 31005
+IE9keXM= 31006
+w6Zy 31007
+IGZvdQ== 31008
+IGVuam95bWVudA== 31009
+0LDRgtGL 31010
+IHd5Z2zEhWRh 31011
+0LDQu9GM0L3QsNGP 31012
+IFByb3RlY3Q= 31013
+IG1veQ== 31014
+IGNsYXc= 31015
+IHN1c3BpY2lvbg== 31016
+IHNhY3JpZmljZWQ= 31017
+IGdvc3Rv 31018
+Qmln 31019
+IGFnZ3Jlc3NpdmVseQ== 31020
+IHZvcm5l 31021
+IGJsYW1lZA== 31022
+IFNlaHI= 31023
+16TXqA== 31024
+Y2l0bw== 31025
+IHNlYWxz 31026
+IG11amVy 31027
+IFdlaXJk 31028
+IGZvcmVucw== 31029
+IGNvbnRyaWJ1dGVz 31030
+ZXN0cmE= 31031
+IHBvZw== 31032
+TE9M 31033
+IGhhY2VybG8= 31034
+0L7RgtGM 31035
+ZmljdGlvbg== 31036
+Nzk= 31037
+zrvOvw== 31038
+INGC0L7QsQ== 31039
+IEdT 31040
+IENsYXJh 31041
+aXRleg== 31042
+IGFkdm9jYXRpbmc= 31043
+IO2UhOs= 31044
+c3VuZw== 31045
+IHZlcnRpY2Vz 31046
+IG5hdmlnYXRpbmc= 31047
+IGV1cm9ww6k= 31048
+IHNsb3dlZA== 31049
+IGZvcmVncm91bmQ= 31050
+IEluZHVzdHJpYWw= 31051
+IGFkb3Jl 31052
+7Iut 31053
+IGNyw6llcg== 31054
+Y2huaXR0 31055
+IHVuYXdhcmU= 31056
+IGN1cmx5 31057
+ZW50YXI= 31058
+IGxlcg== 31059
+IHByb2hpYml0ZWQ= 31060
+IEhlcm9lcw== 31061
+IFJlZWQ= 31062
+dWNh 31063
+IHNtb2s= 31064
+IGt1bm5h 31065
+emVpdGln 31066
+aW1tZW4= 31067
+IEx1bg== 31068
+INCw0LHRgdC+0LvRjtGC 31069
+IGRlZ2xp 31070
+IHZpbGxhZ2Vycw== 31071
+IHByZXNldA== 31072
+emVwdA== 31073
+dWRz 31074
+IGVtaXQ= 31075
+IOuJ 31076
+64qU7KeA 31077
+0L3QsNC60L4= 31078
+IG9zw7Ni 31079
+IDE5Njk= 31080
+INCQ0YA= 31081
+IG1hbmNobWFs 31082
+IEJyb2Nr 31083
+IG1hbnRyYQ== 31084
+IFdJTA== 31085
+YmFjaA== 31086
+aW7DpA== 31087
+ZWxhcw== 31088
+a2Vsbg== 31089
+IGRpc2NpcGxl 31090
+IHF1YWxj 31091
+IGRlaHlk 31092
+7J20652864qU 31093
+QWY= 31094
+7ISx7J20 31095
+Unlhbg== 31096
+IHB1cHBldA== 31097
+INC00YDRg9Cz0LjQtQ== 31098
+IHJ1ZA== 31099
+IHBlbmRpbmc= 31100
+UGx1cw== 31101
+IOyViuydhA== 31102
+IGLhu4s= 31103
+IFNlZ2E= 31104
+w6dl 31105
+IHByb2dyYW1tZXI= 31106
+Ymxp 31107
+IHVubA== 31108
+IGVuc2xhdmVk 31109
+IHNvY2nDqXTDqQ== 31110
+xIFo 31111
+IGluaGVyaXRhbmNl 31112
+IEJhbmds 31113
+ZXJtYWlk 31114
+IHByYWN0aXRpb25lcg== 31115
+IFN0YWxpbg== 31116
+IFVzZXI= 31117
+Y2libGU= 31118
+IGNhcmRpYWM= 31119
+IEtvcmVhbnM= 31120
+IGR1bXBlZA== 31121
+INeU15nXlA== 31122
+w6Fpcw== 31123
+IGh5ZHJhdWxpYw== 31124
+b3VidGVkbHk= 31125
+IFBpdA== 31126
+IHBpY25pYw== 31127
+IGJlaMO2dmVy 31128
+INGB0LzQvtCz 31129
+IGJyYWtpbmc= 31130
+dXRhcg== 31131
+IOyEuOs= 31132
+dWJs 31133
+IMO8eg== 31134
+IG1hamVzdHk= 31135
+IGJlcnM= 31136
+dXRhYmxl 31137
+IGhvdHRlcg== 31138
+24zZhg== 31139
+IGJpYXNlcw== 31140
+IHN1YmplY3RlZA== 31141
+IG5hdWdodHk= 31142
+IGNpcmN1cw== 31143
+IEltbWVkaQ== 31144
+IFN0ZWZhbg== 31145
+IFRyaXBsZQ== 31146
+ZW5r 31147
+IHdpdA== 31148
+IHJlY3ljbGU= 31149
+ZW1pZQ== 31150
+ZGF0ZWQ= 31151
+IHVubG9hZA== 31152
+IHBvcHVsYQ== 31153
+Y2hpbg== 31154
+IHlpZWxkcw== 31155
+IGVuZ2xpc2g= 31156
+IEJvbm5pZQ== 31157
+IHNwaWRlcnM= 31158
+w4E= 31159
+IGVyb3Npb24= 31160
+IE5JQ0s= 31161
+0LjRj9GF 31162
+IGltcGFydA== 31163
+INC60L3QuA== 31164
+IHJlc29sdXRpb25z 31165
+IGxpdGhpdW0= 31166
+IGNvbnZlcmdlbmNl 31167
+IFRhcmE= 31168
+INC00LLQtQ== 31169
+dGhz 31170
+IENpbmR5 31171
+IERJRQ== 31172
+IGFzc3VyYW5jZQ== 31173
+INC+0L/QuNGB 31174
+IGJ1Y2tldHM= 31175
+IGN1ZXM= 31176
+IFF1aWV0 31177
+IHNpbWlsYXJpdHk= 31178
+IGZvdW5kYXRpb25hbA== 31179
+IE1pbmlzdA== 31180
+IHBpYW4= 31181
+IGNlbnRy 31182
+IG51bWI= 31183
+IG1vbmtz 31184
+dWpvdXJk 31185
+ZW56aWU= 31186
+IHNrYXRlYm9hcmQ= 31187
+IGRsYXRlZ28= 31188
+INGB0L7Rgg== 31189
+IEFF 31190
+IG1hc3RlcnBpZWNl 31191
+IFNvbG9tb24= 31192
+IFJlZGRpdA== 31193
+IHJpb3Q= 31194
+YWJs 31195
+IEpheno= 31196
+IGVsZWN0cm9tYWduZXRpYw== 31197
+IGluc2VjdXJl 31198
+IENvbXBldA== 31199
+Z2VyaWVz 31200
+0L7QsdC+0LQ= 31201
+oNeV 31202
+8J+S 31203
+IHNlbmF0b3Jz 31204
+IEJyaXNiYW5l 31205
+IEFsYg== 31206
+dXR0ZXJpbmc= 31207
+IEFsbG93 31208
+emVybw== 31209
+IHBhaQ== 31210
+INCQ0LvQtdC60YE= 31211
+IERpc3BsYXk= 31212
+IEJsYWRl 31213
+IEFwcHM= 31214
+IHDDpA== 31215
+INC00LXRgdGP 31216
+IHF1ZWxsYQ== 31217
+IEdhbw== 31218
+0LXQvdC90YvRhQ== 31219
+IHNwb2lsZXJz 31220
+IGdhbGxvbnM= 31221
+INmE2Yo= 31222
+IFppb24= 31223
+b25pZQ== 31224
+cmFndA== 31225
+IENoYW5k 31226
+IOuzkQ== 31227
+IGJsdW50 31228
+IHVzdQ== 31229
+IEthZA== 31230
+cmFrdA== 31231
+IGNpbmVtYXRpYw== 31232
+IGFtbXVuaXRpb24= 31233
+cmVuZQ== 31234
+IGZvdXJ0ZWVu 31235
+IENhcm4= 31236
+Y3JpdA== 31237
+IHRlbnVyZQ== 31238
+dnU= 31239
+IHByaW5jaXBhbG1lbnRl 31240
+IGFsbGVlbg== 31241
+IGtvbXBsZXR0 31242
+IGTDvG55 31243
+SmFtZXM= 31244
+IHJlY2VwdG9y 31245
+IG9uZXNlbGY= 31246
+Z3VydQ== 31247
+IG1lcmNoYW50 31248
+bGluZXNz 31249
+IG92ZXJsb29rZWQ= 31250
+IGhhcm1vbmlj 31251
+aWVzbw== 31252
+15XXng== 31253
+Y29sbQ== 31254
+INC/0YDQvtC10LrRgg== 31255
+IEFkYQ== 31256
+2KfYsw== 31257
+VGlt 31258
+IHJlY3VycmluZw== 31259
+IHByb2NlZWRz 31260
+IFBhcnRpY3VsYXJseQ== 31261
+IERvd25sb2Fk 31262
+ZXRyaWNhbA== 31263
+IG1hdHJpY2Vz 31264
+IHByb3llY3Rv 31265
+YW5jaWVz 31266
+IFVobQ== 31267
+IGNhdmVz 31268
+IOyWtOugpA== 31269
+IExlYWY= 31270
+INC+0LHRi9GH 31271
+IOydtOycoA== 31272
+RXVyb3Bl 31273
+IHTEhQ== 31274
+IHB1bHM= 31275
+IHRha2llZ28= 31276
+0J3QtQ== 31277
+R1U= 31278
+IGZvcnM= 31279
+z4HOsw== 31280
+IGZvdG9z 31281
+ICkp 31282
+IOuppOs= 31283
+IGFxdWlsbw== 31284
+IEt1cmQ= 31285
+77iP 31286
+cHRpYw== 31287
+IERvcnQ= 31288
+IG1pc2VyeQ== 31289
+YXVzbw== 31290
+Y2h1Y2tsaW5n 31291
+IFJpZGdl 31292
+IO2WiOyKteuLiOuLpA== 31293
+ICoqKg== 31294
+IEhtbW0= 31295
+IGdlb2dyYXBoaWM= 31296
+IGFueXM= 31297
+IHRhbHZleg== 31298
+IHNrZWxldA== 31299
+IHNpZ25hdHVyZXM= 31300
+IGxpdGVycw== 31301
+kOuptA== 31302
+INGB0LLQvtC10LPQvg== 31303
+IHNraWluZw== 31304
+INCc0L7RgQ== 31305
+IGFkb3B0aW5n 31306
+IGhhZnQ= 31307
+IHN5bW1ldHJpYw== 31308
+IExpcXU= 31309
+IHRoeXJvaWQ= 31310
+IG1pc2lu 31311
+bHVkZQ== 31312
+IGh1bGw= 31313
+IFhE 31314
+IEd1c3Q= 31315
+emVpY2g= 31316
+IHZpYnJhdGlvbnM= 31317
+IGVzZW1w 31318
+INCy0YHRjg== 31319
+IFF1ZW0= 31320
+IMO8YnJpZw== 31321
+IFNrZQ== 31322
+IEx5bmNo 31323
+cm9vbXM= 31324
+YXJ0ZXQ= 31325
+ZmVzdA== 31326
+IGZyw7xoZXI= 31327
+IGx1cmU= 31328
+IOyVjOyVhA== 31329
+IFdJTg== 31330
+IFJZQU4= 31331
+INC60L7RgtC+0YDRg9GO 31332
+IEthc2g= 31333
+INeU154= 31334
+IHNhZmVn 31335
+IEhhbGxlbHVqYWg= 31336
+INC00LLRg9GF 31337
+IHN0YXBsZQ== 31338
+IHNlZGltZW50 31339
+IEFjdHM= 31340
+IGJsYW1pbmc= 31341
+IG1haW5sYW5k 31342
+IHNwb3J0aW5n 31343
+IGRlY29yYXRpb25z 31344
+IGV4ZWN1dGluZw== 31345
+IHBhcmFu 31346
+IERvbGxhcg== 31347
+IHByb2plY3Rpb25z 31348
+IGNvbW1pc3Npb25lZA== 31349
+IGJvdXI= 31350
+w7Zt 31351
+IHN0ZWFtZWQ= 31352
+IOutmA== 31353
+IHBldHJvbA== 31354
+IGNlbHVsYXI= 31355
+IEh1bmdhcnk= 31356
+IHJlbnRlZA== 31357
+INCy0LDRgNC4 31358
+YmJpZQ== 31359
+IHPDqWN1cg== 31360
+w7xsbA== 31361
+IHN3aW5ncw== 31362
+YmV0d2Vlbg== 31363
+INC40YI= 31364
+ZXN0cm8= 31365
+IG5pZW1hbmQ= 31366
+IOyCvA== 31367
+IFBhcmRvbg== 31368
+ZXNzZXM= 31369
+IE1JRA== 31370
+IGNlbnRyYWxpemVk 31371
+IEFsaWVu 31372
+Y3Vsb3M= 31373
+IGNyaXNl 31374
+IGNsYXNzZQ== 31375
+YmVpdGV0 31376
+acSfaQ== 31377
+IHdoYWxlcw== 31378
+IHBlcmltZXRlcg== 31379
+IHR5aW5n 31380
+IHN0cm9ueQ== 31381
+IGxpa2V3aXNl 31382
+IFB1bmNo 31383
+RGE= 31384
+IEJhcHRpc3Q= 31385
+IHNvcnRpbmc= 31386
+IGl2 31387
+IO2VqQ== 31388
+IHJlaGFi 31389
+IGV0YQ== 31390
+cml2ZXI= 31391
+IHNhaQ== 31392
+b2R1cw== 31393
+IGVzc2F5ZXI= 31394
+IHR1cnRsZXM= 31395
+IEhhenJhdA== 31396
+IGZhYnJpY3M= 31397
+IGNhdml0eQ== 31398
+IHBvbmlld2HFvA== 31399
+IHNjaGxlY2h0 31400
+IHNhbHNh 31401
+xZ9la2vDvHI= 31402
+IHNlYXRpbmc= 31403
+IGVjb25vbWlzdHM= 31404
+IG1hbmc= 31405
+IHNlZ3VpbnRl 31406
+IHJhbmc= 31407
+IHJhdGlvcw== 31408
+IGNvbnN0ZWxs 31409
+IGxvbmd0ZW1wcw== 31410
+dWF0aW5n 31411
+IHNwb2lsZWQ= 31412
+IHJlY2lwaWVudHM= 31413
+IHNuaXBlcg== 31414
+7Iq164uI6rmM 31415
+IHdw 31416
+IExJTktF 31417
+IGZsYXJl 31418
+IEFkcmk= 31419
+w7Fhcw== 31420
+IGJhY2ts 31421
+bcOkw58= 31422
+IEJlbmQ= 31423
+IHdvcmtsb2Fkcw== 31424
+INGB0YPQvw== 31425
+IDE5NzU= 31426
+0LjQvNGB0Y8= 31427
+0LDQvdC1 31428
+INC80L7QvQ== 31429
+IGFzcGlyYXRpb25z 31430
+IEFlcg== 31431
+INCz0L7QstC+0YDQuNGC0Yw= 31432
+IFFpYW4= 31433
+IGNvbXByb21pc2Vk 31434
+IHlvbGs= 31435
+0LvQsNGB0YI= 31436
+IGhlbWVu 31437
+cm92ZQ== 31438
+ZGVucw== 31439
+INC60L7QvNC80LXQvdGC 31440
+IC0tLQ== 31441
+IGZsdW9yZXM= 31442
+0L3QvtGB 31443
+IExpdmVycG9vbA== 31444
+INGB0L7QsdC+0Lk= 31445
+IFp3ZQ== 31446
+IGx1bWlu 31447
+IE9H 31448
+4bg= 31449
+aG9sbQ== 31450
+cHJvZml0cw== 31451
+U04= 31452
+IHByb3BvcnRpb25z 31453
+IG1pY2E= 31454
+IEJvaA== 31455
+IEF0bGFz 31456
+IHVuc3VyZQ== 31457
+IHRvdXJpbmc= 31458
+IG5pZWQ= 31459
+IHTEmQ== 31460
+IGltcGVyYXRpdmU= 31461
+IGRlbWVr 31462
+IFNoZXJpZmY= 31463
+cmFuY2U= 31464
+IGhvbWVsYW5k 31465
+IEhhaWw= 31466
+IEdhbno= 31467
+eW1t 31468
+TW9u 31469
+dmlkYQ== 31470
+IGRlc2Fycm9sbA== 31471
+IGludHJpZ3Vpbmc= 31472
+IEh1Z28= 31473
+6aw= 31474
+0LDRhg== 31475
+IFdpxJlj 31476
+YXR0ZWQ= 31477
+IOyVhOuLiOqzoA== 31478
+IFZhcmk= 31479
+w6Fk 31480
+IHN1cnJlYWw= 31481
+IGRpc3Bhcml0aWVz 31482
+IG3Dsw== 31483
+dWxsZW4= 31484
+IOyeiOuLpOqzoA== 31485
+INC/0L7QttCw0LvRg9C50YHRgtCw 31486
+IG1haW5z 31487
+IGVqZWN0 31488
+IG1ldGhhbmU= 31489
+IG1hcmdpbmFsaXplZA== 31490
+IGNoaWxsaQ== 31491
+csOocw== 31492
+IHllbQ== 31493
+IENodW4= 31494
+IGRlYnRz 31495
+IGRvd25sb2FkaW5n 31496
+IEF0aGVucw== 31497
+aXNpZXJ1bmc= 31498
+cnlu 31499
+IHRla24= 31500
+IFF1aW5kaQ== 31501
+IHRhcmFm 31502
+IGjDqQ== 31503
+IGNvbnNjaW91c2x5 31504
+IGZpeGVz 31505
+dWNrbGU= 31506
+bWF5xLFu 31507
+IGZyZWk= 31508
+IHNwYQ== 31509
+IOynhO2WiQ== 31510
+INin2YTYsA== 31511
+INGD0Lo= 31512
+bGV0dA== 31513
+IG9sbXXFnw== 31514
+IGNoZWVzeQ== 31515
+4Liy4LiB 31516
+bmFpcmU= 31517
+IHdpZGVu 31518
+IGxpZW4= 31519
+IGVzY2FwaW5n 31520
+aWdncw== 31521
+IEJsaWNr 31522
+Y8SF 31523
+IOyEnOs= 31524
+INeU16E= 31525
+INCy0L/QtdGA 31526
+b3Bob25l 31527
+aWVsbA== 31528
+IFNVQlNDUkk= 31529
+IGxpb25z 31530
+IOq3uOqygw== 31531
+IGluc3BpcmVz 31532
+IGd1YXJhbnRlZXM= 31533
+IGNvbWXDp2E= 31534
+IEdyb3dpbmc= 31535
+IG5lZ2xpZw== 31536
+IEZyYW5rZg== 31537
+IGdlZ2ViZW4= 31538
+IMSR4bqndQ== 31539
+IGVuZGxpY2g= 31540
+IOyNqA== 31541
+IFRU 31542
+IExpdGg= 31543
+z4DOsQ== 31544
+YXN0ZXJu 31545
+IEF6ZXI= 31546
+IGx1bmFy 31547
+aGlj 31548
+INC90LDRgNC+0LQ= 31549
+IG5lbmh1bQ== 31550
+IFNhbHZhZG9y 31551
+IFByb2dyZXNz 31552
+IHByaXZpbGVnZXM= 31553
+IOuPmeyViA== 31554
+IGFudGFnb24= 31555
+IEltcGY= 31556
+IGRlc2N1Yg== 31557
+IExlaQ== 31558
+IOyDiOuhnA== 31559
+0YfQtQ== 31560
+IGTDs2xhcmVz 31561
+IE1lZ2hhbg== 31562
+IFdpcmU= 31563
+dG9v 31564
+YXlpbmc= 31565
+dXNj 31566
+IHR1ZA== 31567
+IGFwcGVhbHM= 31568
+ZWR1Yw== 31569
+IHBhbmU= 31570
+IGpp 31571
+IGRlY2tz 31572
+IEFsdGVy 31573
+7ISk 31574
+IHByb2R1Y3Rpb25z 31575
+IFdJTExJQU0= 31576
+IGltcGxpZWQ= 31577
+IGZ1bGZpbGxtZW50 31578
+IEFhaA== 31579
+IHNhamE= 31580
+eHVz 31581
+IM6azrHOuQ== 31582
+w6Bz 31583
+dWNjaA== 31584
+0L7QutC+ 31585
+IERpc2NvcmQ= 31586
+IFNZ 31587
+anNr 31588
+IFdhbGxhY2U= 31589
+dW5jdGlvbg== 31590
+RGFuaWVs 31591
+IGvDtnQ= 31592
+aWphaA== 31593
+IG1hcmNoZQ== 31594
+IGRpc2dy 31595
+IG11bmdraW4= 31596
+IGFsbWE= 31597
+s7U= 31598
+IGV4dGVuc2l2ZWx5 31599
+IEZsb3Jlbg== 31600
+IEFsbGlzb24= 31601
+2YrZhQ== 31602
+IGp1dmVu 31603
+IFJlbmFpc3NhbmNl 31604
+IGZ1bmRyYWlzaW5n 31605
+IENoYW9z 31606
+IHBhcmFseQ== 31607
+IG5hcnJhdG9y 31608
+IGVjb3N5c3RlbXM= 31609
+QXNo 31610
+IG1pdGlnYXRpb24= 31611
+IEF1am91cmQ= 31612
+IElkZWU= 31613
+ISw= 31614
+IMK9 31615
+IGxhbmRsb3Jk 31616
+IGRlZmVjdHM= 31617
+IGFjcmU= 31618
+dWxzaXZl 31619
+IGFsZ2Fl 31620
+cGVr 31621
+IGVtYmE= 31622
+IFJvYw== 31623
+a3NvbQ== 31624
+w6RjaGU= 31625
+IGxldWs= 31626
+IGxldmVyYWdpbmc= 31627
+IOq3uOugh+yngA== 31628
+IFBhbG0= 31629
+IMOkdmVu 31630
+IGxpcw== 31631
+IEluc3A= 31632
+IFJpdGE= 31633
+IEFiYg== 31634
+aXRobQ== 31635
+IHN1cGVydmlzaW9u 31636
+IHJldmlzaXQ= 31637
+IHBpxJk= 31638
+IGV1aA== 31639
+IGZhZGVz 31640
+IG1vdHRv 31641
+0LXQt9C2 31642
+IFNoaW0= 31643
+IHJlbGV2YW5jZQ== 31644
+IG9v 31645
+IG9zdGF0 31646
+bmljYQ== 31647
+IGNob2l4 31648
+IEZhY3VsdHk= 31649
+IOykkeyXkA== 31650
+IEFib3Zl 31651
+INC90LXQsdC+0LvRjNGI 31652
+IHNlcXVlbmNpbmc= 31653
+IG51dHJpZW50 31654
+IGNvbnF1ZXJlZA== 31655
+IGRpZ2VzdGl2ZQ== 31656
+IGJhY2tkcm9w 31657
+IExvcmk= 31658
+YWlsYWJsZQ== 31659
+R2FtZQ== 31660
+IG5lZ2xlY3RlZA== 31661
+b21vcnBo 31662
+aWxsYWg= 31663
+IGtuZQ== 31664
+IHNpaXTDpA== 31665
+IHdvcmtzcGFjZQ== 31666
+IFZlbmljZQ== 31667
+IEtuZQ== 31668
+0YnQvg== 31669
+hYA= 31670
+IEhhc3M= 31671
+IHZpdGE= 31672
+nbzrqbQ= 31673
+IGxheXM= 31674
+w6puY2lhcw== 31675
+w6lyaWNh 31676
+IExs 31677
+IENvY2E= 31678
+IFdIWQ== 31679
+IHJvdXRpbmc= 31680
+IHBlcm1pc3Npb25z 31681
+IGRpbmdz 31682
+cHJlbmQ= 31683
+cHJvZ3JhbQ== 31684
+IGNyb2NvZA== 31685
+YnJhbA== 31686
+QUFBQUFBQUE= 31687
+YWdpdA== 31688
+IE7DpA== 31689
+IGdla29tbWVu 31690
+YXR0ZW4= 31691
+IHJlZmVyZW5jZWQ= 31692
+IHBhaXJpbmc= 31693
+IFBhcnRuZXI= 31694
+IENvcm9uYXZpcnVz 31695
+0ZbRgQ== 31696
+INeU15M= 31697
+IGVzcGVjw61maWM= 31698
+YXJzaQ== 31699
+cXVlbGxl 31700
+IHNwb250YW5lb3Vz 31701
+IOqyg+ydhA== 31702
+INCf0L7RgdC70LU= 31703
+INin2YTYrw== 31704
+IFNob3V0 31705
+INC90LDQuw== 31706
+IGRpc2d1aXNl 31707
+IEpvcmQ= 31708
+IHdlZQ== 31709
+IG1pZWpzYw== 31710
+IHNlcnVt 31711
+IHBsYWlzaXI= 31712
+IGNyZWRpYmxl 31713
+IGLDpQ== 31714
+IEFK 31715
+bWFyZXM= 31716
+IHJvZHM= 31717
+IGVyYW4= 31718
+IHDDpMOk 31719
+IFVB 31720
+IFVua25vd24= 31721
+INmE2YU= 31722
+IFJhYmJp 31723
+IGxhYXQ= 31724
+IGhhaXJzdHlsZQ== 31725
+INi6 31726
+IGNhY2g= 31727
+IFdyaXRpbmc= 31728
+0L7Rh9C60Lg= 31729
+YWJhZA== 31730
+IHN0cmFpZ2h0ZW4= 31731
+LS0i 31732
+d2lmZQ== 31733
+IGhvdHRlc3Q= 31734
+IHB1bnlh 31735
+IEZhc2hpb24= 31736
+Z3JpZmY= 31737
+IFFS 31738
+b3RjaA== 31739
+INCc0L7QttC10YI= 31740
+Q2xvdWQ= 31741
+IFN0cmlrZQ== 31742
+IEhlaW4= 31743
+IGxlaQ== 31744
+IEZsb3c= 31745
+d2Vncw== 31746
+IGhhYnI= 31747
+bmFobWU= 31748
+zIE= 31749
+IHBsZWFzaW5n 31750
+b3BwaW5n 31751
+IOq1rOuPhQ== 31752
+IGRyYW4= 31753
+IGJhbmdz 31754
+IDc5 31755
+IHNrZXQ= 31756
+IGNhdmFs 31757
+IE1hY3Jvbg== 31758
+IHdlaWdodGVk 31759
+IG11dGVk 31760
+IG51ZXN0cmFz 31761
+RUVQ 31762
+IG1hdGhlbWF0aWM= 31763
+IE1SSQ== 31764
+YWd1cw== 31765
+IHRoZXJhcGllcw== 31766
+zrjOtQ== 31767
+IHVucGw= 31768
+IGNvbW1lbmNlcg== 31769
+ZnVsbA== 31770
+IHRvd2Vscw== 31771
+IHBydWU= 31772
+IGxpY2Vuc2Vz 31773
+15vXldec 31774
+INCf0L7Rh9C10LzRgw== 31775
+IHBvaW50bGVzcw== 31776
+Qnll 31777
+IGVsaWdpYmlsaXR5 31778
+IHNjcmFwZQ== 31779
+IGFidXNpdmU= 31780
+IE1hbnQ= 31781
+IGpldW5lcw== 31782
+dGFs 31783
+IFByaW5jaXA= 31784
+IE9ydGhvZG94 31785
+IG1lbG9k 31786
+INC80LDRgtC10YDQuA== 31787
+IHByb3NlY3V0b3I= 31788
+IG9waW9pZA== 31789
+INGD0LLQtdGA 31790
+IEJlZW4= 31791
+IOygkeyihQ== 31792
+IGR5bmFzdHk= 31793
+IGFqdWRh 31794
+IGVudHJlZw== 31795
+IHdlaWdoZWQ= 31796
+IGV1cmU= 31797
+IEJlbQ== 31798
+IGFibm9ybWFs 31799
+ODI= 31800
+IEpS 31801
+IEFrdA== 31802
+IEJyaQ== 31803
+w7p0 31804
+IHN0YWdu 31805
+ISo= 31806
+IHdlZ2Vu 31807
+IGxlYWtpbmc= 31808
+IFdvcmRz 31809
+IE1hdQ== 31810
+IHZ1ZQ== 31811
+IExpYW0= 31812
+0LDQvdC40LXQvA== 31813
+IGNsaW5pY2lhbnM= 31814
+IFB1bXA= 31815
+IGbDtnJzdA== 31816
+Py4uLg== 31817
+IGF1dG9tb3RpdmU= 31818
+IE93ZW4= 31819
+enVzYWdlbg== 31820
+IEh1bmRyZWQ= 31821
+IGRlY2VudHJhbGl6ZWQ= 31822
+IGJ1bGJz 31823
+INec15s= 31824
+IHByb3ZpbmNlcw== 31825
+IE1pbGFu 31826
+ODE= 31827
+a2Fz 31828
+IOuTow== 31829
+IGZvcsOnYQ== 31830
+IHJpZ2h0bHk= 31831
+csSF 31832
+IHZlbnVlcw== 31833
+IHdhaQ== 31834
+IHByZWRpY3Rpbmc= 31835
+IFdpRmk= 31836
+IOq2geq4iA== 31837
+2LHZiA== 31838
+INeU15Y= 31839
+Y2VudHVyeQ== 31840
+IGdyYWR1YWw= 31841
+IFByb2JsZW1l 31842
+IOyXhQ== 31843
+IGNvcGluZw== 31844
+IEJydXM= 31845
+IHBlYW51dHM= 31846
+aXJ0c2NoYWZ0 31847
+INC30LDQuw== 31848
+IFRyb3k= 31849
+IHNwZXJt 31850
+IE1pdGFy 31851
+IFTDvHJraXll 31852
+Z3JhbmQ= 31853
+pq0= 31854
+INee16E= 31855
+IHBhbnM= 31856
+IEtub3dsZWRnZQ== 31857
+YmVybHk= 31858
+INCV0LPQvg== 31859
+IGRhbmNlZA== 31860
+IEZyb3N0 31861
+IEJ1cmc= 31862
+IGJpdGluZw== 31863
+7KCV7J2E 31864
+bWVhbA== 31865
+IGhlcm9pYw== 31866
+IG1vdGhlcmJvYXJk 31867
+IExpY2h0 31868
+bGxhbg== 31869
+0LDQudC9 31870
+INGA0Y/QtA== 31871
+IOC5gOC4 31872
+b25lbg== 31873
+aXJpZQ== 31874
+QXJ0 31875
+cmFuZw== 31876
+zr3Otw== 31877
+IG5ld2Jvcm4= 31878
+IGFtaXM= 31879
+INin2YjYsQ== 31880
+IHNvcGhvbQ== 31881
+IENhcmVmdWw= 31882
+IHByb3NwZWN0cw== 31883
+ZW5zZW4= 31884
+IHRocmlsbA== 31885
+IFZp4buHdA== 31886
+QWRhbQ== 31887
+cml0aW9u 31888
+ZW50cmlj 31889
+dWRlbg== 31890
+IGNlcnRpZmljYXRlcw== 31891
+IGFzaGVz 31892
+cGxheWluZw== 31893
+IHNhZGVjZQ== 31894
+IG9zdA== 31895
+IGFpcnBsYW5lcw== 31896
+0YDQvtC6 31897
+b25lcg== 31898
+IG1hZ25lc2l1bQ== 31899
+IGdvZGRhbW4= 31900
+IDE5NzI= 31901
+IFNjaHVsZQ== 31902
+IHRlbWF0 31903
+IHBhcnRvdXQ= 31904
+4K+C 31905
+IGludmU= 31906
+IFNjaWVudGlzdHM= 31907
+IEh1ZHNvbg== 31908
+d2lubmluZw== 31909
+Y2Vrc2lu 31910
+IGNvbmdyZXNzaW9uYWw= 31911
+b3J1 31912
+IHJvcGVz 31913
+0LLQtdC0 31914
+IG1hZHJl 31915
+IGZlcnJ5 31916
+IENvaGVu 31917
+IFByZWQ= 31918
+IHZhZ3k= 31919
+INCx0LXRgdC/ 31920
+IG11bHRpbQ== 31921
+IGRyYWluYWdl 31922
+IHNpbXVsYXRvcg== 31923
+Z2lnZ2xlcw== 31924
+IFN0YWRpdW0= 31925
+0L7QsdGJ 31926
+IG5vdGljZXM= 31927
+IGNyYXdsaW5n 31928
+IGdyb3VwZQ== 31929
+IGt0b8Wb 31930
+IFlvZ2E= 31931
+IG1lZGlkYQ== 31932
+INGF0LLQsNGC 31933
+IExpdGU= 31934
+IHJhdg== 31935
+b3JhbWE= 31936
+IGRpc2NvcmQ= 31937
+IERJUkU= 31938
+IHRlaA== 31939
+IE51cnM= 31940
+IHBpdGNoZWQ= 31941
+IGJhcmtpbmc= 31942
+IENva2U= 31943
+d2lhZA== 31944
+IHBvcHVsYXRlZA== 31945
+cGVsbGVk 31946
+INCx0L7Qsw== 31947
+IHBld25v 31948
+IEN1YmU= 31949
+IHJlY3J1aXRlZA== 31950
+IENhcmE= 31951
+xLHEn8SxbsSx 31952
+aW1hdGVk 31953
+INGI0LrQvtC7 31954
+aWNpb25hbA== 31955
+INC/0YDQvtGE 31956
+IGNvbnRhbWluYXRpb24= 31957
+IMO6bHRpbW9z 31958
+IGZlYXJmdWw= 31959
+IGVsZXBoYW50cw== 31960
+dXNp 31961
+IGlUdW5lcw== 31962
+IFN3YW1p 31963
+6rw= 31964
+IOyEpOuqhQ== 31965
+IFJpY2hhcmRz 31966
+IG1hZ25ldHM= 31967
+IFJpY2h0dW5n 31968
+IExlZ2lvbg== 31969
+IGtpdHR5 31970
+IGtpc3NlZA== 31971
+IHdhdGVyaW5n 31972
+IGNvbm8= 31973
+IFBhbGVzdGluZQ== 31974
+aWRpcg== 31975
+IG1hemU= 31976
+IGZsdWlkcw== 31977
+IFByb2R1Y2Vy 31978
+IEtyc25h 31979
+bGFm 31980
+INeQ15U= 31981
+IG1pZXN6 31982
+IFhpbmc= 31983
+b2ludGVk 31984
+c2Vpbg== 31985
+IEZ1aw== 31986
+IERlcHJlc3Npb24= 31987
+IER1dHk= 31988
+IFBhbnRoZXI= 31989
+IHN1bmQ= 31990
+IHJlZmVyZQ== 31991
+IGV4Y2x1c2lvbg== 31992
+IG5hdmFs 31993
+IFdpbnN0b24= 31994
+IHNsb2dhbg== 31995
+IGh5cG90aGV0aWNhbA== 31996
+IGVsZXZhdGU= 31997
+66C5 31998
+IGNhYmXDp2E= 31999
+IEdlc3VuZA== 32000
+bWV0ZXI= 32001
+IOyVhOuLiOuptA== 32002
+IGNsb3VkeQ== 32003
+4oCmPw== 32004
+IFNjaHJpdHQ= 32005
+IEpT 32006
+7I0= 32007
+IFNwcmluZ3M= 32008
+IEJhdHRlcg== 32009
+t7A= 32010
+IHRhaWxvcg== 32011
+IFBUU0Q= 32012
+IEdlbnQ= 32013
+IGJhxJ8= 32014
+IHNwYXR1bGE= 32015
+IGNyYXk= 32016
+IExlZ2lzbA== 32017
+IHPDug== 32018
+IGxldmU= 32019
+4Liy4Lih 32020
+IGVyYWQ= 32021
+IGRvbmc= 32022
+IGRlcm0= 32023
+IEJhbmtz 32024
+aWNobw== 32025
+IEZyYW56 32026
+cmF2ZWw= 32027
+0L7Qu9C+ 32028
+IGZsdXRl 32029
+IEVr 32030
+IGpveWZ1bA== 32031
+IGNoYXNlZA== 32032
+IExhcmdl 32033
+T3Zlcg== 32034
+IGVudHJlcHJlbmV1cmlhbA== 32035
+IGNvbnNpZGVycw== 32036
+0YPQtdC8 32037
+b3Bh 32038
+IGRvcm1pcg== 32039
+IEVsZW1lbnRhcnk= 32040
+IHByenlwYWQ= 32041
+0YPRgdC60LA= 32042
+INC+0YfQtdGA 32043
+dWdlbmU= 32044
+IHRlbmlkbw== 32045
+IGx1Z2FyZXM= 32046
+66U= 32047
+INGH0LDRgdGC 32048
+IHNhbw== 32049
+IGJyYWlk 32050
+IFZlcmU= 32051
+IFJlaWNo 32052
+IFBvc3M= 32053
+IGluYW4= 32054
+d2FuZA== 32055
+cmVm 32056
+IG1vbnRyZXI= 32057
+IDE5ODE= 32058
+YXPEsW5kYQ== 32059
+IGNocm9tZQ== 32060
+IFRyaW5pdHk= 32061
+IGV4cGxvaXRhdGlvbg== 32062
+IFNlbnNl 32063
+IENNUw== 32064
+IE5vYmxl 32065
+IOyEoO2DnQ== 32066
+IHN3ZWxsaW5n 32067
+ZWxlY3Ryb25pYw== 32068
+XT8= 32069
+IGJydXNoaW5n 32070
+IGxpcXVpZGl0eQ== 32071
+IEhvb2s= 32072
+IENvbm5vcg== 32073
+IEFsdW0= 32074
+IGd1Y2tlbg== 32075
+c3VpdGU= 32076
+IHdpZWxl 32077
+IGJhcnJlbHM= 32078
+IFJlZ2Vs 32079
+IE1lbnQ= 32080
+IFRyaXA= 32081
+IEJydXNo 32082
+IEVyaWs= 32083
+dXJhdGU= 32084
+yZly 32085
+IEN5cg== 32086
+b3VibGU= 32087
+IEJlY2Nh 32088
+IHBhc3N3b3Jkcw== 32089
+xbE= 32090
+Ym9yZw== 32091
+IHZlbmRv 32092
+IENsYXVz 32093
+IEZheg== 32094
+aW5kZXN0 32095
+IGRlY2Vhc2Vk 32096
+IGNvbXBhcmlzb25z 32097
+IExDRA== 32098
+IFBvcms= 32099
+IGV2ZW50dWFs 32100
+IHBhdHJlb24= 32101
+IGluYWJpbGl0eQ== 32102
+IGV4dGluY3Rpb24= 32103
+IOyii+yVhO2VmOuKlA== 32104
+INGB0L7RgQ== 32105
+YWp1 32106
+INeR15DX 32107
+IHNvZm9ydA== 32108
+IGRlc3RpbmVk 32109
+IFJpbg== 32110
+IG1vdXRocw== 32111
+IE5hdMO8cmxpY2g= 32112
+IHByZXNlcnZpbmc= 32113
+IGxpbXA= 32114
+b2N1c2Vk 32115
+0LjQvdCz 32116
+IGV4cG9zaW5n 32117
+IM6+ 32118
+640= 32119
+bGF1Z2g= 32120
+IGhpc3M= 32121
+IGluZGll 32122
+IGRldGFs 32123
+0YDQsNCy0YHRgtCy 32124
+IHRyw6pu 32125
+IG9nbmk= 32126
+IHNpbXBsZW1lbnRl 32127
+IDE5Nzg= 32128
+IGdvbw== 32129
+IDE5Njc= 32130
+IGdlbnVn 32131
+aMO2 32132
+IGhpc3TDsw== 32133
+IGxvYnN0ZXI= 32134
+Y2VuZG8= 32135
+IHRlaWw= 32136
+IGFsbGV2aQ== 32137
+MDAwMA== 32138
+T0xE 32139
+IHBlc29z 32140
+IGJvbnVzZXM= 32141
+IGFtaQ== 32142
+IHJldml2YWw= 32143
+IEhvcnNl 32144
+IHNhY2s= 32145
+VGFsaw== 32146
+IG11bGhlcg== 32147
+INC/0L7RgdGC0L7Rj9C9 32148
+IEhvb2Q= 32149
+SHVo 32150
+IOu2gQ== 32151
+IGh5dW5n 32152
+IE1lZXRpbmc= 32153
+IGltcG9ydGE= 32154
+IOywvuyVhA== 32155
+IFZlcm4= 32156
+IHN0cmlwcGVk 32157
+IHJlZnVzZXM= 32158
+IHF1YWxpZmljYXRpb25z 32159
+b3Bs 32160
+gOuPhA== 32161
+aXjDrQ== 32162
+IGRpYWI= 32163
+aXRpbWU= 32164
+Zmxvd3M= 32165
+IGluYWM= 32166
+IEdvbmc= 32167
+IG1lYW5pbmdsZXNz 32168
+IGNvdXJhZ2VvdXM= 32169
+IG1pY3JvYmk= 32170
+YXp5 32171
+aGlzdA== 32172
+IHZvbHVudGVlcmluZw== 32173
+VklF 32174
+IHZpb2xhdGVk 32175
+IHN5bXBhdGh5 32176
+IEVkaXQ= 32177
+ZWxlY3RyaWM= 32178
+cHJvZHVjdA== 32179
+IHBhbmRlbWlh 32180
+IGdlb21ldHJpYw== 32181
+IENvbnZlcnM= 32182
+Z3Jl 32183
+IGdsdXQ= 32184
+aXN0ZWQ= 32185
+INin2YTZgw== 32186
+IENoYWlu 32187
+IFByZXNlbnQ= 32188
+IFlpbg== 32189
+INGB0L7Qsw== 32190
+IFZsb2c= 32191
+IOyWtOuouA== 32192
+IGRvbm4= 32193
+IGhpdGNo 32194
+dWNraW5n 32195
+d2FsZA== 32196
+cmlzaw== 32197
+IGhhcmk= 32198
+IEtlbnM= 32199
+IElkb2w= 32200
+INCy0L3QuNC80LDQvdC40LU= 32201
+IHRvZGQ= 32202
+IHNtYXNoZWQ= 32203
+IGludmFyaQ== 32204
+INC60L7QvdGC0YA= 32205
+IGF1dGlzdGlj 32206
+7J6l64uY 32207
+UmVz 32208
+0LTRiw== 32209
+Y2hhdQ== 32210
+IHNlbHY= 32211
+IGjDpHR0ZW4= 32212
+4KS/ 32213
+IGV4cGVjdHM= 32214
+z4HOtw== 32215
+IGHDp8Sxaw== 32216
+IEhUVFA= 32217
+bGXFnw== 32218
+IHN3ZWVwaW5n 32219
+IEJldGE= 32220
+IGNvdW50ZXJwYXJ0cw== 32221
+YWJpbGU= 32222
+IFNpbXM= 32223
+Q3M= 32224
+IHJlcGFy 32225
+c3F1 32226
+IHByb3ZpbmNpYWw= 32227
+IHNoYXJlaG9sZGVycw== 32228
+IHJ1bnRlcg== 32229
+IGdlZGFjaHQ= 32230
+IFRlZW4= 32231
+IGdyYW5kcw== 32232
+YWdsZXM= 32233
+IHJvY2t5 32234
+dmVucw== 32235
+IHJpdmFscw== 32236
+dW5hbA== 32237
+IHJlYWN0cw== 32238
+66k= 32239
+IG1lcmN1cnk= 32240
+IEx1aWdp 32241
+INC+0LM= 32242
+IEpVU1Q= 32243
+IGxvZA== 32244
+IGNvcnRleA== 32245
+d2ln 32246
+IGxha2g= 32247
+7KSR7JeQ 32248
+IFZpYw== 32249
+IE11bmQ= 32250
+IG1hcHBlZA== 32251
+IERlbGw= 32252
+IERydWNr 32253
+IGxpZmVz 32254
+0LDQu9GM0L3QvtC1 32255
+aXZpZHVhbA== 32256
+YWTEsW0= 32257
+IGF0cmF2 32258
+IEZsdWc= 32259
+IEtsZWlu 32260
+6rGw7JW8 32261
+4Lir4LiZ 32262
+IGFwcGxp 32263
+4K6+Pw== 32264
+w7x5b3J1bQ== 32265
+INC40L3RgtC10YDQtdGB0L3Qvg== 32266
+IGRpc2luZmVjdA== 32267
+Pi0= 32268
+IGNoYW1wYWduZQ== 32269
+IGtsYQ== 32270
+b3BlcnM= 32271
+VHJhbnM= 32272
+IERlc2VydA== 32273
+IGN1bHRpdmF0ZQ== 32274
+IEZ1Y2tpbmc= 32275
+aWRlbGl0eQ== 32276
+INGC0LDQvQ== 32277
+IGluY3Vi 32278
+IHRlbXU= 32279
+IGxlYXJuZXI= 32280
+Zm91bmRlcg== 32281
+IFN5bA== 32282
+IGZhdG8= 32283
+emllcg== 32284
+IOyXhuydtA== 32285
+IOyIqA== 32286
+IHBzeWNobw== 32287
+INGC0LXQu9C10YQ= 32288
+IHJlZ2FyZGU= 32289
+IHJlcHJlc2VudGF0aW9ucw== 32290
+IGxpdGlnYXRpb24= 32291
+IHNwYW5u 32292
+dWx0cw== 32293
+Ymlvcg== 32294
+IFN1cnZleQ== 32295
+IExFRHM= 32296
+IHRyw6Q= 32297
+IGzDqm4= 32298
+IGFudGlveGlk 32299
+0LXRgNC+0Lw= 32300
+IGluZHVjdGlvbg== 32301
+IGZvb2xlZA== 32302
+w6R0emxpY2g= 32303
+INCz0L7QstC+0YDRj9GC 32304
+IEZhY3Q= 32305
+dW1iYWk= 32306
+IHdpZ2dsZQ== 32307
+Tk9VTg== 32308
+IGTDqXZlbG9wcA== 32309
+IENsYXJv 32310
+IOy4 32311
+66w= 32312
+IGFjY3VtdWxhdGU= 32313
+IG1haW50YWlucw== 32314
+64Q= 32315
+IEZpZ2h0ZXI= 32316
+7Yag 32317
+IG1hdGlu 32318
+IGNvdXBvbg== 32319
+IHN0dW50 32320
+IGRlYnV0ZWQ= 32321
+IHByYWc= 32322
+0LjQstCw0LXQvA== 32323
+NzM= 32324
+IGV4cHJlcw== 32325
+IOyYpOu5oA== 32326
+INC/0LXRgNGB0L7QvQ== 32327
+IGNhbGN1bHVz 32328
+IGFicnVwdA== 32329
+IEluc3BlY3Rvcg== 32330
+b3VydA== 32331
+xbpuaWVq 32332
+aW50ZW5zZQ== 32333
+QmE= 32334
+IGxvdW5nZQ== 32335
+IGFzdGhtYQ== 32336
+IEhpw6c= 32337
+qrs= 32338
+IGVkaXRvcmlhbA== 32339
+IHNlaXpl 32340
+IGvEsXI= 32341
+IG1vdXZl 32342
+IHRpZXJyYQ== 32343
+IHRlc3Rvc3Rlcm9uZQ== 32344
+IHJo 32345
+IEtpbmdzdG9u 32346
+RUxMRQ== 32347
+IFJlcHJlc2VudGF0aXZl 32348
+IDE5NzQ= 32349
+IGliYQ== 32350
+VHM= 32351
+IHNvcnRh 32352
+ICg/KQ== 32353
+INiq2Yg= 32354
+IOuCtOugpA== 32355
+IGJla29tbXQ= 32356
+IHNwaXJpdHVhbGx5 32357
+IGRpc3RvcnRlZA== 32358
+TWFk 32359
+IHJlaW0= 32360
+w6FuaA== 32361
+IE90dG9tYW4= 32362
+IFJlbGln 32363
+IEVscw== 32364
+IHJldGFpbmVk 32365
+IExhdWdocw== 32366
+IFNBUw== 32367
+INC60L7Qu9C40YfQtdGB0YLQstC+ 32368
+15XXqteo 32369
+IGlubm92YXRl 32370
+IGtvcms= 32371
+INGA0LDRgdGB0LrQsNC30YvQsg== 32372
+b25kZXJl 32373
+aXZp 32374
+YXll 32375
+b3VudHk= 32376
+INC/0L7Qu9GD0YfQsNC10YLRgdGP 32377
+IGJ1bnM= 32378
+IHnDvHpkZW4= 32379
+IHN1cmdlcmllcw== 32380
+2KPZhg== 32381
+IGJhbmtydXB0Y3k= 32382
+d2VsdA== 32383
+IHNpYW1v 32384
+IGRhcmtlc3Q= 32385
+IEhhbm4= 32386
+Z2dh 32387
+IGZvcm1hcw== 32388
+IERq 32389
+bmFtZWQ= 32390
+IHNoaWVsZHM= 32391
+dWVsbGVy 32392
+IEZldw== 32393
+IGxhY2U= 32394
+IGZ1cmlvdXM= 32395
+IFlV 32396
+IHNvY2lldGFs 32397
+IGp1ZGdlbWVudA== 32398
+IERvcw== 32399
+IGphYg== 32400
+bGF3cw== 32401
+IHJlaW52ZW50 32402
+IEthdGhlcmluZQ== 32403
+IENob2k= 32404
+YWRvd3M= 32405
+IHJhbnM= 32406
+b2Rlbg== 32407
+IE1pZHdlc3Q= 32408
+bsSxbg== 32409
+IGRlcG9ydA== 32410
+IERpcA== 32411
+IGF0ZW5jacOzbg== 32412
+IENvdXJ0bmV5 32413
+aXZpZGFk 32414
+INqp24E= 32415
+IGVmZmljYWN5 32416
+IEJyb29rcw== 32417
+IHJlZmVycmFs 32418
+INC60L7QvdGG 32419
+IG1hbGljaW91cw== 32420
+IGtpcg== 32421
+IEdvZGRlc3M= 32422
+IGZ1bmt5 32423
+IGludGVyaW0= 32424
+IEvDtnJwZXI= 32425
+IOyWvOun 32426
+a3Vy 32427
+INC60LvQuA== 32428
+IHRydWNz 32429
+Z2VzZXR6 32430
+IHp1Zw== 32431
+IEdsw7xjaw== 32432
+IE1pbnV0ZQ== 32433
+IHByZXN0aWdpb3Vz 32434
+IG5pZXo= 32435
+IGNvbmNlbnRyYXRpb25z 32436
+0LvQsNGB0YLQuA== 32437
+IFNpcw== 32438
+IFZpdGFtaW4= 32439
+a292 32440
+IFBCUw== 32441
+INC90LXQtQ== 32442
+IHJldGFpbGVycw== 32443
+IGNvbnZlbnRpb25z 32444
+IFNhbWFudGhh 32445
+IHByb3VkbHk= 32446
+Sm9yZGFu 32447
+IEpBU09O 32448
+YXRr 32449
+IHRyaXN0ZQ== 32450
+IHN0w6Ry 32451
+IHJlaXRlcmF0ZQ== 32452
+IHBvc3Rlcmlvcg== 32453
+IDE5NzM= 32454
+IFBpbmU= 32455
+IEp1bGlldA== 32456
+IHBlZGly 32457
+a2ls 32458
+IG92ZXJsYXBwaW5n 32459
+IGV4Y2x1ZGU= 32460
+IGVjb27Ds20= 32461
+IGFjY2VwdHM= 32462
+IFN0ZXI= 32463
+IOyatOuPmQ== 32464
+ZXN0YWI= 32465
+IHR1Zw== 32466
+YXJn 32467
+IGxpdnJv 32468
+2KfYtQ== 32469
+IHNlYW1z 32470
+IGJ1cmF5YQ== 32471
+IGVsbG8= 32472
+IFRN 32473
+IFBhdw== 32474
+IEluZGV4 32475
+RXhj 32476
+IGluc3BpcmF0aW9uYWw= 32477
+IGR1bms= 32478
+YWt0ZXI= 32479
+IGNvbmRpdGlvbmVy 32480
+IFNhbHV0 32481
+xYJlYw== 32482
+IOyJvQ== 32483
+INGD0LfQvdCw 32484
+IFJvbWVv 32485
+ZnJ1aXQ= 32486
+IFlP 32487
+IGNo4buJ 32488
+0LHRgw== 32489
+Ym9ucw== 32490
+IHJlcHJvZHVjdGl2ZQ== 32491
+IG9yYWRh 32492
+IO2aqA== 32493
+IHRlbnRhcg== 32494
+IG1hw7FhbmE= 32495
+IHNvbHZlbnQ= 32496
+SmVzc2ljYQ== 32497
+IExlZ2Fs 32498
+IHR1YQ== 32499
+IHNpYw== 32500
+IEVR 32501
+YXVrZWU= 32502
+7Iuc64uk 32503
+IMWedQ== 32504
+IGFkaGVyZQ== 32505
+IFR1bA== 32506
+IOCuhg== 32507
+IHRleHRib29rcw== 32508
+IEZpZnRo 32509
+IGV4cGVyaQ== 32510
+IGNoaWM= 32511
+IGhlYXA= 32512
+aW5lbHk= 32513
+YXRyYQ== 32514
+VHdv 32515
+IGhlbGVtYWFs 32516
+IGZyZW4= 32517
+IGJpc2hlcg== 32518
+2KfYtA== 32519
+IOyEoOyDnQ== 32520
+IFRhZ2Vz 32521
+IHPhu7E= 32522
+IGJ1bGxpZWQ= 32523
+2KQ= 32524
+IGJlbmVmaXRlZA== 32525
+IFByZXZpb3VzbHk= 32526
+INGN0YTRhA== 32527
+2Y0= 32528
+IHNlbmF0ZQ== 32529
+IE1vcm0= 32530
+aWprZQ== 32531
+IEZsdQ== 32532
+IGluY29ycG9yYXRpbmc= 32533
+amFjaw== 32534
+INC/0LjRgg== 32535
+IGltcGx5 32536
+IGhhY2tz 32537
+IFJJQ0g= 32538
+INC60LLQsNGA 32539
+INC/0YDQtdC60YDQsNGB 32540
+IGRlcGVuZGVuY3k= 32541
+IOyaqQ== 32542
+IOyxhQ== 32543
+IHfDpGhyZW5k 32544
+IHN1bGxh 32545
+IFBpdHRzYnVyZ2g= 32546
+IGVzZW1waW8= 32547
+vOuhnA== 32548
+cHJvdA== 32549
+IFJvc2Vu 32550
+IEluZGVwZW5kZW5jZQ== 32551
+IHBhcnNsZXk= 32552
+aWVnZW4= 32553
+IGhhdw== 32554
+IGFxdWVsbA== 32555
+IENBUA== 32556
+INGA0LDQsdC+0YLQsNGC0Yw= 32557
+IENsaWZm 32558
+aW9uYXI= 32559
+IHNlY3VyaW5n 32560
+zr3OtQ== 32561
+IHV0aWxpcw== 32562
+IGNvdWxl 32563
+IFBpbmc= 32564
+IHRyZWs= 32565
+IGZhaw== 32566
+IGVub3JtZQ== 32567
+IOyLqw== 32568
+IGRvdWJsaW5n 32569
+INC90YDQsNCy0LjRgtGB0Y8= 32570
+IGhlZA== 32571
+aG92ZW4= 32572
+IFN0YW5kaW5n 32573
+IG3DrW4= 32574
+IEppbWlu 32575
+IG1vbmFyY2g= 32576
+IGNva2U= 32577
+IG1y 32578
+IGNsaWM= 32579
+w40= 32580
+IGltcGVhY2htZW50 32581
+IGR1cmFiaWxpdHk= 32582
+IHZhcmlvcw== 32583
+IGNvbW1lcmNpYWxz 32584
+IGdyZWV0aW5ncw== 32585
+IFJp 32586
+IEFwcHJlY2k= 32587
+7J6I64qU 32588
+IHLDqXN1bHQ= 32589
+w6lydA== 32590
+IHNhbHV0ZQ== 32591
+IHBvZGVyaWE= 32592
+IHN1bnJpc2U= 32593
+dmVjaw== 32594
+IHJlbHVjdGFudA== 32595
+IGNvbW1pc3Npb25lcg== 32596
+w6J0ZQ== 32597
+IEtlbm55 32598
+IFNpcmk= 32599
+IOuKmA== 32600
+IEVF 32601
+IHVuY2g= 32602
+0LrQvtC9 32603
+INin2YTYpQ== 32604
+IGJlbHRz 32605
+IGhhc3M= 32606
+INC80L7Rjw== 32607
+IGRpc3BsYWNlZA== 32608
+IGFicmE= 32609
+zq3Ouw== 32610
+IHNjcmF0Y2hlcw== 32611
+IGNvbWV0 32612
+IGF1dGhvcml6YXRpb24= 32613
+IExMQw== 32614
+IHByb2R1aw== 32615
+IHJlaGFiaWxpdGF0aW9u 32616
+5Z4= 32617
+0ZbRhw== 32618
+dWRpbmc= 32619
+b2xpdA== 32620
+IDEwNQ== 32621
+IGV4cGFuZHM= 32622
+IGFsdHJp 32623
+IEtvbW1lbnQ= 32624
+IGFuZg== 32625
+UGw= 32626
+IE1hbmE= 32627
+ZmVk 32628
+IGJyaQ== 32629
+IG9yYQ== 32630
+R3M= 32631
+IEd1cg== 32632
+dWNrbGFuZA== 32633
+IGp1bmN0aW9u 32634
+IGlyb25pYw== 32635
+IEZlZWQ= 32636
+IHByYWt0 32637
+IEhhbW1lcg== 32638
+jOuPhA== 32639
+IFRyYWN5 32640
+IEFzaWRl 32641
+0L3QtdCz0L4= 32642
+INC40YHQv9C+0LvRjNC30L7QstCw0YLRjA== 32643
+IHphag== 32644
+IGVxdWl0YWJsZQ== 32645
+IGN1cmI= 32646
+IGRlcml2YXRpdmVz 32647
+IHB1cHBpZXM= 32648
+IEtlbm5ldGg= 32649
+IENvbXBs 32650
+aWdyYW0= 32651
+IEdhcmNpYQ== 32652
+KSI= 32653
+IEhhcmJvcg== 32654
+ZXN0aWFs 32655
+IGVycw== 32656
+5rk= 32657
+IHVud2FudGVk 32658
+IGJlbGFuZw== 32659
+0LDQs9C+ 32660
+ZW1i 32661
+ZG9z 32662
+IOyZnOs= 32663
+IEJ1ZGdldA== 32664
+IGJhdHRsaW5n 32665
+2K3Yqg== 32666
+a29r 32667
+0L3QsNGH0LDQu9Cw 32668
+IHBsYWc= 32669
+IGNhbnRpZGFk 32670
+IGdydXBvcw== 32671
+IHBsdWdpbnM= 32672
+bGVyaW5p 32673
+INC40LzQtdC10YI= 32674
+IHNvenVzYWdlbg== 32675
+b2xpY3M= 32676
+IHB1ZWJsbw== 32677
+IHJlbWluaXM= 32678
+csOkbg== 32679
+IE1vcnJpc29u 32680
+IGxpbmhh 32681
+IGJyZWF0aHM= 32682
+IFRhc3Rl 32683
+IGVuZnJlbnQ= 32684
+IERvY2tlcg== 32685
+INC00LXQvQ== 32686
+IGV0aG5pY2l0eQ== 32687
+IHdvYg== 32688
+IHN1ZmZlcnM= 32689
+IHRyYW5zaXRpb25pbmc= 32690
+IFJhbmdl 32691
+xJlkenk= 32692
+INC60LDRgg== 32693
+IHN5bmVy 32694
+IGRvbnV0 32695
+IHByb2JhYmlsaXRpZXM= 32696
+IE9tYXI= 32697
+V2hpY2g= 32698
+dWlzaA== 32699
+aXNpbg== 32700
+IGRlbW9z 32701
+IOyggOq4sA== 32702
+IOuYkeqwmQ== 32703
+INC10LTQuNC9 32704
+IGNlcnZl 32705
+IGpva2E= 32706
+SUFO 32707
+IGtpbG9tZXRlcg== 32708
+IGhvcml6b250YWxseQ== 32709
+IEJoYWc= 32710
+IC0+ 32711
+IE1vbml0b3I= 32712
+IGtub3dsZWRnZWFibGU= 32713
+IGZhdg== 32714
+IHBpbm5lZA== 32715
+IGVCYXk= 32716
+aWNrZXI= 32717
+IOyeoOq5kOunjA== 32718
+IFhpYW9taQ== 32719
+IGNhcGl0 32720
+IG5w 32721
+IDE5NjU= 32722
+aG9l 32723
+IG5vaw== 32724
+IFNhZ2U= 32725
+INC90LXQu9GM0LfRjw== 32726
+IFRvdw== 32727
+Z2Ft 32728
+IGRpY2Vu 32729
+IFNVQlNDUklCRQ== 32730
+IHJlYm9vdA== 32731
+IHBhag== 32732
+IOuztOyXrOs= 32733
+IHRoaWNrZW4= 32734
+IFJlYWxpdHk= 32735
+aWTDpG4= 32736
+TmE= 32737
+IOqyg+ydgA== 32738
+ISEp 32739
+IHJvdXRpbmVz 32740
+INC+0LTQvdC+0LPQvg== 32741
+IGV4dGluZw== 32742
+IOymnQ== 32743
+IHN1bGZ1cg== 32744
+IGNhcnZl 32745
+IGFzdGVyb2lk 32746
+IFdhcnJpb3I= 32747
+IHBob3RvZ3JhcGhlcnM= 32748
+IHBlbGw= 32749
+IGNyb3Nzb3Zlcg== 32750
+IGhhY2Vtb3M= 32751
+IE5lag== 32752
+IHNldHRsaW5n 32753
+IGlybQ== 32754
+IEJvb2tz 32755
+aWVudMO0dA== 32756
+IGVzcGFjaW8= 32757
+IFNjaG9sYXJz 32758
+IGRvb21lZA== 32759
+IElSUw== 32760
+d29obA== 32761
+IHNlZ3Vl 32762
+IOuIhOqwgA== 32763
+IHByYXRpYw== 32764
+QlQ= 32765
+IENvbnNpZGVyaW5n 32766
+IEJ1ZmZhbG8= 32767
+IHRyYWluaW5ncw== 32768
+IGdlYnJ1 32769
+IEdsZWljaA== 32770
+IHBpcmF0ZXM= 32771
+IGVudmVsb3A= 32772
+IHJlb3Blbg== 32773
+aW1hdA== 32774
+IHRlZQ== 32775
+IHN1ZWQ= 32776
+ZmVo 32777
+INeU16c= 32778
+IGRpZXRz 32779
+IGp1bnRvcw== 32780
+YXN0bw== 32781
+IG1pc3VuZGVyc3Rvb2Q= 32782
+IHJ1aW0= 32783
+IGNsYXNzaWZ5 32784
+INC/0YDQvtC00YPQug== 32785
+IGluc2U= 32786
+IGlsbHVzdHJhdGVk 32787
+IGNvcnJvc2lvbg== 32788
+IGFjY3JlZA== 32789
+IEF1bnRpZQ== 32790
+INC/0YDQuNCy0LXRgg== 32791
+IExJVkU= 32792
+IHJlaw== 32793
+IHJlY2VpcHQ= 32794
+IEJhcmJpZQ== 32795
+IFNuYWtl 32796
+dHVybg== 32797
+SmVmZg== 32798
+lYQ= 32799
+Vk9JQ0VPVkVS 32800
+Y29sbA== 32801
+IHJ1bm5lcnM= 32802
+7KCc6w== 32803
+b3Nvcw== 32804
+bW9vbg== 32805
+IGtleW5vdGU= 32806
+IEluc3RpdA== 32807
+U1BFQUs= 32808
+IHBsdWdz 32809
+IGN1cnY= 32810
+IFl1cmk= 32811
+IFRoZXJlcw== 32812
+IFBz 32813
+IM68z4DOvw== 32814
+IGNvbnZlcnRlcg== 32815
+IHJlZmluZQ== 32816
+IGJhZGFzcw== 32817
+IM6/zrk= 32818
+IHJlZ2Vu 32819
+YXp6aQ== 32820
+2YrZgQ== 32821
+IHNlaXplZA== 32822
+IGnDp2Vy 32823
+aWxlZQ== 32824
+IHVwc3RyZWFt 32825
+IGJ1ZHM= 32826
+IHBpbQ== 32827
+IO2VmOujqA== 32828
+IGFsbHVkZWQ= 32829
+IHRoZW1lZA== 32830
+IGNvbnNpc3Rpbmc= 32831
+IGJvbnM= 32832
+dW51eg== 32833
+INC/0YDQvtCy0L7QtA== 32834
+IExvdmVseQ== 32835
+4KWL 32836
+IHBhcmFjaA== 32837
+IFN0YWF0cw== 32838
+IHNlbGVjdGl2ZQ== 32839
+IGZhc2U= 32840
+IEdlb3JnZXQ= 32841
+IGNvY2FpbmU= 32842
+IHJlcHJvZHVjdGlvbg== 32843
+IExhcmE= 32844
+IExE 32845
+IGdo 32846
+Sm9u 32847
+IGzDpQ== 32848
+IOuRkOs= 32849
+IHR5cGVk 32850
+IEJhbmE= 32851
+65Oc6w== 32852
+IHNhdm9yeQ== 32853
+IFpvbWI= 32854
+c3RhbmRlbg== 32855
+IHBlZGVzdHJpYW4= 32856
+IGRpZmbDqXJlbnRz 32857
+IOyLuA== 32858
+IGNvbXBsYWluZWQ= 32859
+INCa0YLQvg== 32860
+INec16Q= 32861
+YWxpxZtteQ== 32862
+IG1vcnRhcg== 32863
+IHZlcmRpY3Q= 32864
+IHN1ZmljaWVudGU= 32865
+IE1pbGxpb24= 32866
+bWl0dGVs 32867
+aW5hbHM= 32868
+INin2YTYrg== 32869
+0LDRjtGB0Yw= 32870
+IG1pxJlkenk= 32871
+IE9sZQ== 32872
+IGludmVydA== 32873
+Y3p5xIc= 32874
+0L7Qt9C80L7QttC90L4= 32875
+c3RhcnRlcg== 32876
+IGF1ZGl0b3I= 32877
+IFNjb3V0 32878
+Y2hpZW4= 32879
+IFN2ZXJpZ2U= 32880
+dWZmbGVk 32881
+IHplaG4= 32882
+IEF1Y2tsYW5k 32883
+IGFyZ2VudA== 32884
+IDE5NzY= 32885
+IEhvZQ== 32886
+IGJvdGhlcnM= 32887
+IHNvY2lhbGlzdA== 32888
+IHBsaWVycw== 32889
+IGVtZXJnZW4= 32890
+IFhQ 32891
+0LXRgNC+0LI= 32892
+TW9yZQ== 32893
+IExldmk= 32894
+IEFuZGVycw== 32895
+aWJpbGlkYWQ= 32896
+IFBhcmVudHM= 32897
+IGluZHVjZWQ= 32898
+7Ja07KQ= 32899
+IGJhbGFuY2Vz 32900
+INCy0YvRiA== 32901
+IHN1Ym1hcmluZQ== 32902
+U3RhcnQ= 32903
+IGRyaWVz 32904
+IHZvbHZlcg== 32905
+IHRpY2tpbmc= 32906
+Y290dA== 32907
+IGZhag== 32908
+cHLDqXM= 32909
+IFNhYmI= 32910
+INC30LDRhw== 32911
+INC/0L7QutGD0L8= 32912
+IGJhcHRpemVk 32913
+IEJyaWxsaWFudA== 32914
+INCR0L7Qsw== 32915
+IG1vdHM= 32916
+Yml0cw== 32917
+IGxhdHRpY2U= 32918
+IGNvcmlhbmRlcg== 32919
+IHJlc2lkZW5jeQ== 32920
+eW5j 32921
+IHBpZXJ3c3p5 32922
+IEtub2Nr 32923
+IFphcA== 32924
+INCV0LI= 32925
+6rKs 32926
+IHVuZXZlbg== 32927
+IEphcw== 32928
+b2Rvcg== 32929
+NzQ= 32930
+IFNpdGU= 32931
+IGFjb250ZWNldQ== 32932
+eW1wdA== 32933
+IHRyaWxvZ3k= 32934
+IGxhbnRlcm4= 32935
+IFp1Y2tlcg== 32936
+dmFyaQ== 32937
+d2VsbGluZw== 32938
+IFBvdGF0bw== 32939
+Z29tZXJ5 32940
+IHJlYWN0ZWQ= 32941
+IENocm9u 32942
+IGplZGU= 32943
+YmVlbGQ= 32944
+IHR3ZW50 32945
+IGxhY3Q= 32946
+IHLDqXNl 32947
+IHJlbGVudA== 32948
+IGZ1cm5hY2U= 32949
+IHdpZGdldA== 32950
+IGVhcnRocXVha2Vz 32951
+IEFkanVzdA== 32952
+aWxpdA== 32953
+INij2Yg= 32954
+IGhlYXJpbmdz 32955
+IGRlZmVuZGFudA== 32956
+aXJzaW5peg== 32957
+IGJhc2s= 32958
+Y2ph 32959
+nKg= 32960
+IHJpZmxlcw== 32961
+IGluc3RhbA== 32962
+IEZvcmdpdmU= 32963
+cGljYWw= 32964
+INCe0YfQtdC90Yw= 32965
+IHBldGl0ZXM= 32966
+IGhw 32967
+IHJlbm93bmVk 32968
+IElubg== 32969
+IOyjvOyEuOyalA== 32970
+IGVtcGhhc2l6ZWQ= 32971
+IOyeiOyjoA== 32972
+IOqyg+ycvOuhnA== 32973
+xZM= 32974
+Z2lsaQ== 32975
+RGF2ZQ== 32976
+IGV4aGF1c3Rpbmc= 32977
+xYJ1Zw== 32978
+IHNjaGVtYQ== 32979
+zrzOrA== 32980
+Y3ljbA== 32981
+IGF1dGFudA== 32982
+IHBhcmNlbA== 32983
+IG1hdGVyaWE= 32984
+IEJlcnJ5 32985
+INGB0LDQvNC4 32986
+IGV4dHJhY3RlZA== 32987
+IFNheWluZw== 32988
+aXNtYXRpYw== 32989
+INC/0L7Qv9GA0L7QsQ== 32990
+IG5ldXJvbg== 32991
+Z3JhcGg= 32992
+nOuptA== 32993
+IGVuY2xvc3VyZQ== 32994
+IEpvaGFubg== 32995
+IGFmdGVybWF0aA== 32996
+0YLQvtCx 32997
+IHXFvHk= 32998
+IHNhbXA= 32999
+MzYw 33000
+IE1laQ== 33001
+IHRhY28= 33002
+IHJlY2VwdG9ycw== 33003
+IHB1bmNoZXM= 33004
+IEhvamU= 33005
+INmH2YbYpw== 33006
+PSIj 33007
+IEFuZ3VsYXI= 33008
+IG11c2lxdWU= 33009
+IHJvbA== 33010
+IMOx 33011
+c3RlcnJlaWNo 33012
+IGNsYW0= 33013
+IFRyZWFzdXJ5 33014
+Y2hlbWljYWw= 33015
+IGFwYXI= 33016
+IGFwcGVuZA== 33017
+IGZvcmJpZA== 33018
+IEhhbWJ1cmc= 33019
+0LDQutC+0LI= 33020
+IOq4iA== 33021
+aWxkYQ== 33022
+IHByZXBhcmF0aW9ucw== 33023
+IG1vZ8SF 33024
+IGNhbWlubw== 33025
+RXJpYw== 33026
+IEJsaW5k 33027
+IERpc2NvdmVyeQ== 33028
+7Lig 33029
+IGludGVycHJldGVy 33030
+IGJyZWQ= 33031
+IFBzYWxt 33032
+IGRlZmVuZGVk 33033
+7Ims 33034
+IEVyZmFocg== 33035
+IFBlYWNo 33036
+IG1vb25z 33037
+IE9zdA== 33038
+IHNww6ljaWFs 33039
+IGFycml2ZXI= 33040
+IFdpcw== 33041
+dWNp 33042
+IHJvYm90aWNz 33043
+SVZF 33044
+IHNpZWdl 33045
+YXJsYQ== 33046
+IHNlcGFyYXRlcw== 33047
+IFRD 33048
+7Y+w 33049
+cXVpc2l0ZQ== 33050
+IHBhcmVudGhlc2Vz 33051
+0LjQutC1 33052
+IHRyb3Vz 33053
+INGB0LjQu9GM 33054
+IGJlZXJz 33055
+INC/0LvQsNGC 33056
+IHNvbGE= 33057
+IGTDqHM= 33058
+bWluZ2hhbQ== 33059
+aWt0ZQ== 33060
+IG9vcHM= 33061
+IHR3aXRjaA== 33062
+z4g= 33063
+IFNob3VsZG4= 33064
+dXZyZQ== 33065
+IGxlZXI= 33066
+Y3JpcHRpb25z 33067
+IGV5ZXNoYWRvdw== 33068
+IEd1bw== 33069
+IFBvd2VsbA== 33070
+IHN1cHVlc3Rv 33071
+IGFuYQ== 33072
+cmFscw== 33073
+IE1vbnRyZWFs 33074
+IHN1cmZpbmc= 33075
+INCf0LXRgNCy 33076
+157XlQ== 33077
+IG1pbGxpc2Vjb25kcw== 33078
+IHN1YnVyYnM= 33079
+IHBsYW5ldGE= 33080
+0YPRiNC60LA= 33081
+aHJsaWNo 33082
+IEhZ 33083
+INiz25I= 33084
+IE1N 33085
+IEVmZg== 33086
+IEhT 33087
+YW5zb24= 33088
+IOyngeygkQ== 33089
+IHN1bw== 33090
+IGRlcGxveWluZw== 33091
+IGt1bnQ= 33092
+dGVyaW5n 33093
+IGVyZWN0 33094
+7J6l7J20 33095
+IOydjOyLnQ== 33096
+IHNwZWNpbWVu 33097
+IS4uLg== 33098
+IGxpZ25l 33099
+IGtvbnN0 33100
+YWRlcXU= 33101
+IOyDge2DnA== 33102
+IGFjY2Vzc2Vk 33103
+IFBvbGU= 33104
+a2lsbA== 33105
+IOuyhOs= 33106
+IGF1dGhlbnRpY2l0eQ== 33107
+IGFwcGVsbGU= 33108
+dWxsZQ== 33109
+IHJldmlzaW9u 33110
+IGdvYXRz 33111
+0LPQu9C4 33112
+IHBhdQ== 33113
+IFJhbmdlcg== 33114
+IEltYWc= 33115
+YXV0aG9y 33116
+IGV2ZQ== 33117
+IE1lc3Nlbmdlcg== 33118
+IG5heQ== 33119
+IHdob2xlcw== 33120
+w6R0dGU= 33121
+IG9ud2FyZHM= 33122
+IERlcG9pcw== 33123
+IO2RnO2YhA== 33124
+IFNBUlM= 33125
+IHdzenlzdGtpY2g= 33126
+IGRlc3RydQ== 33127
+dW1iaW5n 33128
+IGNvbXBhdGliaWxpdHk= 33129
+IG1pc2luZm9ybWF0aW9u 33130
+b2RvcmU= 33131
+IEZhdm9y 33132
+ZWtv 33133
+j4w= 33134
+d2F1a2Vl 33135
+IFRlYWNoaW5n 33136
+IEtP 33137
+IGJldHRpbmc= 33138
+IHF1ZXN0cw== 33139
+IHZpdnJl 33140
+INC80YPQt9GL 33141
+IHNhZ2E= 33142
+IHN3ZWxs 33143
+IGdlaGU= 33144
+INC+0YDQs9Cw0L3QuNC3 33145
+IGdpZGU= 33146
+IEdyb3Nz 33147
+IGRhbGVq 33148
+IGNsYXdz 33149
+4buZYw== 33150
+IHByZWp1ZGljZQ== 33151
+IGluc2lnbg== 33152
+aWhvb2Q= 33153
+IHBsZWQ= 33154
+IGTDs25kZQ== 33155
+IFBvbGl0aWNhbA== 33156
+IHByZW1pc2Vz 33157
+dW5kZXJ0 33158
+2LnYqg== 33159
+b25uZW4= 33160
+IGVzcGHDp28= 33161
+IGbDqQ== 33162
+IEhhcnJpc29u 33163
+IENlbnN1cw== 33164
+IGNhcmRpbw== 33165
+IGRpeQ== 33166
+IG1pbGlldQ== 33167
+IGpvdXJuw6ll 33168
+IFJlbGVhc2U= 33169
+TklF 33170
+IE11aw== 33171
+aWTDqWU= 33172
+4buNaQ== 33173
+IGnDp2luZGU= 33174
+npk= 33175
+IHJlc29uYXRl 33176
+IG1vbGVz 33177
+IEZseWluZw== 33178
+IEdsb3JpYQ== 33179
+IFBhc3Rvcg== 33180
+IEFyZW5h 33181
+Tk9O 33182
+0L7Qu9C+0LI= 33183
+IGFsbMOt 33184
+b21hdA== 33185
+7Ja064+E 33186
+IGNhcmFjdGVyw61zdA== 33187
+IGRlY2xpbmluZw== 33188
+0ZbRjw== 33189
+YW5jbw== 33190
+IEluZm9ybQ== 33191
+IGJhcmdhaW4= 33192
+IGJ1c2hlcw== 33193
+IE5hdHVyYWxseQ== 33194
+IHJlY2h0cw== 33195
+IFRlbnNvcg== 33196
+IFBhdHJpY2lh 33197
+IHByaW5jaXBpbw== 33198
+IE11bWJhaQ== 33199
+IHdvbWI= 33200
+IG5vc3RyYQ== 33201
+IGRpbGVtbWE= 33202
+IGlyZ2VuZHdhbm4= 33203
+IDE5NjQ= 33204
+IGVuZXJnw61h 33205
+INC90LDRgA== 33206
+IHNlZ3JlZ2F0aW9u 33207
+IEF0aGxldA== 33208
+IMK7LA== 33209
+IHllbmk= 33210
+IFNlaXQ= 33211
+IHZlbm9t 33212
+IGRha2lrYQ== 33213
+IOuPjOs= 33214
+IMOJbA== 33215
+IGZ1cw== 33216
+IE1vZw== 33217
+pr3ri4jri6Q= 33218
+IHJlbWFy 33219
+IFRlZGR5 33220
+IGJyZWFzdHM= 33221
+aWNhbnM= 33222
+a2Fw 33223
+IGjGoW4= 33224
+IEpQ 33225
+IHJlc3VycmVjdA== 33226
+IOyduOs= 33227
+aGVyaWNhbA== 33228
+IGZvdG9ncmFm 33229
+IEpvc8Op 33230
+IGxpdmVsaWhvb2Q= 33231
+IGJpYmxp 33232
+dGVyaQ== 33233
+IHZvcnN0ZWxsZW4= 33234
+IEFBQQ== 33235
+IGFzc2Vzc2luZw== 33236
+WUE= 33237
+IHNwbGVuZA== 33238
+IGV4Y2F2 33239
+IGJhcHRpc20= 33240
+eWxs 33241
+d293 33242
+TWFj 33243
+IHBsYXN0aWNz 33244
+dGVva2Jva2tp 33245
+IGludMOpcmVzc2FudA== 33246
+IGNvbW1hbmRlZA== 33247
+IGZhbW91c2x5 33248
+INCY0LvQuA== 33249
+IE1hbnVlbA== 33250
+IHNvdXRod2VzdA== 33251
+IGRlZm9ybWF0aW9u 33252
+w61jdWxv 33253
+INC90LDRhdC+0LTQuNGC0YHRjw== 33254
+IFBhdHRlcg== 33255
+ZGVncmVl 33256
+IGN6xJlzdG8= 33257
+Ii0= 33258
+IOyFiw== 33259
+IG1hbmdlcg== 33260
+IFRydXN0ZWU= 33261
+gOumrA== 33262
+IHB1bnRvcw== 33263
+aXZhYmxl 33264
+IHZvbGF0aWxl 33265
+IOuKkA== 33266
+IGluc3RhYmlsaXR5 33267
+IGNpZWw= 33268
+Y2nEhQ== 33269
+IHB1cml0eQ== 33270
+0L3QvtGB0YI= 33271
+U2ls 33272
+ZWRhcg== 33273
+Tk9VTkNFUg== 33274
+IHNwZWxsZWQ= 33275
+R0VS 33276
+IHNhbmN0dWFyeQ== 33277
+IGFjY2VsZXJhdGluZw== 33278
+IHNjb3V0 33279
+INC/0YDQtdCy 33280
+ZmFocmVu 33281
+IOuCmOyYqA== 33282
+IHBvY3rEhXQ= 33283
+IE1ldQ== 33284
+a2Fhcg== 33285
+s7Tqs6A= 33286
+YWtyYQ== 33287
+RG93bg== 33288
+IMOEcg== 33289
+IEVsaXRl 33290
+IGFsbG9ucw== 33291
+IG1heW9ubmFpc2U= 33292
+IFN1c3RhaW4= 33293
+cHJpc2luZ2x5 33294
+IHN1cGVydmlz 33295
+IOq3uOugh+yjoA== 33296
+IHVuZW1wbG95ZWQ= 33297
+IGZyZXNobHk= 33298
+INee16I= 33299
+IERo 33300
+IHRhY2tsaW5n 33301
+IG9ncg== 33302
+IOy0iOs= 33303
+IGxvZnQ= 33304
+YXJhaA== 33305
+IEFpcmw= 33306
+IERpcg== 33307
+INCc0L7QttC90L4= 33308
+IGJvb2tpbmc= 33309
+IENSQQ== 33310
+IGh0dHBz 33311
+IGNob2tl 33312
+IGdvd24= 33313
+IG5vaXRl 33314
+IHphYw== 33315
+aXN0b2w= 33316
+IHNlY3Jl 33317
+IHJlc2VtYmxlcw== 33318
+IGN1YWQ= 33319
+7IKs6rCA 33320
+c2hvdw== 33321
+IGJsYW5j 33322
+IGFndQ== 33323
+IFByaW50 33324
+YXN0ZWQ= 33325
+IFdlYXRoZXI= 33326
+aXBs 33327
+IG9ic2N1cmU= 33328
+IGNvbnRl 33329
+b3VnaHM= 33330
+KTs= 33331
+IERhbWU= 33332
+IGNsYXJpZmljYXRpb24= 33333
+IGludGltYWN5 33334
+IHVwaG9sZA== 33335
+IE1pcnJvcg== 33336
+IHdhZ29u 33337
+eGlkZQ== 33338
+IGNsb2c= 33339
+YXBwZXI= 33340
+IEltbWVkaWF0ZWx5 33341
+w7pkZQ== 33342
+IHRvdWNoZG93bg== 33343
+IHJvb2Z0 33344
+0LDRiNCw 33345
+IMOnxLFrdA== 33346
+IGxhaXNzZXI= 33347
+IFVucmVhbA== 33348
+ZW5zaXRpdmU= 33349
+IDEyMw== 33350
+IHBsYXN0ZXI= 33351
+IGR1Y2tz 33352
+IGV0bWU= 33353
+IGJpc2hvcA== 33354
+YnJldmk= 33355
+IGJpYw== 33356
+IHJ1bnRpbWU= 33357
+IGFtYml0aW9ucw== 33358
+0LzQsNGC 33359
+IFdlaW4= 33360
+IE1hcmk= 33361
+IO2KuOs= 33362
+IHJlc29sdmVy 33363
+IG5nw6B5 33364
+IFJpc2U= 33365
+IENydXM= 33366
+IG1lcmNoYW5kaXNl 33367
+IGVsaQ== 33368
+IHN0YXRld2lkZQ== 33369
+IG93bA== 33370
+IHR3aXN0aW5n 33371
+IGNvbnRhbWluYXRlZA== 33372
+IENvbW1lcmNl 33373
+aHl0aG0= 33374
+IMOI 33375
+IOyLpOs= 33376
+IG11c3N0ZQ== 33377
+dWly 33378
+IHN1bXM= 33379
+IFNvbWV3aGVyZQ== 33380
+IGthbWk= 33381
+IGFpcmVk 33382
+IEFORFJFVw== 33383
+IOq6 33384
+IHZpZW5kbw== 33385
+IGFudGlib2R5 33386
+IGFic29sdW1lbnQ= 33387
+IHByb3Rlc3RlcnM= 33388
+IFF1w6liZWM= 33389
+c3RhZHQ= 33390
+U2hhdW4= 33391
+IGNoYW1iZXJz 33392
+IFdlYXI= 33393
+IEVmZmVjdHM= 33394
+IGhhemFyZHM= 33395
+IG5laQ== 33396
+IGNvcmF6w7Nu 33397
+IOG8 33398
+IFNH 33399
+lKk= 33400
+IOyXreyLnA== 33401
+IGNvbWZ5 33402
+IENvZHk= 33403
+IHBlbnNhbmRv 33404
+IGdhbnNrYQ== 33405
+IEFjcm9zcw== 33406
+w7ZsbGln 33407
+YWJ5dGU= 33408
+IHdlZGdl 33409
+IGthbGlhbg== 33410
+IHNpZ3Vl 33411
+ZW5kZXM= 33412
+IEdyb8Of 33413
+IHV0aWxpc2Vy 33414
+IGZsb3du 33415
+0LDQvdC40Y4= 33416
+IGxldmFy 33417
+cmVzdHJpYWw= 33418
+IGlsbHVzdHJhdGlvbnM= 33419
+IGFzbMSxbmRh 33420
+QkxFRVA= 33421
+INC00L7RgdGC 33422
+IHR1cnJldA== 33423
+IHN1aXRjYXNl 33424
+emnEmWtp 33425
+IHNrZXRjaGVz 33426
+IGFjcmVk 33427
+IFJlaQ== 33428
+IHRzdW4= 33429
+IFNhZw== 33430
+IHRoaXJkcw== 33431
+IEtJUkJZ 33432
+cmFp 33433
+IGh1bWFub3M= 33434
+IHJlY29tbWVuZHM= 33435
+IGV4dHJhb3JkaW5hcmlseQ== 33436
+IGNvbW1lbmNlbWVudA== 33437
+S04= 33438
+b3Bleg== 33439
+INeR16k= 33440
+IGxldGhhbA== 33441
+IEVzdGFtb3M= 33442
+IGluc3BlY3Rvcg== 33443
+IFNlb2s= 33444
+ZXVu 33445
+IG9mZnNob3Jl 33446
+IGdldHRpbg== 33447
+eWVhcnM= 33448
+IFNpbGVuY2U= 33449
+IE5hdHVy 33450
+dXB1bg== 33451
+IHRyenk= 33452
+IG5vZ2V0 33453
+IGhhbWJ1cmdlcg== 33454
+IFByYWlzZQ== 33455
+w6luZA== 33456
+IDE5NzE= 33457
+eWxpZQ== 33458
+a3JpdA== 33459
+IOyDneqwgeydtA== 33460
+IG1vbWVudG9z 33461
+IGVzdMOp 33462
+IGRpc3NlbWlu 33463
+IGdpZ3M= 33464
+IGRlc2Fm 33465
+IGF2aXM= 33466
+IFpvbw== 33467
+IOyViuydgA== 33468
+aMOkbmc= 33469
+aGFrZQ== 33470
+IEJpc20= 33471
+IHJldGhpbms= 33472
+IE1hbGNvbG0= 33473
+IGlkZW50aWZpZXM= 33474
+bG93ZXI= 33475
+aXhlbA== 33476
+IHR2w6U= 33477
+a2Vk 33478
+aWVyeg== 33479
+IMO2ZmZlbnRsaWNo 33480
+IHByb2NsYWlt 33481
+c29vbg== 33482
+bG9s 33483
+IGxvaQ== 33484
+IGJpdHRlbg== 33485
+cm9sbG8= 33486
+IHNlcm1vbg== 33487
+IGVzcXU= 33488
+IGphY2tldHM= 33489
+IGdyw6FmaWM= 33490
+INC/0L7QutCw0LfRi9Cy 33491
+IGNhYmV6YQ== 33492
+Y2hvZHpp 33493
+IHBlbHZpcw== 33494
+IG5vc3RhbGdpYQ== 33495
+IGJyZXc= 33496
+IHNob3J0Y3V0cw== 33497
+IEFkZW3DoXM= 33498
+IHN1cGVyZmljaWFs 33499
+IGJvY2E= 33500
+aW1lbnRvcw== 33501
+IHNwcm91dHM= 33502
+IEpvbmFz 33503
+IEZsb3JlbmNl 33504
+c3RhdGlj 33505
+ZGF1Z2h0ZXI= 33506
+Kik= 33507
+xYJieQ== 33508
+ZmFzaGlvbg== 33509
+IEdpbmdlcg== 33510
+IOunpOs= 33511
+IGh1c3RsZQ== 33512
+dXRvcw== 33513
+INGC0Y/Qtg== 33514
+IEzDtnM= 33515
+16nXmded 33516
+YW55Y2g= 33517
+dHViZXI= 33518
+IHRpZHk= 33519
+IGZyb250YWw= 33520
+IHdoaXNrZXk= 33521
+IGh1bWlk 33522
+IM6f 33523
+IHJpZGdl 33524
+IG1hcmlu 33525
+IGJpZW50w7R0 33526
+IENhcnJpZQ== 33527
+Y2h3 33528
+IHRhaHVu 33529
+IEVyZ2Vi 33530
+RlI= 33531
+IOygleu2gA== 33532
+IFNvbGRpZXI= 33533
+IGVubGlnaHRlbm1lbnQ= 33534
+IGV4YW1pbmluZw== 33535
+IE5vdHJl 33536
+IGVyYW0= 33537
+IFN1bm55 33538
+IGxheWVyZWQ= 33539
+IERhenU= 33540
+cmFkZXM= 33541
+INC90LDRiNC10Lk= 33542
+IHRpbWJlcg== 33543
+IG1hbm5lcnM= 33544
+IEJpcm1pbmdoYW0= 33545
+IG1pbmlhdHVyZQ== 33546
+b21ldGVycw== 33547
+IGZpbGxlcg== 33548
+IFJpcA== 33549
+IEtvbWI= 33550
+b3duZXI= 33551
+7L8= 33552
+aWRpYW4= 33553
+IGRlbcOhcw== 33554
+INmI2Ko= 33555
+IHByZWNhdXRpb25z 33556
+IGdvdmVybm8= 33557
+emVsZg== 33558
+IENvbXBsZXRl 33559
+IFBoYW50b20= 33560
+INC90LXQtw== 33561
+INC60LDRgNGC 33562
+IEFudHdvcnQ= 33563
+IFBmaXplcg== 33564
+IEZyYW5jbw== 33565
+IHfFgg== 33566
+IGZyaWc= 33567
+ZXNwZXI= 33568
+IGthbGU= 33569
+IGZpbG1tYWtlcg== 33570
+IGt1cnQ= 33571
+IGludmFsaWQ= 33572
+YXJlbGxh 33573
+xINuZw== 33574
+cmFtZW50bw== 33575
+IG51dHJpdGlvbmFs 33576
+IGRpY3RhdG9ycw== 33577
+IGFmaW4= 33578
+IGZ1enp5 33579
+IEdpbmE= 33580
+w7N0 33581
+IEV4dHJlbWFkdXJh 33582
+IGRlbW9uc3RyYXRpb25z 33583
+IE1vbnRnb21lcnk= 33584
+7ZW07ISk 33585
+IEdhbmRoaQ== 33586
+IHJldW5pb24= 33587
+IGpha2nFmw== 33588
+IFp1Zw== 33589
+T1VHSA== 33590
+bGlmdGluZw== 33591
+IOCy 33592
+4bmb4bmj 33593
+ZWI= 33594
+IFdPVw== 33595
+IFNoaXZh 33596
+b21ldHJ5 33597
+IHdpbGRseQ== 33598
+IHRlbmRlZA== 33599
+IG1lZ2Fw 33600
+7LKY 33601
+IG5hdXNl 33602
+IGdlcmVr 33603
+IE1hcmNlbA== 33604
+IG5lc3Rl 33605
+2K7YsQ== 33606
+IGZlaA== 33607
+c3VzcGVuc2VmdWw= 33608
+IFdyZXN0bGU= 33609
+IFBhbGVzdGluaWFucw== 33610
+IEdPUkQ= 33611
+aXlldA== 33612
+INGA0LDQtNC4 33613
+IHZlcnN1Y2hlbg== 33614
+IHRyYW5zaXN0b3I= 33615
+INCf0YDQvtGB0YLQvg== 33616
+INC/0L7QvdGA0LDQsg== 33617
+IHJoeW1l 33618
+IFZlcm1vbnQ= 33619
+cGxhdHo= 33620
+IMSwxZ90ZQ== 33621
+IEhhZw== 33622
+INCY0Lw= 33623
+INGA0LDRgdGB0LrQsNC3 33624
+IG1ldHJvcw== 33625
+IEluZmluaXR5 33626
+d29sZg== 33627
+aWJhbA== 33628
+ZnRpZw== 33629
+INqG 33630
+IO2YueyLnA== 33631
+IG9nZ2k= 33632
+IGRpc3Bvc2l0 33633
+INC/0YDQuNC7 33634
+INCy0YvQv9C+0Ls= 33635
+IHRow7Rp 33636
+IEtFTk4= 33637
+IGhhbmRpbmc= 33638
+YWN0dXM= 33639
+IHRhY29z 33640
+IGZvcm1lcmx5 33641
+IENvcmludGhpYW5z 33642
+0YbRltGX 33643
+IHBhZHJl 33644
+IGNvbmdyZWdhdGlvbg== 33645
+5pE= 33646
+ZmVydA== 33647
+IHN1Ymly 33648
+YWlzZXI= 33649
+cXVh 33650
+YXJhb2g= 33651
+IEN1cnJ5 33652
+IOyViuuKlA== 33653
+0LXQu9GO 33654
+IGZ1c3M= 33655
+IGJvb3R5 33656
+IGxvd3M= 33657
+IGhvbW1lcw== 33658
+IE1I 33659
+IERpc25leWxhbmQ= 33660
+d2VudA== 33661
+IHJlc2lkdWU= 33662
+IGJlZXBpbmc= 33663
+w6R0dGE= 33664
+IG1vdWxk 33665
+IFByb2pla3Q= 33666
+c3RhbGs= 33667
+IGFydGlmYWN0 33668
+IEFudHJhZw== 33669
+IEFNRA== 33670
+IENyeXB0 33671
+IOuplA== 33672
+IEZlbGlwZQ== 33673
+IENPQg== 33674
+ZWx1 33675
+IHNlbGZpZXM= 33676
+IFNhbnRp 33677
+Y2h1dHo= 33678
+INCj0LrRgNCw0Zc= 33679
+Z2VzYW10 33680
+IGZsb2Nr 33681
+amF6 33682
+cGxhaW4= 33683
+IHdyaW5rbGVz 33684
+IHJlYWlz 33685
+IHBhbGpvbg== 33686
+IGVtcG93ZXJtZW50 33687
+IGF0dGVuZGVlcw== 33688
+cHBh 33689
+IG5lZGVu 33690
+0L7QvdGL 33691
+IHRpbWVmcmFtZQ== 33692
+IENoZXJyeQ== 33693
+IGlkw6ll 33694
+IGdhZw== 33695
+IGRvbmtleQ== 33696
+IMO0bmc= 33697
+IEhhcmU= 33698
+IEthcmE= 33699
+IGFjb21wYW4= 33700
+cGxhY2Vz 33701
+aW1pZW50b3M= 33702
+IEhhbW0= 33703
+0LHQuA== 33704
+dWJlbg== 33705
+aWxpeW9y 33706
+IHRoaXJzdA== 33707
+IGtyeQ== 33708
+IEdlb3JnZXRvd24= 33709
+16DXlA== 33710
+IG9yY2g= 33711
+IGhlYXJ0YmVhdA== 33712
+IHRyYW5zZm9ybWF0aW9ucw== 33713
+ZXN0b25lcw== 33714
+IEtI 33715
+IGNhcnRvb25z 33716
+IGFuY2k= 33717
+IHdvcnRobGVzcw== 33718
+IHRhaWxvcmVk 33719
+cHU= 33720
+QW1lcmljYW5z 33721
+IHBpbGVz 33722
+IE1vbmtleQ== 33723
+IGJhc2lu 33724
+IFRlbXBlcg== 33725
+IFBhaW50 33726
+IHB1bmNoaW5n 33727
+IGJhaWs= 33728
+IE9ha2xhbmQ= 33729
+dnJl 33730
+xZ9hbGxhaA== 33731
+eWRk 33732
+IGNhc3VhbGx5 33733
+b2R1 33734
+IGNvZGVk 33735
+IE5vcndlZ2lhbg== 33736
+IFZpbmNl 33737
+IHByZW1hdHVyZQ== 33738
+IFByb21pc2U= 33739
+0LXQutGB0YI= 33740
+IGRldmFzdGF0ZWQ= 33741
+IFByZW1pdW0= 33742
+IFBhcmFt 33743
+IMOWeWxl 33744
+dW11eg== 33745
+UE8= 33746
+cmF0b3Jz 33747
+IGxhbXBz 33748
+IHRlcnJpdG9yaWFs 33749
+IGJhY2tib25l 33750
+bGlzdGVk 33751
+RFk= 33752
+INin2YTYsQ== 33753
+IHB1cnN1ZWQ= 33754
+IENvbW1vbnM= 33755
+IOqzoQ== 33756
+bG9ja3M= 33757
+ZWRvcg== 33758
+IGNvbmNlaXZlZA== 33759
+Z2VyZQ== 33760
+IGRpc2FwcGVhcmluZw== 33761
+IFN1bGw= 33762
+IOyXsOs= 33763
+IGhvZmZl 33764
+IGRldG94 33765
+7ZSM 33766
+IHJldGly 33767
+IOuBneuC 33768
+IHBlcmd1bnRh 33769
+IEJPWQ== 33770
+IHBlbm4= 33771
+aMOpcw== 33772
+aG9u 33773
+IGNhdGFzdHJvcGhpYw== 33774
+IGF1c3Q= 33775
+IHRvcnNv 33776
+IOyWtOuKkA== 33777
+IOyCrOuejOuTpOydtA== 33778
+IG1hcnZlbG91cw== 33779
+IEhhcmxleQ== 33780
+YWNoaW5l 33781
+IHRp4bq/ 33782
+aXR0bw== 33783
+IEnDrW0= 33784
+eWxvbg== 33785
+IHNodXRkb3du 33786
+Licn 33787
+IGFwb2xvZ2llcw== 33788
+IENvbW11bmljYXRpb24= 33789
+INCz0L7QstC+0YDRjg== 33790
+4oSi 33791
+w612ZWlz 33792
+YWN1bg== 33793
+IHJldGFpbmluZw== 33794
+IGNvbnRyYWRpY3Rpb24= 33795
+IEFEQU0= 33796
+Q09N 33797
+QnJ5YW4= 33798
+IE1vbnNpZXVy 33799
+IGFkYXB0aW5n 33800
+0KjQkA== 33801
+IFNjcg== 33802
+w6RuZGVydA== 33803
+IHBsYXVz 33804
+IG9uc2V0 33805
+IGFzc2lzdGFudHM= 33806
+IHZhbHZlcw== 33807
+IHNjYXR0ZXI= 33808
+IFJ1c3Q= 33809
+YXdpYQ== 33810
+IHJlYWRpbmVzcw== 33811
+IHBhaXM= 33812
+IGJpYmxl 33813
+IGFtYmllbnRl 33814
+INCw0LzQtdGA0LjQug== 33815
+IHVuY29uZA== 33816
+IGthbGs= 33817
+IG1vYw== 33818
+dW5u 33819
+IGFjdHU= 33820
+IGh1bW1pbmc= 33821
+aXNzaW1v 33822
+IFBhdHJvbA== 33823
+Z293 33824
+IFRIRVk= 33825
+IEJvZGVu 33826
+IEJpZQ== 33827
+IHJlZWw= 33828
+INGD0YHQu9C+0LI= 33829
+IGVuZGVhdm9y 33830
+IFBlcmlvZA== 33831
+dXN0b21lZA== 33832
+bWFscw== 33833
+YWxvbg== 33834
+Qm94 33835
+IM+DzrHPgg== 33836
+IG9tZGF0 33837
+IGFsdHJl 33838
+IEhlaA== 33839
+a2Fk 33840
+IHByb3RlY3Rvcg== 33841
+IGRvbWluYW5jZQ== 33842
+b2R5bmFtaWM= 33843
+IGNvbW11bmljYXRlZA== 33844
+a8O2 33845
+IHByZWRlY2Vzc29y 33846
+IEx1aw== 33847
+IEZsb3dlcg== 33848
+cG9xdWU= 33849
+0YLQuNGA0L7Qsg== 33850
+IHJldHJvc3BlY3Q= 33851
+IGRlY2lzaXZl 33852
+IGV4ZW1wZWw= 33853
+e1w= 33854
+IFLDvGNr 33855
+cml0ZQ== 33856
+IFpldXM= 33857
+IGNhbG9yaWU= 33858
+IGF0dHJhY3Rpb25z 33859
+IEhpbnRlcg== 33860
+IHVobQ== 33861
+IO2MkA== 33862
+IHJ1bGVycw== 33863
+IGRpc2NvdXJhZ2Vk 33864
+IGFjb250ZWNlcg== 33865
+IGFjY2VudHM= 33866
+IE9wdGlt 33867
+IEFsZw== 33868
+a2lkcw== 33869
+MjAyMQ== 33870
+IExpbmRzYXk= 33871
+IGZpbG1tYWtlcnM= 33872
+cHJvd2Fk 33873
+IHRlcnVn 33874
+64u0 33875
+IFNvbW1lcg== 33876
+MjAxOA== 33877
+IGJvcnJvd2luZw== 33878
+IFRyYW5zZmVy 33879
+0L3QvtC/ 33880
+YXJpYXM= 33881
+IGhlYWRwaG9uZQ== 33882
+7Lyc 33883
+IHRyYW5zbGF0aW5n 33884
+IGF1Zmdl 33885
+4K6q4K6f 33886
+d2Vpcw== 33887
+YXZhbnQ= 33888
+cGFpZA== 33889
+YmFieQ== 33890
+IHRvdWdoZXN0 33891
+IHJlcGVhdHM= 33892
+IFRlcmVzYQ== 33893
+TG9yZA== 33894
+IGFjYWJhcg== 33895
+IFJpZGU= 33896
+ZGly 33897
+IGxlbmc= 33898
+IGR3YQ== 33899
+IGhlYWRhY2hlcw== 33900
+IG7hu69h 33901
+INC90LDRgdGC0L7Rj9GJ 33902
+IGJvaWxz 33903
+IGxvbmdpbmc= 33904
+cmlhcw== 33905
+w7NyaW8= 33906
+IFBhcmFkaXNl 33907
+IFNlw7Fvcg== 33908
+ZXJkZW0= 33909
+IHJlaW5zdA== 33910
+IHNhbGFyaWVz 33911
+IGluc2VjdXJpdHk= 33912
+xYJvxZtjaQ== 33913
+INCw0LHRgdC+0LvRjtGC0L3Qvg== 33914
+aW5rZW4= 33915
+IEVkZHk= 33916
+dWRvcw== 33917
+IGR1bW15 33918
+0JrQsNC6 33919
+c2l4 33920
+IGluYm94 33921
+4bqp 33922
+UGVvcGxl 33923
+4buTbmc= 33924
+IG9yZ2FuaXplcnM= 33925
+ZmluZA== 33926
+IMO8bA== 33927
+IENPTQ== 33928
+xbxh 33929
+d2VpbGU= 33930
+Q29tbWVudGFyeQ== 33931
+7Yq466W8 33932
+IE1pdHRlbA== 33933
+a3Vz 33934
+4KSo 33935
+aXJhbA== 33936
+IGdhcm1lbnQ= 33937
+zrnOus6s 33938
+IHN0b29s 33939
+cGF5ZXJz 33940
+IHNoaW1tZXI= 33941
+IE9sbGll 33942
+IEplxbxlbGk= 33943
+IDE5Nzc= 33944
+IGpldXg= 33945
+IGV4dGluY3Q= 33946
+IFRyYW5zcG9ydGF0aW9u 33947
+IE1ha2Vy 33948
+IGpvaG4= 33949
+IHJpY2hlc3Q= 33950
+IHRyYXVtYXQ= 33951
+IGxpZWdlbg== 33952
+tOulvA== 33953
+IHVucmVzdA== 33954
+IFN0cmF3 33955
+IGNvbWE= 33956
+IEtyaXN0ZW4= 33957
+INCa0L7QvdC10YfQvdC+ 33958
+IEJyeWNl 33959
+INGP0LrRlg== 33960
+IHBlYXJscw== 33961
+INC/0L7QvdC40LzQsNGO 33962
+IGFkZGl0aW9ucw== 33963
+IGFzeW1wdA== 33964
+INC80LXQvdGM0YjQtQ== 33965
+IHNjYW5z 33966
+Q2hpbGQ= 33967
+IEhpZGU= 33968
+0LrRg9GO 33969
+ZXRhcw== 33970
+IGRhbms= 33971
+IHBsZWFz 33972
+IGVzc2F5cw== 33973
+IGpldHM= 33974
+INCy0LXQtA== 33975
+IHBvc2l0aXZlcw== 33976
+aG9m 33977
+LSk= 33978
+enpv 33979
+IHN0YXJ0ZXJz 33980
+IHNtaWxlZA== 33981
+IDE5NDQ= 33982
+cXVpZXJh 33983
+IHJvaw== 33984
+IHB1ZXN0bw== 33985
+Tmljbw== 33986
+IHNpbXVsYXRpb25z 33987
+IOC2 33988
+IGludHJpZ3VlZA== 33989
+IE92ZXJ3YXRjaA== 33990
+c2lnaA== 33991
+YmFp 33992
+IOunkOqzoA== 33993
+aWTDqQ== 33994
+IGNyYWJz 33995
+4bqtcA== 33996
+IElyYXFp 33997
+7J2066W8 33998
+0YLRjw== 33999
+IFNvcGhpYQ== 34000
+IEROUw== 34001
+IMO2bmVtbGk= 34002
+IEx1bw== 34003
+naQ= 34004
+IENvdW5zZWw= 34005
+bGlnZW4= 34006
+0LDQvdGM0YjQtQ== 34007
+IHRydW1wZXQ= 34008
+IGRhcGF0 34009
+IEpN 34010
+IEVWRVJZ 34011
+IExheWVy 34012
+IGPDtA== 34013
+0L3QsNC7 34014
+IEpvbw== 34015
+IEhhY2s= 34016
+IHN1bnQ= 34017
+IExlb25hcmQ= 34018
+IEZpcmViYXNl 34019
+w6RuZ2Vy 34020
+IGV4cGxvZGluZw== 34021
+dm95 34022
+IOymkA== 34023
+INGB0LXRgNGM 34024
+IHNldmVyaXR5 34025
+IGJlc3RpbW0= 34026
+IHRpcmluZw== 34027
+IHByb2N1cmVtZW50 34028
+IGRpcGxvbWFjeQ== 34029
+IGRlY29yYXRpdmU= 34030
+INmK2Kc= 34031
+IHBlbmV0cmF0aW9u 34032
+1as= 34033
+IG91dHJpZ2h0 34034
+RU5F 34035
+IFVuaQ== 34036
+b2RsZXM= 34037
+IHplcm9z 34038
+IGRlbGlnaHRmdWw= 34039
+am0= 34040
+IGRvcG8= 34041
+IHBvc2l0aXZpdHk= 34042
+IFZJU1RB 34043
+IFJlc291cmNl 34044
+7YOA6w== 34045
+0YjQuNC1 34046
+Q2FybA== 34047
+IHBpcGluZw== 34048
+IGNob3BwaW5n 34049
+IEdhbnpl 34050
+w7xzcw== 34051
+IEFv 34052
+IHNoYXR0ZXJlZA== 34053
+IERldGVjdGl2ZQ== 34054
+IHVuZG91YnRlZGx5 34055
+IGhhbGx1Yw== 34056
+IGVuY2g= 34057
+0YvRh9C90L4= 34058
+0YPQu9GP0YA= 34059
+aXNlc3Rp 34060
+IHBlZGFscw== 34061
+IGR1cnVt 34062
+pO2U 34063
+bGFpbWVy 34064
+IHByb3ByZQ== 34065
+Q3U= 34066
+IHRyYW5zbGF0b3I= 34067
+IGNhxYI= 34068
+IOq3uOqxuA== 34069
+IGNhxYJ5 34070
+VUE= 34071
+IHJldmlzZWQ= 34072
+INC/0L7QtNC+0LE= 34073
+IEFydGljbGU= 34074
+IEhhaXRp 34075
+IMOT 34076
+IEN0cmw= 34077
+IHJvem0= 34078
+bGFpdA== 34079
+IGxldHp0ZQ== 34080
+aXNwZXJpbmc= 34081
+ZGlzcGxheQ== 34082
+IGFsdW1pbml1bQ== 34083
+IHBhbGFicmFz 34084
+IGNvbm9jZXI= 34085
+IHppdHRlbg== 34086
+IGRpcmln 34087
+IGJyYWluc3Rvcm0= 34088
+IHdpZmk= 34089
+IFBhcnRpY2lw 34090
+IHZpZXdwb2ludA== 34091
+IFF1YW4= 34092
+IGhpZXJhcmNo 34093
+V2VsY29tZQ== 34094
+IG9mZmVu 34095
+IFJlY292ZXJ5 34096
+Z2Fubw== 34097
+V291bGQ= 34098
+IHJlcHJv 34099
+IHBlcmNlcHRpb25z 34100
+IGRlbWFzaQ== 34101
+IEJhbmdsYWRlc2g= 34102
+IEluY3JlZGlibGU= 34103
+IGxldHp0 34104
+IGJlaGF2aW5n 34105
+IGFzdG9uaXNoaW5n 34106
+IOKG 34107
+IOuCqOyekA== 34108
+IEdPUkRPTg== 34109
+Q0FS 34110
+PyEi 34111
+IFByZXN0 34112
+IOunnuyVhOyalA== 34113
+IHRhbmQ= 34114
+IGxhc2g= 34115
+54o= 34116
+aWZpY2FudA== 34117
+IGludG9sZXI= 34118
+INCz0LXRgNC+ 34119
+IHRldQ== 34120
+YXNv 34121
+INGB0L7QstC10YI= 34122
+IHRyYXZlbGVycw== 34123
+IFN5bmQ= 34124
+INCy0LXRgNGB 34125
+Rm9uZGE= 34126
+YWTEsQ== 34127
+IHRyYW5zY3JpcHRpb24= 34128
+IHRpdGFuaXVt 34129
+IHR3aXN0cw== 34130
+IGdlYXJib3g= 34131
+ZW5zYXRpb24= 34132
+ZmF0 34133
+Q29sbA== 34134
+IENvbW1vbndlYWx0aA== 34135
+em9u 34136
+IFBvbGl6ZWk= 34137
+IEFQUExBVVNF 34138
+ZnJ5 34139
+IEp1ZGE= 34140
+ZXN0ZWVt 34141
+IHNvY2s= 34142
+IEp1Z2VuZA== 34143
+INC60YHRgtCw0YLQuA== 34144
+IERybw== 34145
+IHByb2NoYWluZQ== 34146
+IGxpa3NvbQ== 34147
+IEVuZXJnaWU= 34148
+IE1hcmluYQ== 34149
+IDIzMA== 34150
+IOqwgOyEnA== 34151
+dW1waW5n 34152
+IGxvbmU= 34153
+IGZvbnRz 34154
+IGJ1c2luZXNzbWFu 34155
+IHBseQ== 34156
+IGRvZQ== 34157
+Z3JpZA== 34158
+IE1pbHdhdWtlZQ== 34159
+IEVkZW4= 34160
+ISIu 34161
+INuM24E= 34162
+b2dlbnM= 34163
+IHRlYXNlcg== 34164
+IHF1acOpbg== 34165
+IGluY2VudGl2 34166
+Z292ZXJu 34167
+IGNoaWxkY2FyZQ== 34168
+IHNuZWFrZXJz 34169
+IGltcHJpc29uZWQ= 34170
+wq4= 34171
+0LjRgtC10YHRjA== 34172
+YW5idWw= 34173
+IHJlZ2Fpbg== 34174
+IHRyYW5xdWls 34175
+UmVkbmVy 34176
+SUZB 34177
+IGlkZW9sb2dpY2Fs 34178
+IG1heW9yw61h 34179
+IGJ1cmVhdQ== 34180
+ZXRlcm0= 34181
+IERJRA== 34182
+7Iq3 34183
+IHdhdmluZw== 34184
+IGJlYg== 34185
+IMOhcg== 34186
+INC60LI= 34187
+IGVudm95 34188
+YW51dA== 34189
+0LjQutGD 34190
+IEVudmlyb25tZW50 34191
+IEFzc2Fzcw== 34192
+IEJyZWFk 34193
+INCi0YPRgg== 34194
+IHN0YWlyY2FzZQ== 34195
+IERpc2Vhc2U= 34196
+IGF1Y3Vu 34197
+IOuLiA== 34198
+IGNvbmZyb250YXRpb24= 34199
+IDE5NDE= 34200
+IGlyb255 34201
+IHdvcnNo 34202
+IGZpY2s= 34203
+IE5hb21p 34204
+IGJhY2tzaWRl 34205
+aWV1eA== 34206
+S2Fw 34207
+IHZlZGVyZQ== 34208
+IGxlbmd0aHk= 34209
+IGJyZWFrZXI= 34210
+IFJvbGxl 34211
+IHByZWRhdG9y 34212
+IG5vc3Nvcw== 34213
+IGFkdmVydGlzZQ== 34214
+0YDQvtC00LU= 34215
+UmVkbmVyd2VjaHNlbA== 34216
+cmV0ZW4= 34217
+IGNvbGxlY3RvcnM= 34218
+xLHEn8SxbcSxeg== 34219
+IHRyaWc= 34220
+IGF4ZXM= 34221
+aW50ZXJz 34222
+IHBlbmFsdGllcw== 34223
+IE9zbWFu 34224
+IEplbm5h 34225
+IGZsYWtlcw== 34226
+IHRyYWluZXJz 34227
+IHN0dW5uZWQ= 34228
+IFNjcm9sbA== 34229
+IFBpcA== 34230
+INC90LDRgdGC 34231
+IG5ow6A= 34232
+IFNtYWNr 34233
+4bqrbg== 34234
+cmF0b3M= 34235
+INGA0LDQsdC+0YLRiw== 34236
+IHVjeg== 34237
+IExlbW9u 34238
+IFNpbmQ= 34239
+IHBzeWNoaWM= 34240
+IEFiZw== 34241
+IG1hbW1hbHM= 34242
+IGltbWVyc2l2ZQ== 34243
+IGJvdHM= 34244
+IHZlcnNjaGllZGVuZQ== 34245
+IGdlcmFs 34246
+IGZvbGxvd2Vy 34247
+IHNlZ3VyaWRhZA== 34248
+IGltbWVyc2Vk 34249
+ZmVpdG8= 34250
+Y3Jvc3M= 34251
+IMO2bGQ= 34252
+7YOE 34253
+INeU15nXkA== 34254
+IEppYW4= 34255
+IGJpbGl5b3I= 34256
+YXJlYQ== 34257
+IGthZg== 34258
+IGdvZHQ= 34259
+IOuwqeyGoQ== 34260
+IGRldHJpbWVudA== 34261
+0ZbQuw== 34262
+IMSRw6J1 34263
+IGNobG9yaWRl 34264
+w7hyZQ== 34265
+bGVp 34266
+IG1vbnRl 34267
+IGRpZmbDqXJlbnRlcw== 34268
+4K+BLg== 34269
+IGNhcmVnaXZlcnM= 34270
+IGluYWRlcXU= 34271
+IGZhcmV3ZWxs 34272
+INGC0LjQv9Cw 34273
+b250ZWM= 34274
+IEVwaA== 34275
+SEhI 34276
+IFRvZG9z 34277
+INCh0KjQkA== 34278
+IHRyb3Y= 34279
+IGxpZ2U= 34280
+IGPDtG5n 34281
+IENpdg== 34282
+IGNhcGF6 34283
+IFZhbGxhaGk= 34284
+IHF1ZXN0ZQ== 34285
+IHJlcGxpY2E= 34286
+2LPYqA== 34287
+em5h 34288
+INGB0LvRg9C2 34289
+IFBU 34290
+d2F2ZQ== 34291
+aWVuaQ== 34292
+IHJlbGllZA== 34293
+ZGV2ZWxvcA== 34294
+IGRlbWU= 34295
+IEFtYW4= 34296
+IFsuLi5d 34297
+IGNvbXBsaW1lbnRz 34298
+dWFpcw== 34299
+IO2MqA== 34300
+IHNtZWxsaW5n 34301
+IGRhZHVyY2g= 34302
+2YjYqg== 34303
+IG9yYW5nZXM= 34304
+INC70LDQuQ== 34305
+IHN0YWJpbGl6YXRpb24= 34306
+IGFwcGxpYW5jZXM= 34307
+IGht 34308
+g5DrqbQ= 34309
+b2R5bmFtaWNz 34310
+IGNpxJk= 34311
+IENvdHQ= 34312
+TU9O 34313
+IE1hbmc= 34314
+IGFsbGVyZGluZ3M= 34315
+zrnOus6u 34316
+c2hvdHM= 34317
+IHRz 34318
+IEfDtnI= 34319
+IENIQVI= 34320
+IDoo 34321
+IHdyYXRo 34322
+IGZpcXVl 34323
+IGbDvGhyZW4= 34324
+IHRlc3RhbWVudA== 34325
+IF5e 34326
+4bmb4bmj4bmHYQ== 34327
+QUxE 34328
+IHRleHRv 34329
+IERvZ3M= 34330
+IHNpYg== 34331
+IHBhdGhldGlj 34332
+b2Nrcw== 34333
+IHJhZGljYWxseQ== 34334
+IE1PUkU= 34335
+IEpBTUVT 34336
+IGluZ2w= 34337
+IFRlY2huaWNhbA== 34338
+IHBvcmNo 34339
+IFVU 34340
+INC+0LHRj9C30LDRgtC10LvRjNC90L4= 34341
+IHJlbmV3YWw= 34342
+IGFlc3RoZXRpY3M= 34343
+aWt1bQ== 34344
+IGJldmVyYWdl 34345
+ZGVybg== 34346
+IHByZWRpY3RpdmU= 34347
+IGNodXk= 34348
+IFJlZ2FyZGluZw== 34349
+IEZvcndhcmQ= 34350
+INmI2YQ= 34351
+IGNvbnRleHR1YWw= 34352
+IGR3YXJm 34353
+IHByZWhl 34354
+IGdvdmVybmVk 34355
+hYQ= 34356
+IHRyYWJhbGhhcg== 34357
+IG5lZ8OzY2lv 34358
+INCx0L7Qu9GM0YjQvtC5 34359
+0LXRh9Cw0YI= 34360
+INC00YPRhQ== 34361
+IGZsb29kcw== 34362
+IGJvd2xpbmc= 34363
+IE9C 34364
+IEjDpHI= 34365
+IGdyYWRpbmc= 34366
+7KO864qU 34367
+IGdhcnM= 34368
+ZGxpbmc= 34369
+IHJhaw== 34370
+64g= 34371
+Y3JlYXQ= 34372
+INGJ0LU= 34373
+IG5laWdoYm91cnM= 34374
+Zm9vZA== 34375
+UXVlcnk= 34376
+IGhlcm9pbg== 34377
+aWNlcHM= 34378
+IEtpbmRh 34379
+TkVU 34380
+IG1hcmk= 34381
+IGltaXRhdGU= 34382
+IGFjaHRlcg== 34383
+IHNldHRsZW1lbnRz 34384
+cmFyZQ== 34385
+Y2Npb25lcw== 34386
+IOuTnA== 34387
+IGZpaw== 34388
+aXR1bmc= 34389
+INC80LDQutGB0LjQvA== 34390
+IGVsZg== 34391
+IGRhbGxh 34392
+IFBvbHNjZQ== 34393
+IFB1bA== 34394
+0KfRgtC+ 34395
+IE1vcmdlbg== 34396
+2K3ZhQ== 34397
+IHN1cHJlbWFjeQ== 34398
+IGt5cw== 34399
+IEh1cnJpY2FuZQ== 34400
+IEdUQQ== 34401
+IEZlaA== 34402
+IGZpbmFsbWVudGU= 34403
+bXVuZA== 34404
+IEtyaWU= 34405
+w6lwb3F1ZQ== 34406
+IFR1Y2tlcg== 34407
+SVRU 34408
+IGx1cg== 34409
+IGRpcHBpbmc= 34410
+w6R2 34411
+IGVlcnN0ZQ== 34412
+IEZsaW50 34413
+YmlsZHVuZw== 34414
+4Li54LmJ 34415
+IHRvaW0= 34416
+IHByYWN5 34417
+IHRyYW5zZm9ybXM= 34418
+IHNwZWVkaW5n 34419
+IHByZXNlbnRlcg== 34420
+IGZlbGxvd3M= 34421
+ZmlsbGVk 34422
+aWV6YQ== 34423
+IGFkdmlzaW5n 34424
+IEludGVydmlldw== 34425
+0LjQs9GA 34426
+d2Vocg== 34427
+IERhbnRl 34428
+cHR1cmU= 34429
+iOusuA== 34430
+r7jr 34431
+kJA= 34432
+IENvdW50ZXI= 34433
+IGNyaXN0 34434
+IOynnA== 34435
+IGpldW5l 34436
+INGB0YLRgNCw0Yg= 34437
+IG1pZcSH 34438
+IHR1dG9y 34439
+IG1hc2FsYQ== 34440
+IHBvd2RlcmVk 34441
+IG5hdQ== 34442
+IEZyZWRlcmljaw== 34443
+IGJpbGxpbmc= 34444
+IEVpc2Vu 34445
+INC00L7QsdGA 34446
+IG1lc3Q= 34447
+5r0= 34448
+IHNuaXBw 34449
+IG1vbm8= 34450
+IEFsbw== 34451
+IE1lcmN5 34452
+w6lyaWVuY2U= 34453
+IGNhc3VhbHRpZXM= 34454
+IEFOTk9VTkNFUg== 34455
+IHRvY2Fy 34456
+IGJhY3RlcmlhbA== 34457
+SG8= 34458
+IHN0cmVhaw== 34459
+IEpFTk4= 34460
+IHBsYXN0 34461
+0YHQu9C10LQ= 34462
+IHJlYXBw 34463
+IHBheWNoZWNr 34464
+IG1pbmVycw== 34465
+aGFidA== 34466
+IEphcA== 34467
+0L3Rg9GC 34468
+IHJlZGVtcHRpb24= 34469
+IHF1aXI= 34470
+aG5saWNo 34471
+IGFjY3VtdWxhdGlvbg== 34472
+IHNob3Zl 34473
+IGFkcmVuYWxpbmU= 34474
+TWFrZQ== 34475
+IEhlcm4= 34476
+b3NzaW5n 34477
+IFZpbA== 34478
+dWJieQ== 34479
+aGVydHo= 34480
+YnJlYWtz 34481
+IHNwdXI= 34482
+IERhaGE= 34483
+VVNUSU4= 34484
+IGNvbnRpbnVlcg== 34485
+IFNhdWw= 34486
+IO2PrQ== 34487
+IOuQmOuptA== 34488
+IOunkOyUgA== 34489
+INC+0LY= 34490
+IHN1c3BlY3Rz 34491
+IGxhcXVlbGxl 34492
+IE11Y2hhcw== 34493
+IHbDtmxsaWc= 34494
+dWxlbg== 34495
+IGltcHJlcw== 34496
+IGxvYmI= 34497
+ZW5lZQ== 34498
+INC90LDQtg== 34499
+VGE= 34500
+IHLDqWFsaXTDqQ== 34501
+IFJleA== 34502
+IGhhcnZlc3Rpbmc= 34503
+IGVzdHI= 34504
+5rY= 34505
+b3NwYWNl 34506
+T1NT 34507
+IGRpc3R1cmJhbmNl 34508
+YXNzaWM= 34509
+IElzYWI= 34510
+IGTDqWNvdXY= 34511
+IEhhbXBzaGlyZQ== 34512
+IG9ybmFtZW50 34513
+IGx1w7Ru 34514
+IFVX 34515
+IGrEhQ== 34516
+IHJlc3BlY3Rv 34517
+IGNvbXVuaWRhZA== 34518
+IGNvbWlnbw== 34519
+YWduYQ== 34520
+IGludHJpbnNpYw== 34521
+IEFsdW1uaQ== 34522
+IHNlc2xlcmk= 34523
+IGVzdGltYXRpb24= 34524
+4oCU4oCU 34525
+IHByb2R1aXQ= 34526
+44CC44CN 34527
+INCy0YA= 34528
+IHdoaXJs 34529
+IGFjY2Vz 34530
+w6d1 34531
+IHZhcmlhYmlsaXR5 34532
+IHZvZGth 34533
+aXRzdQ== 34534
+IGludGVybnNoaXBz 34535
+IGFsbG9jYXRl 34536
+UlI= 34537
+7ZuI 34538
+IGluc3RydWN0aW9uYWw= 34539
+dGFudA== 34540
+IOCuheCupA== 34541
+IGludml0ZXM= 34542
+IGhhaw== 34543
+IHNjYXJlcw== 34544
+IGVjbGlwc2U= 34545
+0L/QvtCy 34546
+0LrQvtC70Yw= 34547
+YXRpdmFz 34548
+IHN0YWJiZWQ= 34549
+IERPTQ== 34550
+cm9vdHM= 34551
+IFBpY3R1cmU= 34552
+7Zi8 34553
+IENIQQ== 34554
+aWVj 34555
+xLHEsQ== 34556
+aGFub2w= 34557
+IG1pc3VuZGVyc3RhbmQ= 34558
+UmF5 34559
+IHJvYWRtYXA= 34560
+b2N1bWVudGVk 34561
+aXppb25l 34562
+IE9saXZl 34563
+cmlmdA== 34564
+INeU16A= 34565
+bGVzdA== 34566
+Ozs= 34567
+IEVB 34568
+0L7QtNGD 34569
+IGhvYmJpZXM= 34570
+IGJ1cmlhbA== 34571
+0KQ= 34572
+bGVnZQ== 34573
+IEhK 34574
+IG9iamVjdGlvbg== 34575
+Y3Rvcnk= 34576
+IGluY3JlbWVudGFs 34577
+IGd5bW4= 34578
+IGVwaWRlbWk= 34579
+0YHRi9C7 34580
+w5E= 34581
+IGFkdmFuY2VtZW50 34582
+IHBhcmNo 34583
+TmV3cw== 34584
+IGF5cg== 34585
+0LvQsNC8 34586
+INec16k= 34587
+IGRpcGxvbWE= 34588
+IHJvYmJlZA== 34589
+T25seQ== 34590
+IGluY3Vy 34591
+IGNoYW50aW5n 34592
+IO2VtOuPhA== 34593
+IHJpY2hlcw== 34594
+IENhcm1lbg== 34595
+IG5vc3Rybw== 34596
+zrvOrQ== 34597
+IFBvd2Rlcg== 34598
+4LmA4Lir 34599
+IOyeiOycvOuptA== 34600
+IGdlcsOnZWt0ZW4= 34601
+IFBpa2FjaHU= 34602
+0LXQvNC+0L0= 34603
+T0xM 34604
+IHBsYW5ldGFyeQ== 34605
+IHNsb3dz 34606
+IGNsb2Nrd2lzZQ== 34607
+YWxpb24= 34608
+IOyM 34609
+IHZlcm4= 34610
+IGhvbW1l 34611
+IGVuZHBvaW50 34612
+IGlubm9jZW5jZQ== 34613
+IGVsZW1lbnRvcw== 34614
+IHNvcGhvbW9yZQ== 34615
+IG5vdGlvbnM= 34616
+IENvdWxkbg== 34617
+cHVy 34618
+IHphdA== 34619
+IG9ic2Vzcw== 34620
+IG1vdGl2bw== 34621
+IEt1Yg== 34622
+IERydWc= 34623
+QW50 34624
+IFBsYXllcnM= 34625
+IEh1bWFucw== 34626
+IG1lbGVl 34627
+IFdpbGRsaWZl 34628
+IFZQ 34629
+IHZvbGNhbmlj 34630
+IGNvbWlu 34631
+IEd1YW5n 34632
+IM+EzrnPgg== 34633
+INC+0YHQvtCx0LXQvdC90L4= 34634
+IFNpemU= 34635
+TGlzdGVu 34636
+IEFhYQ== 34637
+YXBwcm8= 34638
+IGJhcmJhcg== 34639
+IFBhcmtpbnNvbg== 34640
+0L3Rj9GC0Yw= 34641
+IHVuZGVyZXN0aW1hdGU= 34642
+IHN1YnN0aXR1dGlvbg== 34643
+IGNvc21ldGlj 34644
+IHdpbGxlbg== 34645
+IGJlaWRl 34646
+YW5uaQ== 34647
+IGNvbmRpdGlvbmVk 34648
+IERlYmJpZQ== 34649
+IGlzdG8= 34650
+IEVkd2FyZHM= 34651
+7JuM7JqU 34652
+INGC0L7Qsg== 34653
+IGFiYnJldmk= 34654
+IE3DvG4= 34655
+IFByaW5j 34656
+IExpYW5n 34657
+IHN0aW5r 34658
+IHJhZGlvYWN0aXZl 34659
+IGFjb250ZWM= 34660
+IHVuY29u 34661
+IFR1cmJv 34662
+IGtpc3Nlcw== 34663
+0LXRgtGA0L7Qsg== 34664
+IGZyb250aWVy 34665
+IFNweQ== 34666
+IEJlbGFydXM= 34667
+IENCUw== 34668
+4buX 34669
+YW1vdG8= 34670
+7ZWc642w 34671
+INGB0YLRgNC+ 34672
+IEVuZmlu 34673
+IGJyZWFkdGg= 34674
+IENhZmU= 34675
+IERhZsO8cg== 34676
+IEJvdXI= 34677
+YXJhcw== 34678
+IGJsdWVwcmludA== 34679
+YW7EsQ== 34680
+IGNvbnN0YW50cw== 34681
+IGF0dGFja2Vy 34682
+IEZvcm11bGE= 34683
+emHEhw== 34684
+IHNvd2ll 34685
+IGV5ZWJyb3c= 34686
+b2Jvb2s= 34687
+IHNldHplbg== 34688
+b25zaWRlcg== 34689
+YXduaW5n 34690
+IHPDtnlsZXll 34691
+IGludmFkZWQ= 34692
+IHByb25vdW5z 34693
+IGRvYnJ5 34694
+U2k= 34695
+INCl0L7Rgg== 34696
+IHZvbGxleWJhbGw= 34697
+IGxhbWVudA== 34698
+aXNjaGVz 34699
+YXJtZQ== 34700
+YXBp 34701
+IFdpa2k= 34702
+0LvQuNGI 34703
+IGthc2lo 34704
+IHBlc3M= 34705
+INGE0L7Rgg== 34706
+IFN1bA== 34707
+IHBzZXVkbw== 34708
+IG1lbW8= 34709
+IOyXsOyKtQ== 34710
+INC00L7Qu9C70LDRgNC+0LI= 34711
+INC/0LXRgNC10Lw= 34712
+IFJlYWNo 34713
+bWlyYWw= 34714
+YWx0ZWQ= 34715
+IHN0YXR1dA== 34716
+cmVhZGluZw== 34717
+IHPDtnlsZWQ= 34718
+IExpbmRzZXk= 34719
+IEFobWFk 34720
+67aA6w== 34721
+INCh0LXQs9C+0LTQvdGP 34722
+IHByenlnb3Q= 34723
+IGh5c3Rlcg== 34724
+VVJF 34725
+IE5laWdo 34726
+UmVwb3J0ZXI= 34727
+IEJ1bnU= 34728
+IFRyZWF0eQ== 34729
+IFJhbms= 34730
+IEZhbWU= 34731
+aW5pc2hlZA== 34732
+IGdlYXJlZA== 34733
+IGNvbXBvc2U= 34734
+b2RpYQ== 34735
+IExvbg== 34736
+IGplc3RlxZtteQ== 34737
+IERJUkVDVE9S 34738
+IGVsa2Fhcg== 34739
+IFZpZWw= 34740
+15DXqQ== 34741
+eW50aGlh 34742
+IG3DqHJl 34743
+IFRvbWF0bw== 34744
+IGV4YXRhbWVudGU= 34745
+bmnEmQ== 34746
+IEZyZWk= 34747
+IERpZg== 34748
+IG9wZW5pbmdz 34749
+IGdyYXBoaWNhbA== 34750
+INGD0LTQvtCx 34751
+INCy0YHQvw== 34752
+IFdlZWtseQ== 34753
+0LXQstCw 34754
+IGhhbmdz 34755
+IHVuc2FmZQ== 34756
+IGVtYmxlbQ== 34757
+IEtvbGxlZ2lubmVu 34758
+YWxheQ== 34759
+IGtzaQ== 34760
+IGhpZGVz 34761
+IG9sbWF5 34762
+IGVudHN0ZQ== 34763
+IGFydGhyaXRpcw== 34764
+w59lcmRlbQ== 34765
+IGJpbm5lbg== 34766
+IGxpc3RlbnM= 34767
+IEhlc3M= 34768
+IExvdWlzZQ== 34769
+bGRlbg== 34770
+0LXQvdGB 34771
+IFZlcnNpb24= 34772
+IEFncmljdWx0dXJl 34773
+7Iqk66W8 34774
+0LzQsNC9 34775
+64Sk7JqU 34776
+IHdpbmVz 34777
+IElORg== 34778
+cnVs 34779
+IEpL 34780
+xLF5b3JsYXI= 34781
+c2hpZWxk 34782
+cmVhdGg= 34783
+IHRlcnVz 34784
+IEx1bQ== 34785
+IGFudGljaXBhdGlvbg== 34786
+IGFjY3VzdG9tZWQ= 34787
+IE1pbmE= 34788
+IHdpZWxk 34789
+aW/DqA== 34790
+bWVyYQ== 34791
+IGNvdW50ZG93bg== 34792
+IGNsaW5n 34793
+IGNvbW1lbmQ= 34794
+IGZha3Rpc2t0 34795
+IGRlZmVuc2Vz 34796
+IGNvY2twaXQ= 34797
+INC60L7QvNCw0L3QtA== 34798
+IGRpc2h3YXM= 34799
+IFRoYW5vcw== 34800
+IGtpZG5leXM= 34801
+IHNlaGU= 34802
+IG1pY3JvYmVz 34803
+IGN1ZmY= 34804
+INCy0YvRgdC+0Lo= 34805
+IFNwaWN5 34806
+4K614K6w 34807
+Y3VsdXM= 34808
+b3Jj 34809
+aXhlcw== 34810
+IENyZWRpdA== 34811
+IHJhag== 34812
+IGJyaW5ndA== 34813
+IE5pc3M= 34814
+IGdyaW0= 34815
+IFNPTA== 34816
+IHRlbmlt 34817
+IFN1ZGFu 34818
+IFNwYXJ0 34819
+IHByb21vdGVz 34820
+IE5vc3Nh 34821
+INGB0L7RgdGC0L7Rj9C90Lg= 34822
+IOywqQ== 34823
+IHVuY29udA== 34824
+IExpYmVyYWw= 34825
+INCi0L7Qu9GM0LrQvg== 34826
+IFZpZWxl 34827
+IGt0w7NyZWo= 34828
+ICoqKio= 34829
+TWF4 34830
+INCn0YLQvtCx0Ys= 34831
+MzUw 34832
+IO2YvOyekA== 34833
+IOu2hOuTpOydtA== 34834
+IHdhcnA= 34835
+IHRlbmdh 34836
+IHN5bXBhdGhldGlj 34837
+IGJpemk= 34838
+IFphY2s= 34839
+aWVkbw== 34840
+IOuJtOw= 34841
+cGllbA== 34842
+INGC0L7Quw== 34843
+IHNjYWxlZA== 34844
+IFBFVEVS 34845
+IENPTU0= 34846
+IENhbWU= 34847
+IGNhdGFzdHJvcGhl 34848
+IHN3ZWF0eQ== 34849
+aWdyYXRpb24= 34850
+IHN0dWZmaW5n 34851
+IM+Azr/Ou8+N 34852
+IERyaXZlcg== 34853
+enlzdA== 34854
+VGVjaA== 34855
+IGFzc2Vzc2Vk 34856
+IFN1cmZhY2U= 34857
+xLFyxLFt 34858
+c3Vy 34859
+bGVyd2VpbGU= 34860
+INC00L7Qsw== 34861
+IHNodXR0aW5n 34862
+IGZyYWN0aW9ucw== 34863
+INGB0L7Quw== 34864
+ZXZlcnlvbmU= 34865
+IGVybg== 34866
+INCd0L7Qsg== 34867
+IGRlZmVuZGVycw== 34868
+IHZlcnN1Y2h0 34869
+IHBvbGl0eQ== 34870
+INCf0L7QvQ== 34871
+dmVyc3TDpG5k 34872
+IGJyb3dzZXJz 34873
+IHRyYW5zZm9ybWF0aXZl 34874
+IGRpY3RhdGU= 34875
+IExFR08= 34876
+IG5pbmd1bmE= 34877
+6rSR 34878
+IHBpeno= 34879
+IEhhcm9sZA== 34880
+IExvcGV6 34881
+2r7bjA== 34882
+YW7EsXo= 34883
+YXRjaGV0 34884
+2YrYqg== 34885
+IGxlcm5lbg== 34886
+IOq3gOyXrA== 34887
+IGhvdXNlZA== 34888
+IGNsZWFuc2U= 34889
+IFdBVA== 34890
+bGFyYXRpb24= 34891
+IGJ5dGVz 34892
+IHR1Y2tlZA== 34893
+IGZhdWx0cw== 34894
+0LTQvg== 34895
+Rlg= 34896
+IOyWvOuniOuCmA== 34897
+IGRlZm9ybQ== 34898
+IGNvbnRyYWN0aW5n 34899
+IFRJTUU= 34900
+aXJzZQ== 34901
+IG5lYmVu 34902
+IGNlcmM= 34903
+IEFybXN0cm9uZw== 34904
+IHRlc3Rlcg== 34905
+IHBhcmZhaXQ= 34906
+IGplYWxvdXN5 34907
+IHRveGlucw== 34908
+IGRpc2JlbA== 34909
+0YPRgNGL 34910
+aW1wcmVzc2lvbg== 34911
+IHByb3N0YXRl 34912
+IGZpcmV3YWxs 34913
+IGNsYXNzaWNz 34914
+0LXRh9GM 34915
+IHNvY2lhbGlzbQ== 34916
+IGdyYWNpb3Vz 34917
+INGB0L3QvtCy0LA= 34918
+INC00L3Rjw== 34919
+IGJ1cm5lcg== 34920
+IE1pbm9y 34921
+IOyasOumrOs= 34922
+IGplZGVz 34923
+IGNvbnRpbnV1bQ== 34924
+IGhvdHM= 34925
+IG9jY3VycmVuY2U= 34926
+IGFkbWluaXN0ZXJlZA== 34927
+INC30LDQvNC10YI= 34928
+IGhlc2l0YXRpb24= 34929
+IGRyaWxscw== 34930
+ZXJjYQ== 34931
+INCy0YLQvtGA0L7QuQ== 34932
+IHN0ZWFkaWx5 34933
+IGluc2FubGFy 34934
+IGloYW4= 34935
+7ZE= 34936
+IGhlbHBlcg== 34937
+IFNlbmlu 34938
+0L7QstCw0L3QuNC1 34939
+IEVSSUM= 34940
+Ymxh 34941
+IEFjYWRlbWlj 34942
+IGh1bWFuaXRpZXM= 34943
+YmxhY2s= 34944
+dW1weQ== 34945
+b3J0ZXg= 34946
+IOygiOs= 34947
+INil2YY= 34948
+IGRpc2Nsb3Nl 34949
+IEVsaWphaA== 34950
+IM67zq0= 34951
+IFF1ZXI= 34952
+2KjZhA== 34953
+VGVsbA== 34954
+YXJsZQ== 34955
+0ZbRgA== 34956
+IGF1Z21lbnRlZA== 34957
+IOu5hOyKtw== 34958
+IGFuZHJvaWQ= 34959
+4KSk 34960
+YXJtYQ== 34961
+IHN6ZXI= 34962
+Z2VvcmQ= 34963
+IGdlZWs= 34964
+IHlldXg= 34965
+IHBvbmc= 34966
+IHRvcnR1cmVk 34967
+IEJhdGg= 34968
+emln 34969
+YXNvbmFibGU= 34970
+IG5ldHM= 34971
+IGJhcnU= 34972
+IEZsYXQ= 34973
+IFZhdGVy 34974
+IFRlcnJvcg== 34975
+IEF2bw== 34976
+IGNlcmVtb25pZXM= 34977
+cm9l 34978
+2YHYsw== 34979
+T3Bz 34980
+IGh5dmlu 34981
+IGFwcmVzZW50 34982
+b2xvcg== 34983
+INC40LPRgNGL 34984
+b3J0b24= 34985
+IOq3uOuerA== 34986
+IGxvb2tpbg== 34987
+IFRZ 34988
+IE1pbnQ= 34989
+QWRk 34990
+IG1pdGU= 34991
+IFNtb2tl 34992
+IG5vdGE= 34993
+IG1vc3M= 34994
+IEFiZW5k 34995
+IOy7qA== 34996
+IGV4YWdnZXJhdGVk 34997
+ZmlyZXM= 34998
+IHJlZGlzdA== 34999
+ZmZpdGk= 35000
+IG9wZW5uZXNz 35001
+6rCQ7J20 35002
+ZW5kZXU= 35003
+0LXQvdC90L7QuQ== 35004
+V2F0Y2g= 35005
+IGF2YXRhcg== 35006
+IFBleQ== 35007
+dXJ1bg== 35008
+IHNlbnph 35009
+IOyngOyXrQ== 35010
+IE5hdG9taWFzdA== 35011
+IGVtZXJnZW5jZQ== 35012
+cmF5cw== 35013
+IGNyYWZ0ZWQ= 35014
+Z2FyeQ== 35015
+w7xuZw== 35016
+LSI= 35017
+IGhhY2tlZA== 35018
+IHN0cmF5 35019
+ZW5jaWU= 35020
+ZW1v 35021
+IGNvbWVu 35022
+IEvEsXo= 35023
+IEphc21pbmU= 35024
+IEhpbmRp 35025
+bWFuYXM= 35026
+IGluZmluaXRlbHk= 35027
+ZW1vbg== 35028
+7J24642w7JqU 35029
+amFr 35030
+IHJvYXJpbmc= 35031
+w6lyaXF1ZQ== 35032
+c3dlaXNl 35033
+IFJvbGV4 35034
+IFN0dWFydA== 35035
+Ym5i 35036
+IGRpYWdub3Nl 35037
+IGNvaGVyZW50 35038
+IE1K 35039
+IHBpa2U= 35040
+bGF2 35041
+IG9yY2hlc3RyYWw= 35042
+0LDRgdGC0Lg= 35043
+IHRlcm1pbmFy 35044
+IGdhdGhlcmluZ3M= 35045
+IGNvbXBsaWFudA== 35046
+IHVwZ3JhZGluZw== 35047
+IHJlZ3VsYXRvcg== 35048
+IGxhbsOn 35049
+IG1lcmNoYW50cw== 35050
+dGF3YQ== 35051
+IG1vbml0b3JlZA== 35052
+IHJlbmRyZQ== 35053
+IHVudGVyd2Vncw== 35054
+YW5ndWFyZA== 35055
+Z2FyZA== 35056
+IEJlbG93 35057
+ZHVpbm8= 35058
+INCm0LU= 35059
+IGltcGVkYW5jZQ== 35060
+7Jyh 35061
+IGFrdHVlbGw= 35062
+IFZhdGlj 35063
+IHN0ZXdhcmRz 35064
+IGJyaWdodGVzdA== 35065
+IGtlbm4= 35066
+IGthdQ== 35067
+IE1hdHJpeA== 35068
+IEJhcms= 35069
+IPCfkQ== 35070
+IHRhcGVy 35071
+IGNhc2lubw== 35072
+16jXlA== 35073
+eXNpY2Fs 35074
+IGJ1aWxkZXJz 35075
+IGN6xYJvd2ll 35076
+IE5lcGFs 35077
+ICEi 35078
+IHRlcm1l 35079
+IGlubnljaA== 35080
+IG1hdGhz 35081
+IGRyYWZ0ZWQ= 35082
+IEJhbGs= 35083
+IGhlc2l0YW50 35084
+IHZvbHRhcg== 35085
+IHJldml2ZQ== 35086
+INGE0LjQu9GM0LzQsA== 35087
+IGFzc2Fzc2lu 35088
+IFNvbHV0aW9ucw== 35089
+IGR1ZWw= 35090
+IGJlYXJpbmdz 35091
+4LiE4Liw 35092
+IHJvb2tpZQ== 35093
+aWthdA== 35094
+IGJpc2N1aXRz 35095
+IGNvcmRz 35096
+0YPQstCw0YLQuA== 35097
+QVJJTg== 35098
+IHByb2dyZXNzaW5n 35099
+IEdpcg== 35100
+IHBlbmV0cmF0ZQ== 35101
+IFN0b3JhZ2U= 35102
+ZWlnaHQ= 35103
+INGC0YDRgw== 35104
+IGRvbsOtdA== 35105
+IHNpemlu 35106
+IG91dGRhdGVk 35107
+INC90LDRiNC4 35108
+IGFmZmly 35109
+IHNwb29ucw== 35110
+IG9uaQ== 35111
+IGZsYW5r 35112
+IEdvbA== 35113
+aMOj 35114
+IHDDqXJp 35115
+IGhvbm9yYWJsZQ== 35116
+IEJyZWF0aGU= 35117
+c2NlbmVz 35118
+IG9idmlhbWVudGU= 35119
+0LjQutGB 35120
+INep157X 35121
+IHNtb290aGll 35122
+nojr 35123
+IGRpbWU= 35124
+IO2WiOyWtOyalA== 35125
+IGFwcGVs 35126
+IENhdGhvbGljcw== 35127
+IHNpbmdsZXM= 35128
+IGxhdGVu 35129
+IMOnw7xua8O8 35130
+IFZhZGVy 35131
+IHZhcmTEsQ== 35132
+IElzdGFuYnVs 35133
+Z3LDqQ== 35134
+IEVsc2E= 35135
+w6ts 35136
+IGludmVjZQ== 35137
+IGNyYW5l 35138
+IG9iZQ== 35139
+IFNoYXJr 35140
+IHNtYWNr 35141
+IHJlc3RvcmluZw== 35142
+Llw= 35143
+IOu5oOs= 35144
+IGZhZGVk 35145
+dW1iZXJz 35146
+U2luZ2luZw== 35147
+IGRlcHJlc3Npbmc= 35148
+dGhlc3Q= 35149
+IFdhaHI= 35150
+IG11bHRpdHVkZQ== 35151
+0YDQsNCy0YHRgtCy0YPQudGC0LU= 35152
+cmlqaw== 35153
+ZWth 35154
+IGNvbXBsZXRlcw== 35155
+IFdlbGxz 35156
+IHJveQ== 35157
+IFByYXk= 35158
+IEthbGF1 35159
+aXppbg== 35160
+aWHFgmVt 35161
+IGxvY29t 35162
+IE5hc2h2aWxsZQ== 35163
+IFBlbnRhZ29u 35164
+66+4 35165
+IE5FVw== 35166
+xIXEhw== 35167
+w61zcw== 35168
+IG1hcnJ5aW5n 35169
+IGZldWQ= 35170
+7ZmV 35171
+KSE= 35172
+IE9wZXJhdGlvbnM= 35173
+0YPRlA== 35174
+IG1vamU= 35175
+IGluc3RydWN0ZWQ= 35176
+IOuIhOq1rA== 35177
+INeU15I= 35178
+INC/0L7QvNC+0YnRjNGO 35179
+IHNhYmlh 35180
+7JWY7Ja07JqU 35181
+cGxhbmU= 35182
+cHJp 35183
+INC/0L7Qu9C90L7RgdGC0YzRjg== 35184
+IEtpdHR5 35185
+IHByw7Nwcmlv 35186
+ZWRlcmU= 35187
+IGludGVyZXNhbnRl 35188
+INC00LU= 35189
+IGNvbmRlbnNlZA== 35190
+IGF2ZW50 35191
+VE9S 35192
+IGdyZWFzeQ== 35193
+QVJL 35194
+b3J0YQ== 35195
+QUo= 35196
+IGRpc3JlZw== 35197
+IGNvcnJlY3Rpb25z 35198
+IHN0ZXJv 35199
+IGluZmx1ZW56YQ== 35200
+IGRlc3Nlcw== 35201
+IGJhbGxvdHM= 35202
+IG1lZ2V0 35203
+IG1hZmlh 35204
+IGLDtmw= 35205
+bm9zdA== 35206
+INGB0YLQsNGC0Yw= 35207
+IHJlc3BvbmRlcg== 35208
+IGhpbnRlbg== 35209
+Z3Jhdg== 35210
+4Lit4Liw 35211
+eW5jaHJvbg== 35212
+IHZpZW5z 35213
+IHNhbW8= 35214
+IGR0 35215
+cGFubnQ= 35216
+IMWbd2lhdA== 35217
+INC30LDQv9C40YE= 35218
+IG1lcmdlZA== 35219
+IGtlcA== 35220
+IG1pc2xlYWRpbmc= 35221
+IGRpZ2Ftb3M= 35222
+IGFtbW9u 35223
+Y2hldA== 35224
+IOqwgOyguA== 35225
+IHVuaQ== 35226
+IOuQmOuKlOuNsA== 35227
+INC90LDQv9GA0LDQsg== 35228
+INC60L7RgtC+0YDQvtCz0L4= 35229
+IGFuaW1hdGU= 35230
+15XXkNc= 35231
+0LXRgNCy 35232
+IG1pbmNlZA== 35233
+IGthdW0= 35234
+z4DOtQ== 35235
+0LvQtdCz 35236
+ZXhpc3Rpbmc= 35237
+IHBsYXRhZm9ybQ== 35238
+IEtSSVM= 35239
+7Jug 35240
+IEZhbWlsaWVu 35241
+IExpYnlh 35242
+IGJpb2RpdmVyc2l0eQ== 35243
+IGlkaW90cw== 35244
+aXJkaQ== 35245
+IHN6eWI= 35246
+IFJvbGxpbmc= 35247
+w7xjaHQ= 35248
+INGD0LTQuNCy 35249
+0YHRg9C0 35250
+IHJlYWxpemFy 35251
+IGNhbm5lZA== 35252
+INGA0LDQvQ== 35253
+IG1ldGFib2xpYw== 35254
+IEJlZWY= 35255
+IGtpbGth 35256
+0LvRjtGB 35257
+IHJlZ2lzdHJ5 35258
+0LzQvtGC0YDQuNGC0LU= 35259
+IHZpZWzDpA== 35260
+IG9kYw== 35261
+IGNvbmRlbW5lZA== 35262
+ZmFs 35263
+IERpbA== 35264
+d2/Fm2Np 35265
+QXc= 35266
+IHN0YXRpc3RpY2FsbHk= 35267
+IHNvZ2Vu 35268
+IEJFVEg= 35269
+IHNoYXZpbmc= 35270
+b2NhbA== 35271
+IEZ1bm55 35272
+IHBlYWNlZnVsbHk= 35273
+IGFkZGljdGl2ZQ== 35274
+IEluc2VydA== 35275
+bGF1Zg== 35276
+IGV4cGVyaWVuY2lh 35277
+0LjRgtC10LvRjw== 35278
+w61nZW4= 35279
+w6FnaW5h 35280
+IGFiZG9tZW4= 35281
+7ZWc64uk 35282
+aWN1cw== 35283
+aW1hbmE= 35284
+7I2o 35285
+YXJjaGluZw== 35286
+IGtvbmtyZXQ= 35287
+7JWY6w== 35288
+0LXQutCw 35289
+b3VmbA== 35290
+aXZlbA== 35291
+IG51ZGU= 35292
+w6h0cmVz 35293
+IG1vbnNpZXVy 35294
+IGNsYXNo 35295
+IHRoZXJhcGlzdHM= 35296
+IGN1YmVk 35297
+IHJldHJvdXZlcg== 35298
+IHdhdmVmb3Jt 35299
+IHBvdGVt 35300
+IEZvcm1lcg== 35301
+aXNpw7Nu 35302
+INeQ150= 35303
+dW5kb3M= 35304
+IE1laW51bmc= 35305
+2LXZhA== 35306
+IEp1ZGU= 35307
+IG7DpXI= 35308
+IExlb25hcmRv 35309
+IENyaXN0bw== 35310
+IEdPVA== 35311
+0YHRgtGA0YPQug== 35312
+TEFO 35313
+IGfDpW5n 35314
+IGTDqWI= 35315
+IEZyYW5rZnVydA== 35316
+IGNyYXBweQ== 35317
+IGxpbA== 35318
+YW5uw6ll 35319
+INC80LXRgdGC0LU= 35320
+UkVU 35321
+IE5lcg== 35322
+IENPU1RB 35323
+IGplZGVt 35324
+IGN1cnRhaW5z 35325
+IGl0ZXJhdGlvbnM= 35326
+IHVuYXY= 35327
+IHBsYXF1ZQ== 35328
+b3J1bQ== 35329
+IM62 35330
+IG7Dum1lcm9z 35331
+IGRlc2Fw 35332
+sr0= 35333
+IGNvbXBpbGVk 35334
+IHJlZmxl 35335
+IHJhbmtpbmdz 35336
+IHJlcGFpcmVk 35337
+INCd0LDQv9GA 35338
+IGRvd25sb2Fkcw== 35339
+IGFybW91cg== 35340
+INeZ15XXqteo 35341
+IGxvbmdldml0eQ== 35342
+IFRPTkVS 35343
+INC60L7QvNC80LXQvdGC0LDRgA== 35344
+IGN6ZWdv 35345
+IG5vdGlmeQ== 35346
+IGFpcnBvcnRz 35347
+IGVuZHVyaW5n 35348
+bGV0dGU= 35349
+IGFwcGFyYXQ= 35350
+IGhhYmls 35351
+4buHYw== 35352
+bmFk 35353
+SUNP 35354
+IEJyYWg= 35355
+IHNlZ8O6bg== 35356
+IGdvdmVybm9ycw== 35357
+a2FoYQ== 35358
+IFNjaGx1c3M= 35359
+IG9kcG93aWVk 35360
+aXJ0aW5n 35361
+IHJlbXBs 35362
+IEFib3JpZ2luYWw= 35363
+aWRlbnRhbGx5 35364
+IGVuaGFuY2luZw== 35365
+bGljdGluZw== 35366
+IEhhd2FpaWFu 35367
+IHN0cml2aW5n 35368
+IE5pZXQ= 35369
+IHpuYWN6eQ== 35370
+IG9iZWRpZW5jZQ== 35371
+IG7DpWdvdA== 35372
+IGV4cGlyZWQ= 35373
+IDE5MTg= 35374
+cHJlc2VudGVk 35375
+IHByb3dhZA== 35376
+IFRlcnI= 35377
+IFByaW5jZXRvbg== 35378
+IG1vcmdlbg== 35379
+IGF0dHJhY3Rpbmc= 35380
+IFNpZ21h 35381
+aWduZXI= 35382
+IFJlY2h0cw== 35383
+IFBla2k= 35384
+IG1ldGh5 35385
+IGhhbW0= 35386
+IGRpcmVpdG8= 35387
+IGRlbGVnYXRpb24= 35388
+0LjQstCw0Y7Rgg== 35389
+IGdpbg== 35390
+WW91bmc= 35391
+IGRlcGVuZGVuY2llcw== 35392
+IEJyYWRsZXk= 35393
+YnVkcw== 35394
+IGZpcw== 35395
+IHB5dGFuaWU= 35396
+IGludGVyY29ubmVjdGVk 35397
+IGVtYmFpeG8= 35398
+IFNhcw== 35399
+IHJ1aA== 35400
+IFNpY2h0 35401
+U3Vy 35402
+IHN1cGVyYg== 35403
+IFNhYmJhdGg= 35404
+IERhbmdlcg== 35405
+a29s 35406
+IGhvdQ== 35407
+c3VwcA== 35408
+IE5hY2lvbmFs 35409
+IHN1Y2Nlc3Npb24= 35410
+IHbDoQ== 35411
+IE1hw59uYWhtZW4= 35412
+IEplc3NpZQ== 35413
+IElkYWhv 35414
+Zm9yZXN0 35415
+hZg= 35416
+INee15M= 35417
+INij2Yo= 35418
+IHN3ZWV0aGVhcnQ= 35419
+IG5lYXRseQ== 35420
+IEV2YW5nZWw= 35421
+6rOh 35422
+IFN1aXRl 35423
+w7pibGljYQ== 35424
+INGD0LvQuA== 35425
+IEFubm91bmNlcg== 35426
+bGlnaA== 35427
+IHNlbnNhdGlvbnM= 35428
+IHNoZWx0ZXJz 35429
+IGhhcnQ= 35430
+IHNxdWVlemluZw== 35431
+IFJpdmVycw== 35432
+IENvb2tpbmc= 35433
+7LGF 35434
+cGVyc29uYWw= 35435
+IG1hbm9z 35436
+0ZHRgtGB0Y8= 35437
+d2lq 35438
+IGdvZ2c= 35439
+IE1pbGxp 35440
+IEZQ 35441
+w7xuc3Q= 35442
+IExT 35443
+IHNwcmF5aW5n 35444
+IGZhdXg= 35445
+IGF1dG9ncmFwaA== 35446
+b2xvZ2lj 35447
+IHRvcm1lbnQ= 35448
+IGVuY3J5cHRlZA== 35449
+4buF 35450
+IGVzdHJl 35451
+4LE= 35452
+IHN0dW1ibGVk 35453
+IGFpZGVy 35454
+IHNhYmVu 35455
+eHRlcg== 35456
+IENpdGllcw== 35457
+IFTDvHJr 35458
+64ul 35459
+Y2hpbmU= 35460
+IHRvcHBpbmc= 35461
+IHBvaXNvbmVk 35462
+IFJvbWFuaWE= 35463
+15PXmQ== 35464
+gOuhnA== 35465
+INC/0L7RgNGP0LQ= 35466
+IGNoaXJwaW5n 35467
+IOyZhOs= 35468
+15HXog== 35469
+IGN1YW50bw== 35470
+IGRvbmF0aW5n 35471
+IFJlZ2VudA== 35472
+IEJlcnVm 35473
+IGRpc3RyYWN0aW5n 35474
+IHN0YW1pbmE= 35475
+IERhcnJlbg== 35476
+IOy2lQ== 35477
+bGlzdHM= 35478
+ZGFs 35479
+Y2h1c3M= 35480
+IGVjb25vbWlzdA== 35481
+b3JndA== 35482
+IGlzdGl5b3J1bQ== 35483
+IFN1cnByaXNl 35484
+IEhhbw== 35485
+IOy1nOqzoA== 35486
+IEdX 35487
+IElubmVy 35488
+IHF1aWVyZW4= 35489
+IG1pbmRlZA== 35490
+IHN1cGVyY29tcHV0ZXI= 35491
+IGRpYWdyYW1z 35492
+7Yqc6w== 35493
+6rKg7Ja0 35494
+INC+0LHRitGP0YE= 35495
+IGVzdGFiYW4= 35496
+IGRlc3Ryb3lz 35497
+IEJyZWFraW5n 35498
+IGthcsSxxZ8= 35499
+IHJlYnVpbGRpbmc= 35500
+nOuMgA== 35501
+0LvQuNCy0L4= 35502
+IFNhdWNl 35503
+IEZ1c2lvbg== 35504
+15XXntc= 35505
+IFF1aW5u 35506
+IGdhdWNoZQ== 35507
+INmI2KM= 35508
+IMg= 35509
+IHRlY2hubw== 35510
+IGRpc3BhdGNo 35511
+IGHFn2s= 35512
+IGVpbnplbA== 35513
+IEdtYWls 35514
+554= 35515
+IOqwnOyduA== 35516
+INGB0LXQvNGM 35517
+IGpvdXJuZXlz 35518
+IGlodA== 35519
+IGZpYnJl 35520
+IGRyYW1hcw== 35521
+b3VjaGVk 35522
+IHJlbmFtZQ== 35523
+INC+0L/QtdGA 35524
+IHBvbw== 35525
+IERydQ== 35526
+INC40YLQvtCz 35527
+IHphc3Q= 35528
+IGNveg== 35529
+IHp1Y2No 35530
+IG9idGFpbmluZw== 35531
+IGNvbW11dGU= 35532
+IHN1Ym1lcg== 35533
+IFZpc2g= 35534
+IFJhYmI= 35535
+b2dn 35536
+IGh1dA== 35537
+7ZaI7Ja0 35538
+ZXJlbWk= 35539
+IM68zrE= 35540
+IGRpc2t1dA== 35541
+INCx0YPQug== 35542
+IGltcGFpcmVk 35543
+ZGVwZW5k 35544
+INmI2Kc= 35545
+INGA0YPQug== 35546
+INCx0LDRgA== 35547
+IG94aWRhdGlvbg== 35548
+IHNpdHVhw6fDo28= 35549
+yZlu 35550
+dcOnw6Nv 35551
+IHNhZ3Rl 35552
+IFNFUg== 35553
+IENha2U= 35554
+IHR1cm1lcmlj 35555
+IEthaw== 35556
+YnVuZw== 35557
+IEvhuZvhuaPhuYdh 35558
+IHBvaXNvbmluZw== 35559
+IHNsaXBwaW5n 35560
+IFNheXM= 35561
+w7JuZw== 35562
+wqs= 35563
+IENsYXVkaWE= 35564
+IENoYXJhY3Rlcg== 35565
+0L3QuNGG 35566
+Y29hdA== 35567
+IHByb2dyZXNzZWQ= 35568
+IEZlcmd1cw== 35569
+IOyYpOuK 35570
+IG9hdA== 35571
+b3JkYWJsZQ== 35572
+IExleQ== 35573
+IEhlcmF1cw== 35574
+IHJlc3VsdGFkb3M= 35575
+IEtheWxh 35576
+IHJpZmY= 35577
+IGNoZWdvdQ== 35578
+IHhp 35579
+IHNwYWNpb3Vz 35580
+IHJlY29nbmlzZWQ= 35581
+IGVjaA== 35582
+IFRpZQ== 35583
+IGxhdW5jaGVy 35584
+Smlt 35585
+IHN1cHByZXNzaW9u 35586
+IEltcG9zc2libGU= 35587
+IGd1aXRhcnM= 35588
+IEZvdXJpZXI= 35589
+0LjRh9C10YHQutC40Lk= 35590
+IFRoZXJhcA== 35591
+IEthZg== 35592
+Y2VudGVyZWQ= 35593
+INGB0L7QvtGC0LLQtdGC 35594
+IGtsaW0= 35595
+IGNhcmJvaHlkcmF0ZXM= 35596
+aWduYW50 35597
+IEFzdHJvbg== 35598
+IGVtcGxl 35599
+IGRyYXN0aWM= 35600
+INC80LjRgNC1 35601
+0LLQuNC9 35602
+dXc= 35603
+IHByZXR0aWVy 35604
+IGRvbnV0cw== 35605
+IEF0aGVuYQ== 35606
+IGRpc3NlcnQ= 35607
+IHBsYW50ZQ== 35608
+IHVyYW5pdW0= 35609
+7J2M6w== 35610
+YXLDqQ== 35611
+IHJ6ZWN6 35612
+IGRpc3BsYXlpbmc= 35613
+IHNhcmM= 35614
+csOjbw== 35615
+IHRhbXBvY28= 35616
+IHBoaWxvc29waGVycw== 35617
+IFJlY2h0 35618
+IGNvbWVudGFyaW9z 35619
+eXNl 35620
+IOycpA== 35621
+IG1pc2U= 35622
+IEdpbg== 35623
+INC90L7QvA== 35624
+IEZST00= 35625
+bGluZXI= 35626
+YXRpZg== 35627
+IHNwb8WCZWM= 35628
+eGE= 35629
+INGC0YDRg9C0 35630
+IHdhZw== 35631
+6riw7JeQ 35632
+IE1H 35633
+IG9mZnNwcmluZw== 35634
+IFVuZGVyc3RhbmRpbmc= 35635
+T1JB 35636
+IHdoaXJyaW5n 35637
+IHN1cnJlbmQ= 35638
+IHBva2Vy 35639
+IG1vbnVtZW50cw== 35640
+IOKZqQ== 35641
+IG9yZ2FuaXNlZA== 35642
+IFNvemlhbA== 35643
+IEZhY3Rvcnk= 35644
+0YXQsA== 35645
+IHJlc2VtYmxl 35646
+0LfQtA== 35647
+IGV4cGxvc2lvbnM= 35648
+IHBheXJvbGw= 35649
+IG9tbg== 35650
+IEpvcmdl 35651
+zrnPgw== 35652
+IGZyYWN0dXJl 35653
+IHBlcnNlY3V0aW9u 35654
+IGRlbWFpcw== 35655
+RUNI 35656
+LCk= 35657
+IGNyaWFy 35658
+IEpPU0g= 35659
+IGRlbW9ncmFwaGljcw== 35660
+IDE2MDA= 35661
+IGN1cnJlbmNpZXM= 35662
+IFRpcHM= 35663
+IFJlZmVy 35664
+IERhbmNpbmc= 35665
+IGluY29uc2lzdGVudA== 35666
+IGRlaA== 35667
+IGltbWVucw== 35668
+IG1laXN0 35669
+IGltcGF0aWVudA== 35670
+IGJlaGF2ZXM= 35671
+IOuCtOyaqQ== 35672
+IGJhY2tzdG9yeQ== 35673
+IGFncmVlaW5n 35674
+IMWB 35675
+aWhpbg== 35676
+IHRlbXBlcmF0dXJh 35677
+IEJhY2tncm91bmQ= 35678
+IG51dHplbg== 35679
+IOuFuQ== 35680
+IE3DpG5uZXI= 35681
+IGNvbGxhYm9yYXRpb25z 35682
+IEtvcw== 35683
+IG5pZ2h0bWFyZXM= 35684
+65Ox 35685
+IFF1ZWVuc2xhbmQ= 35686
+IGFzc29jaWF0ZXM= 35687
+IEtvaw== 35688
+IGZhY3RvcmlhbA== 35689
+IEh5dW5n 35690
+IOq3uOuLpOydjA== 35691
+IGZpbGhv 35692
+IGVsw6l0 35693
+IO2WieuztQ== 35694
+sLE= 35695
+IGdlZnVuZGVu 35696
+IHNlbWljb25kdQ== 35697
+IGNvdW5zZWxvcnM= 35698
+IFVwcGVy 35699
+IEF1Yg== 35700
+aWNrZXJz 35701
+VmVy 35702
+IG5vcnRod2VzdA== 35703
+IE1haW50ZW5hbnQ= 35704
+IExha2Vz 35705
+0LDRj9Cy 35706
+aW50w6k= 35707
+7LC9 35708
+INCz0LDQtw== 35709
+IGdpb3Ju 35710
+IGRpZ2l0YWxseQ== 35711
+IENpcmN1aXQ= 35712
+7LyA 35713
+IGNoZWVyZnVs 35714
+IFBldGVyc29u 35715
+IERhbmlzaA== 35716
+YXRpdm9z 35717
+IGxpa2Vu 35718
+IGhhcmJvcg== 35719
+0LDQu9C40YHRgg== 35720
+eGU= 35721
+IGN1cmxz 35722
+IFJob2Q= 35723
+RW5k 35724
+IEVU 35725
+IGFjcXVhaW50 35726
+IEtlbHZpbg== 35727
+IHRyaWY= 35728
+IEF3YXk= 35729
+7J6Q64qU 35730
+dnM= 35731
+IHDDoWdpbmE= 35732
+IGlubGV0 35733
+IFNhbnRvcw== 35734
+IOyasOyZgA== 35735
+IHlhcMSxeW9yc3Vu 35736
+dGhlbWU= 35737
+IHNvdWZm 35738
+IGluamVjdGVk 35739
+IHDDs8W6bmllag== 35740
+aXZlcnNv 35741
+YW1wZWQ= 35742
+IGRhaGVy 35743
+IGRhZ2dlcg== 35744
+INC70Y7QsdC40Lw= 35745
+IHR1bW15 35746
+IGVubGlnaHRlbmVk 35747
+Y2VudHM= 35748
+IERhaA== 35749
+IGN1ZXN0 35750
+SUxZ 35751
+INeR16g= 35752
+IGJhbmdpbmc= 35753
+IEVtaWw= 35754
+IENsZXI= 35755
+IEJvcmRlcg== 35756
+0LjQttGD 35757
+IHByZXNlbnRlcnM= 35758
+IFNUVUQ= 35759
+Y29pbnM= 35760
+IO2ZjQ== 35761
+IHBlcmtz 35762
+IHBhcmFw 35763
+IGNlcnRhaW5lcw== 35764
+IExvcmU= 35765
+w7ZzdA== 35766
+IE1BUlRJTg== 35767
+IGJpb3M= 35768
+IHdoZXJlYnk= 35769
+dmVydHM= 35770
+IE1pcmFuZGE= 35771
+IHN0aXA= 35772
+YW5kZXo= 35773
+15vXnA== 35774
+dWppbg== 35775
+IOq+ 35776
+IGFsbGVyZ2llcw== 35777
+cGxhdGU= 35778
+IHlhcMSxbA== 35779
+IHVuZGVydGFrZQ== 35780
+IOuCmOqwgA== 35781
+UGFydA== 35782
+IGvEsXrEsW0= 35783
+aGd1cnU= 35784
+IEpvaG5z 35785
+IGV5ZWxhc2hlcw== 35786
+IGRyYWluZWQ= 35787
+IHN0w6Vy 35788
+IEphZGU= 35789
+IGNhbGVuZA== 35790
+ZmlsbQ== 35791
+IG1lc2E= 35792
+IGx1ZHppZQ== 35793
+IGF0dHJhY3Rz 35794
+IGp1aWNlcw== 35795
+INC60LjQuw== 35796
+IG5pZXV3ZQ== 35797
+IG1lbmNpb24= 35798
+IGlnbml0aW9u 35799
+IGJsYWRkZXI= 35800
+YW5kYWFn 35801
+IEV4dGVuc2lvbg== 35802
+7YKo 35803
+ZmVlZA== 35804
+INmI2Yc= 35805
+IHNwdW4= 35806
+IHTDpHQ= 35807
+0L7RgNC+0YI= 35808
+dHlhcmQ= 35809
+cm9uaWNz 35810
+IEh1Z2U= 35811
+0YPQttC0 35812
+c3RyaW5n 35813
+IHVuanVzdA== 35814
+IHByYXdu 35815
+IGZyb3N0aW5n 35816
+IGRpc2FwcGVhcmFuY2U= 35817
+aW9zYQ== 35818
+IGNhcmRp 35819
+IFByaWVzdA== 35820
+IGNpZW50w61maWM= 35821
+INCS0LDRgQ== 35822
+IOu2gO2DgQ== 35823
+IHRoaWV2ZXM= 35824
+IHBoeXNpcXVl 35825
+IEV1Z2VuZQ== 35826
+INCx0LvQuNC3 35827
+IG1vbm9wb2x5 35828
+IGJpb2dyYXBoeQ== 35829
+IGhvxZ8= 35830
+IHTDtg== 35831
+bWFj 35832
+IHNob2Nrcw== 35833
+7IS46w== 35834
+aGl0 35835
+IHNudWc= 35836
+IGluY2w= 35837
+IGRlZGlj 35838
+IHVsdHJhcw== 35839
+INC40LfQstC10YHRgg== 35840
+IHV0aWxpemF0aW9u 35841
+INGB0L7QstC10YDRiNC10L3QvdC+ 35842
+IHNlcnZp 35843
+c3RhZw== 35844
+MTgw 35845
+IHNld2Vy 35846
+IENob2ljZQ== 35847
+IGRpc2NoYXJnZWQ= 35848
+IEpE 35849
+0L7Qu9C10YI= 35850
+INC60LLQsNGA0YLQuA== 35851
+IHRlbGVzY29w 35852
+IEplxZtsaQ== 35853
+IE5hbmE= 35854
+Y2FsZQ== 35855
+INGC0L7QvQ== 35856
+bW1t 35857
+IGdlaGFidA== 35858
+64Kg 35859
+4LiZ4LiZ 35860
+IGV0aGVy 35861
+IHplbg== 35862
+IHJlc2VhcmNoZWQ= 35863
+IEN6eWxp 35864
+d29ya2Vycw== 35865
+IOqyveywsA== 35866
+IHNoZXJpZmY= 35867
+YWxsbw== 35868
+IHRpcG9z 35869
+IHByb3NlY3V0aW9u 35870
+IGZyb2dz 35871
+IGZhbHQ= 35872
+amQ= 35873
+IO2MlA== 35874
+IGZpbHRlcmVk 35875
+IE9mdA== 35876
+IOyN 35877
+IGRpc2Zy 35878
+IE11c3Rhbmc= 35879
+IHdvYWg= 35880
+IFJFQUxMWQ== 35881
+INC80L7Qs9C70Lg= 35882
+IGVudHJhZGE= 35883
+INC40LPRgNCw 35884
+IG1peGVz 35885
+INCw0LLRgtC+0LzQvtCx 35886
+0Jk= 35887
+IHNoaW4= 35888
+IHBhcmFub3JtYWw= 35889
+IHNvbWVwbGFjZQ== 35890
+IGRpc2hvbg== 35891
+ZXRhYW4= 35892
+IGZ1ZXJ0ZQ== 35893
+2bk= 35894
+IGRvb20= 35895
+7Iic 35896
+IGV4aXN0ZW50aWFs 35897
+IGJ1bGQ= 35898
+IFNESw== 35899
+INC/0YDQsNCy0LTQsA== 35900
+IHR1cm5vdmVy 35901
+IOyXrOq4sOyXkA== 35902
+IOCkuQ== 35903
+IG1vZGVsZWQ= 35904
+IGJ1Z8O8bg== 35905
+IGV4cGVyaW1lbnRhdGlvbg== 35906
+IG1vcm5pbmdz 35907
+IG1lZG8= 35908
+U3Rldmll 35909
+IHBsYXlhYmxl 35910
+IGFpcmxpbmVz 35911
+Z21lbnRz 35912
+IOq4sOu2hA== 35913
+IFRvbWI= 35914
+IE1WUA== 35915
+QVVESUVOQ0U= 35916
+IGNoZWNrb3V0 35917
+IHBhc3N0 35918
+IGJlaXNwaWVs 35919
+IExpbmtz 35920
+aGVhdnk= 35921
+IHF1ZXN0aW9uYWJsZQ== 35922
+IOyTsOs= 35923
+IHNpbGw= 35924
+IG1hbmlwdWxhdGVk 35925
+IExvcmVu 35926
+IOycvA== 35927
+IHZlcmdl 35928
+w6Fr 35929
+SUVT 35930
+IHNhYm90 35931
+IEN1c3RvbWVy 35932
+YWxlxbx5 35933
+IG5vbWluZWU= 35934
+IEdhZA== 35935
+IG5vdXZlbGxlcw== 35936
+IFNQRQ== 35937
+aXN0bGluZw== 35938
+IG92YWw= 35939
+0L7QsdGA0LDQtg== 35940
+aWZ0eQ== 35941
+IGJlemVs 35942
+eWV0 35943
+IGZyZWlnaHQ= 35944
+IEhhbsSxbQ== 35945
+csOtYQ== 35946
+IHpvbmluZw== 35947
+IGluZGVt 35948
+IELDvA== 35949
+IGZlbWluaXNt 35950
+IHZvaXg= 35951
+IG9maWNpYWw= 35952
+IGRpeW9ydW0= 35953
+u5A= 35954
+IGFyb3Nl 35955
+IHBhcmFy 35956
+7J247KeA 35957
+IE1hcnRpbmU= 35958
+IExlY3Q= 35959
+IHJlc3Rlcg== 35960
+IGRyb3duaW5n 35961
+dXlh 35962
+Y2lkYQ== 35963
+IEFyaWVs 35964
+IDAy 35965
+INeU15Q= 35966
+IFdlcnQ= 35967
+0KLRiw== 35968
+IHdpZG93 35969
+IHBhcmNobWVudA== 35970
+IGNvdHRhZ2U= 35971
+IFhM 35972
+IFNsYWNr 35973
+IE5FUw== 35974
+IHJvYmU= 35975
+IGdpbW0= 35976
+IGNhbWluaG8= 35977
+IEhhcnBlcg== 35978
+IGNpdHJ1cw== 35979
+IGZpcmVmaWdodGVycw== 35980
+IGRvcGFtaW5l 35981
+ZWxldHM= 35982
+IGRlbW9jcmF0 35983
+7KCc66Gc 35984
+IHBsYXliYWNr 35985
+b2o= 35986
+INC/0YDQvtC6 35987
+IFN1bGxpdmFu 35988
+c2VtYmxl 35989
+IFdvcnRo 35990
+IE11c3RhZmE= 35991
+4Liy4Lij 35992
+IG1ldHM= 35993
+0LvQvtGB0Yw= 35994
+IGluZXJ0aWE= 35995
+IHVuaWZvcm1z 35996
+w6lyaW8= 35997
+15XXqNeU 35998
+w6ludA== 35999
+IOCukg== 36000
+INGB0LDQvNGL0YU= 36001
+IHZvdWxhaXM= 36002
+IFppbW1lcg== 36003
+6rKg6w== 36004
+INC90L7RgQ== 36005
+ZW5jaWFz 36006
+IHJlbGFjacOzbg== 36007
+IOqxuOs= 36008
+IGZhY3Rpb24= 36009
+IGdvc3A= 36010
+0L/QvtC70L7Qtg== 36011
+bmFw 36012
+aGFr 36013
+IHByb2NlZWRpbmdz 36014
+IOyGlA== 36015
+7JWE64uI 36016
+IOyekOq4sA== 36017
+IHdlcmQ= 36018
+IHNvZg== 36019
+IHNjaGxpbQ== 36020
+IGZsYXZvcmVk 36021
+IHF1YWRyYXRpYw== 36022
+IEJvb3Q= 36023
+IHB1YmxpY2l0eQ== 36024
+IENhcm8= 36025
+ID8i 36026
+0L3QuNGG0LA= 36027
+bWFuaWE= 36028
+IFNVUg== 36029
+IEJVUg== 36030
+bGFuY2U= 36031
+w6l0aWNh 36032
+IHpvYmFjenk= 36033
+IHRyaW8= 36034
+c2FtYQ== 36035
+IHRhxZ8= 36036
+IGFzeW1t 36037
+cmVzc2Vy 36038
+INiq2Lk= 36039
+INC/0LXRgQ== 36040
+IGJlZ2lubmluZ3M= 36041
+bGFkxLFt 36042
+INCx0YvRgdGC0YA= 36043
+IG1vbw== 36044
+IEdlbmV2YQ== 36045
+ZXJ1cw== 36046
+Ym9yYWg= 36047
+IHJlZnVzaW5n 36048
+YnVsbA== 36049
+IFdhaXRpbmc= 36050
+IEluZGl2aWR1YWw= 36051
+IGFub255bQ== 36052
+aW1lbnM= 36053
+IG1lZGlkYXM= 36054
+IGZyYWdyYW50 36055
+IGRpcmVjdGVtZW50 36056
+IOyVhOuniA== 36057
+dXJpYQ== 36058
+IHNwaGVyaWNhbA== 36059
+IGFiZ2U= 36060
+IFZpY3Rvcmlhbg== 36061
+IHNwZWN0YWNsZQ== 36062
+IFJvZHJpZ3Vleg== 36063
+IG9jdXA= 36064
+IE7DpHI= 36065
+bWFya3M= 36066
+bmd1bG8= 36067
+IEx1Y2k= 36068
+IHNob3V0ZWQ= 36069
+IHJlZ3VsYXRvcnM= 36070
+xJ9pbmk= 36071
+IGRpc2VudA== 36072
+INGA0YvQvQ== 36073
+64Ko 36074
+IOyCtOs= 36075
+IHByb2Jsw6htZXM= 36076
+IEZpbmdlcg== 36077
+YXNzZW1ibGU= 36078
+IHBlYXI= 36079
+IGRyb2l0ZQ== 36080
+IEV2ZXJ5d2hlcmU= 36081
+dGFt 36082
+0L7RgtC40LI= 36083
+0LLQvtC5 36084
+b3JkaW5hdGU= 36085
+IExhaw== 36086
+IG3hu5tp 36087
+IFRlbGV2aXNpb24= 36088
+IGV4cG9uZW50aWFsbHk= 36089
+YXZhcw== 36090
+IGJsZXY= 36091
+IE1U 36092
+Q29ubmVsbA== 36093
+IOq1reuvvA== 36094
+INGB0LLQvtC40Lw= 36095
+IGFjaGE= 36096
+IER5bmFzdHk= 36097
+Smlu 36098
+IHRvcmU= 36099
+IGZsb3I= 36100
+INC80L3QvtCz0LjQtQ== 36101
+b3dhbg== 36102
+YmFo 36103
+IOyjhA== 36104
+IENlbGE= 36105
+IOy1nOq3vA== 36106
+IHBlcm1ldHRyZQ== 36107
+IGFicmFz 36108
+IHZlcnN0ZWhlbg== 36109
+IGVzY29ydA== 36110
+IFRoZW0= 36111
+w6Rya2U= 36112
+cG9ydGVy 36113
+IGthaGthaGE= 36114
+IGhlY3Q= 36115
+IGRhdQ== 36116
+d2Fo 36117
+b2x2ZQ== 36118
+IEFnZXM= 36119
+c2NoYWZ0 36120
+IFN0ZWxs 36121
+bmVsbGU= 36122
+IEVuc3VpdGU= 36123
+INCS0YHQtdC8 36124
+IGNyw6lk 36125
+IFBQ 36126
+bG9yZHM= 36127
+Z3J1bnRpbmc= 36128
+IGNvbnRyYWN0aW9u 36129
+R290 36130
+IGFjcXVpcmluZw== 36131
+IHNvcHI= 36132
+IHBvaXNvbm91cw== 36133
+Uk5B 36134
+IGFuYXI= 36135
+IEhvZg== 36136
+Jyk= 36137
+IHJlbWFya2FibHk= 36138
+IGludGVybmFjaW9uYWw= 36139
+w7xja2U= 36140
+aW5xdQ== 36141
+IGR1eQ== 36142
+IGJlYXN0cw== 36143
+IExBTg== 36144
+IHByZWNlZGVudA== 36145
+IFJQTQ== 36146
+IHNlbG9u 36147
+IG1vcnRl 36148
+IGNvbWXDp291 36149
+0Y/Qu9Cw 36150
+IGludGVycHJldGluZw== 36151
+IEJ1cmtl 36152
+0YLRgNCw 36153
+IOydtOufrA== 36154
+IHBlc3NpbQ== 36155
+IE5vaw== 36156
+7Yyd 36157
+RmVtYWxl 36158
+IOyLpO0= 36159
+mYA= 36160
+IHN0aW11bGF0aW9u 36161
+IHNsaWNr 36162
+IOqwgOuKlA== 36163
+INC60LDQtw== 36164
+IEhCTw== 36165
+IHBhcGllcg== 36166
+IGvDtm5udGVu 36167
+0YPQsdC70Lg= 36168
+IENvbnN0YW50 36169
+U1BFQUtJTkc= 36170
+IGt0w7NyxIU= 36171
+IGNvc21ldGljcw== 36172
+IFRyZW5k 36173
+IHJvYmJlcnk= 36174
+IHRpdHQ= 36175
+IGdqb3J0 36176
+IGRpZXRhcnk= 36177
+oIw= 36178
+IEtpcmJ5 36179
+INC/0YDQuNC80LXRgNC90L4= 36180
+IHF1YWxpZmljYXRpb24= 36181
+IOyViQ== 36182
+IGNhYmluZXRz 36183
+IGh0dHA= 36184
+IEVyaWNh 36185
+IGRpc2FkdmFudGFnZXM= 36186
+IGNoYXR0ZXJpbmc= 36187
+eXo= 36188
+ZmVpdA== 36189
+IGd1aWxk 36190
+IEVURg== 36191
+IERyYWdvbnM= 36192
+IEhFUkU= 36193
+dmVudGg= 36194
+2YTYp9mF 36195
+IG1hcmNow6k= 36196
+RGFt 36197
+IHBob3Rvbg== 36198
+IGVzdGFibGU= 36199
+TWFn 36200
+IG9saGFy 36201
+IGNvdXBsaW5n 36202
+IEhpbGZl 36203
+IFdpemFyZA== 36204
+INC80LDQu9C+ 36205
+aGVscA== 36206
+IGzDrW5lYQ== 36207
+IOyr 36208
+IHN0YW5kYWxvbmU= 36209
+IG1vcmFsZQ== 36210
+IHp3ZWl0ZQ== 36211
+w6RocnQ= 36212
+IGRvdHRlZA== 36213
+IGRyaXBwaW5n 36214
+IEZsYWc= 36215
+cm9ja2V0 36216
+cmF0ZWd5 36217
+aXJpbQ== 36218
+IO2VmOuptOyEnA== 36219
+IHNvZ2VuYW4= 36220
+IFVubw== 36221
+IFNjaHV0eg== 36222
+IGVzdGlsbw== 36223
+IFN1YnM= 36224
+IERhaXN5 36225
+0J3QtdGC 36226
+Jy4uLg== 36227
+IHBsYXRpbnVt 36228
+IGJpcmw= 36229
+IFNvdmk= 36230
+IHZpb2xhdGU= 36231
+0YPQtdGC0YHRjw== 36232
+cmlsbA== 36233
+IHRyYXo= 36234
+IHNuaXA= 36235
+IGN1bXBs 36236
+4Lit4LiB 36237
+IGN1aw== 36238
+IFBhcmxhbWVudA== 36239
+IGh5cGVydA== 36240
+IHB1bHA= 36241
+IHRvbmd1ZXM= 36242
+YXR0bw== 36243
+IGJ1c2Nh 36244
+aWhu 36245
+RVJP 36246
+INmK2Lk= 36247
+IHZhcmlhcw== 36248
+IE1hcmlhbg== 36249
+IGJvdW5kZWQ= 36250
+IHBpdGNoaW5n 36251
+IGRlZmljaWVuY3k= 36252
+IEJsZXNzZWQ= 36253
+IEV4ZXJj 36254
+dWNocw== 36255
+IG5oxrBuZw== 36256
+IHJhcGVk 36257
+aGFsZXM= 36258
+IG1hbGE= 36259
+cGlj 36260
+IDQwMQ== 36261
+xZtuaWVq 36262
+YXJpbmE= 36263
+65Ok7J2E 36264
+b3R0aQ== 36265
+INC00L7Qu9Cz0L4= 36266
+IHRyYWNrZXI= 36267
+IFNoZWxieQ== 36268
+IHZhbmlzaGVk 36269
+IGJha2VyeQ== 36270
+S2FwxLE= 36271
+SmVzdXM= 36272
+IEtS 36273
+Sk8= 36274
+hbg= 36275
+IGRpc2Nz 36276
+7ISv 36277
+7KeA6w== 36278
+15nXpg== 36279
+ZW1hcnk= 36280
+S2VuZHJh 36281
+IHnDvGs= 36282
+w7xja3Q= 36283
+IHZheg== 36284
+IGt1cA== 36285
+YWt0dQ== 36286
+INGB0L/QsNGB0LjQsdC+ 36287
+IGFpaw== 36288
+IG51cnNlcnk= 36289
+IGVuZGFuZ2VyZWQ= 36290
+w6ptZW1lbnQ= 36291
+ZW1hdGljcw== 36292
+IHJlc3BvbmRlcnM= 36293
+IFJlcHJlc2VudGF0aXZlcw== 36294
+IHNjdWxwdHVyZXM= 36295
+aWdrZWl0ZW4= 36296
+IGRlcGw= 36297
+IGludGVycHJldGF0aW9ucw== 36298
+IGRlYWRsaW5lcw== 36299
+IDE5NDI= 36300
+w5c= 36301
+IHN1Z2Fycw== 36302
+ZW11 36303
+bGl2ZWx5 36304
+IHJlY3JlYXRpb25hbA== 36305
+IGRpc3RvcnQ= 36306
+IHVuZGVyc2NvcmU= 36307
+IHVucXVvdGU= 36308
+IHNhZmVzdA== 36309
+IHN3b2xsZW4= 36310
+IGFuYWx5c2Vz 36311
+IGNvbW1lbmPDqQ== 36312
+YW5kaW4= 36313
+INCl0L7RgNC+0YjQvg== 36314
+IGRpYXJy 36315
+emllc3Q= 36316
+IHRvb3RoYnJ1c2g= 36317
+dWF0aW9ucw== 36318
+IGNhZGU= 36319
+IGJhY2tsYXNo 36320
+aGluZA== 36321
+IHJpc3F1ZQ== 36322
+emVzcw== 36323
+IOydtOyVvOq4sA== 36324
+IGVzcGVyYXI= 36325
+IHRyYW5zbGF0aW9ucw== 36326
+aW9uZWQ= 36327
+Z3JvYW5z 36328
+INC/0YPRgg== 36329
+IGdlbmV0aWNhbGx5 36330
+IGhhcHBpZXN0 36331
+IHdlcms= 36332
+YXRvb24= 36333
+IG11c2k= 36334
+IGZ1bsOnw6Nv 36335
+IOyeheuLiOuLpA== 36336
+INGA0LDQuQ== 36337
+IGJldm9y 36338
+QkxBTks= 36339
+IHJlcGVudGFuY2U= 36340
+UHV0 36341
+IHBvdHJ6ZWI= 36342
+IHNhbGE= 36343
+IGNhbXBh 36344
+V0VS 36345
+IGRlY8OtYQ== 36346
+IHPDqWN1cml0w6k= 36347
+IEFwcHJlY2lhdGU= 36348
+0YfQuA== 36349
+IFJhbmRvbQ== 36350
+67OE 36351
+a2Fo 36352
+IG3Dtmo= 36353
+IHPDpGdlcg== 36354
+INeZ15vXldec 36355
+IDE5MA== 36356
+eHR1cmVz 36357
+RXU= 36358
+IGfDpA== 36359
+INeR16o= 36360
+IENyb2F0 36361
+YXBv 36362
+UExF 36363
+IHBlcnNpc3RlbmNl 36364
+IGJsZW5kcw== 36365
+IHRyZWZmZW4= 36366
+IFNhbnRpYWdv 36367
+eWRpYQ== 36368
+YWxkbw== 36369
+IFRlbnNvckZsb3c= 36370
+IER1YWw= 36371
+IGNoaWZm 36372
+7Je0 36373
+IGNvbnRyYWN0ZWQ= 36374
+IHNlZ3JlZw== 36375
+IEZhaXJ5 36376
+IHdpc2VseQ== 36377
+IHZ1bG5lcmFiaWxpdGllcw== 36378
+IGhhbmRoZWxk 36379
+IGdhZGdldHM= 36380
+IGJvxZ8= 36381
+IFBvcHVsYXI= 36382
+IGN1cnZhdHVyZQ== 36383
+66y4 36384
+IE1BUlk= 36385
+7J207Io= 36386
+IGZvcm11bGF0aW9u 36387
+IGNlbGVyeQ== 36388
+IGJsdXJyeQ== 36389
+IFRT 36390
+YWxleg== 36391
+IHdz 36392
+IHByb2dyYW1t 36393
+IFN0YWNr 36394
+IEpJTQ== 36395
+0L7QstCw0LvQuA== 36396
+xLFsbA== 36397
+IHDDqHJl 36398
+IEthbnll 36399
+IERlbGF3YXJl 36400
+IGRhdW50aW5n 36401
+INCx0LXRgQ== 36402
+IFN0dXBpZA== 36403
+Ymln 36404
+ZmZpY2lhbA== 36405
+IHByZWNpcGl0YXRpb24= 36406
+IHBsdW5n 36407
+4bulYw== 36408
+YnVyc2U= 36409
+IGRhcmxl 36410
+IGNyaXBw 36411
+IHBpb25lZXI= 36412
+IGRpc3B1dA== 36413
+IHNlYW4= 36414
+IHJlc2lzdG9y 36415
+IGFsbGVpbg== 36416
+aXBwbGVz 36417
+YXJlbA== 36418
+IGVuZG9ycw== 36419
+enVzdA== 36420
+INGA0LXQsdGP0YLQsA== 36421
+ZWRlZA== 36422
+IOy5tOuplOs= 36423
+IGxsZXZh 36424
+IGtlbm50 36425
+INCx0LDQuw== 36426
+IERvY3VtZW50 36427
+IEtuaWdodHM= 36428
+IGJ1Y2tsZQ== 36429
+IOyJrA== 36430
+IGFsaw== 36431
+IEV2ZXJ5ZGF5 36432
+YXR0ZXJz 36433
+IHRvaWxldHM= 36434
+IGp1Z2Fy 36435
+IOyeiOyngA== 36436
+IGdlbmF1c28= 36437
+IExhbmRlc3JlZ2llcnVuZw== 36438
+aWpl 36439
+IHRyYWlsZXJz 36440
+IFRpZ2Vycw== 36441
+IGdpdHRp 36442
+IGZvcmdpdmluZw== 36443
+IGNvbmN1cnJlbnQ= 36444
+IFZ1 36445
+IO2Kue2eiA== 36446
+IEJST1dO 36447
+b3VuZGVk 36448
+Ijs= 36449
+IHRyZW1i 36450
+IHRpZXQ= 36451
+INGA0LXQttC40Lw= 36452
+IG51dHNoZWxs 36453
+0LXQu9C40Yc= 36454
+IGxvc2Vycw== 36455
+cmljdGluZw== 36456
+IHJlZGVlbQ== 36457
+ZGVmaW5lZA== 36458
+TmljZQ== 36459
+IGJyb2FkYmFuZA== 36460
+S08= 36461
+IHRlYXNpbmc= 36462
+IHBhcnRpc2Fu 36463
+xLFtYQ== 36464
+IOyerOuvuA== 36465
+IEpvdXJuZXk= 36466
+IHNsb3Blcw== 36467
+dW5pbmc= 36468
+Z3J1bnRz 36469
+IHTDpGxs 36470
+IHVuY292ZXJlZA== 36471
+IG15xZtsxJk= 36472
+IEVzdGhlcg== 36473
+IEhlYWx0aHk= 36474
+IOuwkQ== 36475
+csOpZQ== 36476
+IHBvbGFyaXphdGlvbg== 36477
+IGZsYXY= 36478
+IGNhbWJpYXI= 36479
+IHly 36480
+IFJhbmNo 36481
+IHNwbGl0cw== 36482
+IHRyb3V2w6k= 36483
+IHJlY29yZGVy 36484
+IGTDqXBhcnQ= 36485
+2YjYqA== 36486
+IEtyeQ== 36487
+IGludGVyZXNzYW50 36488
+IGVkZXJpbQ== 36489
+xZt3aWFk 36490
+aWxhdGVyYWw= 36491
+d3JpZ2h0 36492
+IHBvdXJyYQ== 36493
+w6p0ZXI= 36494
+IGNhbWVs 36495
+4Z4= 36496
+IHJhcGlkZW1lbnQ= 36497
+IG1lag== 36498
+IHN0aWZmbmVzcw== 36499
+QURBUw== 36500
+IGRpZmZlcnM= 36501
+IGFsb3Q= 36502
+IFNpZw== 36503
+0Y/RgtC10LvRjA== 36504
+IGFic3RyYWN0aW9u 36505
+IGtlaW5lcg== 36506
+Z3J1cHA= 36507
+IFNoZXJsb2Nr 36508
+7ZiU 36509
+IGNpdGU= 36510
+IG92ZXJmbG93 36511
+IHThuqFp 36512
+w7pjYXI= 36513
+YnVsYQ== 36514
+IGNvbmp1bnRv 36515
+IENJ 36516
+IG1vZGVyYXRvcg== 36517
+IGluZGlyZWN0bHk= 36518
+IGFsbGVpbmU= 36519
+4oI= 36520
+0YjQuNCx 36521
+INCx0LDQsQ== 36522
+IGRhbmFjaA== 36523
+IDE5Mzk= 36524
+IHByb21ldA== 36525
+IGRlc3RpbmF0aW9ucw== 36526
+IElsbHVzdA== 36527
+zrnOus+M 36528
+IHNhYmVz 36529
+IGhlaA== 36530
+IEdlc2V0emVudA== 36531
+IE1peg== 36532
+0LXQvdC60L4= 36533
+IE15cw== 36534
+0Kw= 36535
+IEp1ZGFpc20= 36536
+IG11c3RhY2hl 36537
+IHN0aW1tdA== 36538
+IEdhemE= 36539
+IHZvbHRl 36540
+IG51bw== 36541
+IG3Ds24= 36542
+IENvbXB1dA== 36543
+4Li54LmI 36544
+IFJhZGk= 36545
+IGV4Y2VwdGlvbmFsbHk= 36546
+IGFzc3VtZXM= 36547
+aW5mb3Jt 36548
+IHNocmluZQ== 36549
+IGltcGxpY2F0aW9u 36550
+IEZpdHo= 36551
+IS4= 36552
+IGx0 36553
+IGFsbG95 36554
+IGV0aGlj 36555
+IG1vbmFzdGVyeQ== 36556
+7Iuc7KOg 36557
+aWNhw6fDo28= 36558
+IGNvb3JkaW5hdGluZw== 36559
+IE1vdG8= 36560
+IG92ZXJsb29r 36561
+IGNob2lz 36562
+IGFudGliaW90aWM= 36563
+IE1pbm5l 36564
+IEJK 36565
+IEFwYQ== 36566
+b3JpYW4= 36567
+IHNwaWxsZWQ= 36568
+SmFt 36569
+IGh1c2JhbmRz 36570
+IGNyZWF0aW9ucw== 36571
+IGHDsQ== 36572
+w7xzc2Vs 36573
+IOydtOyaqQ== 36574
+IGFuYWx5c2U= 36575
+cm9zZQ== 36576
+IHB1bmNoZWQ= 36577
+IHByZXNxdWU= 36578
+IGFzdHJvbm9teQ== 36579
+IHNjaHdpZXJpZw== 36580
+IEVib2xh 36581
+IGNpcw== 36582
+IGFjZXQ= 36583
+IEZY 36584
+ZW5kcmU= 36585
+IOydjOyVhQ== 36586
+IHdlYnBhZ2U= 36587
+IGZyZWFrZWQ= 36588
+IGxhdHRl 36589
+IOy/oA== 36590
+IOuouOs= 36591
+TmV2ZXI= 36592
+R3Jh 36593
+7ZmU66W8 36594
+ZXllZA== 36595
+IOuwnOudvA== 36596
+IGVzcGVyYQ== 36597
+IGFwYXJlY2U= 36598
+cmHDp8Ojbw== 36599
+IGRpc3J1cHRpdmU= 36600
+IEpvaW50 36601
+dXJvdXM= 36602
+cmVhcw== 36603
+IHF1ZXLDrWE= 36604
+IGRpc3RyaWJ1dGlvbnM= 36605
+IGV4cG9uZW50 36606
+7LmY66W8 36607
+IGRs 36608
+emhvdQ== 36609
+IEhlYXJpbmc= 36610
+IENyYXc= 36611
+IGZsb2F0cw== 36612
+b3VuY2Vk 36613
+TGFi 36614
+V29ybGQ= 36615
+IGJ1cmRlbnM= 36616
+IGF1dGhvcml0YXJpYW4= 36617
+IEJvbHQ= 36618
+INC+0LTQvdGD 36619
+IHBpZ2Vvbg== 36620
+IGRpc3RyYWN0aW9ucw== 36621
+IEhlcmF1c2ZvcmRlcg== 36622
+IHplc3Q= 36623
+ZXNj 36624
+IHNoYWtlcw== 36625
+YXRhcw== 36626
+INmF2LQ= 36627
+aG9sZXM= 36628
+IHRoaW5rZXJz 36629
+YWx0YQ== 36630
+IGFyY2hl 36631
+IFN1aw== 36632
+YW5oYQ== 36633
+IHRlbXB0aW5n 36634
+IHlvdXR1YmVy 36635
+IHbDrA== 36636
+IGR6aWHFgmE= 36637
+IFZhdGljYW4= 36638
+UGFyaw== 36639
+IHN1cGVycw== 36640
+IE5pa2tp 36641
+64qQ6w== 36642
+b3Jhbmc= 36643
+cmFtaWVudA== 36644
+IOqwluqzoA== 36645
+IGRlc3NlcnRz 36646
+IGF2ZXJl 36647
+IEdyZWdvcnk= 36648
+IOuTpOyWtOyY 36649
+IGNvc3Rpbmc= 36650
+IENsaW5pYw== 36651
+IHJlYmVscw== 36652
+IE1vYg== 36653
+IGJ1bmxhcg== 36654
+IFlvdXJz 36655
+ZXJ0aW1l 36656
+IHJldGFsaQ== 36657
+bWFyYQ== 36658
+YXR1cw== 36659
+YWxsZXM= 36660
+INC00YA= 36661
+INC00LjRgQ== 36662
+IGRpc2NvdW50cw== 36663
+IEdVWQ== 36664
+INC60LDQutC+0LU= 36665
+IEV4cGVyaW1lbnQ= 36666
+cmVtZW50 36667
+IFhpYW5n 36668
+IGJhdGU= 36669
+V0U= 36670
+IHNwZWNpYWxpemU= 36671
+IGRlaXR5 36672
+IExva2k= 36673
+bWFn 36674
+IE5pdA== 36675
+V2VzdA== 36676
+IG1hdGVybmFs 36677
+IHF1aXM= 36678
+YnJva2Vu 36679
+IGxhc2Vycw== 36680
+IGhha2s= 36681
+IEFuZ2Vscw== 36682
+IG1hc3Rlcnk= 36683
+YW50aXM= 36684
+VGlmZmFueQ== 36685
+ZWVl 36686
+55E= 36687
+b3JlbQ== 36688
+IGluYWNj 36689
+IGp1cmlzZGljdGlvbnM= 36690
+IEthcmRhc2g= 36691
+SWw= 36692
+IFNpbm4= 36693
+IGF0aGxldGljcw== 36694
+Y8SZ 36695
+IGxvb3NlbHk= 36696
+IGRpZXRh 36697
+QWc= 36698
+ID8/ 36699
+IOuMgO2RnA== 36700
+IHN1cGVydg== 36701
+IG51dHJpdA== 36702
+IGRyaWZ0aW5n 36703
+IOyEoOyDneuLmA== 36704
+INC/0L7QvdGP0Ls= 36705
+IFZpY3Rvcnk= 36706
+2YTYqQ== 36707
+15XXoNeU 36708
+INC/0LjRiA== 36709
+IHNoYXZlZA== 36710
+IG1lc3VyZQ== 36711
+b25kZW4= 36712
+2YPYsQ== 36713
+IGV4aWxl 36714
+IERlc2Rl 36715
+IFBpbnRlcmVzdA== 36716
+IGF0dGFjaG1lbnRz 36717
+IGhvbWJyZXM= 36718
+IGZpbmVz 36719
+IOyEuOyDgQ== 36720
+IHNsZWVwcw== 36721
+IFRhY28= 36722
+IElSQQ== 36723
+cmlvcw== 36724
+IG9sbA== 36725
+ZXRlcw== 36726
+IHVudXQ= 36727
+ZmFzaGlvbmVk 36728
+IHRyZWJhbGw= 36729
+IE5lYXJseQ== 36730
+INGA0LXQsNC70YzQvdC+ 36731
+IGNoaWw= 36732
+xJ9h 36733
+IE1FTA== 36734
+cm9zY29w 36735
+IENH 36736
+IHZlbmdl 36737
+IGRpc2h3YXNoZXI= 36738
+YWxnaWM= 36739
+IG1vZGlmaWVy 36740
+IGVtYmFzc3k= 36741
+dGltZXI= 36742
+ZW1pY3M= 36743
+IGludHJpY2F0ZQ== 36744
+IGV2ZXQ= 36745
+IOuMgOuwlQ== 36746
+IGlzb3Q= 36747
+INC90LDRg9GH 36748
+IFF1aXo= 36749
+cmVzbw== 36750
+zrTPjg== 36751
+IHllbGxlZA== 36752
+IGZlZGVy 36753
+RUxMRVI= 36754
+IGV4Y2VlZGVk 36755
+b25hcw== 36756
+aWNhbm8= 36757
+INC20LjQstC+0YI= 36758
+IE1hbw== 36759
+IEthenV0bw== 36760
+IOOFi+OFi+OFi+OFiw== 36761
+IGZyb250bGluZQ== 36762
+IEh1bmdhcmlhbg== 36763
+IMO8YmVyYWxs 36764
+YXdhdA== 36765
+IGdyaXBz 36766
+acOnw7Vlcw== 36767
+YXJueWE= 36768
+IM2h 36769
+IHNlaWQ= 36770
+IGFuYWs= 36771
+IGFjYWJvdQ== 36772
+7ZWR 36773
+IG5vdG9yaW91cw== 36774
+IEdvZHppbGxh 36775
+IG92ZXJjb21pbmc= 36776
+IFBlbmQ= 36777
+IG9sYWJpbGly 36778
+w7xsbWU= 36779
+IGVyaGFsdGVu 36780
+6re5 36781
+IE1ldGVy 36782
+IHN0YWFu 36783
+T2w= 36784
+IGNoYXRz 36785
+IEJ1ZW5vcw== 36786
+w612ZQ== 36787
+YWx1YWJsZQ== 36788
+IHN0cmF0ZWdpY2FsbHk= 36789
+IGNvbXByaXNlZA== 36790
+INC/0LXRgNGB0L7QvdCw0LY= 36791
+IHdhbm4= 36792
+IENlbg== 36793
+0L3QuNGC0LU= 36794
+n4E= 36795
+INGC0L7QsdC+0Lk= 36796
+aWFk 36797
+IGthcmRlxZ9pbQ== 36798
+IENvbmdyZXNzbWFu 36799
+cmVhbWluZw== 36800
+aG9tbWU= 36801
+IGNvbW11bmF1dA== 36802
+IGFsY29ob2xpYw== 36803
+IHBpY2tsZWQ= 36804
+IGFjb3Jk 36805
+cG9zaXRpb24= 36806
+ZWfDs2w= 36807
+IHRyb3VibGluZw== 36808
+IE1hcmNoZWc= 36809
+IHp1bWluZGVzdA== 36810
+IHNlYW1sZXNzbHk= 36811
+IG9sdW4= 36812
+IFRWcw== 36813
+INC/0YDQsNC60YLQuNGH0LXRgdC60Lg= 36814
+IGJhY2tlbmQ= 36815
+aWRhYmxl 36816
+IGdhZGdldA== 36817
+IGZhw6dv 36818
+IE1hcmNoZWdpYW5p 36819
+IOuwpA== 36820
+IGFjY2lkZW50YWw= 36821
+IExQ 36822
+IGVsZGVzdA== 36823
+IEFkbWlyYWw= 36824
+IG7Eg20= 36825
+bGV2ZXI= 36826
+IHBhc3RlbA== 36827
+IGZvbmRv 36828
+Q29ubmll 36829
+IHRlcmNlcg== 36830
+IHBhY3Q= 36831
+IE1vbnRl 36832
+IG1lYXRz 36833
+IFNNUw== 36834
+IEF1c3RyYWxpYW5z 36835
+57w= 36836
+UmhldHQ= 36837
+IGV4YWN0ZW1lbnQ= 36838
+IOu5vA== 36839
+IE1PRA== 36840
+56E= 36841
+IFJhcHQ= 36842
+IE5vY2g= 36843
+IGFib3J0 36844
+IE5hdmFs 36845
+IEZ1amk= 36846
+SU5URVI= 36847
+INC90L7QstGL0Lk= 36848
+IG1pZWpzY2U= 36849
+IElDVQ== 36850
+IEdyYWR1YXRl 36851
+IEdsZW4= 36852
+YXJkaQ== 36853
+IMiY 36854
+IHNvbGRlcg== 36855
+IHByb2Zlc3Npb25z 36856
+IG9ydGhvZw== 36857
+b21u 36858
+aW50cm9kdQ== 36859
+IERlbmlzZQ== 36860
+7J6Q66W8 36861
+IGNvcnJlc3BvbmRlbmNl 36862
+QU1B 36863
+IGluZmxpY3Q= 36864
+IGZhbmQ= 36865
+IEfDvA== 36866
+INGH0LXRgg== 36867
+IHRyYWNlZA== 36868
+IHBhdGVudHM= 36869
+IGFtYnVzaA== 36870
+IGxvdHRh 36871
+ZmZlcg== 36872
+IFdhZ25lcg== 36873
+IGltcGVyc29u 36874
+IGV4dHLDqm1lbWVudA== 36875
+2YLYqg== 36876
+Y29uZHVjdA== 36877
+QXR0 36878
+IE11ZWxsZXI= 36879
+IEFsaWNpYQ== 36880
+IGN5Yw== 36881
+IGhhY2tlcg== 36882
+IHR5cw== 36883
+IGhhaWw= 36884
+INC30LDRj9Cy 36885
+IHBhc3Nv 36886
+IOy2lOqwgA== 36887
+IM6I 36888
+IHBhY2thZ2Vk 36889
+IEN5bnRoaWE= 36890
+aGVldA== 36891
+IE5pc3Nhbg== 36892
+IFF1ZXN0bw== 36893
+6ag= 36894
+ZGlk 36895
+IM68zrnOsQ== 36896
+IEVsbGlz 36897
+IEFuYWx5c2lz 36898
+Y2Vtb3M= 36899
+IGFzZWc= 36900
+IE15c3Rlcg== 36901
+IENhbw== 36902
+IHR1dg== 36903
+IEluZHVzdHJ5 36904
+7KO86rOg 36905
+b3RhbA== 36906
+IHBlcXVlw7Fv 36907
+YnJhcw== 36908
+IGNvbXByZWhlbmQ= 36909
+IFNpbXBzb24= 36910
+0YHRgtCy0LjQtQ== 36911
+b2NyYWN5 36912
+0LjRh9C10YHQutC4 36913
+IE11c2g= 36914
+IExhdXJpZQ== 36915
+IHRyaWFuZ3VsYXI= 36916
+IFByZXNlbnRz 36917
+IEt1bmRlbg== 36918
+IElzcw== 36919
+IERlY2s= 36920
+4buDbg== 36921
+IERhcmtuZXNz 36922
+IGluZmxhbW1hdG9yeQ== 36923
+ZXJlbWlhaA== 36924
+IHdhcm1lZA== 36925
+dmV5YXJk 36926
+IE1lbW9yeQ== 36927
+ZXR0eQ== 36928
+IHRheHBheWVycw== 36929
+4LiT 36930
+2KE= 36931
+IHByYWN0aXNl 36932
+64us6w== 36933
+IGRyaWxsZWQ= 36934
+bcO8xZ8= 36935
+bG9nbw== 36936
+IEZhY2g= 36937
+pOuhnA== 36938
+IMO8YnJpZ2Vucw== 36939
+IGtvbm50ZW4= 36940
+IG5vcm1hbG1lbnRl 36941
+IGFyZ3Vlcw== 36942
+aWxpbmd1YWw= 36943
+sOulvA== 36944
+ZWdhbA== 36945
+IHRyYXZhaWxs 36946
+b3Z5 36947
+0LDRgtC+ 36948
+IHJ1dGg= 36949
+IExpZ2h0cw== 36950
+IGNvbnNpc3RlZA== 36951
+15HXqNeZ150= 36952
+IHN0ZXJlb3R5cGU= 36953
+IHBheWVy 36954
+IFJlZQ== 36955
+IEFpcmJuYg== 36956
+IGRyb3duZWQ= 36957
+IFpvZQ== 36958
+IGNhbm9weQ== 36959
+IGJhcnI= 36960
+INC90L7Rhw== 36961
+IHBhZ2Fu 36962
+IGphcnM= 36963
+IHLDqg== 36964
+ZXJ2ZXI= 36965
+aWViZW4= 36966
+IGVzcGVjdA== 36967
+IEZp 36968
+IHVud2lsbGluZw== 36969
+IHRlY2huaWNpYW4= 36970
+4bq3dA== 36971
+bWVtYmVy 36972
+IENhbmFs 36973
+2LPZhQ== 36974
+IGxpZWJlcg== 36975
+IGluZmVyZW5jZQ== 36976
+IGhvbm9yaW5n 36977
+IENhbXBhaWdu 36978
+IGxpbmVhZ2U= 36979
+IFN0cmVzcw== 36980
+IHZpY3Rvcmllcw== 36981
+IGRlamE= 36982
+16M= 36983
+w6p0ZXM= 36984
+YmxpY2s= 36985
+INC80LXQvdC10LU= 36986
+b3Rocw== 36987
+IENvdXBsZQ== 36988
+SmFzb24= 36989
+IE5pY29sYXM= 36990
+0LXQutGB 36991
+bGli 36992
+IGhlcnJhbWllbnQ= 36993
+INeQ15XXnteo 36994
+INCy0LjQtNC40Lw= 36995
+bWlsbGltZXRlcg== 36996
+IHNpbGhvdWV0dGU= 36997
+IGRyaXZld2F5 36998
+IGNoZXJpc2g= 36999
+44Wg44Wg 37000
+IHJhbnNvbQ== 37001
+IGludGVyZGlzY2lwbGluYXJ5 37002
+IFBvcnRhbA== 37003
+IHRyYWc= 37004
+dGhvb2Q= 37005
+IHRlZGlvdXM= 37006
+IGdsb3NzeQ== 37007
+IHByw6lwYXI= 37008
+IENheQ== 37009
+IFRvb2s= 37010
+IEJvdHRvbQ== 37011
+IHppZw== 37012
+5as= 37013
+cmVwcmVzZW50ZWQ= 37014
+4LmA4Lil4Lii 37015
+IGRlc2Fycm9sbG8= 37016
+7ISc6w== 37017
+IHZpc2Nvcw== 37018
+IG1pbGxpZ3JhbQ== 37019
+IEd1bmQ= 37020
+IGZlcm1lbnQ= 37021
+ZHJ1bQ== 37022
+IGRyYXdlcnM= 37023
+TGF1Z2g= 37024
+IHBlbG9z 37025
+IHBhdmVtZW50 37026
+IG1lbW9pcg== 37027
+YXZhaXQ= 37028
+IDIwNTA= 37029
+pOulvA== 37030
+IHJhesOzbg== 37031
+IGZsb3VyaXNo 37032
+IHN0ZXJu 37033
+IENodW5n 37034
+IHNlcnBlbnQ= 37035
+IEdlbnRsZW1lbg== 37036
+a29vaw== 37037
+IGx1dA== 37038
+aW1wb3J0ZQ== 37039
+cGFyZW50 37040
+IHdzeg== 37041
+IHNjcmVl 37042
+IE1pdGFyYmVpdGVy 37043
+bXV0 37044
+IOyWmOq4sOulvA== 37045
+IHNlbWJsZQ== 37046
+IE9X 37047
+IGludmVzdGlnYXRvcg== 37048
+IENoZXJ5bA== 37049
+IEdlcmFsZA== 37050
+IHByZXJl 37051
+IGNvbXBhcmVz 37052
+bnl0 37053
+IGRpZmVyZW7Dp2E= 37054
+Py0= 37055
+IHF1w6E= 37056
+16jXmQ== 37057
+U2Vu 37058
+IGhlcHM= 37059
+IGdyYXR1aXQ= 37060
+IGNvbnNvcnQ= 37061
+IFNUT1A= 37062
+IFByb3Rlc3RhbnQ= 37063
+IGVsZWN0cm9kZQ== 37064
+4pc= 37065
+IHNlY3VyZWx5 37066
+0LjRh9C10YHQutC+0Lk= 37067
+IHTDpMOk 37068
+IHJlZ2lzdGVycw== 37069
+IEhlYXZlbmx5 37070
+b2dseQ== 37071
+aXNzw6Q= 37072
+IFBoeXNpY3M= 37073
+IE1lcmtlbA== 37074
+IHLDqXY= 37075
+IGVyYXNlZA== 37076
+IFNhY3JhbWVudG8= 37077
+IGNvZmZpbg== 37078
+IGV4YWNlcg== 37079
+IGxhbno= 37080
+IHBvZXRz 37081
+dWxpZg== 37082
+IOy5mOs= 37083
+IE5lcmQ= 37084
+IE5DVA== 37085
+IEhvdXI= 37086
+bmVobWVy 37087
+npjrj4Q= 37088
+IFByaW5jaQ== 37089
+U3c= 37090
+bWllcw== 37091
+YXJtZWQ= 37092
+IEJlYXRsZXM= 37093
+IHByb3BhZ2F0aW9u 37094
+IGV4Y2hhbmdlZA== 37095
+IGN1bXVsYXRpdmU= 37096
+IOynkeyXkA== 37097
+IGRlZmVhdGluZw== 37098
+YmVscw== 37099
+IHdlcw== 37100
+IE9keXNzZXk= 37101
+YXZpb3I= 37102
+IOychOyXkA== 37103
+IGJyaXQ= 37104
+IGhpam8= 37105
+REFZ 37106
+INin2YTYqtmK 37107
+INCh0LXRgNCz 37108
+0YPQutCw 37109
+ZWRzacSZ 37110
+IGltcG9z 37111
+IGVsbGFz 37112
+IGZpcmVhcm1z 37113
+IE5S 37114
+INeR15A= 37115
+INCf0L7QutCw 37116
+YXdp 37117
+IOyEseqztQ== 37118
+IHB1cGlscw== 37119
+IFRhY2s= 37120
+IGZyYXNl 37121
+IFNoaXA= 37122
+IHN0YWQ= 37123
+IEdyZWF0ZXI= 37124
+dW51bg== 37125
+aW1tdW5n 37126
+Z3Jvd24= 37127
+IE5YVA== 37128
+IEFtZXJpY2Fz 37129
+Zm94 37130
+IG1hbnRlbg== 37131
+INGB0L7Qug== 37132
+IHJpa3Q= 37133
+bGVjdHJpYw== 37134
+ZGVlcA== 37135
+INC30L3QsNC10YjRjA== 37136
+IGJlbnV0 37137
+IEluZnJhc3Q= 37138
+IEVtaXI= 37139
+INC+0YLQv9GA0LDQsg== 37140
+IEtpbWNoaQ== 37141
+IEZpbm5pc2g= 37142
+tOyggQ== 37143
+aW5haXJl 37144
+IG9pa2U= 37145
+IGhvc3RhZ2U= 37146
+IEJ1dHRvbg== 37147
+2YLZig== 37148
+ZWtpbmc= 37149
+IEthemFraA== 37150
+IGNvbWZvcnRpbmc= 37151
+IHNvZw== 37152
+IGdyZWV0ZWQ= 37153
+Z3VpdGFy 37154
+cGF5ZXI= 37155
+IHJlbGF0aW9uYWw= 37156
+IGNvbnN0cnVpcg== 37157
+b3BpYW4= 37158
+IFZvbHVtZQ== 37159
+aWV0aA== 37160
+0YHRgtCy0L7QvA== 37161
+dXJyZWN0aW9u 37162
+bGnFm215 37163
+IGhlbWlzcGhlcmU= 37164
+IEJlYW4= 37165
+SUdO 37166
+IGvDtnTDvA== 37167
+IEZhbGxvdXQ= 37168
+IGJyYWNl 37169
+z4DOrA== 37170
+IEhBUw== 37171
+IGfDqQ== 37172
+IGNoYXJhY3Rlcml6ZQ== 37173
+4bq3Yw== 37174
+IE1pbGt5 37175
+IHR1bW9ycw== 37176
+IG51aXQ= 37177
+IEdheg== 37178
+IOyeiOuLpOuKlA== 37179
+INCz0LDRgA== 37180
+ZXNzbWVudA== 37181
+IEFiZQ== 37182
+IOu9kQ== 37183
+IEVpbnNhdHo= 37184
+SklO 37185
+asOk 37186
+Q3J5 37187
+IFByb21pc2Vk 37188
+INGB0LXRgNC0 37189
+b2t1cw== 37190
+IHNjYWxhYmxl 37191
+INC/0L7RgdC80L7RgtGA0LXRgtGM 37192
+w7xja2xpY2g= 37193
+IHJlYWxpc20= 37194
+IG1heW8= 37195
+IGp1dmVuaWxl 37196
+IGhlYWRsaWdodHM= 37197
+IGfDtnLDvMWf 37198
+IFJlZm9ybQ== 37199
+IGhhbHZlcw== 37200
+Y3puZQ== 37201
+IGJyZWFrdXA= 37202
+xbxlag== 37203
+IHLDpHR0 37204
+RGF5 37205
+IOydvOuzuA== 37206
+IG11ZXJ0ZQ== 37207
+IHR1bmVz 37208
+IFNtaWxl 37209
+cmVjb3Jk 37210
+IHJlY2hlcmNoZQ== 37211
+YXRpc2ZpZWQ= 37212
+IHBvemk= 37213
+IGNlbGVicmF0aW9ucw== 37214
+aXNleHVhbA== 37215
+IFJPQg== 37216
+dGhpcmRz 37217
+IEZvcnR1bmU= 37218
+INGC0L7QuQ== 37219
+IGJyYW5kZWQ= 37220
+bG9v 37221
+IGR1ZA== 37222
+IHJhbmRvbWl6ZWQ= 37223
+IGNvbWJpbg== 37224
+aWVyYW4= 37225
+Y3plbmlh 37226
+IGN1cmF0b3I= 37227
+IGFydGVyeQ== 37228
+INGD0Yg= 37229
+INGH0LjRgg== 37230
+IHN1YnNpZGllcw== 37231
+IGJsb3Nzb20= 37232
+IFR3aWxpZ2h0 37233
+IGh5dsOk 37234
+IFBvbXBl 37235
+IENpc2Nv 37236
+INCf0YDQvg== 37237
+IGJpcmk= 37238
+IGdlcm4= 37239
+IHJlYnVpbHQ= 37240
+IHdjemU= 37241
+IGJlbmVmaWNp 37242
+IGRydW1tZXI= 37243
+IHNvbGlkcw== 37244
+IGRpeW9yc3Vu 37245
+bGF0ZWQ= 37246
+IG11ZGR5 37247
+IGhvbG9n 37248
+IGNsYXBz 37249
+IFJpbmdz 37250
+IE9rZXk= 37251
+IEJyYXZl 37252
+IHZhbHVhdGlvbg== 37253
+IG1pZ3JhbnQ= 37254
+IGludGVybWl0dA== 37255
+IGVpZ2VuZQ== 37256
+aWxpYXJ5 37257
+bWFya3Q= 37258
+a3I= 37259
+IFJpYg== 37260
+4buZaQ== 37261
+IGFjY3VzYXRpb25z 37262
+IGFyYWI= 37263
+d2FzaA== 37264
+IEJhcmR6bw== 37265
+IHVnaA== 37266
+ZXN0ZXJz 37267
+b3BocmVu 37268
+IGFsaW1lbnRvcw== 37269
+IFV6 37270
+1oI= 37271
+IDY1MA== 37272
+INC/0YDQuNC10YU= 37273
+Rkk= 37274
+IHNhbXBhaQ== 37275
+IHBhcmzDqQ== 37276
+aGVzaW9u 37277
+IHPEsXI= 37278
+IGFwcGFyYXR1cw== 37279
+IGNvcnJlbGF0ZWQ= 37280
+IFByaW5jaXBhbA== 37281
+IGNvcnI= 37282
+IE9mZmljaWFs 37283
+0LjRh9C10YHQutC40LU= 37284
+IHRlcm1pbmFscw== 37285
+U2hvdWxk 37286
+IHZhY3Vu 37287
+IHN0ZWxsdA== 37288
+IG1vb2k= 37289
+ZXR6dW5n 37290
+INC60YDQsA== 37291
+IGRhaQ== 37292
+INC/0L7Qtg== 37293
+VGVhbQ== 37294
+IFBQRQ== 37295
+INCe0YE= 37296
+IExlYWg= 37297
+IEl2eQ== 37298
+eXN0 37299
+IHVoaGg= 37300
+IG5pZ2h0dGltZQ== 37301
+IHRyZW5keQ== 37302
+IHNlY3VyaXRpZXM= 37303
+IGNvbnRpbmVudHM= 37304
+IGZpcnN0aGFuZA== 37305
+IFZlcm9u 37306
+IOuCrg== 37307
+IGJyb3dzaW5n 37308
+IENhZGE= 37309
+dHJv 37310
+IHRyYW1w 37311
+cmVpYg== 37312
+IGVyc3RtYWw= 37313
+aXJsZXI= 37314
+IHBzaWM= 37315
+IGdldGly 37316
+IE5Q 37317
+IGR6aWVjaQ== 37318
+0L7QsdGA0LDQtw== 37319
+IG1hZ2ljaWFu 37320
+IHNjcnV0aW55 37321
+IHNsYWI= 37322
+IE9U 37323
+aXN0eQ== 37324
+aXJpZXM= 37325
+b3Jlc3Q= 37326
+IHRhc2tlZA== 37327
+IG1vcmFsbHk= 37328
+7JW87KeA 37329
+dXN0ZXJlZA== 37330
+IGZvb2xz 37331
+IGlycmVzcG9ucw== 37332
+IGVpbmY= 37333
+IHZp4buHYw== 37334
+IHNjb3I= 37335
+IHBpbGxvd3M= 37336
+IEdlZ2Vu 37337
+IHR1dHRl 37338
+IHF1YXJ0ZXJseQ== 37339
+IGRpZG50 37340
+IEd5bQ== 37341
+IEV0aGVy 37342
+INir 37343
+0LvQuNGI0LrQvtC8 37344
+IHNpZ25hbGluZw== 37345
+IE5vZGU= 37346
+IERvbmNz 37347
+IHlhaA== 37348
+IEthbmFs 37349
+IGZhZGluZw== 37350
+ZXRpbg== 37351
+IGluZmx1ZW5jZXJz 37352
+IG1lZGFscw== 37353
+IGVuZ2luZWVyZWQ= 37354
+IGZlcm1lbnRlZA== 37355
+6rKg7KeA66eM 37356
+IEJlZXRob3Zlbg== 37357
+157XqQ== 37358
+aW5lbnRhbA== 37359
+IOyVjOugpA== 37360
+w7x0ZmVu 37361
+YWxueWE= 37362
+IG92ZXJl 37363
+IGRlbmt0 37364
+0LDQutGC0LXRgA== 37365
+IOKY 37366
+IG5lY2VzaXQ= 37367
+IGdlbmVyYXRvcnM= 37368
+Z3Jhc3M= 37369
+INC/0L7QtNGD0Lw= 37370
+bGllw59lbg== 37371
+QmFy 37372
+nOuPmQ== 37373
+INC00LXRgtC10Lk= 37374
+IHN1Y2tpbmc= 37375
+IHN0ZW5jaWw= 37376
+IHByaW1v 37377
+IEJyZWF0aA== 37378
+c3Ryb20= 37379
+IGltbWVuc2VseQ== 37380
+IGFwcHJlaA== 37381
+7KCV7J20 37382
+UG9w 37383
+IGpvbmc= 37384
+IEdpdWw= 37385
+IEFESEQ= 37386
+IGjDtnJlbg== 37387
+IGVsbw== 37388
+aXZlbnQ= 37389
+IHJ1cw== 37390
+IG91dHJhZ2VvdXM= 37391
+IG1hc3RlcmVk 37392
+IOy7pA== 37393
+2YjZgQ== 37394
+aXBlcw== 37395
+IFJ1ZHk= 37396
+SmFjb2I= 37397
+IGJ1bGxpc2g= 37398
+IHRhcHBlZA== 37399
+IGZhdWQ= 37400
+aXpvcGhyZW4= 37401
+INGB0L7RhQ== 37402
+IERhcmxpbmc= 37403
+IDE5NjM= 37404
+IFByZXZlbnRpb24= 37405
+spQ= 37406
+IGFiZG9taW5hbA== 37407
+c3RvbmVz 37408
+IGF2YWllbnQ= 37409
+4buVaQ== 37410
+bWFrZQ== 37411
+IHNhcmU= 37412
+IEluc3RhbnQ= 37413
+0LrQsNC8 37414
+IGtlZXBlcg== 37415
+IGJsYW5rZXRz 37416
+IHN3ZWF0cw== 37417
+IE1pbm5lYXBvbGlz 37418
+IGdlbm9tbWVu 37419
+IGZhc3Rlbg== 37420
+IEJydXNzZWxz 37421
+IGNhZmV0ZXI= 37422
+IGFic29yYmluZw== 37423
+IGhhZ28= 37424
+IEVsbW8= 37425
+IGd1c3Rv 37426
+IFlhcA== 37427
+TcO6c2ljYQ== 37428
+IHRlcnQ= 37429
+IGJhbmRh 37430
+IG1pbHk= 37431
+IHRoZXJlYWZ0ZXI= 37432
+IFN0b2NraG9sbQ== 37433
+IENhcnNvbg== 37434
+IGNhbGlicmF0aW9u 37435
+YXZhxZ8= 37436
+YW5zYQ== 37437
+aWtrZQ== 37438
+IGZvcmVzZWU= 37439
+IHF1YWxjaGU= 37440
+IGRlc3Rl 37441
+5qQ= 37442
+w7xuw7x6 37443
+IGZvcmdl 37444
+RGlz 37445
+ZXN0ZW4= 37446
+IM60zrnOsQ== 37447
+IGVuY2Fwcw== 37448
+IEdlc3By 37449
+IGNoZXJjaGVy 37450
+aWNrZXRz 37451
+0YLQvtGA0Ys= 37452
+Q3I= 37453
+INCi0LDQutC20LU= 37454
+IHJhYmJpdHM= 37455
+IERvdA== 37456
+aGVpdGVu 37457
+IGNhdXNhbA== 37458
+IEZvc3Rlcg== 37459
+YWrEhWM= 37460
+IGJlcmVpdA== 37461
+IGF5dWRhcg== 37462
+c29uZw== 37463
+Y29tYg== 37464
+IGZyaW5nZQ== 37465
+IGN5YmVyc2VjdXJpdHk= 37466
+IOucqA== 37467
+IGtpZXI= 37468
+IGJlc2Now6RmdA== 37469
+INC60L7QvdGG0LU= 37470
+IGZhY2lsaXQ= 37471
+IE5hbWVu 37472
+IGJpbGF0ZXJhbA== 37473
+dHg= 37474
+IFdpc3NlbnNjaGFmdA== 37475
+IG51YW5jZXM= 37476
+IHJpcHBpbmc= 37477
+IGZ5 37478
+IFNpY2hlcmhlaXQ= 37479
+IEdoYW5h 37480
+b2xvbg== 37481
+IHRvcHBlZA== 37482
+IE1vcm9jY28= 37483
+IHJhZGlhbA== 37484
+IExFRQ== 37485
+IEFuZHJlYXM= 37486
+ZWRk 37487
+IOyXtOs= 37488
+IEFpcmxpbmVz 37489
+IHZhbG9yZXM= 37490
+6rec 37491
+SHk= 37492
+INC30LDQtNCw0Yc= 37493
+IEtlbmRhbGw= 37494
+INGF0LDRgA== 37495
+IFZhbXA= 37496
+IHB5dGhvbg== 37497
+IG1hbmFnZWFibGU= 37498
+IEdlbnRl 37499
+b2lzZQ== 37500
+aWNpYXJ5 37501
+IGltcG9zcw== 37502
+IEJ1bm55 37503
+aWVzdGE= 37504
+QW5kcmV3 37505
+IHNlcnQ= 37506
+IENlYw== 37507
+enphcmVsbGE= 37508
+IGF1dG9tb2JpbGU= 37509
+IFRpZXJl 37510
+YWxsb3dz 37511
+IOuwgA== 37512
+IFNjb3Jw 37513
+IEplbGx5 37514
+YWdhcmE= 37515
+IFN0cmV0Y2g= 37516
+IHJlZGVm 37517
+IGV4YWNlcmI= 37518
+IFNIQQ== 37519
+w6lm 37520
+b3JzYQ== 37521
+IGZsYXdlZA== 37522
+IE5vZWw= 37523
+PyE/ 37524
+IHByb2NlbnQ= 37525
+IG1lbnN0cnU= 37526
+INC/0YDQvtGH 37527
+IGluZmFudHM= 37528
+8J+OtQ== 37529
+cGF1c2U= 37530
+IFJhY2luZw== 37531
+IDE5NDg= 37532
+IHN1cGVyaW50ZW5kZW50 37533
+aWRvcmVz 37534
+aWR5 37535
+YnJhaGlt 37536
+IHVubHVja3k= 37537
+IHBlcms= 37538
+YW5jaQ== 37539
+IOunjOuCmA== 37540
+INCc0L7RgdC60LI= 37541
+IGZpbmFucw== 37542
+IGRpZmVyZW5jaWE= 37543
+oIjsnbQ= 37544
+T1JZ 37545
+IFRhYw== 37546
+24zYpw== 37547
+IGRlc2Vt 37548
+INCy0LDQttC90L4= 37549
+IEpV 37550
+IOyeiOyeluyVhOyalA== 37551
+IM6d 37552
+IGluZm9ybWF0aW9ucw== 37553
+IEhFTA== 37554
+aHN0 37555
+INC/0L7Qs9C+0LLQvtGA 37556
+IHZvaXR1cmU= 37557
+IHJldXM= 37558
+w6RuZGln 37559
+INC/0L7RhdC+0LY= 37560
+amluZw== 37561
+IGRydQ== 37562
+YWx0cmE= 37563
+IHByb2R1aXRz 37564
+IGtpdGU= 37565
+IGV5ZWJhbGw= 37566
+IEJlbHQ= 37567
+IFJlc3RhdXJhbnQ= 37568
+IGdhbWI= 37569
+IHBvcnJpZGdl 37570
+aXR0ZXJz 37571
+IGNvbnZlcnRz 37572
+IHlhcmTEsW0= 37573
+IG3DoXhpbW8= 37574
+d2lydHNjaGFmdA== 37575
+IO2VmOuCmOs= 37576
+IOykgA== 37577
+IGljZWJlcmc= 37578
+IHZvcmJlaQ== 37579
+IDI1Ng== 37580
+b2NyYXRpYw== 37581
+IHJlY2tsZXNz 37582
+b25uZXI= 37583
+IG3DunM= 37584
+IGxvZ2ljYWxseQ== 37585
+IFByaXNvbg== 37586
+IE5ldHo= 37587
+IHZhY2FudA== 37588
+IG5pbW10 37589
+IEhBUlI= 37590
+INC30L7Qsg== 37591
+IERlZQ== 37592
+cmluZ2U= 37593
+bmllc3Q= 37594
+IFJ1bGVz 37595
+7Iqk65+9 37596
+Y3Vzc2lvbnM= 37597
+IGZsb3JhbA== 37598
+IGNvbnN0cmFpbmVk 37599
+IGRpZmZlcmVudGlhdGlvbg== 37600
+IFF1ZWJlYw== 37601
+INuB24zaug== 37602
+IHDDumJsaWNh 37603
+aXRlbA== 37604
+IGFjY29tbW9kYXRpb25z 37605
+IEdyw7w= 37606
+7Zw= 37607
+IHBpY2tsZXM= 37608
+0LjRh9C10YHQutC40YU= 37609
+IGNvbW1pc3Npb25z 37610
+IEJhZWs= 37611
+IMOnb2N1xJ8= 37612
+IE1lZGl1bQ== 37613
+IHBlcmlvZGljYWxseQ== 37614
+IHdvbmRlcmZ1bGx5 37615
+IHN0YWZmaW5n 37616
+7JuQ6w== 37617
+cmlyZQ== 37618
+Zmxl 37619
+IE1jTA== 37620
+INGC0LXQvw== 37621
+INC/0LXRgNC10Lo= 37622
+0L3QvtC70L7Qsw== 37623
+IO2BrOqyjA== 37624
+IHByb3NwZXJvdXM= 37625
+IFNwaXJpdHVhbA== 37626
+IENoaWNr 37627
+RElB 37628
+INCf0YDQuNCy0LXRgg== 37629
+IHBlcsOt 37630
+0YzRjtGC 37631
+IGNvbnN1bHRhbnRz 37632
+IEVhcmw= 37633
+IHJ1aW5pbmc= 37634
+0L7RgNC1 37635
+IHBlbnNlcg== 37636
+IHRha2llag== 37637
+IHN0cmVuZ3RoZW5lZA== 37638
+IExpcXVpZA== 37639
+0L7QvdC10YY= 37640
+0LDQstCw0YLRjA== 37641
+IGNhbWVy 37642
+IGRpc2FncmVlbWVudA== 37643
+IGJhdGhpbmc= 37644
+IFlvc2g= 37645
+YWFs 37646
+cHJlY2hlbg== 37647
+UklTQURBUw== 37648
+IHN1cGVyc3Rhcg== 37649
+0LvRj9GC0Yw= 37650
+IG5pYg== 37651
+IFRoZXJt 37652
+IERBTklFTA== 37653
+IHBhdw== 37654
+IGxpcXVpZHM= 37655
+IGNhcGFjaXQ= 37656
+YXJrZW4= 37657
+IHZhZ2luYQ== 37658
+IG1hc2hlZA== 37659
+IGVtZXJnZXM= 37660
+eXNjeQ== 37661
+IHVucmVsYXRlZA== 37662
+IEd1aWxk 37663
+IGludmVydGVk 37664
+aXRpdmVz 37665
+VHJh 37666
+IGJlZ3I= 37667
+IGFsdGU= 37668
+7KeV 37669
+INGA0LDQt9GA0LDQsdC+0YI= 37670
+ZmluZGVy 37671
+INC00LDQu9C10LU= 37672
+INCx0LvQsNCz0L7QtNCw0YA= 37673
+d2Fsa2Vy 37674
+IGNyYXRlcg== 37675
+YXNzYWRvcnM= 37676
+cmVuY2Vz 37677
+aW5za2k= 37678
+IEtJTQ== 37679
+IEVsbGlvdA== 37680
+MjAxNw== 37681
+IFNy 37682
+aW5rYQ== 37683
+YW5vdg== 37684
+IOyemOuquw== 37685
+IHByb3ByaWV0YXJ5 37686
+ZGlzcGxheXN0eWxl 37687
+INGB0LjQvA== 37688
+INC40LfQsQ== 37689
+IFBhbmVs 37690
+IGluc3RpbmN0cw== 37691
+IENvbW11bmljYXRpb25z 37692
+bWlkdA== 37693
+IOunjOuTpOyWtA== 37694
+INGB0LvQvtCy0LA= 37695
+IEdpbGJlcnQ= 37696
+0KLQsNC6 37697
+dm9vcmJlZWxk 37698
+0LXRjtGB0Yw= 37699
+YXJ5bg== 37700
+cXVleg== 37701
+IGRhcnQ= 37702
+0ZbRiA== 37703
+IEh1dA== 37704
+U2Fs 37705
+IHNvdXRoZWFzdA== 37706
+IHBlc3RpY2lkZXM= 37707
+IGhlbGljb3B0ZXJz 37708
+IGVuZHVyZWQ= 37709
+aWFkYQ== 37710
+IGJyZXdpbmc= 37711
+7Jes6w== 37712
+INGB0LLQvtCx0L7QtA== 37713
+IFNhaW50cw== 37714
+IEZyYW7Dp2Fpcw== 37715
+IEVjb25vbWljcw== 37716
+IGRpc2xvYw== 37717
+b3Bob2JpYQ== 37718
+Q2FtZXI= 37719
+IG5lZ290aWF0ZWQ= 37720
+INGB0YLQsNC70Lg= 37721
+7Iqk7YE= 37722
+b2dpZQ== 37723
+IHRzdW5hbWk= 37724
+IHBlZWxlZA== 37725
+IG1vdGl2YXRpb25z 37726
+b3N0YXQ= 37727
+Zmxhbg== 37728
+IERBQw== 37729
+IGthdg== 37730
+J1JF 37731
+IFBlYXJzb24= 37732
+YmJl 37733
+Y3plbmll 37734
+IGF0ZW7Dp8Ojbw== 37735
+7Ya166C5 37736
+INGD0LTQsNGA 37737
+IGludHJvZHVjdG9yeQ== 37738
+IEljaQ== 37739
+64yA6w== 37740
+YWthdA== 37741
+IHRyZW5jaA== 37742
+IHByb2NlZWRlZA== 37743
+IENvaW4= 37744
+IGRlcmVjaG8= 37745
+IFJlZGU= 37746
+0LDQvdC90YvQuQ== 37747
+IGluY2FyY2VyYXRlZA== 37748
+IFJpY2htb25k 37749
+Um9jaw== 37750
+IFBhdg== 37751
+IEthcm1h 37752
+dWdlcw== 37753
+IGNvbnRlw7o= 37754
+67mE 37755
+IOq3uOunjA== 37756
+IEdvbmU= 37757
+IHdzcMOzxYI= 37758
+IFJhaG1lbg== 37759
+dW5rZW4= 37760
+IOykkeyalO2VnA== 37761
+IGli 37762
+IGF0dGFjaGluZw== 37763
+SGF5 37764
+IHN1a2E= 37765
+7I25 37766
+IHBpdm90YWw= 37767
+IFJlc3BlY3Q= 37768
+w61kYQ== 37769
+SUI= 37770
+IFZlcmFudHdvcnQ= 37771
+d2lldA== 37772
+IGZvcmVuc2lj 37773
+0YDQuNGB0YI= 37774
+INC/0YDQuNC90YbQuNC/0LU= 37775
+IG1hcmtpbmdz 37776
+IGtldHRsZQ== 37777
+IE9wZXJh 37778
+IERvY3RvcnM= 37779
+IHNocmVkZGVk 37780
+IHJlY3Vlcg== 37781
+IHZpZ2ls 37782
+IEZhaWw= 37783
+IGVudHJldg== 37784
+INC00YPRiA== 37785
+IG91dGJyZWFrcw== 37786
+IM+Azr8= 37787
+IHJvZ3Vl 37788
+YW5nbGVk 37789
+IHllYXJseQ== 37790
+IENyZWVk 37791
+IHdhbQ== 37792
+IGxvdHVz 37793
+6rO86w== 37794
+44CB44CB 37795
+IFNwaXQ= 37796
+IEl0dQ== 37797
+IHN0cmFpbnM= 37798
+IHN0YW1wZWQ= 37799
+IHBsYWludA== 37800
+IHBvdGlvbg== 37801
+IGNvbnNvbGlkYXRpb24= 37802
+0L7Rh9C60YM= 37803
+IHZsb2dnaW5n 37804
+IHNsYXRl 37805
+IEF1ZnQ= 37806
+IEluY29y 37807
+4burbmc= 37808
+p5A= 37809
+ZW5o 37810
+IGhlacOf 37811
+IGRvbWVzdA== 37812
+IFN0cm9t 37813
+YWtpcw== 37814
+IGZyYWdlbg== 37815
+IGZpbmVy 37816
+IFN1Zw== 37817
+IHVwaGlsbA== 37818
+IMOpw6lu 37819
+4oCmKQ== 37820
+INGB0L7Qvw== 37821
+IENvcmV5 37822
+IHNpZWJpZQ== 37823
+IG11c2U= 37824
+IGNsb3Zlcw== 37825
+IHBvdXM= 37826
+IEZpbmFueg== 37827
+IFJvdXRl 37828
+YW1hdA== 37829
+IG11dHVhbGx5 37830
+INCy0L3Rg9GC0YDQuA== 37831
+IFNlbGVuYQ== 37832
+65Q= 37833
+IEdhdXNzaWFu 37834
+67aA7YSw 37835
+INeR15s= 37836
+IGVqZXJj 37837
+a2Vh 37838
+IEdlcnJ5 37839
+IFNpYw== 37840
+IDE5NjY= 37841
+aWVzZQ== 37842
+IGZvc3NpbHM= 37843
+IGVzdGFk 37844
+IEthbmU= 37845
+Y2nEhw== 37846
+IOycoO2KnOs= 37847
+INC/0LDQvA== 37848
+IENydWlzZQ== 37849
+aW50w6lyaWV1cg== 37850
+IGJla2FubnQ= 37851
+IFBvZGU= 37852
+IGRlbWFuZGVy 37853
+UmVt 37854
+IGludmFkZQ== 37855
+IGRlY29yYXRpbmc= 37856
+cm9waWM= 37857
+IGNvd2JveQ== 37858
+IFBob3Rv 37859
+b3BvbGl0 37860
+IOy7rOufrOs= 37861
+IHJlYXA= 37862
+IGhhbmR3cml0aW5n 37863
+4LmE4Lij 37864
+IOua 37865
+INio2LnYrw== 37866
+IE10 37867
+2YA= 37868
+IHNwYWNlc2hpcA== 37869
+IG5hdGlvbmFsaXNt 37870
+IGNvdW5jaWxz 37871
+IEdyaWZmaW4= 37872
+IEFobWVk 37873
+IGNsaWNo 37874
+IE9M 37875
+d2w= 37876
+IFBpbG90 37877
+IGFjcm9ueW0= 37878
+IGdlbHM= 37879
+IGVsZWN0cm9seQ== 37880
+6JM= 37881
+INC80L3QvtC5 37882
+IGVwaXNvZA== 37883
+IERpZXNlcw== 37884
+IEFUUA== 37885
+IGVkaXlvcnVt 37886
+IGV4cHJlc3Nlcw== 37887
+IGV4aGliaXRz 37888
+Q29tbQ== 37889
+INC60YDRg9C/ 37890
+IG1hdGFy 37891
+IDIwMjU= 37892
+IEFydGVt 37893
+dmFzaXZl 37894
+csOg 37895
+IGJlxZ8= 37896
+IGxpemFyZA== 37897
+IGZpbGxl 37898
+IOyniOusuA== 37899
+INC80L7RiQ== 37900
+IHTDvHI= 37901
+IGN1bHByaXQ= 37902
+IHdvdmVu 37903
+IEFOWQ== 37904
+bmlt 37905
+IHRheQ== 37906
+IHByb21pbg== 37907
+IGFjb21wYQ== 37908
+IGlkw6k= 37909
+IGJvaWxlcg== 37910
+IFRoZW1lbg== 37911
+IGF2ZW51ZQ== 37912
+IE11ZA== 37913
+INC90L7QstGL0LU= 37914
+IHdpdG5lc3Npbmc= 37915
+IGxhbmNl 37916
+IENIQU4= 37917
+IEJldmVy 37918
+2KrZhQ== 37919
+IGNoZW1vdGhlcmFweQ== 37920
+S2luZw== 37921
+IGLEmWTEmQ== 37922
+IGF0dWFs 37923
+IHRpdmU= 37924
+IHRhbGtpbg== 37925
+IHF1ZWRhcg== 37926
+aWXDnw== 37927
+ZWRlbA== 37928
+IOyWtOygnA== 37929
+IGpvZ2Fy 37930
+IMO2cg== 37931
+IHVuZGVydGFraW5n 37932
+IFN0cmVuZ3Ro 37933
+IG1pbGjDtWVz 37934
+IFdpbmU= 37935
+IE1vbHQ= 37936
+IHVuZGVybWluZQ== 37937
+IEFyY2hpdmVz 37938
+dmFuYQ== 37939
+bWVyY2lhbA== 37940
+TUM= 37941
+IGNhc3Rl 37942
+0L/RgA== 37943
+IGxlZ2lzbGF0b3Jz 37944
+dWxhdG9ycw== 37945
+w6puaW8= 37946
+IOuNsOs= 37947
+INGF0L7RgtC40YLQtQ== 37948
+INC90LXQug== 37949
+IHN1cm4= 37950
+IGNvbnNjaQ== 37951
+IFBPVw== 37952
+IGN1bGluYXJ5 37953
+IEtBVA== 37954
+IEZvbGtz 37955
+0YvQstCw0LXQvA== 37956
+INCy0L7Qug== 37957
+c2VydmljZQ== 37958
+cHRz 37959
+INC/0L7QsdC10LQ= 37960
+IHRlbnRz 37961
+IG5vcmQ= 37962
+U1RF 37963
+IHJlcHVibGljYW4= 37964
+IHd5aw== 37965
+IG1pbmlvbnM= 37966
+IG1lbWFuZw== 37967
+amVzdA== 37968
+IGNvbXBhcmF0aXZl 37969
+IHR5bGU= 37970
+Y2FyYm9u 37971
+YmVkaW5ndA== 37972
+a3Nlbg== 37973
+IG5lZ2F0aXZpdHk= 37974
+IHNqw6Rsdg== 37975
+IGTDug== 37976
+IHJlY2FsbGVk 37977
+Y3Jh 37978
+IFRhZGE= 37979
+INGA0YPQutC4 37980
+INC+0L/RgNC10LTQtdC7 37981
+IHByb2NyYXN0 37982
+IGpvZ29z 37983
+IE9v 37984
+IEhlYXJ0cw== 37985
+IMOpY2g= 37986
+IGtzacSFxbw= 37987
+IGNvYXJzZQ== 37988
+IFR1YmU= 37989
+IEdyZWVucw== 37990
+IMOpbg== 37991
+IGR1bWJiZWxs 37992
+INGC0Lg= 37993
+IHF1ZXJlcg== 37994
+2KfYrQ== 37995
+z4POtc65 37996
+INC/0YDQsNCy0LjQu9GM0L3Qvg== 37997
+INC/0LDQvw== 37998
+IGNvbXByYQ== 37999
+IHTDqXI= 38000
+IEFudGVz 38001
+IG9wdGltdW0= 38002
+IGJpc2N1aXQ= 38003
+zrrOuQ== 38004
+YWN6ZWdv 38005
+IOyLnOqwhOydtA== 38006
+IE1hcmluZXM= 38007
+dmVybw== 38008
+IHZhY2NpbmF0aW9ucw== 38009
+IHBldHR5 38010
+cml0ZXJz 38011
+INCw0Ls= 38012
+Y291bnRyeQ== 38013
+IGNvdW50ZXJz 38014
+IGF0dGVuZGFudA== 38015
+IEh1aQ== 38016
+Y2th 38017
+0YHRgtCy0LXQvdC90YvQuQ== 38018
+Z3V5 38019
+IHRyaWNrZWQ= 38020
+IFJFRA== 38021
+IHRocmlsbGluZw== 38022
+z4DOv865 38023
+IHBpZ2d5 38024
+IGFudW5jaQ== 38025
+T1JURVI= 38026
+IFZhbHVl 38027
+IHJvbmQ= 38028
+IEFEQQ== 38029
+IHBvc2Vy 38030
+aG9yZXM= 38031
+IFJvbGFuZA== 38032
+k68= 38033
+IG5vaXI= 38034
+INep15DX 38035
+67Cc 38036
+aWVtYW5k 38037
+INC/0L7RgtC10YA= 38038
+6rOz 38039
+IOqxsQ== 38040
+IGZvcm1hdHRpbmc= 38041
+IExlZA== 38042
+IGtpbGxlcnM= 38043
+IMSR4bqleQ== 38044
+IGhhYXI= 38045
+YWdhaW4= 38046
+ITwv 38047
+IHNvbWV0aGlu 38048
+IGNvdWdoaW5n 38049
+IG5hdmU= 38050
+IHByb3NwZWN0aXZl 38051
+IEhL 38052
+IFJlc2N1ZQ== 38053
+bWF5YmU= 38054
+Z2dlcg== 38055
+INGA0LDQsdC+0YLRgw== 38056
+15XXnNed 38057
+dGFpbHM= 38058
+7ZWY7ZWY 38059
+IGV5ZWxpZA== 38060
+IGN1c3RvbWl6YXRpb24= 38061
+YXZpbGlvbg== 38062
+IHByb2NoYWlu 38063
+IGdsYXpl 38064
+U2lt 38065
+INC+0L/QsNGB 38066
+IG1vc3F1aXRvZXM= 38067
+IGZlbnQ= 38068
+IGNhcGFjaXRpZXM= 38069
+IGFwb3N0bGVz 38070
+IGFsdHVyYQ== 38071
+IOusuw== 38072
+IHNlcm9udA== 38073
+IEFueXRpbWU= 38074
+pbTripQ= 38075
+IGNvc3BsYXk= 38076
+IHNwYWM= 38077
+IHNhbWVu 38078
+dWNj 38079
+acOocmVz 38080
+IHNpYmxpbmc= 38081
+IENvY2s= 38082
+IOuPhQ== 38083
+INC/0YDQtdC00YHRgtCw0LLQu9GP 38084
+IGluc3RhbGxtZW50 38085
+IGRpamU= 38086
+IE1DVQ== 38087
+IEVI 38088
+IE5pbmc= 38089
+IHByZXBhcmVz 38090
+IGh5cG9jcg== 38091
+cHR5 38092
+IGthZMSxbg== 38093
+IEZyb3plbg== 38094
+aGF1bA== 38095
+IEt5bGll 38096
+IHNodWZmbGU= 38097
+IGVsZW1lbnRhbA== 38098
+IGF1w59lcg== 38099
+IEtOT1c= 38100
+IEFMSVNTQQ== 38101
+WkE= 38102
+7LKg 38103
+IHJlY2l0ZQ== 38104
+IHNjcmli 38105
+IDExNQ== 38106
+IHN0YXJyZWQ= 38107
+IGxlcXVlbA== 38108
+IGJyZXdlcg== 38109
+IE9wcG9ydHVu 38110
+IHLDpA== 38111
+IGNob3BzdGlja3M= 38112
+IEthaA== 38113
+IEV0aGlvcGlh 38114
+IGhhbmRtYWRl 38115
+IGVyZm9sZw== 38116
+IER6 38117
+aXR0ZW5z 38118
+0LLQsNC7 38119
+zrfOvQ== 38120
+YnJpbmdlbg== 38121
+IHVucGx1Zw== 38122
+IG9mZnM= 38123
+IGhlcm1hbg== 38124
+bGllZA== 38125
+YXNvbmlj 38126
+IFNlcmJpYQ== 38127
+IEd1YXRlbQ== 38128
+IC4uLiI= 38129
+IGVycmVpY2hlbg== 38130
+IGFtYmlndW91cw== 38131
+IFdoaXRuZXk= 38132
+enVm 38133
+TUFORA== 38134
+oLU= 38135
+IHNxdWVlemVk 38136
+eWFz 38137
+IFNob2Nr 38138
+IHV0aWxpc2U= 38139
+dWtv 38140
+Ym9sdA== 38141
+IG1vdGlm 38142
+IGlubWF0ZXM= 38143
+IGNvcnJ1cHRlZA== 38144
+IGNvbmNyZXQ= 38145
+IENyaXRpY2Fs 38146
+IFNpbmdpbmc= 38147
+INGE0YPQvdC6 38148
+bm92YQ== 38149
+cmViYmU= 38150
+ZHQ= 38151
+VW5pcw== 38152
+IHdlYmNhbQ== 38153
+IGNhbW91Zmw= 38154
+S2Vu 38155
+IGxhd3N1aXRz 38156
+IENvbnN1bWVy 38157
+IHJlY29sbA== 38158
+IGtsZWluZXI= 38159
+IEZJRkE= 38160
+IDE5NjI= 38161
+IG1hbGFk 38162
+IOywvQ== 38163
+IMOldA== 38164
+IGluZmx1ZW5jZXI= 38165
+IEFydGlzdA== 38166
+c3Rp 38167
+4Lin4Lii 38168
+eXPFgg== 38169
+IEJpYW4= 38170
+iOuEpA== 38171
+IGZpcmVwbGFjZQ== 38172
+IEFwcGxpY2F0aW9u 38173
+IG1uaWVq 38174
+IGFjaWRpYw== 38175
+IE1vcm1vbg== 38176
+c3Nh 38177
+IHNuZWFreQ== 38178
+IG9qb3M= 38179
+IHZvdWQ= 38180
+IERhaQ== 38181
+IGdyYXNzcm9vdHM= 38182
+IFVuYmVsaWV2YWJsZQ== 38183
+IEdhYmU= 38184
+IEV4dHJlbWU= 38185
+IGhhc3NsZQ== 38186
+IGNvYg== 38187
+bXVtYmxpbmc= 38188
+UGFzcw== 38189
+jOufrA== 38190
+IHN5c3RlbWF0aWNhbGx5 38191
+IHNldmVudGVlbg== 38192
+z4DOtc65 38193
+4pmh 38194
+INC60L7Rgg== 38195
+IHNlbmRpcmk= 38196
+IGJhdGhyb29tcw== 38197
+IFN0ZXJu 38198
+IEFyZHVpbm8= 38199
+6Lk= 38200
+Y3JpYmluZw== 38201
+IHJlb3BlbmluZw== 38202
+IGNlcnY= 38203
+cGVl 38204
+QVJJ 38205
+IGNhZHJl 38206
+IEFuY2g= 38207
+TGVl 38208
+IE1BWA== 38209
+IG3DpG5u 38210
+IGNob3Jlcw== 38211
+IGFkZXNzbw== 38212
+IE5pZw== 38213
+IGRpc3NlcnRhdGlvbg== 38214
+IFZheQ== 38215
+U1RBTEs= 38216
+0LDQutCw 38217
+YXZhdA== 38218
+IHB1bmt0 38219
+IHBhZGRpbmc= 38220
+IFRlbXBs 38221
+IGVqZQ== 38222
+IO2EsA== 38223
+IGF6dA== 38224
+IOuMgO2GteuguQ== 38225
+IHJlYXJyYW5nZQ== 38226
+w6FjaA== 38227
+IOyCrOuejOuTpA== 38228
+IGZyZWFraW4= 38229
+Y3JpcmU= 38230
+IOy7pOs= 38231
+IEV4cGxhaW4= 38232
+IM+Ez4nOvQ== 38233
+IGJvZGlseQ== 38234
+IExlaXN0 38235
+IHNpZ3Vp 38236
+IGJ1bmtlcg== 38237
+IGF6dWw= 38238
+IEhhdXNo 38239
+U3Vi 38240
+INCQ0L3QtA== 38241
+INC60YDQsNC5 38242
+IGlsbGVnYWxseQ== 38243
+IE11eQ== 38244
+IEZlaQ== 38245
+IEJhbmFuYQ== 38246
+IHNjaG9sYXJseQ== 38247
+IFByenk= 38248
+IE1vc3M= 38249
+IEZpbHRlcg== 38250
+IOyWtOuWoQ== 38251
+IE1heHdlbGw= 38252
+dGVuc2U= 38253
+IGxvbmdpdHVk 38254
+IGxhbmdzYW0= 38255
+INee16c= 38256
+c21pdGg= 38257
+aXphZGE= 38258
+INC90L7RgNC80LDQu9GM0L3Qvg== 38259
+IFZvbGw= 38260
+IEVsZW5h 38261
+INGF0L7RgtGM 38262
+IERhYmVp 38263
+IGNvbnNlcnZhdGl2ZXM= 38264
+IHByw7Nwcmlh 38265
+IERpZXNlcg== 38266
+IEJyZW5kYQ== 38267
+b29raWU= 38268
+IGJhbmM= 38269
+7J207KY= 38270
+7JuD7J2M 38271
+IGtlaA== 38272
+IHdlZGRpbmdz 38273
+IHRodW5kZXJzdG9ybQ== 38274
+IENvb3JkaW4= 38275
+7IiY6rCA 38276
+IHByemVjaQ== 38277
+T1NTVEFMSw== 38278
+bWFhbg== 38279
+IOqxtOs= 38280
+INio2Yc= 38281
+IMW8YWQ= 38282
+IHlhY2h0 38283
+IGfDtnQ= 38284
+IGJsZWFjaA== 38285
+IHNob3J0ZW4= 38286
+INGB0YLQsNC70L4= 38287
+dXNhbg== 38288
+IOyekOyXsA== 38289
+IGRlcnM= 38290
+eGlz 38291
+jZTri4g= 38292
+IHF1YW50aWRhZGU= 38293
+IG9wcHJlc3NlZA== 38294
+INC30LDQutC+0L3Rhw== 38295
+INGH0LXRgtGL 38296
+INCd0LDQv9GA0LjQvNC10YA= 38297
+dWxw 38298
+2YLZiNmE 38299
+0L7Rh9C1 38300
+zqzOuw== 38301
+emVuaXU= 38302
+IGZvcm1hdGlvbnM= 38303
+IHNwYXJrZWQ= 38304
+IEVudHdpY2tsdW5n 38305
+YWxscw== 38306
+IHZpdmly 38307
+IGV4cGlyYXRpb24= 38308
+b3RpbmU= 38309
+INCn0LXRgA== 38310
+IFR1cm5pbmc= 38311
+IHRhcmlmZnM= 38312
+IG5hc3TEmXA= 38313
+IGFiaWRl 38314
+aWtzaQ== 38315
+IGZsYXNoZXM= 38316
+IGRpc3B1dGVz 38317
+IOyytA== 38318
+IG1lcmFr 38319
+IGVub3Jtb3VzbHk= 38320
+emFobA== 38321
+IGbDvGhydA== 38322
+0LLQvtC9 38323
+INC30LDQstC40YE= 38324
+IHBlcnNldmVyYW5jZQ== 38325
+IGRpdmlkZW5kcw== 38326
+IGNvbnRlc3RhbnRz 38327
+IHByb3N6xJk= 38328
+IEZyYW5rZW4= 38329
+IGV4cGxvcmVy 38330
+IGJ1ZmZhbG8= 38331
+4oCV 38332
+IGVjb2xvZ3k= 38333
+IHNjYWxhcg== 38334
+IGNyYW4= 38335
+zrXPhM6xzrk= 38336
+xbx5xIc= 38337
+IOyalOs= 38338
+IGdpYQ== 38339
+IEdvZw== 38340
+IFByaXY= 38341
+IOunkOydhA== 38342
+IFJlYXNvbg== 38343
+cmFrdGlvbg== 38344
+IERlYm9yYWg= 38345
+IGtpdHRlbg== 38346
+IEVkaW4= 38347
+cGllag== 38348
+IOuLtA== 38349
+IG3DoXF1 38350
+IGJpZGRpbmc= 38351
+IGFmZmluaXR5 38352
+IGFpa2E= 38353
+Zm9saw== 38354
+IENvbnNl 38355
+IGRldXRzY2hlbg== 38356
+6IY= 38357
+IGRlYml0 38358
+xLHEn8Sxbg== 38359
+aXNlbA== 38360
+IOykkeq1rQ== 38361
+IOutkOqwgA== 38362
+IHRydXN0d29ydGh5 38363
+IFN0YXJ0ZWQ= 38364
+w7xyZA== 38365
+INC/0L7QvdGP0YLQvdC+ 38366
+IHNjaWVudGlmaWNhbGx5 38367
+UG9kcw== 38368
+Q1JPU1NUQUxL 38369
+IHByZWd1bnRhcw== 38370
+IGNhbG1pbmc= 38371
+IFByZW1pZXJl 38372
+15vXqQ== 38373
+INGF0L7Qu9C+0LQ= 38374
+IGNhcGl0YQ== 38375
+IHRvbWE= 38376
+IG11cm0= 38377
+IGZ1ZXJ6YQ== 38378
+IEhhbmk= 38379
+w7xm 38380
+YXJsb3M= 38381
+IGjDpHVm 38382
+IG9zb2J5 38383
+amVnbw== 38384
+INC/0LjRgQ== 38385
+IGNhbG1seQ== 38386
+aWRldA== 38387
+YnVjaA== 38388
+Z29uZQ== 38389
+IHZpc2Nvc2l0eQ== 38390
+IG1vZGFs 38391
+IGdlc2Ft 38392
+IEh6 38393
+IG11bmljaXBhbGl0aWVz 38394
+IGNpcmN1bGF0aW5n 38395
+b2xpbmE= 38396
+U2hv 38397
+IEJlbmVk 38398
+b2x1 38399
+IHJlc3Rz 38400
+IGzDpW5n 38401
+INCe0LTQvdCw0LrQvg== 38402
+IHByemV3 38403
+IHBlcHA= 38404
+IG1hcnJpYWdlcw== 38405
+IEJJRw== 38406
+YW5kYW4= 38407
+IG1hZ2ljYWxseQ== 38408
+IGJhYnlz 38409
+IOuMkw== 38410
+IGhhY2tlcnM= 38411
+QmFieQ== 38412
+IE1vbnN0 38413
+IGNpZXI= 38414
+IEFyYWJz 38415
+INC80LDQs9Cw0Lc= 38416
+IEluZG9uZXNpYW4= 38417
+IE1hcmt0 38418
+IGRhY2h0ZQ== 38419
+IFNjaMO8bGVy 38420
+IFZORA== 38421
+IHNwaWVsdA== 38422
+IHBlcmx1 38423
+INC/0YDQvtGF0L7QtA== 38424
+IHNhbHRlZA== 38425
+IGltcHJvdmlz 38426
+IEluc3Ry 38427
+dmVsbWVudGU= 38428
+IG5lc3M= 38429
+IGZ1bmd1cw== 38430
+IGNvbGxhYm9yYXRvcnM= 38431
+IFZpcnVz 38432
+ZXN0YXI= 38433
+IHByb2plY3Rvcg== 38434
+INCf0YDQsNCy 38435
+IGFnaWxpdHk= 38436
+15nXoNeV 38437
+ZXJlbA== 38438
+INCy0L7Qt9Cy 38439
+INCx0LDQtw== 38440
+IENhdGh5 38441
+xJ91 38442
+INCz0L7QstC+0YDQuNC7 38443
+YmlsaXR5 38444
+IExhbmM= 38445
+IEtpbWJlcmx5 38446
+IEJyaWVm 38447
+IHV0dmVjaw== 38448
+IGdvZ2dsZXM= 38449
+IHByZXNjaG9vbA== 38450
+QVRIRVI= 38451
+IG1vdGl2ZXM= 38452
+IEJvbmc= 38453
+RVg= 38454
+IGNoaWxseQ== 38455
+IEFkdmlzb3J5 38456
+4oCL4oCL 38457
+INC60L7RgtC+0YDQvtC8 38458
+IHRyYWl0b3I= 38459
+IGRlbWFzaWFkbw== 38460
+INGG0LXQvQ== 38461
+INC80L7QuA== 38462
+IG11bHRpZg== 38463
+7JSs 38464
+IEFsZXhpcw== 38465
+IHppZXQ= 38466
+IFJhbWE= 38467
+YnJhbmNl 38468
+IHNhbmN0aW9u 38469
+aXRvdXM= 38470
+15XXmg== 38471
+IOuztOuC 38472
+0YHRgtCw0L3QvtCy 38473
+INGA0LXRgQ== 38474
+IENodXJjaGlsbA== 38475
+INC/0YDQtdC3 38476
+IElP 38477
+IEdlZQ== 38478
+IEdhdGhlcg== 38479
+YXRvcmk= 38480
+VHlsZXI= 38481
+INC90LXQvNC90L7Qtg== 38482
+IGLDpWRl 38483
+IEtpbGxlcg== 38484
+IHR1YmVy 38485
+IFJhbWFkYW4= 38486
+4b8= 38487
+aWVodA== 38488
+IHN0cmFuZ2VseQ== 38489
+0LvRgw== 38490
+IHJlZGVzaWdu 38491
+IGluY3VtYg== 38492
+IGJlcmFiZXI= 38493
+IFZvbGtzd2FnZW4= 38494
+bWV0YWw= 38495
+ZHp5 38496
+cGNpw7Nu 38497
+IOyViuyVhA== 38498
+IEdvb2RuZXNz 38499
+0LjQstCw0LXRgtGB0Y8= 38500
+YmFobg== 38501
+IEFudGFyY3RpY2E= 38502
+0LXQutGC0L7RgA== 38503
+IGhvbWVvd25lcnM= 38504
+emVpZ3Q= 38505
+IO2YhOyerA== 38506
+7KeA64+E 38507
+IGdlb2dyYXBoaWNhbA== 38508
+dGhpbmtpbmc= 38509
+IGdvc3Rh 38510
+IEltYW0= 38511
+dWxpZmxvd2Vy 38512
+ZGFn 38513
+YW5udA== 38514
+YWtvdg== 38515
+IGRvd253YXJkcw== 38516
+7LK06rCA 38517
+Q1VCRQ== 38518
+INCa0YHRgtCw0YLQuA== 38519
+INC/0L7Qu9C+0LI= 38520
+IHBsYXRlYXU= 38521
+4bil 38522
+IGNobG9yaW5l 38523
+IGFjY2VsZXJhdG9y 38524
+IHNvbHZlcw== 38525
+IEdyYXNz 38526
+cGlhbm8= 38527
+INqp2Kc= 38528
+INio2Ko= 38529
+IFJvY2hlc3Rlcg== 38530
+INmH2Yo= 38531
+IGNvbGxlY3Rz 38532
+jZTrnbw= 38533
+IENoZWVy 38534
+bGluZ2Vu 38535
+INGA0LDQt9Cz 38536
+IGFtw6lyaWM= 38537
+aHRh 38538
+RUNU 38539
+IGFydGlmaWM= 38540
+IFBheVBhbA== 38541
+aGFuYQ== 38542
+U3RlcGhlbg== 38543
+IEdlc3Q= 38544
+cGhhbHQ= 38545
+IHJlcGxpY2F0aW9u 38546
+IFdpbGxpZQ== 38547
+IG5ldXRy 38548
+IGlycmF0aW9uYWw= 38549
+IGRhZG9z 38550
+IEFpZA== 38551
+a2Ft 38552
+YW50ZXI= 38553
+INC00YPQttC1 38554
+IGRldG9u 38555
+IGhhcmU= 38556
+IGJldHM= 38557
+YmFnYWk= 38558
+IHN0YWluZWQ= 38559
+IHBsYXVzaWJsZQ== 38560
+IHBlZWxpbmc= 38561
+IGNyw610 38562
+IGdyb3Rl 38563
+7Law 38564
+pbTqsow= 38565
+YWx0ZXQ= 38566
+UGhvbmU= 38567
+Rmls 38568
+U1FM 38569
+IGdlZmFsbGVu 38570
+IHNhw7pkZQ== 38571
+IFRhbWls 38572
+Y291cw== 38573
+INCz0LvQsNCy0L3QvtC1 38574
+IGF0cmF2w6lz 38575
+dXNzaWE= 38576
+IHp3ZWl0ZW4= 38577
+IEVsdmlz 38578
+IG1vdmVy 38579
+IGxpbWl0ZQ== 38580
+YXJleg== 38581
+pbTqs6A= 38582
+IEtyYW5rZW4= 38583
+w7xyZQ== 38584
+IOyViuyVhOyalA== 38585
+IHRow6BuaA== 38586
+IHByb2ZvdW5kbHk= 38587
+IGJlZHJvb21z 38588
+IHRvb3RocGFzdGU= 38589
+IEFjY2VwdA== 38590
+w6l0aWNv 38591
+IGvDvMOn 38592
+IEFyeQ== 38593
+YWRpbg== 38594
+IGdyYW51bGFy 38595
+ZWN0ZWQ= 38596
+IG1lbmphZGk= 38597
+IGNvbXBldGVuY2U= 38598
+ZG9j 38599
+IHNwYXJrbGluZw== 38600
+IOyii+ydhA== 38601
+IGNvbnN0cnVjdGluZw== 38602
+IGFtdXNlbWVudA== 38603
+IEluc3VyYW5jZQ== 38604
+IEZldWVy 38605
+IHJlbm92YXRpb24= 38606
+c3VjaA== 38607
+cGxhdA== 38608
+IHByb3N0aA== 38609
+IGJleQ== 38610
+IENvbXBsZXRlbHk= 38611
+IHpvZA== 38612
+YWxu 38613
+VmljdA== 38614
+IGNvbmZpcm1z 38615
+w6R0eg== 38616
+4pY= 38617
+aGFtbWVy 38618
+INC30L3QsNC10YI= 38619
+IGFkbWlyZWQ= 38620
+oOulvA== 38621
+IEZydWl0 38622
+ZXJ0ZW4= 38623
+IG5pZWNl 38624
+IFRpbnk= 38625
+IHBsdW1iaW5n 38626
+ZXJtYQ== 38627
+INC70LXQs9C60L4= 38628
+IHdpbmRzaGllbGQ= 38629
+INGB0LzQtdGA 38630
+IGJ6dw== 38631
+IGFib2xpdGlvbg== 38632
+IFNhZGhndXJ1 38633
+IHByZWFjaGVk 38634
+IENyZWF0aW5n 38635
+cGVyZWQ= 38636
+IHZvbG9udA== 38637
+IHF1aW50 38638
+IHByaW50ZXJz 38639
+IG5lZ3Jv 38640
+IGdyb3NzZQ== 38641
+IFRoeQ== 38642
+IEZlbGxvd3M= 38643
+IHN0YW5pZQ== 38644
+IG5ld2NvbQ== 38645
+IEh1ZQ== 38646
+IEZyZXVuZGU= 38647
+IENvbnN0cnVjdGlvbg== 38648
+IGFkdmVyc2l0eQ== 38649
+IG5lZ2F0aXZlcw== 38650
+IGhhemFyZG91cw== 38651
+IGNvbXBlbGxlZA== 38652
+IHdvaw== 38653
+IE95 38654
+0L/QsA== 38655
+qqjr 38656
+IHJlbmRleg== 38657
+IG92ZXJj 38658
+IHdlYXZpbmc= 38659
+INC40LTQtdGC 38660
+IHByb3NlY3V0b3Jz 38661
+IGF1ZGlvYm9vaw== 38662
+IGFuY2VzdG9y 38663
+IHVuZGVyZ29pbmc= 38664
+IHBvdW5kaW5n 38665
+IO2SgA== 38666
+IOy2pA== 38667
+IHR1bGVl 38668
+IOyXtOw= 38669
+IHpvYWxz 38670
+IG5laW4= 38671
+IG9rZQ== 38672
+IEpveWNl 38673
+IG51ZA== 38674
+IGRpbGlnZW5jZQ== 38675
+IExhYnM= 38676
+IHZlbnRz 38677
+IGFuY2VzdHJhbA== 38678
+4Lir4Lih 38679
+INC80YPQttGH 38680
+IG5vbcOpcw== 38681
+d2FsaQ== 38682
+cWluZw== 38683
+IE11bHRpcGxl 38684
+IENvbnN1bHQ= 38685
+IGlzdGVkaQ== 38686
+IERveQ== 38687
+YWthaA== 38688
+IGRpc2NpcGxpbmVk 38689
+IGFsdGVybmF0aW5n 38690
+55I= 38691
+IHZlcm1l 38692
+INC+0Yk= 38693
+IHRvdGE= 38694
+IFByYWc= 38695
+IHN3b3Ju 38696
+IGJlYmVy 38697
+IEF1ZmdhYmU= 38698
+7Jq06w== 38699
+IHl1cA== 38700
+IHJlY2xhaW0= 38701
+b251dA== 38702
+IGF1Y3VuZQ== 38703
+IGFtcGg= 38704
+IMWbd2ll 38705
+IGFh 38706
+aXNjb3Zlcg== 38707
+IEFyZw== 38708
+Y2llxbw= 38709
+IGRlc3Nhcw== 38710
+IFfDpGg= 38711
+4bu5 38712
+INC00LDQstC90L4= 38713
+IHNpbGVudGx5 38714
+YXJj 38715
+IO2bhOuztA== 38716
+IHR3ZWV0aW5n 38717
+IE9uZA== 38718
+pqzrqbQ= 38719
+IGJvd2Vs 38720
+7IWo7Ja07JqU 38721
+T1NF 38722
+IHByb3Bpbw== 38723
+IEt1bnN0 38724
+a3VuZw== 38725
+IGRvbm7DqWVz 38726
+IEhvcml6b24= 38727
+IEZyb2c= 38728
+IGFyaXN0 38729
+w6Js 38730
+INC60L7Qtg== 38731
+IHNlZ3VuZG9z 38732
+IFNob3J0bHk= 38733
+IENyb3dk 38734
+aXJhbg== 38735
+IHfFgmHFm2Np 38736
+IExhYw== 38737
+aWRlbnRl 38738
+IOqwgOyekA== 38739
+IGxlbg== 38740
+IFNVUw== 38741
+IE1vdG9ycw== 38742
+IFRyZW50 38743
+b21pZQ== 38744
+IHRyYW5zbWl0dGVy 38745
+IEFzc2Fk 38746
+IHBzeWNoaWF0cmlj 38747
+INC20LjRgtGM 38748
+IG91dGxpbmVz 38749
+IGVmZmVjdGl2ZW1lbnQ= 38750
+IFJlbGlnaW9u 38751
+cHJlaA== 38752
+INC00L7Qu9C20L3QsA== 38753
+IM2hwrA= 38754
+IENvbnNlcnZhdGlvbg== 38755
+IOG7 38756
+INC30LDQuQ== 38757
+IHJlc2lkZQ== 38758
+IGNvbXBsZXRv 38759
+S0VO 38760
+IOuCmOyYpOuKlA== 38761
+IHN1YnVyYmFu 38762
+IHLDqXBvbmRyZQ== 38763
+INGA0LDQt9C70LjRhw== 38764
+IGdhbGxlcmllcw== 38765
+IHJhcHQ= 38766
+KS4uLg== 38767
+IGNydWVsdHk= 38768
+IFZNd2FyZQ== 38769
+7Yis 38770
+IGhhecSxcg== 38771
+IGdyb3VwaW5n 38772
+IFJpZGVy 38773
+IHN5bGxhYmxl 38774
+IGJlaXNwaWVsc3dlaXNl 38775
+IHNhZmVndWFyZA== 38776
+IHBlbMOtY3VsYQ== 38777
+YXJ0aQ== 38778
+INCh0L4= 38779
+IGNoZWdh 38780
+INC60L7QvNGD 38781
+IHNlaXNt 38782
+IGhhcm1sZXNz 38783
+IFdhcnJpb3Jz 38784
+INC/0YE= 38785
+IHNoYW1lbGVzcw== 38786
+IEJhdW0= 38787
+aW5zdGFsbA== 38788
+IHRvb2xraXQ= 38789
+IHBpcGVsaW5lcw== 38790
+IHB1c3N5 38791
+IGNvbmNlYWw= 38792
+IHByb3Rlc3Rpbmc= 38793
+b2Nob25k 38794
+IGR1YQ== 38795
+IFBvc2U= 38796
+IGhlbGl1bQ== 38797
+IFVY 38798
+aWtsZQ== 38799
+IFN1ZmY= 38800
+IOyEuOqzhA== 38801
+aW5nZXJz 38802
+INGB0LvRg9GH0LDQuQ== 38803
+IGRlc2NlbmRpbmc= 38804
+IG1vbnRhZ2U= 38805
+SGlnaA== 38806
+IOydtOyW 38807
+IElkaQ== 38808
+INeR16E= 38809
+IGV4cHJlc3NpdmU= 38810
+INC/0L7Qu9C10Lc= 38811
+IHBvbmU= 38812
+IGFkb2xlc2NlbnQ= 38813
+0LDQvdC90YvQtQ== 38814
+IGFzc2Fzc2luYXRpb24= 38815
+d2Vpc2Vu 38816
+ZW1hdGljYWxseQ== 38817
+YXV0aA== 38818
+IHVyZw== 38819
+IGdhbmhhcg== 38820
+IGZ1bmRv 38821
+IFJob2Rl 38822
+INC40YHRgtC+0YDQuNC4 38823
+IGNvbXBhcnRpbA== 38824
+IGRpbWluaXNoZWQ= 38825
+IGFwcHJlbnRpY2U= 38826
+INCR0YPQtA== 38827
+IHBob3RvbnM= 38828
+IGPDs2Q= 38829
+b25haw== 38830
+IGFkZWxhbnRl 38831
+IGNodQ== 38832
+b3BpYw== 38833
+IGFpeMOt 38834
+ZWRkYXI= 38835
+IENvbmdyYXRz 38836
+bW9y 38837
+IHJlc2VydmF0aW9ucw== 38838
+IFRvYnk= 38839
+IEtlcm4= 38840
+IHJhemVt 38841
+IGZvcmdlZA== 38842
+IGhvcnJpZnlpbmc= 38843
+2YrYuQ== 38844
+IEpvaW5pbmc= 38845
+IEF1dGg= 38846
+ZGFo 38847
+IGNvbnNpZw== 38848
+IGludGltaWRhdGVk 38849
+IHBlcmlwaGVyYWw= 38850
+IG1lbm8= 38851
+IGRldGVjdGluZw== 38852
+IHRlb3I= 38853
+IHRhZ2dlZA== 38854
+IG5vc3RhbGdpYw== 38855
+IOuvuOyViA== 38856
+IHZlcmRp 38857
+IGxhYmVsaW5n 38858
+0L/QvtC0 38859
+YXN0ZXM= 38860
+IHZpc3Q= 38861
+IGN5dA== 38862
+IGZsaXBz 38863
+0YDQuNC3 38864
+YmFsYW5jZWQ= 38865
+INC+0YjQuNCx 38866
+IGRlc3Rpbg== 38867
+bGFzc2U= 38868
+ZXJlaQ== 38869
+IGthbG8= 38870
+IGFycXU= 38871
+IHBsYW5v 38872
+IG9yZGluYW5jZQ== 38873
+IGNvbXBpbGF0aW9u 38874
+IFZvY8Oqcw== 38875
+IEVjbw== 38876
+IOy2lOyynA== 38877
+IGVuY2ltYQ== 38878
+IEdhcnJldHQ= 38879
+IENvcmQ= 38880
+w7Zsa2Vy 38881
+IEFycm93 38882
+IHByb3RvbnM= 38883
+LOKAiw== 38884
+IOyymOs= 38885
+IHNjYW5k 38886
+IGJlaWdl 38887
+Y29uZw== 38888
+IGJpa2luZw== 38889
+IFRM 38890
+0YPQvdC0 38891
+IOyGlOyngQ== 38892
+IFZpbGxh 38893
+IEpBQ0s= 38894
+IMO2xJ9yZW4= 38895
+IHRlbWFz 38896
+IEt5dW5n 38897
+SmVubg== 38898
+IGN1ZA== 38899
+IGltcG9zaW5n 38900
+IGNvbW1hbmRtZW50cw== 38901
+IE1lYW5z 38902
+IETDpHI= 38903
+IHJlY29tZW5k 38904
+IGRpc3Bvc2l0aW9u 38905
+2KfZhw== 38906
+IHRodQ== 38907
+IHJlZHVjdGlvbnM= 38908
+IGRpdQ== 38909
+INeV15DX 38910
+INC40YHRgdC70LXQtA== 38911
+dGhyZW4= 38912
+IGxhZG9z 38913
+IFJC 38914
+aXhlZA== 38915
+IOyP 38916
+RnI= 38917
+c3RpbGw= 38918
+IG9sbWFz 38919
+Q0hVQ0s= 38920
+IO2GoA== 38921
+IEluZGVwZW5kZW50 38922
+0JLQng== 38923
+IHBpdHM= 38924
+IHVuZGVydGFrZW4= 38925
+IGbDuHI= 38926
+IE5hdw== 38927
+IOyekeyXhQ== 38928
+IHNoZXBoZXJk 38929
+IGxhbmd1ZQ== 38930
+IEphYg== 38931
+IERydW0= 38932
+IEVsZWt0 38933
+4buRdA== 38934
+IOydtOyqvQ== 38935
+IGJlZ2lubmVu 38936
+IEZ1cnk= 38937
+4buDdQ== 38938
+c2VjdGlvbnM= 38939
+IHNwcmF5ZWQ= 38940
+IG3DoXI= 38941
+IFZvbHQ= 38942
+IFNlb25n 38943
+0LjRgtC10Ls= 38944
+ZHVjdGlvbg== 38945
+YXNhbg== 38946
+IGp1ZGdtZW50cw== 38947
+aW1hYW4= 38948
+nteq 38949
+IHNpZW50bw== 38950
+IEFDVA== 38951
+IEJI 38952
+ZGV2 38953
+IOyii+yVhO2VmA== 38954
+IGpvcm4= 38955
+SVNUSU4= 38956
+IHJvYXI= 38957
+IGltbWVyc2lvbg== 38958
+YWZmbGVz 38959
+IHRyYWluZWU= 38960
+IEJpbGxib2FyZA== 38961
+cmVzc2Vz 38962
+IFdhcm0= 38963
+IFJvYmVydG8= 38964
+IHV0aWxpeno= 38965
+IElnb3I= 38966
+IHJhc2g= 38967
+IGFuYWx5dGlj 38968
+aXJhbQ== 38969
+IHN5bW1ldHJpY2Fs 38970
+IGxpZmVzcGFu 38971
+IGVhdGVy 38972
+IEJsb29tYmVyZw== 38973
+YXRlcmlhbA== 38974
+IOuvvw== 38975
+IGlzdGVy 38976
+IGludmFsdWFibGU= 38977
+IGFzc2lzdGluZw== 38978
+IHNoYWNr 38979
+zrzOsc+EzrE= 38980
+amlz 38981
+ZW5peg== 38982
+INC/0YDQtdC00LvQvtC2 38983
+IGRlY2xhcmluZw== 38984
+IFZpa2luZw== 38985
+IEFzc2lt 38986
+IGV4cGVuZGl0dXJl 38987
+IHBvc2luZw== 38988
+IE9udW4= 38989
+IGluaWM= 38990
+0LDRjtGC0Yw= 38991
+cmV2 38992
+IG1pZWRv 38993
+IGZpbHRoeQ== 38994
+IElC 38995
+IERpc2NvdmVy 38996
+aWNodGV0 38997
+bWlsbGlvbg== 38998
+toTrk6TsnbQ= 38999
+IGFtYmlndQ== 39000
+IEZseW5u 39001
+YmFyZHppZWo= 39002
+IGluY29tcA== 39003
+0LDQstC90L4= 39004
+emlh 39005
+IGluZmx1ZW5jaW5n 39006
+IHdvcmxkbHk= 39007
+IFNhbGVzZm9yY2U= 39008
+emV0 39009
+IHBhcnRpY3VsaWVy 39010
+IEtvY2g= 39011
+IDE5NDM= 39012
+IHRvbmVy 39013
+INGN0LrRgdC/0LXRgA== 39014
+IHN1c2NyaQ== 39015
+IHRyaWdnZXJpbmc= 39016
+SUNFUw== 39017
+7Iqk6rCA 39018
+zrTOsQ== 39019
+0YDQsNCx0L7Rgg== 39020
+IGFmdGVyd2FyZA== 39021
+cGluZQ== 39022
+IElM 39023
+YXJldGg= 39024
+INC/0LDQuw== 39025
+IHNha2Vy 39026
+IDE5NDc= 39027
+QUY= 39028
+dXlvcnN1bg== 39029
+IOyKpOs= 39030
+IHF1YW50aWZ5 39031
+IG1lbnRvcnNoaXA= 39032
+IGxsZWdh 39033
+IFRhbWFyYQ== 39034
+IG9wdGltaXppbmc= 39035
+IGZyb250cw== 39036
+b3N0ZXJz 39037
+IGVzcXVlcg== 39038
+IHN1Ym1pc3Npb25z 39039
+IGFubmlo 39040
+IHN1Y3Rpb24= 39041
+bHVlbmNl 39042
+Y2hpZWRlbg== 39043
+SU5HUw== 39044
+INeR15Q= 39045
+INGB0YbQtdC9 39046
+IHdpZWx1 39047
+IG9iamV0bw== 39048
+IGJvb2Jz 39049
+IEdlc2Now6RmdA== 39050
+IGVhcmJ1ZHM= 39051
+INGA0LDQvdGM0YjQtQ== 39052
+IHJvdXRpbmVseQ== 39053
+IGNvbGxhZ2Vu 39054
+0L7QtNGL 39055
+IENpbm5hbW9u 39056
+IGJhaXg= 39057
+2K/ZhQ== 39058
+ZnJhZ2U= 39059
+INC60L3QvtC/ 39060
+IGRlY2VwdGlvbg== 39061
+IHVuZXhwZWN0ZWRseQ== 39062
+IHNtZWxsZWQ= 39063
+IGxvb3M= 39064
+IGhpZ2hsaWdodGVy 39065
+IOq4sOuzuA== 39066
+IEdsYXNnb3c= 39067
+b3dhbmE= 39068
+bW4= 39069
+IEplcmVtaWFo 39070
+IERhdGFi 39071
+aWV0ZQ== 39072
+IGJhdw== 39073
+IHByb3BpYQ== 39074
+IHByb3ByaQ== 39075
+T09PT09PT08= 39076
+aW5rZXI= 39077
+IHBlcnR1cmI= 39078
+IEZha2U= 39079
+7J207JY= 39080
+aW1taW5n 39081
+IHVuZG9jdW1lbnRlZA== 39082
+IHRyYWJhamFuZG8= 39083
+IHJvYW0= 39084
+INC00L7Qu9C20L3Qvg== 39085
+IGFyYmU= 39086
+IGFuaQ== 39087
+YXRhbA== 39088
+IGFyYWRh 39089
+IEFuZGE= 39090
+IOybgA== 39091
+IEJyYW5jaA== 39092
+b2lyZXM= 39093
+IG91dHNpZGVy 39094
+ZG9sbGFy 39095
+aXNzZXM= 39096
+YmVhbnM= 39097
+IEdpZw== 39098
+cmFkb3M= 39099
+IFN1dA== 39100
+IExhbmNl 39101
+ZWRzacSZYmlvcg== 39102
+IGNvbGE= 39103
+b25lbnRz 39104
+IHJlY29uc2lkZXI= 39105
+IG1vbmRv 39106
+IHVuc3VjY2Vzcw== 39107
+IEvDpA== 39108
+IHJlZ2Vs 39109
+IGJpc29n 39110
+ZXR1cw== 39111
+IHVucmF2ZWw= 39112
+IHN3ZWV0aWU= 39113
+IHJlcHLDqXNlbnQ= 39114
+b3VyaW5n 39115
+IGdyb3VuZHdhdGVy 39116
+IEJldw== 39117
+IHNjcmF0Y2hlZA== 39118
+IGNhc3NldHRl 39119
+IGNpZGVy 39120
+cGlz 39121
+INGB0LDQvNCw 39122
+IGdsb2JhbGl6YXRpb24= 39123
+IGRlZ3JhZGF0aW9u 39124
+IGRlZ2VuZXI= 39125
+IFJvc2ll 39126
+aWNrdA== 39127
+IG92ZXJ3ZWlnaHQ= 39128
+IE1FTQ== 39129
+IGd1YXJkaWFucw== 39130
+IGNvbnNlYw== 39131
+SG1t 39132
+INC/0L7RgtGA0LXQsQ== 39133
+IG1ldmE= 39134
+IGdyYWZmaXRp 39135
+IGZsaXJ0 39136
+IEJQ 39137
+IGp1c3Rv 39138
+IFRob3VzYW5kcw== 39139
+n6zsmrQ= 39140
+Lio= 39141
+IFJBVw== 39142
+IGZsdW9y 39143
+aXlp 39144
+YW50YWw= 39145
+amVk 39146
+IFNoZW5n 39147
+IEVsaXNl 39148
+IENoYXJnZQ== 39149
+7J207Yq4 39150
+IGNvbmVz 39151
+bmllcw== 39152
+Z2lh 39153
+INC90LDRh9Cw0LvQsA== 39154
+IERoYXJtYQ== 39155
+IOuLpOyWkQ== 39156
+IGZhdm9ycw== 39157
+IFRydW5n 39158
+aGV0dG8= 39159
+IHBvenc= 39160
+IGxvbmdv 39161
+IGtlbHU= 39162
+IGRpZ2VzdGlvbg== 39163
+IEVpZw== 39164
+IFRIRVJF 39165
+IHRpZXJz 39166
+IHN1bms= 39167
+IG15c3RpY2Fs 39168
+enVi 39169
+IMOJdA== 39170
+IGFudGljaXBhdGluZw== 39171
+IFZpbmU= 39172
+WVk= 39173
+IGNvbmNlbnRyYXRpbmc= 39174
+IEFncmVlbWVudA== 39175
+INC+0LrQvtC70L4= 39176
+IGxpZHQ= 39177
+IFlhbw== 39178
+INGB0LvQuNGI0LrQvtC8 39179
+csOt 39180
+SVNUSU5DVA== 39181
+IE9GRklD 39182
+IHNvYWtpbmc= 39183
+IHNpaWhlbg== 39184
+IHJlZmVyZW5jaW5n 39185
+IFRhbXBh 39186
+YW5leQ== 39187
+IHJlc3B1ZXN0YQ== 39188
+IENvYWxpdGlvbg== 39189
+INGB0L7Qs9C70LDRgQ== 39190
+YW5raW5k 39191
+IOub 39192
+IFl1bW15 39193
+67Cw 39194
+IG9uYw== 39195
+dWnDp8Ojbw== 39196
+IHRoZW8= 39197
+IG11cmFs 39198
+IFRlYWNoZXJz 39199
+IHdhaXRz 39200
+IHJlbnRpbmc= 39201
+IEhhcm1vbg== 39202
+IGXFnw== 39203
+IE11bmljaA== 39204
+7Zmc 39205
+7Ja8 39206
+Y2FyZHM= 39207
+IHJvdWdl 39208
+IG7Dqm4= 39209
+Y2x1Yg== 39210
+IHVuc2Vlbg== 39211
+IGRlcHJlY2k= 39212
+IGNvbXB1dGVk 39213
+IHdpcGluZw== 39214
+IEVsbGk= 39215
+aWRlbnRpZmllZA== 39216
+IGNsdXR0ZXI= 39217
+cm9sZXVt 39218
+IHRlbGVm 39219
+IGxldmVsaW5n 39220
+IFdvb2R5 39221
+IEd1cw== 39222
+IEJlbm5ldHQ= 39223
+IHNpdGlv 39224
+acWC 39225
+IHBvc3Nlc3Npb25z 39226
+IE5hdGFzaGE= 39227
+b2xkb3du 39228
+INGB0L7QvtCx0Yk= 39229
+IExpYw== 39230
+IOunjOuToA== 39231
+IGxvcnNxdWU= 39232
+d2Vo 39233
+INC80LDQvA== 39234
+bGl0ZXI= 39235
+YWRvbW8= 39236
+IGZpbmk= 39237
+z47Pgg== 39238
+INGD0LHQuNC5 39239
+IGluZGlzcA== 39240
+IHRlbGV2aXM= 39241
+IHDDoQ== 39242
+IENyZW8= 39243
+w61sbA== 39244
+IGd1cg== 39245
+IE1BTA== 39246
+INGA0LDQt9C90YvRhQ== 39247
+IHppZWhlbg== 39248
+IGZhc2hpb25lZA== 39249
+IGRlYmF0aW5n 39250
+IFNvdXA= 39251
+IFByb3ZpbmNl 39252
+6re466CH 39253
+IGltcHJvcGVy 39254
+IGltYWdlbg== 39255
+INGB0LTQtdC70LDQuw== 39256
+IGxvZ29z 39257
+IGV2ZW50bw== 39258
+4bqjbw== 39259
+bGFyZGE= 39260
+INC90LDQt9GL0LLQsNC10YLRgdGP 39261
+IHZlcmY= 39262
+IHNjcmVlbnNob3Rz 39263
+15XXk9ei 39264
+IEF1cm9yYQ== 39265
+IEJhbGk= 39266
+dGVyZWQ= 39267
+IGNvbnRhZ2lvdXM= 39268
+IGNvbXBhcnRpcg== 39269
+dmVuaWRvcw== 39270
+cmlrZQ== 39271
+INCy0YvQs9C70Y/QtNC40YI= 39272
+IGZyZWVkb21z 39273
+bmljYXM= 39274
+oKTshJw= 39275
+IHJlZHV6 39276
+IEVjdQ== 39277
+IGFib25u 39278
+IFNFw5E= 39279
+IEJpdGNo 39280
+IHByb2pldG8= 39281
+0LjRh9C90L4= 39282
+ZXR0cmU= 39283
+QU5OQQ== 39284
+dGhhbms= 39285
+IEFP 39286
+YXJuaXNo 39287
+aWXDn2Vu 39288
+IHJpcHBsZQ== 39289
+IHBhbnRyeQ== 39290
+IEdI 39291
+zrPOsQ== 39292
+IOydtOuyiOyXkA== 39293
+IHZhbGlkYXRlZA== 39294
+IGJydXNoZWQ= 39295
+IEVtaW4= 39296
+IERhcnRo 39297
+ZXNpbg== 39298
+LC4= 39299
+IHZhbGxl 39300
+IGplcnNleQ== 39301
+dWxhbg== 39302
+UmVhZA== 39303
+IFJhbmdlcnM= 39304
+IHNvb3RoaW5n 39305
+IGNvbXBsZW1lbnRhcnk= 39306
+IFZlcmtlaHI= 39307
+YWNha3Q= 39308
+IGJhdGh0 39309
+IE5E 39310
+U29u 39311
+IO2ZlOyepQ== 39312
+IEF2aQ== 39313
+IFNBTA== 39314
+YWlzc2U= 39315
+IHNlbWFpbmVz 39316
+IFN1cnY= 39317
+d2llcg== 39318
+INCy0LjQtNC10Ls= 39319
+IHNpZXRl 39320
+lOuPhA== 39321
+IFJhbXNheQ== 39322
+IFF1ZWVuc2Jvcm91Z2g= 39323
+IE1lbmdl 39324
+IEZvb2Rz 39325
+IHRoZW9sb2dpY2Fs 39326
+IFsj 39327
+INCy0L7QvdC4 39328
+IGltbWlu 39329
+aW9zaXR5 39330
+IEFiZ2VvcmQ= 39331
+IEFjaG8= 39332
+IMOU 39333
+IHN0YWlucw== 39334
+IHJlYWxpc3RpY2FsbHk= 39335
+IGZhc2hpb25hYmxl 39336
+IENFT3M= 39337
+IFNraWxs 39338
+INCy0LbQtQ== 39339
+IGRldmVy 39340
+IFBsdWc= 39341
+5qo= 39342
+UG9k 39343
+IGxvYWY= 39344
+IGdlYnJhY2h0 39345
+IGFic29yYnM= 39346
+IEdyYW5ueQ== 39347
+IG1hbHdhcmU= 39348
+YWfEmQ== 39349
+IGNpdmlsaXphdGlvbnM= 39350
+IM+B 39351
+IGjDpGx0 39352
+0KHQog== 39353
+Z3JlYXQ= 39354
+IGxheWVyaW5n 39355
+c2luZ3M= 39356
+INCy0ZbQvQ== 39357
+IHJlY29nbml6YWJsZQ== 39358
+IHdvag== 39359
+IHdldGVu 39360
+zrPOvw== 39361
+U3R1ZGVudA== 39362
+IGTDqWZpbg== 39363
+cGxlYXNl 39364
+ZW5jaA== 39365
+IGF0dGlj 39366
+IE90dGF3YQ== 39367
+IG9wdGVk 39368
+IGNhcHRpdg== 39369
+IG3Fgg== 39370
+IFlB 39371
+IFdhbmQ= 39372
+IGJvdW50eQ== 39373
+IDI3MA== 39374
+IHNwZWN1bGF0ZQ== 39375
+IGVuaGFuY2VtZW50 39376
+IGNvbW1vZGl0aWVz 39377
+IE1pbHRvbg== 39378
+ZWo= 39379
+YWxvbQ== 39380
+RGFz 39381
+IGNvb2xkb3du 39382
+16jXkNec 39383
+INeQ16Q= 39384
+IHdjemXFm25pZWo= 39385
+IGVsb25n 39386
+IGRpb2Rl 39387
+aW5hw6fDo28= 39388
+IElyaXM= 39389
+IEli 39390
+IHN1bW1vbmVk 39391
+IHJlc3Bl 39392
+IFJhY2g= 39393
+IMK7Og== 39394
+IHZ1cg== 39395
+IG1vdmltZW50bw== 39396
+IGZsdWVudA== 39397
+IEV2b2x1dGlvbg== 39398
+IEJ1dHQ= 39399
+aWZpY2FjacOzbg== 39400
+lJTslrQ= 39401
+INGN0L3QtdGA0LM= 39402
+IG1hbmlwdWxhdGluZw== 39403
+IHBvc2l0aXY= 39404
+0LzQvtGB 39405
+IHdpeg== 39406
+IGludG94 39407
+zq3PgQ== 39408
+0LXQvNGB0Y8= 39409
+aXZlc3Nl 39410
+aW1pemk= 39411
+IOyauA== 39412
+IGtub2Nrcw== 39413
+IGNvbmdlc3Rpb24= 39414
+IElkZWFsbHk= 39415
+IEhvbGRpbmc= 39416
+IHBvYnJl 39417
+IEpVTA== 39418
+IOu2hOuTpOydgA== 39419
+IM6xzro= 39420
+IEZlcmd1c29u 39421
+IExhYm9yYXRvcnk= 39422
+cmljaHRlbg== 39423
+cm9waHk= 39424
+cHJvZHVjdGlvbg== 39425
+YXNzdW5n 39426
+SVRB 39427
+IHNpw6hjbGU= 39428
+16jXqg== 39429
+Y2lzaW9u 39430
+INek15Q= 39431
+IElyZW5l 39432
+YW5jYQ== 39433
+IOyCrOqzoA== 39434
+IHBpbnBvaW50 39435
+IGRlc2lnbmF0aW9u 39436
+xZ9hbQ== 39437
+bMSxxZ8= 39438
+YWF0 39439
+IG7DpWdyYQ== 39440
+IG15dGhpY2Fs 39441
+IERlY2xhcmF0aW9u 39442
+IOyeoeyVhA== 39443
+IGJ5dGU= 39444
+LuKZqg== 39445
+RGVs 39446
+IO2NvA== 39447
+IG51dHJpdGlvdXM= 39448
+INGA0YPQsdC70LXQuQ== 39449
+U0FZ 39450
+TWFzdGVy 39451
+INGE0L7RgtC+0LPRgNCw0YQ= 39452
+IOuSpOyXkA== 39453
+IG5laA== 39454
+IGRva3VtZW50 39455
+IGN6YXN1 39456
+IGNvbnRpbnVh 39457
+IFNpbGVudA== 39458
+IHRlbnNvcg== 39459
+IHRhbnRh 39460
+IGlyZ2VuZHdv 39461
+IExFVA== 39462
+IFNoYWt0 39463
+bGFtYQ== 39464
+Y2hsYWc= 39465
+IGRpbmdlbg== 39466
+0YHRgtGA0LA= 39467
+IGVocmxpY2g= 39468
+IE1hY2h0 39469
+cmVscw== 39470
+w6BjaWVz 39471
+dmlkZW8= 39472
+IG5hdHVyYWxl 39473
+IFNURVZF 39474
+dW1t 39475
+QkFDSw== 39476
+IDcyMA== 39477
+IG1vbWVuY2ll 39478
+IFN3YW4= 39479
+IHRlY2huaWNpYW5z 39480
+IGdlZWhy 39481
+IE1lbmQ= 39482
+UmVn 39483
+IHNjYWZm 39484
+IGFpZGU= 39485
+IOuztOuKlA== 39486
+IHByZXNzZXM= 39487
+bGVyZGU= 39488
+XCc= 39489
+IHVsdHJhc291bmQ= 39490
+IGRpc2NsYWltZXI= 39491
+IE1pdHM= 39492
+IEhvbGlkYXk= 39493
+IGV4dGVybmFsbHk= 39494
+IEZhdGU= 39495
+SU5P 39496
+IENhdHM= 39497
+67CV 39498
+dW1v 39499
+Y29udHJvbA== 39500
+IHRoZUNVQkU= 39501
+dGlj 39502
+aWVydW5ncw== 39503
+INC30L3QsNC60L7QvA== 39504
+IGZyZWVzdHlsZQ== 39505
+TUFOREFSSU4= 39506
+IGlzZQ== 39507
+YXVydXM= 39508
+IFN0cmF0ZWd5 39509
+IEJlYW0= 39510
+csOkZ2U= 39511
+IGV4cGxvaXRlZA== 39512
+aWRpcw== 39513
+IGNoaW1l 39514
+IFBlbmluc3VsYQ== 39515
+IG1lcml0cw== 39516
+IGFsdHJv 39517
+IFRPUA== 39518
+IFNlbnM= 39519
+IEthbnQ= 39520
+b3Jhcw== 39521
+IHJveWFsdHk= 39522
+IElERQ== 39523
+cmFjeQ== 39524
+IFRIT00= 39525
+b21vcw== 39526
+IGzDpG5nZXI= 39527
+IG51bWJlcmVk 39528
+VW0= 39529
+IE5peWU= 39530
+zrjOtw== 39531
+enlrYQ== 39532
+bGltZQ== 39533
+IFBlcnNvbmVu 39534
+IHZhbGlkaXR5 39535
+IGNvbnRyYXQ= 39536
+IENvbWlj 39537
+w6dvbnM= 39538
+IEhlaWRp 39539
+IHpn 39540
+IHJlbmFtZWQ= 39541
+IGN1bWlu 39542
+IEpG 39543
+aW5lbA== 39544
+IGVuZm9yY2Vk 39545
+IGNoYW1h 39546
+0LvQuNGH0L3Qvg== 39547
+4bq7 39548
+INC00LXQvdC10LM= 39549
+IHByb2Z1bmQ= 39550
+IHBlbHZpYw== 39551
+IHBhbGF2cmE= 39552
+IGV4dHJhcw== 39553
+IGFua2xlcw== 39554
+7JeQ7ISc64+E 39555
+IFRG 39556
+IGluc2FuZWx5 39557
+INC80Y/RgQ== 39558
+IHLDqXBvbnNl 39559
+IGfDtnN0ZXI= 39560
+IEJCUQ== 39561
+INGD0YfQsNGB0YI= 39562
+IHNoYWtlbg== 39563
+IGFsbW9uZHM= 39564
+ZGlzaA== 39565
+IFBH 39566
+IEJsaXp6YXJk 39567
+0YzQvtCz0L4= 39568
+IOOF 39569
+IGtuYXBw 39570
+VG9v 39571
+IHVuZGU= 39572
+IG1vdW50cw== 39573
+0L7QvNC40L3QsA== 39574
+IG5vcnRoZWFzdA== 39575
+IGNlbnNvcnNoaXA= 39576
+0Y/RgtGM0YHRjw== 39577
+bHI= 39578
+IGxhd21ha2Vycw== 39579
+IHPDpWRhbg== 39580
+IGluc2lkZXI= 39581
+IGNsZWFudXA= 39582
+IE5hZGE= 39583
+w7Nj 39584
+IGhhcnZlc3RlZA== 39585
+IERlc3B1w6lz 39586
+7ZqN 39587
+IHJlZHVuZGFudA== 39588
+RU5B 39589
+IGRlbGVnYXRl 39590
+IGJ1cmc= 39591
+IEFsaXNvbg== 39592
+IGNlbGVzdGlhbA== 39593
+IHNpbm5lcnM= 39594
+IG1hcnR5cg== 39595
+IFBlcm0= 39596
+IHNwZWNpbWVucw== 39597
+IG1pdG9jaG9uZA== 39598
+IG1hcmF2aWw= 39599
+IGNhdmFscnk= 39600
+IGFycmF5cw== 39601
+IGFubmV4 39602
+IGxhYm9yYXRvcmllcw== 39603
+IEJ5eg== 39604
+IGF0YWM= 39605
+INGB0LvQvtC20L3Qvg== 39606
+IHRvcGw= 39607
+IGdlcmk= 39608
+IENvbWJhdA== 39609
+0YHRj9GC 39610
+ZWtlbg== 39611
+INCS0LvQsNC0 39612
+IGFqdXN0 39613
+IG1hcnF1ZQ== 39614
+IGxvb2tvdXQ= 39615
+IExvbA== 39616
+IHJvb2Z0b3A= 39617
+IE9yaW9u 39618
+INCx0L7QuQ== 39619
+IGhlYXJ0YnJlYWtpbmc= 39620
+IGRldHRv 39621
+emg= 39622
+w6R0dGVy 39623
+Y2VyYQ== 39624
+IGhlYXRz 39625
+IGFudGlxdQ== 39626
+IHVuZmluaXNoZWQ= 39627
+IEthenU= 39628
+xLFsxLE= 39629
+IHNsaWdodGVzdA== 39630
+bGVv 39631
+IHbDpXJh 39632
+IHZlcnNjaGllZGVuZW4= 39633
+IGxvdGlvbg== 39634
+0YjQtdCz0L4= 39635
+Y3Rpb25hbA== 39636
+IOydtOyg 39637
+ZHJhZ29u 39638
+IHJlc29uYXRlcw== 39639
+IGlubQ== 39640
+YXZpYw== 39641
+IGZ1bGZpbA== 39642
+IOq4sOuMgA== 39643
+IGp1c3RhbWVudGU= 39644
+INC00L7RgdGC0YPQvw== 39645
+IOq3uOqxtA== 39646
+IHJlY29uY2lsZQ== 39647
+IFNjaMO2bg== 39648
+IEF2b2lk 39649
+6rmA 39650
+J0Q= 39651
+IGNvbmZpbmVtZW50 39652
+IO2R 39653
+IG1vdGl2YXRpbmc= 39654
+IEJyaXR0YW55 39655
+IHNjcmVhbWVk 39656
+b2JqZWN0 39657
+IGRlY3JlZQ== 39658
+IHRyYXZhaWxsZQ== 39659
+aXNzaWJsZQ== 39660
+IGJ1c3RlZA== 39661
+cHJvY2Vzcw== 39662
+IG1hc3NhY3Jl 39663
+IG5naMSp 39664
+aWx5bg== 39665
+INCy0YDQvtC00LU= 39666
+IHBvZXRpYw== 39667
+IG5o4bqldA== 39668
+IGlyb25pY2FsbHk= 39669
+dXN1 39670
+bmlv 39671
+IHN0YWdpbmc= 39672
+b21lZGljYWw= 39673
+bGVhc2Vk 39674
+IOyDiOuhnOyatA== 39675
+IE5a 39676
+YWN0aW5n 39677
+IEJhdHRsZWZpZWxk 39678
+cGxheWZ1bA== 39679
+Vmk= 39680
+IHNlw7FvcmE= 39681
+IHByb21wdHM= 39682
+bGljaGtlaXQ= 39683
+IMOnxLFrYXI= 39684
+amlhbmc= 39685
+IHBpY2t5 39686
+IENhdmU= 39687
+IG1pcmFjdWxvdXM= 39688
+IEh1Z2hlcw== 39689
+MjAxNg== 39690
+IHh1 39691
+IERvcm90aHk= 39692
+IHZpcnR1ZXM= 39693
+IHJldHJhY3Q= 39694
+IHR5cg== 39695
+IGNoYXJpc21hdGlj 39696
+IGJvbGE= 39697
+6bw= 39698
+IOunkOyUgOs= 39699
+IHBhcmVudGFs 39700
+IG1pbGxpb25haXJl 39701
+YXJpYXQ= 39702
+IGludm9rZQ== 39703
+xbxlbmll 39704
+IGV4dHJlbWVz 39705
+IEFrdQ== 39706
+aXZpZGFkZQ== 39707
+IO+3ug== 39708
+IOyLnOyyrQ== 39709
+IEdhcmxpYw== 39710
+UklB 39711
+INC00L7RgQ== 39712
+IFBvbnQ= 39713
+IG1pbGo= 39714
+ZWxsaQ== 39715
+IHJhY2tldA== 39716
+IGNvbXBldGl0 39717
+IFdoaXM= 39718
+IHJlYWx0 39719
+aWdubWVudA== 39720
+ZXN0cmU= 39721
+IHBlcm5haA== 39722
+IE9wZW5pbmc= 39723
+IEZT 39724
+IERlbW9rcmF0ZW4= 39725
+YWNlbWVudHM= 39726
+IHdvcmxkdmlldw== 39727
+IHBsYXlvZmZz 39728
+IENBRA== 39729
+IMOpdGFudA== 39730
+IHllbWVr 39731
+IHNlbnRpbWVudHM= 39732
+b2RlbA== 39733
+YnVzdGVy 39734
+YcWf 39735
+IEtZ 39736
+Y3rEmQ== 39737
+IHNjaMO2bmU= 39738
+YXBl 39739
+IFJhc3BiZXJyeQ== 39740
+IGNyZWRpdGVk 39741
+IEhpZGRlbg== 39742
+IHNhdXNhZ2Vz 39743
+cnVjZQ== 39744
+IEJldg== 39745
+aWxhbnRybw== 39746
+IHBva2Vtb24= 39747
+IOqwgOqyqQ== 39748
+IHByb2NlZWRpbmc= 39749
+IHZlaW8= 39750
+IDE3NQ== 39751
+6Lg= 39752
+bWF4 39753
+IGZyYXRlcg== 39754
+7KCE7JeQ 39755
+IGVnZW50 39756
+IDI1MDA= 39757
+dXNjaA== 39758
+VHViZQ== 39759
+IGFtcGxpZnk= 39760
+IHByYXdk 39761
+IG9kb3I= 39762
+IFNjYW4= 39763
+IHBsb3R0aW5n 39764
+aXRobWV0aWM= 39765
+IHJlc2lnbmVk 39766
+IFNDT1RU 39767
+IHN0ZXJlb3R5 39768
+IGRvYWJsZQ== 39769
+IENvbXBsZXg= 39770
+2YHZig== 39771
+dMSxbQ== 39772
+0YDQuNCz 39773
+bGFyZGFu 39774
+ZXNv 39775
+REVO 39776
+IGhvb2RpZQ== 39777
+IENBVA== 39778
+2KfYtw== 39779
+IGJvbmRlZA== 39780
+IEJ1cm5z 39781
+0L7Qv9Cw0YE= 39782
+IHLEmQ== 39783
+zrXOuc6x 39784
+INC+0YLQtNC10LvRjA== 39785
+IHRpbWVsZXNz 39786
+IFZpag== 39787
+IFBhbmFtYQ== 39788
+IHJlb3JnYW4= 39789
+IFTDpA== 39790
+IFBsdXRv 39791
+T3Jhbmdl 39792
+INC/0L7QudC0 39793
+IEJyaXN0b2w= 39794
+dWNlZA== 39795
+IOuQmOyWtA== 39796
+IHVuYmVkaW5ndA== 39797
+YWRsZQ== 39798
+IHZvbHVudGVlcmVk 39799
+IG1pZWxp 39800
+IEVkaW5idXJnaA== 39801
+aWthbA== 39802
+IGFsdGVu 39803
+IEFyc2Vu 39804
+IG1vdXZlbWVudA== 39805
+IGFudGlxdWU= 39806
+IGJo 39807
+IEhlcnM= 39808
+IHNhdXRl 39809
+IGFzcGlyZQ== 39810
+IHNwaGVyZXM= 39811
+IFdhbQ== 39812
+4bqvbQ== 39813
+IHdpcGVz 39814
+IDI4MA== 39815
+IFZlaA== 39816
+IGNvbG9jYQ== 39817
+0LDRhA== 39818
+INCy0L7Qt9C80L7QttC90L7RgdGC0Yw= 39819
+IHBoeXNpb2xvZ2ljYWw= 39820
+aHdh 39821
+ZXR1 39822
+IHByb2xvbmdlZA== 39823
+IGV4cGVyacOqbmNpYQ== 39824
+INCy0LjQtNC90L4= 39825
+IHF1YXJhbnQ= 39826
+IHB1ZWRhbg== 39827
+6JQ= 39828
+dmluZQ== 39829
+IFVTREE= 39830
+cGhlbQ== 39831
+IGZvcm1pZGFibGU= 39832
+IGZsYXR0ZXI= 39833
+7Ja07KeA 39834
+IGLDqW4= 39835
+4LmB4LiV 39836
+IOusvOuhoA== 39837
+IGZhY3Rpb25z 39838
+IExlYXZpbmc= 39839
+INeQ16rXlA== 39840
+IEV4cGVydA== 39841
+ZGlv 39842
+IFZlcmQ= 39843
+IHNpbnQ= 39844
+2YbYrw== 39845
+bnVtYmVy 39846
+IG93ZWQ= 39847
+IGluZHVjZQ== 39848
+IEZyZWRkaWU= 39849
+YWJv 39850
+IEZpbGlwaW5v 39851
+r7zr 39852
+YmVsaWV2YWJseQ== 39853
+YXRobG9u 39854
+YW1hYW4= 39855
+IGRldmVuaXI= 39856
+IEdvcw== 39857
+IEplbmtpbnM= 39858
+YmFpdA== 39859
+IGJpbnM= 39860
+IE1JQ0g= 39861
+dXlvcnVt 39862
+aWdyYWRl 39863
+aXNzbw== 39864
+IOyXtA== 39865
+IOyVhOu5oA== 39866
+IGRpYXJyaGVh 39867
+IHRvcm5hcg== 39868
+YWRkaW4= 39869
+IHVuZ2Vmw6Rocg== 39870
+IHJlc3Ryb29t 39871
+IHBzeWNoaWF0cmlzdA== 39872
+IEtpY2tzdGFydGVy 39873
+IGdlcmE= 39874
+IGFscmVk 39875
+IFdyYXA= 39876
+z4zPgw== 39877
+IHNpbm5lcg== 39878
+Q0hFRVJJTkc= 39879
+IGtpbG93 39880
+IGRldGVybWluYW50 39881
+IGRlbW9uaWM= 39882
+aWRlbmNlcw== 39883
+Y2hhcw== 39884
+IERlZA== 39885
+IHN0dW1ibGU= 39886
+IFVycw== 39887
+IGRlY2VpdmVk 39888
+IFRFUg== 39889
+IEPDsw== 39890
+ZWxsZWQ= 39891
+IG5vdHdlbmQ= 39892
+IOyngOq4iOq5jOyngA== 39893
+IHBhcnRpZG8= 39894
+IGRlc2NlbmRlZA== 39895
+IHZhcmTEsXI= 39896
+IGVuYWN0ZWQ= 39897
+IGN6xJnFm2Np 39898
+IHRyYWluZWVz 39899
+IGF1ZGlibGU= 39900
+IG1hbGY= 39901
+IHZlbw== 39902
+w6xu 39903
+IEdQQQ== 39904
+IEFwcGU= 39905
+IHJ1dA== 39906
+IENhcmxh 39907
+a2FjaA== 39908
+IHNhdmlvcg== 39909
+aXRjaGVk 39910
+IGNsaW1heA== 39911
+0LDRgtC10LvRjw== 39912
+IE1jQ29ubmVsbA== 39913
+0L7Qu9GP 39914
+ZXJleWU= 39915
+INGB0L7Qt9C9 39916
+IGNhYm8= 39917
+IFNuZQ== 39918
+IEFmZm9yZGFibGU= 39919
+IHNhcsOg 39920
+IGxlZ2l0aW1hY3k= 39921
+IHNjYXJjZQ== 39922
+Li4uPC8= 39923
+IDEwOA== 39924
+IGFjdW0= 39925
+IEZyYW5rbHk= 39926
+IHJhZGlhdG9y 39927
+IGdlbmVyYWxz 39928
+IGRpdmlkZXM= 39929
+IGNoZWVzZWNha2U= 39930
+IHNvcmNlcg== 39931
+IG1pc2NvbmNlcHRpb24= 39932
+IGhhcmRzaGlwcw== 39933
+IE9uZVBsdXM= 39934
+w7x5b3JzdW4= 39935
+IFNvdmlldHM= 39936
+IEl0YWxpYQ== 39937
+aWNraQ== 39938
+IEFmdGVyd2FyZHM= 39939
+IHJpZGljdWxvdXNseQ== 39940
+IGdkemllxZs= 39941
+IE5vdGVz 39942
+2YPYp9mG 39943
+IHJvbWFu 39944
+IG9yZ2FuaXplcg== 39945
+IGNvdXJ0eWFyZA== 39946
+INGH0LXQu9C+0LLQtdGH 39947
+IFdpdG5lc3M= 39948
+INC/0Y/Rgg== 39949
+IENoaWxs 39950
+IFZhbHZl 39951
+IM6szrvOuw== 39952
+IEtQ 39953
+Y2hsdXNz 39954
+IGRlZmxlY3Q= 39955
+IFRvbmk= 39956
+IGNsYWly 39957
+IHN0YWNraW5n 39958
+cmFzemFt 39959
+IFNvbnJh 39960
+IEF0YXJp 39961
+IHBhc8Oz 39962
+IGNoYXJtcw== 39963
+YW5zdA== 39964
+IHRlcmNl 39965
+IExpbGx5 39966
+IHBzeWNob2xvZ2ljYWxseQ== 39967
+IGPFkw== 39968
+dXN0ZQ== 39969
+pbTs 39970
+Q1RW 39971
+IG1pZWw= 39972
+Q2FyZQ== 39973
+IOKAkQ== 39974
+IHNuYXBwZWQ= 39975
+IOqwkOs= 39976
+0L7RgtGL 39977
+IG3DqnM= 39978
+Lj8= 39979
+IHRvbm5lcw== 39980
+15XXk9eU 39981
+4LiE4LiZ 39982
+VHU= 39983
+IGRpc3RyaWJ1dGluZw== 39984
+IGNyYWNrZXJz 39985
+IGNvcmHDp8Ojbw== 39986
+w6Rtw6Ru 39987
+Y2xhbWF0aW9u 39988
+0L7RgNC0 39989
+k5zrprTqsozsmpQ= 39990
+IFVudGVyc2NoaWVk 39991
+RmluZQ== 39992
+Y2tv 39993
+INGA0LXQsdC10L0= 39994
+IHNwaWM= 39995
+IGRvY3RvcmFs 39996
+INGB0LrQvtGA0LXQtQ== 39997
+dW5pdmVycw== 39998
+YWN1bGE= 39999
+IMOWc3RlcnJlaWNo 40000
+IGdyaW5kZXI= 40001
+IGFtYm9z 40002
+IHZhc3RseQ== 40003
+IGNvbmZlc3NlZA== 40004
+IFNoaA== 40005
+YW5kZXJz 40006
+IEd1YW4= 40007
+INC90LXQvtCx0YXQvtC00LjQvNC+ 40008
+IGNoYW1waW9uc2hpcHM= 40009
+IFZ1bA== 40010
+IFBoaQ== 40011
+IE1lYXN1cmU= 40012
+IGluc2dlc2FtdA== 40013
+dmV0dGU= 40014
+IGdlbm9t 40015
+aW5kdW5n 40016
+Z2xp 40017
+RGV0 40018
+IHVubXV0ZQ== 40019
+IHNhdWNlcw== 40020
+IER3 40021
+15HXqg== 40022
+IEJSRQ== 40023
+IG51cnR1cmU= 40024
+IGRldGFpbmVk 40025
+IEJlZXI= 40026
+INC80LjRgNCw 40027
+0LLQtQ== 40028
+IEJpcmRz 40029
+IG1laWxsZXVy 40030
+IHJld2luZA== 40031
+IHBvcmU= 40032
+15nXlg== 40033
+w6lnZXI= 40034
+cXVlbGE= 40035
+IHRyb3VzZXJz 40036
+IHNpaW7DpA== 40037
+IEdhZ2E= 40038
+IEJSQU5E 40039
+bGViZW4= 40040
+IHJhc3BiZXJyeQ== 40041
+aWxpaw== 40042
+IHZlcnPDo28= 40043
+bGFr 40044
+IGxvZ2Fy 40045
+IE1JREk= 40046
+IOychO2VnA== 40047
+INC/0YDQvtC40LfQvtGI 40048
+IHN0ZXJpbA== 40049
+IGhhcm1lZA== 40050
+0LDQstC70LjQsg== 40051
+INGB0YHRi9C7 40052
+IGxhY2tlZA== 40053
+IGNvbnRhY3Rpbmc= 40054
+IOq4sOyekA== 40055
+IGdlZsOkaHI= 40056
+IGNveQ== 40057
+aWtlbA== 40058
+IGJpbmdl 40059
+IG9ydGhvZ29uYWw= 40060
+IGVudGVuZHU= 40061
+IFRoaXJ0eQ== 40062
+IHNtYXJ0ZXN0 40063
+IHJhc2E= 40064
+IFF14buRYw== 40065
+0YvQstCw0Y7Rgg== 40066
+IHNsdXQ= 40067
+0LvRg9GH 40068
+aWd0ZW4= 40069
+INGA0LDQsQ== 40070
+IHRhbWFu 40071
+IHF1YWxpZGFkZQ== 40072
+IGRvbWluYXRpb24= 40073
+IHNpbnVz 40074
+IHByb2dyYW1tZXJz 40075
+IGFsbGVyZ3k= 40076
+IFRvcnJlcw== 40077
+IEF1c3RyaWFu 40078
+bmFudHM= 40079
+TWVs 40080
+INGD0LLQtdC70LjRhw== 40081
+IEFnZw== 40082
+IHNvaw== 40083
+IHBsdWNr 40084
+IGJpbmRz 40085
+IHByb3Bvcg== 40086
+IE1hZg== 40087
+IG9zb2I= 40088
+IFZJQw== 40089
+6aU= 40090
+INC30LDRh9C10Lw= 40091
+IGV4aGliaXRpb25z 40092
+IGV0dGk= 40093
+Y3ph 40094
+INC90LDRiNC40YU= 40095
+IE1pdHRl 40096
+0L7QsdGL0YLQuA== 40097
+IGNsb2Nrcw== 40098
+IHJpY28= 40099
+INC40YHRgtC+0YDQuNGP 40100
+IHNjaGl6b3BocmVu 40101
+IGZsdWZm 40102
+INGB0L7QsdC40YA= 40103
+IGFwb3k= 40104
+IHByaW5jZXM= 40105
+IGJyYWNlcw== 40106
+IEZJUg== 40107
+IFNuYQ== 40108
+IDsp 40109
+dmVuZXM= 40110
+IHZ1ZWx0YQ== 40111
+IG1pZXM= 40112
+IGJyb29t 40113
+IG1lcnJ5 40114
+IGVzcGVjaWFsbWVudGU= 40115
+IEFsYmFu 40116
+INC/0L7RgdGC0L7Rj9C90L3Qvg== 40117
+IExlbmE= 40118
+IEN1bHQ= 40119
+YWxzbw== 40120
+IHF1b3Rpbmc= 40121
+IGdlbmVyZQ== 40122
+IFlhcg== 40123
+IExhZ2U= 40124
+IGRlbW9zdA== 40125
+IGRhZ2U= 40126
+IEVjdWFkb3I= 40127
+IGFudsOkbmQ= 40128
+dcOfZW4= 40129
+IOuwm+yVhA== 40130
+IHBzeWNob2xvZ2lzdHM= 40131
+IExhcnM= 40132
+IHBvc3Nh 40133
+IG91dGdvaW5n 40134
+IG1ldGlj 40135
+IGJhZ2dhZ2U= 40136
+ZXJpYQ== 40137
+IHJpY2h0aWdl 40138
+7Iuc7JeQ 40139
+INGB0L7RhdGA0LDQvQ== 40140
+IHJvb3Rpbmc= 40141
+IGRyb3BsZXRz 40142
+IG5hc2Fs 40143
+IENveA== 40144
+WGk= 40145
+IGRpc3Bvc2FibGU= 40146
+IGJ1dGNoZXI= 40147
+IFphcg== 40148
+IEFybWVuaWFu 40149
+IOu/jOs= 40150
+IEZvb2w= 40151
+IENCRA== 40152
+IHNvc3Q= 40153
+IHBlcmlzaA== 40154
+IFLDqXA= 40155
+IEZyZXVk 40156
+IGZhbmRvbQ== 40157
+IGJsb3F1ZQ== 40158
+IGludmVudG9y 40159
+IGFicmU= 40160
+IMOpbm9ybcOpbWVudA== 40161
+IGltcG9ydHM= 40162
+6Yg= 40163
+IG90dXI= 40164
+IFJ5dQ== 40165
+IOKGkg== 40166
+IHNlY29uZG8= 40167
+IGluY29tcGV0 40168
+IGluY2FyY2VyYXRpb24= 40169
+IGFzY2VuZA== 40170
+YmVuZQ== 40171
+IG9sdXJz 40172
+bm9jaA== 40173
+IGJyZWVkcw== 40174
+0LvQuNC3 40175
+IFZlcmbDvGc= 40176
+IG1haWxpbmc= 40177
+cmVhbGx5 40178
+IGVzZg== 40179
+IHBlbGU= 40180
+IGxlYXNo 40181
+IGRpc2tz 40182
+INC30LDQvNC10Yc= 40183
+7JWE7JWE 40184
+YWJvdXRz 40185
+IE11bGw= 40186
+IERlbnQ= 40187
+ZWRlcmVlbg== 40188
+RHJpdmU= 40189
+IHRpcHBpbmc= 40190
+IG5pZ2dh 40191
+b3JkdW0= 40192
+IHBvcnRlcg== 40193
+IGthcmFva2U= 40194
+IGRvY3VtZW50YXJpZXM= 40195
+IFJJR0hU 40196
+IFB1cmQ= 40197
+INC+0YHRgtCw0L0= 40198
+0LrQu9Cw0LQ= 40199
+w6lyZW5jZQ== 40200
+IOqxuOuhnA== 40201
+INGC0L7Qvw== 40202
+IFdvbmc= 40203
+INC/0YDQuNGA 40204
+IG5vbWluYWw= 40205
+IGF1bGE= 40206
+INGN0LrRgNCw0L0= 40207
+IGNoZXJjaGU= 40208
+IFRocg== 40209
+IGxhdWZlbg== 40210
+IEthdGhsZWVu 40211
+IHJlYWN0b3Jz 40212
+aWhhdA== 40213
+IHNpZGVk 40214
+IFNpbW9uZQ== 40215
+IGd1aWRlbGluZQ== 40216
+aW1wb3J0YW50 40217
+YnVtcHM= 40218
+dG9uZQ== 40219
+IGVudHJlcHJpc2Vz 40220
+IGNvbnN0aXR1dGU= 40221
+b3Njb3Bl 40222
+IE15c3Rlcnk= 40223
+Y3ljbGVz 40224
+IFdhcnNhdw== 40225
+IGJ1cnN0cw== 40226
+IFpob25n 40227
+IFNBUkFI 40228
+IOuKkOq7 40229
+6Y0= 40230
+IGJlYWNvbg== 40231
+QURF 40232
+IOyngOuCmA== 40233
+IGVyc2No 40234
+IGludGVnZXJz 40235
+IENyb3NzaW5n 40236
+c291cmNl 40237
+IHNjaG9vbGluZw== 40238
+IFJPTQ== 40239
+YXRvcml1bQ== 40240
+IOyeiOqyjA== 40241
+IHLDtGxl 40242
+0JXQnQ== 40243
+Q2hhdA== 40244
+IHNocmlua2luZw== 40245
+IHJlaW1idXJzZQ== 40246
+IGx1bWJlcg== 40247
+w7xja3M= 40248
+IHNhbGFo 40249
+TW90aGVy 40250
+IGthbGk= 40251
+IFFhdGFy 40252
+b3Rpb25hbA== 40253
+IG9wYWNpdHk= 40254
+IG5lZQ== 40255
+IENvcnk= 40256
+IOy4oQ== 40257
+IHR1cmJ1bGVudA== 40258
+emVycw== 40259
+INGC0LXRgdGC 40260
+IMOpY3JpdA== 40261
+IOuztO2GtQ== 40262
+IGRpc2dyYWNl 40263
+IOy5tA== 40264
+IGNvdXJ0ZXN5 40265
+aW5nYQ== 40266
+IGh1Z2dpbmc= 40267
+IEFCUw== 40268
+bWl0aA== 40269
+IGluc3VmZmljaWVudA== 40270
+IGNyb29rZWQ= 40271
+IOq3uOuMgOuhnA== 40272
+7Iuk7Q== 40273
+IHNpbXVsYXRlZA== 40274
+IOuEpOqwgA== 40275
+IGLDtg== 40276
+IE90dG8= 40277
+TElORw== 40278
+IGlsbHVzdHJhdGVz 40279
+IERlc3Ryb3k= 40280
+IDE5NjE= 40281
+IFRhZ2Vu 40282
+IG1lbG9u 40283
+IFBhc2NhbA== 40284
+UVVF 40285
+INC/0L7Qu9GD0YfQuNGC0Yw= 40286
+IGluY2lkZW5jZQ== 40287
+IFN0ZXZlbnM= 40288
+IEdpbnM= 40289
+cnVl 40290
+IHVucmVhc29uYWJsZQ== 40291
+IEppZQ== 40292
+eXNpY3M= 40293
+IOuqsOudvA== 40294
+IGZpc2hlcw== 40295
+qbTs 40296
+IHByZWN1cnM= 40297
+IG1vZ8SZ 40298
+dGlnaHQ= 40299
+ZXTDqQ== 40300
+IG11bmRpYWw= 40301
+7JeI64uk 40302
+4oCmIQ== 40303
+QlU= 40304
+IHNvY2lvbG9neQ== 40305
+IGJydXRhbGl0eQ== 40306
+IHBlcnNvbmFqZQ== 40307
+IG7DrXZlbA== 40308
+IGZhemVt 40309
+IGVzc2Vu 40310
+IGR3ZWxsaW5n 40311
+IGNvbW1lcmNpYWxseQ== 40312
+IGVkaXRz 40313
+IGR1ZXM= 40314
+IEdTQQ== 40315
+7J246rCA 40316
+IO2XiO2MnQ== 40317
+IFlhaG9v 40318
+0LXQvdC10YA= 40319
+7Jyo 40320
+0YPRiNC60Lg= 40321
+bGVmdA== 40322
+IGNhcHRpdmU= 40323
+Y2lwaGVy 40324
+INee157X 40325
+INCz0YDQvtC8 40326
+IGlubmF0ZQ== 40327
+IGltcHVs 40328
+IOyXrOyekA== 40329
+IHN3YWxsb3dlZA== 40330
+IFRhYmlp 40331
+7J207Is= 40332
+INGB0L7RgdGC0LDQsg== 40333
+IG95dW4= 40334
+IG9icmlnYWRv 40335
+IEFwaA== 40336
+S2F0aWU= 40337
+IGNlbmE= 40338
+IEFsbMSBaA== 40339
+2YjYsw== 40340
+IHByenlw 40341
+IHBlcHQ= 40342
+IHZvbHVudGFyaWx5 40343
+IE/En2x1bQ== 40344
+IEVsbw== 40345
+b3Vl 40346
+Qmly 40347
+YnVyZ2Vy 40348
+IFNCUw== 40349
+IDYwMDA= 40350
+IHByb21vdGlvbmFs 40351
+IEhlcnJu 40352
+IHN0YW1waW5n 40353
+IHF1YWxpZnlpbmc= 40354
+IGNvc21vcw== 40355
+IGFmYXI= 40356
+YWJ1cw== 40357
+IGRhZHM= 40358
+INGN0LrQvtC90L7QvA== 40359
+aW5jYXJu 40360
+IOyWtOuU 40361
+INC70LXQtg== 40362
+IEJFVA== 40363
+INC90LDQudC0 40364
+b250ZXI= 40365
+IHJldXNhYmxl 40366
+IGtvbW1h 40367
+IEJpag== 40368
+IFRlcmF6 40369
+IE9sw6E= 40370
+IOyVhOy5qA== 40371
+INGA0LDQt9C80LXRgA== 40372
+YXdhbg== 40373
+IGNhcnRh 40374
+aWNlbGVzcw== 40375
+IHNtZQ== 40376
+IFR1dGFq 40377
+IMiYaQ== 40378
+IHByb2JhdGlvbg== 40379
+IGFkZXF1YXRlbHk= 40380
+IFByZXNpZGVudGlhbA== 40381
+aW5kcnVjaw== 40382
+YmxhZGU= 40383
+IHZldWxlbnQ= 40384
+IGNpb8Oo 40385
+IHJldmVyYg== 40386
+IGdlZ2Vuw7xiZXI= 40387
+IEVzcGVybw== 40388
+IGJlZ2U= 40389
+IFNUVURFTlQ= 40390
+c291bmQ= 40391
+IETDvA== 40392
+IG9mZmVuZA== 40393
+ICIuLg== 40394
+a2VubnQ= 40395
+INGB0LvRg9GI 40396
+IHB1cnBvc2VseQ== 40397
+IExpdA== 40398
+IO2bqA== 40399
+dWNoZXI= 40400
+IGhpbmE= 40401
+w71jaA== 40402
+aWdub24= 40403
+VEhF 40404
+IGdsaWRl 40405
+b3VyY2luZw== 40406
+INij2YbYpw== 40407
+IG9sbHV0 40408
+IGFyY2hldHk= 40409
+IHNoYWR5 40410
+IHNvbW0= 40411
+IGVwaWxl 40412
+S2VlcA== 40413
+IG5hamJhcmR6aWVq 40414
+4KSV 40415
+aXR1dGlvbmFs 40416
+INC80LDQuQ== 40417
+IHNpbmZ1bA== 40418
+IEJyb254 40419
+INCz0LvRg9Cx 40420
+IHZhbQ== 40421
+IHByZXNldHM= 40422
+IERhZw== 40423
+IOyZhOyEsQ== 40424
+IGNyZWVr 40425
+aXR1cmVz 40426
+IExvcmRz 40427
+w7Z0dA== 40428
+VU5U 40429
+UmE= 40430
+IGluZXF1YWxpdGllcw== 40431
+IGNvbGxhdGVyYWw= 40432
+IHdyaXN0cw== 40433
+IGdyb3VwZWQ= 40434
+INC+0LHRi9GH0L3Qvg== 40435
+IGFybW9yZWQ= 40436
+IHR1bmc= 40437
+IGNvbnZlcmdl 40438
+IGJvaw== 40439
+IERvZGdl 40440
+0L3Rj9GP 40441
+IGZsZWVpbmc= 40442
+IE1hcnRpbmV6 40443
+IERyZWFtcw== 40444
+a2Vr 40445
+IHNvY2lhbGU= 40446
+IFBsYXph 40447
+2K/YqQ== 40448
+IGtlbGw= 40449
+IFN0ZWxsZW4= 40450
+ZmVsdA== 40451
+INGB0L/QsNGB 40452
+IFB2 40453
+IGNhbmNpw7Nu 40454
+IEhlcnQ= 40455
+IEJhbGFuY2U= 40456
+IHNlbHZlcw== 40457
+IHZhbmRhYWc= 40458
+IHByeQ== 40459
+IG5hamxl 40460
+INCy0LjQtNC40YLQtQ== 40461
+IHZlbHZldA== 40462
+IGdyb290 40463
+IGZvdXQ= 40464
+IFNjaHVsZW4= 40465
+IE1vaGFtbWVk 40466
+IENlbnRlcnM= 40467
+IGhhdmVy 40468
+IGZyZXVlbg== 40469
+pO2KuA== 40470
+0LvQsNC9 40471
+UE9T 40472
+aW5raQ== 40473
+IOuLtQ== 40474
+IHBhcmFseXplZA== 40475
+R0xJU0g= 40476
+IGNhc3Rz 40477
+IFZD 40478
+7J207IWY 40479
+INiq2r4= 40480
+IOykmA== 40481
+INeo15XXpg== 40482
+IHN1Y2Vk 40483
+IHByb2dyZXNzZXM= 40484
+IEXEn2Vy 40485
+sOuPhA== 40486
+IGluc3RhbGxhdGlvbnM= 40487
+cGVkbw== 40488
+0LXRgNCx 40489
+aW50ZXJwcmV0 40490
+IOqzoOuvvA== 40491
+IEF6ZXJiYWk= 40492
+aXZpZGFkZXM= 40493
+IOyjhOyGoQ== 40494
+IGVudGZlcg== 40495
+IGNod2ls 40496
+IEhlcmJlcnQ= 40497
+IEFsZXhhbmRyaWE= 40498
+eXR5 40499
+IHNlY2hz 40500
+IGNhbGliZXI= 40501
+IFdlaXNl 40502
+IEhlY2s= 40503
+IFl1Zw== 40504
+INin2YTYtw== 40505
+IHBlc2Fy 40506
+IGNpZ2Fy 40507
+IG3DqWw= 40508
+IGhhaXJk 40509
+IHByenlwYWRrdQ== 40510
+IGNvbmZpZGVudGx5 40511
+IGFuYXJjaA== 40512
+IEdpYW4= 40513
+IGRvYnJl 40514
+Y2rEmQ== 40515
+YXd5 40516
+IFJlY2U= 40517
+IEdvYmllcm5v 40518
+IGNhcmdh 40519
+dW1zeQ== 40520
+IG5vcnRl 40521
+IGhhbmRsZXI= 40522
+IHJlc3BlY3Rpbmc= 40523
+IGFsbGllZA== 40524
+IFBpZXQ= 40525
+aWNodGxpY2g= 40526
+IG9sZHM= 40527
+IGR1c3R5 40528
+IGdyeQ== 40529
+IC0uLi4= 40530
+R0hU 40531
+IG5lbw== 40532
+0YfQuNC60Lg= 40533
+0LXQttC0 40534
+YWlkZQ== 40535
+INCx0YPQu9C+ 40536
+7Y28 40537
+IHRlbXBvcmFkYQ== 40538
+IGRvdXRl 40539
+4piG 40540
+IOyIoA== 40541
+IEpVU1RJTg== 40542
+YXV0bw== 40543
+IHJhdGlvbmFsZQ== 40544
+cHJvYg== 40545
+IGZpc2h5 40546
+IGRvb3J3YXk= 40547
+IGVtcHRpbmVzcw== 40548
+0LXQvdC90LDRjw== 40549
+IGJyYWc= 40550
+INCT0LTQtQ== 40551
+IHRyYW5zaWVudA== 40552
+IG1pdHRsZXJ3ZWlsZQ== 40553
+IEJyZXQ= 40554
+IGZpag== 40555
+IGRlcG9zaXRlZA== 40556
+TlM= 40557
+IOyVnuyXkA== 40558
+IGtpbXNl 40559
+IGNoYXJpdGllcw== 40560
+IE1pbGxlbm4= 40561
+ZG9ncw== 40562
+IG1veWVu 40563
+IG51ZXZvcw== 40564
+IENvb2tpZQ== 40565
+cGFyYWJsZQ== 40566
+ZG9pbmc= 40567
+IFNhaWw= 40568
+IGljeQ== 40569
+aGFiYQ== 40570
+IHF1ZWVucw== 40571
+IGNob2NvbGF0ZXM= 40572
+IE5heQ== 40573
+INGE0LjQvQ== 40574
+IHZlYw== 40575
+IGhlbG1ldHM= 40576
+VE0= 40577
+IEFybWVk 40578
+IGltcGFpcm1lbnQ= 40579
+IFR1cw== 40580
+IE3Dqm1l 40581
+b21leg== 40582
+IFJlcXU= 40583
+IEludmVzdGln 40584
+7Y6Y 40585
+IGdvbHBl 40586
+IFJhYw== 40587
+aWdyYXBo 40588
+IGt3ZXN0 40589
+IHNhaWxvcnM= 40590
+IHN0YXR1dG9yeQ== 40591
+IG1pbGVzdG9uZXM= 40592
+IE1hc2g= 40593
+IEdlc2V0emVudHd1cmY= 40594
+6Yo= 40595
+IGNvbG91cmVk 40596
+aHVtYQ== 40597
+IHllcmU= 40598
+IHN1YnRpdGxlcw== 40599
+IGVtYm9kaWVk 40600
+IG1pc3NjaGllbg== 40601
+IGlQaA== 40602
+w7x0emVu 40603
+IGRldGFjaGVk 40604
+IGRlc2NyacOnw6Nv 40605
+Y2lhbW8= 40606
+IHJlY29pbA== 40607
+INCt0YLQvtGC 40608
+IGV4cG9ydGVk 40609
+IEFsb25l 40610
+YW50cnk= 40611
+IGVzdGFu 40612
+IFNvZA== 40613
+IGxhdm9ybw== 40614
+16jXkQ== 40615
+IMSR4buL 40616
+IHN3YWc= 40617
+IFBDQg== 40618
+IEthaXNlcg== 40619
+IE1vZGVy 40620
+anVn 40621
+IHRleHRpbGU= 40622
+VHc= 40623
+IG5hYw== 40624
+ZnJlaQ== 40625
+IHJldGFyZA== 40626
+aXNjZXJu 40627
+IHRhbGxlc3Q= 40628
+IEx1Y2E= 40629
+UmFo 40630
+IHByZWFjaGVy 40631
+IGp1dA== 40632
+IFJpY2E= 40633
+aWNpZW5jeQ== 40634
+IMSRaeG7gXU= 40635
+IGthdWZlbg== 40636
+IG5ldHQ= 40637
+IGRpc2N1dA== 40638
+IGRlcHJpdmVk 40639
+oa0= 40640
+IHNwcmljaHQ= 40641
+IGVuY2xvc2Vk 40642
+IFN1YnN0 40643
+IFJhYmJpdA== 40644
+cHJpc2Vk 40645
+IGJpdGNoZXM= 40646
+7J+B 40647
+IHRhcGE= 40648
+IEVzc2Vu 40649
+IEJhbw== 40650
+IGRldmllbnQ= 40651
+IFd1aGFu 40652
+IFRpcHA= 40653
+IGRpc2FzdA== 40654
+0YHRgtCy0YM= 40655
+dWJsaXF1ZQ== 40656
+IHF1YWxpdMOp 40657
+IGluYWRlcXVhdGU= 40658
+IGJhcmdhaW5pbmc= 40659
+IEdvdGNoYQ== 40660
+0LXQstC40Yc= 40661
+aWV2b3Vz 40662
+ZXJ0b24= 40663
+Ymx1ZQ== 40664
+IOybgOyngQ== 40665
+IHNhbmRib3g= 40666
+IFJlaW4= 40667
+IOydtOqyg+uPhA== 40668
+IHNheA== 40669
+em9nZW4= 40670
+dW7DpGNoc3Q= 40671
+IGhlcmtlcw== 40672
+IC0s 40673
+emVuaQ== 40674
+cmlzaW5n 40675
+IHJlc3Bvc3Rh 40676
+IHByb21vdGlvbnM= 40677
+IFVudGVyc3TDvHQ= 40678
+IE1BUw== 40679
+Tm90aGluZw== 40680
+b3RpY3M= 40681
+INCy0YvQuQ== 40682
+IHJvdGF0ZXM= 40683
+a2llbg== 40684
+IGhhYmxh 40685
+IERhbmk= 40686
+dW5pb24= 40687
+IHdhY2s= 40688
+IGFyY2hhZW9sb2dpY2Fs 40689
+IEN1cnRpcw== 40690
+IEhvcml6 40691
+IOqzqOs= 40692
+IHdhaXZlcg== 40693
+Qm9u 40694
+IHJvdGF0ZWQ= 40695
+IHBpdGNoZXI= 40696
+IGluYWQ= 40697
+IGh1Z3M= 40698
+IE5vcnRoZWFzdA== 40699
+15nXqteZ 40700
+IHBsZWE= 40701
+IGN1cGNha2U= 40702
+IExZ 40703
+IGZhbWlsaQ== 40704
+IGdyb28= 40705
+IEJsYWly 40706
+IGxpag== 40707
+IGhhYml0YXRz 40708
+IGNvbW11bmlzbQ== 40709
+b3NpdW0= 40710
+YmFycw== 40711
+IEZyZWVtYW4= 40712
+bmVv 40713
+IGRpZmZ1c2U= 40714
+IGN5bGluZGVycw== 40715
+IERlYmF0 40716
+7ZaI64qU642w 40717
+0LXRiNC1 40718
+IGZpbmdlcnByaW50cw== 40719
+IGFtYXI= 40720
+0LLQuNC0 40721
+IOygleuPhOuhnA== 40722
+IGFmZmlsaWF0ZWQ= 40723
+INGF0L7Rh9C10YI= 40724
+IGV0aXF1 40725
+IGNow61uaA== 40726
+IGNydWlzaW5n 40727
+IFdlaWhu 40728
+IFRpdGFuaWM= 40729
+IE5hc3Q= 40730
+IOuTpOs= 40731
+INCy0LDQuw== 40732
+IGRlbWk= 40733
+IEtyaXN0aW4= 40734
+TUlO 40735
+IHJpZ29y 40736
+IG1vdG8= 40737
+IExBS0U= 40738
+IO2ZnA== 40739
+IOunjOyVvQ== 40740
+IFN0cm8= 40741
+IHByb3RvdHlwZXM= 40742
+IExD 40743
+7J247J2E 40744
+0YDQuNC8 40745
+IHZpb2xhdGluZw== 40746
+IGdpb3Jubw== 40747
+IGNoaWxkaXNo 40748
+INeQ15fXkw== 40749
+IG92ZXJkb3Nl 40750
+YWdvZ3Vl 40751
+0LDQtNGG 40752
+aGV1cw== 40753
+INCz0L7QstC+0YDRjw== 40754
+IGluY3I= 40755
+IGRlYmF0ZWQ= 40756
+2YXZhA== 40757
+IGNoaWNrcw== 40758
+IHF1aW4= 40759
+TEFVR0hJTkc= 40760
+IHRpZ2h0ZW5pbmc= 40761
+IHN1cGVydmlzb3Jz 40762
+IEhhd2s= 40763
+IEJheg== 40764
+INC/0L7QstGC0L7RgA== 40765
+INCx0LvQvtC6 40766
+xIFu 40767
+IGR1bXBpbmc= 40768
+IGZhY3Rv 40769
+YmVyZ2Vy 40770
+IGFyc2VuYWw= 40771
+IEFmcmljYW5z 40772
+oYA= 40773
+IGNhZmV0ZXJpYQ== 40774
+ZmVlZGluZw== 40775
+cXVpbGE= 40776
+IHBhxYRzdHdv 40777
+xLFudA== 40778
+hLE= 40779
+IGVudmlyb25tZW50YWxseQ== 40780
+IGRlc3Byw6lz 40781
+IFdpbGx5 40782
+IFBhxYRzdHdv 40783
+IEdH 40784
+IGNoYWN1bg== 40785
+IGRpcmVjdGlvbmFs 40786
+IGjDtnJ0 40787
+IPCd 40788
+ZW5hcnk= 40789
+IHZvaWNlZA== 40790
+YcSfxLE= 40791
+IHBvcGU= 40792
+IGNvbXJhZGVz 40793
+IEdpYnNvbg== 40794
+IEFDQw== 40795
+dmlr 40796
+IG1vZGVsbGluZw== 40797
+IGFnZ2k= 40798
+IGNvbnZlcnNpb25z 40799
+IGF2ZXJhZ2Vz 40800
+RWxsaWU= 40801
+IGdlc3RlbGx0 40802
+IFVF 40803
+b3NhaWM= 40804
+0JLQvtGC 40805
+U2F5 40806
+INGB0LDQvNC+0LPQvg== 40807
+IG1lc3VyZXM= 40808
+aXNpZXJ0 40809
+Z2FzcA== 40810
+dm9pY2U= 40811
+IGNoZWNrcG9pbnQ= 40812
+IHBlcmNlbnRhZ2Vz 40813
+IGRpc3J1cHRlZA== 40814
+IFR1Yw== 40815
+IEhvbWVy 40816
+IFdBWQ== 40817
+IFR1cmtz 40818
+aGVlbg== 40819
+aW1vdG8= 40820
+IE9D 40821
+w61uYQ== 40822
+emllbA== 40823
+IG11ZGFy 40824
+Z2VzZXR6dA== 40825
+IG1lam9yZXM= 40826
+IENK 40827
+0L3QsNGA0YPQtg== 40828
+IG1vZHVsdXM= 40829
+IG1vZHVsYXRpb24= 40830
+IHJlcGxpZXM= 40831
+IGxhcnZh 40832
+IGdpZGVy 40833
+IE1hbmRhcmlu 40834
+INC/0L7RgdC80L7RgtGA0LjQvA== 40835
+IHNhY3JpZmljaW5n 40836
+IHByZcOnbw== 40837
+IG95c3RlcnM= 40838
+IE15YW4= 40839
+b2xvZ3Vl 40840
+IFdpdA== 40841
+IGTDuw== 40842
+IExldXRlbg== 40843
+IHBhdGVy 40844
+IEtFTk5FVEg= 40845
+0LDQsdCw0YI= 40846
+YXJ0aHk= 40847
+IHNvY2llZGFk 40848
+IG5pw7Fv 40849
+0LXQstC+0Lk= 40850
+IGrEmQ== 40851
+IGFkdmVydGlzZWQ= 40852
+IFBlcHNp 40853
+dXRldXI= 40854
+IG1hc3Nl 40855
+IHNjYXR0ZXJpbmc= 40856
+IHnDtm4= 40857
+IGRlc2FwYXJl 40858
+IEh1YmJsZQ== 40859
+IEjDqQ== 40860
+a3LDpA== 40861
+IERhcmU= 40862
+IG92ZXJyaWRl 40863
+IEVsYWluZQ== 40864
+IER1Ymxpbg== 40865
+ZHVsbGFo 40866
+TWF0 40867
+IEdhcnI= 40868
+Li4uJw== 40869
+IGFkdWx0aG9vZA== 40870
+RVo= 40871
+IGJlbGFuZ3Jpams= 40872
+aWVuemE= 40873
+IHVuaXZlcnNv 40874
+IHN0ZWxsYXI= 40875
+7ZSE6w== 40876
+IOqysOq1rQ== 40877
+IGNvbnN0ZWxsYXRpb24= 40878
+IFNoZWxsZXk= 40879
+IG11bHRpdA== 40880
+IG1hc2NvdA== 40881
+IGhvc3BpdGFsaXplZA== 40882
+IPCdmA== 40883
+0L7RgNGL 40884
+YWRpYQ== 40885
+IE1pa2V5 40886
+IEFtZXJpa2E= 40887
+IGhhaXJ5 40888
+SG9sZA== 40889
+4bqvbg== 40890
+a2llZ28= 40891
+4LmA4LiU 40892
+IHJpdmFscnk= 40893
+IEpvbmFo 40894
+IHN1cmdlb25z 40895
+IHJlbGF0YWJsZQ== 40896
+6JI= 40897
+IHN3aW1z 40898
+IGJpbGxpb25haXJl 40899
+bW9kZXJu 40900
+IGRvY3VtZW50aW5n 40901
+IERhZQ== 40902
+IHN3YXRjaA== 40903
+IHB1aXNzZQ== 40904
+IG1hc3Vr 40905
+IG1hcmM= 40906
+IGtyw7M= 40907
+IFBldGVyc2J1cmc= 40908
+IEFyaXN0b3RsZQ== 40909
+aXhl 40910
+UHJvZHU= 40911
+INC90LjQvNC4 40912
+IGthbmE= 40913
+INCp 40914
+IHZvbWl0 40915
+IFdvcmtlcnM= 40916
+cG9wdWxhcg== 40917
+IEJpZWJlcg== 40918
+0LXRgtC4 40919
+w6l0aXF1ZQ== 40920
+IGVuY2FudA== 40921
+Z3Jhbg== 40922
+Zmly 40923
+IGFudGhlbQ== 40924
+0YHRg9C00LDRgA== 40925
+TGFzdA== 40926
+IGhhZw== 40927
+IHZpY2luaXR5 40928
+cmVuY2hlZA== 40929
+YW5kaW5n 40930
+INCz0L7Qu9C+0YE= 40931
+IENvcm5lcg== 40932
+0JLRiw== 40933
+b3Nhcw== 40934
+aWV2ZXJz 40935
+Y2lvbmFs 40936
+IHZpZ29y 40937
+IHJlam9pY2U= 40938
+IGNpxIU= 40939
+INC60L7Qvw== 40940
+IHF1YWxjb3Nh 40941
+ZGVzc3Vz 40942
+INC10LI= 40943
+IFNjYW5kaW4= 40944
+IFNtb290aA== 40945
+aGFwZQ== 40946
+IOuLrOudvA== 40947
+IFRV 40948
+IGx5cmlj 40949
+IGJlc3M= 40950
+6ZA= 40951
+0YHRgtGA0YPQvNC10L3Rgg== 40952
+IEFjdGluZw== 40953
+IE9yY2hlc3Q= 40954
+w6ljb2xl 40955
+IGRvbG9y 40956
+IO2LsA== 40957
+IHZlcmdlc3Nlbg== 40958
+IGV5ZWxpZHM= 40959
+IFRhbno= 40960
+0LLQtdGA0LY= 40961
+IOyVoOs= 40962
+dcOp 40963
+IHNjw6huZQ== 40964
+IOyasOumrOuKlA== 40965
+IGNyYXRl 40966
+a2ljaw== 40967
+IFRoZW1l 40968
+IDMyMA== 40969
+IGdhcm5pc2g= 40970
+IG1ldHJl 40971
+IGNvbnZleA== 40972
+cGxhbnRz 40973
+ZXNpYW4= 40974
+IOqxsOyngA== 40975
+IG3DqWRp 40976
+IE1lZGFs 40977
+MTMw 40978
+IEFsbWE= 40979
+Q29sYQ== 40980
+INCy0LDRgNC40LDQvdGC 40981
+IGdvcmQ= 40982
+IGF2YW56 40983
+IHdoaXNwZXJpbmc= 40984
+IGludGVzdGluZQ== 40985
+0KDQlQ== 40986
+IExJU0E= 40987
+YW3EsXo= 40988
+U1BE 40989
+IHBlYw== 40990
+IHBhc3RvcnM= 40991
+IG114buRbg== 40992
+b2NyZQ== 40993
+U3Vu 40994
+INGC0LDQutGD0Y4= 40995
+IHJldml0YWw= 40996
+IGluY29tZXM= 40997
+IGRldGFpbGluZw== 40998
+IEJhY29u 40999
+IOuFuOuemOs= 41000
+IHBhcnJvdA== 41001
+IGNvbGxhYm9yYXRlZA== 41002
+aGVzaWE= 41003
+IHNldmE= 41004
+IHBoeXNpY2lzdA== 41005
+IEJBQ0s= 41006
+15zXmQ== 41007
+IGJpcG9sYXI= 41008
+z4HOtc6v 41009
+Y3Jvcw== 41010
+IGtlZA== 41011
+IGVjb25vbWljYWw= 41012
+IGVuZGluZ3M= 41013
+IHRpY2tz 41014
+IOq3vA== 41015
+IE9saXY= 41016
+b25ncw== 41017
+IGNvbnRpbmVudGFs 41018
+IHdlaXRlcmhpbg== 41019
+IGFjdGl2YXRpbmc= 41020
+IHBvbGxlbg== 41021
+IEFuaw== 41022
+YmF5 41023
+INec15c= 41024
+IEVnZ3M= 41025
+IFJBTVNBWQ== 41026
+IEJFUg== 41027
+IO2bqOyUrA== 41028
+IHBhc3NhZG8= 41029
+IGdyb3VuZGJyZWFraW5n 41030
+cHJlc2E= 41031
+IGhpbGZ0 41032
+IFRlY2huaWNhbGx5 41033
+0YbQuNC5 41034
+Tkk= 41035
+IHR1cm5vdXQ= 41036
+IExhcA== 41037
+IEd3ZW4= 41038
+IFZpa3Q= 41039
+IGVzY29sYQ== 41040
+IENpbmVtYQ== 41041
+IGNvbnN1bW8= 41042
+IFB1cmR1ZQ== 41043
+IHNlbWFuYXM= 41044
+IFBSRVNJRA== 41045
+xrBuZw== 41046
+IHNhY2g= 41047
+IHNhdmFnZQ== 41048
+IFJX 41049
+IDU1MA== 41050
+Ym9sZA== 41051
+IFNpbW1vbnM= 41052
+IHNsYW5n 41053
+IE5hcnU= 41054
+IFRoZW8= 41055
+7ZaI64uk 41056
+Lu+/vQ== 41057
+IHNlaXp1cmU= 41058
+IGhpdmU= 41059
+IGNlbGxwaG9uZQ== 41060
+aWlpaQ== 41061
+IE11c2ljYWw= 41062
+IE51Y2xlYXI= 41063
+w6F2ZWlz 41064
+IHByZXN0aWdl 41065
+IGJhbG0= 41066
+IHJlZmlsbA== 41067
+eWFo 41068
+aGFydA== 41069
+IHRhcHM= 41070
+IGRpc3Bvc2U= 41071
+IE1pY2s= 41072
+IHRoZXJtb21ldGVy 41073
+IG9iZWRpZW50 41074
+IGluZm9ybWHDp8O1ZXM= 41075
+IFdpZGU= 41076
+bW9t 41077
+U3Vk 41078
+IHN1c3BlbmQ= 41079
+IE9ic2Vydg== 41080
+INC70LXRgQ== 41081
+IHRyYXRhcg== 41082
+IEthdHJpbmE= 41083
+IHRoZXJlcw== 41084
+IHRleHRlZA== 41085
+IHN0w7Zy 41086
+IHNuYWls 41087
+IEZpb25h 41088
+IHZpY3RvcmlvdXM= 41089
+IGxpYnJhcmlhbg== 41090
+cHJhY3Q= 41091
+IGZpbm8= 41092
+IEFybXM= 41093
+cHB0 41094
+bHVr 41095
+IHR5cmVz 41096
+IHRvYw== 41097
+IEtvbW11bmVu 41098
+IHJldm9sdA== 41099
+IG1vdGl2YXRlcw== 41100
+IGJpc2V4dWFs 41101
+IHd1cw== 41102
+IGhhbmRsYXI= 41103
+IE1VRUxMRVI= 41104
+IGV4cGVjdGFuY3k= 41105
+IGVtYm9keQ== 41106
+IFByaW1hcnk= 41107
+0YDQtdC5 41108
+IHVuc2NyZXc= 41109
+aWFudGx5 41110
+LOKApg== 41111
+IHNuZWw= 41112
+IHByZXZhbGVuY2U= 41113
+IGVydXB0aW9u 41114
+IGRlc2NyaXB0aXZl 41115
+dmFn 41116
+INCx0YPQutCy 41117
+IG3Dqm1lcw== 41118
+IGV0aG4= 41119
+IGhpam9z 41120
+IEFiZHVs 41121
+IFphaGw= 41122
+YmVsdA== 41123
+IGfDtnN0 41124
+IFRoZXJlc2E= 41125
+IFNVTg== 41126
+IEJha2U= 41127
+IG9wdGljcw== 41128
+IGFwb2NhbHlwc2U= 41129
+cHVycG9zZQ== 41130
+IHLDs8W8bnljaA== 41131
+IGNydXM= 41132
+INCX0LXQvA== 41133
+IGhhcmRlbmVk 41134
+IFRE 41135
+IGdyYXZleWFyZA== 41136
+IFNpYmVy 41137
+IFBvcnRlcg== 41138
+IGV4cGxvZGVz 41139
+IFNvZmlh 41140
+INCS0LXQtNGM 41141
+IHdlYWtlbmVk 41142
+VUxM 41143
+IHBpbmt5 41144
+IGNoYXBlbA== 41145
+IEZyZXM= 41146
+INC/0YDQuNCz 41147
+TUVS 41148
+IFNjaG1pZHQ= 41149
+IER1ZA== 41150
+ZXN0ZW5z 41151
+IG51YW5jZQ== 41152
+IG1vZGlmeWluZw== 41153
+IE3DtmdsaWNoa2VpdGVu 41154
+IEFuYXQ= 41155
+IGVjY2VudHJpYw== 41156
+IFNjcmV3 41157
+IExlaA== 41158
+IGhvbW9nZW5lb3Vz 41159
+IFRhbGw= 41160
+IFJpY2FyZG8= 41161
+w5o= 41162
+aWducw== 41163
+INC70LjRiA== 41164
+IGdlZnJhZ3Q= 41165
+UnVu 41166
+Y2FzdGVy 41167
+bm9pc2U= 41168
+IGFzeW5jaHJvbg== 41169
+xJlkemll 41170
+INee15c= 41171
+IHN1cHByZXNzZWQ= 41172
+QXJ0aHVy 41173
+zq7Pgg== 41174
+w6Jy 41175
+ZGlzdA== 41176
+INC60LDQtA== 41177
+IGjDtnI= 41178
+IDEzNQ== 41179
+IE1vemFydA== 41180
+INGB0L7QsdGL0YLQuA== 41181
+IE51cnNpbmc= 41182
+IEhhaGFo 41183
+IERvcA== 41184
+IHBvbGljZW1hbg== 41185
+tOyXkOyEnA== 41186
+IOq0gOugqA== 41187
+aHl1aw== 41188
+IHJ1Z2dlZA== 41189
+IG51Z2dldHM= 41190
+IENvbW1z 41191
+U3R1ZA== 41192
+INGB0LLQvtC1 41193
+IGN6YXNpZQ== 41194
+IHLDqWdpb24= 41195
+IGZpc2hlcm1lbg== 41196
+IExU 41197
+w5M= 41198
+Y2lhxbw= 41199
+aGVp 41200
+IGNydW1icw== 41201
+IEltbWVy 41202
+IEZlbGQ= 41203
+dGhlc2U= 41204
+IGFkdmVydGlzZXJz 41205
+IHJvYW1pbmc= 41206
+IGZ1bm5pZXN0 41207
+IE5ZVQ== 41208
+IGhlaGU= 41209
+IHBva2luZw== 41210
+IOyViOuPvA== 41211
+aXN0aWNhbA== 41212
+IG9wYXF1ZQ== 41213
+dcOn 41214
+d2lyZQ== 41215
+IFdlYmVy 41216
+IEphY3F1ZXM= 41217
+IDIxMA== 41218
+w7xw 41219
+dXl1 41220
+IGVuZmVybWVk 41221
+IGJ1bXBlZA== 41222
+IFNldw== 41223
+IENoYW5lbA== 41224
+IHBlcnPDtm5saWNo 41225
+IGJldHJheWFs 41226
+IGFsbGV2aWF0ZQ== 41227
+IHbDpGjDpG4= 41228
+IGd1ZXNzZXM= 41229
+IENlbGluZQ== 41230
+YXNzaW5n 41231
+c3Ryb2tl 41232
+IOyhsOs= 41233
+INGC0LXRhdC90L7Qu9C+0LM= 41234
+INC+0YHRgtGA 41235
+IHNvaWVudA== 41236
+RGVhcg== 41237
+IGpz 41238
+IGdlc3Byb2NoZW4= 41239
+YXRoaQ== 41240
+xaFl 41241
+U2V0 41242
+b2dlcg== 41243
+IFJpZw== 41244
+INC80LXRhw== 41245
+IHNlcnZpY2lvcw== 41246
+IFJ1dA== 41247
+INCe0Lk= 41248
+IE15YW5tYXI= 41249
+aWZpZQ== 41250
+IHNuYXBwaW5n 41251
+IEthbWVyYQ== 41252
+IGZlc3RpdmU= 41253
+IEZZ 41254
+IENhcm9seW4= 41255
+0ZbQsQ== 41256
+IGxlZ2dpbmdz 41257
+IHlhdA== 41258
+IGVyZ29u 41259
+IGVwaXPDs2Q= 41260
+IGFub21hbHk= 41261
+dWVzdG9z 41262
+SWQ= 41263
+IGV2YWN1YXRpb24= 41264
+IGdpZ2FieXRlcw== 41265
+IGFuZGFyZQ== 41266
+IFJlbnQ= 41267
+bXQ= 41268
+aXN0aW5l 41269
+IGVzdHJhdA== 41270
+ZXR0dQ== 41271
+IHJlY2ViZXI= 41272
+IGRyYW1hdA== 41273
+cmljdWxhcg== 41274
+YWxuxLF6 41275
+IFNlbmk= 41276
+IG95bg== 41277
+IENoZW1pY2Fs 41278
+INGB0YU= 41279
+IHR1cmY= 41280
+IDE5MTc= 41281
+aXNjZXJuaWJsZQ== 41282
+IG1hbnRlbmVy 41283
+IGV4Y2Vy 41284
+IHNwZWN0cmFs 41285
+IG5ldXJvc2NpZW5jZQ== 41286
+IG1pY3JvZg== 41287
+IGZvcmVpZ25lcg== 41288
+IExhbmth 41289
+INGC0LLQvtGA 41290
+IHRvc3NlZA== 41291
+IHBvYmxhY2nDs24= 41292
+IG1hdGVpeA== 41293
+IHNpZWxsw6Q= 41294
+IG90dA== 41295
+IGNvbXB1bHM= 41296
+YWt1a2Fu 41297
+IG1hbmlmZXN0ZWQ= 41298
+IOyTuA== 41299
+IHV0bW9zdA== 41300
+IHJldmVyc2Fs 41301
+IHBsYWNlYm8= 41302
+IGJsYXQ= 41303
+IFN0dW5kZQ== 41304
+bWFuc2hpcA== 41305
+IGF0dGU= 41306
+IOyGjOqwnA== 41307
+IGlzdGVt 41308
+IGFubmF0 41309
+IFBsYXlzdGF0aW9u 41310
+IHphZA== 41311
+IHF1aXR0aW5n 41312
+IGZhbWluZQ== 41313
+IFJvdWdo 41314
+IEZsYW1l 41315
+IGhldXQ= 41316
+IG9wb3J0dW5pZGFk 41317
+IGZhaXNhaXQ= 41318
+IERQ 41319
+IGRpY2llbmRv 41320
+IE1lbGFuaWU= 41321
+IENhcm5l 41322
+bWVn 41323
+cGV0dG8= 41324
+SlVO 41325
+INC70Y7QsdC+0Lk= 41326
+IG9zdGU= 41327
+IEpKb25haw== 41328
+IHRoZWF0cmljYWw= 41329
+IGludmluY2k= 41330
+IGNvbW11bmlvbg== 41331
+dm9jYWw= 41332
+RWg= 41333
+IERldGFpbHM= 41334
+IHN0cm9sbA== 41335
+IFJheW1vbmQ= 41336
+IEFtZWxpYQ== 41337
+kaU= 41338
+IHByb2R1a3Q= 41339
+IG51ZXZhcw== 41340
+IG11c3Ru 41341
+bWF5xLE= 41342
+Y29sb3JlZA== 41343
+ZGVj 41344
+IGhqw6Rs 41345
+IHNlbnRpbWVudGFs 41346
+IHJlYWxtcw== 41347
+IGtyaXQ= 41348
+IHNleHQ= 41349
+IFBzeWNob2xvZ3k= 41350
+aGls 41351
+INC60L7RgNCw0LE= 41352
+IOuCtOydvA== 41353
+IFVuZGVyc3Rvb2Q= 41354
+IEd1dGVu 41355
+IGdhbmdz 41356
+IGV2ZW5pbmdz 41357
+RW50 41358
+IExlZ2FjeQ== 41359
+IENvbmdv 41360
+IGR1cmNoYXVz 41361
+IGJ1b3k= 41362
+ZXJlbGxh 41363
+V0FO 41364
+UHJl 41365
+INGA0LXQtA== 41366
+IENyaXNpcw== 41367
+IOydvOydtA== 41368
+IG1hbnVzY3JpcHRz 41369
+0LXRgtGA 41370
+IG5vbnByb2ZpdHM= 41371
+IGRpY3RhdG9y 41372
+IGJhc2tldHM= 41373
+IElzaA== 41374
+IHBlcnRv 41375
+IGRhdGFzZXRz 41376
+IGFtcGxl 41377
+Z2ViYXV0 41378
+IGNvbnRyaWJ1dG9y 41379
+IGNpYW8= 41380
+IGNvbmZpcm1pbmc= 41381
+IFVDTEE= 41382
+4pms 41383
+INGB0L0= 41384
+IG92ZXJ0dXJu 41385
+IHVucmVhbGlzdGlj 41386
+IFBpZWNl 41387
+b2NhdGU= 41388
+IGbDpGxsdA== 41389
+cG94 41390
+IOuztOyLnOuptA== 41391
+IOuplOs= 41392
+IENyZWF0aW9u 41393
+0Y7QtNCw 41394
+INeU15A= 41395
+IHdoYWNr 41396
+b2xpdGhpYw== 41397
+Y2VseQ== 41398
+INGB0L7QstGA0LXQvA== 41399
+IHNlcXVlbnRpYWw= 41400
+IHByb2Zlc2lvbmFs 41401
+IGNvb2xz 41402
+IHJlcGVudGU= 41403
+IGFpcmU= 41404
+ZW5uZXM= 41405
+cml0b3M= 41406
+INCS0LjQtA== 41407
+IGvDtnI= 41408
+IEJpdHRl 41409
+dWxhcnM= 41410
+IGluY29ycmVjdGx5 41411
+IHNoYXJwbHk= 41412
+IGJvbWJhcmQ= 41413
+64uY7J20 41414
+IGNocm9tb3NvbWU= 41415
+IGFkdmVydGlzZW1lbnRz 41416
+aHVu 41417
+INGJ0L7QsQ== 41418
+INCU0LDQttC1 41419
+IGJhdGh0dWI= 41420
+IFNubw== 41421
+2ZDZkQ== 41422
+IGJ1ZmZldA== 41423
+IEdyaWQ= 41424
+IEJyZXc= 41425
+aXNldA== 41426
+IEltcG9ydGFudA== 41427
+w7xtw7x6 41428
+IHZldG8= 41429
+IFdlcms= 41430
+IFNoYW0= 41431
+a3Jh 41432
+aWxlZW4= 41433
+aGVhcmQ= 41434
+IGRyYWluaW5n 41435
+IGtsYXNz 41436
+IGJha2F5xLFt 41437
+Y3R1cmU= 41438
+YW1vdXI= 41439
+IHNwb25zb3JzaGlw 41440
+IGRpc3RpbGw= 41441
+IHBhdGlv 41442
+IGtvbWI= 41443
+IG92ZXJ3aGVsbWluZ2x5 41444
+IEphbWFpY2E= 41445
+dWl0ZW4= 41446
+TGl0dGxl 41447
+IExPVA== 41448
+dGHEhw== 41449
+IGNvbW1hbmRlcnM= 41450
+IFdhdHRz 41451
+IE9wdGlvbnM= 41452
+7J2066m0 41453
+QUNU 41454
+IGluZGlzcGVucw== 41455
+IEZvcnNjaA== 41456
+b3RvbQ== 41457
+IM6tz4fOtc65 41458
+IHByYWlzaW5n 41459
+IOyYgeyDgeydhA== 41460
+IGFtYW4= 41461
+IGh5cG5vdA== 41462
+dGhtcw== 41463
+IG5hc3plag== 41464
+IG1vdXJuaW5n 41465
+IFNBWQ== 41466
+Y3lq 41467
+INCz0L7RgdGD0LTQsNGA 41468
+IGNhdQ== 41469
+bWVl 41470
+IHRhZGk= 41471
+TWVk 41472
+IGNhbGlkYWQ= 41473
+IHN0cmlwZQ== 41474
+IM61zr0= 41475
+IEthdHk= 41476
+IEVzY2FwZQ== 41477
+IG3DvHNzdGU= 41478
+INin2YTYpw== 41479
+0LrRgg== 41480
+IGpvYmJhcg== 41481
+IEplanU= 41482
+b3Jhcg== 41483
+IFNlcsOh 41484
+IE1lc3Np 41485
+w6F6 41486
+IFRyYW4= 41487
+IHBpZXJjaW5n 41488
+IGFyaXRobWV0aWM= 41489
+IHN0YWdnZXJpbmc= 41490
+IHBsdWdnaW5n 41491
+IEtBUg== 41492
+dmw= 41493
+tOyY 41494
+IFJlZ2llcnVuZw== 41495
+IE9jenl3acWbY2ll 41496
+IEVkZ2Fy 41497
+IGNvbmR1Y3Rpdml0eQ== 41498
+eWVsbGluZw== 41499
+dmFpcw== 41500
+YWRpYW4= 41501
+IGJ1bGt5 41502
+INGB0YDQsNCy 41503
+INC/0YDQvtC8 41504
+IHBhdmVk 41505
+IGJlbmRz 41506
+IFNraWxsc2hhcmU= 41507
+IE1tbW0= 41508
+IEhvcnJvcg== 41509
+IHR1bWI= 41510
+IGdvb2Z5 41511
+IE1lb3c= 41512
+15nXnNeV 41513
+IFdhc3M= 41514
+IFNjYWxl 41515
+IFJhaw== 41516
+IHByb2plY3Rpbmc= 41517
+IGxpbmd1aXN0aWM= 41518
+IFdvcmxkcw== 41519
+ZW5zZW1ibGU= 41520
+IHBlZ2E= 41521
+c3RvcHBhYmxl 41522
+IGltYmFsYW5jZQ== 41523
+IMO4 41524
+IHRocmlsbGVy 41525
+0LrQvtC70YzQutGD 41526
+IGxlZnRvdmVycw== 41527
+IGNhdmVhdA== 41528
+IFNUUg== 41529
+dW5kYWk= 41530
+IHdhdGVyeQ== 41531
+IE1hcmlu 41532
+IGVnZ3BsYW50 41533
+IEpC 41534
+2YXZg9mG 41535
+dmlkaWE= 41536
+IEZJTg== 41537
+aWNhYmxl 41538
+IHBvZG9i 41539
+IGNvaGVzaXZl 41540
+IFZlcmbDvGd1bmc= 41541
+IFBsYXRv 41542
+0LDRgNC40Yk= 41543
+IGtvdA== 41544
+INCf0L7QvA== 41545
+INC00L7QutGD0Lw= 41546
+IGltcGxhbnRz 41547
+aXNzZXo= 41548
+QnJl 41549
+IGdhc3Bz 41550
+IFRFRA== 41551
+cmF0bw== 41552
+Skk= 41553
+IGF2ZW51ZXM= 41554
+IENob25n 41555
+bGFkxLE= 41556
+2LHYtg== 41557
+IGluaWNp 41558
+IFN1YmFydQ== 41559
+4LiL 41560
+IGFjaHQ= 41561
+IEFyY2hpdGVjdHVyZQ== 41562
+INCy0LXRidC4 41563
+IERldk9wcw== 41564
+IHRvcHBpbmdz 41565
+IG9ic29s 41566
+YWluYQ== 41567
+IEJhbmdrb2s= 41568
+ZXN0cnVjdA== 41569
+IGtvYg== 41570
+IOuTrw== 41571
+INGA0LDQt9C90YvQtQ== 41572
+IHJlZQ== 41573
+IGJpanZvb3JiZWVsZA== 41574
+IERlbW9jcmFjeQ== 41575
+4LmA4Lij4Liy 41576
+INC60L7QvdGC 41577
+IHNlw6c= 41578
+IHJhaGF0 41579
+IHBhcmxpYW1lbnRhcnk= 41580
+IEJhc2g= 41581
+emlhxYI= 41582
+SVRDSA== 41583
+IEJ1YmJsZQ== 41584
+a3TDsw== 41585
+V2hvYQ== 41586
+IGZsYXRz 41587
+em5l 41588
+IHNlcnZpY2lv 41589
+IERldw== 41590
+1bjWgg== 41591
+IHVudGVyc3TDvHR6ZW4= 41592
+IFdpbmRz 41593
+IOyWmOuKlA== 41594
+IGV2YWx1YXRpb25z 41595
+IHJlY2E= 41596
+IGVsdmVz 41597
+Y2hlZXI= 41598
+IGphbA== 41599
+IHJlc3RlZA== 41600
+IHF1aWVuZXM= 41601
+IEJyb29rZQ== 41602
+IOuniOydjOyXkA== 41603
+IGludGVu 41604
+IG9hdHM= 41605
+IHJlZmVyZWU= 41606
+IHBuZXVtb25pYQ== 41607
+IGRlbHZl 41608
+cGVhY2U= 41609
+ZW55 41610
+IG1vc3RyYQ== 41611
+IENhbm5vbg== 41612
+z4HOv8+N 41613
+INCQ0Ls= 41614
+IG1vbnVtZW50YWw= 41615
+zr/Pjc68zrU= 41616
+aW1tZXJz 41617
+YXZpYW4= 41618
+INC00LXQu9Cw0LXRgg== 41619
+IHBpdGNoZXM= 41620
+IEdyb3Zl 41621
+IHNlbWluYXJz 41622
+IHLDqWN1cA== 41623
+IFZvb3I= 41624
+IGRldmVu 41625
+IGRC 41626
+IGJvb3N0aW5n 41627
+ZWdhbg== 41628
+IHdlbHQ= 41629
+IEd1YXRlbWFsYQ== 41630
+IG1pbGVhZ2U= 41631
+IGJlaGFuZA== 41632
+IFdhYXI= 41633
+IFN1cmY= 41634
+IGNhdWxpZmxvd2Vy 41635
+IFR5cg== 41636
+IG1pdGVpbmFuZGVy 41637
+IGRhcmluZw== 41638
+IFNpdHRpbmc= 41639
+ZGxlZA== 41640
+IHJlc2VudG1lbnQ= 41641
+bcOkw59pZw== 41642
+IGZpbG1tYWtpbmc= 41643
+d2FydHM= 41644
+dGhvdWdodA== 41645
+b2xvZ2lxdWU= 41646
+IENPUg== 41647
+IGFjY291bnRlZA== 41648
+IGFwZXI= 41649
+IElOVA== 41650
+b2xhcmU= 41651
+IGFjb21wYcOx 41652
+IMahaQ== 41653
+IG1lcm1haWQ= 41654
+IEJlbnRsZXk= 41655
+YXRvcmU= 41656
+IHByZW4= 41657
+IGV0aGFub2w= 41658
+IGFzdHJvbm9tZXJz 41659
+c2VhdA== 41660
+a2VlcGVycw== 41661
+IGV4ZW1wdGlvbg== 41662
+IGFtbw== 41663
+IOuCmOyEnA== 41664
+IGluaGFs 41665
+IGJvd3M= 41666
+0YHQutGD0Y4= 41667
+MzAwMA== 41668
+IGZlcm1lbnRhdGlvbg== 41669
+IHNpbmtz 41670
+IGNvbWVyY2lhbA== 41671
+IHN0dW1w 41672
+IGNlbGU= 41673
+IFNpc3RlcnM= 41674
+IFJlZ2lzdGVy 41675
+IHNvb3J0 41676
+IG5hdG9taWFzdA== 41677
+IOq3uOumvA== 41678
+IMWeZXk= 41679
+IGh5cGVk 41680
+IFJhZmFlbA== 41681
+IEVpcw== 41682
+IEJhc2ls 41683
+IEFzc2Fzc2lu 41684
+IEFkZQ== 41685
+csOlbg== 41686
+IG9ubGFy 41687
+IG1vdmltaWVudG8= 41688
+IGFkZGl0aW9uYWxseQ== 41689
+IHNsaXQ= 41690
+IENocnk= 41691
+IEludGVydmlld2Vy 41692
+15zXpw== 41693
+IGRpc2w= 41694
+IGxpZ2dlcg== 41695
+0YPQutC4 41696
+YmVyaXNo 41697
+INGA0Y/QtNC+0Lw= 41698
+QVJPTg== 41699
+XSws 41700
+IGx1bWnDqHJl 41701
+IG9sdmlk 41702
+IGZyZXVl 41703
+IFRpbmc= 41704
+IEvDtg== 41705
+IGdlbw== 41706
+IGR5ZWQ= 41707
+0YjQtdC5 41708
+IMW8eWNpZQ== 41709
+IGll 41710
+IHRheHBheWVy 41711
+IHBlxYI= 41712
+IGTDqWNpZMOp 41713
+IGPFk3Vy 41714
+IGVudHdpY2tlbHQ= 41715
+IEhR 41716
+S0s= 41717
+b2Rhcg== 41718
+IGhvbmU= 41719
+IGNvbmZpYW5jZQ== 41720
+IGlzc3Vpbmc= 41721
+IGRpYWdub3N0 41722
+IOyehA== 41723
+INC60YDRg9GC 41724
+INC60LDRgQ== 41725
+IMO+ 41726
+IHJlc3RyaWN0aXZl 41727
+IENhc3Rybw== 41728
+IHXEnw== 41729
+IGVtcHJl 41730
+IE1vbw== 41731
+IEZpZ3VyZQ== 41732
+cGhvbmV0aWM= 41733
+UHJvZg== 41734
+INC/0YDQtQ== 41735
+IHRpbHRlZA== 41736
+IE5lZ2F0aXZl 41737
+IExpbWl0ZWQ= 41738
+bWVubw== 41739
+bGFtYXRpb24= 41740
+IHRydXN0ZWVz 41741
+IGludGVuc2VseQ== 41742
+IGHDp8SxbA== 41743
+IFVzZWQ= 41744
+IHp1bA== 41745
+IGFwcHJlY2lhdGl2ZQ== 41746
+IHRpbmM= 41747
+IGNvbnF1ZXN0 41748
+INi52YbYrw== 41749
+IHN1aWNpZGFs 41750
+IG11bGhlcmVz 41751
+IGRldGFjaA== 41752
+IGthbWVyYQ== 41753
+IEFpclBvZHM= 41754
+SU5ESVNUSU5DVA== 41755
+0LPQu9C40Lk= 41756
+IOuDhA== 41757
+IHdyZXN0bGU= 41758
+IGZpcmVhcm0= 41759
+IGxpcmU= 41760
+cHJh 41761
+IGpld2Vscw== 41762
+IENvcm5lbGw= 41763
+IO2VoOqyjOyalA== 41764
+IHN1Y2tlcg== 41765
+IG5vbWJyZXV4 41766
+IEZlcm0= 41767
+7JuQ7J20 41768
+IFBpcw== 41769
+INC40LfRg9GH 41770
+IG1pdGVu 41771
+IGNldg== 41772
+IFVSTHM= 41773
+IENBUw== 41774
+ZmluZGVu 41775
+IGJyYXZlcnk= 41776
+INGB0LvQvtCy0L4= 41777
+IG5lbmh1bWE= 41778
+IGVuY3VlbnRyYQ== 41779
+IFNoaXJsZXk= 41780
+IHBlcmNlcHQ= 41781
+ZnJhbWVz 41782
+IFJvdmVy 41783
+IEFsYmVydGE= 41784
+b2Nj 41785
+IOudvOqzoA== 41786
+IHPDunBlcg== 41787
+IHByZXN1bWU= 41788
+IGdsYW5k 41789
+IHBhY2luZw== 41790
+IG5ldXJvdA== 41791
+IHNubw== 41792
+IHBsb3R0ZWQ= 41793
+IHBhxYRzdHdh 41794
+IE93bmVy 41795
+IERlZmVuY2U= 41796
+cmlkZ2Vz 41797
+IHdhbGxwYXBlcg== 41798
+b25pYW4= 41799
+QnJv 41800
+IEFyaWFuYQ== 41801
+a3J5 41802
+IG5hcnJhdGlvbg== 41803
+IGNyaWFuw6dh 41804
+IEFscmlnaHR5 41805
+IOydvQ== 41806
+IOyTsOqzoA== 41807
+IGxpYmVyYXRlZA== 41808
+IGV4Y2VlZHM= 41809
+IGRvbWluYXRpbmc= 41810
+IGJha8Sxbg== 41811
+bGs= 41812
+IHNsYXBwZWQ= 41813
+0JfQtA== 41814
+dW1lbnRhbA== 41815
+Z2V0dGFibGU= 41816
+IFJveg== 41817
+IEd1bA== 41818
+b3V2ZXJ0 41819
+IHNtYXNoaW5n 41820
+YXp1amU= 41821
+U2ly 41822
+IGdyYXRlZA== 41823
+QVRU 41824
+IGFydGljdWxhdGVk 41825
+IHN0b3Jh 41826
+IGV4dHJhdGVy 41827
+4buJ 41828
+z4PPiQ== 41829
+d2ly 41830
+IE1ldGU= 41831
+SW1w 41832
+IGhvb3I= 41833
+cGhhc2U= 41834
+INGH0YPQtA== 41835
+INCx0YDQsNGC 41836
+IGlkYWc= 41837
+IGNpbnE= 41838
+IGFwYXJlY2Vy 41839
+IElDRQ== 41840
+IHF1aWV0ZXI= 41841
+IGZhbHNjaA== 41842
+YWRpYw== 41843
+INC/0LvRjtGB 41844
+IE1lbnU= 41845
+dXhl 41846
+IFTDtGk= 41847
+IE1JTA== 41848
+IEhhag== 41849
+dmVyYnM= 41850
+IHR1YmluZw== 41851
+IG1hY2hzdA== 41852
+IGRhbGw= 41853
+VGVy 41854
+IGdlbGVu 41855
+IGN1Y3VtYmVycw== 41856
+IHdpZGdldHM= 41857
+IGRldnJhaXQ= 41858
+IG1pa2U= 41859
+IGludHJh 41860
+7ZWt 41861
+IMOF 41862
+IEh1bmQ= 41863
+cXVhcnRlcg== 41864
+IGV3 41865
+IGtlbHVhcg== 41866
+IG1hdHM= 41867
+IFRyaWNr 41868
+IEluZmluaXRl 41869
+nqg= 41870
+IHBlYWM= 41871
+IFByb3Rl 41872
+4KWI 41873
+IDE3MDA= 41874
+IFJhaXM= 41875
+4LmK 41876
+w6RobHQ= 41877
+aWZpY2E= 41878
+YWltZXI= 41879
+YcSH 41880
+IGFrbA== 41881
+IFZvbHZv 41882
+IFR5c29u 41883
+IFJvbmc= 41884
+aXJzaW4= 41885
+IOKZpQ== 41886
+IHBhcm9keQ== 41887
+bmF0aW9uYWw= 41888
+cG9k 41889
+YXlk 41890
+YW1ibGVk 41891
+IGdvdmVybm1lbnRhbA== 41892
+IGNvbmZvcnQ= 41893
+aWNpZGVz 41894
+IG5hc3pl 41895
+IFNoZXBoZXJk 41896
+IEtvbnRha3Q= 41897
+IGRpc3Byb3BvcnRpb25hdGVseQ== 41898
+INC60LvRjtGH 41899
+IHTDrXR1bG8= 41900
+IHNpbmE= 41901
+IGNvbXBvc2l0aW9ucw== 41902
+IFBG 41903
+IHZlcmts 41904
+IHN1aXZyZQ== 41905
+IGFzdGE= 41906
+IHN0YWtlaG9sZGVy 41907
+IHNhbW1h 41908
+IEJMQUNL 41909
+IG5vZGln 41910
+IGxldmE= 41911
+IGp1ZWdvcw== 41912
+IGVybnN0 41913
+IGJvdHRvbXM= 41914
+IFNpZ25hbA== 41915
+IHBvbGx1dA== 41916
+IGR1cmE= 41917
+TXVzaWs= 41918
+INC60L7QvNC90LA= 41919
+INCy0YHQtdC5 41920
+YWx0ZXI= 41921
+IFN0ZWY= 41922
+IEJpZ1F1ZXJ5 41923
+IFZlcmFudHdvcnR1bmc= 41924
+IOuLueyXsA== 41925
+IHF1aXp6 41926
+IExldHRlcg== 41927
+IEludmVzdG1lbnQ= 41928
+0YjRgg== 41929
+kOuNsA== 41930
+IGVuY29kaW5n 41931
+IHTDpG5rZXI= 41932
+IEt3 41933
+YW5uaWU= 41934
+MTEw 41935
+IHp3eQ== 41936
+IOynpw== 41937
+IGRhdw== 41938
+ZXN0w6Q= 41939
+IGRlY2VpdmU= 41940
+IEzDpG5kZXI= 41941
+aXNrbw== 41942
+IHBvZHN0YXc= 41943
+IFBoYXJhb2g= 41944
+7LOk 41945
+w7psdA== 41946
+IHR5w7Y= 41947
+IG11c2lteQ== 41948
+IHBj 41949
+IE5U 41950
+IENvc3Rjbw== 41951
+IM+Dzr/PhQ== 41952
+IHVuaW4= 41953
+cm91bmRz 41954
+IHJlbWluZGVycw== 41955
+IHB1aXNxdQ== 41956
+IGtyaWpnZW4= 41957
+IHdvcmtmbG93cw== 41958
+bmV0ZW4= 41959
+IOuQmOyngA== 41960
+IHNsZWVr 41961
+IGNvd29ya2Vycw== 41962
+YW1pZW50b3M= 41963
+IHdpdGNoZXM= 41964
+YmFhcg== 41965
+ZXRpZXM= 41966
+IHVubmF0dXJhbA== 41967
+IFNpY2s= 41968
+IEVmZW5kaQ== 41969
+amNpZQ== 41970
+IGNoYW1hZG8= 41971
+7JiA7Iq164uI64uk 41972
+IHByemVkc2nEmWJpb3I= 41973
+IGJvb2tzdG9yZQ== 41974
+IOyeoOq5kA== 41975
+IFNlcGFy 41976
+YW5naQ== 41977
+RXZldA== 41978
+IGVtZXJnZW5jaWVz 41979
+IFhNTA== 41980
+0L3QtA== 41981
+pbTrqbQ= 41982
+IOq/iA== 41983
+IOuTpOqzoA== 41984
+IHN1dA== 41985
+IFdpeg== 41986
+IGR5bmFtaWNhbGx5 41987
+b3BlcmF0aW9u 41988
+ZG90 41989
+IGluZWZmaWNpZW50 41990
+Y2xlYXJz 41991
+IG11bmRhbmU= 41992
+IFZlcm9uaWNh 41993
+2LHYqg== 41994
+cG9zZQ== 41995
+cGFp 41996
+IG55bG9u 41997
+IGF1bWVudGFy 41998
+IGFsbHRzw6U= 41999
+dmFr 42000
+IGNhcGFjaWRhZA== 42001
+IFdyZXN0bGluZw== 42002
+IGZlcnRpbGU= 42003
+IG3DqWc= 42004
+IE5hbm8= 42005
+0LDRgtC10LvQuA== 42006
+IOyWtOyp 42007
+IHRvY2E= 42008
+IEVn 42009
+4oE= 42010
+IOyz 42011
+bHVlbnQ= 42012
+IHNvbGVt 42013
+IGNpbmVtYXQ= 42014
+IFF1ZWw= 42015
+IG9yYml0cw== 42016
+IEhhcm0= 42017
+cmljYW5lcw== 42018
+IGJsdXJyZWQ= 42019
+INin2YTYsNmK 42020
+IGppbg== 42021
+IGdyZW5hZGVz 42022
+IGF0cm9j 42023
+IHdoZXJlaW4= 42024
+IHJlcGxlbg== 42025
+IENvbWljcw== 42026
+ZWRhYW4= 42027
+IGRlbmlt 42028
+IGVtYmFycmFzc21lbnQ= 42029
+IEdvbWV6 42030
+IEJ1c2Fu 42031
+aXZpdGllcw== 42032
+IHNhbGl2YQ== 42033
+IG1lcms= 42034
+IGlsZ2lsaQ== 42035
+INC60YDRg9Cz 42036
+IG9jY3VwYXRpb25hbA== 42037
+IFNhaGli 42038
+U3Rh 42039
+IGFkdmlzZXI= 42040
+IFRydWx5 42041
+IFlFQUg= 42042
+IOyeiOuKlOuNsOyalA== 42043
+emV3 42044
+YmFyZW4= 42045
+IHN0b2w= 42046
+IGJlbG9uZ2luZ3M= 42047
+IFJlc2VhcmNoZXJz 42048
+IGVmZW5kaW0= 42049
+z4XPhw== 42050
+xYLEhWN6 42051
+IFVuZw== 42052
+IEp1Yg== 42053
+IGNlcmVicmFs 42054
+4buHdQ== 42055
+INem16g= 42056
+INC/0L7QtNCw0YA= 42057
+IG1hcmNoZWQ= 42058
+IGF3YWtlbg== 42059
+IGFrbw== 42060
+IGFjZXB0 42061
+IGluaXRpYXRpb24= 42062
+bG90 42063
+IHfFgmFz 42064
+IE1vbmdvbA== 42065
+dXRyYWw= 42066
+IHRlbnRhbmc= 42067
+IGludmVyc2lvbg== 42068
+IOydtO2bhA== 42069
+IGxvaw== 42070
+xYJieW0= 42071
+UlM= 42072
+IHN0b3M= 42073
+IGludGVyYWN0cw== 42074
+IENhbGVuZGFy 42075
+IHZhbmlzaA== 42076
+IHBoeXNpb2xvZ3k= 42077
+IGxpbmVhcmx5 42078
+IEpZ 42079
+xJ9hbg== 42080
+ZnVuZGVk 42081
+aXppZXJ0 42082
+IHptaWFu 42083
+IEdyaWxs 42084
+IHVuYmVsaWV2YWJseQ== 42085
+b3RlY2hub2xvZ3k= 42086
+IENhcnM= 42087
+INmG24E= 42088
+IEZvbGdl 42089
+IEJldmVybHk= 42090
+w6Rpc2NoZW4= 42091
+IGF1bWVudG8= 42092
+7JuM7ISc 42093
+IG1haWxib3g= 42094
+IHN0ZWVkcw== 42095
+IFBlYWs= 42096
+IHd5a29y 42097
+IHByYXdkYQ== 42098
+0LjRgtGL 42099
+IGRpc2NvdXJz 42100
+IGFjY3VzZQ== 42101
+Y2Vzc28= 42102
+dWlyZQ== 42103
+INC/0L7Qv9Cw0LQ= 42104
+IHRoYQ== 42105
+IG1lYXN1cmFibGU= 42106
+YmVlcGluZw== 42107
+IElubmVu 42108
+INC/0Y/RgtGM 42109
+IGNvbXBldGVk 42110
+IEl0YWxpYW5z 42111
+IGVuY29udHJh 42112
+IG5pZXc= 42113
+IGZpbHRyYXRpb24= 42114
+INC/0YDQvtGE0LXRgdGB 42115
+IHBhamFtYXM= 42116
+IGNpbGFudHJv 42117
+IFNvYw== 42118
+THVj 42119
+IOq5gOs= 42120
+IE9kZA== 42121
+IGh5ZHJhdGlvbg== 42122
+0LzQvtCy 42123
+IHBseXdvb2Q= 42124
+IENvbXBldGl0aW9u 42125
+0LjQt9C90LXRgQ== 42126
+ZmxpZ2h0 42127
+IEJlaXQ= 42128
+Ym91cmc= 42129
+IGNvaWxz 42130
+IGPDom1lcmE= 42131
+IGFtZW5kZWQ= 42132
+xIFt 42133
+QW5nZWw= 42134
+IFN0YWN5 42135
+Zmxv 42136
+IG5vcm1hbGU= 42137
+IGNvbnNvbmFudA== 42138
+IGFjY29tcGFueWluZw== 42139
+0LrRlg== 42140
+IGlycml0YXRlZA== 42141
+IGbDpXR0 42142
+IGNyb2NvZGlsZQ== 42143
+kJjripQ= 42144
+IGFsYmVpdA== 42145
+IFBoaWxvc29waHk= 42146
+xYY= 42147
+eXRpYw== 42148
+IHLDqGc= 42149
+IGZyYW7Dp2E= 42150
+IGF0dGVudGl2ZQ== 42151
+SGFt 42152
+IGFscmVkZWRvcg== 42153
+c2Vp 42154
+INGB0LLQuNC0 42155
+IGdpbWJhbA== 42156
+IGNoaW5h 42157
+IPCfjrY= 42158
+INCS0LDQvA== 42159
+IHN0aW11bGF0aW5n 42160
+IE9yYQ== 42161
+eXRlcw== 42162
+IGhlZnQ= 42163
+IGhhdGVycw== 42164
+IGNvbXBsZXhlcw== 42165
+IDAz 42166
+csOzZA== 42167
+Y2xlYXI= 42168
+IGJlc3RlaHQ= 42169
+d255 42170
+bW9pbA== 42171
+IHNsb3BweQ== 42172
+IGluc2lnbmlmaWNhbnQ= 42173
+IGR1YmJlZA== 42174
+IOuWoA== 42175
+IGNvbnNpZ28= 42176
+0LvRg9GI0LDQuQ== 42177
+U24= 42178
+INeU16Y= 42179
+IM6M 42180
+IG5hZHppZQ== 42181
+IGZyZXNobWVu 42182
+dGFh 42183
+IHV3YWfEmQ== 42184
+IEZhdm9yaXRl 42185
+IENyaW1pbmFs 42186
+IGV2aWRlbg== 42187
+IHN5bWI= 42188
+TGVz 42189
+IEJlYXU= 42190
+dW5lZA== 42191
+cGxlbWVudA== 42192
+QWM= 42193
+IGRlcm1hdA== 42194
+IE5vbGFu 42195
+0YvQvw== 42196
+IHNpdHQ= 42197
+IGV2ZXJsYXN0aW5n 42198
+IGVzdGF2YW0= 42199
+INC80LjQug== 42200
+IGtow6Fj 42201
+IGludml0 42202
+IHRyZWJsZQ== 42203
+IGppZw== 42204
+bWFuaQ== 42205
+IHR1dm8= 42206
+IFJVUw== 42207
+IEVyZGU= 42208
+IER6acSZa3VqxJk= 42209
+IGJsdWViZXJyaWVz 42210
+a2VsbA== 42211
+YWNpb25z 42212
+0LLQuA== 42213
+TEVU 42214
+IHNwcm91dA== 42215
+IHNwb3I= 42216
+IGLDqm4= 42217
+IE1vbmE= 42218
+IENvbnRhaW4= 42219
+IEtleXM= 42220
+0L7Qt9GP 42221
+IGZ1bmNpw7Nu 42222
+IHJhcHBlbGxl 42223
+IGV2b2x2ZXM= 42224
+IHNjcmFwaW5n 42225
+IGNvbWVudMOhcmlvcw== 42226
+IHByYXRpcXVl 42227
+IGF1eGlsaWFyeQ== 42228
+IFNwb25nZQ== 42229
+0YHQutC40Lw= 42230
+dXZv 42231
+INGB0LDQvNC+ 42232
+IHNhbms= 42233
+IGhpZ2h3YXlz 42234
+IGludmVudGlvbnM= 42235
+INC40L3QvtCz0LTQsA== 42236
+IGNyZWF0aXZlbHk= 42237
+IGJlbmNobWFya3M= 42238
+b25jw6k= 42239
+YWxhbA== 42240
+IHNvdHRv 42241
+IGNhbHZlcw== 42242
+IE1vdg== 42243
+IGxhdmVuZGVy 42244
+IGV5ZWJhbGxz 42245
+IGF3YWl0aW5n 42246
+IFBhdHk= 42247
+2YTZhw== 42248
+IGVtYnJvaWRlcnk= 42249
+IGR1aA== 42250
+IGNhbWFy 42251
+IEJPQg== 42252
+IHNwYWNlZA== 42253
+IGfFgm9z 42254
+0LDQtdC80YHRjw== 42255
+IGVzY2FwZXM= 42256
+IFJvZ3Vl 42257
+emN6 42258
+6J4= 42259
+rOulvA== 42260
+IE1vxbxl 42261
+INC10YHRgtC1 42262
+IEJ1cmFkYQ== 42263
+d2Q= 42264
+dXV1dQ== 42265
+IHNhc2g= 42266
+IEx1Yg== 42267
+IG5vdGVib29rcw== 42268
+IG1hZQ== 42269
+IGNvbmZsaWN0aW5n 42270
+IHN1bW1lcnRpbWU= 42271
+YWNhcw== 42272
+IGJhdWVu 42273
+Ymxvd2luZw== 42274
+4bqhbw== 42275
+IOyWuOygnA== 42276
+IFNlbmhvcg== 42277
+IGlQaG9uZXM= 42278
+IFF1YXJ0ZXI= 42279
+IOygnOuMgOuhnA== 42280
+dcOf 42281
+IOuniOustOs= 42282
+IHNldHRsZXJz 42283
+IGNyZXN0 42284
+IHRyYW5zYw== 42285
+IHJpb3Rz 42286
+IGNsb25lcw== 42287
+IE9wcmFo 42288
+zq/Otg== 42289
+IHBhbHM= 42290
+Li4uLi4uLg== 42291
+INGA0L7RgdGB 42292
+IExhc2Vy 42293
+IHphY3p5 42294
+IHNldmk= 42295
+IHJlZ2VuZXJhdGlvbg== 42296
+7Je8 42297
+d291bGQ= 42298
+IMO8emVyaW5l 42299
+IFN0cmHDn2U= 42300
+IHZlbmdlYW5jZQ== 42301
+IHJlcg== 42302
+IFNhZmFyaQ== 42303
+IEhFWQ== 42304
+IHNhY2Fy 42305
+IGltYWdlbQ== 42306
+IEJ1bmRlc3Q= 42307
+bWVzYW4= 42308
+IFBhc3Rl 42309
+IHNpeno= 42310
+INC/0L7RgdGC0YPQvw== 42311
+15TXlQ== 42312
+dHJhZA== 42313
+IGZyYW7Dp2Fpc2U= 42314
+IEJvdQ== 42315
+IGJhcnJl 42316
+IFpoaQ== 42317
+IEdlZXo= 42318
+aWhhZA== 42319
+IHJlY29ub2M= 42320
+IHBlbGln 42321
+IGluZGljZXM= 42322
+IOuwlOuA 42323
+IGNvbmR1Y3Rpb24= 42324
+IOyVhQ== 42325
+IHpla2Vy 42326
+IGZ1bQ== 42327
+IFfDvHI= 42328
+YnJlYWtlcg== 42329
+IHNwcml0ZQ== 42330
+Q3Jvd2Q= 42331
+IG9wZW5lcg== 42332
+IG9sdg== 42333
+IGJ1ZW5hcw== 42334
+IFNpbGs= 42335
+IEhJTQ== 42336
+a29w 42337
+Y29tcGw= 42338
+IHBvc3Nvbm8= 42339
+s4A= 42340
+IG9zY2lsbGF0b3I= 42341
+IFNpdGg= 42342
+0LDQttC4 42343
+IHJhZnQ= 42344
+aGFsbA== 42345
+IHNjaG5lbGxlcg== 42346
+IGltcG9ydGluZw== 42347
+IGFzc2VtYmxpbmc= 42348
+IHViaXF1 42349
+IGFjdGl2YXRlcw== 42350
+YWNjaQ== 42351
+k5zrpbw= 42352
+IGNvbXBvc2Vycw== 42353
+IEFDTA== 42354
+Q29uZg== 42355
+IOy9mA== 42356
+INC90LXQutC+0YLQvtGA0YvQtQ== 42357
+IGNhbmRpZXM= 42358
+IE11c3M= 42359
+4LmD4LiK 42360
+IGR1ZGE= 42361
+0L3QuNC60L7QvA== 42362
+bWVkZW4= 42363
+IOyWtOuVjA== 42364
+IFllc2h1YQ== 42365
+emFn 42366
+aG9kb3U= 42367
+IGFsb3Vk 42368
+IFBhbG1lcg== 42369
+aW1pemU= 42370
+IG1hcml0aW1l 42371
+IGNvbW11bmFs 42372
+IGJhZGdlcw== 42373
+IHJ1Z2J5 42374
+IG1hcnNobWFsbG93 42375
+IGZpZXJ5 42376
+IGFjY291bnRhbnQ= 42377
+IGFibGE= 42378
+IE1vbnJvZQ== 42379
+IEZvbnQ= 42380
+IEJvb3N0 42381
+IEJhcm5lcw== 42382
+YW5zd2Vy 42383
+IEJ1cm5pbmc= 42384
+IGFuZ2Vm 42385
+IFdlc2xleQ== 42386
+bGxz 42387
+7LU= 42388
+16nXnA== 42389
+aWxpxZtteQ== 42390
+15DXnw== 42391
+YW11cmE= 42392
+IEZ1ag== 42393
+IHBhbmk= 42394
+IFRyb3A= 42395
+YXJiZWl0ZW4= 42396
+IHJ1ZQ== 42397
+IFJhcmU= 42398
+w6RuZ2Vu 42399
+INGB0LzQvtGC0YDQtdGC0Yw= 42400
+INCa0LDRgA== 42401
+IE1UVg== 42402
+Ym9hcmRpbmc= 42403
+XVs= 42404
+IOugiOs= 42405
+c3RhbmJ1bA== 42406
+cGllbHQ= 42407
+IEhhcmR5 42408
+IEVuZ2FnZW1lbnQ= 42409
+IERpZW5zdA== 42410
+IHfDpHJlbg== 42411
+IGZ1ZWdv 42412
+IGVzdHJ1Y3Q= 42413
+IGNhbGFt 42414
+IFJlc3BvbnNl 42415
+IE1vaGFtbWFk 42416
+IHJlc2lzdGluZw== 42417
+IGR1cmFudA== 42418
+IE9MRUQ= 42419
+IHZlcno= 42420
+bcOkbg== 42421
+INmG25I= 42422
+IHBhcmFub2lk 42423
+IEF3YXJl 42424
+IEVuZ2luZWVycw== 42425
+IHByb2NlZHVyYWw= 42426
+IHBlcnNvbm5hZ2U= 42427
+IGZhcmtsxLE= 42428
+Zmxvd2luZw== 42429
+INC80LXRgdGC0LA= 42430
+IEJhcmU= 42431
+aXN0ZW0= 42432
+IHBvY3rEhXRrdQ== 42433
+IHBlcnNvbmFqZXM= 42434
+IOyWtOugtQ== 42435
+rYk= 42436
+INCl0L7RgtGP 42437
+IHVuc2V0dA== 42438
+IEFic29s 42439
+IOG6pXk= 42440
+IE1BWU9S 42441
+0L/QvtC70L3QtQ== 42442
+IGluZm9ybWluZw== 42443
+IGFtcHM= 42444
+0J/RgA== 42445
+IOutlA== 42446
+YWVkYQ== 42447
+INeU15HX 42448
+4bqlbg== 42449
+a2VsaWpr 42450
+IGF0aGVpc3Q= 42451
+IHRyb3V0 42452
+IG5ldWVz 42453
+IE5va2lh 42454
+bWFjaGVu 42455
+IHdob2xlc2FsZQ== 42456
+xLFyZA== 42457
+SW5z 42458
+INGN0L8= 42459
+IHByaWNr 42460
+IEtpbmRlcm4= 42461
+4LiX4Liz 42462
+IGNsYXNzeQ== 42463
+IMOubnQ= 42464
+IFNob3BpZnk= 42465
+INGB0L7RgA== 42466
+INC30LDQutGA0Ys= 42467
+enVr 42468
+IHVuaXZlcnNhbGx5 42469
+IHRlYXNwb29ucw== 42470
+IHJlY291bnQ= 42471
+IG7DpWdvbnRpbmc= 42472
+IFh1ZQ== 42473
+aXNpw6htZQ== 42474
+IHdlYWtlc3Q= 42475
+IHRlxZ9la2vDvHI= 42476
+IG1hdGhlbWF0aWNhbGx5 42477
+IEhvcw== 42478
+IO2VnOuLpA== 42479
+IHBhcnRhZ2Vy 42480
+IERhcnI= 42481
+6ro= 42482
+IM61zro= 42483
+IGdlcm1z 42484
+IGdlbGly 42485
+IGR1bA== 42486
+LC0= 42487
+IOyWuOs= 42488
+INee16Y= 42489
+INGP0YA= 42490
+IHF1b3RpZA== 42491
+IHByenlzeg== 42492
+IGhhcmRuZXNz 42493
+IGFxdWF0aWM= 42494
+IEp1bmdsZQ== 42495
+IFBDUg== 42496
+IEVsaW90 42497
+IG9zdHI= 42498
+IG1hcGE= 42499
+ZXNzw6Q= 42500
+IEdJUg== 42501
+IERyaXZpbmc= 42502
+IFNhbWk= 42503
+IE1lZGllbg== 42504
+IENvbXBhbmllcw== 42505
+IFBoYXJt 42506
+c2VpdHM= 42507
+IFJpbQ== 42508
+IM6/z4DOvw== 42509
+IHdlaXRlcmVu 42510
+IHBpenphcw== 42511
+IEx5ZGlh 42512
+IEhlaWdodHM= 42513
+IHNpbmNlcml0eQ== 42514
+IG5vc3Nhcw== 42515
+IGTFgg== 42516
+IGFsYXJtaW5n 42517
+IENhdWM= 42518
+INGB0LzRi9GB 42519
+ZmFjaW5n 42520
+YmFncw== 42521
+V1c= 42522
+INi02Yo= 42523
+IGNvdXJ0cm9vbQ== 42524
+IFBoaWxsaXA= 42525
+IOqyg+yymOufvA== 42526
+IFNwaWVsZXI= 42527
+IGthbnQ= 42528
+IGFkbWl0dGluZw== 42529
+IGNvbnRhaW5tZW50 42530
+IHJlbW92YWJsZQ== 42531
+IGp1bXBlcg== 42532
+Zm9jdXNlZA== 42533
+INC40YLQvtCz0LU= 42534
+INCi0LXQvA== 42535
+IHZhc2U= 42536
+IFVTQw== 42537
+IE1vbmF0ZQ== 42538
+IEphY29icw== 42539
+IEhPTA== 42540
+aWtlZA== 42541
+ZXJ3ZWlzZQ== 42542
+IGdvb2RpZXM= 42543
+IGhvbWFnZQ== 42544
+15vXqdeZ15U= 42545
+IHF1YWlz 42546
+IGluaWNpYWw= 42547
+IGd1YXJkaW5n 42548
+IGRheno= 42549
+IGNvbWJvcw== 42550
+INGD0L/RgNCw0LI= 42551
+IFRhbGVudA== 42552
+IMOzcg== 42553
+IGludGVybWl0dGVudA== 42554
+IE1jQ2FydGh5 42555
+IHNwYW5z 42556
+IHR5cmU= 42557
+IHF1eQ== 42558
+anV0 42559
+IFplbnQ= 42560
+IGdhdA== 42561
+IHNjYWZmb2xk 42562
+IG5lY2VzYXJpbw== 42563
+IFphaGxlbg== 42564
+IFNBTkQ= 42565
+IFBV 42566
+RXZlcnl0aGluZw== 42567
+LS0tLS0tLS0tLS0tLS0tLQ== 42568
+INCy0LfRj9GC0Yw= 42569
+IHNwYXJrcw== 42570
+IHBlbmR1bHVt 42571
+157Xnw== 42572
+IOyDieq5 42573
+IG11bHRpcGxpZXI= 42574
+INC70LDQtNC90L4= 42575
+dXJhdA== 42576
+IHVwc2V0dGluZw== 42577
+YmFr 42578
+IOy1nOuMgA== 42579
+IGFuw6Fs 42580
+IEpPRQ== 42581
+IGtvc3Rlbg== 42582
+IFBhdHR5 42583
+IEd1aW4= 42584
+Y2tlZA== 42585
+IEVneXB0aWFucw== 42586
+IENpdGl6ZW5z 42587
+16jXmw== 42588
+INCV0YnQtQ== 42589
+INC50L7Qs9C+ 42590
+IHNub3dmbA== 42591
+IGxla2tlcg== 42592
+IGFjb3N0 42593
+IEJhYmU= 42594
+IGdhbWJsZQ== 42595
+IGFkamVjdGl2ZQ== 42596
+0LrQuNC80Lg= 42597
+b3lz 42598
+IG1vbnRyZQ== 42599
+IEh5dW5kYWk= 42600
+IG1vaXN0dXJpemluZw== 42601
+IG1venphcmVsbGE= 42602
+T09P 42603
+IGZhY3VsdA== 42604
+IGRvZXQ= 42605
+IGZlYXJsZXNz 42606
+IGVzcHJlc3Nv 42607
+IGFsbG9yYQ== 42608
+IENpbmM= 42609
+IGNvbnRlw7pkbw== 42610
+IFBlbG9zaQ== 42611
+IG1pbmRlcg== 42612
+cm9vdA== 42613
+IO2VoOs= 42614
+INC/0LDQtA== 42615
+IENhbGxpbmc= 42616
+IENvbmZpZw== 42617
+IENvbnNvbGU= 42618
+aW5za3k= 42619
+w6luZXJnaWU= 42620
+IHNvbGl0YXJ5 42621
+0L7QtNC1 42622
+IGd1YXJkZWQ= 42623
+MTYw 42624
+INC/0YHQuNGF 42625
+IFNoYXA= 42626
+IHRpdHJl 42627
+b2xvZ25l 42628
+INC/0LDRgNGD 42629
+IFBSRQ== 42630
+IGxu 42631
+IE1pdGds 42632
+IENhcnJ5 42633
+IHNwaW5k 42634
+IENhbnRvbg== 42635
+IGtpbmdkb21z 42636
+cmVtbw== 42637
+IHJhZ2luZw== 42638
+IGluY2FwYWJsZQ== 42639
+IFdS 42640
+INGB0L7QsdGB0YLQstC10L0= 42641
+INC60LDQutC40YU= 42642
+IFNIRQ== 42643
+64u57Z6I 42644
+IHNjYXJjaXR5 42645
+IHBlcmRl 42646
+IGV4aXRz 42647
+IFNpbmdlcg== 42648
+IHN1cHBlcg== 42649
+IG11bmljaXBhbGl0eQ== 42650
+IERpdmVyc2l0eQ== 42651
+IHRpcm8= 42652
+aWVscw== 42653
+IGzDrWRlcg== 42654
+IGJsdWZm 42655
+IGF0cmE= 42656
+bHlz 42657
+IG1haGQ= 42658
+IGPDs2RpZ28= 42659
+IEhhcmxlbQ== 42660
+cnVsZQ== 42661
+aWNpdHk= 42662
+IHNpbXBsaXN0aWM= 42663
+IEtvbnN0 42664
+RUxMSQ== 42665
+IGbDtnJzdGE= 42666
+IGNvbnN0aXR1dGVz 42667
+INGB0YLQvtGA0L7QvdGD 42668
+IHVyZ2Vk 42669
+IFBhbmRh 42670
+7LCo6w== 42671
+cmVjZQ== 42672
+IHBhdHJpb3Q= 42673
+IENydXNo 42674
+IHdpbms= 42675
+0L7QudGC0Lg= 42676
+dXJhbsOnYQ== 42677
+IHNlaXp1cmVz 42678
+IGVsZWN0cm9k 42679
+IERvbmtleQ== 42680
+IElV 42681
+IE1PUw== 42682
+IGFsa2Fs 42683
+7LSJ 42684
+YmVzb25kZXJl 42685
+IHBhcmFsbGVscw== 42686
+IGJpdHRlcm5lc3M= 42687
+w6R0dHJl 42688
+ZXNzaW9uYWw= 42689
+IHNveWJlYW4= 42690
+IGNvbGxhYg== 42691
+IFJlcG9ydGluZw== 42692
+INC60L7QvNC/0LDQvdC40Lg= 42693
+IHdzenlzY3k= 42694
+IENydW5jaA== 42695
+aXNlZW4= 42696
+IGFtYmFzc2Fkb3Jz 42697
+IENoZXY= 42698
+0L7QstGL0LU= 42699
+c2Nh 42700
+INGA0LXRiNC40Ls= 42701
+0L7RgtC+ 42702
+IGdsZWljaHplaXRpZw== 42703
+bWVybg== 42704
+w7xzdA== 42705
+IEhhZQ== 42706
+s7TqsqDsirXri4jri6Q= 42707
+IHNob3Jlcw== 42708
+IGRlcHJlc3M= 42709
+IGFob3I= 42710
+IFN0ZXVlcg== 42711
+YWho 42712
+IHJldmlzZQ== 42713
+INGB0LDQvNGL0LU= 42714
+amF0 42715
+IGhlcmJhbA== 42716
+IGN1w6FudA== 42717
+IGJ1bmE= 42718
+bmllanN6ZQ== 42719
+RmluYWxseQ== 42720
+15XXlg== 42721
+Y2pl 42722
+IOyeiOqxsOuToOyalA== 42723
+IOuCmOuI 42724
+IHByemVzdA== 42725
+bGljYQ== 42726
+IER1Y2g= 42727
+0ZbQudGB0Yw= 42728
+cGFzc2Vu 42729
+IHNhdGlzZmllcw== 42730
+IEFkZGl0aW9uYWw= 42731
+IGPDoW1hcmE= 42732
+0LXRh9C10L3QuNC1 42733
+IHBvbXA= 42734
+IOunkOydtA== 42735
+IE1pbGxz 42736
+0LXQstC40LQ= 42737
+IHJlc3BlY3RhYmxl 42738
+IGZpbGFtZW50 42739
+IHZlbmRlcg== 42740
+IG1hdHRlcmVk 42741
+b3VyZQ== 42742
+7Li1 42743
+S29yZWFu 42744
+IGVzdHVkaW8= 42745
+IGNhY3R1cw== 42746
+IFZpdmU= 42747
+IFJhZw== 42748
+IGNvbXBsaXF1w6k= 42749
+INmI24E= 42750
+IHRhbw== 42751
+pr8= 42752
+U2luY2U= 42753
+IGplb3BhcmQ= 42754
+IFNlbGw= 42755
+IOyYmw== 42756
+IGtldG8= 42757
+IGludGVsaWc= 42758
+IEFuZ2Vi 42759
+IHRpZGVu 42760
+IHNvY2lv 42761
+IHJlbWluaXNjZW50 42762
+IGNhcmVnaXZlcg== 42763
+U3BhY2U= 42764
+IEV4ZXJjaXNl 42765
+IEJlY29tZQ== 42766
+w6p0cw== 42767
+YWtr 42768
+IS4u 42769
+INGB0L/RgNC+0YE= 42770
+IM6xz4DOvw== 42771
+IHNob290aW5ncw== 42772
+IGFwZQ== 42773
+IFNhbW15 42774
+IEt1bmc= 42775
+IGN1w6Fs 42776
+IEx1cA== 42777
+INGB0YLRg9C0 42778
+IHN3ZWV0ZXI= 42779
+IGNvbXVt 42780
+IEFkcw== 42781
+aHl1bmc= 42782
+INCx0YPQtNGD0Yk= 42783
+IHdhZmZsZQ== 42784
+IE9yYg== 42785
+IGxhdXQ= 42786
+IGZvcmVjYXN0aW5n 42787
+5ao= 42788
+IHJhcHBpbmc= 42789
+IHByZWZlcnM= 42790
+IGJlbno= 42791
+IG5paw== 42792
+IEJhaG4= 42793
+IHNhbmRpbmc= 42794
+IGltbWluZW50 42795
+INC/0YDQvtCx0LvQtdC80Ys= 42796
+IGRvaXZlbnQ= 42797
+0L7Qu9Cw 42798
+IMW8eWNpYQ== 42799
+aWh1 42800
+IGV4aXN0ZW0= 42801
+IEludGVyaW9y 42802
+IFRha2Vz 42803
+IHRvZGRsZXI= 42804
+IGRpY3RhdG9yc2hpcA== 42805
+IFNtaXRoc29u 42806
+IEFsbGFodQ== 42807
+z47Pgc6x 42808
+7JWY7Iq164uI64uk 42809
+IFZvdGU= 42810
+IFNtZWxscw== 42811
+0L7QtNC90L4= 42812
+IGhpbmRzaWdodA== 42813
+VlI= 42814
+IFBhdGNo 42815
+IEphaHJlcw== 42816
+IHNvdXZlbmly 42817
+IG5ldXRyb24= 42818
+IGxvbmd0aW1l 42819
+IHNheWlu 42820
+YXNha2k= 42821
+INC+0YHRgtCw0L3QvtCy 42822
+IGV4cGVsbGVk 42823
+IGNyeXB0b2N1cnJlbmNpZXM= 42824
+IE11cmRlcg== 42825
+IENpdGl6ZW4= 42826
+V0FZ 42827
+IHBsdQ== 42828
+IGxlbW9uYWRl 42829
+IGNvbnZlbmllbnRseQ== 42830
+IEhJ 42831
+IDIwMjM= 42832
+16nXldeq 42833
+0LDRhtC40L7QvQ== 42834
+IOubsA== 42835
+INmE2YPZhg== 42836
+INC90LXQvNC90L7QttC60L4= 42837
+IHVudXNlZA== 42838
+IG1haW9yaWE= 42839
+IGFzdHJvbG9neQ== 42840
+IERvd250 42841
+Tmljaw== 42842
+IHByZW9jY3Vw 42843
+IGRlbWFpbg== 42844
+157Xog== 42845
+INCy0L7QtNGL 42846
+IFNhbnNrcml0 42847
+IHByw6p0 42848
+IHN0cmFuZGVk 42849
+IHJlZmlu 42850
+INC/0YDQuNC90LjQvA== 42851
+INC/0L7QstC10YDRhQ== 42852
+4K+NPw== 42853
+IHpyb2I= 42854
+IGludGVydHc= 42855
+IERhdmlkc29u 42856
+0LvQtdC90LA= 42857
+INC/0L7QvdGP0YLRjA== 42858
+IFJlbm8= 42859
+INC/0L7Qu9GD0YfQuNC70L7RgdGM 42860
+IGNvcnJlc3BvbmRlbnQ= 42861
+IFVyYW4= 42862
+ZWxzZQ== 42863
+wrfCtw== 42864
+IHR1dG9yaW5n 42865
+IGdyYW5kZGF1Z2h0ZXI= 42866
+bHVkZWQ= 42867
+IHN0ZXNzbw== 42868
+IGjhur90 42869
+IGdlZ2FuZ2Vu 42870
+INCd0JA= 42871
+IGFudGln 42872
+YmFja2dyb3VuZA== 42873
+IGdlZGFhbg== 42874
+IGZhdm9yZWQ= 42875
+IEVtbWFudWVs 42876
+IGlvZA== 42877
+IGNsYW1wcw== 42878
+IGNvbXBsZQ== 42879
+IEFkdmFuY2U= 42880
+IOyeiOqzoOyalA== 42881
+IFJveA== 42882
+IOyXkOs= 42883
+IGludGVzdGluZXM= 42884
+IHBlcmN1c3Npb24= 42885
+IGxlZ2l0aW1hdGVseQ== 42886
+IEV0ZXJuYWw= 42887
+ZmFtaWx5 42888
+YWxvZw== 42889
+QnJhZA== 42890
+0LXQvdC40YLRjA== 42891
+INGB0L3QsNGH0LDQu9Cw 42892
+IGNlcnRh 42893
+IGFra29y 42894
+IM61zrTPjg== 42895
+IG9jdGF2ZQ== 42896
+IFZhYw== 42897
+0LzQvtGC0YDQuA== 42898
+IMOJdGF0cw== 42899
+IGxvbmd1ZQ== 42900
+IGRpc3NvY2k= 42901
+0YDRj9C0 42902
+aGVpbg== 42903
+IHBhbnRhbGxh 42904
+IGluZGljYXRpb25z 42905
+IEx0 42906
+IEdyYWRl 42907
+b2luZQ== 42908
+YnVn 42909
+IFZlcml6b24= 42910
+IEFsw6lt 42911
+IHZpZW5uZW50 42912
+INGH0LjRgdGC 42913
+IEJlbmk= 42914
+IFRzY2g= 42915
+IFRQ 42916
+IGluc3VsdGluZw== 42917
+IFdlaWdodA== 42918
+IGFkYXB0YXRpb25z 42919
+IGhhYsOtYW4= 42920
+IGNsaXF1ZQ== 42921
+b8WbY2k= 42922
+anVuYQ== 42923
+IHN1Y2hlbg== 42924
+IEdvZXM= 42925
+IEV4b2R1cw== 42926
+Q2hv 42927
+IGFudGlz 42928
+IO2MjOs= 42929
+c2V2ZW4= 42930
+INGH0LDRgdC+0LI= 42931
+IGJhbGxpc3RpYw== 42932
+em9ueQ== 42933
+SUNJQQ== 42934
+INC/0YDQtdGB0YI= 42935
+IHNpbXBsZXNtZW50ZQ== 42936
+IENvbGxhYm9y 42937
+RnJlZA== 42938
+INGC0LXQu9C10YTQvtC9 42939
+IFJhdmk= 42940
+7ZW07KQ= 42941
+0L/QtdGA0LI= 42942
+IOyeiOycvOuLiOq5jA== 42943
+IMOzdA== 42944
+IGFsZWc= 42945
+w7pw 42946
+IGRpc3JlZ2FyZA== 42947
+IGluZGVudA== 42948
+Y2xvdWQ= 42949
+Y2hsYWdlbg== 42950
+IGl0ZXJhdGU= 42951
+IGdlbmVyYWxpemVk 42952
+4KS5 42953
+ZWxlcmk= 42954
+IGRpc2FzdHJvdXM= 42955
+INGB0YLQsNC70LA= 42956
+s5E= 42957
+S05PV04= 42958
+IHJpY2huZXNz 42959
+IGNvbnNjaWVudA== 42960
+aWNodHM= 42961
+INGN0LvQtdC8 42962
+2KjYrw== 42963
+aXJlbnM= 42964
+IGhhdW50aW5n 42965
+cnVjdHVyZXM= 42966
+YXR0YWNr 42967
+IGN1cGNha2Vz 42968
+c3F1ZQ== 42969
+IG5hc3plZ28= 42970
+IGFudGhyb3BvbG9neQ== 42971
+Y2hhZQ== 42972
+IGRpc2NvdmVycw== 42973
+IFBlcnNvbmFsaXR5 42974
+IM6kzr8= 42975
+IGRpxJ9lcg== 42976
+INC90LXRkQ== 42977
+IEFuaXRh 42978
+IFvimao= 42979
+IENhcm0= 42980
+IEJlbm55 42981
+7Iqs 42982
+IHB1cGls 42983
+IG9jYXM= 42984
+w6RsbGV0 42985
+asWbxIc= 42986
+YW1lbnRhbA== 42987
+INC+0YLQvdC+0YE= 42988
+IHBpZA== 42989
+IGFybXA= 42990
+UkVF 42991
+INC+0YLQutGA0YvQsg== 42992
+IHVkYQ== 42993
+IFN5bmRyb21l 42994
+IFN0YW5kYXJkcw== 42995
+IHBvaW50ZXJz 42996
+IGVuYW0= 42997
+IFRpZw== 42998
+w616 42999
+INC90LDQvNC4 43000
+IHVuY2hhbmdlZA== 43001
+IHR1cm1vaWw= 43002
+4bupbmc= 43003
+ISEi 43004
+NTAwMA== 43005
+IOusvOyWtOs= 43006
+IG1lcmdpbmc= 43007
+IGVudHNjaGVpZGVu 43008
+Zm9ybWU= 43009
+IHRyaW1tZWQ= 43010
+IGRhcmVk 43011
+IGFzcGlyYXRpb24= 43012
+IE15dGhpY2Fs 43013
+IEhlag== 43014
+IEFsZWo= 43015
+0YbQvg== 43016
+0L7RgtGD 43017
+WmU= 43018
+INC40L3RgdGC0YDRg9C80LXQvdGC 43019
+IFJUWA== 43020
+IGxvY2FsaXplZA== 43021
+IHN1cnJvdW5kcw== 43022
+IGVtcGllemE= 43023
+IGNsYXNl 43024
+IOC4gQ== 43025
+IFJhcGlk 43026
+b21pbm91cw== 43027
+aWdhaWw= 43028
+INGI0LjRgA== 43029
+IGzDpg== 43030
+IHphc2Fk 43031
+IHVuZm9sZGluZw== 43032
+PyE/IQ== 43033
+IOyInOqwhA== 43034
+IFBvbHNraQ== 43035
+IEthdWY= 43036
+IENlbHQ= 43037
+aXRpYw== 43038
+IHRvb2xib3g= 43039
+IFBvY2tldA== 43040
+IOyEnOuhnA== 43041
+IGJlbGtp 43042
+IGFkbWlyYXRpb24= 43043
+cGhy 43044
+IFByb2R1a3Q= 43045
+IFRydWNr 43046
+IGRyYXXDn2Vu 43047
+d2HFgg== 43048
+IEhlYnJld3M= 43049
+IO2VmOqyjA== 43050
+IEFDRQ== 43051
+dXJnZW5jZQ== 43052
+YXVyYWlz 43053
+IGNoYXJpdGFibGU= 43054
+xLF0 43055
+IGFybWFz 43056
+IEdlZGFua2Vu 43057
+cmVhdGluZw== 43058
+cG9ydGU= 43059
+IGltcHJpbnQ= 43060
+ZsOkaA== 43061
+INC/0L7QtNGF0L7QtA== 43062
+IG91dHNldA== 43063
+4Lin4LiB 43064
+0LXQvdC90L7Qs9C+ 43065
+Q2xhc3M= 43066
+IHZhbml0eQ== 43067
+IFZPSUNFUw== 43068
+IDI2MA== 43069
+cmVzaWRlbnQ= 43070
+VVNF 43071
+IOqwgOyatOuNsA== 43072
+6b0= 43073
+IHRocm91Z2hwdXQ= 43074
+IGN1bWE= 43075
+7Jqx 43076
+INC/0LvQvtGJ 43077
+IHBhcnRpcw== 43078
+IEFuaW1hdGlvbg== 43079
+p4jr 43080
+Q3Jl 43081
+w7Z0emxpY2g= 43082
+IG1hZ2c= 43083
+IGNsdW1zeQ== 43084
+IGJvdHRsZW5l 43085
+IGJpcmxpa3Rl 43086
+IEdhbWI= 43087
+INeb158= 43088
+IG1ldHJvcG9saXRhbg== 43089
+T29o 43090
+IG9iamVjdGlvbnM= 43091
+INmF2Ko= 43092
+INC80LXQuw== 43093
+IHJlbW5hbnRz 43094
+IFhhdmllcg== 43095
+UmljaA== 43096
+IG9sc2E= 43097
+IFBpbGw= 43098
+IGdyb2Fucw== 43099
+IE5hcnVob2RvdQ== 43100
+IENvbnRyYWN0 43101
+0LDQtNCw 43102
+bmFp 43103
+INGE0LjQtw== 43104
+IG9wcw== 43105
+4bqhdA== 43106
+IHBhcmFjaHV0ZQ== 43107
+IG5lbGw= 43108
+IEVudHNjaGVpZHVuZw== 43109
+15zXmded 43110
+IHRydXRoZnVs 43111
+IHNoYXJwZXI= 43112
+IGJ1cmVhdWNyYWN5 43113
+Y2FydA== 43114
+INC40L3Rgg== 43115
+d2llaw== 43116
+IHdpbGxpbmdseQ== 43117
+IEhlcm1hbg== 43118
+IG1laHJlcmU= 43119
+IGVsaXRlcw== 43120
+IEFybW9y 43121
+IGVtYm9yYQ== 43122
+IFJlY29nbg== 43123
+INC70Y7QsdC70Y4= 43124
+IEV4Y2VsbGVuY2U= 43125
+aWJlbA== 43126
+IGV4cG9ydGluZw== 43127
+7LK07KCB 43128
+S2VsbHk= 43129
+Q2FtZXJhbWFu 43130
+IHNsaXBz 43131
+IGZpZ3VyYQ== 43132
+IGtvbGw= 43133
+IFBhbmRlbWll 43134
+IHRpbWVk 43135
+bGllw59saWNo 43136
+INee15s= 43137
+IHBlcsOtb2Rv 43138
+aXZhdA== 43139
+IHF1ZXN0aW9ubmFpcmU= 43140
+IHDDqXJpb2Rl 43141
+IHNpZ2hz 43142
+IGFsbGVnaWFuY2U= 43143
+IFhW 43144
+IEtlbnN1a2U= 43145
+IEdlc3VuZGhlaXRz 43146
+IHBvc2l0aXZv 43147
+IEphbmVpcm8= 43148
+IFNFRQ== 43149
+INin2LPYqg== 43150
+IEtlbHNleQ== 43151
+dG9iZXI= 43152
+IM6xzrvOu86s 43153
+IFBhcmVudA== 43154
+IERheXRvbg== 43155
+IEJpbGRlcg== 43156
+b3VyYWdl 43157
+IHNlcmVz 43158
+IG11Y2jDrXNpbW8= 43159
+IFJlYWxt 43160
+IE9GRklDRVI= 43161
+ZXJzb25pYw== 43162
+b255YQ== 43163
+IOq4iQ== 43164
+IGFuY2VzdHJ5 43165
+IEp1cmFzc2lj 43166
+IGNlbnRpZ3JhZGU= 43167
+4bqldQ== 43168
+dWrEhWM= 43169
+bWFucw== 43170
+IHRpbw== 43171
+IE1vxbw= 43172
+IHRyYWdlbg== 43173
+IHN0YXJlZA== 43174
+IHNjaGVtYXRpYw== 43175
+IHBhc3NvdQ== 43176
+IG1lYXRiYWxscw== 43177
+xYJvxZvEhw== 43178
+IHN5bmNocm9ub3Vz 43179
+IHBlcm1pcw== 43180
+YXJpYWw= 43181
+IHplcg== 43182
+IHBhcml0eQ== 43183
+IEF2YXRhcg== 43184
+aW5kZWVy 43185
+ZXN0b24= 43186
+IG1laWTDpG4= 43187
+IENseQ== 43188
+tIk= 43189
+IGVzdHJvZ2Vu 43190
+IGNlbnRpbWV0 43191
+IGNvbnZpY3Rpb25z 43192
+IHBvc3NpYW1v 43193
+IHBlcmR1 43194
+IHBhdGhvZ2Vucw== 43195
+IFF1aW4= 43196
+IFByb2dyYW1z 43197
+IFBvaW50cw== 43198
+cmFtZW50 43199
+cmFpbA== 43200
+IHZ5 43201
+IGdyYWZ0 43202
+IGJhcnQ= 43203
+IExvdHVz 43204
+4Kg= 43205
+IOuztOyLnA== 43206
+cmFtZXI= 43207
+RmF0aGVy 43208
+IOucuw== 43209
+INeU150= 43210
+IHRyYXplcg== 43211
+IHRhcms= 43212
+w6hjZXM= 43213
+Zm9ydGg= 43214
+INGB0LTQtdC70LDQu9C4 43215
+IHp1Y2NoaW5p 43216
+IHdha3R1 43217
+IGVudGVydGFpbmVk 43218
+IE1pbGxpYXJkZW4= 43219
+IHNoYWt5 43220
+IHByemVkZQ== 43221
+uIzr 43222
+IHJldmVyc2libGU= 43223
+IE5BVQ== 43224
+dWlucw== 43225
+w6lyw6p0 43226
+YW5uZW4= 43227
+IEh1bnRpbmc= 43228
+IEZlbGxvdw== 43229
+w6lsaW9y 43230
+IHJvdGF0aW9ucw== 43231
+IGdyYW5ueQ== 43232
+eHRvbg== 43233
+INGB0YLQsNC90L7QstC40YLRgdGP 43234
+INC90LDRh9Cw0Ls= 43235
+IGFydGVyaWVz 43236
+cmnDsw== 43237
+INC/0L7Qu9GM0LfQvtCy 43238
+INCR0Ys= 43239
+IG5vdmVsdHk= 43240
+cG91bmQ= 43241
+IHdlaXJkZXN0 43242
+IGJvaXM= 43243
+w6ltaWU= 43244
+dXBs 43245
+QVRB 43246
+IHRlaGQ= 43247
+IE5pcg== 43248
+c8SxbsSxeg== 43249
+ISIs 43250
+IGltbW9ydA== 43251
+IGVsaw== 43252
+0LDQvdC40Yc= 43253
+IGZhYnJpY2F0aW9u 43254
+IE5vaXNl 43255
+IEF2YW50 43256
+2LHbjA== 43257
+d2F0 43258
+IHdob29zaGluZw== 43259
+INeb15k= 43260
+INCX0L3QsNGH0LjRgg== 43261
+IGNlbnRyaWY= 43262
+YW5zaW5n 43263
+U291bmQ= 43264
+IOudvOs= 43265
+IGNhcHRpb25z 43266
+4LON 43267
+IG9yZ2Fz 43268
+IGRvbHBoaW5z 43269
+IEJsZW5k 43270
+IFRhag== 43271
+IENDVFY= 43272
+IGlub20= 43273
+IGVkaXRpb25z 43274
+IGJ1cm5vdXQ= 43275
+IGLDpHR0cmU= 43276
+IENhc2E= 43277
+b3ZpY2g= 43278
+IG1vbHRlbg== 43279
+IGJsaW5kZm9sZA== 43280
+IEd1ZQ== 43281
+IHNwaW5uZXI= 43282
+IG3DtmdsaWNoc3Q= 43283
+IFbDoA== 43284
+ZW5lY2E= 43285
+IG3DqWRpY28= 43286
+w6FzdGljbw== 43287
+IGFyZA== 43288
+IFN1bmRheXM= 43289
+IFJlbW90ZQ== 43290
+IOyWvOuniA== 43291
+IHRyxrDhu5tj 43292
+7IWo6w== 43293
+IGRvcHA= 43294
+IGJlxJ8= 43295
+aWNhbmE= 43296
+IOuCmOykkeyXkA== 43297
+IGhvbGluZXNz 43298
+ZGlyZWN0 43299
+IOyYge2ZlA== 43300
+IGN1bHBh 43301
+IFN0aXRjaA== 43302
+bGlnaHRseQ== 43303
+0LDQvNC10L0= 43304
+INC80LXRiA== 43305
+INC/0LXRhw== 43306
+IHlodGU= 43307
+b3NwaGVyZQ== 43308
+IOyTsOuKlA== 43309
+w6lr 43310
+IHNlcmlvdXNuZXNz 43311
+IGdhcm1lbnRz 43312
+IGNvbmNpc2U= 43313
+IFNK 43314
+IHZlcmxvcmVu 43315
+IHBhcmVjZXI= 43316
+IFVOQw== 43317
+7Iqk7YOA 43318
+IGVuZmFudA== 43319
+IGJvbWJlcg== 43320
+IEdpZnQ= 43321
+IOyii+uLpA== 43322
+IHJoeXRobXM= 43323
+IEtsYXI= 43324
+b3duaWs= 43325
+IFJldmVyZW5k 43326
+IGVtaXR0ZWQ= 43327
+bGFzc2Vu 43328
+IHJldmVuaXI= 43329
+IGFyaXNpbmc= 43330
+IHByZWNpc2FtZW50ZQ== 43331
+IGludGVycG9s 43332
+IFRlbmVtb3M= 43333
+b2JlZA== 43334
+IHRlY25vbG9naWE= 43335
+IG5lcmVkZQ== 43336
+IFZpc2E= 43337
+IHNhdmE= 43338
+IGVzY3JldmVy 43339
+IGFzc2F1bHRlZA== 43340
+IEZsZWlzY2g= 43341
+IENvdW5jaWxsb3Jz 43342
+IOqwgOq5jA== 43343
+IGJlZ2c= 43344
+IERldmVsb3Blcg== 43345
+IEJyb256ZQ== 43346
+IEJvbnVz 43347
+INeo16c= 43348
+ZmFjdA== 43349
+IGVuZGxlc3NseQ== 43350
+IG1hY2Ft 43351
+IHJ6ZWN6eXdpxZtjaWU= 43352
+IGhvdmVyaW5n 43353
+w6hnZQ== 43354
+IHBvb3Jlc3Q= 43355
+IFNjaGVk 43356
+bWlsZQ== 43357
+aXNzZW1lbnRz 43358
+YWPEgw== 43359
+IOumvQ== 43360
+IHZhY2Npbg== 43361
+IGZ1dHVyaXN0aWM= 43362
+IFdpbmRvdw== 43363
+0L/QsNGA 43364
+INGA0L7RgQ== 43365
+IGxvd2Vycw== 43366
+YWNz 43367
+INCQ0LvQtdC60YHQsNC90LQ= 43368
+IEFsZXJ0 43369
+aWVtZQ== 43370
+IENhdWNhcw== 43371
+IGphd3M= 43372
+IGh1bnRlZA== 43373
+7Je9 43374
+INio2YY= 43375
+INec16DXlQ== 43376
+IHR1cmJpbmVz 43377
+IGx1bXBz 43378
+IEFsbGllcw== 43379
+YWhsdA== 43380
+IHN1YnNjcmlwdGlvbnM= 43381
+IG5vdXZlYXV4 43382
+dWdlcg== 43383
+Ym9uZXM= 43384
+IGJlcnJ5 43385
+IOyEoOusvA== 43386
+IE1hbnVmYWN0 43387
+IEx1bmNo 43388
+6re4656Y 43389
+IGh5ZHJhdGVk 43390
+IGFjaGVp 43391
+IFlheg== 43392
+IFRpYmV0YW4= 43393
+IFF1YW50dW0= 43394
+IEplcm9tZQ== 43395
+INC+0YnRg9GJ 43396
+0L7QstCw0L0= 43397
+bW90aW9u 43398
+IENvbnRyb2xsZXI= 43399
+ZW5lcmdldGlj 43400
+INGB0LrQvtGA0L4= 43401
+IHZvd2Vscw== 43402
+INGD0LbQsNGB 43403
+IGhvb2Y= 43404
+IEJ1bGxldA== 43405
+aW1hZ2lu 43406
+16DXmded 43407
+IGVuZ2FnZW1lbnRz 43408
+IEJsdWVz 43409
+IGHDsWFk 43410
+IGZwcw== 43411
+IGNhdGVycA== 43412
+IHPhu5E= 43413
+IFRyaWJl 43414
+0L/QvtC9 43415
+aWZlcmF0aW9u 43416
+IHJ1bWFo 43417
+IFB1bmo= 43418
+bGFi 43419
+IGNvbXByZWhlbnNpb24= 43420
+YnJpbmdpbmc= 43421
+V28= 43422
+IHRpaw== 43423
+IGFueWhvdw== 43424
+w6F0aWNhcw== 43425
+IHNpdHplbg== 43426
+IGtvbGF5 43427
+IENvbmZlZGVyYXRl 43428
+IENhbGxlZA== 43429
+IG5hc3p5Y2g= 43430
+IGR6acSZa2k= 43431
+IGNsb2Fr 43432
+IEdvb2c= 43433
+IEFzaGU= 43434
+ZW5hbg== 43435
+INC80YvRiA== 43436
+INCy0LXRgg== 43437
+IFNwbw== 43438
+IFNrZXQ= 43439
+IEhlbmRlcnNvbg== 43440
+aWxhaA== 43441
+INCx0LXQt9C+0L/QsNGB 43442
+IHNla2FsaQ== 43443
+7Ja06rCA 43444
+IHNuYXJl 43445
+IHLhurFuZw== 43446
+IGbDtnJzw7Y= 43447
+c3p5Y2g= 43448
+IMO8YmVycw== 43449
+IHN0cmF0w6ln 43450
+IOy6kOs= 43451
+IHJhcHBlcnM= 43452
+IGNlcA== 43453
+IEhhc3Rh 43454
+IGhvcnJpYmx5 43455
+IGZyw7xo 43456
+INio2Lk= 43457
+IG1hbnRsZQ== 43458
+44CF 43459
+ZnVuZGluZw== 43460
+IHp1c3Q= 43461
+IFBlbnM= 43462
+c2Vk 43463
+IO2XpA== 43464
+IGdlcmVraQ== 43465
+IGFsYXJtcw== 43466
+IFdoYQ== 43467
+IE1hcmt1cw== 43468
+YWtzaQ== 43469
+INCQ0LvQtQ== 43470
+a2xvcmU= 43471
+IMOpbmVy 43472
+IHRpbGRl 43473
+Ym94aW5n 43474
+IOyEng== 43475
+IGVuY29udHJhbW9z 43476
+IFBoYXI= 43477
+0L3QsNC60L7QvA== 43478
+w7NzdA== 43479
+IMSwcw== 43480
+IOuLmA== 43481
+IHNxdWF0cw== 43482
+IHByZXRlbmRlZA== 43483
+IGRleg== 43484
+IOq0nOywruyVhA== 43485
+amFjaA== 43486
+65286rOg 43487
+IO2ZleynhA== 43488
+IEFuc2No 43489
+aW1lcms= 43490
+IGNvbmp1Z2F0ZQ== 43491
+IHBlbmluc3VsYQ== 43492
+IGdvcmlsbGE= 43493
+IHBob3RvZ3JhcGhlZA== 43494
+IEF1bnF1ZQ== 43495
+IGVudHJlbg== 43496
+IERldXRzY2hlbg== 43497
+IEFsYWRkaW4= 43498
+IOustOyEnA== 43499
+IFN0ZWxsYQ== 43500
+IEVsZWN0aW9u 43501
+b3V0aW5l 43502
+R3JhbmQ= 43503
+IFdhaw== 43504
+IFNlcmdpbw== 43505
+aG9yc2U= 43506
+YWhvbg== 43507
+IEZhbWlsaWVz 43508
+IGhhdGluZw== 43509
+IEJldHQ= 43510
+4LiZ4Liw4LiE4Liw 43511
+IGN1cmxpbmc= 43512
+IElzcmFlbGlz 43513
+INec15DX 43514
+IE15ZXJz 43515
+IHNjYW5uZWQ= 43516
+IEJFQw== 43517
+aWxlcmk= 43518
+IGNhbGxl 43519
+IE1pbmg= 43520
+IG1pY3Jvbg== 43521
+IGNvbmR1Yw== 43522
+w612 43523
+INCy0L7Qt9GM 43524
+IGFjdGlvbmFibGU= 43525
+IFRydXN0ZWVz 43526
+IHRpZWY= 43527
+IGhlYWRlcnM= 43528
+IGFuaW1hbGVz 43529
+7JuA 43530
+0LvQvtGF 43531
+dW5pdHk= 43532
+bHlh 43533
+IGphbmdhbg== 43534
+IGhhbmk= 43535
+IGNhc2luZw== 43536
+IGrDs3ZlbmVz 43537
+IFNwbGl0 43538
+IENhcmxv 43539
+IEJlaW0= 43540
+IG51YW5jZWQ= 43541
+IHRlZGR5 43542
+IENsYW4= 43543
+w6RjaGVu 43544
+cGllcg== 43545
+INC00L7Qv9C+0LvQvQ== 43546
+IGRpYXBlcg== 43547
+ZWZmZWN0aXZl 43548
+IE5pYWdhcmE= 43549
+IHdhcnQ= 43550
+IGNvcnJv 43551
+IEthbXBm 43552
+enRl 43553
+IGTDqXZlbG9wcGVtZW50 43554
+IGF0dGFja2Vycw== 43555
+IFNoZXJtYW4= 43556
+IDE5MTQ= 43557
+IG1lb3c= 43558
+IFDDpQ== 43559
+7Lo= 43560
+Y2l0 43561
+IGNvdXBl 43562
+IOq3uOuLpOydjOyXkA== 43563
+IGh1bW91cg== 43564
+IGNvbGU= 43565
+IFdhcm5pbmc= 43566
+IFRpbA== 43567
+Y2FsbQ== 43568
+YnVhdA== 43569
+IGNpbmU= 43570
+a2llag== 43571
+S2V2aW4= 43572
+IG1pbGxpZ3JhbXM= 43573
+15PXqA== 43574
+YXJpYW1lbnRl 43575
+IG9ybw== 43576
+IEhvZA== 43577
+ZXJ0b3M= 43578
+IGxpaGF0 43579
+IGZ1bGxlc3Q= 43580
+IGdyYW5kaQ== 43581
+INCx0L7Qug== 43582
+IHdob2xseQ== 43583
+IG1haGRvbGw= 43584
+IGNvbnRyb2xs 43585
+IEJ1bnVu 43586
+IGRpcHBlZA== 43587
+IHJlZ2nDs24= 43588
+INmE2Yg= 43589
+INCx0LDQsw== 43590
+IHByZW1pZXJz 43591
+IGNo4buL 43592
+aWRleg== 43593
+IHF1b3Rh 43594
+IGdoZWU= 43595
+YXJrYW4= 43596
+IGdlbGF0aW4= 43597
+IENsZXJr 43598
+YmJsZXM= 43599
+IFBhaWdl 43600
+IHN0YWdlZA== 43601
+IHNvY2lhaXM= 43602
+IEJpemlt 43603
+IHZlbG9jaWRhZGU= 43604
+IG1hbGFyaWE= 43605
+IHNob3J0ZW5lZA== 43606
+IHNhbHV0 43607
+IEhlaGU= 43608
+IHbhu4s= 43609
+IFRhaXdhbmVzZQ== 43610
+IEFycmk= 43611
+Z3Jlcw== 43612
+KCk= 43613
+cmlhZA== 43614
+kZDr 43615
+IG1hc2N1bGluaXR5 43616
+TFA= 43617
+IOuWoQ== 43618
+IHTDqXJtaW4= 43619
+IFbDpA== 43620
+IFNlaXRlbg== 43621
+IHJlc3BlY3RmdWxseQ== 43622
+w6Fv 43623
+IHRvdGFsZW1lbnQ= 43624
+IHNjcmFwcw== 43625
+IGluZnJpbmc= 43626
+IEJvc2U= 43627
+YW1hcg== 43628
+IEx1aXph 43629
+IEFSTQ== 43630
+INC/0LvQvtGF0L4= 43631
+IG1laWxsw6Q= 43632
+IERpb24= 43633
+IHNvdWhh 43634
+IGdlc2NoYWZmdA== 43635
+IGNvbnZvbHV0aW9u 43636
+IOKAkeKAkQ== 43637
+IDE0NA== 43638
+bGluZ3Q= 43639
+IG3DpG5uaXNr 43640
+IGd1c3RhZG8= 43641
+IGNvaW5lZA== 43642
+IEx1bHU= 43643
+b3BvdA== 43644
+IFByYXllcg== 43645
+IHJvYXN0aW5n 43646
+IGNocm9tb3NvbWVz 43647
+0LXQu9C1 43648
+Qmx1ZQ== 43649
+IEVyZm9sZw== 43650
+INC/0YDQuNC00YPQvA== 43651
+IHJpc2tpbmc= 43652
+IEd1YXJkaWFucw== 43653
+IDIwMjQ= 43654
+w6hzZQ== 43655
+INCx0YPQtNGC0L4= 43656
+IGNvbnNlcnZl 43657
+IEJyaW5naW5n 43658
+IEFzdHJh 43659
+4LmA4LiC 43660
+INC60LDQutGD0Y4= 43661
+cmVzcGFjZQ== 43662
+INCe0L8= 43663
+INCy0L7QutGA0YPQsw== 43664
+IG1hc2tlZA== 43665
+IFNoeQ== 43666
+IE5pbQ== 43667
+ZW5kYXM= 43668
+IO2PrOyduA== 43669
+IOuqqOyWkQ== 43670
+IHZhbGV1cg== 43671
+IE5lZ3Jv 43672
+IENEcw== 43673
+aW5rbGluZw== 43674
+IG1vbnTDs24= 43675
+IEhvbmQ= 43676
+UmVhbA== 43677
+IGZ1bGxuZXNz 43678
+IFdob29wcw== 43679
+IFNoYW5r 43680
+IEJyYW4= 43681
+IHRyYW5zbHVj 43682
+IGVycg== 43683
+IEdhcmRlbnM= 43684
+b3l1 43685
+IGFmZmlybWF0aXZl 43686
+IHBvdHRlcnk= 43687
+bGl2ZQ== 43688
+aWF1 43689
+bW91bnQ= 43690
+IGZsdWN0dWF0aW9ucw== 43691
+w61lbQ== 43692
+IHB1bHNlcw== 43693
+IGNyaWFuw6dhcw== 43694
+zq/Osc+C 43695
+IGJhc3Rh 43696
+RU5OSVM= 43697
+INC60L7RgNC/ 43698
+IEZ1bms= 43699
+w6VydA== 43700
+INC30LDRgtC10Lw= 43701
+IHBhcmFzaXRlcw== 43702
+IGFpcmZsb3c= 43703
+IFh1YW4= 43704
+R8O8bG1l 43705
+IGJsb29taW5n 43706
+IG11bW15 43707
+IGJhbw== 43708
+IENsYXA= 43709
+YW50aWNz 43710
+c2tpbg== 43711
+Y2VudHJpYw== 43712
+YmVmb3Jl 43713
+IFJJQ0hBUkQ= 43714
+IEhhaG4= 43715
+VEFLRQ== 43716
+INGC0YDQtdGC0Yw= 43717
+IHByZXNzdXJlZA== 43718
+IEt1cno= 43719
+aXN0aQ== 43720
+INC90LDRiNC10LPQvg== 43721
+IHNlbWljb25kdWN0b3I= 43722
+IENsaW50 43723
+IHBsdXA= 43724
+IE9yaWdpbg== 43725
+IEV2ZW50cw== 43726
+IOqxseyglQ== 43727
+bXBmZW4= 43728
+TkVZ 43729
+IERX 43730
+IOu2ge2VnA== 43731
+IGluZm9ybXM= 43732
+IGZvcnNr 43733
+IGFtaWdh 43734
+IENpbmNpbm4= 43735
+U3Ry 43736
+IHBhcmlzaA== 43737
+IOy7pO2U 43738
+IHNpemk= 43739
+IHBsYW50YXRpb24= 43740
+IGJsaXZlcg== 43741
+INC/0L7Qu9C40YI= 43742
+IHN1YmRpdg== 43743
+IHJhbnQ= 43744
+IHByaW5jaXBhbHM= 43745
+IGt1bm5l 43746
+w7xnZW4= 43747
+YXJlc3BhY2U= 43748
+IHZhbGxhaGk= 43749
+IGNvbGxhcHNpbmc= 43750
+2KfZhNmF 43751
+IGxpZGVy 43752
+IHRhbWE= 43753
+IGdhZ25lcg== 43754
+cm9sbGU= 43755
+IOunkOyUgOuTnOs= 43756
+IGNhdGhlZHJhbA== 43757
+IFdlYnM= 43758
+IFBvbGl0aWNz 43759
+IERlbmlz 43760
+IHR1bw== 43761
+IHJlZnJhY3Q= 43762
+IGRpc2ludGVncg== 43763
+c3Rlcw== 43764
+INC70Y7QsdC+0LI= 43765
+IHdpbHQ= 43766
+IHRydXN0cw== 43767
+IGtvbXVu 43768
+IEJhc2tldA== 43769
+fiEh 43770
+bmFl 43771
+INCa0L7Quw== 43772
+IHN5bGxhYmxlcw== 43773
+IEhlbnJp 43774
+IE5hYg== 43775
+2YjYuQ== 43776
+IHdu 43777
+IGthbXA= 43778
+IFByYWd1ZQ== 43779
+IEJyZWFrZmFzdA== 43780
+IOq3uOuftA== 43781
+IGNodXQ= 43782
+IDMzMA== 43783
+IEluZHVzdHJpZXM= 43784
+IGnFn2k= 43785
+IEdvbGRtYW4= 43786
+IMSwbnM= 43787
+dXNzYQ== 43788
+aXRoZQ== 43789
+hJA= 43790
+IFNPVU5E 43791
+0LDQu9GM0L3Ri9C8 43792
+Lig= 43793
+INCz0L7RgNCw0Lc= 43794
+IGRhZ2VnZW4= 43795
+IOuu 43796
+IHdhaXRlcg== 43797
+bGVuZ3Ro 43798
+IM+Dz4TOsQ== 43799
+IGNodW5reQ== 43800
+U2E= 43801
+IHJ1c3R5 43802
+IEp1ZGl0aA== 43803
+NzUw 43804
+IGVwb3h5 43805
+7Lmg 43806
+bWV0cm8= 43807
+IHJlamVjdGluZw== 43808
+IHNxdWlzaHk= 43809
+IHBsdXBhcnQ= 43810
+IG3DqXRo 43811
+IGFzcGlyaW5n 43812
+IERyYW1h 43813
+IHVwbGlmdA== 43814
+p4jri6Q= 43815
+Li4uLi4uLi4uLi4uLi4uLg== 43816
+oKTsmpQ= 43817
+IHTDqWNuaWNh 43818
+IHBhc2FuZG8= 43819
+VGhvc2U= 43820
+INGA0LDQt9C00LXQuw== 43821
+IG1lZGlvY3Jl 43822
+IE5pY2tlbA== 43823
+IHN1cGVyaGVyb2Vz 43824
+IG1pc3Npb25hcnk= 43825
+IFBhcmVjZQ== 43826
+IHJvdGF0aW9uYWw= 43827
+IHByZXR0 43828
+IGxhbWE= 43829
+IGNhbnlvbg== 43830
+IGJldGVy 43831
+IFByb3Zvc3Q= 43832
+IGh2aXM= 43833
+IGRlYWN0aXY= 43834
+IEhlbHM= 43835
+cGZsaWNodA== 43836
+U29tZXRoaW5n 43837
+IFBpZXJjZQ== 43838
+IOqygOywsA== 43839
+bHVuZ2Vu 43840
+IHNpemluZw== 43841
+IGxhdGl0dWRl 43842
+IE5vbmV0aGVsZXNz 43843
+b21uaWE= 43844
+IFNhYnJpbmE= 43845
+IER5bmFtaWM= 43846
+b250YQ== 43847
+7IaQ 43848
+IGRpcmVjdGl2ZQ== 43849
+IERlcG90 43850
+IGZ1ZWxlZA== 43851
+IGV4cGlyZQ== 43852
+IGNvbcO6bg== 43853
+IFNleHVhbA== 43854
+IEdvcmU= 43855
+IHJlc3RsZXNz 43856
+IEpBS0U= 43857
+0YLQtdGA0LXRgQ== 43858
+INGC0YDQsNC9 43859
+IEhvbHo= 43860
+IEFjdG9y 43861
+Y2FsbA== 43862
+IGVtYWlsZWQ= 43863
+IFBlYXI= 43864
+0YPQtNC4 43865
+0YDQsNC7 43866
+IG3DoHk= 43867
+IENIRUVSSU5H 43868
+IHJldGFpbGVy 43869
+IHByb3Ry 43870
+IGRpc2NhcmRlZA== 43871
+IEhJUw== 43872
+IGV2YW5nZWxpY2Fs 43873
+IEVsc2U= 43874
+IGV4cGxvcmVz 43875
+IGNyaXRpY2l6aW5n 43876
+aWZpaw== 43877
+IHdoaXBwaW5n 43878
+IG9waXM= 43879
+b3VzZWQ= 43880
+RnJlZQ== 43881
+IO2MrA== 43882
+IG1pY3M= 43883
+cnVubmluZw== 43884
+T2I= 43885
+aXRpw6k= 43886
+IG5lY2VzaXRh 43887
+IERvbWluaWNhbg== 43888
+IEJhZ2g= 43889
+IHRlbmRlbmNpZXM= 43890
+IE1ldHJvcG9saXRhbg== 43891
+xZFs 43892
+INC30L3QsNC10Lw= 43893
+IFphbQ== 43894
+IERlYWRwb29s 43895
+YWxlxbw= 43896
+IGludmVzdGlnYXRpdmU= 43897
+IFByb251bmNpYXRpb24= 43898
+IGVtdWxhdGU= 43899
+IGJhbmNv 43900
+IC3imao= 43901
+IG92ZXJhcmNoaW5n 43902
+bGljaGVz 43903
+INCy0L7Qt9Cy0YDQsNGJ 43904
+IFNjYXJ5 43905
+IEtpYQ== 43906
+cm9udGluZw== 43907
+aW5uZWQ= 43908
+INuB2Yg= 43909
+7IiY66W8 43910
+d2Vs 43911
+IOuzhOuhnA== 43912
+IHVuaW50ZW50aW9u 43913
+YWFT 43914
+IG5pY2VzdA== 43915
+IFRlc3Rpbmc= 43916
+IElTSUw= 43917
+b2dlbm91cw== 43918
+INif 43919
+IGxpZXV0ZW5hbnQ= 43920
+IGJyYXVjaA== 43921
+IFRpcg== 43922
+ZHJpdmU= 43923
+IHRvbGVyYW50 43924
+IHNob290ZXJz 43925
+IOyYiOu7kA== 43926
+b250b24= 43927
+IHRlcmlh 43928
+aWV0ZXQ= 43929
+Um9u 43930
+bGVpZ2g= 43931
+Z2Fl 43932
+IG9sbWFr 43933
+IENsb25l 43934
+c29sZA== 43935
+IHNrZWxldG9ucw== 43936
+IGluY3VtYmVudA== 43937
+0L7QvNC1 43938
+Q09O 43939
+IGxldmVu 43940
+IG1pbGxlbm5pYWxz 43941
+IGVxdWF0b3I= 43942
+IEZlZGVy 43943
+IEFsZXhhbmRyYQ== 43944
+IHZyaWo= 43945
+IEhlYWx0aGNhcmU= 43946
+IO2VkQ== 43947
+IGVtcGhhc2l6aW5n 43948
+IGRpYWxvZ3Vlcw== 43949
+IGNoaWxsZWQ= 43950
+IHByb3c= 43951
+IFBhc3Npb24= 43952
+IExhZGVu 43953
+YXJpZXN0 43954
+YXBocmFn 43955
+IGFkZGl0aXZl 43956
+IFN0YWF0 43957
+IE5lcHQ= 43958
+IEhBTQ== 43959
+4LmA4Lit 43960
+ZGF5cw== 43961
+IO2WiOuNmA== 43962
+IHZvaWxh 43963
+INGF0Ls= 43964
+IERldXRzY2hl 43965
+cXVpcg== 43966
+T3Blbg== 43967
+IHJhbmdlZA== 43968
+IGxldmVycw== 43969
+IE1hbnNpb24= 43970
+cGFyZWQ= 43971
+IFRpdGFucw== 43972
+YXRvaXJl 43973
+IGVuZ2FnZXM= 43974
+eWV6 43975
+bmFkZW4= 43976
+IG9ic3RydWN0 43977
+IEVtbXk= 43978
+sKU= 43979
+IHRyb3Bo 43980
+IHRha2Vhd2F5cw== 43981
+Ky4= 43982
+dHljem5pZQ== 43983
+aMOpc2l0ZXo= 43984
+IHBvZMOtYQ== 43985
+IOyjvOuKlA== 43986
+IGNpdGF0aW9u 43987
+IEFxdWE= 43988
+IGRlYnVnZ2luZw== 43989
+0LLQsNC9 43990
+IOuLueyLoA== 43991
+INin2YTZig== 43992
+IGluc3RhbnRhbmVvdXM= 43993
+IEF1dHVtbg== 43994
+IGtlcGFkYQ== 43995
+IGdldGFu 43996
+aGluaQ== 43997
+eW50aGVzaXM= 43998
+INC/0LXRgNC4 43999
+IE1hY2Vk 44000
+UGFj 44001
+dW50dQ== 44002
+QnJh 44003
+INCz0L7RgNCw0LfQtNC+ 44004
+IDE5NTk= 44005
+INGC0LXQvNC/0LXRgA== 44006
+IHNhbmU= 44007
+IE9VUg== 44008
+YXN1 44009
+IOustOyX 44010
+IHZhbGxleXM= 44011
+IGxpc3Rpbmdz 44012
+IHByemVkc3Rhdw== 44013
+IGd1bW15 44014
+IGNvcnRpc29s 44015
+IE9icmln 44016
+IEFsbGllZA== 44017
+0L7QttGD 44018
+IGfDqW7DqXI= 44019
+IGRvY3M= 44020
+IENoaWxp 44021
+IEFiZHVsbGFo 44022
+S2l0 44023
+IGNvbnRyaWJ1dG9ycw== 44024
+0LPQvtGA 44025
+0LvQtdGA 44026
+IGJpbmRlcg== 44027
+IG1vZMOobGU= 44028
+7YWQ 44029
+IGludGVpcm8= 44030
+bWlz 44031
+ZmVyYQ== 44032
+2KfYsA== 44033
+TWFuaWE= 44034
+IO2ZnOuPmQ== 44035
+IOu0kOyalA== 44036
+IEpheg== 44037
+0ZbQu9GM0LrQuA== 44038
+cmlzaG5h 44039
+IOq1sA== 44040
+IHRhbWFuaG8= 44041
+IGFwcGxpYW5jZQ== 44042
+IFJlc2lzdGFuY2U= 44043
+IExPT0s= 44044
+IEh5cA== 44045
+IEhlaWw= 44046
+RmlyZQ== 44047
+dWp1 44048
+IGhlYWxz 44049
+IG1hbHQ= 44050
+IFZFUlk= 44051
+INGF0L7Rh9C10YjRjA== 44052
+IGxpbmdlcg== 44053
+IE5hcnI= 44054
+IFJlZ3VsYXI= 44055
+IExvb3A= 44056
+IExlbm8= 44057
+IHNvcnRpZQ== 44058
+IFNlcnZl 44059
+IOydtQ== 44060
+IEx1ZWdv 44061
+aXR0w6Q= 44062
+IHVuZGVz 44063
+IHNsaXBwZXJz 44064
+IG9uZGE= 44065
+IMSQw6J5 44066
+IHRhcGVk 44067
+IHRyYXZlcnNl 44068
+IHJlbGF0aXZpdHk= 44069
+IFlvc2hp 44070
+Y2pvbg== 44071
+aWxhdGVk 44072
+YWN0aXZlbHk= 44073
+INCh0L7Qsg== 44074
+IFBPTA== 44075
+0KDQmA== 44076
+aW5mbGFtbQ== 44077
+Y2hlZXJmdWw= 44078
+INee15DX 44079
+ID4+Ww== 44080
+bWluc3Rlcg== 44081
+INCy0LvQuA== 44082
+IGlkZW50aWZpZXI= 44083
+IExhbWJkYQ== 44084
+IHRyb3M= 44085
+IGZsYXdsZXNz 44086
+IGRldHJpbWVudGFs 44087
+IGJ1bmxhcsSx 44088
+V2Fy 44089
+IHJlZ2nDo28= 44090
+IEJpa2U= 44091
+Y2Vzc29ycw== 44092
+IGPDuW5n 44093
+IFJO 44094
+IOq9gw== 44095
+IGvDvMOnw7xr 44096
+IEJlZ2lubmluZw== 44097
+7Zi46w== 44098
+IGdld2U= 44099
+IGRlbm90ZQ== 44100
+IEFsYmVydG8= 44101
+IHByb2Jpb3Q= 44102
+IG9kZQ== 44103
+IG1vbGFy 44104
+IGJ1cnN0aW5n 44105
+YXNzdW1lZA== 44106
+IGZvb3RwcmludHM= 44107
+dmVkYQ== 44108
+IHN0ZXJvaWRz 44109
+IGZsYW1pbmc= 44110
+IEVsbGVy 44111
+IGVya2VubmVu 44112
+w6R0emVu 44113
+IGxpZmVjeWNsZQ== 44114
+IERPVQ== 44115
+IEthcmVuYQ== 44116
+IEd1ZXJyYQ== 44117
+IHNpbmlzdGVy 44118
+IHBvZMOpaXM= 44119
+IHBhcmFi 44120
+IG9rbw== 44121
+IG1hdMOpcmk= 44122
+IGNhcmlj 44123
+c29uYXJv 44124
+IHByYXRpY2FtZW50ZQ== 44125
+0YPRgdCw 44126
+IGNvbXVucXVl 44127
+IHZpZ2lsYW50 44128
+IHJlZ2ltZXM= 44129
+IFNob290aW5n 44130
+IHJhaWRz 44131
+IE5vcmE= 44132
+IFdpZWRlcg== 44133
+bWVucw== 44134
+INGB0L7QtA== 44135
+IOqyveyasOyXkOuKlA== 44136
+INCy0YXQvtC0 44137
+IGF1dG9iaQ== 44138
+IFNjaG4= 44139
+IFJvYmJpZQ== 44140
+IEZpdG5lc3M= 44141
+INC60L7QvdGE 44142
+IHBlbmd1aW4= 44143
+0LzQvtGC0YDRjw== 44144
+INC80LjQvdC40Lw= 44145
+cGxheXM= 44146
+IGRlbGVnYXRlcw== 44147
+TWVy 44148
+IHNpc3RlbQ== 44149
+IE1pY2hhZWxz 44150
+bWFsZQ== 44151
+2KfYuQ== 44152
+IGPDoWNo 44153
+IEjDpA== 44154
+INeZ15XXk9ei 44155
+IHN1cGVycG93ZXI= 44156
+IHN0cm9u 44157
+IHJvdmVy 44158
+IGTDqXBlbmQ= 44159
+IHJldGlyaW5n 44160
+IHZhbXBpcmVz 44161
+IG1lcmRl 44162
+IENoYW5naW5n 44163
+IHRhbWU= 44164
+IHNwb2tlc3BlcnNvbg== 44165
+IGNheQ== 44166
+IGZsaXJ0aW5n 44167
+IEdyw7Y= 44168
+IHfDpHI= 44169
+IHd5Yg== 44170
+IGNvZXVy 44171
+4bqhbmg= 44172
+IOyZgOyEnA== 44173
+IGNvbm5haXM= 44174
+IEh1bmRyZWRz 44175
+IEJlYQ== 44176
+IM6xz4A= 44177
+cHJ1Y2g= 44178
+IHNvY2llZGFkZQ== 44179
+IFdoaWxzdA== 44180
+IEthaXQ= 44181
+ZXNwYWNl 44182
+IGNoaWE= 44183
+IEVybQ== 44184
+IOuwlOq/ 44185
+IGZlbmNlcw== 44186
+IE1vcnRhbA== 44187
+6rKB 44188
+INCz0YDQsNGE 44189
+IEhvbWVsYW5k 44190
+IEpVTg== 44191
+aXNzdA== 44192
+IHBhcmxhcg== 44193
+IHNwb3J0eQ== 44194
+w6lv 44195
+IGRlZXBlbg== 44196
+IEJlaGF2aW9y 44197
+IGVycmFuZA== 44198
+IHJvdGFyeQ== 44199
+IFdlbGxpbmd0b24= 44200
+V2luZA== 44201
+IG1lc2VsYQ== 44202
+4bqjbmc= 44203
+aWVuZGU= 44204
+IGV4Y2VsbA== 44205
+IEdlbml1cw== 44206
+IEVkdWFyZG8= 44207
+IMWfdW51 44208
+IMSwc3RhbmJ1bA== 44209
+IHByb2R1dG8= 44210
+IOOFjuOFjg== 44211
+T0ZG 44212
+IHdvbGx0 44213
+IOuJtOyKpA== 44214
+IGxhc3M= 44215
+IGhlcnR6 44216
+IGFyb21hdGlj 44217
+INC30LLQvtC9 44218
+IGF1dG9j 44219
+IEx1c3Q= 44220
+IDExMg== 44221
+IM6X 44222
+IHJldmlld2Vycw== 44223
+IHJlY2VwdGl2ZQ== 44224
+w6JuZA== 44225
+b2dsbw== 44226
+IOyVhOuLmQ== 44227
+IG5nbw== 44228
+0ZbRgtC4 44229
+w6V0 44230
+Y29ubw== 44231
+IHRla3Jhcg== 44232
+IOyjvOqzoA== 44233
+IGdlbG1pxZ8= 44234
+IGJlZHRpbWU= 44235
+IEFyZ2g= 44236
+QURB 44237
+INCz0L7RgNC+0LTQsA== 44238
+IMSH 44239
+IGFsbGlhbmNlcw== 44240
+Z2lnZ2xpbmc= 44241
+IHllcmRl 44242
+IHNwaWVz 44243
+IGd1dGVz 44244
+w6dp 44245
+IGFsbHRpZA== 44246
+IExhaA== 44247
+npDr 44248
+IGRva8WCYWQ= 44249
+2YjZig== 44250
+IHRveGljaXR5 44251
+IGNhbmNlbGxhdGlvbg== 44252
+IDE5NTg= 44253
+ZHJv 44254
+IOyekeydgA== 44255
+IE1vdG9yb2xh 44256
+IG11bHRpbg== 44257
+IGVudGh1c2lhc3Rz 44258
+IE1pZ2h0eQ== 44259
+IENvY29udXQ= 44260
+OuOAjA== 44261
+IFBpY3R1cmVz 44262
+IHNhbmdyZQ== 44263
+IGJsaW5raW5n 44264
+b2xlc29tZQ== 44265
+IOyKpO2DgOydvA== 44266
+RlA= 44267
+IGJvb21pbmc= 44268
+INC00LXRgdGP0YI= 44269
+IHJhdGNoZXQ= 44270
+IHRpbWVsaW5lcw== 44271
+bGVuZXNz 44272
+IGNhZ2Vz 44273
+IEdvb2RuaWdodA== 44274
+b21ldGltZXM= 44275
+IGN1bm5pbmc= 44276
+IFJpc2s= 44277
+dWxlZA== 44278
+ZGFkZQ== 44279
+IHByYXRh 44280
+IGd1c3RhcsOtYQ== 44281
+YW11cw== 44282
+IEppbnBpbmc= 44283
+IGVzdHJ1dA== 44284
+IGRlc2NvYnJpcg== 44285
+IE3EgQ== 44286
+IEFsbGFu 44287
+INec16c= 44288
+IHByZXNlcnY= 44289
+IFN0cmF3YmVycnk= 44290
+xI8= 44291
+THU= 44292
+IGtybw== 44293
+IFJlcG9ydHM= 44294
+7IWU7JW8 44295
+IHZhbHQ= 44296
+IHBvdXZhaXQ= 44297
+IGFwcGFy 44298
+IEJvbmU= 44299
+IHByZWZlcmFibHk= 44300
+IFJlcMO6YmxpY2E= 44301
+IGhlcnpsaWNo 44302
+IGNoaW1uZXk= 44303
+IMOnZXY= 44304
+IHZpc2Fz 44305
+IHZlcnI= 44306
+IGN1bHRpdmF0aW9u 44307
+IEFybWVuaWE= 44308
+INCy0LTRgNGD0LM= 44309
+IGNvY2tybw== 44310
+cmV0Y2hlZA== 44311
+YXJ0eg== 44312
+INC70Y7QtNGP0Lw= 44313
+IHBvbMOtdGljYXM= 44314
+IFBhbno= 44315
+IEFLQQ== 44316
+IOuIjOufrA== 44317
+IGVycm8= 44318
+IGNhbXBlcg== 44319
+IDEwMg== 44320
+4KS4 44321
+ZG9uZQ== 44322
+IGhvYXJk 44323
+INCf0L7RgtC+0Lw= 44324
+amVvbmc= 44325
+IGRlc3Rh 44326
+cGFr 44327
+IGluaW0= 44328
+IGdyb3dlcnM= 44329
+IE1lc3NhZ2U= 44330
+IGVsZWN0b3I= 44331
+ZW5nYWdl 44332
+IEZvcmJlcw== 44333
+IENpbmNpbm5hdGk= 44334
+IGRpZmbDqXJlbmNl 44335
+ZGY= 44336
+IHNwYXI= 44337
+IGF3YWl0cw== 44338
+IFVTU1I= 44339
+IFJpc2luZw== 44340
+IEhvxZ8= 44341
+IGZvb3Rpbmc= 44342
+IGNvbmRpY2lvbmVz 44343
+0YLQvtGA0L7Qsg== 44344
+IGNsaW5pY2lhbg== 44345
+IERpc2t1c3M= 44346
+16jXkg== 44347
+16U= 44348
+aXRlaXQ= 44349
+Z3Jlbg== 44350
+IGNoYXJpc21h 44351
+IGxldWtl 44352
+IGlycml0YXRpbmc= 44353
+IGNpcmNh 44354
+IFJob2Rlcw== 44355
+IHBpb3I= 44356
+IGhhbmRpY2Fw 44357
+cm95YWJsZQ== 44358
+IHZ1bGw= 44359
+T0c= 44360
+IGluw61jaW8= 44361
+aWVyaQ== 44362
+IHNwbGFzaGluZw== 44363
+IGRlbWlzZQ== 44364
+IGFzc2lzdGly 44365
+0YfRgtC+ 44366
+IGNvdmVydA== 44367
+IEd1ZA== 44368
+4LiJ 44369
+a2zDpHI= 44370
+IOyekOq+uA== 44371
+IHZlcsOkbmRlcnQ= 44372
+IFJFTQ== 44373
+IENvbnZlbg== 44374
+YXRnZQ== 44375
+IHBpZXJ3c3pl 44376
+IGNsZXJneQ== 44377
+bGluZ3Rvbg== 44378
+bGl2 44379
+VlBO 44380
+INGB0L7QttCw0Ls= 44381
+IEhhdGU= 44382
+z4bOvw== 44383
+IFJlc3BvbnM= 44384
+0L7Qt9C0 44385
+IGV0bWVr 44386
+IGNoZW1pbg== 44387
+2YXYqQ== 44388
+IOqwgOyhsQ== 44389
+VHJl 44390
+IHVtYXM= 44391
+IEJ1cnRvbg== 44392
+IHBhdHJpYXJjaA== 44393
+IFNtaXRoc29uaWFu 44394
+pZg= 44395
+TW9vbg== 44396
+QWly 44397
+IG1lZGlvcw== 44398
+IGVyYXNlcg== 44399
+IHdvbGx0ZW4= 44400
+IHBhcmVpbA== 44401
+IEJpbGxpZQ== 44402
+0LXRgNGC0LI= 44403
+IHBhcmxhbWVudA== 44404
+IGFnb255 44405
+IFFVRQ== 44406
+c2VxdWVudGx5 44407
+QW5vdGhlcg== 44408
+IFdoZXc= 44409
+IEFubnVhbA== 44410
+IHNlYmVu 44411
+7IOB7J2E 44412
+dmFsdWVz 44413
+npzrp4w= 44414
+IHNpbm9u 44415
+ZXJlYWw= 44416
+IEVubGlnaHQ= 44417
+IENoZW1pc3RyeQ== 44418
+IENhdGFsdW55YQ== 44419
+IGRvY3Ry 44420
+YW50b24= 44421
+IHN0dWs= 44422
+IFBsYXRl 44423
+IEthcmRhc2hpYW4= 44424
+IGZpbG9z 44425
+IFdldA== 44426
+INC/0L7Qv9GL0YI= 44427
+IHVua25vd25z 44428
+IFNjaG9u 44429
+IEJhbGR3aW4= 44430
+IHRlbGVzY29wZXM= 44431
+IEd1Y2Np 44432
+b3hpZGU= 44433
+IENvbnNlcnZhdGl2ZQ== 44434
+7ISx7J2E 44435
+IGhpbmF1cw== 44436
+UG93ZXI= 44437
+IOqxtOqwlQ== 44438
+IHByZXZhaWw= 44439
+b3JtYW4= 44440
+bWFjaGluZQ== 44441
+IDE5NDY= 44442
+IHVuYmVs 44443
+IHNjaGF1dA== 44444
+IHBpZWw= 44445
+ZWVudGg= 44446
+IG9iamVjdGl2ZWx5 44447
+IGNoYWtyYQ== 44448
+YXVkaW8= 44449
+IGNoaWNvcw== 44450
+IFZhdWx0 44451
+IG1lZGljaW5hbA== 44452
+IFRhaWw= 44453
+V2hpbGU= 44454
+IGFzcGhhbHQ= 44455
+IGZyb3pl 44456
+IEVL 44457
+dW5jaGluZw== 44458
+bm9zaXM= 44459
+MjAxNQ== 44460
+IEdyaQ== 44461
+IG9kZGx5 44462
+IE3DpHI= 44463
+IEFlZw== 44464
+Y29sbw== 44465
+UGFy 44466
+IOuTpOyWtOs= 44467
+IHZpbmRlbg== 44468
+IE9WRVI= 44469
+IGljZWQ= 44470
+IHNjb3Jw 44471
+IGhhYw== 44472
+cXVhbGlmaWVk 44473
+INGD0LLQuNC00LXRgtGM 44474
+ZXJtbw== 44475
+SEVO 44476
+IHNvaQ== 44477
+IG11bHRpcGxlcw== 44478
+IGxheW91dHM= 44479
+IGJsaW5kbmVzcw== 44480
+IEJvd3Nlcg== 44481
+INC/0L7QtNGC 44482
+IMOO 44483
+dmVudGlvbmFs 44484
+IG1hdGE= 44485
+bWFkxLE= 44486
+IGdlZXo= 44487
+IGNhZGVuY2U= 44488
+IHdhxbxuZQ== 44489
+IENocmlzdGll 44490
+dmVuZ2U= 44491
+Q2FsbA== 44492
+IHR1cm5hcm91bmQ= 44493
+IGJsb2I= 44494
+INCv0Lo= 44495
+IFZvaWNlb3Zlcg== 44496
+IHBlcmls 44497
+IEphaW1l 44498
+IEhPWQ== 44499
+bGFuZQ== 44500
+IHNlYmVs 44501
+IER1bw== 44502
+IEhpc3RvcmljYWw= 44503
+IGRuaQ== 44504
+IGdlbWE= 44505
+eWs= 44506
+IHNhYmVt 44507
+4bqvbmc= 44508
+IHZhcnM= 44509
+IFJvbm5pZQ== 44510
+IFJvbmFsZG8= 44511
+IFBlcnF1w6g= 44512
+bnNpbm4= 44513
+aGFpcg== 44514
+IHJlbGVudGxlc3M= 44515
+IGx5bg== 44516
+IHRyYXZlbGVy 44517
+bmluZQ== 44518
+IGFudGlt 44519
+IOy8gA== 44520
+IHNub3diYWxs 44521
+INGF0LDRgNCw0LrRgtC10YA= 44522
+IGludGVybnM= 44523
+IGNvbnN0aXR1ZW5jeQ== 44524
+INCd0LDQvA== 44525
+15zXnA== 44526
+VkVM 44527
+IHZpa3RpZ3Q= 44528
+IGFwb3lv 44529
+2YTYqA== 44530
+IGphcmQ= 44531
+IGhlaWdodGVuZWQ= 44532
+0YDQvtGB0YI= 44533
+IFNNSVRI 44534
+INC00LXQu9Cw 44535
+IHJlcGFpcmluZw== 44536
+IHJpZ3Q= 44537
+IFNoZWlraA== 44538
+IEJyaXRuZXk= 44539
+IGV2ZXJ5dGltZQ== 44540
+IGFkdmVudHVyb3Vz 44541
+b2NrZXk= 44542
+ZXJudA== 44543
+IGF0YXF1ZQ== 44544
+IEFsdGVybmF0aXZlbHk= 44545
+ZWZmZWN0 44546
+IHBhbGF2cmFz 44547
+IEVsbGlvdHQ= 44548
+IHLDqXVzc2k= 44549
+IGh5cGVydGVuc2lvbg== 44550
+IE1hbnVhbA== 44551
+IHByb3BoZXRpYw== 44552
+IGhhbmRj 44553
+0YzQtQ== 44554
+IHJlZnJhaW4= 44555
+IFNxdWlk 44556
+7J6h 44557
+INC60L7QvNCw0L0= 44558
+w6RsbGVu 44559
+IGxsZWfDsw== 44560
+IGJhc2g= 44561
+aW9ueQ== 44562
+INGB0LrQu9Cw0LQ= 44563
+INC60LDQsQ== 44564
+IGNhcmVsZXNz 44565
+IFBvb2w= 44566
+IHRyw6Fz 44567
+IGZpbHM= 44568
+IFNjaHI= 44569
+IHNwcmF3ZA== 44570
+IE1vbmF0ZW4= 44571
+IHVuZm9yZ2V0dGFibGU= 44572
+IENvdHRvbg== 44573
+IGluY29udmVuaWVudA== 44574
+IFJY 44575
+b3Jpcw== 44576
+IGh1bWJsZWQ= 44577
+16rXlw== 44578
+INii2b4= 44579
+IGluY3Jlw60= 44580
+IEtvbW1lbnRhcmU= 44581
+cmFjacOzbg== 44582
+IHZhbnRhZ2U= 44583
+IFNlYWw= 44584
+IOydtOqxsOulvA== 44585
+IGpvdWU= 44586
+IOyYpOuemA== 44587
+INC40YHQv9GL0YI= 44588
+b2Jlbg== 44589
+IGdyYXRl 44590
+IGNvbnRyb2xl 44591
+IFBlcmN5 44592
+xYJhZGE= 44593
+IHNpbXVsdGFuZW91cw== 44594
+IHByb3RvdHk= 44595
+IGdyb8OfZXI= 44596
+IGJld3Vzc3Q= 44597
+aW5pemk= 44598
+IHBhc3NpZXJlbg== 44599
+IEhhcHBpbmVzcw== 44600
+c2hp 44601
+Z2VodA== 44602
+IHN0YXRpb25lZA== 44603
+IEVyZ2Vibmlz 44604
+IGRpcmVjdGFtZW50ZQ== 44605
+IHN1cnZpdmVz 44606
+IHBlcnNvbmVz 44607
+QkVSRw== 44608
+IHZvbWl0aW5n 44609
+IGNvbmhlY2Vy 44610
+IGFkam91cg== 44611
+IENpdmlj 44612
+cGVp 44613
+YnVyc3Q= 44614
+IOuLpOuLiA== 44615
+6Y8= 44616
+IHNsZWQ= 44617
+IHBsYXRhZm9ybWE= 44618
+IFNlY3Q= 44619
+IERlZmlu 44620
+w6lub20= 44621
+Y2huZXQ= 44622
+IHByb2ZpdGFiaWxpdHk= 44623
+IGVycmVpY2h0 44624
+4buPaQ== 44625
+Y2F0aW9u 44626
+IOyngOq4 44627
+IHBlcmRyZQ== 44628
+IGZlbG9ueQ== 44629
+IDE5NTc= 44630
+IHVuc3VjY2Vzc2Z1bA== 44631
+IG5hZ3lvbg== 44632
+IGVsYXN0aWNpdHk= 44633
+IGZhY2FkZQ== 44634
+IGVhcnRobHk= 44635
+INCw0LzQtdGA0LjQutCw0L0= 44636
+IGNvbm4= 44637
+Y2xh 44638
+RHU= 44639
+IHBvbGl0aXF1ZXM= 44640
+IGhhbG8= 44641
+aWFudGVz 44642
+INC80L7QtdC5 44643
+dG9uZXM= 44644
+ZWxpZXI= 44645
+aHRha2luZw== 44646
+IHdpY2h0aWdl 44647
+IGFubm8= 44648
+IExvaw== 44649
+aWxsaW9ucw== 44650
+IHZpdmVy 44651
+IHNvbGNoZW4= 44652
+IHN1Zg== 44653
+IFNhbHo= 44654
+IE52aWRpYQ== 44655
+enVnZQ== 44656
+IFNwaWtl 44657
+VmlkZW8= 44658
+IHR3b3I= 44659
+IEFsYQ== 44660
+IGhhbnlh 44661
+IEFkbQ== 44662
+7J21 44663
+IFBhdGllbnRlbg== 44664
+IE9uaW9u 44665
+IEtvYmU= 44666
+IFNjZW5l 44667
+IFJhc2g= 44668
+0YDQsNGB0YI= 44669
+aXN0YW5p 44670
+R2VuZXJhbA== 44671
+bGV5ZQ== 44672
+aW1iYXA= 44673
+IGNvbmNlYWxlZA== 44674
+IEZyaWRheXM= 44675
+IFdvb2w= 44676
+INC90L7QstGL0YU= 44677
+2LTYsQ== 44678
+IOqysOqzvA== 44679
+IGplZG9jaA== 44680
+tOyLnA== 44681
+k6Trj4Q= 44682
+IOyepeuCnA== 44683
+dWt0 44684
+TG91 44685
+IOuoueyWtA== 44686
+IEV4cGVjdA== 44687
+INC00L7QvNC+0Lk= 44688
+IGlycmVzcG9uc2libGU= 44689
+IGFjZXJjYQ== 44690
+IFp1c3Q= 44691
+16jXmA== 44692
+VUk= 44693
+IHlvdXR1YmVycw== 44694
+IFBvc2l0aXZl 44695
+IHNvY2lvZQ== 44696
+IHNuYXRjaA== 44697
+IHJlZnJlc2hlZA== 44698
+IG5vbWluYXRpb25z 44699
+IFBhdHQ= 44700
+IG9ic29sZXRl 44701
+IGRlbWnFnw== 44702
+b3JtdcWf 44703
+IOyGlOynge2eiA== 44704
+IGZsYQ== 44705
+IGNyYXppZXN0 44706
+IFppZQ== 44707
+IFTDug== 44708
+emVw 44709
+aWNlbQ== 44710
+IOupi+yeiA== 44711
+IGN5bmljYWw= 44712
+IHRyZXNw 44713
+IGNyYXo= 44714
+1aXV 44715
+IG5lbGxl 44716
+IG1waA== 44717
+IE5lcmVk 44718
+IEtvYg== 44719
+IEVjaw== 44720
+qLjri4g= 44721
+SmFu 44722
+INCi0L7Qs9C00LA= 44723
+IGRlY2k= 44724
+IFZvZw== 44725
+IGJ1YmJsaW5n 44726
+w7ph 44727
+IHByb2R1Y3Rvcw== 44728
+aWJlcmFs 44729
+IHJlcGxpY2F0ZWQ= 44730
+IEltcHJvdmU= 44731
+aWxsYXJ5 44732
+Q2hh 44733
+IHLDqWR1 44734
+g5DtlZjrqbQ= 44735
+IGNvbm5vdA== 44736
+IEtyaXQ= 44737
+INC00YPRhdC+0LI= 44738
+IHRyZWFkbWlsbA== 44739
+IFBX 44740
+INC30L7QstGD0YI= 44741
+IGNsYW1z 44742
+IGRyYWZ0aW5n 44743
+IDE5NTY= 44744
+dW50YQ== 44745
+IGV4cGVuZGl0dXJlcw== 44746
+IEhvb3Zlcg== 44747
+V09P 44748
+0YjQtdC1 44749
+IGRlZHVjdGlvbg== 44750
+bW9uYXJ5 44751
+IHJlY2li 44752
+IHBvdm8= 44753
+IOuNlOs= 44754
+IFBBTA== 44755
+IEJsb3c= 44756
+IHd5cA== 44757
+IGRlc3RhYw== 44758
+ZGVhbA== 44759
+R3JhZW1l 44760
+IG7DqWNlc3NhaXJl 44761
+IGRhbW5lZA== 44762
+IDE5Mzg= 44763
+IOyLpOygnOuhnA== 44764
+IHRyb29w 44765
+IGluc2lnaHRmdWw= 44766
+IFRK 44767
+INC+0YHQsg== 44768
+IGZpZGVsaXR5 44769
+IFNraXA= 44770
+IE1heW8= 44771
+66ed 44772
+YXBwZQ== 44773
+IGJsYXM= 44774
+IFdZ 44775
+IEdO 44776
+Y3Rhcg== 44777
+U3U= 44778
+IGN1ZW50 44779
+aGV3cw== 44780
+IGNvcnBzZXM= 44781
+QWJz 44782
+IHdhc3Rld2F0ZXI= 44783
+IGNpZWs= 44784
+IE9udQ== 44785
+IGV4cGxvc2l2ZXM= 44786
+IGFybWE= 44787
+IFNURVBIQU4= 44788
+cG9saXRpaw== 44789
+IE9zYWth 44790
+dGHFgg== 44791
+IHlhcMSxeW9y 44792
+IGl6cXVpZXI= 44793
+IGJlbGV6YQ== 44794
+IFd5YXR0 44795
+IHN1aw== 44796
+IHNwZWNqYWw= 44797
+IGRhbmtl 44798
+d2hpc3RsZQ== 44799
+IGbDrXNpY2E= 44800
+IEhhcnJpZXQ= 44801
+IOyVhO2MjA== 44802
+IHdpbGxrb21tZW4= 44803
+aXBpbmc= 44804
+INGB0LzQvtGC0YDQuNGC0LU= 44805
+INC80L7QttC10YjRjA== 44806
+IGluYWNjdXJhdGU= 44807
+IGFycm9nYW5jZQ== 44808
+IFJlbW8= 44809
+zrPOrA== 44810
+YXNzZWQ= 44811
+IGRlbGl2ZXJpZXM= 44812
+IHN0aW5reQ== 44813
+INC/0LXRgNC10LY= 44814
+amF5 44815
+IHRyYW5zaXRpb25hbA== 44816
+IHJlcmU= 44817
+IE5HT3M= 44818
+IEFUTQ== 44819
+2K7Yqg== 44820
+aW9sb2d5 44821
+INCy0LvQsNC0 44822
+IHNjaG1l 44823
+IFNoaW5l 44824
+7JWh 44825
+cGFudHM= 44826
+IHNlcmdl 44827
+IHNlbmhvcg== 44828
+IGFiZHVjdA== 44829
+IEJyeWFudA== 44830
+VkVT 44831
+IGF3YWtlbmVk 44832
+IExheg== 44833
+cm9wb2xpcw== 44834
+IExhbw== 44835
+IHZpbGxh 44836
+IHN1bW1lcnM= 44837
+IGVudGhhbA== 44838
+IDE5NDk= 44839
+Vmlh 44840
+IOyWtOyo 44841
+IHRlbmRvbg== 44842
+IHZpb2xldA== 44843
+IGludGVsbGVjdHVhbGx5 44844
+IGJvdW5jZWQ= 44845
+YXJhdXM= 44846
+IDE5MTk= 44847
+IHZyYWFn 44848
+IHNwZWw= 44849
+IFNjaHdhcg== 44850
+U2NvdHQ= 44851
+IEluZG8= 44852
+IOunnQ== 44853
+IGNhbm9uaWNhbA== 44854
+IElLRQ== 44855
+IHRoYXTDrXM= 44856
+IG1lbGxhbg== 44857
+aWdtYXQ= 44858
+Q291bGQ= 44859
+Li4uPyk= 44860
+IGZvYXJ0ZQ== 44861
+IEt1bWFy 44862
+cmVuZG8= 44863
+IMOpbMOp 44864
+4LQ= 44865
+dmFsdWF0aW9u 44866
+Y2FzZXM= 44867
+IGludHVpdGl2ZWx5 44868
+aG9uZw== 44869
+ZXR0ZWQ= 44870
+IHNvdXZlbg== 44871
+IG1vcmI= 44872
+IGNvcnM= 44873
+IE5W 44874
+IEhhc2Fu 44875
+aWV2ZWQ= 44876
+IOyngOq4iOydgA== 44877
+IGR1bXBsaW5n 44878
+IGNvbnRyw7RsZQ== 44879
+IGFtYmlndWl0eQ== 44880
+IGNvZw== 44881
+IFNjcmlwdHVyZXM= 44882
+IGNhaQ== 44883
+IGJldmVy 44884
+IGh1aXM= 44885
+IGFpbWU= 44886
+IGVya2zDpHJlbg== 44887
+IExN 44888
+IEZleQ== 44889
+4K6x4K6k 44890
+IHN1cGVydmlzZWQ= 44891
+IGpld2U= 44892
+c3Bs 44893
+INGG0LXQvdGC0YA= 44894
+IGNvbGxpc2lvbnM= 44895
+2YTZgQ== 44896
+IEhvZ3dhcnRz 44897
+IER1cmhhbQ== 44898
+15XXow== 44899
+IHBob3NwaGF0ZQ== 44900
+IG92ZXJzZWU= 44901
+IGluc3BlY3Rpb25z 44902
+IGJyaW5j 44903
+IFphaw== 44904
+IHBheW9mZg== 44905
+IGNoYXVk 44906
+IEh1bmdlcg== 44907
+w6Nvcw== 44908
+dmly 44909
+IGZpYW5jZQ== 44910
+IGJvdWc= 44911
+bGl2ZWQ= 44912
+Y3J5 44913
+IGpvaW50bHk= 44914
+IGdpcmxmcmllbmRz 44915
+IE5leHVz 44916
+pqzqsqDsirXri4jri6Q= 44917
+IEt3YW5n 44918
+xYLEmQ== 44919
+IE5lZGVu 44920
+aWVjZQ== 44921
+IGluc2VydGluZw== 44922
+IE11bW15 44923
+IEdsb2Jl 44924
+IGxlZQ== 44925
+IGdlcm1hbg== 44926
+IGNyZWFtcw== 44927
+YWNobw== 44928
+IGNoxrBh 44929
+IEdhbGlsZQ== 44930
+IGbDvHJz 44931
+IGVzdGl2ZXI= 44932
+Y2lkb3M= 44933
+Q2hyaXN0aWFu 44934
+IGxvcnNxdQ== 44935
+IGN1dGVzdA== 44936
+dmFsZQ== 44937
+INC60YDQtdC/ 44938
+IHdhcnk= 44939
+IHNsaWNpbmc= 44940
+IGVzcGVyYW5kbw== 44941
+IFZhbmRlcg== 44942
+IERlaXhh 44943
+IDE5NTQ= 44944
+IG3Ds3dpxIU= 44945
+0ZbRlA== 44946
+IHRvb2xpbmc= 44947
+IHJlc3Rvcg== 44948
+IHBvc2ljacOzbg== 44949
+IGludGVudGFy 44950
+IEFwYWNoZQ== 44951
+T1VM 44952
+INmI2Kg= 44953
+IG1hdGnDqHJl 44954
+IGxpbmVu 44955
+IGVzdHJhdMOpZw== 44956
+IE11dHRh 44957
+IHBhcnRpbmc= 44958
+IG1pbmltaXppbmc= 44959
+IGFwcHJlbmRyZQ== 44960
+INCw0L3Qs9C70LjQuQ== 44961
+IERvbw== 44962
+IEZpcmVmb3g= 44963
+Y8OzbW8= 44964
+IGdlb3BvbGl0 44965
+IG1ha2Fu 44966
+IG1vZ2VsaWpr 44967
+IM+AzrXPgc65 44968
+IGPhu6k= 44969
+IGluc3RhbGxlcg== 44970
+IGRpYnVq 44971
+IEhlYXRo 44972
+bG9vcA== 44973
+IEJyb2tlbg== 44974
+SFlVTg== 44975
+c2hlbGY= 44976
+IGZpemVy 44977
+IGVuaGFuY2Vz 44978
+INC00L7RgdGC0Lg= 44979
+IFBVQg== 44980
+IEtvbGxlZ2lu 44981
+IGF0dGFpbmVk 44982
+xL4= 44983
+IG1pc3RyZXNz 44984
+IE9mdGVudGltZXM= 44985
+157Xmded 44986
+IGJld2U= 44987
+IFNvcmE= 44988
+cmF1ZW4= 44989
+YmF1bQ== 44990
+IHJvbGxlcnM= 44991
+IG1lcmluZw== 44992
+IFBBQw== 44993
+INC90ZY= 44994
+IFLDqXB1YmxpcXVl 44995
+INGC0YDQsNCy 44996
+IFZhbmd1YXJk 44997
+dWNpb25lcw== 44998
+IOustOuMgA== 44999
+IGdvdXI= 45000
+r6Q= 45001
+IM+J 45002
+IHNhdW5h 45003
+IHBlaW5l 45004
+IFZhbGVyaWU= 45005
+IFNpa2g= 45006
+ZmVuZGltaXo= 45007
+YmVybw== 45008
+INGH0Lg= 45009
+IGRvxZt3aWFk 45010
+IEV1cm9z 45011
+IGNvbW1lbnRhaXJlcw== 45012
+IHR3ZWFrcw== 45013
+IEZhc3Rlcg== 45014
+INGA0LDRgdC6 45015
+IHByb2dyZXNzaXZlbHk= 45016
+IEV1Y2g= 45017
+Ym9ybw== 45018
+IEluZ3JlZA== 45019
+Q2Fw 45020
+IHVuY2hlY2s= 45021
+IOyYpOuluA== 45022
+IHdyZQ== 45023
+IEZU 45024
+w7ZydW5n 45025
+IG1lbW9yaXplZA== 45026
+IERpbm5lcg== 45027
+IFBoZXc= 45028
+b3VibA== 45029
+IHB1dGE= 45030
+IGFkbWl0cw== 45031
+0LXQt9C00LU= 45032
+b3BvZA== 45033
+IHBhbmRh 45034
+IGhpbmdlcw== 45035
+Y2lwZQ== 45036
+IHRyYW5zYWN0 45037
+IHBvZGlh 45038
+IHBpY3M= 45039
+IGNyaXRlcmlvbg== 45040
+IE9yY2hlc3RyYQ== 45041
+IEJsb2c= 45042
+IHNvbGVtbg== 45043
+IFBpeGFy 45044
+VGhyZWU= 45045
+INCy0L3QuNC3 45046
+IFZvbHVudGU= 45047
+IFNhdmFnZQ== 45048
+IFBWQw== 45049
+IENhZg== 45050
+IHd5a29u 45051
+IGdyYWRlcnM= 45052
+IGNyb3VjaA== 45053
+IGNsaWNoZQ== 45054
+IHNveWJlYW5z 45055
+IE1VUg== 45056
+IEdvbnphbGV6 45057
+IE1pbWk= 45058
+IEJvbHNvbmFybw== 45059
+IGRpYXBocmFn 45060
+IGJpbGFuZw== 45061
+65CY64qU 45062
+IHJlZ3VsYXRpbmc= 45063
+TWM= 45064
+SnVkZ2U= 45065
+INC90L7Qtg== 45066
+IGpha8SF 45067
+aXRlc3Nl 45068
+IFdpag== 45069
+IGxhdGE= 45070
+Z3JvYW5pbmc= 45071
+UE9TSU5H 45072
+INeQ15XXqteV 45073
+IGhhZ2E= 45074
+IGdyb3VuZGluZw== 45075
+IHZpb2xlbnRseQ== 45076
+IHRpbGxz 45077
+IGVuZ2Fn 45078
+IEhvbGxvdw== 45079
+INC/0L7Qv9GD0LvRj9GA 45080
+IHdwcm93YWQ= 45081
+IHJlcGxhY2Vz 45082
+IGZsdW9yZXNjZW50 45083
+dXJnaWNhbA== 45084
+aWdnbHk= 45085
+IFRyYWRpdGlvbmFs 45086
+dHRl 45087
+INmE2Yc= 45088
+IHBob3NwaG9ydXM= 45089
+IGFwcm9u 45090
+IFdhdGVycw== 45091
+IEt1bHR1cg== 45092
+0LDQstCw0Lk= 45093
+IG9saXZlcw== 45094
+INeU15DXnA== 45095
+IHRlaWx3ZWlzZQ== 45096
+IHNlbmNpbGw= 45097
+IHByZW5kcw== 45098
+IG5hcnJvd2Vy 45099
+IGrDpHR0ZQ== 45100
+IEluZm9ybWF0aW9uZW4= 45101
+7IOB7J20 45102
+IHN0YXJ2ZQ== 45103
+IGZyaWNr 45104
+IEJld2Vn 45105
+4KSy 45106
+IGRvbHBoaW4= 45107
+IExBVUdIVEVS 45108
+IElOVEVSVklF 45109
+IHlhbmzEscWf 45110
+IHRvcnBlZG8= 45111
+IHNob3J0YWdlcw== 45112
+7J2065Oc 45113
+xLFsZMSx 45114
+IHBhd3M= 45115
+IG96b25l 45116
+IGN1bHRpdmF0ZWQ= 45117
+IEZvdA== 45118
+IG5vdG9y 45119
+0L3QvtC3 45120
+INC60L7RiA== 45121
+IHRvdWNoc2NyZWVu 45122
+IEFsbHk= 45123
+IOunm+yeiOyWtOyalA== 45124
+INCh0LXRgA== 45125
+INCy0L/QvtC70L3QtQ== 45126
+IHBhcHJpa2E= 45127
+IER1c3Rpbg== 45128
+IGVmZWN0bw== 45129
+IG9waW5p 45130
+IG11dXQ= 45131
+IGjhu41j 45132
+IGludGVyamVjdA== 45133
+xJl0 45134
+IGJ1dHRz 45135
+dXJleg== 45136
+IFBpa2U= 45137
+IEhvaw== 45138
+IEd1aW5lYQ== 45139
+IENhdGhlZHJhbA== 45140
+IDE0MDA= 45141
+Q3Jh 45142
+Kyw= 45143
+66eb 45144
+s7Trj4TroZ0= 45145
+YWJ5cmlu 45146
+IHZpZGVvZw== 45147
+INC+0YDRg9C2 45148
+IHXFvg== 45149
+IGJ1c2NhbmRv 45150
+IEFzc2lzdGFuY2U= 45151
+IG1lbGhvcmVz 45152
+7KG0 45153
+IOuBvA== 45154
+IFJK 45155
+INiq2YU= 45156
+IG9taW4= 45157
+IG1vdG9yY3ljbGVz 45158
+IFNhcHA= 45159
+IHN1cHBseWluZw== 45160
+IEFsZ3Vu 45161
+IGFlcm9zcGFjZQ== 45162
+16LXnA== 45163
+b2NjdXA= 45164
+bGVpc3Q= 45165
+IOqxsOuKlA== 45166
+IGNvbXBsZXRh 45167
+YnJlcw== 45168
+ISg= 45169
+INCf0YDQtdC0 45170
+IGRpc2FkdmFudGFnZWQ= 45171
+IEF0dGVuZA== 45172
+IEp1ZGFo 45173
+4buLY2g= 45174
+eWxlbmU= 45175
+YWN0bHk= 45176
+IHNldHVwcw== 45177
+IGFtbW9uaWE= 45178
+IFNjaHdlaXo= 45179
+IFNoYW1l 45180
+IGJhbmRl 45181
+IEZ1ZWw= 45182
+IHRyb3VibGVzb21l 45183
+IG51bWVybw== 45184
+IE1PTQ== 45185
+INC/0YDQtdC00LvQsNCz 45186
+bWVudGlvbmVk 45187
+INCx0L7Qu9GM0YjQvtC1 45188
+IFZpa3Rvcg== 45189
+IFN0eWxlcw== 45190
+IGNydWNpZmllZA== 45191
+cnVjdHVyZWQ= 45192
+ZW52aXJvbg== 45193
+IG1vcmFscw== 45194
+IG1lZGl0YXRpbmc= 45195
+IGF4aWFs 45196
+aXNhbmNl 45197
+IEFic3Q= 45198
+R3JlZW4= 45199
+IOqxtOw= 45200
+IHF1YWRyYW50 45201
+IHBlcmdp 45202
+IGNhbWVyYW1hbg== 45203
+IFNlcXU= 45204
+IHBhdXNlZA== 45205
+IExhdWdoaW5n 45206
+6reA 45207
+Py4u 45208
+IMW7ZQ== 45209
+IHBlcm1pdGly 45210
+IGRldGVjdG9ycw== 45211
+IEhVRA== 45212
+YXZhbA== 45213
+IOyXrOq4sOq5jOyngA== 45214
+IGh1YnM= 45215
+IGJlc3RpbW10 45216
+INCx0YPQtNC10YLQtQ== 45217
+SU5URVJQT1NJTkc= 45218
+IHRlbmdhbg== 45219
+IGNyYXZl 45220
+IEJ1bmRlc3JlZ2llcnVuZw== 45221
+IEJsb29keQ== 45222
+IHVzYWJpbGl0eQ== 45223
+IEVhcw== 45224
+IMSR4buZbmc= 45225
+IDE5NTU= 45226
+IGtyaWVnZW4= 45227
+IGhhYml0dWFs 45228
+IGVzc2VudGlhbHM= 45229
+cmltaW5hbA== 45230
+IHJvb21tYXRlcw== 45231
+INC/0LXRgNC10YXQvtC0 45232
+IG5naGk= 45233
+IG1lbmluZw== 45234
+IFN5bXBob255 45235
+IEh1Zw== 45236
+YWdnaQ== 45237
+IHdpZWQ= 45238
+IG1pdGFk 45239
+dGVlbnRo 45240
+aWRhxIc= 45241
+U2F2ZQ== 45242
+IHJvYmnEhw== 45243
+IGJvdW5jZXM= 45244
+sJbsl5A= 45245
+c3RhcnM= 45246
+IHByYWdtYXRpYw== 45247
+IGNvZ25pdGlvbg== 45248
+IHdyYXBwZXI= 45249
+IHdhcnRlbg== 45250
+YWRo 45251
+IHBlbnNh 45252
+IEhlcnR6 45253
+IG7Emw== 45254
+IFJlaWQ= 45255
+IFBDcw== 45256
+IE1vbGU= 45257
+IC4uLi4u 45258
+IHByZWNpbw== 45259
+IENoYW1waW9uc2hpcHM= 45260
+6rCA6529 45261
+IHbDqXI= 45262
+IGNvcnJpZG9ycw== 45263
+IEVsZWN0cm9uaWM= 45264
+U2w= 45265
+INCw0LvQtQ== 45266
+IG92ZXJ0aHJvdw== 45267
+IGthYnVs 45268
+IFJFUw== 45269
+IEN5YmVycHVuaw== 45270
+0L7Qs9C+0LQ= 45271
+INCd0LDQsg== 45272
+IHdhbg== 45273
+IG1hbmlmZXN0YXRpb25z 45274
+IGN1YWxlcw== 45275
+IFdpc2U= 45276
+IEzDtnN1bmc= 45277
+IGV4Zm9s 45278
+IGVhcm5z 45279
+0YPRgdGC0LjRgtGM 45280
+IHNhcHA= 45281
+IEJyYXVu 45282
+IEJSQU5ET04= 45283
+7LmZ 45284
+IHNhbm8= 45285
+IEZFTA== 45286
+0YvQstCw0LnRgtC10YHRjA== 45287
+0L7QttC00LXQvdC40Y8= 45288
+IHNld24= 45289
+RnVu 45290
+IHJlY2lwcm9jYWw= 45291
+IGV4cGFuc2l2ZQ== 45292
+IFRyYWZmaWM= 45293
+IGt0w7NyZWdv 45294
+INmI2LM= 45295
+IOu5qA== 45296
+cHJvdmU= 45297
+aWdhcmU= 45298
+IGxvaA== 45299
+2KfYtg== 45300
+SG9wZQ== 45301
+IGRldm90ZWVz 45302
+IEdvbQ== 45303
+IHN0ZWFscw== 45304
+IFVtcw== 45305
+IFR3aWNl 45306
+aXlpbQ== 45307
+IHJoeXRobWlj 45308
+IFZvcnRl 45309
+IHByZWZpeA== 45310
+b21pbmF0aW9u 45311
+IGRhdG8= 45312
+IGN1c3RhcmQ= 45313
+IFZPSUNF 45314
+IG1lbnk= 45315
+aXN0b3Jz 45316
+IO2YkQ== 45317
+IOyCtOyVhA== 45318
+IO2DhA== 45319
+IGtvcnQ= 45320
+IGFiYQ== 45321
+IFZlcmE= 45322
+ZXB5 45323
+IOy5tOuplOudvA== 45324
+IHN1Ym1lcmdlZA== 45325
+IENsb2Nr 45326
+IHRodW1ibmFpbHM= 45327
+IGJvYXN0 45328
+IEZhcmU= 45329
+ISFd 45330
+IMWbbQ== 45331
+IGthaWtraQ== 45332
+IFRlY2hub2xvZ2llcw== 45333
+7Jm4 45334
+0LjRgtCw0Lk= 45335
+INCw0YI= 45336
+IGtub2Jz 45337
+IHJlaWNodA== 45338
+xrDhu6NuZw== 45339
+Z2xpbw== 45340
+IOunm+ydtA== 45341
+6rCQ7J2E 45342
+IGpvdGth 45343
+IEhhbmR5 45344
+IEhhYmVu 45345
+bm91cw== 45346
+IGlubGFuZA== 45347
+IGFtYXpvbg== 45348
+aG9vdGluZw== 45349
+U0w= 45350
+IGxlaXN0ZW4= 45351
+fiI= 45352
+IHByb3Zva2U= 45353
+IFR3aXN0 45354
+INeR15c= 45355
+IGRlcGFydGVk 45356
+6rCc66W8 45357
+IGtvbnNl 45358
+IENhcnd5bg== 45359
+7ZWY7Iug 45360
+aWRlbnRhbA== 45361
+RVNDTw== 45362
+IHR0ZW9rYm9ra2k= 45363
+IGRpemVuZG8= 45364
+xLFuZGFraQ== 45365
+aW1hc3U= 45366
+YWZhcg== 45367
+IGxhbmRmaWxs 45368
+IGNvcnJlY3Rpbmc= 45369
+IGNsZWFycw== 45370
+IE51bW1lcg== 45371
+SEFN 45372
+IGNhcnRyaWRnZXM= 45373
+IERpZXNlbA== 45374
+cGFjZWQ= 45375
+IG9ibGl2 45376
+IG1veWVucw== 45377
+IFNpbm5l 45378
+IFByZWlz 45379
+aWxpeg== 45380
+INGB0LzQvtC2 45381
+IGJyb2FkZW4= 45382
+eGVz 45383
+IGNhcmJvaHlkcmF0ZQ== 45384
+7Zi5 45385
+c2Vvaw== 45386
+IGVjaG9lcw== 45387
+IGNlc3M= 45388
+67CU 45389
+INCx0LjQt9C90LXRgQ== 45390
+IGxsYW1hZG8= 45391
+IGVzc2VudA== 45392
+IOydvOuwmA== 45393
+IEFpcmVz 45394
+cGhlbg== 45395
+IHplYnJh 45396
+IHN5bWJvbGlzbQ== 45397
+T25jZQ== 45398
+IHJhY2tz 45399
+IEthZmth 45400
+INGB0LXRgNGM0LXQtw== 45401
+IHNpbm4= 45402
+cGljaW91cw== 45403
+a2Fh 45404
+IG1vdGhlcmZ1Y2tlcg== 45405
+IGFwcHJlbnRpY2VzaGlw 45406
+IHJwbQ== 45407
+IHRheGF0aW9u 45408
+IGZ1cnJ5 45409
+IFNhY3JlZA== 45410
+INGA0LDQt9C8 45411
+cG9yYQ== 45412
+ZW5nZXM= 45413
+IO2XiOs= 45414
+INGB0LjQvQ== 45415
+IHNhbml0aXplcg== 45416
+IGNyaW5nZQ== 45417
+IFNjYQ== 45418
+0L7Rh9C90L4= 45419
+IG9mZXJl 45420
+IG1lbG9kaWVz 45421
+IFZlbHZldA== 45422
+IElocmVy 45423
+IEh5YnJpZA== 45424
+IEdpb3Y= 45425
+IGlyZ2VuZHdhcw== 45426
+IGRlcGVuZGU= 45427
+IFVzZXJz 45428
+IGh1bXA= 45429
+ZHJpdmluZw== 45430
+IHNm 45431
+IHJ1dGhsZXNz 45432
+4LmA4LiE 45433
+IGxlbW9ucw== 45434
+IGbDtnJldA== 45435
+IE9q 45436
+INC80LDQvNCw 45437
+IGludGVycGVyc29uYWw= 45438
+IGdldg== 45439
+IGFibm9ybQ== 45440
+0LjRgdC7 45441
+INC40L3QtA== 45442
+IGtvbnRyb2xs 45443
+IHJlZ3Jlcw== 45444
+IGxlZGdl 45445
+IGVyesOkaGx0 45446
+IFRhY3Q= 45447
+IGFycml2w6k= 45448
+IHN1YnN0YW50aXZl 45449
+IHNwb29uZnVs 45450
+endpc2NoZW4= 45451
+b29vb28= 45452
+IGNvbnRlbmlkbw== 45453
+IGJlc2w= 45454
+4buDbQ== 45455
+a3Rlbg== 45456
+SmFtaWU= 45457
+IHNhbmR5 45458
+4os= 45459
+IHBhc2U= 45460
+IGRldHRl 45461
+IEJlbGdpYW4= 45462
+6rCc6w== 45463
+dWxhcmVz 45464
+cnVk 45465
+aWdvcg== 45466
+IO2MrOs= 45467
+IHJlbWVkaWVz 45468
+IGJsYXN0aW5n 45469
+IFNpY2g= 45470
+INC+0LbQuNC0 45471
+IG1vbnN0cg== 45472
+IG1hbmlmb2xk 45473
+IGdsYXViZW4= 45474
+IEVTVA== 45475
+IHN0cmVhbWxpbmU= 45476
+IGxvYmJ5aW5n 45477
+IEdvdGhpYw== 45478
+dG9pcmU= 45479
+Li4n 45480
+IGTDqW1vY3I= 45481
+INC90LDQsdC70Y7QtA== 45482
+IHdzcMOzbA== 45483
+IGN6xJnFm8SH 45484
+aXPDqXM= 45485
+Z2FuZ2Vu 45486
+IGJlenBpZQ== 45487
+cmVtbGlu 45488
+6rCd 45489
+U3RpbGw= 45490
+IHJlc2lkZXM= 45491
+IGdlbGVjZWs= 45492
+IHTDqWzDqXBob25l 45493
+IHBld24= 45494
+IGxlb3BhcmQ= 45495
+IGNvbXBsaW1lbnRhcnk= 45496
+IGNyaWI= 45497
+IEFuaW1hbHM= 45498
+IGdlaWw= 45499
+ZXNzZWw= 45500
+IGdhcmRlcg== 45501
+IGNhdGNoeQ== 45502
+IEV0cw== 45503
+IENvbW1lcmNpYWw= 45504
+IERFTk5JUw== 45505
+IENvb3JkaW5hdG9y 45506
+IEFiaWdhaWw= 45507
+ZmZmZmZm 45508
+4bqlcA== 45509
+IHBlcXVlw7Fh 45510
+IGluamVjdGlvbnM= 45511
+Y2VrdA== 45512
+IHBoaWxhbnRocm9weQ== 45513
+IHB1Y2s= 45514
+IGNlbGVicmF0ZXM= 45515
+IER1bms= 45516
+IERsYXRlZ28= 45517
+zrTOrg== 45518
+Z3JhZHVhdGU= 45519
+IE1vYmls 45520
+dGlsbA== 45521
+YWNhbQ== 45522
+IHlvbGtz 45523
+IHRhbmdsZWQ= 45524
+IG1hbmlhYw== 45525
+IG9ibGlnZWQ= 45526
+IExhaW5r 45527
+IHZlcmRlcg== 45528
+IERhbW9u 45529
+IG11dGFudA== 45530
+IGhvcHBpbmc= 45531
+IHJlaW5z 45532
+IGludmVydGVy 45533
+IGNvbnRlbXB0 45534
+16DXoQ== 45535
+bGVhcm5pbmc= 45536
+TWlzcw== 45537
+INCT0L7RgQ== 45538
+IE1leWVy 45539
+6ruY7ISc 45540
+15XXoNeZ150= 45541
+YXNraW5n 45542
+IHRyaW1taW5n 45543
+IHRyZWFzdXJ5 45544
+IHNlbnRl 45545
+QXVzdA== 45546
+IFVudGVyc3TDvHR6dW5n 45547
+IENvbWVkeQ== 45548
+IEFuYWtpbg== 45549
+6bk= 45550
+0YDRg9GC 45551
+IEhhcmk= 45552
+b2dyYXBoZXJz 45553
+IG9hdG1lYWw= 45554
+IEJvdHM= 45555
+INC/0LDQu9GM 45556
+IGFja25vd2xlZGdlbWVudA== 45557
+eGlj 45558
+IOq0gOyLrA== 45559
+Z2FzcGluZw== 45560
+IHRlcnJhY2U= 45561
+IG9ybmFtZW50cw== 45562
+IE1FUg== 45563
+Y29tbWl0dGVl 45564
+IOyXhuyKteuLiOuLpA== 45565
+IHJpag== 45566
+6bM= 45567
+16bXnQ== 45568
+bGVtZQ== 45569
+IGxpYmVydGllcw== 45570
+IGZlbGxhcw== 45571
+IENvcHBlcg== 45572
+YmVuY2g= 45573
+IElkZWE= 45574
+4buNbg== 45575
+0YjQsA== 45576
+IHZlcnNpw7Nu 45577
+z4TOv8+N 45578
+INCc0Lg= 45579
+INC/0YDQuNC70L7Qtg== 45580
+IGJveGVy 45581
+IFRhbm5lcg== 45582
+IE1veQ== 45583
+7LmY64qU 45584
+VGhy 45585
+IHRpbmhhbQ== 45586
+IHBvbGlzaGluZw== 45587
+IGNvbnNlcXVlbnRseQ== 45588
+IGFtZW5pdGllcw== 45589
+IEtJ 45590
+IEdSRUVO 45591
+IEZyYW5raWU= 45592
+0L3QuNGC 45593
+aXR0ZWw= 45594
+0YHQutC+0LU= 45595
+dXJzZWQ= 45596
+IHVwYnJpbmdpbmc= 45597
+IHRo4bup 45598
+IOyLneycvOuhnA== 45599
+IHdoaW0= 45600
+IGNoaW5lc2U= 45601
+Y29uZmlkZW5jZQ== 45602
+IEplZGVy 45603
+YWpjaWU= 45604
+IFRvdXM= 45605
+IFBvd2Vycw== 45606
+4burYQ== 45607
+b3RoZXJtYWw= 45608
+INCy0YvRiNC1 45609
+cmFsZQ== 45610
+2KfYrg== 45611
+IOyngOybkA== 45612
+IMOpcGlzb2Rl 45613
+IHN1bHBo 45614
+IGVuY2FyYQ== 45615
+a3JhZnQ= 45616
+YWxhcsSx 45617
+IENvbWVz 45618
+IGRpdnVs 45619
+IFJ1ZG9scGg= 45620
+IE11c2U= 45621
+IHV0ZW5z 45622
+IOyekOyjvA== 45623
+IHBhbmE= 45624
+IFZlZ2V0YQ== 45625
+IFBIUA== 45626
+IE5TQQ== 45627
+ZW50aW4= 45628
+IENhcm5lZ2ll 45629
+2KfZig== 45630
+acSZY3k= 45631
+SGFycnk= 45632
+IGbEsXI= 45633
+0KHQvw== 45634
+IGdsYWRseQ== 45635
+IGF2ZXJhZ2luZw== 45636
+7ZWY6rKg7Iq164uI64uk 45637
+0LvRj9GO0YLRgdGP 45638
+INCc0LXQvdGP 45639
+IHF1b3RhdGlvbg== 45640
+cmlyZXM= 45641
+aXRjaGVucw== 45642
+YXllZA== 45643
+IHVuYXR0 45644
+IFBlcmV6 45645
+INC+0YLQvNC10YI= 45646
+IHRhY3RpbGU= 45647
+IEV1aA== 45648
+aXNpbmk= 45649
+YnVo 45650
+IGhhdMSxcg== 45651
+IOyeiOycvA== 45652
+IHBvbGljeW1ha2Vycw== 45653
+s7TshLjsmpQ= 45654
+YWPEsQ== 45655
+IM66zrk= 45656
+IHJlZ2lzdGVyaW5n 45657
+cmV0bw== 45658
+IFNwcmlua2xl 45659
+IEdyYW1teQ== 45660
+YXh0ZXI= 45661
+INCx0Lg= 45662
+IHNpdHRlcg== 45663
+IHByZWRpYw== 45664
+IHRoaW5seQ== 45665
+IHN0cnVt 45666
+IGFnZ3Jhdg== 45667
+IGFoYQ== 45668
+2LHYrA== 45669
+bWVsbG93 45670
+IGNvbnN0YW50ZQ== 45671
+IExhdXQ= 45672
+aXN0b24= 45673
+IHRyYW5zaXRpb25lZA== 45674
+IENhbWJvZGlh 45675
+YXJ0ZWQ= 45676
+IG1pc2Y= 45677
+IFB1bmt0ZQ== 45678
+jOuToA== 45679
+IHRyZW1ibGluZw== 45680
+IGdlc3Bhbm50 45681
+INi52YTZitmH 45682
+INC90LjQutCw0LrQuNGF 45683
+IOu2gOuTnOs= 45684
+INGA0LDQt9Cy0LjRgg== 45685
+IGl0Y2h5 45686
+IGNpZW50bw== 45687
+IHBsYWlucw== 45688
+IGtpdHRlbnM= 45689
+IGJhY2tsb2c= 45690
+IFByZXNpZGluZw== 45691
+cHRh 45692
+IGhhdm9j 45693
+IERhcnJpbg== 45694
+INCb0Y7QsQ== 45695
+IHNlZ3JlZ2F0ZWQ= 45696
+IGdoZXR0bw== 45697
+IGVybGVidA== 45698
+IGRydWdpZWo= 45699
+IFNpeHQ= 45700
+4Lij4Liw 45701
+dWVuY2lh 45702
+IO2VmOq4sA== 45703
+IOuGjQ== 45704
+IHJvYmk= 45705
+IHBpb25lZXJz 45706
+IG1pbGxpYXJkcw== 45707
+IFdpdGNoZXI= 45708
+IOustOyXhw== 45709
+b3Jybw== 45710
+bWFzcw== 45711
+IGRpdmVyZ2VuY2U= 45712
+IFJpdmVyYQ== 45713
+IE5vb2RsZXM= 45714
+IGVuZHJvaXQ= 45715
+IEtvc3Rlbg== 45716
+INC00YDRg9Cz0LA= 45717
+IG3DrW5pbW8= 45718
+IEthemFraHN0YW4= 45719
+2KrZhw== 45720
+INCy0L7Qt9C00YM= 45721
+IGdlc2NocmllYmVu 45722
+IE5pbA== 45723
+0YHQutC4 45724
+IEZyw7xo 45725
+IGJldmVyYWdlcw== 45726
+IEdvbg== 45727
+QXJpbg== 45728
+IEludHJv 45729
+b2NhbHlwdGlj 45730
+IGV4aGF1c3Rpb24= 45731
+IFN0YXR1cw== 45732
+IEJhdHRlcnk= 45733
+w6lzeg== 45734
+o7zr 45735
+YWlyeQ== 45736
+IOuztOyXrOuTnOs= 45737
+IGRpc3Bhcml0eQ== 45738
+2Yw= 45739
+IFR1Y3Nvbg== 45740
+IGJyaWdodGx5 45741
+cHJvYmxlbQ== 45742
+IGJpb21hc3M= 45743
+p4k= 45744
+IGh1cmRsZQ== 45745
+IHdhdmVsZW5ndGhz 45746
+IDw8 45747
+IHRlYW1lZA== 45748
+RkZGRg== 45749
+IFNsaW0= 45750
+b21pYWw= 45751
+IHVudmVpbGVk 45752
+IFZlcmVpbg== 45753
+2YLYtw== 45754
+ZXN0cnk= 45755
+IGNsw6Fz 45756
+IGNoZWRkYXI= 45757
+IGFjY3VzaW5n 45758
+IFNjaWVudGlmaWM= 45759
+INCx0YPQtNC1 45760
+IEN5cnVz 45761
+zrXPhM61 45762
+hpPqs6A= 45763
+IOuzhA== 45764
+IGN1cmQ= 45765
+IHJlZmVycmFscw== 45766
+c2hpZnQ= 45767
+bmlrw7N3 45768
+IG1pZXI= 45769
+IGNvbmZyb250aW5n 45770
+6rKD64+E 45771
+YXds 45772
+IHRyeWlu 45773
+IOq3uOuemOyalA== 45774
+IGNoaWFy 45775
+IOyYpOuKmOuPhA== 45776
+ZXNxdWU= 45777
+IG1pc21vcw== 45778
+IFNoYWs= 45779
+IHNvY2lhdXg= 45780
+IHBpxZ8= 45781
+IGtpxZ9p 45782
+IGN5YW4= 45783
+aGF5 45784
+YmV3 45785
+Ym9k 45786
+IM65 45787
+IE1haW5seQ== 45788
+0Y7RgtGM 45789
+aGFiaXR1ZGU= 45790
+INGB0L/QvtC60L7QuQ== 45791
+IHByZWNvbg== 45792
+IE1hbmR5 45793
+8J+kow== 45794
+aWxsb3M= 45795
+IGdydXBw 45796
+IGNydW1ibGU= 45797
+IGNvbnN0cnVjdG9y 45798
+ZXJ2aWNlcw== 45799
+IGxpZ2h0aG91c2U= 45800
+IENvbmNlcHQ= 45801
+0LDQvdGC0Lg= 45802
+YWx0cm8= 45803
+aG9wZQ== 45804
+IEFsbGVn 45805
+7Ja066W8 45806
+cGllY2Vz 45807
+b3VudGVy 45808
+IO2VmOuLiOq5jA== 45809
+IOyduO2EsOs= 45810
+IHbDqXJpdGFibGU= 45811
+IHRocmVhZGVk 45812
+YmxpbmQ= 45813
+gpjrnbw= 45814
+IHRyYXlz 45815
+IEVkaXNvbg== 45816
+IMOWeg== 45817
+IFN0ZXZpZQ== 45818
+IGxlbmRlcg== 45819
+IGJyaWdhZGU= 45820
+IGRldXRzY2hl 45821
+bXVmZmxlZA== 45822
+YmFydA== 45823
+IGluc2FuaXR5 45824
+IHNhdnZ5 45825
+IHNlbnNhdGlvbmFs 45826
+IGRlcmVjaG9z 45827
+IE1Y 45828
+INC/0YDQtdC/ 45829
+IHRocmVhdGVucw== 45830
+IHJlYWx0w6A= 45831
+IGluZGljYXRpdmU= 45832
+IGNob3Bz 45833
+IGJlbmVmaXRpbmc= 45834
+IFZlcm5vbg== 45835
+IFN0cmFuZA== 45836
+bnVu 45837
+cXVlbnRseQ== 45838
+MTAx 45839
+IGVlbA== 45840
+7IiZ 45841
+cmludHM= 45842
+INmF2LM= 45843
+INio2K8= 45844
+INC/0L7RgdGC0YDQvg== 45845
+IHlhcG3EscWf 45846
+IG9sbWFzxLE= 45847
+IGllZGVyZWVu 45848
+b2zDqQ== 45849
+a2Vm 45850
+IOuwnOyDnQ== 45851
+IHJhaW5lZA== 45852
+IGFsbWlnaHR5 45853
+INCy0YvQtA== 45854
+IENQUg== 45855
+RnJl 45856
+IGluaGFiaXRlZA== 45857
+IGFyYmV0cw== 45858
+IGFraW4= 45859
+0LDRgdGC0LI= 45860
+dmFuaWE= 45861
+IGjDpHVmaWc= 45862
+IE1hdHRl 45863
+c29ycnk= 45864
+SmVubnk= 45865
+INCz0YDQsNC0 45866
+IHdoaXQ= 45867
+IGJyb2tlcnM= 45868
+IGhpbmU= 45869
+YXN0ZW4= 45870
+INCz0YDRgw== 45871
+TUI= 45872
+IFBSSQ== 45873
+U2Fi 45874
+IHdyZXN0bGVy 45875
+IGZhY2lsaXRhdGluZw== 45876
+IGVoa8Ok 45877
+IENyZWQ= 45878
+IDEyNw== 45879
+IG5vdGhpbg== 45880
+IG1hbmRhdGVk 45881
+0YPRgtGB0YLQsg== 45882
+RnJhbms= 45883
+IHdvcnM= 45884
+IGR6aWXFhA== 45885
+IFVuZGVyZ3JvdW5k 45886
+IHpuYWpkdQ== 45887
+IELDpA== 45888
+IFByaW56aXA= 45889
+0LDRgtC10LvQtdC5 45890
+IHZldGVyaW5hcg== 45891
+IHNwbGVuZGlk 45892
+IHJvenA= 45893
+IHBzeWNob3BhdGg= 45894
+aWdvbg== 45895
+IGhvcHM= 45896
+IGPhuqdu 45897
+IFhpYW4= 45898
+IHRyb2lzacOobWU= 45899
+IHByb2R1Y3Rv 45900
+IGRlxJ9lcg== 45901
+IENvbnRpbnVpbmc= 45902
+0LjQstCw0Ls= 45903
+Y8Sxaw== 45904
+IG1vaXN0dXJpemVy 45905
+V2hpdGU= 45906
+IHNpaXM= 45907
+IEV2ZXJlc3Q= 45908
+aWVuY2Vk 45909
+IGPhuqNt 45910
+IEphcG9u 45911
+tOyghA== 45912
+IHRlbsOtYW4= 45913
+IGVuY2FudGE= 45914
+TW0= 45915
+IGRyb3Bkb3du 45916
+IEl5YQ== 45917
+s7TrqbQ= 45918
+IHdvcmRpbmc= 45919
+IFNxdWVlemU= 45920
+IE1hcGxl 45921
+IGNsYXJpZmllZA== 45922
+IE11bmljaXA= 45923
+IFJvdWdl 45924
+IE5pY2tp 45925
+IEdvbw== 45926
+dm9sdA== 45927
+dGVr 45928
+ZmVjdHVyZQ== 45929
+ZnJlZA== 45930
+YXJyaXZl 45931
+dGV6 45932
+RXA= 45933
+IG9icmFz 45934
+IFZJRA== 45935
+IFJpdg== 45936
+IE1vZGk= 45937
+aWJl 45938
+IGFjb250ZWNlbmRv 45939
+IGltaXRhdGlvbg== 45940
+IGNhbW91ZmxhZ2U= 45941
+IHNwYW5uaW5n 45942
+IFNFQ1JFVA== 45943
+IE9yZW8= 45944
+7IaM66as 45945
+IGh1bmNo 45946
+IGNhxYJl 45947
+IHNwb250YW5lb3VzbHk= 45948
+IFBlcmQ= 45949
+IGV0YXA= 45950
+IEhvbGU= 45951
+IERpc2FiaWxpdHk= 45952
+IGFmdGVybGlmZQ== 45953
+IHRlc3RpZmllZA== 45954
+IHByZXN1cA== 45955
+IHBldHJvbGV1bQ== 45956
+IGNvbnRyYXJpbw== 45957
+IEFzc2Vzc21lbnQ= 45958
+xJ9sdQ== 45959
+IHBlc3Rz 45960
+IGRpbGln 45961
+INCy0YHRgtGA0LXRgg== 45962
+IGNvbnPDqXF1 45963
+IGNhbm5vbnM= 45964
+IGNhbm9l 45965
+IE1pbGU= 45966
+IGNpdG95 45967
+IGJlZ2dlZA== 45968
+IE1pbm5pZQ== 45969
+xYJ5Y2g= 45970
+IHByaW5jaXBl 45971
+z4DPjM69 45972
+bW5pZWo= 45973
+IHdlcnQ= 45974
+IOuLpOuTpA== 45975
+YW5zZQ== 45976
+IHVuY2xlcw== 45977
+IHByb3ZvY2F0aXZl 45978
+IGludGVyc2VjdGlvbnM= 45979
+IGRlbW9jcmF0cw== 45980
+IEp1bGl1cw== 45981
+0LjQvdC60Lg= 45982
+eWd1c2Fs 45983
+INec15U= 45984
+IGdqb3JkZQ== 45985
+IGdhc2tldA== 45986
+IEJvY2s= 45987
+IMSwbg== 45988
+YnJlYXQ= 45989
+IEVxdWl0eQ== 45990
+YXJkxLE= 45991
+INC60LDQvdCw0LvQtQ== 45992
+INC00L3QtdC5 45993
+IHThu5tp 45994
+IGZpeHR1cmU= 45995
+IGFidXNlcw== 45996
+IHZheWE= 45997
+IG91dmVydA== 45998
+IG11bHRpY3VsdHVyYWw= 45999
+IGNvbnRleHRv 46000
+IFNlc2FtZQ== 46001
+IGTDqXBs 46002
+IGNvbnNvbW0= 46003
+IFBhcnRl 46004
+IHBlbQ== 46005
+IENvbmFu 46006
+INCx0ZbQu9GM 46007
+IHBlcnN1YWRlZA== 46008
+IGRyYWlucw== 46009
+TW9v 46010
+Rk9SRQ== 46011
+INCx0LDRgg== 46012
+IGZvZA== 46013
+IFByb2R1Y3Rz 46014
+7KeE7Kec 46015
+ICJb 46016
+IFdpY2s= 46017
+IE5hcnV0bw== 46018
+0L3QsNC70Lg= 46019
+cnl3 46020
+IGxvZGdl 46021
+IGluaA== 46022
+IHZvbnRhZGU= 46023
+IGRpag== 46024
+IEplc8O6cw== 46025
+TG9va2luZw== 46026
+IGZvcmVhcm0= 46027
+IEludGVncmF0aW9u 46028
+IEhBUlJJUw== 46029
+IHRvb2xiYXI= 46030
+bGVhZGVy 46031
+IHNlbGRvbQ== 46032
+INCx0YDQvtGB 46033
+IEtvb2s= 46034
+0L7QvdC0 46035
+IG1vbm9wb2w= 46036
+IG1pbGxldA== 46037
+IGxpcmE= 46038
+IEFzaWFucw== 46039
+IDE4OTA= 46040
+Y2nEn2lt 46041
+IGVkZW4= 46042
+IElLRUE= 46043
+IE5laWdoYm9y 46044
+IEthenV5YQ== 46045
+w7xk 46046
+IHBzeWNoZWRlbA== 46047
+IGVudmlzaW9uZWQ= 46048
+IO+3uw== 46049
+IHd1bmRlcg== 46050
+IEJ1bGdhcmlh 46051
+QnJpZA== 46052
+IG1hcnJvdw== 46053
+IGRlcGljdGlvbg== 46054
+IFRpbg== 46055
+IFBoYXJpc2U= 46056
+IGVpbnppZ2U= 46057
+IGJsaW5kbHk= 46058
+IGRlZmVucw== 46059
+RGlyZQ== 46060
+IHZpYnJhdGluZw== 46061
+IHRyb2xscw== 46062
+IGRpc3Jlc3BlY3RmdWw= 46063
+IHdvZA== 46064
+IHN0aW11bGk= 46065
+IGNyZWVwaW5n 46066
+IGNsYWlyZW1lbnQ= 46067
+IHNjYXJpZXN0 46068
+IGTDqWNvdXZyaXI= 46069
+IDEwNA== 46070
+INCy0LXRgNGF 46071
+IMWCYXQ= 46072
+IHLDs8W8bmU= 46073
+IGJhcmxleQ== 46074
+IFJlcGw= 46075
+IFR3ZQ== 46076
+a2tl 46077
+IFJlZG1p 46078
+IE1ldHJvaWQ= 46079
+IM6uz4TOsc69 46080
+Q2hlY2s= 46081
+IFNFTg== 46082
+IGlkbw== 46083
+0YLQvtGA0LjQuA== 46084
+w7Nw 46085
+VU5LTk9XTg== 46086
+IMOkbmRlcm4= 46087
+IEp1aWNl 46088
+IEdlc2ljaHQ= 46089
+INC90LDRgdGC0L7Qu9GM0LrQvg== 46090
+7YOV 46091
+wq0= 46092
+ZXhoYWxlcw== 46093
+IOy0iQ== 46094
+IGpzZW0= 46095
+z4DPic+C 46096
+IGl0dA== 46097
+66qF7J20 46098
+IHJlbWl4 46099
+IGJsb3Nzb21z 46100
+IFJlbmVl 46101
+aXNhdGlvbnM= 46102
+7Iqk7YSw 46103
+IOuztOydtOuKlA== 46104
+dWVzdGFz 46105
+b3BlZGlh 46106
+IEFpbQ== 46107
+7J207KaI 46108
+c2NlbmU= 46109
+IGxlYWthZ2U= 46110
+dWNrdA== 46111
+U2Fk 46112
+QXNr 46113
+IHN1c3BlbnNl 46114
+IGltcG9zdA== 46115
+IFN0cmF0ZWdpYw== 46116
+IEl0w61z 46117
+4oCM 46118
+IGtleWJvYXJkcw== 46119
+IGFtdXNpbmc= 46120
+b2dy 46121
+aWRlcm1hbg== 46122
+npY= 46123
+INCy0LjQttGD 46124
+IGRpcHM= 46125
+IGFwb2xvZ2l6ZWQ= 46126
+IFNUQVI= 46127
+IGVzY3VlbGE= 46128
+IENoaW5n 46129
+0L3QtdC90LjRjw== 46130
+IOu2gOu2hOydtA== 46131
+IEZsZWV0 46132
+IHNhbWI= 46133
+IGVudHNwcmVjaGVuZA== 46134
+IGVsZWN0cm9kZXM= 46135
+IEZyZWloZWl0 46136
+IFNocmlt 46137
+acOfZQ== 46138
+IHNlbGVjdGlvbnM= 46139
+IGZvcmRp 46140
+IGRvc3M= 46141
+0Y/Rhw== 46142
+IGRpc2NyaW1pbmF0ZQ== 46143
+IEF1w59lcmRlbQ== 46144
+IGRlc2Vudm9sdg== 46145
+IEludGVybmFs 46146
+IEJlbmVkaWN0 46147
+IFNoaXY= 46148
+TWlzc3k= 46149
+INC+0LHQvdCw0YDRg9C2 46150
+INC90LDRgdGC0YDQvg== 46151
+IGNvbnRyb2xhcg== 46152
+IExpYQ== 46153
+IG9waW9pZHM= 46154
+YW50dQ== 46155
+IGN1cGJvYXJk 46156
+0LPQtQ== 46157
+YWNodHM= 46158
+IGN1cmF0ZWQ= 46159
+IHhlbQ== 46160
+IHdlYXJ5 46161
+IGJyZXRocmVu 46162
+IGJ1ZGdldGluZw== 46163
+IHBvdXJ0YW50 46164
+YWlzaWE= 46165
+INC+0YLQstC10Yc= 46166
+IEdJUw== 46167
+zrzOsc65 46168
+INep15TXldeQ 46169
+IHNhdWQ= 46170
+IGzhu5s= 46171
+0JXQog== 46172
+dWJpbmU= 46173
+INC90YPQttC10L0= 46174
+IGtpZG5hcHBpbmc= 46175
+IGJyYXQ= 46176
+IFRlcnJl 46177
+IE1vbmV0 46178
+IOuniOyKpO2B 46179
+IGZsYXNoeQ== 46180
+IElTQk4= 46181
+IGZyZWVsYW5jZQ== 46182
+aWFnZQ== 46183
+IGp1bmdl 46184
+7Lap 46185
+Y2VyYWw= 46186
+INGC0L7Rh9C60Lg= 46187
+IGZvcm11bGF0ZQ== 46188
+IEZFUg== 46189
+IERhcnRtb3V0aA== 46190
+7Jy866m07ISc 46191
+b3dpxIU= 46192
+IOuUlOyekA== 46193
+IHJlZ2ltZW50 46194
+IG1ldGFib2xpc21v 46195
+IFBhcnI= 46196
+IOy2qeu2hA== 46197
+IHNhbml0eQ== 46198
+IExhbA== 46199
+IEfDtg== 46200
+IEdsYQ== 46201
+IHByb3Rv 46202
+IG1pY3Jvc2NvcGlj 46203
+IGthbmc= 46204
+IFNjYWxpYQ== 46205
+IHB1Zw== 46206
+IFNjb3Jl 46207
+IFNhdmFubmFo 46208
+IGdhcmRl 46209
+IE5PUg== 46210
+IHNjaGVpbnQ= 46211
+IHDDs8WC 46212
+IGNvcnJp 46213
+IGJydXRl 46214
+IMWCYWQ= 46215
+IHN1Y2NlZWRpbmc= 46216
+IGJpY3ljbGVz 46217
+Tm9u 46218
+IHNlZWtlcnM= 46219
+IHVuY29uZGl0aW9uYWw= 46220
+IHJoeW1lcw== 46221
+IEdhcmFnZQ== 46222
+IGludm9pY2U= 46223
+IGNhbnZp 46224
+bmVjaw== 46225
+IGN1c3RvbWl6YWJsZQ== 46226
+aXJpdHVhbA== 46227
+UXVlZW4= 46228
+7ZWY7Iuc64qU 46229
+IHBvd2VybGVzcw== 46230
+IGNzYWs= 46231
+aXNvZnQ= 46232
+IOygle2ZlQ== 46233
+IG5ow6Ju 46234
+IE1BTkQ= 46235
+IEhhZg== 46236
+IHJldm9sdmVz 46237
+b3Zhbg== 46238
+YXJvbw== 46239
+IEdyaW5k 46240
+IGluZGlzcGVuc2FibGU= 46241
+IGNvbnN1bHRlZA== 46242
+IENsaW5pY2Fs 46243
+QWNj 46244
+IG9saG9z 46245
+IG1vbnRlcg== 46246
+IEhhbmE= 46247
+ZXRhaA== 46248
+IHZhYW4= 46249
+IHRpZ2Vycw== 46250
+IGNhdWN1cw== 46251
+8J+Ygg== 46252
+s7TsnpA= 46253
+cG93ZXJz 46254
+aXVtcw== 46255
+IO2GoOs= 46256
+IHRyYWRpY2lvbmFs 46257
+IHJlc29uYXRlZA== 46258
+IOyLoOq4sA== 46259
+dGhlbQ== 46260
+Um9iZXJ0 46261
+IGVsZW1lbnRv 46262
+IGFudGlk 46263
+INC+0LHRgQ== 46264
+IG5hdGl2ZXM= 46265
+IGxvY2E= 46266
+b3dtZW50 46267
+IFRpZ2h0 46268
+IG1lbGFu 46269
+IE51ZQ== 46270
+YW1pcw== 46271
+IHNvcmdlbg== 46272
+YXPEsW5h 46273
+SG9tZQ== 46274
+IFBVQkc= 46275
+IGF3ZnVsbHk= 46276
+IFNob3Jl 46277
+IFBlcmNow6k= 46278
+IExhdQ== 46279
+IENpbmRlcmVsbGE= 46280
+IENoZXN0 46281
+IHNlbWFudGlj 46282
+IGRlc2VydGVk 46283
+IE1vbW8= 46284
+IEhlcm5hbmRleg== 46285
+Z2VuZXM= 46286
+IEFkdWx0 46287
+0LjRh9C10YHQutC+0LPQvg== 46288
+b3NoaW1h 46289
+IGNhcmFjdGVyw61zdGljYXM= 46290
+IEtM 46291
+tOyepQ== 46292
+b2Nhcg== 46293
+IGZlaGx0 46294
+IGRydWs= 46295
+IFBvcHB5 46296
+RU5HTElTSA== 46297
+IFZlcmdsZWljaA== 46298
+QnJpZW4= 46299
+IHJlY29tcA== 46300
+INGB0LQ= 46301
+IG1lcmdlcg== 46302
+IG1hcmtldGVycw== 46303
+IGhvbmV5bW9vbg== 46304
+IHBlbnNv 46305
+IGJlbGxp 46306
+0LXRgtGD 46307
+IGJhbmtlcg== 46308
+Q2FtZXJh 46309
+IFN0YWxs 46310
+IFN0YW1w 46311
+IEJpdGU= 46312
+0LXQttC00LU= 46313
+IHPDvHI= 46314
+IGfDvMOn 46315
+IFBhc3NvdmVy 46316
+IEJ1Z8O8bg== 46317
+INGB0L7QttCw0LvQtdC90LjRjg== 46318
+INC90LjQtw== 46319
+IG1hbnVyZQ== 46320
+IGdsYWNpZXI= 46321
+UkFZ 46322
+dGVycm9y 46323
+IHNhbGFkcw== 46324
+IGh1cnJpY2FuZXM= 46325
+IERlc2lnbmVy 46326
+YXRvcmlv 46327
+IGZhY3R1YWw= 46328
+IFRhbW15 46329
+INC30LLRg9GH 46330
+IGludHJvZHVjdGlvbnM= 46331
+IGhvdXNla2VlcGluZw== 46332
+IGhhbmdlcg== 46333
+64uY6w== 46334
+YWt0ZQ== 46335
+IENvbGE= 46336
+J10= 46337
+IEdlbmRlcg== 46338
+0L7RgNC+0L0= 46339
+aXBzZQ== 46340
+aWNpYXM= 46341
+IHN1Y2Nlc3NpdmU= 46342
+IHBvbGl0aWM= 46343
+IGjDtmhlcg== 46344
+IFFpYW8= 46345
+IEdpbW1l 46346
+INC70L7Qtg== 46347
+IHNlYg== 46348
+IFdlaXRlcg== 46349
+IFNha3VyYQ== 46350
+IEJvdWxkZXI= 46351
+IEFtw6lyaWNh 46352
+cGXFgm5pZQ== 46353
+IHRlY25vbG9nw61h 46354
+aXNob3Bz 46355
+ZnVy 46356
+IG1vb25saWdodA== 46357
+IGRpc3BlcnNlZA== 46358
+IHJleg== 46359
+0LXQvdC90L7QtQ== 46360
+0LDQu9GM0L3Rg9GO 46361
+IFR3ZWx2ZQ== 46362
+IEhPUg== 46363
+7Iuk7Z6I 46364
+aWxhZ2U= 46365
+IHNoYWRlZA== 46366
+IHJlc3VtZXM= 46367
+IFBlYW51dA== 46368
+IE1JTEw= 46369
+YXBvbnM= 46370
+IFVGQw== 46371
+IFNvbGU= 46372
+IGpveXN0aWNr 46373
+IE9saXZpZXI= 46374
+d2FybWluZw== 46375
+IHN5bGxhYnVz 46376
+INC+0LHRidC1 46377
+IGhp4buHbg== 46378
+IGZlc3Rh 46379
+IGNyYWRsZQ== 46380
+IFphYw== 46381
+IHJlbWVtYnJhbmNl 46382
+IOqwmeyVhOyEnA== 46383
+IHBpxJlr 46384
+IGNvZXhpc3Q= 46385
+IFZJSQ== 46386
+IMOhcmVhcw== 46387
+IHV3YcW8 46388
+IG9ic2VydmVycw== 46389
+IG3DpG5uaXNrb3I= 46390
+Y29vbg== 46391
+IERBTQ== 46392
+IG5hc3p5bQ== 46393
+IGFsbGlnYXRvcg== 46394
+IEZyZWV6ZQ== 46395
+IEVzdGF0ZQ== 46396
+INGC0YDQsNC00Lg= 46397
+IHVuZGVyY292ZXI= 46398
+IG5pZXM= 46399
+IEZlaGxlcg== 46400
+cGxpbg== 46401
+IEthYnVs 46402
+aWxhdGU= 46403
+IOqzoOyWkQ== 46404
+IG1vcA== 46405
+7IS8 46406
+IGFuZGVyZXI= 46407
+IEtFTEw= 46408
+0L7QutC4 46409
+INC20LXRgdGC 46410
+IGdyYXppbmc= 46411
+IGRhw60= 46412
+IGNhcGl0YWxpemU= 46413
+IGFwZXg= 46414
+IG51cnR1cmluZw== 46415
+IGNvcnRhcg== 46416
+IGNvbnRyYWM= 46417
+xLFtxLF6xLE= 46418
+IHRhbmRlbQ== 46419
+Z2VtZW50 46420
+INGB0LjRgdGC0LXQvNCw 46421
+IG1hbnF1ZQ== 46422
+aWFqxIU= 46423
+V09S 46424
+INin2Kg= 46425
+IGNhcnRz 46426
+QU5P 46427
+IOuwm+qzoA== 46428
+IENlbmE= 46429
+IEJpb2xvZ3k= 46430
+aWRhcg== 46431
+IGHFvA== 46432
+ZXJuZQ== 46433
+YW51 46434
+IHRoYW5rZWQ= 46435
+IHN1Ym1hcmluZXM= 46436
+IG1hbmlj 46437
+INC80L7Qtw== 46438
+aW5zdGFudA== 46439
+ZXNzZW50aWFs 46440
+IHNhbXVyYWk= 46441
+IHBhc3Rp 46442
+IGFsYW4= 46443
+IGJyb2No 46444
+IGJha2Vy 46445
+IEd1aWxs 46446
+qLw= 46447
+IHdpdGhkcmF3bg== 46448
+64ud 46449
+UGVyZmVjdA== 46450
+cXVlbmN5 46451
+IHN0cmVhbWxpbmVk 46452
+IDEzMDA= 46453
+tOuPhA== 46454
+IOuWoOs= 46455
+IGh2YWQ= 46456
+IHZlcmJhbGx5 46457
+IEtvbnM= 46458
+IOyhsOyLrA== 46459
+IGRpZXo= 46460
+IGNodWNrbGluZw== 46461
+IE1paA== 46462
+IHJhbGxpZXM= 46463
+IG1hbnRlcg== 46464
+IGVhcm5lc3Q= 46465
+c3VwZXI= 46466
+IGdlY2U= 46467
+IFJlbmQ= 46468
+IEdlcmFkZQ== 46469
+amVuaWdlbg== 46470
+IFZhbGw= 46471
+IOyeiOuCmA== 46472
+INGB0LrQsNC30LDQu9Cw 46473
+IHRyYWJhbGg= 46474
+INC90LDRiNC10Lw= 46475
+INC80LXRhQ== 46476
+aWtpdA== 46477
+IG5vdW5z 46478
+IG5ldXJvbG9naWNhbA== 46479
+IG1vdGl2YXRpb25hbA== 46480
+IE1jTWFob24= 46481
+IEZpbmlzaGVk 46482
+IOuztOydtA== 46483
+IEZpZWxkcw== 46484
+IGFkb2xlc2NlbnRz 46485
+IFRpc2No 46486
+IE5lYmVu 46487
+IEZsb3dlcnM= 46488
+IEVuZXJn 46489
+IGRpcmV0 46490
+IFRoaQ== 46491
+IFBpY2Fz 46492
+IGF2ZXRl 46493
+IEZvcnM= 46494
+IENoYXBlbA== 46495
+TsOjbw== 46496
+RXQ= 46497
+INGB0L7QtNC10YDQtg== 46498
+cmVubw== 46499
+IHN2ZW4= 46500
+IGRvc3TEmXA= 46501
+bmVl 46502
+IFNuYXBkcmFnb24= 46503
+IElEcw== 46504
+7JWY64qU642w 46505
+16jXmg== 46506
+IHN1bmZsb3dlcg== 46507
+IHBlcnBldHVhbA== 46508
+IGtuaWdodHM= 46509
+IGdpcmQ= 46510
+IFRvbGQ= 46511
+IHZvbGNhbm9lcw== 46512
+IGFkdmVyc2FyeQ== 46513
+IEVjb25vbXk= 46514
+IGV4dHJhcG9s 46515
+IGJsdWV0b290aA== 46516
+IHpvb21pbmc= 46517
+IHNreXM= 46518
+IGdlbmlhbA== 46519
+w61jdWxvcw== 46520
+YW1icmU= 46521
+INC80LXRgA== 46522
+IHRlZW55 46523
+IHN0cmVzc2luZw== 46524
+7JWM 46525
+T05Z 46526
+IHRyYW5zbHVjZW50 46527
+IHJvdW5kaW5n 46528
+IGdydWVz 46529
+15nXoNeU 46530
+YXByw6hz 46531
+IHBydWViYQ== 46532
+IHBvbHlnb24= 46533
+IGJsdWViZXJyeQ== 46534
+IFByb2dyYW1t 46535
+IHRyZW5jaGVz 46536
+IHNlYmFnYWk= 46537
+IHBhbGF0ZQ== 46538
+IGxhdWRl 46539
+IGJlaGF2ZWQ= 46540
+IGxvbmdpdHVkaW5hbA== 46541
+IE1vZHVsZQ== 46542
+IGFkbWly 46543
+zrvOuQ== 46544
+R3JlZw== 46545
+IHd5c3Q= 46546
+IHByb3BhZ2F0ZQ== 46547
+IG1vbGRz 46548
+IFR1Yg== 46549
+IExvdWQ= 46550
+dXN0bw== 46551
+IHVuc3RvcHBhYmxl 46552
+IHJlaW5mb3JjaW5n 46553
+INC/0YDQvtCx0LvQtdC80LA= 46554
+IHBvdGVuY2lhbA== 46555
+IGhlbXA= 46556
+7J6U 46557
+4KSv 46558
+IG9wdGlj 46559
+IGVyZm9sZ3JlaWNo 46560
+0YHRiw== 46561
+0L7Qu9GM0YjQtQ== 46562
+dXJzdA== 46563
+IFBvaXM= 46564
+IHJlc3BvbmRlbnRz 46565
+IG5laG1l 46566
+IEV4dGVybmFs 46567
+b2xhdGU= 46568
+SHl1bg== 46569
+IHF1YXJ0eg== 46570
+IG1hdGhlbWF0aWNpYW4= 46571
+IGLDoXNpY2FtZW50ZQ== 46572
+IGFpbA== 46573
+7KCc66W8 46574
+YXR0dXR0bw== 46575
+IG5vb2l0 46576
+IGFmZmxpY3Q= 46577
+IE9sZ2E= 46578
+INC90LDRgg== 46579
+IGRpdGVz 46580
+IHJlYWxpZGFkZQ== 46581
+IGvDpG4= 46582
+IHVuaXF1ZW5lc3M= 46583
+IHBhZHJlcw== 46584
+IHN1YnNpZGk= 46585
+IHBpZ2VvbnM= 46586
+zrLOsQ== 46587
+c3RhZA== 46588
+IGRlcmVu 46589
+INCh0LvQtdC0 46590
+ZG9v 46591
+INC+0L/QuNGB0LDQvdC40Lg= 46592
+IGFtYmVy 46593
+IGdvb3NlYnVtcHM= 46594
+IGZyw6Vnb3I= 46595
+IFZpdGFs 46596
+IElzcmFlbGl0ZXM= 46597
+d2Fzc2Vy 46598
+SXNu 46599
+IGNvbW1pdHM= 46600
+IFNURVZFTg== 46601
+IEJldsO2bGtlcg== 46602
+dWl0aXZl 46603
+IGxlZ2Vu 46604
+IGJydWs= 46605
+0LjRgNC+0LLQsNC9 46606
+eW5lbg== 46607
+aGVsbQ== 46608
+IGdlbmVyYXRpb25hbA== 46609
+IEzDpG5kZXJu 46610
+zr/Ouc+Az4zOvQ== 46611
+dXp1 46612
+IGNhbGxlcg== 46613
+0L7QvdGM 46614
+w7xtw7w= 46615
+IGJlc2Fy 46616
+IHBsYXRz 46617
+IG1pZ3JhdGVk 46618
+IGphcA== 46619
+IFdBUg== 46620
+IGRpc3NlY3Q= 46621
+IFp1c2No 46622
+IFplaXRlbg== 46623
+IExpb25z 46624
+IERG 46625
+4pQ= 46626
+0LrQuNCy 46627
+IHBlZGVzdHJpYW5z 46628
+IE1hcmlseW4= 46629
+ZG9jaw== 46630
+IHlodA== 46631
+IHJlaW5jYXJu 46632
+IFNvbm8= 46633
+IEdyb3d0aA== 46634
+0YPRgdC+0LI= 46635
+IGR1bmdlb25z 46636
+IGJhZ3Vz 46637
+a2ljaA== 46638
+INGD0LrRgNCw0Zc= 46639
+IEtlbGxlcg== 46640
+Y2hlbWlzdHJ5 46641
+SmFwYW5lc2U= 46642
+IHdpbGxzdA== 46643
+IGRlY29tcG9zaXRpb24= 46644
+INGB0YLQtdC9 46645
+IHJldml2ZWQ= 46646
+7ZWZ6rWQ 46647
+IMWT 46648
+7Iu4 46649
+aXBweQ== 46650
+IGhvdXJseQ== 46651
+asOkbg== 46652
+IFdvcmtzaG9w 46653
+nbzshJw= 46654
+IGN1YXJ0bw== 46655
+IHBhdHJpbQ== 46656
+IEJ1cmNo 46657
+IOyeiOq4sA== 46658
+IGhlcGF0 46659
+IGjDoG5n 46660
+IOuMgO2VtA== 46661
+INCy0LDRiNC4 46662
+IHJld29yaw== 46663
+IHBhcnNl 46664
+IMOnxLFrdMSx 46665
+IFNheA== 46666
+IE1vbmdv 46667
+IEFhYWg= 46668
+cmFtYmxl 46669
+REo= 46670
+IHN0YWJpbGl6ZWQ= 46671
+IFNwZWVjaA== 46672
+Qm9va3M= 46673
+IGh1cmRsZXM= 46674
+IFdP 46675
+IExhbWJvcmc= 46676
+IDE5MzM= 46677
+IHZvcmJlcmU= 46678
+IGNsaW5pY2FsbHk= 46679
+IGJyZWF0aHRha2luZw== 46680
+IEdhdGV3YXk= 46681
+0L/QtdGA0LLRi9GF 46682
+dXRlcnM= 46683
+IOu5tQ== 46684
+IHlldGVy 46685
+IHB1bGxleQ== 46686
+IG11ZmZpbg== 46687
+IFByZWZlcg== 46688
+IFBlbmNl 46689
+IGluZm9ybWHDp8Ojbw== 46690
+7Iqk7Yq46w== 46691
+IFR1cnRsZQ== 46692
+IFJlZ2luYQ== 46693
+IExvYWQ= 46694
+ZG9lcw== 46695
+cGFuemU= 46696
+uJQ= 46697
+IG1pbmE= 46698
+IExhdGlub3M= 46699
+YW1tZXJz 46700
+IFRvcnQ= 46701
+IEJleW9uY2U= 46702
+0LjQvNC+0YHRgtC4 46703
+INCy0L7Qv9GA0L7RgdGL 46704
+IGJ1bHVu 46705
+aW5law== 46706
+YmVyZWljaA== 46707
+IHBhc3R1cmU= 46708
+IE9B 46709
+IE1lbHQ= 46710
+IEV0dA== 46711
+IERZ 46712
+IG9id29obA== 46713
+IGxlYWd1ZXM= 46714
+0YLQtdGB0Yw= 46715
+INC60YPRgQ== 46716
+IHZvcnM= 46717
+IHRvcHA= 46718
+b2dyYXBoaWNhbA== 46719
+YXNzdA== 46720
+IGxpbmRv 46721
+IOuwne2YlA== 46722
+IHLDqWZs 46723
+IGNsaW1icw== 46724
+IHZhcnNh 46725
+IG1ldGh5bA== 46726
+IEthcmVyZQ== 46727
+xrDhu58= 46728
+UmFk 46729
+IHByZXBhcmVkbmVzcw== 46730
+0L7QvdGH 46731
+IE9E 46732
+IENHSQ== 46733
+IOCkrg== 46734
+IHNwZWVjaGxlc3M= 46735
+IGxhc2Np 46736
+IGJvbGFn 46737
+INGF0L7Rh9C10YLRgdGP 46738
+IGdyaWV2aW5n 46739
+IEpvaGFubmVz 46740
+IENhcnJvbGw= 46741
+YWRha2k= 46742
+iKzr 46743
+IHPFgnU= 46744
+IGlubmVyaGFsYg== 46745
+IGd5bW5hc3RpY3M= 46746
+0L/RgNC4 46747
+aWZpcXVlcw== 46748
+IGthcmF0ZQ== 46749
+IGRvbXU= 46750
+T1RIRVI= 46751
+IGRlbWFuZMOp 46752
+IGJvb2tsZXQ= 46753
+IEt5b3Rv 46754
+IHdvaA== 46755
+IE1hcsOtYQ== 46756
+dmlvbGVudA== 46757
+SkU= 46758
+IGzDs2c= 46759
+IGJydXRhbGx5 46760
+Y290 46761
+INmF24w= 46762
+IFdhcnN6 46763
+d29s 46764
+IG1pa8Ok 46765
+IFByb25vdW5jZQ== 46766
+IEJyZW5kYW4= 46767
+IHJvdXA= 46768
+IGl0YWxpYW5v 46769
+INC60L7QvNC/0YzRjtGC 46770
+IHVyZ2luZw== 46771
+ZWRlcw== 46772
+IGNhcmJvbm8= 46773
+IFJpY2hhcmRzb24= 46774
+INCd0LDRhw== 46775
+IFRyYWluZXI= 46776
+IENyaW1lYQ== 46777
+IGRpYXBlcnM= 46778
+IGNvdmV0 46779
+IE1haGFy 46780
+IEh1dGNo 46781
+IEF1c3c= 46782
+YmVydHk= 46783
+IGluZGlmZmVyZW50 46784
+0LrRgNC10YI= 46785
+dWxkYWRl 46786
+IGhhcm1z 46787
+otmG 46788
+bGVzaWE= 46789
+IGdpbw== 46790
+IE1pc3RyZXNz 46791
+IEtub3g= 46792
+IEZSRUU= 46793
+IOujqOs= 46794
+INC90LDRiNCw 46795
+IGludmluY2libGU= 46796
+IG1haWRlbg== 46797
+IEplZXo= 46798
+IGJyZXZl 46799
+cG9sZQ== 46800
+IGNyaXRpY2lzbXM= 46801
+IFJ1c2lh 46802
+4KSu 46803
+cGhpbg== 46804
+IENvbXBhcmU= 46805
+IEJPTg== 46806
+IHNuZWFraW5n 46807
+IFJhaWxz 46808
+IEdlcmFs 46809
+IDE5NTM= 46810
+SG9sYQ== 46811
+INC+0L/Ri9GC 46812
+IHJhaW5mb3Jlc3Q= 46813
+IGJlbHVt 46814
+IE9iaQ== 46815
+IElTUw== 46816
+INCh0LI= 46817
+IGJsb25k 46818
+IHd6Z2w= 46819
+IHBvd2llZHppYcWC 46820
+IGNob2tpbmc= 46821
+IFNvbmdz 46822
+IEJpcmF6 46823
+IHllbGxz 46824
+IHN0eWxpc3Q= 46825
+z4zPhM61 46826
+IHNjaHJlaWJlbg== 46827
+IEphdw== 46828
+IEVsZXZlbg== 46829
+IFJpZg== 46830
+Ly4= 46831
+IOyYpOuenOunjA== 46832
+IHRyZWF0aWVz 46833
+dWZmZWQ= 46834
+IOKIkg== 46835
+IHJvb2Zz 46836
+4LmA4Liq 46837
+IOu7 46838
+IHNwYXJrbGU= 46839
+IEtpZXY= 46840
+IEFyZ3U= 46841
+ZXJlY2h0 46842
+INCd0LDQtNC+ 46843
+IEZJTA== 46844
+IG1vbHRh 46845
+IERldmk= 46846
+IGNhbXBl 46847
+IGJlbmV2b2w= 46848
+IFRvdWdo 46849
+IG1vaW0= 46850
+IGV2YWN1YXRl 46851
+IGVycmFkbw== 46852
+0YDRg9Cz0L4= 46853
+IO2OmA== 46854
+IM6TzrnOsQ== 46855
+IHdlYWtlbg== 46856
+IGlsbHVtaW5hdGVk 46857
+IHNpZ2xv 46858
+IFZhY2M= 46859
+0LjQtdC5 46860
+YWxpcw== 46861
+INGD0YHRgtGA0L7QuQ== 46862
+IGRvbmE= 46863
+xYJvcw== 46864
+w7xtYW4= 46865
+IHByb2R1Y2Npw7Nu 46866
+IGNsb3Q= 46867
+IE1hbmdv 46868
+IHVuZWFzeQ== 46869
+IHNodXRz 46870
+IEV4YW1wbGVz 46871
+dmVsbA== 46872
+ZWJl 46873
+IHByb21wdGx5 46874
+IFRlbGVz 46875
+INC/0YDQvtGI0Ls= 46876
+IHB1ZXJ0YQ== 46877
+IMO8YmVyemV1Zw== 46878
+IGNvY2g= 46879
+c29jaWFs 46880
+IEJlbnNvbg== 46881
+IE1ldGg= 46882
+IEV4cGVk 46883
+IHN1cHBsZW1lbnRhbA== 46884
+IGNvbmNlaXZl 46885
+INeY15XXkQ== 46886
+IGNhcHRpdml0eQ== 46887
+j5nslYg= 46888
+INGF0YPQtA== 46889
+Zm9ybWluZw== 46890
+IHVwbG9hZHM= 46891
+IHR1cmJ1bGVuY2U= 46892
+am9pbnQ= 46893
+IHNhdGlzZmFjdG9yeQ== 46894
+IEFuaW1l 46895
+IHdhc2hlcw== 46896
+IGxpYmVyYWxz 46897
+IFN1bnNoaW5l 46898
+IFJFQUw= 46899
+dWJsaWs= 46900
+YmluYXJ5 46901
+VG9ueQ== 46902
+IHBvbGFyaXplZA== 46903
+IGVucmljaGVk 46904
+dGFraW5n 46905
+IOuBneuCmA== 46906
+IHBsZWFzdXJlcw== 46907
+IGV4dGVybWlu 46908
+aW5lc2U= 46909
+YXRs 46910
+dsOkcg== 46911
+0LDRgNGL 46912
+IG15xZs= 46913
+bmFycmF0b3I= 46914
+INC+0LTQvdC+0Lw= 46915
+IG5handpxJk= 46916
+IG1vYmlsaXpl 46917
+IG1pbGxvcg== 46918
+IGF0YQ== 46919
+IHBvbMOtdGljbw== 46920
+IHBsZWFk 46921
+IHBhaW50ZXJz 46922
+IFNvdw== 46923
+0L7RhA== 46924
+IOyYm+uCoA== 46925
+INGH0YLQvtCx 46926
+IHNhYm9y 46927
+IFVuZGVydA== 46928
+IEpFUlJZ 46929
+xaHDrQ== 46930
+IOuwluyXkA== 46931
+IHByw6ljw6lk 46932
+IGFubm90YXRpb24= 46933
+IEluYXVkaWJsZQ== 46934
+IHRleHR1cmVk 46935
+IGZpc2hlcm1hbg== 46936
+dm9yZGFu 46937
+aWNoZXJ1bmc= 46938
+IOyggeydtA== 46939
+IGdlemVpZ3Q= 46940
+IG1hbmRhdGVz 46941
+IGJlYWs= 46942
+IFRXTw== 46943
+IEFrYmFy 46944
+aWxpYW4= 46945
+IHRp4bq/cA== 46946
+IHN1cGVyaW9yaXR5 46947
+aW5rdQ== 46948
+IGx5cw== 46949
+IEZDQw== 46950
+IENQQQ== 46951
+dXN0ZXJpbmc= 46952
+bmljb3M= 46953
+YW5qYQ== 46954
+IGNoaWxscw== 46955
+IENhZ2U= 46956
+IHNlYWxpbmc= 46957
+IHNhw6c= 46958
+IGRlZGFucw== 46959
+IEFsZ2Vy 46960
+IHNwZXppZQ== 46961
+IGNvbG9zcw== 46962
+xLF5xLE= 46963
+Y2xvY2t3aXNl 46964
+IGV4YWN0YW1lbnRl 46965
+IGllbWFuZA== 46966
+YW3EsQ== 46967
+IG1hbmRhcg== 46968
+cmFq 46969
+ZmFjZWQ= 46970
+YWd1YQ== 46971
+IOq5lOs= 46972
+IGluc2Jlc29uZGVyZQ== 46973
+IGRyaXp6bGU= 46974
+IGRpbWluaXNo 46975
+IFlvZGE= 46976
+QUk= 46977
+IGJpbG1peW9ydW0= 46978
+IE1NQQ== 46979
+YXRlZ29yeQ== 46980
+INC/0LXRgNC10L8= 46981
+IHBhcnRpY2lwYXI= 46982
+IG5vcm1hbGl6ZWQ= 46983
+IGNvbXBsZXhpdGllcw== 46984
+0LDRgNC+0LI= 46985
+bWlzdA== 46986
+aWNoYQ== 46987
+R3JvdXA= 46988
+IHJlc2lsaWVuY3k= 46989
+IG5vZ2xl 46990
+IENOQw== 46991
+cHLDvA== 46992
+IHBoeXNpY2lzdHM= 46993
+0L3QvtC6 46994
+TEk= 46995
+IHN0dWZmcw== 46996
+IHNpc3RlbWFz 46997
+IGludGVyZmVyaW5n 46998
+IE1hcnZpbg== 46999
+w6lyY2l0bw== 47000
+IOyXhuqzoA== 47001
+IHNvbmlj 47002
+IGVxdWl2 47003
+IGFib3Jk 47004
+IFJhbWVu 47005
+IDA5 47006
+bWVkaW0= 47007
+YXRpcXVlcw== 47008
+INC00LXQu9Cw0Y7Rgg== 47009
+IHVuYW5pbW91c2x5 47010
+IHNraXJ0cw== 47011
+IO2KueuzhA== 47012
+IFByaXg= 47013
+a2FtaQ== 47014
+IGZydWl0aW9u 47015
+IGJpcnRoZGF5cw== 47016
+0LjQutC+0Lw= 47017
+IGluYXVndXJhbA== 47018
+IGNvcnJlbGF0ZQ== 47019
+IFRvcnk= 47020
+IOuCmOyB 47021
+IGRldw== 47022
+IFByZWNpcw== 47023
+aWhp 47024
+IOusuOygnOqwgA== 47025
+IGNpdGluZw== 47026
+IExhbmE= 47027
+IEthZw== 47028
+IHBsYXl0aHJvdWdo 47029
+IFByb3RvY29s 47030
+ZnJpc3Q= 47031
+aG92YWg= 47032
+IG1lcmNpZnVs 47033
+IGJpbGluZ3VhbA== 47034
+IEd1aXRhcg== 47035
+cmg= 47036
+IGdsYW1vcm91cw== 47037
+IFZpa2luZ3M= 47038
+IE9vb29o 47039
+7ZWY64qU642w 47040
+IFVnYW5kYQ== 47041
+IGNvbGxhcHNlcw== 47042
+ZW50cnk= 47043
+IGFudGlveGlkYW50cw== 47044
+64KY6w== 47045
+0YjQsNGP 47046
+IHRyaXZpYQ== 47047
+IGfDpGxsZXI= 47048
+IGZ1bmdp 47049
+IG1pbGtz 47050
+IGRpY2h0 47051
+zrzOtw== 47052
+cG9rZQ== 47053
+INCy0YvQv9GD0YHQug== 47054
+IGZlZWRlcg== 47055
+IEFsY29ob2w= 47056
+aG93ZXI= 47057
+IGRlc2VydmluZw== 47058
+IFJlYmVs 47059
+aW9zaXM= 47060
+IDEwMw== 47061
+IGhhbmRvdXQ= 47062
+IGVubQ== 47063
+IGxhbmRsb3Jkcw== 47064
+IGdlb2xvZ3k= 47065
+cmlscw== 47066
+IGNvYnJh 47067
+IFZvbGQ= 47068
+IFBhbmNo 47069
+IEdSRUc= 47070
+IHByb3Nz 47071
+IGJyYWNlbGV0cw== 47072
+IFZlZ2E= 47073
+IHJvenVt 47074
+0LDQt9C0 47075
+IEx5bmQ= 47076
+IEhvbm9ycw== 47077
+IHN1cnJlbmRlcmVk 47078
+IGxpYnJhcmlhbnM= 47079
+MTI1 47080
+INGB0LjQsw== 47081
+IHVuaWZvcm1seQ== 47082
+IEVhZ2xlcw== 47083
+7JWZ 47084
+0LjRgtCw0L0= 47085
+YW5kaWQ= 47086
+IOygiOuMgA== 47087
+INi2 47088
+IGFycmVzdHM= 47089
+IENTVg== 47090
+IEF6ZXJiYWlqYW4= 47091
+b3J0aWM= 47092
+IERY 47093
+IEFkdmVudHVyZXM= 47094
+IGFidXM= 47095
+IEZhdQ== 47096
+IHNjaGxpbW0= 47097
+IHJhdHRsaW5n 47098
+IGNvbnN1bWVz 47099
+IFRvbGtpZW4= 47100
+IHJlc3VycmVjdGVk 47101
+IFhZ 47102
+7Yq46rCA 47103
+INCy0YvRgdGC0YPQvw== 47104
+IEFuZ2ll 47105
+xbxlbmlh 47106
+TWlj 47107
+IFNoZWlsYQ== 47108
+YWNodGV0 47109
+IG92ZXJzdA== 47110
+IGzDog== 47111
+IGluZWZmZWN0aXZl 47112
+IHdpY2h0aWdlcg== 47113
+IHZpbm8= 47114
+IHB1bQ== 47115
+IGFuZ2xlZA== 47116
+IFBpb25l 47117
+IE3hu7k= 47118
+d2/Fm8SH 47119
+ZHJhdw== 47120
+4Lix4LmI 47121
+bWFya2V0cw== 47122
+IGNhZmVz 47123
+IENlbQ== 47124
+4p2k 47125
+IFN1aXQ= 47126
+TUs= 47127
+IGVtcGhhc2l6ZXM= 47128
+IHRvcnRpbGxh 47129
+IG1lam9yYXI= 47130
+IFN1cnZpdg== 47131
+Y2FzdGluZw== 47132
+IGVkdWNhY2nDs24= 47133
+IEd1bQ== 47134
+dWVseQ== 47135
+IOyXrOq4sOuKlA== 47136
+IHN0cmV0Y2h5 47137
+ZW7Dp2E= 47138
+IHdpdGhob2xk 47139
+IGV4aXRpbmc= 47140
+IGVudGhhbHB5 47141
+IFRyYW5zaXQ= 47142
+xLFsbcSxxZ8= 47143
+YWxpZXM= 47144
+IHNhbHZhcg== 47145
+IGxlYW5lZA== 47146
+IGdyb8OfZXM= 47147
+IGZpdHQ= 47148
+0LDQutC4 47149
+U2FyYWg= 47150
+IGhvc3RlbA== 47151
+IGZpbmdlcm5h 47152
+IG5hZHppZWrEmQ== 47153
+d2l2ZXM= 47154
+UmVj 47155
+IHNwb29s 47156
+0LDRgtC+0LI= 47157
+IEVuZW15 47158
+IGZ1cnk= 47159
+IGRldHRh 47160
+IEZheQ== 47161
+0Y/RjtGC 47162
+IGFwcm94aW1hZGFtZW50ZQ== 47163
+IHNpbG9z 47164
+IG1hZ2lzdA== 47165
+IGNyZWU= 47166
+IEtyYW5r 47167
+IERPV04= 47168
+IHN0YXJ0bGVk 47169
+IHJlYm9ybg== 47170
+IFVtd2VsdA== 47171
+IFN1emFubmU= 47172
+0L3QuNGG0Ys= 47173
+b3V0ZXo= 47174
+IEpBQw== 47175
+eWFyZHM= 47176
+cmFkYXM= 47177
+cmF1 47178
+aXB0cw== 47179
+aGFpbA== 47180
+IHBhcmFncmFwaHM= 47181
+IG1lZ2xpbw== 47182
+IGlzb2xhdGluZw== 47183
+IGFjZWl0ZQ== 47184
+IEhhcnNo 47185
+IGN5c3Q= 47186
+IEJsb2NrY2hhaW4= 47187
+INGF0L7RgNC+0YjQuNC5 47188
+IHZpcnR1b3Vz 47189
+IGludmVzdGlnYWNpw7Nu 47190
+IGRldm9pcg== 47191
+IG1hc3R1cmI= 47192
+IFNhbGU= 47193
+2YrYsdip 47194
+IM6n 47195
+IFN0cmHDn2Vu 47196
+IGRpa2s= 47197
+IGFmb3Jl 47198
+IEp1bmdrb29r 47199
+IGNob2NpYcW8 47200
+IERlYmF0dGU= 47201
+IHdlaXJkbHk= 47202
+IHZpYWpl 47203
+cmVnaXN0 47204
+SGVscA== 47205
+IGtpbmRlcmVu 47206
+IGZvcm11bGF0ZWQ= 47207
+IGVuZmlt 47208
+IFRvd2FyZHM= 47209
+0LrQvtGX 47210
+aXZlcmluZw== 47211
+INC00LXRgtC4 47212
+Y2hhcmdlcg== 47213
+IHB1cmw= 47214
+IGFjYWRlbWljYWxseQ== 47215
+IE51cnNl 47216
+IGRlbGV0aW5n 47217
+YXlv 47218
+IHJlZnVzYWw= 47219
+IGRlcGljdHM= 47220
+IERyYWN1bGE= 47221
+IHRvYXN0ZWQ= 47222
+IFpvbWJpZQ== 47223
+IFN1cGVyaW9y 47224
+IEJvbGQ= 47225
+IHF1aXp6ZXM= 47226
+IGdsZQ== 47227
+NDUw 47228
+IGNvbWXDp28= 47229
+eW5u 47230
+IHZlcnN0 47231
+IE9sYWY= 47232
+IHBvbW9j 47233
+IFNhc2s= 47234
+65g= 47235
+IFRDUA== 47236
+IFByb3BlcnR5 47237
+7ZWY7KOg 47238
+4Lic4Lih 47239
+Ym9vbQ== 47240
+YXJvcw== 47241
+INGA0L7RgdGB0LjQuQ== 47242
+INCx0YvQstCw0LXRgg== 47243
+IOydtOyVvOq4sOulvA== 47244
+IGNvbWJpZW4= 47245
+dmFjYw== 47246
+IGViZW5mYWxscw== 47247
+cGFyYQ== 47248
+INC30Lw= 47249
+IGRlc3BlcmF0aW9u 47250
+b3JkcmU= 47251
+INep15zXmQ== 47252
+IGdlbmVyb3VzbHk= 47253
+INCe0Lo= 47254
+IG9yYml0aW5n 47255
+Pjwv 47256
+IGVzcMOt 47257
+IENPUA== 47258
+dmlzaWJsZQ== 47259
+INC/0YDQtdGB0YLRg9C/ 47260
+IHN0aXRjaGVk 47261
+4K+ILg== 47262
+IGxhdGVudA== 47263
+IFByYWI= 47264
+IE1jTg== 47265
+IEhlYWxpbmc= 47266
+IEN1cmlvc2l0eQ== 47267
+Y2VydA== 47268
+IOuvvOyjvA== 47269
+IHBhdGllbnRseQ== 47270
+IFlU 47271
+Zm9yZWlnbg== 47272
+IHbhuqtu 47273
+IGluZHVzdHJp 47274
+IGNvY2t0YWlscw== 47275
+IGJyaWdodGVu 47276
+IGNvbnNvbGlkYXRlZA== 47277
+0LDRgNC0 47278
+bHRyeQ== 47279
+IGdyaWxsZQ== 47280
+IGJvbmE= 47281
+IGRpbGlnZW50bHk= 47282
+IFdyZXN0bGVNYW5pYQ== 47283
+ZXJrdA== 47284
+ZW5lcmd5 47285
+OTk5 47286
+4K6V4K61 47287
+IHRvdGU= 47288
+aW9ubw== 47289
+RElP 47290
+IHNjaGl6b3BocmVuaWE= 47291
+IHBvc3Rwb25lZA== 47292
+IFFpdQ== 47293
+IM+Dz4XOvQ== 47294
+IHpkasSZ 47295
+IHNwYW5uZW5k 47296
+IERJUw== 47297
+UmVs 47298
+IHJoaW4= 47299
+aW1tdW5l 47300
+T2xk 47301
+IHBsw7Z0emxpY2g= 47302
+IG1vdW5k 47303
+IGFzdHJvbm9taWNhbA== 47304
+IEd1aWQ= 47305
+IEN1bA== 47306
+SEk= 47307
+IMWg 47308
+IHJlcG8= 47309
+IE1hdXJpY2U= 47310
+IGJhbmRpdHM= 47311
+IERlc2t0b3A= 47312
+w6Rzcw== 47313
+ZnRh 47314
+IGxpY2VuY2U= 47315
+IGltYWdpbmFy 47316
+IEVudHJlcHJlbmU= 47317
+eG8= 47318
+IOunm+yeiOuKlA== 47319
+INeU15E= 47320
+IHB1bXBraW5z 47321
+IGthbnNzYQ== 47322
+IGrEmXp5 47323
+IGNvbW11bmF1dMOp 47324
+YsO8cg== 47325
+IGVyaMO2 47326
+IFdvbHZlcg== 47327
+IFNoYXJpbmc= 47328
+IHBha2Fp 47329
+IGluc3VsdGVk 47330
+0JzRiw== 47331
+0L7Rlw== 47332
+IGNvbnNpc3Rl 47333
+IHlvdW5nc3RlcnM= 47334
+IGdsZWljaGVu 47335
+d2VkZXI= 47336
+IG1vdGU= 47337
+IGNsYXVzZXM= 47338
+w6l0YXQ= 47339
+cHJ1cw== 47340
+IHdhc3Q= 47341
+IENyaXNw 47342
+IG9mZmVuZGVycw== 47343
+IGNvbnZlY3Rpb24= 47344
+IGNvbmZpYW4= 47345
+b2xsb3c= 47346
+YW1ldA== 47347
+INGX0YU= 47348
+ZmZpY2llbmN5 47349
+IHVuZ2xhdWI= 47350
+aWdhbnM= 47351
+IG1hcmtldGVk 47352
+IFZBTg== 47353
+IHByb2NsYWltZWQ= 47354
+IGPDqWx1bGFz 47355
+IGNvbGxpZGU= 47356
+IE9jdWx1cw== 47357
+YWRvcmU= 47358
+Smk= 47359
+IHN1c3RhaW5pbmc= 47360
+IEZhc2M= 47361
+IHNldHp0 47362
+IG5vc2FsdHJlcw== 47363
+TW9zdA== 47364
+INCy0Yc= 47365
+IG5hdWM= 47366
+IEJoYXI= 47367
+IHnDqnU= 47368
+IHRpbWVzdA== 47369
+IHBlcnRhbWE= 47370
+aXJtaQ== 47371
+IHp3cg== 47372
+IHZlcmJlc3M= 47373
+IHZvcnRleA== 47374
+IFNUQUNL 47375
+2KvYsQ== 47376
+uYTr 47377
+lJTsmKQ= 47378
+IGxpbmthZ2U= 47379
+IEZyYXNlcg== 47380
+ZW5hcmlv 47381
+IOudvOuKlA== 47382
+IOyEoOuwsA== 47383
+aHRoYWw= 47384
+IOq5jA== 47385
+IEtow7RuZw== 47386
+w4M= 47387
+IHNjcmFtYmxlZA== 47388
+IEVpbms= 47389
+IG1pY3Jvb3JnYW4= 47390
+IG5hcmNpc3Npc3Q= 47391
+IEtvbWJhdA== 47392
+IOunoQ== 47393
+IEFHQQ== 47394
+IHBlcmZla3Q= 47395
+IFNlcmll 47396
+ZGV0ZXJt 47397
+LSc= 47398
+IHBvbnl0YWls 47399
+IGtvc2th 47400
+7JM= 47401
+IG9iZWM= 47402
+IGNoZXN0cw== 47403
+dmVlcg== 47404
+IHVwcmlzaW5n 47405
+IHN0b2tlZA== 47406
+YXNzb2Np 47407
+IHByb2R1w6fDo28= 47408
+IFNoYXBl 47409
+7KCc6rCA 47410
+IOuUsA== 47411
+IGpvbg== 47412
+IGluYWR2ZXJ0 47413
+YW50YXM= 47414
+INC90LDQutC+0L3QtdGG 47415
+IEFyc2VuYWw= 47416
+IHByb3RlZw== 47417
+IGxpYmVydMOp 47418
+IGdsYXJl 47419
+IHZlcmVpbg== 47420
+IGluc2VydHM= 47421
+IEphbmE= 47422
+IHd5ZGFqZQ== 47423
+xYJ1bQ== 47424
+ICUu 47425
+b3JpZ2luZQ== 47426
+IHN5bmFnb2d1ZQ== 47427
+IGZhbGxhaXQ= 47428
+IGRpc29iZWQ= 47429
+IGFudGlj 47430
+IEN5Y2w= 47431
+IGFzeW5jaHJvbm91cw== 47432
+IOuyjOyNqA== 47433
+IGdlc3VuZA== 47434
+IGdhZ24= 47435
+IHBlYQ== 47436
+IGdyaW4= 47437
+w6lzdA== 47438
+IHNhdWM= 47439
+IE3DpGQ= 47440
+7ZW064+E 47441
+cHBz 47442
+IM61z4DOuQ== 47443
+IHBldXBsZQ== 47444
+IGRlYmVu 47445
+IEJyZWU= 47446
+INGA0L7Qu9GM 47447
+INC60LDQutC40Lw= 47448
+IMO6dGls 47449
+IGRpc3RyaWJ1dG9y 47450
+0LDQu9GL 47451
+IHN3b2rEhQ== 47452
+IGZvbGtsb3Jl 47453
+IHJlY2VpdmVycw== 47454
+IE1PTw== 47455
+Ymlucw== 47456
+YXN0cmU= 47457
+7JWI6w== 47458
+IOuEo+qzoA== 47459
+IG11bHRpbWVkaWE= 47460
+IGdlYmF1dA== 47461
+0L7QstGL0YU= 47462
+w6N5 47463
+IGRhbmU= 47464
+b2tvbA== 47465
+ZW1pdGlzbQ== 47466
+T05FWQ== 47467
+IHlhxJ8= 47468
+IGNoYXVmZg== 47469
+IGVzZnVlcg== 47470
+xINu 47471
+ZXJ0YXM= 47472
+IGZvbmN0aW9ubmU= 47473
+b21pbmE= 47474
+IGl2b3J5 47475
+IFlvdXR1YmVy 47476
+IFNreXdhbGtlcg== 47477
+0LjRh9C10YHQutCw0Y8= 47478
+dG9p 47479
+IHZleWE= 47480
+IGdlbGVybnQ= 47481
+IGNoYW5jZWxsb3I= 47482
+IFN0YXRpc3RpY3M= 47483
+IHdlbGRlZA== 47484
+IG9uZGFu 47485
+IFNlaQ== 47486
+IG1lZGljYWxseQ== 47487
+IGVuZXJnaXplZA== 47488
+IFZpYQ== 47489
+INCy0LjQug== 47490
+IHVuaW50ZXI= 47491
+IGhpZ2huZXNz 47492
+IO2MlOs= 47493
+IGFtcGxpZmllZA== 47494
+IFNlcmdleQ== 47495
+IE1pbnM= 47496
+d2FybQ== 47497
+cGVsbA== 47498
+b3BoaWxl 47499
+IGjDqA== 47500
+IEJlbG8= 47501
+IFNrZXRjaA== 47502
+IGNoYXJhY3Rlcml6YXRpb24= 47503
+YW5zZW4= 47504
+INGC0YPRgA== 47505
+IOOFi+OFi+OFiw== 47506
+Tm90ZQ== 47507
+IGtvxZ8= 47508
+IGNpZXJ0 47509
+Zmx1 47510
+IGJhaHQ= 47511
+IERvd250b3du 47512
+IENSSVM= 47513
+b2RpZQ== 47514
+MTQw 47515
+IGxpdHJlcw== 47516
+IGdyaWV2 47517
+IOyUqOqwgA== 47518
+IHN1Y2NlZWRz 47519
+IF9f 47520
+ZW50aW5n 47521
+IHZpbW9z 47522
+IHPDrA== 47523
+ZGVmZW5zZQ== 47524
+IE1jRA== 47525
+IE1hcmlvbg== 47526
+IERvbnQ= 47527
+IEREUg== 47528
+IExhemFy 47529
+IERBUg== 47530
+IGt1dg== 47531
+S24= 47532
+IHNlbWJsYQ== 47533
+IGFpcmJvcm5l 47534
+IFZpb2xlbmNl 47535
+65CQ 47536
+IHJlc3RyYWludA== 47537
+IHdoaXN0bGVz 47538
+IHNjb2xkZWQ= 47539
+IGFjY2Vzbw== 47540
+IGFic29sdXRhbWVudGU= 47541
+IFR5bA== 47542
+IFNhcA== 47543
+toDrtoQ= 47544
+aXTDpHRlbg== 47545
+YWRlbQ== 47546
+IMO9 47547
+IHByZXNjcmliZQ== 47548
+IE1hZ2U= 47549
+IEhlbGVuYQ== 47550
+dnQ= 47551
+IHZpZW5lbg== 47552
+IHNuZWV6 47553
+IG1vbMOp 47554
+xrDhu59uZw== 47555
+IHRyYW5zcG9ydGluZw== 47556
+IExlYW4= 47557
+IGt1bmc= 47558
+0YPRgNCw 47559
+z4TOrQ== 47560
+dXRjaGVz 47561
+b25kZXJz 47562
+bGl5b3I= 47563
+TmF0 47564
+IHppag== 47565
+IG1hbW1hbA== 47566
+IGvDpHl0 47567
+IEpvYW5uYQ== 47568
+c2VudA== 47569
+IOCkuA== 47570
+IHZlc3RlZA== 47571
+IEVyZmFocnVuZw== 47572
+b2tlZQ== 47573
+IGNsaXBwaW5n 47574
+IExpc3RlbmluZw== 47575
+ICgj 47576
+ZsO2 47577
+IHZpZGFyZQ== 47578
+IGJyaXR0bGU= 47579
+IFNUQVJU 47580
+IERhbWFz 47581
+IFlvZw== 47582
+Z2FydA== 47583
+IHZlcmxpZXI= 47584
+IGhlYXJ0ZmVsdA== 47585
+IGRvxZvEhw== 47586
+7LmY6rCA 47587
+LsK7 47588
+IG1heGltYWw= 47589
+IGRpc3RpbnRvcw== 47590
+IOyZnOuDkO2VmOuptA== 47591
+IHNhaWxlZA== 47592
+IGNvbnZleWVk 47593
+IFRpbmRlcg== 47594
+IFNVUEVS 47595
+0L3QuNGG0YM= 47596
+Y29udHJvbGxlZA== 47597
+IGZ1bno= 47598
+IGJhc3RhcmRz 47599
+IEdpbnNidXJn 47600
+IG51b3Zv 47601
+IFBlcmU= 47602
+IEpFUw== 47603
+IERpbmdlbg== 47604
+IEJldHM= 47605
+dW1iYQ== 47606
+YWNjacOzbg== 47607
+IOyeiOyngOunjA== 47608
+IHJldHJh 47609
+IExhdXJlbnQ= 47610
+IHBvenk= 47611
+IGdyb292ZXM= 47612
+IG3DoXF1aW5h 47613
+IG1pbmlvbg== 47614
+IGRlaW5lbg== 47615
+IFNoYXVu 47616
+15nXmQ== 47617
+IGhvbm9yYXJ5 47618
+b3NhdXJ1cw== 47619
+IHplaXQ= 47620
+IGVzcGVjaWU= 47621
+IEJDRQ== 47622
+0LDRgtC1 47623
+SnVzdGlu 47624
+IFdoZWVscw== 47625
+IOydtO2VtA== 47626
+INio2YrZhg== 47627
+IHByb3B1bHNpb24= 47628
+IHBlcmNlYmVy 47629
+IE5ld21hbg== 47630
+5bQ= 47631
+Y3Vsb3Npcw== 47632
+TWk= 47633
+INCw0LrQutGD 47634
+IG1hc3RlcmluZw== 47635
+IGzDpGg= 47636
+IGZpc3Rz 47637
+IG1hcmluYWRl 47638
+TGlsbHk= 47639
+IOuFuOugpQ== 47640
+IFlI 47641
+IHVyZ2VudGx5 47642
+IGluZm9ybWF0aW9uYWw= 47643
+IGFjb3Jkbw== 47644
+aXp6eQ== 47645
+7J207Ja0 47646
+aW1hcg== 47647
+IOuCmOyYpOs= 47648
+IHR3ZW50aWVz 47649
+IHJhc3A= 47650
+IGJ1bXB5 47651
+2KjYqQ== 47652
+d29ya2Vy 47653
+IHF1aWNrZXN0 47654
+IGF0dGFjaGVz 47655
+0LLQuNCz 47656
+IOuCmO2DgOs= 47657
+IHB1cmVl 47658
+IG92ZXJzaXplZA== 47659
+IHN0aXJyZWQ= 47660
+IGpha2lt 47661
+IGhvbWljaWRl 47662
+aXNjaWxsYQ== 47663
+IOyxmQ== 47664
+IHNwZWN1bGF0aXZl 47665
+IGFzc2lzdHM= 47666
+bWFpbg== 47667
+asOkaHI= 47668
+aW5kZXQ= 47669
+IMWfdXI= 47670
+IGZvcmVjYXN0cw== 47671
+IGRpdmVyc2lvbg== 47672
+IHRhcmU= 47673
+IG9nbA== 47674
+IE9yZ2FuaXNhdGlvbg== 47675
+IENoZXZ5 47676
+IGJhamE= 47677
+YW5kxLFy 47678
+INmI2YTYpw== 47679
+IHJhZGlhbnQ= 47680
+IGxpYWlzb24= 47681
+IGRlbW9rcmF0 47682
+IE1BUkM= 47683
+z4DOv8+F 47684
+IHJ1bnQ= 47685
+IHByw6ljaXM= 47686
+IGdldmVu 47687
+IHbDqWhpYw== 47688
+IEpFU1M= 47689
+U1RS 47690
+IOyWmOs= 47691
+IHZpc2lvbmFyeQ== 47692
+IGJ1cmFkYW4= 47693
+IHJlYmlydGg= 47694
+IGV4aGliaXRlZA== 47695
+IE1ldGFsbA== 47696
+b2xpZQ== 47697
+ZWx5bg== 47698
+IGZsYXZvdXJz 47699
+IGVzY3JpdG8= 47700
+IERlbGV0ZQ== 47701
+IOyVjOyVmOyWtA== 47702
+INGD0LrRgNCw0ZfQvQ== 47703
+IGludGVycnVwdGluZw== 47704
+IGlkZW50aWZpYw== 47705
+IFN1enVraQ== 47706
+IExhbmRpbmc= 47707
+YW5kaQ== 47708
+IGVzdHJhbg== 47709
+IGNvdWxldXI= 47710
+IGFncmFk 47711
+IFNueQ== 47712
+IOCuh+Cusg== 47713
+IGFuZGVy 47714
+IHJ1YQ== 47715
+IHByaXNl 47716
+IGxhdXJl 47717
+IO2KgA== 47718
+IG1vZGVyYXRpb24= 47719
+IGVyZmFocmVu 47720
+IGRlY29uc3Q= 47721
+IFJlZXNl 47722
+IFBL 47723
+ZXRvcw== 47724
+IEdyYXZpdHk= 47725
+IEVyZW4= 47726
+IG92ZXJib2FyZA== 47727
+IG3DvHNzdA== 47728
+IEVtYWls 47729
+0LXRgNC8 47730
+eWRp 47731
+acSZZHp5 47732
+IExPVQ== 47733
+IEZ1w59iYWxs 47734
+IFJE 47735
+YWx0cw== 47736
+IOyKpO2KuOs= 47737
+INCa0YDQsNGB 47738
+IHRlbGV2 47739
+INGA0L4= 47740
+IHJlc2lnbmF0aW9u 47741
+IGppbmdsZQ== 47742
+IFN0dWRpZW4= 47743
+IElY 47744
+IFNlbnRpbmVs 47745
+IFBhbmc= 47746
+6YQ= 47747
+SmFrZQ== 47748
+IHBlcnNvbmFnZW0= 47749
+IG3DqWRpYQ== 47750
+IENoZXJu 47751
+YW50aWNhbGx5 47752
+IHRo4budaQ== 47753
+IHBhcmFseXNpcw== 47754
+IGphcGFuZXNl 47755
+IGNvbmV4 47756
+IGVmaWM= 47757
+IHVuZGVyc2lkZQ== 47758
+IG5lb2w= 47759
+IGZpYW4= 47760
+0LjQvNC+0YHRgtGM 47761
+IHF1aXJreQ== 47762
+IHBpc3Rh 47763
+IENsZW1lbnQ= 47764
+bm90aGluZw== 47765
+INC/0L7QtdGF 47766
+IGhvcnJlbmQ= 47767
+IGNvbnNvbGlkYXRl 47768
+cGxveXM= 47769
+ZW1ha2Vy 47770
+SmVubmlmZXI= 47771
+IG51bcOpcm8= 47772
+IGZhbW9zbw== 47773
+IE5lcHR1bmU= 47774
+IO2WiOyWtA== 47775
+INC/0YDQtdC30LjQtA== 47776
+IHNpdGNvbQ== 47777
+IHNlcmlv 47778
+IG11ZQ== 47779
+IGdsYW5kcw== 47780
+IGLDtnJqYXI= 47781
+IFlK 47782
+IFJpb3Q= 47783
+cGFyYWd1cw== 47784
+IHNlZ3VyYW7Dp2E= 47785
+IGltbWF0dXJl 47786
+IE1hZG9ubmE= 47787
+4LiN 47788
+IGxpbmdlcmluZw== 47789
+IGFjZXNzbw== 47790
+IE9yaWVudA== 47791
+IFJlY29tbQ== 47792
+IGNvbXBsYWM= 47793
+Zm91bmRlZA== 47794
+YXR0ZW5k 47795
+IGNpZWxv 47796
+IFpoYW4= 47797
+bmFpcmVz 47798
+Y2Nv 47799
+INeQ16A= 47800
+IHN0YXRh 47801
+IGNvbnRyYWRpY3Rvcnk= 47802
+IFPDqQ== 47803
+IFNBTg== 47804
+IENvbm5pZQ== 47805
+IOuLueyLnA== 47806
+INGB0LDQvNC+0Lk= 47807
+IG1hamVzdGlj 47808
+IFBlbmd1aW4= 47809
+IENPTUU= 47810
+w61jaW9z 47811
+cGVybw== 47812
+IG1n 47813
+IGZhdWM= 47814
+IGNvcnJlcg== 47815
+IEdvdHRlcw== 47816
+IEFuZ2xv 47817
+SGFy 47818
+4buXaQ== 47819
+IHZpdGVzc2U= 47820
+IGFubm91bmNlcg== 47821
+IE9tYWhh 47822
+a3Vt 47823
+IHNwYXJlZA== 47824
+INGA0LDQt9Cw 47825
+INC/0L7Qu9GD0YfQuNGC0YHRjw== 47826
+IHTDpGjDpG4= 47827
+INC/0L7QvdCw0LQ= 47828
+IHBlcnRhaW5pbmc= 47829
+IFJhdGU= 47830
+aWVybg== 47831
+R29sZA== 47832
+IHRlc3Rl 47833
+IGRlxJ9pbGQ= 47834
+IGRhbXBpbmc= 47835
+IFBhcnRuZXJzaGlw 47836
+enlzdGE= 47837
+Z2VsZA== 47838
+IHNtb2tlcw== 47839
+IE1hcnJpYWdl 47840
+7Kq97JeQ 47841
+aXNjZQ== 47842
+IHRyeW5h 47843
+IERpcmVjdG9yeQ== 47844
+IOuCmOyYrA== 47845
+IHNoYW1lZnVs 47846
+IG1lbnRyZQ== 47847
+IGFzc2lnbmluZw== 47848
+IHJlcGVydG9pcmU= 47849
+IG9iamV0b3M= 47850
+IHVuZGVyd29ybGQ= 47851
+IGVuZGVhdm9ycw== 47852
+IGlnbml0ZQ== 47853
+INmI2Kw= 47854
+IGV4cGVyaWVudA== 47855
+INCX0LDQvw== 47856
+INC30LDQutC70Y7Rhw== 47857
+IHZvbHRhZ2Vz 47858
+IG5pZWdv 47859
+IGRlZmljaXRz 47860
+IGJ1ZW5vcw== 47861
+IFNsZWVwaW5n 47862
+IFNhbGVt 47863
+IHVubG9ja2luZw== 47864
+IGludGVyYWN0ZWQ= 47865
+IGVudGVuZGV1 47866
+IFN1cGVyaW50ZW5kZW50 47867
+IHN6Y3plZ8OzbA== 47868
+IHF1YXM= 47869
+IHBhbGluZw== 47870
+IGtobw== 47871
+2KjYrQ== 47872
+IGNvbGFib3I= 47873
+INC/0YDQuNCz0L7RgtC+0LI= 47874
+IG1hdXY= 47875
+IEp1ZGFz 47876
+IEFzc2lzdA== 47877
+INGC0LXRgNGA0Lg= 47878
+INC90LDRgdC60L7Qu9GM0LrQvg== 47879
+IHN1YnNpZHk= 47880
+IEVtYmFzc3k= 47881
+IGRhZ2Vu 47882
+IFNhbnRv 47883
+16nXldeR 47884
+IGFicnVwdGx5 47885
+IEFkYXB0 47886
+IHZhYWs= 47887
+IHBvc3RhbA== 47888
+IGludmVzdGly 47889
+IGZpcXVlaQ== 47890
+IGRvd250aW1l 47891
+IFdlYmI= 47892
+IE5DQUE= 47893
+IEVzdG95 47894
+0L7Qu9C+0YI= 47895
+IOyCrOqxtA== 47896
+IG5hdGlvbmFsaXN0 47897
+IEthdGhyeW4= 47898
+IEtvcA== 47899
+6ao= 47900
+U2Vhbg== 47901
+T05B 47902
+IEJq 47903
+16LXnQ== 47904
+w61i 47905
+aWRhbWVudGU= 47906
+INCz0LvQsNC30LA= 47907
+IHVubmll 47908
+IGdlbWFha3Q= 47909
+IElOVEVSVklFV0VS 47910
+IEhhdXQ= 47911
+zq/Ovw== 47912
+Z2VvaXM= 47913
+d3lkZA== 47914
+INC60L7Qu9C4 47915
+IHRpZ2h0ZW5lZA== 47916
+IHBsYW5uZXJz 47917
+IGhlcnVt 47918
+IGfDtnLDvG4= 47919
+IGVsZWN0cm9uaWNhbGx5 47920
+IGNlcmFt 47921
+IOuLpOyWke2VnA== 47922
+IGVwaWxlcHN5 47923
+IGXEnw== 47924
+bGlucw== 47925
+IFNoaW55 47926
+INGB0L7Qu9C9 47927
+IG1hY2Fyb24= 47928
+IGltcGFjdG8= 47929
+IFZlZ2Fu 47930
+emXFhA== 47931
+IFJhcGhh 47932
+IFBhcnM= 47933
+IExFTw== 47934
+Y8O8 47935
+INec15TXmdeV16o= 47936
+IMOkaG5saWNo 47937
+IGZsb3Nz 47938
+IEFa 47939
+IG3DtmNodGVu 47940
+IGdyb29taW5n 47941
+IGdyYXNzZXM= 47942
+cmFuY2g= 47943
+IHJlY2liaXI= 47944
+IGJvdW5jeQ== 47945
+IEhvYmJ5 47946
+IHZpa3RpZw== 47947
+IGJlZ2l0dQ== 47948
+IFBpY2Fzc28= 47949
+IEt1c2g= 47950
+66qo 47951
+IG9ic3RydWN0aW9u 47952
+IOu2hOychA== 47953
+IG1pY3JvYg== 47954
+IFdlc3RtaW5zdGVy 47955
+cm9wcw== 47956
+ZHVs 47957
+IGRldm8= 47958
+IExlaHJlcg== 47959
+IEFkdmlzb3I= 47960
+dWNrZW4= 47961
+INCx0YPQvA== 47962
+IGZsYXR0ZXJpbmc= 47963
+IFRydW1hbg== 47964
+IFNlbXByZQ== 47965
+IE1jQ2Fpbg== 47966
+IEhpbmR1cw== 47967
+SnVsaWE= 47968
+IHdhdGVyc2hlZA== 47969
+IGx1c2g= 47970
+7KCE6w== 47971
+QmVmb3Jl 47972
+INCS0YLQvtGA 47973
+IFNhYVM= 47974
+IHNpdHp0 47975
+IGJlZXRsZQ== 47976
+IEVzc2VudGlhbA== 47977
+ZW5rbw== 47978
+IOuVjOuPhA== 47979
+IHJldnZpbmc= 47980
+IHBvb3Jlcg== 47981
+IGNvZXJj 47982
+IGlkZWU= 47983
+IGNvw7s= 47984
+YWxldA== 47985
+IHpkcm93 47986
+IGZlbmRlcg== 47987
+Z3Jvd3Ro 47988
+RElORw== 47989
+IHpkZQ== 47990
+RU5UUw== 47991
+IGZhY2V0cw== 47992
+dXNoaW1h 47993
+IMWfZWg= 47994
+IHBhcmFzaXRl 47995
+IGxhcHNl 47996
+IE1lZXI= 47997
+IEt1bmQ= 47998
+IHNsb2c= 47999
+IGJydW5jaA== 48000
+IENoYXJ0 48001
+YXJ6 48002
+IE1VUw== 48003
+IG9mZmVuc2Vz 48004
+IGluZ2zDqXM= 48005
+IGZvbGlhZ2U= 48006
+b3BsYW4= 48007
+QXV0 48008
+IEphY3F1 48009
+dGFr 48010
+aWVtYnJl 48011
+IHhlbg== 48012
+IG5vbWluZWVz 48013
+IGJpb21lZGljYWw= 48014
+w6lzdXM= 48015
+IGVzdHV2 48016
+z4TPjA== 48017
+QVRIQU4= 48018
+IO2VnOuNsA== 48019
+IGhlZWQ= 48020
+Y3Jvc3N0YWxr 48021
+QmlsbA== 48022
+IHNwb3VzZXM= 48023
+INGB0Y7Qtg== 48024
+IHZlcnNv 48025
+IFN2ZW4= 48026
+IENhdQ== 48027
+Y3V6 48028
+IOuztOyEuOyalA== 48029
+INGF0L7Qt9GP 48030
+IG1vY2tpbmc= 48031
+IE9uYQ== 48032
+IETDoQ== 48033
+IGZydWl0ZnVs 48034
+IGJhbnF1ZXQ= 48035
+dWRkaW5n 48036
+aW5jdGlvbnM= 48037
+ZGVydA== 48038
+c3Vk 48039
+IGRlc2Nvbg== 48040
+IEpD 48041
+IMKn 48042
+IHB1Ymxp 48043
+64iI 48044
+IGVudHNjaGllZGVu 48045
+IFJPSQ== 48046
+IOyDneqyvA== 48047
+IGvDpHl0dA== 48048
+eWFuaQ== 48049
+c2hhdw== 48050
+IHVubGVhc2g= 48051
+IG1hbm5l 48052
+IGhpc3RvZ3JhbQ== 48053
+4Lit4Liw4LmE4Lij 48054
+IGdu 48055
+IGZlbGxh 48056
+IGVpbmdlcw== 48057
+IEJ1aWx0 48058
+IHJlcHJlc2VudGE= 48059
+IHB1bmlzaGluZw== 48060
+IG91dHNpZGVycw== 48061
+0L3Rg9GC0YzRgdGP 48062
+Y3VycmVudA== 48063
+IGZhbWlsaWFyaXR5 48064
+INC00LjQsg== 48065
+IHByb2pldHM= 48066
+IGFxdWVsZXM= 48067
+IEdsdWU= 48068
+dGhvc2U= 48069
+IGluY2VwdGlvbg== 48070
+IGFxdWVsbG9z 48071
+IGlsbHVzaW9ucw== 48072
+IGF0dGVuZHM= 48073
+cmVzZQ== 48074
+IHN3YXJt 48075
+IHN3YWI= 48076
+IHJlZ2FyZGV6 48077
+IHBvc2nDp8Ojbw== 48078
+IGFraGly 48079
+IGV4dHJhY3Rpbmc= 48080
+IGFuZWNkb3Rl 48081
+IFRhbGU= 48082
+INCy0LjQvQ== 48083
+IGFiZ2Vz 48084
+IG9sdcWf 48085
+IGNvbXBsaWNhZG8= 48086
+IGNvdmFyaQ== 48087
+0ZbRgtGM 48088
+RGVy 48089
+INeZ15Q= 48090
+Rm9ybQ== 48091
+IOyWtOyojOuToA== 48092
+IHJlYWRhYmxl 48093
+IGluaGliaXQ= 48094
+IGRlY2lwaGVy 48095
+IEFuZ3J5 48096
+cGc= 48097
+4K614K6k 48098
+INGB0L7QsdGB0YLQstC10L3QvdC+ 48099
+IHNhbWg= 48100
+IGVzY3I= 48101
+IGVuY29tcGFzc2Vz 48102
+IGF1c3Rlcg== 48103
+IGNvbmZpc2M= 48104
+IE1hbmRhbA== 48105
+IH0= 48106
+YXRjaGVy 48107
+PSM= 48108
+INC60LjQvdC+ 48109
+IHN0YWw= 48110
+bHVuZ3M= 48111
+IHZvbGU= 48112
+IHJlcXVpcw== 48113
+IHDDqW4= 48114
+IGxlY3R1cmVy 48115
+IGluc2NyaXB0aW9u 48116
+IGNlcnZpY2Fs 48117
+IFRyZWFzdXJl 48118
+IEpX 48119
+Y29taW5ncw== 48120
+IGV5ZXNpZ2h0 48121
+IFRhaWxz 48122
+w61zaW1v 48123
+IHdvcmtzaGVldA== 48124
+IHN3aWZ0bHk= 48125
+IGNvbm9z 48126
+IGVsaW1pbmF0ZXM= 48127
+IEJsYXpl 48128
+0LDQu9C+0LM= 48129
+IHBpY3R1cmVk 48130
+IGdpcmFmZmU= 48131
+IExvZ2lj 48132
+IGVucmljaG1lbnQ= 48133
+Rml0 48134
+IHVuaW50ZW5kZWQ= 48135
+IHBlcnNlY3V0ZWQ= 48136
+YWthcA== 48137
+67CY 48138
+IGJhcmJlcg== 48139
+IGFyYmVpdGV0 48140
+IFN1cnByaXNpbmdseQ== 48141
+IEF1dG9i 48142
+dW5rdQ== 48143
+cHJvdg== 48144
+IExvY2g= 48145
+b2J5bA== 48146
+INC/0L7QtNCz0L7RgtC+0LI= 48147
+IMOpY29ub21pcXVl 48148
+IHBhdHQ= 48149
+IGNlYXNlZA== 48150
+INGB0L/QuNGB 48151
+IG51Y2xlaQ== 48152
+IGlzdGU= 48153
+IFdhZw== 48154
+IHp1cGXFgm5pZQ== 48155
+IHByb3ZlcmI= 48156
+IEFow60= 48157
+bGlhbW8= 48158
+IHJlbGlhYmx5 48159
+IHBpaw== 48160
+IFRyYWRpbmc= 48161
+IENvbGVtYW4= 48162
+IM6xzr3OsQ== 48163
+IG1hZ2FyaQ== 48164
+IFBISUw= 48165
+IHNoZWRkaW5n 48166
+b2huZXI= 48167
+IHBvcm5vZ3JhcGh5 48168
+IGJlbmVmaWNpYXJpZXM= 48169
+4oCi 48170
+ZW5pbg== 48171
+IHJlc29sdmluZw== 48172
+INGB0L/QvtGA0YI= 48173
+INCx0LXQsw== 48174
+IG5lY3Rhcg== 48175
+dWx0dXJh 48176
+aW1zaWNhbA== 48177
+jIDrpbw= 48178
+IHZpc8Ojbw== 48179
+w7/Dv8O/w7/Dv8O/w7/Dvw== 48180
+YXR0Zm9ybQ== 48181
+IOunnuuKlA== 48182
+IHBpbGdyaW1hZ2U= 48183
+IG1hdGluZw== 48184
+IFJlYXBlcg== 48185
+IEJyZWY= 48186
+INeR15M= 48187
+IG5vdmFtZW50ZQ== 48188
+IGdyaWxsaW5n 48189
+IFdpcmVsZXNz 48190
+IFJvbWFuaWFu 48191
+0ps= 48192
+7Jyg6w== 48193
+aGFpdA== 48194
+IEJvcmE= 48195
+QVJSWQ== 48196
+IGh5cG90aGVzZXM= 48197
+aWt1dA== 48198
+IOyVhOuyhA== 48199
+INGW0Lc= 48200
+IG5hdGlvbmFsZQ== 48201
+2KrZiQ== 48202
+w7xsbHQ= 48203
+IMOpbMOpbWVudHM= 48204
+IFdhcmU= 48205
+ICgt 48206
+0LDQu9GM0L3QvtC8 48207
+IGluZGljdA== 48208
+IFN0b25lcw== 48209
+ZXhwbG9zaW9u 48210
+IOuDhOyDiA== 48211
+IGZlbGlj 48212
+IGp1ZGljaWFyeQ== 48213
+IGluY2FybmF0aW9u 48214
+IGlubmluZw== 48215
+IGZvcm11bA== 48216
+IHNoaXBtZW50 48217
+IHJlaW5kZWVy 48218
+INC+0LfQvdCw0Yc= 48219
+IGVudm9s 48220
+dW5keQ== 48221
+INC30L3QsNGC0Yw= 48222
+INCy0LjQtNC10LvQuA== 48223
+IGV4Y2x1ZGluZw== 48224
+ZGVhdGg= 48225
+IGJlcm0= 48226
+IHNvcHJhdHR1dHRv 48227
+IGRlYmlkbw== 48228
+IFppZw== 48229
+IE92 48230
+IEtFVklO 48231
+IFBhbGU= 48232
+IE1pcmU= 48233
+IGFuZGFy 48234
+aW5jbHVkaW5n 48235
+IHN3YXBwZWQ= 48236
+IG1pc2NvbmNlcHRpb25z 48237
+IHNwb25n 48238
+csOpYWw= 48239
+IG9yYml0YWxz 48240
+IGhhc2h0YWdz 48241
+b3JpdA== 48242
+IG1hdXZhaXM= 48243
+0LjRgdCw 48244
+IGxpdnJlcw== 48245
+IElQUw== 48246
+IDA0 48247
+w7Zn 48248
+aW5zdHI= 48249
+INCy0L3QtdGI 48250
+IGhpY2U= 48251
+aXPDqWU= 48252
+IG93ZXM= 48253
+IGVzaW1lcms= 48254
+IFVI 48255
+IGlycml0YXRpb24= 48256
+IGdpZ2dsZXM= 48257
+IGNvbG9uaWFsaXNt 48258
+IEJsaXNz 48259
+c3RyaW5ncw== 48260
+IHJldW5pdGVk 48261
+IFBzYWtp 48262
+d2FjaA== 48263
+IGNsaWZmcw== 48264
+IEZhbHNl 48265
+w6Rn 48266
+cGlwZQ== 48267
+IHdob3BwaW5n 48268
+IG1lcmluZ3Vl 48269
+IGJ1bmc= 48270
+aW5kdXN0cmll 48271
+IGxlY2hl 48272
+IExveQ== 48273
+IGRyaWU= 48274
+IHBhc3NhdA== 48275
+IG9sZWg= 48276
+IGPDqXU= 48277
+IEdhYnJpZQ== 48278
+IHJlZWZz 48279
+IGJvbWJlcnM= 48280
+IGVwaXPDs2Rpbw== 48281
+IFJ1Zw== 48282
+IFByb3Nl 48283
+b25vcw== 48284
+IG9iZXNl 48285
+IGdvb2c= 48286
+IHBpYWNl 48287
+Zmxhbnplbg== 48288
+IGZsYXBz 48289
+IEFsdG8= 48290
+Rmlu 48291
+IHJlc2l6ZQ== 48292
+6re4656o 48293
+TmF0aGFu 48294
+nojroKQ= 48295
+INGC0LDQuQ== 48296
+IE5GVA== 48297
+IHNuZWV6ZQ== 48298
+IHNocm91ZA== 48299
+acOp 48300
+IHZlcmFtZW50ZQ== 48301
+IGNhc2NhZGU= 48302
+IE9vaw== 48303
+7JeG7J20 48304
+IGluZnVzZWQ= 48305
+ZnBz 48306
+Y2VudGVy 48307
+IGdyYXBwbGluZw== 48308
+IFdvaG51bmc= 48309
+IFR1bWI= 48310
+IEltbWE= 48311
+IER1eWd1c2Fs 48312
+0LXQvdGC0Lg= 48313
+IHN0ZXdhcmRzaGlw 48314
+IGhhcnA= 48315
+IGVuZG9yc2Vk 48316
+xLFsYW4= 48317
+INC+0LTQvdC40Lw= 48318
+IGNvbXBldGVuY3k= 48319
+IGJlcnQ= 48320
+IFRhbGVz 48321
+IHJoZQ== 48322
+IG9oaA== 48323
+IOqwhOuLqA== 48324
+IG1STkE= 48325
+IGdhbmdzdGVy 48326
+IFJ1bm5lcg== 48327
+0LXQvdC90YvQvA== 48328
+cGhvcmlh 48329
+IHfFgmHFm2Npd2ll 48330
+IHF1YXJ0bw== 48331
+IG9yZ2FuaXNl 48332
+IFZldA== 48333
+UGFk 48334
+INmF2Ks= 48335
+IHN0aW5rcw== 48336
+IER1bA== 48337
+dWVt 48338
+aXNpZWo= 48339
+VG9w 48340
+IHR1c3Nlbg== 48341
+IEVmZW5kaW1peg== 48342
+IEJvdWxl 48343
+IFNsb3Zlbg== 48344
+IEzDtg== 48345
+0ZHQtw== 48346
+0YDQuNC/ 48347
+Y2F2ZQ== 48348
+IGJvw64= 48349
+IGFwb2xvZ2lzZQ== 48350
+IE1hcmx5 48351
+IEV4cG9ydA== 48352
+IENhaXRsaW4= 48353
+IHRhdmFsbGE= 48354
+IGVudGFpbHM= 48355
+IGJyb20= 48356
+IENvcGVuaA== 48357
+IHdhbG51dA== 48358
+IGluc2lzdHM= 48359
+IGN14buZYw== 48360
+IFF1aXQ= 48361
+IERldmljZQ== 48362
+15LXnQ== 48363
+IERPVA== 48364
+IHZlbG9jaWRhZA== 48365
+TElF 48366
+Q29vbA== 48367
+IHNhbml0YXRpb24= 48368
+IG9saG8= 48369
+IEVC 48370
+IO2ZleyLpO2eiA== 48371
+INCc0LjRhQ== 48372
+IHp1aw== 48373
+IHN1cm5hbWU= 48374
+IFNjaHVsZA== 48375
+cnVmZg== 48376
+Y3VsdHVyYWw= 48377
+INGB0YLQvtC70YzQutC+ 48378
+jOuNsA== 48379
+IHRvcnRv 48380
+IGJhY2t1cHM= 48381
+0YDQuNC5 48382
+cmVsYXg= 48383
+IHN5bmVyZ3k= 48384
+IGJ1ZmZz 48385
+IGFwbw== 48386
+IFdlbGxuZXNz 48387
+cm91bmRlZA== 48388
+IHVuaXZlcnNlcw== 48389
+IGZlcmE= 48390
+IHN0YW5kYnk= 48391
+IFNpbHZh 48392
+IEpJ 48393
+ZW5zb3JlZA== 48394
+IOyXhuuLpA== 48395
+INCQ0LI= 48396
+INC+0YLQtNC10Ls= 48397
+IGbDuA== 48398
+IFJvY2tlZg== 48399
+IENvbXBhc3M= 48400
+IEJlYXJz 48401
+VHVybg== 48402
+IHRo4buxYw== 48403
+IHBvc3NpYmlsZQ== 48404
+IGVzdGVt 48405
+IENyb2F0aWE= 48406
+IHTDpHTDpA== 48407
+IENBTA== 48408
+4LmA4Lie 48409
+INGB0YLRgNCw0YU= 48410
+IHNhbHRz 48411
+IG1pbmltYWxpc3Q= 48412
+IGluY29ycG9yYXRlcw== 48413
+INmG24HbjNq6 48414
+YWNhbw== 48415
+IHNsYW1tZWQ= 48416
+IGNhbWE= 48417
+VGV4dA== 48418
+ISEhISEh 48419
+IGFsY2Fueg== 48420
+w6ltYQ== 48421
+IGluY2Vuc2U= 48422
+IGhhcmRlbg== 48423
+IGdyYW50aW5n 48424
+IE5haQ== 48425
+IEZpcm1h 48426
+IGh5cG9j 48427
+am9i 48428
+IFJI 48429
+enVy 48430
+0LjQu9GP 48431
+IMW6 48432
+IGRhcmVz 48433
+YW5o 48434
+IOunjO2BvA== 48435
+IGN1ZXN0acOzbg== 48436
+IExpbWE= 48437
+IGFzc3VudG8= 48438
+IElQTw== 48439
+IEJlbmdhbA== 48440
+IEJpZXI= 48441
+IHBzeWNoZQ== 48442
+IGFjcXVhaW50ZWQ= 48443
+IEfDvG4= 48444
+0L7Qt9C4 48445
+xZtjacSF 48446
+QUc= 48447
+IG1hbGZ1bmN0aW9u 48448
+IGFzdGVyb2lkcw== 48449
+aXJleg== 48450
+YW1vcnBo 48451
+INGB0L7RgtGA0YPQtA== 48452
+IGZyZXNod2F0ZXI= 48453
+IGFycmFu 48454
+INC/0YDRiw== 48455
+0L3QvtCz 48456
+IGRpYWJldGlj 48457
+INmC2KfZhA== 48458
+IG9wcHJlc3M= 48459
+IGNhcGFjaXRhbmNl 48460
+cGVyZm9ybWFuY2U= 48461
+Y3JhdGVz 48462
+IGFwb3N0bGU= 48463
+IEpFTg== 48464
+T1VMRA== 48465
+SW50cm8= 48466
+IHN0YWxscw== 48467
+IEFCT1VU 48468
+Y3RpY2FtZW50ZQ== 48469
+IGRpbGlnZW50 48470
+IG1hbmlmZXN0cw== 48471
+IFBha2lzdGFuaQ== 48472
+ICgn 48473
+= 48474
+6ZM= 48475
+6bI= 48476
+55g= 48477
+6Jw= 48478
+6bg= 48479
+6a4= 48480
+6Jo= 48481
+6J0= 48482
+6a8= 48483
+6JU= 48484
+6KQ= 48485
+6bA= 48486
+6J8= 48487
+56M= 48488
+6KA= 48489
+6ZE= 48490
+6bU= 48491
+6bE= 48492
+6bQ= 48493
+5bU= 48494
+54A= 48495
+6Z4= 48496
+6bY= 48497
+6Y4= 48498
+6ac= 48499
+6bc= 48500
+5aw= 48501
+6Ys= 48502
+5bY= 48503
+5qs= 48504
+55M= 48505
+77w= 48506
+46k= 48507
+770= 48508
+45Y= 48509
+45c= 48510
+5IE= 48511
+8KA= 48512
+4oCY 48513
+4oCZ 48514
+4oCc 48515
+4oCd 48516
+4oCn 48517
+4oSD 48518
+4peL 48519
+44CD 48520
+44CG 48521
+44CH 48522
+44CI 48523
+44CJ 48524
+44CS 48525
+44Cd 48526
+44Ce 48527
+44GB 48528
+44GC 48529
+44GD 48530
+44GE 48531
+44GF 48532
+44GG 48533
+44GH 48534
+44GI 48535
+44GJ 48536
+44GK 48537
+44GL 48538
+44GM 48539
+44GN 48540
+44GO 48541
+44GP 48542
+44GQ 48543
+44GR 48544
+44GS 48545
+44GT 48546
+44GU 48547
+44GV 48548
+44GW 48549
+44GX 48550
+44GY 48551
+44GZ 48552
+44Ga 48553
+44Gb 48554
+44Gc 48555
+44Gd 48556
+44Ge 48557
+44Gf 48558
+44Gg 48559
+44Gh 48560
+44Gi 48561
+44Gj 48562
+44Gk 48563
+44Gl 48564
+44Gm 48565
+44Gn 48566
+44Go 48567
+44Gp 48568
+44Gq 48569
+44Gr 48570
+44Gs 48571
+44Gt 48572
+44Gu 48573
+44Gv 48574
+44Gw 48575
+44Gx 48576
+44Gy 48577
+44Gz 48578
+44G0 48579
+44G1 48580
+44G2 48581
+44G3 48582
+44G4 48583
+44G5 48584
+44G6 48585
+44G7 48586
+44G8 48587
+44G9 48588
+44G+ 48589
+44G/ 48590
+44KA 48591
+44KB 48592
+44KC 48593
+44KD 48594
+44KE 48595
+44KF 48596
+44KG 48597
+44KH 48598
+44KI 48599
+44KJ 48600
+44KK 48601
+44KL 48602
+44KM 48603
+44KN 48604
+44KP 48605
+44KQ 48606
+44KR 48607
+44KS 48608
+44KT 48609
+44Kd 48610
+44Ke 48611
+44Kh 48612
+44Ki 48613
+44Kj 48614
+44Kk 48615
+44Kl 48616
+44Km 48617
+44Kn 48618
+44Ko 48619
+44Kp 48620
+44Kq 48621
+44Kr 48622
+44Ks 48623
+44Kt 48624
+44Ku 48625
+44Kv 48626
+44Kw 48627
+44Kx 48628
+44Ky 48629
+44Kz 48630
+44K0 48631
+44K1 48632
+44K2 48633
+44K3 48634
+44K4 48635
+44K5 48636
+44K6 48637
+44K7 48638
+44K8 48639
+44K9 48640
+44K+ 48641
+44K/ 48642
+44OA 48643
+44OB 48644
+44OC 48645
+44OD 48646
+44OE 48647
+44OF 48648
+44OG 48649
+44OH 48650
+44OI 48651
+44OJ 48652
+44OK 48653
+44OL 48654
+44OM 48655
+44ON 48656
+44OO 48657
+44OP 48658
+44OQ 48659
+44OR 48660
+44OS 48661
+44OT 48662
+44OU 48663
+44OV 48664
+44OW 48665
+44OX 48666
+44OY 48667
+44OZ 48668
+44Oa 48669
+44Ob 48670
+44Oc 48671
+44Od 48672
+44Oe 48673
+44Of 48674
+44Og 48675
+44Oh 48676
+44Oi 48677
+44Oj 48678
+44Ok 48679
+44Ol 48680
+44Om 48681
+44On 48682
+44Oo 48683
+44Op 48684
+44Oq 48685
+44Or 48686
+44Os 48687
+44Ot 48688
+44Ou 48689
+44Ov 48690
+44Ow 48691
+44Ox 48692
+44Oy 48693
+44Oz 48694
+44O0 48695
+44O1 48696
+44O2 48697
+44O7 48698
+44O8 48699
+44O+ 48700
+45at 48701
+45eO 48702
+46mS 48703
+46mn 48704
+5IGv 48705
+5LiA 48706
+5LiB 48707
+5LiD 48708
+5LiH 48709
+5LiI 48710
+5LiJ 48711
+5LiK 48712
+5LiL 48713
+5LiN 48714
+5LiO 48715
+5LiQ 48716
+5LiR 48717
+5LiT 48718
+5LiU 48719
+5LiV 48720
+5LiW 48721
+5LiX 48722
+5LiY 48723
+5LiZ 48724
+5Lia 48725
+5Lib 48726
+5Lic 48727
+5Lid 48728
+5Lie 48729
+5Lif 48730
+5Lih 48731
+5Lii 48732
+5Lik 48733
+5Lil 48734
+5Lim 48735
+5Lin 48736
+5Lio 48737
+5Liq 48738
+5Lir 48739
+5Lit 48740
+5Liw 48741
+5Liy 48742
+5Li0 48743
+5Li2 48744
+5Li4 48745
+5Li5 48746
+5Li6 48747
+5Li7 48748
+5Li8 48749
+5Li9 48750
+5Li+ 48751
+5LmC 48752
+5LmD 48753
+5LmF 48754
+5LmH 48755
+5LmI 48756
+5LmJ 48757
+5LmL 48758
+5LmM 48759
+5LmN 48760
+5LmO 48761
+5LmP 48762
+5LmQ 48763
+5LmS 48764
+5LmT 48765
+5LmU 48766
+5LmW 48767
+5LmX 48768
+5LmY 48769
+5LmZ 48770
+5Lmc 48771
+5Lmd 48772
+5Lme 48773
+5Lmf 48774
+5Lmg 48775
+5Lmh 48776
+5Lmi 48777
+5Lmm 48778
+5Lmp 48779
+5Lmq 48780
+5Lmw 48781
+5Lmx 48782
+5Lmz 48783
+5Lm4 48784
+5Lm+ 48785
+5LqA 48786
+5LqB 48787
+5LqC 48788
+5LqG 48789
+5LqI 48790
+5LqJ 48791
+5LqL 48792
+5LqM 48793
+5LqN 48794
+5LqO 48795
+5LqP 48796
+5LqR 48797
+5LqS 48798
+5LqT 48799
+5LqU 48800
+5LqV 48801
+5LqY 48802
+5LqZ 48803
+5Lqa 48804
+5Lqb 48805
+5Lqc 48806
+5Lqe 48807
+5Lqf 48808
+5Lqh 48809
+5Lqi 48810
+5Lqk 48811
+5Lql 48812
+5Lqm 48813
+5Lqn 48814
+5Lqo 48815
+5Lqp 48816
+5Lqr 48817
+5Lqs 48818
+5Lqt 48819
+5Lqu 48820
+5Lqw 48821
+5Lqy 48822
+5Lqz 48823
+5Lq1 48824
+5Lq2 48825
+5Lq5 48826
+5Lq6 48827
+5Lq/ 48828
+5LuA 48829
+5LuB 48830
+5LuC 48831
+5LuD 48832
+5LuE 48833
+5LuF 48834
+5LuG 48835
+5LuH 48836
+5LuJ 48837
+5LuK 48838
+5LuL 48839
+5LuN 48840
+5LuO 48841
+5LuP 48842
+5LuR 48843
+5LuT 48844
+5LuU 48845
+5LuV 48846
+5LuW 48847
+5LuX 48848
+5LuY 48849
+5LuZ 48850
+5Lud 48851
+5Lue 48852
+5Luf 48853
+5Luh 48854
+5Luj 48855
+5Luk 48856
+5Lul 48857
+5Luo 48858
+5Luq 48859
+5Lur 48860
+5Lus 48861
+5Lut 48862
+5Luu 48863
+5Luw 48864
+5Luy 48865
+5Luz 48866
+5Lu1 48867
+5Lu2 48868
+5Lu3 48869
+5Lu7 48870
+5Lu9 48871
+5Lu/ 48872
+5LyB 48873
+5LyD 48874
+5LyE 48875
+5LyJ 48876
+5LyK 48877
+5LyL 48878
+5LyN 48879
+5LyO 48880
+5LyP 48881
+5LyQ 48882
+5LyR 48883
+5LyX 48884
+5LyY 48885
+5LyZ 48886
+5Lya 48887
+5Lyb 48888
+5Lyc 48889
+5Lyd 48890
+5Lye 48891
+5Lyf 48892
+5Lyg 48893
+5Lyi 48894
+5Lyk 48895
+5Lyl 48896
+5Lym 48897
+5Lyn 48898
+5Lyq 48899
+5Lyr 48900
+5Lyv 48901
+5Lyw 48902
+5Lyx 48903
+5Lyy 48904
+5Ly0 48905
+5Ly2 48906
+5Ly3 48907
+5Ly4 48908
+5Ly6 48909
+5Ly8 48910
+5Ly9 48911
+5L2D 48912
+5L2G 48913
+5L2H 48914
+5L2I 48915
+5L2J 48916
+5L2N 48917
+5L2O 48918
+5L2P 48919
+5L2Q 48920
+5L2R 48921
+5L2T 48922
+5L2U 48923
+5L2V 48924
+5L2X 48925
+5L2Y 48926
+5L2Z 48927
+5L2a 48928
+5L2b 48929
+5L2c 48930
+5L2d 48931
+5L2e 48932
+5L2f 48933
+5L2g 48934
+5L2i 48935
+5L2j 48936
+5L2k 48937
+5L2l 48938
+5L2p 48939
+5L2s 48940
+5L2v 48941
+5L2w 48942
+5L2z 48943
+5L21 48944
+5L22 48945
+5L23 48946
+5L26 48947
+5L27 48948
+5L28 48949
+5L2+ 48950
+5L2/ 48951
+5L6C 48952
+5L6D 48953
+5L6E 48954
+5L6G 48955
+5L6I 48956
+5L6J 48957
+5L6L 48958
+5L6N 48959
+5L6P 48960
+5L6R 48961
+5L6T 48962
+5L6U 48963
+5L6X 48964
+5L6Y 48965
+5L6b 48966
+5L6d 48967
+5L6g 48968
+5L6h 48969
+5L6j 48970
+5L6l 48971
+5L6m 48972
+5L6n 48973
+5L6o 48974
+5L6p 48975
+5L6q 48976
+5L6s 48977
+5L6t 48978
+5L6u 48979
+5L6v 48980
+5L61 48981
+5L62 48982
+5L63 48983
+5L6/ 48984
+5L+C 48985
+5L+D 48986
+5L+E 48987
+5L+F 48988
+5L+K 48989
+5L+O 48990
+5L+P 48991
+5L+Q 48992
+5L+R 48993
+5L+U 48994
+5L+X 48995
+5L+Y 48996
+5L+a 48997
+5L+b 48998
+5L+c 48999
+5L+d 49000
+5L+e 49001
+5L+f 49002
+5L+g 49003
+5L+h 49004
+5L+i 49005
+5L+j 49006
+5L+k 49007
+5L+m 49008
+5L+o 49009
+5L+p 49010
+5L+q 49011
+5L+s 49012
+5L+t 49013
+5L+u 49014
+5L+v 49015
+5L+x 49016
+5L+z 49017
+5L+1 49018
+5L+2 49019
+5L+4 49020
+5L+6 49021
+5L++ 49022
+5YCF 49023
+5YCG 49024
+5YCJ 49025
+5YCL 49026
+5YCM 49027
+5YCN 49028
+5YCP 49029
+5YCR 49030
+5YCS 49031
+5YCT 49032
+5YCU 49033
+5YCW 49034
+5YCY 49035
+5YCZ 49036
+5YCa 49037
+5YCc 49038
+5YCf 49039
+5YCh 49040
+5YCi 49041
+5YCj 49042
+5YCk 49043
+5YCl 49044
+5YCm 49045
+5YCn 49046
+5YCo 49047
+5YCp 49048
+5YCq 49049
+5YCr 49050
+5YCs 49051
+5YCt 49052
+5YCu 49053
+5YC2 49054
+5YC5 49055
+5YC6 49056
+5YC8 49057
+5YC+ 49058
+5YGB 49059
+5YGD 49060
+5YGH 49061
+5YGI 49062
+5YGJ 49063
+5YGM 49064
+5YGO 49065
+5YGP 49066
+5YGT 49067
+5YGV 49068
+5YGW 49069
+5YGa 49070
+5YGc 49071
+5YGl 49072
+5YGs 49073
+5YGt 49074
+5YGx 49075
+5YGy 49076
+5YG0 49077
+5YG1 49078
+5YG2 49079
+5YG3 49080
+5YG4 49081
+5YG7 49082
+5YG9 49083
+5YG+ 49084
+5YG/ 49085
+5YKA 49086
+5YKF 49087
+5YKI 49088
+5YKN 49089
+5YKR 49090
+5YKV 49091
+5YKY 49092
+5YKZ 49093
+5YKc 49094
+5YKi 49095
+5YKj 49096
+5YKl 49097
+5YKn 49098
+5YKo 49099
+5YKp 49100
+5YKr 49101
+5YKs 49102
+5YKt 49103
+5YKy 49104
+5YKz 49105
+5YK0 49106
+5YK1 49107
+5YK3 49108
+5YK6 49109
+5YK7 49110
+5YK+ 49111
+5YOC 49112
+5YOF 49113
+5YOG 49114
+5YOJ 49115
+5YOK 49116
+5YON 49117
+5YOP 49118
+5YOR 49119
+5YOU 49120
+5YOV 49121
+5YOW 49122
+5YOa 49123
+5YOe 49124
+5YOj 49125
+5YOl 49126
+5YOm 49127
+5YOn 49128
+5YOs 49129
+5YOt 49130
+5YOu 49131
+5YOx 49132
+5YOz 49133
+5YO1 49134
+5YO5 49135
+5YO7 49136
+5YSA 49137
+5YSB 49138
+5YSC 49139
+5YSE 49140
+5YSG 49141
+5YSI 49142
+5YSJ 49143
+5YSL 49144
+5YSS 49145
+5YSU 49146
+5YSV 49147
+5YSY 49148
+5YSa 49149
+5YSf 49150
+5YSh 49151
+5YSq 49152
+5YSy 49153
+5YS3 49154
+5YS6 49155
+5YS8 49156
+5YS/ 49157
+5YWA 49158
+5YWB 49159
+5YWD 49160
+5YWE 49161
+5YWF 49162
+5YWG 49163
+5YWH 49164
+5YWI 49165
+5YWJ 49166
+5YWK 49167
+5YWL 49168
+5YWM 49169
+5YWN 49170
+5YWO 49171
+5YWQ 49172
+5YWR 49173
+5YWS 49174
+5YWU 49175
+5YWV 49176
+5YWW 49177
+5YWa 49178
+5YWc 49179
+5YWi 49180
+5YWl 49181
+5YWn 49182
+5YWo 49183
+5YWp 49184
+5YWq 49185
+5YWr 49186
+5YWs 49187
+5YWt 49188
+5YWu 49189
+5YWw 49190
+5YWx 49191
+5YWz 49192
+5YW0 49193
+5YW1 49194
+5YW2 49195
+5YW3 49196
+5YW4 49197
+5YW5 49198
+5YW7 49199
+5YW8 49200
+5YW9 49201
+5YaA 49202
+5YaC 49203
+5YaF 49204
+5YaG 49205
+5YaH 49206
+5YaI 49207
+5YaJ 49208
+5YaK 49209
+5YaM 49210
+5YaN 49211
+5YaP 49212
+5YaR 49213
+5YaS 49214
+5YaV 49215
+5YaW 49216
+5YaX 49217
+5YaZ 49218
+5Yaa 49219
+5Yab 49220
+5Yac 49221
+5Yag 49222
+5Yai 49223
+5Yak 49224
+5Yal 49225
+5Yam 49226
+5Yan 49227
+5Yao 49228
+5Yap 49229
+5Yaq 49230
+5Yar 49231
+5Yas 49232
+5Yau 49233
+5Yav 49234
+5Yaw 49235
+5Yax 49236
+5Yay 49237
+5Yaz 49238
+5Ya0 49239
+5Ya1 49240
+5Ya2 49241
+5Ya3 49242
+5Ya7 49243
+5Ya8 49244
+5Ya9 49245
+5Ya/ 49246
+5YeA 49247
+5YeE 49248
+5YeG 49249
+5YeH 49250
+5YeI 49251
+5YeJ 49252
+5YeK 49253
+5YeL 49254
+5YeM 49255
+5YeN 49256
+5YeP 49257
+5YeR 49258
+5YeW 49259
+5Yeb 49260
+5Yec 49261
+5Yed 49262
+5Yeg 49263
+5Yeh 49264
+5Yei 49265
+5Yek 49266
+5Yem 49267
+5Yen 49268
+5Yep 49269
+5Yeq 49270
+5Yer 49271
+5Yet 49272
+5Yev 49273
+5Yew 49274
+5Yex 49275
+5Yez 49276
+5Ye1 49277
+5Ye2 49278
+5Ye4 49279
+5Ye5 49280
+5Ye6 49281
+5Ye7 49282
+5Ye8 49283
+5Ye9 49284
+5Ye/ 49285
+5YiA 49286
+5YiB 49287
+5YiD 49288
+5YiG 49289
+5YiH 49290
+5YiI 49291
+5YiK 49292
+5YiN 49293
+5YiO 49294
+5YiR 49295
+5YiS 49296
+5YiU 49297
+5YiX 49298
+5YiY 49299
+5YiZ 49300
+5Yia 49301
+5Yib 49302
+5Yid 49303
+5Yig 49304
+5Yik 49305
+5Yil 49306
+5Yio 49307
+5Yip 49308
+5Yiq 49309
+5Yir 49310
+5Yis 49311
+5Yit 49312
+5Yiu 49313
+5Yiw 49314
+5Yiz 49315
+5Yi2 49316
+5Yi3 49317
+5Yi4 49318
+5Yi5 49319
+5Yi6 49320
+5Yi7 49321
+5Yi8 49322
+5Yi9 49323
+5Yi/ 49324
+5YmA 49325
+5YmB 49326
+5YmC 49327
+5YmD 49328
+5YmF 49329
+5YmH 49330
+5YmK 49331
+5YmL 49332
+5YmM 49333
+5YmN 49334
+5YmO 49335
+5YmQ 49336
+5YmR 49337
+5YmU 49338
+5YmV 49339
+5YmW 49340
+5Ymb 49341
+5Ymc 49342
+5Ymd 49343
+5Yme 49344
+5Ymf 49345
+5Ymh 49346
+5Ymj 49347
+5Ymk 49348
+5Yml 49349
+5Ymn 49350
+5Ymp 49351
+5Ymq 49352
+5Ymv 49353
+5Ymw 49354
+5Ymx 49355
+5Ymy 49356
+5Ym0 49357
+5Ym1 49358
+5Ym9 49359
+5Ym/ 49360
+5YqC 49361
+5YqD 49362
+5YqH 49363
+5YqI 49364
+5YqJ 49365
+5YqN 49366
+5YqP 49367
+5YqR 49368
+5YqS 49369
+5YqT 49370
+5YqU 49371
+5Yqb 49372
+5Yqd 49373
+5Yqe 49374
+5Yqf 49375
+5Yqg 49376
+5Yqh 49377
+5Yqi 49378
+5Yqj 49379
+5Yqo 49380
+5Yqp 49381
+5Yqq 49382
+5Yqr 49383
+5Yqs 49384
+5Yqt 49385
+5Yqx 49386
+5Yqy 49387
+5Yqz 49388
+5Yq0 49389
+5Yq1 49390
+5Yq5 49391
+5Yq8 49392
+5Yq+ 49393
+5Yq/ 49394
+5YuB 49395
+5YuD 49396
+5YuF 49397
+5YuH 49398
+5YuJ 49399
+5YuL 49400
+5YuN 49401
+5YuQ 49402
+5YuS 49403
+5YuV 49404
+5YuW 49405
+5YuX 49406
+5YuY 49407
+5YuZ 49408
+5Yub 49409
+5Yud 49410
+5Yue 49411
+5Yuf 49412
+5Yug 49413
+5Yui 49414
+5Yuj 49415
+5Yuk 49416
+5Yum 49417
+5Yun 49418
+5Yuu 49419
+5Yuw 49420
+5Yuy 49421
+5Yuz 49422
+5Yu1 49423
+5Yu4 49424
+5Yu5 49425
+5Yu6 49426
+5Yu7 49427
+5Yu+ 49428
+5Yu/ 49429
+5YyA 49430
+5YyB 49431
+5YyC 49432
+5YyF 49433
+5YyG 49434
+5YyI 49435
+5YyN 49436
+5YyP 49437
+5YyQ 49438
+5YyV 49439
+5YyW 49440
+5YyX 49441
+5YyZ 49442
+5Yya 49443
+5Yyc 49444
+5Yyd 49445
+5Yyg 49446
+5Yyh 49447
+5Yyj 49448
+5Yyq 49449
+5Yyu 49450
+5Yyv 49451
+5Yyx 49452
+5Yy4 49453
+5Yy5 49454
+5Yy6 49455
+5Yy7 49456
+5Yy8 49457
+5Yy9 49458
+5Yy+ 49459
+5Yy/ 49460
+5Y2A 49461
+5Y2B 49462
+5Y2D 49463
+5Y2F 49464
+5Y2G 49465
+5Y2H 49466
+5Y2I 49467
+5Y2J 49468
+5Y2K 49469
+5Y2N 49470
+5Y2O 49471
+5Y2P 49472
+5Y2R 49473
+5Y2S 49474
+5Y2T 49475
+5Y2U 49476
+5Y2V 49477
+5Y2W 49478
+5Y2X 49479
+5Y2Y 49480
+5Y2a 49481
+5Y2c 49482
+5Y2e 49483
+5Y2f 49484
+5Y2g 49485
+5Y2h 49486
+5Y2i 49487
+5Y2j 49488
+5Y2k 49489
+5Y2m 49490
+5Y2n 49491
+5Y2r 49492
+5Y2s 49493
+5Y2u 49494
+5Y2v 49495
+5Y2w 49496
+5Y2x 49497
+5Y2y 49498
+5Y2z 49499
+5Y20 49500
+5Y21 49501
+5Y23 49502
+5Y24 49503
+5Y26 49504
+5Y27 49505
+5Y2/ 49506
+5Y6C 49507
+5Y6E 49508
+5Y6F 49509
+5Y6G 49510
+5Y6J 49511
+5Y6L 49512
+5Y6M 49513
+5Y6N 49514
+5Y6Q 49515
+5Y6T 49516
+5Y6U 49517
+5Y6V 49518
+5Y6W 49519
+5Y6Y 49520
+5Y6a 49521
+5Y6d 49522
+5Y6f 49523
+5Y6g 49524
+5Y6i 49525
+5Y6j 49526
+5Y6l 49527
+5Y6m 49528
+5Y6o 49529
+5Y6p 49530
+5Y6t 49531
+5Y6u 49532
+5Y6w 49533
+5Y6y 49534
+5Y6z 49535
+5Y62 49536
+5Y67 49537
+5Y6/ 49538
+5Y+B 49539
+5Y+C 49540
+5Y+D 49541
+5Y+G 49542
+5Y+I 49543
+5Y+J 49544
+5Y+K 49545
+5Y+L 49546
+5Y+M 49547
+5Y+N 49548
+5Y+O 49549
+5Y+R 49550
+5Y+U 49551
+5Y+W 49552
+5Y+X 49553
+5Y+Y 49554
+5Y+Z 49555
+5Y+b 49556
+5Y+f 49557
+5Y+g 49558
+5Y+h 49559
+5Y+i 49560
+5Y+j 49561
+5Y+k 49562
+5Y+l 49563
+5Y+m 49564
+5Y+o 49565
+5Y+p 49566
+5Y+q 49567
+5Y+r 49568
+5Y+s 49569
+5Y+t 49570
+5Y+u 49571
+5Y+v 49572
+5Y+w 49573
+5Y+x 49574
+5Y+y 49575
+5Y+z 49576
+5Y+1 49577
+5Y+2 49578
+5Y+3 49579
+5Y+4 49580
+5Y+5 49581
+5Y+6 49582
+5Y+7 49583
+5Y+8 49584
+5Y+9 49585
+5ZCB 49586
+5ZCD 49587
+5ZCE 49588
+5ZCF 49589
+5ZCG 49590
+5ZCH 49591
+5ZCI 49592
+5ZCJ 49593
+5ZCK 49594
+5ZCL 49595
+5ZCM 49596
+5ZCN 49597
+5ZCO 49598
+5ZCP 49599
+5ZCQ 49600
+5ZCR 49601
+5ZCS 49602
+5ZCT 49603
+5ZCU 49604
+5ZCV 49605
+5ZCW 49606
+5ZCX 49607
+5ZCZ 49608
+5ZCb 49609
+5ZCd 49610
+5ZCe 49611
+5ZCf 49612
+5ZCg 49613
+5ZCh 49614
+5ZCj 49615
+5ZCl 49616
+5ZCm 49617
+5ZCn 49618
+5ZCo 49619
+5ZCp 49620
+5ZCr 49621
+5ZCs 49622
+5ZCt 49623
+5ZCu 49624
+5ZCv 49625
+5ZCx 49626
+5ZCy 49627
+5ZCz 49628
+5ZC0 49629
+5ZC1 49630
+5ZC2 49631
+5ZC4 49632
+5ZC5 49633
+5ZC7 49634
+5ZC8 49635
+5ZC9 49636
+5ZC+ 49637
+5ZC/ 49638
+5ZGA 49639
+5ZGC 49640
+5ZGD 49641
+5ZGG 49642
+5ZGI 49643
+5ZGJ 49644
+5ZGK 49645
+5ZGL 49646
+5ZGO 49647
+5ZGQ 49648
+5ZGR 49649
+5ZGS 49650
+5ZGT 49651
+5ZGU 49652
+5ZGV 49653
+5ZGW 49654
+5ZGX 49655
+5ZGY 49656
+5ZGZ 49657
+5ZGb 49658
+5ZGc 49659
+5ZGf 49660
+5ZGi 49661
+5ZGj 49662
+5ZGk 49663
+5ZGm 49664
+5ZGo 49665
+5ZGq 49666
+5ZGr 49667
+5ZGv 49668
+5ZGw 49669
+5ZGx 49670
+5ZGy 49671
+5ZGz 49672
+5ZG1 49673
+5ZG2 49674
+5ZG3 49675
+5ZG4 49676
+5ZG7 49677
+5ZG8 49678
+5ZG9 49679
+5ZKA 49680
+5ZKB 49681
+5ZKC 49682
+5ZKE 49683
+5ZKG 49684
+5ZKJ 49685
+5ZKL 49686
+5ZKM 49687
+5ZKO 49688
+5ZKP 49689
+5ZKQ 49690
+5ZKS 49691
+5ZKU 49692
+5ZKV 49693
+5ZKW 49694
+5ZKX 49695
+5ZKY 49696
+5ZKZ 49697
+5ZKa 49698
+5ZKb 49699
+5ZKd 49700
+5ZKj 49701
+5ZKk 49702
+5ZKm 49703
+5ZKn 49704
+5ZKo 49705
+5ZKp 49706
+5ZKq 49707
+5ZKr 49708
+5ZKs 49709
+5ZKt 49710
+5ZKv 49711
+5ZKx 49712
+5ZKy 49713
+5ZKz 49714
+5ZK0 49715
+5ZK1 49716
+5ZK2 49717
+5ZK3 49718
+5ZK4 49719
+5ZK5 49720
+5ZK7 49721
+5ZK9 49722
+5ZK+ 49723
+5ZK/ 49724
+5ZOA 49725
+5ZOB 49726
+5ZOC 49727
+5ZOD 49728
+5ZOE 49729
+5ZOG 49730
+5ZOH 49731
+5ZOI 49732
+5ZOJ 49733
+5ZOL 49734
+5ZOM 49735
+5ZON 49736
+5ZOO 49737
+5ZOP 49738
+5ZOQ 49739
+5ZOR 49740
+5ZOS 49741
+5ZOT 49742
+5ZOU 49743
+5ZOV 49744
+5ZOW 49745
+5ZOX 49746
+5ZOY 49747
+5ZOZ 49748
+5ZOa 49749
+5ZOc 49750
+5ZOd 49751
+5ZOe 49752
+5ZOf 49753
+5ZOg 49754
+5ZOh 49755
+5ZOl 49756
+5ZOm 49757
+5ZOn 49758
+5ZOo 49759
+5ZOp 49760
+5ZOq 49761
+5ZOt 49762
+5ZOu 49763
+5ZOy 49764
+5ZO6 49765
+5ZO8 49766
+5ZO9 49767
+5ZSB 49768
+5ZSE 49769
+5ZSG 49770
+5ZSH 49771
+5ZSJ 49772
+5ZSP 49773
+5ZSQ 49774
+5ZSR 49775
+5ZSU 49776
+5ZSW 49777
+5ZSb 49778
+5ZSe 49779
+5ZSg 49780
+5ZSi 49781
+5ZSj 49782
+5ZSk 49783
+5ZSn 49784
+5ZSq 49785
+5ZSs 49786
+5ZSu 49787
+5ZSv 49788
+5ZSw 49789
+5ZSx 49790
+5ZSz 49791
+5ZS1 49792
+5ZS3 49793
+5ZS4 49794
+5ZS7 49795
+5ZS8 49796
+5ZS+ 49797
+5ZS/ 49798
+5ZWA 49799
+5ZWB 49800
+5ZWC 49801
+5ZWD 49802
+5ZWE 49803
+5ZWG 49804
+5ZWJ 49805
+5ZWK 49806
+5ZWL 49807
+5ZWP 49808
+5ZWQ 49809
+5ZWT 49810
+5ZWV 49811
+5ZWW 49812
+5ZWc 49813
+5ZWe 49814
+5ZWf 49815
+5ZWh 49816
+5ZWk 49817
+5ZWl 49818
+5ZWm 49819
+5ZWn 49820
+5ZWp 49821
+5ZWq 49822
+5ZWr 49823
+5ZWs 49824
+5ZWt 49825
+5ZWu 49826
+5ZWv 49827
+5ZWw 49828
+5ZWx 49829
+5ZWy 49830
+5ZW1 49831
+5ZW2 49832
+5ZW3 49833
+5ZW4 49834
+5ZW7 49835
+5ZW8 49836
+5ZW+ 49837
+5ZaA 49838
+5ZaB 49839
+5ZaC 49840
+5ZaD 49841
+5ZaE 49842
+5ZaG 49843
+5ZaH 49844
+5ZaI 49845
+5ZaJ 49846
+5ZaK 49847
+5ZaL 49848
+5ZaO 49849
+5ZaP 49850
+5ZaQ 49851
+5ZaR 49852
+5ZaS 49853
+5ZaT 49854
+5ZaU 49855
+5ZaY 49856
+5ZaZ 49857
+5Zaa 49858
+5Zab 49859
+5Zac 49860
+5Zad 49861
+5Zaf 49862
+5Zak 49863
+5Zan 49864
+5Zao 49865
+5Zap 49866
+5Zaq 49867
+5Zar 49868
+5Zas 49869
+5Zau 49870
+5Zaw 49871
+5Zax 49872
+5Zay 49873
+5Zaz 49874
+5Za1 49875
+5Za2 49876
+5Za3 49877
+5Za5 49878
+5Za6 49879
+5Za7 49880
+5Za9 49881
+5Za+ 49882
+5ZeE 49883
+5ZeF 49884
+5ZeG 49885
+5ZeH 49886
+5ZeJ 49887
+5ZeM 49888
+5ZeN 49889
+5ZeO 49890
+5ZeQ 49891
+5ZeR 49892
+5ZeS 49893
+5ZeT 49894
+5ZeU 49895
+5ZeW 49896
+5ZeZ 49897
+5Zea 49898
+5Zeb 49899
+5Zec 49900
+5Zed 49901
+5Zee 49902
+5Zef 49903
+5Zeh 49904
+5Zej 49905
+5Zek 49906
+5Zel 49907
+5Zem 49908
+5Zeo 49909
+5Zeq 49910
+5Zer 49911
+5Zes 49912
+5Zeu 49913
+5Zev 49914
+5Zew 49915
+5Zex 49916
+5Zey 49917
+5Zez 49918
+5Ze1 49919
+5Ze2 49920
+5Ze3 49921
+5Ze7 49922
+5Ze9 49923
+5Ze+ 49924
+5ZiA 49925
+5ZiB 49926
+5ZiF 49927
+5ZiG 49928
+5ZiI 49929
+5ZiJ 49930
+5ZiM 49931
+5ZiO 49932
+5ZiP 49933
+5ZiU 49934
+5ZiW 49935
+5ZiX 49936
+5ZiY 49937
+5Zia 49938
+5Zib 49939
+5Zie 49940
+5Zif 49941
+5Zih 49942
+5Zii 49943
+5Zij 49944
+5Zik 49945
+5Zil 49946
+5Zin 49947
+5Zip 49948
+5Zis 49949
+5Zit 49950
+5Ziv 49951
+5Ziw 49952
+5Zix 49953
+5Ziy 49954
+5Zi0 49955
+5Zi2 49956
+5Zi4 49957
+5Zi5 49958
+5Zi7 49959
+5Zi/ 49960
+5ZmC 49961
+5ZmD 49962
+5ZmM 49963
+5ZmN 49964
+5ZmO 49965
+5ZmT 49966
+5ZmU 49967
+5ZmX 49968
+5ZmY 49969
+5ZmZ 49970
+5Zmb 49971
+5Zmc 49972
+5Zmg 49973
+5Zmi 49974
+5Zmk 49975
+5Zmo 49976
+5Zmp 49977
+5Zmq 49978
+5Zmr 49979
+5Zms 49980
+5Zmx 49981
+5Zm0 49982
+5Zm2 49983
+5Zm4 49984
+5Zm5 49985
+5Zm6 49986
+5Zm7 49987
+5Zm8 49988
+5ZqA 49989
+5ZqF 49990
+5ZqG 49991
+5ZqH 49992
+5ZqK 49993
+5ZqO 49994
+5ZqP 49995
+5ZqQ 49996
+5ZqS 49997
+5ZqT 49998
+5ZqU 49999
+5ZqV 50000
+5ZqX 50001
+5Zqc 50002
+5Zqf 50003
+5Zqg 50004
+5Zqi 50005
+5Zqj 50006
+5Zql 50007
+5Zqo 50008
+5Zqp 50009
+5Zqt 50010
+5Zqu 50011
+5Zqv 50012
+5Zq0 50013
+5Zq2 50014
+5Zq3 50015
+5Zq8 50016
+5ZuA 50017
+5ZuB 50018
+5ZuC 50019
+5ZuD 50020
+5ZuI 50021
+5ZuJ 50022
+5ZuK 50023
+5ZuN 50024
+5ZuO 50025
+5ZuR 50026
+5ZuU 50027
+5ZuW 50028
+5ZuX 50029
+5Zua 50030
+5Zub 50031
+5Zud 50032
+5Zue 50033
+5Zuf 50034
+5Zug 50035
+5Zuh 50036
+5Zui 50037
+5Zuj 50038
+5Zuk 50039
+5Zun 50040
+5Zuo 50041
+5Zur 50042
+5Zut 50043
+5Zuu 50044
+5Zuv 50045
+5Zuw 50046
+5Zux 50047
+5Zuy 50048
+5Zuz 50049
+5Zu0 50050
+5Zu1 50051
+5Zu3 50052
+5Zu5 50053
+5Zu6 50054
+5Zu9 50055
+5Zu+ 50056
+5Zu/ 50057
+5ZyA 50058
+5ZyD 50059
+5ZyE 50060
+5ZyG 50061
+5ZyI 50062
+5ZyJ 50063
+5ZyK 50064
+5ZyL 50065
+5ZyN 50066
+5ZyP 50067
+5ZyS 50068
+5ZyT 50069
+5ZyW 50070
+5ZyY 50071
+5Zyc 50072
+5Zyf 50073
+5Zyj 50074
+5Zym 50075
+5Zyn 50076
+5Zyo 50077
+5Zyp 50078
+5Zyq 50079
+5Zys 50080
+5Zyt 50081
+5Zyu 50082
+5Zyv 50083
+5Zyw 50084
+5Zyz 50085
+5Zy3 50086
+5Zy5 50087
+5Zy6 50088
+5Zy7 50089
+5Zy+ 50090
+5Z2A 50091
+5Z2C 50092
+5Z2H 50093
+5Z2K 50094
+5Z2N 50095
+5Z2O 50096
+5Z2P 50097
+5Z2Q 50098
+5Z2R 50099
+5Z2S 50100
+5Z2X 50101
+5Z2a 50102
+5Z2b 50103
+5Z2c 50104
+5Z2d 50105
+5Z2e 50106
+5Z2f 50107
+5Z2g 50108
+5Z2h 50109
+5Z2k 50110
+5Z2m 50111
+5Z2o 50112
+5Z2p 50113
+5Z2q 50114
+5Z2t 50115
+5Z2u 50116
+5Z2v 50117
+5Z2z 50118
+5Z23 50119
+5Z27 50120
+5Z28 50121
+5Z6C 50122
+5Z6D 50123
+5Z6E 50124
+5Z6F 50125
+5Z6G 50126
+5Z6H 50127
+5Z6I 50128
+5Z6L 50129
+5Z6M 50130
+5Z6N 50131
+5Z6S 50132
+5Z6T 50133
+5Z6V 50134
+5Z6X 50135
+5Z6a 50136
+5Z6b 50137
+5Z6d 50138
+5Z6g 50139
+5Z6h 50140
+5Z6i 50141
+5Z6j 50142
+5Z6k 50143
+5Z6m 50144
+5Z6n 50145
+5Z6p 50146
+5Z6q 50147
+5Z6r 50148
+5Z6t 50149
+5Z6u 50150
+5Z6v 50151
+5Z6w 50152
+5Z6x 50153
+5Z6z 50154
+5Z60 50155
+5Z61 50156
+5Z64 50157
+5Z+A 50158
+5Z+C 50159
+5Z+D 50160
+5Z+G 50161
+5Z+H 50162
+5Z+L 50163
+5Z+M 50164
+5Z+O 50165
+5Z+P 50166
+5Z+S 50167
+5Z+T 50168
+5Z+U 50169
+5Z+V 50170
+5Z+W 50171
+5Z+X 50172
+5Z+Z 50173
+5Z+a 50174
+5Z+c 50175
+5Z+d 50176
+5Z+f 50177
+5Z+g 50178
+5Z+j 50179
+5Z+k 50180
+5Z+q 50181
+5Z+t 50182
+5Z+y 50183
+5Z+0 50184
+5Z+1 50185
+5Z+3 50186
+5Z+4 50187
+5Z+5 50188
+5Z+6 50189
+5Z+8 50190
+5Z+9 50191
+5aCA 50192
+5aCC 50193
+5aCD 50194
+5aCF 50195
+5aCG 50196
+5aCH 50197
+5aCK 50198
+5aCL 50199
+5aCM 50200
+5aCN 50201
+5aCO 50202
+5aCQ 50203
+5aCR 50204
+5aCV 50205
+5aCZ 50206
+5aCd 50207
+5aCe 50208
+5aCg 50209
+5aCh 50210
+5aCi 50211
+5aCk 50212
+5aCo 50213
+5aCq 50214
+5aCv 50215
+5aCw 50216
+5aCx 50217
+5aC0 50218
+5aC1 50219
+5aC6 50220
+5aC8 50221
+5aGA 50222
+5aGB 50223
+5aGE 50224
+5aGK 50225
+5aGL 50226
+5aGM 50227
+5aGN 50228
+5aGR 50229
+5aGS 50230
+5aGU 50231
+5aGX 50232
+5aGY 50233
+5aGZ 50234
+5aGa 50235
+5aGe 50236
+5aGi 50237
+5aGp 50238
+5aGr 50239
+5aGs 50240
+5aGy 50241
+5aG1 50242
+5aG5 50243
+5aG9 50244
+5aG+ 50245
+5aKA 50246
+5aKB 50247
+5aKD 50248
+5aKF 50249
+5aKJ 50250
+5aKK 50251
+5aKO 50252
+5aKS 50253
+5aKT 50254
+5aKV 50255
+5aKX 50256
+5aKZ 50257
+5aKa 50258
+5aKc 50259
+5aKe 50260
+5aKf 50261
+5aKh 50262
+5aKo 50263
+5aKp 50264
+5aKr 50265
+5aKu 50266
+5aKz 50267
+5aK5 50268
+5aK6 50269
+5aK7 50270
+5aK8 50271
+5aK+ 50272
+5aOB 50273
+5aOF 50274
+5aOH 50275
+5aOK 50276
+5aOM 50277
+5aOR 50278
+5aOT 50279
+5aOV 50280
+5aOX 50281
+5aOY 50282
+5aOc 50283
+5aOe 50284
+5aOf 50285
+5aOk 50286
+5aOp 50287
+5aOr 50288
+5aOs 50289
+5aOu 50290
+5aOv 50291
+5aOw 50292
+5aOx 50293
+5aOy 50294
+5aOz 50295
+5aO2 50296
+5aO3 50297
+5aO4 50298
+5aO5 50299
+5aO6 50300
+5aO7 50301
+5aO9 50302
+5aSE 50303
+5aSH 50304
+5aSJ 50305
+5aSK 50306
+5aSM 50307
+5aSN 50308
+5aSP 50309
+5aSQ 50310
+5aSU 50311
+5aSV 50312
+5aSW 50313
+5aSZ 50314
+5aSa 50315
+5aSb 50316
+5aSc 50317
+5aSf 50318
+5aSg 50319
+5aSi 50320
+5aSk 50321
+5aSl 50322
+5aSn 50323
+5aSp 50324
+5aSq 50325
+5aSr 50326
+5aSs 50327
+5aSt 50328
+5aSu 50329
+5aSv 50330
+5aSx 50331
+5aS0 50332
+5aS3 50333
+5aS4 50334
+5aS5 50335
+5aS6 50336
+5aS8 50337
+5aS+ 50338
+5aWB 50339
+5aWC 50340
+5aWE 50341
+5aWH 50342
+5aWI 50343
+5aWJ 50344
+5aWL 50345
+5aWM 50346
+5aWO 50347
+5aWP 50348
+5aWQ 50349
+5aWR 50350
+5aWU 50351
+5aWV 50352
+5aWW 50353
+5aWX 50354
+5aWY 50355
+5aWa 50356
+5aWg 50357
+5aWi 50358
+5aWl 50359
+5aWn 50360
+5aWo 50361
+5aWq 50362
+5aWs 50363
+5aWt 50364
+5aWu 50365
+5aWz 50366
+5aW0 50367
+5aW2 50368
+5aW4 50369
+5aW5 50370
+5aW9 50371
+5aaB 50372
+5aaC 50373
+5aaD 50374
+5aaE 50375
+5aaG 50376
+5aaH 50377
+5aaI 50378
+5aaK 50379
+5aaN 50380
+5aaS 50381
+5aaT 50382
+5aaW 50383
+5aaX 50384
+5aaZ 50385
+5aad 50386
+5aae 50387
+5aaj 50388
+5aak 50389
+5aal 50390
+5aan 50391
+5aao 50392
+5aap 50393
+5aaq 50394
+5aar 50395
+5aas 50396
+5aau 50397
+5aav 50398
+5aax 50399
+5aay 50400
+5aa5 50401
+5aa6 50402
+5aa7 50403
+5aa9 50404
+5aa+ 50405
+5aeG 50406
+5aeJ 50407
+5aeK 50408
+5aeL 50409
+5aeN 50410
+5aeQ 50411
+5aeR 50412
+5aeS 50413
+5aeT 50414
+5aeU 50415
+5aeX 50416
+5aeY 50417
+5aea 50418
+5aec 50419
+5aed 50420
+5aej 50421
+5aek 50422
+5ael 50423
+5aem 50424
+5aeo 50425
+5aep 50426
+5aeq 50427
+5aer 50428
+5aes 50429
+5aeu 50430
+5ae2 50431
+5ae4 50432
+5ae5 50433
+5ae6 50434
+5ae7 50435
+5ae/ 50436
+5aiB 50437
+5aiD 50438
+5aiE 50439
+5aiF 50440
+5aiG 50441
+5aiH 50442
+5aiI 50443
+5aiJ 50444
+5aiM 50445
+5aiR 50446
+5aiT 50447
+5aiY 50448
+5aia 50449
+5aib 50450
+5aic 50451
+5aif 50452
+5aig 50453
+5aih 50454
+5aij 50455
+5ail 50456
+5aip 50457
+5aiv 50458
+5aix 50459
+5aiy 50460
+5ai0 50461
+5ai2 50462
+5ai8 50463
+5ai/ 50464
+5amA 50465
+5amB 50466
+5amE 50467
+5amG 50468
+5amJ 50469
+5amK 50470
+5amM 50471
+5amV 50472
+5ama 50473
+5ami 50474
+5amm 50475
+5amn 50476
+5amq 50477
+5ams 50478
+5am0 50479
+5am1 50480
+5am2 50481
+5am3 50482
+5am6 50483
+5am/ 50484
+5aqS 50485
+5aqW 50486
+5aqa 50487
+5aqb 50488
+5aqc 50489
+5aqe 50490
+5aqq 50491
+5aqx 50492
+5aqy 50493
+5aqz 50494
+5aq1 50495
+5aq8 50496
+5aq9 50497
+5aq+ 50498
+5auB 50499
+5auC 50500
+5auE 50501
+5auJ 50502
+5auL 50503
+5auM 50504
+5auS 50505
+5auU 50506
+5auW 50507
+5auX 50508
+5auY 50509
+5aua 50510
+5auh 50511
+5auj 50512
+5auk 50513
+5aum 50514
+5aup 50515
+5auq 50516
+5aur 50517
+5auw 50518
+5aux 50519
+5auy 50520
+5au1 50521
+5au7 50522
+5ayF 50523
+5ayJ 50524
+5ayL 50525
+5ayM 50526
+5ayW 50527
+5ayX 50528
+5ayb 50529
+5ayi 50530
+5ayl 50531
+5ayq 50532
+5ays 50533
+5ayw 50534
+5ayy 50535
+5ay0 50536
+5ay2 50537
+5ay3 50538
+5ay4 50539
+5a2A 50540
+5a2D 50541
+5a2Q 50542
+5a2R 50543
+5a2T 50544
+5a2U 50545
+5a2V 50546
+5a2W 50547
+5a2X 50548
+5a2Y 50549
+5a2Z 50550
+5a2a 50551
+5a2b 50552
+5a2c 50553
+5a2d 50554
+5a2f 50555
+5a2i 50556
+5a2j 50557
+5a2k 50558
+5a2l 50559
+5a2m 50560
+5a2p 50561
+5a2q 50562
+5a2r 50563
+5a2s 50564
+5a2t 50565
+5a2w 50566
+5a2x 50567
+5a2z 50568
+5a21 50569
+5a24 50570
+5a26 50571
+5a29 50572
+5a6B 50573
+5a6D 50574
+5a6E 50575
+5a6F 50576
+5a6H 50577
+5a6I 50578
+5a6J 50579
+5a6L 50580
+5a6M 50581
+5a6N 50582
+5a6P 50583
+5a6T 50584
+5a6V 50585
+5a6X 50586
+5a6Y 50587
+5a6Z 50588
+5a6a 50589
+5a6b 50590
+5a6c 50591
+5a6d 50592
+5a6e 50593
+5a6f 50594
+5a6g 50595
+5a6h 50596
+5a6i 50597
+5a6j 50598
+5a6k 50599
+5a6l 50600
+5a6m 50601
+5a6q 50602
+5a6r 50603
+5a6u 50604
+5a6w 50605
+5a6z 50606
+5a60 50607
+5a61 50608
+5a62 50609
+5a64 50610
+5a65 50611
+5a69 50612
+5a6+ 50613
+5a6/ 50614
+5a+C 50615
+5a+E 50616
+5a+F 50617
+5a+G 50618
+5a+H 50619
+5a+M 50620
+5a+Q 50621
+5a+S 50622
+5a+T 50623
+5a+U 50624
+5a+V 50625
+5a+W 50626
+5a+b 50627
+5a+d 50628
+5a+e 50629
+5a+f 50630
+5a+g 50631
+5a+h 50632
+5a+i 50633
+5a+k 50634
+5a+l 50635
+5a+m 50636
+5a+n 50637
+5a+o 50638
+5a+p 50639
+5a+r 50640
+5a+s 50641
+5a+u 50642
+5a+w 50643
+5a+z 50644
+5a+1 50645
+5a+2 50646
+5a+4 50647
+5a+5 50648
+5a+6 50649
+5a+7 50650
+5a+8 50651
+5a++ 50652
+5a+/ 50653
+5bCB 50654
+5bCC 50655
+5bCE 50656
+5bCF 50657
+5bCG 50658
+5bCH 50659
+5bCI 50660
+5bCJ 50661
+5bCK 50662
+5bCL 50663
+5bCN 50664
+5bCO 50665
+5bCP 50666
+5bCR 50667
+5bCT 50668
+5bCU 50669
+5bCV 50670
+5bCW 50671
+5bCY 50672
+5bCZ 50673
+5bCa 50674
+5bCd 50675
+5bCe 50676
+5bCg 50677
+5bCh 50678
+5bCi 50679
+5bCk 50680
+5bCl 50681
+5bCn 50682
+5bCo 50683
+5bCq 50684
+5bCs 50685
+5bCt 50686
+5bCw 50687
+5bCx 50688
+5bC0 50689
+5bC3 50690
+5bC4 50691
+5bC5 50692
+5bC6 50693
+5bC7 50694
+5bC8 50695
+5bC9 50696
+5bC+ 50697
+5bC/ 50698
+5bGA 50699
+5bGB 50700
+5bGC 50701
+5bGE 50702
+5bGF 50703
+5bGG 50704
+5bGI 50705
+5bGJ 50706
+5bGK 50707
+5bGL 50708
+5bGM 50709
+5bGN 50710
+5bGO 50711
+5bGP 50712
+5bGQ 50713
+5bGR 50714
+5bGT 50715
+5bGV 50716
+5bGY 50717
+5bGZ 50718
+5bGc 50719
+5bGe 50720
+5bGg 50721
+5bGh 50722
+5bGi 50723
+5bGj 50724
+5bGk 50725
+5bGl 50726
+5bGm 50727
+5bGs 50728
+5bGv 50729
+5bGx 50730
+5bG5 50731
+5bG6 50732
+5bG/ 50733
+5bKA 50734
+5bKB 50735
+5bKC 50736
+5bKI 50737
+5bKM 50738
+5bKN 50739
+5bKQ 50740
+5bKR 50741
+5bKU 50742
+5bKV 50743
+5bKW 50744
+5bKX 50745
+5bKY 50746
+5bKZ 50747
+5bKa 50748
+5bKb 50749
+5bKc 50750
+5bKe 50751
+5bKh 50752
+5bKi 50753
+5bKo 50754
+5bKp 50755
+5bKr 50756
+5bKs 50757
+5bKt 50758
+5bKx 50759
+5bKz 50760
+5bK1 50761
+5bK3 50762
+5bK4 50763
+5bK7 50764
+5bK8 50765
+5bK9 50766
+5bK/ 50767
+5bOB 50768
+5bOE 50769
+5bOF 50770
+5bOH 50771
+5bOL 50772
+5bOS 50773
+5bOZ 50774
+5bOg 50775
+5bOh 50776
+5bOj 50777
+5bOk 50778
+5bOl 50779
+5bOm 50780
+5bOo 50781
+5bOq 50782
+5bOt 50783
+5bOv 50784
+5bOw 50785
+5bO2 50786
+5bO7 50787
+5bO9 50788
+5bSC 50789
+5bSD 50790
+5bSG 50791
+5bSH 50792
+5bSL 50793
+5bSO 50794
+5bSQ 50795
+5bSR 50796
+5bSS 50797
+5bSU 50798
+5bSV 50799
+5bSW 50800
+5bSX 50801
+5bSZ 50802
+5bSa 50803
+5bSb 50804
+5bSe 50805
+5bSf 50806
+5bSi 50807
+5bSk 50808
+5bSm 50809
+5bSn 50810
+5bSp 50811
+5bSt 50812
+5bSu 50813
+5bS0 50814
+5bS9 50815
+5bS+ 50816
+5bWC 50817
+5bWH 50818
+5bWK 50819
+5bWL 50820
+5bWM 50821
+5bWO 50822
+5bWQ 50823
+5bWS 50824
+5bWW 50825
+5bWY 50826
+5bWb 50827
+5bWc 50828
+5bWo 50829
+5bWp 50830
+5bWq 50831
+5bWr 50832
+5bWs 50833
+5bWv 50834
+5bW0 50835
+5baC 50836
+5baD 50837
+5baE 50838
+5baH 50839
+5baL 50840
+5baM 50841
+5baT 50842
+5baW 50843
+5baZ 50844
+5bad 50845
+5bas 50846
+5bau 50847
+5ba3 50848
+5ba6 50849
+5ba8 50850
+5ba9 50851
+5beC 50852
+5beF 50853
+5beH 50854
+5beJ 50855
+5beM 50856
+5beN 50857
+5beS 50858
+5beT 50859
+5beW 50860
+5bed 50861
+5bee 50862
+5beh 50863
+5bei 50864
+5bej 50865
+5bel 50866
+5bem 50867
+5ben 50868
+5beo 50869
+5bep 50870
+5ber 50871
+5beu 50872
+5bev 50873
+5bex 50874
+5bey 50875
+5bez 50876
+5be0 50877
+5be1 50878
+5be3 50879
+5be6 50880
+5be7 50881
+5be9 50882
+5be+ 50883
+5be/ 50884
+5biA 50885
+5biB 50886
+5biC 50887
+5biD 50888
+5biF 50889
+5biG 50890
+5biI 50891
+5biM 50892
+5biP 50893
+5biQ 50894
+5biR 50895
+5biU 50896
+5biV 50897
+5biW 50898
+5biY 50899
+5biZ 50900
+5bia 50901
+5bib 50902
+5bic 50903
+5bid 50904
+5bil 50905
+5bim 50906
+5bin 50907
+5bir 50908
+5bit 50909
+5biu 50910
+5biv 50911
+5biw 50912
+5biz 50913
+5bi2 50914
+5bi3 50915
+5bi4 50916
+5bi7 50917
+5bi8 50918
+5bi9 50919
+5bmA 50920
+5bmC 50921
+5bmE 50922
+5bmF 50923
+5bmH 50924
+5bmM 50925
+5bmU 50926
+5bmV 50927
+5bmX 50928
+5bmb 50929
+5bmf 50930
+5bmh 50931
+5bmi 50932
+5bmj 50933
+5bmr 50934
+5bmy 50935
+5bmz 50936
+5bm0 50937
+5bm2 50938
+5bm3 50939
+5bm4 50940
+5bm5 50941
+5bm6 50942
+5bm7 50943
+5bm8 50944
+5bm9 50945
+5bm+ 50946
+5bm/ 50947
+5bqB 50948
+5bqD 50949
+5bqE 50950
+5bqF 50951
+5bqG 50952
+5bqH 50953
+5bqK 50954
+5bqP 50955
+5bqQ 50956
+5bqR 50957
+5bqT 50958
+5bqU 50959
+5bqV 50960
+5bqW 50961
+5bqX 50962
+5bqZ 50963
+5bqa 50964
+5bqc 50965
+5bqe 50966
+5bqf 50967
+5bqg 50968
+5bqm 50969
+5bqn 50970
+5bqr 50971
+5bqt 50972
+5bq1 50973
+5bq2 50974
+5bq3 50975
+5bq4 50976
+5bq5 50977
+5bq+ 50978
+5buB 50979
+5buC 50980
+5buD 50981
+5buG 50982
+5buI 50983
+5buJ 50984
+5buK 50985
+5buL 50986
+5buR 50987
+5buT 50988
+5buW 50989
+5bua 50990
+5bub 50991
+5bud 50992
+5buf 50993
+5bug 50994
+5bui 50995
+5buj 50996
+5buo 50997
+5buq 50998
+5bus 50999
+5buz 51000
+5bu0 51001
+5bu2 51002
+5bu3 51003
+5bu6 51004
+5bu7 51005
+5bu8 51006
+5bu/ 51007
+5byA 51008
+5byB 51009
+5byC 51010
+5byD 51011
+5byE 51012
+5byH 51013
+5byI 51014
+5byJ 51015
+5byK 51016
+5byL 51017
+5byP 51018
+5byQ 51019
+5byR 51020
+5byT 51021
+5byU 51022
+5byV 51023
+5byW 51024
+5byX 51025
+5byY 51026
+5byb 51027
+5byf 51028
+5byg 51029
+5byi 51030
+5byl 51031
+5bym 51032
+5byn 51033
+5byo 51034
+5byp 51035
+5byt 51036
+5byv 51037
+5byx 51038
+5by1 51039
+5by3 51040
+5by5 51041
+5by6 51042
+5by8 51043
+5by+ 51044
+5b2A 51045
+5b2F 51046
+5b2G 51047
+5b2I 51048
+5b2K 51049
+5b2M 51050
+5b2O 51051
+5b2R 51052
+5b2S 51053
+5b2T 51054
+5b2V 51055
+5b2W 51056
+5b2X 51057
+5b2Y 51058
+5b2Z 51059
+5b2d 51060
+5b2f 51061
+5b2h 51062
+5b2i 51063
+5b2k 51064
+5b2l 51065
+5b2m 51066
+5b2n 51067
+5b2p 51068
+5b2q 51069
+5b2r 51070
+5b2s 51071
+5b2t 51072
+5b2w 51073
+5b2x 51074
+5b2z 51075
+5b23 51076
+5b25 51077
+5b26 51078
+5b27 51079
+5b28 51080
+5b2/ 51081
+5b6A 51082
+5b6B 51083
+5b6C 51084
+5b6D 51085
+5b6E 51086
+5b6F 51087
+5b6H 51088
+5b6I 51089
+5b6J 51090
+5b6K 51091
+5b6L 51092
+5b6M 51093
+5b6N 51094
+5b6Q 51095
+5b6R 51096
+5b6S 51097
+5b6T 51098
+5b6V 51099
+5b6X 51100
+5b6Y 51101
+5b6Z 51102
+5b6c 51103
+5b6e 51104
+5b6g 51105
+5b6h 51106
+5b6o 51107
+5b6p 51108
+5b6q 51109
+5b6s 51110
+5b6t 51111
+5b6u 51112
+5b6z 51113
+5b60 51114
+5b61 51115
+5b63 51116
+5b65 51117
+5b68 51118
+5b69 51119
+5b+D 51120
+5b+E 51121
+5b+F 51122
+5b+G 51123
+5b+L 51124
+5b+M 51125
+5b+N 51126
+5b+P 51127
+5b+Q 51128
+5b+R 51129
+5b+S 51130
+5b+W 51131
+5b+X 51132
+5b+Y 51133
+5b+Z 51134
+5b+c 51135
+5b+d 51136
+5b+g 51137
+5b+h 51138
+5b+k 51139
+5b+n 51140
+5b+q 51141
+5b+r 51142
+5b+w 51143
+5b+x 51144
+5b+1 51145
+5b+4 51146
+5b+7 51147
+5b+9 51148
+5b++ 51149
+5b+/ 51150
+5oCA 51151
+5oCB 51152
+5oCC 51153
+5oCD 51154
+5oCE 51155
+5oCF 51156
+5oCG 51157
+5oCN 51158
+5oCO 51159
+5oCP 51160
+5oCS 51161
+5oCU 51162
+5oCV 51163
+5oCW 51164
+5oCX 51165
+5oCZ 51166
+5oCb 51167
+5oCc 51168
+5oCd 51169
+5oCg 51170
+5oCh 51171
+5oCl 51172
+5oCm 51173
+5oCn 51174
+5oCo 51175
+5oCp 51176
+5oCq 51177
+5oCr 51178
+5oCv 51179
+5oCx 51180
+5oC1 51181
+5oC5 51182
+5oC7 51183
+5oC8 51184
+5oC/ 51185
+5oGB 51186
+5oGC 51187
+5oGD 51188
+5oGG 51189
+5oGL 51190
+5oGN 51191
+5oGQ 51192
+5oGS 51193
+5oGT 51194
+5oGV 51195
+5oGZ 51196
+5oGa 51197
+5oGf 51198
+5oGi 51199
+5oGj 51200
+5oGk 51201
+5oGl 51202
+5oGo 51203
+5oGp 51204
+5oGq 51205
+5oGr 51206
+5oGs 51207
+5oGt 51208
+5oGv 51209
+5oGw 51210
+5oGz 51211
+5oG1 51212
+5oG2 51213
+5oG4 51214
+5oG5 51215
+5oG6 51216
+5oG7 51217
+5oG8 51218
+5oG9 51219
+5oG/ 51220
+5oKD 51221
+5oKE 51222
+5oKF 51223
+5oKG 51224
+5oKJ 51225
+5oKL 51226
+5oKM 51227
+5oKN 51228
+5oKS 51229
+5oKU 51230
+5oKV 51231
+5oKW 51232
+5oKa 51233
+5oKb 51234
+5oKd 51235
+5oKf 51236
+5oKg 51237
+5oKj 51238
+5oKm 51239
+5oKn 51240
+5oKo 51241
+5oKp 51242
+5oKq 51243
+5oKr 51244
+5oKs 51245
+5oKt 51246
+5oKv 51247
+5oKx 51248
+5oKy 51249
+5oKz 51250
+5oK0 51251
+5oK1 51252
+5oK2 51253
+5oK4 51254
+5oK7 51255
+5oK8 51256
+5oK9 51257
+5oOF 51258
+5oOG 51259
+5oOH 51260
+5oOK 51261
+5oOL 51262
+5oOO 51263
+5oOR 51264
+5oOV 51265
+5oOY 51266
+5oOZ 51267
+5oOa 51268
+5oOc 51269
+5oOf 51270
+5oOg 51271
+5oOh 51272
+5oOj 51273
+5oOm 51274
+5oOn 51275
+5oOo 51276
+5oOp 51277
+5oOr 51278
+5oOs 51279
+5oOt 51280
+5oOu 51281
+5oOv 51282
+5oOw 51283
+5oOx 51284
+5oOz 51285
+5oO0 51286
+5oO2 51287
+5oO5 51288
+5oO6 51289
+5oO7 51290
+5oSA 51291
+5oSB 51292
+5oSG 51293
+5oSH 51294
+5oSI 51295
+5oSJ 51296
+5oSK 51297
+5oSN 51298
+5oSO 51299
+5oSP 51300
+5oSU 51301
+5oSV 51302
+5oSa 51303
+5oSb 51304
+5oSf 51305
+5oSg 51306
+5oSj 51307
+5oSk 51308
+5oSm 51309
+5oSn 51310
+5oSr 51311
+5oSs 51312
+5oS0 51313
+5oS3 51314
+5oS8 51315
+5oS+ 51316
+5oS/ 51317
+5oWC 51318
+5oWE 51319
+5oWG 51320
+5oWH 51321
+5oWI 51322
+5oWK 51323
+5oWL 51324
+5oWM 51325
+5oWO 51326
+5oWR 51327
+5oWT 51328
+5oWV 51329
+5oWY 51330
+5oWZ 51331
+5oWa 51332
+5oWc 51333
+5oWd 51334
+5oWf 51335
+5oWi 51336
+5oWj 51337
+5oWl 51338
+5oWn 51339
+5oWo 51340
+5oWr 51341
+5oWu 51342
+5oWw 51343
+5oWz 51344
+5oW0 51345
+5oW1 51346
+5oW2 51347
+5oW3 51348
+5oW+ 51349
+5oaC 51350
+5oaK 51351
+5oaL 51352
+5oaN 51353
+5oaO 51354
+5oaQ 51355
+5oaR 51356
+5oaU 51357
+5oaW 51358
+5oaZ 51359
+5oaa 51360
+5oak 51361
+5oan 51362
+5oao 51363
+5oap 51364
+5oar 51365
+5oas 51366
+5oau 51367
+5oay 51368
+5oa2 51369
+5oa3 51370
+5oa6 51371
+5oa+ 51372
+5oeC 51373
+5oeD 51374
+5oeG 51375
+5oeH 51376
+5oeI 51377
+5oeJ 51378
+5oeK 51379
+5oeL 51380
+5oeQ 51381
+5oeR 51382
+5oeS 51383
+5oej 51384
+5oem 51385
+5oey 51386
+5oe1 51387
+5oe2 51388
+5oe3 51389
+5oe4 51390
+5oe6 51391
+5oe8 51392
+5oe+ 51393
+5oe/ 51394
+5oiA 51395
+5oiG 51396
+5oiH 51397
+5oiI 51398
+5oiK 51399
+5oiM 51400
+5oiN 51401
+5oiO 51402
+5oiP 51403
+5oiQ 51404
+5oiR 51405
+5oiS 51406
+5oiU 51407
+5oiV 51408
+5oiW 51409
+5oiX 51410
+5oiY 51411
+5oia 51412
+5oib 51413
+5oie 51414
+5oif 51415
+5oih 51416
+5oii 51417
+5oil 51418
+5oim 51419
+5oiq 51420
+5ois 51421
+5oiu 51422
+5oiv 51423
+5oiw 51424
+5oiy 51425
+5oiz 51426
+5oi0 51427
+5oi2 51428
+5oi3 51429
+5oi4 51430
+5oi7 51431
+5oi9 51432
+5oi+ 51433
+5oi/ 51434
+5omA 51435
+5omB 51436
+5omD 51437
+5omG 51438
+5omH 51439
+5omI 51440
+5omJ 51441
+5omL 51442
+5omN 51443
+5omO 51444
+5omQ 51445
+5omR 51446
+5omS 51447
+5omT 51448
+5omU 51449
+5omV 51450
+5omY 51451
+5omb 51452
+5ome 51453
+5omg 51454
+5omj 51455
+5omk 51456
+5oml 51457
+5omm 51458
+5omn 51459
+5omo 51460
+5omp 51461
+5omq 51462
+5omr 51463
+5oms 51464
+5omt 51465
+5omu 51466
+5omv 51467
+5omw 51468
+5omx 51469
+5omz 51470
+5om2 51471
+5om5 51472
+5om8 51473
+5om+ 51474
+5om/ 51475
+5oqA 51476
+5oqD 51477
+5oqE 51478
+5oqF 51479
+5oqJ 51480
+5oqK 51481
+5oqM 51482
+5oqR 51483
+5oqS 51484
+5oqT 51485
+5oqU 51486
+5oqV 51487
+5oqW 51488
+5oqX 51489
+5oqY 51490
+5oqa 51491
+5oqb 51492
+5oqc 51493
+5oqe 51494
+5oqf 51495
+5oqg 51496
+5oqh 51497
+5oqi 51498
+5oqk 51499
+5oql 51500
+5oqo 51501
+5oqr 51502
+5oqs 51503
+5oqx 51504
+5oq1 51505
+5oq5 51506
+5oq7 51507
+5oq8 51508
+5oq9 51509
+5oq+ 51510
+5oq/ 51511
+5ouC 51512
+5ouD 51513
+5ouE 51514
+5ouF 51515
+5ouG 51516
+5ouH 51517
+5ouI 51518
+5ouJ 51519
+5ouK 51520
+5ouL 51521
+5ouM 51522
+5ouN 51523
+5ouO 51524
+5ouQ 51525
+5ouS 51526
+5ouT 51527
+5ouU 51528
+5ouW 51529
+5ouX 51530
+5ouY 51531
+5ouZ 51532
+5oua 51533
+5oub 51534
+5ouc 51535
+5oud 51536
+5ouf 51537
+5oug 51538
+5ouh 51539
+5oui 51540
+5ouj 51541
+5ouk 51542
+5oul 51543
+5oum 51544
+5oun 51545
+5ouo 51546
+5oup 51547
+5ous 51548
+5out 51549
+5ouu 51550
+5ouv 51551
+5oux 51552
+5ouz 51553
+5ou0 51554
+5ou1 51555
+5ou2 51556
+5ou3 51557
+5ou8 51558
+5ou9 51559
+5ou+ 51560
+5ou/ 51561
+5oyB 51562
+5oyC 51563
+5oyH 51564
+5oyI 51565
+5oyJ 51566
+5oyM 51567
+5oyO 51568
+5oyR 51569
+5oyS 51570
+5oyW 51571
+5oyZ 51572
+5oya 51573
+5oyb 51574
+5oyd 51575
+5oye 51576
+5oyf 51577
+5oyg 51578
+5oyh 51579
+5oyi 51580
+5oyj 51581
+5oyk 51582
+5oyl 51583
+5oyo 51584
+5oyq 51585
+5oyr 51586
+5oyv 51587
+5oyy 51588
+5oy5 51589
+5oy6 51590
+5oy8 51591
+5oy9 51592
+5oy+ 51593
+5oy/ 51594
+5o2C 51595
+5o2F 51596
+5o2G 51597
+5o2J 51598
+5o2K 51599
+5o2L 51600
+5o2M 51601
+5o2N 51602
+5o2O 51603
+5o2P 51604
+5o2Q 51605
+5o2V 51606
+5o2X 51607
+5o2c 51608
+5o2e 51609
+5o2f 51610
+5o2h 51611
+5o2i 51612
+5o2j 51613
+5o2n 51614
+5o2o 51615
+5o2p 51616
+5o2r 51617
+5o2t 51618
+5o2u 51619
+5o2v 51620
+5o2x 51621
+5o2y 51622
+5o22 51623
+5o23 51624
+5o26 51625
+5o27 51626
+5o6A 51627
+5o6C 51628
+5o6D 51629
+5o6H 51630
+5o6I 51631
+5o6J 51632
+5o6K 51633
+5o6M 51634
+5o6O 51635
+5o6P 51636
+5o6Q 51637
+5o6S 51638
+5o6V 51639
+5o6W 51640
+5o6Y 51641
+5o6Z 51642
+5o6b 51643
+5o6e 51644
+5o6f 51645
+5o6g 51646
+5o6h 51647
+5o6i 51648
+5o6j 51649
+5o6l 51650
+5o6n 51651
+5o6o 51652
+5o6p 51653
+5o6q 51654
+5o6s 51655
+5o6u 51656
+5o6w 51657
+5o6y 51658
+5o6z 51659
+5o60 51660
+5o63 51661
+5o64 51662
+5o65 51663
+5o66 51664
+5o67 51665
+5o68 51666
+5o6+ 51667
+5o+A 51668
+5o+D 51669
+5o+E 51670
+5o+G 51671
+5o+I 51672
+5o+J 51673
+5o+N 51674
+5o+O 51675
+5o+P 51676
+5o+Q 51677
+5o+S 51678
+5o+W 51679
+5o+a 51680
+5o+b 51681
+5o+e 51682
+5o+g 51683
+5o+h 51684
+5o+j 51685
+5o+p 51686
+5o+q 51687
+5o+t 51688
+5o+u 51689
+5o+y 51690
+5o+0 51691
+5o+1 51692
+5o+2 51693
+5o+4 51694
+5o+6 51695
+5o+8 51696
+5o+9 51697
+5o+/ 51698
+5pCA 51699
+5pCB 51700
+5pCC 51701
+5pCF 51702
+5pCG 51703
+5pCL 51704
+5pCM 51705
+5pCN 51706
+5pCP 51707
+5pCQ 51708
+5pCT 51709
+5pCU 51710
+5pCW 51711
+5pCX 51712
+5pCb 51713
+5pCc 51714
+5pCe 51715
+5pCg 51716
+5pCh 51717
+5pCi 51718
+5pCj 51719
+5pCm 51720
+5pCn 51721
+5pCo 51722
+5pCq 51723
+5pCs 51724
+5pCt 51725
+5pC0 51726
+5pC1 51727
+5pC2 51728
+5pC5 51729
+5pC6 51730
+5pC9 51731
+5pC+ 51732
+5pGB 51733
+5pGC 51734
+5pGE 51735
+5pGF 51736
+5pGG 51737
+5pGH 51738
+5pGI 51739
+5pGK 51740
+5pGQ 51741
+5pGS 51742
+5pGU 51743
+5pGY 51744
+5pGb 51745
+5pGe 51746
+5pGf 51747
+5pGn 51748
+5pGp 51749
+5pGt 51750
+5pGv 51751
+5pG1 51752
+5pG4 51753
+5pG5 51754
+5pG6 51755
+5pG9 51756
+5pKC 51757
+5pKD 51758
+5pKE 51759
+5pKF 51760
+5pKH 51761
+5pKI 51762
+5pKQ 51763
+5pKR 51764
+5pKS 51765
+5pKT 51766
+5pKV 51767
+5pKW 51768
+5pKZ 51769
+5pKa 51770
+5pKe 51771
+5pKk 51772
+5pKl 51773
+5pKp 51774
+5pKr 51775
+5pKs 51776
+5pKt 51777
+5pKu 51778
+5pKw 51779
+5pKy 51780
+5pKz 51781
+5pK0 51782
+5pK1 51783
+5pK3 51784
+5pK4 51785
+5pK5 51786
+5pK6 51787
+5pK7 51788
+5pK8 51789
+5pK/ 51790
+5pOA 51791
+5pOB 51792
+5pOC 51793
+5pOF 51794
+5pOH 51795
+5pOK 51796
+5pOL 51797
+5pON 51798
+5pOO 51799
+5pOQ 51800
+5pOS 51801
+5pOU 51802
+5pOX 51803
+5pOY 51804
+5pOa 51805
+5pOe 51806
+5pOg 51807
+5pOh 51808
+5pOi 51809
+5pOj 51810
+5pOk 51811
+5pOm 51812
+5pOn 51813
+5pOs 51814
+5pOv 51815
+5pOw 51816
+5pOx 51817
+5pOy 51818
+5pOz 51819
+5pO0 51820
+5pO2 51821
+5pO6 51822
+5pO7 51823
+5pO9 51824
+5pO+ 51825
+5pSA 51826
+5pSG 51827
+5pSJ 51828
+5pSP 51829
+5pSS 51830
+5pSU 51831
+5pSY 51832
+5pSZ 51833
+5pSc 51834
+5pSd 51835
+5pSe 51836
+5pSj 51837
+5pSk 51838
+5pSl 51839
+5pSq 51840
+5pSr 51841
+5pSs 51842
+5pSu 51843
+5pSv 51844
+5pSw 51845
+5pS0 51846
+5pS1 51847
+5pS2 51848
+5pS3 51849
+5pS4 51850
+5pS5 51851
+5pS7 51852
+5pS+ 51853
+5pS/ 51854
+5pWF 51855
+5pWI 51856
+5pWJ 51857
+5pWM 51858
+5pWN 51859
+5pWP 51860
+5pWR 51861
+5pWV 51862
+5pWW 51863
+5pWX 51864
+5pWY 51865
+5pWZ 51866
+5pWb 51867
+5pWd 51868
+5pWe 51869
+5pWi 51870
+5pWj 51871
+5pWm 51872
+5pWr 51873
+5pWs 51874
+5pWw 51875
+5pWy 51876
+5pW0 51877
+5pW1 51878
+5pW3 51879
+5pW4 51880
+5paB 51881
+5paC 51882
+5paD 51883
+5paH 51884
+5paJ 51885
+5paL 51886
+5paM 51887
+5paO 51888
+5paQ 51889
+5paR 51890
+5paT 51891
+5paV 51892
+5paX 51893
+5paZ 51894
+5pab 51895
+5pac 51896
+5pad 51897
+5paf 51898
+5pah 51899
+5pak 51900
+5pal 51901
+5pan 51902
+5pap 51903
+5par 51904
+5pas 51905
+5pat 51906
+5pav 51907
+5paw 51908
+5pa3 51909
+5pa5 51910
+5pa8 51911
+5pa9 51912
+5pa/ 51913
+5peB 51914
+5peD 51915
+5peE 51916
+5peF 51917
+5peG 51918
+5peJ 51919
+5peL 51920
+5peM 51921
+5peO 51922
+5peP 51923
+5peS 51924
+5peW 51925
+5peX 51926
+5peb 51927
+5peg 51928
+5pei 51929
+5pel 51930
+5pem 51931
+5pen 51932
+5peo 51933
+5pep 51934
+5pes 51935
+5pet 51936
+5peu 51937
+5pev 51938
+5pew 51939
+5pex 51940
+5pez 51941
+5pe2 51942
+5pe3 51943
+5pe4 51944
+5pe6 51945
+5pe7 51946
+5pe8 51947
+5piA 51948
+5piC 51949
+5piD 51950
+5piG 51951
+5piH 51952
+5piJ 51953
+5piK 51954
+5piM 51955
+5piO 51956
+5piP 51957
+5piT 51958
+5piU 51959
+5piV 51960
+5piZ 51961
+5pia 51962
+5pid 51963
+5pie 51964
+5pif 51965
+5pig 51966
+5pil 51967
+5pin 51968
+5pio 51969
+5pit 51970
+5piv 51971
+5pix 51972
+5pi0 51973
+5pi1 51974
+5pi2 51975
+5pi6 51976
+5pi8 51977
+5pi+ 51978
+5pmB 51979
+5pmC 51980
+5pmD 51981
+5pmE 51982
+5pmJ 51983
+5pmL 51984
+5pmM 51985
+5pmP 51986
+5pmS 51987
+5pmT 51988
+5pmU 51989
+5pmV 51990
+5pmW 51991
+5pmX 51992
+5pma 51993
+5pmd 51994
+5pme 51995
+5pmf 51996
+5pmh 51997
+5pmi 51998
+5pmk 51999
+5pml 52000
+5pmm 52001
+5pmn 52002
+5pmo 52003
+5pmp 52004
+5pms 52005
+5pmu 52006
+5pmv 52007
+5pmw 52008
+5pmz 52009
+5pm0 52010
+5pm2 52011
+5pm3 52012
+5pm6 52013
+5pm7 52014
+5pm+ 52015
+5pqB 52016
+5pqC 52017
+5pqE 52018
+5pqH 52019
+5pqI 52020
+5pqJ 52021
+5pqM 52022
+5pqN 52023
+5pqO 52024
+5pqR 52025
+5pqW 52026
+5pqX 52027
+5pqY 52028
+5pqd 52029
+5pqi 52030
+5pqm 52031
+5pqn 52032
+5pqo 52033
+5pqr 52034
+5pqu 52035
+5pqy 52036
+5pq0 52037
+5pq5 52038
+5pq7 52039
+5pq8 52040
+5pq+ 52041
+5puE 52042
+5puG 52043
+5puH 52044
+5puI 52045
+5puJ 52046
+5puM 52047
+5puW 52048
+5puZ 52049
+5pua 52050
+5pub 52051
+5puc 52052
+5pud 52053
+5pug 52054
+5puh 52055
+5pum 52056
+5pup 52057
+5pus 52058
+5puw 52059
+5pux 52060
+5puy 52061
+5puz 52062
+5pu0 52063
+5pu1 52064
+5pu3 52065
+5pu4 52066
+5pu5 52067
+5pu6 52068
+5pu8 52069
+5pu9 52070
+5pu+ 52071
+5pu/ 52072
+5pyA 52073
+5pyD 52074
+5pyI 52075
+5pyJ 52076
+5pyK 52077
+5pyL 52078
+5pyN 52079
+5pyP 52080
+5pyQ 52081
+5pyT 52082
+5pyU 52083
+5pyV 52084
+5pyX 52085
+5pyb 52086
+5pyd 52087
+5pyf 52088
+5pym 52089
+5pyn 52090
+5pyo 52091
+5pyq 52092
+5pyr 52093
+5pys 52094
+5pyt 52095
+5pyu 52096
+5pyv 52097
+5pyx 52098
+5py0 52099
+5py1 52100
+5py2 52101
+5py4 52102
+5py6 52103
+5py9 52104
+5py+ 52105
+5p2A 52106
+5p2B 52107
+5p2C 52108
+5p2D 52109
+5p2G 52110
+5p2I 52111
+5p2J 52112
+5p2M 52113
+5p2O 52114
+5p2P 52115
+5p2Q 52116
+5p2R 52117
+5p2T 52118
+5p2W 52119
+5p2Z 52120
+5p2c 52121
+5p2e 52122
+5p2f 52123
+5p2g 52124
+5p2h 52125
+5p2i 52126
+5p2j 52127
+5p2k 52128
+5p2l 52129
+5p2o 52130
+5p2p 52131
+5p2q 52132
+5p2s 52133
+5p2t 52134
+5p2u 52135
+5p2v 52136
+5p2w 52137
+5p2x 52138
+5p2y 52139
+5p2z 52140
+5p21 52141
+5p23 52142
+5p27 52143
+5p28 52144
+5p2+ 52145
+5p2/ 52146
+5p6B 52147
+5p6E 52148
+5p6F 52149
+5p6H 52150
+5p6J 52151
+5p6L 52152
+5p6M 52153
+5p6O 52154
+5p6Q 52155
+5p6T 52156
+5p6V 52157
+5p6X 52158
+5p6Y 52159
+5p6a 52160
+5p6c 52161
+5p6d 52162
+5p6e 52163
+5p6g 52164
+5p6h 52165
+5p6i 52166
+5p6j 52167
+5p6l 52168
+5p6m 52169
+5p6n 52170
+5p6o 52171
+5p6q 52172
+5p6r 52173
+5p6t 52174
+5p6v 52175
+5p6w 52176
+5p6x 52177
+5p6y 52178
+5p6z 52179
+5p60 52180
+5p62 52181
+5p63 52182
+5p64 52183
+5p65 52184
+5p+A 52185
+5p+B 52186
+5p+D 52187
+5p+E 52188
+5p+H 52189
+5p+K 52190
+5p+P 52191
+5p+Q 52192
+5p+R 52193
+5p+S 52194
+5p+T 52195
+5p+U 52196
+5p+Y 52197
+5p+a 52198
+5p+c 52199
+5p+d 52200
+5p+e 52201
+5p+g 52202
+5p+i 52203
+5p+k 52204
+5p+l 52205
+5p+n 52206
+5p+p 52207
+5p+s 52208
+5p+v 52209
+5p+w 52210
+5p+x 52211
+5p+z 52212
+5p+0 52213
+5p+1 52214
+5p+3 52215
+5p+4 52216
+5p+5 52217
+5p+7 52218
+5p+9 52219
+5p++ 52220
+5p+/ 52221
+5qCA 52222
+5qCC 52223
+5qCD 52224
+5qCE 52225
+5qCF 52226
+5qCH 52227
+5qCI 52228
+5qCJ 52229
+5qCK 52230
+5qCL 52231
+5qCM 52232
+5qCO 52233
+5qCP 52234
+5qCR 52235
+5qCT 52236
+5qCW 52237
+5qCX 52238
+5qCY 52239
+5qCd 52240
+5qCe 52241
+5qCf 52242
+5qCh 52243
+5qCi 52244
+5qCn 52245
+5qCp 52246
+5qCq 52247
+5qCr 52248
+5qCs 52249
+5qCx 52250
+5qCy 52251
+5qCz 52252
+5qC0 52253
+5qC3 52254
+5qC4 52255
+5qC5 52256
+5qC7 52257
+5qC8 52258
+5qC9 52259
+5qC+ 52260
+5qGA 52261
+5qGB 52262
+5qGC 52263
+5qGD 52264
+5qGE 52265
+5qGF 52266
+5qGG 52267
+5qGI 52268
+5qGJ 52269
+5qGM 52270
+5qGO 52271
+5qGQ 52272
+5qGR 52273
+5qGT 52274
+5qGU 52275
+5qGV 52276
+5qGW 52277
+5qGZ 52278
+5qGc 52279
+5qGd 52280
+5qGf 52281
+5qGg 52282
+5qGh 52283
+5qGi 52284
+5qGj 52285
+5qGk 52286
+5qGl 52287
+5qGm 52288
+5qGn 52289
+5qGo 52290
+5qGp 52291
+5qGr 52292
+5qGv 52293
+5qG0 52294
+5qG2 52295
+5qG3 52296
+5qG8 52297
+5qG+ 52298
+5qG/ 52299
+5qKB 52300
+5qKD 52301
+5qKF 52302
+5qKG 52303
+5qKI 52304
+5qKP 52305
+5qKT 52306
+5qKU 52307
+5qKX 52308
+5qKb 52309
+5qKd 52310
+5qKf 52311
+5qKg 52312
+5qKh 52313
+5qKi 52314
+5qKj 52315
+5qKm 52316
+5qKn 52317
+5qKo 52318
+5qKq 52319
+5qKr 52320
+5qKt 52321
+5qKv 52322
+5qKw 52323
+5qKx 52324
+5qKz 52325
+5qK1 52326
+5qK2 52327
+5qK5 52328
+5qK6 52329
+5qK7 52330
+5qK8 52331
+5qK/ 52332
+5qOA 52333
+5qOC 52334
+5qOE 52335
+5qOF 52336
+5qOI 52337
+5qOJ 52338
+5qOK 52339
+5qOL 52340
+5qON 52341
+5qOS 52342
+5qOU 52343
+5qOV 52344
+5qOX 52345
+5qOY 52346
+5qOa 52347
+5qOf 52348
+5qOg 52349
+5qOh 52350
+5qOj 52351
+5qOn 52352
+5qOo 52353
+5qOs 52354
+5qOu 52355
+5qOv 52356
+5qOw 52357
+5qOx 52358
+5qOy 52359
+5qO1 52360
+5qO2 52361
+5qO3 52362
+5qO5 52363
+5qO6 52364
+5qO7 52365
+5qO8 52366
+5qSA 52367
+5qSB 52368
+5qSF 52369
+5qSI 52370
+5qSK 52371
+5qSL 52372
+5qSM 52373
+5qSN 52374
+5qSO 52375
+5qSP 52376
+5qSQ 52377
+5qSS 52378
+5qSZ 52379
+5qSa 52380
+5qSb 52381
+5qSc 52382
+5qSf 52383
+5qSh 52384
+5qSj 52385
+5qSk 52386
+5qSl 52387
+5qSo 52388
+5qSq 52389
+5qSt 52390
+5qSw 52391
+5qS0 52392
+5qS5 52393
+5qS9 52394
+5qS/ 52395
+5qWA 52396
+5qWC 52397
+5qWJ 52398
+5qWK 52399
+5qWT 52400
+5qWU 52401
+5qWV 52402
+5qWX 52403
+5qWZ 52404
+5qWa 52405
+5qWc 52406
+5qWd 52407
+5qWe 52408
+5qWg 52409
+5qWh 52410
+5qWi 52411
+5qWj 52412
+5qWm 52413
+5qWq 52414
+5qWr 52415
+5qWt 52416
+5qWu 52417
+5qWv 52418
+5qWx 52419
+5qWz 52420
+5qW1 52421
+5qW3 52422
+5qW4 52423
+5qW5 52424
+5qW8 52425
+5qW9 52426
+5qW+ 52427
+5qaB 52428
+5qaC 52429
+5qaE 52430
+5qaG 52431
+5qaH 52432
+5qaI 52433
+5qaJ 52434
+5qaK 52435
+5qaO 52436
+5qaR 52437
+5qaU 52438
+5qaV 52439
+5qaW 52440
+5qaY 52441
+5qab 52442
+5qac 52443
+5qan 52444
+5qao 52445
+5qar 52446
+5qat 52447
+5qau 52448
+5qay 52449
+5qa0 52450
+5qa3 52451
+5qa7 52452
+5qa8 52453
+5qa+ 52454
+5qa/ 52455
+5qeB 52456
+5qeD 52457
+5qeF 52458
+5qeH 52459
+5qeK 52460
+5qeL 52461
+5qeM 52462
+5qeN 52463
+5qeO 52464
+5qeQ 52465
+5qeT 52466
+5qeW 52467
+5qeY 52468
+5qeZ 52469
+5qeb 52470
+5qef 52471
+5qen 52472
+5qeo 52473
+5qep 52474
+5qer 52475
+5qet 52476
+5qey 52477
+5qez 52478
+5qe7 52479
+5qe9 52480
+5qe/ 52481
+5qiA 52482
+5qiB 52483
+5qiC 52484
+5qiF 52485
+5qiJ 52486
+5qiK 52487
+5qiL 52488
+5qiR 52489
+5qiS 52490
+5qiT 52491
+5qiW 52492
+5qiX 52493
+5qiY 52494
+5qiZ 52495
+5qib 52496
+5qie 52497
+5qif 52498
+5qih 52499
+5qij 52500
+5qio 52501
+5qip 52502
+5qiq 52503
+5qir 52504
+5qiu 52505
+5qiv 52506
+5qix 52507
+5qi1 52508
+5qi4 52509
+5qi5 52510
+5qi6 52511
+5qi9 52512
+5qi+ 52513
+5qmE 52514
+5qmF 52515
+5qmH 52516
+5qmI 52517
+5qmL 52518
+5qmQ 52519
+5qmY 52520
+5qmZ 52521
+5qmb 52522
+5qme 52523
+5qmf 52524
+5qmh 52525
+5qmi 52526
+5qml 52527
+5qmm 52528
+5qmr 52529
+5qmx 52530
+5qmy 52531
+5qm5 52532
+5qm8 52533
+5qm/ 52534
+5qqA 52535
+5qqE 52536
+5qqJ 52537
+5qqN 52538
+5qqO 52539
+5qqQ 52540
+5qqU 52541
+5qqX 52542
+5qqc 52543
+5qqe 52544
+5qqg 52545
+5qqi 52546
+5qqj 52547
+5qqp 52548
+5qqq 52549
+5qqr 52550
+5qqs 52551
+5qqu 52552
+5qqv 52553
+5qqz 52554
+5qq4 52555
+5qq7 52556
+5qq/ 52557
+5quB 52558
+5quC 52559
+5quD 52560
+5quI 52561
+5quT 52562
+5qua 52563
+5qub 52564
+5que 52565
+5quf 52566
+5qun 52567
+5quo 52568
+5quq 52569
+5qu6 52570
+5qu7 52571
+5qyE 52572
+5qyF 52573
+5qyK 52574
+5qyS 52575
+5qyW 52576
+5qyd 52577
+5qye 52578
+5qyg 52579
+5qyh 52580
+5qyi 52581
+5qyj 52582
+5qyk 52583
+5qyn 52584
+5qyy 52585
+5qy3 52586
+5qy4 52587
+5qy5 52588
+5qy6 52589
+5qy7 52590
+5qy9 52591
+5qy+ 52592
+5q2A 52593
+5q2D 52594
+5q2G 52595
+5q2H 52596
+5q2J 52597
+5q2M 52598
+5q2O 52599
+5q2Q 52600
+5q2T 52601
+5q2U 52602
+5q2Y 52603
+5q2Z 52604
+5q2h 52605
+5q2i 52606
+5q2j 52607
+5q2k 52608
+5q2l 52609
+5q2m 52610
+5q2n 52611
+5q2p 52612
+5q2q 52613
+5q2t 52614
+5q2v 52615
+5q2y 52616
+5q2z 52617
+5q20 52618
+5q23 52619
+5q24 52620
+5q25 52621
+5q27 52622
+5q28 52623
+5q2/ 52624
+5q6B 52625
+5q6C 52626
+5q6D 52627
+5q6E 52628
+5q6G 52629
+5q6H 52630
+5q6J 52631
+5q6K 52632
+5q6L 52633
+5q6N 52634
+5q6S 52635
+5q6T 52636
+5q6W 52637
+5q6Y 52638
+5q6a 52639
+5q6b 52640
+5q6h 52641
+5q6q 52642
+5q6t 52643
+5q6u 52644
+5q6v 52645
+5q6y 52646
+5q6z 52647
+5q60 52648
+5q61 52649
+5q63 52650
+5q66 52651
+5q67 52652
+5q68 52653
+5q69 52654
+5q6/ 52655
+5q+A 52656
+5q+B 52657
+5q+C 52658
+5q+F 52659
+5q+G 52660
+5q+L 52661
+5q+N 52662
+5q+O 52663
+5q+P 52664
+5q+Q 52665
+5q+S 52666
+5q+T 52667
+5q+U 52668
+5q+V 52669
+5q+W 52670
+5q+X 52671
+5q+Y 52672
+5q+Z 52673
+5q+b 52674
+5q+f 52675
+5q+h 52676
+5q+r 52677
+5q+s 52678
+5q+v 52679
+5q+z 52680
+5q+5 52681
+5q+9 52682
+5rCF 52683
+5rCG 52684
+5rCH 52685
+5rCI 52686
+5rCN 52687
+5rCP 52688
+5rCQ 52689
+5rCR 52690
+5rCT 52691
+5rCU 52692
+5rCV 52693
+5rCW 52694
+5rCX 52695
+5rCY 52696
+5rCZ 52697
+5rCa 52698
+5rCb 52699
+5rCf 52700
+5rCh 52701
+5rCi 52702
+5rCj 52703
+5rCk 52704
+5rCm 52705
+5rCn 52706
+5rCo 52707
+5rCp 52708
+5rCq 52709
+5rCu 52710
+5rCv 52711
+5rCw 52712
+5rCy 52713
+5rC0 52714
+5rC3 52715
+5rC4 52716
+5rC5 52717
+5rC9 52718
+5rC+ 52719
+5rGA 52720
+5rGB 52721
+5rGC 52722
+5rGG 52723
+5rGH 52724
+5rGI 52725
+5rGJ 52726
+5rGK 52727
+5rGO 52728
+5rGQ 52729
+5rGU 52730
+5rGV 52731
+5rGX 52732
+5rGa 52733
+5rGb 52734
+5rGc 52735
+5rGd 52736
+5rGe 52737
+5rGf 52738
+5rGg 52739
+5rGh 52740
+5rGk 52741
+5rGo 52742
+5rGp 52743
+5rGq 52744
+5rGt 52745
+5rGw 52746
+5rGy 52747
+5rG0 52748
+5rG1 52749
+5rG2 52750
+5rG5 52751
+5rG6 52752
+5rG9 52753
+5rG+ 52754
+5rKB 52755
+5rKC 52756
+5rKD 52757
+5rKE 52758
+5rKF 52759
+5rKG 52760
+5rKI 52761
+5rKJ 52762
+5rKM 52763
+5rKP 52764
+5rKQ 52765
+5rKS 52766
+5rKT 52767
+5rKU 52768
+5rKW 52769
+5rKZ 52770
+5rKa 52771
+5rKb 52772
+5rKf 52773
+5rKh 52774
+5rKi 52775
+5rKj 52776
+5rKk 52777
+5rKl 52778
+5rKm 52779
+5rKn 52780
+5rKo 52781
+5rKp 52782
+5rKq 52783
+5rKr 52784
+5rKt 52785
+5rKu 52786
+5rKx 52787
+5rKz 52788
+5rK4 52789
+5rK5 52790
+5rK7 52791
+5rK8 52792
+5rK9 52793
+5rK+ 52794
+5rK/ 52795
+5rOB 52796
+5rOD 52797
+5rOE 52798
+5rOF 52799
+5rOG 52800
+5rOJ 52801
+5rOK 52802
+5rOM 52803
+5rOQ 52804
+5rOS 52805
+5rOT 52806
+5rOU 52807
+5rOV 52808
+5rOW 52809
+5rOX 52810
+5rOZ 52811
+5rOa 52812
+5rOb 52813
+5rOe 52814
+5rOg 52815
+5rOh 52816
+5rOi 52817
+5rOj 52818
+5rOl 52819
+5rOo 52820
+5rOq 52821
+5rOr 52822
+5rOs 52823
+5rOu 52824
+5rOv 52825
+5rOw 52826
+5rOx 52827
+5rOz 52828
+5rO1 52829
+5rO3 52830
+5rO4 52831
+5rO6 52832
+5rO7 52833
+5rO8 52834
+5rO9 52835
+5rO+ 52836
+5rSB 52837
+5rSE 52838
+5rSH 52839
+5rSL 52840
+5rSM 52841
+5rSO 52842
+5rSP 52843
+5rSR 52844
+5rSS 52845
+5rSX 52846
+5rSZ 52847
+5rSb 52848
+5rSe 52849
+5rSf 52850
+5rSj 52851
+5rSl 52852
+5rSn 52853
+5rSp 52854
+5rSq 52855
+5rSu 52856
+5rSx 52857
+5rSy 52858
+5rSz 52859
+5rS1 52860
+5rS2 52861
+5rS4 52862
+5rS5 52863
+5rS6 52864
+5rS7 52865
+5rS8 52866
+5rS9 52867
+5rS+ 52868
+5rS/ 52869
+5rWB 52870
+5rWD 52871
+5rWE 52872
+5rWF 52873
+5rWG 52874
+5rWH 52875
+5rWI 52876
+5rWJ 52877
+5rWK 52878
+5rWL 52879
+5rWN 52880
+5rWO 52881
+5rWP 52882
+5rWQ 52883
+5rWR 52884
+5rWS 52885
+5rWT 52886
+5rWU 52887
+5rWZ 52888
+5rWa 52889
+5rWb 52890
+5rWc 52891
+5rWe 52892
+5rWg 52893
+5rWj 52894
+5rWl 52895
+5rWm 52896
+5rWp 52897
+5rWq 52898
+5rWs 52899
+5rWu 52900
+5rWv 52901
+5rW0 52902
+5rW3 52903
+5rW4 52904
+5rW5 52905
+5raC 52906
+5raF 52907
+5raI 52908
+5raJ 52909
+5raM 52910
+5raO 52911
+5raR 52912
+5raT 52913
+5raU 52914
+5raV 52915
+5raW 52916
+5raY 52917
+5raZ 52918
+5rab 52919
+5rac 52920
+5rad 52921
+5rae 52922
+5raf 52923
+5rag 52924
+5rah 52925
+5raj 52926
+5rak 52927
+5ram 52928
+5ran 52929
+5rao 52930
+5rap 52931
+5raq 52932
+5rau 52933
+5rav 52934
+5ray 52935
+5ra1 52936
+5ra4 52937
+5ra8 52938
+5ra/ 52939
+5reA 52940
+5reE 52941
+5reF 52942
+5reG 52943
+5reH 52944
+5reL 52945
+5reM 52946
+5reP 52947
+5reR 52948
+5reS 52949
+5reW 52950
+5reY 52951
+5reZ 52952
+5rea 52953
+5rec 52954
+5red 52955
+5ree 52956
+5reg 52957
+5reh 52958
+5rek 52959
+5rem 52960
+5reo 52961
+5req 52962
+5rer 52963
+5res 52964
+5reu 52965
+5rev 52966
+5rex 52967
+5rez 52968
+5re1 52969
+5re3 52970
+5re4 52971
+5re5 52972
+5re6 52973
+5re7 52974
+5re8 52975
+5riA 52976
+5riF 52977
+5riH 52978
+5riI 52979
+5riJ 52980
+5riK 52981
+5riL 52982
+5riM 52983
+5riN 52984
+5riO 52985
+5riQ 52986
+5riR 52987
+5riT 52988
+5riU 52989
+5riV 52990
+5riX 52991
+5riZ 52992
+5ria 52993
+5rib 52994
+5rid 52995
+5rif 52996
+5rig 52997
+5rih 52998
+5rij 52999
+5rik 53000
+5ril 53001
+5rim 53002
+5rip 53003
+5rir 53004
+5ris 53005
+5rit 53006
+5riv 53007
+5riy 53008
+5ri0 53009
+5ri4 53010
+5ri6 53011
+5ri+ 53012
+5rmD 53013
+5rmE 53014
+5rmJ 53015
+5rmK 53016
+5rmN 53017
+5rmO 53018
+5rmT 53019
+5rmU 53020
+5rmW 53021
+5rmY 53022
+5rmb 53023
+5rmc 53024
+5rmd 53025
+5rmf 53026
+5rmj 53027
+5rmn 53028
+5rmr 53029
+5rmu 53030
+5rmv 53031
+5rmy 53032
+5rm0 53033
+5rm2 53034
+5rm+ 53035
+5rm/ 53036
+5rqA 53037
+5rqC 53038
+5rqD 53039
+5rqF 53040
+5rqG 53041
+5rqH 53042
+5rqJ 53043
+5rqM 53044
+5rqN 53045
+5rqP 53046
+5rqQ 53047
+5rqW 53048
+5rqY 53049
+5rqc 53050
+5rqd 53051
+5rqf 53052
+5rqi 53053
+5rql 53054
+5rqn 53055
+5rqq 53056
+5rqr 53057
+5rqv 53058
+5rqx 53059
+5rqy 53060
+5rq0 53061
+5rq2 53062
+5rq3 53063
+5rq6 53064
+5rq9 53065
+5ruB 53066
+5ruC 53067
+5ruE 53068
+5ruF 53069
+5ruH 53070
+5ruI 53071
+5ruJ 53072
+5ruL 53073
+5ruM 53074
+5ruP 53075
+5ruR 53076
+5ruT 53077
+5ruU 53078
+5ruV 53079
+5ruX 53080
+5ruY 53081
+5rua 53082
+5rud 53083
+5rue 53084
+5ruf 53085
+5rug 53086
+5ruh 53087
+5rui 53088
+5ruk 53089
+5rul 53090
+5rum 53091
+5ruo 53092
+5rup 53093
+5rus 53094
+5ruv 53095
+5ruy 53096
+5ru0 53097
+5ru4 53098
+5ru5 53099
+5ru+ 53100
+5ru/ 53101
+5ryB 53102
+5ryC 53103
+5ryG 53104
+5ryI 53105
+5ryJ 53106
+5ryP 53107
+5ryR 53108
+5ryT 53109
+5ryU 53110
+5ryV 53111
+5ryg 53112
+5ryi 53113
+5ryj 53114
+5ryp 53115
+5ryq 53116
+5ryr 53117
+5rys 53118
+5ryt 53119
+5ryv 53120
+5ryx 53121
+5ryy 53122
+5ryz 53123
+5ry0 53124
+5ry2 53125
+5ry3 53126
+5ry4 53127
+5ry8 53128
+5ry+ 53129
+5ry/ 53130
+5r2F 53131
+5r2G 53132
+5r2H 53133
+5r2L 53134
+5r2N 53135
+5r2O 53136
+5r2P 53137
+5r2R 53138
+5r2U 53139
+5r2Y 53140
+5r2b 53141
+5r2c 53142
+5r2e 53143
+5r2f 53144
+5r2h 53145
+5r2i 53146
+5r2k 53147
+5r2l 53148
+5r2m 53149
+5r2p 53150
+5r2s 53151
+5r2t 53152
+5r2u 53153
+5r2w 53154
+5r2y 53155
+5r20 53156
+5r21 53157
+5r24 53158
+5r26 53159
+5r28 53160
+5r6A 53161
+5r6B 53162
+5r6E 53163
+5r6G 53164
+5r6I 53165
+5r6M 53166
+5r6N 53167
+5r6O 53168
+5r6X 53169
+5r6c 53170
+5r6h 53171
+5r6j 53172
+5r6k 53173
+5r6n 53174
+5r6q 53175
+5r6x 53176
+5r6z 53177
+5r60 53178
+5r62 53179
+5r65 53180
+5r+A 53181
+5r+B 53182
+5r+C 53183
+5r+D 53184
+5r+G 53185
+5r+J 53186
+5r+R 53187
+5r+S 53188
+5r+V 53189
+5r+Y 53190
+5r+b 53191
+5r+e 53192
+5r+f 53193
+5r+g 53194
+5r+h 53195
+5r+k 53196
+5r+p 53197
+5r+r 53198
+5r+s 53199
+5r+u 53200
+5r+v 53201
+5r+x 53202
+5r+2 53203
+5r+6 53204
+5r++ 53205
+54CJ 53206
+54CL 53207
+54CN 53208
+54CP 53209
+54CR 53210
+54CV 53211
+54Ca 53212
+54Cb 53213
+54Cd 53214
+54Ce 53215
+54Cf 53216
+54Cj 53217
+54Cm 53218
+54Cn 53219
+54Cs 53220
+54Cw 53221
+54Cy 53222
+54C5 53223
+54C+ 53224
+54GM 53225
+54GP 53226
+54GR 53227
+54GY 53228
+54Ge 53229
+54Gj 53230
+54Gr 53231
+54Gt 53232
+54Gv 53233
+54Gw 53234
+54G1 53235
+54G2 53236
+54G4 53237
+54G8 53238
+54G9 53239
+54G+ 53240
+54G/ 53241
+54KA 53242
+54KB 53243
+54KF 53244
+54KG 53245
+54KJ 53246
+54KK 53247
+54KO 53248
+54KS 53249
+54KU 53250
+54KV 53251
+54KW 53252
+54KZ 53253
+54Kc 53254
+54Kd 53255
+54Kf 53256
+54Kk 53257
+54Kq 53258
+54Kr 53259
+54Ks 53260
+54Kt 53261
+54Ku 53262
+54Kv 53263
+54Kz 53264
+54K0 53265
+54K3 53266
+54K4 53267
+54K5 53268
+54K6 53269
+54K8 53270
+54K9 53271
+54OA 53272
+54OB 53273
+54OC 53274
+54OD 53275
+54OI 53276
+54OK 53277
+54OP 53278
+54OW 53279
+54OY 53280
+54OZ 53281
+54Ob 53282
+54Oc 53283
+54Of 53284
+54Ok 53285
+54Om 53286
+54On 53287
+54Oo 53288
+54Op 53289
+54Or 53290
+54Os 53291
+54Ot 53292
+54Ov 53293
+54Ox 53294
+54O3 53295
+54O5 53296
+54O9 53297
+54SJ 53298
+54SK 53299
+54SQ 53300
+54ST 53301
+54SU 53302
+54SV 53303
+54SW 53304
+54SX 53305
+54SY 53306
+54SZ 53307
+54Sa 53308
+54Sc 53309
+54Sh 53310
+54Sm 53311
+54Sv 53312
+54Sw 53313
+54Sx 53314
+54S2 53315
+54S8 53316
+54WF 53317
+54WJ 53318
+54WK 53319
+54WL 53320
+54WM 53321
+54WO 53322
+54WV 53323
+54WW 53324
+54WZ 53325
+54Wc 53326
+54We 53327
+54Wf 53328
+54Wi 53329
+54Wk 53330
+54Wl 53331
+54Wm 53332
+54Wn 53333
+54Wo 53334
+54Wp 53335
+54Ws 53336
+54Wu 53337
+54Wy 53338
+54Wz 53339
+54W4 53340
+54W6 53341
+54W9 53342
+54W/ 53343
+54aE 53344
+54aK 53345
+54aP 53346
+54aU 53347
+54aY 53348
+54aZ 53349
+54ac 53350
+54af 53351
+54ag 53352
+54ao 53353
+54as 53354
+54ax 53355
+54az 53356
+54a1 53357
+54a5 53358
+54a+ 53359
+54eD 53360
+54eE 53361
+54eI 53362
+54eJ 53363
+54eK 53364
+54eL 53365
+54eO 53366
+54eQ 53367
+54eS 53368
+54eU 53369
+54eV 53370
+54eX 53371
+54eZ 53372
+54ea 53373
+54ef 53374
+54eg 53375
+54el 53376
+54em 53377
+54en 53378
+54et 53379
+54eu 53380
+54e1 53381
+54e5 53382
+54e7 53383
+54e8 53384
+54e/ 53385
+54iG 53386
+54iN 53387
+54iQ 53388
+54ib 53389
+54io 53390
+54iq 53391
+54is 53392
+54it 53393
+54iw 53394
+54ix 53395
+54iy 53396
+54i1 53397
+54i2 53398
+54i3 53399
+54i4 53400
+54i5 53401
+54i6 53402
+54i7 53403
+54i8 53404
+54i9 53405
+54i+ 53406
+54i/ 53407
+54mA 53408
+54mB 53409
+54mC 53410
+54mG 53411
+54mH 53412
+54mI 53413
+54mM 53414
+54mN 53415
+54mS 53416
+54mV 53417
+54mW 53418
+54mY 53419
+54mZ 53420
+54mb 53421
+54md 53422
+54mf 53423
+54mg 53424
+54mh 53425
+54mi 53426
+54mk 53427
+54mm 53428
+54mn 53429
+54mp 53430
+54mu 53431
+54mv 53432
+54my 53433
+54m0 53434
+54m1 53435
+54m4 53436
+54m5 53437
+54m6 53438
+54m9 53439
+54m+ 53440
+54qA 53441
+54qB 53442
+54qC 53443
+54qE 53444
+54qH 53445
+54qK 53446
+54qL 53447
+54qN 53448
+54qS 53449
+54qW 53450
+54qf 53451
+54qg 53452
+54qi 53453
+54qn 53454
+54qs 53455
+54qv 53456
+54qw 53457
+54qz 53458
+54q0 53459
+54q2 53460
+54q3 53461
+54q4 53462
+54q5 53463
+54q8 53464
+54uA 53465
+54uC 53466
+54uE 53467
+54uG 53468
+54uI 53469
+54uM 53470
+54uN 53471
+54uO 53472
+54uQ 53473
+54uS 53474
+54uX 53475
+54uZ 53476
+54ub 53477
+54ud 53478
+54ue 53479
+54ug 53480
+54uh 53481
+54ui 53482
+54uo 53483
+54up 53484
+54us 53485
+54ut 53486
+54uu 53487
+54uw 53488
+54ux 53489
+54uy 53490
+54u0 53491
+54u3 53492
+54u4 53493
+54u5 53494
+54u7 53495
+54u8 53496
+54u9 53497
+54yB 53498
+54yH 53499
+54yK 53500
+54yO 53501
+54yV 53502
+54yW 53503
+54yX 53504
+54yZ 53505
+54yb 53506
+54yc 53507
+54yd 53508
+54ye 53509
+54yf 53510
+54yh 53511
+54yi 53512
+54yl 53513
+54yo 53514
+54yp 53515
+54yq 53516
+54yr 53517
+54ys 53518
+54yu 53519
+54yv 53520
+54yx 53521
+54yy 53522
+54y0 53523
+54y2 53524
+54y3 53525
+54y5 53526
+54y+ 53527
+54y/ 53528
+542E 53529
+542F 53530
+542O 53531
+542P 53532
+542Q 53533
+542S 53534
+542X 53535
+542g 53536
+542j 53537
+542o 53538
+542q 53539
+542s 53540
+542t 53541
+542w 53542
+542y 53543
+5421 53544
+5424 53545
+5426 53546
+5427 53547
+542+ 53548
+546E 53549
+546H 53550
+546J 53551
+546L 53552
+546O 53553
+546R 53554
+546V 53555
+546W 53556
+546Y 53557
+546Z 53558
+546a 53559
+546b 53560
+546f 53561
+546g 53562
+546i 53563
+546l 53564
+546m 53565
+546p 53566
+546r 53567
+546u 53568
+546v 53569
+546w 53570
+546y 53571
+546z 53572
+5463 53573
+5465 53574
+5466 53575
+5467 53576
+54+A 53577
+54+C 53578
+54+F 53579
+54+I 53580
+54+J 53581
+54+K 53582
+54+N 53583
+54+O 53584
+54+P 53585
+54+Q 53586
+54+R 53587
+54+Z 53588
+54+c 53589
+54+e 53590
+54+g 53591
+54+j 53592
+54+l 53593
+54+m 53594
+54+n 53595
+54+p 53596
+54+q 53597
+54+r 53598
+54+t 53599
+54+u 53600
+54+w 53601
+54+y 53602
+54+4 53603
+54+6 53604
+54+9 53605
+54++ 53606
+55CD 53607
+55CF 53608
+55CG 53609
+55CH 53610
+55CJ 53611
+55CK 53612
+55CN 53613
+55CO 53614
+55CP 53615
+55CQ 53616
+55Ca 53617
+55Cb 53618
+55Ci 53619
+55Ck 53620
+55Cl 53621
+55Cm 53622
+55Co 53623
+55Cq 53624
+55Cs 53625
+55Cu 53626
+55Cv 53627
+55Cw 53628
+55Cy 53629
+55Cz 53630
+55C0 53631
+55C1 53632
+55C2 53633
+55C6 53634
+55C8 53635
+55GA 53636
+55GB 53637
+55GE 53638
+55GV 53639
+55GX 53640
+55GZ 53641
+55Ga 53642
+55Gb 53643
+55Gc 53644
+55Ge 53645
+55Gf 53646
+55Gg 53647
+55Gj 53648
+55Gk 53649
+55Gp 53650
+55Gq 53651
+55Gt 53652
+55Gu 53653
+55Gv 53654
+55Gw 53655
+55Gx 53656
+55Gz 53657
+55G0 53658
+55G2 53659
+55G3 53660
+55G+ 53661
+55KA 53662
+55KB 53663
+55KD 53664
+55KH 53665
+55KI 53666
+55KL 53667
+55KO 53668
+55KQ 53669
+55KY 53670
+55Kc 53671
+55Kd 53672
+55Ke 53673
+55Kf 53674
+55Kg 53675
+55Kj 53676
+55Kn 53677
+55Ko 53678
+55Kp 53679
+55Kw 53680
+55K6 53681
+55K9 53682
+55OK 53683
+55OP 53684
+55OS 53685
+55OU 53686
+55OY 53687
+55Oc 53688
+55Og 53689
+55Oi 53690
+55Oj 53691
+55Ok 53692
+55Om 53693
+55Or 53694
+55Ou 53695
+55Ov 53696
+55O0 53697
+55O2 53698
+55O3 53699
+55O/ 53700
+55SD 53701
+55SE 53702
+55SM 53703
+55SN 53704
+55SO 53705
+55SP 53706
+55SR 53707
+55ST 53708
+55SV 53709
+55SY 53710
+55SZ 53711
+55Sa 53712
+55Sc 53713
+55Sf 53714
+55Si 53715
+55Sj 53716
+55Sl 53717
+55Sm 53718
+55So 53719
+55Sp 53720
+55Sq 53721
+55Sr 53722
+55Ss 53723
+55St 53724
+55Sv 53725
+55Sw 53726
+55Sx 53727
+55Sy 53728
+55Sz 53729
+55S0 53730
+55S1 53731
+55S3 53732
+55S4 53733
+55S6 53734
+55S7 53735
+55S+ 53736
+55WA 53737
+55WF 53738
+55WI 53739
+55WJ 53740
+55WK 53741
+55WL 53742
+55WM 53743
+55WO 53744
+55WP 53745
+55WR 53746
+55WU 53747
+55WZ 53748
+55Wa 53749
+55Wb 53750
+55Wc 53751
+55Wd 53752
+55Wg 53753
+55Wi 53754
+55Wk 53755
+55Wl 53756
+55Wm 53757
+55Wq 53758
+55Wr 53759
+55Ws 53760
+55Wt 53761
+55Wv 53762
+55Ww 53763
+55Wy 53764
+55Wz 53765
+55W0 53766
+55W1 53767
+55W2 53768
+55W3 53769
+55W4 53770
+55W5 53771
+55W/ 53772
+55aD 53773
+55aG 53774
+55aH 53775
+55aK 53776
+55aL 53777
+55aN 53778
+55aO 53779
+55aP 53780
+55aR 53781
+55aU 53782
+55aW 53783
+55aX 53784
+55aZ 53785
+55aa 53786
+55ad 53787
+55af 53788
+55ag 53789
+55ah 53790
+55aj 53791
+55ak 53792
+55al 53793
+55ar 53794
+55as 53795
+55at 53796
+55au 53797
+55av 53798
+55aw 53799
+55ax 53800
+55ay 53801
+55az 53802
+55a0 53803
+55a1 53804
+55a4 53805
+55a5 53806
+55a8 53807
+55a9 53808
+55a+ 53809
+55eC 53810
+55eD 53811
+55eE 53812
+55eF 53813
+55eH 53814
+55eI 53815
+55eJ 53816
+55eK 53817
+55eN 53818
+55eS 53819
+55eU 53820
+55eV 53821
+55eY 53822
+55eZ 53823
+55eb 53824
+55ee 53825
+55ei 53826
+55ej 53827
+55ek 53828
+55em 53829
+55en 53830
+55eo 53831
+55ep 53832
+55eq 53833
+55er 53834
+55ew 53835
+55ex 53836
+55ey 53837
+55ez 53838
+55e0 53839
+55e5 53840
+55e6 53841
+55e8 53842
+55e+ 53843
+55e/ 53844
+55iA 53845
+55iB 53846
+55iF 53847
+55iG 53848
+55iK 53849
+55iL 53850
+55iM 53851
+55iN 53852
+55iT 53853
+55iV 53854
+55iX 53855
+55iY 53856
+55iZ 53857
+55id 53858
+55if 53859
+55ig 53860
+55ih 53861
+55ii 53862
+55ik 53863
+55im 53864
+55in 53865
+55ip 53866
+55iq 53867
+55ir 53868
+55iw 53869
+55iz 53870
+55i0 53871
+55i1 53872
+55i4 53873
+55i7 53874
+55i8 53875
+55i+ 53876
+55i/ 53877
+55mA 53878
+55mC 53879
+55mD 53880
+55mG 53881
+55mH 53882
+55mI 53883
+55mM 53884
+55mN 53885
+55mO 53886
+55mS 53887
+55mU 53888
+55mW 53889
+55mY 53890
+55mc 53891
+55me 53892
+55mh 53893
+55mi 53894
+55mj 53895
+55ml 53896
+55mn 53897
+55mp 53898
+55mq 53899
+55mr 53900
+55ms 53901
+55mu 53902
+55mv 53903
+55mw 53904
+55mx 53905
+55my 53906
+55m4 53907
+55m6 53908
+55m7 53909
+55m8 53910
+55m9 53911
+55m+ 53912
+55qA 53913
+55qC 53914
+55qE 53915
+55qG 53916
+55qH 53917
+55qI 53918
+55qL 53919
+55qM 53920
+55qO 53921
+55qQ 53922
+55qR 53923
+55qT 53924
+55qV 53925
+55qW 53926
+55qZ 53927
+55qa 53928
+55qd 53929
+55qe 53930
+55qk 53931
+55qm 53932
+55qu 53933
+55qv 53934
+55qw 53935
+55qx 53936
+55qy 53937
+55q0 53938
+55q3 53939
+55q4 53940
+55q5 53941
+55q6 53942
+55q/ 53943
+55uC 53944
+55uD 53945
+55uF 53946
+55uG 53947
+55uI 53948
+55uJ 53949
+55uK 53950
+55uN 53951
+55uO 53952
+55uP 53953
+55uQ 53954
+55uR 53955
+55uS 53956
+55uU 53957
+55uW 53958
+55uX 53959
+55uY 53960
+55ub 53961
+55uc 53962
+55ue 53963
+55uf 53964
+55uh 53965
+55uj 53966
+55uk 53967
+55ul 53968
+55un 53969
+55up 53970
+55uq 53971
+55uu 53972
+55uv 53973
+55ux 53974
+55uy 53975
+55u0 53976
+55u4 53977
+55u5 53978
+55u7 53979
+55u8 53980
+55u+ 53981
+55yA 53982
+55yB 53983
+55yE 53984
+55yH 53985
+55yI 53986
+55yJ 53987
+55yL 53988
+55yM 53989
+55yN 53990
+55yZ 53991
+55ya 53992
+55yb 53993
+55yc 53994
+55ye 53995
+55yf 53996
+55yg 53997
+55yl 53998
+55ym 53999
+55yo 54000
+55yp 54001
+55ys 54002
+55yt 54003
+55yv 54004
+55y1 54005
+55y2 54006
+55y3 54007
+55y4 54008
+55y6 54009
+55y8 54010
+55y+ 54011
+552A 54012
+552B 54013
+552D 54014
+552G 54015
+552H 54016
+552Q 54017
+552R 54018
+552a 54019
+552b 54020
+552c 54021
+552e 54022
+552f 54023
+552h 54024
+552i 54025
+552j 54026
+552l 54027
+552m 54028
+552o 54029
+552r 54030
+552s 54031
+552x 54032
+5525 54033
+5526 54034
+5529 54035
+552+ 54036
+552/ 54037
+556A 54038
+556E 54039
+556F 54040
+556G 54041
+556H 54042
+556L 54043
+556M 54044
+556O 54045
+556R 54046
+556S 54047
+556T 54048
+556e 54049
+556f 54050
+556g 54051
+556i 54052
+556l 54053
+556n 54054
+556p 54055
+556q 54056
+556s 54057
+556t 54058
+556w 54059
+556z 54060
+5561 54061
+5567 54062
+5568 54063
+5569 54064
+556+ 54065
+556/ 54066
+55+H 54067
+55+N 54068
+55+T 54069
+55+X 54070
+55+a 54071
+55+b 54072
+55+c 54073
+55+i 54074
+55+j 54075
+55+l 54076
+55+n 54077
+55+p 54078
+55+r 54079
+55+s 54080
+55+t 54081
+55+u 54082
+55+v 54083
+55+x 54084
+55+z 54085
+55+2 54086
+55+4 54087
+55+8 54088
+55+9 54089
+55++ 54090
+55+/ 54091
+56CA 54092
+56CB 54093
+56CC 54094
+56CK 54095
+56CM 54096
+56CN 54097
+56CS 54098
+56CU 54099
+56CV 54100
+56CW 54101
+56CX 54102
+56CY 54103
+56Ca 54104
+56Cc 54105
+56Cd 54106
+56Cf 54107
+56Cg 54108
+56Cj 54109
+56Cl 54110
+56Cm 54111
+56Cn 54112
+56Cp 54113
+56Cr 54114
+56Cs 54115
+56Ct 54116
+56Cw 54117
+56Cy 54118
+56C0 54119
+56C1 54120
+56C3 54121
+56C4 54122
+56C6 54123
+56C7 54124
+56C8 54125
+56C+ 54126
+56C/ 54127
+56GA 54128
+56GB 54129
+56GF 54130
+56GM 54131
+56GQ 54132
+56GS 54133
+56GV 54134
+56GW 54135
+56GX 54136
+56Ga 54137
+56Gd 54138
+56Ge 54139
+56Gq 54140
+56Gr 54141
+56Gs 54142
+56Gt 54143
+56Gu 54144
+56Gv 54145
+56Gy 54146
+56G0 54147
+56G3 54148
+56G8 54149
+56G/ 54150
+56KB 54151
+56KG 54152
+56KH 54153
+56KJ 54154
+56KM 54155
+56KN 54156
+56KO 54157
+56KR 54158
+56KT 54159
+56KV 54160
+56KX 54161
+56KY 54162
+56Ka 54163
+56Kb 54164
+56Kc 54165
+56Kf 54166
+56Kh 54167
+56Kj 54168
+56Kn 54169
+56Kp 54170
+56Kq 54171
+56Kw 54172
+56Kx 54173
+56Ky 54174
+56Kz 54175
+56K0 54176
+56K6 54177
+56K8 54178
+56K+ 54179
+56OB 54180
+56OF 54181
+56OG 54182
+56OJ 54183
+56OK 54184
+56OL 54185
+56OQ 54186
+56OR 54187
+56OS 54188
+56OU 54189
+56OV 54190
+56OZ 54191
+56Oa 54192
+56Oh 54193
+56On 54194
+56Oo 54195
+56Os 54196
+56Ov 54197
+56Oy 54198
+56O0 54199
+56O3 54200
+56O6 54201
+56O7 54202
+56O+ 54203
+56SB 54204
+56SF 54205
+56SM 54206
+56SO 54207
+56SS 54208
+56ST 54209
+56SZ 54210
+56Se 54211
+56Sm 54212
+56Sq 54213
+56Sr 54214
+56Ss 54215
+56S0 54216
+56S6 54217
+56S7 54218
+56S8 54219
+56S9 54220
+56S+ 54221
+56WA 54222
+56WB 54223
+56WG 54224
+56WH 54225
+56WI 54226
+56WJ 54227
+56WO 54228
+56WP 54229
+56WQ 54230
+56WT 54231
+56WW 54232
+56WX 54233
+56Wa 54234
+56Wb 54235
+56Wc 54236
+56Wd 54237
+56We 54238
+56Wf 54239
+56Wg 54240
+56Wi 54241
+56Wl 54242
+56Wn 54243
+56Wo 54244
+56Wt 54245
+56Wv 54246
+56W3 54247
+56W4 54248
+56W6 54249
+56W8 54250
+56W/ 54251
+56aA 54252
+56aB 54253
+56aE 54254
+56aF 54255
+56aK 54256
+56aN 54257
+56aO 54258
+56aP 54259
+56ab 54260
+56al 54261
+56am 54262
+56an 54263
+56ao 54264
+56ap 54265
+56aq 54266
+56au 54267
+56aw 54268
+56ax 54269
+56az 54270
+56a5 54271
+56a6 54272
+56a7 54273
+56a9 54274
+56a+ 54275
+56a/ 54276
+56eA 54277
+56eB 54278
+56eC 54279
+56eD 54280
+56eG 54281
+56eJ 54282
+56eL 54283
+56eN 54284
+56eP 54285
+56eR 54286
+56eS 54287
+56eV 54288
+56eY 54289
+56ef 54290
+56eh 54291
+56ej 54292
+56ek 54293
+56em 54294
+56en 54295
+56ep 54296
+56er 54297
+56es 54298
+56et 54299
+56ev 54300
+56ew 54301
+56e4 54302
+56e7 54303
+56e9 54304
+56e+ 54305
+56iA 54306
+56iC 54307
+56iF 54308
+56iI 54309
+56iK 54310
+56iL 54311
+56iN 54312
+56iO 54313
+56iU 54314
+56iX 54315
+56iZ 54316
+56ia 54317
+56ic 54318
+56ie 54319
+56if 54320
+56ig 54321
+56ij 54322
+56iu 54323
+56ix 54324
+56iy 54325
+56iz 54326
+56i3 54327
+56i5 54328
+56i7 54329
+56i8 54330
+56i9 54331
+56i/ 54332
+56mA 54333
+56mC 54334
+56mG 54335
+56mH 54336
+56mM 54337
+56mN 54338
+56mO 54339
+56mP 54340
+56mQ 54341
+56mR 54342
+56mX 54343
+56mh 54344
+56mi 54345
+56mj 54346
+56mp 54347
+56mr 54348
+56mw 54349
+56m0 54350
+56m2 54351
+56m3 54352
+56m4 54353
+56m5 54354
+56m6 54355
+56m9 54356
+56m/ 54357
+56qA 54358
+56qB 54359
+56qD 54360
+56qE 54361
+56qI 54362
+56qN 54363
+56qO 54364
+56qR 54365
+56qS 54366
+56qT 54367
+56qV 54368
+56qW 54369
+56qX 54370
+56qY 54371
+56qc 54372
+56qd 54373
+56qf 54374
+56qg 54375
+56qh 54376
+56qj 54377
+56ql 54378
+56qm 54379
+56qo 54380
+56qp 54381
+56qq 54382
+56qt 54383
+56qu 54384
+56qv 54385
+56qz 54386
+56q2 54387
+56q4 54388
+56q6 54389
+56q+ 54390
+56q/ 54391
+56uD 54392
+56uE 54393
+56uF 54394
+56uH 54395
+56uI 54396
+56uK 54397
+56uL 54398
+56uR 54399
+56uW 54400
+56uZ 54401
+56uc 54402
+56ud 54403
+56ue 54404
+56uf 54405
+56ug 54406
+56ui 54407
+56uj 54408
+56ul 54409
+56um 54410
+56uq 54411
+56ut 54412
+56uv 54413
+56uy 54414
+56u2 54415
+56u5 54416
+56u6 54417
+56u9 54418
+56u/ 54419
+56yC 54420
+56yD 54421
+56yE 54422
+56yG 54423
+56yI 54424
+56yK 54425
+56yL 54426
+56yP 54427
+56yR 54428
+56yU 54429
+56yV 54430
+56yW 54431
+56yY 54432
+56yZ 54433
+56yb 54434
+56ye 54435
+56yg 54436
+56yk 54437
+56yl 54438
+56ym 54439
+56yo 54440
+56yp 54441
+56yq 54442
+56yr 54443
+56ys 54444
+56yz 54445
+56y1 54446
+56y4 54447
+56y5 54448
+56y6 54449
+56y8 54450
+56y+ 54451
+562F 54452
+562G 54453
+562H 54454
+562I 54455
+562J 54456
+562K 54457
+562L 54458
+562M 54459
+562N 54460
+562P 54461
+562Q 54462
+562R 54463
+562S 54464
+562U 54465
+562W 54466
+562Y 54467
+562a 54468
+562b 54469
+562c 54470
+562d 54471
+562g 54472
+562i 54473
+562l 54474
+562n 54475
+562s 54476
+562u 54477
+562w 54478
+562x 54479
+562y 54480
+5621 54481
+5623 54482
+5625 54483
+5626 54484
+5627 54485
+5628 54486
+562+ 54487
+566A 54488
+566F 54489
+566G 54490
+566H 54491
+566L 54492
+566N 54493
+566P 54494
+566Q 54495
+566S 54496
+566T 54497
+566U 54498
+566V 54499
+566X 54500
+566Z 54501
+566a 54502
+566c 54503
+566d 54504
+566f 54505
+566h 54506
+566i 54507
+566l 54508
+566m 54509
+566n 54510
+566o 54511
+566p 54512
+566q 54513
+566r 54514
+566s 54515
+566t 54516
+566x 54517
+5660 54518
+5664 54519
+5667 54520
+5668 54521
+566+ 54522
+56+A 54523
+56+B 54524
+56+E 54525
+56+G 54526
+56+H 54527
+56+J 54528
+56+L 54529
+56+M 54530
+56+R 54531
+56+T 54532
+56+Z 54533
+56+a 54534
+56+d 54535
+56+g 54536
+56+h 54537
+56+k 54538
+56+l 54539
+56+m 54540
+56+p 54541
+56+q 54542
+56+t 54543
+56+u 54544
+56+x 54545
+56+z 54546
+56+2 54547
+56+3 54548
+56+8 54549
+56++ 54550
+57CA 54551
+57CD 54552
+57CH 54553
+57CL 54554
+57CM 54555
+57CN 54556
+57CP 54557
+57CR 54558
+57CS 54559
+57CT 54560
+57CU 54561
+57CW 54562
+57CX 54563
+57Cf 54564
+57Ch 54565
+57Cj 54566
+57Cm 54567
+57Cn 54568
+57Cq 54569
+57Cr 54570
+57Cw 54571
+57C3 54572
+57C4 54573
+57C6 54574
+57C9 54575
+57C+ 54576
+57C/ 54577
+57GA 54578
+57GB 54579
+57GD 54580
+57GH 54581
+57GM 54582
+57GN 54583
+57GP 54584
+57GQ 54585
+57GT 54586
+57GU 54587
+57GW 54588
+57Gf 54589
+57Gg 54590
+57Gj 54591
+57Gk 54592
+57Gs 54593
+57Gu 54594
+57Gy 54595
+57Gz 54596
+57G7 54597
+57G8 54598
+57G9 54599
+57G+ 54600
+57KB 54601
+57KC 54602
+57KD 54603
+57KJ 54604
+57KL 54605
+57KN 54606
+57KR 54607
+57KS 54608
+57KV 54609
+57KX 54610
+57KY 54611
+57Kb 54612
+57Kd 54613
+57Kf 54614
+57Ki 54615
+57Kk 54616
+57Kl 54617
+57Km 54618
+57Kn 54619
+57Kq 54620
+57Kt 54621
+57Ku 54622
+57Kx 54623
+57Ky 54624
+57Kz 54625
+57K1 54626
+57K5 54627
+57K8 54628
+57K9 54629
+57K+ 54630
+57OA 54631
+57OB 54632
+57OF 54633
+57OK 54634
+57OM 54635
+57ON 54636
+57OO 54637
+57OS 54638
+57OV 54639
+57OW 54640
+57OX 54641
+57OY 54642
+57OZ 54643
+57Oc 54644
+57Oe 54645
+57Of 54646
+57Og 54647
+57Oi 54648
+57On 54649
+57Oo 54650
+57Ov 54651
+57Oy 54652
+57O2 54653
+57O4 54654
+57O6 54655
+57O7 54656
+57O+ 54657
+57SA 54658
+57SC 54659
+57SE 54660
+57SF 54661
+57SG 54662
+57SJ 54663
+57SK 54664
+57SL 54665
+57SN 54666
+57SQ 54667
+57SU 54668
+57SX 54669
+57SY 54670
+57SZ 54671
+57Sa 54672
+57Sb 54673
+57Sc 54674
+57Sg 54675
+57Sh 54676
+57Si 54677
+57Sn 54678
+57Sr 54679
+57Ss 54680
+57Su 54681
+57Sv 54682
+57Sw 54683
+57Sy 54684
+57Sz 54685
+57S1 54686
+57S5 54687
+57S6 54688
+57WA 54689
+57WC 54690
+57WD 54691
+57WE 54692
+57WF 54693
+57WG 54694
+57WL 54695
+57WM 54696
+57WO 54697
+57WP 54698
+57WQ 54699
+57WV 54700
+57WW 54701
+57Wb 54702
+57Wc 54703
+57We 54704
+57Wh 54705
+57Wi 54706
+57Wj 54707
+57Wm 54708
+57Wo 54709
+57Wu 54710
+57Wv 54711
+57Wx 54712
+57Wy 54713
+57Wz 54714
+57W1 54715
+57W2 54716
+57W5 54717
+57W6 54718
+57W9 54719
+57aB 54720
+57aJ 54721
+57aP 54722
+57aT 54723
+57aZ 54724
+57aa 54725
+57ab 54726
+57ac 54727
+57af 54728
+57ag 54729
+57ai 54730
+57aj 54731
+57am 54732
+57as 54733
+57at 54734
+57au 54735
+57av 54736
+57aw 54737
+57ax 54738
+57ay 54739
+57a0 54740
+57a1 54741
+57a2 54742
+57a4 54743
+57a6 54744
+57a7 54745
+57a9 54746
+57a+ 54747
+57a/ 54748
+57eH 54749
+57eK 54750
+57eL 54751
+57eP 54752
+57eR 54753
+57eS 54754
+57eY 54755
+57ea 54756
+57eb 54757
+57ed 54758
+57ee 54759
+57eg 54760
+57eh 54761
+57ej 54762
+57ek 54763
+57eo 54764
+57ep 54765
+57es 54766
+57ev 54767
+57ey 54768
+57e0 54769
+57e7 54770
+57iB 54771
+57iE 54772
+57iF 54773
+57iI 54774
+57iJ 54775
+57iK 54776
+57iL 54777
+57iS 54778
+57ib 54779
+57ie 54780
+57if 54781
+57ig 54782
+57ii 54783
+57ij 54784
+57im 54785
+57ir 54786
+57iu 54787
+57ix 54788
+57iy 54789
+57i0 54790
+57i1 54791
+57i3 54792
+57i5 54793
+57i6 54794
+57i7 54795
+57i9 54796
+57i+ 54797
+57mB 54798
+57mD 54799
+57mG 54800
+57mH 54801
+57mK 54802
+57mL 54803
+57mN 54804
+57mR 54805
+57mU 54806
+57mV 54807
+57mW 54808
+57mZ 54809
+57ma 54810
+57md 54811
+57me 54812
+57mh 54813
+57mn 54814
+57mp 54815
+57mq 54816
+57mr 54817
+57mt 54818
+57mw 54819
+57mz 54820
+57m5 54821
+57m7 54822
+57m8 54823
+57m9 54824
+57qC 54825
+57qH 54826
+57qI 54827
+57qM 54828
+57qN 54829
+57qP 54830
+57qQ 54831
+57qS 54832
+57qT 54833
+57qU 54834
+57qW 54835
+57qb 54836
+57qc 54837
+57qg 54838
+57qh 54839
+57qi 54840
+57qj 54841
+57qk 54842
+57ql 54843
+57qm 54844
+57qn 54845
+57qo 54846
+57qp 54847
+57qq 54848
+57qr 54849
+57qs 54850
+57qt 54851
+57qu 54852
+57qv 54853
+57qw 54854
+57qx 54855
+57qy 54856
+57qz 54857
+57q1 54858
+57q2 54859
+57q3 54860
+57q4 54861
+57q5 54862
+57q6 54863
+57q7 54864
+57q9 54865
+57q+ 54866
+57q/ 54867
+57uA 54868
+57uB 54869
+57uC 54870
+57uD 54871
+57uE 54872
+57uF 54873
+57uG 54874
+57uH 54875
+57uI 54876
+57uJ 54877
+57uK 54878
+57uL 54879
+57uM 54880
+57uN 54881
+57uO 54882
+57uP 54883
+57uQ 54884
+57uR 54885
+57uS 54886
+57uT 54887
+57uU 54888
+57uV 54889
+57uX 54890
+57uY 54891
+57uZ 54892
+57ua 54893
+57ub 54894
+57uc 54895
+57ud 54896
+57ue 54897
+57uf 54898
+57uh 54899
+57ui 54900
+57uj 54901
+57uk 54902
+57ul 54903
+57um 54904
+57un 54905
+57uo 54906
+57up 54907
+57uq 54908
+57ur 54909
+57ut 54910
+57uu 54911
+57uv 54912
+57uw 54913
+57ux 54914
+57uy 54915
+57uz 54916
+57u0 54917
+57u1 54918
+57u2 54919
+57u3 54920
+57u4 54921
+57u6 54922
+57u7 54923
+57u8 54924
+57u9 54925
+57u+ 54926
+57u/ 54927
+57yA 54928
+57yB 54929
+57yC 54930
+57yD 54931
+57yE 54932
+57yF 54933
+57yG 54934
+57yH 54935
+57yI 54936
+57yJ 54937
+57yK 54938
+57yL 54939
+57yM 54940
+57yO 54941
+57yR 54942
+57yS 54943
+57yT 54944
+57yU 54945
+57yV 54946
+57yW 54947
+57yX 54948
+57yY 54949
+57yZ 54950
+57ya 54951
+57yb 54952
+57yc 54953
+57yd 54954
+57yf 54955
+57yg 54956
+57yi 54957
+57yj 54958
+57yk 54959
+57yl 54960
+57ym 54961
+57yn 54962
+57yo 54963
+57yp 54964
+57yq 54965
+57yr 54966
+57ys 54967
+57yt 54968
+57yu 54969
+57yv 54970
+57yw 54971
+57yx 54972
+57yz 54973
+57y0 54974
+57y1 54975
+57y2 54976
+57y4 54977
+57y6 54978
+572C 54979
+572E 54980
+572F 54981
+572H 54982
+572M 54983
+572N 54984
+572Q 54985
+572R 54986
+572U 54987
+572V 54988
+572X 54989
+572Y 54990
+572a 54991
+572f 54992
+572g 54993
+572h 54994
+572i 54995
+572n 54996
+572o 54997
+572p 54998
+572q 54999
+572r 55000
+572u 55001
+572w 55002
+572y 55003
+5720 55004
+5721 55005
+5723 55006
+5725 55007
+572+ 55008
+576B 55009
+576C 55010
+576D 55011
+576F 55012
+576G 55013
+576H 55014
+576I 55015
+576K 55016
+576M 55017
+576O 55018
+576R 55019
+576U 55020
+576W 55021
+576X 55022
+576a 55023
+576d 55024
+576e 55025
+576f 55026
+576h 55027
+576j 55028
+576k 55029
+576n 55030
+576o 55031
+576p 55032
+576v 55033
+576w 55034
+576y 55035
+5762 55036
+5764 55037
+5765 55038
+5769 55039
+576/ 55040
+57+A 55041
+57+B 55042
+57+D 55043
+57+F 55044
+57+G 55045
+57+K 55046
+57+M 55047
+57+O 55048
+57+S 55049
+57+U 55050
+57+V 55051
+57+Y 55052
+57+a 55053
+57+b 55054
+57+f 55055
+57+g 55056
+57+h 55057
+57+l 55058
+57+m 55059
+57+p 55060
+57+r 55061
+57+u 55062
+57+w 55063
+57+x 55064
+57+z 55065
+57+5 55066
+57+7 55067
+57+8 55068
+6ICA 55069
+6ICB 55070
+6ICD 55071
+6ICE 55072
+6ICF 55073
+6ICG 55074
+6ICL 55075
+6ICM 55076
+6ICN 55077
+6ICQ 55078
+6ICS 55079
+6ICV 55080
+6ICX 55081
+6ICY 55082
+6ICZ 55083
+6ICc 55084
+6ICm 55085
+6ICn 55086
+6ICo 55087
+6ICq 55088
+6ICx 55089
+6ICz 55090
+6IC1 55091
+6IC2 55092
+6IC3 55093
+6IC4 55094
+6IC7 55095
+6IC9 55096
+6IC/ 55097
+6IGC 55098
+6IGD 55099
+6IGG 55100
+6IGK 55101
+6IGL 55102
+6IGM 55103
+6IGN 55104
+6IGS 55105
+6IGU 55106
+6IGW 55107
+6IGY 55108
+6IGa 55109
+6IGe 55110
+6IGf 55111
+6IGh 55112
+6IGi 55113
+6IGo 55114
+6IGp 55115
+6IGq 55116
+6IGv 55117
+6IGw 55118
+6IGy 55119
+6IGz 55120
+6IG0 55121
+6IG1 55122
+6IG2 55123
+6IG3 55124
+6IG9 55125
+6IG+ 55126
+6IG/ 55127
+6IKD 55128
+6IKE 55129
+6IKF 55130
+6IKG 55131
+6IKH 55132
+6IKJ 55133
+6IKL 55134
+6IKM 55135
+6IKP 55136
+6IKT 55137
+6IKW 55138
+6IKY 55139
+6IKa 55140
+6IKb 55141
+6IKd 55142
+6IKf 55143
+6IKg 55144
+6IKh 55145
+6IKi 55146
+6IKk 55147
+6IKl 55148
+6IKp 55149
+6IKq 55150
+6IKr 55151
+6IKs 55152
+6IKt 55153
+6IKu 55154
+6IKv 55155
+6IKx 55156
+6IKy 55157
+6IK0 55158
+6IK3 55159
+6IK4 55160
+6IK6 55161
+6IK8 55162
+6IK9 55163
+6IK+ 55164
+6IK/ 55165
+6IOA 55166
+6IOB 55167
+6IOD 55168
+6IOE 55169
+6IOG 55170
+6IOM 55171
+6ION 55172
+6IOO 55173
+6IOW 55174
+6IOX 55175
+6IOZ 55176
+6IOa 55177
+6IOb 55178
+6IOc 55179
+6IOd 55180
+6IOe 55181
+6IOh 55182
+6IOk 55183
+6IOl 55184
+6IOn 55185
+6IOq 55186
+6IOr 55187
+6IOs 55188
+6IOt 55189
+6IOv 55190
+6IOw 55191
+6IOx 55192
+6IOz 55193
+6IO0 55194
+6IO2 55195
+6IO4 55196
+6IO6 55197
+6IO8 55198
+6IO9 55199
+6ISB 55200
+6ISC 55201
+6ISF 55202
+6ISG 55203
+6ISH 55204
+6ISI 55205
+6ISJ 55206
+6ISK 55207
+6ISN 55208
+6ISP 55209
+6ISQ 55210
+6ISR 55211
+6ISS 55212
+6IST 55213
+6ISU 55214
+6ISW 55215
+6ISY 55216
+6ISa 55217
+6ISb 55218
+6ISj 55219
+6ISp 55220
+6ISr 55221
+6ISs 55222
+6ISv 55223
+6ISw 55224
+6ISx 55225
+6ISy 55226
+6ISz 55227
+6IS3 55228
+6IS4 55229
+6IS5 55230
+6IS+ 55231
+6IWG 55232
+6IWI 55233
+6IWK 55234
+6IWL 55235
+6IWM 55236
+6IWO 55237
+6IWQ 55238
+6IWR 55239
+6IWT 55240
+6IWU 55241
+6IWV 55242
+6IWY 55243
+6IWZ 55244
+6IWa 55245
+6IWf 55246
+6IWg 55247
+6IWl 55248
+6IWm 55249
+6IWn 55250
+6IWp 55251
+6IWr 55252
+6IWt 55253
+6IWu 55254
+6IWw 55255
+6IWx 55256
+6IWz 55257
+6IW0 55258
+6IW4 55259
+6IW5 55260
+6IW6 55261
+6IW7 55262
+6IW8 55263
+6IW+ 55264
+6IW/ 55265
+6IaA 55266
+6IaC 55267
+6IaD 55268
+6IaI 55269
+6IaK 55270
+6IaP 55271
+6IaR 55272
+6IaV 55273
+6IaY 55274
+6Iaa 55275
+6Iab 55276
+6Iac 55277
+6Iad 55278
+6Iag 55279
+6Iaj 55280
+6Iam 55281
+6Iao 55282
+6Iap 55283
+6Iaz 55284
+6Ia0 55285
+6Ia1 55286
+6Ia6 55287
+6Ia7 55288
+6Ia9 55289
+6Ia+ 55290
+6Ia/ 55291
+6IeA 55292
+6IeC 55293
+6IeD 55294
+6IeG 55295
+6IeH 55296
+6IeI 55297
+6IeJ 55298
+6IeK 55299
+6IeM 55300
+6IeN 55301
+6IeR 55302
+6IeT 55303
+6IeY 55304
+6IeZ 55305
+6Iea 55306
+6Iec 55307
+6Ief 55308
+6Iej 55309
+6Iel 55310
+6Ien 55311
+6Ieo 55312
+6Ieq 55313
+6Ies 55314
+6Iet 55315
+6Iez 55316
+6Ie0 55317
+6Ie6 55318
+6Ie7 55319
+6Ie8 55320
+6Ie+ 55321
+6IiA 55322
+6IiB 55323
+6IiC 55324
+6IiF 55325
+6IiG 55326
+6IiH 55327
+6IiI 55328
+6IiJ 55329
+6IiK 55330
+6IiM 55331
+6IiN 55332
+6IiO 55333
+6IiQ 55334
+6IiS 55335
+6IiU 55336
+6IiW 55337
+6IiX 55338
+6IiY 55339
+6Iib 55340
+6Iic 55341
+6Iie 55342
+6Iif 55343
+6Iii 55344
+6Iij 55345
+6Iip 55346
+6Iiq 55347
+6Iir 55348
+6Iis 55349
+6Iiu 55350
+6Iiv 55351
+6Iiw 55352
+6Iix 55353
+6Iiy 55354
+6Iiz 55355
+6Ii1 55356
+6Ii2 55357
+6Ii3 55358
+6Ii4 55359
+6Ii5 55360
+6Ii+ 55361
+6ImA 55362
+6ImE 55363
+6ImH 55364
+6ImJ 55365
+6ImL 55366
+6ImP 55367
+6ImY 55368
+6ImZ 55369
+6Imf 55370
+6Imk 55371
+6Imm 55372
+6Imo 55373
+6Imq 55374
+6Imr 55375
+6Imu 55376
+6Imv 55377
+6Imw 55378
+6Imx 55379
+6Imy 55380
+6Imz 55381
+6Im2 55382
+6Im3 55383
+6Im4 55384
+6Im5 55385
+6Im6 55386
+6Im9 55387
+6Im+ 55388
+6Im/ 55389
+6IqC 55390
+6IqD 55391
+6IqI 55392
+6IqK 55393
+6IqL 55394
+6IqN 55395
+6IqO 55396
+6IqR 55397
+6IqS 55398
+6IqX 55399
+6IqY 55400
+6IqZ 55401
+6Iqc 55402
+6Iqd 55403
+6Iqf 55404
+6Iqh 55405
+6Iqj 55406
+6Iqk 55407
+6Iql 55408
+6Iqm 55409
+6Iqo 55410
+6Iqp 55411
+6Iqq 55412
+6Iqr 55413
+6Iqs 55414
+6Iqt 55415
+6Iqu 55416
+6Iqv 55417
+6Iqw 55418
+6Iqx 55419
+6Iqz 55420
+6Iq3 55421
+6Iq4 55422
+6Iq5 55423
+6Iq7 55424
+6Iq9 55425
+6Iq+ 55426
+6IuB 55427
+6IuE 55428
+6IuF 55429
+6IuH 55430
+6IuI 55431
+6IuK 55432
+6IuL 55433
+6IuM 55434
+6IuN 55435
+6IuO 55436
+6IuP 55437
+6IuR 55438
+6IuS 55439
+6IuT 55440
+6IuU 55441
+6IuV 55442
+6IuX 55443
+6IuY 55444
+6IuZ 55445
+6Iub 55446
+6Iuc 55447
+6Iue 55448
+6Iuf 55449
+6Iuh 55450
+6Iui 55451
+6Iuj 55452
+6Iuk 55453
+6Iul 55454
+6Ium 55455
+6Iun 55456
+6Iur 55457
+6Iuv 55458
+6Iux 55459
+6Iuz 55460
+6Iu0 55461
+6Iu3 55462
+6Iu5 55463
+6Iu6 55464
+6Iu7 55465
+6Iu8 55466
+6Iu+ 55467
+6IyB 55468
+6IyC 55469
+6IyD 55470
+6IyE 55471
+6IyF 55472
+6IyG 55473
+6IyI 55474
+6IyJ 55475
+6IyM 55476
+6IyO 55477
+6IyP 55478
+6IyR 55479
+6IyU 55480
+6IyV 55481
+6IyX 55482
+6IyY 55483
+6Iya 55484
+6Iyb 55485
+6Iyc 55486
+6Iyj 55487
+6Iyn 55488
+6Iyo 55489
+6Iyr 55490
+6Iys 55491
+6Iyt 55492
+6Iyv 55493
+6Iyx 55494
+6Iyy 55495
+6Iyz 55496
+6Iy0 55497
+6Iy1 55498
+6Iy2 55499
+6Iy4 55500
+6Iy5 55501
+6Iy6 55502
+6Iy8 55503
+6I2A 55504
+6I2D 55505
+6I2F 55506
+6I2G 55507
+6I2H 55508
+6I2J 55509
+6I2K 55510
+6I2P 55511
+6I2Q 55512
+6I2R 55513
+6I2S 55514
+6I2U 55515
+6I2Y 55516
+6I2Z 55517
+6I2a 55518
+6I2b 55519
+6I2c 55520
+6I2e 55521
+6I2f 55522
+6I2g 55523
+6I2h 55524
+6I2j 55525
+6I2k 55526
+6I2l 55527
+6I2m 55528
+6I2n 55529
+6I2o 55530
+6I2p 55531
+6I2q 55532
+6I2r 55533
+6I2s 55534
+6I2v 55535
+6I2z 55536
+6I23 55537
+6I24 55538
+6I27 55539
+6I28 55540
+6I29 55541
+6I6F 55542
+6I6G 55543
+6I6H 55544
+6I6J 55545
+6I6K 55546
+6I6O 55547
+6I6S 55548
+6I6T 55549
+6I6V 55550
+6I6W 55551
+6I6Y 55552
+6I6Z 55553
+6I6a 55554
+6I6c 55555
+6I6e 55556
+6I6f 55557
+6I6g 55558
+6I6i 55559
+6I6n 55560
+6I6o 55561
+6I6p 55562
+6I6q 55563
+6I6r 55564
+6I6x 55565
+6I6y 55566
+6I6z 55567
+6I60 55568
+6I61 55569
+6I62 55570
+6I63 55571
+6I64 55572
+6I65 55573
+6I66 55574
+6I68 55575
+6I69 55576
+6I+A 55577
+6I+B 55578
+6I+F 55579
+6I+H 55580
+6I+J 55581
+6I+K 55582
+6I+M 55583
+6I+O 55584
+6I+P 55585
+6I+T 55586
+6I+U 55587
+6I+W 55588
+6I+Y 55589
+6I+c 55590
+6I+f 55591
+6I+g 55592
+6I+h 55593
+6I+l 55594
+6I+o 55595
+6I+p 55596
+6I+q 55597
+6I+r 55598
+6I+v 55599
+6I+w 55600
+6I+x 55601
+6I+y 55602
+6I+0 55603
+6I+4 55604
+6I+5 55605
+6I+9 55606
+6JCB 55607
+6JCD 55608
+6JCE 55609
+6JCG 55610
+6JCK 55611
+6JCL 55612
+6JCM 55613
+6JCN 55614
+6JCO 55615
+6JCP 55616
+6JCT 55617
+6JCY 55618
+6JCc 55619
+6JCd 55620
+6JCg 55621
+6JCi 55622
+6JCk 55623
+6JCl 55624
+6JCm 55625
+6JCn 55626
+6JCo 55627
+6JCp 55628
+6JCs 55629
+6JCx 55630
+6JC1 55631
+6JC4 55632
+6JC5 55633
+6JC8 55634
+6JC9 55635
+6JGG 55636
+6JGI 55637
+6JGJ 55638
+6JGO 55639
+6JGR 55640
+6JGX 55641
+6JGZ 55642
+6JGa 55643
+6JGb 55644
+6JGh 55645
+6JGj 55646
+6JGm 55647
+6JGp 55648
+6JGr 55649
+6JGs 55650
+6JGt 55651
+6JGv 55652
+6JGx 55653
+6JGz 55654
+6JG1 55655
+6JG2 55656
+6JG3 55657
+6JG4 55658
+6JG5 55659
+6JG6 55660
+6JKC 55661
+6JKE 55662
+6JKL 55663
+6JKM 55664
+6JKO 55665
+6JKQ 55666
+6JKU 55667
+6JKX 55668
+6JKZ 55669
+6JKc 55670
+6JKf 55671
+6JKh 55672
+6JKv 55673
+6JKy 55674
+6JK0 55675
+6JK4 55676
+6JK5 55677
+6JK6 55678
+6JK7 55679
+6JK8 55680
+6JK9 55681
+6JK/ 55682
+6JOA 55683
+6JOB 55684
+6JOC 55685
+6JOE 55686
+6JOG 55687
+6JOJ 55688
+6JOK 55689
+6JOL 55690
+6JON 55691
+6JOQ 55692
+6JOR 55693
+6JOT 55694
+6JOW 55695
+6JOZ 55696
+6JOa 55697
+6JOd 55698
+6JOf 55699
+6JOg 55700
+6JOj 55701
+6JOl 55702
+6JOm 55703
+6JOs 55704
+6JOu 55705
+6JO0 55706
+6JO8 55707
+6JO/ 55708
+6JSA 55709
+6JSM 55710
+6JSR 55711
+6JST 55712
+6JSU 55713
+6JSV 55714
+6JSX 55715
+6JSa 55716
+6JSf 55717
+6JSh 55718
+6JSj 55719
+6JSk 55720
+6JSl 55721
+6JSm 55722
+6JSr 55723
+6JSs 55724
+6JSt 55725
+6JS1 55726
+6JS3 55727
+6JS4 55728
+6JS5 55729
+6JS6 55730
+6JS7 55731
+6JS8 55732
+6JS9 55733
+6JWB 55734
+6JWD 55735
+6JWI 55736
+6JWJ 55737
+6JWK 55738
+6JWL 55739
+6JWO 55740
+6JWW 55741
+6JWX 55742
+6JWZ 55743
+6JWa 55744
+6JWe 55745
+6JWj 55746
+6JWk 55747
+6JWo 55748
+6JWp 55749
+6JWq 55750
+6JWr 55751
+6JWt 55752
+6JWy 55753
+6JW0 55754
+6JW3 55755
+6JW6 55756
+6JW7 55757
+6JW+ 55758
+6JaA 55759
+6JaB 55760
+6JaE 55761
+6JaF 55762
+6JaH 55763
+6JaI 55764
+6JaK 55765
+6JaP 55766
+6JaQ 55767
+6JaR 55768
+6JaU 55769
+6JaX 55770
+6JaZ 55771
+6Jab 55772
+6Jac 55773
+6Jai 55774
+6Jak 55775
+6Jam 55776
+6Jao 55777
+6Jap 55778
+6Jaq 55779
+6Jar 55780
+6Jas 55781
+6Jat 55782
+6Jau 55783
+6Jav 55784
+6Jaw 55785
+6Jaz 55786
+6Ja3 55787
+6Ja5 55788
+6Ja6 55789
+6JeB 55790
+6JeJ 55791
+6JeN 55792
+6JeP 55793
+6JeQ 55794
+6JeT 55795
+6JeV 55796
+6Jec 55797
+6Jed 55798
+6Jeg 55799
+6Jek 55800
+6Jel 55801
+6Jem 55802
+6Jep 55803
+6Jeq 55804
+6Je3 55805
+6Je5 55806
+6Je6 55807
+6Je7 55808
+6Je/ 55809
+6JiC 55810
+6JiF 55811
+6JiG 55812
+6JiH 55813
+6JiK 55814
+6JiL 55815
+6JiR 55816
+6JiT 55817
+6JiW 55818
+6JiX 55819
+6JiY 55820
+6Jia 55821
+6Jin 55822
+6Jit 55823
+6Jiw 55824
+6Ji4 55825
+6Ji8 55826
+6Ji/ 55827
+6JmO 55828
+6JmP 55829
+6JmQ 55830
+6JmR 55831
+6JmT 55832
+6JmU 55833
+6JmV 55834
+6Jma 55835
+6Jmb 55836
+6Jmc 55837
+6Jme 55838
+6Jmf 55839
+6Jmi 55840
+6Jmn 55841
+6Jmr 55842
+6Jms 55843
+6Jmu 55844
+6Jmx 55845
+6Jm1 55846
+6Jm5 55847
+6Jm6 55848
+6Jm7 55849
+6Jm9 55850
+6Jm+ 55851
+6Jm/ 55852
+6JqA 55853
+6JqB 55854
+6JqC 55855
+6JqK 55856
+6JqL 55857
+6JqM 55858
+6JqN 55859
+6JqT 55860
+6JqV 55861
+6Jqc 55862
+6Jqd 55863
+6Jqh 55864
+6Jqj 55865
+6Jqk 55866
+6Jqn 55867
+6Jqo 55868
+6Jqp 55869
+6Jqq 55870
+6Jqr 55871
+6Jqs 55872
+6Jqv 55873
+6Jqw 55874
+6Jqx 55875
+6Jqz 55876
+6Jq0 55877
+6Jq1 55878
+6Jq2 55879
+6JuA 55880
+6JuE 55881
+6JuG 55882
+6JuH 55883
+6JuJ 55884
+6JuK 55885
+6JuL 55886
+6JuN 55887
+6JuO 55888
+6JuP 55889
+6JuQ 55890
+6JuR 55891
+6JuU 55892
+6JuY 55893
+6JuZ 55894
+6Jub 55895
+6Jue 55896
+6Juf 55897
+6Juk 55898
+6Jup 55899
+6Jus 55900
+6Jut 55901
+6Juu 55902
+6Juv 55903
+6Juw 55904
+6Jux 55905
+6Juy 55906
+6Juz 55907
+6Ju0 55908
+6Ju4 55909
+6Ju5 55910
+6Ju7 55911
+6Ju8 55912
+6Ju9 55913
+6Ju+ 55914
+6JyA 55915
+6JyC 55916
+6JyD 55917
+6JyG 55918
+6JyH 55919
+6JyI 55920
+6JyJ 55921
+6JyK 55922
+6JyN 55923
+6JyO 55924
+6JyR 55925
+6JyS 55926
+6JyT 55927
+6JyV 55928
+6JyX 55929
+6JyY 55930
+6Jya 55931
+6Jyb 55932
+6Jyc 55933
+6Jye 55934
+6Jyh 55935
+6Jyi 55936
+6Jyj 55937
+6Jyl 55938
+6Jyp 55939
+6Jyu 55940
+6Jyx 55941
+6Jy0 55942
+6Jy3 55943
+6Jy7 55944
+6Jy+ 55945
+6Jy/ 55946
+6J2H 55947
+6J2I 55948
+6J2J 55949
+6J2L 55950
+6J2M 55951
+6J2O 55952
+6J2T 55953
+6J2V 55954
+6J2X 55955
+6J2Y 55956
+6J2Z 55957
+6J2f 55958
+6J2g 55959
+6J2j 55960
+6J2l 55961
+6J2m 55962
+6J2o 55963
+6J2u 55964
+6J2w 55965
+6J2y 55966
+6J20 55967
+6J22 55968
+6J24 55969
+6J27 55970
+6J28 55971
+6J29 55972
+6J2+ 55973
+6J2/ 55974
+6J6C 55975
+6J6D 55976
+6J6F 55977
+6J6I 55978
+6J6L 55979
+6J6N 55980
+6J6e 55981
+6J6f 55982
+6J6g 55983
+6J6i 55984
+6J6o 55985
+6J6r 55986
+6J6s 55987
+6J6t 55988
+6J6v 55989
+6J6z 55990
+6J61 55991
+6J66 55992
+6J67 55993
+6J69 55994
+6J+A 55995
+6J+E 55996
+6J+G 55997
+6J+H 55998
+6J+K 55999
+6J+L 56000
+6J+Q 56001
+6J+R 56002
+6J+S 56003
+6J+b 56004
+6J+c 56005
+6J+f 56006
+6J+g 56007
+6J+l 56008
+6J+q 56009
+6J+s 56010
+6J+t 56011
+6J+u 56012
+6J+v 56013
+6J+y 56014
+6J+2 56015
+6J+3 56016
+6J+5 56017
+6J+7 56018
+6J++ 56019
+6KCD 56020
+6KCF 56021
+6KCK 56022
+6KCL 56023
+6KCN 56024
+6KCO 56025
+6KCP 56026
+6KCR 56027
+6KCT 56028
+6KCV 56029
+6KCW 56030
+6KCb 56031
+6KCc 56032
+6KCf 56033
+6KCh 56034
+6KCi 56035
+6KCj 56036
+6KCn 56037
+6KCx 56038
+6KCy 56039
+6KC2 56040
+6KC5 56041
+6KC7 56042
+6KC8 56043
+6KGA 56044
+6KGE 56045
+6KGF 56046
+6KGG 56047
+6KGM 56048
+6KGN 56049
+6KGO 56050
+6KGS 56051
+6KGT 56052
+6KGU 56053
+6KGX 56054
+6KGZ 56055
+6KGb 56056
+6KGd 56057
+6KGe 56058
+6KGh 56059
+6KGi 56060
+6KGj 56061
+6KGl 56062
+6KGo 56063
+6KGp 56064
+6KGr 56065
+6KGs 56066
+6KGu 56067
+6KGv 56068
+6KGw 56069
+6KGy 56070
+6KG1 56071
+6KG3 56072
+6KG9 56073
+6KG+ 56074
+6KG/ 56075
+6KKB 56076
+6KKC 56077
+6KKE 56078
+6KKF 56079
+6KKG 56080
+6KKI 56081
+6KKL 56082
+6KKN 56083
+6KKS 56084
+6KKT 56085
+6KKW 56086
+6KKX 56087
+6KKZ 56088
+6KKc 56089
+6KKd 56090
+6KKe 56091
+6KKi 56092
+6KKk 56093
+6KKq 56094
+6KKr 56095
+6KKt 56096
+6KKw 56097
+6KKx 56098
+6KK0 56099
+6KK1 56100
+6KK3 56101
+6KK8 56102
+6KK/ 56103
+6KOB 56104
+6KOC 56105
+6KOD 56106
+6KOE 56107
+6KOF 56108
+6KOG 56109
+6KOH 56110
+6KOJ 56111
+6KOK 56112
+6KOO 56113
+6KOP 56114
+6KOS 56115
+6KOU 56116
+6KOV 56117
+6KOY 56118
+6KOZ 56119
+6KOb 56120
+6KOc 56121
+6KOd 56122
+6KOf 56123
+6KOh 56124
+6KOi 56125
+6KOk 56126
+6KOl 56127
+6KOo 56128
+6KOw 56129
+6KOx 56130
+6KOy 56131
+6KOz 56132
+6KO0 56133
+6KO4 56134
+6KO5 56135
+6KO9 56136
+6KO+ 56137
+6KSA 56138
+6KSC 56139
+6KSE 56140
+6KSH 56141
+6KSK 56142
+6KSM 56143
+6KSQ 56144
+6KSS 56145
+6KST 56146
+6KSU 56147
+6KSZ 56148
+6KSa 56149
+6KSb 56150
+6KSe 56151
+6KSh 56152
+6KSl 56153
+6KSq 56154
+6KSr 56155
+6KSt 56156
+6KSw 56157
+6KSy 56158
+6KS0 56159
+6KS2 56160
+6KS4 56161
+6KS7 56162
+6KWA 56163
+6KWB 56164
+6KWE 56165
+6KWW 56166
+6KWc 56167
+6KWe 56168
+6KWf 56169
+6KWg 56170
+6KWk 56171
+6KWm 56172
+6KWq 56173
+6KWv 56174
+6KWy 56175
+6KW0 56176
+6KW3 56177
+6KW7 56178
+6KW/ 56179
+6KaB 56180
+6KaD 56181
+6KaG 56182
+6KaH 56183
+6KaL 56184
+6KaP 56185
+6KaT 56186
+6KaW 56187
+6KaX 56188
+6KaY 56189
+6Kaa 56190
+6Kah 56191
+6Kan 56192
+6Kap 56193
+6Kaq 56194
+6Kav 56195
+6Kay 56196
+6Kaz 56197
+6Ka3 56198
+6Ka6 56199
+6Ka9 56200
+6Ka/ 56201
+6KeA 56202
+6KeB 56203
+6KeC 56204
+6KeE 56205
+6KeF 56206
+6KeG 56207
+6KeH 56208
+6KeI 56209
+6KeJ 56210
+6KeK 56211
+6KeM 56212
+6KeO 56213
+6KeP 56214
+6KeQ 56215
+6KeR 56216
+6KeS 56217
+6Kea 56218
+6Kec 56219
+6Ked 56220
+6Kee 56221
+6Kej 56222
+6Kel 56223
+6Kem 56224
+6Ker 56225
+6Kev 56226
+6Kex 56227
+6Kez 56228
+6Ke0 56229
+6Ke4 56230
+6KiA 56231
+6KiC 56232
+6KiD 56233
+6KiH 56234
+6KiI 56235
+6KiK 56236
+6KiM 56237
+6KiO 56238
+6KiT 56239
+6KiV 56240
+6KiX 56241
+6KiY 56242
+6Kia 56243
+6Kib 56244
+6Kid 56245
+6Kif 56246
+6Kii 56247
+6Kij 56248
+6Kil 56249
+6Kiq 56250
+6Kit 56251
+6Kix 56252
+6Kiz 56253
+6Ki0 56254
+6Ki2 56255
+6Ki6 56256
+6Ki7 56257
+6Ki8 56258
+6Ki+ 56259
+6KmB 56260
+6KmI 56261
+6KmQ 56262
+6KmS 56263
+6KmU 56264
+6KmV 56265
+6Kmb 56266
+6Kmd 56267
+6Kme 56268
+6Kmg 56269
+6Kmi 56270
+6Kmj 56271
+6Kmm 56272
+6Kmp 56273
+6Kmr 56274
+6Kms 56275
+6Kmt 56276
+6Kmu 56277
+6Kmw 56278
+6Kmx 56279
+6Kmy 56280
+6Kmz 56281
+6Km5 56282
+6KqC 56283
+6KqE 56284
+6KqF 56285
+6KqH 56286
+6KqJ 56287
+6KqK 56288
+6KqM 56289
+6KqN 56290
+6KqR 56291
+6KqS 56292
+6KqT 56293
+6KqV 56294
+6KqY 56295
+6Kqe 56296
+6Kqg 56297
+6Kqh 56298
+6Kqj 56299
+6Kqk 56300
+6Kql 56301
+6Kqm 56302
+6Kqo 56303
+6Kqq 56304
+6Kqs 56305
+6Kqt 56306
+6Kqw 56307
+6Kqy 56308
+6Kq5 56309
+6Kq8 56310
+6Kq/ 56311
+6KuC 56312
+6KuE 56313
+6KuH 56314
+6KuL 56315
+6KuM 56316
+6KuN 56317
+6KuP 56318
+6KuS 56319
+6KuW 56320
+6KuX 56321
+6Kua 56322
+6Kub 56323
+6Kuc 56324
+6Kuh 56325
+6Kui 56326
+6Kuk 56327
+6Kum 56328
+6Kun 56329
+6Kur 56330
+6Kut 56331
+6Kuu 56332
+6Kux 56333
+6Kuz 56334
+6Ku3 56335
+6Ku4 56336
+6Ku6 56337
+6Ku8 56338
+6Ku+ 56339
+6KyA 56340
+6KyB 56341
+6KyC 56342
+6KyE 56343
+6KyH 56344
+6KyK 56345
+6KyO 56346
+6KyQ 56347
+6KyU 56348
+6KyW 56349
+6KyX 56350
+6KyZ 56351
+6Kya 56352
+6Kyb 56353
+6Kyd 56354
+6Kyg 56355
+6Kyh 56356
+6Kym 56357
+6Kyo 56358
+6Kyp 56359
+6Kyr 56360
+6Kys 56361
+6Kyz 56362
+6Ky3 56363
+6Ky5 56364
+6K2B 56365
+6K2J 56366
+6K2O 56367
+6K2P 56368
+6K2W 56369
+6K2Y 56370
+6K2a 56371
+6K2c 56372
+6K2e 56373
+6K2f 56374
+6K2m 56375
+6K2r 56376
+6K2s 56377
+6K2v 56378
+6K2w 56379
+6K2y 56380
+6K20 56381
+6K23 56382
+6K29 56383
+6K6A 56384
+6K6D 56385
+6K6K 56386
+6K6M 56387
+6K6O 56388
+6K6Q 56389
+6K6S 56390
+6K6T 56391
+6K6W 56392
+6K6a 56393
+6K6g 56394
+6K6h 56395
+6K6i 56396
+6K6j 56397
+6K6k 56398
+6K6l 56399
+6K6m 56400
+6K6n 56401
+6K6o 56402
+6K6p 56403
+6K6q 56404
+6K6r 56405
+6K6t 56406
+6K6u 56407
+6K6v 56408
+6K6w 56409
+6K6y 56410
+6K6z 56411
+6K60 56412
+6K61 56413
+6K62 56414
+6K63 56415
+6K64 56416
+6K65 56417
+6K66 56418
+6K68 56419
+6K69 56420
+6K6+ 56421
+6K6/ 56422
+6K+A 56423
+6K+B 56424
+6K+C 56425
+6K+D 56426
+6K+E 56427
+6K+F 56428
+6K+G 56429
+6K+I 56430
+6K+J 56431
+6K+K 56432
+6K+L 56433
+6K+M 56434
+6K+N 56435
+6K+O 56436
+6K+P 56437
+6K+Q 56438
+6K+R 56439
+6K+S 56440
+6K+T 56441
+6K+U 56442
+6K+V 56443
+6K+W 56444
+6K+X 56445
+6K+Y 56446
+6K+Z 56447
+6K+a 56448
+6K+b 56449
+6K+c 56450
+6K+d 56451
+6K+e 56452
+6K+f 56453
+6K+g 56454
+6K+h 56455
+6K+i 56456
+6K+j 56457
+6K+k 56458
+6K+l 56459
+6K+m 56460
+6K+n 56461
+6K+o 56462
+6K+p 56463
+6K+r 56464
+6K+s 56465
+6K+t 56466
+6K+u 56467
+6K+v 56468
+6K+w 56469
+6K+x 56470
+6K+y 56471
+6K+z 56472
+6K+0 56473
+6K+1 56474
+6K+2 56475
+6K+3 56476
+6K+4 56477
+6K+5 56478
+6K+6 56479
+6K+7 56480
+6K+8 56481
+6K+9 56482
+6K++ 56483
+6K+/ 56484
+6LCA 56485
+6LCB 56486
+6LCC 56487
+6LCD 56488
+6LCE 56489
+6LCF 56490
+6LCG 56491
+6LCI 56492
+6LCK 56493
+6LCL 56494
+6LCM 56495
+6LCN 56496
+6LCO 56497
+6LCP 56498
+6LCQ 56499
+6LCR 56500
+6LCS 56501
+6LCT 56502
+6LCU 56503
+6LCV 56504
+6LCW 56505
+6LCX 56506
+6LCY 56507
+6LCZ 56508
+6LCa 56509
+6LCb 56510
+6LCc 56511
+6LCd 56512
+6LCe 56513
+6LCf 56514
+6LCg 56515
+6LCh 56516
+6LCi 56517
+6LCj 56518
+6LCk 56519
+6LCl 56520
+6LCm 56521
+6LCn 56522
+6LCo 56523
+6LCp 56524
+6LCq 56525
+6LCs 56526
+6LCt 56527
+6LCu 56528
+6LCv 56529
+6LCw 56530
+6LCx 56531
+6LCy 56532
+6LCz 56533
+6LC0 56534
+6LC1 56535
+6LC2 56536
+6LC3 56537
+6LC6 56538
+6LC/ 56539
+6LGB 56540
+6LGF 56541
+6LGG 56542
+6LGH 56543
+6LGI 56544
+6LGJ 56545
+6LGK 56546
+6LGM 56547
+6LGO 56548
+6LGQ 56549
+6LGU 56550
+6LGV 56551
+6LGa 56552
+6LGb 56553
+6LGd 56554
+6LGh 56555
+6LGi 56556
+6LGo 56557
+6LGq 56558
+6LGr 56559
+6LGs 56560
+6LGz 56561
+6LG4 56562
+6LG5 56563
+6LG6 56564
+6LKC 56565
+6LKF 56566
+6LKJ 56567
+6LKK 56568
+6LKM 56569
+6LKT 56570
+6LKU 56571
+6LKY 56572
+6LKd 56573
+6LKe 56574
+6LKg 56575
+6LKh 56576
+6LKi 56577
+6LKn 56578
+6LKo 56579
+6LKp 56580
+6LKq 56581
+6LKr 56582
+6LKs 56583
+6LKv 56584
+6LKw 56585
+6LKy 56586
+6LKz 56587
+6LK0 56588
+6LK2 56589
+6LK3 56590
+6LK4 56591
+6LK7 56592
+6LK8 56593
+6LK9 56594
+6LK/ 56595
+6LOA 56596
+6LOB 56597
+6LOC 56598
+6LOD 56599
+6LOE 56600
+6LOH 56601
+6LOI 56602
+6LOK 56603
+6LOO 56604
+6LOR 56605
+6LOS 56606
+6LOT 56607
+6LOa 56608
+6LOb 56609
+6LOc 56610
+6LOe 56611
+6LOg 56612
+6LOi 56613
+6LOj 56614
+6LOk 56615
+6LOm 56616
+6LOq 56617
+6LOs 56618
+6LOt 56619
+6LO0 56620
+6LO6 56621
+6LO8 56622
+6LO9 56623
+6LSE 56624
+6LSF 56625
+6LSH 56626
+6LSI 56627
+6LSK 56628
+6LSL 56629
+6LSP 56630
+6LSQ 56631
+6LST 56632
+6LSU 56633
+6LSW 56634
+6LSd 56635
+6LSe 56636
+6LSf 56637
+6LSh 56638
+6LSi 56639
+6LSj 56640
+6LSk 56641
+6LSl 56642
+6LSm 56643
+6LSn 56644
+6LSo 56645
+6LSp 56646
+6LSq 56647
+6LSr 56648
+6LSs 56649
+6LSt 56650
+6LSu 56651
+6LSv 56652
+6LSw 56653
+6LSx 56654
+6LSy 56655
+6LS0 56656
+6LS1 56657
+6LS2 56658
+6LS3 56659
+6LS4 56660
+6LS5 56661
+6LS6 56662
+6LS7 56663
+6LS8 56664
+6LS9 56665
+6LS+ 56666
+6LS/ 56667
+6LWB 56668
+6LWC 56669
+6LWD 56670
+6LWE 56671
+6LWF 56672
+6LWI 56673
+6LWJ 56674
+6LWK 56675
+6LWL 56676
+6LWM 56677
+6LWN 56678
+6LWO 56679
+6LWP 56680
+6LWQ 56681
+6LWT 56682
+6LWU 56683
+6LWV 56684
+6LWW 56685
+6LWY 56686
+6LWZ 56687
+6LWa 56688
+6LWb 56689
+6LWc 56690
+6LWd 56691
+6LWe 56692
+6LWf 56693
+6LWg 56694
+6LWh 56695
+6LWi 56696
+6LWj 56697
+6LWk 56698
+6LWm 56699
+6LWn 56700
+6LWq 56701
+6LWr 56702
+6LWs 56703
+6LWt 56704
+6LWw 56705
+6LWz 56706
+6LW0 56707
+6LW1 56708
+6LW2 56709
+6LW3 56710
+6LaB 56711
+6LaE 56712
+6LaF 56713
+6LaK 56714
+6LaL 56715
+6LaU 56716
+6LaV 56717
+6LaZ 56718
+6Laf 56719
+6Laj 56720
+6Lao 56721
+6Laz 56722
+6La0 56723
+6La1 56724
+6La4 56725
+6La6 56726
+6La+ 56727
+6La/ 56728
+6LeC 56729
+6LeD 56730
+6LeE 56731
+6LeG 56732
+6LeL 56733
+6LeM 56734
+6LeO 56735
+6LeP 56736
+6LeR 56737
+6LeW 56738
+6LeX 56739
+6Lea 56740
+6Leb 56741
+6Led 56742
+6Lee 56743
+6Lef 56744
+6Leh 56745
+6Lej 56746
+6Lek 56747
+6Leo 56748
+6Lep 56749
+6Leq 56750
+6Ler 56751
+6Les 56752
+6Lev 56753
+6Lex 56754
+6Lez 56755
+6Le1 56756
+6Le2 56757
+6Le3 56758
+6Le4 56759
+6Le5 56760
+6Le6 56761
+6Le7 56762
+6Le8 56763
+6Le9 56764
+6Le/ 56765
+6LiF 56766
+6LiJ 56767
+6LiK 56768
+6LiM 56769
+6LiO 56770
+6LiP 56771
+6LiQ 56772
+6LiU 56773
+6LiW 56774
+6Lid 56775
+6Lie 56776
+6Lif 56777
+6Lih 56778
+6Lii 56779
+6Lij 56780
+6Lim 56781
+6Lin 56782
+6Lip 56783
+6Liq 56784
+6Lis 56785
+6Lit 56786
+6Liu 56787
+6Liv 56788
+6Liw 56789
+6Lix 56790
+6Li0 56791
+6Li1 56792
+6Li5 56793
+6Li6 56794
+6Li9 56795
+6LmA 56796
+6LmB 56797
+6LmC 56798
+6LmE 56799
+6LmH 56800
+6LmI 56801
+6LmJ 56802
+6LmK 56803
+6LmL 56804
+6LmM 56805
+6LmQ 56806
+6LmR 56807
+6LmS 56808
+6LmV 56809
+6LmZ 56810
+6Lma 56811
+6Lmf 56812
+6Lmg 56813
+6Lmh 56814
+6Lmj 56815
+6Lmk 56816
+6Lmm 56817
+6Lmp 56818
+6Lms 56819
+6Lmt 56820
+6Lmv 56821
+6Lmw 56822
+6Lmy 56823
+6Lm0 56824
+6Lm2 56825
+6Lm8 56826
+6Lm9 56827
+6Lm/ 56828
+6LqB 56829
+6LqE 56830
+6LqF 56831
+6LqH 56832
+6LqK 56833
+6LqN 56834
+6LqP 56835
+6LqQ 56836
+6LqR 56837
+6LqT 56838
+6LqU 56839
+6LqZ 56840
+6Lqc 56841
+6Lqh 56842
+6Lqq 56843
+6Lqr 56844
+6Lqs 56845
+6Lqv 56846
+6Lqw 56847
+6Lqx 56848
+6Lqy 56849
+6Lq6 56850
+6Lq+ 56851
+6LuA 56852
+6LuI 56853
+6LuK 56854
+6LuL 56855
+6LuM 56856
+6LuN 56857
+6LuO 56858
+6LuS 56859
+6Lub 56860
+6Luf 56861
+6Lui 56862
+6Luj 56863
+6Lur 56864
+6Lu4 56865
+6Lu7 56866
+6Lu8 56867
+6Lu9 56868
+6LyD 56869
+6LyJ 56870
+6LyK 56871
+6LyM 56872
+6LyS 56873
+6LyT 56874
+6LyU 56875
+6LyV 56876
+6Lyb 56877
+6Lyc 56878
+6Lyd 56879
+6Lym 56880
+6Lyp 56881
+6Lyq 56882
+6Lyv 56883
+6Lyz 56884
+6Ly2 56885
+6Ly4 56886
+6Ly7 56887
+6Ly+ 56888
+6Ly/ 56889
+6L2C 56890
+6L2E 56891
+6L2F 56892
+6L2G 56893
+6L2J 56894
+6L2M 56895
+6L2N 56896
+6L2O 56897
+6L2X 56898
+6L2f 56899
+6L2h 56900
+6L2i 56901
+6L2j 56902
+6L2k 56903
+6L2m 56904
+6L2n 56905
+6L2o 56906
+6L2p 56907
+6L2r 56908
+6L2s 56909
+6L2t 56910
+6L2u 56911
+6L2v 56912
+6L2w 56913
+6L2x 56914
+6L2y 56915
+6L2z 56916
+6L20 56917
+6L21 56918
+6L22 56919
+6L24 56920
+6L25 56921
+6L26 56922
+6L27 56923
+6L28 56924
+6L29 56925
+6L2+ 56926
+6L2/ 56927
+6L6C 56928
+6L6D 56929
+6L6E 56930
+6L6F 56931
+6L6G 56932
+6L6H 56933
+6L6I 56934
+6L6J 56935
+6L6K 56936
+6L6L 56937
+6L6N 56938
+6L6O 56939
+6L6P 56940
+6L6Q 56941
+6L6R 56942
+6L6T 56943
+6L6U 56944
+6L6V 56945
+6L6W 56946
+6L6X 56947
+6L6Y 56948
+6L6Z 56949
+6L6a 56950
+6L6b 56951
+6L6c 56952
+6L6e 56953
+6L6f 56954
+6L6j 56955
+6L6m 56956
+6L6o 56957
+6L6p 56958
+6L6r 56959
+6L6t 56960
+6L6u 56961
+6L6v 56962
+6L6w 56963
+6L6x 56964
+6L6y 56965
+6L63 56966
+6L65 56967
+6L66 56968
+6L67 56969
+6L68 56970
+6L69 56971
+6L6+ 56972
+6L6/ 56973
+6L+B 56974
+6L+C 56975
+6L+E 56976
+6L+F 56977
+6L+H 56978
+6L+I 56979
+6L+O 56980
+6L+Q 56981
+6L+R 56982
+6L+T 56983
+6L+U 56984
+6L+V 56985
+6L+Y 56986
+6L+Z 56987
+6L+a 56988
+6L+b 56989
+6L+c 56990
+6L+d 56991
+6L+e 56992
+6L+f 56993
+6L+g 56994
+6L+i 56995
+6L+k 56996
+6L+l 56997
+6L+m 56998
+6L+o 56999
+6L+p 57000
+6L+q 57001
+6L+r 57002
+6L+t 57003
+6L+u 57004
+6L+w 57005
+6L+z 57006
+6L+0 57007
+6L+3 57008
+6L+4 57009
+6L+5 57010
+6L+9 57011
+6YCA 57012
+6YCB 57013
+6YCC 57014
+6YCD 57015
+6YCE 57016
+6YCF 57017
+6YCG 57018
+6YCJ 57019
+6YCK 57020
+6YCL 57021
+6YCN 57022
+6YCP 57023
+6YCQ 57024
+6YCR 57025
+6YCS 57026
+6YCT 57027
+6YCU 57028
+6YCV 57029
+6YCW 57030
+6YCX 57031
+6YCZ 57032
+6YCa 57033
+6YCb 57034
+6YCd 57035
+6YCe 57036
+6YCf 57037
+6YCg 57038
+6YCh 57039
+6YCi 57040
+6YCj 57041
+6YCm 57042
+6YCu 57043
+6YCv 57044
+6YCx 57045
+6YCy 57046
+6YC1 57047
+6YC2 57048
+6YC4 57049
+6YC7 57050
+6YC8 57051
+6YC+ 57052
+6YGB 57053
+6YGC 57054
+6YGE 57055
+6YGF 57056
+6YGH 57057
+6YGJ 57058
+6YGK 57059
+6YGL 57060
+6YGN 57061
+6YGO 57062
+6YGP 57063
+6YGQ 57064
+6YGR 57065
+6YGS 57066
+6YGT 57067
+6YGU 57068
+6YGV 57069
+6YGW 57070
+6YGX 57071
+6YGY 57072
+6YGZ 57073
+6YGb 57074
+6YGc 57075
+6YGe 57076
+6YGg 57077
+6YGh 57078
+6YGi 57079
+6YGj 57080
+6YGl 57081
+6YGo 57082
+6YGp 57083
+6YGt 57084
+6YGu 57085
+6YGv 57086
+6YGy 57087
+6YG0 57088
+6YG1 57089
+6YG3 57090
+6YG4 57091
+6YG5 57092
+6YG6 57093
+6YG8 57094
+6YG9 57095
+6YG/ 57096
+6YKA 57097
+6YKB 57098
+6YKC 57099
+6YKD 57100
+6YKE 57101
+6YKF 57102
+6YKH 57103
+6YKI 57104
+6YKJ 57105
+6YKK 57106
+6YKL 57107
+6YKP 57108
+6YKR 57109
+6YKT 57110
+6YKV 57111
+6YKX 57112
+6YKY 57113
+6YKZ 57114
+6YKb 57115
+6YKd 57116
+6YKg 57117
+6YKh 57118
+6YKi 57119
+6YKj 57120
+6YKm 57121
+6YKo 57122
+6YKq 57123
+6YKs 57124
+6YKu 57125
+6YKv 57126
+6YKw 57127
+6YKx 57128
+6YKz 57129
+6YK0 57130
+6YK1 57131
+6YK2 57132
+6YK4 57133
+6YK5 57134
+6YK6 57135
+6YK7 57136
+6YK9 57137
+6YK+ 57138
+6YOB 57139
+6YOD 57140
+6YOE 57141
+6YOF 57142
+6YOH 57143
+6YOK 57144
+6YOO 57145
+6YOP 57146
+6YOQ 57147
+6YOR 57148
+6YOT 57149
+6YOV 57150
+6YOb 57151
+6YOc 57152
+6YOd 57153
+6YOe 57154
+6YOh 57155
+6YOi 57156
+6YOk 57157
+6YOm 57158
+6YOn 57159
+6YOo 57160
+6YOr 57161
+6YOt 57162
+6YOv 57163
+6YO0 57164
+6YO1 57165
+6YO3 57166
+6YO4 57167
+6YO9 57168
+6YO+ 57169
+6YO/ 57170
+6YSC 57171
+6YSE 57172
+6YSF 57173
+6YSJ 57174
+6YSP 57175
+6YSS 57176
+6YSY 57177
+6YSZ 57178
+6YSa 57179
+6YSc 57180
+6YSe 57181
+6YSg 57182
+6YSi 57183
+6YSj 57184
+6YSn 57185
+6YSp 57186
+6YSr 57187
+6YSt 57188
+6YSu 57189
+6YSv 57190
+6YSw 57191
+6YSx 57192
+6YSy 57193
+6YS6 57194
+6YWG 57195
+6YWJ 57196
+6YWK 57197
+6YWL 57198
+6YWM 57199
+6YWN 57200
+6YWO 57201
+6YWQ 57202
+6YWS 57203
+6YWU 57204
+6YWX 57205
+6YWY 57206
+6YWa 57207
+6YWd 57208
+6YWe 57209
+6YWh 57210
+6YWi 57211
+6YWj 57212
+6YWk 57213
+6YWl 57214
+6YWp 57215
+6YWq 57216
+6YWr 57217
+6YWs 57218
+6YWu 57219
+6YWv 57220
+6YWw 57221
+6YWx 57222
+6YW1 57223
+6YW2 57224
+6YW3 57225
+6YW4 57226
+6YW5 57227
+6YW6 57228
+6YW9 57229
+6YW/ 57230
+6YaC 57231
+6YaF 57232
+6YaH 57233
+6YaJ 57234
+6YaL 57235
+6YaM 57236
+6YaN 57237
+6YaQ 57238
+6YaS 57239
+6YaX 57240
+6Yaa 57241
+6Yab 57242
+6Yac 57243
+6Yae 57244
+6Yai 57245
+6Yaj 57246
+6Yak 57247
+6Yaq 57248
+6Yar 57249
+6Yas 57250
+6Yat 57251
+6Yau 57252
+6Yav 57253
+6Yay 57254
+6Ya0 57255
+6Ya1 57256
+6Ya4 57257
+6Ya6 57258
+6YeA 57259
+6YeB 57260
+6YeG 57261
+6YeH 57262
+6YeI 57263
+6YeJ 57264
+6YeK 57265
+6YeL 57266
+6YeM 57267
+6YeN 57268
+6YeO 57269
+6YeP 57270
+6YeQ 57271
+6YeR 57272
+6YeY 57273
+6Yeb 57274
+6Yec 57275
+6Yed 57276
+6Yeh 57277
+6Yej 57278
+6Yem 57279
+6Yen 57280
+6Yet 57281
+6Ye1 57282
+6Ye2 57283
+6Ye8 57284
+6Ye/ 57285
+6YiH 57286
+6YiN 57287
+6YiO 57288
+6YiR 57289
+6YiU 57290
+6YiV 57291
+6Yie 57292
+6Yip 57293
+6Yiq 57294
+6Yiz 57295
+6Yi0 57296
+6Yi3 57297
+6Yi/ 57298
+6YmE 57299
+6YmI 57300
+6YmJ 57301
+6YmL 57302
+6YmP 57303
+6YmX 57304
+6Ymb 57305
+6Yme 57306
+6Ymi 57307
+6Ymk 57308
+6Ymm 57309
+6Ymx 57310
+6Ym0 57311
+6Ym+ 57312
+6YqA 57313
+6YqD 57314
+6YqF 57315
+6YqR 57316
+6YqT 57317
+6YqV 57318
+6YqW 57319
+6YqY 57320
+6Yqa 57321
+6Yqb 57322
+6Yqc 57323
+6Yqt 57324
+6Yqu 57325
+6Yqz 57326
+6Yq2 57327
+6Yq3 57328
+6Yq5 57329
+6YuG 57330
+6YuI 57331
+6YuP 57332
+6YuQ 57333
+6YuS 57334
+6Yuk 57335
+6Yup 57336
+6Yuq 57337
+6Yut 57338
+6Yuy 57339
+6Yuz 57340
+6Yu4 57341
+6Yu5 57342
+6Yu6 57343
+6Yu8 57344
+6YyE 57345
+6YyG 57346
+6YyP 57347
+6YyQ 57348
+6YyY 57349
+6YyZ 57350
+6Yya 57351
+6Yye 57352
+6Yyg 57353
+6Yyi 57354
+6Yyj 57355
+6Yym 57356
+6Yyo 57357
+6Yyr 57358
+6Yys 57359
+6Yyu 57360
+6Yyv 57361
+6Yyy 57362
+6Yy1 57363
+6Yy2 57364
+6Yy6 57365
+6Yy+ 57366
+6Y2J 57367
+6Y2K 57368
+6Y2L 57369
+6Y2N 57370
+6Y2S 57371
+6Y2U 57372
+6Y2b 57373
+6Y2c 57374
+6Y2l 57375
+6Y2q 57376
+6Y2s 57377
+6Y2u 57378
+6Y21 57379
+6Y28 57380
+6Y2+ 57381
+6Y6J 57382
+6Y6K 57383
+6Y6M 57384
+6Y6P 57385
+6Y6U 57386
+6Y6W 57387
+6Y6X 57388
+6Y6a 57389
+6Y6b 57390
+6Y6n 57391
+6Y6s 57392
+6Y6t 57393
+6Y6u 57394
+6Y6w 57395
+6Y64 57396
+6Y65 57397
+6Y+D 57398
+6Y+I 57399
+6Y+K 57400
+6Y+Q 57401
+6Y+R 57402
+6Y+W 57403
+6Y+X 57404
+6Y+Y 57405
+6Y+d 57406
+6Y+f 57407
+6Y+h 57408
+6Y+k 57409
+6Y+l 57410
+6Y+o 57411
+6ZCD 57412
+6ZCH 57413
+6ZCO 57414
+6ZCQ 57415
+6ZCU 57416
+6ZCY 57417
+6ZCZ 57418
+6ZCa 57419
+6ZCh 57420
+6ZCn 57421
+6ZCu 57422
+6ZCy 57423
+6ZC1 57424
+6ZC2 57425
+6ZC4 57426
+6ZC6 57427
+6ZGB 57428
+6ZGE 57429
+6ZGK 57430
+6ZGR 57431
+6ZGS 57432
+6ZGT 57433
+6ZGb 57434
+6ZGe 57435
+6ZGg 57436
+6ZGi 57437
+6ZGq 57438
+6ZGr 57439
+6ZGw 57440
+6ZGx 57441
+6ZGy 57442
+6ZG1 57443
+6ZG3 57444
+6ZG5 57445
+6ZG8 57446
+6ZG9 57447
+6ZG/ 57448
+6ZKH 57449
+6ZKI 57450
+6ZKJ 57451
+6ZKK 57452
+6ZKL 57453
+6ZKM 57454
+6ZKN 57455
+6ZKO 57456
+6ZKP 57457
+6ZKQ 57458
+6ZKS 57459
+6ZKT 57460
+6ZKU 57461
+6ZKV 57462
+6ZKX 57463
+6ZKY 57464
+6ZKZ 57465
+6ZKa 57466
+6ZKb 57467
+6ZKc 57468
+6ZKd 57469
+6ZKe 57470
+6ZKf 57471
+6ZKg 57472
+6ZKh 57473
+6ZKi 57474
+6ZKj 57475
+6ZKk 57476
+6ZKl 57477
+6ZKm 57478
+6ZKn 57479
+6ZKo 57480
+6ZKp 57481
+6ZKq 57482
+6ZKr 57483
+6ZKs 57484
+6ZKu 57485
+6ZKv 57486
+6ZKw 57487
+6ZKx 57488
+6ZKy 57489
+6ZKz 57490
+6ZK0 57491
+6ZK1 57492
+6ZK5 57493
+6ZK6 57494
+6ZK7 57495
+6ZK8 57496
+6ZK9 57497
+6ZK+ 57498
+6ZK/ 57499
+6ZOA 57500
+6ZOB 57501
+6ZOC 57502
+6ZOD 57503
+6ZOE 57504
+6ZOF 57505
+6ZOG 57506
+6ZOI 57507
+6ZOJ 57508
+6ZOK 57509
+6ZOL 57510
+6ZOM 57511
+6ZON 57512
+6ZOO 57513
+6ZOQ 57514
+6ZOR 57515
+6ZOS 57516
+6ZOT 57517
+6ZOW 57518
+6ZOX 57519
+6ZOZ 57520
+6ZOa 57521
+6ZOb 57522
+6ZOc 57523
+6ZOd 57524
+6ZOe 57525
+6ZOf 57526
+6ZOg 57527
+6ZOh 57528
+6ZOi 57529
+6ZOj 57530
+6ZOk 57531
+6ZOl 57532
+6ZOm 57533
+6ZOn 57534
+6ZOo 57535
+6ZOp 57536
+6ZOq 57537
+6ZOr 57538
+6ZOs 57539
+6ZOt 57540
+6ZOu 57541
+6ZOv 57542
+6ZOw 57543
+6ZOx 57544
+6ZOy 57545
+6ZOz 57546
+6ZO1 57547
+6ZO2 57548
+6ZO3 57549
+6ZO4 57550
+6ZO6 57551
+6ZO8 57552
+6ZO+ 57553
+6ZO/ 57554
+6ZSA 57555
+6ZSB 57556
+6ZSC 57557
+6ZSD 57558
+6ZSE 57559
+6ZSF 57560
+6ZSG 57561
+6ZSH 57562
+6ZSI 57563
+6ZSJ 57564
+6ZSK 57565
+6ZSL 57566
+6ZSM 57567
+6ZSN 57568
+6ZSP 57569
+6ZSQ 57570
+6ZSR 57571
+6ZSS 57572
+6ZSU 57573
+6ZSV 57574
+6ZSW 57575
+6ZSX 57576
+6ZSY 57577
+6ZSZ 57578
+6ZSa 57579
+6ZSb 57580
+6ZSc 57581
+6ZSd 57582
+6ZSe 57583
+6ZSf 57584
+6ZSh 57585
+6ZSi 57586
+6ZSj 57587
+6ZSk 57588
+6ZSl 57589
+6ZSm 57590
+6ZSo 57591
+6ZSp 57592
+6ZSr 57593
+6ZSs 57594
+6ZSt 57595
+6ZSu 57596
+6ZSv 57597
+6ZSw 57598
+6ZSx 57599
+6ZSy 57600
+6ZS0 57601
+6ZS1 57602
+6ZS2 57603
+6ZS3 57604
+6ZS4 57605
+6ZS5 57606
+6ZS6 57607
+6ZS7 57608
+6ZWA 57609
+6ZWB 57610
+6ZWC 57611
+6ZWF 57612
+6ZWG 57613
+6ZWH 57614
+6ZWJ 57615
+6ZWK 57616
+6ZWL 57617
+6ZWM 57618
+6ZWN 57619
+6ZWP 57620
+6ZWQ 57621
+6ZWR 57622
+6ZWS 57623
+6ZWU 57624
+6ZWV 57625
+6ZWW 57626
+6ZWX 57627
+6ZWY 57628
+6ZWb 57629
+6ZWc 57630
+6ZWd 57631
+6ZWe 57632
+6ZWf 57633
+6ZWg 57634
+6ZWh 57635
+6ZWi 57636
+6ZWj 57637
+6ZWl 57638
+6ZWn 57639
+6ZWp 57640
+6ZWq 57641
+6ZWr 57642
+6ZWs 57643
+6ZWt 57644
+6ZWv 57645
+6ZWw 57646
+6ZWx 57647
+6ZWy 57648
+6ZWz 57649
+6ZW1 57650
+6ZW2 57651
+6ZW3 57652
+6ZW/ 57653
+6ZaA 57654
+6ZaC 57655
+6ZaD 57656
+6ZaH 57657
+6ZaJ 57658
+6ZaK 57659
+6ZaL 57660
+6ZaO 57661
+6ZaP 57662
+6ZaR 57663
+6ZaS 57664
+6ZaT 57665
+6ZaU 57666
+6ZaW 57667
+6ZaY 57668
+6Zaf 57669
+6Zai 57670
+6Zaj 57671
+6Zak 57672
+6Zal 57673
+6Zam 57674
+6Zao 57675
+6Zat 57676
+6Zax 57677
+6Zay 57678
+6Za5 57679
+6Za7 57680
+6Za8 57681
+6Za+ 57682
+6ZeD 57683
+6ZeG 57684
+6ZeH 57685
+6ZeK 57686
+6ZeM 57687
+6ZeN 57688
+6ZeV 57689
+6ZeW 57690
+6ZeY 57691
+6Zec 57692
+6Zeh 57693
+6Zei 57694
+6Zel 57695
+6Zeo 57696
+6Zep 57697
+6Zeq 57698
+6Zer 57699
+6Zet 57700
+6Zeu 57701
+6Zev 57702
+6Zew 57703
+6Zex 57704
+6Zey 57705
+6Zez 57706
+6Ze0 57707
+6Ze1 57708
+6Ze3 57709
+6Ze4 57710
+6Ze5 57711
+6Ze6 57712
+6Ze7 57713
+6Ze8 57714
+6Ze9 57715
+6Ze+ 57716
+6Ze/ 57717
+6ZiA 57718
+6ZiB 57719
+6ZiC 57720
+6ZiD 57721
+6ZiE 57722
+6ZiF 57723
+6ZiG 57724
+6ZiH 57725
+6ZiI 57726
+6ZiJ 57727
+6ZiK 57728
+6ZiL 57729
+6ZiM 57730
+6ZiN 57731
+6ZiO 57732
+6ZiP 57733
+6ZiQ 57734
+6ZiR 57735
+6ZiS 57736
+6ZiU 57737
+6ZiV 57738
+6ZiW 57739
+6ZiX 57740
+6ZiZ 57741
+6Zia 57742
+6Zic 57743
+6Zid 57744
+6Zif 57745
+6Zih 57746
+6Zio 57747
+6Ziq 57748
+6Zis 57749
+6Ziu 57750
+6Zix 57751
+6Ziy 57752
+6Ziz 57753
+6Zi0 57754
+6Zi1 57755
+6Zi2 57756
+6Zi7 57757
+6Zi8 57758
+6Zi9 57759
+6Zi/ 57760
+6ZmA 57761
+6ZmC 57762
+6ZmE 57763
+6ZmF 57764
+6ZmG 57765
+6ZmH 57766
+6ZmI 57767
+6ZmJ 57768
+6ZmL 57769
+6ZmM 57770
+6ZmN 57771
+6ZmQ 57772
+6ZmU 57773
+6ZmV 57774
+6Zmb 57775
+6Zmc 57776
+6Zmd 57777
+6Zme 57778
+6Zmf 57779
+6Zmh 57780
+6Zmi 57781
+6Zmj 57782
+6Zmk 57783
+6Zml 57784
+6Zmm 57785
+6Zmn 57786
+6Zmo 57787
+6Zmp 57788
+6Zmq 57789
+6Zms 57790
+6Zmw 57791
+6Zmy 57792
+6Zmz 57793
+6Zm1 57794
+6Zm2 57795
+6Zm3 57796
+6Zm4 57797
+6Zm6 57798
+6Zm9 57799
+6ZqF 57800
+6ZqG 57801
+6ZqI 57802
+6ZqK 57803
+6ZqL 57804
+6ZqN 57805
+6ZqO 57806
+6ZqP 57807
+6ZqQ 57808
+6ZqU 57809
+6ZqV 57810
+6ZqX 57811
+6ZqY 57812
+6ZqZ 57813
+6Zqb 57814
+6Zqc 57815
+6Zqg 57816
+6Zqj 57817
+6Zqn 57818
+6Zqo 57819
+6Zqq 57820
+6Zqw 57821
+6Zqx 57822
+6Zqy 57823
+6Zqz 57824
+6Zq2 57825
+6Zq3 57826
+6Zq4 57827
+6Zq5 57828
+6Zq7 57829
+6Zq8 57830
+6Zq9 57831
+6Zq+ 57832
+6ZuA 57833
+6ZuB 57834
+6ZuE 57835
+6ZuF 57836
+6ZuG 57837
+6ZuH 57838
+6ZuJ 57839
+6ZuL 57840
+6ZuM 57841
+6ZuN 57842
+6ZuO 57843
+6ZuP 57844
+6ZuR 57845
+6ZuS 57846
+6ZuV 57847
+6ZuW 57848
+6ZuZ 57849
+6Zub 57850
+6Zuc 57851
+6Zue 57852
+6Zug 57853
+6Zui 57854
+6Zuj 57855
+6Zuo 57856
+6Zup 57857
+6Zuq 57858
+6Zur 57859
+6Zuv 57860
+6Zuw 57861
+6Zux 57862
+6Zuy 57863
+6Zuz 57864
+6Zu2 57865
+6Zu3 57866
+6Zu5 57867
+6Zu7 57868
+6Zu+ 57869
+6ZyA 57870
+6ZyB 57871
+6ZyE 57872
+6ZyG 57873
+6ZyH 57874
+6ZyI 57875
+6ZyJ 57876
+6ZyK 57877
+6ZyN 57878
+6ZyO 57879
+6ZyP 57880
+6ZyR 57881
+6ZyT 57882
+6ZyW 57883
+6ZyZ 57884
+6Zyc 57885
+6Zye 57886
+6Zyn 57887
+6Zyq 57888
+6Zyr 57889
+6Zyt 57890
+6Zyw 57891
+6Zyy 57892
+6Zy4 57893
+6Zy5 57894
+6Zy9 57895
+6Zy+ 57896
+6Z2C 57897
+6Z2E 57898
+6Z2G 57899
+6Z2I 57900
+6Z2J 57901
+6Z2R 57902
+6Z2S 57903
+6Z2T 57904
+6Z2W 57905
+6Z2Z 57906
+6Z2a 57907
+6Z2b 57908
+6Z2c 57909
+6Z2e 57910
+6Z2g 57911
+6Z2h 57912
+6Z2i 57913
+6Z2l 57914
+6Z2m 57915
+6Z2o 57916
+6Z2p 57917
+6Z2r 57918
+6Z2t 57919
+6Z2x 57920
+6Z2z 57921
+6Z20 57922
+6Z22 57923
+6Z24 57924
+6Z25 57925
+6Z26 57926
+6Z28 57927
+6Z6B 57928
+6Z6E 57929
+6Z6F 57930
+6Z6G 57931
+6Z6L 57932
+6Z6N 57933
+6Z6P 57934
+6Z6Q 57935
+6Z6R 57936
+6Z6Y 57937
+6Z6c 57938
+6Z6g 57939
+6Z6j 57940
+6Z6l 57941
+6Z6m 57942
+6Z6n 57943
+6Z6o 57944
+6Z6r 57945
+6Z6s 57946
+6Z6t 57947
+6Z60 57948
+6Z62 57949
+6Z+C 57950
+6Z+D 57951
+6Z+L 57952
+6Z+M 57953
+6Z+T 57954
+6Z+Y 57955
+6Z+c 57956
+6Z+h 57957
+6Z+m 57958
+6Z+n 57959
+6Z+p 57960
+6Z+q 57961
+6Z+r 57962
+6Z+s 57963
+6Z+t 57964
+6Z+u 57965
+6Z+y 57966
+6Z+z 57967
+6Z+1 57968
+6Z+2 57969
+6Z+7 57970
+6Z+/ 57971
+6aCB 57972
+6aCC 57973
+6aCD 57974
+6aCF 57975
+6aCG 57976
+6aCI 57977
+6aCM 57978
+6aCP 57979
+6aCQ 57980
+6aCR 57981
+6aCS 57982
+6aCT 57983
+6aCX 57984
+6aCY 57985
+6aCa 57986
+6aCg 57987
+6aCh 57988
+6aCk 57989
+6aCr 57990
+6aCs 57991
+6aCt 57992
+6aCw 57993
+6aC0 57994
+6aC3 57995
+6aC4 57996
+6aC5 57997
+6aC7 57998
+6aC8 57999
+6aC9 58000
+6aGG 58001
+6aGL 58002
+6aGM 58003
+6aGN 58004
+6aGO 58005
+6aGP 58006
+6aGS 58007
+6aGU 58008
+6aGV 58009
+6aGX 58010
+6aGY 58011
+6aGb 58012
+6aGe 58013
+6aGn 58014
+6aGr 58015
+6aGs 58016
+6aGv 58017
+6aGw 58018
+6aGx 58019
+6aGz 58020
+6aG0 58021
+6aG1 58022
+6aG2 58023
+6aG3 58024
+6aG4 58025
+6aG5 58026
+6aG6 58027
+6aG7 58028
+6aG8 58029
+6aG9 58030
+6aG+ 58031
+6aG/ 58032
+6aKA 58033
+6aKB 58034
+6aKC 58035
+6aKE 58036
+6aKF 58037
+6aKG 58038
+6aKH 58039
+6aKI 58040
+6aKJ 58041
+6aKK 58042
+6aKL 58043
+6aKM 58044
+6aKN 58045
+6aKO 58046
+6aKP 58047
+6aKQ 58048
+6aKR 58049
+6aKT 58050
+6aKU 58051
+6aKW 58052
+6aKX 58053
+6aKY 58054
+6aKZ 58055
+6aKa 58056
+6aKb 58057
+6aKc 58058
+6aKd 58059
+6aKe 58060
+6aKf 58061
+6aKg 58062
+6aKh 58063
+6aKi 58064
+6aKk 58065
+6aKm 58066
+6aKn 58067
+6aKo 58068
+6aKq 58069
+6aKv 58070
+6aKx 58071
+6aK2 58072
+6aOE 58073
+6aOG 58074
+6aOK 58075
+6aOO 58076
+6aOQ 58077
+6aOS 58078
+6aOT 58079
+6aOV 58080
+6aOW 58081
+6aOX 58082
+6aOY 58083
+6aOZ 58084
+6aOa 58085
+6aOb 58086
+6aOe 58087
+6aOf 58088
+6aOi 58089
+6aOn 58090
+6aOo 58091
+6aOp 58092
+6aOq 58093
+6aOr 58094
+6aOt 58095
+6aOv 58096
+6aOy 58097
+6aO0 58098
+6aO8 58099
+6aO9 58100
+6aO+ 58101
+6aSD 58102
+6aSF 58103
+6aSJ 58104
+6aSK 58105
+6aSM 58106
+6aSN 58107
+6aSQ 58108
+6aSR 58109
+6aSS 58110
+6aST 58111
+6aSY 58112
+6aSa 58113
+6aSd 58114
+6aSe 58115
+6aSg 58116
+6aSh 58117
+6aSo 58118
+6aSs 58119
+6aSu 58120
+6aS1 58121
+6aS4 58122
+6aWC 58123
+6aWF 58124
+6aWJ 58125
+6aWL 58126
+6aWM 58127
+6aWQ 58128
+6aWR 58129
+6aWS 58130
+6aWU 58131
+6aWV 58132
+6aWX 58133
+6aWk 58134
+6aWl 58135
+6aWn 58136
+6aWo 58137
+6aWq 58138
+6aWs 58139
+6aWt 58140
+6aWu 58141
+6aWv 58142
+6aWw 58143
+6aWx 58144
+6aWy 58145
+6aW0 58146
+6aW1 58147
+6aW2 58148
+6aW3 58149
+6aW4 58150
+6aW5 58151
+6aW6 58152
+6aW8 58153
+6aW9 58154
+6aW/ 58155
+6aaA 58156
+6aaB 58157
+6aaD 58158
+6aaE 58159
+6aaF 58160
+6aaG 58161
+6aaH 58162
+6aaI 58163
+6aaK 58164
+6aaL 58165
+6aaN 58166
+6aaP 58167
+6aaQ 58168
+6aaR 58169
+6aaS 58170
+6aaT 58171
+6aaU 58172
+6aaV 58173
+6aaW 58174
+6aaX 58175
+6aaY 58176
+6aaZ 58177
+6aal 58178
+6aao 58179
+6aas 58180
+6aat 58181
+6aau 58182
+6aaz 58183
+6aa0 58184
+6aa/ 58185
+6aeB 58186
+6aeD 58187
+6aeE 58188
+6aeF 58189
+6aeG 58190
+6aeI 58191
+6aeQ 58192
+6aeR 58193
+6aeS 58194
+6aeV 58195
+6aeY 58196
+6aeb 58197
+6aed 58198
+6aef 58199
+6aei 58200
+6aet 58201
+6aeu 58202
+6aex 58203
+6ae4 58204
+6ae7 58205
+6ae/ 58206
+6aiB 58207
+6aiO 58208
+6aiP 58209
+6aiR 58210
+6aiS 58211
+6aiT 58212
+6aiV 58213
+6aiW 58214
+6aiZ 58215
+6aig 58216
+6aio 58217
+6aiu 58218
+6aiw 58219
+6ai3 58220
+6ai+ 58221
+6amA 58222
+6amD 58223
+6amF 58224
+6amN 58225
+6amV 58226
+6amX 58227
+6ama 58228
+6amb 58229
+6amf 58230
+6ami 58231
+6aml 58232
+6amp 58233
+6amq 58234
+6amr 58235
+6ams 58236
+6amt 58237
+6amu 58238
+6amv 58239
+6amw 58240
+6amx 58241
+6amz 58242
+6am0 58243
+6am2 58244
+6am3 58245
+6am4 58246
+6am5 58247
+6am6 58248
+6am7 58249
+6am8 58250
+6am9 58251
+6am+ 58252
+6am/ 58253
+6aqA 58254
+6aqB 58255
+6aqC 58256
+6aqE 58257
+6aqF 58258
+6aqG 58259
+6aqH 58260
+6aqI 58261
+6aqK 58262
+6aqL 58263
+6aqM 58264
+6aqN 58265
+6aqO 58266
+6aqP 58267
+6aqQ 58268
+6aqR 58269
+6aqS 58270
+6aqT 58271
+6aqV 58272
+6aqW 58273
+6aqX 58274
+6aqY 58275
+6aqa 58276
+6aqb 58277
+6aqc 58278
+6aqd 58279
+6aqe 58280
+6aqf 58281
+6aqg 58282
+6aqh 58283
+6aqi 58284
+6aqk 58285
+6aql 58286
+6aqn 58287
+6aqo 58288
+6aqv 58289
+6aqw 58290
+6aq2 58291
+6aq3 58292
+6aq4 58293
+6aq6 58294
+6aq8 58295
+6auA 58296
+6auB 58297
+6auC 58298
+6auD 58299
+6auE 58300
+6auF 58301
+6auI 58302
+6auL 58303
+6auM 58304
+6auO 58305
+6auP 58306
+6auR 58307
+6auS 58308
+6auT 58309
+6auU 58310
+6auY 58311
+6auZ 58312
+6auf 58313
+6auh 58314
+6aui 58315
+6auj 58316
+6aum 58317
+6auq 58318
+6aur 58319
+6aut 58320
+6auu 58321
+6auv 58322
+6aux 58323
+6au0 58324
+6au3 58325
+6au5 58326
+6au7 58327
+6au9 58328
+6ayD 58329
+6ayE 58330
+6ayG 58331
+6ayN 58332
+6ayP 58333
+6ayQ 58334
+6ayT 58335
+6ayY 58336
+6aya 58337
+6ayf 58338
+6ayi 58339
+6ayj 58340
+6ayl 58341
+6ayn 58342
+6ayo 58343
+6ayp 58344
+6ayu 58345
+6ayv 58346
+6ayx 58347
+6ayy 58348
+6ay2 58349
+6ay7 58350
+6ay8 58351
+6a2B 58352
+6a2C 58353
+6a2D 58354
+6a2E 58355
+6a2F 58356
+6a2H 58357
+6a2I 58358
+6a2J 58359
+6a2N 58360
+6a2O 58361
+6a2P 58362
+6a2R 58363
+6a2U 58364
+6a2Y 58365
+6a2a 58366
+6a2f 58367
+6a2j 58368
+6a2m 58369
+6a2s 58370
+6a2v 58371
+6a2z 58372
+6a20 58373
+6a21 58374
+6a6A 58375
+6a6D 58376
+6a6E 58377
+6a6J 58378
+6a6K 58379
+6a6L 58380
+6a6O 58381
+6a6P 58382
+6a6Q 58383
+6a6R 58384
+6a6S 58385
+6a6T 58386
+6a6U 58387
+6a6W 58388
+6a6X 58389
+6a6f 58390
+6a6g 58391
+6a6n 58392
+6a6o 58393
+6a6q 58394
+6a6r 58395
+6a6t 58396
+6a6u 58397
+6a6x 58398
+6a6y 58399
+6a60 58400
+6a64 58401
+6a65 58402
+6a+A 58403
+6a+G 58404
+6a+H 58405
+6a+J 58406
+6a+K 58407
+6a+O 58408
+6a+P 58409
+6a+R 58410
+6a+S 58411
+6a+U 58412
+6a+W 58413
+6a+b 58414
+6a+h 58415
+6a+i 58416
+6a+j 58417
+6a+l 58418
+6a+n 58419
+6a+o 58420
+6a+q 58421
+6a+w 58422
+6a+x 58423
+6a+y 58424
+6a+z 58425
+6a+1 58426
+6a+3 58427
+6a+9 58428
+6bCE 58429
+6bCG 58430
+6bCI 58431
+6bCJ 58432
+6bCK 58433
+6bCM 58434
+6bCN 58435
+6bCQ 58436
+6bCR 58437
+6bCS 58438
+6bCT 58439
+6bCU 58440
+6bCV 58441
+6bCW 58442
+6bCY 58443
+6bCZ 58444
+6bCa 58445
+6bCh 58446
+6bCk 58447
+6bCl 58448
+6bCn 58449
+6bCt 58450
+6bCu 58451
+6bCv 58452
+6bCw 58453
+6bCy 58454
+6bC2 58455
+6bC5 58456
+6bC6 58457
+6bC7 58458
+6bC+ 58459
+6bGA 58460
+6bGF 58461
+6bGG 58462
+6bGH 58463
+6bGI 58464
+6bGP 58465
+6bGS 58466
+6bGT 58467
+6bGX 58468
+6bGY 58469
+6bGa 58470
+6bGd 58471
+6bGg 58472
+6bGn 58473
+6bGu 58474
+6bGw 58475
+6bG1 58476
+6bG2 58477
+6bG4 58478
+6bG8 58479
+6bG/ 58480
+6bKA 58481
+6bKB 58482
+6bKC 58483
+6bKF 58484
+6bKH 58485
+6bKI 58486
+6bKK 58487
+6bKL 58488
+6bKM 58489
+6bKN 58490
+6bKQ 58491
+6bKR 58492
+6bKS 58493
+6bKU 58494
+6bKW 58495
+6bKa 58496
+6bKb 58497
+6bKc 58498
+6bKe 58499
+6bKf 58500
+6bKg 58501
+6bKh 58502
+6bKi 58503
+6bKj 58504
+6bKk 58505
+6bKl 58506
+6bKm 58507
+6bKn 58508
+6bKo 58509
+6bKp 58510
+6bKr 58511
+6bKt 58512
+6bKu 58513
+6bKx 58514
+6bKy 58515
+6bKz 58516
+6bK0 58517
+6bK1 58518
+6bK2 58519
+6bK3 58520
+6bK4 58521
+6bK7 58522
+6bK8 58523
+6bK9 58524
+6bOA 58525
+6bOD 58526
+6bOE 58527
+6bOF 58528
+6bOH 58529
+6bOJ 58530
+6bOK 58531
+6bOM 58532
+6bON 58533
+6bOO 58534
+6bOP 58535
+6bOQ 58536
+6bOU 58537
+6bOV 58538
+6bOW 58539
+6bOX 58540
+6bOZ 58541
+6bOc 58542
+6bOd 58543
+6bOe 58544
+6bOf 58545
+6bOj 58546
+6bOl 58547
+6bOn 58548
+6bOp 58549
+6bOr 58550
+6bOw 58551
+6bOz 58552
+6bO0 58553
+6bO2 58554
+6bSC 58555
+6bSD 58556
+6bSG 58557
+6bSH 58558
+6bSI 58559
+6bSJ 58560
+6bSL 58561
+6bSO 58562
+6bSQ 58563
+6bSS 58564
+6bSb 58565
+6bSe 58566
+6bSf 58567
+6bSj 58568
+6bSm 58569
+6bSo 58570
+6bSr 58571
+6bSs 58572
+6bSy 58573
+6bS7 58574
+6bS+ 58575
+6bS/ 58576
+6bWA 58577
+6bWE 58578
+6bWG 58579
+6bWH 58580
+6bWQ 58581
+6bWR 58582
+6bWZ 58583
+6bWc 58584
+6bWd 58585
+6bWe 58586
+6bWf 58587
+6bWg 58588
+6bWh 58589
+6bWk 58590
+6bWr 58591
+6bWs 58592
+6bWv 58593
+6bWw 58594
+6bWy 58595
+6bW6 58596
+6bW8 58597
+6baH 58598
+6baJ 58599
+6baP 58600
+6baS 58601
+6baW 58602
+6baX 58603
+6baa 58604
+6bah 58605
+6bak 58606
+6bap 58607
+6bar 58608
+6bav 58609
+6bay 58610
+6ba0 58611
+6ba4 58612
+6ba6 58613
+6ba7 58614
+6beB 58615
+6beC 58616
+6beD 58617
+6beG 58618
+6beT 58619
+6beZ 58620
+6bea 58621
+6beg 58622
+6bem 58623
+6bet 58624
+6bev 58625
+6bey 58626
+6be4 58627
+6be5 58628
+6be6 58629
+6be9 58630
+6biC 58631
+6biZ 58632
+6bia 58633
+6bib 58634
+6bie 58635
+6bif 58636
+6big 58637
+6bih 58638
+6bii 58639
+6bij 58640
+6bil 58641
+6bim 58642
+6bio 58643
+6bip 58644
+6biq 58645
+6bir 58646
+6bis 58647
+6bit 58648
+6biu 58649
+6biv 58650
+6biw 58651
+6bix 58652
+6biz 58653
+6bi1 58654
+6bi2 58655
+6bi3 58656
+6bi4 58657
+6bi5 58658
+6bi9 58659
+6bi+ 58660
+6bi/ 58661
+6bmA 58662
+6bmB 58663
+6bmC 58664
+6bmD 58665
+6bmE 58666
+6bmF 58667
+6bmG 58668
+6bmH 58669
+6bmI 58670
+6bmJ 58671
+6bmK 58672
+6bmL 58673
+6bmM 58674
+6bmO 58675
+6bmP 58676
+6bmR 58677
+6bmV 58678
+6bmW 58679
+6bmX 58680
+6bmY 58681
+6bma 58682
+6bmc 58683
+6bme 58684
+6bmj 58685
+6bmk 58686
+6bmm 58687
+6bmn 58688
+6bmo 58689
+6bmp 58690
+6bmq 58691
+6bmr 58692
+6bms 58693
+6bmt 58694
+6bmu 58695
+6bmv 58696
+6bmw 58697
+6bmx 58698
+6bmz 58699
+6bm1 58700
+6bm4 58701
+6bm5 58702
+6bm9 58703
+6bm+ 58704
+6bm/ 58705
+6bqB 58706
+6bqC 58707
+6bqH 58708
+6bqI 58709
+6bqL 58710
+6bqS 58711
+6bqT 58712
+6bqV 58713
+6bqX 58714
+6bqd 58715
+6bqe 58716
+6bqf 58717
+6bql 58718
+6bqm 58719
+6bqp 58720
+6bqq 58721
+6bqt 58722
+6bq0 58723
+6bq1 58724
+6bq4 58725
+6bq5 58726
+6bq6 58727
+6bq7 58728
+6bq8 58729
+6bq9 58730
+6bq+ 58731
+6bq/ 58732
+6buD 58733
+6buE 58734
+6buJ 58735
+6buM 58736
+6buN 58737
+6buO 58738
+6buP 58739
+6buQ 58740
+6buR 58741
+6buS 58742
+6buU 58743
+6buY 58744
+6buZ 58745
+6bub 58746
+6buc 58747
+6bud 58748
+6bue 58749
+6buf 58750
+6bug 58751
+6buh 58752
+6bui 58753
+6bul 58754
+6bun 58755
+6buo 58756
+6bup 58757
+6buv 58758
+6bu0 58759
+6bu5 58760
+6bu7 58761
+6bu8 58762
+6bu9 58763
+6byH 58764
+6byI 58765
+6byL 58766
+6byN 58767
+6byO 58768
+6byQ 58769
+6byT 58770
+6byV 58771
+6byZ 58772
+6byg 58773
+6byh 58774
+6byi 58775
+6byp 58776
+6byq 58777
+6bys 58778
+6byv 58779
+6byx 58780
+6by5 58781
+6by7 58782
+6by+ 58783
+6b2B 58784
+6b2J 58785
+6b2K 58786
+6b2L 58787
+6b2O 58788
+6b2Q 58789
+6b2R 58790
+6b2S 58791
+6b2U 58792
+6b2f 58793
+6b2h 58794
+6b2i 58795
+6b2j 58796
+6b2m 58797
+6b2n 58798
+6b2q 58799
+6b2s 58800
+6b2y 58801
+6b23 58802
+6b2/ 58803
+6b6D 58804
+6b6E 58805
+6b6F 58806
+6b6H 58807
+6b6I 58808
+6b6J 58809
+6b6K 58810
+6b6L 58811
+6b6M 58812
+6b6N 58813
+6b6Q 58814
+6b6V 58815
+6b6Z 58816
+6b6a 58817
+6b6b 58818
+6b6c 58819
+6b6d 58820
+6b6f 58821
+6b6g 58822
+6b6i 58823
+77yB 58824
+77yI 58825
+77yJ 58826
+77yM 58827
+77yN 58828
+77ya 58829
+77yb 58830
+77yf 58831
+772Y 58832
+772Z 58833
+8KCx 58834
+8KCxgQ== 58835
diff --git a/projects/llm_framework/main_cosy_voice/scripts/tokenizer/tokenizer.py b/projects/llm_framework/main_cosy_voice/scripts/tokenizer/tokenizer.py
new file mode 100755
index 00000000..43fb39a2
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/scripts/tokenizer/tokenizer.py
@@ -0,0 +1,279 @@
+import base64
+import os
+from functools import lru_cache
+from typing import Optional
+import torch
+from transformers import AutoTokenizer
+from whisper.tokenizer import Tokenizer
+
+import tiktoken
+
+LANGUAGES = {
+    "en": "english",
+    "zh": "chinese",
+    "de": "german",
+    "es": "spanish",
+    "ru": "russian",
+    "ko": "korean",
+    "fr": "french",
+    "ja": "japanese",
+    "pt": "portuguese",
+    "tr": "turkish",
+    "pl": "polish",
+    "ca": "catalan",
+    "nl": "dutch",
+    "ar": "arabic",
+    "sv": "swedish",
+    "it": "italian",
+    "id": "indonesian",
+    "hi": "hindi",
+    "fi": "finnish",
+    "vi": "vietnamese",
+    "he": "hebrew",
+    "uk": "ukrainian",
+    "el": "greek",
+    "ms": "malay",
+    "cs": "czech",
+    "ro": "romanian",
+    "da": "danish",
+    "hu": "hungarian",
+    "ta": "tamil",
+    "no": "norwegian",
+    "th": "thai",
+    "ur": "urdu",
+    "hr": "croatian",
+    "bg": "bulgarian",
+    "lt": "lithuanian",
+    "la": "latin",
+    "mi": "maori",
+    "ml": "malayalam",
+    "cy": "welsh",
+    "sk": "slovak",
+    "te": "telugu",
+    "fa": "persian",
+    "lv": "latvian",
+    "bn": "bengali",
+    "sr": "serbian",
+    "az": "azerbaijani",
+    "sl": "slovenian",
+    "kn": "kannada",
+    "et": "estonian",
+    "mk": "macedonian",
+    "br": "breton",
+    "eu": "basque",
+    "is": "icelandic",
+    "hy": "armenian",
+    "ne": "nepali",
+    "mn": "mongolian",
+    "bs": "bosnian",
+    "kk": "kazakh",
+    "sq": "albanian",
+    "sw": "swahili",
+    "gl": "galician",
+    "mr": "marathi",
+    "pa": "punjabi",
+    "si": "sinhala",
+    "km": "khmer",
+    "sn": "shona",
+    "yo": "yoruba",
+    "so": "somali",
+    "af": "afrikaans",
+    "oc": "occitan",
+    "ka": "georgian",
+    "be": "belarusian",
+    "tg": "tajik",
+    "sd": "sindhi",
+    "gu": "gujarati",
+    "am": "amharic",
+    "yi": "yiddish",
+    "lo": "lao",
+    "uz": "uzbek",
+    "fo": "faroese",
+    "ht": "haitian creole",
+    "ps": "pashto",
+    "tk": "turkmen",
+    "nn": "nynorsk",
+    "mt": "maltese",
+    "sa": "sanskrit",
+    "lb": "luxembourgish",
+    "my": "myanmar",
+    "bo": "tibetan",
+    "tl": "tagalog",
+    "mg": "malagasy",
+    "as": "assamese",
+    "tt": "tatar",
+    "haw": "hawaiian",
+    "ln": "lingala",
+    "ha": "hausa",
+    "ba": "bashkir",
+    "jw": "javanese",
+    "su": "sundanese",
+    "yue": "cantonese",
+    "minnan": "minnan",
+    "wuyu": "wuyu",
+    "dialect": "dialect",
+    "zh/en": "zh/en",
+    "en/zh": "en/zh",
+}
+
+# language code lookup by name, with a few language aliases
+TO_LANGUAGE_CODE = {
+    **{language: code for code, language in LANGUAGES.items()},
+    "burmese": "my",
+    "valencian": "ca",
+    "flemish": "nl",
+    "haitian": "ht",
+    "letzeburgesch": "lb",
+    "pushto": "ps",
+    "panjabi": "pa",
+    "moldavian": "ro",
+    "moldovan": "ro",
+    "sinhalese": "si",
+    "castilian": "es",
+    "mandarin": "zh",
+}
+
+AUDIO_EVENT = {
+    "ASR": "ASR",
+    "AED": "AED",
+    "SER": "SER",
+    "Speech": "Speech",
+    "/Speech": "/Speech",
+    "BGM": "BGM",
+    "/BGM": "/BGM",
+    "Laughter": "Laughter",
+    "/Laughter": "/Laughter",
+    "Applause": "Applause",
+    "/Applause": "/Applause",
+}
+
+EMOTION = {
+    "HAPPY": "HAPPY",
+    "SAD": "SAD",
+    "ANGRY": "ANGRY",
+    "NEUTRAL": "NEUTRAL",
+}
+
+TTS_Vocal_Token = {
+    "TTS/B": "TTS/B",
+    "TTS/O": "TTS/O",
+    "TTS/Q": "TTS/Q",
+    "TTS/A": "TTS/A",
+    "TTS/CO": "TTS/CO",
+    "TTS/CL": "TTS/CL",
+    "TTS/H": "TTS/H",
+    **{f"TTS/SP{i:02d}": f"TTS/SP{i:02d}" for i in range(1, 14)}
+}
+
+
+@lru_cache(maxsize=None)
+def get_encoding(name: str = "gpt2", num_languages: int = 99):
+    vocab_path = os.path.join(os.path.dirname(__file__), "assets", f"{name}.tiktoken")
+    ranks = {
+        base64.b64decode(token): int(rank)
+        for token, rank in (line.split() for line in open(vocab_path) if line)
+    }
+    n_vocab = len(ranks)
+    special_tokens = {}
+
+    specials = [
+        "<|endoftext|>",
+        "<|startoftranscript|>",
+        *[f"<|{lang}|>" for lang in list(LANGUAGES.keys())[:num_languages]],
+        *[f"<|{audio_event}|>" for audio_event in list(AUDIO_EVENT.keys())],
+        *[f"<|{emotion}|>" for emotion in list(EMOTION.keys())],
+        "<|translate|>",
+        "<|transcribe|>",
+        "<|startoflm|>",
+        "<|startofprev|>",
+        "<|nospeech|>",
+        "<|notimestamps|>",
+        *[f"<|SPECIAL_TOKEN_{i}|>" for i in range(1, 31)],        # register special tokens for ASR
+        *[f"<|{tts}|>" for tts in list(TTS_Vocal_Token.keys())],  # register special tokens for TTS
+        *[f"<|{i * 0.02:.2f}|>" for i in range(1501)],
+    ]
+
+    for token in specials:
+        special_tokens[token] = n_vocab
+        n_vocab += 1
+
+    return tiktoken.Encoding(
+        name=os.path.basename(vocab_path),
+        explicit_n_vocab=n_vocab,
+        pat_str=r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""",
+        mergeable_ranks=ranks,
+        special_tokens=special_tokens,
+    )
+
+
+@lru_cache(maxsize=None)
+def get_tokenizer(
+    multilingual: bool,
+    *,
+    num_languages: int = 99,
+    language: Optional[str] = None,
+    task: Optional[str] = None,  # Literal["transcribe", "translate", None]
+) -> Tokenizer:
+    if language is not None:
+        language = language.lower()
+        if language not in LANGUAGES:
+            if language in TO_LANGUAGE_CODE:
+                language = TO_LANGUAGE_CODE[language]
+            else:
+                raise ValueError(f"Unsupported language: {language}")
+
+    if multilingual:
+        encoding_name = "multilingual_zh_ja_yue_char_del"
+        language = language or "en"
+        task = task or "transcribe"
+    else:
+        encoding_name = "gpt2"
+        language = None
+        task = None
+
+    encoding = get_encoding(name=encoding_name, num_languages=num_languages)
+
+    return Tokenizer(
+        encoding=encoding, num_languages=num_languages, language=language, task=task
+    )
+
+
+class QwenTokenizer():
+    def __init__(self, token_path, skip_special_tokens=True):
+        super().__init__()
+        # NOTE: non-chat model, all these special tokens keep randomly initialized.
+        special_tokens = {
+            'eos_token': '<|endoftext|>',
+            'pad_token': '<|endoftext|>',
+            'additional_special_tokens': [
+                '<|im_start|>', '<|im_end|>', '<|endofprompt|>',
+                '[breath]', '<strong>', '</strong>', '[noise]',
+                '[laughter]', '[cough]', '[clucking]', '[accent]',
+                '[quick_breath]',
+                "<laughter>", "</laughter>",
+                "[hissing]", "[sigh]", "[vocalized-noise]",
+                "[lipsmack]", "[mn]"
+            ]
+        }
+        self.special_tokens = special_tokens
+        self.tokenizer = AutoTokenizer.from_pretrained(token_path)
+        self.tokenizer.add_special_tokens(special_tokens)
+        self.skip_special_tokens = skip_special_tokens
+
+    def encode(self, text, **kwargs):
+        tokens = self.tokenizer([text], return_tensors="pt")
+        tokens = tokens["input_ids"][0].cpu().tolist()
+        return tokens
+
+    def decode(self, tokens):
+        tokens = torch.tensor(tokens, dtype=torch.int64)
+        text = self.tokenizer.batch_decode([tokens], skip_special_tokens=self.skip_special_tokens)[0]
+        return text
+
+
+@lru_cache(maxsize=None)
+def get_qwen_tokenizer(
+    token_path: str,
+    skip_special_tokens: bool
+) -> QwenTokenizer:
+    return QwenTokenizer(token_path=token_path, skip_special_tokens=skip_special_tokens)
diff --git a/projects/llm_framework/main_cosy_voice/scripts/tokenizer_CosyVoice2-0.5B-ax650.py b/projects/llm_framework/main_cosy_voice/scripts/tokenizer_CosyVoice2-0.5B-ax650.py
new file mode 100755
index 00000000..2aa93bf5
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/scripts/tokenizer_CosyVoice2-0.5B-ax650.py
@@ -0,0 +1,122 @@
+from transformers import AutoTokenizer, PreTrainedTokenizerFast
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+from tokenizer.tokenizer import get_qwen_tokenizer
+
+class Tokenizer_Http():
+
+    def __init__(self, model_id):
+        
+        self.tokenizer = get_qwen_tokenizer(model_id, True)
+
+    def encode(self, prompt):
+    
+        token_ids = self.tokenizer.encode(prompt, allowed_special="all")
+        return token_ids
+
+    def decode(self, token_ids):
+        return self.tokenizer.decode(token_ids)
+
+    @property
+    def bos_id(self):
+        return 0
+
+    @property
+    def eos_id(self):
+        return 1773
+
+    # @property
+    # def bos_token(self):
+    #     return self.tokenizer.bos_token
+
+    @property
+    def eos_token(self):
+        return "<|eot_id|>"
+
+class Request(BaseHTTPRequestHandler):
+    #通过类继承，新定义类
+    timeout = 5
+    server_version = 'Apache'
+
+    def do_GET(self):
+        print(self.path)
+        #在新类中定义get的内容（当客户端向该服务端使用get请求时，本服务端将如下运行）
+        self.send_response(200)
+        self.send_header("type", "get")  #设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == '/bos_id':
+            bos_id = tokenizer.bos_id
+            # print(bos_id)
+            # to json
+            if bos_id is None:
+                msg = json.dumps({'bos_id': -1})
+            else:
+                msg = json.dumps({'bos_id': bos_id})
+        elif self.path == '/eos_id':
+            eos_id = tokenizer.eos_id
+            if eos_id is None:
+                msg = json.dumps({'eos_id': -1})
+            else:
+                msg = json.dumps({'eos_id': eos_id})
+        else:
+            msg = 'error'
+
+        print(msg)
+        msg = str(msg).encode()  #转为str再转为byte格式
+
+        self.wfile.write(msg)  #将byte格式的信息返回给客户端
+
+    def do_POST(self):
+        #在新类中定义post的内容（当客户端向该服务端使用post请求时，本服务端将如下运行）
+        data = self.rfile.read(int(
+            self.headers['content-length']))  #获取从客户端传入的参数（byte格式）
+        data = data.decode()  #将byte格式转为str格式
+
+        self.send_response(200)
+        self.send_header("type", "post")  #设置响应头，可省略或设置多个
+        self.end_headers()
+
+        if self.path == '/encode':
+            req = json.loads(data)
+            prompt = req['text']
+
+            token_ids = tokenizer.encode(prompt)
+            if token_ids is None:
+                msg = json.dumps({'token_ids': -1})
+            else:
+                msg = json.dumps({'token_ids': token_ids})
+
+        elif self.path == '/decode':
+            req = json.loads(data)
+            token_ids = req['token_ids']
+            text = tokenizer.decode(token_ids)
+            if text is None:
+                msg = json.dumps({'text': ""})
+            else:
+                msg = json.dumps({'text': text})
+        else:
+            msg = 'error'
+        print(msg)
+        msg = str(msg).encode()  #转为str再转为byte格式
+
+        self.wfile.write(msg)  #将byte格式的信息返回给客户端
+
+
+if __name__ == "__main__":
+
+    args = argparse.ArgumentParser()
+    args.add_argument('--host', type=str, default='localhost')
+    args.add_argument('--port', type=int, default=12345)
+    args.add_argument('--model_id', type=str, default='tokenizer')
+
+    args = args.parse_args()
+
+    tokenizer = Tokenizer_Http(args.model_id)
+
+
+    host = (args.host, args.port)  #设定地址与端口号，'localhost'等价于'127.0.0.1'
+    print('http://%s:%s' % host)
+    server = HTTPServer(host, Request)  #根据地址端口号和新定义的类，创建服务器实例
+    server.serve_forever()  #开启服务
diff --git a/projects/llm_framework/main_cosy_voice/src/main.cpp b/projects/llm_framework/main_cosy_voice/src/main.cpp
new file mode 100644
index 00000000..a3abbe1d
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/main.cpp
@@ -0,0 +1,885 @@
+/*
+ * SPDX-FileCopyrightText: 2024 M5Stack Technology CO LTD
+ *
+ * SPDX-License-Identifier: MIT
+ */
+#include "StackFlow.h"
+#include "runner/LLM.hpp"
+#include "runner/Token2wav.hpp"
+
+#include <signal.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+#include <base64.h>
+#include <fstream>
+#include <stdexcept>
+#include <semaphore.h>
+#include "../../../../SDK/components/utilities/include/sample_log.h"
+#include "thread_safe_list.h"
+using namespace StackFlows;
+#ifdef ENABLE_BACKWARD
+#define BACKWARD_HAS_DW 1
+#include "backward.hpp"
+#include "backward.h"
+#endif
+
+#include "runner/utils/wav.hpp"
+
+#define MAX_TASK_NUM 2
+
+int main_exit_flage = 0;
+static void __sigint(int iSigNo)
+{
+    SLOGW("llm_cosy_voice will be exit!");
+    main_exit_flage = 1;
+}
+
+static std::string base_model_path_;
+static std::string base_model_config_path_;
+
+typedef std::function<void(const std::string &data, bool finish)> task_callback_t;
+
+#define CONFIG_AUTO_SET(obj, key)             \
+    if (config_body.contains(#key))           \
+        mode_config_.key = config_body[#key]; \
+    else if (obj.contains(#key))              \
+        mode_config_.key = obj[#key];
+
+class llm_task {
+private:
+    static std::atomic<unsigned int> next_port_;
+    std::atomic_bool tokenizer_server_flage_;
+    unsigned int port_;
+    pid_t tokenizer_pid_ = -1;
+    std::mutex g_buffer_mutex;
+    std::condition_variable g_buffer_cv;
+    std::atomic<bool> g_llm_finished{false};
+    std::atomic<bool> g_stop{false};
+    TokenBuffer g_token_buffer;
+    Token2Wav lToken2Wav;
+
+    std::vector<int> prompt_text_token;
+    std::vector<unsigned short> prompt_text_embeds;
+    std::vector<int> prompt_speech_token;
+    std::vector<unsigned short> prompt_speech_embeds;
+
+    std::vector<float> prompt_feat;
+    std::vector<float> prompt_speech_embeds_flow;
+    std::vector<float> spk_embeds;
+
+public:
+    enum inference_status { INFERENCE_NONE = 0, INFERENCE_RUNNING };
+    LLMAttrType mode_config_;
+    std::unique_ptr<LLM> lLaMa_;
+    std::string model_;
+    std::string response_format_;
+    std::vector<std::string> inputs_;
+    std::string prompt_;
+    std::string last_reply;
+    std::vector<unsigned short> prompt_data;
+    std::vector<int> tokens_ids, tokens_diff;
+    std::vector<std::vector<unsigned short>> k_caches, v_caches;
+    std::string kvcache_path;
+    int precompute_len = 0;
+    std::vector<int> _token_ids;
+    task_callback_t out_callback_;
+    bool enoutput_;
+    bool enstream_;
+
+    std::unique_ptr<std::thread> inference_run_;
+    thread_safe::list<std::string> async_list_;
+
+    void set_output(task_callback_t out_callback)
+    {
+        out_callback_ = out_callback;
+    }
+
+    bool parse_config(const nlohmann::json &config_body)
+    {
+        try {
+            model_           = config_body.at("model");
+            response_format_ = config_body.at("response_format");
+            enoutput_        = config_body.at("enoutput");
+            prompt_          = config_body.at("prompt");
+
+            if (config_body.contains("input")) {
+                if (config_body["input"].is_string()) {
+                    inputs_.push_back(config_body["input"].get<std::string>());
+                } else if (config_body["input"].is_array()) {
+                    for (auto _in : config_body["input"]) {
+                        inputs_.push_back(_in.get<std::string>());
+                    }
+                }
+            }
+        } catch (...) {
+            SLOGE("setup config_body error");
+            return true;
+        }
+        enstream_ = (response_format_.find("stream") != std::string::npos);
+        return false;
+    }
+
+    void prepare_kvcache_folder(const std::string &kvcache_path)
+    {
+        try {
+            if (!std::filesystem::exists(kvcache_path)) {
+                std::filesystem::create_directories(kvcache_path);
+            }
+
+            if (std::filesystem::exists(kvcache_path) && std::filesystem::is_directory(kvcache_path)) {
+                for (const auto &entry : std::filesystem::directory_iterator(kvcache_path)) {
+                    std::filesystem::remove_all(entry.path());
+                }
+            }
+        } catch (const std::exception &e) {
+            ALOGI("prepare_kvcache_folder: skip clear/create due to error: %s", e.what());
+        }
+    }
+
+    int load_model(const nlohmann::json &config_body)
+    {
+        if (parse_config(config_body)) {
+            return -1;
+        }
+        nlohmann::json file_body;
+        std::list<std::string> config_file_paths =
+            get_config_file_paths(base_model_path_, base_model_config_path_, model_);
+        try {
+            for (auto file_name : config_file_paths) {
+                std::ifstream config_file(file_name);
+                if (!config_file.is_open()) {
+                    SLOGW("config file :%s miss", file_name.c_str());
+                    continue;
+                }
+                SLOGI("config file :%s read", file_name.c_str());
+                config_file >> file_body;
+                config_file.close();
+                break;
+            }
+            if (file_body.empty()) {
+                SLOGE("all config file miss");
+                return -2;
+            }
+            std::string base_model = base_model_path_ + model_ + "/";
+            SLOGI("base_model %s", base_model.c_str());
+
+            CONFIG_AUTO_SET(file_body["mode_param"], tokenizer_type);
+            CONFIG_AUTO_SET(file_body["mode_param"], filename_tokenizer_model);
+            CONFIG_AUTO_SET(file_body["mode_param"], url_tokenizer_model);
+            CONFIG_AUTO_SET(file_body["mode_param"], filename_post_axmodel);
+            CONFIG_AUTO_SET(file_body["mode_param"], filename_tokens_embed);
+            CONFIG_AUTO_SET(file_body["mode_param"], filename_llm_embed);
+            CONFIG_AUTO_SET(file_body["mode_param"], filename_speech_embed);
+            CONFIG_AUTO_SET(file_body["mode_param"], filename_decoder_axmodel);
+            CONFIG_AUTO_SET(file_body["mode_param"], token2wav_axmodel_dir);
+            CONFIG_AUTO_SET(file_body["mode_param"], n_timesteps);
+            CONFIG_AUTO_SET(file_body["mode_param"], prompt_files);
+            CONFIG_AUTO_SET(file_body["mode_param"], template_filename_axmodel);
+            CONFIG_AUTO_SET(file_body["mode_param"], b_bos);
+            CONFIG_AUTO_SET(file_body["mode_param"], b_eos);
+            CONFIG_AUTO_SET(file_body["mode_param"], axmodel_num);
+            CONFIG_AUTO_SET(file_body["mode_param"], tokens_embed_num);
+            CONFIG_AUTO_SET(file_body["mode_param"], tokens_embed_size);
+            CONFIG_AUTO_SET(file_body["mode_param"], b_use_mmap_load_embed);
+            CONFIG_AUTO_SET(file_body["mode_param"], b_dynamic_load_axmodel_layer);
+            CONFIG_AUTO_SET(file_body["mode_param"], max_token_len);
+            {
+                auto has_http = [](const std::string &s) { return s.find("http") != std::string::npos; };
+
+                auto find_tokenizer_file = [this]() -> std::string {
+                    const std::string base = "/opt/m5stack/scripts/";
+                    const std::string a    = base + model_ + "_tokenizer.py";
+                    if (file_exists(a)) return a;
+                    const std::string b = base + "tokenizer_" + model_ + ".py";
+                    if (file_exists(b)) return b;
+                    SLOGE("%s or %s not found!", a.c_str(), b.c_str());
+                    return {};
+                };
+
+                auto start_tokenizer_server = [&](const std::string &tokenizer_file) {
+                    if (tokenizer_file.empty()) return;
+                    if (tokenizer_server_flage_.load()) return;
+
+                    tokenizer_pid_ = fork();
+                    if (tokenizer_pid_ == 0) {
+                        setenv("PYTHONPATH", "/opt/m5stack/lib/cosy_voice/site-packages", 1);
+                        const std::string port_str = std::to_string(port_);
+                        const std::string model_id = base_model + "tokenizer";
+
+                        execl("/usr/bin/python3", "python3", tokenizer_file.c_str(), "--host", "localhost", "--port",
+                              port_str.c_str(), "--model_id", model_id.c_str(), (char *)nullptr);
+
+                        perror("execl failed");
+                        _exit(1);
+                    }
+
+                    tokenizer_server_flage_.store(true);
+                    SLOGI("port_=%s model_id=%s content=%s", std::to_string(port_).c_str(),
+                          (base_model + std::string("tokenizer")).c_str(), prompt_.c_str());
+
+                    std::this_thread::sleep_for(std::chrono::seconds(15));
+                };
+
+                auto process_field = [&](std::string &field, const char *name_for_log) -> bool {
+                    if (!has_http(field)) return false;
+
+                    field                            = "http://localhost:" + std::to_string(port_);
+                    const std::string tokenizer_file = find_tokenizer_file();
+                    start_tokenizer_server(tokenizer_file);
+                    SLOGI("%s: %s", name_for_log, field.c_str());
+                    return true;
+                };
+
+                if (!process_field(mode_config_.filename_tokenizer_model, "filename_tokenizer_model") &&
+                    !process_field(mode_config_.url_tokenizer_model, "url_tokenizer_model")) {
+                    mode_config_.filename_tokenizer_model = base_model + mode_config_.filename_tokenizer_model;
+                    SLOGI("filename_tokenizer_model: %s", mode_config_.filename_tokenizer_model.c_str());
+                }
+            }
+            mode_config_.filename_tokens_embed     = base_model + mode_config_.filename_tokens_embed;
+            mode_config_.filename_post_axmodel     = base_model + mode_config_.filename_post_axmodel;
+            mode_config_.filename_llm_embed        = base_model + mode_config_.filename_llm_embed;
+            mode_config_.template_filename_axmodel = base_model + mode_config_.template_filename_axmodel;
+            mode_config_.filename_speech_embed     = base_model + mode_config_.filename_speech_embed;
+            mode_config_.filename_decoder_axmodel  = base_model + mode_config_.filename_decoder_axmodel;
+            mode_config_.token2wav_axmodel_dir     = base_model + mode_config_.token2wav_axmodel_dir;
+            mode_config_.prompt_files              = base_model + mode_config_.prompt_files;
+            mode_config_.runing_callback = [this](int *p_token, int n_token, const char *p_str, float token_per_sec,
+                                                  void *reserve) {
+                if (this->out_callback_) {
+                    this->out_callback_(std::string(p_str), false);
+                }
+            };
+
+            readtxt(mode_config_.prompt_files + "/prompt_text.txt", prompt_text_token);
+            readtxt(mode_config_.prompt_files + "/llm_prompt_speech_token.txt", prompt_speech_token);
+            readtxt(mode_config_.prompt_files + "/prompt_speech_feat.txt", prompt_feat);
+            readtxt<float>(mode_config_.prompt_files + "/flow_embedding.txt", spk_embeds);
+
+            lLaMa_ = std::make_unique<LLM>();
+            if (!lLaMa_->Init(mode_config_)) {
+                lLaMa_->Deinit();
+                lLaMa_.reset();
+                return -2;
+            }
+            if (!lToken2Wav.Init(mode_config_.token2wav_axmodel_dir, mode_config_.n_timesteps)) {
+                return -1;
+            }
+            SLOGE();
+            lLaMa_->TextToken2Embeds(prompt_text_token, prompt_text_embeds);
+            SLOGE();
+            lLaMa_->SpeechToken2Embeds(prompt_speech_token, prompt_speech_embeds);
+            SLOGE();
+            lToken2Wav.SpeechToken2Embeds(prompt_speech_token, prompt_speech_embeds_flow);
+
+        } catch (...) {
+            SLOGE("config false");
+            return -3;
+        }
+        return 0;
+    }
+
+    std::string prompt_complete(const std::string &input)
+    {
+        std::ostringstream oss_prompt;
+        switch (mode_config_.tokenizer_type) {
+            case TKT_LLaMa:
+                oss_prompt << "<|user|>\n" << input << "</s><|assistant|>\n";
+                break;
+            case TKT_MINICPM:
+                oss_prompt << "<用户>" << input << "<AI>";
+                break;
+            case TKT_Phi3:
+                oss_prompt << input << " ";
+                break;
+            case TKT_Qwen:
+                oss_prompt << "<|im_start|>system\n" << prompt_ << ".<|im_end|>";
+                oss_prompt << "\n<|im_start|>user\n" << input << "<|im_end|>\n<|im_start|>assistant\n";
+                break;
+            case TKT_HTTP:
+            default:
+                oss_prompt << input;
+                break;
+        }
+        SLOGI("prompt_complete:%s", oss_prompt.str().c_str());
+        return oss_prompt.str();
+    }
+
+    void reset()
+    {
+        g_llm_finished = false;
+        g_token_buffer.erase(g_token_buffer.begin(), g_token_buffer.end());
+        lToken2Wav.reset();
+    }
+
+    int tts(const std::string &text, const std::vector<unsigned short> &prompt_text_embeds,
+            const std::vector<unsigned short> &prompt_speech_embeds, const std::vector<float> &prompt_feat,
+            const std::vector<float> &prompt_speech_embeds_flow, std::vector<float> &spk_embeds)
+    {
+        std::vector<float> output;
+        timer time_total;
+        time_total.start();
+        try {
+            auto llm_thread_func = [this, &text, &prompt_text_embeds, &prompt_speech_embeds]() {
+                lLaMa_->Run(text, prompt_text_embeds, prompt_speech_embeds, g_token_buffer, g_buffer_mutex, g_buffer_cv,
+                            g_llm_finished);
+            };
+
+            // Start the LLM in a separate thread
+            std::thread llm_thread(llm_thread_func);
+
+            int token_offset     = 0;
+            int prompt_token_len = prompt_speech_embeds_flow.size() / lToken2Wav.flow_embed_size;
+            if (prompt_token_len < 75) {
+                SLOGE("Error, prompt speech token len %d < 75", prompt_token_len);
+                return -1;
+            }
+            // int prompt_token_align_len = int(prompt_token_len / lToken2Wav.token_hop_len) * lToken2Wav.token_hop_len;
+            int prompt_token_align_len = 75;  // only support 75 now
+
+            std::vector<float> prompt_speech_embeds_flow1;
+            prompt_speech_embeds_flow1.insert(prompt_speech_embeds_flow1.begin(), prompt_speech_embeds_flow.begin(),
+                                              prompt_speech_embeds_flow.begin() + prompt_token_align_len * 512);
+
+            std::vector<float> prompt_feat1;
+            prompt_feat1.insert(prompt_feat1.begin(), prompt_feat.begin(),
+                                prompt_feat.begin() + prompt_token_align_len * 2 * 80);
+
+            int promot_token_pad = 0;
+            int this_token_hop_len;
+            int i = 0;
+            while (true) {
+                this_token_hop_len =
+                    (token_offset == 0) ? lToken2Wav.token_hop_len + promot_token_pad : lToken2Wav.token_hop_len;
+
+                std::unique_lock<std::mutex> lock(g_buffer_mutex);
+
+                g_buffer_cv.wait(lock, [&] {
+                    return (g_token_buffer.size() - token_offset >=
+                            this_token_hop_len + lToken2Wav.pre_lookahead_len) ||
+                           g_llm_finished.load() || g_stop.load();
+                });
+
+                if (g_stop) {
+                    lock.unlock();
+                    break;
+                } else if (g_token_buffer.size() - token_offset >= this_token_hop_len + lToken2Wav.pre_lookahead_len) {
+                    std::vector<SpeechToken> token;
+                    int start = token_offset - std::min(int(token_offset / lToken2Wav.token_hop_len),
+                                                        lToken2Wav.max_infer_chunk_num - 1) *
+                                                   lToken2Wav.token_hop_len;
+                    int end = token_offset + this_token_hop_len + lToken2Wav.pre_lookahead_len;
+
+                    token.insert(token.end(), g_token_buffer.begin() + start, g_token_buffer.begin() + end);
+
+                    lock.unlock();
+                    std::cout << "[Main/Token2Wav Thread] Processing batch of " << token.size() << " tokens...\n";
+                    auto speech = lToken2Wav.infer(token, prompt_speech_embeds_flow1, prompt_feat1, spk_embeds,
+                                                   token_offset, false);
+                    token_offset += this_token_hop_len;
+                    output.insert(output.end(), speech.begin(), speech.end());
+                    std::string path = "output_" + std::to_string(i) + ".wav";
+
+                    saveVectorAsWavFloat(speech, path, 24000, 1);
+                    i += 1;
+
+                }
+
+                else if (g_llm_finished.load()) {
+                    std::cout << "[Main/Token2Wav Thread] Buffer is empty and LLM finished. Exiting.\n";
+                    lock.unlock();
+                    break;
+                }
+                // Check exit condition: Buffer is empty and LLM is done
+                else {
+                    // This else branch is technically not needed because the wait condition
+                    // ensures we only get here if one of the conditions is true.
+                    // But it's good practice to structure logic clearly.
+                    // In this specific loop, we will always process if we wake up.
+                    lock.unlock();  // Make sure to unlock if not processing
+                }
+            }
+
+            // Wait for the LLM thread to finish
+            if (llm_thread.joinable()) {
+                llm_thread.join();
+            }
+
+            if (g_stop) {
+                g_token_buffer.erase(g_token_buffer.begin(), g_token_buffer.end());
+                return 1;
+            }
+
+            std::vector<SpeechToken> token;
+            int start = g_token_buffer.size() - std::min(int(g_token_buffer.size() / lToken2Wav.token_hop_len),
+                                                         lToken2Wav.max_infer_chunk_num - 1) *
+                                                    lToken2Wav.token_hop_len;
+            token.insert(token.end(), g_token_buffer.begin() + start, g_token_buffer.end());
+            auto speech = lToken2Wav.infer(token, prompt_speech_embeds_flow1, prompt_feat1, spk_embeds,
+                                           token_offset - start, true);
+            // TODO: 另起一个线程处理生成的音频
+            output.insert(output.end(), speech.begin(), speech.end());
+            std::string path = "output_" + std::to_string(i) + ".wav";
+            saveVectorAsWavFloat(speech, path, 24000, 1);
+            saveVectorAsWavFloat(output, "output.wav", 24000, 1);
+
+            SLOGI("tts total use time: %.3f s", time_total.cost() / 1000);
+            reset();
+            std::cout << "\nVoice generation pipeline completed.\n";
+
+        } catch (const std::exception &e) {
+            std::cerr << "Error in pipeline: " << e.what() << std::endl;
+            return 1;
+        }
+
+        return 0;
+    }
+
+    void run()
+    {
+        std::string par;
+        for (;;) {
+            {
+                par = async_list_.get();
+                if (par.empty()) break;
+                inference(par);
+            }
+        }
+    }
+
+    int inference_async(const std::string &msg)
+    {
+        if (msg.empty()) return -1;
+        if (async_list_.size() < 3) {
+            std::string par = msg;
+            async_list_.put(par);
+        } else {
+            SLOGE("inference list is full\n");
+        }
+        return async_list_.size();
+    }
+
+    void inference(const std::string &msg)
+    {
+        try {
+            // std::string out = lLaMa_->Run(prompt_complete(msg));
+            // if (out_callback_) out_callback_(out, true);
+            tts(msg, prompt_text_embeds, prompt_speech_embeds, prompt_feat, prompt_speech_embeds_flow, spk_embeds);
+            std::string out = "finish";
+            if (out_callback_) out_callback_(out, true);
+
+        } catch (...) {
+            SLOGW("lLaMa_->Run have error!");
+        }
+    }
+
+    bool pause()
+    {
+        if (lLaMa_) lLaMa_->Stop();
+        return true;
+    }
+
+    bool delete_model()
+    {
+        if (tokenizer_pid_ != -1) {
+            kill(tokenizer_pid_, SIGTERM);
+            waitpid(tokenizer_pid_, nullptr, 0);
+            tokenizer_pid_ = -1;
+        }
+        lLaMa_->Deinit();
+        lLaMa_.reset();
+        return true;
+    }
+
+    static unsigned int getNextPort()
+    {
+        unsigned int port = next_port_++;
+        if (port > 8089) {
+            next_port_ = 8080;
+            port       = 8080;
+        }
+        return port;
+    }
+
+    llm_task(const std::string &workid) : tokenizer_server_flage_(false), port_(getNextPort())
+    {
+        inference_run_ = std::make_unique<std::thread>(std::bind(&llm_task::run, this));
+    }
+
+    void start()
+    {
+        if (!inference_run_) {
+            inference_run_ = std::make_unique<std::thread>(std::bind(&llm_task::run, this));
+        }
+    }
+
+    void stop()
+    {
+        if (inference_run_) {
+            std::string par;
+            async_list_.put(par);
+            if (lLaMa_) lLaMa_->Stop();
+            inference_run_->join();
+            inference_run_.reset();
+        }
+    }
+
+    ~llm_task()
+    {
+        stop();
+        if (tokenizer_pid_ != -1) {
+            kill(tokenizer_pid_, SIGTERM);
+            waitpid(tokenizer_pid_, nullptr, WNOHANG);
+        }
+        if (lLaMa_) {
+            lLaMa_->Deinit();
+        }
+    }
+};
+
+std::atomic<unsigned int> llm_task::next_port_{8080};
+
+#undef CONFIG_AUTO_SET
+
+class llm_cosy_voice : public StackFlow {
+private:
+    std::unordered_map<int, std::shared_ptr<llm_task>> llm_task_;
+
+public:
+    llm_cosy_voice() : StackFlow("cosy_voice")
+    {
+    }
+
+    void task_output(const std::weak_ptr<llm_task> llm_task_obj_weak,
+                     const std::weak_ptr<llm_channel_obj> llm_channel_weak, const std::string &data, bool finish)
+    {
+        auto llm_task_obj = llm_task_obj_weak.lock();
+        auto llm_channel  = llm_channel_weak.lock();
+        if (!(llm_task_obj && llm_channel)) {
+            return;
+        }
+        SLOGI("send:%s", data.c_str());
+        if (llm_channel->enstream_) {
+            static int count = 0;
+            nlohmann::json data_body;
+            data_body["index"] = count++;
+            data_body["delta"] = data;
+            if (!finish)
+                data_body["delta"] = data;
+            else
+                data_body["delta"] = std::string("");
+            data_body["finish"] = finish;
+            if (finish) count = 0;
+            SLOGI("send stream");
+            llm_channel->send(llm_task_obj->response_format_, data_body, LLM_NO_ERROR);
+        } else if (finish) {
+            SLOGI("send utf-8");
+            llm_channel->send(llm_task_obj->response_format_, data, LLM_NO_ERROR);
+        }
+    }
+
+    void task_pause(const std::weak_ptr<llm_task> llm_task_obj_weak,
+                    const std::weak_ptr<llm_channel_obj> llm_channel_weak)
+    {
+        auto llm_task_obj = llm_task_obj_weak.lock();
+        auto llm_channel  = llm_channel_weak.lock();
+        if (!(llm_task_obj && llm_channel)) {
+            return;
+        }
+        llm_task_obj->lLaMa_->Stop();
+    }
+
+    void pause(const std::string &work_id, const std::string &object, const std::string &data) override
+    {
+        SLOGI("llm_asr::work:%s", data.c_str());
+
+        nlohmann::json error_body;
+        int work_id_num = sample_get_work_id_num(work_id);
+        if (llm_task_.find(work_id_num) == llm_task_.end()) {
+            error_body["code"]    = -6;
+            error_body["message"] = "Unit Does Not Exist";
+            send("None", "None", error_body, work_id);
+            return;
+        }
+        task_pause(llm_task_[work_id_num], get_channel(work_id_num));
+        send("None", "None", LLM_NO_ERROR, work_id);
+    }
+
+    void task_user_data(const std::weak_ptr<llm_task> llm_task_obj_weak,
+                        const std::weak_ptr<llm_channel_obj> llm_channel_weak, const std::string &object,
+                        const std::string &data)
+    {
+        nlohmann::json error_body;
+        auto llm_task_obj = llm_task_obj_weak.lock();
+        auto llm_channel  = llm_channel_weak.lock();
+        if (!(llm_task_obj && llm_channel)) {
+            error_body["code"]    = -11;
+            error_body["message"] = "Model run failed.";
+            send("None", "None", error_body, unit_name_);
+            return;
+        }
+        const std::string *next_data = &data;
+        int ret;
+        std::string tmp_msg1;
+        if (object.find("stream") != std::string::npos) {
+            static std::unordered_map<int, std::string> stream_buff;
+            try {
+                if (decode_stream(data, tmp_msg1, stream_buff)) {
+                    return;
+                };
+            } catch (...) {
+                stream_buff.clear();
+                error_body["code"]    = -25;
+                error_body["message"] = "Stream data index error.";
+                send("None", "None", error_body, unit_name_);
+                return;
+            }
+            next_data = &tmp_msg1;
+        }
+        std::string tmp_msg2;
+        if (object.find("base64") != std::string::npos) {
+            ret = decode_base64((*next_data), tmp_msg2);
+            if (ret == -1) {
+                error_body["code"]    = -23;
+                error_body["message"] = "Base64 decoding error.";
+                send("None", "None", error_body, unit_name_);
+                return;
+            }
+            next_data = &tmp_msg2;
+        }
+        llm_task_obj->inference_async(sample_unescapeString(*next_data));
+    }
+
+    void task_asr_data(const std::weak_ptr<llm_task> llm_task_obj_weak,
+                       const std::weak_ptr<llm_channel_obj> llm_channel_weak, const std::string &object,
+                       const std::string &data)
+    {
+        auto llm_task_obj = llm_task_obj_weak.lock();
+        auto llm_channel  = llm_channel_weak.lock();
+        if (!(llm_task_obj && llm_channel)) {
+            return;
+        }
+        if (object.find("stream") != std::string::npos) {
+            if (sample_json_str_get(data, "finish") == "true") {
+                llm_task_obj->inference_async(sample_json_str_get(data, "delta"));
+            }
+        } else {
+            llm_task_obj->inference_async(data);
+        }
+    }
+
+    void kws_awake(const std::weak_ptr<llm_task> llm_task_obj_weak,
+                   const std::weak_ptr<llm_channel_obj> llm_channel_weak, const std::string &object,
+                   const std::string &data)
+    {
+        auto llm_task_obj = llm_task_obj_weak.lock();
+        auto llm_channel  = llm_channel_weak.lock();
+        if (!(llm_task_obj && llm_channel)) {
+            return;
+        }
+        llm_task_obj->lLaMa_->Stop();
+    }
+
+    int setup(const std::string &work_id, const std::string &object, const std::string &data) override
+    {
+        nlohmann::json error_body;
+        if ((llm_task_channel_.size() - 1) == MAX_TASK_NUM) {
+            error_body["code"]    = -21;
+            error_body["message"] = "task full";
+            send("None", "None", error_body, "llm");
+            return -1;
+        }
+
+        int work_id_num   = sample_get_work_id_num(work_id);
+        auto llm_channel  = get_channel(work_id);
+        auto llm_task_obj = std::make_shared<llm_task>(work_id);
+
+        nlohmann::json config_body;
+        try {
+            config_body = nlohmann::json::parse(data);
+        } catch (...) {
+            SLOGE("setup json format error.");
+            error_body["code"]    = -2;
+            error_body["message"] = "json format error.";
+            send("None", "None", error_body, "kws");
+            return -2;
+        }
+        int ret = llm_task_obj->load_model(config_body);
+        if (ret == 0) {
+            llm_channel->set_output(llm_task_obj->enoutput_);
+            llm_channel->set_stream(llm_task_obj->enstream_);
+
+            llm_task_obj->set_output(
+                std::bind(&llm_cosy_voice::task_output, this, std::weak_ptr<llm_task>(llm_task_obj),
+                          std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1, std::placeholders::_2));
+
+            for (const auto input : llm_task_obj->inputs_) {
+                if (input.find("llm") != std::string::npos) {
+                    llm_channel->subscriber_work_id(
+                        "", std::bind(&llm_cosy_voice::task_user_data, this, std::weak_ptr<llm_task>(llm_task_obj),
+                                      std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
+                                      std::placeholders::_2));
+                } else if ((input.find("asr") != std::string::npos) || (input.find("whisper") != std::string::npos)) {
+                    llm_channel->subscriber_work_id(
+                        input, std::bind(&llm_cosy_voice::task_asr_data, this, std::weak_ptr<llm_task>(llm_task_obj),
+                                         std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
+                                         std::placeholders::_2));
+                } else if (input.find("kws") != std::string::npos) {
+                    llm_channel->subscriber_work_id(
+                        input, std::bind(&llm_cosy_voice::kws_awake, this, std::weak_ptr<llm_task>(llm_task_obj),
+                                         std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
+                                         std::placeholders::_2));
+                }
+            }
+            llm_task_[work_id_num] = llm_task_obj;
+            SLOGI("load_mode success");
+            send("None", "None", LLM_NO_ERROR, work_id);
+            return 0;
+        } else {
+            SLOGE("load_mode Failed");
+            error_body["code"]    = -5;
+            error_body["message"] = "Model loading failed.";
+            send("None", "None", error_body, "llm");
+            return -1;
+        }
+    }
+
+    void link(const std::string &work_id, const std::string &object, const std::string &data) override
+    {
+        SLOGI("llm_llm::link:%s", data.c_str());
+        int ret = 1;
+        nlohmann::json error_body;
+        int work_id_num = sample_get_work_id_num(work_id);
+        if (llm_task_.find(work_id_num) == llm_task_.end()) {
+            error_body["code"]    = -6;
+            error_body["message"] = "Unit Does Not Exist";
+            send("None", "None", error_body, work_id);
+            return;
+        }
+        auto llm_channel  = get_channel(work_id);
+        auto llm_task_obj = llm_task_[work_id_num];
+        if (data.find("asr") != std::string::npos) {
+            ret = llm_channel->subscriber_work_id(
+                data,
+                std::bind(&llm_cosy_voice::task_asr_data, this, std::weak_ptr<llm_task>(llm_task_obj),
+                          std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1, std::placeholders::_2));
+            llm_task_obj->inputs_.push_back(data);
+        } else if (data.find("kws") != std::string::npos) {
+            ret = llm_channel->subscriber_work_id(
+                data,
+                std::bind(&llm_cosy_voice::kws_awake, this, std::weak_ptr<llm_task>(llm_task_obj),
+                          std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1, std::placeholders::_2));
+            llm_task_obj->inputs_.push_back(data);
+        }
+        if (ret) {
+            error_body["code"]    = -20;
+            error_body["message"] = "link false";
+            send("None", "None", error_body, work_id);
+            return;
+        } else {
+            send("None", "None", LLM_NO_ERROR, work_id);
+        }
+    }
+
+    void unlink(const std::string &work_id, const std::string &object, const std::string &data) override
+    {
+        SLOGI("llm_llm::unlink:%s", data.c_str());
+        int ret = 0;
+        nlohmann::json error_body;
+        int work_id_num = sample_get_work_id_num(work_id);
+        if (llm_task_.find(work_id_num) == llm_task_.end()) {
+            error_body["code"]    = -6;
+            error_body["message"] = "Unit Does Not Exist";
+            send("None", "None", error_body, work_id);
+            return;
+        }
+        auto llm_channel = get_channel(work_id);
+        llm_channel->stop_subscriber_work_id(data);
+        auto llm_task_obj = llm_task_[work_id_num];
+        for (auto it = llm_task_obj->inputs_.begin(); it != llm_task_obj->inputs_.end();) {
+            if (*it == data) {
+                it = llm_task_obj->inputs_.erase(it);
+            } else {
+                ++it;
+            }
+        }
+        send("None", "None", LLM_NO_ERROR, work_id);
+    }
+
+    void taskinfo(const std::string &work_id, const std::string &object, const std::string &data) override
+    {
+        SLOGI("llm_llm::taskinfo:%s", data.c_str());
+        // int ret = 0;
+        nlohmann::json req_body;
+        int work_id_num = sample_get_work_id_num(work_id);
+        if (WORK_ID_NONE == work_id_num) {
+            std::vector<std::string> task_list;
+            std::transform(llm_task_channel_.begin(), llm_task_channel_.end(), std::back_inserter(task_list),
+                           [](const auto task_channel) { return task_channel.second->work_id_; });
+            req_body = task_list;
+            send("llm.tasklist", req_body, LLM_NO_ERROR, work_id);
+        } else {
+            if (llm_task_.find(work_id_num) == llm_task_.end()) {
+                req_body["code"]    = -6;
+                req_body["message"] = "Unit Does Not Exist";
+                send("None", "None", req_body, work_id);
+                return;
+            }
+            auto llm_task_obj           = llm_task_[work_id_num];
+            req_body["model"]           = llm_task_obj->model_;
+            req_body["response_format"] = llm_task_obj->response_format_;
+            req_body["enoutput"]        = llm_task_obj->enoutput_;
+            req_body["inputs"]          = llm_task_obj->inputs_;
+            send("llm.taskinfo", req_body, LLM_NO_ERROR, work_id);
+        }
+    }
+
+    int exit(const std::string &work_id, const std::string &object, const std::string &data) override
+    {
+        SLOGI("llm_llm::exit:%s", data.c_str());
+
+        nlohmann::json error_body;
+        int work_id_num = sample_get_work_id_num(work_id);
+        if (llm_task_.find(work_id_num) == llm_task_.end()) {
+            error_body["code"]    = -6;
+            error_body["message"] = "Unit Does Not Exist";
+            send("None", "None", error_body, work_id);
+            return -1;
+        }
+        llm_task_[work_id_num]->stop();
+        auto llm_channel = get_channel(work_id_num);
+        llm_channel->stop_subscriber("");
+        llm_task_.erase(work_id_num);
+        send("None", "None", LLM_NO_ERROR, work_id);
+        return 0;
+    }
+
+    ~llm_cosy_voice()
+    {
+        while (1) {
+            auto iteam = llm_task_.begin();
+            if (iteam == llm_task_.end()) {
+                break;
+            }
+            iteam->second->stop();
+            get_channel(iteam->first)->stop_subscriber("");
+            iteam->second.reset();
+            llm_task_.erase(iteam->first);
+        }
+    }
+};
+
+int main(int argc, char *argv[])
+{
+    signal(SIGTERM, __sigint);
+    signal(SIGINT, __sigint);
+    mkdir("/tmp/llm", 0777);
+    llm_cosy_voice llm;
+    while (!main_exit_flage) {
+        sleep(1);
+    }
+    llm.llm_firework_exit();
+    return 0;
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp b/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
new file mode 100644
index 00000000..01a93e40
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
@@ -0,0 +1,705 @@
+#pragma once
+#include <iostream>
+#include <thread>
+#include <mutex>
+#include <condition_variable>
+#include <deque>
+#include <vector>
+#include <atomic>
+#include <string>
+#include <algorithm>
+#include <cmath>
+#include <numeric>
+#include "bfloat16.hpp"
+#include "Tokenizer/Tokenizer.hpp"
+#include "LLMEmbedSelector.hpp"
+#include "ax_model_runner/ax_model_runner_ax650.hpp"
+#include "ax_cmm_utils.hpp"
+#include "cqdm.h"
+#include "timer.hpp"
+#include "ax_sys_api.h"
+#include "utils/sampling.hpp"
+#include "utils/utils.hpp"
+
+using SpeechToken = int;
+// The container for speech tokens. std::deque is efficient for front/back operations.
+using TokenBuffer = std::deque<SpeechToken>;
+
+typedef std::function<void(int *, int, const char *, float, void *)> LLMRuningCallback;
+// typedef void (*LLMRuningCallback)(int *p_token, int n_token, float token_per_sec, void *reserve);
+
+struct LLMAttrType {
+    std::string template_filename_axmodel = "tinyllama-int8/tinyllama_l%d.axmodel";
+    int axmodel_num                       = 22;
+
+    std::string filename_post_axmodel = "tinyllama-int8/tinyllama_post.axmodel";
+    std::string filename_decoder_axmodel;
+    std::string token2wav_axmodel_dir;
+    std::string prompt_files;
+
+    int prefill_token_num     = 96;  // auto calc
+    int prefill_max_token_num = 512;
+    std::vector<int> prefill_max_kv_cache_num_grp;
+    int precompute_len = 0;
+    int prefill_grpid  = -1;
+
+    TokenizerType tokenizer_type         = TKT_HTTP;
+    std::string filename_tokenizer_model = "http://127.0.0.1:12345";
+    std::string url_tokenizer_model      = "http://127.0.0.1:12345";
+    bool b_bos = false, b_eos = false;
+    std::string filename_tokens_embed = "tinyllama.model.embed_tokens.weight.bfloat16.bin";
+    std::string filename_llm_embed    = "tinyllama.model.embed_tokens.weight.bfloat16.bin";
+    std::string filename_speech_embed = "tinyllama.model.embed_tokens.weight.bfloat16.bin";
+    int tokens_embed_num              = 151936;
+    int tokens_embed_size             = 896;
+
+    int llm_embed_num     = 2;
+    int llm_embed_size    = 896;
+    int speech_embed_num  = 6564;
+    int speech_embed_size = 896;
+
+    int max_token_len = 127;  // auto calc
+
+    int kv_cache_num  = 1024;  // auto calc
+    int kv_cache_size = 256;   // auto calc
+
+    bool b_use_mmap_load_embed        = false;
+    bool b_dynamic_load_axmodel_layer = false;
+
+    bool b_use_mmap_load_layer = true;
+
+    bool b_use_topk = false;
+
+    int n_timesteps;
+
+    // bool b_live_print = true;
+    LLMRuningCallback runing_callback = nullptr;
+    void *reserve                     = nullptr;
+};
+
+class LLM {
+private:
+    std::shared_ptr<BaseTokenizer> tokenizer;
+    LLaMaEmbedSelector embed_selector;
+    LLaMaEmbedSelector llm_embed_selector;
+    LLaMaEmbedSelector speech_embed_selector;
+
+    LLMAttrType _attr;
+
+    struct LLMLayer {
+        ax_runner_ax650 layer;
+        std::string filename;
+        MMap layer_buffer;
+        std::vector<char> layer_buffer_vec;
+    };
+
+    std::vector<LLMLayer> llama_layers;
+    ax_runner_ax650 llama_post;
+    ax_runner_ax650 llm_decoder;
+
+    // int prefill_grpid = 1;
+    int decode_grpid = 0;
+    bool b_stop      = false;
+    int min_len      = -1;
+    int max_len      = -1;
+
+public:
+    bool Init(LLMAttrType attr)
+    {
+        ALOGI("LLM init start");
+        t_cqdm cqdm = create_cqdm(attr.axmodel_num + 3, 32);
+        this->_attr = attr;
+        tokenizer   = CreateTokenizer(attr.tokenizer_type);
+        if (!tokenizer->Init(attr.filename_tokenizer_model, attr.b_bos, attr.b_eos)) {
+            ALOGE("tokenizer.Init(%s, %d, %d) failed", attr.filename_tokenizer_model.c_str(), attr.b_bos, attr.b_eos);
+            return false;
+        }
+        update_cqdm(&cqdm, 0, "count", "tokenizer init ok");
+
+        if (!embed_selector.Init(attr.filename_tokens_embed, attr.tokens_embed_num, attr.tokens_embed_size,
+                                 attr.b_use_mmap_load_embed)) {
+            ALOGE("embed_selector.Init(%s, %d, %d) failed", attr.filename_tokens_embed.c_str(), attr.tokens_embed_num,
+                  attr.tokens_embed_size);
+            return false;
+        }
+        if (!llm_embed_selector.Init(attr.filename_llm_embed, attr.llm_embed_num, attr.llm_embed_size,
+                                     attr.b_use_mmap_load_embed)) {
+            ALOGE("llm_embed_selector.Init(%s, %d, %d) failed", attr.filename_llm_embed.c_str(), attr.llm_embed_num,
+                  attr.llm_embed_size);
+            return false;
+        }
+        if (!speech_embed_selector.Init(attr.filename_speech_embed, attr.speech_embed_num, attr.speech_embed_size,
+                                        attr.b_use_mmap_load_embed)) {
+            ALOGE("speech_embed_selector.Init(%s, %d, %d) failed", attr.filename_tokens_embed.c_str(),
+                  attr.speech_embed_num, attr.speech_embed_size);
+            return false;
+        }
+        update_cqdm(&cqdm, 1, "count", "embed_selector init ok");
+
+        llama_layers.resize(attr.axmodel_num);
+        // prefill_layers.resize(attr.prefill_axmodel_num);
+        ALOGI("attr.axmodel_num:%d", attr.axmodel_num);
+        char axmodel_path[1024];
+        for (int i = 0; i < attr.axmodel_num; i++) {
+            sprintf(axmodel_path, attr.template_filename_axmodel.c_str(), i);
+            llama_layers[i].filename = axmodel_path;
+
+            if (!attr.b_dynamic_load_axmodel_layer) {
+                int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), false);
+                if (ret != 0) {
+                    ALOGE("init axmodel(%s) failed", llama_layers[i].filename.c_str());
+                    return false;
+                }
+                int remain_cmm = get_remaining_cmm_size();
+                sprintf(axmodel_path, "init %d axmodel ok,remain_cmm(%d MB)", i, remain_cmm);
+                update_cqdm(&cqdm, i + 2, "count", axmodel_path);
+            } else {
+                if (!attr.b_use_mmap_load_layer) {
+                    if (!read_file(llama_layers[i].filename, llama_layers[i].layer_buffer_vec)) {
+                        ALOGE("read_file(%s) failed", llama_layers[i].filename.c_str());
+                        return false;
+                    }
+                } else {
+                    llama_layers[i].layer_buffer.open_file(llama_layers[i].filename.c_str());
+                }
+
+                sprintf(axmodel_path, "read_file %s ok", llama_layers[i].filename.c_str());
+                update_cqdm(&cqdm, i + 2, "count", axmodel_path);
+            }
+        }
+
+        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), false);
+        if (ret != 0) {
+            ALOGE("init post axmodel(%s) failed", attr.filename_post_axmodel.c_str());
+            return false;
+        }
+        ret = llm_decoder.init(attr.filename_decoder_axmodel.c_str(), false);
+        if (ret != 0) {
+            ALOGE("init llm decoder axmodel(%s) failed", attr.filename_decoder_axmodel.c_str());
+            return false;
+        }
+
+        int remain_cmm = get_remaining_cmm_size();
+        sprintf(axmodel_path, "init post axmodel ok,remain_cmm(%d MB)", remain_cmm);
+        update_cqdm(&cqdm, attr.axmodel_num + 2, "count", axmodel_path);
+
+        if (attr.b_dynamic_load_axmodel_layer) {
+            // 加载第一层获取shape信息
+            auto &layer = llama_layers[0];
+            int ret;
+            if (_attr.b_use_mmap_load_layer) {
+                ret = layer.layer.init((char *)layer.layer_buffer.data(), layer.layer_buffer.size());
+            } else {
+                ret = layer.layer.init(layer.layer_buffer_vec.data(), layer.layer_buffer_vec.size());
+            }
+            if (ret != 0) {
+                ALOGE("init axmodel(%s) failed", layer.filename.c_str());
+            }
+        }
+
+        {
+            _attr.max_token_len = llama_layers[0].layer.get_input("mask").nSize / sizeof(unsigned short) - 1;
+            printf("\n");
+            ALOGI("max_token_len : %d", _attr.max_token_len);
+            // auto &input_k_cache = llama_layers[0].layer.get_input("K_cache");
+            // auto &output_k_cache_out = llama_layers[0].layer.get_output("K_cache_out");
+            _attr.kv_cache_size = llama_layers[0].layer.get_output("K_cache_out").nSize / sizeof(unsigned short);
+            _attr.kv_cache_num =
+                llama_layers[0].layer.get_input("K_cache").nSize / _attr.kv_cache_size / sizeof(unsigned short);
+            ALOGI("kv_cache_size : %d, kv_cache_num: %d", _attr.kv_cache_size, _attr.kv_cache_num);
+            if (_attr.max_token_len > _attr.kv_cache_num) {
+                ALOGE("max_token_len(%d) > kv_cache_num(%d)", _attr.max_token_len, _attr.kv_cache_num);
+                return false;
+            }
+
+            _attr.prefill_token_num = llama_layers[0].layer.get_input(1, "indices").vShape[1];
+            ALOGI("prefill_token_num : %d", _attr.prefill_token_num);
+            for (size_t i = 0; i < llama_layers[0].layer.get_num_input_groups() - 1; i++) {
+                int prefill_max_kv_cache_num = llama_layers[0].layer.get_input(i + 1, "K_cache").vShape[1];
+                ALOGI("grp: %ld, prefill_max_token_num : %d", i + 1, prefill_max_kv_cache_num);
+                _attr.prefill_max_kv_cache_num_grp.push_back(prefill_max_kv_cache_num);
+            }
+            _attr.prefill_max_token_num =
+                _attr.prefill_max_kv_cache_num_grp[_attr.prefill_max_kv_cache_num_grp.size() - 1];
+            ALOGI("prefill_max_token_num : %d", _attr.prefill_max_token_num);
+        }
+        if (attr.b_dynamic_load_axmodel_layer) {
+            for (int i = 0; i < attr.axmodel_num; i++) {
+                auto &layer = llama_layers[i];
+                layer.layer.deinit();
+            }
+        }
+
+        // Reset();
+        ALOGI("LLM init ok");
+        return true;
+    }
+
+    LLMAttrType *getAttr()
+    {
+        return &_attr;
+    }
+
+    void Deinit()
+    {
+        for (int i = 0; i < _attr.axmodel_num; i++) {
+            llama_layers[i].layer.release();
+        }
+        llama_post.release();
+        llm_decoder.release();
+        embed_selector.Deinit();
+        llm_embed_selector.Deinit();
+        speech_embed_selector.Deinit();
+    }
+
+    void Stop()
+    {
+        b_stop = true;
+    }
+
+    int TextToken2Embeds(std::vector<int> &token_ids, std::vector<unsigned short> &token_embeds)
+    {
+        if (token_embeds.empty() || token_embeds.size() < token_ids.size() * _attr.tokens_embed_size) {
+            token_embeds.resize(token_ids.size() * _attr.tokens_embed_size);
+        }
+
+        for (size_t i = 0; i < token_ids.size(); i++) {
+            embed_selector.getByIndex(token_ids[i], token_embeds.data() + i * _attr.tokens_embed_size);
+        }
+        return token_embeds.size();
+    }
+
+    int SpeechToken2Embeds(std::vector<int> &token_ids, std::vector<unsigned short> &token_embeds)
+    {
+        if (token_embeds.empty() || token_embeds.size() < token_ids.size() * _attr.speech_embed_size) {
+            token_embeds.resize(token_ids.size() * _attr.speech_embed_size);
+        }
+
+        for (size_t i = 0; i < token_ids.size(); i++) {
+            speech_embed_selector.getByIndex(token_ids[i], token_embeds.data() + i * _attr.speech_embed_size);
+        }
+        return token_embeds.size();
+    }
+
+    int Encode(std::vector<unsigned short> &out_embed, std::vector<std::vector<int>> &position_ids, std::string text,
+               const std::vector<unsigned short> &prompt_text_embeds, const std::vector<unsigned short> &prompt_speech_embeds)
+    {
+        // std::vector<int> prompt_ids = tokenizer->Encode(prompt_text, true);
+        ImageInfo img_info;
+        img_info.img_prompt       = false;
+        std::vector<int> text_ids = tokenizer->Encode(text, img_info);
+        int prompt_ids_size       = prompt_text_embeds.size() / _attr.tokens_embed_size;
+        int total_size = prompt_ids_size + text_ids.size() + 2 + prompt_speech_embeds.size() / _attr.speech_embed_size;
+        if (total_size > _attr.prefill_max_token_num) {
+            ALOGE("input embeding size(%d) > prefill_max_token_num(%d)", total_size, _attr.prefill_max_token_num);
+            return -1;
+        }
+        out_embed.resize(total_size * _attr.tokens_embed_size);
+
+        llm_embed_selector.getByIndex(0, out_embed.data() + 0 * _attr.tokens_embed_size);
+
+        // for (size_t i = 0; i < prompt_ids_size; i++)
+        // {
+        //     embed_selector.getByIndex(prompt_ids[i], out_embed.data() + (1+i) * _attr.tokens_embed_size);
+        // }
+        memcpy(out_embed.data() + _attr.tokens_embed_size, prompt_text_embeds.data(),
+               prompt_text_embeds.size() * sizeof(unsigned short));
+
+        for (size_t i = 0; i < text_ids.size(); i++) {
+            embed_selector.getByIndex(text_ids[i],
+                                      out_embed.data() + (1 + prompt_ids_size + i) * _attr.tokens_embed_size);
+        }
+
+        llm_embed_selector.getByIndex(
+            1, out_embed.data() + (1 + prompt_ids_size + text_ids.size()) * _attr.tokens_embed_size);
+
+        // for (size_t i = 0; i < prompt_speech_tokens.size(); i++)
+        // {
+        //     speech_embed_selector.getByIndex(prompt_speech_tokens[i], out_embed.data() +
+        //     (1+prompt_ids_size+text_ids.size()+1) * _attr.tokens_embed_size);
+        // }
+
+        memcpy(out_embed.data() + (1 + prompt_ids_size + text_ids.size() + 1) * _attr.tokens_embed_size,
+               prompt_speech_embeds.data(), prompt_speech_embeds.size() * sizeof(unsigned short));
+
+        std::vector<int> pos_ids;
+        for (size_t i = 0; i < total_size; i++) {
+            pos_ids.push_back(i);
+        }
+        position_ids.push_back(pos_ids);
+
+        min_len = text_ids.size() * 2;
+        max_len = text_ids.size() * 20;
+        return 0;
+    }
+
+    int Run(const std::string &input_str, const std::vector<unsigned short> &prompt_text_embeds,
+            const std::vector<unsigned short> &prompt_speech_embeds, TokenBuffer &token_buffer,
+            std::mutex &buffer_mutex, std::condition_variable &buffer_cv, std::atomic<bool> &llm_finished)
+    {
+        std::vector<unsigned short> text_embed;
+        std::vector<std::vector<int>> position_ids;
+        Encode(text_embed, position_ids, input_str, prompt_text_embeds, prompt_speech_embeds);
+        return Run(text_embed, position_ids, token_buffer, buffer_mutex, buffer_cv, llm_finished);
+    }
+
+    int Run(std::vector<unsigned short> &text_embed, std::vector<std::vector<int>> &position_ids,
+            TokenBuffer &token_buffer, std::mutex &buffer_mutex, std::condition_variable &buffer_cv,
+            std::atomic<bool> &llm_finished)
+    {
+        b_stop = false;
+        std::string final_out;
+
+        bfloat16 bf16 = -65536.f;
+        std::vector<unsigned short> mask(_attr.kv_cache_num + 1, bf16.data);
+        std::vector<unsigned short> embed(_attr.speech_embed_size, 0);
+
+        std::vector<int> cached_token;
+        std::vector<int> token_ids;
+
+        // std::vector<unsigned short> embed_tmp(_attr.prefill_token_num * _attr.tokens_embed_size, 0);
+        int input_embed_num   = text_embed.size() / _attr.tokens_embed_size;
+        int prefill_split_num = ceil((double)input_embed_num / _attr.prefill_token_num);
+        ALOGI("input token num : %d, prefill_split_num : %d", input_embed_num, prefill_split_num);
+        if (input_embed_num > _attr.prefill_max_token_num) {
+            ALOGE("input token num(%d) > prefill_max_token_num(%d)", input_embed_num, _attr.prefill_max_token_num);
+            return -1;
+        }
+
+        int kv_cache_num;
+        mask[_attr.kv_cache_num] = 0;
+        for (size_t i = 0; i < input_embed_num; i++) {
+            mask[i] = 0;
+        }
+        timer t_cost;
+        timer ttft_timer;
+        ttft_timer.start();
+
+        int max_pos_id = 0;
+        for (size_t p = 0; p < prefill_split_num; p++) {
+            if (b_stop) {
+                break;
+            }
+            _attr.prefill_grpid = p + 1;
+            kv_cache_num        = p * _attr.prefill_token_num;
+            std::vector<unsigned short> mask_tmp;
+            mask_tmp.resize(1 * _attr.prefill_token_num * (kv_cache_num + _attr.prefill_token_num), bf16.data);
+            int input_num_token = _attr.prefill_token_num;
+            if (p == prefill_split_num - 1) {
+                input_num_token = input_embed_num - p * _attr.prefill_token_num;
+            }
+
+            ALOGI("input_num_token:%d", input_num_token);
+            for (size_t i = 0; i < _attr.prefill_token_num; i++) {
+                if (i < input_num_token) {
+                    int mask_current_start = kv_cache_num;
+                    auto mask_ptr          = mask_tmp.data() + i * (kv_cache_num + _attr.prefill_token_num);
+
+                    for (int j = 0; j < _attr.precompute_len + p * _attr.prefill_token_num; j++) {
+                        mask_ptr[j] = 0;
+                    }
+
+                    for (int j = mask_current_start; j < mask_current_start + i + 1; j++) {
+                        mask_ptr[j] = 0;
+                    }
+                }
+            }
+
+            std::vector<unsigned short> embed_tmp(_attr.prefill_token_num * _attr.tokens_embed_size, 0);
+            if (p == (prefill_split_num - 1)) {
+                memcpy(
+                    embed_tmp.data(), text_embed.data() + p * _attr.prefill_token_num * _attr.tokens_embed_size,
+                    (input_embed_num - p * _attr.prefill_token_num) * _attr.tokens_embed_size * sizeof(unsigned short));
+            } else {
+                memcpy(embed_tmp.data(), text_embed.data() + p * _attr.prefill_token_num * _attr.tokens_embed_size,
+                       _attr.prefill_token_num * _attr.tokens_embed_size * sizeof(unsigned short));
+            }
+
+            for (unsigned int m = 0; m < _attr.axmodel_num; m++) {
+                if (b_stop) {
+                    break;
+                }
+
+                auto &layer = llama_layers[m];
+
+                if (_attr.b_dynamic_load_axmodel_layer) {
+                    int ret;
+                    if (_attr.b_use_mmap_load_layer) {
+                        ret = layer.layer.init((char *)layer.layer_buffer.data(), layer.layer_buffer.size());
+                    } else {
+                        ret = layer.layer.init(layer.layer_buffer_vec.data(), layer.layer_buffer_vec.size());
+                    }
+                    if (ret != 0) {
+                        ALOGE("init axmodel(%s) failed", layer.filename.c_str());
+                    }
+                }
+
+                // set indices
+                auto &input_indices             = layer.layer.get_input(_attr.prefill_grpid, "indices");
+                unsigned int *input_indices_ptr = (unsigned int *)input_indices.pVirAddr;
+                memset(input_indices_ptr, 0, input_indices.nSize);
+                // ALOGI("position_ids");
+                for (unsigned int i = 0; i < position_ids.size(); i++) {
+                    for (unsigned int j = _attr.precompute_len + p * _attr.prefill_token_num, jj = 0;
+                         j < _attr.precompute_len + (p + 1) * _attr.prefill_token_num; j++, jj++) {
+                        if (j < position_ids[i].size()) {
+                            input_indices_ptr[i * _attr.prefill_token_num + jj] = position_ids[i][j];
+                            if (position_ids[i][j] > max_pos_id) {
+                                max_pos_id = position_ids[i][j];
+                            }
+                        }
+                    }
+                }
+
+                // set mask
+                auto &input_mask = layer.layer.get_input(_attr.prefill_grpid, "mask");
+                memcpy((void *)input_mask.pVirAddr, (void *)mask_tmp.data(), mask_tmp.size() * sizeof(unsigned short));
+                // set input
+                auto &input_input = layer.layer.get_input(_attr.prefill_grpid, "input");
+                memcpy((void *)input_input.pVirAddr, embed_tmp.data(), embed_tmp.size() * sizeof(unsigned short));
+
+                layer.layer.inference(_attr.prefill_grpid);
+
+                auto &input_decoder_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
+                auto &input_decoder_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
+
+                auto &output_k_cache = layer.layer.get_output(_attr.prefill_grpid, "K_cache_out");
+                auto &output_v_cache = layer.layer.get_output(_attr.prefill_grpid, "V_cache_out");
+
+                int kv_offset = (_attr.precompute_len + p * _attr.prefill_token_num) * _attr.kv_cache_size;
+
+                memcpy((unsigned short *)input_decoder_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
+                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+
+                memcpy((unsigned short *)input_decoder_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
+                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+
+                for (int gid = _attr.prefill_grpid + 1; gid < prefill_split_num + 1; gid++) {
+                    auto &input_prefill_k_cache = layer.layer.get_input(gid, "K_cache");
+                    memcpy((unsigned short *)input_prefill_k_cache.pVirAddr + kv_offset,
+                           (void *)output_k_cache.pVirAddr,
+                           sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+                }
+
+                for (int gid = _attr.prefill_grpid + 1; gid < prefill_split_num + 1; gid++) {
+                    auto &input_prefill_v_cache = layer.layer.get_input(gid, "V_cache");
+                    memcpy((unsigned short *)input_prefill_v_cache.pVirAddr + kv_offset,
+                           (void *)output_v_cache.pVirAddr,
+                           sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+                }
+
+                auto &output = layer.layer.get_output(_attr.prefill_grpid, "output");
+                memcpy(embed_tmp.data(), (void *)output.pVirAddr, embed_tmp.size() * sizeof(unsigned short));
+
+                if (_attr.b_dynamic_load_axmodel_layer) {
+                    layer.layer.deinit();
+                }
+            }
+            if (p == (prefill_split_num - 1)) {
+                memcpy(embed.data(),
+                       embed_tmp.data() + (input_embed_num - p * _attr.prefill_token_num - 1) * _attr.tokens_embed_size,
+                       _attr.tokens_embed_size * sizeof(unsigned short));
+            }
+        }
+
+        int next_token = -1;
+        t_cqdm cqdm    = create_cqdm(_attr.max_token_len, 32);
+        int max_index;
+        std::vector<float> scores(_attr.speech_embed_num, 0.0f);
+        {
+            // post process
+            auto &input = llama_post.get_input(0);
+            memcpy((void *)input.pVirAddr, embed.data(), embed.size() * sizeof(unsigned short));
+            llama_post.inference();
+            if (_attr.b_use_topk) {
+                AX_SYS_MinvalidateCache(llama_post.get_output("indices").phyAddr,
+                                        llama_post.get_output("indices").pVirAddr,
+                                        llama_post.get_output("indices").nSize);
+                max_index = *(int *)llama_post.get_output("indices").pVirAddr;
+            } else {
+                auto &output_post        = llama_post.get_output("output_norm");  // 1 means get rmsnorm output
+                unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
+                std::vector<float> logits(output_post.nSize / sizeof(unsigned short));
+                for (int i = 0; i < output_post.nSize / sizeof(unsigned short); i++) {
+                    unsigned int proc = post_out[i] << 16;
+                    logits[i]         = *reinterpret_cast<float *>(&proc);
+                }
+
+                auto &input_decoder = llm_decoder.get_input(0);
+                memcpy(input_decoder.pVirAddr, logits.data(), logits.size() * sizeof(float));
+                llm_decoder.inference();
+
+                auto &output_decoder = llm_decoder.get_output(0);
+                float *post_decoder  = (float *)output_decoder.pVirAddr;
+
+                memcpy(scores.data(), post_decoder, output_decoder.nSize);
+                max_index = sampling::sampling_ids(scores, cached_token, _attr.speech_embed_num - 3, true);
+            }
+            next_token = max_index;
+
+            if (max_index >= _attr.speech_embed_num - 3) {
+                llm_finished = true;
+                buffer_cv.notify_all();
+                ALOGI("hit eos, llm finished");
+                return -1;
+            }
+
+            token_ids.push_back(max_index);
+            cached_token.push_back(max_index);
+            {
+                std::lock_guard<std::mutex> lock(buffer_mutex);
+                token_buffer.push_back(max_index);
+            }
+            buffer_cv.notify_one();
+            // ALOGI("token_buffer push %d", max_index);
+            ALOGI("ttft: %.2f ms", ttft_timer.cost());
+        }
+        t_cost.start();
+
+        bool b_hit_eos = false;
+
+        for (unsigned int indices = max_pos_id + 1; indices - max_pos_id < max_len; indices++) {
+            if (b_stop) {
+                break;
+            }
+
+            speech_embed_selector.getByIndex(next_token, embed.data());
+            memcpy((void *)llama_layers[0].layer.get_input(decode_grpid, "input").pVirAddr, embed.data(),
+                   llama_layers[0].layer.get_input(decode_grpid, "input").nSize);
+
+            {
+                std::vector<float> float_embeds(embed.size());
+                for (int i = 0; i < embed.size(); i++) {
+                    unsigned int proc = embed[i] << 16;
+                    float_embeds[i]   = *reinterpret_cast<float *>(&proc);
+                }
+            }
+
+            for (int m = 0; m < _attr.axmodel_num; m++) {
+                if (b_stop) {
+                    break;
+                }
+
+                auto &layer = llama_layers[m];
+
+                if (_attr.b_dynamic_load_axmodel_layer) {
+                    int ret;
+                    if (_attr.b_use_mmap_load_layer) {
+                        ret = layer.layer.init((char *)layer.layer_buffer.data(), layer.layer_buffer.size());
+                    } else {
+                        ret = layer.layer.init(layer.layer_buffer_vec.data(), layer.layer_buffer_vec.size());
+                    }
+                    if (ret != 0) {
+                        ALOGE("init axmodel(%s) failed", layer.filename.c_str());
+                    }
+                }
+
+                auto &input_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
+                auto &input_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
+
+                auto &input_indices = layer.layer.get_input(decode_grpid, "indices");
+                memcpy((void *)input_indices.pVirAddr, &indices, sizeof(indices));
+
+                auto &input_mask = layer.layer.get_input(decode_grpid, "mask");
+                memcpy((void *)input_mask.pVirAddr, mask.data(), mask.size() * sizeof(unsigned short));
+
+                layer.layer.inference(decode_grpid);
+
+                auto &output_k_cache = layer.layer.get_output(decode_grpid, "K_cache_out");
+                memcpy((unsigned short *)input_k_cache.pVirAddr + indices * _attr.kv_cache_size,
+                       (void *)output_k_cache.pVirAddr, output_k_cache.nSize);
+
+                auto &output_v_cache = layer.layer.get_output(decode_grpid, "V_cache_out");
+                memcpy((unsigned short *)input_v_cache.pVirAddr + indices * _attr.kv_cache_size,
+                       (void *)output_v_cache.pVirAddr, output_v_cache.nSize);
+
+                if (m == _attr.axmodel_num - 1) {
+                    memcpy((void *)llama_post.get_input(0).pVirAddr,
+                           (void *)layer.layer.get_output(decode_grpid, "output").pVirAddr,
+                           llama_post.get_input(0).nSize);
+                } else if (m < _attr.axmodel_num - 1) {
+                    memcpy((void *)llama_layers[m + 1].layer.get_input(decode_grpid, "input").pVirAddr,
+                           (void *)layer.layer.get_output(decode_grpid, "output").pVirAddr,
+                           layer.layer.get_input(decode_grpid, "input").nSize);
+                }
+            }
+
+            mask[indices] = 0;
+            {
+                llama_post.inference();
+
+                auto &output_post        = llama_post.get_output("output_norm");  // 1 means get rmsnorm output
+                unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
+                std::vector<float> logits(output_post.nSize / sizeof(unsigned short));
+                for (int i = 0; i < output_post.nSize / sizeof(unsigned short); i++) {
+                    unsigned int proc = post_out[i] << 16;
+                    logits[i]         = *reinterpret_cast<float *>(&proc);
+                }
+
+                auto &input_decoder = llm_decoder.get_input(0);
+                memcpy(input_decoder.pVirAddr, logits.data(), logits.size() * sizeof(float));
+                llm_decoder.inference();
+
+                auto &output_decoder = llm_decoder.get_output(0);
+                float *post_decoder  = (float *)output_decoder.pVirAddr;
+
+                memcpy(scores.data(), post_decoder, output_decoder.nSize);
+
+                bool ignore_eos = false;
+                if (indices < min_len) {
+                    ignore_eos = true;
+                }
+
+                max_index  = sampling::sampling_ids(scores, cached_token, _attr.speech_embed_num - 3, ignore_eos);
+                next_token = max_index;
+
+                if (max_index == _attr.speech_embed_num - 3) {
+                    b_hit_eos    = true;
+                    llm_finished = true;
+                    buffer_cv.notify_all();
+                    ALOGI("hit eos, llm finished");
+                    break;
+                }
+
+                if (max_index < _attr.speech_embed_num - 3) {
+                    token_ids.push_back(max_index);
+                    cached_token.push_back(max_index);
+                    {
+                        std::lock_guard<std::mutex> lock(buffer_mutex);
+                        token_buffer.push_back(max_index);
+                    }
+                    buffer_cv.notify_one();
+                    // ALOGI("token_buffer push %d", max_index);
+                }
+            }
+
+            if (b_hit_eos) {
+                llm_finished = true;
+                buffer_cv.notify_all();
+                ALOGI("hit eos, llm finished");
+                break;
+            }
+        }
+
+        llm_finished = true;
+        buffer_cv.notify_all();
+        ALOGI("llm finished");
+
+        printf("\n\n");
+        fflush(stdout);
+        float t_cost_ms = t_cost.cost();
+        ALOGI("total decode tokens:%d", cached_token.size());
+        ALOGN("hit eos, decode avg %.2f token/s\n", cached_token.size() / (t_cost_ms / 1000));
+
+        for (size_t i = 0; i < _attr.axmodel_num; i++) {
+            for (size_t j = 0; j < llama_layers[i].layer.get_num_input_groups(); j++) {
+                memset((void *)llama_layers[i].layer.get_input(j, "K_cache").pVirAddr, 0,
+                       llama_layers[i].layer.get_input(j, "K_cache").nSize);
+                memset((void *)llama_layers[i].layer.get_input(j, "V_cache").pVirAddr, 0,
+                       llama_layers[i].layer.get_input(j, "V_cache").nSize);
+            }
+        }
+
+        return cached_token.size();
+    }
+};
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/LLMEmbedSelector.hpp b/projects/llm_framework/main_cosy_voice/src/runner/LLMEmbedSelector.hpp
new file mode 100644
index 00000000..422e1166
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/LLMEmbedSelector.hpp
@@ -0,0 +1,110 @@
+#pragma once
+#include <string.h>
+#include <string>
+#include <vector>
+#include <fstream>
+
+#include "sample_log.h"
+
+#include "memory_utils.hpp"
+
+class LLaMaEmbedSelector
+{
+    MMap _embed_map;
+    std::vector<unsigned short> _embeds;
+    unsigned int _token_num, _embed_size;
+    bool _use_mmap = false;
+
+public:
+    bool Init(std::string embed_path, unsigned int token_num, unsigned int embed_size, bool use_mmap = false)
+    {
+        _token_num = token_num;
+        _embed_size = embed_size;
+        _use_mmap = use_mmap;
+        if (use_mmap)
+        {
+            ALOGI("LLaMaEmbedSelector use mmap");
+            if (!_embed_map.open_file(embed_path.c_str()))
+            {
+                ALOGE("embed file(%s) open failed", embed_path.c_str());
+                return false;
+            }
+        }
+        else
+        {
+            std::ifstream fin(embed_path);
+            if (!fin.is_open())
+            {
+                ALOGE("embed file(%s) open failed", embed_path.c_str());
+                return false;
+            }
+
+            // get file size
+            fin.seekg(0, std::ios::end);
+            int file_size = fin.tellg();
+            fin.seekg(0, std::ios::beg);
+            if (file_size != token_num * embed_size * 2)
+            {
+                ALOGE("embed file(%s) size(%d) not equal token_num(%d) * embed_size(%d) * 2", embed_path.c_str(), file_size, token_num, embed_size);
+                return false;
+            }
+
+            _embeds.resize(token_num * embed_size);
+            fin.read((char *)_embeds.data(), file_size);
+            fin.close();
+        }
+
+        return true;
+    }
+
+    void Deinit()
+    {
+        _embed_map.close_file();
+        _embeds.clear();
+    }
+
+    void getByIndex(unsigned int index, std::vector<unsigned short> &embed)
+    {
+        if (index >= _token_num)
+        {
+            ALOGE("index(%d) > token_num(%d)", index, _token_num);
+            return;
+        }
+        embed.resize(_embed_size);
+        if (_use_mmap)
+        {
+            unsigned short *ptr = (unsigned short *)_embed_map.data();
+            memcpy(embed.data(), ptr + index * _embed_size, _embed_size * sizeof(unsigned short));
+        }
+        else
+        {
+            memcpy(embed.data(), _embeds.data() + index * _embed_size, _embed_size * sizeof(unsigned short));
+        }
+    }
+
+    void getByIndex(unsigned int index, unsigned short *embed)
+    {
+        if (index >= _token_num)
+        {
+            ALOGE("index(%d) > token_num(%d)", index, _token_num);
+            return;
+        }
+        
+        if (_use_mmap)
+        {
+            unsigned short *ptr = (unsigned short *)_embed_map.data();
+            memcpy(embed, ptr + index * _embed_size, _embed_size * sizeof(unsigned short));
+        }
+        else
+        {
+            memcpy(embed, _embeds.data() + index * _embed_size, _embed_size * sizeof(unsigned short));
+        }
+    }
+
+    std::vector<unsigned short> getByIndex(unsigned int index)
+    {
+        std::vector<unsigned short> embed;
+        this->getByIndex(index, embed);
+        return embed;
+    }
+};
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/LLMPostprocess.hpp b/projects/llm_framework/main_cosy_voice/src/runner/LLMPostprocess.hpp
new file mode 100644
index 00000000..b7d50156
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/LLMPostprocess.hpp
@@ -0,0 +1,363 @@
+#pragma once
+#include <iostream>
+#include <vector>
+#include <random>
+#include <algorithm>
+#include <numeric>
+#include <cmath>
+#include <unordered_set>
+#include "utils/json.hpp"
+#include "utils/sample_log.h"
+
+class LLMPostprocess
+{
+private:
+    // 	控制随机性
+    void apply_temperature(std::vector<float> &logits, float temperature)
+    {
+        for (float &logit : logits)
+        {
+            logit /= temperature;
+        }
+    }
+
+    // 防止重复
+    void apply_repetition_penalty(std::vector<float> &logits, const std::vector<int> &history, float penalty)
+    {
+        for (int token : history)
+        {
+            if (token < logits.size())
+            {
+                logits[token] = logits[token] < 0 ? logits[token] * penalty : logits[token] / penalty;
+            }
+        }
+    }
+
+    void apply_repetition_penalty(std::vector<float> &logits,
+                                  const std::vector<int> &generated_tokens,
+                                  float repetition_penalty,
+                                  int penalty_window)
+    {
+        if (repetition_penalty == 1.0f || generated_tokens.empty())
+        {
+            return; // 如果 penalty = 1.0 或者没有生成 token，则不进行修改
+        }
+
+        int start_idx = std::max(0, (int)generated_tokens.size() - penalty_window);
+        std::unordered_set<int> recent_tokens(generated_tokens.begin() + start_idx, generated_tokens.end());
+
+        for (int token : recent_tokens)
+        {
+            if (token < 0 || token >= logits.size())
+                continue;
+
+            if (logits[token] > 0)
+            {
+                logits[token] /= std::sqrt(repetition_penalty);
+            }
+            else
+            {
+                logits[token] *= std::sqrt(repetition_penalty);
+            }
+        }
+    }
+
+    // 增强多样性
+    void apply_diversity_penalty(std::vector<float> &logits, const std::vector<int> &common_phrases, float penalty)
+    {
+        for (int token : common_phrases)
+        {
+            if (token < logits.size())
+            {
+                logits[token] *= penalty;
+            }
+        }
+    }
+
+    // Softmax function
+    std::vector<float> softmax(const std::vector<float> &logits)
+    {
+        std::vector<float> probs(logits.size());
+        float max_logit = *std::max_element(logits.begin(), logits.end());
+        float sum = 0.0f;
+
+        for (size_t i = 0; i < logits.size(); ++i)
+        {
+            probs[i] = std::exp(logits[i] - max_logit);
+            sum += probs[i];
+        }
+
+        for (float &p : probs)
+        {
+            p /= sum;
+        }
+
+        return probs;
+    }
+
+    // 	动态裁剪低概率 token
+    int faster_top_p_sampling(const std::vector<float> &logits, float top_p)
+    {
+        // 计算softmax
+        std::vector<float> probs = softmax(logits);
+
+        // 构建最大堆（概率和索引的配对）
+        std::vector<std::pair<float, size_t>> prob_index;
+        prob_index.reserve(logits.size());
+        for (size_t i = 0; i < logits.size(); ++i)
+        {
+            prob_index.emplace_back(probs[i], i);
+        }
+        auto cmp = [](const auto &a, const auto &b)
+        { return a.first < b.first; };
+        std::make_heap(prob_index.begin(), prob_index.end(), cmp);
+
+        // 提取top-p元素
+        std::vector<size_t> filtered_indices;
+        std::vector<float> filtered_probs;
+        float cumulative_prob = 0.0f;
+
+        while (!prob_index.empty() && cumulative_prob < top_p)
+        {
+            std::pop_heap(prob_index.begin(), prob_index.end(), cmp);
+            auto [prob, index] = prob_index.back();
+            prob_index.pop_back();
+
+            cumulative_prob += prob;
+            filtered_indices.push_back(index);
+            filtered_probs.push_back(prob);
+
+            if (cumulative_prob >= top_p)
+                break;
+        }
+
+        // 处理边缘情况（概率全零时返回第一个元素）
+        if (filtered_indices.empty())
+            return 0;
+
+        // 使用thread_local随机数生成器（线程安全）
+        static thread_local std::mt19937 gen(std::random_device{}());
+        std::discrete_distribution<int> dist(filtered_probs.begin(), filtered_probs.end());
+        return filtered_indices[dist(gen)];
+    }
+    int top_p_sampling(const std::vector<float> &logits, float top_p)
+    {
+        std::vector<float> probs = softmax(logits);
+
+        // Sort indices by probability in descending order
+        std::vector<size_t> indices(logits.size());
+        std::iota(indices.begin(), indices.end(), 0);
+        std::sort(indices.begin(), indices.end(), [&](size_t i, size_t j)
+                  { return probs[i] > probs[j]; });
+
+        // Compute cumulative probabilities
+        float cumulative_prob = 0.0f;
+        size_t cut_off = 0;
+        for (; cut_off < indices.size(); ++cut_off)
+        {
+            cumulative_prob += probs[indices[cut_off]];
+            if (cumulative_prob >= top_p)
+                break;
+        }
+
+        // Keep only the top-p probabilities
+        std::vector<size_t> filtered_indices(indices.begin(), indices.begin() + cut_off + 1);
+        std::vector<float> filtered_probs(filtered_indices.size());
+        for (size_t i = 0; i < filtered_indices.size(); ++i)
+        {
+            filtered_probs[i] = probs[filtered_indices[i]];
+        }
+
+        // Normalize the probabilities
+        float filtered_sum = std::accumulate(filtered_probs.begin(), filtered_probs.end(), 0.0f);
+        for (float &p : filtered_probs)
+        {
+            p /= filtered_sum;
+        }
+
+        // Sample from the filtered distribution
+        std::random_device rd;
+        std::mt19937 gen(rd());
+        std::discrete_distribution<int> dist(filtered_probs.begin(), filtered_probs.end());
+        return filtered_indices[dist(gen)];
+    }
+
+    // 限制候选 token 数
+    int top_k_sampling(const std::vector<float> &logits, int k)
+    {
+        // std::vector<float> probs = softmax(logits);
+
+        // 获取 top-k 索引
+        std::vector<size_t> indices(logits.size());
+        std::iota(indices.begin(), indices.end(), 0);
+        std::partial_sort(indices.begin(), indices.begin() + k, indices.end(), [&](size_t i, size_t j)
+                          { return logits[i] > logits[j]; });
+
+        // 仅保留 top-k 概率
+        std::vector<size_t> filtered_indices(indices.begin(), indices.begin() + k);
+        std::vector<float> filtered_probs(k);
+        for (size_t i = 0; i < k; ++i)
+        {
+            filtered_probs[i] = logits[filtered_indices[i]];
+        }
+        filtered_probs = softmax(filtered_probs);
+
+        // 归一化
+        float sum = std::accumulate(filtered_probs.begin(), filtered_probs.end(), 0.0f);
+        for (float &p : filtered_probs)
+        {
+            p /= sum;
+        }
+
+        // 采样
+        std::random_device rd;
+        std::mt19937 gen(rd());
+        std::discrete_distribution<int> dist(filtered_probs.begin(), filtered_probs.end());
+        return filtered_indices[dist(gen)];
+    }
+
+    bool enable_temperature = false;
+    float temperature = 1.0f;
+
+    bool enable_repetition_penalty = false;
+    float repetition_penalty = 1.0f;
+    int penalty_window = 20;
+
+    bool enable_diversity_penalty = false;
+    std::vector<int> common_phrases;
+    float diversity_penalty = 1.0f;
+
+    bool enable_top_p_sampling = false;
+    float top_p = 1.0f;
+
+    bool enable_top_k_sampling = false;
+    int top_k = 1;
+
+public:
+    LLMPostprocess() {}
+
+    void set_temperature(bool enable, float temperature)
+    {
+        enable_temperature = enable;
+        this->temperature = temperature;
+    }
+
+    void set_repetition_penalty(bool enable, float penalty, int penalty_window)
+    {
+        enable_repetition_penalty = enable;
+        this->repetition_penalty = penalty;
+        this->penalty_window = penalty_window;
+    }
+
+    void set_diversity_penalty(bool enable, const std::vector<int> &common_phrases, float penalty)
+    {
+        enable_diversity_penalty = enable;
+        this->common_phrases = common_phrases;
+        this->diversity_penalty = penalty;
+    }
+
+    void set_top_p_sampling(bool enable, float top_p)
+    {
+        enable_top_k_sampling = false;
+        enable_top_p_sampling = enable;
+        this->top_p = top_p;
+    }
+
+    void set_top_k_sampling(bool enable, int top_k)
+    {
+        enable_top_p_sampling = false;
+        enable_top_k_sampling = enable;
+        this->top_k = top_k;
+    }
+
+    bool load_config(std::string config_path)
+    {
+        std::ifstream config_file(config_path);
+        if (!config_file.is_open())
+        {
+            ALOGE("config file(%s) open failed", config_path.c_str());
+            return false;
+        }
+        nlohmann::json config = nlohmann::json::parse(config_file);
+        ALOGI("load config: \n%s\n", config.dump(4).c_str());
+
+        enable_temperature = config["enable_temperature"];
+        temperature = config["temperature"];
+
+        enable_repetition_penalty = config["enable_repetition_penalty"];
+        repetition_penalty = config["repetition_penalty"];
+        penalty_window = config["penalty_window"];
+
+        enable_top_p_sampling = config["enable_top_p_sampling"];
+        top_p = config["top_p"];
+
+        enable_top_k_sampling = config["enable_top_k_sampling"];
+        top_k = config["top_k"];
+        return true;
+    }
+
+    bool load_config(const nlohmann::json& config)
+    {
+        if (config.is_null()) {
+            ALOGE("config is null or invalid");
+            return false;
+        }
+
+        ALOGI("load config: \n%s\n", config.dump(4).c_str());
+
+        if (config.contains("enable_temperature")) {
+            enable_temperature = config["enable_temperature"].get<bool>();
+        }
+        if (config.contains("temperature")) {
+            temperature = config["temperature"].get<float>();
+        }
+
+        if (config.contains("enable_repetition_penalty")) {
+            enable_repetition_penalty = config["enable_repetition_penalty"].get<bool>();
+        }
+        if (config.contains("repetition_penalty")) {
+            repetition_penalty = config["repetition_penalty"].get<float>();
+        }
+        if (config.contains("penalty_window")) {
+            penalty_window = config["penalty_window"].get<int>();
+        }
+
+        if (config.contains("enable_top_p_sampling")) {
+            enable_top_p_sampling = config["enable_top_p_sampling"].get<bool>();
+        }
+        if (config.contains("top_p")) {
+            top_p = config["top_p"].get<float>();
+        }
+
+        if (config.contains("enable_top_k_sampling")) {
+            enable_top_k_sampling = config["enable_top_k_sampling"].get<bool>();
+        }
+        if (config.contains("top_k")) {
+            top_k = config["top_k"].get<int>();
+        }
+
+        return true;
+    }
+
+    int apply(std::vector<float> &logits, const std::vector<int> &history)
+    {
+        if (enable_temperature)
+            apply_temperature(logits, temperature);
+        if (enable_repetition_penalty)
+            apply_repetition_penalty(logits, history, repetition_penalty, penalty_window);
+        if (enable_diversity_penalty)
+            apply_diversity_penalty(logits, common_phrases, diversity_penalty);
+
+        if (enable_top_p_sampling)
+            return faster_top_p_sampling(logits, top_p);
+        else if (enable_top_k_sampling)
+            return top_k_sampling(logits, top_k);
+        else
+        {
+            // 最大值
+            float max_logit = *std::max_element(logits.begin(), logits.end());
+            int max_index = std::distance(logits.begin(), std::max_element(logits.begin(), logits.end()));
+            return max_index;
+        }
+    }
+};
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp b/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp
new file mode 100644
index 00000000..475d7440
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp
@@ -0,0 +1,635 @@
+#pragma once
+#include <string>
+#include <algorithm>
+#include <cmath>
+#include <numeric>
+#include <iostream>
+#include <unordered_map>
+#include <vector>
+#include <cstddef> // For size_t
+#include <stdexcept> // For std::invalid_argument
+#include "bfloat16.hpp"
+#include "Tokenizer/Tokenizer.hpp"
+#include "LLMEmbedSelector.hpp"
+#include "ax_model_runner/ax_model_runner_ax650.hpp"
+#include "utils/utils.hpp"
+#include "utils/slice_3d.h"
+#include "utils/concat_3d.h"
+#include "ax_cmm_utils.hpp"
+#include "cqdm.h"
+#include "timer.hpp"
+// #include "opencv2/opencv.hpp"
+#include "ax_sys_api.h"
+
+class Token2Wav
+{
+public:
+    int flow_embed_num = 6561;
+    int flow_embed_size = 512;
+    int token_mel_ratio = 2;
+    int token_hop_len = 25;
+    int max_infer_chunk_num = 3;
+    int mel_cache_len = 8;
+    int source_cache_len = mel_cache_len * 480;
+    int pre_lookahead_len = 3;
+    float inference_cfg_rate = 0.7;
+
+private:
+    ax_runner_ax650 flow_encoder_28;
+    ax_runner_ax650 flow_encoder_53;
+    ax_runner_ax650 flow_encoder_78;
+    ax_runner_ax650 flow_encoder_50_final;
+
+    ax_runner_ax650 flow_estimator_200;
+    ax_runner_ax650 flow_estimator_250;
+    ax_runner_ax650 flow_estimator_300;
+
+    ax_runner_ax650 hift_50_first;
+    ax_runner_ax650 hift_58;
+
+    std::vector<float> rand_noise;
+    std::vector<float> t_span;
+
+    LLaMaEmbedSelector flow_embed_selector;
+    
+    std::unordered_map<std::string, std::vector<float>> hift_cache_dict;
+    std::vector<float> speech_window; // np.hamming(2 * 8 * 480)
+
+    int init_noise(std::string model_dir)
+    {
+        return readtxt(model_dir+"/rand_noise_1_80_300.txt", rand_noise);
+    }
+
+    int init_speech_window(std::string model_dir)
+    {
+        return readtxt(model_dir+"/speech_window_2x8x480.txt", speech_window);
+    }
+
+    int init_tspan(int n_timesteps)
+    {
+        // std::vector<float> t_span_10 = {0.0000, 0.0123, 0.0489, 0.1090, 0.1910, 0.2929, 0.4122, 0.5460, 0.6910,0.8436, 1.0000};
+        // std::vector<float> t_span_7 = {0.0000, 0.1429, 0.2857, 0.4286, 0.5714, 0.7143, 0.8571, 1.0000}; // n_timesteps = 7
+        // std::vector<float> t_span_6 = {0.0000, 0.1667, 0.3333, 0.5000, 0.6667, 0.8333, 1.0000};       // n_timesteps = 6
+        // std::vector<float> t_span_5 = {0.0000, 0.2000, 0.4000, 0.6000, 0.8000, 1.0000 };
+        // std::vector<float> t_span_4 = {0.0000, 0.2500, 0.5000, 0.7500, 1.0000};
+
+        if(n_timesteps <4)
+        {
+            return -1;
+        }
+
+        n_timesteps = n_timesteps;
+        t_span = linspace(0.0, 1.0, n_timesteps + 1);
+        return 0;
+    }
+
+public:
+    bool Init(std::string model_dir, int n_timesteps)
+    {
+        int ret;
+
+        ret = init_tspan(n_timesteps);
+        if(ret != 0){
+            ALOGE("init_tspan failed, n_timesteps:%d", n_timesteps);
+            return false;
+        }
+
+        ret = init_noise(model_dir);
+        if(ret != 0){
+            ALOGE("init rand noise(%s) failed", "rand_noise_1_80_300.txt");
+            return false;
+        }
+
+        ret = init_speech_window(model_dir);
+        if(ret != 0){
+            ALOGE("init speech_window(%s) failed", "speech_window_2x8x480.txt");
+            return false;
+        }
+
+        if (!flow_embed_selector.Init((model_dir+"/flow.input_embedding.float16.bin").c_str(), flow_embed_num, flow_embed_size, false))
+        {
+            ALOGE("flow_embed_selector.Init(%s, %d, %d) failed", (model_dir+"/flow.input_embedding.float16.bin").c_str(),flow_embed_num, flow_embed_size);
+            return false;
+        }
+
+        ret = flow_encoder_28.init((model_dir+"/flow_encoder_28.axmodel").c_str(), false);
+        if (ret != 0)
+        {
+            ALOGE("init axmodel(%s) failed", (model_dir+"/flow_encoder_28.axmodel").c_str());
+            return false;
+        }
+
+        ret = flow_encoder_53.init((model_dir+"/flow_encoder_53.axmodel").c_str(), false);
+        if (ret != 0)
+        {
+            ALOGE("init axmodel(%s) failed", (model_dir+"/flow_encoder_53.axmodel").c_str());
+            return false;
+        }
+
+        ret = flow_encoder_78.init((model_dir+"/flow_encoder_78.axmodel").c_str(), false);
+        if (ret != 0)
+        {
+            ALOGE("init axmodel(%s) failed", (model_dir+"/flow_encoder_78.axmodel").c_str());
+            return false;
+        }
+
+        ret = flow_encoder_50_final.init((model_dir+"/flow_encoder_50_final.axmodel").c_str(), false);
+        if (ret != 0)
+        {
+            ALOGE("init axmodel(%s) failed", (model_dir+"/flow_encoder_50_final.axmodel").c_str());
+            return false;
+        }
+
+        ret = flow_estimator_200.init((model_dir+"/flow_estimator_200.axmodel").c_str(), false);
+        if (ret != 0)
+        {
+            ALOGE("init axmodel(%s) failed", (model_dir+"/flow_estimator_200.axmodel").c_str());
+            return false;
+        }
+
+        ret = flow_estimator_250.init((model_dir+"/flow_estimator_250.axmodel").c_str(), false);
+        if (ret != 0)
+        {
+            ALOGE("init axmodel(%s) failed", (model_dir+"/flow_estimator_250.axmodel").c_str());
+            return false;
+        }
+
+        ret = flow_estimator_300.init((model_dir+"/flow_estimator_300.axmodel").c_str(), false);
+        if (ret != 0)
+        {
+            ALOGE("init axmodel(%s) failed", (model_dir+"/flow_estimator_300.axmodel").c_str());
+            return false;
+        }
+
+        ret = hift_50_first.init((model_dir+"/hift_50_first.axmodel").c_str(), false);
+        if (ret != 0)
+        {
+            ALOGE("init axmodel(%s) failed", (model_dir+"/hift_50_first.axmodel").c_str());
+            return false;
+        }
+
+        ret = hift_58.init((model_dir+"/hift_58.axmodel").c_str(), false);
+        if (ret != 0)
+        {
+            ALOGE("init axmodel(%s) failed", (model_dir+"/hift_58.axmodel").c_str());
+            return false;
+        }
+
+        ALOGI("Token2Wav init ok");
+        return true;
+    }
+
+    void Deinit()
+    {
+        flow_encoder_28.release();
+        flow_encoder_53.release();
+        flow_encoder_78.release();
+        flow_encoder_50_final.release();
+        flow_estimator_200.release();
+        flow_estimator_250.release();
+        flow_estimator_300.release();
+        hift_50_first.release();
+        hift_58.release();
+        flow_embed_selector.Deinit();
+    }
+
+    int SpeechToken2Embeds(std::vector<int> & token_ids,  std::vector<float> &token_embeds)
+    {   
+        if(token_embeds.empty() || token_embeds.size() != token_ids.size()* flow_embed_size)
+        {
+            token_embeds.resize(token_ids.size()* flow_embed_size);
+        }
+        std::vector<unsigned short> speech_embeds_one(flow_embed_size);
+        for (size_t i = 0; i < token_ids.size(); i++)
+        {
+            flow_embed_selector.getByIndex(token_ids[i], speech_embeds_one.data());
+            for (int j = 0; j < flow_embed_size; j++)
+                {
+                    unsigned int proc = speech_embeds_one[j] << 16;
+                    token_embeds[i * flow_embed_size + j] = *reinterpret_cast<float *>(&proc);
+                }
+        }
+        return token_embeds.size();
+    }
+
+    int infer_flow_encoder(
+        std::vector<float> & token_embeds, std::vector<float> & prompt_feat, std::vector<float> & spk_embeds, int token_len, bool finalize,
+        std::vector<float> & mu, std::vector<float> & spks, std::vector<float> & cond
+        )
+    {
+        ax_runner_ax650 * model;
+        if(!finalize)
+        {
+            if(token_len == 28)
+            {
+                model = &flow_encoder_28;
+            }else if(token_len == 53)
+            {
+                model = &flow_encoder_53;
+            }else if(token_len == 78)
+            {
+                model = &flow_encoder_78;
+            }else{
+                return -1;
+            }
+        }else if(token_len == 50){
+            model = &flow_encoder_50_final;
+        }else{
+            return -1;
+        }
+
+        void * p = model->get_input("token_embedding").pVirAddr;
+        memcpy(p, token_embeds.data(), token_embeds.size() * sizeof(float));
+        p = model->get_input("prompt_feat").pVirAddr;
+        memcpy(p, prompt_feat.data(), prompt_feat.size() * sizeof(float));
+        p = model->get_input("embedding").pVirAddr;
+        memcpy(p, spk_embeds.data(), spk_embeds.size() * sizeof(float));
+
+        model->inference();
+
+        
+        auto &output_mu = model->get_output("mu");
+        if(mu.empty())
+        {
+            mu.resize(output_mu.nSize / sizeof(float));
+        }
+        memcpy(mu.data(), output_mu.pVirAddr, output_mu.nSize);
+
+        auto &output_spks = model->get_output("spks");
+        if(spks.empty())
+        {
+            spks.resize(output_spks.nSize / sizeof(float));
+        }
+        memcpy(spks.data(), output_spks.pVirAddr, output_spks.nSize);
+
+        auto &output_cond = model->get_output("cond");
+        if(cond.empty())
+        {
+            cond.resize(output_cond.nSize / sizeof(float));
+        }
+        memcpy(cond.data(), output_cond.pVirAddr, output_cond.nSize);
+
+        return 0;
+    }
+
+    int infer_flow_estimator(
+        std::vector<float> & x, std::vector<float> & mask, std::vector<float> & t,
+        std::vector<float> & mu, std::vector<float> & spks, std::vector<float> & cond, 
+        std::vector<float> & dphi_dt
+        )
+    {
+        ax_runner_ax650 * model;
+        int len = x.size()/(2*80);
+        if(len == 200){
+            model = &flow_estimator_200;
+        }else if(len == 250){
+            model = &flow_estimator_250;
+        }else if(len == 300){
+            model = &flow_estimator_300;
+        }else{
+            return -1;
+        }
+
+        void * p = model->get_input("x").pVirAddr;
+        memcpy(p, x.data(), x.size() * sizeof(float));
+        p = model->get_input("mask").pVirAddr;
+        memcpy(p, mask.data(), mask.size() * sizeof(float));
+        p = model->get_input("t").pVirAddr;
+        memcpy(p, t.data(), t.size() * sizeof(float));
+        p = model->get_input("mu").pVirAddr;
+        memcpy(p, mu.data(), mu.size() * sizeof(float));
+        p = model->get_input("spks").pVirAddr;
+        memcpy(p, spks.data(), spks.size() * sizeof(float));
+        p = model->get_input("cond").pVirAddr;
+        memcpy(p, cond.data(), cond.size() * sizeof(float));
+
+        model->inference();
+
+        auto &output_dphi_dt = model->get_output("y");
+        if(dphi_dt.empty() || dphi_dt.size() != output_dphi_dt.nSize / sizeof(float))
+        {
+            dphi_dt.resize(output_dphi_dt.nSize / sizeof(float));
+        }
+        memcpy(dphi_dt.data(), output_dphi_dt.pVirAddr, output_dphi_dt.nSize);
+
+        return 0;
+    }
+
+    int infer_hift(std::vector<float> &mel, std::vector<float> &cache_source, 
+                    std::vector<float> & tts_speech, std::vector<float> & tts_source)
+    {
+        ax_runner_ax650 * model;
+        int len = mel.size()/(80);
+        
+        if(len == 50 && cache_source.empty())
+        { 
+            model = &hift_50_first;
+        }else if(len == 58 && !cache_source.empty())
+        {
+            model = &hift_58;
+        }else
+        {
+            ALOGE("invalid size: %d", len);
+            return -1;
+        }
+
+        void * p = model->get_input("mel").pVirAddr;
+        memcpy(p, mel.data(), mel.size() * sizeof(float));
+        if(!cache_source.empty())
+        {
+            p = model->get_input("hift_cache_source").pVirAddr;
+            memcpy(p, cache_source.data(), cache_source.size() * sizeof(float));
+        }
+
+        model->inference();
+
+        auto &output_speech = model->get_output("audio");
+        if(tts_speech.empty() || tts_speech.size() != output_speech.nSize / sizeof(float))
+        {
+            tts_speech.resize(output_speech.nSize / sizeof(float));
+        }
+        memcpy(tts_speech.data(), output_speech.pVirAddr, output_speech.nSize);
+
+        auto &output_source = model->get_output("x");
+        if(tts_source.empty() || tts_source.size() != output_source.nSize / sizeof(float))
+        {
+            tts_source.resize(output_source.nSize / sizeof(float));
+        }
+        memcpy(tts_source.data(), output_source.pVirAddr, output_source.nSize);
+        
+        return 0;
+    }
+
+    int infer_flow_decoder_solve_euler(
+        std::vector<float> & x,  std::vector<float> & mu, std::vector<float> & spks, std::vector<float> & cond, std::vector<float> & mask,
+        std::vector<float> & mel
+    )
+    {
+        int len = mu.size()/80;
+
+        float t = t_span[0];
+        float dt = t_span[1] - t_span[0];
+
+        std::vector<float> x_in(2*80*len, 0);
+        std::vector<float> mask_in(2*1*len, 0);
+        std::vector<float> mu_in(2*80*len,0);
+        std::vector<float> t_in(2,0);
+        std::vector<float> spks_in(2*80, 0);
+        std::vector<float> cond_in(2*80*len, 0);
+        for(int step=1; step<t_span.size(); step++)
+        {   
+            memcpy(x_in.data(), x.data(), x.size() * sizeof(float));
+            memcpy(x_in.data()+x.size(),  x.data(), x.size() * sizeof(float));
+
+            memcpy(mask_in.data(), mask.data(), mask.size() * sizeof(float));
+            memcpy(mask_in.data()+mask.size(), mask.data(), mask.size() * sizeof(float));
+
+            memcpy(mu_in.data(), mu.data(), mu.size() * sizeof(float));
+
+            t_in[0] = t;
+            t_in[1] = t;
+
+            memcpy(spks_in.data(), spks.data(), spks.size() * sizeof(float));
+            memcpy(cond_in.data(), cond.data(), cond.size() * sizeof(float));
+
+            std::vector<float> dphi_dt;
+            int ret = infer_flow_estimator(x_in, mask_in, t_in, mu_in, spks_in, cond_in, dphi_dt);
+            if(ret != 0)
+            {
+                return ret;
+            }
+
+            for(int i=0; i<80*len; i++)
+            {
+
+                dphi_dt[i] = (1.0 + inference_cfg_rate) * dphi_dt[i] - inference_cfg_rate * dphi_dt[80 * len + i];
+                x[i] = x[i] + dt * dphi_dt[i];
+            }
+            
+            t = t + dt;
+
+            if(step < t_span.size()-1)
+            {
+                dt = t_span[step+1] - t;
+            }
+            else{
+                if(mel.empty() || mel.size()!=x.size())
+                {
+                    mel.resize(x.size());
+                }
+                memcpy(mel.data(), x.data(), x.size() * sizeof(float));
+            }
+
+        }
+
+        return 0;
+    }
+
+    int infer_flow_decoder(
+        std::vector<float> & mu, std::vector<float> & spks, std::vector<float> & cond, std::vector<float> & mask,
+        std::vector<float> & mel
+    )
+    {
+        std::vector<float> z;
+        z.insert(z.end(), rand_noise.begin(), rand_noise.begin() + mu.size());
+
+        int ret = infer_flow_decoder_solve_euler(z, mu, spks, cond, mask, mel);
+        return ret;
+    }
+
+    std::vector<float> infer_flow(
+        std::vector<float> & token_embeds, std::vector<float> & prompt_feat, std::vector<float> & spk_embeds, int token_len, bool finalize
+    )
+    {
+        int ret; 
+        int len;
+        std::vector<float> mu;
+        std::vector<float> spks;
+        std::vector<float> cond;
+
+        ret = infer_flow_encoder(token_embeds, prompt_feat, spk_embeds, token_len, finalize, mu, spks, cond);
+        if(ret != 0)
+        {
+            return std::vector<float>{};
+        }
+
+        len = mu.size()/80;
+
+        std::vector<float> mask(len, 1.0);
+        std::vector<float> all_mel;
+        
+        ret = infer_flow_decoder(mu, spks, cond, mask, all_mel);
+        if(ret != 0)
+        {
+            return std::vector<float>{};
+        }
+        
+        int len_mel1 = prompt_feat.size()/80;
+        int len_mel2 = all_mel.size()/80 - len_mel1;
+        
+        std::vector<float> mel(len_mel2 * 80, 0);
+        auto result = slice_3d_last_dim_from<float>(all_mel, 1, 80, all_mel.size()/80, len_mel1);
+        
+        return result;
+    }
+
+    void fade_in_out(std::vector<float>& fade_in_mel_data,
+                 const std::vector<float>& fade_out_mel_data,
+                 const std::vector<float>& window) {
+
+        // --- Constants based on window = np.hamming(2 * 8 * 480) ---
+        const size_t WINDOW_SIZE = 2 * 8 * 480; // 7680
+        const size_t MEL_OVERLAP_LEN = WINDOW_SIZE / 2; // 3840
+        // dim0 is implicitly 1 for both inputs
+        size_t dim1_in = fade_in_mel_data.size();
+        size_t dim1_out = fade_out_mel_data.size();
+        // --- Input Validation ---
+        // For 2D arrays [1, L], the 1D vector size is just L.
+        
+        if (window.size() != WINDOW_SIZE) {
+            throw std::invalid_argument("window size (" + std::to_string(window.size()) +
+                                        ") does not match expected size (7680).");
+        }
+        // Check if input arrays have enough elements for the overlap
+        if (dim1_in < MEL_OVERLAP_LEN) {
+            throw std::invalid_argument("fade_in_mel_data's column count (" + std::to_string(dim1_in) +
+                                        ") is smaller than mel_overlap_len (" + std::to_string(MEL_OVERLAP_LEN) + ").");
+        }
+        if (dim1_out < MEL_OVERLAP_LEN) {
+            throw std::invalid_argument("fade_out_mel_data's column count (" + std::to_string(dim1_out) +
+                                        ") is smaller than mel_overlap_len (" + std::to_string(MEL_OVERLAP_LEN) + ").");
+        }
+
+        // --- Perform Fade In/Out ---
+        // Since dim0=1, we only have one "row" to process.
+        // Iterate through the overlapping elements in the column dimension.
+        for (size_t i = 0; i < MEL_OVERLAP_LEN; ++i) {
+            // Indices are simply 'i' for the start of fade_in_mel
+            // and 'dim1_out - MEL_OVERLAP_LEN + i' for the end of fade_out_mel
+            const size_t in_idx = i;
+            const size_t out_idx = dim1_out - MEL_OVERLAP_LEN + i;
+
+            // Perform the weighted sum: result = in_val * win_in + out_val * win_out
+            // in_val = fade_in_mel_data[in_idx]
+            // out_val = fade_out_mel_data[out_idx]
+            // win_in = window[i]
+            // win_out = window[MEL_OVERLAP_LEN + i]
+            fade_in_mel_data[in_idx] = fade_in_mel_data[in_idx] * window[i] +
+                                    fade_out_mel_data[out_idx] * window[MEL_OVERLAP_LEN + i];
+        }
+        // fade_in_mel_data is now modified in-place with the faded result.
+    }
+
+    void reset()
+    {
+        std::unordered_map<std::string, std::vector<float>>().swap(hift_cache_dict);
+    }
+
+    std::vector<float> infer(std::vector<int> & text_speech_token, std::vector<float> & prompt_speech_embeds, std::vector<float> & prompt_feat,  
+                std::vector<float> & spk_embeds, int token_offset, bool finalize)
+    {
+        int ret = 0;
+        std::vector<float> speech_embeds( text_speech_token.size()*flow_embed_size + prompt_speech_embeds.size(), 0.0f);
+        std::vector<unsigned short> speech_embeds_one(flow_embed_size, 0);
+
+        memcpy(speech_embeds.data(), prompt_speech_embeds.data(), prompt_speech_embeds.size() * sizeof(float));
+
+        for (size_t i = 0; i < text_speech_token.size(); i++)
+        {
+            flow_embed_selector.getByIndex(text_speech_token[i], speech_embeds_one.data());
+
+            for (int j = 0; j < flow_embed_size; j++)
+                {
+                    unsigned int proc = speech_embeds_one[j] << 16;
+                    speech_embeds[prompt_speech_embeds.size() + i * flow_embed_size + j] = *reinterpret_cast<float *>(&proc);
+                }
+        }
+
+        std::vector<float> mel;
+    
+        mel = infer_flow(speech_embeds, prompt_feat, spk_embeds, text_speech_token.size(), finalize);   
+
+        std::vector<float> tts_mel;
+        int neg_offset=0, start;
+        if(finalize)
+        {
+            neg_offset = token_offset * token_mel_ratio - mel.size()/80;
+            start = - token_hop_len * token_mel_ratio;
+        }
+        else{
+            start = std::min( int(token_offset / token_hop_len), max_infer_chunk_num-1) * token_hop_len * token_mel_ratio;
+        }
+        
+        tts_mel = slice_3d_last_dim_from<float>(mel, 1, 80, mel.size()/80, start);
+        
+        std::vector<float> hift_cache_source;
+        std::vector<float> tts_mel1;
+        std::vector<float> speech, source, tts_speech;
+        if (!hift_cache_dict.empty())
+        {
+            auto hift_cache_mel = hift_cache_dict["mel"];
+            hift_cache_source = hift_cache_dict["source"];
+            tts_mel1 = concat_3d_dim2<float>(hift_cache_mel, 1, 80, hift_cache_mel.size()/80, tts_mel, 1, 80, tts_mel.size()/80);
+        }
+        else{
+            tts_mel1 = tts_mel;
+        }   
+        ret = infer_hift(tts_mel1, hift_cache_source, speech, source);
+        
+        if(ret != 0){
+            ALOGE("failed");
+            return std::vector<float>{};
+        }
+
+        if(!finalize)
+        {
+            
+            if(!hift_cache_dict.empty())
+            {
+                fade_in_out(speech, hift_cache_dict["speech"], speech_window);
+            }
+
+            hift_cache_dict["mel"] = slice_3d_last_dim_from<float>(tts_mel1, 1, 80, tts_mel1.size()/80, -mel_cache_len);
+
+            int offset = speech.size();
+            if(speech.size() > source_cache_len)
+            {
+                offset = source_cache_len;
+            }
+
+            hift_cache_dict["source"].assign(source.end()-offset, source.end());
+            hift_cache_dict["speech"].assign(speech.end()-offset, speech.end());
+            tts_speech.assign(speech.begin(), speech.end()-offset);
+
+        }
+        else{
+
+            if(speech.size() < source_cache_len){
+                tts_speech.assign(speech.begin(), speech.end());
+            }
+            else if (- neg_offset*480 >= source_cache_len)
+            {
+                tts_speech.assign(speech.end() + neg_offset*480, speech.end());
+
+                if(!hift_cache_dict.empty())
+                {
+                    fade_in_out(tts_speech, hift_cache_dict["speech"], speech_window);
+                }
+            }
+            else{
+                tts_speech.assign(speech.end()-source_cache_len, speech.end());
+
+                if(!hift_cache_dict.empty())
+                {
+                    fade_in_out(tts_speech, hift_cache_dict["speech"], speech_window);
+                }
+
+                int offset = speech.size() + neg_offset*480 - (speech.size() - source_cache_len);
+                tts_speech.assign(tts_speech.begin() + offset, tts_speech.end());
+            }
+            
+        }
+
+        return tts_speech;
+    }
+};
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/QwenTokenizer.cpp b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/QwenTokenizer.cpp
new file mode 100644
index 00000000..3b723ea1
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/QwenTokenizer.cpp
@@ -0,0 +1,127 @@
+#include "QwenTokenizer.hpp"
+
+#include <sstream>
+#include <fstream>
+#include <unordered_map>
+
+#include "sample_log.h"
+#include "base64.h"
+
+static const std::string PAT_STR = R"((?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?:$|[^\S])|\s+)";
+
+static std::pair<std::string, int> _parse(const std::string &line)
+{
+    auto pos = line.find(" ");
+    if (pos == std::string::npos)
+    {
+        throw std::runtime_error("invalid encoder line: " + line);
+    }
+
+    auto token = base64::decode({line.data(), pos});
+    int rank = 0;
+    try
+    {
+        rank = std::stoul(line.substr(pos + 1));
+    }
+    catch (const std::exception &)
+    {
+        throw std::runtime_error("invalid encoder rank: " + line);
+    }
+
+    return {std::move(token), rank};
+}
+
+QwenTokenizer::QwenTokenizer(const std::string &tiktoken_path, const QwenConfig &config)
+{
+    std::ifstream file(tiktoken_path);
+    if (!file)
+    {
+        throw std::runtime_error("failed to open encoder file: " + tiktoken_path);
+    }
+
+    ankerl::unordered_dense::map<std::string, int> encoder;
+    std::string line;
+    while (std::getline(file, line))
+    {
+        auto [token, rank] = _parse(line);
+
+        if (!encoder.emplace(std::move(token), rank).second)
+        {
+            throw std::runtime_error("duplicate item: " + line);
+        }
+    }
+
+    std::vector<std::string> special_tokens_s{"<|endoftext|>", "<|im_start|>", "<|im_end|>"};
+    char buffer[14];
+    for (size_t i = 0; i < 205; i++)
+    {
+        snprintf(buffer, 14, "<|extra_%zu|>", i);
+        special_tokens_s.push_back(buffer);
+    }
+    size_t encoder_size = encoder.size();
+    ankerl::unordered_dense::map<std::string, int> special_tokens;
+    special_tokens.reserve(special_tokens_s.size());
+    for (size_t i = 0; i < special_tokens_s.size(); i++)
+    {
+        special_tokens[special_tokens_s[i]] = encoder_size + i;
+    }
+
+    tokenizer = tiktoken::tiktoken(std::move(encoder), special_tokens, PAT_STR);
+    eos_token_id = config.eos_token_id;
+    im_start_id = config.im_start_id;
+    im_end_id = config.im_end_id;
+}
+
+auto QwenTokenizer::build_prompt(const std::vector<std::string> &history) const -> std::string
+{
+    if (history.size() % 2 == 1)
+    {
+        ALOGE("invalid history size %d", history.size());
+        return "";
+    }
+
+    std::ostringstream oss_prompt;
+    oss_prompt << "<|im_start|>system\nYou are a helpful assistant.<|im_end|>";
+    for (size_t i = 0; i < history.size() - 1; i += 2)
+    {
+        oss_prompt << "\n<|im_start|>user\n"
+                   << history[i] << "<|im_end|>\n<|im_start|>" << history[i + 1] << "<|im_end|>";
+    }
+    oss_prompt << "\n<|im_start|>user\n"
+               << history.back() << "<|im_end|>\n<|im_start|>assistant\n";
+
+    return oss_prompt.str();
+}
+
+auto QwenTokenizer::encode(const std::string &text, int max_length) const -> std::vector<int>
+{
+    auto ids = tokenizer.encode(text);
+    if ((int)ids.size() > max_length)
+    {
+        ids.erase(ids.begin(), ids.end() - max_length);
+    }
+    return ids;
+}
+
+auto QwenTokenizer::decode(const std::vector<int> &ids) const -> std::string
+{
+    std::vector<int> normal_ids(ids);
+    normal_ids.erase(std::remove_if(normal_ids.begin(), normal_ids.end(), [this](int id)
+                                    { return is_special_id(id); }),
+                     normal_ids.end());
+    auto text = tokenizer.decode(normal_ids);
+    return text;
+}
+
+auto QwenTokenizer::encode_history(
+    const std::vector<std::string> &history, int max_length) const -> std::vector<int>
+{
+    std::string prompt = build_prompt(history);
+    std::vector<int> input_ids = encode(prompt, max_length);
+    return input_ids;
+}
+
+auto QwenTokenizer::is_special_id(int id) const -> bool
+{
+    return id == eos_token_id || id == im_start_id || id == im_end_id;
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/QwenTokenizer.hpp b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/QwenTokenizer.hpp
new file mode 100644
index 00000000..a217f9c2
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/QwenTokenizer.hpp
@@ -0,0 +1,34 @@
+#pragma once
+#include <string>
+#include <vector>
+#include "tiktoken.h"
+
+struct QwenConfig
+{
+    // int pad_token_id;
+    // for tokenizer
+    int eos_token_id = 151643;
+    int im_start_id = 151644;
+    int im_end_id = 151645;
+};
+
+class QwenTokenizer
+{
+public:
+    QwenTokenizer(const std::string &tiktoken_path, const QwenConfig &config);
+
+    auto encode(const std::string &text, int max_length) const -> std::vector<int>;
+
+    auto decode(const std::vector<int> &ids) const -> std::string;
+
+    auto encode_history(const std::vector<std::string> &history, int max_length) const -> std::vector<int>;
+
+    auto build_prompt(const std::vector<std::string> &history) const -> std::string;
+
+    auto is_special_id(int id) const -> bool;
+
+    tiktoken::tiktoken tokenizer;
+    int eos_token_id;
+    int im_start_id;
+    int im_end_id;
+};
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp
new file mode 100644
index 00000000..f676efdd
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp
@@ -0,0 +1,912 @@
+#include "Tokenizer.hpp"
+
+#include "sentencepiece_processor.h"
+#include "builtin_pb/sentencepiece.pb.h"
+
+#include "QwenTokenizer.hpp"
+
+// #include "chatglm.h"
+
+#include "httplib.h"
+#include "http_utils.hpp"
+#include "json.hpp"
+
+#include "sample_log.h"
+#include "string_utility.hpp"
+#include "memory_utils.hpp"
+
+#include <iostream>
+#include <unistd.h>
+#include <fcntl.h>
+#include <sys/wait.h>
+#include <cstring>
+#include <csignal>
+
+class TokenizerLLaMa : public BaseTokenizer {
+protected:
+    sentencepiece::SentencePieceProcessor sp;
+    bool _b_bos, _b_eos;
+
+private:
+    /* data */
+public:
+    bool Init(std::string model_path, bool b_bos = true, bool b_eos = false) override
+    {
+        auto ret = sp.Load(model_path);
+        if (!ret.ok()) {
+            ALOGE("%s", ret.error_message());
+            return false;
+        }
+
+        this->_b_bos = b_bos;
+        this->_b_eos = b_eos;
+        return ret.ok();
+    }
+
+    bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) override
+    {
+        auto ret = sp.Encode(input, &output);
+        if (!ret.ok()) {
+            ALOGE("%s", ret.error_message());
+            return false;
+        }
+        if (_b_bos) {
+            output.insert(output.begin(), sp.bos_id());
+        }
+        if (_b_eos) {
+            output.push_back(sp.eos_id());
+        }
+        return true;
+    }
+
+    std::vector<int> Encode(std::string input, ImageInfo img_info) override
+    {
+        std::vector<int> output;
+        Encode(input, output, img_info);
+        return output;
+    }
+
+    std::string Decode(const std::vector<int> &input) override
+    {
+        sentencepiece::SentencePieceText spt;
+        sp.Decode(input, &spt);
+        std::string out = spt.pieces()[0].piece();
+        if (*(unsigned short *)out.data() == 38626) {
+            return " " + spt.text();
+        } else {
+            return spt.text();
+        }
+    }
+
+    int GetBosID() override
+    {
+        return sp.bos_id();
+    }
+
+    int GetEosID() override
+    {
+        return sp.eos_id();
+    }
+    std::string apply_chat_template() override
+    {
+        std::ostringstream oss_prompt;
+        int messages_len = messages_.size();
+        for (auto &message : messages_) {
+            messages_len--;
+            switch (message.first) {
+                case ROLE_USER: {
+                    oss_prompt << "<|user|>\n" << message.second << "</s>";
+                } break;
+                case ROLE_SYSTEM:
+                    break;
+                case ROLE_ASSISTANT:
+                    break;
+                case ROLE_ASSISTANT_HELP: {
+                    if (messages_len == 0) {
+                        oss_prompt << "<|assistant|>\n";
+                    }
+                } break;
+                default:
+                    break;
+            }
+        }
+        return oss_prompt.str();
+    }
+};
+
+class TokenizerMINICPM : public TokenizerLLaMa {
+public:
+    std::string Decode(const std::vector<int> &input) override
+    {
+        sentencepiece::SentencePieceText spt;
+        sp.Decode(input, &spt);
+        return spt.text();
+    }
+    std::string apply_chat_template() override
+    {
+        std::ostringstream oss_prompt;
+        int messages_len = messages_.size();
+        for (auto &message : messages_) {
+            messages_len--;
+            switch (message.first) {
+                case ROLE_USER: {
+                    oss_prompt << "<用户>" << message.second;
+                } break;
+                case ROLE_SYSTEM:
+                    break;
+                case ROLE_ASSISTANT:
+                    break;
+                case ROLE_ASSISTANT_HELP: {
+                    if (messages_len == 0) {
+                        oss_prompt << "<AI>";
+                    }
+                } break;
+                default:
+                    break;
+            }
+        }
+        return oss_prompt.str();
+    }
+};
+
+class TokenizerPhi3 : public BaseTokenizer {
+    sentencepiece::SentencePieceProcessor sp;
+    bool _b_bos, _b_eos;
+
+private:
+    /* data */
+public:
+    bool Init(std::string model_path, bool b_bos = true, bool b_eos = false) override
+    {
+        auto ret = sp.Load(model_path);
+        if (!ret.ok()) {
+            ALOGE("%s", ret.error_message());
+            return false;
+        }
+
+        this->_b_bos = b_bos;
+        this->_b_eos = b_eos;
+        return ret.ok();
+    }
+
+    bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) override
+    {
+        auto ret = sp.Encode(input, &output);
+        if (!ret.ok()) {
+            ALOGE("%s", ret.error_message());
+            return false;
+        }
+        output.insert(output.begin(), 32010);  //"<|user|>"
+        output.push_back(32007);               //"<|end|>"
+        output.push_back(32001);               //"<|assistant|>"
+        if (_b_bos) {
+            output.insert(output.begin(), sp.bos_id());
+        }
+        if (_b_eos) {
+            output.push_back(sp.eos_id());
+        }
+        return true;
+    }
+
+    std::vector<int> Encode(std::string input, ImageInfo img_info) override
+    {
+        std::vector<int> output;
+        Encode(input, output, img_info);
+        return output;
+    }
+
+    std::string Decode(const std::vector<int> &input) override
+    {
+        sentencepiece::SentencePieceText spt;
+        sp.Decode(input, &spt);
+        std::string out = spt.pieces()[0].piece();
+        if (*(unsigned short *)out.data() == 38626) {
+            return " " + spt.text();
+        } else {
+            return spt.text();
+        }
+    }
+
+    int GetBosID() override
+    {
+        return sp.bos_id();
+    }
+
+    int GetEosID() override
+    {
+        return 32007;
+    }
+
+    bool isEnd(int id) override
+    {
+        return id == GetEosID() || id > 31999;
+    }
+
+    std::string apply_chat_template() override
+    {
+        std::ostringstream oss_prompt;
+        int messages_len = messages_.size();
+        for (auto &message : messages_) {
+            messages_len--;
+            switch (message.first) {
+                case ROLE_USER: {
+                    oss_prompt << message.second;
+                } break;
+                case ROLE_SYSTEM:
+                    break;
+                case ROLE_ASSISTANT:
+                    break;
+                case ROLE_ASSISTANT_HELP: {
+                    if (messages_len == 0) {
+                        oss_prompt << " ";
+                    }
+                } break;
+                default:
+                    break;
+            }
+        }
+        return oss_prompt.str();
+    }
+};
+
+class TokenizerQwen : public BaseTokenizer {
+    std::shared_ptr<QwenTokenizer> sp;
+    bool _b_bos, _b_eos;
+
+private:
+    /* data */
+public:
+    bool Init(std::string model_path, bool b_bos = true, bool b_eos = false) override
+    {
+        if (!file_exist(model_path)) {
+            ALOGE("tokenizer model file(%s) not exist", model_path.c_str());
+            return false;
+        }
+
+        sp.reset(new QwenTokenizer(model_path, QwenConfig()));
+
+        this->_b_bos = b_bos;
+        this->_b_eos = b_eos;
+        return true;
+    }
+
+    bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) override
+    {
+        if (_b_bos) {
+            // input += "<|im_start|>";
+        }
+        if (_b_eos) {
+            input += "<|endoftext|>";
+        }
+        output = sp->encode(input, 1024);
+
+        return true;
+    }
+
+    std::vector<int> Encode(std::string input, ImageInfo img_info) override
+    {
+        std::vector<int> output;
+        Encode(input, output, img_info);
+        return output;
+    }
+
+    std::string Decode(const std::vector<int> &input) override
+    {
+        return sp->decode(input);
+    }
+
+    int GetBosID() override
+    {
+        return -1;
+    }
+
+    int GetEosID() override
+    {
+        return sp->eos_token_id;
+    }
+    std::string apply_chat_template() override
+    {
+        std::ostringstream oss_prompt;
+        int messages_len = messages_.size();
+        for (auto &message : messages_) {
+            messages_len--;
+            switch (message.first) {
+                case ROLE_USER: {
+                    oss_prompt << "<|im_start|>user\n" << message.second << "<|im_end|>\n";
+                } break;
+                case ROLE_SYSTEM: {
+                    oss_prompt << "<|im_start|>system\n" << message.second << ".<|im_end|>\n";
+                } break;
+                case ROLE_ASSISTANT: {
+                    oss_prompt << "<|im_start|>assistant\n" << message.second << ".<|im_end|>\n";
+                } break;
+                case ROLE_ASSISTANT_HELP: {
+                    if (messages_len == 0) {
+                        oss_prompt << "<|im_start|>assistant\n";
+                    }
+                } break;
+                default:
+                    break;
+            }
+        }
+        return oss_prompt.str();
+    }
+};
+
+class Tokenizer_Http : public BaseTokenizer {
+    std::shared_ptr<httplib::Client> cli;
+    bool _b_bos, _b_eos;
+
+    std::string base_url;
+
+    int bos_id, eos_id;
+
+    std::string uid;
+
+private:
+    /* data */
+public:
+    bool Init(std::string model_path = "http://localhost:8080", bool b_bos = true, bool b_eos = false) override
+    {
+        base_url = model_path;
+        try {
+            cli = std::make_shared<httplib::Client>(base_url);
+            cli->set_connection_timeout(1);
+            cli->set_read_timeout(1);
+            cli->set_write_timeout(1);
+            {
+                auto ret = cli->Get("/bos_id");
+                auto rep = ret.value();
+                if (rep.status != 200) {
+                    ALOGE("get bos_id failed, status: %d", rep.status);
+                    return false;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                bos_id           = j["bos_id"];
+            }
+
+            {
+                auto ret = cli->Get("/eos_id");
+                auto rep = ret.value();
+                if (rep.status != 200) {
+                    ALOGE("get eos_id failed, status: %d", rep.status);
+                    return false;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                eos_id           = j["eos_id"];
+            }
+            printf("bos_id: %d, eos_id: %d\n", bos_id, eos_id);
+        } catch (const std::exception &e) {
+            std::cerr << e.what() << '\n';
+            return false;
+        }
+
+        this->_b_bos = b_bos;
+        this->_b_eos = b_eos;
+        return true;
+    }
+
+    bool Init(std::string model_path = "http://localhost:8080") override
+    {
+        base_url = model_path;
+        if (!test_connect_http(base_url, 10)) {
+            ALOGE("connect %s failed", base_url.c_str());
+            return false;
+        } else {
+            ALOGI("connect %s ok", base_url.c_str());
+        }
+
+        cli = std::make_shared<httplib::Client>(base_url);
+        cli->set_connection_timeout(10);
+        cli->set_read_timeout(10);
+        cli->set_write_timeout(10);
+
+        int try_count = 10;
+        int count     = try_count;
+        while (count-- > 0) {
+            try {
+                auto ret = cli->Get("/get_uid");
+                auto rep = ret.value();
+                if (rep.status != 200) {
+                    ALOGE("get uid failed, status: %d", rep.status);
+                    return false;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                uid              = j["uid"];
+                ALOGI("uid: %s", uid.c_str());
+                break;
+            } catch (const std::exception &e) {
+                std::cerr << e.what() << '\n';
+            }
+            std::this_thread::sleep_for(std::chrono::seconds(1));
+            ALOGE("get uid failed, try again %d/%d", count, try_count);
+        }
+
+        count = 10;
+        while (count-- > 0) {
+            try {
+                auto ret = cli->Get("/bos_id?uid=" + uid);
+                auto rep = ret.value();
+                if (rep.status != 200) {
+                    ALOGE("get bos_id failed, status: %d", rep.status);
+                    return false;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                bos_id           = j["bos_id"];
+                break;
+            } catch (const std::exception &e) {
+                std::cerr << e.what() << '\n';
+            }
+            std::this_thread::sleep_for(std::chrono::seconds(1));
+            ALOGE("get bos_id failed, try again %d/%d", count, try_count);
+        }
+
+        count = 10;
+        while (count-- > 0) {
+            try {
+                auto ret = cli->Get("/eos_id?uid=" + uid);
+                auto rep = ret.value();
+                if (rep.status != 200) {
+                    ALOGE("get eos_id failed, status: %d", rep.status);
+                    return false;
+                }
+                nlohmann::json j = nlohmann::json::parse(rep.body);
+                eos_id           = j["eos_id"];
+                break;
+            } catch (const std::exception &e) {
+                std::cerr << e.what() << '\n';
+            }
+            std::this_thread::sleep_for(std::chrono::seconds(1));
+            ALOGE("get eos_id failed, try again %d/%d", count, try_count);
+        }
+
+        printf("bos_id: %d, eos_id: %d\n", bos_id, eos_id);
+
+        return true;
+    }
+
+    bool Encode(std::string input, std::string last_reply, std::vector<int> &tokens, std::vector<int> &tokens_diff,
+                ImageInfo img_info) override
+    {
+        nlohmann::json j;
+        j["uid"]  = uid;
+        j["text"] = input;
+        if (!last_reply.empty() and last_reply != "") {
+            j["last_reply"] = last_reply;
+        }
+
+        j["img_prompt"] = img_info.img_prompt;
+        auto ret        = cli->Post("/encode", j.dump(), "application/json");
+        auto rep        = ret.value();
+        if (rep.status != 200) {
+            ALOGE("encode failed, status: %d", rep.status);
+            return false;
+        }
+        nlohmann::json j2;
+        try {
+            j2 = nlohmann::json::parse(rep.body);
+        } catch (const std::exception &e) {
+            ALOGE("json parse failed: %s", e.what());
+            ALOGE("%s", rep.body.c_str());
+            return false;
+        }
+
+        std::vector<int> _token_ids   = j2["token_ids"];
+        std::vector<int> _tokens_diff = j2["diff"];
+
+        tokens      = _token_ids;
+        tokens_diff = _tokens_diff;
+
+        return true;
+    }
+
+    bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) override
+    {
+        nlohmann::json j;
+        j["text"]       = input;
+        j["img_prompt"] = img_info.img_prompt;
+        auto ret        = cli->Post("/encode", j.dump(), "application/json");
+        auto rep        = ret.value();
+        if (rep.status != 200) {
+            ALOGE("encode failed, status: %d", rep.status);
+            return false;
+        }
+        nlohmann::json j2;
+        try {
+            j2 = nlohmann::json::parse(rep.body);
+        } catch (const std::exception &e) {
+            ALOGE("json parse failed: %s", e.what());
+            ALOGE("%s", rep.body.c_str());
+            return false;
+        }
+
+        std::vector<int> out = j2["token_ids"];
+        output               = out;
+
+        if (_b_bos) {
+            output.insert(output.begin(), bos_id);
+        }
+        if (_b_eos) {
+            output.push_back(eos_id);
+        }
+
+        return true;
+    }
+
+    bool Reset(std::string system_prompt, std::vector<int> &tokens) override
+    {
+        nlohmann::json j;
+        j["uid"] = uid;
+        if (!system_prompt.empty() and system_prompt != "") {
+            j["system_prompt"] = system_prompt;
+        }
+
+        auto ret = cli->Post("/reset", j.dump(), "application/json");
+        auto rep = ret.value();
+        if (rep.status != 200) {
+            ALOGE("reset failed, status: %d", rep.status);
+            return false;
+        }
+        nlohmann::json j_rep        = nlohmann::json::parse(rep.body);
+        std::vector<int> _token_ids = j_rep["token_ids"];
+        tokens                      = _token_ids;
+        return true;
+    }
+
+    std::vector<int> Encode(std::string input, ImageInfo img_info) override
+    {
+        std::vector<int> output;
+        Encode(input, output, img_info);
+        return output;
+    }
+
+    std::string Decode(const std::vector<int> &input) override
+    {
+        int cnt             = 2;
+        std::string out_str = "";
+        while (cnt--) {
+            nlohmann::json j;
+            j["token_ids"] = input;
+            j["uid"]       = uid;
+            auto ret       = cli->Post("/decode", j.dump(), "application/json");
+            auto rep       = ret.value();
+            if (rep.status != 200) {
+                ALOGE("decode failed, status: %d, try again", rep.status);
+                ALOGE("%s", rep.body.c_str());
+                usleep(1000 * 1000);
+                continue;
+            }
+            try {
+                nlohmann::json j2 = nlohmann::json::parse(rep.body);
+                out_str           = j2["text"];
+                break;
+            } catch (const std::exception &e) {
+                ALOGE("json parse failed: %s, try again", e.what());
+                ALOGE("%s", rep.body.c_str());
+                usleep(1000 * 1000);
+                continue;
+            }
+        }
+        return out_str;
+    }
+
+    int GetBosID() override
+    {
+        return bos_id;
+    }
+
+    int GetEosID() override
+    {
+        return eos_id;
+    }
+
+    std::string apply_chat_template() override
+    {
+        std::ostringstream oss_prompt;
+        int messages_len = messages_.size();
+        for (auto &message : messages_) {
+            messages_len--;
+            switch (message.first) {
+                case ROLE_USER: {
+                    oss_prompt << message.second;
+                } break;
+                case ROLE_SYSTEM: {
+                } break;
+                case ROLE_ASSISTANT: {
+                } break;
+                case ROLE_ASSISTANT_HELP: {
+                } break;
+                default:
+                    break;
+            }
+        }
+        return oss_prompt.str();
+    }
+};
+
+class ResultObj {
+private:
+    /* data */
+public:
+    bool success;
+    nlohmann::json result;
+    ResultObj(bool _success, const nlohmann::json &_result)
+    {
+        success = _success;
+        result  = _result;
+    }
+};
+
+class Tokenizer_Auto : public BaseTokenizer {
+    bool _b_bos, _b_eos;
+    std::string base_url;
+    int bos_id, eos_id;
+    pid_t pid       = 0;
+    int pipe_c2p[2] = {0};  // C++ -> Python
+    int pipe_p2c[2] = {0};  // Python -> C++
+private:
+    ResultObj readPython(int id, int timeout = 10000)
+    {
+        std::string message;
+        message.reserve(1056);
+        static char buffer[1056];
+        while (timeout > 0) {
+            ssize_t bytesRead = read(pipe_p2c[0], buffer, sizeof(buffer) - 1);
+            if (bytesRead > 0) {
+                for (int i = 0; i < bytesRead; i++) {
+                    message += buffer[i];
+                    if (buffer[i] == '\n') {
+                        ALOGI("readPython:%s", message.c_str());
+                        int pos = message.size() - 12;
+                        if ((message.size() > 12) && (message.substr(pos) == "rpccontinue\n")) {
+                            message.erase(pos);
+                            continue;
+                        }
+                        nlohmann::json j2;
+                        try {
+                            j2 = nlohmann::json::parse(message);
+                            if (j2["id"].is_number_integer() && j2["id"] == id) {
+                                return ResultObj(true, j2["result"]);
+                            } else if (j2["id"].is_null()) {
+                                std::string errormesg;
+                                if (j2["result"].is_string()) {
+                                    errormesg = j2["result"];
+                                }
+                                ALOGE("python runtime error: %s", errormesg.c_str());
+                                return ResultObj(false, j2);
+                            }
+                        } catch (const nlohmann::json::parse_error &e) {
+                            ALOGE("JSON parse error: %s", e.what());
+                            return ResultObj(false, "");
+                        } catch (const nlohmann::json::type_error &e) {
+                            ALOGE("JSON type error: %s", e.what());
+                            return ResultObj(false, "");
+                        } catch (const nlohmann::json::out_of_range &e) {
+                            ALOGE("JSON out of range error: %s", e.what());
+                            return ResultObj(false, "");
+                        } catch (const std::exception &e) {
+                            ALOGE("Standard exception: %s", e.what());
+                            return ResultObj(false, "");
+                        } catch (...) {
+                            ALOGE("Unknown error occurred while parsing JSON");
+                            return ResultObj(false, "");
+                        }
+                    }
+                }
+            } else {
+                break;
+            }
+            usleep(10000);
+            timeout -= 10;
+        }
+        return ResultObj(false, nullptr);
+    }
+
+    ResultObj callPython(nlohmann::json &rpcobj)
+    {
+        static int id        = 0;
+        rpcobj["jsonrpc"]    = "2.0";
+        rpcobj["id"]         = ++id;
+        std::string _rawdata = rpcobj.dump(-1, ' ', true);
+        _rawdata += "\n";
+        std::string rawdata;
+        rawdata.reserve(_rawdata.size());
+        for (int i = 0; i < _rawdata.length(); i++) {
+            if ((_rawdata.length() - i > 4) && (_rawdata[i] == '\\') && (_rawdata[i + 1] == '\\') &&
+                (_rawdata[i + 2] == 'u')) {
+                rawdata += '\\';
+                rawdata += 'u';
+                i += 2;
+            } else {
+                rawdata += _rawdata[i];
+            }
+        }
+        const char *delstr = "\"DELSELF\":0";
+        size_t pos         = rawdata.find(delstr);
+        if (pos != std::string::npos) {
+            rawdata.erase(pos, strlen(delstr));
+        }
+        int start_pos = 0;
+        for (;;) {
+            if (rawdata.length() - start_pos > 1024) {
+                // ALOGE("callPython write:%.1024s", rawdata.c_str() + start_pos);
+                write(pipe_c2p[1], rawdata.c_str() + start_pos, 1024);
+                write(pipe_c2p[1], "rpccontinue\n", 12);
+                start_pos += 1024;
+                continue;
+            }
+            // ALOGE("callPython write:%.*s", rawdata.length() - start_pos, rawdata.c_str() + start_pos);
+            write(pipe_c2p[1], rawdata.c_str() + start_pos, rawdata.length() - start_pos);
+            break;
+        }
+        ALOGI("callPython write:%s", rawdata.c_str());
+        return readPython(id);
+    }
+
+public:
+    ~Tokenizer_Auto()
+    {
+        if (pid > 0) {
+            kill(pid, SIGTERM);
+            waitpid(pid, nullptr, 0);
+        }
+    }
+
+    bool Init(std::string model_path = "tokenizer", bool b_bos = true, bool b_eos = false) override
+    {
+        // ALOGE("Tokenizer_Auto model_path:%s", model_path.c_str());
+        std::string tokenizer_script = model_path;
+        std::string model_id         = model_path.substr(0, model_path.find("_"));
+        struct stat st;
+        if (stat(tokenizer_script.c_str(), &st) == 0) {
+            if (S_ISDIR(st.st_mode)) {
+                tokenizer_script = "/opt/m5stack/scripts/llm-llm_tokenizer_auto.py";
+            }
+        } else {
+            return false;
+        }
+        if (pipe(pipe_c2p) == -1 || pipe(pipe_p2c) == -1) {
+            perror("pipe");
+            return false;
+        }
+        pid = fork();
+        if (pid == -1) {
+            perror("fork");
+            return false;
+        }
+        if (pid == 0) {
+            close(pipe_c2p[1]);
+            close(pipe_p2c[0]);
+            dup2(pipe_c2p[0], STDIN_FILENO);
+            dup2(pipe_p2c[1], STDOUT_FILENO);
+            close(pipe_c2p[0]);
+            close(pipe_p2c[1]);
+            execlp("/usr/bin/python3", "python3", tokenizer_script.c_str(), "--model_id", model_id.c_str(), nullptr);
+            perror("execlp");
+            exit(EXIT_FAILURE);
+        }
+        close(pipe_c2p[0]);
+        close(pipe_p2c[1]);
+        auto ret = readPython(0, 15000);
+        if (ret.success) {
+            bos_id       = ret.result["bos_id"].is_number_integer() ? (int)ret.result["bos_id"] : 0;
+            eos_id       = ret.result["eos_id"].is_number_integer() ? (int)ret.result["eos_id"] : 0;
+            this->_b_bos = b_bos;
+            this->_b_eos = b_eos;
+        }
+        return ret.success;
+    }
+
+    bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) override
+    {
+        nlohmann::json rpcobj;
+        rpcobj["method"] = "encode";
+        rpcobj["params"] = nlohmann::json::array({nlohmann::json::array({input}), {{"DELSELF", 0}}});
+        auto ret         = callPython(rpcobj);
+        if (ret.success) {
+            if (ret.result.is_array()) {
+                output = ret.result.get<std::vector<int>>();
+            }
+        }
+        return ret.success;
+    }
+
+    std::vector<int> Encode(std::string input, ImageInfo img_info) override
+    {
+        std::vector<int> output;
+        Encode(input, output, img_info);
+        return output;
+    }
+
+    std::string Decode(const std::vector<int> &input) override
+    {
+        nlohmann::json rpcobj;
+        rpcobj["method"] = "decode";
+        rpcobj["params"] = nlohmann::json::array({nlohmann::json::array({input}), {{"DELSELF", 0}}});
+        auto ret         = callPython(rpcobj);
+        if (ret.success) {
+            return ret.result.get<std::string>();
+        }
+        return "";
+    }
+
+    int GetBosID() override
+    {
+        return bos_id;
+    }
+
+    int GetEosID() override
+    {
+        return eos_id;
+    }
+    std::string apply_chat_template() override
+    {
+        nlohmann::json messages_list = nlohmann::json::array();
+        for (auto &message : messages_) {
+            nlohmann::json msg;
+            switch (message.first) {
+                case ROLE_USER: {
+                    // {"role": "user", "content": prompt}
+                    msg["role"]    = "user";
+                    msg["content"] = (std::string)message.second;
+                    messages_list.push_back(msg);
+                } break;
+                case ROLE_SYSTEM: {
+                    msg["role"]    = "system";
+                    msg["content"] = message.second;
+                    messages_list.push_back(msg);
+                } break;
+                case ROLE_TOOL: {
+                    msg["role"]    = "tool";
+                    msg["content"] = message.second;
+                    messages_list.push_back(msg);
+                } break;
+                case ROLE_IPYTHON: {
+                    msg["role"]    = "ipython";
+                    msg["content"] = message.second;
+                    messages_list.push_back(msg);
+                } break;
+                case ROLE_ASSISTANT: {
+                    msg["role"]    = "assistant";
+                    msg["content"] = message.second;
+                    messages_list.push_back(msg);
+                } break;
+                case ROLE_ASSISTANT_HELP: {
+                } break;
+                default:
+                    break;
+            }
+        }
+        nlohmann::json rpcobj;
+        rpcobj["method"] = "apply_chat_template";
+        rpcobj["params"] = nlohmann::json::array(
+            {nlohmann::json::array({messages_list}), {{"tokenize", false}, {"add_generation_prompt", true}}});
+        auto ret = callPython(rpcobj);
+        if (ret.success) {
+            std::string out_str = (std::string)ret.result;
+            // ALOGE("out_str:%s", out_str.c_str());
+            return out_str;
+        }
+        return "";
+    }
+};
+
+std::shared_ptr<BaseTokenizer> CreateTokenizer(TokenizerType type)
+{
+    switch (type) {
+        case TKT_LLaMa:
+            return std::make_shared<TokenizerLLaMa>();
+        case TKT_MINICPM:
+            return std::make_shared<TokenizerMINICPM>();
+        case TKT_HTTP:
+            return std::make_shared<Tokenizer_Http>();
+        case TKT_Qwen:
+            return std::make_shared<TokenizerQwen>();
+        case TKT_Phi3:
+            return std::make_shared<TokenizerPhi3>();
+        case TKT_AUTO:
+            return std::make_shared<Tokenizer_Auto>();
+        default:
+            return nullptr;
+    }
+}
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.hpp b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.hpp
new file mode 100644
index 00000000..ab545999
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.hpp
@@ -0,0 +1,80 @@
+#pragma once
+#include <string>
+#include <vector>
+#include <memory>
+#include <list>
+#include <utility>
+#include <iostream>
+enum TokenizerType { TKT_LLaMa, TKT_Qwen, TKT_HTTP, TKT_Phi3, TKT_MINICPM, TKT_AUTO, TKT_END };
+
+enum TokenizeRole {
+    ROLE_USER,           // 用户输入
+    ROLE_SYSTEM,         // 提示词
+    ROLE_TOOL,           // 工具
+    ROLE_IPYTHON,        // 工具
+    ROLE_ASSISTANT,      // 助手回复
+    ROLE_ASSISTANT_HELP  // 询问句
+};
+
+struct ImageInfo {
+    int imgsz       = 448;
+    int num_img     = 1;
+    bool img_prompt = false;
+};
+
+class BaseTokenizer {
+public:
+    std::list<std::pair<TokenizeRole, std::string>> messages_;
+
+    void messages_clean()
+    {
+        messages_.clear();
+    }
+
+    virtual ~BaseTokenizer() = default;
+
+    virtual bool Init(std::string model_path)
+    {
+        return false;
+    };
+
+    virtual bool Init(std::string model_path, bool b_bos, bool b_eos)
+    {
+        return false;
+    };
+
+    virtual bool Reset(std::string system_prompt, std::vector<int> &tokens)
+    {
+        return false;
+    };
+
+    virtual bool Encode(std::string input, std::string last_reply, std::vector<int> &tokens,
+                        std::vector<int> &tokens_diff, ImageInfo img_info)
+    {
+        return false;
+    };
+
+    virtual bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) = 0;
+    virtual std::vector<int> Encode(std::string input, ImageInfo img_info)               = 0;
+    virtual std::string Decode(const std::vector<int> &input)                            = 0;
+    virtual int GetBosID()                                                               = 0;
+    virtual int GetEosID()                                                               = 0;
+
+    virtual std::string apply_chat_template() = 0;
+
+    virtual std::string messages_complete(TokenizeRole role, const std::string &content = "")
+    {
+        messages_.push_back(std::make_pair(role, content));
+        if (ROLE_ASSISTANT_HELP == role)
+            return apply_chat_template();
+        else
+            return "";
+    }
+
+    virtual bool isEnd(int id)
+    {
+        return id == GetEosID();
+    }
+};
+
+std::shared_ptr<BaseTokenizer> CreateTokenizer(TokenizerType type);
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/base64.h b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/base64.h
new file mode 100644
index 00000000..c42b4828
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/base64.h
@@ -0,0 +1,54 @@
+#pragma once
+
+#include <string>
+#include <string_view>
+#include <stdexcept>
+
+namespace base64
+{
+
+    static auto pos_of_char(const unsigned char chr) -> size_t
+    {
+        if (chr >= 'A' && chr <= 'Z')
+            return chr - 'A';
+        else if (chr >= 'a' && chr <= 'z')
+            return chr - 'a' + ('Z' - 'A') + 1;
+        else if (chr >= '0' && chr <= '9')
+            return chr - '0' + ('Z' - 'A') + ('z' - 'a') + 2;
+        else if (chr == '+' || chr == '-')
+            return 62;
+        else if (chr == '/' || chr == '_')
+            return 63;
+        else
+            throw std::runtime_error("Input is not valid base64-encoded data.");
+    }
+
+    inline auto decode(std::string_view s) -> std::string
+    {
+        if (s.empty())
+            throw std::runtime_error("empty input");
+        size_t length = s.length();
+        size_t idx = 0;
+
+        std::string out;
+        out.reserve(length / 4 * 3);
+
+        while (idx < length)
+        {
+            size_t pos_of_char_1 = pos_of_char(s.at(idx + 1));
+            out.push_back(static_cast<std::string::value_type>(((pos_of_char(s.at(idx + 0))) << 2) + ((pos_of_char_1 & 0x30) >> 4)));
+            if ((idx + 2 < length) && s.at(idx + 2) != '=' && s.at(idx + 2) != '.')
+            {
+                size_t pos_of_char_2 = pos_of_char(s.at(idx + 2));
+                out.push_back(static_cast<std::string::value_type>(((pos_of_char_1 & 0x0f) << 4) + ((pos_of_char_2 & 0x3c) >> 2)));
+                if ((idx + 3 < length) && s.at(idx + 3) != '=' && s.at(idx + 3) != '.')
+                {
+                    out.push_back(static_cast<std::string::value_type>(((pos_of_char_2 & 0x03) << 6) + pos_of_char(s.at(idx + 3))));
+                }
+            }
+            idx += 4;
+        }
+        return out;
+    }
+
+} // namespace base64
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/tiktoken.h b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/tiktoken.h
new file mode 100644
index 00000000..33df900d
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/tiktoken.h
@@ -0,0 +1,327 @@
+#pragma once
+
+#include <re2/re2.h>
+#include "unordered_dense.h"
+
+#include <cassert>
+#include <limits>
+#include <optional>
+#include <regex>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+namespace tiktoken
+{
+
+	static auto _byte_pair_merge(
+		const std::string &piece,
+		const ankerl::unordered_dense::map<std::string, int> &ranks,
+		std::function<int(int, int)> func) -> std::vector<int>
+	{
+		std::vector<std::pair<int, int>> parts;
+		parts.reserve(piece.size() + 1);
+		for (auto idx = 0U; idx < piece.size() + 1; ++idx)
+		{
+			parts.emplace_back(idx, std::numeric_limits<int>::max());
+		}
+
+		auto get_rank = [&piece, &ranks](
+							const std::vector<std::pair<int, int>> &parts,
+							int start_idx,
+							int skip) -> std::optional<int>
+		{
+			if (start_idx + skip + 2 < parts.size())
+			{
+				auto s = parts[start_idx].first;
+				auto e = parts[start_idx + skip + 2].first;
+				auto key = piece.substr(s, e - s);
+				auto iter = ranks.find(key);
+				if (iter != ranks.end())
+				{
+					return iter->second;
+				}
+			}
+			return std::nullopt;
+		};
+
+		for (auto i = 0U; i < parts.size() - 2; ++i)
+		{
+			auto rank = get_rank(parts, i, 0);
+			if (rank)
+			{
+				assert(*rank != std::numeric_limits<int>::max());
+				parts[i].second = *rank;
+			}
+		}
+
+		while (true)
+		{
+			if (parts.size() == 1)
+				break;
+
+			auto min_rank = std::make_pair<int, int>(std::numeric_limits<int>::max(), 0);
+			for (auto i = 0U; i < parts.size() - 1; ++i)
+			{
+				auto rank = parts[i].second;
+				if (rank < min_rank.first)
+				{
+					min_rank = {rank, i};
+				}
+			}
+
+			if (min_rank.first != std::numeric_limits<int>::max())
+			{
+				auto i = min_rank.second;
+				auto rank = get_rank(parts, i, 1);
+				if (rank)
+				{
+					parts[i].second = *rank;
+				}
+				else
+				{
+					parts[i].second = std::numeric_limits<int>::max();
+				}
+				if (i > 0)
+				{
+					auto rank = get_rank(parts, i - 1, 1);
+					if (rank)
+					{
+						parts[i - 1].second = *rank;
+					}
+					else
+					{
+						parts[i - 1].second = std::numeric_limits<int>::max();
+					}
+				}
+
+				parts.erase(parts.begin() + (i + 1));
+			}
+			else
+			{
+				break;
+			}
+		}
+		std::vector<int> out;
+		out.reserve(parts.size() - 1);
+		for (auto i = 0U; i < parts.size() - 1; ++i)
+		{
+			out.push_back(func(parts[i].first, parts[i + 1].first));
+		}
+		return out;
+	}
+
+	static auto byte_pair_encode(
+		const std::string &piece,
+		const ankerl::unordered_dense::map<std::string, int> &ranks) -> std::vector<int>
+	{
+		if (piece.size() == 1)
+		{
+			return {ranks.at(piece)};
+		}
+
+		auto func = [&piece, &ranks](int start, int stop) -> int
+		{
+			std::string key = piece.substr(start, stop - start);
+			return ranks.at(key);
+		};
+
+		return _byte_pair_merge(piece, ranks, func);
+	}
+
+	class tiktoken
+	{
+	public:
+		tiktoken() = default;
+		tiktoken(
+			ankerl::unordered_dense::map<std::string, int> encoder,
+			ankerl::unordered_dense::map<std::string, int> special_encoder,
+			const std::string &pattern)
+		{
+			regex_ = std::make_unique<re2::RE2>("(" + pattern + ")");
+
+			std::string special_pattern;
+			for (const auto &item : special_encoder)
+			{
+				if (!special_pattern.empty())
+				{
+					special_pattern += "|";
+				}
+				special_pattern += re2::RE2::QuoteMeta(item.first);
+			}
+			if (special_pattern.empty())
+			{
+				special_regex_ = nullptr;
+			}
+			else
+			{
+				special_regex_ = std::make_unique<re2::RE2>("(" + special_pattern + ")");
+			}
+
+			encoder_ = std::move(encoder);
+			special_tokens_encoder = std::move(special_encoder);
+
+			for (const auto &[k, v] : encoder_)
+			{
+				decoder_.emplace(v, k);
+			}
+			// 不知道有什么用 debug会报错，先注释了再说
+			// assert(encoder_.size() != decoder_.size() && "Encoder and decoder must be of equal length; maybe you had duplicate token indices in your encoder?");
+
+			for (const auto &[k, v] : special_tokens_encoder)
+			{
+				special_tokens_decoder.emplace(v, k);
+			}
+		}
+
+		auto encode_ordinary(const std::string &text) const -> std::vector<int>
+		{
+			return _encode_ordinary_native(text);
+		}
+
+		auto encode(const std::string &text) const -> std::vector<int>
+		{
+			return _encode_native(text, special_tokens_encoder).first;
+		}
+
+		auto encode_single_piece(const std::string &text) const -> std::vector<int>
+		{
+			auto iter = encoder_.find(text);
+			if (iter != encoder_.end())
+			{
+				return {iter->second};
+			}
+			return byte_pair_encode(text, encoder_);
+		}
+
+		auto decode(const std::vector<int> &tokens) const -> std::string
+		{
+			return _decode_native(tokens);
+		}
+
+	private:
+		auto split_with_allowed_special_token(
+			re2::StringPiece &input,
+			const ankerl::unordered_dense::map<std::string, int> &allowed_special) const -> std::pair<std::optional<std::string>, re2::StringPiece>
+		{
+			if (special_regex_ == nullptr)
+				return {std::nullopt, input};
+
+			auto start = input.begin();
+			std::string special;
+			while (true)
+			{
+				if (!re2::RE2::FindAndConsume(&input, *special_regex_, &special))
+				{
+					break;
+				}
+
+				if (allowed_special.count(special) == 1)
+				{
+					return {std::move(special), re2::StringPiece(start, input.begin() - start - special.size())};
+				}
+			}
+
+			return {std::nullopt, input};
+		}
+
+		auto _encode_ordinary_native(const std::string &text) const -> std::vector<int>
+		{
+			std::vector<int> ret;
+			re2::StringPiece input(text);
+
+			std::string piece;
+			while (re2::RE2::FindAndConsume(&input, *regex_, &piece))
+			{
+				auto iter = encoder_.find(piece);
+				if (iter != encoder_.end())
+				{
+					ret.push_back(iter->second);
+					continue;
+				}
+				auto tokens = byte_pair_encode(piece, encoder_);
+				ret.insert(ret.end(), tokens.begin(), tokens.end());
+			}
+			return ret;
+		}
+
+		auto _encode_native(
+			const std::string &text,
+			const ankerl::unordered_dense::map<std::string, int> &allowed_special) const -> std::pair<std::vector<int>, int>
+		{
+			std::vector<int> ret;
+			int last_piece_token_len = 0;
+			re2::StringPiece input(text);
+
+			while (true)
+			{
+				auto [special, sub_input] = split_with_allowed_special_token(input, allowed_special);
+				std::string piece;
+				while (re2::RE2::FindAndConsume(&sub_input, *regex_, &piece))
+				{
+					auto iter = encoder_.find(piece);
+					if (iter != encoder_.end())
+					{
+						last_piece_token_len = 1;
+						ret.push_back(iter->second);
+						continue;
+					}
+					auto tokens = byte_pair_encode(piece, encoder_);
+					last_piece_token_len = tokens.size();
+					ret.insert(ret.end(), tokens.begin(), tokens.end());
+				}
+
+				if (special)
+				{
+					int token = special_tokens_encoder.at(*special);
+					ret.push_back(token);
+					last_piece_token_len = 0;
+				}
+				else
+				{
+					break;
+				}
+			}
+
+			return {ret, last_piece_token_len};
+		}
+
+		auto _decode_native(const std::vector<int> &tokens) const -> std::string
+		{
+			std::string ret;
+			ret.reserve(tokens.size() * 2);
+			for (auto token : tokens)
+			{
+				std::string token_bytes;
+				auto iter = decoder_.find(token);
+				if (iter != decoder_.end())
+				{
+					token_bytes = iter->second;
+				}
+				else
+				{
+					iter = special_tokens_decoder.find(token);
+					if (iter != special_tokens_decoder.end())
+					{
+						token_bytes = iter->second;
+					}
+					else
+					{
+						throw std::runtime_error("unknown token: " + std::to_string(token));
+					}
+				}
+				ret += token_bytes;
+			}
+			return ret;
+		}
+
+		ankerl::unordered_dense::map<std::string, int> encoder_;
+		ankerl::unordered_dense::map<std::string, int> special_tokens_encoder;
+		ankerl::unordered_dense::map<int, std::string> decoder_;
+		ankerl::unordered_dense::map<int, std::string> special_tokens_decoder;
+		std::unique_ptr<re2::RE2> regex_;
+		std::unique_ptr<re2::RE2> special_regex_;
+	};
+
+} // namespace tiktoken
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/unordered_dense.h b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/unordered_dense.h
new file mode 100644
index 00000000..af23690b
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/unordered_dense.h
@@ -0,0 +1,2240 @@
+///////////////////////// ankerl::unordered_dense::{map, set} /////////////////////////
+
+// A fast & densely stored hashmap and hashset based on robin-hood backward shift deletion.
+// Version 4.1.2
+// https://github.com/martinus/unordered_dense
+//
+// Licensed under the MIT License <http://opensource.org/licenses/MIT>.
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2022-2023 Martin Leitner-Ankerl <martin.ankerl@gmail.com>
+//
+// Permission is hereby granted, free of charge, to any person obtaining a copy
+// of this software and associated documentation files (the "Software"), to deal
+// in the Software without restriction, including without limitation the rights
+// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+// copies of the Software, and to permit persons to whom the Software is
+// furnished to do so, subject to the following conditions:
+//
+// The above copyright notice and this permission notice shall be included in all
+// copies or substantial portions of the Software.
+//
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+// SOFTWARE.
+
+#ifndef ANKERL_UNORDERED_DENSE_H
+#define ANKERL_UNORDERED_DENSE_H
+
+// see https://semver.org/spec/v2.0.0.html
+#define ANKERL_UNORDERED_DENSE_VERSION_MAJOR 4 // NOLINT(cppcoreguidelines-macro-usage) incompatible API changes
+#define ANKERL_UNORDERED_DENSE_VERSION_MINOR 1 // NOLINT(cppcoreguidelines-macro-usage) backwards compatible functionality
+#define ANKERL_UNORDERED_DENSE_VERSION_PATCH 2 // NOLINT(cppcoreguidelines-macro-usage) backwards compatible bug fixes
+
+// API versioning with inline namespace, see https://www.foonathan.net/2018/11/inline-namespaces/
+
+// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
+#define ANKERL_UNORDERED_DENSE_VERSION_CONCAT1(major, minor, patch) v##major##_##minor##_##patch
+// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
+#define ANKERL_UNORDERED_DENSE_VERSION_CONCAT(major, minor, patch) ANKERL_UNORDERED_DENSE_VERSION_CONCAT1(major, minor, patch)
+#define ANKERL_UNORDERED_DENSE_NAMESPACE   \
+    ANKERL_UNORDERED_DENSE_VERSION_CONCAT( \
+        ANKERL_UNORDERED_DENSE_VERSION_MAJOR, ANKERL_UNORDERED_DENSE_VERSION_MINOR, ANKERL_UNORDERED_DENSE_VERSION_PATCH)
+
+#if defined(_MSVC_LANG)
+#define ANKERL_UNORDERED_DENSE_CPP_VERSION _MSVC_LANG
+#else
+#define ANKERL_UNORDERED_DENSE_CPP_VERSION __cplusplus
+#endif
+
+#if defined(__GNUC__)
+// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
+#define ANKERL_UNORDERED_DENSE_PACK(decl) decl __attribute__((__packed__))
+#elif defined(_MSC_VER)
+// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
+#define ANKERL_UNORDERED_DENSE_PACK(decl) __pragma(pack(push, 1)) decl __pragma(pack(pop))
+#endif
+
+// exceptions
+#if defined(__cpp_exceptions) || defined(__EXCEPTIONS) || defined(_CPPUNWIND)
+#define ANKERL_UNORDERED_DENSE_HAS_EXCEPTIONS() 1 // NOLINT(cppcoreguidelines-macro-usage)
+#else
+#define ANKERL_UNORDERED_DENSE_HAS_EXCEPTIONS() 0 // NOLINT(cppcoreguidelines-macro-usage)
+#endif
+#ifdef _MSC_VER
+#define ANKERL_UNORDERED_DENSE_NOINLINE __declspec(noinline)
+#else
+#define ANKERL_UNORDERED_DENSE_NOINLINE __attribute__((noinline))
+#endif
+
+// defined in unordered_dense.cpp
+#if !defined(ANKERL_UNORDERED_DENSE_EXPORT)
+#define ANKERL_UNORDERED_DENSE_EXPORT
+#endif
+
+#if ANKERL_UNORDERED_DENSE_CPP_VERSION < 201703L
+#error ankerl::unordered_dense requires C++17 or higher
+#else
+#include <array>            // for array
+#include <cstdint>          // for uint64_t, uint32_t, uint8_t, UINT64_C
+#include <cstring>          // for size_t, memcpy, memset
+#include <functional>       // for equal_to, hash
+#include <initializer_list> // for initializer_list
+#include <iterator>         // for pair, distance
+#include <limits>           // for numeric_limits
+#include <memory>           // for allocator, allocator_traits, shared_ptr
+#include <stdexcept>        // for out_of_range
+#include <string>           // for basic_string
+#include <string_view>      // for basic_string_view, hash
+#include <tuple>            // for forward_as_tuple
+#include <type_traits>      // for enable_if_t, declval, conditional_t, ena...
+#include <utility>          // for forward, exchange, pair, as_const, piece...
+#include <vector>           // for vector
+#if ANKERL_UNORDERED_DENSE_HAS_EXCEPTIONS() == 0
+#include <cstdlib> // for abort
+#endif
+
+#if defined(__has_include)
+#if __has_include(<memory_resource>)
+#define ANKERL_UNORDERED_DENSE_PMR std::pmr // NOLINT(cppcoreguidelines-macro-usage)
+#include <memory_resource>                  // for polymorphic_allocator
+#elif __has_include(<experimental/memory_resource>)
+#define ANKERL_UNORDERED_DENSE_PMR std::experimental::pmr // NOLINT(cppcoreguidelines-macro-usage)
+#include <experimental/memory_resource>                   // for polymorphic_allocator
+#endif
+#endif
+
+#if defined(_MSC_VER) && defined(_M_X64)
+#include <intrin.h>
+#pragma intrinsic(_umul128)
+#endif
+
+#if defined(__GNUC__) || defined(__INTEL_COMPILER) || defined(__clang__)
+#define ANKERL_UNORDERED_DENSE_LIKELY(x) __builtin_expect(x, 1)   // NOLINT(cppcoreguidelines-macro-usage)
+#define ANKERL_UNORDERED_DENSE_UNLIKELY(x) __builtin_expect(x, 0) // NOLINT(cppcoreguidelines-macro-usage)
+#else
+#define ANKERL_UNORDERED_DENSE_LIKELY(x) (x)   // NOLINT(cppcoreguidelines-macro-usage)
+#define ANKERL_UNORDERED_DENSE_UNLIKELY(x) (x) // NOLINT(cppcoreguidelines-macro-usage)
+#endif
+
+namespace ankerl::unordered_dense
+{
+    inline namespace ANKERL_UNORDERED_DENSE_NAMESPACE
+    {
+
+        namespace detail
+        {
+
+#if ANKERL_UNORDERED_DENSE_HAS_EXCEPTIONS()
+
+            // make sure this is not inlined as it is slow and dramatically enlarges code, thus making other
+            // inlinings more difficult. Throws are also generally the slow path.
+            [[noreturn]] inline ANKERL_UNORDERED_DENSE_NOINLINE void on_error_key_not_found()
+            {
+                throw std::out_of_range("ankerl::unordered_dense::map::at(): key not found");
+            }
+            [[noreturn]] inline ANKERL_UNORDERED_DENSE_NOINLINE void on_error_bucket_overflow()
+            {
+                throw std::overflow_error("ankerl::unordered_dense: reached max bucket size, cannot increase size");
+            }
+            [[noreturn]] inline ANKERL_UNORDERED_DENSE_NOINLINE void on_error_too_many_elements()
+            {
+                throw std::out_of_range("ankerl::unordered_dense::map::replace(): too many elements");
+            }
+
+#else
+
+            [[noreturn]] inline void on_error_key_not_found()
+            {
+                abort();
+            }
+            [[noreturn]] inline void on_error_bucket_overflow()
+            {
+                abort();
+            }
+            [[noreturn]] inline void on_error_too_many_elements()
+            {
+                abort();
+            }
+
+#endif
+
+        } // namespace detail
+
+        // hash ///////////////////////////////////////////////////////////////////////
+
+        // This is a stripped-down implementation of wyhash: https://github.com/wangyi-fudan/wyhash
+        // No big-endian support (because different values on different machines don't matter),
+        // hardcodes seed and the secret, reformats the code, and clang-tidy fixes.
+        namespace detail::wyhash
+        {
+
+            inline void mum(uint64_t *a, uint64_t *b)
+            {
+#if defined(__SIZEOF_INT128__)
+                __uint128_t r = *a;
+                r *= *b;
+                *a = static_cast<uint64_t>(r);
+                *b = static_cast<uint64_t>(r >> 64U);
+#elif defined(_MSC_VER) && defined(_M_X64)
+                *a = _umul128(*a, *b, b);
+#else
+                uint64_t ha = *a >> 32U;
+                uint64_t hb = *b >> 32U;
+                uint64_t la = static_cast<uint32_t>(*a);
+                uint64_t lb = static_cast<uint32_t>(*b);
+                uint64_t hi{};
+                uint64_t lo{};
+                uint64_t rh = ha * hb;
+                uint64_t rm0 = ha * lb;
+                uint64_t rm1 = hb * la;
+                uint64_t rl = la * lb;
+                uint64_t t = rl + (rm0 << 32U);
+                auto c = static_cast<uint64_t>(t < rl);
+                lo = t + (rm1 << 32U);
+                c += static_cast<uint64_t>(lo < t);
+                hi = rh + (rm0 >> 32U) + (rm1 >> 32U) + c;
+                *a = lo;
+                *b = hi;
+#endif
+            }
+
+            // multiply and xor mix function, aka MUM
+            [[nodiscard]] inline auto mix(uint64_t a, uint64_t b) -> uint64_t
+            {
+                mum(&a, &b);
+                return a ^ b;
+            }
+
+            // read functions. WARNING: we don't care about endianness, so results are different on big endian!
+            [[nodiscard]] inline auto r8(const uint8_t *p) -> uint64_t
+            {
+                uint64_t v{};
+                std::memcpy(&v, p, 8U);
+                return v;
+            }
+
+            [[nodiscard]] inline auto r4(const uint8_t *p) -> uint64_t
+            {
+                uint32_t v{};
+                std::memcpy(&v, p, 4);
+                return v;
+            }
+
+            // reads 1, 2, or 3 bytes
+            [[nodiscard]] inline auto r3(const uint8_t *p, size_t k) -> uint64_t
+            {
+                return (static_cast<uint64_t>(p[0]) << 16U) | (static_cast<uint64_t>(p[k >> 1U]) << 8U) | p[k - 1];
+            }
+
+            [[maybe_unused]] [[nodiscard]] inline auto hash(void const *key, size_t len) -> uint64_t
+            {
+                static constexpr auto secret = std::array{UINT64_C(0xa0761d6478bd642f),
+                                                          UINT64_C(0xe7037ed1a0b428db),
+                                                          UINT64_C(0x8ebc6af09c88c6e3),
+                                                          UINT64_C(0x589965cc75374cc3)};
+
+                auto const *p = static_cast<uint8_t const *>(key);
+                uint64_t seed = secret[0];
+                uint64_t a{};
+                uint64_t b{};
+                if (ANKERL_UNORDERED_DENSE_LIKELY(len <= 16))
+                {
+                    if (ANKERL_UNORDERED_DENSE_LIKELY(len >= 4))
+                    {
+                        a = (r4(p) << 32U) | r4(p + ((len >> 3U) << 2U));
+                        b = (r4(p + len - 4) << 32U) | r4(p + len - 4 - ((len >> 3U) << 2U));
+                    }
+                    else if (ANKERL_UNORDERED_DENSE_LIKELY(len > 0))
+                    {
+                        a = r3(p, len);
+                        b = 0;
+                    }
+                    else
+                    {
+                        a = 0;
+                        b = 0;
+                    }
+                }
+                else
+                {
+                    size_t i = len;
+                    if (ANKERL_UNORDERED_DENSE_UNLIKELY(i > 48))
+                    {
+                        uint64_t see1 = seed;
+                        uint64_t see2 = seed;
+                        do
+                        {
+                            seed = mix(r8(p) ^ secret[1], r8(p + 8) ^ seed);
+                            see1 = mix(r8(p + 16) ^ secret[2], r8(p + 24) ^ see1);
+                            see2 = mix(r8(p + 32) ^ secret[3], r8(p + 40) ^ see2);
+                            p += 48;
+                            i -= 48;
+                        } while (ANKERL_UNORDERED_DENSE_LIKELY(i > 48));
+                        seed ^= see1 ^ see2;
+                    }
+                    while (ANKERL_UNORDERED_DENSE_UNLIKELY(i > 16))
+                    {
+                        seed = mix(r8(p) ^ secret[1], r8(p + 8) ^ seed);
+                        i -= 16;
+                        p += 16;
+                    }
+                    a = r8(p + i - 16);
+                    b = r8(p + i - 8);
+                }
+
+                return mix(secret[1] ^ len, mix(a ^ secret[1], b ^ seed));
+            }
+
+            [[nodiscard]] inline auto hash(uint64_t x) -> uint64_t
+            {
+                return detail::wyhash::mix(x, UINT64_C(0x9E3779B97F4A7C15));
+            }
+
+        } // namespace detail::wyhash
+
+        ANKERL_UNORDERED_DENSE_EXPORT template <typename T, typename Enable = void>
+        struct hash
+        {
+            auto operator()(T const &obj) const noexcept(noexcept(std::declval<std::hash<T>>().operator()(std::declval<T const &>())))
+                -> uint64_t
+            {
+                return std::hash<T>{}(obj);
+            }
+        };
+
+        template <typename CharT>
+        struct hash<std::basic_string<CharT>>
+        {
+            using is_avalanching = void;
+            auto operator()(std::basic_string<CharT> const &str) const noexcept -> uint64_t
+            {
+                return detail::wyhash::hash(str.data(), sizeof(CharT) * str.size());
+            }
+        };
+
+        template <typename CharT>
+        struct hash<std::basic_string_view<CharT>>
+        {
+            using is_avalanching = void;
+            auto operator()(std::basic_string_view<CharT> const &sv) const noexcept -> uint64_t
+            {
+                return detail::wyhash::hash(sv.data(), sizeof(CharT) * sv.size());
+            }
+        };
+
+        template <class T>
+        struct hash<T *>
+        {
+            using is_avalanching = void;
+            auto operator()(T *ptr) const noexcept -> uint64_t
+            {
+                // NOLINTNEXTLINE(cppcoreguidelines-pro-type-reinterpret-cast)
+                return detail::wyhash::hash(reinterpret_cast<uintptr_t>(ptr));
+            }
+        };
+
+        template <class T>
+        struct hash<std::unique_ptr<T>>
+        {
+            using is_avalanching = void;
+            auto operator()(std::unique_ptr<T> const &ptr) const noexcept -> uint64_t
+            {
+                // NOLINTNEXTLINE(cppcoreguidelines-pro-type-reinterpret-cast)
+                return detail::wyhash::hash(reinterpret_cast<uintptr_t>(ptr.get()));
+            }
+        };
+
+        template <class T>
+        struct hash<std::shared_ptr<T>>
+        {
+            using is_avalanching = void;
+            auto operator()(std::shared_ptr<T> const &ptr) const noexcept -> uint64_t
+            {
+                // NOLINTNEXTLINE(cppcoreguidelines-pro-type-reinterpret-cast)
+                return detail::wyhash::hash(reinterpret_cast<uintptr_t>(ptr.get()));
+            }
+        };
+
+        template <typename Enum>
+        struct hash<Enum, typename std::enable_if<std::is_enum<Enum>::value>::type>
+        {
+            using is_avalanching = void;
+            auto operator()(Enum e) const noexcept -> uint64_t
+            {
+                using underlying = typename std::underlying_type_t<Enum>;
+                return detail::wyhash::hash(static_cast<underlying>(e));
+            }
+        };
+
+// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
+#define ANKERL_UNORDERED_DENSE_HASH_STATICCAST(T)                    \
+    template <>                                                      \
+    struct hash<T>                                                   \
+    {                                                                \
+        using is_avalanching = void;                                 \
+        auto operator()(T const &obj) const noexcept -> uint64_t     \
+        {                                                            \
+            return detail::wyhash::hash(static_cast<uint64_t>(obj)); \
+        }                                                            \
+    }
+
+#if defined(__GNUC__) && !defined(__clang__)
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wuseless-cast"
+#endif
+        // see https://en.cppreference.com/w/cpp/utility/hash
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(bool);
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(char);
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(signed char);
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(unsigned char);
+#if ANKERL_UNORDERED_DENSE_CPP_VERSION >= 202002L && defined(__cpp_char8_t)
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(char8_t);
+#endif
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(char16_t);
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(char32_t);
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(wchar_t);
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(short);
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(unsigned short);
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(int);
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(unsigned int);
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(long);
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(long long);
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(unsigned long);
+        ANKERL_UNORDERED_DENSE_HASH_STATICCAST(unsigned long long);
+
+#if defined(__GNUC__) && !defined(__clang__)
+#pragma GCC diagnostic pop
+#endif
+
+        // bucket_type //////////////////////////////////////////////////////////
+
+        namespace bucket_type
+        {
+
+            struct standard
+            {
+                static constexpr uint32_t dist_inc = 1U << 8U;             // skip 1 byte fingerprint
+                static constexpr uint32_t fingerprint_mask = dist_inc - 1; // mask for 1 byte of fingerprint
+
+                uint32_t m_dist_and_fingerprint; // upper 3 byte: distance to original bucket. lower byte: fingerprint from hash
+                uint32_t m_value_idx;            // index into the m_values vector.
+            };
+
+            ANKERL_UNORDERED_DENSE_PACK(struct big {
+                static constexpr uint32_t dist_inc = 1U << 8U;             // skip 1 byte fingerprint
+                static constexpr uint32_t fingerprint_mask = dist_inc - 1; // mask for 1 byte of fingerprint
+
+                uint32_t m_dist_and_fingerprint; // upper 3 byte: distance to original bucket. lower byte: fingerprint from hash
+                size_t m_value_idx;              // index into the m_values vector.
+            });
+
+        } // namespace bucket_type
+
+        namespace detail
+        {
+
+            struct nonesuch
+            {
+            };
+
+            template <class Default, class AlwaysVoid, template <class...> class Op, class... Args>
+            struct detector
+            {
+                using value_t = std::false_type;
+                using type = Default;
+            };
+
+            template <class Default, template <class...> class Op, class... Args>
+            struct detector<Default, std::void_t<Op<Args...>>, Op, Args...>
+            {
+                using value_t = std::true_type;
+                using type = Op<Args...>;
+            };
+
+            template <template <class...> class Op, class... Args>
+            using is_detected = typename detail::detector<detail::nonesuch, void, Op, Args...>::value_t;
+
+            template <template <class...> class Op, class... Args>
+            constexpr bool is_detected_v = is_detected<Op, Args...>::value;
+
+            template <typename T>
+            using detect_avalanching = typename T::is_avalanching;
+
+            template <typename T>
+            using detect_is_transparent = typename T::is_transparent;
+
+            template <typename T>
+            using detect_iterator = typename T::iterator;
+
+            template <typename T>
+            using detect_reserve = decltype(std::declval<T &>().reserve(size_t{}));
+
+            // enable_if helpers
+
+            template <typename Mapped>
+            constexpr bool is_map_v = !std::is_void_v<Mapped>;
+
+            // clang-format off
+template <typename Hash, typename KeyEqual>
+constexpr bool is_transparent_v = is_detected_v<detect_is_transparent, Hash> && is_detected_v<detect_is_transparent, KeyEqual>;
+            // clang-format on
+
+            template <typename From, typename To1, typename To2>
+            constexpr bool is_neither_convertible_v = !std::is_convertible_v<From, To1> && !std::is_convertible_v<From, To2>;
+
+            template <typename T>
+            constexpr bool has_reserve = is_detected_v<detect_reserve, T>;
+
+            // base type for map has mapped_type
+            template <class T>
+            struct base_table_type_map
+            {
+                using mapped_type = T;
+            };
+
+            // base type for set doesn't have mapped_type
+            struct base_table_type_set
+            {
+            };
+
+        } // namespace detail
+
+        // Very much like std::deque, but faster for indexing (in most cases). As of now this doesn't implement the full std::vector
+        // API, but merely what's necessary to work as an underlying container for ankerl::unordered_dense::{map, set}.
+        // It allocates blocks of equal size and puts them into the m_blocks vector. That means it can grow simply by adding a new
+        // block to the back of m_blocks, and doesn't double its size like an std::vector. The disadvantage is that memory is not
+        // linear and thus there is one more indirection necessary for indexing.
+        template <typename T, typename Allocator = std::allocator<T>, size_t MaxSegmentSizeBytes = 4096>
+        class segmented_vector
+        {
+            template <bool IsConst>
+            class iter_t;
+
+        public:
+            using allocator_type = Allocator;
+            using pointer = typename std::allocator_traits<allocator_type>::pointer;
+            using const_pointer = typename std::allocator_traits<allocator_type>::const_pointer;
+            using difference_type = typename std::allocator_traits<allocator_type>::difference_type;
+            using value_type = T;
+            using size_type = std::size_t;
+            using reference = T &;
+            using const_reference = T const &;
+            using iterator = iter_t<false>;
+            using const_iterator = iter_t<true>;
+
+        private:
+            using vec_alloc = typename std::allocator_traits<Allocator>::template rebind_alloc<pointer>;
+            std::vector<pointer, vec_alloc> m_blocks{};
+            size_t m_size{};
+
+            // Calculates the maximum number for x in  (s << x) <= max_val
+            static constexpr auto num_bits_closest(size_t max_val, size_t s) -> size_t
+            {
+                auto f = size_t{0};
+                while (s << (f + 1) <= max_val)
+                {
+                    ++f;
+                }
+                return f;
+            }
+
+            using self_t = segmented_vector<T, Allocator, MaxSegmentSizeBytes>;
+            static constexpr auto num_bits = num_bits_closest(MaxSegmentSizeBytes, sizeof(T));
+            static constexpr auto num_elements_in_block = 1U << num_bits;
+            static constexpr auto mask = num_elements_in_block - 1U;
+
+            /**
+             * Iterator class doubles as const_iterator and iterator
+             */
+            template <bool IsConst>
+            class iter_t
+            {
+                using ptr_t = typename std::conditional_t<IsConst, segmented_vector::const_pointer const *, segmented_vector::pointer *>;
+                ptr_t m_data{};
+                size_t m_idx{};
+
+                template <bool B>
+                friend class iter_t;
+
+            public:
+                using difference_type = segmented_vector::difference_type;
+                using value_type = T;
+                using reference = typename std::conditional_t<IsConst, value_type const &, value_type &>;
+                using pointer = typename std::conditional_t<IsConst, segmented_vector::const_pointer, segmented_vector::pointer>;
+                using iterator_category = std::forward_iterator_tag;
+
+                iter_t() noexcept = default;
+
+                template <bool OtherIsConst, typename = typename std::enable_if<IsConst && !OtherIsConst>::type>
+                // NOLINTNEXTLINE(google-explicit-constructor,hicpp-explicit-conversions)
+                constexpr iter_t(iter_t<OtherIsConst> const &other) noexcept
+                    : m_data(other.m_data), m_idx(other.m_idx)
+                {
+                }
+
+                constexpr iter_t(ptr_t data, size_t idx) noexcept
+                    : m_data(data), m_idx(idx) {}
+
+                template <bool OtherIsConst, typename = typename std::enable_if<IsConst && !OtherIsConst>::type>
+                constexpr auto operator=(iter_t<OtherIsConst> const &other) noexcept -> iter_t &
+                {
+                    m_data = other.m_data;
+                    m_idx = other.m_idx;
+                    return *this;
+                }
+
+                constexpr auto operator++() noexcept -> iter_t &
+                {
+                    ++m_idx;
+                    return *this;
+                }
+
+                constexpr auto operator+(difference_type diff) noexcept -> iter_t
+                {
+                    return {m_data, static_cast<size_t>(static_cast<difference_type>(m_idx) + diff)};
+                }
+
+                template <bool OtherIsConst>
+                constexpr auto operator-(iter_t<OtherIsConst> const &other) noexcept -> difference_type
+                {
+                    return static_cast<difference_type>(m_idx) - static_cast<difference_type>(other.m_idx);
+                }
+
+                constexpr auto operator*() const noexcept -> reference
+                {
+                    return m_data[m_idx >> num_bits][m_idx & mask];
+                }
+
+                constexpr auto operator->() const noexcept -> pointer
+                {
+                    return &m_data[m_idx >> num_bits][m_idx & mask];
+                }
+
+                template <bool O>
+                constexpr auto operator==(iter_t<O> const &o) const noexcept -> bool
+                {
+                    return m_idx == o.m_idx;
+                }
+
+                template <bool O>
+                constexpr auto operator!=(iter_t<O> const &o) const noexcept -> bool
+                {
+                    return !(*this == o);
+                }
+            };
+
+            // slow path: need to allocate a new segment every once in a while
+            void increase_capacity()
+            {
+                auto ba = Allocator(m_blocks.get_allocator());
+                pointer block = std::allocator_traits<Allocator>::allocate(ba, num_elements_in_block);
+                m_blocks.push_back(block);
+            }
+
+            // Moves everything from other
+            void append_everything_from(segmented_vector &&other)
+            {
+                reserve(size() + other.size());
+                for (auto &&o : other)
+                {
+                    emplace_back(std::move(o));
+                }
+            }
+
+            // Copies everything from other
+            void append_everything_from(segmented_vector const &other)
+            {
+                reserve(size() + other.size());
+                for (auto const &o : other)
+                {
+                    emplace_back(o);
+                }
+            }
+
+            void dealloc()
+            {
+                auto ba = Allocator(m_blocks.get_allocator());
+                for (auto ptr : m_blocks)
+                {
+                    std::allocator_traits<Allocator>::deallocate(ba, ptr, num_elements_in_block);
+                }
+            }
+
+            [[nodiscard]] static constexpr auto calc_num_blocks_for_capacity(size_t capacity)
+            {
+                return (capacity + num_elements_in_block - 1U) / num_elements_in_block;
+            }
+
+        public:
+            segmented_vector() = default;
+
+            // NOLINTNEXTLINE(google-explicit-constructor,hicpp-explicit-conversions)
+            segmented_vector(Allocator alloc)
+                : m_blocks(vec_alloc(alloc)) {}
+
+            segmented_vector(segmented_vector &&other, Allocator alloc)
+                : segmented_vector(alloc)
+            {
+                *this = std::move(other);
+            }
+
+            segmented_vector(segmented_vector const &other, Allocator alloc)
+                : m_blocks(vec_alloc(alloc))
+            {
+                append_everything_from(other);
+            }
+
+            segmented_vector(segmented_vector &&other) noexcept
+                : segmented_vector(std::move(other), get_allocator()) {}
+
+            segmented_vector(segmented_vector const &other)
+            {
+                append_everything_from(other);
+            }
+
+            auto operator=(segmented_vector const &other) -> segmented_vector &
+            {
+                if (this == &other)
+                {
+                    return *this;
+                }
+                clear();
+                append_everything_from(other);
+                return *this;
+            }
+
+            auto operator=(segmented_vector &&other) noexcept -> segmented_vector &
+            {
+                clear();
+                dealloc();
+                if (other.get_allocator() == get_allocator())
+                {
+                    m_blocks = std::move(other.m_blocks);
+                    m_size = std::exchange(other.m_size, {});
+                }
+                else
+                {
+                    // make sure to construct with other's allocator!
+                    m_blocks = std::vector<pointer, vec_alloc>(vec_alloc(other.get_allocator()));
+                    append_everything_from(std::move(other));
+                }
+                return *this;
+            }
+
+            ~segmented_vector()
+            {
+                clear();
+                dealloc();
+            }
+
+            [[nodiscard]] constexpr auto size() const -> size_t
+            {
+                return m_size;
+            }
+
+            [[nodiscard]] constexpr auto capacity() const -> size_t
+            {
+                return m_blocks.size() * num_elements_in_block;
+            }
+
+            // Indexing is highly performance critical
+            [[nodiscard]] constexpr auto operator[](size_t i) const noexcept -> T const &
+            {
+                return m_blocks[i >> num_bits][i & mask];
+            }
+
+            [[nodiscard]] constexpr auto operator[](size_t i) noexcept -> T &
+            {
+                return m_blocks[i >> num_bits][i & mask];
+            }
+
+            [[nodiscard]] constexpr auto begin() -> iterator
+            {
+                return {m_blocks.data(), 0U};
+            }
+            [[nodiscard]] constexpr auto begin() const -> const_iterator
+            {
+                return {m_blocks.data(), 0U};
+            }
+            [[nodiscard]] constexpr auto cbegin() const -> const_iterator
+            {
+                return {m_blocks.data(), 0U};
+            }
+
+            [[nodiscard]] constexpr auto end() -> iterator
+            {
+                return {m_blocks.data(), m_size};
+            }
+            [[nodiscard]] constexpr auto end() const -> const_iterator
+            {
+                return {m_blocks.data(), m_size};
+            }
+            [[nodiscard]] constexpr auto cend() const -> const_iterator
+            {
+                return {m_blocks.data(), m_size};
+            }
+
+            [[nodiscard]] constexpr auto back() -> reference
+            {
+                return operator[](m_size - 1);
+            }
+            [[nodiscard]] constexpr auto back() const -> const_reference
+            {
+                return operator[](m_size - 1);
+            }
+
+            void pop_back()
+            {
+                back().~T();
+                --m_size;
+            }
+
+            [[nodiscard]] auto empty() const
+            {
+                return 0 == m_size;
+            }
+
+            void reserve(size_t new_capacity)
+            {
+                m_blocks.reserve(calc_num_blocks_for_capacity(new_capacity));
+                while (new_capacity > capacity())
+                {
+                    increase_capacity();
+                }
+            }
+
+            [[nodiscard]] auto get_allocator() const -> allocator_type
+            {
+                return allocator_type{m_blocks.get_allocator()};
+            }
+
+            template <class... Args>
+            auto emplace_back(Args &&...args) -> reference
+            {
+                if (m_size == capacity())
+                {
+                    increase_capacity();
+                }
+                auto *ptr = static_cast<void *>(&operator[](m_size));
+                auto &ref = *new (ptr) T(std::forward<Args>(args)...);
+                ++m_size;
+                return ref;
+            }
+
+            void clear()
+            {
+                if constexpr (!std::is_trivially_destructible_v<T>)
+                {
+                    for (size_t i = 0, s = size(); i < s; ++i)
+                    {
+                        operator[](i).~T();
+                    }
+                }
+                m_size = 0;
+            }
+
+            void shrink_to_fit()
+            {
+                auto ba = Allocator(m_blocks.get_allocator());
+                auto num_blocks_required = calc_num_blocks_for_capacity(m_size);
+                while (m_blocks.size() > num_blocks_required)
+                {
+                    std::allocator_traits<Allocator>::deallocate(ba, m_blocks.back(), num_elements_in_block);
+                    m_blocks.pop_back();
+                }
+                m_blocks.shrink_to_fit();
+            }
+        };
+
+        namespace detail
+        {
+
+            // This is it, the table. Doubles as map and set, and uses `void` for T when its used as a set.
+            template <class Key,
+                      class T, // when void, treat it as a set.
+                      class Hash,
+                      class KeyEqual,
+                      class AllocatorOrContainer,
+                      class Bucket,
+                      bool IsSegmented>
+            class table : public std::conditional_t<is_map_v<T>, base_table_type_map<T>, base_table_type_set>
+            {
+                using underlying_value_type = typename std::conditional_t<is_map_v<T>, std::pair<Key, T>, Key>;
+                using underlying_container_type = std::conditional_t<IsSegmented,
+                                                                     segmented_vector<underlying_value_type, AllocatorOrContainer>,
+                                                                     std::vector<underlying_value_type, AllocatorOrContainer>>;
+
+            public:
+                using value_container_type = std::
+                    conditional_t<is_detected_v<detect_iterator, AllocatorOrContainer>, AllocatorOrContainer, underlying_container_type>;
+
+            private:
+                using bucket_alloc =
+                    typename std::allocator_traits<typename value_container_type::allocator_type>::template rebind_alloc<Bucket>;
+                using bucket_alloc_traits = std::allocator_traits<bucket_alloc>;
+
+                static constexpr uint8_t initial_shifts = 64 - 3; // 2^(64-m_shift) number of buckets
+                static constexpr float default_max_load_factor = 0.8F;
+
+            public:
+                using key_type = Key;
+                using value_type = typename value_container_type::value_type;
+                using size_type = typename value_container_type::size_type;
+                using difference_type = typename value_container_type::difference_type;
+                using hasher = Hash;
+                using key_equal = KeyEqual;
+                using allocator_type = typename value_container_type::allocator_type;
+                using reference = typename value_container_type::reference;
+                using const_reference = typename value_container_type::const_reference;
+                using pointer = typename value_container_type::pointer;
+                using const_pointer = typename value_container_type::const_pointer;
+                using const_iterator = typename value_container_type::const_iterator;
+                using iterator = std::conditional_t<is_map_v<T>, typename value_container_type::iterator, const_iterator>;
+                using bucket_type = Bucket;
+
+            private:
+                using value_idx_type = decltype(Bucket::m_value_idx);
+                using dist_and_fingerprint_type = decltype(Bucket::m_dist_and_fingerprint);
+
+                static_assert(std::is_trivially_destructible_v<Bucket>, "assert there's no need to call destructor / std::destroy");
+                static_assert(std::is_trivially_copyable_v<Bucket>, "assert we can just memset / memcpy");
+
+                value_container_type m_values{}; // Contains all the key-value pairs in one densely stored container. No holes.
+                using bucket_pointer = typename std::allocator_traits<bucket_alloc>::pointer;
+                bucket_pointer m_buckets{};
+                size_t m_num_buckets = 0;
+                size_t m_max_bucket_capacity = 0;
+                float m_max_load_factor = default_max_load_factor;
+                Hash m_hash{};
+                KeyEqual m_equal{};
+                uint8_t m_shifts = initial_shifts;
+
+                [[nodiscard]] auto next(value_idx_type bucket_idx) const -> value_idx_type
+                {
+                    return ANKERL_UNORDERED_DENSE_UNLIKELY(bucket_idx + 1U == m_num_buckets)
+                               ? 0
+                               : static_cast<value_idx_type>(bucket_idx + 1U);
+                }
+
+                // Helper to access bucket through pointer types
+                [[nodiscard]] static constexpr auto at(bucket_pointer bucket_ptr, size_t offset) -> Bucket &
+                {
+                    return *(bucket_ptr + static_cast<typename std::allocator_traits<bucket_alloc>::difference_type>(offset));
+                }
+
+                // use the dist_inc and dist_dec functions so that uint16_t types work without warning
+                [[nodiscard]] static constexpr auto dist_inc(dist_and_fingerprint_type x) -> dist_and_fingerprint_type
+                {
+                    return static_cast<dist_and_fingerprint_type>(x + Bucket::dist_inc);
+                }
+
+                [[nodiscard]] static constexpr auto dist_dec(dist_and_fingerprint_type x) -> dist_and_fingerprint_type
+                {
+                    return static_cast<dist_and_fingerprint_type>(x - Bucket::dist_inc);
+                }
+
+                // The goal of mixed_hash is to always produce a high quality 64bit hash.
+                template <typename K>
+                [[nodiscard]] constexpr auto mixed_hash(K const &key) const -> uint64_t
+                {
+                    if constexpr (is_detected_v<detect_avalanching, Hash>)
+                    {
+                        // we know that the hash is good because is_avalanching.
+                        if constexpr (sizeof(decltype(m_hash(key))) < sizeof(uint64_t))
+                        {
+                            // 32bit hash and is_avalanching => multiply with a constant to avalanche bits upwards
+                            return m_hash(key) * UINT64_C(0x9ddfea08eb382d69);
+                        }
+                        else
+                        {
+                            // 64bit and is_avalanching => only use the hash itself.
+                            return m_hash(key);
+                        }
+                    }
+                    else
+                    {
+                        // not is_avalanching => apply wyhash
+                        return wyhash::hash(m_hash(key));
+                    }
+                }
+
+                [[nodiscard]] constexpr auto dist_and_fingerprint_from_hash(uint64_t hash) const -> dist_and_fingerprint_type
+                {
+                    return Bucket::dist_inc | (static_cast<dist_and_fingerprint_type>(hash) & Bucket::fingerprint_mask);
+                }
+
+                [[nodiscard]] constexpr auto bucket_idx_from_hash(uint64_t hash) const -> value_idx_type
+                {
+                    return static_cast<value_idx_type>(hash >> m_shifts);
+                }
+
+                [[nodiscard]] static constexpr auto get_key(value_type const &vt) -> key_type const &
+                {
+                    if constexpr (is_map_v<T>)
+                    {
+                        return vt.first;
+                    }
+                    else
+                    {
+                        return vt;
+                    }
+                }
+
+                template <typename K>
+                [[nodiscard]] auto next_while_less(K const &key) const -> Bucket
+                {
+                    auto hash = mixed_hash(key);
+                    auto dist_and_fingerprint = dist_and_fingerprint_from_hash(hash);
+                    auto bucket_idx = bucket_idx_from_hash(hash);
+
+                    while (dist_and_fingerprint < at(m_buckets, bucket_idx).m_dist_and_fingerprint)
+                    {
+                        dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+                        bucket_idx = next(bucket_idx);
+                    }
+                    return {dist_and_fingerprint, bucket_idx};
+                }
+
+                void place_and_shift_up(Bucket bucket, value_idx_type place)
+                {
+                    while (0 != at(m_buckets, place).m_dist_and_fingerprint)
+                    {
+                        bucket = std::exchange(at(m_buckets, place), bucket);
+                        bucket.m_dist_and_fingerprint = dist_inc(bucket.m_dist_and_fingerprint);
+                        place = next(place);
+                    }
+                    at(m_buckets, place) = bucket;
+                }
+
+                [[nodiscard]] static constexpr auto calc_num_buckets(uint8_t shifts) -> size_t
+                {
+                    return (std::min)(max_bucket_count(), size_t{1} << (64U - shifts));
+                }
+
+                [[nodiscard]] constexpr auto calc_shifts_for_size(size_t s) const -> uint8_t
+                {
+                    auto shifts = initial_shifts;
+                    while (shifts > 0 && static_cast<size_t>(static_cast<float>(calc_num_buckets(shifts)) * max_load_factor()) < s)
+                    {
+                        --shifts;
+                    }
+                    return shifts;
+                }
+
+                // assumes m_values has data, m_buckets=m_buckets_end=nullptr, m_shifts is INITIAL_SHIFTS
+                void copy_buckets(table const &other)
+                {
+                    if (!empty())
+                    {
+                        m_shifts = other.m_shifts;
+                        allocate_buckets_from_shift();
+                        std::memcpy(m_buckets, other.m_buckets, sizeof(Bucket) * bucket_count());
+                    }
+                }
+
+                /**
+                 * True when no element can be added any more without increasing the size
+                 */
+                [[nodiscard]] auto is_full() const -> bool
+                {
+                    return size() >= m_max_bucket_capacity;
+                }
+
+                void deallocate_buckets()
+                {
+                    auto ba = bucket_alloc(m_values.get_allocator());
+                    if (nullptr != m_buckets)
+                    {
+                        bucket_alloc_traits::deallocate(ba, m_buckets, bucket_count());
+                        m_buckets = nullptr;
+                    }
+                    m_num_buckets = 0;
+                    m_max_bucket_capacity = 0;
+                }
+
+                void allocate_buckets_from_shift()
+                {
+                    auto ba = bucket_alloc(m_values.get_allocator());
+                    m_num_buckets = calc_num_buckets(m_shifts);
+                    m_buckets = bucket_alloc_traits::allocate(ba, m_num_buckets);
+                    if (m_num_buckets == max_bucket_count())
+                    {
+                        // reached the maximum, make sure we can use each bucket
+                        m_max_bucket_capacity = max_bucket_count();
+                    }
+                    else
+                    {
+                        m_max_bucket_capacity = static_cast<value_idx_type>(static_cast<float>(m_num_buckets) * max_load_factor());
+                    }
+                }
+
+                void clear_buckets()
+                {
+                    if (m_buckets != nullptr)
+                    {
+                        std::memset(&*m_buckets, 0, sizeof(Bucket) * bucket_count());
+                    }
+                }
+
+                void clear_and_fill_buckets_from_values()
+                {
+                    clear_buckets();
+                    for (value_idx_type value_idx = 0, end_idx = static_cast<value_idx_type>(m_values.size()); value_idx < end_idx;
+                         ++value_idx)
+                    {
+                        auto const &key = get_key(m_values[value_idx]);
+                        auto [dist_and_fingerprint, bucket] = next_while_less(key);
+
+                        // we know for certain that key has not yet been inserted, so no need to check it.
+                        place_and_shift_up({dist_and_fingerprint, value_idx}, bucket);
+                    }
+                }
+
+                void increase_size()
+                {
+                    if (ANKERL_UNORDERED_DENSE_UNLIKELY(m_max_bucket_capacity == max_bucket_count()))
+                    {
+                        on_error_bucket_overflow();
+                    }
+                    --m_shifts;
+                    deallocate_buckets();
+                    allocate_buckets_from_shift();
+                    clear_and_fill_buckets_from_values();
+                }
+
+                void do_erase(value_idx_type bucket_idx)
+                {
+                    auto const value_idx_to_remove = at(m_buckets, bucket_idx).m_value_idx;
+
+                    // shift down until either empty or an element with correct spot is found
+                    auto next_bucket_idx = next(bucket_idx);
+                    while (at(m_buckets, next_bucket_idx).m_dist_and_fingerprint >= Bucket::dist_inc * 2)
+                    {
+                        at(m_buckets, bucket_idx) = {dist_dec(at(m_buckets, next_bucket_idx).m_dist_and_fingerprint),
+                                                     at(m_buckets, next_bucket_idx).m_value_idx};
+                        bucket_idx = std::exchange(next_bucket_idx, next(next_bucket_idx));
+                    }
+                    at(m_buckets, bucket_idx) = {};
+
+                    // update m_values
+                    if (value_idx_to_remove != m_values.size() - 1)
+                    {
+                        // no luck, we'll have to replace the value with the last one and update the index accordingly
+                        auto &val = m_values[value_idx_to_remove];
+                        val = std::move(m_values.back());
+
+                        // update the values_idx of the moved entry. No need to play the info game, just look until we find the values_idx
+                        auto mh = mixed_hash(get_key(val));
+                        bucket_idx = bucket_idx_from_hash(mh);
+
+                        auto const values_idx_back = static_cast<value_idx_type>(m_values.size() - 1);
+                        while (values_idx_back != at(m_buckets, bucket_idx).m_value_idx)
+                        {
+                            bucket_idx = next(bucket_idx);
+                        }
+                        at(m_buckets, bucket_idx).m_value_idx = value_idx_to_remove;
+                    }
+                    m_values.pop_back();
+                }
+
+                template <typename K>
+                auto do_erase_key(K &&key) -> size_t
+                {
+                    if (empty())
+                    {
+                        return 0;
+                    }
+
+                    auto [dist_and_fingerprint, bucket_idx] = next_while_less(key);
+
+                    while (dist_and_fingerprint == at(m_buckets, bucket_idx).m_dist_and_fingerprint &&
+                           !m_equal(key, get_key(m_values[at(m_buckets, bucket_idx).m_value_idx])))
+                    {
+                        dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+                        bucket_idx = next(bucket_idx);
+                    }
+
+                    if (dist_and_fingerprint != at(m_buckets, bucket_idx).m_dist_and_fingerprint)
+                    {
+                        return 0;
+                    }
+                    do_erase(bucket_idx);
+                    return 1;
+                }
+
+                template <class K, class M>
+                auto do_insert_or_assign(K &&key, M &&mapped) -> std::pair<iterator, bool>
+                {
+                    auto it_isinserted = try_emplace(std::forward<K>(key), std::forward<M>(mapped));
+                    if (!it_isinserted.second)
+                    {
+                        it_isinserted.first->second = std::forward<M>(mapped);
+                    }
+                    return it_isinserted;
+                }
+
+                template <typename K, typename... Args>
+                auto do_place_element(dist_and_fingerprint_type dist_and_fingerprint, value_idx_type bucket_idx, K &&key, Args &&...args)
+                    -> std::pair<iterator, bool>
+                {
+
+                    // emplace the new value. If that throws an exception, no harm done; index is still in a valid state
+                    m_values.emplace_back(std::piecewise_construct,
+                                          std::forward_as_tuple(std::forward<K>(key)),
+                                          std::forward_as_tuple(std::forward<Args>(args)...));
+
+                    // place element and shift up until we find an empty spot
+                    auto value_idx = static_cast<value_idx_type>(m_values.size() - 1);
+                    place_and_shift_up({dist_and_fingerprint, value_idx}, bucket_idx);
+                    return {begin() + static_cast<difference_type>(value_idx), true};
+                }
+
+                template <typename K, typename... Args>
+                auto do_try_emplace(K &&key, Args &&...args) -> std::pair<iterator, bool>
+                {
+                    if (ANKERL_UNORDERED_DENSE_UNLIKELY(is_full()))
+                    {
+                        increase_size();
+                    }
+
+                    auto hash = mixed_hash(key);
+                    auto dist_and_fingerprint = dist_and_fingerprint_from_hash(hash);
+                    auto bucket_idx = bucket_idx_from_hash(hash);
+
+                    while (true)
+                    {
+                        auto *bucket = &at(m_buckets, bucket_idx);
+                        if (dist_and_fingerprint == bucket->m_dist_and_fingerprint)
+                        {
+                            if (m_equal(key, get_key(m_values[bucket->m_value_idx])))
+                            {
+                                return {begin() + static_cast<difference_type>(bucket->m_value_idx), false};
+                            }
+                        }
+                        else if (dist_and_fingerprint > bucket->m_dist_and_fingerprint)
+                        {
+                            return do_place_element(dist_and_fingerprint, bucket_idx, std::forward<K>(key), std::forward<Args>(args)...);
+                        }
+                        dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+                        bucket_idx = next(bucket_idx);
+                    }
+                }
+
+                template <typename K>
+                auto do_find(K const &key) -> iterator
+                {
+                    if (ANKERL_UNORDERED_DENSE_UNLIKELY(empty()))
+                    {
+                        return end();
+                    }
+
+                    auto mh = mixed_hash(key);
+                    auto dist_and_fingerprint = dist_and_fingerprint_from_hash(mh);
+                    auto bucket_idx = bucket_idx_from_hash(mh);
+                    auto *bucket = &at(m_buckets, bucket_idx);
+
+                    // unrolled loop. *Always* check a few directly, then enter the loop. This is faster.
+                    if (dist_and_fingerprint == bucket->m_dist_and_fingerprint && m_equal(key, get_key(m_values[bucket->m_value_idx])))
+                    {
+                        return begin() + static_cast<difference_type>(bucket->m_value_idx);
+                    }
+                    dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+                    bucket_idx = next(bucket_idx);
+                    bucket = &at(m_buckets, bucket_idx);
+
+                    if (dist_and_fingerprint == bucket->m_dist_and_fingerprint && m_equal(key, get_key(m_values[bucket->m_value_idx])))
+                    {
+                        return begin() + static_cast<difference_type>(bucket->m_value_idx);
+                    }
+                    dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+                    bucket_idx = next(bucket_idx);
+                    bucket = &at(m_buckets, bucket_idx);
+
+                    while (true)
+                    {
+                        if (dist_and_fingerprint == bucket->m_dist_and_fingerprint)
+                        {
+                            if (m_equal(key, get_key(m_values[bucket->m_value_idx])))
+                            {
+                                return begin() + static_cast<difference_type>(bucket->m_value_idx);
+                            }
+                        }
+                        else if (dist_and_fingerprint > bucket->m_dist_and_fingerprint)
+                        {
+                            return end();
+                        }
+                        dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+                        bucket_idx = next(bucket_idx);
+                        bucket = &at(m_buckets, bucket_idx);
+                    }
+                }
+
+                template <typename K>
+                auto do_find(K const &key) const -> const_iterator
+                {
+                    return const_cast<table *>(this)->do_find(key); // NOLINT(cppcoreguidelines-pro-type-const-cast)
+                }
+
+                template <typename K, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
+                auto do_at(K const &key) -> Q &
+                {
+                    if (auto it = find(key); ANKERL_UNORDERED_DENSE_LIKELY(end() != it))
+                    {
+                        return it->second;
+                    }
+                    on_error_key_not_found();
+                }
+
+                template <typename K, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
+                auto do_at(K const &key) const -> Q const &
+                {
+                    return const_cast<table *>(this)->at(key); // NOLINT(cppcoreguidelines-pro-type-const-cast)
+                }
+
+            public:
+                table()
+                    : table(0) {}
+
+                explicit table(size_t bucket_count,
+                               Hash const &hash = Hash(),
+                               KeyEqual const &equal = KeyEqual(),
+                               allocator_type const &alloc_or_container = allocator_type())
+                    : m_values(alloc_or_container), m_hash(hash), m_equal(equal)
+                {
+                    if (0 != bucket_count)
+                    {
+                        reserve(bucket_count);
+                    }
+                }
+
+                table(size_t bucket_count, allocator_type const &alloc)
+                    : table(bucket_count, Hash(), KeyEqual(), alloc) {}
+
+                table(size_t bucket_count, Hash const &hash, allocator_type const &alloc)
+                    : table(bucket_count, hash, KeyEqual(), alloc) {}
+
+                explicit table(allocator_type const &alloc)
+                    : table(0, Hash(), KeyEqual(), alloc) {}
+
+                template <class InputIt>
+                table(InputIt first,
+                      InputIt last,
+                      size_type bucket_count = 0,
+                      Hash const &hash = Hash(),
+                      KeyEqual const &equal = KeyEqual(),
+                      allocator_type const &alloc = allocator_type())
+                    : table(bucket_count, hash, equal, alloc)
+                {
+                    insert(first, last);
+                }
+
+                template <class InputIt>
+                table(InputIt first, InputIt last, size_type bucket_count, allocator_type const &alloc)
+                    : table(first, last, bucket_count, Hash(), KeyEqual(), alloc) {}
+
+                template <class InputIt>
+                table(InputIt first, InputIt last, size_type bucket_count, Hash const &hash, allocator_type const &alloc)
+                    : table(first, last, bucket_count, hash, KeyEqual(), alloc) {}
+
+                table(table const &other)
+                    : table(other, other.m_values.get_allocator()) {}
+
+                table(table const &other, allocator_type const &alloc)
+                    : m_values(other.m_values, alloc), m_max_load_factor(other.m_max_load_factor), m_hash(other.m_hash), m_equal(other.m_equal)
+                {
+                    copy_buckets(other);
+                }
+
+                table(table &&other) noexcept
+                    : table(std::move(other), other.m_values.get_allocator()) {}
+
+                table(table &&other, allocator_type const &alloc) noexcept
+                    : m_values(alloc)
+                {
+                    *this = std::move(other);
+                }
+
+                table(std::initializer_list<value_type> ilist,
+                      size_t bucket_count = 0,
+                      Hash const &hash = Hash(),
+                      KeyEqual const &equal = KeyEqual(),
+                      allocator_type const &alloc = allocator_type())
+                    : table(bucket_count, hash, equal, alloc)
+                {
+                    insert(ilist);
+                }
+
+                table(std::initializer_list<value_type> ilist, size_type bucket_count, allocator_type const &alloc)
+                    : table(ilist, bucket_count, Hash(), KeyEqual(), alloc) {}
+
+                table(std::initializer_list<value_type> init, size_type bucket_count, Hash const &hash, allocator_type const &alloc)
+                    : table(init, bucket_count, hash, KeyEqual(), alloc) {}
+
+                ~table()
+                {
+                    if (nullptr != m_buckets)
+                    {
+                        auto ba = bucket_alloc(m_values.get_allocator());
+                        bucket_alloc_traits::deallocate(ba, m_buckets, bucket_count());
+                    }
+                }
+
+                auto operator=(table const &other) -> table &
+                {
+                    if (&other != this)
+                    {
+                        deallocate_buckets(); // deallocate before m_values is set (might have another allocator)
+                        m_values = other.m_values;
+                        m_max_load_factor = other.m_max_load_factor;
+                        m_hash = other.m_hash;
+                        m_equal = other.m_equal;
+                        m_shifts = initial_shifts;
+                        copy_buckets(other);
+                    }
+                    return *this;
+                }
+
+                auto operator=(table &&other) noexcept(
+                    noexcept(std::is_nothrow_move_assignable_v<value_container_type> && std::is_nothrow_move_assignable_v<Hash> &&
+                             std::is_nothrow_move_assignable_v<KeyEqual>)) -> table &
+                {
+                    if (&other != this)
+                    {
+                        deallocate_buckets(); // deallocate before m_values is set (might have another allocator)
+                        m_values = std::move(other.m_values);
+                        other.m_values.clear();
+
+                        // we can only reuse m_buckets when both maps have the same allocator!
+                        if (get_allocator() == other.get_allocator())
+                        {
+                            m_buckets = std::exchange(other.m_buckets, nullptr);
+                            m_num_buckets = std::exchange(other.m_num_buckets, 0);
+                            m_max_bucket_capacity = std::exchange(other.m_max_bucket_capacity, 0);
+                            m_shifts = std::exchange(other.m_shifts, initial_shifts);
+                            m_max_load_factor = std::exchange(other.m_max_load_factor, default_max_load_factor);
+                            m_hash = std::exchange(other.m_hash, {});
+                            m_equal = std::exchange(other.m_equal, {});
+                        }
+                        else
+                        {
+                            // set max_load_factor *before* copying the other's buckets, so we have the same
+                            // behavior
+                            m_max_load_factor = other.m_max_load_factor;
+
+                            // copy_buckets sets m_buckets, m_num_buckets, m_max_bucket_capacity, m_shifts
+                            copy_buckets(other);
+                            // clear's the other's buckets so other is now already usable.
+                            other.clear_buckets();
+                            m_hash = other.m_hash;
+                            m_equal = other.m_equal;
+                        }
+                        // map "other" is now already usable, it's empty.
+                    }
+                    return *this;
+                }
+
+                auto operator=(std::initializer_list<value_type> ilist) -> table &
+                {
+                    clear();
+                    insert(ilist);
+                    return *this;
+                }
+
+                auto get_allocator() const noexcept -> allocator_type
+                {
+                    return m_values.get_allocator();
+                }
+
+                // iterators //////////////////////////////////////////////////////////////
+
+                auto begin() noexcept -> iterator
+                {
+                    return m_values.begin();
+                }
+
+                auto begin() const noexcept -> const_iterator
+                {
+                    return m_values.begin();
+                }
+
+                auto cbegin() const noexcept -> const_iterator
+                {
+                    return m_values.cbegin();
+                }
+
+                auto end() noexcept -> iterator
+                {
+                    return m_values.end();
+                }
+
+                auto cend() const noexcept -> const_iterator
+                {
+                    return m_values.cend();
+                }
+
+                auto end() const noexcept -> const_iterator
+                {
+                    return m_values.end();
+                }
+
+                // capacity ///////////////////////////////////////////////////////////////
+
+                [[nodiscard]] auto empty() const noexcept -> bool
+                {
+                    return m_values.empty();
+                }
+
+                [[nodiscard]] auto size() const noexcept -> size_t
+                {
+                    return m_values.size();
+                }
+
+                [[nodiscard]] static constexpr auto max_size() noexcept -> size_t
+                {
+                    if constexpr ((std::numeric_limits<value_idx_type>::max)() == (std::numeric_limits<size_t>::max)())
+                    {
+                        return size_t{1} << (sizeof(value_idx_type) * 8 - 1);
+                    }
+                    else
+                    {
+                        return size_t{1} << (sizeof(value_idx_type) * 8);
+                    }
+                }
+
+                // modifiers //////////////////////////////////////////////////////////////
+
+                void clear()
+                {
+                    m_values.clear();
+                    clear_buckets();
+                }
+
+                auto insert(value_type const &value) -> std::pair<iterator, bool>
+                {
+                    return emplace(value);
+                }
+
+                auto insert(value_type &&value) -> std::pair<iterator, bool>
+                {
+                    return emplace(std::move(value));
+                }
+
+                template <class P, std::enable_if_t<std::is_constructible_v<value_type, P &&>, bool> = true>
+                auto insert(P &&value) -> std::pair<iterator, bool>
+                {
+                    return emplace(std::forward<P>(value));
+                }
+
+                auto insert(const_iterator /*hint*/, value_type const &value) -> iterator
+                {
+                    return insert(value).first;
+                }
+
+                auto insert(const_iterator /*hint*/, value_type &&value) -> iterator
+                {
+                    return insert(std::move(value)).first;
+                }
+
+                template <class P, std::enable_if_t<std::is_constructible_v<value_type, P &&>, bool> = true>
+                auto insert(const_iterator /*hint*/, P &&value) -> iterator
+                {
+                    return insert(std::forward<P>(value)).first;
+                }
+
+                template <class InputIt>
+                void insert(InputIt first, InputIt last)
+                {
+                    while (first != last)
+                    {
+                        insert(*first);
+                        ++first;
+                    }
+                }
+
+                void insert(std::initializer_list<value_type> ilist)
+                {
+                    insert(ilist.begin(), ilist.end());
+                }
+
+                // nonstandard API: *this is emptied.
+                // Also see "A Standard flat_map" https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2022/p0429r9.pdf
+                auto extract() && -> value_container_type
+                {
+                    return std::move(m_values);
+                }
+
+                // nonstandard API:
+                // Discards the internally held container and replaces it with the one passed. Erases non-unique elements.
+                auto replace(value_container_type &&container)
+                {
+                    if (ANKERL_UNORDERED_DENSE_UNLIKELY(container.size() > max_size()))
+                    {
+                        on_error_too_many_elements();
+                    }
+                    auto shifts = calc_shifts_for_size(container.size());
+                    if (0 == m_num_buckets || shifts < m_shifts || container.get_allocator() != m_values.get_allocator())
+                    {
+                        m_shifts = shifts;
+                        deallocate_buckets();
+                        allocate_buckets_from_shift();
+                    }
+                    clear_buckets();
+
+                    m_values = std::move(container);
+
+                    // can't use clear_and_fill_buckets_from_values() because container elements might not be unique
+                    auto value_idx = value_idx_type{};
+
+                    // loop until we reach the end of the container. duplicated entries will be replaced with back().
+                    while (value_idx != static_cast<value_idx_type>(m_values.size()))
+                    {
+                        auto const &key = get_key(m_values[value_idx]);
+
+                        auto hash = mixed_hash(key);
+                        auto dist_and_fingerprint = dist_and_fingerprint_from_hash(hash);
+                        auto bucket_idx = bucket_idx_from_hash(hash);
+
+                        bool key_found = false;
+                        while (true)
+                        {
+                            auto const &bucket = at(m_buckets, bucket_idx);
+                            if (dist_and_fingerprint > bucket.m_dist_and_fingerprint)
+                            {
+                                break;
+                            }
+                            if (dist_and_fingerprint == bucket.m_dist_and_fingerprint &&
+                                m_equal(key, get_key(m_values[bucket.m_value_idx])))
+                            {
+                                key_found = true;
+                                break;
+                            }
+                            dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+                            bucket_idx = next(bucket_idx);
+                        }
+
+                        if (key_found)
+                        {
+                            if (value_idx != static_cast<value_idx_type>(m_values.size() - 1))
+                            {
+                                m_values[value_idx] = std::move(m_values.back());
+                            }
+                            m_values.pop_back();
+                        }
+                        else
+                        {
+                            place_and_shift_up({dist_and_fingerprint, value_idx}, bucket_idx);
+                            ++value_idx;
+                        }
+                    }
+                }
+
+                template <class M, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
+                auto insert_or_assign(Key const &key, M &&mapped) -> std::pair<iterator, bool>
+                {
+                    return do_insert_or_assign(key, std::forward<M>(mapped));
+                }
+
+                template <class M, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
+                auto insert_or_assign(Key &&key, M &&mapped) -> std::pair<iterator, bool>
+                {
+                    return do_insert_or_assign(std::move(key), std::forward<M>(mapped));
+                }
+
+                template <typename K,
+                          typename M,
+                          typename Q = T,
+                          typename H = Hash,
+                          typename KE = KeyEqual,
+                          std::enable_if_t<is_map_v<Q> && is_transparent_v<H, KE>, bool> = true>
+                auto insert_or_assign(K &&key, M &&mapped) -> std::pair<iterator, bool>
+                {
+                    return do_insert_or_assign(std::forward<K>(key), std::forward<M>(mapped));
+                }
+
+                template <class M, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
+                auto insert_or_assign(const_iterator /*hint*/, Key const &key, M &&mapped) -> iterator
+                {
+                    return do_insert_or_assign(key, std::forward<M>(mapped)).first;
+                }
+
+                template <class M, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
+                auto insert_or_assign(const_iterator /*hint*/, Key &&key, M &&mapped) -> iterator
+                {
+                    return do_insert_or_assign(std::move(key), std::forward<M>(mapped)).first;
+                }
+
+                template <typename K,
+                          typename M,
+                          typename Q = T,
+                          typename H = Hash,
+                          typename KE = KeyEqual,
+                          std::enable_if_t<is_map_v<Q> && is_transparent_v<H, KE>, bool> = true>
+                auto insert_or_assign(const_iterator /*hint*/, K &&key, M &&mapped) -> iterator
+                {
+                    return do_insert_or_assign(std::forward<K>(key), std::forward<M>(mapped)).first;
+                }
+
+                // Single arguments for unordered_set can be used without having to construct the value_type
+                template <class K,
+                          typename Q = T,
+                          typename H = Hash,
+                          typename KE = KeyEqual,
+                          std::enable_if_t<!is_map_v<Q> && is_transparent_v<H, KE>, bool> = true>
+                auto emplace(K &&key) -> std::pair<iterator, bool>
+                {
+                    if (is_full())
+                    {
+                        increase_size();
+                    }
+
+                    auto hash = mixed_hash(key);
+                    auto dist_and_fingerprint = dist_and_fingerprint_from_hash(hash);
+                    auto bucket_idx = bucket_idx_from_hash(hash);
+
+                    while (dist_and_fingerprint <= at(m_buckets, bucket_idx).m_dist_and_fingerprint)
+                    {
+                        if (dist_and_fingerprint == at(m_buckets, bucket_idx).m_dist_and_fingerprint &&
+                            m_equal(key, m_values[at(m_buckets, bucket_idx).m_value_idx]))
+                        {
+                            // found it, return without ever actually creating anything
+                            return {begin() + static_cast<difference_type>(at(m_buckets, bucket_idx).m_value_idx), false};
+                        }
+                        dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+                        bucket_idx = next(bucket_idx);
+                    }
+
+                    // value is new, insert element first, so when exception happens we are in a valid state
+                    m_values.emplace_back(std::forward<K>(key));
+                    // now place the bucket and shift up until we find an empty spot
+                    auto value_idx = static_cast<value_idx_type>(m_values.size() - 1);
+                    place_and_shift_up({dist_and_fingerprint, value_idx}, bucket_idx);
+                    return {begin() + static_cast<difference_type>(value_idx), true};
+                }
+
+                template <class... Args>
+                auto emplace(Args &&...args) -> std::pair<iterator, bool>
+                {
+                    if (is_full())
+                    {
+                        increase_size();
+                    }
+
+                    // we have to instantiate the value_type to be able to access the key.
+                    // 1. emplace_back the object so it is constructed. 2. If the key is already there, pop it later in the loop.
+                    auto &key = get_key(m_values.emplace_back(std::forward<Args>(args)...));
+                    auto hash = mixed_hash(key);
+                    auto dist_and_fingerprint = dist_and_fingerprint_from_hash(hash);
+                    auto bucket_idx = bucket_idx_from_hash(hash);
+
+                    while (dist_and_fingerprint <= at(m_buckets, bucket_idx).m_dist_and_fingerprint)
+                    {
+                        if (dist_and_fingerprint == at(m_buckets, bucket_idx).m_dist_and_fingerprint &&
+                            m_equal(key, get_key(m_values[at(m_buckets, bucket_idx).m_value_idx])))
+                        {
+                            m_values.pop_back(); // value was already there, so get rid of it
+                            return {begin() + static_cast<difference_type>(at(m_buckets, bucket_idx).m_value_idx), false};
+                        }
+                        dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+                        bucket_idx = next(bucket_idx);
+                    }
+
+                    // value is new, place the bucket and shift up until we find an empty spot
+                    auto value_idx = static_cast<value_idx_type>(m_values.size() - 1);
+                    place_and_shift_up({dist_and_fingerprint, value_idx}, bucket_idx);
+
+                    return {begin() + static_cast<difference_type>(value_idx), true};
+                }
+
+                template <class... Args>
+                auto emplace_hint(const_iterator /*hint*/, Args &&...args) -> iterator
+                {
+                    return emplace(std::forward<Args>(args)...).first;
+                }
+
+                template <class... Args, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
+                auto try_emplace(Key const &key, Args &&...args) -> std::pair<iterator, bool>
+                {
+                    return do_try_emplace(key, std::forward<Args>(args)...);
+                }
+
+                template <class... Args, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
+                auto try_emplace(Key &&key, Args &&...args) -> std::pair<iterator, bool>
+                {
+                    return do_try_emplace(std::move(key), std::forward<Args>(args)...);
+                }
+
+                template <class... Args, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
+                auto try_emplace(const_iterator /*hint*/, Key const &key, Args &&...args) -> iterator
+                {
+                    return do_try_emplace(key, std::forward<Args>(args)...).first;
+                }
+
+                template <class... Args, typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
+                auto try_emplace(const_iterator /*hint*/, Key &&key, Args &&...args) -> iterator
+                {
+                    return do_try_emplace(std::move(key), std::forward<Args>(args)...).first;
+                }
+
+                template <
+                    typename K,
+                    typename... Args,
+                    typename Q = T,
+                    typename H = Hash,
+                    typename KE = KeyEqual,
+                    std::enable_if_t<is_map_v<Q> && is_transparent_v<H, KE> && is_neither_convertible_v<K &&, iterator, const_iterator>,
+                                     bool> = true>
+                auto try_emplace(K &&key, Args &&...args) -> std::pair<iterator, bool>
+                {
+                    return do_try_emplace(std::forward<K>(key), std::forward<Args>(args)...);
+                }
+
+                template <
+                    typename K,
+                    typename... Args,
+                    typename Q = T,
+                    typename H = Hash,
+                    typename KE = KeyEqual,
+                    std::enable_if_t<is_map_v<Q> && is_transparent_v<H, KE> && is_neither_convertible_v<K &&, iterator, const_iterator>,
+                                     bool> = true>
+                auto try_emplace(const_iterator /*hint*/, K &&key, Args &&...args) -> iterator
+                {
+                    return do_try_emplace(std::forward<K>(key), std::forward<Args>(args)...).first;
+                }
+
+                auto erase(iterator it) -> iterator
+                {
+                    auto hash = mixed_hash(get_key(*it));
+                    auto bucket_idx = bucket_idx_from_hash(hash);
+
+                    auto const value_idx_to_remove = static_cast<value_idx_type>(it - cbegin());
+                    while (at(m_buckets, bucket_idx).m_value_idx != value_idx_to_remove)
+                    {
+                        bucket_idx = next(bucket_idx);
+                    }
+
+                    do_erase(bucket_idx);
+                    return begin() + static_cast<difference_type>(value_idx_to_remove);
+                }
+
+                template <typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
+                auto erase(const_iterator it) -> iterator
+                {
+                    return erase(begin() + (it - cbegin()));
+                }
+
+                auto erase(const_iterator first, const_iterator last) -> iterator
+                {
+                    auto const idx_first = first - cbegin();
+                    auto const idx_last = last - cbegin();
+                    auto const first_to_last = std::distance(first, last);
+                    auto const last_to_end = std::distance(last, cend());
+
+                    // remove elements from left to right which moves elements from the end back
+                    auto const mid = idx_first + (std::min)(first_to_last, last_to_end);
+                    auto idx = idx_first;
+                    while (idx != mid)
+                    {
+                        erase(begin() + idx);
+                        ++idx;
+                    }
+
+                    // all elements from the right are moved, now remove the last element until all done
+                    idx = idx_last;
+                    while (idx != mid)
+                    {
+                        --idx;
+                        erase(begin() + idx);
+                    }
+
+                    return begin() + idx_first;
+                }
+
+                auto erase(Key const &key) -> size_t
+                {
+                    return do_erase_key(key);
+                }
+
+                template <class K, class H = Hash, class KE = KeyEqual, std::enable_if_t<is_transparent_v<H, KE>, bool> = true>
+                auto erase(K &&key) -> size_t
+                {
+                    return do_erase_key(std::forward<K>(key));
+                }
+
+                void swap(table &other) noexcept(noexcept(std::is_nothrow_swappable_v<value_container_type> &&
+                                                          std::is_nothrow_swappable_v<Hash> && std::is_nothrow_swappable_v<KeyEqual>))
+                {
+                    using std::swap;
+                    swap(other, *this);
+                }
+
+                // lookup /////////////////////////////////////////////////////////////////
+
+                template <typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
+                auto at(key_type const &key) -> Q &
+                {
+                    return do_at(key);
+                }
+
+                template <typename K,
+                          typename Q = T,
+                          typename H = Hash,
+                          typename KE = KeyEqual,
+                          std::enable_if_t<is_map_v<Q> && is_transparent_v<H, KE>, bool> = true>
+                auto at(K const &key) -> Q &
+                {
+                    return do_at(key);
+                }
+
+                template <typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
+                auto at(key_type const &key) const -> Q const &
+                {
+                    return do_at(key);
+                }
+
+                template <typename K,
+                          typename Q = T,
+                          typename H = Hash,
+                          typename KE = KeyEqual,
+                          std::enable_if_t<is_map_v<Q> && is_transparent_v<H, KE>, bool> = true>
+                auto at(K const &key) const -> Q const &
+                {
+                    return do_at(key);
+                }
+
+                template <typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
+                auto operator[](Key const &key) -> Q &
+                {
+                    return try_emplace(key).first->second;
+                }
+
+                template <typename Q = T, std::enable_if_t<is_map_v<Q>, bool> = true>
+                auto operator[](Key &&key) -> Q &
+                {
+                    return try_emplace(std::move(key)).first->second;
+                }
+
+                template <typename K,
+                          typename Q = T,
+                          typename H = Hash,
+                          typename KE = KeyEqual,
+                          std::enable_if_t<is_map_v<Q> && is_transparent_v<H, KE>, bool> = true>
+                auto operator[](K &&key) -> Q &
+                {
+                    return try_emplace(std::forward<K>(key)).first->second;
+                }
+
+                auto count(Key const &key) const -> size_t
+                {
+                    return find(key) == end() ? 0 : 1;
+                }
+
+                template <class K, class H = Hash, class KE = KeyEqual, std::enable_if_t<is_transparent_v<H, KE>, bool> = true>
+                auto count(K const &key) const -> size_t
+                {
+                    return find(key) == end() ? 0 : 1;
+                }
+
+                auto find(Key const &key) -> iterator
+                {
+                    return do_find(key);
+                }
+
+                auto find(Key const &key) const -> const_iterator
+                {
+                    return do_find(key);
+                }
+
+                template <class K, class H = Hash, class KE = KeyEqual, std::enable_if_t<is_transparent_v<H, KE>, bool> = true>
+                auto find(K const &key) -> iterator
+                {
+                    return do_find(key);
+                }
+
+                template <class K, class H = Hash, class KE = KeyEqual, std::enable_if_t<is_transparent_v<H, KE>, bool> = true>
+                auto find(K const &key) const -> const_iterator
+                {
+                    return do_find(key);
+                }
+
+                auto contains(Key const &key) const -> bool
+                {
+                    return find(key) != end();
+                }
+
+                template <class K, class H = Hash, class KE = KeyEqual, std::enable_if_t<is_transparent_v<H, KE>, bool> = true>
+                auto contains(K const &key) const -> bool
+                {
+                    return find(key) != end();
+                }
+
+                auto equal_range(Key const &key) -> std::pair<iterator, iterator>
+                {
+                    auto it = do_find(key);
+                    return {it, it == end() ? end() : it + 1};
+                }
+
+                auto equal_range(const Key &key) const -> std::pair<const_iterator, const_iterator>
+                {
+                    auto it = do_find(key);
+                    return {it, it == end() ? end() : it + 1};
+                }
+
+                template <class K, class H = Hash, class KE = KeyEqual, std::enable_if_t<is_transparent_v<H, KE>, bool> = true>
+                auto equal_range(K const &key) -> std::pair<iterator, iterator>
+                {
+                    auto it = do_find(key);
+                    return {it, it == end() ? end() : it + 1};
+                }
+
+                template <class K, class H = Hash, class KE = KeyEqual, std::enable_if_t<is_transparent_v<H, KE>, bool> = true>
+                auto equal_range(K const &key) const -> std::pair<const_iterator, const_iterator>
+                {
+                    auto it = do_find(key);
+                    return {it, it == end() ? end() : it + 1};
+                }
+
+                // bucket interface ///////////////////////////////////////////////////////
+
+                auto bucket_count() const noexcept -> size_t
+                { // NOLINT(modernize-use-nodiscard)
+                    return m_num_buckets;
+                }
+
+                static constexpr auto max_bucket_count() noexcept -> size_t
+                { // NOLINT(modernize-use-nodiscard)
+                    return max_size();
+                }
+
+                // hash policy ////////////////////////////////////////////////////////////
+
+                [[nodiscard]] auto load_factor() const -> float
+                {
+                    return bucket_count() ? static_cast<float>(size()) / static_cast<float>(bucket_count()) : 0.0F;
+                }
+
+                [[nodiscard]] auto max_load_factor() const -> float
+                {
+                    return m_max_load_factor;
+                }
+
+                void max_load_factor(float ml)
+                {
+                    m_max_load_factor = ml;
+                    if (m_num_buckets != max_bucket_count())
+                    {
+                        m_max_bucket_capacity = static_cast<value_idx_type>(static_cast<float>(bucket_count()) * max_load_factor());
+                    }
+                }
+
+                void rehash(size_t count)
+                {
+                    count = (std::min)(count, max_size());
+                    auto shifts = calc_shifts_for_size((std::max)(count, size()));
+                    if (shifts != m_shifts)
+                    {
+                        m_shifts = shifts;
+                        deallocate_buckets();
+                        m_values.shrink_to_fit();
+                        allocate_buckets_from_shift();
+                        clear_and_fill_buckets_from_values();
+                    }
+                }
+
+                void reserve(size_t capa)
+                {
+                    capa = (std::min)(capa, max_size());
+                    if constexpr (has_reserve<value_container_type>)
+                    {
+                        // std::deque doesn't have reserve(). Make sure we only call when available
+                        m_values.reserve(capa);
+                    }
+                    auto shifts = calc_shifts_for_size((std::max)(capa, size()));
+                    if (0 == m_num_buckets || shifts < m_shifts)
+                    {
+                        m_shifts = shifts;
+                        deallocate_buckets();
+                        allocate_buckets_from_shift();
+                        clear_and_fill_buckets_from_values();
+                    }
+                }
+
+                // observers //////////////////////////////////////////////////////////////
+
+                auto hash_function() const -> hasher
+                {
+                    return m_hash;
+                }
+
+                auto key_eq() const -> key_equal
+                {
+                    return m_equal;
+                }
+
+                // nonstandard API: expose the underlying values container
+                [[nodiscard]] auto values() const noexcept -> value_container_type const &
+                {
+                    return m_values;
+                }
+
+                // non-member functions ///////////////////////////////////////////////////
+
+                friend auto operator==(table const &a, table const &b) -> bool
+                {
+                    if (&a == &b)
+                    {
+                        return true;
+                    }
+                    if (a.size() != b.size())
+                    {
+                        return false;
+                    }
+                    for (auto const &b_entry : b)
+                    {
+                        auto it = a.find(get_key(b_entry));
+                        if constexpr (is_map_v<T>)
+                        {
+                            // map: check that key is here, then also check that value is the same
+                            if (a.end() == it || !(b_entry.second == it->second))
+                            {
+                                return false;
+                            }
+                        }
+                        else
+                        {
+                            // set: only check that the key is here
+                            if (a.end() == it)
+                            {
+                                return false;
+                            }
+                        }
+                    }
+                    return true;
+                }
+
+                friend auto operator!=(table const &a, table const &b) -> bool
+                {
+                    return !(a == b);
+                }
+            };
+
+        } // namespace detail
+
+        ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
+                                                class T,
+                                                class Hash = hash<Key>,
+                                                class KeyEqual = std::equal_to<Key>,
+                                                class AllocatorOrContainer = std::allocator<std::pair<Key, T>>,
+                                                class Bucket = bucket_type::standard>
+        using map = detail::table<Key, T, Hash, KeyEqual, AllocatorOrContainer, Bucket, false>;
+
+        ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
+                                                class T,
+                                                class Hash = hash<Key>,
+                                                class KeyEqual = std::equal_to<Key>,
+                                                class AllocatorOrContainer = std::allocator<std::pair<Key, T>>,
+                                                class Bucket = bucket_type::standard>
+        using segmented_map = detail::table<Key, T, Hash, KeyEqual, AllocatorOrContainer, Bucket, true>;
+
+        ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
+                                                class Hash = hash<Key>,
+                                                class KeyEqual = std::equal_to<Key>,
+                                                class AllocatorOrContainer = std::allocator<Key>,
+                                                class Bucket = bucket_type::standard>
+        using set = detail::table<Key, void, Hash, KeyEqual, AllocatorOrContainer, Bucket, false>;
+
+        ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
+                                                class Hash = hash<Key>,
+                                                class KeyEqual = std::equal_to<Key>,
+                                                class AllocatorOrContainer = std::allocator<Key>,
+                                                class Bucket = bucket_type::standard>
+        using segmented_set = detail::table<Key, void, Hash, KeyEqual, AllocatorOrContainer, Bucket, true>;
+
+#if defined(ANKERL_UNORDERED_DENSE_PMR)
+
+        namespace pmr
+        {
+
+            ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
+                                                    class T,
+                                                    class Hash = hash<Key>,
+                                                    class KeyEqual = std::equal_to<Key>,
+                                                    class Bucket = bucket_type::standard>
+            using map =
+                detail::table<Key, T, Hash, KeyEqual, ANKERL_UNORDERED_DENSE_PMR::polymorphic_allocator<std::pair<Key, T>>, Bucket, false>;
+
+            ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
+                                                    class T,
+                                                    class Hash = hash<Key>,
+                                                    class KeyEqual = std::equal_to<Key>,
+                                                    class Bucket = bucket_type::standard>
+            using segmented_map =
+                detail::table<Key, T, Hash, KeyEqual, ANKERL_UNORDERED_DENSE_PMR::polymorphic_allocator<std::pair<Key, T>>, Bucket, true>;
+
+            ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
+                                                    class Hash = hash<Key>,
+                                                    class KeyEqual = std::equal_to<Key>,
+                                                    class Bucket = bucket_type::standard>
+            using set = detail::table<Key, void, Hash, KeyEqual, ANKERL_UNORDERED_DENSE_PMR::polymorphic_allocator<Key>, Bucket, false>;
+
+            ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
+                                                    class Hash = hash<Key>,
+                                                    class KeyEqual = std::equal_to<Key>,
+                                                    class Bucket = bucket_type::standard>
+            using segmented_set =
+                detail::table<Key, void, Hash, KeyEqual, ANKERL_UNORDERED_DENSE_PMR::polymorphic_allocator<Key>, Bucket, true>;
+
+        } // namespace pmr
+
+#endif
+
+        // deduction guides ///////////////////////////////////////////////////////////
+
+        // deduction guides for alias templates are only possible since C++20
+        // see https://en.cppreference.com/w/cpp/language/class_template_argument_deduction
+
+    } // namespace ANKERL_UNORDERED_DENSE_NAMESPACE
+} // namespace ankerl::unordered_dense
+
+// std extensions /////////////////////////////////////////////////////////////
+
+namespace std
+{ // NOLINT(cert-dcl58-cpp)
+
+    ANKERL_UNORDERED_DENSE_EXPORT template <class Key,
+                                            class T,
+                                            class Hash,
+                                            class KeyEqual,
+                                            class AllocatorOrContainer,
+                                            class Bucket,
+                                            class Pred,
+                                            bool IsSegmented>
+    // NOLINTNEXTLINE(cert-dcl58-cpp)
+    auto erase_if(ankerl::unordered_dense::detail::table<Key, T, Hash, KeyEqual, AllocatorOrContainer, Bucket, IsSegmented> &map,
+                  Pred pred) -> size_t
+    {
+        using map_t = ankerl::unordered_dense::detail::table<Key, T, Hash, KeyEqual, AllocatorOrContainer, Bucket, IsSegmented>;
+
+        // going back to front because erase() invalidates the end iterator
+        auto const old_size = map.size();
+        auto idx = old_size;
+        while (idx)
+        {
+            --idx;
+            auto it = map.begin() + static_cast<typename map_t::difference_type>(idx);
+            if (pred(*it))
+            {
+                map.erase(it);
+            }
+        }
+
+        return old_size - map.size();
+    }
+
+} // namespace std
+
+#endif
+#endif
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner.hpp b/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner.hpp
new file mode 100644
index 00000000..c9ed4fa3
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner.hpp
@@ -0,0 +1,157 @@
+#pragma once
+#include <vector>
+#include <string>
+#include <map>
+#include <stdexcept>
+
+typedef enum _color_space_e
+{
+    axdl_color_space_unknown,
+    axdl_color_space_nv12,
+    axdl_color_space_nv21,
+    axdl_color_space_bgr,
+    axdl_color_space_rgb,
+} ax_color_space_e;
+
+typedef struct _image_t
+{
+    unsigned long long int pPhy;
+    void *pVir;
+    unsigned int nSize;
+    unsigned int nWidth;
+    unsigned int nHeight;
+    ax_color_space_e eDtype;
+    union
+    {
+        int tStride_H, tStride_W, tStride_C;
+    };
+} ax_image_t;
+
+typedef struct
+{
+    std::string sName;
+    unsigned int nIdx;
+    std::vector<unsigned int> vShape;
+    int nSize;
+    unsigned long phyAddr;
+    void *pVirAddr;
+} ax_runner_tensor_t;
+
+class ax_runner_base
+{
+protected:
+    std::vector<ax_runner_tensor_t> moutput_tensors;
+    std::vector<ax_runner_tensor_t> minput_tensors;
+
+    std::vector<std::vector<ax_runner_tensor_t>> mgroup_output_tensors;
+    std::vector<std::vector<ax_runner_tensor_t>> mgroup_input_tensors;
+
+    std::map<std::string, ax_runner_tensor_t> map_output_tensors;
+    std::map<std::string, ax_runner_tensor_t> map_input_tensors;
+
+    std::map<std::string, std::vector<ax_runner_tensor_t>> map_group_output_tensors;
+    std::map<std::string, std::vector<ax_runner_tensor_t>> map_group_input_tensors;
+
+public:
+    virtual int init(const char *model_file, bool use_mmap = false) = 0;
+    virtual int init(char *model_buffer, size_t model_size) = 0;
+
+    virtual void deinit() = 0;
+
+    int get_num_inputs() { return minput_tensors.size(); };
+    int get_num_outputs() { return moutput_tensors.size(); };
+
+    int get_num_input_groups() { return mgroup_input_tensors.size(); };
+    int get_num_output_groups() { return mgroup_output_tensors.size(); };
+
+    const ax_runner_tensor_t &get_input(int idx) { return minput_tensors[idx]; }
+    const ax_runner_tensor_t *get_inputs_ptr() { return minput_tensors.data(); }
+    const ax_runner_tensor_t &get_input(std::string name)
+    {
+        if (map_input_tensors.size() == 0)
+        {
+            for (size_t i = 0; i < minput_tensors.size(); i++)
+            {
+                map_input_tensors[minput_tensors[i].sName] = minput_tensors[i];
+            }
+        }
+        if (map_input_tensors.find(name) == map_input_tensors.end())
+        {
+            throw std::runtime_error("input tensor not found: " + name);
+        }
+
+        return map_input_tensors[name];
+    }
+
+    const ax_runner_tensor_t &get_input(int grpid, int idx) { return mgroup_input_tensors[grpid][idx]; }
+    const ax_runner_tensor_t *get_inputs_ptr(int grpid) { return mgroup_input_tensors[grpid].data(); }
+    const ax_runner_tensor_t &get_input(int grpid, std::string name)
+    {
+        if (map_group_input_tensors.size() == 0)
+        {
+            for (size_t i = 0; i < mgroup_input_tensors.size(); i++)
+            {
+                for (size_t j = 0; j < mgroup_input_tensors[i].size(); j++)
+                {
+                    map_group_input_tensors[mgroup_input_tensors[i][j].sName].push_back(mgroup_input_tensors[i][j]);
+                }
+            }
+        }
+        if (map_group_input_tensors.find(name) == map_group_input_tensors.end())
+        {
+            throw std::runtime_error("input tensor not found: " + name);
+        }
+        return map_group_input_tensors[name][grpid];
+        // return map_input_tensors[name];
+    }
+
+    const ax_runner_tensor_t &get_output(int idx) { return moutput_tensors[idx]; }
+    const ax_runner_tensor_t *get_outputs_ptr() { return moutput_tensors.data(); }
+    const ax_runner_tensor_t &get_output(std::string name)
+    {
+        if (map_output_tensors.size() == 0)
+        {
+            for (size_t i = 0; i < moutput_tensors.size(); i++)
+            {
+                map_output_tensors[moutput_tensors[i].sName] = moutput_tensors[i];
+            }
+        }
+        if (map_output_tensors.find(name) == map_output_tensors.end())
+        {
+            throw std::runtime_error("output tensor not found: " + name);
+        }
+
+        return map_output_tensors[name];
+    }
+
+    const ax_runner_tensor_t &get_output(int grpid, int idx) { return mgroup_output_tensors[grpid][idx]; }
+    const ax_runner_tensor_t *get_outputs_ptr(int grpid) { return mgroup_output_tensors[grpid].data(); }
+    const ax_runner_tensor_t &get_output(int grpid, std::string name)
+    {
+        if (map_group_output_tensors.size() == 0)
+        {
+            for (size_t i = 0; i < mgroup_output_tensors.size(); i++)
+            {
+                for (size_t j = 0; j < mgroup_output_tensors[i].size(); j++)
+                {
+                    map_group_output_tensors[mgroup_output_tensors[i][j].sName].push_back(mgroup_output_tensors[i][j]);
+                }
+            }
+        }
+        if (map_group_output_tensors.find(name) == map_group_output_tensors.end())
+        {
+            throw std::runtime_error("input tensor not found: " + name);
+        }
+        return map_group_output_tensors[name][grpid];
+    }
+
+    virtual int inference() = 0;
+    virtual int inference(int grpid) = 0;
+
+    int operator()()
+    {
+        return inference();
+    }
+};
+
+// int ax_cmmcpy(unsigned long long int dst, unsigned long long int src, int size);
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.cpp b/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.cpp
new file mode 100644
index 00000000..a974f0cb
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.cpp
@@ -0,0 +1,454 @@
+#include "ax_model_runner_ax650.hpp"
+#include "string.h"
+#include "fstream"
+#include "memory"
+// #include "utilities/file.hpp"
+#include <ax_sys_api.h>
+#include <ax_ivps_api.h>
+#include <ax_engine_api.h>
+#include <fcntl.h>
+#include "memory_utils.hpp"
+#include "sample_log.h"
+
+#define AX_CMM_ALIGN_SIZE 128
+
+const char *AX_CMM_SESSION_NAME = "npu";
+
+typedef enum
+{
+    AX_ENGINE_ABST_DEFAULT = 0,
+    AX_ENGINE_ABST_CACHED = 1,
+} AX_ENGINE_ALLOC_BUFFER_STRATEGY_T;
+
+typedef std::pair<AX_ENGINE_ALLOC_BUFFER_STRATEGY_T, AX_ENGINE_ALLOC_BUFFER_STRATEGY_T> INPUT_OUTPUT_ALLOC_STRATEGY;
+
+static void print_io_info(AX_ENGINE_IO_INFO_T *io_info)
+{
+    static std::map<AX_ENGINE_DATA_TYPE_T, const char *> data_type = {
+        {AX_ENGINE_DT_UNKNOWN, "UNKNOWN"},
+        {AX_ENGINE_DT_UINT8, "UINT8"},
+        {AX_ENGINE_DT_UINT16, "UINT16"},
+        {AX_ENGINE_DT_FLOAT32, "FLOAT32"},
+        {AX_ENGINE_DT_SINT16, "SINT16"},
+        {AX_ENGINE_DT_SINT8, "SINT8"},
+        {AX_ENGINE_DT_SINT32, "SINT32"},
+        {AX_ENGINE_DT_UINT32, "UINT32"},
+        {AX_ENGINE_DT_FLOAT64, "FLOAT64"},
+        {AX_ENGINE_DT_UINT10_PACKED, "UINT10_PACKED"},
+        {AX_ENGINE_DT_UINT12_PACKED, "UINT12_PACKED"},
+        {AX_ENGINE_DT_UINT14_PACKED, "UINT14_PACKED"},
+        {AX_ENGINE_DT_UINT16_PACKED, "UINT16_PACKED"},
+    };
+
+    static std::map<AX_ENGINE_COLOR_SPACE_T, const char *> color_type = {
+        {AX_ENGINE_CS_FEATUREMAP, "FEATUREMAP"},
+        {AX_ENGINE_CS_RAW8, "RAW8"},
+        {AX_ENGINE_CS_RAW10, "RAW10"},
+        {AX_ENGINE_CS_RAW12, "RAW12"},
+        {AX_ENGINE_CS_RAW14, "RAW14"},
+        {AX_ENGINE_CS_RAW16, "RAW16"},
+        {AX_ENGINE_CS_NV12, "NV12"},
+        {AX_ENGINE_CS_NV21, "NV21"},
+        {AX_ENGINE_CS_RGB, "RGB"},
+        {AX_ENGINE_CS_BGR, "BGR"},
+        {AX_ENGINE_CS_RGBA, "RGBA"},
+        {AX_ENGINE_CS_GRAY, "GRAY"},
+        {AX_ENGINE_CS_YUV444, "YUV444"},
+    };
+    printf("\ninput size: %d\n", io_info->nInputSize);
+    for (uint32_t i = 0; i < io_info->nInputSize; ++i)
+    {
+        // print shape info,like [batchsize x channel x height x width]
+        auto &info = io_info->pInputs[i];
+        printf("    name: \e[1;32m%8s", info.pName);
+
+        std::string dt = "unknown";
+        if (data_type.find(info.eDataType) != data_type.end())
+        {
+            dt = data_type[info.eDataType];
+            printf(" \e[1;34m[%s] ", dt.c_str());
+        }
+        else
+        {
+            printf(" \e[1;31m[%s] ", dt.c_str());
+        }
+
+        std::string ct = "unknown";
+        if (info.pExtraMeta && color_type.find(info.pExtraMeta->eColorSpace) != color_type.end())
+        {
+            ct = color_type[info.pExtraMeta->eColorSpace];
+            printf("\e[1;34m[%s]", ct.c_str());
+        }
+        else
+        {
+            printf("\e[1;31m[%s]", ct.c_str());
+        }
+        printf(" \n        \e[1;31m");
+
+        for (AX_U8 s = 0; s < info.nShapeSize; s++)
+        {
+            printf("%d", info.pShape[s]);
+            if (s != info.nShapeSize - 1)
+            {
+                printf(" x ");
+            }
+        }
+        printf("\e[0m\n\n");
+    }
+
+    printf("\noutput size: %d\n", io_info->nOutputSize);
+    for (uint32_t i = 0; i < io_info->nOutputSize; ++i)
+    {
+        // print shape info,like [batchsize x channel x height x width]
+        auto &info = io_info->pOutputs[i];
+        printf("    name: \e[1;32m%8s \e[1;34m[%s]\e[0m\n        \e[1;31m", info.pName, data_type[info.eDataType]);
+        for (AX_U8 s = 0; s < info.nShapeSize; s++)
+        {
+            printf("%d", info.pShape[s]);
+            if (s != info.nShapeSize - 1)
+            {
+                printf(" x ");
+            }
+        }
+        printf("\e[0m\n\n");
+    }
+}
+
+void free_io_index(AX_ENGINE_IO_BUFFER_T *io_buf, int index)
+{
+    for (int i = 0; i < index; ++i)
+    {
+        AX_ENGINE_IO_BUFFER_T *pBuf = io_buf + i;
+        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+    }
+}
+
+void free_io(AX_ENGINE_IO_T *io)
+{
+    for (size_t j = 0; j < io->nInputSize; ++j)
+    {
+        AX_ENGINE_IO_BUFFER_T *pBuf = io->pInputs + j;
+        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+    }
+    for (size_t j = 0; j < io->nOutputSize; ++j)
+    {
+        AX_ENGINE_IO_BUFFER_T *pBuf = io->pOutputs + j;
+        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+    }
+    delete[] io->pInputs;
+    delete[] io->pOutputs;
+}
+
+static inline int prepare_io(AX_ENGINE_IO_INFO_T *info, AX_ENGINE_IO_T *io_data, INPUT_OUTPUT_ALLOC_STRATEGY strategy)
+{
+    memset(io_data, 0, sizeof(*io_data));
+    io_data->pInputs = new AX_ENGINE_IO_BUFFER_T[info->nInputSize];
+    io_data->nInputSize = info->nInputSize;
+
+    auto ret = 0;
+    for (uint i = 0; i < info->nInputSize; ++i)
+    {
+        auto meta = info->pInputs[i];
+        auto buffer = &io_data->pInputs[i];
+        if (strategy.first == AX_ENGINE_ABST_CACHED)
+        {
+            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        }
+        else
+        {
+            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        }
+
+        if (ret != 0)
+        {
+            free_io_index(io_data->pInputs, i);
+            fprintf(stderr, "Allocate input{%d} { phy: %p, vir: %p, size: %lu Bytes }. fail \n", i, (void *)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+            return ret;
+        }
+        memset(buffer->pVirAddr, 0, meta.nSize);
+        // fprintf(stderr, "Allocate input{%d} { phy: %p, vir: %p, size: %lu Bytes }. \n", i, (void*)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+    }
+
+    io_data->pOutputs = new AX_ENGINE_IO_BUFFER_T[info->nOutputSize];
+    io_data->nOutputSize = info->nOutputSize;
+    for (uint i = 0; i < info->nOutputSize; ++i)
+    {
+        auto meta = info->pOutputs[i];
+        auto buffer = &io_data->pOutputs[i];
+        buffer->nSize = meta.nSize;
+        if (strategy.second == AX_ENGINE_ABST_CACHED)
+        {
+            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        }
+        else
+        {
+            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        }
+        if (ret != 0)
+        {
+            fprintf(stderr, "Allocate output{%d} { phy: %p, vir: %p, size: %lu Bytes }. fail \n", i, (void *)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+            free_io_index(io_data->pInputs, io_data->nInputSize);
+            free_io_index(io_data->pOutputs, i);
+            return ret;
+        }
+        memset(buffer->pVirAddr, 0, meta.nSize);
+        // fprintf(stderr, "Allocate output{%d} { phy: %p, vir: %p, size: %lu Bytes }.\n", i, (void*)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+    }
+
+    return 0;
+}
+
+struct ax_joint_runner_ax650_handle_t
+{
+    AX_ENGINE_HANDLE handle;
+    AX_ENGINE_CONTEXT_T context;
+    std::vector<AX_ENGINE_IO_INFO_T *> io_info;
+    std::vector<AX_ENGINE_IO_T> io_data;
+
+    // int algo_width, algo_height;
+    // int algo_colorformat;
+};
+
+int ax_runner_ax650::sub_init()
+{
+    // 4. create context
+    int ret = AX_ENGINE_CreateContext(m_handle->handle);
+    if (0 != ret)
+    {
+        ALOGE("AX_ENGINE_CreateContext");
+        return ret;
+    }
+    ret = AX_ENGINE_CreateContextV2(m_handle->handle, &m_handle->context);
+    if (0 != ret)
+    {
+        ALOGE("AX_ENGINE_CreateContextV2");
+        return ret;
+    }
+    // fprintf(stdout, "Engine creating context is done.\n");
+
+    // 5. set io
+    AX_U32 io_count = 0;
+    ret = AX_ENGINE_GetGroupIOInfoCount(m_handle->handle, &io_count);
+    if (0 != ret)
+    {
+        ALOGE("AX_ENGINE_GetGroupIOInfoCount");
+        return ret;
+    }
+    // ALOGI("io_count=%d", io_count);
+
+    m_handle->io_info.resize(io_count);
+    m_handle->io_data.resize(io_count);
+    mgroup_input_tensors.resize(io_count);
+    mgroup_output_tensors.resize(io_count);
+
+    // fprintf(stdout, "Engine get io info is done. \n");
+
+    // 6. alloc io
+    if (!_parepare_io)
+    {
+        for (size_t grpid = 0; grpid < io_count; grpid++)
+        {
+            AX_ENGINE_IO_INFO_T *io_info = nullptr;
+            ret = AX_ENGINE_GetGroupIOInfo(m_handle->handle, grpid, &io_info);
+            if (0 != ret)
+            {
+                ALOGE("AX_ENGINE_GetIOInfo");
+                return ret;
+            }
+            // print_io_info(io_info);
+
+            m_handle->io_info[grpid] = io_info;
+
+            ret = prepare_io(m_handle->io_info[grpid], &m_handle->io_data[grpid], std::make_pair(AX_ENGINE_ABST_DEFAULT, AX_ENGINE_ABST_CACHED));
+            if (0 != ret)
+            {
+                ALOGE("prepare_io grpid=%d", grpid);
+                return ret;
+            }
+        }
+
+        for (size_t grpid = 0; grpid < io_count; grpid++)
+        {
+            auto &io_info = m_handle->io_info[grpid];
+            auto &io_data = m_handle->io_data[grpid];
+            for (size_t i = 0; i < io_info->nOutputSize; i++)
+            {
+                ax_runner_tensor_t tensor;
+                tensor.nIdx = i;
+                tensor.sName = std::string(io_info->pOutputs[i].pName);
+                tensor.nSize = io_info->pOutputs[i].nSize;
+                for (size_t j = 0; j < io_info->pOutputs[i].nShapeSize; j++)
+                {
+                    tensor.vShape.push_back(io_info->pOutputs[i].pShape[j]);
+                }
+                tensor.phyAddr = io_data.pOutputs[i].phyAddr;
+                tensor.pVirAddr = io_data.pOutputs[i].pVirAddr;
+                mgroup_output_tensors[grpid].push_back(tensor);
+            }
+
+            for (size_t i = 0; i < io_info->nInputSize; i++)
+            {
+                ax_runner_tensor_t tensor;
+                tensor.nIdx = i;
+                tensor.sName = std::string(io_info->pInputs[i].pName);
+                tensor.nSize = io_info->pInputs[i].nSize;
+                for (size_t j = 0; j < io_info->pInputs[i].nShapeSize; j++)
+                {
+                    tensor.vShape.push_back(io_info->pInputs[i].pShape[j]);
+                }
+                tensor.phyAddr = io_data.pInputs[i].phyAddr;
+                tensor.pVirAddr = io_data.pInputs[i].pVirAddr;
+                mgroup_input_tensors[grpid].push_back(tensor);
+            }
+        }
+
+        moutput_tensors = mgroup_output_tensors[0];
+        minput_tensors = mgroup_input_tensors[0];
+
+        _parepare_io = true;
+    }
+    else
+    {
+    }
+
+    return ret;
+}
+
+int ax_runner_ax650::init(const char *model_file, bool use_mmap)
+{
+    if (use_mmap)
+    {
+        MMap model_buffer(model_file);
+        if (!model_buffer.data())
+        {
+            ALOGE("mmap");
+            return -1;
+        }
+        auto ret = init((char *)model_buffer.data(), model_buffer.size());
+        model_buffer.close_file();
+        return ret;
+    }
+    else
+    {
+        char *model_buffer;
+        size_t len;
+        if (!read_file(model_file, &model_buffer, &len))
+        {
+            ALOGE("read_file");
+            return -1;
+        }
+        auto ret = init(model_buffer, len);
+        delete[] model_buffer;
+        return ret;
+    }
+}
+
+int ax_runner_ax650::init(char *model_buffer, size_t model_size)
+{
+    if (!m_handle)
+    {
+        m_handle = new ax_joint_runner_ax650_handle_t;
+    }
+
+    static bool b_init = false;
+    if (!b_init)
+    {
+        // 1. init engine
+        AX_ENGINE_NPU_ATTR_T npu_attr;
+        memset(&npu_attr, 0, sizeof(npu_attr));
+        npu_attr.eHardMode = AX_ENGINE_VIRTUAL_NPU_DISABLE;
+        AX_SYS_Init();
+        auto ret = AX_ENGINE_Init(&npu_attr);
+        if (0 != ret)
+        {
+            return ret;
+        }
+        b_init = true;
+    }
+
+    // 3. create handle
+
+    int ret = AX_ENGINE_CreateHandle(&m_handle->handle, model_buffer, model_size);
+    if (0 != ret)
+    {
+        ALOGE("AX_ENGINE_CreateHandle");
+        return ret;
+    }
+    // fprintf(stdout, "Engine creating handle is done.\n");
+
+    return sub_init();
+}
+
+void ax_runner_ax650::release()
+{
+    if (m_handle && m_handle->handle)
+    {
+        for (size_t i = 0; i < m_handle->io_data.size(); i++)
+        {
+            /* code */
+            free_io(&m_handle->io_data[i]);
+        }
+
+        AX_ENGINE_DestroyHandle(m_handle->handle);
+        m_handle->handle = nullptr;
+    }
+
+    if (m_handle)
+    {
+        delete m_handle;
+        m_handle = nullptr;
+    }
+
+    moutput_tensors.clear();
+    minput_tensors.clear();
+    map_input_tensors.clear();
+    map_output_tensors.clear();
+
+    mgroup_output_tensors.clear();
+    mgroup_input_tensors.clear();
+    map_group_input_tensors.clear();
+    map_group_output_tensors.clear();
+
+    // AX_ENGINE_Deinit();
+}
+
+void ax_runner_ax650::deinit()
+{
+    if (m_handle && m_handle->handle)
+    {
+        // free_io(&m_handle->io_data);
+        // mtensors.clear();
+        // minput_tensors.clear();
+        // map_input_tensors.clear();
+        // map_tensors.clear();
+        AX_ENGINE_DestroyHandle(m_handle->handle);
+        m_handle->handle = nullptr;
+        // delete m_handle;
+        // m_handle = nullptr;
+    }
+
+    // AX_ENGINE_Deinit();
+}
+
+int ax_runner_ax650::inference()
+{
+    int ret = AX_ENGINE_RunSync(m_handle->handle, &m_handle->io_data[0]);
+    for (size_t i = 0; i < get_num_outputs(); i++)
+    {
+        auto &tensor = get_output(i);
+        AX_SYS_MinvalidateCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
+    }
+    return ret;
+}
+
+int ax_runner_ax650::inference(int grpid)
+{
+    int ret = AX_ENGINE_RunGroupIOSync(m_handle->handle, m_handle->context, grpid, &m_handle->io_data[grpid]);
+
+    for (size_t i = 0; i < get_num_outputs(); i++)
+    {
+        auto &tensor = get_output(grpid, i);
+        AX_SYS_MinvalidateCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
+    }
+    return ret;
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.hpp b/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.hpp
new file mode 100644
index 00000000..351fb04a
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.hpp
@@ -0,0 +1,22 @@
+#pragma once
+#include "ax_model_runner.hpp"
+
+class ax_runner_ax650 : public ax_runner_base
+{
+protected:
+    struct ax_joint_runner_ax650_handle_t *m_handle = nullptr;
+
+    bool _parepare_io = false;
+
+    int sub_init();
+
+public:
+    int init(const char *model_file, bool use_mmap = false) override;
+    int init(char *model_buffer, size_t model_size) override;
+
+    void release();
+    void deinit() override;
+
+    int inference() override;
+    int inference(int grpid) override;
+};
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/ax_cmm_utils.hpp b/projects/llm_framework/main_cosy_voice/src/runner/utils/ax_cmm_utils.hpp
new file mode 100644
index 00000000..aaa6d6b9
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/ax_cmm_utils.hpp
@@ -0,0 +1,43 @@
+#pragma once
+#include <string.h>
+#include <vector>
+#include <fstream>
+#include <regex>
+
+// #include "sample_log.h"
+
+static std::string exec_cmd(std::string cmd)
+{
+    FILE *pipe = popen(cmd.c_str(), "r");
+    if (!pipe)
+    {
+        return "";
+    }
+    char buffer[128];
+    std::string result = "";
+    while (!feof(pipe))
+    {
+        if (fgets(buffer, 128, pipe) != NULL)
+        {
+            result += buffer;
+        }
+    }
+    pclose(pipe);
+    return result;
+}
+
+static int get_remaining_cmm_size()
+{
+    std::string cmd = "cat /proc/ax_proc/mem_cmm_info |grep 'total size'";
+    std::string result = exec_cmd(cmd);
+
+    std::regex pattern("remain=(\\d+)KB\\((\\d+)MB \\+ (\\d+)KB\\)");
+    std::smatch match;
+    if (std::regex_search(result, match, pattern))
+    {
+        int remain_kb = std::stoi(match[1]);
+        int remain_mb = std::stoi(match[2]);
+        return remain_mb;
+    }
+    return -1;
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/bfloat16.hpp b/projects/llm_framework/main_cosy_voice/src/runner/utils/bfloat16.hpp
new file mode 100644
index 00000000..157e07d1
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/bfloat16.hpp
@@ -0,0 +1,71 @@
+#pragma once
+#include <vector>
+
+struct bfloat16
+{
+    unsigned short int data;
+
+public:
+    bfloat16()
+    {
+        data = 0;
+    }
+    bfloat16(float float_val)
+    {
+        *this = float_val;
+    }
+
+    bfloat16(unsigned short int _data)
+    {
+        data = _data;
+    }
+
+    float fp32()
+    {
+        return *this;
+    }
+
+    // cast to float
+    operator float()
+    {
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wuninitialized"
+        unsigned int proc = data << 16;
+        return *reinterpret_cast<float *>(&proc);
+#pragma GCC diagnostic pop
+    }
+    // cast to bfloat16
+    bfloat16 &operator=(float float_val)
+    {
+        data = (*reinterpret_cast<unsigned int *>(&float_val)) >> 16;
+        return *this;
+    }
+};
+
+static std::vector<std::pair<int, float>> topk_bfloat16(unsigned short *arr, int size, int k)
+{
+    std::vector<std::pair<int, float>> result;
+
+    // Create a vector of pairs with index and value
+    std::vector<std::pair<int, float>> indexedValues;
+    indexedValues.reserve(size);
+    for (int i = 0; i < size; ++i)
+    {
+        indexedValues.push_back(std::make_pair(i, bfloat16(arr[i])));
+    }
+
+    // Sort the vector based on the values in descending order
+    std::sort(indexedValues.begin(), indexedValues.end(),
+              [](const std::pair<int, float> &a, const std::pair<int, float> &b)
+              {
+                  return a.second > b.second;
+              });
+
+    // Take the top k elements
+    for (int i = 0; i < k; ++i)
+    {
+        result.push_back(indexedValues[i]);
+    }
+
+    return result;
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/concat_3d.h b/projects/llm_framework/main_cosy_voice/src/runner/utils/concat_3d.h
new file mode 100644
index 00000000..19559ee2
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/concat_3d.h
@@ -0,0 +1,96 @@
+#include <vector>
+#include <cstddef> // For size_t
+#include <stdexcept> // For std::invalid_argument
+#include <algorithm> // For std::copy
+#include <type_traits> // For std::enable_if, std::is_arithmetic
+
+/**
+ * @brief Concatenates two 3D arrays along the last dimension (dim=2).
+ *        Input and output data are represented as 1D std::vectors in row-major order.
+ *
+ * @tparam T The data type of the elements (e.g., float, int).
+ *          Must be an arithmetic type.
+ * @param hift_cache_mel_data The first input 1D vector (data for hift_cache_mel).
+ * @param dim0_h The size of the first dimension for hift_cache_mel.
+ * @param dim1_h The size of the second dimension for hift_cache_mel.
+ * @param dim2_h The size of the third dimension for hift_cache_mel.
+ * @param tts_mel_data The second input 1D vector (data for tts_mel).
+ * @param dim0_t The size of the first dimension for tts_mel.
+ * @param dim1_t The size of the second dimension for tts_mel.
+ * @param dim2_t The size of the third dimension for tts_mel.
+ * @return std::vector<T> A new 1D vector containing the concatenated data.
+ * @throws std::invalid_argument If dimensions are incompatible.
+ */
+template <typename T>
+typename std::enable_if<std::is_arithmetic<T>::value, std::vector<T>>::type
+concat_3d_dim2(const std::vector<T>& hift_cache_mel_data,
+               size_t dim0_h, size_t dim1_h, size_t dim2_h,
+               const std::vector<T>& tts_mel_data,
+               size_t dim0_t, size_t dim1_t, size_t dim2_t) {
+
+    // --- Input Validation ---
+    if (hift_cache_mel_data.size() != dim0_h * dim1_h * dim2_h) {
+        throw std::invalid_argument("hift_cache_mel_data size does not match provided dimensions.");
+    }
+    if (tts_mel_data.size() != dim0_t * dim1_t * dim2_t) {
+        throw std::invalid_argument("tts_mel_data size does not match provided dimensions.");
+    }
+    if (dim0_h != dim0_t || dim1_h != dim1_t) {
+        throw std::invalid_argument("First two dimensions (dim0, dim1) of input arrays must match for concatenation along dim=2.");
+    }
+
+    const size_t dim0 = dim0_h; // == dim0_t
+    const size_t dim1 = dim1_h; // == dim1_t
+    const size_t dim2_result = dim2_h + dim2_t;
+
+    // Handle case where result would be empty
+    if (dim0 == 0 || dim1 == 0 || dim2_result == 0) {
+        return std::vector<T>();
+    }
+
+    // --- Calculate Result Size and Reserve Memory ---
+    const size_t result_size = dim0 * dim1 * dim2_result;
+    std::vector<T> result_data;
+    result_data.reserve(result_size); // Pre-allocate for efficiency
+
+    // --- Calculate Strides (Row-Major Order) ---
+    const size_t stride_dim2_h = 1;
+    const size_t stride_dim1_h = dim2_h * stride_dim2_h;
+    const size_t stride_dim0_h = dim1_h * stride_dim1_h;
+
+    const size_t stride_dim2_t = 1;
+    const size_t stride_dim1_t = dim2_t * stride_dim2_t;
+    const size_t stride_dim0_t = dim1_t * stride_dim1_t;
+
+    const size_t stride_dim2_result = 1;
+    const size_t stride_dim1_result = dim2_result * stride_dim2_result;
+    const size_t stride_dim0_result = dim1 * stride_dim1_result; // dim1 == dim1_h == dim1_t
+
+    // --- Perform Concatenation ---
+    result_data.resize(result_size); // Resize once for direct access
+    size_t result_idx = 0;
+
+    for (size_t n = 0; n < dim0; ++n) {
+        for (size_t c = 0; c < dim1; ++c) {
+            // --- Copy slice from hift_cache_mel [n][c][:] ---
+            const size_t hift_base_offset = n * stride_dim0_h + c * stride_dim1_h;
+            const size_t hift_slice_end_offset = hift_base_offset + dim2_h * stride_dim2_h;
+            // Use std::copy for efficient block copy
+            std::copy(hift_cache_mel_data.data() + hift_base_offset,
+                      hift_cache_mel_data.data() + hift_slice_end_offset,
+                      result_data.data() + result_idx);
+            result_idx += dim2_h;
+
+            // --- Copy slice from tts_mel [n][c][:] ---
+            const size_t tts_base_offset = n * stride_dim0_t + c * stride_dim1_t;
+            const size_t tts_slice_end_offset = tts_base_offset + dim2_t * stride_dim2_t;
+            // Use std::copy for efficient block copy
+            std::copy(tts_mel_data.data() + tts_base_offset,
+                      tts_mel_data.data() + tts_slice_end_offset,
+                      result_data.data() + result_idx);
+            result_idx += dim2_t;
+        }
+    }
+
+    return result_data; // RVO should apply
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/cqdm.cpp b/projects/llm_framework/main_cosy_voice/src/runner/utils/cqdm.cpp
new file mode 100644
index 00000000..79a19798
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/cqdm.cpp
@@ -0,0 +1,48 @@
+#include "cqdm.h"
+
+// Returns the current time in milliseconds
+long long get_msec_now()
+{
+    struct timeval time;
+
+    if (gettimeofday(&time, NULL) == -1)
+        return (0);
+    return (long long)time.tv_sec * 1000LL + (long long)time.tv_usec / 1000LL;
+}
+
+t_cqdm create_cqdm(int total, int size)
+{
+    t_cqdm cqdm;
+
+    cqdm.total = total;
+    cqdm.count = 0;
+    cqdm.average_time = 0;
+    cqdm.last_time = get_msec_now();
+    cqdm.total_time = 0;
+    cqdm.size = size;
+    return (cqdm);
+}
+
+void update_cqdm(t_cqdm *cqdm, int x, const char *unit, const char *log_str)
+{
+    long long now;
+    double temp;
+
+    now = get_msec_now();
+
+    cqdm->count++;
+    cqdm->total_time += ((now - cqdm->last_time) / 1000);
+    cqdm->average_time = cqdm->total_time / cqdm->count;
+    cqdm->last_time = now;
+
+    temp = ((double)(x + 1) / (double)cqdm->total);
+
+    fprintf(stdout, "\033[1;30;33m\r%3d%% | ", (int)(temp * 100));
+    for (int i = 0; i < (temp * cqdm->size); i++)
+        fprintf(stdout, "█");
+    for (int i = 0; i < (cqdm->size - temp * cqdm->size); i++)
+        fprintf(stdout, " ");
+    fprintf(stdout, " | %3d / %3d [%2.2fs<%2.2fs, %2.2f %s/s] \033[1;30;32m%s\033[0m", x + 1, cqdm->total, cqdm->total_time,
+            cqdm->average_time * cqdm->total, 1 / cqdm->average_time, unit, log_str);
+    fflush(stdout);
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/cqdm.h b/projects/llm_framework/main_cosy_voice/src/runner/utils/cqdm.h
new file mode 100644
index 00000000..c35fb72f
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/cqdm.h
@@ -0,0 +1,28 @@
+#ifndef CQDM_H
+#define CQDM_H
+
+#include <stdio.h>
+#include <sys/time.h>
+#include <unistd.h>
+
+// Structure to hold progress bar information
+typedef struct s_cqdm
+{
+    int total;
+    int count;
+    int size;
+    double average_time;
+    double last_time;
+    double total_time;
+} t_cqdm;
+#ifdef __cplusplus
+extern "C"
+{
+#endif
+    t_cqdm create_cqdm(int total, int size);
+    void update_cqdm(t_cqdm *cqdm, int x, const char *unit, const char *log_str);
+#ifdef __cplusplus
+}
+#endif
+
+#endif
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/http_utils.hpp b/projects/llm_framework/main_cosy_voice/src/runner/utils/http_utils.hpp
new file mode 100644
index 00000000..5041b6a1
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/http_utils.hpp
@@ -0,0 +1,102 @@
+#pragma once
+#include <iostream>
+#include <string>
+#include <cstring>
+#include <sys/socket.h>
+#include <arpa/inet.h>
+#include <unistd.h>
+#include <chrono>
+#include <thread>
+
+/**
+ * @brief Attempts to establish a TCP connection to a specified host and port.
+ *
+ * This function creates a socket and tries to connect to a server specified by
+ * the host and port parameters. It returns true if the connection is successful,
+ * otherwise it returns false and outputs an error message to standard error.
+ *
+ * @param host The IP address of the server to connect to.
+ * @param port The port number of the server to connect to.
+ *
+ * @return true if the connection is successfully established, false otherwise.
+ */
+static bool test_connect(const std::string &host, int port)
+{
+    int sock = socket(AF_INET, SOCK_STREAM, 0);
+    if (sock < 0)
+    {
+        // std::cerr << "Socket creation failed\n";
+        return false;
+    }
+
+    sockaddr_in server_addr{};
+    server_addr.sin_family = AF_INET;
+    server_addr.sin_port = htons(port);
+
+    if (inet_pton(AF_INET, host.c_str(), &server_addr.sin_addr) <= 0)
+    {
+        // std::cerr << "IP address conversion failed\n";
+        close(sock);
+        return false;
+    }
+
+    if (connect(sock, (struct sockaddr *)&server_addr, sizeof(server_addr)) < 0)
+    {
+        // std::cerr << "Connection failed\n";
+        close(sock);
+        return false;
+    }
+
+    close(sock);
+    return true;
+}
+
+/**
+ * @brief Attempts to establish an HTTP connection to a specified URL with a timeout.
+ *
+ * This function parses the provided HTTP URL to extract the host and port information,
+ * and attempts to establish a TCP connection using the `test_connect` function.
+ * It retries the connection until the specified timeout is reached.
+ *
+ * @param http_url The HTTP URL of the server to connect to.
+ * @param timeout The maximum number of seconds to keep attempting the connection.
+ *
+ * @return true if the connection is successfully established within the timeout period,
+ * false otherwise.
+ */
+static bool test_connect_http(const std::string &http_url, int timeout)
+{
+    size_t pos = http_url.find("://");
+    if (pos == std::string::npos)
+        return false;
+    std::string host = http_url.substr(pos + 3);
+    pos = host.find('/');
+    if (pos != std::string::npos)
+        host = host.substr(0, pos);
+    pos = host.find(':');
+    int port = 80;
+    if (pos != std::string::npos)
+    {
+        port = std::stoi(host.substr(pos + 1));
+        host = host.substr(0, pos);
+    }
+    else
+    {
+        return false;
+    }
+
+    if (host == "localhost")
+        host = "127.0.0.1";
+    int tmp = timeout;
+    while (timeout--)
+    {
+        if (test_connect(host, port))
+            return true;
+        std::this_thread::sleep_for(std::chrono::seconds(1));
+        printf("\033[1;30;31m"
+               "connect failed %s, try again in %2d/%2d \n"
+               "\033[0m",
+               http_url.c_str(), timeout, tmp);
+    }
+    return false;
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/httplib.h b/projects/llm_framework/main_cosy_voice/src/runner/utils/httplib.h
new file mode 100644
index 00000000..4e79f11f
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/httplib.h
@@ -0,0 +1,9483 @@
+//
+//  httplib.h
+//
+//  Copyright (c) 2024 Yuji Hirose. All rights reserved.
+//  MIT License
+//
+
+#ifndef CPPHTTPLIB_HTTPLIB_H
+#define CPPHTTPLIB_HTTPLIB_H
+
+#define CPPHTTPLIB_VERSION "0.15.3"
+
+/*
+ * Configuration
+ */
+
+#ifndef CPPHTTPLIB_KEEPALIVE_TIMEOUT_SECOND
+#define CPPHTTPLIB_KEEPALIVE_TIMEOUT_SECOND 5
+#endif
+
+#ifndef CPPHTTPLIB_KEEPALIVE_MAX_COUNT
+#define CPPHTTPLIB_KEEPALIVE_MAX_COUNT 5
+#endif
+
+#ifndef CPPHTTPLIB_CONNECTION_TIMEOUT_SECOND
+#define CPPHTTPLIB_CONNECTION_TIMEOUT_SECOND 300
+#endif
+
+#ifndef CPPHTTPLIB_CONNECTION_TIMEOUT_USECOND
+#define CPPHTTPLIB_CONNECTION_TIMEOUT_USECOND 0
+#endif
+
+#ifndef CPPHTTPLIB_READ_TIMEOUT_SECOND
+#define CPPHTTPLIB_READ_TIMEOUT_SECOND 5
+#endif
+
+#ifndef CPPHTTPLIB_READ_TIMEOUT_USECOND
+#define CPPHTTPLIB_READ_TIMEOUT_USECOND 0
+#endif
+
+#ifndef CPPHTTPLIB_WRITE_TIMEOUT_SECOND
+#define CPPHTTPLIB_WRITE_TIMEOUT_SECOND 5
+#endif
+
+#ifndef CPPHTTPLIB_WRITE_TIMEOUT_USECOND
+#define CPPHTTPLIB_WRITE_TIMEOUT_USECOND 0
+#endif
+
+#ifndef CPPHTTPLIB_IDLE_INTERVAL_SECOND
+#define CPPHTTPLIB_IDLE_INTERVAL_SECOND 0
+#endif
+
+#ifndef CPPHTTPLIB_IDLE_INTERVAL_USECOND
+#ifdef _WIN32
+#define CPPHTTPLIB_IDLE_INTERVAL_USECOND 10000
+#else
+#define CPPHTTPLIB_IDLE_INTERVAL_USECOND 0
+#endif
+#endif
+
+#ifndef CPPHTTPLIB_REQUEST_URI_MAX_LENGTH
+#define CPPHTTPLIB_REQUEST_URI_MAX_LENGTH 8192
+#endif
+
+#ifndef CPPHTTPLIB_HEADER_MAX_LENGTH
+#define CPPHTTPLIB_HEADER_MAX_LENGTH 8192
+#endif
+
+#ifndef CPPHTTPLIB_REDIRECT_MAX_COUNT
+#define CPPHTTPLIB_REDIRECT_MAX_COUNT 20
+#endif
+
+#ifndef CPPHTTPLIB_MULTIPART_FORM_DATA_FILE_MAX_COUNT
+#define CPPHTTPLIB_MULTIPART_FORM_DATA_FILE_MAX_COUNT 1024
+#endif
+
+#ifndef CPPHTTPLIB_PAYLOAD_MAX_LENGTH
+#define CPPHTTPLIB_PAYLOAD_MAX_LENGTH ((std::numeric_limits<size_t>::max)())
+#endif
+
+#ifndef CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH
+#define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 8192
+#endif
+
+#ifndef CPPHTTPLIB_RANGE_MAX_COUNT
+#define CPPHTTPLIB_RANGE_MAX_COUNT 1024
+#endif
+
+#ifndef CPPHTTPLIB_TCP_NODELAY
+#define CPPHTTPLIB_TCP_NODELAY false
+#endif
+
+#ifndef CPPHTTPLIB_RECV_BUFSIZ
+#define CPPHTTPLIB_RECV_BUFSIZ size_t(16384u)
+#endif
+
+#ifndef CPPHTTPLIB_COMPRESSION_BUFSIZ
+#define CPPHTTPLIB_COMPRESSION_BUFSIZ size_t(16384u)
+#endif
+
+#ifndef CPPHTTPLIB_THREAD_POOL_COUNT
+#define CPPHTTPLIB_THREAD_POOL_COUNT                                           \
+  ((std::max)(8u, std::thread::hardware_concurrency() > 0                      \
+                      ? std::thread::hardware_concurrency() - 1                \
+                      : 0))
+#endif
+
+#ifndef CPPHTTPLIB_RECV_FLAGS
+#define CPPHTTPLIB_RECV_FLAGS 0
+#endif
+
+#ifndef CPPHTTPLIB_SEND_FLAGS
+#define CPPHTTPLIB_SEND_FLAGS 0
+#endif
+
+#ifndef CPPHTTPLIB_LISTEN_BACKLOG
+#define CPPHTTPLIB_LISTEN_BACKLOG 5
+#endif
+
+/*
+ * Headers
+ */
+
+#ifdef _WIN32
+#ifndef _CRT_SECURE_NO_WARNINGS
+#define _CRT_SECURE_NO_WARNINGS
+#endif //_CRT_SECURE_NO_WARNINGS
+
+#ifndef _CRT_NONSTDC_NO_DEPRECATE
+#define _CRT_NONSTDC_NO_DEPRECATE
+#endif //_CRT_NONSTDC_NO_DEPRECATE
+
+#if defined(_MSC_VER)
+#if _MSC_VER < 1900
+#error Sorry, Visual Studio versions prior to 2015 are not supported
+#endif
+
+#pragma comment(lib, "ws2_32.lib")
+
+#ifdef _WIN64
+using ssize_t = __int64;
+#else
+using ssize_t = long;
+#endif
+#endif // _MSC_VER
+
+#ifndef S_ISREG
+#define S_ISREG(m) (((m) & S_IFREG) == S_IFREG)
+#endif // S_ISREG
+
+#ifndef S_ISDIR
+#define S_ISDIR(m) (((m) & S_IFDIR) == S_IFDIR)
+#endif // S_ISDIR
+
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif // NOMINMAX
+
+#include <io.h>
+#include <winsock2.h>
+#include <ws2tcpip.h>
+
+#ifndef WSA_FLAG_NO_HANDLE_INHERIT
+#define WSA_FLAG_NO_HANDLE_INHERIT 0x80
+#endif
+
+using socket_t = SOCKET;
+#ifdef CPPHTTPLIB_USE_POLL
+#define poll(fds, nfds, timeout) WSAPoll(fds, nfds, timeout)
+#endif
+
+#else // not _WIN32
+
+#include <arpa/inet.h>
+#if !defined(_AIX) && !defined(__MVS__)
+#include <ifaddrs.h>
+#endif
+#ifdef __MVS__
+#include <strings.h>
+#ifndef NI_MAXHOST
+#define NI_MAXHOST 1025
+#endif
+#endif
+#include <net/if.h>
+#include <netdb.h>
+#include <netinet/in.h>
+#ifdef __linux__
+#include <resolv.h>
+#endif
+#include <netinet/tcp.h>
+#ifdef CPPHTTPLIB_USE_POLL
+#include <poll.h>
+#endif
+#include <csignal>
+#include <pthread.h>
+#include <sys/mman.h>
+#include <sys/select.h>
+#include <sys/socket.h>
+#include <sys/un.h>
+#include <unistd.h>
+
+using socket_t = int;
+#ifndef INVALID_SOCKET
+#define INVALID_SOCKET (-1)
+#endif
+#endif //_WIN32
+
+#include <algorithm>
+#include <array>
+#include <atomic>
+#include <cassert>
+#include <cctype>
+#include <climits>
+#include <condition_variable>
+#include <cstring>
+#include <errno.h>
+#include <exception>
+#include <fcntl.h>
+#include <fstream>
+#include <functional>
+#include <iomanip>
+#include <iostream>
+#include <list>
+#include <map>
+#include <memory>
+#include <mutex>
+#include <random>
+#include <regex>
+#include <set>
+#include <sstream>
+#include <string>
+#include <sys/stat.h>
+#include <thread>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+#ifdef _WIN32
+#include <wincrypt.h>
+
+// these are defined in wincrypt.h and it breaks compilation if BoringSSL is
+// used
+#undef X509_NAME
+#undef X509_CERT_PAIR
+#undef X509_EXTENSIONS
+#undef PKCS7_SIGNER_INFO
+
+#ifdef _MSC_VER
+#pragma comment(lib, "crypt32.lib")
+#endif
+#elif defined(CPPHTTPLIB_USE_CERTS_FROM_MACOSX_KEYCHAIN) && defined(__APPLE__)
+#include <TargetConditionals.h>
+#if TARGET_OS_OSX
+#include <CoreFoundation/CoreFoundation.h>
+#include <Security/Security.h>
+#endif // TARGET_OS_OSX
+#endif // _WIN32
+
+#include <openssl/err.h>
+#include <openssl/evp.h>
+#include <openssl/ssl.h>
+#include <openssl/x509v3.h>
+
+#if defined(_WIN32) && defined(OPENSSL_USE_APPLINK)
+#include <openssl/applink.c>
+#endif
+
+#include <iostream>
+#include <sstream>
+
+#if OPENSSL_VERSION_NUMBER < 0x30000000L
+#error Sorry, OpenSSL versions prior to 3.0.0 are not supported
+#endif
+
+#endif
+
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+#include <zlib.h>
+#endif
+
+#ifdef CPPHTTPLIB_BROTLI_SUPPORT
+#include <brotli/decode.h>
+#include <brotli/encode.h>
+#endif
+
+/*
+ * Declaration
+ */
+namespace httplib {
+
+namespace detail {
+
+/*
+ * Backport std::make_unique from C++14.
+ *
+ * NOTE: This code came up with the following stackoverflow post:
+ * https://stackoverflow.com/questions/10149840/c-arrays-and-make-unique
+ *
+ */
+
+template <class T, class... Args>
+typename std::enable_if<!std::is_array<T>::value, std::unique_ptr<T>>::type
+make_unique(Args &&...args) {
+  return std::unique_ptr<T>(new T(std::forward<Args>(args)...));
+}
+
+template <class T>
+typename std::enable_if<std::is_array<T>::value, std::unique_ptr<T>>::type
+make_unique(std::size_t n) {
+  typedef typename std::remove_extent<T>::type RT;
+  return std::unique_ptr<T>(new RT[n]);
+}
+
+struct ci {
+  bool operator()(const std::string &s1, const std::string &s2) const {
+    return std::lexicographical_compare(s1.begin(), s1.end(), s2.begin(),
+                                        s2.end(),
+                                        [](unsigned char c1, unsigned char c2) {
+                                          return ::tolower(c1) < ::tolower(c2);
+                                        });
+  }
+};
+
+// This is based on
+// "http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4189".
+
+struct scope_exit {
+  explicit scope_exit(std::function<void(void)> &&f)
+      : exit_function(std::move(f)), execute_on_destruction{true} {}
+
+  scope_exit(scope_exit &&rhs) noexcept
+      : exit_function(std::move(rhs.exit_function)),
+        execute_on_destruction{rhs.execute_on_destruction} {
+    rhs.release();
+  }
+
+  ~scope_exit() {
+    if (execute_on_destruction) { this->exit_function(); }
+  }
+
+  void release() { this->execute_on_destruction = false; }
+
+private:
+  scope_exit(const scope_exit &) = delete;
+  void operator=(const scope_exit &) = delete;
+  scope_exit &operator=(scope_exit &&) = delete;
+
+  std::function<void(void)> exit_function;
+  bool execute_on_destruction;
+};
+
+} // namespace detail
+
+enum StatusCode {
+  // Information responses
+  Continue_100 = 100,
+  SwitchingProtocol_101 = 101,
+  Processing_102 = 102,
+  EarlyHints_103 = 103,
+
+  // Successful responses
+  OK_200 = 200,
+  Created_201 = 201,
+  Accepted_202 = 202,
+  NonAuthoritativeInformation_203 = 203,
+  NoContent_204 = 204,
+  ResetContent_205 = 205,
+  PartialContent_206 = 206,
+  MultiStatus_207 = 207,
+  AlreadyReported_208 = 208,
+  IMUsed_226 = 226,
+
+  // Redirection messages
+  MultipleChoices_300 = 300,
+  MovedPermanently_301 = 301,
+  Found_302 = 302,
+  SeeOther_303 = 303,
+  NotModified_304 = 304,
+  UseProxy_305 = 305,
+  unused_306 = 306,
+  TemporaryRedirect_307 = 307,
+  PermanentRedirect_308 = 308,
+
+  // Client error responses
+  BadRequest_400 = 400,
+  Unauthorized_401 = 401,
+  PaymentRequired_402 = 402,
+  Forbidden_403 = 403,
+  NotFound_404 = 404,
+  MethodNotAllowed_405 = 405,
+  NotAcceptable_406 = 406,
+  ProxyAuthenticationRequired_407 = 407,
+  RequestTimeout_408 = 408,
+  Conflict_409 = 409,
+  Gone_410 = 410,
+  LengthRequired_411 = 411,
+  PreconditionFailed_412 = 412,
+  PayloadTooLarge_413 = 413,
+  UriTooLong_414 = 414,
+  UnsupportedMediaType_415 = 415,
+  RangeNotSatisfiable_416 = 416,
+  ExpectationFailed_417 = 417,
+  ImATeapot_418 = 418,
+  MisdirectedRequest_421 = 421,
+  UnprocessableContent_422 = 422,
+  Locked_423 = 423,
+  FailedDependency_424 = 424,
+  TooEarly_425 = 425,
+  UpgradeRequired_426 = 426,
+  PreconditionRequired_428 = 428,
+  TooManyRequests_429 = 429,
+  RequestHeaderFieldsTooLarge_431 = 431,
+  UnavailableForLegalReasons_451 = 451,
+
+  // Server error responses
+  InternalServerError_500 = 500,
+  NotImplemented_501 = 501,
+  BadGateway_502 = 502,
+  ServiceUnavailable_503 = 503,
+  GatewayTimeout_504 = 504,
+  HttpVersionNotSupported_505 = 505,
+  VariantAlsoNegotiates_506 = 506,
+  InsufficientStorage_507 = 507,
+  LoopDetected_508 = 508,
+  NotExtended_510 = 510,
+  NetworkAuthenticationRequired_511 = 511,
+};
+
+using Headers = std::multimap<std::string, std::string, detail::ci>;
+
+using Params = std::multimap<std::string, std::string>;
+using Match = std::smatch;
+
+using Progress = std::function<bool(uint64_t current, uint64_t total)>;
+
+struct Response;
+using ResponseHandler = std::function<bool(const Response &response)>;
+
+struct MultipartFormData {
+  std::string name;
+  std::string content;
+  std::string filename;
+  std::string content_type;
+};
+using MultipartFormDataItems = std::vector<MultipartFormData>;
+using MultipartFormDataMap = std::multimap<std::string, MultipartFormData>;
+
+class DataSink {
+public:
+  DataSink() : os(&sb_), sb_(*this) {}
+
+  DataSink(const DataSink &) = delete;
+  DataSink &operator=(const DataSink &) = delete;
+  DataSink(DataSink &&) = delete;
+  DataSink &operator=(DataSink &&) = delete;
+
+  std::function<bool(const char *data, size_t data_len)> write;
+  std::function<bool()> is_writable;
+  std::function<void()> done;
+  std::function<void(const Headers &trailer)> done_with_trailer;
+  std::ostream os;
+
+private:
+  class data_sink_streambuf final : public std::streambuf {
+  public:
+    explicit data_sink_streambuf(DataSink &sink) : sink_(sink) {}
+
+  protected:
+    std::streamsize xsputn(const char *s, std::streamsize n) override {
+      sink_.write(s, static_cast<size_t>(n));
+      return n;
+    }
+
+  private:
+    DataSink &sink_;
+  };
+
+  data_sink_streambuf sb_;
+};
+
+using ContentProvider =
+    std::function<bool(size_t offset, size_t length, DataSink &sink)>;
+
+using ContentProviderWithoutLength =
+    std::function<bool(size_t offset, DataSink &sink)>;
+
+using ContentProviderResourceReleaser = std::function<void(bool success)>;
+
+struct MultipartFormDataProvider {
+  std::string name;
+  ContentProviderWithoutLength provider;
+  std::string filename;
+  std::string content_type;
+};
+using MultipartFormDataProviderItems = std::vector<MultipartFormDataProvider>;
+
+using ContentReceiverWithProgress =
+    std::function<bool(const char *data, size_t data_length, uint64_t offset,
+                       uint64_t total_length)>;
+
+using ContentReceiver =
+    std::function<bool(const char *data, size_t data_length)>;
+
+using MultipartContentHeader =
+    std::function<bool(const MultipartFormData &file)>;
+
+class ContentReader {
+public:
+  using Reader = std::function<bool(ContentReceiver receiver)>;
+  using MultipartReader = std::function<bool(MultipartContentHeader header,
+                                             ContentReceiver receiver)>;
+
+  ContentReader(Reader reader, MultipartReader multipart_reader)
+      : reader_(std::move(reader)),
+        multipart_reader_(std::move(multipart_reader)) {}
+
+  bool operator()(MultipartContentHeader header,
+                  ContentReceiver receiver) const {
+    return multipart_reader_(std::move(header), std::move(receiver));
+  }
+
+  bool operator()(ContentReceiver receiver) const {
+    return reader_(std::move(receiver));
+  }
+
+  Reader reader_;
+  MultipartReader multipart_reader_;
+};
+
+using Range = std::pair<ssize_t, ssize_t>;
+using Ranges = std::vector<Range>;
+
+struct Request {
+  std::string method;
+  std::string path;
+  Headers headers;
+  std::string body;
+
+  std::string remote_addr;
+  int remote_port = -1;
+  std::string local_addr;
+  int local_port = -1;
+
+  // for server
+  std::string version;
+  std::string target;
+  Params params;
+  MultipartFormDataMap files;
+  Ranges ranges;
+  Match matches;
+  std::unordered_map<std::string, std::string> path_params;
+
+  // for client
+  ResponseHandler response_handler;
+  ContentReceiverWithProgress content_receiver;
+  Progress progress;
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  const SSL *ssl = nullptr;
+#endif
+
+  bool has_header(const std::string &key) const;
+  std::string get_header_value(const std::string &key, size_t id = 0) const;
+  uint64_t get_header_value_u64(const std::string &key, size_t id = 0) const;
+  size_t get_header_value_count(const std::string &key) const;
+  void set_header(const std::string &key, const std::string &val);
+
+  bool has_param(const std::string &key) const;
+  std::string get_param_value(const std::string &key, size_t id = 0) const;
+  size_t get_param_value_count(const std::string &key) const;
+
+  bool is_multipart_form_data() const;
+
+  bool has_file(const std::string &key) const;
+  MultipartFormData get_file_value(const std::string &key) const;
+  std::vector<MultipartFormData> get_file_values(const std::string &key) const;
+
+  // private members...
+  size_t redirect_count_ = CPPHTTPLIB_REDIRECT_MAX_COUNT;
+  size_t content_length_ = 0;
+  ContentProvider content_provider_;
+  bool is_chunked_content_provider_ = false;
+  size_t authorization_count_ = 0;
+};
+
+struct Response {
+  std::string version;
+  int status = -1;
+  std::string reason;
+  Headers headers;
+  std::string body;
+  std::string location; // Redirect location
+
+  bool has_header(const std::string &key) const;
+  std::string get_header_value(const std::string &key, size_t id = 0) const;
+  uint64_t get_header_value_u64(const std::string &key, size_t id = 0) const;
+  size_t get_header_value_count(const std::string &key) const;
+  void set_header(const std::string &key, const std::string &val);
+
+  void set_redirect(const std::string &url, int status = StatusCode::Found_302);
+  void set_content(const char *s, size_t n, const std::string &content_type);
+  void set_content(const std::string &s, const std::string &content_type);
+  void set_content(std::string &&s, const std::string &content_type);
+
+  void set_content_provider(
+      size_t length, const std::string &content_type, ContentProvider provider,
+      ContentProviderResourceReleaser resource_releaser = nullptr);
+
+  void set_content_provider(
+      const std::string &content_type, ContentProviderWithoutLength provider,
+      ContentProviderResourceReleaser resource_releaser = nullptr);
+
+  void set_chunked_content_provider(
+      const std::string &content_type, ContentProviderWithoutLength provider,
+      ContentProviderResourceReleaser resource_releaser = nullptr);
+
+  Response() = default;
+  Response(const Response &) = default;
+  Response &operator=(const Response &) = default;
+  Response(Response &&) = default;
+  Response &operator=(Response &&) = default;
+  ~Response() {
+    if (content_provider_resource_releaser_) {
+      content_provider_resource_releaser_(content_provider_success_);
+    }
+  }
+
+  // private members...
+  size_t content_length_ = 0;
+  ContentProvider content_provider_;
+  ContentProviderResourceReleaser content_provider_resource_releaser_;
+  bool is_chunked_content_provider_ = false;
+  bool content_provider_success_ = false;
+};
+
+class Stream {
+public:
+  virtual ~Stream() = default;
+
+  virtual bool is_readable() const = 0;
+  virtual bool is_writable() const = 0;
+
+  virtual ssize_t read(char *ptr, size_t size) = 0;
+  virtual ssize_t write(const char *ptr, size_t size) = 0;
+  virtual void get_remote_ip_and_port(std::string &ip, int &port) const = 0;
+  virtual void get_local_ip_and_port(std::string &ip, int &port) const = 0;
+  virtual socket_t socket() const = 0;
+
+  template <typename... Args>
+  ssize_t write_format(const char *fmt, const Args &...args);
+  ssize_t write(const char *ptr);
+  ssize_t write(const std::string &s);
+};
+
+class TaskQueue {
+public:
+  TaskQueue() = default;
+  virtual ~TaskQueue() = default;
+
+  virtual bool enqueue(std::function<void()> fn) = 0;
+  virtual void shutdown() = 0;
+
+  virtual void on_idle() {}
+};
+
+class ThreadPool final : public TaskQueue {
+public:
+  explicit ThreadPool(size_t n, size_t mqr = 0)
+      : shutdown_(false), max_queued_requests_(mqr) {
+    while (n) {
+      threads_.emplace_back(worker(*this));
+      n--;
+    }
+  }
+
+  ThreadPool(const ThreadPool &) = delete;
+  ~ThreadPool() override = default;
+
+  bool enqueue(std::function<void()> fn) override {
+    {
+      std::unique_lock<std::mutex> lock(mutex_);
+      if (max_queued_requests_ > 0 && jobs_.size() >= max_queued_requests_) {
+        return false;
+      }
+      jobs_.push_back(std::move(fn));
+    }
+
+    cond_.notify_one();
+    return true;
+  }
+
+  void shutdown() override {
+    // Stop all worker threads...
+    {
+      std::unique_lock<std::mutex> lock(mutex_);
+      shutdown_ = true;
+    }
+
+    cond_.notify_all();
+
+    // Join...
+    for (auto &t : threads_) {
+      t.join();
+    }
+  }
+
+private:
+  struct worker {
+    explicit worker(ThreadPool &pool) : pool_(pool) {}
+
+    void operator()() {
+      for (;;) {
+        std::function<void()> fn;
+        {
+          std::unique_lock<std::mutex> lock(pool_.mutex_);
+
+          pool_.cond_.wait(
+              lock, [&] { return !pool_.jobs_.empty() || pool_.shutdown_; });
+
+          if (pool_.shutdown_ && pool_.jobs_.empty()) { break; }
+
+          fn = pool_.jobs_.front();
+          pool_.jobs_.pop_front();
+        }
+
+        assert(true == static_cast<bool>(fn));
+        fn();
+      }
+    }
+
+    ThreadPool &pool_;
+  };
+  friend struct worker;
+
+  std::vector<std::thread> threads_;
+  std::list<std::function<void()>> jobs_;
+
+  bool shutdown_;
+  size_t max_queued_requests_ = 0;
+
+  std::condition_variable cond_;
+  std::mutex mutex_;
+};
+
+using Logger = std::function<void(const Request &, const Response &)>;
+
+using SocketOptions = std::function<void(socket_t sock)>;
+
+void default_socket_options(socket_t sock);
+
+const char *status_message(int status);
+
+std::string get_bearer_token_auth(const Request &req);
+
+namespace detail {
+
+class MatcherBase {
+public:
+  virtual ~MatcherBase() = default;
+
+  // Match request path and populate its matches and
+  virtual bool match(Request &request) const = 0;
+};
+
+/**
+ * Captures parameters in request path and stores them in Request::path_params
+ *
+ * Capture name is a substring of a pattern from : to /.
+ * The rest of the pattern is matched agains the request path directly
+ * Parameters are captured starting from the next character after
+ * the end of the last matched static pattern fragment until the next /.
+ *
+ * Example pattern:
+ * "/path/fragments/:capture/more/fragments/:second_capture"
+ * Static fragments:
+ * "/path/fragments/", "more/fragments/"
+ *
+ * Given the following request path:
+ * "/path/fragments/:1/more/fragments/:2"
+ * the resulting capture will be
+ * {{"capture", "1"}, {"second_capture", "2"}}
+ */
+class PathParamsMatcher final : public MatcherBase {
+public:
+  PathParamsMatcher(const std::string &pattern);
+
+  bool match(Request &request) const override;
+
+private:
+  static constexpr char marker = ':';
+  // Treat segment separators as the end of path parameter capture
+  // Does not need to handle query parameters as they are parsed before path
+  // matching
+  static constexpr char separator = '/';
+
+  // Contains static path fragments to match against, excluding the '/' after
+  // path params
+  // Fragments are separated by path params
+  std::vector<std::string> static_fragments_;
+  // Stores the names of the path parameters to be used as keys in the
+  // Request::path_params map
+  std::vector<std::string> param_names_;
+};
+
+/**
+ * Performs std::regex_match on request path
+ * and stores the result in Request::matches
+ *
+ * Note that regex match is performed directly on the whole request.
+ * This means that wildcard patterns may match multiple path segments with /:
+ * "/begin/(.*)/end" will match both "/begin/middle/end" and "/begin/1/2/end".
+ */
+class RegexMatcher final : public MatcherBase {
+public:
+  RegexMatcher(const std::string &pattern) : regex_(pattern) {}
+
+  bool match(Request &request) const override;
+
+private:
+  std::regex regex_;
+};
+
+ssize_t write_headers(Stream &strm, const Headers &headers);
+
+} // namespace detail
+
+class Server {
+public:
+  using Handler = std::function<void(const Request &, Response &)>;
+
+  using ExceptionHandler =
+      std::function<void(const Request &, Response &, std::exception_ptr ep)>;
+
+  enum class HandlerResponse {
+    Handled,
+    Unhandled,
+  };
+  using HandlerWithResponse =
+      std::function<HandlerResponse(const Request &, Response &)>;
+
+  using HandlerWithContentReader = std::function<void(
+      const Request &, Response &, const ContentReader &content_reader)>;
+
+  using Expect100ContinueHandler =
+      std::function<int(const Request &, Response &)>;
+
+  Server();
+
+  virtual ~Server();
+
+  virtual bool is_valid() const;
+
+  Server &Get(const std::string &pattern, Handler handler);
+  Server &Post(const std::string &pattern, Handler handler);
+  Server &Post(const std::string &pattern, HandlerWithContentReader handler);
+  Server &Put(const std::string &pattern, Handler handler);
+  Server &Put(const std::string &pattern, HandlerWithContentReader handler);
+  Server &Patch(const std::string &pattern, Handler handler);
+  Server &Patch(const std::string &pattern, HandlerWithContentReader handler);
+  Server &Delete(const std::string &pattern, Handler handler);
+  Server &Delete(const std::string &pattern, HandlerWithContentReader handler);
+  Server &Options(const std::string &pattern, Handler handler);
+
+  bool set_base_dir(const std::string &dir,
+                    const std::string &mount_point = std::string());
+  bool set_mount_point(const std::string &mount_point, const std::string &dir,
+                       Headers headers = Headers());
+  bool remove_mount_point(const std::string &mount_point);
+  Server &set_file_extension_and_mimetype_mapping(const std::string &ext,
+                                                  const std::string &mime);
+  Server &set_default_file_mimetype(const std::string &mime);
+  Server &set_file_request_handler(Handler handler);
+
+  template <class ErrorHandlerFunc>
+  Server &set_error_handler(ErrorHandlerFunc &&handler) {
+    return set_error_handler_core(
+        std::forward<ErrorHandlerFunc>(handler),
+        std::is_convertible<ErrorHandlerFunc, HandlerWithResponse>{});
+  }
+
+  Server &set_exception_handler(ExceptionHandler handler);
+  Server &set_pre_routing_handler(HandlerWithResponse handler);
+  Server &set_post_routing_handler(Handler handler);
+
+  Server &set_expect_100_continue_handler(Expect100ContinueHandler handler);
+  Server &set_logger(Logger logger);
+
+  Server &set_address_family(int family);
+  Server &set_tcp_nodelay(bool on);
+  Server &set_socket_options(SocketOptions socket_options);
+
+  Server &set_default_headers(Headers headers);
+  Server &
+  set_header_writer(std::function<ssize_t(Stream &, Headers &)> const &writer);
+
+  Server &set_keep_alive_max_count(size_t count);
+  Server &set_keep_alive_timeout(time_t sec);
+
+  Server &set_read_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  Server &set_read_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  Server &set_write_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  Server &set_write_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  Server &set_idle_interval(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  Server &set_idle_interval(const std::chrono::duration<Rep, Period> &duration);
+
+  Server &set_payload_max_length(size_t length);
+
+  bool bind_to_port(const std::string &host, int port, int socket_flags = 0);
+  int bind_to_any_port(const std::string &host, int socket_flags = 0);
+  bool listen_after_bind();
+
+  bool listen(const std::string &host, int port, int socket_flags = 0);
+
+  bool is_running() const;
+  void wait_until_ready() const;
+  void stop();
+
+  std::function<TaskQueue *(void)> new_task_queue;
+
+protected:
+  bool process_request(Stream &strm, bool close_connection,
+                       bool &connection_closed,
+                       const std::function<void(Request &)> &setup_request);
+
+  std::atomic<socket_t> svr_sock_{INVALID_SOCKET};
+  size_t keep_alive_max_count_ = CPPHTTPLIB_KEEPALIVE_MAX_COUNT;
+  time_t keep_alive_timeout_sec_ = CPPHTTPLIB_KEEPALIVE_TIMEOUT_SECOND;
+  time_t read_timeout_sec_ = CPPHTTPLIB_READ_TIMEOUT_SECOND;
+  time_t read_timeout_usec_ = CPPHTTPLIB_READ_TIMEOUT_USECOND;
+  time_t write_timeout_sec_ = CPPHTTPLIB_WRITE_TIMEOUT_SECOND;
+  time_t write_timeout_usec_ = CPPHTTPLIB_WRITE_TIMEOUT_USECOND;
+  time_t idle_interval_sec_ = CPPHTTPLIB_IDLE_INTERVAL_SECOND;
+  time_t idle_interval_usec_ = CPPHTTPLIB_IDLE_INTERVAL_USECOND;
+  size_t payload_max_length_ = CPPHTTPLIB_PAYLOAD_MAX_LENGTH;
+
+private:
+  using Handlers =
+      std::vector<std::pair<std::unique_ptr<detail::MatcherBase>, Handler>>;
+  using HandlersForContentReader =
+      std::vector<std::pair<std::unique_ptr<detail::MatcherBase>,
+                            HandlerWithContentReader>>;
+
+  static std::unique_ptr<detail::MatcherBase>
+  make_matcher(const std::string &pattern);
+
+  Server &set_error_handler_core(HandlerWithResponse handler, std::true_type);
+  Server &set_error_handler_core(Handler handler, std::false_type);
+
+  socket_t create_server_socket(const std::string &host, int port,
+                                int socket_flags,
+                                SocketOptions socket_options) const;
+  int bind_internal(const std::string &host, int port, int socket_flags);
+  bool listen_internal();
+
+  bool routing(Request &req, Response &res, Stream &strm);
+  bool handle_file_request(const Request &req, Response &res,
+                           bool head = false);
+  bool dispatch_request(Request &req, Response &res,
+                        const Handlers &handlers) const;
+  bool dispatch_request_for_content_reader(
+      Request &req, Response &res, ContentReader content_reader,
+      const HandlersForContentReader &handlers) const;
+
+  bool parse_request_line(const char *s, Request &req) const;
+  void apply_ranges(const Request &req, Response &res,
+                    std::string &content_type, std::string &boundary) const;
+  bool write_response(Stream &strm, bool close_connection, Request &req,
+                      Response &res);
+  bool write_response_with_content(Stream &strm, bool close_connection,
+                                   const Request &req, Response &res);
+  bool write_response_core(Stream &strm, bool close_connection,
+                           const Request &req, Response &res,
+                           bool need_apply_ranges);
+  bool write_content_with_provider(Stream &strm, const Request &req,
+                                   Response &res, const std::string &boundary,
+                                   const std::string &content_type);
+  bool read_content(Stream &strm, Request &req, Response &res);
+  bool
+  read_content_with_content_receiver(Stream &strm, Request &req, Response &res,
+                                     ContentReceiver receiver,
+                                     MultipartContentHeader multipart_header,
+                                     ContentReceiver multipart_receiver);
+  bool read_content_core(Stream &strm, Request &req, Response &res,
+                         ContentReceiver receiver,
+                         MultipartContentHeader multipart_header,
+                         ContentReceiver multipart_receiver) const;
+
+  virtual bool process_and_close_socket(socket_t sock);
+
+  std::atomic<bool> is_running_{false};
+  std::atomic<bool> done_{false};
+
+  struct MountPointEntry {
+    std::string mount_point;
+    std::string base_dir;
+    Headers headers;
+  };
+  std::vector<MountPointEntry> base_dirs_;
+  std::map<std::string, std::string> file_extension_and_mimetype_map_;
+  std::string default_file_mimetype_ = "application/octet-stream";
+  Handler file_request_handler_;
+
+  Handlers get_handlers_;
+  Handlers post_handlers_;
+  HandlersForContentReader post_handlers_for_content_reader_;
+  Handlers put_handlers_;
+  HandlersForContentReader put_handlers_for_content_reader_;
+  Handlers patch_handlers_;
+  HandlersForContentReader patch_handlers_for_content_reader_;
+  Handlers delete_handlers_;
+  HandlersForContentReader delete_handlers_for_content_reader_;
+  Handlers options_handlers_;
+
+  HandlerWithResponse error_handler_;
+  ExceptionHandler exception_handler_;
+  HandlerWithResponse pre_routing_handler_;
+  Handler post_routing_handler_;
+  Expect100ContinueHandler expect_100_continue_handler_;
+
+  Logger logger_;
+
+  int address_family_ = AF_UNSPEC;
+  bool tcp_nodelay_ = CPPHTTPLIB_TCP_NODELAY;
+  SocketOptions socket_options_ = default_socket_options;
+
+  Headers default_headers_;
+  std::function<ssize_t(Stream &, Headers &)> header_writer_ =
+      detail::write_headers;
+};
+
+enum class Error {
+  Success = 0,
+  Unknown,
+  Connection,
+  BindIPAddress,
+  Read,
+  Write,
+  ExceedRedirectCount,
+  Canceled,
+  SSLConnection,
+  SSLLoadingCerts,
+  SSLServerVerification,
+  UnsupportedMultipartBoundaryChars,
+  Compression,
+  ConnectionTimeout,
+  ProxyConnection,
+
+  // For internal use only
+  SSLPeerCouldBeClosed_,
+};
+
+std::string to_string(Error error);
+
+std::ostream &operator<<(std::ostream &os, const Error &obj);
+
+class Result {
+public:
+  Result() = default;
+  Result(std::unique_ptr<Response> &&res, Error err,
+         Headers &&request_headers = Headers{})
+      : res_(std::move(res)), err_(err),
+        request_headers_(std::move(request_headers)) {}
+  // Response
+  operator bool() const { return res_ != nullptr; }
+  bool operator==(std::nullptr_t) const { return res_ == nullptr; }
+  bool operator!=(std::nullptr_t) const { return res_ != nullptr; }
+  const Response &value() const { return *res_; }
+  Response &value() { return *res_; }
+  const Response &operator*() const { return *res_; }
+  Response &operator*() { return *res_; }
+  const Response *operator->() const { return res_.get(); }
+  Response *operator->() { return res_.get(); }
+
+  // Error
+  Error error() const { return err_; }
+
+  // Request Headers
+  bool has_request_header(const std::string &key) const;
+  std::string get_request_header_value(const std::string &key,
+                                       size_t id = 0) const;
+  uint64_t get_request_header_value_u64(const std::string &key,
+                                        size_t id = 0) const;
+  size_t get_request_header_value_count(const std::string &key) const;
+
+private:
+  std::unique_ptr<Response> res_;
+  Error err_ = Error::Unknown;
+  Headers request_headers_;
+};
+
+class ClientImpl {
+public:
+  explicit ClientImpl(const std::string &host);
+
+  explicit ClientImpl(const std::string &host, int port);
+
+  explicit ClientImpl(const std::string &host, int port,
+                      const std::string &client_cert_path,
+                      const std::string &client_key_path);
+
+  virtual ~ClientImpl();
+
+  virtual bool is_valid() const;
+
+  Result Get(const std::string &path);
+  Result Get(const std::string &path, const Headers &headers);
+  Result Get(const std::string &path, Progress progress);
+  Result Get(const std::string &path, const Headers &headers,
+             Progress progress);
+  Result Get(const std::string &path, ContentReceiver content_receiver);
+  Result Get(const std::string &path, const Headers &headers,
+             ContentReceiver content_receiver);
+  Result Get(const std::string &path, ContentReceiver content_receiver,
+             Progress progress);
+  Result Get(const std::string &path, const Headers &headers,
+             ContentReceiver content_receiver, Progress progress);
+  Result Get(const std::string &path, ResponseHandler response_handler,
+             ContentReceiver content_receiver);
+  Result Get(const std::string &path, const Headers &headers,
+             ResponseHandler response_handler,
+             ContentReceiver content_receiver);
+  Result Get(const std::string &path, ResponseHandler response_handler,
+             ContentReceiver content_receiver, Progress progress);
+  Result Get(const std::string &path, const Headers &headers,
+             ResponseHandler response_handler, ContentReceiver content_receiver,
+             Progress progress);
+
+  Result Get(const std::string &path, const Params &params,
+             const Headers &headers, Progress progress = nullptr);
+  Result Get(const std::string &path, const Params &params,
+             const Headers &headers, ContentReceiver content_receiver,
+             Progress progress = nullptr);
+  Result Get(const std::string &path, const Params &params,
+             const Headers &headers, ResponseHandler response_handler,
+             ContentReceiver content_receiver, Progress progress = nullptr);
+
+  Result Head(const std::string &path);
+  Result Head(const std::string &path, const Headers &headers);
+
+  Result Post(const std::string &path);
+  Result Post(const std::string &path, const Headers &headers);
+  Result Post(const std::string &path, const char *body, size_t content_length,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers, const char *body,
+              size_t content_length, const std::string &content_type);
+  Result Post(const std::string &path, const std::string &body,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers,
+              const std::string &body, const std::string &content_type);
+  Result Post(const std::string &path, size_t content_length,
+              ContentProvider content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path,
+              ContentProviderWithoutLength content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers,
+              size_t content_length, ContentProvider content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers,
+              ContentProviderWithoutLength content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Params &params);
+  Result Post(const std::string &path, const Headers &headers,
+              const Params &params);
+  Result Post(const std::string &path, const MultipartFormDataItems &items);
+  Result Post(const std::string &path, const Headers &headers,
+              const MultipartFormDataItems &items);
+  Result Post(const std::string &path, const Headers &headers,
+              const MultipartFormDataItems &items, const std::string &boundary);
+  Result Post(const std::string &path, const Headers &headers,
+              const MultipartFormDataItems &items,
+              const MultipartFormDataProviderItems &provider_items);
+
+  Result Put(const std::string &path);
+  Result Put(const std::string &path, const char *body, size_t content_length,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers, const char *body,
+             size_t content_length, const std::string &content_type);
+  Result Put(const std::string &path, const std::string &body,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers,
+             const std::string &body, const std::string &content_type);
+  Result Put(const std::string &path, size_t content_length,
+             ContentProvider content_provider, const std::string &content_type);
+  Result Put(const std::string &path,
+             ContentProviderWithoutLength content_provider,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers,
+             size_t content_length, ContentProvider content_provider,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers,
+             ContentProviderWithoutLength content_provider,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Params &params);
+  Result Put(const std::string &path, const Headers &headers,
+             const Params &params);
+  Result Put(const std::string &path, const MultipartFormDataItems &items);
+  Result Put(const std::string &path, const Headers &headers,
+             const MultipartFormDataItems &items);
+  Result Put(const std::string &path, const Headers &headers,
+             const MultipartFormDataItems &items, const std::string &boundary);
+  Result Put(const std::string &path, const Headers &headers,
+             const MultipartFormDataItems &items,
+             const MultipartFormDataProviderItems &provider_items);
+
+  Result Patch(const std::string &path);
+  Result Patch(const std::string &path, const char *body, size_t content_length,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               const char *body, size_t content_length,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const std::string &body,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               const std::string &body, const std::string &content_type);
+  Result Patch(const std::string &path, size_t content_length,
+               ContentProvider content_provider,
+               const std::string &content_type);
+  Result Patch(const std::string &path,
+               ContentProviderWithoutLength content_provider,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               size_t content_length, ContentProvider content_provider,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               ContentProviderWithoutLength content_provider,
+               const std::string &content_type);
+
+  Result Delete(const std::string &path);
+  Result Delete(const std::string &path, const Headers &headers);
+  Result Delete(const std::string &path, const char *body,
+                size_t content_length, const std::string &content_type);
+  Result Delete(const std::string &path, const Headers &headers,
+                const char *body, size_t content_length,
+                const std::string &content_type);
+  Result Delete(const std::string &path, const std::string &body,
+                const std::string &content_type);
+  Result Delete(const std::string &path, const Headers &headers,
+                const std::string &body, const std::string &content_type);
+
+  Result Options(const std::string &path);
+  Result Options(const std::string &path, const Headers &headers);
+
+  bool send(Request &req, Response &res, Error &error);
+  Result send(const Request &req);
+
+  void stop();
+
+  std::string host() const;
+  int port() const;
+
+  size_t is_socket_open() const;
+  socket_t socket() const;
+
+  void set_hostname_addr_map(std::map<std::string, std::string> addr_map);
+
+  void set_default_headers(Headers headers);
+
+  void
+  set_header_writer(std::function<ssize_t(Stream &, Headers &)> const &writer);
+
+  void set_address_family(int family);
+  void set_tcp_nodelay(bool on);
+  void set_socket_options(SocketOptions socket_options);
+
+  void set_connection_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  void
+  set_connection_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  void set_read_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  void set_read_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  void set_write_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  void set_write_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  void set_basic_auth(const std::string &username, const std::string &password);
+  void set_bearer_token_auth(const std::string &token);
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void set_digest_auth(const std::string &username,
+                       const std::string &password);
+#endif
+
+  void set_keep_alive(bool on);
+  void set_follow_location(bool on);
+
+  void set_url_encode(bool on);
+
+  void set_compress(bool on);
+
+  void set_decompress(bool on);
+
+  void set_interface(const std::string &intf);
+
+  void set_proxy(const std::string &host, int port);
+  void set_proxy_basic_auth(const std::string &username,
+                            const std::string &password);
+  void set_proxy_bearer_token_auth(const std::string &token);
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void set_proxy_digest_auth(const std::string &username,
+                             const std::string &password);
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void set_ca_cert_path(const std::string &ca_cert_file_path,
+                        const std::string &ca_cert_dir_path = std::string());
+  void set_ca_cert_store(X509_STORE *ca_cert_store);
+  X509_STORE *create_ca_cert_store(const char *ca_cert, std::size_t size) const;
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void enable_server_certificate_verification(bool enabled);
+#endif
+
+  void set_logger(Logger logger);
+
+protected:
+  struct Socket {
+    socket_t sock = INVALID_SOCKET;
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+    SSL *ssl = nullptr;
+#endif
+
+    bool is_open() const { return sock != INVALID_SOCKET; }
+  };
+
+  virtual bool create_and_connect_socket(Socket &socket, Error &error);
+
+  // All of:
+  //   shutdown_ssl
+  //   shutdown_socket
+  //   close_socket
+  // should ONLY be called when socket_mutex_ is locked.
+  // Also, shutdown_ssl and close_socket should also NOT be called concurrently
+  // with a DIFFERENT thread sending requests using that socket.
+  virtual void shutdown_ssl(Socket &socket, bool shutdown_gracefully);
+  void shutdown_socket(Socket &socket) const;
+  void close_socket(Socket &socket);
+
+  bool process_request(Stream &strm, Request &req, Response &res,
+                       bool close_connection, Error &error);
+
+  bool write_content_with_provider(Stream &strm, const Request &req,
+                                   Error &error) const;
+
+  void copy_settings(const ClientImpl &rhs);
+
+  // Socket endpoint information
+  const std::string host_;
+  const int port_;
+  const std::string host_and_port_;
+
+  // Current open socket
+  Socket socket_;
+  mutable std::mutex socket_mutex_;
+  std::recursive_mutex request_mutex_;
+
+  // These are all protected under socket_mutex
+  size_t socket_requests_in_flight_ = 0;
+  std::thread::id socket_requests_are_from_thread_ = std::thread::id();
+  bool socket_should_be_closed_when_request_is_done_ = false;
+
+  // Hostname-IP map
+  std::map<std::string, std::string> addr_map_;
+
+  // Default headers
+  Headers default_headers_;
+
+  // Header writer
+  std::function<ssize_t(Stream &, Headers &)> header_writer_ =
+      detail::write_headers;
+
+  // Settings
+  std::string client_cert_path_;
+  std::string client_key_path_;
+
+  time_t connection_timeout_sec_ = CPPHTTPLIB_CONNECTION_TIMEOUT_SECOND;
+  time_t connection_timeout_usec_ = CPPHTTPLIB_CONNECTION_TIMEOUT_USECOND;
+  time_t read_timeout_sec_ = CPPHTTPLIB_READ_TIMEOUT_SECOND;
+  time_t read_timeout_usec_ = CPPHTTPLIB_READ_TIMEOUT_USECOND;
+  time_t write_timeout_sec_ = CPPHTTPLIB_WRITE_TIMEOUT_SECOND;
+  time_t write_timeout_usec_ = CPPHTTPLIB_WRITE_TIMEOUT_USECOND;
+
+  std::string basic_auth_username_;
+  std::string basic_auth_password_;
+  std::string bearer_token_auth_token_;
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  std::string digest_auth_username_;
+  std::string digest_auth_password_;
+#endif
+
+  bool keep_alive_ = false;
+  bool follow_location_ = false;
+
+  bool url_encode_ = true;
+
+  int address_family_ = AF_UNSPEC;
+  bool tcp_nodelay_ = CPPHTTPLIB_TCP_NODELAY;
+  SocketOptions socket_options_ = nullptr;
+
+  bool compress_ = false;
+  bool decompress_ = true;
+
+  std::string interface_;
+
+  std::string proxy_host_;
+  int proxy_port_ = -1;
+
+  std::string proxy_basic_auth_username_;
+  std::string proxy_basic_auth_password_;
+  std::string proxy_bearer_token_auth_token_;
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  std::string proxy_digest_auth_username_;
+  std::string proxy_digest_auth_password_;
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  std::string ca_cert_file_path_;
+  std::string ca_cert_dir_path_;
+
+  X509_STORE *ca_cert_store_ = nullptr;
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  bool server_certificate_verification_ = true;
+#endif
+
+  Logger logger_;
+
+private:
+  bool send_(Request &req, Response &res, Error &error);
+  Result send_(Request &&req);
+
+  socket_t create_client_socket(Error &error) const;
+  bool read_response_line(Stream &strm, const Request &req,
+                          Response &res) const;
+  bool write_request(Stream &strm, Request &req, bool close_connection,
+                     Error &error);
+  bool redirect(Request &req, Response &res, Error &error);
+  bool handle_request(Stream &strm, Request &req, Response &res,
+                      bool close_connection, Error &error);
+  std::unique_ptr<Response> send_with_content_provider(
+      Request &req, const char *body, size_t content_length,
+      ContentProvider content_provider,
+      ContentProviderWithoutLength content_provider_without_length,
+      const std::string &content_type, Error &error);
+  Result send_with_content_provider(
+      const std::string &method, const std::string &path,
+      const Headers &headers, const char *body, size_t content_length,
+      ContentProvider content_provider,
+      ContentProviderWithoutLength content_provider_without_length,
+      const std::string &content_type);
+  ContentProviderWithoutLength get_multipart_content_provider(
+      const std::string &boundary, const MultipartFormDataItems &items,
+      const MultipartFormDataProviderItems &provider_items) const;
+
+  std::string adjust_host_string(const std::string &host) const;
+
+  virtual bool process_socket(const Socket &socket,
+                              std::function<bool(Stream &strm)> callback);
+  virtual bool is_ssl() const;
+};
+
+class Client {
+public:
+  // Universal interface
+  explicit Client(const std::string &scheme_host_port);
+
+  explicit Client(const std::string &scheme_host_port,
+                  const std::string &client_cert_path,
+                  const std::string &client_key_path);
+
+  // HTTP only interface
+  explicit Client(const std::string &host, int port);
+
+  explicit Client(const std::string &host, int port,
+                  const std::string &client_cert_path,
+                  const std::string &client_key_path);
+
+  Client(Client &&) = default;
+
+  ~Client();
+
+  bool is_valid() const;
+
+  Result Get(const std::string &path);
+  Result Get(const std::string &path, const Headers &headers);
+  Result Get(const std::string &path, Progress progress);
+  Result Get(const std::string &path, const Headers &headers,
+             Progress progress);
+  Result Get(const std::string &path, ContentReceiver content_receiver);
+  Result Get(const std::string &path, const Headers &headers,
+             ContentReceiver content_receiver);
+  Result Get(const std::string &path, ContentReceiver content_receiver,
+             Progress progress);
+  Result Get(const std::string &path, const Headers &headers,
+             ContentReceiver content_receiver, Progress progress);
+  Result Get(const std::string &path, ResponseHandler response_handler,
+             ContentReceiver content_receiver);
+  Result Get(const std::string &path, const Headers &headers,
+             ResponseHandler response_handler,
+             ContentReceiver content_receiver);
+  Result Get(const std::string &path, const Headers &headers,
+             ResponseHandler response_handler, ContentReceiver content_receiver,
+             Progress progress);
+  Result Get(const std::string &path, ResponseHandler response_handler,
+             ContentReceiver content_receiver, Progress progress);
+
+  Result Get(const std::string &path, const Params &params,
+             const Headers &headers, Progress progress = nullptr);
+  Result Get(const std::string &path, const Params &params,
+             const Headers &headers, ContentReceiver content_receiver,
+             Progress progress = nullptr);
+  Result Get(const std::string &path, const Params &params,
+             const Headers &headers, ResponseHandler response_handler,
+             ContentReceiver content_receiver, Progress progress = nullptr);
+
+  Result Head(const std::string &path);
+  Result Head(const std::string &path, const Headers &headers);
+
+  Result Post(const std::string &path);
+  Result Post(const std::string &path, const Headers &headers);
+  Result Post(const std::string &path, const char *body, size_t content_length,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers, const char *body,
+              size_t content_length, const std::string &content_type);
+  Result Post(const std::string &path, const std::string &body,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers,
+              const std::string &body, const std::string &content_type);
+  Result Post(const std::string &path, size_t content_length,
+              ContentProvider content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path,
+              ContentProviderWithoutLength content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers,
+              size_t content_length, ContentProvider content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers,
+              ContentProviderWithoutLength content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Params &params);
+  Result Post(const std::string &path, const Headers &headers,
+              const Params &params);
+  Result Post(const std::string &path, const MultipartFormDataItems &items);
+  Result Post(const std::string &path, const Headers &headers,
+              const MultipartFormDataItems &items);
+  Result Post(const std::string &path, const Headers &headers,
+              const MultipartFormDataItems &items, const std::string &boundary);
+  Result Post(const std::string &path, const Headers &headers,
+              const MultipartFormDataItems &items,
+              const MultipartFormDataProviderItems &provider_items);
+
+  Result Put(const std::string &path);
+  Result Put(const std::string &path, const char *body, size_t content_length,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers, const char *body,
+             size_t content_length, const std::string &content_type);
+  Result Put(const std::string &path, const std::string &body,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers,
+             const std::string &body, const std::string &content_type);
+  Result Put(const std::string &path, size_t content_length,
+             ContentProvider content_provider, const std::string &content_type);
+  Result Put(const std::string &path,
+             ContentProviderWithoutLength content_provider,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers,
+             size_t content_length, ContentProvider content_provider,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers,
+             ContentProviderWithoutLength content_provider,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Params &params);
+  Result Put(const std::string &path, const Headers &headers,
+             const Params &params);
+  Result Put(const std::string &path, const MultipartFormDataItems &items);
+  Result Put(const std::string &path, const Headers &headers,
+             const MultipartFormDataItems &items);
+  Result Put(const std::string &path, const Headers &headers,
+             const MultipartFormDataItems &items, const std::string &boundary);
+  Result Put(const std::string &path, const Headers &headers,
+             const MultipartFormDataItems &items,
+             const MultipartFormDataProviderItems &provider_items);
+
+  Result Patch(const std::string &path);
+  Result Patch(const std::string &path, const char *body, size_t content_length,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               const char *body, size_t content_length,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const std::string &body,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               const std::string &body, const std::string &content_type);
+  Result Patch(const std::string &path, size_t content_length,
+               ContentProvider content_provider,
+               const std::string &content_type);
+  Result Patch(const std::string &path,
+               ContentProviderWithoutLength content_provider,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               size_t content_length, ContentProvider content_provider,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               ContentProviderWithoutLength content_provider,
+               const std::string &content_type);
+
+  Result Delete(const std::string &path);
+  Result Delete(const std::string &path, const Headers &headers);
+  Result Delete(const std::string &path, const char *body,
+                size_t content_length, const std::string &content_type);
+  Result Delete(const std::string &path, const Headers &headers,
+                const char *body, size_t content_length,
+                const std::string &content_type);
+  Result Delete(const std::string &path, const std::string &body,
+                const std::string &content_type);
+  Result Delete(const std::string &path, const Headers &headers,
+                const std::string &body, const std::string &content_type);
+
+  Result Options(const std::string &path);
+  Result Options(const std::string &path, const Headers &headers);
+
+  bool send(Request &req, Response &res, Error &error);
+  Result send(const Request &req);
+
+  void stop();
+
+  std::string host() const;
+  int port() const;
+
+  size_t is_socket_open() const;
+  socket_t socket() const;
+
+  void set_hostname_addr_map(std::map<std::string, std::string> addr_map);
+
+  void set_default_headers(Headers headers);
+
+  void
+  set_header_writer(std::function<ssize_t(Stream &, Headers &)> const &writer);
+
+  void set_address_family(int family);
+  void set_tcp_nodelay(bool on);
+  void set_socket_options(SocketOptions socket_options);
+
+  void set_connection_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  void
+  set_connection_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  void set_read_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  void set_read_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  void set_write_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  void set_write_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  void set_basic_auth(const std::string &username, const std::string &password);
+  void set_bearer_token_auth(const std::string &token);
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void set_digest_auth(const std::string &username,
+                       const std::string &password);
+#endif
+
+  void set_keep_alive(bool on);
+  void set_follow_location(bool on);
+
+  void set_url_encode(bool on);
+
+  void set_compress(bool on);
+
+  void set_decompress(bool on);
+
+  void set_interface(const std::string &intf);
+
+  void set_proxy(const std::string &host, int port);
+  void set_proxy_basic_auth(const std::string &username,
+                            const std::string &password);
+  void set_proxy_bearer_token_auth(const std::string &token);
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void set_proxy_digest_auth(const std::string &username,
+                             const std::string &password);
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void enable_server_certificate_verification(bool enabled);
+#endif
+
+  void set_logger(Logger logger);
+
+  // SSL
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void set_ca_cert_path(const std::string &ca_cert_file_path,
+                        const std::string &ca_cert_dir_path = std::string());
+
+  void set_ca_cert_store(X509_STORE *ca_cert_store);
+  void load_ca_cert_store(const char *ca_cert, std::size_t size);
+
+  long get_openssl_verify_result() const;
+
+  SSL_CTX *ssl_context() const;
+#endif
+
+private:
+  std::unique_ptr<ClientImpl> cli_;
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  bool is_ssl_ = false;
+#endif
+};
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+class SSLServer : public Server {
+public:
+  SSLServer(const char *cert_path, const char *private_key_path,
+            const char *client_ca_cert_file_path = nullptr,
+            const char *client_ca_cert_dir_path = nullptr,
+            const char *private_key_password = nullptr);
+
+  SSLServer(X509 *cert, EVP_PKEY *private_key,
+            X509_STORE *client_ca_cert_store = nullptr);
+
+  SSLServer(
+      const std::function<bool(SSL_CTX &ssl_ctx)> &setup_ssl_ctx_callback);
+
+  ~SSLServer() override;
+
+  bool is_valid() const override;
+
+  SSL_CTX *ssl_context() const;
+
+private:
+  bool process_and_close_socket(socket_t sock) override;
+
+  SSL_CTX *ctx_;
+  std::mutex ctx_mutex_;
+};
+
+class SSLClient final : public ClientImpl {
+public:
+  explicit SSLClient(const std::string &host);
+
+  explicit SSLClient(const std::string &host, int port);
+
+  explicit SSLClient(const std::string &host, int port,
+                     const std::string &client_cert_path,
+                     const std::string &client_key_path,
+                     const std::string &private_key_password = std::string());
+
+  explicit SSLClient(const std::string &host, int port, X509 *client_cert,
+                     EVP_PKEY *client_key,
+                     const std::string &private_key_password = std::string());
+
+  ~SSLClient() override;
+
+  bool is_valid() const override;
+
+  void set_ca_cert_store(X509_STORE *ca_cert_store);
+  void load_ca_cert_store(const char *ca_cert, std::size_t size);
+
+  long get_openssl_verify_result() const;
+
+  SSL_CTX *ssl_context() const;
+
+private:
+  bool create_and_connect_socket(Socket &socket, Error &error) override;
+  void shutdown_ssl(Socket &socket, bool shutdown_gracefully) override;
+  void shutdown_ssl_impl(Socket &socket, bool shutdown_gracefully);
+
+  bool process_socket(const Socket &socket,
+                      std::function<bool(Stream &strm)> callback) override;
+  bool is_ssl() const override;
+
+  bool connect_with_proxy(Socket &sock, Response &res, bool &success,
+                          Error &error);
+  bool initialize_ssl(Socket &socket, Error &error);
+
+  bool load_certs();
+
+  bool verify_host(X509 *server_cert) const;
+  bool verify_host_with_subject_alt_name(X509 *server_cert) const;
+  bool verify_host_with_common_name(X509 *server_cert) const;
+  bool check_host_name(const char *pattern, size_t pattern_len) const;
+
+  SSL_CTX *ctx_;
+  std::mutex ctx_mutex_;
+  std::once_flag initialize_cert_;
+
+  std::vector<std::string> host_components_;
+
+  long verify_result_ = 0;
+
+  friend class ClientImpl;
+};
+#endif
+
+/*
+ * Implementation of template methods.
+ */
+
+namespace detail {
+
+template <typename T, typename U>
+inline void duration_to_sec_and_usec(const T &duration, U callback) {
+  auto sec = std::chrono::duration_cast<std::chrono::seconds>(duration).count();
+  auto usec = std::chrono::duration_cast<std::chrono::microseconds>(
+                  duration - std::chrono::seconds(sec))
+                  .count();
+  callback(static_cast<time_t>(sec), static_cast<time_t>(usec));
+}
+
+inline uint64_t get_header_value_u64(const Headers &headers,
+                                     const std::string &key, size_t id,
+                                     uint64_t def) {
+  auto rng = headers.equal_range(key);
+  auto it = rng.first;
+  std::advance(it, static_cast<ssize_t>(id));
+  if (it != rng.second) {
+    return std::strtoull(it->second.data(), nullptr, 10);
+  }
+  return def;
+}
+
+} // namespace detail
+
+inline uint64_t Request::get_header_value_u64(const std::string &key,
+                                              size_t id) const {
+  return detail::get_header_value_u64(headers, key, id, 0);
+}
+
+inline uint64_t Response::get_header_value_u64(const std::string &key,
+                                               size_t id) const {
+  return detail::get_header_value_u64(headers, key, id, 0);
+}
+
+template <typename... Args>
+inline ssize_t Stream::write_format(const char *fmt, const Args &...args) {
+  const auto bufsiz = 2048;
+  std::array<char, bufsiz> buf{};
+
+  auto sn = snprintf(buf.data(), buf.size() - 1, fmt, args...);
+  if (sn <= 0) { return sn; }
+
+  auto n = static_cast<size_t>(sn);
+
+  if (n >= buf.size() - 1) {
+    std::vector<char> glowable_buf(buf.size());
+
+    while (n >= glowable_buf.size() - 1) {
+      glowable_buf.resize(glowable_buf.size() * 2);
+      n = static_cast<size_t>(
+          snprintf(&glowable_buf[0], glowable_buf.size() - 1, fmt, args...));
+    }
+    return write(&glowable_buf[0], n);
+  } else {
+    return write(buf.data(), n);
+  }
+}
+
+inline void default_socket_options(socket_t sock) {
+  int yes = 1;
+#ifdef _WIN32
+  setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
+             reinterpret_cast<const char *>(&yes), sizeof(yes));
+  setsockopt(sock, SOL_SOCKET, SO_EXCLUSIVEADDRUSE,
+             reinterpret_cast<const char *>(&yes), sizeof(yes));
+#else
+#ifdef SO_REUSEPORT
+  setsockopt(sock, SOL_SOCKET, SO_REUSEPORT,
+             reinterpret_cast<const void *>(&yes), sizeof(yes));
+#else
+  setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
+             reinterpret_cast<const void *>(&yes), sizeof(yes));
+#endif
+#endif
+}
+
+inline const char *status_message(int status) {
+  switch (status) {
+  case StatusCode::Continue_100: return "Continue";
+  case StatusCode::SwitchingProtocol_101: return "Switching Protocol";
+  case StatusCode::Processing_102: return "Processing";
+  case StatusCode::EarlyHints_103: return "Early Hints";
+  case StatusCode::OK_200: return "OK";
+  case StatusCode::Created_201: return "Created";
+  case StatusCode::Accepted_202: return "Accepted";
+  case StatusCode::NonAuthoritativeInformation_203:
+    return "Non-Authoritative Information";
+  case StatusCode::NoContent_204: return "No Content";
+  case StatusCode::ResetContent_205: return "Reset Content";
+  case StatusCode::PartialContent_206: return "Partial Content";
+  case StatusCode::MultiStatus_207: return "Multi-Status";
+  case StatusCode::AlreadyReported_208: return "Already Reported";
+  case StatusCode::IMUsed_226: return "IM Used";
+  case StatusCode::MultipleChoices_300: return "Multiple Choices";
+  case StatusCode::MovedPermanently_301: return "Moved Permanently";
+  case StatusCode::Found_302: return "Found";
+  case StatusCode::SeeOther_303: return "See Other";
+  case StatusCode::NotModified_304: return "Not Modified";
+  case StatusCode::UseProxy_305: return "Use Proxy";
+  case StatusCode::unused_306: return "unused";
+  case StatusCode::TemporaryRedirect_307: return "Temporary Redirect";
+  case StatusCode::PermanentRedirect_308: return "Permanent Redirect";
+  case StatusCode::BadRequest_400: return "Bad Request";
+  case StatusCode::Unauthorized_401: return "Unauthorized";
+  case StatusCode::PaymentRequired_402: return "Payment Required";
+  case StatusCode::Forbidden_403: return "Forbidden";
+  case StatusCode::NotFound_404: return "Not Found";
+  case StatusCode::MethodNotAllowed_405: return "Method Not Allowed";
+  case StatusCode::NotAcceptable_406: return "Not Acceptable";
+  case StatusCode::ProxyAuthenticationRequired_407:
+    return "Proxy Authentication Required";
+  case StatusCode::RequestTimeout_408: return "Request Timeout";
+  case StatusCode::Conflict_409: return "Conflict";
+  case StatusCode::Gone_410: return "Gone";
+  case StatusCode::LengthRequired_411: return "Length Required";
+  case StatusCode::PreconditionFailed_412: return "Precondition Failed";
+  case StatusCode::PayloadTooLarge_413: return "Payload Too Large";
+  case StatusCode::UriTooLong_414: return "URI Too Long";
+  case StatusCode::UnsupportedMediaType_415: return "Unsupported Media Type";
+  case StatusCode::RangeNotSatisfiable_416: return "Range Not Satisfiable";
+  case StatusCode::ExpectationFailed_417: return "Expectation Failed";
+  case StatusCode::ImATeapot_418: return "I'm a teapot";
+  case StatusCode::MisdirectedRequest_421: return "Misdirected Request";
+  case StatusCode::UnprocessableContent_422: return "Unprocessable Content";
+  case StatusCode::Locked_423: return "Locked";
+  case StatusCode::FailedDependency_424: return "Failed Dependency";
+  case StatusCode::TooEarly_425: return "Too Early";
+  case StatusCode::UpgradeRequired_426: return "Upgrade Required";
+  case StatusCode::PreconditionRequired_428: return "Precondition Required";
+  case StatusCode::TooManyRequests_429: return "Too Many Requests";
+  case StatusCode::RequestHeaderFieldsTooLarge_431:
+    return "Request Header Fields Too Large";
+  case StatusCode::UnavailableForLegalReasons_451:
+    return "Unavailable For Legal Reasons";
+  case StatusCode::NotImplemented_501: return "Not Implemented";
+  case StatusCode::BadGateway_502: return "Bad Gateway";
+  case StatusCode::ServiceUnavailable_503: return "Service Unavailable";
+  case StatusCode::GatewayTimeout_504: return "Gateway Timeout";
+  case StatusCode::HttpVersionNotSupported_505:
+    return "HTTP Version Not Supported";
+  case StatusCode::VariantAlsoNegotiates_506: return "Variant Also Negotiates";
+  case StatusCode::InsufficientStorage_507: return "Insufficient Storage";
+  case StatusCode::LoopDetected_508: return "Loop Detected";
+  case StatusCode::NotExtended_510: return "Not Extended";
+  case StatusCode::NetworkAuthenticationRequired_511:
+    return "Network Authentication Required";
+
+  default:
+  case StatusCode::InternalServerError_500: return "Internal Server Error";
+  }
+}
+
+inline std::string get_bearer_token_auth(const Request &req) {
+  if (req.has_header("Authorization")) {
+    static std::string BearerHeaderPrefix = "Bearer ";
+    return req.get_header_value("Authorization")
+        .substr(BearerHeaderPrefix.length());
+  }
+  return "";
+}
+
+template <class Rep, class Period>
+inline Server &
+Server::set_read_timeout(const std::chrono::duration<Rep, Period> &duration) {
+  detail::duration_to_sec_and_usec(
+      duration, [&](time_t sec, time_t usec) { set_read_timeout(sec, usec); });
+  return *this;
+}
+
+template <class Rep, class Period>
+inline Server &
+Server::set_write_timeout(const std::chrono::duration<Rep, Period> &duration) {
+  detail::duration_to_sec_and_usec(
+      duration, [&](time_t sec, time_t usec) { set_write_timeout(sec, usec); });
+  return *this;
+}
+
+template <class Rep, class Period>
+inline Server &
+Server::set_idle_interval(const std::chrono::duration<Rep, Period> &duration) {
+  detail::duration_to_sec_and_usec(
+      duration, [&](time_t sec, time_t usec) { set_idle_interval(sec, usec); });
+  return *this;
+}
+
+inline std::string to_string(const Error error) {
+  switch (error) {
+  case Error::Success: return "Success (no error)";
+  case Error::Connection: return "Could not establish connection";
+  case Error::BindIPAddress: return "Failed to bind IP address";
+  case Error::Read: return "Failed to read connection";
+  case Error::Write: return "Failed to write connection";
+  case Error::ExceedRedirectCount: return "Maximum redirect count exceeded";
+  case Error::Canceled: return "Connection handling canceled";
+  case Error::SSLConnection: return "SSL connection failed";
+  case Error::SSLLoadingCerts: return "SSL certificate loading failed";
+  case Error::SSLServerVerification: return "SSL server verification failed";
+  case Error::UnsupportedMultipartBoundaryChars:
+    return "Unsupported HTTP multipart boundary characters";
+  case Error::Compression: return "Compression failed";
+  case Error::ConnectionTimeout: return "Connection timed out";
+  case Error::ProxyConnection: return "Proxy connection failed";
+  case Error::Unknown: return "Unknown";
+  default: break;
+  }
+
+  return "Invalid";
+}
+
+inline std::ostream &operator<<(std::ostream &os, const Error &obj) {
+  os << to_string(obj);
+  os << " (" << static_cast<std::underlying_type<Error>::type>(obj) << ')';
+  return os;
+}
+
+inline uint64_t Result::get_request_header_value_u64(const std::string &key,
+                                                     size_t id) const {
+  return detail::get_header_value_u64(request_headers_, key, id, 0);
+}
+
+template <class Rep, class Period>
+inline void ClientImpl::set_connection_timeout(
+    const std::chrono::duration<Rep, Period> &duration) {
+  detail::duration_to_sec_and_usec(duration, [&](time_t sec, time_t usec) {
+    set_connection_timeout(sec, usec);
+  });
+}
+
+template <class Rep, class Period>
+inline void ClientImpl::set_read_timeout(
+    const std::chrono::duration<Rep, Period> &duration) {
+  detail::duration_to_sec_and_usec(
+      duration, [&](time_t sec, time_t usec) { set_read_timeout(sec, usec); });
+}
+
+template <class Rep, class Period>
+inline void ClientImpl::set_write_timeout(
+    const std::chrono::duration<Rep, Period> &duration) {
+  detail::duration_to_sec_and_usec(
+      duration, [&](time_t sec, time_t usec) { set_write_timeout(sec, usec); });
+}
+
+template <class Rep, class Period>
+inline void Client::set_connection_timeout(
+    const std::chrono::duration<Rep, Period> &duration) {
+  cli_->set_connection_timeout(duration);
+}
+
+template <class Rep, class Period>
+inline void
+Client::set_read_timeout(const std::chrono::duration<Rep, Period> &duration) {
+  cli_->set_read_timeout(duration);
+}
+
+template <class Rep, class Period>
+inline void
+Client::set_write_timeout(const std::chrono::duration<Rep, Period> &duration) {
+  cli_->set_write_timeout(duration);
+}
+
+/*
+ * Forward declarations and types that will be part of the .h file if split into
+ * .h + .cc.
+ */
+
+std::string hosted_at(const std::string &hostname);
+
+void hosted_at(const std::string &hostname, std::vector<std::string> &addrs);
+
+std::string append_query_params(const std::string &path, const Params &params);
+
+std::pair<std::string, std::string> make_range_header(const Ranges &ranges);
+
+std::pair<std::string, std::string>
+make_basic_authentication_header(const std::string &username,
+                                 const std::string &password,
+                                 bool is_proxy = false);
+
+namespace detail {
+
+std::string encode_query_param(const std::string &value);
+
+std::string decode_url(const std::string &s, bool convert_plus_to_space);
+
+void read_file(const std::string &path, std::string &out);
+
+std::string trim_copy(const std::string &s);
+
+void split(const char *b, const char *e, char d,
+           std::function<void(const char *, const char *)> fn);
+
+void split(const char *b, const char *e, char d, size_t m,
+           std::function<void(const char *, const char *)> fn);
+
+bool process_client_socket(socket_t sock, time_t read_timeout_sec,
+                           time_t read_timeout_usec, time_t write_timeout_sec,
+                           time_t write_timeout_usec,
+                           std::function<bool(Stream &)> callback);
+
+socket_t create_client_socket(
+    const std::string &host, const std::string &ip, int port,
+    int address_family, bool tcp_nodelay, SocketOptions socket_options,
+    time_t connection_timeout_sec, time_t connection_timeout_usec,
+    time_t read_timeout_sec, time_t read_timeout_usec, time_t write_timeout_sec,
+    time_t write_timeout_usec, const std::string &intf, Error &error);
+
+const char *get_header_value(const Headers &headers, const std::string &key,
+                             size_t id = 0, const char *def = nullptr);
+
+std::string params_to_query_str(const Params &params);
+
+void parse_query_text(const std::string &s, Params &params);
+
+bool parse_multipart_boundary(const std::string &content_type,
+                              std::string &boundary);
+
+bool parse_range_header(const std::string &s, Ranges &ranges);
+
+int close_socket(socket_t sock);
+
+ssize_t send_socket(socket_t sock, const void *ptr, size_t size, int flags);
+
+ssize_t read_socket(socket_t sock, void *ptr, size_t size, int flags);
+
+enum class EncodingType { None = 0, Gzip, Brotli };
+
+EncodingType encoding_type(const Request &req, const Response &res);
+
+class BufferStream final : public Stream {
+public:
+  BufferStream() = default;
+  ~BufferStream() override = default;
+
+  bool is_readable() const override;
+  bool is_writable() const override;
+  ssize_t read(char *ptr, size_t size) override;
+  ssize_t write(const char *ptr, size_t size) override;
+  void get_remote_ip_and_port(std::string &ip, int &port) const override;
+  void get_local_ip_and_port(std::string &ip, int &port) const override;
+  socket_t socket() const override;
+
+  const std::string &get_buffer() const;
+
+private:
+  std::string buffer;
+  size_t position = 0;
+};
+
+class compressor {
+public:
+  virtual ~compressor() = default;
+
+  typedef std::function<bool(const char *data, size_t data_len)> Callback;
+  virtual bool compress(const char *data, size_t data_length, bool last,
+                        Callback callback) = 0;
+};
+
+class decompressor {
+public:
+  virtual ~decompressor() = default;
+
+  virtual bool is_valid() const = 0;
+
+  typedef std::function<bool(const char *data, size_t data_len)> Callback;
+  virtual bool decompress(const char *data, size_t data_length,
+                          Callback callback) = 0;
+};
+
+class nocompressor final : public compressor {
+public:
+  ~nocompressor() override = default;
+
+  bool compress(const char *data, size_t data_length, bool /*last*/,
+                Callback callback) override;
+};
+
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+class gzip_compressor final : public compressor {
+public:
+  gzip_compressor();
+  ~gzip_compressor() override;
+
+  bool compress(const char *data, size_t data_length, bool last,
+                Callback callback) override;
+
+private:
+  bool is_valid_ = false;
+  z_stream strm_;
+};
+
+class gzip_decompressor final : public decompressor {
+public:
+  gzip_decompressor();
+  ~gzip_decompressor() override;
+
+  bool is_valid() const override;
+
+  bool decompress(const char *data, size_t data_length,
+                  Callback callback) override;
+
+private:
+  bool is_valid_ = false;
+  z_stream strm_;
+};
+#endif
+
+#ifdef CPPHTTPLIB_BROTLI_SUPPORT
+class brotli_compressor final : public compressor {
+public:
+  brotli_compressor();
+  ~brotli_compressor();
+
+  bool compress(const char *data, size_t data_length, bool last,
+                Callback callback) override;
+
+private:
+  BrotliEncoderState *state_ = nullptr;
+};
+
+class brotli_decompressor final : public decompressor {
+public:
+  brotli_decompressor();
+  ~brotli_decompressor();
+
+  bool is_valid() const override;
+
+  bool decompress(const char *data, size_t data_length,
+                  Callback callback) override;
+
+private:
+  BrotliDecoderResult decoder_r;
+  BrotliDecoderState *decoder_s = nullptr;
+};
+#endif
+
+// NOTE: until the read size reaches `fixed_buffer_size`, use `fixed_buffer`
+// to store data. The call can set memory on stack for performance.
+class stream_line_reader {
+public:
+  stream_line_reader(Stream &strm, char *fixed_buffer,
+                     size_t fixed_buffer_size);
+  const char *ptr() const;
+  size_t size() const;
+  bool end_with_crlf() const;
+  bool getline();
+
+private:
+  void append(char c);
+
+  Stream &strm_;
+  char *fixed_buffer_;
+  const size_t fixed_buffer_size_;
+  size_t fixed_buffer_used_size_ = 0;
+  std::string glowable_buffer_;
+};
+
+class mmap {
+public:
+  mmap(const char *path);
+  ~mmap();
+
+  bool open(const char *path);
+  void close();
+
+  bool is_open() const;
+  size_t size() const;
+  const char *data() const;
+
+private:
+#if defined(_WIN32)
+  HANDLE hFile_;
+  HANDLE hMapping_;
+#else
+  int fd_;
+#endif
+  size_t size_;
+  void *addr_;
+};
+
+} // namespace detail
+
+// ----------------------------------------------------------------------------
+
+/*
+ * Implementation that will be part of the .cc file if split into .h + .cc.
+ */
+
+namespace detail {
+
+inline bool is_hex(char c, int &v) {
+  if (0x20 <= c && isdigit(c)) {
+    v = c - '0';
+    return true;
+  } else if ('A' <= c && c <= 'F') {
+    v = c - 'A' + 10;
+    return true;
+  } else if ('a' <= c && c <= 'f') {
+    v = c - 'a' + 10;
+    return true;
+  }
+  return false;
+}
+
+inline bool from_hex_to_i(const std::string &s, size_t i, size_t cnt,
+                          int &val) {
+  if (i >= s.size()) { return false; }
+
+  val = 0;
+  for (; cnt; i++, cnt--) {
+    if (!s[i]) { return false; }
+    auto v = 0;
+    if (is_hex(s[i], v)) {
+      val = val * 16 + v;
+    } else {
+      return false;
+    }
+  }
+  return true;
+}
+
+inline std::string from_i_to_hex(size_t n) {
+  static const auto charset = "0123456789abcdef";
+  std::string ret;
+  do {
+    ret = charset[n & 15] + ret;
+    n >>= 4;
+  } while (n > 0);
+  return ret;
+}
+
+inline size_t to_utf8(int code, char *buff) {
+  if (code < 0x0080) {
+    buff[0] = static_cast<char>(code & 0x7F);
+    return 1;
+  } else if (code < 0x0800) {
+    buff[0] = static_cast<char>(0xC0 | ((code >> 6) & 0x1F));
+    buff[1] = static_cast<char>(0x80 | (code & 0x3F));
+    return 2;
+  } else if (code < 0xD800) {
+    buff[0] = static_cast<char>(0xE0 | ((code >> 12) & 0xF));
+    buff[1] = static_cast<char>(0x80 | ((code >> 6) & 0x3F));
+    buff[2] = static_cast<char>(0x80 | (code & 0x3F));
+    return 3;
+  } else if (code < 0xE000) { // D800 - DFFF is invalid...
+    return 0;
+  } else if (code < 0x10000) {
+    buff[0] = static_cast<char>(0xE0 | ((code >> 12) & 0xF));
+    buff[1] = static_cast<char>(0x80 | ((code >> 6) & 0x3F));
+    buff[2] = static_cast<char>(0x80 | (code & 0x3F));
+    return 3;
+  } else if (code < 0x110000) {
+    buff[0] = static_cast<char>(0xF0 | ((code >> 18) & 0x7));
+    buff[1] = static_cast<char>(0x80 | ((code >> 12) & 0x3F));
+    buff[2] = static_cast<char>(0x80 | ((code >> 6) & 0x3F));
+    buff[3] = static_cast<char>(0x80 | (code & 0x3F));
+    return 4;
+  }
+
+  // NOTREACHED
+  return 0;
+}
+
+// NOTE: This code came up with the following stackoverflow post:
+// https://stackoverflow.com/questions/180947/base64-decode-snippet-in-c
+inline std::string base64_encode(const std::string &in) {
+  static const auto lookup =
+      "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
+
+  std::string out;
+  out.reserve(in.size());
+
+  auto val = 0;
+  auto valb = -6;
+
+  for (auto c : in) {
+    val = (val << 8) + static_cast<uint8_t>(c);
+    valb += 8;
+    while (valb >= 0) {
+      out.push_back(lookup[(val >> valb) & 0x3F]);
+      valb -= 6;
+    }
+  }
+
+  if (valb > -6) { out.push_back(lookup[((val << 8) >> (valb + 8)) & 0x3F]); }
+
+  while (out.size() % 4) {
+    out.push_back('=');
+  }
+
+  return out;
+}
+
+inline bool is_file(const std::string &path) {
+#ifdef _WIN32
+  return _access_s(path.c_str(), 0) == 0;
+#else
+  struct stat st;
+  return stat(path.c_str(), &st) >= 0 && S_ISREG(st.st_mode);
+#endif
+}
+
+inline bool is_dir(const std::string &path) {
+  struct stat st;
+  return stat(path.c_str(), &st) >= 0 && S_ISDIR(st.st_mode);
+}
+
+inline bool is_valid_path(const std::string &path) {
+  size_t level = 0;
+  size_t i = 0;
+
+  // Skip slash
+  while (i < path.size() && path[i] == '/') {
+    i++;
+  }
+
+  while (i < path.size()) {
+    // Read component
+    auto beg = i;
+    while (i < path.size() && path[i] != '/') {
+      if (path[i] == '\0') {
+        return false;
+      } else if (path[i] == '\\') {
+        return false;
+      }
+      i++;
+    }
+
+    auto len = i - beg;
+    assert(len > 0);
+
+    if (!path.compare(beg, len, ".")) {
+      ;
+    } else if (!path.compare(beg, len, "..")) {
+      if (level == 0) { return false; }
+      level--;
+    } else {
+      level++;
+    }
+
+    // Skip slash
+    while (i < path.size() && path[i] == '/') {
+      i++;
+    }
+  }
+
+  return true;
+}
+
+inline std::string encode_query_param(const std::string &value) {
+  std::ostringstream escaped;
+  escaped.fill('0');
+  escaped << std::hex;
+
+  for (auto c : value) {
+    if (std::isalnum(static_cast<uint8_t>(c)) || c == '-' || c == '_' ||
+        c == '.' || c == '!' || c == '~' || c == '*' || c == '\'' || c == '(' ||
+        c == ')') {
+      escaped << c;
+    } else {
+      escaped << std::uppercase;
+      escaped << '%' << std::setw(2)
+              << static_cast<int>(static_cast<unsigned char>(c));
+      escaped << std::nouppercase;
+    }
+  }
+
+  return escaped.str();
+}
+
+inline std::string encode_url(const std::string &s) {
+  std::string result;
+  result.reserve(s.size());
+
+  for (size_t i = 0; s[i]; i++) {
+    switch (s[i]) {
+    case ' ': result += "%20"; break;
+    case '+': result += "%2B"; break;
+    case '\r': result += "%0D"; break;
+    case '\n': result += "%0A"; break;
+    case '\'': result += "%27"; break;
+    case ',': result += "%2C"; break;
+    // case ':': result += "%3A"; break; // ok? probably...
+    case ';': result += "%3B"; break;
+    default:
+      auto c = static_cast<uint8_t>(s[i]);
+      if (c >= 0x80) {
+        result += '%';
+        char hex[4];
+        auto len = snprintf(hex, sizeof(hex) - 1, "%02X", c);
+        assert(len == 2);
+        result.append(hex, static_cast<size_t>(len));
+      } else {
+        result += s[i];
+      }
+      break;
+    }
+  }
+
+  return result;
+}
+
+inline std::string decode_url(const std::string &s,
+                              bool convert_plus_to_space) {
+  std::string result;
+
+  for (size_t i = 0; i < s.size(); i++) {
+    if (s[i] == '%' && i + 1 < s.size()) {
+      if (s[i + 1] == 'u') {
+        auto val = 0;
+        if (from_hex_to_i(s, i + 2, 4, val)) {
+          // 4 digits Unicode codes
+          char buff[4];
+          size_t len = to_utf8(val, buff);
+          if (len > 0) { result.append(buff, len); }
+          i += 5; // 'u0000'
+        } else {
+          result += s[i];
+        }
+      } else {
+        auto val = 0;
+        if (from_hex_to_i(s, i + 1, 2, val)) {
+          // 2 digits hex codes
+          result += static_cast<char>(val);
+          i += 2; // '00'
+        } else {
+          result += s[i];
+        }
+      }
+    } else if (convert_plus_to_space && s[i] == '+') {
+      result += ' ';
+    } else {
+      result += s[i];
+    }
+  }
+
+  return result;
+}
+
+inline void read_file(const std::string &path, std::string &out) {
+  std::ifstream fs(path, std::ios_base::binary);
+  fs.seekg(0, std::ios_base::end);
+  auto size = fs.tellg();
+  fs.seekg(0);
+  out.resize(static_cast<size_t>(size));
+  fs.read(&out[0], static_cast<std::streamsize>(size));
+}
+
+inline std::string file_extension(const std::string &path) {
+  std::smatch m;
+  static auto re = std::regex("\\.([a-zA-Z0-9]+)$");
+  if (std::regex_search(path, m, re)) { return m[1].str(); }
+  return std::string();
+}
+
+inline bool is_space_or_tab(char c) { return c == ' ' || c == '\t'; }
+
+inline std::pair<size_t, size_t> trim(const char *b, const char *e, size_t left,
+                                      size_t right) {
+  while (b + left < e && is_space_or_tab(b[left])) {
+    left++;
+  }
+  while (right > 0 && is_space_or_tab(b[right - 1])) {
+    right--;
+  }
+  return std::make_pair(left, right);
+}
+
+inline std::string trim_copy(const std::string &s) {
+  auto r = trim(s.data(), s.data() + s.size(), 0, s.size());
+  return s.substr(r.first, r.second - r.first);
+}
+
+inline std::string trim_double_quotes_copy(const std::string &s) {
+  if (s.length() >= 2 && s.front() == '"' && s.back() == '"') {
+    return s.substr(1, s.size() - 2);
+  }
+  return s;
+}
+
+inline void split(const char *b, const char *e, char d,
+                  std::function<void(const char *, const char *)> fn) {
+  return split(b, e, d, (std::numeric_limits<size_t>::max)(), std::move(fn));
+}
+
+inline void split(const char *b, const char *e, char d, size_t m,
+                  std::function<void(const char *, const char *)> fn) {
+  size_t i = 0;
+  size_t beg = 0;
+  size_t count = 1;
+
+  while (e ? (b + i < e) : (b[i] != '\0')) {
+    if (b[i] == d && count < m) {
+      auto r = trim(b, e, beg, i);
+      if (r.first < r.second) { fn(&b[r.first], &b[r.second]); }
+      beg = i + 1;
+      count++;
+    }
+    i++;
+  }
+
+  if (i) {
+    auto r = trim(b, e, beg, i);
+    if (r.first < r.second) { fn(&b[r.first], &b[r.second]); }
+  }
+}
+
+inline stream_line_reader::stream_line_reader(Stream &strm, char *fixed_buffer,
+                                              size_t fixed_buffer_size)
+    : strm_(strm), fixed_buffer_(fixed_buffer),
+      fixed_buffer_size_(fixed_buffer_size) {}
+
+inline const char *stream_line_reader::ptr() const {
+  if (glowable_buffer_.empty()) {
+    return fixed_buffer_;
+  } else {
+    return glowable_buffer_.data();
+  }
+}
+
+inline size_t stream_line_reader::size() const {
+  if (glowable_buffer_.empty()) {
+    return fixed_buffer_used_size_;
+  } else {
+    return glowable_buffer_.size();
+  }
+}
+
+inline bool stream_line_reader::end_with_crlf() const {
+  auto end = ptr() + size();
+  return size() >= 2 && end[-2] == '\r' && end[-1] == '\n';
+}
+
+inline bool stream_line_reader::getline() {
+  fixed_buffer_used_size_ = 0;
+  glowable_buffer_.clear();
+
+  for (size_t i = 0;; i++) {
+    char byte;
+    auto n = strm_.read(&byte, 1);
+
+    if (n < 0) {
+      return false;
+    } else if (n == 0) {
+      if (i == 0) {
+        return false;
+      } else {
+        break;
+      }
+    }
+
+    append(byte);
+
+    if (byte == '\n') { break; }
+  }
+
+  return true;
+}
+
+inline void stream_line_reader::append(char c) {
+  if (fixed_buffer_used_size_ < fixed_buffer_size_ - 1) {
+    fixed_buffer_[fixed_buffer_used_size_++] = c;
+    fixed_buffer_[fixed_buffer_used_size_] = '\0';
+  } else {
+    if (glowable_buffer_.empty()) {
+      assert(fixed_buffer_[fixed_buffer_used_size_] == '\0');
+      glowable_buffer_.assign(fixed_buffer_, fixed_buffer_used_size_);
+    }
+    glowable_buffer_ += c;
+  }
+}
+
+inline mmap::mmap(const char *path)
+#if defined(_WIN32)
+    : hFile_(NULL), hMapping_(NULL)
+#else
+    : fd_(-1)
+#endif
+      ,
+      size_(0), addr_(nullptr) {
+  open(path);
+}
+
+inline mmap::~mmap() { close(); }
+
+inline bool mmap::open(const char *path) {
+  close();
+
+#if defined(_WIN32)
+  std::wstring wpath;
+  for (size_t i = 0; i < strlen(path); i++) {
+    wpath += path[i];
+  }
+
+  hFile_ = ::CreateFile2(wpath.c_str(), GENERIC_READ, FILE_SHARE_READ,
+                         OPEN_EXISTING, NULL);
+
+  if (hFile_ == INVALID_HANDLE_VALUE) { return false; }
+
+  LARGE_INTEGER size{};
+  if (!::GetFileSizeEx(hFile_, &size)) { return false; }
+  size_ = static_cast<size_t>(size.QuadPart);
+
+  hMapping_ =
+      ::CreateFileMappingFromApp(hFile_, NULL, PAGE_READONLY, size_, NULL);
+
+  if (hMapping_ == NULL) {
+    close();
+    return false;
+  }
+
+  addr_ = ::MapViewOfFileFromApp(hMapping_, FILE_MAP_READ, 0, 0);
+#else
+  fd_ = ::open(path, O_RDONLY);
+  if (fd_ == -1) { return false; }
+
+  struct stat sb;
+  if (fstat(fd_, &sb) == -1) {
+    close();
+    return false;
+  }
+  size_ = static_cast<size_t>(sb.st_size);
+
+  addr_ = ::mmap(NULL, size_, PROT_READ, MAP_PRIVATE, fd_, 0);
+#endif
+
+  if (addr_ == nullptr) {
+    close();
+    return false;
+  }
+
+  return true;
+}
+
+inline bool mmap::is_open() const { return addr_ != nullptr; }
+
+inline size_t mmap::size() const { return size_; }
+
+inline const char *mmap::data() const {
+  return static_cast<const char *>(addr_);
+}
+
+inline void mmap::close() {
+#if defined(_WIN32)
+  if (addr_) {
+    ::UnmapViewOfFile(addr_);
+    addr_ = nullptr;
+  }
+
+  if (hMapping_) {
+    ::CloseHandle(hMapping_);
+    hMapping_ = NULL;
+  }
+
+  if (hFile_ != INVALID_HANDLE_VALUE) {
+    ::CloseHandle(hFile_);
+    hFile_ = INVALID_HANDLE_VALUE;
+  }
+#else
+  if (addr_ != nullptr) {
+    munmap(addr_, size_);
+    addr_ = nullptr;
+  }
+
+  if (fd_ != -1) {
+    ::close(fd_);
+    fd_ = -1;
+  }
+#endif
+  size_ = 0;
+}
+inline int close_socket(socket_t sock) {
+#ifdef _WIN32
+  return closesocket(sock);
+#else
+  return close(sock);
+#endif
+}
+
+template <typename T> inline ssize_t handle_EINTR(T fn) {
+  ssize_t res = 0;
+  while (true) {
+    res = fn();
+    if (res < 0 && errno == EINTR) { continue; }
+    break;
+  }
+  return res;
+}
+
+inline ssize_t read_socket(socket_t sock, void *ptr, size_t size, int flags) {
+  return handle_EINTR([&]() {
+    return recv(sock,
+#ifdef _WIN32
+                static_cast<char *>(ptr), static_cast<int>(size),
+#else
+                ptr, size,
+#endif
+                flags);
+  });
+}
+
+inline ssize_t send_socket(socket_t sock, const void *ptr, size_t size,
+                           int flags) {
+  return handle_EINTR([&]() {
+    return send(sock,
+#ifdef _WIN32
+                static_cast<const char *>(ptr), static_cast<int>(size),
+#else
+                ptr, size,
+#endif
+                flags);
+  });
+}
+
+inline ssize_t select_read(socket_t sock, time_t sec, time_t usec) {
+#ifdef CPPHTTPLIB_USE_POLL
+  struct pollfd pfd_read;
+  pfd_read.fd = sock;
+  pfd_read.events = POLLIN;
+
+  auto timeout = static_cast<int>(sec * 1000 + usec / 1000);
+
+  return handle_EINTR([&]() { return poll(&pfd_read, 1, timeout); });
+#else
+#ifndef _WIN32
+  if (sock >= FD_SETSIZE) { return -1; }
+#endif
+
+  fd_set fds;
+  FD_ZERO(&fds);
+  FD_SET(sock, &fds);
+
+  timeval tv;
+  tv.tv_sec = static_cast<long>(sec);
+  tv.tv_usec = static_cast<decltype(tv.tv_usec)>(usec);
+
+  return handle_EINTR([&]() {
+    return select(static_cast<int>(sock + 1), &fds, nullptr, nullptr, &tv);
+  });
+#endif
+}
+
+inline ssize_t select_write(socket_t sock, time_t sec, time_t usec) {
+#ifdef CPPHTTPLIB_USE_POLL
+  struct pollfd pfd_read;
+  pfd_read.fd = sock;
+  pfd_read.events = POLLOUT;
+
+  auto timeout = static_cast<int>(sec * 1000 + usec / 1000);
+
+  return handle_EINTR([&]() { return poll(&pfd_read, 1, timeout); });
+#else
+#ifndef _WIN32
+  if (sock >= FD_SETSIZE) { return -1; }
+#endif
+
+  fd_set fds;
+  FD_ZERO(&fds);
+  FD_SET(sock, &fds);
+
+  timeval tv;
+  tv.tv_sec = static_cast<long>(sec);
+  tv.tv_usec = static_cast<decltype(tv.tv_usec)>(usec);
+
+  return handle_EINTR([&]() {
+    return select(static_cast<int>(sock + 1), nullptr, &fds, nullptr, &tv);
+  });
+#endif
+}
+
+inline Error wait_until_socket_is_ready(socket_t sock, time_t sec,
+                                        time_t usec) {
+#ifdef CPPHTTPLIB_USE_POLL
+  struct pollfd pfd_read;
+  pfd_read.fd = sock;
+  pfd_read.events = POLLIN | POLLOUT;
+
+  auto timeout = static_cast<int>(sec * 1000 + usec / 1000);
+
+  auto poll_res = handle_EINTR([&]() { return poll(&pfd_read, 1, timeout); });
+
+  if (poll_res == 0) { return Error::ConnectionTimeout; }
+
+  if (poll_res > 0 && pfd_read.revents & (POLLIN | POLLOUT)) {
+    auto error = 0;
+    socklen_t len = sizeof(error);
+    auto res = getsockopt(sock, SOL_SOCKET, SO_ERROR,
+                          reinterpret_cast<char *>(&error), &len);
+    auto successful = res >= 0 && !error;
+    return successful ? Error::Success : Error::Connection;
+  }
+
+  return Error::Connection;
+#else
+#ifndef _WIN32
+  if (sock >= FD_SETSIZE) { return Error::Connection; }
+#endif
+
+  fd_set fdsr;
+  FD_ZERO(&fdsr);
+  FD_SET(sock, &fdsr);
+
+  auto fdsw = fdsr;
+  auto fdse = fdsr;
+
+  timeval tv;
+  tv.tv_sec = static_cast<long>(sec);
+  tv.tv_usec = static_cast<decltype(tv.tv_usec)>(usec);
+
+  auto ret = handle_EINTR([&]() {
+    return select(static_cast<int>(sock + 1), &fdsr, &fdsw, &fdse, &tv);
+  });
+
+  if (ret == 0) { return Error::ConnectionTimeout; }
+
+  if (ret > 0 && (FD_ISSET(sock, &fdsr) || FD_ISSET(sock, &fdsw))) {
+    auto error = 0;
+    socklen_t len = sizeof(error);
+    auto res = getsockopt(sock, SOL_SOCKET, SO_ERROR,
+                          reinterpret_cast<char *>(&error), &len);
+    auto successful = res >= 0 && !error;
+    return successful ? Error::Success : Error::Connection;
+  }
+  return Error::Connection;
+#endif
+}
+
+inline bool is_socket_alive(socket_t sock) {
+  const auto val = detail::select_read(sock, 0, 0);
+  if (val == 0) {
+    return true;
+  } else if (val < 0 && errno == EBADF) {
+    return false;
+  }
+  char buf[1];
+  return detail::read_socket(sock, &buf[0], sizeof(buf), MSG_PEEK) > 0;
+}
+
+class SocketStream final : public Stream {
+public:
+  SocketStream(socket_t sock, time_t read_timeout_sec, time_t read_timeout_usec,
+               time_t write_timeout_sec, time_t write_timeout_usec);
+  ~SocketStream() override;
+
+  bool is_readable() const override;
+  bool is_writable() const override;
+  ssize_t read(char *ptr, size_t size) override;
+  ssize_t write(const char *ptr, size_t size) override;
+  void get_remote_ip_and_port(std::string &ip, int &port) const override;
+  void get_local_ip_and_port(std::string &ip, int &port) const override;
+  socket_t socket() const override;
+
+private:
+  socket_t sock_;
+  time_t read_timeout_sec_;
+  time_t read_timeout_usec_;
+  time_t write_timeout_sec_;
+  time_t write_timeout_usec_;
+
+  std::vector<char> read_buff_;
+  size_t read_buff_off_ = 0;
+  size_t read_buff_content_size_ = 0;
+
+  static const size_t read_buff_size_ = 1024l * 4;
+};
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+class SSLSocketStream final : public Stream {
+public:
+  SSLSocketStream(socket_t sock, SSL *ssl, time_t read_timeout_sec,
+                  time_t read_timeout_usec, time_t write_timeout_sec,
+                  time_t write_timeout_usec);
+  ~SSLSocketStream() override;
+
+  bool is_readable() const override;
+  bool is_writable() const override;
+  ssize_t read(char *ptr, size_t size) override;
+  ssize_t write(const char *ptr, size_t size) override;
+  void get_remote_ip_and_port(std::string &ip, int &port) const override;
+  void get_local_ip_and_port(std::string &ip, int &port) const override;
+  socket_t socket() const override;
+
+private:
+  socket_t sock_;
+  SSL *ssl_;
+  time_t read_timeout_sec_;
+  time_t read_timeout_usec_;
+  time_t write_timeout_sec_;
+  time_t write_timeout_usec_;
+};
+#endif
+
+inline bool keep_alive(socket_t sock, time_t keep_alive_timeout_sec) {
+  using namespace std::chrono;
+  auto start = steady_clock::now();
+  while (true) {
+    auto val = select_read(sock, 0, 10000);
+    if (val < 0) {
+      return false;
+    } else if (val == 0) {
+      auto current = steady_clock::now();
+      auto duration = duration_cast<milliseconds>(current - start);
+      auto timeout = keep_alive_timeout_sec * 1000;
+      if (duration.count() > timeout) { return false; }
+      std::this_thread::sleep_for(std::chrono::milliseconds(1));
+    } else {
+      return true;
+    }
+  }
+}
+
+template <typename T>
+inline bool
+process_server_socket_core(const std::atomic<socket_t> &svr_sock, socket_t sock,
+                           size_t keep_alive_max_count,
+                           time_t keep_alive_timeout_sec, T callback) {
+  assert(keep_alive_max_count > 0);
+  auto ret = false;
+  auto count = keep_alive_max_count;
+  while (svr_sock != INVALID_SOCKET && count > 0 &&
+         keep_alive(sock, keep_alive_timeout_sec)) {
+    auto close_connection = count == 1;
+    auto connection_closed = false;
+    ret = callback(close_connection, connection_closed);
+    if (!ret || connection_closed) { break; }
+    count--;
+  }
+  return ret;
+}
+
+template <typename T>
+inline bool
+process_server_socket(const std::atomic<socket_t> &svr_sock, socket_t sock,
+                      size_t keep_alive_max_count,
+                      time_t keep_alive_timeout_sec, time_t read_timeout_sec,
+                      time_t read_timeout_usec, time_t write_timeout_sec,
+                      time_t write_timeout_usec, T callback) {
+  return process_server_socket_core(
+      svr_sock, sock, keep_alive_max_count, keep_alive_timeout_sec,
+      [&](bool close_connection, bool &connection_closed) {
+        SocketStream strm(sock, read_timeout_sec, read_timeout_usec,
+                          write_timeout_sec, write_timeout_usec);
+        return callback(strm, close_connection, connection_closed);
+      });
+}
+
+inline bool process_client_socket(socket_t sock, time_t read_timeout_sec,
+                                  time_t read_timeout_usec,
+                                  time_t write_timeout_sec,
+                                  time_t write_timeout_usec,
+                                  std::function<bool(Stream &)> callback) {
+  SocketStream strm(sock, read_timeout_sec, read_timeout_usec,
+                    write_timeout_sec, write_timeout_usec);
+  return callback(strm);
+}
+
+inline int shutdown_socket(socket_t sock) {
+#ifdef _WIN32
+  return shutdown(sock, SD_BOTH);
+#else
+  return shutdown(sock, SHUT_RDWR);
+#endif
+}
+
+template <typename BindOrConnect>
+socket_t create_socket(const std::string &host, const std::string &ip, int port,
+                       int address_family, int socket_flags, bool tcp_nodelay,
+                       SocketOptions socket_options,
+                       BindOrConnect bind_or_connect) {
+  // Get address info
+  const char *node = nullptr;
+  struct addrinfo hints;
+  struct addrinfo *result;
+
+  memset(&hints, 0, sizeof(struct addrinfo));
+  hints.ai_socktype = SOCK_STREAM;
+  hints.ai_protocol = 0;
+
+  if (!ip.empty()) {
+    node = ip.c_str();
+    // Ask getaddrinfo to convert IP in c-string to address
+    hints.ai_family = AF_UNSPEC;
+    hints.ai_flags = AI_NUMERICHOST;
+  } else {
+    if (!host.empty()) { node = host.c_str(); }
+    hints.ai_family = address_family;
+    hints.ai_flags = socket_flags;
+  }
+
+#ifndef _WIN32
+  if (hints.ai_family == AF_UNIX) {
+    const auto addrlen = host.length();
+    if (addrlen > sizeof(sockaddr_un::sun_path)) { return INVALID_SOCKET; }
+
+    auto sock = socket(hints.ai_family, hints.ai_socktype, hints.ai_protocol);
+    if (sock != INVALID_SOCKET) {
+      sockaddr_un addr{};
+      addr.sun_family = AF_UNIX;
+      std::copy(host.begin(), host.end(), addr.sun_path);
+
+      hints.ai_addr = reinterpret_cast<sockaddr *>(&addr);
+      hints.ai_addrlen = static_cast<socklen_t>(
+          sizeof(addr) - sizeof(addr.sun_path) + addrlen);
+
+      fcntl(sock, F_SETFD, FD_CLOEXEC);
+      if (socket_options) { socket_options(sock); }
+
+      if (!bind_or_connect(sock, hints)) {
+        close_socket(sock);
+        sock = INVALID_SOCKET;
+      }
+    }
+    return sock;
+  }
+#endif
+
+  auto service = std::to_string(port);
+
+  if (getaddrinfo(node, service.c_str(), &hints, &result)) {
+#if defined __linux__ && !defined __ANDROID__
+    res_init();
+#endif
+    return INVALID_SOCKET;
+  }
+
+  for (auto rp = result; rp; rp = rp->ai_next) {
+    // Create a socket
+#ifdef _WIN32
+    auto sock =
+        WSASocketW(rp->ai_family, rp->ai_socktype, rp->ai_protocol, nullptr, 0,
+                   WSA_FLAG_NO_HANDLE_INHERIT | WSA_FLAG_OVERLAPPED);
+    /**
+     * Since the WSA_FLAG_NO_HANDLE_INHERIT is only supported on Windows 7 SP1
+     * and above the socket creation fails on older Windows Systems.
+     *
+     * Let's try to create a socket the old way in this case.
+     *
+     * Reference:
+     * https://docs.microsoft.com/en-us/windows/win32/api/winsock2/nf-winsock2-wsasocketa
+     *
+     * WSA_FLAG_NO_HANDLE_INHERIT:
+     * This flag is supported on Windows 7 with SP1, Windows Server 2008 R2 with
+     * SP1, and later
+     *
+     */
+    if (sock == INVALID_SOCKET) {
+      sock = socket(rp->ai_family, rp->ai_socktype, rp->ai_protocol);
+    }
+#else
+    auto sock = socket(rp->ai_family, rp->ai_socktype, rp->ai_protocol);
+#endif
+    if (sock == INVALID_SOCKET) { continue; }
+
+#ifndef _WIN32
+    if (fcntl(sock, F_SETFD, FD_CLOEXEC) == -1) {
+      close_socket(sock);
+      continue;
+    }
+#endif
+
+    if (tcp_nodelay) {
+      auto yes = 1;
+#ifdef _WIN32
+      setsockopt(sock, IPPROTO_TCP, TCP_NODELAY,
+                 reinterpret_cast<const char *>(&yes), sizeof(yes));
+#else
+      setsockopt(sock, IPPROTO_TCP, TCP_NODELAY,
+                 reinterpret_cast<const void *>(&yes), sizeof(yes));
+#endif
+    }
+
+    if (socket_options) { socket_options(sock); }
+
+    if (rp->ai_family == AF_INET6) {
+      auto no = 0;
+#ifdef _WIN32
+      setsockopt(sock, IPPROTO_IPV6, IPV6_V6ONLY,
+                 reinterpret_cast<const char *>(&no), sizeof(no));
+#else
+      setsockopt(sock, IPPROTO_IPV6, IPV6_V6ONLY,
+                 reinterpret_cast<const void *>(&no), sizeof(no));
+#endif
+    }
+
+    // bind or connect
+    if (bind_or_connect(sock, *rp)) {
+      freeaddrinfo(result);
+      return sock;
+    }
+
+    close_socket(sock);
+  }
+
+  freeaddrinfo(result);
+  return INVALID_SOCKET;
+}
+
+inline void set_nonblocking(socket_t sock, bool nonblocking) {
+#ifdef _WIN32
+  auto flags = nonblocking ? 1UL : 0UL;
+  ioctlsocket(sock, FIONBIO, &flags);
+#else
+  auto flags = fcntl(sock, F_GETFL, 0);
+  fcntl(sock, F_SETFL,
+        nonblocking ? (flags | O_NONBLOCK) : (flags & (~O_NONBLOCK)));
+#endif
+}
+
+inline bool is_connection_error() {
+#ifdef _WIN32
+  return WSAGetLastError() != WSAEWOULDBLOCK;
+#else
+  return errno != EINPROGRESS;
+#endif
+}
+
+inline bool bind_ip_address(socket_t sock, const std::string &host) {
+  struct addrinfo hints;
+  struct addrinfo *result;
+
+  memset(&hints, 0, sizeof(struct addrinfo));
+  hints.ai_family = AF_UNSPEC;
+  hints.ai_socktype = SOCK_STREAM;
+  hints.ai_protocol = 0;
+
+  if (getaddrinfo(host.c_str(), "0", &hints, &result)) { return false; }
+
+  auto ret = false;
+  for (auto rp = result; rp; rp = rp->ai_next) {
+    const auto &ai = *rp;
+    if (!::bind(sock, ai.ai_addr, static_cast<socklen_t>(ai.ai_addrlen))) {
+      ret = true;
+      break;
+    }
+  }
+
+  freeaddrinfo(result);
+  return ret;
+}
+
+#if !defined _WIN32 && !defined ANDROID && !defined _AIX && !defined __MVS__
+#define USE_IF2IP
+#endif
+
+#ifdef USE_IF2IP
+inline std::string if2ip(int address_family, const std::string &ifn) {
+  struct ifaddrs *ifap;
+  getifaddrs(&ifap);
+  std::string addr_candidate;
+  for (auto ifa = ifap; ifa; ifa = ifa->ifa_next) {
+    if (ifa->ifa_addr && ifn == ifa->ifa_name &&
+        (AF_UNSPEC == address_family ||
+         ifa->ifa_addr->sa_family == address_family)) {
+      if (ifa->ifa_addr->sa_family == AF_INET) {
+        auto sa = reinterpret_cast<struct sockaddr_in *>(ifa->ifa_addr);
+        char buf[INET_ADDRSTRLEN];
+        if (inet_ntop(AF_INET, &sa->sin_addr, buf, INET_ADDRSTRLEN)) {
+          freeifaddrs(ifap);
+          return std::string(buf, INET_ADDRSTRLEN);
+        }
+      } else if (ifa->ifa_addr->sa_family == AF_INET6) {
+        auto sa = reinterpret_cast<struct sockaddr_in6 *>(ifa->ifa_addr);
+        if (!IN6_IS_ADDR_LINKLOCAL(&sa->sin6_addr)) {
+          char buf[INET6_ADDRSTRLEN] = {};
+          if (inet_ntop(AF_INET6, &sa->sin6_addr, buf, INET6_ADDRSTRLEN)) {
+            // equivalent to mac's IN6_IS_ADDR_UNIQUE_LOCAL
+            auto s6_addr_head = sa->sin6_addr.s6_addr[0];
+            if (s6_addr_head == 0xfc || s6_addr_head == 0xfd) {
+              addr_candidate = std::string(buf, INET6_ADDRSTRLEN);
+            } else {
+              freeifaddrs(ifap);
+              return std::string(buf, INET6_ADDRSTRLEN);
+            }
+          }
+        }
+      }
+    }
+  }
+  freeifaddrs(ifap);
+  return addr_candidate;
+}
+#endif
+
+inline socket_t create_client_socket(
+    const std::string &host, const std::string &ip, int port,
+    int address_family, bool tcp_nodelay, SocketOptions socket_options,
+    time_t connection_timeout_sec, time_t connection_timeout_usec,
+    time_t read_timeout_sec, time_t read_timeout_usec, time_t write_timeout_sec,
+    time_t write_timeout_usec, const std::string &intf, Error &error) {
+  auto sock = create_socket(
+      host, ip, port, address_family, 0, tcp_nodelay, std::move(socket_options),
+      [&](socket_t sock2, struct addrinfo &ai) -> bool {
+        if (!intf.empty()) {
+#ifdef USE_IF2IP
+          auto ip_from_if = if2ip(address_family, intf);
+          if (ip_from_if.empty()) { ip_from_if = intf; }
+          if (!bind_ip_address(sock2, ip_from_if)) {
+            error = Error::BindIPAddress;
+            return false;
+          }
+#endif
+        }
+
+        set_nonblocking(sock2, true);
+
+        auto ret =
+            ::connect(sock2, ai.ai_addr, static_cast<socklen_t>(ai.ai_addrlen));
+
+        if (ret < 0) {
+          if (is_connection_error()) {
+            error = Error::Connection;
+            return false;
+          }
+          error = wait_until_socket_is_ready(sock2, connection_timeout_sec,
+                                             connection_timeout_usec);
+          if (error != Error::Success) { return false; }
+        }
+
+        set_nonblocking(sock2, false);
+
+        {
+#ifdef _WIN32
+          auto timeout = static_cast<uint32_t>(read_timeout_sec * 1000 +
+                                               read_timeout_usec / 1000);
+          setsockopt(sock2, SOL_SOCKET, SO_RCVTIMEO,
+                     reinterpret_cast<const char *>(&timeout), sizeof(timeout));
+#else
+          timeval tv;
+          tv.tv_sec = static_cast<long>(read_timeout_sec);
+          tv.tv_usec = static_cast<decltype(tv.tv_usec)>(read_timeout_usec);
+          setsockopt(sock2, SOL_SOCKET, SO_RCVTIMEO,
+                     reinterpret_cast<const void *>(&tv), sizeof(tv));
+#endif
+        }
+        {
+
+#ifdef _WIN32
+          auto timeout = static_cast<uint32_t>(write_timeout_sec * 1000 +
+                                               write_timeout_usec / 1000);
+          setsockopt(sock2, SOL_SOCKET, SO_SNDTIMEO,
+                     reinterpret_cast<const char *>(&timeout), sizeof(timeout));
+#else
+          timeval tv;
+          tv.tv_sec = static_cast<long>(write_timeout_sec);
+          tv.tv_usec = static_cast<decltype(tv.tv_usec)>(write_timeout_usec);
+          setsockopt(sock2, SOL_SOCKET, SO_SNDTIMEO,
+                     reinterpret_cast<const void *>(&tv), sizeof(tv));
+#endif
+        }
+
+        error = Error::Success;
+        return true;
+      });
+
+  if (sock != INVALID_SOCKET) {
+    error = Error::Success;
+  } else {
+    if (error == Error::Success) { error = Error::Connection; }
+  }
+
+  return sock;
+}
+
+inline bool get_ip_and_port(const struct sockaddr_storage &addr,
+                            socklen_t addr_len, std::string &ip, int &port) {
+  if (addr.ss_family == AF_INET) {
+    port = ntohs(reinterpret_cast<const struct sockaddr_in *>(&addr)->sin_port);
+  } else if (addr.ss_family == AF_INET6) {
+    port =
+        ntohs(reinterpret_cast<const struct sockaddr_in6 *>(&addr)->sin6_port);
+  } else {
+    return false;
+  }
+
+  std::array<char, NI_MAXHOST> ipstr{};
+  if (getnameinfo(reinterpret_cast<const struct sockaddr *>(&addr), addr_len,
+                  ipstr.data(), static_cast<socklen_t>(ipstr.size()), nullptr,
+                  0, NI_NUMERICHOST)) {
+    return false;
+  }
+
+  ip = ipstr.data();
+  return true;
+}
+
+inline void get_local_ip_and_port(socket_t sock, std::string &ip, int &port) {
+  struct sockaddr_storage addr;
+  socklen_t addr_len = sizeof(addr);
+  if (!getsockname(sock, reinterpret_cast<struct sockaddr *>(&addr),
+                   &addr_len)) {
+    get_ip_and_port(addr, addr_len, ip, port);
+  }
+}
+
+inline void get_remote_ip_and_port(socket_t sock, std::string &ip, int &port) {
+  struct sockaddr_storage addr;
+  socklen_t addr_len = sizeof(addr);
+
+  if (!getpeername(sock, reinterpret_cast<struct sockaddr *>(&addr),
+                   &addr_len)) {
+#ifndef _WIN32
+    if (addr.ss_family == AF_UNIX) {
+#if defined(__linux__)
+      struct ucred ucred;
+      socklen_t len = sizeof(ucred);
+      if (getsockopt(sock, SOL_SOCKET, SO_PEERCRED, &ucred, &len) == 0) {
+        port = ucred.pid;
+      }
+#elif defined(SOL_LOCAL) && defined(SO_PEERPID) // __APPLE__
+      pid_t pid;
+      socklen_t len = sizeof(pid);
+      if (getsockopt(sock, SOL_LOCAL, SO_PEERPID, &pid, &len) == 0) {
+        port = pid;
+      }
+#endif
+      return;
+    }
+#endif
+    get_ip_and_port(addr, addr_len, ip, port);
+  }
+}
+
+inline constexpr unsigned int str2tag_core(const char *s, size_t l,
+                                           unsigned int h) {
+  return (l == 0)
+             ? h
+             : str2tag_core(
+                   s + 1, l - 1,
+                   // Unsets the 6 high bits of h, therefore no overflow happens
+                   (((std::numeric_limits<unsigned int>::max)() >> 6) &
+                    h * 33) ^
+                       static_cast<unsigned char>(*s));
+}
+
+inline unsigned int str2tag(const std::string &s) {
+  return str2tag_core(s.data(), s.size(), 0);
+}
+
+namespace udl {
+
+inline constexpr unsigned int operator"" _t(const char *s, size_t l) {
+  return str2tag_core(s, l, 0);
+}
+
+} // namespace udl
+
+inline std::string
+find_content_type(const std::string &path,
+                  const std::map<std::string, std::string> &user_data,
+                  const std::string &default_content_type) {
+  auto ext = file_extension(path);
+
+  auto it = user_data.find(ext);
+  if (it != user_data.end()) { return it->second; }
+
+  using udl::operator""_t;
+
+  switch (str2tag(ext)) {
+  default: return default_content_type;
+
+  case "css"_t: return "text/css";
+  case "csv"_t: return "text/csv";
+  case "htm"_t:
+  case "html"_t: return "text/html";
+  case "js"_t:
+  case "mjs"_t: return "text/javascript";
+  case "txt"_t: return "text/plain";
+  case "vtt"_t: return "text/vtt";
+
+  case "apng"_t: return "image/apng";
+  case "avif"_t: return "image/avif";
+  case "bmp"_t: return "image/bmp";
+  case "gif"_t: return "image/gif";
+  case "png"_t: return "image/png";
+  case "svg"_t: return "image/svg+xml";
+  case "webp"_t: return "image/webp";
+  case "ico"_t: return "image/x-icon";
+  case "tif"_t: return "image/tiff";
+  case "tiff"_t: return "image/tiff";
+  case "jpg"_t:
+  case "jpeg"_t: return "image/jpeg";
+
+  case "mp4"_t: return "video/mp4";
+  case "mpeg"_t: return "video/mpeg";
+  case "webm"_t: return "video/webm";
+
+  case "mp3"_t: return "audio/mp3";
+  case "mpga"_t: return "audio/mpeg";
+  case "weba"_t: return "audio/webm";
+  case "wav"_t: return "audio/wave";
+
+  case "otf"_t: return "font/otf";
+  case "ttf"_t: return "font/ttf";
+  case "woff"_t: return "font/woff";
+  case "woff2"_t: return "font/woff2";
+
+  case "7z"_t: return "application/x-7z-compressed";
+  case "atom"_t: return "application/atom+xml";
+  case "pdf"_t: return "application/pdf";
+  case "json"_t: return "application/json";
+  case "rss"_t: return "application/rss+xml";
+  case "tar"_t: return "application/x-tar";
+  case "xht"_t:
+  case "xhtml"_t: return "application/xhtml+xml";
+  case "xslt"_t: return "application/xslt+xml";
+  case "xml"_t: return "application/xml";
+  case "gz"_t: return "application/gzip";
+  case "zip"_t: return "application/zip";
+  case "wasm"_t: return "application/wasm";
+  }
+}
+
+inline bool can_compress_content_type(const std::string &content_type) {
+  using udl::operator""_t;
+
+  auto tag = str2tag(content_type);
+
+  switch (tag) {
+  case "image/svg+xml"_t:
+  case "application/javascript"_t:
+  case "application/json"_t:
+  case "application/xml"_t:
+  case "application/protobuf"_t:
+  case "application/xhtml+xml"_t: return true;
+
+  default:
+    return !content_type.rfind("text/", 0) && tag != "text/event-stream"_t;
+  }
+}
+
+inline EncodingType encoding_type(const Request &req, const Response &res) {
+  auto ret =
+      detail::can_compress_content_type(res.get_header_value("Content-Type"));
+  if (!ret) { return EncodingType::None; }
+
+  const auto &s = req.get_header_value("Accept-Encoding");
+  (void)(s);
+
+#ifdef CPPHTTPLIB_BROTLI_SUPPORT
+  // TODO: 'Accept-Encoding' has br, not br;q=0
+  ret = s.find("br") != std::string::npos;
+  if (ret) { return EncodingType::Brotli; }
+#endif
+
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+  // TODO: 'Accept-Encoding' has gzip, not gzip;q=0
+  ret = s.find("gzip") != std::string::npos;
+  if (ret) { return EncodingType::Gzip; }
+#endif
+
+  return EncodingType::None;
+}
+
+inline bool nocompressor::compress(const char *data, size_t data_length,
+                                   bool /*last*/, Callback callback) {
+  if (!data_length) { return true; }
+  return callback(data, data_length);
+}
+
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+inline gzip_compressor::gzip_compressor() {
+  std::memset(&strm_, 0, sizeof(strm_));
+  strm_.zalloc = Z_NULL;
+  strm_.zfree = Z_NULL;
+  strm_.opaque = Z_NULL;
+
+  is_valid_ = deflateInit2(&strm_, Z_DEFAULT_COMPRESSION, Z_DEFLATED, 31, 8,
+                           Z_DEFAULT_STRATEGY) == Z_OK;
+}
+
+inline gzip_compressor::~gzip_compressor() { deflateEnd(&strm_); }
+
+inline bool gzip_compressor::compress(const char *data, size_t data_length,
+                                      bool last, Callback callback) {
+  assert(is_valid_);
+
+  do {
+    constexpr size_t max_avail_in =
+        (std::numeric_limits<decltype(strm_.avail_in)>::max)();
+
+    strm_.avail_in = static_cast<decltype(strm_.avail_in)>(
+        (std::min)(data_length, max_avail_in));
+    strm_.next_in = const_cast<Bytef *>(reinterpret_cast<const Bytef *>(data));
+
+    data_length -= strm_.avail_in;
+    data += strm_.avail_in;
+
+    auto flush = (last && data_length == 0) ? Z_FINISH : Z_NO_FLUSH;
+    auto ret = Z_OK;
+
+    std::array<char, CPPHTTPLIB_COMPRESSION_BUFSIZ> buff{};
+    do {
+      strm_.avail_out = static_cast<uInt>(buff.size());
+      strm_.next_out = reinterpret_cast<Bytef *>(buff.data());
+
+      ret = deflate(&strm_, flush);
+      if (ret == Z_STREAM_ERROR) { return false; }
+
+      if (!callback(buff.data(), buff.size() - strm_.avail_out)) {
+        return false;
+      }
+    } while (strm_.avail_out == 0);
+
+    assert((flush == Z_FINISH && ret == Z_STREAM_END) ||
+           (flush == Z_NO_FLUSH && ret == Z_OK));
+    assert(strm_.avail_in == 0);
+  } while (data_length > 0);
+
+  return true;
+}
+
+inline gzip_decompressor::gzip_decompressor() {
+  std::memset(&strm_, 0, sizeof(strm_));
+  strm_.zalloc = Z_NULL;
+  strm_.zfree = Z_NULL;
+  strm_.opaque = Z_NULL;
+
+  // 15 is the value of wbits, which should be at the maximum possible value
+  // to ensure that any gzip stream can be decoded. The offset of 32 specifies
+  // that the stream type should be automatically detected either gzip or
+  // deflate.
+  is_valid_ = inflateInit2(&strm_, 32 + 15) == Z_OK;
+}
+
+inline gzip_decompressor::~gzip_decompressor() { inflateEnd(&strm_); }
+
+inline bool gzip_decompressor::is_valid() const { return is_valid_; }
+
+inline bool gzip_decompressor::decompress(const char *data, size_t data_length,
+                                          Callback callback) {
+  assert(is_valid_);
+
+  auto ret = Z_OK;
+
+  do {
+    constexpr size_t max_avail_in =
+        (std::numeric_limits<decltype(strm_.avail_in)>::max)();
+
+    strm_.avail_in = static_cast<decltype(strm_.avail_in)>(
+        (std::min)(data_length, max_avail_in));
+    strm_.next_in = const_cast<Bytef *>(reinterpret_cast<const Bytef *>(data));
+
+    data_length -= strm_.avail_in;
+    data += strm_.avail_in;
+
+    std::array<char, CPPHTTPLIB_COMPRESSION_BUFSIZ> buff{};
+    while (strm_.avail_in > 0 && ret == Z_OK) {
+      strm_.avail_out = static_cast<uInt>(buff.size());
+      strm_.next_out = reinterpret_cast<Bytef *>(buff.data());
+
+      ret = inflate(&strm_, Z_NO_FLUSH);
+
+      assert(ret != Z_STREAM_ERROR);
+      switch (ret) {
+      case Z_NEED_DICT:
+      case Z_DATA_ERROR:
+      case Z_MEM_ERROR: inflateEnd(&strm_); return false;
+      }
+
+      if (!callback(buff.data(), buff.size() - strm_.avail_out)) {
+        return false;
+      }
+    }
+
+    if (ret != Z_OK && ret != Z_STREAM_END) { return false; }
+
+  } while (data_length > 0);
+
+  return true;
+}
+#endif
+
+#ifdef CPPHTTPLIB_BROTLI_SUPPORT
+inline brotli_compressor::brotli_compressor() {
+  state_ = BrotliEncoderCreateInstance(nullptr, nullptr, nullptr);
+}
+
+inline brotli_compressor::~brotli_compressor() {
+  BrotliEncoderDestroyInstance(state_);
+}
+
+inline bool brotli_compressor::compress(const char *data, size_t data_length,
+                                        bool last, Callback callback) {
+  std::array<uint8_t, CPPHTTPLIB_COMPRESSION_BUFSIZ> buff{};
+
+  auto operation = last ? BROTLI_OPERATION_FINISH : BROTLI_OPERATION_PROCESS;
+  auto available_in = data_length;
+  auto next_in = reinterpret_cast<const uint8_t *>(data);
+
+  for (;;) {
+    if (last) {
+      if (BrotliEncoderIsFinished(state_)) { break; }
+    } else {
+      if (!available_in) { break; }
+    }
+
+    auto available_out = buff.size();
+    auto next_out = buff.data();
+
+    if (!BrotliEncoderCompressStream(state_, operation, &available_in, &next_in,
+                                     &available_out, &next_out, nullptr)) {
+      return false;
+    }
+
+    auto output_bytes = buff.size() - available_out;
+    if (output_bytes) {
+      callback(reinterpret_cast<const char *>(buff.data()), output_bytes);
+    }
+  }
+
+  return true;
+}
+
+inline brotli_decompressor::brotli_decompressor() {
+  decoder_s = BrotliDecoderCreateInstance(0, 0, 0);
+  decoder_r = decoder_s ? BROTLI_DECODER_RESULT_NEEDS_MORE_INPUT
+                        : BROTLI_DECODER_RESULT_ERROR;
+}
+
+inline brotli_decompressor::~brotli_decompressor() {
+  if (decoder_s) { BrotliDecoderDestroyInstance(decoder_s); }
+}
+
+inline bool brotli_decompressor::is_valid() const { return decoder_s; }
+
+inline bool brotli_decompressor::decompress(const char *data,
+                                            size_t data_length,
+                                            Callback callback) {
+  if (decoder_r == BROTLI_DECODER_RESULT_SUCCESS ||
+      decoder_r == BROTLI_DECODER_RESULT_ERROR) {
+    return 0;
+  }
+
+  auto next_in = reinterpret_cast<const uint8_t *>(data);
+  size_t avail_in = data_length;
+  size_t total_out;
+
+  decoder_r = BROTLI_DECODER_RESULT_NEEDS_MORE_OUTPUT;
+
+  std::array<char, CPPHTTPLIB_COMPRESSION_BUFSIZ> buff{};
+  while (decoder_r == BROTLI_DECODER_RESULT_NEEDS_MORE_OUTPUT) {
+    char *next_out = buff.data();
+    size_t avail_out = buff.size();
+
+    decoder_r = BrotliDecoderDecompressStream(
+        decoder_s, &avail_in, &next_in, &avail_out,
+        reinterpret_cast<uint8_t **>(&next_out), &total_out);
+
+    if (decoder_r == BROTLI_DECODER_RESULT_ERROR) { return false; }
+
+    if (!callback(buff.data(), buff.size() - avail_out)) { return false; }
+  }
+
+  return decoder_r == BROTLI_DECODER_RESULT_SUCCESS ||
+         decoder_r == BROTLI_DECODER_RESULT_NEEDS_MORE_INPUT;
+}
+#endif
+
+inline bool has_header(const Headers &headers, const std::string &key) {
+  return headers.find(key) != headers.end();
+}
+
+inline const char *get_header_value(const Headers &headers,
+                                    const std::string &key, size_t id,
+                                    const char *def) {
+  auto rng = headers.equal_range(key);
+  auto it = rng.first;
+  std::advance(it, static_cast<ssize_t>(id));
+  if (it != rng.second) { return it->second.c_str(); }
+  return def;
+}
+
+inline bool compare_case_ignore(const std::string &a, const std::string &b) {
+  if (a.size() != b.size()) { return false; }
+  for (size_t i = 0; i < b.size(); i++) {
+    if (::tolower(a[i]) != ::tolower(b[i])) { return false; }
+  }
+  return true;
+}
+
+template <typename T>
+inline bool parse_header(const char *beg, const char *end, T fn) {
+  // Skip trailing spaces and tabs.
+  while (beg < end && is_space_or_tab(end[-1])) {
+    end--;
+  }
+
+  auto p = beg;
+  while (p < end && *p != ':') {
+    p++;
+  }
+
+  if (p == end) { return false; }
+
+  auto key_end = p;
+
+  if (*p++ != ':') { return false; }
+
+  while (p < end && is_space_or_tab(*p)) {
+    p++;
+  }
+
+  if (p < end) {
+    auto key_len = key_end - beg;
+    if (!key_len) { return false; }
+
+    auto key = std::string(beg, key_end);
+    auto val = compare_case_ignore(key, "Location")
+                   ? std::string(p, end)
+                   : decode_url(std::string(p, end), false);
+    fn(key, val);
+    return true;
+  }
+
+  return false;
+}
+
+inline bool read_headers(Stream &strm, Headers &headers) {
+  const auto bufsiz = 2048;
+  char buf[bufsiz];
+  stream_line_reader line_reader(strm, buf, bufsiz);
+
+  for (;;) {
+    if (!line_reader.getline()) { return false; }
+
+    // Check if the line ends with CRLF.
+    auto line_terminator_len = 2;
+    if (line_reader.end_with_crlf()) {
+      // Blank line indicates end of headers.
+      if (line_reader.size() == 2) { break; }
+#ifdef CPPHTTPLIB_ALLOW_LF_AS_LINE_TERMINATOR
+    } else {
+      // Blank line indicates end of headers.
+      if (line_reader.size() == 1) { break; }
+      line_terminator_len = 1;
+    }
+#else
+    } else {
+      continue; // Skip invalid line.
+    }
+#endif
+
+    if (line_reader.size() > CPPHTTPLIB_HEADER_MAX_LENGTH) { return false; }
+
+    // Exclude line terminator
+    auto end = line_reader.ptr() + line_reader.size() - line_terminator_len;
+
+    parse_header(line_reader.ptr(), end,
+                 [&](const std::string &key, const std::string &val) {
+                   headers.emplace(key, val);
+                 });
+  }
+
+  return true;
+}
+
+inline bool read_content_with_length(Stream &strm, uint64_t len,
+                                     Progress progress,
+                                     ContentReceiverWithProgress out) {
+  char buf[CPPHTTPLIB_RECV_BUFSIZ];
+
+  uint64_t r = 0;
+  while (r < len) {
+    auto read_len = static_cast<size_t>(len - r);
+    auto n = strm.read(buf, (std::min)(read_len, CPPHTTPLIB_RECV_BUFSIZ));
+    if (n <= 0) { return false; }
+
+    if (!out(buf, static_cast<size_t>(n), r, len)) { return false; }
+    r += static_cast<uint64_t>(n);
+
+    if (progress) {
+      if (!progress(r, len)) { return false; }
+    }
+  }
+
+  return true;
+}
+
+inline void skip_content_with_length(Stream &strm, uint64_t len) {
+  char buf[CPPHTTPLIB_RECV_BUFSIZ];
+  uint64_t r = 0;
+  while (r < len) {
+    auto read_len = static_cast<size_t>(len - r);
+    auto n = strm.read(buf, (std::min)(read_len, CPPHTTPLIB_RECV_BUFSIZ));
+    if (n <= 0) { return; }
+    r += static_cast<uint64_t>(n);
+  }
+}
+
+inline bool read_content_without_length(Stream &strm,
+                                        ContentReceiverWithProgress out) {
+  char buf[CPPHTTPLIB_RECV_BUFSIZ];
+  uint64_t r = 0;
+  for (;;) {
+    auto n = strm.read(buf, CPPHTTPLIB_RECV_BUFSIZ);
+    if (n <= 0) { return true; }
+
+    if (!out(buf, static_cast<size_t>(n), r, 0)) { return false; }
+    r += static_cast<uint64_t>(n);
+  }
+
+  return true;
+}
+
+template <typename T>
+inline bool read_content_chunked(Stream &strm, T &x,
+                                 ContentReceiverWithProgress out) {
+  const auto bufsiz = 16;
+  char buf[bufsiz];
+
+  stream_line_reader line_reader(strm, buf, bufsiz);
+
+  if (!line_reader.getline()) { return false; }
+
+  unsigned long chunk_len;
+  while (true) {
+    char *end_ptr;
+
+    chunk_len = std::strtoul(line_reader.ptr(), &end_ptr, 16);
+
+    if (end_ptr == line_reader.ptr()) { return false; }
+    if (chunk_len == ULONG_MAX) { return false; }
+
+    if (chunk_len == 0) { break; }
+
+    if (!read_content_with_length(strm, chunk_len, nullptr, out)) {
+      return false;
+    }
+
+    if (!line_reader.getline()) { return false; }
+
+    if (strcmp(line_reader.ptr(), "\r\n") != 0) { return false; }
+
+    if (!line_reader.getline()) { return false; }
+  }
+
+  assert(chunk_len == 0);
+
+  // Trailer
+  if (!line_reader.getline()) { return false; }
+
+  while (strcmp(line_reader.ptr(), "\r\n") != 0) {
+    if (line_reader.size() > CPPHTTPLIB_HEADER_MAX_LENGTH) { return false; }
+
+    // Exclude line terminator
+    constexpr auto line_terminator_len = 2;
+    auto end = line_reader.ptr() + line_reader.size() - line_terminator_len;
+
+    parse_header(line_reader.ptr(), end,
+                 [&](const std::string &key, const std::string &val) {
+                   x.headers.emplace(key, val);
+                 });
+
+    if (!line_reader.getline()) { return false; }
+  }
+
+  return true;
+}
+
+inline bool is_chunked_transfer_encoding(const Headers &headers) {
+  return compare_case_ignore(
+      get_header_value(headers, "Transfer-Encoding", 0, ""), "chunked");
+}
+
+template <typename T, typename U>
+bool prepare_content_receiver(T &x, int &status,
+                              ContentReceiverWithProgress receiver,
+                              bool decompress, U callback) {
+  if (decompress) {
+    std::string encoding = x.get_header_value("Content-Encoding");
+    std::unique_ptr<decompressor> decompressor;
+
+    if (encoding == "gzip" || encoding == "deflate") {
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+      decompressor = detail::make_unique<gzip_decompressor>();
+#else
+      status = StatusCode::UnsupportedMediaType_415;
+      return false;
+#endif
+    } else if (encoding.find("br") != std::string::npos) {
+#ifdef CPPHTTPLIB_BROTLI_SUPPORT
+      decompressor = detail::make_unique<brotli_decompressor>();
+#else
+      status = StatusCode::UnsupportedMediaType_415;
+      return false;
+#endif
+    }
+
+    if (decompressor) {
+      if (decompressor->is_valid()) {
+        ContentReceiverWithProgress out = [&](const char *buf, size_t n,
+                                              uint64_t off, uint64_t len) {
+          return decompressor->decompress(buf, n,
+                                          [&](const char *buf2, size_t n2) {
+                                            return receiver(buf2, n2, off, len);
+                                          });
+        };
+        return callback(std::move(out));
+      } else {
+        status = StatusCode::InternalServerError_500;
+        return false;
+      }
+    }
+  }
+
+  ContentReceiverWithProgress out = [&](const char *buf, size_t n, uint64_t off,
+                                        uint64_t len) {
+    return receiver(buf, n, off, len);
+  };
+  return callback(std::move(out));
+}
+
+template <typename T>
+bool read_content(Stream &strm, T &x, size_t payload_max_length, int &status,
+                  Progress progress, ContentReceiverWithProgress receiver,
+                  bool decompress) {
+  return prepare_content_receiver(
+      x, status, std::move(receiver), decompress,
+      [&](const ContentReceiverWithProgress &out) {
+        auto ret = true;
+        auto exceed_payload_max_length = false;
+
+        if (is_chunked_transfer_encoding(x.headers)) {
+          ret = read_content_chunked(strm, x, out);
+        } else if (!has_header(x.headers, "Content-Length")) {
+          ret = read_content_without_length(strm, out);
+        } else {
+          auto len = get_header_value_u64(x.headers, "Content-Length", 0, 0);
+          if (len > payload_max_length) {
+            exceed_payload_max_length = true;
+            skip_content_with_length(strm, len);
+            ret = false;
+          } else if (len > 0) {
+            ret = read_content_with_length(strm, len, std::move(progress), out);
+          }
+        }
+
+        if (!ret) {
+          status = exceed_payload_max_length ? StatusCode::PayloadTooLarge_413
+                                             : StatusCode::BadRequest_400;
+        }
+        return ret;
+      });
+} // namespace detail
+
+inline ssize_t write_headers(Stream &strm, const Headers &headers) {
+  ssize_t write_len = 0;
+  for (const auto &x : headers) {
+    auto len =
+        strm.write_format("%s: %s\r\n", x.first.c_str(), x.second.c_str());
+    if (len < 0) { return len; }
+    write_len += len;
+  }
+  auto len = strm.write("\r\n");
+  if (len < 0) { return len; }
+  write_len += len;
+  return write_len;
+}
+
+inline bool write_data(Stream &strm, const char *d, size_t l) {
+  size_t offset = 0;
+  while (offset < l) {
+    auto length = strm.write(d + offset, l - offset);
+    if (length < 0) { return false; }
+    offset += static_cast<size_t>(length);
+  }
+  return true;
+}
+
+template <typename T>
+inline bool write_content(Stream &strm, const ContentProvider &content_provider,
+                          size_t offset, size_t length, T is_shutting_down,
+                          Error &error) {
+  size_t end_offset = offset + length;
+  auto ok = true;
+  DataSink data_sink;
+
+  data_sink.write = [&](const char *d, size_t l) -> bool {
+    if (ok) {
+      if (strm.is_writable() && write_data(strm, d, l)) {
+        offset += l;
+      } else {
+        ok = false;
+      }
+    }
+    return ok;
+  };
+
+  data_sink.is_writable = [&]() -> bool { return strm.is_writable(); };
+
+  while (offset < end_offset && !is_shutting_down()) {
+    if (!strm.is_writable()) {
+      error = Error::Write;
+      return false;
+    } else if (!content_provider(offset, end_offset - offset, data_sink)) {
+      error = Error::Canceled;
+      return false;
+    } else if (!ok) {
+      error = Error::Write;
+      return false;
+    }
+  }
+
+  error = Error::Success;
+  return true;
+}
+
+template <typename T>
+inline bool write_content(Stream &strm, const ContentProvider &content_provider,
+                          size_t offset, size_t length,
+                          const T &is_shutting_down) {
+  auto error = Error::Success;
+  return write_content(strm, content_provider, offset, length, is_shutting_down,
+                       error);
+}
+
+template <typename T>
+inline bool
+write_content_without_length(Stream &strm,
+                             const ContentProvider &content_provider,
+                             const T &is_shutting_down) {
+  size_t offset = 0;
+  auto data_available = true;
+  auto ok = true;
+  DataSink data_sink;
+
+  data_sink.write = [&](const char *d, size_t l) -> bool {
+    if (ok) {
+      offset += l;
+      if (!strm.is_writable() || !write_data(strm, d, l)) { ok = false; }
+    }
+    return ok;
+  };
+
+  data_sink.is_writable = [&]() -> bool { return strm.is_writable(); };
+
+  data_sink.done = [&](void) { data_available = false; };
+
+  while (data_available && !is_shutting_down()) {
+    if (!strm.is_writable()) {
+      return false;
+    } else if (!content_provider(offset, 0, data_sink)) {
+      return false;
+    } else if (!ok) {
+      return false;
+    }
+  }
+  return true;
+}
+
+template <typename T, typename U>
+inline bool
+write_content_chunked(Stream &strm, const ContentProvider &content_provider,
+                      const T &is_shutting_down, U &compressor, Error &error) {
+  size_t offset = 0;
+  auto data_available = true;
+  auto ok = true;
+  DataSink data_sink;
+
+  data_sink.write = [&](const char *d, size_t l) -> bool {
+    if (ok) {
+      data_available = l > 0;
+      offset += l;
+
+      std::string payload;
+      if (compressor.compress(d, l, false,
+                              [&](const char *data, size_t data_len) {
+                                payload.append(data, data_len);
+                                return true;
+                              })) {
+        if (!payload.empty()) {
+          // Emit chunked response header and footer for each chunk
+          auto chunk =
+              from_i_to_hex(payload.size()) + "\r\n" + payload + "\r\n";
+          if (!strm.is_writable() ||
+              !write_data(strm, chunk.data(), chunk.size())) {
+            ok = false;
+          }
+        }
+      } else {
+        ok = false;
+      }
+    }
+    return ok;
+  };
+
+  data_sink.is_writable = [&]() -> bool { return strm.is_writable(); };
+
+  auto done_with_trailer = [&](const Headers *trailer) {
+    if (!ok) { return; }
+
+    data_available = false;
+
+    std::string payload;
+    if (!compressor.compress(nullptr, 0, true,
+                             [&](const char *data, size_t data_len) {
+                               payload.append(data, data_len);
+                               return true;
+                             })) {
+      ok = false;
+      return;
+    }
+
+    if (!payload.empty()) {
+      // Emit chunked response header and footer for each chunk
+      auto chunk = from_i_to_hex(payload.size()) + "\r\n" + payload + "\r\n";
+      if (!strm.is_writable() ||
+          !write_data(strm, chunk.data(), chunk.size())) {
+        ok = false;
+        return;
+      }
+    }
+
+    static const std::string done_marker("0\r\n");
+    if (!write_data(strm, done_marker.data(), done_marker.size())) {
+      ok = false;
+    }
+
+    // Trailer
+    if (trailer) {
+      for (const auto &kv : *trailer) {
+        std::string field_line = kv.first + ": " + kv.second + "\r\n";
+        if (!write_data(strm, field_line.data(), field_line.size())) {
+          ok = false;
+        }
+      }
+    }
+
+    static const std::string crlf("\r\n");
+    if (!write_data(strm, crlf.data(), crlf.size())) { ok = false; }
+  };
+
+  data_sink.done = [&](void) { done_with_trailer(nullptr); };
+
+  data_sink.done_with_trailer = [&](const Headers &trailer) {
+    done_with_trailer(&trailer);
+  };
+
+  while (data_available && !is_shutting_down()) {
+    if (!strm.is_writable()) {
+      error = Error::Write;
+      return false;
+    } else if (!content_provider(offset, 0, data_sink)) {
+      error = Error::Canceled;
+      return false;
+    } else if (!ok) {
+      error = Error::Write;
+      return false;
+    }
+  }
+
+  error = Error::Success;
+  return true;
+}
+
+template <typename T, typename U>
+inline bool write_content_chunked(Stream &strm,
+                                  const ContentProvider &content_provider,
+                                  const T &is_shutting_down, U &compressor) {
+  auto error = Error::Success;
+  return write_content_chunked(strm, content_provider, is_shutting_down,
+                               compressor, error);
+}
+
+template <typename T>
+inline bool redirect(T &cli, Request &req, Response &res,
+                     const std::string &path, const std::string &location,
+                     Error &error) {
+  Request new_req = req;
+  new_req.path = path;
+  new_req.redirect_count_ -= 1;
+
+  if (res.status == StatusCode::SeeOther_303 &&
+      (req.method != "GET" && req.method != "HEAD")) {
+    new_req.method = "GET";
+    new_req.body.clear();
+    new_req.headers.clear();
+  }
+
+  Response new_res;
+
+  auto ret = cli.send(new_req, new_res, error);
+  if (ret) {
+    req = new_req;
+    res = new_res;
+
+    if (res.location.empty()) { res.location = location; }
+  }
+  return ret;
+}
+
+inline std::string params_to_query_str(const Params &params) {
+  std::string query;
+
+  for (auto it = params.begin(); it != params.end(); ++it) {
+    if (it != params.begin()) { query += "&"; }
+    query += it->first;
+    query += "=";
+    query += encode_query_param(it->second);
+  }
+  return query;
+}
+
+inline void parse_query_text(const std::string &s, Params &params) {
+  std::set<std::string> cache;
+  split(s.data(), s.data() + s.size(), '&', [&](const char *b, const char *e) {
+    std::string kv(b, e);
+    if (cache.find(kv) != cache.end()) { return; }
+    cache.insert(kv);
+
+    std::string key;
+    std::string val;
+    split(b, e, '=', [&](const char *b2, const char *e2) {
+      if (key.empty()) {
+        key.assign(b2, e2);
+      } else {
+        val.assign(b2, e2);
+      }
+    });
+
+    if (!key.empty()) {
+      params.emplace(decode_url(key, true), decode_url(val, true));
+    }
+  });
+}
+
+inline bool parse_multipart_boundary(const std::string &content_type,
+                                     std::string &boundary) {
+  auto boundary_keyword = "boundary=";
+  auto pos = content_type.find(boundary_keyword);
+  if (pos == std::string::npos) { return false; }
+  auto end = content_type.find(';', pos);
+  auto beg = pos + strlen(boundary_keyword);
+  boundary = trim_double_quotes_copy(content_type.substr(beg, end - beg));
+  return !boundary.empty();
+}
+
+inline void parse_disposition_params(const std::string &s, Params &params) {
+  std::set<std::string> cache;
+  split(s.data(), s.data() + s.size(), ';', [&](const char *b, const char *e) {
+    std::string kv(b, e);
+    if (cache.find(kv) != cache.end()) { return; }
+    cache.insert(kv);
+
+    std::string key;
+    std::string val;
+    split(b, e, '=', [&](const char *b2, const char *e2) {
+      if (key.empty()) {
+        key.assign(b2, e2);
+      } else {
+        val.assign(b2, e2);
+      }
+    });
+
+    if (!key.empty()) {
+      params.emplace(trim_double_quotes_copy((key)),
+                     trim_double_quotes_copy((val)));
+    }
+  });
+}
+
+#ifdef CPPHTTPLIB_NO_EXCEPTIONS
+inline bool parse_range_header(const std::string &s, Ranges &ranges) {
+#else
+inline bool parse_range_header(const std::string &s, Ranges &ranges) try {
+#endif
+  auto is_valid = [](const std::string &str) {
+    return std::all_of(str.cbegin(), str.cend(),
+                       [](unsigned char c) { return std::isdigit(c); });
+  };
+
+  if (s.size() > 7 && s.compare(0, 6, "bytes=") == 0) {
+    const auto pos = static_cast<size_t>(6);
+    const auto len = static_cast<size_t>(s.size() - 6);
+    auto all_valid_ranges = true;
+    split(&s[pos], &s[pos + len], ',', [&](const char *b, const char *e) {
+      if (!all_valid_ranges) { return; }
+
+      const auto it = std::find(b, e, '-');
+      if (it == e) {
+        all_valid_ranges = false;
+        return;
+      }
+
+      const auto lhs = std::string(b, it);
+      const auto rhs = std::string(it + 1, e);
+      if (!is_valid(lhs) || !is_valid(rhs)) {
+        all_valid_ranges = false;
+        return;
+      }
+
+      const auto first =
+          static_cast<ssize_t>(lhs.empty() ? -1 : std::stoll(lhs));
+      const auto last =
+          static_cast<ssize_t>(rhs.empty() ? -1 : std::stoll(rhs));
+      if ((first == -1 && last == -1) ||
+          (first != -1 && last != -1 && first > last)) {
+        all_valid_ranges = false;
+        return;
+      }
+
+      ranges.emplace_back(first, last);
+    });
+    return all_valid_ranges && !ranges.empty();
+  }
+  return false;
+#ifdef CPPHTTPLIB_NO_EXCEPTIONS
+}
+#else
+} catch (...) { return false; }
+#endif
+
+class MultipartFormDataParser {
+public:
+  MultipartFormDataParser() = default;
+
+  void set_boundary(std::string &&boundary) {
+    boundary_ = boundary;
+    dash_boundary_crlf_ = dash_ + boundary_ + crlf_;
+    crlf_dash_boundary_ = crlf_ + dash_ + boundary_;
+  }
+
+  bool is_valid() const { return is_valid_; }
+
+  bool parse(const char *buf, size_t n, const ContentReceiver &content_callback,
+             const MultipartContentHeader &header_callback) {
+
+    buf_append(buf, n);
+
+    while (buf_size() > 0) {
+      switch (state_) {
+      case 0: { // Initial boundary
+        buf_erase(buf_find(dash_boundary_crlf_));
+        if (dash_boundary_crlf_.size() > buf_size()) { return true; }
+        if (!buf_start_with(dash_boundary_crlf_)) { return false; }
+        buf_erase(dash_boundary_crlf_.size());
+        state_ = 1;
+        break;
+      }
+      case 1: { // New entry
+        clear_file_info();
+        state_ = 2;
+        break;
+      }
+      case 2: { // Headers
+        auto pos = buf_find(crlf_);
+        if (pos > CPPHTTPLIB_HEADER_MAX_LENGTH) { return false; }
+        while (pos < buf_size()) {
+          // Empty line
+          if (pos == 0) {
+            if (!header_callback(file_)) {
+              is_valid_ = false;
+              return false;
+            }
+            buf_erase(crlf_.size());
+            state_ = 3;
+            break;
+          }
+
+          const auto header = buf_head(pos);
+
+          if (!parse_header(header.data(), header.data() + header.size(),
+                            [&](const std::string &, const std::string &) {})) {
+            is_valid_ = false;
+            return false;
+          }
+
+          static const std::string header_content_type = "Content-Type:";
+
+          if (start_with_case_ignore(header, header_content_type)) {
+            file_.content_type =
+                trim_copy(header.substr(header_content_type.size()));
+          } else {
+            static const std::regex re_content_disposition(
+                R"~(^Content-Disposition:\s*form-data;\s*(.*)$)~",
+                std::regex_constants::icase);
+
+            std::smatch m;
+            if (std::regex_match(header, m, re_content_disposition)) {
+              Params params;
+              parse_disposition_params(m[1], params);
+
+              auto it = params.find("name");
+              if (it != params.end()) {
+                file_.name = it->second;
+              } else {
+                is_valid_ = false;
+                return false;
+              }
+
+              it = params.find("filename");
+              if (it != params.end()) { file_.filename = it->second; }
+
+              it = params.find("filename*");
+              if (it != params.end()) {
+                // Only allow UTF-8 enconnding...
+                static const std::regex re_rfc5987_encoding(
+                    R"~(^UTF-8''(.+?)$)~", std::regex_constants::icase);
+
+                std::smatch m2;
+                if (std::regex_match(it->second, m2, re_rfc5987_encoding)) {
+                  file_.filename = decode_url(m2[1], false); // override...
+                } else {
+                  is_valid_ = false;
+                  return false;
+                }
+              }
+            }
+          }
+          buf_erase(pos + crlf_.size());
+          pos = buf_find(crlf_);
+        }
+        if (state_ != 3) { return true; }
+        break;
+      }
+      case 3: { // Body
+        if (crlf_dash_boundary_.size() > buf_size()) { return true; }
+        auto pos = buf_find(crlf_dash_boundary_);
+        if (pos < buf_size()) {
+          if (!content_callback(buf_data(), pos)) {
+            is_valid_ = false;
+            return false;
+          }
+          buf_erase(pos + crlf_dash_boundary_.size());
+          state_ = 4;
+        } else {
+          auto len = buf_size() - crlf_dash_boundary_.size();
+          if (len > 0) {
+            if (!content_callback(buf_data(), len)) {
+              is_valid_ = false;
+              return false;
+            }
+            buf_erase(len);
+          }
+          return true;
+        }
+        break;
+      }
+      case 4: { // Boundary
+        if (crlf_.size() > buf_size()) { return true; }
+        if (buf_start_with(crlf_)) {
+          buf_erase(crlf_.size());
+          state_ = 1;
+        } else {
+          if (dash_.size() > buf_size()) { return true; }
+          if (buf_start_with(dash_)) {
+            buf_erase(dash_.size());
+            is_valid_ = true;
+            buf_erase(buf_size()); // Remove epilogue
+          } else {
+            return true;
+          }
+        }
+        break;
+      }
+      }
+    }
+
+    return true;
+  }
+
+private:
+  void clear_file_info() {
+    file_.name.clear();
+    file_.filename.clear();
+    file_.content_type.clear();
+  }
+
+  bool start_with_case_ignore(const std::string &a,
+                              const std::string &b) const {
+    if (a.size() < b.size()) { return false; }
+    for (size_t i = 0; i < b.size(); i++) {
+      if (::tolower(a[i]) != ::tolower(b[i])) { return false; }
+    }
+    return true;
+  }
+
+  const std::string dash_ = "--";
+  const std::string crlf_ = "\r\n";
+  std::string boundary_;
+  std::string dash_boundary_crlf_;
+  std::string crlf_dash_boundary_;
+
+  size_t state_ = 0;
+  bool is_valid_ = false;
+  MultipartFormData file_;
+
+  // Buffer
+  bool start_with(const std::string &a, size_t spos, size_t epos,
+                  const std::string &b) const {
+    if (epos - spos < b.size()) { return false; }
+    for (size_t i = 0; i < b.size(); i++) {
+      if (a[i + spos] != b[i]) { return false; }
+    }
+    return true;
+  }
+
+  size_t buf_size() const { return buf_epos_ - buf_spos_; }
+
+  const char *buf_data() const { return &buf_[buf_spos_]; }
+
+  std::string buf_head(size_t l) const { return buf_.substr(buf_spos_, l); }
+
+  bool buf_start_with(const std::string &s) const {
+    return start_with(buf_, buf_spos_, buf_epos_, s);
+  }
+
+  size_t buf_find(const std::string &s) const {
+    auto c = s.front();
+
+    size_t off = buf_spos_;
+    while (off < buf_epos_) {
+      auto pos = off;
+      while (true) {
+        if (pos == buf_epos_) { return buf_size(); }
+        if (buf_[pos] == c) { break; }
+        pos++;
+      }
+
+      auto remaining_size = buf_epos_ - pos;
+      if (s.size() > remaining_size) { return buf_size(); }
+
+      if (start_with(buf_, pos, buf_epos_, s)) { return pos - buf_spos_; }
+
+      off = pos + 1;
+    }
+
+    return buf_size();
+  }
+
+  void buf_append(const char *data, size_t n) {
+    auto remaining_size = buf_size();
+    if (remaining_size > 0 && buf_spos_ > 0) {
+      for (size_t i = 0; i < remaining_size; i++) {
+        buf_[i] = buf_[buf_spos_ + i];
+      }
+    }
+    buf_spos_ = 0;
+    buf_epos_ = remaining_size;
+
+    if (remaining_size + n > buf_.size()) { buf_.resize(remaining_size + n); }
+
+    for (size_t i = 0; i < n; i++) {
+      buf_[buf_epos_ + i] = data[i];
+    }
+    buf_epos_ += n;
+  }
+
+  void buf_erase(size_t size) { buf_spos_ += size; }
+
+  std::string buf_;
+  size_t buf_spos_ = 0;
+  size_t buf_epos_ = 0;
+};
+
+inline std::string to_lower(const char *beg, const char *end) {
+  std::string out;
+  auto it = beg;
+  while (it != end) {
+    out += static_cast<char>(::tolower(*it));
+    it++;
+  }
+  return out;
+}
+
+inline std::string random_string(size_t length) {
+  static const char data[] =
+      "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
+
+  // std::random_device might actually be deterministic on some
+  // platforms, but due to lack of support in the c++ standard library,
+  // doing better requires either some ugly hacks or breaking portability.
+  static std::random_device seed_gen;
+
+  // Request 128 bits of entropy for initialization
+  static std::seed_seq seed_sequence{seed_gen(), seed_gen(), seed_gen(),
+                                     seed_gen()};
+
+  static std::mt19937 engine(seed_sequence);
+
+  std::string result;
+  for (size_t i = 0; i < length; i++) {
+    result += data[engine() % (sizeof(data) - 1)];
+  }
+  return result;
+}
+
+inline std::string make_multipart_data_boundary() {
+  return "--cpp-httplib-multipart-data-" + detail::random_string(16);
+}
+
+inline bool is_multipart_boundary_chars_valid(const std::string &boundary) {
+  auto valid = true;
+  for (size_t i = 0; i < boundary.size(); i++) {
+    auto c = boundary[i];
+    if (!std::isalnum(c) && c != '-' && c != '_') {
+      valid = false;
+      break;
+    }
+  }
+  return valid;
+}
+
+template <typename T>
+inline std::string
+serialize_multipart_formdata_item_begin(const T &item,
+                                        const std::string &boundary) {
+  std::string body = "--" + boundary + "\r\n";
+  body += "Content-Disposition: form-data; name=\"" + item.name + "\"";
+  if (!item.filename.empty()) {
+    body += "; filename=\"" + item.filename + "\"";
+  }
+  body += "\r\n";
+  if (!item.content_type.empty()) {
+    body += "Content-Type: " + item.content_type + "\r\n";
+  }
+  body += "\r\n";
+
+  return body;
+}
+
+inline std::string serialize_multipart_formdata_item_end() { return "\r\n"; }
+
+inline std::string
+serialize_multipart_formdata_finish(const std::string &boundary) {
+  return "--" + boundary + "--\r\n";
+}
+
+inline std::string
+serialize_multipart_formdata_get_content_type(const std::string &boundary) {
+  return "multipart/form-data; boundary=" + boundary;
+}
+
+inline std::string
+serialize_multipart_formdata(const MultipartFormDataItems &items,
+                             const std::string &boundary, bool finish = true) {
+  std::string body;
+
+  for (const auto &item : items) {
+    body += serialize_multipart_formdata_item_begin(item, boundary);
+    body += item.content + serialize_multipart_formdata_item_end();
+  }
+
+  if (finish) { body += serialize_multipart_formdata_finish(boundary); }
+
+  return body;
+}
+
+inline bool range_error(Request &req, Response &res) {
+  if (!req.ranges.empty() && 200 <= res.status && res.status < 300) {
+    ssize_t contant_len = static_cast<ssize_t>(
+        res.content_length_ ? res.content_length_ : res.body.size());
+
+    ssize_t prev_first_pos = -1;
+    ssize_t prev_last_pos = -1;
+    size_t overwrapping_count = 0;
+
+    // NOTE: The following Range check is based on '14.2. Range' in RFC 9110
+    // 'HTTP Semantics' to avoid potential denial-of-service attacks.
+    // https://www.rfc-editor.org/rfc/rfc9110#section-14.2
+
+    // Too many ranges
+    if (req.ranges.size() > CPPHTTPLIB_RANGE_MAX_COUNT) { return true; }
+
+    for (auto &r : req.ranges) {
+      auto &first_pos = r.first;
+      auto &last_pos = r.second;
+
+      if (first_pos == -1 && last_pos == -1) {
+        first_pos = 0;
+        last_pos = contant_len;
+      }
+
+      if (first_pos == -1) {
+        first_pos = contant_len - last_pos;
+        last_pos = contant_len - 1;
+      }
+
+      if (last_pos == -1) { last_pos = contant_len - 1; }
+
+      // Range must be within content length
+      if (!(0 <= first_pos && first_pos <= last_pos &&
+            last_pos <= contant_len - 1)) {
+        return true;
+      }
+
+      // Ranges must be in ascending order
+      if (first_pos <= prev_first_pos) { return true; }
+
+      // Request must not have more than two overlapping ranges
+      if (first_pos <= prev_last_pos) {
+        overwrapping_count++;
+        if (overwrapping_count > 2) { return true; }
+      }
+
+      prev_first_pos = (std::max)(prev_first_pos, first_pos);
+      prev_last_pos = (std::max)(prev_last_pos, last_pos);
+    }
+  }
+
+  return false;
+}
+
+inline std::pair<size_t, size_t>
+get_range_offset_and_length(Range r, size_t content_length) {
+  assert(r.first != -1 && r.second != -1);
+  assert(0 <= r.first && r.first < static_cast<ssize_t>(content_length));
+  assert(r.first <= r.second &&
+         r.second < static_cast<ssize_t>(content_length));
+
+  return std::make_pair(r.first, static_cast<size_t>(r.second - r.first) + 1);
+}
+
+inline std::string make_content_range_header_field(
+    const std::pair<size_t, size_t> &offset_and_length, size_t content_length) {
+  auto st = offset_and_length.first;
+  auto ed = st + offset_and_length.second - 1;
+
+  std::string field = "bytes ";
+  field += std::to_string(st);
+  field += "-";
+  field += std::to_string(ed);
+  field += "/";
+  field += std::to_string(content_length);
+  return field;
+}
+
+template <typename SToken, typename CToken, typename Content>
+bool process_multipart_ranges_data(const Request &req,
+                                   const std::string &boundary,
+                                   const std::string &content_type,
+                                   size_t content_length, SToken stoken,
+                                   CToken ctoken, Content content) {
+  for (size_t i = 0; i < req.ranges.size(); i++) {
+    ctoken("--");
+    stoken(boundary);
+    ctoken("\r\n");
+    if (!content_type.empty()) {
+      ctoken("Content-Type: ");
+      stoken(content_type);
+      ctoken("\r\n");
+    }
+
+    auto offset_and_length =
+        get_range_offset_and_length(req.ranges[i], content_length);
+
+    ctoken("Content-Range: ");
+    stoken(make_content_range_header_field(offset_and_length, content_length));
+    ctoken("\r\n");
+    ctoken("\r\n");
+
+    if (!content(offset_and_length.first, offset_and_length.second)) {
+      return false;
+    }
+    ctoken("\r\n");
+  }
+
+  ctoken("--");
+  stoken(boundary);
+  ctoken("--");
+
+  return true;
+}
+
+inline void make_multipart_ranges_data(const Request &req, Response &res,
+                                       const std::string &boundary,
+                                       const std::string &content_type,
+                                       size_t content_length,
+                                       std::string &data) {
+  process_multipart_ranges_data(
+      req, boundary, content_type, content_length,
+      [&](const std::string &token) { data += token; },
+      [&](const std::string &token) { data += token; },
+      [&](size_t offset, size_t length) {
+        assert(offset + length <= content_length);
+        data += res.body.substr(offset, length);
+        return true;
+      });
+}
+
+inline size_t get_multipart_ranges_data_length(const Request &req,
+                                               const std::string &boundary,
+                                               const std::string &content_type,
+                                               size_t content_length) {
+  size_t data_length = 0;
+
+  process_multipart_ranges_data(
+      req, boundary, content_type, content_length,
+      [&](const std::string &token) { data_length += token.size(); },
+      [&](const std::string &token) { data_length += token.size(); },
+      [&](size_t /*offset*/, size_t length) {
+        data_length += length;
+        return true;
+      });
+
+  return data_length;
+}
+
+template <typename T>
+inline bool
+write_multipart_ranges_data(Stream &strm, const Request &req, Response &res,
+                            const std::string &boundary,
+                            const std::string &content_type,
+                            size_t content_length, const T &is_shutting_down) {
+  return process_multipart_ranges_data(
+      req, boundary, content_type, content_length,
+      [&](const std::string &token) { strm.write(token); },
+      [&](const std::string &token) { strm.write(token); },
+      [&](size_t offset, size_t length) {
+        return write_content(strm, res.content_provider_, offset, length,
+                             is_shutting_down);
+      });
+}
+
+inline bool expect_content(const Request &req) {
+  if (req.method == "POST" || req.method == "PUT" || req.method == "PATCH" ||
+      req.method == "PRI" || req.method == "DELETE") {
+    return true;
+  }
+  // TODO: check if Content-Length is set
+  return false;
+}
+
+inline bool has_crlf(const std::string &s) {
+  auto p = s.c_str();
+  while (*p) {
+    if (*p == '\r' || *p == '\n') { return true; }
+    p++;
+  }
+  return false;
+}
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline std::string message_digest(const std::string &s, const EVP_MD *algo) {
+  auto context = std::unique_ptr<EVP_MD_CTX, decltype(&EVP_MD_CTX_free)>(
+      EVP_MD_CTX_new(), EVP_MD_CTX_free);
+
+  unsigned int hash_length = 0;
+  unsigned char hash[EVP_MAX_MD_SIZE];
+
+  EVP_DigestInit_ex(context.get(), algo, nullptr);
+  EVP_DigestUpdate(context.get(), s.c_str(), s.size());
+  EVP_DigestFinal_ex(context.get(), hash, &hash_length);
+
+  std::stringstream ss;
+  for (auto i = 0u; i < hash_length; ++i) {
+    ss << std::hex << std::setw(2) << std::setfill('0')
+       << static_cast<unsigned int>(hash[i]);
+  }
+
+  return ss.str();
+}
+
+inline std::string MD5(const std::string &s) {
+  return message_digest(s, EVP_md5());
+}
+
+inline std::string SHA_256(const std::string &s) {
+  return message_digest(s, EVP_sha256());
+}
+
+inline std::string SHA_512(const std::string &s) {
+  return message_digest(s, EVP_sha512());
+}
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+#ifdef _WIN32
+// NOTE: This code came up with the following stackoverflow post:
+// https://stackoverflow.com/questions/9507184/can-openssl-on-windows-use-the-system-certificate-store
+inline bool load_system_certs_on_windows(X509_STORE *store) {
+  auto hStore = CertOpenSystemStoreW((HCRYPTPROV_LEGACY)NULL, L"ROOT");
+  if (!hStore) { return false; }
+
+  auto result = false;
+  PCCERT_CONTEXT pContext = NULL;
+  while ((pContext = CertEnumCertificatesInStore(hStore, pContext)) !=
+         nullptr) {
+    auto encoded_cert =
+        static_cast<const unsigned char *>(pContext->pbCertEncoded);
+
+    auto x509 = d2i_X509(NULL, &encoded_cert, pContext->cbCertEncoded);
+    if (x509) {
+      X509_STORE_add_cert(store, x509);
+      X509_free(x509);
+      result = true;
+    }
+  }
+
+  CertFreeCertificateContext(pContext);
+  CertCloseStore(hStore, 0);
+
+  return result;
+}
+#elif defined(CPPHTTPLIB_USE_CERTS_FROM_MACOSX_KEYCHAIN) && defined(__APPLE__)
+#if TARGET_OS_OSX
+template <typename T>
+using CFObjectPtr =
+    std::unique_ptr<typename std::remove_pointer<T>::type, void (*)(CFTypeRef)>;
+
+inline void cf_object_ptr_deleter(CFTypeRef obj) {
+  if (obj) { CFRelease(obj); }
+}
+
+inline bool retrieve_certs_from_keychain(CFObjectPtr<CFArrayRef> &certs) {
+  CFStringRef keys[] = {kSecClass, kSecMatchLimit, kSecReturnRef};
+  CFTypeRef values[] = {kSecClassCertificate, kSecMatchLimitAll,
+                        kCFBooleanTrue};
+
+  CFObjectPtr<CFDictionaryRef> query(
+      CFDictionaryCreate(nullptr, reinterpret_cast<const void **>(keys), values,
+                         sizeof(keys) / sizeof(keys[0]),
+                         &kCFTypeDictionaryKeyCallBacks,
+                         &kCFTypeDictionaryValueCallBacks),
+      cf_object_ptr_deleter);
+
+  if (!query) { return false; }
+
+  CFTypeRef security_items = nullptr;
+  if (SecItemCopyMatching(query.get(), &security_items) != errSecSuccess ||
+      CFArrayGetTypeID() != CFGetTypeID(security_items)) {
+    return false;
+  }
+
+  certs.reset(reinterpret_cast<CFArrayRef>(security_items));
+  return true;
+}
+
+inline bool retrieve_root_certs_from_keychain(CFObjectPtr<CFArrayRef> &certs) {
+  CFArrayRef root_security_items = nullptr;
+  if (SecTrustCopyAnchorCertificates(&root_security_items) != errSecSuccess) {
+    return false;
+  }
+
+  certs.reset(root_security_items);
+  return true;
+}
+
+inline bool add_certs_to_x509_store(CFArrayRef certs, X509_STORE *store) {
+  auto result = false;
+  for (auto i = 0; i < CFArrayGetCount(certs); ++i) {
+    const auto cert = reinterpret_cast<const __SecCertificate *>(
+        CFArrayGetValueAtIndex(certs, i));
+
+    if (SecCertificateGetTypeID() != CFGetTypeID(cert)) { continue; }
+
+    CFDataRef cert_data = nullptr;
+    if (SecItemExport(cert, kSecFormatX509Cert, 0, nullptr, &cert_data) !=
+        errSecSuccess) {
+      continue;
+    }
+
+    CFObjectPtr<CFDataRef> cert_data_ptr(cert_data, cf_object_ptr_deleter);
+
+    auto encoded_cert = static_cast<const unsigned char *>(
+        CFDataGetBytePtr(cert_data_ptr.get()));
+
+    auto x509 =
+        d2i_X509(NULL, &encoded_cert, CFDataGetLength(cert_data_ptr.get()));
+
+    if (x509) {
+      X509_STORE_add_cert(store, x509);
+      X509_free(x509);
+      result = true;
+    }
+  }
+
+  return result;
+}
+
+inline bool load_system_certs_on_macos(X509_STORE *store) {
+  auto result = false;
+  CFObjectPtr<CFArrayRef> certs(nullptr, cf_object_ptr_deleter);
+  if (retrieve_certs_from_keychain(certs) && certs) {
+    result = add_certs_to_x509_store(certs.get(), store);
+  }
+
+  if (retrieve_root_certs_from_keychain(certs) && certs) {
+    result = add_certs_to_x509_store(certs.get(), store) || result;
+  }
+
+  return result;
+}
+#endif // TARGET_OS_OSX
+#endif // _WIN32
+#endif // CPPHTTPLIB_OPENSSL_SUPPORT
+
+#ifdef _WIN32
+class WSInit {
+public:
+  WSInit() {
+    WSADATA wsaData;
+    if (WSAStartup(0x0002, &wsaData) == 0) is_valid_ = true;
+  }
+
+  ~WSInit() {
+    if (is_valid_) WSACleanup();
+  }
+
+  bool is_valid_ = false;
+};
+
+static WSInit wsinit_;
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline std::pair<std::string, std::string> make_digest_authentication_header(
+    const Request &req, const std::map<std::string, std::string> &auth,
+    size_t cnonce_count, const std::string &cnonce, const std::string &username,
+    const std::string &password, bool is_proxy = false) {
+  std::string nc;
+  {
+    std::stringstream ss;
+    ss << std::setfill('0') << std::setw(8) << std::hex << cnonce_count;
+    nc = ss.str();
+  }
+
+  std::string qop;
+  if (auth.find("qop") != auth.end()) {
+    qop = auth.at("qop");
+    if (qop.find("auth-int") != std::string::npos) {
+      qop = "auth-int";
+    } else if (qop.find("auth") != std::string::npos) {
+      qop = "auth";
+    } else {
+      qop.clear();
+    }
+  }
+
+  std::string algo = "MD5";
+  if (auth.find("algorithm") != auth.end()) { algo = auth.at("algorithm"); }
+
+  std::string response;
+  {
+    auto H = algo == "SHA-256"   ? detail::SHA_256
+             : algo == "SHA-512" ? detail::SHA_512
+                                 : detail::MD5;
+
+    auto A1 = username + ":" + auth.at("realm") + ":" + password;
+
+    auto A2 = req.method + ":" + req.path;
+    if (qop == "auth-int") { A2 += ":" + H(req.body); }
+
+    if (qop.empty()) {
+      response = H(H(A1) + ":" + auth.at("nonce") + ":" + H(A2));
+    } else {
+      response = H(H(A1) + ":" + auth.at("nonce") + ":" + nc + ":" + cnonce +
+                   ":" + qop + ":" + H(A2));
+    }
+  }
+
+  auto opaque = (auth.find("opaque") != auth.end()) ? auth.at("opaque") : "";
+
+  auto field = "Digest username=\"" + username + "\", realm=\"" +
+               auth.at("realm") + "\", nonce=\"" + auth.at("nonce") +
+               "\", uri=\"" + req.path + "\", algorithm=" + algo +
+               (qop.empty() ? ", response=\""
+                            : ", qop=" + qop + ", nc=" + nc + ", cnonce=\"" +
+                                  cnonce + "\", response=\"") +
+               response + "\"" +
+               (opaque.empty() ? "" : ", opaque=\"" + opaque + "\"");
+
+  auto key = is_proxy ? "Proxy-Authorization" : "Authorization";
+  return std::make_pair(key, field);
+}
+#endif
+
+inline bool parse_www_authenticate(const Response &res,
+                                   std::map<std::string, std::string> &auth,
+                                   bool is_proxy) {
+  auto auth_key = is_proxy ? "Proxy-Authenticate" : "WWW-Authenticate";
+  if (res.has_header(auth_key)) {
+    static auto re = std::regex(R"~((?:(?:,\s*)?(.+?)=(?:"(.*?)"|([^,]*))))~");
+    auto s = res.get_header_value(auth_key);
+    auto pos = s.find(' ');
+    if (pos != std::string::npos) {
+      auto type = s.substr(0, pos);
+      if (type == "Basic") {
+        return false;
+      } else if (type == "Digest") {
+        s = s.substr(pos + 1);
+        auto beg = std::sregex_iterator(s.begin(), s.end(), re);
+        for (auto i = beg; i != std::sregex_iterator(); ++i) {
+          const auto &m = *i;
+          auto key = s.substr(static_cast<size_t>(m.position(1)),
+                              static_cast<size_t>(m.length(1)));
+          auto val = m.length(2) > 0
+                         ? s.substr(static_cast<size_t>(m.position(2)),
+                                    static_cast<size_t>(m.length(2)))
+                         : s.substr(static_cast<size_t>(m.position(3)),
+                                    static_cast<size_t>(m.length(3)));
+          auth[key] = val;
+        }
+        return true;
+      }
+    }
+  }
+  return false;
+}
+
+class ContentProviderAdapter {
+public:
+  explicit ContentProviderAdapter(
+      ContentProviderWithoutLength &&content_provider)
+      : content_provider_(content_provider) {}
+
+  bool operator()(size_t offset, size_t, DataSink &sink) {
+    return content_provider_(offset, sink);
+  }
+
+private:
+  ContentProviderWithoutLength content_provider_;
+};
+
+} // namespace detail
+
+inline std::string hosted_at(const std::string &hostname) {
+  std::vector<std::string> addrs;
+  hosted_at(hostname, addrs);
+  if (addrs.empty()) { return std::string(); }
+  return addrs[0];
+}
+
+inline void hosted_at(const std::string &hostname,
+                      std::vector<std::string> &addrs) {
+  struct addrinfo hints;
+  struct addrinfo *result;
+
+  memset(&hints, 0, sizeof(struct addrinfo));
+  hints.ai_family = AF_UNSPEC;
+  hints.ai_socktype = SOCK_STREAM;
+  hints.ai_protocol = 0;
+
+  if (getaddrinfo(hostname.c_str(), nullptr, &hints, &result)) {
+#if defined __linux__ && !defined __ANDROID__
+    res_init();
+#endif
+    return;
+  }
+
+  for (auto rp = result; rp; rp = rp->ai_next) {
+    const auto &addr =
+        *reinterpret_cast<struct sockaddr_storage *>(rp->ai_addr);
+    std::string ip;
+    auto dummy = -1;
+    if (detail::get_ip_and_port(addr, sizeof(struct sockaddr_storage), ip,
+                                dummy)) {
+      addrs.push_back(ip);
+    }
+  }
+
+  freeaddrinfo(result);
+}
+
+inline std::string append_query_params(const std::string &path,
+                                       const Params &params) {
+  std::string path_with_query = path;
+  const static std::regex re("[^?]+\\?.*");
+  auto delm = std::regex_match(path, re) ? '&' : '?';
+  path_with_query += delm + detail::params_to_query_str(params);
+  return path_with_query;
+}
+
+// Header utilities
+inline std::pair<std::string, std::string>
+make_range_header(const Ranges &ranges) {
+  std::string field = "bytes=";
+  auto i = 0;
+  for (const auto &r : ranges) {
+    if (i != 0) { field += ", "; }
+    if (r.first != -1) { field += std::to_string(r.first); }
+    field += '-';
+    if (r.second != -1) { field += std::to_string(r.second); }
+    i++;
+  }
+  return std::make_pair("Range", std::move(field));
+}
+
+inline std::pair<std::string, std::string>
+make_basic_authentication_header(const std::string &username,
+                                 const std::string &password, bool is_proxy) {
+  auto field = "Basic " + detail::base64_encode(username + ":" + password);
+  auto key = is_proxy ? "Proxy-Authorization" : "Authorization";
+  return std::make_pair(key, std::move(field));
+}
+
+inline std::pair<std::string, std::string>
+make_bearer_token_authentication_header(const std::string &token,
+                                        bool is_proxy = false) {
+  auto field = "Bearer " + token;
+  auto key = is_proxy ? "Proxy-Authorization" : "Authorization";
+  return std::make_pair(key, std::move(field));
+}
+
+// Request implementation
+inline bool Request::has_header(const std::string &key) const {
+  return detail::has_header(headers, key);
+}
+
+inline std::string Request::get_header_value(const std::string &key,
+                                             size_t id) const {
+  return detail::get_header_value(headers, key, id, "");
+}
+
+inline size_t Request::get_header_value_count(const std::string &key) const {
+  auto r = headers.equal_range(key);
+  return static_cast<size_t>(std::distance(r.first, r.second));
+}
+
+inline void Request::set_header(const std::string &key,
+                                const std::string &val) {
+  if (!detail::has_crlf(key) && !detail::has_crlf(val)) {
+    headers.emplace(key, val);
+  }
+}
+
+inline bool Request::has_param(const std::string &key) const {
+  return params.find(key) != params.end();
+}
+
+inline std::string Request::get_param_value(const std::string &key,
+                                            size_t id) const {
+  auto rng = params.equal_range(key);
+  auto it = rng.first;
+  std::advance(it, static_cast<ssize_t>(id));
+  if (it != rng.second) { return it->second; }
+  return std::string();
+}
+
+inline size_t Request::get_param_value_count(const std::string &key) const {
+  auto r = params.equal_range(key);
+  return static_cast<size_t>(std::distance(r.first, r.second));
+}
+
+inline bool Request::is_multipart_form_data() const {
+  const auto &content_type = get_header_value("Content-Type");
+  return !content_type.rfind("multipart/form-data", 0);
+}
+
+inline bool Request::has_file(const std::string &key) const {
+  return files.find(key) != files.end();
+}
+
+inline MultipartFormData Request::get_file_value(const std::string &key) const {
+  auto it = files.find(key);
+  if (it != files.end()) { return it->second; }
+  return MultipartFormData();
+}
+
+inline std::vector<MultipartFormData>
+Request::get_file_values(const std::string &key) const {
+  std::vector<MultipartFormData> values;
+  auto rng = files.equal_range(key);
+  for (auto it = rng.first; it != rng.second; it++) {
+    values.push_back(it->second);
+  }
+  return values;
+}
+
+// Response implementation
+inline bool Response::has_header(const std::string &key) const {
+  return headers.find(key) != headers.end();
+}
+
+inline std::string Response::get_header_value(const std::string &key,
+                                              size_t id) const {
+  return detail::get_header_value(headers, key, id, "");
+}
+
+inline size_t Response::get_header_value_count(const std::string &key) const {
+  auto r = headers.equal_range(key);
+  return static_cast<size_t>(std::distance(r.first, r.second));
+}
+
+inline void Response::set_header(const std::string &key,
+                                 const std::string &val) {
+  if (!detail::has_crlf(key) && !detail::has_crlf(val)) {
+    headers.emplace(key, val);
+  }
+}
+
+inline void Response::set_redirect(const std::string &url, int stat) {
+  if (!detail::has_crlf(url)) {
+    set_header("Location", url);
+    if (300 <= stat && stat < 400) {
+      this->status = stat;
+    } else {
+      this->status = StatusCode::Found_302;
+    }
+  }
+}
+
+inline void Response::set_content(const char *s, size_t n,
+                                  const std::string &content_type) {
+  body.assign(s, n);
+
+  auto rng = headers.equal_range("Content-Type");
+  headers.erase(rng.first, rng.second);
+  set_header("Content-Type", content_type);
+}
+
+inline void Response::set_content(const std::string &s,
+                                  const std::string &content_type) {
+  set_content(s.data(), s.size(), content_type);
+}
+
+inline void Response::set_content(std::string &&s,
+                                  const std::string &content_type) {
+  body = std::move(s);
+
+  auto rng = headers.equal_range("Content-Type");
+  headers.erase(rng.first, rng.second);
+  set_header("Content-Type", content_type);
+}
+
+inline void Response::set_content_provider(
+    size_t in_length, const std::string &content_type, ContentProvider provider,
+    ContentProviderResourceReleaser resource_releaser) {
+  set_header("Content-Type", content_type);
+  content_length_ = in_length;
+  if (in_length > 0) { content_provider_ = std::move(provider); }
+  content_provider_resource_releaser_ = std::move(resource_releaser);
+  is_chunked_content_provider_ = false;
+}
+
+inline void Response::set_content_provider(
+    const std::string &content_type, ContentProviderWithoutLength provider,
+    ContentProviderResourceReleaser resource_releaser) {
+  set_header("Content-Type", content_type);
+  content_length_ = 0;
+  content_provider_ = detail::ContentProviderAdapter(std::move(provider));
+  content_provider_resource_releaser_ = std::move(resource_releaser);
+  is_chunked_content_provider_ = false;
+}
+
+inline void Response::set_chunked_content_provider(
+    const std::string &content_type, ContentProviderWithoutLength provider,
+    ContentProviderResourceReleaser resource_releaser) {
+  set_header("Content-Type", content_type);
+  content_length_ = 0;
+  content_provider_ = detail::ContentProviderAdapter(std::move(provider));
+  content_provider_resource_releaser_ = std::move(resource_releaser);
+  is_chunked_content_provider_ = true;
+}
+
+// Result implementation
+inline bool Result::has_request_header(const std::string &key) const {
+  return request_headers_.find(key) != request_headers_.end();
+}
+
+inline std::string Result::get_request_header_value(const std::string &key,
+                                                    size_t id) const {
+  return detail::get_header_value(request_headers_, key, id, "");
+}
+
+inline size_t
+Result::get_request_header_value_count(const std::string &key) const {
+  auto r = request_headers_.equal_range(key);
+  return static_cast<size_t>(std::distance(r.first, r.second));
+}
+
+// Stream implementation
+inline ssize_t Stream::write(const char *ptr) {
+  return write(ptr, strlen(ptr));
+}
+
+inline ssize_t Stream::write(const std::string &s) {
+  return write(s.data(), s.size());
+}
+
+namespace detail {
+
+// Socket stream implementation
+inline SocketStream::SocketStream(socket_t sock, time_t read_timeout_sec,
+                                  time_t read_timeout_usec,
+                                  time_t write_timeout_sec,
+                                  time_t write_timeout_usec)
+    : sock_(sock), read_timeout_sec_(read_timeout_sec),
+      read_timeout_usec_(read_timeout_usec),
+      write_timeout_sec_(write_timeout_sec),
+      write_timeout_usec_(write_timeout_usec), read_buff_(read_buff_size_, 0) {}
+
+inline SocketStream::~SocketStream() = default;
+
+inline bool SocketStream::is_readable() const {
+  return select_read(sock_, read_timeout_sec_, read_timeout_usec_) > 0;
+}
+
+inline bool SocketStream::is_writable() const {
+  return select_write(sock_, write_timeout_sec_, write_timeout_usec_) > 0 &&
+         is_socket_alive(sock_);
+}
+
+inline ssize_t SocketStream::read(char *ptr, size_t size) {
+#ifdef _WIN32
+  size =
+      (std::min)(size, static_cast<size_t>((std::numeric_limits<int>::max)()));
+#else
+  size = (std::min)(size,
+                    static_cast<size_t>((std::numeric_limits<ssize_t>::max)()));
+#endif
+
+  if (read_buff_off_ < read_buff_content_size_) {
+    auto remaining_size = read_buff_content_size_ - read_buff_off_;
+    if (size <= remaining_size) {
+      memcpy(ptr, read_buff_.data() + read_buff_off_, size);
+      read_buff_off_ += size;
+      return static_cast<ssize_t>(size);
+    } else {
+      memcpy(ptr, read_buff_.data() + read_buff_off_, remaining_size);
+      read_buff_off_ += remaining_size;
+      return static_cast<ssize_t>(remaining_size);
+    }
+  }
+
+  if (!is_readable()) { return -1; }
+
+  read_buff_off_ = 0;
+  read_buff_content_size_ = 0;
+
+  if (size < read_buff_size_) {
+    auto n = read_socket(sock_, read_buff_.data(), read_buff_size_,
+                         CPPHTTPLIB_RECV_FLAGS);
+    if (n <= 0) {
+      return n;
+    } else if (n <= static_cast<ssize_t>(size)) {
+      memcpy(ptr, read_buff_.data(), static_cast<size_t>(n));
+      return n;
+    } else {
+      memcpy(ptr, read_buff_.data(), size);
+      read_buff_off_ = size;
+      read_buff_content_size_ = static_cast<size_t>(n);
+      return static_cast<ssize_t>(size);
+    }
+  } else {
+    return read_socket(sock_, ptr, size, CPPHTTPLIB_RECV_FLAGS);
+  }
+}
+
+inline ssize_t SocketStream::write(const char *ptr, size_t size) {
+  if (!is_writable()) { return -1; }
+
+#if defined(_WIN32) && !defined(_WIN64)
+  size =
+      (std::min)(size, static_cast<size_t>((std::numeric_limits<int>::max)()));
+#endif
+
+  return send_socket(sock_, ptr, size, CPPHTTPLIB_SEND_FLAGS);
+}
+
+inline void SocketStream::get_remote_ip_and_port(std::string &ip,
+                                                 int &port) const {
+  return detail::get_remote_ip_and_port(sock_, ip, port);
+}
+
+inline void SocketStream::get_local_ip_and_port(std::string &ip,
+                                                int &port) const {
+  return detail::get_local_ip_and_port(sock_, ip, port);
+}
+
+inline socket_t SocketStream::socket() const { return sock_; }
+
+// Buffer stream implementation
+inline bool BufferStream::is_readable() const { return true; }
+
+inline bool BufferStream::is_writable() const { return true; }
+
+inline ssize_t BufferStream::read(char *ptr, size_t size) {
+#if defined(_MSC_VER) && _MSC_VER < 1910
+  auto len_read = buffer._Copy_s(ptr, size, size, position);
+#else
+  auto len_read = buffer.copy(ptr, size, position);
+#endif
+  position += static_cast<size_t>(len_read);
+  return static_cast<ssize_t>(len_read);
+}
+
+inline ssize_t BufferStream::write(const char *ptr, size_t size) {
+  buffer.append(ptr, size);
+  return static_cast<ssize_t>(size);
+}
+
+inline void BufferStream::get_remote_ip_and_port(std::string & /*ip*/,
+                                                 int & /*port*/) const {}
+
+inline void BufferStream::get_local_ip_and_port(std::string & /*ip*/,
+                                                int & /*port*/) const {}
+
+inline socket_t BufferStream::socket() const { return 0; }
+
+inline const std::string &BufferStream::get_buffer() const { return buffer; }
+
+inline PathParamsMatcher::PathParamsMatcher(const std::string &pattern) {
+  // One past the last ending position of a path param substring
+  std::size_t last_param_end = 0;
+
+#ifndef CPPHTTPLIB_NO_EXCEPTIONS
+  // Needed to ensure that parameter names are unique during matcher
+  // construction
+  // If exceptions are disabled, only last duplicate path
+  // parameter will be set
+  std::unordered_set<std::string> param_name_set;
+#endif
+
+  while (true) {
+    const auto marker_pos = pattern.find(marker, last_param_end);
+    if (marker_pos == std::string::npos) { break; }
+
+    static_fragments_.push_back(
+        pattern.substr(last_param_end, marker_pos - last_param_end));
+
+    const auto param_name_start = marker_pos + 1;
+
+    auto sep_pos = pattern.find(separator, param_name_start);
+    if (sep_pos == std::string::npos) { sep_pos = pattern.length(); }
+
+    auto param_name =
+        pattern.substr(param_name_start, sep_pos - param_name_start);
+
+#ifndef CPPHTTPLIB_NO_EXCEPTIONS
+    if (param_name_set.find(param_name) != param_name_set.cend()) {
+      std::string msg = "Encountered path parameter '" + param_name +
+                        "' multiple times in route pattern '" + pattern + "'.";
+      throw std::invalid_argument(msg);
+    }
+#endif
+
+    param_names_.push_back(std::move(param_name));
+
+    last_param_end = sep_pos + 1;
+  }
+
+  if (last_param_end < pattern.length()) {
+    static_fragments_.push_back(pattern.substr(last_param_end));
+  }
+}
+
+inline bool PathParamsMatcher::match(Request &request) const {
+  request.matches = std::smatch();
+  request.path_params.clear();
+  request.path_params.reserve(param_names_.size());
+
+  // One past the position at which the path matched the pattern last time
+  std::size_t starting_pos = 0;
+  for (size_t i = 0; i < static_fragments_.size(); ++i) {
+    const auto &fragment = static_fragments_[i];
+
+    if (starting_pos + fragment.length() > request.path.length()) {
+      return false;
+    }
+
+    // Avoid unnecessary allocation by using strncmp instead of substr +
+    // comparison
+    if (std::strncmp(request.path.c_str() + starting_pos, fragment.c_str(),
+                     fragment.length()) != 0) {
+      return false;
+    }
+
+    starting_pos += fragment.length();
+
+    // Should only happen when we have a static fragment after a param
+    // Example: '/users/:id/subscriptions'
+    // The 'subscriptions' fragment here does not have a corresponding param
+    if (i >= param_names_.size()) { continue; }
+
+    auto sep_pos = request.path.find(separator, starting_pos);
+    if (sep_pos == std::string::npos) { sep_pos = request.path.length(); }
+
+    const auto &param_name = param_names_[i];
+
+    request.path_params.emplace(
+        param_name, request.path.substr(starting_pos, sep_pos - starting_pos));
+
+    // Mark everythin up to '/' as matched
+    starting_pos = sep_pos + 1;
+  }
+  // Returns false if the path is longer than the pattern
+  return starting_pos >= request.path.length();
+}
+
+inline bool RegexMatcher::match(Request &request) const {
+  request.path_params.clear();
+  return std::regex_match(request.path, request.matches, regex_);
+}
+
+} // namespace detail
+
+// HTTP server implementation
+inline Server::Server()
+    : new_task_queue(
+          [] { return new ThreadPool(CPPHTTPLIB_THREAD_POOL_COUNT); }) {
+#ifndef _WIN32
+  signal(SIGPIPE, SIG_IGN);
+#endif
+}
+
+inline Server::~Server() = default;
+
+inline std::unique_ptr<detail::MatcherBase>
+Server::make_matcher(const std::string &pattern) {
+  if (pattern.find("/:") != std::string::npos) {
+    return detail::make_unique<detail::PathParamsMatcher>(pattern);
+  } else {
+    return detail::make_unique<detail::RegexMatcher>(pattern);
+  }
+}
+
+inline Server &Server::Get(const std::string &pattern, Handler handler) {
+  get_handlers_.emplace_back(make_matcher(pattern), std::move(handler));
+  return *this;
+}
+
+inline Server &Server::Post(const std::string &pattern, Handler handler) {
+  post_handlers_.emplace_back(make_matcher(pattern), std::move(handler));
+  return *this;
+}
+
+inline Server &Server::Post(const std::string &pattern,
+                            HandlerWithContentReader handler) {
+  post_handlers_for_content_reader_.emplace_back(make_matcher(pattern),
+                                                 std::move(handler));
+  return *this;
+}
+
+inline Server &Server::Put(const std::string &pattern, Handler handler) {
+  put_handlers_.emplace_back(make_matcher(pattern), std::move(handler));
+  return *this;
+}
+
+inline Server &Server::Put(const std::string &pattern,
+                           HandlerWithContentReader handler) {
+  put_handlers_for_content_reader_.emplace_back(make_matcher(pattern),
+                                                std::move(handler));
+  return *this;
+}
+
+inline Server &Server::Patch(const std::string &pattern, Handler handler) {
+  patch_handlers_.emplace_back(make_matcher(pattern), std::move(handler));
+  return *this;
+}
+
+inline Server &Server::Patch(const std::string &pattern,
+                             HandlerWithContentReader handler) {
+  patch_handlers_for_content_reader_.emplace_back(make_matcher(pattern),
+                                                  std::move(handler));
+  return *this;
+}
+
+inline Server &Server::Delete(const std::string &pattern, Handler handler) {
+  delete_handlers_.emplace_back(make_matcher(pattern), std::move(handler));
+  return *this;
+}
+
+inline Server &Server::Delete(const std::string &pattern,
+                              HandlerWithContentReader handler) {
+  delete_handlers_for_content_reader_.emplace_back(make_matcher(pattern),
+                                                   std::move(handler));
+  return *this;
+}
+
+inline Server &Server::Options(const std::string &pattern, Handler handler) {
+  options_handlers_.emplace_back(make_matcher(pattern), std::move(handler));
+  return *this;
+}
+
+inline bool Server::set_base_dir(const std::string &dir,
+                                 const std::string &mount_point) {
+  return set_mount_point(mount_point, dir);
+}
+
+inline bool Server::set_mount_point(const std::string &mount_point,
+                                    const std::string &dir, Headers headers) {
+  if (detail::is_dir(dir)) {
+    std::string mnt = !mount_point.empty() ? mount_point : "/";
+    if (!mnt.empty() && mnt[0] == '/') {
+      base_dirs_.push_back({mnt, dir, std::move(headers)});
+      return true;
+    }
+  }
+  return false;
+}
+
+inline bool Server::remove_mount_point(const std::string &mount_point) {
+  for (auto it = base_dirs_.begin(); it != base_dirs_.end(); ++it) {
+    if (it->mount_point == mount_point) {
+      base_dirs_.erase(it);
+      return true;
+    }
+  }
+  return false;
+}
+
+inline Server &
+Server::set_file_extension_and_mimetype_mapping(const std::string &ext,
+                                                const std::string &mime) {
+  file_extension_and_mimetype_map_[ext] = mime;
+  return *this;
+}
+
+inline Server &Server::set_default_file_mimetype(const std::string &mime) {
+  default_file_mimetype_ = mime;
+  return *this;
+}
+
+inline Server &Server::set_file_request_handler(Handler handler) {
+  file_request_handler_ = std::move(handler);
+  return *this;
+}
+
+inline Server &Server::set_error_handler_core(HandlerWithResponse handler,
+                                              std::true_type) {
+  error_handler_ = std::move(handler);
+  return *this;
+}
+
+inline Server &Server::set_error_handler_core(Handler handler,
+                                              std::false_type) {
+  error_handler_ = [handler](const Request &req, Response &res) {
+    handler(req, res);
+    return HandlerResponse::Handled;
+  };
+  return *this;
+}
+
+inline Server &Server::set_exception_handler(ExceptionHandler handler) {
+  exception_handler_ = std::move(handler);
+  return *this;
+}
+
+inline Server &Server::set_pre_routing_handler(HandlerWithResponse handler) {
+  pre_routing_handler_ = std::move(handler);
+  return *this;
+}
+
+inline Server &Server::set_post_routing_handler(Handler handler) {
+  post_routing_handler_ = std::move(handler);
+  return *this;
+}
+
+inline Server &Server::set_logger(Logger logger) {
+  logger_ = std::move(logger);
+  return *this;
+}
+
+inline Server &
+Server::set_expect_100_continue_handler(Expect100ContinueHandler handler) {
+  expect_100_continue_handler_ = std::move(handler);
+  return *this;
+}
+
+inline Server &Server::set_address_family(int family) {
+  address_family_ = family;
+  return *this;
+}
+
+inline Server &Server::set_tcp_nodelay(bool on) {
+  tcp_nodelay_ = on;
+  return *this;
+}
+
+inline Server &Server::set_socket_options(SocketOptions socket_options) {
+  socket_options_ = std::move(socket_options);
+  return *this;
+}
+
+inline Server &Server::set_default_headers(Headers headers) {
+  default_headers_ = std::move(headers);
+  return *this;
+}
+
+inline Server &Server::set_header_writer(
+    std::function<ssize_t(Stream &, Headers &)> const &writer) {
+  header_writer_ = writer;
+  return *this;
+}
+
+inline Server &Server::set_keep_alive_max_count(size_t count) {
+  keep_alive_max_count_ = count;
+  return *this;
+}
+
+inline Server &Server::set_keep_alive_timeout(time_t sec) {
+  keep_alive_timeout_sec_ = sec;
+  return *this;
+}
+
+inline Server &Server::set_read_timeout(time_t sec, time_t usec) {
+  read_timeout_sec_ = sec;
+  read_timeout_usec_ = usec;
+  return *this;
+}
+
+inline Server &Server::set_write_timeout(time_t sec, time_t usec) {
+  write_timeout_sec_ = sec;
+  write_timeout_usec_ = usec;
+  return *this;
+}
+
+inline Server &Server::set_idle_interval(time_t sec, time_t usec) {
+  idle_interval_sec_ = sec;
+  idle_interval_usec_ = usec;
+  return *this;
+}
+
+inline Server &Server::set_payload_max_length(size_t length) {
+  payload_max_length_ = length;
+  return *this;
+}
+
+inline bool Server::bind_to_port(const std::string &host, int port,
+                                 int socket_flags) {
+  return bind_internal(host, port, socket_flags) >= 0;
+}
+inline int Server::bind_to_any_port(const std::string &host, int socket_flags) {
+  return bind_internal(host, 0, socket_flags);
+}
+
+inline bool Server::listen_after_bind() {
+  auto se = detail::scope_exit([&]() { done_ = true; });
+  return listen_internal();
+}
+
+inline bool Server::listen(const std::string &host, int port,
+                           int socket_flags) {
+  auto se = detail::scope_exit([&]() { done_ = true; });
+  return bind_to_port(host, port, socket_flags) && listen_internal();
+}
+
+inline bool Server::is_running() const { return is_running_; }
+
+inline void Server::wait_until_ready() const {
+  while (!is_running() && !done_) {
+    std::this_thread::sleep_for(std::chrono::milliseconds{1});
+  }
+}
+
+inline void Server::stop() {
+  if (is_running_) {
+    assert(svr_sock_ != INVALID_SOCKET);
+    std::atomic<socket_t> sock(svr_sock_.exchange(INVALID_SOCKET));
+    detail::shutdown_socket(sock);
+    detail::close_socket(sock);
+  }
+}
+
+inline bool Server::parse_request_line(const char *s, Request &req) const {
+  auto len = strlen(s);
+  if (len < 2 || s[len - 2] != '\r' || s[len - 1] != '\n') { return false; }
+  len -= 2;
+
+  {
+    size_t count = 0;
+
+    detail::split(s, s + len, ' ', [&](const char *b, const char *e) {
+      switch (count) {
+      case 0: req.method = std::string(b, e); break;
+      case 1: req.target = std::string(b, e); break;
+      case 2: req.version = std::string(b, e); break;
+      default: break;
+      }
+      count++;
+    });
+
+    if (count != 3) { return false; }
+  }
+
+  static const std::set<std::string> methods{
+      "GET",     "HEAD",    "POST",  "PUT",   "DELETE",
+      "CONNECT", "OPTIONS", "TRACE", "PATCH", "PRI"};
+
+  if (methods.find(req.method) == methods.end()) { return false; }
+
+  if (req.version != "HTTP/1.1" && req.version != "HTTP/1.0") { return false; }
+
+  {
+    // Skip URL fragment
+    for (size_t i = 0; i < req.target.size(); i++) {
+      if (req.target[i] == '#') {
+        req.target.erase(i);
+        break;
+      }
+    }
+
+    size_t count = 0;
+
+    detail::split(req.target.data(), req.target.data() + req.target.size(), '?',
+                  2, [&](const char *b, const char *e) {
+                    switch (count) {
+                    case 0:
+                      req.path = detail::decode_url(std::string(b, e), false);
+                      break;
+                    case 1: {
+                      if (e - b > 0) {
+                        detail::parse_query_text(std::string(b, e), req.params);
+                      }
+                      break;
+                    }
+                    default: break;
+                    }
+                    count++;
+                  });
+
+    if (count > 2) { return false; }
+  }
+
+  return true;
+}
+
+inline bool Server::write_response(Stream &strm, bool close_connection,
+                                   Request &req, Response &res) {
+  // NOTE: `req.ranges` should be empty, otherwise it will be applied
+  // incorrectly to the error content.
+  req.ranges.clear();
+  return write_response_core(strm, close_connection, req, res, false);
+}
+
+inline bool Server::write_response_with_content(Stream &strm,
+                                                bool close_connection,
+                                                const Request &req,
+                                                Response &res) {
+  return write_response_core(strm, close_connection, req, res, true);
+}
+
+inline bool Server::write_response_core(Stream &strm, bool close_connection,
+                                        const Request &req, Response &res,
+                                        bool need_apply_ranges) {
+  assert(res.status != -1);
+
+  if (400 <= res.status && error_handler_ &&
+      error_handler_(req, res) == HandlerResponse::Handled) {
+    need_apply_ranges = true;
+  }
+
+  std::string content_type;
+  std::string boundary;
+  if (need_apply_ranges) { apply_ranges(req, res, content_type, boundary); }
+
+  // Prepare additional headers
+  if (close_connection || req.get_header_value("Connection") == "close") {
+    res.set_header("Connection", "close");
+  } else {
+    std::stringstream ss;
+    ss << "timeout=" << keep_alive_timeout_sec_
+       << ", max=" << keep_alive_max_count_;
+    res.set_header("Keep-Alive", ss.str());
+  }
+
+  if (!res.has_header("Content-Type") &&
+      (!res.body.empty() || res.content_length_ > 0 || res.content_provider_)) {
+    res.set_header("Content-Type", "text/plain");
+  }
+
+  if (!res.has_header("Content-Length") && res.body.empty() &&
+      !res.content_length_ && !res.content_provider_) {
+    res.set_header("Content-Length", "0");
+  }
+
+  if (!res.has_header("Accept-Ranges") && req.method == "HEAD") {
+    res.set_header("Accept-Ranges", "bytes");
+  }
+
+  if (post_routing_handler_) { post_routing_handler_(req, res); }
+
+  // Response line and headers
+  {
+    detail::BufferStream bstrm;
+
+    if (!bstrm.write_format("HTTP/1.1 %d %s\r\n", res.status,
+                            status_message(res.status))) {
+      return false;
+    }
+
+    if (!header_writer_(bstrm, res.headers)) { return false; }
+
+    // Flush buffer
+    auto &data = bstrm.get_buffer();
+    detail::write_data(strm, data.data(), data.size());
+  }
+
+  // Body
+  auto ret = true;
+  if (req.method != "HEAD") {
+    if (!res.body.empty()) {
+      if (!detail::write_data(strm, res.body.data(), res.body.size())) {
+        ret = false;
+      }
+    } else if (res.content_provider_) {
+      if (write_content_with_provider(strm, req, res, boundary, content_type)) {
+        res.content_provider_success_ = true;
+      } else {
+        ret = false;
+      }
+    }
+  }
+
+  // Log
+  if (logger_) { logger_(req, res); }
+
+  return ret;
+}
+
+inline bool
+Server::write_content_with_provider(Stream &strm, const Request &req,
+                                    Response &res, const std::string &boundary,
+                                    const std::string &content_type) {
+  auto is_shutting_down = [this]() {
+    return this->svr_sock_ == INVALID_SOCKET;
+  };
+
+  if (res.content_length_ > 0) {
+    if (req.ranges.empty()) {
+      return detail::write_content(strm, res.content_provider_, 0,
+                                   res.content_length_, is_shutting_down);
+    } else if (req.ranges.size() == 1) {
+      auto offset_and_length = detail::get_range_offset_and_length(
+          req.ranges[0], res.content_length_);
+
+      return detail::write_content(strm, res.content_provider_,
+                                   offset_and_length.first,
+                                   offset_and_length.second, is_shutting_down);
+    } else {
+      return detail::write_multipart_ranges_data(
+          strm, req, res, boundary, content_type, res.content_length_,
+          is_shutting_down);
+    }
+  } else {
+    if (res.is_chunked_content_provider_) {
+      auto type = detail::encoding_type(req, res);
+
+      std::unique_ptr<detail::compressor> compressor;
+      if (type == detail::EncodingType::Gzip) {
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+        compressor = detail::make_unique<detail::gzip_compressor>();
+#endif
+      } else if (type == detail::EncodingType::Brotli) {
+#ifdef CPPHTTPLIB_BROTLI_SUPPORT
+        compressor = detail::make_unique<detail::brotli_compressor>();
+#endif
+      } else {
+        compressor = detail::make_unique<detail::nocompressor>();
+      }
+      assert(compressor != nullptr);
+
+      return detail::write_content_chunked(strm, res.content_provider_,
+                                           is_shutting_down, *compressor);
+    } else {
+      return detail::write_content_without_length(strm, res.content_provider_,
+                                                  is_shutting_down);
+    }
+  }
+}
+
+inline bool Server::read_content(Stream &strm, Request &req, Response &res) {
+  MultipartFormDataMap::iterator cur;
+  auto file_count = 0;
+  if (read_content_core(
+          strm, req, res,
+          // Regular
+          [&](const char *buf, size_t n) {
+            if (req.body.size() + n > req.body.max_size()) { return false; }
+            req.body.append(buf, n);
+            return true;
+          },
+          // Multipart
+          [&](const MultipartFormData &file) {
+            if (file_count++ == CPPHTTPLIB_MULTIPART_FORM_DATA_FILE_MAX_COUNT) {
+              return false;
+            }
+            cur = req.files.emplace(file.name, file);
+            return true;
+          },
+          [&](const char *buf, size_t n) {
+            auto &content = cur->second.content;
+            if (content.size() + n > content.max_size()) { return false; }
+            content.append(buf, n);
+            return true;
+          })) {
+    const auto &content_type = req.get_header_value("Content-Type");
+    if (!content_type.find("application/x-www-form-urlencoded")) {
+      if (req.body.size() > CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH) {
+        res.status = StatusCode::PayloadTooLarge_413; // NOTE: should be 414?
+        return false;
+      }
+      detail::parse_query_text(req.body, req.params);
+    }
+    return true;
+  }
+  return false;
+}
+
+inline bool Server::read_content_with_content_receiver(
+    Stream &strm, Request &req, Response &res, ContentReceiver receiver,
+    MultipartContentHeader multipart_header,
+    ContentReceiver multipart_receiver) {
+  return read_content_core(strm, req, res, std::move(receiver),
+                           std::move(multipart_header),
+                           std::move(multipart_receiver));
+}
+
+inline bool
+Server::read_content_core(Stream &strm, Request &req, Response &res,
+                          ContentReceiver receiver,
+                          MultipartContentHeader multipart_header,
+                          ContentReceiver multipart_receiver) const {
+  detail::MultipartFormDataParser multipart_form_data_parser;
+  ContentReceiverWithProgress out;
+
+  if (req.is_multipart_form_data()) {
+    const auto &content_type = req.get_header_value("Content-Type");
+    std::string boundary;
+    if (!detail::parse_multipart_boundary(content_type, boundary)) {
+      res.status = StatusCode::BadRequest_400;
+      return false;
+    }
+
+    multipart_form_data_parser.set_boundary(std::move(boundary));
+    out = [&](const char *buf, size_t n, uint64_t /*off*/, uint64_t /*len*/) {
+      /* For debug
+      size_t pos = 0;
+      while (pos < n) {
+        auto read_size = (std::min)<size_t>(1, n - pos);
+        auto ret = multipart_form_data_parser.parse(
+            buf + pos, read_size, multipart_receiver, multipart_header);
+        if (!ret) { return false; }
+        pos += read_size;
+      }
+      return true;
+      */
+      return multipart_form_data_parser.parse(buf, n, multipart_receiver,
+                                              multipart_header);
+    };
+  } else {
+    out = [receiver](const char *buf, size_t n, uint64_t /*off*/,
+                     uint64_t /*len*/) { return receiver(buf, n); };
+  }
+
+  if (req.method == "DELETE" && !req.has_header("Content-Length")) {
+    return true;
+  }
+
+  if (!detail::read_content(strm, req, payload_max_length_, res.status, nullptr,
+                            out, true)) {
+    return false;
+  }
+
+  if (req.is_multipart_form_data()) {
+    if (!multipart_form_data_parser.is_valid()) {
+      res.status = StatusCode::BadRequest_400;
+      return false;
+    }
+  }
+
+  return true;
+}
+
+inline bool Server::handle_file_request(const Request &req, Response &res,
+                                        bool head) {
+  for (const auto &entry : base_dirs_) {
+    // Prefix match
+    if (!req.path.compare(0, entry.mount_point.size(), entry.mount_point)) {
+      std::string sub_path = "/" + req.path.substr(entry.mount_point.size());
+      if (detail::is_valid_path(sub_path)) {
+        auto path = entry.base_dir + sub_path;
+        if (path.back() == '/') { path += "index.html"; }
+
+        if (detail::is_file(path)) {
+          for (const auto &kv : entry.headers) {
+            res.set_header(kv.first, kv.second);
+          }
+
+          auto mm = std::make_shared<detail::mmap>(path.c_str());
+          if (!mm->is_open()) { return false; }
+
+          res.set_content_provider(
+              mm->size(),
+              detail::find_content_type(path, file_extension_and_mimetype_map_,
+                                        default_file_mimetype_),
+              [mm](size_t offset, size_t length, DataSink &sink) -> bool {
+                sink.write(mm->data() + offset, length);
+                return true;
+              });
+
+          if (!head && file_request_handler_) {
+            file_request_handler_(req, res);
+          }
+
+          return true;
+        }
+      }
+    }
+  }
+  return false;
+}
+
+inline socket_t
+Server::create_server_socket(const std::string &host, int port,
+                             int socket_flags,
+                             SocketOptions socket_options) const {
+  return detail::create_socket(
+      host, std::string(), port, address_family_, socket_flags, tcp_nodelay_,
+      std::move(socket_options),
+      [](socket_t sock, struct addrinfo &ai) -> bool {
+        if (::bind(sock, ai.ai_addr, static_cast<socklen_t>(ai.ai_addrlen))) {
+          return false;
+        }
+        if (::listen(sock, CPPHTTPLIB_LISTEN_BACKLOG)) { return false; }
+        return true;
+      });
+}
+
+inline int Server::bind_internal(const std::string &host, int port,
+                                 int socket_flags) {
+  if (!is_valid()) { return -1; }
+
+  svr_sock_ = create_server_socket(host, port, socket_flags, socket_options_);
+  if (svr_sock_ == INVALID_SOCKET) { return -1; }
+
+  if (port == 0) {
+    struct sockaddr_storage addr;
+    socklen_t addr_len = sizeof(addr);
+    if (getsockname(svr_sock_, reinterpret_cast<struct sockaddr *>(&addr),
+                    &addr_len) == -1) {
+      return -1;
+    }
+    if (addr.ss_family == AF_INET) {
+      return ntohs(reinterpret_cast<struct sockaddr_in *>(&addr)->sin_port);
+    } else if (addr.ss_family == AF_INET6) {
+      return ntohs(reinterpret_cast<struct sockaddr_in6 *>(&addr)->sin6_port);
+    } else {
+      return -1;
+    }
+  } else {
+    return port;
+  }
+}
+
+inline bool Server::listen_internal() {
+  auto ret = true;
+  is_running_ = true;
+  auto se = detail::scope_exit([&]() { is_running_ = false; });
+
+  {
+    std::unique_ptr<TaskQueue> task_queue(new_task_queue());
+
+    while (svr_sock_ != INVALID_SOCKET) {
+#ifndef _WIN32
+      if (idle_interval_sec_ > 0 || idle_interval_usec_ > 0) {
+#endif
+        auto val = detail::select_read(svr_sock_, idle_interval_sec_,
+                                       idle_interval_usec_);
+        if (val == 0) { // Timeout
+          task_queue->on_idle();
+          continue;
+        }
+#ifndef _WIN32
+      }
+#endif
+      socket_t sock = accept(svr_sock_, nullptr, nullptr);
+
+      if (sock == INVALID_SOCKET) {
+        if (errno == EMFILE) {
+          // The per-process limit of open file descriptors has been reached.
+          // Try to accept new connections after a short sleep.
+          std::this_thread::sleep_for(std::chrono::milliseconds(1));
+          continue;
+        } else if (errno == EINTR || errno == EAGAIN) {
+          continue;
+        }
+        if (svr_sock_ != INVALID_SOCKET) {
+          detail::close_socket(svr_sock_);
+          ret = false;
+        } else {
+          ; // The server socket was closed by user.
+        }
+        break;
+      }
+
+      {
+#ifdef _WIN32
+        auto timeout = static_cast<uint32_t>(read_timeout_sec_ * 1000 +
+                                             read_timeout_usec_ / 1000);
+        setsockopt(sock, SOL_SOCKET, SO_RCVTIMEO,
+                   reinterpret_cast<const char *>(&timeout), sizeof(timeout));
+#else
+        timeval tv;
+        tv.tv_sec = static_cast<long>(read_timeout_sec_);
+        tv.tv_usec = static_cast<decltype(tv.tv_usec)>(read_timeout_usec_);
+        setsockopt(sock, SOL_SOCKET, SO_RCVTIMEO,
+                   reinterpret_cast<const void *>(&tv), sizeof(tv));
+#endif
+      }
+      {
+
+#ifdef _WIN32
+        auto timeout = static_cast<uint32_t>(write_timeout_sec_ * 1000 +
+                                             write_timeout_usec_ / 1000);
+        setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO,
+                   reinterpret_cast<const char *>(&timeout), sizeof(timeout));
+#else
+        timeval tv;
+        tv.tv_sec = static_cast<long>(write_timeout_sec_);
+        tv.tv_usec = static_cast<decltype(tv.tv_usec)>(write_timeout_usec_);
+        setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO,
+                   reinterpret_cast<const void *>(&tv), sizeof(tv));
+#endif
+      }
+
+      if (!task_queue->enqueue(
+              [this, sock]() { process_and_close_socket(sock); })) {
+        detail::shutdown_socket(sock);
+        detail::close_socket(sock);
+      }
+    }
+
+    task_queue->shutdown();
+  }
+
+  return ret;
+}
+
+inline bool Server::routing(Request &req, Response &res, Stream &strm) {
+  if (pre_routing_handler_ &&
+      pre_routing_handler_(req, res) == HandlerResponse::Handled) {
+    return true;
+  }
+
+  // File handler
+  auto is_head_request = req.method == "HEAD";
+  if ((req.method == "GET" || is_head_request) &&
+      handle_file_request(req, res, is_head_request)) {
+    return true;
+  }
+
+  if (detail::expect_content(req)) {
+    // Content reader handler
+    {
+      ContentReader reader(
+          [&](ContentReceiver receiver) {
+            return read_content_with_content_receiver(
+                strm, req, res, std::move(receiver), nullptr, nullptr);
+          },
+          [&](MultipartContentHeader header, ContentReceiver receiver) {
+            return read_content_with_content_receiver(strm, req, res, nullptr,
+                                                      std::move(header),
+                                                      std::move(receiver));
+          });
+
+      if (req.method == "POST") {
+        if (dispatch_request_for_content_reader(
+                req, res, std::move(reader),
+                post_handlers_for_content_reader_)) {
+          return true;
+        }
+      } else if (req.method == "PUT") {
+        if (dispatch_request_for_content_reader(
+                req, res, std::move(reader),
+                put_handlers_for_content_reader_)) {
+          return true;
+        }
+      } else if (req.method == "PATCH") {
+        if (dispatch_request_for_content_reader(
+                req, res, std::move(reader),
+                patch_handlers_for_content_reader_)) {
+          return true;
+        }
+      } else if (req.method == "DELETE") {
+        if (dispatch_request_for_content_reader(
+                req, res, std::move(reader),
+                delete_handlers_for_content_reader_)) {
+          return true;
+        }
+      }
+    }
+
+    // Read content into `req.body`
+    if (!read_content(strm, req, res)) { return false; }
+  }
+
+  // Regular handler
+  if (req.method == "GET" || req.method == "HEAD") {
+    return dispatch_request(req, res, get_handlers_);
+  } else if (req.method == "POST") {
+    return dispatch_request(req, res, post_handlers_);
+  } else if (req.method == "PUT") {
+    return dispatch_request(req, res, put_handlers_);
+  } else if (req.method == "DELETE") {
+    return dispatch_request(req, res, delete_handlers_);
+  } else if (req.method == "OPTIONS") {
+    return dispatch_request(req, res, options_handlers_);
+  } else if (req.method == "PATCH") {
+    return dispatch_request(req, res, patch_handlers_);
+  }
+
+  res.status = StatusCode::BadRequest_400;
+  return false;
+}
+
+inline bool Server::dispatch_request(Request &req, Response &res,
+                                     const Handlers &handlers) const {
+  for (const auto &x : handlers) {
+    const auto &matcher = x.first;
+    const auto &handler = x.second;
+
+    if (matcher->match(req)) {
+      handler(req, res);
+      return true;
+    }
+  }
+  return false;
+}
+
+inline void Server::apply_ranges(const Request &req, Response &res,
+                                 std::string &content_type,
+                                 std::string &boundary) const {
+  if (req.ranges.size() > 1 && res.status == StatusCode::PartialContent_206) {
+    auto it = res.headers.find("Content-Type");
+    if (it != res.headers.end()) {
+      content_type = it->second;
+      res.headers.erase(it);
+    }
+
+    boundary = detail::make_multipart_data_boundary();
+
+    res.set_header("Content-Type",
+                   "multipart/byteranges; boundary=" + boundary);
+  }
+
+  auto type = detail::encoding_type(req, res);
+
+  if (res.body.empty()) {
+    if (res.content_length_ > 0) {
+      size_t length = 0;
+      if (req.ranges.empty() || res.status != StatusCode::PartialContent_206) {
+        length = res.content_length_;
+      } else if (req.ranges.size() == 1) {
+        auto offset_and_length = detail::get_range_offset_and_length(
+            req.ranges[0], res.content_length_);
+
+        length = offset_and_length.second;
+
+        auto content_range = detail::make_content_range_header_field(
+            offset_and_length, res.content_length_);
+        res.set_header("Content-Range", content_range);
+      } else {
+        length = detail::get_multipart_ranges_data_length(
+            req, boundary, content_type, res.content_length_);
+      }
+      res.set_header("Content-Length", std::to_string(length));
+    } else {
+      if (res.content_provider_) {
+        if (res.is_chunked_content_provider_) {
+          res.set_header("Transfer-Encoding", "chunked");
+          if (type == detail::EncodingType::Gzip) {
+            res.set_header("Content-Encoding", "gzip");
+          } else if (type == detail::EncodingType::Brotli) {
+            res.set_header("Content-Encoding", "br");
+          }
+        }
+      }
+    }
+  } else {
+    if (req.ranges.empty() || res.status != StatusCode::PartialContent_206) {
+      ;
+    } else if (req.ranges.size() == 1) {
+      auto offset_and_length =
+          detail::get_range_offset_and_length(req.ranges[0], res.body.size());
+      auto offset = offset_and_length.first;
+      auto length = offset_and_length.second;
+
+      auto content_range = detail::make_content_range_header_field(
+          offset_and_length, res.body.size());
+      res.set_header("Content-Range", content_range);
+
+      assert(offset + length <= res.body.size());
+      res.body = res.body.substr(offset, length);
+    } else {
+      std::string data;
+      detail::make_multipart_ranges_data(req, res, boundary, content_type,
+                                         res.body.size(), data);
+      res.body.swap(data);
+    }
+
+    if (type != detail::EncodingType::None) {
+      std::unique_ptr<detail::compressor> compressor;
+      std::string content_encoding;
+
+      if (type == detail::EncodingType::Gzip) {
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+        compressor = detail::make_unique<detail::gzip_compressor>();
+        content_encoding = "gzip";
+#endif
+      } else if (type == detail::EncodingType::Brotli) {
+#ifdef CPPHTTPLIB_BROTLI_SUPPORT
+        compressor = detail::make_unique<detail::brotli_compressor>();
+        content_encoding = "br";
+#endif
+      }
+
+      if (compressor) {
+        std::string compressed;
+        if (compressor->compress(res.body.data(), res.body.size(), true,
+                                 [&](const char *data, size_t data_len) {
+                                   compressed.append(data, data_len);
+                                   return true;
+                                 })) {
+          res.body.swap(compressed);
+          res.set_header("Content-Encoding", content_encoding);
+        }
+      }
+    }
+
+    auto length = std::to_string(res.body.size());
+    res.set_header("Content-Length", length);
+  }
+}
+
+inline bool Server::dispatch_request_for_content_reader(
+    Request &req, Response &res, ContentReader content_reader,
+    const HandlersForContentReader &handlers) const {
+  for (const auto &x : handlers) {
+    const auto &matcher = x.first;
+    const auto &handler = x.second;
+
+    if (matcher->match(req)) {
+      handler(req, res, content_reader);
+      return true;
+    }
+  }
+  return false;
+}
+
+inline bool
+Server::process_request(Stream &strm, bool close_connection,
+                        bool &connection_closed,
+                        const std::function<void(Request &)> &setup_request) {
+  std::array<char, 2048> buf{};
+
+  detail::stream_line_reader line_reader(strm, buf.data(), buf.size());
+
+  // Connection has been closed on client
+  if (!line_reader.getline()) { return false; }
+
+  Request req;
+
+  Response res;
+  res.version = "HTTP/1.1";
+  res.headers = default_headers_;
+
+#ifdef _WIN32
+  // TODO: Increase FD_SETSIZE statically (libzmq), dynamically (MySQL).
+#else
+#ifndef CPPHTTPLIB_USE_POLL
+  // Socket file descriptor exceeded FD_SETSIZE...
+  if (strm.socket() >= FD_SETSIZE) {
+    Headers dummy;
+    detail::read_headers(strm, dummy);
+    res.status = StatusCode::InternalServerError_500;
+    return write_response(strm, close_connection, req, res);
+  }
+#endif
+#endif
+
+  // Check if the request URI doesn't exceed the limit
+  if (line_reader.size() > CPPHTTPLIB_REQUEST_URI_MAX_LENGTH) {
+    Headers dummy;
+    detail::read_headers(strm, dummy);
+    res.status = StatusCode::UriTooLong_414;
+    return write_response(strm, close_connection, req, res);
+  }
+
+  // Request line and headers
+  if (!parse_request_line(line_reader.ptr(), req) ||
+      !detail::read_headers(strm, req.headers)) {
+    res.status = StatusCode::BadRequest_400;
+    return write_response(strm, close_connection, req, res);
+  }
+
+  if (req.get_header_value("Connection") == "close") {
+    connection_closed = true;
+  }
+
+  if (req.version == "HTTP/1.0" &&
+      req.get_header_value("Connection") != "Keep-Alive") {
+    connection_closed = true;
+  }
+
+  strm.get_remote_ip_and_port(req.remote_addr, req.remote_port);
+  req.set_header("REMOTE_ADDR", req.remote_addr);
+  req.set_header("REMOTE_PORT", std::to_string(req.remote_port));
+
+  strm.get_local_ip_and_port(req.local_addr, req.local_port);
+  req.set_header("LOCAL_ADDR", req.local_addr);
+  req.set_header("LOCAL_PORT", std::to_string(req.local_port));
+
+  if (req.has_header("Range")) {
+    const auto &range_header_value = req.get_header_value("Range");
+    if (!detail::parse_range_header(range_header_value, req.ranges)) {
+      res.status = StatusCode::RangeNotSatisfiable_416;
+      return write_response(strm, close_connection, req, res);
+    }
+  }
+
+  if (setup_request) { setup_request(req); }
+
+  if (req.get_header_value("Expect") == "100-continue") {
+    int status = StatusCode::Continue_100;
+    if (expect_100_continue_handler_) {
+      status = expect_100_continue_handler_(req, res);
+    }
+    switch (status) {
+    case StatusCode::Continue_100:
+    case StatusCode::ExpectationFailed_417:
+      strm.write_format("HTTP/1.1 %d %s\r\n\r\n", status,
+                        status_message(status));
+      break;
+    default: return write_response(strm, close_connection, req, res);
+    }
+  }
+
+  // Routing
+  auto routed = false;
+#ifdef CPPHTTPLIB_NO_EXCEPTIONS
+  routed = routing(req, res, strm);
+#else
+  try {
+    routed = routing(req, res, strm);
+  } catch (std::exception &e) {
+    if (exception_handler_) {
+      auto ep = std::current_exception();
+      exception_handler_(req, res, ep);
+      routed = true;
+    } else {
+      res.status = StatusCode::InternalServerError_500;
+      std::string val;
+      auto s = e.what();
+      for (size_t i = 0; s[i]; i++) {
+        switch (s[i]) {
+        case '\r': val += "\\r"; break;
+        case '\n': val += "\\n"; break;
+        default: val += s[i]; break;
+        }
+      }
+      res.set_header("EXCEPTION_WHAT", val);
+    }
+  } catch (...) {
+    if (exception_handler_) {
+      auto ep = std::current_exception();
+      exception_handler_(req, res, ep);
+      routed = true;
+    } else {
+      res.status = StatusCode::InternalServerError_500;
+      res.set_header("EXCEPTION_WHAT", "UNKNOWN");
+    }
+  }
+#endif
+  if (routed) {
+    if (res.status == -1) {
+      res.status = req.ranges.empty() ? StatusCode::OK_200
+                                      : StatusCode::PartialContent_206;
+    }
+
+    if (detail::range_error(req, res)) {
+      res.body.clear();
+      res.content_length_ = 0;
+      res.content_provider_ = nullptr;
+      res.status = StatusCode::RangeNotSatisfiable_416;
+      return write_response(strm, close_connection, req, res);
+    }
+
+    return write_response_with_content(strm, close_connection, req, res);
+  } else {
+    if (res.status == -1) { res.status = StatusCode::NotFound_404; }
+
+    return write_response(strm, close_connection, req, res);
+  }
+}
+
+inline bool Server::is_valid() const { return true; }
+
+inline bool Server::process_and_close_socket(socket_t sock) {
+  auto ret = detail::process_server_socket(
+      svr_sock_, sock, keep_alive_max_count_, keep_alive_timeout_sec_,
+      read_timeout_sec_, read_timeout_usec_, write_timeout_sec_,
+      write_timeout_usec_,
+      [this](Stream &strm, bool close_connection, bool &connection_closed) {
+        return process_request(strm, close_connection, connection_closed,
+                               nullptr);
+      });
+
+  detail::shutdown_socket(sock);
+  detail::close_socket(sock);
+  return ret;
+}
+
+// HTTP client implementation
+inline ClientImpl::ClientImpl(const std::string &host)
+    : ClientImpl(host, 80, std::string(), std::string()) {}
+
+inline ClientImpl::ClientImpl(const std::string &host, int port)
+    : ClientImpl(host, port, std::string(), std::string()) {}
+
+inline ClientImpl::ClientImpl(const std::string &host, int port,
+                              const std::string &client_cert_path,
+                              const std::string &client_key_path)
+    : host_(host), port_(port),
+      host_and_port_(adjust_host_string(host) + ":" + std::to_string(port)),
+      client_cert_path_(client_cert_path), client_key_path_(client_key_path) {}
+
+inline ClientImpl::~ClientImpl() {
+  std::lock_guard<std::mutex> guard(socket_mutex_);
+  shutdown_socket(socket_);
+  close_socket(socket_);
+}
+
+inline bool ClientImpl::is_valid() const { return true; }
+
+inline void ClientImpl::copy_settings(const ClientImpl &rhs) {
+  client_cert_path_ = rhs.client_cert_path_;
+  client_key_path_ = rhs.client_key_path_;
+  connection_timeout_sec_ = rhs.connection_timeout_sec_;
+  read_timeout_sec_ = rhs.read_timeout_sec_;
+  read_timeout_usec_ = rhs.read_timeout_usec_;
+  write_timeout_sec_ = rhs.write_timeout_sec_;
+  write_timeout_usec_ = rhs.write_timeout_usec_;
+  basic_auth_username_ = rhs.basic_auth_username_;
+  basic_auth_password_ = rhs.basic_auth_password_;
+  bearer_token_auth_token_ = rhs.bearer_token_auth_token_;
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  digest_auth_username_ = rhs.digest_auth_username_;
+  digest_auth_password_ = rhs.digest_auth_password_;
+#endif
+  keep_alive_ = rhs.keep_alive_;
+  follow_location_ = rhs.follow_location_;
+  url_encode_ = rhs.url_encode_;
+  address_family_ = rhs.address_family_;
+  tcp_nodelay_ = rhs.tcp_nodelay_;
+  socket_options_ = rhs.socket_options_;
+  compress_ = rhs.compress_;
+  decompress_ = rhs.decompress_;
+  interface_ = rhs.interface_;
+  proxy_host_ = rhs.proxy_host_;
+  proxy_port_ = rhs.proxy_port_;
+  proxy_basic_auth_username_ = rhs.proxy_basic_auth_username_;
+  proxy_basic_auth_password_ = rhs.proxy_basic_auth_password_;
+  proxy_bearer_token_auth_token_ = rhs.proxy_bearer_token_auth_token_;
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  proxy_digest_auth_username_ = rhs.proxy_digest_auth_username_;
+  proxy_digest_auth_password_ = rhs.proxy_digest_auth_password_;
+#endif
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  ca_cert_file_path_ = rhs.ca_cert_file_path_;
+  ca_cert_dir_path_ = rhs.ca_cert_dir_path_;
+  ca_cert_store_ = rhs.ca_cert_store_;
+#endif
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  server_certificate_verification_ = rhs.server_certificate_verification_;
+#endif
+  logger_ = rhs.logger_;
+}
+
+inline socket_t ClientImpl::create_client_socket(Error &error) const {
+  if (!proxy_host_.empty() && proxy_port_ != -1) {
+    return detail::create_client_socket(
+        proxy_host_, std::string(), proxy_port_, address_family_, tcp_nodelay_,
+        socket_options_, connection_timeout_sec_, connection_timeout_usec_,
+        read_timeout_sec_, read_timeout_usec_, write_timeout_sec_,
+        write_timeout_usec_, interface_, error);
+  }
+
+  // Check is custom IP specified for host_
+  std::string ip;
+  auto it = addr_map_.find(host_);
+  if (it != addr_map_.end()) { ip = it->second; }
+
+  return detail::create_client_socket(
+      host_, ip, port_, address_family_, tcp_nodelay_, socket_options_,
+      connection_timeout_sec_, connection_timeout_usec_, read_timeout_sec_,
+      read_timeout_usec_, write_timeout_sec_, write_timeout_usec_, interface_,
+      error);
+}
+
+inline bool ClientImpl::create_and_connect_socket(Socket &socket,
+                                                  Error &error) {
+  auto sock = create_client_socket(error);
+  if (sock == INVALID_SOCKET) { return false; }
+  socket.sock = sock;
+  return true;
+}
+
+inline void ClientImpl::shutdown_ssl(Socket & /*socket*/,
+                                     bool /*shutdown_gracefully*/) {
+  // If there are any requests in flight from threads other than us, then it's
+  // a thread-unsafe race because individual ssl* objects are not thread-safe.
+  assert(socket_requests_in_flight_ == 0 ||
+         socket_requests_are_from_thread_ == std::this_thread::get_id());
+}
+
+inline void ClientImpl::shutdown_socket(Socket &socket) const {
+  if (socket.sock == INVALID_SOCKET) { return; }
+  detail::shutdown_socket(socket.sock);
+}
+
+inline void ClientImpl::close_socket(Socket &socket) {
+  // If there are requests in flight in another thread, usually closing
+  // the socket will be fine and they will simply receive an error when
+  // using the closed socket, but it is still a bug since rarely the OS
+  // may reassign the socket id to be used for a new socket, and then
+  // suddenly they will be operating on a live socket that is different
+  // than the one they intended!
+  assert(socket_requests_in_flight_ == 0 ||
+         socket_requests_are_from_thread_ == std::this_thread::get_id());
+
+  // It is also a bug if this happens while SSL is still active
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  assert(socket.ssl == nullptr);
+#endif
+  if (socket.sock == INVALID_SOCKET) { return; }
+  detail::close_socket(socket.sock);
+  socket.sock = INVALID_SOCKET;
+}
+
+inline bool ClientImpl::read_response_line(Stream &strm, const Request &req,
+                                           Response &res) const {
+  std::array<char, 2048> buf{};
+
+  detail::stream_line_reader line_reader(strm, buf.data(), buf.size());
+
+  if (!line_reader.getline()) { return false; }
+
+#ifdef CPPHTTPLIB_ALLOW_LF_AS_LINE_TERMINATOR
+  const static std::regex re("(HTTP/1\\.[01]) (\\d{3})(?: (.*?))?\r?\n");
+#else
+  const static std::regex re("(HTTP/1\\.[01]) (\\d{3})(?: (.*?))?\r\n");
+#endif
+
+  std::cmatch m;
+  if (!std::regex_match(line_reader.ptr(), m, re)) {
+    return req.method == "CONNECT";
+  }
+  res.version = std::string(m[1]);
+  res.status = std::stoi(std::string(m[2]));
+  res.reason = std::string(m[3]);
+
+  // Ignore '100 Continue'
+  while (res.status == StatusCode::Continue_100) {
+    if (!line_reader.getline()) { return false; } // CRLF
+    if (!line_reader.getline()) { return false; } // next response line
+
+    if (!std::regex_match(line_reader.ptr(), m, re)) { return false; }
+    res.version = std::string(m[1]);
+    res.status = std::stoi(std::string(m[2]));
+    res.reason = std::string(m[3]);
+  }
+
+  return true;
+}
+
+inline bool ClientImpl::send(Request &req, Response &res, Error &error) {
+  std::lock_guard<std::recursive_mutex> request_mutex_guard(request_mutex_);
+  auto ret = send_(req, res, error);
+  if (error == Error::SSLPeerCouldBeClosed_) {
+    assert(!ret);
+    ret = send_(req, res, error);
+  }
+  return ret;
+}
+
+inline bool ClientImpl::send_(Request &req, Response &res, Error &error) {
+  {
+    std::lock_guard<std::mutex> guard(socket_mutex_);
+
+    // Set this to false immediately - if it ever gets set to true by the end of
+    // the request, we know another thread instructed us to close the socket.
+    socket_should_be_closed_when_request_is_done_ = false;
+
+    auto is_alive = false;
+    if (socket_.is_open()) {
+      is_alive = detail::is_socket_alive(socket_.sock);
+      if (!is_alive) {
+        // Attempt to avoid sigpipe by shutting down nongracefully if it seems
+        // like the other side has already closed the connection Also, there
+        // cannot be any requests in flight from other threads since we locked
+        // request_mutex_, so safe to close everything immediately
+        const bool shutdown_gracefully = false;
+        shutdown_ssl(socket_, shutdown_gracefully);
+        shutdown_socket(socket_);
+        close_socket(socket_);
+      }
+    }
+
+    if (!is_alive) {
+      if (!create_and_connect_socket(socket_, error)) { return false; }
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+      // TODO: refactoring
+      if (is_ssl()) {
+        auto &scli = static_cast<SSLClient &>(*this);
+        if (!proxy_host_.empty() && proxy_port_ != -1) {
+          auto success = false;
+          if (!scli.connect_with_proxy(socket_, res, success, error)) {
+            return success;
+          }
+        }
+
+        if (!scli.initialize_ssl(socket_, error)) { return false; }
+      }
+#endif
+    }
+
+    // Mark the current socket as being in use so that it cannot be closed by
+    // anyone else while this request is ongoing, even though we will be
+    // releasing the mutex.
+    if (socket_requests_in_flight_ > 1) {
+      assert(socket_requests_are_from_thread_ == std::this_thread::get_id());
+    }
+    socket_requests_in_flight_ += 1;
+    socket_requests_are_from_thread_ = std::this_thread::get_id();
+  }
+
+  for (const auto &header : default_headers_) {
+    if (req.headers.find(header.first) == req.headers.end()) {
+      req.headers.insert(header);
+    }
+  }
+
+  auto ret = false;
+  auto close_connection = !keep_alive_;
+
+  auto se = detail::scope_exit([&]() {
+    // Briefly lock mutex in order to mark that a request is no longer ongoing
+    std::lock_guard<std::mutex> guard(socket_mutex_);
+    socket_requests_in_flight_ -= 1;
+    if (socket_requests_in_flight_ <= 0) {
+      assert(socket_requests_in_flight_ == 0);
+      socket_requests_are_from_thread_ = std::thread::id();
+    }
+
+    if (socket_should_be_closed_when_request_is_done_ || close_connection ||
+        !ret) {
+      shutdown_ssl(socket_, true);
+      shutdown_socket(socket_);
+      close_socket(socket_);
+    }
+  });
+
+  ret = process_socket(socket_, [&](Stream &strm) {
+    return handle_request(strm, req, res, close_connection, error);
+  });
+
+  if (!ret) {
+    if (error == Error::Success) { error = Error::Unknown; }
+  }
+
+  return ret;
+}
+
+inline Result ClientImpl::send(const Request &req) {
+  auto req2 = req;
+  return send_(std::move(req2));
+}
+
+inline Result ClientImpl::send_(Request &&req) {
+  auto res = detail::make_unique<Response>();
+  auto error = Error::Success;
+  auto ret = send(req, *res, error);
+  return Result{ret ? std::move(res) : nullptr, error, std::move(req.headers)};
+}
+
+inline bool ClientImpl::handle_request(Stream &strm, Request &req,
+                                       Response &res, bool close_connection,
+                                       Error &error) {
+  if (req.path.empty()) {
+    error = Error::Connection;
+    return false;
+  }
+
+  auto req_save = req;
+
+  bool ret;
+
+  if (!is_ssl() && !proxy_host_.empty() && proxy_port_ != -1) {
+    auto req2 = req;
+    req2.path = "http://" + host_and_port_ + req.path;
+    ret = process_request(strm, req2, res, close_connection, error);
+    req = req2;
+    req.path = req_save.path;
+  } else {
+    ret = process_request(strm, req, res, close_connection, error);
+  }
+
+  if (!ret) { return false; }
+
+  if (res.get_header_value("Connection") == "close" ||
+      (res.version == "HTTP/1.0" && res.reason != "Connection established")) {
+    // TODO this requires a not-entirely-obvious chain of calls to be correct
+    // for this to be safe.
+
+    // This is safe to call because handle_request is only called by send_
+    // which locks the request mutex during the process. It would be a bug
+    // to call it from a different thread since it's a thread-safety issue
+    // to do these things to the socket if another thread is using the socket.
+    std::lock_guard<std::mutex> guard(socket_mutex_);
+    shutdown_ssl(socket_, true);
+    shutdown_socket(socket_);
+    close_socket(socket_);
+  }
+
+  if (300 < res.status && res.status < 400 && follow_location_) {
+    req = req_save;
+    ret = redirect(req, res, error);
+  }
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  if ((res.status == StatusCode::Unauthorized_401 ||
+       res.status == StatusCode::ProxyAuthenticationRequired_407) &&
+      req.authorization_count_ < 5) {
+    auto is_proxy = res.status == StatusCode::ProxyAuthenticationRequired_407;
+    const auto &username =
+        is_proxy ? proxy_digest_auth_username_ : digest_auth_username_;
+    const auto &password =
+        is_proxy ? proxy_digest_auth_password_ : digest_auth_password_;
+
+    if (!username.empty() && !password.empty()) {
+      std::map<std::string, std::string> auth;
+      if (detail::parse_www_authenticate(res, auth, is_proxy)) {
+        Request new_req = req;
+        new_req.authorization_count_ += 1;
+        new_req.headers.erase(is_proxy ? "Proxy-Authorization"
+                                       : "Authorization");
+        new_req.headers.insert(detail::make_digest_authentication_header(
+            req, auth, new_req.authorization_count_, detail::random_string(10),
+            username, password, is_proxy));
+
+        Response new_res;
+
+        ret = send(new_req, new_res, error);
+        if (ret) { res = new_res; }
+      }
+    }
+  }
+#endif
+
+  return ret;
+}
+
+inline bool ClientImpl::redirect(Request &req, Response &res, Error &error) {
+  if (req.redirect_count_ == 0) {
+    error = Error::ExceedRedirectCount;
+    return false;
+  }
+
+  auto location = res.get_header_value("location");
+  if (location.empty()) { return false; }
+
+  const static std::regex re(
+      R"((?:(https?):)?(?://(?:\[([\d:]+)\]|([^:/?#]+))(?::(\d+))?)?([^?#]*)(\?[^#]*)?(?:#.*)?)");
+
+  std::smatch m;
+  if (!std::regex_match(location, m, re)) { return false; }
+
+  auto scheme = is_ssl() ? "https" : "http";
+
+  auto next_scheme = m[1].str();
+  auto next_host = m[2].str();
+  if (next_host.empty()) { next_host = m[3].str(); }
+  auto port_str = m[4].str();
+  auto next_path = m[5].str();
+  auto next_query = m[6].str();
+
+  auto next_port = port_;
+  if (!port_str.empty()) {
+    next_port = std::stoi(port_str);
+  } else if (!next_scheme.empty()) {
+    next_port = next_scheme == "https" ? 443 : 80;
+  }
+
+  if (next_scheme.empty()) { next_scheme = scheme; }
+  if (next_host.empty()) { next_host = host_; }
+  if (next_path.empty()) { next_path = "/"; }
+
+  auto path = detail::decode_url(next_path, true) + next_query;
+
+  if (next_scheme == scheme && next_host == host_ && next_port == port_) {
+    return detail::redirect(*this, req, res, path, location, error);
+  } else {
+    if (next_scheme == "https") {
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+      SSLClient cli(next_host, next_port);
+      cli.copy_settings(*this);
+      if (ca_cert_store_) { cli.set_ca_cert_store(ca_cert_store_); }
+      return detail::redirect(cli, req, res, path, location, error);
+#else
+      return false;
+#endif
+    } else {
+      ClientImpl cli(next_host, next_port);
+      cli.copy_settings(*this);
+      return detail::redirect(cli, req, res, path, location, error);
+    }
+  }
+}
+
+inline bool ClientImpl::write_content_with_provider(Stream &strm,
+                                                    const Request &req,
+                                                    Error &error) const {
+  auto is_shutting_down = []() { return false; };
+
+  if (req.is_chunked_content_provider_) {
+    // TODO: Brotli support
+    std::unique_ptr<detail::compressor> compressor;
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+    if (compress_) {
+      compressor = detail::make_unique<detail::gzip_compressor>();
+    } else
+#endif
+    {
+      compressor = detail::make_unique<detail::nocompressor>();
+    }
+
+    return detail::write_content_chunked(strm, req.content_provider_,
+                                         is_shutting_down, *compressor, error);
+  } else {
+    return detail::write_content(strm, req.content_provider_, 0,
+                                 req.content_length_, is_shutting_down, error);
+  }
+}
+
+inline bool ClientImpl::write_request(Stream &strm, Request &req,
+                                      bool close_connection, Error &error) {
+  // Prepare additional headers
+  if (close_connection) {
+    if (!req.has_header("Connection")) {
+      req.set_header("Connection", "close");
+    }
+  }
+
+  if (!req.has_header("Host")) {
+    if (is_ssl()) {
+      if (port_ == 443) {
+        req.set_header("Host", host_);
+      } else {
+        req.set_header("Host", host_and_port_);
+      }
+    } else {
+      if (port_ == 80) {
+        req.set_header("Host", host_);
+      } else {
+        req.set_header("Host", host_and_port_);
+      }
+    }
+  }
+
+  if (!req.has_header("Accept")) { req.set_header("Accept", "*/*"); }
+
+#ifndef CPPHTTPLIB_NO_DEFAULT_USER_AGENT
+  if (!req.has_header("User-Agent")) {
+    auto agent = std::string("cpp-httplib/") + CPPHTTPLIB_VERSION;
+    req.set_header("User-Agent", agent);
+  }
+#endif
+
+  if (req.body.empty()) {
+    if (req.content_provider_) {
+      if (!req.is_chunked_content_provider_) {
+        if (!req.has_header("Content-Length")) {
+          auto length = std::to_string(req.content_length_);
+          req.set_header("Content-Length", length);
+        }
+      }
+    } else {
+      if (req.method == "POST" || req.method == "PUT" ||
+          req.method == "PATCH") {
+        req.set_header("Content-Length", "0");
+      }
+    }
+  } else {
+    if (!req.has_header("Content-Type")) {
+      req.set_header("Content-Type", "text/plain");
+    }
+
+    if (!req.has_header("Content-Length")) {
+      auto length = std::to_string(req.body.size());
+      req.set_header("Content-Length", length);
+    }
+  }
+
+  if (!basic_auth_password_.empty() || !basic_auth_username_.empty()) {
+    if (!req.has_header("Authorization")) {
+      req.headers.insert(make_basic_authentication_header(
+          basic_auth_username_, basic_auth_password_, false));
+    }
+  }
+
+  if (!proxy_basic_auth_username_.empty() &&
+      !proxy_basic_auth_password_.empty()) {
+    if (!req.has_header("Proxy-Authorization")) {
+      req.headers.insert(make_basic_authentication_header(
+          proxy_basic_auth_username_, proxy_basic_auth_password_, true));
+    }
+  }
+
+  if (!bearer_token_auth_token_.empty()) {
+    if (!req.has_header("Authorization")) {
+      req.headers.insert(make_bearer_token_authentication_header(
+          bearer_token_auth_token_, false));
+    }
+  }
+
+  if (!proxy_bearer_token_auth_token_.empty()) {
+    if (!req.has_header("Proxy-Authorization")) {
+      req.headers.insert(make_bearer_token_authentication_header(
+          proxy_bearer_token_auth_token_, true));
+    }
+  }
+
+  // Request line and headers
+  {
+    detail::BufferStream bstrm;
+
+    const auto &path = url_encode_ ? detail::encode_url(req.path) : req.path;
+    bstrm.write_format("%s %s HTTP/1.1\r\n", req.method.c_str(), path.c_str());
+
+    header_writer_(bstrm, req.headers);
+
+    // Flush buffer
+    auto &data = bstrm.get_buffer();
+    if (!detail::write_data(strm, data.data(), data.size())) {
+      error = Error::Write;
+      return false;
+    }
+  }
+
+  // Body
+  if (req.body.empty()) {
+    return write_content_with_provider(strm, req, error);
+  }
+
+  if (!detail::write_data(strm, req.body.data(), req.body.size())) {
+    error = Error::Write;
+    return false;
+  }
+
+  return true;
+}
+
+inline std::unique_ptr<Response> ClientImpl::send_with_content_provider(
+    Request &req, const char *body, size_t content_length,
+    ContentProvider content_provider,
+    ContentProviderWithoutLength content_provider_without_length,
+    const std::string &content_type, Error &error) {
+  if (!content_type.empty()) { req.set_header("Content-Type", content_type); }
+
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+  if (compress_) { req.set_header("Content-Encoding", "gzip"); }
+#endif
+
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+  if (compress_ && !content_provider_without_length) {
+    // TODO: Brotli support
+    detail::gzip_compressor compressor;
+
+    if (content_provider) {
+      auto ok = true;
+      size_t offset = 0;
+      DataSink data_sink;
+
+      data_sink.write = [&](const char *data, size_t data_len) -> bool {
+        if (ok) {
+          auto last = offset + data_len == content_length;
+
+          auto ret = compressor.compress(
+              data, data_len, last,
+              [&](const char *compressed_data, size_t compressed_data_len) {
+                req.body.append(compressed_data, compressed_data_len);
+                return true;
+              });
+
+          if (ret) {
+            offset += data_len;
+          } else {
+            ok = false;
+          }
+        }
+        return ok;
+      };
+
+      while (ok && offset < content_length) {
+        if (!content_provider(offset, content_length - offset, data_sink)) {
+          error = Error::Canceled;
+          return nullptr;
+        }
+      }
+    } else {
+      if (!compressor.compress(body, content_length, true,
+                               [&](const char *data, size_t data_len) {
+                                 req.body.append(data, data_len);
+                                 return true;
+                               })) {
+        error = Error::Compression;
+        return nullptr;
+      }
+    }
+  } else
+#endif
+  {
+    if (content_provider) {
+      req.content_length_ = content_length;
+      req.content_provider_ = std::move(content_provider);
+      req.is_chunked_content_provider_ = false;
+    } else if (content_provider_without_length) {
+      req.content_length_ = 0;
+      req.content_provider_ = detail::ContentProviderAdapter(
+          std::move(content_provider_without_length));
+      req.is_chunked_content_provider_ = true;
+      req.set_header("Transfer-Encoding", "chunked");
+    } else {
+      req.body.assign(body, content_length);
+    }
+  }
+
+  auto res = detail::make_unique<Response>();
+  return send(req, *res, error) ? std::move(res) : nullptr;
+}
+
+inline Result ClientImpl::send_with_content_provider(
+    const std::string &method, const std::string &path, const Headers &headers,
+    const char *body, size_t content_length, ContentProvider content_provider,
+    ContentProviderWithoutLength content_provider_without_length,
+    const std::string &content_type) {
+  Request req;
+  req.method = method;
+  req.headers = headers;
+  req.path = path;
+
+  auto error = Error::Success;
+
+  auto res = send_with_content_provider(
+      req, body, content_length, std::move(content_provider),
+      std::move(content_provider_without_length), content_type, error);
+
+  return Result{std::move(res), error, std::move(req.headers)};
+}
+
+inline std::string
+ClientImpl::adjust_host_string(const std::string &host) const {
+  if (host.find(':') != std::string::npos) { return "[" + host + "]"; }
+  return host;
+}
+
+inline bool ClientImpl::process_request(Stream &strm, Request &req,
+                                        Response &res, bool close_connection,
+                                        Error &error) {
+  // Send request
+  if (!write_request(strm, req, close_connection, error)) { return false; }
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  if (is_ssl()) {
+    auto is_proxy_enabled = !proxy_host_.empty() && proxy_port_ != -1;
+    if (!is_proxy_enabled) {
+      char buf[1];
+      if (SSL_peek(socket_.ssl, buf, 1) == 0 &&
+          SSL_get_error(socket_.ssl, 0) == SSL_ERROR_ZERO_RETURN) {
+        error = Error::SSLPeerCouldBeClosed_;
+        return false;
+      }
+    }
+  }
+#endif
+
+  // Receive response and headers
+  if (!read_response_line(strm, req, res) ||
+      !detail::read_headers(strm, res.headers)) {
+    error = Error::Read;
+    return false;
+  }
+
+  // Body
+  if ((res.status != StatusCode::NoContent_204) && req.method != "HEAD" &&
+      req.method != "CONNECT") {
+    auto redirect = 300 < res.status && res.status < 400 && follow_location_;
+
+    if (req.response_handler && !redirect) {
+      if (!req.response_handler(res)) {
+        error = Error::Canceled;
+        return false;
+      }
+    }
+
+    auto out =
+        req.content_receiver
+            ? static_cast<ContentReceiverWithProgress>(
+                  [&](const char *buf, size_t n, uint64_t off, uint64_t len) {
+                    if (redirect) { return true; }
+                    auto ret = req.content_receiver(buf, n, off, len);
+                    if (!ret) { error = Error::Canceled; }
+                    return ret;
+                  })
+            : static_cast<ContentReceiverWithProgress>(
+                  [&](const char *buf, size_t n, uint64_t /*off*/,
+                      uint64_t /*len*/) {
+                    if (res.body.size() + n > res.body.max_size()) {
+                      return false;
+                    }
+                    res.body.append(buf, n);
+                    return true;
+                  });
+
+    auto progress = [&](uint64_t current, uint64_t total) {
+      if (!req.progress || redirect) { return true; }
+      auto ret = req.progress(current, total);
+      if (!ret) { error = Error::Canceled; }
+      return ret;
+    };
+
+    int dummy_status;
+    if (!detail::read_content(strm, res, (std::numeric_limits<size_t>::max)(),
+                              dummy_status, std::move(progress), std::move(out),
+                              decompress_)) {
+      if (error != Error::Canceled) { error = Error::Read; }
+      return false;
+    }
+  }
+
+  // Log
+  if (logger_) { logger_(req, res); }
+
+  return true;
+}
+
+inline ContentProviderWithoutLength ClientImpl::get_multipart_content_provider(
+    const std::string &boundary, const MultipartFormDataItems &items,
+    const MultipartFormDataProviderItems &provider_items) const {
+  size_t cur_item = 0;
+  size_t cur_start = 0;
+  // cur_item and cur_start are copied to within the std::function and maintain
+  // state between successive calls
+  return [&, cur_item, cur_start](size_t offset,
+                                  DataSink &sink) mutable -> bool {
+    if (!offset && !items.empty()) {
+      sink.os << detail::serialize_multipart_formdata(items, boundary, false);
+      return true;
+    } else if (cur_item < provider_items.size()) {
+      if (!cur_start) {
+        const auto &begin = detail::serialize_multipart_formdata_item_begin(
+            provider_items[cur_item], boundary);
+        offset += begin.size();
+        cur_start = offset;
+        sink.os << begin;
+      }
+
+      DataSink cur_sink;
+      auto has_data = true;
+      cur_sink.write = sink.write;
+      cur_sink.done = [&]() { has_data = false; };
+
+      if (!provider_items[cur_item].provider(offset - cur_start, cur_sink)) {
+        return false;
+      }
+
+      if (!has_data) {
+        sink.os << detail::serialize_multipart_formdata_item_end();
+        cur_item++;
+        cur_start = 0;
+      }
+      return true;
+    } else {
+      sink.os << detail::serialize_multipart_formdata_finish(boundary);
+      sink.done();
+      return true;
+    }
+  };
+}
+
+inline bool
+ClientImpl::process_socket(const Socket &socket,
+                           std::function<bool(Stream &strm)> callback) {
+  return detail::process_client_socket(
+      socket.sock, read_timeout_sec_, read_timeout_usec_, write_timeout_sec_,
+      write_timeout_usec_, std::move(callback));
+}
+
+inline bool ClientImpl::is_ssl() const { return false; }
+
+inline Result ClientImpl::Get(const std::string &path) {
+  return Get(path, Headers(), Progress());
+}
+
+inline Result ClientImpl::Get(const std::string &path, Progress progress) {
+  return Get(path, Headers(), std::move(progress));
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Headers &headers) {
+  return Get(path, headers, Progress());
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Headers &headers,
+                              Progress progress) {
+  Request req;
+  req.method = "GET";
+  req.path = path;
+  req.headers = headers;
+  req.progress = std::move(progress);
+
+  return send_(std::move(req));
+}
+
+inline Result ClientImpl::Get(const std::string &path,
+                              ContentReceiver content_receiver) {
+  return Get(path, Headers(), nullptr, std::move(content_receiver), nullptr);
+}
+
+inline Result ClientImpl::Get(const std::string &path,
+                              ContentReceiver content_receiver,
+                              Progress progress) {
+  return Get(path, Headers(), nullptr, std::move(content_receiver),
+             std::move(progress));
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Headers &headers,
+                              ContentReceiver content_receiver) {
+  return Get(path, headers, nullptr, std::move(content_receiver), nullptr);
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Headers &headers,
+                              ContentReceiver content_receiver,
+                              Progress progress) {
+  return Get(path, headers, nullptr, std::move(content_receiver),
+             std::move(progress));
+}
+
+inline Result ClientImpl::Get(const std::string &path,
+                              ResponseHandler response_handler,
+                              ContentReceiver content_receiver) {
+  return Get(path, Headers(), std::move(response_handler),
+             std::move(content_receiver), nullptr);
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Headers &headers,
+                              ResponseHandler response_handler,
+                              ContentReceiver content_receiver) {
+  return Get(path, headers, std::move(response_handler),
+             std::move(content_receiver), nullptr);
+}
+
+inline Result ClientImpl::Get(const std::string &path,
+                              ResponseHandler response_handler,
+                              ContentReceiver content_receiver,
+                              Progress progress) {
+  return Get(path, Headers(), std::move(response_handler),
+             std::move(content_receiver), std::move(progress));
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Headers &headers,
+                              ResponseHandler response_handler,
+                              ContentReceiver content_receiver,
+                              Progress progress) {
+  Request req;
+  req.method = "GET";
+  req.path = path;
+  req.headers = headers;
+  req.response_handler = std::move(response_handler);
+  req.content_receiver =
+      [content_receiver](const char *data, size_t data_length,
+                         uint64_t /*offset*/, uint64_t /*total_length*/) {
+        return content_receiver(data, data_length);
+      };
+  req.progress = std::move(progress);
+
+  return send_(std::move(req));
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Params &params,
+                              const Headers &headers, Progress progress) {
+  if (params.empty()) { return Get(path, headers); }
+
+  std::string path_with_query = append_query_params(path, params);
+  return Get(path_with_query, headers, std::move(progress));
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Params &params,
+                              const Headers &headers,
+                              ContentReceiver content_receiver,
+                              Progress progress) {
+  return Get(path, params, headers, nullptr, std::move(content_receiver),
+             std::move(progress));
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Params &params,
+                              const Headers &headers,
+                              ResponseHandler response_handler,
+                              ContentReceiver content_receiver,
+                              Progress progress) {
+  if (params.empty()) {
+    return Get(path, headers, std::move(response_handler),
+               std::move(content_receiver), std::move(progress));
+  }
+
+  std::string path_with_query = append_query_params(path, params);
+  return Get(path_with_query, headers, std::move(response_handler),
+             std::move(content_receiver), std::move(progress));
+}
+
+inline Result ClientImpl::Head(const std::string &path) {
+  return Head(path, Headers());
+}
+
+inline Result ClientImpl::Head(const std::string &path,
+                               const Headers &headers) {
+  Request req;
+  req.method = "HEAD";
+  req.headers = headers;
+  req.path = path;
+
+  return send_(std::move(req));
+}
+
+inline Result ClientImpl::Post(const std::string &path) {
+  return Post(path, std::string(), std::string());
+}
+
+inline Result ClientImpl::Post(const std::string &path,
+                               const Headers &headers) {
+  return Post(path, headers, nullptr, 0, std::string());
+}
+
+inline Result ClientImpl::Post(const std::string &path, const char *body,
+                               size_t content_length,
+                               const std::string &content_type) {
+  return Post(path, Headers(), body, content_length, content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Headers &headers,
+                               const char *body, size_t content_length,
+                               const std::string &content_type) {
+  return send_with_content_provider("POST", path, headers, body, content_length,
+                                    nullptr, nullptr, content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const std::string &body,
+                               const std::string &content_type) {
+  return Post(path, Headers(), body, content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Headers &headers,
+                               const std::string &body,
+                               const std::string &content_type) {
+  return send_with_content_provider("POST", path, headers, body.data(),
+                                    body.size(), nullptr, nullptr,
+                                    content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Params &params) {
+  return Post(path, Headers(), params);
+}
+
+inline Result ClientImpl::Post(const std::string &path, size_t content_length,
+                               ContentProvider content_provider,
+                               const std::string &content_type) {
+  return Post(path, Headers(), content_length, std::move(content_provider),
+              content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path,
+                               ContentProviderWithoutLength content_provider,
+                               const std::string &content_type) {
+  return Post(path, Headers(), std::move(content_provider), content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Headers &headers,
+                               size_t content_length,
+                               ContentProvider content_provider,
+                               const std::string &content_type) {
+  return send_with_content_provider("POST", path, headers, nullptr,
+                                    content_length, std::move(content_provider),
+                                    nullptr, content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Headers &headers,
+                               ContentProviderWithoutLength content_provider,
+                               const std::string &content_type) {
+  return send_with_content_provider("POST", path, headers, nullptr, 0, nullptr,
+                                    std::move(content_provider), content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Headers &headers,
+                               const Params &params) {
+  auto query = detail::params_to_query_str(params);
+  return Post(path, headers, query, "application/x-www-form-urlencoded");
+}
+
+inline Result ClientImpl::Post(const std::string &path,
+                               const MultipartFormDataItems &items) {
+  return Post(path, Headers(), items);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Headers &headers,
+                               const MultipartFormDataItems &items) {
+  const auto &boundary = detail::make_multipart_data_boundary();
+  const auto &content_type =
+      detail::serialize_multipart_formdata_get_content_type(boundary);
+  const auto &body = detail::serialize_multipart_formdata(items, boundary);
+  return Post(path, headers, body, content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Headers &headers,
+                               const MultipartFormDataItems &items,
+                               const std::string &boundary) {
+  if (!detail::is_multipart_boundary_chars_valid(boundary)) {
+    return Result{nullptr, Error::UnsupportedMultipartBoundaryChars};
+  }
+
+  const auto &content_type =
+      detail::serialize_multipart_formdata_get_content_type(boundary);
+  const auto &body = detail::serialize_multipart_formdata(items, boundary);
+  return Post(path, headers, body, content_type);
+}
+
+inline Result
+ClientImpl::Post(const std::string &path, const Headers &headers,
+                 const MultipartFormDataItems &items,
+                 const MultipartFormDataProviderItems &provider_items) {
+  const auto &boundary = detail::make_multipart_data_boundary();
+  const auto &content_type =
+      detail::serialize_multipart_formdata_get_content_type(boundary);
+  return send_with_content_provider(
+      "POST", path, headers, nullptr, 0, nullptr,
+      get_multipart_content_provider(boundary, items, provider_items),
+      content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path) {
+  return Put(path, std::string(), std::string());
+}
+
+inline Result ClientImpl::Put(const std::string &path, const char *body,
+                              size_t content_length,
+                              const std::string &content_type) {
+  return Put(path, Headers(), body, content_length, content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Headers &headers,
+                              const char *body, size_t content_length,
+                              const std::string &content_type) {
+  return send_with_content_provider("PUT", path, headers, body, content_length,
+                                    nullptr, nullptr, content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const std::string &body,
+                              const std::string &content_type) {
+  return Put(path, Headers(), body, content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Headers &headers,
+                              const std::string &body,
+                              const std::string &content_type) {
+  return send_with_content_provider("PUT", path, headers, body.data(),
+                                    body.size(), nullptr, nullptr,
+                                    content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, size_t content_length,
+                              ContentProvider content_provider,
+                              const std::string &content_type) {
+  return Put(path, Headers(), content_length, std::move(content_provider),
+             content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path,
+                              ContentProviderWithoutLength content_provider,
+                              const std::string &content_type) {
+  return Put(path, Headers(), std::move(content_provider), content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Headers &headers,
+                              size_t content_length,
+                              ContentProvider content_provider,
+                              const std::string &content_type) {
+  return send_with_content_provider("PUT", path, headers, nullptr,
+                                    content_length, std::move(content_provider),
+                                    nullptr, content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Headers &headers,
+                              ContentProviderWithoutLength content_provider,
+                              const std::string &content_type) {
+  return send_with_content_provider("PUT", path, headers, nullptr, 0, nullptr,
+                                    std::move(content_provider), content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Params &params) {
+  return Put(path, Headers(), params);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Headers &headers,
+                              const Params &params) {
+  auto query = detail::params_to_query_str(params);
+  return Put(path, headers, query, "application/x-www-form-urlencoded");
+}
+
+inline Result ClientImpl::Put(const std::string &path,
+                              const MultipartFormDataItems &items) {
+  return Put(path, Headers(), items);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Headers &headers,
+                              const MultipartFormDataItems &items) {
+  const auto &boundary = detail::make_multipart_data_boundary();
+  const auto &content_type =
+      detail::serialize_multipart_formdata_get_content_type(boundary);
+  const auto &body = detail::serialize_multipart_formdata(items, boundary);
+  return Put(path, headers, body, content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Headers &headers,
+                              const MultipartFormDataItems &items,
+                              const std::string &boundary) {
+  if (!detail::is_multipart_boundary_chars_valid(boundary)) {
+    return Result{nullptr, Error::UnsupportedMultipartBoundaryChars};
+  }
+
+  const auto &content_type =
+      detail::serialize_multipart_formdata_get_content_type(boundary);
+  const auto &body = detail::serialize_multipart_formdata(items, boundary);
+  return Put(path, headers, body, content_type);
+}
+
+inline Result
+ClientImpl::Put(const std::string &path, const Headers &headers,
+                const MultipartFormDataItems &items,
+                const MultipartFormDataProviderItems &provider_items) {
+  const auto &boundary = detail::make_multipart_data_boundary();
+  const auto &content_type =
+      detail::serialize_multipart_formdata_get_content_type(boundary);
+  return send_with_content_provider(
+      "PUT", path, headers, nullptr, 0, nullptr,
+      get_multipart_content_provider(boundary, items, provider_items),
+      content_type);
+}
+inline Result ClientImpl::Patch(const std::string &path) {
+  return Patch(path, std::string(), std::string());
+}
+
+inline Result ClientImpl::Patch(const std::string &path, const char *body,
+                                size_t content_length,
+                                const std::string &content_type) {
+  return Patch(path, Headers(), body, content_length, content_type);
+}
+
+inline Result ClientImpl::Patch(const std::string &path, const Headers &headers,
+                                const char *body, size_t content_length,
+                                const std::string &content_type) {
+  return send_with_content_provider("PATCH", path, headers, body,
+                                    content_length, nullptr, nullptr,
+                                    content_type);
+}
+
+inline Result ClientImpl::Patch(const std::string &path,
+                                const std::string &body,
+                                const std::string &content_type) {
+  return Patch(path, Headers(), body, content_type);
+}
+
+inline Result ClientImpl::Patch(const std::string &path, const Headers &headers,
+                                const std::string &body,
+                                const std::string &content_type) {
+  return send_with_content_provider("PATCH", path, headers, body.data(),
+                                    body.size(), nullptr, nullptr,
+                                    content_type);
+}
+
+inline Result ClientImpl::Patch(const std::string &path, size_t content_length,
+                                ContentProvider content_provider,
+                                const std::string &content_type) {
+  return Patch(path, Headers(), content_length, std::move(content_provider),
+               content_type);
+}
+
+inline Result ClientImpl::Patch(const std::string &path,
+                                ContentProviderWithoutLength content_provider,
+                                const std::string &content_type) {
+  return Patch(path, Headers(), std::move(content_provider), content_type);
+}
+
+inline Result ClientImpl::Patch(const std::string &path, const Headers &headers,
+                                size_t content_length,
+                                ContentProvider content_provider,
+                                const std::string &content_type) {
+  return send_with_content_provider("PATCH", path, headers, nullptr,
+                                    content_length, std::move(content_provider),
+                                    nullptr, content_type);
+}
+
+inline Result ClientImpl::Patch(const std::string &path, const Headers &headers,
+                                ContentProviderWithoutLength content_provider,
+                                const std::string &content_type) {
+  return send_with_content_provider("PATCH", path, headers, nullptr, 0, nullptr,
+                                    std::move(content_provider), content_type);
+}
+
+inline Result ClientImpl::Delete(const std::string &path) {
+  return Delete(path, Headers(), std::string(), std::string());
+}
+
+inline Result ClientImpl::Delete(const std::string &path,
+                                 const Headers &headers) {
+  return Delete(path, headers, std::string(), std::string());
+}
+
+inline Result ClientImpl::Delete(const std::string &path, const char *body,
+                                 size_t content_length,
+                                 const std::string &content_type) {
+  return Delete(path, Headers(), body, content_length, content_type);
+}
+
+inline Result ClientImpl::Delete(const std::string &path,
+                                 const Headers &headers, const char *body,
+                                 size_t content_length,
+                                 const std::string &content_type) {
+  Request req;
+  req.method = "DELETE";
+  req.headers = headers;
+  req.path = path;
+
+  if (!content_type.empty()) { req.set_header("Content-Type", content_type); }
+  req.body.assign(body, content_length);
+
+  return send_(std::move(req));
+}
+
+inline Result ClientImpl::Delete(const std::string &path,
+                                 const std::string &body,
+                                 const std::string &content_type) {
+  return Delete(path, Headers(), body.data(), body.size(), content_type);
+}
+
+inline Result ClientImpl::Delete(const std::string &path,
+                                 const Headers &headers,
+                                 const std::string &body,
+                                 const std::string &content_type) {
+  return Delete(path, headers, body.data(), body.size(), content_type);
+}
+
+inline Result ClientImpl::Options(const std::string &path) {
+  return Options(path, Headers());
+}
+
+inline Result ClientImpl::Options(const std::string &path,
+                                  const Headers &headers) {
+  Request req;
+  req.method = "OPTIONS";
+  req.headers = headers;
+  req.path = path;
+
+  return send_(std::move(req));
+}
+
+inline void ClientImpl::stop() {
+  std::lock_guard<std::mutex> guard(socket_mutex_);
+
+  // If there is anything ongoing right now, the ONLY thread-safe thing we can
+  // do is to shutdown_socket, so that threads using this socket suddenly
+  // discover they can't read/write any more and error out. Everything else
+  // (closing the socket, shutting ssl down) is unsafe because these actions are
+  // not thread-safe.
+  if (socket_requests_in_flight_ > 0) {
+    shutdown_socket(socket_);
+
+    // Aside from that, we set a flag for the socket to be closed when we're
+    // done.
+    socket_should_be_closed_when_request_is_done_ = true;
+    return;
+  }
+
+  // Otherwise, still holding the mutex, we can shut everything down ourselves
+  shutdown_ssl(socket_, true);
+  shutdown_socket(socket_);
+  close_socket(socket_);
+}
+
+inline std::string ClientImpl::host() const { return host_; }
+
+inline int ClientImpl::port() const { return port_; }
+
+inline size_t ClientImpl::is_socket_open() const {
+  std::lock_guard<std::mutex> guard(socket_mutex_);
+  return socket_.is_open();
+}
+
+inline socket_t ClientImpl::socket() const { return socket_.sock; }
+
+inline void ClientImpl::set_connection_timeout(time_t sec, time_t usec) {
+  connection_timeout_sec_ = sec;
+  connection_timeout_usec_ = usec;
+}
+
+inline void ClientImpl::set_read_timeout(time_t sec, time_t usec) {
+  read_timeout_sec_ = sec;
+  read_timeout_usec_ = usec;
+}
+
+inline void ClientImpl::set_write_timeout(time_t sec, time_t usec) {
+  write_timeout_sec_ = sec;
+  write_timeout_usec_ = usec;
+}
+
+inline void ClientImpl::set_basic_auth(const std::string &username,
+                                       const std::string &password) {
+  basic_auth_username_ = username;
+  basic_auth_password_ = password;
+}
+
+inline void ClientImpl::set_bearer_token_auth(const std::string &token) {
+  bearer_token_auth_token_ = token;
+}
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline void ClientImpl::set_digest_auth(const std::string &username,
+                                        const std::string &password) {
+  digest_auth_username_ = username;
+  digest_auth_password_ = password;
+}
+#endif
+
+inline void ClientImpl::set_keep_alive(bool on) { keep_alive_ = on; }
+
+inline void ClientImpl::set_follow_location(bool on) { follow_location_ = on; }
+
+inline void ClientImpl::set_url_encode(bool on) { url_encode_ = on; }
+
+inline void
+ClientImpl::set_hostname_addr_map(std::map<std::string, std::string> addr_map) {
+  addr_map_ = std::move(addr_map);
+}
+
+inline void ClientImpl::set_default_headers(Headers headers) {
+  default_headers_ = std::move(headers);
+}
+
+inline void ClientImpl::set_header_writer(
+    std::function<ssize_t(Stream &, Headers &)> const &writer) {
+  header_writer_ = writer;
+}
+
+inline void ClientImpl::set_address_family(int family) {
+  address_family_ = family;
+}
+
+inline void ClientImpl::set_tcp_nodelay(bool on) { tcp_nodelay_ = on; }
+
+inline void ClientImpl::set_socket_options(SocketOptions socket_options) {
+  socket_options_ = std::move(socket_options);
+}
+
+inline void ClientImpl::set_compress(bool on) { compress_ = on; }
+
+inline void ClientImpl::set_decompress(bool on) { decompress_ = on; }
+
+inline void ClientImpl::set_interface(const std::string &intf) {
+  interface_ = intf;
+}
+
+inline void ClientImpl::set_proxy(const std::string &host, int port) {
+  proxy_host_ = host;
+  proxy_port_ = port;
+}
+
+inline void ClientImpl::set_proxy_basic_auth(const std::string &username,
+                                             const std::string &password) {
+  proxy_basic_auth_username_ = username;
+  proxy_basic_auth_password_ = password;
+}
+
+inline void ClientImpl::set_proxy_bearer_token_auth(const std::string &token) {
+  proxy_bearer_token_auth_token_ = token;
+}
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline void ClientImpl::set_proxy_digest_auth(const std::string &username,
+                                              const std::string &password) {
+  proxy_digest_auth_username_ = username;
+  proxy_digest_auth_password_ = password;
+}
+
+inline void ClientImpl::set_ca_cert_path(const std::string &ca_cert_file_path,
+                                         const std::string &ca_cert_dir_path) {
+  ca_cert_file_path_ = ca_cert_file_path;
+  ca_cert_dir_path_ = ca_cert_dir_path;
+}
+
+inline void ClientImpl::set_ca_cert_store(X509_STORE *ca_cert_store) {
+  if (ca_cert_store && ca_cert_store != ca_cert_store_) {
+    ca_cert_store_ = ca_cert_store;
+  }
+}
+
+inline X509_STORE *ClientImpl::create_ca_cert_store(const char *ca_cert,
+                                                    std::size_t size) const {
+  auto mem = BIO_new_mem_buf(ca_cert, static_cast<int>(size));
+  if (!mem) { return nullptr; }
+
+  auto inf = PEM_X509_INFO_read_bio(mem, nullptr, nullptr, nullptr);
+  if (!inf) {
+    BIO_free_all(mem);
+    return nullptr;
+  }
+
+  auto cts = X509_STORE_new();
+  if (cts) {
+    for (auto i = 0; i < static_cast<int>(sk_X509_INFO_num(inf)); i++) {
+      auto itmp = sk_X509_INFO_value(inf, i);
+      if (!itmp) { continue; }
+
+      if (itmp->x509) { X509_STORE_add_cert(cts, itmp->x509); }
+      if (itmp->crl) { X509_STORE_add_crl(cts, itmp->crl); }
+    }
+  }
+
+  sk_X509_INFO_pop_free(inf, X509_INFO_free);
+  BIO_free_all(mem);
+  return cts;
+}
+
+inline void ClientImpl::enable_server_certificate_verification(bool enabled) {
+  server_certificate_verification_ = enabled;
+}
+#endif
+
+inline void ClientImpl::set_logger(Logger logger) {
+  logger_ = std::move(logger);
+}
+
+/*
+ * SSL Implementation
+ */
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+namespace detail {
+
+template <typename U, typename V>
+inline SSL *ssl_new(socket_t sock, SSL_CTX *ctx, std::mutex &ctx_mutex,
+                    U SSL_connect_or_accept, V setup) {
+  SSL *ssl = nullptr;
+  {
+    std::lock_guard<std::mutex> guard(ctx_mutex);
+    ssl = SSL_new(ctx);
+  }
+
+  if (ssl) {
+    set_nonblocking(sock, true);
+    auto bio = BIO_new_socket(static_cast<int>(sock), BIO_NOCLOSE);
+    BIO_set_nbio(bio, 1);
+    SSL_set_bio(ssl, bio, bio);
+
+    if (!setup(ssl) || SSL_connect_or_accept(ssl) != 1) {
+      SSL_shutdown(ssl);
+      {
+        std::lock_guard<std::mutex> guard(ctx_mutex);
+        SSL_free(ssl);
+      }
+      set_nonblocking(sock, false);
+      return nullptr;
+    }
+    BIO_set_nbio(bio, 0);
+    set_nonblocking(sock, false);
+  }
+
+  return ssl;
+}
+
+inline void ssl_delete(std::mutex &ctx_mutex, SSL *ssl,
+                       bool shutdown_gracefully) {
+  // sometimes we may want to skip this to try to avoid SIGPIPE if we know
+  // the remote has closed the network connection
+  // Note that it is not always possible to avoid SIGPIPE, this is merely a
+  // best-efforts.
+  if (shutdown_gracefully) { SSL_shutdown(ssl); }
+
+  std::lock_guard<std::mutex> guard(ctx_mutex);
+  SSL_free(ssl);
+}
+
+template <typename U>
+bool ssl_connect_or_accept_nonblocking(socket_t sock, SSL *ssl,
+                                       U ssl_connect_or_accept,
+                                       time_t timeout_sec,
+                                       time_t timeout_usec) {
+  auto res = 0;
+  while ((res = ssl_connect_or_accept(ssl)) != 1) {
+    auto err = SSL_get_error(ssl, res);
+    switch (err) {
+    case SSL_ERROR_WANT_READ:
+      if (select_read(sock, timeout_sec, timeout_usec) > 0) { continue; }
+      break;
+    case SSL_ERROR_WANT_WRITE:
+      if (select_write(sock, timeout_sec, timeout_usec) > 0) { continue; }
+      break;
+    default: break;
+    }
+    return false;
+  }
+  return true;
+}
+
+template <typename T>
+inline bool process_server_socket_ssl(
+    const std::atomic<socket_t> &svr_sock, SSL *ssl, socket_t sock,
+    size_t keep_alive_max_count, time_t keep_alive_timeout_sec,
+    time_t read_timeout_sec, time_t read_timeout_usec, time_t write_timeout_sec,
+    time_t write_timeout_usec, T callback) {
+  return process_server_socket_core(
+      svr_sock, sock, keep_alive_max_count, keep_alive_timeout_sec,
+      [&](bool close_connection, bool &connection_closed) {
+        SSLSocketStream strm(sock, ssl, read_timeout_sec, read_timeout_usec,
+                             write_timeout_sec, write_timeout_usec);
+        return callback(strm, close_connection, connection_closed);
+      });
+}
+
+template <typename T>
+inline bool
+process_client_socket_ssl(SSL *ssl, socket_t sock, time_t read_timeout_sec,
+                          time_t read_timeout_usec, time_t write_timeout_sec,
+                          time_t write_timeout_usec, T callback) {
+  SSLSocketStream strm(sock, ssl, read_timeout_sec, read_timeout_usec,
+                       write_timeout_sec, write_timeout_usec);
+  return callback(strm);
+}
+
+class SSLInit {
+public:
+  SSLInit() {
+    OPENSSL_init_ssl(
+        OPENSSL_INIT_LOAD_SSL_STRINGS | OPENSSL_INIT_LOAD_CRYPTO_STRINGS, NULL);
+  }
+};
+
+// SSL socket stream implementation
+inline SSLSocketStream::SSLSocketStream(socket_t sock, SSL *ssl,
+                                        time_t read_timeout_sec,
+                                        time_t read_timeout_usec,
+                                        time_t write_timeout_sec,
+                                        time_t write_timeout_usec)
+    : sock_(sock), ssl_(ssl), read_timeout_sec_(read_timeout_sec),
+      read_timeout_usec_(read_timeout_usec),
+      write_timeout_sec_(write_timeout_sec),
+      write_timeout_usec_(write_timeout_usec) {
+  SSL_clear_mode(ssl, SSL_MODE_AUTO_RETRY);
+}
+
+inline SSLSocketStream::~SSLSocketStream() = default;
+
+inline bool SSLSocketStream::is_readable() const {
+  return detail::select_read(sock_, read_timeout_sec_, read_timeout_usec_) > 0;
+}
+
+inline bool SSLSocketStream::is_writable() const {
+  return select_write(sock_, write_timeout_sec_, write_timeout_usec_) > 0 &&
+         is_socket_alive(sock_);
+}
+
+inline ssize_t SSLSocketStream::read(char *ptr, size_t size) {
+  if (SSL_pending(ssl_) > 0) {
+    return SSL_read(ssl_, ptr, static_cast<int>(size));
+  } else if (is_readable()) {
+    auto ret = SSL_read(ssl_, ptr, static_cast<int>(size));
+    if (ret < 0) {
+      auto err = SSL_get_error(ssl_, ret);
+      auto n = 1000;
+#ifdef _WIN32
+      while (--n >= 0 && (err == SSL_ERROR_WANT_READ ||
+                          (err == SSL_ERROR_SYSCALL &&
+                           WSAGetLastError() == WSAETIMEDOUT))) {
+#else
+      while (--n >= 0 && err == SSL_ERROR_WANT_READ) {
+#endif
+        if (SSL_pending(ssl_) > 0) {
+          return SSL_read(ssl_, ptr, static_cast<int>(size));
+        } else if (is_readable()) {
+          std::this_thread::sleep_for(std::chrono::milliseconds(1));
+          ret = SSL_read(ssl_, ptr, static_cast<int>(size));
+          if (ret >= 0) { return ret; }
+          err = SSL_get_error(ssl_, ret);
+        } else {
+          return -1;
+        }
+      }
+    }
+    return ret;
+  }
+  return -1;
+}
+
+inline ssize_t SSLSocketStream::write(const char *ptr, size_t size) {
+  if (is_writable()) {
+    auto handle_size = static_cast<int>(
+        std::min<size_t>(size, (std::numeric_limits<int>::max)()));
+
+    auto ret = SSL_write(ssl_, ptr, static_cast<int>(handle_size));
+    if (ret < 0) {
+      auto err = SSL_get_error(ssl_, ret);
+      auto n = 1000;
+#ifdef _WIN32
+      while (--n >= 0 && (err == SSL_ERROR_WANT_WRITE ||
+                          (err == SSL_ERROR_SYSCALL &&
+                           WSAGetLastError() == WSAETIMEDOUT))) {
+#else
+      while (--n >= 0 && err == SSL_ERROR_WANT_WRITE) {
+#endif
+        if (is_writable()) {
+          std::this_thread::sleep_for(std::chrono::milliseconds(1));
+          ret = SSL_write(ssl_, ptr, static_cast<int>(handle_size));
+          if (ret >= 0) { return ret; }
+          err = SSL_get_error(ssl_, ret);
+        } else {
+          return -1;
+        }
+      }
+    }
+    return ret;
+  }
+  return -1;
+}
+
+inline void SSLSocketStream::get_remote_ip_and_port(std::string &ip,
+                                                    int &port) const {
+  detail::get_remote_ip_and_port(sock_, ip, port);
+}
+
+inline void SSLSocketStream::get_local_ip_and_port(std::string &ip,
+                                                   int &port) const {
+  detail::get_local_ip_and_port(sock_, ip, port);
+}
+
+inline socket_t SSLSocketStream::socket() const { return sock_; }
+
+static SSLInit sslinit_;
+
+} // namespace detail
+
+// SSL HTTP server implementation
+inline SSLServer::SSLServer(const char *cert_path, const char *private_key_path,
+                            const char *client_ca_cert_file_path,
+                            const char *client_ca_cert_dir_path,
+                            const char *private_key_password) {
+  ctx_ = SSL_CTX_new(TLS_server_method());
+
+  if (ctx_) {
+    SSL_CTX_set_options(ctx_,
+                        SSL_OP_NO_COMPRESSION |
+                            SSL_OP_NO_SESSION_RESUMPTION_ON_RENEGOTIATION);
+
+    SSL_CTX_set_min_proto_version(ctx_, TLS1_1_VERSION);
+
+    if (private_key_password != nullptr && (private_key_password[0] != '\0')) {
+      SSL_CTX_set_default_passwd_cb_userdata(
+          ctx_,
+          reinterpret_cast<void *>(const_cast<char *>(private_key_password)));
+    }
+
+    if (SSL_CTX_use_certificate_chain_file(ctx_, cert_path) != 1 ||
+        SSL_CTX_use_PrivateKey_file(ctx_, private_key_path, SSL_FILETYPE_PEM) !=
+            1) {
+      SSL_CTX_free(ctx_);
+      ctx_ = nullptr;
+    } else if (client_ca_cert_file_path || client_ca_cert_dir_path) {
+      SSL_CTX_load_verify_locations(ctx_, client_ca_cert_file_path,
+                                    client_ca_cert_dir_path);
+
+      SSL_CTX_set_verify(
+          ctx_, SSL_VERIFY_PEER | SSL_VERIFY_FAIL_IF_NO_PEER_CERT, nullptr);
+    }
+  }
+}
+
+inline SSLServer::SSLServer(X509 *cert, EVP_PKEY *private_key,
+                            X509_STORE *client_ca_cert_store) {
+  ctx_ = SSL_CTX_new(TLS_server_method());
+
+  if (ctx_) {
+    SSL_CTX_set_options(ctx_,
+                        SSL_OP_NO_COMPRESSION |
+                            SSL_OP_NO_SESSION_RESUMPTION_ON_RENEGOTIATION);
+
+    SSL_CTX_set_min_proto_version(ctx_, TLS1_1_VERSION);
+
+    if (SSL_CTX_use_certificate(ctx_, cert) != 1 ||
+        SSL_CTX_use_PrivateKey(ctx_, private_key) != 1) {
+      SSL_CTX_free(ctx_);
+      ctx_ = nullptr;
+    } else if (client_ca_cert_store) {
+      SSL_CTX_set_cert_store(ctx_, client_ca_cert_store);
+
+      SSL_CTX_set_verify(
+          ctx_, SSL_VERIFY_PEER | SSL_VERIFY_FAIL_IF_NO_PEER_CERT, nullptr);
+    }
+  }
+}
+
+inline SSLServer::SSLServer(
+    const std::function<bool(SSL_CTX &ssl_ctx)> &setup_ssl_ctx_callback) {
+  ctx_ = SSL_CTX_new(TLS_method());
+  if (ctx_) {
+    if (!setup_ssl_ctx_callback(*ctx_)) {
+      SSL_CTX_free(ctx_);
+      ctx_ = nullptr;
+    }
+  }
+}
+
+inline SSLServer::~SSLServer() {
+  if (ctx_) { SSL_CTX_free(ctx_); }
+}
+
+inline bool SSLServer::is_valid() const { return ctx_; }
+
+inline SSL_CTX *SSLServer::ssl_context() const { return ctx_; }
+
+inline bool SSLServer::process_and_close_socket(socket_t sock) {
+  auto ssl = detail::ssl_new(
+      sock, ctx_, ctx_mutex_,
+      [&](SSL *ssl2) {
+        return detail::ssl_connect_or_accept_nonblocking(
+            sock, ssl2, SSL_accept, read_timeout_sec_, read_timeout_usec_);
+      },
+      [](SSL * /*ssl2*/) { return true; });
+
+  auto ret = false;
+  if (ssl) {
+    ret = detail::process_server_socket_ssl(
+        svr_sock_, ssl, sock, keep_alive_max_count_, keep_alive_timeout_sec_,
+        read_timeout_sec_, read_timeout_usec_, write_timeout_sec_,
+        write_timeout_usec_,
+        [this, ssl](Stream &strm, bool close_connection,
+                    bool &connection_closed) {
+          return process_request(strm, close_connection, connection_closed,
+                                 [&](Request &req) { req.ssl = ssl; });
+        });
+
+    // Shutdown gracefully if the result seemed successful, non-gracefully if
+    // the connection appeared to be closed.
+    const bool shutdown_gracefully = ret;
+    detail::ssl_delete(ctx_mutex_, ssl, shutdown_gracefully);
+  }
+
+  detail::shutdown_socket(sock);
+  detail::close_socket(sock);
+  return ret;
+}
+
+// SSL HTTP client implementation
+inline SSLClient::SSLClient(const std::string &host)
+    : SSLClient(host, 443, std::string(), std::string()) {}
+
+inline SSLClient::SSLClient(const std::string &host, int port)
+    : SSLClient(host, port, std::string(), std::string()) {}
+
+inline SSLClient::SSLClient(const std::string &host, int port,
+                            const std::string &client_cert_path,
+                            const std::string &client_key_path,
+                            const std::string &private_key_password)
+    : ClientImpl(host, port, client_cert_path, client_key_path) {
+  ctx_ = SSL_CTX_new(TLS_client_method());
+
+  detail::split(&host_[0], &host_[host_.size()], '.',
+                [&](const char *b, const char *e) {
+                  host_components_.emplace_back(b, e);
+                });
+
+  if (!client_cert_path.empty() && !client_key_path.empty()) {
+    if (!private_key_password.empty()) {
+      SSL_CTX_set_default_passwd_cb_userdata(
+          ctx_, reinterpret_cast<void *>(
+                    const_cast<char *>(private_key_password.c_str())));
+    }
+
+    if (SSL_CTX_use_certificate_file(ctx_, client_cert_path.c_str(),
+                                     SSL_FILETYPE_PEM) != 1 ||
+        SSL_CTX_use_PrivateKey_file(ctx_, client_key_path.c_str(),
+                                    SSL_FILETYPE_PEM) != 1) {
+      SSL_CTX_free(ctx_);
+      ctx_ = nullptr;
+    }
+  }
+}
+
+inline SSLClient::SSLClient(const std::string &host, int port,
+                            X509 *client_cert, EVP_PKEY *client_key,
+                            const std::string &private_key_password)
+    : ClientImpl(host, port) {
+  ctx_ = SSL_CTX_new(TLS_client_method());
+
+  detail::split(&host_[0], &host_[host_.size()], '.',
+                [&](const char *b, const char *e) {
+                  host_components_.emplace_back(b, e);
+                });
+
+  if (client_cert != nullptr && client_key != nullptr) {
+    if (!private_key_password.empty()) {
+      SSL_CTX_set_default_passwd_cb_userdata(
+          ctx_, reinterpret_cast<void *>(
+                    const_cast<char *>(private_key_password.c_str())));
+    }
+
+    if (SSL_CTX_use_certificate(ctx_, client_cert) != 1 ||
+        SSL_CTX_use_PrivateKey(ctx_, client_key) != 1) {
+      SSL_CTX_free(ctx_);
+      ctx_ = nullptr;
+    }
+  }
+}
+
+inline SSLClient::~SSLClient() {
+  if (ctx_) { SSL_CTX_free(ctx_); }
+  // Make sure to shut down SSL since shutdown_ssl will resolve to the
+  // base function rather than the derived function once we get to the
+  // base class destructor, and won't free the SSL (causing a leak).
+  shutdown_ssl_impl(socket_, true);
+}
+
+inline bool SSLClient::is_valid() const { return ctx_; }
+
+inline void SSLClient::set_ca_cert_store(X509_STORE *ca_cert_store) {
+  if (ca_cert_store) {
+    if (ctx_) {
+      if (SSL_CTX_get_cert_store(ctx_) != ca_cert_store) {
+        // Free memory allocated for old cert and use new store `ca_cert_store`
+        SSL_CTX_set_cert_store(ctx_, ca_cert_store);
+      }
+    } else {
+      X509_STORE_free(ca_cert_store);
+    }
+  }
+}
+
+inline void SSLClient::load_ca_cert_store(const char *ca_cert,
+                                          std::size_t size) {
+  set_ca_cert_store(ClientImpl::create_ca_cert_store(ca_cert, size));
+}
+
+inline long SSLClient::get_openssl_verify_result() const {
+  return verify_result_;
+}
+
+inline SSL_CTX *SSLClient::ssl_context() const { return ctx_; }
+
+inline bool SSLClient::create_and_connect_socket(Socket &socket, Error &error) {
+  return is_valid() && ClientImpl::create_and_connect_socket(socket, error);
+}
+
+// Assumes that socket_mutex_ is locked and that there are no requests in flight
+inline bool SSLClient::connect_with_proxy(Socket &socket, Response &res,
+                                          bool &success, Error &error) {
+  success = true;
+  Response proxy_res;
+  if (!detail::process_client_socket(
+          socket.sock, read_timeout_sec_, read_timeout_usec_,
+          write_timeout_sec_, write_timeout_usec_, [&](Stream &strm) {
+            Request req2;
+            req2.method = "CONNECT";
+            req2.path = host_and_port_;
+            return process_request(strm, req2, proxy_res, false, error);
+          })) {
+    // Thread-safe to close everything because we are assuming there are no
+    // requests in flight
+    shutdown_ssl(socket, true);
+    shutdown_socket(socket);
+    close_socket(socket);
+    success = false;
+    return false;
+  }
+
+  if (proxy_res.status == StatusCode::ProxyAuthenticationRequired_407) {
+    if (!proxy_digest_auth_username_.empty() &&
+        !proxy_digest_auth_password_.empty()) {
+      std::map<std::string, std::string> auth;
+      if (detail::parse_www_authenticate(proxy_res, auth, true)) {
+        proxy_res = Response();
+        if (!detail::process_client_socket(
+                socket.sock, read_timeout_sec_, read_timeout_usec_,
+                write_timeout_sec_, write_timeout_usec_, [&](Stream &strm) {
+                  Request req3;
+                  req3.method = "CONNECT";
+                  req3.path = host_and_port_;
+                  req3.headers.insert(detail::make_digest_authentication_header(
+                      req3, auth, 1, detail::random_string(10),
+                      proxy_digest_auth_username_, proxy_digest_auth_password_,
+                      true));
+                  return process_request(strm, req3, proxy_res, false, error);
+                })) {
+          // Thread-safe to close everything because we are assuming there are
+          // no requests in flight
+          shutdown_ssl(socket, true);
+          shutdown_socket(socket);
+          close_socket(socket);
+          success = false;
+          return false;
+        }
+      }
+    }
+  }
+
+  // If status code is not 200, proxy request is failed.
+  // Set error to ProxyConnection and return proxy response
+  // as the response of the request
+  if (proxy_res.status != StatusCode::OK_200) {
+    error = Error::ProxyConnection;
+    res = std::move(proxy_res);
+    // Thread-safe to close everything because we are assuming there are
+    // no requests in flight
+    shutdown_ssl(socket, true);
+    shutdown_socket(socket);
+    close_socket(socket);
+    return false;
+  }
+
+  return true;
+}
+
+inline bool SSLClient::load_certs() {
+  auto ret = true;
+
+  std::call_once(initialize_cert_, [&]() {
+    std::lock_guard<std::mutex> guard(ctx_mutex_);
+    if (!ca_cert_file_path_.empty()) {
+      if (!SSL_CTX_load_verify_locations(ctx_, ca_cert_file_path_.c_str(),
+                                         nullptr)) {
+        ret = false;
+      }
+    } else if (!ca_cert_dir_path_.empty()) {
+      if (!SSL_CTX_load_verify_locations(ctx_, nullptr,
+                                         ca_cert_dir_path_.c_str())) {
+        ret = false;
+      }
+    } else {
+      auto loaded = false;
+#ifdef _WIN32
+      loaded =
+          detail::load_system_certs_on_windows(SSL_CTX_get_cert_store(ctx_));
+#elif defined(CPPHTTPLIB_USE_CERTS_FROM_MACOSX_KEYCHAIN) && defined(__APPLE__)
+#if TARGET_OS_OSX
+      loaded = detail::load_system_certs_on_macos(SSL_CTX_get_cert_store(ctx_));
+#endif // TARGET_OS_OSX
+#endif // _WIN32
+      if (!loaded) { SSL_CTX_set_default_verify_paths(ctx_); }
+    }
+  });
+
+  return ret;
+}
+
+inline bool SSLClient::initialize_ssl(Socket &socket, Error &error) {
+  auto ssl = detail::ssl_new(
+      socket.sock, ctx_, ctx_mutex_,
+      [&](SSL *ssl2) {
+        if (server_certificate_verification_) {
+          if (!load_certs()) {
+            error = Error::SSLLoadingCerts;
+            return false;
+          }
+          SSL_set_verify(ssl2, SSL_VERIFY_NONE, nullptr);
+        }
+
+        if (!detail::ssl_connect_or_accept_nonblocking(
+                socket.sock, ssl2, SSL_connect, connection_timeout_sec_,
+                connection_timeout_usec_)) {
+          error = Error::SSLConnection;
+          return false;
+        }
+
+        if (server_certificate_verification_) {
+          verify_result_ = SSL_get_verify_result(ssl2);
+
+          if (verify_result_ != X509_V_OK) {
+            error = Error::SSLServerVerification;
+            return false;
+          }
+
+          auto server_cert = SSL_get1_peer_certificate(ssl2);
+
+          if (server_cert == nullptr) {
+            error = Error::SSLServerVerification;
+            return false;
+          }
+
+          if (!verify_host(server_cert)) {
+            X509_free(server_cert);
+            error = Error::SSLServerVerification;
+            return false;
+          }
+          X509_free(server_cert);
+        }
+
+        return true;
+      },
+      [&](SSL *ssl2) {
+        // NOTE: Direct call instead of using the OpenSSL macro to suppress
+        // -Wold-style-cast warning
+        // SSL_set_tlsext_host_name(ssl2, host_.c_str());
+        SSL_ctrl(ssl2, SSL_CTRL_SET_TLSEXT_HOSTNAME, TLSEXT_NAMETYPE_host_name,
+                 static_cast<void *>(const_cast<char *>(host_.c_str())));
+        return true;
+      });
+
+  if (ssl) {
+    socket.ssl = ssl;
+    return true;
+  }
+
+  shutdown_socket(socket);
+  close_socket(socket);
+  return false;
+}
+
+inline void SSLClient::shutdown_ssl(Socket &socket, bool shutdown_gracefully) {
+  shutdown_ssl_impl(socket, shutdown_gracefully);
+}
+
+inline void SSLClient::shutdown_ssl_impl(Socket &socket,
+                                         bool shutdown_gracefully) {
+  if (socket.sock == INVALID_SOCKET) {
+    assert(socket.ssl == nullptr);
+    return;
+  }
+  if (socket.ssl) {
+    detail::ssl_delete(ctx_mutex_, socket.ssl, shutdown_gracefully);
+    socket.ssl = nullptr;
+  }
+  assert(socket.ssl == nullptr);
+}
+
+inline bool
+SSLClient::process_socket(const Socket &socket,
+                          std::function<bool(Stream &strm)> callback) {
+  assert(socket.ssl);
+  return detail::process_client_socket_ssl(
+      socket.ssl, socket.sock, read_timeout_sec_, read_timeout_usec_,
+      write_timeout_sec_, write_timeout_usec_, std::move(callback));
+}
+
+inline bool SSLClient::is_ssl() const { return true; }
+
+inline bool SSLClient::verify_host(X509 *server_cert) const {
+  /* Quote from RFC2818 section 3.1 "Server Identity"
+
+     If a subjectAltName extension of type dNSName is present, that MUST
+     be used as the identity. Otherwise, the (most specific) Common Name
+     field in the Subject field of the certificate MUST be used. Although
+     the use of the Common Name is existing practice, it is deprecated and
+     Certification Authorities are encouraged to use the dNSName instead.
+
+     Matching is performed using the matching rules specified by
+     [RFC2459].  If more than one identity of a given type is present in
+     the certificate (e.g., more than one dNSName name, a match in any one
+     of the set is considered acceptable.) Names may contain the wildcard
+     character * which is considered to match any single domain name
+     component or component fragment. E.g., *.a.com matches foo.a.com but
+     not bar.foo.a.com. f*.com matches foo.com but not bar.com.
+
+     In some cases, the URI is specified as an IP address rather than a
+     hostname. In this case, the iPAddress subjectAltName must be present
+     in the certificate and must exactly match the IP in the URI.
+
+  */
+  return verify_host_with_subject_alt_name(server_cert) ||
+         verify_host_with_common_name(server_cert);
+}
+
+inline bool
+SSLClient::verify_host_with_subject_alt_name(X509 *server_cert) const {
+  auto ret = false;
+
+  auto type = GEN_DNS;
+
+  struct in6_addr addr6 {};
+  struct in_addr addr {};
+  size_t addr_len = 0;
+
+#ifndef __MINGW32__
+  if (inet_pton(AF_INET6, host_.c_str(), &addr6)) {
+    type = GEN_IPADD;
+    addr_len = sizeof(struct in6_addr);
+  } else if (inet_pton(AF_INET, host_.c_str(), &addr)) {
+    type = GEN_IPADD;
+    addr_len = sizeof(struct in_addr);
+  }
+#endif
+
+  auto alt_names = static_cast<const struct stack_st_GENERAL_NAME *>(
+      X509_get_ext_d2i(server_cert, NID_subject_alt_name, nullptr, nullptr));
+
+  if (alt_names) {
+    auto dsn_matched = false;
+    auto ip_matched = false;
+
+    auto count = sk_GENERAL_NAME_num(alt_names);
+
+    for (decltype(count) i = 0; i < count && !dsn_matched; i++) {
+      auto val = sk_GENERAL_NAME_value(alt_names, i);
+      if (val->type == type) {
+        auto name =
+            reinterpret_cast<const char *>(ASN1_STRING_get0_data(val->d.ia5));
+        auto name_len = static_cast<size_t>(ASN1_STRING_length(val->d.ia5));
+
+        switch (type) {
+        case GEN_DNS: dsn_matched = check_host_name(name, name_len); break;
+
+        case GEN_IPADD:
+          if (!memcmp(&addr6, name, addr_len) ||
+              !memcmp(&addr, name, addr_len)) {
+            ip_matched = true;
+          }
+          break;
+        }
+      }
+    }
+
+    if (dsn_matched || ip_matched) { ret = true; }
+  }
+
+  GENERAL_NAMES_free(const_cast<STACK_OF(GENERAL_NAME) *>(
+      reinterpret_cast<const STACK_OF(GENERAL_NAME) *>(alt_names)));
+  return ret;
+}
+
+inline bool SSLClient::verify_host_with_common_name(X509 *server_cert) const {
+  const auto subject_name = X509_get_subject_name(server_cert);
+
+  if (subject_name != nullptr) {
+    char name[BUFSIZ];
+    auto name_len = X509_NAME_get_text_by_NID(subject_name, NID_commonName,
+                                              name, sizeof(name));
+
+    if (name_len != -1) {
+      return check_host_name(name, static_cast<size_t>(name_len));
+    }
+  }
+
+  return false;
+}
+
+inline bool SSLClient::check_host_name(const char *pattern,
+                                       size_t pattern_len) const {
+  if (host_.size() == pattern_len && host_ == pattern) { return true; }
+
+  // Wildcard match
+  // https://bugs.launchpad.net/ubuntu/+source/firefox-3.0/+bug/376484
+  std::vector<std::string> pattern_components;
+  detail::split(&pattern[0], &pattern[pattern_len], '.',
+                [&](const char *b, const char *e) {
+                  pattern_components.emplace_back(b, e);
+                });
+
+  if (host_components_.size() != pattern_components.size()) { return false; }
+
+  auto itr = pattern_components.begin();
+  for (const auto &h : host_components_) {
+    auto &p = *itr;
+    if (p != h && p != "*") {
+      auto partial_match = (p.size() > 0 && p[p.size() - 1] == '*' &&
+                            !p.compare(0, p.size() - 1, h));
+      if (!partial_match) { return false; }
+    }
+    ++itr;
+  }
+
+  return true;
+}
+#endif
+
+// Universal client implementation
+inline Client::Client(const std::string &scheme_host_port)
+    : Client(scheme_host_port, std::string(), std::string()) {}
+
+inline Client::Client(const std::string &scheme_host_port,
+                      const std::string &client_cert_path,
+                      const std::string &client_key_path) {
+  const static std::regex re(
+      R"((?:([a-z]+):\/\/)?(?:\[([\d:]+)\]|([^:/?#]+))(?::(\d+))?)");
+
+  std::smatch m;
+  if (std::regex_match(scheme_host_port, m, re)) {
+    auto scheme = m[1].str();
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+    if (!scheme.empty() && (scheme != "http" && scheme != "https")) {
+#else
+    if (!scheme.empty() && scheme != "http") {
+#endif
+#ifndef CPPHTTPLIB_NO_EXCEPTIONS
+      std::string msg = "'" + scheme + "' scheme is not supported.";
+      throw std::invalid_argument(msg);
+#endif
+      return;
+    }
+
+    auto is_ssl = scheme == "https";
+
+    auto host = m[2].str();
+    if (host.empty()) { host = m[3].str(); }
+
+    auto port_str = m[4].str();
+    auto port = !port_str.empty() ? std::stoi(port_str) : (is_ssl ? 443 : 80);
+
+    if (is_ssl) {
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+      cli_ = detail::make_unique<SSLClient>(host, port, client_cert_path,
+                                            client_key_path);
+      is_ssl_ = is_ssl;
+#endif
+    } else {
+      cli_ = detail::make_unique<ClientImpl>(host, port, client_cert_path,
+                                             client_key_path);
+    }
+  } else {
+    cli_ = detail::make_unique<ClientImpl>(scheme_host_port, 80,
+                                           client_cert_path, client_key_path);
+  }
+}
+
+inline Client::Client(const std::string &host, int port)
+    : cli_(detail::make_unique<ClientImpl>(host, port)) {}
+
+inline Client::Client(const std::string &host, int port,
+                      const std::string &client_cert_path,
+                      const std::string &client_key_path)
+    : cli_(detail::make_unique<ClientImpl>(host, port, client_cert_path,
+                                           client_key_path)) {}
+
+inline Client::~Client() = default;
+
+inline bool Client::is_valid() const {
+  return cli_ != nullptr && cli_->is_valid();
+}
+
+inline Result Client::Get(const std::string &path) { return cli_->Get(path); }
+inline Result Client::Get(const std::string &path, const Headers &headers) {
+  return cli_->Get(path, headers);
+}
+inline Result Client::Get(const std::string &path, Progress progress) {
+  return cli_->Get(path, std::move(progress));
+}
+inline Result Client::Get(const std::string &path, const Headers &headers,
+                          Progress progress) {
+  return cli_->Get(path, headers, std::move(progress));
+}
+inline Result Client::Get(const std::string &path,
+                          ContentReceiver content_receiver) {
+  return cli_->Get(path, std::move(content_receiver));
+}
+inline Result Client::Get(const std::string &path, const Headers &headers,
+                          ContentReceiver content_receiver) {
+  return cli_->Get(path, headers, std::move(content_receiver));
+}
+inline Result Client::Get(const std::string &path,
+                          ContentReceiver content_receiver, Progress progress) {
+  return cli_->Get(path, std::move(content_receiver), std::move(progress));
+}
+inline Result Client::Get(const std::string &path, const Headers &headers,
+                          ContentReceiver content_receiver, Progress progress) {
+  return cli_->Get(path, headers, std::move(content_receiver),
+                   std::move(progress));
+}
+inline Result Client::Get(const std::string &path,
+                          ResponseHandler response_handler,
+                          ContentReceiver content_receiver) {
+  return cli_->Get(path, std::move(response_handler),
+                   std::move(content_receiver));
+}
+inline Result Client::Get(const std::string &path, const Headers &headers,
+                          ResponseHandler response_handler,
+                          ContentReceiver content_receiver) {
+  return cli_->Get(path, headers, std::move(response_handler),
+                   std::move(content_receiver));
+}
+inline Result Client::Get(const std::string &path,
+                          ResponseHandler response_handler,
+                          ContentReceiver content_receiver, Progress progress) {
+  return cli_->Get(path, std::move(response_handler),
+                   std::move(content_receiver), std::move(progress));
+}
+inline Result Client::Get(const std::string &path, const Headers &headers,
+                          ResponseHandler response_handler,
+                          ContentReceiver content_receiver, Progress progress) {
+  return cli_->Get(path, headers, std::move(response_handler),
+                   std::move(content_receiver), std::move(progress));
+}
+inline Result Client::Get(const std::string &path, const Params &params,
+                          const Headers &headers, Progress progress) {
+  return cli_->Get(path, params, headers, std::move(progress));
+}
+inline Result Client::Get(const std::string &path, const Params &params,
+                          const Headers &headers,
+                          ContentReceiver content_receiver, Progress progress) {
+  return cli_->Get(path, params, headers, std::move(content_receiver),
+                   std::move(progress));
+}
+inline Result Client::Get(const std::string &path, const Params &params,
+                          const Headers &headers,
+                          ResponseHandler response_handler,
+                          ContentReceiver content_receiver, Progress progress) {
+  return cli_->Get(path, params, headers, std::move(response_handler),
+                   std::move(content_receiver), std::move(progress));
+}
+
+inline Result Client::Head(const std::string &path) { return cli_->Head(path); }
+inline Result Client::Head(const std::string &path, const Headers &headers) {
+  return cli_->Head(path, headers);
+}
+
+inline Result Client::Post(const std::string &path) { return cli_->Post(path); }
+inline Result Client::Post(const std::string &path, const Headers &headers) {
+  return cli_->Post(path, headers);
+}
+inline Result Client::Post(const std::string &path, const char *body,
+                           size_t content_length,
+                           const std::string &content_type) {
+  return cli_->Post(path, body, content_length, content_type);
+}
+inline Result Client::Post(const std::string &path, const Headers &headers,
+                           const char *body, size_t content_length,
+                           const std::string &content_type) {
+  return cli_->Post(path, headers, body, content_length, content_type);
+}
+inline Result Client::Post(const std::string &path, const std::string &body,
+                           const std::string &content_type) {
+  return cli_->Post(path, body, content_type);
+}
+inline Result Client::Post(const std::string &path, const Headers &headers,
+                           const std::string &body,
+                           const std::string &content_type) {
+  return cli_->Post(path, headers, body, content_type);
+}
+inline Result Client::Post(const std::string &path, size_t content_length,
+                           ContentProvider content_provider,
+                           const std::string &content_type) {
+  return cli_->Post(path, content_length, std::move(content_provider),
+                    content_type);
+}
+inline Result Client::Post(const std::string &path,
+                           ContentProviderWithoutLength content_provider,
+                           const std::string &content_type) {
+  return cli_->Post(path, std::move(content_provider), content_type);
+}
+inline Result Client::Post(const std::string &path, const Headers &headers,
+                           size_t content_length,
+                           ContentProvider content_provider,
+                           const std::string &content_type) {
+  return cli_->Post(path, headers, content_length, std::move(content_provider),
+                    content_type);
+}
+inline Result Client::Post(const std::string &path, const Headers &headers,
+                           ContentProviderWithoutLength content_provider,
+                           const std::string &content_type) {
+  return cli_->Post(path, headers, std::move(content_provider), content_type);
+}
+inline Result Client::Post(const std::string &path, const Params &params) {
+  return cli_->Post(path, params);
+}
+inline Result Client::Post(const std::string &path, const Headers &headers,
+                           const Params &params) {
+  return cli_->Post(path, headers, params);
+}
+inline Result Client::Post(const std::string &path,
+                           const MultipartFormDataItems &items) {
+  return cli_->Post(path, items);
+}
+inline Result Client::Post(const std::string &path, const Headers &headers,
+                           const MultipartFormDataItems &items) {
+  return cli_->Post(path, headers, items);
+}
+inline Result Client::Post(const std::string &path, const Headers &headers,
+                           const MultipartFormDataItems &items,
+                           const std::string &boundary) {
+  return cli_->Post(path, headers, items, boundary);
+}
+inline Result
+Client::Post(const std::string &path, const Headers &headers,
+             const MultipartFormDataItems &items,
+             const MultipartFormDataProviderItems &provider_items) {
+  return cli_->Post(path, headers, items, provider_items);
+}
+inline Result Client::Put(const std::string &path) { return cli_->Put(path); }
+inline Result Client::Put(const std::string &path, const char *body,
+                          size_t content_length,
+                          const std::string &content_type) {
+  return cli_->Put(path, body, content_length, content_type);
+}
+inline Result Client::Put(const std::string &path, const Headers &headers,
+                          const char *body, size_t content_length,
+                          const std::string &content_type) {
+  return cli_->Put(path, headers, body, content_length, content_type);
+}
+inline Result Client::Put(const std::string &path, const std::string &body,
+                          const std::string &content_type) {
+  return cli_->Put(path, body, content_type);
+}
+inline Result Client::Put(const std::string &path, const Headers &headers,
+                          const std::string &body,
+                          const std::string &content_type) {
+  return cli_->Put(path, headers, body, content_type);
+}
+inline Result Client::Put(const std::string &path, size_t content_length,
+                          ContentProvider content_provider,
+                          const std::string &content_type) {
+  return cli_->Put(path, content_length, std::move(content_provider),
+                   content_type);
+}
+inline Result Client::Put(const std::string &path,
+                          ContentProviderWithoutLength content_provider,
+                          const std::string &content_type) {
+  return cli_->Put(path, std::move(content_provider), content_type);
+}
+inline Result Client::Put(const std::string &path, const Headers &headers,
+                          size_t content_length,
+                          ContentProvider content_provider,
+                          const std::string &content_type) {
+  return cli_->Put(path, headers, content_length, std::move(content_provider),
+                   content_type);
+}
+inline Result Client::Put(const std::string &path, const Headers &headers,
+                          ContentProviderWithoutLength content_provider,
+                          const std::string &content_type) {
+  return cli_->Put(path, headers, std::move(content_provider), content_type);
+}
+inline Result Client::Put(const std::string &path, const Params &params) {
+  return cli_->Put(path, params);
+}
+inline Result Client::Put(const std::string &path, const Headers &headers,
+                          const Params &params) {
+  return cli_->Put(path, headers, params);
+}
+inline Result Client::Put(const std::string &path,
+                          const MultipartFormDataItems &items) {
+  return cli_->Put(path, items);
+}
+inline Result Client::Put(const std::string &path, const Headers &headers,
+                          const MultipartFormDataItems &items) {
+  return cli_->Put(path, headers, items);
+}
+inline Result Client::Put(const std::string &path, const Headers &headers,
+                          const MultipartFormDataItems &items,
+                          const std::string &boundary) {
+  return cli_->Put(path, headers, items, boundary);
+}
+inline Result
+Client::Put(const std::string &path, const Headers &headers,
+            const MultipartFormDataItems &items,
+            const MultipartFormDataProviderItems &provider_items) {
+  return cli_->Put(path, headers, items, provider_items);
+}
+inline Result Client::Patch(const std::string &path) {
+  return cli_->Patch(path);
+}
+inline Result Client::Patch(const std::string &path, const char *body,
+                            size_t content_length,
+                            const std::string &content_type) {
+  return cli_->Patch(path, body, content_length, content_type);
+}
+inline Result Client::Patch(const std::string &path, const Headers &headers,
+                            const char *body, size_t content_length,
+                            const std::string &content_type) {
+  return cli_->Patch(path, headers, body, content_length, content_type);
+}
+inline Result Client::Patch(const std::string &path, const std::string &body,
+                            const std::string &content_type) {
+  return cli_->Patch(path, body, content_type);
+}
+inline Result Client::Patch(const std::string &path, const Headers &headers,
+                            const std::string &body,
+                            const std::string &content_type) {
+  return cli_->Patch(path, headers, body, content_type);
+}
+inline Result Client::Patch(const std::string &path, size_t content_length,
+                            ContentProvider content_provider,
+                            const std::string &content_type) {
+  return cli_->Patch(path, content_length, std::move(content_provider),
+                     content_type);
+}
+inline Result Client::Patch(const std::string &path,
+                            ContentProviderWithoutLength content_provider,
+                            const std::string &content_type) {
+  return cli_->Patch(path, std::move(content_provider), content_type);
+}
+inline Result Client::Patch(const std::string &path, const Headers &headers,
+                            size_t content_length,
+                            ContentProvider content_provider,
+                            const std::string &content_type) {
+  return cli_->Patch(path, headers, content_length, std::move(content_provider),
+                     content_type);
+}
+inline Result Client::Patch(const std::string &path, const Headers &headers,
+                            ContentProviderWithoutLength content_provider,
+                            const std::string &content_type) {
+  return cli_->Patch(path, headers, std::move(content_provider), content_type);
+}
+inline Result Client::Delete(const std::string &path) {
+  return cli_->Delete(path);
+}
+inline Result Client::Delete(const std::string &path, const Headers &headers) {
+  return cli_->Delete(path, headers);
+}
+inline Result Client::Delete(const std::string &path, const char *body,
+                             size_t content_length,
+                             const std::string &content_type) {
+  return cli_->Delete(path, body, content_length, content_type);
+}
+inline Result Client::Delete(const std::string &path, const Headers &headers,
+                             const char *body, size_t content_length,
+                             const std::string &content_type) {
+  return cli_->Delete(path, headers, body, content_length, content_type);
+}
+inline Result Client::Delete(const std::string &path, const std::string &body,
+                             const std::string &content_type) {
+  return cli_->Delete(path, body, content_type);
+}
+inline Result Client::Delete(const std::string &path, const Headers &headers,
+                             const std::string &body,
+                             const std::string &content_type) {
+  return cli_->Delete(path, headers, body, content_type);
+}
+inline Result Client::Options(const std::string &path) {
+  return cli_->Options(path);
+}
+inline Result Client::Options(const std::string &path, const Headers &headers) {
+  return cli_->Options(path, headers);
+}
+
+inline bool Client::send(Request &req, Response &res, Error &error) {
+  return cli_->send(req, res, error);
+}
+
+inline Result Client::send(const Request &req) { return cli_->send(req); }
+
+inline void Client::stop() { cli_->stop(); }
+
+inline std::string Client::host() const { return cli_->host(); }
+
+inline int Client::port() const { return cli_->port(); }
+
+inline size_t Client::is_socket_open() const { return cli_->is_socket_open(); }
+
+inline socket_t Client::socket() const { return cli_->socket(); }
+
+inline void
+Client::set_hostname_addr_map(std::map<std::string, std::string> addr_map) {
+  cli_->set_hostname_addr_map(std::move(addr_map));
+}
+
+inline void Client::set_default_headers(Headers headers) {
+  cli_->set_default_headers(std::move(headers));
+}
+
+inline void Client::set_header_writer(
+    std::function<ssize_t(Stream &, Headers &)> const &writer) {
+  cli_->set_header_writer(writer);
+}
+
+inline void Client::set_address_family(int family) {
+  cli_->set_address_family(family);
+}
+
+inline void Client::set_tcp_nodelay(bool on) { cli_->set_tcp_nodelay(on); }
+
+inline void Client::set_socket_options(SocketOptions socket_options) {
+  cli_->set_socket_options(std::move(socket_options));
+}
+
+inline void Client::set_connection_timeout(time_t sec, time_t usec) {
+  cli_->set_connection_timeout(sec, usec);
+}
+
+inline void Client::set_read_timeout(time_t sec, time_t usec) {
+  cli_->set_read_timeout(sec, usec);
+}
+
+inline void Client::set_write_timeout(time_t sec, time_t usec) {
+  cli_->set_write_timeout(sec, usec);
+}
+
+inline void Client::set_basic_auth(const std::string &username,
+                                   const std::string &password) {
+  cli_->set_basic_auth(username, password);
+}
+inline void Client::set_bearer_token_auth(const std::string &token) {
+  cli_->set_bearer_token_auth(token);
+}
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline void Client::set_digest_auth(const std::string &username,
+                                    const std::string &password) {
+  cli_->set_digest_auth(username, password);
+}
+#endif
+
+inline void Client::set_keep_alive(bool on) { cli_->set_keep_alive(on); }
+inline void Client::set_follow_location(bool on) {
+  cli_->set_follow_location(on);
+}
+
+inline void Client::set_url_encode(bool on) { cli_->set_url_encode(on); }
+
+inline void Client::set_compress(bool on) { cli_->set_compress(on); }
+
+inline void Client::set_decompress(bool on) { cli_->set_decompress(on); }
+
+inline void Client::set_interface(const std::string &intf) {
+  cli_->set_interface(intf);
+}
+
+inline void Client::set_proxy(const std::string &host, int port) {
+  cli_->set_proxy(host, port);
+}
+inline void Client::set_proxy_basic_auth(const std::string &username,
+                                         const std::string &password) {
+  cli_->set_proxy_basic_auth(username, password);
+}
+inline void Client::set_proxy_bearer_token_auth(const std::string &token) {
+  cli_->set_proxy_bearer_token_auth(token);
+}
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline void Client::set_proxy_digest_auth(const std::string &username,
+                                          const std::string &password) {
+  cli_->set_proxy_digest_auth(username, password);
+}
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline void Client::enable_server_certificate_verification(bool enabled) {
+  cli_->enable_server_certificate_verification(enabled);
+}
+#endif
+
+inline void Client::set_logger(Logger logger) {
+  cli_->set_logger(std::move(logger));
+}
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline void Client::set_ca_cert_path(const std::string &ca_cert_file_path,
+                                     const std::string &ca_cert_dir_path) {
+  cli_->set_ca_cert_path(ca_cert_file_path, ca_cert_dir_path);
+}
+
+inline void Client::set_ca_cert_store(X509_STORE *ca_cert_store) {
+  if (is_ssl_) {
+    static_cast<SSLClient &>(*cli_).set_ca_cert_store(ca_cert_store);
+  } else {
+    cli_->set_ca_cert_store(ca_cert_store);
+  }
+}
+
+inline void Client::load_ca_cert_store(const char *ca_cert, std::size_t size) {
+  set_ca_cert_store(cli_->create_ca_cert_store(ca_cert, size));
+}
+
+inline long Client::get_openssl_verify_result() const {
+  if (is_ssl_) {
+    return static_cast<SSLClient &>(*cli_).get_openssl_verify_result();
+  }
+  return -1; // NOTE: -1 doesn't match any of X509_V_ERR_???
+}
+
+inline SSL_CTX *Client::ssl_context() const {
+  if (is_ssl_) { return static_cast<SSLClient &>(*cli_).ssl_context(); }
+  return nullptr;
+}
+#endif
+
+// ----------------------------------------------------------------------------
+
+} // namespace httplib
+
+#if defined(_WIN32) && defined(CPPHTTPLIB_USE_POLL)
+#undef poll
+#endif
+
+#endif // CPPHTTPLIB_HTTPLIB_H
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/json.hpp b/projects/llm_framework/main_cosy_voice/src/runner/utils/json.hpp
new file mode 100644
index 00000000..a858728c
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/json.hpp
@@ -0,0 +1,24766 @@
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+/****************************************************************************\
+ * Note on documentation: The source files contain links to the online      *
+ * documentation of the public API at https://json.nlohmann.me. This URL    *
+ * contains the most recent documentation and should also be applicable to  *
+ * previous versions; documentation for deprecated functions is not         *
+ * removed, but marked deprecated. See "Generate documentation" section in  *
+ * file docs/README.md.                                                     *
+\****************************************************************************/
+
+#ifndef INCLUDE_NLOHMANN_JSON_HPP_
+#define INCLUDE_NLOHMANN_JSON_HPP_
+
+#include <algorithm> // all_of, find, for_each
+#include <cstddef> // nullptr_t, ptrdiff_t, size_t
+#include <functional> // hash, less
+#include <initializer_list> // initializer_list
+#ifndef JSON_NO_IO
+    #include <iosfwd> // istream, ostream
+#endif  // JSON_NO_IO
+#include <iterator> // random_access_iterator_tag
+#include <memory> // unique_ptr
+#include <string> // string, stoi, to_string
+#include <utility> // declval, forward, move, pair, swap
+#include <vector> // vector
+
+// #include <nlohmann/adl_serializer.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <utility>
+
+// #include <nlohmann/detail/abi_macros.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// This file contains all macro definitions affecting or depending on the ABI
+
+#ifndef JSON_SKIP_LIBRARY_VERSION_CHECK
+    #if defined(NLOHMANN_JSON_VERSION_MAJOR) && defined(NLOHMANN_JSON_VERSION_MINOR) && defined(NLOHMANN_JSON_VERSION_PATCH)
+        #if NLOHMANN_JSON_VERSION_MAJOR != 3 || NLOHMANN_JSON_VERSION_MINOR != 11 || NLOHMANN_JSON_VERSION_PATCH != 3
+            #warning "Already included a different version of the library!"
+        #endif
+    #endif
+#endif
+
+#define NLOHMANN_JSON_VERSION_MAJOR 3   // NOLINT(modernize-macro-to-enum)
+#define NLOHMANN_JSON_VERSION_MINOR 11  // NOLINT(modernize-macro-to-enum)
+#define NLOHMANN_JSON_VERSION_PATCH 3   // NOLINT(modernize-macro-to-enum)
+
+#ifndef JSON_DIAGNOSTICS
+    #define JSON_DIAGNOSTICS 0
+#endif
+
+#ifndef JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
+    #define JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON 0
+#endif
+
+#if JSON_DIAGNOSTICS
+    #define NLOHMANN_JSON_ABI_TAG_DIAGNOSTICS _diag
+#else
+    #define NLOHMANN_JSON_ABI_TAG_DIAGNOSTICS
+#endif
+
+#if JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
+    #define NLOHMANN_JSON_ABI_TAG_LEGACY_DISCARDED_VALUE_COMPARISON _ldvcmp
+#else
+    #define NLOHMANN_JSON_ABI_TAG_LEGACY_DISCARDED_VALUE_COMPARISON
+#endif
+
+#ifndef NLOHMANN_JSON_NAMESPACE_NO_VERSION
+    #define NLOHMANN_JSON_NAMESPACE_NO_VERSION 0
+#endif
+
+// Construct the namespace ABI tags component
+#define NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b) json_abi ## a ## b
+#define NLOHMANN_JSON_ABI_TAGS_CONCAT(a, b) \
+    NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b)
+
+#define NLOHMANN_JSON_ABI_TAGS                                       \
+    NLOHMANN_JSON_ABI_TAGS_CONCAT(                                   \
+            NLOHMANN_JSON_ABI_TAG_DIAGNOSTICS,                       \
+            NLOHMANN_JSON_ABI_TAG_LEGACY_DISCARDED_VALUE_COMPARISON)
+
+// Construct the namespace version component
+#define NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT_EX(major, minor, patch) \
+    _v ## major ## _ ## minor ## _ ## patch
+#define NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT(major, minor, patch) \
+    NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT_EX(major, minor, patch)
+
+#if NLOHMANN_JSON_NAMESPACE_NO_VERSION
+#define NLOHMANN_JSON_NAMESPACE_VERSION
+#else
+#define NLOHMANN_JSON_NAMESPACE_VERSION                                 \
+    NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT(NLOHMANN_JSON_VERSION_MAJOR, \
+                                           NLOHMANN_JSON_VERSION_MINOR, \
+                                           NLOHMANN_JSON_VERSION_PATCH)
+#endif
+
+// Combine namespace components
+#define NLOHMANN_JSON_NAMESPACE_CONCAT_EX(a, b) a ## b
+#define NLOHMANN_JSON_NAMESPACE_CONCAT(a, b) \
+    NLOHMANN_JSON_NAMESPACE_CONCAT_EX(a, b)
+
+#ifndef NLOHMANN_JSON_NAMESPACE
+#define NLOHMANN_JSON_NAMESPACE               \
+    nlohmann::NLOHMANN_JSON_NAMESPACE_CONCAT( \
+            NLOHMANN_JSON_ABI_TAGS,           \
+            NLOHMANN_JSON_NAMESPACE_VERSION)
+#endif
+
+#ifndef NLOHMANN_JSON_NAMESPACE_BEGIN
+#define NLOHMANN_JSON_NAMESPACE_BEGIN                \
+    namespace nlohmann                               \
+    {                                                \
+    inline namespace NLOHMANN_JSON_NAMESPACE_CONCAT( \
+                NLOHMANN_JSON_ABI_TAGS,              \
+                NLOHMANN_JSON_NAMESPACE_VERSION)     \
+    {
+#endif
+
+#ifndef NLOHMANN_JSON_NAMESPACE_END
+#define NLOHMANN_JSON_NAMESPACE_END                                     \
+    }  /* namespace (inline namespace) NOLINT(readability/namespace) */ \
+    }  // namespace nlohmann
+#endif
+
+// #include <nlohmann/detail/conversions/from_json.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <algorithm> // transform
+#include <array> // array
+#include <forward_list> // forward_list
+#include <iterator> // inserter, front_inserter, end
+#include <map> // map
+#include <string> // string
+#include <tuple> // tuple, make_tuple
+#include <type_traits> // is_arithmetic, is_same, is_enum, underlying_type, is_convertible
+#include <unordered_map> // unordered_map
+#include <utility> // pair, declval
+#include <valarray> // valarray
+
+// #include <nlohmann/detail/exceptions.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstddef> // nullptr_t
+#include <exception> // exception
+#if JSON_DIAGNOSTICS
+    #include <numeric> // accumulate
+#endif
+#include <stdexcept> // runtime_error
+#include <string> // to_string
+#include <vector> // vector
+
+// #include <nlohmann/detail/value_t.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <array> // array
+#include <cstddef> // size_t
+#include <cstdint> // uint8_t
+#include <string> // string
+
+// #include <nlohmann/detail/macro_scope.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <utility> // declval, pair
+// #include <nlohmann/detail/meta/detected.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <type_traits>
+
+// #include <nlohmann/detail/meta/void_t.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+template<typename ...Ts> struct make_void
+{
+    using type = void;
+};
+template<typename ...Ts> using void_t = typename make_void<Ts...>::type;
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+// https://en.cppreference.com/w/cpp/experimental/is_detected
+struct nonesuch
+{
+    nonesuch() = delete;
+    ~nonesuch() = delete;
+    nonesuch(nonesuch const&) = delete;
+    nonesuch(nonesuch const&&) = delete;
+    void operator=(nonesuch const&) = delete;
+    void operator=(nonesuch&&) = delete;
+};
+
+template<class Default,
+         class AlwaysVoid,
+         template<class...> class Op,
+         class... Args>
+struct detector
+{
+    using value_t = std::false_type;
+    using type = Default;
+};
+
+template<class Default, template<class...> class Op, class... Args>
+struct detector<Default, void_t<Op<Args...>>, Op, Args...>
+{
+    using value_t = std::true_type;
+    using type = Op<Args...>;
+};
+
+template<template<class...> class Op, class... Args>
+using is_detected = typename detector<nonesuch, void, Op, Args...>::value_t;
+
+template<template<class...> class Op, class... Args>
+struct is_detected_lazy : is_detected<Op, Args...> { };
+
+template<template<class...> class Op, class... Args>
+using detected_t = typename detector<nonesuch, void, Op, Args...>::type;
+
+template<class Default, template<class...> class Op, class... Args>
+using detected_or = detector<Default, void, Op, Args...>;
+
+template<class Default, template<class...> class Op, class... Args>
+using detected_or_t = typename detected_or<Default, Op, Args...>::type;
+
+template<class Expected, template<class...> class Op, class... Args>
+using is_detected_exact = std::is_same<Expected, detected_t<Op, Args...>>;
+
+template<class To, template<class...> class Op, class... Args>
+using is_detected_convertible =
+    std::is_convertible<detected_t<Op, Args...>, To>;
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/thirdparty/hedley/hedley.hpp>
+
+
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-FileCopyrightText: 2016-2021 Evan Nemerson <evan@nemerson.com>
+// SPDX-License-Identifier: MIT
+
+/* Hedley - https://nemequ.github.io/hedley
+ * Created by Evan Nemerson <evan@nemerson.com>
+ */
+
+#if !defined(JSON_HEDLEY_VERSION) || (JSON_HEDLEY_VERSION < 15)
+#if defined(JSON_HEDLEY_VERSION)
+    #undef JSON_HEDLEY_VERSION
+#endif
+#define JSON_HEDLEY_VERSION 15
+
+#if defined(JSON_HEDLEY_STRINGIFY_EX)
+    #undef JSON_HEDLEY_STRINGIFY_EX
+#endif
+#define JSON_HEDLEY_STRINGIFY_EX(x) #x
+
+#if defined(JSON_HEDLEY_STRINGIFY)
+    #undef JSON_HEDLEY_STRINGIFY
+#endif
+#define JSON_HEDLEY_STRINGIFY(x) JSON_HEDLEY_STRINGIFY_EX(x)
+
+#if defined(JSON_HEDLEY_CONCAT_EX)
+    #undef JSON_HEDLEY_CONCAT_EX
+#endif
+#define JSON_HEDLEY_CONCAT_EX(a,b) a##b
+
+#if defined(JSON_HEDLEY_CONCAT)
+    #undef JSON_HEDLEY_CONCAT
+#endif
+#define JSON_HEDLEY_CONCAT(a,b) JSON_HEDLEY_CONCAT_EX(a,b)
+
+#if defined(JSON_HEDLEY_CONCAT3_EX)
+    #undef JSON_HEDLEY_CONCAT3_EX
+#endif
+#define JSON_HEDLEY_CONCAT3_EX(a,b,c) a##b##c
+
+#if defined(JSON_HEDLEY_CONCAT3)
+    #undef JSON_HEDLEY_CONCAT3
+#endif
+#define JSON_HEDLEY_CONCAT3(a,b,c) JSON_HEDLEY_CONCAT3_EX(a,b,c)
+
+#if defined(JSON_HEDLEY_VERSION_ENCODE)
+    #undef JSON_HEDLEY_VERSION_ENCODE
+#endif
+#define JSON_HEDLEY_VERSION_ENCODE(major,minor,revision) (((major) * 1000000) + ((minor) * 1000) + (revision))
+
+#if defined(JSON_HEDLEY_VERSION_DECODE_MAJOR)
+    #undef JSON_HEDLEY_VERSION_DECODE_MAJOR
+#endif
+#define JSON_HEDLEY_VERSION_DECODE_MAJOR(version) ((version) / 1000000)
+
+#if defined(JSON_HEDLEY_VERSION_DECODE_MINOR)
+    #undef JSON_HEDLEY_VERSION_DECODE_MINOR
+#endif
+#define JSON_HEDLEY_VERSION_DECODE_MINOR(version) (((version) % 1000000) / 1000)
+
+#if defined(JSON_HEDLEY_VERSION_DECODE_REVISION)
+    #undef JSON_HEDLEY_VERSION_DECODE_REVISION
+#endif
+#define JSON_HEDLEY_VERSION_DECODE_REVISION(version) ((version) % 1000)
+
+#if defined(JSON_HEDLEY_GNUC_VERSION)
+    #undef JSON_HEDLEY_GNUC_VERSION
+#endif
+#if defined(__GNUC__) && defined(__GNUC_PATCHLEVEL__)
+    #define JSON_HEDLEY_GNUC_VERSION JSON_HEDLEY_VERSION_ENCODE(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__)
+#elif defined(__GNUC__)
+    #define JSON_HEDLEY_GNUC_VERSION JSON_HEDLEY_VERSION_ENCODE(__GNUC__, __GNUC_MINOR__, 0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_VERSION_CHECK)
+    #undef JSON_HEDLEY_GNUC_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_GNUC_VERSION)
+    #define JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_GNUC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_MSVC_VERSION)
+    #undef JSON_HEDLEY_MSVC_VERSION
+#endif
+#if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 140000000) && !defined(__ICL)
+    #define JSON_HEDLEY_MSVC_VERSION JSON_HEDLEY_VERSION_ENCODE(_MSC_FULL_VER / 10000000, (_MSC_FULL_VER % 10000000) / 100000, (_MSC_FULL_VER % 100000) / 100)
+#elif defined(_MSC_FULL_VER) && !defined(__ICL)
+    #define JSON_HEDLEY_MSVC_VERSION JSON_HEDLEY_VERSION_ENCODE(_MSC_FULL_VER / 1000000, (_MSC_FULL_VER % 1000000) / 10000, (_MSC_FULL_VER % 10000) / 10)
+#elif defined(_MSC_VER) && !defined(__ICL)
+    #define JSON_HEDLEY_MSVC_VERSION JSON_HEDLEY_VERSION_ENCODE(_MSC_VER / 100, _MSC_VER % 100, 0)
+#endif
+
+#if defined(JSON_HEDLEY_MSVC_VERSION_CHECK)
+    #undef JSON_HEDLEY_MSVC_VERSION_CHECK
+#endif
+#if !defined(JSON_HEDLEY_MSVC_VERSION)
+    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (0)
+#elif defined(_MSC_VER) && (_MSC_VER >= 1400)
+    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (_MSC_FULL_VER >= ((major * 10000000) + (minor * 100000) + (patch)))
+#elif defined(_MSC_VER) && (_MSC_VER >= 1200)
+    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (_MSC_FULL_VER >= ((major * 1000000) + (minor * 10000) + (patch)))
+#else
+    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (_MSC_VER >= ((major * 100) + (minor)))
+#endif
+
+#if defined(JSON_HEDLEY_INTEL_VERSION)
+    #undef JSON_HEDLEY_INTEL_VERSION
+#endif
+#if defined(__INTEL_COMPILER) && defined(__INTEL_COMPILER_UPDATE) && !defined(__ICL)
+    #define JSON_HEDLEY_INTEL_VERSION JSON_HEDLEY_VERSION_ENCODE(__INTEL_COMPILER / 100, __INTEL_COMPILER % 100, __INTEL_COMPILER_UPDATE)
+#elif defined(__INTEL_COMPILER) && !defined(__ICL)
+    #define JSON_HEDLEY_INTEL_VERSION JSON_HEDLEY_VERSION_ENCODE(__INTEL_COMPILER / 100, __INTEL_COMPILER % 100, 0)
+#endif
+
+#if defined(JSON_HEDLEY_INTEL_VERSION_CHECK)
+    #undef JSON_HEDLEY_INTEL_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_INTEL_VERSION)
+    #define JSON_HEDLEY_INTEL_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_INTEL_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_INTEL_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_INTEL_CL_VERSION)
+    #undef JSON_HEDLEY_INTEL_CL_VERSION
+#endif
+#if defined(__INTEL_COMPILER) && defined(__INTEL_COMPILER_UPDATE) && defined(__ICL)
+    #define JSON_HEDLEY_INTEL_CL_VERSION JSON_HEDLEY_VERSION_ENCODE(__INTEL_COMPILER, __INTEL_COMPILER_UPDATE, 0)
+#endif
+
+#if defined(JSON_HEDLEY_INTEL_CL_VERSION_CHECK)
+    #undef JSON_HEDLEY_INTEL_CL_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_INTEL_CL_VERSION)
+    #define JSON_HEDLEY_INTEL_CL_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_INTEL_CL_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_INTEL_CL_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_PGI_VERSION)
+    #undef JSON_HEDLEY_PGI_VERSION
+#endif
+#if defined(__PGI) && defined(__PGIC__) && defined(__PGIC_MINOR__) && defined(__PGIC_PATCHLEVEL__)
+    #define JSON_HEDLEY_PGI_VERSION JSON_HEDLEY_VERSION_ENCODE(__PGIC__, __PGIC_MINOR__, __PGIC_PATCHLEVEL__)
+#endif
+
+#if defined(JSON_HEDLEY_PGI_VERSION_CHECK)
+    #undef JSON_HEDLEY_PGI_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_PGI_VERSION)
+    #define JSON_HEDLEY_PGI_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_PGI_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_PGI_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_SUNPRO_VERSION)
+    #undef JSON_HEDLEY_SUNPRO_VERSION
+#endif
+#if defined(__SUNPRO_C) && (__SUNPRO_C > 0x1000)
+    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((((__SUNPRO_C >> 16) & 0xf) * 10) + ((__SUNPRO_C >> 12) & 0xf), (((__SUNPRO_C >> 8) & 0xf) * 10) + ((__SUNPRO_C >> 4) & 0xf), (__SUNPRO_C & 0xf) * 10)
+#elif defined(__SUNPRO_C)
+    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((__SUNPRO_C >> 8) & 0xf, (__SUNPRO_C >> 4) & 0xf, (__SUNPRO_C) & 0xf)
+#elif defined(__SUNPRO_CC) && (__SUNPRO_CC > 0x1000)
+    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((((__SUNPRO_CC >> 16) & 0xf) * 10) + ((__SUNPRO_CC >> 12) & 0xf), (((__SUNPRO_CC >> 8) & 0xf) * 10) + ((__SUNPRO_CC >> 4) & 0xf), (__SUNPRO_CC & 0xf) * 10)
+#elif defined(__SUNPRO_CC)
+    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((__SUNPRO_CC >> 8) & 0xf, (__SUNPRO_CC >> 4) & 0xf, (__SUNPRO_CC) & 0xf)
+#endif
+
+#if defined(JSON_HEDLEY_SUNPRO_VERSION_CHECK)
+    #undef JSON_HEDLEY_SUNPRO_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_SUNPRO_VERSION)
+    #define JSON_HEDLEY_SUNPRO_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_SUNPRO_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_SUNPRO_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_EMSCRIPTEN_VERSION)
+    #undef JSON_HEDLEY_EMSCRIPTEN_VERSION
+#endif
+#if defined(__EMSCRIPTEN__)
+    #define JSON_HEDLEY_EMSCRIPTEN_VERSION JSON_HEDLEY_VERSION_ENCODE(__EMSCRIPTEN_major__, __EMSCRIPTEN_minor__, __EMSCRIPTEN_tiny__)
+#endif
+
+#if defined(JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK)
+    #undef JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_EMSCRIPTEN_VERSION)
+    #define JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_EMSCRIPTEN_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_ARM_VERSION)
+    #undef JSON_HEDLEY_ARM_VERSION
+#endif
+#if defined(__CC_ARM) && defined(__ARMCOMPILER_VERSION)
+    #define JSON_HEDLEY_ARM_VERSION JSON_HEDLEY_VERSION_ENCODE(__ARMCOMPILER_VERSION / 1000000, (__ARMCOMPILER_VERSION % 1000000) / 10000, (__ARMCOMPILER_VERSION % 10000) / 100)
+#elif defined(__CC_ARM) && defined(__ARMCC_VERSION)
+    #define JSON_HEDLEY_ARM_VERSION JSON_HEDLEY_VERSION_ENCODE(__ARMCC_VERSION / 1000000, (__ARMCC_VERSION % 1000000) / 10000, (__ARMCC_VERSION % 10000) / 100)
+#endif
+
+#if defined(JSON_HEDLEY_ARM_VERSION_CHECK)
+    #undef JSON_HEDLEY_ARM_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_ARM_VERSION)
+    #define JSON_HEDLEY_ARM_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_ARM_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_ARM_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_IBM_VERSION)
+    #undef JSON_HEDLEY_IBM_VERSION
+#endif
+#if defined(__ibmxl__)
+    #define JSON_HEDLEY_IBM_VERSION JSON_HEDLEY_VERSION_ENCODE(__ibmxl_version__, __ibmxl_release__, __ibmxl_modification__)
+#elif defined(__xlC__) && defined(__xlC_ver__)
+    #define JSON_HEDLEY_IBM_VERSION JSON_HEDLEY_VERSION_ENCODE(__xlC__ >> 8, __xlC__ & 0xff, (__xlC_ver__ >> 8) & 0xff)
+#elif defined(__xlC__)
+    #define JSON_HEDLEY_IBM_VERSION JSON_HEDLEY_VERSION_ENCODE(__xlC__ >> 8, __xlC__ & 0xff, 0)
+#endif
+
+#if defined(JSON_HEDLEY_IBM_VERSION_CHECK)
+    #undef JSON_HEDLEY_IBM_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_IBM_VERSION)
+    #define JSON_HEDLEY_IBM_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_IBM_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_IBM_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_VERSION)
+    #undef JSON_HEDLEY_TI_VERSION
+#endif
+#if \
+    defined(__TI_COMPILER_VERSION__) && \
+    ( \
+      defined(__TMS470__) || defined(__TI_ARM__) || \
+      defined(__MSP430__) || \
+      defined(__TMS320C2000__) \
+    )
+#if (__TI_COMPILER_VERSION__ >= 16000000)
+    #define JSON_HEDLEY_TI_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+#endif
+
+#if defined(JSON_HEDLEY_TI_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_VERSION)
+    #define JSON_HEDLEY_TI_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL2000_VERSION)
+    #undef JSON_HEDLEY_TI_CL2000_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && defined(__TMS320C2000__)
+    #define JSON_HEDLEY_TI_CL2000_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL2000_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_CL2000_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_CL2000_VERSION)
+    #define JSON_HEDLEY_TI_CL2000_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL2000_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_CL2000_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL430_VERSION)
+    #undef JSON_HEDLEY_TI_CL430_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && defined(__MSP430__)
+    #define JSON_HEDLEY_TI_CL430_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL430_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_CL430_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_CL430_VERSION)
+    #define JSON_HEDLEY_TI_CL430_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL430_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_CL430_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_ARMCL_VERSION)
+    #undef JSON_HEDLEY_TI_ARMCL_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && (defined(__TMS470__) || defined(__TI_ARM__))
+    #define JSON_HEDLEY_TI_ARMCL_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_ARMCL_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_ARMCL_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_ARMCL_VERSION)
+    #define JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_ARMCL_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL6X_VERSION)
+    #undef JSON_HEDLEY_TI_CL6X_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && defined(__TMS320C6X__)
+    #define JSON_HEDLEY_TI_CL6X_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL6X_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_CL6X_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_CL6X_VERSION)
+    #define JSON_HEDLEY_TI_CL6X_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL6X_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_CL6X_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL7X_VERSION)
+    #undef JSON_HEDLEY_TI_CL7X_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && defined(__C7000__)
+    #define JSON_HEDLEY_TI_CL7X_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL7X_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_CL7X_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_CL7X_VERSION)
+    #define JSON_HEDLEY_TI_CL7X_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL7X_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_CL7X_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_CLPRU_VERSION)
+    #undef JSON_HEDLEY_TI_CLPRU_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && defined(__PRU__)
+    #define JSON_HEDLEY_TI_CLPRU_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_CLPRU_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_CLPRU_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_CLPRU_VERSION)
+    #define JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CLPRU_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_CRAY_VERSION)
+    #undef JSON_HEDLEY_CRAY_VERSION
+#endif
+#if defined(_CRAYC)
+    #if defined(_RELEASE_PATCHLEVEL)
+        #define JSON_HEDLEY_CRAY_VERSION JSON_HEDLEY_VERSION_ENCODE(_RELEASE_MAJOR, _RELEASE_MINOR, _RELEASE_PATCHLEVEL)
+    #else
+        #define JSON_HEDLEY_CRAY_VERSION JSON_HEDLEY_VERSION_ENCODE(_RELEASE_MAJOR, _RELEASE_MINOR, 0)
+    #endif
+#endif
+
+#if defined(JSON_HEDLEY_CRAY_VERSION_CHECK)
+    #undef JSON_HEDLEY_CRAY_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_CRAY_VERSION)
+    #define JSON_HEDLEY_CRAY_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_CRAY_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_CRAY_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_IAR_VERSION)
+    #undef JSON_HEDLEY_IAR_VERSION
+#endif
+#if defined(__IAR_SYSTEMS_ICC__)
+    #if __VER__ > 1000
+        #define JSON_HEDLEY_IAR_VERSION JSON_HEDLEY_VERSION_ENCODE((__VER__ / 1000000), ((__VER__ / 1000) % 1000), (__VER__ % 1000))
+    #else
+        #define JSON_HEDLEY_IAR_VERSION JSON_HEDLEY_VERSION_ENCODE(__VER__ / 100, __VER__ % 100, 0)
+    #endif
+#endif
+
+#if defined(JSON_HEDLEY_IAR_VERSION_CHECK)
+    #undef JSON_HEDLEY_IAR_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_IAR_VERSION)
+    #define JSON_HEDLEY_IAR_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_IAR_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_IAR_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TINYC_VERSION)
+    #undef JSON_HEDLEY_TINYC_VERSION
+#endif
+#if defined(__TINYC__)
+    #define JSON_HEDLEY_TINYC_VERSION JSON_HEDLEY_VERSION_ENCODE(__TINYC__ / 1000, (__TINYC__ / 100) % 10, __TINYC__ % 100)
+#endif
+
+#if defined(JSON_HEDLEY_TINYC_VERSION_CHECK)
+    #undef JSON_HEDLEY_TINYC_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TINYC_VERSION)
+    #define JSON_HEDLEY_TINYC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TINYC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TINYC_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_DMC_VERSION)
+    #undef JSON_HEDLEY_DMC_VERSION
+#endif
+#if defined(__DMC__)
+    #define JSON_HEDLEY_DMC_VERSION JSON_HEDLEY_VERSION_ENCODE(__DMC__ >> 8, (__DMC__ >> 4) & 0xf, __DMC__ & 0xf)
+#endif
+
+#if defined(JSON_HEDLEY_DMC_VERSION_CHECK)
+    #undef JSON_HEDLEY_DMC_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_DMC_VERSION)
+    #define JSON_HEDLEY_DMC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_DMC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_DMC_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_COMPCERT_VERSION)
+    #undef JSON_HEDLEY_COMPCERT_VERSION
+#endif
+#if defined(__COMPCERT_VERSION__)
+    #define JSON_HEDLEY_COMPCERT_VERSION JSON_HEDLEY_VERSION_ENCODE(__COMPCERT_VERSION__ / 10000, (__COMPCERT_VERSION__ / 100) % 100, __COMPCERT_VERSION__ % 100)
+#endif
+
+#if defined(JSON_HEDLEY_COMPCERT_VERSION_CHECK)
+    #undef JSON_HEDLEY_COMPCERT_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_COMPCERT_VERSION)
+    #define JSON_HEDLEY_COMPCERT_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_COMPCERT_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_COMPCERT_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_PELLES_VERSION)
+    #undef JSON_HEDLEY_PELLES_VERSION
+#endif
+#if defined(__POCC__)
+    #define JSON_HEDLEY_PELLES_VERSION JSON_HEDLEY_VERSION_ENCODE(__POCC__ / 100, __POCC__ % 100, 0)
+#endif
+
+#if defined(JSON_HEDLEY_PELLES_VERSION_CHECK)
+    #undef JSON_HEDLEY_PELLES_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_PELLES_VERSION)
+    #define JSON_HEDLEY_PELLES_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_PELLES_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_PELLES_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_MCST_LCC_VERSION)
+    #undef JSON_HEDLEY_MCST_LCC_VERSION
+#endif
+#if defined(__LCC__) && defined(__LCC_MINOR__)
+    #define JSON_HEDLEY_MCST_LCC_VERSION JSON_HEDLEY_VERSION_ENCODE(__LCC__ / 100, __LCC__ % 100, __LCC_MINOR__)
+#endif
+
+#if defined(JSON_HEDLEY_MCST_LCC_VERSION_CHECK)
+    #undef JSON_HEDLEY_MCST_LCC_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_MCST_LCC_VERSION)
+    #define JSON_HEDLEY_MCST_LCC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_MCST_LCC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_MCST_LCC_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_VERSION)
+    #undef JSON_HEDLEY_GCC_VERSION
+#endif
+#if \
+    defined(JSON_HEDLEY_GNUC_VERSION) && \
+    !defined(__clang__) && \
+    !defined(JSON_HEDLEY_INTEL_VERSION) && \
+    !defined(JSON_HEDLEY_PGI_VERSION) && \
+    !defined(JSON_HEDLEY_ARM_VERSION) && \
+    !defined(JSON_HEDLEY_CRAY_VERSION) && \
+    !defined(JSON_HEDLEY_TI_VERSION) && \
+    !defined(JSON_HEDLEY_TI_ARMCL_VERSION) && \
+    !defined(JSON_HEDLEY_TI_CL430_VERSION) && \
+    !defined(JSON_HEDLEY_TI_CL2000_VERSION) && \
+    !defined(JSON_HEDLEY_TI_CL6X_VERSION) && \
+    !defined(JSON_HEDLEY_TI_CL7X_VERSION) && \
+    !defined(JSON_HEDLEY_TI_CLPRU_VERSION) && \
+    !defined(__COMPCERT__) && \
+    !defined(JSON_HEDLEY_MCST_LCC_VERSION)
+    #define JSON_HEDLEY_GCC_VERSION JSON_HEDLEY_GNUC_VERSION
+#endif
+
+#if defined(JSON_HEDLEY_GCC_VERSION_CHECK)
+    #undef JSON_HEDLEY_GCC_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_GCC_VERSION)
+    #define JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_GCC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_ATTRIBUTE)
+    #undef JSON_HEDLEY_HAS_ATTRIBUTE
+#endif
+#if \
+  defined(__has_attribute) && \
+  ( \
+    (!defined(JSON_HEDLEY_IAR_VERSION) || JSON_HEDLEY_IAR_VERSION_CHECK(8,5,9)) \
+  )
+#  define JSON_HEDLEY_HAS_ATTRIBUTE(attribute) __has_attribute(attribute)
+#else
+#  define JSON_HEDLEY_HAS_ATTRIBUTE(attribute) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_ATTRIBUTE)
+    #undef JSON_HEDLEY_GNUC_HAS_ATTRIBUTE
+#endif
+#if defined(__has_attribute)
+    #define JSON_HEDLEY_GNUC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_HAS_ATTRIBUTE(attribute)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_ATTRIBUTE)
+    #undef JSON_HEDLEY_GCC_HAS_ATTRIBUTE
+#endif
+#if defined(__has_attribute)
+    #define JSON_HEDLEY_GCC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_HAS_ATTRIBUTE(attribute)
+#else
+    #define JSON_HEDLEY_GCC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_CPP_ATTRIBUTE)
+    #undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE
+#endif
+#if \
+    defined(__has_cpp_attribute) && \
+    defined(__cplusplus) && \
+    (!defined(JSON_HEDLEY_SUNPRO_VERSION) || JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0))
+    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE(attribute) __has_cpp_attribute(attribute)
+#else
+    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE(attribute) (0)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS)
+    #undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS
+#endif
+#if !defined(__cplusplus) || !defined(__has_cpp_attribute)
+    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(ns,attribute) (0)
+#elif \
+    !defined(JSON_HEDLEY_PGI_VERSION) && \
+    !defined(JSON_HEDLEY_IAR_VERSION) && \
+    (!defined(JSON_HEDLEY_SUNPRO_VERSION) || JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0)) && \
+    (!defined(JSON_HEDLEY_MSVC_VERSION) || JSON_HEDLEY_MSVC_VERSION_CHECK(19,20,0))
+    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(ns,attribute) JSON_HEDLEY_HAS_CPP_ATTRIBUTE(ns::attribute)
+#else
+    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(ns,attribute) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE)
+    #undef JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE
+#endif
+#if defined(__has_cpp_attribute) && defined(__cplusplus)
+    #define JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) __has_cpp_attribute(attribute)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE)
+    #undef JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE
+#endif
+#if defined(__has_cpp_attribute) && defined(__cplusplus)
+    #define JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) __has_cpp_attribute(attribute)
+#else
+    #define JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_BUILTIN)
+    #undef JSON_HEDLEY_HAS_BUILTIN
+#endif
+#if defined(__has_builtin)
+    #define JSON_HEDLEY_HAS_BUILTIN(builtin) __has_builtin(builtin)
+#else
+    #define JSON_HEDLEY_HAS_BUILTIN(builtin) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_BUILTIN)
+    #undef JSON_HEDLEY_GNUC_HAS_BUILTIN
+#endif
+#if defined(__has_builtin)
+    #define JSON_HEDLEY_GNUC_HAS_BUILTIN(builtin,major,minor,patch) __has_builtin(builtin)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_BUILTIN(builtin,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_BUILTIN)
+    #undef JSON_HEDLEY_GCC_HAS_BUILTIN
+#endif
+#if defined(__has_builtin)
+    #define JSON_HEDLEY_GCC_HAS_BUILTIN(builtin,major,minor,patch) __has_builtin(builtin)
+#else
+    #define JSON_HEDLEY_GCC_HAS_BUILTIN(builtin,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_FEATURE)
+    #undef JSON_HEDLEY_HAS_FEATURE
+#endif
+#if defined(__has_feature)
+    #define JSON_HEDLEY_HAS_FEATURE(feature) __has_feature(feature)
+#else
+    #define JSON_HEDLEY_HAS_FEATURE(feature) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_FEATURE)
+    #undef JSON_HEDLEY_GNUC_HAS_FEATURE
+#endif
+#if defined(__has_feature)
+    #define JSON_HEDLEY_GNUC_HAS_FEATURE(feature,major,minor,patch) __has_feature(feature)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_FEATURE(feature,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_FEATURE)
+    #undef JSON_HEDLEY_GCC_HAS_FEATURE
+#endif
+#if defined(__has_feature)
+    #define JSON_HEDLEY_GCC_HAS_FEATURE(feature,major,minor,patch) __has_feature(feature)
+#else
+    #define JSON_HEDLEY_GCC_HAS_FEATURE(feature,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_EXTENSION)
+    #undef JSON_HEDLEY_HAS_EXTENSION
+#endif
+#if defined(__has_extension)
+    #define JSON_HEDLEY_HAS_EXTENSION(extension) __has_extension(extension)
+#else
+    #define JSON_HEDLEY_HAS_EXTENSION(extension) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_EXTENSION)
+    #undef JSON_HEDLEY_GNUC_HAS_EXTENSION
+#endif
+#if defined(__has_extension)
+    #define JSON_HEDLEY_GNUC_HAS_EXTENSION(extension,major,minor,patch) __has_extension(extension)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_EXTENSION(extension,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_EXTENSION)
+    #undef JSON_HEDLEY_GCC_HAS_EXTENSION
+#endif
+#if defined(__has_extension)
+    #define JSON_HEDLEY_GCC_HAS_EXTENSION(extension,major,minor,patch) __has_extension(extension)
+#else
+    #define JSON_HEDLEY_GCC_HAS_EXTENSION(extension,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE)
+    #undef JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE
+#endif
+#if defined(__has_declspec_attribute)
+    #define JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE(attribute) __has_declspec_attribute(attribute)
+#else
+    #define JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE(attribute) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE)
+    #undef JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE
+#endif
+#if defined(__has_declspec_attribute)
+    #define JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) __has_declspec_attribute(attribute)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE)
+    #undef JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE
+#endif
+#if defined(__has_declspec_attribute)
+    #define JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) __has_declspec_attribute(attribute)
+#else
+    #define JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_WARNING)
+    #undef JSON_HEDLEY_HAS_WARNING
+#endif
+#if defined(__has_warning)
+    #define JSON_HEDLEY_HAS_WARNING(warning) __has_warning(warning)
+#else
+    #define JSON_HEDLEY_HAS_WARNING(warning) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_WARNING)
+    #undef JSON_HEDLEY_GNUC_HAS_WARNING
+#endif
+#if defined(__has_warning)
+    #define JSON_HEDLEY_GNUC_HAS_WARNING(warning,major,minor,patch) __has_warning(warning)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_WARNING(warning,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_WARNING)
+    #undef JSON_HEDLEY_GCC_HAS_WARNING
+#endif
+#if defined(__has_warning)
+    #define JSON_HEDLEY_GCC_HAS_WARNING(warning,major,minor,patch) __has_warning(warning)
+#else
+    #define JSON_HEDLEY_GCC_HAS_WARNING(warning,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if \
+    (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)) || \
+    defined(__clang__) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,0,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(18,4,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,7,0) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(2,0,1) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,1,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,0,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_CRAY_VERSION_CHECK(5,0,0) || \
+    JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,17) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(8,0,0) || \
+    (JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) && defined(__C99_PRAGMA_OPERATOR))
+    #define JSON_HEDLEY_PRAGMA(value) _Pragma(#value)
+#elif JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0)
+    #define JSON_HEDLEY_PRAGMA(value) __pragma(value)
+#else
+    #define JSON_HEDLEY_PRAGMA(value)
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_PUSH)
+    #undef JSON_HEDLEY_DIAGNOSTIC_PUSH
+#endif
+#if defined(JSON_HEDLEY_DIAGNOSTIC_POP)
+    #undef JSON_HEDLEY_DIAGNOSTIC_POP
+#endif
+#if defined(__clang__)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("clang diagnostic push")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("clang diagnostic pop")
+#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("warning(push)")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("warning(pop)")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,6,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("GCC diagnostic push")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("GCC diagnostic pop")
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH __pragma(warning(push))
+    #define JSON_HEDLEY_DIAGNOSTIC_POP __pragma(warning(pop))
+#elif JSON_HEDLEY_ARM_VERSION_CHECK(5,6,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("push")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("pop")
+#elif \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,4,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,1,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("diag_push")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("diag_pop")
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(2,90,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("warning(push)")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("warning(pop)")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH
+    #define JSON_HEDLEY_DIAGNOSTIC_POP
+#endif
+
+/* JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_ is for
+   HEDLEY INTERNAL USE ONLY.  API subject to change without notice. */
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_
+#endif
+#if defined(__cplusplus)
+#  if JSON_HEDLEY_HAS_WARNING("-Wc++98-compat")
+#    if JSON_HEDLEY_HAS_WARNING("-Wc++17-extensions")
+#      if JSON_HEDLEY_HAS_WARNING("-Wc++1z-extensions")
+#        define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(xpr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wc++98-compat\"") \
+    _Pragma("clang diagnostic ignored \"-Wc++17-extensions\"") \
+    _Pragma("clang diagnostic ignored \"-Wc++1z-extensions\"") \
+    xpr \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#      else
+#        define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(xpr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wc++98-compat\"") \
+    _Pragma("clang diagnostic ignored \"-Wc++17-extensions\"") \
+    xpr \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#      endif
+#    else
+#      define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(xpr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wc++98-compat\"") \
+    xpr \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#    endif
+#  endif
+#endif
+#if !defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(x) x
+#endif
+
+#if defined(JSON_HEDLEY_CONST_CAST)
+    #undef JSON_HEDLEY_CONST_CAST
+#endif
+#if defined(__cplusplus)
+#  define JSON_HEDLEY_CONST_CAST(T, expr) (const_cast<T>(expr))
+#elif \
+  JSON_HEDLEY_HAS_WARNING("-Wcast-qual") || \
+  JSON_HEDLEY_GCC_VERSION_CHECK(4,6,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+#  define JSON_HEDLEY_CONST_CAST(T, expr) (__extension__ ({ \
+        JSON_HEDLEY_DIAGNOSTIC_PUSH \
+        JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL \
+        ((T) (expr)); \
+        JSON_HEDLEY_DIAGNOSTIC_POP \
+    }))
+#else
+#  define JSON_HEDLEY_CONST_CAST(T, expr) ((T) (expr))
+#endif
+
+#if defined(JSON_HEDLEY_REINTERPRET_CAST)
+    #undef JSON_HEDLEY_REINTERPRET_CAST
+#endif
+#if defined(__cplusplus)
+    #define JSON_HEDLEY_REINTERPRET_CAST(T, expr) (reinterpret_cast<T>(expr))
+#else
+    #define JSON_HEDLEY_REINTERPRET_CAST(T, expr) ((T) (expr))
+#endif
+
+#if defined(JSON_HEDLEY_STATIC_CAST)
+    #undef JSON_HEDLEY_STATIC_CAST
+#endif
+#if defined(__cplusplus)
+    #define JSON_HEDLEY_STATIC_CAST(T, expr) (static_cast<T>(expr))
+#else
+    #define JSON_HEDLEY_STATIC_CAST(T, expr) ((T) (expr))
+#endif
+
+#if defined(JSON_HEDLEY_CPP_CAST)
+    #undef JSON_HEDLEY_CPP_CAST
+#endif
+#if defined(__cplusplus)
+#  if JSON_HEDLEY_HAS_WARNING("-Wold-style-cast")
+#    define JSON_HEDLEY_CPP_CAST(T, expr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wold-style-cast\"") \
+    ((T) (expr)) \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#  elif JSON_HEDLEY_IAR_VERSION_CHECK(8,3,0)
+#    define JSON_HEDLEY_CPP_CAST(T, expr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("diag_suppress=Pe137") \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#  else
+#    define JSON_HEDLEY_CPP_CAST(T, expr) ((T) (expr))
+#  endif
+#else
+#  define JSON_HEDLEY_CPP_CAST(T, expr) (expr)
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wdeprecated-declarations")
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("clang diagnostic ignored \"-Wdeprecated-declarations\"")
+#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("warning(disable:1478 1786)")
+#elif JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED __pragma(warning(disable:1478 1786))
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(20,7,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1215,1216,1444,1445")
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1215,1444")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,3,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("GCC diagnostic ignored \"-Wdeprecated-declarations\"")
+#elif JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED __pragma(warning(disable:4996))
+#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1215,1444")
+#elif \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1291,1718")
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) && !defined(__cplusplus)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("error_messages(off,E_DEPRECATED_ATT,E_DEPRECATED_ATT_MESS)")
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) && defined(__cplusplus)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("error_messages(off,symdeprecated,symdeprecated2)")
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress=Pe1444,Pe1215")
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(2,90,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("warn(disable:2241)")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wunknown-pragmas")
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("clang diagnostic ignored \"-Wunknown-pragmas\"")
+#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("warning(disable:161)")
+#elif JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS __pragma(warning(disable:161))
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 1675")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,3,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("GCC diagnostic ignored \"-Wunknown-pragmas\"")
+#elif JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS __pragma(warning(disable:4068))
+#elif \
+    JSON_HEDLEY_TI_VERSION_CHECK(16,9,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,0,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,3,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 163")
+#elif JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 163")
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress=Pe161")
+#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 161")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wunknown-attributes")
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("clang diagnostic ignored \"-Wunknown-attributes\"")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,6,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("GCC diagnostic ignored \"-Wdeprecated-declarations\"")
+#elif JSON_HEDLEY_INTEL_VERSION_CHECK(17,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("warning(disable:1292)")
+#elif JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES __pragma(warning(disable:1292))
+#elif JSON_HEDLEY_MSVC_VERSION_CHECK(19,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES __pragma(warning(disable:5030))
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(20,7,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1097,1098")
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1097")
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,14,0) && defined(__cplusplus)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("error_messages(off,attrskipunsup)")
+#elif \
+    JSON_HEDLEY_TI_VERSION_CHECK(18,1,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,3,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1173")
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress=Pe1097")
+#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1097")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wcast-qual")
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL _Pragma("clang diagnostic ignored \"-Wcast-qual\"")
+#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL _Pragma("warning(disable:2203 2331)")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(3,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL _Pragma("GCC diagnostic ignored \"-Wcast-qual\"")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wunused-function")
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION _Pragma("clang diagnostic ignored \"-Wunused-function\"")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION _Pragma("GCC diagnostic ignored \"-Wunused-function\"")
+#elif JSON_HEDLEY_MSVC_VERSION_CHECK(1,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION __pragma(warning(disable:4505))
+#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION _Pragma("diag_suppress 3142")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION
+#endif
+
+#if defined(JSON_HEDLEY_DEPRECATED)
+    #undef JSON_HEDLEY_DEPRECATED
+#endif
+#if defined(JSON_HEDLEY_DEPRECATED_FOR)
+    #undef JSON_HEDLEY_DEPRECATED_FOR
+#endif
+#if \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(14,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DEPRECATED(since) __declspec(deprecated("Since " # since))
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __declspec(deprecated("Since " #since "; use " #replacement))
+#elif \
+    (JSON_HEDLEY_HAS_EXTENSION(attribute_deprecated_with_message) && !defined(JSON_HEDLEY_IAR_VERSION)) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,5,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(5,6,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(18,1,0) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(18,1,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,3,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,3,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_DEPRECATED(since) __attribute__((__deprecated__("Since " #since)))
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __attribute__((__deprecated__("Since " #since "; use " #replacement)))
+#elif defined(__cplusplus) && (__cplusplus >= 201402L)
+    #define JSON_HEDLEY_DEPRECATED(since) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[deprecated("Since " #since)]])
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[deprecated("Since " #since "; use " #replacement)]])
+#elif \
+    JSON_HEDLEY_HAS_ATTRIBUTE(deprecated) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \
+    JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
+    #define JSON_HEDLEY_DEPRECATED(since) __attribute__((__deprecated__))
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __attribute__((__deprecated__))
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
+    JSON_HEDLEY_PELLES_VERSION_CHECK(6,50,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DEPRECATED(since) __declspec(deprecated)
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __declspec(deprecated)
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_DEPRECATED(since) _Pragma("deprecated")
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) _Pragma("deprecated")
+#else
+    #define JSON_HEDLEY_DEPRECATED(since)
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement)
+#endif
+
+#if defined(JSON_HEDLEY_UNAVAILABLE)
+    #undef JSON_HEDLEY_UNAVAILABLE
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(warning) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,3,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_UNAVAILABLE(available_since) __attribute__((__warning__("Not available until " #available_since)))
+#else
+    #define JSON_HEDLEY_UNAVAILABLE(available_since)
+#endif
+
+#if defined(JSON_HEDLEY_WARN_UNUSED_RESULT)
+    #undef JSON_HEDLEY_WARN_UNUSED_RESULT
+#endif
+#if defined(JSON_HEDLEY_WARN_UNUSED_RESULT_MSG)
+    #undef JSON_HEDLEY_WARN_UNUSED_RESULT_MSG
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(warn_unused_result) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0) && defined(__cplusplus)) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT __attribute__((__warn_unused_result__))
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) __attribute__((__warn_unused_result__))
+#elif (JSON_HEDLEY_HAS_CPP_ATTRIBUTE(nodiscard) >= 201907L)
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard]])
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard(msg)]])
+#elif JSON_HEDLEY_HAS_CPP_ATTRIBUTE(nodiscard)
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard]])
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard]])
+#elif defined(_Check_return_) /* SAL */
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT _Check_return_
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) _Check_return_
+#else
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg)
+#endif
+
+#if defined(JSON_HEDLEY_SENTINEL)
+    #undef JSON_HEDLEY_SENTINEL
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(sentinel) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,0,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(5,4,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_SENTINEL(position) __attribute__((__sentinel__(position)))
+#else
+    #define JSON_HEDLEY_SENTINEL(position)
+#endif
+
+#if defined(JSON_HEDLEY_NO_RETURN)
+    #undef JSON_HEDLEY_NO_RETURN
+#endif
+#if JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_NO_RETURN __noreturn
+#elif \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_NO_RETURN __attribute__((__noreturn__))
+#elif defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L
+    #define JSON_HEDLEY_NO_RETURN _Noreturn
+#elif defined(__cplusplus) && (__cplusplus >= 201103L)
+    #define JSON_HEDLEY_NO_RETURN JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[noreturn]])
+#elif \
+    JSON_HEDLEY_HAS_ATTRIBUTE(noreturn) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,2,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
+    #define JSON_HEDLEY_NO_RETURN __attribute__((__noreturn__))
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
+    #define JSON_HEDLEY_NO_RETURN _Pragma("does_not_return")
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_NO_RETURN __declspec(noreturn)
+#elif JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,0,0) && defined(__cplusplus)
+    #define JSON_HEDLEY_NO_RETURN _Pragma("FUNC_NEVER_RETURNS;")
+#elif JSON_HEDLEY_COMPCERT_VERSION_CHECK(3,2,0)
+    #define JSON_HEDLEY_NO_RETURN __attribute((noreturn))
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(9,0,0)
+    #define JSON_HEDLEY_NO_RETURN __declspec(noreturn)
+#else
+    #define JSON_HEDLEY_NO_RETURN
+#endif
+
+#if defined(JSON_HEDLEY_NO_ESCAPE)
+    #undef JSON_HEDLEY_NO_ESCAPE
+#endif
+#if JSON_HEDLEY_HAS_ATTRIBUTE(noescape)
+    #define JSON_HEDLEY_NO_ESCAPE __attribute__((__noescape__))
+#else
+    #define JSON_HEDLEY_NO_ESCAPE
+#endif
+
+#if defined(JSON_HEDLEY_UNREACHABLE)
+    #undef JSON_HEDLEY_UNREACHABLE
+#endif
+#if defined(JSON_HEDLEY_UNREACHABLE_RETURN)
+    #undef JSON_HEDLEY_UNREACHABLE_RETURN
+#endif
+#if defined(JSON_HEDLEY_ASSUME)
+    #undef JSON_HEDLEY_ASSUME
+#endif
+#if \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_ASSUME(expr) __assume(expr)
+#elif JSON_HEDLEY_HAS_BUILTIN(__builtin_assume)
+    #define JSON_HEDLEY_ASSUME(expr) __builtin_assume(expr)
+#elif \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(4,0,0)
+    #if defined(__cplusplus)
+        #define JSON_HEDLEY_ASSUME(expr) std::_nassert(expr)
+    #else
+        #define JSON_HEDLEY_ASSUME(expr) _nassert(expr)
+    #endif
+#endif
+#if \
+    (JSON_HEDLEY_HAS_BUILTIN(__builtin_unreachable) && (!defined(JSON_HEDLEY_ARM_VERSION))) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,5,0) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(18,10,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(13,1,5) || \
+    JSON_HEDLEY_CRAY_VERSION_CHECK(10,0,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_UNREACHABLE() __builtin_unreachable()
+#elif defined(JSON_HEDLEY_ASSUME)
+    #define JSON_HEDLEY_UNREACHABLE() JSON_HEDLEY_ASSUME(0)
+#endif
+#if !defined(JSON_HEDLEY_ASSUME)
+    #if defined(JSON_HEDLEY_UNREACHABLE)
+        #define JSON_HEDLEY_ASSUME(expr) JSON_HEDLEY_STATIC_CAST(void, ((expr) ? 1 : (JSON_HEDLEY_UNREACHABLE(), 1)))
+    #else
+        #define JSON_HEDLEY_ASSUME(expr) JSON_HEDLEY_STATIC_CAST(void, expr)
+    #endif
+#endif
+#if defined(JSON_HEDLEY_UNREACHABLE)
+    #if  \
+        JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,0) || \
+        JSON_HEDLEY_TI_CL6X_VERSION_CHECK(4,0,0)
+        #define JSON_HEDLEY_UNREACHABLE_RETURN(value) return (JSON_HEDLEY_STATIC_CAST(void, JSON_HEDLEY_ASSUME(0)), (value))
+    #else
+        #define JSON_HEDLEY_UNREACHABLE_RETURN(value) JSON_HEDLEY_UNREACHABLE()
+    #endif
+#else
+    #define JSON_HEDLEY_UNREACHABLE_RETURN(value) return (value)
+#endif
+#if !defined(JSON_HEDLEY_UNREACHABLE)
+    #define JSON_HEDLEY_UNREACHABLE() JSON_HEDLEY_ASSUME(0)
+#endif
+
+JSON_HEDLEY_DIAGNOSTIC_PUSH
+#if JSON_HEDLEY_HAS_WARNING("-Wpedantic")
+    #pragma clang diagnostic ignored "-Wpedantic"
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wc++98-compat-pedantic") && defined(__cplusplus)
+    #pragma clang diagnostic ignored "-Wc++98-compat-pedantic"
+#endif
+#if JSON_HEDLEY_GCC_HAS_WARNING("-Wvariadic-macros",4,0,0)
+    #if defined(__clang__)
+        #pragma clang diagnostic ignored "-Wvariadic-macros"
+    #elif defined(JSON_HEDLEY_GCC_VERSION)
+        #pragma GCC diagnostic ignored "-Wvariadic-macros"
+    #endif
+#endif
+#if defined(JSON_HEDLEY_NON_NULL)
+    #undef JSON_HEDLEY_NON_NULL
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(nonnull) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,3,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0)
+    #define JSON_HEDLEY_NON_NULL(...) __attribute__((__nonnull__(__VA_ARGS__)))
+#else
+    #define JSON_HEDLEY_NON_NULL(...)
+#endif
+JSON_HEDLEY_DIAGNOSTIC_POP
+
+#if defined(JSON_HEDLEY_PRINTF_FORMAT)
+    #undef JSON_HEDLEY_PRINTF_FORMAT
+#endif
+#if defined(__MINGW32__) && JSON_HEDLEY_GCC_HAS_ATTRIBUTE(format,4,4,0) && !defined(__USE_MINGW_ANSI_STDIO)
+    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __attribute__((__format__(ms_printf, string_idx, first_to_check)))
+#elif defined(__MINGW32__) && JSON_HEDLEY_GCC_HAS_ATTRIBUTE(format,4,4,0) && defined(__USE_MINGW_ANSI_STDIO)
+    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __attribute__((__format__(gnu_printf, string_idx, first_to_check)))
+#elif \
+    JSON_HEDLEY_HAS_ATTRIBUTE(format) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(5,6,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __attribute__((__format__(__printf__, string_idx, first_to_check)))
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(6,0,0)
+    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __declspec(vaformat(printf,string_idx,first_to_check))
+#else
+    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check)
+#endif
+
+#if defined(JSON_HEDLEY_CONSTEXPR)
+    #undef JSON_HEDLEY_CONSTEXPR
+#endif
+#if defined(__cplusplus)
+    #if __cplusplus >= 201103L
+        #define JSON_HEDLEY_CONSTEXPR JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(constexpr)
+    #endif
+#endif
+#if !defined(JSON_HEDLEY_CONSTEXPR)
+    #define JSON_HEDLEY_CONSTEXPR
+#endif
+
+#if defined(JSON_HEDLEY_PREDICT)
+    #undef JSON_HEDLEY_PREDICT
+#endif
+#if defined(JSON_HEDLEY_LIKELY)
+    #undef JSON_HEDLEY_LIKELY
+#endif
+#if defined(JSON_HEDLEY_UNLIKELY)
+    #undef JSON_HEDLEY_UNLIKELY
+#endif
+#if defined(JSON_HEDLEY_UNPREDICTABLE)
+    #undef JSON_HEDLEY_UNPREDICTABLE
+#endif
+#if JSON_HEDLEY_HAS_BUILTIN(__builtin_unpredictable)
+    #define JSON_HEDLEY_UNPREDICTABLE(expr) __builtin_unpredictable((expr))
+#endif
+#if \
+  (JSON_HEDLEY_HAS_BUILTIN(__builtin_expect_with_probability) && !defined(JSON_HEDLEY_PGI_VERSION)) || \
+  JSON_HEDLEY_GCC_VERSION_CHECK(9,0,0) || \
+  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+#  define JSON_HEDLEY_PREDICT(expr, value, probability) __builtin_expect_with_probability(  (expr), (value), (probability))
+#  define JSON_HEDLEY_PREDICT_TRUE(expr, probability)   __builtin_expect_with_probability(!!(expr),    1   , (probability))
+#  define JSON_HEDLEY_PREDICT_FALSE(expr, probability)  __builtin_expect_with_probability(!!(expr),    0   , (probability))
+#  define JSON_HEDLEY_LIKELY(expr)                      __builtin_expect                 (!!(expr),    1                  )
+#  define JSON_HEDLEY_UNLIKELY(expr)                    __builtin_expect                 (!!(expr),    0                  )
+#elif \
+  (JSON_HEDLEY_HAS_BUILTIN(__builtin_expect) && !defined(JSON_HEDLEY_INTEL_CL_VERSION)) || \
+  JSON_HEDLEY_GCC_VERSION_CHECK(3,0,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+  (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0) && defined(__cplusplus)) || \
+  JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+  JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+  JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+  JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,7,0) || \
+  JSON_HEDLEY_TI_CL430_VERSION_CHECK(3,1,0) || \
+  JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,1,0) || \
+  JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,1,0) || \
+  JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+  JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+  JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,27) || \
+  JSON_HEDLEY_CRAY_VERSION_CHECK(8,1,0) || \
+  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+#  define JSON_HEDLEY_PREDICT(expr, expected, probability) \
+    (((probability) >= 0.9) ? __builtin_expect((expr), (expected)) : (JSON_HEDLEY_STATIC_CAST(void, expected), (expr)))
+#  define JSON_HEDLEY_PREDICT_TRUE(expr, probability) \
+    (__extension__ ({ \
+        double hedley_probability_ = (probability); \
+        ((hedley_probability_ >= 0.9) ? __builtin_expect(!!(expr), 1) : ((hedley_probability_ <= 0.1) ? __builtin_expect(!!(expr), 0) : !!(expr))); \
+    }))
+#  define JSON_HEDLEY_PREDICT_FALSE(expr, probability) \
+    (__extension__ ({ \
+        double hedley_probability_ = (probability); \
+        ((hedley_probability_ >= 0.9) ? __builtin_expect(!!(expr), 0) : ((hedley_probability_ <= 0.1) ? __builtin_expect(!!(expr), 1) : !!(expr))); \
+    }))
+#  define JSON_HEDLEY_LIKELY(expr)   __builtin_expect(!!(expr), 1)
+#  define JSON_HEDLEY_UNLIKELY(expr) __builtin_expect(!!(expr), 0)
+#else
+#  define JSON_HEDLEY_PREDICT(expr, expected, probability) (JSON_HEDLEY_STATIC_CAST(void, expected), (expr))
+#  define JSON_HEDLEY_PREDICT_TRUE(expr, probability) (!!(expr))
+#  define JSON_HEDLEY_PREDICT_FALSE(expr, probability) (!!(expr))
+#  define JSON_HEDLEY_LIKELY(expr) (!!(expr))
+#  define JSON_HEDLEY_UNLIKELY(expr) (!!(expr))
+#endif
+#if !defined(JSON_HEDLEY_UNPREDICTABLE)
+    #define JSON_HEDLEY_UNPREDICTABLE(expr) JSON_HEDLEY_PREDICT(expr, 1, 0.5)
+#endif
+
+#if defined(JSON_HEDLEY_MALLOC)
+    #undef JSON_HEDLEY_MALLOC
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(malloc) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(12,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_MALLOC __attribute__((__malloc__))
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
+    #define JSON_HEDLEY_MALLOC _Pragma("returns_new_memory")
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(14,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_MALLOC __declspec(restrict)
+#else
+    #define JSON_HEDLEY_MALLOC
+#endif
+
+#if defined(JSON_HEDLEY_PURE)
+    #undef JSON_HEDLEY_PURE
+#endif
+#if \
+  JSON_HEDLEY_HAS_ATTRIBUTE(pure) || \
+  JSON_HEDLEY_GCC_VERSION_CHECK(2,96,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+  JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+  JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+  JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+  JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+  (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+  (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+  (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+  (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+  JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+  JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+  JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
+  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+#  define JSON_HEDLEY_PURE __attribute__((__pure__))
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
+#  define JSON_HEDLEY_PURE _Pragma("does_not_write_global_data")
+#elif defined(__cplusplus) && \
+    ( \
+      JSON_HEDLEY_TI_CL430_VERSION_CHECK(2,0,1) || \
+      JSON_HEDLEY_TI_CL6X_VERSION_CHECK(4,0,0) || \
+      JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) \
+    )
+#  define JSON_HEDLEY_PURE _Pragma("FUNC_IS_PURE;")
+#else
+#  define JSON_HEDLEY_PURE
+#endif
+
+#if defined(JSON_HEDLEY_CONST)
+    #undef JSON_HEDLEY_CONST
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(const) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(2,5,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_CONST __attribute__((__const__))
+#elif \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
+    #define JSON_HEDLEY_CONST _Pragma("no_side_effect")
+#else
+    #define JSON_HEDLEY_CONST JSON_HEDLEY_PURE
+#endif
+
+#if defined(JSON_HEDLEY_RESTRICT)
+    #undef JSON_HEDLEY_RESTRICT
+#endif
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) && !defined(__cplusplus)
+    #define JSON_HEDLEY_RESTRICT restrict
+#elif \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(14,0,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,4) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,1,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,14,0) && defined(__cplusplus)) || \
+    JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) || \
+    defined(__clang__) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_RESTRICT __restrict
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,3,0) && !defined(__cplusplus)
+    #define JSON_HEDLEY_RESTRICT _Restrict
+#else
+    #define JSON_HEDLEY_RESTRICT
+#endif
+
+#if defined(JSON_HEDLEY_INLINE)
+    #undef JSON_HEDLEY_INLINE
+#endif
+#if \
+    (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)) || \
+    (defined(__cplusplus) && (__cplusplus >= 199711L))
+    #define JSON_HEDLEY_INLINE inline
+#elif \
+    defined(JSON_HEDLEY_GCC_VERSION) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(6,2,0)
+    #define JSON_HEDLEY_INLINE __inline__
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(12,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,1,0) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(3,1,0) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,0,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_INLINE __inline
+#else
+    #define JSON_HEDLEY_INLINE
+#endif
+
+#if defined(JSON_HEDLEY_ALWAYS_INLINE)
+    #undef JSON_HEDLEY_ALWAYS_INLINE
+#endif
+#if \
+  JSON_HEDLEY_HAS_ATTRIBUTE(always_inline) || \
+  JSON_HEDLEY_GCC_VERSION_CHECK(4,0,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+  JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+  JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+  JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+  JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+  (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+  (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+  (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+  (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+  JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+  JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \
+  JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
+#  define JSON_HEDLEY_ALWAYS_INLINE __attribute__((__always_inline__)) JSON_HEDLEY_INLINE
+#elif \
+  JSON_HEDLEY_MSVC_VERSION_CHECK(12,0,0) || \
+  JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+#  define JSON_HEDLEY_ALWAYS_INLINE __forceinline
+#elif defined(__cplusplus) && \
+    ( \
+      JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+      JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+      JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+      JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,1,0) || \
+      JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+      JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) \
+    )
+#  define JSON_HEDLEY_ALWAYS_INLINE _Pragma("FUNC_ALWAYS_INLINE;")
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+#  define JSON_HEDLEY_ALWAYS_INLINE _Pragma("inline=forced")
+#else
+#  define JSON_HEDLEY_ALWAYS_INLINE JSON_HEDLEY_INLINE
+#endif
+
+#if defined(JSON_HEDLEY_NEVER_INLINE)
+    #undef JSON_HEDLEY_NEVER_INLINE
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(noinline) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,0,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \
+    JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
+    #define JSON_HEDLEY_NEVER_INLINE __attribute__((__noinline__))
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_NEVER_INLINE __declspec(noinline)
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(10,2,0)
+    #define JSON_HEDLEY_NEVER_INLINE _Pragma("noinline")
+#elif JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,0,0) && defined(__cplusplus)
+    #define JSON_HEDLEY_NEVER_INLINE _Pragma("FUNC_CANNOT_INLINE;")
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_NEVER_INLINE _Pragma("inline=never")
+#elif JSON_HEDLEY_COMPCERT_VERSION_CHECK(3,2,0)
+    #define JSON_HEDLEY_NEVER_INLINE __attribute((noinline))
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(9,0,0)
+    #define JSON_HEDLEY_NEVER_INLINE __declspec(noinline)
+#else
+    #define JSON_HEDLEY_NEVER_INLINE
+#endif
+
+#if defined(JSON_HEDLEY_PRIVATE)
+    #undef JSON_HEDLEY_PRIVATE
+#endif
+#if defined(JSON_HEDLEY_PUBLIC)
+    #undef JSON_HEDLEY_PUBLIC
+#endif
+#if defined(JSON_HEDLEY_IMPORT)
+    #undef JSON_HEDLEY_IMPORT
+#endif
+#if defined(_WIN32) || defined(__CYGWIN__)
+#  define JSON_HEDLEY_PRIVATE
+#  define JSON_HEDLEY_PUBLIC   __declspec(dllexport)
+#  define JSON_HEDLEY_IMPORT   __declspec(dllimport)
+#else
+#  if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(visibility) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,3,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(13,1,0) || \
+    ( \
+      defined(__TI_EABI__) && \
+      ( \
+        (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+        JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) \
+      ) \
+    ) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+#    define JSON_HEDLEY_PRIVATE __attribute__((__visibility__("hidden")))
+#    define JSON_HEDLEY_PUBLIC  __attribute__((__visibility__("default")))
+#  else
+#    define JSON_HEDLEY_PRIVATE
+#    define JSON_HEDLEY_PUBLIC
+#  endif
+#  define JSON_HEDLEY_IMPORT    extern
+#endif
+
+#if defined(JSON_HEDLEY_NO_THROW)
+    #undef JSON_HEDLEY_NO_THROW
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(nothrow) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,3,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_NO_THROW __attribute__((__nothrow__))
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(13,1,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0)
+    #define JSON_HEDLEY_NO_THROW __declspec(nothrow)
+#else
+    #define JSON_HEDLEY_NO_THROW
+#endif
+
+#if defined(JSON_HEDLEY_FALL_THROUGH)
+    #undef JSON_HEDLEY_FALL_THROUGH
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(fallthrough) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(7,0,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_FALL_THROUGH __attribute__((__fallthrough__))
+#elif JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(clang,fallthrough)
+    #define JSON_HEDLEY_FALL_THROUGH JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[clang::fallthrough]])
+#elif JSON_HEDLEY_HAS_CPP_ATTRIBUTE(fallthrough)
+    #define JSON_HEDLEY_FALL_THROUGH JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[fallthrough]])
+#elif defined(__fallthrough) /* SAL */
+    #define JSON_HEDLEY_FALL_THROUGH __fallthrough
+#else
+    #define JSON_HEDLEY_FALL_THROUGH
+#endif
+
+#if defined(JSON_HEDLEY_RETURNS_NON_NULL)
+    #undef JSON_HEDLEY_RETURNS_NON_NULL
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(returns_nonnull) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_RETURNS_NON_NULL __attribute__((__returns_nonnull__))
+#elif defined(_Ret_notnull_) /* SAL */
+    #define JSON_HEDLEY_RETURNS_NON_NULL _Ret_notnull_
+#else
+    #define JSON_HEDLEY_RETURNS_NON_NULL
+#endif
+
+#if defined(JSON_HEDLEY_ARRAY_PARAM)
+    #undef JSON_HEDLEY_ARRAY_PARAM
+#endif
+#if \
+    defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) && \
+    !defined(__STDC_NO_VLA__) && \
+    !defined(__cplusplus) && \
+    !defined(JSON_HEDLEY_PGI_VERSION) && \
+    !defined(JSON_HEDLEY_TINYC_VERSION)
+    #define JSON_HEDLEY_ARRAY_PARAM(name) (name)
+#else
+    #define JSON_HEDLEY_ARRAY_PARAM(name)
+#endif
+
+#if defined(JSON_HEDLEY_IS_CONSTANT)
+    #undef JSON_HEDLEY_IS_CONSTANT
+#endif
+#if defined(JSON_HEDLEY_REQUIRE_CONSTEXPR)
+    #undef JSON_HEDLEY_REQUIRE_CONSTEXPR
+#endif
+/* JSON_HEDLEY_IS_CONSTEXPR_ is for
+   HEDLEY INTERNAL USE ONLY.  API subject to change without notice. */
+#if defined(JSON_HEDLEY_IS_CONSTEXPR_)
+    #undef JSON_HEDLEY_IS_CONSTEXPR_
+#endif
+#if \
+    JSON_HEDLEY_HAS_BUILTIN(__builtin_constant_p) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,19) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(13,1,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,1,0) || \
+    (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0) && !defined(__cplusplus)) || \
+    JSON_HEDLEY_CRAY_VERSION_CHECK(8,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_IS_CONSTANT(expr) __builtin_constant_p(expr)
+#endif
+#if !defined(__cplusplus)
+#  if \
+       JSON_HEDLEY_HAS_BUILTIN(__builtin_types_compatible_p) || \
+       JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) || \
+       JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+       JSON_HEDLEY_IBM_VERSION_CHECK(13,1,0) || \
+       JSON_HEDLEY_CRAY_VERSION_CHECK(8,1,0) || \
+       JSON_HEDLEY_ARM_VERSION_CHECK(5,4,0) || \
+       JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,24)
+#if defined(__INTPTR_TYPE__)
+    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) __builtin_types_compatible_p(__typeof__((1 ? (void*) ((__INTPTR_TYPE__) ((expr) * 0)) : (int*) 0)), int*)
+#else
+    #include <stdint.h>
+    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) __builtin_types_compatible_p(__typeof__((1 ? (void*) ((intptr_t) ((expr) * 0)) : (int*) 0)), int*)
+#endif
+#  elif \
+       ( \
+          defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) && \
+          !defined(JSON_HEDLEY_SUNPRO_VERSION) && \
+          !defined(JSON_HEDLEY_PGI_VERSION) && \
+          !defined(JSON_HEDLEY_IAR_VERSION)) || \
+       (JSON_HEDLEY_HAS_EXTENSION(c_generic_selections) && !defined(JSON_HEDLEY_IAR_VERSION)) || \
+       JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0) || \
+       JSON_HEDLEY_INTEL_VERSION_CHECK(17,0,0) || \
+       JSON_HEDLEY_IBM_VERSION_CHECK(12,1,0) || \
+       JSON_HEDLEY_ARM_VERSION_CHECK(5,3,0)
+#if defined(__INTPTR_TYPE__)
+    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) _Generic((1 ? (void*) ((__INTPTR_TYPE__) ((expr) * 0)) : (int*) 0), int*: 1, void*: 0)
+#else
+    #include <stdint.h>
+    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) _Generic((1 ? (void*) ((intptr_t) * 0) : (int*) 0), int*: 1, void*: 0)
+#endif
+#  elif \
+       defined(JSON_HEDLEY_GCC_VERSION) || \
+       defined(JSON_HEDLEY_INTEL_VERSION) || \
+       defined(JSON_HEDLEY_TINYC_VERSION) || \
+       defined(JSON_HEDLEY_TI_ARMCL_VERSION) || \
+       JSON_HEDLEY_TI_CL430_VERSION_CHECK(18,12,0) || \
+       defined(JSON_HEDLEY_TI_CL2000_VERSION) || \
+       defined(JSON_HEDLEY_TI_CL6X_VERSION) || \
+       defined(JSON_HEDLEY_TI_CL7X_VERSION) || \
+       defined(JSON_HEDLEY_TI_CLPRU_VERSION) || \
+       defined(__clang__)
+#    define JSON_HEDLEY_IS_CONSTEXPR_(expr) ( \
+        sizeof(void) != \
+        sizeof(*( \
+                  1 ? \
+                  ((void*) ((expr) * 0L) ) : \
+((struct { char v[sizeof(void) * 2]; } *) 1) \
+                ) \
+              ) \
+                                            )
+#  endif
+#endif
+#if defined(JSON_HEDLEY_IS_CONSTEXPR_)
+    #if !defined(JSON_HEDLEY_IS_CONSTANT)
+        #define JSON_HEDLEY_IS_CONSTANT(expr) JSON_HEDLEY_IS_CONSTEXPR_(expr)
+    #endif
+    #define JSON_HEDLEY_REQUIRE_CONSTEXPR(expr) (JSON_HEDLEY_IS_CONSTEXPR_(expr) ? (expr) : (-1))
+#else
+    #if !defined(JSON_HEDLEY_IS_CONSTANT)
+        #define JSON_HEDLEY_IS_CONSTANT(expr) (0)
+    #endif
+    #define JSON_HEDLEY_REQUIRE_CONSTEXPR(expr) (expr)
+#endif
+
+#if defined(JSON_HEDLEY_BEGIN_C_DECLS)
+    #undef JSON_HEDLEY_BEGIN_C_DECLS
+#endif
+#if defined(JSON_HEDLEY_END_C_DECLS)
+    #undef JSON_HEDLEY_END_C_DECLS
+#endif
+#if defined(JSON_HEDLEY_C_DECL)
+    #undef JSON_HEDLEY_C_DECL
+#endif
+#if defined(__cplusplus)
+    #define JSON_HEDLEY_BEGIN_C_DECLS extern "C" {
+    #define JSON_HEDLEY_END_C_DECLS }
+    #define JSON_HEDLEY_C_DECL extern "C"
+#else
+    #define JSON_HEDLEY_BEGIN_C_DECLS
+    #define JSON_HEDLEY_END_C_DECLS
+    #define JSON_HEDLEY_C_DECL
+#endif
+
+#if defined(JSON_HEDLEY_STATIC_ASSERT)
+    #undef JSON_HEDLEY_STATIC_ASSERT
+#endif
+#if \
+  !defined(__cplusplus) && ( \
+      (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)) || \
+      (JSON_HEDLEY_HAS_FEATURE(c_static_assert) && !defined(JSON_HEDLEY_INTEL_CL_VERSION)) || \
+      JSON_HEDLEY_GCC_VERSION_CHECK(6,0,0) || \
+      JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+      defined(_Static_assert) \
+    )
+#  define JSON_HEDLEY_STATIC_ASSERT(expr, message) _Static_assert(expr, message)
+#elif \
+  (defined(__cplusplus) && (__cplusplus >= 201103L)) || \
+  JSON_HEDLEY_MSVC_VERSION_CHECK(16,0,0) || \
+  JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+#  define JSON_HEDLEY_STATIC_ASSERT(expr, message) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(static_assert(expr, message))
+#else
+#  define JSON_HEDLEY_STATIC_ASSERT(expr, message)
+#endif
+
+#if defined(JSON_HEDLEY_NULL)
+    #undef JSON_HEDLEY_NULL
+#endif
+#if defined(__cplusplus)
+    #if __cplusplus >= 201103L
+        #define JSON_HEDLEY_NULL JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(nullptr)
+    #elif defined(NULL)
+        #define JSON_HEDLEY_NULL NULL
+    #else
+        #define JSON_HEDLEY_NULL JSON_HEDLEY_STATIC_CAST(void*, 0)
+    #endif
+#elif defined(NULL)
+    #define JSON_HEDLEY_NULL NULL
+#else
+    #define JSON_HEDLEY_NULL ((void*) 0)
+#endif
+
+#if defined(JSON_HEDLEY_MESSAGE)
+    #undef JSON_HEDLEY_MESSAGE
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wunknown-pragmas")
+#  define JSON_HEDLEY_MESSAGE(msg) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS \
+    JSON_HEDLEY_PRAGMA(message msg) \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#elif \
+  JSON_HEDLEY_GCC_VERSION_CHECK(4,4,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(message msg)
+#elif JSON_HEDLEY_CRAY_VERSION_CHECK(5,0,0)
+#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(_CRI message msg)
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(message(msg))
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(2,0,0)
+#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(message(msg))
+#else
+#  define JSON_HEDLEY_MESSAGE(msg)
+#endif
+
+#if defined(JSON_HEDLEY_WARNING)
+    #undef JSON_HEDLEY_WARNING
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wunknown-pragmas")
+#  define JSON_HEDLEY_WARNING(msg) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS \
+    JSON_HEDLEY_PRAGMA(clang warning msg) \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#elif \
+  JSON_HEDLEY_GCC_VERSION_CHECK(4,8,0) || \
+  JSON_HEDLEY_PGI_VERSION_CHECK(18,4,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+#  define JSON_HEDLEY_WARNING(msg) JSON_HEDLEY_PRAGMA(GCC warning msg)
+#elif \
+  JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0) || \
+  JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+#  define JSON_HEDLEY_WARNING(msg) JSON_HEDLEY_PRAGMA(message(msg))
+#else
+#  define JSON_HEDLEY_WARNING(msg) JSON_HEDLEY_MESSAGE(msg)
+#endif
+
+#if defined(JSON_HEDLEY_REQUIRE)
+    #undef JSON_HEDLEY_REQUIRE
+#endif
+#if defined(JSON_HEDLEY_REQUIRE_MSG)
+    #undef JSON_HEDLEY_REQUIRE_MSG
+#endif
+#if JSON_HEDLEY_HAS_ATTRIBUTE(diagnose_if)
+#  if JSON_HEDLEY_HAS_WARNING("-Wgcc-compat")
+#    define JSON_HEDLEY_REQUIRE(expr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wgcc-compat\"") \
+    __attribute__((diagnose_if(!(expr), #expr, "error"))) \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#    define JSON_HEDLEY_REQUIRE_MSG(expr,msg) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wgcc-compat\"") \
+    __attribute__((diagnose_if(!(expr), msg, "error"))) \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#  else
+#    define JSON_HEDLEY_REQUIRE(expr) __attribute__((diagnose_if(!(expr), #expr, "error")))
+#    define JSON_HEDLEY_REQUIRE_MSG(expr,msg) __attribute__((diagnose_if(!(expr), msg, "error")))
+#  endif
+#else
+#  define JSON_HEDLEY_REQUIRE(expr)
+#  define JSON_HEDLEY_REQUIRE_MSG(expr,msg)
+#endif
+
+#if defined(JSON_HEDLEY_FLAGS)
+    #undef JSON_HEDLEY_FLAGS
+#endif
+#if JSON_HEDLEY_HAS_ATTRIBUTE(flag_enum) && (!defined(__cplusplus) || JSON_HEDLEY_HAS_WARNING("-Wbitfield-enum-conversion"))
+    #define JSON_HEDLEY_FLAGS __attribute__((__flag_enum__))
+#else
+    #define JSON_HEDLEY_FLAGS
+#endif
+
+#if defined(JSON_HEDLEY_FLAGS_CAST)
+    #undef JSON_HEDLEY_FLAGS_CAST
+#endif
+#if JSON_HEDLEY_INTEL_VERSION_CHECK(19,0,0)
+#  define JSON_HEDLEY_FLAGS_CAST(T, expr) (__extension__ ({ \
+        JSON_HEDLEY_DIAGNOSTIC_PUSH \
+        _Pragma("warning(disable:188)") \
+        ((T) (expr)); \
+        JSON_HEDLEY_DIAGNOSTIC_POP \
+    }))
+#else
+#  define JSON_HEDLEY_FLAGS_CAST(T, expr) JSON_HEDLEY_STATIC_CAST(T, expr)
+#endif
+
+#if defined(JSON_HEDLEY_EMPTY_BASES)
+    #undef JSON_HEDLEY_EMPTY_BASES
+#endif
+#if \
+    (JSON_HEDLEY_MSVC_VERSION_CHECK(19,0,23918) && !JSON_HEDLEY_MSVC_VERSION_CHECK(20,0,0)) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_EMPTY_BASES __declspec(empty_bases)
+#else
+    #define JSON_HEDLEY_EMPTY_BASES
+#endif
+
+/* Remaining macros are deprecated. */
+
+#if defined(JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK)
+    #undef JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK
+#endif
+#if defined(__clang__)
+    #define JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK(major,minor,patch) (0)
+#else
+    #define JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK(major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_CLANG_HAS_ATTRIBUTE)
+    #undef JSON_HEDLEY_CLANG_HAS_ATTRIBUTE
+#endif
+#define JSON_HEDLEY_CLANG_HAS_ATTRIBUTE(attribute) JSON_HEDLEY_HAS_ATTRIBUTE(attribute)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE)
+    #undef JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE
+#endif
+#define JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE(attribute) JSON_HEDLEY_HAS_CPP_ATTRIBUTE(attribute)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_BUILTIN)
+    #undef JSON_HEDLEY_CLANG_HAS_BUILTIN
+#endif
+#define JSON_HEDLEY_CLANG_HAS_BUILTIN(builtin) JSON_HEDLEY_HAS_BUILTIN(builtin)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_FEATURE)
+    #undef JSON_HEDLEY_CLANG_HAS_FEATURE
+#endif
+#define JSON_HEDLEY_CLANG_HAS_FEATURE(feature) JSON_HEDLEY_HAS_FEATURE(feature)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_EXTENSION)
+    #undef JSON_HEDLEY_CLANG_HAS_EXTENSION
+#endif
+#define JSON_HEDLEY_CLANG_HAS_EXTENSION(extension) JSON_HEDLEY_HAS_EXTENSION(extension)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_DECLSPEC_DECLSPEC_ATTRIBUTE)
+    #undef JSON_HEDLEY_CLANG_HAS_DECLSPEC_DECLSPEC_ATTRIBUTE
+#endif
+#define JSON_HEDLEY_CLANG_HAS_DECLSPEC_ATTRIBUTE(attribute) JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE(attribute)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_WARNING)
+    #undef JSON_HEDLEY_CLANG_HAS_WARNING
+#endif
+#define JSON_HEDLEY_CLANG_HAS_WARNING(warning) JSON_HEDLEY_HAS_WARNING(warning)
+
+#endif /* !defined(JSON_HEDLEY_VERSION) || (JSON_HEDLEY_VERSION < X) */
+
+
+// This file contains all internal macro definitions (except those affecting ABI)
+// You MUST include macro_unscope.hpp at the end of json.hpp to undef all of them
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+
+// exclude unsupported compilers
+#if !defined(JSON_SKIP_UNSUPPORTED_COMPILER_CHECK)
+    #if defined(__clang__)
+        #if (__clang_major__ * 10000 + __clang_minor__ * 100 + __clang_patchlevel__) < 30400
+            #error "unsupported Clang version - see https://github.com/nlohmann/json#supported-compilers"
+        #endif
+    #elif defined(__GNUC__) && !(defined(__ICC) || defined(__INTEL_COMPILER))
+        #if (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__) < 40800
+            #error "unsupported GCC version - see https://github.com/nlohmann/json#supported-compilers"
+        #endif
+    #endif
+#endif
+
+// C++ language standard detection
+// if the user manually specified the used c++ version this is skipped
+#if !defined(JSON_HAS_CPP_20) && !defined(JSON_HAS_CPP_17) && !defined(JSON_HAS_CPP_14) && !defined(JSON_HAS_CPP_11)
+    #if (defined(__cplusplus) && __cplusplus >= 202002L) || (defined(_MSVC_LANG) && _MSVC_LANG >= 202002L)
+        #define JSON_HAS_CPP_20
+        #define JSON_HAS_CPP_17
+        #define JSON_HAS_CPP_14
+    #elif (defined(__cplusplus) && __cplusplus >= 201703L) || (defined(_HAS_CXX17) && _HAS_CXX17 == 1) // fix for issue #464
+        #define JSON_HAS_CPP_17
+        #define JSON_HAS_CPP_14
+    #elif (defined(__cplusplus) && __cplusplus >= 201402L) || (defined(_HAS_CXX14) && _HAS_CXX14 == 1)
+        #define JSON_HAS_CPP_14
+    #endif
+    // the cpp 11 flag is always specified because it is the minimal required version
+    #define JSON_HAS_CPP_11
+#endif
+
+#ifdef __has_include
+    #if __has_include(<version>)
+        #include <version>
+    #endif
+#endif
+
+#if !defined(JSON_HAS_FILESYSTEM) && !defined(JSON_HAS_EXPERIMENTAL_FILESYSTEM)
+    #ifdef JSON_HAS_CPP_17
+        #if defined(__cpp_lib_filesystem)
+            #define JSON_HAS_FILESYSTEM 1
+        #elif defined(__cpp_lib_experimental_filesystem)
+            #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 1
+        #elif !defined(__has_include)
+            #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 1
+        #elif __has_include(<filesystem>)
+            #define JSON_HAS_FILESYSTEM 1
+        #elif __has_include(<experimental/filesystem>)
+            #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 1
+        #endif
+
+        // std::filesystem does not work on MinGW GCC 8: https://sourceforge.net/p/mingw-w64/bugs/737/
+        #if defined(__MINGW32__) && defined(__GNUC__) && __GNUC__ == 8
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+
+        // no filesystem support before GCC 8: https://en.cppreference.com/w/cpp/compiler_support
+        #if defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 8
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+
+        // no filesystem support before Clang 7: https://en.cppreference.com/w/cpp/compiler_support
+        #if defined(__clang_major__) && __clang_major__ < 7
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+
+        // no filesystem support before MSVC 19.14: https://en.cppreference.com/w/cpp/compiler_support
+        #if defined(_MSC_VER) && _MSC_VER < 1914
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+
+        // no filesystem support before iOS 13
+        #if defined(__IPHONE_OS_VERSION_MIN_REQUIRED) && __IPHONE_OS_VERSION_MIN_REQUIRED < 130000
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+
+        // no filesystem support before macOS Catalina
+        #if defined(__MAC_OS_X_VERSION_MIN_REQUIRED) && __MAC_OS_X_VERSION_MIN_REQUIRED < 101500
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+    #endif
+#endif
+
+#ifndef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+    #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 0
+#endif
+
+#ifndef JSON_HAS_FILESYSTEM
+    #define JSON_HAS_FILESYSTEM 0
+#endif
+
+#ifndef JSON_HAS_THREE_WAY_COMPARISON
+    #if defined(__cpp_impl_three_way_comparison) && __cpp_impl_three_way_comparison >= 201907L \
+        && defined(__cpp_lib_three_way_comparison) && __cpp_lib_three_way_comparison >= 201907L
+        #define JSON_HAS_THREE_WAY_COMPARISON 1
+    #else
+        #define JSON_HAS_THREE_WAY_COMPARISON 0
+    #endif
+#endif
+
+#ifndef JSON_HAS_RANGES
+    // ranges header shipping in GCC 11.1.0 (released 2021-04-27) has syntax error
+    #if defined(__GLIBCXX__) && __GLIBCXX__ == 20210427
+        #define JSON_HAS_RANGES 0
+    #elif defined(__cpp_lib_ranges)
+        #define JSON_HAS_RANGES 1
+    #else
+        #define JSON_HAS_RANGES 0
+    #endif
+#endif
+
+#ifndef JSON_HAS_STATIC_RTTI
+    #if !defined(_HAS_STATIC_RTTI) || _HAS_STATIC_RTTI != 0
+        #define JSON_HAS_STATIC_RTTI 1
+    #else
+        #define JSON_HAS_STATIC_RTTI 0
+    #endif
+#endif
+
+#ifdef JSON_HAS_CPP_17
+    #define JSON_INLINE_VARIABLE inline
+#else
+    #define JSON_INLINE_VARIABLE
+#endif
+
+#if JSON_HEDLEY_HAS_ATTRIBUTE(no_unique_address)
+    #define JSON_NO_UNIQUE_ADDRESS [[no_unique_address]]
+#else
+    #define JSON_NO_UNIQUE_ADDRESS
+#endif
+
+// disable documentation warnings on clang
+#if defined(__clang__)
+    #pragma clang diagnostic push
+    #pragma clang diagnostic ignored "-Wdocumentation"
+    #pragma clang diagnostic ignored "-Wdocumentation-unknown-command"
+#endif
+
+// allow disabling exceptions
+#if (defined(__cpp_exceptions) || defined(__EXCEPTIONS) || defined(_CPPUNWIND)) && !defined(JSON_NOEXCEPTION)
+    #define JSON_THROW(exception) throw exception
+    #define JSON_TRY try
+    #define JSON_CATCH(exception) catch(exception)
+    #define JSON_INTERNAL_CATCH(exception) catch(exception)
+#else
+    #include <cstdlib>
+    #define JSON_THROW(exception) std::abort()
+    #define JSON_TRY if(true)
+    #define JSON_CATCH(exception) if(false)
+    #define JSON_INTERNAL_CATCH(exception) if(false)
+#endif
+
+// override exception macros
+#if defined(JSON_THROW_USER)
+    #undef JSON_THROW
+    #define JSON_THROW JSON_THROW_USER
+#endif
+#if defined(JSON_TRY_USER)
+    #undef JSON_TRY
+    #define JSON_TRY JSON_TRY_USER
+#endif
+#if defined(JSON_CATCH_USER)
+    #undef JSON_CATCH
+    #define JSON_CATCH JSON_CATCH_USER
+    #undef JSON_INTERNAL_CATCH
+    #define JSON_INTERNAL_CATCH JSON_CATCH_USER
+#endif
+#if defined(JSON_INTERNAL_CATCH_USER)
+    #undef JSON_INTERNAL_CATCH
+    #define JSON_INTERNAL_CATCH JSON_INTERNAL_CATCH_USER
+#endif
+
+// allow overriding assert
+#if !defined(JSON_ASSERT)
+    #include <cassert> // assert
+    #define JSON_ASSERT(x) assert(x)
+#endif
+
+// allow to access some private functions (needed by the test suite)
+#if defined(JSON_TESTS_PRIVATE)
+    #define JSON_PRIVATE_UNLESS_TESTED public
+#else
+    #define JSON_PRIVATE_UNLESS_TESTED private
+#endif
+
+/*!
+@brief macro to briefly define a mapping between an enum and JSON
+@def NLOHMANN_JSON_SERIALIZE_ENUM
+@since version 3.4.0
+*/
+#define NLOHMANN_JSON_SERIALIZE_ENUM(ENUM_TYPE, ...)                                            \
+    template<typename BasicJsonType>                                                            \
+    inline void to_json(BasicJsonType& j, const ENUM_TYPE& e)                                   \
+    {                                                                                           \
+        static_assert(std::is_enum<ENUM_TYPE>::value, #ENUM_TYPE " must be an enum!");          \
+        static const std::pair<ENUM_TYPE, BasicJsonType> m[] = __VA_ARGS__;                     \
+        auto it = std::find_if(std::begin(m), std::end(m),                                      \
+                               [e](const std::pair<ENUM_TYPE, BasicJsonType>& ej_pair) -> bool  \
+        {                                                                                       \
+            return ej_pair.first == e;                                                          \
+        });                                                                                     \
+        j = ((it != std::end(m)) ? it : std::begin(m))->second;                                 \
+    }                                                                                           \
+    template<typename BasicJsonType>                                                            \
+    inline void from_json(const BasicJsonType& j, ENUM_TYPE& e)                                 \
+    {                                                                                           \
+        static_assert(std::is_enum<ENUM_TYPE>::value, #ENUM_TYPE " must be an enum!");          \
+        static const std::pair<ENUM_TYPE, BasicJsonType> m[] = __VA_ARGS__;                     \
+        auto it = std::find_if(std::begin(m), std::end(m),                                      \
+                               [&j](const std::pair<ENUM_TYPE, BasicJsonType>& ej_pair) -> bool \
+        {                                                                                       \
+            return ej_pair.second == j;                                                         \
+        });                                                                                     \
+        e = ((it != std::end(m)) ? it : std::begin(m))->first;                                  \
+    }
+
+// Ugly macros to avoid uglier copy-paste when specializing basic_json. They
+// may be removed in the future once the class is split.
+
+#define NLOHMANN_BASIC_JSON_TPL_DECLARATION                                \
+    template<template<typename, typename, typename...> class ObjectType,   \
+             template<typename, typename...> class ArrayType,              \
+             class StringType, class BooleanType, class NumberIntegerType, \
+             class NumberUnsignedType, class NumberFloatType,              \
+             template<typename> class AllocatorType,                       \
+             template<typename, typename = void> class JSONSerializer,     \
+             class BinaryType,                                             \
+             class CustomBaseClass>
+
+#define NLOHMANN_BASIC_JSON_TPL                                            \
+    basic_json<ObjectType, ArrayType, StringType, BooleanType,             \
+    NumberIntegerType, NumberUnsignedType, NumberFloatType,                \
+    AllocatorType, JSONSerializer, BinaryType, CustomBaseClass>
+
+// Macros to simplify conversion from/to types
+
+#define NLOHMANN_JSON_EXPAND( x ) x
+#define NLOHMANN_JSON_GET_MACRO(_1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52, _53, _54, _55, _56, _57, _58, _59, _60, _61, _62, _63, _64, NAME,...) NAME
+#define NLOHMANN_JSON_PASTE(...) NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_GET_MACRO(__VA_ARGS__, \
+        NLOHMANN_JSON_PASTE64, \
+        NLOHMANN_JSON_PASTE63, \
+        NLOHMANN_JSON_PASTE62, \
+        NLOHMANN_JSON_PASTE61, \
+        NLOHMANN_JSON_PASTE60, \
+        NLOHMANN_JSON_PASTE59, \
+        NLOHMANN_JSON_PASTE58, \
+        NLOHMANN_JSON_PASTE57, \
+        NLOHMANN_JSON_PASTE56, \
+        NLOHMANN_JSON_PASTE55, \
+        NLOHMANN_JSON_PASTE54, \
+        NLOHMANN_JSON_PASTE53, \
+        NLOHMANN_JSON_PASTE52, \
+        NLOHMANN_JSON_PASTE51, \
+        NLOHMANN_JSON_PASTE50, \
+        NLOHMANN_JSON_PASTE49, \
+        NLOHMANN_JSON_PASTE48, \
+        NLOHMANN_JSON_PASTE47, \
+        NLOHMANN_JSON_PASTE46, \
+        NLOHMANN_JSON_PASTE45, \
+        NLOHMANN_JSON_PASTE44, \
+        NLOHMANN_JSON_PASTE43, \
+        NLOHMANN_JSON_PASTE42, \
+        NLOHMANN_JSON_PASTE41, \
+        NLOHMANN_JSON_PASTE40, \
+        NLOHMANN_JSON_PASTE39, \
+        NLOHMANN_JSON_PASTE38, \
+        NLOHMANN_JSON_PASTE37, \
+        NLOHMANN_JSON_PASTE36, \
+        NLOHMANN_JSON_PASTE35, \
+        NLOHMANN_JSON_PASTE34, \
+        NLOHMANN_JSON_PASTE33, \
+        NLOHMANN_JSON_PASTE32, \
+        NLOHMANN_JSON_PASTE31, \
+        NLOHMANN_JSON_PASTE30, \
+        NLOHMANN_JSON_PASTE29, \
+        NLOHMANN_JSON_PASTE28, \
+        NLOHMANN_JSON_PASTE27, \
+        NLOHMANN_JSON_PASTE26, \
+        NLOHMANN_JSON_PASTE25, \
+        NLOHMANN_JSON_PASTE24, \
+        NLOHMANN_JSON_PASTE23, \
+        NLOHMANN_JSON_PASTE22, \
+        NLOHMANN_JSON_PASTE21, \
+        NLOHMANN_JSON_PASTE20, \
+        NLOHMANN_JSON_PASTE19, \
+        NLOHMANN_JSON_PASTE18, \
+        NLOHMANN_JSON_PASTE17, \
+        NLOHMANN_JSON_PASTE16, \
+        NLOHMANN_JSON_PASTE15, \
+        NLOHMANN_JSON_PASTE14, \
+        NLOHMANN_JSON_PASTE13, \
+        NLOHMANN_JSON_PASTE12, \
+        NLOHMANN_JSON_PASTE11, \
+        NLOHMANN_JSON_PASTE10, \
+        NLOHMANN_JSON_PASTE9, \
+        NLOHMANN_JSON_PASTE8, \
+        NLOHMANN_JSON_PASTE7, \
+        NLOHMANN_JSON_PASTE6, \
+        NLOHMANN_JSON_PASTE5, \
+        NLOHMANN_JSON_PASTE4, \
+        NLOHMANN_JSON_PASTE3, \
+        NLOHMANN_JSON_PASTE2, \
+        NLOHMANN_JSON_PASTE1)(__VA_ARGS__))
+#define NLOHMANN_JSON_PASTE2(func, v1) func(v1)
+#define NLOHMANN_JSON_PASTE3(func, v1, v2) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE2(func, v2)
+#define NLOHMANN_JSON_PASTE4(func, v1, v2, v3) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE3(func, v2, v3)
+#define NLOHMANN_JSON_PASTE5(func, v1, v2, v3, v4) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE4(func, v2, v3, v4)
+#define NLOHMANN_JSON_PASTE6(func, v1, v2, v3, v4, v5) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE5(func, v2, v3, v4, v5)
+#define NLOHMANN_JSON_PASTE7(func, v1, v2, v3, v4, v5, v6) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE6(func, v2, v3, v4, v5, v6)
+#define NLOHMANN_JSON_PASTE8(func, v1, v2, v3, v4, v5, v6, v7) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE7(func, v2, v3, v4, v5, v6, v7)
+#define NLOHMANN_JSON_PASTE9(func, v1, v2, v3, v4, v5, v6, v7, v8) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE8(func, v2, v3, v4, v5, v6, v7, v8)
+#define NLOHMANN_JSON_PASTE10(func, v1, v2, v3, v4, v5, v6, v7, v8, v9) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE9(func, v2, v3, v4, v5, v6, v7, v8, v9)
+#define NLOHMANN_JSON_PASTE11(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE10(func, v2, v3, v4, v5, v6, v7, v8, v9, v10)
+#define NLOHMANN_JSON_PASTE12(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE11(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11)
+#define NLOHMANN_JSON_PASTE13(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE12(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12)
+#define NLOHMANN_JSON_PASTE14(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE13(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13)
+#define NLOHMANN_JSON_PASTE15(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE14(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14)
+#define NLOHMANN_JSON_PASTE16(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE15(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15)
+#define NLOHMANN_JSON_PASTE17(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE16(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16)
+#define NLOHMANN_JSON_PASTE18(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE17(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17)
+#define NLOHMANN_JSON_PASTE19(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE18(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18)
+#define NLOHMANN_JSON_PASTE20(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE19(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19)
+#define NLOHMANN_JSON_PASTE21(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE20(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20)
+#define NLOHMANN_JSON_PASTE22(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE21(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21)
+#define NLOHMANN_JSON_PASTE23(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE22(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22)
+#define NLOHMANN_JSON_PASTE24(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE23(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23)
+#define NLOHMANN_JSON_PASTE25(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE24(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24)
+#define NLOHMANN_JSON_PASTE26(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE25(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25)
+#define NLOHMANN_JSON_PASTE27(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE26(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26)
+#define NLOHMANN_JSON_PASTE28(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE27(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27)
+#define NLOHMANN_JSON_PASTE29(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE28(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28)
+#define NLOHMANN_JSON_PASTE30(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE29(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29)
+#define NLOHMANN_JSON_PASTE31(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE30(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30)
+#define NLOHMANN_JSON_PASTE32(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE31(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31)
+#define NLOHMANN_JSON_PASTE33(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE32(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32)
+#define NLOHMANN_JSON_PASTE34(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE33(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33)
+#define NLOHMANN_JSON_PASTE35(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE34(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34)
+#define NLOHMANN_JSON_PASTE36(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE35(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35)
+#define NLOHMANN_JSON_PASTE37(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE36(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36)
+#define NLOHMANN_JSON_PASTE38(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE37(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37)
+#define NLOHMANN_JSON_PASTE39(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE38(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38)
+#define NLOHMANN_JSON_PASTE40(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE39(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39)
+#define NLOHMANN_JSON_PASTE41(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE40(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40)
+#define NLOHMANN_JSON_PASTE42(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE41(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41)
+#define NLOHMANN_JSON_PASTE43(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE42(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42)
+#define NLOHMANN_JSON_PASTE44(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE43(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43)
+#define NLOHMANN_JSON_PASTE45(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE44(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44)
+#define NLOHMANN_JSON_PASTE46(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE45(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45)
+#define NLOHMANN_JSON_PASTE47(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE46(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46)
+#define NLOHMANN_JSON_PASTE48(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE47(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47)
+#define NLOHMANN_JSON_PASTE49(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE48(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48)
+#define NLOHMANN_JSON_PASTE50(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE49(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49)
+#define NLOHMANN_JSON_PASTE51(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE50(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50)
+#define NLOHMANN_JSON_PASTE52(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE51(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51)
+#define NLOHMANN_JSON_PASTE53(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE52(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52)
+#define NLOHMANN_JSON_PASTE54(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE53(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53)
+#define NLOHMANN_JSON_PASTE55(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE54(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54)
+#define NLOHMANN_JSON_PASTE56(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE55(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55)
+#define NLOHMANN_JSON_PASTE57(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE56(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56)
+#define NLOHMANN_JSON_PASTE58(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE57(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57)
+#define NLOHMANN_JSON_PASTE59(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE58(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58)
+#define NLOHMANN_JSON_PASTE60(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE59(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59)
+#define NLOHMANN_JSON_PASTE61(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE60(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60)
+#define NLOHMANN_JSON_PASTE62(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE61(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61)
+#define NLOHMANN_JSON_PASTE63(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE62(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62)
+#define NLOHMANN_JSON_PASTE64(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62, v63) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE63(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62, v63)
+
+#define NLOHMANN_JSON_TO(v1) nlohmann_json_j[#v1] = nlohmann_json_t.v1;
+#define NLOHMANN_JSON_FROM(v1) nlohmann_json_j.at(#v1).get_to(nlohmann_json_t.v1);
+#define NLOHMANN_JSON_FROM_WITH_DEFAULT(v1) nlohmann_json_t.v1 = nlohmann_json_j.value(#v1, nlohmann_json_default_obj.v1);
+
+/*!
+@brief macro
+@def NLOHMANN_DEFINE_TYPE_INTRUSIVE
+@since version 3.9.0
+*/
+#define NLOHMANN_DEFINE_TYPE_INTRUSIVE(Type, ...)  \
+    friend void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \
+    friend void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM, __VA_ARGS__)) }
+
+#define NLOHMANN_DEFINE_TYPE_INTRUSIVE_WITH_DEFAULT(Type, ...)  \
+    friend void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \
+    friend void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { const Type nlohmann_json_default_obj{}; NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM_WITH_DEFAULT, __VA_ARGS__)) }
+
+#define NLOHMANN_DEFINE_TYPE_INTRUSIVE_ONLY_SERIALIZE(Type, ...)  \
+    friend void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) }
+
+/*!
+@brief macro
+@def NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE
+@since version 3.9.0
+*/
+#define NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE(Type, ...)  \
+    inline void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \
+    inline void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM, __VA_ARGS__)) }
+
+#define NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE_ONLY_SERIALIZE(Type, ...)  \
+    inline void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) }
+
+#define NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE_WITH_DEFAULT(Type, ...)  \
+    inline void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \
+    inline void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { const Type nlohmann_json_default_obj{}; NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM_WITH_DEFAULT, __VA_ARGS__)) }
+
+// inspired from https://stackoverflow.com/a/26745591
+// allows to call any std function as if (e.g. with begin):
+// using std::begin; begin(x);
+//
+// it allows using the detected idiom to retrieve the return type
+// of such an expression
+#define NLOHMANN_CAN_CALL_STD_FUNC_IMPL(std_name)                                 \
+    namespace detail {                                                            \
+    using std::std_name;                                                          \
+    \
+    template<typename... T>                                                       \
+    using result_of_##std_name = decltype(std_name(std::declval<T>()...));        \
+    }                                                                             \
+    \
+    namespace detail2 {                                                           \
+    struct std_name##_tag                                                         \
+    {                                                                             \
+    };                                                                            \
+    \
+    template<typename... T>                                                       \
+    std_name##_tag std_name(T&&...);                                              \
+    \
+    template<typename... T>                                                       \
+    using result_of_##std_name = decltype(std_name(std::declval<T>()...));        \
+    \
+    template<typename... T>                                                       \
+    struct would_call_std_##std_name                                              \
+    {                                                                             \
+        static constexpr auto const value = ::nlohmann::detail::                  \
+                                            is_detected_exact<std_name##_tag, result_of_##std_name, T...>::value; \
+    };                                                                            \
+    } /* namespace detail2 */ \
+    \
+    template<typename... T>                                                       \
+    struct would_call_std_##std_name : detail2::would_call_std_##std_name<T...>   \
+    {                                                                             \
+    }
+
+#ifndef JSON_USE_IMPLICIT_CONVERSIONS
+    #define JSON_USE_IMPLICIT_CONVERSIONS 1
+#endif
+
+#if JSON_USE_IMPLICIT_CONVERSIONS
+    #define JSON_EXPLICIT
+#else
+    #define JSON_EXPLICIT explicit
+#endif
+
+#ifndef JSON_DISABLE_ENUM_SERIALIZATION
+    #define JSON_DISABLE_ENUM_SERIALIZATION 0
+#endif
+
+#ifndef JSON_USE_GLOBAL_UDLS
+    #define JSON_USE_GLOBAL_UDLS 1
+#endif
+
+#if JSON_HAS_THREE_WAY_COMPARISON
+    #include <compare> // partial_ordering
+#endif
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+///////////////////////////
+// JSON type enumeration //
+///////////////////////////
+
+/*!
+@brief the JSON type enumeration
+
+This enumeration collects the different JSON types. It is internally used to
+distinguish the stored values, and the functions @ref basic_json::is_null(),
+@ref basic_json::is_object(), @ref basic_json::is_array(),
+@ref basic_json::is_string(), @ref basic_json::is_boolean(),
+@ref basic_json::is_number() (with @ref basic_json::is_number_integer(),
+@ref basic_json::is_number_unsigned(), and @ref basic_json::is_number_float()),
+@ref basic_json::is_discarded(), @ref basic_json::is_primitive(), and
+@ref basic_json::is_structured() rely on it.
+
+@note There are three enumeration entries (number_integer, number_unsigned, and
+number_float), because the library distinguishes these three types for numbers:
+@ref basic_json::number_unsigned_t is used for unsigned integers,
+@ref basic_json::number_integer_t is used for signed integers, and
+@ref basic_json::number_float_t is used for floating-point numbers or to
+approximate integers which do not fit in the limits of their respective type.
+
+@sa see @ref basic_json::basic_json(const value_t value_type) -- create a JSON
+value with the default value for a given type
+
+@since version 1.0.0
+*/
+enum class value_t : std::uint8_t
+{
+    null,             ///< null value
+    object,           ///< object (unordered set of name/value pairs)
+    array,            ///< array (ordered collection of values)
+    string,           ///< string value
+    boolean,          ///< boolean value
+    number_integer,   ///< number value (signed integer)
+    number_unsigned,  ///< number value (unsigned integer)
+    number_float,     ///< number value (floating-point)
+    binary,           ///< binary array (ordered collection of bytes)
+    discarded         ///< discarded by the parser callback function
+};
+
+/*!
+@brief comparison operator for JSON types
+
+Returns an ordering that is similar to Python:
+- order: null < boolean < number < object < array < string < binary
+- furthermore, each type is not smaller than itself
+- discarded values are not comparable
+- binary is represented as a b"" string in python and directly comparable to a
+  string; however, making a binary array directly comparable with a string would
+  be surprising behavior in a JSON file.
+
+@since version 1.0.0
+*/
+#if JSON_HAS_THREE_WAY_COMPARISON
+    inline std::partial_ordering operator<=>(const value_t lhs, const value_t rhs) noexcept // *NOPAD*
+#else
+    inline bool operator<(const value_t lhs, const value_t rhs) noexcept
+#endif
+{
+    static constexpr std::array<std::uint8_t, 9> order = {{
+            0 /* null */, 3 /* object */, 4 /* array */, 5 /* string */,
+            1 /* boolean */, 2 /* integer */, 2 /* unsigned */, 2 /* float */,
+            6 /* binary */
+        }
+    };
+
+    const auto l_index = static_cast<std::size_t>(lhs);
+    const auto r_index = static_cast<std::size_t>(rhs);
+#if JSON_HAS_THREE_WAY_COMPARISON
+    if (l_index < order.size() && r_index < order.size())
+    {
+        return order[l_index] <=> order[r_index]; // *NOPAD*
+    }
+    return std::partial_ordering::unordered;
+#else
+    return l_index < order.size() && r_index < order.size() && order[l_index] < order[r_index];
+#endif
+}
+
+// GCC selects the built-in operator< over an operator rewritten from
+// a user-defined spaceship operator
+// Clang, MSVC, and ICC select the rewritten candidate
+// (see GCC bug https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105200)
+#if JSON_HAS_THREE_WAY_COMPARISON && defined(__GNUC__)
+inline bool operator<(const value_t lhs, const value_t rhs) noexcept
+{
+    return std::is_lt(lhs <=> rhs); // *NOPAD*
+}
+#endif
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/string_escape.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/*!
+@brief replace all occurrences of a substring by another string
+
+@param[in,out] s  the string to manipulate; changed so that all
+               occurrences of @a f are replaced with @a t
+@param[in]     f  the substring to replace with @a t
+@param[in]     t  the string to replace @a f
+
+@pre The search string @a f must not be empty. **This precondition is
+enforced with an assertion.**
+
+@since version 2.0.0
+*/
+template<typename StringType>
+inline void replace_substring(StringType& s, const StringType& f,
+                              const StringType& t)
+{
+    JSON_ASSERT(!f.empty());
+    for (auto pos = s.find(f);                // find first occurrence of f
+            pos != StringType::npos;          // make sure f was found
+            s.replace(pos, f.size(), t),      // replace with t, and
+            pos = s.find(f, pos + t.size()))  // find next occurrence of f
+    {}
+}
+
+/*!
+ * @brief string escaping as described in RFC 6901 (Sect. 4)
+ * @param[in] s string to escape
+ * @return    escaped string
+ *
+ * Note the order of escaping "~" to "~0" and "/" to "~1" is important.
+ */
+template<typename StringType>
+inline StringType escape(StringType s)
+{
+    replace_substring(s, StringType{"~"}, StringType{"~0"});
+    replace_substring(s, StringType{"/"}, StringType{"~1"});
+    return s;
+}
+
+/*!
+ * @brief string unescaping as described in RFC 6901 (Sect. 4)
+ * @param[in] s string to unescape
+ * @return    unescaped string
+ *
+ * Note the order of escaping "~1" to "/" and "~0" to "~" is important.
+ */
+template<typename StringType>
+static void unescape(StringType& s)
+{
+    replace_substring(s, StringType{"~1"}, StringType{"/"});
+    replace_substring(s, StringType{"~0"}, StringType{"~"});
+}
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/input/position_t.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstddef> // size_t
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/// struct to capture the start position of the current token
+struct position_t
+{
+    /// the total number of characters read
+    std::size_t chars_read_total = 0;
+    /// the number of characters read in the current line
+    std::size_t chars_read_current_line = 0;
+    /// the number of lines read
+    std::size_t lines_read = 0;
+
+    /// conversion to size_t to preserve SAX interface
+    constexpr operator size_t() const
+    {
+        return chars_read_total;
+    }
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-FileCopyrightText: 2018 The Abseil Authors
+// SPDX-License-Identifier: MIT
+
+
+
+#include <array> // array
+#include <cstddef> // size_t
+#include <type_traits> // conditional, enable_if, false_type, integral_constant, is_constructible, is_integral, is_same, remove_cv, remove_reference, true_type
+#include <utility> // index_sequence, make_index_sequence, index_sequence_for
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+template<typename T>
+using uncvref_t = typename std::remove_cv<typename std::remove_reference<T>::type>::type;
+
+#ifdef JSON_HAS_CPP_14
+
+// the following utilities are natively available in C++14
+using std::enable_if_t;
+using std::index_sequence;
+using std::make_index_sequence;
+using std::index_sequence_for;
+
+#else
+
+// alias templates to reduce boilerplate
+template<bool B, typename T = void>
+using enable_if_t = typename std::enable_if<B, T>::type;
+
+// The following code is taken from https://github.com/abseil/abseil-cpp/blob/10cb35e459f5ecca5b2ff107635da0bfa41011b4/absl/utility/utility.h
+// which is part of Google Abseil (https://github.com/abseil/abseil-cpp), licensed under the Apache License 2.0.
+
+//// START OF CODE FROM GOOGLE ABSEIL
+
+// integer_sequence
+//
+// Class template representing a compile-time integer sequence. An instantiation
+// of `integer_sequence<T, Ints...>` has a sequence of integers encoded in its
+// type through its template arguments (which is a common need when
+// working with C++11 variadic templates). `absl::integer_sequence` is designed
+// to be a drop-in replacement for C++14's `std::integer_sequence`.
+//
+// Example:
+//
+//   template< class T, T... Ints >
+//   void user_function(integer_sequence<T, Ints...>);
+//
+//   int main()
+//   {
+//     // user_function's `T` will be deduced to `int` and `Ints...`
+//     // will be deduced to `0, 1, 2, 3, 4`.
+//     user_function(make_integer_sequence<int, 5>());
+//   }
+template <typename T, T... Ints>
+struct integer_sequence
+{
+    using value_type = T;
+    static constexpr std::size_t size() noexcept
+    {
+        return sizeof...(Ints);
+    }
+};
+
+// index_sequence
+//
+// A helper template for an `integer_sequence` of `size_t`,
+// `absl::index_sequence` is designed to be a drop-in replacement for C++14's
+// `std::index_sequence`.
+template <size_t... Ints>
+using index_sequence = integer_sequence<size_t, Ints...>;
+
+namespace utility_internal
+{
+
+template <typename Seq, size_t SeqSize, size_t Rem>
+struct Extend;
+
+// Note that SeqSize == sizeof...(Ints). It's passed explicitly for efficiency.
+template <typename T, T... Ints, size_t SeqSize>
+struct Extend<integer_sequence<T, Ints...>, SeqSize, 0>
+{
+    using type = integer_sequence < T, Ints..., (Ints + SeqSize)... >;
+};
+
+template <typename T, T... Ints, size_t SeqSize>
+struct Extend<integer_sequence<T, Ints...>, SeqSize, 1>
+{
+    using type = integer_sequence < T, Ints..., (Ints + SeqSize)..., 2 * SeqSize >;
+};
+
+// Recursion helper for 'make_integer_sequence<T, N>'.
+// 'Gen<T, N>::type' is an alias for 'integer_sequence<T, 0, 1, ... N-1>'.
+template <typename T, size_t N>
+struct Gen
+{
+    using type =
+        typename Extend < typename Gen < T, N / 2 >::type, N / 2, N % 2 >::type;
+};
+
+template <typename T>
+struct Gen<T, 0>
+{
+    using type = integer_sequence<T>;
+};
+
+}  // namespace utility_internal
+
+// Compile-time sequences of integers
+
+// make_integer_sequence
+//
+// This template alias is equivalent to
+// `integer_sequence<int, 0, 1, ..., N-1>`, and is designed to be a drop-in
+// replacement for C++14's `std::make_integer_sequence`.
+template <typename T, T N>
+using make_integer_sequence = typename utility_internal::Gen<T, N>::type;
+
+// make_index_sequence
+//
+// This template alias is equivalent to `index_sequence<0, 1, ..., N-1>`,
+// and is designed to be a drop-in replacement for C++14's
+// `std::make_index_sequence`.
+template <size_t N>
+using make_index_sequence = make_integer_sequence<size_t, N>;
+
+// index_sequence_for
+//
+// Converts a typename pack into an index sequence of the same length, and
+// is designed to be a drop-in replacement for C++14's
+// `std::index_sequence_for()`
+template <typename... Ts>
+using index_sequence_for = make_index_sequence<sizeof...(Ts)>;
+
+//// END OF CODE FROM GOOGLE ABSEIL
+
+#endif
+
+// dispatch utility (taken from ranges-v3)
+template<unsigned N> struct priority_tag : priority_tag < N - 1 > {};
+template<> struct priority_tag<0> {};
+
+// taken from ranges-v3
+template<typename T>
+struct static_const
+{
+    static JSON_INLINE_VARIABLE constexpr T value{};
+};
+
+#ifndef JSON_HAS_CPP_17
+    template<typename T>
+    constexpr T static_const<T>::value;
+#endif
+
+template<typename T, typename... Args>
+inline constexpr std::array<T, sizeof...(Args)> make_array(Args&& ... args)
+{
+    return std::array<T, sizeof...(Args)> {{static_cast<T>(std::forward<Args>(args))...}};
+}
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <limits> // numeric_limits
+#include <type_traits> // false_type, is_constructible, is_integral, is_same, true_type
+#include <utility> // declval
+#include <tuple> // tuple
+#include <string> // char_traits
+
+// #include <nlohmann/detail/iterators/iterator_traits.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <iterator> // random_access_iterator_tag
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+// #include <nlohmann/detail/meta/void_t.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+template<typename It, typename = void>
+struct iterator_types {};
+
+template<typename It>
+struct iterator_types <
+    It,
+    void_t<typename It::difference_type, typename It::value_type, typename It::pointer,
+    typename It::reference, typename It::iterator_category >>
+{
+    using difference_type = typename It::difference_type;
+    using value_type = typename It::value_type;
+    using pointer = typename It::pointer;
+    using reference = typename It::reference;
+    using iterator_category = typename It::iterator_category;
+};
+
+// This is required as some compilers implement std::iterator_traits in a way that
+// doesn't work with SFINAE. See https://github.com/nlohmann/json/issues/1341.
+template<typename T, typename = void>
+struct iterator_traits
+{
+};
+
+template<typename T>
+struct iterator_traits < T, enable_if_t < !std::is_pointer<T>::value >>
+            : iterator_types<T>
+{
+};
+
+template<typename T>
+struct iterator_traits<T*, enable_if_t<std::is_object<T>::value>>
+{
+    using iterator_category = std::random_access_iterator_tag;
+    using value_type = T;
+    using difference_type = ptrdiff_t;
+    using pointer = T*;
+    using reference = T&;
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/call_std/begin.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+NLOHMANN_CAN_CALL_STD_FUNC_IMPL(begin);
+
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/meta/call_std/end.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+NLOHMANN_CAN_CALL_STD_FUNC_IMPL(end);
+
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/detected.hpp>
+
+// #include <nlohmann/json_fwd.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+#ifndef INCLUDE_NLOHMANN_JSON_FWD_HPP_
+    #define INCLUDE_NLOHMANN_JSON_FWD_HPP_
+
+    #include <cstdint> // int64_t, uint64_t
+    #include <map> // map
+    #include <memory> // allocator
+    #include <string> // string
+    #include <vector> // vector
+
+    // #include <nlohmann/detail/abi_macros.hpp>
+
+
+    /*!
+    @brief namespace for Niels Lohmann
+    @see https://github.com/nlohmann
+    @since version 1.0.0
+    */
+    NLOHMANN_JSON_NAMESPACE_BEGIN
+
+    /*!
+    @brief default JSONSerializer template argument
+
+    This serializer ignores the template arguments and uses ADL
+    ([argument-dependent lookup](https://en.cppreference.com/w/cpp/language/adl))
+    for serialization.
+    */
+    template<typename T = void, typename SFINAE = void>
+    struct adl_serializer;
+
+    /// a class to store JSON values
+    /// @sa https://json.nlohmann.me/api/basic_json/
+    template<template<typename U, typename V, typename... Args> class ObjectType =
+    std::map,
+    template<typename U, typename... Args> class ArrayType = std::vector,
+    class StringType = std::string, class BooleanType = bool,
+    class NumberIntegerType = std::int64_t,
+    class NumberUnsignedType = std::uint64_t,
+    class NumberFloatType = double,
+    template<typename U> class AllocatorType = std::allocator,
+    template<typename T, typename SFINAE = void> class JSONSerializer =
+    adl_serializer,
+    class BinaryType = std::vector<std::uint8_t>, // cppcheck-suppress syntaxError
+    class CustomBaseClass = void>
+    class basic_json;
+
+    /// @brief JSON Pointer defines a string syntax for identifying a specific value within a JSON document
+    /// @sa https://json.nlohmann.me/api/json_pointer/
+    template<typename RefStringType>
+    class json_pointer;
+
+    /*!
+    @brief default specialization
+    @sa https://json.nlohmann.me/api/json/
+    */
+    using json = basic_json<>;
+
+    /// @brief a minimal map-like container that preserves insertion order
+    /// @sa https://json.nlohmann.me/api/ordered_map/
+    template<class Key, class T, class IgnoredLess, class Allocator>
+    struct ordered_map;
+
+    /// @brief specialization that maintains the insertion order of object keys
+    /// @sa https://json.nlohmann.me/api/ordered_json/
+    using ordered_json = basic_json<nlohmann::ordered_map>;
+
+    NLOHMANN_JSON_NAMESPACE_END
+
+#endif  // INCLUDE_NLOHMANN_JSON_FWD_HPP_
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+/*!
+@brief detail namespace with internal helper functions
+
+This namespace collects functions that should not be exposed,
+implementations of some @ref basic_json methods, and meta-programming helpers.
+
+@since version 2.1.0
+*/
+namespace detail
+{
+
+/////////////
+// helpers //
+/////////////
+
+// Note to maintainers:
+//
+// Every trait in this file expects a non CV-qualified type.
+// The only exceptions are in the 'aliases for detected' section
+// (i.e. those of the form: decltype(T::member_function(std::declval<T>())))
+//
+// In this case, T has to be properly CV-qualified to constraint the function arguments
+// (e.g. to_json(BasicJsonType&, const T&))
+
+template<typename> struct is_basic_json : std::false_type {};
+
+NLOHMANN_BASIC_JSON_TPL_DECLARATION
+struct is_basic_json<NLOHMANN_BASIC_JSON_TPL> : std::true_type {};
+
+// used by exceptions create() member functions
+// true_type for pointer to possibly cv-qualified basic_json or std::nullptr_t
+// false_type otherwise
+template<typename BasicJsonContext>
+struct is_basic_json_context :
+    std::integral_constant < bool,
+    is_basic_json<typename std::remove_cv<typename std::remove_pointer<BasicJsonContext>::type>::type>::value
+    || std::is_same<BasicJsonContext, std::nullptr_t>::value >
+{};
+
+//////////////////////
+// json_ref helpers //
+//////////////////////
+
+template<typename>
+class json_ref;
+
+template<typename>
+struct is_json_ref : std::false_type {};
+
+template<typename T>
+struct is_json_ref<json_ref<T>> : std::true_type {};
+
+//////////////////////////
+// aliases for detected //
+//////////////////////////
+
+template<typename T>
+using mapped_type_t = typename T::mapped_type;
+
+template<typename T>
+using key_type_t = typename T::key_type;
+
+template<typename T>
+using value_type_t = typename T::value_type;
+
+template<typename T>
+using difference_type_t = typename T::difference_type;
+
+template<typename T>
+using pointer_t = typename T::pointer;
+
+template<typename T>
+using reference_t = typename T::reference;
+
+template<typename T>
+using iterator_category_t = typename T::iterator_category;
+
+template<typename T, typename... Args>
+using to_json_function = decltype(T::to_json(std::declval<Args>()...));
+
+template<typename T, typename... Args>
+using from_json_function = decltype(T::from_json(std::declval<Args>()...));
+
+template<typename T, typename U>
+using get_template_function = decltype(std::declval<T>().template get<U>());
+
+// trait checking if JSONSerializer<T>::from_json(json const&, udt&) exists
+template<typename BasicJsonType, typename T, typename = void>
+struct has_from_json : std::false_type {};
+
+// trait checking if j.get<T> is valid
+// use this trait instead of std::is_constructible or std::is_convertible,
+// both rely on, or make use of implicit conversions, and thus fail when T
+// has several constructors/operator= (see https://github.com/nlohmann/json/issues/958)
+template <typename BasicJsonType, typename T>
+struct is_getable
+{
+    static constexpr bool value = is_detected<get_template_function, const BasicJsonType&, T>::value;
+};
+
+template<typename BasicJsonType, typename T>
+struct has_from_json < BasicJsonType, T, enable_if_t < !is_basic_json<T>::value >>
+{
+    using serializer = typename BasicJsonType::template json_serializer<T, void>;
+
+    static constexpr bool value =
+        is_detected_exact<void, from_json_function, serializer,
+        const BasicJsonType&, T&>::value;
+};
+
+// This trait checks if JSONSerializer<T>::from_json(json const&) exists
+// this overload is used for non-default-constructible user-defined-types
+template<typename BasicJsonType, typename T, typename = void>
+struct has_non_default_from_json : std::false_type {};
+
+template<typename BasicJsonType, typename T>
+struct has_non_default_from_json < BasicJsonType, T, enable_if_t < !is_basic_json<T>::value >>
+{
+    using serializer = typename BasicJsonType::template json_serializer<T, void>;
+
+    static constexpr bool value =
+        is_detected_exact<T, from_json_function, serializer,
+        const BasicJsonType&>::value;
+};
+
+// This trait checks if BasicJsonType::json_serializer<T>::to_json exists
+// Do not evaluate the trait when T is a basic_json type, to avoid template instantiation infinite recursion.
+template<typename BasicJsonType, typename T, typename = void>
+struct has_to_json : std::false_type {};
+
+template<typename BasicJsonType, typename T>
+struct has_to_json < BasicJsonType, T, enable_if_t < !is_basic_json<T>::value >>
+{
+    using serializer = typename BasicJsonType::template json_serializer<T, void>;
+
+    static constexpr bool value =
+        is_detected_exact<void, to_json_function, serializer, BasicJsonType&,
+        T>::value;
+};
+
+template<typename T>
+using detect_key_compare = typename T::key_compare;
+
+template<typename T>
+struct has_key_compare : std::integral_constant<bool, is_detected<detect_key_compare, T>::value> {};
+
+// obtains the actual object key comparator
+template<typename BasicJsonType>
+struct actual_object_comparator
+{
+    using object_t = typename BasicJsonType::object_t;
+    using object_comparator_t = typename BasicJsonType::default_object_comparator_t;
+    using type = typename std::conditional < has_key_compare<object_t>::value,
+          typename object_t::key_compare, object_comparator_t>::type;
+};
+
+template<typename BasicJsonType>
+using actual_object_comparator_t = typename actual_object_comparator<BasicJsonType>::type;
+
+/////////////////
+// char_traits //
+/////////////////
+
+// Primary template of char_traits calls std char_traits
+template<typename T>
+struct char_traits : std::char_traits<T>
+{};
+
+// Explicitly define char traits for unsigned char since it is not standard
+template<>
+struct char_traits<unsigned char> : std::char_traits<char>
+{
+    using char_type = unsigned char;
+    using int_type = uint64_t;
+
+    // Redefine to_int_type function
+    static int_type to_int_type(char_type c) noexcept
+    {
+        return static_cast<int_type>(c);
+    }
+
+    static char_type to_char_type(int_type i) noexcept
+    {
+        return static_cast<char_type>(i);
+    }
+
+    static constexpr int_type eof() noexcept
+    {
+        return static_cast<int_type>(EOF);
+    }
+};
+
+// Explicitly define char traits for signed char since it is not standard
+template<>
+struct char_traits<signed char> : std::char_traits<char>
+{
+    using char_type = signed char;
+    using int_type = uint64_t;
+
+    // Redefine to_int_type function
+    static int_type to_int_type(char_type c) noexcept
+    {
+        return static_cast<int_type>(c);
+    }
+
+    static char_type to_char_type(int_type i) noexcept
+    {
+        return static_cast<char_type>(i);
+    }
+
+    static constexpr int_type eof() noexcept
+    {
+        return static_cast<int_type>(EOF);
+    }
+};
+
+///////////////////
+// is_ functions //
+///////////////////
+
+// https://en.cppreference.com/w/cpp/types/conjunction
+template<class...> struct conjunction : std::true_type { };
+template<class B> struct conjunction<B> : B { };
+template<class B, class... Bn>
+struct conjunction<B, Bn...>
+: std::conditional<static_cast<bool>(B::value), conjunction<Bn...>, B>::type {};
+
+// https://en.cppreference.com/w/cpp/types/negation
+template<class B> struct negation : std::integral_constant < bool, !B::value > { };
+
+// Reimplementation of is_constructible and is_default_constructible, due to them being broken for
+// std::pair and std::tuple until LWG 2367 fix (see https://cplusplus.github.io/LWG/lwg-defects.html#2367).
+// This causes compile errors in e.g. clang 3.5 or gcc 4.9.
+template <typename T>
+struct is_default_constructible : std::is_default_constructible<T> {};
+
+template <typename T1, typename T2>
+struct is_default_constructible<std::pair<T1, T2>>
+            : conjunction<is_default_constructible<T1>, is_default_constructible<T2>> {};
+
+template <typename T1, typename T2>
+struct is_default_constructible<const std::pair<T1, T2>>
+            : conjunction<is_default_constructible<T1>, is_default_constructible<T2>> {};
+
+template <typename... Ts>
+struct is_default_constructible<std::tuple<Ts...>>
+            : conjunction<is_default_constructible<Ts>...> {};
+
+template <typename... Ts>
+struct is_default_constructible<const std::tuple<Ts...>>
+            : conjunction<is_default_constructible<Ts>...> {};
+
+template <typename T, typename... Args>
+struct is_constructible : std::is_constructible<T, Args...> {};
+
+template <typename T1, typename T2>
+struct is_constructible<std::pair<T1, T2>> : is_default_constructible<std::pair<T1, T2>> {};
+
+template <typename T1, typename T2>
+struct is_constructible<const std::pair<T1, T2>> : is_default_constructible<const std::pair<T1, T2>> {};
+
+template <typename... Ts>
+struct is_constructible<std::tuple<Ts...>> : is_default_constructible<std::tuple<Ts...>> {};
+
+template <typename... Ts>
+struct is_constructible<const std::tuple<Ts...>> : is_default_constructible<const std::tuple<Ts...>> {};
+
+template<typename T, typename = void>
+struct is_iterator_traits : std::false_type {};
+
+template<typename T>
+struct is_iterator_traits<iterator_traits<T>>
+{
+  private:
+    using traits = iterator_traits<T>;
+
+  public:
+    static constexpr auto value =
+        is_detected<value_type_t, traits>::value &&
+        is_detected<difference_type_t, traits>::value &&
+        is_detected<pointer_t, traits>::value &&
+        is_detected<iterator_category_t, traits>::value &&
+        is_detected<reference_t, traits>::value;
+};
+
+template<typename T>
+struct is_range
+{
+  private:
+    using t_ref = typename std::add_lvalue_reference<T>::type;
+
+    using iterator = detected_t<result_of_begin, t_ref>;
+    using sentinel = detected_t<result_of_end, t_ref>;
+
+    // to be 100% correct, it should use https://en.cppreference.com/w/cpp/iterator/input_or_output_iterator
+    // and https://en.cppreference.com/w/cpp/iterator/sentinel_for
+    // but reimplementing these would be too much work, as a lot of other concepts are used underneath
+    static constexpr auto is_iterator_begin =
+        is_iterator_traits<iterator_traits<iterator>>::value;
+
+  public:
+    static constexpr bool value = !std::is_same<iterator, nonesuch>::value && !std::is_same<sentinel, nonesuch>::value && is_iterator_begin;
+};
+
+template<typename R>
+using iterator_t = enable_if_t<is_range<R>::value, result_of_begin<decltype(std::declval<R&>())>>;
+
+template<typename T>
+using range_value_t = value_type_t<iterator_traits<iterator_t<T>>>;
+
+// The following implementation of is_complete_type is taken from
+// https://blogs.msdn.microsoft.com/vcblog/2015/12/02/partial-support-for-expression-sfinae-in-vs-2015-update-1/
+// and is written by Xiang Fan who agreed to using it in this library.
+
+template<typename T, typename = void>
+struct is_complete_type : std::false_type {};
+
+template<typename T>
+struct is_complete_type<T, decltype(void(sizeof(T)))> : std::true_type {};
+
+template<typename BasicJsonType, typename CompatibleObjectType,
+         typename = void>
+struct is_compatible_object_type_impl : std::false_type {};
+
+template<typename BasicJsonType, typename CompatibleObjectType>
+struct is_compatible_object_type_impl <
+    BasicJsonType, CompatibleObjectType,
+    enable_if_t < is_detected<mapped_type_t, CompatibleObjectType>::value&&
+    is_detected<key_type_t, CompatibleObjectType>::value >>
+{
+    using object_t = typename BasicJsonType::object_t;
+
+    // macOS's is_constructible does not play well with nonesuch...
+    static constexpr bool value =
+        is_constructible<typename object_t::key_type,
+        typename CompatibleObjectType::key_type>::value &&
+        is_constructible<typename object_t::mapped_type,
+        typename CompatibleObjectType::mapped_type>::value;
+};
+
+template<typename BasicJsonType, typename CompatibleObjectType>
+struct is_compatible_object_type
+    : is_compatible_object_type_impl<BasicJsonType, CompatibleObjectType> {};
+
+template<typename BasicJsonType, typename ConstructibleObjectType,
+         typename = void>
+struct is_constructible_object_type_impl : std::false_type {};
+
+template<typename BasicJsonType, typename ConstructibleObjectType>
+struct is_constructible_object_type_impl <
+    BasicJsonType, ConstructibleObjectType,
+    enable_if_t < is_detected<mapped_type_t, ConstructibleObjectType>::value&&
+    is_detected<key_type_t, ConstructibleObjectType>::value >>
+{
+    using object_t = typename BasicJsonType::object_t;
+
+    static constexpr bool value =
+        (is_default_constructible<ConstructibleObjectType>::value &&
+         (std::is_move_assignable<ConstructibleObjectType>::value ||
+          std::is_copy_assignable<ConstructibleObjectType>::value) &&
+         (is_constructible<typename ConstructibleObjectType::key_type,
+          typename object_t::key_type>::value &&
+          std::is_same <
+          typename object_t::mapped_type,
+          typename ConstructibleObjectType::mapped_type >::value)) ||
+        (has_from_json<BasicJsonType,
+         typename ConstructibleObjectType::mapped_type>::value ||
+         has_non_default_from_json <
+         BasicJsonType,
+         typename ConstructibleObjectType::mapped_type >::value);
+};
+
+template<typename BasicJsonType, typename ConstructibleObjectType>
+struct is_constructible_object_type
+    : is_constructible_object_type_impl<BasicJsonType,
+      ConstructibleObjectType> {};
+
+template<typename BasicJsonType, typename CompatibleStringType>
+struct is_compatible_string_type
+{
+    static constexpr auto value =
+        is_constructible<typename BasicJsonType::string_t, CompatibleStringType>::value;
+};
+
+template<typename BasicJsonType, typename ConstructibleStringType>
+struct is_constructible_string_type
+{
+    // launder type through decltype() to fix compilation failure on ICPC
+#ifdef __INTEL_COMPILER
+    using laundered_type = decltype(std::declval<ConstructibleStringType>());
+#else
+    using laundered_type = ConstructibleStringType;
+#endif
+
+    static constexpr auto value =
+        conjunction <
+        is_constructible<laundered_type, typename BasicJsonType::string_t>,
+        is_detected_exact<typename BasicJsonType::string_t::value_type,
+        value_type_t, laundered_type >>::value;
+};
+
+template<typename BasicJsonType, typename CompatibleArrayType, typename = void>
+struct is_compatible_array_type_impl : std::false_type {};
+
+template<typename BasicJsonType, typename CompatibleArrayType>
+struct is_compatible_array_type_impl <
+    BasicJsonType, CompatibleArrayType,
+    enable_if_t <
+    is_detected<iterator_t, CompatibleArrayType>::value&&
+    is_iterator_traits<iterator_traits<detected_t<iterator_t, CompatibleArrayType>>>::value&&
+// special case for types like std::filesystem::path whose iterator's value_type are themselves
+// c.f. https://github.com/nlohmann/json/pull/3073
+    !std::is_same<CompatibleArrayType, detected_t<range_value_t, CompatibleArrayType>>::value >>
+{
+    static constexpr bool value =
+        is_constructible<BasicJsonType,
+        range_value_t<CompatibleArrayType>>::value;
+};
+
+template<typename BasicJsonType, typename CompatibleArrayType>
+struct is_compatible_array_type
+    : is_compatible_array_type_impl<BasicJsonType, CompatibleArrayType> {};
+
+template<typename BasicJsonType, typename ConstructibleArrayType, typename = void>
+struct is_constructible_array_type_impl : std::false_type {};
+
+template<typename BasicJsonType, typename ConstructibleArrayType>
+struct is_constructible_array_type_impl <
+    BasicJsonType, ConstructibleArrayType,
+    enable_if_t<std::is_same<ConstructibleArrayType,
+    typename BasicJsonType::value_type>::value >>
+            : std::true_type {};
+
+template<typename BasicJsonType, typename ConstructibleArrayType>
+struct is_constructible_array_type_impl <
+    BasicJsonType, ConstructibleArrayType,
+    enable_if_t < !std::is_same<ConstructibleArrayType,
+    typename BasicJsonType::value_type>::value&&
+    !is_compatible_string_type<BasicJsonType, ConstructibleArrayType>::value&&
+    is_default_constructible<ConstructibleArrayType>::value&&
+(std::is_move_assignable<ConstructibleArrayType>::value ||
+ std::is_copy_assignable<ConstructibleArrayType>::value)&&
+is_detected<iterator_t, ConstructibleArrayType>::value&&
+is_iterator_traits<iterator_traits<detected_t<iterator_t, ConstructibleArrayType>>>::value&&
+is_detected<range_value_t, ConstructibleArrayType>::value&&
+// special case for types like std::filesystem::path whose iterator's value_type are themselves
+// c.f. https://github.com/nlohmann/json/pull/3073
+!std::is_same<ConstructibleArrayType, detected_t<range_value_t, ConstructibleArrayType>>::value&&
+        is_complete_type <
+        detected_t<range_value_t, ConstructibleArrayType >>::value >>
+{
+    using value_type = range_value_t<ConstructibleArrayType>;
+
+    static constexpr bool value =
+        std::is_same<value_type,
+        typename BasicJsonType::array_t::value_type>::value ||
+        has_from_json<BasicJsonType,
+        value_type>::value ||
+        has_non_default_from_json <
+        BasicJsonType,
+        value_type >::value;
+};
+
+template<typename BasicJsonType, typename ConstructibleArrayType>
+struct is_constructible_array_type
+    : is_constructible_array_type_impl<BasicJsonType, ConstructibleArrayType> {};
+
+template<typename RealIntegerType, typename CompatibleNumberIntegerType,
+         typename = void>
+struct is_compatible_integer_type_impl : std::false_type {};
+
+template<typename RealIntegerType, typename CompatibleNumberIntegerType>
+struct is_compatible_integer_type_impl <
+    RealIntegerType, CompatibleNumberIntegerType,
+    enable_if_t < std::is_integral<RealIntegerType>::value&&
+    std::is_integral<CompatibleNumberIntegerType>::value&&
+    !std::is_same<bool, CompatibleNumberIntegerType>::value >>
+{
+    // is there an assert somewhere on overflows?
+    using RealLimits = std::numeric_limits<RealIntegerType>;
+    using CompatibleLimits = std::numeric_limits<CompatibleNumberIntegerType>;
+
+    static constexpr auto value =
+        is_constructible<RealIntegerType,
+        CompatibleNumberIntegerType>::value &&
+        CompatibleLimits::is_integer &&
+        RealLimits::is_signed == CompatibleLimits::is_signed;
+};
+
+template<typename RealIntegerType, typename CompatibleNumberIntegerType>
+struct is_compatible_integer_type
+    : is_compatible_integer_type_impl<RealIntegerType,
+      CompatibleNumberIntegerType> {};
+
+template<typename BasicJsonType, typename CompatibleType, typename = void>
+struct is_compatible_type_impl: std::false_type {};
+
+template<typename BasicJsonType, typename CompatibleType>
+struct is_compatible_type_impl <
+    BasicJsonType, CompatibleType,
+    enable_if_t<is_complete_type<CompatibleType>::value >>
+{
+    static constexpr bool value =
+        has_to_json<BasicJsonType, CompatibleType>::value;
+};
+
+template<typename BasicJsonType, typename CompatibleType>
+struct is_compatible_type
+    : is_compatible_type_impl<BasicJsonType, CompatibleType> {};
+
+template<typename T1, typename T2>
+struct is_constructible_tuple : std::false_type {};
+
+template<typename T1, typename... Args>
+struct is_constructible_tuple<T1, std::tuple<Args...>> : conjunction<is_constructible<T1, Args>...> {};
+
+template<typename BasicJsonType, typename T>
+struct is_json_iterator_of : std::false_type {};
+
+template<typename BasicJsonType>
+struct is_json_iterator_of<BasicJsonType, typename BasicJsonType::iterator> : std::true_type {};
+
+template<typename BasicJsonType>
+struct is_json_iterator_of<BasicJsonType, typename BasicJsonType::const_iterator> : std::true_type
+{};
+
+// checks if a given type T is a template specialization of Primary
+template<template <typename...> class Primary, typename T>
+struct is_specialization_of : std::false_type {};
+
+template<template <typename...> class Primary, typename... Args>
+struct is_specialization_of<Primary, Primary<Args...>> : std::true_type {};
+
+template<typename T>
+using is_json_pointer = is_specialization_of<::nlohmann::json_pointer, uncvref_t<T>>;
+
+// checks if A and B are comparable using Compare functor
+template<typename Compare, typename A, typename B, typename = void>
+struct is_comparable : std::false_type {};
+
+template<typename Compare, typename A, typename B>
+struct is_comparable<Compare, A, B, void_t<
+decltype(std::declval<Compare>()(std::declval<A>(), std::declval<B>())),
+decltype(std::declval<Compare>()(std::declval<B>(), std::declval<A>()))
+>> : std::true_type {};
+
+template<typename T>
+using detect_is_transparent = typename T::is_transparent;
+
+// type trait to check if KeyType can be used as object key (without a BasicJsonType)
+// see is_usable_as_basic_json_key_type below
+template<typename Comparator, typename ObjectKeyType, typename KeyTypeCVRef, bool RequireTransparentComparator = true,
+         bool ExcludeObjectKeyType = RequireTransparentComparator, typename KeyType = uncvref_t<KeyTypeCVRef>>
+using is_usable_as_key_type = typename std::conditional <
+                              is_comparable<Comparator, ObjectKeyType, KeyTypeCVRef>::value
+                              && !(ExcludeObjectKeyType && std::is_same<KeyType,
+                                   ObjectKeyType>::value)
+                              && (!RequireTransparentComparator
+                                  || is_detected <detect_is_transparent, Comparator>::value)
+                              && !is_json_pointer<KeyType>::value,
+                              std::true_type,
+                              std::false_type >::type;
+
+// type trait to check if KeyType can be used as object key
+// true if:
+//   - KeyType is comparable with BasicJsonType::object_t::key_type
+//   - if ExcludeObjectKeyType is true, KeyType is not BasicJsonType::object_t::key_type
+//   - the comparator is transparent or RequireTransparentComparator is false
+//   - KeyType is not a JSON iterator or json_pointer
+template<typename BasicJsonType, typename KeyTypeCVRef, bool RequireTransparentComparator = true,
+         bool ExcludeObjectKeyType = RequireTransparentComparator, typename KeyType = uncvref_t<KeyTypeCVRef>>
+using is_usable_as_basic_json_key_type = typename std::conditional <
+        is_usable_as_key_type<typename BasicJsonType::object_comparator_t,
+        typename BasicJsonType::object_t::key_type, KeyTypeCVRef,
+        RequireTransparentComparator, ExcludeObjectKeyType>::value
+        && !is_json_iterator_of<BasicJsonType, KeyType>::value,
+        std::true_type,
+        std::false_type >::type;
+
+template<typename ObjectType, typename KeyType>
+using detect_erase_with_key_type = decltype(std::declval<ObjectType&>().erase(std::declval<KeyType>()));
+
+// type trait to check if object_t has an erase() member functions accepting KeyType
+template<typename BasicJsonType, typename KeyType>
+using has_erase_with_key_type = typename std::conditional <
+                                is_detected <
+                                detect_erase_with_key_type,
+                                typename BasicJsonType::object_t, KeyType >::value,
+                                std::true_type,
+                                std::false_type >::type;
+
+// a naive helper to check if a type is an ordered_map (exploits the fact that
+// ordered_map inherits capacity() from std::vector)
+template <typename T>
+struct is_ordered_map
+{
+    using one = char;
+
+    struct two
+    {
+        char x[2]; // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+    };
+
+    template <typename C> static one test( decltype(&C::capacity) ) ;
+    template <typename C> static two test(...);
+
+    enum { value = sizeof(test<T>(nullptr)) == sizeof(char) }; // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+};
+
+// to avoid useless casts (see https://github.com/nlohmann/json/issues/2893#issuecomment-889152324)
+template < typename T, typename U, enable_if_t < !std::is_same<T, U>::value, int > = 0 >
+T conditional_static_cast(U value)
+{
+    return static_cast<T>(value);
+}
+
+template<typename T, typename U, enable_if_t<std::is_same<T, U>::value, int> = 0>
+T conditional_static_cast(U value)
+{
+    return value;
+}
+
+template<typename... Types>
+using all_integral = conjunction<std::is_integral<Types>...>;
+
+template<typename... Types>
+using all_signed = conjunction<std::is_signed<Types>...>;
+
+template<typename... Types>
+using all_unsigned = conjunction<std::is_unsigned<Types>...>;
+
+// there's a disjunction trait in another PR; replace when merged
+template<typename... Types>
+using same_sign = std::integral_constant < bool,
+      all_signed<Types...>::value || all_unsigned<Types...>::value >;
+
+template<typename OfType, typename T>
+using never_out_of_range = std::integral_constant < bool,
+      (std::is_signed<OfType>::value && (sizeof(T) < sizeof(OfType)))
+      || (same_sign<OfType, T>::value && sizeof(OfType) == sizeof(T)) >;
+
+template<typename OfType, typename T,
+         bool OfTypeSigned = std::is_signed<OfType>::value,
+         bool TSigned = std::is_signed<T>::value>
+struct value_in_range_of_impl2;
+
+template<typename OfType, typename T>
+struct value_in_range_of_impl2<OfType, T, false, false>
+{
+    static constexpr bool test(T val)
+    {
+        using CommonType = typename std::common_type<OfType, T>::type;
+        return static_cast<CommonType>(val) <= static_cast<CommonType>((std::numeric_limits<OfType>::max)());
+    }
+};
+
+template<typename OfType, typename T>
+struct value_in_range_of_impl2<OfType, T, true, false>
+{
+    static constexpr bool test(T val)
+    {
+        using CommonType = typename std::common_type<OfType, T>::type;
+        return static_cast<CommonType>(val) <= static_cast<CommonType>((std::numeric_limits<OfType>::max)());
+    }
+};
+
+template<typename OfType, typename T>
+struct value_in_range_of_impl2<OfType, T, false, true>
+{
+    static constexpr bool test(T val)
+    {
+        using CommonType = typename std::common_type<OfType, T>::type;
+        return val >= 0 && static_cast<CommonType>(val) <= static_cast<CommonType>((std::numeric_limits<OfType>::max)());
+    }
+};
+
+template<typename OfType, typename T>
+struct value_in_range_of_impl2<OfType, T, true, true>
+{
+    static constexpr bool test(T val)
+    {
+        using CommonType = typename std::common_type<OfType, T>::type;
+        return static_cast<CommonType>(val) >= static_cast<CommonType>((std::numeric_limits<OfType>::min)())
+               && static_cast<CommonType>(val) <= static_cast<CommonType>((std::numeric_limits<OfType>::max)());
+    }
+};
+
+template<typename OfType, typename T,
+         bool NeverOutOfRange = never_out_of_range<OfType, T>::value,
+         typename = detail::enable_if_t<all_integral<OfType, T>::value>>
+struct value_in_range_of_impl1;
+
+template<typename OfType, typename T>
+struct value_in_range_of_impl1<OfType, T, false>
+{
+    static constexpr bool test(T val)
+    {
+        return value_in_range_of_impl2<OfType, T>::test(val);
+    }
+};
+
+template<typename OfType, typename T>
+struct value_in_range_of_impl1<OfType, T, true>
+{
+    static constexpr bool test(T /*val*/)
+    {
+        return true;
+    }
+};
+
+template<typename OfType, typename T>
+inline constexpr bool value_in_range_of(T val)
+{
+    return value_in_range_of_impl1<OfType, T>::test(val);
+}
+
+template<bool Value>
+using bool_constant = std::integral_constant<bool, Value>;
+
+///////////////////////////////////////////////////////////////////////////////
+// is_c_string
+///////////////////////////////////////////////////////////////////////////////
+
+namespace impl
+{
+
+template<typename T>
+inline constexpr bool is_c_string()
+{
+    using TUnExt = typename std::remove_extent<T>::type;
+    using TUnCVExt = typename std::remove_cv<TUnExt>::type;
+    using TUnPtr = typename std::remove_pointer<T>::type;
+    using TUnCVPtr = typename std::remove_cv<TUnPtr>::type;
+    return
+        (std::is_array<T>::value && std::is_same<TUnCVExt, char>::value)
+        || (std::is_pointer<T>::value && std::is_same<TUnCVPtr, char>::value);
+}
+
+}  // namespace impl
+
+// checks whether T is a [cv] char */[cv] char[] C string
+template<typename T>
+struct is_c_string : bool_constant<impl::is_c_string<T>()> {};
+
+template<typename T>
+using is_c_string_uncvref = is_c_string<uncvref_t<T>>;
+
+///////////////////////////////////////////////////////////////////////////////
+// is_transparent
+///////////////////////////////////////////////////////////////////////////////
+
+namespace impl
+{
+
+template<typename T>
+inline constexpr bool is_transparent()
+{
+    return is_detected<detect_is_transparent, T>::value;
+}
+
+}  // namespace impl
+
+// checks whether T has a member named is_transparent
+template<typename T>
+struct is_transparent : bool_constant<impl::is_transparent<T>()> {};
+
+///////////////////////////////////////////////////////////////////////////////
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/string_concat.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstring> // strlen
+#include <string> // string
+#include <utility> // forward
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/detected.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+inline std::size_t concat_length()
+{
+    return 0;
+}
+
+template<typename... Args>
+inline std::size_t concat_length(const char* cstr, const Args& ... rest);
+
+template<typename StringType, typename... Args>
+inline std::size_t concat_length(const StringType& str, const Args& ... rest);
+
+template<typename... Args>
+inline std::size_t concat_length(const char /*c*/, const Args& ... rest)
+{
+    return 1 + concat_length(rest...);
+}
+
+template<typename... Args>
+inline std::size_t concat_length(const char* cstr, const Args& ... rest)
+{
+    // cppcheck-suppress ignoredReturnValue
+    return ::strlen(cstr) + concat_length(rest...);
+}
+
+template<typename StringType, typename... Args>
+inline std::size_t concat_length(const StringType& str, const Args& ... rest)
+{
+    return str.size() + concat_length(rest...);
+}
+
+template<typename OutStringType>
+inline void concat_into(OutStringType& /*out*/)
+{}
+
+template<typename StringType, typename Arg>
+using string_can_append = decltype(std::declval<StringType&>().append(std::declval < Arg && > ()));
+
+template<typename StringType, typename Arg>
+using detect_string_can_append = is_detected<string_can_append, StringType, Arg>;
+
+template<typename StringType, typename Arg>
+using string_can_append_op = decltype(std::declval<StringType&>() += std::declval < Arg && > ());
+
+template<typename StringType, typename Arg>
+using detect_string_can_append_op = is_detected<string_can_append_op, StringType, Arg>;
+
+template<typename StringType, typename Arg>
+using string_can_append_iter = decltype(std::declval<StringType&>().append(std::declval<const Arg&>().begin(), std::declval<const Arg&>().end()));
+
+template<typename StringType, typename Arg>
+using detect_string_can_append_iter = is_detected<string_can_append_iter, StringType, Arg>;
+
+template<typename StringType, typename Arg>
+using string_can_append_data = decltype(std::declval<StringType&>().append(std::declval<const Arg&>().data(), std::declval<const Arg&>().size()));
+
+template<typename StringType, typename Arg>
+using detect_string_can_append_data = is_detected<string_can_append_data, StringType, Arg>;
+
+template < typename OutStringType, typename Arg, typename... Args,
+           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
+                         && detect_string_can_append_op<OutStringType, Arg>::value, int > = 0 >
+inline void concat_into(OutStringType& out, Arg && arg, Args && ... rest);
+
+template < typename OutStringType, typename Arg, typename... Args,
+           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
+                         && !detect_string_can_append_op<OutStringType, Arg>::value
+                         && detect_string_can_append_iter<OutStringType, Arg>::value, int > = 0 >
+inline void concat_into(OutStringType& out, const Arg& arg, Args && ... rest);
+
+template < typename OutStringType, typename Arg, typename... Args,
+           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
+                         && !detect_string_can_append_op<OutStringType, Arg>::value
+                         && !detect_string_can_append_iter<OutStringType, Arg>::value
+                         && detect_string_can_append_data<OutStringType, Arg>::value, int > = 0 >
+inline void concat_into(OutStringType& out, const Arg& arg, Args && ... rest);
+
+template<typename OutStringType, typename Arg, typename... Args,
+         enable_if_t<detect_string_can_append<OutStringType, Arg>::value, int> = 0>
+inline void concat_into(OutStringType& out, Arg && arg, Args && ... rest)
+{
+    out.append(std::forward<Arg>(arg));
+    concat_into(out, std::forward<Args>(rest)...);
+}
+
+template < typename OutStringType, typename Arg, typename... Args,
+           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
+                         && detect_string_can_append_op<OutStringType, Arg>::value, int > >
+inline void concat_into(OutStringType& out, Arg&& arg, Args&& ... rest)
+{
+    out += std::forward<Arg>(arg);
+    concat_into(out, std::forward<Args>(rest)...);
+}
+
+template < typename OutStringType, typename Arg, typename... Args,
+           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
+                         && !detect_string_can_append_op<OutStringType, Arg>::value
+                         && detect_string_can_append_iter<OutStringType, Arg>::value, int > >
+inline void concat_into(OutStringType& out, const Arg& arg, Args&& ... rest)
+{
+    out.append(arg.begin(), arg.end());
+    concat_into(out, std::forward<Args>(rest)...);
+}
+
+template < typename OutStringType, typename Arg, typename... Args,
+           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
+                         && !detect_string_can_append_op<OutStringType, Arg>::value
+                         && !detect_string_can_append_iter<OutStringType, Arg>::value
+                         && detect_string_can_append_data<OutStringType, Arg>::value, int > >
+inline void concat_into(OutStringType& out, const Arg& arg, Args&& ... rest)
+{
+    out.append(arg.data(), arg.size());
+    concat_into(out, std::forward<Args>(rest)...);
+}
+
+template<typename OutStringType = std::string, typename... Args>
+inline OutStringType concat(Args && ... args)
+{
+    OutStringType str;
+    str.reserve(concat_length(args...));
+    concat_into(str, std::forward<Args>(args)...);
+    return str;
+}
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+////////////////
+// exceptions //
+////////////////
+
+/// @brief general exception of the @ref basic_json class
+/// @sa https://json.nlohmann.me/api/basic_json/exception/
+class exception : public std::exception
+{
+  public:
+    /// returns the explanatory string
+    const char* what() const noexcept override
+    {
+        return m.what();
+    }
+
+    /// the id of the exception
+    const int id; // NOLINT(cppcoreguidelines-non-private-member-variables-in-classes)
+
+  protected:
+    JSON_HEDLEY_NON_NULL(3)
+    exception(int id_, const char* what_arg) : id(id_), m(what_arg) {} // NOLINT(bugprone-throw-keyword-missing)
+
+    static std::string name(const std::string& ename, int id_)
+    {
+        return concat("[json.exception.", ename, '.', std::to_string(id_), "] ");
+    }
+
+    static std::string diagnostics(std::nullptr_t /*leaf_element*/)
+    {
+        return "";
+    }
+
+    template<typename BasicJsonType>
+    static std::string diagnostics(const BasicJsonType* leaf_element)
+    {
+#if JSON_DIAGNOSTICS
+        std::vector<std::string> tokens;
+        for (const auto* current = leaf_element; current != nullptr && current->m_parent != nullptr; current = current->m_parent)
+        {
+            switch (current->m_parent->type())
+            {
+                case value_t::array:
+                {
+                    for (std::size_t i = 0; i < current->m_parent->m_data.m_value.array->size(); ++i)
+                    {
+                        if (&current->m_parent->m_data.m_value.array->operator[](i) == current)
+                        {
+                            tokens.emplace_back(std::to_string(i));
+                            break;
+                        }
+                    }
+                    break;
+                }
+
+                case value_t::object:
+                {
+                    for (const auto& element : *current->m_parent->m_data.m_value.object)
+                    {
+                        if (&element.second == current)
+                        {
+                            tokens.emplace_back(element.first.c_str());
+                            break;
+                        }
+                    }
+                    break;
+                }
+
+                case value_t::null: // LCOV_EXCL_LINE
+                case value_t::string: // LCOV_EXCL_LINE
+                case value_t::boolean: // LCOV_EXCL_LINE
+                case value_t::number_integer: // LCOV_EXCL_LINE
+                case value_t::number_unsigned: // LCOV_EXCL_LINE
+                case value_t::number_float: // LCOV_EXCL_LINE
+                case value_t::binary: // LCOV_EXCL_LINE
+                case value_t::discarded: // LCOV_EXCL_LINE
+                default:   // LCOV_EXCL_LINE
+                    break; // LCOV_EXCL_LINE
+            }
+        }
+
+        if (tokens.empty())
+        {
+            return "";
+        }
+
+        auto str = std::accumulate(tokens.rbegin(), tokens.rend(), std::string{},
+                                   [](const std::string & a, const std::string & b)
+        {
+            return concat(a, '/', detail::escape(b));
+        });
+        return concat('(', str, ") ");
+#else
+        static_cast<void>(leaf_element);
+        return "";
+#endif
+    }
+
+  private:
+    /// an exception object as storage for error messages
+    std::runtime_error m;
+};
+
+/// @brief exception indicating a parse error
+/// @sa https://json.nlohmann.me/api/basic_json/parse_error/
+class parse_error : public exception
+{
+  public:
+    /*!
+    @brief create a parse error exception
+    @param[in] id_       the id of the exception
+    @param[in] pos       the position where the error occurred (or with
+                         chars_read_total=0 if the position cannot be
+                         determined)
+    @param[in] what_arg  the explanatory string
+    @return parse_error object
+    */
+    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
+    static parse_error create(int id_, const position_t& pos, const std::string& what_arg, BasicJsonContext context)
+    {
+        const std::string w = concat(exception::name("parse_error", id_), "parse error",
+                                     position_string(pos), ": ", exception::diagnostics(context), what_arg);
+        return {id_, pos.chars_read_total, w.c_str()};
+    }
+
+    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
+    static parse_error create(int id_, std::size_t byte_, const std::string& what_arg, BasicJsonContext context)
+    {
+        const std::string w = concat(exception::name("parse_error", id_), "parse error",
+                                     (byte_ != 0 ? (concat(" at byte ", std::to_string(byte_))) : ""),
+                                     ": ", exception::diagnostics(context), what_arg);
+        return {id_, byte_, w.c_str()};
+    }
+
+    /*!
+    @brief byte index of the parse error
+
+    The byte index of the last read character in the input file.
+
+    @note For an input with n bytes, 1 is the index of the first character and
+          n+1 is the index of the terminating null byte or the end of file.
+          This also holds true when reading a byte vector (CBOR or MessagePack).
+    */
+    const std::size_t byte;
+
+  private:
+    parse_error(int id_, std::size_t byte_, const char* what_arg)
+        : exception(id_, what_arg), byte(byte_) {}
+
+    static std::string position_string(const position_t& pos)
+    {
+        return concat(" at line ", std::to_string(pos.lines_read + 1),
+                      ", column ", std::to_string(pos.chars_read_current_line));
+    }
+};
+
+/// @brief exception indicating errors with iterators
+/// @sa https://json.nlohmann.me/api/basic_json/invalid_iterator/
+class invalid_iterator : public exception
+{
+  public:
+    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
+    static invalid_iterator create(int id_, const std::string& what_arg, BasicJsonContext context)
+    {
+        const std::string w = concat(exception::name("invalid_iterator", id_), exception::diagnostics(context), what_arg);
+        return {id_, w.c_str()};
+    }
+
+  private:
+    JSON_HEDLEY_NON_NULL(3)
+    invalid_iterator(int id_, const char* what_arg)
+        : exception(id_, what_arg) {}
+};
+
+/// @brief exception indicating executing a member function with a wrong type
+/// @sa https://json.nlohmann.me/api/basic_json/type_error/
+class type_error : public exception
+{
+  public:
+    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
+    static type_error create(int id_, const std::string& what_arg, BasicJsonContext context)
+    {
+        const std::string w = concat(exception::name("type_error", id_), exception::diagnostics(context), what_arg);
+        return {id_, w.c_str()};
+    }
+
+  private:
+    JSON_HEDLEY_NON_NULL(3)
+    type_error(int id_, const char* what_arg) : exception(id_, what_arg) {}
+};
+
+/// @brief exception indicating access out of the defined range
+/// @sa https://json.nlohmann.me/api/basic_json/out_of_range/
+class out_of_range : public exception
+{
+  public:
+    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
+    static out_of_range create(int id_, const std::string& what_arg, BasicJsonContext context)
+    {
+        const std::string w = concat(exception::name("out_of_range", id_), exception::diagnostics(context), what_arg);
+        return {id_, w.c_str()};
+    }
+
+  private:
+    JSON_HEDLEY_NON_NULL(3)
+    out_of_range(int id_, const char* what_arg) : exception(id_, what_arg) {}
+};
+
+/// @brief exception indicating other library errors
+/// @sa https://json.nlohmann.me/api/basic_json/other_error/
+class other_error : public exception
+{
+  public:
+    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
+    static other_error create(int id_, const std::string& what_arg, BasicJsonContext context)
+    {
+        const std::string w = concat(exception::name("other_error", id_), exception::diagnostics(context), what_arg);
+        return {id_, w.c_str()};
+    }
+
+  private:
+    JSON_HEDLEY_NON_NULL(3)
+    other_error(int id_, const char* what_arg) : exception(id_, what_arg) {}
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/identity_tag.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+// dispatching helper struct
+template <class T> struct identity_tag {};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/meta/std_fs.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+#if JSON_HAS_EXPERIMENTAL_FILESYSTEM
+#include <experimental/filesystem>
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+namespace std_fs = std::experimental::filesystem;
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+#elif JSON_HAS_FILESYSTEM
+#include <filesystem>
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+namespace std_fs = std::filesystem;
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+#endif
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, typename std::nullptr_t& n)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_null()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be null, but is ", j.type_name()), &j));
+    }
+    n = nullptr;
+}
+
+// overloads for basic_json template parameters
+template < typename BasicJsonType, typename ArithmeticType,
+           enable_if_t < std::is_arithmetic<ArithmeticType>::value&&
+                         !std::is_same<ArithmeticType, typename BasicJsonType::boolean_t>::value,
+                         int > = 0 >
+void get_arithmetic_value(const BasicJsonType& j, ArithmeticType& val)
+{
+    switch (static_cast<value_t>(j))
+    {
+        case value_t::number_unsigned:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_unsigned_t*>());
+            break;
+        }
+        case value_t::number_integer:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_integer_t*>());
+            break;
+        }
+        case value_t::number_float:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_float_t*>());
+            break;
+        }
+
+        case value_t::null:
+        case value_t::object:
+        case value_t::array:
+        case value_t::string:
+        case value_t::boolean:
+        case value_t::binary:
+        case value_t::discarded:
+        default:
+            JSON_THROW(type_error::create(302, concat("type must be number, but is ", j.type_name()), &j));
+    }
+}
+
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, typename BasicJsonType::boolean_t& b)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_boolean()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be boolean, but is ", j.type_name()), &j));
+    }
+    b = *j.template get_ptr<const typename BasicJsonType::boolean_t*>();
+}
+
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, typename BasicJsonType::string_t& s)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_string()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be string, but is ", j.type_name()), &j));
+    }
+    s = *j.template get_ptr<const typename BasicJsonType::string_t*>();
+}
+
+template <
+    typename BasicJsonType, typename StringType,
+    enable_if_t <
+        std::is_assignable<StringType&, const typename BasicJsonType::string_t>::value
+        && is_detected_exact<typename BasicJsonType::string_t::value_type, value_type_t, StringType>::value
+        && !std::is_same<typename BasicJsonType::string_t, StringType>::value
+        && !is_json_ref<StringType>::value, int > = 0 >
+inline void from_json(const BasicJsonType& j, StringType& s)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_string()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be string, but is ", j.type_name()), &j));
+    }
+
+    s = *j.template get_ptr<const typename BasicJsonType::string_t*>();
+}
+
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, typename BasicJsonType::number_float_t& val)
+{
+    get_arithmetic_value(j, val);
+}
+
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, typename BasicJsonType::number_unsigned_t& val)
+{
+    get_arithmetic_value(j, val);
+}
+
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, typename BasicJsonType::number_integer_t& val)
+{
+    get_arithmetic_value(j, val);
+}
+
+#if !JSON_DISABLE_ENUM_SERIALIZATION
+template<typename BasicJsonType, typename EnumType,
+         enable_if_t<std::is_enum<EnumType>::value, int> = 0>
+inline void from_json(const BasicJsonType& j, EnumType& e)
+{
+    typename std::underlying_type<EnumType>::type val;
+    get_arithmetic_value(j, val);
+    e = static_cast<EnumType>(val);
+}
+#endif  // JSON_DISABLE_ENUM_SERIALIZATION
+
+// forward_list doesn't have an insert method
+template<typename BasicJsonType, typename T, typename Allocator,
+         enable_if_t<is_getable<BasicJsonType, T>::value, int> = 0>
+inline void from_json(const BasicJsonType& j, std::forward_list<T, Allocator>& l)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
+    }
+    l.clear();
+    std::transform(j.rbegin(), j.rend(),
+                   std::front_inserter(l), [](const BasicJsonType & i)
+    {
+        return i.template get<T>();
+    });
+}
+
+// valarray doesn't have an insert method
+template<typename BasicJsonType, typename T,
+         enable_if_t<is_getable<BasicJsonType, T>::value, int> = 0>
+inline void from_json(const BasicJsonType& j, std::valarray<T>& l)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
+    }
+    l.resize(j.size());
+    std::transform(j.begin(), j.end(), std::begin(l),
+                   [](const BasicJsonType & elem)
+    {
+        return elem.template get<T>();
+    });
+}
+
+template<typename BasicJsonType, typename T, std::size_t N>
+auto from_json(const BasicJsonType& j, T (&arr)[N])  // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+-> decltype(j.template get<T>(), void())
+{
+    for (std::size_t i = 0; i < N; ++i)
+    {
+        arr[i] = j.at(i).template get<T>();
+    }
+}
+
+template<typename BasicJsonType>
+inline void from_json_array_impl(const BasicJsonType& j, typename BasicJsonType::array_t& arr, priority_tag<3> /*unused*/)
+{
+    arr = *j.template get_ptr<const typename BasicJsonType::array_t*>();
+}
+
+template<typename BasicJsonType, typename T, std::size_t N>
+auto from_json_array_impl(const BasicJsonType& j, std::array<T, N>& arr,
+                          priority_tag<2> /*unused*/)
+-> decltype(j.template get<T>(), void())
+{
+    for (std::size_t i = 0; i < N; ++i)
+    {
+        arr[i] = j.at(i).template get<T>();
+    }
+}
+
+template<typename BasicJsonType, typename ConstructibleArrayType,
+         enable_if_t<
+             std::is_assignable<ConstructibleArrayType&, ConstructibleArrayType>::value,
+             int> = 0>
+auto from_json_array_impl(const BasicJsonType& j, ConstructibleArrayType& arr, priority_tag<1> /*unused*/)
+-> decltype(
+    arr.reserve(std::declval<typename ConstructibleArrayType::size_type>()),
+    j.template get<typename ConstructibleArrayType::value_type>(),
+    void())
+{
+    using std::end;
+
+    ConstructibleArrayType ret;
+    ret.reserve(j.size());
+    std::transform(j.begin(), j.end(),
+                   std::inserter(ret, end(ret)), [](const BasicJsonType & i)
+    {
+        // get<BasicJsonType>() returns *this, this won't call a from_json
+        // method when value_type is BasicJsonType
+        return i.template get<typename ConstructibleArrayType::value_type>();
+    });
+    arr = std::move(ret);
+}
+
+template<typename BasicJsonType, typename ConstructibleArrayType,
+         enable_if_t<
+             std::is_assignable<ConstructibleArrayType&, ConstructibleArrayType>::value,
+             int> = 0>
+inline void from_json_array_impl(const BasicJsonType& j, ConstructibleArrayType& arr,
+                                 priority_tag<0> /*unused*/)
+{
+    using std::end;
+
+    ConstructibleArrayType ret;
+    std::transform(
+        j.begin(), j.end(), std::inserter(ret, end(ret)),
+        [](const BasicJsonType & i)
+    {
+        // get<BasicJsonType>() returns *this, this won't call a from_json
+        // method when value_type is BasicJsonType
+        return i.template get<typename ConstructibleArrayType::value_type>();
+    });
+    arr = std::move(ret);
+}
+
+template < typename BasicJsonType, typename ConstructibleArrayType,
+           enable_if_t <
+               is_constructible_array_type<BasicJsonType, ConstructibleArrayType>::value&&
+               !is_constructible_object_type<BasicJsonType, ConstructibleArrayType>::value&&
+               !is_constructible_string_type<BasicJsonType, ConstructibleArrayType>::value&&
+               !std::is_same<ConstructibleArrayType, typename BasicJsonType::binary_t>::value&&
+               !is_basic_json<ConstructibleArrayType>::value,
+               int > = 0 >
+auto from_json(const BasicJsonType& j, ConstructibleArrayType& arr)
+-> decltype(from_json_array_impl(j, arr, priority_tag<3> {}),
+j.template get<typename ConstructibleArrayType::value_type>(),
+void())
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
+    }
+
+    from_json_array_impl(j, arr, priority_tag<3> {});
+}
+
+template < typename BasicJsonType, typename T, std::size_t... Idx >
+std::array<T, sizeof...(Idx)> from_json_inplace_array_impl(BasicJsonType&& j,
+        identity_tag<std::array<T, sizeof...(Idx)>> /*unused*/, index_sequence<Idx...> /*unused*/)
+{
+    return { { std::forward<BasicJsonType>(j).at(Idx).template get<T>()... } };
+}
+
+template < typename BasicJsonType, typename T, std::size_t N >
+auto from_json(BasicJsonType&& j, identity_tag<std::array<T, N>> tag)
+-> decltype(from_json_inplace_array_impl(std::forward<BasicJsonType>(j), tag, make_index_sequence<N> {}))
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
+    }
+
+    return from_json_inplace_array_impl(std::forward<BasicJsonType>(j), tag, make_index_sequence<N> {});
+}
+
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, typename BasicJsonType::binary_t& bin)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_binary()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be binary, but is ", j.type_name()), &j));
+    }
+
+    bin = *j.template get_ptr<const typename BasicJsonType::binary_t*>();
+}
+
+template<typename BasicJsonType, typename ConstructibleObjectType,
+         enable_if_t<is_constructible_object_type<BasicJsonType, ConstructibleObjectType>::value, int> = 0>
+inline void from_json(const BasicJsonType& j, ConstructibleObjectType& obj)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_object()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be object, but is ", j.type_name()), &j));
+    }
+
+    ConstructibleObjectType ret;
+    const auto* inner_object = j.template get_ptr<const typename BasicJsonType::object_t*>();
+    using value_type = typename ConstructibleObjectType::value_type;
+    std::transform(
+        inner_object->begin(), inner_object->end(),
+        std::inserter(ret, ret.begin()),
+        [](typename BasicJsonType::object_t::value_type const & p)
+    {
+        return value_type(p.first, p.second.template get<typename ConstructibleObjectType::mapped_type>());
+    });
+    obj = std::move(ret);
+}
+
+// overload for arithmetic types, not chosen for basic_json template arguments
+// (BooleanType, etc..); note: Is it really necessary to provide explicit
+// overloads for boolean_t etc. in case of a custom BooleanType which is not
+// an arithmetic type?
+template < typename BasicJsonType, typename ArithmeticType,
+           enable_if_t <
+               std::is_arithmetic<ArithmeticType>::value&&
+               !std::is_same<ArithmeticType, typename BasicJsonType::number_unsigned_t>::value&&
+               !std::is_same<ArithmeticType, typename BasicJsonType::number_integer_t>::value&&
+               !std::is_same<ArithmeticType, typename BasicJsonType::number_float_t>::value&&
+               !std::is_same<ArithmeticType, typename BasicJsonType::boolean_t>::value,
+               int > = 0 >
+inline void from_json(const BasicJsonType& j, ArithmeticType& val)
+{
+    switch (static_cast<value_t>(j))
+    {
+        case value_t::number_unsigned:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_unsigned_t*>());
+            break;
+        }
+        case value_t::number_integer:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_integer_t*>());
+            break;
+        }
+        case value_t::number_float:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_float_t*>());
+            break;
+        }
+        case value_t::boolean:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::boolean_t*>());
+            break;
+        }
+
+        case value_t::null:
+        case value_t::object:
+        case value_t::array:
+        case value_t::string:
+        case value_t::binary:
+        case value_t::discarded:
+        default:
+            JSON_THROW(type_error::create(302, concat("type must be number, but is ", j.type_name()), &j));
+    }
+}
+
+template<typename BasicJsonType, typename... Args, std::size_t... Idx>
+std::tuple<Args...> from_json_tuple_impl_base(BasicJsonType&& j, index_sequence<Idx...> /*unused*/)
+{
+    return std::make_tuple(std::forward<BasicJsonType>(j).at(Idx).template get<Args>()...);
+}
+
+template < typename BasicJsonType, class A1, class A2 >
+std::pair<A1, A2> from_json_tuple_impl(BasicJsonType&& j, identity_tag<std::pair<A1, A2>> /*unused*/, priority_tag<0> /*unused*/)
+{
+    return {std::forward<BasicJsonType>(j).at(0).template get<A1>(),
+            std::forward<BasicJsonType>(j).at(1).template get<A2>()};
+}
+
+template<typename BasicJsonType, typename A1, typename A2>
+inline void from_json_tuple_impl(BasicJsonType&& j, std::pair<A1, A2>& p, priority_tag<1> /*unused*/)
+{
+    p = from_json_tuple_impl(std::forward<BasicJsonType>(j), identity_tag<std::pair<A1, A2>> {}, priority_tag<0> {});
+}
+
+template<typename BasicJsonType, typename... Args>
+std::tuple<Args...> from_json_tuple_impl(BasicJsonType&& j, identity_tag<std::tuple<Args...>> /*unused*/, priority_tag<2> /*unused*/)
+{
+    return from_json_tuple_impl_base<BasicJsonType, Args...>(std::forward<BasicJsonType>(j), index_sequence_for<Args...> {});
+}
+
+template<typename BasicJsonType, typename... Args>
+inline void from_json_tuple_impl(BasicJsonType&& j, std::tuple<Args...>& t, priority_tag<3> /*unused*/)
+{
+    t = from_json_tuple_impl_base<BasicJsonType, Args...>(std::forward<BasicJsonType>(j), index_sequence_for<Args...> {});
+}
+
+template<typename BasicJsonType, typename TupleRelated>
+auto from_json(BasicJsonType&& j, TupleRelated&& t)
+-> decltype(from_json_tuple_impl(std::forward<BasicJsonType>(j), std::forward<TupleRelated>(t), priority_tag<3> {}))
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
+    }
+
+    return from_json_tuple_impl(std::forward<BasicJsonType>(j), std::forward<TupleRelated>(t), priority_tag<3> {});
+}
+
+template < typename BasicJsonType, typename Key, typename Value, typename Compare, typename Allocator,
+           typename = enable_if_t < !std::is_constructible <
+                                        typename BasicJsonType::string_t, Key >::value >>
+inline void from_json(const BasicJsonType& j, std::map<Key, Value, Compare, Allocator>& m)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
+    }
+    m.clear();
+    for (const auto& p : j)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!p.is_array()))
+        {
+            JSON_THROW(type_error::create(302, concat("type must be array, but is ", p.type_name()), &j));
+        }
+        m.emplace(p.at(0).template get<Key>(), p.at(1).template get<Value>());
+    }
+}
+
+template < typename BasicJsonType, typename Key, typename Value, typename Hash, typename KeyEqual, typename Allocator,
+           typename = enable_if_t < !std::is_constructible <
+                                        typename BasicJsonType::string_t, Key >::value >>
+inline void from_json(const BasicJsonType& j, std::unordered_map<Key, Value, Hash, KeyEqual, Allocator>& m)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
+    }
+    m.clear();
+    for (const auto& p : j)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!p.is_array()))
+        {
+            JSON_THROW(type_error::create(302, concat("type must be array, but is ", p.type_name()), &j));
+        }
+        m.emplace(p.at(0).template get<Key>(), p.at(1).template get<Value>());
+    }
+}
+
+#if JSON_HAS_FILESYSTEM || JSON_HAS_EXPERIMENTAL_FILESYSTEM
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, std_fs::path& p)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_string()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be string, but is ", j.type_name()), &j));
+    }
+    p = *j.template get_ptr<const typename BasicJsonType::string_t*>();
+}
+#endif
+
+struct from_json_fn
+{
+    template<typename BasicJsonType, typename T>
+    auto operator()(const BasicJsonType& j, T&& val) const
+    noexcept(noexcept(from_json(j, std::forward<T>(val))))
+    -> decltype(from_json(j, std::forward<T>(val)))
+    {
+        return from_json(j, std::forward<T>(val));
+    }
+};
+
+}  // namespace detail
+
+#ifndef JSON_HAS_CPP_17
+/// namespace to hold default `from_json` function
+/// to see why this is required:
+/// http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4381.html
+namespace // NOLINT(cert-dcl59-cpp,fuchsia-header-anon-namespaces,google-build-namespaces)
+{
+#endif
+JSON_INLINE_VARIABLE constexpr const auto& from_json = // NOLINT(misc-definitions-in-headers)
+    detail::static_const<detail::from_json_fn>::value;
+#ifndef JSON_HAS_CPP_17
+}  // namespace
+#endif
+
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/conversions/to_json.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <algorithm> // copy
+#include <iterator> // begin, end
+#include <string> // string
+#include <tuple> // tuple, get
+#include <type_traits> // is_same, is_constructible, is_floating_point, is_enum, underlying_type
+#include <utility> // move, forward, declval, pair
+#include <valarray> // valarray
+#include <vector> // vector
+
+// #include <nlohmann/detail/iterators/iteration_proxy.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstddef> // size_t
+#include <iterator> // input_iterator_tag
+#include <string> // string, to_string
+#include <tuple> // tuple_size, get, tuple_element
+#include <utility> // move
+
+#if JSON_HAS_RANGES
+    #include <ranges> // enable_borrowed_range
+#endif
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+template<typename string_type>
+void int_to_string( string_type& target, std::size_t value )
+{
+    // For ADL
+    using std::to_string;
+    target = to_string(value);
+}
+template<typename IteratorType> class iteration_proxy_value
+{
+  public:
+    using difference_type = std::ptrdiff_t;
+    using value_type = iteration_proxy_value;
+    using pointer = value_type *;
+    using reference = value_type &;
+    using iterator_category = std::input_iterator_tag;
+    using string_type = typename std::remove_cv< typename std::remove_reference<decltype( std::declval<IteratorType>().key() ) >::type >::type;
+
+  private:
+    /// the iterator
+    IteratorType anchor{};
+    /// an index for arrays (used to create key names)
+    std::size_t array_index = 0;
+    /// last stringified array index
+    mutable std::size_t array_index_last = 0;
+    /// a string representation of the array index
+    mutable string_type array_index_str = "0";
+    /// an empty string (to return a reference for primitive values)
+    string_type empty_str{};
+
+  public:
+    explicit iteration_proxy_value() = default;
+    explicit iteration_proxy_value(IteratorType it, std::size_t array_index_ = 0)
+    noexcept(std::is_nothrow_move_constructible<IteratorType>::value
+             && std::is_nothrow_default_constructible<string_type>::value)
+        : anchor(std::move(it))
+        , array_index(array_index_)
+    {}
+
+    iteration_proxy_value(iteration_proxy_value const&) = default;
+    iteration_proxy_value& operator=(iteration_proxy_value const&) = default;
+    // older GCCs are a bit fussy and require explicit noexcept specifiers on defaulted functions
+    iteration_proxy_value(iteration_proxy_value&&)
+    noexcept(std::is_nothrow_move_constructible<IteratorType>::value
+             && std::is_nothrow_move_constructible<string_type>::value) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor,cppcoreguidelines-noexcept-move-operations)
+    iteration_proxy_value& operator=(iteration_proxy_value&&)
+    noexcept(std::is_nothrow_move_assignable<IteratorType>::value
+             && std::is_nothrow_move_assignable<string_type>::value) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor,cppcoreguidelines-noexcept-move-operations)
+    ~iteration_proxy_value() = default;
+
+    /// dereference operator (needed for range-based for)
+    const iteration_proxy_value& operator*() const
+    {
+        return *this;
+    }
+
+    /// increment operator (needed for range-based for)
+    iteration_proxy_value& operator++()
+    {
+        ++anchor;
+        ++array_index;
+
+        return *this;
+    }
+
+    iteration_proxy_value operator++(int)& // NOLINT(cert-dcl21-cpp)
+    {
+        auto tmp = iteration_proxy_value(anchor, array_index);
+        ++anchor;
+        ++array_index;
+        return tmp;
+    }
+
+    /// equality operator (needed for InputIterator)
+    bool operator==(const iteration_proxy_value& o) const
+    {
+        return anchor == o.anchor;
+    }
+
+    /// inequality operator (needed for range-based for)
+    bool operator!=(const iteration_proxy_value& o) const
+    {
+        return anchor != o.anchor;
+    }
+
+    /// return key of the iterator
+    const string_type& key() const
+    {
+        JSON_ASSERT(anchor.m_object != nullptr);
+
+        switch (anchor.m_object->type())
+        {
+            // use integer array index as key
+            case value_t::array:
+            {
+                if (array_index != array_index_last)
+                {
+                    int_to_string( array_index_str, array_index );
+                    array_index_last = array_index;
+                }
+                return array_index_str;
+            }
+
+            // use key from the object
+            case value_t::object:
+                return anchor.key();
+
+            // use an empty key for all primitive types
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                return empty_str;
+        }
+    }
+
+    /// return value of the iterator
+    typename IteratorType::reference value() const
+    {
+        return anchor.value();
+    }
+};
+
+/// proxy class for the items() function
+template<typename IteratorType> class iteration_proxy
+{
+  private:
+    /// the container to iterate
+    typename IteratorType::pointer container = nullptr;
+
+  public:
+    explicit iteration_proxy() = default;
+
+    /// construct iteration proxy from a container
+    explicit iteration_proxy(typename IteratorType::reference cont) noexcept
+        : container(&cont) {}
+
+    iteration_proxy(iteration_proxy const&) = default;
+    iteration_proxy& operator=(iteration_proxy const&) = default;
+    iteration_proxy(iteration_proxy&&) noexcept = default;
+    iteration_proxy& operator=(iteration_proxy&&) noexcept = default;
+    ~iteration_proxy() = default;
+
+    /// return iterator begin (needed for range-based for)
+    iteration_proxy_value<IteratorType> begin() const noexcept
+    {
+        return iteration_proxy_value<IteratorType>(container->begin());
+    }
+
+    /// return iterator end (needed for range-based for)
+    iteration_proxy_value<IteratorType> end() const noexcept
+    {
+        return iteration_proxy_value<IteratorType>(container->end());
+    }
+};
+
+// Structured Bindings Support
+// For further reference see https://blog.tartanllama.xyz/structured-bindings/
+// And see https://github.com/nlohmann/json/pull/1391
+template<std::size_t N, typename IteratorType, enable_if_t<N == 0, int> = 0>
+auto get(const nlohmann::detail::iteration_proxy_value<IteratorType>& i) -> decltype(i.key())
+{
+    return i.key();
+}
+// Structured Bindings Support
+// For further reference see https://blog.tartanllama.xyz/structured-bindings/
+// And see https://github.com/nlohmann/json/pull/1391
+template<std::size_t N, typename IteratorType, enable_if_t<N == 1, int> = 0>
+auto get(const nlohmann::detail::iteration_proxy_value<IteratorType>& i) -> decltype(i.value())
+{
+    return i.value();
+}
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// The Addition to the STD Namespace is required to add
+// Structured Bindings Support to the iteration_proxy_value class
+// For further reference see https://blog.tartanllama.xyz/structured-bindings/
+// And see https://github.com/nlohmann/json/pull/1391
+namespace std
+{
+
+#if defined(__clang__)
+    // Fix: https://github.com/nlohmann/json/issues/1401
+    #pragma clang diagnostic push
+    #pragma clang diagnostic ignored "-Wmismatched-tags"
+#endif
+template<typename IteratorType>
+class tuple_size<::nlohmann::detail::iteration_proxy_value<IteratorType>> // NOLINT(cert-dcl58-cpp)
+            : public std::integral_constant<std::size_t, 2> {};
+
+template<std::size_t N, typename IteratorType>
+class tuple_element<N, ::nlohmann::detail::iteration_proxy_value<IteratorType >> // NOLINT(cert-dcl58-cpp)
+{
+  public:
+    using type = decltype(
+                     get<N>(std::declval <
+                            ::nlohmann::detail::iteration_proxy_value<IteratorType >> ()));
+};
+#if defined(__clang__)
+    #pragma clang diagnostic pop
+#endif
+
+}  // namespace std
+
+#if JSON_HAS_RANGES
+    template <typename IteratorType>
+    inline constexpr bool ::std::ranges::enable_borrowed_range<::nlohmann::detail::iteration_proxy<IteratorType>> = true;
+#endif
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/std_fs.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+//////////////////
+// constructors //
+//////////////////
+
+/*
+ * Note all external_constructor<>::construct functions need to call
+ * j.m_data.m_value.destroy(j.m_data.m_type) to avoid a memory leak in case j contains an
+ * allocated value (e.g., a string). See bug issue
+ * https://github.com/nlohmann/json/issues/2865 for more information.
+ */
+
+template<value_t> struct external_constructor;
+
+template<>
+struct external_constructor<value_t::boolean>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::boolean_t b) noexcept
+    {
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::boolean;
+        j.m_data.m_value = b;
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::string>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, const typename BasicJsonType::string_t& s)
+    {
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::string;
+        j.m_data.m_value = s;
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::string_t&& s)
+    {
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::string;
+        j.m_data.m_value = std::move(s);
+        j.assert_invariant();
+    }
+
+    template < typename BasicJsonType, typename CompatibleStringType,
+               enable_if_t < !std::is_same<CompatibleStringType, typename BasicJsonType::string_t>::value,
+                             int > = 0 >
+    static void construct(BasicJsonType& j, const CompatibleStringType& str)
+    {
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::string;
+        j.m_data.m_value.string = j.template create<typename BasicJsonType::string_t>(str);
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::binary>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, const typename BasicJsonType::binary_t& b)
+    {
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::binary;
+        j.m_data.m_value = typename BasicJsonType::binary_t(b);
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::binary_t&& b)
+    {
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::binary;
+        j.m_data.m_value = typename BasicJsonType::binary_t(std::move(b));
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::number_float>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::number_float_t val) noexcept
+    {
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::number_float;
+        j.m_data.m_value = val;
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::number_unsigned>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::number_unsigned_t val) noexcept
+    {
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::number_unsigned;
+        j.m_data.m_value = val;
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::number_integer>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::number_integer_t val) noexcept
+    {
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::number_integer;
+        j.m_data.m_value = val;
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::array>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, const typename BasicJsonType::array_t& arr)
+    {
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::array;
+        j.m_data.m_value = arr;
+        j.set_parents();
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::array_t&& arr)
+    {
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::array;
+        j.m_data.m_value = std::move(arr);
+        j.set_parents();
+        j.assert_invariant();
+    }
+
+    template < typename BasicJsonType, typename CompatibleArrayType,
+               enable_if_t < !std::is_same<CompatibleArrayType, typename BasicJsonType::array_t>::value,
+                             int > = 0 >
+    static void construct(BasicJsonType& j, const CompatibleArrayType& arr)
+    {
+        using std::begin;
+        using std::end;
+
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::array;
+        j.m_data.m_value.array = j.template create<typename BasicJsonType::array_t>(begin(arr), end(arr));
+        j.set_parents();
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, const std::vector<bool>& arr)
+    {
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::array;
+        j.m_data.m_value = value_t::array;
+        j.m_data.m_value.array->reserve(arr.size());
+        for (const bool x : arr)
+        {
+            j.m_data.m_value.array->push_back(x);
+            j.set_parent(j.m_data.m_value.array->back());
+        }
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType, typename T,
+             enable_if_t<std::is_convertible<T, BasicJsonType>::value, int> = 0>
+    static void construct(BasicJsonType& j, const std::valarray<T>& arr)
+    {
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::array;
+        j.m_data.m_value = value_t::array;
+        j.m_data.m_value.array->resize(arr.size());
+        if (arr.size() > 0)
+        {
+            std::copy(std::begin(arr), std::end(arr), j.m_data.m_value.array->begin());
+        }
+        j.set_parents();
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::object>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, const typename BasicJsonType::object_t& obj)
+    {
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::object;
+        j.m_data.m_value = obj;
+        j.set_parents();
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::object_t&& obj)
+    {
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::object;
+        j.m_data.m_value = std::move(obj);
+        j.set_parents();
+        j.assert_invariant();
+    }
+
+    template < typename BasicJsonType, typename CompatibleObjectType,
+               enable_if_t < !std::is_same<CompatibleObjectType, typename BasicJsonType::object_t>::value, int > = 0 >
+    static void construct(BasicJsonType& j, const CompatibleObjectType& obj)
+    {
+        using std::begin;
+        using std::end;
+
+        j.m_data.m_value.destroy(j.m_data.m_type);
+        j.m_data.m_type = value_t::object;
+        j.m_data.m_value.object = j.template create<typename BasicJsonType::object_t>(begin(obj), end(obj));
+        j.set_parents();
+        j.assert_invariant();
+    }
+};
+
+/////////////
+// to_json //
+/////////////
+
+template<typename BasicJsonType, typename T,
+         enable_if_t<std::is_same<T, typename BasicJsonType::boolean_t>::value, int> = 0>
+inline void to_json(BasicJsonType& j, T b) noexcept
+{
+    external_constructor<value_t::boolean>::construct(j, b);
+}
+
+template < typename BasicJsonType, typename BoolRef,
+           enable_if_t <
+               ((std::is_same<std::vector<bool>::reference, BoolRef>::value
+                 && !std::is_same <std::vector<bool>::reference, typename BasicJsonType::boolean_t&>::value)
+                || (std::is_same<std::vector<bool>::const_reference, BoolRef>::value
+                    && !std::is_same <detail::uncvref_t<std::vector<bool>::const_reference>,
+                                      typename BasicJsonType::boolean_t >::value))
+               && std::is_convertible<const BoolRef&, typename BasicJsonType::boolean_t>::value, int > = 0 >
+inline void to_json(BasicJsonType& j, const BoolRef& b) noexcept
+{
+    external_constructor<value_t::boolean>::construct(j, static_cast<typename BasicJsonType::boolean_t>(b));
+}
+
+template<typename BasicJsonType, typename CompatibleString,
+         enable_if_t<std::is_constructible<typename BasicJsonType::string_t, CompatibleString>::value, int> = 0>
+inline void to_json(BasicJsonType& j, const CompatibleString& s)
+{
+    external_constructor<value_t::string>::construct(j, s);
+}
+
+template<typename BasicJsonType>
+inline void to_json(BasicJsonType& j, typename BasicJsonType::string_t&& s)
+{
+    external_constructor<value_t::string>::construct(j, std::move(s));
+}
+
+template<typename BasicJsonType, typename FloatType,
+         enable_if_t<std::is_floating_point<FloatType>::value, int> = 0>
+inline void to_json(BasicJsonType& j, FloatType val) noexcept
+{
+    external_constructor<value_t::number_float>::construct(j, static_cast<typename BasicJsonType::number_float_t>(val));
+}
+
+template<typename BasicJsonType, typename CompatibleNumberUnsignedType,
+         enable_if_t<is_compatible_integer_type<typename BasicJsonType::number_unsigned_t, CompatibleNumberUnsignedType>::value, int> = 0>
+inline void to_json(BasicJsonType& j, CompatibleNumberUnsignedType val) noexcept
+{
+    external_constructor<value_t::number_unsigned>::construct(j, static_cast<typename BasicJsonType::number_unsigned_t>(val));
+}
+
+template<typename BasicJsonType, typename CompatibleNumberIntegerType,
+         enable_if_t<is_compatible_integer_type<typename BasicJsonType::number_integer_t, CompatibleNumberIntegerType>::value, int> = 0>
+inline void to_json(BasicJsonType& j, CompatibleNumberIntegerType val) noexcept
+{
+    external_constructor<value_t::number_integer>::construct(j, static_cast<typename BasicJsonType::number_integer_t>(val));
+}
+
+#if !JSON_DISABLE_ENUM_SERIALIZATION
+template<typename BasicJsonType, typename EnumType,
+         enable_if_t<std::is_enum<EnumType>::value, int> = 0>
+inline void to_json(BasicJsonType& j, EnumType e) noexcept
+{
+    using underlying_type = typename std::underlying_type<EnumType>::type;
+    static constexpr value_t integral_value_t = std::is_unsigned<underlying_type>::value ? value_t::number_unsigned : value_t::number_integer;
+    external_constructor<integral_value_t>::construct(j, static_cast<underlying_type>(e));
+}
+#endif  // JSON_DISABLE_ENUM_SERIALIZATION
+
+template<typename BasicJsonType>
+inline void to_json(BasicJsonType& j, const std::vector<bool>& e)
+{
+    external_constructor<value_t::array>::construct(j, e);
+}
+
+template < typename BasicJsonType, typename CompatibleArrayType,
+           enable_if_t < is_compatible_array_type<BasicJsonType,
+                         CompatibleArrayType>::value&&
+                         !is_compatible_object_type<BasicJsonType, CompatibleArrayType>::value&&
+                         !is_compatible_string_type<BasicJsonType, CompatibleArrayType>::value&&
+                         !std::is_same<typename BasicJsonType::binary_t, CompatibleArrayType>::value&&
+                         !is_basic_json<CompatibleArrayType>::value,
+                         int > = 0 >
+inline void to_json(BasicJsonType& j, const CompatibleArrayType& arr)
+{
+    external_constructor<value_t::array>::construct(j, arr);
+}
+
+template<typename BasicJsonType>
+inline void to_json(BasicJsonType& j, const typename BasicJsonType::binary_t& bin)
+{
+    external_constructor<value_t::binary>::construct(j, bin);
+}
+
+template<typename BasicJsonType, typename T,
+         enable_if_t<std::is_convertible<T, BasicJsonType>::value, int> = 0>
+inline void to_json(BasicJsonType& j, const std::valarray<T>& arr)
+{
+    external_constructor<value_t::array>::construct(j, std::move(arr));
+}
+
+template<typename BasicJsonType>
+inline void to_json(BasicJsonType& j, typename BasicJsonType::array_t&& arr)
+{
+    external_constructor<value_t::array>::construct(j, std::move(arr));
+}
+
+template < typename BasicJsonType, typename CompatibleObjectType,
+           enable_if_t < is_compatible_object_type<BasicJsonType, CompatibleObjectType>::value&& !is_basic_json<CompatibleObjectType>::value, int > = 0 >
+inline void to_json(BasicJsonType& j, const CompatibleObjectType& obj)
+{
+    external_constructor<value_t::object>::construct(j, obj);
+}
+
+template<typename BasicJsonType>
+inline void to_json(BasicJsonType& j, typename BasicJsonType::object_t&& obj)
+{
+    external_constructor<value_t::object>::construct(j, std::move(obj));
+}
+
+template <
+    typename BasicJsonType, typename T, std::size_t N,
+    enable_if_t < !std::is_constructible<typename BasicJsonType::string_t,
+                  const T(&)[N]>::value, // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+                  int > = 0 >
+inline void to_json(BasicJsonType& j, const T(&arr)[N]) // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+{
+    external_constructor<value_t::array>::construct(j, arr);
+}
+
+template < typename BasicJsonType, typename T1, typename T2, enable_if_t < std::is_constructible<BasicJsonType, T1>::value&& std::is_constructible<BasicJsonType, T2>::value, int > = 0 >
+inline void to_json(BasicJsonType& j, const std::pair<T1, T2>& p)
+{
+    j = { p.first, p.second };
+}
+
+// for https://github.com/nlohmann/json/pull/1134
+template<typename BasicJsonType, typename T,
+         enable_if_t<std::is_same<T, iteration_proxy_value<typename BasicJsonType::iterator>>::value, int> = 0>
+inline void to_json(BasicJsonType& j, const T& b)
+{
+    j = { {b.key(), b.value()} };
+}
+
+template<typename BasicJsonType, typename Tuple, std::size_t... Idx>
+inline void to_json_tuple_impl(BasicJsonType& j, const Tuple& t, index_sequence<Idx...> /*unused*/)
+{
+    j = { std::get<Idx>(t)... };
+}
+
+template<typename BasicJsonType, typename T, enable_if_t<is_constructible_tuple<BasicJsonType, T>::value, int > = 0>
+inline void to_json(BasicJsonType& j, const T& t)
+{
+    to_json_tuple_impl(j, t, make_index_sequence<std::tuple_size<T>::value> {});
+}
+
+#if JSON_HAS_FILESYSTEM || JSON_HAS_EXPERIMENTAL_FILESYSTEM
+template<typename BasicJsonType>
+inline void to_json(BasicJsonType& j, const std_fs::path& p)
+{
+    j = p.string();
+}
+#endif
+
+struct to_json_fn
+{
+    template<typename BasicJsonType, typename T>
+    auto operator()(BasicJsonType& j, T&& val) const noexcept(noexcept(to_json(j, std::forward<T>(val))))
+    -> decltype(to_json(j, std::forward<T>(val)), void())
+    {
+        return to_json(j, std::forward<T>(val));
+    }
+};
+}  // namespace detail
+
+#ifndef JSON_HAS_CPP_17
+/// namespace to hold default `to_json` function
+/// to see why this is required:
+/// http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4381.html
+namespace // NOLINT(cert-dcl59-cpp,fuchsia-header-anon-namespaces,google-build-namespaces)
+{
+#endif
+JSON_INLINE_VARIABLE constexpr const auto& to_json = // NOLINT(misc-definitions-in-headers)
+    detail::static_const<detail::to_json_fn>::value;
+#ifndef JSON_HAS_CPP_17
+}  // namespace
+#endif
+
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/meta/identity_tag.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+/// @sa https://json.nlohmann.me/api/adl_serializer/
+template<typename ValueType, typename>
+struct adl_serializer
+{
+    /// @brief convert a JSON value to any value type
+    /// @sa https://json.nlohmann.me/api/adl_serializer/from_json/
+    template<typename BasicJsonType, typename TargetType = ValueType>
+    static auto from_json(BasicJsonType && j, TargetType& val) noexcept(
+        noexcept(::nlohmann::from_json(std::forward<BasicJsonType>(j), val)))
+    -> decltype(::nlohmann::from_json(std::forward<BasicJsonType>(j), val), void())
+    {
+        ::nlohmann::from_json(std::forward<BasicJsonType>(j), val);
+    }
+
+    /// @brief convert a JSON value to any value type
+    /// @sa https://json.nlohmann.me/api/adl_serializer/from_json/
+    template<typename BasicJsonType, typename TargetType = ValueType>
+    static auto from_json(BasicJsonType && j) noexcept(
+    noexcept(::nlohmann::from_json(std::forward<BasicJsonType>(j), detail::identity_tag<TargetType> {})))
+    -> decltype(::nlohmann::from_json(std::forward<BasicJsonType>(j), detail::identity_tag<TargetType> {}))
+    {
+        return ::nlohmann::from_json(std::forward<BasicJsonType>(j), detail::identity_tag<TargetType> {});
+    }
+
+    /// @brief convert any value type to a JSON value
+    /// @sa https://json.nlohmann.me/api/adl_serializer/to_json/
+    template<typename BasicJsonType, typename TargetType = ValueType>
+    static auto to_json(BasicJsonType& j, TargetType && val) noexcept(
+        noexcept(::nlohmann::to_json(j, std::forward<TargetType>(val))))
+    -> decltype(::nlohmann::to_json(j, std::forward<TargetType>(val)), void())
+    {
+        ::nlohmann::to_json(j, std::forward<TargetType>(val));
+    }
+};
+
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/byte_container_with_subtype.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstdint> // uint8_t, uint64_t
+#include <tuple> // tie
+#include <utility> // move
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+/// @brief an internal type for a backed binary type
+/// @sa https://json.nlohmann.me/api/byte_container_with_subtype/
+template<typename BinaryType>
+class byte_container_with_subtype : public BinaryType
+{
+  public:
+    using container_type = BinaryType;
+    using subtype_type = std::uint64_t;
+
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
+    byte_container_with_subtype() noexcept(noexcept(container_type()))
+        : container_type()
+    {}
+
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
+    byte_container_with_subtype(const container_type& b) noexcept(noexcept(container_type(b)))
+        : container_type(b)
+    {}
+
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
+    byte_container_with_subtype(container_type&& b) noexcept(noexcept(container_type(std::move(b))))
+        : container_type(std::move(b))
+    {}
+
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
+    byte_container_with_subtype(const container_type& b, subtype_type subtype_) noexcept(noexcept(container_type(b)))
+        : container_type(b)
+        , m_subtype(subtype_)
+        , m_has_subtype(true)
+    {}
+
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
+    byte_container_with_subtype(container_type&& b, subtype_type subtype_) noexcept(noexcept(container_type(std::move(b))))
+        : container_type(std::move(b))
+        , m_subtype(subtype_)
+        , m_has_subtype(true)
+    {}
+
+    bool operator==(const byte_container_with_subtype& rhs) const
+    {
+        return std::tie(static_cast<const BinaryType&>(*this), m_subtype, m_has_subtype) ==
+               std::tie(static_cast<const BinaryType&>(rhs), rhs.m_subtype, rhs.m_has_subtype);
+    }
+
+    bool operator!=(const byte_container_with_subtype& rhs) const
+    {
+        return !(rhs == *this);
+    }
+
+    /// @brief sets the binary subtype
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/set_subtype/
+    void set_subtype(subtype_type subtype_) noexcept
+    {
+        m_subtype = subtype_;
+        m_has_subtype = true;
+    }
+
+    /// @brief return the binary subtype
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/subtype/
+    constexpr subtype_type subtype() const noexcept
+    {
+        return m_has_subtype ? m_subtype : static_cast<subtype_type>(-1);
+    }
+
+    /// @brief return whether the value has a subtype
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/has_subtype/
+    constexpr bool has_subtype() const noexcept
+    {
+        return m_has_subtype;
+    }
+
+    /// @brief clears the binary subtype
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/clear_subtype/
+    void clear_subtype() noexcept
+    {
+        m_subtype = 0;
+        m_has_subtype = false;
+    }
+
+  private:
+    subtype_type m_subtype = 0;
+    bool m_has_subtype = false;
+};
+
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/conversions/from_json.hpp>
+
+// #include <nlohmann/detail/conversions/to_json.hpp>
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/hash.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstdint> // uint8_t
+#include <cstddef> // size_t
+#include <functional> // hash
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+// boost::hash_combine
+inline std::size_t combine(std::size_t seed, std::size_t h) noexcept
+{
+    seed ^= h + 0x9e3779b9 + (seed << 6U) + (seed >> 2U);
+    return seed;
+}
+
+/*!
+@brief hash a JSON value
+
+The hash function tries to rely on std::hash where possible. Furthermore, the
+type of the JSON value is taken into account to have different hash values for
+null, 0, 0U, and false, etc.
+
+@tparam BasicJsonType basic_json specialization
+@param j JSON value to hash
+@return hash value of j
+*/
+template<typename BasicJsonType>
+std::size_t hash(const BasicJsonType& j)
+{
+    using string_t = typename BasicJsonType::string_t;
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+
+    const auto type = static_cast<std::size_t>(j.type());
+    switch (j.type())
+    {
+        case BasicJsonType::value_t::null:
+        case BasicJsonType::value_t::discarded:
+        {
+            return combine(type, 0);
+        }
+
+        case BasicJsonType::value_t::object:
+        {
+            auto seed = combine(type, j.size());
+            for (const auto& element : j.items())
+            {
+                const auto h = std::hash<string_t> {}(element.key());
+                seed = combine(seed, h);
+                seed = combine(seed, hash(element.value()));
+            }
+            return seed;
+        }
+
+        case BasicJsonType::value_t::array:
+        {
+            auto seed = combine(type, j.size());
+            for (const auto& element : j)
+            {
+                seed = combine(seed, hash(element));
+            }
+            return seed;
+        }
+
+        case BasicJsonType::value_t::string:
+        {
+            const auto h = std::hash<string_t> {}(j.template get_ref<const string_t&>());
+            return combine(type, h);
+        }
+
+        case BasicJsonType::value_t::boolean:
+        {
+            const auto h = std::hash<bool> {}(j.template get<bool>());
+            return combine(type, h);
+        }
+
+        case BasicJsonType::value_t::number_integer:
+        {
+            const auto h = std::hash<number_integer_t> {}(j.template get<number_integer_t>());
+            return combine(type, h);
+        }
+
+        case BasicJsonType::value_t::number_unsigned:
+        {
+            const auto h = std::hash<number_unsigned_t> {}(j.template get<number_unsigned_t>());
+            return combine(type, h);
+        }
+
+        case BasicJsonType::value_t::number_float:
+        {
+            const auto h = std::hash<number_float_t> {}(j.template get<number_float_t>());
+            return combine(type, h);
+        }
+
+        case BasicJsonType::value_t::binary:
+        {
+            auto seed = combine(type, j.get_binary().size());
+            const auto h = std::hash<bool> {}(j.get_binary().has_subtype());
+            seed = combine(seed, h);
+            seed = combine(seed, static_cast<std::size_t>(j.get_binary().subtype()));
+            for (const auto byte : j.get_binary())
+            {
+                seed = combine(seed, std::hash<std::uint8_t> {}(byte));
+            }
+            return seed;
+        }
+
+        default:                   // LCOV_EXCL_LINE
+            JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+            return 0;              // LCOV_EXCL_LINE
+    }
+}
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/input/binary_reader.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <algorithm> // generate_n
+#include <array> // array
+#include <cmath> // ldexp
+#include <cstddef> // size_t
+#include <cstdint> // uint8_t, uint16_t, uint32_t, uint64_t
+#include <cstdio> // snprintf
+#include <cstring> // memcpy
+#include <iterator> // back_inserter
+#include <limits> // numeric_limits
+#include <string> // char_traits, string
+#include <utility> // make_pair, move
+#include <vector> // vector
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/input/input_adapters.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <array> // array
+#include <cstddef> // size_t
+#include <cstring> // strlen
+#include <iterator> // begin, end, iterator_traits, random_access_iterator_tag, distance, next
+#include <memory> // shared_ptr, make_shared, addressof
+#include <numeric> // accumulate
+#include <string> // string, char_traits
+#include <type_traits> // enable_if, is_base_of, is_pointer, is_integral, remove_pointer
+#include <utility> // pair, declval
+
+#ifndef JSON_NO_IO
+    #include <cstdio>   // FILE *
+    #include <istream>  // istream
+#endif                  // JSON_NO_IO
+
+// #include <nlohmann/detail/iterators/iterator_traits.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/// the supported input formats
+enum class input_format_t { json, cbor, msgpack, ubjson, bson, bjdata };
+
+////////////////////
+// input adapters //
+////////////////////
+
+#ifndef JSON_NO_IO
+/*!
+Input adapter for stdio file access. This adapter read only 1 byte and do not use any
+ buffer. This adapter is a very low level adapter.
+*/
+class file_input_adapter
+{
+  public:
+    using char_type = char;
+
+    JSON_HEDLEY_NON_NULL(2)
+    explicit file_input_adapter(std::FILE* f) noexcept
+        : m_file(f)
+    {
+        JSON_ASSERT(m_file != nullptr);
+    }
+
+    // make class move-only
+    file_input_adapter(const file_input_adapter&) = delete;
+    file_input_adapter(file_input_adapter&&) noexcept = default;
+    file_input_adapter& operator=(const file_input_adapter&) = delete;
+    file_input_adapter& operator=(file_input_adapter&&) = delete;
+    ~file_input_adapter() = default;
+
+    std::char_traits<char>::int_type get_character() noexcept
+    {
+        return std::fgetc(m_file);
+    }
+
+  private:
+    /// the file pointer to read from
+    std::FILE* m_file;
+};
+
+/*!
+Input adapter for a (caching) istream. Ignores a UFT Byte Order Mark at
+beginning of input. Does not support changing the underlying std::streambuf
+in mid-input. Maintains underlying std::istream and std::streambuf to support
+subsequent use of standard std::istream operations to process any input
+characters following those used in parsing the JSON input.  Clears the
+std::istream flags; any input errors (e.g., EOF) will be detected by the first
+subsequent call for input from the std::istream.
+*/
+class input_stream_adapter
+{
+  public:
+    using char_type = char;
+
+    ~input_stream_adapter()
+    {
+        // clear stream flags; we use underlying streambuf I/O, do not
+        // maintain ifstream flags, except eof
+        if (is != nullptr)
+        {
+            is->clear(is->rdstate() & std::ios::eofbit);
+        }
+    }
+
+    explicit input_stream_adapter(std::istream& i)
+        : is(&i), sb(i.rdbuf())
+    {}
+
+    // delete because of pointer members
+    input_stream_adapter(const input_stream_adapter&) = delete;
+    input_stream_adapter& operator=(input_stream_adapter&) = delete;
+    input_stream_adapter& operator=(input_stream_adapter&&) = delete;
+
+    input_stream_adapter(input_stream_adapter&& rhs) noexcept
+        : is(rhs.is), sb(rhs.sb)
+    {
+        rhs.is = nullptr;
+        rhs.sb = nullptr;
+    }
+
+    // std::istream/std::streambuf use std::char_traits<char>::to_int_type, to
+    // ensure that std::char_traits<char>::eof() and the character 0xFF do not
+    // end up as the same value, e.g. 0xFFFFFFFF.
+    std::char_traits<char>::int_type get_character()
+    {
+        auto res = sb->sbumpc();
+        // set eof manually, as we don't use the istream interface.
+        if (JSON_HEDLEY_UNLIKELY(res == std::char_traits<char>::eof()))
+        {
+            is->clear(is->rdstate() | std::ios::eofbit);
+        }
+        return res;
+    }
+
+  private:
+    /// the associated input stream
+    std::istream* is = nullptr;
+    std::streambuf* sb = nullptr;
+};
+#endif  // JSON_NO_IO
+
+// General-purpose iterator-based adapter. It might not be as fast as
+// theoretically possible for some containers, but it is extremely versatile.
+template<typename IteratorType>
+class iterator_input_adapter
+{
+  public:
+    using char_type = typename std::iterator_traits<IteratorType>::value_type;
+
+    iterator_input_adapter(IteratorType first, IteratorType last)
+        : current(std::move(first)), end(std::move(last))
+    {}
+
+    typename char_traits<char_type>::int_type get_character()
+    {
+        if (JSON_HEDLEY_LIKELY(current != end))
+        {
+            auto result = char_traits<char_type>::to_int_type(*current);
+            std::advance(current, 1);
+            return result;
+        }
+
+        return char_traits<char_type>::eof();
+    }
+
+  private:
+    IteratorType current;
+    IteratorType end;
+
+    template<typename BaseInputAdapter, size_t T>
+    friend struct wide_string_input_helper;
+
+    bool empty() const
+    {
+        return current == end;
+    }
+};
+
+template<typename BaseInputAdapter, size_t T>
+struct wide_string_input_helper;
+
+template<typename BaseInputAdapter>
+struct wide_string_input_helper<BaseInputAdapter, 4>
+{
+    // UTF-32
+    static void fill_buffer(BaseInputAdapter& input,
+                            std::array<std::char_traits<char>::int_type, 4>& utf8_bytes,
+                            size_t& utf8_bytes_index,
+                            size_t& utf8_bytes_filled)
+    {
+        utf8_bytes_index = 0;
+
+        if (JSON_HEDLEY_UNLIKELY(input.empty()))
+        {
+            utf8_bytes[0] = std::char_traits<char>::eof();
+            utf8_bytes_filled = 1;
+        }
+        else
+        {
+            // get the current character
+            const auto wc = input.get_character();
+
+            // UTF-32 to UTF-8 encoding
+            if (wc < 0x80)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
+                utf8_bytes_filled = 1;
+            }
+            else if (wc <= 0x7FF)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xC0u | ((static_cast<unsigned int>(wc) >> 6u) & 0x1Fu));
+                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
+                utf8_bytes_filled = 2;
+            }
+            else if (wc <= 0xFFFF)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xE0u | ((static_cast<unsigned int>(wc) >> 12u) & 0x0Fu));
+                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 6u) & 0x3Fu));
+                utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
+                utf8_bytes_filled = 3;
+            }
+            else if (wc <= 0x10FFFF)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xF0u | ((static_cast<unsigned int>(wc) >> 18u) & 0x07u));
+                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 12u) & 0x3Fu));
+                utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 6u) & 0x3Fu));
+                utf8_bytes[3] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
+                utf8_bytes_filled = 4;
+            }
+            else
+            {
+                // unknown character
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
+                utf8_bytes_filled = 1;
+            }
+        }
+    }
+};
+
+template<typename BaseInputAdapter>
+struct wide_string_input_helper<BaseInputAdapter, 2>
+{
+    // UTF-16
+    static void fill_buffer(BaseInputAdapter& input,
+                            std::array<std::char_traits<char>::int_type, 4>& utf8_bytes,
+                            size_t& utf8_bytes_index,
+                            size_t& utf8_bytes_filled)
+    {
+        utf8_bytes_index = 0;
+
+        if (JSON_HEDLEY_UNLIKELY(input.empty()))
+        {
+            utf8_bytes[0] = std::char_traits<char>::eof();
+            utf8_bytes_filled = 1;
+        }
+        else
+        {
+            // get the current character
+            const auto wc = input.get_character();
+
+            // UTF-16 to UTF-8 encoding
+            if (wc < 0x80)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
+                utf8_bytes_filled = 1;
+            }
+            else if (wc <= 0x7FF)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xC0u | ((static_cast<unsigned int>(wc) >> 6u)));
+                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
+                utf8_bytes_filled = 2;
+            }
+            else if (0xD800 > wc || wc >= 0xE000)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xE0u | ((static_cast<unsigned int>(wc) >> 12u)));
+                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 6u) & 0x3Fu));
+                utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
+                utf8_bytes_filled = 3;
+            }
+            else
+            {
+                if (JSON_HEDLEY_UNLIKELY(!input.empty()))
+                {
+                    const auto wc2 = static_cast<unsigned int>(input.get_character());
+                    const auto charcode = 0x10000u + (((static_cast<unsigned int>(wc) & 0x3FFu) << 10u) | (wc2 & 0x3FFu));
+                    utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xF0u | (charcode >> 18u));
+                    utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((charcode >> 12u) & 0x3Fu));
+                    utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | ((charcode >> 6u) & 0x3Fu));
+                    utf8_bytes[3] = static_cast<std::char_traits<char>::int_type>(0x80u | (charcode & 0x3Fu));
+                    utf8_bytes_filled = 4;
+                }
+                else
+                {
+                    utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
+                    utf8_bytes_filled = 1;
+                }
+            }
+        }
+    }
+};
+
+// Wraps another input adapter to convert wide character types into individual bytes.
+template<typename BaseInputAdapter, typename WideCharType>
+class wide_string_input_adapter
+{
+  public:
+    using char_type = char;
+
+    wide_string_input_adapter(BaseInputAdapter base)
+        : base_adapter(base) {}
+
+    typename std::char_traits<char>::int_type get_character() noexcept
+    {
+        // check if buffer needs to be filled
+        if (utf8_bytes_index == utf8_bytes_filled)
+        {
+            fill_buffer<sizeof(WideCharType)>();
+
+            JSON_ASSERT(utf8_bytes_filled > 0);
+            JSON_ASSERT(utf8_bytes_index == 0);
+        }
+
+        // use buffer
+        JSON_ASSERT(utf8_bytes_filled > 0);
+        JSON_ASSERT(utf8_bytes_index < utf8_bytes_filled);
+        return utf8_bytes[utf8_bytes_index++];
+    }
+
+  private:
+    BaseInputAdapter base_adapter;
+
+    template<size_t T>
+    void fill_buffer()
+    {
+        wide_string_input_helper<BaseInputAdapter, T>::fill_buffer(base_adapter, utf8_bytes, utf8_bytes_index, utf8_bytes_filled);
+    }
+
+    /// a buffer for UTF-8 bytes
+    std::array<std::char_traits<char>::int_type, 4> utf8_bytes = {{0, 0, 0, 0}};
+
+    /// index to the utf8_codes array for the next valid byte
+    std::size_t utf8_bytes_index = 0;
+    /// number of valid bytes in the utf8_codes array
+    std::size_t utf8_bytes_filled = 0;
+};
+
+template<typename IteratorType, typename Enable = void>
+struct iterator_input_adapter_factory
+{
+    using iterator_type = IteratorType;
+    using char_type = typename std::iterator_traits<iterator_type>::value_type;
+    using adapter_type = iterator_input_adapter<iterator_type>;
+
+    static adapter_type create(IteratorType first, IteratorType last)
+    {
+        return adapter_type(std::move(first), std::move(last));
+    }
+};
+
+template<typename T>
+struct is_iterator_of_multibyte
+{
+    using value_type = typename std::iterator_traits<T>::value_type;
+    enum
+    {
+        value = sizeof(value_type) > 1
+    };
+};
+
+template<typename IteratorType>
+struct iterator_input_adapter_factory<IteratorType, enable_if_t<is_iterator_of_multibyte<IteratorType>::value>>
+{
+    using iterator_type = IteratorType;
+    using char_type = typename std::iterator_traits<iterator_type>::value_type;
+    using base_adapter_type = iterator_input_adapter<iterator_type>;
+    using adapter_type = wide_string_input_adapter<base_adapter_type, char_type>;
+
+    static adapter_type create(IteratorType first, IteratorType last)
+    {
+        return adapter_type(base_adapter_type(std::move(first), std::move(last)));
+    }
+};
+
+// General purpose iterator-based input
+template<typename IteratorType>
+typename iterator_input_adapter_factory<IteratorType>::adapter_type input_adapter(IteratorType first, IteratorType last)
+{
+    using factory_type = iterator_input_adapter_factory<IteratorType>;
+    return factory_type::create(first, last);
+}
+
+// Convenience shorthand from container to iterator
+// Enables ADL on begin(container) and end(container)
+// Encloses the using declarations in namespace for not to leak them to outside scope
+
+namespace container_input_adapter_factory_impl
+{
+
+using std::begin;
+using std::end;
+
+template<typename ContainerType, typename Enable = void>
+struct container_input_adapter_factory {};
+
+template<typename ContainerType>
+struct container_input_adapter_factory< ContainerType,
+       void_t<decltype(begin(std::declval<ContainerType>()), end(std::declval<ContainerType>()))>>
+       {
+           using adapter_type = decltype(input_adapter(begin(std::declval<ContainerType>()), end(std::declval<ContainerType>())));
+
+           static adapter_type create(const ContainerType& container)
+{
+    return input_adapter(begin(container), end(container));
+}
+       };
+
+}  // namespace container_input_adapter_factory_impl
+
+template<typename ContainerType>
+typename container_input_adapter_factory_impl::container_input_adapter_factory<ContainerType>::adapter_type input_adapter(const ContainerType& container)
+{
+    return container_input_adapter_factory_impl::container_input_adapter_factory<ContainerType>::create(container);
+}
+
+#ifndef JSON_NO_IO
+// Special cases with fast paths
+inline file_input_adapter input_adapter(std::FILE* file)
+{
+    return file_input_adapter(file);
+}
+
+inline input_stream_adapter input_adapter(std::istream& stream)
+{
+    return input_stream_adapter(stream);
+}
+
+inline input_stream_adapter input_adapter(std::istream&& stream)
+{
+    return input_stream_adapter(stream);
+}
+#endif  // JSON_NO_IO
+
+using contiguous_bytes_input_adapter = decltype(input_adapter(std::declval<const char*>(), std::declval<const char*>()));
+
+// Null-delimited strings, and the like.
+template < typename CharT,
+           typename std::enable_if <
+               std::is_pointer<CharT>::value&&
+               !std::is_array<CharT>::value&&
+               std::is_integral<typename std::remove_pointer<CharT>::type>::value&&
+               sizeof(typename std::remove_pointer<CharT>::type) == 1,
+               int >::type = 0 >
+contiguous_bytes_input_adapter input_adapter(CharT b)
+{
+    auto length = std::strlen(reinterpret_cast<const char*>(b));
+    const auto* ptr = reinterpret_cast<const char*>(b);
+    return input_adapter(ptr, ptr + length);
+}
+
+template<typename T, std::size_t N>
+auto input_adapter(T (&array)[N]) -> decltype(input_adapter(array, array + N)) // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+{
+    return input_adapter(array, array + N);
+}
+
+// This class only handles inputs of input_buffer_adapter type.
+// It's required so that expressions like {ptr, len} can be implicitly cast
+// to the correct adapter.
+class span_input_adapter
+{
+  public:
+    template < typename CharT,
+               typename std::enable_if <
+                   std::is_pointer<CharT>::value&&
+                   std::is_integral<typename std::remove_pointer<CharT>::type>::value&&
+                   sizeof(typename std::remove_pointer<CharT>::type) == 1,
+                   int >::type = 0 >
+    span_input_adapter(CharT b, std::size_t l)
+        : ia(reinterpret_cast<const char*>(b), reinterpret_cast<const char*>(b) + l) {}
+
+    template<class IteratorType,
+             typename std::enable_if<
+                 std::is_same<typename iterator_traits<IteratorType>::iterator_category, std::random_access_iterator_tag>::value,
+                 int>::type = 0>
+    span_input_adapter(IteratorType first, IteratorType last)
+        : ia(input_adapter(first, last)) {}
+
+    contiguous_bytes_input_adapter&& get()
+    {
+        return std::move(ia); // NOLINT(hicpp-move-const-arg,performance-move-const-arg)
+    }
+
+  private:
+    contiguous_bytes_input_adapter ia;
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/input/json_sax.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstddef>
+#include <string> // string
+#include <utility> // move
+#include <vector> // vector
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+/*!
+@brief SAX interface
+
+This class describes the SAX interface used by @ref nlohmann::json::sax_parse.
+Each function is called in different situations while the input is parsed. The
+boolean return value informs the parser whether to continue processing the
+input.
+*/
+template<typename BasicJsonType>
+struct json_sax
+{
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+
+    /*!
+    @brief a null value was read
+    @return whether parsing should proceed
+    */
+    virtual bool null() = 0;
+
+    /*!
+    @brief a boolean value was read
+    @param[in] val  boolean value
+    @return whether parsing should proceed
+    */
+    virtual bool boolean(bool val) = 0;
+
+    /*!
+    @brief an integer number was read
+    @param[in] val  integer value
+    @return whether parsing should proceed
+    */
+    virtual bool number_integer(number_integer_t val) = 0;
+
+    /*!
+    @brief an unsigned integer number was read
+    @param[in] val  unsigned integer value
+    @return whether parsing should proceed
+    */
+    virtual bool number_unsigned(number_unsigned_t val) = 0;
+
+    /*!
+    @brief a floating-point number was read
+    @param[in] val  floating-point value
+    @param[in] s    raw token value
+    @return whether parsing should proceed
+    */
+    virtual bool number_float(number_float_t val, const string_t& s) = 0;
+
+    /*!
+    @brief a string value was read
+    @param[in] val  string value
+    @return whether parsing should proceed
+    @note It is safe to move the passed string value.
+    */
+    virtual bool string(string_t& val) = 0;
+
+    /*!
+    @brief a binary value was read
+    @param[in] val  binary value
+    @return whether parsing should proceed
+    @note It is safe to move the passed binary value.
+    */
+    virtual bool binary(binary_t& val) = 0;
+
+    /*!
+    @brief the beginning of an object was read
+    @param[in] elements  number of object elements or -1 if unknown
+    @return whether parsing should proceed
+    @note binary formats may report the number of elements
+    */
+    virtual bool start_object(std::size_t elements) = 0;
+
+    /*!
+    @brief an object key was read
+    @param[in] val  object key
+    @return whether parsing should proceed
+    @note It is safe to move the passed string.
+    */
+    virtual bool key(string_t& val) = 0;
+
+    /*!
+    @brief the end of an object was read
+    @return whether parsing should proceed
+    */
+    virtual bool end_object() = 0;
+
+    /*!
+    @brief the beginning of an array was read
+    @param[in] elements  number of array elements or -1 if unknown
+    @return whether parsing should proceed
+    @note binary formats may report the number of elements
+    */
+    virtual bool start_array(std::size_t elements) = 0;
+
+    /*!
+    @brief the end of an array was read
+    @return whether parsing should proceed
+    */
+    virtual bool end_array() = 0;
+
+    /*!
+    @brief a parse error occurred
+    @param[in] position    the position in the input where the error occurs
+    @param[in] last_token  the last read token
+    @param[in] ex          an exception object describing the error
+    @return whether parsing should proceed (must return false)
+    */
+    virtual bool parse_error(std::size_t position,
+                             const std::string& last_token,
+                             const detail::exception& ex) = 0;
+
+    json_sax() = default;
+    json_sax(const json_sax&) = default;
+    json_sax(json_sax&&) noexcept = default;
+    json_sax& operator=(const json_sax&) = default;
+    json_sax& operator=(json_sax&&) noexcept = default;
+    virtual ~json_sax() = default;
+};
+
+namespace detail
+{
+/*!
+@brief SAX implementation to create a JSON value from SAX events
+
+This class implements the @ref json_sax interface and processes the SAX events
+to create a JSON value which makes it basically a DOM parser. The structure or
+hierarchy of the JSON value is managed by the stack `ref_stack` which contains
+a pointer to the respective array or object for each recursion depth.
+
+After successful parsing, the value that is passed by reference to the
+constructor contains the parsed value.
+
+@tparam BasicJsonType  the JSON type
+*/
+template<typename BasicJsonType>
+class json_sax_dom_parser
+{
+  public:
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+
+    /*!
+    @param[in,out] r  reference to a JSON value that is manipulated while
+                       parsing
+    @param[in] allow_exceptions_  whether parse errors yield exceptions
+    */
+    explicit json_sax_dom_parser(BasicJsonType& r, const bool allow_exceptions_ = true)
+        : root(r), allow_exceptions(allow_exceptions_)
+    {}
+
+    // make class move-only
+    json_sax_dom_parser(const json_sax_dom_parser&) = delete;
+    json_sax_dom_parser(json_sax_dom_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    json_sax_dom_parser& operator=(const json_sax_dom_parser&) = delete;
+    json_sax_dom_parser& operator=(json_sax_dom_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    ~json_sax_dom_parser() = default;
+
+    bool null()
+    {
+        handle_value(nullptr);
+        return true;
+    }
+
+    bool boolean(bool val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_integer(number_integer_t val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_unsigned(number_unsigned_t val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_float(number_float_t val, const string_t& /*unused*/)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool string(string_t& val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool binary(binary_t& val)
+    {
+        handle_value(std::move(val));
+        return true;
+    }
+
+    bool start_object(std::size_t len)
+    {
+        ref_stack.push_back(handle_value(BasicJsonType::value_t::object));
+
+        if (JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
+        {
+            JSON_THROW(out_of_range::create(408, concat("excessive object size: ", std::to_string(len)), ref_stack.back()));
+        }
+
+        return true;
+    }
+
+    bool key(string_t& val)
+    {
+        JSON_ASSERT(!ref_stack.empty());
+        JSON_ASSERT(ref_stack.back()->is_object());
+
+        // add null at given key and store the reference for later
+        object_element = &(ref_stack.back()->m_data.m_value.object->operator[](val));
+        return true;
+    }
+
+    bool end_object()
+    {
+        JSON_ASSERT(!ref_stack.empty());
+        JSON_ASSERT(ref_stack.back()->is_object());
+
+        ref_stack.back()->set_parents();
+        ref_stack.pop_back();
+        return true;
+    }
+
+    bool start_array(std::size_t len)
+    {
+        ref_stack.push_back(handle_value(BasicJsonType::value_t::array));
+
+        if (JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
+        {
+            JSON_THROW(out_of_range::create(408, concat("excessive array size: ", std::to_string(len)), ref_stack.back()));
+        }
+
+        return true;
+    }
+
+    bool end_array()
+    {
+        JSON_ASSERT(!ref_stack.empty());
+        JSON_ASSERT(ref_stack.back()->is_array());
+
+        ref_stack.back()->set_parents();
+        ref_stack.pop_back();
+        return true;
+    }
+
+    template<class Exception>
+    bool parse_error(std::size_t /*unused*/, const std::string& /*unused*/,
+                     const Exception& ex)
+    {
+        errored = true;
+        static_cast<void>(ex);
+        if (allow_exceptions)
+        {
+            JSON_THROW(ex);
+        }
+        return false;
+    }
+
+    constexpr bool is_errored() const
+    {
+        return errored;
+    }
+
+  private:
+    /*!
+    @invariant If the ref stack is empty, then the passed value will be the new
+               root.
+    @invariant If the ref stack contains a value, then it is an array or an
+               object to which we can add elements
+    */
+    template<typename Value>
+    JSON_HEDLEY_RETURNS_NON_NULL
+    BasicJsonType* handle_value(Value&& v)
+    {
+        if (ref_stack.empty())
+        {
+            root = BasicJsonType(std::forward<Value>(v));
+            return &root;
+        }
+
+        JSON_ASSERT(ref_stack.back()->is_array() || ref_stack.back()->is_object());
+
+        if (ref_stack.back()->is_array())
+        {
+            ref_stack.back()->m_data.m_value.array->emplace_back(std::forward<Value>(v));
+            return &(ref_stack.back()->m_data.m_value.array->back());
+        }
+
+        JSON_ASSERT(ref_stack.back()->is_object());
+        JSON_ASSERT(object_element);
+        *object_element = BasicJsonType(std::forward<Value>(v));
+        return object_element;
+    }
+
+    /// the parsed JSON value
+    BasicJsonType& root;
+    /// stack to model hierarchy of values
+    std::vector<BasicJsonType*> ref_stack {};
+    /// helper to hold the reference for the next object element
+    BasicJsonType* object_element = nullptr;
+    /// whether a syntax error occurred
+    bool errored = false;
+    /// whether to throw exceptions in case of errors
+    const bool allow_exceptions = true;
+};
+
+template<typename BasicJsonType>
+class json_sax_dom_callback_parser
+{
+  public:
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+    using parser_callback_t = typename BasicJsonType::parser_callback_t;
+    using parse_event_t = typename BasicJsonType::parse_event_t;
+
+    json_sax_dom_callback_parser(BasicJsonType& r,
+                                 const parser_callback_t cb,
+                                 const bool allow_exceptions_ = true)
+        : root(r), callback(cb), allow_exceptions(allow_exceptions_)
+    {
+        keep_stack.push_back(true);
+    }
+
+    // make class move-only
+    json_sax_dom_callback_parser(const json_sax_dom_callback_parser&) = delete;
+    json_sax_dom_callback_parser(json_sax_dom_callback_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    json_sax_dom_callback_parser& operator=(const json_sax_dom_callback_parser&) = delete;
+    json_sax_dom_callback_parser& operator=(json_sax_dom_callback_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    ~json_sax_dom_callback_parser() = default;
+
+    bool null()
+    {
+        handle_value(nullptr);
+        return true;
+    }
+
+    bool boolean(bool val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_integer(number_integer_t val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_unsigned(number_unsigned_t val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_float(number_float_t val, const string_t& /*unused*/)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool string(string_t& val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool binary(binary_t& val)
+    {
+        handle_value(std::move(val));
+        return true;
+    }
+
+    bool start_object(std::size_t len)
+    {
+        // check callback for object start
+        const bool keep = callback(static_cast<int>(ref_stack.size()), parse_event_t::object_start, discarded);
+        keep_stack.push_back(keep);
+
+        auto val = handle_value(BasicJsonType::value_t::object, true);
+        ref_stack.push_back(val.second);
+
+        // check object limit
+        if (ref_stack.back() && JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
+        {
+            JSON_THROW(out_of_range::create(408, concat("excessive object size: ", std::to_string(len)), ref_stack.back()));
+        }
+
+        return true;
+    }
+
+    bool key(string_t& val)
+    {
+        BasicJsonType k = BasicJsonType(val);
+
+        // check callback for key
+        const bool keep = callback(static_cast<int>(ref_stack.size()), parse_event_t::key, k);
+        key_keep_stack.push_back(keep);
+
+        // add discarded value at given key and store the reference for later
+        if (keep && ref_stack.back())
+        {
+            object_element = &(ref_stack.back()->m_data.m_value.object->operator[](val) = discarded);
+        }
+
+        return true;
+    }
+
+    bool end_object()
+    {
+        if (ref_stack.back())
+        {
+            if (!callback(static_cast<int>(ref_stack.size()) - 1, parse_event_t::object_end, *ref_stack.back()))
+            {
+                // discard object
+                *ref_stack.back() = discarded;
+            }
+            else
+            {
+                ref_stack.back()->set_parents();
+            }
+        }
+
+        JSON_ASSERT(!ref_stack.empty());
+        JSON_ASSERT(!keep_stack.empty());
+        ref_stack.pop_back();
+        keep_stack.pop_back();
+
+        if (!ref_stack.empty() && ref_stack.back() && ref_stack.back()->is_structured())
+        {
+            // remove discarded value
+            for (auto it = ref_stack.back()->begin(); it != ref_stack.back()->end(); ++it)
+            {
+                if (it->is_discarded())
+                {
+                    ref_stack.back()->erase(it);
+                    break;
+                }
+            }
+        }
+
+        return true;
+    }
+
+    bool start_array(std::size_t len)
+    {
+        const bool keep = callback(static_cast<int>(ref_stack.size()), parse_event_t::array_start, discarded);
+        keep_stack.push_back(keep);
+
+        auto val = handle_value(BasicJsonType::value_t::array, true);
+        ref_stack.push_back(val.second);
+
+        // check array limit
+        if (ref_stack.back() && JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
+        {
+            JSON_THROW(out_of_range::create(408, concat("excessive array size: ", std::to_string(len)), ref_stack.back()));
+        }
+
+        return true;
+    }
+
+    bool end_array()
+    {
+        bool keep = true;
+
+        if (ref_stack.back())
+        {
+            keep = callback(static_cast<int>(ref_stack.size()) - 1, parse_event_t::array_end, *ref_stack.back());
+            if (keep)
+            {
+                ref_stack.back()->set_parents();
+            }
+            else
+            {
+                // discard array
+                *ref_stack.back() = discarded;
+            }
+        }
+
+        JSON_ASSERT(!ref_stack.empty());
+        JSON_ASSERT(!keep_stack.empty());
+        ref_stack.pop_back();
+        keep_stack.pop_back();
+
+        // remove discarded value
+        if (!keep && !ref_stack.empty() && ref_stack.back()->is_array())
+        {
+            ref_stack.back()->m_data.m_value.array->pop_back();
+        }
+
+        return true;
+    }
+
+    template<class Exception>
+    bool parse_error(std::size_t /*unused*/, const std::string& /*unused*/,
+                     const Exception& ex)
+    {
+        errored = true;
+        static_cast<void>(ex);
+        if (allow_exceptions)
+        {
+            JSON_THROW(ex);
+        }
+        return false;
+    }
+
+    constexpr bool is_errored() const
+    {
+        return errored;
+    }
+
+  private:
+    /*!
+    @param[in] v  value to add to the JSON value we build during parsing
+    @param[in] skip_callback  whether we should skip calling the callback
+               function; this is required after start_array() and
+               start_object() SAX events, because otherwise we would call the
+               callback function with an empty array or object, respectively.
+
+    @invariant If the ref stack is empty, then the passed value will be the new
+               root.
+    @invariant If the ref stack contains a value, then it is an array or an
+               object to which we can add elements
+
+    @return pair of boolean (whether value should be kept) and pointer (to the
+            passed value in the ref_stack hierarchy; nullptr if not kept)
+    */
+    template<typename Value>
+    std::pair<bool, BasicJsonType*> handle_value(Value&& v, const bool skip_callback = false)
+    {
+        JSON_ASSERT(!keep_stack.empty());
+
+        // do not handle this value if we know it would be added to a discarded
+        // container
+        if (!keep_stack.back())
+        {
+            return {false, nullptr};
+        }
+
+        // create value
+        auto value = BasicJsonType(std::forward<Value>(v));
+
+        // check callback
+        const bool keep = skip_callback || callback(static_cast<int>(ref_stack.size()), parse_event_t::value, value);
+
+        // do not handle this value if we just learnt it shall be discarded
+        if (!keep)
+        {
+            return {false, nullptr};
+        }
+
+        if (ref_stack.empty())
+        {
+            root = std::move(value);
+            return {true, & root};
+        }
+
+        // skip this value if we already decided to skip the parent
+        // (https://github.com/nlohmann/json/issues/971#issuecomment-413678360)
+        if (!ref_stack.back())
+        {
+            return {false, nullptr};
+        }
+
+        // we now only expect arrays and objects
+        JSON_ASSERT(ref_stack.back()->is_array() || ref_stack.back()->is_object());
+
+        // array
+        if (ref_stack.back()->is_array())
+        {
+            ref_stack.back()->m_data.m_value.array->emplace_back(std::move(value));
+            return {true, & (ref_stack.back()->m_data.m_value.array->back())};
+        }
+
+        // object
+        JSON_ASSERT(ref_stack.back()->is_object());
+        // check if we should store an element for the current key
+        JSON_ASSERT(!key_keep_stack.empty());
+        const bool store_element = key_keep_stack.back();
+        key_keep_stack.pop_back();
+
+        if (!store_element)
+        {
+            return {false, nullptr};
+        }
+
+        JSON_ASSERT(object_element);
+        *object_element = std::move(value);
+        return {true, object_element};
+    }
+
+    /// the parsed JSON value
+    BasicJsonType& root;
+    /// stack to model hierarchy of values
+    std::vector<BasicJsonType*> ref_stack {};
+    /// stack to manage which values to keep
+    std::vector<bool> keep_stack {}; // NOLINT(readability-redundant-member-init)
+    /// stack to manage which object keys to keep
+    std::vector<bool> key_keep_stack {}; // NOLINT(readability-redundant-member-init)
+    /// helper to hold the reference for the next object element
+    BasicJsonType* object_element = nullptr;
+    /// whether a syntax error occurred
+    bool errored = false;
+    /// callback function
+    const parser_callback_t callback = nullptr;
+    /// whether to throw exceptions in case of errors
+    const bool allow_exceptions = true;
+    /// a discarded value for the callback
+    BasicJsonType discarded = BasicJsonType::value_t::discarded;
+};
+
+template<typename BasicJsonType>
+class json_sax_acceptor
+{
+  public:
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+
+    bool null()
+    {
+        return true;
+    }
+
+    bool boolean(bool /*unused*/)
+    {
+        return true;
+    }
+
+    bool number_integer(number_integer_t /*unused*/)
+    {
+        return true;
+    }
+
+    bool number_unsigned(number_unsigned_t /*unused*/)
+    {
+        return true;
+    }
+
+    bool number_float(number_float_t /*unused*/, const string_t& /*unused*/)
+    {
+        return true;
+    }
+
+    bool string(string_t& /*unused*/)
+    {
+        return true;
+    }
+
+    bool binary(binary_t& /*unused*/)
+    {
+        return true;
+    }
+
+    bool start_object(std::size_t /*unused*/ = static_cast<std::size_t>(-1))
+    {
+        return true;
+    }
+
+    bool key(string_t& /*unused*/)
+    {
+        return true;
+    }
+
+    bool end_object()
+    {
+        return true;
+    }
+
+    bool start_array(std::size_t /*unused*/ = static_cast<std::size_t>(-1))
+    {
+        return true;
+    }
+
+    bool end_array()
+    {
+        return true;
+    }
+
+    bool parse_error(std::size_t /*unused*/, const std::string& /*unused*/, const detail::exception& /*unused*/)
+    {
+        return false;
+    }
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/input/lexer.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <array> // array
+#include <clocale> // localeconv
+#include <cstddef> // size_t
+#include <cstdio> // snprintf
+#include <cstdlib> // strtof, strtod, strtold, strtoll, strtoull
+#include <initializer_list> // initializer_list
+#include <string> // char_traits, string
+#include <utility> // move
+#include <vector> // vector
+
+// #include <nlohmann/detail/input/input_adapters.hpp>
+
+// #include <nlohmann/detail/input/position_t.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+///////////
+// lexer //
+///////////
+
+template<typename BasicJsonType>
+class lexer_base
+{
+  public:
+    /// token types for the parser
+    enum class token_type
+    {
+        uninitialized,    ///< indicating the scanner is uninitialized
+        literal_true,     ///< the `true` literal
+        literal_false,    ///< the `false` literal
+        literal_null,     ///< the `null` literal
+        value_string,     ///< a string -- use get_string() for actual value
+        value_unsigned,   ///< an unsigned integer -- use get_number_unsigned() for actual value
+        value_integer,    ///< a signed integer -- use get_number_integer() for actual value
+        value_float,      ///< an floating point number -- use get_number_float() for actual value
+        begin_array,      ///< the character for array begin `[`
+        begin_object,     ///< the character for object begin `{`
+        end_array,        ///< the character for array end `]`
+        end_object,       ///< the character for object end `}`
+        name_separator,   ///< the name separator `:`
+        value_separator,  ///< the value separator `,`
+        parse_error,      ///< indicating a parse error
+        end_of_input,     ///< indicating the end of the input buffer
+        literal_or_value  ///< a literal or the begin of a value (only for diagnostics)
+    };
+
+    /// return name of values of type token_type (only used for errors)
+    JSON_HEDLEY_RETURNS_NON_NULL
+    JSON_HEDLEY_CONST
+    static const char* token_type_name(const token_type t) noexcept
+    {
+        switch (t)
+        {
+            case token_type::uninitialized:
+                return "<uninitialized>";
+            case token_type::literal_true:
+                return "true literal";
+            case token_type::literal_false:
+                return "false literal";
+            case token_type::literal_null:
+                return "null literal";
+            case token_type::value_string:
+                return "string literal";
+            case token_type::value_unsigned:
+            case token_type::value_integer:
+            case token_type::value_float:
+                return "number literal";
+            case token_type::begin_array:
+                return "'['";
+            case token_type::begin_object:
+                return "'{'";
+            case token_type::end_array:
+                return "']'";
+            case token_type::end_object:
+                return "'}'";
+            case token_type::name_separator:
+                return "':'";
+            case token_type::value_separator:
+                return "','";
+            case token_type::parse_error:
+                return "<parse error>";
+            case token_type::end_of_input:
+                return "end of input";
+            case token_type::literal_or_value:
+                return "'[', '{', or a literal";
+            // LCOV_EXCL_START
+            default: // catch non-enum values
+                return "unknown token";
+                // LCOV_EXCL_STOP
+        }
+    }
+};
+/*!
+@brief lexical analysis
+
+This class organizes the lexical analysis during JSON deserialization.
+*/
+template<typename BasicJsonType, typename InputAdapterType>
+class lexer : public lexer_base<BasicJsonType>
+{
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using char_type = typename InputAdapterType::char_type;
+    using char_int_type = typename char_traits<char_type>::int_type;
+
+  public:
+    using token_type = typename lexer_base<BasicJsonType>::token_type;
+
+    explicit lexer(InputAdapterType&& adapter, bool ignore_comments_ = false) noexcept
+        : ia(std::move(adapter))
+        , ignore_comments(ignore_comments_)
+        , decimal_point_char(static_cast<char_int_type>(get_decimal_point()))
+    {}
+
+    // delete because of pointer members
+    lexer(const lexer&) = delete;
+    lexer(lexer&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    lexer& operator=(lexer&) = delete;
+    lexer& operator=(lexer&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    ~lexer() = default;
+
+  private:
+    /////////////////////
+    // locales
+    /////////////////////
+
+    /// return the locale-dependent decimal point
+    JSON_HEDLEY_PURE
+    static char get_decimal_point() noexcept
+    {
+        const auto* loc = localeconv();
+        JSON_ASSERT(loc != nullptr);
+        return (loc->decimal_point == nullptr) ? '.' : *(loc->decimal_point);
+    }
+
+    /////////////////////
+    // scan functions
+    /////////////////////
+
+    /*!
+    @brief get codepoint from 4 hex characters following `\u`
+
+    For input "\u c1 c2 c3 c4" the codepoint is:
+      (c1 * 0x1000) + (c2 * 0x0100) + (c3 * 0x0010) + c4
+    = (c1 << 12) + (c2 << 8) + (c3 << 4) + (c4 << 0)
+
+    Furthermore, the possible characters '0'..'9', 'A'..'F', and 'a'..'f'
+    must be converted to the integers 0x0..0x9, 0xA..0xF, 0xA..0xF, resp. The
+    conversion is done by subtracting the offset (0x30, 0x37, and 0x57)
+    between the ASCII value of the character and the desired integer value.
+
+    @return codepoint (0x0000..0xFFFF) or -1 in case of an error (e.g. EOF or
+            non-hex character)
+    */
+    int get_codepoint()
+    {
+        // this function only makes sense after reading `\u`
+        JSON_ASSERT(current == 'u');
+        int codepoint = 0;
+
+        const auto factors = { 12u, 8u, 4u, 0u };
+        for (const auto factor : factors)
+        {
+            get();
+
+            if (current >= '0' && current <= '9')
+            {
+                codepoint += static_cast<int>((static_cast<unsigned int>(current) - 0x30u) << factor);
+            }
+            else if (current >= 'A' && current <= 'F')
+            {
+                codepoint += static_cast<int>((static_cast<unsigned int>(current) - 0x37u) << factor);
+            }
+            else if (current >= 'a' && current <= 'f')
+            {
+                codepoint += static_cast<int>((static_cast<unsigned int>(current) - 0x57u) << factor);
+            }
+            else
+            {
+                return -1;
+            }
+        }
+
+        JSON_ASSERT(0x0000 <= codepoint && codepoint <= 0xFFFF);
+        return codepoint;
+    }
+
+    /*!
+    @brief check if the next byte(s) are inside a given range
+
+    Adds the current byte and, for each passed range, reads a new byte and
+    checks if it is inside the range. If a violation was detected, set up an
+    error message and return false. Otherwise, return true.
+
+    @param[in] ranges  list of integers; interpreted as list of pairs of
+                       inclusive lower and upper bound, respectively
+
+    @pre The passed list @a ranges must have 2, 4, or 6 elements; that is,
+         1, 2, or 3 pairs. This precondition is enforced by an assertion.
+
+    @return true if and only if no range violation was detected
+    */
+    bool next_byte_in_range(std::initializer_list<char_int_type> ranges)
+    {
+        JSON_ASSERT(ranges.size() == 2 || ranges.size() == 4 || ranges.size() == 6);
+        add(current);
+
+        for (auto range = ranges.begin(); range != ranges.end(); ++range)
+        {
+            get();
+            if (JSON_HEDLEY_LIKELY(*range <= current && current <= *(++range))) // NOLINT(bugprone-inc-dec-in-conditions)
+            {
+                add(current);
+            }
+            else
+            {
+                error_message = "invalid string: ill-formed UTF-8 byte";
+                return false;
+            }
+        }
+
+        return true;
+    }
+
+    /*!
+    @brief scan a string literal
+
+    This function scans a string according to Sect. 7 of RFC 8259. While
+    scanning, bytes are escaped and copied into buffer token_buffer. Then the
+    function returns successfully, token_buffer is *not* null-terminated (as it
+    may contain \0 bytes), and token_buffer.size() is the number of bytes in the
+    string.
+
+    @return token_type::value_string if string could be successfully scanned,
+            token_type::parse_error otherwise
+
+    @note In case of errors, variable error_message contains a textual
+          description.
+    */
+    token_type scan_string()
+    {
+        // reset token_buffer (ignore opening quote)
+        reset();
+
+        // we entered the function by reading an open quote
+        JSON_ASSERT(current == '\"');
+
+        while (true)
+        {
+            // get next character
+            switch (get())
+            {
+                // end of file while parsing string
+                case char_traits<char_type>::eof():
+                {
+                    error_message = "invalid string: missing closing quote";
+                    return token_type::parse_error;
+                }
+
+                // closing quote
+                case '\"':
+                {
+                    return token_type::value_string;
+                }
+
+                // escapes
+                case '\\':
+                {
+                    switch (get())
+                    {
+                        // quotation mark
+                        case '\"':
+                            add('\"');
+                            break;
+                        // reverse solidus
+                        case '\\':
+                            add('\\');
+                            break;
+                        // solidus
+                        case '/':
+                            add('/');
+                            break;
+                        // backspace
+                        case 'b':
+                            add('\b');
+                            break;
+                        // form feed
+                        case 'f':
+                            add('\f');
+                            break;
+                        // line feed
+                        case 'n':
+                            add('\n');
+                            break;
+                        // carriage return
+                        case 'r':
+                            add('\r');
+                            break;
+                        // tab
+                        case 't':
+                            add('\t');
+                            break;
+
+                        // unicode escapes
+                        case 'u':
+                        {
+                            const int codepoint1 = get_codepoint();
+                            int codepoint = codepoint1; // start with codepoint1
+
+                            if (JSON_HEDLEY_UNLIKELY(codepoint1 == -1))
+                            {
+                                error_message = "invalid string: '\\u' must be followed by 4 hex digits";
+                                return token_type::parse_error;
+                            }
+
+                            // check if code point is a high surrogate
+                            if (0xD800 <= codepoint1 && codepoint1 <= 0xDBFF)
+                            {
+                                // expect next \uxxxx entry
+                                if (JSON_HEDLEY_LIKELY(get() == '\\' && get() == 'u'))
+                                {
+                                    const int codepoint2 = get_codepoint();
+
+                                    if (JSON_HEDLEY_UNLIKELY(codepoint2 == -1))
+                                    {
+                                        error_message = "invalid string: '\\u' must be followed by 4 hex digits";
+                                        return token_type::parse_error;
+                                    }
+
+                                    // check if codepoint2 is a low surrogate
+                                    if (JSON_HEDLEY_LIKELY(0xDC00 <= codepoint2 && codepoint2 <= 0xDFFF))
+                                    {
+                                        // overwrite codepoint
+                                        codepoint = static_cast<int>(
+                                                        // high surrogate occupies the most significant 22 bits
+                                                        (static_cast<unsigned int>(codepoint1) << 10u)
+                                                        // low surrogate occupies the least significant 15 bits
+                                                        + static_cast<unsigned int>(codepoint2)
+                                                        // there is still the 0xD800, 0xDC00 and 0x10000 noise
+                                                        // in the result, so we have to subtract with:
+                                                        // (0xD800 << 10) + DC00 - 0x10000 = 0x35FDC00
+                                                        - 0x35FDC00u);
+                                    }
+                                    else
+                                    {
+                                        error_message = "invalid string: surrogate U+D800..U+DBFF must be followed by U+DC00..U+DFFF";
+                                        return token_type::parse_error;
+                                    }
+                                }
+                                else
+                                {
+                                    error_message = "invalid string: surrogate U+D800..U+DBFF must be followed by U+DC00..U+DFFF";
+                                    return token_type::parse_error;
+                                }
+                            }
+                            else
+                            {
+                                if (JSON_HEDLEY_UNLIKELY(0xDC00 <= codepoint1 && codepoint1 <= 0xDFFF))
+                                {
+                                    error_message = "invalid string: surrogate U+DC00..U+DFFF must follow U+D800..U+DBFF";
+                                    return token_type::parse_error;
+                                }
+                            }
+
+                            // result of the above calculation yields a proper codepoint
+                            JSON_ASSERT(0x00 <= codepoint && codepoint <= 0x10FFFF);
+
+                            // translate codepoint into bytes
+                            if (codepoint < 0x80)
+                            {
+                                // 1-byte characters: 0xxxxxxx (ASCII)
+                                add(static_cast<char_int_type>(codepoint));
+                            }
+                            else if (codepoint <= 0x7FF)
+                            {
+                                // 2-byte characters: 110xxxxx 10xxxxxx
+                                add(static_cast<char_int_type>(0xC0u | (static_cast<unsigned int>(codepoint) >> 6u)));
+                                add(static_cast<char_int_type>(0x80u | (static_cast<unsigned int>(codepoint) & 0x3Fu)));
+                            }
+                            else if (codepoint <= 0xFFFF)
+                            {
+                                // 3-byte characters: 1110xxxx 10xxxxxx 10xxxxxx
+                                add(static_cast<char_int_type>(0xE0u | (static_cast<unsigned int>(codepoint) >> 12u)));
+                                add(static_cast<char_int_type>(0x80u | ((static_cast<unsigned int>(codepoint) >> 6u) & 0x3Fu)));
+                                add(static_cast<char_int_type>(0x80u | (static_cast<unsigned int>(codepoint) & 0x3Fu)));
+                            }
+                            else
+                            {
+                                // 4-byte characters: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
+                                add(static_cast<char_int_type>(0xF0u | (static_cast<unsigned int>(codepoint) >> 18u)));
+                                add(static_cast<char_int_type>(0x80u | ((static_cast<unsigned int>(codepoint) >> 12u) & 0x3Fu)));
+                                add(static_cast<char_int_type>(0x80u | ((static_cast<unsigned int>(codepoint) >> 6u) & 0x3Fu)));
+                                add(static_cast<char_int_type>(0x80u | (static_cast<unsigned int>(codepoint) & 0x3Fu)));
+                            }
+
+                            break;
+                        }
+
+                        // other characters after escape
+                        default:
+                            error_message = "invalid string: forbidden character after backslash";
+                            return token_type::parse_error;
+                    }
+
+                    break;
+                }
+
+                // invalid control characters
+                case 0x00:
+                {
+                    error_message = "invalid string: control character U+0000 (NUL) must be escaped to \\u0000";
+                    return token_type::parse_error;
+                }
+
+                case 0x01:
+                {
+                    error_message = "invalid string: control character U+0001 (SOH) must be escaped to \\u0001";
+                    return token_type::parse_error;
+                }
+
+                case 0x02:
+                {
+                    error_message = "invalid string: control character U+0002 (STX) must be escaped to \\u0002";
+                    return token_type::parse_error;
+                }
+
+                case 0x03:
+                {
+                    error_message = "invalid string: control character U+0003 (ETX) must be escaped to \\u0003";
+                    return token_type::parse_error;
+                }
+
+                case 0x04:
+                {
+                    error_message = "invalid string: control character U+0004 (EOT) must be escaped to \\u0004";
+                    return token_type::parse_error;
+                }
+
+                case 0x05:
+                {
+                    error_message = "invalid string: control character U+0005 (ENQ) must be escaped to \\u0005";
+                    return token_type::parse_error;
+                }
+
+                case 0x06:
+                {
+                    error_message = "invalid string: control character U+0006 (ACK) must be escaped to \\u0006";
+                    return token_type::parse_error;
+                }
+
+                case 0x07:
+                {
+                    error_message = "invalid string: control character U+0007 (BEL) must be escaped to \\u0007";
+                    return token_type::parse_error;
+                }
+
+                case 0x08:
+                {
+                    error_message = "invalid string: control character U+0008 (BS) must be escaped to \\u0008 or \\b";
+                    return token_type::parse_error;
+                }
+
+                case 0x09:
+                {
+                    error_message = "invalid string: control character U+0009 (HT) must be escaped to \\u0009 or \\t";
+                    return token_type::parse_error;
+                }
+
+                case 0x0A:
+                {
+                    error_message = "invalid string: control character U+000A (LF) must be escaped to \\u000A or \\n";
+                    return token_type::parse_error;
+                }
+
+                case 0x0B:
+                {
+                    error_message = "invalid string: control character U+000B (VT) must be escaped to \\u000B";
+                    return token_type::parse_error;
+                }
+
+                case 0x0C:
+                {
+                    error_message = "invalid string: control character U+000C (FF) must be escaped to \\u000C or \\f";
+                    return token_type::parse_error;
+                }
+
+                case 0x0D:
+                {
+                    error_message = "invalid string: control character U+000D (CR) must be escaped to \\u000D or \\r";
+                    return token_type::parse_error;
+                }
+
+                case 0x0E:
+                {
+                    error_message = "invalid string: control character U+000E (SO) must be escaped to \\u000E";
+                    return token_type::parse_error;
+                }
+
+                case 0x0F:
+                {
+                    error_message = "invalid string: control character U+000F (SI) must be escaped to \\u000F";
+                    return token_type::parse_error;
+                }
+
+                case 0x10:
+                {
+                    error_message = "invalid string: control character U+0010 (DLE) must be escaped to \\u0010";
+                    return token_type::parse_error;
+                }
+
+                case 0x11:
+                {
+                    error_message = "invalid string: control character U+0011 (DC1) must be escaped to \\u0011";
+                    return token_type::parse_error;
+                }
+
+                case 0x12:
+                {
+                    error_message = "invalid string: control character U+0012 (DC2) must be escaped to \\u0012";
+                    return token_type::parse_error;
+                }
+
+                case 0x13:
+                {
+                    error_message = "invalid string: control character U+0013 (DC3) must be escaped to \\u0013";
+                    return token_type::parse_error;
+                }
+
+                case 0x14:
+                {
+                    error_message = "invalid string: control character U+0014 (DC4) must be escaped to \\u0014";
+                    return token_type::parse_error;
+                }
+
+                case 0x15:
+                {
+                    error_message = "invalid string: control character U+0015 (NAK) must be escaped to \\u0015";
+                    return token_type::parse_error;
+                }
+
+                case 0x16:
+                {
+                    error_message = "invalid string: control character U+0016 (SYN) must be escaped to \\u0016";
+                    return token_type::parse_error;
+                }
+
+                case 0x17:
+                {
+                    error_message = "invalid string: control character U+0017 (ETB) must be escaped to \\u0017";
+                    return token_type::parse_error;
+                }
+
+                case 0x18:
+                {
+                    error_message = "invalid string: control character U+0018 (CAN) must be escaped to \\u0018";
+                    return token_type::parse_error;
+                }
+
+                case 0x19:
+                {
+                    error_message = "invalid string: control character U+0019 (EM) must be escaped to \\u0019";
+                    return token_type::parse_error;
+                }
+
+                case 0x1A:
+                {
+                    error_message = "invalid string: control character U+001A (SUB) must be escaped to \\u001A";
+                    return token_type::parse_error;
+                }
+
+                case 0x1B:
+                {
+                    error_message = "invalid string: control character U+001B (ESC) must be escaped to \\u001B";
+                    return token_type::parse_error;
+                }
+
+                case 0x1C:
+                {
+                    error_message = "invalid string: control character U+001C (FS) must be escaped to \\u001C";
+                    return token_type::parse_error;
+                }
+
+                case 0x1D:
+                {
+                    error_message = "invalid string: control character U+001D (GS) must be escaped to \\u001D";
+                    return token_type::parse_error;
+                }
+
+                case 0x1E:
+                {
+                    error_message = "invalid string: control character U+001E (RS) must be escaped to \\u001E";
+                    return token_type::parse_error;
+                }
+
+                case 0x1F:
+                {
+                    error_message = "invalid string: control character U+001F (US) must be escaped to \\u001F";
+                    return token_type::parse_error;
+                }
+
+                // U+0020..U+007F (except U+0022 (quote) and U+005C (backspace))
+                case 0x20:
+                case 0x21:
+                case 0x23:
+                case 0x24:
+                case 0x25:
+                case 0x26:
+                case 0x27:
+                case 0x28:
+                case 0x29:
+                case 0x2A:
+                case 0x2B:
+                case 0x2C:
+                case 0x2D:
+                case 0x2E:
+                case 0x2F:
+                case 0x30:
+                case 0x31:
+                case 0x32:
+                case 0x33:
+                case 0x34:
+                case 0x35:
+                case 0x36:
+                case 0x37:
+                case 0x38:
+                case 0x39:
+                case 0x3A:
+                case 0x3B:
+                case 0x3C:
+                case 0x3D:
+                case 0x3E:
+                case 0x3F:
+                case 0x40:
+                case 0x41:
+                case 0x42:
+                case 0x43:
+                case 0x44:
+                case 0x45:
+                case 0x46:
+                case 0x47:
+                case 0x48:
+                case 0x49:
+                case 0x4A:
+                case 0x4B:
+                case 0x4C:
+                case 0x4D:
+                case 0x4E:
+                case 0x4F:
+                case 0x50:
+                case 0x51:
+                case 0x52:
+                case 0x53:
+                case 0x54:
+                case 0x55:
+                case 0x56:
+                case 0x57:
+                case 0x58:
+                case 0x59:
+                case 0x5A:
+                case 0x5B:
+                case 0x5D:
+                case 0x5E:
+                case 0x5F:
+                case 0x60:
+                case 0x61:
+                case 0x62:
+                case 0x63:
+                case 0x64:
+                case 0x65:
+                case 0x66:
+                case 0x67:
+                case 0x68:
+                case 0x69:
+                case 0x6A:
+                case 0x6B:
+                case 0x6C:
+                case 0x6D:
+                case 0x6E:
+                case 0x6F:
+                case 0x70:
+                case 0x71:
+                case 0x72:
+                case 0x73:
+                case 0x74:
+                case 0x75:
+                case 0x76:
+                case 0x77:
+                case 0x78:
+                case 0x79:
+                case 0x7A:
+                case 0x7B:
+                case 0x7C:
+                case 0x7D:
+                case 0x7E:
+                case 0x7F:
+                {
+                    add(current);
+                    break;
+                }
+
+                // U+0080..U+07FF: bytes C2..DF 80..BF
+                case 0xC2:
+                case 0xC3:
+                case 0xC4:
+                case 0xC5:
+                case 0xC6:
+                case 0xC7:
+                case 0xC8:
+                case 0xC9:
+                case 0xCA:
+                case 0xCB:
+                case 0xCC:
+                case 0xCD:
+                case 0xCE:
+                case 0xCF:
+                case 0xD0:
+                case 0xD1:
+                case 0xD2:
+                case 0xD3:
+                case 0xD4:
+                case 0xD5:
+                case 0xD6:
+                case 0xD7:
+                case 0xD8:
+                case 0xD9:
+                case 0xDA:
+                case 0xDB:
+                case 0xDC:
+                case 0xDD:
+                case 0xDE:
+                case 0xDF:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!next_byte_in_range({0x80, 0xBF})))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+0800..U+0FFF: bytes E0 A0..BF 80..BF
+                case 0xE0:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0xA0, 0xBF, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+1000..U+CFFF: bytes E1..EC 80..BF 80..BF
+                // U+E000..U+FFFF: bytes EE..EF 80..BF 80..BF
+                case 0xE1:
+                case 0xE2:
+                case 0xE3:
+                case 0xE4:
+                case 0xE5:
+                case 0xE6:
+                case 0xE7:
+                case 0xE8:
+                case 0xE9:
+                case 0xEA:
+                case 0xEB:
+                case 0xEC:
+                case 0xEE:
+                case 0xEF:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0xBF, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+D000..U+D7FF: bytes ED 80..9F 80..BF
+                case 0xED:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0x9F, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+10000..U+3FFFF F0 90..BF 80..BF 80..BF
+                case 0xF0:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x90, 0xBF, 0x80, 0xBF, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+40000..U+FFFFF F1..F3 80..BF 80..BF 80..BF
+                case 0xF1:
+                case 0xF2:
+                case 0xF3:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0xBF, 0x80, 0xBF, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+100000..U+10FFFF F4 80..8F 80..BF 80..BF
+                case 0xF4:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0x8F, 0x80, 0xBF, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // remaining bytes (80..C1 and F5..FF) are ill-formed
+                default:
+                {
+                    error_message = "invalid string: ill-formed UTF-8 byte";
+                    return token_type::parse_error;
+                }
+            }
+        }
+    }
+
+    /*!
+     * @brief scan a comment
+     * @return whether comment could be scanned successfully
+     */
+    bool scan_comment()
+    {
+        switch (get())
+        {
+            // single-line comments skip input until a newline or EOF is read
+            case '/':
+            {
+                while (true)
+                {
+                    switch (get())
+                    {
+                        case '\n':
+                        case '\r':
+                        case char_traits<char_type>::eof():
+                        case '\0':
+                            return true;
+
+                        default:
+                            break;
+                    }
+                }
+            }
+
+            // multi-line comments skip input until */ is read
+            case '*':
+            {
+                while (true)
+                {
+                    switch (get())
+                    {
+                        case char_traits<char_type>::eof():
+                        case '\0':
+                        {
+                            error_message = "invalid comment; missing closing '*/'";
+                            return false;
+                        }
+
+                        case '*':
+                        {
+                            switch (get())
+                            {
+                                case '/':
+                                    return true;
+
+                                default:
+                                {
+                                    unget();
+                                    continue;
+                                }
+                            }
+                        }
+
+                        default:
+                            continue;
+                    }
+                }
+            }
+
+            // unexpected character after reading '/'
+            default:
+            {
+                error_message = "invalid comment; expecting '/' or '*' after '/'";
+                return false;
+            }
+        }
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    static void strtof(float& f, const char* str, char** endptr) noexcept
+    {
+        f = std::strtof(str, endptr);
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    static void strtof(double& f, const char* str, char** endptr) noexcept
+    {
+        f = std::strtod(str, endptr);
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    static void strtof(long double& f, const char* str, char** endptr) noexcept
+    {
+        f = std::strtold(str, endptr);
+    }
+
+    /*!
+    @brief scan a number literal
+
+    This function scans a string according to Sect. 6 of RFC 8259.
+
+    The function is realized with a deterministic finite state machine derived
+    from the grammar described in RFC 8259. Starting in state "init", the
+    input is read and used to determined the next state. Only state "done"
+    accepts the number. State "error" is a trap state to model errors. In the
+    table below, "anything" means any character but the ones listed before.
+
+    state    | 0        | 1-9      | e E      | +       | -       | .        | anything
+    ---------|----------|----------|----------|---------|---------|----------|-----------
+    init     | zero     | any1     | [error]  | [error] | minus   | [error]  | [error]
+    minus    | zero     | any1     | [error]  | [error] | [error] | [error]  | [error]
+    zero     | done     | done     | exponent | done    | done    | decimal1 | done
+    any1     | any1     | any1     | exponent | done    | done    | decimal1 | done
+    decimal1 | decimal2 | decimal2 | [error]  | [error] | [error] | [error]  | [error]
+    decimal2 | decimal2 | decimal2 | exponent | done    | done    | done     | done
+    exponent | any2     | any2     | [error]  | sign    | sign    | [error]  | [error]
+    sign     | any2     | any2     | [error]  | [error] | [error] | [error]  | [error]
+    any2     | any2     | any2     | done     | done    | done    | done     | done
+
+    The state machine is realized with one label per state (prefixed with
+    "scan_number_") and `goto` statements between them. The state machine
+    contains cycles, but any cycle can be left when EOF is read. Therefore,
+    the function is guaranteed to terminate.
+
+    During scanning, the read bytes are stored in token_buffer. This string is
+    then converted to a signed integer, an unsigned integer, or a
+    floating-point number.
+
+    @return token_type::value_unsigned, token_type::value_integer, or
+            token_type::value_float if number could be successfully scanned,
+            token_type::parse_error otherwise
+
+    @note The scanner is independent of the current locale. Internally, the
+          locale's decimal point is used instead of `.` to work with the
+          locale-dependent converters.
+    */
+    token_type scan_number()  // lgtm [cpp/use-of-goto]
+    {
+        // reset token_buffer to store the number's bytes
+        reset();
+
+        // the type of the parsed number; initially set to unsigned; will be
+        // changed if minus sign, decimal point or exponent is read
+        token_type number_type = token_type::value_unsigned;
+
+        // state (init): we just found out we need to scan a number
+        switch (current)
+        {
+            case '-':
+            {
+                add(current);
+                goto scan_number_minus;
+            }
+
+            case '0':
+            {
+                add(current);
+                goto scan_number_zero;
+            }
+
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any1;
+            }
+
+            // all other characters are rejected outside scan_number()
+            default:            // LCOV_EXCL_LINE
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        }
+
+scan_number_minus:
+        // state: we just parsed a leading minus sign
+        number_type = token_type::value_integer;
+        switch (get())
+        {
+            case '0':
+            {
+                add(current);
+                goto scan_number_zero;
+            }
+
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any1;
+            }
+
+            default:
+            {
+                error_message = "invalid number; expected digit after '-'";
+                return token_type::parse_error;
+            }
+        }
+
+scan_number_zero:
+        // state: we just parse a zero (maybe with a leading minus sign)
+        switch (get())
+        {
+            case '.':
+            {
+                add(decimal_point_char);
+                goto scan_number_decimal1;
+            }
+
+            case 'e':
+            case 'E':
+            {
+                add(current);
+                goto scan_number_exponent;
+            }
+
+            default:
+                goto scan_number_done;
+        }
+
+scan_number_any1:
+        // state: we just parsed a number 0-9 (maybe with a leading minus sign)
+        switch (get())
+        {
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any1;
+            }
+
+            case '.':
+            {
+                add(decimal_point_char);
+                goto scan_number_decimal1;
+            }
+
+            case 'e':
+            case 'E':
+            {
+                add(current);
+                goto scan_number_exponent;
+            }
+
+            default:
+                goto scan_number_done;
+        }
+
+scan_number_decimal1:
+        // state: we just parsed a decimal point
+        number_type = token_type::value_float;
+        switch (get())
+        {
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_decimal2;
+            }
+
+            default:
+            {
+                error_message = "invalid number; expected digit after '.'";
+                return token_type::parse_error;
+            }
+        }
+
+scan_number_decimal2:
+        // we just parsed at least one number after a decimal point
+        switch (get())
+        {
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_decimal2;
+            }
+
+            case 'e':
+            case 'E':
+            {
+                add(current);
+                goto scan_number_exponent;
+            }
+
+            default:
+                goto scan_number_done;
+        }
+
+scan_number_exponent:
+        // we just parsed an exponent
+        number_type = token_type::value_float;
+        switch (get())
+        {
+            case '+':
+            case '-':
+            {
+                add(current);
+                goto scan_number_sign;
+            }
+
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any2;
+            }
+
+            default:
+            {
+                error_message =
+                    "invalid number; expected '+', '-', or digit after exponent";
+                return token_type::parse_error;
+            }
+        }
+
+scan_number_sign:
+        // we just parsed an exponent sign
+        switch (get())
+        {
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any2;
+            }
+
+            default:
+            {
+                error_message = "invalid number; expected digit after exponent sign";
+                return token_type::parse_error;
+            }
+        }
+
+scan_number_any2:
+        // we just parsed a number after the exponent or exponent sign
+        switch (get())
+        {
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any2;
+            }
+
+            default:
+                goto scan_number_done;
+        }
+
+scan_number_done:
+        // unget the character after the number (we only read it to know that
+        // we are done scanning a number)
+        unget();
+
+        char* endptr = nullptr; // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+        errno = 0;
+
+        // try to parse integers first and fall back to floats
+        if (number_type == token_type::value_unsigned)
+        {
+            const auto x = std::strtoull(token_buffer.data(), &endptr, 10);
+
+            // we checked the number format before
+            JSON_ASSERT(endptr == token_buffer.data() + token_buffer.size());
+
+            if (errno == 0)
+            {
+                value_unsigned = static_cast<number_unsigned_t>(x);
+                if (value_unsigned == x)
+                {
+                    return token_type::value_unsigned;
+                }
+            }
+        }
+        else if (number_type == token_type::value_integer)
+        {
+            const auto x = std::strtoll(token_buffer.data(), &endptr, 10);
+
+            // we checked the number format before
+            JSON_ASSERT(endptr == token_buffer.data() + token_buffer.size());
+
+            if (errno == 0)
+            {
+                value_integer = static_cast<number_integer_t>(x);
+                if (value_integer == x)
+                {
+                    return token_type::value_integer;
+                }
+            }
+        }
+
+        // this code is reached if we parse a floating-point number or if an
+        // integer conversion above failed
+        strtof(value_float, token_buffer.data(), &endptr);
+
+        // we checked the number format before
+        JSON_ASSERT(endptr == token_buffer.data() + token_buffer.size());
+
+        return token_type::value_float;
+    }
+
+    /*!
+    @param[in] literal_text  the literal text to expect
+    @param[in] length        the length of the passed literal text
+    @param[in] return_type   the token type to return on success
+    */
+    JSON_HEDLEY_NON_NULL(2)
+    token_type scan_literal(const char_type* literal_text, const std::size_t length,
+                            token_type return_type)
+    {
+        JSON_ASSERT(char_traits<char_type>::to_char_type(current) == literal_text[0]);
+        for (std::size_t i = 1; i < length; ++i)
+        {
+            if (JSON_HEDLEY_UNLIKELY(char_traits<char_type>::to_char_type(get()) != literal_text[i]))
+            {
+                error_message = "invalid literal";
+                return token_type::parse_error;
+            }
+        }
+        return return_type;
+    }
+
+    /////////////////////
+    // input management
+    /////////////////////
+
+    /// reset token_buffer; current character is beginning of token
+    void reset() noexcept
+    {
+        token_buffer.clear();
+        token_string.clear();
+        token_string.push_back(char_traits<char_type>::to_char_type(current));
+    }
+
+    /*
+    @brief get next character from the input
+
+    This function provides the interface to the used input adapter. It does
+    not throw in case the input reached EOF, but returns a
+    `char_traits<char>::eof()` in that case.  Stores the scanned characters
+    for use in error messages.
+
+    @return character read from the input
+    */
+    char_int_type get()
+    {
+        ++position.chars_read_total;
+        ++position.chars_read_current_line;
+
+        if (next_unget)
+        {
+            // just reset the next_unget variable and work with current
+            next_unget = false;
+        }
+        else
+        {
+            current = ia.get_character();
+        }
+
+        if (JSON_HEDLEY_LIKELY(current != char_traits<char_type>::eof()))
+        {
+            token_string.push_back(char_traits<char_type>::to_char_type(current));
+        }
+
+        if (current == '\n')
+        {
+            ++position.lines_read;
+            position.chars_read_current_line = 0;
+        }
+
+        return current;
+    }
+
+    /*!
+    @brief unget current character (read it again on next get)
+
+    We implement unget by setting variable next_unget to true. The input is not
+    changed - we just simulate ungetting by modifying chars_read_total,
+    chars_read_current_line, and token_string. The next call to get() will
+    behave as if the unget character is read again.
+    */
+    void unget()
+    {
+        next_unget = true;
+
+        --position.chars_read_total;
+
+        // in case we "unget" a newline, we have to also decrement the lines_read
+        if (position.chars_read_current_line == 0)
+        {
+            if (position.lines_read > 0)
+            {
+                --position.lines_read;
+            }
+        }
+        else
+        {
+            --position.chars_read_current_line;
+        }
+
+        if (JSON_HEDLEY_LIKELY(current != char_traits<char_type>::eof()))
+        {
+            JSON_ASSERT(!token_string.empty());
+            token_string.pop_back();
+        }
+    }
+
+    /// add a character to token_buffer
+    void add(char_int_type c)
+    {
+        token_buffer.push_back(static_cast<typename string_t::value_type>(c));
+    }
+
+  public:
+    /////////////////////
+    // value getters
+    /////////////////////
+
+    /// return integer value
+    constexpr number_integer_t get_number_integer() const noexcept
+    {
+        return value_integer;
+    }
+
+    /// return unsigned integer value
+    constexpr number_unsigned_t get_number_unsigned() const noexcept
+    {
+        return value_unsigned;
+    }
+
+    /// return floating-point value
+    constexpr number_float_t get_number_float() const noexcept
+    {
+        return value_float;
+    }
+
+    /// return current string value (implicitly resets the token; useful only once)
+    string_t& get_string()
+    {
+        return token_buffer;
+    }
+
+    /////////////////////
+    // diagnostics
+    /////////////////////
+
+    /// return position of last read token
+    constexpr position_t get_position() const noexcept
+    {
+        return position;
+    }
+
+    /// return the last read token (for errors only).  Will never contain EOF
+    /// (an arbitrary value that is not a valid char value, often -1), because
+    /// 255 may legitimately occur.  May contain NUL, which should be escaped.
+    std::string get_token_string() const
+    {
+        // escape control characters
+        std::string result;
+        for (const auto c : token_string)
+        {
+            if (static_cast<unsigned char>(c) <= '\x1F')
+            {
+                // escape control characters
+                std::array<char, 9> cs{{}};
+                static_cast<void>((std::snprintf)(cs.data(), cs.size(), "<U+%.4X>", static_cast<unsigned char>(c))); // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+                result += cs.data();
+            }
+            else
+            {
+                // add character as is
+                result.push_back(static_cast<std::string::value_type>(c));
+            }
+        }
+
+        return result;
+    }
+
+    /// return syntax error message
+    JSON_HEDLEY_RETURNS_NON_NULL
+    constexpr const char* get_error_message() const noexcept
+    {
+        return error_message;
+    }
+
+    /////////////////////
+    // actual scanner
+    /////////////////////
+
+    /*!
+    @brief skip the UTF-8 byte order mark
+    @return true iff there is no BOM or the correct BOM has been skipped
+    */
+    bool skip_bom()
+    {
+        if (get() == 0xEF)
+        {
+            // check if we completely parse the BOM
+            return get() == 0xBB && get() == 0xBF;
+        }
+
+        // the first character is not the beginning of the BOM; unget it to
+        // process is later
+        unget();
+        return true;
+    }
+
+    void skip_whitespace()
+    {
+        do
+        {
+            get();
+        }
+        while (current == ' ' || current == '\t' || current == '\n' || current == '\r');
+    }
+
+    token_type scan()
+    {
+        // initially, skip the BOM
+        if (position.chars_read_total == 0 && !skip_bom())
+        {
+            error_message = "invalid BOM; must be 0xEF 0xBB 0xBF if given";
+            return token_type::parse_error;
+        }
+
+        // read next character and ignore whitespace
+        skip_whitespace();
+
+        // ignore comments
+        while (ignore_comments && current == '/')
+        {
+            if (!scan_comment())
+            {
+                return token_type::parse_error;
+            }
+
+            // skip following whitespace
+            skip_whitespace();
+        }
+
+        switch (current)
+        {
+            // structural characters
+            case '[':
+                return token_type::begin_array;
+            case ']':
+                return token_type::end_array;
+            case '{':
+                return token_type::begin_object;
+            case '}':
+                return token_type::end_object;
+            case ':':
+                return token_type::name_separator;
+            case ',':
+                return token_type::value_separator;
+
+            // literals
+            case 't':
+            {
+                std::array<char_type, 4> true_literal = {{static_cast<char_type>('t'), static_cast<char_type>('r'), static_cast<char_type>('u'), static_cast<char_type>('e')}};
+                return scan_literal(true_literal.data(), true_literal.size(), token_type::literal_true);
+            }
+            case 'f':
+            {
+                std::array<char_type, 5> false_literal = {{static_cast<char_type>('f'), static_cast<char_type>('a'), static_cast<char_type>('l'), static_cast<char_type>('s'), static_cast<char_type>('e')}};
+                return scan_literal(false_literal.data(), false_literal.size(), token_type::literal_false);
+            }
+            case 'n':
+            {
+                std::array<char_type, 4> null_literal = {{static_cast<char_type>('n'), static_cast<char_type>('u'), static_cast<char_type>('l'), static_cast<char_type>('l')}};
+                return scan_literal(null_literal.data(), null_literal.size(), token_type::literal_null);
+            }
+
+            // string
+            case '\"':
+                return scan_string();
+
+            // number
+            case '-':
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+                return scan_number();
+
+            // end of input (the null byte is needed when parsing from
+            // string literals)
+            case '\0':
+            case char_traits<char_type>::eof():
+                return token_type::end_of_input;
+
+            // error
+            default:
+                error_message = "invalid literal";
+                return token_type::parse_error;
+        }
+    }
+
+  private:
+    /// input adapter
+    InputAdapterType ia;
+
+    /// whether comments should be ignored (true) or signaled as errors (false)
+    const bool ignore_comments = false;
+
+    /// the current character
+    char_int_type current = char_traits<char_type>::eof();
+
+    /// whether the next get() call should just return current
+    bool next_unget = false;
+
+    /// the start position of the current token
+    position_t position {};
+
+    /// raw input token string (for error messages)
+    std::vector<char_type> token_string {};
+
+    /// buffer for variable-length tokens (numbers, strings)
+    string_t token_buffer {};
+
+    /// a description of occurred lexer errors
+    const char* error_message = "";
+
+    // number values
+    number_integer_t value_integer = 0;
+    number_unsigned_t value_unsigned = 0;
+    number_float_t value_float = 0;
+
+    /// the decimal point
+    const char_int_type decimal_point_char = '.';
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/is_sax.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstdint> // size_t
+#include <utility> // declval
+#include <string> // string
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+// #include <nlohmann/detail/meta/detected.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+template<typename T>
+using null_function_t = decltype(std::declval<T&>().null());
+
+template<typename T>
+using boolean_function_t =
+    decltype(std::declval<T&>().boolean(std::declval<bool>()));
+
+template<typename T, typename Integer>
+using number_integer_function_t =
+    decltype(std::declval<T&>().number_integer(std::declval<Integer>()));
+
+template<typename T, typename Unsigned>
+using number_unsigned_function_t =
+    decltype(std::declval<T&>().number_unsigned(std::declval<Unsigned>()));
+
+template<typename T, typename Float, typename String>
+using number_float_function_t = decltype(std::declval<T&>().number_float(
+                                    std::declval<Float>(), std::declval<const String&>()));
+
+template<typename T, typename String>
+using string_function_t =
+    decltype(std::declval<T&>().string(std::declval<String&>()));
+
+template<typename T, typename Binary>
+using binary_function_t =
+    decltype(std::declval<T&>().binary(std::declval<Binary&>()));
+
+template<typename T>
+using start_object_function_t =
+    decltype(std::declval<T&>().start_object(std::declval<std::size_t>()));
+
+template<typename T, typename String>
+using key_function_t =
+    decltype(std::declval<T&>().key(std::declval<String&>()));
+
+template<typename T>
+using end_object_function_t = decltype(std::declval<T&>().end_object());
+
+template<typename T>
+using start_array_function_t =
+    decltype(std::declval<T&>().start_array(std::declval<std::size_t>()));
+
+template<typename T>
+using end_array_function_t = decltype(std::declval<T&>().end_array());
+
+template<typename T, typename Exception>
+using parse_error_function_t = decltype(std::declval<T&>().parse_error(
+        std::declval<std::size_t>(), std::declval<const std::string&>(),
+        std::declval<const Exception&>()));
+
+template<typename SAX, typename BasicJsonType>
+struct is_sax
+{
+  private:
+    static_assert(is_basic_json<BasicJsonType>::value,
+                  "BasicJsonType must be of type basic_json<...>");
+
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+    using exception_t = typename BasicJsonType::exception;
+
+  public:
+    static constexpr bool value =
+        is_detected_exact<bool, null_function_t, SAX>::value &&
+        is_detected_exact<bool, boolean_function_t, SAX>::value &&
+        is_detected_exact<bool, number_integer_function_t, SAX, number_integer_t>::value &&
+        is_detected_exact<bool, number_unsigned_function_t, SAX, number_unsigned_t>::value &&
+        is_detected_exact<bool, number_float_function_t, SAX, number_float_t, string_t>::value &&
+        is_detected_exact<bool, string_function_t, SAX, string_t>::value &&
+        is_detected_exact<bool, binary_function_t, SAX, binary_t>::value &&
+        is_detected_exact<bool, start_object_function_t, SAX>::value &&
+        is_detected_exact<bool, key_function_t, SAX, string_t>::value &&
+        is_detected_exact<bool, end_object_function_t, SAX>::value &&
+        is_detected_exact<bool, start_array_function_t, SAX>::value &&
+        is_detected_exact<bool, end_array_function_t, SAX>::value &&
+        is_detected_exact<bool, parse_error_function_t, SAX, exception_t>::value;
+};
+
+template<typename SAX, typename BasicJsonType>
+struct is_sax_static_asserts
+{
+  private:
+    static_assert(is_basic_json<BasicJsonType>::value,
+                  "BasicJsonType must be of type basic_json<...>");
+
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+    using exception_t = typename BasicJsonType::exception;
+
+  public:
+    static_assert(is_detected_exact<bool, null_function_t, SAX>::value,
+                  "Missing/invalid function: bool null()");
+    static_assert(is_detected_exact<bool, boolean_function_t, SAX>::value,
+                  "Missing/invalid function: bool boolean(bool)");
+    static_assert(is_detected_exact<bool, boolean_function_t, SAX>::value,
+                  "Missing/invalid function: bool boolean(bool)");
+    static_assert(
+        is_detected_exact<bool, number_integer_function_t, SAX,
+        number_integer_t>::value,
+        "Missing/invalid function: bool number_integer(number_integer_t)");
+    static_assert(
+        is_detected_exact<bool, number_unsigned_function_t, SAX,
+        number_unsigned_t>::value,
+        "Missing/invalid function: bool number_unsigned(number_unsigned_t)");
+    static_assert(is_detected_exact<bool, number_float_function_t, SAX,
+                  number_float_t, string_t>::value,
+                  "Missing/invalid function: bool number_float(number_float_t, const string_t&)");
+    static_assert(
+        is_detected_exact<bool, string_function_t, SAX, string_t>::value,
+        "Missing/invalid function: bool string(string_t&)");
+    static_assert(
+        is_detected_exact<bool, binary_function_t, SAX, binary_t>::value,
+        "Missing/invalid function: bool binary(binary_t&)");
+    static_assert(is_detected_exact<bool, start_object_function_t, SAX>::value,
+                  "Missing/invalid function: bool start_object(std::size_t)");
+    static_assert(is_detected_exact<bool, key_function_t, SAX, string_t>::value,
+                  "Missing/invalid function: bool key(string_t&)");
+    static_assert(is_detected_exact<bool, end_object_function_t, SAX>::value,
+                  "Missing/invalid function: bool end_object()");
+    static_assert(is_detected_exact<bool, start_array_function_t, SAX>::value,
+                  "Missing/invalid function: bool start_array(std::size_t)");
+    static_assert(is_detected_exact<bool, end_array_function_t, SAX>::value,
+                  "Missing/invalid function: bool end_array()");
+    static_assert(
+        is_detected_exact<bool, parse_error_function_t, SAX, exception_t>::value,
+        "Missing/invalid function: bool parse_error(std::size_t, const "
+        "std::string&, const exception&)");
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/// how to treat CBOR tags
+enum class cbor_tag_handler_t
+{
+    error,   ///< throw a parse_error exception in case of a tag
+    ignore,  ///< ignore tags
+    store    ///< store tags as binary type
+};
+
+/*!
+@brief determine system byte order
+
+@return true if and only if system's byte order is little endian
+
+@note from https://stackoverflow.com/a/1001328/266378
+*/
+static inline bool little_endianness(int num = 1) noexcept
+{
+    return *reinterpret_cast<char*>(&num) == 1;
+}
+
+///////////////////
+// binary reader //
+///////////////////
+
+/*!
+@brief deserialization of CBOR, MessagePack, and UBJSON values
+*/
+template<typename BasicJsonType, typename InputAdapterType, typename SAX = json_sax_dom_parser<BasicJsonType>>
+class binary_reader
+{
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+    using json_sax_t = SAX;
+    using char_type = typename InputAdapterType::char_type;
+    using char_int_type = typename char_traits<char_type>::int_type;
+
+  public:
+    /*!
+    @brief create a binary reader
+
+    @param[in] adapter  input adapter to read from
+    */
+    explicit binary_reader(InputAdapterType&& adapter, const input_format_t format = input_format_t::json) noexcept : ia(std::move(adapter)), input_format(format)
+    {
+        (void)detail::is_sax_static_asserts<SAX, BasicJsonType> {};
+    }
+
+    // make class move-only
+    binary_reader(const binary_reader&) = delete;
+    binary_reader(binary_reader&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    binary_reader& operator=(const binary_reader&) = delete;
+    binary_reader& operator=(binary_reader&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    ~binary_reader() = default;
+
+    /*!
+    @param[in] format  the binary format to parse
+    @param[in] sax_    a SAX event processor
+    @param[in] strict  whether to expect the input to be consumed completed
+    @param[in] tag_handler  how to treat CBOR tags
+
+    @return whether parsing was successful
+    */
+    JSON_HEDLEY_NON_NULL(3)
+    bool sax_parse(const input_format_t format,
+                   json_sax_t* sax_,
+                   const bool strict = true,
+                   const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
+    {
+        sax = sax_;
+        bool result = false;
+
+        switch (format)
+        {
+            case input_format_t::bson:
+                result = parse_bson_internal();
+                break;
+
+            case input_format_t::cbor:
+                result = parse_cbor_internal(true, tag_handler);
+                break;
+
+            case input_format_t::msgpack:
+                result = parse_msgpack_internal();
+                break;
+
+            case input_format_t::ubjson:
+            case input_format_t::bjdata:
+                result = parse_ubjson_internal();
+                break;
+
+            case input_format_t::json: // LCOV_EXCL_LINE
+            default:            // LCOV_EXCL_LINE
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        }
+
+        // strict mode: next byte must be EOF
+        if (result && strict)
+        {
+            if (input_format == input_format_t::ubjson || input_format == input_format_t::bjdata)
+            {
+                get_ignore_noop();
+            }
+            else
+            {
+                get();
+            }
+
+            if (JSON_HEDLEY_UNLIKELY(current != char_traits<char_type>::eof()))
+            {
+                return sax->parse_error(chars_read, get_token_string(), parse_error::create(110, chars_read,
+                                        exception_message(input_format, concat("expected end of input; last byte: 0x", get_token_string()), "value"), nullptr));
+            }
+        }
+
+        return result;
+    }
+
+  private:
+    //////////
+    // BSON //
+    //////////
+
+    /*!
+    @brief Reads in a BSON-object and passes it to the SAX-parser.
+    @return whether a valid BSON-value was passed to the SAX parser
+    */
+    bool parse_bson_internal()
+    {
+        std::int32_t document_size{};
+        get_number<std::int32_t, true>(input_format_t::bson, document_size);
+
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(static_cast<std::size_t>(-1))))
+        {
+            return false;
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(!parse_bson_element_list(/*is_array*/false)))
+        {
+            return false;
+        }
+
+        return sax->end_object();
+    }
+
+    /*!
+    @brief Parses a C-style string from the BSON input.
+    @param[in,out] result  A reference to the string variable where the read
+                            string is to be stored.
+    @return `true` if the \x00-byte indicating the end of the string was
+             encountered before the EOF; false` indicates an unexpected EOF.
+    */
+    bool get_bson_cstr(string_t& result)
+    {
+        auto out = std::back_inserter(result);
+        while (true)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::bson, "cstring")))
+            {
+                return false;
+            }
+            if (current == 0x00)
+            {
+                return true;
+            }
+            *out++ = static_cast<typename string_t::value_type>(current);
+        }
+    }
+
+    /*!
+    @brief Parses a zero-terminated string of length @a len from the BSON
+           input.
+    @param[in] len  The length (including the zero-byte at the end) of the
+                    string to be read.
+    @param[in,out] result  A reference to the string variable where the read
+                            string is to be stored.
+    @tparam NumberType The type of the length @a len
+    @pre len >= 1
+    @return `true` if the string was successfully parsed
+    */
+    template<typename NumberType>
+    bool get_bson_string(const NumberType len, string_t& result)
+    {
+        if (JSON_HEDLEY_UNLIKELY(len < 1))
+        {
+            auto last_token = get_token_string();
+            return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                    exception_message(input_format_t::bson, concat("string length must be at least 1, is ", std::to_string(len)), "string"), nullptr));
+        }
+
+        return get_string(input_format_t::bson, len - static_cast<NumberType>(1), result) && get() != char_traits<char_type>::eof();
+    }
+
+    /*!
+    @brief Parses a byte array input of length @a len from the BSON input.
+    @param[in] len  The length of the byte array to be read.
+    @param[in,out] result  A reference to the binary variable where the read
+                            array is to be stored.
+    @tparam NumberType The type of the length @a len
+    @pre len >= 0
+    @return `true` if the byte array was successfully parsed
+    */
+    template<typename NumberType>
+    bool get_bson_binary(const NumberType len, binary_t& result)
+    {
+        if (JSON_HEDLEY_UNLIKELY(len < 0))
+        {
+            auto last_token = get_token_string();
+            return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                    exception_message(input_format_t::bson, concat("byte array length cannot be negative, is ", std::to_string(len)), "binary"), nullptr));
+        }
+
+        // All BSON binary values have a subtype
+        std::uint8_t subtype{};
+        get_number<std::uint8_t>(input_format_t::bson, subtype);
+        result.set_subtype(subtype);
+
+        return get_binary(input_format_t::bson, len, result);
+    }
+
+    /*!
+    @brief Read a BSON document element of the given @a element_type.
+    @param[in] element_type The BSON element type, c.f. http://bsonspec.org/spec.html
+    @param[in] element_type_parse_position The position in the input stream,
+               where the `element_type` was read.
+    @warning Not all BSON element types are supported yet. An unsupported
+             @a element_type will give rise to a parse_error.114:
+             Unsupported BSON record type 0x...
+    @return whether a valid BSON-object/array was passed to the SAX parser
+    */
+    bool parse_bson_element_internal(const char_int_type element_type,
+                                     const std::size_t element_type_parse_position)
+    {
+        switch (element_type)
+        {
+            case 0x01: // double
+            {
+                double number{};
+                return get_number<double, true>(input_format_t::bson, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 0x02: // string
+            {
+                std::int32_t len{};
+                string_t value;
+                return get_number<std::int32_t, true>(input_format_t::bson, len) && get_bson_string(len, value) && sax->string(value);
+            }
+
+            case 0x03: // object
+            {
+                return parse_bson_internal();
+            }
+
+            case 0x04: // array
+            {
+                return parse_bson_array();
+            }
+
+            case 0x05: // binary
+            {
+                std::int32_t len{};
+                binary_t value;
+                return get_number<std::int32_t, true>(input_format_t::bson, len) && get_bson_binary(len, value) && sax->binary(value);
+            }
+
+            case 0x08: // boolean
+            {
+                return sax->boolean(get() != 0);
+            }
+
+            case 0x0A: // null
+            {
+                return sax->null();
+            }
+
+            case 0x10: // int32
+            {
+                std::int32_t value{};
+                return get_number<std::int32_t, true>(input_format_t::bson, value) && sax->number_integer(value);
+            }
+
+            case 0x12: // int64
+            {
+                std::int64_t value{};
+                return get_number<std::int64_t, true>(input_format_t::bson, value) && sax->number_integer(value);
+            }
+
+            default: // anything else not supported (yet)
+            {
+                std::array<char, 3> cr{{}};
+                static_cast<void>((std::snprintf)(cr.data(), cr.size(), "%.2hhX", static_cast<unsigned char>(element_type))); // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+                const std::string cr_str{cr.data()};
+                return sax->parse_error(element_type_parse_position, cr_str,
+                                        parse_error::create(114, element_type_parse_position, concat("Unsupported BSON record type 0x", cr_str), nullptr));
+            }
+        }
+    }
+
+    /*!
+    @brief Read a BSON element list (as specified in the BSON-spec)
+
+    The same binary layout is used for objects and arrays, hence it must be
+    indicated with the argument @a is_array which one is expected
+    (true --> array, false --> object).
+
+    @param[in] is_array Determines if the element list being read is to be
+                        treated as an object (@a is_array == false), or as an
+                        array (@a is_array == true).
+    @return whether a valid BSON-object/array was passed to the SAX parser
+    */
+    bool parse_bson_element_list(const bool is_array)
+    {
+        string_t key;
+
+        while (auto element_type = get())
+        {
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::bson, "element list")))
+            {
+                return false;
+            }
+
+            const std::size_t element_type_parse_position = chars_read;
+            if (JSON_HEDLEY_UNLIKELY(!get_bson_cstr(key)))
+            {
+                return false;
+            }
+
+            if (!is_array && !sax->key(key))
+            {
+                return false;
+            }
+
+            if (JSON_HEDLEY_UNLIKELY(!parse_bson_element_internal(element_type, element_type_parse_position)))
+            {
+                return false;
+            }
+
+            // get_bson_cstr only appends
+            key.clear();
+        }
+
+        return true;
+    }
+
+    /*!
+    @brief Reads an array from the BSON input and passes it to the SAX-parser.
+    @return whether a valid BSON-array was passed to the SAX parser
+    */
+    bool parse_bson_array()
+    {
+        std::int32_t document_size{};
+        get_number<std::int32_t, true>(input_format_t::bson, document_size);
+
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(static_cast<std::size_t>(-1))))
+        {
+            return false;
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(!parse_bson_element_list(/*is_array*/true)))
+        {
+            return false;
+        }
+
+        return sax->end_array();
+    }
+
+    //////////
+    // CBOR //
+    //////////
+
+    /*!
+    @param[in] get_char  whether a new character should be retrieved from the
+                         input (true) or whether the last read character should
+                         be considered instead (false)
+    @param[in] tag_handler how CBOR tags should be treated
+
+    @return whether a valid CBOR value was passed to the SAX parser
+    */
+    bool parse_cbor_internal(const bool get_char,
+                             const cbor_tag_handler_t tag_handler)
+    {
+        switch (get_char ? get() : current)
+        {
+            // EOF
+            case char_traits<char_type>::eof():
+                return unexpect_eof(input_format_t::cbor, "value");
+
+            // Integer 0x00..0x17 (0..23)
+            case 0x00:
+            case 0x01:
+            case 0x02:
+            case 0x03:
+            case 0x04:
+            case 0x05:
+            case 0x06:
+            case 0x07:
+            case 0x08:
+            case 0x09:
+            case 0x0A:
+            case 0x0B:
+            case 0x0C:
+            case 0x0D:
+            case 0x0E:
+            case 0x0F:
+            case 0x10:
+            case 0x11:
+            case 0x12:
+            case 0x13:
+            case 0x14:
+            case 0x15:
+            case 0x16:
+            case 0x17:
+                return sax->number_unsigned(static_cast<number_unsigned_t>(current));
+
+            case 0x18: // Unsigned integer (one-byte uint8_t follows)
+            {
+                std::uint8_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
+            }
+
+            case 0x19: // Unsigned integer (two-byte uint16_t follows)
+            {
+                std::uint16_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
+            }
+
+            case 0x1A: // Unsigned integer (four-byte uint32_t follows)
+            {
+                std::uint32_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
+            }
+
+            case 0x1B: // Unsigned integer (eight-byte uint64_t follows)
+            {
+                std::uint64_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
+            }
+
+            // Negative integer -1-0x00..-1-0x17 (-1..-24)
+            case 0x20:
+            case 0x21:
+            case 0x22:
+            case 0x23:
+            case 0x24:
+            case 0x25:
+            case 0x26:
+            case 0x27:
+            case 0x28:
+            case 0x29:
+            case 0x2A:
+            case 0x2B:
+            case 0x2C:
+            case 0x2D:
+            case 0x2E:
+            case 0x2F:
+            case 0x30:
+            case 0x31:
+            case 0x32:
+            case 0x33:
+            case 0x34:
+            case 0x35:
+            case 0x36:
+            case 0x37:
+                return sax->number_integer(static_cast<std::int8_t>(0x20 - 1 - current));
+
+            case 0x38: // Negative integer (one-byte uint8_t follows)
+            {
+                std::uint8_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1) - number);
+            }
+
+            case 0x39: // Negative integer -1-n (two-byte uint16_t follows)
+            {
+                std::uint16_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1) - number);
+            }
+
+            case 0x3A: // Negative integer -1-n (four-byte uint32_t follows)
+            {
+                std::uint32_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1) - number);
+            }
+
+            case 0x3B: // Negative integer -1-n (eight-byte uint64_t follows)
+            {
+                std::uint64_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1)
+                        - static_cast<number_integer_t>(number));
+            }
+
+            // Binary data (0x00..0x17 bytes follow)
+            case 0x40:
+            case 0x41:
+            case 0x42:
+            case 0x43:
+            case 0x44:
+            case 0x45:
+            case 0x46:
+            case 0x47:
+            case 0x48:
+            case 0x49:
+            case 0x4A:
+            case 0x4B:
+            case 0x4C:
+            case 0x4D:
+            case 0x4E:
+            case 0x4F:
+            case 0x50:
+            case 0x51:
+            case 0x52:
+            case 0x53:
+            case 0x54:
+            case 0x55:
+            case 0x56:
+            case 0x57:
+            case 0x58: // Binary data (one-byte uint8_t for n follows)
+            case 0x59: // Binary data (two-byte uint16_t for n follow)
+            case 0x5A: // Binary data (four-byte uint32_t for n follow)
+            case 0x5B: // Binary data (eight-byte uint64_t for n follow)
+            case 0x5F: // Binary data (indefinite length)
+            {
+                binary_t b;
+                return get_cbor_binary(b) && sax->binary(b);
+            }
+
+            // UTF-8 string (0x00..0x17 bytes follow)
+            case 0x60:
+            case 0x61:
+            case 0x62:
+            case 0x63:
+            case 0x64:
+            case 0x65:
+            case 0x66:
+            case 0x67:
+            case 0x68:
+            case 0x69:
+            case 0x6A:
+            case 0x6B:
+            case 0x6C:
+            case 0x6D:
+            case 0x6E:
+            case 0x6F:
+            case 0x70:
+            case 0x71:
+            case 0x72:
+            case 0x73:
+            case 0x74:
+            case 0x75:
+            case 0x76:
+            case 0x77:
+            case 0x78: // UTF-8 string (one-byte uint8_t for n follows)
+            case 0x79: // UTF-8 string (two-byte uint16_t for n follow)
+            case 0x7A: // UTF-8 string (four-byte uint32_t for n follow)
+            case 0x7B: // UTF-8 string (eight-byte uint64_t for n follow)
+            case 0x7F: // UTF-8 string (indefinite length)
+            {
+                string_t s;
+                return get_cbor_string(s) && sax->string(s);
+            }
+
+            // array (0x00..0x17 data items follow)
+            case 0x80:
+            case 0x81:
+            case 0x82:
+            case 0x83:
+            case 0x84:
+            case 0x85:
+            case 0x86:
+            case 0x87:
+            case 0x88:
+            case 0x89:
+            case 0x8A:
+            case 0x8B:
+            case 0x8C:
+            case 0x8D:
+            case 0x8E:
+            case 0x8F:
+            case 0x90:
+            case 0x91:
+            case 0x92:
+            case 0x93:
+            case 0x94:
+            case 0x95:
+            case 0x96:
+            case 0x97:
+                return get_cbor_array(
+                           conditional_static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x1Fu), tag_handler);
+
+            case 0x98: // array (one-byte uint8_t for n follows)
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_array(static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0x99: // array (two-byte uint16_t for n follow)
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_array(static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0x9A: // array (four-byte uint32_t for n follow)
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_array(conditional_static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0x9B: // array (eight-byte uint64_t for n follow)
+            {
+                std::uint64_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_array(conditional_static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0x9F: // array (indefinite length)
+                return get_cbor_array(static_cast<std::size_t>(-1), tag_handler);
+
+            // map (0x00..0x17 pairs of data items follow)
+            case 0xA0:
+            case 0xA1:
+            case 0xA2:
+            case 0xA3:
+            case 0xA4:
+            case 0xA5:
+            case 0xA6:
+            case 0xA7:
+            case 0xA8:
+            case 0xA9:
+            case 0xAA:
+            case 0xAB:
+            case 0xAC:
+            case 0xAD:
+            case 0xAE:
+            case 0xAF:
+            case 0xB0:
+            case 0xB1:
+            case 0xB2:
+            case 0xB3:
+            case 0xB4:
+            case 0xB5:
+            case 0xB6:
+            case 0xB7:
+                return get_cbor_object(conditional_static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x1Fu), tag_handler);
+
+            case 0xB8: // map (one-byte uint8_t for n follows)
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_object(static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0xB9: // map (two-byte uint16_t for n follow)
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_object(static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0xBA: // map (four-byte uint32_t for n follow)
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_object(conditional_static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0xBB: // map (eight-byte uint64_t for n follow)
+            {
+                std::uint64_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_object(conditional_static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0xBF: // map (indefinite length)
+                return get_cbor_object(static_cast<std::size_t>(-1), tag_handler);
+
+            case 0xC6: // tagged item
+            case 0xC7:
+            case 0xC8:
+            case 0xC9:
+            case 0xCA:
+            case 0xCB:
+            case 0xCC:
+            case 0xCD:
+            case 0xCE:
+            case 0xCF:
+            case 0xD0:
+            case 0xD1:
+            case 0xD2:
+            case 0xD3:
+            case 0xD4:
+            case 0xD8: // tagged item (1 bytes follow)
+            case 0xD9: // tagged item (2 bytes follow)
+            case 0xDA: // tagged item (4 bytes follow)
+            case 0xDB: // tagged item (8 bytes follow)
+            {
+                switch (tag_handler)
+                {
+                    case cbor_tag_handler_t::error:
+                    {
+                        auto last_token = get_token_string();
+                        return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                                exception_message(input_format_t::cbor, concat("invalid byte: 0x", last_token), "value"), nullptr));
+                    }
+
+                    case cbor_tag_handler_t::ignore:
+                    {
+                        // ignore binary subtype
+                        switch (current)
+                        {
+                            case 0xD8:
+                            {
+                                std::uint8_t subtype_to_ignore{};
+                                get_number(input_format_t::cbor, subtype_to_ignore);
+                                break;
+                            }
+                            case 0xD9:
+                            {
+                                std::uint16_t subtype_to_ignore{};
+                                get_number(input_format_t::cbor, subtype_to_ignore);
+                                break;
+                            }
+                            case 0xDA:
+                            {
+                                std::uint32_t subtype_to_ignore{};
+                                get_number(input_format_t::cbor, subtype_to_ignore);
+                                break;
+                            }
+                            case 0xDB:
+                            {
+                                std::uint64_t subtype_to_ignore{};
+                                get_number(input_format_t::cbor, subtype_to_ignore);
+                                break;
+                            }
+                            default:
+                                break;
+                        }
+                        return parse_cbor_internal(true, tag_handler);
+                    }
+
+                    case cbor_tag_handler_t::store:
+                    {
+                        binary_t b;
+                        // use binary subtype and store in binary container
+                        switch (current)
+                        {
+                            case 0xD8:
+                            {
+                                std::uint8_t subtype{};
+                                get_number(input_format_t::cbor, subtype);
+                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
+                                break;
+                            }
+                            case 0xD9:
+                            {
+                                std::uint16_t subtype{};
+                                get_number(input_format_t::cbor, subtype);
+                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
+                                break;
+                            }
+                            case 0xDA:
+                            {
+                                std::uint32_t subtype{};
+                                get_number(input_format_t::cbor, subtype);
+                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
+                                break;
+                            }
+                            case 0xDB:
+                            {
+                                std::uint64_t subtype{};
+                                get_number(input_format_t::cbor, subtype);
+                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
+                                break;
+                            }
+                            default:
+                                return parse_cbor_internal(true, tag_handler);
+                        }
+                        get();
+                        return get_cbor_binary(b) && sax->binary(b);
+                    }
+
+                    default:                 // LCOV_EXCL_LINE
+                        JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+                        return false;        // LCOV_EXCL_LINE
+                }
+            }
+
+            case 0xF4: // false
+                return sax->boolean(false);
+
+            case 0xF5: // true
+                return sax->boolean(true);
+
+            case 0xF6: // null
+                return sax->null();
+
+            case 0xF9: // Half-Precision Float (two-byte IEEE 754)
+            {
+                const auto byte1_raw = get();
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "number")))
+                {
+                    return false;
+                }
+                const auto byte2_raw = get();
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "number")))
+                {
+                    return false;
+                }
+
+                const auto byte1 = static_cast<unsigned char>(byte1_raw);
+                const auto byte2 = static_cast<unsigned char>(byte2_raw);
+
+                // code from RFC 7049, Appendix D, Figure 3:
+                // As half-precision floating-point numbers were only added
+                // to IEEE 754 in 2008, today's programming platforms often
+                // still only have limited support for them. It is very
+                // easy to include at least decoding support for them even
+                // without such support. An example of a small decoder for
+                // half-precision floating-point numbers in the C language
+                // is shown in Fig. 3.
+                const auto half = static_cast<unsigned int>((byte1 << 8u) + byte2);
+                const double val = [&half]
+                {
+                    const int exp = (half >> 10u) & 0x1Fu;
+                    const unsigned int mant = half & 0x3FFu;
+                    JSON_ASSERT(0 <= exp&& exp <= 32);
+                    JSON_ASSERT(mant <= 1024);
+                    switch (exp)
+                    {
+                        case 0:
+                            return std::ldexp(mant, -24);
+                        case 31:
+                            return (mant == 0)
+                            ? std::numeric_limits<double>::infinity()
+                            : std::numeric_limits<double>::quiet_NaN();
+                        default:
+                            return std::ldexp(mant + 1024, exp - 25);
+                    }
+                }();
+                return sax->number_float((half & 0x8000u) != 0
+                                         ? static_cast<number_float_t>(-val)
+                                         : static_cast<number_float_t>(val), "");
+            }
+
+            case 0xFA: // Single-Precision Float (four-byte IEEE 754)
+            {
+                float number{};
+                return get_number(input_format_t::cbor, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 0xFB: // Double-Precision Float (eight-byte IEEE 754)
+            {
+                double number{};
+                return get_number(input_format_t::cbor, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            default: // anything else (0xFF is handled inside the other types)
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                        exception_message(input_format_t::cbor, concat("invalid byte: 0x", last_token), "value"), nullptr));
+            }
+        }
+    }
+
+    /*!
+    @brief reads a CBOR string
+
+    This function first reads starting bytes to determine the expected
+    string length and then copies this number of bytes into a string.
+    Additionally, CBOR's strings with indefinite lengths are supported.
+
+    @param[out] result  created string
+
+    @return whether string creation completed
+    */
+    bool get_cbor_string(string_t& result)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "string")))
+        {
+            return false;
+        }
+
+        switch (current)
+        {
+            // UTF-8 string (0x00..0x17 bytes follow)
+            case 0x60:
+            case 0x61:
+            case 0x62:
+            case 0x63:
+            case 0x64:
+            case 0x65:
+            case 0x66:
+            case 0x67:
+            case 0x68:
+            case 0x69:
+            case 0x6A:
+            case 0x6B:
+            case 0x6C:
+            case 0x6D:
+            case 0x6E:
+            case 0x6F:
+            case 0x70:
+            case 0x71:
+            case 0x72:
+            case 0x73:
+            case 0x74:
+            case 0x75:
+            case 0x76:
+            case 0x77:
+            {
+                return get_string(input_format_t::cbor, static_cast<unsigned int>(current) & 0x1Fu, result);
+            }
+
+            case 0x78: // UTF-8 string (one-byte uint8_t for n follows)
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
+            }
+
+            case 0x79: // UTF-8 string (two-byte uint16_t for n follow)
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
+            }
+
+            case 0x7A: // UTF-8 string (four-byte uint32_t for n follow)
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
+            }
+
+            case 0x7B: // UTF-8 string (eight-byte uint64_t for n follow)
+            {
+                std::uint64_t len{};
+                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
+            }
+
+            case 0x7F: // UTF-8 string (indefinite length)
+            {
+                while (get() != 0xFF)
+                {
+                    string_t chunk;
+                    if (!get_cbor_string(chunk))
+                    {
+                        return false;
+                    }
+                    result.append(chunk);
+                }
+                return true;
+            }
+
+            default:
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read,
+                                        exception_message(input_format_t::cbor, concat("expected length specification (0x60-0x7B) or indefinite string type (0x7F); last byte: 0x", last_token), "string"), nullptr));
+            }
+        }
+    }
+
+    /*!
+    @brief reads a CBOR byte array
+
+    This function first reads starting bytes to determine the expected
+    byte array length and then copies this number of bytes into the byte array.
+    Additionally, CBOR's byte arrays with indefinite lengths are supported.
+
+    @param[out] result  created byte array
+
+    @return whether byte array creation completed
+    */
+    bool get_cbor_binary(binary_t& result)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "binary")))
+        {
+            return false;
+        }
+
+        switch (current)
+        {
+            // Binary data (0x00..0x17 bytes follow)
+            case 0x40:
+            case 0x41:
+            case 0x42:
+            case 0x43:
+            case 0x44:
+            case 0x45:
+            case 0x46:
+            case 0x47:
+            case 0x48:
+            case 0x49:
+            case 0x4A:
+            case 0x4B:
+            case 0x4C:
+            case 0x4D:
+            case 0x4E:
+            case 0x4F:
+            case 0x50:
+            case 0x51:
+            case 0x52:
+            case 0x53:
+            case 0x54:
+            case 0x55:
+            case 0x56:
+            case 0x57:
+            {
+                return get_binary(input_format_t::cbor, static_cast<unsigned int>(current) & 0x1Fu, result);
+            }
+
+            case 0x58: // Binary data (one-byte uint8_t for n follows)
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::cbor, len) &&
+                       get_binary(input_format_t::cbor, len, result);
+            }
+
+            case 0x59: // Binary data (two-byte uint16_t for n follow)
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::cbor, len) &&
+                       get_binary(input_format_t::cbor, len, result);
+            }
+
+            case 0x5A: // Binary data (four-byte uint32_t for n follow)
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::cbor, len) &&
+                       get_binary(input_format_t::cbor, len, result);
+            }
+
+            case 0x5B: // Binary data (eight-byte uint64_t for n follow)
+            {
+                std::uint64_t len{};
+                return get_number(input_format_t::cbor, len) &&
+                       get_binary(input_format_t::cbor, len, result);
+            }
+
+            case 0x5F: // Binary data (indefinite length)
+            {
+                while (get() != 0xFF)
+                {
+                    binary_t chunk;
+                    if (!get_cbor_binary(chunk))
+                    {
+                        return false;
+                    }
+                    result.insert(result.end(), chunk.begin(), chunk.end());
+                }
+                return true;
+            }
+
+            default:
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read,
+                                        exception_message(input_format_t::cbor, concat("expected length specification (0x40-0x5B) or indefinite binary array type (0x5F); last byte: 0x", last_token), "binary"), nullptr));
+            }
+        }
+    }
+
+    /*!
+    @param[in] len  the length of the array or static_cast<std::size_t>(-1) for an
+                    array of indefinite size
+    @param[in] tag_handler how CBOR tags should be treated
+    @return whether array creation completed
+    */
+    bool get_cbor_array(const std::size_t len,
+                        const cbor_tag_handler_t tag_handler)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(len)))
+        {
+            return false;
+        }
+
+        if (len != static_cast<std::size_t>(-1))
+        {
+            for (std::size_t i = 0; i < len; ++i)
+            {
+                if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(true, tag_handler)))
+                {
+                    return false;
+                }
+            }
+        }
+        else
+        {
+            while (get() != 0xFF)
+            {
+                if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(false, tag_handler)))
+                {
+                    return false;
+                }
+            }
+        }
+
+        return sax->end_array();
+    }
+
+    /*!
+    @param[in] len  the length of the object or static_cast<std::size_t>(-1) for an
+                    object of indefinite size
+    @param[in] tag_handler how CBOR tags should be treated
+    @return whether object creation completed
+    */
+    bool get_cbor_object(const std::size_t len,
+                         const cbor_tag_handler_t tag_handler)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(len)))
+        {
+            return false;
+        }
+
+        if (len != 0)
+        {
+            string_t key;
+            if (len != static_cast<std::size_t>(-1))
+            {
+                for (std::size_t i = 0; i < len; ++i)
+                {
+                    get();
+                    if (JSON_HEDLEY_UNLIKELY(!get_cbor_string(key) || !sax->key(key)))
+                    {
+                        return false;
+                    }
+
+                    if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(true, tag_handler)))
+                    {
+                        return false;
+                    }
+                    key.clear();
+                }
+            }
+            else
+            {
+                while (get() != 0xFF)
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!get_cbor_string(key) || !sax->key(key)))
+                    {
+                        return false;
+                    }
+
+                    if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(true, tag_handler)))
+                    {
+                        return false;
+                    }
+                    key.clear();
+                }
+            }
+        }
+
+        return sax->end_object();
+    }
+
+    /////////////
+    // MsgPack //
+    /////////////
+
+    /*!
+    @return whether a valid MessagePack value was passed to the SAX parser
+    */
+    bool parse_msgpack_internal()
+    {
+        switch (get())
+        {
+            // EOF
+            case char_traits<char_type>::eof():
+                return unexpect_eof(input_format_t::msgpack, "value");
+
+            // positive fixint
+            case 0x00:
+            case 0x01:
+            case 0x02:
+            case 0x03:
+            case 0x04:
+            case 0x05:
+            case 0x06:
+            case 0x07:
+            case 0x08:
+            case 0x09:
+            case 0x0A:
+            case 0x0B:
+            case 0x0C:
+            case 0x0D:
+            case 0x0E:
+            case 0x0F:
+            case 0x10:
+            case 0x11:
+            case 0x12:
+            case 0x13:
+            case 0x14:
+            case 0x15:
+            case 0x16:
+            case 0x17:
+            case 0x18:
+            case 0x19:
+            case 0x1A:
+            case 0x1B:
+            case 0x1C:
+            case 0x1D:
+            case 0x1E:
+            case 0x1F:
+            case 0x20:
+            case 0x21:
+            case 0x22:
+            case 0x23:
+            case 0x24:
+            case 0x25:
+            case 0x26:
+            case 0x27:
+            case 0x28:
+            case 0x29:
+            case 0x2A:
+            case 0x2B:
+            case 0x2C:
+            case 0x2D:
+            case 0x2E:
+            case 0x2F:
+            case 0x30:
+            case 0x31:
+            case 0x32:
+            case 0x33:
+            case 0x34:
+            case 0x35:
+            case 0x36:
+            case 0x37:
+            case 0x38:
+            case 0x39:
+            case 0x3A:
+            case 0x3B:
+            case 0x3C:
+            case 0x3D:
+            case 0x3E:
+            case 0x3F:
+            case 0x40:
+            case 0x41:
+            case 0x42:
+            case 0x43:
+            case 0x44:
+            case 0x45:
+            case 0x46:
+            case 0x47:
+            case 0x48:
+            case 0x49:
+            case 0x4A:
+            case 0x4B:
+            case 0x4C:
+            case 0x4D:
+            case 0x4E:
+            case 0x4F:
+            case 0x50:
+            case 0x51:
+            case 0x52:
+            case 0x53:
+            case 0x54:
+            case 0x55:
+            case 0x56:
+            case 0x57:
+            case 0x58:
+            case 0x59:
+            case 0x5A:
+            case 0x5B:
+            case 0x5C:
+            case 0x5D:
+            case 0x5E:
+            case 0x5F:
+            case 0x60:
+            case 0x61:
+            case 0x62:
+            case 0x63:
+            case 0x64:
+            case 0x65:
+            case 0x66:
+            case 0x67:
+            case 0x68:
+            case 0x69:
+            case 0x6A:
+            case 0x6B:
+            case 0x6C:
+            case 0x6D:
+            case 0x6E:
+            case 0x6F:
+            case 0x70:
+            case 0x71:
+            case 0x72:
+            case 0x73:
+            case 0x74:
+            case 0x75:
+            case 0x76:
+            case 0x77:
+            case 0x78:
+            case 0x79:
+            case 0x7A:
+            case 0x7B:
+            case 0x7C:
+            case 0x7D:
+            case 0x7E:
+            case 0x7F:
+                return sax->number_unsigned(static_cast<number_unsigned_t>(current));
+
+            // fixmap
+            case 0x80:
+            case 0x81:
+            case 0x82:
+            case 0x83:
+            case 0x84:
+            case 0x85:
+            case 0x86:
+            case 0x87:
+            case 0x88:
+            case 0x89:
+            case 0x8A:
+            case 0x8B:
+            case 0x8C:
+            case 0x8D:
+            case 0x8E:
+            case 0x8F:
+                return get_msgpack_object(conditional_static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x0Fu));
+
+            // fixarray
+            case 0x90:
+            case 0x91:
+            case 0x92:
+            case 0x93:
+            case 0x94:
+            case 0x95:
+            case 0x96:
+            case 0x97:
+            case 0x98:
+            case 0x99:
+            case 0x9A:
+            case 0x9B:
+            case 0x9C:
+            case 0x9D:
+            case 0x9E:
+            case 0x9F:
+                return get_msgpack_array(conditional_static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x0Fu));
+
+            // fixstr
+            case 0xA0:
+            case 0xA1:
+            case 0xA2:
+            case 0xA3:
+            case 0xA4:
+            case 0xA5:
+            case 0xA6:
+            case 0xA7:
+            case 0xA8:
+            case 0xA9:
+            case 0xAA:
+            case 0xAB:
+            case 0xAC:
+            case 0xAD:
+            case 0xAE:
+            case 0xAF:
+            case 0xB0:
+            case 0xB1:
+            case 0xB2:
+            case 0xB3:
+            case 0xB4:
+            case 0xB5:
+            case 0xB6:
+            case 0xB7:
+            case 0xB8:
+            case 0xB9:
+            case 0xBA:
+            case 0xBB:
+            case 0xBC:
+            case 0xBD:
+            case 0xBE:
+            case 0xBF:
+            case 0xD9: // str 8
+            case 0xDA: // str 16
+            case 0xDB: // str 32
+            {
+                string_t s;
+                return get_msgpack_string(s) && sax->string(s);
+            }
+
+            case 0xC0: // nil
+                return sax->null();
+
+            case 0xC2: // false
+                return sax->boolean(false);
+
+            case 0xC3: // true
+                return sax->boolean(true);
+
+            case 0xC4: // bin 8
+            case 0xC5: // bin 16
+            case 0xC6: // bin 32
+            case 0xC7: // ext 8
+            case 0xC8: // ext 16
+            case 0xC9: // ext 32
+            case 0xD4: // fixext 1
+            case 0xD5: // fixext 2
+            case 0xD6: // fixext 4
+            case 0xD7: // fixext 8
+            case 0xD8: // fixext 16
+            {
+                binary_t b;
+                return get_msgpack_binary(b) && sax->binary(b);
+            }
+
+            case 0xCA: // float 32
+            {
+                float number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 0xCB: // float 64
+            {
+                double number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 0xCC: // uint 8
+            {
+                std::uint8_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
+            }
+
+            case 0xCD: // uint 16
+            {
+                std::uint16_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
+            }
+
+            case 0xCE: // uint 32
+            {
+                std::uint32_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
+            }
+
+            case 0xCF: // uint 64
+            {
+                std::uint64_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
+            }
+
+            case 0xD0: // int 8
+            {
+                std::int8_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
+            }
+
+            case 0xD1: // int 16
+            {
+                std::int16_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
+            }
+
+            case 0xD2: // int 32
+            {
+                std::int32_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
+            }
+
+            case 0xD3: // int 64
+            {
+                std::int64_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
+            }
+
+            case 0xDC: // array 16
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::msgpack, len) && get_msgpack_array(static_cast<std::size_t>(len));
+            }
+
+            case 0xDD: // array 32
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::msgpack, len) && get_msgpack_array(conditional_static_cast<std::size_t>(len));
+            }
+
+            case 0xDE: // map 16
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::msgpack, len) && get_msgpack_object(static_cast<std::size_t>(len));
+            }
+
+            case 0xDF: // map 32
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::msgpack, len) && get_msgpack_object(conditional_static_cast<std::size_t>(len));
+            }
+
+            // negative fixint
+            case 0xE0:
+            case 0xE1:
+            case 0xE2:
+            case 0xE3:
+            case 0xE4:
+            case 0xE5:
+            case 0xE6:
+            case 0xE7:
+            case 0xE8:
+            case 0xE9:
+            case 0xEA:
+            case 0xEB:
+            case 0xEC:
+            case 0xED:
+            case 0xEE:
+            case 0xEF:
+            case 0xF0:
+            case 0xF1:
+            case 0xF2:
+            case 0xF3:
+            case 0xF4:
+            case 0xF5:
+            case 0xF6:
+            case 0xF7:
+            case 0xF8:
+            case 0xF9:
+            case 0xFA:
+            case 0xFB:
+            case 0xFC:
+            case 0xFD:
+            case 0xFE:
+            case 0xFF:
+                return sax->number_integer(static_cast<std::int8_t>(current));
+
+            default: // anything else
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                        exception_message(input_format_t::msgpack, concat("invalid byte: 0x", last_token), "value"), nullptr));
+            }
+        }
+    }
+
+    /*!
+    @brief reads a MessagePack string
+
+    This function first reads starting bytes to determine the expected
+    string length and then copies this number of bytes into a string.
+
+    @param[out] result  created string
+
+    @return whether string creation completed
+    */
+    bool get_msgpack_string(string_t& result)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::msgpack, "string")))
+        {
+            return false;
+        }
+
+        switch (current)
+        {
+            // fixstr
+            case 0xA0:
+            case 0xA1:
+            case 0xA2:
+            case 0xA3:
+            case 0xA4:
+            case 0xA5:
+            case 0xA6:
+            case 0xA7:
+            case 0xA8:
+            case 0xA9:
+            case 0xAA:
+            case 0xAB:
+            case 0xAC:
+            case 0xAD:
+            case 0xAE:
+            case 0xAF:
+            case 0xB0:
+            case 0xB1:
+            case 0xB2:
+            case 0xB3:
+            case 0xB4:
+            case 0xB5:
+            case 0xB6:
+            case 0xB7:
+            case 0xB8:
+            case 0xB9:
+            case 0xBA:
+            case 0xBB:
+            case 0xBC:
+            case 0xBD:
+            case 0xBE:
+            case 0xBF:
+            {
+                return get_string(input_format_t::msgpack, static_cast<unsigned int>(current) & 0x1Fu, result);
+            }
+
+            case 0xD9: // str 8
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::msgpack, len) && get_string(input_format_t::msgpack, len, result);
+            }
+
+            case 0xDA: // str 16
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::msgpack, len) && get_string(input_format_t::msgpack, len, result);
+            }
+
+            case 0xDB: // str 32
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::msgpack, len) && get_string(input_format_t::msgpack, len, result);
+            }
+
+            default:
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read,
+                                        exception_message(input_format_t::msgpack, concat("expected length specification (0xA0-0xBF, 0xD9-0xDB); last byte: 0x", last_token), "string"), nullptr));
+            }
+        }
+    }
+
+    /*!
+    @brief reads a MessagePack byte array
+
+    This function first reads starting bytes to determine the expected
+    byte array length and then copies this number of bytes into a byte array.
+
+    @param[out] result  created byte array
+
+    @return whether byte array creation completed
+    */
+    bool get_msgpack_binary(binary_t& result)
+    {
+        // helper function to set the subtype
+        auto assign_and_return_true = [&result](std::int8_t subtype)
+        {
+            result.set_subtype(static_cast<std::uint8_t>(subtype));
+            return true;
+        };
+
+        switch (current)
+        {
+            case 0xC4: // bin 8
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_binary(input_format_t::msgpack, len, result);
+            }
+
+            case 0xC5: // bin 16
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_binary(input_format_t::msgpack, len, result);
+            }
+
+            case 0xC6: // bin 32
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_binary(input_format_t::msgpack, len, result);
+            }
+
+            case 0xC7: // ext 8
+            {
+                std::uint8_t len{};
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, len, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xC8: // ext 16
+            {
+                std::uint16_t len{};
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, len, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xC9: // ext 32
+            {
+                std::uint32_t len{};
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, len, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xD4: // fixext 1
+            {
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, 1, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xD5: // fixext 2
+            {
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, 2, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xD6: // fixext 4
+            {
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, 4, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xD7: // fixext 8
+            {
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, 8, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xD8: // fixext 16
+            {
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, 16, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            default:           // LCOV_EXCL_LINE
+                return false;  // LCOV_EXCL_LINE
+        }
+    }
+
+    /*!
+    @param[in] len  the length of the array
+    @return whether array creation completed
+    */
+    bool get_msgpack_array(const std::size_t len)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(len)))
+        {
+            return false;
+        }
+
+        for (std::size_t i = 0; i < len; ++i)
+        {
+            if (JSON_HEDLEY_UNLIKELY(!parse_msgpack_internal()))
+            {
+                return false;
+            }
+        }
+
+        return sax->end_array();
+    }
+
+    /*!
+    @param[in] len  the length of the object
+    @return whether object creation completed
+    */
+    bool get_msgpack_object(const std::size_t len)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(len)))
+        {
+            return false;
+        }
+
+        string_t key;
+        for (std::size_t i = 0; i < len; ++i)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!get_msgpack_string(key) || !sax->key(key)))
+            {
+                return false;
+            }
+
+            if (JSON_HEDLEY_UNLIKELY(!parse_msgpack_internal()))
+            {
+                return false;
+            }
+            key.clear();
+        }
+
+        return sax->end_object();
+    }
+
+    ////////////
+    // UBJSON //
+    ////////////
+
+    /*!
+    @param[in] get_char  whether a new character should be retrieved from the
+                         input (true, default) or whether the last read
+                         character should be considered instead
+
+    @return whether a valid UBJSON value was passed to the SAX parser
+    */
+    bool parse_ubjson_internal(const bool get_char = true)
+    {
+        return get_ubjson_value(get_char ? get_ignore_noop() : current);
+    }
+
+    /*!
+    @brief reads a UBJSON string
+
+    This function is either called after reading the 'S' byte explicitly
+    indicating a string, or in case of an object key where the 'S' byte can be
+    left out.
+
+    @param[out] result   created string
+    @param[in] get_char  whether a new character should be retrieved from the
+                         input (true, default) or whether the last read
+                         character should be considered instead
+
+    @return whether string creation completed
+    */
+    bool get_ubjson_string(string_t& result, const bool get_char = true)
+    {
+        if (get_char)
+        {
+            get();  // TODO(niels): may we ignore N here?
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "value")))
+        {
+            return false;
+        }
+
+        switch (current)
+        {
+            case 'U':
+            {
+                std::uint8_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            case 'i':
+            {
+                std::int8_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            case 'I':
+            {
+                std::int16_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            case 'l':
+            {
+                std::int32_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            case 'L':
+            {
+                std::int64_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            case 'u':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint16_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            case 'm':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint32_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            case 'M':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint64_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            default:
+                break;
+        }
+        auto last_token = get_token_string();
+        std::string message;
+
+        if (input_format != input_format_t::bjdata)
+        {
+            message = "expected length type specification (U, i, I, l, L); last byte: 0x" + last_token;
+        }
+        else
+        {
+            message = "expected length type specification (U, i, u, I, m, l, M, L); last byte: 0x" + last_token;
+        }
+        return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read, exception_message(input_format, message, "string"), nullptr));
+    }
+
+    /*!
+    @param[out] dim  an integer vector storing the ND array dimensions
+    @return whether reading ND array size vector is successful
+    */
+    bool get_ubjson_ndarray_size(std::vector<size_t>& dim)
+    {
+        std::pair<std::size_t, char_int_type> size_and_type;
+        size_t dimlen = 0;
+        bool no_ndarray = true;
+
+        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_type(size_and_type, no_ndarray)))
+        {
+            return false;
+        }
+
+        if (size_and_type.first != npos)
+        {
+            if (size_and_type.second != 0)
+            {
+                if (size_and_type.second != 'N')
+                {
+                    for (std::size_t i = 0; i < size_and_type.first; ++i)
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_value(dimlen, no_ndarray, size_and_type.second)))
+                        {
+                            return false;
+                        }
+                        dim.push_back(dimlen);
+                    }
+                }
+            }
+            else
+            {
+                for (std::size_t i = 0; i < size_and_type.first; ++i)
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_value(dimlen, no_ndarray)))
+                    {
+                        return false;
+                    }
+                    dim.push_back(dimlen);
+                }
+            }
+        }
+        else
+        {
+            while (current != ']')
+            {
+                if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_value(dimlen, no_ndarray, current)))
+                {
+                    return false;
+                }
+                dim.push_back(dimlen);
+                get_ignore_noop();
+            }
+        }
+        return true;
+    }
+
+    /*!
+    @param[out] result  determined size
+    @param[in,out] is_ndarray  for input, `true` means already inside an ndarray vector
+                               or ndarray dimension is not allowed; `false` means ndarray
+                               is allowed; for output, `true` means an ndarray is found;
+                               is_ndarray can only return `true` when its initial value
+                               is `false`
+    @param[in] prefix  type marker if already read, otherwise set to 0
+
+    @return whether size determination completed
+    */
+    bool get_ubjson_size_value(std::size_t& result, bool& is_ndarray, char_int_type prefix = 0)
+    {
+        if (prefix == 0)
+        {
+            prefix = get_ignore_noop();
+        }
+
+        switch (prefix)
+        {
+            case 'U':
+            {
+                std::uint8_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                result = static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case 'i':
+            {
+                std::int8_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                if (number < 0)
+                {
+                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read,
+                                            exception_message(input_format, "count in an optimized container must be positive", "size"), nullptr));
+                }
+                result = static_cast<std::size_t>(number); // NOLINT(bugprone-signed-char-misuse,cert-str34-c): number is not a char
+                return true;
+            }
+
+            case 'I':
+            {
+                std::int16_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                if (number < 0)
+                {
+                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read,
+                                            exception_message(input_format, "count in an optimized container must be positive", "size"), nullptr));
+                }
+                result = static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case 'l':
+            {
+                std::int32_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                if (number < 0)
+                {
+                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read,
+                                            exception_message(input_format, "count in an optimized container must be positive", "size"), nullptr));
+                }
+                result = static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case 'L':
+            {
+                std::int64_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                if (number < 0)
+                {
+                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read,
+                                            exception_message(input_format, "count in an optimized container must be positive", "size"), nullptr));
+                }
+                if (!value_in_range_of<std::size_t>(number))
+                {
+                    return sax->parse_error(chars_read, get_token_string(), out_of_range::create(408,
+                                            exception_message(input_format, "integer value overflow", "size"), nullptr));
+                }
+                result = static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case 'u':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint16_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                result = static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case 'm':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint32_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                result = conditional_static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case 'M':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint64_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                if (!value_in_range_of<std::size_t>(number))
+                {
+                    return sax->parse_error(chars_read, get_token_string(), out_of_range::create(408,
+                                            exception_message(input_format, "integer value overflow", "size"), nullptr));
+                }
+                result = detail::conditional_static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case '[':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                if (is_ndarray) // ndarray dimensional vector can only contain integers, and can not embed another array
+                {
+                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read, exception_message(input_format, "ndarray dimensional vector is not allowed", "size"), nullptr));
+                }
+                std::vector<size_t> dim;
+                if (JSON_HEDLEY_UNLIKELY(!get_ubjson_ndarray_size(dim)))
+                {
+                    return false;
+                }
+                if (dim.size() == 1 || (dim.size() == 2 && dim.at(0) == 1)) // return normal array size if 1D row vector
+                {
+                    result = dim.at(dim.size() - 1);
+                    return true;
+                }
+                if (!dim.empty())  // if ndarray, convert to an object in JData annotated array format
+                {
+                    for (auto i : dim) // test if any dimension in an ndarray is 0, if so, return a 1D empty container
+                    {
+                        if ( i == 0 )
+                        {
+                            result = 0;
+                            return true;
+                        }
+                    }
+
+                    string_t key = "_ArraySize_";
+                    if (JSON_HEDLEY_UNLIKELY(!sax->start_object(3) || !sax->key(key) || !sax->start_array(dim.size())))
+                    {
+                        return false;
+                    }
+                    result = 1;
+                    for (auto i : dim)
+                    {
+                        result *= i;
+                        if (result == 0 || result == npos) // because dim elements shall not have zeros, result = 0 means overflow happened; it also can't be npos as it is used to initialize size in get_ubjson_size_type()
+                        {
+                            return sax->parse_error(chars_read, get_token_string(), out_of_range::create(408, exception_message(input_format, "excessive ndarray size caused overflow", "size"), nullptr));
+                        }
+                        if (JSON_HEDLEY_UNLIKELY(!sax->number_unsigned(static_cast<number_unsigned_t>(i))))
+                        {
+                            return false;
+                        }
+                    }
+                    is_ndarray = true;
+                    return sax->end_array();
+                }
+                result = 0;
+                return true;
+            }
+
+            default:
+                break;
+        }
+        auto last_token = get_token_string();
+        std::string message;
+
+        if (input_format != input_format_t::bjdata)
+        {
+            message = "expected length type specification (U, i, I, l, L) after '#'; last byte: 0x" + last_token;
+        }
+        else
+        {
+            message = "expected length type specification (U, i, u, I, m, l, M, L) after '#'; last byte: 0x" + last_token;
+        }
+        return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read, exception_message(input_format, message, "size"), nullptr));
+    }
+
+    /*!
+    @brief determine the type and size for a container
+
+    In the optimized UBJSON format, a type and a size can be provided to allow
+    for a more compact representation.
+
+    @param[out] result  pair of the size and the type
+    @param[in] inside_ndarray  whether the parser is parsing an ND array dimensional vector
+
+    @return whether pair creation completed
+    */
+    bool get_ubjson_size_type(std::pair<std::size_t, char_int_type>& result, bool inside_ndarray = false)
+    {
+        result.first = npos; // size
+        result.second = 0; // type
+        bool is_ndarray = false;
+
+        get_ignore_noop();
+
+        if (current == '$')
+        {
+            result.second = get();  // must not ignore 'N', because 'N' maybe the type
+            if (input_format == input_format_t::bjdata
+                    && JSON_HEDLEY_UNLIKELY(std::binary_search(bjd_optimized_type_markers.begin(), bjd_optimized_type_markers.end(), result.second)))
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                        exception_message(input_format, concat("marker 0x", last_token, " is not a permitted optimized array type"), "type"), nullptr));
+            }
+
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "type")))
+            {
+                return false;
+            }
+
+            get_ignore_noop();
+            if (JSON_HEDLEY_UNLIKELY(current != '#'))
+            {
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "value")))
+                {
+                    return false;
+                }
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                        exception_message(input_format, concat("expected '#' after type information; last byte: 0x", last_token), "size"), nullptr));
+            }
+
+            const bool is_error = get_ubjson_size_value(result.first, is_ndarray);
+            if (input_format == input_format_t::bjdata && is_ndarray)
+            {
+                if (inside_ndarray)
+                {
+                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(112, chars_read,
+                                            exception_message(input_format, "ndarray can not be recursive", "size"), nullptr));
+                }
+                result.second |= (1 << 8); // use bit 8 to indicate ndarray, all UBJSON and BJData markers should be ASCII letters
+            }
+            return is_error;
+        }
+
+        if (current == '#')
+        {
+            const bool is_error = get_ubjson_size_value(result.first, is_ndarray);
+            if (input_format == input_format_t::bjdata && is_ndarray)
+            {
+                return sax->parse_error(chars_read, get_token_string(), parse_error::create(112, chars_read,
+                                        exception_message(input_format, "ndarray requires both type and size", "size"), nullptr));
+            }
+            return is_error;
+        }
+
+        return true;
+    }
+
+    /*!
+    @param prefix  the previously read or set type prefix
+    @return whether value creation completed
+    */
+    bool get_ubjson_value(const char_int_type prefix)
+    {
+        switch (prefix)
+        {
+            case char_traits<char_type>::eof():  // EOF
+                return unexpect_eof(input_format, "value");
+
+            case 'T':  // true
+                return sax->boolean(true);
+            case 'F':  // false
+                return sax->boolean(false);
+
+            case 'Z':  // null
+                return sax->null();
+
+            case 'U':
+            {
+                std::uint8_t number{};
+                return get_number(input_format, number) && sax->number_unsigned(number);
+            }
+
+            case 'i':
+            {
+                std::int8_t number{};
+                return get_number(input_format, number) && sax->number_integer(number);
+            }
+
+            case 'I':
+            {
+                std::int16_t number{};
+                return get_number(input_format, number) && sax->number_integer(number);
+            }
+
+            case 'l':
+            {
+                std::int32_t number{};
+                return get_number(input_format, number) && sax->number_integer(number);
+            }
+
+            case 'L':
+            {
+                std::int64_t number{};
+                return get_number(input_format, number) && sax->number_integer(number);
+            }
+
+            case 'u':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint16_t number{};
+                return get_number(input_format, number) && sax->number_unsigned(number);
+            }
+
+            case 'm':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint32_t number{};
+                return get_number(input_format, number) && sax->number_unsigned(number);
+            }
+
+            case 'M':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint64_t number{};
+                return get_number(input_format, number) && sax->number_unsigned(number);
+            }
+
+            case 'h':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                const auto byte1_raw = get();
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "number")))
+                {
+                    return false;
+                }
+                const auto byte2_raw = get();
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "number")))
+                {
+                    return false;
+                }
+
+                const auto byte1 = static_cast<unsigned char>(byte1_raw);
+                const auto byte2 = static_cast<unsigned char>(byte2_raw);
+
+                // code from RFC 7049, Appendix D, Figure 3:
+                // As half-precision floating-point numbers were only added
+                // to IEEE 754 in 2008, today's programming platforms often
+                // still only have limited support for them. It is very
+                // easy to include at least decoding support for them even
+                // without such support. An example of a small decoder for
+                // half-precision floating-point numbers in the C language
+                // is shown in Fig. 3.
+                const auto half = static_cast<unsigned int>((byte2 << 8u) + byte1);
+                const double val = [&half]
+                {
+                    const int exp = (half >> 10u) & 0x1Fu;
+                    const unsigned int mant = half & 0x3FFu;
+                    JSON_ASSERT(0 <= exp&& exp <= 32);
+                    JSON_ASSERT(mant <= 1024);
+                    switch (exp)
+                    {
+                        case 0:
+                            return std::ldexp(mant, -24);
+                        case 31:
+                            return (mant == 0)
+                            ? std::numeric_limits<double>::infinity()
+                            : std::numeric_limits<double>::quiet_NaN();
+                        default:
+                            return std::ldexp(mant + 1024, exp - 25);
+                    }
+                }();
+                return sax->number_float((half & 0x8000u) != 0
+                                         ? static_cast<number_float_t>(-val)
+                                         : static_cast<number_float_t>(val), "");
+            }
+
+            case 'd':
+            {
+                float number{};
+                return get_number(input_format, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 'D':
+            {
+                double number{};
+                return get_number(input_format, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 'H':
+            {
+                return get_ubjson_high_precision_number();
+            }
+
+            case 'C':  // char
+            {
+                get();
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "char")))
+                {
+                    return false;
+                }
+                if (JSON_HEDLEY_UNLIKELY(current > 127))
+                {
+                    auto last_token = get_token_string();
+                    return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read,
+                                            exception_message(input_format, concat("byte after 'C' must be in range 0x00..0x7F; last byte: 0x", last_token), "char"), nullptr));
+                }
+                string_t s(1, static_cast<typename string_t::value_type>(current));
+                return sax->string(s);
+            }
+
+            case 'S':  // string
+            {
+                string_t s;
+                return get_ubjson_string(s) && sax->string(s);
+            }
+
+            case '[':  // array
+                return get_ubjson_array();
+
+            case '{':  // object
+                return get_ubjson_object();
+
+            default: // anything else
+                break;
+        }
+        auto last_token = get_token_string();
+        return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read, exception_message(input_format, "invalid byte: 0x" + last_token, "value"), nullptr));
+    }
+
+    /*!
+    @return whether array creation completed
+    */
+    bool get_ubjson_array()
+    {
+        std::pair<std::size_t, char_int_type> size_and_type;
+        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_type(size_and_type)))
+        {
+            return false;
+        }
+
+        // if bit-8 of size_and_type.second is set to 1, encode bjdata ndarray as an object in JData annotated array format (https://github.com/NeuroJSON/jdata):
+        // {"_ArrayType_" : "typeid", "_ArraySize_" : [n1, n2, ...], "_ArrayData_" : [v1, v2, ...]}
+
+        if (input_format == input_format_t::bjdata && size_and_type.first != npos && (size_and_type.second & (1 << 8)) != 0)
+        {
+            size_and_type.second &= ~(static_cast<char_int_type>(1) << 8);  // use bit 8 to indicate ndarray, here we remove the bit to restore the type marker
+            auto it = std::lower_bound(bjd_types_map.begin(), bjd_types_map.end(), size_and_type.second, [](const bjd_type & p, char_int_type t)
+            {
+                return p.first < t;
+            });
+            string_t key = "_ArrayType_";
+            if (JSON_HEDLEY_UNLIKELY(it == bjd_types_map.end() || it->first != size_and_type.second))
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                        exception_message(input_format, "invalid byte: 0x" + last_token, "type"), nullptr));
+            }
+
+            string_t type = it->second; // sax->string() takes a reference
+            if (JSON_HEDLEY_UNLIKELY(!sax->key(key) || !sax->string(type)))
+            {
+                return false;
+            }
+
+            if (size_and_type.second == 'C')
+            {
+                size_and_type.second = 'U';
+            }
+
+            key = "_ArrayData_";
+            if (JSON_HEDLEY_UNLIKELY(!sax->key(key) || !sax->start_array(size_and_type.first) ))
+            {
+                return false;
+            }
+
+            for (std::size_t i = 0; i < size_and_type.first; ++i)
+            {
+                if (JSON_HEDLEY_UNLIKELY(!get_ubjson_value(size_and_type.second)))
+                {
+                    return false;
+                }
+            }
+
+            return (sax->end_array() && sax->end_object());
+        }
+
+        if (size_and_type.first != npos)
+        {
+            if (JSON_HEDLEY_UNLIKELY(!sax->start_array(size_and_type.first)))
+            {
+                return false;
+            }
+
+            if (size_and_type.second != 0)
+            {
+                if (size_and_type.second != 'N')
+                {
+                    for (std::size_t i = 0; i < size_and_type.first; ++i)
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_value(size_and_type.second)))
+                        {
+                            return false;
+                        }
+                    }
+                }
+            }
+            else
+            {
+                for (std::size_t i = 0; i < size_and_type.first; ++i)
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal()))
+                    {
+                        return false;
+                    }
+                }
+            }
+        }
+        else
+        {
+            if (JSON_HEDLEY_UNLIKELY(!sax->start_array(static_cast<std::size_t>(-1))))
+            {
+                return false;
+            }
+
+            while (current != ']')
+            {
+                if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal(false)))
+                {
+                    return false;
+                }
+                get_ignore_noop();
+            }
+        }
+
+        return sax->end_array();
+    }
+
+    /*!
+    @return whether object creation completed
+    */
+    bool get_ubjson_object()
+    {
+        std::pair<std::size_t, char_int_type> size_and_type;
+        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_type(size_and_type)))
+        {
+            return false;
+        }
+
+        // do not accept ND-array size in objects in BJData
+        if (input_format == input_format_t::bjdata && size_and_type.first != npos && (size_and_type.second & (1 << 8)) != 0)
+        {
+            auto last_token = get_token_string();
+            return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                    exception_message(input_format, "BJData object does not support ND-array size in optimized format", "object"), nullptr));
+        }
+
+        string_t key;
+        if (size_and_type.first != npos)
+        {
+            if (JSON_HEDLEY_UNLIKELY(!sax->start_object(size_and_type.first)))
+            {
+                return false;
+            }
+
+            if (size_and_type.second != 0)
+            {
+                for (std::size_t i = 0; i < size_and_type.first; ++i)
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!get_ubjson_string(key) || !sax->key(key)))
+                    {
+                        return false;
+                    }
+                    if (JSON_HEDLEY_UNLIKELY(!get_ubjson_value(size_and_type.second)))
+                    {
+                        return false;
+                    }
+                    key.clear();
+                }
+            }
+            else
+            {
+                for (std::size_t i = 0; i < size_and_type.first; ++i)
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!get_ubjson_string(key) || !sax->key(key)))
+                    {
+                        return false;
+                    }
+                    if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal()))
+                    {
+                        return false;
+                    }
+                    key.clear();
+                }
+            }
+        }
+        else
+        {
+            if (JSON_HEDLEY_UNLIKELY(!sax->start_object(static_cast<std::size_t>(-1))))
+            {
+                return false;
+            }
+
+            while (current != '}')
+            {
+                if (JSON_HEDLEY_UNLIKELY(!get_ubjson_string(key, false) || !sax->key(key)))
+                {
+                    return false;
+                }
+                if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal()))
+                {
+                    return false;
+                }
+                get_ignore_noop();
+                key.clear();
+            }
+        }
+
+        return sax->end_object();
+    }
+
+    // Note, no reader for UBJSON binary types is implemented because they do
+    // not exist
+
+    bool get_ubjson_high_precision_number()
+    {
+        // get size of following number string
+        std::size_t size{};
+        bool no_ndarray = true;
+        auto res = get_ubjson_size_value(size, no_ndarray);
+        if (JSON_HEDLEY_UNLIKELY(!res))
+        {
+            return res;
+        }
+
+        // get number string
+        std::vector<char> number_vector;
+        for (std::size_t i = 0; i < size; ++i)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "number")))
+            {
+                return false;
+            }
+            number_vector.push_back(static_cast<char>(current));
+        }
+
+        // parse number string
+        using ia_type = decltype(detail::input_adapter(number_vector));
+        auto number_lexer = detail::lexer<BasicJsonType, ia_type>(detail::input_adapter(number_vector), false);
+        const auto result_number = number_lexer.scan();
+        const auto number_string = number_lexer.get_token_string();
+        const auto result_remainder = number_lexer.scan();
+
+        using token_type = typename detail::lexer_base<BasicJsonType>::token_type;
+
+        if (JSON_HEDLEY_UNLIKELY(result_remainder != token_type::end_of_input))
+        {
+            return sax->parse_error(chars_read, number_string, parse_error::create(115, chars_read,
+                                    exception_message(input_format, concat("invalid number text: ", number_lexer.get_token_string()), "high-precision number"), nullptr));
+        }
+
+        switch (result_number)
+        {
+            case token_type::value_integer:
+                return sax->number_integer(number_lexer.get_number_integer());
+            case token_type::value_unsigned:
+                return sax->number_unsigned(number_lexer.get_number_unsigned());
+            case token_type::value_float:
+                return sax->number_float(number_lexer.get_number_float(), std::move(number_string));
+            case token_type::uninitialized:
+            case token_type::literal_true:
+            case token_type::literal_false:
+            case token_type::literal_null:
+            case token_type::value_string:
+            case token_type::begin_array:
+            case token_type::begin_object:
+            case token_type::end_array:
+            case token_type::end_object:
+            case token_type::name_separator:
+            case token_type::value_separator:
+            case token_type::parse_error:
+            case token_type::end_of_input:
+            case token_type::literal_or_value:
+            default:
+                return sax->parse_error(chars_read, number_string, parse_error::create(115, chars_read,
+                                        exception_message(input_format, concat("invalid number text: ", number_lexer.get_token_string()), "high-precision number"), nullptr));
+        }
+    }
+
+    ///////////////////////
+    // Utility functions //
+    ///////////////////////
+
+    /*!
+    @brief get next character from the input
+
+    This function provides the interface to the used input adapter. It does
+    not throw in case the input reached EOF, but returns a -'ve valued
+    `char_traits<char_type>::eof()` in that case.
+
+    @return character read from the input
+    */
+    char_int_type get()
+    {
+        ++chars_read;
+        return current = ia.get_character();
+    }
+
+    /*!
+    @return character read from the input after ignoring all 'N' entries
+    */
+    char_int_type get_ignore_noop()
+    {
+        do
+        {
+            get();
+        }
+        while (current == 'N');
+
+        return current;
+    }
+
+    /*
+    @brief read a number from the input
+
+    @tparam NumberType the type of the number
+    @param[in] format   the current format (for diagnostics)
+    @param[out] result  number of type @a NumberType
+
+    @return whether conversion completed
+
+    @note This function needs to respect the system's endianness, because
+          bytes in CBOR, MessagePack, and UBJSON are stored in network order
+          (big endian) and therefore need reordering on little endian systems.
+          On the other hand, BSON and BJData use little endian and should reorder
+          on big endian systems.
+    */
+    template<typename NumberType, bool InputIsLittleEndian = false>
+    bool get_number(const input_format_t format, NumberType& result)
+    {
+        // step 1: read input into array with system's byte order
+        std::array<std::uint8_t, sizeof(NumberType)> vec{};
+        for (std::size_t i = 0; i < sizeof(NumberType); ++i)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(format, "number")))
+            {
+                return false;
+            }
+
+            // reverse byte order prior to conversion if necessary
+            if (is_little_endian != (InputIsLittleEndian || format == input_format_t::bjdata))
+            {
+                vec[sizeof(NumberType) - i - 1] = static_cast<std::uint8_t>(current);
+            }
+            else
+            {
+                vec[i] = static_cast<std::uint8_t>(current); // LCOV_EXCL_LINE
+            }
+        }
+
+        // step 2: convert array into number of type T and return
+        std::memcpy(&result, vec.data(), sizeof(NumberType));
+        return true;
+    }
+
+    /*!
+    @brief create a string by reading characters from the input
+
+    @tparam NumberType the type of the number
+    @param[in] format the current format (for diagnostics)
+    @param[in] len number of characters to read
+    @param[out] result string created by reading @a len bytes
+
+    @return whether string creation completed
+
+    @note We can not reserve @a len bytes for the result, because @a len
+          may be too large. Usually, @ref unexpect_eof() detects the end of
+          the input before we run out of string memory.
+    */
+    template<typename NumberType>
+    bool get_string(const input_format_t format,
+                    const NumberType len,
+                    string_t& result)
+    {
+        bool success = true;
+        for (NumberType i = 0; i < len; i++)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(format, "string")))
+            {
+                success = false;
+                break;
+            }
+            result.push_back(static_cast<typename string_t::value_type>(current));
+        }
+        return success;
+    }
+
+    /*!
+    @brief create a byte array by reading bytes from the input
+
+    @tparam NumberType the type of the number
+    @param[in] format the current format (for diagnostics)
+    @param[in] len number of bytes to read
+    @param[out] result byte array created by reading @a len bytes
+
+    @return whether byte array creation completed
+
+    @note We can not reserve @a len bytes for the result, because @a len
+          may be too large. Usually, @ref unexpect_eof() detects the end of
+          the input before we run out of memory.
+    */
+    template<typename NumberType>
+    bool get_binary(const input_format_t format,
+                    const NumberType len,
+                    binary_t& result)
+    {
+        bool success = true;
+        for (NumberType i = 0; i < len; i++)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(format, "binary")))
+            {
+                success = false;
+                break;
+            }
+            result.push_back(static_cast<std::uint8_t>(current));
+        }
+        return success;
+    }
+
+    /*!
+    @param[in] format   the current format (for diagnostics)
+    @param[in] context  further context information (for diagnostics)
+    @return whether the last read character is not EOF
+    */
+    JSON_HEDLEY_NON_NULL(3)
+    bool unexpect_eof(const input_format_t format, const char* context) const
+    {
+        if (JSON_HEDLEY_UNLIKELY(current == char_traits<char_type>::eof()))
+        {
+            return sax->parse_error(chars_read, "<end of file>",
+                                    parse_error::create(110, chars_read, exception_message(format, "unexpected end of input", context), nullptr));
+        }
+        return true;
+    }
+
+    /*!
+    @return a string representation of the last read byte
+    */
+    std::string get_token_string() const
+    {
+        std::array<char, 3> cr{{}};
+        static_cast<void>((std::snprintf)(cr.data(), cr.size(), "%.2hhX", static_cast<unsigned char>(current))); // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+        return std::string{cr.data()};
+    }
+
+    /*!
+    @param[in] format   the current format
+    @param[in] detail   a detailed error message
+    @param[in] context  further context information
+    @return a message string to use in the parse_error exceptions
+    */
+    std::string exception_message(const input_format_t format,
+                                  const std::string& detail,
+                                  const std::string& context) const
+    {
+        std::string error_msg = "syntax error while parsing ";
+
+        switch (format)
+        {
+            case input_format_t::cbor:
+                error_msg += "CBOR";
+                break;
+
+            case input_format_t::msgpack:
+                error_msg += "MessagePack";
+                break;
+
+            case input_format_t::ubjson:
+                error_msg += "UBJSON";
+                break;
+
+            case input_format_t::bson:
+                error_msg += "BSON";
+                break;
+
+            case input_format_t::bjdata:
+                error_msg += "BJData";
+                break;
+
+            case input_format_t::json: // LCOV_EXCL_LINE
+            default:            // LCOV_EXCL_LINE
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        }
+
+        return concat(error_msg, ' ', context, ": ", detail);
+    }
+
+  private:
+    static JSON_INLINE_VARIABLE constexpr std::size_t npos = static_cast<std::size_t>(-1);
+
+    /// input adapter
+    InputAdapterType ia;
+
+    /// the current character
+    char_int_type current = char_traits<char_type>::eof();
+
+    /// the number of characters read
+    std::size_t chars_read = 0;
+
+    /// whether we can assume little endianness
+    const bool is_little_endian = little_endianness();
+
+    /// input format
+    const input_format_t input_format = input_format_t::json;
+
+    /// the SAX parser
+    json_sax_t* sax = nullptr;
+
+    // excluded markers in bjdata optimized type
+#define JSON_BINARY_READER_MAKE_BJD_OPTIMIZED_TYPE_MARKERS_ \
+    make_array<char_int_type>('F', 'H', 'N', 'S', 'T', 'Z', '[', '{')
+
+#define JSON_BINARY_READER_MAKE_BJD_TYPES_MAP_ \
+    make_array<bjd_type>(                      \
+    bjd_type{'C', "char"},                     \
+    bjd_type{'D', "double"},                   \
+    bjd_type{'I', "int16"},                    \
+    bjd_type{'L', "int64"},                    \
+    bjd_type{'M', "uint64"},                   \
+    bjd_type{'U', "uint8"},                    \
+    bjd_type{'d', "single"},                   \
+    bjd_type{'i', "int8"},                     \
+    bjd_type{'l', "int32"},                    \
+    bjd_type{'m', "uint32"},                   \
+    bjd_type{'u', "uint16"})
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    // lookup tables
+    // NOLINTNEXTLINE(cppcoreguidelines-non-private-member-variables-in-classes)
+    const decltype(JSON_BINARY_READER_MAKE_BJD_OPTIMIZED_TYPE_MARKERS_) bjd_optimized_type_markers =
+        JSON_BINARY_READER_MAKE_BJD_OPTIMIZED_TYPE_MARKERS_;
+
+    using bjd_type = std::pair<char_int_type, string_t>;
+    // NOLINTNEXTLINE(cppcoreguidelines-non-private-member-variables-in-classes)
+    const decltype(JSON_BINARY_READER_MAKE_BJD_TYPES_MAP_) bjd_types_map =
+        JSON_BINARY_READER_MAKE_BJD_TYPES_MAP_;
+
+#undef JSON_BINARY_READER_MAKE_BJD_OPTIMIZED_TYPE_MARKERS_
+#undef JSON_BINARY_READER_MAKE_BJD_TYPES_MAP_
+};
+
+#ifndef JSON_HAS_CPP_17
+    template<typename BasicJsonType, typename InputAdapterType, typename SAX>
+    constexpr std::size_t binary_reader<BasicJsonType, InputAdapterType, SAX>::npos;
+#endif
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/input/input_adapters.hpp>
+
+// #include <nlohmann/detail/input/lexer.hpp>
+
+// #include <nlohmann/detail/input/parser.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cmath> // isfinite
+#include <cstdint> // uint8_t
+#include <functional> // function
+#include <string> // string
+#include <utility> // move
+#include <vector> // vector
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/input/input_adapters.hpp>
+
+// #include <nlohmann/detail/input/json_sax.hpp>
+
+// #include <nlohmann/detail/input/lexer.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/is_sax.hpp>
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+////////////
+// parser //
+////////////
+
+enum class parse_event_t : std::uint8_t
+{
+    /// the parser read `{` and started to process a JSON object
+    object_start,
+    /// the parser read `}` and finished processing a JSON object
+    object_end,
+    /// the parser read `[` and started to process a JSON array
+    array_start,
+    /// the parser read `]` and finished processing a JSON array
+    array_end,
+    /// the parser read a key of a value in an object
+    key,
+    /// the parser finished reading a JSON value
+    value
+};
+
+template<typename BasicJsonType>
+using parser_callback_t =
+    std::function<bool(int /*depth*/, parse_event_t /*event*/, BasicJsonType& /*parsed*/)>;
+
+/*!
+@brief syntax analysis
+
+This class implements a recursive descent parser.
+*/
+template<typename BasicJsonType, typename InputAdapterType>
+class parser
+{
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using lexer_t = lexer<BasicJsonType, InputAdapterType>;
+    using token_type = typename lexer_t::token_type;
+
+  public:
+    /// a parser reading from an input adapter
+    explicit parser(InputAdapterType&& adapter,
+                    const parser_callback_t<BasicJsonType> cb = nullptr,
+                    const bool allow_exceptions_ = true,
+                    const bool skip_comments = false)
+        : callback(cb)
+        , m_lexer(std::move(adapter), skip_comments)
+        , allow_exceptions(allow_exceptions_)
+    {
+        // read first token
+        get_token();
+    }
+
+    /*!
+    @brief public parser interface
+
+    @param[in] strict      whether to expect the last token to be EOF
+    @param[in,out] result  parsed JSON value
+
+    @throw parse_error.101 in case of an unexpected token
+    @throw parse_error.102 if to_unicode fails or surrogate error
+    @throw parse_error.103 if to_unicode fails
+    */
+    void parse(const bool strict, BasicJsonType& result)
+    {
+        if (callback)
+        {
+            json_sax_dom_callback_parser<BasicJsonType> sdp(result, callback, allow_exceptions);
+            sax_parse_internal(&sdp);
+
+            // in strict mode, input must be completely read
+            if (strict && (get_token() != token_type::end_of_input))
+            {
+                sdp.parse_error(m_lexer.get_position(),
+                                m_lexer.get_token_string(),
+                                parse_error::create(101, m_lexer.get_position(),
+                                                    exception_message(token_type::end_of_input, "value"), nullptr));
+            }
+
+            // in case of an error, return discarded value
+            if (sdp.is_errored())
+            {
+                result = value_t::discarded;
+                return;
+            }
+
+            // set top-level value to null if it was discarded by the callback
+            // function
+            if (result.is_discarded())
+            {
+                result = nullptr;
+            }
+        }
+        else
+        {
+            json_sax_dom_parser<BasicJsonType> sdp(result, allow_exceptions);
+            sax_parse_internal(&sdp);
+
+            // in strict mode, input must be completely read
+            if (strict && (get_token() != token_type::end_of_input))
+            {
+                sdp.parse_error(m_lexer.get_position(),
+                                m_lexer.get_token_string(),
+                                parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_of_input, "value"), nullptr));
+            }
+
+            // in case of an error, return discarded value
+            if (sdp.is_errored())
+            {
+                result = value_t::discarded;
+                return;
+            }
+        }
+
+        result.assert_invariant();
+    }
+
+    /*!
+    @brief public accept interface
+
+    @param[in] strict  whether to expect the last token to be EOF
+    @return whether the input is a proper JSON text
+    */
+    bool accept(const bool strict = true)
+    {
+        json_sax_acceptor<BasicJsonType> sax_acceptor;
+        return sax_parse(&sax_acceptor, strict);
+    }
+
+    template<typename SAX>
+    JSON_HEDLEY_NON_NULL(2)
+    bool sax_parse(SAX* sax, const bool strict = true)
+    {
+        (void)detail::is_sax_static_asserts<SAX, BasicJsonType> {};
+        const bool result = sax_parse_internal(sax);
+
+        // strict mode: next byte must be EOF
+        if (result && strict && (get_token() != token_type::end_of_input))
+        {
+            return sax->parse_error(m_lexer.get_position(),
+                                    m_lexer.get_token_string(),
+                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_of_input, "value"), nullptr));
+        }
+
+        return result;
+    }
+
+  private:
+    template<typename SAX>
+    JSON_HEDLEY_NON_NULL(2)
+    bool sax_parse_internal(SAX* sax)
+    {
+        // stack to remember the hierarchy of structured values we are parsing
+        // true = array; false = object
+        std::vector<bool> states;
+        // value to avoid a goto (see comment where set to true)
+        bool skip_to_state_evaluation = false;
+
+        while (true)
+        {
+            if (!skip_to_state_evaluation)
+            {
+                // invariant: get_token() was called before each iteration
+                switch (last_token)
+                {
+                    case token_type::begin_object:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(static_cast<std::size_t>(-1))))
+                        {
+                            return false;
+                        }
+
+                        // closing } -> we are done
+                        if (get_token() == token_type::end_object)
+                        {
+                            if (JSON_HEDLEY_UNLIKELY(!sax->end_object()))
+                            {
+                                return false;
+                            }
+                            break;
+                        }
+
+                        // parse key
+                        if (JSON_HEDLEY_UNLIKELY(last_token != token_type::value_string))
+                        {
+                            return sax->parse_error(m_lexer.get_position(),
+                                                    m_lexer.get_token_string(),
+                                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::value_string, "object key"), nullptr));
+                        }
+                        if (JSON_HEDLEY_UNLIKELY(!sax->key(m_lexer.get_string())))
+                        {
+                            return false;
+                        }
+
+                        // parse separator (:)
+                        if (JSON_HEDLEY_UNLIKELY(get_token() != token_type::name_separator))
+                        {
+                            return sax->parse_error(m_lexer.get_position(),
+                                                    m_lexer.get_token_string(),
+                                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::name_separator, "object separator"), nullptr));
+                        }
+
+                        // remember we are now inside an object
+                        states.push_back(false);
+
+                        // parse values
+                        get_token();
+                        continue;
+                    }
+
+                    case token_type::begin_array:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(static_cast<std::size_t>(-1))))
+                        {
+                            return false;
+                        }
+
+                        // closing ] -> we are done
+                        if (get_token() == token_type::end_array)
+                        {
+                            if (JSON_HEDLEY_UNLIKELY(!sax->end_array()))
+                            {
+                                return false;
+                            }
+                            break;
+                        }
+
+                        // remember we are now inside an array
+                        states.push_back(true);
+
+                        // parse values (no need to call get_token)
+                        continue;
+                    }
+
+                    case token_type::value_float:
+                    {
+                        const auto res = m_lexer.get_number_float();
+
+                        if (JSON_HEDLEY_UNLIKELY(!std::isfinite(res)))
+                        {
+                            return sax->parse_error(m_lexer.get_position(),
+                                                    m_lexer.get_token_string(),
+                                                    out_of_range::create(406, concat("number overflow parsing '", m_lexer.get_token_string(), '\''), nullptr));
+                        }
+
+                        if (JSON_HEDLEY_UNLIKELY(!sax->number_float(res, m_lexer.get_string())))
+                        {
+                            return false;
+                        }
+
+                        break;
+                    }
+
+                    case token_type::literal_false:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->boolean(false)))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::literal_null:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->null()))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::literal_true:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->boolean(true)))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::value_integer:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->number_integer(m_lexer.get_number_integer())))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::value_string:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->string(m_lexer.get_string())))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::value_unsigned:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->number_unsigned(m_lexer.get_number_unsigned())))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::parse_error:
+                    {
+                        // using "uninitialized" to avoid "expected" message
+                        return sax->parse_error(m_lexer.get_position(),
+                                                m_lexer.get_token_string(),
+                                                parse_error::create(101, m_lexer.get_position(), exception_message(token_type::uninitialized, "value"), nullptr));
+                    }
+                    case token_type::end_of_input:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(m_lexer.get_position().chars_read_total == 1))
+                        {
+                            return sax->parse_error(m_lexer.get_position(),
+                                                    m_lexer.get_token_string(),
+                                                    parse_error::create(101, m_lexer.get_position(),
+                                                            "attempting to parse an empty input; check that your input string or stream contains the expected JSON", nullptr));
+                        }
+
+                        return sax->parse_error(m_lexer.get_position(),
+                                                m_lexer.get_token_string(),
+                                                parse_error::create(101, m_lexer.get_position(), exception_message(token_type::literal_or_value, "value"), nullptr));
+                    }
+                    case token_type::uninitialized:
+                    case token_type::end_array:
+                    case token_type::end_object:
+                    case token_type::name_separator:
+                    case token_type::value_separator:
+                    case token_type::literal_or_value:
+                    default: // the last token was unexpected
+                    {
+                        return sax->parse_error(m_lexer.get_position(),
+                                                m_lexer.get_token_string(),
+                                                parse_error::create(101, m_lexer.get_position(), exception_message(token_type::literal_or_value, "value"), nullptr));
+                    }
+                }
+            }
+            else
+            {
+                skip_to_state_evaluation = false;
+            }
+
+            // we reached this line after we successfully parsed a value
+            if (states.empty())
+            {
+                // empty stack: we reached the end of the hierarchy: done
+                return true;
+            }
+
+            if (states.back())  // array
+            {
+                // comma -> next value
+                if (get_token() == token_type::value_separator)
+                {
+                    // parse a new value
+                    get_token();
+                    continue;
+                }
+
+                // closing ]
+                if (JSON_HEDLEY_LIKELY(last_token == token_type::end_array))
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!sax->end_array()))
+                    {
+                        return false;
+                    }
+
+                    // We are done with this array. Before we can parse a
+                    // new value, we need to evaluate the new state first.
+                    // By setting skip_to_state_evaluation to false, we
+                    // are effectively jumping to the beginning of this if.
+                    JSON_ASSERT(!states.empty());
+                    states.pop_back();
+                    skip_to_state_evaluation = true;
+                    continue;
+                }
+
+                return sax->parse_error(m_lexer.get_position(),
+                                        m_lexer.get_token_string(),
+                                        parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_array, "array"), nullptr));
+            }
+
+            // states.back() is false -> object
+
+            // comma -> next value
+            if (get_token() == token_type::value_separator)
+            {
+                // parse key
+                if (JSON_HEDLEY_UNLIKELY(get_token() != token_type::value_string))
+                {
+                    return sax->parse_error(m_lexer.get_position(),
+                                            m_lexer.get_token_string(),
+                                            parse_error::create(101, m_lexer.get_position(), exception_message(token_type::value_string, "object key"), nullptr));
+                }
+
+                if (JSON_HEDLEY_UNLIKELY(!sax->key(m_lexer.get_string())))
+                {
+                    return false;
+                }
+
+                // parse separator (:)
+                if (JSON_HEDLEY_UNLIKELY(get_token() != token_type::name_separator))
+                {
+                    return sax->parse_error(m_lexer.get_position(),
+                                            m_lexer.get_token_string(),
+                                            parse_error::create(101, m_lexer.get_position(), exception_message(token_type::name_separator, "object separator"), nullptr));
+                }
+
+                // parse values
+                get_token();
+                continue;
+            }
+
+            // closing }
+            if (JSON_HEDLEY_LIKELY(last_token == token_type::end_object))
+            {
+                if (JSON_HEDLEY_UNLIKELY(!sax->end_object()))
+                {
+                    return false;
+                }
+
+                // We are done with this object. Before we can parse a
+                // new value, we need to evaluate the new state first.
+                // By setting skip_to_state_evaluation to false, we
+                // are effectively jumping to the beginning of this if.
+                JSON_ASSERT(!states.empty());
+                states.pop_back();
+                skip_to_state_evaluation = true;
+                continue;
+            }
+
+            return sax->parse_error(m_lexer.get_position(),
+                                    m_lexer.get_token_string(),
+                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_object, "object"), nullptr));
+        }
+    }
+
+    /// get next token from lexer
+    token_type get_token()
+    {
+        return last_token = m_lexer.scan();
+    }
+
+    std::string exception_message(const token_type expected, const std::string& context)
+    {
+        std::string error_msg = "syntax error ";
+
+        if (!context.empty())
+        {
+            error_msg += concat("while parsing ", context, ' ');
+        }
+
+        error_msg += "- ";
+
+        if (last_token == token_type::parse_error)
+        {
+            error_msg += concat(m_lexer.get_error_message(), "; last read: '",
+                                m_lexer.get_token_string(), '\'');
+        }
+        else
+        {
+            error_msg += concat("unexpected ", lexer_t::token_type_name(last_token));
+        }
+
+        if (expected != token_type::uninitialized)
+        {
+            error_msg += concat("; expected ", lexer_t::token_type_name(expected));
+        }
+
+        return error_msg;
+    }
+
+  private:
+    /// callback function
+    const parser_callback_t<BasicJsonType> callback = nullptr;
+    /// the type of the last read token
+    token_type last_token = token_type::uninitialized;
+    /// the lexer
+    lexer_t m_lexer;
+    /// whether to throw exceptions in case of errors
+    const bool allow_exceptions = true;
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/iterators/internal_iterator.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+// #include <nlohmann/detail/iterators/primitive_iterator.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstddef> // ptrdiff_t
+#include <limits>  // numeric_limits
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/*
+@brief an iterator for primitive JSON types
+
+This class models an iterator for primitive JSON types (boolean, number,
+string). It's only purpose is to allow the iterator/const_iterator classes
+to "iterate" over primitive values. Internally, the iterator is modeled by
+a `difference_type` variable. Value begin_value (`0`) models the begin,
+end_value (`1`) models past the end.
+*/
+class primitive_iterator_t
+{
+  private:
+    using difference_type = std::ptrdiff_t;
+    static constexpr difference_type begin_value = 0;
+    static constexpr difference_type end_value = begin_value + 1;
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    /// iterator as signed integer type
+    difference_type m_it = (std::numeric_limits<std::ptrdiff_t>::min)();
+
+  public:
+    constexpr difference_type get_value() const noexcept
+    {
+        return m_it;
+    }
+
+    /// set iterator to a defined beginning
+    void set_begin() noexcept
+    {
+        m_it = begin_value;
+    }
+
+    /// set iterator to a defined past the end
+    void set_end() noexcept
+    {
+        m_it = end_value;
+    }
+
+    /// return whether the iterator can be dereferenced
+    constexpr bool is_begin() const noexcept
+    {
+        return m_it == begin_value;
+    }
+
+    /// return whether the iterator is at end
+    constexpr bool is_end() const noexcept
+    {
+        return m_it == end_value;
+    }
+
+    friend constexpr bool operator==(primitive_iterator_t lhs, primitive_iterator_t rhs) noexcept
+    {
+        return lhs.m_it == rhs.m_it;
+    }
+
+    friend constexpr bool operator<(primitive_iterator_t lhs, primitive_iterator_t rhs) noexcept
+    {
+        return lhs.m_it < rhs.m_it;
+    }
+
+    primitive_iterator_t operator+(difference_type n) noexcept
+    {
+        auto result = *this;
+        result += n;
+        return result;
+    }
+
+    friend constexpr difference_type operator-(primitive_iterator_t lhs, primitive_iterator_t rhs) noexcept
+    {
+        return lhs.m_it - rhs.m_it;
+    }
+
+    primitive_iterator_t& operator++() noexcept
+    {
+        ++m_it;
+        return *this;
+    }
+
+    primitive_iterator_t operator++(int)& noexcept // NOLINT(cert-dcl21-cpp)
+    {
+        auto result = *this;
+        ++m_it;
+        return result;
+    }
+
+    primitive_iterator_t& operator--() noexcept
+    {
+        --m_it;
+        return *this;
+    }
+
+    primitive_iterator_t operator--(int)& noexcept // NOLINT(cert-dcl21-cpp)
+    {
+        auto result = *this;
+        --m_it;
+        return result;
+    }
+
+    primitive_iterator_t& operator+=(difference_type n) noexcept
+    {
+        m_it += n;
+        return *this;
+    }
+
+    primitive_iterator_t& operator-=(difference_type n) noexcept
+    {
+        m_it -= n;
+        return *this;
+    }
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/*!
+@brief an iterator value
+
+@note This structure could easily be a union, but MSVC currently does not allow
+unions members with complex constructors, see https://github.com/nlohmann/json/pull/105.
+*/
+template<typename BasicJsonType> struct internal_iterator
+{
+    /// iterator for JSON objects
+    typename BasicJsonType::object_t::iterator object_iterator {};
+    /// iterator for JSON arrays
+    typename BasicJsonType::array_t::iterator array_iterator {};
+    /// generic iterator for all other types
+    primitive_iterator_t primitive_iterator {};
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/iterators/iter_impl.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <iterator> // iterator, random_access_iterator_tag, bidirectional_iterator_tag, advance, next
+#include <type_traits> // conditional, is_const, remove_const
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/iterators/internal_iterator.hpp>
+
+// #include <nlohmann/detail/iterators/primitive_iterator.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+// forward declare, to be able to friend it later on
+template<typename IteratorType> class iteration_proxy;
+template<typename IteratorType> class iteration_proxy_value;
+
+/*!
+@brief a template for a bidirectional iterator for the @ref basic_json class
+This class implements a both iterators (iterator and const_iterator) for the
+@ref basic_json class.
+@note An iterator is called *initialized* when a pointer to a JSON value has
+      been set (e.g., by a constructor or a copy assignment). If the iterator is
+      default-constructed, it is *uninitialized* and most methods are undefined.
+      **The library uses assertions to detect calls on uninitialized iterators.**
+@requirement The class satisfies the following concept requirements:
+-
+[BidirectionalIterator](https://en.cppreference.com/w/cpp/named_req/BidirectionalIterator):
+  The iterator that can be moved can be moved in both directions (i.e.
+  incremented and decremented).
+@since version 1.0.0, simplified in version 2.0.9, change to bidirectional
+       iterators in version 3.0.0 (see https://github.com/nlohmann/json/issues/593)
+*/
+template<typename BasicJsonType>
+class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-special-member-functions)
+{
+    /// the iterator with BasicJsonType of different const-ness
+    using other_iter_impl = iter_impl<typename std::conditional<std::is_const<BasicJsonType>::value, typename std::remove_const<BasicJsonType>::type, const BasicJsonType>::type>;
+    /// allow basic_json to access private members
+    friend other_iter_impl;
+    friend BasicJsonType;
+    friend iteration_proxy<iter_impl>;
+    friend iteration_proxy_value<iter_impl>;
+
+    using object_t = typename BasicJsonType::object_t;
+    using array_t = typename BasicJsonType::array_t;
+    // make sure BasicJsonType is basic_json or const basic_json
+    static_assert(is_basic_json<typename std::remove_const<BasicJsonType>::type>::value,
+                  "iter_impl only accepts (const) basic_json");
+    // superficial check for the LegacyBidirectionalIterator named requirement
+    static_assert(std::is_base_of<std::bidirectional_iterator_tag, std::bidirectional_iterator_tag>::value
+                  &&  std::is_base_of<std::bidirectional_iterator_tag, typename std::iterator_traits<typename array_t::iterator>::iterator_category>::value,
+                  "basic_json iterator assumes array and object type iterators satisfy the LegacyBidirectionalIterator named requirement.");
+
+  public:
+    /// The std::iterator class template (used as a base class to provide typedefs) is deprecated in C++17.
+    /// The C++ Standard has never required user-defined iterators to derive from std::iterator.
+    /// A user-defined iterator should provide publicly accessible typedefs named
+    /// iterator_category, value_type, difference_type, pointer, and reference.
+    /// Note that value_type is required to be non-const, even for constant iterators.
+    using iterator_category = std::bidirectional_iterator_tag;
+
+    /// the type of the values when the iterator is dereferenced
+    using value_type = typename BasicJsonType::value_type;
+    /// a type to represent differences between iterators
+    using difference_type = typename BasicJsonType::difference_type;
+    /// defines a pointer to the type iterated over (value_type)
+    using pointer = typename std::conditional<std::is_const<BasicJsonType>::value,
+          typename BasicJsonType::const_pointer,
+          typename BasicJsonType::pointer>::type;
+    /// defines a reference to the type iterated over (value_type)
+    using reference =
+        typename std::conditional<std::is_const<BasicJsonType>::value,
+        typename BasicJsonType::const_reference,
+        typename BasicJsonType::reference>::type;
+
+    iter_impl() = default;
+    ~iter_impl() = default;
+    iter_impl(iter_impl&&) noexcept = default;
+    iter_impl& operator=(iter_impl&&) noexcept = default;
+
+    /*!
+    @brief constructor for a given JSON instance
+    @param[in] object  pointer to a JSON object for this iterator
+    @pre object != nullptr
+    @post The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    explicit iter_impl(pointer object) noexcept : m_object(object)
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_data.m_type)
+        {
+            case value_t::object:
+            {
+                m_it.object_iterator = typename object_t::iterator();
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_it.array_iterator = typename array_t::iterator();
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                m_it.primitive_iterator = primitive_iterator_t();
+                break;
+            }
+        }
+    }
+
+    /*!
+    @note The conventional copy constructor and copy assignment are implicitly
+          defined. Combined with the following converting constructor and
+          assignment, they support: (1) copy from iterator to iterator, (2)
+          copy from const iterator to const iterator, and (3) conversion from
+          iterator to const iterator. However conversion from const iterator
+          to iterator is not defined.
+    */
+
+    /*!
+    @brief const copy constructor
+    @param[in] other const iterator to copy from
+    @note This copy constructor had to be defined explicitly to circumvent a bug
+          occurring on msvc v19.0 compiler (VS 2015) debug build. For more
+          information refer to: https://github.com/nlohmann/json/issues/1608
+    */
+    iter_impl(const iter_impl<const BasicJsonType>& other) noexcept
+        : m_object(other.m_object), m_it(other.m_it)
+    {}
+
+    /*!
+    @brief converting assignment
+    @param[in] other const iterator to copy from
+    @return const/non-const iterator
+    @note It is not checked whether @a other is initialized.
+    */
+    iter_impl& operator=(const iter_impl<const BasicJsonType>& other) noexcept
+    {
+        if (&other != this)
+        {
+            m_object = other.m_object;
+            m_it = other.m_it;
+        }
+        return *this;
+    }
+
+    /*!
+    @brief converting constructor
+    @param[in] other  non-const iterator to copy from
+    @note It is not checked whether @a other is initialized.
+    */
+    iter_impl(const iter_impl<typename std::remove_const<BasicJsonType>::type>& other) noexcept
+        : m_object(other.m_object), m_it(other.m_it)
+    {}
+
+    /*!
+    @brief converting assignment
+    @param[in] other  non-const iterator to copy from
+    @return const/non-const iterator
+    @note It is not checked whether @a other is initialized.
+    */
+    iter_impl& operator=(const iter_impl<typename std::remove_const<BasicJsonType>::type>& other) noexcept // NOLINT(cert-oop54-cpp)
+    {
+        m_object = other.m_object;
+        m_it = other.m_it;
+        return *this;
+    }
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    /*!
+    @brief set the iterator to the first value
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    void set_begin() noexcept
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_data.m_type)
+        {
+            case value_t::object:
+            {
+                m_it.object_iterator = m_object->m_data.m_value.object->begin();
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_it.array_iterator = m_object->m_data.m_value.array->begin();
+                break;
+            }
+
+            case value_t::null:
+            {
+                // set to end so begin()==end() is true: null is empty
+                m_it.primitive_iterator.set_end();
+                break;
+            }
+
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                m_it.primitive_iterator.set_begin();
+                break;
+            }
+        }
+    }
+
+    /*!
+    @brief set the iterator past the last value
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    void set_end() noexcept
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_data.m_type)
+        {
+            case value_t::object:
+            {
+                m_it.object_iterator = m_object->m_data.m_value.object->end();
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_it.array_iterator = m_object->m_data.m_value.array->end();
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                m_it.primitive_iterator.set_end();
+                break;
+            }
+        }
+    }
+
+  public:
+    /*!
+    @brief return a reference to the value pointed to by the iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    reference operator*() const
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_data.m_type)
+        {
+            case value_t::object:
+            {
+                JSON_ASSERT(m_it.object_iterator != m_object->m_data.m_value.object->end());
+                return m_it.object_iterator->second;
+            }
+
+            case value_t::array:
+            {
+                JSON_ASSERT(m_it.array_iterator != m_object->m_data.m_value.array->end());
+                return *m_it.array_iterator;
+            }
+
+            case value_t::null:
+                JSON_THROW(invalid_iterator::create(214, "cannot get value", m_object));
+
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                if (JSON_HEDLEY_LIKELY(m_it.primitive_iterator.is_begin()))
+                {
+                    return *m_object;
+                }
+
+                JSON_THROW(invalid_iterator::create(214, "cannot get value", m_object));
+            }
+        }
+    }
+
+    /*!
+    @brief dereference the iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    pointer operator->() const
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_data.m_type)
+        {
+            case value_t::object:
+            {
+                JSON_ASSERT(m_it.object_iterator != m_object->m_data.m_value.object->end());
+                return &(m_it.object_iterator->second);
+            }
+
+            case value_t::array:
+            {
+                JSON_ASSERT(m_it.array_iterator != m_object->m_data.m_value.array->end());
+                return &*m_it.array_iterator;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                if (JSON_HEDLEY_LIKELY(m_it.primitive_iterator.is_begin()))
+                {
+                    return m_object;
+                }
+
+                JSON_THROW(invalid_iterator::create(214, "cannot get value", m_object));
+            }
+        }
+    }
+
+    /*!
+    @brief post-increment (it++)
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl operator++(int)& // NOLINT(cert-dcl21-cpp)
+    {
+        auto result = *this;
+        ++(*this);
+        return result;
+    }
+
+    /*!
+    @brief pre-increment (++it)
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl& operator++()
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_data.m_type)
+        {
+            case value_t::object:
+            {
+                std::advance(m_it.object_iterator, 1);
+                break;
+            }
+
+            case value_t::array:
+            {
+                std::advance(m_it.array_iterator, 1);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                ++m_it.primitive_iterator;
+                break;
+            }
+        }
+
+        return *this;
+    }
+
+    /*!
+    @brief post-decrement (it--)
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl operator--(int)& // NOLINT(cert-dcl21-cpp)
+    {
+        auto result = *this;
+        --(*this);
+        return result;
+    }
+
+    /*!
+    @brief pre-decrement (--it)
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl& operator--()
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_data.m_type)
+        {
+            case value_t::object:
+            {
+                std::advance(m_it.object_iterator, -1);
+                break;
+            }
+
+            case value_t::array:
+            {
+                std::advance(m_it.array_iterator, -1);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                --m_it.primitive_iterator;
+                break;
+            }
+        }
+
+        return *this;
+    }
+
+    /*!
+    @brief comparison: equal
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    template < typename IterImpl, detail::enable_if_t < (std::is_same<IterImpl, iter_impl>::value || std::is_same<IterImpl, other_iter_impl>::value), std::nullptr_t > = nullptr >
+    bool operator==(const IterImpl& other) const
+    {
+        // if objects are not the same, the comparison is undefined
+        if (JSON_HEDLEY_UNLIKELY(m_object != other.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(212, "cannot compare iterators of different containers", m_object));
+        }
+
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_data.m_type)
+        {
+            case value_t::object:
+                return (m_it.object_iterator == other.m_it.object_iterator);
+
+            case value_t::array:
+                return (m_it.array_iterator == other.m_it.array_iterator);
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                return (m_it.primitive_iterator == other.m_it.primitive_iterator);
+        }
+    }
+
+    /*!
+    @brief comparison: not equal
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    template < typename IterImpl, detail::enable_if_t < (std::is_same<IterImpl, iter_impl>::value || std::is_same<IterImpl, other_iter_impl>::value), std::nullptr_t > = nullptr >
+    bool operator!=(const IterImpl& other) const
+    {
+        return !operator==(other);
+    }
+
+    /*!
+    @brief comparison: smaller
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    bool operator<(const iter_impl& other) const
+    {
+        // if objects are not the same, the comparison is undefined
+        if (JSON_HEDLEY_UNLIKELY(m_object != other.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(212, "cannot compare iterators of different containers", m_object));
+        }
+
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_data.m_type)
+        {
+            case value_t::object:
+                JSON_THROW(invalid_iterator::create(213, "cannot compare order of object iterators", m_object));
+
+            case value_t::array:
+                return (m_it.array_iterator < other.m_it.array_iterator);
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                return (m_it.primitive_iterator < other.m_it.primitive_iterator);
+        }
+    }
+
+    /*!
+    @brief comparison: less than or equal
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    bool operator<=(const iter_impl& other) const
+    {
+        return !other.operator < (*this);
+    }
+
+    /*!
+    @brief comparison: greater than
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    bool operator>(const iter_impl& other) const
+    {
+        return !operator<=(other);
+    }
+
+    /*!
+    @brief comparison: greater than or equal
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    bool operator>=(const iter_impl& other) const
+    {
+        return !operator<(other);
+    }
+
+    /*!
+    @brief add to iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl& operator+=(difference_type i)
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_data.m_type)
+        {
+            case value_t::object:
+                JSON_THROW(invalid_iterator::create(209, "cannot use offsets with object iterators", m_object));
+
+            case value_t::array:
+            {
+                std::advance(m_it.array_iterator, i);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                m_it.primitive_iterator += i;
+                break;
+            }
+        }
+
+        return *this;
+    }
+
+    /*!
+    @brief subtract from iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl& operator-=(difference_type i)
+    {
+        return operator+=(-i);
+    }
+
+    /*!
+    @brief add to iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl operator+(difference_type i) const
+    {
+        auto result = *this;
+        result += i;
+        return result;
+    }
+
+    /*!
+    @brief addition of distance and iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    friend iter_impl operator+(difference_type i, const iter_impl& it)
+    {
+        auto result = it;
+        result += i;
+        return result;
+    }
+
+    /*!
+    @brief subtract from iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl operator-(difference_type i) const
+    {
+        auto result = *this;
+        result -= i;
+        return result;
+    }
+
+    /*!
+    @brief return difference
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    difference_type operator-(const iter_impl& other) const
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_data.m_type)
+        {
+            case value_t::object:
+                JSON_THROW(invalid_iterator::create(209, "cannot use offsets with object iterators", m_object));
+
+            case value_t::array:
+                return m_it.array_iterator - other.m_it.array_iterator;
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                return m_it.primitive_iterator - other.m_it.primitive_iterator;
+        }
+    }
+
+    /*!
+    @brief access to successor
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    reference operator[](difference_type n) const
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_data.m_type)
+        {
+            case value_t::object:
+                JSON_THROW(invalid_iterator::create(208, "cannot use operator[] for object iterators", m_object));
+
+            case value_t::array:
+                return *std::next(m_it.array_iterator, n);
+
+            case value_t::null:
+                JSON_THROW(invalid_iterator::create(214, "cannot get value", m_object));
+
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                if (JSON_HEDLEY_LIKELY(m_it.primitive_iterator.get_value() == -n))
+                {
+                    return *m_object;
+                }
+
+                JSON_THROW(invalid_iterator::create(214, "cannot get value", m_object));
+            }
+        }
+    }
+
+    /*!
+    @brief return the key of an object iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    const typename object_t::key_type& key() const
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        if (JSON_HEDLEY_LIKELY(m_object->is_object()))
+        {
+            return m_it.object_iterator->first;
+        }
+
+        JSON_THROW(invalid_iterator::create(207, "cannot use key() for non-object iterators", m_object));
+    }
+
+    /*!
+    @brief return the value of an iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    reference value() const
+    {
+        return operator*();
+    }
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    /// associated JSON instance
+    pointer m_object = nullptr;
+    /// the actual iterator of the associated instance
+    internal_iterator<typename std::remove_const<BasicJsonType>::type> m_it {};
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/iterators/iteration_proxy.hpp>
+
+// #include <nlohmann/detail/iterators/json_reverse_iterator.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstddef> // ptrdiff_t
+#include <iterator> // reverse_iterator
+#include <utility> // declval
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+//////////////////////
+// reverse_iterator //
+//////////////////////
+
+/*!
+@brief a template for a reverse iterator class
+
+@tparam Base the base iterator type to reverse. Valid types are @ref
+iterator (to create @ref reverse_iterator) and @ref const_iterator (to
+create @ref const_reverse_iterator).
+
+@requirement The class satisfies the following concept requirements:
+-
+[BidirectionalIterator](https://en.cppreference.com/w/cpp/named_req/BidirectionalIterator):
+  The iterator that can be moved can be moved in both directions (i.e.
+  incremented and decremented).
+- [OutputIterator](https://en.cppreference.com/w/cpp/named_req/OutputIterator):
+  It is possible to write to the pointed-to element (only if @a Base is
+  @ref iterator).
+
+@since version 1.0.0
+*/
+template<typename Base>
+class json_reverse_iterator : public std::reverse_iterator<Base>
+{
+  public:
+    using difference_type = std::ptrdiff_t;
+    /// shortcut to the reverse iterator adapter
+    using base_iterator = std::reverse_iterator<Base>;
+    /// the reference type for the pointed-to element
+    using reference = typename Base::reference;
+
+    /// create reverse iterator from iterator
+    explicit json_reverse_iterator(const typename base_iterator::iterator_type& it) noexcept
+        : base_iterator(it) {}
+
+    /// create reverse iterator from base class
+    explicit json_reverse_iterator(const base_iterator& it) noexcept : base_iterator(it) {}
+
+    /// post-increment (it++)
+    json_reverse_iterator operator++(int)& // NOLINT(cert-dcl21-cpp)
+    {
+        return static_cast<json_reverse_iterator>(base_iterator::operator++(1));
+    }
+
+    /// pre-increment (++it)
+    json_reverse_iterator& operator++()
+    {
+        return static_cast<json_reverse_iterator&>(base_iterator::operator++());
+    }
+
+    /// post-decrement (it--)
+    json_reverse_iterator operator--(int)& // NOLINT(cert-dcl21-cpp)
+    {
+        return static_cast<json_reverse_iterator>(base_iterator::operator--(1));
+    }
+
+    /// pre-decrement (--it)
+    json_reverse_iterator& operator--()
+    {
+        return static_cast<json_reverse_iterator&>(base_iterator::operator--());
+    }
+
+    /// add to iterator
+    json_reverse_iterator& operator+=(difference_type i)
+    {
+        return static_cast<json_reverse_iterator&>(base_iterator::operator+=(i));
+    }
+
+    /// add to iterator
+    json_reverse_iterator operator+(difference_type i) const
+    {
+        return static_cast<json_reverse_iterator>(base_iterator::operator+(i));
+    }
+
+    /// subtract from iterator
+    json_reverse_iterator operator-(difference_type i) const
+    {
+        return static_cast<json_reverse_iterator>(base_iterator::operator-(i));
+    }
+
+    /// return difference
+    difference_type operator-(const json_reverse_iterator& other) const
+    {
+        return base_iterator(*this) - base_iterator(other);
+    }
+
+    /// access to successor
+    reference operator[](difference_type n) const
+    {
+        return *(this->operator+(n));
+    }
+
+    /// return the key of an object iterator
+    auto key() const -> decltype(std::declval<Base>().key())
+    {
+        auto it = --this->base();
+        return it.key();
+    }
+
+    /// return the value of an iterator
+    reference value() const
+    {
+        auto it = --this->base();
+        return it.operator * ();
+    }
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/iterators/primitive_iterator.hpp>
+
+// #include <nlohmann/detail/json_custom_base_class.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <type_traits> // conditional, is_same
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/*!
+@brief Default base class of the @ref basic_json class.
+
+So that the correct implementations of the copy / move ctors / assign operators
+of @ref basic_json do not require complex case distinctions
+(no base class / custom base class used as customization point),
+@ref basic_json always has a base class.
+By default, this class is used because it is empty and thus has no effect
+on the behavior of @ref basic_json.
+*/
+struct json_default_base {};
+
+template<class T>
+using json_base_class = typename std::conditional <
+                        std::is_same<T, void>::value,
+                        json_default_base,
+                        T
+                        >::type;
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/json_pointer.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <algorithm> // all_of
+#include <cctype> // isdigit
+#include <cerrno> // errno, ERANGE
+#include <cstdlib> // strtoull
+#ifndef JSON_NO_IO
+    #include <iosfwd> // ostream
+#endif  // JSON_NO_IO
+#include <limits> // max
+#include <numeric> // accumulate
+#include <string> // string
+#include <utility> // move
+#include <vector> // vector
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+// #include <nlohmann/detail/string_escape.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+/// @brief JSON Pointer defines a string syntax for identifying a specific value within a JSON document
+/// @sa https://json.nlohmann.me/api/json_pointer/
+template<typename RefStringType>
+class json_pointer
+{
+    // allow basic_json to access private members
+    NLOHMANN_BASIC_JSON_TPL_DECLARATION
+    friend class basic_json;
+
+    template<typename>
+    friend class json_pointer;
+
+    template<typename T>
+    struct string_t_helper
+    {
+        using type = T;
+    };
+
+    NLOHMANN_BASIC_JSON_TPL_DECLARATION
+    struct string_t_helper<NLOHMANN_BASIC_JSON_TPL>
+    {
+        using type = StringType;
+    };
+
+  public:
+    // for backwards compatibility accept BasicJsonType
+    using string_t = typename string_t_helper<RefStringType>::type;
+
+    /// @brief create JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/json_pointer/
+    explicit json_pointer(const string_t& s = "")
+        : reference_tokens(split(s))
+    {}
+
+    /// @brief return a string representation of the JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/to_string/
+    string_t to_string() const
+    {
+        return std::accumulate(reference_tokens.begin(), reference_tokens.end(),
+                               string_t{},
+                               [](const string_t& a, const string_t& b)
+        {
+            return detail::concat(a, '/', detail::escape(b));
+        });
+    }
+
+    /// @brief return a string representation of the JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_string/
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, to_string())
+    operator string_t() const
+    {
+        return to_string();
+    }
+
+#ifndef JSON_NO_IO
+    /// @brief write string representation of the JSON pointer to stream
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ltlt/
+    friend std::ostream& operator<<(std::ostream& o, const json_pointer& ptr)
+    {
+        o << ptr.to_string();
+        return o;
+    }
+#endif
+
+    /// @brief append another JSON pointer at the end of this JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slasheq/
+    json_pointer& operator/=(const json_pointer& ptr)
+    {
+        reference_tokens.insert(reference_tokens.end(),
+                                ptr.reference_tokens.begin(),
+                                ptr.reference_tokens.end());
+        return *this;
+    }
+
+    /// @brief append an unescaped reference token at the end of this JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slasheq/
+    json_pointer& operator/=(string_t token)
+    {
+        push_back(std::move(token));
+        return *this;
+    }
+
+    /// @brief append an array index at the end of this JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slasheq/
+    json_pointer& operator/=(std::size_t array_idx)
+    {
+        return *this /= std::to_string(array_idx);
+    }
+
+    /// @brief create a new JSON pointer by appending the right JSON pointer at the end of the left JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slash/
+    friend json_pointer operator/(const json_pointer& lhs,
+                                  const json_pointer& rhs)
+    {
+        return json_pointer(lhs) /= rhs;
+    }
+
+    /// @brief create a new JSON pointer by appending the unescaped token at the end of the JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slash/
+    friend json_pointer operator/(const json_pointer& lhs, string_t token) // NOLINT(performance-unnecessary-value-param)
+    {
+        return json_pointer(lhs) /= std::move(token);
+    }
+
+    /// @brief create a new JSON pointer by appending the array-index-token at the end of the JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slash/
+    friend json_pointer operator/(const json_pointer& lhs, std::size_t array_idx)
+    {
+        return json_pointer(lhs) /= array_idx;
+    }
+
+    /// @brief returns the parent of this JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/parent_pointer/
+    json_pointer parent_pointer() const
+    {
+        if (empty())
+        {
+            return *this;
+        }
+
+        json_pointer res = *this;
+        res.pop_back();
+        return res;
+    }
+
+    /// @brief remove last reference token
+    /// @sa https://json.nlohmann.me/api/json_pointer/pop_back/
+    void pop_back()
+    {
+        if (JSON_HEDLEY_UNLIKELY(empty()))
+        {
+            JSON_THROW(detail::out_of_range::create(405, "JSON pointer has no parent", nullptr));
+        }
+
+        reference_tokens.pop_back();
+    }
+
+    /// @brief return last reference token
+    /// @sa https://json.nlohmann.me/api/json_pointer/back/
+    const string_t& back() const
+    {
+        if (JSON_HEDLEY_UNLIKELY(empty()))
+        {
+            JSON_THROW(detail::out_of_range::create(405, "JSON pointer has no parent", nullptr));
+        }
+
+        return reference_tokens.back();
+    }
+
+    /// @brief append an unescaped token at the end of the reference pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/push_back/
+    void push_back(const string_t& token)
+    {
+        reference_tokens.push_back(token);
+    }
+
+    /// @brief append an unescaped token at the end of the reference pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/push_back/
+    void push_back(string_t&& token)
+    {
+        reference_tokens.push_back(std::move(token));
+    }
+
+    /// @brief return whether pointer points to the root document
+    /// @sa https://json.nlohmann.me/api/json_pointer/empty/
+    bool empty() const noexcept
+    {
+        return reference_tokens.empty();
+    }
+
+  private:
+    /*!
+    @param[in] s  reference token to be converted into an array index
+
+    @return integer representation of @a s
+
+    @throw parse_error.106  if an array index begins with '0'
+    @throw parse_error.109  if an array index begins not with a digit
+    @throw out_of_range.404 if string @a s could not be converted to an integer
+    @throw out_of_range.410 if an array index exceeds size_type
+    */
+    template<typename BasicJsonType>
+    static typename BasicJsonType::size_type array_index(const string_t& s)
+    {
+        using size_type = typename BasicJsonType::size_type;
+
+        // error condition (cf. RFC 6901, Sect. 4)
+        if (JSON_HEDLEY_UNLIKELY(s.size() > 1 && s[0] == '0'))
+        {
+            JSON_THROW(detail::parse_error::create(106, 0, detail::concat("array index '", s, "' must not begin with '0'"), nullptr));
+        }
+
+        // error condition (cf. RFC 6901, Sect. 4)
+        if (JSON_HEDLEY_UNLIKELY(s.size() > 1 && !(s[0] >= '1' && s[0] <= '9')))
+        {
+            JSON_THROW(detail::parse_error::create(109, 0, detail::concat("array index '", s, "' is not a number"), nullptr));
+        }
+
+        const char* p = s.c_str();
+        char* p_end = nullptr;
+        errno = 0; // strtoull doesn't reset errno
+        const unsigned long long res = std::strtoull(p, &p_end, 10); // NOLINT(runtime/int)
+        if (p == p_end // invalid input or empty string
+                || errno == ERANGE // out of range
+                || JSON_HEDLEY_UNLIKELY(static_cast<std::size_t>(p_end - p) != s.size())) // incomplete read
+        {
+            JSON_THROW(detail::out_of_range::create(404, detail::concat("unresolved reference token '", s, "'"), nullptr));
+        }
+
+        // only triggered on special platforms (like 32bit), see also
+        // https://github.com/nlohmann/json/pull/2203
+        if (res >= static_cast<unsigned long long>((std::numeric_limits<size_type>::max)()))  // NOLINT(runtime/int)
+        {
+            JSON_THROW(detail::out_of_range::create(410, detail::concat("array index ", s, " exceeds size_type"), nullptr));   // LCOV_EXCL_LINE
+        }
+
+        return static_cast<size_type>(res);
+    }
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    json_pointer top() const
+    {
+        if (JSON_HEDLEY_UNLIKELY(empty()))
+        {
+            JSON_THROW(detail::out_of_range::create(405, "JSON pointer has no parent", nullptr));
+        }
+
+        json_pointer result = *this;
+        result.reference_tokens = {reference_tokens[0]};
+        return result;
+    }
+
+  private:
+    /*!
+    @brief create and return a reference to the pointed to value
+
+    @complexity Linear in the number of reference tokens.
+
+    @throw parse_error.109 if array index is not a number
+    @throw type_error.313 if value cannot be unflattened
+    */
+    template<typename BasicJsonType>
+    BasicJsonType& get_and_create(BasicJsonType& j) const
+    {
+        auto* result = &j;
+
+        // in case no reference tokens exist, return a reference to the JSON value
+        // j which will be overwritten by a primitive value
+        for (const auto& reference_token : reference_tokens)
+        {
+            switch (result->type())
+            {
+                case detail::value_t::null:
+                {
+                    if (reference_token == "0")
+                    {
+                        // start a new array if reference token is 0
+                        result = &result->operator[](0);
+                    }
+                    else
+                    {
+                        // start a new object otherwise
+                        result = &result->operator[](reference_token);
+                    }
+                    break;
+                }
+
+                case detail::value_t::object:
+                {
+                    // create an entry in the object
+                    result = &result->operator[](reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    // create an entry in the array
+                    result = &result->operator[](array_index<BasicJsonType>(reference_token));
+                    break;
+                }
+
+                /*
+                The following code is only reached if there exists a reference
+                token _and_ the current value is primitive. In this case, we have
+                an error situation, because primitive values may only occur as
+                single value; that is, with an empty list of reference tokens.
+                */
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                    JSON_THROW(detail::type_error::create(313, "invalid value to unflatten", &j));
+            }
+        }
+
+        return *result;
+    }
+
+    /*!
+    @brief return a reference to the pointed to value
+
+    @note This version does not throw if a value is not present, but tries to
+          create nested values instead. For instance, calling this function
+          with pointer `"/this/that"` on a null value is equivalent to calling
+          `operator[]("this").operator[]("that")` on that value, effectively
+          changing the null value to an object.
+
+    @param[in] ptr  a JSON value
+
+    @return reference to the JSON value pointed to by the JSON pointer
+
+    @complexity Linear in the length of the JSON pointer.
+
+    @throw parse_error.106   if an array index begins with '0'
+    @throw parse_error.109   if an array index was not a number
+    @throw out_of_range.404  if the JSON pointer can not be resolved
+    */
+    template<typename BasicJsonType>
+    BasicJsonType& get_unchecked(BasicJsonType* ptr) const
+    {
+        for (const auto& reference_token : reference_tokens)
+        {
+            // convert null values to arrays or objects before continuing
+            if (ptr->is_null())
+            {
+                // check if reference token is a number
+                const bool nums =
+                    std::all_of(reference_token.begin(), reference_token.end(),
+                                [](const unsigned char x)
+                {
+                    return std::isdigit(x);
+                });
+
+                // change value to array for numbers or "-" or to object otherwise
+                *ptr = (nums || reference_token == "-")
+                       ? detail::value_t::array
+                       : detail::value_t::object;
+            }
+
+            switch (ptr->type())
+            {
+                case detail::value_t::object:
+                {
+                    // use unchecked object access
+                    ptr = &ptr->operator[](reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    if (reference_token == "-")
+                    {
+                        // explicitly treat "-" as index beyond the end
+                        ptr = &ptr->operator[](ptr->m_data.m_value.array->size());
+                    }
+                    else
+                    {
+                        // convert array index to number; unchecked access
+                        ptr = &ptr->operator[](array_index<BasicJsonType>(reference_token));
+                    }
+                    break;
+                }
+
+                case detail::value_t::null:
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                    JSON_THROW(detail::out_of_range::create(404, detail::concat("unresolved reference token '", reference_token, "'"), ptr));
+            }
+        }
+
+        return *ptr;
+    }
+
+    /*!
+    @throw parse_error.106   if an array index begins with '0'
+    @throw parse_error.109   if an array index was not a number
+    @throw out_of_range.402  if the array index '-' is used
+    @throw out_of_range.404  if the JSON pointer can not be resolved
+    */
+    template<typename BasicJsonType>
+    BasicJsonType& get_checked(BasicJsonType* ptr) const
+    {
+        for (const auto& reference_token : reference_tokens)
+        {
+            switch (ptr->type())
+            {
+                case detail::value_t::object:
+                {
+                    // note: at performs range check
+                    ptr = &ptr->at(reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
+                    {
+                        // "-" always fails the range check
+                        JSON_THROW(detail::out_of_range::create(402, detail::concat(
+                                "array index '-' (", std::to_string(ptr->m_data.m_value.array->size()),
+                                ") is out of range"), ptr));
+                    }
+
+                    // note: at performs range check
+                    ptr = &ptr->at(array_index<BasicJsonType>(reference_token));
+                    break;
+                }
+
+                case detail::value_t::null:
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                    JSON_THROW(detail::out_of_range::create(404, detail::concat("unresolved reference token '", reference_token, "'"), ptr));
+            }
+        }
+
+        return *ptr;
+    }
+
+    /*!
+    @brief return a const reference to the pointed to value
+
+    @param[in] ptr  a JSON value
+
+    @return const reference to the JSON value pointed to by the JSON
+    pointer
+
+    @throw parse_error.106   if an array index begins with '0'
+    @throw parse_error.109   if an array index was not a number
+    @throw out_of_range.402  if the array index '-' is used
+    @throw out_of_range.404  if the JSON pointer can not be resolved
+    */
+    template<typename BasicJsonType>
+    const BasicJsonType& get_unchecked(const BasicJsonType* ptr) const
+    {
+        for (const auto& reference_token : reference_tokens)
+        {
+            switch (ptr->type())
+            {
+                case detail::value_t::object:
+                {
+                    // use unchecked object access
+                    ptr = &ptr->operator[](reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
+                    {
+                        // "-" cannot be used for const access
+                        JSON_THROW(detail::out_of_range::create(402, detail::concat("array index '-' (", std::to_string(ptr->m_data.m_value.array->size()), ") is out of range"), ptr));
+                    }
+
+                    // use unchecked array access
+                    ptr = &ptr->operator[](array_index<BasicJsonType>(reference_token));
+                    break;
+                }
+
+                case detail::value_t::null:
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                    JSON_THROW(detail::out_of_range::create(404, detail::concat("unresolved reference token '", reference_token, "'"), ptr));
+            }
+        }
+
+        return *ptr;
+    }
+
+    /*!
+    @throw parse_error.106   if an array index begins with '0'
+    @throw parse_error.109   if an array index was not a number
+    @throw out_of_range.402  if the array index '-' is used
+    @throw out_of_range.404  if the JSON pointer can not be resolved
+    */
+    template<typename BasicJsonType>
+    const BasicJsonType& get_checked(const BasicJsonType* ptr) const
+    {
+        for (const auto& reference_token : reference_tokens)
+        {
+            switch (ptr->type())
+            {
+                case detail::value_t::object:
+                {
+                    // note: at performs range check
+                    ptr = &ptr->at(reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
+                    {
+                        // "-" always fails the range check
+                        JSON_THROW(detail::out_of_range::create(402, detail::concat(
+                                "array index '-' (", std::to_string(ptr->m_data.m_value.array->size()),
+                                ") is out of range"), ptr));
+                    }
+
+                    // note: at performs range check
+                    ptr = &ptr->at(array_index<BasicJsonType>(reference_token));
+                    break;
+                }
+
+                case detail::value_t::null:
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                    JSON_THROW(detail::out_of_range::create(404, detail::concat("unresolved reference token '", reference_token, "'"), ptr));
+            }
+        }
+
+        return *ptr;
+    }
+
+    /*!
+    @throw parse_error.106   if an array index begins with '0'
+    @throw parse_error.109   if an array index was not a number
+    */
+    template<typename BasicJsonType>
+    bool contains(const BasicJsonType* ptr) const
+    {
+        for (const auto& reference_token : reference_tokens)
+        {
+            switch (ptr->type())
+            {
+                case detail::value_t::object:
+                {
+                    if (!ptr->contains(reference_token))
+                    {
+                        // we did not find the key in the object
+                        return false;
+                    }
+
+                    ptr = &ptr->operator[](reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
+                    {
+                        // "-" always fails the range check
+                        return false;
+                    }
+                    if (JSON_HEDLEY_UNLIKELY(reference_token.size() == 1 && !("0" <= reference_token && reference_token <= "9")))
+                    {
+                        // invalid char
+                        return false;
+                    }
+                    if (JSON_HEDLEY_UNLIKELY(reference_token.size() > 1))
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!('1' <= reference_token[0] && reference_token[0] <= '9')))
+                        {
+                            // first char should be between '1' and '9'
+                            return false;
+                        }
+                        for (std::size_t i = 1; i < reference_token.size(); i++)
+                        {
+                            if (JSON_HEDLEY_UNLIKELY(!('0' <= reference_token[i] && reference_token[i] <= '9')))
+                            {
+                                // other char should be between '0' and '9'
+                                return false;
+                            }
+                        }
+                    }
+
+                    const auto idx = array_index<BasicJsonType>(reference_token);
+                    if (idx >= ptr->size())
+                    {
+                        // index out of range
+                        return false;
+                    }
+
+                    ptr = &ptr->operator[](idx);
+                    break;
+                }
+
+                case detail::value_t::null:
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                {
+                    // we do not expect primitive values if there is still a
+                    // reference token to process
+                    return false;
+                }
+            }
+        }
+
+        // no reference token left means we found a primitive value
+        return true;
+    }
+
+    /*!
+    @brief split the string input to reference tokens
+
+    @note This function is only called by the json_pointer constructor.
+          All exceptions below are documented there.
+
+    @throw parse_error.107  if the pointer is not empty or begins with '/'
+    @throw parse_error.108  if character '~' is not followed by '0' or '1'
+    */
+    static std::vector<string_t> split(const string_t& reference_string)
+    {
+        std::vector<string_t> result;
+
+        // special case: empty reference string -> no reference tokens
+        if (reference_string.empty())
+        {
+            return result;
+        }
+
+        // check if nonempty reference string begins with slash
+        if (JSON_HEDLEY_UNLIKELY(reference_string[0] != '/'))
+        {
+            JSON_THROW(detail::parse_error::create(107, 1, detail::concat("JSON pointer must be empty or begin with '/' - was: '", reference_string, "'"), nullptr));
+        }
+
+        // extract the reference tokens:
+        // - slash: position of the last read slash (or end of string)
+        // - start: position after the previous slash
+        for (
+            // search for the first slash after the first character
+            std::size_t slash = reference_string.find_first_of('/', 1),
+            // set the beginning of the first reference token
+            start = 1;
+            // we can stop if start == 0 (if slash == string_t::npos)
+            start != 0;
+            // set the beginning of the next reference token
+            // (will eventually be 0 if slash == string_t::npos)
+            start = (slash == string_t::npos) ? 0 : slash + 1,
+            // find next slash
+            slash = reference_string.find_first_of('/', start))
+        {
+            // use the text between the beginning of the reference token
+            // (start) and the last slash (slash).
+            auto reference_token = reference_string.substr(start, slash - start);
+
+            // check reference tokens are properly escaped
+            for (std::size_t pos = reference_token.find_first_of('~');
+                    pos != string_t::npos;
+                    pos = reference_token.find_first_of('~', pos + 1))
+            {
+                JSON_ASSERT(reference_token[pos] == '~');
+
+                // ~ must be followed by 0 or 1
+                if (JSON_HEDLEY_UNLIKELY(pos == reference_token.size() - 1 ||
+                                         (reference_token[pos + 1] != '0' &&
+                                          reference_token[pos + 1] != '1')))
+                {
+                    JSON_THROW(detail::parse_error::create(108, 0, "escape character '~' must be followed with '0' or '1'", nullptr));
+                }
+            }
+
+            // finally, store the reference token
+            detail::unescape(reference_token);
+            result.push_back(reference_token);
+        }
+
+        return result;
+    }
+
+  private:
+    /*!
+    @param[in] reference_string  the reference string to the current value
+    @param[in] value             the value to consider
+    @param[in,out] result        the result object to insert values to
+
+    @note Empty objects or arrays are flattened to `null`.
+    */
+    template<typename BasicJsonType>
+    static void flatten(const string_t& reference_string,
+                        const BasicJsonType& value,
+                        BasicJsonType& result)
+    {
+        switch (value.type())
+        {
+            case detail::value_t::array:
+            {
+                if (value.m_data.m_value.array->empty())
+                {
+                    // flatten empty array as null
+                    result[reference_string] = nullptr;
+                }
+                else
+                {
+                    // iterate array and use index as reference string
+                    for (std::size_t i = 0; i < value.m_data.m_value.array->size(); ++i)
+                    {
+                        flatten(detail::concat(reference_string, '/', std::to_string(i)),
+                                value.m_data.m_value.array->operator[](i), result);
+                    }
+                }
+                break;
+            }
+
+            case detail::value_t::object:
+            {
+                if (value.m_data.m_value.object->empty())
+                {
+                    // flatten empty object as null
+                    result[reference_string] = nullptr;
+                }
+                else
+                {
+                    // iterate object and use keys as reference string
+                    for (const auto& element : *value.m_data.m_value.object)
+                    {
+                        flatten(detail::concat(reference_string, '/', detail::escape(element.first)), element.second, result);
+                    }
+                }
+                break;
+            }
+
+            case detail::value_t::null:
+            case detail::value_t::string:
+            case detail::value_t::boolean:
+            case detail::value_t::number_integer:
+            case detail::value_t::number_unsigned:
+            case detail::value_t::number_float:
+            case detail::value_t::binary:
+            case detail::value_t::discarded:
+            default:
+            {
+                // add primitive value with its reference string
+                result[reference_string] = value;
+                break;
+            }
+        }
+    }
+
+    /*!
+    @param[in] value  flattened JSON
+
+    @return unflattened JSON
+
+    @throw parse_error.109 if array index is not a number
+    @throw type_error.314  if value is not an object
+    @throw type_error.315  if object values are not primitive
+    @throw type_error.313  if value cannot be unflattened
+    */
+    template<typename BasicJsonType>
+    static BasicJsonType
+    unflatten(const BasicJsonType& value)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!value.is_object()))
+        {
+            JSON_THROW(detail::type_error::create(314, "only objects can be unflattened", &value));
+        }
+
+        BasicJsonType result;
+
+        // iterate the JSON object values
+        for (const auto& element : *value.m_data.m_value.object)
+        {
+            if (JSON_HEDLEY_UNLIKELY(!element.second.is_primitive()))
+            {
+                JSON_THROW(detail::type_error::create(315, "values in object must be primitive", &element.second));
+            }
+
+            // assign value to reference pointed to by JSON pointer; Note that if
+            // the JSON pointer is "" (i.e., points to the whole value), function
+            // get_and_create returns a reference to result itself. An assignment
+            // will then create a primitive value.
+            json_pointer(element.first).get_and_create(result) = element.second;
+        }
+
+        return result;
+    }
+
+    // can't use conversion operator because of ambiguity
+    json_pointer<string_t> convert() const&
+    {
+        json_pointer<string_t> result;
+        result.reference_tokens = reference_tokens;
+        return result;
+    }
+
+    json_pointer<string_t> convert()&&
+    {
+        json_pointer<string_t> result;
+        result.reference_tokens = std::move(reference_tokens);
+        return result;
+    }
+
+  public:
+#if JSON_HAS_THREE_WAY_COMPARISON
+    /// @brief compares two JSON pointers for equality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_eq/
+    template<typename RefStringTypeRhs>
+    bool operator==(const json_pointer<RefStringTypeRhs>& rhs) const noexcept
+    {
+        return reference_tokens == rhs.reference_tokens;
+    }
+
+    /// @brief compares JSON pointer and string for equality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_eq/
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.2, operator==(json_pointer))
+    bool operator==(const string_t& rhs) const
+    {
+        return *this == json_pointer(rhs);
+    }
+
+    /// @brief 3-way compares two JSON pointers
+    template<typename RefStringTypeRhs>
+    std::strong_ordering operator<=>(const json_pointer<RefStringTypeRhs>& rhs) const noexcept // *NOPAD*
+    {
+        return  reference_tokens <=> rhs.reference_tokens; // *NOPAD*
+    }
+#else
+    /// @brief compares two JSON pointers for equality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_eq/
+    template<typename RefStringTypeLhs, typename RefStringTypeRhs>
+    // NOLINTNEXTLINE(readability-redundant-declaration)
+    friend bool operator==(const json_pointer<RefStringTypeLhs>& lhs,
+                           const json_pointer<RefStringTypeRhs>& rhs) noexcept;
+
+    /// @brief compares JSON pointer and string for equality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_eq/
+    template<typename RefStringTypeLhs, typename StringType>
+    // NOLINTNEXTLINE(readability-redundant-declaration)
+    friend bool operator==(const json_pointer<RefStringTypeLhs>& lhs,
+                           const StringType& rhs);
+
+    /// @brief compares string and JSON pointer for equality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_eq/
+    template<typename RefStringTypeRhs, typename StringType>
+    // NOLINTNEXTLINE(readability-redundant-declaration)
+    friend bool operator==(const StringType& lhs,
+                           const json_pointer<RefStringTypeRhs>& rhs);
+
+    /// @brief compares two JSON pointers for inequality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_ne/
+    template<typename RefStringTypeLhs, typename RefStringTypeRhs>
+    // NOLINTNEXTLINE(readability-redundant-declaration)
+    friend bool operator!=(const json_pointer<RefStringTypeLhs>& lhs,
+                           const json_pointer<RefStringTypeRhs>& rhs) noexcept;
+
+    /// @brief compares JSON pointer and string for inequality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_ne/
+    template<typename RefStringTypeLhs, typename StringType>
+    // NOLINTNEXTLINE(readability-redundant-declaration)
+    friend bool operator!=(const json_pointer<RefStringTypeLhs>& lhs,
+                           const StringType& rhs);
+
+    /// @brief compares string and JSON pointer for inequality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_ne/
+    template<typename RefStringTypeRhs, typename StringType>
+    // NOLINTNEXTLINE(readability-redundant-declaration)
+    friend bool operator!=(const StringType& lhs,
+                           const json_pointer<RefStringTypeRhs>& rhs);
+
+    /// @brief compares two JSON pointer for less-than
+    template<typename RefStringTypeLhs, typename RefStringTypeRhs>
+    // NOLINTNEXTLINE(readability-redundant-declaration)
+    friend bool operator<(const json_pointer<RefStringTypeLhs>& lhs,
+                          const json_pointer<RefStringTypeRhs>& rhs) noexcept;
+#endif
+
+  private:
+    /// the reference tokens
+    std::vector<string_t> reference_tokens;
+};
+
+#if !JSON_HAS_THREE_WAY_COMPARISON
+// functions cannot be defined inside class due to ODR violations
+template<typename RefStringTypeLhs, typename RefStringTypeRhs>
+inline bool operator==(const json_pointer<RefStringTypeLhs>& lhs,
+                       const json_pointer<RefStringTypeRhs>& rhs) noexcept
+{
+    return lhs.reference_tokens == rhs.reference_tokens;
+}
+
+template<typename RefStringTypeLhs,
+         typename StringType = typename json_pointer<RefStringTypeLhs>::string_t>
+JSON_HEDLEY_DEPRECATED_FOR(3.11.2, operator==(json_pointer, json_pointer))
+inline bool operator==(const json_pointer<RefStringTypeLhs>& lhs,
+                       const StringType& rhs)
+{
+    return lhs == json_pointer<RefStringTypeLhs>(rhs);
+}
+
+template<typename RefStringTypeRhs,
+         typename StringType = typename json_pointer<RefStringTypeRhs>::string_t>
+JSON_HEDLEY_DEPRECATED_FOR(3.11.2, operator==(json_pointer, json_pointer))
+inline bool operator==(const StringType& lhs,
+                       const json_pointer<RefStringTypeRhs>& rhs)
+{
+    return json_pointer<RefStringTypeRhs>(lhs) == rhs;
+}
+
+template<typename RefStringTypeLhs, typename RefStringTypeRhs>
+inline bool operator!=(const json_pointer<RefStringTypeLhs>& lhs,
+                       const json_pointer<RefStringTypeRhs>& rhs) noexcept
+{
+    return !(lhs == rhs);
+}
+
+template<typename RefStringTypeLhs,
+         typename StringType = typename json_pointer<RefStringTypeLhs>::string_t>
+JSON_HEDLEY_DEPRECATED_FOR(3.11.2, operator!=(json_pointer, json_pointer))
+inline bool operator!=(const json_pointer<RefStringTypeLhs>& lhs,
+                       const StringType& rhs)
+{
+    return !(lhs == rhs);
+}
+
+template<typename RefStringTypeRhs,
+         typename StringType = typename json_pointer<RefStringTypeRhs>::string_t>
+JSON_HEDLEY_DEPRECATED_FOR(3.11.2, operator!=(json_pointer, json_pointer))
+inline bool operator!=(const StringType& lhs,
+                       const json_pointer<RefStringTypeRhs>& rhs)
+{
+    return !(lhs == rhs);
+}
+
+template<typename RefStringTypeLhs, typename RefStringTypeRhs>
+inline bool operator<(const json_pointer<RefStringTypeLhs>& lhs,
+                      const json_pointer<RefStringTypeRhs>& rhs) noexcept
+{
+    return lhs.reference_tokens < rhs.reference_tokens;
+}
+#endif
+
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/json_ref.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <initializer_list>
+#include <utility>
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+template<typename BasicJsonType>
+class json_ref
+{
+  public:
+    using value_type = BasicJsonType;
+
+    json_ref(value_type&& value)
+        : owned_value(std::move(value))
+    {}
+
+    json_ref(const value_type& value)
+        : value_ref(&value)
+    {}
+
+    json_ref(std::initializer_list<json_ref> init)
+        : owned_value(init)
+    {}
+
+    template <
+        class... Args,
+        enable_if_t<std::is_constructible<value_type, Args...>::value, int> = 0 >
+    json_ref(Args && ... args)
+        : owned_value(std::forward<Args>(args)...)
+    {}
+
+    // class should be movable only
+    json_ref(json_ref&&) noexcept = default;
+    json_ref(const json_ref&) = delete;
+    json_ref& operator=(const json_ref&) = delete;
+    json_ref& operator=(json_ref&&) = delete;
+    ~json_ref() = default;
+
+    value_type moved_or_copied() const
+    {
+        if (value_ref == nullptr)
+        {
+            return std::move(owned_value);
+        }
+        return *value_ref;
+    }
+
+    value_type const& operator*() const
+    {
+        return value_ref ? *value_ref : owned_value;
+    }
+
+    value_type const* operator->() const
+    {
+        return &** this;
+    }
+
+  private:
+    mutable value_type owned_value = nullptr;
+    value_type const* value_ref = nullptr;
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+// #include <nlohmann/detail/string_escape.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/output/binary_writer.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <algorithm> // reverse
+#include <array> // array
+#include <map> // map
+#include <cmath> // isnan, isinf
+#include <cstdint> // uint8_t, uint16_t, uint32_t, uint64_t
+#include <cstring> // memcpy
+#include <limits> // numeric_limits
+#include <string> // string
+#include <utility> // move
+#include <vector> // vector
+
+// #include <nlohmann/detail/input/binary_reader.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/output/output_adapters.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <algorithm> // copy
+#include <cstddef> // size_t
+#include <iterator> // back_inserter
+#include <memory> // shared_ptr, make_shared
+#include <string> // basic_string
+#include <vector> // vector
+
+#ifndef JSON_NO_IO
+    #include <ios>      // streamsize
+    #include <ostream>  // basic_ostream
+#endif  // JSON_NO_IO
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/// abstract output adapter interface
+template<typename CharType> struct output_adapter_protocol
+{
+    virtual void write_character(CharType c) = 0;
+    virtual void write_characters(const CharType* s, std::size_t length) = 0;
+    virtual ~output_adapter_protocol() = default;
+
+    output_adapter_protocol() = default;
+    output_adapter_protocol(const output_adapter_protocol&) = default;
+    output_adapter_protocol(output_adapter_protocol&&) noexcept = default;
+    output_adapter_protocol& operator=(const output_adapter_protocol&) = default;
+    output_adapter_protocol& operator=(output_adapter_protocol&&) noexcept = default;
+};
+
+/// a type to simplify interfaces
+template<typename CharType>
+using output_adapter_t = std::shared_ptr<output_adapter_protocol<CharType>>;
+
+/// output adapter for byte vectors
+template<typename CharType, typename AllocatorType = std::allocator<CharType>>
+class output_vector_adapter : public output_adapter_protocol<CharType>
+{
+  public:
+    explicit output_vector_adapter(std::vector<CharType, AllocatorType>& vec) noexcept
+        : v(vec)
+    {}
+
+    void write_character(CharType c) override
+    {
+        v.push_back(c);
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    void write_characters(const CharType* s, std::size_t length) override
+    {
+        v.insert(v.end(), s, s + length);
+    }
+
+  private:
+    std::vector<CharType, AllocatorType>& v;
+};
+
+#ifndef JSON_NO_IO
+/// output adapter for output streams
+template<typename CharType>
+class output_stream_adapter : public output_adapter_protocol<CharType>
+{
+  public:
+    explicit output_stream_adapter(std::basic_ostream<CharType>& s) noexcept
+        : stream(s)
+    {}
+
+    void write_character(CharType c) override
+    {
+        stream.put(c);
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    void write_characters(const CharType* s, std::size_t length) override
+    {
+        stream.write(s, static_cast<std::streamsize>(length));
+    }
+
+  private:
+    std::basic_ostream<CharType>& stream;
+};
+#endif  // JSON_NO_IO
+
+/// output adapter for basic_string
+template<typename CharType, typename StringType = std::basic_string<CharType>>
+class output_string_adapter : public output_adapter_protocol<CharType>
+{
+  public:
+    explicit output_string_adapter(StringType& s) noexcept
+        : str(s)
+    {}
+
+    void write_character(CharType c) override
+    {
+        str.push_back(c);
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    void write_characters(const CharType* s, std::size_t length) override
+    {
+        str.append(s, length);
+    }
+
+  private:
+    StringType& str;
+};
+
+template<typename CharType, typename StringType = std::basic_string<CharType>>
+class output_adapter
+{
+  public:
+    template<typename AllocatorType = std::allocator<CharType>>
+    output_adapter(std::vector<CharType, AllocatorType>& vec)
+        : oa(std::make_shared<output_vector_adapter<CharType, AllocatorType>>(vec)) {}
+
+#ifndef JSON_NO_IO
+    output_adapter(std::basic_ostream<CharType>& s)
+        : oa(std::make_shared<output_stream_adapter<CharType>>(s)) {}
+#endif  // JSON_NO_IO
+
+    output_adapter(StringType& s)
+        : oa(std::make_shared<output_string_adapter<CharType, StringType>>(s)) {}
+
+    operator output_adapter_t<CharType>()
+    {
+        return oa;
+    }
+
+  private:
+    output_adapter_t<CharType> oa = nullptr;
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+///////////////////
+// binary writer //
+///////////////////
+
+/*!
+@brief serialization to CBOR and MessagePack values
+*/
+template<typename BasicJsonType, typename CharType>
+class binary_writer
+{
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+
+  public:
+    /*!
+    @brief create a binary writer
+
+    @param[in] adapter  output adapter to write to
+    */
+    explicit binary_writer(output_adapter_t<CharType> adapter) : oa(std::move(adapter))
+    {
+        JSON_ASSERT(oa);
+    }
+
+    /*!
+    @param[in] j  JSON value to serialize
+    @pre       j.type() == value_t::object
+    */
+    void write_bson(const BasicJsonType& j)
+    {
+        switch (j.type())
+        {
+            case value_t::object:
+            {
+                write_bson_object(*j.m_data.m_value.object);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::array:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                JSON_THROW(type_error::create(317, concat("to serialize to BSON, top-level type must be object, but is ", j.type_name()), &j));
+            }
+        }
+    }
+
+    /*!
+    @param[in] j  JSON value to serialize
+    */
+    void write_cbor(const BasicJsonType& j)
+    {
+        switch (j.type())
+        {
+            case value_t::null:
+            {
+                oa->write_character(to_char_type(0xF6));
+                break;
+            }
+
+            case value_t::boolean:
+            {
+                oa->write_character(j.m_data.m_value.boolean
+                                    ? to_char_type(0xF5)
+                                    : to_char_type(0xF4));
+                break;
+            }
+
+            case value_t::number_integer:
+            {
+                if (j.m_data.m_value.number_integer >= 0)
+                {
+                    // CBOR does not differentiate between positive signed
+                    // integers and unsigned integers. Therefore, we used the
+                    // code from the value_t::number_unsigned case here.
+                    if (j.m_data.m_value.number_integer <= 0x17)
+                    {
+                        write_number(static_cast<std::uint8_t>(j.m_data.m_value.number_integer));
+                    }
+                    else if (j.m_data.m_value.number_integer <= (std::numeric_limits<std::uint8_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x18));
+                        write_number(static_cast<std::uint8_t>(j.m_data.m_value.number_integer));
+                    }
+                    else if (j.m_data.m_value.number_integer <= (std::numeric_limits<std::uint16_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x19));
+                        write_number(static_cast<std::uint16_t>(j.m_data.m_value.number_integer));
+                    }
+                    else if (j.m_data.m_value.number_integer <= (std::numeric_limits<std::uint32_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x1A));
+                        write_number(static_cast<std::uint32_t>(j.m_data.m_value.number_integer));
+                    }
+                    else
+                    {
+                        oa->write_character(to_char_type(0x1B));
+                        write_number(static_cast<std::uint64_t>(j.m_data.m_value.number_integer));
+                    }
+                }
+                else
+                {
+                    // The conversions below encode the sign in the first
+                    // byte, and the value is converted to a positive number.
+                    const auto positive_number = -1 - j.m_data.m_value.number_integer;
+                    if (j.m_data.m_value.number_integer >= -24)
+                    {
+                        write_number(static_cast<std::uint8_t>(0x20 + positive_number));
+                    }
+                    else if (positive_number <= (std::numeric_limits<std::uint8_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x38));
+                        write_number(static_cast<std::uint8_t>(positive_number));
+                    }
+                    else if (positive_number <= (std::numeric_limits<std::uint16_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x39));
+                        write_number(static_cast<std::uint16_t>(positive_number));
+                    }
+                    else if (positive_number <= (std::numeric_limits<std::uint32_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x3A));
+                        write_number(static_cast<std::uint32_t>(positive_number));
+                    }
+                    else
+                    {
+                        oa->write_character(to_char_type(0x3B));
+                        write_number(static_cast<std::uint64_t>(positive_number));
+                    }
+                }
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                if (j.m_data.m_value.number_unsigned <= 0x17)
+                {
+                    write_number(static_cast<std::uint8_t>(j.m_data.m_value.number_unsigned));
+                }
+                else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x18));
+                    write_number(static_cast<std::uint8_t>(j.m_data.m_value.number_unsigned));
+                }
+                else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x19));
+                    write_number(static_cast<std::uint16_t>(j.m_data.m_value.number_unsigned));
+                }
+                else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x1A));
+                    write_number(static_cast<std::uint32_t>(j.m_data.m_value.number_unsigned));
+                }
+                else
+                {
+                    oa->write_character(to_char_type(0x1B));
+                    write_number(static_cast<std::uint64_t>(j.m_data.m_value.number_unsigned));
+                }
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                if (std::isnan(j.m_data.m_value.number_float))
+                {
+                    // NaN is 0xf97e00 in CBOR
+                    oa->write_character(to_char_type(0xF9));
+                    oa->write_character(to_char_type(0x7E));
+                    oa->write_character(to_char_type(0x00));
+                }
+                else if (std::isinf(j.m_data.m_value.number_float))
+                {
+                    // Infinity is 0xf97c00, -Infinity is 0xf9fc00
+                    oa->write_character(to_char_type(0xf9));
+                    oa->write_character(j.m_data.m_value.number_float > 0 ? to_char_type(0x7C) : to_char_type(0xFC));
+                    oa->write_character(to_char_type(0x00));
+                }
+                else
+                {
+                    write_compact_float(j.m_data.m_value.number_float, detail::input_format_t::cbor);
+                }
+                break;
+            }
+
+            case value_t::string:
+            {
+                // step 1: write control byte and the string length
+                const auto N = j.m_data.m_value.string->size();
+                if (N <= 0x17)
+                {
+                    write_number(static_cast<std::uint8_t>(0x60 + N));
+                }
+                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x78));
+                    write_number(static_cast<std::uint8_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x79));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x7A));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+                // LCOV_EXCL_START
+                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x7B));
+                    write_number(static_cast<std::uint64_t>(N));
+                }
+                // LCOV_EXCL_STOP
+
+                // step 2: write the string
+                oa->write_characters(
+                    reinterpret_cast<const CharType*>(j.m_data.m_value.string->c_str()),
+                    j.m_data.m_value.string->size());
+                break;
+            }
+
+            case value_t::array:
+            {
+                // step 1: write control byte and the array size
+                const auto N = j.m_data.m_value.array->size();
+                if (N <= 0x17)
+                {
+                    write_number(static_cast<std::uint8_t>(0x80 + N));
+                }
+                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x98));
+                    write_number(static_cast<std::uint8_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x99));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x9A));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+                // LCOV_EXCL_START
+                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x9B));
+                    write_number(static_cast<std::uint64_t>(N));
+                }
+                // LCOV_EXCL_STOP
+
+                // step 2: write each element
+                for (const auto& el : *j.m_data.m_value.array)
+                {
+                    write_cbor(el);
+                }
+                break;
+            }
+
+            case value_t::binary:
+            {
+                if (j.m_data.m_value.binary->has_subtype())
+                {
+                    if (j.m_data.m_value.binary->subtype() <= (std::numeric_limits<std::uint8_t>::max)())
+                    {
+                        write_number(static_cast<std::uint8_t>(0xd8));
+                        write_number(static_cast<std::uint8_t>(j.m_data.m_value.binary->subtype()));
+                    }
+                    else if (j.m_data.m_value.binary->subtype() <= (std::numeric_limits<std::uint16_t>::max)())
+                    {
+                        write_number(static_cast<std::uint8_t>(0xd9));
+                        write_number(static_cast<std::uint16_t>(j.m_data.m_value.binary->subtype()));
+                    }
+                    else if (j.m_data.m_value.binary->subtype() <= (std::numeric_limits<std::uint32_t>::max)())
+                    {
+                        write_number(static_cast<std::uint8_t>(0xda));
+                        write_number(static_cast<std::uint32_t>(j.m_data.m_value.binary->subtype()));
+                    }
+                    else if (j.m_data.m_value.binary->subtype() <= (std::numeric_limits<std::uint64_t>::max)())
+                    {
+                        write_number(static_cast<std::uint8_t>(0xdb));
+                        write_number(static_cast<std::uint64_t>(j.m_data.m_value.binary->subtype()));
+                    }
+                }
+
+                // step 1: write control byte and the binary array size
+                const auto N = j.m_data.m_value.binary->size();
+                if (N <= 0x17)
+                {
+                    write_number(static_cast<std::uint8_t>(0x40 + N));
+                }
+                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x58));
+                    write_number(static_cast<std::uint8_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x59));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x5A));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+                // LCOV_EXCL_START
+                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x5B));
+                    write_number(static_cast<std::uint64_t>(N));
+                }
+                // LCOV_EXCL_STOP
+
+                // step 2: write each element
+                oa->write_characters(
+                    reinterpret_cast<const CharType*>(j.m_data.m_value.binary->data()),
+                    N);
+
+                break;
+            }
+
+            case value_t::object:
+            {
+                // step 1: write control byte and the object size
+                const auto N = j.m_data.m_value.object->size();
+                if (N <= 0x17)
+                {
+                    write_number(static_cast<std::uint8_t>(0xA0 + N));
+                }
+                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    oa->write_character(to_char_type(0xB8));
+                    write_number(static_cast<std::uint8_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    oa->write_character(to_char_type(0xB9));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    oa->write_character(to_char_type(0xBA));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+                // LCOV_EXCL_START
+                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    oa->write_character(to_char_type(0xBB));
+                    write_number(static_cast<std::uint64_t>(N));
+                }
+                // LCOV_EXCL_STOP
+
+                // step 2: write each element
+                for (const auto& el : *j.m_data.m_value.object)
+                {
+                    write_cbor(el.first);
+                    write_cbor(el.second);
+                }
+                break;
+            }
+
+            case value_t::discarded:
+            default:
+                break;
+        }
+    }
+
+    /*!
+    @param[in] j  JSON value to serialize
+    */
+    void write_msgpack(const BasicJsonType& j)
+    {
+        switch (j.type())
+        {
+            case value_t::null: // nil
+            {
+                oa->write_character(to_char_type(0xC0));
+                break;
+            }
+
+            case value_t::boolean: // true and false
+            {
+                oa->write_character(j.m_data.m_value.boolean
+                                    ? to_char_type(0xC3)
+                                    : to_char_type(0xC2));
+                break;
+            }
+
+            case value_t::number_integer:
+            {
+                if (j.m_data.m_value.number_integer >= 0)
+                {
+                    // MessagePack does not differentiate between positive
+                    // signed integers and unsigned integers. Therefore, we used
+                    // the code from the value_t::number_unsigned case here.
+                    if (j.m_data.m_value.number_unsigned < 128)
+                    {
+                        // positive fixnum
+                        write_number(static_cast<std::uint8_t>(j.m_data.m_value.number_integer));
+                    }
+                    else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits<std::uint8_t>::max)())
+                    {
+                        // uint 8
+                        oa->write_character(to_char_type(0xCC));
+                        write_number(static_cast<std::uint8_t>(j.m_data.m_value.number_integer));
+                    }
+                    else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits<std::uint16_t>::max)())
+                    {
+                        // uint 16
+                        oa->write_character(to_char_type(0xCD));
+                        write_number(static_cast<std::uint16_t>(j.m_data.m_value.number_integer));
+                    }
+                    else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits<std::uint32_t>::max)())
+                    {
+                        // uint 32
+                        oa->write_character(to_char_type(0xCE));
+                        write_number(static_cast<std::uint32_t>(j.m_data.m_value.number_integer));
+                    }
+                    else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits<std::uint64_t>::max)())
+                    {
+                        // uint 64
+                        oa->write_character(to_char_type(0xCF));
+                        write_number(static_cast<std::uint64_t>(j.m_data.m_value.number_integer));
+                    }
+                }
+                else
+                {
+                    if (j.m_data.m_value.number_integer >= -32)
+                    {
+                        // negative fixnum
+                        write_number(static_cast<std::int8_t>(j.m_data.m_value.number_integer));
+                    }
+                    else if (j.m_data.m_value.number_integer >= (std::numeric_limits<std::int8_t>::min)() &&
+                             j.m_data.m_value.number_integer <= (std::numeric_limits<std::int8_t>::max)())
+                    {
+                        // int 8
+                        oa->write_character(to_char_type(0xD0));
+                        write_number(static_cast<std::int8_t>(j.m_data.m_value.number_integer));
+                    }
+                    else if (j.m_data.m_value.number_integer >= (std::numeric_limits<std::int16_t>::min)() &&
+                             j.m_data.m_value.number_integer <= (std::numeric_limits<std::int16_t>::max)())
+                    {
+                        // int 16
+                        oa->write_character(to_char_type(0xD1));
+                        write_number(static_cast<std::int16_t>(j.m_data.m_value.number_integer));
+                    }
+                    else if (j.m_data.m_value.number_integer >= (std::numeric_limits<std::int32_t>::min)() &&
+                             j.m_data.m_value.number_integer <= (std::numeric_limits<std::int32_t>::max)())
+                    {
+                        // int 32
+                        oa->write_character(to_char_type(0xD2));
+                        write_number(static_cast<std::int32_t>(j.m_data.m_value.number_integer));
+                    }
+                    else if (j.m_data.m_value.number_integer >= (std::numeric_limits<std::int64_t>::min)() &&
+                             j.m_data.m_value.number_integer <= (std::numeric_limits<std::int64_t>::max)())
+                    {
+                        // int 64
+                        oa->write_character(to_char_type(0xD3));
+                        write_number(static_cast<std::int64_t>(j.m_data.m_value.number_integer));
+                    }
+                }
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                if (j.m_data.m_value.number_unsigned < 128)
+                {
+                    // positive fixnum
+                    write_number(static_cast<std::uint8_t>(j.m_data.m_value.number_integer));
+                }
+                else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    // uint 8
+                    oa->write_character(to_char_type(0xCC));
+                    write_number(static_cast<std::uint8_t>(j.m_data.m_value.number_integer));
+                }
+                else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    // uint 16
+                    oa->write_character(to_char_type(0xCD));
+                    write_number(static_cast<std::uint16_t>(j.m_data.m_value.number_integer));
+                }
+                else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    // uint 32
+                    oa->write_character(to_char_type(0xCE));
+                    write_number(static_cast<std::uint32_t>(j.m_data.m_value.number_integer));
+                }
+                else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    // uint 64
+                    oa->write_character(to_char_type(0xCF));
+                    write_number(static_cast<std::uint64_t>(j.m_data.m_value.number_integer));
+                }
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                write_compact_float(j.m_data.m_value.number_float, detail::input_format_t::msgpack);
+                break;
+            }
+
+            case value_t::string:
+            {
+                // step 1: write control byte and the string length
+                const auto N = j.m_data.m_value.string->size();
+                if (N <= 31)
+                {
+                    // fixstr
+                    write_number(static_cast<std::uint8_t>(0xA0 | N));
+                }
+                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    // str 8
+                    oa->write_character(to_char_type(0xD9));
+                    write_number(static_cast<std::uint8_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    // str 16
+                    oa->write_character(to_char_type(0xDA));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    // str 32
+                    oa->write_character(to_char_type(0xDB));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+
+                // step 2: write the string
+                oa->write_characters(
+                    reinterpret_cast<const CharType*>(j.m_data.m_value.string->c_str()),
+                    j.m_data.m_value.string->size());
+                break;
+            }
+
+            case value_t::array:
+            {
+                // step 1: write control byte and the array size
+                const auto N = j.m_data.m_value.array->size();
+                if (N <= 15)
+                {
+                    // fixarray
+                    write_number(static_cast<std::uint8_t>(0x90 | N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    // array 16
+                    oa->write_character(to_char_type(0xDC));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    // array 32
+                    oa->write_character(to_char_type(0xDD));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+
+                // step 2: write each element
+                for (const auto& el : *j.m_data.m_value.array)
+                {
+                    write_msgpack(el);
+                }
+                break;
+            }
+
+            case value_t::binary:
+            {
+                // step 0: determine if the binary type has a set subtype to
+                // determine whether or not to use the ext or fixext types
+                const bool use_ext = j.m_data.m_value.binary->has_subtype();
+
+                // step 1: write control byte and the byte string length
+                const auto N = j.m_data.m_value.binary->size();
+                if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    std::uint8_t output_type{};
+                    bool fixed = true;
+                    if (use_ext)
+                    {
+                        switch (N)
+                        {
+                            case 1:
+                                output_type = 0xD4; // fixext 1
+                                break;
+                            case 2:
+                                output_type = 0xD5; // fixext 2
+                                break;
+                            case 4:
+                                output_type = 0xD6; // fixext 4
+                                break;
+                            case 8:
+                                output_type = 0xD7; // fixext 8
+                                break;
+                            case 16:
+                                output_type = 0xD8; // fixext 16
+                                break;
+                            default:
+                                output_type = 0xC7; // ext 8
+                                fixed = false;
+                                break;
+                        }
+
+                    }
+                    else
+                    {
+                        output_type = 0xC4; // bin 8
+                        fixed = false;
+                    }
+
+                    oa->write_character(to_char_type(output_type));
+                    if (!fixed)
+                    {
+                        write_number(static_cast<std::uint8_t>(N));
+                    }
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    const std::uint8_t output_type = use_ext
+                                                     ? 0xC8 // ext 16
+                                                     : 0xC5; // bin 16
+
+                    oa->write_character(to_char_type(output_type));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    const std::uint8_t output_type = use_ext
+                                                     ? 0xC9 // ext 32
+                                                     : 0xC6; // bin 32
+
+                    oa->write_character(to_char_type(output_type));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+
+                // step 1.5: if this is an ext type, write the subtype
+                if (use_ext)
+                {
+                    write_number(static_cast<std::int8_t>(j.m_data.m_value.binary->subtype()));
+                }
+
+                // step 2: write the byte string
+                oa->write_characters(
+                    reinterpret_cast<const CharType*>(j.m_data.m_value.binary->data()),
+                    N);
+
+                break;
+            }
+
+            case value_t::object:
+            {
+                // step 1: write control byte and the object size
+                const auto N = j.m_data.m_value.object->size();
+                if (N <= 15)
+                {
+                    // fixmap
+                    write_number(static_cast<std::uint8_t>(0x80 | (N & 0xF)));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    // map 16
+                    oa->write_character(to_char_type(0xDE));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    // map 32
+                    oa->write_character(to_char_type(0xDF));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+
+                // step 2: write each element
+                for (const auto& el : *j.m_data.m_value.object)
+                {
+                    write_msgpack(el.first);
+                    write_msgpack(el.second);
+                }
+                break;
+            }
+
+            case value_t::discarded:
+            default:
+                break;
+        }
+    }
+
+    /*!
+    @param[in] j  JSON value to serialize
+    @param[in] use_count   whether to use '#' prefixes (optimized format)
+    @param[in] use_type    whether to use '$' prefixes (optimized format)
+    @param[in] add_prefix  whether prefixes need to be used for this value
+    @param[in] use_bjdata  whether write in BJData format, default is false
+    */
+    void write_ubjson(const BasicJsonType& j, const bool use_count,
+                      const bool use_type, const bool add_prefix = true,
+                      const bool use_bjdata = false)
+    {
+        switch (j.type())
+        {
+            case value_t::null:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(to_char_type('Z'));
+                }
+                break;
+            }
+
+            case value_t::boolean:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(j.m_data.m_value.boolean
+                                        ? to_char_type('T')
+                                        : to_char_type('F'));
+                }
+                break;
+            }
+
+            case value_t::number_integer:
+            {
+                write_number_with_ubjson_prefix(j.m_data.m_value.number_integer, add_prefix, use_bjdata);
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                write_number_with_ubjson_prefix(j.m_data.m_value.number_unsigned, add_prefix, use_bjdata);
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                write_number_with_ubjson_prefix(j.m_data.m_value.number_float, add_prefix, use_bjdata);
+                break;
+            }
+
+            case value_t::string:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(to_char_type('S'));
+                }
+                write_number_with_ubjson_prefix(j.m_data.m_value.string->size(), true, use_bjdata);
+                oa->write_characters(
+                    reinterpret_cast<const CharType*>(j.m_data.m_value.string->c_str()),
+                    j.m_data.m_value.string->size());
+                break;
+            }
+
+            case value_t::array:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(to_char_type('['));
+                }
+
+                bool prefix_required = true;
+                if (use_type && !j.m_data.m_value.array->empty())
+                {
+                    JSON_ASSERT(use_count);
+                    const CharType first_prefix = ubjson_prefix(j.front(), use_bjdata);
+                    const bool same_prefix = std::all_of(j.begin() + 1, j.end(),
+                                                         [this, first_prefix, use_bjdata](const BasicJsonType & v)
+                    {
+                        return ubjson_prefix(v, use_bjdata) == first_prefix;
+                    });
+
+                    std::vector<CharType> bjdx = {'[', '{', 'S', 'H', 'T', 'F', 'N', 'Z'}; // excluded markers in bjdata optimized type
+
+                    if (same_prefix && !(use_bjdata && std::find(bjdx.begin(), bjdx.end(), first_prefix) != bjdx.end()))
+                    {
+                        prefix_required = false;
+                        oa->write_character(to_char_type('$'));
+                        oa->write_character(first_prefix);
+                    }
+                }
+
+                if (use_count)
+                {
+                    oa->write_character(to_char_type('#'));
+                    write_number_with_ubjson_prefix(j.m_data.m_value.array->size(), true, use_bjdata);
+                }
+
+                for (const auto& el : *j.m_data.m_value.array)
+                {
+                    write_ubjson(el, use_count, use_type, prefix_required, use_bjdata);
+                }
+
+                if (!use_count)
+                {
+                    oa->write_character(to_char_type(']'));
+                }
+
+                break;
+            }
+
+            case value_t::binary:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(to_char_type('['));
+                }
+
+                if (use_type && !j.m_data.m_value.binary->empty())
+                {
+                    JSON_ASSERT(use_count);
+                    oa->write_character(to_char_type('$'));
+                    oa->write_character('U');
+                }
+
+                if (use_count)
+                {
+                    oa->write_character(to_char_type('#'));
+                    write_number_with_ubjson_prefix(j.m_data.m_value.binary->size(), true, use_bjdata);
+                }
+
+                if (use_type)
+                {
+                    oa->write_characters(
+                        reinterpret_cast<const CharType*>(j.m_data.m_value.binary->data()),
+                        j.m_data.m_value.binary->size());
+                }
+                else
+                {
+                    for (size_t i = 0; i < j.m_data.m_value.binary->size(); ++i)
+                    {
+                        oa->write_character(to_char_type('U'));
+                        oa->write_character(j.m_data.m_value.binary->data()[i]);
+                    }
+                }
+
+                if (!use_count)
+                {
+                    oa->write_character(to_char_type(']'));
+                }
+
+                break;
+            }
+
+            case value_t::object:
+            {
+                if (use_bjdata && j.m_data.m_value.object->size() == 3 && j.m_data.m_value.object->find("_ArrayType_") != j.m_data.m_value.object->end() && j.m_data.m_value.object->find("_ArraySize_") != j.m_data.m_value.object->end() && j.m_data.m_value.object->find("_ArrayData_") != j.m_data.m_value.object->end())
+                {
+                    if (!write_bjdata_ndarray(*j.m_data.m_value.object, use_count, use_type))  // decode bjdata ndarray in the JData format (https://github.com/NeuroJSON/jdata)
+                    {
+                        break;
+                    }
+                }
+
+                if (add_prefix)
+                {
+                    oa->write_character(to_char_type('{'));
+                }
+
+                bool prefix_required = true;
+                if (use_type && !j.m_data.m_value.object->empty())
+                {
+                    JSON_ASSERT(use_count);
+                    const CharType first_prefix = ubjson_prefix(j.front(), use_bjdata);
+                    const bool same_prefix = std::all_of(j.begin(), j.end(),
+                                                         [this, first_prefix, use_bjdata](const BasicJsonType & v)
+                    {
+                        return ubjson_prefix(v, use_bjdata) == first_prefix;
+                    });
+
+                    std::vector<CharType> bjdx = {'[', '{', 'S', 'H', 'T', 'F', 'N', 'Z'}; // excluded markers in bjdata optimized type
+
+                    if (same_prefix && !(use_bjdata && std::find(bjdx.begin(), bjdx.end(), first_prefix) != bjdx.end()))
+                    {
+                        prefix_required = false;
+                        oa->write_character(to_char_type('$'));
+                        oa->write_character(first_prefix);
+                    }
+                }
+
+                if (use_count)
+                {
+                    oa->write_character(to_char_type('#'));
+                    write_number_with_ubjson_prefix(j.m_data.m_value.object->size(), true, use_bjdata);
+                }
+
+                for (const auto& el : *j.m_data.m_value.object)
+                {
+                    write_number_with_ubjson_prefix(el.first.size(), true, use_bjdata);
+                    oa->write_characters(
+                        reinterpret_cast<const CharType*>(el.first.c_str()),
+                        el.first.size());
+                    write_ubjson(el.second, use_count, use_type, prefix_required, use_bjdata);
+                }
+
+                if (!use_count)
+                {
+                    oa->write_character(to_char_type('}'));
+                }
+
+                break;
+            }
+
+            case value_t::discarded:
+            default:
+                break;
+        }
+    }
+
+  private:
+    //////////
+    // BSON //
+    //////////
+
+    /*!
+    @return The size of a BSON document entry header, including the id marker
+            and the entry name size (and its null-terminator).
+    */
+    static std::size_t calc_bson_entry_header_size(const string_t& name, const BasicJsonType& j)
+    {
+        const auto it = name.find(static_cast<typename string_t::value_type>(0));
+        if (JSON_HEDLEY_UNLIKELY(it != BasicJsonType::string_t::npos))
+        {
+            JSON_THROW(out_of_range::create(409, concat("BSON key cannot contain code point U+0000 (at byte ", std::to_string(it), ")"), &j));
+            static_cast<void>(j);
+        }
+
+        return /*id*/ 1ul + name.size() + /*zero-terminator*/1u;
+    }
+
+    /*!
+    @brief Writes the given @a element_type and @a name to the output adapter
+    */
+    void write_bson_entry_header(const string_t& name,
+                                 const std::uint8_t element_type)
+    {
+        oa->write_character(to_char_type(element_type)); // boolean
+        oa->write_characters(
+            reinterpret_cast<const CharType*>(name.c_str()),
+            name.size() + 1u);
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and boolean value @a value
+    */
+    void write_bson_boolean(const string_t& name,
+                            const bool value)
+    {
+        write_bson_entry_header(name, 0x08);
+        oa->write_character(value ? to_char_type(0x01) : to_char_type(0x00));
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and double value @a value
+    */
+    void write_bson_double(const string_t& name,
+                           const double value)
+    {
+        write_bson_entry_header(name, 0x01);
+        write_number<double>(value, true);
+    }
+
+    /*!
+    @return The size of the BSON-encoded string in @a value
+    */
+    static std::size_t calc_bson_string_size(const string_t& value)
+    {
+        return sizeof(std::int32_t) + value.size() + 1ul;
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and string value @a value
+    */
+    void write_bson_string(const string_t& name,
+                           const string_t& value)
+    {
+        write_bson_entry_header(name, 0x02);
+
+        write_number<std::int32_t>(static_cast<std::int32_t>(value.size() + 1ul), true);
+        oa->write_characters(
+            reinterpret_cast<const CharType*>(value.c_str()),
+            value.size() + 1);
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and null value
+    */
+    void write_bson_null(const string_t& name)
+    {
+        write_bson_entry_header(name, 0x0A);
+    }
+
+    /*!
+    @return The size of the BSON-encoded integer @a value
+    */
+    static std::size_t calc_bson_integer_size(const std::int64_t value)
+    {
+        return (std::numeric_limits<std::int32_t>::min)() <= value && value <= (std::numeric_limits<std::int32_t>::max)()
+               ? sizeof(std::int32_t)
+               : sizeof(std::int64_t);
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and integer @a value
+    */
+    void write_bson_integer(const string_t& name,
+                            const std::int64_t value)
+    {
+        if ((std::numeric_limits<std::int32_t>::min)() <= value && value <= (std::numeric_limits<std::int32_t>::max)())
+        {
+            write_bson_entry_header(name, 0x10); // int32
+            write_number<std::int32_t>(static_cast<std::int32_t>(value), true);
+        }
+        else
+        {
+            write_bson_entry_header(name, 0x12); // int64
+            write_number<std::int64_t>(static_cast<std::int64_t>(value), true);
+        }
+    }
+
+    /*!
+    @return The size of the BSON-encoded unsigned integer in @a j
+    */
+    static constexpr std::size_t calc_bson_unsigned_size(const std::uint64_t value) noexcept
+    {
+        return (value <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
+               ? sizeof(std::int32_t)
+               : sizeof(std::int64_t);
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and unsigned @a value
+    */
+    void write_bson_unsigned(const string_t& name,
+                             const BasicJsonType& j)
+    {
+        if (j.m_data.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
+        {
+            write_bson_entry_header(name, 0x10 /* int32 */);
+            write_number<std::int32_t>(static_cast<std::int32_t>(j.m_data.m_value.number_unsigned), true);
+        }
+        else if (j.m_data.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int64_t>::max)()))
+        {
+            write_bson_entry_header(name, 0x12 /* int64 */);
+            write_number<std::int64_t>(static_cast<std::int64_t>(j.m_data.m_value.number_unsigned), true);
+        }
+        else
+        {
+            JSON_THROW(out_of_range::create(407, concat("integer number ", std::to_string(j.m_data.m_value.number_unsigned), " cannot be represented by BSON as it does not fit int64"), &j));
+        }
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and object @a value
+    */
+    void write_bson_object_entry(const string_t& name,
+                                 const typename BasicJsonType::object_t& value)
+    {
+        write_bson_entry_header(name, 0x03); // object
+        write_bson_object(value);
+    }
+
+    /*!
+    @return The size of the BSON-encoded array @a value
+    */
+    static std::size_t calc_bson_array_size(const typename BasicJsonType::array_t& value)
+    {
+        std::size_t array_index = 0ul;
+
+        const std::size_t embedded_document_size = std::accumulate(std::begin(value), std::end(value), static_cast<std::size_t>(0), [&array_index](std::size_t result, const typename BasicJsonType::array_t::value_type & el)
+        {
+            return result + calc_bson_element_size(std::to_string(array_index++), el);
+        });
+
+        return sizeof(std::int32_t) + embedded_document_size + 1ul;
+    }
+
+    /*!
+    @return The size of the BSON-encoded binary array @a value
+    */
+    static std::size_t calc_bson_binary_size(const typename BasicJsonType::binary_t& value)
+    {
+        return sizeof(std::int32_t) + value.size() + 1ul;
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and array @a value
+    */
+    void write_bson_array(const string_t& name,
+                          const typename BasicJsonType::array_t& value)
+    {
+        write_bson_entry_header(name, 0x04); // array
+        write_number<std::int32_t>(static_cast<std::int32_t>(calc_bson_array_size(value)), true);
+
+        std::size_t array_index = 0ul;
+
+        for (const auto& el : value)
+        {
+            write_bson_element(std::to_string(array_index++), el);
+        }
+
+        oa->write_character(to_char_type(0x00));
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and binary value @a value
+    */
+    void write_bson_binary(const string_t& name,
+                           const binary_t& value)
+    {
+        write_bson_entry_header(name, 0x05);
+
+        write_number<std::int32_t>(static_cast<std::int32_t>(value.size()), true);
+        write_number(value.has_subtype() ? static_cast<std::uint8_t>(value.subtype()) : static_cast<std::uint8_t>(0x00));
+
+        oa->write_characters(reinterpret_cast<const CharType*>(value.data()), value.size());
+    }
+
+    /*!
+    @brief Calculates the size necessary to serialize the JSON value @a j with its @a name
+    @return The calculated size for the BSON document entry for @a j with the given @a name.
+    */
+    static std::size_t calc_bson_element_size(const string_t& name,
+            const BasicJsonType& j)
+    {
+        const auto header_size = calc_bson_entry_header_size(name, j);
+        switch (j.type())
+        {
+            case value_t::object:
+                return header_size + calc_bson_object_size(*j.m_data.m_value.object);
+
+            case value_t::array:
+                return header_size + calc_bson_array_size(*j.m_data.m_value.array);
+
+            case value_t::binary:
+                return header_size + calc_bson_binary_size(*j.m_data.m_value.binary);
+
+            case value_t::boolean:
+                return header_size + 1ul;
+
+            case value_t::number_float:
+                return header_size + 8ul;
+
+            case value_t::number_integer:
+                return header_size + calc_bson_integer_size(j.m_data.m_value.number_integer);
+
+            case value_t::number_unsigned:
+                return header_size + calc_bson_unsigned_size(j.m_data.m_value.number_unsigned);
+
+            case value_t::string:
+                return header_size + calc_bson_string_size(*j.m_data.m_value.string);
+
+            case value_t::null:
+                return header_size + 0ul;
+
+            // LCOV_EXCL_START
+            case value_t::discarded:
+            default:
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert)
+                return 0ul;
+                // LCOV_EXCL_STOP
+        }
+    }
+
+    /*!
+    @brief Serializes the JSON value @a j to BSON and associates it with the
+           key @a name.
+    @param name The name to associate with the JSON entity @a j within the
+                current BSON document
+    */
+    void write_bson_element(const string_t& name,
+                            const BasicJsonType& j)
+    {
+        switch (j.type())
+        {
+            case value_t::object:
+                return write_bson_object_entry(name, *j.m_data.m_value.object);
+
+            case value_t::array:
+                return write_bson_array(name, *j.m_data.m_value.array);
+
+            case value_t::binary:
+                return write_bson_binary(name, *j.m_data.m_value.binary);
+
+            case value_t::boolean:
+                return write_bson_boolean(name, j.m_data.m_value.boolean);
+
+            case value_t::number_float:
+                return write_bson_double(name, j.m_data.m_value.number_float);
+
+            case value_t::number_integer:
+                return write_bson_integer(name, j.m_data.m_value.number_integer);
+
+            case value_t::number_unsigned:
+                return write_bson_unsigned(name, j);
+
+            case value_t::string:
+                return write_bson_string(name, *j.m_data.m_value.string);
+
+            case value_t::null:
+                return write_bson_null(name);
+
+            // LCOV_EXCL_START
+            case value_t::discarded:
+            default:
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert)
+                return;
+                // LCOV_EXCL_STOP
+        }
+    }
+
+    /*!
+    @brief Calculates the size of the BSON serialization of the given
+           JSON-object @a j.
+    @param[in] value  JSON value to serialize
+    @pre       value.type() == value_t::object
+    */
+    static std::size_t calc_bson_object_size(const typename BasicJsonType::object_t& value)
+    {
+        const std::size_t document_size = std::accumulate(value.begin(), value.end(), static_cast<std::size_t>(0),
+                                          [](size_t result, const typename BasicJsonType::object_t::value_type & el)
+        {
+            return result += calc_bson_element_size(el.first, el.second);
+        });
+
+        return sizeof(std::int32_t) + document_size + 1ul;
+    }
+
+    /*!
+    @param[in] value  JSON value to serialize
+    @pre       value.type() == value_t::object
+    */
+    void write_bson_object(const typename BasicJsonType::object_t& value)
+    {
+        write_number<std::int32_t>(static_cast<std::int32_t>(calc_bson_object_size(value)), true);
+
+        for (const auto& el : value)
+        {
+            write_bson_element(el.first, el.second);
+        }
+
+        oa->write_character(to_char_type(0x00));
+    }
+
+    //////////
+    // CBOR //
+    //////////
+
+    static constexpr CharType get_cbor_float_prefix(float /*unused*/)
+    {
+        return to_char_type(0xFA);  // Single-Precision Float
+    }
+
+    static constexpr CharType get_cbor_float_prefix(double /*unused*/)
+    {
+        return to_char_type(0xFB);  // Double-Precision Float
+    }
+
+    /////////////
+    // MsgPack //
+    /////////////
+
+    static constexpr CharType get_msgpack_float_prefix(float /*unused*/)
+    {
+        return to_char_type(0xCA);  // float 32
+    }
+
+    static constexpr CharType get_msgpack_float_prefix(double /*unused*/)
+    {
+        return to_char_type(0xCB);  // float 64
+    }
+
+    ////////////
+    // UBJSON //
+    ////////////
+
+    // UBJSON: write number (floating point)
+    template<typename NumberType, typename std::enable_if<
+                 std::is_floating_point<NumberType>::value, int>::type = 0>
+    void write_number_with_ubjson_prefix(const NumberType n,
+                                         const bool add_prefix,
+                                         const bool use_bjdata)
+    {
+        if (add_prefix)
+        {
+            oa->write_character(get_ubjson_float_prefix(n));
+        }
+        write_number(n, use_bjdata);
+    }
+
+    // UBJSON: write number (unsigned integer)
+    template<typename NumberType, typename std::enable_if<
+                 std::is_unsigned<NumberType>::value, int>::type = 0>
+    void write_number_with_ubjson_prefix(const NumberType n,
+                                         const bool add_prefix,
+                                         const bool use_bjdata)
+    {
+        if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int8_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('i'));  // int8
+            }
+            write_number(static_cast<std::uint8_t>(n), use_bjdata);
+        }
+        else if (n <= (std::numeric_limits<std::uint8_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('U'));  // uint8
+            }
+            write_number(static_cast<std::uint8_t>(n), use_bjdata);
+        }
+        else if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int16_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('I'));  // int16
+            }
+            write_number(static_cast<std::int16_t>(n), use_bjdata);
+        }
+        else if (use_bjdata && n <= static_cast<uint64_t>((std::numeric_limits<uint16_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('u'));  // uint16 - bjdata only
+            }
+            write_number(static_cast<std::uint16_t>(n), use_bjdata);
+        }
+        else if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('l'));  // int32
+            }
+            write_number(static_cast<std::int32_t>(n), use_bjdata);
+        }
+        else if (use_bjdata && n <= static_cast<uint64_t>((std::numeric_limits<uint32_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('m'));  // uint32 - bjdata only
+            }
+            write_number(static_cast<std::uint32_t>(n), use_bjdata);
+        }
+        else if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int64_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('L'));  // int64
+            }
+            write_number(static_cast<std::int64_t>(n), use_bjdata);
+        }
+        else if (use_bjdata && n <= (std::numeric_limits<uint64_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('M'));  // uint64 - bjdata only
+            }
+            write_number(static_cast<std::uint64_t>(n), use_bjdata);
+        }
+        else
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('H'));  // high-precision number
+            }
+
+            const auto number = BasicJsonType(n).dump();
+            write_number_with_ubjson_prefix(number.size(), true, use_bjdata);
+            for (std::size_t i = 0; i < number.size(); ++i)
+            {
+                oa->write_character(to_char_type(static_cast<std::uint8_t>(number[i])));
+            }
+        }
+    }
+
+    // UBJSON: write number (signed integer)
+    template < typename NumberType, typename std::enable_if <
+                   std::is_signed<NumberType>::value&&
+                   !std::is_floating_point<NumberType>::value, int >::type = 0 >
+    void write_number_with_ubjson_prefix(const NumberType n,
+                                         const bool add_prefix,
+                                         const bool use_bjdata)
+    {
+        if ((std::numeric_limits<std::int8_t>::min)() <= n && n <= (std::numeric_limits<std::int8_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('i'));  // int8
+            }
+            write_number(static_cast<std::int8_t>(n), use_bjdata);
+        }
+        else if (static_cast<std::int64_t>((std::numeric_limits<std::uint8_t>::min)()) <= n && n <= static_cast<std::int64_t>((std::numeric_limits<std::uint8_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('U'));  // uint8
+            }
+            write_number(static_cast<std::uint8_t>(n), use_bjdata);
+        }
+        else if ((std::numeric_limits<std::int16_t>::min)() <= n && n <= (std::numeric_limits<std::int16_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('I'));  // int16
+            }
+            write_number(static_cast<std::int16_t>(n), use_bjdata);
+        }
+        else if (use_bjdata && (static_cast<std::int64_t>((std::numeric_limits<std::uint16_t>::min)()) <= n && n <= static_cast<std::int64_t>((std::numeric_limits<std::uint16_t>::max)())))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('u'));  // uint16 - bjdata only
+            }
+            write_number(static_cast<uint16_t>(n), use_bjdata);
+        }
+        else if ((std::numeric_limits<std::int32_t>::min)() <= n && n <= (std::numeric_limits<std::int32_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('l'));  // int32
+            }
+            write_number(static_cast<std::int32_t>(n), use_bjdata);
+        }
+        else if (use_bjdata && (static_cast<std::int64_t>((std::numeric_limits<std::uint32_t>::min)()) <= n && n <= static_cast<std::int64_t>((std::numeric_limits<std::uint32_t>::max)())))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('m'));  // uint32 - bjdata only
+            }
+            write_number(static_cast<uint32_t>(n), use_bjdata);
+        }
+        else if ((std::numeric_limits<std::int64_t>::min)() <= n && n <= (std::numeric_limits<std::int64_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('L'));  // int64
+            }
+            write_number(static_cast<std::int64_t>(n), use_bjdata);
+        }
+        // LCOV_EXCL_START
+        else
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('H'));  // high-precision number
+            }
+
+            const auto number = BasicJsonType(n).dump();
+            write_number_with_ubjson_prefix(number.size(), true, use_bjdata);
+            for (std::size_t i = 0; i < number.size(); ++i)
+            {
+                oa->write_character(to_char_type(static_cast<std::uint8_t>(number[i])));
+            }
+        }
+        // LCOV_EXCL_STOP
+    }
+
+    /*!
+    @brief determine the type prefix of container values
+    */
+    CharType ubjson_prefix(const BasicJsonType& j, const bool use_bjdata) const noexcept
+    {
+        switch (j.type())
+        {
+            case value_t::null:
+                return 'Z';
+
+            case value_t::boolean:
+                return j.m_data.m_value.boolean ? 'T' : 'F';
+
+            case value_t::number_integer:
+            {
+                if ((std::numeric_limits<std::int8_t>::min)() <= j.m_data.m_value.number_integer && j.m_data.m_value.number_integer <= (std::numeric_limits<std::int8_t>::max)())
+                {
+                    return 'i';
+                }
+                if ((std::numeric_limits<std::uint8_t>::min)() <= j.m_data.m_value.number_integer && j.m_data.m_value.number_integer <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    return 'U';
+                }
+                if ((std::numeric_limits<std::int16_t>::min)() <= j.m_data.m_value.number_integer && j.m_data.m_value.number_integer <= (std::numeric_limits<std::int16_t>::max)())
+                {
+                    return 'I';
+                }
+                if (use_bjdata && ((std::numeric_limits<std::uint16_t>::min)() <= j.m_data.m_value.number_integer && j.m_data.m_value.number_integer <= (std::numeric_limits<std::uint16_t>::max)()))
+                {
+                    return 'u';
+                }
+                if ((std::numeric_limits<std::int32_t>::min)() <= j.m_data.m_value.number_integer && j.m_data.m_value.number_integer <= (std::numeric_limits<std::int32_t>::max)())
+                {
+                    return 'l';
+                }
+                if (use_bjdata && ((std::numeric_limits<std::uint32_t>::min)() <= j.m_data.m_value.number_integer && j.m_data.m_value.number_integer <= (std::numeric_limits<std::uint32_t>::max)()))
+                {
+                    return 'm';
+                }
+                if ((std::numeric_limits<std::int64_t>::min)() <= j.m_data.m_value.number_integer && j.m_data.m_value.number_integer <= (std::numeric_limits<std::int64_t>::max)())
+                {
+                    return 'L';
+                }
+                // anything else is treated as high-precision number
+                return 'H'; // LCOV_EXCL_LINE
+            }
+
+            case value_t::number_unsigned:
+            {
+                if (j.m_data.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int8_t>::max)()))
+                {
+                    return 'i';
+                }
+                if (j.m_data.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::uint8_t>::max)()))
+                {
+                    return 'U';
+                }
+                if (j.m_data.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int16_t>::max)()))
+                {
+                    return 'I';
+                }
+                if (use_bjdata && j.m_data.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::uint16_t>::max)()))
+                {
+                    return 'u';
+                }
+                if (j.m_data.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
+                {
+                    return 'l';
+                }
+                if (use_bjdata && j.m_data.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::uint32_t>::max)()))
+                {
+                    return 'm';
+                }
+                if (j.m_data.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int64_t>::max)()))
+                {
+                    return 'L';
+                }
+                if (use_bjdata && j.m_data.m_value.number_unsigned <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    return 'M';
+                }
+                // anything else is treated as high-precision number
+                return 'H'; // LCOV_EXCL_LINE
+            }
+
+            case value_t::number_float:
+                return get_ubjson_float_prefix(j.m_data.m_value.number_float);
+
+            case value_t::string:
+                return 'S';
+
+            case value_t::array: // fallthrough
+            case value_t::binary:
+                return '[';
+
+            case value_t::object:
+                return '{';
+
+            case value_t::discarded:
+            default:  // discarded values
+                return 'N';
+        }
+    }
+
+    static constexpr CharType get_ubjson_float_prefix(float /*unused*/)
+    {
+        return 'd';  // float 32
+    }
+
+    static constexpr CharType get_ubjson_float_prefix(double /*unused*/)
+    {
+        return 'D';  // float 64
+    }
+
+    /*!
+    @return false if the object is successfully converted to a bjdata ndarray, true if the type or size is invalid
+    */
+    bool write_bjdata_ndarray(const typename BasicJsonType::object_t& value, const bool use_count, const bool use_type)
+    {
+        std::map<string_t, CharType> bjdtype = {{"uint8", 'U'},  {"int8", 'i'},  {"uint16", 'u'}, {"int16", 'I'},
+            {"uint32", 'm'}, {"int32", 'l'}, {"uint64", 'M'}, {"int64", 'L'}, {"single", 'd'}, {"double", 'D'}, {"char", 'C'}
+        };
+
+        string_t key = "_ArrayType_";
+        auto it = bjdtype.find(static_cast<string_t>(value.at(key)));
+        if (it == bjdtype.end())
+        {
+            return true;
+        }
+        CharType dtype = it->second;
+
+        key = "_ArraySize_";
+        std::size_t len = (value.at(key).empty() ? 0 : 1);
+        for (const auto& el : value.at(key))
+        {
+            len *= static_cast<std::size_t>(el.m_data.m_value.number_unsigned);
+        }
+
+        key = "_ArrayData_";
+        if (value.at(key).size() != len)
+        {
+            return true;
+        }
+
+        oa->write_character('[');
+        oa->write_character('$');
+        oa->write_character(dtype);
+        oa->write_character('#');
+
+        key = "_ArraySize_";
+        write_ubjson(value.at(key), use_count, use_type, true,  true);
+
+        key = "_ArrayData_";
+        if (dtype == 'U' || dtype == 'C')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::uint8_t>(el.m_data.m_value.number_unsigned), true);
+            }
+        }
+        else if (dtype == 'i')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::int8_t>(el.m_data.m_value.number_integer), true);
+            }
+        }
+        else if (dtype == 'u')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::uint16_t>(el.m_data.m_value.number_unsigned), true);
+            }
+        }
+        else if (dtype == 'I')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::int16_t>(el.m_data.m_value.number_integer), true);
+            }
+        }
+        else if (dtype == 'm')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::uint32_t>(el.m_data.m_value.number_unsigned), true);
+            }
+        }
+        else if (dtype == 'l')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::int32_t>(el.m_data.m_value.number_integer), true);
+            }
+        }
+        else if (dtype == 'M')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::uint64_t>(el.m_data.m_value.number_unsigned), true);
+            }
+        }
+        else if (dtype == 'L')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::int64_t>(el.m_data.m_value.number_integer), true);
+            }
+        }
+        else if (dtype == 'd')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<float>(el.m_data.m_value.number_float), true);
+            }
+        }
+        else if (dtype == 'D')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<double>(el.m_data.m_value.number_float), true);
+            }
+        }
+        return false;
+    }
+
+    ///////////////////////
+    // Utility functions //
+    ///////////////////////
+
+    /*
+    @brief write a number to output input
+    @param[in] n number of type @a NumberType
+    @param[in] OutputIsLittleEndian Set to true if output data is
+                                 required to be little endian
+    @tparam NumberType the type of the number
+
+    @note This function needs to respect the system's endianness, because bytes
+          in CBOR, MessagePack, and UBJSON are stored in network order (big
+          endian) and therefore need reordering on little endian systems.
+          On the other hand, BSON and BJData use little endian and should reorder
+          on big endian systems.
+    */
+    template<typename NumberType>
+    void write_number(const NumberType n, const bool OutputIsLittleEndian = false)
+    {
+        // step 1: write number to array of length NumberType
+        std::array<CharType, sizeof(NumberType)> vec{};
+        std::memcpy(vec.data(), &n, sizeof(NumberType));
+
+        // step 2: write array to output (with possible reordering)
+        if (is_little_endian != OutputIsLittleEndian)
+        {
+            // reverse byte order prior to conversion if necessary
+            std::reverse(vec.begin(), vec.end());
+        }
+
+        oa->write_characters(vec.data(), sizeof(NumberType));
+    }
+
+    void write_compact_float(const number_float_t n, detail::input_format_t format)
+    {
+#ifdef __GNUC__
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wfloat-equal"
+#endif
+        if (static_cast<double>(n) >= static_cast<double>(std::numeric_limits<float>::lowest()) &&
+                static_cast<double>(n) <= static_cast<double>((std::numeric_limits<float>::max)()) &&
+                static_cast<double>(static_cast<float>(n)) == static_cast<double>(n))
+        {
+            oa->write_character(format == detail::input_format_t::cbor
+                                ? get_cbor_float_prefix(static_cast<float>(n))
+                                : get_msgpack_float_prefix(static_cast<float>(n)));
+            write_number(static_cast<float>(n));
+        }
+        else
+        {
+            oa->write_character(format == detail::input_format_t::cbor
+                                ? get_cbor_float_prefix(n)
+                                : get_msgpack_float_prefix(n));
+            write_number(n);
+        }
+#ifdef __GNUC__
+#pragma GCC diagnostic pop
+#endif
+    }
+
+  public:
+    // The following to_char_type functions are implement the conversion
+    // between uint8_t and CharType. In case CharType is not unsigned,
+    // such a conversion is required to allow values greater than 128.
+    // See <https://github.com/nlohmann/json/issues/1286> for a discussion.
+    template < typename C = CharType,
+               enable_if_t < std::is_signed<C>::value && std::is_signed<char>::value > * = nullptr >
+    static constexpr CharType to_char_type(std::uint8_t x) noexcept
+    {
+        return *reinterpret_cast<char*>(&x);
+    }
+
+    template < typename C = CharType,
+               enable_if_t < std::is_signed<C>::value && std::is_unsigned<char>::value > * = nullptr >
+    static CharType to_char_type(std::uint8_t x) noexcept
+    {
+        static_assert(sizeof(std::uint8_t) == sizeof(CharType), "size of CharType must be equal to std::uint8_t");
+        static_assert(std::is_trivial<CharType>::value, "CharType must be trivial");
+        CharType result;
+        std::memcpy(&result, &x, sizeof(x));
+        return result;
+    }
+
+    template<typename C = CharType,
+             enable_if_t<std::is_unsigned<C>::value>* = nullptr>
+    static constexpr CharType to_char_type(std::uint8_t x) noexcept
+    {
+        return x;
+    }
+
+    template < typename InputCharType, typename C = CharType,
+               enable_if_t <
+                   std::is_signed<C>::value &&
+                   std::is_signed<char>::value &&
+                   std::is_same<char, typename std::remove_cv<InputCharType>::type>::value
+                   > * = nullptr >
+    static constexpr CharType to_char_type(InputCharType x) noexcept
+    {
+        return x;
+    }
+
+  private:
+    /// whether we can assume little endianness
+    const bool is_little_endian = little_endianness();
+
+    /// the output
+    output_adapter_t<CharType> oa = nullptr;
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/output/output_adapters.hpp>
+
+// #include <nlohmann/detail/output/serializer.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2008-2009 Björn Hoehrmann <bjoern@hoehrmann.de>
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <algorithm> // reverse, remove, fill, find, none_of
+#include <array> // array
+#include <clocale> // localeconv, lconv
+#include <cmath> // labs, isfinite, isnan, signbit
+#include <cstddef> // size_t, ptrdiff_t
+#include <cstdint> // uint8_t
+#include <cstdio> // snprintf
+#include <limits> // numeric_limits
+#include <string> // string, char_traits
+#include <iomanip> // setfill, setw
+#include <type_traits> // is_same
+#include <utility> // move
+
+// #include <nlohmann/detail/conversions/to_chars.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2009 Florian Loitsch <https://florian.loitsch.com/>
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <array> // array
+#include <cmath>   // signbit, isfinite
+#include <cstdint> // intN_t, uintN_t
+#include <cstring> // memcpy, memmove
+#include <limits> // numeric_limits
+#include <type_traits> // conditional
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/*!
+@brief implements the Grisu2 algorithm for binary to decimal floating-point
+conversion.
+
+This implementation is a slightly modified version of the reference
+implementation which may be obtained from
+http://florian.loitsch.com/publications (bench.tar.gz).
+
+The code is distributed under the MIT license, Copyright (c) 2009 Florian Loitsch.
+
+For a detailed description of the algorithm see:
+
+[1] Loitsch, "Printing Floating-Point Numbers Quickly and Accurately with
+    Integers", Proceedings of the ACM SIGPLAN 2010 Conference on Programming
+    Language Design and Implementation, PLDI 2010
+[2] Burger, Dybvig, "Printing Floating-Point Numbers Quickly and Accurately",
+    Proceedings of the ACM SIGPLAN 1996 Conference on Programming Language
+    Design and Implementation, PLDI 1996
+*/
+namespace dtoa_impl
+{
+
+template<typename Target, typename Source>
+Target reinterpret_bits(const Source source)
+{
+    static_assert(sizeof(Target) == sizeof(Source), "size mismatch");
+
+    Target target;
+    std::memcpy(&target, &source, sizeof(Source));
+    return target;
+}
+
+struct diyfp // f * 2^e
+{
+    static constexpr int kPrecision = 64; // = q
+
+    std::uint64_t f = 0;
+    int e = 0;
+
+    constexpr diyfp(std::uint64_t f_, int e_) noexcept : f(f_), e(e_) {}
+
+    /*!
+    @brief returns x - y
+    @pre x.e == y.e and x.f >= y.f
+    */
+    static diyfp sub(const diyfp& x, const diyfp& y) noexcept
+    {
+        JSON_ASSERT(x.e == y.e);
+        JSON_ASSERT(x.f >= y.f);
+
+        return {x.f - y.f, x.e};
+    }
+
+    /*!
+    @brief returns x * y
+    @note The result is rounded. (Only the upper q bits are returned.)
+    */
+    static diyfp mul(const diyfp& x, const diyfp& y) noexcept
+    {
+        static_assert(kPrecision == 64, "internal error");
+
+        // Computes:
+        //  f = round((x.f * y.f) / 2^q)
+        //  e = x.e + y.e + q
+
+        // Emulate the 64-bit * 64-bit multiplication:
+        //
+        // p = u * v
+        //   = (u_lo + 2^32 u_hi) (v_lo + 2^32 v_hi)
+        //   = (u_lo v_lo         ) + 2^32 ((u_lo v_hi         ) + (u_hi v_lo         )) + 2^64 (u_hi v_hi         )
+        //   = (p0                ) + 2^32 ((p1                ) + (p2                )) + 2^64 (p3                )
+        //   = (p0_lo + 2^32 p0_hi) + 2^32 ((p1_lo + 2^32 p1_hi) + (p2_lo + 2^32 p2_hi)) + 2^64 (p3                )
+        //   = (p0_lo             ) + 2^32 (p0_hi + p1_lo + p2_lo                      ) + 2^64 (p1_hi + p2_hi + p3)
+        //   = (p0_lo             ) + 2^32 (Q                                          ) + 2^64 (H                 )
+        //   = (p0_lo             ) + 2^32 (Q_lo + 2^32 Q_hi                           ) + 2^64 (H                 )
+        //
+        // (Since Q might be larger than 2^32 - 1)
+        //
+        //   = (p0_lo + 2^32 Q_lo) + 2^64 (Q_hi + H)
+        //
+        // (Q_hi + H does not overflow a 64-bit int)
+        //
+        //   = p_lo + 2^64 p_hi
+
+        const std::uint64_t u_lo = x.f & 0xFFFFFFFFu;
+        const std::uint64_t u_hi = x.f >> 32u;
+        const std::uint64_t v_lo = y.f & 0xFFFFFFFFu;
+        const std::uint64_t v_hi = y.f >> 32u;
+
+        const std::uint64_t p0 = u_lo * v_lo;
+        const std::uint64_t p1 = u_lo * v_hi;
+        const std::uint64_t p2 = u_hi * v_lo;
+        const std::uint64_t p3 = u_hi * v_hi;
+
+        const std::uint64_t p0_hi = p0 >> 32u;
+        const std::uint64_t p1_lo = p1 & 0xFFFFFFFFu;
+        const std::uint64_t p1_hi = p1 >> 32u;
+        const std::uint64_t p2_lo = p2 & 0xFFFFFFFFu;
+        const std::uint64_t p2_hi = p2 >> 32u;
+
+        std::uint64_t Q = p0_hi + p1_lo + p2_lo;
+
+        // The full product might now be computed as
+        //
+        // p_hi = p3 + p2_hi + p1_hi + (Q >> 32)
+        // p_lo = p0_lo + (Q << 32)
+        //
+        // But in this particular case here, the full p_lo is not required.
+        // Effectively we only need to add the highest bit in p_lo to p_hi (and
+        // Q_hi + 1 does not overflow).
+
+        Q += std::uint64_t{1} << (64u - 32u - 1u); // round, ties up
+
+        const std::uint64_t h = p3 + p2_hi + p1_hi + (Q >> 32u);
+
+        return {h, x.e + y.e + 64};
+    }
+
+    /*!
+    @brief normalize x such that the significand is >= 2^(q-1)
+    @pre x.f != 0
+    */
+    static diyfp normalize(diyfp x) noexcept
+    {
+        JSON_ASSERT(x.f != 0);
+
+        while ((x.f >> 63u) == 0)
+        {
+            x.f <<= 1u;
+            x.e--;
+        }
+
+        return x;
+    }
+
+    /*!
+    @brief normalize x such that the result has the exponent E
+    @pre e >= x.e and the upper e - x.e bits of x.f must be zero.
+    */
+    static diyfp normalize_to(const diyfp& x, const int target_exponent) noexcept
+    {
+        const int delta = x.e - target_exponent;
+
+        JSON_ASSERT(delta >= 0);
+        JSON_ASSERT(((x.f << delta) >> delta) == x.f);
+
+        return {x.f << delta, target_exponent};
+    }
+};
+
+struct boundaries
+{
+    diyfp w;
+    diyfp minus;
+    diyfp plus;
+};
+
+/*!
+Compute the (normalized) diyfp representing the input number 'value' and its
+boundaries.
+
+@pre value must be finite and positive
+*/
+template<typename FloatType>
+boundaries compute_boundaries(FloatType value)
+{
+    JSON_ASSERT(std::isfinite(value));
+    JSON_ASSERT(value > 0);
+
+    // Convert the IEEE representation into a diyfp.
+    //
+    // If v is denormal:
+    //      value = 0.F * 2^(1 - bias) = (          F) * 2^(1 - bias - (p-1))
+    // If v is normalized:
+    //      value = 1.F * 2^(E - bias) = (2^(p-1) + F) * 2^(E - bias - (p-1))
+
+    static_assert(std::numeric_limits<FloatType>::is_iec559,
+                  "internal error: dtoa_short requires an IEEE-754 floating-point implementation");
+
+    constexpr int      kPrecision = std::numeric_limits<FloatType>::digits; // = p (includes the hidden bit)
+    constexpr int      kBias      = std::numeric_limits<FloatType>::max_exponent - 1 + (kPrecision - 1);
+    constexpr int      kMinExp    = 1 - kBias;
+    constexpr std::uint64_t kHiddenBit = std::uint64_t{1} << (kPrecision - 1); // = 2^(p-1)
+
+    using bits_type = typename std::conditional<kPrecision == 24, std::uint32_t, std::uint64_t >::type;
+
+    const auto bits = static_cast<std::uint64_t>(reinterpret_bits<bits_type>(value));
+    const std::uint64_t E = bits >> (kPrecision - 1);
+    const std::uint64_t F = bits & (kHiddenBit - 1);
+
+    const bool is_denormal = E == 0;
+    const diyfp v = is_denormal
+                    ? diyfp(F, kMinExp)
+                    : diyfp(F + kHiddenBit, static_cast<int>(E) - kBias);
+
+    // Compute the boundaries m- and m+ of the floating-point value
+    // v = f * 2^e.
+    //
+    // Determine v- and v+, the floating-point predecessor and successor if v,
+    // respectively.
+    //
+    //      v- = v - 2^e        if f != 2^(p-1) or e == e_min                (A)
+    //         = v - 2^(e-1)    if f == 2^(p-1) and e > e_min                (B)
+    //
+    //      v+ = v + 2^e
+    //
+    // Let m- = (v- + v) / 2 and m+ = (v + v+) / 2. All real numbers _strictly_
+    // between m- and m+ round to v, regardless of how the input rounding
+    // algorithm breaks ties.
+    //
+    //      ---+-------------+-------------+-------------+-------------+---  (A)
+    //         v-            m-            v             m+            v+
+    //
+    //      -----------------+------+------+-------------+-------------+---  (B)
+    //                       v-     m-     v             m+            v+
+
+    const bool lower_boundary_is_closer = F == 0 && E > 1;
+    const diyfp m_plus = diyfp(2 * v.f + 1, v.e - 1);
+    const diyfp m_minus = lower_boundary_is_closer
+                          ? diyfp(4 * v.f - 1, v.e - 2)  // (B)
+                          : diyfp(2 * v.f - 1, v.e - 1); // (A)
+
+    // Determine the normalized w+ = m+.
+    const diyfp w_plus = diyfp::normalize(m_plus);
+
+    // Determine w- = m- such that e_(w-) = e_(w+).
+    const diyfp w_minus = diyfp::normalize_to(m_minus, w_plus.e);
+
+    return {diyfp::normalize(v), w_minus, w_plus};
+}
+
+// Given normalized diyfp w, Grisu needs to find a (normalized) cached
+// power-of-ten c, such that the exponent of the product c * w = f * 2^e lies
+// within a certain range [alpha, gamma] (Definition 3.2 from [1])
+//
+//      alpha <= e = e_c + e_w + q <= gamma
+//
+// or
+//
+//      f_c * f_w * 2^alpha <= f_c 2^(e_c) * f_w 2^(e_w) * 2^q
+//                          <= f_c * f_w * 2^gamma
+//
+// Since c and w are normalized, i.e. 2^(q-1) <= f < 2^q, this implies
+//
+//      2^(q-1) * 2^(q-1) * 2^alpha <= c * w * 2^q < 2^q * 2^q * 2^gamma
+//
+// or
+//
+//      2^(q - 2 + alpha) <= c * w < 2^(q + gamma)
+//
+// The choice of (alpha,gamma) determines the size of the table and the form of
+// the digit generation procedure. Using (alpha,gamma)=(-60,-32) works out well
+// in practice:
+//
+// The idea is to cut the number c * w = f * 2^e into two parts, which can be
+// processed independently: An integral part p1, and a fractional part p2:
+//
+//      f * 2^e = ( (f div 2^-e) * 2^-e + (f mod 2^-e) ) * 2^e
+//              = (f div 2^-e) + (f mod 2^-e) * 2^e
+//              = p1 + p2 * 2^e
+//
+// The conversion of p1 into decimal form requires a series of divisions and
+// modulos by (a power of) 10. These operations are faster for 32-bit than for
+// 64-bit integers, so p1 should ideally fit into a 32-bit integer. This can be
+// achieved by choosing
+//
+//      -e >= 32   or   e <= -32 := gamma
+//
+// In order to convert the fractional part
+//
+//      p2 * 2^e = p2 / 2^-e = d[-1] / 10^1 + d[-2] / 10^2 + ...
+//
+// into decimal form, the fraction is repeatedly multiplied by 10 and the digits
+// d[-i] are extracted in order:
+//
+//      (10 * p2) div 2^-e = d[-1]
+//      (10 * p2) mod 2^-e = d[-2] / 10^1 + ...
+//
+// The multiplication by 10 must not overflow. It is sufficient to choose
+//
+//      10 * p2 < 16 * p2 = 2^4 * p2 <= 2^64.
+//
+// Since p2 = f mod 2^-e < 2^-e,
+//
+//      -e <= 60   or   e >= -60 := alpha
+
+constexpr int kAlpha = -60;
+constexpr int kGamma = -32;
+
+struct cached_power // c = f * 2^e ~= 10^k
+{
+    std::uint64_t f;
+    int e;
+    int k;
+};
+
+/*!
+For a normalized diyfp w = f * 2^e, this function returns a (normalized) cached
+power-of-ten c = f_c * 2^e_c, such that the exponent of the product w * c
+satisfies (Definition 3.2 from [1])
+
+     alpha <= e_c + e + q <= gamma.
+*/
+inline cached_power get_cached_power_for_binary_exponent(int e)
+{
+    // Now
+    //
+    //      alpha <= e_c + e + q <= gamma                                    (1)
+    //      ==> f_c * 2^alpha <= c * 2^e * 2^q
+    //
+    // and since the c's are normalized, 2^(q-1) <= f_c,
+    //
+    //      ==> 2^(q - 1 + alpha) <= c * 2^(e + q)
+    //      ==> 2^(alpha - e - 1) <= c
+    //
+    // If c were an exact power of ten, i.e. c = 10^k, one may determine k as
+    //
+    //      k = ceil( log_10( 2^(alpha - e - 1) ) )
+    //        = ceil( (alpha - e - 1) * log_10(2) )
+    //
+    // From the paper:
+    // "In theory the result of the procedure could be wrong since c is rounded,
+    //  and the computation itself is approximated [...]. In practice, however,
+    //  this simple function is sufficient."
+    //
+    // For IEEE double precision floating-point numbers converted into
+    // normalized diyfp's w = f * 2^e, with q = 64,
+    //
+    //      e >= -1022      (min IEEE exponent)
+    //           -52        (p - 1)
+    //           -52        (p - 1, possibly normalize denormal IEEE numbers)
+    //           -11        (normalize the diyfp)
+    //         = -1137
+    //
+    // and
+    //
+    //      e <= +1023      (max IEEE exponent)
+    //           -52        (p - 1)
+    //           -11        (normalize the diyfp)
+    //         = 960
+    //
+    // This binary exponent range [-1137,960] results in a decimal exponent
+    // range [-307,324]. One does not need to store a cached power for each
+    // k in this range. For each such k it suffices to find a cached power
+    // such that the exponent of the product lies in [alpha,gamma].
+    // This implies that the difference of the decimal exponents of adjacent
+    // table entries must be less than or equal to
+    //
+    //      floor( (gamma - alpha) * log_10(2) ) = 8.
+    //
+    // (A smaller distance gamma-alpha would require a larger table.)
+
+    // NB:
+    // Actually this function returns c, such that -60 <= e_c + e + 64 <= -34.
+
+    constexpr int kCachedPowersMinDecExp = -300;
+    constexpr int kCachedPowersDecStep = 8;
+
+    static constexpr std::array<cached_power, 79> kCachedPowers =
+    {
+        {
+            { 0xAB70FE17C79AC6CA, -1060, -300 },
+            { 0xFF77B1FCBEBCDC4F, -1034, -292 },
+            { 0xBE5691EF416BD60C, -1007, -284 },
+            { 0x8DD01FAD907FFC3C,  -980, -276 },
+            { 0xD3515C2831559A83,  -954, -268 },
+            { 0x9D71AC8FADA6C9B5,  -927, -260 },
+            { 0xEA9C227723EE8BCB,  -901, -252 },
+            { 0xAECC49914078536D,  -874, -244 },
+            { 0x823C12795DB6CE57,  -847, -236 },
+            { 0xC21094364DFB5637,  -821, -228 },
+            { 0x9096EA6F3848984F,  -794, -220 },
+            { 0xD77485CB25823AC7,  -768, -212 },
+            { 0xA086CFCD97BF97F4,  -741, -204 },
+            { 0xEF340A98172AACE5,  -715, -196 },
+            { 0xB23867FB2A35B28E,  -688, -188 },
+            { 0x84C8D4DFD2C63F3B,  -661, -180 },
+            { 0xC5DD44271AD3CDBA,  -635, -172 },
+            { 0x936B9FCEBB25C996,  -608, -164 },
+            { 0xDBAC6C247D62A584,  -582, -156 },
+            { 0xA3AB66580D5FDAF6,  -555, -148 },
+            { 0xF3E2F893DEC3F126,  -529, -140 },
+            { 0xB5B5ADA8AAFF80B8,  -502, -132 },
+            { 0x87625F056C7C4A8B,  -475, -124 },
+            { 0xC9BCFF6034C13053,  -449, -116 },
+            { 0x964E858C91BA2655,  -422, -108 },
+            { 0xDFF9772470297EBD,  -396, -100 },
+            { 0xA6DFBD9FB8E5B88F,  -369,  -92 },
+            { 0xF8A95FCF88747D94,  -343,  -84 },
+            { 0xB94470938FA89BCF,  -316,  -76 },
+            { 0x8A08F0F8BF0F156B,  -289,  -68 },
+            { 0xCDB02555653131B6,  -263,  -60 },
+            { 0x993FE2C6D07B7FAC,  -236,  -52 },
+            { 0xE45C10C42A2B3B06,  -210,  -44 },
+            { 0xAA242499697392D3,  -183,  -36 },
+            { 0xFD87B5F28300CA0E,  -157,  -28 },
+            { 0xBCE5086492111AEB,  -130,  -20 },
+            { 0x8CBCCC096F5088CC,  -103,  -12 },
+            { 0xD1B71758E219652C,   -77,   -4 },
+            { 0x9C40000000000000,   -50,    4 },
+            { 0xE8D4A51000000000,   -24,   12 },
+            { 0xAD78EBC5AC620000,     3,   20 },
+            { 0x813F3978F8940984,    30,   28 },
+            { 0xC097CE7BC90715B3,    56,   36 },
+            { 0x8F7E32CE7BEA5C70,    83,   44 },
+            { 0xD5D238A4ABE98068,   109,   52 },
+            { 0x9F4F2726179A2245,   136,   60 },
+            { 0xED63A231D4C4FB27,   162,   68 },
+            { 0xB0DE65388CC8ADA8,   189,   76 },
+            { 0x83C7088E1AAB65DB,   216,   84 },
+            { 0xC45D1DF942711D9A,   242,   92 },
+            { 0x924D692CA61BE758,   269,  100 },
+            { 0xDA01EE641A708DEA,   295,  108 },
+            { 0xA26DA3999AEF774A,   322,  116 },
+            { 0xF209787BB47D6B85,   348,  124 },
+            { 0xB454E4A179DD1877,   375,  132 },
+            { 0x865B86925B9BC5C2,   402,  140 },
+            { 0xC83553C5C8965D3D,   428,  148 },
+            { 0x952AB45CFA97A0B3,   455,  156 },
+            { 0xDE469FBD99A05FE3,   481,  164 },
+            { 0xA59BC234DB398C25,   508,  172 },
+            { 0xF6C69A72A3989F5C,   534,  180 },
+            { 0xB7DCBF5354E9BECE,   561,  188 },
+            { 0x88FCF317F22241E2,   588,  196 },
+            { 0xCC20CE9BD35C78A5,   614,  204 },
+            { 0x98165AF37B2153DF,   641,  212 },
+            { 0xE2A0B5DC971F303A,   667,  220 },
+            { 0xA8D9D1535CE3B396,   694,  228 },
+            { 0xFB9B7CD9A4A7443C,   720,  236 },
+            { 0xBB764C4CA7A44410,   747,  244 },
+            { 0x8BAB8EEFB6409C1A,   774,  252 },
+            { 0xD01FEF10A657842C,   800,  260 },
+            { 0x9B10A4E5E9913129,   827,  268 },
+            { 0xE7109BFBA19C0C9D,   853,  276 },
+            { 0xAC2820D9623BF429,   880,  284 },
+            { 0x80444B5E7AA7CF85,   907,  292 },
+            { 0xBF21E44003ACDD2D,   933,  300 },
+            { 0x8E679C2F5E44FF8F,   960,  308 },
+            { 0xD433179D9C8CB841,   986,  316 },
+            { 0x9E19DB92B4E31BA9,  1013,  324 },
+        }
+    };
+
+    // This computation gives exactly the same results for k as
+    //      k = ceil((kAlpha - e - 1) * 0.30102999566398114)
+    // for |e| <= 1500, but doesn't require floating-point operations.
+    // NB: log_10(2) ~= 78913 / 2^18
+    JSON_ASSERT(e >= -1500);
+    JSON_ASSERT(e <=  1500);
+    const int f = kAlpha - e - 1;
+    const int k = (f * 78913) / (1 << 18) + static_cast<int>(f > 0);
+
+    const int index = (-kCachedPowersMinDecExp + k + (kCachedPowersDecStep - 1)) / kCachedPowersDecStep;
+    JSON_ASSERT(index >= 0);
+    JSON_ASSERT(static_cast<std::size_t>(index) < kCachedPowers.size());
+
+    const cached_power cached = kCachedPowers[static_cast<std::size_t>(index)];
+    JSON_ASSERT(kAlpha <= cached.e + e + 64);
+    JSON_ASSERT(kGamma >= cached.e + e + 64);
+
+    return cached;
+}
+
+/*!
+For n != 0, returns k, such that pow10 := 10^(k-1) <= n < 10^k.
+For n == 0, returns 1 and sets pow10 := 1.
+*/
+inline int find_largest_pow10(const std::uint32_t n, std::uint32_t& pow10)
+{
+    // LCOV_EXCL_START
+    if (n >= 1000000000)
+    {
+        pow10 = 1000000000;
+        return 10;
+    }
+    // LCOV_EXCL_STOP
+    if (n >= 100000000)
+    {
+        pow10 = 100000000;
+        return  9;
+    }
+    if (n >= 10000000)
+    {
+        pow10 = 10000000;
+        return  8;
+    }
+    if (n >= 1000000)
+    {
+        pow10 = 1000000;
+        return  7;
+    }
+    if (n >= 100000)
+    {
+        pow10 = 100000;
+        return  6;
+    }
+    if (n >= 10000)
+    {
+        pow10 = 10000;
+        return  5;
+    }
+    if (n >= 1000)
+    {
+        pow10 = 1000;
+        return  4;
+    }
+    if (n >= 100)
+    {
+        pow10 = 100;
+        return  3;
+    }
+    if (n >= 10)
+    {
+        pow10 = 10;
+        return  2;
+    }
+
+    pow10 = 1;
+    return 1;
+}
+
+inline void grisu2_round(char* buf, int len, std::uint64_t dist, std::uint64_t delta,
+                         std::uint64_t rest, std::uint64_t ten_k)
+{
+    JSON_ASSERT(len >= 1);
+    JSON_ASSERT(dist <= delta);
+    JSON_ASSERT(rest <= delta);
+    JSON_ASSERT(ten_k > 0);
+
+    //               <--------------------------- delta ---->
+    //                                  <---- dist --------->
+    // --------------[------------------+-------------------]--------------
+    //               M-                 w                   M+
+    //
+    //                                  ten_k
+    //                                <------>
+    //                                       <---- rest ---->
+    // --------------[------------------+----+--------------]--------------
+    //                                  w    V
+    //                                       = buf * 10^k
+    //
+    // ten_k represents a unit-in-the-last-place in the decimal representation
+    // stored in buf.
+    // Decrement buf by ten_k while this takes buf closer to w.
+
+    // The tests are written in this order to avoid overflow in unsigned
+    // integer arithmetic.
+
+    while (rest < dist
+            && delta - rest >= ten_k
+            && (rest + ten_k < dist || dist - rest > rest + ten_k - dist))
+    {
+        JSON_ASSERT(buf[len - 1] != '0');
+        buf[len - 1]--;
+        rest += ten_k;
+    }
+}
+
+/*!
+Generates V = buffer * 10^decimal_exponent, such that M- <= V <= M+.
+M- and M+ must be normalized and share the same exponent -60 <= e <= -32.
+*/
+inline void grisu2_digit_gen(char* buffer, int& length, int& decimal_exponent,
+                             diyfp M_minus, diyfp w, diyfp M_plus)
+{
+    static_assert(kAlpha >= -60, "internal error");
+    static_assert(kGamma <= -32, "internal error");
+
+    // Generates the digits (and the exponent) of a decimal floating-point
+    // number V = buffer * 10^decimal_exponent in the range [M-, M+]. The diyfp's
+    // w, M- and M+ share the same exponent e, which satisfies alpha <= e <= gamma.
+    //
+    //               <--------------------------- delta ---->
+    //                                  <---- dist --------->
+    // --------------[------------------+-------------------]--------------
+    //               M-                 w                   M+
+    //
+    // Grisu2 generates the digits of M+ from left to right and stops as soon as
+    // V is in [M-,M+].
+
+    JSON_ASSERT(M_plus.e >= kAlpha);
+    JSON_ASSERT(M_plus.e <= kGamma);
+
+    std::uint64_t delta = diyfp::sub(M_plus, M_minus).f; // (significand of (M+ - M-), implicit exponent is e)
+    std::uint64_t dist  = diyfp::sub(M_plus, w      ).f; // (significand of (M+ - w ), implicit exponent is e)
+
+    // Split M+ = f * 2^e into two parts p1 and p2 (note: e < 0):
+    //
+    //      M+ = f * 2^e
+    //         = ((f div 2^-e) * 2^-e + (f mod 2^-e)) * 2^e
+    //         = ((p1        ) * 2^-e + (p2        )) * 2^e
+    //         = p1 + p2 * 2^e
+
+    const diyfp one(std::uint64_t{1} << -M_plus.e, M_plus.e);
+
+    auto p1 = static_cast<std::uint32_t>(M_plus.f >> -one.e); // p1 = f div 2^-e (Since -e >= 32, p1 fits into a 32-bit int.)
+    std::uint64_t p2 = M_plus.f & (one.f - 1);                    // p2 = f mod 2^-e
+
+    // 1)
+    //
+    // Generate the digits of the integral part p1 = d[n-1]...d[1]d[0]
+
+    JSON_ASSERT(p1 > 0);
+
+    std::uint32_t pow10{};
+    const int k = find_largest_pow10(p1, pow10);
+
+    //      10^(k-1) <= p1 < 10^k, pow10 = 10^(k-1)
+    //
+    //      p1 = (p1 div 10^(k-1)) * 10^(k-1) + (p1 mod 10^(k-1))
+    //         = (d[k-1]         ) * 10^(k-1) + (p1 mod 10^(k-1))
+    //
+    //      M+ = p1                                             + p2 * 2^e
+    //         = d[k-1] * 10^(k-1) + (p1 mod 10^(k-1))          + p2 * 2^e
+    //         = d[k-1] * 10^(k-1) + ((p1 mod 10^(k-1)) * 2^-e + p2) * 2^e
+    //         = d[k-1] * 10^(k-1) + (                         rest) * 2^e
+    //
+    // Now generate the digits d[n] of p1 from left to right (n = k-1,...,0)
+    //
+    //      p1 = d[k-1]...d[n] * 10^n + d[n-1]...d[0]
+    //
+    // but stop as soon as
+    //
+    //      rest * 2^e = (d[n-1]...d[0] * 2^-e + p2) * 2^e <= delta * 2^e
+
+    int n = k;
+    while (n > 0)
+    {
+        // Invariants:
+        //      M+ = buffer * 10^n + (p1 + p2 * 2^e)    (buffer = 0 for n = k)
+        //      pow10 = 10^(n-1) <= p1 < 10^n
+        //
+        const std::uint32_t d = p1 / pow10;  // d = p1 div 10^(n-1)
+        const std::uint32_t r = p1 % pow10;  // r = p1 mod 10^(n-1)
+        //
+        //      M+ = buffer * 10^n + (d * 10^(n-1) + r) + p2 * 2^e
+        //         = (buffer * 10 + d) * 10^(n-1) + (r + p2 * 2^e)
+        //
+        JSON_ASSERT(d <= 9);
+        buffer[length++] = static_cast<char>('0' + d); // buffer := buffer * 10 + d
+        //
+        //      M+ = buffer * 10^(n-1) + (r + p2 * 2^e)
+        //
+        p1 = r;
+        n--;
+        //
+        //      M+ = buffer * 10^n + (p1 + p2 * 2^e)
+        //      pow10 = 10^n
+        //
+
+        // Now check if enough digits have been generated.
+        // Compute
+        //
+        //      p1 + p2 * 2^e = (p1 * 2^-e + p2) * 2^e = rest * 2^e
+        //
+        // Note:
+        // Since rest and delta share the same exponent e, it suffices to
+        // compare the significands.
+        const std::uint64_t rest = (std::uint64_t{p1} << -one.e) + p2;
+        if (rest <= delta)
+        {
+            // V = buffer * 10^n, with M- <= V <= M+.
+
+            decimal_exponent += n;
+
+            // We may now just stop. But instead look if the buffer could be
+            // decremented to bring V closer to w.
+            //
+            // pow10 = 10^n is now 1 ulp in the decimal representation V.
+            // The rounding procedure works with diyfp's with an implicit
+            // exponent of e.
+            //
+            //      10^n = (10^n * 2^-e) * 2^e = ulp * 2^e
+            //
+            const std::uint64_t ten_n = std::uint64_t{pow10} << -one.e;
+            grisu2_round(buffer, length, dist, delta, rest, ten_n);
+
+            return;
+        }
+
+        pow10 /= 10;
+        //
+        //      pow10 = 10^(n-1) <= p1 < 10^n
+        // Invariants restored.
+    }
+
+    // 2)
+    //
+    // The digits of the integral part have been generated:
+    //
+    //      M+ = d[k-1]...d[1]d[0] + p2 * 2^e
+    //         = buffer            + p2 * 2^e
+    //
+    // Now generate the digits of the fractional part p2 * 2^e.
+    //
+    // Note:
+    // No decimal point is generated: the exponent is adjusted instead.
+    //
+    // p2 actually represents the fraction
+    //
+    //      p2 * 2^e
+    //          = p2 / 2^-e
+    //          = d[-1] / 10^1 + d[-2] / 10^2 + ...
+    //
+    // Now generate the digits d[-m] of p1 from left to right (m = 1,2,...)
+    //
+    //      p2 * 2^e = d[-1]d[-2]...d[-m] * 10^-m
+    //                      + 10^-m * (d[-m-1] / 10^1 + d[-m-2] / 10^2 + ...)
+    //
+    // using
+    //
+    //      10^m * p2 = ((10^m * p2) div 2^-e) * 2^-e + ((10^m * p2) mod 2^-e)
+    //                = (                   d) * 2^-e + (                   r)
+    //
+    // or
+    //      10^m * p2 * 2^e = d + r * 2^e
+    //
+    // i.e.
+    //
+    //      M+ = buffer + p2 * 2^e
+    //         = buffer + 10^-m * (d + r * 2^e)
+    //         = (buffer * 10^m + d) * 10^-m + 10^-m * r * 2^e
+    //
+    // and stop as soon as 10^-m * r * 2^e <= delta * 2^e
+
+    JSON_ASSERT(p2 > delta);
+
+    int m = 0;
+    for (;;)
+    {
+        // Invariant:
+        //      M+ = buffer * 10^-m + 10^-m * (d[-m-1] / 10 + d[-m-2] / 10^2 + ...) * 2^e
+        //         = buffer * 10^-m + 10^-m * (p2                                 ) * 2^e
+        //         = buffer * 10^-m + 10^-m * (1/10 * (10 * p2)                   ) * 2^e
+        //         = buffer * 10^-m + 10^-m * (1/10 * ((10*p2 div 2^-e) * 2^-e + (10*p2 mod 2^-e)) * 2^e
+        //
+        JSON_ASSERT(p2 <= (std::numeric_limits<std::uint64_t>::max)() / 10);
+        p2 *= 10;
+        const std::uint64_t d = p2 >> -one.e;     // d = (10 * p2) div 2^-e
+        const std::uint64_t r = p2 & (one.f - 1); // r = (10 * p2) mod 2^-e
+        //
+        //      M+ = buffer * 10^-m + 10^-m * (1/10 * (d * 2^-e + r) * 2^e
+        //         = buffer * 10^-m + 10^-m * (1/10 * (d + r * 2^e))
+        //         = (buffer * 10 + d) * 10^(-m-1) + 10^(-m-1) * r * 2^e
+        //
+        JSON_ASSERT(d <= 9);
+        buffer[length++] = static_cast<char>('0' + d); // buffer := buffer * 10 + d
+        //
+        //      M+ = buffer * 10^(-m-1) + 10^(-m-1) * r * 2^e
+        //
+        p2 = r;
+        m++;
+        //
+        //      M+ = buffer * 10^-m + 10^-m * p2 * 2^e
+        // Invariant restored.
+
+        // Check if enough digits have been generated.
+        //
+        //      10^-m * p2 * 2^e <= delta * 2^e
+        //              p2 * 2^e <= 10^m * delta * 2^e
+        //                    p2 <= 10^m * delta
+        delta *= 10;
+        dist  *= 10;
+        if (p2 <= delta)
+        {
+            break;
+        }
+    }
+
+    // V = buffer * 10^-m, with M- <= V <= M+.
+
+    decimal_exponent -= m;
+
+    // 1 ulp in the decimal representation is now 10^-m.
+    // Since delta and dist are now scaled by 10^m, we need to do the
+    // same with ulp in order to keep the units in sync.
+    //
+    //      10^m * 10^-m = 1 = 2^-e * 2^e = ten_m * 2^e
+    //
+    const std::uint64_t ten_m = one.f;
+    grisu2_round(buffer, length, dist, delta, p2, ten_m);
+
+    // By construction this algorithm generates the shortest possible decimal
+    // number (Loitsch, Theorem 6.2) which rounds back to w.
+    // For an input number of precision p, at least
+    //
+    //      N = 1 + ceil(p * log_10(2))
+    //
+    // decimal digits are sufficient to identify all binary floating-point
+    // numbers (Matula, "In-and-Out conversions").
+    // This implies that the algorithm does not produce more than N decimal
+    // digits.
+    //
+    //      N = 17 for p = 53 (IEEE double precision)
+    //      N = 9  for p = 24 (IEEE single precision)
+}
+
+/*!
+v = buf * 10^decimal_exponent
+len is the length of the buffer (number of decimal digits)
+The buffer must be large enough, i.e. >= max_digits10.
+*/
+JSON_HEDLEY_NON_NULL(1)
+inline void grisu2(char* buf, int& len, int& decimal_exponent,
+                   diyfp m_minus, diyfp v, diyfp m_plus)
+{
+    JSON_ASSERT(m_plus.e == m_minus.e);
+    JSON_ASSERT(m_plus.e == v.e);
+
+    //  --------(-----------------------+-----------------------)--------    (A)
+    //          m-                      v                       m+
+    //
+    //  --------------------(-----------+-----------------------)--------    (B)
+    //                      m-          v                       m+
+    //
+    // First scale v (and m- and m+) such that the exponent is in the range
+    // [alpha, gamma].
+
+    const cached_power cached = get_cached_power_for_binary_exponent(m_plus.e);
+
+    const diyfp c_minus_k(cached.f, cached.e); // = c ~= 10^-k
+
+    // The exponent of the products is = v.e + c_minus_k.e + q and is in the range [alpha,gamma]
+    const diyfp w       = diyfp::mul(v,       c_minus_k);
+    const diyfp w_minus = diyfp::mul(m_minus, c_minus_k);
+    const diyfp w_plus  = diyfp::mul(m_plus,  c_minus_k);
+
+    //  ----(---+---)---------------(---+---)---------------(---+---)----
+    //          w-                      w                       w+
+    //          = c*m-                  = c*v                   = c*m+
+    //
+    // diyfp::mul rounds its result and c_minus_k is approximated too. w, w- and
+    // w+ are now off by a small amount.
+    // In fact:
+    //
+    //      w - v * 10^k < 1 ulp
+    //
+    // To account for this inaccuracy, add resp. subtract 1 ulp.
+    //
+    //  --------+---[---------------(---+---)---------------]---+--------
+    //          w-  M-                  w                   M+  w+
+    //
+    // Now any number in [M-, M+] (bounds included) will round to w when input,
+    // regardless of how the input rounding algorithm breaks ties.
+    //
+    // And digit_gen generates the shortest possible such number in [M-, M+].
+    // Note that this does not mean that Grisu2 always generates the shortest
+    // possible number in the interval (m-, m+).
+    const diyfp M_minus(w_minus.f + 1, w_minus.e);
+    const diyfp M_plus (w_plus.f  - 1, w_plus.e );
+
+    decimal_exponent = -cached.k; // = -(-k) = k
+
+    grisu2_digit_gen(buf, len, decimal_exponent, M_minus, w, M_plus);
+}
+
+/*!
+v = buf * 10^decimal_exponent
+len is the length of the buffer (number of decimal digits)
+The buffer must be large enough, i.e. >= max_digits10.
+*/
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1)
+void grisu2(char* buf, int& len, int& decimal_exponent, FloatType value)
+{
+    static_assert(diyfp::kPrecision >= std::numeric_limits<FloatType>::digits + 3,
+                  "internal error: not enough precision");
+
+    JSON_ASSERT(std::isfinite(value));
+    JSON_ASSERT(value > 0);
+
+    // If the neighbors (and boundaries) of 'value' are always computed for double-precision
+    // numbers, all float's can be recovered using strtod (and strtof). However, the resulting
+    // decimal representations are not exactly "short".
+    //
+    // The documentation for 'std::to_chars' (https://en.cppreference.com/w/cpp/utility/to_chars)
+    // says "value is converted to a string as if by std::sprintf in the default ("C") locale"
+    // and since sprintf promotes floats to doubles, I think this is exactly what 'std::to_chars'
+    // does.
+    // On the other hand, the documentation for 'std::to_chars' requires that "parsing the
+    // representation using the corresponding std::from_chars function recovers value exactly". That
+    // indicates that single precision floating-point numbers should be recovered using
+    // 'std::strtof'.
+    //
+    // NB: If the neighbors are computed for single-precision numbers, there is a single float
+    //     (7.0385307e-26f) which can't be recovered using strtod. The resulting double precision
+    //     value is off by 1 ulp.
+#if 0 // NOLINT(readability-avoid-unconditional-preprocessor-if)
+    const boundaries w = compute_boundaries(static_cast<double>(value));
+#else
+    const boundaries w = compute_boundaries(value);
+#endif
+
+    grisu2(buf, len, decimal_exponent, w.minus, w.w, w.plus);
+}
+
+/*!
+@brief appends a decimal representation of e to buf
+@return a pointer to the element following the exponent.
+@pre -1000 < e < 1000
+*/
+JSON_HEDLEY_NON_NULL(1)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* append_exponent(char* buf, int e)
+{
+    JSON_ASSERT(e > -1000);
+    JSON_ASSERT(e <  1000);
+
+    if (e < 0)
+    {
+        e = -e;
+        *buf++ = '-';
+    }
+    else
+    {
+        *buf++ = '+';
+    }
+
+    auto k = static_cast<std::uint32_t>(e);
+    if (k < 10)
+    {
+        // Always print at least two digits in the exponent.
+        // This is for compatibility with printf("%g").
+        *buf++ = '0';
+        *buf++ = static_cast<char>('0' + k);
+    }
+    else if (k < 100)
+    {
+        *buf++ = static_cast<char>('0' + k / 10);
+        k %= 10;
+        *buf++ = static_cast<char>('0' + k);
+    }
+    else
+    {
+        *buf++ = static_cast<char>('0' + k / 100);
+        k %= 100;
+        *buf++ = static_cast<char>('0' + k / 10);
+        k %= 10;
+        *buf++ = static_cast<char>('0' + k);
+    }
+
+    return buf;
+}
+
+/*!
+@brief prettify v = buf * 10^decimal_exponent
+
+If v is in the range [10^min_exp, 10^max_exp) it will be printed in fixed-point
+notation. Otherwise it will be printed in exponential notation.
+
+@pre min_exp < 0
+@pre max_exp > 0
+*/
+JSON_HEDLEY_NON_NULL(1)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* format_buffer(char* buf, int len, int decimal_exponent,
+                           int min_exp, int max_exp)
+{
+    JSON_ASSERT(min_exp < 0);
+    JSON_ASSERT(max_exp > 0);
+
+    const int k = len;
+    const int n = len + decimal_exponent;
+
+    // v = buf * 10^(n-k)
+    // k is the length of the buffer (number of decimal digits)
+    // n is the position of the decimal point relative to the start of the buffer.
+
+    if (k <= n && n <= max_exp)
+    {
+        // digits[000]
+        // len <= max_exp + 2
+
+        std::memset(buf + k, '0', static_cast<size_t>(n) - static_cast<size_t>(k));
+        // Make it look like a floating-point number (#362, #378)
+        buf[n + 0] = '.';
+        buf[n + 1] = '0';
+        return buf + (static_cast<size_t>(n) + 2);
+    }
+
+    if (0 < n && n <= max_exp)
+    {
+        // dig.its
+        // len <= max_digits10 + 1
+
+        JSON_ASSERT(k > n);
+
+        std::memmove(buf + (static_cast<size_t>(n) + 1), buf + n, static_cast<size_t>(k) - static_cast<size_t>(n));
+        buf[n] = '.';
+        return buf + (static_cast<size_t>(k) + 1U);
+    }
+
+    if (min_exp < n && n <= 0)
+    {
+        // 0.[000]digits
+        // len <= 2 + (-min_exp - 1) + max_digits10
+
+        std::memmove(buf + (2 + static_cast<size_t>(-n)), buf, static_cast<size_t>(k));
+        buf[0] = '0';
+        buf[1] = '.';
+        std::memset(buf + 2, '0', static_cast<size_t>(-n));
+        return buf + (2U + static_cast<size_t>(-n) + static_cast<size_t>(k));
+    }
+
+    if (k == 1)
+    {
+        // dE+123
+        // len <= 1 + 5
+
+        buf += 1;
+    }
+    else
+    {
+        // d.igitsE+123
+        // len <= max_digits10 + 1 + 5
+
+        std::memmove(buf + 2, buf + 1, static_cast<size_t>(k) - 1);
+        buf[1] = '.';
+        buf += 1 + static_cast<size_t>(k);
+    }
+
+    *buf++ = 'e';
+    return append_exponent(buf, n - 1);
+}
+
+}  // namespace dtoa_impl
+
+/*!
+@brief generates a decimal representation of the floating-point number value in [first, last).
+
+The format of the resulting decimal representation is similar to printf's %g
+format. Returns an iterator pointing past-the-end of the decimal representation.
+
+@note The input number must be finite, i.e. NaN's and Inf's are not supported.
+@note The buffer must be large enough.
+@note The result is NOT null-terminated.
+*/
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+char* to_chars(char* first, const char* last, FloatType value)
+{
+    static_cast<void>(last); // maybe unused - fix warning
+    JSON_ASSERT(std::isfinite(value));
+
+    // Use signbit(value) instead of (value < 0) since signbit works for -0.
+    if (std::signbit(value))
+    {
+        value = -value;
+        *first++ = '-';
+    }
+
+#ifdef __GNUC__
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wfloat-equal"
+#endif
+    if (value == 0) // +-0
+    {
+        *first++ = '0';
+        // Make it look like a floating-point number (#362, #378)
+        *first++ = '.';
+        *first++ = '0';
+        return first;
+    }
+#ifdef __GNUC__
+#pragma GCC diagnostic pop
+#endif
+
+    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10);
+
+    // Compute v = buffer * 10^decimal_exponent.
+    // The decimal digits are stored in the buffer, which needs to be interpreted
+    // as an unsigned decimal integer.
+    // len is the length of the buffer, i.e. the number of decimal digits.
+    int len = 0;
+    int decimal_exponent = 0;
+    dtoa_impl::grisu2(first, len, decimal_exponent, value);
+
+    JSON_ASSERT(len <= std::numeric_limits<FloatType>::max_digits10);
+
+    // Format the buffer like printf("%.*g", prec, value)
+    constexpr int kMinExp = -4;
+    // Use digits10 here to increase compatibility with version 2.
+    constexpr int kMaxExp = std::numeric_limits<FloatType>::digits10;
+
+    JSON_ASSERT(last - first >= kMaxExp + 2);
+    JSON_ASSERT(last - first >= 2 + (-kMinExp - 1) + std::numeric_limits<FloatType>::max_digits10);
+    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10 + 6);
+
+    return dtoa_impl::format_buffer(first, len, decimal_exponent, kMinExp, kMaxExp);
+}
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/output/binary_writer.hpp>
+
+// #include <nlohmann/detail/output/output_adapters.hpp>
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+///////////////////
+// serialization //
+///////////////////
+
+/// how to treat decoding errors
+enum class error_handler_t
+{
+    strict,  ///< throw a type_error exception in case of invalid UTF-8
+    replace, ///< replace invalid UTF-8 sequences with U+FFFD
+    ignore   ///< ignore invalid UTF-8 sequences
+};
+
+template<typename BasicJsonType>
+class serializer
+{
+    using string_t = typename BasicJsonType::string_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using binary_char_t = typename BasicJsonType::binary_t::value_type;
+    static constexpr std::uint8_t UTF8_ACCEPT = 0;
+    static constexpr std::uint8_t UTF8_REJECT = 1;
+
+  public:
+    /*!
+    @param[in] s  output stream to serialize to
+    @param[in] ichar  indentation character to use
+    @param[in] error_handler_  how to react on decoding errors
+    */
+    serializer(output_adapter_t<char> s, const char ichar,
+               error_handler_t error_handler_ = error_handler_t::strict)
+        : o(std::move(s))
+        , loc(std::localeconv())
+        , thousands_sep(loc->thousands_sep == nullptr ? '\0' : std::char_traits<char>::to_char_type(* (loc->thousands_sep)))
+        , decimal_point(loc->decimal_point == nullptr ? '\0' : std::char_traits<char>::to_char_type(* (loc->decimal_point)))
+        , indent_char(ichar)
+        , indent_string(512, indent_char)
+        , error_handler(error_handler_)
+    {}
+
+    // delete because of pointer members
+    serializer(const serializer&) = delete;
+    serializer& operator=(const serializer&) = delete;
+    serializer(serializer&&) = delete;
+    serializer& operator=(serializer&&) = delete;
+    ~serializer() = default;
+
+    /*!
+    @brief internal implementation of the serialization function
+
+    This function is called by the public member function dump and organizes
+    the serialization internally. The indentation level is propagated as
+    additional parameter. In case of arrays and objects, the function is
+    called recursively.
+
+    - strings and object keys are escaped using `escape_string()`
+    - integer numbers are converted implicitly via `operator<<`
+    - floating-point numbers are converted to a string using `"%g"` format
+    - binary values are serialized as objects containing the subtype and the
+      byte array
+
+    @param[in] val               value to serialize
+    @param[in] pretty_print      whether the output shall be pretty-printed
+    @param[in] ensure_ascii If @a ensure_ascii is true, all non-ASCII characters
+    in the output are escaped with `\uXXXX` sequences, and the result consists
+    of ASCII characters only.
+    @param[in] indent_step       the indent level
+    @param[in] current_indent    the current indent level (only used internally)
+    */
+    void dump(const BasicJsonType& val,
+              const bool pretty_print,
+              const bool ensure_ascii,
+              const unsigned int indent_step,
+              const unsigned int current_indent = 0)
+    {
+        switch (val.m_data.m_type)
+        {
+            case value_t::object:
+            {
+                if (val.m_data.m_value.object->empty())
+                {
+                    o->write_characters("{}", 2);
+                    return;
+                }
+
+                if (pretty_print)
+                {
+                    o->write_characters("{\n", 2);
+
+                    // variable to hold indentation for recursive calls
+                    const auto new_indent = current_indent + indent_step;
+                    if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
+                    {
+                        indent_string.resize(indent_string.size() * 2, ' ');
+                    }
+
+                    // first n-1 elements
+                    auto i = val.m_data.m_value.object->cbegin();
+                    for (std::size_t cnt = 0; cnt < val.m_data.m_value.object->size() - 1; ++cnt, ++i)
+                    {
+                        o->write_characters(indent_string.c_str(), new_indent);
+                        o->write_character('\"');
+                        dump_escaped(i->first, ensure_ascii);
+                        o->write_characters("\": ", 3);
+                        dump(i->second, true, ensure_ascii, indent_step, new_indent);
+                        o->write_characters(",\n", 2);
+                    }
+
+                    // last element
+                    JSON_ASSERT(i != val.m_data.m_value.object->cend());
+                    JSON_ASSERT(std::next(i) == val.m_data.m_value.object->cend());
+                    o->write_characters(indent_string.c_str(), new_indent);
+                    o->write_character('\"');
+                    dump_escaped(i->first, ensure_ascii);
+                    o->write_characters("\": ", 3);
+                    dump(i->second, true, ensure_ascii, indent_step, new_indent);
+
+                    o->write_character('\n');
+                    o->write_characters(indent_string.c_str(), current_indent);
+                    o->write_character('}');
+                }
+                else
+                {
+                    o->write_character('{');
+
+                    // first n-1 elements
+                    auto i = val.m_data.m_value.object->cbegin();
+                    for (std::size_t cnt = 0; cnt < val.m_data.m_value.object->size() - 1; ++cnt, ++i)
+                    {
+                        o->write_character('\"');
+                        dump_escaped(i->first, ensure_ascii);
+                        o->write_characters("\":", 2);
+                        dump(i->second, false, ensure_ascii, indent_step, current_indent);
+                        o->write_character(',');
+                    }
+
+                    // last element
+                    JSON_ASSERT(i != val.m_data.m_value.object->cend());
+                    JSON_ASSERT(std::next(i) == val.m_data.m_value.object->cend());
+                    o->write_character('\"');
+                    dump_escaped(i->first, ensure_ascii);
+                    o->write_characters("\":", 2);
+                    dump(i->second, false, ensure_ascii, indent_step, current_indent);
+
+                    o->write_character('}');
+                }
+
+                return;
+            }
+
+            case value_t::array:
+            {
+                if (val.m_data.m_value.array->empty())
+                {
+                    o->write_characters("[]", 2);
+                    return;
+                }
+
+                if (pretty_print)
+                {
+                    o->write_characters("[\n", 2);
+
+                    // variable to hold indentation for recursive calls
+                    const auto new_indent = current_indent + indent_step;
+                    if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
+                    {
+                        indent_string.resize(indent_string.size() * 2, ' ');
+                    }
+
+                    // first n-1 elements
+                    for (auto i = val.m_data.m_value.array->cbegin();
+                            i != val.m_data.m_value.array->cend() - 1; ++i)
+                    {
+                        o->write_characters(indent_string.c_str(), new_indent);
+                        dump(*i, true, ensure_ascii, indent_step, new_indent);
+                        o->write_characters(",\n", 2);
+                    }
+
+                    // last element
+                    JSON_ASSERT(!val.m_data.m_value.array->empty());
+                    o->write_characters(indent_string.c_str(), new_indent);
+                    dump(val.m_data.m_value.array->back(), true, ensure_ascii, indent_step, new_indent);
+
+                    o->write_character('\n');
+                    o->write_characters(indent_string.c_str(), current_indent);
+                    o->write_character(']');
+                }
+                else
+                {
+                    o->write_character('[');
+
+                    // first n-1 elements
+                    for (auto i = val.m_data.m_value.array->cbegin();
+                            i != val.m_data.m_value.array->cend() - 1; ++i)
+                    {
+                        dump(*i, false, ensure_ascii, indent_step, current_indent);
+                        o->write_character(',');
+                    }
+
+                    // last element
+                    JSON_ASSERT(!val.m_data.m_value.array->empty());
+                    dump(val.m_data.m_value.array->back(), false, ensure_ascii, indent_step, current_indent);
+
+                    o->write_character(']');
+                }
+
+                return;
+            }
+
+            case value_t::string:
+            {
+                o->write_character('\"');
+                dump_escaped(*val.m_data.m_value.string, ensure_ascii);
+                o->write_character('\"');
+                return;
+            }
+
+            case value_t::binary:
+            {
+                if (pretty_print)
+                {
+                    o->write_characters("{\n", 2);
+
+                    // variable to hold indentation for recursive calls
+                    const auto new_indent = current_indent + indent_step;
+                    if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
+                    {
+                        indent_string.resize(indent_string.size() * 2, ' ');
+                    }
+
+                    o->write_characters(indent_string.c_str(), new_indent);
+
+                    o->write_characters("\"bytes\": [", 10);
+
+                    if (!val.m_data.m_value.binary->empty())
+                    {
+                        for (auto i = val.m_data.m_value.binary->cbegin();
+                                i != val.m_data.m_value.binary->cend() - 1; ++i)
+                        {
+                            dump_integer(*i);
+                            o->write_characters(", ", 2);
+                        }
+                        dump_integer(val.m_data.m_value.binary->back());
+                    }
+
+                    o->write_characters("],\n", 3);
+                    o->write_characters(indent_string.c_str(), new_indent);
+
+                    o->write_characters("\"subtype\": ", 11);
+                    if (val.m_data.m_value.binary->has_subtype())
+                    {
+                        dump_integer(val.m_data.m_value.binary->subtype());
+                    }
+                    else
+                    {
+                        o->write_characters("null", 4);
+                    }
+                    o->write_character('\n');
+                    o->write_characters(indent_string.c_str(), current_indent);
+                    o->write_character('}');
+                }
+                else
+                {
+                    o->write_characters("{\"bytes\":[", 10);
+
+                    if (!val.m_data.m_value.binary->empty())
+                    {
+                        for (auto i = val.m_data.m_value.binary->cbegin();
+                                i != val.m_data.m_value.binary->cend() - 1; ++i)
+                        {
+                            dump_integer(*i);
+                            o->write_character(',');
+                        }
+                        dump_integer(val.m_data.m_value.binary->back());
+                    }
+
+                    o->write_characters("],\"subtype\":", 12);
+                    if (val.m_data.m_value.binary->has_subtype())
+                    {
+                        dump_integer(val.m_data.m_value.binary->subtype());
+                        o->write_character('}');
+                    }
+                    else
+                    {
+                        o->write_characters("null}", 5);
+                    }
+                }
+                return;
+            }
+
+            case value_t::boolean:
+            {
+                if (val.m_data.m_value.boolean)
+                {
+                    o->write_characters("true", 4);
+                }
+                else
+                {
+                    o->write_characters("false", 5);
+                }
+                return;
+            }
+
+            case value_t::number_integer:
+            {
+                dump_integer(val.m_data.m_value.number_integer);
+                return;
+            }
+
+            case value_t::number_unsigned:
+            {
+                dump_integer(val.m_data.m_value.number_unsigned);
+                return;
+            }
+
+            case value_t::number_float:
+            {
+                dump_float(val.m_data.m_value.number_float);
+                return;
+            }
+
+            case value_t::discarded:
+            {
+                o->write_characters("<discarded>", 11);
+                return;
+            }
+
+            case value_t::null:
+            {
+                o->write_characters("null", 4);
+                return;
+            }
+
+            default:            // LCOV_EXCL_LINE
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        }
+    }
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    /*!
+    @brief dump escaped string
+
+    Escape a string by replacing certain special characters by a sequence of an
+    escape character (backslash) and another character and other control
+    characters by a sequence of "\u" followed by a four-digit hex
+    representation. The escaped string is written to output stream @a o.
+
+    @param[in] s  the string to escape
+    @param[in] ensure_ascii  whether to escape non-ASCII characters with
+                             \uXXXX sequences
+
+    @complexity Linear in the length of string @a s.
+    */
+    void dump_escaped(const string_t& s, const bool ensure_ascii)
+    {
+        std::uint32_t codepoint{};
+        std::uint8_t state = UTF8_ACCEPT;
+        std::size_t bytes = 0;  // number of bytes written to string_buffer
+
+        // number of bytes written at the point of the last valid byte
+        std::size_t bytes_after_last_accept = 0;
+        std::size_t undumped_chars = 0;
+
+        for (std::size_t i = 0; i < s.size(); ++i)
+        {
+            const auto byte = static_cast<std::uint8_t>(s[i]);
+
+            switch (decode(state, codepoint, byte))
+            {
+                case UTF8_ACCEPT:  // decode found a new code point
+                {
+                    switch (codepoint)
+                    {
+                        case 0x08: // backspace
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = 'b';
+                            break;
+                        }
+
+                        case 0x09: // horizontal tab
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = 't';
+                            break;
+                        }
+
+                        case 0x0A: // newline
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = 'n';
+                            break;
+                        }
+
+                        case 0x0C: // formfeed
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = 'f';
+                            break;
+                        }
+
+                        case 0x0D: // carriage return
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = 'r';
+                            break;
+                        }
+
+                        case 0x22: // quotation mark
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = '\"';
+                            break;
+                        }
+
+                        case 0x5C: // reverse solidus
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = '\\';
+                            break;
+                        }
+
+                        default:
+                        {
+                            // escape control characters (0x00..0x1F) or, if
+                            // ensure_ascii parameter is used, non-ASCII characters
+                            if ((codepoint <= 0x1F) || (ensure_ascii && (codepoint >= 0x7F)))
+                            {
+                                if (codepoint <= 0xFFFF)
+                                {
+                                    // NOLINTNEXTLINE(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+                                    static_cast<void>((std::snprintf)(string_buffer.data() + bytes, 7, "\\u%04x",
+                                                                      static_cast<std::uint16_t>(codepoint)));
+                                    bytes += 6;
+                                }
+                                else
+                                {
+                                    // NOLINTNEXTLINE(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+                                    static_cast<void>((std::snprintf)(string_buffer.data() + bytes, 13, "\\u%04x\\u%04x",
+                                                                      static_cast<std::uint16_t>(0xD7C0u + (codepoint >> 10u)),
+                                                                      static_cast<std::uint16_t>(0xDC00u + (codepoint & 0x3FFu))));
+                                    bytes += 12;
+                                }
+                            }
+                            else
+                            {
+                                // copy byte to buffer (all previous bytes
+                                // been copied have in default case above)
+                                string_buffer[bytes++] = s[i];
+                            }
+                            break;
+                        }
+                    }
+
+                    // write buffer and reset index; there must be 13 bytes
+                    // left, as this is the maximal number of bytes to be
+                    // written ("\uxxxx\uxxxx\0") for one code point
+                    if (string_buffer.size() - bytes < 13)
+                    {
+                        o->write_characters(string_buffer.data(), bytes);
+                        bytes = 0;
+                    }
+
+                    // remember the byte position of this accept
+                    bytes_after_last_accept = bytes;
+                    undumped_chars = 0;
+                    break;
+                }
+
+                case UTF8_REJECT:  // decode found invalid UTF-8 byte
+                {
+                    switch (error_handler)
+                    {
+                        case error_handler_t::strict:
+                        {
+                            JSON_THROW(type_error::create(316, concat("invalid UTF-8 byte at index ", std::to_string(i), ": 0x", hex_bytes(byte | 0)), nullptr));
+                        }
+
+                        case error_handler_t::ignore:
+                        case error_handler_t::replace:
+                        {
+                            // in case we saw this character the first time, we
+                            // would like to read it again, because the byte
+                            // may be OK for itself, but just not OK for the
+                            // previous sequence
+                            if (undumped_chars > 0)
+                            {
+                                --i;
+                            }
+
+                            // reset length buffer to the last accepted index;
+                            // thus removing/ignoring the invalid characters
+                            bytes = bytes_after_last_accept;
+
+                            if (error_handler == error_handler_t::replace)
+                            {
+                                // add a replacement character
+                                if (ensure_ascii)
+                                {
+                                    string_buffer[bytes++] = '\\';
+                                    string_buffer[bytes++] = 'u';
+                                    string_buffer[bytes++] = 'f';
+                                    string_buffer[bytes++] = 'f';
+                                    string_buffer[bytes++] = 'f';
+                                    string_buffer[bytes++] = 'd';
+                                }
+                                else
+                                {
+                                    string_buffer[bytes++] = detail::binary_writer<BasicJsonType, char>::to_char_type('\xEF');
+                                    string_buffer[bytes++] = detail::binary_writer<BasicJsonType, char>::to_char_type('\xBF');
+                                    string_buffer[bytes++] = detail::binary_writer<BasicJsonType, char>::to_char_type('\xBD');
+                                }
+
+                                // write buffer and reset index; there must be 13 bytes
+                                // left, as this is the maximal number of bytes to be
+                                // written ("\uxxxx\uxxxx\0") for one code point
+                                if (string_buffer.size() - bytes < 13)
+                                {
+                                    o->write_characters(string_buffer.data(), bytes);
+                                    bytes = 0;
+                                }
+
+                                bytes_after_last_accept = bytes;
+                            }
+
+                            undumped_chars = 0;
+
+                            // continue processing the string
+                            state = UTF8_ACCEPT;
+                            break;
+                        }
+
+                        default:            // LCOV_EXCL_LINE
+                            JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+                    }
+                    break;
+                }
+
+                default:  // decode found yet incomplete multi-byte code point
+                {
+                    if (!ensure_ascii)
+                    {
+                        // code point will not be escaped - copy byte to buffer
+                        string_buffer[bytes++] = s[i];
+                    }
+                    ++undumped_chars;
+                    break;
+                }
+            }
+        }
+
+        // we finished processing the string
+        if (JSON_HEDLEY_LIKELY(state == UTF8_ACCEPT))
+        {
+            // write buffer
+            if (bytes > 0)
+            {
+                o->write_characters(string_buffer.data(), bytes);
+            }
+        }
+        else
+        {
+            // we finish reading, but do not accept: string was incomplete
+            switch (error_handler)
+            {
+                case error_handler_t::strict:
+                {
+                    JSON_THROW(type_error::create(316, concat("incomplete UTF-8 string; last byte: 0x", hex_bytes(static_cast<std::uint8_t>(s.back() | 0))), nullptr));
+                }
+
+                case error_handler_t::ignore:
+                {
+                    // write all accepted bytes
+                    o->write_characters(string_buffer.data(), bytes_after_last_accept);
+                    break;
+                }
+
+                case error_handler_t::replace:
+                {
+                    // write all accepted bytes
+                    o->write_characters(string_buffer.data(), bytes_after_last_accept);
+                    // add a replacement character
+                    if (ensure_ascii)
+                    {
+                        o->write_characters("\\ufffd", 6);
+                    }
+                    else
+                    {
+                        o->write_characters("\xEF\xBF\xBD", 3);
+                    }
+                    break;
+                }
+
+                default:            // LCOV_EXCL_LINE
+                    JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+            }
+        }
+    }
+
+  private:
+    /*!
+    @brief count digits
+
+    Count the number of decimal (base 10) digits for an input unsigned integer.
+
+    @param[in] x  unsigned integer number to count its digits
+    @return    number of decimal digits
+    */
+    inline unsigned int count_digits(number_unsigned_t x) noexcept
+    {
+        unsigned int n_digits = 1;
+        for (;;)
+        {
+            if (x < 10)
+            {
+                return n_digits;
+            }
+            if (x < 100)
+            {
+                return n_digits + 1;
+            }
+            if (x < 1000)
+            {
+                return n_digits + 2;
+            }
+            if (x < 10000)
+            {
+                return n_digits + 3;
+            }
+            x = x / 10000u;
+            n_digits += 4;
+        }
+    }
+
+    /*!
+     * @brief convert a byte to a uppercase hex representation
+     * @param[in] byte byte to represent
+     * @return representation ("00".."FF")
+     */
+    static std::string hex_bytes(std::uint8_t byte)
+    {
+        std::string result = "FF";
+        constexpr const char* nibble_to_hex = "0123456789ABCDEF";
+        result[0] = nibble_to_hex[byte / 16];
+        result[1] = nibble_to_hex[byte % 16];
+        return result;
+    }
+
+    // templates to avoid warnings about useless casts
+    template <typename NumberType, enable_if_t<std::is_signed<NumberType>::value, int> = 0>
+    bool is_negative_number(NumberType x)
+    {
+        return x < 0;
+    }
+
+    template < typename NumberType, enable_if_t <std::is_unsigned<NumberType>::value, int > = 0 >
+    bool is_negative_number(NumberType /*unused*/)
+    {
+        return false;
+    }
+
+    /*!
+    @brief dump an integer
+
+    Dump a given integer to output stream @a o. Works internally with
+    @a number_buffer.
+
+    @param[in] x  integer number (signed or unsigned) to dump
+    @tparam NumberType either @a number_integer_t or @a number_unsigned_t
+    */
+    template < typename NumberType, detail::enable_if_t <
+                   std::is_integral<NumberType>::value ||
+                   std::is_same<NumberType, number_unsigned_t>::value ||
+                   std::is_same<NumberType, number_integer_t>::value ||
+                   std::is_same<NumberType, binary_char_t>::value,
+                   int > = 0 >
+    void dump_integer(NumberType x)
+    {
+        static constexpr std::array<std::array<char, 2>, 100> digits_to_99
+        {
+            {
+                {{'0', '0'}}, {{'0', '1'}}, {{'0', '2'}}, {{'0', '3'}}, {{'0', '4'}}, {{'0', '5'}}, {{'0', '6'}}, {{'0', '7'}}, {{'0', '8'}}, {{'0', '9'}},
+                {{'1', '0'}}, {{'1', '1'}}, {{'1', '2'}}, {{'1', '3'}}, {{'1', '4'}}, {{'1', '5'}}, {{'1', '6'}}, {{'1', '7'}}, {{'1', '8'}}, {{'1', '9'}},
+                {{'2', '0'}}, {{'2', '1'}}, {{'2', '2'}}, {{'2', '3'}}, {{'2', '4'}}, {{'2', '5'}}, {{'2', '6'}}, {{'2', '7'}}, {{'2', '8'}}, {{'2', '9'}},
+                {{'3', '0'}}, {{'3', '1'}}, {{'3', '2'}}, {{'3', '3'}}, {{'3', '4'}}, {{'3', '5'}}, {{'3', '6'}}, {{'3', '7'}}, {{'3', '8'}}, {{'3', '9'}},
+                {{'4', '0'}}, {{'4', '1'}}, {{'4', '2'}}, {{'4', '3'}}, {{'4', '4'}}, {{'4', '5'}}, {{'4', '6'}}, {{'4', '7'}}, {{'4', '8'}}, {{'4', '9'}},
+                {{'5', '0'}}, {{'5', '1'}}, {{'5', '2'}}, {{'5', '3'}}, {{'5', '4'}}, {{'5', '5'}}, {{'5', '6'}}, {{'5', '7'}}, {{'5', '8'}}, {{'5', '9'}},
+                {{'6', '0'}}, {{'6', '1'}}, {{'6', '2'}}, {{'6', '3'}}, {{'6', '4'}}, {{'6', '5'}}, {{'6', '6'}}, {{'6', '7'}}, {{'6', '8'}}, {{'6', '9'}},
+                {{'7', '0'}}, {{'7', '1'}}, {{'7', '2'}}, {{'7', '3'}}, {{'7', '4'}}, {{'7', '5'}}, {{'7', '6'}}, {{'7', '7'}}, {{'7', '8'}}, {{'7', '9'}},
+                {{'8', '0'}}, {{'8', '1'}}, {{'8', '2'}}, {{'8', '3'}}, {{'8', '4'}}, {{'8', '5'}}, {{'8', '6'}}, {{'8', '7'}}, {{'8', '8'}}, {{'8', '9'}},
+                {{'9', '0'}}, {{'9', '1'}}, {{'9', '2'}}, {{'9', '3'}}, {{'9', '4'}}, {{'9', '5'}}, {{'9', '6'}}, {{'9', '7'}}, {{'9', '8'}}, {{'9', '9'}},
+            }
+        };
+
+        // special case for "0"
+        if (x == 0)
+        {
+            o->write_character('0');
+            return;
+        }
+
+        // use a pointer to fill the buffer
+        auto buffer_ptr = number_buffer.begin(); // NOLINT(llvm-qualified-auto,readability-qualified-auto,cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+
+        number_unsigned_t abs_value;
+
+        unsigned int n_chars{};
+
+        if (is_negative_number(x))
+        {
+            *buffer_ptr = '-';
+            abs_value = remove_sign(static_cast<number_integer_t>(x));
+
+            // account one more byte for the minus sign
+            n_chars = 1 + count_digits(abs_value);
+        }
+        else
+        {
+            abs_value = static_cast<number_unsigned_t>(x);
+            n_chars = count_digits(abs_value);
+        }
+
+        // spare 1 byte for '\0'
+        JSON_ASSERT(n_chars < number_buffer.size() - 1);
+
+        // jump to the end to generate the string from backward,
+        // so we later avoid reversing the result
+        buffer_ptr += n_chars;
+
+        // Fast int2ascii implementation inspired by "Fastware" talk by Andrei Alexandrescu
+        // See: https://www.youtube.com/watch?v=o4-CwDo2zpg
+        while (abs_value >= 100)
+        {
+            const auto digits_index = static_cast<unsigned>((abs_value % 100));
+            abs_value /= 100;
+            *(--buffer_ptr) = digits_to_99[digits_index][1];
+            *(--buffer_ptr) = digits_to_99[digits_index][0];
+        }
+
+        if (abs_value >= 10)
+        {
+            const auto digits_index = static_cast<unsigned>(abs_value);
+            *(--buffer_ptr) = digits_to_99[digits_index][1];
+            *(--buffer_ptr) = digits_to_99[digits_index][0];
+        }
+        else
+        {
+            *(--buffer_ptr) = static_cast<char>('0' + abs_value);
+        }
+
+        o->write_characters(number_buffer.data(), n_chars);
+    }
+
+    /*!
+    @brief dump a floating-point number
+
+    Dump a given floating-point number to output stream @a o. Works internally
+    with @a number_buffer.
+
+    @param[in] x  floating-point number to dump
+    */
+    void dump_float(number_float_t x)
+    {
+        // NaN / inf
+        if (!std::isfinite(x))
+        {
+            o->write_characters("null", 4);
+            return;
+        }
+
+        // If number_float_t is an IEEE-754 single or double precision number,
+        // use the Grisu2 algorithm to produce short numbers which are
+        // guaranteed to round-trip, using strtof and strtod, resp.
+        //
+        // NB: The test below works if <long double> == <double>.
+        static constexpr bool is_ieee_single_or_double
+            = (std::numeric_limits<number_float_t>::is_iec559 && std::numeric_limits<number_float_t>::digits == 24 && std::numeric_limits<number_float_t>::max_exponent == 128) ||
+              (std::numeric_limits<number_float_t>::is_iec559 && std::numeric_limits<number_float_t>::digits == 53 && std::numeric_limits<number_float_t>::max_exponent == 1024);
+
+        dump_float(x, std::integral_constant<bool, is_ieee_single_or_double>());
+    }
+
+    void dump_float(number_float_t x, std::true_type /*is_ieee_single_or_double*/)
+    {
+        auto* begin = number_buffer.data();
+        auto* end = ::nlohmann::detail::to_chars(begin, begin + number_buffer.size(), x);
+
+        o->write_characters(begin, static_cast<size_t>(end - begin));
+    }
+
+    void dump_float(number_float_t x, std::false_type /*is_ieee_single_or_double*/)
+    {
+        // get number of digits for a float -> text -> float round-trip
+        static constexpr auto d = std::numeric_limits<number_float_t>::max_digits10;
+
+        // the actual conversion
+        // NOLINTNEXTLINE(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+        std::ptrdiff_t len = (std::snprintf)(number_buffer.data(), number_buffer.size(), "%.*g", d, x);
+
+        // negative value indicates an error
+        JSON_ASSERT(len > 0);
+        // check if buffer was large enough
+        JSON_ASSERT(static_cast<std::size_t>(len) < number_buffer.size());
+
+        // erase thousands separator
+        if (thousands_sep != '\0')
+        {
+            // NOLINTNEXTLINE(readability-qualified-auto,llvm-qualified-auto): std::remove returns an iterator, see https://github.com/nlohmann/json/issues/3081
+            const auto end = std::remove(number_buffer.begin(), number_buffer.begin() + len, thousands_sep);
+            std::fill(end, number_buffer.end(), '\0');
+            JSON_ASSERT((end - number_buffer.begin()) <= len);
+            len = (end - number_buffer.begin());
+        }
+
+        // convert decimal point to '.'
+        if (decimal_point != '\0' && decimal_point != '.')
+        {
+            // NOLINTNEXTLINE(readability-qualified-auto,llvm-qualified-auto): std::find returns an iterator, see https://github.com/nlohmann/json/issues/3081
+            const auto dec_pos = std::find(number_buffer.begin(), number_buffer.end(), decimal_point);
+            if (dec_pos != number_buffer.end())
+            {
+                *dec_pos = '.';
+            }
+        }
+
+        o->write_characters(number_buffer.data(), static_cast<std::size_t>(len));
+
+        // determine if we need to append ".0"
+        const bool value_is_int_like =
+            std::none_of(number_buffer.begin(), number_buffer.begin() + len + 1,
+                         [](char c)
+        {
+            return c == '.' || c == 'e';
+        });
+
+        if (value_is_int_like)
+        {
+            o->write_characters(".0", 2);
+        }
+    }
+
+    /*!
+    @brief check whether a string is UTF-8 encoded
+
+    The function checks each byte of a string whether it is UTF-8 encoded. The
+    result of the check is stored in the @a state parameter. The function must
+    be called initially with state 0 (accept). State 1 means the string must
+    be rejected, because the current byte is not allowed. If the string is
+    completely processed, but the state is non-zero, the string ended
+    prematurely; that is, the last byte indicated more bytes should have
+    followed.
+
+    @param[in,out] state  the state of the decoding
+    @param[in,out] codep  codepoint (valid only if resulting state is UTF8_ACCEPT)
+    @param[in] byte       next byte to decode
+    @return               new state
+
+    @note The function has been edited: a std::array is used.
+
+    @copyright Copyright (c) 2008-2009 Bjoern Hoehrmann <bjoern@hoehrmann.de>
+    @sa http://bjoern.hoehrmann.de/utf-8/decoder/dfa/
+    */
+    static std::uint8_t decode(std::uint8_t& state, std::uint32_t& codep, const std::uint8_t byte) noexcept
+    {
+        static const std::array<std::uint8_t, 400> utf8d =
+        {
+            {
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 00..1F
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 20..3F
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 40..5F
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 60..7F
+                1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, // 80..9F
+                7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, // A0..BF
+                8, 8, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // C0..DF
+                0xA, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x4, 0x3, 0x3, // E0..EF
+                0xB, 0x6, 0x6, 0x6, 0x5, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, // F0..FF
+                0x0, 0x1, 0x2, 0x3, 0x5, 0x8, 0x7, 0x1, 0x1, 0x1, 0x4, 0x6, 0x1, 0x1, 0x1, 0x1, // s0..s0
+                1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 0, 1, 1, 1, 1, 1, 1, // s1..s2
+                1, 2, 1, 1, 1, 1, 1, 2, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, // s3..s4
+                1, 2, 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 3, 1, 3, 1, 1, 1, 1, 1, 1, // s5..s6
+                1, 3, 1, 1, 1, 1, 1, 3, 1, 3, 1, 1, 1, 1, 1, 1, 1, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 // s7..s8
+            }
+        };
+
+        JSON_ASSERT(byte < utf8d.size());
+        const std::uint8_t type = utf8d[byte];
+
+        codep = (state != UTF8_ACCEPT)
+                ? (byte & 0x3fu) | (codep << 6u)
+                : (0xFFu >> type) & (byte);
+
+        const std::size_t index = 256u + static_cast<size_t>(state) * 16u + static_cast<size_t>(type);
+        JSON_ASSERT(index < utf8d.size());
+        state = utf8d[index];
+        return state;
+    }
+
+    /*
+     * Overload to make the compiler happy while it is instantiating
+     * dump_integer for number_unsigned_t.
+     * Must never be called.
+     */
+    number_unsigned_t remove_sign(number_unsigned_t x)
+    {
+        JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        return x; // LCOV_EXCL_LINE
+    }
+
+    /*
+     * Helper function for dump_integer
+     *
+     * This function takes a negative signed integer and returns its absolute
+     * value as unsigned integer. The plus/minus shuffling is necessary as we can
+     * not directly remove the sign of an arbitrary signed integer as the
+     * absolute values of INT_MIN and INT_MAX are usually not the same. See
+     * #1708 for details.
+     */
+    inline number_unsigned_t remove_sign(number_integer_t x) noexcept
+    {
+        JSON_ASSERT(x < 0 && x < (std::numeric_limits<number_integer_t>::max)()); // NOLINT(misc-redundant-expression)
+        return static_cast<number_unsigned_t>(-(x + 1)) + 1;
+    }
+
+  private:
+    /// the output of the serializer
+    output_adapter_t<char> o = nullptr;
+
+    /// a (hopefully) large enough character buffer
+    std::array<char, 64> number_buffer{{}};
+
+    /// the locale
+    const std::lconv* loc = nullptr;
+    /// the locale's thousand separator character
+    const char thousands_sep = '\0';
+    /// the locale's decimal point character
+    const char decimal_point = '\0';
+
+    /// string buffer
+    std::array<char, 512> string_buffer{{}};
+
+    /// the indentation character
+    const char indent_char;
+    /// the indentation string
+    string_t indent_string;
+
+    /// error_handler how to react on decoding errors
+    const error_handler_t error_handler;
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/value_t.hpp>
+
+// #include <nlohmann/json_fwd.hpp>
+
+// #include <nlohmann/ordered_map.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <functional> // equal_to, less
+#include <initializer_list> // initializer_list
+#include <iterator> // input_iterator_tag, iterator_traits
+#include <memory> // allocator
+#include <stdexcept> // for out_of_range
+#include <type_traits> // enable_if, is_convertible
+#include <utility> // pair
+#include <vector> // vector
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+/// ordered_map: a minimal map-like container that preserves insertion order
+/// for use within nlohmann::basic_json<ordered_map>
+template <class Key, class T, class IgnoredLess = std::less<Key>,
+          class Allocator = std::allocator<std::pair<const Key, T>>>
+                  struct ordered_map : std::vector<std::pair<const Key, T>, Allocator>
+{
+    using key_type = Key;
+    using mapped_type = T;
+    using Container = std::vector<std::pair<const Key, T>, Allocator>;
+    using iterator = typename Container::iterator;
+    using const_iterator = typename Container::const_iterator;
+    using size_type = typename Container::size_type;
+    using value_type = typename Container::value_type;
+#ifdef JSON_HAS_CPP_14
+    using key_compare = std::equal_to<>;
+#else
+    using key_compare = std::equal_to<Key>;
+#endif
+
+    // Explicit constructors instead of `using Container::Container`
+    // otherwise older compilers choke on it (GCC <= 5.5, xcode <= 9.4)
+    ordered_map() noexcept(noexcept(Container())) : Container{} {}
+    explicit ordered_map(const Allocator& alloc) noexcept(noexcept(Container(alloc))) : Container{alloc} {}
+    template <class It>
+    ordered_map(It first, It last, const Allocator& alloc = Allocator())
+        : Container{first, last, alloc} {}
+    ordered_map(std::initializer_list<value_type> init, const Allocator& alloc = Allocator() )
+        : Container{init, alloc} {}
+
+    std::pair<iterator, bool> emplace(const key_type& key, T&& t)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return {it, false};
+            }
+        }
+        Container::emplace_back(key, std::forward<T>(t));
+        return {std::prev(this->end()), true};
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    std::pair<iterator, bool> emplace(KeyType && key, T && t)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return {it, false};
+            }
+        }
+        Container::emplace_back(std::forward<KeyType>(key), std::forward<T>(t));
+        return {std::prev(this->end()), true};
+    }
+
+    T& operator[](const key_type& key)
+    {
+        return emplace(key, T{}).first->second;
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    T & operator[](KeyType && key)
+    {
+        return emplace(std::forward<KeyType>(key), T{}).first->second;
+    }
+
+    const T& operator[](const key_type& key) const
+    {
+        return at(key);
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    const T & operator[](KeyType && key) const
+    {
+        return at(std::forward<KeyType>(key));
+    }
+
+    T& at(const key_type& key)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return it->second;
+            }
+        }
+
+        JSON_THROW(std::out_of_range("key not found"));
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    T & at(KeyType && key) // NOLINT(cppcoreguidelines-missing-std-forward)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return it->second;
+            }
+        }
+
+        JSON_THROW(std::out_of_range("key not found"));
+    }
+
+    const T& at(const key_type& key) const
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return it->second;
+            }
+        }
+
+        JSON_THROW(std::out_of_range("key not found"));
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    const T & at(KeyType && key) const // NOLINT(cppcoreguidelines-missing-std-forward)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return it->second;
+            }
+        }
+
+        JSON_THROW(std::out_of_range("key not found"));
+    }
+
+    size_type erase(const key_type& key)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                // Since we cannot move const Keys, re-construct them in place
+                for (auto next = it; ++next != this->end(); ++it)
+                {
+                    it->~value_type(); // Destroy but keep allocation
+                    new (&*it) value_type{std::move(*next)};
+                }
+                Container::pop_back();
+                return 1;
+            }
+        }
+        return 0;
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    size_type erase(KeyType && key) // NOLINT(cppcoreguidelines-missing-std-forward)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                // Since we cannot move const Keys, re-construct them in place
+                for (auto next = it; ++next != this->end(); ++it)
+                {
+                    it->~value_type(); // Destroy but keep allocation
+                    new (&*it) value_type{std::move(*next)};
+                }
+                Container::pop_back();
+                return 1;
+            }
+        }
+        return 0;
+    }
+
+    iterator erase(iterator pos)
+    {
+        return erase(pos, std::next(pos));
+    }
+
+    iterator erase(iterator first, iterator last)
+    {
+        if (first == last)
+        {
+            return first;
+        }
+
+        const auto elements_affected = std::distance(first, last);
+        const auto offset = std::distance(Container::begin(), first);
+
+        // This is the start situation. We need to delete elements_affected
+        // elements (3 in this example: e, f, g), and need to return an
+        // iterator past the last deleted element (h in this example).
+        // Note that offset is the distance from the start of the vector
+        // to first. We will need this later.
+
+        // [ a, b, c, d, e, f, g, h, i, j ]
+        //               ^        ^
+        //             first    last
+
+        // Since we cannot move const Keys, we re-construct them in place.
+        // We start at first and re-construct (viz. copy) the elements from
+        // the back of the vector. Example for first iteration:
+
+        //               ,--------.
+        //               v        |   destroy e and re-construct with h
+        // [ a, b, c, d, e, f, g, h, i, j ]
+        //               ^        ^
+        //               it       it + elements_affected
+
+        for (auto it = first; std::next(it, elements_affected) != Container::end(); ++it)
+        {
+            it->~value_type(); // destroy but keep allocation
+            new (&*it) value_type{std::move(*std::next(it, elements_affected))}; // "move" next element to it
+        }
+
+        // [ a, b, c, d, h, i, j, h, i, j ]
+        //               ^        ^
+        //             first    last
+
+        // remove the unneeded elements at the end of the vector
+        Container::resize(this->size() - static_cast<size_type>(elements_affected));
+
+        // [ a, b, c, d, h, i, j ]
+        //               ^        ^
+        //             first    last
+
+        // first is now pointing past the last deleted element, but we cannot
+        // use this iterator, because it may have been invalidated by the
+        // resize call. Instead, we can return begin() + offset.
+        return Container::begin() + offset;
+    }
+
+    size_type count(const key_type& key) const
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return 1;
+            }
+        }
+        return 0;
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    size_type count(KeyType && key) const // NOLINT(cppcoreguidelines-missing-std-forward)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return 1;
+            }
+        }
+        return 0;
+    }
+
+    iterator find(const key_type& key)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return it;
+            }
+        }
+        return Container::end();
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    iterator find(KeyType && key) // NOLINT(cppcoreguidelines-missing-std-forward)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return it;
+            }
+        }
+        return Container::end();
+    }
+
+    const_iterator find(const key_type& key) const
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return it;
+            }
+        }
+        return Container::end();
+    }
+
+    std::pair<iterator, bool> insert( value_type&& value )
+    {
+        return emplace(value.first, std::move(value.second));
+    }
+
+    std::pair<iterator, bool> insert( const value_type& value )
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, value.first))
+            {
+                return {it, false};
+            }
+        }
+        Container::push_back(value);
+        return {--this->end(), true};
+    }
+
+    template<typename InputIt>
+    using require_input_iter = typename std::enable_if<std::is_convertible<typename std::iterator_traits<InputIt>::iterator_category,
+            std::input_iterator_tag>::value>::type;
+
+    template<typename InputIt, typename = require_input_iter<InputIt>>
+    void insert(InputIt first, InputIt last)
+    {
+        for (auto it = first; it != last; ++it)
+        {
+            insert(*it);
+        }
+    }
+
+private:
+    JSON_NO_UNIQUE_ADDRESS key_compare m_compare = key_compare();
+};
+
+NLOHMANN_JSON_NAMESPACE_END
+
+
+#if defined(JSON_HAS_CPP_17)
+    #if JSON_HAS_STATIC_RTTI
+        #include <any>
+    #endif
+    #include <string_view>
+#endif
+
+/*!
+@brief namespace for Niels Lohmann
+@see https://github.com/nlohmann
+@since version 1.0.0
+*/
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+/*!
+@brief a class to store JSON values
+
+@internal
+@invariant The member variables @a m_value and @a m_type have the following
+relationship:
+- If `m_type == value_t::object`, then `m_value.object != nullptr`.
+- If `m_type == value_t::array`, then `m_value.array != nullptr`.
+- If `m_type == value_t::string`, then `m_value.string != nullptr`.
+The invariants are checked by member function assert_invariant().
+
+@note ObjectType trick from https://stackoverflow.com/a/9860911
+@endinternal
+
+@since version 1.0.0
+
+@nosubgrouping
+*/
+NLOHMANN_BASIC_JSON_TPL_DECLARATION
+class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-special-member-functions)
+    : public ::nlohmann::detail::json_base_class<CustomBaseClass>
+{
+  private:
+    template<detail::value_t> friend struct detail::external_constructor;
+
+    template<typename>
+    friend class ::nlohmann::json_pointer;
+    // can be restored when json_pointer backwards compatibility is removed
+    // friend ::nlohmann::json_pointer<StringType>;
+
+    template<typename BasicJsonType, typename InputType>
+    friend class ::nlohmann::detail::parser;
+    friend ::nlohmann::detail::serializer<basic_json>;
+    template<typename BasicJsonType>
+    friend class ::nlohmann::detail::iter_impl;
+    template<typename BasicJsonType, typename CharType>
+    friend class ::nlohmann::detail::binary_writer;
+    template<typename BasicJsonType, typename InputType, typename SAX>
+    friend class ::nlohmann::detail::binary_reader;
+    template<typename BasicJsonType>
+    friend class ::nlohmann::detail::json_sax_dom_parser;
+    template<typename BasicJsonType>
+    friend class ::nlohmann::detail::json_sax_dom_callback_parser;
+    friend class ::nlohmann::detail::exception;
+
+    /// workaround type for MSVC
+    using basic_json_t = NLOHMANN_BASIC_JSON_TPL;
+    using json_base_class_t = ::nlohmann::detail::json_base_class<CustomBaseClass>;
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    // convenience aliases for types residing in namespace detail;
+    using lexer = ::nlohmann::detail::lexer_base<basic_json>;
+
+    template<typename InputAdapterType>
+    static ::nlohmann::detail::parser<basic_json, InputAdapterType> parser(
+        InputAdapterType adapter,
+        detail::parser_callback_t<basic_json>cb = nullptr,
+        const bool allow_exceptions = true,
+        const bool ignore_comments = false
+                                 )
+    {
+        return ::nlohmann::detail::parser<basic_json, InputAdapterType>(std::move(adapter),
+                std::move(cb), allow_exceptions, ignore_comments);
+    }
+
+  private:
+    using primitive_iterator_t = ::nlohmann::detail::primitive_iterator_t;
+    template<typename BasicJsonType>
+    using internal_iterator = ::nlohmann::detail::internal_iterator<BasicJsonType>;
+    template<typename BasicJsonType>
+    using iter_impl = ::nlohmann::detail::iter_impl<BasicJsonType>;
+    template<typename Iterator>
+    using iteration_proxy = ::nlohmann::detail::iteration_proxy<Iterator>;
+    template<typename Base> using json_reverse_iterator = ::nlohmann::detail::json_reverse_iterator<Base>;
+
+    template<typename CharType>
+    using output_adapter_t = ::nlohmann::detail::output_adapter_t<CharType>;
+
+    template<typename InputType>
+    using binary_reader = ::nlohmann::detail::binary_reader<basic_json, InputType>;
+    template<typename CharType> using binary_writer = ::nlohmann::detail::binary_writer<basic_json, CharType>;
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    using serializer = ::nlohmann::detail::serializer<basic_json>;
+
+  public:
+    using value_t = detail::value_t;
+    /// JSON Pointer, see @ref nlohmann::json_pointer
+    using json_pointer = ::nlohmann::json_pointer<StringType>;
+    template<typename T, typename SFINAE>
+    using json_serializer = JSONSerializer<T, SFINAE>;
+    /// how to treat decoding errors
+    using error_handler_t = detail::error_handler_t;
+    /// how to treat CBOR tags
+    using cbor_tag_handler_t = detail::cbor_tag_handler_t;
+    /// helper type for initializer lists of basic_json values
+    using initializer_list_t = std::initializer_list<detail::json_ref<basic_json>>;
+
+    using input_format_t = detail::input_format_t;
+    /// SAX interface type, see @ref nlohmann::json_sax
+    using json_sax_t = json_sax<basic_json>;
+
+    ////////////////
+    // exceptions //
+    ////////////////
+
+    /// @name exceptions
+    /// Classes to implement user-defined exceptions.
+    /// @{
+
+    using exception = detail::exception;
+    using parse_error = detail::parse_error;
+    using invalid_iterator = detail::invalid_iterator;
+    using type_error = detail::type_error;
+    using out_of_range = detail::out_of_range;
+    using other_error = detail::other_error;
+
+    /// @}
+
+    /////////////////////
+    // container types //
+    /////////////////////
+
+    /// @name container types
+    /// The canonic container types to use @ref basic_json like any other STL
+    /// container.
+    /// @{
+
+    /// the type of elements in a basic_json container
+    using value_type = basic_json;
+
+    /// the type of an element reference
+    using reference = value_type&;
+    /// the type of an element const reference
+    using const_reference = const value_type&;
+
+    /// a type to represent differences between iterators
+    using difference_type = std::ptrdiff_t;
+    /// a type to represent container sizes
+    using size_type = std::size_t;
+
+    /// the allocator type
+    using allocator_type = AllocatorType<basic_json>;
+
+    /// the type of an element pointer
+    using pointer = typename std::allocator_traits<allocator_type>::pointer;
+    /// the type of an element const pointer
+    using const_pointer = typename std::allocator_traits<allocator_type>::const_pointer;
+
+    /// an iterator for a basic_json container
+    using iterator = iter_impl<basic_json>;
+    /// a const iterator for a basic_json container
+    using const_iterator = iter_impl<const basic_json>;
+    /// a reverse iterator for a basic_json container
+    using reverse_iterator = json_reverse_iterator<typename basic_json::iterator>;
+    /// a const reverse iterator for a basic_json container
+    using const_reverse_iterator = json_reverse_iterator<typename basic_json::const_iterator>;
+
+    /// @}
+
+    /// @brief returns the allocator associated with the container
+    /// @sa https://json.nlohmann.me/api/basic_json/get_allocator/
+    static allocator_type get_allocator()
+    {
+        return allocator_type();
+    }
+
+    /// @brief returns version information on the library
+    /// @sa https://json.nlohmann.me/api/basic_json/meta/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json meta()
+    {
+        basic_json result;
+
+        result["copyright"] = "(C) 2013-2023 Niels Lohmann";
+        result["name"] = "JSON for Modern C++";
+        result["url"] = "https://github.com/nlohmann/json";
+        result["version"]["string"] =
+            detail::concat(std::to_string(NLOHMANN_JSON_VERSION_MAJOR), '.',
+                           std::to_string(NLOHMANN_JSON_VERSION_MINOR), '.',
+                           std::to_string(NLOHMANN_JSON_VERSION_PATCH));
+        result["version"]["major"] = NLOHMANN_JSON_VERSION_MAJOR;
+        result["version"]["minor"] = NLOHMANN_JSON_VERSION_MINOR;
+        result["version"]["patch"] = NLOHMANN_JSON_VERSION_PATCH;
+
+#ifdef _WIN32
+        result["platform"] = "win32";
+#elif defined __linux__
+        result["platform"] = "linux";
+#elif defined __APPLE__
+        result["platform"] = "apple";
+#elif defined __unix__
+        result["platform"] = "unix";
+#else
+        result["platform"] = "unknown";
+#endif
+
+#if defined(__ICC) || defined(__INTEL_COMPILER)
+        result["compiler"] = {{"family", "icc"}, {"version", __INTEL_COMPILER}};
+#elif defined(__clang__)
+        result["compiler"] = {{"family", "clang"}, {"version", __clang_version__}};
+#elif defined(__GNUC__) || defined(__GNUG__)
+        result["compiler"] = {{"family", "gcc"}, {"version", detail::concat(
+                    std::to_string(__GNUC__), '.',
+                    std::to_string(__GNUC_MINOR__), '.',
+                    std::to_string(__GNUC_PATCHLEVEL__))
+            }
+        };
+#elif defined(__HP_cc) || defined(__HP_aCC)
+        result["compiler"] = "hp"
+#elif defined(__IBMCPP__)
+        result["compiler"] = {{"family", "ilecpp"}, {"version", __IBMCPP__}};
+#elif defined(_MSC_VER)
+        result["compiler"] = {{"family", "msvc"}, {"version", _MSC_VER}};
+#elif defined(__PGI)
+        result["compiler"] = {{"family", "pgcpp"}, {"version", __PGI}};
+#elif defined(__SUNPRO_CC)
+        result["compiler"] = {{"family", "sunpro"}, {"version", __SUNPRO_CC}};
+#else
+        result["compiler"] = {{"family", "unknown"}, {"version", "unknown"}};
+#endif
+
+#if defined(_MSVC_LANG)
+        result["compiler"]["c++"] = std::to_string(_MSVC_LANG);
+#elif defined(__cplusplus)
+        result["compiler"]["c++"] = std::to_string(__cplusplus);
+#else
+        result["compiler"]["c++"] = "unknown";
+#endif
+        return result;
+    }
+
+    ///////////////////////////
+    // JSON value data types //
+    ///////////////////////////
+
+    /// @name JSON value data types
+    /// The data types to store a JSON value. These types are derived from
+    /// the template arguments passed to class @ref basic_json.
+    /// @{
+
+    /// @brief default object key comparator type
+    /// The actual object key comparator type (@ref object_comparator_t) may be
+    /// different.
+    /// @sa https://json.nlohmann.me/api/basic_json/default_object_comparator_t/
+#if defined(JSON_HAS_CPP_14)
+    // use of transparent comparator avoids unnecessary repeated construction of temporaries
+    // in functions involving lookup by key with types other than object_t::key_type (aka. StringType)
+    using default_object_comparator_t = std::less<>;
+#else
+    using default_object_comparator_t = std::less<StringType>;
+#endif
+
+    /// @brief a type for an object
+    /// @sa https://json.nlohmann.me/api/basic_json/object_t/
+    using object_t = ObjectType<StringType,
+          basic_json,
+          default_object_comparator_t,
+          AllocatorType<std::pair<const StringType,
+          basic_json>>>;
+
+    /// @brief a type for an array
+    /// @sa https://json.nlohmann.me/api/basic_json/array_t/
+    using array_t = ArrayType<basic_json, AllocatorType<basic_json>>;
+
+    /// @brief a type for a string
+    /// @sa https://json.nlohmann.me/api/basic_json/string_t/
+    using string_t = StringType;
+
+    /// @brief a type for a boolean
+    /// @sa https://json.nlohmann.me/api/basic_json/boolean_t/
+    using boolean_t = BooleanType;
+
+    /// @brief a type for a number (integer)
+    /// @sa https://json.nlohmann.me/api/basic_json/number_integer_t/
+    using number_integer_t = NumberIntegerType;
+
+    /// @brief a type for a number (unsigned)
+    /// @sa https://json.nlohmann.me/api/basic_json/number_unsigned_t/
+    using number_unsigned_t = NumberUnsignedType;
+
+    /// @brief a type for a number (floating-point)
+    /// @sa https://json.nlohmann.me/api/basic_json/number_float_t/
+    using number_float_t = NumberFloatType;
+
+    /// @brief a type for a packed binary type
+    /// @sa https://json.nlohmann.me/api/basic_json/binary_t/
+    using binary_t = nlohmann::byte_container_with_subtype<BinaryType>;
+
+    /// @brief object key comparator type
+    /// @sa https://json.nlohmann.me/api/basic_json/object_comparator_t/
+    using object_comparator_t = detail::actual_object_comparator_t<basic_json>;
+
+    /// @}
+
+  private:
+
+    /// helper for exception-safe object creation
+    template<typename T, typename... Args>
+    JSON_HEDLEY_RETURNS_NON_NULL
+    static T* create(Args&& ... args)
+    {
+        AllocatorType<T> alloc;
+        using AllocatorTraits = std::allocator_traits<AllocatorType<T>>;
+
+        auto deleter = [&](T * obj)
+        {
+            AllocatorTraits::deallocate(alloc, obj, 1);
+        };
+        std::unique_ptr<T, decltype(deleter)> obj(AllocatorTraits::allocate(alloc, 1), deleter);
+        AllocatorTraits::construct(alloc, obj.get(), std::forward<Args>(args)...);
+        JSON_ASSERT(obj != nullptr);
+        return obj.release();
+    }
+
+    ////////////////////////
+    // JSON value storage //
+    ////////////////////////
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    /*!
+    @brief a JSON value
+
+    The actual storage for a JSON value of the @ref basic_json class. This
+    union combines the different storage types for the JSON value types
+    defined in @ref value_t.
+
+    JSON type | value_t type    | used type
+    --------- | --------------- | ------------------------
+    object    | object          | pointer to @ref object_t
+    array     | array           | pointer to @ref array_t
+    string    | string          | pointer to @ref string_t
+    boolean   | boolean         | @ref boolean_t
+    number    | number_integer  | @ref number_integer_t
+    number    | number_unsigned | @ref number_unsigned_t
+    number    | number_float    | @ref number_float_t
+    binary    | binary          | pointer to @ref binary_t
+    null      | null            | *no value is stored*
+
+    @note Variable-length types (objects, arrays, and strings) are stored as
+    pointers. The size of the union should not exceed 64 bits if the default
+    value types are used.
+
+    @since version 1.0.0
+    */
+    union json_value
+    {
+        /// object (stored with pointer to save storage)
+        object_t* object;
+        /// array (stored with pointer to save storage)
+        array_t* array;
+        /// string (stored with pointer to save storage)
+        string_t* string;
+        /// binary (stored with pointer to save storage)
+        binary_t* binary;
+        /// boolean
+        boolean_t boolean;
+        /// number (integer)
+        number_integer_t number_integer;
+        /// number (unsigned integer)
+        number_unsigned_t number_unsigned;
+        /// number (floating-point)
+        number_float_t number_float;
+
+        /// default constructor (for null values)
+        json_value() = default;
+        /// constructor for booleans
+        json_value(boolean_t v) noexcept : boolean(v) {}
+        /// constructor for numbers (integer)
+        json_value(number_integer_t v) noexcept : number_integer(v) {}
+        /// constructor for numbers (unsigned)
+        json_value(number_unsigned_t v) noexcept : number_unsigned(v) {}
+        /// constructor for numbers (floating-point)
+        json_value(number_float_t v) noexcept : number_float(v) {}
+        /// constructor for empty values of a given type
+        json_value(value_t t)
+        {
+            switch (t)
+            {
+                case value_t::object:
+                {
+                    object = create<object_t>();
+                    break;
+                }
+
+                case value_t::array:
+                {
+                    array = create<array_t>();
+                    break;
+                }
+
+                case value_t::string:
+                {
+                    string = create<string_t>("");
+                    break;
+                }
+
+                case value_t::binary:
+                {
+                    binary = create<binary_t>();
+                    break;
+                }
+
+                case value_t::boolean:
+                {
+                    boolean = static_cast<boolean_t>(false);
+                    break;
+                }
+
+                case value_t::number_integer:
+                {
+                    number_integer = static_cast<number_integer_t>(0);
+                    break;
+                }
+
+                case value_t::number_unsigned:
+                {
+                    number_unsigned = static_cast<number_unsigned_t>(0);
+                    break;
+                }
+
+                case value_t::number_float:
+                {
+                    number_float = static_cast<number_float_t>(0.0);
+                    break;
+                }
+
+                case value_t::null:
+                {
+                    object = nullptr;  // silence warning, see #821
+                    break;
+                }
+
+                case value_t::discarded:
+                default:
+                {
+                    object = nullptr;  // silence warning, see #821
+                    if (JSON_HEDLEY_UNLIKELY(t == value_t::null))
+                    {
+                        JSON_THROW(other_error::create(500, "961c151d2e87f2686a955a9be24d316f1362bf21 3.11.3", nullptr)); // LCOV_EXCL_LINE
+                    }
+                    break;
+                }
+            }
+        }
+
+        /// constructor for strings
+        json_value(const string_t& value) : string(create<string_t>(value)) {}
+
+        /// constructor for rvalue strings
+        json_value(string_t&& value) : string(create<string_t>(std::move(value))) {}
+
+        /// constructor for objects
+        json_value(const object_t& value) : object(create<object_t>(value)) {}
+
+        /// constructor for rvalue objects
+        json_value(object_t&& value) : object(create<object_t>(std::move(value))) {}
+
+        /// constructor for arrays
+        json_value(const array_t& value) : array(create<array_t>(value)) {}
+
+        /// constructor for rvalue arrays
+        json_value(array_t&& value) : array(create<array_t>(std::move(value))) {}
+
+        /// constructor for binary arrays
+        json_value(const typename binary_t::container_type& value) : binary(create<binary_t>(value)) {}
+
+        /// constructor for rvalue binary arrays
+        json_value(typename binary_t::container_type&& value) : binary(create<binary_t>(std::move(value))) {}
+
+        /// constructor for binary arrays (internal type)
+        json_value(const binary_t& value) : binary(create<binary_t>(value)) {}
+
+        /// constructor for rvalue binary arrays (internal type)
+        json_value(binary_t&& value) : binary(create<binary_t>(std::move(value))) {}
+
+        void destroy(value_t t)
+        {
+            if (
+                (t == value_t::object && object == nullptr) ||
+                (t == value_t::array && array == nullptr) ||
+                (t == value_t::string && string == nullptr) ||
+                (t == value_t::binary && binary == nullptr)
+            )
+            {
+                //not initialized (e.g. due to exception in the ctor)
+                return;
+            }
+            if (t == value_t::array || t == value_t::object)
+            {
+                // flatten the current json_value to a heap-allocated stack
+                std::vector<basic_json> stack;
+
+                // move the top-level items to stack
+                if (t == value_t::array)
+                {
+                    stack.reserve(array->size());
+                    std::move(array->begin(), array->end(), std::back_inserter(stack));
+                }
+                else
+                {
+                    stack.reserve(object->size());
+                    for (auto&& it : *object)
+                    {
+                        stack.push_back(std::move(it.second));
+                    }
+                }
+
+                while (!stack.empty())
+                {
+                    // move the last item to local variable to be processed
+                    basic_json current_item(std::move(stack.back()));
+                    stack.pop_back();
+
+                    // if current_item is array/object, move
+                    // its children to the stack to be processed later
+                    if (current_item.is_array())
+                    {
+                        std::move(current_item.m_data.m_value.array->begin(), current_item.m_data.m_value.array->end(), std::back_inserter(stack));
+
+                        current_item.m_data.m_value.array->clear();
+                    }
+                    else if (current_item.is_object())
+                    {
+                        for (auto&& it : *current_item.m_data.m_value.object)
+                        {
+                            stack.push_back(std::move(it.second));
+                        }
+
+                        current_item.m_data.m_value.object->clear();
+                    }
+
+                    // it's now safe that current_item get destructed
+                    // since it doesn't have any children
+                }
+            }
+
+            switch (t)
+            {
+                case value_t::object:
+                {
+                    AllocatorType<object_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, object);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, object, 1);
+                    break;
+                }
+
+                case value_t::array:
+                {
+                    AllocatorType<array_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, array);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, array, 1);
+                    break;
+                }
+
+                case value_t::string:
+                {
+                    AllocatorType<string_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, string);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, string, 1);
+                    break;
+                }
+
+                case value_t::binary:
+                {
+                    AllocatorType<binary_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, binary);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, binary, 1);
+                    break;
+                }
+
+                case value_t::null:
+                case value_t::boolean:
+                case value_t::number_integer:
+                case value_t::number_unsigned:
+                case value_t::number_float:
+                case value_t::discarded:
+                default:
+                {
+                    break;
+                }
+            }
+        }
+    };
+
+  private:
+    /*!
+    @brief checks the class invariants
+
+    This function asserts the class invariants. It needs to be called at the
+    end of every constructor to make sure that created objects respect the
+    invariant. Furthermore, it has to be called each time the type of a JSON
+    value is changed, because the invariant expresses a relationship between
+    @a m_type and @a m_value.
+
+    Furthermore, the parent relation is checked for arrays and objects: If
+    @a check_parents true and the value is an array or object, then the
+    container's elements must have the current value as parent.
+
+    @param[in] check_parents  whether the parent relation should be checked.
+               The value is true by default and should only be set to false
+               during destruction of objects when the invariant does not
+               need to hold.
+    */
+    void assert_invariant(bool check_parents = true) const noexcept
+    {
+        JSON_ASSERT(m_data.m_type != value_t::object || m_data.m_value.object != nullptr);
+        JSON_ASSERT(m_data.m_type != value_t::array || m_data.m_value.array != nullptr);
+        JSON_ASSERT(m_data.m_type != value_t::string || m_data.m_value.string != nullptr);
+        JSON_ASSERT(m_data.m_type != value_t::binary || m_data.m_value.binary != nullptr);
+
+#if JSON_DIAGNOSTICS
+        JSON_TRY
+        {
+            // cppcheck-suppress assertWithSideEffect
+            JSON_ASSERT(!check_parents || !is_structured() || std::all_of(begin(), end(), [this](const basic_json & j)
+            {
+                return j.m_parent == this;
+            }));
+        }
+        JSON_CATCH(...) {} // LCOV_EXCL_LINE
+#endif
+        static_cast<void>(check_parents);
+    }
+
+    void set_parents()
+    {
+#if JSON_DIAGNOSTICS
+        switch (m_data.m_type)
+        {
+            case value_t::array:
+            {
+                for (auto& element : *m_data.m_value.array)
+                {
+                    element.m_parent = this;
+                }
+                break;
+            }
+
+            case value_t::object:
+            {
+                for (auto& element : *m_data.m_value.object)
+                {
+                    element.second.m_parent = this;
+                }
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                break;
+        }
+#endif
+    }
+
+    iterator set_parents(iterator it, typename iterator::difference_type count_set_parents)
+    {
+#if JSON_DIAGNOSTICS
+        for (typename iterator::difference_type i = 0; i < count_set_parents; ++i)
+        {
+            (it + i)->m_parent = this;
+        }
+#else
+        static_cast<void>(count_set_parents);
+#endif
+        return it;
+    }
+
+    reference set_parent(reference j, std::size_t old_capacity = static_cast<std::size_t>(-1))
+    {
+#if JSON_DIAGNOSTICS
+        if (old_capacity != static_cast<std::size_t>(-1))
+        {
+            // see https://github.com/nlohmann/json/issues/2838
+            JSON_ASSERT(type() == value_t::array);
+            if (JSON_HEDLEY_UNLIKELY(m_data.m_value.array->capacity() != old_capacity))
+            {
+                // capacity has changed: update all parents
+                set_parents();
+                return j;
+            }
+        }
+
+        // ordered_json uses a vector internally, so pointers could have
+        // been invalidated; see https://github.com/nlohmann/json/issues/2962
+#ifdef JSON_HEDLEY_MSVC_VERSION
+#pragma warning(push )
+#pragma warning(disable : 4127) // ignore warning to replace if with if constexpr
+#endif
+        if (detail::is_ordered_map<object_t>::value)
+        {
+            set_parents();
+            return j;
+        }
+#ifdef JSON_HEDLEY_MSVC_VERSION
+#pragma warning( pop )
+#endif
+
+        j.m_parent = this;
+#else
+        static_cast<void>(j);
+        static_cast<void>(old_capacity);
+#endif
+        return j;
+    }
+
+  public:
+    //////////////////////////
+    // JSON parser callback //
+    //////////////////////////
+
+    /// @brief parser event types
+    /// @sa https://json.nlohmann.me/api/basic_json/parse_event_t/
+    using parse_event_t = detail::parse_event_t;
+
+    /// @brief per-element parser callback type
+    /// @sa https://json.nlohmann.me/api/basic_json/parser_callback_t/
+    using parser_callback_t = detail::parser_callback_t<basic_json>;
+
+    //////////////////
+    // constructors //
+    //////////////////
+
+    /// @name constructors and destructors
+    /// Constructors of class @ref basic_json, copy/move constructor, copy
+    /// assignment, static functions creating objects, and the destructor.
+    /// @{
+
+    /// @brief create an empty value with a given type
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(const value_t v)
+        : m_data(v)
+    {
+        assert_invariant();
+    }
+
+    /// @brief create a null object
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(std::nullptr_t = nullptr) noexcept // NOLINT(bugprone-exception-escape)
+        : basic_json(value_t::null)
+    {
+        assert_invariant();
+    }
+
+    /// @brief create a JSON value from compatible types
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    template < typename CompatibleType,
+               typename U = detail::uncvref_t<CompatibleType>,
+               detail::enable_if_t <
+                   !detail::is_basic_json<U>::value && detail::is_compatible_type<basic_json_t, U>::value, int > = 0 >
+    basic_json(CompatibleType && val) noexcept(noexcept( // NOLINT(bugprone-forwarding-reference-overload,bugprone-exception-escape)
+                JSONSerializer<U>::to_json(std::declval<basic_json_t&>(),
+                                           std::forward<CompatibleType>(val))))
+    {
+        JSONSerializer<U>::to_json(*this, std::forward<CompatibleType>(val));
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief create a JSON value from an existing one
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    template < typename BasicJsonType,
+               detail::enable_if_t <
+                   detail::is_basic_json<BasicJsonType>::value&& !std::is_same<basic_json, BasicJsonType>::value, int > = 0 >
+    basic_json(const BasicJsonType& val)
+    {
+        using other_boolean_t = typename BasicJsonType::boolean_t;
+        using other_number_float_t = typename BasicJsonType::number_float_t;
+        using other_number_integer_t = typename BasicJsonType::number_integer_t;
+        using other_number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+        using other_string_t = typename BasicJsonType::string_t;
+        using other_object_t = typename BasicJsonType::object_t;
+        using other_array_t = typename BasicJsonType::array_t;
+        using other_binary_t = typename BasicJsonType::binary_t;
+
+        switch (val.type())
+        {
+            case value_t::boolean:
+                JSONSerializer<other_boolean_t>::to_json(*this, val.template get<other_boolean_t>());
+                break;
+            case value_t::number_float:
+                JSONSerializer<other_number_float_t>::to_json(*this, val.template get<other_number_float_t>());
+                break;
+            case value_t::number_integer:
+                JSONSerializer<other_number_integer_t>::to_json(*this, val.template get<other_number_integer_t>());
+                break;
+            case value_t::number_unsigned:
+                JSONSerializer<other_number_unsigned_t>::to_json(*this, val.template get<other_number_unsigned_t>());
+                break;
+            case value_t::string:
+                JSONSerializer<other_string_t>::to_json(*this, val.template get_ref<const other_string_t&>());
+                break;
+            case value_t::object:
+                JSONSerializer<other_object_t>::to_json(*this, val.template get_ref<const other_object_t&>());
+                break;
+            case value_t::array:
+                JSONSerializer<other_array_t>::to_json(*this, val.template get_ref<const other_array_t&>());
+                break;
+            case value_t::binary:
+                JSONSerializer<other_binary_t>::to_json(*this, val.template get_ref<const other_binary_t&>());
+                break;
+            case value_t::null:
+                *this = nullptr;
+                break;
+            case value_t::discarded:
+                m_data.m_type = value_t::discarded;
+                break;
+            default:            // LCOV_EXCL_LINE
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        }
+        JSON_ASSERT(m_data.m_type == val.type());
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief create a container (array or object) from an initializer list
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(initializer_list_t init,
+               bool type_deduction = true,
+               value_t manual_type = value_t::array)
+    {
+        // check if each element is an array with two elements whose first
+        // element is a string
+        bool is_an_object = std::all_of(init.begin(), init.end(),
+                                        [](const detail::json_ref<basic_json>& element_ref)
+        {
+            // The cast is to ensure op[size_type] is called, bearing in mind size_type may not be int;
+            // (many string types can be constructed from 0 via its null-pointer guise, so we get a
+            // broken call to op[key_type], the wrong semantics and a 4804 warning on Windows)
+            return element_ref->is_array() && element_ref->size() == 2 && (*element_ref)[static_cast<size_type>(0)].is_string();
+        });
+
+        // adjust type if type deduction is not wanted
+        if (!type_deduction)
+        {
+            // if array is wanted, do not create an object though possible
+            if (manual_type == value_t::array)
+            {
+                is_an_object = false;
+            }
+
+            // if object is wanted but impossible, throw an exception
+            if (JSON_HEDLEY_UNLIKELY(manual_type == value_t::object && !is_an_object))
+            {
+                JSON_THROW(type_error::create(301, "cannot create object from initializer list", nullptr));
+            }
+        }
+
+        if (is_an_object)
+        {
+            // the initializer list is a list of pairs -> create object
+            m_data.m_type = value_t::object;
+            m_data.m_value = value_t::object;
+
+            for (auto& element_ref : init)
+            {
+                auto element = element_ref.moved_or_copied();
+                m_data.m_value.object->emplace(
+                    std::move(*((*element.m_data.m_value.array)[0].m_data.m_value.string)),
+                    std::move((*element.m_data.m_value.array)[1]));
+            }
+        }
+        else
+        {
+            // the initializer list describes an array -> create array
+            m_data.m_type = value_t::array;
+            m_data.m_value.array = create<array_t>(init.begin(), init.end());
+        }
+
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief explicitly create a binary array (without subtype)
+    /// @sa https://json.nlohmann.me/api/basic_json/binary/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json binary(const typename binary_t::container_type& init)
+    {
+        auto res = basic_json();
+        res.m_data.m_type = value_t::binary;
+        res.m_data.m_value = init;
+        return res;
+    }
+
+    /// @brief explicitly create a binary array (with subtype)
+    /// @sa https://json.nlohmann.me/api/basic_json/binary/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json binary(const typename binary_t::container_type& init, typename binary_t::subtype_type subtype)
+    {
+        auto res = basic_json();
+        res.m_data.m_type = value_t::binary;
+        res.m_data.m_value = binary_t(init, subtype);
+        return res;
+    }
+
+    /// @brief explicitly create a binary array
+    /// @sa https://json.nlohmann.me/api/basic_json/binary/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json binary(typename binary_t::container_type&& init)
+    {
+        auto res = basic_json();
+        res.m_data.m_type = value_t::binary;
+        res.m_data.m_value = std::move(init);
+        return res;
+    }
+
+    /// @brief explicitly create a binary array (with subtype)
+    /// @sa https://json.nlohmann.me/api/basic_json/binary/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json binary(typename binary_t::container_type&& init, typename binary_t::subtype_type subtype)
+    {
+        auto res = basic_json();
+        res.m_data.m_type = value_t::binary;
+        res.m_data.m_value = binary_t(std::move(init), subtype);
+        return res;
+    }
+
+    /// @brief explicitly create an array from an initializer list
+    /// @sa https://json.nlohmann.me/api/basic_json/array/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json array(initializer_list_t init = {})
+    {
+        return basic_json(init, false, value_t::array);
+    }
+
+    /// @brief explicitly create an object from an initializer list
+    /// @sa https://json.nlohmann.me/api/basic_json/object/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json object(initializer_list_t init = {})
+    {
+        return basic_json(init, false, value_t::object);
+    }
+
+    /// @brief construct an array with count copies of given value
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(size_type cnt, const basic_json& val):
+        m_data{cnt, val}
+    {
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief construct a JSON container given an iterator range
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    template < class InputIT, typename std::enable_if <
+                   std::is_same<InputIT, typename basic_json_t::iterator>::value ||
+                   std::is_same<InputIT, typename basic_json_t::const_iterator>::value, int >::type = 0 >
+    basic_json(InputIT first, InputIT last)
+    {
+        JSON_ASSERT(first.m_object != nullptr);
+        JSON_ASSERT(last.m_object != nullptr);
+
+        // make sure iterator fits the current value
+        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(201, "iterators are not compatible", nullptr));
+        }
+
+        // copy type from first iterator
+        m_data.m_type = first.m_object->m_data.m_type;
+
+        // check if iterator range is complete for primitive values
+        switch (m_data.m_type)
+        {
+            case value_t::boolean:
+            case value_t::number_float:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::string:
+            {
+                if (JSON_HEDLEY_UNLIKELY(!first.m_it.primitive_iterator.is_begin()
+                                         || !last.m_it.primitive_iterator.is_end()))
+                {
+                    JSON_THROW(invalid_iterator::create(204, "iterators out of range", first.m_object));
+                }
+                break;
+            }
+
+            case value_t::null:
+            case value_t::object:
+            case value_t::array:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                break;
+        }
+
+        switch (m_data.m_type)
+        {
+            case value_t::number_integer:
+            {
+                m_data.m_value.number_integer = first.m_object->m_data.m_value.number_integer;
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                m_data.m_value.number_unsigned = first.m_object->m_data.m_value.number_unsigned;
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                m_data.m_value.number_float = first.m_object->m_data.m_value.number_float;
+                break;
+            }
+
+            case value_t::boolean:
+            {
+                m_data.m_value.boolean = first.m_object->m_data.m_value.boolean;
+                break;
+            }
+
+            case value_t::string:
+            {
+                m_data.m_value = *first.m_object->m_data.m_value.string;
+                break;
+            }
+
+            case value_t::object:
+            {
+                m_data.m_value.object = create<object_t>(first.m_it.object_iterator,
+                                        last.m_it.object_iterator);
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_data.m_value.array = create<array_t>(first.m_it.array_iterator,
+                                                       last.m_it.array_iterator);
+                break;
+            }
+
+            case value_t::binary:
+            {
+                m_data.m_value = *first.m_object->m_data.m_value.binary;
+                break;
+            }
+
+            case value_t::null:
+            case value_t::discarded:
+            default:
+                JSON_THROW(invalid_iterator::create(206, detail::concat("cannot construct with iterators from ", first.m_object->type_name()), first.m_object));
+        }
+
+        set_parents();
+        assert_invariant();
+    }
+
+    ///////////////////////////////////////
+    // other constructors and destructor //
+    ///////////////////////////////////////
+
+    template<typename JsonRef,
+             detail::enable_if_t<detail::conjunction<detail::is_json_ref<JsonRef>,
+                                 std::is_same<typename JsonRef::value_type, basic_json>>::value, int> = 0 >
+    basic_json(const JsonRef& ref) : basic_json(ref.moved_or_copied()) {}
+
+    /// @brief copy constructor
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(const basic_json& other)
+        : json_base_class_t(other)
+    {
+        m_data.m_type = other.m_data.m_type;
+        // check of passed value is valid
+        other.assert_invariant();
+
+        switch (m_data.m_type)
+        {
+            case value_t::object:
+            {
+                m_data.m_value = *other.m_data.m_value.object;
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_data.m_value = *other.m_data.m_value.array;
+                break;
+            }
+
+            case value_t::string:
+            {
+                m_data.m_value = *other.m_data.m_value.string;
+                break;
+            }
+
+            case value_t::boolean:
+            {
+                m_data.m_value = other.m_data.m_value.boolean;
+                break;
+            }
+
+            case value_t::number_integer:
+            {
+                m_data.m_value = other.m_data.m_value.number_integer;
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                m_data.m_value = other.m_data.m_value.number_unsigned;
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                m_data.m_value = other.m_data.m_value.number_float;
+                break;
+            }
+
+            case value_t::binary:
+            {
+                m_data.m_value = *other.m_data.m_value.binary;
+                break;
+            }
+
+            case value_t::null:
+            case value_t::discarded:
+            default:
+                break;
+        }
+
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief move constructor
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(basic_json&& other) noexcept
+        : json_base_class_t(std::forward<json_base_class_t>(other)),
+          m_data(std::move(other.m_data))
+    {
+        // check that passed value is valid
+        other.assert_invariant(false);
+
+        // invalidate payload
+        other.m_data.m_type = value_t::null;
+        other.m_data.m_value = {};
+
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief copy assignment
+    /// @sa https://json.nlohmann.me/api/basic_json/operator=/
+    basic_json& operator=(basic_json other) noexcept (
+        std::is_nothrow_move_constructible<value_t>::value&&
+        std::is_nothrow_move_assignable<value_t>::value&&
+        std::is_nothrow_move_constructible<json_value>::value&&
+        std::is_nothrow_move_assignable<json_value>::value&&
+        std::is_nothrow_move_assignable<json_base_class_t>::value
+    )
+    {
+        // check that passed value is valid
+        other.assert_invariant();
+
+        using std::swap;
+        swap(m_data.m_type, other.m_data.m_type);
+        swap(m_data.m_value, other.m_data.m_value);
+        json_base_class_t::operator=(std::move(other));
+
+        set_parents();
+        assert_invariant();
+        return *this;
+    }
+
+    /// @brief destructor
+    /// @sa https://json.nlohmann.me/api/basic_json/~basic_json/
+    ~basic_json() noexcept
+    {
+        assert_invariant(false);
+    }
+
+    /// @}
+
+  public:
+    ///////////////////////
+    // object inspection //
+    ///////////////////////
+
+    /// @name object inspection
+    /// Functions to inspect the type of a JSON value.
+    /// @{
+
+    /// @brief serialization
+    /// @sa https://json.nlohmann.me/api/basic_json/dump/
+    string_t dump(const int indent = -1,
+                  const char indent_char = ' ',
+                  const bool ensure_ascii = false,
+                  const error_handler_t error_handler = error_handler_t::strict) const
+    {
+        string_t result;
+        serializer s(detail::output_adapter<char, string_t>(result), indent_char, error_handler);
+
+        if (indent >= 0)
+        {
+            s.dump(*this, true, ensure_ascii, static_cast<unsigned int>(indent));
+        }
+        else
+        {
+            s.dump(*this, false, ensure_ascii, 0);
+        }
+
+        return result;
+    }
+
+    /// @brief return the type of the JSON value (explicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/type/
+    constexpr value_t type() const noexcept
+    {
+        return m_data.m_type;
+    }
+
+    /// @brief return whether type is primitive
+    /// @sa https://json.nlohmann.me/api/basic_json/is_primitive/
+    constexpr bool is_primitive() const noexcept
+    {
+        return is_null() || is_string() || is_boolean() || is_number() || is_binary();
+    }
+
+    /// @brief return whether type is structured
+    /// @sa https://json.nlohmann.me/api/basic_json/is_structured/
+    constexpr bool is_structured() const noexcept
+    {
+        return is_array() || is_object();
+    }
+
+    /// @brief return whether value is null
+    /// @sa https://json.nlohmann.me/api/basic_json/is_null/
+    constexpr bool is_null() const noexcept
+    {
+        return m_data.m_type == value_t::null;
+    }
+
+    /// @brief return whether value is a boolean
+    /// @sa https://json.nlohmann.me/api/basic_json/is_boolean/
+    constexpr bool is_boolean() const noexcept
+    {
+        return m_data.m_type == value_t::boolean;
+    }
+
+    /// @brief return whether value is a number
+    /// @sa https://json.nlohmann.me/api/basic_json/is_number/
+    constexpr bool is_number() const noexcept
+    {
+        return is_number_integer() || is_number_float();
+    }
+
+    /// @brief return whether value is an integer number
+    /// @sa https://json.nlohmann.me/api/basic_json/is_number_integer/
+    constexpr bool is_number_integer() const noexcept
+    {
+        return m_data.m_type == value_t::number_integer || m_data.m_type == value_t::number_unsigned;
+    }
+
+    /// @brief return whether value is an unsigned integer number
+    /// @sa https://json.nlohmann.me/api/basic_json/is_number_unsigned/
+    constexpr bool is_number_unsigned() const noexcept
+    {
+        return m_data.m_type == value_t::number_unsigned;
+    }
+
+    /// @brief return whether value is a floating-point number
+    /// @sa https://json.nlohmann.me/api/basic_json/is_number_float/
+    constexpr bool is_number_float() const noexcept
+    {
+        return m_data.m_type == value_t::number_float;
+    }
+
+    /// @brief return whether value is an object
+    /// @sa https://json.nlohmann.me/api/basic_json/is_object/
+    constexpr bool is_object() const noexcept
+    {
+        return m_data.m_type == value_t::object;
+    }
+
+    /// @brief return whether value is an array
+    /// @sa https://json.nlohmann.me/api/basic_json/is_array/
+    constexpr bool is_array() const noexcept
+    {
+        return m_data.m_type == value_t::array;
+    }
+
+    /// @brief return whether value is a string
+    /// @sa https://json.nlohmann.me/api/basic_json/is_string/
+    constexpr bool is_string() const noexcept
+    {
+        return m_data.m_type == value_t::string;
+    }
+
+    /// @brief return whether value is a binary array
+    /// @sa https://json.nlohmann.me/api/basic_json/is_binary/
+    constexpr bool is_binary() const noexcept
+    {
+        return m_data.m_type == value_t::binary;
+    }
+
+    /// @brief return whether value is discarded
+    /// @sa https://json.nlohmann.me/api/basic_json/is_discarded/
+    constexpr bool is_discarded() const noexcept
+    {
+        return m_data.m_type == value_t::discarded;
+    }
+
+    /// @brief return the type of the JSON value (implicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_value_t/
+    constexpr operator value_t() const noexcept
+    {
+        return m_data.m_type;
+    }
+
+    /// @}
+
+  private:
+    //////////////////
+    // value access //
+    //////////////////
+
+    /// get a boolean (explicit)
+    boolean_t get_impl(boolean_t* /*unused*/) const
+    {
+        if (JSON_HEDLEY_LIKELY(is_boolean()))
+        {
+            return m_data.m_value.boolean;
+        }
+
+        JSON_THROW(type_error::create(302, detail::concat("type must be boolean, but is ", type_name()), this));
+    }
+
+    /// get a pointer to the value (object)
+    object_t* get_impl_ptr(object_t* /*unused*/) noexcept
+    {
+        return is_object() ? m_data.m_value.object : nullptr;
+    }
+
+    /// get a pointer to the value (object)
+    constexpr const object_t* get_impl_ptr(const object_t* /*unused*/) const noexcept
+    {
+        return is_object() ? m_data.m_value.object : nullptr;
+    }
+
+    /// get a pointer to the value (array)
+    array_t* get_impl_ptr(array_t* /*unused*/) noexcept
+    {
+        return is_array() ? m_data.m_value.array : nullptr;
+    }
+
+    /// get a pointer to the value (array)
+    constexpr const array_t* get_impl_ptr(const array_t* /*unused*/) const noexcept
+    {
+        return is_array() ? m_data.m_value.array : nullptr;
+    }
+
+    /// get a pointer to the value (string)
+    string_t* get_impl_ptr(string_t* /*unused*/) noexcept
+    {
+        return is_string() ? m_data.m_value.string : nullptr;
+    }
+
+    /// get a pointer to the value (string)
+    constexpr const string_t* get_impl_ptr(const string_t* /*unused*/) const noexcept
+    {
+        return is_string() ? m_data.m_value.string : nullptr;
+    }
+
+    /// get a pointer to the value (boolean)
+    boolean_t* get_impl_ptr(boolean_t* /*unused*/) noexcept
+    {
+        return is_boolean() ? &m_data.m_value.boolean : nullptr;
+    }
+
+    /// get a pointer to the value (boolean)
+    constexpr const boolean_t* get_impl_ptr(const boolean_t* /*unused*/) const noexcept
+    {
+        return is_boolean() ? &m_data.m_value.boolean : nullptr;
+    }
+
+    /// get a pointer to the value (integer number)
+    number_integer_t* get_impl_ptr(number_integer_t* /*unused*/) noexcept
+    {
+        return is_number_integer() ? &m_data.m_value.number_integer : nullptr;
+    }
+
+    /// get a pointer to the value (integer number)
+    constexpr const number_integer_t* get_impl_ptr(const number_integer_t* /*unused*/) const noexcept
+    {
+        return is_number_integer() ? &m_data.m_value.number_integer : nullptr;
+    }
+
+    /// get a pointer to the value (unsigned number)
+    number_unsigned_t* get_impl_ptr(number_unsigned_t* /*unused*/) noexcept
+    {
+        return is_number_unsigned() ? &m_data.m_value.number_unsigned : nullptr;
+    }
+
+    /// get a pointer to the value (unsigned number)
+    constexpr const number_unsigned_t* get_impl_ptr(const number_unsigned_t* /*unused*/) const noexcept
+    {
+        return is_number_unsigned() ? &m_data.m_value.number_unsigned : nullptr;
+    }
+
+    /// get a pointer to the value (floating-point number)
+    number_float_t* get_impl_ptr(number_float_t* /*unused*/) noexcept
+    {
+        return is_number_float() ? &m_data.m_value.number_float : nullptr;
+    }
+
+    /// get a pointer to the value (floating-point number)
+    constexpr const number_float_t* get_impl_ptr(const number_float_t* /*unused*/) const noexcept
+    {
+        return is_number_float() ? &m_data.m_value.number_float : nullptr;
+    }
+
+    /// get a pointer to the value (binary)
+    binary_t* get_impl_ptr(binary_t* /*unused*/) noexcept
+    {
+        return is_binary() ? m_data.m_value.binary : nullptr;
+    }
+
+    /// get a pointer to the value (binary)
+    constexpr const binary_t* get_impl_ptr(const binary_t* /*unused*/) const noexcept
+    {
+        return is_binary() ? m_data.m_value.binary : nullptr;
+    }
+
+    /*!
+    @brief helper function to implement get_ref()
+
+    This function helps to implement get_ref() without code duplication for
+    const and non-const overloads
+
+    @tparam ThisType will be deduced as `basic_json` or `const basic_json`
+
+    @throw type_error.303 if ReferenceType does not match underlying value
+    type of the current JSON
+    */
+    template<typename ReferenceType, typename ThisType>
+    static ReferenceType get_ref_impl(ThisType& obj)
+    {
+        // delegate the call to get_ptr<>()
+        auto* ptr = obj.template get_ptr<typename std::add_pointer<ReferenceType>::type>();
+
+        if (JSON_HEDLEY_LIKELY(ptr != nullptr))
+        {
+            return *ptr;
+        }
+
+        JSON_THROW(type_error::create(303, detail::concat("incompatible ReferenceType for get_ref, actual type is ", obj.type_name()), &obj));
+    }
+
+  public:
+    /// @name value access
+    /// Direct access to the stored value of a JSON value.
+    /// @{
+
+    /// @brief get a pointer value (implicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/get_ptr/
+    template<typename PointerType, typename std::enable_if<
+                 std::is_pointer<PointerType>::value, int>::type = 0>
+    auto get_ptr() noexcept -> decltype(std::declval<basic_json_t&>().get_impl_ptr(std::declval<PointerType>()))
+    {
+        // delegate the call to get_impl_ptr<>()
+        return get_impl_ptr(static_cast<PointerType>(nullptr));
+    }
+
+    /// @brief get a pointer value (implicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/get_ptr/
+    template < typename PointerType, typename std::enable_if <
+                   std::is_pointer<PointerType>::value&&
+                   std::is_const<typename std::remove_pointer<PointerType>::type>::value, int >::type = 0 >
+    constexpr auto get_ptr() const noexcept -> decltype(std::declval<const basic_json_t&>().get_impl_ptr(std::declval<PointerType>()))
+    {
+        // delegate the call to get_impl_ptr<>() const
+        return get_impl_ptr(static_cast<PointerType>(nullptr));
+    }
+
+  private:
+    /*!
+    @brief get a value (explicit)
+
+    Explicit type conversion between the JSON value and a compatible value
+    which is [CopyConstructible](https://en.cppreference.com/w/cpp/named_req/CopyConstructible)
+    and [DefaultConstructible](https://en.cppreference.com/w/cpp/named_req/DefaultConstructible).
+    The value is converted by calling the @ref json_serializer<ValueType>
+    `from_json()` method.
+
+    The function is equivalent to executing
+    @code {.cpp}
+    ValueType ret;
+    JSONSerializer<ValueType>::from_json(*this, ret);
+    return ret;
+    @endcode
+
+    This overloads is chosen if:
+    - @a ValueType is not @ref basic_json,
+    - @ref json_serializer<ValueType> has a `from_json()` method of the form
+      `void from_json(const basic_json&, ValueType&)`, and
+    - @ref json_serializer<ValueType> does not have a `from_json()` method of
+      the form `ValueType from_json(const basic_json&)`
+
+    @tparam ValueType the returned value type
+
+    @return copy of the JSON value, converted to @a ValueType
+
+    @throw what @ref json_serializer<ValueType> `from_json()` method throws
+
+    @liveexample{The example below shows several conversions from JSON values
+    to other types. There a few things to note: (1) Floating-point numbers can
+    be converted to integers\, (2) A JSON array can be converted to a standard
+    `std::vector<short>`\, (3) A JSON object can be converted to C++
+    associative containers such as `std::unordered_map<std::string\,
+    json>`.,get__ValueType_const}
+
+    @since version 2.1.0
+    */
+    template < typename ValueType,
+               detail::enable_if_t <
+                   detail::is_default_constructible<ValueType>::value&&
+                   detail::has_from_json<basic_json_t, ValueType>::value,
+                   int > = 0 >
+    ValueType get_impl(detail::priority_tag<0> /*unused*/) const noexcept(noexcept(
+                JSONSerializer<ValueType>::from_json(std::declval<const basic_json_t&>(), std::declval<ValueType&>())))
+    {
+        auto ret = ValueType();
+        JSONSerializer<ValueType>::from_json(*this, ret);
+        return ret;
+    }
+
+    /*!
+    @brief get a value (explicit); special case
+
+    Explicit type conversion between the JSON value and a compatible value
+    which is **not** [CopyConstructible](https://en.cppreference.com/w/cpp/named_req/CopyConstructible)
+    and **not** [DefaultConstructible](https://en.cppreference.com/w/cpp/named_req/DefaultConstructible).
+    The value is converted by calling the @ref json_serializer<ValueType>
+    `from_json()` method.
+
+    The function is equivalent to executing
+    @code {.cpp}
+    return JSONSerializer<ValueType>::from_json(*this);
+    @endcode
+
+    This overloads is chosen if:
+    - @a ValueType is not @ref basic_json and
+    - @ref json_serializer<ValueType> has a `from_json()` method of the form
+      `ValueType from_json(const basic_json&)`
+
+    @note If @ref json_serializer<ValueType> has both overloads of
+    `from_json()`, this one is chosen.
+
+    @tparam ValueType the returned value type
+
+    @return copy of the JSON value, converted to @a ValueType
+
+    @throw what @ref json_serializer<ValueType> `from_json()` method throws
+
+    @since version 2.1.0
+    */
+    template < typename ValueType,
+               detail::enable_if_t <
+                   detail::has_non_default_from_json<basic_json_t, ValueType>::value,
+                   int > = 0 >
+    ValueType get_impl(detail::priority_tag<1> /*unused*/) const noexcept(noexcept(
+                JSONSerializer<ValueType>::from_json(std::declval<const basic_json_t&>())))
+    {
+        return JSONSerializer<ValueType>::from_json(*this);
+    }
+
+    /*!
+    @brief get special-case overload
+
+    This overloads converts the current @ref basic_json in a different
+    @ref basic_json type
+
+    @tparam BasicJsonType == @ref basic_json
+
+    @return a copy of *this, converted into @a BasicJsonType
+
+    @complexity Depending on the implementation of the called `from_json()`
+                method.
+
+    @since version 3.2.0
+    */
+    template < typename BasicJsonType,
+               detail::enable_if_t <
+                   detail::is_basic_json<BasicJsonType>::value,
+                   int > = 0 >
+    BasicJsonType get_impl(detail::priority_tag<2> /*unused*/) const
+    {
+        return *this;
+    }
+
+    /*!
+    @brief get special-case overload
+
+    This overloads avoids a lot of template boilerplate, it can be seen as the
+    identity method
+
+    @tparam BasicJsonType == @ref basic_json
+
+    @return a copy of *this
+
+    @complexity Constant.
+
+    @since version 2.1.0
+    */
+    template<typename BasicJsonType,
+             detail::enable_if_t<
+                 std::is_same<BasicJsonType, basic_json_t>::value,
+                 int> = 0>
+    basic_json get_impl(detail::priority_tag<3> /*unused*/) const
+    {
+        return *this;
+    }
+
+    /*!
+    @brief get a pointer value (explicit)
+    @copydoc get()
+    */
+    template<typename PointerType,
+             detail::enable_if_t<
+                 std::is_pointer<PointerType>::value,
+                 int> = 0>
+    constexpr auto get_impl(detail::priority_tag<4> /*unused*/) const noexcept
+    -> decltype(std::declval<const basic_json_t&>().template get_ptr<PointerType>())
+    {
+        // delegate the call to get_ptr
+        return get_ptr<PointerType>();
+    }
+
+  public:
+    /*!
+    @brief get a (pointer) value (explicit)
+
+    Performs explicit type conversion between the JSON value and a compatible value if required.
+
+    - If the requested type is a pointer to the internally stored JSON value that pointer is returned.
+    No copies are made.
+
+    - If the requested type is the current @ref basic_json, or a different @ref basic_json convertible
+    from the current @ref basic_json.
+
+    - Otherwise the value is converted by calling the @ref json_serializer<ValueType> `from_json()`
+    method.
+
+    @tparam ValueTypeCV the provided value type
+    @tparam ValueType the returned value type
+
+    @return copy of the JSON value, converted to @tparam ValueType if necessary
+
+    @throw what @ref json_serializer<ValueType> `from_json()` method throws if conversion is required
+
+    @since version 2.1.0
+    */
+    template < typename ValueTypeCV, typename ValueType = detail::uncvref_t<ValueTypeCV>>
+#if defined(JSON_HAS_CPP_14)
+    constexpr
+#endif
+    auto get() const noexcept(
+    noexcept(std::declval<const basic_json_t&>().template get_impl<ValueType>(detail::priority_tag<4> {})))
+    -> decltype(std::declval<const basic_json_t&>().template get_impl<ValueType>(detail::priority_tag<4> {}))
+    {
+        // we cannot static_assert on ValueTypeCV being non-const, because
+        // there is support for get<const basic_json_t>(), which is why we
+        // still need the uncvref
+        static_assert(!std::is_reference<ValueTypeCV>::value,
+                      "get() cannot be used with reference types, you might want to use get_ref()");
+        return get_impl<ValueType>(detail::priority_tag<4> {});
+    }
+
+    /*!
+    @brief get a pointer value (explicit)
+
+    Explicit pointer access to the internally stored JSON value. No copies are
+    made.
+
+    @warning The pointer becomes invalid if the underlying JSON object
+    changes.
+
+    @tparam PointerType pointer type; must be a pointer to @ref array_t, @ref
+    object_t, @ref string_t, @ref boolean_t, @ref number_integer_t,
+    @ref number_unsigned_t, or @ref number_float_t.
+
+    @return pointer to the internally stored JSON value if the requested
+    pointer type @a PointerType fits to the JSON value; `nullptr` otherwise
+
+    @complexity Constant.
+
+    @liveexample{The example below shows how pointers to internal values of a
+    JSON value can be requested. Note that no type conversions are made and a
+    `nullptr` is returned if the value and the requested pointer type does not
+    match.,get__PointerType}
+
+    @sa see @ref get_ptr() for explicit pointer-member access
+
+    @since version 1.0.0
+    */
+    template<typename PointerType, typename std::enable_if<
+                 std::is_pointer<PointerType>::value, int>::type = 0>
+    auto get() noexcept -> decltype(std::declval<basic_json_t&>().template get_ptr<PointerType>())
+    {
+        // delegate the call to get_ptr
+        return get_ptr<PointerType>();
+    }
+
+    /// @brief get a value (explicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/get_to/
+    template < typename ValueType,
+               detail::enable_if_t <
+                   !detail::is_basic_json<ValueType>::value&&
+                   detail::has_from_json<basic_json_t, ValueType>::value,
+                   int > = 0 >
+    ValueType & get_to(ValueType& v) const noexcept(noexcept(
+                JSONSerializer<ValueType>::from_json(std::declval<const basic_json_t&>(), v)))
+    {
+        JSONSerializer<ValueType>::from_json(*this, v);
+        return v;
+    }
+
+    // specialization to allow calling get_to with a basic_json value
+    // see https://github.com/nlohmann/json/issues/2175
+    template<typename ValueType,
+             detail::enable_if_t <
+                 detail::is_basic_json<ValueType>::value,
+                 int> = 0>
+    ValueType & get_to(ValueType& v) const
+    {
+        v = *this;
+        return v;
+    }
+
+    template <
+        typename T, std::size_t N,
+        typename Array = T (&)[N], // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+        detail::enable_if_t <
+            detail::has_from_json<basic_json_t, Array>::value, int > = 0 >
+    Array get_to(T (&v)[N]) const // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+    noexcept(noexcept(JSONSerializer<Array>::from_json(
+                          std::declval<const basic_json_t&>(), v)))
+    {
+        JSONSerializer<Array>::from_json(*this, v);
+        return v;
+    }
+
+    /// @brief get a reference value (implicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/get_ref/
+    template<typename ReferenceType, typename std::enable_if<
+                 std::is_reference<ReferenceType>::value, int>::type = 0>
+    ReferenceType get_ref()
+    {
+        // delegate call to get_ref_impl
+        return get_ref_impl<ReferenceType>(*this);
+    }
+
+    /// @brief get a reference value (implicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/get_ref/
+    template < typename ReferenceType, typename std::enable_if <
+                   std::is_reference<ReferenceType>::value&&
+                   std::is_const<typename std::remove_reference<ReferenceType>::type>::value, int >::type = 0 >
+    ReferenceType get_ref() const
+    {
+        // delegate call to get_ref_impl
+        return get_ref_impl<ReferenceType>(*this);
+    }
+
+    /*!
+    @brief get a value (implicit)
+
+    Implicit type conversion between the JSON value and a compatible value.
+    The call is realized by calling @ref get() const.
+
+    @tparam ValueType non-pointer type compatible to the JSON value, for
+    instance `int` for JSON integer numbers, `bool` for JSON booleans, or
+    `std::vector` types for JSON arrays. The character type of @ref string_t
+    as well as an initializer list of this type is excluded to avoid
+    ambiguities as these types implicitly convert to `std::string`.
+
+    @return copy of the JSON value, converted to type @a ValueType
+
+    @throw type_error.302 in case passed type @a ValueType is incompatible
+    to the JSON value type (e.g., the JSON value is of type boolean, but a
+    string is requested); see example below
+
+    @complexity Linear in the size of the JSON value.
+
+    @liveexample{The example below shows several conversions from JSON values
+    to other types. There a few things to note: (1) Floating-point numbers can
+    be converted to integers\, (2) A JSON array can be converted to a standard
+    `std::vector<short>`\, (3) A JSON object can be converted to C++
+    associative containers such as `std::unordered_map<std::string\,
+    json>`.,operator__ValueType}
+
+    @since version 1.0.0
+    */
+    template < typename ValueType, typename std::enable_if <
+                   detail::conjunction <
+                       detail::negation<std::is_pointer<ValueType>>,
+                       detail::negation<std::is_same<ValueType, std::nullptr_t>>,
+                       detail::negation<std::is_same<ValueType, detail::json_ref<basic_json>>>,
+                                        detail::negation<std::is_same<ValueType, typename string_t::value_type>>,
+                                        detail::negation<detail::is_basic_json<ValueType>>,
+                                        detail::negation<std::is_same<ValueType, std::initializer_list<typename string_t::value_type>>>,
+#if defined(JSON_HAS_CPP_17) && (defined(__GNUC__) || (defined(_MSC_VER) && _MSC_VER >= 1910 && _MSC_VER <= 1914))
+                                                detail::negation<std::is_same<ValueType, std::string_view>>,
+#endif
+#if defined(JSON_HAS_CPP_17) && JSON_HAS_STATIC_RTTI
+                                                detail::negation<std::is_same<ValueType, std::any>>,
+#endif
+                                                detail::is_detected_lazy<detail::get_template_function, const basic_json_t&, ValueType>
+                                                >::value, int >::type = 0 >
+                                        JSON_EXPLICIT operator ValueType() const
+    {
+        // delegate the call to get<>() const
+        return get<ValueType>();
+    }
+
+    /// @brief get a binary value
+    /// @sa https://json.nlohmann.me/api/basic_json/get_binary/
+    binary_t& get_binary()
+    {
+        if (!is_binary())
+        {
+            JSON_THROW(type_error::create(302, detail::concat("type must be binary, but is ", type_name()), this));
+        }
+
+        return *get_ptr<binary_t*>();
+    }
+
+    /// @brief get a binary value
+    /// @sa https://json.nlohmann.me/api/basic_json/get_binary/
+    const binary_t& get_binary() const
+    {
+        if (!is_binary())
+        {
+            JSON_THROW(type_error::create(302, detail::concat("type must be binary, but is ", type_name()), this));
+        }
+
+        return *get_ptr<const binary_t*>();
+    }
+
+    /// @}
+
+    ////////////////////
+    // element access //
+    ////////////////////
+
+    /// @name element access
+    /// Access to the JSON value.
+    /// @{
+
+    /// @brief access specified array element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    reference at(size_type idx)
+    {
+        // at only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            JSON_TRY
+            {
+                return set_parent(m_data.m_value.array->at(idx));
+            }
+            JSON_CATCH (std::out_of_range&)
+            {
+                // create better exception explanation
+                JSON_THROW(out_of_range::create(401, detail::concat("array index ", std::to_string(idx), " is out of range"), this));
+            }
+        }
+        else
+        {
+            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
+        }
+    }
+
+    /// @brief access specified array element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    const_reference at(size_type idx) const
+    {
+        // at only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            JSON_TRY
+            {
+                return m_data.m_value.array->at(idx);
+            }
+            JSON_CATCH (std::out_of_range&)
+            {
+                // create better exception explanation
+                JSON_THROW(out_of_range::create(401, detail::concat("array index ", std::to_string(idx), " is out of range"), this));
+            }
+        }
+        else
+        {
+            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
+        }
+    }
+
+    /// @brief access specified object element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    reference at(const typename object_t::key_type& key)
+    {
+        // at only works for objects
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
+        }
+
+        auto it = m_data.m_value.object->find(key);
+        if (it == m_data.m_value.object->end())
+        {
+            JSON_THROW(out_of_range::create(403, detail::concat("key '", key, "' not found"), this));
+        }
+        return set_parent(it->second);
+    }
+
+    /// @brief access specified object element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
+    reference at(KeyType && key)
+    {
+        // at only works for objects
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
+        }
+
+        auto it = m_data.m_value.object->find(std::forward<KeyType>(key));
+        if (it == m_data.m_value.object->end())
+        {
+            JSON_THROW(out_of_range::create(403, detail::concat("key '", string_t(std::forward<KeyType>(key)), "' not found"), this));
+        }
+        return set_parent(it->second);
+    }
+
+    /// @brief access specified object element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    const_reference at(const typename object_t::key_type& key) const
+    {
+        // at only works for objects
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
+        }
+
+        auto it = m_data.m_value.object->find(key);
+        if (it == m_data.m_value.object->end())
+        {
+            JSON_THROW(out_of_range::create(403, detail::concat("key '", key, "' not found"), this));
+        }
+        return it->second;
+    }
+
+    /// @brief access specified object element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
+    const_reference at(KeyType && key) const
+    {
+        // at only works for objects
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
+        }
+
+        auto it = m_data.m_value.object->find(std::forward<KeyType>(key));
+        if (it == m_data.m_value.object->end())
+        {
+            JSON_THROW(out_of_range::create(403, detail::concat("key '", string_t(std::forward<KeyType>(key)), "' not found"), this));
+        }
+        return it->second;
+    }
+
+    /// @brief access specified array element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    reference operator[](size_type idx)
+    {
+        // implicitly convert null value to an empty array
+        if (is_null())
+        {
+            m_data.m_type = value_t::array;
+            m_data.m_value.array = create<array_t>();
+            assert_invariant();
+        }
+
+        // operator[] only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            // fill up array with null values if given idx is outside range
+            if (idx >= m_data.m_value.array->size())
+            {
+#if JSON_DIAGNOSTICS
+                // remember array size & capacity before resizing
+                const auto old_size = m_data.m_value.array->size();
+                const auto old_capacity = m_data.m_value.array->capacity();
+#endif
+                m_data.m_value.array->resize(idx + 1);
+
+#if JSON_DIAGNOSTICS
+                if (JSON_HEDLEY_UNLIKELY(m_data.m_value.array->capacity() != old_capacity))
+                {
+                    // capacity has changed: update all parents
+                    set_parents();
+                }
+                else
+                {
+                    // set parent for values added above
+                    set_parents(begin() + static_cast<typename iterator::difference_type>(old_size), static_cast<typename iterator::difference_type>(idx + 1 - old_size));
+                }
+#endif
+                assert_invariant();
+            }
+
+            return m_data.m_value.array->operator[](idx);
+        }
+
+        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a numeric argument with ", type_name()), this));
+    }
+
+    /// @brief access specified array element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    const_reference operator[](size_type idx) const
+    {
+        // const operator[] only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            return m_data.m_value.array->operator[](idx);
+        }
+
+        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a numeric argument with ", type_name()), this));
+    }
+
+    /// @brief access specified object element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    reference operator[](typename object_t::key_type key)
+    {
+        // implicitly convert null value to an empty object
+        if (is_null())
+        {
+            m_data.m_type = value_t::object;
+            m_data.m_value.object = create<object_t>();
+            assert_invariant();
+        }
+
+        // operator[] only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            auto result = m_data.m_value.object->emplace(std::move(key), nullptr);
+            return set_parent(result.first->second);
+        }
+
+        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a string argument with ", type_name()), this));
+    }
+
+    /// @brief access specified object element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    const_reference operator[](const typename object_t::key_type& key) const
+    {
+        // const operator[] only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            auto it = m_data.m_value.object->find(key);
+            JSON_ASSERT(it != m_data.m_value.object->end());
+            return it->second;
+        }
+
+        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a string argument with ", type_name()), this));
+    }
+
+    // these two functions resolve a (const) char * ambiguity affecting Clang and MSVC
+    // (they seemingly cannot be constrained to resolve the ambiguity)
+    template<typename T>
+    reference operator[](T* key)
+    {
+        return operator[](typename object_t::key_type(key));
+    }
+
+    template<typename T>
+    const_reference operator[](T* key) const
+    {
+        return operator[](typename object_t::key_type(key));
+    }
+
+    /// @brief access specified object element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int > = 0 >
+    reference operator[](KeyType && key)
+    {
+        // implicitly convert null value to an empty object
+        if (is_null())
+        {
+            m_data.m_type = value_t::object;
+            m_data.m_value.object = create<object_t>();
+            assert_invariant();
+        }
+
+        // operator[] only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            auto result = m_data.m_value.object->emplace(std::forward<KeyType>(key), nullptr);
+            return set_parent(result.first->second);
+        }
+
+        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a string argument with ", type_name()), this));
+    }
+
+    /// @brief access specified object element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int > = 0 >
+    const_reference operator[](KeyType && key) const
+    {
+        // const operator[] only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            auto it = m_data.m_value.object->find(std::forward<KeyType>(key));
+            JSON_ASSERT(it != m_data.m_value.object->end());
+            return it->second;
+        }
+
+        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a string argument with ", type_name()), this));
+    }
+
+  private:
+    template<typename KeyType>
+    using is_comparable_with_object_key = detail::is_comparable <
+        object_comparator_t, const typename object_t::key_type&, KeyType >;
+
+    template<typename ValueType>
+    using value_return_type = std::conditional <
+        detail::is_c_string_uncvref<ValueType>::value,
+        string_t, typename std::decay<ValueType>::type >;
+
+  public:
+    /// @brief access specified object element with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    template < class ValueType, detail::enable_if_t <
+                   !detail::is_transparent<object_comparator_t>::value
+                   && detail::is_getable<basic_json_t, ValueType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    ValueType value(const typename object_t::key_type& key, const ValueType& default_value) const
+    {
+        // value only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            // if key is found, return value and given default value otherwise
+            const auto it = find(key);
+            if (it != end())
+            {
+                return it->template get<ValueType>();
+            }
+
+            return default_value;
+        }
+
+        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
+    }
+
+    /// @brief access specified object element with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    template < class ValueType, class ReturnType = typename value_return_type<ValueType>::type,
+               detail::enable_if_t <
+                   !detail::is_transparent<object_comparator_t>::value
+                   && detail::is_getable<basic_json_t, ReturnType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    ReturnType value(const typename object_t::key_type& key, ValueType && default_value) const
+    {
+        // value only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            // if key is found, return value and given default value otherwise
+            const auto it = find(key);
+            if (it != end())
+            {
+                return it->template get<ReturnType>();
+            }
+
+            return std::forward<ValueType>(default_value);
+        }
+
+        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
+    }
+
+    /// @brief access specified object element with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    template < class ValueType, class KeyType, detail::enable_if_t <
+                   detail::is_transparent<object_comparator_t>::value
+                   && !detail::is_json_pointer<KeyType>::value
+                   && is_comparable_with_object_key<KeyType>::value
+                   && detail::is_getable<basic_json_t, ValueType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    ValueType value(KeyType && key, const ValueType& default_value) const
+    {
+        // value only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            // if key is found, return value and given default value otherwise
+            const auto it = find(std::forward<KeyType>(key));
+            if (it != end())
+            {
+                return it->template get<ValueType>();
+            }
+
+            return default_value;
+        }
+
+        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
+    }
+
+    /// @brief access specified object element via JSON Pointer with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    template < class ValueType, class KeyType, class ReturnType = typename value_return_type<ValueType>::type,
+               detail::enable_if_t <
+                   detail::is_transparent<object_comparator_t>::value
+                   && !detail::is_json_pointer<KeyType>::value
+                   && is_comparable_with_object_key<KeyType>::value
+                   && detail::is_getable<basic_json_t, ReturnType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    ReturnType value(KeyType && key, ValueType && default_value) const
+    {
+        // value only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            // if key is found, return value and given default value otherwise
+            const auto it = find(std::forward<KeyType>(key));
+            if (it != end())
+            {
+                return it->template get<ReturnType>();
+            }
+
+            return std::forward<ValueType>(default_value);
+        }
+
+        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
+    }
+
+    /// @brief access specified object element via JSON Pointer with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    template < class ValueType, detail::enable_if_t <
+                   detail::is_getable<basic_json_t, ValueType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    ValueType value(const json_pointer& ptr, const ValueType& default_value) const
+    {
+        // value only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            // if pointer resolves a value, return it or use default value
+            JSON_TRY
+            {
+                return ptr.get_checked(this).template get<ValueType>();
+            }
+            JSON_INTERNAL_CATCH (out_of_range&)
+            {
+                return default_value;
+            }
+        }
+
+        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
+    }
+
+    /// @brief access specified object element via JSON Pointer with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    template < class ValueType, class ReturnType = typename value_return_type<ValueType>::type,
+               detail::enable_if_t <
+                   detail::is_getable<basic_json_t, ReturnType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    ReturnType value(const json_pointer& ptr, ValueType && default_value) const
+    {
+        // value only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            // if pointer resolves a value, return it or use default value
+            JSON_TRY
+            {
+                return ptr.get_checked(this).template get<ReturnType>();
+            }
+            JSON_INTERNAL_CATCH (out_of_range&)
+            {
+                return std::forward<ValueType>(default_value);
+            }
+        }
+
+        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
+    }
+
+    template < class ValueType, class BasicJsonType, detail::enable_if_t <
+                   detail::is_basic_json<BasicJsonType>::value
+                   && detail::is_getable<basic_json_t, ValueType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
+    ValueType value(const ::nlohmann::json_pointer<BasicJsonType>& ptr, const ValueType& default_value) const
+    {
+        return value(ptr.convert(), default_value);
+    }
+
+    template < class ValueType, class BasicJsonType, class ReturnType = typename value_return_type<ValueType>::type,
+               detail::enable_if_t <
+                   detail::is_basic_json<BasicJsonType>::value
+                   && detail::is_getable<basic_json_t, ReturnType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
+    ReturnType value(const ::nlohmann::json_pointer<BasicJsonType>& ptr, ValueType && default_value) const
+    {
+        return value(ptr.convert(), std::forward<ValueType>(default_value));
+    }
+
+    /// @brief access the first element
+    /// @sa https://json.nlohmann.me/api/basic_json/front/
+    reference front()
+    {
+        return *begin();
+    }
+
+    /// @brief access the first element
+    /// @sa https://json.nlohmann.me/api/basic_json/front/
+    const_reference front() const
+    {
+        return *cbegin();
+    }
+
+    /// @brief access the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/back/
+    reference back()
+    {
+        auto tmp = end();
+        --tmp;
+        return *tmp;
+    }
+
+    /// @brief access the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/back/
+    const_reference back() const
+    {
+        auto tmp = cend();
+        --tmp;
+        return *tmp;
+    }
+
+    /// @brief remove element given an iterator
+    /// @sa https://json.nlohmann.me/api/basic_json/erase/
+    template < class IteratorType, detail::enable_if_t <
+                   std::is_same<IteratorType, typename basic_json_t::iterator>::value ||
+                   std::is_same<IteratorType, typename basic_json_t::const_iterator>::value, int > = 0 >
+    IteratorType erase(IteratorType pos)
+    {
+        // make sure iterator fits the current value
+        if (JSON_HEDLEY_UNLIKELY(this != pos.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", this));
+        }
+
+        IteratorType result = end();
+
+        switch (m_data.m_type)
+        {
+            case value_t::boolean:
+            case value_t::number_float:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::string:
+            case value_t::binary:
+            {
+                if (JSON_HEDLEY_UNLIKELY(!pos.m_it.primitive_iterator.is_begin()))
+                {
+                    JSON_THROW(invalid_iterator::create(205, "iterator out of range", this));
+                }
+
+                if (is_string())
+                {
+                    AllocatorType<string_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_data.m_value.string);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_data.m_value.string, 1);
+                    m_data.m_value.string = nullptr;
+                }
+                else if (is_binary())
+                {
+                    AllocatorType<binary_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_data.m_value.binary);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_data.m_value.binary, 1);
+                    m_data.m_value.binary = nullptr;
+                }
+
+                m_data.m_type = value_t::null;
+                assert_invariant();
+                break;
+            }
+
+            case value_t::object:
+            {
+                result.m_it.object_iterator = m_data.m_value.object->erase(pos.m_it.object_iterator);
+                break;
+            }
+
+            case value_t::array:
+            {
+                result.m_it.array_iterator = m_data.m_value.array->erase(pos.m_it.array_iterator);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::discarded:
+            default:
+                JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this));
+        }
+
+        return result;
+    }
+
+    /// @brief remove elements given an iterator range
+    /// @sa https://json.nlohmann.me/api/basic_json/erase/
+    template < class IteratorType, detail::enable_if_t <
+                   std::is_same<IteratorType, typename basic_json_t::iterator>::value ||
+                   std::is_same<IteratorType, typename basic_json_t::const_iterator>::value, int > = 0 >
+    IteratorType erase(IteratorType first, IteratorType last)
+    {
+        // make sure iterator fits the current value
+        if (JSON_HEDLEY_UNLIKELY(this != first.m_object || this != last.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(203, "iterators do not fit current value", this));
+        }
+
+        IteratorType result = end();
+
+        switch (m_data.m_type)
+        {
+            case value_t::boolean:
+            case value_t::number_float:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::string:
+            case value_t::binary:
+            {
+                if (JSON_HEDLEY_LIKELY(!first.m_it.primitive_iterator.is_begin()
+                                       || !last.m_it.primitive_iterator.is_end()))
+                {
+                    JSON_THROW(invalid_iterator::create(204, "iterators out of range", this));
+                }
+
+                if (is_string())
+                {
+                    AllocatorType<string_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_data.m_value.string);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_data.m_value.string, 1);
+                    m_data.m_value.string = nullptr;
+                }
+                else if (is_binary())
+                {
+                    AllocatorType<binary_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_data.m_value.binary);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_data.m_value.binary, 1);
+                    m_data.m_value.binary = nullptr;
+                }
+
+                m_data.m_type = value_t::null;
+                assert_invariant();
+                break;
+            }
+
+            case value_t::object:
+            {
+                result.m_it.object_iterator = m_data.m_value.object->erase(first.m_it.object_iterator,
+                                              last.m_it.object_iterator);
+                break;
+            }
+
+            case value_t::array:
+            {
+                result.m_it.array_iterator = m_data.m_value.array->erase(first.m_it.array_iterator,
+                                             last.m_it.array_iterator);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::discarded:
+            default:
+                JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this));
+        }
+
+        return result;
+    }
+
+  private:
+    template < typename KeyType, detail::enable_if_t <
+                   detail::has_erase_with_key_type<basic_json_t, KeyType>::value, int > = 0 >
+    size_type erase_internal(KeyType && key)
+    {
+        // this erase only works for objects
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this));
+        }
+
+        return m_data.m_value.object->erase(std::forward<KeyType>(key));
+    }
+
+    template < typename KeyType, detail::enable_if_t <
+                   !detail::has_erase_with_key_type<basic_json_t, KeyType>::value, int > = 0 >
+    size_type erase_internal(KeyType && key)
+    {
+        // this erase only works for objects
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this));
+        }
+
+        const auto it = m_data.m_value.object->find(std::forward<KeyType>(key));
+        if (it != m_data.m_value.object->end())
+        {
+            m_data.m_value.object->erase(it);
+            return 1;
+        }
+        return 0;
+    }
+
+  public:
+
+    /// @brief remove element from a JSON object given a key
+    /// @sa https://json.nlohmann.me/api/basic_json/erase/
+    size_type erase(const typename object_t::key_type& key)
+    {
+        // the indirection via erase_internal() is added to avoid making this
+        // function a template and thus de-rank it during overload resolution
+        return erase_internal(key);
+    }
+
+    /// @brief remove element from a JSON object given a key
+    /// @sa https://json.nlohmann.me/api/basic_json/erase/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
+    size_type erase(KeyType && key)
+    {
+        return erase_internal(std::forward<KeyType>(key));
+    }
+
+    /// @brief remove element from a JSON array given an index
+    /// @sa https://json.nlohmann.me/api/basic_json/erase/
+    void erase(const size_type idx)
+    {
+        // this erase only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            if (JSON_HEDLEY_UNLIKELY(idx >= size()))
+            {
+                JSON_THROW(out_of_range::create(401, detail::concat("array index ", std::to_string(idx), " is out of range"), this));
+            }
+
+            m_data.m_value.array->erase(m_data.m_value.array->begin() + static_cast<difference_type>(idx));
+        }
+        else
+        {
+            JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this));
+        }
+    }
+
+    /// @}
+
+    ////////////
+    // lookup //
+    ////////////
+
+    /// @name lookup
+    /// @{
+
+    /// @brief find an element in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/find/
+    iterator find(const typename object_t::key_type& key)
+    {
+        auto result = end();
+
+        if (is_object())
+        {
+            result.m_it.object_iterator = m_data.m_value.object->find(key);
+        }
+
+        return result;
+    }
+
+    /// @brief find an element in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/find/
+    const_iterator find(const typename object_t::key_type& key) const
+    {
+        auto result = cend();
+
+        if (is_object())
+        {
+            result.m_it.object_iterator = m_data.m_value.object->find(key);
+        }
+
+        return result;
+    }
+
+    /// @brief find an element in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/find/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
+    iterator find(KeyType && key)
+    {
+        auto result = end();
+
+        if (is_object())
+        {
+            result.m_it.object_iterator = m_data.m_value.object->find(std::forward<KeyType>(key));
+        }
+
+        return result;
+    }
+
+    /// @brief find an element in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/find/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
+    const_iterator find(KeyType && key) const
+    {
+        auto result = cend();
+
+        if (is_object())
+        {
+            result.m_it.object_iterator = m_data.m_value.object->find(std::forward<KeyType>(key));
+        }
+
+        return result;
+    }
+
+    /// @brief returns the number of occurrences of a key in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/count/
+    size_type count(const typename object_t::key_type& key) const
+    {
+        // return 0 for all nonobject types
+        return is_object() ? m_data.m_value.object->count(key) : 0;
+    }
+
+    /// @brief returns the number of occurrences of a key in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/count/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
+    size_type count(KeyType && key) const
+    {
+        // return 0 for all nonobject types
+        return is_object() ? m_data.m_value.object->count(std::forward<KeyType>(key)) : 0;
+    }
+
+    /// @brief check the existence of an element in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/contains/
+    bool contains(const typename object_t::key_type& key) const
+    {
+        return is_object() && m_data.m_value.object->find(key) != m_data.m_value.object->end();
+    }
+
+    /// @brief check the existence of an element in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/contains/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
+    bool contains(KeyType && key) const
+    {
+        return is_object() && m_data.m_value.object->find(std::forward<KeyType>(key)) != m_data.m_value.object->end();
+    }
+
+    /// @brief check the existence of an element in a JSON object given a JSON pointer
+    /// @sa https://json.nlohmann.me/api/basic_json/contains/
+    bool contains(const json_pointer& ptr) const
+    {
+        return ptr.contains(this);
+    }
+
+    template<typename BasicJsonType, detail::enable_if_t<detail::is_basic_json<BasicJsonType>::value, int> = 0>
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
+    bool contains(const typename ::nlohmann::json_pointer<BasicJsonType>& ptr) const
+    {
+        return ptr.contains(this);
+    }
+
+    /// @}
+
+    ///////////////
+    // iterators //
+    ///////////////
+
+    /// @name iterators
+    /// @{
+
+    /// @brief returns an iterator to the first element
+    /// @sa https://json.nlohmann.me/api/basic_json/begin/
+    iterator begin() noexcept
+    {
+        iterator result(this);
+        result.set_begin();
+        return result;
+    }
+
+    /// @brief returns an iterator to the first element
+    /// @sa https://json.nlohmann.me/api/basic_json/begin/
+    const_iterator begin() const noexcept
+    {
+        return cbegin();
+    }
+
+    /// @brief returns a const iterator to the first element
+    /// @sa https://json.nlohmann.me/api/basic_json/cbegin/
+    const_iterator cbegin() const noexcept
+    {
+        const_iterator result(this);
+        result.set_begin();
+        return result;
+    }
+
+    /// @brief returns an iterator to one past the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/end/
+    iterator end() noexcept
+    {
+        iterator result(this);
+        result.set_end();
+        return result;
+    }
+
+    /// @brief returns an iterator to one past the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/end/
+    const_iterator end() const noexcept
+    {
+        return cend();
+    }
+
+    /// @brief returns an iterator to one past the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/cend/
+    const_iterator cend() const noexcept
+    {
+        const_iterator result(this);
+        result.set_end();
+        return result;
+    }
+
+    /// @brief returns an iterator to the reverse-beginning
+    /// @sa https://json.nlohmann.me/api/basic_json/rbegin/
+    reverse_iterator rbegin() noexcept
+    {
+        return reverse_iterator(end());
+    }
+
+    /// @brief returns an iterator to the reverse-beginning
+    /// @sa https://json.nlohmann.me/api/basic_json/rbegin/
+    const_reverse_iterator rbegin() const noexcept
+    {
+        return crbegin();
+    }
+
+    /// @brief returns an iterator to the reverse-end
+    /// @sa https://json.nlohmann.me/api/basic_json/rend/
+    reverse_iterator rend() noexcept
+    {
+        return reverse_iterator(begin());
+    }
+
+    /// @brief returns an iterator to the reverse-end
+    /// @sa https://json.nlohmann.me/api/basic_json/rend/
+    const_reverse_iterator rend() const noexcept
+    {
+        return crend();
+    }
+
+    /// @brief returns a const reverse iterator to the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/crbegin/
+    const_reverse_iterator crbegin() const noexcept
+    {
+        return const_reverse_iterator(cend());
+    }
+
+    /// @brief returns a const reverse iterator to one before the first
+    /// @sa https://json.nlohmann.me/api/basic_json/crend/
+    const_reverse_iterator crend() const noexcept
+    {
+        return const_reverse_iterator(cbegin());
+    }
+
+  public:
+    /// @brief wrapper to access iterator member functions in range-based for
+    /// @sa https://json.nlohmann.me/api/basic_json/items/
+    /// @deprecated This function is deprecated since 3.1.0 and will be removed in
+    ///             version 4.0.0 of the library. Please use @ref items() instead;
+    ///             that is, replace `json::iterator_wrapper(j)` with `j.items()`.
+    JSON_HEDLEY_DEPRECATED_FOR(3.1.0, items())
+    static iteration_proxy<iterator> iterator_wrapper(reference ref) noexcept
+    {
+        return ref.items();
+    }
+
+    /// @brief wrapper to access iterator member functions in range-based for
+    /// @sa https://json.nlohmann.me/api/basic_json/items/
+    /// @deprecated This function is deprecated since 3.1.0 and will be removed in
+    ///         version 4.0.0 of the library. Please use @ref items() instead;
+    ///         that is, replace `json::iterator_wrapper(j)` with `j.items()`.
+    JSON_HEDLEY_DEPRECATED_FOR(3.1.0, items())
+    static iteration_proxy<const_iterator> iterator_wrapper(const_reference ref) noexcept
+    {
+        return ref.items();
+    }
+
+    /// @brief helper to access iterator member functions in range-based for
+    /// @sa https://json.nlohmann.me/api/basic_json/items/
+    iteration_proxy<iterator> items() noexcept
+    {
+        return iteration_proxy<iterator>(*this);
+    }
+
+    /// @brief helper to access iterator member functions in range-based for
+    /// @sa https://json.nlohmann.me/api/basic_json/items/
+    iteration_proxy<const_iterator> items() const noexcept
+    {
+        return iteration_proxy<const_iterator>(*this);
+    }
+
+    /// @}
+
+    //////////////
+    // capacity //
+    //////////////
+
+    /// @name capacity
+    /// @{
+
+    /// @brief checks whether the container is empty.
+    /// @sa https://json.nlohmann.me/api/basic_json/empty/
+    bool empty() const noexcept
+    {
+        switch (m_data.m_type)
+        {
+            case value_t::null:
+            {
+                // null values are empty
+                return true;
+            }
+
+            case value_t::array:
+            {
+                // delegate call to array_t::empty()
+                return m_data.m_value.array->empty();
+            }
+
+            case value_t::object:
+            {
+                // delegate call to object_t::empty()
+                return m_data.m_value.object->empty();
+            }
+
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                // all other types are nonempty
+                return false;
+            }
+        }
+    }
+
+    /// @brief returns the number of elements
+    /// @sa https://json.nlohmann.me/api/basic_json/size/
+    size_type size() const noexcept
+    {
+        switch (m_data.m_type)
+        {
+            case value_t::null:
+            {
+                // null values are empty
+                return 0;
+            }
+
+            case value_t::array:
+            {
+                // delegate call to array_t::size()
+                return m_data.m_value.array->size();
+            }
+
+            case value_t::object:
+            {
+                // delegate call to object_t::size()
+                return m_data.m_value.object->size();
+            }
+
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                // all other types have size 1
+                return 1;
+            }
+        }
+    }
+
+    /// @brief returns the maximum possible number of elements
+    /// @sa https://json.nlohmann.me/api/basic_json/max_size/
+    size_type max_size() const noexcept
+    {
+        switch (m_data.m_type)
+        {
+            case value_t::array:
+            {
+                // delegate call to array_t::max_size()
+                return m_data.m_value.array->max_size();
+            }
+
+            case value_t::object:
+            {
+                // delegate call to object_t::max_size()
+                return m_data.m_value.object->max_size();
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                // all other types have max_size() == size()
+                return size();
+            }
+        }
+    }
+
+    /// @}
+
+    ///////////////
+    // modifiers //
+    ///////////////
+
+    /// @name modifiers
+    /// @{
+
+    /// @brief clears the contents
+    /// @sa https://json.nlohmann.me/api/basic_json/clear/
+    void clear() noexcept
+    {
+        switch (m_data.m_type)
+        {
+            case value_t::number_integer:
+            {
+                m_data.m_value.number_integer = 0;
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                m_data.m_value.number_unsigned = 0;
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                m_data.m_value.number_float = 0.0;
+                break;
+            }
+
+            case value_t::boolean:
+            {
+                m_data.m_value.boolean = false;
+                break;
+            }
+
+            case value_t::string:
+            {
+                m_data.m_value.string->clear();
+                break;
+            }
+
+            case value_t::binary:
+            {
+                m_data.m_value.binary->clear();
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_data.m_value.array->clear();
+                break;
+            }
+
+            case value_t::object:
+            {
+                m_data.m_value.object->clear();
+                break;
+            }
+
+            case value_t::null:
+            case value_t::discarded:
+            default:
+                break;
+        }
+    }
+
+    /// @brief add an object to an array
+    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
+    void push_back(basic_json&& val)
+    {
+        // push_back only works for null objects or arrays
+        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_array())))
+        {
+            JSON_THROW(type_error::create(308, detail::concat("cannot use push_back() with ", type_name()), this));
+        }
+
+        // transform null object into an array
+        if (is_null())
+        {
+            m_data.m_type = value_t::array;
+            m_data.m_value = value_t::array;
+            assert_invariant();
+        }
+
+        // add element to array (move semantics)
+        const auto old_capacity = m_data.m_value.array->capacity();
+        m_data.m_value.array->push_back(std::move(val));
+        set_parent(m_data.m_value.array->back(), old_capacity);
+        // if val is moved from, basic_json move constructor marks it null, so we do not call the destructor
+    }
+
+    /// @brief add an object to an array
+    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
+    reference operator+=(basic_json&& val)
+    {
+        push_back(std::move(val));
+        return *this;
+    }
+
+    /// @brief add an object to an array
+    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
+    void push_back(const basic_json& val)
+    {
+        // push_back only works for null objects or arrays
+        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_array())))
+        {
+            JSON_THROW(type_error::create(308, detail::concat("cannot use push_back() with ", type_name()), this));
+        }
+
+        // transform null object into an array
+        if (is_null())
+        {
+            m_data.m_type = value_t::array;
+            m_data.m_value = value_t::array;
+            assert_invariant();
+        }
+
+        // add element to array
+        const auto old_capacity = m_data.m_value.array->capacity();
+        m_data.m_value.array->push_back(val);
+        set_parent(m_data.m_value.array->back(), old_capacity);
+    }
+
+    /// @brief add an object to an array
+    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
+    reference operator+=(const basic_json& val)
+    {
+        push_back(val);
+        return *this;
+    }
+
+    /// @brief add an object to an object
+    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
+    void push_back(const typename object_t::value_type& val)
+    {
+        // push_back only works for null objects or objects
+        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_object())))
+        {
+            JSON_THROW(type_error::create(308, detail::concat("cannot use push_back() with ", type_name()), this));
+        }
+
+        // transform null object into an object
+        if (is_null())
+        {
+            m_data.m_type = value_t::object;
+            m_data.m_value = value_t::object;
+            assert_invariant();
+        }
+
+        // add element to object
+        auto res = m_data.m_value.object->insert(val);
+        set_parent(res.first->second);
+    }
+
+    /// @brief add an object to an object
+    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
+    reference operator+=(const typename object_t::value_type& val)
+    {
+        push_back(val);
+        return *this;
+    }
+
+    /// @brief add an object to an object
+    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
+    void push_back(initializer_list_t init)
+    {
+        if (is_object() && init.size() == 2 && (*init.begin())->is_string())
+        {
+            basic_json&& key = init.begin()->moved_or_copied();
+            push_back(typename object_t::value_type(
+                          std::move(key.get_ref<string_t&>()), (init.begin() + 1)->moved_or_copied()));
+        }
+        else
+        {
+            push_back(basic_json(init));
+        }
+    }
+
+    /// @brief add an object to an object
+    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
+    reference operator+=(initializer_list_t init)
+    {
+        push_back(init);
+        return *this;
+    }
+
+    /// @brief add an object to an array
+    /// @sa https://json.nlohmann.me/api/basic_json/emplace_back/
+    template<class... Args>
+    reference emplace_back(Args&& ... args)
+    {
+        // emplace_back only works for null objects or arrays
+        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_array())))
+        {
+            JSON_THROW(type_error::create(311, detail::concat("cannot use emplace_back() with ", type_name()), this));
+        }
+
+        // transform null object into an array
+        if (is_null())
+        {
+            m_data.m_type = value_t::array;
+            m_data.m_value = value_t::array;
+            assert_invariant();
+        }
+
+        // add element to array (perfect forwarding)
+        const auto old_capacity = m_data.m_value.array->capacity();
+        m_data.m_value.array->emplace_back(std::forward<Args>(args)...);
+        return set_parent(m_data.m_value.array->back(), old_capacity);
+    }
+
+    /// @brief add an object to an object if key does not exist
+    /// @sa https://json.nlohmann.me/api/basic_json/emplace/
+    template<class... Args>
+    std::pair<iterator, bool> emplace(Args&& ... args)
+    {
+        // emplace only works for null objects or arrays
+        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_object())))
+        {
+            JSON_THROW(type_error::create(311, detail::concat("cannot use emplace() with ", type_name()), this));
+        }
+
+        // transform null object into an object
+        if (is_null())
+        {
+            m_data.m_type = value_t::object;
+            m_data.m_value = value_t::object;
+            assert_invariant();
+        }
+
+        // add element to array (perfect forwarding)
+        auto res = m_data.m_value.object->emplace(std::forward<Args>(args)...);
+        set_parent(res.first->second);
+
+        // create result iterator and set iterator to the result of emplace
+        auto it = begin();
+        it.m_it.object_iterator = res.first;
+
+        // return pair of iterator and boolean
+        return {it, res.second};
+    }
+
+    /// Helper for insertion of an iterator
+    /// @note: This uses std::distance to support GCC 4.8,
+    ///        see https://github.com/nlohmann/json/pull/1257
+    template<typename... Args>
+    iterator insert_iterator(const_iterator pos, Args&& ... args)
+    {
+        iterator result(this);
+        JSON_ASSERT(m_data.m_value.array != nullptr);
+
+        auto insert_pos = std::distance(m_data.m_value.array->begin(), pos.m_it.array_iterator);
+        m_data.m_value.array->insert(pos.m_it.array_iterator, std::forward<Args>(args)...);
+        result.m_it.array_iterator = m_data.m_value.array->begin() + insert_pos;
+
+        // This could have been written as:
+        // result.m_it.array_iterator = m_data.m_value.array->insert(pos.m_it.array_iterator, cnt, val);
+        // but the return value of insert is missing in GCC 4.8, so it is written this way instead.
+
+        set_parents();
+        return result;
+    }
+
+    /// @brief inserts element into array
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    iterator insert(const_iterator pos, const basic_json& val)
+    {
+        // insert only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            // check if iterator pos fits to this JSON value
+            if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
+            {
+                JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", this));
+            }
+
+            // insert to array and return iterator
+            return insert_iterator(pos, val);
+        }
+
+        JSON_THROW(type_error::create(309, detail::concat("cannot use insert() with ", type_name()), this));
+    }
+
+    /// @brief inserts element into array
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    iterator insert(const_iterator pos, basic_json&& val)
+    {
+        return insert(pos, val);
+    }
+
+    /// @brief inserts copies of element into array
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    iterator insert(const_iterator pos, size_type cnt, const basic_json& val)
+    {
+        // insert only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            // check if iterator pos fits to this JSON value
+            if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
+            {
+                JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", this));
+            }
+
+            // insert to array and return iterator
+            return insert_iterator(pos, cnt, val);
+        }
+
+        JSON_THROW(type_error::create(309, detail::concat("cannot use insert() with ", type_name()), this));
+    }
+
+    /// @brief inserts range of elements into array
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    iterator insert(const_iterator pos, const_iterator first, const_iterator last)
+    {
+        // insert only works for arrays
+        if (JSON_HEDLEY_UNLIKELY(!is_array()))
+        {
+            JSON_THROW(type_error::create(309, detail::concat("cannot use insert() with ", type_name()), this));
+        }
+
+        // check if iterator pos fits to this JSON value
+        if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
+        {
+            JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", this));
+        }
+
+        // check if range iterators belong to the same JSON object
+        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(210, "iterators do not fit", this));
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(first.m_object == this))
+        {
+            JSON_THROW(invalid_iterator::create(211, "passed iterators may not belong to container", this));
+        }
+
+        // insert to array and return iterator
+        return insert_iterator(pos, first.m_it.array_iterator, last.m_it.array_iterator);
+    }
+
+    /// @brief inserts elements from initializer list into array
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    iterator insert(const_iterator pos, initializer_list_t ilist)
+    {
+        // insert only works for arrays
+        if (JSON_HEDLEY_UNLIKELY(!is_array()))
+        {
+            JSON_THROW(type_error::create(309, detail::concat("cannot use insert() with ", type_name()), this));
+        }
+
+        // check if iterator pos fits to this JSON value
+        if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
+        {
+            JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", this));
+        }
+
+        // insert to array and return iterator
+        return insert_iterator(pos, ilist.begin(), ilist.end());
+    }
+
+    /// @brief inserts range of elements into object
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    void insert(const_iterator first, const_iterator last)
+    {
+        // insert only works for objects
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(309, detail::concat("cannot use insert() with ", type_name()), this));
+        }
+
+        // check if range iterators belong to the same JSON object
+        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(210, "iterators do not fit", this));
+        }
+
+        // passed iterators must belong to objects
+        if (JSON_HEDLEY_UNLIKELY(!first.m_object->is_object()))
+        {
+            JSON_THROW(invalid_iterator::create(202, "iterators first and last must point to objects", this));
+        }
+
+        m_data.m_value.object->insert(first.m_it.object_iterator, last.m_it.object_iterator);
+    }
+
+    /// @brief updates a JSON object from another object, overwriting existing keys
+    /// @sa https://json.nlohmann.me/api/basic_json/update/
+    void update(const_reference j, bool merge_objects = false)
+    {
+        update(j.begin(), j.end(), merge_objects);
+    }
+
+    /// @brief updates a JSON object from another object, overwriting existing keys
+    /// @sa https://json.nlohmann.me/api/basic_json/update/
+    void update(const_iterator first, const_iterator last, bool merge_objects = false)
+    {
+        // implicitly convert null value to an empty object
+        if (is_null())
+        {
+            m_data.m_type = value_t::object;
+            m_data.m_value.object = create<object_t>();
+            assert_invariant();
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(312, detail::concat("cannot use update() with ", type_name()), this));
+        }
+
+        // check if range iterators belong to the same JSON object
+        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(210, "iterators do not fit", this));
+        }
+
+        // passed iterators must belong to objects
+        if (JSON_HEDLEY_UNLIKELY(!first.m_object->is_object()))
+        {
+            JSON_THROW(type_error::create(312, detail::concat("cannot use update() with ", first.m_object->type_name()), first.m_object));
+        }
+
+        for (auto it = first; it != last; ++it)
+        {
+            if (merge_objects && it.value().is_object())
+            {
+                auto it2 = m_data.m_value.object->find(it.key());
+                if (it2 != m_data.m_value.object->end())
+                {
+                    it2->second.update(it.value(), true);
+                    continue;
+                }
+            }
+            m_data.m_value.object->operator[](it.key()) = it.value();
+#if JSON_DIAGNOSTICS
+            m_data.m_value.object->operator[](it.key()).m_parent = this;
+#endif
+        }
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(reference other) noexcept (
+        std::is_nothrow_move_constructible<value_t>::value&&
+        std::is_nothrow_move_assignable<value_t>::value&&
+        std::is_nothrow_move_constructible<json_value>::value&& // NOLINT(cppcoreguidelines-noexcept-swap,performance-noexcept-swap)
+        std::is_nothrow_move_assignable<json_value>::value
+    )
+    {
+        std::swap(m_data.m_type, other.m_data.m_type);
+        std::swap(m_data.m_value, other.m_data.m_value);
+
+        set_parents();
+        other.set_parents();
+        assert_invariant();
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    friend void swap(reference left, reference right) noexcept (
+        std::is_nothrow_move_constructible<value_t>::value&&
+        std::is_nothrow_move_assignable<value_t>::value&&
+        std::is_nothrow_move_constructible<json_value>::value&& // NOLINT(cppcoreguidelines-noexcept-swap,performance-noexcept-swap)
+        std::is_nothrow_move_assignable<json_value>::value
+    )
+    {
+        left.swap(right);
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(array_t& other) // NOLINT(bugprone-exception-escape,cppcoreguidelines-noexcept-swap,performance-noexcept-swap)
+    {
+        // swap only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            using std::swap;
+            swap(*(m_data.m_value.array), other);
+        }
+        else
+        {
+            JSON_THROW(type_error::create(310, detail::concat("cannot use swap(array_t&) with ", type_name()), this));
+        }
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(object_t& other) // NOLINT(bugprone-exception-escape,cppcoreguidelines-noexcept-swap,performance-noexcept-swap)
+    {
+        // swap only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            using std::swap;
+            swap(*(m_data.m_value.object), other);
+        }
+        else
+        {
+            JSON_THROW(type_error::create(310, detail::concat("cannot use swap(object_t&) with ", type_name()), this));
+        }
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(string_t& other) // NOLINT(bugprone-exception-escape,cppcoreguidelines-noexcept-swap,performance-noexcept-swap)
+    {
+        // swap only works for strings
+        if (JSON_HEDLEY_LIKELY(is_string()))
+        {
+            using std::swap;
+            swap(*(m_data.m_value.string), other);
+        }
+        else
+        {
+            JSON_THROW(type_error::create(310, detail::concat("cannot use swap(string_t&) with ", type_name()), this));
+        }
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(binary_t& other) // NOLINT(bugprone-exception-escape,cppcoreguidelines-noexcept-swap,performance-noexcept-swap)
+    {
+        // swap only works for strings
+        if (JSON_HEDLEY_LIKELY(is_binary()))
+        {
+            using std::swap;
+            swap(*(m_data.m_value.binary), other);
+        }
+        else
+        {
+            JSON_THROW(type_error::create(310, detail::concat("cannot use swap(binary_t&) with ", type_name()), this));
+        }
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(typename binary_t::container_type& other) // NOLINT(bugprone-exception-escape)
+    {
+        // swap only works for strings
+        if (JSON_HEDLEY_LIKELY(is_binary()))
+        {
+            using std::swap;
+            swap(*(m_data.m_value.binary), other);
+        }
+        else
+        {
+            JSON_THROW(type_error::create(310, detail::concat("cannot use swap(binary_t::container_type&) with ", type_name()), this));
+        }
+    }
+
+    /// @}
+
+    //////////////////////////////////////////
+    // lexicographical comparison operators //
+    //////////////////////////////////////////
+
+    /// @name lexicographical comparison operators
+    /// @{
+
+    // note parentheses around operands are necessary; see
+    // https://github.com/nlohmann/json/issues/1530
+#define JSON_IMPLEMENT_OPERATOR(op, null_result, unordered_result, default_result)                       \
+    const auto lhs_type = lhs.type();                                                                    \
+    const auto rhs_type = rhs.type();                                                                    \
+    \
+    if (lhs_type == rhs_type) /* NOLINT(readability/braces) */                                           \
+    {                                                                                                    \
+        switch (lhs_type)                                                                                \
+        {                                                                                                \
+            case value_t::array:                                                                         \
+                return (*lhs.m_data.m_value.array) op (*rhs.m_data.m_value.array);                                     \
+                \
+            case value_t::object:                                                                        \
+                return (*lhs.m_data.m_value.object) op (*rhs.m_data.m_value.object);                                   \
+                \
+            case value_t::null:                                                                          \
+                return (null_result);                                                                    \
+                \
+            case value_t::string:                                                                        \
+                return (*lhs.m_data.m_value.string) op (*rhs.m_data.m_value.string);                                   \
+                \
+            case value_t::boolean:                                                                       \
+                return (lhs.m_data.m_value.boolean) op (rhs.m_data.m_value.boolean);                                   \
+                \
+            case value_t::number_integer:                                                                \
+                return (lhs.m_data.m_value.number_integer) op (rhs.m_data.m_value.number_integer);                     \
+                \
+            case value_t::number_unsigned:                                                               \
+                return (lhs.m_data.m_value.number_unsigned) op (rhs.m_data.m_value.number_unsigned);                   \
+                \
+            case value_t::number_float:                                                                  \
+                return (lhs.m_data.m_value.number_float) op (rhs.m_data.m_value.number_float);                         \
+                \
+            case value_t::binary:                                                                        \
+                return (*lhs.m_data.m_value.binary) op (*rhs.m_data.m_value.binary);                                   \
+                \
+            case value_t::discarded:                                                                     \
+            default:                                                                                     \
+                return (unordered_result);                                                               \
+        }                                                                                                \
+    }                                                                                                    \
+    else if (lhs_type == value_t::number_integer && rhs_type == value_t::number_float)                   \
+    {                                                                                                    \
+        return static_cast<number_float_t>(lhs.m_data.m_value.number_integer) op rhs.m_data.m_value.number_float;      \
+    }                                                                                                    \
+    else if (lhs_type == value_t::number_float && rhs_type == value_t::number_integer)                   \
+    {                                                                                                    \
+        return lhs.m_data.m_value.number_float op static_cast<number_float_t>(rhs.m_data.m_value.number_integer);      \
+    }                                                                                                    \
+    else if (lhs_type == value_t::number_unsigned && rhs_type == value_t::number_float)                  \
+    {                                                                                                    \
+        return static_cast<number_float_t>(lhs.m_data.m_value.number_unsigned) op rhs.m_data.m_value.number_float;     \
+    }                                                                                                    \
+    else if (lhs_type == value_t::number_float && rhs_type == value_t::number_unsigned)                  \
+    {                                                                                                    \
+        return lhs.m_data.m_value.number_float op static_cast<number_float_t>(rhs.m_data.m_value.number_unsigned);     \
+    }                                                                                                    \
+    else if (lhs_type == value_t::number_unsigned && rhs_type == value_t::number_integer)                \
+    {                                                                                                    \
+        return static_cast<number_integer_t>(lhs.m_data.m_value.number_unsigned) op rhs.m_data.m_value.number_integer; \
+    }                                                                                                    \
+    else if (lhs_type == value_t::number_integer && rhs_type == value_t::number_unsigned)                \
+    {                                                                                                    \
+        return lhs.m_data.m_value.number_integer op static_cast<number_integer_t>(rhs.m_data.m_value.number_unsigned); \
+    }                                                                                                    \
+    else if(compares_unordered(lhs, rhs))\
+    {\
+        return (unordered_result);\
+    }\
+    \
+    return (default_result);
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    // returns true if:
+    // - any operand is NaN and the other operand is of number type
+    // - any operand is discarded
+    // in legacy mode, discarded values are considered ordered if
+    // an operation is computed as an odd number of inverses of others
+    static bool compares_unordered(const_reference lhs, const_reference rhs, bool inverse = false) noexcept
+    {
+        if ((lhs.is_number_float() && std::isnan(lhs.m_data.m_value.number_float) && rhs.is_number())
+                || (rhs.is_number_float() && std::isnan(rhs.m_data.m_value.number_float) && lhs.is_number()))
+        {
+            return true;
+        }
+#if JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
+        return (lhs.is_discarded() || rhs.is_discarded()) && !inverse;
+#else
+        static_cast<void>(inverse);
+        return lhs.is_discarded() || rhs.is_discarded();
+#endif
+    }
+
+  private:
+    bool compares_unordered(const_reference rhs, bool inverse = false) const noexcept
+    {
+        return compares_unordered(*this, rhs, inverse);
+    }
+
+  public:
+#if JSON_HAS_THREE_WAY_COMPARISON
+    /// @brief comparison: equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
+    bool operator==(const_reference rhs) const noexcept
+    {
+#ifdef __GNUC__
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wfloat-equal"
+#endif
+        const_reference lhs = *this;
+        JSON_IMPLEMENT_OPERATOR( ==, true, false, false)
+#ifdef __GNUC__
+#pragma GCC diagnostic pop
+#endif
+    }
+
+    /// @brief comparison: equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
+    template<typename ScalarType>
+    requires std::is_scalar_v<ScalarType>
+    bool operator==(ScalarType rhs) const noexcept
+    {
+        return *this == basic_json(rhs);
+    }
+
+    /// @brief comparison: not equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ne/
+    bool operator!=(const_reference rhs) const noexcept
+    {
+        if (compares_unordered(rhs, true))
+        {
+            return false;
+        }
+        return !operator==(rhs);
+    }
+
+    /// @brief comparison: 3-way
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_spaceship/
+    std::partial_ordering operator<=>(const_reference rhs) const noexcept // *NOPAD*
+    {
+        const_reference lhs = *this;
+        // default_result is used if we cannot compare values. In that case,
+        // we compare types.
+        JSON_IMPLEMENT_OPERATOR(<=>, // *NOPAD*
+                                std::partial_ordering::equivalent,
+                                std::partial_ordering::unordered,
+                                lhs_type <=> rhs_type) // *NOPAD*
+    }
+
+    /// @brief comparison: 3-way
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_spaceship/
+    template<typename ScalarType>
+    requires std::is_scalar_v<ScalarType>
+    std::partial_ordering operator<=>(ScalarType rhs) const noexcept // *NOPAD*
+    {
+        return *this <=> basic_json(rhs); // *NOPAD*
+    }
+
+#if JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
+    // all operators that are computed as an odd number of inverses of others
+    // need to be overloaded to emulate the legacy comparison behavior
+
+    /// @brief comparison: less than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, undef JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON)
+    bool operator<=(const_reference rhs) const noexcept
+    {
+        if (compares_unordered(rhs, true))
+        {
+            return false;
+        }
+        return !(rhs < *this);
+    }
+
+    /// @brief comparison: less than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
+    template<typename ScalarType>
+    requires std::is_scalar_v<ScalarType>
+    bool operator<=(ScalarType rhs) const noexcept
+    {
+        return *this <= basic_json(rhs);
+    }
+
+    /// @brief comparison: greater than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, undef JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON)
+    bool operator>=(const_reference rhs) const noexcept
+    {
+        if (compares_unordered(rhs, true))
+        {
+            return false;
+        }
+        return !(*this < rhs);
+    }
+
+    /// @brief comparison: greater than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
+    template<typename ScalarType>
+    requires std::is_scalar_v<ScalarType>
+    bool operator>=(ScalarType rhs) const noexcept
+    {
+        return *this >= basic_json(rhs);
+    }
+#endif
+#else
+    /// @brief comparison: equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
+    friend bool operator==(const_reference lhs, const_reference rhs) noexcept
+    {
+#ifdef __GNUC__
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wfloat-equal"
+#endif
+        JSON_IMPLEMENT_OPERATOR( ==, true, false, false)
+#ifdef __GNUC__
+#pragma GCC diagnostic pop
+#endif
+    }
+
+    /// @brief comparison: equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator==(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs == basic_json(rhs);
+    }
+
+    /// @brief comparison: equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator==(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) == rhs;
+    }
+
+    /// @brief comparison: not equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ne/
+    friend bool operator!=(const_reference lhs, const_reference rhs) noexcept
+    {
+        if (compares_unordered(lhs, rhs, true))
+        {
+            return false;
+        }
+        return !(lhs == rhs);
+    }
+
+    /// @brief comparison: not equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ne/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator!=(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs != basic_json(rhs);
+    }
+
+    /// @brief comparison: not equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ne/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator!=(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) != rhs;
+    }
+
+    /// @brief comparison: less than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_lt/
+    friend bool operator<(const_reference lhs, const_reference rhs) noexcept
+    {
+        // default_result is used if we cannot compare values. In that case,
+        // we compare types. Note we have to call the operator explicitly,
+        // because MSVC has problems otherwise.
+        JSON_IMPLEMENT_OPERATOR( <, false, false, operator<(lhs_type, rhs_type))
+    }
+
+    /// @brief comparison: less than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_lt/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator<(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs < basic_json(rhs);
+    }
+
+    /// @brief comparison: less than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_lt/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator<(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) < rhs;
+    }
+
+    /// @brief comparison: less than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
+    friend bool operator<=(const_reference lhs, const_reference rhs) noexcept
+    {
+        if (compares_unordered(lhs, rhs, true))
+        {
+            return false;
+        }
+        return !(rhs < lhs);
+    }
+
+    /// @brief comparison: less than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator<=(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs <= basic_json(rhs);
+    }
+
+    /// @brief comparison: less than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator<=(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) <= rhs;
+    }
+
+    /// @brief comparison: greater than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_gt/
+    friend bool operator>(const_reference lhs, const_reference rhs) noexcept
+    {
+        // double inverse
+        if (compares_unordered(lhs, rhs))
+        {
+            return false;
+        }
+        return !(lhs <= rhs);
+    }
+
+    /// @brief comparison: greater than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_gt/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator>(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs > basic_json(rhs);
+    }
+
+    /// @brief comparison: greater than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_gt/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator>(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) > rhs;
+    }
+
+    /// @brief comparison: greater than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
+    friend bool operator>=(const_reference lhs, const_reference rhs) noexcept
+    {
+        if (compares_unordered(lhs, rhs, true))
+        {
+            return false;
+        }
+        return !(lhs < rhs);
+    }
+
+    /// @brief comparison: greater than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator>=(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs >= basic_json(rhs);
+    }
+
+    /// @brief comparison: greater than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator>=(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) >= rhs;
+    }
+#endif
+
+#undef JSON_IMPLEMENT_OPERATOR
+
+    /// @}
+
+    ///////////////////
+    // serialization //
+    ///////////////////
+
+    /// @name serialization
+    /// @{
+#ifndef JSON_NO_IO
+    /// @brief serialize to stream
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ltlt/
+    friend std::ostream& operator<<(std::ostream& o, const basic_json& j)
+    {
+        // read width member and use it as indentation parameter if nonzero
+        const bool pretty_print = o.width() > 0;
+        const auto indentation = pretty_print ? o.width() : 0;
+
+        // reset width to 0 for subsequent calls to this stream
+        o.width(0);
+
+        // do the actual serialization
+        serializer s(detail::output_adapter<char>(o), o.fill());
+        s.dump(j, pretty_print, false, static_cast<unsigned int>(indentation));
+        return o;
+    }
+
+    /// @brief serialize to stream
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ltlt/
+    /// @deprecated This function is deprecated since 3.0.0 and will be removed in
+    ///             version 4.0.0 of the library. Please use
+    ///             operator<<(std::ostream&, const basic_json&) instead; that is,
+    ///             replace calls like `j >> o;` with `o << j;`.
+    JSON_HEDLEY_DEPRECATED_FOR(3.0.0, operator<<(std::ostream&, const basic_json&))
+    friend std::ostream& operator>>(const basic_json& j, std::ostream& o)
+    {
+        return o << j;
+    }
+#endif  // JSON_NO_IO
+    /// @}
+
+    /////////////////////
+    // deserialization //
+    /////////////////////
+
+    /// @name deserialization
+    /// @{
+
+    /// @brief deserialize from a compatible input
+    /// @sa https://json.nlohmann.me/api/basic_json/parse/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json parse(InputType&& i,
+                            const parser_callback_t cb = nullptr,
+                            const bool allow_exceptions = true,
+                            const bool ignore_comments = false)
+    {
+        basic_json result;
+        parser(detail::input_adapter(std::forward<InputType>(i)), cb, allow_exceptions, ignore_comments).parse(true, result);
+        return result;
+    }
+
+    /// @brief deserialize from a pair of character iterators
+    /// @sa https://json.nlohmann.me/api/basic_json/parse/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json parse(IteratorType first,
+                            IteratorType last,
+                            const parser_callback_t cb = nullptr,
+                            const bool allow_exceptions = true,
+                            const bool ignore_comments = false)
+    {
+        basic_json result;
+        parser(detail::input_adapter(std::move(first), std::move(last)), cb, allow_exceptions, ignore_comments).parse(true, result);
+        return result;
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, parse(ptr, ptr + len))
+    static basic_json parse(detail::span_input_adapter&& i,
+                            const parser_callback_t cb = nullptr,
+                            const bool allow_exceptions = true,
+                            const bool ignore_comments = false)
+    {
+        basic_json result;
+        parser(i.get(), cb, allow_exceptions, ignore_comments).parse(true, result);
+        return result;
+    }
+
+    /// @brief check if the input is valid JSON
+    /// @sa https://json.nlohmann.me/api/basic_json/accept/
+    template<typename InputType>
+    static bool accept(InputType&& i,
+                       const bool ignore_comments = false)
+    {
+        return parser(detail::input_adapter(std::forward<InputType>(i)), nullptr, false, ignore_comments).accept(true);
+    }
+
+    /// @brief check if the input is valid JSON
+    /// @sa https://json.nlohmann.me/api/basic_json/accept/
+    template<typename IteratorType>
+    static bool accept(IteratorType first, IteratorType last,
+                       const bool ignore_comments = false)
+    {
+        return parser(detail::input_adapter(std::move(first), std::move(last)), nullptr, false, ignore_comments).accept(true);
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, accept(ptr, ptr + len))
+    static bool accept(detail::span_input_adapter&& i,
+                       const bool ignore_comments = false)
+    {
+        return parser(i.get(), nullptr, false, ignore_comments).accept(true);
+    }
+
+    /// @brief generate SAX events
+    /// @sa https://json.nlohmann.me/api/basic_json/sax_parse/
+    template <typename InputType, typename SAX>
+    JSON_HEDLEY_NON_NULL(2)
+    static bool sax_parse(InputType&& i, SAX* sax,
+                          input_format_t format = input_format_t::json,
+                          const bool strict = true,
+                          const bool ignore_comments = false)
+    {
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        return format == input_format_t::json
+               ? parser(std::move(ia), nullptr, true, ignore_comments).sax_parse(sax, strict)
+               : detail::binary_reader<basic_json, decltype(ia), SAX>(std::move(ia), format).sax_parse(format, sax, strict);
+    }
+
+    /// @brief generate SAX events
+    /// @sa https://json.nlohmann.me/api/basic_json/sax_parse/
+    template<class IteratorType, class SAX>
+    JSON_HEDLEY_NON_NULL(3)
+    static bool sax_parse(IteratorType first, IteratorType last, SAX* sax,
+                          input_format_t format = input_format_t::json,
+                          const bool strict = true,
+                          const bool ignore_comments = false)
+    {
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        return format == input_format_t::json
+               ? parser(std::move(ia), nullptr, true, ignore_comments).sax_parse(sax, strict)
+               : detail::binary_reader<basic_json, decltype(ia), SAX>(std::move(ia), format).sax_parse(format, sax, strict);
+    }
+
+    /// @brief generate SAX events
+    /// @sa https://json.nlohmann.me/api/basic_json/sax_parse/
+    /// @deprecated This function is deprecated since 3.8.0 and will be removed in
+    ///             version 4.0.0 of the library. Please use
+    ///             sax_parse(ptr, ptr + len) instead.
+    template <typename SAX>
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, sax_parse(ptr, ptr + len, ...))
+    JSON_HEDLEY_NON_NULL(2)
+    static bool sax_parse(detail::span_input_adapter&& i, SAX* sax,
+                          input_format_t format = input_format_t::json,
+                          const bool strict = true,
+                          const bool ignore_comments = false)
+    {
+        auto ia = i.get();
+        return format == input_format_t::json
+               // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+               ? parser(std::move(ia), nullptr, true, ignore_comments).sax_parse(sax, strict)
+               // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+               : detail::binary_reader<basic_json, decltype(ia), SAX>(std::move(ia), format).sax_parse(format, sax, strict);
+    }
+#ifndef JSON_NO_IO
+    /// @brief deserialize from stream
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_gtgt/
+    /// @deprecated This stream operator is deprecated since 3.0.0 and will be removed in
+    ///             version 4.0.0 of the library. Please use
+    ///             operator>>(std::istream&, basic_json&) instead; that is,
+    ///             replace calls like `j << i;` with `i >> j;`.
+    JSON_HEDLEY_DEPRECATED_FOR(3.0.0, operator>>(std::istream&, basic_json&))
+    friend std::istream& operator<<(basic_json& j, std::istream& i)
+    {
+        return operator>>(i, j);
+    }
+
+    /// @brief deserialize from stream
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_gtgt/
+    friend std::istream& operator>>(std::istream& i, basic_json& j)
+    {
+        parser(detail::input_adapter(i)).parse(false, j);
+        return i;
+    }
+#endif  // JSON_NO_IO
+    /// @}
+
+    ///////////////////////////
+    // convenience functions //
+    ///////////////////////////
+
+    /// @brief return the type as string
+    /// @sa https://json.nlohmann.me/api/basic_json/type_name/
+    JSON_HEDLEY_RETURNS_NON_NULL
+    const char* type_name() const noexcept
+    {
+        switch (m_data.m_type)
+        {
+            case value_t::null:
+                return "null";
+            case value_t::object:
+                return "object";
+            case value_t::array:
+                return "array";
+            case value_t::string:
+                return "string";
+            case value_t::boolean:
+                return "boolean";
+            case value_t::binary:
+                return "binary";
+            case value_t::discarded:
+                return "discarded";
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            default:
+                return "number";
+        }
+    }
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    //////////////////////
+    // member variables //
+    //////////////////////
+
+    struct data
+    {
+        /// the type of the current element
+        value_t m_type = value_t::null;
+
+        /// the value of the current element
+        json_value m_value = {};
+
+        data(const value_t v)
+            : m_type(v), m_value(v)
+        {
+        }
+
+        data(size_type cnt, const basic_json& val)
+            : m_type(value_t::array)
+        {
+            m_value.array = create<array_t>(cnt, val);
+        }
+
+        data() noexcept = default;
+        data(data&&) noexcept = default;
+        data(const data&) noexcept = delete;
+        data& operator=(data&&) noexcept = delete;
+        data& operator=(const data&) noexcept = delete;
+
+        ~data() noexcept
+        {
+            m_value.destroy(m_type);
+        }
+    };
+
+    data m_data = {};
+
+#if JSON_DIAGNOSTICS
+    /// a pointer to a parent value (for debugging purposes)
+    basic_json* m_parent = nullptr;
+#endif
+
+    //////////////////////////////////////////
+    // binary serialization/deserialization //
+    //////////////////////////////////////////
+
+    /// @name binary serialization/deserialization support
+    /// @{
+
+  public:
+    /// @brief create a CBOR serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_cbor/
+    static std::vector<std::uint8_t> to_cbor(const basic_json& j)
+    {
+        std::vector<std::uint8_t> result;
+        to_cbor(j, result);
+        return result;
+    }
+
+    /// @brief create a CBOR serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_cbor/
+    static void to_cbor(const basic_json& j, detail::output_adapter<std::uint8_t> o)
+    {
+        binary_writer<std::uint8_t>(o).write_cbor(j);
+    }
+
+    /// @brief create a CBOR serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_cbor/
+    static void to_cbor(const basic_json& j, detail::output_adapter<char> o)
+    {
+        binary_writer<char>(o).write_cbor(j);
+    }
+
+    /// @brief create a MessagePack serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_msgpack/
+    static std::vector<std::uint8_t> to_msgpack(const basic_json& j)
+    {
+        std::vector<std::uint8_t> result;
+        to_msgpack(j, result);
+        return result;
+    }
+
+    /// @brief create a MessagePack serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_msgpack/
+    static void to_msgpack(const basic_json& j, detail::output_adapter<std::uint8_t> o)
+    {
+        binary_writer<std::uint8_t>(o).write_msgpack(j);
+    }
+
+    /// @brief create a MessagePack serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_msgpack/
+    static void to_msgpack(const basic_json& j, detail::output_adapter<char> o)
+    {
+        binary_writer<char>(o).write_msgpack(j);
+    }
+
+    /// @brief create a UBJSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_ubjson/
+    static std::vector<std::uint8_t> to_ubjson(const basic_json& j,
+            const bool use_size = false,
+            const bool use_type = false)
+    {
+        std::vector<std::uint8_t> result;
+        to_ubjson(j, result, use_size, use_type);
+        return result;
+    }
+
+    /// @brief create a UBJSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_ubjson/
+    static void to_ubjson(const basic_json& j, detail::output_adapter<std::uint8_t> o,
+                          const bool use_size = false, const bool use_type = false)
+    {
+        binary_writer<std::uint8_t>(o).write_ubjson(j, use_size, use_type);
+    }
+
+    /// @brief create a UBJSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_ubjson/
+    static void to_ubjson(const basic_json& j, detail::output_adapter<char> o,
+                          const bool use_size = false, const bool use_type = false)
+    {
+        binary_writer<char>(o).write_ubjson(j, use_size, use_type);
+    }
+
+    /// @brief create a BJData serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_bjdata/
+    static std::vector<std::uint8_t> to_bjdata(const basic_json& j,
+            const bool use_size = false,
+            const bool use_type = false)
+    {
+        std::vector<std::uint8_t> result;
+        to_bjdata(j, result, use_size, use_type);
+        return result;
+    }
+
+    /// @brief create a BJData serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_bjdata/
+    static void to_bjdata(const basic_json& j, detail::output_adapter<std::uint8_t> o,
+                          const bool use_size = false, const bool use_type = false)
+    {
+        binary_writer<std::uint8_t>(o).write_ubjson(j, use_size, use_type, true, true);
+    }
+
+    /// @brief create a BJData serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_bjdata/
+    static void to_bjdata(const basic_json& j, detail::output_adapter<char> o,
+                          const bool use_size = false, const bool use_type = false)
+    {
+        binary_writer<char>(o).write_ubjson(j, use_size, use_type, true, true);
+    }
+
+    /// @brief create a BSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_bson/
+    static std::vector<std::uint8_t> to_bson(const basic_json& j)
+    {
+        std::vector<std::uint8_t> result;
+        to_bson(j, result);
+        return result;
+    }
+
+    /// @brief create a BSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_bson/
+    static void to_bson(const basic_json& j, detail::output_adapter<std::uint8_t> o)
+    {
+        binary_writer<std::uint8_t>(o).write_bson(j);
+    }
+
+    /// @brief create a BSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_bson/
+    static void to_bson(const basic_json& j, detail::output_adapter<char> o)
+    {
+        binary_writer<char>(o).write_bson(j);
+    }
+
+    /// @brief create a JSON value from an input in CBOR format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_cbor/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_cbor(InputType&& i,
+                                const bool strict = true,
+                                const bool allow_exceptions = true,
+                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::cbor).sax_parse(input_format_t::cbor, &sdp, strict, tag_handler);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in CBOR format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_cbor/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_cbor(IteratorType first, IteratorType last,
+                                const bool strict = true,
+                                const bool allow_exceptions = true,
+                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::cbor).sax_parse(input_format_t::cbor, &sdp, strict, tag_handler);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    template<typename T>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_cbor(ptr, ptr + len))
+    static basic_json from_cbor(const T* ptr, std::size_t len,
+                                const bool strict = true,
+                                const bool allow_exceptions = true,
+                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
+    {
+        return from_cbor(ptr, ptr + len, strict, allow_exceptions, tag_handler);
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_cbor(ptr, ptr + len))
+    static basic_json from_cbor(detail::span_input_adapter&& i,
+                                const bool strict = true,
+                                const bool allow_exceptions = true,
+                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = i.get();
+        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::cbor).sax_parse(input_format_t::cbor, &sdp, strict, tag_handler);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in MessagePack format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_msgpack/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_msgpack(InputType&& i,
+                                   const bool strict = true,
+                                   const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::msgpack).sax_parse(input_format_t::msgpack, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in MessagePack format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_msgpack/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_msgpack(IteratorType first, IteratorType last,
+                                   const bool strict = true,
+                                   const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::msgpack).sax_parse(input_format_t::msgpack, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    template<typename T>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_msgpack(ptr, ptr + len))
+    static basic_json from_msgpack(const T* ptr, std::size_t len,
+                                   const bool strict = true,
+                                   const bool allow_exceptions = true)
+    {
+        return from_msgpack(ptr, ptr + len, strict, allow_exceptions);
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_msgpack(ptr, ptr + len))
+    static basic_json from_msgpack(detail::span_input_adapter&& i,
+                                   const bool strict = true,
+                                   const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = i.get();
+        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::msgpack).sax_parse(input_format_t::msgpack, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in UBJSON format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_ubjson/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_ubjson(InputType&& i,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::ubjson).sax_parse(input_format_t::ubjson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in UBJSON format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_ubjson/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_ubjson(IteratorType first, IteratorType last,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::ubjson).sax_parse(input_format_t::ubjson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    template<typename T>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_ubjson(ptr, ptr + len))
+    static basic_json from_ubjson(const T* ptr, std::size_t len,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        return from_ubjson(ptr, ptr + len, strict, allow_exceptions);
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_ubjson(ptr, ptr + len))
+    static basic_json from_ubjson(detail::span_input_adapter&& i,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = i.get();
+        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::ubjson).sax_parse(input_format_t::ubjson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in BJData format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_bjdata/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_bjdata(InputType&& i,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::bjdata).sax_parse(input_format_t::bjdata, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in BJData format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_bjdata/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_bjdata(IteratorType first, IteratorType last,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::bjdata).sax_parse(input_format_t::bjdata, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in BSON format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_bson/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_bson(InputType&& i,
+                                const bool strict = true,
+                                const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::bson).sax_parse(input_format_t::bson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in BSON format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_bson/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_bson(IteratorType first, IteratorType last,
+                                const bool strict = true,
+                                const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::bson).sax_parse(input_format_t::bson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    template<typename T>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_bson(ptr, ptr + len))
+    static basic_json from_bson(const T* ptr, std::size_t len,
+                                const bool strict = true,
+                                const bool allow_exceptions = true)
+    {
+        return from_bson(ptr, ptr + len, strict, allow_exceptions);
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_bson(ptr, ptr + len))
+    static basic_json from_bson(detail::span_input_adapter&& i,
+                                const bool strict = true,
+                                const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = i.get();
+        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::bson).sax_parse(input_format_t::bson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+    /// @}
+
+    //////////////////////////
+    // JSON Pointer support //
+    //////////////////////////
+
+    /// @name JSON Pointer functions
+    /// @{
+
+    /// @brief access specified element via JSON Pointer
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    reference operator[](const json_pointer& ptr)
+    {
+        return ptr.get_unchecked(this);
+    }
+
+    template<typename BasicJsonType, detail::enable_if_t<detail::is_basic_json<BasicJsonType>::value, int> = 0>
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
+    reference operator[](const ::nlohmann::json_pointer<BasicJsonType>& ptr)
+    {
+        return ptr.get_unchecked(this);
+    }
+
+    /// @brief access specified element via JSON Pointer
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    const_reference operator[](const json_pointer& ptr) const
+    {
+        return ptr.get_unchecked(this);
+    }
+
+    template<typename BasicJsonType, detail::enable_if_t<detail::is_basic_json<BasicJsonType>::value, int> = 0>
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
+    const_reference operator[](const ::nlohmann::json_pointer<BasicJsonType>& ptr) const
+    {
+        return ptr.get_unchecked(this);
+    }
+
+    /// @brief access specified element via JSON Pointer
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    reference at(const json_pointer& ptr)
+    {
+        return ptr.get_checked(this);
+    }
+
+    template<typename BasicJsonType, detail::enable_if_t<detail::is_basic_json<BasicJsonType>::value, int> = 0>
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
+    reference at(const ::nlohmann::json_pointer<BasicJsonType>& ptr)
+    {
+        return ptr.get_checked(this);
+    }
+
+    /// @brief access specified element via JSON Pointer
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    const_reference at(const json_pointer& ptr) const
+    {
+        return ptr.get_checked(this);
+    }
+
+    template<typename BasicJsonType, detail::enable_if_t<detail::is_basic_json<BasicJsonType>::value, int> = 0>
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
+    const_reference at(const ::nlohmann::json_pointer<BasicJsonType>& ptr) const
+    {
+        return ptr.get_checked(this);
+    }
+
+    /// @brief return flattened JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/flatten/
+    basic_json flatten() const
+    {
+        basic_json result(value_t::object);
+        json_pointer::flatten("", *this, result);
+        return result;
+    }
+
+    /// @brief unflatten a previously flattened JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/unflatten/
+    basic_json unflatten() const
+    {
+        return json_pointer::unflatten(*this);
+    }
+
+    /// @}
+
+    //////////////////////////
+    // JSON Patch functions //
+    //////////////////////////
+
+    /// @name JSON Patch functions
+    /// @{
+
+    /// @brief applies a JSON patch in-place without copying the object
+    /// @sa https://json.nlohmann.me/api/basic_json/patch/
+    void patch_inplace(const basic_json& json_patch)
+    {
+        basic_json& result = *this;
+        // the valid JSON Patch operations
+        enum class patch_operations {add, remove, replace, move, copy, test, invalid};
+
+        const auto get_op = [](const std::string & op)
+        {
+            if (op == "add")
+            {
+                return patch_operations::add;
+            }
+            if (op == "remove")
+            {
+                return patch_operations::remove;
+            }
+            if (op == "replace")
+            {
+                return patch_operations::replace;
+            }
+            if (op == "move")
+            {
+                return patch_operations::move;
+            }
+            if (op == "copy")
+            {
+                return patch_operations::copy;
+            }
+            if (op == "test")
+            {
+                return patch_operations::test;
+            }
+
+            return patch_operations::invalid;
+        };
+
+        // wrapper for "add" operation; add value at ptr
+        const auto operation_add = [&result](json_pointer & ptr, basic_json val)
+        {
+            // adding to the root of the target document means replacing it
+            if (ptr.empty())
+            {
+                result = val;
+                return;
+            }
+
+            // make sure the top element of the pointer exists
+            json_pointer const top_pointer = ptr.top();
+            if (top_pointer != ptr)
+            {
+                result.at(top_pointer);
+            }
+
+            // get reference to parent of JSON pointer ptr
+            const auto last_path = ptr.back();
+            ptr.pop_back();
+            // parent must exist when performing patch add per RFC6902 specs
+            basic_json& parent = result.at(ptr);
+
+            switch (parent.m_data.m_type)
+            {
+                case value_t::null:
+                case value_t::object:
+                {
+                    // use operator[] to add value
+                    parent[last_path] = val;
+                    break;
+                }
+
+                case value_t::array:
+                {
+                    if (last_path == "-")
+                    {
+                        // special case: append to back
+                        parent.push_back(val);
+                    }
+                    else
+                    {
+                        const auto idx = json_pointer::template array_index<basic_json_t>(last_path);
+                        if (JSON_HEDLEY_UNLIKELY(idx > parent.size()))
+                        {
+                            // avoid undefined behavior
+                            JSON_THROW(out_of_range::create(401, detail::concat("array index ", std::to_string(idx), " is out of range"), &parent));
+                        }
+
+                        // default case: insert add offset
+                        parent.insert(parent.begin() + static_cast<difference_type>(idx), val);
+                    }
+                    break;
+                }
+
+                // if there exists a parent it cannot be primitive
+                case value_t::string: // LCOV_EXCL_LINE
+                case value_t::boolean: // LCOV_EXCL_LINE
+                case value_t::number_integer: // LCOV_EXCL_LINE
+                case value_t::number_unsigned: // LCOV_EXCL_LINE
+                case value_t::number_float: // LCOV_EXCL_LINE
+                case value_t::binary: // LCOV_EXCL_LINE
+                case value_t::discarded: // LCOV_EXCL_LINE
+                default:            // LCOV_EXCL_LINE
+                    JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+            }
+        };
+
+        // wrapper for "remove" operation; remove value at ptr
+        const auto operation_remove = [this, & result](json_pointer & ptr)
+        {
+            // get reference to parent of JSON pointer ptr
+            const auto last_path = ptr.back();
+            ptr.pop_back();
+            basic_json& parent = result.at(ptr);
+
+            // remove child
+            if (parent.is_object())
+            {
+                // perform range check
+                auto it = parent.find(last_path);
+                if (JSON_HEDLEY_LIKELY(it != parent.end()))
+                {
+                    parent.erase(it);
+                }
+                else
+                {
+                    JSON_THROW(out_of_range::create(403, detail::concat("key '", last_path, "' not found"), this));
+                }
+            }
+            else if (parent.is_array())
+            {
+                // note erase performs range check
+                parent.erase(json_pointer::template array_index<basic_json_t>(last_path));
+            }
+        };
+
+        // type check: top level value must be an array
+        if (JSON_HEDLEY_UNLIKELY(!json_patch.is_array()))
+        {
+            JSON_THROW(parse_error::create(104, 0, "JSON patch must be an array of objects", &json_patch));
+        }
+
+        // iterate and apply the operations
+        for (const auto& val : json_patch)
+        {
+            // wrapper to get a value for an operation
+            const auto get_value = [&val](const std::string & op,
+                                          const std::string & member,
+                                          bool string_type) -> basic_json &
+            {
+                // find value
+                auto it = val.m_data.m_value.object->find(member);
+
+                // context-sensitive error message
+                const auto error_msg = (op == "op") ? "operation" : detail::concat("operation '", op, '\''); // NOLINT(bugprone-unused-local-non-trivial-variable)
+
+                // check if desired value is present
+                if (JSON_HEDLEY_UNLIKELY(it == val.m_data.m_value.object->end()))
+                {
+                    // NOLINTNEXTLINE(performance-inefficient-string-concatenation)
+                    JSON_THROW(parse_error::create(105, 0, detail::concat(error_msg, " must have member '", member, "'"), &val));
+                }
+
+                // check if result is of type string
+                if (JSON_HEDLEY_UNLIKELY(string_type && !it->second.is_string()))
+                {
+                    // NOLINTNEXTLINE(performance-inefficient-string-concatenation)
+                    JSON_THROW(parse_error::create(105, 0, detail::concat(error_msg, " must have string member '", member, "'"), &val));
+                }
+
+                // no error: return value
+                return it->second;
+            };
+
+            // type check: every element of the array must be an object
+            if (JSON_HEDLEY_UNLIKELY(!val.is_object()))
+            {
+                JSON_THROW(parse_error::create(104, 0, "JSON patch must be an array of objects", &val));
+            }
+
+            // collect mandatory members
+            const auto op = get_value("op", "op", true).template get<std::string>();
+            const auto path = get_value(op, "path", true).template get<std::string>();
+            json_pointer ptr(path);
+
+            switch (get_op(op))
+            {
+                case patch_operations::add:
+                {
+                    operation_add(ptr, get_value("add", "value", false));
+                    break;
+                }
+
+                case patch_operations::remove:
+                {
+                    operation_remove(ptr);
+                    break;
+                }
+
+                case patch_operations::replace:
+                {
+                    // the "path" location must exist - use at()
+                    result.at(ptr) = get_value("replace", "value", false);
+                    break;
+                }
+
+                case patch_operations::move:
+                {
+                    const auto from_path = get_value("move", "from", true).template get<std::string>();
+                    json_pointer from_ptr(from_path);
+
+                    // the "from" location must exist - use at()
+                    basic_json const v = result.at(from_ptr);
+
+                    // The move operation is functionally identical to a
+                    // "remove" operation on the "from" location, followed
+                    // immediately by an "add" operation at the target
+                    // location with the value that was just removed.
+                    operation_remove(from_ptr);
+                    operation_add(ptr, v);
+                    break;
+                }
+
+                case patch_operations::copy:
+                {
+                    const auto from_path = get_value("copy", "from", true).template get<std::string>();
+                    const json_pointer from_ptr(from_path);
+
+                    // the "from" location must exist - use at()
+                    basic_json const v = result.at(from_ptr);
+
+                    // The copy is functionally identical to an "add"
+                    // operation at the target location using the value
+                    // specified in the "from" member.
+                    operation_add(ptr, v);
+                    break;
+                }
+
+                case patch_operations::test:
+                {
+                    bool success = false;
+                    JSON_TRY
+                    {
+                        // check if "value" matches the one at "path"
+                        // the "path" location must exist - use at()
+                        success = (result.at(ptr) == get_value("test", "value", false));
+                    }
+                    JSON_INTERNAL_CATCH (out_of_range&)
+                    {
+                        // ignore out of range errors: success remains false
+                    }
+
+                    // throw an exception if test fails
+                    if (JSON_HEDLEY_UNLIKELY(!success))
+                    {
+                        JSON_THROW(other_error::create(501, detail::concat("unsuccessful: ", val.dump()), &val));
+                    }
+
+                    break;
+                }
+
+                case patch_operations::invalid:
+                default:
+                {
+                    // op must be "add", "remove", "replace", "move", "copy", or
+                    // "test"
+                    JSON_THROW(parse_error::create(105, 0, detail::concat("operation value '", op, "' is invalid"), &val));
+                }
+            }
+        }
+    }
+
+    /// @brief applies a JSON patch to a copy of the current object
+    /// @sa https://json.nlohmann.me/api/basic_json/patch/
+    basic_json patch(const basic_json& json_patch) const
+    {
+        basic_json result = *this;
+        result.patch_inplace(json_patch);
+        return result;
+    }
+
+    /// @brief creates a diff as a JSON patch
+    /// @sa https://json.nlohmann.me/api/basic_json/diff/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json diff(const basic_json& source, const basic_json& target,
+                           const std::string& path = "")
+    {
+        // the patch
+        basic_json result(value_t::array);
+
+        // if the values are the same, return empty patch
+        if (source == target)
+        {
+            return result;
+        }
+
+        if (source.type() != target.type())
+        {
+            // different types: replace value
+            result.push_back(
+            {
+                {"op", "replace"}, {"path", path}, {"value", target}
+            });
+            return result;
+        }
+
+        switch (source.type())
+        {
+            case value_t::array:
+            {
+                // first pass: traverse common elements
+                std::size_t i = 0;
+                while (i < source.size() && i < target.size())
+                {
+                    // recursive call to compare array values at index i
+                    auto temp_diff = diff(source[i], target[i], detail::concat(path, '/', std::to_string(i)));
+                    result.insert(result.end(), temp_diff.begin(), temp_diff.end());
+                    ++i;
+                }
+
+                // We now reached the end of at least one array
+                // in a second pass, traverse the remaining elements
+
+                // remove my remaining elements
+                const auto end_index = static_cast<difference_type>(result.size());
+                while (i < source.size())
+                {
+                    // add operations in reverse order to avoid invalid
+                    // indices
+                    result.insert(result.begin() + end_index, object(
+                    {
+                        {"op", "remove"},
+                        {"path", detail::concat(path, '/', std::to_string(i))}
+                    }));
+                    ++i;
+                }
+
+                // add other remaining elements
+                while (i < target.size())
+                {
+                    result.push_back(
+                    {
+                        {"op", "add"},
+                        {"path", detail::concat(path, "/-")},
+                        {"value", target[i]}
+                    });
+                    ++i;
+                }
+
+                break;
+            }
+
+            case value_t::object:
+            {
+                // first pass: traverse this object's elements
+                for (auto it = source.cbegin(); it != source.cend(); ++it)
+                {
+                    // escape the key name to be used in a JSON patch
+                    const auto path_key = detail::concat(path, '/', detail::escape(it.key()));
+
+                    if (target.find(it.key()) != target.end())
+                    {
+                        // recursive call to compare object values at key it
+                        auto temp_diff = diff(it.value(), target[it.key()], path_key);
+                        result.insert(result.end(), temp_diff.begin(), temp_diff.end());
+                    }
+                    else
+                    {
+                        // found a key that is not in o -> remove it
+                        result.push_back(object(
+                        {
+                            {"op", "remove"}, {"path", path_key}
+                        }));
+                    }
+                }
+
+                // second pass: traverse other object's elements
+                for (auto it = target.cbegin(); it != target.cend(); ++it)
+                {
+                    if (source.find(it.key()) == source.end())
+                    {
+                        // found a key that is not in this -> add it
+                        const auto path_key = detail::concat(path, '/', detail::escape(it.key()));
+                        result.push_back(
+                        {
+                            {"op", "add"}, {"path", path_key},
+                            {"value", it.value()}
+                        });
+                    }
+                }
+
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                // both primitive type: replace value
+                result.push_back(
+                {
+                    {"op", "replace"}, {"path", path}, {"value", target}
+                });
+                break;
+            }
+        }
+
+        return result;
+    }
+    /// @}
+
+    ////////////////////////////////
+    // JSON Merge Patch functions //
+    ////////////////////////////////
+
+    /// @name JSON Merge Patch functions
+    /// @{
+
+    /// @brief applies a JSON Merge Patch
+    /// @sa https://json.nlohmann.me/api/basic_json/merge_patch/
+    void merge_patch(const basic_json& apply_patch)
+    {
+        if (apply_patch.is_object())
+        {
+            if (!is_object())
+            {
+                *this = object();
+            }
+            for (auto it = apply_patch.begin(); it != apply_patch.end(); ++it)
+            {
+                if (it.value().is_null())
+                {
+                    erase(it.key());
+                }
+                else
+                {
+                    operator[](it.key()).merge_patch(it.value());
+                }
+            }
+        }
+        else
+        {
+            *this = apply_patch;
+        }
+    }
+
+    /// @}
+};
+
+/// @brief user-defined to_string function for JSON values
+/// @sa https://json.nlohmann.me/api/basic_json/to_string/
+NLOHMANN_BASIC_JSON_TPL_DECLARATION
+std::string to_string(const NLOHMANN_BASIC_JSON_TPL& j)
+{
+    return j.dump();
+}
+
+inline namespace literals
+{
+inline namespace json_literals
+{
+
+/// @brief user-defined string literal for JSON values
+/// @sa https://json.nlohmann.me/api/basic_json/operator_literal_json/
+JSON_HEDLEY_NON_NULL(1)
+#if !defined(JSON_HEDLEY_GCC_VERSION) || JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0)
+    inline nlohmann::json operator ""_json(const char* s, std::size_t n)
+#else
+    inline nlohmann::json operator "" _json(const char* s, std::size_t n)
+#endif
+{
+    return nlohmann::json::parse(s, s + n);
+}
+
+/// @brief user-defined string literal for JSON pointer
+/// @sa https://json.nlohmann.me/api/basic_json/operator_literal_json_pointer/
+JSON_HEDLEY_NON_NULL(1)
+#if !defined(JSON_HEDLEY_GCC_VERSION) || JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0)
+    inline nlohmann::json::json_pointer operator ""_json_pointer(const char* s, std::size_t n)
+#else
+    inline nlohmann::json::json_pointer operator "" _json_pointer(const char* s, std::size_t n)
+#endif
+{
+    return nlohmann::json::json_pointer(std::string(s, n));
+}
+
+}  // namespace json_literals
+}  // namespace literals
+NLOHMANN_JSON_NAMESPACE_END
+
+///////////////////////
+// nonmember support //
+///////////////////////
+
+namespace std // NOLINT(cert-dcl58-cpp)
+{
+
+/// @brief hash value for JSON objects
+/// @sa https://json.nlohmann.me/api/basic_json/std_hash/
+NLOHMANN_BASIC_JSON_TPL_DECLARATION
+struct hash<nlohmann::NLOHMANN_BASIC_JSON_TPL> // NOLINT(cert-dcl58-cpp)
+{
+    std::size_t operator()(const nlohmann::NLOHMANN_BASIC_JSON_TPL& j) const
+    {
+        return nlohmann::detail::hash(j);
+    }
+};
+
+// specialization for std::less<value_t>
+template<>
+struct less< ::nlohmann::detail::value_t> // do not remove the space after '<', see https://github.com/nlohmann/json/pull/679
+{
+    /*!
+    @brief compare two value_t enum values
+    @since version 3.0.0
+    */
+    bool operator()(::nlohmann::detail::value_t lhs,
+                    ::nlohmann::detail::value_t rhs) const noexcept
+    {
+#if JSON_HAS_THREE_WAY_COMPARISON
+        return std::is_lt(lhs <=> rhs); // *NOPAD*
+#else
+        return ::nlohmann::detail::operator<(lhs, rhs);
+#endif
+    }
+};
+
+// C++20 prohibit function specialization in the std namespace.
+#ifndef JSON_HAS_CPP_20
+
+/// @brief exchanges the values of two JSON objects
+/// @sa https://json.nlohmann.me/api/basic_json/std_swap/
+NLOHMANN_BASIC_JSON_TPL_DECLARATION
+inline void swap(nlohmann::NLOHMANN_BASIC_JSON_TPL& j1, nlohmann::NLOHMANN_BASIC_JSON_TPL& j2) noexcept(  // NOLINT(readability-inconsistent-declaration-parameter-name, cert-dcl58-cpp)
+    is_nothrow_move_constructible<nlohmann::NLOHMANN_BASIC_JSON_TPL>::value&&                          // NOLINT(misc-redundant-expression,cppcoreguidelines-noexcept-swap,performance-noexcept-swap)
+    is_nothrow_move_assignable<nlohmann::NLOHMANN_BASIC_JSON_TPL>::value)
+{
+    j1.swap(j2);
+}
+
+#endif
+
+}  // namespace std
+
+#if JSON_USE_GLOBAL_UDLS
+    #if !defined(JSON_HEDLEY_GCC_VERSION) || JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0)
+        using nlohmann::literals::json_literals::operator ""_json; // NOLINT(misc-unused-using-decls,google-global-names-in-headers)
+        using nlohmann::literals::json_literals::operator ""_json_pointer; //NOLINT(misc-unused-using-decls,google-global-names-in-headers)
+    #else
+        using nlohmann::literals::json_literals::operator "" _json; // NOLINT(misc-unused-using-decls,google-global-names-in-headers)
+        using nlohmann::literals::json_literals::operator "" _json_pointer; //NOLINT(misc-unused-using-decls,google-global-names-in-headers)
+    #endif
+#endif
+
+// #include <nlohmann/detail/macro_unscope.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// restore clang diagnostic settings
+#if defined(__clang__)
+    #pragma clang diagnostic pop
+#endif
+
+// clean up
+#undef JSON_ASSERT
+#undef JSON_INTERNAL_CATCH
+#undef JSON_THROW
+#undef JSON_PRIVATE_UNLESS_TESTED
+#undef NLOHMANN_BASIC_JSON_TPL_DECLARATION
+#undef NLOHMANN_BASIC_JSON_TPL
+#undef JSON_EXPLICIT
+#undef NLOHMANN_CAN_CALL_STD_FUNC_IMPL
+#undef JSON_INLINE_VARIABLE
+#undef JSON_NO_UNIQUE_ADDRESS
+#undef JSON_DISABLE_ENUM_SERIALIZATION
+#undef JSON_USE_GLOBAL_UDLS
+
+#ifndef JSON_TEST_KEEP_MACROS
+    #undef JSON_CATCH
+    #undef JSON_TRY
+    #undef JSON_HAS_CPP_11
+    #undef JSON_HAS_CPP_14
+    #undef JSON_HAS_CPP_17
+    #undef JSON_HAS_CPP_20
+    #undef JSON_HAS_FILESYSTEM
+    #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+    #undef JSON_HAS_THREE_WAY_COMPARISON
+    #undef JSON_HAS_RANGES
+    #undef JSON_HAS_STATIC_RTTI
+    #undef JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
+#endif
+
+// #include <nlohmann/thirdparty/hedley/hedley_undef.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.3
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#undef JSON_HEDLEY_ALWAYS_INLINE
+#undef JSON_HEDLEY_ARM_VERSION
+#undef JSON_HEDLEY_ARM_VERSION_CHECK
+#undef JSON_HEDLEY_ARRAY_PARAM
+#undef JSON_HEDLEY_ASSUME
+#undef JSON_HEDLEY_BEGIN_C_DECLS
+#undef JSON_HEDLEY_CLANG_HAS_ATTRIBUTE
+#undef JSON_HEDLEY_CLANG_HAS_BUILTIN
+#undef JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE
+#undef JSON_HEDLEY_CLANG_HAS_DECLSPEC_DECLSPEC_ATTRIBUTE
+#undef JSON_HEDLEY_CLANG_HAS_EXTENSION
+#undef JSON_HEDLEY_CLANG_HAS_FEATURE
+#undef JSON_HEDLEY_CLANG_HAS_WARNING
+#undef JSON_HEDLEY_COMPCERT_VERSION
+#undef JSON_HEDLEY_COMPCERT_VERSION_CHECK
+#undef JSON_HEDLEY_CONCAT
+#undef JSON_HEDLEY_CONCAT3
+#undef JSON_HEDLEY_CONCAT3_EX
+#undef JSON_HEDLEY_CONCAT_EX
+#undef JSON_HEDLEY_CONST
+#undef JSON_HEDLEY_CONSTEXPR
+#undef JSON_HEDLEY_CONST_CAST
+#undef JSON_HEDLEY_CPP_CAST
+#undef JSON_HEDLEY_CRAY_VERSION
+#undef JSON_HEDLEY_CRAY_VERSION_CHECK
+#undef JSON_HEDLEY_C_DECL
+#undef JSON_HEDLEY_DEPRECATED
+#undef JSON_HEDLEY_DEPRECATED_FOR
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION
+#undef JSON_HEDLEY_DIAGNOSTIC_POP
+#undef JSON_HEDLEY_DIAGNOSTIC_PUSH
+#undef JSON_HEDLEY_DMC_VERSION
+#undef JSON_HEDLEY_DMC_VERSION_CHECK
+#undef JSON_HEDLEY_EMPTY_BASES
+#undef JSON_HEDLEY_EMSCRIPTEN_VERSION
+#undef JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK
+#undef JSON_HEDLEY_END_C_DECLS
+#undef JSON_HEDLEY_FLAGS
+#undef JSON_HEDLEY_FLAGS_CAST
+#undef JSON_HEDLEY_GCC_HAS_ATTRIBUTE
+#undef JSON_HEDLEY_GCC_HAS_BUILTIN
+#undef JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE
+#undef JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE
+#undef JSON_HEDLEY_GCC_HAS_EXTENSION
+#undef JSON_HEDLEY_GCC_HAS_FEATURE
+#undef JSON_HEDLEY_GCC_HAS_WARNING
+#undef JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK
+#undef JSON_HEDLEY_GCC_VERSION
+#undef JSON_HEDLEY_GCC_VERSION_CHECK
+#undef JSON_HEDLEY_GNUC_HAS_ATTRIBUTE
+#undef JSON_HEDLEY_GNUC_HAS_BUILTIN
+#undef JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE
+#undef JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE
+#undef JSON_HEDLEY_GNUC_HAS_EXTENSION
+#undef JSON_HEDLEY_GNUC_HAS_FEATURE
+#undef JSON_HEDLEY_GNUC_HAS_WARNING
+#undef JSON_HEDLEY_GNUC_VERSION
+#undef JSON_HEDLEY_GNUC_VERSION_CHECK
+#undef JSON_HEDLEY_HAS_ATTRIBUTE
+#undef JSON_HEDLEY_HAS_BUILTIN
+#undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE
+#undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS
+#undef JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE
+#undef JSON_HEDLEY_HAS_EXTENSION
+#undef JSON_HEDLEY_HAS_FEATURE
+#undef JSON_HEDLEY_HAS_WARNING
+#undef JSON_HEDLEY_IAR_VERSION
+#undef JSON_HEDLEY_IAR_VERSION_CHECK
+#undef JSON_HEDLEY_IBM_VERSION
+#undef JSON_HEDLEY_IBM_VERSION_CHECK
+#undef JSON_HEDLEY_IMPORT
+#undef JSON_HEDLEY_INLINE
+#undef JSON_HEDLEY_INTEL_CL_VERSION
+#undef JSON_HEDLEY_INTEL_CL_VERSION_CHECK
+#undef JSON_HEDLEY_INTEL_VERSION
+#undef JSON_HEDLEY_INTEL_VERSION_CHECK
+#undef JSON_HEDLEY_IS_CONSTANT
+#undef JSON_HEDLEY_IS_CONSTEXPR_
+#undef JSON_HEDLEY_LIKELY
+#undef JSON_HEDLEY_MALLOC
+#undef JSON_HEDLEY_MCST_LCC_VERSION
+#undef JSON_HEDLEY_MCST_LCC_VERSION_CHECK
+#undef JSON_HEDLEY_MESSAGE
+#undef JSON_HEDLEY_MSVC_VERSION
+#undef JSON_HEDLEY_MSVC_VERSION_CHECK
+#undef JSON_HEDLEY_NEVER_INLINE
+#undef JSON_HEDLEY_NON_NULL
+#undef JSON_HEDLEY_NO_ESCAPE
+#undef JSON_HEDLEY_NO_RETURN
+#undef JSON_HEDLEY_NO_THROW
+#undef JSON_HEDLEY_NULL
+#undef JSON_HEDLEY_PELLES_VERSION
+#undef JSON_HEDLEY_PELLES_VERSION_CHECK
+#undef JSON_HEDLEY_PGI_VERSION
+#undef JSON_HEDLEY_PGI_VERSION_CHECK
+#undef JSON_HEDLEY_PREDICT
+#undef JSON_HEDLEY_PRINTF_FORMAT
+#undef JSON_HEDLEY_PRIVATE
+#undef JSON_HEDLEY_PUBLIC
+#undef JSON_HEDLEY_PURE
+#undef JSON_HEDLEY_REINTERPRET_CAST
+#undef JSON_HEDLEY_REQUIRE
+#undef JSON_HEDLEY_REQUIRE_CONSTEXPR
+#undef JSON_HEDLEY_REQUIRE_MSG
+#undef JSON_HEDLEY_RESTRICT
+#undef JSON_HEDLEY_RETURNS_NON_NULL
+#undef JSON_HEDLEY_SENTINEL
+#undef JSON_HEDLEY_STATIC_ASSERT
+#undef JSON_HEDLEY_STATIC_CAST
+#undef JSON_HEDLEY_STRINGIFY
+#undef JSON_HEDLEY_STRINGIFY_EX
+#undef JSON_HEDLEY_SUNPRO_VERSION
+#undef JSON_HEDLEY_SUNPRO_VERSION_CHECK
+#undef JSON_HEDLEY_TINYC_VERSION
+#undef JSON_HEDLEY_TINYC_VERSION_CHECK
+#undef JSON_HEDLEY_TI_ARMCL_VERSION
+#undef JSON_HEDLEY_TI_ARMCL_VERSION_CHECK
+#undef JSON_HEDLEY_TI_CL2000_VERSION
+#undef JSON_HEDLEY_TI_CL2000_VERSION_CHECK
+#undef JSON_HEDLEY_TI_CL430_VERSION
+#undef JSON_HEDLEY_TI_CL430_VERSION_CHECK
+#undef JSON_HEDLEY_TI_CL6X_VERSION
+#undef JSON_HEDLEY_TI_CL6X_VERSION_CHECK
+#undef JSON_HEDLEY_TI_CL7X_VERSION
+#undef JSON_HEDLEY_TI_CL7X_VERSION_CHECK
+#undef JSON_HEDLEY_TI_CLPRU_VERSION
+#undef JSON_HEDLEY_TI_CLPRU_VERSION_CHECK
+#undef JSON_HEDLEY_TI_VERSION
+#undef JSON_HEDLEY_TI_VERSION_CHECK
+#undef JSON_HEDLEY_UNAVAILABLE
+#undef JSON_HEDLEY_UNLIKELY
+#undef JSON_HEDLEY_UNPREDICTABLE
+#undef JSON_HEDLEY_UNREACHABLE
+#undef JSON_HEDLEY_UNREACHABLE_RETURN
+#undef JSON_HEDLEY_VERSION
+#undef JSON_HEDLEY_VERSION_DECODE_MAJOR
+#undef JSON_HEDLEY_VERSION_DECODE_MINOR
+#undef JSON_HEDLEY_VERSION_DECODE_REVISION
+#undef JSON_HEDLEY_VERSION_ENCODE
+#undef JSON_HEDLEY_WARNING
+#undef JSON_HEDLEY_WARN_UNUSED_RESULT
+#undef JSON_HEDLEY_WARN_UNUSED_RESULT_MSG
+#undef JSON_HEDLEY_FALL_THROUGH
+
+
+
+#endif  // INCLUDE_NLOHMANN_JSON_HPP_
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/memory_utils.cpp b/projects/llm_framework/main_cosy_voice/src/runner/utils/memory_utils.cpp
new file mode 100644
index 00000000..5ee81b72
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/memory_utils.cpp
@@ -0,0 +1,65 @@
+#include "memory_utils.hpp"
+
+bool file_exist(const std::string &path)
+{
+    auto flag = false;
+
+    std::fstream fs(path, std::ios::in | std::ios::binary);
+    flag = fs.is_open();
+    fs.close();
+
+    return flag;
+}
+
+bool read_file(const std::string &path, std::vector<char> &data)
+{
+
+    std::fstream fs(path, std::ios::in | std::ios::binary);
+
+    if (!fs.is_open())
+    {
+        return false;
+    }
+
+    // get file size
+    fs.seekg(0, std::ios::end);
+    size_t file_size = fs.tellg();
+    fs.seekg(0, std::ios::beg);
+
+    if (file_size == 0)
+    {
+        return false;
+    }
+
+    data.resize(file_size);
+    fs.read(data.data(), file_size);
+    // data.insert(data.end(), std::istreambuf_iterator<char>(fs), std::istreambuf_iterator<char>());
+
+    fs.close();
+
+    return true;
+}
+
+bool read_file(const std::string &path, char **data, size_t *len)
+{
+    FILE *fp = fopen(path.c_str(), "rb");
+
+    if (!fp)
+    {
+        return false;
+    }
+
+    fseek(fp, 0, SEEK_END);
+
+    *len = ftell(fp);
+
+    fseek(fp, 0, SEEK_SET);
+
+    *data = new char[*len];
+
+    fread(*data, *len, 1, fp);
+
+    fclose(fp);
+
+    return true;
+}
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/memory_utils.hpp b/projects/llm_framework/main_cosy_voice/src/runner/utils/memory_utils.hpp
new file mode 100644
index 00000000..37965b14
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/memory_utils.hpp
@@ -0,0 +1,74 @@
+#pragma once
+#include <sys/mman.h>
+#include <fcntl.h>
+#include <stdio.h>
+#include <fstream>
+#include <vector>
+
+bool file_exist(const std::string &path);
+
+bool read_file(const std::string &path, std::vector<char> &data);
+bool read_file(const std::string &path, char **data, size_t *len);
+class MMap
+{
+private:
+    void *_add = nullptr;
+    int _size;
+
+public:
+    MMap() {}
+    MMap(const char *file)
+    {
+        open_file(file);
+    }
+    ~MMap()
+    {
+        close_file();
+    }
+
+    bool open_file(const char *file)
+    {
+        _add = _mmap(file, &_size);
+        if (!_add)
+        {
+            return false;
+        }
+        return true;
+    }
+
+    void close_file()
+    {
+        if (_add)
+        {
+            munmap(_add, _size);
+            _add = nullptr;
+            _size = 0;
+        }
+    }
+
+    size_t size()
+    {
+        return _size;
+    }
+
+    void *data()
+    {
+        return _add;
+    }
+
+    static void *_mmap(const char *model_file, int *model_size)
+    {
+        auto *file_fp = fopen(model_file, "r");
+        if (!file_fp)
+        {
+
+            return nullptr;
+        }
+        fseek(file_fp, 0, SEEK_END);
+        *model_size = ftell(file_fp);
+        fclose(file_fp);
+        int fd = open(model_file, O_RDWR, 0644);
+        void *mmap_add = mmap(NULL, *model_size, PROT_WRITE, MAP_SHARED, fd, 0);
+        return mmap_add;
+    }
+};
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/sample_log.h b/projects/llm_framework/main_cosy_voice/src/runner/utils/sample_log.h
new file mode 100755
index 00000000..f03fb635
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/sample_log.h
@@ -0,0 +1,65 @@
+/**********************************************************************************
+ *
+ * Copyright (c) 2019-2022 Beijing AXera Technology Co., Ltd. All Rights Reserved.
+ *
+ * This source file is the property of Beijing AXera Technology Co., Ltd. and
+ * may not be copied or distributed in any isomorphic form without the prior
+ * written consent of Beijing AXera Technology Co., Ltd.
+ *
+ **********************************************************************************/
+#ifndef _SAMPLE_LOG_H_
+#define _SAMPLE_LOG_H_
+
+#include <stdio.h>
+
+typedef enum
+{
+    SAMPLE_LOG_MIN = -1,
+    SAMPLE_LOG_EMERGENCY = 0,
+    SAMPLE_LOG_ALERT = 1,
+    SAMPLE_LOG_CRITICAL = 2,
+    SAMPLE_LOG_ERROR = 3,
+    SAMPLE_LOG_WARN = 4,
+    SAMPLE_LOG_NOTICE = 5,
+    SAMPLE_LOG_INFO = 6,
+    SAMPLE_LOG_DEBUG = 7,
+    SAMPLE_LOG_MAX
+} SAMPLE_LOG_LEVEL_E;
+
+static SAMPLE_LOG_LEVEL_E log_level = SAMPLE_LOG_INFO;
+
+#if 1
+#define MACRO_BLACK "\033[1;30;30m"
+#define MACRO_RED "\033[1;30;31m"
+#define MACRO_GREEN "\033[1;30;32m"
+#define MACRO_YELLOW "\033[1;30;33m"
+#define MACRO_BLUE "\033[1;30;34m"
+#define MACRO_PURPLE "\033[1;30;35m"
+#define MACRO_WHITE "\033[1;30;37m"
+#define MACRO_END "\033[0m"
+#else
+#define MACRO_BLACK
+#define MACRO_RED
+#define MACRO_GREEN
+#define MACRO_YELLOW
+#define MACRO_BLUE
+#define MACRO_PURPLE
+#define MACRO_WHITE
+#define MACRO_END
+#endif
+
+#define ALOGE(fmt, ...) printf(MACRO_RED "[E][%32s][%4d]: " fmt MACRO_END "\n", __FUNCTION__, __LINE__, ##__VA_ARGS__)
+#define ALOGW(fmt, ...)               \
+    if (log_level >= SAMPLE_LOG_WARN) \
+    printf(MACRO_YELLOW "[W][%32s][%4d]: " fmt MACRO_END "\n", __FUNCTION__, __LINE__, ##__VA_ARGS__)
+#define ALOGI(fmt, ...)               \
+    if (log_level >= SAMPLE_LOG_INFO) \
+    printf(MACRO_GREEN "[I][%32s][%4d]: " fmt MACRO_END "\n", __FUNCTION__, __LINE__, ##__VA_ARGS__)
+#define ALOGD(fmt, ...)                \
+    if (log_level >= SAMPLE_LOG_DEBUG) \
+    printf(MACRO_WHITE "[D][%32s][%4d]: " fmt MACRO_END "\n", __FUNCTION__, __LINE__, ##__VA_ARGS__)
+#define ALOGN(fmt, ...)                 \
+    if (log_level >= SAMPLE_LOG_NOTICE) \
+    printf(MACRO_PURPLE "[N][%32s][%4d]: " fmt MACRO_END "\n", __FUNCTION__, __LINE__, ##__VA_ARGS__)
+
+#endif /* _SAMPLE_LOG_H_ */
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/sampling.hpp b/projects/llm_framework/main_cosy_voice/src/runner/utils/sampling.hpp
new file mode 100644
index 00000000..77460d86
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/sampling.hpp
@@ -0,0 +1,200 @@
+// sampling.h
+#ifndef SAMPLING_H
+#define SAMPLING_H
+
+#include <vector>
+#include <random>
+#include <algorithm>
+#include <cmath>
+#include <limits>
+#include <numeric>
+#include <iostream> // For debugging, can be removed
+#include <stdexcept>
+
+namespace sampling {
+    
+    // Numerically stable softmax implementation [[32]]
+    std::vector<float> softmax_stable(const std::vector<float>& logits) {
+        if (logits.empty()) return {};
+
+        // Find the maximum value for numerical stability [[32]]
+        float max_val = *std::max_element(logits.begin(), logits.end());
+
+        std::vector<float> exp_values(logits.size());
+        float sum_exp = 0.0f;
+        for (size_t i = 0; i < logits.size(); ++i) {
+            exp_values[i] = std::exp(logits[i] - max_val); // Subtract max for stability [[32]]
+            sum_exp += exp_values[i];
+        }
+
+        // Normalize
+        if (sum_exp > 0.0f) {
+            for (float& val : exp_values) {
+                val /= sum_exp;
+            }
+        } else {
+            // Handle case where all logits are very negative (sum_exp ~ 0)
+            // Assign uniform probability
+            float uniform_prob = 1.0f / static_cast<float>(exp_values.size());
+            for (float& val : exp_values) {
+                val = uniform_prob;
+            }
+        }
+        return exp_values;
+    }
+
+    // Sort indices based on values in descending order [[29]]
+    std::vector<size_t> sort_indices_desc(const std::vector<float>& v) {
+        std::vector<size_t> idx(v.size());
+        std::iota(idx.begin(), idx.end(), 0); // Fill idx with 0, 1, ..., v.size()-1 [[23]]
+
+        // Sort indices based on the corresponding values in 'v' in descending order [[24]]
+        std::stable_sort(idx.begin(), idx.end(),
+                [&v](size_t i1, size_t i2) {return v[i1] > v[i2];}); // [[29]]
+
+        return idx;
+    }
+
+    // Multinomial sampling with replacement [[45]]
+    int sample_multinomial(const std::vector<float>& probabilities, std::mt19937& gen) {
+        if (probabilities.empty()) {
+            throw std::invalid_argument("Cannot sample from an empty probability distribution.");
+        }
+
+        std::discrete_distribution<int> dist(probabilities.begin(), probabilities.end()); // [[13]]
+        return dist(gen);
+    }
+
+    // --- Core Sampling Functions ---
+
+    // Nucleus (Top-p) Sampling with Top-k filtering
+    int nucleus_sampling(const std::vector<float>& weighted_scores, float top_p = 0.8f, int top_k = 25) {
+        if (weighted_scores.empty()) {
+            throw std::invalid_argument("weighted_scores cannot be empty.");
+        }
+
+        // 1. Apply softmax to get probabilities
+        std::vector<float> probs = softmax_stable(weighted_scores);
+
+        // 2. Get sorted indices (descending) [[29]]
+        std::vector<size_t> sorted_indices = sort_indices_desc(probs);
+
+        // 3. Apply Top-p and Top-k filtering
+        std::vector<float> filtered_probs;
+        std::vector<size_t> filtered_indices;
+        float cum_prob = 0.0f;
+
+        int actual_top_k = std::min(top_k, static_cast<int>(sorted_indices.size()));
+
+        for (int i = 0; i < actual_top_k; ++i) {
+            size_t idx = sorted_indices[i];
+            float prob = probs[idx];
+            if (cum_prob < top_p && static_cast<int>(filtered_probs.size()) < top_k) {
+                cum_prob += prob;
+                filtered_probs.push_back(prob);
+                filtered_indices.push_back(idx);
+            } else {
+                break; // Stop if cumulative probability exceeds top_p or reached top_k limit
+            }
+        }
+
+        if (filtered_probs.empty()) {
+            // This can happen if the first element's probability is >= top_p
+            // or if top_k is 0. Fall back to sampling from the top element.
+            // Or if all probabilities were 0 (handled by softmax_stable).
+            filtered_probs.push_back(1.0f);
+            filtered_indices.push_back(sorted_indices[0]);
+        }
+
+        // 4. Re-normalize the filtered probabilities (like PyTorch does)
+        // This is crucial to match Python behavior
+        float sum_filtered = std::accumulate(filtered_probs.begin(), filtered_probs.end(), 0.0f);
+        if (sum_filtered > 0.0f) {
+            for (float& prob : filtered_probs) {
+                prob /= sum_filtered;
+            }
+        }
+
+        // 5. Sample from the filtered distribution
+        static std::random_device rd; // Static to seed once
+        static std::mt19937 gen(rd());
+        int sampled_index_in_filtered = sample_multinomial(filtered_probs, gen);
+
+        // 6. Return the original index
+        return static_cast<int>(filtered_indices[sampled_index_in_filtered]);
+    }
+
+    // Random Sampling (equivalent to Top-1 with temperature -> 1)
+    int random_sampling(const std::vector<float>& weighted_scores) {
+        if (weighted_scores.empty()) {
+            throw std::invalid_argument("weighted_scores cannot be empty.");
+        }
+        std::vector<float> probs = softmax_stable(weighted_scores);
+
+        static std::random_device rd;
+        static std::mt19937 gen(rd());
+        return sample_multinomial(probs, gen);
+    }
+
+    // Repetition-Aware Sampling (RAS)
+    int ras_sampling(const std::vector<float>& weighted_scores,
+                    const std::vector<int>& decoded_tokens,
+                    int speech_token_size, // Assuming this is passed or part of context
+                    float top_p = 0.8f, int top_k = 25,
+                    int win_size = 10, float tau_r = 0.1f) {
+
+        // 1. Perform Nucleus Sampling
+        int top_id = nucleus_sampling(weighted_scores, top_p, top_k);
+
+        // 2. Check for repetition
+        int rep_num = 0;
+        int window_start = std::max(0, static_cast<int>(decoded_tokens.size()) - win_size);
+        for (size_t i = window_start; i < decoded_tokens.size(); ++i) {
+            if (decoded_tokens[i] == top_id) {
+                rep_num++;
+            }
+        }
+
+        // 3. If repetition threshold is met, fallback to random sampling
+        if (rep_num >= static_cast<int>(win_size * tau_r)) {
+            top_id = random_sampling(weighted_scores);
+        }
+
+        return top_id;
+    }
+
+    // Main sampling function with EOS handling
+    int sampling_ids(const std::vector<float>& weighted_scores,
+                    const std::vector<int>& decoded_tokens,
+                    int speech_token_size, // Assuming this is passed or part of context
+                    bool ignore_eos = true,
+                    int max_trials = 100) {
+
+        static std::random_device rd;
+        static std::mt19937 gen(rd()); // Static for efficiency in loops
+
+        int num_trials = 0;
+        int top_id = -1; // Initialize
+
+        while (true) {
+            top_id = ras_sampling(weighted_scores, decoded_tokens, speech_token_size);
+
+            // Check EOS condition
+            if (!ignore_eos || (speech_token_size < 0 || top_id != speech_token_size)) {
+                break; // Accept the sample if EOS is not ignored, or if it's not the EOS token
+            }
+
+            num_trials++;
+            if (num_trials > max_trials) {
+                throw std::runtime_error("sampling reaches max_trials " + std::to_string(max_trials) +
+                                        " and still gets eos when ignore_eos is True, check your input!");
+            }
+        }
+        return top_id;
+    }
+
+
+
+} // namespace sampling
+
+#endif // SAMPLING_H
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/slice_3d.h b/projects/llm_framework/main_cosy_voice/src/runner/utils/slice_3d.h
new file mode 100644
index 00000000..a98ec73c
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/slice_3d.h
@@ -0,0 +1,44 @@
+// src/slice_3d.h
+#ifndef SLICE_3D_H
+#define SLICE_3D_H
+
+#include <vector>
+#include <cstddef>
+#include <type_traits>
+
+// Template declarations
+template <typename T>
+typename std::enable_if<std::is_integral<T>::value && std::is_signed<T>::value, size_t>::type
+normalize_slice_index(T index, size_t dim_size);
+
+template <typename T>
+typename std::enable_if<std::is_arithmetic<T>::value, std::vector<T>>::type
+slice_3d_optimized(std::vector<T>& data_1d,
+                   size_t dim0, size_t dim1, size_t dim2,
+                   int start0, int stop0,
+                   int start1, int stop1,
+                   int start2, int stop2);
+
+template <typename T>
+typename std::enable_if<std::is_arithmetic<T>::value, std::vector<T>>::type
+slice_3d_last_dim_from(std::vector<T>& data_1d,
+                       size_t dim0, size_t dim1, size_t dim2,
+                       int start2);
+
+// 这个函数有bug
+// template <typename T>
+// typename std::enable_if<std::is_arithmetic<T>::value, std::vector<T>>::type
+// slice_3d_last_dim_last_n(std::vector<T>& data_1d,
+//                          size_t dim0, size_t dim1, size_t dim2,
+//                          size_t n);
+
+template <typename T>
+typename std::enable_if<std::is_arithmetic<T>::value, std::vector<T>>::type
+slice_3d_last_dim_range(std::vector<T>& data_1d,
+                        size_t dim0, size_t dim1, size_t dim2,
+                        int start2, int stop2);
+
+// Include the implementation for templates
+#include "slice_3d.tpp"
+
+#endif // SLICE_3D_H
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/slice_3d.tpp b/projects/llm_framework/main_cosy_voice/src/runner/utils/slice_3d.tpp
new file mode 100644
index 00000000..bfcd71ad
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/slice_3d.tpp
@@ -0,0 +1,253 @@
+// src/slice_3d.tpp
+#ifndef SLICE_3D_TPP
+#define SLICE_3D_TPP
+
+#include "slice_3d.h"
+#include <algorithm>
+#include <stdexcept>
+#include <cstring>
+#include <cassert>
+#include <iostream>
+#include <fstream>
+#include <string>
+#include <sstream>
+#include <iomanip> // For std::setprecision
+
+// --- Helper for index normalization ---
+// Handles ALL negative indices, including -1, according to Python rules.
+// -1 means the last element, -2 the second to last, etc.
+template <typename T>
+typename std::enable_if<std::is_integral<T>::value && std::is_signed<T>::value, size_t>::type
+normalize_slice_index(T index, size_t dim_size) {
+    if (index < 0) {
+        // In Python, negative indices are relative to the end.
+        // e.g., -1 is the last element, -2 is second to last.
+        // We clamp the result to be >= 0.
+        return static_cast<size_t>(std::max(static_cast<T>(0), static_cast<T>(dim_size) + index));
+    } else {
+        // Clamp positive indices to dim_size (Python allows index==dim_size for stop)
+        return static_cast<size_t>(std::min(static_cast<size_t>(index), dim_size));
+    }
+}
+
+// --- Main optimized slicing function ---
+template <typename T>
+typename std::enable_if<std::is_arithmetic<T>::value, std::vector<T>>::type
+slice_3d_optimized(std::vector<T>& data_1d,
+                   size_t dim0, size_t dim1, size_t dim2,
+                   int start0, int stop0,
+                   int start1, int stop1,
+                   int start2, int stop2) {
+
+    // --- Input Validation ---
+    const size_t total_elements = dim0 * dim1 * dim2;
+    if (data_1d.size() != total_elements) {
+        throw std::invalid_argument("Data size does not match provided dimensions.");
+    }
+
+    // --- Normalize ALL Indices using the same logic ---
+    // This correctly handles ALL negative indices, including stop==-1.
+    // There is NO special case for stop==-1 anymore.
+    const size_t norm_start0 = normalize_slice_index(start0, dim0);
+    const size_t norm_stop0 = normalize_slice_index(stop0, dim0); // Treat -1 as "last element index"
+
+    const size_t norm_start1 = normalize_slice_index(start1, dim1);
+    const size_t norm_stop1 = normalize_slice_index(stop1, dim1); // Treat -1 as "last element index"
+
+    const size_t norm_start2 = normalize_slice_index(start2, dim2);
+    const size_t norm_stop2 = normalize_slice_index(stop2, dim2); // Treat -1 as "last element index"
+
+    // --- Calculate Final Slice Boundaries ---
+    // Ensure start <= stop (handles user errors like [5:3] and empty slices like [3:3])
+    // Python slicing returns an empty list if start >= stop.
+    // const size_t final_start0 = std::min(norm_start0, norm_stop0);
+    // const size_t final_stop0 = std::max(norm_start0, norm_stop0); // Ensure stop >= start
+
+    // const size_t final_start1 = std::min(norm_start1, norm_stop1);
+    // const size_t final_stop1 = std::max(norm_start1, norm_stop1); // Ensure stop >= start
+
+    // const size_t final_start2 = std::min(norm_start2, norm_stop2);
+    // const size_t final_stop2 = std::max(norm_start2, norm_stop2); // Ensure stop >= start
+
+    // --- Calculate Effective Slice Lengths ---
+    const size_t final_start0 = norm_start0;
+    const size_t final_stop0 = norm_stop0;
+    const size_t slice_len0 = (final_start0 < final_stop0) ? (final_stop0 - final_start0) : 0;
+
+    const size_t final_start1 = norm_start1;
+    const size_t final_stop1 = norm_stop1;
+    const size_t slice_len1 = (final_start1 < final_stop1) ? (final_stop1 - final_start1) : 0;
+
+    const size_t final_start2 = norm_start2;
+    const size_t final_stop2 = norm_stop2;
+    const size_t slice_len2 = (final_start2 < final_stop2) ? (final_stop2 - final_start2) : 0;
+
+    // --- Handle Empty Result ---
+    if (slice_len0 == 0 || slice_len1 == 0 || slice_len2 == 0) {
+        return std::vector<T>();
+    }
+
+    // --- Calculate Result Size ---
+    const size_t result_size = slice_len0 * slice_len1 * slice_len2;
+    std::vector<T> result_1d(result_size);
+
+    // --- Calculate Strides (Row-Major Order) ---
+    const size_t stride_dim2 = 1;
+    const size_t stride_dim1 = dim2 * stride_dim2;
+    const size_t stride_dim0 = dim1 * stride_dim1;
+
+    // --- Optimization Strategy ---
+
+    // 1. Slice only in the last dimension (dim2)
+    // Pattern: data[:, :, start2:stop2]
+    if (slice_len0 == dim0 && slice_len1 == dim1) {
+        const size_t elements_to_copy_per_row = slice_len2;
+        size_t result_idx = 0;
+        for (size_t n = 0; n < dim0; ++n) {
+            for (size_t c = 0; c < dim1; ++c) {
+                const size_t src_start_idx = n * stride_dim0 + c * stride_dim1 + final_start2;
+                std::copy(data_1d.data() + src_start_idx,
+                          data_1d.data() + src_start_idx + elements_to_copy_per_row,
+                          result_1d.data() + result_idx);
+                result_idx += elements_to_copy_per_row;
+            }
+        }
+        return result_1d;
+    }
+
+    // 2. Slice only in the middle dimension (dim1) and full last dimension (dim2)
+    // Pattern: data[:, start1:stop1, :] (when start2==0 && stop2==dim2 implicitly)
+    if (slice_len0 == dim0 && slice_len2 == dim2) {
+        // Check if the slice on dim2 is indeed the full dimension
+        if (final_start2 == 0 && final_stop2 == dim2) {
+             const size_t elements_to_copy_per_n_block = slice_len1 * dim2;
+             size_t result_idx = 0;
+             for (size_t n = 0; n < dim0; ++n) {
+                 const size_t src_start_idx = n * stride_dim0 + final_start1 * stride_dim1;
+                 std::copy(data_1d.data() + src_start_idx,
+                           data_1d.data() + src_start_idx + elements_to_copy_per_n_block,
+                           result_1d.data() + result_idx);
+                 result_idx += elements_to_copy_per_n_block;
+             }
+             return result_1d;
+        }
+        // If slicing dim2 too in a non-optimized way, fall through to general case.
+    }
+
+    // 3. Slice only in the first dimension (dim0) and full middle/last dimensions
+    // Pattern: data[start0:stop0, :, :] (when start1==0, stop1==dim1, start2==0, stop2==dim2 implicitly)
+    if (slice_len1 == dim1 && slice_len2 == dim2) {
+        // Check if the slices on dim1 and dim2 are indeed the full dimensions
+        if (final_start1 == 0 && final_stop1 == dim1 &&
+            final_start2 == 0 && final_stop2 == dim2) {
+             const size_t elements_to_copy = slice_len0 * dim1 * dim2;
+             const size_t src_start_idx = final_start0 * stride_dim0;
+             std::copy(data_1d.data() + src_start_idx,
+                       data_1d.data() + src_start_idx + elements_to_copy,
+                       result_1d.data());
+             return result_1d;
+        }
+        // If slicing dim1 or dim2 too, fall through to general case.
+    }
+
+    // --- General Case (Fallback for complex or non-contiguous slices) ---
+    // This handles all other cases.
+    size_t result_idx = 0;
+    for (size_t n = final_start0; n < final_stop0; ++n) {
+        const size_t base_offset_n = n * stride_dim0;
+        for (size_t c = final_start1; c < final_stop1; ++c) {
+            const size_t base_offset_nc = base_offset_n + c * stride_dim1;
+            for (size_t l = final_start2; l < final_stop2; ++l) {
+                const size_t original_index = base_offset_nc + l * stride_dim2;
+                result_1d[result_idx++] = data_1d[original_index];
+            }
+        }
+    }
+
+    return result_1d;
+}
+
+// --- Convenience Wrappers ---
+// These wrappers now pass dimX instead of -1 to signify "the end of the dimension".
+
+template <typename T>
+typename std::enable_if<std::is_arithmetic<T>::value, std::vector<T>>::type
+slice_3d_last_dim_from(std::vector<T>& data_1d,
+                       size_t dim0, size_t dim1, size_t dim2,
+                       int start2) {
+    // Slice all of dim0 and dim1, and from start2 to the end of dim2.
+    // Pass dim2 as stop2 to indicate the end.
+    return slice_3d_optimized(data_1d, dim0, dim1, dim2,
+                              0, static_cast<int>(dim0), // [:] for dim0
+                              0, static_cast<int>(dim1), // [:] for dim1
+                              start2, static_cast<int>(dim2)); // [start2:] for dim2
+}
+
+template <typename T>
+typename std::enable_if<std::is_arithmetic<T>::value, std::vector<T>>::type
+slice_3d_last_dim_last_n(std::vector<T>& data_1d,
+                         size_t dim0, size_t dim1, size_t dim2,
+                         size_t n) {
+    // Calculate the correct start index for the last N elements.
+    // Handle case where n >= dim2.
+    int start2 = (n >= dim2) ? 0 : static_cast<int>(dim2 - n);
+    // Slice all of dim0 and dim1, and from start2 to the end of dim2.
+    // Pass dim2 as stop2 to indicate the end.
+    return slice_3d_optimized(data_1d, dim0, dim1, dim2,
+                              0, static_cast<int>(dim0), // [:] for dim0
+                              0, static_cast<int>(dim1), // [:] for dim1
+                              start2, static_cast<int>(dim2)); // [-n:] -> [dim2-n:] for dim2
+}
+
+template <typename T>
+typename std::enable_if<std::is_arithmetic<T>::value, std::vector<T>>::type
+slice_3d_last_dim_range(std::vector<T>& data_1d,
+                        size_t dim0, size_t dim1, size_t dim2,
+                        int start2, int stop2) {
+    // Slice all of dim0 and dim1, and the specified range in dim2.
+    return slice_3d_optimized(data_1d, dim0, dim1, dim2,
+                              0, static_cast<int>(dim0), // [:] for dim0
+                              0, static_cast<int>(dim1), // [:] for dim1
+                              start2, stop2); // [start2:stop2] for dim2
+}
+
+// --- Utility to save vector to file ---
+template<typename T>
+void save_vector_to_file(const std::vector<T>& vec, const std::string& filename) {
+    std::ofstream file(filename);
+    if (!file.is_open()) {
+        throw std::runtime_error("Could not open file for writing: " + filename);
+    }
+    file << std::fixed << std::setprecision(10); // Set precision for floats
+    for (size_t i = 0; i < vec.size(); ++i) {
+        file << vec[i];
+        if (i < vec.size() - 1) {
+            file << "\n"; // Newline separated
+        }
+    }
+    file.close();
+}
+
+// --- Utility to load vector from file ---
+template<typename T>
+std::vector<T> load_vector_from_file(const std::string& filename) {
+     std::ifstream file(filename);
+    if (!file.is_open()) {
+        throw std::runtime_error("Could not open file for reading: " + filename);
+    }
+    std::vector<T> vec;
+    std::string line;
+    while(std::getline(file, line)) {
+        if (!line.empty()) {
+             std::stringstream ss(line);
+             T value;
+             ss >> value;
+             vec.push_back(value);
+        }
+    }
+    file.close();
+    return vec;
+}
+
+
+#endif // SLICE_3D_TPP
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/string_utility.hpp b/projects/llm_framework/main_cosy_voice/src/runner/utils/string_utility.hpp
new file mode 100644
index 00000000..12749a20
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/string_utility.hpp
@@ -0,0 +1,217 @@
+
+#pragma once
+
+#include <cstring>
+#include <algorithm>
+#include <sstream>
+#include <iomanip>
+#include <vector>
+
+template <typename Str>
+struct tokenizer
+{
+    tokenizer(Str const& str)
+        : _string(str), _offset(0)
+    {}
+    tokenizer(Str const& str, Str const& delimiters)
+        : _string(str), _offset(0), _delimiters(delimiters)
+    {}
+
+    bool next_token()
+    {
+        return next_token(_delimiters);
+    }
+
+    bool next_token(Str const& delimiters)
+    {
+        size_t i = _string.find_first_not_of(delimiters, _offset);
+        if (i == Str::npos)
+        {
+            _offset = _string.length();
+            return false;
+        }
+
+        size_t j = _string.find_first_of(delimiters, i);
+        if (j == Str::npos) {
+            _token = _string.substr(i);
+            _offset = _string.length();
+            return true;
+        }
+
+        _token = _string.substr(i, j - i);
+        _offset = j;
+        return true;
+    }
+
+    const Str get_token() const
+    {
+        return _token;
+    }
+
+    void reset()
+    {
+        _offset = 0;
+    }
+
+    size_t _offset;
+    const Str _string;
+    Str _token;
+    Str _delimiters;
+};
+
+template <typename Str>
+struct string_utility
+{
+
+    static Str to_upper(const Str& str)
+    {
+        Str temp(str);
+        std::transform(temp.begin(), temp.end(), temp.begin(), ::toupper);
+        return temp;
+    }
+
+    static Str to_lower(const Str& str)
+    {
+        Str temp(str);
+        std::transform(temp.begin(), temp.end(), temp.begin(), ::tolower);
+        return temp;
+    }
+
+    static Str trim_left(const Str& str)
+    {
+        Str temp(str);
+        auto it = temp.begin();
+        for (it = temp.begin(); it != temp.end(); it++) {
+            if (!isspace(*it)) {
+                break;
+            }
+        }
+        if (it == temp.end()) {
+            temp.clear();
+        } else {
+            temp.erase(temp.begin(), it);
+        }
+        return temp;
+    }
+
+    static Str trim_right(const Str& str)
+    {
+        Str temp(str);
+        for (auto it = temp.end() - 1; ;it--) {
+            if (!isspace(*it)) {
+                temp.erase(it + 1, temp.end());
+                break;
+            }
+            if (it == temp.begin()) {
+                temp.clear();
+                break;
+            }
+        }
+        return temp;
+    }
+
+    static Str trim(const Str& str)
+    {
+        Str temp = trim_left(str);
+        return trim_right(temp);
+    }
+
+    static bool starts_with(Str const & value, Str const & starting)
+    {
+        if (starting.size() > value.size()) return false;
+        return std::equal(starting.begin(), starting.end(), value.begin());
+    }
+
+    static bool ends_with(Str const & value, Str const & ending)
+    {
+        if (ending.size() > value.size()) return false;
+        return std::equal(ending.rbegin(), ending.rend(), value.rbegin());
+    }
+
+    static bool equals_ignore_case(const Str& str1, const Str& str2)
+    {
+        return to_lower(str1) == to_lower(str2);
+    }
+
+    static bool istarts_with(Str const & value, Str const & starting)
+    {
+        if (starting.size() > value.size()) return false;
+        Str temp = value.substr(0, starting.size());
+        return equals_ignore_case(to_lower(starting), to_lower(temp));
+    }
+
+    static bool iends_with(Str const & value, Str const & ending)
+    {
+        if (ending.size() > value.size()) return false;
+        Str temp = value.substr(value.size() - ending.size(), ending.size());
+        return equals_ignore_case(to_lower(ending), to_lower(temp));
+    }
+
+    template <typename T>
+    static T from_string(const Str& str)
+    {
+        T obj;
+        std::basic_istringstream<typename Str::value_type> temp(str);
+        temp >> obj;
+        return obj;
+    }
+
+    static bool from_string(const Str& str)
+    {
+        bool obj;
+        std::basic_istringstream<typename Str::value_type> temp(str);
+        temp >> std::boolalpha >> obj;
+        return obj;
+    }
+
+    template <typename T>
+    static T from_hex_string(const Str& str)
+    {
+        T obj;
+        std::basic_istringstream<typename Str::value_type> temp(str);
+        temp >> std::hex >> obj;
+        return obj;
+    }
+
+    template <typename T>
+    static Str to_string(const T& var)
+    {
+        std::basic_ostringstream<typename Str::value_type> temp;
+        temp << var;
+        return temp.str();
+    }
+
+    static Str to_string(const bool& var)
+    {
+        std::basic_ostringstream<typename Str::value_type> temp;
+        temp << std::boolalpha << var;
+        return temp.str();
+    }
+
+    template <typename T>
+    static Str to_hex_string(const T& var, int width)
+    {
+        std::basic_ostringstream<typename Str::value_type> temp;
+        temp << std::hex;
+        if(width > 0)
+        {
+            temp << std::setw(width) << std::setfill<typename Str::value_type>('0');
+        }
+        temp << var;
+        return temp.str();
+    }
+
+    static std::vector<Str> split(Str const& str, Str const& delimiters)
+    {
+        std::vector<Str> ss;
+        tokenizer<Str> token(str, delimiters);
+        while (token.next_token())
+        {
+            ss.push_back(token.get_token());
+        }
+        return ss;
+    }
+};
+
+typedef string_utility<std::string> string_utility_a;
+typedef string_utility<std::wstring> string_utility_w;
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/timer.hpp b/projects/llm_framework/main_cosy_voice/src/runner/utils/timer.hpp
new file mode 100644
index 00000000..2bdf971e
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/timer.hpp
@@ -0,0 +1,57 @@
+/*
+ * AXERA is pleased to support the open source community by making ax-samples available.
+ *
+ * Copyright (c) 2022, AXERA Semiconductor (Shanghai) Co., Ltd. All rights reserved.
+ *
+ * Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+ * in compliance with the License. You may obtain a copy of the License at
+ *
+ * https://opensource.org/licenses/BSD-3-Clause
+ *
+ * Unless required by applicable law or agreed to in writing, software distributed
+ * under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+ * CONDITIONS OF ANY KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations under the License.
+ */
+
+/*
+ * Author: ls.wang
+ */
+
+#pragma once
+
+#include <chrono>
+
+class timer
+{
+private:
+    std::chrono::system_clock::time_point start_time, end_time;
+
+public:
+    timer()
+    {
+        start();
+    }
+
+    void start()
+    {
+        stop();
+        this->start_time = this->end_time;
+    }
+
+    void stop()
+    {
+#ifdef _MSC_VER
+        this->end_time = std::chrono::system_clock::now();
+#else
+        this->end_time = std::chrono::high_resolution_clock::now();
+#endif
+    }
+
+    float cost()
+    {
+        this->stop();
+        auto ms = std::chrono::duration_cast<std::chrono::microseconds>(this->end_time - this->start_time).count();
+        return static_cast<float>(ms) / 1000.f;
+    }
+};
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/utils.hpp b/projects/llm_framework/main_cosy_voice/src/runner/utils/utils.hpp
new file mode 100644
index 00000000..58e9bb0d
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/utils.hpp
@@ -0,0 +1,97 @@
+#pragma once
+
+#include <fstream>
+#include <vector>
+#include <iomanip>
+#include <type_traits> // 用于类型检查
+#include <iostream>
+#include <fstream>
+#include <vector>
+#include <string>
+#include <sstream>
+#include <type_traits>
+#include <stdexcept> // 用于异常处理
+
+// 函数模板：支持任意元素类型的vector
+template <typename T>
+void savetxt(const std::string& filename, 
+            const std::vector<T>& data,
+            char delimiter = ' ', 
+            int precision = 6)  // 默认精度调整为6位
+{
+    std::ofstream outfile(filename);
+    if (!outfile.is_open()) {
+        throw std::runtime_error("无法打开文件: " + filename);
+    }
+
+    // 保存流的原始格式状态
+    std::ios_base::fmtflags original_flags = outfile.flags();
+
+    // 仅对浮点类型启用科学计数法
+    if constexpr (std::is_floating_point_v<T>) {
+        outfile << std::scientific << std::setprecision(precision);
+    }
+
+    // 优化输出：先输出第一个元素，再循环输出后续元素
+    if (!data.empty()) {
+        outfile << data[0];
+        for (size_t i = 1; i < data.size(); ++i) {
+            outfile << delimiter << data[i];
+        }
+    }
+
+    // 恢复流的原始格式状态
+    outfile.flags(original_flags);
+    outfile.close();
+}
+
+
+
+// 函数模板：读取文本文件到 vector<T>
+template <typename T>
+int readtxt(const std::string& filename, std::vector<T>& data) {
+    std::ifstream file(filename);
+    if (!file.is_open()) {
+        std::cerr << "错误：无法打开文件 " << filename << std::endl;
+        return -1;
+    }
+
+    
+    std::string line;
+    while (std::getline(file, line)) {
+        // 跳过空行
+        if (line.empty()) continue;
+
+        std::istringstream iss(line);
+        T value;
+        // 解析行中的每个数值
+        while (iss >> value) {
+            data.push_back(value);
+            // 跳过分隔符（兼容逗号、分号等）
+            if (iss.peek() == ',' || iss.peek() == ';') iss.ignore();
+        }
+    }
+    file.close();
+    return 0;
+}
+
+std::vector<float> linspace(float start, float end, std::size_t num_steps) {
+    if (num_steps == 0) {
+        return {}; // 返回空向量
+    }
+    if (num_steps == 1) {
+        return {start}; // 如果只需要一个点，返回起始值
+    }
+
+    std::vector<float> result(num_steps);
+    float step_size = (end - start) / (num_steps - 1); // 计算步长
+
+    for (std::size_t i = 0; i < num_steps; ++i) {
+        result[i] = start + i * step_size;
+    }
+
+    // 确保最后一个值精确等于 end，避免浮点数精度带来的误差
+    result.back() = end;
+
+    return result;
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/wav.hpp b/projects/llm_framework/main_cosy_voice/src/runner/utils/wav.hpp
new file mode 100644
index 00000000..dbbd979b
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/wav.hpp
@@ -0,0 +1,160 @@
+#include <vector>
+#include <fstream>
+#include <iostream>
+#include <cstdint>
+#include <cstring> // for memcpy
+#include <algorithm> // for std::clamp (C++17)
+
+// 将 32 位整数以小端序写入文件
+void write_little_endian_32(std::ofstream& file, uint32_t value) {
+    file.put(static_cast<char>(value & 0xFF));
+    file.put(static_cast<char>((value >> 8) & 0xFF));
+    file.put(static_cast<char>((value >> 16) & 0xFF));
+    file.put(static_cast<char>((value >> 24) & 0xFF));
+}
+
+// 将 16 位整数以小端序写入文件
+void write_little_endian_16(std::ofstream& file, uint16_t value) {
+    file.put(static_cast<char>(value & 0xFF));
+    file.put(static_cast<char>((value >> 8) & 0xFF));
+}
+
+// 将 32 位浮点数以小端序写入文件 (IEEE 754)
+void write_little_endian_float32(std::ofstream& file, float value) {
+    // IEEE 754 float 在大多数系统上已经是小端序
+    // 但为了确保跨平台兼容性，我们显式处理字节序
+    static_assert(sizeof(float) == 4, "Float must be 32 bits");
+    char bytes[4];
+    std::memcpy(bytes, &value, 4);
+    // bytes[0] 是最低有效字节 (LSB)
+    file.put(bytes[0]);
+    file.put(bytes[1]);
+    file.put(bytes[2]);
+    file.put(bytes[3]);
+}
+
+// 将 float 范围 [-1.0, 1.0] 转换为 16-bit signed integer [-32768, 32767]
+int16_t float_to_int16(float sample) {
+    // 确保输入在有效范围内
+    // C++17 std::clamp, 或手动实现
+    // sample = std::clamp(sample, -1.0f, 1.0f);
+    if (sample > 1.0f) sample = 1.0f;
+    if (sample < -1.0f) sample = -1.0f;
+    
+    // 转换
+    return static_cast<int16_t>(sample * 32767.0f);
+}
+
+
+bool saveVectorAsWavFloat(const std::vector<float>& audio_data, const std::string& filename, int sample_rate, int channels) {
+    if (audio_data.empty() || channels <= 0 || sample_rate <= 0) {
+        std::cerr << "Invalid input parameters." << std::endl;
+        return false;
+    }
+
+    std::ofstream file(filename, std::ios::binary);
+    if (!file.is_open()) {
+        std::cerr << "Could not open file " << filename << " for writing." << std::endl;
+        return false;
+    }
+
+    // --- 计算关键参数 ---
+    uint32_t num_samples = static_cast<uint32_t>(audio_data.size());
+    uint16_t bits_per_sample = 32; // IEEE Float
+    uint32_t byte_rate = sample_rate * channels * (bits_per_sample / 8);
+    uint16_t block_align = channels * (bits_per_sample / 8);
+    uint32_t data_chunk_size = num_samples * (bits_per_sample / 8);
+    uint32_t riff_chunk_size = 4 + (8 + 16) + (8 + data_chunk_size); // fmt chunk is 24 bytes including header
+
+    // --- 写入 RIFF Header ---
+    file.write("RIFF", 4);
+    write_little_endian_32(file, riff_chunk_size);
+    file.write("WAVE", 4);
+
+    // --- 写入 fmt Subchunk ---
+    file.write("fmt ", 4);
+    write_little_endian_32(file, 16); // Subchunk1Size for PCM
+    write_little_endian_16(file, 3); // AudioFormat: 3 = IEEE Float
+    write_little_endian_16(file, static_cast<uint16_t>(channels));
+    write_little_endian_32(file, sample_rate);
+    write_little_endian_32(file, byte_rate);
+    write_little_endian_16(file, block_align);
+    write_little_endian_16(file, bits_per_sample);
+
+    // --- 写入 data Subchunk ---
+    file.write("data", 4);
+    write_little_endian_32(file, data_chunk_size);
+
+    // --- 写入音频数据 ---
+    for (const float& sample : audio_data) {
+        write_little_endian_float32(file, sample);
+    }
+
+    if (file.fail()) {
+        std::cerr << "Error occurred while writing audio data." << std::endl;
+        file.close();
+        return false;
+    }
+
+    file.close();
+    std::cout << "Successfully saved audio to " << filename << " (32-bit Float PCM)." << std::endl;
+    return true;
+}
+
+bool saveVectorAsWavInt16(const std::vector<float>& audio_data, const std::string& filename, int sample_rate, int channels) {
+    if (audio_data.empty() || channels <= 0 || sample_rate <= 0) {
+        std::cerr << "Invalid input parameters." << std::endl;
+        return false;
+    }
+
+    std::ofstream file(filename, std::ios::binary);
+    if (!file.is_open()) {
+        std::cerr << "Could not open file " << filename << " for writing." << std::endl;
+        return false;
+    }
+
+    // --- 计算关键参数 ---
+    uint32_t num_samples = static_cast<uint32_t>(audio_data.size());
+    uint16_t bits_per_sample = 16; // 16-bit signed integer
+    uint32_t byte_rate = sample_rate * channels * (bits_per_sample / 8);
+    uint16_t block_align = channels * (bits_per_sample / 8);
+    uint32_t data_chunk_size = num_samples * (bits_per_sample / 8);
+    uint32_t riff_chunk_size = 4 + (8 + 16) + (8 + data_chunk_size); // fmt chunk is 24 bytes including header
+
+    // --- 写入 RIFF Header ---
+    file.write("RIFF", 4);
+    write_little_endian_32(file, riff_chunk_size);
+    file.write("WAVE", 4);
+
+    // --- 写入 fmt Subchunk ---
+    file.write("fmt ", 4);
+    write_little_endian_32(file, 16); // Subchunk1Size for PCM
+    write_little_endian_16(file, 1); // AudioFormat: 1 = PCM (Integer)
+    write_little_endian_16(file, static_cast<uint16_t>(channels));
+    write_little_endian_32(file, sample_rate);
+    write_little_endian_32(file, byte_rate);
+    write_little_endian_16(file, block_align);
+    write_little_endian_16(file, bits_per_sample);
+
+    // --- 写入 data Subchunk ---
+    file.write("data", 4);
+    write_little_endian_32(file, data_chunk_size);
+
+    // --- 写入音频数据 ---
+    for (const float& sample : audio_data) {
+        int16_t int_sample = float_to_int16(sample);
+        write_little_endian_16(file, static_cast<uint16_t>(int_sample)); // int16_t -> uint16_t for raw bytes
+    }
+
+    if (file.fail()) {
+        std::cerr << "Error occurred while writing audio data." << std::endl;
+        file.close();
+        return false;
+    }
+
+    file.close();
+    std::cout << "Successfully saved audio to " << filename << " (16-bit Integer PCM)." << std::endl;
+    return true;
+}
+
+

From 2d064fdfa68b26732d81fd649566ca2bd6bb0cc0 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Mon, 15 Sep 2025 18:53:40 +0800
Subject: [PATCH 41/79] [update] update cosy_voice

---
 .../main_cosy_voice/src/main.cpp              | 95 +++++++++----------
 1 file changed, 46 insertions(+), 49 deletions(-)

diff --git a/projects/llm_framework/main_cosy_voice/src/main.cpp b/projects/llm_framework/main_cosy_voice/src/main.cpp
index a3abbe1d..121eb2da 100644
--- a/projects/llm_framework/main_cosy_voice/src/main.cpp
+++ b/projects/llm_framework/main_cosy_voice/src/main.cpp
@@ -102,7 +102,6 @@ class llm_task {
             model_           = config_body.at("model");
             response_format_ = config_body.at("response_format");
             enoutput_        = config_body.at("enoutput");
-            prompt_          = config_body.at("prompt");
 
             if (config_body.contains("input")) {
                 if (config_body["input"].is_string()) {
@@ -327,7 +326,6 @@ class llm_task {
                             g_llm_finished);
             };
 
-            // Start the LLM in a separate thread
             std::thread llm_thread(llm_thread_func);
 
             int token_offset     = 0;
@@ -336,8 +334,7 @@ class llm_task {
                 SLOGE("Error, prompt speech token len %d < 75", prompt_token_len);
                 return -1;
             }
-            // int prompt_token_align_len = int(prompt_token_len / lToken2Wav.token_hop_len) * lToken2Wav.token_hop_len;
-            int prompt_token_align_len = 75;  // only support 75 now
+            int prompt_token_align_len = 75;
 
             std::vector<float> prompt_speech_embeds_flow1;
             prompt_speech_embeds_flow1.insert(prompt_speech_embeds_flow1.begin(), prompt_speech_embeds_flow.begin(),
@@ -381,28 +378,31 @@ class llm_task {
                     token_offset += this_token_hop_len;
                     output.insert(output.end(), speech.begin(), speech.end());
                     std::string path = "output_" + std::to_string(i) + ".wav";
+                    std::vector<int16_t> wav_pcm_data;
+                    wav_pcm_data.resize(speech.size());
+                    for (size_t k = 0; k < speech.size(); ++k) {
+                        float sample = speech[k];
+                        if (sample > 1.0f) sample = 1.0f;
+                        if (sample < -1.0f) sample = -1.0f;
+                        wav_pcm_data[k] = static_cast<int16_t>(sample * 32767.0f);
+                    }
+                    if (out_callback_) {
+                        out_callback_(std::string(reinterpret_cast<char *>(wav_pcm_data.data()),
+                                                  wav_pcm_data.size() * sizeof(int16_t)),
+                                      false);
+                    }
 
                     saveVectorAsWavFloat(speech, path, 24000, 1);
                     i += 1;
-
-                }
-
-                else if (g_llm_finished.load()) {
+                } else if (g_llm_finished.load()) {
                     std::cout << "[Main/Token2Wav Thread] Buffer is empty and LLM finished. Exiting.\n";
                     lock.unlock();
                     break;
-                }
-                // Check exit condition: Buffer is empty and LLM is done
-                else {
-                    // This else branch is technically not needed because the wait condition
-                    // ensures we only get here if one of the conditions is true.
-                    // But it's good practice to structure logic clearly.
-                    // In this specific loop, we will always process if we wake up.
-                    lock.unlock();  // Make sure to unlock if not processing
+                } else {
+                    lock.unlock();
                 }
             }
 
-            // Wait for the LLM thread to finish
             if (llm_thread.joinable()) {
                 llm_thread.join();
             }
@@ -419,11 +419,23 @@ class llm_task {
             token.insert(token.end(), g_token_buffer.begin() + start, g_token_buffer.end());
             auto speech = lToken2Wav.infer(token, prompt_speech_embeds_flow1, prompt_feat1, spk_embeds,
                                            token_offset - start, true);
-            // TODO: 另起一个线程处理生成的音频
             output.insert(output.end(), speech.begin(), speech.end());
             std::string path = "output_" + std::to_string(i) + ".wav";
             saveVectorAsWavFloat(speech, path, 24000, 1);
             saveVectorAsWavFloat(output, "output.wav", 24000, 1);
+            std::vector<int16_t> wav_pcm_data;
+            wav_pcm_data.resize(speech.size());
+            for (size_t k = 0; k < speech.size(); ++k) {
+                float sample = speech[k];
+                if (sample > 1.0f) sample = 1.0f;
+                if (sample < -1.0f) sample = -1.0f;
+                wav_pcm_data[k] = static_cast<int16_t>(sample * 32767.0f);
+            }
+            if (out_callback_) {
+                out_callback_(
+                    std::string(reinterpret_cast<char *>(wav_pcm_data.data()), wav_pcm_data.size() * sizeof(int16_t)),
+                    true);
+            }
 
             SLOGI("tts total use time: %.3f s", time_total.cost() / 1000);
             reset();
@@ -560,23 +572,26 @@ class llm_cosy_voice : public StackFlow {
         if (!(llm_task_obj && llm_channel)) {
             return;
         }
-        SLOGI("send:%s", data.c_str());
+        std::string base64_data;
+        if (!data.empty()) {
+            int len = encode_base64(data, base64_data);
+        }
         if (llm_channel->enstream_) {
             static int count = 0;
             nlohmann::json data_body;
             data_body["index"] = count++;
-            data_body["delta"] = data;
-            if (!finish)
-                data_body["delta"] = data;
+            if (!data.empty())
+                data_body["delta"] = base64_data;
             else
-                data_body["delta"] = std::string("");
+                data_body["delta"] = "";
             data_body["finish"] = finish;
             if (finish) count = 0;
-            SLOGI("send stream");
             llm_channel->send(llm_task_obj->response_format_, data_body, LLM_NO_ERROR);
         } else if (finish) {
-            SLOGI("send utf-8");
-            llm_channel->send(llm_task_obj->response_format_, data, LLM_NO_ERROR);
+            llm_channel->send(llm_task_obj->response_format_, base64_data, LLM_NO_ERROR);
+        }
+        if (llm_task_obj->response_format_.find("sys") != std::string::npos) {
+            unit_call("audio", "queue_play", data);
         }
     }
 
@@ -652,24 +667,6 @@ class llm_cosy_voice : public StackFlow {
         llm_task_obj->inference_async(sample_unescapeString(*next_data));
     }
 
-    void task_asr_data(const std::weak_ptr<llm_task> llm_task_obj_weak,
-                       const std::weak_ptr<llm_channel_obj> llm_channel_weak, const std::string &object,
-                       const std::string &data)
-    {
-        auto llm_task_obj = llm_task_obj_weak.lock();
-        auto llm_channel  = llm_channel_weak.lock();
-        if (!(llm_task_obj && llm_channel)) {
-            return;
-        }
-        if (object.find("stream") != std::string::npos) {
-            if (sample_json_str_get(data, "finish") == "true") {
-                llm_task_obj->inference_async(sample_json_str_get(data, "delta"));
-            }
-        } else {
-            llm_task_obj->inference_async(data);
-        }
-    }
-
     void kws_awake(const std::weak_ptr<llm_task> llm_task_obj_weak,
                    const std::weak_ptr<llm_channel_obj> llm_channel_weak, const std::string &object,
                    const std::string &data)
@@ -716,14 +713,14 @@ class llm_cosy_voice : public StackFlow {
                           std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1, std::placeholders::_2));
 
             for (const auto input : llm_task_obj->inputs_) {
-                if (input.find("llm") != std::string::npos) {
+                if (input.find("tts") != std::string::npos) {
                     llm_channel->subscriber_work_id(
                         "", std::bind(&llm_cosy_voice::task_user_data, this, std::weak_ptr<llm_task>(llm_task_obj),
                                       std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
                                       std::placeholders::_2));
-                } else if ((input.find("asr") != std::string::npos) || (input.find("whisper") != std::string::npos)) {
+                } else if ((input.find("llm") != std::string::npos) || (input.find("vlm") != std::string::npos)) {
                     llm_channel->subscriber_work_id(
-                        input, std::bind(&llm_cosy_voice::task_asr_data, this, std::weak_ptr<llm_task>(llm_task_obj),
+                        input, std::bind(&llm_cosy_voice::task_user_data, this, std::weak_ptr<llm_task>(llm_task_obj),
                                          std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
                                          std::placeholders::_2));
                 } else if (input.find("kws") != std::string::npos) {
@@ -760,10 +757,10 @@ class llm_cosy_voice : public StackFlow {
         }
         auto llm_channel  = get_channel(work_id);
         auto llm_task_obj = llm_task_[work_id_num];
-        if (data.find("asr") != std::string::npos) {
+        if (data.find("llm") != std::string::npos) {
             ret = llm_channel->subscriber_work_id(
                 data,
-                std::bind(&llm_cosy_voice::task_asr_data, this, std::weak_ptr<llm_task>(llm_task_obj),
+                std::bind(&llm_cosy_voice::task_user_data, this, std::weak_ptr<llm_task>(llm_task_obj),
                           std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1, std::placeholders::_2));
             llm_task_obj->inputs_.push_back(data);
         } else if (data.find("kws") != std::string::npos) {

From 52a09b6a63eafa16fa3e1931c8bfb82d8e9ac195 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Wed, 17 Sep 2025 17:15:52 +0800
Subject: [PATCH 42/79] [update] add new kws unit

---
 projects/llm_framework/main_kws_new/Kconfig   |   0
 .../llm_framework/main_kws_new/SConstruct     |  73 ++
 .../llm_framework/main_kws_new/src/main.cpp   | 652 ++++++++++++++++++
 3 files changed, 725 insertions(+)
 create mode 100644 projects/llm_framework/main_kws_new/Kconfig
 create mode 100644 projects/llm_framework/main_kws_new/SConstruct
 create mode 100644 projects/llm_framework/main_kws_new/src/main.cpp

diff --git a/projects/llm_framework/main_kws_new/Kconfig b/projects/llm_framework/main_kws_new/Kconfig
new file mode 100644
index 00000000..e69de29b
diff --git a/projects/llm_framework/main_kws_new/SConstruct b/projects/llm_framework/main_kws_new/SConstruct
new file mode 100644
index 00000000..f42b60be
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/SConstruct
@@ -0,0 +1,73 @@
+import os
+
+Import('env')
+with open(env['PROJECT_TOOL_S']) as f:
+    exec(f.read())
+
+SRCS = Glob('src/*.c*')
+INCLUDE = [ADir('include'), ADir('.')]
+PRIVATE_INCLUDE = []
+REQUIREMENTS = ['pthread', 'dl', 'utilities', 'eventpp', 'StackFlow', 'single_header_libs']
+STATIC_LIB = []
+DYNAMIC_LIB = []
+DEFINITIONS = []
+DEFINITIONS_PRIVATE = []
+LDFLAGS = []
+LINK_SEARCH_PATH = []
+STATIC_FILES = []
+
+python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-kws-python-venv_v1.6.tar.gz", 'm5stack_llm-kws-python-venv_v1.6.tar.gz')
+
+DEFINITIONS += ['-std=c++17']
+LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
+LINK_SEARCH_PATH += [ADir('../static_lib')]
+
+INCLUDE += [ADir('../static_lib/include/sherpa'),
+            ADir('../static_lib/include/sherpa/sherpa-onnx'),
+            ADir('../static_lib/include/sherpa/sherpa-onnx/onnxruntime-src'),
+            ADir('../static_lib/include/sherpa/sherpa-onnx/openfst-src')
+            ]
+
+LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/onnx')]
+LDFLAGS += ['-l:libcargs.a',
+            '-l:libsherpa-onnx-core.a', '-l:libkaldi-native-fbank-core.a',
+            '-l:libkaldi-decoder-core.a', '-l:libssentencepiece_core.a']
+
+REQUIREMENTS += ['onnxruntime']
+
+# STATIC_FILES += [os.path.join(python_venv, 'sherpa-onnx')]
+# STATIC_FILES += Glob('llm-kws_text2token.py')
+# STATIC_FILES += Glob('mode_*.json')
+
+IGNORE_FILES = []
+IGNORE_FILES += ['sherpa-onnx']
+
+import json
+if not os.path.exists('../dist'):
+    os.makedirs('../dist')
+ignore = {'ignore':[]}
+try:
+    with open('../dist/fileignore', 'a+') as f:
+        f.seek(0)
+        ignore = json.load(f)
+except:
+    pass
+ignore['ignore'] += IGNORE_FILES
+ignore['ignore'] = list(set(ignore['ignore']))
+with open('../dist/fileignore', 'w') as f:
+    json.dump(ignore, f, indent=4)
+
+env['COMPONENTS'].append({'target':'llm_kws_new-1.9',
+                          'SRCS':SRCS,
+                          'INCLUDE':INCLUDE,
+                          'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
+                          'REQUIREMENTS':REQUIREMENTS,
+                          'STATIC_LIB':STATIC_LIB,
+                          'DYNAMIC_LIB':DYNAMIC_LIB,
+                          'DEFINITIONS':DEFINITIONS,
+                          'DEFINITIONS_PRIVATE':DEFINITIONS_PRIVATE,
+                          'LDFLAGS':LDFLAGS,
+                          'LINK_SEARCH_PATH':LINK_SEARCH_PATH,
+                          'STATIC_FILES':STATIC_FILES,
+                          'REGISTER':'project'
+                          })
diff --git a/projects/llm_framework/main_kws_new/src/main.cpp b/projects/llm_framework/main_kws_new/src/main.cpp
new file mode 100644
index 00000000..cd824290
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/main.cpp
@@ -0,0 +1,652 @@
+/*
+ * SPDX-FileCopyrightText: 2024 M5Stack Technology CO LTD
+ *
+ * SPDX-License-Identifier: MIT
+ */
+#include "StackFlow.h"
+
+#include <signal.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+#include <base64.h>
+#include <fstream>
+#include <stdexcept>
+#include <thread_safe_list.h>
+#include "../../../../SDK/components/utilities/include/sample_log.h"
+
+#define BUFFER_IMPLEMENTATION
+#include <stdbool.h>
+#include <stdint.h>
+#include "libs/buffer.h"
+
+using namespace StackFlows;
+
+int main_exit_flage = 0;
+static void __sigint(int iSigNo)
+{
+    SLOGW("llm_kws_new will be exit!");
+    main_exit_flage = 1;
+}
+
+static std::string base_model_path_;
+static std::string base_model_config_path_;
+
+typedef std::function<void(const std::string &data, bool finish)> task_callback_t;
+
+#define CONFIG_AUTO_SET(obj, key)             \
+    if (config_body.contains(#key))           \
+        mode_config_.key = config_body[#key]; \
+    else if (obj.contains(#key))              \
+        mode_config_.key = obj[#key];
+
+class llm_task {
+private:
+    int chunk_size                 = 32;
+    float threshold                = 0.9f;
+    int min_continuous_frames      = 5;
+    int count_frames               = 0;
+    long long last_trigger_time_ms = -1e9;
+    long long frame_index_global   = 0;
+    int REFRACTORY_TIME_MS         = 2000;
+
+public:
+    std::string model_;
+    std::string response_format_;
+    std::vector<std::string> inputs_;
+    std::vector<std::string> kws_;
+    bool enoutput_;
+    bool enoutput_json_;
+    bool enstream_;
+    bool enwake_audio_;
+    std::atomic_bool audio_flage_;
+    task_callback_t out_callback_;
+    int delay_audio_frame_ = 10;
+    buffer_t *pcmdata;
+    std::string wake_wav_file_;
+
+    std::function<void(const std::string &)> play_awake_wav;
+
+    bool parse_config(const nlohmann::json &config_body)
+    {
+        try {
+            model_           = config_body.at("model");
+            response_format_ = config_body.at("response_format");
+            enoutput_        = config_body.at("enoutput");
+            if (config_body.contains("enwake_audio")) {
+                enwake_audio_ = config_body["enwake_audio"];
+            } else {
+                enwake_audio_ = true;
+            }
+            if (config_body.contains("input")) {
+                if (config_body["input"].is_string()) {
+                    inputs_.push_back(config_body["input"].get<std::string>());
+                } else if (config_body["input"].is_array()) {
+                    for (auto _in : config_body["input"]) {
+                        inputs_.push_back(_in.get<std::string>());
+                    }
+                }
+            }
+            if (config_body.contains("kws")) {
+                if (config_body["kws"].is_string()) {
+                    kws_.push_back(config_body["kws"].get<std::string>());
+                } else if (config_body["kws"].is_array()) {
+                    for (auto _in : config_body["kws"]) {
+                        kws_.push_back(_in.get<std::string>());
+                    }
+                }
+            }
+            enoutput_json_ = response_format_.find("json") == std::string::npos ? false : true;
+        } catch (...) {
+            SLOGE("setup config_body error");
+            return true;
+        }
+        enstream_ = response_format_.find("stream") == std::string::npos ? false : true;
+        return false;
+    }
+
+    int load_model(const nlohmann::json &config_body)
+    {
+        if (parse_config(config_body)) {
+            return -1;
+        }
+
+        nlohmann::json file_body;
+        std::list<std::string> config_file_paths =
+            get_config_file_paths(base_model_path_, base_model_config_path_, model_);
+        try {
+            for (auto file_name : config_file_paths) {
+                std::ifstream config_file(file_name);
+                if (!config_file.is_open()) {
+                    SLOGW("config file :%s miss", file_name.c_str());
+                    continue;
+                }
+                SLOGI("config file :%s read", file_name.c_str());
+                config_file >> file_body;
+                config_file.close();
+                break;
+            }
+            if (file_body.empty()) {
+                SLOGE("all config file miss");
+                return -2;
+            }
+            std::string base_model = base_model_path_ + model_ + "/";
+            SLOGI("base_model %s", base_model.c_str());
+
+            if (config_body.contains("wake_wav_file"))
+                wake_wav_file_ = config_body["wake_wav_file"];
+            else if (file_body["mode_param"].contains("wake_wav_file"))
+                wake_wav_file_ = file_body["mode_param"]["wake_wav_file"];
+
+        } catch (...) {
+            SLOGE("config file read false");
+            return -3;
+        }
+        return 0;
+    }
+
+    void set_output(task_callback_t out_callback)
+    {
+        out_callback_ = out_callback;
+    }
+
+    bool detect_wakeup(const std::vector<float> &scores)
+    {
+        bool triggered = false;
+        for (auto score : scores) {
+            if (score > threshold) {
+                count_frames++;
+                if (count_frames >= min_continuous_frames) {
+                    long long trigger_time_ms = (frame_index_global - min_continuous_frames + 1) * 10;
+                    if (trigger_time_ms - last_trigger_time_ms >= REFRACTORY_TIME_MS) {
+                        std::cout << "[唤醒触发] 时间: " << trigger_time_ms << " ms\n";
+                        last_trigger_time_ms = trigger_time_ms;
+                        triggered            = true;
+                    }
+                }
+            } else {
+                count_frames = 0;
+            }
+            frame_index_global++;
+        }
+        return triggered;
+    }
+
+    kaldi::Matrix<float> compute_fbank(const std::vector<float> &audio16k)
+    {
+        kaldi::FbankOptions fbank_opts;
+        fbank_opts.mel_opts.num_bins     = feat_dim;
+        fbank_opts.frame_opts.samp_freq  = resample_rate;
+        fbank_opts.frame_opts.dither     = 0.0;
+        fbank_opts.energy_floor          = 0.0;
+        fbank_opts.mel_opts.htk_mode     = false;
+        fbank_opts.use_energy            = false;
+        fbank_opts.frame_opts.snip_edges = false;
+
+        kaldi::Fbank fbank(fbank_opts);
+
+        kaldi::Vector<float> wav(audio16k.size());
+        for (size_t i = 0; i < audio16k.size(); i++) {
+            wav(i) = audio16k[i] * (1 << 15);
+        }
+
+        kaldi::Matrix<float> features;
+        fbank.ComputeFeatures(wav, resample_rate, 1.0, &features);
+        return features;  // Shape: [num_frames, feat_dim]
+    }
+
+    // std::vector<float> run_inference(const std::vector<float> &audio16k)
+    // {
+    //     auto feats = compute_fbank(audio16k);
+
+    //     std::array<int64_t, 3> input_shape{1, feats.NumRows(), feats.NumCols()};
+    //     Ort::Value input_tensor = Ort::Value::CreateTensor<float>(
+    //         memInfo, (float *)feats.Data(), feats.NumRows() * feats.NumCols(), input_shape.data(),
+    //         input_shape.size());
+
+    //     // cache 作为输入
+    //     std::array<int64_t, 4> cache_shape{1, 32, 88, 1};
+    //     Ort::Value cache_tensor = Ort::Value::CreateTensor<float>(
+    //         memInfo, (float *)cache.Data(), cache.NumRows() * cache.NumCols(), cache_shape.data(),
+    //         cache_shape.size());
+
+    //     const char *input_names[]  = {"input", "cache"};
+    //     const char *output_names[] = {"output", "cache_out"};
+
+    //     auto outputs = ort_session->Run(Ort::RunOptions{nullptr}, input_names, &input_tensor,
+    //                                     1,  // TODO: pass cache_tensor too (update input count)
+    //                                     output_names, 2);
+
+    //     auto &out_chunk_tensor = outputs[0];
+    //     float *out_data        = out_chunk_tensor.GetTensorMutableData<float>();
+    //     size_t out_len         = out_chunk_tensor.GetTensorTypeAndShapeInfo().GetElementCount();
+
+    //     return std::vector<float>(out_data, out_data + out_len);
+    // }
+
+    void sys_pcm_on_data(const std::string &raw)
+    {
+        static int count = 0;
+        if (count < delay_audio_frame_) {
+            buffer_write_char(pcmdata, raw.data(), raw.length());
+            count++;
+            return;
+        }
+        buffer_write_char(pcmdata, raw.data(), raw.length());
+        buffer_position_set(pcmdata, 0);
+
+        std::vector<float> floatSamples;
+        {
+            int16_t audio_val;
+            while (buffer_read_i16(pcmdata, &audio_val, 1)) {
+                float normalizedSample = static_cast<float>(audio_val) / INT16_MAX;
+                floatSamples.push_back(normalizedSample);
+            }
+        }
+
+        buffer_resize(pcmdata, 0);
+        count = 0;
+    }
+
+    void trigger()
+    {
+        if (out_callback_) out_callback_("", true);
+    }
+
+    bool delete_model()
+    {
+        return true;
+    }
+
+    llm_task(const std::string &workid) : audio_flage_(false)
+    {
+        pcmdata = buffer_create();
+    }
+
+    void start()
+    {
+    }
+
+    void stop()
+    {
+    }
+
+    ~llm_task()
+    {
+        stop();
+        buffer_destroy(pcmdata);
+    }
+};
+#undef CONFIG_AUTO_SET
+
+class llm_kws : public StackFlow {
+private:
+    enum { EVENT_TRIGGER = EVENT_EXPORT + 1 };
+    int task_count_;
+    std::string audio_url_;
+    std::unordered_map<int, std::shared_ptr<llm_task>> llm_task_;
+
+public:
+    llm_kws() : StackFlow("kws_new")
+    {
+        task_count_ = 1;
+        event_queue_.appendListener(EVENT_TRIGGER, std::bind(&llm_kws::trigger, this, std::placeholders::_1));
+        rpc_ctx_->register_rpc_action(
+            "trigger", [this](pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) -> std::string {
+                this->event_queue_.enqueue(EVENT_TRIGGER,
+                                           std::make_shared<stackflow_data>(data->get_param(0), data->get_param(1)));
+                return LLM_NONE;
+            });
+    }
+
+    void task_output(const std::weak_ptr<llm_task> llm_task_obj_weak,
+                     const std::weak_ptr<llm_channel_obj> llm_channel_weak, const std::string &data, bool finish)
+    {
+        auto llm_task_obj = llm_task_obj_weak.lock();
+        auto llm_channel  = llm_channel_weak.lock();
+        if (!(llm_task_obj && llm_channel)) {
+            return;
+        }
+        std::string tmp_msg1;
+        const std::string *next_data = &data;
+        if (data.empty()) {
+            llm_channel->send(llm_task_obj->response_format_, true, LLM_NO_ERROR);
+            return;
+        }
+        if (finish) {
+            tmp_msg1  = data + ".";
+            next_data = &tmp_msg1;
+        }
+        if (llm_channel->enstream_) {
+            static int count = 0;
+            nlohmann::json data_body;
+            data_body["index"]  = count++;
+            data_body["delta"]  = (*next_data);
+            data_body["finish"] = finish;
+            if (finish) count = 0;
+            SLOGI("send stream:%s", next_data->c_str());
+            llm_channel->send(llm_task_obj->response_format_, data_body, LLM_NO_ERROR);
+        } else if (finish) {
+            SLOGI("send utf-8:%s", next_data->c_str());
+            llm_channel->send(llm_task_obj->response_format_, (*next_data), LLM_NO_ERROR);
+        }
+    }
+
+    void play_awake_wav(const std::string &wav_file)
+    {
+        FILE *fp = fopen(wav_file.c_str(), "rb");
+        if (!fp) {
+            printf("Open %s failed!\n", wav_file.c_str());
+            return;
+        }
+        fseek(fp, 0, SEEK_END);
+        long size = ftell(fp);
+        fseek(fp, 0, SEEK_SET);
+        std::vector<char> wav_data(size);
+        fread(wav_data.data(), 1, wav_data.size(), fp);
+        fclose(fp);
+        int post = 0;
+        for (int i = 0; i < wav_data.size() - 4; i++) {
+            if ((wav_data[i] == 'd') && (wav_data[i + 1] == 'a') && (wav_data[i + 2] == 't') &&
+                (wav_data[i + 3] == 'a')) {
+                post = i + 8;
+                break;
+            }
+        }
+        if (post != 0) {
+            unit_call("audio", "play_raw", std::string((char *)(wav_data.data() + post), size - post));
+        }
+    }
+
+    void task_pause(const std::weak_ptr<llm_task> llm_task_obj_weak,
+                    const std::weak_ptr<llm_channel_obj> llm_channel_weak)
+    {
+        auto llm_task_obj = llm_task_obj_weak.lock();
+        auto llm_channel  = llm_channel_weak.lock();
+        if (!(llm_task_obj && llm_channel)) {
+            return;
+        }
+        if (llm_task_obj->audio_flage_) {
+            if (!audio_url_.empty()) llm_channel->stop_subscriber(audio_url_);
+            llm_task_obj->audio_flage_ = false;
+        }
+    }
+
+    void task_work(const std::weak_ptr<llm_task> llm_task_obj_weak,
+                   const std::weak_ptr<llm_channel_obj> llm_channel_weak)
+    {
+        auto llm_task_obj = llm_task_obj_weak.lock();
+        auto llm_channel  = llm_channel_weak.lock();
+        if (!(llm_task_obj && llm_channel)) {
+            return;
+        }
+        if ((!audio_url_.empty()) && (llm_task_obj->audio_flage_ == false)) {
+            std::weak_ptr<llm_task> _llm_task_obj = llm_task_obj;
+            llm_channel->subscriber(audio_url_, [_llm_task_obj](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
+                _llm_task_obj.lock()->sys_pcm_on_data(raw->string());
+            });
+            llm_task_obj->audio_flage_ = true;
+        }
+    }
+
+    void work(const std::string &work_id, const std::string &object, const std::string &data) override
+    {
+        SLOGI("llm_asr::work:%s", data.c_str());
+
+        nlohmann::json error_body;
+        int work_id_num = sample_get_work_id_num(work_id);
+        if (llm_task_.find(work_id_num) == llm_task_.end()) {
+            error_body["code"]    = -6;
+            error_body["message"] = "Unit Does Not Exist";
+            send("None", "None", error_body, work_id);
+            return;
+        }
+        task_work(llm_task_[work_id_num], get_channel(work_id_num));
+        send("None", "None", LLM_NO_ERROR, work_id);
+    }
+
+    void pause(const std::string &work_id, const std::string &object, const std::string &data) override
+    {
+        SLOGI("llm_asr::work:%s", data.c_str());
+
+        nlohmann::json error_body;
+        int work_id_num = sample_get_work_id_num(work_id);
+        if (llm_task_.find(work_id_num) == llm_task_.end()) {
+            error_body["code"]    = -6;
+            error_body["message"] = "Unit Does Not Exist";
+            send("None", "None", error_body, work_id);
+            return;
+        }
+        task_pause(llm_task_[work_id_num], get_channel(work_id_num));
+        send("None", "None", LLM_NO_ERROR, work_id);
+    }
+
+    void task_user_data(const std::weak_ptr<llm_task> llm_task_obj_weak,
+                        const std::weak_ptr<llm_channel_obj> llm_channel_weak, const std::string &object,
+                        const std::string &data)
+    {
+        nlohmann::json error_body;
+        auto llm_task_obj = llm_task_obj_weak.lock();
+        auto llm_channel  = llm_channel_weak.lock();
+        if (!(llm_task_obj && llm_channel)) {
+            error_body["code"]    = -11;
+            error_body["message"] = "Model run failed.";
+            send("None", "None", error_body, unit_name_);
+            return;
+        }
+        std::string tmp_msg1;
+        const std::string *next_data = &data;
+        int ret;
+        if (object.find("stream") != std::string::npos) {
+            static std::unordered_map<int, std::string> stream_buff;
+            try {
+                if (decode_stream(data, tmp_msg1, stream_buff)) {
+                    return;
+                };
+            } catch (...) {
+                stream_buff.clear();
+                error_body["code"]    = -25;
+                error_body["message"] = "Stream data index error.";
+                send("None", "None", error_body, unit_name_);
+                return;
+            }
+            next_data = &tmp_msg1;
+        }
+        std::string tmp_msg2;
+        if (object.find("base64") != std::string::npos) {
+            ret = decode_base64((*next_data), tmp_msg2);
+            if (ret == -1) {
+                error_body["code"]    = -23;
+                error_body["message"] = "Base64 decoding error.";
+                send("None", "None", error_body, unit_name_);
+                return;
+            }
+            next_data = &tmp_msg2;
+        }
+        llm_task_obj->sys_pcm_on_data((*next_data));
+    }
+
+    int setup(const std::string &work_id, const std::string &object, const std::string &data) override
+    {
+        nlohmann::json error_body;
+        if ((llm_task_channel_.size() - 1) == task_count_) {
+            error_body["code"]    = -21;
+            error_body["message"] = "task full";
+            send("None", "None", error_body, "kws");
+            return -1;
+        }
+
+        int work_id_num                             = sample_get_work_id_num(work_id);
+        auto llm_channel                            = get_channel(work_id);
+        auto llm_task_obj                           = std::make_shared<llm_task>(work_id);
+        std::weak_ptr<llm_task> _llm_task_obj       = llm_task_obj;
+        std::weak_ptr<llm_channel_obj> _llm_channel = llm_channel;
+        nlohmann::json config_body;
+        try {
+            config_body = nlohmann::json::parse(data);
+        } catch (...) {
+            SLOGE("setup json format error.");
+            error_body["code"]    = -2;
+            error_body["message"] = "json format error.";
+            send("None", "None", error_body, "kws");
+            return -2;
+        }
+        int ret = llm_task_obj->load_model(config_body);
+        if (ret == 0) {
+            llm_channel->set_output(llm_task_obj->enoutput_);
+            llm_channel->set_stream(llm_task_obj->enstream_);
+            llm_task_obj->play_awake_wav = std::bind(&llm_kws::play_awake_wav, this, std::placeholders::_1);
+            llm_task_obj->set_output(std::bind(&llm_kws::task_output, this, std::weak_ptr<llm_task>(llm_task_obj),
+                                               std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
+                                               std::placeholders::_2));
+
+            for (const auto input : llm_task_obj->inputs_) {
+                if (input.find("sys") != std::string::npos) {
+                    audio_url_ = unit_call("audio", "cap", "None");
+                    llm_channel->subscriber(audio_url_,
+                                            [_llm_task_obj](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
+                                                auto llm_task_obj = _llm_task_obj.lock();
+                                                if (llm_task_obj) llm_task_obj->sys_pcm_on_data(raw->string());
+                                            });
+                    llm_task_obj->audio_flage_ = true;
+                } else if (input.find("kws") != std::string::npos) {
+                    llm_task_obj->delay_audio_frame_ = 0;
+                    llm_channel->subscriber_work_id(
+                        "", std::bind(&llm_kws::task_user_data, this, std::weak_ptr<llm_task>(llm_task_obj),
+                                      std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
+                                      std::placeholders::_2));
+                }
+            }
+            llm_task_[work_id_num] = llm_task_obj;
+            SLOGI("load_mode success");
+            send("None", "None", LLM_NO_ERROR, work_id);
+            return 0;
+        } else {
+            SLOGE("load_mode Failed");
+            error_body["code"]    = -5;
+            error_body["message"] = "Model loading failed.";
+            send("None", "None", error_body, "kws");
+            return -1;
+        }
+    }
+
+    void taskinfo(const std::string &work_id, const std::string &object, const std::string &data) override
+    {
+        SLOGI("llm_kws::taskinfo:%s", data.c_str());
+        nlohmann::json req_body;
+        int work_id_num = sample_get_work_id_num(work_id);
+        if (WORK_ID_NONE == work_id_num) {
+            std::vector<std::string> task_list;
+            std::transform(llm_task_channel_.begin(), llm_task_channel_.end(), std::back_inserter(task_list),
+                           [](const auto task_channel) { return task_channel.second->work_id_; });
+            req_body = task_list;
+            send("kws.tasklist", req_body, LLM_NO_ERROR, work_id);
+        } else {
+            if (llm_task_.find(work_id_num) == llm_task_.end()) {
+                req_body["code"]    = -6;
+                req_body["message"] = "Unit Does Not Exist";
+                send("None", "None", req_body, work_id);
+                return;
+            }
+            auto llm_task_obj           = llm_task_[work_id_num];
+            req_body["model"]           = llm_task_obj->model_;
+            req_body["response_format"] = llm_task_obj->response_format_;
+            req_body["enoutput"]        = llm_task_obj->enoutput_;
+            req_body["inputs"]          = llm_task_obj->inputs_;
+            send("kws.taskinfo", req_body, LLM_NO_ERROR, work_id);
+        }
+    }
+
+    int exit(const std::string &work_id, const std::string &object, const std::string &data) override
+    {
+        SLOGI("llm_kws::exit:%s", data.c_str());
+        nlohmann::json error_body;
+        int work_id_num = sample_get_work_id_num(work_id);
+        if (llm_task_.find(work_id_num) == llm_task_.end()) {
+            error_body["code"]    = -6;
+            error_body["message"] = "Unit Does Not Exist";
+            send("None", "None", error_body, work_id);
+            return -1;
+        }
+        llm_task_[work_id_num]->stop();
+        auto llm_channel = get_channel(work_id_num);
+        llm_channel->stop_subscriber("");
+        if (llm_task_[work_id_num]->audio_flage_) {
+            unit_call("audio", "cap_stop", "None");
+        }
+        llm_task_.erase(work_id_num);
+        send("None", "None", LLM_NO_ERROR, work_id);
+        return 0;
+    }
+
+    std::string trigger(const std::shared_ptr<void> &arg)
+    {
+        std::shared_ptr<stackflow_data> originalPtr = std::static_pointer_cast<stackflow_data>(arg);
+        std::string zmq_url                         = originalPtr->string(0);
+        std::string data                            = originalPtr->string(1);
+        std::string work_id                         = sample_json_str_get(data, "work_id");
+        if (work_id.length() == 0) {
+            nlohmann::json out_body;
+            out_body["request_id"]       = sample_json_str_get(data, "request_id");
+            out_body["work_id"]          = "kws";
+            out_body["created"]          = time(NULL);
+            out_body["object"]           = "";
+            out_body["data"]             = "";
+            out_body["error"]["code"]    = -2;
+            out_body["error"]["message"] = "json format error.";
+            pzmq _zmq(zmq_url, ZMQ_PUSH);
+            std::string out = out_body.dump();
+            out += "\n";
+            _zmq.send_data(out);
+            return LLM_NONE;
+        }
+
+        int work_id_num = sample_get_work_id_num(work_id);
+        if (llm_task_.find(work_id_num) == llm_task_.end()) {
+            nlohmann::json out_body;
+            out_body["request_id"]       = sample_json_str_get(data, "request_id");
+            out_body["work_id"]          = "kws";
+            out_body["created"]          = time(NULL);
+            out_body["object"]           = "";
+            out_body["data"]             = "";
+            out_body["error"]["code"]    = -6;
+            out_body["error"]["message"] = "Unit Does Not Exist";
+            pzmq _zmq(zmq_url, ZMQ_PUSH);
+            std::string out = out_body.dump();
+            out += "\n";
+            _zmq.send_data(out);
+            return LLM_NONE;
+        }
+        llm_task_[work_id_num]->trigger();
+        return LLM_NONE;
+    }
+
+    ~llm_kws()
+    {
+        while (1) {
+            auto iteam = llm_task_.begin();
+            if (iteam == llm_task_.end()) {
+                break;
+            }
+            iteam->second->stop();
+            if (iteam->second->audio_flage_) {
+                unit_call("audio", "cap_stop", "None");
+            }
+            get_channel(iteam->first)->stop_subscriber("");
+            iteam->second.reset();
+            llm_task_.erase(iteam->first);
+        }
+    }
+};
+
+int main(int argc, char *argv[])
+{
+    signal(SIGTERM, __sigint);
+    signal(SIGINT, __sigint);
+    mkdir("/tmp/llm", 0777);
+    llm_kws llm;
+    while (!main_exit_flage) {
+        sleep(1);
+    }
+    return 0;
+}
\ No newline at end of file

From 01d6715b47562613d0ce435999ad202915c77c43 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Tue, 23 Sep 2025 12:35:15 +0800
Subject: [PATCH 43/79] [update] update cosy_voice & new kws

---
 projects/llm_framework/main/SConstruct        |   3 +-
 .../llm_framework/main_cosy_voice/SConstruct  |   9 +-
 .../models/mode_CosyVoice2-0.5B-ax650.json    |   3 +-
 .../main_cosy_voice/src/main.cpp              |   5 +-
 .../main_cosy_voice/src/runner/Token2wav.hpp  |  97 +++-
 .../llm_framework/main_kws_new/SConstruct     |  17 +-
 .../llm_framework/main_kws_new/src/main.cpp   | 341 +++++++++++--
 .../runner/kaldi-native-fbank/CMakeLists.txt  |   8 +
 .../kaldi-native-fbank/csrc/CMakeLists.txt    | 104 ++++
 .../kaldi-native-fbank/csrc/CPPLINT.cfg       |   1 +
 .../kaldi-native-fbank/csrc/feature-fbank.cc  | 116 +++++
 .../kaldi-native-fbank/csrc/feature-fbank.h   | 136 ++++++
 .../csrc/feature-functions.cc                 |  49 ++
 .../csrc/feature-functions.h                  |  38 ++
 .../kaldi-native-fbank/csrc/feature-mfcc.cc   | 174 +++++++
 .../kaldi-native-fbank/csrc/feature-mfcc.h    | 151 ++++++
 .../csrc/feature-raw-audio-samples.cc         |  38 ++
 .../csrc/feature-raw-audio-samples.h          |  81 ++++
 .../kaldi-native-fbank/csrc/feature-window.cc | 274 +++++++++++
 .../kaldi-native-fbank/csrc/feature-window.h  | 187 +++++++
 .../runner/kaldi-native-fbank/csrc/istft.cc   | 158 ++++++
 .../runner/kaldi-native-fbank/csrc/istft.h    |  45 ++
 .../kaldi-native-fbank/csrc/kaldi-math.cc     |  51 ++
 .../kaldi-native-fbank/csrc/kaldi-math.h      |  52 ++
 .../src/runner/kaldi-native-fbank/csrc/log.cc | 142 ++++++
 .../src/runner/kaldi-native-fbank/csrc/log.h  | 383 +++++++++++++++
 .../csrc/mel-computations.cc                  | 457 ++++++++++++++++++
 .../csrc/mel-computations.h                   | 190 ++++++++
 .../kaldi-native-fbank/csrc/online-feature.cc | 170 +++++++
 .../kaldi-native-fbank/csrc/online-feature.h  | 155 ++++++
 .../runner/kaldi-native-fbank/csrc/rfft.cc    | 112 +++++
 .../src/runner/kaldi-native-fbank/csrc/rfft.h |  58 +++
 .../runner/kaldi-native-fbank/csrc/stft.cc    | 142 ++++++
 .../src/runner/kaldi-native-fbank/csrc/stft.h |  65 +++
 .../kaldi-native-fbank/csrc/test-log.cc       |  71 +++
 .../csrc/test-online-fbank.cc                 |  54 +++
 .../csrc/test-online-feature.cc               |  59 +++
 .../kaldi-native-fbank/csrc/test-rfft.cc      | 101 ++++
 .../csrc/whisper-feature.cc                   | 169 +++++++
 .../kaldi-native-fbank/csrc/whisper-feature.h |  69 +++
 projects/llm_framework/main_kws_new/test.py   | 115 +++++
 projects/llm_framework/tools/llm_pack.py      |  22 +-
 42 files changed, 4589 insertions(+), 83 deletions(-)
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/CMakeLists.txt
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CMakeLists.txt
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CPPLINT.cfg
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.h
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.h
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.h
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.h
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.h
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.h
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.h
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.h
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.h
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.h
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.h
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.h
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-log.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-fbank.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-feature.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-rfft.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.cc
 create mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.h
 create mode 100644 projects/llm_framework/main_kws_new/test.py

diff --git a/projects/llm_framework/main/SConstruct b/projects/llm_framework/main/SConstruct
index ed4cbc5b..69e56047 100644
--- a/projects/llm_framework/main/SConstruct
+++ b/projects/llm_framework/main/SConstruct
@@ -28,7 +28,8 @@ STATIC_FILES += [AFile('../static_lib/sherpa/ncnn/libsherpa-ncnn-core.so'),
                  AFile('../static_lib/wetext/libfst.so.16'),
                  AFile('../static_lib/libonnxruntime.so.1'),
                  AFile('../static_lib/libonnxruntime.so.1.14.0'),
-                 AFile('../static_lib/libzmq.so.5')
+                 AFile('../static_lib/libzmq.so.5'),
+                 AFile('../static_lib/libMNN.so')
                  ]
 
 env['COMPONENTS'].append({'target':'static_file-1.0',
diff --git a/projects/llm_framework/main_cosy_voice/SConstruct b/projects/llm_framework/main_cosy_voice/SConstruct
index f82f28aa..76fae578 100644
--- a/projects/llm_framework/main_cosy_voice/SConstruct
+++ b/projects/llm_framework/main_cosy_voice/SConstruct
@@ -17,7 +17,7 @@ LDFLAGS = []
 LINK_SEARCH_PATH = []
 STATIC_FILES = []
 
-python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-llm-python-venv_v1.7.tar.gz", 'm5stack_llm-llm-python-venv_v1.7.tar.gz')
+python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-cosy-voice-python-venv_v1.7.tar.gz", 'm5stack_llm-cosy-voice-python-venv_v1.7.tar.gz')
 
 # REQUIREMENTS += ['Backward_cpp']
 # DYNAMIC_LIB += [ AFile('../static_lib/libdw.so.1'),
@@ -29,7 +29,7 @@ python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/reso
 DEFINITIONS += ['-O2']
 DEFINITIONS += ['-std=c++17']
 LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
-REQUIREMENTS += ['ax_engine', 'ax_interpreter', 'ax_sys', 'utilities']
+REQUIREMENTS += ['ax_engine', 'ax_interpreter', 'ax_sys', 'utilities', 'MNN']
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 
 
@@ -37,17 +37,16 @@ INCLUDE += [ADir('src/runner'), ADir('src/runner/utils')]
 INCLUDE += [ADir('../static_lib/include/sentencepiece'),
             ADir('../static_lib/include/protobuf-lite'),
             ADir('../static_lib/include/abseil-cpp'),
+            ADir('../static_lib/include/mnn'),
             ADir('../static_lib/include/re2')]
 
 static_file = Glob('../static_lib/module-llm/libabsl_*')
 static_file += [AFile('../static_lib/module-llm/libre2.a'), AFile('../static_lib/module-llm/libsentencepiece.a'), AFile('../static_lib/module-llm/libsentencepiece_train.a')]
 STATIC_LIB += static_file * 4
 
-# STATIC_FILES += [os.path.join(python_venv, 'llm')]
-# STATIC_FILES += Glob('scripts/tokenizer_*.py')
+STATIC_FILES += [os.path.join(python_venv, 'cosy-voice')]
 STATIC_FILES += Glob('scripts/tokenizer*')
 STATIC_FILES += Glob('models/mode_*.json')
-# STATIC_FILES += [AFile('scripts/llm-llm_tokenizer_auto.py')]
 
 IGNORE_FILES = []
 IGNORE_FILES += ['llm']
diff --git a/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json b/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json
index 2955af82..546a1270 100644
--- a/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json
+++ b/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json
@@ -38,7 +38,8 @@
         "b_use_mmap_load_embed": true,
         "b_dynamic_load_axmodel_layer": false,
         "ext_scripts": [
-            "tokenizer_cosyvoice2-0.5B-ax650.py"
+            "tokenizer_CosyVoice2-0.5B-ax650.py",
+            "tokenizer"
         ]
     }
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/main.cpp b/projects/llm_framework/main_cosy_voice/src/main.cpp
index 121eb2da..5fa42c43 100644
--- a/projects/llm_framework/main_cosy_voice/src/main.cpp
+++ b/projects/llm_framework/main_cosy_voice/src/main.cpp
@@ -203,7 +203,7 @@ class llm_task {
 
                     tokenizer_pid_ = fork();
                     if (tokenizer_pid_ == 0) {
-                        setenv("PYTHONPATH", "/opt/m5stack/lib/cosy_voice/site-packages", 1);
+                        setenv("PYTHONPATH", "/opt/m5stack/lib/cosy-voice/site-packages", 1);
                         const std::string port_str = std::to_string(port_);
                         const std::string model_id = base_model + "tokenizer";
 
@@ -266,11 +266,8 @@ class llm_task {
             if (!lToken2Wav.Init(mode_config_.token2wav_axmodel_dir, mode_config_.n_timesteps)) {
                 return -1;
             }
-            SLOGE();
             lLaMa_->TextToken2Embeds(prompt_text_token, prompt_text_embeds);
-            SLOGE();
             lLaMa_->SpeechToken2Embeds(prompt_speech_token, prompt_speech_embeds);
-            SLOGE();
             lToken2Wav.SpeechToken2Embeds(prompt_speech_token, prompt_speech_embeds_flow);
 
         } catch (...) {
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp b/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp
index 475d7440..f0cc3c43 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp
@@ -20,6 +20,9 @@
 #include "timer.hpp"
 // #include "opencv2/opencv.hpp"
 #include "ax_sys_api.h"
+#include "MNN/MNNDefine.h"
+#include "MNN/MNNForwardType.h"
+#include "MNN/Interpreter.hpp"
 
 class Token2Wav
 {
@@ -44,8 +47,15 @@ class Token2Wav
     ax_runner_ax650 flow_estimator_250;
     ax_runner_ax650 flow_estimator_300;
 
-    ax_runner_ax650 hift_50_first;
-    ax_runner_ax650 hift_58;
+    ax_runner_ax650 hift_p2_50_first;
+    ax_runner_ax650 hift_p2_58;
+
+    std::shared_ptr<MNN::Interpreter> hift_p1_50_first = nullptr;
+    std::shared_ptr<MNN::Interpreter> hift_p1_58 = nullptr;
+
+    MNN::Session * sess_hift_p1_50_first = nullptr;
+    MNN::Session * sess_hift_p1_58 = nullptr;
+
 
     std::vector<float> rand_noise;
     std::vector<float> t_span;
@@ -161,20 +171,44 @@ class Token2Wav
             return false;
         }
 
-        ret = hift_50_first.init((model_dir+"/hift_50_first.axmodel").c_str(), false);
+        ret = hift_p2_50_first.init((model_dir+"/hift_p2_50_first.axmodel").c_str(), false);
         if (ret != 0)
         {
-            ALOGE("init axmodel(%s) failed", (model_dir+"/hift_50_first.axmodel").c_str());
+            ALOGE("init axmodel(%s) failed", (model_dir+"/hift_p2_50_first.axmodel").c_str());
             return false;
         }
 
-        ret = hift_58.init((model_dir+"/hift_58.axmodel").c_str(), false);
+        ret = hift_p2_58.init((model_dir+"/hift_p2_58.axmodel").c_str(), false);
         if (ret != 0)
         {
-            ALOGE("init axmodel(%s) failed", (model_dir+"/hift_58.axmodel").c_str());
+            ALOGE("init axmodel(%s) failed", (model_dir+"/hift_p2_58.axmodel").c_str());
             return false;
         }
 
+        MNN::ScheduleConfig config;
+        config.numThread = 2;
+        config.type      = static_cast<MNNForwardType>(MNN_FORWARD_CPU);
+        MNN::BackendConfig backendConfig;
+        backendConfig.precision = (MNN::BackendConfig::PrecisionMode)1;
+        config.backendConfig = &backendConfig;
+
+        hift_p1_50_first = std::shared_ptr<MNN::Interpreter>(MNN::Interpreter::createFromFile( (model_dir+"/hift_p1_50_first.mnn").c_str() ));
+        if(nullptr == hift_p1_50_first)
+        {
+            ALOGE("init mnn model(%s) failed", (model_dir+"/hift_p1_50_first.mnn").c_str());
+            return false;
+        }
+        sess_hift_p1_50_first = hift_p1_50_first->createSession(config);
+
+        hift_p1_58 = std::shared_ptr<MNN::Interpreter>(MNN::Interpreter::createFromFile( (model_dir+"/hift_p1_58.mnn").c_str() ));
+        if(nullptr == hift_p1_58)
+        {
+            ALOGE("init mnn model(%s) failed", (model_dir+"/hift_p1_58.mnn").c_str() );
+            return false;
+        }
+
+        sess_hift_p1_58 = hift_p1_58->createSession(config);
+
         ALOGI("Token2Wav init ok");
         return true;
     }
@@ -188,8 +222,8 @@ class Token2Wav
         flow_estimator_200.release();
         flow_estimator_250.release();
         flow_estimator_300.release();
-        hift_50_first.release();
-        hift_58.release();
+        hift_p2_50_first.release();
+        hift_p2_58.release();
         flow_embed_selector.Deinit();
     }
 
@@ -318,39 +352,66 @@ class Token2Wav
     int infer_hift(std::vector<float> &mel, std::vector<float> &cache_source, 
                     std::vector<float> & tts_speech, std::vector<float> & tts_source)
     {
-        ax_runner_ax650 * model;
+        std::shared_ptr<MNN::Interpreter> model_p1;
+        MNN::Session * sess_p1;
+        ax_runner_ax650 * model_p2;
         int len = mel.size()/(80);
         
         if(len == 50 && cache_source.empty())
         { 
-            model = &hift_50_first;
+            model_p1 = hift_p1_50_first;
+            sess_p1 = sess_hift_p1_50_first;
+            model_p2 = &hift_p2_50_first;
         }else if(len == 58 && !cache_source.empty())
         {
-            model = &hift_58;
+            model_p1 = hift_p1_58;
+            sess_p1 = sess_hift_p1_58;
+            model_p2 = &hift_p2_58;
         }else
         {
             ALOGE("invalid size: %d", len);
             return -1;
         }
 
-        void * p = model->get_input("mel").pVirAddr;
+        std::vector<int> dims{1, 80, len};
+        auto tensor = MNN::Tensor::create<float>(dims, NULL, MNN::Tensor::CAFFE);
+        auto p_tensor   = tensor->host<float>();
+        auto size   = tensor->size();
+        std::memcpy(p_tensor, mel.data(), size);
+        
+        auto inputTensor = model_p1->getSessionInput(sess_p1, nullptr);
+        inputTensor->copyFromHostTensor(tensor);
+        
+        model_p1->runSession(sess_p1);
+        
+        MNN::Tensor *p_out  = model_p1->getSessionOutput(sess_p1, "s");
+        MNN::Tensor out_host(p_out, p_out->getDimensionType());
+        p_out->copyToHostTensor(&out_host);
+        
+        auto p_s = out_host.host<float>();
+
+        void * p = model_p2->get_input("s").pVirAddr;
+        memcpy(p, p_s, len * 480 * sizeof(float));
+        
+        p = model_p2->get_input("mel").pVirAddr;
         memcpy(p, mel.data(), mel.size() * sizeof(float));
+        
         if(!cache_source.empty())
         {
-            p = model->get_input("hift_cache_source").pVirAddr;
+            p = model_p2->get_input("hift_cache_source").pVirAddr;
             memcpy(p, cache_source.data(), cache_source.size() * sizeof(float));
         }
-
-        model->inference();
-
-        auto &output_speech = model->get_output("audio");
+        
+        model_p2->inference();
+        
+        auto &output_speech = model_p2->get_output("audio");
         if(tts_speech.empty() || tts_speech.size() != output_speech.nSize / sizeof(float))
         {
             tts_speech.resize(output_speech.nSize / sizeof(float));
         }
         memcpy(tts_speech.data(), output_speech.pVirAddr, output_speech.nSize);
 
-        auto &output_source = model->get_output("x");
+        auto &output_source = model_p2->get_output(1);
         if(tts_source.empty() || tts_source.size() != output_source.nSize / sizeof(float))
         {
             tts_source.resize(output_source.nSize / sizeof(float));
diff --git a/projects/llm_framework/main_kws_new/SConstruct b/projects/llm_framework/main_kws_new/SConstruct
index f42b60be..bc38274f 100644
--- a/projects/llm_framework/main_kws_new/SConstruct
+++ b/projects/llm_framework/main_kws_new/SConstruct
@@ -22,16 +22,19 @@ DEFINITIONS += ['-std=c++17']
 LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 
-INCLUDE += [ADir('../static_lib/include/sherpa'),
-            ADir('../static_lib/include/sherpa/sherpa-onnx'),
+INCLUDE += [
+    # ADir('../static_lib/include/sherpa'),
+            # ADir('../static_lib/include/sherpa/sherpa-onnx'),
             ADir('../static_lib/include/sherpa/sherpa-onnx/onnxruntime-src'),
-            ADir('../static_lib/include/sherpa/sherpa-onnx/openfst-src')
+            ADir('src/runner'),
+            # ADir('../static_lib/include/sherpa/sherpa-onnx/openfst-src')
             ]
 
-LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/onnx')]
-LDFLAGS += ['-l:libcargs.a',
-            '-l:libsherpa-onnx-core.a', '-l:libkaldi-native-fbank-core.a',
-            '-l:libkaldi-decoder-core.a', '-l:libssentencepiece_core.a']
+LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/fbank')]
+LDFLAGS += [
+            '-l:libkaldi-native-fbank-core.a',
+            '-l:libkissfft-float.a',
+            ]
 
 REQUIREMENTS += ['onnxruntime']
 
diff --git a/projects/llm_framework/main_kws_new/src/main.cpp b/projects/llm_framework/main_kws_new/src/main.cpp
index cd824290..f6664745 100644
--- a/projects/llm_framework/main_kws_new/src/main.cpp
+++ b/projects/llm_framework/main_kws_new/src/main.cpp
@@ -20,6 +20,12 @@
 #include <stdint.h>
 #include "libs/buffer.h"
 
+#include <onnxruntime_cxx_api.h>
+
+// #include <kaldi-native-fbank/csrc/feature-fbank.h>
+#include "kaldi-native-fbank/csrc/online-feature.h"
+// #include <kaldi-native-fbank/csrc/feature-window.h>
+
 using namespace StackFlows;
 
 int main_exit_flage = 0;
@@ -49,6 +55,14 @@ class llm_task {
     long long last_trigger_time_ms = -1e9;
     long long frame_index_global   = 0;
     int REFRACTORY_TIME_MS         = 2000;
+    const int RESAMPLE_RATE        = 16000;
+    const int FEAT_DIM             = 80;
+    std::vector<float> cache;
+    std::unique_ptr<Ort::Session> session;
+    knf::FbankOptions opts_;
+    std::unique_ptr<knf::OnlineFbank> fbank_;
+    Ort::Env env;
+    Ort::SessionOptions session_options;
 
 public:
     std::string model_;
@@ -61,9 +75,10 @@ class llm_task {
     bool enwake_audio_;
     std::atomic_bool audio_flage_;
     task_callback_t out_callback_;
-    int delay_audio_frame_ = 10;
+    int delay_audio_frame_ = 32;
     buffer_t *pcmdata;
     std::string wake_wav_file_;
+    int file_counter = 0;
 
     std::function<void(const std::string &)> play_awake_wav;
 
@@ -133,11 +148,36 @@ class llm_task {
             std::string base_model = base_model_path_ + model_ + "/";
             SLOGI("base_model %s", base_model.c_str());
 
+            std::string model_file = base_model + "kws.onnx";
+
             if (config_body.contains("wake_wav_file"))
                 wake_wav_file_ = config_body["wake_wav_file"];
             else if (file_body["mode_param"].contains("wake_wav_file"))
                 wake_wav_file_ = file_body["mode_param"]["wake_wav_file"];
 
+            session = std::make_unique<Ort::Session>(env, model_file.c_str(), session_options);
+            cache.assign(1 * 32 * 88, 0.0f);
+
+            opts_.frame_opts.samp_freq        = 16000;
+            opts_.frame_opts.frame_length_ms  = 25.0;
+            opts_.frame_opts.frame_shift_ms   = 10.0;
+            opts_.frame_opts.snip_edges       = false;
+            opts_.frame_opts.dither           = 0.0;
+            opts_.frame_opts.preemph_coeff    = 0.97;
+            opts_.frame_opts.remove_dc_offset = true;
+            opts_.frame_opts.window_type      = "povey";
+
+            opts_.mel_opts.num_bins  = 80;
+            opts_.mel_opts.low_freq  = 20;
+            opts_.mel_opts.high_freq = 0;
+
+            opts_.energy_floor = 0.0;
+            opts_.use_energy   = false;
+
+            opts_.raw_energy = true;
+            // use_log_fbank / use_power 由 knf::OnlineFbank 默认控制，一般一致
+
+            fbank_ = std::make_unique<knf::OnlineFbank>(opts_);
         } catch (...) {
             SLOGE("config file read false");
             return -3;
@@ -153,13 +193,14 @@ class llm_task {
     bool detect_wakeup(const std::vector<float> &scores)
     {
         bool triggered = false;
+        SLOGE("%d", scores.size());
         for (auto score : scores) {
+            printf("%f ", score);
             if (score > threshold) {
                 count_frames++;
                 if (count_frames >= min_continuous_frames) {
                     long long trigger_time_ms = (frame_index_global - min_continuous_frames + 1) * 10;
                     if (trigger_time_ms - last_trigger_time_ms >= REFRACTORY_TIME_MS) {
-                        std::cout << "[唤醒触发] 时间: " << trigger_time_ms << " ms\n";
                         last_trigger_time_ms = trigger_time_ms;
                         triggered            = true;
                     }
@@ -169,64 +210,260 @@ class llm_task {
             }
             frame_index_global++;
         }
+        SLOGE("\n");
         return triggered;
     }
 
-    kaldi::Matrix<float> compute_fbank(const std::vector<float> &audio16k)
+    std::vector<std::vector<float>> compute_fbank_kaldi1(const std::vector<float> &waveform, int sample_rate,
+                                                         int num_mel_bins)
+    {
+        fbank_.reset();
+        fbank_ = std::make_unique<knf::OnlineFbank>(opts_);
+
+        fbank_->AcceptWaveform(sample_rate, waveform.data(), waveform.size());
+
+        int num_frames = fbank_->NumFramesReady();
+
+        std::vector<std::vector<float>> features;
+        SLOGE("num_frames %d", num_frames);
+        features.reserve(num_frames);
+
+        for (int i = 0; i < num_frames; ++i) {
+            const float *frame_data = fbank_->GetFrame(i);
+            std::vector<float> frame(frame_data, frame_data + num_mel_bins);
+            features.push_back(std::move(frame));
+        }
+
+        return features;
+    }
+
+    std::vector<std::vector<float>> compute_fbank_kaldi(const std::vector<float> &waveform, int sample_rate,
+                                                        int num_mel_bins)
     {
-        kaldi::FbankOptions fbank_opts;
-        fbank_opts.mel_opts.num_bins     = feat_dim;
-        fbank_opts.frame_opts.samp_freq  = resample_rate;
-        fbank_opts.frame_opts.dither     = 0.0;
-        fbank_opts.energy_floor          = 0.0;
-        fbank_opts.mel_opts.htk_mode     = false;
-        fbank_opts.use_energy            = false;
-        fbank_opts.frame_opts.snip_edges = false;
+        fbank_.reset();
+        fbank_ = std::make_unique<knf::OnlineFbank>(opts_);
+        fbank_->AcceptWaveform(sample_rate, waveform.data(), waveform.size());
+        int num_frames = fbank_->NumFramesReady();
+        std::vector<std::vector<float>> features;
+        SLOGE("num_frames %d", num_frames);
+        features.reserve(num_frames);
+        for (int i = 0; i < num_frames; ++i) {
+            const float *frame_data = fbank_->GetFrame(i);
+            std::vector<float> frame(frame_data, frame_data + num_mel_bins);
+            features.push_back(std::move(frame));
+        }
+
+        // 生成文件名序号（三位数，补零）
+        // std::stringstream ss;
+        // ss << std::setfill('0') << std::setw(3) << file_counter;
+        // std::string file_suffix = ss.str();
+
+        // 保存 waveform 为二进制
+        // std::string waveform_filename = "waveform_" + file_suffix + ".bin";
+        // std::ofstream waveform_bin(waveform_filename, std::ios::binary);
+        // if (waveform_bin.is_open()) {
+        //     size_t size = waveform.size();
+        //     waveform_bin.write(reinterpret_cast<const char *>(&size), sizeof(size));
+        //     waveform_bin.write(reinterpret_cast<const char *>(waveform.data()), sizeof(float) * waveform.size());
+        //     waveform_bin.close();
+        //     SLOGE("Waveform saved to %s", waveform_filename.c_str());
+        // }
+
+        // 保存 features 为二进制
+        // std::string features_filename = "features_" + file_suffix + ".bin";
+        // std::ofstream features_bin(features_filename, std::ios::binary);
+        // if (features_bin.is_open()) {
+        //     size_t rows = features.size();
+        //     size_t cols = (rows > 0) ? features[0].size() : 0;
+        //     features_bin.write(reinterpret_cast<const char *>(&rows), sizeof(rows));
+        //     features_bin.write(reinterpret_cast<const char *>(&cols), sizeof(cols));
+        //     for (const auto &row : features) {
+        //         features_bin.write(reinterpret_cast<const char *>(row.data()), sizeof(float) * row.size());
+        //     }
+        //     features_bin.close();
+        //     SLOGE("Features saved to %s", features_filename.c_str());
+        // }
+
+        // 递增计数器
+        // file_counter++;
+
+        return features;
+    }
+
+    std::vector<std::vector<float>> read_mat_from_bin(int index)
+    {
+        std::stringstream ss;
+        ss << "mat_" << std::setfill('0') << std::setw(3) << index << ".bin";
+        std::string filename = ss.str();
+
+        std::ifstream file(filename, std::ios::binary);
+        if (!file.is_open()) {
+            SLOGE("无法打开文件: %s", filename.c_str());
+            return {};
+        }
 
-        kaldi::Fbank fbank(fbank_opts);
+        // 读取维度信息
+        int64_t rows, cols;
+        file.read(reinterpret_cast<char *>(&rows), sizeof(int64_t));
+        file.read(reinterpret_cast<char *>(&cols), sizeof(int64_t));
 
-        kaldi::Vector<float> wav(audio16k.size());
-        for (size_t i = 0; i < audio16k.size(); i++) {
-            wav(i) = audio16k[i] * (1 << 15);
+        SLOGE("读取mat文件 %s, shape: (%ld, %ld)", filename.c_str(), rows, cols);
+
+        // 读取数据
+        std::vector<float> data(rows * cols);
+        file.read(reinterpret_cast<char *>(data.data()), sizeof(float) * data.size());
+        file.close();
+
+        // 转换为二维向量
+        std::vector<std::vector<float>> result;
+        result.reserve(rows);
+
+        for (int i = 0; i < rows; ++i) {
+            std::vector<float> row;
+            row.reserve(cols);
+            for (int j = 0; j < cols; ++j) {
+                row.push_back(data[i * cols + j]);
+            }
+            result.push_back(std::move(row));
         }
 
-        kaldi::Matrix<float> features;
-        fbank.ComputeFeatures(wav, resample_rate, 1.0, &features);
-        return features;  // Shape: [num_frames, feat_dim]
+        return result;
     }
 
-    // std::vector<float> run_inference(const std::vector<float> &audio16k)
-    // {
-    //     auto feats = compute_fbank(audio16k);
+    std::vector<float> run_inference1(const std::vector<float> &audio_chunk_16k)
+    {
+        auto fbank_feats = compute_fbank_kaldi(audio_chunk_16k, RESAMPLE_RATE, FEAT_DIM);
+        SLOGE("=== FBank Features ===");
+        for (int i = 0; i < fbank_feats.size(); ++i) {
+            SLOGE("Frame %d:", i);
+            std::string frame_str = "";
+            for (int j = 0; j < fbank_feats[i].size(); ++j) {
+                frame_str += std::to_string(fbank_feats[i][j]) + " ";
+            }
+            SLOGE("%s", frame_str.c_str());
+        }
+        SLOGE("fbank_feats.size()=%d", (int)fbank_feats.size());
+        SLOGE("=======================");
+
+        if (fbank_feats.empty()) {
+            return {};
+        }
+
+        int T = fbank_feats.size();
+        std::vector<float> mat_flattened;
+        for (const auto &feat : fbank_feats) {
+            mat_flattened.insert(mat_flattened.end(), feat.begin(), feat.end());
+        }
+
+        std::vector<int64_t> input_shape = {1, static_cast<int64_t>(T), FEAT_DIM};
+        std::vector<int64_t> cache_shape = {1, 32, 88};
 
-    //     std::array<int64_t, 3> input_shape{1, feats.NumRows(), feats.NumCols()};
-    //     Ort::Value input_tensor = Ort::Value::CreateTensor<float>(
-    //         memInfo, (float *)feats.Data(), feats.NumRows() * feats.NumCols(), input_shape.data(),
-    //         input_shape.size());
+        Ort::MemoryInfo memory_info = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
 
-    //     // cache 作为输入
-    //     std::array<int64_t, 4> cache_shape{1, 32, 88, 1};
-    //     Ort::Value cache_tensor = Ort::Value::CreateTensor<float>(
-    //         memInfo, (float *)cache.Data(), cache.NumRows() * cache.NumCols(), cache_shape.data(),
-    //         cache_shape.size());
+        Ort::Value input_tensor = Ort::Value::CreateTensor<float>(
+            memory_info, mat_flattened.data(), mat_flattened.size(), input_shape.data(), input_shape.size());
 
-    //     const char *input_names[]  = {"input", "cache"};
-    //     const char *output_names[] = {"output", "cache_out"};
+        Ort::Value cache_tensor = Ort::Value::CreateTensor<float>(memory_info, cache.data(), cache.size(),
+                                                                  cache_shape.data(), cache_shape.size());
 
-    //     auto outputs = ort_session->Run(Ort::RunOptions{nullptr}, input_names, &input_tensor,
-    //                                     1,  // TODO: pass cache_tensor too (update input count)
-    //                                     output_names, 2);
+        const char *input_names[]  = {"input", "cache"};
+        const char *output_names[] = {"output", "r_cache"};
 
-    //     auto &out_chunk_tensor = outputs[0];
-    //     float *out_data        = out_chunk_tensor.GetTensorMutableData<float>();
-    //     size_t out_len         = out_chunk_tensor.GetTensorTypeAndShapeInfo().GetElementCount();
+        std::vector<Ort::Value> inputs;
+        inputs.push_back(std::move(input_tensor));
+        inputs.push_back(std::move(cache_tensor));
 
-    //     return std::vector<float>(out_data, out_data + out_len);
-    // }
+        auto output_tensors = session->Run(Ort::RunOptions{nullptr}, input_names, inputs.data(), 2, output_names, 2);
+
+        float *out_data       = output_tensors[0].GetTensorMutableData<float>();
+        float *cache_out_data = output_tensors[1].GetTensorMutableData<float>();
+
+        std::vector<int64_t> out_shape = output_tensors[0].GetTensorTypeAndShapeInfo().GetShape();
+        size_t out_size                = 1;
+        for (auto dim : out_shape) out_size *= dim;
+
+        std::vector<float> out_chunk(out_data, out_data + out_size);
+
+        std::copy(cache_out_data, cache_out_data + cache.size(), cache.begin());
+
+        return out_chunk;
+    }
+
+    std::vector<float> run_inference(const std::vector<float> &audio_chunk_16k)
+    {
+        std::vector<std::vector<float>> fbank_feats;
+        static int file_index = -1;
+        // file_index++;
+        if (file_index >= 0) {
+            // 从文件读取
+            fbank_feats = read_mat_from_bin(file_index);
+            SLOGE("=== 从文件读取的 FBank Features ===");
+        } else {
+            // 原来的计算方式
+            fbank_feats = compute_fbank_kaldi(audio_chunk_16k, RESAMPLE_RATE, FEAT_DIM);
+            SLOGE("=== 计算的 FBank Features ===");
+        }
+        if (file_index == 6) file_index = 0;
+
+        // 打印特征用于对比
+        // for (int i = 0; i < fbank_feats.size(); ++i) {
+        //     SLOGE("Frame %d:", i);
+        //     std::string frame_str = "";
+        //     for (int j = 0; j < fbank_feats[i].size(); ++j) {
+        //         frame_str += std::to_string(fbank_feats[i][j]) + " ";
+        //     }
+        //     SLOGE("%s", frame_str.c_str());
+        // }
+        // SLOGE("fbank_feats.size()=%d", (int)fbank_feats.size());
+        // SLOGE("=======================");
+
+        if (fbank_feats.empty()) {
+            return {};
+        }
+
+        int T = fbank_feats.size();
+        std::vector<float> mat_flattened;
+        for (const auto &feat : fbank_feats) {
+            mat_flattened.insert(mat_flattened.end(), feat.begin(), feat.end());
+        }
+
+        std::vector<int64_t> input_shape = {1, static_cast<int64_t>(T), FEAT_DIM};
+        std::vector<int64_t> cache_shape = {1, 32, 88};
+
+        Ort::MemoryInfo memory_info = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
+        Ort::Value input_tensor     = Ort::Value::CreateTensor<float>(
+            memory_info, mat_flattened.data(), mat_flattened.size(), input_shape.data(), input_shape.size());
+        Ort::Value cache_tensor = Ort::Value::CreateTensor<float>(memory_info, cache.data(), cache.size(),
+                                                                  cache_shape.data(), cache_shape.size());
+
+        const char *input_names[]  = {"input", "cache"};
+        const char *output_names[] = {"output", "r_cache"};
+
+        std::vector<Ort::Value> inputs;
+        inputs.push_back(std::move(input_tensor));
+        inputs.push_back(std::move(cache_tensor));
+
+        auto output_tensors = session->Run(Ort::RunOptions{nullptr}, input_names, inputs.data(), 2, output_names, 2);
+
+        float *out_data       = output_tensors[0].GetTensorMutableData<float>();
+        float *cache_out_data = output_tensors[1].GetTensorMutableData<float>();
+
+        std::vector<int64_t> out_shape = output_tensors[0].GetTensorTypeAndShapeInfo().GetShape();
+        size_t out_size                = 1;
+        for (auto dim : out_shape) out_size *= dim;
+
+        std::vector<float> out_chunk(out_data, out_data + out_size);
+        std::copy(cache_out_data, cache_out_data + cache.size(), cache.begin());
+
+        return out_chunk;
+    }
 
     void sys_pcm_on_data(const std::string &raw)
     {
         static int count = 0;
+        static std::ofstream ffout("audio_float.raw", std::ios::binary | std::ios::app);
+        static std::ofstream fout("audio_int16.raw", std::ios::binary | std::ios::app);
+
         if (count < delay_audio_frame_) {
             buffer_write_char(pcmdata, raw.data(), raw.length());
             count++;
@@ -236,16 +473,36 @@ class llm_task {
         buffer_position_set(pcmdata, 0);
 
         std::vector<float> floatSamples;
+        std::vector<int16_t> int16Samples;
         {
             int16_t audio_val;
             while (buffer_read_i16(pcmdata, &audio_val, 1)) {
-                float normalizedSample = static_cast<float>(audio_val) / INT16_MAX;
-                floatSamples.push_back(normalizedSample);
+                int16Samples.push_back(audio_val);
+                float normalizedSample = static_cast<float>(audio_val) / 1.0f;
+                float sample           = static_cast<float>(normalizedSample);
+                floatSamples.push_back(sample);
             }
         }
 
         buffer_resize(pcmdata, 0);
         count = 0;
+
+        if (!floatSamples.empty()) {
+            ffout.write(reinterpret_cast<const char *>(floatSamples.data()), floatSamples.size() * sizeof(float));
+        }
+
+        if (!int16Samples.empty()) {
+            fout.write(reinterpret_cast<const char *>(int16Samples.data()), int16Samples.size() * sizeof(int16_t));
+        }
+
+        auto scores = run_inference(floatSamples);
+        if (detect_wakeup(scores)) {
+            if (enwake_audio_ && (!wake_wav_file_.empty()) && play_awake_wav) {
+                SLOGE("\n\n\n\n\n");
+                play_awake_wav(wake_wav_file_);
+            }
+            if (out_callback_) out_callback_("", true);
+        }
     }
 
     void trigger()
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/CMakeLists.txt b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/CMakeLists.txt
new file mode 100644
index 00000000..20376267
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/CMakeLists.txt
@@ -0,0 +1,8 @@
+add_subdirectory(csrc)
+
+if(KALDI_NATIVE_FBANK_BUILD_PYTHON)
+  message(STATUS "Building Python")
+  add_subdirectory(python)
+else()
+  message(STATUS "Disable building Python")
+endif()
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CMakeLists.txt b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CMakeLists.txt
new file mode 100644
index 00000000..741c2340
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CMakeLists.txt
@@ -0,0 +1,104 @@
+
+include_directories(${PROJECT_SOURCE_DIR})
+set(sources
+  feature-fbank.cc
+  feature-functions.cc
+  feature-mfcc.cc
+  feature-raw-audio-samples.cc
+  feature-window.cc
+  istft.cc
+  kaldi-math.cc
+  mel-computations.cc
+  online-feature.cc
+  rfft.cc
+  stft.cc
+  whisper-feature.cc
+)
+
+if(KALDI_NATIVE_FBANK_ENABLE_CHECK)
+  list(APPEND sources log.cc)
+endif()
+
+add_library(kaldi-native-fbank-core ${sources})
+if(KALDI_NATIVE_FBANK_ENABLE_CHECK)
+  target_compile_definitions(kaldi-native-fbank-core PUBLIC KNF_ENABLE_CHECK=1)
+
+  if(KNF_HAVE_EXECINFO_H)
+    target_compile_definitions(kaldi-native-fbank-core PRIVATE KNF_HAVE_EXECINFO_H=1)
+  endif()
+
+  if(KNF_HAVE_CXXABI_H)
+    target_compile_definitions(kaldi-native-fbank-core PRIVATE KNF_HAVE_CXXABI_H=1)
+  endif()
+endif()
+
+# We are using std::call_once() in log.h,which requires us to link with -pthread
+if(NOT WIN32 AND KALDI_NATIVE_FBANK_ENABLE_CHECK)
+  target_link_libraries(kaldi-native-fbank-core -pthread)
+endif()
+target_link_libraries(kaldi-native-fbank-core kissfft)
+
+if(KALDI_NATIVE_FBANK_BUILD_TESTS)
+  add_executable(test-online-fbank test-online-fbank.cc)
+  target_link_libraries(test-online-fbank kaldi-native-fbank-core)
+endif()
+
+function(kaldi_native_fbank_add_test source)
+  get_filename_component(name ${source} NAME_WE)
+  add_executable(${name} "${source}")
+  target_link_libraries(${name}
+    PRIVATE
+      kaldi-native-fbank-core
+      gtest
+      gtest_main
+  )
+
+  add_test(NAME "Test.${name}"
+    COMMAND
+    $<TARGET_FILE:${name}>
+  )
+endfunction()
+
+# please sort the source files alphabetically
+set(test_srcs
+  # test-online-feature.cc
+  test-log.cc
+  test-rfft.cc
+)
+
+if(KALDI_NATIVE_FBANK_BUILD_TESTS)
+  foreach(source IN LISTS test_srcs)
+    kaldi_native_fbank_add_test(${source})
+  endforeach()
+endif()
+
+install(TARGETS kaldi-native-fbank-core
+  DESTINATION lib
+)
+if(KALDI_NATIVE_FBANK_BUILD_PYTHON AND WIN32)
+  install(TARGETS kaldi-native-fbank-core
+    DESTINATION ..
+  )
+endif()
+
+if(KALDI_NATIVE_FBANK_BUILD_TESTS)
+  install(TARGETS test-online-fbank
+    DESTINATION bin
+  )
+endif()
+
+file(MAKE_DIRECTORY
+  DESTINATION
+    ${PROJECT_BINARY_DIR}/include/kaldi-native-fbank/csrc
+)
+file(GLOB_RECURSE all_headers *.h)
+
+file(COPY
+  ${all_headers}
+  DESTINATION
+    ${PROJECT_BINARY_DIR}/include/kaldi-native-fbank/csrc
+)
+
+install(FILES ${all_headers}
+  DESTINATION include/kaldi-native-fbank/csrc
+)
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CPPLINT.cfg b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CPPLINT.cfg
new file mode 100644
index 00000000..4c1a8fb3
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CPPLINT.cfg
@@ -0,0 +1 @@
+exclude_files=fftsg.cc
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.cc
new file mode 100644
index 00000000..b73aa807
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.cc
@@ -0,0 +1,116 @@
+/**
+ * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// This file is copied/modified from kaldi/src/feat/feature-fbank.cc
+//
+#include "kaldi-native-fbank/csrc/feature-fbank.h"
+
+#include <algorithm>
+#include <limits>
+#include <utility>
+#include <vector>
+
+#include "kaldi-native-fbank/csrc/feature-functions.h"
+#include "kaldi-native-fbank/csrc/kaldi-math.h"
+#include "kaldi-native-fbank/csrc/log.h"
+
+namespace knf {
+
+std::ostream &operator<<(std::ostream &os, const FbankOptions &opts) {
+  os << opts.ToString();
+  return os;
+}
+
+FbankComputer::FbankComputer(const FbankOptions &opts)
+    : opts_(opts), rfft_(opts.frame_opts.PaddedWindowSize()) {
+  if (opts.energy_floor > 0.0f) {
+    log_energy_floor_ = logf(opts.energy_floor);
+  }
+
+  // We'll definitely need the filterbanks info for VTLN warping factor 1.0.
+  // [note: this call caches it.]
+  GetMelBanks(1.0f);
+}
+
+FbankComputer::~FbankComputer() {
+  for (auto iter = mel_banks_.begin(); iter != mel_banks_.end(); ++iter)
+    delete iter->second;
+}
+
+const MelBanks *FbankComputer::GetMelBanks(float vtln_warp) {
+  MelBanks *this_mel_banks = nullptr;
+
+  // std::map<float, MelBanks *>::iterator iter = mel_banks_.find(vtln_warp);
+  auto iter = mel_banks_.find(vtln_warp);
+  if (iter == mel_banks_.end()) {
+    this_mel_banks = new MelBanks(opts_.mel_opts, opts_.frame_opts, vtln_warp);
+    mel_banks_[vtln_warp] = this_mel_banks;
+  } else {
+    this_mel_banks = iter->second;
+  }
+  return this_mel_banks;
+}
+
+void FbankComputer::Compute(float signal_raw_log_energy, float vtln_warp,
+                            std::vector<float> *signal_frame, float *feature) {
+  const MelBanks &mel_banks = *(GetMelBanks(vtln_warp));
+
+  KNF_CHECK_EQ(signal_frame->size(), opts_.frame_opts.PaddedWindowSize());
+
+  // Compute energy after window function (not the raw one).
+  if (opts_.use_energy && !opts_.raw_energy) {
+    signal_raw_log_energy = std::log(
+        std::max<float>(InnerProduct(signal_frame->data(), signal_frame->data(),
+                                     signal_frame->size()),
+                        std::numeric_limits<float>::epsilon()));
+  }
+  rfft_.Compute(signal_frame->data());  // signal_frame is modified in-place
+  ComputePowerSpectrum(signal_frame);
+
+  // Use magnitude instead of power if requested.
+  if (!opts_.use_power) {
+    Sqrt(signal_frame->data(), signal_frame->size() / 2 + 1);
+  }
+
+  int32_t mel_offset = ((opts_.use_energy && !opts_.htk_compat) ? 1 : 0);
+
+  // Its length is opts_.mel_opts.num_bins
+  float *mel_energies = feature + mel_offset;
+
+  // Sum with mel filter banks over the power spectrum
+  mel_banks.Compute(signal_frame->data(), mel_energies);
+
+  if (opts_.use_log_fbank) {
+    // Avoid log of zero (which should be prevented anyway by dithering).
+    for (int32_t i = 0; i != opts_.mel_opts.num_bins; ++i) {
+      auto t = std::max(mel_energies[i], std::numeric_limits<float>::epsilon());
+      mel_energies[i] = std::log(t);
+    }
+  }
+
+  // Copy energy as first value (or the last, if htk_compat == true).
+  if (opts_.use_energy) {
+    if (opts_.energy_floor > 0.0 && signal_raw_log_energy < log_energy_floor_) {
+      signal_raw_log_energy = log_energy_floor_;
+    }
+    int32_t energy_index = opts_.htk_compat ? opts_.mel_opts.num_bins : 0;
+    feature[energy_index] = signal_raw_log_energy;
+  }
+}
+
+}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.h
new file mode 100644
index 00000000..84005e0d
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.h
@@ -0,0 +1,136 @@
+/**
+ * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// This file is copied/modified from kaldi/src/feat/feature-fbank.h
+
+#ifndef KALDI_NATIVE_FBANK_CSRC_FEATURE_FBANK_H_
+#define KALDI_NATIVE_FBANK_CSRC_FEATURE_FBANK_H_
+
+#include <map>
+#include <string>
+#include <vector>
+#include <cstdint>
+#include <sstream>
+
+#include "kaldi-native-fbank/csrc/feature-window.h"
+#include "kaldi-native-fbank/csrc/mel-computations.h"
+#include "kaldi-native-fbank/csrc/rfft.h"
+
+namespace knf {
+
+struct FbankOptions {
+  FrameExtractionOptions frame_opts;
+  MelBanksOptions mel_opts;
+  // append an extra dimension with energy to the filter banks
+  bool use_energy = false;
+  float energy_floor = 0.0f;  // active iff use_energy==true
+
+  // If true, compute log_energy before preemphasis and windowing
+  // If false, compute log_energy after preemphasis ans windowing
+  bool raw_energy = true;  // active iff use_energy==true
+
+  // If true, put energy last (if using energy)
+  // If false, put energy first
+  bool htk_compat = false;  // active iff use_energy==true
+
+  // if true (default), produce log-filterbank, else linear
+  bool use_log_fbank = true;
+
+  // if true (default), use power in filterbank
+  // analysis, else magnitude.
+  bool use_power = true;
+
+  FbankOptions() { mel_opts.num_bins = 23; }
+
+  std::string ToString() const {
+    std::ostringstream os;
+    os << "frame_opts: \n";
+    os << frame_opts << "\n";
+    os << "\n";
+
+    os << "mel_opts: \n";
+    os << mel_opts << "\n";
+
+    os << "use_energy: " << use_energy << "\n";
+    os << "energy_floor: " << energy_floor << "\n";
+    os << "raw_energy: " << raw_energy << "\n";
+    os << "htk_compat: " << htk_compat << "\n";
+    os << "use_log_fbank: " << use_log_fbank << "\n";
+    os << "use_power: " << use_power << "\n";
+    return os.str();
+  }
+};
+
+std::ostream &operator<<(std::ostream &os, const FbankOptions &opts);
+
+class FbankComputer {
+ public:
+  using Options = FbankOptions;
+
+  explicit FbankComputer(const FbankOptions &opts);
+  ~FbankComputer();
+
+  int32_t Dim() const {
+    return opts_.mel_opts.num_bins + (opts_.use_energy ? 1 : 0);
+  }
+
+  // if true, compute log_energy_pre_window but after dithering and dc removal
+  bool NeedRawLogEnergy() const { return opts_.use_energy && opts_.raw_energy; }
+
+  const FrameExtractionOptions &GetFrameOptions() const {
+    return opts_.frame_opts;
+  }
+
+  const FbankOptions &GetOptions() const { return opts_; }
+
+  /**
+     Function that computes one frame of features from
+     one frame of signal.
+
+     @param [in] signal_raw_log_energy The log-energy of the frame of the signal
+         prior to windowing and pre-emphasis, or
+         log(numeric_limits<float>::min()), whichever is greater.  Must be
+         ignored by this function if this class returns false from
+         this->NeedsRawLogEnergy().
+     @param [in] vtln_warp  The VTLN warping factor that the user wants
+         to be applied when computing features for this utterance.  Will
+         normally be 1.0, meaning no warping is to be done.  The value will
+         be ignored for feature types that don't support VLTN, such as
+         spectrogram features.
+     @param [in] signal_frame  One frame of the signal,
+       as extracted using the function ExtractWindow() using the options
+       returned by this->GetFrameOptions().  The function will use the
+       vector as a workspace, which is why it's a non-const pointer.
+     @param [out] feature  Pointer to a vector of size this->Dim(), to which
+         the computed feature will be written. It should be pre-allocated.
+  */
+  void Compute(float signal_raw_log_energy, float vtln_warp,
+               std::vector<float> *signal_frame, float *feature);
+
+ private:
+  const MelBanks *GetMelBanks(float vtln_warp);
+
+  FbankOptions opts_;
+  float log_energy_floor_;
+  std::map<float, MelBanks *> mel_banks_;  // float is VTLN coefficient.
+  Rfft rfft_;
+};
+
+}  // namespace knf
+
+#endif  // KALDI_NATIVE_FBANK_CSRC_FEATURE_FBANK_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.cc
new file mode 100644
index 00000000..00ae4c79
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.cc
@@ -0,0 +1,49 @@
+/**
+ * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// This file is copied/modified from kaldi/src/feat/feature-functions.cc
+
+#include "kaldi-native-fbank/csrc/feature-functions.h"
+
+#include <cstdint>
+#include <vector>
+
+namespace knf {
+
+void ComputePowerSpectrum(std::vector<float> *complex_fft) {
+  int32_t dim = complex_fft->size();
+
+  // now we have in complex_fft, first half of complex spectrum
+  // it's stored as [real0, realN/2, real1, im1, real2, im2, ...]
+
+  float *p = complex_fft->data();
+  int32_t half_dim = dim / 2;
+  float first_energy = p[0] * p[0];
+  float last_energy = p[1] * p[1];  // handle this special case
+
+  for (int32_t i = 1; i < half_dim; ++i) {
+    float real = p[i * 2];
+    float im = p[i * 2 + 1];
+    p[i] = real * real + im * im;
+  }
+  p[0] = first_energy;
+  p[half_dim] = last_energy;  // Will actually never be used, and anyway
+  // if the signal has been bandlimited sensibly this should be zero.
+}
+
+}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.h
new file mode 100644
index 00000000..b2216224
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.h
@@ -0,0 +1,38 @@
+/**
+ * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// This file is copied/modified from kaldi/src/feat/feature-functions.h
+#ifndef KALDI_NATIVE_FBANK_CSRC_FEATURE_FUNCTIONS_H_
+#define KALDI_NATIVE_FBANK_CSRC_FEATURE_FUNCTIONS_H_
+
+#include <vector>
+namespace knf {
+
+// ComputePowerSpectrum converts a complex FFT (as produced by the FFT
+// functions in csrc/rfft.h), and converts it into
+// a power spectrum.  If the complex FFT is a vector of size n (representing
+// half of the complex FFT of a real signal of size n, as described there),
+// this function computes in the first (n/2) + 1 elements of it, the
+// energies of the fft bins from zero to the Nyquist frequency.  Contents of the
+// remaining (n/2) - 1 elements are undefined at output.
+
+void ComputePowerSpectrum(std::vector<float> *complex_fft);
+
+}  // namespace knf
+
+#endif  // KALDI_NATIVE_FBANK_CSRC_FEATURE_FUNCTIONS_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.cc
new file mode 100644
index 00000000..3f38f2fc
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.cc
@@ -0,0 +1,174 @@
+/**
+ * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// This file is copied/modified from kaldi/src/feat/feature-mfcc.cc
+//
+#include "kaldi-native-fbank/csrc/feature-mfcc.h"
+
+#include <algorithm>
+#include <cmath>
+#include <limits>
+#include <utility>
+#include <vector>
+
+#include "kaldi-native-fbank/csrc/feature-functions.h"
+#include "kaldi-native-fbank/csrc/feature-window.h"
+#include "kaldi-native-fbank/csrc/kaldi-math.h"
+#include "kaldi-native-fbank/csrc/log.h"
+
+namespace knf {
+
+static std::vector<float> ComputeDctMatrix(int32_t num_rows, int32_t num_cols) {
+  // this function is copied from
+  // https://github.com/kaldi-asr/kaldi/blob/master/src/matrix/matrix-functions.cc#L592
+
+  std::vector<float> ans(num_rows * num_cols);
+  float *p = ans.data();
+
+  float normalizer = std::sqrt(1.0 / num_cols);  // normalizer for X_0
+
+  for (int32_t i = 0; i != num_cols; ++i) {
+    p[i] = normalizer;
+  }
+
+  normalizer = std::sqrt(2.0 / num_cols);  // normalizer for other elements
+
+  for (int32_t k = 1; k != num_rows; ++k) {
+    for (int32_t n = 0; n != num_cols; ++n) {
+      *(p + k * num_cols + n) =
+          normalizer *
+          std::cos(static_cast<double>(M_PI) / num_cols * (n + 0.5) * k);
+    }
+  }
+
+  return ans;
+}
+
+std::ostream &operator<<(std::ostream &os, const MfccOptions &opts) {
+  os << opts.ToString();
+  return os;
+}
+
+MfccComputer::MfccComputer(const MfccOptions &opts)
+    : opts_(opts),
+      rfft_(opts.frame_opts.PaddedWindowSize()),
+      mel_energies_(opts.mel_opts.num_bins) {
+  if (opts.energy_floor > 0.0f) {
+    log_energy_floor_ = logf(opts.energy_floor);
+  }
+
+  // We'll definitely need the filterbanks info for VTLN warping factor 1.0.
+  // [note: this call caches it.]
+  GetMelBanks(1.0f);
+
+  int32_t num_bins = opts.mel_opts.num_bins;
+
+  KNF_CHECK_LE(opts.num_ceps, num_bins)
+      << "num-ceps cannot be larger than num-mel-bins."
+      << " It should be smaller or equal. You provided num-ceps: "
+      << opts.num_ceps << "  and num-mel-bins: " << num_bins;
+
+  dct_matrix_ = ComputeDctMatrix(opts.num_ceps, num_bins);
+
+  if (opts.cepstral_lifter != 0.0) {
+    lifter_coeffs_ = std::vector<float>(opts.num_ceps);
+    // torch::empty({1, opts.num_ceps}, torch::kFloat32);
+    ComputeLifterCoeffs(opts.cepstral_lifter, &lifter_coeffs_);
+  }
+}
+
+MfccComputer::~MfccComputer() {
+  for (auto iter = mel_banks_.begin(); iter != mel_banks_.end(); ++iter)
+    delete iter->second;
+}
+
+const MelBanks *MfccComputer::GetMelBanks(float vtln_warp) {
+  MelBanks *this_mel_banks = nullptr;
+
+  // std::map<float, MelBanks *>::iterator iter = mel_banks_.find(vtln_warp);
+  auto iter = mel_banks_.find(vtln_warp);
+  if (iter == mel_banks_.end()) {
+    this_mel_banks = new MelBanks(opts_.mel_opts, opts_.frame_opts, vtln_warp);
+    mel_banks_[vtln_warp] = this_mel_banks;
+  } else {
+    this_mel_banks = iter->second;
+  }
+  return this_mel_banks;
+}
+
+void MfccComputer::Compute(float signal_raw_log_energy, float vtln_warp,
+                           std::vector<float> *signal_frame, float *feature) {
+  const MelBanks &mel_banks = *(GetMelBanks(vtln_warp));
+
+  KNF_CHECK_EQ(signal_frame->size(), opts_.frame_opts.PaddedWindowSize());
+
+  // Compute energy after window function (not the raw one).
+  if (opts_.use_energy && !opts_.raw_energy) {
+    signal_raw_log_energy = std::log(
+        std::max<float>(InnerProduct(signal_frame->data(), signal_frame->data(),
+                                     signal_frame->size()),
+                        std::numeric_limits<float>::epsilon()));
+  }
+  rfft_.Compute(signal_frame->data());  // signal_frame is modified in-place
+  ComputePowerSpectrum(signal_frame);
+
+  // Sum with mel filter banks over the power spectrum
+  mel_banks.Compute(signal_frame->data(), mel_energies_.data());
+
+  // Avoid log of zero (which should be prevented anyway by dithering).
+  for (int32_t i = 0; i != opts_.mel_opts.num_bins; ++i) {
+    auto t = std::max(mel_energies_[i], std::numeric_limits<float>::epsilon());
+    mel_energies_[i] = std::log(t);
+  }
+
+  // feature = dct_matrix_ * mel_energies [which now have log]
+  for (int32_t i = 0; i != opts_.num_ceps; ++i) {
+    feature[i] = InnerProduct(dct_matrix_.data() + i * opts_.mel_opts.num_bins,
+                              mel_energies_.data(), opts_.mel_opts.num_bins);
+  }
+
+  if (opts_.cepstral_lifter != 0.0) {
+    for (int32_t i = 0; i != opts_.num_ceps; ++i) {
+      feature[i] *= lifter_coeffs_[i];
+    }
+  }
+
+  if (opts_.use_energy) {
+    if (opts_.energy_floor > 0.0 && signal_raw_log_energy < log_energy_floor_) {
+      signal_raw_log_energy = log_energy_floor_;
+    }
+    feature[0] = signal_raw_log_energy;
+  }
+
+  if (opts_.htk_compat) {
+    double energy = feature[0];
+
+    for (int32_t i = 0; i < opts_.num_ceps - 1; ++i) {
+      feature[i] = feature[i + 1];
+    }
+
+    if (!opts_.use_energy) {
+      energy *= M_SQRT2;  // scale on C0 (actually removing a scale
+    }
+    // we previously added that's part of one common definition of
+    // the cosine transform.)
+    feature[opts_.num_ceps - 1] = static_cast<float>(energy);
+  }
+}
+
+}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.h
new file mode 100644
index 00000000..32dc772e
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.h
@@ -0,0 +1,151 @@
+/**
+ * Copyright 2009-2011  Karel Vesely;  Petr Motlicek;  Saarland University
+ *           2014-2016  Johns Hopkins University (author: Daniel Povey)
+ * Copyright 2024       Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// This file is copied/modified from kaldi/src/feat/feature-mfcc.h
+
+#ifndef KALDI_NATIVE_FBANK_CSRC_FEATURE_MFCC_H_
+#define KALDI_NATIVE_FBANK_CSRC_FEATURE_MFCC_H_
+
+#include <cstdint>
+#include <map>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#include "kaldi-native-fbank/csrc/feature-window.h"
+#include "kaldi-native-fbank/csrc/mel-computations.h"
+#include "kaldi-native-fbank/csrc/rfft.h"
+
+namespace knf {
+
+/// MfccOptions contains basic options for computing MFCC features.
+// (this class is copied from kaldi)
+struct MfccOptions {
+  FrameExtractionOptions frame_opts;
+  MelBanksOptions mel_opts;
+
+  // Number of cepstra in MFCC computation (including C0)
+  int32_t num_ceps = 13;
+
+  // Use energy (not C0) in MFCC computation
+  bool use_energy = true;
+
+  // Floor on energy (absolute, not relative) in MFCC
+  // computation. Only makes a difference if use_energy=true;
+  // only necessary if dither=0.0.
+  // Suggested values: 0.1 or 1.0
+  float energy_floor = 0.0;
+
+  // If true, compute energy before preemphasis and windowing
+  bool raw_energy = true;
+
+  // Constant that controls scaling of MFCCs
+  float cepstral_lifter = 22.0;
+
+  // If true, put energy or C0 last and use a factor of
+  // sqrt(2) on C0.
+  // Warning: not sufficient to get HTK compatible features
+  // (need to change other parameters)
+  bool htk_compat = false;
+
+  MfccOptions() { mel_opts.num_bins = 23; }
+
+  std::string ToString() const {
+    std::ostringstream os;
+    os << "MfccOptions(";
+    os << "frame_opts=" << frame_opts.ToString() << ", ";
+    os << "mel_opts=" << mel_opts.ToString() << ", ";
+
+    os << "num_ceps=" << num_ceps << ", ";
+    os << "use_energy=" << (use_energy ? "True" : "False") << ", ";
+    os << "energy_floor=" << energy_floor << ", ";
+    os << "raw_energy=" << (raw_energy ? "True" : "False") << ", ";
+    os << "cepstral_lifter=" << cepstral_lifter << ", ";
+    os << "htk_compat=" << (htk_compat ? "True" : "False") << ")";
+
+    return os.str();
+  }
+};
+
+std::ostream &operator<<(std::ostream &os, const MfccOptions &opts);
+
+class MfccComputer {
+ public:
+  using Options = MfccOptions;
+
+  explicit MfccComputer(const MfccOptions &opts);
+  ~MfccComputer();
+
+  int32_t Dim() const { return opts_.num_ceps; }
+
+  // if true, compute log_energy_pre_window but after dithering and dc removal
+  bool NeedRawLogEnergy() const { return opts_.use_energy && opts_.raw_energy; }
+
+  const FrameExtractionOptions &GetFrameOptions() const {
+    return opts_.frame_opts;
+  }
+
+  const MfccOptions &GetOptions() const { return opts_; }
+
+  /**
+     Function that computes one frame of features from
+     one frame of signal.
+
+     @param [in] signal_raw_log_energy The log-energy of the frame of the signal
+         prior to windowing and pre-emphasis, or
+         log(numeric_limits<float>::min()), whichever is greater.  Must be
+         ignored by this function if this class returns false from
+         this->NeedsRawLogEnergy().
+     @param [in] vtln_warp  The VTLN warping factor that the user wants
+         to be applied when computing features for this utterance.  Will
+         normally be 1.0, meaning no warping is to be done.  The value will
+         be ignored for feature types that don't support VLTN, such as
+         spectrogram features.
+     @param [in] signal_frame  One frame of the signal,
+       as extracted using the function ExtractWindow() using the options
+       returned by this->GetFrameOptions().  The function will use the
+       vector as a workspace, which is why it's a non-const pointer.
+     @param [out] feature  Pointer to a vector of size this->Dim(), to which
+         the computed feature will be written. It should be pre-allocated.
+  */
+  void Compute(float signal_raw_log_energy, float vtln_warp,
+               std::vector<float> *signal_frame, float *feature);
+
+ private:
+  const MelBanks *GetMelBanks(float vtln_warp);
+
+  MfccOptions opts_;
+  float log_energy_floor_;
+  std::map<float, MelBanks *> mel_banks_;  // float is VTLN coefficient.
+  Rfft rfft_;
+
+  // temp buffer of size num_mel_bins = opts.mel_opts.num_bins
+  std::vector<float> mel_energies_;
+
+  // opts_.num_ceps
+  std::vector<float> lifter_coeffs_;
+
+  // [num_ceps][num_mel_bins]
+  std::vector<float> dct_matrix_;
+};
+
+}  // namespace knf
+
+#endif  // KALDI_NATIVE_FBANK_CSRC_FEATURE_MFCC_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.cc
new file mode 100644
index 00000000..f1ab58ee
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.cc
@@ -0,0 +1,38 @@
+/**
+ * Copyright (c)  2025  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kaldi-native-fbank/csrc/feature-raw-audio-samples.h"
+
+#include <algorithm>
+#include <vector>
+
+namespace knf {
+
+std::ostream &operator<<(std::ostream &os, const RawAudioSamplesOptions &opts) {
+  os << opts.ToString();
+  return os;
+}
+
+void RawAudioSamplesComputer::Compute(float /*signal_raw_log_energy*/,
+                                      float /*vtln_warp*/,
+                                      std::vector<float> *signal_frame,
+                                      float *feature) {
+  std::copy(signal_frame->begin(), signal_frame->end(), feature);
+}
+
+}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.h
new file mode 100644
index 00000000..5cf5e45f
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.h
@@ -0,0 +1,81 @@
+/**
+ * Copyright (c)  2025  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef KALDI_NATIVE_FBANK_CSRC_FEATURE_RAW_AUDIO_SAMPLES_H_
+#define KALDI_NATIVE_FBANK_CSRC_FEATURE_RAW_AUDIO_SAMPLES_H_
+
+#include <cstdint>
+#include <sstream>
+#include <vector>
+
+#include "kaldi-native-fbank/csrc/feature-window.h"
+
+namespace knf {
+
+struct RawAudioSamplesOptions {
+  FrameExtractionOptions frame_opts;
+
+  RawAudioSamplesOptions() {
+    frame_opts.window_type = "rectangular";
+    frame_opts.dither = 0;
+    frame_opts.preemph_coeff = 0;
+    frame_opts.remove_dc_offset = false;
+    frame_opts.round_to_power_of_two = false;
+    frame_opts.snip_edges = true;
+  }
+
+  std::string ToString() const {
+    std::ostringstream os;
+    os << "frame_opts: \n";
+    os << frame_opts << "\n";
+    os << "\n";
+    return os.str();
+  }
+};
+
+std::ostream &operator<<(std::ostream &os, const RawAudioSamplesOptions &opts);
+
+class RawAudioSamplesComputer {
+ public:
+  using Options = RawAudioSamplesOptions;
+
+  explicit RawAudioSamplesComputer(const RawAudioSamplesOptions &opts)
+      : opts_(opts){};
+
+  int32_t Dim() const {
+    return opts_.frame_opts.frame_length_ms * opts_.frame_opts.samp_freq / 1000;
+  }
+
+  bool NeedRawLogEnergy() const { return false; }
+
+  const FrameExtractionOptions &GetFrameOptions() const {
+    return opts_.frame_opts;
+  }
+
+  const RawAudioSamplesOptions &GetOptions() const { return opts_; }
+
+  void Compute(float unused_signal_raw_log_energy, float unused_vtln_warp,
+               std::vector<float> *signal_frame, float *feature);
+
+ private:
+  RawAudioSamplesOptions opts_;
+};
+
+}  // namespace knf
+
+#endif  // KALDI_NATIVE_FBANK_CSRC_FEATURE_RAW_AUDIO_SAMPLES_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.cc
new file mode 100644
index 00000000..047abaf2
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.cc
@@ -0,0 +1,274 @@
+// kaldi-native-fbank/csrc/feature-window.cc
+//
+// Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+
+// This file is copied/modified from kaldi/src/feat/feature-window.cc
+
+#include "kaldi-native-fbank/csrc/feature-window.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstddef>
+#include <limits>
+#include <vector>
+
+#include "kaldi-native-fbank/csrc/kaldi-math.h"
+
+namespace knf {
+
+std::ostream &operator<<(std::ostream &os, const FrameExtractionOptions &opts) {
+  os << opts.ToString();
+  return os;
+}
+
+std::vector<float> GetWindow(const std::string &window_type,
+                             int32_t window_size,
+                             float blackman_coeff /*= 0.42*/) {
+  std::vector<float> window(window_size);
+  int32_t frame_length = window_size;
+  KNF_CHECK_GT(frame_length, 0);
+
+  float *window_data = window.data();
+
+  double a = M_2PI / (frame_length - 1);
+  if (window_type == "hann") {
+    // see https://pytorch.org/docs/stable/generated/torch.hann_window.html
+    // We assume periodic is true
+    a = M_2PI / frame_length;
+  }
+
+  for (int32_t i = 0; i < frame_length; i++) {
+    double i_fl = static_cast<double>(i);
+    if (window_type == "hanning") {
+      window_data[i] = 0.5 - 0.5 * cos(a * i_fl);
+    } else if (window_type == "sine") {
+      // when you are checking ws wikipedia, please
+      // note that 0.5 * a = M_PI/(frame_length-1)
+      window_data[i] = sin(0.5 * a * i_fl);
+    } else if (window_type == "hamming") {
+      window_data[i] = 0.54 - 0.46 * cos(a * i_fl);
+    } else if (window_type == "hann") {
+      window_data[i] = 0.50 - 0.50 * cos(a * i_fl);
+    } else if (window_type == "povey") {
+      // like hamming but goes to zero at edges.
+      window_data[i] = pow(0.5 - 0.5 * cos(a * i_fl), 0.85);
+    } else if (window_type == "rectangular") {
+      window_data[i] = 1.0;
+    } else if (window_type == "blackman") {
+      window_data[i] = blackman_coeff - 0.5 * cos(a * i_fl) +
+                       (0.5 - blackman_coeff) * cos(2 * a * i_fl);
+    } else {
+      fprintf(stderr, "Invalid window type '%s'\n", window_type.c_str());
+      exit(-1);
+    }
+  }
+
+  return window;
+}
+
+FeatureWindowFunction::FeatureWindowFunction(const FrameExtractionOptions &opts)
+    : FeatureWindowFunction(opts.window_type, opts.WindowSize(),
+                            opts.blackman_coeff) {}
+
+FeatureWindowFunction::FeatureWindowFunction(const std::string &window_type,
+                                             int32_t window_size,
+                                             float blackman_coeff /*= 0.42*/)
+    : window_(knf::GetWindow(window_type, window_size, blackman_coeff)) {}
+
+FeatureWindowFunction::FeatureWindowFunction(const std::vector<float> &window)
+    : window_(window) {}
+
+void FeatureWindowFunction::Apply(float *wave) const {
+  int32_t window_size = window_.size();
+  const float *p = window_.data();
+  for (int32_t k = 0; k != window_size; ++k) {
+    wave[k] *= p[k];
+  }
+}
+
+int64_t FirstSampleOfFrame(int32_t frame, const FrameExtractionOptions &opts) {
+  int64_t frame_shift = opts.WindowShift();
+  if (opts.snip_edges) {
+    return frame * frame_shift;
+  } else {
+    int64_t midpoint_of_frame = frame_shift * frame + frame_shift / 2,
+            beginning_of_frame = midpoint_of_frame - opts.WindowSize() / 2;
+    return beginning_of_frame;
+  }
+}
+
+int32_t NumFrames(int64_t num_samples, const FrameExtractionOptions &opts,
+                  bool flush /*= true*/) {
+  int64_t frame_shift = opts.WindowShift();
+  int64_t frame_length = opts.WindowSize();
+  if (opts.snip_edges) {
+    // with --snip-edges=true (the default), we use a HTK-like approach to
+    // determining the number of frames-- all frames have to fit completely into
+    // the waveform, and the first frame begins at sample zero.
+    if (num_samples < frame_length)
+      return 0;
+    else
+      return (1 + ((num_samples - frame_length) / frame_shift));
+    // You can understand the expression above as follows: 'num_samples -
+    // frame_length' is how much room we have to shift the frame within the
+    // waveform; 'frame_shift' is how much we shift it each time; and the ratio
+    // is how many times we can shift it (integer arithmetic rounds down).
+  } else {
+    // if --snip-edges=false, the number of frames is determined by rounding the
+    // (file-length / frame-shift) to the nearest integer.  The point of this
+    // formula is to make the number of frames an obvious and predictable
+    // function of the frame shift and signal length, which makes many
+    // segmentation-related questions simpler.
+    //
+    // Because integer division in C++ rounds toward zero, we add (half the
+    // frame-shift minus epsilon) before dividing, to have the effect of
+    // rounding towards the closest integer.
+    int32_t num_frames = (num_samples + (frame_shift / 2)) / frame_shift;
+
+    if (flush) return num_frames;
+
+    // note: 'end' always means the last plus one, i.e. one past the last.
+    int64_t end_sample_of_last_frame =
+        FirstSampleOfFrame(num_frames - 1, opts) + frame_length;
+
+    // the following code is optimized more for clarity than efficiency.
+    // If flush == false, we can't output frames that extend past the end
+    // of the signal.
+    while (num_frames > 0 && end_sample_of_last_frame > num_samples) {
+      num_frames--;
+      end_sample_of_last_frame -= frame_shift;
+    }
+    return num_frames;
+  }
+}
+
+void ExtractWindow(int64_t sample_offset, const std::vector<float> &wave,
+                   int32_t f, const FrameExtractionOptions &opts,
+                   const FeatureWindowFunction &window_function,
+                   std::vector<float> *window,
+                   float *log_energy_pre_window /*= nullptr*/) {
+  KNF_CHECK(sample_offset >= 0 && wave.size() != 0);
+
+  int32_t frame_length = opts.WindowSize();
+  int32_t frame_length_padded = opts.PaddedWindowSize();
+
+  int64_t num_samples = sample_offset + wave.size();
+  int64_t start_sample = FirstSampleOfFrame(f, opts);
+  int64_t end_sample = start_sample + frame_length;
+
+  if (opts.snip_edges) {
+    KNF_CHECK(start_sample >= sample_offset && end_sample <= num_samples);
+  } else {
+    KNF_CHECK(sample_offset == 0 || start_sample >= sample_offset);
+  }
+
+  if (window->size() != frame_length_padded) {
+    window->resize(frame_length_padded);
+  }
+
+  // wave_start and wave_end are start and end indexes into 'wave', for the
+  // piece of wave that we're trying to extract.
+  int32_t wave_start = int32_t(start_sample - sample_offset);
+  int32_t wave_end = wave_start + frame_length;
+
+  if (wave_start >= 0 && wave_end <= wave.size()) {
+    // the normal case-- no edge effects to consider.
+    std::copy(wave.begin() + wave_start,
+              wave.begin() + wave_start + frame_length, window->data());
+  } else {
+    // Deal with any end effects by reflection, if needed.  This code will only
+    // be reached for about two frames per utterance, so we don't concern
+    // ourselves excessively with efficiency.
+    int32_t wave_dim = wave.size();
+    for (int32_t s = 0; s < frame_length; ++s) {
+      int32_t s_in_wave = s + wave_start;
+      while (s_in_wave < 0 || s_in_wave >= wave_dim) {
+        // reflect around the beginning or end of the wave.
+        // e.g. -1 -> 0, -2 -> 1.
+        // dim -> dim - 1, dim + 1 -> dim - 2.
+        // the code supports repeated reflections, although this
+        // would only be needed in pathological cases.
+        if (s_in_wave < 0)
+          s_in_wave = -s_in_wave - 1;
+        else
+          s_in_wave = 2 * wave_dim - 1 - s_in_wave;
+      }
+      (*window)[s] = wave[s_in_wave];
+    }
+  }
+
+  ProcessWindow(opts, window_function, window->data(), log_energy_pre_window);
+}
+
+static void RemoveDcOffset(float *d, int32_t n) {
+  float sum = 0;
+  for (int32_t i = 0; i != n; ++i) {
+    sum += d[i];
+  }
+
+  float mean = sum / n;
+
+  for (int32_t i = 0; i != n; ++i) {
+    d[i] -= mean;
+  }
+}
+
+float InnerProduct(const float *a, const float *b, int32_t n) {
+  float sum = 0;
+  for (int32_t i = 0; i != n; ++i) {
+    sum += a[i] * b[i];
+  }
+  return sum;
+}
+
+void Dither(float *d, int32_t n, float dither_value) {
+  if (dither_value == 0.0) {
+    return;
+  }
+
+  RandomState rstate;
+  for (int32_t i = 0; i < n; ++i) {
+    d[i] += RandGauss(&rstate) * dither_value;
+  }
+}
+
+static void Preemphasize(float *d, int32_t n, float preemph_coeff) {
+  if (preemph_coeff == 0.0) {
+    return;
+  }
+
+  KNF_CHECK(preemph_coeff >= 0.0 && preemph_coeff <= 1.0);
+
+  for (int32_t i = n - 1; i > 0; --i) {
+    d[i] -= preemph_coeff * d[i - 1];
+  }
+  d[0] -= preemph_coeff * d[0];
+}
+
+void ProcessWindow(const FrameExtractionOptions &opts,
+                   const FeatureWindowFunction &window_function, float *window,
+                   float *log_energy_pre_window /*= nullptr*/) {
+  int32_t frame_length = opts.WindowSize();
+
+  if (opts.dither != 0.0) {
+    Dither(window, frame_length, opts.dither);
+  }
+
+  if (opts.remove_dc_offset) {
+    RemoveDcOffset(window, frame_length);
+  }
+
+  if (log_energy_pre_window != NULL) {
+    float energy = std::max<float>(InnerProduct(window, window, frame_length),
+                                   std::numeric_limits<float>::epsilon());
+    *log_energy_pre_window = std::log(energy);
+  }
+
+  if (opts.preemph_coeff != 0.0) {
+    Preemphasize(window, frame_length, opts.preemph_coeff);
+  }
+
+  window_function.Apply(window);
+}
+
+}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.h
new file mode 100644
index 00000000..b8d1d44c
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.h
@@ -0,0 +1,187 @@
+// kaldi-native-fbank/csrc/feature-window.h
+//
+// Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+
+// This file is copied/modified from kaldi/src/feat/feature-window.h
+
+#ifndef KALDI_NATIVE_FBANK_CSRC_FEATURE_WINDOW_H_
+#define KALDI_NATIVE_FBANK_CSRC_FEATURE_WINDOW_H_
+
+#include <cstdint>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#include "kaldi-native-fbank/csrc/log.h"
+
+namespace knf {
+
+inline int32_t RoundUpToNearestPowerOfTwo(int32_t n) {
+  // copied from kaldi/src/base/kaldi-math.cc
+  KNF_CHECK_GT(n, 0);
+  n--;
+  n |= n >> 1;
+  n |= n >> 2;
+  n |= n >> 4;
+  n |= n >> 8;
+  n |= n >> 16;
+  return n + 1;
+}
+
+struct FrameExtractionOptions {
+  float samp_freq = 16000;
+  float frame_shift_ms = 10.0f;   // in milliseconds.
+  float frame_length_ms = 25.0f;  // in milliseconds.
+
+  float dither = 0.00003f;  // Amount of dithering, 0.0 means no dither.
+                            // Value 0.00003f is equivalent to 1.0 in kaldi.
+
+  float preemph_coeff = 0.97f;        // Preemphasis coefficient.
+  bool remove_dc_offset = true;       // Subtract mean of wave before FFT.
+  std::string window_type = "povey";  // e.g. Hamming window
+  // May be "hamming", "rectangular", "povey", "hanning", "hann", "sine",
+  // "blackman".
+  // "povey" is a window I made to be similar to Hamming but to go to zero at
+  // the edges, it's pow((0.5 - 0.5*cos(n/N*2*pi)), 0.85) I just don't think the
+  // Hamming window makes sense as a windowing function.
+  bool round_to_power_of_two = true;
+  float blackman_coeff = 0.42f;
+  bool snip_edges = true;
+  // bool allow_downsample = false;
+  // bool allow_upsample = false;
+
+  int32_t WindowShift() const {
+    return static_cast<int32_t>(samp_freq * 0.001f * frame_shift_ms);
+  }
+  int32_t WindowSize() const {
+    return static_cast<int32_t>(samp_freq * 0.001f * frame_length_ms);
+  }
+  int32_t PaddedWindowSize() const {
+    return (round_to_power_of_two ? RoundUpToNearestPowerOfTwo(WindowSize())
+                                  : WindowSize());
+  }
+  std::string ToString() const {
+    std::ostringstream os;
+#define KNF_PRINT(x) os << #x << ": " << x << "\n"
+    KNF_PRINT(samp_freq);
+    KNF_PRINT(frame_shift_ms);
+    KNF_PRINT(frame_length_ms);
+    KNF_PRINT(dither);
+    KNF_PRINT(preemph_coeff);
+    KNF_PRINT(remove_dc_offset);
+    KNF_PRINT(window_type);
+    KNF_PRINT(round_to_power_of_two);
+    KNF_PRINT(blackman_coeff);
+    KNF_PRINT(snip_edges);
+    // KNF_PRINT(allow_downsample);
+    // KNF_PRINT(allow_upsample);
+#undef KNF_PRINT
+    return os.str();
+  }
+};
+
+std::ostream &operator<<(std::ostream &os, const FrameExtractionOptions &opts);
+
+class FeatureWindowFunction {
+ public:
+  FeatureWindowFunction() = default;
+  explicit FeatureWindowFunction(const FrameExtractionOptions &opts);
+  FeatureWindowFunction(const std::string &window_type, int32_t window_size,
+                        float blackman_coeff = 0.42);
+
+  explicit FeatureWindowFunction(const std::vector<float> &window);
+
+  /**
+   * @param wave Pointer to a 1-D array of shape [window_size].
+   *             It is modified in-place: wave[i] = wave[i] * window_[i].
+   * @param
+   */
+  void Apply(float *wave) const;
+
+  const std::vector<float> &GetWindow() const { return window_; }
+
+ private:
+  std::vector<float> window_;  // of size opts.WindowSize()
+};
+
+int64_t FirstSampleOfFrame(int32_t frame, const FrameExtractionOptions &opts);
+
+/**
+   This function returns the number of frames that we can extract from a wave
+   file with the given number of samples in it (assumed to have the same
+   sampling rate as specified in 'opts').
+
+      @param [in] num_samples  The number of samples in the wave file.
+      @param [in] opts     The frame-extraction options class
+
+      @param [in] flush   True if we are asserting that this number of samples
+   is 'all there is', false if we expecting more data to possibly come in.  This
+   only makes a difference to the answer
+   if opts.snip_edges== false.  For offline feature extraction you always want
+   flush == true.  In an online-decoding context, once you know (or decide) that
+   no more data is coming in, you'd call it with flush == true at the end to
+   flush out any remaining data.
+*/
+int32_t NumFrames(int64_t num_samples, const FrameExtractionOptions &opts,
+                  bool flush = true);
+
+/*
+  ExtractWindow() extracts a windowed frame of waveform (possibly with a
+  power-of-two, padded size, depending on the config), including all the
+  processing done by ProcessWindow().
+
+  @param [in] sample_offset  If 'wave' is not the entire waveform, but
+                   part of it to the left has been discarded, then the
+                   number of samples prior to 'wave' that we have
+                   already discarded.  Set this to zero if you are
+                   processing the entire waveform in one piece, or
+                   if you get 'no matching function' compilation
+                   errors when updating the code.
+  @param [in] wave  The waveform
+  @param [in] f     The frame index to be extracted, with
+                    0 <= f < NumFrames(sample_offset + wave.Dim(), opts, true)
+  @param [in] opts  The options class to be used
+  @param [in] window_function  The windowing function, as derived from the
+                    options class.
+  @param [out] window  The windowed, possibly-padded waveform to be
+                     extracted.  Will be resized as needed.
+  @param [out] log_energy_pre_window  If non-NULL, the log-energy of
+                   the signal prior to pre-emphasis and multiplying by
+                   the windowing function will be written to here.
+*/
+void ExtractWindow(int64_t sample_offset, const std::vector<float> &wave,
+                   int32_t f, const FrameExtractionOptions &opts,
+                   const FeatureWindowFunction &window_function,
+                   std::vector<float> *window,
+                   float *log_energy_pre_window = nullptr);
+
+/**
+  This function does all the windowing steps after actually
+  extracting the windowed signal: depending on the
+  configuration, it does dithering, dc offset removal,
+  preemphasis, and multiplication by the windowing function.
+   @param [in] opts  The options class to be used
+   @param [in] window_function  The windowing function-- should have
+                    been initialized using 'opts'.
+   @param [in,out] window  A vector of size opts.WindowSize().  Note:
+      it will typically be a sub-vector of a larger vector of size
+      opts.PaddedWindowSize(), with the remaining samples zero,
+      as the FFT code is more efficient if it operates on data with
+      power-of-two size.
+   @param [out]   log_energy_pre_window If non-NULL, then after dithering and
+      DC offset removal, this function will write to this pointer the log of
+      the total energy (i.e. sum-squared) of the frame.
+ */
+void ProcessWindow(const FrameExtractionOptions &opts,
+                   const FeatureWindowFunction &window_function, float *window,
+                   float *log_energy_pre_window = nullptr);
+
+// Compute the inner product of two vectors
+float InnerProduct(const float *a, const float *b, int32_t n);
+
+std::vector<float> GetWindow(const std::string &window_type,
+                             int32_t window_size, float blackman_coeff = 0.42);
+
+}  // namespace knf
+
+#endif  // KALDI_NATIVE_FBANK_CSRC_FEATURE_WINDOW_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.cc
new file mode 100644
index 00000000..e779ac74
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.cc
@@ -0,0 +1,158 @@
+/**
+ * Copyright (c)  2025  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "kaldi-native-fbank/csrc/istft.h"
+
+#include <algorithm>
+#include <cmath>
+#include <sstream>
+#include <string>
+#include <utility>
+
+#include "kaldi-native-fbank/csrc/feature-window.h"
+#include "kaldi-native-fbank/csrc/rfft.h"
+
+namespace knf {
+
+class IStft::Impl {
+ public:
+  explicit Impl(const StftConfig &config) : config_(config) {
+    if (!config.window.empty()) {
+      window_ = std::make_unique<FeatureWindowFunction>(config_.window);
+    } else if (!config.window_type.empty()) {
+      window_ = std::make_unique<FeatureWindowFunction>(config.window_type,
+                                                        config.win_length);
+    }
+  }
+
+  std::vector<float> Compute(const StftResult &stft_result) const {
+    Rfft rfft(config_.n_fft, true);
+
+    int32_t num_samples =
+        config_.n_fft + (stft_result.num_frames - 1) * config_.hop_length;
+
+    std::vector<float> samples(num_samples);
+    for (int32_t i = 0; i < stft_result.num_frames; ++i) {
+      auto x = InverseFFT(stft_result, i, &rfft);
+      OverlapAdd(std::move(x), i, &samples);
+    }
+
+    auto denominator = GetDenominator(stft_result.num_frames);
+
+    for (int32_t i = 0; i < num_samples; ++i) {
+      if (denominator[i]) {
+        samples[i] /= denominator[i];
+      }
+    }
+
+    if (config_.center) {
+      samples = {samples.begin() + config_.n_fft / 2,
+                 samples.end() - config_.n_fft / 2};
+    }
+
+    return samples;
+  }
+
+  std::vector<float> InverseFFT(const StftResult &r, int32_t frame_index,
+                                Rfft *rfft) const {
+    int32_t n_fft = config_.n_fft;
+    int32_t hop_length = config_.hop_length;
+
+    const float *p_real = r.real.data() + frame_index * (n_fft / 2 + 1);
+    const float *p_imag = r.imag.data() + frame_index * (n_fft / 2 + 1);
+    std::vector<float> tmp(n_fft);
+
+    float scale = 1;
+    if (config_.normalized) {
+      scale = std::sqrt(n_fft);
+    }
+
+    for (int32_t i = 0; i < n_fft / 2; ++i) {
+      if (i == 0) {
+        tmp[0] = p_real[0] * scale;
+        tmp[1] = p_real[n_fft / 2] * scale;
+      } else {
+        tmp[2 * i] = p_real[i] * scale;
+        tmp[2 * i + 1] = p_imag[i] * scale;
+      }
+    }
+
+    rfft->Compute(tmp.data());
+
+    scale = 1.0f / n_fft;
+    for (auto &f : tmp) {
+      f *= scale;
+    }
+
+    return tmp;
+  }
+
+  void OverlapAdd(std::vector<float> current_frame, int32_t frame_index,
+                  std::vector<float> *samples) const {
+    if (window_) {
+      window_->Apply(current_frame.data());
+    }
+
+    float *p = samples->data() + frame_index * config_.hop_length;
+    for (int32_t i = 0; i < config_.n_fft; ++i) {
+      p[i] += current_frame[i];
+    }
+  }
+
+  std::vector<float> GetDenominator(int32_t num_frames) const {
+    int32_t num_samples = config_.n_fft + (num_frames - 1) * config_.hop_length;
+    std::vector<float> ans(num_samples);
+    if (!window_) {
+      for (int32_t i = 0; i < num_frames; ++i) {
+        int32_t start = i * config_.hop_length;
+        int32_t end = start + config_.n_fft;
+
+        for (int32_t k = start; k < end; ++k) {
+          ans[k] += 1;
+        }
+      }
+    } else {
+      const auto &w = window_->GetWindow();
+      for (int32_t i = 0; i < num_frames; ++i) {
+        int32_t start = i * config_.hop_length;
+        int32_t end = start + config_.n_fft;
+
+        auto pw = w.data();
+        for (int32_t k = start; k < end; ++k, ++pw) {
+          ans[k] += (*pw) * (*pw);
+        }
+      }
+    }
+
+    return ans;
+  }
+
+ private:
+  StftConfig config_;
+  std::unique_ptr<FeatureWindowFunction> window_;
+};
+
+IStft::IStft(const StftConfig &config)
+    : impl_(std::make_unique<Impl>(config)) {}
+
+IStft::~IStft() = default;
+
+std::vector<float> IStft::Compute(const StftResult &stft_result) const {
+  return impl_->Compute(stft_result);
+}
+
+}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.h
new file mode 100644
index 00000000..e5eba1b9
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.h
@@ -0,0 +1,45 @@
+/**
+ * Copyright (c)  2025  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef KALDI_NATIVE_FBANK_CSRC_ISTFT_H_
+#define KALDI_NATIVE_FBANK_CSRC_ISTFT_H_
+
+#include <cstdint>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "kaldi-native-fbank/csrc/stft.h"
+
+namespace knf {
+
+class IStft {
+ public:
+  explicit IStft(const StftConfig &config);
+  ~IStft();
+  std::vector<float> Compute(const StftResult &stft_result) const;
+
+ private:
+  class Impl;
+  std::unique_ptr<Impl> impl_;
+};
+
+}  // namespace knf
+
+#endif  // KALDI_NATIVE_FBANK_CSRC_ISTFT_H_
+        //
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.cc
new file mode 100644
index 00000000..bcc0dcc3
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.cc
@@ -0,0 +1,51 @@
+// kaldi-native-fbank/csrc/kaldi-math.cc
+//
+// Copyright (c)  2024  Brno University of Technology (authors: Karel Vesely)
+
+// This file is an excerpt from kaldi/src/base/kaldi-math.cc
+
+#include "kaldi-native-fbank/csrc/kaldi-math.h"
+
+#if defined(_POSIX_THREAD_SAFE_FUNCTIONS)
+#include <mutex>  // NOLINT
+#endif
+
+#include <cmath>
+
+namespace knf {
+
+int Rand(struct RandomState *state) {
+#if !defined(_POSIX_THREAD_SAFE_FUNCTIONS)
+  // On Windows and Cygwin, just call Rand()
+  return rand();
+#else
+  if (state) {
+    return rand_r(&(state->seed));
+  } else {
+    static std::mutex _RandMutex;
+    std::lock_guard<std::mutex> lock(_RandMutex);
+    return rand();
+  }
+#endif
+}
+
+RandomState::RandomState() {
+  // we initialize it as Rand() + 27437 instead of just Rand(), because on some
+  // systems, e.g. at the very least Mac OSX Yosemite and later, it seems to be
+  // the case that rand_r when initialized with rand() will give you the exact
+  // same sequence of numbers that rand() will give if you keep calling rand()
+  // after that initial call.  This can cause problems with repeated sequences.
+  // For example if you initialize two RandomState structs one after the other
+  // without calling rand() in between, they would give you the same sequence
+  // offset by one (if we didn't have the "+ 27437" in the code).  27437 is just
+  // a randomly chosen prime number.
+  seed = unsigned(Rand()) + 27437;
+}
+
+void Sqrt(float *in_out, int32_t n) {
+  for (int32_t i = 0; i != n; ++i) {
+    in_out[i] = std::sqrt(in_out[i]);
+  }
+}
+
+}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.h
new file mode 100644
index 00000000..4db3f831
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.h
@@ -0,0 +1,52 @@
+// kaldi-native-fbank/csrc/kaldi-math.h
+//
+// Copyright (c)  2024  Brno University of Technology (authors: Karel Vesely)
+
+// This file is an excerpt from kaldi/src/base/kaldi-math.h
+
+#ifndef KALDI_NATIVE_FBANK_CSRC_KALDI_MATH_H_
+#define KALDI_NATIVE_FBANK_CSRC_KALDI_MATH_H_
+
+#include <cmath>  // logf, sqrtf, cosf
+#include <cstdint>
+#include <cstdlib>  // RAND_MAX
+
+#ifndef M_PI
+#define M_PI 3.1415926535897932384626433832795
+#endif
+
+#ifndef M_2PI
+#define M_2PI 6.283185307179586476925286766559005
+#endif
+
+#ifndef M_SQRT2
+#define M_SQRT2 1.4142135623730950488016887
+#endif
+
+namespace knf {
+
+inline float Log(float x) { return logf(x); }
+
+// Returns a random integer between 0 and RAND_MAX, inclusive
+int Rand(struct RandomState *state = NULL);
+
+// State for thread-safe random number generator
+struct RandomState {
+  RandomState();
+  unsigned seed;
+};
+
+/// Returns a random number strictly between 0 and 1.
+inline float RandUniform(struct RandomState *state = NULL) {
+  return static_cast<float>((Rand(state) + 1.0) / (RAND_MAX + 2.0));
+}
+
+inline float RandGauss(struct RandomState *state = NULL) {
+  return static_cast<float>(sqrtf(-2 * Log(RandUniform(state))) *
+                            cosf(2 * M_PI * RandUniform(state)));
+}
+
+void Sqrt(float *in_out, int32_t n);
+
+}  // namespace knf
+#endif  // KALDI_NATIVE_FBANK_CSRC_KALDI_MATH_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.cc
new file mode 100644
index 00000000..7223337b
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.cc
@@ -0,0 +1,142 @@
+/**
+ * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/*
+ * Stack trace related stuff is from kaldi.
+ * Refer to
+ * https://github.com/kaldi-asr/kaldi/blob/master/src/base/kaldi-error.cc
+ */
+
+#include "kaldi-native-fbank/csrc/log.h"
+
+#ifdef KNF_HAVE_EXECINFO_H
+#include <execinfo.h>  // To get stack trace in error messages.
+#ifdef KNF_HAVE_CXXABI_H
+#include <cxxabi.h>  // For name demangling.
+// Useful to decode the stack trace, but only used if we have execinfo.h
+#endif  // KNF_HAVE_CXXABI_H
+#endif  // KNF_HAVE_EXECINFO_H
+
+#include <stdlib.h>
+
+#include <ctime>
+#include <iomanip>
+#include <string>
+
+namespace knf {
+
+std::string GetDateTimeStr() {
+  std::ostringstream os;
+  std::time_t t = std::time(nullptr);
+  std::tm tm = *std::localtime(&t);
+  os << std::put_time(&tm, "%F %T");  // yyyy-mm-dd hh:mm:ss
+  return os.str();
+}
+
+static bool LocateSymbolRange(const std::string &trace_name, std::size_t *begin,
+                              std::size_t *end) {
+  // Find the first '_' with leading ' ' or '('.
+  *begin = std::string::npos;
+  for (std::size_t i = 1; i < trace_name.size(); ++i) {
+    if (trace_name[i] != '_') {
+      continue;
+    }
+    if (trace_name[i - 1] == ' ' || trace_name[i - 1] == '(') {
+      *begin = i;
+      break;
+    }
+  }
+  if (*begin == std::string::npos) {
+    return false;
+  }
+  *end = trace_name.find_first_of(" +", *begin);
+  return *end != std::string::npos;
+}
+
+#ifdef KNF_HAVE_EXECINFO_H
+static std::string Demangle(const std::string &trace_name) {
+#ifndef KNF_HAVE_CXXABI_H
+  return trace_name;
+#else   // KNF_HAVE_CXXABI_H
+  // Try demangle the symbol. We are trying to support the following formats
+  // produced by different platforms:
+  //
+  // Linux:
+  //   ./kaldi-error-test(_ZN5kaldi13UnitTestErrorEv+0xb) [0x804965d]
+  //
+  // Mac:
+  //   0 server 0x000000010f67614d _ZNK5kaldi13MessageLogger10LogMessageEv + 813
+  //
+  // We want to extract the name e.g., '_ZN5kaldi13UnitTestErrorEv' and
+  // demangle it info a readable name like kaldi::UnitTextError.
+  std::size_t begin, end;
+  if (!LocateSymbolRange(trace_name, &begin, &end)) {
+    return trace_name;
+  }
+  std::string symbol = trace_name.substr(begin, end - begin);
+  int status;
+  char *demangled_name = abi::__cxa_demangle(symbol.c_str(), 0, 0, &status);
+  if (status == 0 && demangled_name != nullptr) {
+    symbol = demangled_name;
+    free(demangled_name);
+  }
+  return trace_name.substr(0, begin) + symbol +
+         trace_name.substr(end, std::string::npos);
+#endif  // KNF_HAVE_CXXABI_H
+}
+#endif  // KNF_HAVE_EXECINFO_H
+
+std::string GetStackTrace() {
+  std::string ans;
+#ifdef KNF_HAVE_EXECINFO_H
+  constexpr const std::size_t kMaxTraceSize = 50;
+  constexpr const std::size_t kMaxTracePrint = 50;  // Must be even.
+                                                    // Buffer for the trace.
+  void *trace[kMaxTraceSize];
+  // Get the trace.
+  std::size_t size = backtrace(trace, kMaxTraceSize);
+  // Get the trace symbols.
+  char **trace_symbol = backtrace_symbols(trace, size);
+  if (trace_symbol == nullptr) return ans;
+
+  // Compose a human-readable backtrace string.
+  ans += "[ Stack-Trace: ]\n";
+  if (size <= kMaxTracePrint) {
+    for (std::size_t i = 0; i < size; ++i) {
+      ans += Demangle(trace_symbol[i]) + "\n";
+    }
+  } else {  // Print out first+last (e.g.) 5.
+    for (std::size_t i = 0; i < kMaxTracePrint / 2; ++i) {
+      ans += Demangle(trace_symbol[i]) + "\n";
+    }
+    ans += ".\n.\n.\n";
+    for (std::size_t i = size - kMaxTracePrint / 2; i < size; ++i) {
+      ans += Demangle(trace_symbol[i]) + "\n";
+    }
+    if (size == kMaxTraceSize)
+      ans += ".\n.\n.\n";  // Stack was too long, probably a bug.
+  }
+
+  // We must free the array of pointers allocated by backtrace_symbols(),
+  // but not the strings themselves.
+  free(trace_symbol);
+#endif  // KNF_HAVE_EXECINFO_H
+  return ans;
+}
+
+}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.h
new file mode 100644
index 00000000..bd21cc35
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.h
@@ -0,0 +1,383 @@
+/**
+ * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// The content in this file is copied/modified from
+// https://github.com/k2-fsa/k2/blob/master/k2/csrc/log.h
+#ifndef KALDI_NATIVE_FBANK_CSRC_LOG_H_
+#define KALDI_NATIVE_FBANK_CSRC_LOG_H_
+
+#include <stdio.h>
+
+#include <mutex>  // NOLINT
+#include <sstream>
+#include <string>
+
+namespace knf {
+
+#if KNF_ENABLE_CHECK
+
+#if defined(NDEBUG)
+constexpr bool kDisableDebug = true;
+#else
+constexpr bool kDisableDebug = false;
+#endif
+
+enum class LogLevel {
+  kTrace = 0,
+  kDebug = 1,
+  kInfo = 2,
+  kWarning = 3,
+  kError = 4,
+  kFatal = 5,  // print message and abort the program
+};
+
+// They are used in KNF_LOG(xxx), so their names
+// do not follow the google c++ code style
+//
+// You can use them in the following way:
+//
+//  KNF_LOG(TRACE) << "some message";
+//  KNF_LOG(DEBUG) << "some message";
+#ifndef _MSC_VER
+constexpr LogLevel TRACE = LogLevel::kTrace;
+constexpr LogLevel DEBUG = LogLevel::kDebug;
+constexpr LogLevel INFO = LogLevel::kInfo;
+constexpr LogLevel WARNING = LogLevel::kWarning;
+constexpr LogLevel ERROR = LogLevel::kError;
+constexpr LogLevel FATAL = LogLevel::kFatal;
+#else
+#define TRACE LogLevel::kTrace
+#define DEBUG LogLevel::kDebug
+#define INFO LogLevel::kInfo
+#define WARNING LogLevel::kWarning
+#define ERROR LogLevel::kError
+#define FATAL LogLevel::kFatal
+#endif
+
+std::string GetStackTrace();
+
+/* Return the current log level.
+
+
+   If the current log level is TRACE, then all logged messages are printed out.
+
+   If the current log level is DEBUG, log messages with "TRACE" level are not
+   shown and all other levels are printed out.
+
+   Similarly, if the current log level is INFO, log message with "TRACE" and
+   "DEBUG" are not shown and all other levels are printed out.
+
+   If it is FATAL, then only FATAL messages are shown.
+ */
+inline LogLevel GetCurrentLogLevel() {
+  static LogLevel log_level = INFO;
+  static std::once_flag init_flag;
+  std::call_once(init_flag, []() {
+    const char *env_log_level = std::getenv("KNF_LOG_LEVEL");
+    if (env_log_level == nullptr) return;
+
+    std::string s = env_log_level;
+    if (s == "TRACE")
+      log_level = TRACE;
+    else if (s == "DEBUG")
+      log_level = DEBUG;
+    else if (s == "INFO")
+      log_level = INFO;
+    else if (s == "WARNING")
+      log_level = WARNING;
+    else if (s == "ERROR")
+      log_level = ERROR;
+    else if (s == "FATAL")
+      log_level = FATAL;
+    else
+      fprintf(stderr,
+              "Unknown KNF_LOG_LEVEL: %s"
+              "\nSupported values are: "
+              "TRACE, DEBUG, INFO, WARNING, ERROR, FATAL",
+              s.c_str());
+  });
+  return log_level;
+}
+
+inline bool EnableAbort() {
+  static std::once_flag init_flag;
+  static bool enable_abort = false;
+  std::call_once(init_flag, []() {
+    enable_abort = (std::getenv("KNF_ABORT") != nullptr);
+  });
+  return enable_abort;
+}
+
+class Logger {
+ public:
+  Logger(const char *filename, const char *func_name, uint32_t line_num,
+         LogLevel level)
+      : filename_(filename),
+        func_name_(func_name),
+        line_num_(line_num),
+        level_(level) {
+    cur_level_ = GetCurrentLogLevel();
+    fprintf(stderr, "here\n");
+    switch (level) {
+      case TRACE:
+        if (cur_level_ <= TRACE) fprintf(stderr, "[T] ");
+        break;
+      case DEBUG:
+        if (cur_level_ <= DEBUG) fprintf(stderr, "[D] ");
+        break;
+      case INFO:
+        if (cur_level_ <= INFO) fprintf(stderr, "[I] ");
+        break;
+      case WARNING:
+        if (cur_level_ <= WARNING) fprintf(stderr, "[W] ");
+        break;
+      case ERROR:
+        if (cur_level_ <= ERROR) fprintf(stderr, "[E] ");
+        break;
+      case FATAL:
+        if (cur_level_ <= FATAL) fprintf(stderr, "[F] ");
+        break;
+    }
+
+    if (cur_level_ <= level_) {
+      fprintf(stderr, "%s:%u:%s ", filename, line_num, func_name);
+    }
+  }
+
+  ~Logger() noexcept(false) {
+    static constexpr const char *kErrMsg = R"(
+    Some bad things happened. Please read the above error messages and stack
+    trace. If you are using Python, the following command may be helpful:
+
+      gdb --args python /path/to/your/code.py
+
+    (You can use `gdb` to debug the code. Please consider compiling
+    a debug version of KNF.).
+
+    If you are unable to fix it, please open an issue at:
+
+      https://github.com/csukuangfj/kaldi-native-fbank/issues/new
+    )";
+    fprintf(stderr, "\n");
+    if (level_ == FATAL) {
+      std::string stack_trace = GetStackTrace();
+      if (!stack_trace.empty()) {
+        fprintf(stderr, "\n\n%s\n", stack_trace.c_str());
+      }
+
+      fflush(nullptr);
+
+#ifndef __ANDROID_API__
+      if (EnableAbort()) {
+        // NOTE: abort() will terminate the program immediately without
+        // printing the Python stack backtrace.
+        abort();
+      }
+
+      throw std::runtime_error(kErrMsg);
+#else
+      abort();
+#endif
+    }
+  }
+
+  const Logger &operator<<(bool b) const {
+    if (cur_level_ <= level_) {
+      fprintf(stderr, b ? "true" : "false");
+    }
+    return *this;
+  }
+
+  const Logger &operator<<(int8_t i) const {
+    if (cur_level_ <= level_) fprintf(stderr, "%d", i);
+    return *this;
+  }
+
+  const Logger &operator<<(const char *s) const {
+    if (cur_level_ <= level_) fprintf(stderr, "%s", s);
+    return *this;
+  }
+
+  const Logger &operator<<(int32_t i) const {
+    if (cur_level_ <= level_) fprintf(stderr, "%d", i);
+    return *this;
+  }
+
+  const Logger &operator<<(uint32_t i) const {
+    if (cur_level_ <= level_) fprintf(stderr, "%u", i);
+    return *this;
+  }
+
+  const Logger &operator<<(uint64_t i) const {
+    if (cur_level_ <= level_)
+      fprintf(stderr, "%llu", (long long unsigned int)i);  // NOLINT
+    return *this;
+  }
+
+  const Logger &operator<<(int64_t i) const {
+    if (cur_level_ <= level_)
+      fprintf(stderr, "%lli", (long long int)i);  // NOLINT
+    return *this;
+  }
+
+  const Logger &operator<<(float f) const {
+    if (cur_level_ <= level_) fprintf(stderr, "%f", f);
+    return *this;
+  }
+
+  const Logger &operator<<(double d) const {
+    if (cur_level_ <= level_) fprintf(stderr, "%f", d);
+    return *this;
+  }
+
+  template <typename T>
+  const Logger &operator<<(const T &t) const {
+    // require T overloads operator<<
+    std::ostringstream os;
+    os << t;
+    return *this << os.str().c_str();
+  }
+
+  // specialization to fix compile error: `stringstream << nullptr` is ambiguous
+  const Logger &operator<<(const std::nullptr_t &null) const {
+    if (cur_level_ <= level_) *this << "(null)";
+    return *this;
+  }
+
+ private:
+  const char *filename_;
+  const char *func_name_;
+  uint32_t line_num_;
+  LogLevel level_;
+  LogLevel cur_level_;
+};
+#endif  // KNF_ENABLE_CHECK
+
+class Voidifier {
+ public:
+#if KNF_ENABLE_CHECK
+  void operator&(const Logger &) const {}
+#endif
+};
+#if !defined(KNF_ENABLE_CHECK)
+template <typename T>
+const Voidifier &operator<<(const Voidifier &v, T &&) {
+  return v;
+}
+#endif
+
+}  // namespace knf
+
+#define KNF_STATIC_ASSERT(x) static_assert(x, "")
+
+#ifdef KNF_ENABLE_CHECK
+
+#if defined(__clang__) || defined(__GNUC__) || defined(__GNUG__) || \
+    defined(__PRETTY_FUNCTION__)
+// for clang and GCC
+#define KNF_FUNC __PRETTY_FUNCTION__
+#else
+// for other compilers
+#define KNF_FUNC __func__
+#endif
+
+#define KNF_CHECK(x)                                                  \
+  (x) ? (void)0                                                       \
+      : ::knf::Voidifier() &                                          \
+            ::knf::Logger(__FILE__, KNF_FUNC, __LINE__, ::knf::FATAL) \
+                << "Check failed: " << #x << " "
+
+// WARNING: x and y may be evaluated multiple times, but this happens only
+// when the check fails. Since the program aborts if it fails, we don't think
+// the extra evaluation of x and y matters.
+//
+// CAUTION: we recommend the following use case:
+//
+//      auto x = Foo();
+//      auto y = Bar();
+//      KNF_CHECK_EQ(x, y) << "Some message";
+//
+//  And please avoid
+//
+//      KNF_CHECK_EQ(Foo(), Bar());
+//
+//  if `Foo()` or `Bar()` causes some side effects, e.g., changing some
+//  local static variables or global variables.
+#define _KNF_CHECK_OP(x, y, op)                                              \
+  ((x)op(y)) ? (void)0                                                       \
+             : ::knf::Voidifier() &                                          \
+                   ::knf::Logger(__FILE__, KNF_FUNC, __LINE__, ::knf::FATAL) \
+                       << "Check failed: " << #x << " " << #op << " " << #y  \
+                       << " (" << (x) << " vs. " << (y) << ") "
+
+#define KNF_CHECK_EQ(x, y) _KNF_CHECK_OP(x, y, ==)
+#define KNF_CHECK_NE(x, y) _KNF_CHECK_OP(x, y, !=)
+#define KNF_CHECK_LT(x, y) _KNF_CHECK_OP(x, y, <)
+#define KNF_CHECK_LE(x, y) _KNF_CHECK_OP(x, y, <=)
+#define KNF_CHECK_GT(x, y) _KNF_CHECK_OP(x, y, >)
+#define KNF_CHECK_GE(x, y) _KNF_CHECK_OP(x, y, >=)
+
+#define KNF_LOG(x) ::knf::Logger(__FILE__, KNF_FUNC, __LINE__, ::knf::x)
+
+// ------------------------------------------------------------
+//       For debug check
+// ------------------------------------------------------------
+// If you define the macro "-D NDEBUG" while compiling kaldi-native-fbank,
+// the following macros are in fact empty and does nothing.
+
+#define KNF_DCHECK(x) ::knf::kDisableDebug ? (void)0 : KNF_CHECK(x)
+
+#define KNF_DCHECK_EQ(x, y) ::knf::kDisableDebug ? (void)0 : KNF_CHECK_EQ(x, y)
+
+#define KNF_DCHECK_NE(x, y) ::knf::kDisableDebug ? (void)0 : KNF_CHECK_NE(x, y)
+
+#define KNF_DCHECK_LT(x, y) ::knf::kDisableDebug ? (void)0 : KNF_CHECK_LT(x, y)
+
+#define KNF_DCHECK_LE(x, y) ::knf::kDisableDebug ? (void)0 : KNF_CHECK_LE(x, y)
+
+#define KNF_DCHECK_GT(x, y) ::knf::kDisableDebug ? (void)0 : KNF_CHECK_GT(x, y)
+
+#define KNF_DCHECK_GE(x, y) ::knf::kDisableDebug ? (void)0 : KNF_CHECK_GE(x, y)
+
+#define KNF_DLOG(x) \
+  ::knf::kDisableDebug ? (void)0 : ::knf::Voidifier() & KNF_LOG(x)
+
+#else
+
+#define KNF_CHECK(x) ::knf::Voidifier()
+#define KNF_LOG(x) ::knf::Voidifier()
+
+#define KNF_CHECK_EQ(x, y) ::knf::Voidifier()
+#define KNF_CHECK_NE(x, y) ::knf::Voidifier()
+#define KNF_CHECK_LT(x, y) ::knf::Voidifier()
+#define KNF_CHECK_LE(x, y) ::knf::Voidifier()
+#define KNF_CHECK_GT(x, y) ::knf::Voidifier()
+#define KNF_CHECK_GE(x, y) ::knf::Voidifier()
+
+#define KNF_DCHECK(x) ::knf::Voidifier()
+#define KNF_DLOG(x) ::knf::Voidifier()
+#define KNF_DCHECK_EQ(x, y) ::knf::Voidifier()
+#define KNF_DCHECK_NE(x, y) ::knf::Voidifier()
+#define KNF_DCHECK_LT(x, y) ::knf::Voidifier()
+#define KNF_DCHECK_LE(x, y) ::knf::Voidifier()
+#define KNF_DCHECK_GT(x, y) ::knf::Voidifier()
+#define KNF_DCHECK_GE(x, y) ::knf::Voidifier()
+
+#endif  // KNF_CHECK_NE
+
+#endif  // KALDI_NATIVE_FBANK_CSRC_LOG_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.cc
new file mode 100644
index 00000000..bdf55412
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.cc
@@ -0,0 +1,457 @@
+/**
+ * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// This file is copied/modified from kaldi/src/feat/mel-computations.cc
+
+#include "kaldi-native-fbank/csrc/mel-computations.h"
+
+#include <stdio.h>
+
+#include <algorithm>
+#include <sstream>
+#include <vector>
+
+#include "kaldi-native-fbank/csrc/feature-window.h"
+#include "kaldi-native-fbank/csrc/kaldi-math.h"
+#include "kaldi-native-fbank/csrc/log.h"
+
+namespace knf {
+
+std::ostream &operator<<(std::ostream &os, const MelBanksOptions &opts) {
+  os << opts.ToString();
+  return os;
+}
+
+float MelBanks::VtlnWarpFreq(
+    float vtln_low_cutoff,  // upper+lower frequency cutoffs for VTLN.
+    float vtln_high_cutoff,
+    float low_freq,  // upper+lower frequency cutoffs in mel computation
+    float high_freq, float vtln_warp_factor, float freq) {
+  /// This computes a VTLN warping function that is not the same as HTK's one,
+  /// but has similar inputs (this function has the advantage of never producing
+  /// empty bins).
+
+  /// This function computes a warp function F(freq), defined between low_freq
+  /// and high_freq inclusive, with the following properties:
+  ///  F(low_freq) == low_freq
+  ///  F(high_freq) == high_freq
+  /// The function is continuous and piecewise linear with two inflection
+  ///   points.
+  /// The lower inflection point (measured in terms of the unwarped
+  ///  frequency) is at frequency l, determined as described below.
+  /// The higher inflection point is at a frequency h, determined as
+  ///   described below.
+  /// If l <= f <= h, then F(f) = f/vtln_warp_factor.
+  /// If the higher inflection point (measured in terms of the unwarped
+  ///   frequency) is at h, then max(h, F(h)) == vtln_high_cutoff.
+  ///   Since (by the last point) F(h) == h/vtln_warp_factor, then
+  ///   max(h, h/vtln_warp_factor) == vtln_high_cutoff, so
+  ///   h = vtln_high_cutoff / max(1, 1/vtln_warp_factor).
+  ///     = vtln_high_cutoff * min(1, vtln_warp_factor).
+  /// If the lower inflection point (measured in terms of the unwarped
+  ///   frequency) is at l, then min(l, F(l)) == vtln_low_cutoff
+  ///   This implies that l = vtln_low_cutoff / min(1, 1/vtln_warp_factor)
+  ///                       = vtln_low_cutoff * max(1, vtln_warp_factor)
+
+  if (freq < low_freq || freq > high_freq)
+    return freq;  // in case this gets called
+  // for out-of-range frequencies, just return the freq.
+
+  KNF_CHECK_GT(vtln_low_cutoff, low_freq);
+  KNF_CHECK_LT(vtln_high_cutoff, high_freq);
+
+  float one = 1.0f;
+  float l = vtln_low_cutoff * std::max(one, vtln_warp_factor);
+  float h = vtln_high_cutoff * std::min(one, vtln_warp_factor);
+  float scale = 1.0f / vtln_warp_factor;
+  float Fl = scale * l;  // F(l);
+  float Fh = scale * h;  // F(h);
+  KNF_CHECK(l > low_freq && h < high_freq);
+  // slope of left part of the 3-piece linear function
+  float scale_left = (Fl - low_freq) / (l - low_freq);
+  // [slope of center part is just "scale"]
+
+  // slope of right part of the 3-piece linear function
+  float scale_right = (high_freq - Fh) / (high_freq - h);
+
+  if (freq < l) {
+    return low_freq + scale_left * (freq - low_freq);
+  } else if (freq < h) {
+    return scale * freq;
+  } else {  // freq >= h
+    return high_freq + scale_right * (freq - high_freq);
+  }
+}
+
+float MelBanks::VtlnWarpMelFreq(
+    float vtln_low_cutoff,  // upper+lower frequency cutoffs for VTLN.
+    float vtln_high_cutoff,
+    float low_freq,  // upper+lower frequency cutoffs in mel computation
+    float high_freq, float vtln_warp_factor, float mel_freq) {
+  return MelScale(VtlnWarpFreq(vtln_low_cutoff, vtln_high_cutoff, low_freq,
+                               high_freq, vtln_warp_factor,
+                               InverseMelScale(mel_freq)));
+}
+
+MelBanks::MelBanks(const MelBanksOptions &opts,
+                   const FrameExtractionOptions &frame_opts,
+                   float vtln_warp_factor) {
+  if (opts.is_librosa) {
+    InitLibrosaMelBanks(opts, frame_opts, vtln_warp_factor);
+  } else {
+    InitKaldiMelBanks(opts, frame_opts, vtln_warp_factor);
+  }
+}
+
+void MelBanks::InitKaldiMelBanks(const MelBanksOptions &opts,
+                                 const FrameExtractionOptions &frame_opts,
+                                 float vtln_warp_factor) {
+  htk_mode_ = opts.htk_mode;
+  int32_t num_bins = opts.num_bins;
+  if (num_bins < 3) {
+    KNF_LOG(FATAL) << "Must have at least 3 mel bins";
+  }
+
+  float sample_freq = frame_opts.samp_freq;
+  int32_t window_length_padded = frame_opts.PaddedWindowSize();
+  KNF_CHECK_EQ(window_length_padded % 2, 0);
+
+  int32_t num_fft_bins = window_length_padded / 2;
+  float nyquist = 0.5f * sample_freq;
+
+  float low_freq = opts.low_freq, high_freq;
+  if (opts.high_freq > 0.0f) {
+    high_freq = opts.high_freq;
+  } else {
+    high_freq = nyquist + opts.high_freq;
+  }
+
+  if (low_freq < 0.0f || low_freq >= nyquist || high_freq <= 0.0f ||
+      high_freq > nyquist || high_freq <= low_freq) {
+    KNF_LOG(FATAL) << "Bad values in options: low-freq " << low_freq
+                   << " and high-freq " << high_freq << " vs. nyquist "
+                   << nyquist;
+  }
+
+  float fft_bin_width = sample_freq / window_length_padded;
+  // fft-bin width [think of it as Nyquist-freq / half-window-length]
+
+  float mel_low_freq = MelScale(low_freq);
+  float mel_high_freq = MelScale(high_freq);
+
+  debug_ = opts.debug_mel;
+
+  // divide by num_bins+1 in next line because of end-effects where the bins
+  // spread out to the sides.
+  float mel_freq_delta = (mel_high_freq - mel_low_freq) / (num_bins + 1);
+
+  float vtln_low = opts.vtln_low, vtln_high = opts.vtln_high;
+  if (vtln_high < 0.0f) {
+    vtln_high += nyquist;
+  }
+
+  if (vtln_warp_factor != 1.0f &&
+      (vtln_low < 0.0f || vtln_low <= low_freq || vtln_low >= high_freq ||
+       vtln_high <= 0.0f || vtln_high >= high_freq || vtln_high <= vtln_low)) {
+    KNF_LOG(FATAL) << "Bad values in options: vtln-low " << vtln_low
+                   << " and vtln-high " << vtln_high << ", versus "
+                   << "low-freq " << low_freq << " and high-freq " << high_freq;
+  }
+
+  bins_.resize(num_bins);
+
+  for (int32_t bin = 0; bin < num_bins; ++bin) {
+    float left_mel = mel_low_freq + bin * mel_freq_delta,
+          center_mel = mel_low_freq + (bin + 1) * mel_freq_delta,
+          right_mel = mel_low_freq + (bin + 2) * mel_freq_delta;
+
+    if (vtln_warp_factor != 1.0f) {
+      left_mel = VtlnWarpMelFreq(vtln_low, vtln_high, low_freq, high_freq,
+                                 vtln_warp_factor, left_mel);
+      center_mel = VtlnWarpMelFreq(vtln_low, vtln_high, low_freq, high_freq,
+                                   vtln_warp_factor, center_mel);
+      right_mel = VtlnWarpMelFreq(vtln_low, vtln_high, low_freq, high_freq,
+                                  vtln_warp_factor, right_mel);
+    }
+
+    // this_bin will be a vector of coefficients that is only
+    // nonzero where this mel bin is active.
+    std::vector<float> this_bin(num_fft_bins);
+
+    int32_t first_index = -1, last_index = -1;
+    for (int32_t i = 0; i < num_fft_bins; ++i) {
+      float freq = (fft_bin_width * i);  // Center frequency of this fft
+                                         // bin.
+      float mel = MelScale(freq);
+      if (mel > left_mel && mel < right_mel) {
+        float weight;
+        if (mel <= center_mel) {
+          weight = (mel - left_mel) / (center_mel - left_mel);
+        } else {
+          weight = (right_mel - mel) / (right_mel - center_mel);
+        }
+        this_bin[i] = weight;
+        if (first_index == -1) {
+          first_index = i;
+        }
+        last_index = i;
+      }
+    }
+    KNF_CHECK(first_index != -1 && last_index >= first_index &&
+              "You may have set num_mel_bins too large.");
+
+    bins_[bin].first = first_index;
+    int32_t size = last_index + 1 - first_index;
+    bins_[bin].second.insert(bins_[bin].second.end(),
+                             this_bin.begin() + first_index,
+                             this_bin.begin() + first_index + size);
+
+    // Replicate a bug in HTK, for testing purposes.
+    if (opts.htk_mode && bin == 0 && mel_low_freq != 0.0f) {
+      bins_[bin].second[0] = 0.0;
+    }
+  }  // for (int32_t bin = 0; bin < num_bins; ++bin) {
+
+  if (debug_) {
+    std::ostringstream os;
+    for (size_t i = 0; i < bins_.size(); i++) {
+      os << "bin " << i << ", offset = " << bins_[i].first << ", vec = ";
+      for (auto k : bins_[i].second) os << k << ", ";
+      os << "\n";
+    }
+    KNF_LOG(INFO) << os.str();
+  }
+}
+
+void MelBanks::InitLibrosaMelBanks(const MelBanksOptions &opts,
+                                   const FrameExtractionOptions &frame_opts,
+                                   float vtln_warp_factor) {
+  htk_mode_ = opts.htk_mode;
+  int32_t num_bins = opts.num_bins;
+  if (num_bins < 3) {
+    KNF_LOG(FATAL) << "Must have at least 3 mel bins";
+  }
+
+  float sample_freq = frame_opts.samp_freq;
+  int32_t window_length_padded = frame_opts.PaddedWindowSize();
+  KNF_CHECK_EQ(window_length_padded % 2, 0);
+
+  int32_t num_fft_bins = window_length_padded / 2;
+  float nyquist = 0.5f * sample_freq;
+
+  float low_freq = opts.low_freq, high_freq;
+  if (opts.high_freq > 0.0f) {
+    high_freq = opts.high_freq;
+  } else {
+    high_freq = nyquist + opts.high_freq;
+  }
+
+  if (low_freq < 0.0f || low_freq >= nyquist || high_freq <= 0.0f ||
+      high_freq > nyquist || high_freq <= low_freq) {
+    KNF_LOG(FATAL) << "Bad values in options: low-freq " << low_freq
+                   << " and high-freq " << high_freq << " vs. nyquist "
+                   << nyquist;
+  }
+
+  float fft_bin_width = sample_freq / window_length_padded;
+
+  float mel_low_freq;
+  float mel_high_freq;
+
+  if (opts.use_slaney_mel_scale) {
+    mel_low_freq = MelScaleSlaney(low_freq);
+    mel_high_freq = MelScaleSlaney(high_freq);
+  } else {
+    mel_low_freq = MelScale(low_freq);
+    mel_high_freq = MelScale(high_freq);
+  }
+
+  debug_ = opts.debug_mel;
+
+  // divide by num_bins+1 in next line because of end-effects where the bins
+  // spread out to the sides.
+  float mel_freq_delta = (mel_high_freq - mel_low_freq) / (num_bins + 1);
+
+  bool slaney_norm = false;
+  if (!opts.norm.empty()) {
+    if (opts.norm != "slaney") {
+      KNF_LOG(FATAL) << "Unsupported norm: " << opts.norm;
+    }
+    slaney_norm = true;
+  }
+
+  bins_.resize(num_bins);
+  for (int32_t bin = 0; bin < num_bins; ++bin) {
+    float left_mel = mel_low_freq + bin * mel_freq_delta;
+    float center_mel = mel_low_freq + (bin + 1) * mel_freq_delta;
+    float right_mel = mel_low_freq + (bin + 2) * mel_freq_delta;
+
+    float left_hz;
+    float center_hz;
+    float right_hz;
+
+    if (opts.use_slaney_mel_scale) {
+      left_hz = InverseMelScaleSlaney(left_mel);
+      center_hz = InverseMelScaleSlaney(center_mel);
+      right_hz = InverseMelScaleSlaney(right_mel);
+    } else {
+      left_hz = InverseMelScale(left_mel);
+      center_hz = InverseMelScale(center_mel);
+      right_hz = InverseMelScale(right_mel);
+    }
+
+    if (opts.floor_to_int_bin) {
+      left_hz *= (window_length_padded + 1.0) / sample_freq;
+      center_hz *= (window_length_padded + 1.0) / sample_freq;
+      right_hz *= (window_length_padded + 1.0) / sample_freq;
+
+      left_hz = static_cast<int32_t>(left_hz);
+      center_hz = static_cast<int32_t>(center_hz);
+      right_hz = static_cast<int32_t>(right_hz);
+    }
+
+    // this_bin will be a vector of coefficients that is only
+    // nonzero where this mel bin is active.
+    //
+    // It is not an error to use num_fft_bins + 1 here. It is different
+    // from Kaldi.
+    std::vector<float> this_bin(num_fft_bins + 1);
+
+    int32_t first_index = -1, last_index = -1;
+    for (int32_t i = 0; i < num_fft_bins + 1; ++i) {
+      float hz;  // Center frequency of this fft bin.
+
+      if (opts.floor_to_int_bin) {
+        hz = i;
+      } else {
+        hz = fft_bin_width * i;
+      }
+
+      if (hz > left_hz && hz < right_hz) {
+        float weight;
+        if (hz <= center_hz) {
+          weight = (hz - left_hz) / (center_hz - left_hz);
+        } else {
+          weight = (right_hz - hz) / (right_hz - center_hz);
+        }
+
+        if (slaney_norm) {
+          weight *= 2 / (right_hz - left_hz);
+        }
+
+        this_bin[i] = weight;
+        if (first_index == -1) {
+          first_index = i;
+        }
+
+        last_index = i;
+      }
+    }  // for (int32_t i = 0; i < num_fft_bins + 1; ++i)
+
+    KNF_CHECK(first_index != -1 && last_index >= first_index &&
+              "You may have set num_mel_bins too large.");
+
+    bins_[bin].first = first_index;
+    int32_t size = last_index + 1 - first_index;
+    bins_[bin].second.insert(bins_[bin].second.end(),
+                             this_bin.begin() + first_index,
+                             this_bin.begin() + first_index + size);
+  }  // for (int32_t bin = 0; bin < num_bins; ++bin)
+
+  if (debug_) {
+    std::ostringstream os;
+    for (size_t i = 0; i < bins_.size(); i++) {
+      os << "bin " << i << ", offset = " << bins_[i].first << ", vec = ";
+      for (auto k : bins_[i].second) os << k << ", ";
+      os << "\n";
+    }
+    fprintf(stderr, "%s\n", os.str().c_str());
+  }
+}
+
+MelBanks::MelBanks(const float *weights, int32_t num_rows, int32_t num_cols)
+    : debug_(false), htk_mode_(false) {
+  bins_.resize(num_rows);
+  for (int32_t bin = 0; bin < num_rows; ++bin) {
+    const float *this_bin = weights + bin * num_cols;
+
+    int32_t first_index = -1, last_index = -1;
+
+    for (int32_t i = 0; i < num_cols; ++i) {
+      if (this_bin[i] == 0) {
+        continue;
+      }
+      if (first_index == -1) first_index = i;
+      last_index = i;
+    }
+
+    KNF_CHECK(first_index != -1 && last_index >= first_index &&
+              "You have an incorrect weight matrix.");
+
+    bins_[bin].first = first_index;
+    int32_t size = last_index + 1 - first_index;
+
+    bins_[bin].second.insert(bins_[bin].second.end(), this_bin + first_index,
+                             this_bin + first_index + size);
+  }
+}
+
+// "power_spectrum" contains fft energies.
+void MelBanks::Compute(const float *power_spectrum,
+                       float *mel_energies_out) const {
+  int32_t num_bins = bins_.size();
+
+  for (int32_t i = 0; i < num_bins; i++) {
+    int32_t offset = bins_[i].first;
+    const auto &v = bins_[i].second;
+    float energy = 0;
+    for (int32_t k = 0; k != v.size(); ++k) {
+      energy += v[k] * power_spectrum[k + offset];
+    }
+
+    // HTK-like flooring- for testing purposes (we prefer dither)
+    if (htk_mode_ && energy < 1.0) {
+      energy = 1.0;
+    }
+
+    mel_energies_out[i] = energy;
+
+    // The following assert was added due to a problem with OpenBlas that
+    // we had at one point (it was a bug in that library).  Just to detect
+    // it early.
+    KNF_CHECK_EQ(energy, energy);  // check that energy is not nan
+  }
+
+  if (debug_) {
+    fprintf(stderr, "MEL BANKS:\n");
+    for (int32_t i = 0; i < num_bins; i++)
+      fprintf(stderr, " %f", mel_energies_out[i]);
+    fprintf(stderr, "\n");
+  }
+}
+
+void ComputeLifterCoeffs(float Q, std::vector<float> *coeffs) {
+  // Compute liftering coefficients (scaling on cepstral coeffs)
+  // coeffs are numbered slightly differently from HTK: the zeroth
+  // index is C0, which is not affected.
+  for (int32_t i = 0; i != static_cast<int32_t>(coeffs->size()); ++i) {
+    (*coeffs)[i] = 1.0 + 0.5 * Q * sin(M_PI * i / Q);
+  }
+}
+
+}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.h
new file mode 100644
index 00000000..29904071
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.h
@@ -0,0 +1,190 @@
+/**
+ * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+// This file is copied/modified from kaldi/src/feat/mel-computations.h
+#ifndef KALDI_NATIVE_FBANK_CSRC_MEL_COMPUTATIONS_H_
+#define KALDI_NATIVE_FBANK_CSRC_MEL_COMPUTATIONS_H_
+
+#include <cmath>
+#include <cstdint>
+#include <sstream>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "kaldi-native-fbank/csrc/feature-window.h"
+
+namespace knf {
+struct FrameExtractionOptions;
+
+struct MelBanksOptions {
+  int32_t num_bins = 25;  // e.g. 25; number of triangular bins
+  float low_freq = 20;    // e.g. 20; lower frequency cutoff
+
+  // an upper frequency cutoff; 0 -> no cutoff, negative
+  // ->added to the Nyquist frequency to get the cutoff.
+  float high_freq = 0;
+
+  float vtln_low = 100;  // vtln lower cutoff of warping function.
+
+  // vtln upper cutoff of warping function: if negative, added
+  // to the Nyquist frequency to get the cutoff.
+  float vtln_high = -500;
+
+  bool debug_mel = false;
+  // htk_mode is a "hidden" config, it does not show up on command line.
+  // Enables more exact compatibility with HTK, for testing purposes.  Affects
+  // mel-energy flooring and reproduces a bug in HTK.
+  bool htk_mode = false;
+
+  // Note that if you set is_librosa, you probably need to set
+  // low_freq to 0.
+  // Please see
+  // https://librosa.org/doc/main/generated/librosa.filters.mel.html
+  bool is_librosa = false;
+
+  // used only when is_librosa=true
+  // Possible values: "", slaney. We don't support a numeric value here, but
+  // it can be added on demand.
+  // See https://librosa.org/doc/main/generated/librosa.filters.mel.html
+  std::string norm = "slaney";
+
+  // used only when is_librosa is true
+  bool use_slaney_mel_scale = true;
+
+  // used only when is_librosa is true
+  bool floor_to_int_bin = false;
+
+  std::string ToString() const {
+    std::ostringstream os;
+    os << "num_bins: " << num_bins << "\n";
+    os << "low_freq: " << low_freq << "\n";
+    os << "high_freq: " << high_freq << "\n";
+    os << "vtln_low: " << vtln_low << "\n";
+    os << "vtln_high: " << vtln_high << "\n";
+    os << "debug_mel: " << debug_mel << "\n";
+    os << "htk_mode: " << htk_mode << "\n";
+    os << "is_librosa: " << is_librosa << "\n";
+    os << "norm: " << norm << "\n";
+    os << "use_slaney_mel_scale: " << use_slaney_mel_scale << "\n";
+    os << "floor_to_int_bin: " << floor_to_int_bin << "\n";
+    return os.str();
+  }
+};
+
+std::ostream &operator<<(std::ostream &os, const MelBanksOptions &opts);
+
+class MelBanks {
+ public:
+  // see also https://en.wikipedia.org/wiki/Mel_scale
+  // htk, mel to hz
+  static inline float InverseMelScale(float mel_freq) {
+    return 700.0f * (expf(mel_freq / 1127.0f) - 1.0f);
+  }
+
+  // htk, hz to mel
+  static inline float MelScale(float freq) {
+    return 1127.0f * logf(1.0f + freq / 700.0f);
+  }
+
+  // slaney, mel to hz
+  static inline float InverseMelScaleSlaney(float mel_freq) {
+    if (mel_freq <= 15) {
+      return 200.0f / 3 * mel_freq;
+    }
+
+    // return 1000 * expf((mel_freq - 15) * logf(6.4f) / 27);
+
+    // Note: log(6.4)/27 = 0.06875177742094911
+
+    return 1000 * expf((mel_freq - 15) * 0.06875177742094911f);
+  }
+
+  // slaney, hz to mel
+  static inline float MelScaleSlaney(float freq) {
+    if (freq <= 1000) {
+      return freq * 3 / 200.0f;
+    }
+
+    // return 15 + 27 * logf(freq / 1000) / logf(6.4f)
+    //
+    // Note: 27/log(6.4) = 14.545078505785561
+
+    return 15 + 14.545078505785561f * logf(freq / 1000);
+  }
+
+  static float VtlnWarpFreq(
+      float vtln_low_cutoff,
+      float vtln_high_cutoff,  // discontinuities in warp func
+      float low_freq,
+      float high_freq,  // upper+lower frequency cutoffs in
+      // the mel computation
+      float vtln_warp_factor, float freq);
+
+  static float VtlnWarpMelFreq(float vtln_low_cutoff, float vtln_high_cutoff,
+                               float low_freq, float high_freq,
+                               float vtln_warp_factor, float mel_freq);
+
+  // TODO(fangjun): Remove vtln_warp_factor
+  MelBanks(const MelBanksOptions &opts,
+           const FrameExtractionOptions &frame_opts, float vtln_warp_factor);
+
+  // Initialize with a 2-d weights matrix
+  // @param weights Pointer to the start address of the matrix
+  // @param num_rows It equls to number of mel bins
+  // @param num_cols It equals to (number of fft bins)/2+1
+  MelBanks(const float *weights, int32_t num_rows, int32_t num_cols);
+
+  /// Compute Mel energies (note: not log energies).
+  /// At input, "fft_energies" contains the FFT energies (not log).
+  ///
+  /// @param fft_energies 1-D array of size num_fft_bins/2+1
+  /// @param mel_energies_out  1-D array of size num_mel_bins
+  void Compute(const float *fft_energies, float *mel_energies_out) const;
+
+  int32_t NumBins() const { return bins_.size(); }
+
+ private:
+  // for kaldi-compatible
+  void InitKaldiMelBanks(const MelBanksOptions &opts,
+                         const FrameExtractionOptions &frame_opts,
+                         float vtln_warp_factor);
+
+  // for librosa-compatible
+  // See https://librosa.org/doc/main/generated/librosa.filters.mel.html
+  void InitLibrosaMelBanks(const MelBanksOptions &opts,
+                           const FrameExtractionOptions &frame_opts,
+                           float vtln_warp_factor);
+
+ private:
+  // the "bins_" vector is a vector, one for each bin, of a pair:
+  // (the first nonzero fft-bin), (the vector of weights).
+  std::vector<std::pair<int32_t, std::vector<float>>> bins_;
+
+  // TODO(fangjun): Remove debug_ and htk_mode_
+  bool debug_ = false;
+  bool htk_mode_ = false;
+};
+
+// Compute liftering coefficients (scaling on cepstral coeffs)
+// coeffs are numbered slightly differently from HTK: the zeroth
+// index is C0, which is not affected.
+void ComputeLifterCoeffs(float Q, std::vector<float> *coeffs);
+
+}  // namespace knf
+
+#endif  // KALDI_NATIVE_FBANK_CSRC_MEL_COMPUTATIONS_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.cc
new file mode 100644
index 00000000..5f96ea43
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.cc
@@ -0,0 +1,170 @@
+/**
+ * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// The content in this file is copied/modified from
+// This file is copied/modified from kaldi/src/feat/online-feature.cc
+
+#include "kaldi-native-fbank/csrc/online-feature.h"
+
+#include <algorithm>
+#include <cstddef>
+#include <utility>
+#include <vector>
+
+#include "kaldi-native-fbank/csrc/feature-window.h"
+#include "kaldi-native-fbank/csrc/log.h"
+
+namespace knf {
+
+RecyclingVector::RecyclingVector(int32_t items_to_hold)
+    : items_to_hold_(items_to_hold == 0 ? -1 : items_to_hold),
+      first_available_index_(0) {}
+
+const float *RecyclingVector::At(int32_t index) const {
+  if (index < first_available_index_) {
+    KNF_LOG(FATAL) << "Attempted to retrieve feature vector that was "
+                      "already removed by the RecyclingVector (index = "
+                   << index << "; "
+                   << "first_available_index = " << first_available_index_
+                   << "; "
+                   << "size = " << Size() << ")";
+  }
+  // 'at' does size checking.
+  return items_.at(index - first_available_index_).data();
+}
+
+void RecyclingVector::PushBack(std::vector<float> item) {
+  // Note: -1 is a larger number when treated as unsigned
+  if (items_.size() == static_cast<size_t>(items_to_hold_)) {
+    items_.pop_front();
+    ++first_available_index_;
+  }
+  items_.push_back(std::move(item));
+}
+
+int32_t RecyclingVector::Size() const {
+  return first_available_index_ + static_cast<int32_t>(items_.size());
+}
+
+// discard the first n frames
+void RecyclingVector::Pop(int32_t n) {
+  for (int32_t i = 0; i < n && !items_.empty(); ++i) {
+    items_.pop_front();
+    ++first_available_index_;
+  }
+}
+
+template <class C>
+OnlineGenericBaseFeature<C>::OnlineGenericBaseFeature(
+    const typename C::Options &opts)
+    : computer_(opts),
+      window_function_(computer_.GetFrameOptions()),
+      input_finished_(false),
+      waveform_offset_(0) {}
+
+template <class C>
+void OnlineGenericBaseFeature<C>::AcceptWaveform(float sampling_rate,
+                                                 const float *waveform,
+                                                 int32_t n) {
+  if (n == 0) {
+    return;  // Nothing to do.
+  }
+
+  if (input_finished_) {
+    KNF_LOG(FATAL) << "AcceptWaveform called after InputFinished() was called.";
+  }
+
+  KNF_CHECK_EQ(sampling_rate, computer_.GetFrameOptions().samp_freq);
+
+  waveform_remainder_.insert(waveform_remainder_.end(), waveform, waveform + n);
+
+  ComputeFeatures();
+}
+
+template <class C>
+void OnlineGenericBaseFeature<C>::InputFinished() {
+  input_finished_ = true;
+  ComputeFeatures();
+}
+
+template <class C>
+void OnlineGenericBaseFeature<C>::ComputeFeatures() {
+  const FrameExtractionOptions &frame_opts = computer_.GetFrameOptions();
+
+  int64_t num_samples_total = waveform_offset_ + waveform_remainder_.size();
+
+  int32_t num_frames_old = features_.Size();
+
+  int32_t num_frames_new =
+      NumFrames(num_samples_total, frame_opts, input_finished_);
+
+  KNF_CHECK_GE(num_frames_new, num_frames_old);
+
+  // note: this online feature-extraction code does not support VTLN.
+  float vtln_warp = 1.0;
+
+  std::vector<float> window;
+  bool need_raw_log_energy = computer_.NeedRawLogEnergy();
+
+  for (int32_t frame = num_frames_old; frame < num_frames_new; ++frame) {
+    std::fill(window.begin(), window.end(), 0);
+    float raw_log_energy = 0.0;
+    ExtractWindow(waveform_offset_, waveform_remainder_, frame, frame_opts,
+                  window_function_, &window,
+                  need_raw_log_energy ? &raw_log_energy : nullptr);
+
+    std::vector<float> this_feature(computer_.Dim());
+
+    computer_.Compute(raw_log_energy, vtln_warp, &window, this_feature.data());
+    features_.PushBack(std::move(this_feature));
+  }
+
+  // OK, we will now discard any portion of the signal that will not be
+  // necessary to compute frames in the future.
+  int64_t first_sample_of_next_frame =
+      FirstSampleOfFrame(num_frames_new, frame_opts);
+
+  int32_t samples_to_discard = first_sample_of_next_frame - waveform_offset_;
+
+  if (samples_to_discard > 0) {
+    // discard the leftmost part of the waveform that we no longer need.
+    int32_t new_num_samples =
+        static_cast<int32_t>(waveform_remainder_.size()) - samples_to_discard;
+
+    if (new_num_samples <= 0) {
+      // odd, but we'll try to handle it.
+      waveform_offset_ += waveform_remainder_.size();
+      waveform_remainder_.resize(0);
+    } else {
+      std::vector<float> new_remainder(new_num_samples);
+
+      std::copy(waveform_remainder_.begin() + samples_to_discard,
+                waveform_remainder_.end(), new_remainder.begin());
+      waveform_offset_ += samples_to_discard;
+
+      waveform_remainder_.swap(new_remainder);
+    }
+  }
+}
+
+template class OnlineGenericBaseFeature<RawAudioSamplesComputer>;
+template class OnlineGenericBaseFeature<FbankComputer>;
+template class OnlineGenericBaseFeature<MfccComputer>;
+template class OnlineGenericBaseFeature<WhisperFeatureComputer>;
+
+}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.h
new file mode 100644
index 00000000..114567cf
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.h
@@ -0,0 +1,155 @@
+/**
+ * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// The content in this file is copied/modified from
+// This file is copied/modified from kaldi/src/feat/online-feature.h
+#ifndef KALDI_NATIVE_FBANK_CSRC_ONLINE_FEATURE_H_
+#define KALDI_NATIVE_FBANK_CSRC_ONLINE_FEATURE_H_
+
+#include <cstdint>
+#include <deque>
+#include <vector>
+
+#include "kaldi-native-fbank/csrc/feature-fbank.h"
+#include "kaldi-native-fbank/csrc/feature-mfcc.h"
+#include "kaldi-native-fbank/csrc/feature-raw-audio-samples.h"
+#include "kaldi-native-fbank/csrc/feature-window.h"
+#include "kaldi-native-fbank/csrc/whisper-feature.h"
+
+namespace knf {
+
+/// This class serves as a storage for feature vectors with an option to limit
+/// the memory usage by removing old elements. The deleted frames indices are
+/// "remembered" so that regardless of the MAX_ITEMS setting, the user always
+/// provides the indices as if no deletion was being performed.
+/// This is useful when processing very long recordings which would otherwise
+/// cause the memory to eventually blow up when the features are not being
+/// removed.
+class RecyclingVector {
+ public:
+  /// By default it does not remove any elements.
+  explicit RecyclingVector(int32_t items_to_hold = -1);
+
+  ~RecyclingVector() = default;
+  RecyclingVector(const RecyclingVector &) = delete;
+  RecyclingVector &operator=(const RecyclingVector &) = delete;
+
+  // The pointer is owned by RecyclingVector
+  // Users should not free it
+  const float *At(int32_t index) const;
+
+  void PushBack(std::vector<float> item);
+
+  /// This method returns the size as if no "recycling" had happened,
+  /// i.e. equivalent to the number of times the PushBack method has been
+  /// called.
+  int32_t Size() const;
+
+  // discard the first n frames
+  void Pop(int32_t n);
+
+ private:
+  std::deque<std::vector<float>> items_;
+  int32_t items_to_hold_;
+  int32_t first_available_index_;
+};
+
+/// This is a templated class for online feature extraction;
+/// it's templated on a class like MfccComputer or PlpComputer
+/// that does the basic feature extraction.
+template <class C>
+class OnlineGenericBaseFeature {
+ public:
+  // Constructor from options class
+  explicit OnlineGenericBaseFeature(const typename C::Options &opts);
+
+  int32_t Dim() const { return computer_.Dim(); }
+
+  float FrameShiftInSeconds() const {
+    return computer_.GetFrameOptions().frame_shift_ms / 1000.0f;
+  }
+
+  int32_t NumFramesReady() const { return features_.Size(); }
+
+  // Note: IsLastFrame() will only ever return true if you have called
+  // InputFinished() (and this frame is the last frame).
+  bool IsLastFrame(int32_t frame) const {
+    return input_finished_ && frame == NumFramesReady() - 1;
+  }
+
+  const float *GetFrame(int32_t frame) const { return features_.At(frame); }
+
+  // This would be called from the application, when you get
+  // more wave data.  Note: the sampling_rate is only provided so
+  // the code can assert that it matches the sampling rate
+  // expected in the options.
+  //
+  // @param sampling_rate The sampling_rate of the input waveform
+  // @param waveform Pointer to a 1-D array of size n
+  // @param n Number of entries in waveform
+  void AcceptWaveform(float sampling_rate, const float *waveform, int32_t n);
+
+  // InputFinished() tells the class you won't be providing any
+  // more waveform.  This will help flush out the last frame or two
+  // of features, in the case where snip-edges == false; it also
+  // affects the return value of IsLastFrame().
+  void InputFinished();
+
+  // discard the first n frames
+  void Pop(int32_t n) { features_.Pop(n); }
+
+ private:
+  // This function computes any additional feature frames that it is possible to
+  // compute from 'waveform_remainder_', which at this point may contain more
+  // than just a remainder-sized quantity (because AcceptWaveform() appends to
+  // waveform_remainder_ before calling this function).  It adds these feature
+  // frames to features_, and shifts off any now-unneeded samples of input from
+  // waveform_remainder_ while incrementing waveform_offset_ by the same amount.
+  void ComputeFeatures();
+
+  C computer_;  // class that does the MFCC or PLP or filterbank computation
+
+  FeatureWindowFunction window_function_;
+
+  // features_ is the Mfcc or Plp or Fbank features that we have already
+  // computed.
+
+  RecyclingVector features_;
+
+  // True if the user has called "InputFinished()"
+  bool input_finished_;
+
+  // waveform_offset_ is the number of samples of waveform that we have
+  // already discarded, i.e. that were prior to 'waveform_remainder_'.
+  int64_t waveform_offset_;
+
+  // waveform_remainder_ is a short piece of waveform that we may need to keep
+  // after extracting all the whole frames we can (whatever length of feature
+  // will be required for the next phase of computation).
+  // It is a 1-D tensor
+  std::vector<float> waveform_remainder_;
+};
+
+using OnlineRawAudioSamples = OnlineGenericBaseFeature<RawAudioSamplesComputer>;
+using OnlineFbank = OnlineGenericBaseFeature<FbankComputer>;
+using OnlineMfcc = OnlineGenericBaseFeature<MfccComputer>;
+using OnlineWhisperFbank = OnlineGenericBaseFeature<WhisperFeatureComputer>;
+
+}  // namespace knf
+
+#endif  // KALDI_NATIVE_FBANK_CSRC_ONLINE_FEATURE_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.cc
new file mode 100644
index 00000000..64ed83c9
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.cc
@@ -0,0 +1,112 @@
+/**
+ * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kaldi-native-fbank/csrc/rfft.h"
+
+#include <algorithm>
+#include <cmath>
+#include <vector>
+
+#include "kaldi-native-fbank/csrc/log.h"
+#include "kiss_fftr.h"
+
+namespace knf {
+
+class Rfft::RfftImpl {
+ public:
+  RfftImpl(int32_t n, bool inverse) : n_(n), inverse_(inverse) {
+    if ((n & 1) != 0) {
+      fprintf(stderr, "n should be even. Given: %d \n", n);
+      exit(-1);
+    }
+
+    if (n < 0) {
+      fprintf(stderr, "n should be positive. Given: %d \n", n);
+      exit(-1);
+    }
+  }
+
+  void Compute(float *in_out) {
+    if (!inverse_) {
+      Forward(in_out);
+    } else {
+      Reverse(in_out);
+    }
+  }
+
+  void Compute(double *in_out) {
+    std::vector<float> f(in_out, in_out + n_);
+
+    Compute(f.data());
+
+    std::copy(f.begin(), f.end(), in_out);
+  }
+
+ private:
+  void Forward(float *in_out) const {
+    kiss_fftr_cfg cfg = kiss_fftr_alloc(n_, 0, nullptr, nullptr);
+
+    std::vector<kiss_fft_cpx> out(n_ / 2 + 1);
+
+    kiss_fftr(cfg, in_out, out.data());
+
+    kiss_fftr_free(cfg);
+
+    in_out[0] = out[0].r;
+    in_out[1] = out[n_ / 2].r;
+
+    for (int32_t i = 1; i < n_ / 2; ++i) {
+      in_out[2 * i] = out[i].r;
+      in_out[2 * i + 1] = out[i].i;
+    }
+  }
+
+  void Reverse(float *in_out) const {
+    std::vector<kiss_fft_cpx> out(n_ / 2 + 1);
+    out[0].r = in_out[0];
+    out[0].i = 0;
+
+    out[n_ / 2].r = in_out[1];
+    out[n_ / 2].i = 0;
+
+    for (int32_t i = 1; i < n_ / 2; ++i) {
+      out[i].r = in_out[2 * i];
+      out[i].i = in_out[2 * i + 1];
+    }
+
+    kiss_fftr_cfg cfg = kiss_fftr_alloc(n_, 1, nullptr, nullptr);
+
+    kiss_fftri(cfg, out.data(), in_out);
+
+    kiss_fftr_free(cfg);
+  }
+
+ private:
+  int32_t n_;
+  bool inverse_ = false;
+};
+
+Rfft::Rfft(int32_t n, bool inverse /*=false*/)
+    : impl_(std::make_unique<RfftImpl>(n, inverse)) {}
+
+Rfft::~Rfft() = default;
+
+void Rfft::Compute(float *in_out) { impl_->Compute(in_out); }
+void Rfft::Compute(double *in_out) { impl_->Compute(in_out); }
+
+}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.h
new file mode 100644
index 00000000..c6fab6d9
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.h
@@ -0,0 +1,58 @@
+/**
+ * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef KALDI_NATIVE_FBANK_CSRC_RFFT_H_
+#define KALDI_NATIVE_FBANK_CSRC_RFFT_H_
+
+#include <cstdint>
+#include <memory>
+
+namespace knf {
+
+// n-point Real discrete Fourier transform
+// where n is even. n >= 2
+//
+//  R[k] = sum_j=0^n-1 in[j]*cos(2*pi*j*k/n), 0<=k<=n/2
+//  I[k] = sum_j=0^n-1 in[j]*sin(2*pi*j*k/n), 0<k<n/2
+class Rfft {
+ public:
+  // @param n Number of fft bins. it should be even.
+  explicit Rfft(int32_t n, bool inverse = false);
+  ~Rfft();
+
+  /** @param in_out A 1-D array of size n.
+   *             On return:
+   *               in_out[0] = R[0]
+   *               in_out[1] = R[n/2]
+   *               for 1 < k < n/2,
+   *                 in_out[2*k] = R[k]
+   *                 in_out[2*k+1] = I[k]
+   *
+   */
+  void Compute(float *in_out);
+  void Compute(double *in_out);
+
+ private:
+  class RfftImpl;
+
+  std::unique_ptr<RfftImpl> impl_;
+};
+
+}  // namespace knf
+
+#endif  // KALDI_NATIVE_FBANK_CSRC_RFFT_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.cc
new file mode 100644
index 00000000..2cdb637f
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.cc
@@ -0,0 +1,142 @@
+/**
+ * Copyright (c)  2025  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "kaldi-native-fbank/csrc/stft.h"
+
+#include <algorithm>
+#include <cmath>
+#include <sstream>
+#include <string>
+
+#include "kaldi-native-fbank/csrc/feature-window.h"
+#include "kaldi-native-fbank/csrc/rfft.h"
+
+namespace knf {
+std::string StftConfig::ToString() const {
+  std::ostringstream os;
+  os << "StftConfig(";
+  os << "n_fft=" << n_fft << ", ";
+  os << "hop_length=" << hop_length << ", ";
+  os << "win_length=" << win_length << ", ";
+  os << "window_type=\"" << window_type << "\", ";
+  os << "center=" << (center ? "True" : "False") << ", ";
+  os << "pad_mode=\"" << pad_mode << "\", ";
+  os << "normalized=" << (normalized ? "True" : "False") << ")";
+  return os.str();
+}
+
+class Stft::Impl {
+ public:
+  explicit Impl(const StftConfig &config) : config_(config) {
+    if (!config.window.empty()) {
+      window_ = std::make_unique<FeatureWindowFunction>(config.window);
+    } else if (!config.window_type.empty()) {
+      window_ = std::make_unique<FeatureWindowFunction>(config.window_type,
+                                                        config.win_length);
+    }
+  }
+
+  StftResult Compute(const float *data, int32_t n) const {
+    int32_t n_fft = config_.n_fft;
+    int32_t hop_length = config_.hop_length;
+
+    std::vector<float> samples;
+    const float *p = data;
+
+    if (config_.center) {
+      samples = Pad(data, n);
+      p = samples.data();
+      n = samples.size();
+    }
+
+    int64_t num_frames = 1 + (n - n_fft) / hop_length;
+
+    Rfft rfft(n_fft);
+
+    StftResult ans;
+    ans.num_frames = num_frames;
+    ans.real.resize(num_frames * (n_fft / 2 + 1));
+    ans.imag.resize(num_frames * (n_fft / 2 + 1));
+
+    std::vector<float> tmp(config_.n_fft);
+
+    for (int32_t i = 0; i < num_frames; ++i) {
+      tmp = {p + i * hop_length, p + i * hop_length + n_fft};
+      if (window_) {
+        window_->Apply(tmp.data());
+      }
+
+      rfft.Compute(tmp.data());
+
+      for (int32_t k = 0; k < n_fft / 2; ++k) {
+        if (k == 0) {
+          ans.real[i * (n_fft / 2 + 1)] = tmp[0];
+          ans.real[i * (n_fft / 2 + 1) + n_fft / 2] = tmp[1];
+        } else {
+          ans.real[i * (n_fft / 2 + 1) + k] = tmp[2 * k];
+
+          ans.imag[i * (n_fft / 2 + 1) + k] = tmp[2 * k + 1];
+        }
+      }
+    }
+
+    if (config_.normalized) {
+      float scale = 1 / std::sqrt(n_fft);
+      for (int32_t i = 0; i < ans.real.size(); ++i) {
+        ans.real[i] *= scale;
+        ans.imag[i] *= scale;
+      }
+    }
+
+    return ans;
+  }
+
+  std::vector<float> Pad(const float *data, int32_t n) const {
+    int32_t pad_amount = config_.n_fft / 2;
+    std::vector<float> ans(n + config_.n_fft);
+    std::copy(data, data + n, ans.begin() + pad_amount);
+    if (config_.pad_mode == "constant") {
+      // do nothing
+    } else if (config_.pad_mode == "reflect") {
+      // left
+      std::copy(data + 1, data + 1 + pad_amount, ans.rend() - pad_amount);
+      std::copy(data + n - pad_amount - 1, data + n - 1, ans.rbegin());
+    } else if (config_.pad_mode == "replicate") {
+      std::fill(ans.begin(), ans.begin() + pad_amount, data[0]);
+      std::fill(ans.end() - pad_amount, ans.end(), data[n - 1]);
+    } else {
+      fprintf(stderr, "Unsupported pad_mode: '%s'. Use 0 padding\n",
+              config_.pad_mode.c_str());
+    }
+
+    return ans;
+  }
+
+ private:
+  StftConfig config_;
+  std::unique_ptr<FeatureWindowFunction> window_;
+};
+
+Stft::Stft(const StftConfig &config) : impl_(std::make_unique<Impl>(config)) {}
+
+Stft::~Stft() = default;
+
+StftResult Stft::Compute(const float *data, int32_t n) const {
+  return impl_->Compute(data, n);
+}
+
+}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.h
new file mode 100644
index 00000000..6c5d2f7b
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.h
@@ -0,0 +1,65 @@
+/**
+ * Copyright (c)  2025  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef KALDI_NATIVE_FBANK_CSRC_STFT_H_
+#define KALDI_NATIVE_FBANK_CSRC_STFT_H_
+
+#include <cstdint>
+#include <memory>
+#include <string>
+#include <vector>
+
+namespace knf {
+
+struct StftConfig {
+  int32_t n_fft;  // should be a power of two
+  int32_t hop_length;
+  int32_t win_length;
+  std::string window_type;
+  bool center = true;
+  std::string pad_mode = "reflect";
+  bool normalized = false;
+
+  // if it is specified, then window_type is ignored
+  std::vector<float> window;
+
+  std::string ToString() const;
+};
+
+struct StftResult {
+  // [num_frames, n_fft/2+1], flattened in row major
+  std::vector<float> real;
+  std::vector<float> imag;
+  int32_t num_frames;
+};
+
+class Stft {
+ public:
+  explicit Stft(const StftConfig &config);
+  ~Stft();
+  StftResult Compute(const float *data, int32_t n) const;
+
+ private:
+  class Impl;
+  std::unique_ptr<Impl> impl_;
+};
+
+}  // namespace knf
+
+#endif  // KALDI_NATIVE_FBANK_CSRC_STFT_H_
+        //
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-log.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-log.cc
new file mode 100644
index 00000000..1e6ef734
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-log.cc
@@ -0,0 +1,71 @@
+/**
+ * Copyright      2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#if KNF_ENABLE_CHECK
+#include "gtest/gtest.h"
+#include "kaldi-native-fbank/csrc/log.h"
+
+namespace knf {
+
+TEST(Log, TestLog) {
+  KNF_LOG(TRACE) << "this is a trace message";
+  KNF_LOG(DEBUG) << "this is a debug message";
+  KNF_LOG(INFO) << "this is an info message";
+  KNF_LOG(WARNING) << "this is a warning message";
+  KNF_LOG(ERROR) << "this is an error message";
+
+  ASSERT_THROW(KNF_LOG(FATAL) << "This will crash the program",
+               std::runtime_error);
+
+  // For debug build
+
+  KNF_DLOG(TRACE) << "this is a trace message for debug build";
+  KNF_DLOG(DEBUG) << "this is a trace message for debug build";
+  KNF_DLOG(INFO) << "this is a trace message for debug build";
+  KNF_DLOG(ERROR) << "this is an error message for debug build";
+  KNF_DLOG(WARNING) << "this is a trace message for debug build";
+
+#if !defined(NDEBUG)
+  ASSERT_THROW(KNF_DLOG(FATAL) << "this is a trace message for debug build",
+               std::runtime_error);
+#endif
+}
+
+TEST(Log, TestCheck) {
+  KNF_CHECK_EQ(1, 1) << "ok";
+  KNF_CHECK_LE(1, 3) << "ok";
+  KNF_CHECK_LT(1, 2) << "ok";
+  KNF_CHECK_GT(2, 1) << "ok";
+  KNF_CHECK_GE(2, 1) << "ok";
+
+  ASSERT_THROW(KNF_CHECK_EQ(2, 1) << "bad things happened", std::runtime_error);
+
+  // for debug build
+  KNF_DCHECK_EQ(1, 1) << "ok";
+  KNF_DCHECK_LE(1, 3) << "ok";
+  KNF_DCHECK_LT(1, 2) << "ok";
+  KNF_DCHECK_GT(2, 1) << "ok";
+  KNF_DCHECK_GE(2, 1) << "ok";
+
+#if !defined(NDEBUG)
+  ASSERT_THROW(KNF_CHECK_EQ(2, 1) << "bad things happened", std::runtime_error);
+#endif
+}
+
+}  // namespace knf
+#endif
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-fbank.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-fbank.cc
new file mode 100644
index 00000000..c2d437a4
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-fbank.cc
@@ -0,0 +1,54 @@
+/**
+ * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <iostream>
+#include <cstdint>
+#include <sstream>
+#include <string>
+
+#include "kaldi-native-fbank/csrc/online-feature.h"
+#include "kaldi-native-fbank/csrc/feature-fbank.h"
+#include "kaldi-native-fbank/csrc/feature-window.h"
+#include "kaldi-native-fbank/csrc/mel-computations.h"
+
+int main() {
+  knf::FbankOptions opts;
+  opts.frame_opts.dither = 0;
+  opts.mel_opts.num_bins = 10;
+
+  knf::OnlineFbank fbank(opts);
+  for (int32_t i = 0; i < 1600; ++i) {
+    float s = (i * i - i / 2) / 32767.;
+    fbank.AcceptWaveform(16000, &s, 1);
+  }
+
+  std::ostringstream os;
+
+  int32_t n = fbank.NumFramesReady();
+  for (int32_t i = 0; i != n; ++i) {
+    const float *frame = fbank.GetFrame(i);
+    for (int32_t k = 0; k != opts.mel_opts.num_bins; ++k) {
+      os << frame[k] << ", ";
+    }
+    os << "\n";
+  }
+
+  std::cout << os.str() << "\n";
+
+  return 0;
+}
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-feature.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-feature.cc
new file mode 100644
index 00000000..bfbe6210
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-feature.cc
@@ -0,0 +1,59 @@
+/**
+ * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "gtest/gtest.h"
+#include "kaldi-native-fbank/csrc/online-feature.h"
+namespace knf {
+
+TEST(RecyclingVector, TestUnlimited) {
+  RecyclingVector v(-1);
+  constexpr int32_t N = 100;
+  for (int32_t i = 0; i != N; ++i) {
+    std::unique_ptr<float[]> p(new float[3]{i, i + 1, i + 2});
+    v.PushBack(std::move(p));
+  }
+  ASSERT_EQ(v.Size(), N);
+
+  for (int32_t i = 0; i != N; ++i) {
+    const float *t = v.At(i);
+    for (int32_t k = 0; k != 3; ++k) {
+      EXPECT_EQ(t[k], (i + k));
+    }
+  }
+}
+
+TEST(RecyclingVector, Testlimited) {
+  constexpr int32_t K = 3;
+  constexpr int32_t N = 10;
+  RecyclingVector v(K);
+  for (int32_t i = 0; i != N; ++i) {
+    std::unique_ptr<float[]> p(new float[3]{i, i + 1, i + 2});
+    v.PushBack(std::move(p));
+  }
+
+  ASSERT_EQ(v.Size(), N);
+
+  for (int32_t i = N - K; i != N; ++i) {
+    const float *t = v.At(i);
+
+    for (int32_t k = 0; k != 3; ++k) {
+      EXPECT_EQ(t[k], (i + k));
+    }
+  }
+}
+}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-rfft.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-rfft.cc
new file mode 100644
index 00000000..e1763e4e
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-rfft.cc
@@ -0,0 +1,101 @@
+/**
+ * Copyright      2022  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <cstdint>
+#include <vector>
+
+#include "gtest/gtest.h"
+#include "kaldi-native-fbank/csrc/rfft.h"
+
+namespace knf {
+
+#if 0
+>>> import torch
+>>> a = torch.tensor([1., -1, 3, 8, 20, 6, 0, 2])
+>>> torch.fft.rfft(a)
+tensor([ 39.0000+0.0000j, -28.1924-2.2929j,  18.0000+5.0000j,  -9.8076+3.7071j,
+          9.0000+0.0000j])
+#endif
+
+TEST(PowerOfTwo, TestRfft) {
+  knf::Rfft fft(8);
+  std::vector<float> original = {1, -1, 3, 8, 20, 6, 0, 2};
+
+  std::vector<float> d = original;
+  fft.Compute(d.data());
+
+  EXPECT_EQ(d[0], 39);
+  EXPECT_EQ(d[1], 9);
+
+  EXPECT_NEAR(d[2], -28.1924, 1e-3);
+  EXPECT_NEAR(d[3], -2.2929, 1e-3);
+
+  EXPECT_NEAR(d[4], 18, 1e-3);
+  EXPECT_NEAR(d[5], 5, 1e-3);
+
+  EXPECT_NEAR(d[6], -9.8076, 1e-3);
+  EXPECT_NEAR(d[7], 3.7071, 1e-3);
+
+  knf::Rfft ifft(8, true);
+  ifft.Compute(d.data());
+
+  for (int32_t i = 0; i < d.size(); ++i) {
+    d[i] /= d.size();  // we need to rescale it by 1/n
+    EXPECT_EQ(d[i], original[i]);
+  }
+}
+
+#if 0
+>>> import torch
+>>> a = torch.tensor([1., -1, 3, 8, 20, 6, 0, 2, 9, 5])
+>>> torch.fft.rfft(a)
+tensor([ 53.0000+0.0000j, -17.3262-8.2290j,  -3.3820+31.7809j,
+         -1.6738-13.3148j,  -5.6180+3.8697j,  13.0000+0.0000j])
+#endif
+TEST(NotPowerOfTwo, TestRfft) {
+  knf::Rfft fft(10);
+  std::vector<float> original = {1, -1, 3, 8, 20, 6, 0, 2, 9, 5};
+
+  std::vector<float> d = original;
+  fft.Compute(d.data());
+
+  EXPECT_EQ(d[0], 53);
+  EXPECT_EQ(d[1], 13);
+
+  EXPECT_NEAR(d[2], -17.3262, 1e-3);
+  EXPECT_NEAR(d[3], -8.2290, 1e-3);
+
+  EXPECT_NEAR(d[4], -3.3820, 1e-3);
+  EXPECT_NEAR(d[5], 31.7809, 1e-3);
+
+  EXPECT_NEAR(d[6], -1.6738, 1e-3);
+  EXPECT_NEAR(d[7], -13.3148, 1e-3);
+
+  EXPECT_NEAR(d[8], -5.6180, 1e-3);
+  EXPECT_NEAR(d[9], 3.8697, 1e-3);
+
+  knf::Rfft ifft(10, true);
+  ifft.Compute(d.data());
+
+  for (int32_t i = 0; i < d.size(); ++i) {
+    d[i] /= d.size();  // we need to rescale it by 1/n
+    EXPECT_NEAR(d[i], original[i], 1e-5);
+  }
+}
+
+}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.cc
new file mode 100644
index 00000000..1b699846
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.cc
@@ -0,0 +1,169 @@
+/**
+ * Copyright (c)  2023  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kaldi-native-fbank/csrc/whisper-feature.h"
+
+#include <cmath>
+#include <string>
+#include <vector>
+
+#include "kaldi-native-fbank/csrc/log.h"
+#include "kaldi-native-fbank/csrc/mel-computations.h"
+
+#ifndef M_2PI
+#define M_2PI 6.283185307179586476925286766559005
+#endif
+
+namespace knf {
+
+std::string WhisperFeatureOptions::ToString() const {
+  std::ostringstream os;
+  os << "WhisperFeatureOptions(";
+  os << "frame_opts=" << frame_opts.ToString() << ", ";
+  os << "dim=" << dim << ")";
+  return os.str();
+}
+
+static void dft(const std::vector<float> &in, std::vector<float> *out) {
+  // this function is modified from
+  // https://github.com/ggerganov/whisper.cpp/blob/master/whisper.cpp#L2353
+  int32_t N = in.size();
+
+  out->resize(N * 2);
+
+  auto M_2PI_over_N = M_2PI / N;
+  for (int32_t k = 0; k < N; ++k) {
+    float re = 0;
+    float im = 0;
+
+    for (int32_t n = 0; n < N; ++n) {
+      float angle = M_2PI_over_N * k * n;
+      re += in[n] * cos(angle);
+      im -= in[n] * sin(angle);
+    }
+
+    (*out)[k * 2 + 0] = re;
+    (*out)[k * 2 + 1] = im;
+  }
+}
+
+// Cooley-Tukey FFT
+// poor man's implementation - use something better
+// input is real-valued
+// output is complex-valued
+static void fft(const std::vector<float> &in, std::vector<float> *out) {
+  // this function is copied from
+  // https://github.com/ggerganov/whisper.cpp/blob/master/whisper.cpp#L2373C1-L2429C1
+
+  int32_t N = in.size();
+  out->resize(N * 2);
+
+  if (N == 1) {
+    (*out)[0] = in[0];
+    (*out)[1] = 0;
+    return;
+  }
+
+  if (N % 2 == 1) {
+    dft(in, out);
+    return;
+  }
+
+  std::vector<float> even;
+  std::vector<float> odd;
+
+  even.reserve(N / 2);
+  odd.reserve(N / 2);
+
+  for (int32_t i = 0; i != N; ++i) {
+    if (i % 2 == 0) {
+      even.push_back(in[i]);
+    } else {
+      odd.push_back(in[i]);
+    }
+  }
+
+  std::vector<float> even_fft;
+  std::vector<float> odd_fft;
+
+  fft(even, &even_fft);
+  fft(odd, &odd_fft);
+
+  for (int32_t k = 0; k < N / 2; ++k) {
+    float theta = M_2PI * k / N;
+
+    float re = cos(theta);
+    float im = -sin(theta);
+
+    float re_odd = odd_fft[2 * k + 0];
+    float im_odd = odd_fft[2 * k + 1];
+
+    (*out)[2 * k + 0] = even_fft[2 * k + 0] + re * re_odd - im * im_odd;
+    (*out)[2 * k + 1] = even_fft[2 * k + 1] + re * im_odd + im * re_odd;
+
+    (*out)[2 * (k + N / 2) + 0] =
+        even_fft[2 * k + 0] - re * re_odd + im * im_odd;
+    (*out)[2 * (k + N / 2) + 1] =
+        even_fft[2 * k + 1] - re * im_odd - im * re_odd;
+  }
+}
+
+WhisperFeatureComputer::WhisperFeatureComputer(
+    const WhisperFeatureOptions &opts /*= {}*/)
+    : opts_(opts) {
+  opts_.frame_opts.samp_freq = 16000;
+  opts_.frame_opts.frame_shift_ms = 10;
+  opts_.frame_opts.frame_length_ms = 25;
+  opts_.frame_opts.dither = 0;
+  opts_.frame_opts.preemph_coeff = 0;
+  opts_.frame_opts.remove_dc_offset = false;
+  opts_.frame_opts.window_type = "hann";
+  opts_.frame_opts.round_to_power_of_two = false;
+  opts_.frame_opts.snip_edges = false;
+
+  MelBanksOptions mel_opts;
+  mel_opts.num_bins = opts_.dim;
+  mel_opts.low_freq = 0;
+  mel_opts.is_librosa = true;
+
+  mel_banks_ = std::make_unique<MelBanks>(mel_opts, opts_.frame_opts, 1.0f);
+}
+
+void WhisperFeatureComputer::Compute(float /*signal_raw_log_energy*/,
+                                     float /*vtln_warp*/,
+                                     std::vector<float> *signal_frame,
+                                     float *feature) {
+  KNF_CHECK_EQ(signal_frame->size(), frame_opts_.PaddedWindowSize());
+  // we have already applied window function to signal_frame before
+  // calling this method
+  std::vector<float> fft_out;
+  fft(*signal_frame, &fft_out);
+
+  int32_t num_fft = signal_frame->size();
+  std::vector<float> power(num_fft / 2 + 1);
+  for (int32_t i = 0; i <= num_fft / 2; ++i) {
+    float re = fft_out[2 * i + 0];
+    float im = fft_out[2 * i + 1];
+    power[i] = re * re + im * im;
+  }
+
+  // feature is pre-allocated by the user
+  mel_banks_->Compute(power.data(), feature);
+}
+
+}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.h
new file mode 100644
index 00000000..57b9cc38
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.h
@@ -0,0 +1,69 @@
+/**
+ * Copyright (c)  2023  Xiaomi Corporation (authors: Fangjun Kuang)
+ *
+ * See LICENSE for clarification regarding multiple authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef KALDI_NATIVE_FBANK_CSRC_WHISPER_FEATURE_H_
+#define KALDI_NATIVE_FBANK_CSRC_WHISPER_FEATURE_H_
+
+#include <cstdint>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "kaldi-native-fbank/csrc/feature-window.h"
+#include "kaldi-native-fbank/csrc/mel-computations.h"
+
+namespace knf {
+
+struct WhisperFeatureOptions {
+  WhisperFeatureOptions(const FrameExtractionOptions &frame_opts = {},
+                        int32_t dim = 80)
+      : frame_opts(frame_opts), dim(dim) {}
+
+  FrameExtractionOptions frame_opts;
+  int32_t dim = 80;
+
+  std::string ToString() const;
+};
+
+class WhisperFeatureComputer {
+ public:
+  // note: opts.frame_opts is ignored and we reset it inside
+  explicit WhisperFeatureComputer(const WhisperFeatureOptions &opts = {});
+
+  int32_t Dim() const { return opts_.dim; }
+
+  const FrameExtractionOptions &GetFrameOptions() const {
+    return opts_.frame_opts;
+  }
+
+  void Compute(float /*signal_raw_log_energy*/, float /*vtln_warp*/,
+               std::vector<float> *signal_frame, float *feature);
+
+  // if true, compute log_energy_pre_window but after dithering and dc removal
+  bool NeedRawLogEnergy() const { return false; }
+
+  using Options = WhisperFeatureOptions;
+
+ private:
+  std::unique_ptr<MelBanks> mel_banks_;
+  WhisperFeatureOptions opts_;
+};
+
+}  // namespace knf
+
+#endif  // KALDI_NATIVE_FBANK_CSRC_WHISPER_FEATURE_H_
diff --git a/projects/llm_framework/main_kws_new/test.py b/projects/llm_framework/main_kws_new/test.py
new file mode 100644
index 00000000..b73f9e7e
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/test.py
@@ -0,0 +1,115 @@
+import time
+import torch
+import torchaudio
+import torchaudio.compliance.kaldi as kaldi
+import numpy as np
+import onnxruntime as ort
+import os
+import glob
+
+frame_index_global = 0
+last_trigger_time_ms = -1e9
+count = 0  # 添加缺失的全局计数器
+
+def detect_wakeup(scores_chunk):
+    global count, last_trigger_time_ms, frame_index_global
+    above = scores_chunk > threshold
+    triggered = False
+    for flag in above:
+        if flag:
+            count += 1
+            if count >= min_continuous_frames:
+                trigger_time_ms = (frame_index_global - min_continuous_frames + 1) * 10
+                if (trigger_time_ms - last_trigger_time_ms) >= 2000:
+                    print(f"[唤醒触发] 时间: {trigger_time_ms} ms")
+                    last_trigger_time_ms = trigger_time_ms
+                    triggered = True
+        else:
+            count = 0
+        frame_index_global += 1
+    return triggered
+
+def save_mat_to_bin(mat_data, index):
+    """保存mat数据为二进制文件，格式与C++兼容"""
+    filename = f"mat_{index:03d}.bin"
+    
+    # mat_data是torch tensor，shape为 [frames, features]
+    mat_np = mat_data.numpy()
+    rows, cols = mat_np.shape
+    
+    with open(filename, 'wb') as f:
+        # 写入维度信息 (int64)
+        f.write(np.array([rows], dtype=np.int64).tobytes())
+        f.write(np.array([cols], dtype=np.int64).tobytes())
+        # 写入数据 (float32)
+        f.write(mat_np.astype(np.float32).tobytes())
+    
+    print(f"保存mat到 {filename}, shape: ({rows}, {cols})")
+    return filename
+
+# ================= 配置参数 =================
+wavfile = "./001.wav"                  # 测试音频路径
+onnx_model = "./t30.onnx"          # ONNX 模型路径
+feat_dim = 80                          # FBank 维度
+resample_rate = 16000                  # 目标采样率
+chunk_size = 32                        # 推理帧数（1 帧=10ms，32帧≈320ms）
+threshold = 0.9                        # 唤醒分数阈值
+min_continuous_frames = 5              # 连续高分帧数(>=此值判定唤醒)
+device = "cpu"
+
+# ================= 加载模型 =================
+ort_sess = ort.InferenceSession(onnx_model)
+
+# ================= 加载音频并预处理 =================
+waveform, sample_rate = torchaudio.load(wavfile)
+# 音量归一
+waveform = waveform * (1 << 15)
+
+cache = torch.zeros(1, 32, 88, dtype=torch.float).numpy()  # 初始化 cache
+scores = []          # 存储每帧分数
+frame_index = 0      # 已处理全局帧数
+save_index = 0
+save_dir = "debug_inputs"
+os.makedirs(save_dir, exist_ok=True)
+
+samples_per_chunk = int(resample_rate * 0.01 * chunk_size)  # 每个推理块采样点数
+total_samples = waveform.shape[1]
+print(f"音频总采样点数: {total_samples}, 每块点数: {samples_per_chunk}")
+
+start_sample = 0
+while start_sample < total_samples:
+    end_sample = start_sample + samples_per_chunk
+    audio_chunk = waveform[:, start_sample:end_sample]  # [1, T]
+    start_sample += samples_per_chunk
+    frame_index += chunk_size  # 累加帧数
+    
+    # ----------- run_inference1 逻辑的 fbank 计算 -----------
+    mat = kaldi.fbank(
+        audio_chunk,
+        num_mel_bins=feat_dim,
+        dither=0.0,
+        energy_floor=0.0,
+        sample_frequency=resample_rate,
+        snip_edges=False
+    )
+    
+    print(f"mat.shape={mat.shape}")
+    
+    # 保存mat为二进制文件
+    save_mat_to_bin(mat, save_index)
+    
+    # 推理
+    onnx_inputs = {
+        "input": mat.unsqueeze(0).numpy(),
+        "cache": cache
+    }
+    onnx_output = ort_sess.run(None, onnx_inputs)
+    out_chunk, cache = onnx_output[0].flatten(), onnx_output[1]  # 更新 cache
+    
+    if detect_wakeup(out_chunk):
+        print(f"[{time.strftime('%H:%M:%S')}] 检测到唤醒词！")
+    
+    scores.extend(out_chunk)
+    save_index += 1
+
+print(f"总共保存了 {save_index} 个mat文件")
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index aa1ecb25..c36b5c51 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -185,10 +185,16 @@ def create_data_deb(package_name, version, src_folder, revision = 'm5stack1', de
             with open(mode_config_file, 'r', encoding='utf-8') as file:
                 data = json.load(file)
             for scripts_file in data['mode_param']['ext_scripts']:
-                tokenizer_py_file = os.path.join(src_folder, scripts_file)
-                if os.path.exists(tokenizer_py_file):
-                    os.makedirs(os.path.join(deb_folder, 'opt/m5stack/scripts'), exist_ok = True)
-                    shutil.copy2(tokenizer_py_file, os.path.join(deb_folder, 'opt/m5stack/scripts', scripts_file))
+                src_path = os.path.join(src_folder, scripts_file)
+                dst_path = os.path.join(deb_folder, 'opt/m5stack/scripts', scripts_file)
+                if os.path.exists(src_path):
+                    if os.path.isdir(src_path):
+                        # 拷贝整个文件夹
+                        shutil.copytree(src_path, dst_path, dirs_exist_ok=True)
+                    else:
+                        # 确保目标目录存在
+                        os.makedirs(os.path.dirname(dst_path), exist_ok=True)
+                        shutil.copy2(src_path, dst_path)
         except:
             pass
     else:
@@ -263,7 +269,11 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         vlm_dir = os.path.join(src_folder, 'vlm')
         if os.path.exists(vlm_dir):
             shutil.copytree(vlm_dir, os.path.join(deb_folder, 'opt/m5stack/lib/vlm'))
-    
+    if package_name == 'llm-cosy-voice':
+        cosy_voice_dir = os.path.join(src_folder, 'cosy-voice')
+        if os.path.exists(cosy_voice_dir):
+            shutil.copytree(cosy_voice_dir, os.path.join(deb_folder, 'opt/m5stack/lib/cosy-voice'))
+
     bin_file_name = package_name.replace("-", "_")
     if version_info != 0.0:
         bin_file_name = package_name.replace("-", "_") + f'-{version}'
@@ -369,6 +379,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-vad':[create_bin_deb,'llm-vad', '1.8', src_folder, revision],
         'llm-whisper':[create_bin_deb,'llm-whisper', '1.8', src_folder, revision],
         'llm-openai-api':[create_bin_deb,'llm-openai-api', '1.7', src_folder, revision],
+        'llm-cosy-voice':[create_bin_deb,'llm-cosy-voice', '1.8', src_folder, revision],
         # keyword spotting Audio file
         'llm-model-audio-en-us':[create_data_deb,'llm-model-audio-en-us', data_version, src_folder, revision],
         'llm-model-audio-zh-cn':[create_data_deb,'llm-model-audio-zh-cn', data_version, src_folder, revision],
@@ -380,6 +391,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         # TTS model
         'llm-model-single-speaker-english-fast':[create_data_deb,'llm-model-single-speaker-english-fast', '0.3', src_folder, revision],
         'llm-model-single-speaker-fast':[create_data_deb,'llm-model-single-speaker-fast', '0.3', src_folder, revision],
+        'llm-model-CosyVoice2-0.5B-ax650':[create_data_deb,'llm-model-CosyVoice2-0.5B-ax650', '0.6', src_folder, revision],
         # VAD model
         'llm-model-silero-vad':[create_data_deb,'llm-model-silero-vad', '0.4', src_folder, revision],
         # MeloTTS model

From 2a5c1394333fa637f0e3ee5a65c32554b5f85629 Mon Sep 17 00:00:00 2001
From: Little Mouse <n15978654429@gmail.com>
Date: Tue, 23 Sep 2025 16:43:34 +0800
Subject: [PATCH 44/79] update static version

---
 projects/llm_framework/SConstruct | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/projects/llm_framework/SConstruct b/projects/llm_framework/SConstruct
index 21134df7..26ed7343 100644
--- a/projects/llm_framework/SConstruct
+++ b/projects/llm_framework/SConstruct
@@ -5,7 +5,7 @@ import shutil
 os.environ['SDK_PATH'] = os.path.normpath(str(Path(os.getcwd())/'..'/'..'/'SDK'))
 os.environ['EXT_COMPONENTS_PATH'] = os.path.normpath(str(Path(os.getcwd())/'..'/'..'/'ext_components'))
 
-version = 'v0.1.0'
+version = 'v0.1.1'
 static_lib = 'static_lib'
 update = False
 

From d48972352a44a367fd6bd5d2cab20f87e780ac2c Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 25 Sep 2025 16:55:32 +0800
Subject: [PATCH 45/79] [update] clean code

---
 .../llm_framework/main_kws_new/SConstruct     |   5 +-
 .../llm_framework/main_kws_new/src/main.cpp   | 198 +-------
 .../runner/kaldi-native-fbank/CMakeLists.txt  |   8 -
 .../kaldi-native-fbank/csrc/CMakeLists.txt    | 104 ----
 .../kaldi-native-fbank/csrc/CPPLINT.cfg       |   1 -
 .../kaldi-native-fbank/csrc/feature-fbank.cc  | 116 -----
 .../kaldi-native-fbank/csrc/feature-fbank.h   | 136 ------
 .../csrc/feature-functions.cc                 |  49 --
 .../csrc/feature-functions.h                  |  38 --
 .../kaldi-native-fbank/csrc/feature-mfcc.cc   | 174 -------
 .../kaldi-native-fbank/csrc/feature-mfcc.h    | 151 ------
 .../csrc/feature-raw-audio-samples.cc         |  38 --
 .../csrc/feature-raw-audio-samples.h          |  81 ----
 .../kaldi-native-fbank/csrc/feature-window.cc | 274 -----------
 .../kaldi-native-fbank/csrc/feature-window.h  | 187 -------
 .../runner/kaldi-native-fbank/csrc/istft.cc   | 158 ------
 .../runner/kaldi-native-fbank/csrc/istft.h    |  45 --
 .../kaldi-native-fbank/csrc/kaldi-math.cc     |  51 --
 .../kaldi-native-fbank/csrc/kaldi-math.h      |  52 --
 .../src/runner/kaldi-native-fbank/csrc/log.cc | 142 ------
 .../src/runner/kaldi-native-fbank/csrc/log.h  | 383 ---------------
 .../csrc/mel-computations.cc                  | 457 ------------------
 .../csrc/mel-computations.h                   | 190 --------
 .../kaldi-native-fbank/csrc/online-feature.cc | 170 -------
 .../kaldi-native-fbank/csrc/online-feature.h  | 155 ------
 .../runner/kaldi-native-fbank/csrc/rfft.cc    | 112 -----
 .../src/runner/kaldi-native-fbank/csrc/rfft.h |  58 ---
 .../runner/kaldi-native-fbank/csrc/stft.cc    | 142 ------
 .../src/runner/kaldi-native-fbank/csrc/stft.h |  65 ---
 .../kaldi-native-fbank/csrc/test-log.cc       |  71 ---
 .../csrc/test-online-fbank.cc                 |  54 ---
 .../csrc/test-online-feature.cc               |  59 ---
 .../kaldi-native-fbank/csrc/test-rfft.cc      | 101 ----
 .../csrc/whisper-feature.cc                   | 169 -------
 .../kaldi-native-fbank/csrc/whisper-feature.h |  69 ---
 projects/llm_framework/main_kws_new/test.py   | 115 -----
 36 files changed, 4 insertions(+), 4374 deletions(-)
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/CMakeLists.txt
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CMakeLists.txt
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CPPLINT.cfg
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.h
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.h
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.h
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.h
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.h
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.h
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.h
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.h
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.h
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.h
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.h
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.h
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-log.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-fbank.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-feature.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-rfft.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.cc
 delete mode 100644 projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.h
 delete mode 100644 projects/llm_framework/main_kws_new/test.py

diff --git a/projects/llm_framework/main_kws_new/SConstruct b/projects/llm_framework/main_kws_new/SConstruct
index bc38274f..2e327d95 100644
--- a/projects/llm_framework/main_kws_new/SConstruct
+++ b/projects/llm_framework/main_kws_new/SConstruct
@@ -23,11 +23,8 @@ LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 
 INCLUDE += [
-    # ADir('../static_lib/include/sherpa'),
-            # ADir('../static_lib/include/sherpa/sherpa-onnx'),
+            ADir('../static_lib/include/sherpa/'),
             ADir('../static_lib/include/sherpa/sherpa-onnx/onnxruntime-src'),
-            ADir('src/runner'),
-            # ADir('../static_lib/include/sherpa/sherpa-onnx/openfst-src')
             ]
 
 LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/fbank')]
diff --git a/projects/llm_framework/main_kws_new/src/main.cpp b/projects/llm_framework/main_kws_new/src/main.cpp
index f6664745..b56f3c60 100644
--- a/projects/llm_framework/main_kws_new/src/main.cpp
+++ b/projects/llm_framework/main_kws_new/src/main.cpp
@@ -22,9 +22,7 @@
 
 #include <onnxruntime_cxx_api.h>
 
-// #include <kaldi-native-fbank/csrc/feature-fbank.h>
 #include "kaldi-native-fbank/csrc/online-feature.h"
-// #include <kaldi-native-fbank/csrc/feature-window.h>
 
 using namespace StackFlows;
 
@@ -214,29 +212,6 @@ class llm_task {
         return triggered;
     }
 
-    std::vector<std::vector<float>> compute_fbank_kaldi1(const std::vector<float> &waveform, int sample_rate,
-                                                         int num_mel_bins)
-    {
-        fbank_.reset();
-        fbank_ = std::make_unique<knf::OnlineFbank>(opts_);
-
-        fbank_->AcceptWaveform(sample_rate, waveform.data(), waveform.size());
-
-        int num_frames = fbank_->NumFramesReady();
-
-        std::vector<std::vector<float>> features;
-        SLOGE("num_frames %d", num_frames);
-        features.reserve(num_frames);
-
-        for (int i = 0; i < num_frames; ++i) {
-            const float *frame_data = fbank_->GetFrame(i);
-            std::vector<float> frame(frame_data, frame_data + num_mel_bins);
-            features.push_back(std::move(frame));
-        }
-
-        return features;
-    }
-
     std::vector<std::vector<float>> compute_fbank_kaldi(const std::vector<float> &waveform, int sample_rate,
                                                         int num_mel_bins)
     {
@@ -244,8 +219,9 @@ class llm_task {
         fbank_ = std::make_unique<knf::OnlineFbank>(opts_);
         fbank_->AcceptWaveform(sample_rate, waveform.data(), waveform.size());
         int num_frames = fbank_->NumFramesReady();
+
         std::vector<std::vector<float>> features;
-        SLOGE("num_frames %d", num_frames);
+
         features.reserve(num_frames);
         for (int i = 0; i < num_frames; ++i) {
             const float *frame_data = fbank_->GetFrame(i);
@@ -253,169 +229,13 @@ class llm_task {
             features.push_back(std::move(frame));
         }
 
-        // 生成文件名序号（三位数，补零）
-        // std::stringstream ss;
-        // ss << std::setfill('0') << std::setw(3) << file_counter;
-        // std::string file_suffix = ss.str();
-
-        // 保存 waveform 为二进制
-        // std::string waveform_filename = "waveform_" + file_suffix + ".bin";
-        // std::ofstream waveform_bin(waveform_filename, std::ios::binary);
-        // if (waveform_bin.is_open()) {
-        //     size_t size = waveform.size();
-        //     waveform_bin.write(reinterpret_cast<const char *>(&size), sizeof(size));
-        //     waveform_bin.write(reinterpret_cast<const char *>(waveform.data()), sizeof(float) * waveform.size());
-        //     waveform_bin.close();
-        //     SLOGE("Waveform saved to %s", waveform_filename.c_str());
-        // }
-
-        // 保存 features 为二进制
-        // std::string features_filename = "features_" + file_suffix + ".bin";
-        // std::ofstream features_bin(features_filename, std::ios::binary);
-        // if (features_bin.is_open()) {
-        //     size_t rows = features.size();
-        //     size_t cols = (rows > 0) ? features[0].size() : 0;
-        //     features_bin.write(reinterpret_cast<const char *>(&rows), sizeof(rows));
-        //     features_bin.write(reinterpret_cast<const char *>(&cols), sizeof(cols));
-        //     for (const auto &row : features) {
-        //         features_bin.write(reinterpret_cast<const char *>(row.data()), sizeof(float) * row.size());
-        //     }
-        //     features_bin.close();
-        //     SLOGE("Features saved to %s", features_filename.c_str());
-        // }
-
-        // 递增计数器
-        // file_counter++;
-
         return features;
     }
 
-    std::vector<std::vector<float>> read_mat_from_bin(int index)
-    {
-        std::stringstream ss;
-        ss << "mat_" << std::setfill('0') << std::setw(3) << index << ".bin";
-        std::string filename = ss.str();
-
-        std::ifstream file(filename, std::ios::binary);
-        if (!file.is_open()) {
-            SLOGE("无法打开文件: %s", filename.c_str());
-            return {};
-        }
-
-        // 读取维度信息
-        int64_t rows, cols;
-        file.read(reinterpret_cast<char *>(&rows), sizeof(int64_t));
-        file.read(reinterpret_cast<char *>(&cols), sizeof(int64_t));
-
-        SLOGE("读取mat文件 %s, shape: (%ld, %ld)", filename.c_str(), rows, cols);
-
-        // 读取数据
-        std::vector<float> data(rows * cols);
-        file.read(reinterpret_cast<char *>(data.data()), sizeof(float) * data.size());
-        file.close();
-
-        // 转换为二维向量
-        std::vector<std::vector<float>> result;
-        result.reserve(rows);
-
-        for (int i = 0; i < rows; ++i) {
-            std::vector<float> row;
-            row.reserve(cols);
-            for (int j = 0; j < cols; ++j) {
-                row.push_back(data[i * cols + j]);
-            }
-            result.push_back(std::move(row));
-        }
-
-        return result;
-    }
-
-    std::vector<float> run_inference1(const std::vector<float> &audio_chunk_16k)
-    {
-        auto fbank_feats = compute_fbank_kaldi(audio_chunk_16k, RESAMPLE_RATE, FEAT_DIM);
-        SLOGE("=== FBank Features ===");
-        for (int i = 0; i < fbank_feats.size(); ++i) {
-            SLOGE("Frame %d:", i);
-            std::string frame_str = "";
-            for (int j = 0; j < fbank_feats[i].size(); ++j) {
-                frame_str += std::to_string(fbank_feats[i][j]) + " ";
-            }
-            SLOGE("%s", frame_str.c_str());
-        }
-        SLOGE("fbank_feats.size()=%d", (int)fbank_feats.size());
-        SLOGE("=======================");
-
-        if (fbank_feats.empty()) {
-            return {};
-        }
-
-        int T = fbank_feats.size();
-        std::vector<float> mat_flattened;
-        for (const auto &feat : fbank_feats) {
-            mat_flattened.insert(mat_flattened.end(), feat.begin(), feat.end());
-        }
-
-        std::vector<int64_t> input_shape = {1, static_cast<int64_t>(T), FEAT_DIM};
-        std::vector<int64_t> cache_shape = {1, 32, 88};
-
-        Ort::MemoryInfo memory_info = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-
-        Ort::Value input_tensor = Ort::Value::CreateTensor<float>(
-            memory_info, mat_flattened.data(), mat_flattened.size(), input_shape.data(), input_shape.size());
-
-        Ort::Value cache_tensor = Ort::Value::CreateTensor<float>(memory_info, cache.data(), cache.size(),
-                                                                  cache_shape.data(), cache_shape.size());
-
-        const char *input_names[]  = {"input", "cache"};
-        const char *output_names[] = {"output", "r_cache"};
-
-        std::vector<Ort::Value> inputs;
-        inputs.push_back(std::move(input_tensor));
-        inputs.push_back(std::move(cache_tensor));
-
-        auto output_tensors = session->Run(Ort::RunOptions{nullptr}, input_names, inputs.data(), 2, output_names, 2);
-
-        float *out_data       = output_tensors[0].GetTensorMutableData<float>();
-        float *cache_out_data = output_tensors[1].GetTensorMutableData<float>();
-
-        std::vector<int64_t> out_shape = output_tensors[0].GetTensorTypeAndShapeInfo().GetShape();
-        size_t out_size                = 1;
-        for (auto dim : out_shape) out_size *= dim;
-
-        std::vector<float> out_chunk(out_data, out_data + out_size);
-
-        std::copy(cache_out_data, cache_out_data + cache.size(), cache.begin());
-
-        return out_chunk;
-    }
-
     std::vector<float> run_inference(const std::vector<float> &audio_chunk_16k)
     {
         std::vector<std::vector<float>> fbank_feats;
-        static int file_index = -1;
-        // file_index++;
-        if (file_index >= 0) {
-            // 从文件读取
-            fbank_feats = read_mat_from_bin(file_index);
-            SLOGE("=== 从文件读取的 FBank Features ===");
-        } else {
-            // 原来的计算方式
-            fbank_feats = compute_fbank_kaldi(audio_chunk_16k, RESAMPLE_RATE, FEAT_DIM);
-            SLOGE("=== 计算的 FBank Features ===");
-        }
-        if (file_index == 6) file_index = 0;
-
-        // 打印特征用于对比
-        // for (int i = 0; i < fbank_feats.size(); ++i) {
-        //     SLOGE("Frame %d:", i);
-        //     std::string frame_str = "";
-        //     for (int j = 0; j < fbank_feats[i].size(); ++j) {
-        //         frame_str += std::to_string(fbank_feats[i][j]) + " ";
-        //     }
-        //     SLOGE("%s", frame_str.c_str());
-        // }
-        // SLOGE("fbank_feats.size()=%d", (int)fbank_feats.size());
-        // SLOGE("=======================");
+        fbank_feats = compute_fbank_kaldi(audio_chunk_16k, RESAMPLE_RATE, FEAT_DIM);
 
         if (fbank_feats.empty()) {
             return {};
@@ -461,9 +281,6 @@ class llm_task {
     void sys_pcm_on_data(const std::string &raw)
     {
         static int count = 0;
-        static std::ofstream ffout("audio_float.raw", std::ios::binary | std::ios::app);
-        static std::ofstream fout("audio_int16.raw", std::ios::binary | std::ios::app);
-
         if (count < delay_audio_frame_) {
             buffer_write_char(pcmdata, raw.data(), raw.length());
             count++;
@@ -487,18 +304,9 @@ class llm_task {
         buffer_resize(pcmdata, 0);
         count = 0;
 
-        if (!floatSamples.empty()) {
-            ffout.write(reinterpret_cast<const char *>(floatSamples.data()), floatSamples.size() * sizeof(float));
-        }
-
-        if (!int16Samples.empty()) {
-            fout.write(reinterpret_cast<const char *>(int16Samples.data()), int16Samples.size() * sizeof(int16_t));
-        }
-
         auto scores = run_inference(floatSamples);
         if (detect_wakeup(scores)) {
             if (enwake_audio_ && (!wake_wav_file_.empty()) && play_awake_wav) {
-                SLOGE("\n\n\n\n\n");
                 play_awake_wav(wake_wav_file_);
             }
             if (out_callback_) out_callback_("", true);
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/CMakeLists.txt b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/CMakeLists.txt
deleted file mode 100644
index 20376267..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/CMakeLists.txt
+++ /dev/null
@@ -1,8 +0,0 @@
-add_subdirectory(csrc)
-
-if(KALDI_NATIVE_FBANK_BUILD_PYTHON)
-  message(STATUS "Building Python")
-  add_subdirectory(python)
-else()
-  message(STATUS "Disable building Python")
-endif()
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CMakeLists.txt b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CMakeLists.txt
deleted file mode 100644
index 741c2340..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CMakeLists.txt
+++ /dev/null
@@ -1,104 +0,0 @@
-
-include_directories(${PROJECT_SOURCE_DIR})
-set(sources
-  feature-fbank.cc
-  feature-functions.cc
-  feature-mfcc.cc
-  feature-raw-audio-samples.cc
-  feature-window.cc
-  istft.cc
-  kaldi-math.cc
-  mel-computations.cc
-  online-feature.cc
-  rfft.cc
-  stft.cc
-  whisper-feature.cc
-)
-
-if(KALDI_NATIVE_FBANK_ENABLE_CHECK)
-  list(APPEND sources log.cc)
-endif()
-
-add_library(kaldi-native-fbank-core ${sources})
-if(KALDI_NATIVE_FBANK_ENABLE_CHECK)
-  target_compile_definitions(kaldi-native-fbank-core PUBLIC KNF_ENABLE_CHECK=1)
-
-  if(KNF_HAVE_EXECINFO_H)
-    target_compile_definitions(kaldi-native-fbank-core PRIVATE KNF_HAVE_EXECINFO_H=1)
-  endif()
-
-  if(KNF_HAVE_CXXABI_H)
-    target_compile_definitions(kaldi-native-fbank-core PRIVATE KNF_HAVE_CXXABI_H=1)
-  endif()
-endif()
-
-# We are using std::call_once() in log.h,which requires us to link with -pthread
-if(NOT WIN32 AND KALDI_NATIVE_FBANK_ENABLE_CHECK)
-  target_link_libraries(kaldi-native-fbank-core -pthread)
-endif()
-target_link_libraries(kaldi-native-fbank-core kissfft)
-
-if(KALDI_NATIVE_FBANK_BUILD_TESTS)
-  add_executable(test-online-fbank test-online-fbank.cc)
-  target_link_libraries(test-online-fbank kaldi-native-fbank-core)
-endif()
-
-function(kaldi_native_fbank_add_test source)
-  get_filename_component(name ${source} NAME_WE)
-  add_executable(${name} "${source}")
-  target_link_libraries(${name}
-    PRIVATE
-      kaldi-native-fbank-core
-      gtest
-      gtest_main
-  )
-
-  add_test(NAME "Test.${name}"
-    COMMAND
-    $<TARGET_FILE:${name}>
-  )
-endfunction()
-
-# please sort the source files alphabetically
-set(test_srcs
-  # test-online-feature.cc
-  test-log.cc
-  test-rfft.cc
-)
-
-if(KALDI_NATIVE_FBANK_BUILD_TESTS)
-  foreach(source IN LISTS test_srcs)
-    kaldi_native_fbank_add_test(${source})
-  endforeach()
-endif()
-
-install(TARGETS kaldi-native-fbank-core
-  DESTINATION lib
-)
-if(KALDI_NATIVE_FBANK_BUILD_PYTHON AND WIN32)
-  install(TARGETS kaldi-native-fbank-core
-    DESTINATION ..
-  )
-endif()
-
-if(KALDI_NATIVE_FBANK_BUILD_TESTS)
-  install(TARGETS test-online-fbank
-    DESTINATION bin
-  )
-endif()
-
-file(MAKE_DIRECTORY
-  DESTINATION
-    ${PROJECT_BINARY_DIR}/include/kaldi-native-fbank/csrc
-)
-file(GLOB_RECURSE all_headers *.h)
-
-file(COPY
-  ${all_headers}
-  DESTINATION
-    ${PROJECT_BINARY_DIR}/include/kaldi-native-fbank/csrc
-)
-
-install(FILES ${all_headers}
-  DESTINATION include/kaldi-native-fbank/csrc
-)
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CPPLINT.cfg b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CPPLINT.cfg
deleted file mode 100644
index 4c1a8fb3..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/CPPLINT.cfg
+++ /dev/null
@@ -1 +0,0 @@
-exclude_files=fftsg.cc
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.cc
deleted file mode 100644
index b73aa807..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.cc
+++ /dev/null
@@ -1,116 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-// This file is copied/modified from kaldi/src/feat/feature-fbank.cc
-//
-#include "kaldi-native-fbank/csrc/feature-fbank.h"
-
-#include <algorithm>
-#include <limits>
-#include <utility>
-#include <vector>
-
-#include "kaldi-native-fbank/csrc/feature-functions.h"
-#include "kaldi-native-fbank/csrc/kaldi-math.h"
-#include "kaldi-native-fbank/csrc/log.h"
-
-namespace knf {
-
-std::ostream &operator<<(std::ostream &os, const FbankOptions &opts) {
-  os << opts.ToString();
-  return os;
-}
-
-FbankComputer::FbankComputer(const FbankOptions &opts)
-    : opts_(opts), rfft_(opts.frame_opts.PaddedWindowSize()) {
-  if (opts.energy_floor > 0.0f) {
-    log_energy_floor_ = logf(opts.energy_floor);
-  }
-
-  // We'll definitely need the filterbanks info for VTLN warping factor 1.0.
-  // [note: this call caches it.]
-  GetMelBanks(1.0f);
-}
-
-FbankComputer::~FbankComputer() {
-  for (auto iter = mel_banks_.begin(); iter != mel_banks_.end(); ++iter)
-    delete iter->second;
-}
-
-const MelBanks *FbankComputer::GetMelBanks(float vtln_warp) {
-  MelBanks *this_mel_banks = nullptr;
-
-  // std::map<float, MelBanks *>::iterator iter = mel_banks_.find(vtln_warp);
-  auto iter = mel_banks_.find(vtln_warp);
-  if (iter == mel_banks_.end()) {
-    this_mel_banks = new MelBanks(opts_.mel_opts, opts_.frame_opts, vtln_warp);
-    mel_banks_[vtln_warp] = this_mel_banks;
-  } else {
-    this_mel_banks = iter->second;
-  }
-  return this_mel_banks;
-}
-
-void FbankComputer::Compute(float signal_raw_log_energy, float vtln_warp,
-                            std::vector<float> *signal_frame, float *feature) {
-  const MelBanks &mel_banks = *(GetMelBanks(vtln_warp));
-
-  KNF_CHECK_EQ(signal_frame->size(), opts_.frame_opts.PaddedWindowSize());
-
-  // Compute energy after window function (not the raw one).
-  if (opts_.use_energy && !opts_.raw_energy) {
-    signal_raw_log_energy = std::log(
-        std::max<float>(InnerProduct(signal_frame->data(), signal_frame->data(),
-                                     signal_frame->size()),
-                        std::numeric_limits<float>::epsilon()));
-  }
-  rfft_.Compute(signal_frame->data());  // signal_frame is modified in-place
-  ComputePowerSpectrum(signal_frame);
-
-  // Use magnitude instead of power if requested.
-  if (!opts_.use_power) {
-    Sqrt(signal_frame->data(), signal_frame->size() / 2 + 1);
-  }
-
-  int32_t mel_offset = ((opts_.use_energy && !opts_.htk_compat) ? 1 : 0);
-
-  // Its length is opts_.mel_opts.num_bins
-  float *mel_energies = feature + mel_offset;
-
-  // Sum with mel filter banks over the power spectrum
-  mel_banks.Compute(signal_frame->data(), mel_energies);
-
-  if (opts_.use_log_fbank) {
-    // Avoid log of zero (which should be prevented anyway by dithering).
-    for (int32_t i = 0; i != opts_.mel_opts.num_bins; ++i) {
-      auto t = std::max(mel_energies[i], std::numeric_limits<float>::epsilon());
-      mel_energies[i] = std::log(t);
-    }
-  }
-
-  // Copy energy as first value (or the last, if htk_compat == true).
-  if (opts_.use_energy) {
-    if (opts_.energy_floor > 0.0 && signal_raw_log_energy < log_energy_floor_) {
-      signal_raw_log_energy = log_energy_floor_;
-    }
-    int32_t energy_index = opts_.htk_compat ? opts_.mel_opts.num_bins : 0;
-    feature[energy_index] = signal_raw_log_energy;
-  }
-}
-
-}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.h
deleted file mode 100644
index 84005e0d..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-fbank.h
+++ /dev/null
@@ -1,136 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-// This file is copied/modified from kaldi/src/feat/feature-fbank.h
-
-#ifndef KALDI_NATIVE_FBANK_CSRC_FEATURE_FBANK_H_
-#define KALDI_NATIVE_FBANK_CSRC_FEATURE_FBANK_H_
-
-#include <map>
-#include <string>
-#include <vector>
-#include <cstdint>
-#include <sstream>
-
-#include "kaldi-native-fbank/csrc/feature-window.h"
-#include "kaldi-native-fbank/csrc/mel-computations.h"
-#include "kaldi-native-fbank/csrc/rfft.h"
-
-namespace knf {
-
-struct FbankOptions {
-  FrameExtractionOptions frame_opts;
-  MelBanksOptions mel_opts;
-  // append an extra dimension with energy to the filter banks
-  bool use_energy = false;
-  float energy_floor = 0.0f;  // active iff use_energy==true
-
-  // If true, compute log_energy before preemphasis and windowing
-  // If false, compute log_energy after preemphasis ans windowing
-  bool raw_energy = true;  // active iff use_energy==true
-
-  // If true, put energy last (if using energy)
-  // If false, put energy first
-  bool htk_compat = false;  // active iff use_energy==true
-
-  // if true (default), produce log-filterbank, else linear
-  bool use_log_fbank = true;
-
-  // if true (default), use power in filterbank
-  // analysis, else magnitude.
-  bool use_power = true;
-
-  FbankOptions() { mel_opts.num_bins = 23; }
-
-  std::string ToString() const {
-    std::ostringstream os;
-    os << "frame_opts: \n";
-    os << frame_opts << "\n";
-    os << "\n";
-
-    os << "mel_opts: \n";
-    os << mel_opts << "\n";
-
-    os << "use_energy: " << use_energy << "\n";
-    os << "energy_floor: " << energy_floor << "\n";
-    os << "raw_energy: " << raw_energy << "\n";
-    os << "htk_compat: " << htk_compat << "\n";
-    os << "use_log_fbank: " << use_log_fbank << "\n";
-    os << "use_power: " << use_power << "\n";
-    return os.str();
-  }
-};
-
-std::ostream &operator<<(std::ostream &os, const FbankOptions &opts);
-
-class FbankComputer {
- public:
-  using Options = FbankOptions;
-
-  explicit FbankComputer(const FbankOptions &opts);
-  ~FbankComputer();
-
-  int32_t Dim() const {
-    return opts_.mel_opts.num_bins + (opts_.use_energy ? 1 : 0);
-  }
-
-  // if true, compute log_energy_pre_window but after dithering and dc removal
-  bool NeedRawLogEnergy() const { return opts_.use_energy && opts_.raw_energy; }
-
-  const FrameExtractionOptions &GetFrameOptions() const {
-    return opts_.frame_opts;
-  }
-
-  const FbankOptions &GetOptions() const { return opts_; }
-
-  /**
-     Function that computes one frame of features from
-     one frame of signal.
-
-     @param [in] signal_raw_log_energy The log-energy of the frame of the signal
-         prior to windowing and pre-emphasis, or
-         log(numeric_limits<float>::min()), whichever is greater.  Must be
-         ignored by this function if this class returns false from
-         this->NeedsRawLogEnergy().
-     @param [in] vtln_warp  The VTLN warping factor that the user wants
-         to be applied when computing features for this utterance.  Will
-         normally be 1.0, meaning no warping is to be done.  The value will
-         be ignored for feature types that don't support VLTN, such as
-         spectrogram features.
-     @param [in] signal_frame  One frame of the signal,
-       as extracted using the function ExtractWindow() using the options
-       returned by this->GetFrameOptions().  The function will use the
-       vector as a workspace, which is why it's a non-const pointer.
-     @param [out] feature  Pointer to a vector of size this->Dim(), to which
-         the computed feature will be written. It should be pre-allocated.
-  */
-  void Compute(float signal_raw_log_energy, float vtln_warp,
-               std::vector<float> *signal_frame, float *feature);
-
- private:
-  const MelBanks *GetMelBanks(float vtln_warp);
-
-  FbankOptions opts_;
-  float log_energy_floor_;
-  std::map<float, MelBanks *> mel_banks_;  // float is VTLN coefficient.
-  Rfft rfft_;
-};
-
-}  // namespace knf
-
-#endif  // KALDI_NATIVE_FBANK_CSRC_FEATURE_FBANK_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.cc
deleted file mode 100644
index 00ae4c79..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.cc
+++ /dev/null
@@ -1,49 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-// This file is copied/modified from kaldi/src/feat/feature-functions.cc
-
-#include "kaldi-native-fbank/csrc/feature-functions.h"
-
-#include <cstdint>
-#include <vector>
-
-namespace knf {
-
-void ComputePowerSpectrum(std::vector<float> *complex_fft) {
-  int32_t dim = complex_fft->size();
-
-  // now we have in complex_fft, first half of complex spectrum
-  // it's stored as [real0, realN/2, real1, im1, real2, im2, ...]
-
-  float *p = complex_fft->data();
-  int32_t half_dim = dim / 2;
-  float first_energy = p[0] * p[0];
-  float last_energy = p[1] * p[1];  // handle this special case
-
-  for (int32_t i = 1; i < half_dim; ++i) {
-    float real = p[i * 2];
-    float im = p[i * 2 + 1];
-    p[i] = real * real + im * im;
-  }
-  p[0] = first_energy;
-  p[half_dim] = last_energy;  // Will actually never be used, and anyway
-  // if the signal has been bandlimited sensibly this should be zero.
-}
-
-}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.h
deleted file mode 100644
index b2216224..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-functions.h
+++ /dev/null
@@ -1,38 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-// This file is copied/modified from kaldi/src/feat/feature-functions.h
-#ifndef KALDI_NATIVE_FBANK_CSRC_FEATURE_FUNCTIONS_H_
-#define KALDI_NATIVE_FBANK_CSRC_FEATURE_FUNCTIONS_H_
-
-#include <vector>
-namespace knf {
-
-// ComputePowerSpectrum converts a complex FFT (as produced by the FFT
-// functions in csrc/rfft.h), and converts it into
-// a power spectrum.  If the complex FFT is a vector of size n (representing
-// half of the complex FFT of a real signal of size n, as described there),
-// this function computes in the first (n/2) + 1 elements of it, the
-// energies of the fft bins from zero to the Nyquist frequency.  Contents of the
-// remaining (n/2) - 1 elements are undefined at output.
-
-void ComputePowerSpectrum(std::vector<float> *complex_fft);
-
-}  // namespace knf
-
-#endif  // KALDI_NATIVE_FBANK_CSRC_FEATURE_FUNCTIONS_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.cc
deleted file mode 100644
index 3f38f2fc..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.cc
+++ /dev/null
@@ -1,174 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-// This file is copied/modified from kaldi/src/feat/feature-mfcc.cc
-//
-#include "kaldi-native-fbank/csrc/feature-mfcc.h"
-
-#include <algorithm>
-#include <cmath>
-#include <limits>
-#include <utility>
-#include <vector>
-
-#include "kaldi-native-fbank/csrc/feature-functions.h"
-#include "kaldi-native-fbank/csrc/feature-window.h"
-#include "kaldi-native-fbank/csrc/kaldi-math.h"
-#include "kaldi-native-fbank/csrc/log.h"
-
-namespace knf {
-
-static std::vector<float> ComputeDctMatrix(int32_t num_rows, int32_t num_cols) {
-  // this function is copied from
-  // https://github.com/kaldi-asr/kaldi/blob/master/src/matrix/matrix-functions.cc#L592
-
-  std::vector<float> ans(num_rows * num_cols);
-  float *p = ans.data();
-
-  float normalizer = std::sqrt(1.0 / num_cols);  // normalizer for X_0
-
-  for (int32_t i = 0; i != num_cols; ++i) {
-    p[i] = normalizer;
-  }
-
-  normalizer = std::sqrt(2.0 / num_cols);  // normalizer for other elements
-
-  for (int32_t k = 1; k != num_rows; ++k) {
-    for (int32_t n = 0; n != num_cols; ++n) {
-      *(p + k * num_cols + n) =
-          normalizer *
-          std::cos(static_cast<double>(M_PI) / num_cols * (n + 0.5) * k);
-    }
-  }
-
-  return ans;
-}
-
-std::ostream &operator<<(std::ostream &os, const MfccOptions &opts) {
-  os << opts.ToString();
-  return os;
-}
-
-MfccComputer::MfccComputer(const MfccOptions &opts)
-    : opts_(opts),
-      rfft_(opts.frame_opts.PaddedWindowSize()),
-      mel_energies_(opts.mel_opts.num_bins) {
-  if (opts.energy_floor > 0.0f) {
-    log_energy_floor_ = logf(opts.energy_floor);
-  }
-
-  // We'll definitely need the filterbanks info for VTLN warping factor 1.0.
-  // [note: this call caches it.]
-  GetMelBanks(1.0f);
-
-  int32_t num_bins = opts.mel_opts.num_bins;
-
-  KNF_CHECK_LE(opts.num_ceps, num_bins)
-      << "num-ceps cannot be larger than num-mel-bins."
-      << " It should be smaller or equal. You provided num-ceps: "
-      << opts.num_ceps << "  and num-mel-bins: " << num_bins;
-
-  dct_matrix_ = ComputeDctMatrix(opts.num_ceps, num_bins);
-
-  if (opts.cepstral_lifter != 0.0) {
-    lifter_coeffs_ = std::vector<float>(opts.num_ceps);
-    // torch::empty({1, opts.num_ceps}, torch::kFloat32);
-    ComputeLifterCoeffs(opts.cepstral_lifter, &lifter_coeffs_);
-  }
-}
-
-MfccComputer::~MfccComputer() {
-  for (auto iter = mel_banks_.begin(); iter != mel_banks_.end(); ++iter)
-    delete iter->second;
-}
-
-const MelBanks *MfccComputer::GetMelBanks(float vtln_warp) {
-  MelBanks *this_mel_banks = nullptr;
-
-  // std::map<float, MelBanks *>::iterator iter = mel_banks_.find(vtln_warp);
-  auto iter = mel_banks_.find(vtln_warp);
-  if (iter == mel_banks_.end()) {
-    this_mel_banks = new MelBanks(opts_.mel_opts, opts_.frame_opts, vtln_warp);
-    mel_banks_[vtln_warp] = this_mel_banks;
-  } else {
-    this_mel_banks = iter->second;
-  }
-  return this_mel_banks;
-}
-
-void MfccComputer::Compute(float signal_raw_log_energy, float vtln_warp,
-                           std::vector<float> *signal_frame, float *feature) {
-  const MelBanks &mel_banks = *(GetMelBanks(vtln_warp));
-
-  KNF_CHECK_EQ(signal_frame->size(), opts_.frame_opts.PaddedWindowSize());
-
-  // Compute energy after window function (not the raw one).
-  if (opts_.use_energy && !opts_.raw_energy) {
-    signal_raw_log_energy = std::log(
-        std::max<float>(InnerProduct(signal_frame->data(), signal_frame->data(),
-                                     signal_frame->size()),
-                        std::numeric_limits<float>::epsilon()));
-  }
-  rfft_.Compute(signal_frame->data());  // signal_frame is modified in-place
-  ComputePowerSpectrum(signal_frame);
-
-  // Sum with mel filter banks over the power spectrum
-  mel_banks.Compute(signal_frame->data(), mel_energies_.data());
-
-  // Avoid log of zero (which should be prevented anyway by dithering).
-  for (int32_t i = 0; i != opts_.mel_opts.num_bins; ++i) {
-    auto t = std::max(mel_energies_[i], std::numeric_limits<float>::epsilon());
-    mel_energies_[i] = std::log(t);
-  }
-
-  // feature = dct_matrix_ * mel_energies [which now have log]
-  for (int32_t i = 0; i != opts_.num_ceps; ++i) {
-    feature[i] = InnerProduct(dct_matrix_.data() + i * opts_.mel_opts.num_bins,
-                              mel_energies_.data(), opts_.mel_opts.num_bins);
-  }
-
-  if (opts_.cepstral_lifter != 0.0) {
-    for (int32_t i = 0; i != opts_.num_ceps; ++i) {
-      feature[i] *= lifter_coeffs_[i];
-    }
-  }
-
-  if (opts_.use_energy) {
-    if (opts_.energy_floor > 0.0 && signal_raw_log_energy < log_energy_floor_) {
-      signal_raw_log_energy = log_energy_floor_;
-    }
-    feature[0] = signal_raw_log_energy;
-  }
-
-  if (opts_.htk_compat) {
-    double energy = feature[0];
-
-    for (int32_t i = 0; i < opts_.num_ceps - 1; ++i) {
-      feature[i] = feature[i + 1];
-    }
-
-    if (!opts_.use_energy) {
-      energy *= M_SQRT2;  // scale on C0 (actually removing a scale
-    }
-    // we previously added that's part of one common definition of
-    // the cosine transform.)
-    feature[opts_.num_ceps - 1] = static_cast<float>(energy);
-  }
-}
-
-}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.h
deleted file mode 100644
index 32dc772e..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-mfcc.h
+++ /dev/null
@@ -1,151 +0,0 @@
-/**
- * Copyright 2009-2011  Karel Vesely;  Petr Motlicek;  Saarland University
- *           2014-2016  Johns Hopkins University (author: Daniel Povey)
- * Copyright 2024       Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-// This file is copied/modified from kaldi/src/feat/feature-mfcc.h
-
-#ifndef KALDI_NATIVE_FBANK_CSRC_FEATURE_MFCC_H_
-#define KALDI_NATIVE_FBANK_CSRC_FEATURE_MFCC_H_
-
-#include <cstdint>
-#include <map>
-#include <sstream>
-#include <string>
-#include <vector>
-
-#include "kaldi-native-fbank/csrc/feature-window.h"
-#include "kaldi-native-fbank/csrc/mel-computations.h"
-#include "kaldi-native-fbank/csrc/rfft.h"
-
-namespace knf {
-
-/// MfccOptions contains basic options for computing MFCC features.
-// (this class is copied from kaldi)
-struct MfccOptions {
-  FrameExtractionOptions frame_opts;
-  MelBanksOptions mel_opts;
-
-  // Number of cepstra in MFCC computation (including C0)
-  int32_t num_ceps = 13;
-
-  // Use energy (not C0) in MFCC computation
-  bool use_energy = true;
-
-  // Floor on energy (absolute, not relative) in MFCC
-  // computation. Only makes a difference if use_energy=true;
-  // only necessary if dither=0.0.
-  // Suggested values: 0.1 or 1.0
-  float energy_floor = 0.0;
-
-  // If true, compute energy before preemphasis and windowing
-  bool raw_energy = true;
-
-  // Constant that controls scaling of MFCCs
-  float cepstral_lifter = 22.0;
-
-  // If true, put energy or C0 last and use a factor of
-  // sqrt(2) on C0.
-  // Warning: not sufficient to get HTK compatible features
-  // (need to change other parameters)
-  bool htk_compat = false;
-
-  MfccOptions() { mel_opts.num_bins = 23; }
-
-  std::string ToString() const {
-    std::ostringstream os;
-    os << "MfccOptions(";
-    os << "frame_opts=" << frame_opts.ToString() << ", ";
-    os << "mel_opts=" << mel_opts.ToString() << ", ";
-
-    os << "num_ceps=" << num_ceps << ", ";
-    os << "use_energy=" << (use_energy ? "True" : "False") << ", ";
-    os << "energy_floor=" << energy_floor << ", ";
-    os << "raw_energy=" << (raw_energy ? "True" : "False") << ", ";
-    os << "cepstral_lifter=" << cepstral_lifter << ", ";
-    os << "htk_compat=" << (htk_compat ? "True" : "False") << ")";
-
-    return os.str();
-  }
-};
-
-std::ostream &operator<<(std::ostream &os, const MfccOptions &opts);
-
-class MfccComputer {
- public:
-  using Options = MfccOptions;
-
-  explicit MfccComputer(const MfccOptions &opts);
-  ~MfccComputer();
-
-  int32_t Dim() const { return opts_.num_ceps; }
-
-  // if true, compute log_energy_pre_window but after dithering and dc removal
-  bool NeedRawLogEnergy() const { return opts_.use_energy && opts_.raw_energy; }
-
-  const FrameExtractionOptions &GetFrameOptions() const {
-    return opts_.frame_opts;
-  }
-
-  const MfccOptions &GetOptions() const { return opts_; }
-
-  /**
-     Function that computes one frame of features from
-     one frame of signal.
-
-     @param [in] signal_raw_log_energy The log-energy of the frame of the signal
-         prior to windowing and pre-emphasis, or
-         log(numeric_limits<float>::min()), whichever is greater.  Must be
-         ignored by this function if this class returns false from
-         this->NeedsRawLogEnergy().
-     @param [in] vtln_warp  The VTLN warping factor that the user wants
-         to be applied when computing features for this utterance.  Will
-         normally be 1.0, meaning no warping is to be done.  The value will
-         be ignored for feature types that don't support VLTN, such as
-         spectrogram features.
-     @param [in] signal_frame  One frame of the signal,
-       as extracted using the function ExtractWindow() using the options
-       returned by this->GetFrameOptions().  The function will use the
-       vector as a workspace, which is why it's a non-const pointer.
-     @param [out] feature  Pointer to a vector of size this->Dim(), to which
-         the computed feature will be written. It should be pre-allocated.
-  */
-  void Compute(float signal_raw_log_energy, float vtln_warp,
-               std::vector<float> *signal_frame, float *feature);
-
- private:
-  const MelBanks *GetMelBanks(float vtln_warp);
-
-  MfccOptions opts_;
-  float log_energy_floor_;
-  std::map<float, MelBanks *> mel_banks_;  // float is VTLN coefficient.
-  Rfft rfft_;
-
-  // temp buffer of size num_mel_bins = opts.mel_opts.num_bins
-  std::vector<float> mel_energies_;
-
-  // opts_.num_ceps
-  std::vector<float> lifter_coeffs_;
-
-  // [num_ceps][num_mel_bins]
-  std::vector<float> dct_matrix_;
-};
-
-}  // namespace knf
-
-#endif  // KALDI_NATIVE_FBANK_CSRC_FEATURE_MFCC_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.cc
deleted file mode 100644
index f1ab58ee..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.cc
+++ /dev/null
@@ -1,38 +0,0 @@
-/**
- * Copyright (c)  2025  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include "kaldi-native-fbank/csrc/feature-raw-audio-samples.h"
-
-#include <algorithm>
-#include <vector>
-
-namespace knf {
-
-std::ostream &operator<<(std::ostream &os, const RawAudioSamplesOptions &opts) {
-  os << opts.ToString();
-  return os;
-}
-
-void RawAudioSamplesComputer::Compute(float /*signal_raw_log_energy*/,
-                                      float /*vtln_warp*/,
-                                      std::vector<float> *signal_frame,
-                                      float *feature) {
-  std::copy(signal_frame->begin(), signal_frame->end(), feature);
-}
-
-}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.h
deleted file mode 100644
index 5cf5e45f..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-raw-audio-samples.h
+++ /dev/null
@@ -1,81 +0,0 @@
-/**
- * Copyright (c)  2025  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef KALDI_NATIVE_FBANK_CSRC_FEATURE_RAW_AUDIO_SAMPLES_H_
-#define KALDI_NATIVE_FBANK_CSRC_FEATURE_RAW_AUDIO_SAMPLES_H_
-
-#include <cstdint>
-#include <sstream>
-#include <vector>
-
-#include "kaldi-native-fbank/csrc/feature-window.h"
-
-namespace knf {
-
-struct RawAudioSamplesOptions {
-  FrameExtractionOptions frame_opts;
-
-  RawAudioSamplesOptions() {
-    frame_opts.window_type = "rectangular";
-    frame_opts.dither = 0;
-    frame_opts.preemph_coeff = 0;
-    frame_opts.remove_dc_offset = false;
-    frame_opts.round_to_power_of_two = false;
-    frame_opts.snip_edges = true;
-  }
-
-  std::string ToString() const {
-    std::ostringstream os;
-    os << "frame_opts: \n";
-    os << frame_opts << "\n";
-    os << "\n";
-    return os.str();
-  }
-};
-
-std::ostream &operator<<(std::ostream &os, const RawAudioSamplesOptions &opts);
-
-class RawAudioSamplesComputer {
- public:
-  using Options = RawAudioSamplesOptions;
-
-  explicit RawAudioSamplesComputer(const RawAudioSamplesOptions &opts)
-      : opts_(opts){};
-
-  int32_t Dim() const {
-    return opts_.frame_opts.frame_length_ms * opts_.frame_opts.samp_freq / 1000;
-  }
-
-  bool NeedRawLogEnergy() const { return false; }
-
-  const FrameExtractionOptions &GetFrameOptions() const {
-    return opts_.frame_opts;
-  }
-
-  const RawAudioSamplesOptions &GetOptions() const { return opts_; }
-
-  void Compute(float unused_signal_raw_log_energy, float unused_vtln_warp,
-               std::vector<float> *signal_frame, float *feature);
-
- private:
-  RawAudioSamplesOptions opts_;
-};
-
-}  // namespace knf
-
-#endif  // KALDI_NATIVE_FBANK_CSRC_FEATURE_RAW_AUDIO_SAMPLES_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.cc
deleted file mode 100644
index 047abaf2..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.cc
+++ /dev/null
@@ -1,274 +0,0 @@
-// kaldi-native-fbank/csrc/feature-window.cc
-//
-// Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
-
-// This file is copied/modified from kaldi/src/feat/feature-window.cc
-
-#include "kaldi-native-fbank/csrc/feature-window.h"
-
-#include <algorithm>
-#include <cmath>
-#include <cstddef>
-#include <limits>
-#include <vector>
-
-#include "kaldi-native-fbank/csrc/kaldi-math.h"
-
-namespace knf {
-
-std::ostream &operator<<(std::ostream &os, const FrameExtractionOptions &opts) {
-  os << opts.ToString();
-  return os;
-}
-
-std::vector<float> GetWindow(const std::string &window_type,
-                             int32_t window_size,
-                             float blackman_coeff /*= 0.42*/) {
-  std::vector<float> window(window_size);
-  int32_t frame_length = window_size;
-  KNF_CHECK_GT(frame_length, 0);
-
-  float *window_data = window.data();
-
-  double a = M_2PI / (frame_length - 1);
-  if (window_type == "hann") {
-    // see https://pytorch.org/docs/stable/generated/torch.hann_window.html
-    // We assume periodic is true
-    a = M_2PI / frame_length;
-  }
-
-  for (int32_t i = 0; i < frame_length; i++) {
-    double i_fl = static_cast<double>(i);
-    if (window_type == "hanning") {
-      window_data[i] = 0.5 - 0.5 * cos(a * i_fl);
-    } else if (window_type == "sine") {
-      // when you are checking ws wikipedia, please
-      // note that 0.5 * a = M_PI/(frame_length-1)
-      window_data[i] = sin(0.5 * a * i_fl);
-    } else if (window_type == "hamming") {
-      window_data[i] = 0.54 - 0.46 * cos(a * i_fl);
-    } else if (window_type == "hann") {
-      window_data[i] = 0.50 - 0.50 * cos(a * i_fl);
-    } else if (window_type == "povey") {
-      // like hamming but goes to zero at edges.
-      window_data[i] = pow(0.5 - 0.5 * cos(a * i_fl), 0.85);
-    } else if (window_type == "rectangular") {
-      window_data[i] = 1.0;
-    } else if (window_type == "blackman") {
-      window_data[i] = blackman_coeff - 0.5 * cos(a * i_fl) +
-                       (0.5 - blackman_coeff) * cos(2 * a * i_fl);
-    } else {
-      fprintf(stderr, "Invalid window type '%s'\n", window_type.c_str());
-      exit(-1);
-    }
-  }
-
-  return window;
-}
-
-FeatureWindowFunction::FeatureWindowFunction(const FrameExtractionOptions &opts)
-    : FeatureWindowFunction(opts.window_type, opts.WindowSize(),
-                            opts.blackman_coeff) {}
-
-FeatureWindowFunction::FeatureWindowFunction(const std::string &window_type,
-                                             int32_t window_size,
-                                             float blackman_coeff /*= 0.42*/)
-    : window_(knf::GetWindow(window_type, window_size, blackman_coeff)) {}
-
-FeatureWindowFunction::FeatureWindowFunction(const std::vector<float> &window)
-    : window_(window) {}
-
-void FeatureWindowFunction::Apply(float *wave) const {
-  int32_t window_size = window_.size();
-  const float *p = window_.data();
-  for (int32_t k = 0; k != window_size; ++k) {
-    wave[k] *= p[k];
-  }
-}
-
-int64_t FirstSampleOfFrame(int32_t frame, const FrameExtractionOptions &opts) {
-  int64_t frame_shift = opts.WindowShift();
-  if (opts.snip_edges) {
-    return frame * frame_shift;
-  } else {
-    int64_t midpoint_of_frame = frame_shift * frame + frame_shift / 2,
-            beginning_of_frame = midpoint_of_frame - opts.WindowSize() / 2;
-    return beginning_of_frame;
-  }
-}
-
-int32_t NumFrames(int64_t num_samples, const FrameExtractionOptions &opts,
-                  bool flush /*= true*/) {
-  int64_t frame_shift = opts.WindowShift();
-  int64_t frame_length = opts.WindowSize();
-  if (opts.snip_edges) {
-    // with --snip-edges=true (the default), we use a HTK-like approach to
-    // determining the number of frames-- all frames have to fit completely into
-    // the waveform, and the first frame begins at sample zero.
-    if (num_samples < frame_length)
-      return 0;
-    else
-      return (1 + ((num_samples - frame_length) / frame_shift));
-    // You can understand the expression above as follows: 'num_samples -
-    // frame_length' is how much room we have to shift the frame within the
-    // waveform; 'frame_shift' is how much we shift it each time; and the ratio
-    // is how many times we can shift it (integer arithmetic rounds down).
-  } else {
-    // if --snip-edges=false, the number of frames is determined by rounding the
-    // (file-length / frame-shift) to the nearest integer.  The point of this
-    // formula is to make the number of frames an obvious and predictable
-    // function of the frame shift and signal length, which makes many
-    // segmentation-related questions simpler.
-    //
-    // Because integer division in C++ rounds toward zero, we add (half the
-    // frame-shift minus epsilon) before dividing, to have the effect of
-    // rounding towards the closest integer.
-    int32_t num_frames = (num_samples + (frame_shift / 2)) / frame_shift;
-
-    if (flush) return num_frames;
-
-    // note: 'end' always means the last plus one, i.e. one past the last.
-    int64_t end_sample_of_last_frame =
-        FirstSampleOfFrame(num_frames - 1, opts) + frame_length;
-
-    // the following code is optimized more for clarity than efficiency.
-    // If flush == false, we can't output frames that extend past the end
-    // of the signal.
-    while (num_frames > 0 && end_sample_of_last_frame > num_samples) {
-      num_frames--;
-      end_sample_of_last_frame -= frame_shift;
-    }
-    return num_frames;
-  }
-}
-
-void ExtractWindow(int64_t sample_offset, const std::vector<float> &wave,
-                   int32_t f, const FrameExtractionOptions &opts,
-                   const FeatureWindowFunction &window_function,
-                   std::vector<float> *window,
-                   float *log_energy_pre_window /*= nullptr*/) {
-  KNF_CHECK(sample_offset >= 0 && wave.size() != 0);
-
-  int32_t frame_length = opts.WindowSize();
-  int32_t frame_length_padded = opts.PaddedWindowSize();
-
-  int64_t num_samples = sample_offset + wave.size();
-  int64_t start_sample = FirstSampleOfFrame(f, opts);
-  int64_t end_sample = start_sample + frame_length;
-
-  if (opts.snip_edges) {
-    KNF_CHECK(start_sample >= sample_offset && end_sample <= num_samples);
-  } else {
-    KNF_CHECK(sample_offset == 0 || start_sample >= sample_offset);
-  }
-
-  if (window->size() != frame_length_padded) {
-    window->resize(frame_length_padded);
-  }
-
-  // wave_start and wave_end are start and end indexes into 'wave', for the
-  // piece of wave that we're trying to extract.
-  int32_t wave_start = int32_t(start_sample - sample_offset);
-  int32_t wave_end = wave_start + frame_length;
-
-  if (wave_start >= 0 && wave_end <= wave.size()) {
-    // the normal case-- no edge effects to consider.
-    std::copy(wave.begin() + wave_start,
-              wave.begin() + wave_start + frame_length, window->data());
-  } else {
-    // Deal with any end effects by reflection, if needed.  This code will only
-    // be reached for about two frames per utterance, so we don't concern
-    // ourselves excessively with efficiency.
-    int32_t wave_dim = wave.size();
-    for (int32_t s = 0; s < frame_length; ++s) {
-      int32_t s_in_wave = s + wave_start;
-      while (s_in_wave < 0 || s_in_wave >= wave_dim) {
-        // reflect around the beginning or end of the wave.
-        // e.g. -1 -> 0, -2 -> 1.
-        // dim -> dim - 1, dim + 1 -> dim - 2.
-        // the code supports repeated reflections, although this
-        // would only be needed in pathological cases.
-        if (s_in_wave < 0)
-          s_in_wave = -s_in_wave - 1;
-        else
-          s_in_wave = 2 * wave_dim - 1 - s_in_wave;
-      }
-      (*window)[s] = wave[s_in_wave];
-    }
-  }
-
-  ProcessWindow(opts, window_function, window->data(), log_energy_pre_window);
-}
-
-static void RemoveDcOffset(float *d, int32_t n) {
-  float sum = 0;
-  for (int32_t i = 0; i != n; ++i) {
-    sum += d[i];
-  }
-
-  float mean = sum / n;
-
-  for (int32_t i = 0; i != n; ++i) {
-    d[i] -= mean;
-  }
-}
-
-float InnerProduct(const float *a, const float *b, int32_t n) {
-  float sum = 0;
-  for (int32_t i = 0; i != n; ++i) {
-    sum += a[i] * b[i];
-  }
-  return sum;
-}
-
-void Dither(float *d, int32_t n, float dither_value) {
-  if (dither_value == 0.0) {
-    return;
-  }
-
-  RandomState rstate;
-  for (int32_t i = 0; i < n; ++i) {
-    d[i] += RandGauss(&rstate) * dither_value;
-  }
-}
-
-static void Preemphasize(float *d, int32_t n, float preemph_coeff) {
-  if (preemph_coeff == 0.0) {
-    return;
-  }
-
-  KNF_CHECK(preemph_coeff >= 0.0 && preemph_coeff <= 1.0);
-
-  for (int32_t i = n - 1; i > 0; --i) {
-    d[i] -= preemph_coeff * d[i - 1];
-  }
-  d[0] -= preemph_coeff * d[0];
-}
-
-void ProcessWindow(const FrameExtractionOptions &opts,
-                   const FeatureWindowFunction &window_function, float *window,
-                   float *log_energy_pre_window /*= nullptr*/) {
-  int32_t frame_length = opts.WindowSize();
-
-  if (opts.dither != 0.0) {
-    Dither(window, frame_length, opts.dither);
-  }
-
-  if (opts.remove_dc_offset) {
-    RemoveDcOffset(window, frame_length);
-  }
-
-  if (log_energy_pre_window != NULL) {
-    float energy = std::max<float>(InnerProduct(window, window, frame_length),
-                                   std::numeric_limits<float>::epsilon());
-    *log_energy_pre_window = std::log(energy);
-  }
-
-  if (opts.preemph_coeff != 0.0) {
-    Preemphasize(window, frame_length, opts.preemph_coeff);
-  }
-
-  window_function.Apply(window);
-}
-
-}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.h
deleted file mode 100644
index b8d1d44c..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/feature-window.h
+++ /dev/null
@@ -1,187 +0,0 @@
-// kaldi-native-fbank/csrc/feature-window.h
-//
-// Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
-
-// This file is copied/modified from kaldi/src/feat/feature-window.h
-
-#ifndef KALDI_NATIVE_FBANK_CSRC_FEATURE_WINDOW_H_
-#define KALDI_NATIVE_FBANK_CSRC_FEATURE_WINDOW_H_
-
-#include <cstdint>
-#include <sstream>
-#include <string>
-#include <vector>
-
-#include "kaldi-native-fbank/csrc/log.h"
-
-namespace knf {
-
-inline int32_t RoundUpToNearestPowerOfTwo(int32_t n) {
-  // copied from kaldi/src/base/kaldi-math.cc
-  KNF_CHECK_GT(n, 0);
-  n--;
-  n |= n >> 1;
-  n |= n >> 2;
-  n |= n >> 4;
-  n |= n >> 8;
-  n |= n >> 16;
-  return n + 1;
-}
-
-struct FrameExtractionOptions {
-  float samp_freq = 16000;
-  float frame_shift_ms = 10.0f;   // in milliseconds.
-  float frame_length_ms = 25.0f;  // in milliseconds.
-
-  float dither = 0.00003f;  // Amount of dithering, 0.0 means no dither.
-                            // Value 0.00003f is equivalent to 1.0 in kaldi.
-
-  float preemph_coeff = 0.97f;        // Preemphasis coefficient.
-  bool remove_dc_offset = true;       // Subtract mean of wave before FFT.
-  std::string window_type = "povey";  // e.g. Hamming window
-  // May be "hamming", "rectangular", "povey", "hanning", "hann", "sine",
-  // "blackman".
-  // "povey" is a window I made to be similar to Hamming but to go to zero at
-  // the edges, it's pow((0.5 - 0.5*cos(n/N*2*pi)), 0.85) I just don't think the
-  // Hamming window makes sense as a windowing function.
-  bool round_to_power_of_two = true;
-  float blackman_coeff = 0.42f;
-  bool snip_edges = true;
-  // bool allow_downsample = false;
-  // bool allow_upsample = false;
-
-  int32_t WindowShift() const {
-    return static_cast<int32_t>(samp_freq * 0.001f * frame_shift_ms);
-  }
-  int32_t WindowSize() const {
-    return static_cast<int32_t>(samp_freq * 0.001f * frame_length_ms);
-  }
-  int32_t PaddedWindowSize() const {
-    return (round_to_power_of_two ? RoundUpToNearestPowerOfTwo(WindowSize())
-                                  : WindowSize());
-  }
-  std::string ToString() const {
-    std::ostringstream os;
-#define KNF_PRINT(x) os << #x << ": " << x << "\n"
-    KNF_PRINT(samp_freq);
-    KNF_PRINT(frame_shift_ms);
-    KNF_PRINT(frame_length_ms);
-    KNF_PRINT(dither);
-    KNF_PRINT(preemph_coeff);
-    KNF_PRINT(remove_dc_offset);
-    KNF_PRINT(window_type);
-    KNF_PRINT(round_to_power_of_two);
-    KNF_PRINT(blackman_coeff);
-    KNF_PRINT(snip_edges);
-    // KNF_PRINT(allow_downsample);
-    // KNF_PRINT(allow_upsample);
-#undef KNF_PRINT
-    return os.str();
-  }
-};
-
-std::ostream &operator<<(std::ostream &os, const FrameExtractionOptions &opts);
-
-class FeatureWindowFunction {
- public:
-  FeatureWindowFunction() = default;
-  explicit FeatureWindowFunction(const FrameExtractionOptions &opts);
-  FeatureWindowFunction(const std::string &window_type, int32_t window_size,
-                        float blackman_coeff = 0.42);
-
-  explicit FeatureWindowFunction(const std::vector<float> &window);
-
-  /**
-   * @param wave Pointer to a 1-D array of shape [window_size].
-   *             It is modified in-place: wave[i] = wave[i] * window_[i].
-   * @param
-   */
-  void Apply(float *wave) const;
-
-  const std::vector<float> &GetWindow() const { return window_; }
-
- private:
-  std::vector<float> window_;  // of size opts.WindowSize()
-};
-
-int64_t FirstSampleOfFrame(int32_t frame, const FrameExtractionOptions &opts);
-
-/**
-   This function returns the number of frames that we can extract from a wave
-   file with the given number of samples in it (assumed to have the same
-   sampling rate as specified in 'opts').
-
-      @param [in] num_samples  The number of samples in the wave file.
-      @param [in] opts     The frame-extraction options class
-
-      @param [in] flush   True if we are asserting that this number of samples
-   is 'all there is', false if we expecting more data to possibly come in.  This
-   only makes a difference to the answer
-   if opts.snip_edges== false.  For offline feature extraction you always want
-   flush == true.  In an online-decoding context, once you know (or decide) that
-   no more data is coming in, you'd call it with flush == true at the end to
-   flush out any remaining data.
-*/
-int32_t NumFrames(int64_t num_samples, const FrameExtractionOptions &opts,
-                  bool flush = true);
-
-/*
-  ExtractWindow() extracts a windowed frame of waveform (possibly with a
-  power-of-two, padded size, depending on the config), including all the
-  processing done by ProcessWindow().
-
-  @param [in] sample_offset  If 'wave' is not the entire waveform, but
-                   part of it to the left has been discarded, then the
-                   number of samples prior to 'wave' that we have
-                   already discarded.  Set this to zero if you are
-                   processing the entire waveform in one piece, or
-                   if you get 'no matching function' compilation
-                   errors when updating the code.
-  @param [in] wave  The waveform
-  @param [in] f     The frame index to be extracted, with
-                    0 <= f < NumFrames(sample_offset + wave.Dim(), opts, true)
-  @param [in] opts  The options class to be used
-  @param [in] window_function  The windowing function, as derived from the
-                    options class.
-  @param [out] window  The windowed, possibly-padded waveform to be
-                     extracted.  Will be resized as needed.
-  @param [out] log_energy_pre_window  If non-NULL, the log-energy of
-                   the signal prior to pre-emphasis and multiplying by
-                   the windowing function will be written to here.
-*/
-void ExtractWindow(int64_t sample_offset, const std::vector<float> &wave,
-                   int32_t f, const FrameExtractionOptions &opts,
-                   const FeatureWindowFunction &window_function,
-                   std::vector<float> *window,
-                   float *log_energy_pre_window = nullptr);
-
-/**
-  This function does all the windowing steps after actually
-  extracting the windowed signal: depending on the
-  configuration, it does dithering, dc offset removal,
-  preemphasis, and multiplication by the windowing function.
-   @param [in] opts  The options class to be used
-   @param [in] window_function  The windowing function-- should have
-                    been initialized using 'opts'.
-   @param [in,out] window  A vector of size opts.WindowSize().  Note:
-      it will typically be a sub-vector of a larger vector of size
-      opts.PaddedWindowSize(), with the remaining samples zero,
-      as the FFT code is more efficient if it operates on data with
-      power-of-two size.
-   @param [out]   log_energy_pre_window If non-NULL, then after dithering and
-      DC offset removal, this function will write to this pointer the log of
-      the total energy (i.e. sum-squared) of the frame.
- */
-void ProcessWindow(const FrameExtractionOptions &opts,
-                   const FeatureWindowFunction &window_function, float *window,
-                   float *log_energy_pre_window = nullptr);
-
-// Compute the inner product of two vectors
-float InnerProduct(const float *a, const float *b, int32_t n);
-
-std::vector<float> GetWindow(const std::string &window_type,
-                             int32_t window_size, float blackman_coeff = 0.42);
-
-}  // namespace knf
-
-#endif  // KALDI_NATIVE_FBANK_CSRC_FEATURE_WINDOW_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.cc
deleted file mode 100644
index e779ac74..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.cc
+++ /dev/null
@@ -1,158 +0,0 @@
-/**
- * Copyright (c)  2025  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-#include "kaldi-native-fbank/csrc/istft.h"
-
-#include <algorithm>
-#include <cmath>
-#include <sstream>
-#include <string>
-#include <utility>
-
-#include "kaldi-native-fbank/csrc/feature-window.h"
-#include "kaldi-native-fbank/csrc/rfft.h"
-
-namespace knf {
-
-class IStft::Impl {
- public:
-  explicit Impl(const StftConfig &config) : config_(config) {
-    if (!config.window.empty()) {
-      window_ = std::make_unique<FeatureWindowFunction>(config_.window);
-    } else if (!config.window_type.empty()) {
-      window_ = std::make_unique<FeatureWindowFunction>(config.window_type,
-                                                        config.win_length);
-    }
-  }
-
-  std::vector<float> Compute(const StftResult &stft_result) const {
-    Rfft rfft(config_.n_fft, true);
-
-    int32_t num_samples =
-        config_.n_fft + (stft_result.num_frames - 1) * config_.hop_length;
-
-    std::vector<float> samples(num_samples);
-    for (int32_t i = 0; i < stft_result.num_frames; ++i) {
-      auto x = InverseFFT(stft_result, i, &rfft);
-      OverlapAdd(std::move(x), i, &samples);
-    }
-
-    auto denominator = GetDenominator(stft_result.num_frames);
-
-    for (int32_t i = 0; i < num_samples; ++i) {
-      if (denominator[i]) {
-        samples[i] /= denominator[i];
-      }
-    }
-
-    if (config_.center) {
-      samples = {samples.begin() + config_.n_fft / 2,
-                 samples.end() - config_.n_fft / 2};
-    }
-
-    return samples;
-  }
-
-  std::vector<float> InverseFFT(const StftResult &r, int32_t frame_index,
-                                Rfft *rfft) const {
-    int32_t n_fft = config_.n_fft;
-    int32_t hop_length = config_.hop_length;
-
-    const float *p_real = r.real.data() + frame_index * (n_fft / 2 + 1);
-    const float *p_imag = r.imag.data() + frame_index * (n_fft / 2 + 1);
-    std::vector<float> tmp(n_fft);
-
-    float scale = 1;
-    if (config_.normalized) {
-      scale = std::sqrt(n_fft);
-    }
-
-    for (int32_t i = 0; i < n_fft / 2; ++i) {
-      if (i == 0) {
-        tmp[0] = p_real[0] * scale;
-        tmp[1] = p_real[n_fft / 2] * scale;
-      } else {
-        tmp[2 * i] = p_real[i] * scale;
-        tmp[2 * i + 1] = p_imag[i] * scale;
-      }
-    }
-
-    rfft->Compute(tmp.data());
-
-    scale = 1.0f / n_fft;
-    for (auto &f : tmp) {
-      f *= scale;
-    }
-
-    return tmp;
-  }
-
-  void OverlapAdd(std::vector<float> current_frame, int32_t frame_index,
-                  std::vector<float> *samples) const {
-    if (window_) {
-      window_->Apply(current_frame.data());
-    }
-
-    float *p = samples->data() + frame_index * config_.hop_length;
-    for (int32_t i = 0; i < config_.n_fft; ++i) {
-      p[i] += current_frame[i];
-    }
-  }
-
-  std::vector<float> GetDenominator(int32_t num_frames) const {
-    int32_t num_samples = config_.n_fft + (num_frames - 1) * config_.hop_length;
-    std::vector<float> ans(num_samples);
-    if (!window_) {
-      for (int32_t i = 0; i < num_frames; ++i) {
-        int32_t start = i * config_.hop_length;
-        int32_t end = start + config_.n_fft;
-
-        for (int32_t k = start; k < end; ++k) {
-          ans[k] += 1;
-        }
-      }
-    } else {
-      const auto &w = window_->GetWindow();
-      for (int32_t i = 0; i < num_frames; ++i) {
-        int32_t start = i * config_.hop_length;
-        int32_t end = start + config_.n_fft;
-
-        auto pw = w.data();
-        for (int32_t k = start; k < end; ++k, ++pw) {
-          ans[k] += (*pw) * (*pw);
-        }
-      }
-    }
-
-    return ans;
-  }
-
- private:
-  StftConfig config_;
-  std::unique_ptr<FeatureWindowFunction> window_;
-};
-
-IStft::IStft(const StftConfig &config)
-    : impl_(std::make_unique<Impl>(config)) {}
-
-IStft::~IStft() = default;
-
-std::vector<float> IStft::Compute(const StftResult &stft_result) const {
-  return impl_->Compute(stft_result);
-}
-
-}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.h
deleted file mode 100644
index e5eba1b9..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/istft.h
+++ /dev/null
@@ -1,45 +0,0 @@
-/**
- * Copyright (c)  2025  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef KALDI_NATIVE_FBANK_CSRC_ISTFT_H_
-#define KALDI_NATIVE_FBANK_CSRC_ISTFT_H_
-
-#include <cstdint>
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "kaldi-native-fbank/csrc/stft.h"
-
-namespace knf {
-
-class IStft {
- public:
-  explicit IStft(const StftConfig &config);
-  ~IStft();
-  std::vector<float> Compute(const StftResult &stft_result) const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace knf
-
-#endif  // KALDI_NATIVE_FBANK_CSRC_ISTFT_H_
-        //
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.cc
deleted file mode 100644
index bcc0dcc3..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.cc
+++ /dev/null
@@ -1,51 +0,0 @@
-// kaldi-native-fbank/csrc/kaldi-math.cc
-//
-// Copyright (c)  2024  Brno University of Technology (authors: Karel Vesely)
-
-// This file is an excerpt from kaldi/src/base/kaldi-math.cc
-
-#include "kaldi-native-fbank/csrc/kaldi-math.h"
-
-#if defined(_POSIX_THREAD_SAFE_FUNCTIONS)
-#include <mutex>  // NOLINT
-#endif
-
-#include <cmath>
-
-namespace knf {
-
-int Rand(struct RandomState *state) {
-#if !defined(_POSIX_THREAD_SAFE_FUNCTIONS)
-  // On Windows and Cygwin, just call Rand()
-  return rand();
-#else
-  if (state) {
-    return rand_r(&(state->seed));
-  } else {
-    static std::mutex _RandMutex;
-    std::lock_guard<std::mutex> lock(_RandMutex);
-    return rand();
-  }
-#endif
-}
-
-RandomState::RandomState() {
-  // we initialize it as Rand() + 27437 instead of just Rand(), because on some
-  // systems, e.g. at the very least Mac OSX Yosemite and later, it seems to be
-  // the case that rand_r when initialized with rand() will give you the exact
-  // same sequence of numbers that rand() will give if you keep calling rand()
-  // after that initial call.  This can cause problems with repeated sequences.
-  // For example if you initialize two RandomState structs one after the other
-  // without calling rand() in between, they would give you the same sequence
-  // offset by one (if we didn't have the "+ 27437" in the code).  27437 is just
-  // a randomly chosen prime number.
-  seed = unsigned(Rand()) + 27437;
-}
-
-void Sqrt(float *in_out, int32_t n) {
-  for (int32_t i = 0; i != n; ++i) {
-    in_out[i] = std::sqrt(in_out[i]);
-  }
-}
-
-}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.h
deleted file mode 100644
index 4db3f831..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/kaldi-math.h
+++ /dev/null
@@ -1,52 +0,0 @@
-// kaldi-native-fbank/csrc/kaldi-math.h
-//
-// Copyright (c)  2024  Brno University of Technology (authors: Karel Vesely)
-
-// This file is an excerpt from kaldi/src/base/kaldi-math.h
-
-#ifndef KALDI_NATIVE_FBANK_CSRC_KALDI_MATH_H_
-#define KALDI_NATIVE_FBANK_CSRC_KALDI_MATH_H_
-
-#include <cmath>  // logf, sqrtf, cosf
-#include <cstdint>
-#include <cstdlib>  // RAND_MAX
-
-#ifndef M_PI
-#define M_PI 3.1415926535897932384626433832795
-#endif
-
-#ifndef M_2PI
-#define M_2PI 6.283185307179586476925286766559005
-#endif
-
-#ifndef M_SQRT2
-#define M_SQRT2 1.4142135623730950488016887
-#endif
-
-namespace knf {
-
-inline float Log(float x) { return logf(x); }
-
-// Returns a random integer between 0 and RAND_MAX, inclusive
-int Rand(struct RandomState *state = NULL);
-
-// State for thread-safe random number generator
-struct RandomState {
-  RandomState();
-  unsigned seed;
-};
-
-/// Returns a random number strictly between 0 and 1.
-inline float RandUniform(struct RandomState *state = NULL) {
-  return static_cast<float>((Rand(state) + 1.0) / (RAND_MAX + 2.0));
-}
-
-inline float RandGauss(struct RandomState *state = NULL) {
-  return static_cast<float>(sqrtf(-2 * Log(RandUniform(state))) *
-                            cosf(2 * M_PI * RandUniform(state)));
-}
-
-void Sqrt(float *in_out, int32_t n);
-
-}  // namespace knf
-#endif  // KALDI_NATIVE_FBANK_CSRC_KALDI_MATH_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.cc
deleted file mode 100644
index 7223337b..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.cc
+++ /dev/null
@@ -1,142 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-/*
- * Stack trace related stuff is from kaldi.
- * Refer to
- * https://github.com/kaldi-asr/kaldi/blob/master/src/base/kaldi-error.cc
- */
-
-#include "kaldi-native-fbank/csrc/log.h"
-
-#ifdef KNF_HAVE_EXECINFO_H
-#include <execinfo.h>  // To get stack trace in error messages.
-#ifdef KNF_HAVE_CXXABI_H
-#include <cxxabi.h>  // For name demangling.
-// Useful to decode the stack trace, but only used if we have execinfo.h
-#endif  // KNF_HAVE_CXXABI_H
-#endif  // KNF_HAVE_EXECINFO_H
-
-#include <stdlib.h>
-
-#include <ctime>
-#include <iomanip>
-#include <string>
-
-namespace knf {
-
-std::string GetDateTimeStr() {
-  std::ostringstream os;
-  std::time_t t = std::time(nullptr);
-  std::tm tm = *std::localtime(&t);
-  os << std::put_time(&tm, "%F %T");  // yyyy-mm-dd hh:mm:ss
-  return os.str();
-}
-
-static bool LocateSymbolRange(const std::string &trace_name, std::size_t *begin,
-                              std::size_t *end) {
-  // Find the first '_' with leading ' ' or '('.
-  *begin = std::string::npos;
-  for (std::size_t i = 1; i < trace_name.size(); ++i) {
-    if (trace_name[i] != '_') {
-      continue;
-    }
-    if (trace_name[i - 1] == ' ' || trace_name[i - 1] == '(') {
-      *begin = i;
-      break;
-    }
-  }
-  if (*begin == std::string::npos) {
-    return false;
-  }
-  *end = trace_name.find_first_of(" +", *begin);
-  return *end != std::string::npos;
-}
-
-#ifdef KNF_HAVE_EXECINFO_H
-static std::string Demangle(const std::string &trace_name) {
-#ifndef KNF_HAVE_CXXABI_H
-  return trace_name;
-#else   // KNF_HAVE_CXXABI_H
-  // Try demangle the symbol. We are trying to support the following formats
-  // produced by different platforms:
-  //
-  // Linux:
-  //   ./kaldi-error-test(_ZN5kaldi13UnitTestErrorEv+0xb) [0x804965d]
-  //
-  // Mac:
-  //   0 server 0x000000010f67614d _ZNK5kaldi13MessageLogger10LogMessageEv + 813
-  //
-  // We want to extract the name e.g., '_ZN5kaldi13UnitTestErrorEv' and
-  // demangle it info a readable name like kaldi::UnitTextError.
-  std::size_t begin, end;
-  if (!LocateSymbolRange(trace_name, &begin, &end)) {
-    return trace_name;
-  }
-  std::string symbol = trace_name.substr(begin, end - begin);
-  int status;
-  char *demangled_name = abi::__cxa_demangle(symbol.c_str(), 0, 0, &status);
-  if (status == 0 && demangled_name != nullptr) {
-    symbol = demangled_name;
-    free(demangled_name);
-  }
-  return trace_name.substr(0, begin) + symbol +
-         trace_name.substr(end, std::string::npos);
-#endif  // KNF_HAVE_CXXABI_H
-}
-#endif  // KNF_HAVE_EXECINFO_H
-
-std::string GetStackTrace() {
-  std::string ans;
-#ifdef KNF_HAVE_EXECINFO_H
-  constexpr const std::size_t kMaxTraceSize = 50;
-  constexpr const std::size_t kMaxTracePrint = 50;  // Must be even.
-                                                    // Buffer for the trace.
-  void *trace[kMaxTraceSize];
-  // Get the trace.
-  std::size_t size = backtrace(trace, kMaxTraceSize);
-  // Get the trace symbols.
-  char **trace_symbol = backtrace_symbols(trace, size);
-  if (trace_symbol == nullptr) return ans;
-
-  // Compose a human-readable backtrace string.
-  ans += "[ Stack-Trace: ]\n";
-  if (size <= kMaxTracePrint) {
-    for (std::size_t i = 0; i < size; ++i) {
-      ans += Demangle(trace_symbol[i]) + "\n";
-    }
-  } else {  // Print out first+last (e.g.) 5.
-    for (std::size_t i = 0; i < kMaxTracePrint / 2; ++i) {
-      ans += Demangle(trace_symbol[i]) + "\n";
-    }
-    ans += ".\n.\n.\n";
-    for (std::size_t i = size - kMaxTracePrint / 2; i < size; ++i) {
-      ans += Demangle(trace_symbol[i]) + "\n";
-    }
-    if (size == kMaxTraceSize)
-      ans += ".\n.\n.\n";  // Stack was too long, probably a bug.
-  }
-
-  // We must free the array of pointers allocated by backtrace_symbols(),
-  // but not the strings themselves.
-  free(trace_symbol);
-#endif  // KNF_HAVE_EXECINFO_H
-  return ans;
-}
-
-}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.h
deleted file mode 100644
index bd21cc35..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/log.h
+++ /dev/null
@@ -1,383 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-// The content in this file is copied/modified from
-// https://github.com/k2-fsa/k2/blob/master/k2/csrc/log.h
-#ifndef KALDI_NATIVE_FBANK_CSRC_LOG_H_
-#define KALDI_NATIVE_FBANK_CSRC_LOG_H_
-
-#include <stdio.h>
-
-#include <mutex>  // NOLINT
-#include <sstream>
-#include <string>
-
-namespace knf {
-
-#if KNF_ENABLE_CHECK
-
-#if defined(NDEBUG)
-constexpr bool kDisableDebug = true;
-#else
-constexpr bool kDisableDebug = false;
-#endif
-
-enum class LogLevel {
-  kTrace = 0,
-  kDebug = 1,
-  kInfo = 2,
-  kWarning = 3,
-  kError = 4,
-  kFatal = 5,  // print message and abort the program
-};
-
-// They are used in KNF_LOG(xxx), so their names
-// do not follow the google c++ code style
-//
-// You can use them in the following way:
-//
-//  KNF_LOG(TRACE) << "some message";
-//  KNF_LOG(DEBUG) << "some message";
-#ifndef _MSC_VER
-constexpr LogLevel TRACE = LogLevel::kTrace;
-constexpr LogLevel DEBUG = LogLevel::kDebug;
-constexpr LogLevel INFO = LogLevel::kInfo;
-constexpr LogLevel WARNING = LogLevel::kWarning;
-constexpr LogLevel ERROR = LogLevel::kError;
-constexpr LogLevel FATAL = LogLevel::kFatal;
-#else
-#define TRACE LogLevel::kTrace
-#define DEBUG LogLevel::kDebug
-#define INFO LogLevel::kInfo
-#define WARNING LogLevel::kWarning
-#define ERROR LogLevel::kError
-#define FATAL LogLevel::kFatal
-#endif
-
-std::string GetStackTrace();
-
-/* Return the current log level.
-
-
-   If the current log level is TRACE, then all logged messages are printed out.
-
-   If the current log level is DEBUG, log messages with "TRACE" level are not
-   shown and all other levels are printed out.
-
-   Similarly, if the current log level is INFO, log message with "TRACE" and
-   "DEBUG" are not shown and all other levels are printed out.
-
-   If it is FATAL, then only FATAL messages are shown.
- */
-inline LogLevel GetCurrentLogLevel() {
-  static LogLevel log_level = INFO;
-  static std::once_flag init_flag;
-  std::call_once(init_flag, []() {
-    const char *env_log_level = std::getenv("KNF_LOG_LEVEL");
-    if (env_log_level == nullptr) return;
-
-    std::string s = env_log_level;
-    if (s == "TRACE")
-      log_level = TRACE;
-    else if (s == "DEBUG")
-      log_level = DEBUG;
-    else if (s == "INFO")
-      log_level = INFO;
-    else if (s == "WARNING")
-      log_level = WARNING;
-    else if (s == "ERROR")
-      log_level = ERROR;
-    else if (s == "FATAL")
-      log_level = FATAL;
-    else
-      fprintf(stderr,
-              "Unknown KNF_LOG_LEVEL: %s"
-              "\nSupported values are: "
-              "TRACE, DEBUG, INFO, WARNING, ERROR, FATAL",
-              s.c_str());
-  });
-  return log_level;
-}
-
-inline bool EnableAbort() {
-  static std::once_flag init_flag;
-  static bool enable_abort = false;
-  std::call_once(init_flag, []() {
-    enable_abort = (std::getenv("KNF_ABORT") != nullptr);
-  });
-  return enable_abort;
-}
-
-class Logger {
- public:
-  Logger(const char *filename, const char *func_name, uint32_t line_num,
-         LogLevel level)
-      : filename_(filename),
-        func_name_(func_name),
-        line_num_(line_num),
-        level_(level) {
-    cur_level_ = GetCurrentLogLevel();
-    fprintf(stderr, "here\n");
-    switch (level) {
-      case TRACE:
-        if (cur_level_ <= TRACE) fprintf(stderr, "[T] ");
-        break;
-      case DEBUG:
-        if (cur_level_ <= DEBUG) fprintf(stderr, "[D] ");
-        break;
-      case INFO:
-        if (cur_level_ <= INFO) fprintf(stderr, "[I] ");
-        break;
-      case WARNING:
-        if (cur_level_ <= WARNING) fprintf(stderr, "[W] ");
-        break;
-      case ERROR:
-        if (cur_level_ <= ERROR) fprintf(stderr, "[E] ");
-        break;
-      case FATAL:
-        if (cur_level_ <= FATAL) fprintf(stderr, "[F] ");
-        break;
-    }
-
-    if (cur_level_ <= level_) {
-      fprintf(stderr, "%s:%u:%s ", filename, line_num, func_name);
-    }
-  }
-
-  ~Logger() noexcept(false) {
-    static constexpr const char *kErrMsg = R"(
-    Some bad things happened. Please read the above error messages and stack
-    trace. If you are using Python, the following command may be helpful:
-
-      gdb --args python /path/to/your/code.py
-
-    (You can use `gdb` to debug the code. Please consider compiling
-    a debug version of KNF.).
-
-    If you are unable to fix it, please open an issue at:
-
-      https://github.com/csukuangfj/kaldi-native-fbank/issues/new
-    )";
-    fprintf(stderr, "\n");
-    if (level_ == FATAL) {
-      std::string stack_trace = GetStackTrace();
-      if (!stack_trace.empty()) {
-        fprintf(stderr, "\n\n%s\n", stack_trace.c_str());
-      }
-
-      fflush(nullptr);
-
-#ifndef __ANDROID_API__
-      if (EnableAbort()) {
-        // NOTE: abort() will terminate the program immediately without
-        // printing the Python stack backtrace.
-        abort();
-      }
-
-      throw std::runtime_error(kErrMsg);
-#else
-      abort();
-#endif
-    }
-  }
-
-  const Logger &operator<<(bool b) const {
-    if (cur_level_ <= level_) {
-      fprintf(stderr, b ? "true" : "false");
-    }
-    return *this;
-  }
-
-  const Logger &operator<<(int8_t i) const {
-    if (cur_level_ <= level_) fprintf(stderr, "%d", i);
-    return *this;
-  }
-
-  const Logger &operator<<(const char *s) const {
-    if (cur_level_ <= level_) fprintf(stderr, "%s", s);
-    return *this;
-  }
-
-  const Logger &operator<<(int32_t i) const {
-    if (cur_level_ <= level_) fprintf(stderr, "%d", i);
-    return *this;
-  }
-
-  const Logger &operator<<(uint32_t i) const {
-    if (cur_level_ <= level_) fprintf(stderr, "%u", i);
-    return *this;
-  }
-
-  const Logger &operator<<(uint64_t i) const {
-    if (cur_level_ <= level_)
-      fprintf(stderr, "%llu", (long long unsigned int)i);  // NOLINT
-    return *this;
-  }
-
-  const Logger &operator<<(int64_t i) const {
-    if (cur_level_ <= level_)
-      fprintf(stderr, "%lli", (long long int)i);  // NOLINT
-    return *this;
-  }
-
-  const Logger &operator<<(float f) const {
-    if (cur_level_ <= level_) fprintf(stderr, "%f", f);
-    return *this;
-  }
-
-  const Logger &operator<<(double d) const {
-    if (cur_level_ <= level_) fprintf(stderr, "%f", d);
-    return *this;
-  }
-
-  template <typename T>
-  const Logger &operator<<(const T &t) const {
-    // require T overloads operator<<
-    std::ostringstream os;
-    os << t;
-    return *this << os.str().c_str();
-  }
-
-  // specialization to fix compile error: `stringstream << nullptr` is ambiguous
-  const Logger &operator<<(const std::nullptr_t &null) const {
-    if (cur_level_ <= level_) *this << "(null)";
-    return *this;
-  }
-
- private:
-  const char *filename_;
-  const char *func_name_;
-  uint32_t line_num_;
-  LogLevel level_;
-  LogLevel cur_level_;
-};
-#endif  // KNF_ENABLE_CHECK
-
-class Voidifier {
- public:
-#if KNF_ENABLE_CHECK
-  void operator&(const Logger &) const {}
-#endif
-};
-#if !defined(KNF_ENABLE_CHECK)
-template <typename T>
-const Voidifier &operator<<(const Voidifier &v, T &&) {
-  return v;
-}
-#endif
-
-}  // namespace knf
-
-#define KNF_STATIC_ASSERT(x) static_assert(x, "")
-
-#ifdef KNF_ENABLE_CHECK
-
-#if defined(__clang__) || defined(__GNUC__) || defined(__GNUG__) || \
-    defined(__PRETTY_FUNCTION__)
-// for clang and GCC
-#define KNF_FUNC __PRETTY_FUNCTION__
-#else
-// for other compilers
-#define KNF_FUNC __func__
-#endif
-
-#define KNF_CHECK(x)                                                  \
-  (x) ? (void)0                                                       \
-      : ::knf::Voidifier() &                                          \
-            ::knf::Logger(__FILE__, KNF_FUNC, __LINE__, ::knf::FATAL) \
-                << "Check failed: " << #x << " "
-
-// WARNING: x and y may be evaluated multiple times, but this happens only
-// when the check fails. Since the program aborts if it fails, we don't think
-// the extra evaluation of x and y matters.
-//
-// CAUTION: we recommend the following use case:
-//
-//      auto x = Foo();
-//      auto y = Bar();
-//      KNF_CHECK_EQ(x, y) << "Some message";
-//
-//  And please avoid
-//
-//      KNF_CHECK_EQ(Foo(), Bar());
-//
-//  if `Foo()` or `Bar()` causes some side effects, e.g., changing some
-//  local static variables or global variables.
-#define _KNF_CHECK_OP(x, y, op)                                              \
-  ((x)op(y)) ? (void)0                                                       \
-             : ::knf::Voidifier() &                                          \
-                   ::knf::Logger(__FILE__, KNF_FUNC, __LINE__, ::knf::FATAL) \
-                       << "Check failed: " << #x << " " << #op << " " << #y  \
-                       << " (" << (x) << " vs. " << (y) << ") "
-
-#define KNF_CHECK_EQ(x, y) _KNF_CHECK_OP(x, y, ==)
-#define KNF_CHECK_NE(x, y) _KNF_CHECK_OP(x, y, !=)
-#define KNF_CHECK_LT(x, y) _KNF_CHECK_OP(x, y, <)
-#define KNF_CHECK_LE(x, y) _KNF_CHECK_OP(x, y, <=)
-#define KNF_CHECK_GT(x, y) _KNF_CHECK_OP(x, y, >)
-#define KNF_CHECK_GE(x, y) _KNF_CHECK_OP(x, y, >=)
-
-#define KNF_LOG(x) ::knf::Logger(__FILE__, KNF_FUNC, __LINE__, ::knf::x)
-
-// ------------------------------------------------------------
-//       For debug check
-// ------------------------------------------------------------
-// If you define the macro "-D NDEBUG" while compiling kaldi-native-fbank,
-// the following macros are in fact empty and does nothing.
-
-#define KNF_DCHECK(x) ::knf::kDisableDebug ? (void)0 : KNF_CHECK(x)
-
-#define KNF_DCHECK_EQ(x, y) ::knf::kDisableDebug ? (void)0 : KNF_CHECK_EQ(x, y)
-
-#define KNF_DCHECK_NE(x, y) ::knf::kDisableDebug ? (void)0 : KNF_CHECK_NE(x, y)
-
-#define KNF_DCHECK_LT(x, y) ::knf::kDisableDebug ? (void)0 : KNF_CHECK_LT(x, y)
-
-#define KNF_DCHECK_LE(x, y) ::knf::kDisableDebug ? (void)0 : KNF_CHECK_LE(x, y)
-
-#define KNF_DCHECK_GT(x, y) ::knf::kDisableDebug ? (void)0 : KNF_CHECK_GT(x, y)
-
-#define KNF_DCHECK_GE(x, y) ::knf::kDisableDebug ? (void)0 : KNF_CHECK_GE(x, y)
-
-#define KNF_DLOG(x) \
-  ::knf::kDisableDebug ? (void)0 : ::knf::Voidifier() & KNF_LOG(x)
-
-#else
-
-#define KNF_CHECK(x) ::knf::Voidifier()
-#define KNF_LOG(x) ::knf::Voidifier()
-
-#define KNF_CHECK_EQ(x, y) ::knf::Voidifier()
-#define KNF_CHECK_NE(x, y) ::knf::Voidifier()
-#define KNF_CHECK_LT(x, y) ::knf::Voidifier()
-#define KNF_CHECK_LE(x, y) ::knf::Voidifier()
-#define KNF_CHECK_GT(x, y) ::knf::Voidifier()
-#define KNF_CHECK_GE(x, y) ::knf::Voidifier()
-
-#define KNF_DCHECK(x) ::knf::Voidifier()
-#define KNF_DLOG(x) ::knf::Voidifier()
-#define KNF_DCHECK_EQ(x, y) ::knf::Voidifier()
-#define KNF_DCHECK_NE(x, y) ::knf::Voidifier()
-#define KNF_DCHECK_LT(x, y) ::knf::Voidifier()
-#define KNF_DCHECK_LE(x, y) ::knf::Voidifier()
-#define KNF_DCHECK_GT(x, y) ::knf::Voidifier()
-#define KNF_DCHECK_GE(x, y) ::knf::Voidifier()
-
-#endif  // KNF_CHECK_NE
-
-#endif  // KALDI_NATIVE_FBANK_CSRC_LOG_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.cc
deleted file mode 100644
index bdf55412..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.cc
+++ /dev/null
@@ -1,457 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-// This file is copied/modified from kaldi/src/feat/mel-computations.cc
-
-#include "kaldi-native-fbank/csrc/mel-computations.h"
-
-#include <stdio.h>
-
-#include <algorithm>
-#include <sstream>
-#include <vector>
-
-#include "kaldi-native-fbank/csrc/feature-window.h"
-#include "kaldi-native-fbank/csrc/kaldi-math.h"
-#include "kaldi-native-fbank/csrc/log.h"
-
-namespace knf {
-
-std::ostream &operator<<(std::ostream &os, const MelBanksOptions &opts) {
-  os << opts.ToString();
-  return os;
-}
-
-float MelBanks::VtlnWarpFreq(
-    float vtln_low_cutoff,  // upper+lower frequency cutoffs for VTLN.
-    float vtln_high_cutoff,
-    float low_freq,  // upper+lower frequency cutoffs in mel computation
-    float high_freq, float vtln_warp_factor, float freq) {
-  /// This computes a VTLN warping function that is not the same as HTK's one,
-  /// but has similar inputs (this function has the advantage of never producing
-  /// empty bins).
-
-  /// This function computes a warp function F(freq), defined between low_freq
-  /// and high_freq inclusive, with the following properties:
-  ///  F(low_freq) == low_freq
-  ///  F(high_freq) == high_freq
-  /// The function is continuous and piecewise linear with two inflection
-  ///   points.
-  /// The lower inflection point (measured in terms of the unwarped
-  ///  frequency) is at frequency l, determined as described below.
-  /// The higher inflection point is at a frequency h, determined as
-  ///   described below.
-  /// If l <= f <= h, then F(f) = f/vtln_warp_factor.
-  /// If the higher inflection point (measured in terms of the unwarped
-  ///   frequency) is at h, then max(h, F(h)) == vtln_high_cutoff.
-  ///   Since (by the last point) F(h) == h/vtln_warp_factor, then
-  ///   max(h, h/vtln_warp_factor) == vtln_high_cutoff, so
-  ///   h = vtln_high_cutoff / max(1, 1/vtln_warp_factor).
-  ///     = vtln_high_cutoff * min(1, vtln_warp_factor).
-  /// If the lower inflection point (measured in terms of the unwarped
-  ///   frequency) is at l, then min(l, F(l)) == vtln_low_cutoff
-  ///   This implies that l = vtln_low_cutoff / min(1, 1/vtln_warp_factor)
-  ///                       = vtln_low_cutoff * max(1, vtln_warp_factor)
-
-  if (freq < low_freq || freq > high_freq)
-    return freq;  // in case this gets called
-  // for out-of-range frequencies, just return the freq.
-
-  KNF_CHECK_GT(vtln_low_cutoff, low_freq);
-  KNF_CHECK_LT(vtln_high_cutoff, high_freq);
-
-  float one = 1.0f;
-  float l = vtln_low_cutoff * std::max(one, vtln_warp_factor);
-  float h = vtln_high_cutoff * std::min(one, vtln_warp_factor);
-  float scale = 1.0f / vtln_warp_factor;
-  float Fl = scale * l;  // F(l);
-  float Fh = scale * h;  // F(h);
-  KNF_CHECK(l > low_freq && h < high_freq);
-  // slope of left part of the 3-piece linear function
-  float scale_left = (Fl - low_freq) / (l - low_freq);
-  // [slope of center part is just "scale"]
-
-  // slope of right part of the 3-piece linear function
-  float scale_right = (high_freq - Fh) / (high_freq - h);
-
-  if (freq < l) {
-    return low_freq + scale_left * (freq - low_freq);
-  } else if (freq < h) {
-    return scale * freq;
-  } else {  // freq >= h
-    return high_freq + scale_right * (freq - high_freq);
-  }
-}
-
-float MelBanks::VtlnWarpMelFreq(
-    float vtln_low_cutoff,  // upper+lower frequency cutoffs for VTLN.
-    float vtln_high_cutoff,
-    float low_freq,  // upper+lower frequency cutoffs in mel computation
-    float high_freq, float vtln_warp_factor, float mel_freq) {
-  return MelScale(VtlnWarpFreq(vtln_low_cutoff, vtln_high_cutoff, low_freq,
-                               high_freq, vtln_warp_factor,
-                               InverseMelScale(mel_freq)));
-}
-
-MelBanks::MelBanks(const MelBanksOptions &opts,
-                   const FrameExtractionOptions &frame_opts,
-                   float vtln_warp_factor) {
-  if (opts.is_librosa) {
-    InitLibrosaMelBanks(opts, frame_opts, vtln_warp_factor);
-  } else {
-    InitKaldiMelBanks(opts, frame_opts, vtln_warp_factor);
-  }
-}
-
-void MelBanks::InitKaldiMelBanks(const MelBanksOptions &opts,
-                                 const FrameExtractionOptions &frame_opts,
-                                 float vtln_warp_factor) {
-  htk_mode_ = opts.htk_mode;
-  int32_t num_bins = opts.num_bins;
-  if (num_bins < 3) {
-    KNF_LOG(FATAL) << "Must have at least 3 mel bins";
-  }
-
-  float sample_freq = frame_opts.samp_freq;
-  int32_t window_length_padded = frame_opts.PaddedWindowSize();
-  KNF_CHECK_EQ(window_length_padded % 2, 0);
-
-  int32_t num_fft_bins = window_length_padded / 2;
-  float nyquist = 0.5f * sample_freq;
-
-  float low_freq = opts.low_freq, high_freq;
-  if (opts.high_freq > 0.0f) {
-    high_freq = opts.high_freq;
-  } else {
-    high_freq = nyquist + opts.high_freq;
-  }
-
-  if (low_freq < 0.0f || low_freq >= nyquist || high_freq <= 0.0f ||
-      high_freq > nyquist || high_freq <= low_freq) {
-    KNF_LOG(FATAL) << "Bad values in options: low-freq " << low_freq
-                   << " and high-freq " << high_freq << " vs. nyquist "
-                   << nyquist;
-  }
-
-  float fft_bin_width = sample_freq / window_length_padded;
-  // fft-bin width [think of it as Nyquist-freq / half-window-length]
-
-  float mel_low_freq = MelScale(low_freq);
-  float mel_high_freq = MelScale(high_freq);
-
-  debug_ = opts.debug_mel;
-
-  // divide by num_bins+1 in next line because of end-effects where the bins
-  // spread out to the sides.
-  float mel_freq_delta = (mel_high_freq - mel_low_freq) / (num_bins + 1);
-
-  float vtln_low = opts.vtln_low, vtln_high = opts.vtln_high;
-  if (vtln_high < 0.0f) {
-    vtln_high += nyquist;
-  }
-
-  if (vtln_warp_factor != 1.0f &&
-      (vtln_low < 0.0f || vtln_low <= low_freq || vtln_low >= high_freq ||
-       vtln_high <= 0.0f || vtln_high >= high_freq || vtln_high <= vtln_low)) {
-    KNF_LOG(FATAL) << "Bad values in options: vtln-low " << vtln_low
-                   << " and vtln-high " << vtln_high << ", versus "
-                   << "low-freq " << low_freq << " and high-freq " << high_freq;
-  }
-
-  bins_.resize(num_bins);
-
-  for (int32_t bin = 0; bin < num_bins; ++bin) {
-    float left_mel = mel_low_freq + bin * mel_freq_delta,
-          center_mel = mel_low_freq + (bin + 1) * mel_freq_delta,
-          right_mel = mel_low_freq + (bin + 2) * mel_freq_delta;
-
-    if (vtln_warp_factor != 1.0f) {
-      left_mel = VtlnWarpMelFreq(vtln_low, vtln_high, low_freq, high_freq,
-                                 vtln_warp_factor, left_mel);
-      center_mel = VtlnWarpMelFreq(vtln_low, vtln_high, low_freq, high_freq,
-                                   vtln_warp_factor, center_mel);
-      right_mel = VtlnWarpMelFreq(vtln_low, vtln_high, low_freq, high_freq,
-                                  vtln_warp_factor, right_mel);
-    }
-
-    // this_bin will be a vector of coefficients that is only
-    // nonzero where this mel bin is active.
-    std::vector<float> this_bin(num_fft_bins);
-
-    int32_t first_index = -1, last_index = -1;
-    for (int32_t i = 0; i < num_fft_bins; ++i) {
-      float freq = (fft_bin_width * i);  // Center frequency of this fft
-                                         // bin.
-      float mel = MelScale(freq);
-      if (mel > left_mel && mel < right_mel) {
-        float weight;
-        if (mel <= center_mel) {
-          weight = (mel - left_mel) / (center_mel - left_mel);
-        } else {
-          weight = (right_mel - mel) / (right_mel - center_mel);
-        }
-        this_bin[i] = weight;
-        if (first_index == -1) {
-          first_index = i;
-        }
-        last_index = i;
-      }
-    }
-    KNF_CHECK(first_index != -1 && last_index >= first_index &&
-              "You may have set num_mel_bins too large.");
-
-    bins_[bin].first = first_index;
-    int32_t size = last_index + 1 - first_index;
-    bins_[bin].second.insert(bins_[bin].second.end(),
-                             this_bin.begin() + first_index,
-                             this_bin.begin() + first_index + size);
-
-    // Replicate a bug in HTK, for testing purposes.
-    if (opts.htk_mode && bin == 0 && mel_low_freq != 0.0f) {
-      bins_[bin].second[0] = 0.0;
-    }
-  }  // for (int32_t bin = 0; bin < num_bins; ++bin) {
-
-  if (debug_) {
-    std::ostringstream os;
-    for (size_t i = 0; i < bins_.size(); i++) {
-      os << "bin " << i << ", offset = " << bins_[i].first << ", vec = ";
-      for (auto k : bins_[i].second) os << k << ", ";
-      os << "\n";
-    }
-    KNF_LOG(INFO) << os.str();
-  }
-}
-
-void MelBanks::InitLibrosaMelBanks(const MelBanksOptions &opts,
-                                   const FrameExtractionOptions &frame_opts,
-                                   float vtln_warp_factor) {
-  htk_mode_ = opts.htk_mode;
-  int32_t num_bins = opts.num_bins;
-  if (num_bins < 3) {
-    KNF_LOG(FATAL) << "Must have at least 3 mel bins";
-  }
-
-  float sample_freq = frame_opts.samp_freq;
-  int32_t window_length_padded = frame_opts.PaddedWindowSize();
-  KNF_CHECK_EQ(window_length_padded % 2, 0);
-
-  int32_t num_fft_bins = window_length_padded / 2;
-  float nyquist = 0.5f * sample_freq;
-
-  float low_freq = opts.low_freq, high_freq;
-  if (opts.high_freq > 0.0f) {
-    high_freq = opts.high_freq;
-  } else {
-    high_freq = nyquist + opts.high_freq;
-  }
-
-  if (low_freq < 0.0f || low_freq >= nyquist || high_freq <= 0.0f ||
-      high_freq > nyquist || high_freq <= low_freq) {
-    KNF_LOG(FATAL) << "Bad values in options: low-freq " << low_freq
-                   << " and high-freq " << high_freq << " vs. nyquist "
-                   << nyquist;
-  }
-
-  float fft_bin_width = sample_freq / window_length_padded;
-
-  float mel_low_freq;
-  float mel_high_freq;
-
-  if (opts.use_slaney_mel_scale) {
-    mel_low_freq = MelScaleSlaney(low_freq);
-    mel_high_freq = MelScaleSlaney(high_freq);
-  } else {
-    mel_low_freq = MelScale(low_freq);
-    mel_high_freq = MelScale(high_freq);
-  }
-
-  debug_ = opts.debug_mel;
-
-  // divide by num_bins+1 in next line because of end-effects where the bins
-  // spread out to the sides.
-  float mel_freq_delta = (mel_high_freq - mel_low_freq) / (num_bins + 1);
-
-  bool slaney_norm = false;
-  if (!opts.norm.empty()) {
-    if (opts.norm != "slaney") {
-      KNF_LOG(FATAL) << "Unsupported norm: " << opts.norm;
-    }
-    slaney_norm = true;
-  }
-
-  bins_.resize(num_bins);
-  for (int32_t bin = 0; bin < num_bins; ++bin) {
-    float left_mel = mel_low_freq + bin * mel_freq_delta;
-    float center_mel = mel_low_freq + (bin + 1) * mel_freq_delta;
-    float right_mel = mel_low_freq + (bin + 2) * mel_freq_delta;
-
-    float left_hz;
-    float center_hz;
-    float right_hz;
-
-    if (opts.use_slaney_mel_scale) {
-      left_hz = InverseMelScaleSlaney(left_mel);
-      center_hz = InverseMelScaleSlaney(center_mel);
-      right_hz = InverseMelScaleSlaney(right_mel);
-    } else {
-      left_hz = InverseMelScale(left_mel);
-      center_hz = InverseMelScale(center_mel);
-      right_hz = InverseMelScale(right_mel);
-    }
-
-    if (opts.floor_to_int_bin) {
-      left_hz *= (window_length_padded + 1.0) / sample_freq;
-      center_hz *= (window_length_padded + 1.0) / sample_freq;
-      right_hz *= (window_length_padded + 1.0) / sample_freq;
-
-      left_hz = static_cast<int32_t>(left_hz);
-      center_hz = static_cast<int32_t>(center_hz);
-      right_hz = static_cast<int32_t>(right_hz);
-    }
-
-    // this_bin will be a vector of coefficients that is only
-    // nonzero where this mel bin is active.
-    //
-    // It is not an error to use num_fft_bins + 1 here. It is different
-    // from Kaldi.
-    std::vector<float> this_bin(num_fft_bins + 1);
-
-    int32_t first_index = -1, last_index = -1;
-    for (int32_t i = 0; i < num_fft_bins + 1; ++i) {
-      float hz;  // Center frequency of this fft bin.
-
-      if (opts.floor_to_int_bin) {
-        hz = i;
-      } else {
-        hz = fft_bin_width * i;
-      }
-
-      if (hz > left_hz && hz < right_hz) {
-        float weight;
-        if (hz <= center_hz) {
-          weight = (hz - left_hz) / (center_hz - left_hz);
-        } else {
-          weight = (right_hz - hz) / (right_hz - center_hz);
-        }
-
-        if (slaney_norm) {
-          weight *= 2 / (right_hz - left_hz);
-        }
-
-        this_bin[i] = weight;
-        if (first_index == -1) {
-          first_index = i;
-        }
-
-        last_index = i;
-      }
-    }  // for (int32_t i = 0; i < num_fft_bins + 1; ++i)
-
-    KNF_CHECK(first_index != -1 && last_index >= first_index &&
-              "You may have set num_mel_bins too large.");
-
-    bins_[bin].first = first_index;
-    int32_t size = last_index + 1 - first_index;
-    bins_[bin].second.insert(bins_[bin].second.end(),
-                             this_bin.begin() + first_index,
-                             this_bin.begin() + first_index + size);
-  }  // for (int32_t bin = 0; bin < num_bins; ++bin)
-
-  if (debug_) {
-    std::ostringstream os;
-    for (size_t i = 0; i < bins_.size(); i++) {
-      os << "bin " << i << ", offset = " << bins_[i].first << ", vec = ";
-      for (auto k : bins_[i].second) os << k << ", ";
-      os << "\n";
-    }
-    fprintf(stderr, "%s\n", os.str().c_str());
-  }
-}
-
-MelBanks::MelBanks(const float *weights, int32_t num_rows, int32_t num_cols)
-    : debug_(false), htk_mode_(false) {
-  bins_.resize(num_rows);
-  for (int32_t bin = 0; bin < num_rows; ++bin) {
-    const float *this_bin = weights + bin * num_cols;
-
-    int32_t first_index = -1, last_index = -1;
-
-    for (int32_t i = 0; i < num_cols; ++i) {
-      if (this_bin[i] == 0) {
-        continue;
-      }
-      if (first_index == -1) first_index = i;
-      last_index = i;
-    }
-
-    KNF_CHECK(first_index != -1 && last_index >= first_index &&
-              "You have an incorrect weight matrix.");
-
-    bins_[bin].first = first_index;
-    int32_t size = last_index + 1 - first_index;
-
-    bins_[bin].second.insert(bins_[bin].second.end(), this_bin + first_index,
-                             this_bin + first_index + size);
-  }
-}
-
-// "power_spectrum" contains fft energies.
-void MelBanks::Compute(const float *power_spectrum,
-                       float *mel_energies_out) const {
-  int32_t num_bins = bins_.size();
-
-  for (int32_t i = 0; i < num_bins; i++) {
-    int32_t offset = bins_[i].first;
-    const auto &v = bins_[i].second;
-    float energy = 0;
-    for (int32_t k = 0; k != v.size(); ++k) {
-      energy += v[k] * power_spectrum[k + offset];
-    }
-
-    // HTK-like flooring- for testing purposes (we prefer dither)
-    if (htk_mode_ && energy < 1.0) {
-      energy = 1.0;
-    }
-
-    mel_energies_out[i] = energy;
-
-    // The following assert was added due to a problem with OpenBlas that
-    // we had at one point (it was a bug in that library).  Just to detect
-    // it early.
-    KNF_CHECK_EQ(energy, energy);  // check that energy is not nan
-  }
-
-  if (debug_) {
-    fprintf(stderr, "MEL BANKS:\n");
-    for (int32_t i = 0; i < num_bins; i++)
-      fprintf(stderr, " %f", mel_energies_out[i]);
-    fprintf(stderr, "\n");
-  }
-}
-
-void ComputeLifterCoeffs(float Q, std::vector<float> *coeffs) {
-  // Compute liftering coefficients (scaling on cepstral coeffs)
-  // coeffs are numbered slightly differently from HTK: the zeroth
-  // index is C0, which is not affected.
-  for (int32_t i = 0; i != static_cast<int32_t>(coeffs->size()); ++i) {
-    (*coeffs)[i] = 1.0 + 0.5 * Q * sin(M_PI * i / Q);
-  }
-}
-
-}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.h
deleted file mode 100644
index 29904071..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/mel-computations.h
+++ /dev/null
@@ -1,190 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-// This file is copied/modified from kaldi/src/feat/mel-computations.h
-#ifndef KALDI_NATIVE_FBANK_CSRC_MEL_COMPUTATIONS_H_
-#define KALDI_NATIVE_FBANK_CSRC_MEL_COMPUTATIONS_H_
-
-#include <cmath>
-#include <cstdint>
-#include <sstream>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "kaldi-native-fbank/csrc/feature-window.h"
-
-namespace knf {
-struct FrameExtractionOptions;
-
-struct MelBanksOptions {
-  int32_t num_bins = 25;  // e.g. 25; number of triangular bins
-  float low_freq = 20;    // e.g. 20; lower frequency cutoff
-
-  // an upper frequency cutoff; 0 -> no cutoff, negative
-  // ->added to the Nyquist frequency to get the cutoff.
-  float high_freq = 0;
-
-  float vtln_low = 100;  // vtln lower cutoff of warping function.
-
-  // vtln upper cutoff of warping function: if negative, added
-  // to the Nyquist frequency to get the cutoff.
-  float vtln_high = -500;
-
-  bool debug_mel = false;
-  // htk_mode is a "hidden" config, it does not show up on command line.
-  // Enables more exact compatibility with HTK, for testing purposes.  Affects
-  // mel-energy flooring and reproduces a bug in HTK.
-  bool htk_mode = false;
-
-  // Note that if you set is_librosa, you probably need to set
-  // low_freq to 0.
-  // Please see
-  // https://librosa.org/doc/main/generated/librosa.filters.mel.html
-  bool is_librosa = false;
-
-  // used only when is_librosa=true
-  // Possible values: "", slaney. We don't support a numeric value here, but
-  // it can be added on demand.
-  // See https://librosa.org/doc/main/generated/librosa.filters.mel.html
-  std::string norm = "slaney";
-
-  // used only when is_librosa is true
-  bool use_slaney_mel_scale = true;
-
-  // used only when is_librosa is true
-  bool floor_to_int_bin = false;
-
-  std::string ToString() const {
-    std::ostringstream os;
-    os << "num_bins: " << num_bins << "\n";
-    os << "low_freq: " << low_freq << "\n";
-    os << "high_freq: " << high_freq << "\n";
-    os << "vtln_low: " << vtln_low << "\n";
-    os << "vtln_high: " << vtln_high << "\n";
-    os << "debug_mel: " << debug_mel << "\n";
-    os << "htk_mode: " << htk_mode << "\n";
-    os << "is_librosa: " << is_librosa << "\n";
-    os << "norm: " << norm << "\n";
-    os << "use_slaney_mel_scale: " << use_slaney_mel_scale << "\n";
-    os << "floor_to_int_bin: " << floor_to_int_bin << "\n";
-    return os.str();
-  }
-};
-
-std::ostream &operator<<(std::ostream &os, const MelBanksOptions &opts);
-
-class MelBanks {
- public:
-  // see also https://en.wikipedia.org/wiki/Mel_scale
-  // htk, mel to hz
-  static inline float InverseMelScale(float mel_freq) {
-    return 700.0f * (expf(mel_freq / 1127.0f) - 1.0f);
-  }
-
-  // htk, hz to mel
-  static inline float MelScale(float freq) {
-    return 1127.0f * logf(1.0f + freq / 700.0f);
-  }
-
-  // slaney, mel to hz
-  static inline float InverseMelScaleSlaney(float mel_freq) {
-    if (mel_freq <= 15) {
-      return 200.0f / 3 * mel_freq;
-    }
-
-    // return 1000 * expf((mel_freq - 15) * logf(6.4f) / 27);
-
-    // Note: log(6.4)/27 = 0.06875177742094911
-
-    return 1000 * expf((mel_freq - 15) * 0.06875177742094911f);
-  }
-
-  // slaney, hz to mel
-  static inline float MelScaleSlaney(float freq) {
-    if (freq <= 1000) {
-      return freq * 3 / 200.0f;
-    }
-
-    // return 15 + 27 * logf(freq / 1000) / logf(6.4f)
-    //
-    // Note: 27/log(6.4) = 14.545078505785561
-
-    return 15 + 14.545078505785561f * logf(freq / 1000);
-  }
-
-  static float VtlnWarpFreq(
-      float vtln_low_cutoff,
-      float vtln_high_cutoff,  // discontinuities in warp func
-      float low_freq,
-      float high_freq,  // upper+lower frequency cutoffs in
-      // the mel computation
-      float vtln_warp_factor, float freq);
-
-  static float VtlnWarpMelFreq(float vtln_low_cutoff, float vtln_high_cutoff,
-                               float low_freq, float high_freq,
-                               float vtln_warp_factor, float mel_freq);
-
-  // TODO(fangjun): Remove vtln_warp_factor
-  MelBanks(const MelBanksOptions &opts,
-           const FrameExtractionOptions &frame_opts, float vtln_warp_factor);
-
-  // Initialize with a 2-d weights matrix
-  // @param weights Pointer to the start address of the matrix
-  // @param num_rows It equls to number of mel bins
-  // @param num_cols It equals to (number of fft bins)/2+1
-  MelBanks(const float *weights, int32_t num_rows, int32_t num_cols);
-
-  /// Compute Mel energies (note: not log energies).
-  /// At input, "fft_energies" contains the FFT energies (not log).
-  ///
-  /// @param fft_energies 1-D array of size num_fft_bins/2+1
-  /// @param mel_energies_out  1-D array of size num_mel_bins
-  void Compute(const float *fft_energies, float *mel_energies_out) const;
-
-  int32_t NumBins() const { return bins_.size(); }
-
- private:
-  // for kaldi-compatible
-  void InitKaldiMelBanks(const MelBanksOptions &opts,
-                         const FrameExtractionOptions &frame_opts,
-                         float vtln_warp_factor);
-
-  // for librosa-compatible
-  // See https://librosa.org/doc/main/generated/librosa.filters.mel.html
-  void InitLibrosaMelBanks(const MelBanksOptions &opts,
-                           const FrameExtractionOptions &frame_opts,
-                           float vtln_warp_factor);
-
- private:
-  // the "bins_" vector is a vector, one for each bin, of a pair:
-  // (the first nonzero fft-bin), (the vector of weights).
-  std::vector<std::pair<int32_t, std::vector<float>>> bins_;
-
-  // TODO(fangjun): Remove debug_ and htk_mode_
-  bool debug_ = false;
-  bool htk_mode_ = false;
-};
-
-// Compute liftering coefficients (scaling on cepstral coeffs)
-// coeffs are numbered slightly differently from HTK: the zeroth
-// index is C0, which is not affected.
-void ComputeLifterCoeffs(float Q, std::vector<float> *coeffs);
-
-}  // namespace knf
-
-#endif  // KALDI_NATIVE_FBANK_CSRC_MEL_COMPUTATIONS_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.cc
deleted file mode 100644
index 5f96ea43..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.cc
+++ /dev/null
@@ -1,170 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-// The content in this file is copied/modified from
-// This file is copied/modified from kaldi/src/feat/online-feature.cc
-
-#include "kaldi-native-fbank/csrc/online-feature.h"
-
-#include <algorithm>
-#include <cstddef>
-#include <utility>
-#include <vector>
-
-#include "kaldi-native-fbank/csrc/feature-window.h"
-#include "kaldi-native-fbank/csrc/log.h"
-
-namespace knf {
-
-RecyclingVector::RecyclingVector(int32_t items_to_hold)
-    : items_to_hold_(items_to_hold == 0 ? -1 : items_to_hold),
-      first_available_index_(0) {}
-
-const float *RecyclingVector::At(int32_t index) const {
-  if (index < first_available_index_) {
-    KNF_LOG(FATAL) << "Attempted to retrieve feature vector that was "
-                      "already removed by the RecyclingVector (index = "
-                   << index << "; "
-                   << "first_available_index = " << first_available_index_
-                   << "; "
-                   << "size = " << Size() << ")";
-  }
-  // 'at' does size checking.
-  return items_.at(index - first_available_index_).data();
-}
-
-void RecyclingVector::PushBack(std::vector<float> item) {
-  // Note: -1 is a larger number when treated as unsigned
-  if (items_.size() == static_cast<size_t>(items_to_hold_)) {
-    items_.pop_front();
-    ++first_available_index_;
-  }
-  items_.push_back(std::move(item));
-}
-
-int32_t RecyclingVector::Size() const {
-  return first_available_index_ + static_cast<int32_t>(items_.size());
-}
-
-// discard the first n frames
-void RecyclingVector::Pop(int32_t n) {
-  for (int32_t i = 0; i < n && !items_.empty(); ++i) {
-    items_.pop_front();
-    ++first_available_index_;
-  }
-}
-
-template <class C>
-OnlineGenericBaseFeature<C>::OnlineGenericBaseFeature(
-    const typename C::Options &opts)
-    : computer_(opts),
-      window_function_(computer_.GetFrameOptions()),
-      input_finished_(false),
-      waveform_offset_(0) {}
-
-template <class C>
-void OnlineGenericBaseFeature<C>::AcceptWaveform(float sampling_rate,
-                                                 const float *waveform,
-                                                 int32_t n) {
-  if (n == 0) {
-    return;  // Nothing to do.
-  }
-
-  if (input_finished_) {
-    KNF_LOG(FATAL) << "AcceptWaveform called after InputFinished() was called.";
-  }
-
-  KNF_CHECK_EQ(sampling_rate, computer_.GetFrameOptions().samp_freq);
-
-  waveform_remainder_.insert(waveform_remainder_.end(), waveform, waveform + n);
-
-  ComputeFeatures();
-}
-
-template <class C>
-void OnlineGenericBaseFeature<C>::InputFinished() {
-  input_finished_ = true;
-  ComputeFeatures();
-}
-
-template <class C>
-void OnlineGenericBaseFeature<C>::ComputeFeatures() {
-  const FrameExtractionOptions &frame_opts = computer_.GetFrameOptions();
-
-  int64_t num_samples_total = waveform_offset_ + waveform_remainder_.size();
-
-  int32_t num_frames_old = features_.Size();
-
-  int32_t num_frames_new =
-      NumFrames(num_samples_total, frame_opts, input_finished_);
-
-  KNF_CHECK_GE(num_frames_new, num_frames_old);
-
-  // note: this online feature-extraction code does not support VTLN.
-  float vtln_warp = 1.0;
-
-  std::vector<float> window;
-  bool need_raw_log_energy = computer_.NeedRawLogEnergy();
-
-  for (int32_t frame = num_frames_old; frame < num_frames_new; ++frame) {
-    std::fill(window.begin(), window.end(), 0);
-    float raw_log_energy = 0.0;
-    ExtractWindow(waveform_offset_, waveform_remainder_, frame, frame_opts,
-                  window_function_, &window,
-                  need_raw_log_energy ? &raw_log_energy : nullptr);
-
-    std::vector<float> this_feature(computer_.Dim());
-
-    computer_.Compute(raw_log_energy, vtln_warp, &window, this_feature.data());
-    features_.PushBack(std::move(this_feature));
-  }
-
-  // OK, we will now discard any portion of the signal that will not be
-  // necessary to compute frames in the future.
-  int64_t first_sample_of_next_frame =
-      FirstSampleOfFrame(num_frames_new, frame_opts);
-
-  int32_t samples_to_discard = first_sample_of_next_frame - waveform_offset_;
-
-  if (samples_to_discard > 0) {
-    // discard the leftmost part of the waveform that we no longer need.
-    int32_t new_num_samples =
-        static_cast<int32_t>(waveform_remainder_.size()) - samples_to_discard;
-
-    if (new_num_samples <= 0) {
-      // odd, but we'll try to handle it.
-      waveform_offset_ += waveform_remainder_.size();
-      waveform_remainder_.resize(0);
-    } else {
-      std::vector<float> new_remainder(new_num_samples);
-
-      std::copy(waveform_remainder_.begin() + samples_to_discard,
-                waveform_remainder_.end(), new_remainder.begin());
-      waveform_offset_ += samples_to_discard;
-
-      waveform_remainder_.swap(new_remainder);
-    }
-  }
-}
-
-template class OnlineGenericBaseFeature<RawAudioSamplesComputer>;
-template class OnlineGenericBaseFeature<FbankComputer>;
-template class OnlineGenericBaseFeature<MfccComputer>;
-template class OnlineGenericBaseFeature<WhisperFeatureComputer>;
-
-}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.h
deleted file mode 100644
index 114567cf..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/online-feature.h
+++ /dev/null
@@ -1,155 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-// The content in this file is copied/modified from
-// This file is copied/modified from kaldi/src/feat/online-feature.h
-#ifndef KALDI_NATIVE_FBANK_CSRC_ONLINE_FEATURE_H_
-#define KALDI_NATIVE_FBANK_CSRC_ONLINE_FEATURE_H_
-
-#include <cstdint>
-#include <deque>
-#include <vector>
-
-#include "kaldi-native-fbank/csrc/feature-fbank.h"
-#include "kaldi-native-fbank/csrc/feature-mfcc.h"
-#include "kaldi-native-fbank/csrc/feature-raw-audio-samples.h"
-#include "kaldi-native-fbank/csrc/feature-window.h"
-#include "kaldi-native-fbank/csrc/whisper-feature.h"
-
-namespace knf {
-
-/// This class serves as a storage for feature vectors with an option to limit
-/// the memory usage by removing old elements. The deleted frames indices are
-/// "remembered" so that regardless of the MAX_ITEMS setting, the user always
-/// provides the indices as if no deletion was being performed.
-/// This is useful when processing very long recordings which would otherwise
-/// cause the memory to eventually blow up when the features are not being
-/// removed.
-class RecyclingVector {
- public:
-  /// By default it does not remove any elements.
-  explicit RecyclingVector(int32_t items_to_hold = -1);
-
-  ~RecyclingVector() = default;
-  RecyclingVector(const RecyclingVector &) = delete;
-  RecyclingVector &operator=(const RecyclingVector &) = delete;
-
-  // The pointer is owned by RecyclingVector
-  // Users should not free it
-  const float *At(int32_t index) const;
-
-  void PushBack(std::vector<float> item);
-
-  /// This method returns the size as if no "recycling" had happened,
-  /// i.e. equivalent to the number of times the PushBack method has been
-  /// called.
-  int32_t Size() const;
-
-  // discard the first n frames
-  void Pop(int32_t n);
-
- private:
-  std::deque<std::vector<float>> items_;
-  int32_t items_to_hold_;
-  int32_t first_available_index_;
-};
-
-/// This is a templated class for online feature extraction;
-/// it's templated on a class like MfccComputer or PlpComputer
-/// that does the basic feature extraction.
-template <class C>
-class OnlineGenericBaseFeature {
- public:
-  // Constructor from options class
-  explicit OnlineGenericBaseFeature(const typename C::Options &opts);
-
-  int32_t Dim() const { return computer_.Dim(); }
-
-  float FrameShiftInSeconds() const {
-    return computer_.GetFrameOptions().frame_shift_ms / 1000.0f;
-  }
-
-  int32_t NumFramesReady() const { return features_.Size(); }
-
-  // Note: IsLastFrame() will only ever return true if you have called
-  // InputFinished() (and this frame is the last frame).
-  bool IsLastFrame(int32_t frame) const {
-    return input_finished_ && frame == NumFramesReady() - 1;
-  }
-
-  const float *GetFrame(int32_t frame) const { return features_.At(frame); }
-
-  // This would be called from the application, when you get
-  // more wave data.  Note: the sampling_rate is only provided so
-  // the code can assert that it matches the sampling rate
-  // expected in the options.
-  //
-  // @param sampling_rate The sampling_rate of the input waveform
-  // @param waveform Pointer to a 1-D array of size n
-  // @param n Number of entries in waveform
-  void AcceptWaveform(float sampling_rate, const float *waveform, int32_t n);
-
-  // InputFinished() tells the class you won't be providing any
-  // more waveform.  This will help flush out the last frame or two
-  // of features, in the case where snip-edges == false; it also
-  // affects the return value of IsLastFrame().
-  void InputFinished();
-
-  // discard the first n frames
-  void Pop(int32_t n) { features_.Pop(n); }
-
- private:
-  // This function computes any additional feature frames that it is possible to
-  // compute from 'waveform_remainder_', which at this point may contain more
-  // than just a remainder-sized quantity (because AcceptWaveform() appends to
-  // waveform_remainder_ before calling this function).  It adds these feature
-  // frames to features_, and shifts off any now-unneeded samples of input from
-  // waveform_remainder_ while incrementing waveform_offset_ by the same amount.
-  void ComputeFeatures();
-
-  C computer_;  // class that does the MFCC or PLP or filterbank computation
-
-  FeatureWindowFunction window_function_;
-
-  // features_ is the Mfcc or Plp or Fbank features that we have already
-  // computed.
-
-  RecyclingVector features_;
-
-  // True if the user has called "InputFinished()"
-  bool input_finished_;
-
-  // waveform_offset_ is the number of samples of waveform that we have
-  // already discarded, i.e. that were prior to 'waveform_remainder_'.
-  int64_t waveform_offset_;
-
-  // waveform_remainder_ is a short piece of waveform that we may need to keep
-  // after extracting all the whole frames we can (whatever length of feature
-  // will be required for the next phase of computation).
-  // It is a 1-D tensor
-  std::vector<float> waveform_remainder_;
-};
-
-using OnlineRawAudioSamples = OnlineGenericBaseFeature<RawAudioSamplesComputer>;
-using OnlineFbank = OnlineGenericBaseFeature<FbankComputer>;
-using OnlineMfcc = OnlineGenericBaseFeature<MfccComputer>;
-using OnlineWhisperFbank = OnlineGenericBaseFeature<WhisperFeatureComputer>;
-
-}  // namespace knf
-
-#endif  // KALDI_NATIVE_FBANK_CSRC_ONLINE_FEATURE_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.cc
deleted file mode 100644
index 64ed83c9..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.cc
+++ /dev/null
@@ -1,112 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include "kaldi-native-fbank/csrc/rfft.h"
-
-#include <algorithm>
-#include <cmath>
-#include <vector>
-
-#include "kaldi-native-fbank/csrc/log.h"
-#include "kiss_fftr.h"
-
-namespace knf {
-
-class Rfft::RfftImpl {
- public:
-  RfftImpl(int32_t n, bool inverse) : n_(n), inverse_(inverse) {
-    if ((n & 1) != 0) {
-      fprintf(stderr, "n should be even. Given: %d \n", n);
-      exit(-1);
-    }
-
-    if (n < 0) {
-      fprintf(stderr, "n should be positive. Given: %d \n", n);
-      exit(-1);
-    }
-  }
-
-  void Compute(float *in_out) {
-    if (!inverse_) {
-      Forward(in_out);
-    } else {
-      Reverse(in_out);
-    }
-  }
-
-  void Compute(double *in_out) {
-    std::vector<float> f(in_out, in_out + n_);
-
-    Compute(f.data());
-
-    std::copy(f.begin(), f.end(), in_out);
-  }
-
- private:
-  void Forward(float *in_out) const {
-    kiss_fftr_cfg cfg = kiss_fftr_alloc(n_, 0, nullptr, nullptr);
-
-    std::vector<kiss_fft_cpx> out(n_ / 2 + 1);
-
-    kiss_fftr(cfg, in_out, out.data());
-
-    kiss_fftr_free(cfg);
-
-    in_out[0] = out[0].r;
-    in_out[1] = out[n_ / 2].r;
-
-    for (int32_t i = 1; i < n_ / 2; ++i) {
-      in_out[2 * i] = out[i].r;
-      in_out[2 * i + 1] = out[i].i;
-    }
-  }
-
-  void Reverse(float *in_out) const {
-    std::vector<kiss_fft_cpx> out(n_ / 2 + 1);
-    out[0].r = in_out[0];
-    out[0].i = 0;
-
-    out[n_ / 2].r = in_out[1];
-    out[n_ / 2].i = 0;
-
-    for (int32_t i = 1; i < n_ / 2; ++i) {
-      out[i].r = in_out[2 * i];
-      out[i].i = in_out[2 * i + 1];
-    }
-
-    kiss_fftr_cfg cfg = kiss_fftr_alloc(n_, 1, nullptr, nullptr);
-
-    kiss_fftri(cfg, out.data(), in_out);
-
-    kiss_fftr_free(cfg);
-  }
-
- private:
-  int32_t n_;
-  bool inverse_ = false;
-};
-
-Rfft::Rfft(int32_t n, bool inverse /*=false*/)
-    : impl_(std::make_unique<RfftImpl>(n, inverse)) {}
-
-Rfft::~Rfft() = default;
-
-void Rfft::Compute(float *in_out) { impl_->Compute(in_out); }
-void Rfft::Compute(double *in_out) { impl_->Compute(in_out); }
-
-}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.h
deleted file mode 100644
index c6fab6d9..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/rfft.h
+++ /dev/null
@@ -1,58 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef KALDI_NATIVE_FBANK_CSRC_RFFT_H_
-#define KALDI_NATIVE_FBANK_CSRC_RFFT_H_
-
-#include <cstdint>
-#include <memory>
-
-namespace knf {
-
-// n-point Real discrete Fourier transform
-// where n is even. n >= 2
-//
-//  R[k] = sum_j=0^n-1 in[j]*cos(2*pi*j*k/n), 0<=k<=n/2
-//  I[k] = sum_j=0^n-1 in[j]*sin(2*pi*j*k/n), 0<k<n/2
-class Rfft {
- public:
-  // @param n Number of fft bins. it should be even.
-  explicit Rfft(int32_t n, bool inverse = false);
-  ~Rfft();
-
-  /** @param in_out A 1-D array of size n.
-   *             On return:
-   *               in_out[0] = R[0]
-   *               in_out[1] = R[n/2]
-   *               for 1 < k < n/2,
-   *                 in_out[2*k] = R[k]
-   *                 in_out[2*k+1] = I[k]
-   *
-   */
-  void Compute(float *in_out);
-  void Compute(double *in_out);
-
- private:
-  class RfftImpl;
-
-  std::unique_ptr<RfftImpl> impl_;
-};
-
-}  // namespace knf
-
-#endif  // KALDI_NATIVE_FBANK_CSRC_RFFT_H_
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.cc
deleted file mode 100644
index 2cdb637f..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.cc
+++ /dev/null
@@ -1,142 +0,0 @@
-/**
- * Copyright (c)  2025  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-#include "kaldi-native-fbank/csrc/stft.h"
-
-#include <algorithm>
-#include <cmath>
-#include <sstream>
-#include <string>
-
-#include "kaldi-native-fbank/csrc/feature-window.h"
-#include "kaldi-native-fbank/csrc/rfft.h"
-
-namespace knf {
-std::string StftConfig::ToString() const {
-  std::ostringstream os;
-  os << "StftConfig(";
-  os << "n_fft=" << n_fft << ", ";
-  os << "hop_length=" << hop_length << ", ";
-  os << "win_length=" << win_length << ", ";
-  os << "window_type=\"" << window_type << "\", ";
-  os << "center=" << (center ? "True" : "False") << ", ";
-  os << "pad_mode=\"" << pad_mode << "\", ";
-  os << "normalized=" << (normalized ? "True" : "False") << ")";
-  return os.str();
-}
-
-class Stft::Impl {
- public:
-  explicit Impl(const StftConfig &config) : config_(config) {
-    if (!config.window.empty()) {
-      window_ = std::make_unique<FeatureWindowFunction>(config.window);
-    } else if (!config.window_type.empty()) {
-      window_ = std::make_unique<FeatureWindowFunction>(config.window_type,
-                                                        config.win_length);
-    }
-  }
-
-  StftResult Compute(const float *data, int32_t n) const {
-    int32_t n_fft = config_.n_fft;
-    int32_t hop_length = config_.hop_length;
-
-    std::vector<float> samples;
-    const float *p = data;
-
-    if (config_.center) {
-      samples = Pad(data, n);
-      p = samples.data();
-      n = samples.size();
-    }
-
-    int64_t num_frames = 1 + (n - n_fft) / hop_length;
-
-    Rfft rfft(n_fft);
-
-    StftResult ans;
-    ans.num_frames = num_frames;
-    ans.real.resize(num_frames * (n_fft / 2 + 1));
-    ans.imag.resize(num_frames * (n_fft / 2 + 1));
-
-    std::vector<float> tmp(config_.n_fft);
-
-    for (int32_t i = 0; i < num_frames; ++i) {
-      tmp = {p + i * hop_length, p + i * hop_length + n_fft};
-      if (window_) {
-        window_->Apply(tmp.data());
-      }
-
-      rfft.Compute(tmp.data());
-
-      for (int32_t k = 0; k < n_fft / 2; ++k) {
-        if (k == 0) {
-          ans.real[i * (n_fft / 2 + 1)] = tmp[0];
-          ans.real[i * (n_fft / 2 + 1) + n_fft / 2] = tmp[1];
-        } else {
-          ans.real[i * (n_fft / 2 + 1) + k] = tmp[2 * k];
-
-          ans.imag[i * (n_fft / 2 + 1) + k] = tmp[2 * k + 1];
-        }
-      }
-    }
-
-    if (config_.normalized) {
-      float scale = 1 / std::sqrt(n_fft);
-      for (int32_t i = 0; i < ans.real.size(); ++i) {
-        ans.real[i] *= scale;
-        ans.imag[i] *= scale;
-      }
-    }
-
-    return ans;
-  }
-
-  std::vector<float> Pad(const float *data, int32_t n) const {
-    int32_t pad_amount = config_.n_fft / 2;
-    std::vector<float> ans(n + config_.n_fft);
-    std::copy(data, data + n, ans.begin() + pad_amount);
-    if (config_.pad_mode == "constant") {
-      // do nothing
-    } else if (config_.pad_mode == "reflect") {
-      // left
-      std::copy(data + 1, data + 1 + pad_amount, ans.rend() - pad_amount);
-      std::copy(data + n - pad_amount - 1, data + n - 1, ans.rbegin());
-    } else if (config_.pad_mode == "replicate") {
-      std::fill(ans.begin(), ans.begin() + pad_amount, data[0]);
-      std::fill(ans.end() - pad_amount, ans.end(), data[n - 1]);
-    } else {
-      fprintf(stderr, "Unsupported pad_mode: '%s'. Use 0 padding\n",
-              config_.pad_mode.c_str());
-    }
-
-    return ans;
-  }
-
- private:
-  StftConfig config_;
-  std::unique_ptr<FeatureWindowFunction> window_;
-};
-
-Stft::Stft(const StftConfig &config) : impl_(std::make_unique<Impl>(config)) {}
-
-Stft::~Stft() = default;
-
-StftResult Stft::Compute(const float *data, int32_t n) const {
-  return impl_->Compute(data, n);
-}
-
-}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.h
deleted file mode 100644
index 6c5d2f7b..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/stft.h
+++ /dev/null
@@ -1,65 +0,0 @@
-/**
- * Copyright (c)  2025  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef KALDI_NATIVE_FBANK_CSRC_STFT_H_
-#define KALDI_NATIVE_FBANK_CSRC_STFT_H_
-
-#include <cstdint>
-#include <memory>
-#include <string>
-#include <vector>
-
-namespace knf {
-
-struct StftConfig {
-  int32_t n_fft;  // should be a power of two
-  int32_t hop_length;
-  int32_t win_length;
-  std::string window_type;
-  bool center = true;
-  std::string pad_mode = "reflect";
-  bool normalized = false;
-
-  // if it is specified, then window_type is ignored
-  std::vector<float> window;
-
-  std::string ToString() const;
-};
-
-struct StftResult {
-  // [num_frames, n_fft/2+1], flattened in row major
-  std::vector<float> real;
-  std::vector<float> imag;
-  int32_t num_frames;
-};
-
-class Stft {
- public:
-  explicit Stft(const StftConfig &config);
-  ~Stft();
-  StftResult Compute(const float *data, int32_t n) const;
-
- private:
-  class Impl;
-  std::unique_ptr<Impl> impl_;
-};
-
-}  // namespace knf
-
-#endif  // KALDI_NATIVE_FBANK_CSRC_STFT_H_
-        //
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-log.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-log.cc
deleted file mode 100644
index 1e6ef734..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-log.cc
+++ /dev/null
@@ -1,71 +0,0 @@
-/**
- * Copyright      2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#if KNF_ENABLE_CHECK
-#include "gtest/gtest.h"
-#include "kaldi-native-fbank/csrc/log.h"
-
-namespace knf {
-
-TEST(Log, TestLog) {
-  KNF_LOG(TRACE) << "this is a trace message";
-  KNF_LOG(DEBUG) << "this is a debug message";
-  KNF_LOG(INFO) << "this is an info message";
-  KNF_LOG(WARNING) << "this is a warning message";
-  KNF_LOG(ERROR) << "this is an error message";
-
-  ASSERT_THROW(KNF_LOG(FATAL) << "This will crash the program",
-               std::runtime_error);
-
-  // For debug build
-
-  KNF_DLOG(TRACE) << "this is a trace message for debug build";
-  KNF_DLOG(DEBUG) << "this is a trace message for debug build";
-  KNF_DLOG(INFO) << "this is a trace message for debug build";
-  KNF_DLOG(ERROR) << "this is an error message for debug build";
-  KNF_DLOG(WARNING) << "this is a trace message for debug build";
-
-#if !defined(NDEBUG)
-  ASSERT_THROW(KNF_DLOG(FATAL) << "this is a trace message for debug build",
-               std::runtime_error);
-#endif
-}
-
-TEST(Log, TestCheck) {
-  KNF_CHECK_EQ(1, 1) << "ok";
-  KNF_CHECK_LE(1, 3) << "ok";
-  KNF_CHECK_LT(1, 2) << "ok";
-  KNF_CHECK_GT(2, 1) << "ok";
-  KNF_CHECK_GE(2, 1) << "ok";
-
-  ASSERT_THROW(KNF_CHECK_EQ(2, 1) << "bad things happened", std::runtime_error);
-
-  // for debug build
-  KNF_DCHECK_EQ(1, 1) << "ok";
-  KNF_DCHECK_LE(1, 3) << "ok";
-  KNF_DCHECK_LT(1, 2) << "ok";
-  KNF_DCHECK_GT(2, 1) << "ok";
-  KNF_DCHECK_GE(2, 1) << "ok";
-
-#if !defined(NDEBUG)
-  ASSERT_THROW(KNF_CHECK_EQ(2, 1) << "bad things happened", std::runtime_error);
-#endif
-}
-
-}  // namespace knf
-#endif
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-fbank.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-fbank.cc
deleted file mode 100644
index c2d437a4..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-fbank.cc
+++ /dev/null
@@ -1,54 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include <iostream>
-#include <cstdint>
-#include <sstream>
-#include <string>
-
-#include "kaldi-native-fbank/csrc/online-feature.h"
-#include "kaldi-native-fbank/csrc/feature-fbank.h"
-#include "kaldi-native-fbank/csrc/feature-window.h"
-#include "kaldi-native-fbank/csrc/mel-computations.h"
-
-int main() {
-  knf::FbankOptions opts;
-  opts.frame_opts.dither = 0;
-  opts.mel_opts.num_bins = 10;
-
-  knf::OnlineFbank fbank(opts);
-  for (int32_t i = 0; i < 1600; ++i) {
-    float s = (i * i - i / 2) / 32767.;
-    fbank.AcceptWaveform(16000, &s, 1);
-  }
-
-  std::ostringstream os;
-
-  int32_t n = fbank.NumFramesReady();
-  for (int32_t i = 0; i != n; ++i) {
-    const float *frame = fbank.GetFrame(i);
-    for (int32_t k = 0; k != opts.mel_opts.num_bins; ++k) {
-      os << frame[k] << ", ";
-    }
-    os << "\n";
-  }
-
-  std::cout << os.str() << "\n";
-
-  return 0;
-}
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-feature.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-feature.cc
deleted file mode 100644
index bfbe6210..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-online-feature.cc
+++ /dev/null
@@ -1,59 +0,0 @@
-/**
- * Copyright (c)  2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include "gtest/gtest.h"
-#include "kaldi-native-fbank/csrc/online-feature.h"
-namespace knf {
-
-TEST(RecyclingVector, TestUnlimited) {
-  RecyclingVector v(-1);
-  constexpr int32_t N = 100;
-  for (int32_t i = 0; i != N; ++i) {
-    std::unique_ptr<float[]> p(new float[3]{i, i + 1, i + 2});
-    v.PushBack(std::move(p));
-  }
-  ASSERT_EQ(v.Size(), N);
-
-  for (int32_t i = 0; i != N; ++i) {
-    const float *t = v.At(i);
-    for (int32_t k = 0; k != 3; ++k) {
-      EXPECT_EQ(t[k], (i + k));
-    }
-  }
-}
-
-TEST(RecyclingVector, Testlimited) {
-  constexpr int32_t K = 3;
-  constexpr int32_t N = 10;
-  RecyclingVector v(K);
-  for (int32_t i = 0; i != N; ++i) {
-    std::unique_ptr<float[]> p(new float[3]{i, i + 1, i + 2});
-    v.PushBack(std::move(p));
-  }
-
-  ASSERT_EQ(v.Size(), N);
-
-  for (int32_t i = N - K; i != N; ++i) {
-    const float *t = v.At(i);
-
-    for (int32_t k = 0; k != 3; ++k) {
-      EXPECT_EQ(t[k], (i + k));
-    }
-  }
-}
-}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-rfft.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-rfft.cc
deleted file mode 100644
index e1763e4e..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/test-rfft.cc
+++ /dev/null
@@ -1,101 +0,0 @@
-/**
- * Copyright      2022  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include <cstdint>
-#include <vector>
-
-#include "gtest/gtest.h"
-#include "kaldi-native-fbank/csrc/rfft.h"
-
-namespace knf {
-
-#if 0
->>> import torch
->>> a = torch.tensor([1., -1, 3, 8, 20, 6, 0, 2])
->>> torch.fft.rfft(a)
-tensor([ 39.0000+0.0000j, -28.1924-2.2929j,  18.0000+5.0000j,  -9.8076+3.7071j,
-          9.0000+0.0000j])
-#endif
-
-TEST(PowerOfTwo, TestRfft) {
-  knf::Rfft fft(8);
-  std::vector<float> original = {1, -1, 3, 8, 20, 6, 0, 2};
-
-  std::vector<float> d = original;
-  fft.Compute(d.data());
-
-  EXPECT_EQ(d[0], 39);
-  EXPECT_EQ(d[1], 9);
-
-  EXPECT_NEAR(d[2], -28.1924, 1e-3);
-  EXPECT_NEAR(d[3], -2.2929, 1e-3);
-
-  EXPECT_NEAR(d[4], 18, 1e-3);
-  EXPECT_NEAR(d[5], 5, 1e-3);
-
-  EXPECT_NEAR(d[6], -9.8076, 1e-3);
-  EXPECT_NEAR(d[7], 3.7071, 1e-3);
-
-  knf::Rfft ifft(8, true);
-  ifft.Compute(d.data());
-
-  for (int32_t i = 0; i < d.size(); ++i) {
-    d[i] /= d.size();  // we need to rescale it by 1/n
-    EXPECT_EQ(d[i], original[i]);
-  }
-}
-
-#if 0
->>> import torch
->>> a = torch.tensor([1., -1, 3, 8, 20, 6, 0, 2, 9, 5])
->>> torch.fft.rfft(a)
-tensor([ 53.0000+0.0000j, -17.3262-8.2290j,  -3.3820+31.7809j,
-         -1.6738-13.3148j,  -5.6180+3.8697j,  13.0000+0.0000j])
-#endif
-TEST(NotPowerOfTwo, TestRfft) {
-  knf::Rfft fft(10);
-  std::vector<float> original = {1, -1, 3, 8, 20, 6, 0, 2, 9, 5};
-
-  std::vector<float> d = original;
-  fft.Compute(d.data());
-
-  EXPECT_EQ(d[0], 53);
-  EXPECT_EQ(d[1], 13);
-
-  EXPECT_NEAR(d[2], -17.3262, 1e-3);
-  EXPECT_NEAR(d[3], -8.2290, 1e-3);
-
-  EXPECT_NEAR(d[4], -3.3820, 1e-3);
-  EXPECT_NEAR(d[5], 31.7809, 1e-3);
-
-  EXPECT_NEAR(d[6], -1.6738, 1e-3);
-  EXPECT_NEAR(d[7], -13.3148, 1e-3);
-
-  EXPECT_NEAR(d[8], -5.6180, 1e-3);
-  EXPECT_NEAR(d[9], 3.8697, 1e-3);
-
-  knf::Rfft ifft(10, true);
-  ifft.Compute(d.data());
-
-  for (int32_t i = 0; i < d.size(); ++i) {
-    d[i] /= d.size();  // we need to rescale it by 1/n
-    EXPECT_NEAR(d[i], original[i], 1e-5);
-  }
-}
-
-}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.cc b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.cc
deleted file mode 100644
index 1b699846..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.cc
+++ /dev/null
@@ -1,169 +0,0 @@
-/**
- * Copyright (c)  2023  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include "kaldi-native-fbank/csrc/whisper-feature.h"
-
-#include <cmath>
-#include <string>
-#include <vector>
-
-#include "kaldi-native-fbank/csrc/log.h"
-#include "kaldi-native-fbank/csrc/mel-computations.h"
-
-#ifndef M_2PI
-#define M_2PI 6.283185307179586476925286766559005
-#endif
-
-namespace knf {
-
-std::string WhisperFeatureOptions::ToString() const {
-  std::ostringstream os;
-  os << "WhisperFeatureOptions(";
-  os << "frame_opts=" << frame_opts.ToString() << ", ";
-  os << "dim=" << dim << ")";
-  return os.str();
-}
-
-static void dft(const std::vector<float> &in, std::vector<float> *out) {
-  // this function is modified from
-  // https://github.com/ggerganov/whisper.cpp/blob/master/whisper.cpp#L2353
-  int32_t N = in.size();
-
-  out->resize(N * 2);
-
-  auto M_2PI_over_N = M_2PI / N;
-  for (int32_t k = 0; k < N; ++k) {
-    float re = 0;
-    float im = 0;
-
-    for (int32_t n = 0; n < N; ++n) {
-      float angle = M_2PI_over_N * k * n;
-      re += in[n] * cos(angle);
-      im -= in[n] * sin(angle);
-    }
-
-    (*out)[k * 2 + 0] = re;
-    (*out)[k * 2 + 1] = im;
-  }
-}
-
-// Cooley-Tukey FFT
-// poor man's implementation - use something better
-// input is real-valued
-// output is complex-valued
-static void fft(const std::vector<float> &in, std::vector<float> *out) {
-  // this function is copied from
-  // https://github.com/ggerganov/whisper.cpp/blob/master/whisper.cpp#L2373C1-L2429C1
-
-  int32_t N = in.size();
-  out->resize(N * 2);
-
-  if (N == 1) {
-    (*out)[0] = in[0];
-    (*out)[1] = 0;
-    return;
-  }
-
-  if (N % 2 == 1) {
-    dft(in, out);
-    return;
-  }
-
-  std::vector<float> even;
-  std::vector<float> odd;
-
-  even.reserve(N / 2);
-  odd.reserve(N / 2);
-
-  for (int32_t i = 0; i != N; ++i) {
-    if (i % 2 == 0) {
-      even.push_back(in[i]);
-    } else {
-      odd.push_back(in[i]);
-    }
-  }
-
-  std::vector<float> even_fft;
-  std::vector<float> odd_fft;
-
-  fft(even, &even_fft);
-  fft(odd, &odd_fft);
-
-  for (int32_t k = 0; k < N / 2; ++k) {
-    float theta = M_2PI * k / N;
-
-    float re = cos(theta);
-    float im = -sin(theta);
-
-    float re_odd = odd_fft[2 * k + 0];
-    float im_odd = odd_fft[2 * k + 1];
-
-    (*out)[2 * k + 0] = even_fft[2 * k + 0] + re * re_odd - im * im_odd;
-    (*out)[2 * k + 1] = even_fft[2 * k + 1] + re * im_odd + im * re_odd;
-
-    (*out)[2 * (k + N / 2) + 0] =
-        even_fft[2 * k + 0] - re * re_odd + im * im_odd;
-    (*out)[2 * (k + N / 2) + 1] =
-        even_fft[2 * k + 1] - re * im_odd - im * re_odd;
-  }
-}
-
-WhisperFeatureComputer::WhisperFeatureComputer(
-    const WhisperFeatureOptions &opts /*= {}*/)
-    : opts_(opts) {
-  opts_.frame_opts.samp_freq = 16000;
-  opts_.frame_opts.frame_shift_ms = 10;
-  opts_.frame_opts.frame_length_ms = 25;
-  opts_.frame_opts.dither = 0;
-  opts_.frame_opts.preemph_coeff = 0;
-  opts_.frame_opts.remove_dc_offset = false;
-  opts_.frame_opts.window_type = "hann";
-  opts_.frame_opts.round_to_power_of_two = false;
-  opts_.frame_opts.snip_edges = false;
-
-  MelBanksOptions mel_opts;
-  mel_opts.num_bins = opts_.dim;
-  mel_opts.low_freq = 0;
-  mel_opts.is_librosa = true;
-
-  mel_banks_ = std::make_unique<MelBanks>(mel_opts, opts_.frame_opts, 1.0f);
-}
-
-void WhisperFeatureComputer::Compute(float /*signal_raw_log_energy*/,
-                                     float /*vtln_warp*/,
-                                     std::vector<float> *signal_frame,
-                                     float *feature) {
-  KNF_CHECK_EQ(signal_frame->size(), frame_opts_.PaddedWindowSize());
-  // we have already applied window function to signal_frame before
-  // calling this method
-  std::vector<float> fft_out;
-  fft(*signal_frame, &fft_out);
-
-  int32_t num_fft = signal_frame->size();
-  std::vector<float> power(num_fft / 2 + 1);
-  for (int32_t i = 0; i <= num_fft / 2; ++i) {
-    float re = fft_out[2 * i + 0];
-    float im = fft_out[2 * i + 1];
-    power[i] = re * re + im * im;
-  }
-
-  // feature is pre-allocated by the user
-  mel_banks_->Compute(power.data(), feature);
-}
-
-}  // namespace knf
diff --git a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.h b/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.h
deleted file mode 100644
index 57b9cc38..00000000
--- a/projects/llm_framework/main_kws_new/src/runner/kaldi-native-fbank/csrc/whisper-feature.h
+++ /dev/null
@@ -1,69 +0,0 @@
-/**
- * Copyright (c)  2023  Xiaomi Corporation (authors: Fangjun Kuang)
- *
- * See LICENSE for clarification regarding multiple authors
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef KALDI_NATIVE_FBANK_CSRC_WHISPER_FEATURE_H_
-#define KALDI_NATIVE_FBANK_CSRC_WHISPER_FEATURE_H_
-
-#include <cstdint>
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "kaldi-native-fbank/csrc/feature-window.h"
-#include "kaldi-native-fbank/csrc/mel-computations.h"
-
-namespace knf {
-
-struct WhisperFeatureOptions {
-  WhisperFeatureOptions(const FrameExtractionOptions &frame_opts = {},
-                        int32_t dim = 80)
-      : frame_opts(frame_opts), dim(dim) {}
-
-  FrameExtractionOptions frame_opts;
-  int32_t dim = 80;
-
-  std::string ToString() const;
-};
-
-class WhisperFeatureComputer {
- public:
-  // note: opts.frame_opts is ignored and we reset it inside
-  explicit WhisperFeatureComputer(const WhisperFeatureOptions &opts = {});
-
-  int32_t Dim() const { return opts_.dim; }
-
-  const FrameExtractionOptions &GetFrameOptions() const {
-    return opts_.frame_opts;
-  }
-
-  void Compute(float /*signal_raw_log_energy*/, float /*vtln_warp*/,
-               std::vector<float> *signal_frame, float *feature);
-
-  // if true, compute log_energy_pre_window but after dithering and dc removal
-  bool NeedRawLogEnergy() const { return false; }
-
-  using Options = WhisperFeatureOptions;
-
- private:
-  std::unique_ptr<MelBanks> mel_banks_;
-  WhisperFeatureOptions opts_;
-};
-
-}  // namespace knf
-
-#endif  // KALDI_NATIVE_FBANK_CSRC_WHISPER_FEATURE_H_
diff --git a/projects/llm_framework/main_kws_new/test.py b/projects/llm_framework/main_kws_new/test.py
deleted file mode 100644
index b73f9e7e..00000000
--- a/projects/llm_framework/main_kws_new/test.py
+++ /dev/null
@@ -1,115 +0,0 @@
-import time
-import torch
-import torchaudio
-import torchaudio.compliance.kaldi as kaldi
-import numpy as np
-import onnxruntime as ort
-import os
-import glob
-
-frame_index_global = 0
-last_trigger_time_ms = -1e9
-count = 0  # 添加缺失的全局计数器
-
-def detect_wakeup(scores_chunk):
-    global count, last_trigger_time_ms, frame_index_global
-    above = scores_chunk > threshold
-    triggered = False
-    for flag in above:
-        if flag:
-            count += 1
-            if count >= min_continuous_frames:
-                trigger_time_ms = (frame_index_global - min_continuous_frames + 1) * 10
-                if (trigger_time_ms - last_trigger_time_ms) >= 2000:
-                    print(f"[唤醒触发] 时间: {trigger_time_ms} ms")
-                    last_trigger_time_ms = trigger_time_ms
-                    triggered = True
-        else:
-            count = 0
-        frame_index_global += 1
-    return triggered
-
-def save_mat_to_bin(mat_data, index):
-    """保存mat数据为二进制文件，格式与C++兼容"""
-    filename = f"mat_{index:03d}.bin"
-    
-    # mat_data是torch tensor，shape为 [frames, features]
-    mat_np = mat_data.numpy()
-    rows, cols = mat_np.shape
-    
-    with open(filename, 'wb') as f:
-        # 写入维度信息 (int64)
-        f.write(np.array([rows], dtype=np.int64).tobytes())
-        f.write(np.array([cols], dtype=np.int64).tobytes())
-        # 写入数据 (float32)
-        f.write(mat_np.astype(np.float32).tobytes())
-    
-    print(f"保存mat到 {filename}, shape: ({rows}, {cols})")
-    return filename
-
-# ================= 配置参数 =================
-wavfile = "./001.wav"                  # 测试音频路径
-onnx_model = "./t30.onnx"          # ONNX 模型路径
-feat_dim = 80                          # FBank 维度
-resample_rate = 16000                  # 目标采样率
-chunk_size = 32                        # 推理帧数（1 帧=10ms，32帧≈320ms）
-threshold = 0.9                        # 唤醒分数阈值
-min_continuous_frames = 5              # 连续高分帧数(>=此值判定唤醒)
-device = "cpu"
-
-# ================= 加载模型 =================
-ort_sess = ort.InferenceSession(onnx_model)
-
-# ================= 加载音频并预处理 =================
-waveform, sample_rate = torchaudio.load(wavfile)
-# 音量归一
-waveform = waveform * (1 << 15)
-
-cache = torch.zeros(1, 32, 88, dtype=torch.float).numpy()  # 初始化 cache
-scores = []          # 存储每帧分数
-frame_index = 0      # 已处理全局帧数
-save_index = 0
-save_dir = "debug_inputs"
-os.makedirs(save_dir, exist_ok=True)
-
-samples_per_chunk = int(resample_rate * 0.01 * chunk_size)  # 每个推理块采样点数
-total_samples = waveform.shape[1]
-print(f"音频总采样点数: {total_samples}, 每块点数: {samples_per_chunk}")
-
-start_sample = 0
-while start_sample < total_samples:
-    end_sample = start_sample + samples_per_chunk
-    audio_chunk = waveform[:, start_sample:end_sample]  # [1, T]
-    start_sample += samples_per_chunk
-    frame_index += chunk_size  # 累加帧数
-    
-    # ----------- run_inference1 逻辑的 fbank 计算 -----------
-    mat = kaldi.fbank(
-        audio_chunk,
-        num_mel_bins=feat_dim,
-        dither=0.0,
-        energy_floor=0.0,
-        sample_frequency=resample_rate,
-        snip_edges=False
-    )
-    
-    print(f"mat.shape={mat.shape}")
-    
-    # 保存mat为二进制文件
-    save_mat_to_bin(mat, save_index)
-    
-    # 推理
-    onnx_inputs = {
-        "input": mat.unsqueeze(0).numpy(),
-        "cache": cache
-    }
-    onnx_output = ort_sess.run(None, onnx_inputs)
-    out_chunk, cache = onnx_output[0].flatten(), onnx_output[1]  # 更新 cache
-    
-    if detect_wakeup(out_chunk):
-        print(f"[{time.strftime('%H:%M:%S')}] 检测到唤醒词！")
-    
-    scores.extend(out_chunk)
-    save_index += 1
-
-print(f"总共保存了 {save_index} 个mat文件")

From 27a16a44f08de28f1386ce7df98c2c7aa5014870 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Sun, 28 Sep 2025 16:41:43 +0800
Subject: [PATCH 46/79] [update] llm-openai-version fix kws

---
 projects/llm_framework/main_kws_new/SConstruct    | 2 +-
 projects/llm_framework/main_openai_api/SConstruct | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/projects/llm_framework/main_kws_new/SConstruct b/projects/llm_framework/main_kws_new/SConstruct
index 2e327d95..2458e81e 100644
--- a/projects/llm_framework/main_kws_new/SConstruct
+++ b/projects/llm_framework/main_kws_new/SConstruct
@@ -23,7 +23,7 @@ LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 
 INCLUDE += [
-            ADir('../static_lib/include/sherpa/'),
+            ADir('../static_lib/include/sherpa/fbank'),
             ADir('../static_lib/include/sherpa/sherpa-onnx/onnxruntime-src'),
             ]
 
diff --git a/projects/llm_framework/main_openai_api/SConstruct b/projects/llm_framework/main_openai_api/SConstruct
index 9098a679..d2851e73 100644
--- a/projects/llm_framework/main_openai_api/SConstruct
+++ b/projects/llm_framework/main_openai_api/SConstruct
@@ -18,7 +18,7 @@ LINK_SEARCH_PATH = []
 STATIC_FILES = []
 
 
-ModuleLLMOpenAIPluginPath = wget_github_commit('https://github.com/m5stack/ModuleLLM-OpenAI-Plugin.git', '9612b20800eb1708e648744e3cb333c4f743811c', True)
+ModuleLLMOpenAIPluginPath = wget_github_commit('https://github.com/m5stack/ModuleLLM-OpenAI-Plugin.git', '748ec04e6dde31e6283921e9635204b4b8f8ede2', True)
 python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-openai-api-python-venv_v1.6.tar.gz", 'm5stack_llm-openai-api-python-venv_v1.6.tar.gz')
 
 

From 7a97143aad0f7f4215f236014e04be9feecfea39 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Mon, 29 Sep 2025 15:15:00 +0800
Subject: [PATCH 47/79] [update] update sdk version & chip name

---
 ext_components/ax_msp/Kconfig                             | 8 ++++----
 ext_components/ax_msp/SConstruct                          | 4 ++--
 .../main_depth_anything/src/EngineWrapper.cpp             | 4 ++--
 .../main_melotts/src/runner/EngineWrapper.cpp             | 4 ++--
 .../main_whisper/src/runner/EngineWrapper.cpp             | 4 ++--
 projects/llm_framework/main_yolo/src/EngineWrapper.cpp    | 4 ++--
 6 files changed, 14 insertions(+), 14 deletions(-)

diff --git a/ext_components/ax_msp/Kconfig b/ext_components/ax_msp/Kconfig
index 73824242..ce8206e1 100644
--- a/ext_components/ax_msp/Kconfig
+++ b/ext_components/ax_msp/Kconfig
@@ -12,8 +12,8 @@ menuconfig AX_MSP_ENABLED
         bool "enable ax620e bsp"
         help
             build by AXERA!
-    config AX_650N_MSP_ENABLED
-        bool "enable ax650n bsp"
+    config AX_650C_MSP_ENABLED
+        bool "enable ax650C bsp"
         help
             build by AXERA!
     config AX_520_MSP_ENABLED
@@ -26,8 +26,8 @@ menuconfig AX_MSP_ENABLED
         prompt "choice common version"
         default MSP_V2_0_0_ENABLED
         depends on AX_MSP_ENABLED
-    config MSP_V3_0_0_ENABLED
-        bool "enable v3.0.0 bsp"
+    config MSP_V3_6_2_ENABLED
+        bool "enable v3.6.2 bsp"
         help
             build by m5stack!
     config MSP_V2_0_0_ENABLED
diff --git a/ext_components/ax_msp/SConstruct b/ext_components/ax_msp/SConstruct
index c4e1ad0d..696c5655 100644
--- a/ext_components/ax_msp/SConstruct
+++ b/ext_components/ax_msp/SConstruct
@@ -55,9 +55,9 @@ if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
                             'REGISTER':'static'
                             })
 
-if 'CONFIG_AX_650N_MSP_ENABLED' in os.environ:
+if 'CONFIG_AX_650C_MSP_ENABLED' in os.environ:
     MSP_PATH = ''
-    if 'CONFIG_MSP_V3_0_0_ENABLED' in os.environ:
+    if 'CONFIG_MSP_V3_6_2_ENABLED' in os.environ:
         MSP_PATH = os.environ.get('EXT_MSP_PATH', check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_ax650n_msp.tar.gz", 'm5stack_ax650n_msp.tar.gz'))
     
     SRCS=[]
diff --git a/projects/llm_framework/main_depth_anything/src/EngineWrapper.cpp b/projects/llm_framework/main_depth_anything/src/EngineWrapper.cpp
index 1aad9678..ab79dcf4 100644
--- a/projects/llm_framework/main_depth_anything/src/EngineWrapper.cpp
+++ b/projects/llm_framework/main_depth_anything/src/EngineWrapper.cpp
@@ -14,7 +14,7 @@
 
 #include <global_config.h>
 
-#if defined(CONFIG_AX_650N_MSP_ENABLED)
+#if defined(CONFIG_AX_650C_MSP_ENABLED)
 static const char *strAlgoModelType[AX_ENGINE_MODEL_TYPE_BUTT] = {"3.6T", "7.2T", "18T"};
 #endif
 
@@ -32,7 +32,7 @@ typedef enum axNPU_TYPE_E {
     AX_BL_VNPU_2   = (1 << 4)  /* running under BIG-LITTLE VNPU2 */
 } AX_NPU_TYPE_E;
 
-#if defined(CONFIG_AX_650N_MSP_ENABLED)
+#if defined(CONFIG_AX_650C_MSP_ENABLED)
 static AX_S32 CheckModelVNpu(const std::string &strModel, const AX_ENGINE_MODEL_TYPE_T &eModelType,
                              const AX_S32 &nNpuType, AX_U32 &nNpuSet)
 {
diff --git a/projects/llm_framework/main_melotts/src/runner/EngineWrapper.cpp b/projects/llm_framework/main_melotts/src/runner/EngineWrapper.cpp
index bc7590af..3ddc9dd2 100644
--- a/projects/llm_framework/main_melotts/src/runner/EngineWrapper.cpp
+++ b/projects/llm_framework/main_melotts/src/runner/EngineWrapper.cpp
@@ -13,7 +13,7 @@
 #include <cstdlib>
 #include <global_config.h>
 
-#if defined(CONFIG_AX_650N_MSP_ENABLED)
+#if defined(CONFIG_AX_650C_MSP_ENABLED)
 static const char *strAlgoModelType[AX_ENGINE_MODEL_TYPE_BUTT] = {"3.6T", "7.2T", "18T"};
 #endif
 
@@ -32,7 +32,7 @@ typedef enum axNPU_TYPE_E {
     AX_BL_VNPU_2 = (1 << 4)    /* running under BIG-LITTLE VNPU2 */
 } AX_NPU_TYPE_E;
 
-#if defined(CONFIG_AX_650N_MSP_ENABLED)
+#if defined(CONFIG_AX_650C_MSP_ENABLED)
 static AX_S32 CheckModelVNpu(const std::string &strModel, const AX_ENGINE_MODEL_TYPE_T &eModelType, const AX_S32 &nNpuType, AX_U32 &nNpuSet) {
     AX_ENGINE_NPU_ATTR_T stNpuAttr;
     memset(&stNpuAttr, 0x00, sizeof(stNpuAttr));
diff --git a/projects/llm_framework/main_whisper/src/runner/EngineWrapper.cpp b/projects/llm_framework/main_whisper/src/runner/EngineWrapper.cpp
index bb0e7f8c..e9574be6 100644
--- a/projects/llm_framework/main_whisper/src/runner/EngineWrapper.cpp
+++ b/projects/llm_framework/main_whisper/src/runner/EngineWrapper.cpp
@@ -14,7 +14,7 @@
 
 #include <global_config.h>
 
-#if defined(CONFIG_AX_650N_MSP_ENABLED)
+#if defined(CONFIG_AX_650C_MSP_ENABLED)
 static const char *strAlgoModelType[AX_ENGINE_MODEL_TYPE_BUTT] = {"3.6T", "7.2T", "18T"};
 #endif
 
@@ -33,7 +33,7 @@ typedef enum axNPU_TYPE_E {
     AX_BL_VNPU_2 = (1 << 4)    /* running under BIG-LITTLE VNPU2 */
 } AX_NPU_TYPE_E;
 
-#if defined(CONFIG_AX_650N_MSP_ENABLED)
+#if defined(CONFIG_AX_650C_MSP_ENABLED)
 static AX_S32 CheckModelVNpu(const std::string &strModel, const AX_ENGINE_MODEL_TYPE_T &eModelType, const AX_S32 &nNpuType, AX_U32 &nNpuSet) {
     AX_ENGINE_NPU_ATTR_T stNpuAttr;
     memset(&stNpuAttr, 0x00, sizeof(stNpuAttr));
diff --git a/projects/llm_framework/main_yolo/src/EngineWrapper.cpp b/projects/llm_framework/main_yolo/src/EngineWrapper.cpp
index 72f16a6f..7b5b97d3 100644
--- a/projects/llm_framework/main_yolo/src/EngineWrapper.cpp
+++ b/projects/llm_framework/main_yolo/src/EngineWrapper.cpp
@@ -15,7 +15,7 @@
 
 #include <global_config.h>
 
-#if defined(CONFIG_AX_650N_MSP_ENABLED)
+#if defined(CONFIG_AX_650C_MSP_ENABLED)
 static const char* strAlgoModelType[AX_ENGINE_MODEL_TYPE_BUTT] = {"3.6T", "7.2T", "18T"};
 #endif
 
@@ -33,7 +33,7 @@ typedef enum axNPU_TYPE_E {
     AX_BL_VNPU_2   = (1 << 4)  /* running under BIG-LITTLE VNPU2 */
 } AX_NPU_TYPE_E;
 
-#if defined(CONFIG_AX_650N_MSP_ENABLED)
+#if defined(CONFIG_AX_650C_MSP_ENABLED)
 static AX_S32 CheckModelVNpu(const std::string& strModel, const AX_ENGINE_MODEL_TYPE_T& eModelType,
                              const AX_S32& nNpuType, AX_U32& nNpuSet)
 {

From 0e14999465285fd9ccfebb989f9926a5ccc2d384 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 16 Oct 2025 17:56:56 +0800
Subject: [PATCH 48/79] [fix] Fix inference issues caused by memory
 synchronization

---
 .../ax_model_runner/ax_model_runner_ax650.cpp   | 17 +++++++++++++++--
 1 file changed, 15 insertions(+), 2 deletions(-)

diff --git a/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner_ax650.cpp b/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner_ax650.cpp
index 94d0b2fb..374261b6 100644
--- a/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner_ax650.cpp
+++ b/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner_ax650.cpp
@@ -432,10 +432,23 @@ void ax_runner_ax650::deinit()
 
 int ax_runner_ax650::inference()
 {
-    return AX_ENGINE_RunSync(m_handle->handle, &m_handle->io_data[0]);
+    int ret = AX_ENGINE_RunSync(m_handle->handle, &m_handle->io_data[0]);
+    for (size_t i = 0; i < get_num_outputs(); i++)
+    {
+        auto &tensor = get_output(i);
+        AX_SYS_MinvalidateCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
+    }
+    return ret;
 }
 
 int ax_runner_ax650::inference(int grpid)
 {
-    return AX_ENGINE_RunGroupIOSync(m_handle->handle, m_handle->context, grpid, &m_handle->io_data[grpid]);
+    int ret = AX_ENGINE_RunGroupIOSync(m_handle->handle, m_handle->context, grpid, &m_handle->io_data[grpid]);
+
+    for (size_t i = 0; i < get_num_outputs(); i++)
+    {
+        auto &tensor = get_output(grpid, i);
+        AX_SYS_MinvalidateCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
+    }
+    return ret;
 }

From f9de469ee541d323d5b90aa6eed278e014a73eb0 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Mon, 3 Nov 2025 11:18:32 +0800
Subject: [PATCH 49/79] [update] update CosyVoice2

---
 .../llm_framework/main_cosy_voice/SConstruct  |   6 +-
 .../models/mode_CosyVoice2-0.5B-ax650.json    |  21 +-
 .../scripts/tokenizer/tokenizer.py            | 230 ++-----
 .../main_cosy_voice/src/main.cpp              | 240 +++++--
 .../main_cosy_voice/src/runner/BaseRunner.cpp |  14 +
 .../main_cosy_voice/src/runner/BaseRunner.hpp |  37 ++
 .../main_cosy_voice/src/runner/LLM.hpp        |  58 +-
 .../src/runner/OnnxWarpper/OnnxWarpper.hpp    | 174 +++++
 .../main_cosy_voice/src/runner/Token2wav.hpp  | 619 +++++++-----------
 .../src/runner/utils/utils.hpp                |  47 +-
 .../main_cosy_voice/src/runner/utils/wav.hpp  | 106 ++-
 projects/llm_framework/tools/llm_pack.py      |  93 +--
 12 files changed, 871 insertions(+), 774 deletions(-)
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/BaseRunner.cpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/BaseRunner.hpp
 create mode 100644 projects/llm_framework/main_cosy_voice/src/runner/OnnxWarpper/OnnxWarpper.hpp

diff --git a/projects/llm_framework/main_cosy_voice/SConstruct b/projects/llm_framework/main_cosy_voice/SConstruct
index 76fae578..d1dd0e43 100644
--- a/projects/llm_framework/main_cosy_voice/SConstruct
+++ b/projects/llm_framework/main_cosy_voice/SConstruct
@@ -29,7 +29,8 @@ python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/reso
 DEFINITIONS += ['-O2']
 DEFINITIONS += ['-std=c++17']
 LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
-REQUIREMENTS += ['ax_engine', 'ax_interpreter', 'ax_sys', 'utilities', 'MNN']
+REQUIREMENTS += ['ax_engine', 'ax_interpreter', 'ax_sys', 'utilities']
+REQUIREMENTS += ['samplerate']
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 
 
@@ -37,13 +38,14 @@ INCLUDE += [ADir('src/runner'), ADir('src/runner/utils')]
 INCLUDE += [ADir('../static_lib/include/sentencepiece'),
             ADir('../static_lib/include/protobuf-lite'),
             ADir('../static_lib/include/abseil-cpp'),
-            ADir('../static_lib/include/mnn'),
+            ADir('../static_lib/include/onnxruntime/core/session'),
             ADir('../static_lib/include/re2')]
 
 static_file = Glob('../static_lib/module-llm/libabsl_*')
 static_file += [AFile('../static_lib/module-llm/libre2.a'), AFile('../static_lib/module-llm/libsentencepiece.a'), AFile('../static_lib/module-llm/libsentencepiece_train.a')]
 STATIC_LIB += static_file * 4
 
+REQUIREMENTS += ['onnxruntime']
 STATIC_FILES += [os.path.join(python_venv, 'cosy-voice')]
 STATIC_FILES += Glob('scripts/tokenizer*')
 STATIC_FILES += Glob('models/mode_*.json')
diff --git a/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json b/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json
index 546a1270..b8baa740 100644
--- a/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json
+++ b/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json
@@ -22,8 +22,25 @@
         "filename_post_axmodel": "qwen2_post.axmodel",
         "filename_decoder_axmodel": "llm_decoder.axmodel",
         "template_filename_axmodel": "qwen2_p128_l%d_together.axmodel",
-        "token2wav_axmodel_dir": "",
-        "prompt_files": "zh_man1",
+        "flow_input_embedding": "flow.input_embedding.float16.bin",
+        "rand_noise": "rand_noise_1_80_300.txt",
+        "speech_window": "speech_window_2x8x480.txt",
+        "flow_encoder_28": "flow_encoder_28.axmodel",
+        "flow_encoder_53": "flow_encoder_53.axmodel",
+        "flow_encoder_78": "flow_encoder_78.axmodel",
+        "flow_encoder_50_final": "flow_encoder_50_final.axmodel",
+        "flow_estimator_200": "flow_estimator_200.axmodel",
+        "flow_estimator_250": "flow_estimator_250.axmodel",
+        "flow_estimator_300": "flow_estimator_300.axmodel",
+        "hift_p2_50_first": "hift_p2_50_first.axmodel",
+        "hift_p2_58": "hift_p2_58.axmodel",
+        "hift_p1_50_first": "hift_p1_50_first.ort",
+        "hift_p1_58": "hift_p1_58.ort",
+        "prompt_dir": "prompt_data",
+        "prompt_text": "prompt_text.txt",
+        "llm_prompt_speech_token": "llm_prompt_speech_token.txt",
+        "prompt_speech_feat": "prompt_speech_feat.txt",
+        "flow_embedding": "flow_embedding.txt",
         "b_use_topk": false,
         "b_bos": false,
         "b_eos": false,
diff --git a/projects/llm_framework/main_cosy_voice/scripts/tokenizer/tokenizer.py b/projects/llm_framework/main_cosy_voice/scripts/tokenizer/tokenizer.py
index 43fb39a2..44bf18bc 100755
--- a/projects/llm_framework/main_cosy_voice/scripts/tokenizer/tokenizer.py
+++ b/projects/llm_framework/main_cosy_voice/scripts/tokenizer/tokenizer.py
@@ -4,168 +4,53 @@
 from typing import Optional
 import torch
 from transformers import AutoTokenizer
-from whisper.tokenizer import Tokenizer
-
 import tiktoken
 
 LANGUAGES = {
-    "en": "english",
-    "zh": "chinese",
-    "de": "german",
-    "es": "spanish",
-    "ru": "russian",
-    "ko": "korean",
-    "fr": "french",
-    "ja": "japanese",
-    "pt": "portuguese",
-    "tr": "turkish",
-    "pl": "polish",
-    "ca": "catalan",
-    "nl": "dutch",
-    "ar": "arabic",
-    "sv": "swedish",
-    "it": "italian",
-    "id": "indonesian",
-    "hi": "hindi",
-    "fi": "finnish",
-    "vi": "vietnamese",
-    "he": "hebrew",
-    "uk": "ukrainian",
-    "el": "greek",
-    "ms": "malay",
-    "cs": "czech",
-    "ro": "romanian",
-    "da": "danish",
-    "hu": "hungarian",
-    "ta": "tamil",
-    "no": "norwegian",
-    "th": "thai",
-    "ur": "urdu",
-    "hr": "croatian",
-    "bg": "bulgarian",
-    "lt": "lithuanian",
-    "la": "latin",
-    "mi": "maori",
-    "ml": "malayalam",
-    "cy": "welsh",
-    "sk": "slovak",
-    "te": "telugu",
-    "fa": "persian",
-    "lv": "latvian",
-    "bn": "bengali",
-    "sr": "serbian",
-    "az": "azerbaijani",
-    "sl": "slovenian",
-    "kn": "kannada",
-    "et": "estonian",
-    "mk": "macedonian",
-    "br": "breton",
-    "eu": "basque",
-    "is": "icelandic",
-    "hy": "armenian",
-    "ne": "nepali",
-    "mn": "mongolian",
-    "bs": "bosnian",
-    "kk": "kazakh",
-    "sq": "albanian",
-    "sw": "swahili",
-    "gl": "galician",
-    "mr": "marathi",
-    "pa": "punjabi",
-    "si": "sinhala",
-    "km": "khmer",
-    "sn": "shona",
-    "yo": "yoruba",
-    "so": "somali",
-    "af": "afrikaans",
-    "oc": "occitan",
-    "ka": "georgian",
-    "be": "belarusian",
-    "tg": "tajik",
-    "sd": "sindhi",
-    "gu": "gujarati",
-    "am": "amharic",
-    "yi": "yiddish",
-    "lo": "lao",
-    "uz": "uzbek",
-    "fo": "faroese",
-    "ht": "haitian creole",
-    "ps": "pashto",
-    "tk": "turkmen",
-    "nn": "nynorsk",
-    "mt": "maltese",
-    "sa": "sanskrit",
-    "lb": "luxembourgish",
-    "my": "myanmar",
-    "bo": "tibetan",
-    "tl": "tagalog",
-    "mg": "malagasy",
-    "as": "assamese",
-    "tt": "tatar",
-    "haw": "hawaiian",
-    "ln": "lingala",
-    "ha": "hausa",
-    "ba": "bashkir",
-    "jw": "javanese",
-    "su": "sundanese",
-    "yue": "cantonese",
-    "minnan": "minnan",
-    "wuyu": "wuyu",
-    "dialect": "dialect",
-    "zh/en": "zh/en",
-    "en/zh": "en/zh",
+    "en": "english", "zh": "chinese", "de": "german", "es": "spanish", "ru": "russian",
+    "ko": "korean", "fr": "french", "ja": "japanese", "pt": "portuguese", "tr": "turkish",
+    "pl": "polish", "ca": "catalan", "nl": "dutch", "ar": "arabic", "sv": "swedish", "it": "italian",
+    "id": "indonesian", "hi": "hindi", "fi": "finnish", "vi": "vietnamese", "he": "hebrew",
+    "uk": "ukrainian", "el": "greek", "ms": "malay", "cs": "czech", "ro": "romanian", "da": "danish",
+    "hu": "hungarian", "ta": "tamil", "no": "norwegian", "th": "thai", "ur": "urdu", "hr": "croatian",
+    "bg": "bulgarian", "lt": "lithuanian", "la": "latin", "mi": "maori", "ml": "malayalam", "cy": "welsh",
+    "sk": "slovak", "te": "telugu", "fa": "persian", "lv": "latvian", "bn": "bengali", "sr": "serbian",
+    "az": "azerbaijani", "sl": "slovenian", "kn": "kannada", "et": "estonian", "mk": "macedonian",
+    "br": "breton", "eu": "basque", "is": "icelandic", "hy": "armenian", "ne": "nepali", "mn": "mongolian",
+    "bs": "bosnian", "kk": "kazakh", "sq": "albanian", "sw": "swahili", "gl": "galician", "mr": "marathi",
+    "pa": "punjabi", "si": "sinhala", "km": "khmer", "sn": "shona", "yo": "yoruba", "so": "somali",
+    "af": "afrikaans", "oc": "occitan", "ka": "georgian", "be": "belarusian", "tg": "tajik",
+    "sd": "sindhi", "gu": "gujarati", "am": "amharic", "yi": "yiddish", "lo": "lao", "uz": "uzbek",
+    "fo": "faroese", "ht": "haitian creole", "ps": "pashto", "tk": "turkmen", "nn": "nynorsk",
+    "mt": "maltese", "sa": "sanskrit", "lb": "luxembourgish", "my": "myanmar", "bo": "tibetan",
+    "tl": "tagalog", "mg": "malagasy", "as": "assamese", "tt": "tatar", "haw": "hawaiian",
+    "ln": "lingala", "ha": "hausa", "ba": "bashkir", "jw": "javanese", "su": "sundanese",
+    "yue": "cantonese", "minnan": "minnan", "wuyu": "wuyu", "dialect": "dialect", "zh/en": "zh/en", "en/zh": "en/zh"
 }
 
-# language code lookup by name, with a few language aliases
 TO_LANGUAGE_CODE = {
     **{language: code for code, language in LANGUAGES.items()},
-    "burmese": "my",
-    "valencian": "ca",
-    "flemish": "nl",
-    "haitian": "ht",
-    "letzeburgesch": "lb",
-    "pushto": "ps",
-    "panjabi": "pa",
-    "moldavian": "ro",
-    "moldovan": "ro",
-    "sinhalese": "si",
-    "castilian": "es",
-    "mandarin": "zh",
+    "burmese": "my", "valencian": "ca", "flemish": "nl", "haitian": "ht", "letzeburgesch": "lb",
+    "pushto": "ps", "panjabi": "pa", "moldavian": "ro", "moldovan": "ro", "sinhalese": "si",
+    "castilian": "es", "mandarin": "zh",
 }
 
 AUDIO_EVENT = {
-    "ASR": "ASR",
-    "AED": "AED",
-    "SER": "SER",
-    "Speech": "Speech",
-    "/Speech": "/Speech",
-    "BGM": "BGM",
-    "/BGM": "/BGM",
-    "Laughter": "Laughter",
-    "/Laughter": "/Laughter",
-    "Applause": "Applause",
-    "/Applause": "/Applause",
+    "ASR": "ASR", "AED": "AED", "SER": "SER", "Speech": "Speech", "/Speech": "/Speech",
+    "BGM": "BGM", "/BGM": "/BGM", "Laughter": "Laughter", "/Laughter": "/Laughter",
+    "Applause": "Applause", "/Applause": "/Applause",
 }
 
 EMOTION = {
-    "HAPPY": "HAPPY",
-    "SAD": "SAD",
-    "ANGRY": "ANGRY",
-    "NEUTRAL": "NEUTRAL",
+    "HAPPY": "HAPPY", "SAD": "SAD", "ANGRY": "ANGRY", "NEUTRAL": "NEUTRAL",
 }
 
 TTS_Vocal_Token = {
-    "TTS/B": "TTS/B",
-    "TTS/O": "TTS/O",
-    "TTS/Q": "TTS/Q",
-    "TTS/A": "TTS/A",
-    "TTS/CO": "TTS/CO",
-    "TTS/CL": "TTS/CL",
-    "TTS/H": "TTS/H",
-    **{f"TTS/SP{i:02d}": f"TTS/SP{i:02d}" for i in range(1, 14)}
+    "TTS/B": "TTS/B", "TTS/O": "TTS/O", "TTS/Q": "TTS/Q", "TTS/A": "TTS/A", "TTS/CO": "TTS/CO",
+    "TTS/CL": "TTS/CL", "TTS/H": "TTS/H", **{f"TTS/SP{i:02d}": f"TTS/SP{i:02d}" for i in range(1, 14)}
 }
 
-
+# ===== 构造 Encoding =====
 @lru_cache(maxsize=None)
 def get_encoding(name: str = "gpt2", num_languages: int = 99):
     vocab_path = os.path.join(os.path.dirname(__file__), "assets", f"{name}.tiktoken")
@@ -175,28 +60,20 @@ def get_encoding(name: str = "gpt2", num_languages: int = 99):
     }
     n_vocab = len(ranks)
     special_tokens = {}
-
     specials = [
-        "<|endoftext|>",
-        "<|startoftranscript|>",
+        "<|endoftext|>", "<|startoftranscript|>",
         *[f"<|{lang}|>" for lang in list(LANGUAGES.keys())[:num_languages]],
         *[f"<|{audio_event}|>" for audio_event in list(AUDIO_EVENT.keys())],
         *[f"<|{emotion}|>" for emotion in list(EMOTION.keys())],
-        "<|translate|>",
-        "<|transcribe|>",
-        "<|startoflm|>",
-        "<|startofprev|>",
-        "<|nospeech|>",
-        "<|notimestamps|>",
-        *[f"<|SPECIAL_TOKEN_{i}|>" for i in range(1, 31)],        # register special tokens for ASR
-        *[f"<|{tts}|>" for tts in list(TTS_Vocal_Token.keys())],  # register special tokens for TTS
+        "<|translate|>", "<|transcribe|>", "<|startoflm|>", "<|startofprev|>",
+        "<|nospeech|>", "<|notimestamps|>",
+        *[f"<|SPECIAL_TOKEN_{i}|>" for i in range(1, 31)],
+        *[f"<|{tts}|>" for tts in list(TTS_Vocal_Token.keys())],
         *[f"<|{i * 0.02:.2f}|>" for i in range(1501)],
     ]
-
     for token in specials:
         special_tokens[token] = n_vocab
         n_vocab += 1
-
     return tiktoken.Encoding(
         name=os.path.basename(vocab_path),
         explicit_n_vocab=n_vocab,
@@ -205,6 +82,16 @@ def get_encoding(name: str = "gpt2", num_languages: int = 99):
         special_tokens=special_tokens,
     )
 
+class SimpleTokenizer:
+    def __init__(self, encoding, num_languages: int = 99, language: Optional[str] = None, task: Optional[str] = None):
+        self.encoding = encoding
+        self.num_languages = num_languages
+        self.language = language
+        self.task = task
+    def encode(self, text: str):
+        return self.encoding.encode(text)
+    def decode(self, tokens: list):
+        return self.encoding.decode(tokens)
 
 @lru_cache(maxsize=None)
 def get_tokenizer(
@@ -212,8 +99,8 @@ def get_tokenizer(
     *,
     num_languages: int = 99,
     language: Optional[str] = None,
-    task: Optional[str] = None,  # Literal["transcribe", "translate", None]
-) -> Tokenizer:
+    task: Optional[str] = None,
+) -> SimpleTokenizer:
     if language is not None:
         language = language.lower()
         if language not in LANGUAGES:
@@ -221,7 +108,6 @@ def get_tokenizer(
                 language = TO_LANGUAGE_CODE[language]
             else:
                 raise ValueError(f"Unsupported language: {language}")
-
     if multilingual:
         encoding_name = "multilingual_zh_ja_yue_char_del"
         language = language or "en"
@@ -230,18 +116,12 @@ def get_tokenizer(
         encoding_name = "gpt2"
         language = None
         task = None
-
     encoding = get_encoding(name=encoding_name, num_languages=num_languages)
-
-    return Tokenizer(
-        encoding=encoding, num_languages=num_languages, language=language, task=task
-    )
-
+    return SimpleTokenizer(encoding=encoding, num_languages=num_languages, language=language, task=task)
 
 class QwenTokenizer():
     def __init__(self, token_path, skip_special_tokens=True):
         super().__init__()
-        # NOTE: non-chat model, all these special tokens keep randomly initialized.
         special_tokens = {
             'eos_token': '<|endoftext|>',
             'pad_token': '<|endoftext|>',
@@ -259,21 +139,13 @@ def __init__(self, token_path, skip_special_tokens=True):
         self.tokenizer = AutoTokenizer.from_pretrained(token_path)
         self.tokenizer.add_special_tokens(special_tokens)
         self.skip_special_tokens = skip_special_tokens
-
     def encode(self, text, **kwargs):
         tokens = self.tokenizer([text], return_tensors="pt")
-        tokens = tokens["input_ids"][0].cpu().tolist()
-        return tokens
-
+        return tokens["input_ids"][0].cpu().tolist()
     def decode(self, tokens):
         tokens = torch.tensor(tokens, dtype=torch.int64)
-        text = self.tokenizer.batch_decode([tokens], skip_special_tokens=self.skip_special_tokens)[0]
-        return text
-
+        return self.tokenizer.batch_decode([tokens], skip_special_tokens=self.skip_special_tokens)[0]
 
 @lru_cache(maxsize=None)
-def get_qwen_tokenizer(
-    token_path: str,
-    skip_special_tokens: bool
-) -> QwenTokenizer:
-    return QwenTokenizer(token_path=token_path, skip_special_tokens=skip_special_tokens)
+def get_qwen_tokenizer(token_path: str, skip_special_tokens: bool) -> QwenTokenizer:
+    return QwenTokenizer(token_path=token_path, skip_special_tokens=skip_special_tokens)
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/main.cpp b/projects/llm_framework/main_cosy_voice/src/main.cpp
index 5fa42c43..32e7c8d1 100644
--- a/projects/llm_framework/main_cosy_voice/src/main.cpp
+++ b/projects/llm_framework/main_cosy_voice/src/main.cpp
@@ -6,6 +6,7 @@
 #include "StackFlow.h"
 #include "runner/LLM.hpp"
 #include "runner/Token2wav.hpp"
+#include "runner/utils/wav.hpp"
 
 #include <signal.h>
 #include <sys/stat.h>
@@ -15,6 +16,7 @@
 #include <base64.h>
 #include <fstream>
 #include <stdexcept>
+#include <samplerate.h>
 #include <semaphore.h>
 #include "../../../../SDK/components/utilities/include/sample_log.h"
 #include "thread_safe_list.h"
@@ -25,8 +27,6 @@ using namespace StackFlows;
 #include "backward.h"
 #endif
 
-#include "runner/utils/wav.hpp"
-
 #define MAX_TASK_NUM 2
 
 int main_exit_flage = 0;
@@ -47,6 +47,12 @@ typedef std::function<void(const std::string &data, bool finish)> task_callback_
     else if (obj.contains(#key))              \
         mode_config_.key = obj[#key];
 
+#define INFER_CONFIG_AUTO_SET(obj, key)             \
+    if (config_body.contains(#key))                 \
+        infer_mode_config_.key = config_body[#key]; \
+    else if (obj.contains(#key))                    \
+        infer_mode_config_.key = obj[#key];
+
 class llm_task {
 private:
     static std::atomic<unsigned int> next_port_;
@@ -72,6 +78,7 @@ class llm_task {
 public:
     enum inference_status { INFERENCE_NONE = 0, INFERENCE_RUNNING };
     LLMAttrType mode_config_;
+    Token2WavAttr infer_mode_config_;
     std::unique_ptr<LLM> lLaMa_;
     std::string model_;
     std::string response_format_;
@@ -163,7 +170,8 @@ class llm_task {
             }
             std::string base_model = base_model_path_ + model_ + "/";
             SLOGI("base_model %s", base_model.c_str());
-
+            CONFIG_AUTO_SET(file_body["mode_param"], mode_rate);
+            CONFIG_AUTO_SET(file_body["mode_param"], audio_rate);
             CONFIG_AUTO_SET(file_body["mode_param"], tokenizer_type);
             CONFIG_AUTO_SET(file_body["mode_param"], filename_tokenizer_model);
             CONFIG_AUTO_SET(file_body["mode_param"], url_tokenizer_model);
@@ -184,6 +192,26 @@ class llm_task {
             CONFIG_AUTO_SET(file_body["mode_param"], b_use_mmap_load_embed);
             CONFIG_AUTO_SET(file_body["mode_param"], b_dynamic_load_axmodel_layer);
             CONFIG_AUTO_SET(file_body["mode_param"], max_token_len);
+            CONFIG_AUTO_SET(file_body["mode_param"], output_path);
+
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], flow_input_embedding);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], rand_noise);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], speech_window);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], flow_encoder_28);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], flow_encoder_53);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], flow_encoder_78);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], flow_encoder_50_final);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], flow_estimator_200);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], flow_estimator_250);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], flow_estimator_300);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], hift_p2_50_first);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], hift_p2_58);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], hift_p1_50_first);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], hift_p1_58);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], prompt_dir);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], prompt_dir);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], prompt_dir);
+            INFER_CONFIG_AUTO_SET(file_body["mode_param"], prompt_dir);
             {
                 auto has_http = [](const std::string &s) { return s.find("http") != std::string::npos; };
 
@@ -243,8 +271,32 @@ class llm_task {
             mode_config_.template_filename_axmodel = base_model + mode_config_.template_filename_axmodel;
             mode_config_.filename_speech_embed     = base_model + mode_config_.filename_speech_embed;
             mode_config_.filename_decoder_axmodel  = base_model + mode_config_.filename_decoder_axmodel;
-            mode_config_.token2wav_axmodel_dir     = base_model + mode_config_.token2wav_axmodel_dir;
-            mode_config_.prompt_files              = base_model + mode_config_.prompt_files;
+
+            infer_mode_config_.flow_input_embedding  = base_model + infer_mode_config_.flow_input_embedding;
+            infer_mode_config_.rand_noise            = base_model + infer_mode_config_.rand_noise;
+            infer_mode_config_.speech_window         = base_model + infer_mode_config_.speech_window;
+            infer_mode_config_.flow_encoder_28       = base_model + infer_mode_config_.flow_encoder_28;
+            infer_mode_config_.flow_encoder_53       = base_model + infer_mode_config_.flow_encoder_53;
+            infer_mode_config_.flow_encoder_78       = base_model + infer_mode_config_.flow_encoder_78;
+            infer_mode_config_.flow_encoder_50_final = base_model + infer_mode_config_.flow_encoder_50_final;
+            infer_mode_config_.flow_estimator_200    = base_model + infer_mode_config_.flow_estimator_200;
+            infer_mode_config_.flow_estimator_250    = base_model + infer_mode_config_.flow_estimator_250;
+            infer_mode_config_.flow_estimator_300    = base_model + infer_mode_config_.flow_estimator_300;
+            infer_mode_config_.hift_p2_50_first      = base_model + infer_mode_config_.hift_p2_50_first;
+            infer_mode_config_.hift_p2_58            = base_model + infer_mode_config_.hift_p2_58;
+            infer_mode_config_.hift_p1_50_first      = base_model + infer_mode_config_.hift_p1_50_first;
+            infer_mode_config_.hift_p1_58            = base_model + infer_mode_config_.hift_p1_58;
+
+            infer_mode_config_.prompt_text =
+                (fs::path(base_model) / infer_mode_config_.prompt_dir / infer_mode_config_.prompt_text).string();
+            infer_mode_config_.llm_prompt_speech_token =
+                (fs::path(base_model) / infer_mode_config_.prompt_dir / infer_mode_config_.llm_prompt_speech_token)
+                    .string();
+            infer_mode_config_.prompt_speech_feat =
+                (fs::path(base_model) / infer_mode_config_.prompt_dir / infer_mode_config_.prompt_speech_feat).string();
+            infer_mode_config_.flow_embedding =
+                (fs::path(base_model) / infer_mode_config_.prompt_dir / infer_mode_config_.flow_embedding).string();
+
             mode_config_.runing_callback = [this](int *p_token, int n_token, const char *p_str, float token_per_sec,
                                                   void *reserve) {
                 if (this->out_callback_) {
@@ -252,10 +304,10 @@ class llm_task {
                 }
             };
 
-            readtxt(mode_config_.prompt_files + "/prompt_text.txt", prompt_text_token);
-            readtxt(mode_config_.prompt_files + "/llm_prompt_speech_token.txt", prompt_speech_token);
-            readtxt(mode_config_.prompt_files + "/prompt_speech_feat.txt", prompt_feat);
-            readtxt<float>(mode_config_.prompt_files + "/flow_embedding.txt", spk_embeds);
+            if (readtxt(infer_mode_config_.prompt_text, prompt_text_token)) return -3;
+            if (readtxt(infer_mode_config_.llm_prompt_speech_token, prompt_speech_token)) return -3;
+            if (readtxt(infer_mode_config_.prompt_speech_feat, prompt_feat)) return -3;
+            if (readtxt<float>(infer_mode_config_.flow_embedding, spk_embeds)) return -3;
 
             lLaMa_ = std::make_unique<LLM>();
             if (!lLaMa_->Init(mode_config_)) {
@@ -263,7 +315,9 @@ class llm_task {
                 lLaMa_.reset();
                 return -2;
             }
-            if (!lToken2Wav.Init(mode_config_.token2wav_axmodel_dir, mode_config_.n_timesteps)) {
+            if (!lToken2Wav.Init(infer_mode_config_)) {
+                lLaMa_->Deinit();
+                lLaMa_.reset();
                 return -1;
             }
             lLaMa_->TextToken2Embeds(prompt_text_token, prompt_text_embeds);
@@ -310,8 +364,47 @@ class llm_task {
         lToken2Wav.reset();
     }
 
-    int tts(const std::string &text, const std::vector<unsigned short> &prompt_text_embeds,
-            const std::vector<unsigned short> &prompt_speech_embeds, const std::vector<float> &prompt_feat,
+    void resample_audio(float *input_buffer, int input_length, float *output_buffer, int *output_length,
+                        double src_ratio)
+    {
+        SRC_STATE *src_state;
+        int error;
+        src_state = src_new(SRC_SINC_FASTEST, 1, &error);
+        if (!src_state) {
+            fprintf(stderr, "Error : src_new() failed: %s\n", src_strerror(error));
+            throw std::string("src_new() failed");
+        }
+        SRC_DATA src_data;
+        src_data.data_in       = input_buffer;
+        src_data.input_frames  = input_length;
+        src_data.src_ratio     = src_ratio;
+        int max_output_length  = (int)(input_length * src_ratio + 1);
+        src_data.data_out      = output_buffer;
+        src_data.output_frames = max_output_length;
+        error                  = src_process(src_state, &src_data);
+        if (error) {
+            fprintf(stderr, "Error : src_process() failed: %s\n", src_strerror(error));
+            src_delete(src_state);
+            throw std::string("src_process() failed");
+        }
+        *output_length = src_data.output_frames_gen;
+        src_delete(src_state);
+    }
+
+    static std::string generateFilename(const fs::path &dir)
+    {
+        auto now             = std::chrono::system_clock::now();
+        std::time_t now_time = std::chrono::system_clock::to_time_t(now);
+        std::tm tm_now;
+        localtime_r(&now_time, &tm_now);
+
+        std::ostringstream oss;
+        oss << "audio_" << std::put_time(&tm_now, "%Y%m%d_%H%M%S") << ".wav";
+        return (dir / oss.str()).string();
+    }
+
+    int tts(const std::string &text, std::vector<unsigned short> &prompt_text_embeds,
+            std::vector<unsigned short> &prompt_speech_embeds, const std::vector<float> &prompt_feat,
             const std::vector<float> &prompt_speech_embeds_flow, std::vector<float> &spk_embeds)
     {
         std::vector<float> output;
@@ -326,9 +419,10 @@ class llm_task {
             std::thread llm_thread(llm_thread_func);
 
             int token_offset     = 0;
-            int prompt_token_len = prompt_speech_embeds_flow.size() / lToken2Wav.flow_embed_size;
+            int prompt_token_len = prompt_speech_embeds_flow.size() / lToken2Wav._attr.flow_embed_size;
             if (prompt_token_len < 75) {
                 SLOGE("Error, prompt speech token len %d < 75", prompt_token_len);
+                if (llm_thread.joinable()) llm_thread.join();
                 return -1;
             }
             int prompt_token_align_len = 75;
@@ -345,54 +439,56 @@ class llm_task {
             int this_token_hop_len;
             int i = 0;
             while (true) {
-                this_token_hop_len =
-                    (token_offset == 0) ? lToken2Wav.token_hop_len + promot_token_pad : lToken2Wav.token_hop_len;
+                this_token_hop_len = (token_offset == 0) ? lToken2Wav._attr.token_hop_len + promot_token_pad
+                                                         : lToken2Wav._attr.token_hop_len;
 
                 std::unique_lock<std::mutex> lock(g_buffer_mutex);
 
                 g_buffer_cv.wait(lock, [&] {
                     return (g_token_buffer.size() - token_offset >=
-                            this_token_hop_len + lToken2Wav.pre_lookahead_len) ||
+                            this_token_hop_len + lToken2Wav._attr.pre_lookahead_len) ||
                            g_llm_finished.load() || g_stop.load();
                 });
 
                 if (g_stop) {
                     lock.unlock();
                     break;
-                } else if (g_token_buffer.size() - token_offset >= this_token_hop_len + lToken2Wav.pre_lookahead_len) {
+                } else if (g_token_buffer.size() - token_offset >=
+                           this_token_hop_len + lToken2Wav._attr.pre_lookahead_len) {
                     std::vector<SpeechToken> token;
-                    int start = token_offset - std::min(int(token_offset / lToken2Wav.token_hop_len),
-                                                        lToken2Wav.max_infer_chunk_num - 1) *
-                                                   lToken2Wav.token_hop_len;
-                    int end = token_offset + this_token_hop_len + lToken2Wav.pre_lookahead_len;
+                    int start = token_offset - std::min(int(token_offset / lToken2Wav._attr.token_hop_len),
+                                                        lToken2Wav._attr.max_infer_chunk_num - 1) *
+                                                   lToken2Wav._attr.token_hop_len;
+                    int end = token_offset + this_token_hop_len + lToken2Wav._attr.pre_lookahead_len;
 
                     token.insert(token.end(), g_token_buffer.begin() + start, g_token_buffer.begin() + end);
 
                     lock.unlock();
-                    std::cout << "[Main/Token2Wav Thread] Processing batch of " << token.size() << " tokens...\n";
                     auto speech = lToken2Wav.infer(token, prompt_speech_embeds_flow1, prompt_feat1, spk_embeds,
                                                    token_offset, false);
                     token_offset += this_token_hop_len;
                     output.insert(output.end(), speech.begin(), speech.end());
-                    std::string path = "output_" + std::to_string(i) + ".wav";
+                    double src_ratio =
+                        static_cast<double>(mode_config_.audio_rate) / static_cast<double>(mode_config_.mode_rate);
+                    std::vector<float> resampled_pcm(static_cast<size_t>(speech.size() * src_ratio + 1));
+                    int resampled_len = 0;
+                    resample_audio(speech.data(), speech.size(), resampled_pcm.data(), &resampled_len, src_ratio);
                     std::vector<int16_t> wav_pcm_data;
-                    wav_pcm_data.resize(speech.size());
-                    for (size_t k = 0; k < speech.size(); ++k) {
-                        float sample = speech[k];
-                        if (sample > 1.0f) sample = 1.0f;
-                        if (sample < -1.0f) sample = -1.0f;
-                        wav_pcm_data[k] = static_cast<int16_t>(sample * 32767.0f);
+                    wav_pcm_data.reserve(resampled_len);
+                    for (int i = 0; i < resampled_len; i++) {
+                        float val = resampled_pcm[i];
+                        if (val > 1.0f) val = 1.0f;
+                        if (val < -1.0f) val = -1.0f;
+                        wav_pcm_data.push_back(static_cast<int16_t>(val * 32767.0f));
                     }
+
                     if (out_callback_) {
                         out_callback_(std::string(reinterpret_cast<char *>(wav_pcm_data.data()),
                                                   wav_pcm_data.size() * sizeof(int16_t)),
                                       false);
                     }
-
-                    saveVectorAsWavFloat(speech, path, 24000, 1);
-                    i += 1;
+                    ++i;
                 } else if (g_llm_finished.load()) {
-                    std::cout << "[Main/Token2Wav Thread] Buffer is empty and LLM finished. Exiting.\n";
                     lock.unlock();
                     break;
                 } else {
@@ -410,34 +506,64 @@ class llm_task {
             }
 
             std::vector<SpeechToken> token;
-            int start = g_token_buffer.size() - std::min(int(g_token_buffer.size() / lToken2Wav.token_hop_len),
-                                                         lToken2Wav.max_infer_chunk_num - 1) *
-                                                    lToken2Wav.token_hop_len;
+            int start = g_token_buffer.size() - std::min(int(g_token_buffer.size() / lToken2Wav._attr.token_hop_len),
+                                                         lToken2Wav._attr.max_infer_chunk_num - 1) *
+                                                    lToken2Wav._attr.token_hop_len;
             token.insert(token.end(), g_token_buffer.begin() + start, g_token_buffer.end());
             auto speech = lToken2Wav.infer(token, prompt_speech_embeds_flow1, prompt_feat1, spk_embeds,
                                            token_offset - start, true);
             output.insert(output.end(), speech.begin(), speech.end());
-            std::string path = "output_" + std::to_string(i) + ".wav";
-            saveVectorAsWavFloat(speech, path, 24000, 1);
-            saveVectorAsWavFloat(output, "output.wav", 24000, 1);
+            double src_ratio =
+                static_cast<double>(mode_config_.audio_rate) / static_cast<double>(mode_config_.mode_rate);
+            std::vector<float> resampled_pcm(static_cast<size_t>(speech.size() * src_ratio + 1));
+            int resampled_len = 0;
+            resample_audio(speech.data(), speech.size(), resampled_pcm.data(), &resampled_len, src_ratio);
+
             std::vector<int16_t> wav_pcm_data;
-            wav_pcm_data.resize(speech.size());
-            for (size_t k = 0; k < speech.size(); ++k) {
-                float sample = speech[k];
-                if (sample > 1.0f) sample = 1.0f;
-                if (sample < -1.0f) sample = -1.0f;
-                wav_pcm_data[k] = static_cast<int16_t>(sample * 32767.0f);
+            wav_pcm_data.reserve(resampled_len);
+            for (int i = 0; i < resampled_len; i++) {
+                float val = resampled_pcm[i];
+                if (val > 1.0f) val = 1.0f;
+                if (val < -1.0f) val = -1.0f;
+                wav_pcm_data.push_back(static_cast<int16_t>(val * 32767.0f));
             }
             if (out_callback_) {
                 out_callback_(
                     std::string(reinterpret_cast<char *>(wav_pcm_data.data()), wav_pcm_data.size() * sizeof(int16_t)),
                     true);
             }
+            if (response_format_.find("file") != std::string::npos) {
+                double src_ratio =
+                    static_cast<double>(mode_config_.audio_rate) / static_cast<double>(mode_config_.mode_rate);
+                std::vector<float> resampled_pcm(static_cast<size_t>(output.size() * src_ratio + 1));
+                int resampled_len = 0;
+                resample_audio(output.data(), output.size(), resampled_pcm.data(), &resampled_len, src_ratio);
+
+                std::vector<int16_t> wav_pcm_data_full;
+                wav_pcm_data_full.reserve(resampled_len);
+                for (int i = 0; i < resampled_len; i++) {
+                    float val = resampled_pcm[i];
+                    if (val > 1.0f) val = 1.0f;
+                    if (val < -1.0f) val = -1.0f;
+                    wav_pcm_data_full.push_back(static_cast<int16_t>(val * 32767.0f));
+                }
+
+                std::string wav_path;
+                if (mode_config_.output_path.empty()) {
+                    wav_path = generateFilename("/tmp");
+                } else {
+                    fs::path user_path(mode_config_.output_path);
+                    if (!user_path.has_filename()) {
+                        wav_path = generateFilename(user_path);
+                    } else {
+                        wav_path = user_path.string();
+                    }
+                }
+                saveVectorAsWavFloat(resampled_pcm, wav_path, mode_config_.audio_rate, 1);
+            }
 
             SLOGI("tts total use time: %.3f s", time_total.cost() / 1000);
             reset();
-            std::cout << "\nVoice generation pipeline completed.\n";
-
         } catch (const std::exception &e) {
             std::cerr << "Error in pipeline: " << e.what() << std::endl;
             return 1;
@@ -505,9 +631,9 @@ class llm_task {
     static unsigned int getNextPort()
     {
         unsigned int port = next_port_++;
-        if (port > 8089) {
-            next_port_ = 8080;
-            port       = 8080;
+        if (port > 8079) {
+            next_port_ = 8070;
+            port       = 8070;
         }
         return port;
     }
@@ -548,7 +674,7 @@ class llm_task {
     }
 };
 
-std::atomic<unsigned int> llm_task::next_port_{8080};
+std::atomic<unsigned int> llm_task::next_port_{8070};
 
 #undef CONFIG_AUTO_SET
 
@@ -605,7 +731,7 @@ class llm_cosy_voice : public StackFlow {
 
     void pause(const std::string &work_id, const std::string &object, const std::string &data) override
     {
-        SLOGI("llm_asr::work:%s", data.c_str());
+        SLOGI("llm_cosy_voice::work:%s", data.c_str());
 
         nlohmann::json error_body;
         int work_id_num = sample_get_work_id_num(work_id);
@@ -742,7 +868,7 @@ class llm_cosy_voice : public StackFlow {
 
     void link(const std::string &work_id, const std::string &object, const std::string &data) override
     {
-        SLOGI("llm_llm::link:%s", data.c_str());
+        SLOGI("llm_cosy_voice::link:%s", data.c_str());
         int ret = 1;
         nlohmann::json error_body;
         int work_id_num = sample_get_work_id_num(work_id);
@@ -779,7 +905,7 @@ class llm_cosy_voice : public StackFlow {
 
     void unlink(const std::string &work_id, const std::string &object, const std::string &data) override
     {
-        SLOGI("llm_llm::unlink:%s", data.c_str());
+        SLOGI("llm_cosy_voice::unlink:%s", data.c_str());
         int ret = 0;
         nlohmann::json error_body;
         int work_id_num = sample_get_work_id_num(work_id);
@@ -804,7 +930,7 @@ class llm_cosy_voice : public StackFlow {
 
     void taskinfo(const std::string &work_id, const std::string &object, const std::string &data) override
     {
-        SLOGI("llm_llm::taskinfo:%s", data.c_str());
+        SLOGI("llm_cosy_voice::taskinfo:%s", data.c_str());
         // int ret = 0;
         nlohmann::json req_body;
         int work_id_num = sample_get_work_id_num(work_id);
@@ -826,13 +952,13 @@ class llm_cosy_voice : public StackFlow {
             req_body["response_format"] = llm_task_obj->response_format_;
             req_body["enoutput"]        = llm_task_obj->enoutput_;
             req_body["inputs"]          = llm_task_obj->inputs_;
-            send("llm.taskinfo", req_body, LLM_NO_ERROR, work_id);
+            send("cosy_voice.taskinfo", req_body, LLM_NO_ERROR, work_id);
         }
     }
 
     int exit(const std::string &work_id, const std::string &object, const std::string &data) override
     {
-        SLOGI("llm_llm::exit:%s", data.c_str());
+        SLOGI("llm_cosy_voice::exit:%s", data.c_str());
 
         nlohmann::json error_body;
         int work_id_num = sample_get_work_id_num(work_id);
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/BaseRunner.cpp b/projects/llm_framework/main_cosy_voice/src/runner/BaseRunner.cpp
new file mode 100644
index 00000000..cea8cf6a
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/BaseRunner.cpp
@@ -0,0 +1,14 @@
+#include "BaseRunner.hpp"
+#include "OnnxWarpper/OnnxWarpper.hpp"
+
+std::shared_ptr<BaseRunner> CreateRunner(RunnerType rt)
+{
+    switch (rt)
+    {
+    case RT_OnnxRunner:
+        return std::make_shared<OnnxRunner>();
+    default:
+        return nullptr;
+    }
+    return nullptr;
+}
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/BaseRunner.hpp b/projects/llm_framework/main_cosy_voice/src/runner/BaseRunner.hpp
new file mode 100644
index 00000000..cd99e147
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/BaseRunner.hpp
@@ -0,0 +1,37 @@
+#pragma once
+#include "string"
+#include "vector"
+#include "memory"
+
+enum RunnerType
+{
+    RT_UNKNOWN,
+    RT_OnnxRunner,
+    RT_END,
+};
+
+struct BaseConfig
+{
+    std::string onnx_model;
+    std::string output_model; // for trt
+    int nthread;
+};
+
+class BaseRunner
+{
+public:
+    virtual int load(BaseConfig &config) = 0;
+    virtual int inference() = 0;
+
+    virtual int getInputCount() = 0;
+    virtual std::vector<size_t> getInputShape(int idx) = 0;
+    virtual std::string getInputName(int idx) = 0;
+    virtual float *getInputPtr(int idx) = 0;
+
+    virtual int getOutputCount() = 0;
+    virtual std::vector<size_t> getOutputShape(int idx) = 0;
+    virtual std::string getOutputName(int idx) = 0;
+    virtual float *getOutputPtr(int idx) = 0;
+};
+
+std::shared_ptr<BaseRunner> CreateRunner(RunnerType rt);
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp b/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
index 01a93e40..8e74d906 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
@@ -31,48 +31,51 @@ typedef std::function<void(int *, int, const char *, float, void *)> LLMRuningCa
 struct LLMAttrType {
     std::string template_filename_axmodel = "tinyllama-int8/tinyllama_l%d.axmodel";
     int axmodel_num                       = 22;
-
-    std::string filename_post_axmodel = "tinyllama-int8/tinyllama_post.axmodel";
+    std::string filename_post_axmodel     = "tinyllama-int8/tinyllama_post.axmodel";
+    std::string filename_tokenizer_model  = "http://127.0.0.1:12345";
     std::string filename_decoder_axmodel;
     std::string token2wav_axmodel_dir;
-    std::string prompt_files;
+    std::string output_path;
 
     int prefill_token_num     = 96;  // auto calc
     int prefill_max_token_num = 512;
-    std::vector<int> prefill_max_kv_cache_num_grp;
-    int precompute_len = 0;
-    int prefill_grpid  = -1;
 
-    TokenizerType tokenizer_type         = TKT_HTTP;
-    std::string filename_tokenizer_model = "http://127.0.0.1:12345";
-    std::string url_tokenizer_model      = "http://127.0.0.1:12345";
-    bool b_bos = false, b_eos = false;
+    std::string prompt_files;
+
+    TokenizerType tokenizer_type    = TKT_HTTP;
+    std::string url_tokenizer_model = "http://127.0.0.1:12345";
+    bool b_bos                      = false;
+    bool b_eos                      = false;
+    std::vector<int> dev_ids        = {0};
+
     std::string filename_tokens_embed = "tinyllama.model.embed_tokens.weight.bfloat16.bin";
     std::string filename_llm_embed    = "tinyllama.model.embed_tokens.weight.bfloat16.bin";
     std::string filename_speech_embed = "tinyllama.model.embed_tokens.weight.bfloat16.bin";
     int tokens_embed_num              = 151936;
     int tokens_embed_size             = 896;
 
+    int max_token_len = 127;
+    int kv_cache_num  = 1024;
+    int kv_cache_size = 256;
+
+    int precompute_len = 0;
+    std::vector<int> prefill_max_kv_cache_num_grp;
+    int prefill_grpid = -1;
+
     int llm_embed_num     = 2;
     int llm_embed_size    = 896;
     int speech_embed_num  = 6564;
     int speech_embed_size = 896;
 
-    int max_token_len = 127;  // auto calc
+    int n_timesteps;
 
-    int kv_cache_num  = 1024;  // auto calc
-    int kv_cache_size = 256;   // auto calc
+    int mode_rate  = 24000;
+    int audio_rate = 48000;
 
     bool b_use_mmap_load_embed        = false;
     bool b_dynamic_load_axmodel_layer = false;
+    bool b_use_mmap_load_layer        = true;
 
-    bool b_use_mmap_load_layer = true;
-
-    bool b_use_topk = false;
-
-    int n_timesteps;
-
-    // bool b_live_print = true;
     LLMRuningCallback runing_callback = nullptr;
     void *reserve                     = nullptr;
 };
@@ -282,7 +285,7 @@ class LLM {
     }
 
     int Encode(std::vector<unsigned short> &out_embed, std::vector<std::vector<int>> &position_ids, std::string text,
-               const std::vector<unsigned short> &prompt_text_embeds, const std::vector<unsigned short> &prompt_speech_embeds)
+               std::vector<unsigned short> &prompt_text_embeds, std::vector<unsigned short> &prompt_speech_embeds)
     {
         // std::vector<int> prompt_ids = tokenizer->Encode(prompt_text, true);
         ImageInfo img_info;
@@ -333,9 +336,9 @@ class LLM {
         return 0;
     }
 
-    int Run(const std::string &input_str, const std::vector<unsigned short> &prompt_text_embeds,
-            const std::vector<unsigned short> &prompt_speech_embeds, TokenBuffer &token_buffer,
-            std::mutex &buffer_mutex, std::condition_variable &buffer_cv, std::atomic<bool> &llm_finished)
+    int Run(std::string input_str, std::vector<unsigned short> &prompt_text_embeds,
+            std::vector<unsigned short> &prompt_speech_embeds, TokenBuffer &token_buffer, std::mutex &buffer_mutex,
+            std::condition_variable &buffer_cv, std::atomic<bool> &llm_finished)
     {
         std::vector<unsigned short> text_embed;
         std::vector<std::vector<int>> position_ids;
@@ -511,12 +514,7 @@ class LLM {
             auto &input = llama_post.get_input(0);
             memcpy((void *)input.pVirAddr, embed.data(), embed.size() * sizeof(unsigned short));
             llama_post.inference();
-            if (_attr.b_use_topk) {
-                AX_SYS_MinvalidateCache(llama_post.get_output("indices").phyAddr,
-                                        llama_post.get_output("indices").pVirAddr,
-                                        llama_post.get_output("indices").nSize);
-                max_index = *(int *)llama_post.get_output("indices").pVirAddr;
-            } else {
+            {
                 auto &output_post        = llama_post.get_output("output_norm");  // 1 means get rmsnorm output
                 unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
                 std::vector<float> logits(output_post.nSize / sizeof(unsigned short));
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/OnnxWarpper/OnnxWarpper.hpp b/projects/llm_framework/main_cosy_voice/src/runner/OnnxWarpper/OnnxWarpper.hpp
new file mode 100644
index 00000000..f08dc1c0
--- /dev/null
+++ b/projects/llm_framework/main_cosy_voice/src/runner/OnnxWarpper/OnnxWarpper.hpp
@@ -0,0 +1,174 @@
+#pragma once
+#include "../BaseRunner.hpp"
+
+#include "onnxruntime_cxx_api.h"
+
+class OnnxRunner : virtual public BaseRunner
+{
+    Ort::Env env;
+    Ort::Session session{nullptr};
+
+    std::vector<std::vector<size_t>> inputs_shape;
+    std::vector<std::string> inputs_name;
+    std::vector<const char *> inputs_name_cstr;
+    std::vector<std::shared_ptr<float>> inputs_data;
+    std::vector<Ort::Value> inputs_tensor;
+
+    std::vector<std::vector<size_t>> outputs_shape;
+    std::vector<std::string> outputs_name;
+    std::vector<const char *> outputs_name_cstr;
+    std::vector<std::shared_ptr<float>> outputs_data;
+    std::vector<Ort::Value> outputs_tensor;
+
+public:
+    int load(BaseConfig &config) override
+    {
+        Ort::SessionOptions session_options;
+        session_options.SetInterOpNumThreads(config.nthread);
+        session_options.SetIntraOpNumThreads(config.nthread);
+        // session_options
+        // TensorRT加速开启，CUDA加速开启
+        // OrtSessionOptionsAppendExecutionProvider_Tensorrt(session_options, 0); // tensorRT
+        // OrtSessionOptionsAppendExecutionProvider_CUDA(session_options, 0);
+        session_options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_ALL);
+        auto memory_info = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeCPU);
+        Ort::AllocatorWithDefaultOptions allocator;
+        session = Ort::Session{env, config.onnx_model.c_str(), session_options};
+        printf("\ninputs: \n");
+        for (size_t i = 0; i < session.GetInputCount(); i++)
+        {
+            auto input_name = std::string(session.GetInputNameAllocated(i, allocator).get());
+            inputs_name.push_back(input_name);
+
+            printf("%20s: ", input_name.c_str());
+            auto input_shape = session.GetInputTypeInfo(i).GetTensorTypeAndShapeInfo().GetShape();
+
+            if (input_shape.size() && input_shape[0] < 0)
+            {
+                input_shape[0] = 1;
+            }
+            std::vector<size_t> tmp_input_shape(input_shape.size());
+            for (size_t j = 0; j < input_shape.size(); j++)
+            {
+                tmp_input_shape[j] = input_shape[j];
+                printf("%ld", tmp_input_shape[j]);
+                if (j < (input_shape.size() - 1))
+                    printf(" x ");
+            }
+            printf("\n");
+            inputs_shape.push_back(tmp_input_shape);
+
+            int len = 1;
+            for (size_t d = 0; d < input_shape.size(); d++)
+            {
+                len *= input_shape[d];
+            }
+            std::shared_ptr<float> data(new float[len], std::default_delete<float[]>());
+            inputs_data.push_back(data);
+
+            inputs_tensor.push_back(Ort::Value::CreateTensor<float>(memory_info, data.get(), len, input_shape.data(), input_shape.size()));
+        }
+        printf("output: \n");
+        for (size_t i = 0; i < session.GetOutputCount(); i++)
+        {
+            auto output_name = std::string(session.GetOutputNameAllocated(i, allocator).get());
+            outputs_name.push_back(output_name);
+
+            printf("%20s: ", output_name.c_str());
+            auto output_shape = session.GetOutputTypeInfo(i).GetTensorTypeAndShapeInfo().GetShape();
+            if (output_shape.size() && output_shape[0] < 0)
+            {
+                output_shape[0] = 1;
+            }
+            std::vector<size_t> tmp_output_shape(output_shape.size());
+            for (size_t j = 0; j < output_shape.size(); j++)
+            {
+                tmp_output_shape[j] = output_shape[j];
+                printf("%ld", tmp_output_shape[j]);
+                if (j < (output_shape.size() - 1))
+                    printf(" x ");
+            }
+            printf("\n");
+            outputs_shape.push_back(tmp_output_shape);
+
+            int len = 1;
+            for (size_t d = 0; d < output_shape.size(); d++)
+            {
+                len *= output_shape[d];
+            }
+            std::shared_ptr<float> data(new float[len], std::default_delete<float[]>());
+            outputs_data.push_back(data);
+
+            outputs_tensor.push_back(Ort::Value::CreateTensor<float>(memory_info, data.get(), len, output_shape.data(), output_shape.size()));
+        }
+
+        inputs_name_cstr.resize(inputs_name.size());
+        for (size_t i = 0; i < inputs_name.size(); i++)
+        {
+            inputs_name_cstr[i] = inputs_name[i].c_str();
+        }
+
+        outputs_name_cstr.resize(outputs_name.size());
+        for (size_t i = 0; i < outputs_name.size(); i++)
+        {
+            outputs_name_cstr[i] = outputs_name[i].c_str();
+        }
+
+        return 0;
+    }
+
+    int inference() override
+    {
+        Ort::RunOptions run_options;
+        // for (size_t i = 0; i < inputs_name_cstr.size(); i++)
+        // {
+        //     printf("%20s: \n", inputs_name_cstr[i]);
+        // }
+        //  for (size_t i = 0; i < outputs_name_cstr.size(); i++)
+        // {
+        //     printf("%20s: \n", outputs_name_cstr[i]);
+        // }
+        session.Run(run_options, inputs_name_cstr.data(), inputs_tensor.data(), inputs_tensor.size(), outputs_name_cstr.data(), outputs_tensor.data(), outputs_tensor.size());
+        return 0;
+    }
+
+    int getInputCount() override
+    {
+        return inputs_shape.size();
+    }
+
+    std::vector<size_t> getInputShape(int idx) override
+    {
+        return inputs_shape[idx];
+    }
+
+    std::string getInputName(int idx) override
+    {
+        return inputs_name[idx];
+    }
+
+    float *getInputPtr(int idx) override
+    {
+        return inputs_data[idx].get();
+    }
+
+    int getOutputCount() override
+    {
+        return outputs_shape.size();
+    }
+
+    std::vector<size_t> getOutputShape(int idx) override
+    {
+        return outputs_shape[idx];
+    }
+
+    std::string getOutputName(int idx) override
+    {
+        return outputs_name[idx];
+    }
+
+    float *getOutputPtr(int idx) override
+    {
+        return outputs_data[idx].get();
+    }
+};
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp b/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp
index f0cc3c43..8fd5f387 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp
@@ -6,8 +6,10 @@
 #include <iostream>
 #include <unordered_map>
 #include <vector>
-#include <cstddef> // For size_t
-#include <stdexcept> // For std::invalid_argument
+#include <cstddef>
+#include <stdexcept>
+#include <memory>
+
 #include "bfloat16.hpp"
 #include "Tokenizer/Tokenizer.hpp"
 #include "LLMEmbedSelector.hpp"
@@ -18,198 +20,126 @@
 #include "ax_cmm_utils.hpp"
 #include "cqdm.h"
 #include "timer.hpp"
-// #include "opencv2/opencv.hpp"
 #include "ax_sys_api.h"
-#include "MNN/MNNDefine.h"
-#include "MNN/MNNForwardType.h"
-#include "MNN/Interpreter.hpp"
+#include "BaseRunner.hpp"
+
+struct Token2WavAttr {
+    int devid                = 0;
+    int flow_embed_num       = 6561;
+    int flow_embed_size      = 512;
+    int token_mel_ratio      = 2;
+    int token_hop_len        = 25;
+    int max_infer_chunk_num  = 3;
+    int mel_cache_len        = 8;
+    int source_cache_len     = mel_cache_len * 480;
+    int pre_lookahead_len    = 3;
+    float inference_cfg_rate = 0.7;
 
-class Token2Wav
-{
+    std::string flow_input_embedding    = "flow.input_embedding.float16.bin";
+    std::string rand_noise              = "rand_noise_1_80_300.txt";
+    std::string speech_window           = "speech_window_2x8x480.txt";
+    std::string flow_encoder_28         = "flow_encoder_28.axmodel";
+    std::string flow_encoder_53         = "flow_encoder_53.axmodel";
+    std::string flow_encoder_78         = "flow_encoder_78.axmodel";
+    std::string flow_encoder_50_final   = "flow_encoder_50_final.axmodel";
+    std::string flow_estimator_200      = "flow_estimator_200.axmodel";
+    std::string flow_estimator_250      = "flow_estimator_250.axmodel";
+    std::string flow_estimator_300      = "flow_estimator_300.axmodel";
+    std::string hift_p2_50_first        = "hift_p2_50_first.axmodel";
+    std::string hift_p2_58              = "hift_p2_58.axmodel";
+    std::string hift_p1_50_first        = "hift_p1_50_first.onnx";
+    std::string hift_p1_58              = "hift_p1_58.onnx";
+    std::string prompt_dir              = "prompt_data/";
+    std::string prompt_text             = "prompt_text.txt";
+    std::string llm_prompt_speech_token = "llm_prompt_speech_token.txt";
+    std::string prompt_speech_feat      = "prompt_speech_feat.txt";
+    std::string flow_embedding          = "flow_embedding.txt";
+    int n_timesteps                     = 10;
+};
+
+class Token2Wav {
 public:
-    int flow_embed_num = 6561;
-    int flow_embed_size = 512;
-    int token_mel_ratio = 2;
-    int token_hop_len = 25;
-    int max_infer_chunk_num = 3;
-    int mel_cache_len = 8;
-    int source_cache_len = mel_cache_len * 480;
-    int pre_lookahead_len = 3;
-    float inference_cfg_rate = 0.7;
+    Token2WavAttr _attr;
 
 private:
     ax_runner_ax650 flow_encoder_28;
     ax_runner_ax650 flow_encoder_53;
     ax_runner_ax650 flow_encoder_78;
     ax_runner_ax650 flow_encoder_50_final;
-
     ax_runner_ax650 flow_estimator_200;
     ax_runner_ax650 flow_estimator_250;
     ax_runner_ax650 flow_estimator_300;
-
     ax_runner_ax650 hift_p2_50_first;
     ax_runner_ax650 hift_p2_58;
-
-    std::shared_ptr<MNN::Interpreter> hift_p1_50_first = nullptr;
-    std::shared_ptr<MNN::Interpreter> hift_p1_58 = nullptr;
-
-    MNN::Session * sess_hift_p1_50_first = nullptr;
-    MNN::Session * sess_hift_p1_58 = nullptr;
-
-
+    std::shared_ptr<BaseRunner> hift_p1_50_first;
+    std::shared_ptr<BaseRunner> hift_p1_58;
     std::vector<float> rand_noise;
     std::vector<float> t_span;
-
     LLaMaEmbedSelector flow_embed_selector;
-    
     std::unordered_map<std::string, std::vector<float>> hift_cache_dict;
-    std::vector<float> speech_window; // np.hamming(2 * 8 * 480)
+    std::vector<float> speech_window;
 
-    int init_noise(std::string model_dir)
+    int init_noise(const std::string &path)
     {
-        return readtxt(model_dir+"/rand_noise_1_80_300.txt", rand_noise);
+        return readtxt(path, rand_noise);
     }
-
-    int init_speech_window(std::string model_dir)
+    int init_speech_window(const std::string &path)
     {
-        return readtxt(model_dir+"/speech_window_2x8x480.txt", speech_window);
+        return readtxt(path, speech_window);
     }
-
     int init_tspan(int n_timesteps)
     {
-        // std::vector<float> t_span_10 = {0.0000, 0.0123, 0.0489, 0.1090, 0.1910, 0.2929, 0.4122, 0.5460, 0.6910,0.8436, 1.0000};
-        // std::vector<float> t_span_7 = {0.0000, 0.1429, 0.2857, 0.4286, 0.5714, 0.7143, 0.8571, 1.0000}; // n_timesteps = 7
-        // std::vector<float> t_span_6 = {0.0000, 0.1667, 0.3333, 0.5000, 0.6667, 0.8333, 1.0000};       // n_timesteps = 6
-        // std::vector<float> t_span_5 = {0.0000, 0.2000, 0.4000, 0.6000, 0.8000, 1.0000 };
-        // std::vector<float> t_span_4 = {0.0000, 0.2500, 0.5000, 0.7500, 1.0000};
-
-        if(n_timesteps <4)
-        {
-            return -1;
-        }
-
-        n_timesteps = n_timesteps;
+        if (n_timesteps < 4) return -1;
         t_span = linspace(0.0, 1.0, n_timesteps + 1);
         return 0;
     }
 
 public:
-    bool Init(std::string model_dir, int n_timesteps)
+    bool Init(const Token2WavAttr &attr)
     {
-        int ret;
-
-        ret = init_tspan(n_timesteps);
-        if(ret != 0){
-            ALOGE("init_tspan failed, n_timesteps:%d", n_timesteps);
-            return false;
-        }
-
-        ret = init_noise(model_dir);
-        if(ret != 0){
-            ALOGE("init rand noise(%s) failed", "rand_noise_1_80_300.txt");
-            return false;
-        }
-
-        ret = init_speech_window(model_dir);
-        if(ret != 0){
-            ALOGE("init speech_window(%s) failed", "speech_window_2x8x480.txt");
-            return false;
-        }
-
-        if (!flow_embed_selector.Init((model_dir+"/flow.input_embedding.float16.bin").c_str(), flow_embed_num, flow_embed_size, false))
-        {
-            ALOGE("flow_embed_selector.Init(%s, %d, %d) failed", (model_dir+"/flow.input_embedding.float16.bin").c_str(),flow_embed_num, flow_embed_size);
-            return false;
-        }
-
-        ret = flow_encoder_28.init((model_dir+"/flow_encoder_28.axmodel").c_str(), false);
-        if (ret != 0)
-        {
-            ALOGE("init axmodel(%s) failed", (model_dir+"/flow_encoder_28.axmodel").c_str());
-            return false;
-        }
+        _attr = attr;
 
-        ret = flow_encoder_53.init((model_dir+"/flow_encoder_53.axmodel").c_str(), false);
-        if (ret != 0)
-        {
-            ALOGE("init axmodel(%s) failed", (model_dir+"/flow_encoder_53.axmodel").c_str());
+        if (init_tspan(_attr.n_timesteps) != 0) {
+            ALOGE("init_tspan failed, n_timesteps:%d", _attr.n_timesteps);
             return false;
         }
-
-        ret = flow_encoder_78.init((model_dir+"/flow_encoder_78.axmodel").c_str(), false);
-        if (ret != 0)
-        {
-            ALOGE("init axmodel(%s) failed", (model_dir+"/flow_encoder_78.axmodel").c_str());
+        if (init_noise(_attr.rand_noise) != 0) {
+            ALOGE("init rand noise failed");
             return false;
         }
-
-        ret = flow_encoder_50_final.init((model_dir+"/flow_encoder_50_final.axmodel").c_str(), false);
-        if (ret != 0)
-        {
-            ALOGE("init axmodel(%s) failed", (model_dir+"/flow_encoder_50_final.axmodel").c_str());
+        if (init_speech_window(_attr.speech_window) != 0) {
+            ALOGE("init speech_window failed");
             return false;
         }
-
-        ret = flow_estimator_200.init((model_dir+"/flow_estimator_200.axmodel").c_str(), false);
-        if (ret != 0)
-        {
-            ALOGE("init axmodel(%s) failed", (model_dir+"/flow_estimator_200.axmodel").c_str());
-            return false;
-        }
-
-        ret = flow_estimator_250.init((model_dir+"/flow_estimator_250.axmodel").c_str(), false);
-        if (ret != 0)
-        {
-            ALOGE("init axmodel(%s) failed", (model_dir+"/flow_estimator_250.axmodel").c_str());
-            return false;
-        }
-
-        ret = flow_estimator_300.init((model_dir+"/flow_estimator_300.axmodel").c_str(), false);
-        if (ret != 0)
-        {
-            ALOGE("init axmodel(%s) failed", (model_dir+"/flow_estimator_300.axmodel").c_str());
-            return false;
-        }
-
-        ret = hift_p2_50_first.init((model_dir+"/hift_p2_50_first.axmodel").c_str(), false);
-        if (ret != 0)
-        {
-            ALOGE("init axmodel(%s) failed", (model_dir+"/hift_p2_50_first.axmodel").c_str());
-            return false;
-        }
-
-        ret = hift_p2_58.init((model_dir+"/hift_p2_58.axmodel").c_str(), false);
-        if (ret != 0)
-        {
-            ALOGE("init axmodel(%s) failed", (model_dir+"/hift_p2_58.axmodel").c_str());
+        if (!flow_embed_selector.Init(_attr.flow_input_embedding.c_str(), _attr.flow_embed_num, _attr.flow_embed_size,
+                                      false)) {
+            ALOGE("flow_embed_selector init failed");
             return false;
         }
+        if (flow_encoder_28.init(_attr.flow_encoder_28.c_str(), _attr.devid) != 0) return false;
+        if (flow_encoder_53.init(_attr.flow_encoder_53.c_str(), _attr.devid) != 0) return false;
+        if (flow_encoder_78.init(_attr.flow_encoder_78.c_str(), _attr.devid) != 0) return false;
+        if (flow_encoder_50_final.init(_attr.flow_encoder_50_final.c_str(), _attr.devid) != 0) return false;
+        if (flow_estimator_200.init(_attr.flow_estimator_200.c_str(), _attr.devid) != 0) return false;
+        if (flow_estimator_250.init(_attr.flow_estimator_250.c_str(), _attr.devid) != 0) return false;
+        if (flow_estimator_300.init(_attr.flow_estimator_300.c_str(), _attr.devid) != 0) return false;
+        if (hift_p2_50_first.init(_attr.hift_p2_50_first.c_str(), _attr.devid) != 0) return false;
+        if (hift_p2_58.init(_attr.hift_p2_58.c_str(), _attr.devid) != 0) return false;
+
+        hift_p1_50_first = CreateRunner(RT_OnnxRunner);
+        if (!hift_p1_50_first) return false;
+        BaseConfig cfg50;
+        cfg50.nthread    = 8;
+        cfg50.onnx_model = _attr.hift_p1_50_first;
+        hift_p1_50_first->load(cfg50);
+
+        hift_p1_58 = CreateRunner(RT_OnnxRunner);
+        if (!hift_p1_58) return false;
+        BaseConfig cfg58;
+        cfg58.nthread    = 8;
+        cfg58.onnx_model = _attr.hift_p1_58;
+        hift_p1_58->load(cfg58);
 
-        MNN::ScheduleConfig config;
-        config.numThread = 2;
-        config.type      = static_cast<MNNForwardType>(MNN_FORWARD_CPU);
-        MNN::BackendConfig backendConfig;
-        backendConfig.precision = (MNN::BackendConfig::PrecisionMode)1;
-        config.backendConfig = &backendConfig;
-
-        hift_p1_50_first = std::shared_ptr<MNN::Interpreter>(MNN::Interpreter::createFromFile( (model_dir+"/hift_p1_50_first.mnn").c_str() ));
-        if(nullptr == hift_p1_50_first)
-        {
-            ALOGE("init mnn model(%s) failed", (model_dir+"/hift_p1_50_first.mnn").c_str());
-            return false;
-        }
-        sess_hift_p1_50_first = hift_p1_50_first->createSession(config);
-
-        hift_p1_58 = std::shared_ptr<MNN::Interpreter>(MNN::Interpreter::createFromFile( (model_dir+"/hift_p1_58.mnn").c_str() ));
-        if(nullptr == hift_p1_58)
-        {
-            ALOGE("init mnn model(%s) failed", (model_dir+"/hift_p1_58.mnn").c_str() );
-            return false;
-        }
-
-        sess_hift_p1_58 = hift_p1_58->createSession(config);
-
-        ALOGI("Token2Wav init ok");
         return true;
     }
 
@@ -227,52 +157,49 @@ class Token2Wav
         flow_embed_selector.Deinit();
     }
 
-    int SpeechToken2Embeds(std::vector<int> & token_ids,  std::vector<float> &token_embeds)
-    {   
-        if(token_embeds.empty() || token_embeds.size() != token_ids.size()* flow_embed_size)
-        {
-            token_embeds.resize(token_ids.size()* flow_embed_size);
+    void SetTimesteps(int n_timesteps)
+    {
+        init_tspan(n_timesteps);
+    }
+
+    int SpeechToken2Embeds(std::vector<int> &token_ids, std::vector<float> &token_embeds)
+    {
+        if (token_embeds.empty() || token_embeds.size() != token_ids.size() * _attr.flow_embed_size) {
+            token_embeds.resize(token_ids.size() * _attr.flow_embed_size);
         }
-        std::vector<unsigned short> speech_embeds_one(flow_embed_size);
-        for (size_t i = 0; i < token_ids.size(); i++)
-        {
+        std::vector<unsigned short> speech_embeds_one(_attr.flow_embed_size);
+        for (size_t i = 0; i < token_ids.size(); i++) {
             flow_embed_selector.getByIndex(token_ids[i], speech_embeds_one.data());
-            for (int j = 0; j < flow_embed_size; j++)
-                {
-                    unsigned int proc = speech_embeds_one[j] << 16;
-                    token_embeds[i * flow_embed_size + j] = *reinterpret_cast<float *>(&proc);
-                }
+            for (int j = 0; j < _attr.flow_embed_size; j++) {
+                unsigned int proc                           = speech_embeds_one[j] << 16;
+                token_embeds[i * _attr.flow_embed_size + j] = *reinterpret_cast<float *>(&proc);
+            }
         }
         return token_embeds.size();
     }
 
-    int infer_flow_encoder(
-        std::vector<float> & token_embeds, std::vector<float> & prompt_feat, std::vector<float> & spk_embeds, int token_len, bool finalize,
-        std::vector<float> & mu, std::vector<float> & spks, std::vector<float> & cond
-        )
+    int infer_flow_encoder(std::vector<float> &token_embeds, std::vector<float> &prompt_feat,
+                           std::vector<float> &spk_embeds, int token_len, bool finalize, std::vector<float> &mu,
+                           std::vector<float> &spks, std::vector<float> &cond)
     {
-        ax_runner_ax650 * model;
-        if(!finalize)
-        {
-            if(token_len == 28)
-            {
+        ax_runner_ax650 *model;
+        if (!finalize) {
+            if (token_len == 28) {
                 model = &flow_encoder_28;
-            }else if(token_len == 53)
-            {
+            } else if (token_len == 53) {
                 model = &flow_encoder_53;
-            }else if(token_len == 78)
-            {
+            } else if (token_len == 78) {
                 model = &flow_encoder_78;
-            }else{
+            } else {
                 return -1;
             }
-        }else if(token_len == 50){
+        } else if (token_len == 50) {
             model = &flow_encoder_50_final;
-        }else{
+        } else {
             return -1;
         }
 
-        void * p = model->get_input("token_embedding").pVirAddr;
+        void *p = model->get_input("token_embedding").pVirAddr;
         memcpy(p, token_embeds.data(), token_embeds.size() * sizeof(float));
         p = model->get_input("prompt_feat").pVirAddr;
         memcpy(p, prompt_feat.data(), prompt_feat.size() * sizeof(float));
@@ -281,24 +208,20 @@ class Token2Wav
 
         model->inference();
 
-        
         auto &output_mu = model->get_output("mu");
-        if(mu.empty())
-        {
+        if (mu.empty()) {
             mu.resize(output_mu.nSize / sizeof(float));
         }
         memcpy(mu.data(), output_mu.pVirAddr, output_mu.nSize);
 
         auto &output_spks = model->get_output("spks");
-        if(spks.empty())
-        {
+        if (spks.empty()) {
             spks.resize(output_spks.nSize / sizeof(float));
         }
         memcpy(spks.data(), output_spks.pVirAddr, output_spks.nSize);
 
         auto &output_cond = model->get_output("cond");
-        if(cond.empty())
-        {
+        if (cond.empty()) {
             cond.resize(output_cond.nSize / sizeof(float));
         }
         memcpy(cond.data(), output_cond.pVirAddr, output_cond.nSize);
@@ -306,25 +229,23 @@ class Token2Wav
         return 0;
     }
 
-    int infer_flow_estimator(
-        std::vector<float> & x, std::vector<float> & mask, std::vector<float> & t,
-        std::vector<float> & mu, std::vector<float> & spks, std::vector<float> & cond, 
-        std::vector<float> & dphi_dt
-        )
+    int infer_flow_estimator(std::vector<float> &x, std::vector<float> &mask, std::vector<float> &t,
+                             std::vector<float> &mu, std::vector<float> &spks, std::vector<float> &cond,
+                             std::vector<float> &dphi_dt)
     {
-        ax_runner_ax650 * model;
-        int len = x.size()/(2*80);
-        if(len == 200){
+        ax_runner_ax650 *model;
+        int len = x.size() / (2 * 80);
+        if (len == 200) {
             model = &flow_estimator_200;
-        }else if(len == 250){
+        } else if (len == 250) {
             model = &flow_estimator_250;
-        }else if(len == 300){
+        } else if (len == 300) {
             model = &flow_estimator_300;
-        }else{
+        } else {
             return -1;
         }
 
-        void * p = model->get_input("x").pVirAddr;
+        void *p = model->get_input("x").pVirAddr;
         memcpy(p, x.data(), x.size() * sizeof(float));
         p = model->get_input("mask").pVirAddr;
         memcpy(p, mask.data(), mask.size() * sizeof(float));
@@ -340,8 +261,7 @@ class Token2Wav
         model->inference();
 
         auto &output_dphi_dt = model->get_output("y");
-        if(dphi_dt.empty() || dphi_dt.size() != output_dphi_dt.nSize / sizeof(float))
-        {
+        if (dphi_dt.empty() || dphi_dt.size() != output_dphi_dt.nSize / sizeof(float)) {
             dphi_dt.resize(output_dphi_dt.nSize / sizeof(float));
         }
         memcpy(dphi_dt.data(), output_dphi_dt.pVirAddr, output_dphi_dt.nSize);
@@ -349,101 +269,76 @@ class Token2Wav
         return 0;
     }
 
-    int infer_hift(std::vector<float> &mel, std::vector<float> &cache_source, 
-                    std::vector<float> & tts_speech, std::vector<float> & tts_source)
+    int infer_hift(std::vector<float> &mel, std::vector<float> &cache_source, std::vector<float> &tts_speech,
+                   std::vector<float> &tts_source)
     {
-        std::shared_ptr<MNN::Interpreter> model_p1;
-        MNN::Session * sess_p1;
-        ax_runner_ax650 * model_p2;
-        int len = mel.size()/(80);
-        
-        if(len == 50 && cache_source.empty())
-        { 
+        std::shared_ptr<BaseRunner> model_p1;
+        ax_runner_ax650 *model_p2;
+        int len = mel.size() / (80);
+
+        if (len == 50 && cache_source.empty()) {
             model_p1 = hift_p1_50_first;
-            sess_p1 = sess_hift_p1_50_first;
             model_p2 = &hift_p2_50_first;
-        }else if(len == 58 && !cache_source.empty())
-        {
+        } else if (len == 58 && !cache_source.empty()) {
             model_p1 = hift_p1_58;
-            sess_p1 = sess_hift_p1_58;
             model_p2 = &hift_p2_58;
-        }else
-        {
+        } else {
             ALOGE("invalid size: %d", len);
             return -1;
         }
 
-        std::vector<int> dims{1, 80, len};
-        auto tensor = MNN::Tensor::create<float>(dims, NULL, MNN::Tensor::CAFFE);
-        auto p_tensor   = tensor->host<float>();
-        auto size   = tensor->size();
-        std::memcpy(p_tensor, mel.data(), size);
-        
-        auto inputTensor = model_p1->getSessionInput(sess_p1, nullptr);
-        inputTensor->copyFromHostTensor(tensor);
-        
-        model_p1->runSession(sess_p1);
-        
-        MNN::Tensor *p_out  = model_p1->getSessionOutput(sess_p1, "s");
-        MNN::Tensor out_host(p_out, p_out->getDimensionType());
-        p_out->copyToHostTensor(&out_host);
-        
-        auto p_s = out_host.host<float>();
-
-        void * p = model_p2->get_input("s").pVirAddr;
-        memcpy(p, p_s, len * 480 * sizeof(float));
-        
-        p = model_p2->get_input("mel").pVirAddr;
+        float *p_input = (float *)model_p1->getInputPtr(0);
+        memcpy(p_input, mel.data(), mel.size() * sizeof(float));
+        model_p1->inference();
+
+        auto p_s = model_p2->get_input("s");
+        memcpy(p_s.pVirAddr, (void *)model_p1->getOutputPtr(0), len * 480 * sizeof(float));
+
+        void *p = (void *)model_p2->get_input("mel").pVirAddr;
         memcpy(p, mel.data(), mel.size() * sizeof(float));
-        
-        if(!cache_source.empty())
-        {
+
+        if (!cache_source.empty()) {
             p = model_p2->get_input("hift_cache_source").pVirAddr;
             memcpy(p, cache_source.data(), cache_source.size() * sizeof(float));
         }
-        
+
         model_p2->inference();
-        
+
         auto &output_speech = model_p2->get_output("audio");
-        if(tts_speech.empty() || tts_speech.size() != output_speech.nSize / sizeof(float))
-        {
+        if (tts_speech.empty() || tts_speech.size() != output_speech.nSize / sizeof(float)) {
             tts_speech.resize(output_speech.nSize / sizeof(float));
         }
         memcpy(tts_speech.data(), output_speech.pVirAddr, output_speech.nSize);
 
         auto &output_source = model_p2->get_output(1);
-        if(tts_source.empty() || tts_source.size() != output_source.nSize / sizeof(float))
-        {
+        if (tts_source.empty() || tts_source.size() != output_source.nSize / sizeof(float)) {
             tts_source.resize(output_source.nSize / sizeof(float));
         }
         memcpy(tts_source.data(), output_source.pVirAddr, output_source.nSize);
-        
+
         return 0;
     }
 
-    int infer_flow_decoder_solve_euler(
-        std::vector<float> & x,  std::vector<float> & mu, std::vector<float> & spks, std::vector<float> & cond, std::vector<float> & mask,
-        std::vector<float> & mel
-    )
+    int infer_flow_decoder_solve_euler(std::vector<float> &x, std::vector<float> &mu, std::vector<float> &spks,
+                                       std::vector<float> &cond, std::vector<float> &mask, std::vector<float> &mel)
     {
-        int len = mu.size()/80;
+        int len = mu.size() / 80;
 
-        float t = t_span[0];
+        float t  = t_span[0];
         float dt = t_span[1] - t_span[0];
 
-        std::vector<float> x_in(2*80*len, 0);
-        std::vector<float> mask_in(2*1*len, 0);
-        std::vector<float> mu_in(2*80*len,0);
-        std::vector<float> t_in(2,0);
-        std::vector<float> spks_in(2*80, 0);
-        std::vector<float> cond_in(2*80*len, 0);
-        for(int step=1; step<t_span.size(); step++)
-        {   
+        std::vector<float> x_in(2 * 80 * len, 0);
+        std::vector<float> mask_in(2 * 1 * len, 0);
+        std::vector<float> mu_in(2 * 80 * len, 0);
+        std::vector<float> t_in(2, 0);
+        std::vector<float> spks_in(2 * 80, 0);
+        std::vector<float> cond_in(2 * 80 * len, 0);
+        for (int step = 1; step < t_span.size(); step++) {
             memcpy(x_in.data(), x.data(), x.size() * sizeof(float));
-            memcpy(x_in.data()+x.size(),  x.data(), x.size() * sizeof(float));
+            memcpy(x_in.data() + x.size(), x.data(), x.size() * sizeof(float));
 
             memcpy(mask_in.data(), mask.data(), mask.size() * sizeof(float));
-            memcpy(mask_in.data()+mask.size(), mask.data(), mask.size() * sizeof(float));
+            memcpy(mask_in.data() + mask.size(), mask.data(), mask.size() * sizeof(float));
 
             memcpy(mu_in.data(), mu.data(), mu.size() * sizeof(float));
 
@@ -455,41 +350,33 @@ class Token2Wav
 
             std::vector<float> dphi_dt;
             int ret = infer_flow_estimator(x_in, mask_in, t_in, mu_in, spks_in, cond_in, dphi_dt);
-            if(ret != 0)
-            {
+            if (ret != 0) {
                 return ret;
             }
 
-            for(int i=0; i<80*len; i++)
-            {
-
-                dphi_dt[i] = (1.0 + inference_cfg_rate) * dphi_dt[i] - inference_cfg_rate * dphi_dt[80 * len + i];
+            for (int i = 0; i < 80 * len; i++) {
+                dphi_dt[i] =
+                    (1.0 + _attr.inference_cfg_rate) * dphi_dt[i] - _attr.inference_cfg_rate * dphi_dt[80 * len + i];
                 x[i] = x[i] + dt * dphi_dt[i];
             }
-            
+
             t = t + dt;
 
-            if(step < t_span.size()-1)
-            {
-                dt = t_span[step+1] - t;
-            }
-            else{
-                if(mel.empty() || mel.size()!=x.size())
-                {
+            if (step < t_span.size() - 1) {
+                dt = t_span[step + 1] - t;
+            } else {
+                if (mel.empty() || mel.size() != x.size()) {
                     mel.resize(x.size());
                 }
                 memcpy(mel.data(), x.data(), x.size() * sizeof(float));
             }
-
         }
 
         return 0;
     }
 
-    int infer_flow_decoder(
-        std::vector<float> & mu, std::vector<float> & spks, std::vector<float> & cond, std::vector<float> & mask,
-        std::vector<float> & mel
-    )
+    int infer_flow_decoder(std::vector<float> &mu, std::vector<float> &spks, std::vector<float> &cond,
+                           std::vector<float> &mask, std::vector<float> &mel)
     {
         std::vector<float> z;
         z.insert(z.end(), rand_noise.begin(), rand_noise.begin() + mu.size());
@@ -498,55 +385,51 @@ class Token2Wav
         return ret;
     }
 
-    std::vector<float> infer_flow(
-        std::vector<float> & token_embeds, std::vector<float> & prompt_feat, std::vector<float> & spk_embeds, int token_len, bool finalize
-    )
+    std::vector<float> infer_flow(std::vector<float> &token_embeds, std::vector<float> &prompt_feat,
+                                  std::vector<float> &spk_embeds, int token_len, bool finalize)
     {
-        int ret; 
+        int ret;
         int len;
         std::vector<float> mu;
         std::vector<float> spks;
         std::vector<float> cond;
 
         ret = infer_flow_encoder(token_embeds, prompt_feat, spk_embeds, token_len, finalize, mu, spks, cond);
-        if(ret != 0)
-        {
+        if (ret != 0) {
             return std::vector<float>{};
         }
 
-        len = mu.size()/80;
+        len = mu.size() / 80;
 
         std::vector<float> mask(len, 1.0);
         std::vector<float> all_mel;
-        
+
         ret = infer_flow_decoder(mu, spks, cond, mask, all_mel);
-        if(ret != 0)
-        {
+        if (ret != 0) {
             return std::vector<float>{};
         }
-        
-        int len_mel1 = prompt_feat.size()/80;
-        int len_mel2 = all_mel.size()/80 - len_mel1;
-        
+
+        int len_mel1 = prompt_feat.size() / 80;
+        int len_mel2 = all_mel.size() / 80 - len_mel1;
+
         std::vector<float> mel(len_mel2 * 80, 0);
-        auto result = slice_3d_last_dim_from<float>(all_mel, 1, 80, all_mel.size()/80, len_mel1);
-        
+        auto result = slice_3d_last_dim_from<float>(all_mel, 1, 80, all_mel.size() / 80, len_mel1);
+
         return result;
     }
 
-    void fade_in_out(std::vector<float>& fade_in_mel_data,
-                 const std::vector<float>& fade_out_mel_data,
-                 const std::vector<float>& window) {
-
+    void fade_in_out(std::vector<float> &fade_in_mel_data, const std::vector<float> &fade_out_mel_data,
+                     const std::vector<float> &window)
+    {
         // --- Constants based on window = np.hamming(2 * 8 * 480) ---
-        const size_t WINDOW_SIZE = 2 * 8 * 480; // 7680
-        const size_t MEL_OVERLAP_LEN = WINDOW_SIZE / 2; // 3840
+        const size_t WINDOW_SIZE     = 2 * 8 * 480;      // 7680
+        const size_t MEL_OVERLAP_LEN = WINDOW_SIZE / 2;  // 3840
         // dim0 is implicitly 1 for both inputs
-        size_t dim1_in = fade_in_mel_data.size();
+        size_t dim1_in  = fade_in_mel_data.size();
         size_t dim1_out = fade_out_mel_data.size();
         // --- Input Validation ---
         // For 2D arrays [1, L], the 1D vector size is just L.
-        
+
         if (window.size() != WINDOW_SIZE) {
             throw std::invalid_argument("window size (" + std::to_string(window.size()) +
                                         ") does not match expected size (7680).");
@@ -567,7 +450,7 @@ class Token2Wav
         for (size_t i = 0; i < MEL_OVERLAP_LEN; ++i) {
             // Indices are simply 'i' for the start of fade_in_mel
             // and 'dim1_out - MEL_OVERLAP_LEN + i' for the end of fade_out_mel
-            const size_t in_idx = i;
+            const size_t in_idx  = i;
             const size_t out_idx = dim1_out - MEL_OVERLAP_LEN + i;
 
             // Perform the weighted sum: result = in_val * win_in + out_val * win_out
@@ -575,8 +458,8 @@ class Token2Wav
             // out_val = fade_out_mel_data[out_idx]
             // win_in = window[i]
             // win_out = window[MEL_OVERLAP_LEN + i]
-            fade_in_mel_data[in_idx] = fade_in_mel_data[in_idx] * window[i] +
-                                    fade_out_mel_data[out_idx] * window[MEL_OVERLAP_LEN + i];
+            fade_in_mel_data[in_idx] =
+                fade_in_mel_data[in_idx] * window[i] + fade_out_mel_data[out_idx] * window[MEL_OVERLAP_LEN + i];
         }
         // fade_in_mel_data is now modified in-place with the faded result.
     }
@@ -586,109 +469,97 @@ class Token2Wav
         std::unordered_map<std::string, std::vector<float>>().swap(hift_cache_dict);
     }
 
-    std::vector<float> infer(std::vector<int> & text_speech_token, std::vector<float> & prompt_speech_embeds, std::vector<float> & prompt_feat,  
-                std::vector<float> & spk_embeds, int token_offset, bool finalize)
+    std::vector<float> infer(std::vector<int> &text_speech_token, std::vector<float> &prompt_speech_embeds,
+                             std::vector<float> &prompt_feat, std::vector<float> &spk_embeds, int token_offset,
+                             bool finalize)
     {
         int ret = 0;
-        std::vector<float> speech_embeds( text_speech_token.size()*flow_embed_size + prompt_speech_embeds.size(), 0.0f);
-        std::vector<unsigned short> speech_embeds_one(flow_embed_size, 0);
+        std::vector<float> speech_embeds(text_speech_token.size() * _attr.flow_embed_size + prompt_speech_embeds.size(),
+                                         0.0f);
+        std::vector<unsigned short> speech_embeds_one(_attr.flow_embed_size, 0);
 
         memcpy(speech_embeds.data(), prompt_speech_embeds.data(), prompt_speech_embeds.size() * sizeof(float));
 
-        for (size_t i = 0; i < text_speech_token.size(); i++)
-        {
+        for (size_t i = 0; i < text_speech_token.size(); i++) {
             flow_embed_selector.getByIndex(text_speech_token[i], speech_embeds_one.data());
 
-            for (int j = 0; j < flow_embed_size; j++)
-                {
-                    unsigned int proc = speech_embeds_one[j] << 16;
-                    speech_embeds[prompt_speech_embeds.size() + i * flow_embed_size + j] = *reinterpret_cast<float *>(&proc);
-                }
+            for (int j = 0; j < _attr.flow_embed_size; j++) {
+                unsigned int proc = speech_embeds_one[j] << 16;
+                speech_embeds[prompt_speech_embeds.size() + i * _attr.flow_embed_size + j] =
+                    *reinterpret_cast<float *>(&proc);
+            }
         }
 
         std::vector<float> mel;
-    
-        mel = infer_flow(speech_embeds, prompt_feat, spk_embeds, text_speech_token.size(), finalize);   
+
+        mel = infer_flow(speech_embeds, prompt_feat, spk_embeds, text_speech_token.size(), finalize);
 
         std::vector<float> tts_mel;
-        int neg_offset=0, start;
-        if(finalize)
-        {
-            neg_offset = token_offset * token_mel_ratio - mel.size()/80;
-            start = - token_hop_len * token_mel_ratio;
-        }
-        else{
-            start = std::min( int(token_offset / token_hop_len), max_infer_chunk_num-1) * token_hop_len * token_mel_ratio;
+        int neg_offset = 0, start;
+        if (finalize) {
+            neg_offset = token_offset * _attr.token_mel_ratio - mel.size() / 80;
+            start      = -_attr.token_hop_len * _attr.token_mel_ratio;
+        } else {
+            start = std::min(int(token_offset / _attr.token_hop_len), _attr.max_infer_chunk_num - 1) *
+                    _attr.token_hop_len * _attr.token_mel_ratio;
         }
-        
-        tts_mel = slice_3d_last_dim_from<float>(mel, 1, 80, mel.size()/80, start);
-        
+
+        tts_mel = slice_3d_last_dim_from<float>(mel, 1, 80, mel.size() / 80, start);
+
         std::vector<float> hift_cache_source;
         std::vector<float> tts_mel1;
         std::vector<float> speech, source, tts_speech;
-        if (!hift_cache_dict.empty())
-        {
+        if (!hift_cache_dict.empty()) {
             auto hift_cache_mel = hift_cache_dict["mel"];
-            hift_cache_source = hift_cache_dict["source"];
-            tts_mel1 = concat_3d_dim2<float>(hift_cache_mel, 1, 80, hift_cache_mel.size()/80, tts_mel, 1, 80, tts_mel.size()/80);
-        }
-        else{
+            hift_cache_source   = hift_cache_dict["source"];
+            tts_mel1 = concat_3d_dim2<float>(hift_cache_mel, 1, 80, hift_cache_mel.size() / 80, tts_mel, 1, 80,
+                                             tts_mel.size() / 80);
+        } else {
             tts_mel1 = tts_mel;
-        }   
+        }
         ret = infer_hift(tts_mel1, hift_cache_source, speech, source);
-        
-        if(ret != 0){
+
+        if (ret != 0) {
             ALOGE("failed");
             return std::vector<float>{};
         }
 
-        if(!finalize)
-        {
-            
-            if(!hift_cache_dict.empty())
-            {
+        if (!finalize) {
+            if (!hift_cache_dict.empty()) {
                 fade_in_out(speech, hift_cache_dict["speech"], speech_window);
             }
 
-            hift_cache_dict["mel"] = slice_3d_last_dim_from<float>(tts_mel1, 1, 80, tts_mel1.size()/80, -mel_cache_len);
+            hift_cache_dict["mel"] =
+                slice_3d_last_dim_from<float>(tts_mel1, 1, 80, tts_mel1.size() / 80, -_attr.mel_cache_len);
 
             int offset = speech.size();
-            if(speech.size() > source_cache_len)
-            {
-                offset = source_cache_len;
+            if (speech.size() > _attr.source_cache_len) {
+                offset = _attr.source_cache_len;
             }
 
-            hift_cache_dict["source"].assign(source.end()-offset, source.end());
-            hift_cache_dict["speech"].assign(speech.end()-offset, speech.end());
-            tts_speech.assign(speech.begin(), speech.end()-offset);
-
-        }
-        else{
+            hift_cache_dict["source"].assign(source.end() - offset, source.end());
+            hift_cache_dict["speech"].assign(speech.end() - offset, speech.end());
+            tts_speech.assign(speech.begin(), speech.end() - offset);
 
-            if(speech.size() < source_cache_len){
+        } else {
+            if (speech.size() < _attr.source_cache_len) {
                 tts_speech.assign(speech.begin(), speech.end());
-            }
-            else if (- neg_offset*480 >= source_cache_len)
-            {
-                tts_speech.assign(speech.end() + neg_offset*480, speech.end());
+            } else if (-neg_offset * 480 >= _attr.source_cache_len) {
+                tts_speech.assign(speech.end() + neg_offset * 480, speech.end());
 
-                if(!hift_cache_dict.empty())
-                {
+                if (!hift_cache_dict.empty()) {
                     fade_in_out(tts_speech, hift_cache_dict["speech"], speech_window);
                 }
-            }
-            else{
-                tts_speech.assign(speech.end()-source_cache_len, speech.end());
+            } else {
+                tts_speech.assign(speech.end() - _attr.source_cache_len, speech.end());
 
-                if(!hift_cache_dict.empty())
-                {
+                if (!hift_cache_dict.empty()) {
                     fade_in_out(tts_speech, hift_cache_dict["speech"], speech_window);
                 }
 
-                int offset = speech.size() + neg_offset*480 - (speech.size() - source_cache_len);
+                int offset = speech.size() + neg_offset * 480 - (speech.size() - _attr.source_cache_len);
                 tts_speech.assign(tts_speech.begin() + offset, tts_speech.end());
             }
-            
         }
 
         return tts_speech;
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/utils.hpp b/projects/llm_framework/main_cosy_voice/src/runner/utils/utils.hpp
index 58e9bb0d..6d077015 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/utils/utils.hpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/utils.hpp
@@ -3,71 +3,52 @@
 #include <fstream>
 #include <vector>
 #include <iomanip>
-#include <type_traits> // 用于类型检查
+#include <type_traits>
 #include <iostream>
 #include <fstream>
 #include <vector>
 #include <string>
 #include <sstream>
 #include <type_traits>
-#include <stdexcept> // 用于异常处理
+#include <stdexcept>
 
-// 函数模板：支持任意元素类型的vector
 template <typename T>
-void savetxt(const std::string& filename, 
-            const std::vector<T>& data,
-            char delimiter = ' ', 
-            int precision = 6)  // 默认精度调整为6位
+void savetxt(const std::string& filename, const std::vector<T>& data, char delimiter = ' ', int precision = 6)
 {
     std::ofstream outfile(filename);
     if (!outfile.is_open()) {
-        throw std::runtime_error("无法打开文件: " + filename);
+        throw std::runtime_error("Unable to open file: " + filename);
     }
-
-    // 保存流的原始格式状态
     std::ios_base::fmtflags original_flags = outfile.flags();
-
-    // 仅对浮点类型启用科学计数法
     if constexpr (std::is_floating_point_v<T>) {
         outfile << std::scientific << std::setprecision(precision);
     }
-
-    // 优化输出：先输出第一个元素，再循环输出后续元素
     if (!data.empty()) {
         outfile << data[0];
         for (size_t i = 1; i < data.size(); ++i) {
             outfile << delimiter << data[i];
         }
     }
-
-    // 恢复流的原始格式状态
     outfile.flags(original_flags);
     outfile.close();
 }
 
-
-
-// 函数模板：读取文本文件到 vector<T>
 template <typename T>
-int readtxt(const std::string& filename, std::vector<T>& data) {
+int readtxt(const std::string& filename, std::vector<T>& data)
+{
     std::ifstream file(filename);
     if (!file.is_open()) {
-        std::cerr << "错误：无法打开文件 " << filename << std::endl;
+        std::cerr << "Error: Unable to open file " << filename << std::endl;
         return -1;
     }
 
-    
     std::string line;
     while (std::getline(file, line)) {
-        // 跳过空行
         if (line.empty()) continue;
-
         std::istringstream iss(line);
         T value;
-        // 解析行中的每个数值
         while (iss >> value) {
             data.push_back(value);
-            // 跳过分隔符（兼容逗号、分号等）
             if (iss.peek() == ',' || iss.peek() == ';') iss.ignore();
         }
     }
@@ -75,23 +56,19 @@ int readtxt(const std::string& filename, std::vector<T>& data) {
     return 0;
 }
 
-std::vector<float> linspace(float start, float end, std::size_t num_steps) {
+std::vector<float> linspace(float start, float end, std::size_t num_steps)
+{
     if (num_steps == 0) {
-        return {}; // 返回空向量
+        return {};
     }
     if (num_steps == 1) {
-        return {start}; // 如果只需要一个点，返回起始值
+        return {start};
     }
-
     std::vector<float> result(num_steps);
-    float step_size = (end - start) / (num_steps - 1); // 计算步长
-
+    float step_size = (end - start) / (num_steps - 1);
     for (std::size_t i = 0; i < num_steps; ++i) {
         result[i] = start + i * step_size;
     }
-
-    // 确保最后一个值精确等于 end，避免浮点数精度带来的误差
     result.back() = end;
-
     return result;
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/utils/wav.hpp b/projects/llm_framework/main_cosy_voice/src/runner/utils/wav.hpp
index dbbd979b..11573c68 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/utils/wav.hpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/utils/wav.hpp
@@ -2,97 +2,101 @@
 #include <fstream>
 #include <iostream>
 #include <cstdint>
-#include <cstring> // for memcpy
-#include <algorithm> // for std::clamp (C++17)
+#include <cstring>
+#include <algorithm>
+#include <filesystem>
+namespace fs = std::filesystem;
 
-// 将 32 位整数以小端序写入文件
-void write_little_endian_32(std::ofstream& file, uint32_t value) {
+void write_little_endian_32(std::ofstream& file, uint32_t value)
+{
     file.put(static_cast<char>(value & 0xFF));
     file.put(static_cast<char>((value >> 8) & 0xFF));
     file.put(static_cast<char>((value >> 16) & 0xFF));
     file.put(static_cast<char>((value >> 24) & 0xFF));
 }
 
-// 将 16 位整数以小端序写入文件
-void write_little_endian_16(std::ofstream& file, uint16_t value) {
+void write_little_endian_16(std::ofstream& file, uint16_t value)
+{
     file.put(static_cast<char>(value & 0xFF));
     file.put(static_cast<char>((value >> 8) & 0xFF));
 }
 
-// 将 32 位浮点数以小端序写入文件 (IEEE 754)
-void write_little_endian_float32(std::ofstream& file, float value) {
-    // IEEE 754 float 在大多数系统上已经是小端序
-    // 但为了确保跨平台兼容性，我们显式处理字节序
+void write_little_endian_float32(std::ofstream& file, float value)
+{
     static_assert(sizeof(float) == 4, "Float must be 32 bits");
     char bytes[4];
     std::memcpy(bytes, &value, 4);
-    // bytes[0] 是最低有效字节 (LSB)
     file.put(bytes[0]);
     file.put(bytes[1]);
     file.put(bytes[2]);
     file.put(bytes[3]);
 }
 
-// 将 float 范围 [-1.0, 1.0] 转换为 16-bit signed integer [-32768, 32767]
-int16_t float_to_int16(float sample) {
-    // 确保输入在有效范围内
-    // C++17 std::clamp, 或手动实现
-    // sample = std::clamp(sample, -1.0f, 1.0f);
+int16_t float_to_int16(float sample)
+{
     if (sample > 1.0f) sample = 1.0f;
     if (sample < -1.0f) sample = -1.0f;
-    
-    // 转换
     return static_cast<int16_t>(sample * 32767.0f);
 }
 
+bool ensure_directory_exists(const std::string& filename)
+{
+    try {
+        fs::path filepath(filename);
+        fs::path dir = filepath.parent_path();
+        if (!dir.empty() && !fs::exists(dir)) {
+            fs::create_directories(dir);
+        }
+        return true;
+    } catch (const fs::filesystem_error& e) {
+        std::cerr << "Failed to create directories: " << e.what() << std::endl;
+        return false;
+    }
+}
 
-bool saveVectorAsWavFloat(const std::vector<float>& audio_data, const std::string& filename, int sample_rate, int channels) {
+bool saveVectorAsWavFloat(const std::vector<float>& audio_data, const std::string& filename, int sample_rate,
+                          int channels)
+{
     if (audio_data.empty() || channels <= 0 || sample_rate <= 0) {
         std::cerr << "Invalid input parameters." << std::endl;
         return false;
     }
 
+    if (!ensure_directory_exists(filename)) return false;
+
     std::ofstream file(filename, std::ios::binary);
     if (!file.is_open()) {
         std::cerr << "Could not open file " << filename << " for writing." << std::endl;
         return false;
     }
 
-    // --- 计算关键参数 ---
-    uint32_t num_samples = static_cast<uint32_t>(audio_data.size());
-    uint16_t bits_per_sample = 32; // IEEE Float
-    uint32_t byte_rate = sample_rate * channels * (bits_per_sample / 8);
-    uint16_t block_align = channels * (bits_per_sample / 8);
+    uint32_t num_samples     = static_cast<uint32_t>(audio_data.size());
+    uint16_t bits_per_sample = 32;  // IEEE Float
+    uint32_t byte_rate       = sample_rate * channels * (bits_per_sample / 8);
+    uint16_t block_align     = channels * (bits_per_sample / 8);
     uint32_t data_chunk_size = num_samples * (bits_per_sample / 8);
-    uint32_t riff_chunk_size = 4 + (8 + 16) + (8 + data_chunk_size); // fmt chunk is 24 bytes including header
+    uint32_t riff_chunk_size = 4 + (8 + 16) + (8 + data_chunk_size);
 
-    // --- 写入 RIFF Header ---
     file.write("RIFF", 4);
     write_little_endian_32(file, riff_chunk_size);
     file.write("WAVE", 4);
-
-    // --- 写入 fmt Subchunk ---
     file.write("fmt ", 4);
-    write_little_endian_32(file, 16); // Subchunk1Size for PCM
-    write_little_endian_16(file, 3); // AudioFormat: 3 = IEEE Float
+    write_little_endian_32(file, 16);
+    write_little_endian_16(file, 3);
     write_little_endian_16(file, static_cast<uint16_t>(channels));
     write_little_endian_32(file, sample_rate);
     write_little_endian_32(file, byte_rate);
     write_little_endian_16(file, block_align);
     write_little_endian_16(file, bits_per_sample);
-
-    // --- 写入 data Subchunk ---
     file.write("data", 4);
     write_little_endian_32(file, data_chunk_size);
 
-    // --- 写入音频数据 ---
     for (const float& sample : audio_data) {
         write_little_endian_float32(file, sample);
     }
 
     if (file.fail()) {
         std::cerr << "Error occurred while writing audio data." << std::endl;
-        file.close();
         return false;
     }
 
@@ -101,60 +105,54 @@ bool saveVectorAsWavFloat(const std::vector<float>& audio_data, const std::strin
     return true;
 }
 
-bool saveVectorAsWavInt16(const std::vector<float>& audio_data, const std::string& filename, int sample_rate, int channels) {
+bool saveVectorAsWavInt16(const std::vector<float>& audio_data, const std::string& filename, int sample_rate,
+                          int channels)
+{
     if (audio_data.empty() || channels <= 0 || sample_rate <= 0) {
         std::cerr << "Invalid input parameters." << std::endl;
         return false;
     }
 
+    if (!ensure_directory_exists(filename)) return false;
+
     std::ofstream file(filename, std::ios::binary);
     if (!file.is_open()) {
         std::cerr << "Could not open file " << filename << " for writing." << std::endl;
         return false;
     }
 
-    // --- 计算关键参数 ---
-    uint32_t num_samples = static_cast<uint32_t>(audio_data.size());
-    uint16_t bits_per_sample = 16; // 16-bit signed integer
-    uint32_t byte_rate = sample_rate * channels * (bits_per_sample / 8);
-    uint16_t block_align = channels * (bits_per_sample / 8);
+    uint32_t num_samples     = static_cast<uint32_t>(audio_data.size());
+    uint16_t bits_per_sample = 16;
+    uint32_t byte_rate       = sample_rate * channels * (bits_per_sample / 8);
+    uint16_t block_align     = channels * (bits_per_sample / 8);
     uint32_t data_chunk_size = num_samples * (bits_per_sample / 8);
-    uint32_t riff_chunk_size = 4 + (8 + 16) + (8 + data_chunk_size); // fmt chunk is 24 bytes including header
+    uint32_t riff_chunk_size = 4 + (8 + 16) + (8 + data_chunk_size);
 
-    // --- 写入 RIFF Header ---
     file.write("RIFF", 4);
     write_little_endian_32(file, riff_chunk_size);
     file.write("WAVE", 4);
-
-    // --- 写入 fmt Subchunk ---
     file.write("fmt ", 4);
-    write_little_endian_32(file, 16); // Subchunk1Size for PCM
-    write_little_endian_16(file, 1); // AudioFormat: 1 = PCM (Integer)
+    write_little_endian_32(file, 16);
+    write_little_endian_16(file, 1);
     write_little_endian_16(file, static_cast<uint16_t>(channels));
     write_little_endian_32(file, sample_rate);
     write_little_endian_32(file, byte_rate);
     write_little_endian_16(file, block_align);
     write_little_endian_16(file, bits_per_sample);
-
-    // --- 写入 data Subchunk ---
     file.write("data", 4);
     write_little_endian_32(file, data_chunk_size);
 
-    // --- 写入音频数据 ---
     for (const float& sample : audio_data) {
         int16_t int_sample = float_to_int16(sample);
-        write_little_endian_16(file, static_cast<uint16_t>(int_sample)); // int16_t -> uint16_t for raw bytes
+        write_little_endian_16(file, static_cast<uint16_t>(int_sample));
     }
 
     if (file.fail()) {
         std::cerr << "Error occurred while writing audio data." << std::endl;
-        file.close();
         return false;
     }
 
     file.close();
     std::cout << "Successfully saved audio to " << filename << " (16-bit Integer PCM)." << std::endl;
     return true;
-}
-
-
+}
\ No newline at end of file
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index c36b5c51..581d1817 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -150,58 +150,67 @@ def create_lib_deb(package_name, version, src_folder, revision = 'm5stack1'):
     shutil.rmtree(deb_folder)
     return package_name + " creat success!"
 
-def create_data_deb(package_name, version, src_folder, revision = 'm5stack1', depends = 'lib-llm (>= 1.6)'):
+def create_data_deb(package_name, version, src_folder, revision='m5stack1', depends='lib-llm (>= 1.6)'):
     deb_file = f"{package_name}_{version}-{revision}_arm64.deb"
-    deb_folder = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'debian-{}'.format(package_name))
+    deb_folder = os.path.join(os.path.dirname(os.path.abspath(__file__)), f'debian-{package_name}')
+    
     if os.path.exists(deb_folder):
         shutil.rmtree(deb_folder)
-    os.makedirs(deb_folder, exist_ok = True)
+    os.makedirs(deb_folder, exist_ok=True)
 
     zip_file = f"m5stack_{package_name[0:4]+package_name[10:]}_{version}_data.tar.gz"
     down_url = f"https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_{package_name[0:4]+package_name[10:]}_{version}_data.tar.gz"
     zip_file_extrpath = f"m5stack_{package_name}_{version}_data"
+
     if not os.path.exists(zip_file_extrpath):
-        # Downloading via HTTP (more common)
         if not os.path.exists(zip_file):
             response = requests.get(down_url)
             if response.status_code == 200:
                 with open(zip_file, 'wb') as file:
                     file.write(response.content)
             else:
-                print("{} down failed".format(down_url))
+                print(f"{down_url} download failed.")
         with tarfile.open(zip_file, 'r:gz') as tar:
             tar.extractall(path=zip_file_extrpath)
-        print("The {} download successful.".format(down_url))
+        print(f"The {down_url} download successful.")
+
     if os.path.exists(zip_file_extrpath):
         shutil.copytree(zip_file_extrpath, os.path.join(deb_folder, 'opt/m5stack/data'))
 
     RED = "\033[31m"
     RESET = "\033[0m"
-    os.makedirs(os.path.join(deb_folder, 'opt/m5stack/data/models'), exist_ok = True)
-    mode_config_file = os.path.join(src_folder,'mode_{}.json'.format(package_name[10:]))
+    os.makedirs(os.path.join(deb_folder, 'opt/m5stack/data/models'), exist_ok=True)
+
+    mode_config_file = os.path.join(src_folder, f'mode_{package_name[10:]}.json')
     if os.path.exists(mode_config_file):
-        shutil.copy2(mode_config_file, os.path.join(deb_folder, 'opt/m5stack/data/models', 'mode_{}.json'.format(package_name[10:])))
+        shutil.copy2(mode_config_file, os.path.join(deb_folder, 'opt/m5stack/data/models', f'mode_{package_name[10:]}.json'))
+
         try:
             with open(mode_config_file, 'r', encoding='utf-8') as file:
                 data = json.load(file)
-            for scripts_file in data['mode_param']['ext_scripts']:
+
+            for scripts_file in data['mode_param'].get('ext_scripts', []):
                 src_path = os.path.join(src_folder, scripts_file)
-                dst_path = os.path.join(deb_folder, 'opt/m5stack/scripts', scripts_file)
-                if os.path.exists(src_path):
-                    if os.path.isdir(src_path):
-                        # 拷贝整个文件夹
-                        shutil.copytree(src_path, dst_path, dirs_exist_ok=True)
-                    else:
-                        # 确保目标目录存在
-                        os.makedirs(os.path.dirname(dst_path), exist_ok=True)
-                        shutil.copy2(src_path, dst_path)
-        except:
-            pass
+                dest_dir = os.path.join(deb_folder, 'opt/m5stack/scripts')
+                os.makedirs(dest_dir, exist_ok=True)
+
+                if os.path.isfile(src_path):
+                    shutil.copy2(src_path, os.path.join(dest_dir, os.path.basename(scripts_file)))
+                    print(f"Copied file: {scripts_file}")
+                elif os.path.isdir(src_path):
+                    shutil.copytree(src_path, os.path.join(dest_dir, os.path.basename(scripts_file)), dirs_exist_ok=True)
+                    print(f"Copied folder: {scripts_file}")
+                else:
+                    print(f"Warning: {scripts_file} not found in {src_folder}")
+
+        except Exception as e:
+            print(f"Error reading ext_scripts: {e}")
+
     else:
-        print(RED, mode_config_file, " miss", RESET)
+        print(RED, mode_config_file, " missing", RESET)
 
-    os.makedirs(os.path.join(deb_folder, 'DEBIAN'), exist_ok = True)
-    with open(os.path.join(deb_folder, 'DEBIAN/control'),'w') as f:
+    os.makedirs(os.path.join(deb_folder, 'DEBIAN'), exist_ok=True)
+    with open(os.path.join(deb_folder, 'DEBIAN/control'), 'w') as f:
         f.write(f'Package: {package_name}\n')
         f.write(f'Version: {version}\n')
         f.write(f'Architecture: arm64\n')
@@ -213,34 +222,34 @@ def create_data_deb(package_name, version, src_folder, revision = 'm5stack1', de
         f.write(f'Homepage: https://www.m5stack.com\n')
         if deb_file.startswith('llm-model-'):
             deb_name = deb_file[:deb_file.find('_')]
-            old_deb_name = deb_name.replace('model-','').lower()
-            f.write(f'Conflicts: {old_deb_name}\n')   
-        f.write(f'Packaged-Date: {datetime.now().strftime("%Y-%m-%d %H:%M:%S")}\n')     
+            old_deb_name = deb_name.replace('model-', '').lower()
+            f.write(f'Conflicts: {old_deb_name}\n')
+        f.write(f'Packaged-Date: {datetime.now().strftime("%Y-%m-%d %H:%M:%S")}\n')
         f.write(f'Description: llm-module\n')
         f.write(f' bsp.\n')
-    with open(os.path.join(deb_folder, 'DEBIAN/postinst'),'w') as f:
-        f.write(f'#!/bin/sh\n')
-        f.write(f'exit 0\n')
-    with open(os.path.join(deb_folder, 'DEBIAN/prerm'),'w') as f:
-        f.write(f'#!/bin/sh\n')
-        f.write(f'exit 0\n')
+
+    with open(os.path.join(deb_folder, 'DEBIAN/postinst'), 'w') as f:
+        f.write('#!/bin/sh\nexit 0\n')
+    with open(os.path.join(deb_folder, 'DEBIAN/prerm'), 'w') as f:
+        f.write('#!/bin/sh\nexit 0\n')
     os.chmod(os.path.join(deb_folder, 'DEBIAN/postinst'), 0o755)
     os.chmod(os.path.join(deb_folder, 'DEBIAN/prerm'), 0o755)
+
     subprocess.run(["dpkg-deb", "-b", deb_folder, deb_file], check=True)
     print(f"Debian package created: {deb_file}")
+
     shutil.rmtree(deb_folder)
     return package_name + " creat success!"
 
 def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', depends = 'lib-llm (>= 1.8)'):
     bin_files = glob.glob(os.path.join(src_folder, package_name.replace("-", "_") + "-*"))
-    version_info = 0.0
-    print(os.path.join(src_folder, package_name + "-*"))
+    version_info = "0"
     if bin_files:
         for bin_file in bin_files:
-            version_info = float(bin_file.split('-')[-1])
-            if float(bin_file.split('-')[-1]) > version_info:
-                version_info = float(bin_file.split('-')[-1])
-        version = str(version_info)
+            ver = bin_file.split('-')[-1]
+            if ver > version_info:
+                version_info = ver
+        version = version_info
 
     deb_file = f"{package_name}_{version}-{revision}_arm64.deb"
     deb_folder = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'debian-{}'.format(package_name))
@@ -273,7 +282,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         cosy_voice_dir = os.path.join(src_folder, 'cosy-voice')
         if os.path.exists(cosy_voice_dir):
             shutil.copytree(cosy_voice_dir, os.path.join(deb_folder, 'opt/m5stack/lib/cosy-voice'))
-
+    
     bin_file_name = package_name.replace("-", "_")
     if version_info != 0.0:
         bin_file_name = package_name.replace("-", "_") + f'-{version}'
@@ -391,7 +400,6 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         # TTS model
         'llm-model-single-speaker-english-fast':[create_data_deb,'llm-model-single-speaker-english-fast', '0.3', src_folder, revision],
         'llm-model-single-speaker-fast':[create_data_deb,'llm-model-single-speaker-fast', '0.3', src_folder, revision],
-        'llm-model-CosyVoice2-0.5B-ax650':[create_data_deb,'llm-model-CosyVoice2-0.5B-ax650', '0.6', src_folder, revision],
         # VAD model
         'llm-model-silero-vad':[create_data_deb,'llm-model-silero-vad', '0.4', src_folder, revision],
         # MeloTTS model
@@ -413,6 +421,9 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-model-melotts-en-us-ax650':[create_data_deb,'llm-model-melotts-en-us-ax650', '0.6', src_folder, revision],
         'llm-model-melotts-ja-jp-ax650':[create_data_deb,'llm-model-melotts-ja-jp-ax650', '0.6', src_folder, revision],
         'llm-model-melotts-es-es-ax650':[create_data_deb,'llm-model-melotts-es-es-ax650', '0.5', src_folder, revision],
+        # CosyVoice2 model
+        ## AX650
+        'llm-model-CosyVoice2-0.5B-ax650':[create_data_deb,'llm-model-CosyVoice2-0.5B-ax650', '0.6', src_folder, revision],
         # Yolo model
         ## AX630C
         'llm-model-yolo11n':[create_data_deb,'llm-model-yolo11n', data_version, src_folder, revision],

From 4e3d7f3666cb9e3ec0eb1889d8a6555b6220ad59 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Tue, 4 Nov 2025 18:13:37 +0800
Subject: [PATCH 50/79] [update] fix llm generate bug

---
 .../main_cosy_voice/src/main.cpp              | 40 +++++--------------
 .../main_cosy_voice/src/runner/LLM.hpp        |  6 ++-
 2 files changed, 14 insertions(+), 32 deletions(-)

diff --git a/projects/llm_framework/main_cosy_voice/src/main.cpp b/projects/llm_framework/main_cosy_voice/src/main.cpp
index 32e7c8d1..a01472e1 100644
--- a/projects/llm_framework/main_cosy_voice/src/main.cpp
+++ b/projects/llm_framework/main_cosy_voice/src/main.cpp
@@ -411,45 +411,42 @@ class llm_task {
         timer time_total;
         time_total.start();
         try {
-            auto llm_thread_func = [this, &text, &prompt_text_embeds, &prompt_speech_embeds]() {
-                lLaMa_->Run(text, prompt_text_embeds, prompt_speech_embeds, g_token_buffer, g_buffer_mutex, g_buffer_cv,
-                            g_llm_finished);
+            int llm_ret          = 0;
+            auto llm_thread_func = [this, &text, &prompt_text_embeds, &prompt_speech_embeds, &llm_ret]() {
+                llm_ret = lLaMa_->Run(text, prompt_text_embeds, prompt_speech_embeds, g_token_buffer, g_buffer_mutex,
+                                      g_buffer_cv, g_llm_finished);
             };
-
             std::thread llm_thread(llm_thread_func);
-
-            int token_offset     = 0;
+            llm_thread.detach();
             int prompt_token_len = prompt_speech_embeds_flow.size() / lToken2Wav._attr.flow_embed_size;
             if (prompt_token_len < 75) {
                 SLOGE("Error, prompt speech token len %d < 75", prompt_token_len);
                 if (llm_thread.joinable()) llm_thread.join();
                 return -1;
             }
+            if (llm_ret == -1) {
+                return llm_ret;
+            }
             int prompt_token_align_len = 75;
-
             std::vector<float> prompt_speech_embeds_flow1;
             prompt_speech_embeds_flow1.insert(prompt_speech_embeds_flow1.begin(), prompt_speech_embeds_flow.begin(),
                                               prompt_speech_embeds_flow.begin() + prompt_token_align_len * 512);
-
             std::vector<float> prompt_feat1;
             prompt_feat1.insert(prompt_feat1.begin(), prompt_feat.begin(),
                                 prompt_feat.begin() + prompt_token_align_len * 2 * 80);
-
             int promot_token_pad = 0;
             int this_token_hop_len;
-            int i = 0;
+            int token_offset = 0;
+            int i            = 0;
             while (true) {
                 this_token_hop_len = (token_offset == 0) ? lToken2Wav._attr.token_hop_len + promot_token_pad
                                                          : lToken2Wav._attr.token_hop_len;
-
                 std::unique_lock<std::mutex> lock(g_buffer_mutex);
-
                 g_buffer_cv.wait(lock, [&] {
                     return (g_token_buffer.size() - token_offset >=
                             this_token_hop_len + lToken2Wav._attr.pre_lookahead_len) ||
                            g_llm_finished.load() || g_stop.load();
                 });
-
                 if (g_stop) {
                     lock.unlock();
                     break;
@@ -460,9 +457,7 @@ class llm_task {
                                                         lToken2Wav._attr.max_infer_chunk_num - 1) *
                                                    lToken2Wav._attr.token_hop_len;
                     int end = token_offset + this_token_hop_len + lToken2Wav._attr.pre_lookahead_len;
-
                     token.insert(token.end(), g_token_buffer.begin() + start, g_token_buffer.begin() + end);
-
                     lock.unlock();
                     auto speech = lToken2Wav.infer(token, prompt_speech_embeds_flow1, prompt_feat1, spk_embeds,
                                                    token_offset, false);
@@ -481,7 +476,6 @@ class llm_task {
                         if (val < -1.0f) val = -1.0f;
                         wav_pcm_data.push_back(static_cast<int16_t>(val * 32767.0f));
                     }
-
                     if (out_callback_) {
                         out_callback_(std::string(reinterpret_cast<char *>(wav_pcm_data.data()),
                                                   wav_pcm_data.size() * sizeof(int16_t)),
@@ -496,10 +490,6 @@ class llm_task {
                 }
             }
 
-            if (llm_thread.joinable()) {
-                llm_thread.join();
-            }
-
             if (g_stop) {
                 g_token_buffer.erase(g_token_buffer.begin(), g_token_buffer.end());
                 return 1;
@@ -518,7 +508,6 @@ class llm_task {
             std::vector<float> resampled_pcm(static_cast<size_t>(speech.size() * src_ratio + 1));
             int resampled_len = 0;
             resample_audio(speech.data(), speech.size(), resampled_pcm.data(), &resampled_len, src_ratio);
-
             std::vector<int16_t> wav_pcm_data;
             wav_pcm_data.reserve(resampled_len);
             for (int i = 0; i < resampled_len; i++) {
@@ -538,7 +527,6 @@ class llm_task {
                 std::vector<float> resampled_pcm(static_cast<size_t>(output.size() * src_ratio + 1));
                 int resampled_len = 0;
                 resample_audio(output.data(), output.size(), resampled_pcm.data(), &resampled_len, src_ratio);
-
                 std::vector<int16_t> wav_pcm_data_full;
                 wav_pcm_data_full.reserve(resampled_len);
                 for (int i = 0; i < resampled_len; i++) {
@@ -547,7 +535,6 @@ class llm_task {
                     if (val < -1.0f) val = -1.0f;
                     wav_pcm_data_full.push_back(static_cast<int16_t>(val * 32767.0f));
                 }
-
                 std::string wav_path;
                 if (mode_config_.output_path.empty()) {
                     wav_path = generateFilename("/tmp");
@@ -561,14 +548,12 @@ class llm_task {
                 }
                 saveVectorAsWavFloat(resampled_pcm, wav_path, mode_config_.audio_rate, 1);
             }
-
             SLOGI("tts total use time: %.3f s", time_total.cost() / 1000);
             reset();
         } catch (const std::exception &e) {
             std::cerr << "Error in pipeline: " << e.what() << std::endl;
             return 1;
         }
-
         return 0;
     }
 
@@ -599,12 +584,7 @@ class llm_task {
     void inference(const std::string &msg)
     {
         try {
-            // std::string out = lLaMa_->Run(prompt_complete(msg));
-            // if (out_callback_) out_callback_(out, true);
             tts(msg, prompt_text_embeds, prompt_speech_embeds, prompt_feat, prompt_speech_embeds_flow, spk_embeds);
-            std::string out = "finish";
-            if (out_callback_) out_callback_(out, true);
-
         } catch (...) {
             SLOGW("lLaMa_->Run have error!");
         }
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp b/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
index 8e74d906..52994ca6 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
@@ -342,7 +342,7 @@ class LLM {
     {
         std::vector<unsigned short> text_embed;
         std::vector<std::vector<int>> position_ids;
-        Encode(text_embed, position_ids, input_str, prompt_text_embeds, prompt_speech_embeds);
+        if (Encode(text_embed, position_ids, input_str, prompt_text_embeds, prompt_speech_embeds)) return -1;
         return Run(text_embed, position_ids, token_buffer, buffer_mutex, buffer_cv, llm_finished);
     }
 
@@ -560,7 +560,9 @@ class LLM {
             if (b_stop) {
                 break;
             }
-
+            if (indices >= _attr.kv_cache_num) {
+                break;
+            }
             speech_embed_selector.getByIndex(next_token, embed.data());
             memcpy((void *)llama_layers[0].layer.get_input(decode_grpid, "input").pVirAddr, embed.data(),
                    llama_layers[0].layer.get_input(decode_grpid, "input").nSize);

From a3d09132dc832d4f8c0bc927c8fc1edada789201 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Wed, 5 Nov 2025 09:15:48 +0800
Subject: [PATCH 51/79] [update] update model config

---
 .../main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json    | 2 ++
 projects/llm_framework/main_cosy_voice/src/main.cpp           | 4 ----
 2 files changed, 2 insertions(+), 4 deletions(-)

diff --git a/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json b/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json
index b8baa740..a4ea966f 100644
--- a/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json
+++ b/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json
@@ -54,6 +54,8 @@
         "n_timesteps": 10,
         "b_use_mmap_load_embed": true,
         "b_dynamic_load_axmodel_layer": false,
+        "precompute_len": 100,
+        "cmm_size": 1227676,
         "ext_scripts": [
             "tokenizer_CosyVoice2-0.5B-ax650.py",
             "tokenizer"
diff --git a/projects/llm_framework/main_cosy_voice/src/main.cpp b/projects/llm_framework/main_cosy_voice/src/main.cpp
index a01472e1..765f91ae 100644
--- a/projects/llm_framework/main_cosy_voice/src/main.cpp
+++ b/projects/llm_framework/main_cosy_voice/src/main.cpp
@@ -182,7 +182,6 @@ class llm_task {
             CONFIG_AUTO_SET(file_body["mode_param"], filename_decoder_axmodel);
             CONFIG_AUTO_SET(file_body["mode_param"], token2wav_axmodel_dir);
             CONFIG_AUTO_SET(file_body["mode_param"], n_timesteps);
-            CONFIG_AUTO_SET(file_body["mode_param"], prompt_files);
             CONFIG_AUTO_SET(file_body["mode_param"], template_filename_axmodel);
             CONFIG_AUTO_SET(file_body["mode_param"], b_bos);
             CONFIG_AUTO_SET(file_body["mode_param"], b_eos);
@@ -209,9 +208,6 @@ class llm_task {
             INFER_CONFIG_AUTO_SET(file_body["mode_param"], hift_p1_50_first);
             INFER_CONFIG_AUTO_SET(file_body["mode_param"], hift_p1_58);
             INFER_CONFIG_AUTO_SET(file_body["mode_param"], prompt_dir);
-            INFER_CONFIG_AUTO_SET(file_body["mode_param"], prompt_dir);
-            INFER_CONFIG_AUTO_SET(file_body["mode_param"], prompt_dir);
-            INFER_CONFIG_AUTO_SET(file_body["mode_param"], prompt_dir);
             {
                 auto has_http = [](const std::string &s) { return s.find("http") != std::string::npos; };
 

From 423427de9901a48b1b038d1be5aa4408b5fc5eba Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Wed, 5 Nov 2025 11:53:49 +0800
Subject: [PATCH 52/79] [update] update model ctx len

---
 .../models/mode_CosyVoice2-0.5B-ax650.json    |  2 +-
 .../main_cosy_voice/src/main.cpp              | 25 +++++++++++++++----
 2 files changed, 21 insertions(+), 6 deletions(-)

diff --git a/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json b/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json
index a4ea966f..d04cad36 100644
--- a/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json
+++ b/projects/llm_framework/main_cosy_voice/models/mode_CosyVoice2-0.5B-ax650.json
@@ -54,7 +54,7 @@
         "n_timesteps": 10,
         "b_use_mmap_load_embed": true,
         "b_dynamic_load_axmodel_layer": false,
-        "precompute_len": 100,
+        "precompute_len": 200,
         "cmm_size": 1227676,
         "ext_scripts": [
             "tokenizer_CosyVoice2-0.5B-ax650.py",
diff --git a/projects/llm_framework/main_cosy_voice/src/main.cpp b/projects/llm_framework/main_cosy_voice/src/main.cpp
index 765f91ae..a089275e 100644
--- a/projects/llm_framework/main_cosy_voice/src/main.cpp
+++ b/projects/llm_framework/main_cosy_voice/src/main.cpp
@@ -15,6 +15,7 @@
 #include <unistd.h>
 #include <base64.h>
 #include <fstream>
+#include <future>
 #include <stdexcept>
 #include <samplerate.h>
 #include <semaphore.h>
@@ -407,22 +408,36 @@ class llm_task {
         timer time_total;
         time_total.start();
         try {
-            int llm_ret          = 0;
-            auto llm_thread_func = [this, &text, &prompt_text_embeds, &prompt_speech_embeds, &llm_ret]() {
+            int llm_ret = 0;
+            std::promise<int> prom;
+            std::future<int> fut = prom.get_future();
+            auto llm_thread_func = [this, &text, &prompt_text_embeds, &prompt_speech_embeds, &llm_ret, &prom]() {
                 llm_ret = lLaMa_->Run(text, prompt_text_embeds, prompt_speech_embeds, g_token_buffer, g_buffer_mutex,
                                       g_buffer_cv, g_llm_finished);
+                prom.set_value(llm_ret);
             };
             std::thread llm_thread(llm_thread_func);
             llm_thread.detach();
+
+            if (fut.wait_for(std::chrono::milliseconds(10)) == std::future_status::ready) {
+                int llm_ret = fut.get();
+                if (llm_ret == -1) {
+                    SLOGE("Error, Generate failed");
+                    if (out_callback_) out_callback_("Error, Generate failed", true);
+                    return llm_ret;
+                }
+            }
+
             int prompt_token_len = prompt_speech_embeds_flow.size() / lToken2Wav._attr.flow_embed_size;
             if (prompt_token_len < 75) {
                 SLOGE("Error, prompt speech token len %d < 75", prompt_token_len);
                 if (llm_thread.joinable()) llm_thread.join();
+                if (out_callback_) {
+                    out_callback_("Error, prompt speech token len %d < 75", true);
+                }
                 return -1;
             }
-            if (llm_ret == -1) {
-                return llm_ret;
-            }
+
             int prompt_token_align_len = 75;
             std::vector<float> prompt_speech_embeds_flow1;
             prompt_speech_embeds_flow1.insert(prompt_speech_embeds_flow1.begin(), prompt_speech_embeds_flow.begin(),

From 3a162591cf5334a0aee6fa872e17408dbeb41093 Mon Sep 17 00:00:00 2001
From: dianjixz <18637716021@163.com>
Date: Fri, 7 Nov 2025 09:13:57 +0800
Subject: [PATCH 53/79] [fix] pzmq close wait

---
 ext_components/StackFlow/stackflow/pzmq.hpp | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/ext_components/StackFlow/stackflow/pzmq.hpp b/ext_components/StackFlow/stackflow/pzmq.hpp
index 5d51ac32..97c902c2 100644
--- a/ext_components/StackFlow/stackflow/pzmq.hpp
+++ b/ext_components/StackFlow/stackflow/pzmq.hpp
@@ -394,6 +394,8 @@ class pzmq {
     }
     void close_zmq()
     {
+        int linger = 0;
+        zmq_setsockopt(zmq_socket_, ZMQ_LINGER, &linger, sizeof(linger));
         zmq_close(zmq_socket_);
         zmq_ctx_destroy(zmq_ctx_);
         if ((mode_ == ZMQ_PUB) || (mode_ == ZMQ_PULL) || (mode_ == ZMQ_RPC_FUN)) {

From 07c964cf2d57593669fd61e2ed43647cae80072a Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 7 Nov 2025 09:24:00 +0800
Subject: [PATCH 54/79] [update] update software version

---
 projects/llm_framework/main_audio/SConstruct  |  2 +-
 .../llm_framework/main_audio/src/main.cpp     | 31 ++++++++++++++++++-
 projects/llm_framework/main_llm/SConstruct    |  2 +-
 .../llm_framework/main_openai_api/SConstruct  |  4 +--
 projects/llm_framework/tools/llm_pack.py      |  6 ++--
 5 files changed, 37 insertions(+), 8 deletions(-)

diff --git a/projects/llm_framework/main_audio/SConstruct b/projects/llm_framework/main_audio/SConstruct
index 2fa5dbbb..444248ea 100644
--- a/projects/llm_framework/main_audio/SConstruct
+++ b/projects/llm_framework/main_audio/SConstruct
@@ -28,7 +28,7 @@ if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
 STATIC_FILES += [AFile('audio.json'), AFile('audio_kit.json')]
 STATIC_FILES += Glob('mode_*.json')
 
-env['COMPONENTS'].append({'target':'llm_audio-1.6',
+env['COMPONENTS'].append({'target':'llm_audio-1.7',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_audio/src/main.cpp b/projects/llm_framework/main_audio/src/main.cpp
index a892e3b2..de6165fd 100644
--- a/projects/llm_framework/main_audio/src/main.cpp
+++ b/projects/llm_framework/main_audio/src/main.cpp
@@ -49,19 +49,48 @@ class llm_audio : public StackFlow {
         self->pub_ctx_->send_data((const char *)data, size);
     }
 
+    static std::string mono_to_stereo_s16le(const std::string &mono_data)
+    {
+        if (mono_data.empty()) return mono_data;
+
+        size_t sample_count = mono_data.size() / 2;
+        std::string stereo_data;
+        stereo_data.reserve(mono_data.size() * 2);
+
+        const char *src = mono_data.data();
+        for (size_t i = 0; i < sample_count; ++i) {
+            stereo_data.append(src + i * 2, 2);
+            stereo_data.append(src + i * 2, 2);
+        }
+        return stereo_data;
+    }
+
     void hw_queue_play(const std::shared_ptr<void> &arg)
     {
         if (audio_clear_flage_) {
             return;
         }
+
         std::shared_ptr<pzmq_data> originalPtr = std::static_pointer_cast<pzmq_data>(arg);
+        std::string audio_data(static_cast<const char *>(originalPtr->data()), originalPtr->size());
+
+        std::string final_data = audio_data;
+        if (play_config.channel == 2) {
+            final_data = mono_to_stereo_s16le(audio_data);
+        }
+
         std::lock_guard<std::mutex> guard(ax_play_mtx);
         ax_play(play_config.card, play_config.device, play_config.volume, play_config.channel, play_config.rate,
-                play_config.bit, originalPtr->data(), originalPtr->size());
+                play_config.bit, final_data.c_str(), final_data.length());
     }
 
     void hw_play(const std::string &audio_data)
     {
+        std::string final_data = audio_data;
+
+        if (play_config.channel == 2) {
+            final_data = mono_to_stereo_s16le(audio_data);
+        }
         std::lock_guard<std::mutex> guard(ax_play_mtx);
         ax_play(play_config.card, play_config.device, play_config.volume, play_config.channel, play_config.rate,
                 play_config.bit, audio_data.c_str(), audio_data.length());
diff --git a/projects/llm_framework/main_llm/SConstruct b/projects/llm_framework/main_llm/SConstruct
index b6a3d191..608ef386 100644
--- a/projects/llm_framework/main_llm/SConstruct
+++ b/projects/llm_framework/main_llm/SConstruct
@@ -66,7 +66,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_llm-1.9',
+env['COMPONENTS'].append({'target':'llm_llm-1.10',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_openai_api/SConstruct b/projects/llm_framework/main_openai_api/SConstruct
index d2851e73..5259c6b7 100644
--- a/projects/llm_framework/main_openai_api/SConstruct
+++ b/projects/llm_framework/main_openai_api/SConstruct
@@ -18,7 +18,7 @@ LINK_SEARCH_PATH = []
 STATIC_FILES = []
 
 
-ModuleLLMOpenAIPluginPath = wget_github_commit('https://github.com/m5stack/ModuleLLM-OpenAI-Plugin.git', '748ec04e6dde31e6283921e9635204b4b8f8ede2', True)
+ModuleLLMOpenAIPluginPath = wget_github_commit('https://github.com/m5stack/ModuleLLM-OpenAI-Plugin.git', '5298be215735f5b1c21bc9225c38d7cb9c1933db', True)
 python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-openai-api-python-venv_v1.6.tar.gz", 'm5stack_llm-openai-api-python-venv_v1.6.tar.gz')
 
 
@@ -52,7 +52,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_openai_api-1.7',
+env['COMPONENTS'].append({'target':'llm_openai_api-1.8',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 581d1817..478c9198 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -374,10 +374,10 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         # Unit
         'lib-llm':[create_lib_deb,'lib-llm', '1.8', src_folder, revision],
         'llm-sys':[create_bin_deb,'llm-sys', '1.6', src_folder, revision],
-        'llm-audio':[create_bin_deb,'llm-audio', '1.6', src_folder, revision],
+        'llm-audio':[create_bin_deb,'llm-audio', '1.7', src_folder, revision],
         'llm-kws':[create_bin_deb,'llm-kws', '1.9', src_folder, revision],
         'llm-asr':[create_bin_deb,'llm-asr', '1.7', src_folder, revision],
-        'llm-llm':[create_bin_deb,'llm-llm', '1.9', src_folder, revision],
+        'llm-llm':[create_bin_deb,'llm-llm', '1.10', src_folder, revision],
         'llm-tts':[create_bin_deb,'llm-tts', '1.6', src_folder, revision],
         'llm-melotts':[create_bin_deb,'llm-melotts', '1.9', src_folder, revision],
         'llm-camera':[create_bin_deb,'llm-camera', '1.9', src_folder, revision, 'lib-llm'],
@@ -387,7 +387,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-depth-anything':[create_bin_deb,'llm-depth-anything', '1.7', src_folder, revision],
         'llm-vad':[create_bin_deb,'llm-vad', '1.8', src_folder, revision],
         'llm-whisper':[create_bin_deb,'llm-whisper', '1.8', src_folder, revision],
-        'llm-openai-api':[create_bin_deb,'llm-openai-api', '1.7', src_folder, revision],
+        'llm-openai-api':[create_bin_deb,'llm-openai-api', '1.8', src_folder, revision],
         'llm-cosy-voice':[create_bin_deb,'llm-cosy-voice', '1.8', src_folder, revision],
         # keyword spotting Audio file
         'llm-model-audio-en-us':[create_data_deb,'llm-model-audio-en-us', data_version, src_folder, revision],

From bd14152dbd589434f5d9d9138caa05607fed127d Mon Sep 17 00:00:00 2001
From: dianjixz <18637716021@163.com>
Date: Mon, 10 Nov 2025 19:02:00 +0800
Subject: [PATCH 55/79] [update] update ax_msp kconfig bsp version,pzmq add
 NotAction dec,static_lib add libfmt

---
 ext_components/StackFlow/stackflow/pzmq.hpp |  2 +-
 ext_components/ax_msp/Kconfig               | 18 ++++++++++++------
 projects/llm_framework/SConstruct           |  2 +-
 projects/llm_framework/config_defaults.mk   |  1 +
 4 files changed, 15 insertions(+), 8 deletions(-)

diff --git a/ext_components/StackFlow/stackflow/pzmq.hpp b/ext_components/StackFlow/stackflow/pzmq.hpp
index 97c902c2..8a29bd12 100644
--- a/ext_components/StackFlow/stackflow/pzmq.hpp
+++ b/ext_components/StackFlow/stackflow/pzmq.hpp
@@ -382,7 +382,7 @@ class pzmq {
                     std::unique_lock<std::mutex> lock(zmq_fun_mtx_);
                     retval = zmq_fun_.at(msg_ptr->string())(this, msg1_ptr);
                 } catch (...) {
-                    retval = "NotAction";
+                    retval = msg_ptr->string() + " NotAction";
                 }
                 zmq_send(zmq_socket_, retval.c_str(), retval.length(), 0);
                 msg1_ptr.reset();
diff --git a/ext_components/ax_msp/Kconfig b/ext_components/ax_msp/Kconfig
index ce8206e1..ef6a29d2 100644
--- a/ext_components/ax_msp/Kconfig
+++ b/ext_components/ax_msp/Kconfig
@@ -25,17 +25,23 @@ menuconfig AX_MSP_ENABLED
     choice
         prompt "choice common version"
         default MSP_V2_0_0_ENABLED
-        depends on AX_MSP_ENABLED
-    config MSP_V3_6_2_ENABLED
-        bool "enable v3.6.2 bsp"
-        help
-            build by m5stack!
+        depends on AX_MSP_ENABLED && AX_620E_MSP_ENABLED
     config MSP_V2_0_0_ENABLED
         bool "enable v2.0.0 bsp"
         help
             build by m5stack!
     endchoice
-    
+
+    choice
+        prompt "choice common version"
+        default MSP_V3_6_2_ENABLED
+        depends on AX_MSP_ENABLED && AX_650C_MSP_ENABLED
+    config MSP_V3_6_2_ENABLED
+        bool "enable v3.6.2 bsp"
+        help
+            build by m5stack!
+    endchoice
+
     config SAMPLE_COMMON_ENABLED
         bool "Enable sample common"
         default n
diff --git a/projects/llm_framework/SConstruct b/projects/llm_framework/SConstruct
index 26ed7343..b6c85dad 100644
--- a/projects/llm_framework/SConstruct
+++ b/projects/llm_framework/SConstruct
@@ -5,7 +5,7 @@ import shutil
 os.environ['SDK_PATH'] = os.path.normpath(str(Path(os.getcwd())/'..'/'..'/'SDK'))
 os.environ['EXT_COMPONENTS_PATH'] = os.path.normpath(str(Path(os.getcwd())/'..'/'..'/'ext_components'))
 
-version = 'v0.1.1'
+version = 'v0.1.2'
 static_lib = 'static_lib'
 update = False
 
diff --git a/projects/llm_framework/config_defaults.mk b/projects/llm_framework/config_defaults.mk
index 09d6a4b6..898c48a5 100644
--- a/projects/llm_framework/config_defaults.mk
+++ b/projects/llm_framework/config_defaults.mk
@@ -23,3 +23,4 @@ CONFIG_SINGLE_HEADER_LIBS_ENABLED=y
 CONFIG_AX_SAMPLES_ENABLED=y
 CONFIG_SIMDJSON_COMPENENT_ENABLED=y
 CONFIG_SAMPLE_COMMON_ENABLED=y
+CONFIG_MODBUS_ENABLED=y
\ No newline at end of file

From 6d7ae904f4ffd45be7a72027ff3f1cf41e4709fd Mon Sep 17 00:00:00 2001
From: dianjixz <18637716021@163.com>
Date: Mon, 10 Nov 2025 19:02:36 +0800
Subject: [PATCH 56/79] [add] ax650_ec_proxy

---
 .../llm_framework/main_ax650_ec_proxy/Kconfig |   0
 .../main_ax650_ec_proxy/SConstruct            |  58 ++
 .../main_ax650_ec_proxy/src/cmdline.hpp       | 732 +++++++++++++
 .../main_ax650_ec_proxy/src/ec_cli_main.cpp   | 317 ++++++
 .../main_ax650_ec_proxy/src/ec_proxy_main.cpp | 963 ++++++++++++++++++
 5 files changed, 2070 insertions(+)
 create mode 100644 projects/llm_framework/main_ax650_ec_proxy/Kconfig
 create mode 100644 projects/llm_framework/main_ax650_ec_proxy/SConstruct
 create mode 100644 projects/llm_framework/main_ax650_ec_proxy/src/cmdline.hpp
 create mode 100644 projects/llm_framework/main_ax650_ec_proxy/src/ec_cli_main.cpp
 create mode 100644 projects/llm_framework/main_ax650_ec_proxy/src/ec_proxy_main.cpp

diff --git a/projects/llm_framework/main_ax650_ec_proxy/Kconfig b/projects/llm_framework/main_ax650_ec_proxy/Kconfig
new file mode 100644
index 00000000..e69de29b
diff --git a/projects/llm_framework/main_ax650_ec_proxy/SConstruct b/projects/llm_framework/main_ax650_ec_proxy/SConstruct
new file mode 100644
index 00000000..1a908a3a
--- /dev/null
+++ b/projects/llm_framework/main_ax650_ec_proxy/SConstruct
@@ -0,0 +1,58 @@
+import os
+
+Import('env')
+with open(env['PROJECT_TOOL_S']) as f:
+    exec(f.read())
+
+if "CONFIG_AX_650C_MSP_ENABLED" not in os.environ:
+    Return('env')
+
+SRCS = []
+INCLUDE = []
+PRIVATE_INCLUDE = []
+REQUIREMENTS = ['pthread', 'utilities', 'eventpp', 'StackFlow', 'modbus_component', 'fmt']
+STATIC_LIB = []
+DYNAMIC_LIB = []
+DEFINITIONS = []
+DEFINITIONS_PRIVATE = []
+LDFLAGS = []
+LINK_SEARCH_PATH = []
+STATIC_FILES = []
+
+LINK_SEARCH_PATH += [ADir('../static_lib')]
+
+LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
+
+STATIC_FILES += []
+
+
+SRCS = [AFile('src/ec_proxy_main.cpp')]
+env['COMPONENTS'].append({'target':'ax650_ec_proxy-1.0',
+                          'SRCS':SRCS,
+                          'INCLUDE':INCLUDE,
+                          'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
+                          'REQUIREMENTS':REQUIREMENTS,
+                          'STATIC_LIB':STATIC_LIB,
+                          'DYNAMIC_LIB':DYNAMIC_LIB,
+                          'DEFINITIONS':DEFINITIONS,
+                          'DEFINITIONS_PRIVATE':DEFINITIONS_PRIVATE,
+                          'LDFLAGS':LDFLAGS,
+                          'LINK_SEARCH_PATH':LINK_SEARCH_PATH,
+                          'STATIC_FILES':STATIC_FILES,
+                          'REGISTER':'project'
+                          })
+SRCS = [AFile('src/ec_cli_main.cpp')]
+env['COMPONENTS'].append({'target':'ax650_ec_cli-1.0',
+                          'SRCS':SRCS,
+                          'INCLUDE':INCLUDE,
+                          'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
+                          'REQUIREMENTS':REQUIREMENTS,
+                          'STATIC_LIB':STATIC_LIB,
+                          'DYNAMIC_LIB':DYNAMIC_LIB,
+                          'DEFINITIONS':DEFINITIONS,
+                          'DEFINITIONS_PRIVATE':DEFINITIONS_PRIVATE,
+                          'LDFLAGS':LDFLAGS,
+                          'LINK_SEARCH_PATH':LINK_SEARCH_PATH,
+                          'STATIC_FILES':STATIC_FILES,
+                          'REGISTER':'project'
+                          })
diff --git a/projects/llm_framework/main_ax650_ec_proxy/src/cmdline.hpp b/projects/llm_framework/main_ax650_ec_proxy/src/cmdline.hpp
new file mode 100644
index 00000000..ea8b59e2
--- /dev/null
+++ b/projects/llm_framework/main_ax650_ec_proxy/src/cmdline.hpp
@@ -0,0 +1,732 @@
+/*
+  Copyright (c) 2009, Hideyuki Tanaka
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+  modification, are permitted provided that the following conditions are met:
+  * Redistributions of source code must retain the above copyright
+  notice, this list of conditions and the following disclaimer.
+  * Redistributions in binary form must reproduce the above copyright
+  notice, this list of conditions and the following disclaimer in the
+  documentation and/or other materials provided with the distribution.
+  * Neither the name of the <organization> nor the
+  names of its contributors may be used to endorse or promote products
+  derived from this software without specific prior written permission.
+
+  THIS SOFTWARE IS PROVIDED BY <copyright holder> ''AS IS'' AND ANY
+  EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+  WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+  DISCLAIMED. IN NO EVENT SHALL <copyright holder> BE LIABLE FOR ANY
+  DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+  (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+  LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
+  ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+  (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+  SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#pragma once
+
+#include <cxxabi.h>
+
+#include <algorithm>
+#include <cstdlib>
+#include <cstring>
+#include <iostream>
+#include <map>
+#include <sstream>
+#include <stdexcept>
+#include <string>
+#include <typeinfo>
+#include <vector>
+
+namespace cmdline {
+
+namespace detail {
+
+template <typename Target, typename Source, bool Same>
+class lexical_cast_t {
+ public:
+  static Target cast(const Source &arg) {
+    Target ret;
+    std::stringstream ss;
+    if (!(ss << arg && ss >> ret && ss.eof())) throw std::bad_cast();
+
+    return ret;
+  }
+};
+
+template <typename Target, typename Source>
+class lexical_cast_t<Target, Source, true> {
+ public:
+  static Target cast(const Source &arg) { return arg; }
+};
+
+template <typename Source>
+class lexical_cast_t<std::string, Source, false> {
+ public:
+  static std::string cast(const Source &arg) {
+    std::ostringstream ss;
+    ss << arg;
+    return ss.str();
+  }
+};
+
+template <typename Target>
+class lexical_cast_t<Target, std::string, false> {
+ public:
+  static Target cast(const std::string &arg) {
+    Target ret;
+    std::istringstream ss(arg);
+    if (!(ss >> ret && ss.eof())) throw std::bad_cast();
+    return ret;
+  }
+};
+
+template <typename T1, typename T2>
+struct is_same {
+  static const bool value = false;
+};
+
+template <typename T>
+struct is_same<T, T> {
+  static const bool value = true;
+};
+
+template <typename Target, typename Source>
+Target lexical_cast(const Source &arg) {
+  return lexical_cast_t<Target, Source,
+                        detail::is_same<Target, Source>::value>::cast(arg);
+}
+
+static inline std::string demangle(const std::string &name) {
+  int status = 0;
+  char *p = abi::__cxa_demangle(name.c_str(), 0, 0, &status);
+  std::string ret(p);
+  free(p);
+  return ret;
+}
+
+template <class T>
+std::string readable_typename() {
+  return demangle(typeid(T).name());
+}
+
+template <class T>
+std::string default_value(T def) {
+  return detail::lexical_cast<std::string>(def);
+}
+
+template <>
+inline std::string readable_typename<std::string>() {
+  return "string";
+}
+
+}  // namespace detail
+
+//-----
+
+class cmdline_error : public std::exception {
+ public:
+  cmdline_error(const std::string &msg) : msg(msg) {}
+  ~cmdline_error() throw() {}
+  const char *what() const throw() { return msg.c_str(); }
+
+ private:
+  std::string msg;
+};
+
+template <class T>
+struct default_reader {
+  T operator()(const std::string &str) { return detail::lexical_cast<T>(str); }
+};
+
+template <class T>
+struct range_reader {
+  range_reader(const T &low, const T &high) : low(low), high(high) {}
+  T operator()(const std::string &s) const {
+    T ret = default_reader<T>()(s);
+    if (!(ret >= low && ret <= high))
+      throw cmdline::cmdline_error("range_error");
+    return ret;
+  }
+
+ private:
+  T low, high;
+};
+
+template <class T>
+range_reader<T> range(const T &low, const T &high) {
+  return range_reader<T>(low, high);
+}
+
+template <class T>
+struct oneof_reader {
+  T operator()(const std::string &s) {
+    T ret = default_reader<T>()(s);
+    if (std::find(alt.begin(), alt.end(), ret) == alt.end())
+      throw cmdline_error("");
+    return ret;
+  }
+  void add(const T &v) { alt.push_back(v); }
+
+ private:
+  std::vector<T> alt;
+};
+
+template <class T>
+oneof_reader<T> oneof(T a1) {
+  oneof_reader<T> ret;
+  ret.add(a1);
+  return ret;
+}
+
+template <class T>
+oneof_reader<T> oneof(T a1, T a2) {
+  oneof_reader<T> ret;
+  ret.add(a1);
+  ret.add(a2);
+  return ret;
+}
+
+template <class T>
+oneof_reader<T> oneof(T a1, T a2, T a3) {
+  oneof_reader<T> ret;
+  ret.add(a1);
+  ret.add(a2);
+  ret.add(a3);
+  return ret;
+}
+
+template <class T>
+oneof_reader<T> oneof(T a1, T a2, T a3, T a4) {
+  oneof_reader<T> ret;
+  ret.add(a1);
+  ret.add(a2);
+  ret.add(a3);
+  ret.add(a4);
+  return ret;
+}
+
+template <class T>
+oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5) {
+  oneof_reader<T> ret;
+  ret.add(a1);
+  ret.add(a2);
+  ret.add(a3);
+  ret.add(a4);
+  ret.add(a5);
+  return ret;
+}
+
+template <class T>
+oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5, T a6) {
+  oneof_reader<T> ret;
+  ret.add(a1);
+  ret.add(a2);
+  ret.add(a3);
+  ret.add(a4);
+  ret.add(a5);
+  ret.add(a6);
+  return ret;
+}
+
+template <class T>
+oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5, T a6, T a7) {
+  oneof_reader<T> ret;
+  ret.add(a1);
+  ret.add(a2);
+  ret.add(a3);
+  ret.add(a4);
+  ret.add(a5);
+  ret.add(a6);
+  ret.add(a7);
+  return ret;
+}
+
+template <class T>
+oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5, T a6, T a7, T a8) {
+  oneof_reader<T> ret;
+  ret.add(a1);
+  ret.add(a2);
+  ret.add(a3);
+  ret.add(a4);
+  ret.add(a5);
+  ret.add(a6);
+  ret.add(a7);
+  ret.add(a8);
+  return ret;
+}
+
+template <class T>
+oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5, T a6, T a7, T a8, T a9) {
+  oneof_reader<T> ret;
+  ret.add(a1);
+  ret.add(a2);
+  ret.add(a3);
+  ret.add(a4);
+  ret.add(a5);
+  ret.add(a6);
+  ret.add(a7);
+  ret.add(a8);
+  ret.add(a9);
+  return ret;
+}
+
+template <class T>
+oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5, T a6, T a7, T a8, T a9,
+                      T a10) {
+  oneof_reader<T> ret;
+  ret.add(a1);
+  ret.add(a2);
+  ret.add(a3);
+  ret.add(a4);
+  ret.add(a5);
+  ret.add(a6);
+  ret.add(a7);
+  ret.add(a8);
+  ret.add(a9);
+  ret.add(a10);
+  return ret;
+}
+
+//-----
+
+class parser {
+ public:
+  parser() {}
+  ~parser() {
+    for (std::map<std::string, option_base *>::iterator p = options.begin();
+         p != options.end(); p++)
+      delete p->second;
+  }
+
+  void add(const std::string &name, char short_name = 0,
+           const std::string &desc = "") {
+    if (options.count(name))
+      throw cmdline_error("multiple definition: " + name);
+    options[name] = new option_without_value(name, short_name, desc);
+    ordered.push_back(options[name]);
+  }
+
+  template <class T>
+  void add(const std::string &name, char short_name = 0,
+           const std::string &desc = "", bool need = true, const T def = T()) {
+    add(name, short_name, desc, need, def, default_reader<T>());
+  }
+
+  template <class T, class F>
+  void add(const std::string &name, char short_name = 0,
+           const std::string &desc = "", bool need = true, const T def = T(),
+           F reader = F()) {
+    if (options.count(name))
+      throw cmdline_error("multiple definition: " + name);
+    options[name] = new option_with_value_with_reader<T, F>(
+        name, short_name, need, def, desc, reader);
+    ordered.push_back(options[name]);
+  }
+
+  void footer(const std::string &f) { ftr = f; }
+
+  void set_program_name(const std::string &name) { prog_name = name; }
+
+  bool exist(const std::string &name) const {
+    if (options.count(name) == 0)
+      throw cmdline_error("there is no flag: --" + name);
+    return options.find(name)->second->has_set();
+  }
+
+  template <class T>
+  const T &get(const std::string &name) const {
+    if (options.count(name) == 0)
+      throw cmdline_error("there is no flag: --" + name);
+    const option_with_value<T> *p =
+        dynamic_cast<const option_with_value<T> *>(options.find(name)->second);
+    if (p == NULL) throw cmdline_error("type mismatch flag '" + name + "'");
+    return p->get();
+  }
+
+  const std::vector<std::string> &rest() const { return others; }
+
+  bool parse(const std::string &arg) {
+    std::vector<std::string> args;
+
+    std::string buf;
+    bool in_quote = false;
+    for (std::string::size_type i = 0; i < arg.length(); i++) {
+      if (arg[i] == '\"') {
+        in_quote = !in_quote;
+        continue;
+      }
+
+      if (arg[i] == ' ' && !in_quote) {
+        args.push_back(buf);
+        buf = "";
+        continue;
+      }
+
+      if (arg[i] == '\\') {
+        i++;
+        if (i >= arg.length()) {
+          errors.push_back("unexpected occurrence of '\\' at end of string");
+          return false;
+        }
+      }
+
+      buf += arg[i];
+    }
+
+    if (in_quote) {
+      errors.push_back("quote is not closed");
+      return false;
+    }
+
+    if (buf.length() > 0) args.push_back(buf);
+
+    for (size_t i = 0; i < args.size(); i++)
+      std::cout << "\"" << args[i] << "\"" << std::endl;
+
+    return parse(args);
+  }
+
+  bool parse(const std::vector<std::string> &args) {
+    int argc = static_cast<int>(args.size());
+    std::vector<const char *> argv(argc);
+
+    for (int i = 0; i < argc; i++) argv[i] = args[i].c_str();
+
+    return parse(argc, &argv[0]);
+  }
+
+  bool parse(int argc, const char *const argv[]) {
+    errors.clear();
+    others.clear();
+
+    if (argc < 1) {
+      errors.push_back("argument number must be longer than 0");
+      return false;
+    }
+    if (prog_name == "") prog_name = argv[0];
+
+    std::map<char, std::string> lookup;
+    for (std::map<std::string, option_base *>::iterator p = options.begin();
+         p != options.end(); p++) {
+      if (p->first.length() == 0) continue;
+      char initial = p->second->short_name();
+      if (initial) {
+        if (lookup.count(initial) > 0) {
+          lookup[initial] = "";
+          errors.push_back(std::string("short option '") + initial +
+                           "' is ambiguous");
+          return false;
+        } else
+          lookup[initial] = p->first;
+      }
+    }
+
+    for (int i = 1; i < argc; i++) {
+      if (strncmp(argv[i], "--", 2) == 0) {
+        const char *p = strchr(argv[i] + 2, '=');
+        if (p) {
+          std::string name(argv[i] + 2, p);
+          std::string val(p + 1);
+          set_option(name, val);
+        } else {
+          std::string name(argv[i] + 2);
+          if (options.count(name) == 0) {
+            errors.push_back("undefined option: --" + name);
+            continue;
+          }
+          if (options[name]->has_value()) {
+            if (i + 1 >= argc) {
+              errors.push_back("option needs value: --" + name);
+              continue;
+            } else {
+              i++;
+              set_option(name, argv[i]);
+            }
+          } else {
+            set_option(name);
+          }
+        }
+      } else if (strncmp(argv[i], "-", 1) == 0) {
+        if (!argv[i][1]) continue;
+        char last = argv[i][1];
+        for (int j = 2; argv[i][j]; j++) {
+          last = argv[i][j];
+          if (lookup.count(argv[i][j - 1]) == 0) {
+            errors.push_back(std::string("undefined short option: -") +
+                             argv[i][j - 1]);
+            continue;
+          }
+          if (lookup[argv[i][j - 1]] == "") {
+            errors.push_back(std::string("ambiguous short option: -") +
+                             argv[i][j - 1]);
+            continue;
+          }
+          set_option(lookup[argv[i][j - 1]]);
+        }
+
+        if (lookup.count(last) == 0) {
+          errors.push_back(std::string("undefined short option: -") + last);
+          continue;
+        }
+        if (lookup[last] == "") {
+          errors.push_back(std::string("ambiguous short option: -") + last);
+          continue;
+        }
+
+        if (i + 1 < argc && options[lookup[last]]->has_value()) {
+          set_option(lookup[last], argv[i + 1]);
+          i++;
+        } else {
+          set_option(lookup[last]);
+        }
+      } else {
+        others.push_back(argv[i]);
+      }
+    }
+
+    for (std::map<std::string, option_base *>::iterator p = options.begin();
+         p != options.end(); p++)
+      if (!p->second->valid())
+        errors.push_back("need option: --" + std::string(p->first));
+
+    return errors.size() == 0;
+  }
+
+  void parse_check(const std::string &arg) {
+    if (!options.count("help")) add("help", '?', "print this message");
+    check(0, parse(arg));
+  }
+
+  void parse_check(const std::vector<std::string> &args) {
+    if (!options.count("help")) add("help", '?', "print this message");
+    check(args.size(), parse(args));
+  }
+
+  void parse_check(int argc, char *argv[]) {
+    if (!options.count("help")) add("help", '?', "print this message");
+    check(argc, parse(argc, argv));
+  }
+
+  std::string error() const { return errors.size() > 0 ? errors[0] : ""; }
+
+  std::string error_full() const {
+    std::ostringstream oss;
+    for (size_t i = 0; i < errors.size(); i++) oss << errors[i] << std::endl;
+    return oss.str();
+  }
+
+  std::string usage() const {
+    std::ostringstream oss;
+    oss << "usage: " << prog_name << " ";
+    for (size_t i = 0; i < ordered.size(); i++) {
+      if (ordered[i]->must()) oss << ordered[i]->short_description() << " ";
+    }
+
+    oss << "[options] ... " << ftr << std::endl;
+    oss << "options:" << std::endl;
+
+    size_t max_width = 0;
+    for (size_t i = 0; i < ordered.size(); i++) {
+      max_width = std::max(max_width, ordered[i]->name().length());
+    }
+    for (size_t i = 0; i < ordered.size(); i++) {
+      if (ordered[i]->short_name()) {
+        oss << "  -" << ordered[i]->short_name() << ", ";
+      } else {
+        oss << "      ";
+      }
+
+      oss << "--" << ordered[i]->name();
+      for (size_t j = ordered[i]->name().length(); j < max_width + 4; j++)
+        oss << ' ';
+      oss << ordered[i]->description() << std::endl;
+    }
+    return oss.str();
+  }
+
+ private:
+  void check(int argc, bool ok) {
+    if ((argc == 1 && !ok) || exist("help")) {
+      std::cerr << usage();
+      exit(0);
+    }
+
+    if (!ok) {
+      std::cerr << error() << std::endl << usage();
+      exit(1);
+    }
+  }
+
+  void set_option(const std::string &name) {
+    if (options.count(name) == 0) {
+      errors.push_back("undefined option: --" + name);
+      return;
+    }
+    if (!options[name]->set()) {
+      errors.push_back("option needs value: --" + name);
+      return;
+    }
+  }
+
+  void set_option(const std::string &name, const std::string &value) {
+    if (options.count(name) == 0) {
+      errors.push_back("undefined option: --" + name);
+      return;
+    }
+    if (!options[name]->set(value)) {
+      errors.push_back("option value is invalid: --" + name + "=" + value);
+      return;
+    }
+  }
+
+  class option_base {
+   public:
+    virtual ~option_base() {}
+
+    virtual bool has_value() const = 0;
+    virtual bool set() = 0;
+    virtual bool set(const std::string &value) = 0;
+    virtual bool has_set() const = 0;
+    virtual bool valid() const = 0;
+    virtual bool must() const = 0;
+
+    virtual const std::string &name() const = 0;
+    virtual char short_name() const = 0;
+    virtual const std::string &description() const = 0;
+    virtual std::string short_description() const = 0;
+  };
+
+  class option_without_value : public option_base {
+   public:
+    option_without_value(const std::string &name, char short_name,
+                         const std::string &desc)
+        : nam(name), snam(short_name), desc(desc), has(false) {}
+    ~option_without_value() {}
+
+    bool has_value() const { return false; }
+
+    bool set() {
+      has = true;
+      return true;
+    }
+
+    bool set(const std::string &) { return false; }
+
+    bool has_set() const { return has; }
+
+    bool valid() const { return true; }
+
+    bool must() const { return false; }
+
+    const std::string &name() const { return nam; }
+
+    char short_name() const { return snam; }
+
+    const std::string &description() const { return desc; }
+
+    std::string short_description() const { return "--" + nam; }
+
+   private:
+    std::string nam;
+    char snam;
+    std::string desc;
+    bool has;
+  };
+
+  template <class T>
+  class option_with_value : public option_base {
+   public:
+    option_with_value(const std::string &name, char short_name, bool need,
+                      const T &def, const std::string &desc)
+        : nam(name),
+          snam(short_name),
+          need(need),
+          has(false),
+          def(def),
+          actual(def) {
+      this->desc = full_description(desc);
+    }
+    ~option_with_value() {}
+
+    const T &get() const { return actual; }
+
+    bool has_value() const { return true; }
+
+    bool set() { return false; }
+
+    bool set(const std::string &value) {
+      try {
+        actual = read(value);
+        has = true;
+      } catch (const std::exception &e) {
+        return false;
+      }
+      return true;
+    }
+
+    bool has_set() const { return has; }
+
+    bool valid() const {
+      if (need && !has) return false;
+      return true;
+    }
+
+    bool must() const { return need; }
+
+    const std::string &name() const { return nam; }
+
+    char short_name() const { return snam; }
+
+    const std::string &description() const { return desc; }
+
+    std::string short_description() const {
+      return "--" + nam + "=" + detail::readable_typename<T>();
+    }
+
+   protected:
+    std::string full_description(const std::string &desc) {
+      return desc + " (" + detail::readable_typename<T>() +
+             (need ? "" : " [=" + detail::default_value<T>(def) + "]") + ")";
+    }
+
+    virtual T read(const std::string &s) = 0;
+
+    std::string nam;
+    char snam;
+    bool need;
+    std::string desc;
+
+    bool has;
+    T def;
+    T actual;
+  };
+
+  template <class T, class F>
+  class option_with_value_with_reader : public option_with_value<T> {
+   public:
+    option_with_value_with_reader(const std::string &name, char short_name,
+                                  bool need, const T def,
+                                  const std::string &desc, F reader)
+        : option_with_value<T>(name, short_name, need, def, desc),
+          reader(reader) {}
+
+   private:
+    T read(const std::string &s) { return reader(s); }
+
+    F reader;
+  };
+
+  std::map<std::string, option_base *> options;
+  std::vector<option_base *> ordered;
+  std::string ftr;
+
+  std::string prog_name;
+  std::vector<std::string> others;
+
+  std::vector<std::string> errors;
+};
+
+}  // namespace cmdline
diff --git a/projects/llm_framework/main_ax650_ec_proxy/src/ec_cli_main.cpp b/projects/llm_framework/main_ax650_ec_proxy/src/ec_cli_main.cpp
new file mode 100644
index 00000000..1e47073e
--- /dev/null
+++ b/projects/llm_framework/main_ax650_ec_proxy/src/ec_cli_main.cpp
@@ -0,0 +1,317 @@
+#include <iostream>
+#include <modbus.h>
+#include "pzmq.hpp"
+#include "cmdline.hpp"
+#include <../../static_lib/include/fmt/core.h>
+#include <../../static_lib/include/fmt/format.h>
+#include <cstdlib>
+#include <list>
+#include <vector>
+#include "json.hpp"
+int condition = 1;
+const char *rpc_socket_path;
+const char *pub_socket_path;
+
+void info_fun(cmdline::parser &a, std::string set_fun, std::string get_fun)
+{
+    if (a.exist("data") || a.exist("DataRaw")) {
+        if (set_fun.length() == 0) return;
+        std::string DataRaw = a.get<std::string>("DataRaw");
+        if (DataRaw.empty()) {
+            DataRaw = fmt::format("{{ \"data\": {} }}", a.get<std::string>("data"));
+        }
+        StackFlows::pzmq Context(rpc_socket_path);
+        Context.call_rpc_action(set_fun, DataRaw,
+                                [set_fun](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
+                                    if (data->size() > 0) {
+                                        std::cout << data->string() << std::endl;
+                                    } else {
+                                        std::cout << "call " << set_fun << " faile!" << std::endl;
+                                    }
+                                });
+    } else {
+        if (get_fun.length() == 0) {
+            std::cout << "please set data: -d or -D" << std::endl;
+            return;
+        }
+        StackFlows::pzmq Context(rpc_socket_path);
+        Context.call_rpc_action(get_fun, "{}",
+                                [get_fun](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
+                                    if (data->size() > 0) {
+                                        std::cout << data->string() << std::endl;
+                                    } else {
+                                        std::cout << "call " << get_fun << " faile! " << std::endl;
+                                    }
+                                });
+    }
+}
+
+void get_info_fun(cmdline::parser &a, std::string fun)
+{
+    StackFlows::pzmq Context(rpc_socket_path);
+    Context.call_rpc_action(fun, "{\"data\":\"None\"}",
+                            [fun](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
+                                if (data->size() > 0) {
+                                    std::cout << data->string() << std::endl;
+                                } else {
+                                    std::cout << "call " << fun << " faile!" << std::endl;
+                                }
+                            });
+}
+
+void set_info_fun(cmdline::parser &a, std::string fun)
+{
+    if (a.exist("data") || a.exist("DataRaw")) {
+        std::string DataRaw = a.get<std::string>("DataRaw");
+        if (DataRaw.empty()) {
+            DataRaw = fmt::format("{{ \"data\": {} }}", a.get<std::string>("data"));
+        }
+        StackFlows::pzmq Context(rpc_socket_path);
+        Context.call_rpc_action(fun, DataRaw,
+                                [fun](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
+                                    if (data->size() > 0) {
+                                        std::cout << data->string() << std::endl;
+                                    } else {
+                                        std::cout << "call " << fun << "faile!" << std::endl;
+                                    }
+                                });
+    }
+}
+
+void fan_speed_fun(cmdline::parser &a)
+{
+    StackFlows::pzmq Context(rpc_socket_path);
+    Context.call_rpc_action("fan_get_speed", "{}",
+                            [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
+                                if (data->size() > 0) {
+                                    std::cout << data->string() << std::endl;
+                                } else {
+                                    std::cout << "call fan_get_speed faile!" << std::endl;
+                                }
+                            });
+}
+
+void fan_pwm_fun(cmdline::parser &a)
+{
+    if (a.exist("data") || a.exist("DataRaw")) {
+        std::string DataRaw = a.get<std::string>("DataRaw");
+        if (DataRaw.empty()) {
+            DataRaw = fmt::format("{{ \"data\": {} }}", a.get<std::string>("data"));
+        }
+        StackFlows::pzmq Context(rpc_socket_path);
+        Context.call_rpc_action("fan_set_pwm", DataRaw,
+                                [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
+                                    if (data->size() > 0) {
+                                        std::cout << data->string() << std::endl;
+                                    } else {
+                                        std::cout << "call fan_set_pwm faile!" << std::endl;
+                                    }
+                                });
+    } else {
+        StackFlows::pzmq Context(rpc_socket_path);
+        Context.call_rpc_action("fan_get_pwm", "{}",
+                                [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
+                                    if (data->size() > 0) {
+                                        std::cout << data->string() << std::endl;
+                                    } else {
+                                        std::cout << "call fan_get_pwm faile!" << std::endl;
+                                    }
+                                });
+    }
+}
+
+void rgb_fun(cmdline::parser &a)
+{
+    if (a.exist("data") || a.exist("DataRaw")) {
+        std::string DataRaw = a.get<std::string>("DataRaw");
+        if (DataRaw.empty()) {
+            DataRaw = fmt::format("{{ \"data\": {} }}", a.get<std::string>("data"));
+        }
+        StackFlows::pzmq Context(rpc_socket_path);
+        Context.call_rpc_action("rgb_set_mode", DataRaw,
+                                [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
+                                    if (data->size() > 0) {
+                                        std::cout << data->string() << std::endl;
+                                    } else {
+                                        std::cout << "call rgb_set_mode faile!" << std::endl;
+                                    }
+                                });
+    } else {
+        StackFlows::pzmq Context(rpc_socket_path);
+        Context.call_rpc_action("rgb_get_mode", "{}",
+                                [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
+                                    if (data->size() > 0) {
+                                        std::cout << data->string() << std::endl;
+                                    } else {
+                                        std::cout << "call rgb_get_mode faile!" << std::endl;
+                                    }
+                                });
+    }
+}
+
+void exec_fun(cmdline::parser &a)
+{
+    if (a.exist("list")) {
+        StackFlows::pzmq Context(rpc_socket_path);
+        std::cout << "list" << std::endl;
+        Context.call_rpc_action("list_action", "None",
+                                [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
+                                    if (data->size() > 0) {
+                                        try {
+                                            nlohmann::json _data = nlohmann::json::parse(data->string());
+                                            std::cout << _data.dump(4) << std::endl;
+                                        } catch (const nlohmann::json::parse_error &e) {
+                                            std::cerr << "解析失败：" << e.what() << std::endl;
+                                        }
+                                    } else {
+                                        std::cout << "call list_action faile!" << std::endl;
+                                    }
+                                });
+        return;
+    }
+
+    const std::string fun  = a.get<std::string>("fun");
+    const std::string data = a.get<std::string>("data");
+    if (fun.length() == 0) {
+        std::cout << "fun is empty! cli exec -f <fun> -d <data>" << std::endl;
+        return;
+    }
+
+    StackFlows::pzmq Context(rpc_socket_path);
+    Context.call_rpc_action(fun, data, [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
+        if (data->size() > 0) {
+            std::cout << "" << data->string() << std::endl;
+        } else {
+            std::cout << "call ax650_ec_prox faile!" << std::endl;
+        }
+    });
+}
+
+void echo_fun(cmdline::parser &a)
+{
+    if (a.exist("button")) {
+        StackFlows::pzmq Context(pub_socket_path, ZMQ_SUB,
+                                 [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
+                                     std::cout << data->string() << std::endl;
+                                 });
+        while (condition) {
+            usleep(100 * 1000);
+        }
+        exit(0);
+    }
+}
+
+void signalHandler(int signum)
+{
+    condition = 0;
+    std::exit(0);  // 这会调用全局对象的析构函数
+}
+
+struct call_fun {
+    std::string fun;
+    char flage;
+    std::string dec;
+    std::function<void(cmdline::parser &a)> fun_call;
+};
+
+int main(int argc, char *argv[])
+{
+    if (argc < 2) {
+    error_print_exit:
+        std::cout << R"help(
+用法：git [--version] [--help] 
+           <命令> [<参数>]
+
+这些是各种场合常见的 cli 命令：
+
+直接操作 ec （参见：cli help device）
+   device    直接执行一个操作，使用指令传递参数
+
+自定义操作 ec （参见：git help exec）
+   exec       执行一个原始函数调用
+
+订阅 prox 的通道数据（参见：git help echo）
+   echo      显示 prox 的通道数据
+
+命令 'cli help -a' 和 'cli help -g' 显示可用的子命令和一些概念帮助。
+查看 'cli help <命令>' 或 'cli help <概念>' 以获取给定子命令或概念的
+帮助。
+)help" << std::endl;
+        return 0;
+    }
+    signal(SIGTERM, signalHandler);
+    signal(SIGINT, signalHandler);
+    rpc_socket_path = getenv("AX650C_EC_PROXY_RPC_SOCKET");
+    if (rpc_socket_path == NULL) {
+        rpc_socket_path = "ipc:///tmp/rpc.ec_prox";
+    }
+    pub_socket_path = getenv("AX650C_EC_PROXY_PUB_SOCKET");
+    if (pub_socket_path == NULL) {
+        pub_socket_path = "ipc:///tmp/llm/ec_prox.event.socket";
+    }
+    if (std::string(argv[1]) == "device") {
+        cmdline::parser a;
+        // clang-format off
+        std::list<struct call_fun> cmd_list = {
+            {"rgb", 'r', "rgb_mode", std::bind(&info_fun, std::placeholders::_1, "rgb_set_mode", "rgb_get_mode")},
+            {"rgb_size", 0, "rgb_size", std::bind(&info_fun, std::placeholders::_1, "rgb_set_size", "rgb_get_size")},
+            {"rgb_get_color", 0, "rgb_get_color", std::bind(&info_fun, std::placeholders::_1, "rgb_get_color", "")},
+            {"rgb_set_color", 0, "rgb set color ,example: cli --rgb_set_color -d '{\"rgb_index\":0,\"rgb_color\":255}'", std::bind(&info_fun, std::placeholders::_1, "rgb_set_color", "")},
+            {"fan", 'f', "fan_pwm", std::bind(&info_fun, std::placeholders::_1, "fan_set_pwm", "fan_get_pwm")},
+            {"fanspeed", 'F', "fan_speed", std::bind(&info_fun, std::placeholders::_1, "", "fan_get_speed")},
+            {"board", 'B', "board_get_power_info", std::bind(&info_fun, std::placeholders::_1, "", "board_get_power_info")},
+            {"ip", 'i', "ip fun", std::bind(&info_fun, std::placeholders::_1, "ip_set", "ip_get")},
+            {"lcd", 'l', "lcd fun", std::bind(&info_fun, std::placeholders::_1, "lcd_set_mode", "lcd_get_mode")},
+            {"lcd_ram", 0, "lcd_ram", std::bind(&info_fun, std::placeholders::_1, "lcd_set_ram", "")},
+            {"lcd_echo", 'E', "lcd_echo", std::bind(&info_fun, std::placeholders::_1, "lcd_echo", "")},
+            {"vddcpu", 'c', "vddcpu fun", std::bind(&info_fun, std::placeholders::_1, "vddcpu_set", "vddcpu_get")},
+            {"modbus_speed", 0, "modbus_speed", std::bind(&info_fun, std::placeholders::_1, "modbus_set_speed", "modbus_get_speed")},
+            {"ext_power", 'p', "ext_power fun", std::bind(&info_fun, std::placeholders::_1, "ext_power", "")},
+            {"board_power", 'b', "board_power fun", std::bind(&info_fun, std::placeholders::_1, "board_power", "")},
+            {"pcie0", 0, "pcie0_set_switch fun", std::bind(&info_fun, std::placeholders::_1, "pcie0_set_switch", "")},
+            {"pcie1", 0, "pcie1_set_switch fun", std::bind(&info_fun, std::placeholders::_1, "pcie1_set_switch", "")},
+            {"gl3510_reset", 0, "gl3510_reset fun", std::bind(&info_fun, std::placeholders::_1, "gl3510_reset", "")},
+            {"usbds1_big", 0, "usbds1_set_big_power fun", std::bind(&info_fun, std::placeholders::_1, "usbds1_set_big_power", "")},
+            {"usbds2_big", 0, "usbds2_set_big_power fun", std::bind(&info_fun, std::placeholders::_1, "usbds2_set_big_power", "")},
+            {"usbds1", 0, "usbds1_set_switch fun", std::bind(&info_fun, std::placeholders::_1, "usbds1_set_switch", "")},
+            {"usbds2", 0, "usbds2_set_switch fun", std::bind(&info_fun, std::placeholders::_1, "usbds2_set_switch", "")},
+            {"usbds3", 0, "usbds3_set_switch fun", std::bind(&info_fun, std::placeholders::_1, "usbds3_set_switch", "")},
+            {"grove_uart", 0, "grove_uart_set_switch fun", std::bind(&info_fun, std::placeholders::_1, "grove_uart_set_switch", "")},
+            {"grove_iic", 0, "grove_iic_set_switch fun", std::bind(&info_fun, std::placeholders::_1, "grove_iic_set_switch", "")},
+            {"flash_switch", 0, "flash_save_switch fun", std::bind(&info_fun, std::placeholders::_1, "flash_save_switch", "")},
+            {"flash_value", 0, "flash_save_value_config fun", std::bind(&info_fun, std::placeholders::_1, "flash_save_value_config", "")},
+            {"poweroff", 0, "poweroff fun", std::bind(&info_fun, std::placeholders::_1, "poweroff", "")},
+
+        };
+        // clang-format on
+        for (auto &cmd : cmd_list) {
+            a.add(cmd.fun, cmd.flage, cmd.dec);
+        }
+        a.add<std::string>("data", 'd', "call param", false);
+        a.add<std::string>("DataRaw", 'D', "call param raw", false);
+        a.parse_check(argc, argv);
+        for (auto &cmd : cmd_list) {
+            if (a.exist(cmd.fun)) {
+                cmd.fun_call(a);
+            }
+        }
+    } else if (std::string(argv[1]) == "exec") {
+        cmdline::parser a;
+        a.add("list", 'l', "list call function");
+        a.add<std::string>("fun", 'f', "call ax650c_ec_proxy function", false);
+        a.add<std::string>("data", 'd', "call ax650c_ec_proxy function input data", false);
+        a.parse_check(argc, argv);
+        exec_fun(a);
+    } else if (std::string(argv[1]) == "echo") {
+        cmdline::parser a;
+        a.add("button", 'b', "button event");
+        a.parse_check(argc, argv);
+        echo_fun(a);
+        std::cout << a.usage() << std::endl;
+        return 0;
+    } else {
+        goto error_print_exit;
+    }
+
+    return 0;
+}
diff --git a/projects/llm_framework/main_ax650_ec_proxy/src/ec_proxy_main.cpp b/projects/llm_framework/main_ax650_ec_proxy/src/ec_proxy_main.cpp
new file mode 100644
index 00000000..38c40fdb
--- /dev/null
+++ b/projects/llm_framework/main_ax650_ec_proxy/src/ec_proxy_main.cpp
@@ -0,0 +1,963 @@
+/*
+ * SPDX-FileCopyrightText: 2024 M5Stack Technology CO LTD
+ *
+ * SPDX-License-Identifier: MIT
+ */
+#include "StackFlow.h"
+#include <signal.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+#include <base64.h>
+#include <fstream>
+#include <iostream>
+#include <stdexcept>
+#include <modbus.h>
+#include <arpa/inet.h>
+// #include "cmdline.hpp"
+#include <../../static_lib/include/fmt/core.h>
+#include <../../static_lib/include/fmt/format.h>
+#include "../../../../SDK/components/utilities/include/sample_log.h"
+
+int main_exit_flage = 0;
+static void __sigint(int iSigNo)
+{
+    SLOGW("llm_ec_prox will be exit!");
+    main_exit_flage = 1;
+}
+
+#define CONFIG_AUTO_SET(obj, key)             \
+    if (config_body.contains(#key))           \
+        mode_config_.key = config_body[#key]; \
+    else if (obj.contains(#key))              \
+        mode_config_.key = obj[#key];
+
+using namespace StackFlows;
+
+class llm_ec_prox : public StackFlow {
+private:
+    enum { EVENT_LOAD_CONFIG = EVENT_EXPORT + 1, EVENT_QUEUE_PLAY };
+    std::string ec_prox_event_channel = "ipc:///tmp/llm/ec_prox.event.socket";
+    modbus_t *modbus_ctx;
+    std::mutex modbus_mtx_;
+    typedef struct {
+        int reg_index;    // 16位寄存器索引
+        int byte_offset;  // 在寄存器中字节偏移（0-低字节，1-高字节）
+    } RegPosition;
+    std::unique_ptr<pzmq> pub_ctx_;
+    std::string StackFlow_pack_temp =
+        R"format({{ "created":{},"data":{},"object":{},"request_id":{},"work_id":"ax650_ec_proxy" }})format";
+    std::string StackFlow_pack_all_temp =
+        R"format({{ "created":{},"data":{},"error":{{"code":{},"message":{}}},"object":{},"request_id":{},"work_id":"ax650_ec_proxy" }})format";
+    std::string return_success_result(std::string data, std::string object = "\"\"", std::string request_id = "\"\"")
+    {
+        return fmt::format(StackFlow_pack_temp, std::time(nullptr), data, object, request_id);
+    }
+
+    static llm_ec_prox *self;
+
+    void loop()
+    {
+        uint8_t butt = 0;
+        while (!main_exit_flage) {
+            {
+                std::unique_lock<std::mutex> lock(this->modbus_mtx_);
+                uint8_t tab_rp_bits[2];
+                int rc = modbus_read_input_bits(modbus_ctx, 4, 1, tab_rp_bits);
+                if (rc == -1) {
+                    fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
+                } else {
+                    if (tab_rp_bits[0] != butt) {
+                        if (tab_rp_bits[0]) {
+                            pub_ctx_->send_data(return_success_result("1"));
+                        } else {
+                            pub_ctx_->send_data(return_success_result("0"));
+                        }
+                        butt = tab_rp_bits[0];
+                    }
+                }
+                // std::cout << "butt:" << (int)butt << std::endl;
+            }
+            usleep(20 * 1000);
+        }
+    }
+
+    std::string board_get_power_info(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        // mbpoll -m rtu -b 115200 -P none -a 1 -r 1 -t 3 -c 13 -l 10 /dev/ttyS3
+        uint16_t tab_reg[12];
+        int rc = modbus_read_input_registers(modbus_ctx, 0, 12, tab_reg);
+        if (rc == -1) {
+            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
+        } else {
+            return return_success_result(fmt::format(
+                R"format({{ "pcie0_mv": {} , "pcie0_ma": {} , "pcie1_mv": {} , "pcie1_ma": {} , "usb1_mv": {} , "usb1_ma": {} , "usb2_mv": {} , "usb2_ma": {} , "VDD5V_mv": {} , "VDD5V_ma": {} , "EXT5V_mv": {} , "EXT5V_ma": {} }})format",
+                tab_reg[0], tab_reg[1], tab_reg[2], tab_reg[3], tab_reg[4], tab_reg[5], tab_reg[6], tab_reg[7],
+                tab_reg[8], tab_reg[9], tab_reg[10], tab_reg[11]));
+        }
+        return std::string("Error");
+    }
+
+    std::string fan_get_speed(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        // mbpoll -m rtu -b 115200 -P none -a 1 -r 1 -t 3 -c 13 -l 10 /dev/ttyS3
+        uint16_t tab_reg[2];
+        int rc = modbus_read_input_registers(modbus_ctx, 12, 1, tab_reg);
+        if (rc == -1) {
+            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
+        } else {
+            return return_success_result(fmt::format("{}", tab_reg[0]));
+        }
+        return std::string("Error");
+    }
+    std::string fan_get_pwm(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2];
+        int rc = modbus_read_registers(modbus_ctx, 1, 1, tab_reg);
+        if (rc == -1) {
+            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
+        } else {
+            return return_success_result(fmt::format("{}", tab_reg[0]));
+        }
+        return std::string("Error");
+    }
+    std::string fan_set_pwm(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2];
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_register(modbus_ctx, 1, tab_reg[0]);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+        }
+        return return_success_result("\"ok\"");
+    }
+
+    RegPosition get_rgb_reg_pos(int rgb_index)
+    {
+        RegPosition res;
+        int offset      = rgb_index * 3;  // 每个RGB占3字节
+        res.reg_index   = offset / 2;
+        res.byte_offset = offset % 2;
+        return res;
+    }
+    std::string rgb_get_mode(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2];
+        int rc = modbus_read_registers(modbus_ctx, 11, 1, tab_reg);
+        if (rc == -1) {
+            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
+        } else {
+            return return_success_result(fmt::format("{}", tab_reg[0]));
+        }
+        return std::string("Error");
+    }
+    std::string rgb_set_mode(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2];
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_register(modbus_ctx, 11, tab_reg[0]);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+    std::string rgb_get_size(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2];
+        int rc = modbus_read_registers(modbus_ctx, 10, 1, tab_reg);
+        if (rc == -1) {
+            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
+        } else {
+            return return_success_result(fmt::format("{}", tab_reg[0]));
+        }
+        return std::string("Error");
+    }
+    std::string rgb_set_size(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2];
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_register(modbus_ctx, 10, tab_reg[0]);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+    std::string rgb_get_color(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        int rgb_index = 0;
+        int rgb_color = 0;
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            rgb_index               = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        uint16_t tab_reg[2];
+        RegPosition reg_pos = get_rgb_reg_pos(rgb_index);
+        {
+            int rc = modbus_read_registers(modbus_ctx, reg_pos.reg_index + 12, 2, tab_reg);
+            if (rc == -1) {
+                fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
+            } else {
+                for (int i = 0; i < 2; i++) tab_reg[i] = ntohs(tab_reg[i]);
+                uint8_t *rgb_color_reg = (uint8_t *)tab_reg;
+                rgb_color              = rgb_color_reg[reg_pos.byte_offset] << 16;
+                rgb_color |= rgb_color_reg[reg_pos.byte_offset + 1] << 8;
+                rgb_color |= rgb_color_reg[reg_pos.byte_offset + 2];
+                return return_success_result(fmt::format("{}", rgb_color));
+            }
+        }
+        return std::string("Error");
+    }
+    std::string rgb_set_color(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        int rgb_index = 0;
+        int rgb_color = 0;
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            rgb_index               = (int)fan_data["data"]["rgb_index"];
+            rgb_color               = (int)fan_data["data"]["rgb_color"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        uint16_t tab_reg[2];
+        RegPosition reg_pos = get_rgb_reg_pos(rgb_index);
+        {
+            int rc = modbus_read_registers(modbus_ctx, reg_pos.reg_index + 12, 2, tab_reg);
+            if (rc == -1) {
+                fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
+            }
+            for (int i = 0; i < 2; i++) tab_reg[i] = ntohs(tab_reg[i]);
+        }
+        {
+            uint8_t *rgb_color_reg                 = (uint8_t *)tab_reg;
+            rgb_color_reg[reg_pos.byte_offset]     = (rgb_color & 0xff0000) >> 16;
+            rgb_color_reg[reg_pos.byte_offset + 1] = (rgb_color & 0xff00) >> 8;
+            rgb_color_reg[reg_pos.byte_offset + 2] = (rgb_color & 0xff);
+        }
+        for (int i = 0; i < 2; i++) tab_reg[i] = htons(tab_reg[i]);
+        {
+            int rc = modbus_write_registers(modbus_ctx, reg_pos.reg_index + 12, 2, tab_reg);
+            if (rc == -1) {
+                fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+                return std::string("Error");
+            }
+        }
+        return return_success_result("\"ok\"");
+    }
+
+    std::string lcd_get_mode(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2];
+        int rc = modbus_read_registers(modbus_ctx, 86, 1, tab_reg);
+        if (rc == -1) {
+            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
+        } else {
+            return return_success_result(fmt::format("{}", tab_reg[0]));
+        }
+        return std::string("Error");
+    }
+    std::string lcd_set_mode(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2];
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_register(modbus_ctx, 86, tab_reg[0]);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+    std::string lcd_set_ram(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        std::string rgb_ram_base64;
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            rgb_ram_base64          = fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        uint16_t *rgb_ram = (uint16_t *)malloc(((BASE64_DECODE_OUT_SIZE(rgb_ram_base64.length()) / 2) + 1) * 2);
+        int ram_size      = base64_decode(rgb_ram_base64.c_str(), rgb_ram_base64.length(), (uint8_t *)rgb_ram);
+        ram_size          = ram_size > 1870 ? 1870 : ram_size;
+        int rc            = modbus_write_registers(modbus_ctx, 88, ram_size / 2, rgb_ram);
+        free(rgb_ram);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return std::string("Ok");
+    }
+
+    std::string lcd_echo(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        std::string put_strs;
+        uint16_t tab_reg = 0;
+        uint8_t *c_reg   = (uint8_t *)&tab_reg;
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            put_strs                = fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        auto it = put_strs.begin();
+        while (it != put_strs.end()) {
+            c_reg[0] = *it;
+            ++it;
+            if (it != put_strs.end()) {
+                c_reg[1] = *it;
+                ++it;
+            } else {
+                c_reg[1] = 0;
+            }
+            int rc = modbus_write_register(modbus_ctx, 87, tab_reg);
+            if (rc == -1) {
+                fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+                return std::string("Error");
+            }
+        }
+        return return_success_result("\"ok\"");
+    }
+
+    std::string ip_get(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint32_t value      = 0;
+        uint16_t tab_reg[2] = {0};
+
+        int rc = modbus_read_registers(modbus_ctx, 5, 2, tab_reg);
+        if (rc == -1) {
+            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
+        } else {
+            for (int i = 0; i < 2; i++) tab_reg[i] = ntohs(tab_reg[i]);
+            value = ((uint32_t)tab_reg[0] << 16) | tab_reg[1];
+            char ip_str[INET_ADDRSTRLEN];
+            inet_ntop(AF_INET, &value, ip_str, INET_ADDRSTRLEN);
+            return return_success_result(fmt::format("\"{}\"", ip_str));
+        }
+        return std::string("Error");
+    }
+    std::string ip_set(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint32_t value;
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            std::string ip_str      = fan_data["data"];
+            if (inet_pton(AF_INET, ip_str.c_str(), &value) != 1) {
+                return std::string("Error : 无效的 IP 地址\n");
+            }
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        uint16_t tab_reg[2];
+        tab_reg[0] = (value >> 16) & 0xFFFF;
+        tab_reg[1] = value & 0xFFFF;
+        for (int i = 0; i < 2; i++) tab_reg[i] = htons(tab_reg[i]);
+        int rc = modbus_write_registers(modbus_ctx, 5, 2, tab_reg);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+
+    std::string vddcpu_get(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+
+        int rc = modbus_read_registers(modbus_ctx, 2, 1, tab_reg);
+        if (rc == -1) {
+            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
+        } else {
+            return return_success_result(fmt::format("{}", tab_reg[0]));
+        }
+        return std::string("Error");
+    }
+    std::string vddcpu_set(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_registers(modbus_ctx, 2, 1, tab_reg);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+    std::string modbus_set_speed(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        int modbus_speed;
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            modbus_speed            = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        set_modbus_speed(modbus_speed);
+        return return_success_result("\"ok\"");
+    }
+    std::string modbus_get_speed(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        return return_success_result(fmt::format("{}", _modbus_speed));
+    }
+    void set_modbus_speed(int modbus_speed)
+    {
+        uint16_t tab_reg[2];
+        tab_reg[0] = (modbus_speed >> 16) & 0xFFFF;
+        tab_reg[1] = modbus_speed & 0xFFFF;
+        int rc     = modbus_write_registers(modbus_ctx, 3, 2, tab_reg);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            std::exit(-1);
+        }
+        modbus_close(modbus_ctx);
+        modbus_free(modbus_ctx);
+        modbus_ctx = NULL;
+        usleep(100 * 1000);
+        modbus_ctx = modbus_new_rtu("/dev/ttyS3", modbus_speed, 'N', 8, 1);
+        if (modbus_ctx == NULL) {
+            fprintf(stderr, "Unable to create the context\n");
+            std::exit(-1);
+        }
+        modbus_set_slave(modbus_ctx, 1);  // 设置 Modbus 从站地址为1
+        if (modbus_connect(modbus_ctx) == -1) {
+            fprintf(stderr, "Connection failed: %s\n", modbus_strerror(errno));
+            modbus_free(modbus_ctx);
+            std::exit(-1);
+        }
+        _modbus_speed = modbus_speed;
+    }
+    int _modbus_speed = 115200;
+
+    void _init_modbus()
+    {
+        modbus_ctx = modbus_new_rtu("/dev/ttyS3", 115200, 'N', 8, 1);
+        if (modbus_ctx == NULL) {
+            fprintf(stderr, "Unable to create the context\n");
+            std::exit(-1);
+        }
+        modbus_set_slave(modbus_ctx, 1);  // 设置 Modbus 从站地址为1
+        if (modbus_connect(modbus_ctx) == -1) {
+            fprintf(stderr, "Connection failed: %s\n", modbus_strerror(errno));
+            modbus_free(modbus_ctx);
+            std::exit(-1);
+        }
+        uint8_t tab_rp_bits[2];
+        int rc = modbus_read_input_bits(modbus_ctx, 4, 1, tab_rp_bits);
+        if (rc == -1) {
+            _modbus_speed = 1152000;
+            modbus_ctx    = modbus_new_rtu("/dev/ttyS3", _modbus_speed, 'N', 8, 1);
+            if (modbus_ctx == NULL) {
+                fprintf(stderr, "Unable to create the context\n");
+                std::exit(-1);
+            }
+            modbus_set_slave(modbus_ctx, 1);  // 设置 Modbus 从站地址为1
+            if (modbus_connect(modbus_ctx) == -1) {
+                fprintf(stderr, "Connection failed: %s\n", modbus_strerror(errno));
+                modbus_free(modbus_ctx);
+                std::exit(-1);
+            }
+            uint8_t tab_rp_bits[2];
+            int rc = modbus_read_input_bits(modbus_ctx, 4, 1, tab_rp_bits);
+            if (rc == -1) {
+                fprintf(stderr, "modbus 115200/1152000 速度探测失败: %s\n", modbus_strerror(errno));
+                std::exit(-1);
+            }
+        } else {
+            set_modbus_speed(1152000);
+        }
+    }
+
+    std::string ext_poweroff(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_bit(modbus_ctx, 0, 1);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+    std::string board_poweroff(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_bit(modbus_ctx, 1, 1);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+
+    std::string pcie0_set_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_bit(modbus_ctx, 4, 1);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+    std::string pcie1_set_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_bit(modbus_ctx, 5, 1);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+
+    std::string gl3510_reset(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_bit(modbus_ctx, 6, 1);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+    std::string usbds1_set_big_power(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_bit(modbus_ctx, 7, 1);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+    std::string usbds2_set_big_power(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_bit(modbus_ctx, 8, 1);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+    std::string usbds1_set_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_bit(modbus_ctx, 9, 1);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+    std::string usbds2_set_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_bit(modbus_ctx, 10, 1);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+    std::string usbds3_set_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_bit(modbus_ctx, 11, 1);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+
+    std::string grove_uart_set_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_bit(modbus_ctx, 12, 1);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+    std::string grove_iic_set_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_bit(modbus_ctx, 13, 1);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+
+    std::string flash_save_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_bit(modbus_ctx, 14, 1);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+
+    std::string flash_save_value_config(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_bit(modbus_ctx, 15, 1);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+
+    std::string poweroff(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
+    {
+        uint16_t tab_reg[2] = {0};
+        try {
+            nlohmann::json fan_data = nlohmann::json::parse(data->string());
+            tab_reg[0]              = (int)fan_data["data"];
+        } catch (const nlohmann::json::parse_error &e) {
+            std::cerr << "解析失败：" << e.what() << std::endl;
+            return std::string("Error");
+        } catch (...) {
+            std::cout << "data 解析异常" << std::endl;
+            return std::string("Error");
+        }
+        int rc = modbus_write_register(modbus_ctx, 0, tab_reg[0] >= 1000 ? tab_reg[0] : 1000);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        rc = modbus_write_bit(modbus_ctx, 16, 1);
+        if (rc == -1) {
+            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            return std::string("Error");
+        }
+        return return_success_result("\"ok\"");
+    }
+
+public:
+    llm_ec_prox() : StackFlow("ec_prox")
+    {
+        // setup("", "audio.play", "{\"None\":\"None\"}");
+        self = this;
+        // prob modbus speed
+        _init_modbus();
+
+        pub_ctx_ = std::make_unique<pzmq>(ec_prox_event_channel, ZMQ_PUB);
+
+        // clang-format off
+#define REGISTER_RPC_ACTION(name, func) \
+        rpc_ctx_->register_rpc_action(name, [this](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data){ \
+            std::unique_lock<std::mutex> lock(this->modbus_mtx_); \
+            return this->func(_pzmq, data); \
+        });
+
+        REGISTER_RPC_ACTION("fan_get_speed",                fan_get_speed);
+        REGISTER_RPC_ACTION("fan_get_pwm",                  fan_get_pwm);
+        REGISTER_RPC_ACTION("fan_set_pwm",                  fan_set_pwm);
+
+        REGISTER_RPC_ACTION("rgb_get_mode",                 rgb_get_mode);
+        REGISTER_RPC_ACTION("rgb_set_mode",                 rgb_set_mode);
+        REGISTER_RPC_ACTION("rgb_get_size",                 rgb_get_size);
+        REGISTER_RPC_ACTION("rgb_set_size",                 rgb_set_size);
+        REGISTER_RPC_ACTION("rgb_get_color",                rgb_get_color);
+        REGISTER_RPC_ACTION("rgb_set_color",                rgb_set_color);
+        
+        REGISTER_RPC_ACTION("board_get_power_info",         board_get_power_info);
+
+        REGISTER_RPC_ACTION("ip_get",                       ip_get);
+        REGISTER_RPC_ACTION("ip_set",                       ip_set);
+
+        REGISTER_RPC_ACTION("lcd_get_mode",                 lcd_get_mode);
+        REGISTER_RPC_ACTION("lcd_set_mode",                 lcd_set_mode);
+        REGISTER_RPC_ACTION("lcd_set_ram",                  lcd_set_ram);
+        REGISTER_RPC_ACTION("lcd_echo",                     lcd_echo);
+        
+
+        REGISTER_RPC_ACTION("vddcpu_get",                   vddcpu_get);
+        REGISTER_RPC_ACTION("vddcpu_set",                   vddcpu_set);
+
+        REGISTER_RPC_ACTION("modbus_get_speed",             modbus_get_speed);
+        REGISTER_RPC_ACTION("modbus_set_speed",             modbus_set_speed);
+
+        REGISTER_RPC_ACTION("ext_poweroff",                   ext_poweroff);
+        REGISTER_RPC_ACTION("board_poweroff",                 board_poweroff);
+        REGISTER_RPC_ACTION("pcie0_set_switch",                 pcie0_set_switch);
+        REGISTER_RPC_ACTION("pcie1_set_switch",                 pcie1_set_switch);
+        REGISTER_RPC_ACTION("gl3510_reset",                 gl3510_reset);
+        REGISTER_RPC_ACTION("usbds1_set_big_power",                 usbds1_set_big_power);
+        REGISTER_RPC_ACTION("usbds2_set_big_power",                 usbds2_set_big_power);
+        REGISTER_RPC_ACTION("usbds1_set_switch",                 usbds1_set_switch);
+        REGISTER_RPC_ACTION("usbds2_set_switch",                 usbds2_set_switch);
+        REGISTER_RPC_ACTION("usbds3_set_switch",                 usbds3_set_switch);
+        REGISTER_RPC_ACTION("grove_uart_set_switch",                 grove_uart_set_switch);
+        REGISTER_RPC_ACTION("grove_iic_set_switch",                 grove_iic_set_switch);
+        REGISTER_RPC_ACTION("flash_save_switch",                 flash_save_switch);
+        REGISTER_RPC_ACTION("flash_save_value_config",                 flash_save_value_config);
+        REGISTER_RPC_ACTION("poweroff",                 poweroff);
+
+
+
+
+#undef REGISTER_RPC_ACTION
+        // clang-format on
+    }
+    int setup(const std::string &work_id, const std::string &object, const std::string &data) override
+    {
+        send(std::string("None"), std::string("None"), std::string(""), unit_name_);
+        return -1;
+    }
+    void ax650_ec_prox_exit()
+    {
+        pub_ctx_.reset();
+        for (int i = 0; i < 3; i++) {
+            uint32_t value = 115200;
+            uint16_t tab_reg[2];
+            tab_reg[0] = (value >> 16) & 0xFFFF;
+            tab_reg[1] = value & 0xFFFF;
+            int rc     = modbus_write_registers(modbus_ctx, 3, 2, tab_reg);
+            if (rc == -1) {
+                fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
+            } else {
+                break;
+            }
+        }
+        modbus_close(modbus_ctx);
+        modbus_free(modbus_ctx);
+        std::cout << "all exit!" << std::endl;
+    }
+    ~llm_ec_prox()
+    {
+        ax650_ec_prox_exit();
+    }
+};
+
+llm_ec_prox *llm_ec_prox::self;
+int main(int argc, char *argv[])
+{
+    signal(SIGTERM, __sigint);
+    signal(SIGINT, __sigint);
+    mkdir("/tmp/llm", 0777);
+    llm_ec_prox ec_prox;
+    while (!main_exit_flage) {
+        sleep(1);
+    }
+    ec_prox.llm_firework_exit();
+    return 0;
+}

From 0d3e36fb49f415ea47ca336cdd8d6f430b268b6e Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 14 Nov 2025 14:18:10 +0800
Subject: [PATCH 57/79] [update] vlm support qwen3-vl model, add qwen3-vl-2b
 model. update pzmq close timeout to 1s.

---
 ext_components/StackFlow/stackflow/pzmq.hpp   |   2 +-
 .../models/mode_qwen3-vl-2B-Int4-ax650.json   |  54 ++
 .../tokenizer_qwen3-vl-2B-Int4-ax650.py       | 195 +++++
 projects/llm_framework/main_vlm/src/main.cpp  | 172 +++-
 .../llm_framework/main_vlm/src/runner/LLM.hpp | 761 ++++++++++++++++++
 .../src/runner/Tokenizer/Tokenizer.cpp        | 254 +++---
 .../src/runner/Tokenizer/Tokenizer.hpp        |  56 +-
 .../main_vlm/src/runner/utils/files.cpp       |  46 ++
 .../main_vlm/src/runner/utils/files.hpp       |  17 +
 .../src/runner/utils/image_processor.cpp      | 162 ++++
 .../src/runner/utils/image_processor.hpp      |  14 +
 .../main_vlm/src/runner/utils/mrope.cpp       | 272 +++++++
 .../main_vlm/src/runner/utils/mrope.hpp       |  33 +
 projects/llm_framework/tools/llm_pack.py      |   1 +
 14 files changed, 1834 insertions(+), 205 deletions(-)
 create mode 100644 projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax650.json
 create mode 100644 projects/llm_framework/main_vlm/scripts/tokenizer_qwen3-vl-2B-Int4-ax650.py
 create mode 100644 projects/llm_framework/main_vlm/src/runner/utils/files.cpp
 create mode 100644 projects/llm_framework/main_vlm/src/runner/utils/files.hpp
 create mode 100644 projects/llm_framework/main_vlm/src/runner/utils/image_processor.cpp
 create mode 100644 projects/llm_framework/main_vlm/src/runner/utils/image_processor.hpp
 create mode 100644 projects/llm_framework/main_vlm/src/runner/utils/mrope.cpp
 create mode 100644 projects/llm_framework/main_vlm/src/runner/utils/mrope.hpp

diff --git a/ext_components/StackFlow/stackflow/pzmq.hpp b/ext_components/StackFlow/stackflow/pzmq.hpp
index 8a29bd12..3adbab55 100644
--- a/ext_components/StackFlow/stackflow/pzmq.hpp
+++ b/ext_components/StackFlow/stackflow/pzmq.hpp
@@ -394,7 +394,7 @@ class pzmq {
     }
     void close_zmq()
     {
-        int linger = 0;
+        int linger = 1000;
         zmq_setsockopt(zmq_socket_, ZMQ_LINGER, &linger, sizeof(linger));
         zmq_close(zmq_socket_);
         zmq_ctx_destroy(zmq_ctx_);
diff --git a/projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax650.json b/projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax650.json
new file mode 100644
index 00000000..77470565
--- /dev/null
+++ b/projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax650.json
@@ -0,0 +1,54 @@
+{
+    "mode": "qwen3-vl-2B-Int4-ax650",
+    "type": "vlm",
+    "homepage": "https://huggingface.co/AXERA-TECH/Qwen3-VL-2B-Instruct",
+    "capabilities": [
+        "text_generation",
+        "chat"
+    ],
+    "input_type": [
+        "vlm.chat_completion",
+        "vlm.chat_completion.stream"
+    ],
+    "output_type": [
+        "vlm.utf-8",
+        "vlm.utf-8.stream"
+    ],
+    "mode_param": {
+        "tokenizer_type": 2,
+        "url_tokenizer_model": "http://localhost:8080",
+        "filename_tokens_embed": "model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel": "qwen3_vl_text_post.axmodel",
+        "template_filename_axmodel": "qwen3_vl_text_p128_l%d_together.axmodel",
+        "filename_image_encoder_axmodel": "Qwen3-VL-2B-Instruct_vision.axmodel",
+        "enable_temperature": true,
+        "temperature": 0.7,
+        "enable_top_p_sampling": false,
+        "top_p": 0.9,
+        "enable_top_k_sampling": true,
+        "top_k": 40,
+        "enable_repetition_penalty": false,
+        "repetition_penalty": 1.1,
+        "penalty_window": 50,
+        "axmodel_num": 28,
+        "tokens_embed_num": 151936,
+        "tokens_embed_size": 2048,
+        "b_use_mmap_load_embed": true,
+        "b_video": false,
+        "vision_config.temporal_patch_size": 2,
+        "vision_config.tokens_per_second": 2,
+        "vision_config.spatial_merge_size": 2,
+        "vision_config.patch_size": 16,
+        "vision_config.height": 384,
+        "vision_config.width": 384,
+        "vision_config.fps": 1,
+        "image_token_id": 151655,
+        "video_token_id": 151656,
+        "vision_start_token_id": 151652,
+        "precompute_len": 0,
+        "cmm_size": 1919044,
+        "ext_scripts": [
+            "tokenizer_qwen3-vl-2B-Int4-ax650.py"
+        ]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/scripts/tokenizer_qwen3-vl-2B-Int4-ax650.py b/projects/llm_framework/main_vlm/scripts/tokenizer_qwen3-vl-2B-Int4-ax650.py
new file mode 100644
index 00000000..ed6f1aae
--- /dev/null
+++ b/projects/llm_framework/main_vlm/scripts/tokenizer_qwen3-vl-2B-Int4-ax650.py
@@ -0,0 +1,195 @@
+from transformers import AutoTokenizer, PreTrainedTokenizerFast
+from transformers.tokenization_utils_base import AddedToken
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+
+
+class Tokenizer_Http:
+    def __init__(self, model_id, system_content="You are a helpful assistant."):
+        self.tokenizer = AutoTokenizer.from_pretrained(
+            model_id,
+            trust_remote_code=True,
+            use_fast=False
+        )
+        self.token_ids_cache = []
+        self.system_content = system_content
+
+    def encode(self, content):
+        text = [
+            f'<|im_start|>system\n{self.system_content}<|im_end|>\n'
+            f'<|im_start|>user\n{content}<|im_end|>\n'
+            f'<|im_start|>assistant\n'
+        ]
+        input_ids = self.tokenizer(text)
+        return input_ids["input_ids"][0]
+
+    def encode_vpm_image(self, content="Describe this image.", num_img=1, img_token_num=256):
+        imgs_token = (
+            '<|vision_start|>'
+            + '<|image_pad|>' * img_token_num
+            + '<|vision_end|>'
+        )
+        imgs_token *= num_img
+        text = (
+            f'<|im_start|>system\n{self.system_content}<|im_end|>\n'
+            f'<|im_start|>user\n{imgs_token}{content}<|im_end|>\n'
+            f'<|im_start|>assistant\n'
+        )
+
+        output_kwargs = {
+            'text_kwargs': {'padding': True, 'return_tensors': 'pt'},
+            'images_kwargs': {'return_tensors': 'pt'},
+            'audio_kwargs': {'padding': True, 'return_tensors': 'pt'},
+            'videos_kwargs': {'fps': 2.0, 'return_tensors': 'pt'},
+            'common_kwargs': {'return_tensors': 'pt'},
+        }
+
+        text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
+        return text_inputs["input_ids"].tolist()[0]
+
+
+    def encode_vpm_video(self, content="Describe this image.", num_img=1, img_token_num=256):
+        imgs_token = (
+            '<|vision_start|>'
+            + '<|video_pad|>' * img_token_num * num_img
+            + '<|vision_end|>'
+        )
+
+        text = (
+            f'<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n'
+            f'<|im_start|>user\n{imgs_token}{content}<|im_end|>\n'
+            f'<|im_start|>assistant\n'
+        )
+
+        output_kwargs = {
+            'text_kwargs': {'padding': True, 'return_tensors': 'pt'},
+            'images_kwargs': {'return_tensors': 'pt'},
+            'audio_kwargs': {'padding': True, 'return_tensors': 'pt'},
+            'videos_kwargs': {'fps': 2.0, 'return_tensors': 'pt'},
+            'common_kwargs': {'return_tensors': 'pt'},
+        }
+
+        text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
+        return text_inputs["input_ids"].tolist()[0]
+    
+    def decode(self, token_ids):
+        self.token_ids_cache += token_ids
+        text = self.tokenizer.decode(self.token_ids_cache)
+        if "\ufffd" in text:
+            print("text 中包含非法字符")
+            return ""
+        else:
+            self.token_ids_cache.clear()
+            return text
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+
+    @property
+    def bos_token(self):
+        return self.tokenizer.bos_token
+
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+
+    @property
+    def img_start_token(self):
+        return self.tokenizer.encode("<|vision_start|>")[0]
+
+    @property
+    def img_context_token(self):
+        return self.tokenizer.encode("<|image_pad|>")[0]
+
+
+class Request(BaseHTTPRequestHandler):
+    timeout = 5
+    server_version = 'Apache'
+
+    def do_GET(self):
+        print(self.path)
+        self.send_response(200)
+        self.send_header("type", "get")
+        self.end_headers()
+
+        if self.path == '/bos_id':
+            bos_id = tokenizer.bos_id
+            msg = json.dumps({'bos_id': -1 if bos_id is None else bos_id})
+        elif self.path == '/eos_id':
+            eos_id = tokenizer.eos_id
+            msg = json.dumps({'eos_id': -1 if eos_id is None else eos_id})
+        elif self.path == '/img_start_token':
+            img_start_token = tokenizer.img_start_token
+            msg = json.dumps({'img_start_token': -1 if img_start_token is None else img_start_token})
+        elif self.path == '/img_context_token':
+            img_context_token = tokenizer.img_context_token
+            msg = json.dumps({'img_context_token': -1 if img_context_token is None else img_context_token})
+        else:
+            msg = 'error'
+
+        print(msg)
+        msg = str(msg).encode()
+        self.wfile.write(msg)
+
+    def do_POST(self):
+        data = self.rfile.read(int(self.headers['content-length']))
+        req = json.loads(data.decode())
+
+        if self.path == "/encode":
+            prompt = req['text']
+            b_img_prompt = req.get('img_prompt', False)
+            img_type = req.get('img_type', 'image')  # 默认 image
+
+            if b_img_prompt:
+                if img_type == 'image':
+                    token_ids = tokenizer.encode_vpm_image(
+                        prompt,
+                        req.get("num_img", 1),
+                        req.get("img_token_num", 256)
+                    )
+                elif img_type == 'video':
+                    token_ids = tokenizer.encode_vpm_video(
+                        prompt,
+                        req.get("num_img", 1),
+                        req.get("img_token_num", 256)
+                    )
+                else:
+                    token_ids = tokenizer.encode(prompt)  # fallback
+            else:
+                token_ids = tokenizer.encode(prompt)
+
+            msg = json.dumps({'token_ids': -1 if token_ids is None else token_ids})
+
+        elif self.path == "/decode":
+            req = json.loads(data.decode())
+            token_ids = req['token_ids']
+            text = tokenizer.decode(token_ids)
+            msg = json.dumps({'text': "" if text is None else text})
+        else:
+            msg = 'error'
+
+        self.send_response(200)
+        self.end_headers()
+        self.wfile.write(str(msg).encode())
+
+
+if __name__ == "__main__":
+    args = argparse.ArgumentParser()
+    args.add_argument('--host', type=str, default='localhost')
+    args.add_argument('--port', type=int, default=8080)
+    args.add_argument('--model_id', type=str, default='tokenizer')
+    args.add_argument('--content', type=str, default='You are a helpful assistant.')
+    args = args.parse_args()
+
+    tokenizer = Tokenizer_Http(args.model_id, system_content=args.content)
+    host = (args.host, args.port)
+
+    print(f"http://{args.host}:{args.port}")
+    server = HTTPServer(host, Request)
+    server.serve_forever()
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/main.cpp b/projects/llm_framework/main_vlm/src/main.cpp
index 7b5d5ead..7397ff6b 100644
--- a/projects/llm_framework/main_vlm/src/main.cpp
+++ b/projects/llm_framework/main_vlm/src/main.cpp
@@ -42,17 +42,27 @@ typedef std::function<void(const std::string &data, bool finish)> task_callback_
     else if (obj.contains(#key))              \
         mode_config_.key = obj[#key];
 
+#define QWEN_CONFIG_AUTO_SET(obj, key)             \
+    if (config_body.contains(#key))                \
+        qwen_mode_config_.key = config_body[#key]; \
+    else if (obj.contains(#key))                   \
+        qwen_mode_config_.key = obj[#key];
+
 class llm_task {
 private:
     static std::atomic<unsigned int> next_port_;
     std::atomic_bool tokenizer_server_flage_;
     unsigned int port_;
     pid_t tokenizer_pid_ = -1;
+    enum class ModelType { Unknown = 0, Qwen, InternVL, InternVL_CTX };
+    ModelType model_type_ = ModelType::Unknown;
 
 public:
     LLMAttrType mode_config_;
+    Config qwen_mode_config_;
     std::unique_ptr<LLM> lLaMa_;
     std::unique_ptr<LLM_CTX> lLaMa_ctx_;
+    std::unique_ptr<LLM_Qwen> qwen_;
     std::string model_;
     std::string response_format_;
     std::vector<std::string> inputs_;
@@ -61,6 +71,10 @@ class llm_task {
     std::vector<std::vector<unsigned char>> images_data;
     std::vector<cv::Mat> mats;
     std::vector<unsigned short> img_embed;
+    std::vector<std::vector<unsigned short>> imgs_embed;
+    std::vector<std::vector<float>> deepstack_features;
+    std::vector<int> visual_pos_mask;
+    std::vector<std::vector<int>> position_ids;
     std::string prompt_;
     std::string last_reply;
     std::vector<int> tokens_ids, tokens_diff;
@@ -121,6 +135,28 @@ class llm_task {
         }
     }
 
+    void vlm_reset()
+    {
+        std::vector<unsigned short>().swap(prompt_data_);
+
+        for (auto &inner_vec : imgs_embed) {
+            std::vector<unsigned short>().swap(inner_vec);
+        }
+        std::vector<std::vector<unsigned short>>().swap(imgs_embed);
+
+        for (auto &inner_vec : deepstack_features) {
+            std::vector<float>().swap(inner_vec);
+        }
+        std::vector<std::vector<float>>().swap(deepstack_features);
+
+        std::vector<int>().swap(visual_pos_mask);
+
+        for (auto &inner_vec : position_ids) {
+            std::vector<int>().swap(inner_vec);
+        }
+        std::vector<std::vector<int>>().swap(position_ids);
+    }
+
     int load_model(const nlohmann::json &config_body)
     {
         if (parse_config(config_body)) {
@@ -179,7 +215,19 @@ class llm_task {
             CONFIG_AUTO_SET(file_body["mode_param"], vpm_width);
             CONFIG_AUTO_SET(file_body["mode_param"], vpm_height);
             CONFIG_AUTO_SET(file_body["mode_param"], precompute_len);
-
+            CONFIG_AUTO_SET(file_body["mode_param"], b_video);
+
+            QWEN_CONFIG_AUTO_SET(file_body["mode_param"], vision_config.temporal_patch_size);
+            QWEN_CONFIG_AUTO_SET(file_body["mode_param"], vision_config.tokens_per_second);
+            QWEN_CONFIG_AUTO_SET(file_body["mode_param"], vision_config.spatial_merge_size);
+            QWEN_CONFIG_AUTO_SET(file_body["mode_param"], vision_config.patch_size);
+            QWEN_CONFIG_AUTO_SET(file_body["mode_param"], vision_config.width);
+            QWEN_CONFIG_AUTO_SET(file_body["mode_param"], vision_config.height);
+            QWEN_CONFIG_AUTO_SET(file_body["mode_param"], vision_config.fps);
+
+            QWEN_CONFIG_AUTO_SET(file_body["mode_param"], image_token_id);
+            QWEN_CONFIG_AUTO_SET(file_body["mode_param"], video_token_id);
+            QWEN_CONFIG_AUTO_SET(file_body["mode_param"], vision_start_token_id);
             {
                 auto has_http = [](const std::string &s) { return s.find("http") != std::string::npos; };
 
@@ -196,10 +244,9 @@ class llm_task {
                 auto start_tokenizer_server = [&](const std::string &tokenizer_file) {
                     if (tokenizer_file.empty()) return;
                     if (tokenizer_server_flage_.load()) return;
-
                     tokenizer_pid_ = fork();
                     if (tokenizer_pid_ == 0) {
-                        setenv("PYTHONPATH", "/opt/m5stack/lib/llm/site-packages", 1);
+                        setenv("PYTHONPATH", "/opt/m5stack/lib/vlm/site-packages", 1);
                         const std::string port_str = std::to_string(port_);
                         const std::string model_id = base_model + "tokenizer";
 
@@ -234,11 +281,25 @@ class llm_task {
                     SLOGI("filename_tokenizer_model: %s", mode_config_.filename_tokenizer_model.c_str());
                 }
             }
-            mode_config_.filename_tokens_embed          = base_model + mode_config_.filename_tokens_embed;
-            mode_config_.filename_post_axmodel          = base_model + mode_config_.filename_post_axmodel;
-            mode_config_.template_filename_axmodel      = base_model + mode_config_.template_filename_axmodel;
+
+            {
+                std::string encoder_name = mode_config_.filename_image_encoder_axmodel;
+                std::transform(encoder_name.begin(), encoder_name.end(), encoder_name.begin(), ::tolower);
+
+                if (encoder_name.find("qwen3") != std::string::npos)
+                    model_type_ = ModelType::Qwen;
+                else if (encoder_name.find("internvl3") != std::string::npos && mode_config_.precompute_len > 0)
+                    model_type_ = ModelType::InternVL_CTX;
+                else if (encoder_name.find("internvl3") != std::string::npos)
+                    model_type_ = ModelType::InternVL;
+                else
+                    model_type_ = ModelType::Unknown;
+            }
+            mode_config_.filename_tokens_embed           = base_model + mode_config_.filename_tokens_embed;
+            mode_config_.filename_post_axmodel           = base_model + mode_config_.filename_post_axmodel;
+            mode_config_.filename_image_encoder_axmodel  = base_model + mode_config_.filename_image_encoder_axmodel;
+            mode_config_.template_filename_axmodel       = base_model + mode_config_.template_filename_axmodel;
             mode_config_.filename_vpm_resampler_axmodedl = base_model + mode_config_.filename_vpm_resampler_axmodedl;
-            mode_config_.filename_image_encoder_axmodel = base_model + mode_config_.filename_image_encoder_axmodel;
             mode_config_.runing_callback = [this](int *p_token, int n_token, const char *p_str, float token_per_sec,
                                                   void *reserve) {
                 if (this->out_callback_) {
@@ -246,20 +307,40 @@ class llm_task {
                 }
             };
 
-            if (mode_config_.precompute_len > 0) {
-                lLaMa_ctx_ = std::make_unique<LLM_CTX>();
-                if (!lLaMa_ctx_->Init(mode_config_)) {
-                    lLaMa_ctx_->Deinit();
-                    lLaMa_ctx_.reset();
-                    return -2;
+            switch (model_type_) {
+                case ModelType::InternVL: {
+                    lLaMa_ = std::make_unique<LLM>();
+                    if (!lLaMa_->Init(mode_config_)) {
+                        lLaMa_->Deinit();
+                        lLaMa_.reset();
+                        return -2;
+                    }
+                    break;
                 }
-            } else {
-                lLaMa_ = std::make_unique<LLM>();
-                if (!lLaMa_->Init(mode_config_)) {
-                    lLaMa_->Deinit();
-                    lLaMa_.reset();
-                    return -2;
+
+                case ModelType::InternVL_CTX: {
+                    lLaMa_ctx_ = std::make_unique<LLM_CTX>();
+                    if (!lLaMa_ctx_->Init(mode_config_)) {
+                        lLaMa_ctx_->Deinit();
+                        lLaMa_ctx_.reset();
+                        return -2;
+                    }
+                    break;
+                }
+
+                case ModelType::Qwen: {
+                    qwen_ = std::make_unique<LLM_Qwen>();
+                    if (!qwen_->Init(mode_config_)) {
+                        qwen_->Deinit();
+                        qwen_.reset();
+                        return -2;
+                    }
+                    break;
                 }
+                default:
+                    ALOGE("Unknown model type in filename_image_encoder_axmodel: %s",
+                          mode_config_.filename_image_encoder_axmodel.c_str());
+                    return -3;
             }
 
             if (lLaMa_ctx_) {
@@ -421,6 +502,33 @@ class llm_task {
                     if (out_callback_) out_callback_(last_reply, true);
                 }
             }
+
+            if (qwen_) {
+                if (images_data.empty()) {
+                    qwen_->Encode(prompt_data_, position_ids, qwen_mode_config_, prompt_complete(msg));
+                    last_reply = qwen_->Run(prompt_data_, position_ids, deepstack_features, visual_pos_mask);
+                    if (out_callback_) out_callback_(last_reply, true);
+                } else {
+                    for (const auto &img_buf : images_data) {
+                        cv::Mat src = cv::imdecode(img_buf, cv::IMREAD_COLOR);
+                        if (src.empty()) {
+                            std::cerr << "Decode failed!" << std::endl;
+                            continue;
+                        }
+                        mats.push_back(src);
+                    }
+                    images_data.clear();
+                    if (mats.empty()) return;
+                    std::vector<std::vector<unsigned short>> all_embeds;
+                    qwen_->EncodeImage(mats, mode_config_.b_video, qwen_mode_config_, all_embeds, deepstack_features);
+                    mats.clear();
+                    qwen_->Encode(all_embeds, mode_config_.b_video, prompt_data_, position_ids, visual_pos_mask,
+                                  qwen_mode_config_, prompt_complete(msg));
+                    last_reply = qwen_->Run(prompt_data_, position_ids, deepstack_features, visual_pos_mask);
+                    if (out_callback_) out_callback_(last_reply, true);
+                    vlm_reset();
+                }
+            }
         } catch (...) {
             SLOGW("lLaMa_->Run have error!");
         }
@@ -428,7 +536,9 @@ class llm_task {
 
     bool pause()
     {
-        lLaMa_->Stop();
+        if (lLaMa_) lLaMa_->Stop();
+        if (lLaMa_ctx_) lLaMa_ctx_->Stop();
+        if (qwen_) qwen_->Stop();
         return true;
     }
 
@@ -439,8 +549,10 @@ class llm_task {
             waitpid(tokenizer_pid_, nullptr, 0);
             tokenizer_pid_ = -1;
         }
-        lLaMa_->Deinit();
-        lLaMa_.reset();
+        if (lLaMa_) lLaMa_->Deinit();
+        if (lLaMa_) lLaMa_.reset();
+        if (qwen_) qwen_->Deinit();
+        if (qwen_) qwen_.reset();
         return true;
     }
 
@@ -476,6 +588,12 @@ class llm_task {
         if (lLaMa_) {
             lLaMa_->Deinit();
         }
+        if (lLaMa_ctx_) {
+            lLaMa_ctx_->Deinit();
+        }
+        if (qwen_) {
+            qwen_->Deinit();
+        }
     }
 };
 
@@ -530,12 +648,14 @@ class llm_vlm : public StackFlow {
         if (!(llm_task_obj && llm_channel)) {
             return;
         }
-        llm_task_obj->lLaMa_->Stop();
+        if (llm_task_obj->lLaMa_) llm_task_obj->lLaMa_->Stop();
+        if (llm_task_obj->lLaMa_ctx_) llm_task_obj->lLaMa_ctx_->Stop();
+        if (llm_task_obj->qwen_) llm_task_obj->qwen_->Stop();
     }
 
     void pause(const std::string &work_id, const std::string &object, const std::string &data) override
     {
-        SLOGI("llm_asr::work:%s", data.c_str());
+        SLOGI("llm_vlm::work:%s", data.c_str());
 
         nlohmann::json error_body;
         int work_id_num = sample_get_work_id_num(work_id);
@@ -626,7 +746,9 @@ class llm_vlm : public StackFlow {
         if (!(llm_task_obj && llm_channel)) {
             return;
         }
-        llm_task_obj->lLaMa_->Stop();
+        if (llm_task_obj->lLaMa_) llm_task_obj->lLaMa_->Stop();
+        if (llm_task_obj->lLaMa_ctx_) llm_task_obj->lLaMa_ctx_->Stop();
+        if (llm_task_obj->qwen_) llm_task_obj->qwen_->Stop();
     }
 
     void task_camera_data(const std::weak_ptr<llm_task> llm_task_obj_weak,
diff --git a/projects/llm_framework/main_vlm/src/runner/LLM.hpp b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
index 0caf4641..6a3b90fe 100644
--- a/projects/llm_framework/main_vlm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
@@ -4,6 +4,8 @@
 #include <cmath>
 #include <numeric>
 #include "bfloat16.hpp"
+#include "image_processor.hpp"
+#include "mrope.hpp"
 #include "Tokenizer/Tokenizer.hpp"
 #include "LLMEmbedSelector.hpp"
 #include "ax_model_runner/ax_model_runner_ax650.hpp"
@@ -78,6 +80,7 @@ struct LLMAttrType {
     bool b_use_mmap_load_embed        = false;
     bool b_dynamic_load_axmodel_layer = false;
     bool b_use_mmap_load_layer        = true;
+    bool b_video                      = false;
 
     LLMRuningCallback runing_callback = nullptr;
     void *reserve                     = nullptr;
@@ -1710,6 +1713,764 @@ class LLM_CTX {
 
         final_out = tokenizer->Decode(token_ids);
 
+        return final_out;
+    }
+};
+
+class LLM_Qwen {
+private:
+    std::shared_ptr<BaseTokenizer> tokenizer;
+    LLaMaEmbedSelector embed_selector;
+
+    LLMAttrType _attr;
+
+    struct LLMLayer {
+        ax_runner_ax650 layer;
+        std::string filename;
+        MMap layer_buffer;
+        std::vector<char> layer_buffer_vec;
+    };
+
+    std::vector<LLMLayer> llama_layers;
+    ax_runner_ax650 llama_post;
+    ax_runner_ax650 image_encoder;
+    int deepstack_features_num = 3;
+    // int prefill_grpid = 1;
+    int decode_grpid = 0;
+
+    bool b_stop = false;
+
+    LLMPostprocess postprocess;
+    static int post_process(LLMPostprocess &postprocess, unsigned short *p, int n, std::vector<int> &history,
+                            float *val = 0)
+    {
+        std::vector<float> logits(n);
+        for (int i = 0; i < n; i++) {
+            unsigned int proc = p[i] << 16;
+            logits[i]         = *reinterpret_cast<float *>(&proc);
+        }
+
+        return postprocess.apply(logits, history);
+    }
+
+public:
+    bool Init(LLMAttrType attr)
+    {
+        ALOGI("LLM init start");
+        int remain_cmm = get_remaining_cmm_size();
+        ALOGI("Total CMM:%d MB", remain_cmm);
+
+        t_cqdm cqdm = create_cqdm(attr.axmodel_num + 3, 32);
+        this->_attr = attr;
+        tokenizer   = CreateTokenizer(attr.tokenizer_type);
+        if (!tokenizer->Init(attr.url_tokenizer_model, attr.b_bos, attr.b_eos)) {
+            ALOGE("tokenizer.Init(%s, %d, %d) failed", attr.url_tokenizer_model.c_str(), attr.b_bos, attr.b_eos);
+            return false;
+        }
+        update_cqdm(&cqdm, 0, "count", "tokenizer init ok");
+
+        if (!embed_selector.Init(attr.filename_tokens_embed, attr.tokens_embed_num, attr.tokens_embed_size,
+                                 attr.b_use_mmap_load_embed)) {
+            ALOGE("embed_selector.Init(%s, %d, %d) failed", attr.filename_tokens_embed.c_str(), attr.tokens_embed_num,
+                  attr.tokens_embed_size);
+            return false;
+        }
+        update_cqdm(&cqdm, 1, "count", "embed_selector init ok");
+
+        llama_layers.resize(attr.axmodel_num);
+
+        ALOGI("attr.axmodel_num:%d", attr.axmodel_num);
+        char axmodel_path[1024];
+        for (int i = 0; i < attr.axmodel_num; i++) {
+            sprintf(axmodel_path, attr.template_filename_axmodel.c_str(), i);
+            llama_layers[i].filename = axmodel_path;
+
+            if (!attr.b_dynamic_load_axmodel_layer) {
+                int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), false);
+                if (ret != 0) {
+                    ALOGE("init axmodel(%s) failed", llama_layers[i].filename.c_str());
+                    return false;
+                }
+                int remain_cmm = get_remaining_cmm_size();
+                sprintf(axmodel_path, "init %d axmodel ok,remain_cmm(%d MB)", i, remain_cmm);
+                update_cqdm(&cqdm, i + 2, "count", axmodel_path);
+            } else {
+                if (!attr.b_use_mmap_load_layer) {
+                    if (!read_file(llama_layers[i].filename, llama_layers[i].layer_buffer_vec)) {
+                        ALOGE("read_file(%s) failed", llama_layers[i].filename.c_str());
+                        return false;
+                    }
+                } else {
+                    llama_layers[i].layer_buffer.open_file(llama_layers[i].filename.c_str());
+                }
+
+                sprintf(axmodel_path, "read_file %s ok", llama_layers[i].filename.c_str());
+                update_cqdm(&cqdm, i + 2, "count", axmodel_path);
+            }
+        }
+
+        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), false);
+        if (ret != 0) {
+            ALOGE("init post axmodel(%s) failed", attr.filename_post_axmodel.c_str());
+            return false;
+        }
+        remain_cmm = get_remaining_cmm_size();
+        sprintf(axmodel_path, "init post axmodel ok,remain_cmm(%d MB)", remain_cmm);
+        update_cqdm(&cqdm, attr.axmodel_num + 2, "count", axmodel_path);
+
+        ret = image_encoder.init(attr.filename_image_encoder_axmodel.c_str(), false);
+        if (ret != 0) {
+            ALOGE("init image_encoder axmodel(%s) failed", attr.filename_image_encoder_axmodel.c_str());
+            return false;
+        }
+
+        remain_cmm = get_remaining_cmm_size();
+        sprintf(axmodel_path, "init vpm axmodel ok,remain_cmm(%d MB)", remain_cmm);
+        update_cqdm(&cqdm, attr.axmodel_num + 3, "count", axmodel_path);
+
+        _attr.IMAGE_CONTEXT_TOKEN = tokenizer->GetImgContextID();
+        _attr.IMAGE_START_TOKEN   = tokenizer->GetImgStartID();
+
+        ALOGI("IMAGE_CONTEXT_TOKEN: %d, IMAGE_START_TOKEN: %d", _attr.IMAGE_CONTEXT_TOKEN, _attr.IMAGE_START_TOKEN);
+
+        _attr.IMAGE_ENCODER_INPUT_NCHW = -1;
+        for (size_t i = 1; i < image_encoder.get_input(0).vShape.size(); i++) {
+            if (image_encoder.get_input(0).vShape[i] == 3) {
+                if (i == 1) {
+                    _attr.IMAGE_ENCODER_INPUT_NCHW = 1;
+                } else if (i == 3) {
+                    _attr.IMAGE_ENCODER_INPUT_NCHW = 0;
+                }
+            }
+        }
+        if (_attr.IMAGE_ENCODER_INPUT_NCHW == -1) {
+            ALOGE("image encoder input nchw or nhwc not found");
+            return false;
+        }
+
+        if (_attr.IMAGE_ENCODER_INPUT_NCHW == 1) {
+            ALOGE("Qwen2.5_VL Image Encoder just support NHWC");
+            return false;
+        }
+
+        int output_elem_size = 1;
+        for (int i = 0; i < image_encoder.get_output(0).vShape.size(); i++) {
+            output_elem_size *= image_encoder.get_output(0).vShape[i];
+        }
+
+        if (output_elem_size * 2 == image_encoder.get_output(0).nSize) {
+            _attr.IMAGE_ENCODER_OUTPUT_BF16 = 1;
+            ALOGI("image encoder output bf16");
+        } else if (output_elem_size * 4 == image_encoder.get_output(0).nSize) {
+            _attr.IMAGE_ENCODER_OUTPUT_BF16 = 0;
+            ALOGI("image encoder output float32");
+        } else {
+            ALOGE("image encoder output not support");
+            return false;
+        }
+
+        if (attr.b_dynamic_load_axmodel_layer) {
+            auto &layer = llama_layers[0];
+            int ret;
+            if (_attr.b_use_mmap_load_layer) {
+                ret = layer.layer.init((char *)layer.layer_buffer.data(), layer.layer_buffer.size());
+            } else {
+                ret = layer.layer.init(layer.layer_buffer_vec.data(), layer.layer_buffer_vec.size());
+            }
+            if (ret != 0) {
+                ALOGE("init axmodel(%s) failed", layer.filename.c_str());
+            }
+        }
+
+        {
+            _attr.max_token_len = llama_layers[0].layer.get_input("mask").nSize / sizeof(unsigned short) - 1;
+            printf("\n");
+            ALOGI("max_token_len : %d", _attr.max_token_len);
+            _attr.kv_cache_size = llama_layers[0].layer.get_output("K_cache_out").nSize / sizeof(unsigned short);
+            _attr.kv_cache_num =
+                llama_layers[0].layer.get_input("K_cache").nSize / _attr.kv_cache_size / sizeof(unsigned short);
+            ALOGI("kv_cache_size : %d, kv_cache_num: %d", _attr.kv_cache_size, _attr.kv_cache_num);
+            if (_attr.max_token_len > _attr.kv_cache_num) {
+                ALOGE("max_token_len(%d) > kv_cache_num(%d)", _attr.max_token_len, _attr.kv_cache_num);
+                return false;
+            }
+
+            _attr.prefill_token_num = llama_layers[0].layer.get_input(1, "indices").vShape[1];
+            ALOGI("prefill_token_num : %d", _attr.prefill_token_num);
+            for (size_t i = 0; i < llama_layers[0].layer.get_num_input_groups() - 1; i++) {
+                int prefill_max_kv_cache_num = llama_layers[0].layer.get_input(i + 1, "K_cache").vShape[1];
+                ALOGI("grp: %ld, prefill_max_token_num : %d", i + 1, prefill_max_kv_cache_num);
+                _attr.prefill_max_kv_cache_num_grp.push_back(prefill_max_kv_cache_num);
+            }
+            _attr.prefill_max_token_num =
+                _attr.prefill_max_kv_cache_num_grp[_attr.prefill_max_kv_cache_num_grp.size() - 1];
+            ALOGI("prefill_max_token_num : %d", _attr.prefill_max_token_num);
+        }
+        if (attr.b_dynamic_load_axmodel_layer) {
+            for (int i = 0; i < attr.axmodel_num; i++) {
+                auto &layer = llama_layers[i];
+                layer.layer.deinit();
+            }
+        }
+
+        nlohmann::json dynamic_config;
+
+        dynamic_config["enable_temperature"] = _attr.enable_temperature;
+        dynamic_config["temperature"]        = _attr.temperature;
+
+        dynamic_config["enable_repetition_penalty"] = _attr.enable_repetition_penalty;
+        dynamic_config["repetition_penalty"]        = _attr.repetition_penalty;
+        dynamic_config["penalty_window"]            = _attr.penalty_window;
+
+        dynamic_config["enable_top_p_sampling"] = _attr.enable_top_p_sampling;
+        dynamic_config["top_p"]                 = _attr.top_p;
+
+        dynamic_config["enable_top_k_sampling"] = _attr.enable_top_k_sampling;
+        dynamic_config["top_k"]                 = _attr.top_k;
+
+        if (!postprocess.load_config(attr.post_config_path)) {
+            ALOGW("load postprocess config(%s) failed", attr.post_config_path.c_str());
+        }
+
+        if (!postprocess.load_config(dynamic_config)) {
+            ALOGW("load postprocess config(%s) failed", dynamic_config.dump(4).c_str());
+        }
+
+        ALOGI("LLM init ok");
+        remain_cmm = get_remaining_cmm_size();
+        ALOGI("Left CMM:%d MB", remain_cmm);
+        return true;
+    }
+
+    LLMAttrType *getAttr()
+    {
+        return &_attr;
+    }
+
+    LLMPostprocess *getPostprocess()
+    {
+        return &postprocess;
+    }
+
+    void Deinit()
+    {
+        for (int i = 0; i < _attr.axmodel_num; i++) {
+            llama_layers[i].layer.release();
+        }
+        llama_post.release();
+        image_encoder.release();
+        embed_selector.Deinit();
+    }
+
+    void Stop()
+    {
+        b_stop = true;
+    }
+
+    int EncodeImage(std::vector<cv::Mat> &src, bool b_video, Config &cfg,
+                    std::vector<std::vector<unsigned short>> &out_embed,
+                    std::vector<std::vector<float>> &deepstack_features)
+    {
+        int temporal_patch_size = cfg.vision_config.temporal_patch_size;
+        int merge_size          = cfg.vision_config.spatial_merge_size;
+        int patch_size          = cfg.vision_config.patch_size;
+        int ret;
+        timer t;
+        t.start();
+
+        unsigned int grid_h = cfg.vision_config.height / cfg.vision_config.patch_size;
+        unsigned int grid_w = cfg.vision_config.width / cfg.vision_config.patch_size;
+        std::vector<std::vector<unsigned char>> pixel_values;
+        int w = cfg.vision_config.width, h = cfg.vision_config.height;
+        int channel = src[0].channels();
+        int hwc     = grid_h * grid_w * temporal_patch_size * patch_size * patch_size * channel;
+
+        if (!b_video) {
+            for (int i = 0; i < src.size(); i++) {
+                std::vector<std::vector<unsigned char>> img_values;
+                std::vector<cv::Mat> si{src[i]};
+                Qwen2VideoProcessor(si, img_values, h, w, temporal_patch_size, merge_size, patch_size);
+                pixel_values.push_back(img_values[0]);
+            }
+            for (size_t i = 0; i < pixel_values.size(); i++) {
+                cfg.image_grid_thw.push_back({1, static_cast<int>(grid_h), static_cast<int>(grid_w)});
+            }
+        } else {
+            Qwen2VideoProcessor(src, pixel_values, h, w, temporal_patch_size, merge_size, patch_size);
+            cfg.video_grid_thw = {
+                {static_cast<int>(pixel_values.size()), static_cast<int>(grid_h), static_cast<int>(grid_w)}};
+        }
+
+        ALOGI("pixel_values size %d", pixel_values.size());
+        ALOGI("grid_h %d grid_w %d", grid_h, grid_w);
+        int cnt = 0;
+        if (out_embed.empty()) {
+            out_embed.resize(pixel_values.size());
+        }
+
+        deepstack_features.clear();
+        for (int i = 0; i < pixel_values.size(); i++) {
+            void *data = image_encoder.get_input(0).pVirAddr;
+            memcpy(data, pixel_values[i].data(), hwc);
+            image_encoder.inference();
+
+            size_t size = image_encoder.get_output(0).nSize / sizeof(float);
+            if (out_embed[i].empty()) {
+                out_embed[i].resize(size);
+            }
+
+            float *output_data = (float *)image_encoder.get_output(0).pVirAddr;
+            for (size_t j = 0; j < size; j++) {
+                out_embed[i][j] = bfloat16(output_data[j]).data;
+            }
+
+            for (int j = 0; j < deepstack_features_num; j++) {
+                size_t size = image_encoder.get_output(j + 1).nSize / sizeof(float);
+                std::vector<float> feature(size);
+
+                float *output_data = (float *)image_encoder.get_output(j + 1).pVirAddr;
+                for (size_t k = 0; k < size; k++) {
+                    feature[k] = output_data[k];
+                }
+
+                // 将不同image 的 deepstack feature 拼接到一起
+                if (i == 0) {
+                    deepstack_features.push_back(feature);
+                } else {
+                    deepstack_features[j].insert(deepstack_features[j].end(), feature.begin(), feature.end());
+                }
+            }
+        }
+
+        ALOGI("image encode time : %f ms, size : %d", t.cost(), out_embed.size());
+        return 0;
+    }
+
+    int GetPositionIds(std::vector<int> &input_ids, std::vector<std::vector<int>> &position_ids, Config &cfg)
+    {
+        position_ids = get_rope_index(cfg, input_ids, cfg.image_grid_thw, cfg.video_grid_thw);
+        return 0;
+    }
+
+    int Encode(std::vector<unsigned short> &out_embed, std::vector<std::vector<int>> &position_ids, Config &cfg,
+               std::string prompt = "What is in the image?")
+    {
+        ImageInfo img_info;
+        img_info.img_prompt        = false;
+        std::vector<int> input_ids = tokenizer->Encode(prompt, img_info);
+        if (input_ids.size() > _attr.prefill_max_token_num) {
+            ALOGE("input_ids(%d) > prefill_max_token_num(%d)", input_ids.size(), _attr.prefill_max_token_num);
+            return -1;
+        }
+        out_embed.resize(input_ids.size() * _attr.tokens_embed_size);
+
+        for (size_t i = 0; i < input_ids.size(); i++) {
+            embed_selector.getByIndex(input_ids[i], out_embed.data() + i * _attr.tokens_embed_size);
+        }
+
+        cfg.image_grid_thw.clear();
+        cfg.video_grid_thw.clear();
+        GetPositionIds(input_ids, position_ids, cfg);
+        return 0;
+    }
+
+    int Encode(std::vector<std::vector<unsigned short>> &img_embed, bool b_video,
+               std::vector<unsigned short> &out_embed, std::vector<std::vector<int>> &position_ids,
+               std::vector<int> &visual_pos_mask, Config &cfg, std::string prompt = "What is in the image?")
+    {
+        ImageInfo img_info;
+        img_info.img_prompt        = true;
+        img_info.type              = b_video ? ImgType::Video : ImgType::Image;
+        img_info.img_token_num     = img_embed[0].size() / _attr.tokens_embed_size;
+        img_info.num_img           = img_embed.size();
+        std::vector<int> input_ids = tokenizer->Encode(prompt, img_info);
+        ALOGI("input_ids size:%d", input_ids.size());
+        std::vector<int> offsets;
+        int vision_start_token_id = cfg.vision_start_token_id;
+        for (size_t i = 0; i < input_ids.size() - 1; i++) {
+            if (input_ids[i] == vision_start_token_id) {
+                int offset = i + 1;
+                ALOGI("offset %d", offset);
+                offsets.push_back(offset);
+            }
+        }
+
+        visual_pos_mask.resize(input_ids.size());
+        for (size_t i = 0; i < input_ids.size(); i++) {
+            if (input_ids[i] == cfg.image_token_id || input_ids[i] == cfg.video_token_id) {
+                visual_pos_mask[i] = 1;
+            } else {
+                visual_pos_mask[i] = 0;
+            }
+        }
+
+        if (input_ids.size() > _attr.prefill_max_token_num) {
+            ALOGE("input_ids(%ld) > prefill_max_token_num(%d)", input_ids.size(), _attr.prefill_max_token_num);
+            return -1;
+        }
+        out_embed.resize(input_ids.size() * _attr.tokens_embed_size);
+
+        for (size_t i = 0; i < input_ids.size(); i++) {
+            embed_selector.getByIndex(input_ids[i], out_embed.data() + i * _attr.tokens_embed_size);
+        }
+        ALOGI("img_embed.size:%d, %d", img_embed.size(), img_embed[0].size());
+
+        if (offsets.size() == 1 && img_embed.size() > 1) {
+            for (int i = 1; i < img_embed.size(); i++) {
+                offsets.push_back(offsets[i - 1] + img_embed[i - 1].size() / _attr.tokens_embed_size);
+                ALOGI("offset:%d", offsets[i - 1] + img_embed[i - 1].size() / _attr.tokens_embed_size);
+            }
+        }
+
+        for (int i = 0; i < img_embed.size(); i++) {
+            memcpy(out_embed.data() + offsets[i] * _attr.tokens_embed_size, img_embed[i].data(),
+                   img_embed[i].size() * sizeof(unsigned short));
+        }
+
+        ALOGI("out_embed size:%d", out_embed.size());
+        ALOGI("input_ids size %d", input_ids.size());
+        GetPositionIds(input_ids, position_ids, cfg);
+        ALOGI("position_ids size:%d", position_ids[0].size());
+        return 0;
+    }
+
+    std::string Run(std::vector<unsigned short> &test_embed, std::vector<std::vector<int>> &position_ids,
+                    std::vector<std::vector<float>> &deepstack_features, std::vector<int> &visual_pos_mask)
+    {
+        b_stop = false;
+        std::string final_out;
+
+        bfloat16 bf16 = -65536.f;
+        std::vector<unsigned short> mask(_attr.kv_cache_num + 1, bf16.data);
+        std::vector<unsigned short> embed(_attr.tokens_embed_size, 0);
+
+        std::vector<int> cached_token;
+        std::vector<int> token_ids;
+
+        int input_embed_num   = test_embed.size() / _attr.tokens_embed_size;
+        int prefill_split_num = ceil((double)input_embed_num / _attr.prefill_token_num);
+        ALOGI("input token num : %d, prefill_split_num : %d", input_embed_num, prefill_split_num);
+        if (input_embed_num > _attr.prefill_max_token_num) {
+            ALOGE("input token num(%d) > prefill_max_token_num(%d)", input_embed_num, _attr.prefill_max_token_num);
+            return "";
+        }
+
+        int kv_cache_num;
+        mask[_attr.kv_cache_num] = 0;
+        for (size_t i = 0; i < input_embed_num; i++) {
+            mask[i] = 0;
+        }
+        timer t_cost;
+        timer ttft_timer;
+        ttft_timer.start();
+
+        int max_pos_id = 0;
+        for (size_t p = 0; p < prefill_split_num; p++) {
+            if (b_stop) {
+                break;
+            }
+            _attr.prefill_grpid = p + 1;
+            kv_cache_num        = p * _attr.prefill_token_num;
+            std::vector<unsigned short> mask_tmp;
+            mask_tmp.resize(1 * _attr.prefill_token_num * (kv_cache_num + _attr.prefill_token_num), bf16.data);
+            int input_num_token = _attr.prefill_token_num;
+            if (p == prefill_split_num - 1) {
+                input_num_token = input_embed_num - p * _attr.prefill_token_num;
+            }
+
+            ALOGI("input_num_token:%d", input_num_token);
+            for (size_t i = 0; i < _attr.prefill_token_num; i++) {
+                if (i < input_num_token) {
+                    int mask_current_start = kv_cache_num;
+                    auto mask_ptr          = mask_tmp.data() + i * (kv_cache_num + _attr.prefill_token_num);
+
+                    for (int j = 0; j < _attr.precompute_len + p * _attr.prefill_token_num; j++) {
+                        mask_ptr[j] = 0;
+                    }
+
+                    for (int j = mask_current_start; j < mask_current_start + i + 1; j++) {
+                        mask_ptr[j] = 0;
+                    }
+                }
+            }
+
+            std::vector<unsigned short> embed_tmp(_attr.prefill_token_num * _attr.tokens_embed_size, 0);
+            int start, offset;
+
+            start = p * _attr.prefill_token_num;
+            if (p == (prefill_split_num - 1)) {
+                offset = (input_embed_num - p * _attr.prefill_token_num);
+                memcpy(embed_tmp.data(), test_embed.data() + start * _attr.tokens_embed_size,
+                       offset * _attr.tokens_embed_size * sizeof(unsigned short));
+            } else {
+                offset = _attr.prefill_token_num;
+                memcpy(embed_tmp.data(), test_embed.data() + start * _attr.tokens_embed_size,
+                       offset * _attr.tokens_embed_size * sizeof(unsigned short));
+            }
+
+            int start_deepstack_feat = 0, offset_deepstack_feat = 0;
+            if (!visual_pos_mask.empty()) {
+                for (int j = 0; j < start; j++) {
+                    start_deepstack_feat += visual_pos_mask[j];
+                }
+                for (int j = start; j < start + offset; j++) {
+                    offset_deepstack_feat += visual_pos_mask[j];
+                }
+            }
+
+            for (unsigned int m = 0; m < _attr.axmodel_num; m++) {
+                if (b_stop) {
+                    break;
+                }
+
+                auto &layer = llama_layers[m];
+
+                if (_attr.b_dynamic_load_axmodel_layer) {
+                    int ret;
+                    if (_attr.b_use_mmap_load_layer) {
+                        ret = layer.layer.init((char *)layer.layer_buffer.data(), layer.layer_buffer.size());
+                    } else {
+                        ret = layer.layer.init(layer.layer_buffer_vec.data(), layer.layer_buffer_vec.size());
+                    }
+                    if (ret != 0) {
+                        ALOGE("init axmodel(%s) failed", layer.filename.c_str());
+                    }
+                }
+
+                // set indices
+                auto &input_indices             = layer.layer.get_input(_attr.prefill_grpid, "indices");
+                unsigned int *input_indices_ptr = (unsigned int *)input_indices.pVirAddr;
+                memset(input_indices_ptr, 0, input_indices.nSize);
+                for (unsigned int i = 0; i < position_ids.size(); i++) {
+                    for (unsigned int j = _attr.precompute_len + p * _attr.prefill_token_num, jj = 0;
+                         j < _attr.precompute_len + (p + 1) * _attr.prefill_token_num; j++, jj++) {
+                        if (j < position_ids[i].size()) {
+                            input_indices_ptr[i * _attr.prefill_token_num + jj] = position_ids[i][j];
+                            if (position_ids[i][j] > max_pos_id) {
+                                max_pos_id = position_ids[i][j];
+                            }
+                        }
+                    }
+                }
+
+                // set mask
+                auto &input_mask = layer.layer.get_input(_attr.prefill_grpid, "mask");
+                memcpy((void *)input_mask.pVirAddr, (void *)mask_tmp.data(), mask_tmp.size() * sizeof(unsigned short));
+                // set input
+                auto &input_input = layer.layer.get_input(_attr.prefill_grpid, "input");
+                memcpy((void *)input_input.pVirAddr, embed_tmp.data(), embed_tmp.size() * sizeof(unsigned short));
+
+                layer.layer.inference(_attr.prefill_grpid);
+
+                auto &input_decoder_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
+                auto &input_decoder_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
+
+                auto &output_k_cache = layer.layer.get_output(_attr.prefill_grpid, "K_cache_out");
+                auto &output_v_cache = layer.layer.get_output(_attr.prefill_grpid, "V_cache_out");
+
+                int kv_offset = (_attr.precompute_len + p * _attr.prefill_token_num) * _attr.kv_cache_size;
+
+                memcpy((unsigned short *)input_decoder_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
+                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+
+                memcpy((unsigned short *)input_decoder_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
+                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+
+                for (int gid = _attr.prefill_grpid + 1; gid < prefill_split_num + 1; gid++) {
+                    auto &input_prefill_k_cache = layer.layer.get_input(gid, "K_cache");
+
+                    memcpy((unsigned short *)input_prefill_k_cache.pVirAddr + kv_offset,
+                           (void *)output_k_cache.pVirAddr,
+                           sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+                }
+
+                for (int gid = _attr.prefill_grpid + 1; gid < prefill_split_num + 1; gid++) {
+                    auto &input_prefill_v_cache = layer.layer.get_input(gid, "V_cache");
+
+                    memcpy((unsigned short *)input_prefill_v_cache.pVirAddr + kv_offset,
+                           (void *)output_v_cache.pVirAddr,
+                           sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+                }
+
+                auto &output = layer.layer.get_output(_attr.prefill_grpid, "output");
+
+                memcpy(embed_tmp.data(), (void *)output.pVirAddr, embed_tmp.size() * sizeof(unsigned short));
+
+                if (!visual_pos_mask.empty() && m < deepstack_features.size()) {
+                    int k = 0;
+                    for (int j = start, k = start_deepstack_feat; j < start + offset; j++) {
+                        if (visual_pos_mask[j] == 0) {
+                            continue;
+                        }
+                        for (int di = 0; di < _attr.tokens_embed_size; di++) {
+                            // bfloat16 to float32
+                            unsigned int tmp_bf16_1 = embed_tmp[(j - start) * _attr.tokens_embed_size + di] << 16;
+                            float tmp_fp32_1        = *reinterpret_cast<float *>(&tmp_bf16_1);
+
+                            float tmp_fp32_2 = deepstack_features[m][k * _attr.tokens_embed_size + di];
+                            // float32 to bfloat16
+                            embed_tmp[(j - start) * _attr.tokens_embed_size + di] =
+                                bfloat16(tmp_fp32_1 + tmp_fp32_2).data;
+                        }
+                        k++;
+                    }
+                }
+
+                if (_attr.b_dynamic_load_axmodel_layer) {
+                    layer.layer.deinit();
+                }
+            }
+            if (p == (prefill_split_num - 1)) {
+                memcpy(embed.data(),
+                       embed_tmp.data() + (input_embed_num - p * _attr.prefill_token_num - 1) * _attr.tokens_embed_size,
+                       _attr.tokens_embed_size * sizeof(unsigned short));
+            }
+        }
+
+        int next_token = -1;
+        t_cqdm cqdm    = create_cqdm(_attr.max_token_len, 32);
+
+        {
+            auto &input = llama_post.get_input(0);
+            memcpy((void *)input.pVirAddr, embed.data(), embed.size() * sizeof(unsigned short));
+            llama_post.inference();
+
+            int max_index;
+
+            auto &output_post = llama_post.get_output(0);
+            memcpy(output_post.pVirAddr, (void *)output_post.pVirAddr, output_post.nSize);
+            unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
+            float max_val            = -MAXFLOAT;
+            max_index                = post_process(postprocess, post_out, _attr.tokens_embed_num, token_ids, &max_val);
+
+            next_token = max_index;
+
+            token_ids.push_back(max_index);
+            cached_token.push_back(max_index);
+            ALOGI("ttft: %.2f ms", ttft_timer.cost());
+        }
+        t_cost.start();
+
+        bool b_hit_eos = false;
+
+        for (unsigned int indices = max_pos_id + 1; indices < _attr.max_token_len; indices++) {
+            if (b_stop) {
+                break;
+            }
+
+            embed_selector.getByIndex(next_token, embed);
+
+            memcpy((void *)llama_layers[0].layer.get_input(decode_grpid, "input").pVirAddr, embed.data(),
+                   llama_layers[0].layer.get_input(decode_grpid, "input").nSize);
+
+            for (int m = 0; m < _attr.axmodel_num; m++) {
+                if (b_stop) {
+                    break;
+                }
+
+                auto &layer = llama_layers[m];
+
+                if (_attr.b_dynamic_load_axmodel_layer) {
+                    int ret;
+                    if (_attr.b_use_mmap_load_layer) {
+                        ret = layer.layer.init((char *)layer.layer_buffer.data(), layer.layer_buffer.size());
+                    } else {
+                        ret = layer.layer.init(layer.layer_buffer_vec.data(), layer.layer_buffer_vec.size());
+                    }
+                    if (ret != 0) {
+                        ALOGE("init axmodel(%s) failed", layer.filename.c_str());
+                    }
+                }
+
+                auto &input_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
+                auto &input_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
+
+                auto &input_indices = layer.layer.get_input(decode_grpid, "indices");
+                memcpy((void *)input_indices.pVirAddr, &indices, sizeof(indices));
+
+                auto &input_mask = layer.layer.get_input(decode_grpid, "mask");
+                memcpy((void *)input_mask.pVirAddr, mask.data(), mask.size() * sizeof(unsigned short));
+
+                layer.layer.inference(decode_grpid);
+
+                auto &output_k_cache = layer.layer.get_output(decode_grpid, "K_cache_out");
+                memcpy((unsigned short *)input_k_cache.pVirAddr + indices * _attr.kv_cache_size,
+                       (void *)output_k_cache.pVirAddr, output_k_cache.nSize);
+
+                auto &output_v_cache = layer.layer.get_output(decode_grpid, "V_cache_out");
+                memcpy((unsigned short *)input_v_cache.pVirAddr + indices * _attr.kv_cache_size,
+                       (void *)output_v_cache.pVirAddr, output_v_cache.nSize);
+
+                if (m == _attr.axmodel_num - 1) {
+                    memcpy((void *)llama_post.get_input(0).pVirAddr,
+                           (void *)layer.layer.get_output(decode_grpid, "output").pVirAddr,
+                           llama_post.get_input(0).nSize);
+                } else if (m < _attr.axmodel_num - 1) {
+                    memcpy((void *)llama_layers[m + 1].layer.get_input(decode_grpid, "input").pVirAddr,
+                           (void *)layer.layer.get_output(decode_grpid, "output").pVirAddr,
+                           layer.layer.get_input(decode_grpid, "input").nSize);
+                }
+            }
+            mask[indices] = 0;
+            {
+                llama_post.inference();
+
+                auto &output_post = llama_post.get_output(0);
+                memcpy(output_post.pVirAddr, (void *)output_post.pVirAddr, output_post.nSize);
+                unsigned short *post_out = (unsigned short *)output_post.pVirAddr;
+                float max_val            = -MAXFLOAT;
+                auto max_index = post_process(postprocess, post_out, _attr.tokens_embed_num, token_ids, nullptr);
+
+                next_token = max_index;
+
+                if (tokenizer->isEnd(max_index)) {
+                    if (cached_token.size() && _attr.runing_callback) {
+                        float t_cost_ms     = t_cost.cost();
+                        float token_per_sec = token_ids.size() / (t_cost_ms / 1000);
+                        auto tmp_out        = tokenizer->Decode(cached_token);
+                        _attr.runing_callback(cached_token.data(), cached_token.size(), tmp_out.c_str(), token_per_sec,
+                                              _attr.reserve);
+                        cached_token.clear();
+                    }
+                    b_hit_eos = true;
+                    break;
+                }
+                token_ids.push_back(max_index);
+
+                if (_attr.runing_callback) {
+                    cached_token.push_back(max_index);
+                    if (cached_token.size() >= 3) {
+                        float t_cost_ms     = t_cost.cost();
+                        float token_per_sec = token_ids.size() / (t_cost_ms / 1000);
+                        auto tmp_out        = tokenizer->Decode(cached_token);
+                        _attr.runing_callback(cached_token.data(), cached_token.size(), tmp_out.c_str(), token_per_sec,
+                                              _attr.reserve);
+                        cached_token.clear();
+                    }
+                }
+            }
+
+            if (_attr.runing_callback == nullptr) update_cqdm(&cqdm, indices, "token", "");
+            if (b_hit_eos) {
+                break;
+            }
+        }
+        printf("\n\n");
+        fflush(stdout);
+        float t_cost_ms = t_cost.cost();
+        ALOGN("hit eos,avg %.2f token/s\n", token_ids.size() / (t_cost_ms / 1000));
+
+        final_out = tokenizer->Decode(token_ids);
+
+        for (size_t i = 0; i < _attr.axmodel_num; i++) {
+            for (size_t j = 0; j < llama_layers[i].layer.get_num_input_groups(); j++) {
+                memset((void *)llama_layers[i].layer.get_input(j, "K_cache").pVirAddr, 0,
+                       llama_layers[i].layer.get_input(j, "K_cache").nSize);
+                memset((void *)llama_layers[i].layer.get_input(j, "V_cache").pVirAddr, 0,
+                       llama_layers[i].layer.get_input(j, "V_cache").nSize);
+            }
+        }
+
         return final_out;
     }
 };
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp b/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp
index 38974cb6..918951bd 100644
--- a/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp
+++ b/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp
@@ -8,8 +8,7 @@
 #include "string_utility.hpp"
 #include "memory_utils.hpp"
 
-class Tokenizer_Http : public BaseTokenizer
-{
+class Tokenizer_Http : public BaseTokenizer {
     std::shared_ptr<httplib::Client> cli;
     bool _b_bos, _b_eos;
 
@@ -26,13 +25,10 @@ class Tokenizer_Http : public BaseTokenizer
     bool Init(std::string model_path) override
     {
         base_url = model_path;
-        if (!test_connect_http(base_url, 10))
-        {
+        if (!test_connect_http(base_url, 10)) {
             ALOGE("connect %s failed", base_url.c_str());
             return false;
-        }
-        else
-        {
+        } else {
             ALOGI("connect %s ok", base_url.c_str());
         }
 
@@ -42,25 +38,20 @@ class Tokenizer_Http : public BaseTokenizer
         cli->set_write_timeout(10);
 
         int try_count = 10;
-        int count = try_count;
-        while (count-- > 0)
-        {
-            try
-            {
+        int count     = try_count;
+        while (count-- > 0) {
+            try {
                 auto ret = cli->Get("/get_uid");
                 auto rep = ret.value();
-                if (rep.status != 200)
-                {
+                if (rep.status != 200) {
                     ALOGE("get uid failed, status: %d", rep.status);
                     return false;
                 }
                 nlohmann::json j = nlohmann::json::parse(rep.body);
-                uid = j["uid"];
+                uid              = j["uid"];
                 ALOGI("uid: %s", uid.c_str());
                 break;
-            }
-            catch (const std::exception &e)
-            {
+            } catch (const std::exception &e) {
                 std::cerr << e.what() << '\n';
             }
             std::this_thread::sleep_for(std::chrono::seconds(1));
@@ -68,23 +59,18 @@ class Tokenizer_Http : public BaseTokenizer
         }
 
         count = 10;
-        while (count-- > 0)
-        {
-            try
-            {
+        while (count-- > 0) {
+            try {
                 auto ret = cli->Get("/bos_id?uid=" + uid);
                 auto rep = ret.value();
-                if (rep.status != 200)
-                {
+                if (rep.status != 200) {
                     ALOGE("get bos_id failed, status: %d", rep.status);
                     return false;
                 }
                 nlohmann::json j = nlohmann::json::parse(rep.body);
-                bos_id = j["bos_id"];
+                bos_id           = j["bos_id"];
                 break;
-            }
-            catch (const std::exception &e)
-            {
+            } catch (const std::exception &e) {
                 std::cerr << e.what() << '\n';
             }
             std::this_thread::sleep_for(std::chrono::seconds(1));
@@ -92,23 +78,18 @@ class Tokenizer_Http : public BaseTokenizer
         }
 
         count = 10;
-        while (count-- > 0)
-        {
-            try
-            {
+        while (count-- > 0) {
+            try {
                 auto ret = cli->Get("/eos_id?uid=" + uid);
                 auto rep = ret.value();
-                if (rep.status != 200)
-                {
+                if (rep.status != 200) {
                     ALOGE("get eos_id failed, status: %d", rep.status);
                     return false;
                 }
                 nlohmann::json j = nlohmann::json::parse(rep.body);
-                eos_id = j["eos_id"];
+                eos_id           = j["eos_id"];
                 break;
-            }
-            catch (const std::exception &e)
-            {
+            } catch (const std::exception &e) {
                 std::cerr << e.what() << '\n';
             }
             std::this_thread::sleep_for(std::chrono::seconds(1));
@@ -116,28 +97,23 @@ class Tokenizer_Http : public BaseTokenizer
         }
 
         count = 10;
-        while (count-- > 0)
-        {
-            try
-            {
+        while (count-- > 0) {
+            try {
                 auto ret = cli->Get("/img_start_token?uid=" + uid);
                 if (!ret) {
                     ALOGE("get img_start_token failed, no response");
                     continue;
                 }
                 auto rep = ret.value();
-                if (rep.status != 200)
-                {
+                if (rep.status != 200) {
                     ALOGE("get img_start_token failed, status: %d", rep.status);
                     continue;
                 }
                 nlohmann::json j = nlohmann::json::parse(rep.body);
-                img_start_token = j["img_start_token"];
+                img_start_token  = j["img_start_token"];
                 ALOGI("img_start_token: %d", img_start_token);
                 break;
-            }
-            catch (const std::exception &e)
-            {
+            } catch (const std::exception &e) {
                 std::cerr << "Exception: " << e.what() << '\n';
             }
             std::this_thread::sleep_for(std::chrono::seconds(1));
@@ -145,28 +121,23 @@ class Tokenizer_Http : public BaseTokenizer
         }
 
         count = 10;
-        while (count-- > 0)
-        {
-            try
-            {
+        while (count-- > 0) {
+            try {
                 auto ret = cli->Get("/img_context_token?uid=" + uid);
                 if (!ret) {
                     ALOGE("get img_context_token failed, no response");
                     continue;
                 }
                 auto rep = ret.value();
-                if (rep.status != 200)
-                {
+                if (rep.status != 200) {
                     ALOGE("get img_context_token failed, status: %d", rep.status);
                     continue;
                 }
-                nlohmann::json j = nlohmann::json::parse(rep.body);
+                nlohmann::json j  = nlohmann::json::parse(rep.body);
                 img_context_token = j["img_context_token"];
                 ALOGI("img_context_token: %d", img_context_token);
                 break;
-            }
-            catch (const std::exception &e)
-            {
+            } catch (const std::exception &e) {
                 std::cerr << "Exception: " << e.what() << '\n';
             }
             std::this_thread::sleep_for(std::chrono::seconds(1));
@@ -182,8 +153,14 @@ class Tokenizer_Http : public BaseTokenizer
     bool Init(std::string model_path = "http://localhost:8080", bool b_bos = true, bool b_eos = false) override
     {
         base_url = model_path;
-        try
-        {
+        if (!test_connect_http(base_url, 10)) {
+            ALOGE("connect %s failed", base_url.c_str());
+            return false;
+        } else {
+            ALOGI("connect %s ok", base_url.c_str());
+        }
+
+        try {
             cli = std::make_shared<httplib::Client>(base_url);
             cli->set_connection_timeout(1);
             cli->set_read_timeout(1);
@@ -191,30 +168,26 @@ class Tokenizer_Http : public BaseTokenizer
             {
                 auto ret = cli->Get("/bos_id");
                 auto rep = ret.value();
-                if (rep.status != 200)
-                {
+                if (rep.status != 200) {
                     ALOGE("get bos_id failed, status: %d", rep.status);
                     return false;
                 }
                 nlohmann::json j = nlohmann::json::parse(rep.body);
-                bos_id = j["bos_id"];
+                bos_id           = j["bos_id"];
             }
 
             {
                 auto ret = cli->Get("/eos_id");
                 auto rep = ret.value();
-                if (rep.status != 200)
-                {
+                if (rep.status != 200) {
                     ALOGE("get eos_id failed, status: %d", rep.status);
                     return false;
                 }
                 nlohmann::json j = nlohmann::json::parse(rep.body);
-                eos_id = j["eos_id"];
+                eos_id           = j["eos_id"];
             }
             printf("bos_id: %d, eos_id: %d\n", bos_id, eos_id);
-        }
-        catch (const std::exception &e)
-        {
+        } catch (const std::exception &e) {
             std::cerr << e.what() << '\n';
             return false;
         }
@@ -227,18 +200,14 @@ class Tokenizer_Http : public BaseTokenizer
     bool Init_new(std::string model_path, bool b_bos, bool b_eos) override
     {
         base_url = model_path;
-        if (!test_connect_http(base_url, 10))
-        {
+        if (!test_connect_http(base_url, 10)) {
             ALOGE("connect %s failed", base_url.c_str());
             return false;
-        }
-        else
-        {
+        } else {
             ALOGI("connect %s ok", base_url.c_str());
         }
 
-        try
-        {
+        try {
             cli = std::make_shared<httplib::Client>(base_url);
             cli->set_connection_timeout(1);
             cli->set_read_timeout(1);
@@ -246,56 +215,50 @@ class Tokenizer_Http : public BaseTokenizer
             {
                 auto ret = cli->Get("/bos_id");
                 auto rep = ret.value();
-                if (rep.status != 200)
-                {
+                if (rep.status != 200) {
                     ALOGE("get bos_id failed, status: %d", rep.status);
                     return false;
                 }
                 nlohmann::json j = nlohmann::json::parse(rep.body);
-                bos_id = j["bos_id"];
+                bos_id           = j["bos_id"];
             }
 
             {
                 auto ret = cli->Get("/eos_id");
                 auto rep = ret.value();
-                if (rep.status != 200)
-                {
+                if (rep.status != 200) {
                     ALOGE("get eos_id failed, status: %d", rep.status);
                     return false;
                 }
                 nlohmann::json j = nlohmann::json::parse(rep.body);
-                eos_id = j["eos_id"];
+                eos_id           = j["eos_id"];
             }
             ALOGI("bos_id: %d, eos_id: %d", bos_id, eos_id);
 
             {
                 auto ret = cli->Get("/img_start_token");
                 auto rep = ret.value();
-                if (rep.status != 200)
-                {
+                if (rep.status != 200) {
                     ALOGE("get img_start_token failed, status: %d", rep.status);
                     return false;
                 }
                 nlohmann::json j = nlohmann::json::parse(rep.body);
-                img_start_token = j["img_start_token"];
+                img_start_token  = j["img_start_token"];
             }
             ALOGI("img_start_token: %d", img_start_token);
 
             {
                 auto ret = cli->Get("/img_context_token");
                 auto rep = ret.value();
-                if (rep.status != 200)
-                {
+                if (rep.status != 200) {
                     ALOGE("get img_context_token failed, status: %d", rep.status);
                     return false;
                 }
-                nlohmann::json j = nlohmann::json::parse(rep.body);
+                nlohmann::json j  = nlohmann::json::parse(rep.body);
                 img_context_token = j["img_context_token"];
             }
             ALOGI("img_context_token: %d", img_context_token);
-        }
-        catch (const std::exception &e)
-        {
+        } catch (const std::exception &e) {
             std::cerr << e.what() << '\n';
             return false;
         }
@@ -309,59 +272,53 @@ class Tokenizer_Http : public BaseTokenizer
     {
         nlohmann::json j;
         j["uid"] = uid;
-        if (!system_prompt.empty() and system_prompt != "")
-        {
+        if (!system_prompt.empty() and system_prompt != "") {
             j["system_prompt"] = system_prompt;
         }
 
         auto ret = cli->Post("/reset", j.dump(), "application/json");
         auto rep = ret.value();
-        if (rep.status != 200)
-        {
+        if (rep.status != 200) {
             ALOGE("reset failed, status: %d", rep.status);
             return false;
         }
-        nlohmann::json j_rep = nlohmann::json::parse(rep.body);
+        nlohmann::json j_rep        = nlohmann::json::parse(rep.body);
         std::vector<int> _token_ids = j_rep["token_ids"];
-        tokens = _token_ids;
+        tokens                      = _token_ids;
         return true;
     }
 
-    bool Encode(std::string input, std::string last_reply, std::vector<int> &tokens, std::vector<int> &tokens_diff, ImageInfo img_info) override
+    bool Encode(std::string input, std::string last_reply, std::vector<int> &tokens, std::vector<int> &tokens_diff,
+                ImageInfo img_info) override
     {
         nlohmann::json j;
-        j["uid"] = uid;
-        j["text"] = input;
+        j["uid"]        = uid;
+        j["text"]       = input;
         j["img_prompt"] = img_info.img_prompt;
-        j["imgsz"] = img_info.imgsz;
-        j["num_img"] = img_info.num_img;
-        if (!last_reply.empty() and last_reply != "")
-        {
+        j["imgsz"]      = img_info.imgsz;
+        j["num_img"]    = img_info.num_img;
+        if (!last_reply.empty() and last_reply != "") {
             j["last_reply"] = last_reply;
         }
         auto ret = cli->Post("/encode", j.dump(), "application/json");
         auto rep = ret.value();
-        if (rep.status != 200)
-        {
+        if (rep.status != 200) {
             ALOGE("encode failed, status: %d", rep.status);
             return false;
         }
         nlohmann::json j2;
-        try
-        {
+        try {
             j2 = nlohmann::json::parse(rep.body);
-        }
-        catch (const std::exception &e)
-        {
+        } catch (const std::exception &e) {
             ALOGE("json parse failed: %s", e.what());
             ALOGE("%s", rep.body.c_str());
             return false;
         }
 
-        std::vector<int> _token_ids = j2["token_ids"];
+        std::vector<int> _token_ids   = j2["token_ids"];
         std::vector<int> _tokens_diff = j2["diff"];
 
-        tokens = _token_ids;
+        tokens      = _token_ids;
         tokens_diff = _tokens_diff;
 
         return true;
@@ -370,38 +327,38 @@ class Tokenizer_Http : public BaseTokenizer
     bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) override
     {
         nlohmann::json j;
-        j["text"] = input;
-        j["img_prompt"] = img_info.img_prompt;
-        j["imgsz"] = img_info.imgsz;
-        j["num_img"] = img_info.num_img;
+        j["text"]          = input;
+        j["img_prompt"]    = img_info.img_prompt;
+        j["imgsz"]         = img_info.imgsz;
+        j["num_img"]       = img_info.num_img;
+        j["img_token_num"] = img_info.img_token_num;
+        if (img_info.type == ImgType::Image)
+            j["img_type"] = "image";
+        else if (img_info.type == ImgType::Video)
+            j["img_type"] = "video";
+
         auto ret = cli->Post("/encode", j.dump(), "application/json");
         auto rep = ret.value();
-        if (rep.status != 200)
-        {
+        if (rep.status != 200) {
             ALOGE("encode failed, status: %d", rep.status);
             return false;
         }
         nlohmann::json j2;
-        try
-        {
+        try {
             j2 = nlohmann::json::parse(rep.body);
-        }
-        catch (const std::exception &e)
-        {
+        } catch (const std::exception &e) {
             ALOGE("json parse failed: %s", e.what());
             ALOGE("%s", rep.body.c_str());
             return false;
         }
 
         std::vector<int> out = j2["token_ids"];
-        output = out;
+        output               = out;
         // output = sp->encode(input, 1024);
-        if (_b_bos)
-        {
+        if (_b_bos) {
             output.insert(output.begin(), bos_id);
         }
-        if (_b_eos)
-        {
+        if (_b_eos) {
             output.push_back(eos_id);
         }
 
@@ -415,7 +372,8 @@ class Tokenizer_Http : public BaseTokenizer
         return output;
     }
 
-    std::vector<int> Encode_ctx(std::string input, ImageInfo img_info, std::vector<int> &tokens_ids, std::vector<int> &tokens_diff) override
+    std::vector<int> Encode_ctx(std::string input, ImageInfo img_info, std::vector<int> &tokens_ids,
+                                std::vector<int> &tokens_diff) override
     {
         std::vector<int> output;
         Encode(input, "", output, tokens_diff, img_info);
@@ -424,30 +382,25 @@ class Tokenizer_Http : public BaseTokenizer
 
     std::string Decode(const std::vector<int> input) override
     {
-        int cnt = 2;
+        int cnt             = 2;
         std::string out_str = "";
-        while (cnt--)
-        {
+        while (cnt--) {
             nlohmann::json j;
             j["token_ids"] = input;
-            j["uid"] = uid;
-            auto ret = cli->Post("/decode", j.dump(), "application/json");
-            auto rep = ret.value();
-            if (rep.status != 200)
-            {
+            j["uid"]       = uid;
+            auto ret       = cli->Post("/decode", j.dump(), "application/json");
+            auto rep       = ret.value();
+            if (rep.status != 200) {
                 ALOGE("decode failed, status: %d, try again", rep.status);
                 ALOGE("%s", rep.body.c_str());
                 usleep(1000 * 1000);
                 continue;
             }
-            try
-            {
+            try {
                 nlohmann::json j2 = nlohmann::json::parse(rep.body);
-                out_str = j2["text"];
+                out_str           = j2["text"];
                 break;
-            }
-            catch (const std::exception &e)
-            {
+            } catch (const std::exception &e) {
                 ALOGE("json parse failed: %s, try again", e.what());
                 ALOGE("%s", rep.body.c_str());
                 usleep(1000 * 1000);
@@ -480,12 +433,11 @@ class Tokenizer_Http : public BaseTokenizer
 
 std::shared_ptr<BaseTokenizer> CreateTokenizer(TokenizerType type)
 {
-    switch (type)
-    {
-    case TKT_HTTP:
-        return std::make_shared<Tokenizer_Http>();
-    default:
-        ALOGE("unknown tokenizer type: %d", type);
-        return nullptr;
+    switch (type) {
+        case TKT_HTTP:
+            return std::make_shared<Tokenizer_Http>();
+        default:
+            ALOGE("unknown tokenizer type: %d", type);
+            return nullptr;
     }
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.hpp b/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.hpp
index e82bdfd8..df30e615 100644
--- a/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.hpp
@@ -3,40 +3,40 @@
 #include <vector>
 #include <memory>
 
-enum TokenizerType
-{
-    TKT_LLaMa,
-    TKT_Qwen,
-    TKT_HTTP,
-    TKT_Phi3,
-    TKT_END
-};
+enum TokenizerType { TKT_LLaMa, TKT_Qwen, TKT_HTTP, TKT_Phi3, TKT_END };
+
+enum class ImgType { None, Image, Video };
 
-struct ImageInfo
-{
-    int imgsz = 448;
-    int num_img = 1;
-    bool img_prompt = false;
+struct ImageInfo {
+    int imgsz         = 448;
+    int num_img       = 1;
+    bool img_prompt   = false;
+    int img_token_num = -1;
+    ImgType type      = ImgType::None;
 };
 
-class BaseTokenizer
-{
+class BaseTokenizer {
 public:
-    virtual bool Init(std::string model_path) = 0;
-    virtual bool Init(std::string model_path, bool b_bos, bool b_eos) = 0;
-    virtual bool Init_new(std::string model_path, bool b_bos, bool b_eos) = 0;
-    virtual bool Reset(std::string system_prompt, std::vector<int> &tokens) = 0;
-    virtual bool Encode(std::string input, std::string last_reply, std::vector<int> &tokens, std::vector<int> &tokens_diff, ImageInfo img_info) = 0;
+    virtual bool Init(std::string model_path)                                            = 0;
+    virtual bool Init(std::string model_path, bool b_bos, bool b_eos)                    = 0;
+    virtual bool Init_new(std::string model_path, bool b_bos, bool b_eos)                = 0;
+    virtual bool Reset(std::string system_prompt, std::vector<int> &tokens)              = 0;
+    virtual bool Encode(std::string input, std::string last_reply, std::vector<int> &tokens,
+                        std::vector<int> &tokens_diff, ImageInfo img_info)               = 0;
     virtual bool Encode(std::string input, std::vector<int> &output, ImageInfo img_info) = 0;
-    virtual std::vector<int> Encode(std::string input, ImageInfo img_info) = 0;
-    virtual std::vector<int> Encode_ctx(std::string input, ImageInfo img_info, std::vector<int> &tokens_ids, std::vector<int> &tokens_diff) = 0;
-    virtual std::string Decode(const std::vector<int> input) = 0;
-    virtual int GetBosID() = 0;
-    virtual int GetEosID() = 0;
-    virtual int GetImgStartID() = 0;
-    virtual int GetImgContextID() = 0;
+    virtual std::vector<int> Encode(std::string input, ImageInfo img_info)               = 0;
+    virtual std::vector<int> Encode_ctx(std::string input, ImageInfo img_info, std::vector<int> &tokens_ids,
+                                        std::vector<int> &tokens_diff)                   = 0;
+    virtual std::string Decode(const std::vector<int> input)                             = 0;
+    virtual int GetBosID()                                                               = 0;
+    virtual int GetEosID()                                                               = 0;
+    virtual int GetImgStartID()                                                          = 0;
+    virtual int GetImgContextID()                                                        = 0;
 
-    virtual bool isEnd(int id) { return id == GetEosID(); }
+    virtual bool isEnd(int id)
+    {
+        return id == GetEosID();
+    }
 };
 
 std::shared_ptr<BaseTokenizer> CreateTokenizer(TokenizerType type);
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/utils/files.cpp b/projects/llm_framework/main_vlm/src/runner/utils/files.cpp
new file mode 100644
index 00000000..0222cea4
--- /dev/null
+++ b/projects/llm_framework/main_vlm/src/runner/utils/files.cpp
@@ -0,0 +1,46 @@
+#include <sys/stat.h>  
+#include <dirent.h>
+#include <unistd.h>
+#include <string>
+#include <vector>
+#include <iostream>
+#include <algorithm>
+
+bool is_directory(const std::string& path) {
+    struct stat st;
+    if (stat(path.c_str(), &st) != 0) return false;  
+    return S_ISDIR(st.st_mode);  
+}
+
+bool is_file(const std::string& path) {
+    struct stat st;
+    if (stat(path.c_str(), &st) != 0) return false;
+    return S_ISREG(st.st_mode);  
+}
+
+std::vector<std::string> list_files(const std::string& directory) {
+    std::vector<std::string> files;
+    DIR* dir = opendir(directory.c_str());
+    if (!dir) {
+        std::cerr << "无法打开目录: " << directory << std::endl;
+        return files;
+    }
+
+    struct dirent* entry;
+    while ((entry = readdir(dir)) != nullptr) {
+        std::string name = entry->d_name;
+        if (name == "." || name == "..") continue;
+
+        // 拼接完整路径并检查是否为普通文件
+        std::string full_path = directory + "/" + name;
+        struct stat st;
+        if (stat(full_path.c_str(), &st) == 0 && S_ISREG(st.st_mode)) {
+            files.push_back(full_path);
+        }
+    }
+    closedir(dir);
+
+    // 按文件名升序排序
+    std::sort(files.begin(), files.end());
+    return files;
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/utils/files.hpp b/projects/llm_framework/main_vlm/src/runner/utils/files.hpp
new file mode 100644
index 00000000..79f70608
--- /dev/null
+++ b/projects/llm_framework/main_vlm/src/runner/utils/files.hpp
@@ -0,0 +1,17 @@
+#ifndef _FILES_H_
+#define _FILES_H_
+
+#include <sys/stat.h>  
+#include <dirent.h>
+#include <unistd.h>
+#include <string>
+#include <vector>
+#include <iostream>
+#include <algorithm>
+
+bool is_directory(const std::string& path);
+bool is_file(const std::string& path);
+std::vector<std::string> list_files(const std::string& directory);
+
+
+#endif
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/utils/image_processor.cpp b/projects/llm_framework/main_vlm/src/runner/utils/image_processor.cpp
new file mode 100644
index 00000000..c9efb096
--- /dev/null
+++ b/projects/llm_framework/main_vlm/src/runner/utils/image_processor.cpp
@@ -0,0 +1,162 @@
+#include <vector>
+#include <math.h>
+#include <opencv2/opencv.hpp>
+#include "files.hpp"
+#include "image_processor.hpp"
+#include <iostream>
+
+std::vector<cv::Mat> ReadImages(std::string path){
+    std::vector<cv::Mat> src;
+
+    if(is_file(path)){
+        cv::Mat img = cv::imread(path, cv::IMREAD_COLOR);
+        src.push_back(img);
+    }
+    else if(is_directory(path)){
+        auto paths = list_files(path);
+        
+        for(auto &p : paths){
+            std::cout<<p<<std::endl;
+            cv::Mat img = cv::imread(p, cv::IMREAD_COLOR);
+            src.push_back(img);
+        }
+    }
+    else{
+        std::cerr << "错误的路径: " << path << std::endl;
+    }
+
+    return src;
+}
+
+std::pair<int, int> SmartResize(int height, int width, int factor){
+    int h_bar = height/factor;
+    int w_bar = width/factor;
+
+    h_bar *= factor;
+    w_bar *= factor;
+    return {h_bar, w_bar};
+}
+
+void normalizeMeanStd(cv::Mat& image) {
+    // 确保输入图像是浮点类型（避免整数溢出）
+    cv::Mat floatImage;
+    image.convertTo(floatImage, CV_32F);  // 转换为32位浮点格式 <button class="citation-flag" data-index="1">
+
+    // 计算均值和标准差
+    cv::Scalar mean, stddev;
+    cv::meanStdDev(floatImage, mean, stddev);  // 计算均值和标准差 <button class="citation-flag" data-index="2">
+
+    // 避免除以零：如果标准差为0，设置为一个小值（如1e-6）
+    for (int i = 0; i < floatImage.channels(); ++i) {
+        if (stddev[i] < 1e-6) {
+            stddev[i] = 1e-6;
+        }
+    }
+
+    // 归一化：减去均值并除以标准差
+    floatImage -= mean;  // 减去均值 <button class="citation-flag" data-index="4">
+    floatImage /= stddev;  // 除以标准差 <button class="citation-flag" data-index="5">
+
+    // 将结果转换回原始数据类型（如8位无符号整数）
+    floatImage.convertTo(image, image.type());  // 转换回原始格式 <button class="citation-flag" data-index="6">
+}
+
+int Qwen2VideoProcessor( std::vector<cv::Mat>& src, std::vector<std::vector<unsigned char>>& output, 
+                            int tgt_h, int tgt_w,
+                            int temporal_patch_size, int merge_size, int patch_size){
+
+    if(src.empty()){
+        return 0;
+    }
+
+    int height = src[0].rows;
+    int width = src[0].cols;
+
+    // auto [tgt_h, tgt_w] = SmartResize(height, width, 28);
+
+    cv::Size size(tgt_w, tgt_h);
+    std::vector<cv::Mat> imgs_resized;
+    
+    for(auto& img: src){
+        cv::Mat img_rs;
+        if(img.cols!=tgt_w || img.rows!=tgt_h){
+            cv::resize(img, img_rs, size, 0, 0, cv::INTER_CUBIC);
+        }else{
+            img_rs = img;
+        }
+        
+        cv::cvtColor(img_rs, img_rs, cv::COLOR_BGR2RGB);
+        imgs_resized.push_back(img_rs);
+    }
+    
+    if(imgs_resized.empty()){
+        return 0;
+    }
+
+    if(imgs_resized.size()%2!=0){
+        imgs_resized.push_back(imgs_resized.back());
+    }
+
+    std::vector<unsigned char> patches;
+    patches.resize( imgs_resized.size()* tgt_w*tgt_h* 3);
+    for(size_t i=0; i<imgs_resized.size(); ++i){
+        memcpy(patches.data()+i*tgt_w*tgt_h*3, imgs_resized[i].data, tgt_w*tgt_h* 3);
+    }
+
+    int grid_t = imgs_resized.size() / temporal_patch_size;
+    int channel = imgs_resized[0].channels();
+    int grid_h = tgt_h/patch_size;
+    int grid_w = tgt_w/patch_size;
+
+    // channel = patches.shape[3]
+    // patches = patches.reshape(
+    //     grid_t,                     # 0
+    //     self.temporal_patch_size,   # 1
+    //     grid_h // self.merge_size,  # 2
+    //     self.merge_size,            # 3
+    //     self.patch_size,            # 4
+    //     grid_w // self.merge_size,  # 5
+    //     self.merge_size,            # 6
+    //     self.patch_size,            # 7
+    //     channel                     # 8
+    // )   
+    // patches = patches.transpose(0, 2, 5, 3, 6, 1, 4, 7, 8 )
+
+    for(size_t d0=0; d0<grid_t; d0++){
+        std::vector<unsigned char> out_t;
+        for(size_t d2=0; d2<grid_h/merge_size; d2++){
+            for(size_t d5=0; d5<grid_w/merge_size; d5++){
+                for(size_t d3=0; d3<merge_size; d3++ ){
+                    for(size_t d6=0; d6<merge_size; d6++){
+                        for(size_t d1=0; d1<temporal_patch_size; d1++){
+                            for(size_t d4=0; d4<patch_size; d4++){
+                                for(size_t d7=0; d7<patch_size; d7++){
+                                    for(size_t d8=0; d8<channel; d8++){
+                                        size_t idx = d0*temporal_patch_size*grid_h*patch_size*grid_w*patch_size*channel;
+                                        idx += d1*grid_h*patch_size*grid_w*patch_size*channel;
+                                        idx += d2*merge_size*patch_size*grid_w*patch_size*channel;
+                                        idx += d3*patch_size*grid_w*patch_size*channel;
+                                        idx += d4*grid_w*patch_size*channel;
+                                        idx += d5*merge_size*patch_size*channel;
+                                        idx += d6*patch_size*channel;
+                                        idx += d7*channel;
+                                        idx += d8;
+
+                                        out_t.push_back(patches[idx]);
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+        }
+        output.push_back(out_t);
+    }
+
+    // std::vector<size_t> ret={grid_t, grid_h*grid_w, temporal_patch_size*patch_size*patch_size, channel};
+    // return ret;
+    return 0;
+
+}
+
diff --git a/projects/llm_framework/main_vlm/src/runner/utils/image_processor.hpp b/projects/llm_framework/main_vlm/src/runner/utils/image_processor.hpp
new file mode 100644
index 00000000..0b345754
--- /dev/null
+++ b/projects/llm_framework/main_vlm/src/runner/utils/image_processor.hpp
@@ -0,0 +1,14 @@
+
+#ifndef _IMAGE_PROCESSOR_H_
+#define _IMAGE_PROCESSOR_H_
+
+#include <string>
+#include <vector>
+#include <opencv2/opencv.hpp>
+
+std::vector<cv::Mat> ReadImages(std::string path);
+int Qwen2VideoProcessor( std::vector<cv::Mat>& src, std::vector<std::vector<unsigned char>>& output,
+                        int tgt_h, int tgt_w,
+                        int temporal_patch_size=2, int merge_size=2, int patch_size=14);
+
+#endif
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/utils/mrope.cpp b/projects/llm_framework/main_vlm/src/runner/utils/mrope.cpp
new file mode 100644
index 00000000..5e98aff5
--- /dev/null
+++ b/projects/llm_framework/main_vlm/src/runner/utils/mrope.cpp
@@ -0,0 +1,272 @@
+#include <vector>
+#include <algorithm>
+#include <optional>
+#include <cassert>
+#include "mrope.hpp"
+
+#include <iostream>
+#include <vector>
+#include <numeric>  // std::iota
+
+#include <vector>
+#include <algorithm>
+#include <limits>  // 用于std::numeric_limits
+#include <stdexcept>  // 用于异常处理
+
+int findMaxIn2DVector(const std::vector<std::vector<int>>& vec) {
+    if (vec.empty()) {
+        throw std::invalid_argument("输入二维vector为空");
+    }
+
+    int max_value = std::numeric_limits<int>::min();  // 初始化为最小值
+    bool has_elements = false;
+
+    for (const auto& subvec : vec) {
+        if (!subvec.empty()) {
+            has_elements = true;
+            // 使用std::max_element获取子vector的最大值
+            int sub_max = *std::max_element(subvec.begin(), subvec.end());
+            if (sub_max > max_value) {
+                max_value = sub_max;
+            }
+        }
+    }
+
+    if (!has_elements) {
+        throw std::invalid_argument("所有子vector均为空");  // 处理全空子vector
+    }
+
+    return max_value;
+}
+
+// 生成范围序列 [0, text_len-1]
+std::vector<int> generateRange(int text_len, int start) {
+    std::vector<int> range(text_len);
+    std::iota(range.begin(), range.end(), start);  // 填充从0开始的序列
+    return range;
+}
+
+// 扩展为多行矩阵
+std::vector<std::vector<int>> expandToMatrix(const std::vector<int>& range, int rows) {
+    std::vector<std::vector<int>> matrix(rows, range);  // 每一行都是range的副本
+    return matrix;
+}
+
+// 生成多维索引 (unused in Qwen3, but kept for completeness)
+std::vector<std::vector<int>> generateIndices(int grid_t, int grid_h, int grid_w) {
+    std::vector<std::vector<int>> indices(3, std::vector<int>(grid_t * grid_h * grid_w));
+
+    int idx = 0;
+    for (int t = 0; t < grid_t; ++t) {
+        for (int h = 0; h < grid_h; ++h) {
+            for (int w = 0; w < grid_w; ++w) {
+                indices[0][idx] = t;  // 时间索引
+                indices[1][idx] = h;  // 高度索引
+                indices[2][idx] = w;  // 宽度索引
+                ++idx;
+            }
+        }
+    }
+
+    return indices;
+}
+
+// Qwen3-VL specific: Preprocess video_grid_thw by repeating each row t times and setting t=1
+std::vector<std::vector<int>> preprocessVideoGrid(const std::vector<std::vector<int>>& video_grid_thw) {
+    std::vector<std::vector<int>> processed;
+    for (const auto& grid : video_grid_thw) {
+        if (grid.size() != 3) {
+            throw std::invalid_argument("Invalid grid format");
+        }
+        int t = grid[0];
+        // Repeat the row t times
+        for (int i = 0; i < t; ++i) {
+            std::vector<int> repeated_grid = {1, grid[1], grid[2]};  // Set t=1
+            processed.push_back(repeated_grid);
+        }
+    }
+    return processed;
+}
+
+std::vector<std::vector<int>> get_rope_index(
+    const Config& config,
+    const std::vector<int>& input_ids,
+    const std::vector<std::vector<int>>& image_grid_thw,
+    const std::vector<std::vector<int>>& video_grid_thw
+) {
+    const int spatial_merge_size = config.vision_config.spatial_merge_size;
+    const int image_token_id = config.image_token_id;
+    const int video_token_id = config.video_token_id;
+    const int vision_start_token_id = config.vision_start_token_id;
+
+    std::vector<std::vector<int>> position_ids(3);
+
+    // Preprocess video_grid_thw for Qwen3-VL (split into single-frame segments with timestamps)
+    auto processed_video_grid = preprocessVideoGrid(video_grid_thw);
+
+    // Handle pure text case
+    if (input_ids.empty() || (image_grid_thw.empty() && video_grid_thw.empty())) {
+        int b = 0;
+        for (int i = 0; i < 3; ++i) {
+            std::vector<int> seq(input_ids.size());
+            // 手动实现递增序列
+            for (size_t j = 0; j < seq.size(); ++j) {
+                seq[j] = j;
+            }
+            position_ids[i].insert(position_ids[i].end(), seq.begin(), seq.end());
+        }
+        return position_ids;
+    }
+
+    // Multimodal case (batch_size=1, single sequence)
+    const auto& ids = input_ids;
+    // Assume full mask for simplicity (as in Qwen3 fallback)
+    const auto mask = std::vector<int>(ids.size(), 1);
+
+    // Filter valid tokens (masked)
+    std::vector<int> filtered_ids;
+    for (size_t i = 0; i < ids.size(); ++i) {
+        if (mask[i]) filtered_ids.push_back(ids[i]);
+    }
+
+    int image_nums = 0, video_nums = 0;
+
+    for (size_t i = 0; i < filtered_ids.size() - 1; ++i) {
+        if (filtered_ids[i] == vision_start_token_id) {
+            if (filtered_ids[i + 1] == config.image_token_id) {
+                image_nums++;
+            }
+            if (filtered_ids[i + 1] == config.video_token_id) {
+                video_nums++;
+            }
+        }
+    }
+
+    int image_index = 0, video_index = 0;
+    std::vector<std::vector<int>> batch_pos(3);
+    int st = 0;
+    int remain_images = image_nums;
+    int remain_videos = video_nums;
+    std::vector<std::vector<std::vector<int>>> llm_pos_ids_list;
+
+    // Loop over vision blocks (images + videos, now with processed_video_grid)
+    for (size_t i_ = 0; i_ < static_cast<size_t>(image_nums + video_nums); ++i_) {
+
+        int ed_image = filtered_ids.size() + 1;
+        int ed_video = filtered_ids.size() + 1;
+
+        if (remain_images > 0) {
+            for (size_t j = st; j < filtered_ids.size(); ++j) {
+                if (filtered_ids[j] == config.image_token_id) {
+                    ed_image = static_cast<int>(j);
+                    break;
+                }
+            }
+        }
+
+        if (remain_videos > 0) {
+            for (size_t j = st; j < filtered_ids.size(); ++j) {
+                if (filtered_ids[j] == config.video_token_id) {
+                    ed_video = static_cast<int>(j);
+                    break;
+                }
+            }
+        }
+
+        int t, h, w;
+        int ed;
+
+        if (ed_image < ed_video) {
+            // Image
+            t = image_grid_thw[image_index][0];
+            h = image_grid_thw[image_index][1];
+            w = image_grid_thw[image_index][2];
+            image_index += 1;
+            remain_images -= 1;
+            ed = ed_image;
+        } else {
+            // Video (using processed grid, each entry has t=1)
+            t = processed_video_grid[video_index][0];  // 1
+            h = processed_video_grid[video_index][1];
+            w = processed_video_grid[video_index][2];
+            video_index += 1;
+            remain_videos -= 1;
+            ed = ed_video;
+        }
+
+        int llm_grid_t = t;  // For videos, t=1
+        int llm_grid_h = h / spatial_merge_size;
+        int llm_grid_w = w / spatial_merge_size;
+
+        int text_len = ed - st;
+
+        int st_idx;
+        if (llm_pos_ids_list.empty()) {
+            st_idx = 0;
+        } else {
+            st_idx = findMaxIn2DVector(llm_pos_ids_list.back()) + 1;
+        }
+        auto range = generateRange(text_len, st_idx);
+        auto expanded_matrix = expandToMatrix(range, 3);
+
+        llm_pos_ids_list.push_back(expanded_matrix);
+
+        // For Qwen3: t_index always starts from 0 (no scaling, timestamps handle time)
+        std::vector<int> t_index;
+        for (int ti = 0; ti < llm_grid_t; ++ti) {  // llm_grid_t=1 for videos/images typically
+            for (int hw = 0; hw < llm_grid_h * llm_grid_w; ++hw) {
+                t_index.push_back(ti + text_len + st_idx);  // No second_per_grid_t scaling; ti is 0 for t=1
+            }
+        }
+
+        std::vector<int> h_index;
+        for (int ti = 0; ti < llm_grid_t; ++ti) {
+            for (int hi = 0; hi < llm_grid_h; ++hi) {
+                for (int wi = 0; wi < llm_grid_w; ++wi) {
+                    h_index.push_back(hi + text_len + st_idx);
+                }
+            }
+        }
+
+        std::vector<int> w_index;
+        for (int ti = 0; ti < llm_grid_t; ++ti) {
+            for (int hi = 0; hi < llm_grid_h; ++hi) {
+                for (int wi = 0; wi < llm_grid_w; ++wi) {
+                    w_index.push_back(wi + text_len + st_idx);
+                }
+            }
+        }
+
+        std::vector<std::vector<int>> thw_idx;
+        thw_idx.push_back(t_index);
+        thw_idx.push_back(h_index);
+        thw_idx.push_back(w_index);
+        llm_pos_ids_list.push_back(thw_idx);
+
+        st = ed + llm_grid_t * llm_grid_h * llm_grid_w;
+
+        // Append remaining text if any
+        if (st < static_cast<int>(filtered_ids.size())) {
+            if (llm_pos_ids_list.empty()) {
+                st_idx = 0;
+            } else {
+                st_idx = findMaxIn2DVector(llm_pos_ids_list.back()) + 1;
+            }
+
+            text_len = static_cast<int>(filtered_ids.size()) - st;
+
+            range = generateRange(text_len, st_idx);
+            expanded_matrix = expandToMatrix(range, 3);
+            llm_pos_ids_list.push_back(expanded_matrix);
+        }
+    }
+
+    // Concatenate all position lists
+    for (const auto& item : llm_pos_ids_list) {
+        for (size_t pi = 0; pi < position_ids.size(); ++pi) {
+            position_ids[pi].insert(position_ids[pi].end(), item[pi].begin(), item[pi].end());
+        }
+    }
+
+    return position_ids;
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/utils/mrope.hpp b/projects/llm_framework/main_vlm/src/runner/utils/mrope.hpp
new file mode 100644
index 00000000..988effa6
--- /dev/null
+++ b/projects/llm_framework/main_vlm/src/runner/utils/mrope.hpp
@@ -0,0 +1,33 @@
+#ifndef MROPE_QWEN3_H
+#define MROPE_QWEN3_H
+
+#include <vector>
+
+// Forward declaration of Config (assume defined in mrope.hpp or utils.hpp)
+struct Config {
+    struct VisionConfig {
+        int temporal_patch_size;
+        int tokens_per_second;
+        int spatial_merge_size;
+        int patch_size;
+        int width;
+        int height;
+        int fps;
+    };
+    VisionConfig vision_config;
+    int image_token_id;
+    int video_token_id;
+    int vision_start_token_id;
+
+    std::vector<std::vector<int>> image_grid_thw;   // auto calc
+    std::vector<std::vector<int>> video_grid_thw;   // auto calc
+};
+
+std::vector<std::vector<int>> get_rope_index(
+    const Config& config,
+    const std::vector<int>& input_ids,
+    const std::vector<std::vector<int>>& image_grid_thw,
+    const std::vector<std::vector<int>>& video_grid_thw
+);
+
+#endif // MROPE_QWEN3_H
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 478c9198..5c864c0b 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -471,6 +471,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-model-qwen2.5-1.5B-Int4-ax650':[create_data_deb,'llm-model-qwen2.5-1.5B-Int4-ax650', '0.4', src_folder, revision],
         'llm-model-qwen2.5-3B-Int4-ax650':[create_data_deb,'llm-model-qwen2.5-3B-Int4-ax650', '0.4', src_folder, revision],
         'llm-model-qwen2.5-7B-Int4-ax650':[create_data_deb,'llm-model-qwen2.5-7B-Int4-ax650', '0.4', src_folder, revision],
+        'llm-model-qwen3-vl-2B-Int4-ax650':[create_data_deb,'llm-model-qwen3-vl-2B-Int4-ax650', '0.5', src_folder, revision],
         # Llama model
         'llm-model-llama3.2-1B-prefill-ax630c':[create_data_deb,'llm-model-llama3.2-1B-prefill-ax630c', data_version, src_folder, revision],
         'llm-model-llama3.2-1B-p256-ax630c':[create_data_deb,'llm-model-llama3.2-1B-p256-ax630c', '0.4', src_folder, revision],

From 324f04de194d74b4e4966ca598712cf88649e5d6 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 21 Nov 2025 09:47:03 +0800
Subject: [PATCH 58/79] [update] update llm-vlm version & model config

---
 .../src/runner/Tokenizer/Tokenizer.cpp             |  2 +-
 .../main_llm/src/runner/Tokenizer/Tokenizer.cpp    | 14 ++++++++++----
 projects/llm_framework/main_vlm/SConstruct         |  7 +++++--
 .../models/mode_qwen3-vl-2B-Int4-ax650.json        |  2 +-
 projects/llm_framework/main_vlm/src/main.cpp       |  2 ++
 projects/llm_framework/main_vlm/src/runner/LLM.hpp |  4 +++-
 .../main_vlm/src/runner/Tokenizer/Tokenizer.cpp    |  6 +++---
 projects/llm_framework/tools/llm_pack.py           |  2 +-
 8 files changed, 26 insertions(+), 13 deletions(-)

diff --git a/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp
index f676efdd..5c34cb4c 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp
@@ -389,7 +389,7 @@ class Tokenizer_Http : public BaseTokenizer {
     bool Init(std::string model_path = "http://localhost:8080") override
     {
         base_url = model_path;
-        if (!test_connect_http(base_url, 10)) {
+        if (!test_connect_http(base_url, 20)) {
             ALOGE("connect %s failed", base_url.c_str());
             return false;
         } else {
diff --git a/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.cpp b/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.cpp
index f676efdd..0829ffaf 100644
--- a/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.cpp
+++ b/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.cpp
@@ -349,11 +349,17 @@ class Tokenizer_Http : public BaseTokenizer {
     bool Init(std::string model_path = "http://localhost:8080", bool b_bos = true, bool b_eos = false) override
     {
         base_url = model_path;
+        if (!test_connect_http(base_url, 20)) {
+            ALOGE("connect %s failed", base_url.c_str());
+            return false;
+        } else {
+            ALOGI("connect %s ok", base_url.c_str());
+        }
         try {
             cli = std::make_shared<httplib::Client>(base_url);
-            cli->set_connection_timeout(1);
-            cli->set_read_timeout(1);
-            cli->set_write_timeout(1);
+            cli->set_connection_timeout(10);
+            cli->set_read_timeout(10);
+            cli->set_write_timeout(10);
             {
                 auto ret = cli->Get("/bos_id");
                 auto rep = ret.value();
@@ -389,7 +395,7 @@ class Tokenizer_Http : public BaseTokenizer {
     bool Init(std::string model_path = "http://localhost:8080") override
     {
         base_url = model_path;
-        if (!test_connect_http(base_url, 10)) {
+        if (!test_connect_http(base_url, 20)) {
             ALOGE("connect %s failed", base_url.c_str());
             return false;
         } else {
diff --git a/projects/llm_framework/main_vlm/SConstruct b/projects/llm_framework/main_vlm/SConstruct
index 2bb327ba..19239976 100644
--- a/projects/llm_framework/main_vlm/SConstruct
+++ b/projects/llm_framework/main_vlm/SConstruct
@@ -17,7 +17,10 @@ LDFLAGS = []
 LINK_SEARCH_PATH = []
 STATIC_FILES = []
 
-python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-vlm-python-venv_v1.6.tar.gz", 'm5stack_llm-vlm-python-venv_v1.6.tar.gz')
+if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
+    python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-vlm-python-venv_v1.6.tar.gz", 'm5stack_llm-vlm-python-venv_v1.6.tar.gz')
+else:
+    python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-vlm-python-venv_v1.7.tar.gz", 'm5stack_llm-vlm-python-venv_v1.7.tar.gz')
 
 DEFINITIONS += ['-O2']
 DEFINITIONS += ['-std=c++17']
@@ -73,7 +76,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_vlm-1.9',
+env['COMPONENTS'].append({'target':'llm_vlm-1.10',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax650.json b/projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax650.json
index 77470565..a0f01d6c 100644
--- a/projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax650.json
+++ b/projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax650.json
@@ -46,7 +46,7 @@
         "video_token_id": 151656,
         "vision_start_token_id": 151652,
         "precompute_len": 0,
-        "cmm_size": 1919044,
+        "cmm_size": 3582336,
         "ext_scripts": [
             "tokenizer_qwen3-vl-2B-Int4-ax650.py"
         ]
diff --git a/projects/llm_framework/main_vlm/src/main.cpp b/projects/llm_framework/main_vlm/src/main.cpp
index 7397ff6b..d4694e69 100644
--- a/projects/llm_framework/main_vlm/src/main.cpp
+++ b/projects/llm_framework/main_vlm/src/main.cpp
@@ -480,6 +480,7 @@ class llm_task {
                     std::vector<std::vector<unsigned short>> all_embeds;
                     if (auto ret = lLaMa_ctx_->Encode(mats, all_embeds); ret != 0) {
                         ALOGE("lLaMaCtx.Encode failed");
+                        if (out_callback_) out_callback_("Encode failed", true);
                         return;
                     }
                     mats.clear();
@@ -487,6 +488,7 @@ class llm_task {
                             lLaMa_ctx_->Encode(all_embeds, prompt_data_, prompt_complete(msg), tokens_ids, tokens_diff);
                         ret != 0) {
                         ALOGE("lLaMaCtx.Encode failed");
+                        if (out_callback_) out_callback_("Encode failed", true);
                         return;
                     }
                     if (auto ret = lLaMa_ctx_->SetKVCache(k_caches, v_caches, precompute_len, tokens_diff.size());
diff --git a/projects/llm_framework/main_vlm/src/runner/LLM.hpp b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
index 6a3b90fe..bf56bf84 100644
--- a/projects/llm_framework/main_vlm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
@@ -3,6 +3,8 @@
 #include <algorithm>
 #include <cmath>
 #include <numeric>
+#include <atomic>
+
 #include "bfloat16.hpp"
 #include "image_processor.hpp"
 #include "mrope.hpp"
@@ -49,7 +51,7 @@ struct LLMAttrType {
     TokenizerType tokenizer_type         = TKT_LLaMa;
     std::string filename_tokenizer_model = "tokenizer.model";
     std::string url_tokenizer_model;
-    bool b_bos                        = true;
+    bool b_bos                        = false;
     bool b_eos                        = false;
     std::string filename_tokens_embed = "tinyllama.model.embed_tokens.weight.bfloat16.bin";
     int tokens_embed_num              = 32000;
diff --git a/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp b/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp
index 918951bd..54881cdb 100644
--- a/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp
+++ b/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp
@@ -25,7 +25,7 @@ class Tokenizer_Http : public BaseTokenizer {
     bool Init(std::string model_path) override
     {
         base_url = model_path;
-        if (!test_connect_http(base_url, 10)) {
+        if (!test_connect_http(base_url, 20)) {
             ALOGE("connect %s failed", base_url.c_str());
             return false;
         } else {
@@ -153,7 +153,7 @@ class Tokenizer_Http : public BaseTokenizer {
     bool Init(std::string model_path = "http://localhost:8080", bool b_bos = true, bool b_eos = false) override
     {
         base_url = model_path;
-        if (!test_connect_http(base_url, 10)) {
+        if (!test_connect_http(base_url, 30)) {
             ALOGE("connect %s failed", base_url.c_str());
             return false;
         } else {
@@ -200,7 +200,7 @@ class Tokenizer_Http : public BaseTokenizer {
     bool Init_new(std::string model_path, bool b_bos, bool b_eos) override
     {
         base_url = model_path;
-        if (!test_connect_http(base_url, 10)) {
+        if (!test_connect_http(base_url, 30)) {
             ALOGE("connect %s failed", base_url.c_str());
             return false;
         } else {
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 5c864c0b..1206dae3 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -381,7 +381,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-tts':[create_bin_deb,'llm-tts', '1.6', src_folder, revision],
         'llm-melotts':[create_bin_deb,'llm-melotts', '1.9', src_folder, revision],
         'llm-camera':[create_bin_deb,'llm-camera', '1.9', src_folder, revision, 'lib-llm'],
-        'llm-vlm':[create_bin_deb,'llm-vlm', '1.9', src_folder, revision],
+        'llm-vlm':[create_bin_deb,'llm-vlm', '1.10', src_folder, revision],
         'llm-yolo':[create_bin_deb,'llm-yolo', '1.9', src_folder, revision],
         'llm-skel':[create_bin_deb,'llm-skel', version, src_folder, revision],
         'llm-depth-anything':[create_bin_deb,'llm-depth-anything', '1.7', src_folder, revision],

From bd4c03e9a4d4b9ea7bfa5a42b4ebcdc789e7fc94 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 21 Nov 2025 14:55:09 +0800
Subject: [PATCH 59/79] [fix] Fix cosyvoice Deinit bug

---
 .../main_cosy_voice/src/main.cpp              | 44 ++++++++++---------
 .../main_cosy_voice/src/runner/Token2wav.hpp  |  2 +-
 projects/llm_framework/main_llm/src/main.cpp  | 17 +++++--
 .../llm_framework/main_openai_api/SConstruct  |  4 +-
 projects/llm_framework/tools/llm_pack.py      |  2 +-
 5 files changed, 41 insertions(+), 28 deletions(-)

diff --git a/projects/llm_framework/main_cosy_voice/src/main.cpp b/projects/llm_framework/main_cosy_voice/src/main.cpp
index a089275e..d4adcd3d 100644
--- a/projects/llm_framework/main_cosy_voice/src/main.cpp
+++ b/projects/llm_framework/main_cosy_voice/src/main.cpp
@@ -65,7 +65,6 @@ class llm_task {
     std::atomic<bool> g_llm_finished{false};
     std::atomic<bool> g_stop{false};
     TokenBuffer g_token_buffer;
-    Token2Wav lToken2Wav;
 
     std::vector<int> prompt_text_token;
     std::vector<unsigned short> prompt_text_embeds;
@@ -81,6 +80,7 @@ class llm_task {
     LLMAttrType mode_config_;
     Token2WavAttr infer_mode_config_;
     std::unique_ptr<LLM> lLaMa_;
+    std::unique_ptr<Token2Wav> lToken2Wav_;
     std::string model_;
     std::string response_format_;
     std::vector<std::string> inputs_;
@@ -312,14 +312,15 @@ class llm_task {
                 lLaMa_.reset();
                 return -2;
             }
-            if (!lToken2Wav.Init(infer_mode_config_)) {
+            lToken2Wav_ = std::make_unique<Token2Wav>();
+            if (!lToken2Wav_->Init(infer_mode_config_)) {
                 lLaMa_->Deinit();
                 lLaMa_.reset();
                 return -1;
             }
             lLaMa_->TextToken2Embeds(prompt_text_token, prompt_text_embeds);
             lLaMa_->SpeechToken2Embeds(prompt_speech_token, prompt_speech_embeds);
-            lToken2Wav.SpeechToken2Embeds(prompt_speech_token, prompt_speech_embeds_flow);
+            lToken2Wav_->SpeechToken2Embeds(prompt_speech_token, prompt_speech_embeds_flow);
 
         } catch (...) {
             SLOGE("config false");
@@ -358,7 +359,7 @@ class llm_task {
     {
         g_llm_finished = false;
         g_token_buffer.erase(g_token_buffer.begin(), g_token_buffer.end());
-        lToken2Wav.reset();
+        lToken2Wav_->clear();
     }
 
     void resample_audio(float *input_buffer, int input_length, float *output_buffer, int *output_length,
@@ -428,7 +429,7 @@ class llm_task {
                 }
             }
 
-            int prompt_token_len = prompt_speech_embeds_flow.size() / lToken2Wav._attr.flow_embed_size;
+            int prompt_token_len = prompt_speech_embeds_flow.size() / lToken2Wav_->_attr.flow_embed_size;
             if (prompt_token_len < 75) {
                 SLOGE("Error, prompt speech token len %d < 75", prompt_token_len);
                 if (llm_thread.joinable()) llm_thread.join();
@@ -450,28 +451,28 @@ class llm_task {
             int token_offset = 0;
             int i            = 0;
             while (true) {
-                this_token_hop_len = (token_offset == 0) ? lToken2Wav._attr.token_hop_len + promot_token_pad
-                                                         : lToken2Wav._attr.token_hop_len;
+                this_token_hop_len = (token_offset == 0) ? lToken2Wav_->_attr.token_hop_len + promot_token_pad
+                                                         : lToken2Wav_->_attr.token_hop_len;
                 std::unique_lock<std::mutex> lock(g_buffer_mutex);
                 g_buffer_cv.wait(lock, [&] {
                     return (g_token_buffer.size() - token_offset >=
-                            this_token_hop_len + lToken2Wav._attr.pre_lookahead_len) ||
+                            this_token_hop_len + lToken2Wav_->_attr.pre_lookahead_len) ||
                            g_llm_finished.load() || g_stop.load();
                 });
                 if (g_stop) {
                     lock.unlock();
                     break;
                 } else if (g_token_buffer.size() - token_offset >=
-                           this_token_hop_len + lToken2Wav._attr.pre_lookahead_len) {
+                           this_token_hop_len + lToken2Wav_->_attr.pre_lookahead_len) {
                     std::vector<SpeechToken> token;
-                    int start = token_offset - std::min(int(token_offset / lToken2Wav._attr.token_hop_len),
-                                                        lToken2Wav._attr.max_infer_chunk_num - 1) *
-                                                   lToken2Wav._attr.token_hop_len;
-                    int end = token_offset + this_token_hop_len + lToken2Wav._attr.pre_lookahead_len;
+                    int start = token_offset - std::min(int(token_offset / lToken2Wav_->_attr.token_hop_len),
+                                                        lToken2Wav_->_attr.max_infer_chunk_num - 1) *
+                                                   lToken2Wav_->_attr.token_hop_len;
+                    int end = token_offset + this_token_hop_len + lToken2Wav_->_attr.pre_lookahead_len;
                     token.insert(token.end(), g_token_buffer.begin() + start, g_token_buffer.begin() + end);
                     lock.unlock();
-                    auto speech = lToken2Wav.infer(token, prompt_speech_embeds_flow1, prompt_feat1, spk_embeds,
-                                                   token_offset, false);
+                    auto speech = lToken2Wav_->infer(token, prompt_speech_embeds_flow1, prompt_feat1, spk_embeds,
+                                                     token_offset, false);
                     token_offset += this_token_hop_len;
                     output.insert(output.end(), speech.begin(), speech.end());
                     double src_ratio =
@@ -507,12 +508,12 @@ class llm_task {
             }
 
             std::vector<SpeechToken> token;
-            int start = g_token_buffer.size() - std::min(int(g_token_buffer.size() / lToken2Wav._attr.token_hop_len),
-                                                         lToken2Wav._attr.max_infer_chunk_num - 1) *
-                                                    lToken2Wav._attr.token_hop_len;
+            int start = g_token_buffer.size() - std::min(int(g_token_buffer.size() / lToken2Wav_->_attr.token_hop_len),
+                                                         lToken2Wav_->_attr.max_infer_chunk_num - 1) *
+                                                    lToken2Wav_->_attr.token_hop_len;
             token.insert(token.end(), g_token_buffer.begin() + start, g_token_buffer.end());
-            auto speech = lToken2Wav.infer(token, prompt_speech_embeds_flow1, prompt_feat1, spk_embeds,
-                                           token_offset - start, true);
+            auto speech = lToken2Wav_->infer(token, prompt_speech_embeds_flow1, prompt_feat1, spk_embeds,
+                                             token_offset - start, true);
             output.insert(output.end(), speech.begin(), speech.end());
             double src_ratio =
                 static_cast<double>(mode_config_.audio_rate) / static_cast<double>(mode_config_.mode_rate);
@@ -662,6 +663,9 @@ class llm_task {
         if (lLaMa_) {
             lLaMa_->Deinit();
         }
+        if (lToken2Wav_) {
+            lToken2Wav_->Deinit();
+        }
     }
 };
 
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp b/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp
index 8fd5f387..0eb0450c 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp
@@ -464,7 +464,7 @@ class Token2Wav {
         // fade_in_mel_data is now modified in-place with the faded result.
     }
 
-    void reset()
+    void clear()
     {
         std::unordered_map<std::string, std::vector<float>>().swap(hift_cache_dict);
     }
diff --git a/projects/llm_framework/main_llm/src/main.cpp b/projects/llm_framework/main_llm/src/main.cpp
index a845299a..8ce63c1f 100644
--- a/projects/llm_framework/main_llm/src/main.cpp
+++ b/projects/llm_framework/main_llm/src/main.cpp
@@ -404,6 +404,7 @@ class llm_task {
     bool pause()
     {
         if (lLaMa_) lLaMa_->Stop();
+        if (lLaMa_ctx_) lLaMa_ctx_->Stop();
         return true;
     }
 
@@ -414,8 +415,10 @@ class llm_task {
             waitpid(tokenizer_pid_, nullptr, 0);
             tokenizer_pid_ = -1;
         }
-        lLaMa_->Deinit();
-        lLaMa_.reset();
+        if (lLaMa_) lLaMa_->Deinit();
+        if (lLaMa_) lLaMa_.reset();
+        if (lLaMa_ctx_) lLaMa_ctx_->Deinit();
+        if (lLaMa_ctx_) lLaMa_ctx_.reset();
         return true;
     }
 
@@ -447,6 +450,7 @@ class llm_task {
             std::string par;
             async_list_.put(par);
             if (lLaMa_) lLaMa_->Stop();
+            if (lLaMa_ctx_) lLaMa_ctx_->Stop();
             inference_run_->join();
             inference_run_.reset();
         }
@@ -462,6 +466,9 @@ class llm_task {
         if (lLaMa_) {
             lLaMa_->Deinit();
         }
+        if (lLaMa_ctx_) {
+            lLaMa_ctx_->Deinit();
+        }
     }
 };
 
@@ -514,7 +521,8 @@ class llm_llm : public StackFlow {
         if (!(llm_task_obj && llm_channel)) {
             return;
         }
-        llm_task_obj->lLaMa_->Stop();
+        if (llm_task_obj->lLaMa_) llm_task_obj->lLaMa_->Stop();
+        if (llm_task_obj->lLaMa_ctx_) llm_task_obj->lLaMa_ctx_->Stop();
     }
 
     void pause(const std::string &work_id, const std::string &object, const std::string &data) override
@@ -605,7 +613,8 @@ class llm_llm : public StackFlow {
         if (!(llm_task_obj && llm_channel)) {
             return;
         }
-        llm_task_obj->lLaMa_->Stop();
+        if (llm_task_obj->lLaMa_) llm_task_obj->lLaMa_->Stop();
+        if (llm_task_obj->lLaMa_ctx_) llm_task_obj->lLaMa_ctx_->Stop();
     }
 
     int setup(const std::string &work_id, const std::string &object, const std::string &data) override
diff --git a/projects/llm_framework/main_openai_api/SConstruct b/projects/llm_framework/main_openai_api/SConstruct
index 5259c6b7..0bde3482 100644
--- a/projects/llm_framework/main_openai_api/SConstruct
+++ b/projects/llm_framework/main_openai_api/SConstruct
@@ -18,7 +18,7 @@ LINK_SEARCH_PATH = []
 STATIC_FILES = []
 
 
-ModuleLLMOpenAIPluginPath = wget_github_commit('https://github.com/m5stack/ModuleLLM-OpenAI-Plugin.git', '5298be215735f5b1c21bc9225c38d7cb9c1933db', True)
+ModuleLLMOpenAIPluginPath = wget_github_commit('https://github.com/m5stack/ModuleLLM-OpenAI-Plugin.git', 'a8f54b0430c478896b45828f612d0d8b0a6f2fa1', True)
 python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-openai-api-python-venv_v1.6.tar.gz", 'm5stack_llm-openai-api-python-venv_v1.6.tar.gz')
 
 
@@ -52,7 +52,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_openai_api-1.8',
+env['COMPONENTS'].append({'target':'llm_openai_api-1.9',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 1206dae3..3c9721d9 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -387,7 +387,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-depth-anything':[create_bin_deb,'llm-depth-anything', '1.7', src_folder, revision],
         'llm-vad':[create_bin_deb,'llm-vad', '1.8', src_folder, revision],
         'llm-whisper':[create_bin_deb,'llm-whisper', '1.8', src_folder, revision],
-        'llm-openai-api':[create_bin_deb,'llm-openai-api', '1.8', src_folder, revision],
+        'llm-openai-api':[create_bin_deb,'llm-openai-api', '1.9', src_folder, revision],
         'llm-cosy-voice':[create_bin_deb,'llm-cosy-voice', '1.8', src_folder, revision],
         # keyword spotting Audio file
         'llm-model-audio-en-us':[create_data_deb,'llm-model-audio-en-us', data_version, src_folder, revision],

From 9e70ce86908aeb20949d47d29c13e53244c092e2 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 21 Nov 2025 17:35:59 +0800
Subject: [PATCH 60/79] [update] update llm-llm llm-cosyvoice version

---
 projects/llm_framework/main_cosy_voice/SConstruct | 2 +-
 projects/llm_framework/main_llm/SConstruct        | 2 +-
 projects/llm_framework/tools/llm_pack.py          | 4 ++--
 3 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/projects/llm_framework/main_cosy_voice/SConstruct b/projects/llm_framework/main_cosy_voice/SConstruct
index d1dd0e43..499edf79 100644
--- a/projects/llm_framework/main_cosy_voice/SConstruct
+++ b/projects/llm_framework/main_cosy_voice/SConstruct
@@ -68,7 +68,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_cosy_voice-1.8',
+env['COMPONENTS'].append({'target':'llm_cosy_voice-1.9',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_llm/SConstruct b/projects/llm_framework/main_llm/SConstruct
index 608ef386..d77a4fb7 100644
--- a/projects/llm_framework/main_llm/SConstruct
+++ b/projects/llm_framework/main_llm/SConstruct
@@ -66,7 +66,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_llm-1.10',
+env['COMPONENTS'].append({'target':'llm_llm-1.11',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 3c9721d9..f5614f22 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -377,7 +377,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-audio':[create_bin_deb,'llm-audio', '1.7', src_folder, revision],
         'llm-kws':[create_bin_deb,'llm-kws', '1.9', src_folder, revision],
         'llm-asr':[create_bin_deb,'llm-asr', '1.7', src_folder, revision],
-        'llm-llm':[create_bin_deb,'llm-llm', '1.10', src_folder, revision],
+        'llm-llm':[create_bin_deb,'llm-llm', '1.11', src_folder, revision],
         'llm-tts':[create_bin_deb,'llm-tts', '1.6', src_folder, revision],
         'llm-melotts':[create_bin_deb,'llm-melotts', '1.9', src_folder, revision],
         'llm-camera':[create_bin_deb,'llm-camera', '1.9', src_folder, revision, 'lib-llm'],
@@ -388,7 +388,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-vad':[create_bin_deb,'llm-vad', '1.8', src_folder, revision],
         'llm-whisper':[create_bin_deb,'llm-whisper', '1.8', src_folder, revision],
         'llm-openai-api':[create_bin_deb,'llm-openai-api', '1.9', src_folder, revision],
-        'llm-cosy-voice':[create_bin_deb,'llm-cosy-voice', '1.8', src_folder, revision],
+        'llm-cosy-voice':[create_bin_deb,'llm-cosy-voice', '1.9', src_folder, revision],
         # keyword spotting Audio file
         'llm-model-audio-en-us':[create_data_deb,'llm-model-audio-en-us', data_version, src_folder, revision],
         'llm-model-audio-zh-cn':[create_data_deb,'llm-model-audio-zh-cn', data_version, src_folder, revision],

From 87123285d3fca681f4ac6b1899716889568305ca Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Wed, 26 Nov 2025 19:00:01 +0800
Subject: [PATCH 61/79] [update] Add qwen3-vl-2B-Init4-ax630c model

---
 .../models/mode_qwen3-vl-2B-Int4-ax630c.json  |  54 ++++++
 .../tokenizer_qwen3-vl-2B-Int4-ax630c.py      | 162 ++++++++++++++++++
 .../src/runner/Tokenizer/Tokenizer.cpp        |  12 +-
 projects/llm_framework/tools/llm_pack.py      |   1 +
 4 files changed, 223 insertions(+), 6 deletions(-)
 create mode 100644 projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax630c.json
 create mode 100644 projects/llm_framework/main_vlm/scripts/tokenizer_qwen3-vl-2B-Int4-ax630c.py

diff --git a/projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax630c.json b/projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax630c.json
new file mode 100644
index 00000000..fce6c7c5
--- /dev/null
+++ b/projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax630c.json
@@ -0,0 +1,54 @@
+{
+    "mode": "qwen3-vl-2B-Int4-ax630c",
+    "type": "vlm",
+    "homepage": "https://huggingface.co/AXERA-TECH/Qwen3-VL-2B-Instruct",
+    "capabilities": [
+        "text_generation",
+        "chat"
+    ],
+    "input_type": [
+        "vlm.chat_completion",
+        "vlm.chat_completion.stream"
+    ],
+    "output_type": [
+        "vlm.utf-8",
+        "vlm.utf-8.stream"
+    ],
+    "mode_param": {
+        "tokenizer_type": 2,
+        "url_tokenizer_model": "http://localhost:8080",
+        "filename_tokens_embed": "model.embed_tokens.weight.bfloat16.bin",
+        "filename_post_axmodel": "qwen3_vl_text_post.axmodel",
+        "template_filename_axmodel": "qwen3_vl_text_p64_l%d_together.axmodel",
+        "filename_image_encoder_axmodel": "Qwen3-VL-2B-Instruct_vision_u8_384_ax630c.axmodel",
+        "enable_temperature": true,
+        "temperature": 0.7,
+        "enable_top_p_sampling": false,
+        "top_p": 0.9,
+        "enable_top_k_sampling": true,
+        "top_k": 40,
+        "enable_repetition_penalty": false,
+        "repetition_penalty": 1.1,
+        "penalty_window": 50,
+        "axmodel_num": 28,
+        "tokens_embed_num": 151936,
+        "tokens_embed_size": 2048,
+        "b_use_mmap_load_embed": true,
+        "b_video": false,
+        "vision_config.temporal_patch_size": 2,
+        "vision_config.tokens_per_second": 2,
+        "vision_config.spatial_merge_size": 2,
+        "vision_config.patch_size": 16,
+        "vision_config.height": 384,
+        "vision_config.width": 384,
+        "vision_config.fps": 1,
+        "image_token_id": 151655,
+        "video_token_id": 151656,
+        "vision_start_token_id": 151652,
+        "precompute_len": 0,
+        "cmm_size": 3582336,
+        "ext_scripts": [
+            "tokenizer_qwen3-vl-2B-Int4-ax630c.py"
+        ]
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/scripts/tokenizer_qwen3-vl-2B-Int4-ax630c.py b/projects/llm_framework/main_vlm/scripts/tokenizer_qwen3-vl-2B-Int4-ax630c.py
new file mode 100644
index 00000000..d9ac5098
--- /dev/null
+++ b/projects/llm_framework/main_vlm/scripts/tokenizer_qwen3-vl-2B-Int4-ax630c.py
@@ -0,0 +1,162 @@
+from transformers import AutoTokenizer, PreTrainedTokenizerFast
+from transformers.tokenization_utils_base import AddedToken
+from http.server import HTTPServer, BaseHTTPRequestHandler
+import json
+import argparse
+
+class Tokenizer_Http:
+    def __init__(self, model_id, system_content="You are a helpful assistant."):
+        self.tokenizer = AutoTokenizer.from_pretrained(
+            model_id,
+            trust_remote_code=True,
+            use_fast=False
+        )
+        self.token_ids_cache = []
+        self.system_content = system_content
+
+    def encode(self, content):
+        text = [
+            f'<|im_start|>system\n{self.system_content}<|im_end|>\n'
+            f'<|im_start|>user\n{content}<|im_end|>\n'
+            f'<|im_start|>assistant\n'
+        ]
+        input_ids = self.tokenizer(text)
+        return input_ids["input_ids"][0]
+
+    def encode_vpm_image(self, content="Describe this image.", num_img=1, img_token_num=256):
+        imgs_token = (
+            '<|vision_start|>'
+            + '<|image_pad|>' * img_token_num
+            + '<|vision_end|>'
+        )
+        imgs_token *= num_img
+        text = (
+            f'<|im_start|>system\n{self.system_content}<|im_end|>\n'
+            f'<|im_start|>user\n{imgs_token}{content}<|im_end|>\n'
+            f'<|im_start|>assistant\n'
+        )
+        text_inputs = self.tokenizer([text])
+        return text_inputs["input_ids"][0]
+
+    def encode_vpm_video(self, content="Describe this image.", num_img=1, img_token_num=256):
+        imgs_token = (
+            '<|vision_start|>'
+            + '<|video_pad|>' * img_token_num * num_img
+            + '<|vision_end|>'
+        )
+        text = (
+            f'<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n'
+            f'<|im_start|>user\n{imgs_token}{content}<|im_end|>\n'
+            f'<|im_start|>assistant\n'
+        )
+        text_inputs = self.tokenizer([text])
+        return text_inputs["input_ids"][0]
+    
+    def decode(self, token_ids):
+        self.token_ids_cache += token_ids
+        text = self.tokenizer.decode(self.token_ids_cache)
+        if "\ufffd" in text:
+            print("text 中包含非法字符")
+            return ""
+        else:
+            self.token_ids_cache.clear()
+            return text
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+
+    @property
+    def bos_token(self):
+        return self.tokenizer.bos_token
+
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+
+    @property
+    def img_start_token(self):
+        return self.tokenizer.encode("<|vision_start|>")[0]
+
+    @property
+    def img_context_token(self):
+        return self.tokenizer.encode("<|image_pad|>")[0]
+
+class Request(BaseHTTPRequestHandler):
+    timeout = 5
+    server_version = 'Apache'
+
+    def do_GET(self):
+        print(self.path)
+        self.send_response(200)
+        self.send_header("type", "get")
+        self.end_headers()
+        if self.path == '/bos_id':
+            bos_id = tokenizer.bos_id
+            msg = json.dumps({'bos_id': -1 if bos_id is None else bos_id})
+        elif self.path == '/eos_id':
+            eos_id = tokenizer.eos_id
+            msg = json.dumps({'eos_id': -1 if eos_id is None else eos_id})
+        elif self.path == '/img_start_token':
+            img_start_token = tokenizer.img_start_token
+            msg = json.dumps({'img_start_token': -1 if img_start_token is None else img_start_token})
+        elif self.path == '/img_context_token':
+            img_context_token = tokenizer.img_context_token
+            msg = json.dumps({'img_context_token': -1 if img_context_token is None else img_context_token})
+        else:
+            msg = 'error'
+        print(msg)
+        msg = str(msg).encode()
+        self.wfile.write(msg)
+
+    def do_POST(self):
+        data = self.rfile.read(int(self.headers['content-length']))
+        req = json.loads(data.decode())
+        if self.path == "/encode":
+            prompt = req['text']
+            b_img_prompt = req.get('img_prompt', False)
+            img_type = req.get('img_type', 'image')
+            if b_img_prompt:
+                if img_type == 'image':
+                    token_ids = tokenizer.encode_vpm_image(
+                        prompt,
+                        req.get("num_img", 1),
+                        req.get("img_token_num", 256)
+                    )
+                elif img_type == 'video':
+                    token_ids = tokenizer.encode_vpm_video(
+                        prompt,
+                        req.get("num_img", 1),
+                        req.get("img_token_num", 256)
+                    )
+                else:
+                    token_ids = tokenizer.encode(prompt)
+            else:
+                token_ids = tokenizer.encode(prompt)
+            msg = json.dumps({'token_ids': -1 if token_ids is None else token_ids})
+        elif self.path == "/decode":
+            token_ids = req['token_ids']
+            text = tokenizer.decode(token_ids)
+            msg = json.dumps({'text': "" if text is None else text})
+        else:
+            msg = 'error'
+        self.send_response(200)
+        self.end_headers()
+        self.wfile.write(str(msg).encode())
+
+if __name__ == "__main__":
+    args = argparse.ArgumentParser()
+    args.add_argument('--host', type=str, default='localhost')
+    args.add_argument('--port', type=int, default=8080)
+    args.add_argument('--model_id', type=str, default='tokenizer')
+    args.add_argument('--content', type=str, default='You are a helpful assistant.')
+    args = args.parse_args()
+    tokenizer = Tokenizer_Http(args.model_id, system_content=args.content)
+    host = (args.host, args.port)
+    print(f"http://{args.host}:{args.port}")
+    server = HTTPServer(host, Request)
+    server.serve_forever()
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp b/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp
index 54881cdb..8877a2b9 100644
--- a/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp
+++ b/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp
@@ -162,9 +162,9 @@ class Tokenizer_Http : public BaseTokenizer {
 
         try {
             cli = std::make_shared<httplib::Client>(base_url);
-            cli->set_connection_timeout(1);
-            cli->set_read_timeout(1);
-            cli->set_write_timeout(1);
+            cli->set_connection_timeout(10);
+            cli->set_read_timeout(10);
+            cli->set_write_timeout(10);
             {
                 auto ret = cli->Get("/bos_id");
                 auto rep = ret.value();
@@ -209,9 +209,9 @@ class Tokenizer_Http : public BaseTokenizer {
 
         try {
             cli = std::make_shared<httplib::Client>(base_url);
-            cli->set_connection_timeout(1);
-            cli->set_read_timeout(1);
-            cli->set_write_timeout(1);
+            cli->set_connection_timeout(10);
+            cli->set_read_timeout(10);
+            cli->set_write_timeout(10);
             {
                 auto ret = cli->Get("/bos_id");
                 auto rep = ret.value();
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index f5614f22..8229e549 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -465,6 +465,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-model-qwen2.5-1.5B-Int4-ax630c':[create_data_deb,'llm-model-qwen2.5-1.5B-Int4-ax630c', '0.4', src_folder, revision],
         'llm-model-qwen2.5-coder-0.5B-ax630c':[create_data_deb,'llm-model-qwen2.5-coder-0.5B-ax630c', data_version, src_folder, revision],
         'llm-model-qwen3-0.6B-ax630c':[create_data_deb,'llm-model-qwen3-0.6B-ax630c', '0.4', src_folder, revision],
+        'llm-model-qwen3-vl-2B-Int4-ax630c':[create_data_deb,'llm-model-qwen3-vl-2B-Int4-ax630c', '0.5', src_folder, revision],
         ## AX650
         'llm-model-qwen2.5-0.5B-Int4-ax650':[create_data_deb,'llm-model-qwen2.5-0.5B-Int4-ax650', '0.4', src_folder, revision],
         'llm-model-qwen2.5-HA-0.5B-ctx-ax650':[create_data_deb,'llm-model-qwen2.5-HA-0.5B-ctx-ax650', '0.5', src_folder, revision],

From cc1087facacaf16ba61e268e41041c45c80fa65a Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 27 Nov 2025 17:29:20 +0800
Subject: [PATCH 62/79] [update] fix postprocess Div zero bug, update
 llm-openai-api, update lllm-vlm

---
 .../main_llm/src/runner/LLMPostprocess.hpp    | 165 +++++++-----------
 .../llm_framework/main_openai_api/SConstruct  |   2 +-
 projects/llm_framework/main_vlm/SConstruct    |   2 +-
 .../models/mode_qwen3-vl-2B-Int4-ax630c.json  |   2 +-
 projects/llm_framework/main_vlm/src/main.cpp  |   2 +
 .../llm_framework/main_vlm/src/runner/LLM.hpp |   1 +
 .../main_vlm/src/runner/LLMPostprocess.hpp    | 165 +++++++-----------
 7 files changed, 126 insertions(+), 213 deletions(-)

diff --git a/projects/llm_framework/main_llm/src/runner/LLMPostprocess.hpp b/projects/llm_framework/main_llm/src/runner/LLMPostprocess.hpp
index b7d50156..08d6b4e7 100644
--- a/projects/llm_framework/main_llm/src/runner/LLMPostprocess.hpp
+++ b/projects/llm_framework/main_llm/src/runner/LLMPostprocess.hpp
@@ -9,66 +9,50 @@
 #include "utils/json.hpp"
 #include "utils/sample_log.h"
 
-class LLMPostprocess
-{
+class LLMPostprocess {
 private:
-    // 	控制随机性
     void apply_temperature(std::vector<float> &logits, float temperature)
     {
-        for (float &logit : logits)
-        {
+        if (temperature == 0.0f) temperature = 0.01f;
+        for (float &logit : logits) {
             logit /= temperature;
         }
     }
 
-    // 防止重复
     void apply_repetition_penalty(std::vector<float> &logits, const std::vector<int> &history, float penalty)
     {
-        for (int token : history)
-        {
-            if (token < logits.size())
-            {
+        for (int token : history) {
+            if (token < logits.size()) {
                 logits[token] = logits[token] < 0 ? logits[token] * penalty : logits[token] / penalty;
             }
         }
     }
 
-    void apply_repetition_penalty(std::vector<float> &logits,
-                                  const std::vector<int> &generated_tokens,
-                                  float repetition_penalty,
-                                  int penalty_window)
+    void apply_repetition_penalty(std::vector<float> &logits, const std::vector<int> &generated_tokens,
+                                  float repetition_penalty, int penalty_window)
     {
-        if (repetition_penalty == 1.0f || generated_tokens.empty())
-        {
-            return; // 如果 penalty = 1.0 或者没有生成 token，则不进行修改
+        if (repetition_penalty == 1.0f || generated_tokens.empty()) {
+            return;
         }
 
         int start_idx = std::max(0, (int)generated_tokens.size() - penalty_window);
         std::unordered_set<int> recent_tokens(generated_tokens.begin() + start_idx, generated_tokens.end());
 
-        for (int token : recent_tokens)
-        {
-            if (token < 0 || token >= logits.size())
-                continue;
+        for (int token : recent_tokens) {
+            if (token < 0 || token >= logits.size()) continue;
 
-            if (logits[token] > 0)
-            {
+            if (logits[token] > 0) {
                 logits[token] /= std::sqrt(repetition_penalty);
-            }
-            else
-            {
+            } else {
                 logits[token] *= std::sqrt(repetition_penalty);
             }
         }
     }
 
-    // 增强多样性
     void apply_diversity_penalty(std::vector<float> &logits, const std::vector<int> &common_phrases, float penalty)
     {
-        for (int token : common_phrases)
-        {
-            if (token < logits.size())
-            {
+        for (int token : common_phrases) {
+            if (token < logits.size()) {
                 logits[token] *= penalty;
             }
         }
@@ -79,46 +63,37 @@ class LLMPostprocess
     {
         std::vector<float> probs(logits.size());
         float max_logit = *std::max_element(logits.begin(), logits.end());
-        float sum = 0.0f;
+        float sum       = 0.0f;
 
-        for (size_t i = 0; i < logits.size(); ++i)
-        {
+        for (size_t i = 0; i < logits.size(); ++i) {
             probs[i] = std::exp(logits[i] - max_logit);
             sum += probs[i];
         }
 
-        for (float &p : probs)
-        {
+        for (float &p : probs) {
             p /= sum;
         }
 
         return probs;
     }
 
-    // 	动态裁剪低概率 token
     int faster_top_p_sampling(const std::vector<float> &logits, float top_p)
     {
-        // 计算softmax
         std::vector<float> probs = softmax(logits);
 
-        // 构建最大堆（概率和索引的配对）
         std::vector<std::pair<float, size_t>> prob_index;
         prob_index.reserve(logits.size());
-        for (size_t i = 0; i < logits.size(); ++i)
-        {
+        for (size_t i = 0; i < logits.size(); ++i) {
             prob_index.emplace_back(probs[i], i);
         }
-        auto cmp = [](const auto &a, const auto &b)
-        { return a.first < b.first; };
+        auto cmp = [](const auto &a, const auto &b) { return a.first < b.first; };
         std::make_heap(prob_index.begin(), prob_index.end(), cmp);
 
-        // 提取top-p元素
         std::vector<size_t> filtered_indices;
         std::vector<float> filtered_probs;
         float cumulative_prob = 0.0f;
 
-        while (!prob_index.empty() && cumulative_prob < top_p)
-        {
+        while (!prob_index.empty() && cumulative_prob < top_p) {
             std::pop_heap(prob_index.begin(), prob_index.end(), cmp);
             auto [prob, index] = prob_index.back();
             prob_index.pop_back();
@@ -127,15 +102,11 @@ class LLMPostprocess
             filtered_indices.push_back(index);
             filtered_probs.push_back(prob);
 
-            if (cumulative_prob >= top_p)
-                break;
+            if (cumulative_prob >= top_p) break;
         }
 
-        // 处理边缘情况（概率全零时返回第一个元素）
-        if (filtered_indices.empty())
-            return 0;
+        if (filtered_indices.empty()) return 0;
 
-        // 使用thread_local随机数生成器（线程安全）
         static thread_local std::mt19937 gen(std::random_device{}());
         std::discrete_distribution<int> dist(filtered_probs.begin(), filtered_probs.end());
         return filtered_indices[dist(gen)];
@@ -147,31 +118,26 @@ class LLMPostprocess
         // Sort indices by probability in descending order
         std::vector<size_t> indices(logits.size());
         std::iota(indices.begin(), indices.end(), 0);
-        std::sort(indices.begin(), indices.end(), [&](size_t i, size_t j)
-                  { return probs[i] > probs[j]; });
+        std::sort(indices.begin(), indices.end(), [&](size_t i, size_t j) { return probs[i] > probs[j]; });
 
         // Compute cumulative probabilities
         float cumulative_prob = 0.0f;
-        size_t cut_off = 0;
-        for (; cut_off < indices.size(); ++cut_off)
-        {
+        size_t cut_off        = 0;
+        for (; cut_off < indices.size(); ++cut_off) {
             cumulative_prob += probs[indices[cut_off]];
-            if (cumulative_prob >= top_p)
-                break;
+            if (cumulative_prob >= top_p) break;
         }
 
         // Keep only the top-p probabilities
         std::vector<size_t> filtered_indices(indices.begin(), indices.begin() + cut_off + 1);
         std::vector<float> filtered_probs(filtered_indices.size());
-        for (size_t i = 0; i < filtered_indices.size(); ++i)
-        {
+        for (size_t i = 0; i < filtered_indices.size(); ++i) {
             filtered_probs[i] = probs[filtered_indices[i]];
         }
 
         // Normalize the probabilities
         float filtered_sum = std::accumulate(filtered_probs.begin(), filtered_probs.end(), 0.0f);
-        for (float &p : filtered_probs)
-        {
+        for (float &p : filtered_probs) {
             p /= filtered_sum;
         }
 
@@ -182,34 +148,27 @@ class LLMPostprocess
         return filtered_indices[dist(gen)];
     }
 
-    // 限制候选 token 数
     int top_k_sampling(const std::vector<float> &logits, int k)
     {
         // std::vector<float> probs = softmax(logits);
 
-        // 获取 top-k 索引
         std::vector<size_t> indices(logits.size());
         std::iota(indices.begin(), indices.end(), 0);
-        std::partial_sort(indices.begin(), indices.begin() + k, indices.end(), [&](size_t i, size_t j)
-                          { return logits[i] > logits[j]; });
+        std::partial_sort(indices.begin(), indices.begin() + k, indices.end(),
+                          [&](size_t i, size_t j) { return logits[i] > logits[j]; });
 
-        // 仅保留 top-k 概率
         std::vector<size_t> filtered_indices(indices.begin(), indices.begin() + k);
         std::vector<float> filtered_probs(k);
-        for (size_t i = 0; i < k; ++i)
-        {
+        for (size_t i = 0; i < k; ++i) {
             filtered_probs[i] = logits[filtered_indices[i]];
         }
         filtered_probs = softmax(filtered_probs);
 
-        // 归一化
         float sum = std::accumulate(filtered_probs.begin(), filtered_probs.end(), 0.0f);
-        for (float &p : filtered_probs)
-        {
+        for (float &p : filtered_probs) {
             p /= sum;
         }
 
-        // 采样
         std::random_device rd;
         std::mt19937 gen(rd());
         std::discrete_distribution<int> dist(filtered_probs.begin(), filtered_probs.end());
@@ -217,64 +176,65 @@ class LLMPostprocess
     }
 
     bool enable_temperature = false;
-    float temperature = 1.0f;
+    float temperature       = 1.0f;
 
     bool enable_repetition_penalty = false;
-    float repetition_penalty = 1.0f;
-    int penalty_window = 20;
+    float repetition_penalty       = 1.0f;
+    int penalty_window             = 20;
 
     bool enable_diversity_penalty = false;
     std::vector<int> common_phrases;
     float diversity_penalty = 1.0f;
 
     bool enable_top_p_sampling = false;
-    float top_p = 1.0f;
+    float top_p                = 1.0f;
 
     bool enable_top_k_sampling = false;
-    int top_k = 1;
+    int top_k                  = 1;
 
 public:
-    LLMPostprocess() {}
+    LLMPostprocess()
+    {
+    }
 
     void set_temperature(bool enable, float temperature)
     {
         enable_temperature = enable;
-        this->temperature = temperature;
+        this->temperature  = temperature;
     }
 
     void set_repetition_penalty(bool enable, float penalty, int penalty_window)
     {
         enable_repetition_penalty = enable;
-        this->repetition_penalty = penalty;
-        this->penalty_window = penalty_window;
+        this->repetition_penalty  = penalty;
+        this->penalty_window      = penalty_window;
     }
 
     void set_diversity_penalty(bool enable, const std::vector<int> &common_phrases, float penalty)
     {
         enable_diversity_penalty = enable;
-        this->common_phrases = common_phrases;
-        this->diversity_penalty = penalty;
+        this->common_phrases     = common_phrases;
+        this->diversity_penalty  = penalty;
     }
 
     void set_top_p_sampling(bool enable, float top_p)
     {
         enable_top_k_sampling = false;
         enable_top_p_sampling = enable;
-        this->top_p = top_p;
+        this->top_p           = top_p;
     }
 
     void set_top_k_sampling(bool enable, int top_k)
     {
         enable_top_p_sampling = false;
         enable_top_k_sampling = enable;
-        this->top_k = top_k;
+        this->top_k           = top_k;
     }
 
     bool load_config(std::string config_path)
     {
         std::ifstream config_file(config_path);
-        if (!config_file.is_open())
-        {
+        if (!config_file.is_open()) {
             ALOGE("config file(%s) open failed", config_path.c_str());
             return false;
         }
@@ -282,21 +242,21 @@ class LLMPostprocess
         ALOGI("load config: \n%s\n", config.dump(4).c_str());
 
         enable_temperature = config["enable_temperature"];
-        temperature = config["temperature"];
+        temperature        = config["temperature"];
 
         enable_repetition_penalty = config["enable_repetition_penalty"];
-        repetition_penalty = config["repetition_penalty"];
-        penalty_window = config["penalty_window"];
+        repetition_penalty        = config["repetition_penalty"];
+        penalty_window            = config["penalty_window"];
 
         enable_top_p_sampling = config["enable_top_p_sampling"];
-        top_p = config["top_p"];
+        top_p                 = config["top_p"];
 
         enable_top_k_sampling = config["enable_top_k_sampling"];
-        top_k = config["top_k"];
+        top_k                 = config["top_k"];
         return true;
     }
 
-    bool load_config(const nlohmann::json& config)
+    bool load_config(const nlohmann::json &config)
     {
         if (config.is_null()) {
             ALOGE("config is null or invalid");
@@ -341,22 +301,17 @@ class LLMPostprocess
 
     int apply(std::vector<float> &logits, const std::vector<int> &history)
     {
-        if (enable_temperature)
-            apply_temperature(logits, temperature);
-        if (enable_repetition_penalty)
-            apply_repetition_penalty(logits, history, repetition_penalty, penalty_window);
-        if (enable_diversity_penalty)
-            apply_diversity_penalty(logits, common_phrases, diversity_penalty);
+        if (enable_temperature) apply_temperature(logits, temperature);
+        if (enable_repetition_penalty) apply_repetition_penalty(logits, history, repetition_penalty, penalty_window);
+        if (enable_diversity_penalty) apply_diversity_penalty(logits, common_phrases, diversity_penalty);
 
         if (enable_top_p_sampling)
             return faster_top_p_sampling(logits, top_p);
         else if (enable_top_k_sampling)
             return top_k_sampling(logits, top_k);
-        else
-        {
-            // 最大值
+        else {
             float max_logit = *std::max_element(logits.begin(), logits.end());
-            int max_index = std::distance(logits.begin(), std::max_element(logits.begin(), logits.end()));
+            int max_index   = std::distance(logits.begin(), std::max_element(logits.begin(), logits.end()));
             return max_index;
         }
     }
diff --git a/projects/llm_framework/main_openai_api/SConstruct b/projects/llm_framework/main_openai_api/SConstruct
index 0bde3482..f6e45d1d 100644
--- a/projects/llm_framework/main_openai_api/SConstruct
+++ b/projects/llm_framework/main_openai_api/SConstruct
@@ -18,7 +18,7 @@ LINK_SEARCH_PATH = []
 STATIC_FILES = []
 
 
-ModuleLLMOpenAIPluginPath = wget_github_commit('https://github.com/m5stack/ModuleLLM-OpenAI-Plugin.git', 'a8f54b0430c478896b45828f612d0d8b0a6f2fa1', True)
+ModuleLLMOpenAIPluginPath = wget_github_commit('https://github.com/m5stack/ModuleLLM-OpenAI-Plugin.git', '50efecaf566a04c1786b7eda2208a0cdefd53dde', True)
 python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-openai-api-python-venv_v1.6.tar.gz", 'm5stack_llm-openai-api-python-venv_v1.6.tar.gz')
 
 
diff --git a/projects/llm_framework/main_vlm/SConstruct b/projects/llm_framework/main_vlm/SConstruct
index 19239976..2d05a3b2 100644
--- a/projects/llm_framework/main_vlm/SConstruct
+++ b/projects/llm_framework/main_vlm/SConstruct
@@ -18,7 +18,7 @@ LINK_SEARCH_PATH = []
 STATIC_FILES = []
 
 if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
-    python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-vlm-python-venv_v1.6.tar.gz", 'm5stack_llm-vlm-python-venv_v1.6.tar.gz')
+    python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-vlm-python-venv_v1.8.tar.gz", 'm5stack_llm-vlm-python-venv_v1.8.tar.gz')
 else:
     python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-vlm-python-venv_v1.7.tar.gz", 'm5stack_llm-vlm-python-venv_v1.7.tar.gz')
 
diff --git a/projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax630c.json b/projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax630c.json
index fce6c7c5..72f1d249 100644
--- a/projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax630c.json
+++ b/projects/llm_framework/main_vlm/models/mode_qwen3-vl-2B-Int4-ax630c.json
@@ -46,7 +46,7 @@
         "video_token_id": 151656,
         "vision_start_token_id": 151652,
         "precompute_len": 0,
-        "cmm_size": 3582336,
+        "cmm_size": 3098624,
         "ext_scripts": [
             "tokenizer_qwen3-vl-2B-Int4-ax630c.py"
         ]
diff --git a/projects/llm_framework/main_vlm/src/main.cpp b/projects/llm_framework/main_vlm/src/main.cpp
index d4694e69..9b552d36 100644
--- a/projects/llm_framework/main_vlm/src/main.cpp
+++ b/projects/llm_framework/main_vlm/src/main.cpp
@@ -553,6 +553,8 @@ class llm_task {
         }
         if (lLaMa_) lLaMa_->Deinit();
         if (lLaMa_) lLaMa_.reset();
+        if (lLaMa_ctx_) lLaMa_ctx_->Deinit();
+        if (lLaMa_ctx_) lLaMa_ctx_.reset();
         if (qwen_) qwen_->Deinit();
         if (qwen_) qwen_.reset();
         return true;
diff --git a/projects/llm_framework/main_vlm/src/runner/LLM.hpp b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
index bf56bf84..59dd87e8 100644
--- a/projects/llm_framework/main_vlm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
@@ -859,6 +859,7 @@ class LLM_CTX {
             llama_layers[i].layer.release();
         }
         llama_post.release();
+        image_encoder.release();
         embed_selector.Deinit();
     }
 
diff --git a/projects/llm_framework/main_vlm/src/runner/LLMPostprocess.hpp b/projects/llm_framework/main_vlm/src/runner/LLMPostprocess.hpp
index b7d50156..08d6b4e7 100644
--- a/projects/llm_framework/main_vlm/src/runner/LLMPostprocess.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/LLMPostprocess.hpp
@@ -9,66 +9,50 @@
 #include "utils/json.hpp"
 #include "utils/sample_log.h"
 
-class LLMPostprocess
-{
+class LLMPostprocess {
 private:
-    // 	控制随机性
     void apply_temperature(std::vector<float> &logits, float temperature)
     {
-        for (float &logit : logits)
-        {
+        if (temperature == 0.0f) temperature = 0.01f;
+        for (float &logit : logits) {
             logit /= temperature;
         }
     }
 
-    // 防止重复
     void apply_repetition_penalty(std::vector<float> &logits, const std::vector<int> &history, float penalty)
     {
-        for (int token : history)
-        {
-            if (token < logits.size())
-            {
+        for (int token : history) {
+            if (token < logits.size()) {
                 logits[token] = logits[token] < 0 ? logits[token] * penalty : logits[token] / penalty;
             }
         }
     }
 
-    void apply_repetition_penalty(std::vector<float> &logits,
-                                  const std::vector<int> &generated_tokens,
-                                  float repetition_penalty,
-                                  int penalty_window)
+    void apply_repetition_penalty(std::vector<float> &logits, const std::vector<int> &generated_tokens,
+                                  float repetition_penalty, int penalty_window)
     {
-        if (repetition_penalty == 1.0f || generated_tokens.empty())
-        {
-            return; // 如果 penalty = 1.0 或者没有生成 token，则不进行修改
+        if (repetition_penalty == 1.0f || generated_tokens.empty()) {
+            return;
         }
 
         int start_idx = std::max(0, (int)generated_tokens.size() - penalty_window);
         std::unordered_set<int> recent_tokens(generated_tokens.begin() + start_idx, generated_tokens.end());
 
-        for (int token : recent_tokens)
-        {
-            if (token < 0 || token >= logits.size())
-                continue;
+        for (int token : recent_tokens) {
+            if (token < 0 || token >= logits.size()) continue;
 
-            if (logits[token] > 0)
-            {
+            if (logits[token] > 0) {
                 logits[token] /= std::sqrt(repetition_penalty);
-            }
-            else
-            {
+            } else {
                 logits[token] *= std::sqrt(repetition_penalty);
             }
         }
     }
 
-    // 增强多样性
     void apply_diversity_penalty(std::vector<float> &logits, const std::vector<int> &common_phrases, float penalty)
     {
-        for (int token : common_phrases)
-        {
-            if (token < logits.size())
-            {
+        for (int token : common_phrases) {
+            if (token < logits.size()) {
                 logits[token] *= penalty;
             }
         }
@@ -79,46 +63,37 @@ class LLMPostprocess
     {
         std::vector<float> probs(logits.size());
         float max_logit = *std::max_element(logits.begin(), logits.end());
-        float sum = 0.0f;
+        float sum       = 0.0f;
 
-        for (size_t i = 0; i < logits.size(); ++i)
-        {
+        for (size_t i = 0; i < logits.size(); ++i) {
             probs[i] = std::exp(logits[i] - max_logit);
             sum += probs[i];
         }
 
-        for (float &p : probs)
-        {
+        for (float &p : probs) {
             p /= sum;
         }
 
         return probs;
     }
 
-    // 	动态裁剪低概率 token
     int faster_top_p_sampling(const std::vector<float> &logits, float top_p)
     {
-        // 计算softmax
         std::vector<float> probs = softmax(logits);
 
-        // 构建最大堆（概率和索引的配对）
         std::vector<std::pair<float, size_t>> prob_index;
         prob_index.reserve(logits.size());
-        for (size_t i = 0; i < logits.size(); ++i)
-        {
+        for (size_t i = 0; i < logits.size(); ++i) {
             prob_index.emplace_back(probs[i], i);
         }
-        auto cmp = [](const auto &a, const auto &b)
-        { return a.first < b.first; };
+        auto cmp = [](const auto &a, const auto &b) { return a.first < b.first; };
         std::make_heap(prob_index.begin(), prob_index.end(), cmp);
 
-        // 提取top-p元素
         std::vector<size_t> filtered_indices;
         std::vector<float> filtered_probs;
         float cumulative_prob = 0.0f;
 
-        while (!prob_index.empty() && cumulative_prob < top_p)
-        {
+        while (!prob_index.empty() && cumulative_prob < top_p) {
             std::pop_heap(prob_index.begin(), prob_index.end(), cmp);
             auto [prob, index] = prob_index.back();
             prob_index.pop_back();
@@ -127,15 +102,11 @@ class LLMPostprocess
             filtered_indices.push_back(index);
             filtered_probs.push_back(prob);
 
-            if (cumulative_prob >= top_p)
-                break;
+            if (cumulative_prob >= top_p) break;
         }
 
-        // 处理边缘情况（概率全零时返回第一个元素）
-        if (filtered_indices.empty())
-            return 0;
+        if (filtered_indices.empty()) return 0;
 
-        // 使用thread_local随机数生成器（线程安全）
         static thread_local std::mt19937 gen(std::random_device{}());
         std::discrete_distribution<int> dist(filtered_probs.begin(), filtered_probs.end());
         return filtered_indices[dist(gen)];
@@ -147,31 +118,26 @@ class LLMPostprocess
         // Sort indices by probability in descending order
         std::vector<size_t> indices(logits.size());
         std::iota(indices.begin(), indices.end(), 0);
-        std::sort(indices.begin(), indices.end(), [&](size_t i, size_t j)
-                  { return probs[i] > probs[j]; });
+        std::sort(indices.begin(), indices.end(), [&](size_t i, size_t j) { return probs[i] > probs[j]; });
 
         // Compute cumulative probabilities
         float cumulative_prob = 0.0f;
-        size_t cut_off = 0;
-        for (; cut_off < indices.size(); ++cut_off)
-        {
+        size_t cut_off        = 0;
+        for (; cut_off < indices.size(); ++cut_off) {
             cumulative_prob += probs[indices[cut_off]];
-            if (cumulative_prob >= top_p)
-                break;
+            if (cumulative_prob >= top_p) break;
         }
 
         // Keep only the top-p probabilities
         std::vector<size_t> filtered_indices(indices.begin(), indices.begin() + cut_off + 1);
         std::vector<float> filtered_probs(filtered_indices.size());
-        for (size_t i = 0; i < filtered_indices.size(); ++i)
-        {
+        for (size_t i = 0; i < filtered_indices.size(); ++i) {
             filtered_probs[i] = probs[filtered_indices[i]];
         }
 
         // Normalize the probabilities
         float filtered_sum = std::accumulate(filtered_probs.begin(), filtered_probs.end(), 0.0f);
-        for (float &p : filtered_probs)
-        {
+        for (float &p : filtered_probs) {
             p /= filtered_sum;
         }
 
@@ -182,34 +148,27 @@ class LLMPostprocess
         return filtered_indices[dist(gen)];
     }
 
-    // 限制候选 token 数
     int top_k_sampling(const std::vector<float> &logits, int k)
     {
         // std::vector<float> probs = softmax(logits);
 
-        // 获取 top-k 索引
         std::vector<size_t> indices(logits.size());
         std::iota(indices.begin(), indices.end(), 0);
-        std::partial_sort(indices.begin(), indices.begin() + k, indices.end(), [&](size_t i, size_t j)
-                          { return logits[i] > logits[j]; });
+        std::partial_sort(indices.begin(), indices.begin() + k, indices.end(),
+                          [&](size_t i, size_t j) { return logits[i] > logits[j]; });
 
-        // 仅保留 top-k 概率
         std::vector<size_t> filtered_indices(indices.begin(), indices.begin() + k);
         std::vector<float> filtered_probs(k);
-        for (size_t i = 0; i < k; ++i)
-        {
+        for (size_t i = 0; i < k; ++i) {
             filtered_probs[i] = logits[filtered_indices[i]];
         }
         filtered_probs = softmax(filtered_probs);
 
-        // 归一化
         float sum = std::accumulate(filtered_probs.begin(), filtered_probs.end(), 0.0f);
-        for (float &p : filtered_probs)
-        {
+        for (float &p : filtered_probs) {
             p /= sum;
         }
 
-        // 采样
         std::random_device rd;
         std::mt19937 gen(rd());
         std::discrete_distribution<int> dist(filtered_probs.begin(), filtered_probs.end());
@@ -217,64 +176,65 @@ class LLMPostprocess
     }
 
     bool enable_temperature = false;
-    float temperature = 1.0f;
+    float temperature       = 1.0f;
 
     bool enable_repetition_penalty = false;
-    float repetition_penalty = 1.0f;
-    int penalty_window = 20;
+    float repetition_penalty       = 1.0f;
+    int penalty_window             = 20;
 
     bool enable_diversity_penalty = false;
     std::vector<int> common_phrases;
     float diversity_penalty = 1.0f;
 
     bool enable_top_p_sampling = false;
-    float top_p = 1.0f;
+    float top_p                = 1.0f;
 
     bool enable_top_k_sampling = false;
-    int top_k = 1;
+    int top_k                  = 1;
 
 public:
-    LLMPostprocess() {}
+    LLMPostprocess()
+    {
+    }
 
     void set_temperature(bool enable, float temperature)
     {
         enable_temperature = enable;
-        this->temperature = temperature;
+        this->temperature  = temperature;
     }
 
     void set_repetition_penalty(bool enable, float penalty, int penalty_window)
     {
         enable_repetition_penalty = enable;
-        this->repetition_penalty = penalty;
-        this->penalty_window = penalty_window;
+        this->repetition_penalty  = penalty;
+        this->penalty_window      = penalty_window;
     }
 
     void set_diversity_penalty(bool enable, const std::vector<int> &common_phrases, float penalty)
     {
         enable_diversity_penalty = enable;
-        this->common_phrases = common_phrases;
-        this->diversity_penalty = penalty;
+        this->common_phrases     = common_phrases;
+        this->diversity_penalty  = penalty;
     }
 
     void set_top_p_sampling(bool enable, float top_p)
     {
         enable_top_k_sampling = false;
         enable_top_p_sampling = enable;
-        this->top_p = top_p;
+        this->top_p           = top_p;
     }
 
     void set_top_k_sampling(bool enable, int top_k)
     {
         enable_top_p_sampling = false;
         enable_top_k_sampling = enable;
-        this->top_k = top_k;
+        this->top_k           = top_k;
     }
 
     bool load_config(std::string config_path)
     {
         std::ifstream config_file(config_path);
-        if (!config_file.is_open())
-        {
+        if (!config_file.is_open()) {
             ALOGE("config file(%s) open failed", config_path.c_str());
             return false;
         }
@@ -282,21 +242,21 @@ class LLMPostprocess
         ALOGI("load config: \n%s\n", config.dump(4).c_str());
 
         enable_temperature = config["enable_temperature"];
-        temperature = config["temperature"];
+        temperature        = config["temperature"];
 
         enable_repetition_penalty = config["enable_repetition_penalty"];
-        repetition_penalty = config["repetition_penalty"];
-        penalty_window = config["penalty_window"];
+        repetition_penalty        = config["repetition_penalty"];
+        penalty_window            = config["penalty_window"];
 
         enable_top_p_sampling = config["enable_top_p_sampling"];
-        top_p = config["top_p"];
+        top_p                 = config["top_p"];
 
         enable_top_k_sampling = config["enable_top_k_sampling"];
-        top_k = config["top_k"];
+        top_k                 = config["top_k"];
         return true;
     }
 
-    bool load_config(const nlohmann::json& config)
+    bool load_config(const nlohmann::json &config)
     {
         if (config.is_null()) {
             ALOGE("config is null or invalid");
@@ -341,22 +301,17 @@ class LLMPostprocess
 
     int apply(std::vector<float> &logits, const std::vector<int> &history)
     {
-        if (enable_temperature)
-            apply_temperature(logits, temperature);
-        if (enable_repetition_penalty)
-            apply_repetition_penalty(logits, history, repetition_penalty, penalty_window);
-        if (enable_diversity_penalty)
-            apply_diversity_penalty(logits, common_phrases, diversity_penalty);
+        if (enable_temperature) apply_temperature(logits, temperature);
+        if (enable_repetition_penalty) apply_repetition_penalty(logits, history, repetition_penalty, penalty_window);
+        if (enable_diversity_penalty) apply_diversity_penalty(logits, common_phrases, diversity_penalty);
 
         if (enable_top_p_sampling)
             return faster_top_p_sampling(logits, top_p);
         else if (enable_top_k_sampling)
             return top_k_sampling(logits, top_k);
-        else
-        {
-            // 最大值
+        else {
             float max_logit = *std::max_element(logits.begin(), logits.end());
-            int max_index = std::distance(logits.begin(), std::max_element(logits.begin(), logits.end()));
+            int max_index   = std::distance(logits.begin(), std::max_element(logits.begin(), logits.end()));
             return max_index;
         }
     }

From 4bc10a7932fcd0b237bbb37f0b0ac2068cdd0c91 Mon Sep 17 00:00:00 2001
From: dianjixz <18637716021@163.com>
Date: Wed, 3 Dec 2025 15:20:14 +0800
Subject: [PATCH 63/79] [fix] pzmq creat error

---
 ext_components/StackFlow/stackflow/pzmq.hpp | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/ext_components/StackFlow/stackflow/pzmq.hpp b/ext_components/StackFlow/stackflow/pzmq.hpp
index 3adbab55..322c463c 100644
--- a/ext_components/StackFlow/stackflow/pzmq.hpp
+++ b/ext_components/StackFlow/stackflow/pzmq.hpp
@@ -127,7 +127,7 @@ class pzmq {
         if ((url[0] != 'i') && (url[1] != 'p')) {
             rpc_url_head_.clear();
         }
-        if (mode_ != ZMQ_RPC_FUN) creat(url, raw_call);
+        creat(url, raw_call);
     }
     void set_timeout(int ms)
     {
@@ -179,8 +179,6 @@ class pzmq {
         if (zmq_fun_.empty()) {
             std::string url = rpc_url_head_ + rpc_server_;
             mode_           = ZMQ_RPC_FUN;
-            zmq_fun_["list_action"] =
-                std::bind(&pzmq::_rpc_list_action, this, std::placeholders::_1, std::placeholders::_2);
             ret = creat(url);
         }
         zmq_fun_[action] = raw_call;
@@ -330,6 +328,7 @@ class pzmq {
     inline int creat_rep(const std::string &url, const msg_callback_fun &raw_call)
     {
         int ret     = zmq_bind(zmq_socket_, url.c_str());
+        zmq_fun_["list_action"] = std::bind(&pzmq::_rpc_list_action, this, std::placeholders::_1, std::placeholders::_2);
         flage_      = false;
         zmq_thread_ = std::make_unique<std::thread>(std::bind(&pzmq::zmq_event_loop, this, raw_call));
         return ret;

From f12314e91ef3641521b0ca198cf474428277c004 Mon Sep 17 00:00:00 2001
From: dianjixz <18637716021@163.com>
Date: Wed, 3 Dec 2025 15:20:29 +0800
Subject: [PATCH 64/79] [update] del ec_prox

---
 .../llm_framework/main_ax650_ec_proxy/Kconfig |   0
 .../main_ax650_ec_proxy/SConstruct            |  58 --
 .../main_ax650_ec_proxy/src/cmdline.hpp       | 732 -------------
 .../main_ax650_ec_proxy/src/ec_cli_main.cpp   | 317 ------
 .../main_ax650_ec_proxy/src/ec_proxy_main.cpp | 963 ------------------
 5 files changed, 2070 deletions(-)
 delete mode 100644 projects/llm_framework/main_ax650_ec_proxy/Kconfig
 delete mode 100644 projects/llm_framework/main_ax650_ec_proxy/SConstruct
 delete mode 100644 projects/llm_framework/main_ax650_ec_proxy/src/cmdline.hpp
 delete mode 100644 projects/llm_framework/main_ax650_ec_proxy/src/ec_cli_main.cpp
 delete mode 100644 projects/llm_framework/main_ax650_ec_proxy/src/ec_proxy_main.cpp

diff --git a/projects/llm_framework/main_ax650_ec_proxy/Kconfig b/projects/llm_framework/main_ax650_ec_proxy/Kconfig
deleted file mode 100644
index e69de29b..00000000
diff --git a/projects/llm_framework/main_ax650_ec_proxy/SConstruct b/projects/llm_framework/main_ax650_ec_proxy/SConstruct
deleted file mode 100644
index 1a908a3a..00000000
--- a/projects/llm_framework/main_ax650_ec_proxy/SConstruct
+++ /dev/null
@@ -1,58 +0,0 @@
-import os
-
-Import('env')
-with open(env['PROJECT_TOOL_S']) as f:
-    exec(f.read())
-
-if "CONFIG_AX_650C_MSP_ENABLED" not in os.environ:
-    Return('env')
-
-SRCS = []
-INCLUDE = []
-PRIVATE_INCLUDE = []
-REQUIREMENTS = ['pthread', 'utilities', 'eventpp', 'StackFlow', 'modbus_component', 'fmt']
-STATIC_LIB = []
-DYNAMIC_LIB = []
-DEFINITIONS = []
-DEFINITIONS_PRIVATE = []
-LDFLAGS = []
-LINK_SEARCH_PATH = []
-STATIC_FILES = []
-
-LINK_SEARCH_PATH += [ADir('../static_lib')]
-
-LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
-
-STATIC_FILES += []
-
-
-SRCS = [AFile('src/ec_proxy_main.cpp')]
-env['COMPONENTS'].append({'target':'ax650_ec_proxy-1.0',
-                          'SRCS':SRCS,
-                          'INCLUDE':INCLUDE,
-                          'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
-                          'REQUIREMENTS':REQUIREMENTS,
-                          'STATIC_LIB':STATIC_LIB,
-                          'DYNAMIC_LIB':DYNAMIC_LIB,
-                          'DEFINITIONS':DEFINITIONS,
-                          'DEFINITIONS_PRIVATE':DEFINITIONS_PRIVATE,
-                          'LDFLAGS':LDFLAGS,
-                          'LINK_SEARCH_PATH':LINK_SEARCH_PATH,
-                          'STATIC_FILES':STATIC_FILES,
-                          'REGISTER':'project'
-                          })
-SRCS = [AFile('src/ec_cli_main.cpp')]
-env['COMPONENTS'].append({'target':'ax650_ec_cli-1.0',
-                          'SRCS':SRCS,
-                          'INCLUDE':INCLUDE,
-                          'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
-                          'REQUIREMENTS':REQUIREMENTS,
-                          'STATIC_LIB':STATIC_LIB,
-                          'DYNAMIC_LIB':DYNAMIC_LIB,
-                          'DEFINITIONS':DEFINITIONS,
-                          'DEFINITIONS_PRIVATE':DEFINITIONS_PRIVATE,
-                          'LDFLAGS':LDFLAGS,
-                          'LINK_SEARCH_PATH':LINK_SEARCH_PATH,
-                          'STATIC_FILES':STATIC_FILES,
-                          'REGISTER':'project'
-                          })
diff --git a/projects/llm_framework/main_ax650_ec_proxy/src/cmdline.hpp b/projects/llm_framework/main_ax650_ec_proxy/src/cmdline.hpp
deleted file mode 100644
index ea8b59e2..00000000
--- a/projects/llm_framework/main_ax650_ec_proxy/src/cmdline.hpp
+++ /dev/null
@@ -1,732 +0,0 @@
-/*
-  Copyright (c) 2009, Hideyuki Tanaka
-  All rights reserved.
-
-  Redistribution and use in source and binary forms, with or without
-  modification, are permitted provided that the following conditions are met:
-  * Redistributions of source code must retain the above copyright
-  notice, this list of conditions and the following disclaimer.
-  * Redistributions in binary form must reproduce the above copyright
-  notice, this list of conditions and the following disclaimer in the
-  documentation and/or other materials provided with the distribution.
-  * Neither the name of the <organization> nor the
-  names of its contributors may be used to endorse or promote products
-  derived from this software without specific prior written permission.
-
-  THIS SOFTWARE IS PROVIDED BY <copyright holder> ''AS IS'' AND ANY
-  EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
-  WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
-  DISCLAIMED. IN NO EVENT SHALL <copyright holder> BE LIABLE FOR ANY
-  DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
-  (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
-  LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
-  ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-  (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
-  SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-*/
-
-#pragma once
-
-#include <cxxabi.h>
-
-#include <algorithm>
-#include <cstdlib>
-#include <cstring>
-#include <iostream>
-#include <map>
-#include <sstream>
-#include <stdexcept>
-#include <string>
-#include <typeinfo>
-#include <vector>
-
-namespace cmdline {
-
-namespace detail {
-
-template <typename Target, typename Source, bool Same>
-class lexical_cast_t {
- public:
-  static Target cast(const Source &arg) {
-    Target ret;
-    std::stringstream ss;
-    if (!(ss << arg && ss >> ret && ss.eof())) throw std::bad_cast();
-
-    return ret;
-  }
-};
-
-template <typename Target, typename Source>
-class lexical_cast_t<Target, Source, true> {
- public:
-  static Target cast(const Source &arg) { return arg; }
-};
-
-template <typename Source>
-class lexical_cast_t<std::string, Source, false> {
- public:
-  static std::string cast(const Source &arg) {
-    std::ostringstream ss;
-    ss << arg;
-    return ss.str();
-  }
-};
-
-template <typename Target>
-class lexical_cast_t<Target, std::string, false> {
- public:
-  static Target cast(const std::string &arg) {
-    Target ret;
-    std::istringstream ss(arg);
-    if (!(ss >> ret && ss.eof())) throw std::bad_cast();
-    return ret;
-  }
-};
-
-template <typename T1, typename T2>
-struct is_same {
-  static const bool value = false;
-};
-
-template <typename T>
-struct is_same<T, T> {
-  static const bool value = true;
-};
-
-template <typename Target, typename Source>
-Target lexical_cast(const Source &arg) {
-  return lexical_cast_t<Target, Source,
-                        detail::is_same<Target, Source>::value>::cast(arg);
-}
-
-static inline std::string demangle(const std::string &name) {
-  int status = 0;
-  char *p = abi::__cxa_demangle(name.c_str(), 0, 0, &status);
-  std::string ret(p);
-  free(p);
-  return ret;
-}
-
-template <class T>
-std::string readable_typename() {
-  return demangle(typeid(T).name());
-}
-
-template <class T>
-std::string default_value(T def) {
-  return detail::lexical_cast<std::string>(def);
-}
-
-template <>
-inline std::string readable_typename<std::string>() {
-  return "string";
-}
-
-}  // namespace detail
-
-//-----
-
-class cmdline_error : public std::exception {
- public:
-  cmdline_error(const std::string &msg) : msg(msg) {}
-  ~cmdline_error() throw() {}
-  const char *what() const throw() { return msg.c_str(); }
-
- private:
-  std::string msg;
-};
-
-template <class T>
-struct default_reader {
-  T operator()(const std::string &str) { return detail::lexical_cast<T>(str); }
-};
-
-template <class T>
-struct range_reader {
-  range_reader(const T &low, const T &high) : low(low), high(high) {}
-  T operator()(const std::string &s) const {
-    T ret = default_reader<T>()(s);
-    if (!(ret >= low && ret <= high))
-      throw cmdline::cmdline_error("range_error");
-    return ret;
-  }
-
- private:
-  T low, high;
-};
-
-template <class T>
-range_reader<T> range(const T &low, const T &high) {
-  return range_reader<T>(low, high);
-}
-
-template <class T>
-struct oneof_reader {
-  T operator()(const std::string &s) {
-    T ret = default_reader<T>()(s);
-    if (std::find(alt.begin(), alt.end(), ret) == alt.end())
-      throw cmdline_error("");
-    return ret;
-  }
-  void add(const T &v) { alt.push_back(v); }
-
- private:
-  std::vector<T> alt;
-};
-
-template <class T>
-oneof_reader<T> oneof(T a1) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3, T a4) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  ret.add(a4);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  ret.add(a4);
-  ret.add(a5);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5, T a6) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  ret.add(a4);
-  ret.add(a5);
-  ret.add(a6);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5, T a6, T a7) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  ret.add(a4);
-  ret.add(a5);
-  ret.add(a6);
-  ret.add(a7);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5, T a6, T a7, T a8) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  ret.add(a4);
-  ret.add(a5);
-  ret.add(a6);
-  ret.add(a7);
-  ret.add(a8);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5, T a6, T a7, T a8, T a9) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  ret.add(a4);
-  ret.add(a5);
-  ret.add(a6);
-  ret.add(a7);
-  ret.add(a8);
-  ret.add(a9);
-  return ret;
-}
-
-template <class T>
-oneof_reader<T> oneof(T a1, T a2, T a3, T a4, T a5, T a6, T a7, T a8, T a9,
-                      T a10) {
-  oneof_reader<T> ret;
-  ret.add(a1);
-  ret.add(a2);
-  ret.add(a3);
-  ret.add(a4);
-  ret.add(a5);
-  ret.add(a6);
-  ret.add(a7);
-  ret.add(a8);
-  ret.add(a9);
-  ret.add(a10);
-  return ret;
-}
-
-//-----
-
-class parser {
- public:
-  parser() {}
-  ~parser() {
-    for (std::map<std::string, option_base *>::iterator p = options.begin();
-         p != options.end(); p++)
-      delete p->second;
-  }
-
-  void add(const std::string &name, char short_name = 0,
-           const std::string &desc = "") {
-    if (options.count(name))
-      throw cmdline_error("multiple definition: " + name);
-    options[name] = new option_without_value(name, short_name, desc);
-    ordered.push_back(options[name]);
-  }
-
-  template <class T>
-  void add(const std::string &name, char short_name = 0,
-           const std::string &desc = "", bool need = true, const T def = T()) {
-    add(name, short_name, desc, need, def, default_reader<T>());
-  }
-
-  template <class T, class F>
-  void add(const std::string &name, char short_name = 0,
-           const std::string &desc = "", bool need = true, const T def = T(),
-           F reader = F()) {
-    if (options.count(name))
-      throw cmdline_error("multiple definition: " + name);
-    options[name] = new option_with_value_with_reader<T, F>(
-        name, short_name, need, def, desc, reader);
-    ordered.push_back(options[name]);
-  }
-
-  void footer(const std::string &f) { ftr = f; }
-
-  void set_program_name(const std::string &name) { prog_name = name; }
-
-  bool exist(const std::string &name) const {
-    if (options.count(name) == 0)
-      throw cmdline_error("there is no flag: --" + name);
-    return options.find(name)->second->has_set();
-  }
-
-  template <class T>
-  const T &get(const std::string &name) const {
-    if (options.count(name) == 0)
-      throw cmdline_error("there is no flag: --" + name);
-    const option_with_value<T> *p =
-        dynamic_cast<const option_with_value<T> *>(options.find(name)->second);
-    if (p == NULL) throw cmdline_error("type mismatch flag '" + name + "'");
-    return p->get();
-  }
-
-  const std::vector<std::string> &rest() const { return others; }
-
-  bool parse(const std::string &arg) {
-    std::vector<std::string> args;
-
-    std::string buf;
-    bool in_quote = false;
-    for (std::string::size_type i = 0; i < arg.length(); i++) {
-      if (arg[i] == '\"') {
-        in_quote = !in_quote;
-        continue;
-      }
-
-      if (arg[i] == ' ' && !in_quote) {
-        args.push_back(buf);
-        buf = "";
-        continue;
-      }
-
-      if (arg[i] == '\\') {
-        i++;
-        if (i >= arg.length()) {
-          errors.push_back("unexpected occurrence of '\\' at end of string");
-          return false;
-        }
-      }
-
-      buf += arg[i];
-    }
-
-    if (in_quote) {
-      errors.push_back("quote is not closed");
-      return false;
-    }
-
-    if (buf.length() > 0) args.push_back(buf);
-
-    for (size_t i = 0; i < args.size(); i++)
-      std::cout << "\"" << args[i] << "\"" << std::endl;
-
-    return parse(args);
-  }
-
-  bool parse(const std::vector<std::string> &args) {
-    int argc = static_cast<int>(args.size());
-    std::vector<const char *> argv(argc);
-
-    for (int i = 0; i < argc; i++) argv[i] = args[i].c_str();
-
-    return parse(argc, &argv[0]);
-  }
-
-  bool parse(int argc, const char *const argv[]) {
-    errors.clear();
-    others.clear();
-
-    if (argc < 1) {
-      errors.push_back("argument number must be longer than 0");
-      return false;
-    }
-    if (prog_name == "") prog_name = argv[0];
-
-    std::map<char, std::string> lookup;
-    for (std::map<std::string, option_base *>::iterator p = options.begin();
-         p != options.end(); p++) {
-      if (p->first.length() == 0) continue;
-      char initial = p->second->short_name();
-      if (initial) {
-        if (lookup.count(initial) > 0) {
-          lookup[initial] = "";
-          errors.push_back(std::string("short option '") + initial +
-                           "' is ambiguous");
-          return false;
-        } else
-          lookup[initial] = p->first;
-      }
-    }
-
-    for (int i = 1; i < argc; i++) {
-      if (strncmp(argv[i], "--", 2) == 0) {
-        const char *p = strchr(argv[i] + 2, '=');
-        if (p) {
-          std::string name(argv[i] + 2, p);
-          std::string val(p + 1);
-          set_option(name, val);
-        } else {
-          std::string name(argv[i] + 2);
-          if (options.count(name) == 0) {
-            errors.push_back("undefined option: --" + name);
-            continue;
-          }
-          if (options[name]->has_value()) {
-            if (i + 1 >= argc) {
-              errors.push_back("option needs value: --" + name);
-              continue;
-            } else {
-              i++;
-              set_option(name, argv[i]);
-            }
-          } else {
-            set_option(name);
-          }
-        }
-      } else if (strncmp(argv[i], "-", 1) == 0) {
-        if (!argv[i][1]) continue;
-        char last = argv[i][1];
-        for (int j = 2; argv[i][j]; j++) {
-          last = argv[i][j];
-          if (lookup.count(argv[i][j - 1]) == 0) {
-            errors.push_back(std::string("undefined short option: -") +
-                             argv[i][j - 1]);
-            continue;
-          }
-          if (lookup[argv[i][j - 1]] == "") {
-            errors.push_back(std::string("ambiguous short option: -") +
-                             argv[i][j - 1]);
-            continue;
-          }
-          set_option(lookup[argv[i][j - 1]]);
-        }
-
-        if (lookup.count(last) == 0) {
-          errors.push_back(std::string("undefined short option: -") + last);
-          continue;
-        }
-        if (lookup[last] == "") {
-          errors.push_back(std::string("ambiguous short option: -") + last);
-          continue;
-        }
-
-        if (i + 1 < argc && options[lookup[last]]->has_value()) {
-          set_option(lookup[last], argv[i + 1]);
-          i++;
-        } else {
-          set_option(lookup[last]);
-        }
-      } else {
-        others.push_back(argv[i]);
-      }
-    }
-
-    for (std::map<std::string, option_base *>::iterator p = options.begin();
-         p != options.end(); p++)
-      if (!p->second->valid())
-        errors.push_back("need option: --" + std::string(p->first));
-
-    return errors.size() == 0;
-  }
-
-  void parse_check(const std::string &arg) {
-    if (!options.count("help")) add("help", '?', "print this message");
-    check(0, parse(arg));
-  }
-
-  void parse_check(const std::vector<std::string> &args) {
-    if (!options.count("help")) add("help", '?', "print this message");
-    check(args.size(), parse(args));
-  }
-
-  void parse_check(int argc, char *argv[]) {
-    if (!options.count("help")) add("help", '?', "print this message");
-    check(argc, parse(argc, argv));
-  }
-
-  std::string error() const { return errors.size() > 0 ? errors[0] : ""; }
-
-  std::string error_full() const {
-    std::ostringstream oss;
-    for (size_t i = 0; i < errors.size(); i++) oss << errors[i] << std::endl;
-    return oss.str();
-  }
-
-  std::string usage() const {
-    std::ostringstream oss;
-    oss << "usage: " << prog_name << " ";
-    for (size_t i = 0; i < ordered.size(); i++) {
-      if (ordered[i]->must()) oss << ordered[i]->short_description() << " ";
-    }
-
-    oss << "[options] ... " << ftr << std::endl;
-    oss << "options:" << std::endl;
-
-    size_t max_width = 0;
-    for (size_t i = 0; i < ordered.size(); i++) {
-      max_width = std::max(max_width, ordered[i]->name().length());
-    }
-    for (size_t i = 0; i < ordered.size(); i++) {
-      if (ordered[i]->short_name()) {
-        oss << "  -" << ordered[i]->short_name() << ", ";
-      } else {
-        oss << "      ";
-      }
-
-      oss << "--" << ordered[i]->name();
-      for (size_t j = ordered[i]->name().length(); j < max_width + 4; j++)
-        oss << ' ';
-      oss << ordered[i]->description() << std::endl;
-    }
-    return oss.str();
-  }
-
- private:
-  void check(int argc, bool ok) {
-    if ((argc == 1 && !ok) || exist("help")) {
-      std::cerr << usage();
-      exit(0);
-    }
-
-    if (!ok) {
-      std::cerr << error() << std::endl << usage();
-      exit(1);
-    }
-  }
-
-  void set_option(const std::string &name) {
-    if (options.count(name) == 0) {
-      errors.push_back("undefined option: --" + name);
-      return;
-    }
-    if (!options[name]->set()) {
-      errors.push_back("option needs value: --" + name);
-      return;
-    }
-  }
-
-  void set_option(const std::string &name, const std::string &value) {
-    if (options.count(name) == 0) {
-      errors.push_back("undefined option: --" + name);
-      return;
-    }
-    if (!options[name]->set(value)) {
-      errors.push_back("option value is invalid: --" + name + "=" + value);
-      return;
-    }
-  }
-
-  class option_base {
-   public:
-    virtual ~option_base() {}
-
-    virtual bool has_value() const = 0;
-    virtual bool set() = 0;
-    virtual bool set(const std::string &value) = 0;
-    virtual bool has_set() const = 0;
-    virtual bool valid() const = 0;
-    virtual bool must() const = 0;
-
-    virtual const std::string &name() const = 0;
-    virtual char short_name() const = 0;
-    virtual const std::string &description() const = 0;
-    virtual std::string short_description() const = 0;
-  };
-
-  class option_without_value : public option_base {
-   public:
-    option_without_value(const std::string &name, char short_name,
-                         const std::string &desc)
-        : nam(name), snam(short_name), desc(desc), has(false) {}
-    ~option_without_value() {}
-
-    bool has_value() const { return false; }
-
-    bool set() {
-      has = true;
-      return true;
-    }
-
-    bool set(const std::string &) { return false; }
-
-    bool has_set() const { return has; }
-
-    bool valid() const { return true; }
-
-    bool must() const { return false; }
-
-    const std::string &name() const { return nam; }
-
-    char short_name() const { return snam; }
-
-    const std::string &description() const { return desc; }
-
-    std::string short_description() const { return "--" + nam; }
-
-   private:
-    std::string nam;
-    char snam;
-    std::string desc;
-    bool has;
-  };
-
-  template <class T>
-  class option_with_value : public option_base {
-   public:
-    option_with_value(const std::string &name, char short_name, bool need,
-                      const T &def, const std::string &desc)
-        : nam(name),
-          snam(short_name),
-          need(need),
-          has(false),
-          def(def),
-          actual(def) {
-      this->desc = full_description(desc);
-    }
-    ~option_with_value() {}
-
-    const T &get() const { return actual; }
-
-    bool has_value() const { return true; }
-
-    bool set() { return false; }
-
-    bool set(const std::string &value) {
-      try {
-        actual = read(value);
-        has = true;
-      } catch (const std::exception &e) {
-        return false;
-      }
-      return true;
-    }
-
-    bool has_set() const { return has; }
-
-    bool valid() const {
-      if (need && !has) return false;
-      return true;
-    }
-
-    bool must() const { return need; }
-
-    const std::string &name() const { return nam; }
-
-    char short_name() const { return snam; }
-
-    const std::string &description() const { return desc; }
-
-    std::string short_description() const {
-      return "--" + nam + "=" + detail::readable_typename<T>();
-    }
-
-   protected:
-    std::string full_description(const std::string &desc) {
-      return desc + " (" + detail::readable_typename<T>() +
-             (need ? "" : " [=" + detail::default_value<T>(def) + "]") + ")";
-    }
-
-    virtual T read(const std::string &s) = 0;
-
-    std::string nam;
-    char snam;
-    bool need;
-    std::string desc;
-
-    bool has;
-    T def;
-    T actual;
-  };
-
-  template <class T, class F>
-  class option_with_value_with_reader : public option_with_value<T> {
-   public:
-    option_with_value_with_reader(const std::string &name, char short_name,
-                                  bool need, const T def,
-                                  const std::string &desc, F reader)
-        : option_with_value<T>(name, short_name, need, def, desc),
-          reader(reader) {}
-
-   private:
-    T read(const std::string &s) { return reader(s); }
-
-    F reader;
-  };
-
-  std::map<std::string, option_base *> options;
-  std::vector<option_base *> ordered;
-  std::string ftr;
-
-  std::string prog_name;
-  std::vector<std::string> others;
-
-  std::vector<std::string> errors;
-};
-
-}  // namespace cmdline
diff --git a/projects/llm_framework/main_ax650_ec_proxy/src/ec_cli_main.cpp b/projects/llm_framework/main_ax650_ec_proxy/src/ec_cli_main.cpp
deleted file mode 100644
index 1e47073e..00000000
--- a/projects/llm_framework/main_ax650_ec_proxy/src/ec_cli_main.cpp
+++ /dev/null
@@ -1,317 +0,0 @@
-#include <iostream>
-#include <modbus.h>
-#include "pzmq.hpp"
-#include "cmdline.hpp"
-#include <../../static_lib/include/fmt/core.h>
-#include <../../static_lib/include/fmt/format.h>
-#include <cstdlib>
-#include <list>
-#include <vector>
-#include "json.hpp"
-int condition = 1;
-const char *rpc_socket_path;
-const char *pub_socket_path;
-
-void info_fun(cmdline::parser &a, std::string set_fun, std::string get_fun)
-{
-    if (a.exist("data") || a.exist("DataRaw")) {
-        if (set_fun.length() == 0) return;
-        std::string DataRaw = a.get<std::string>("DataRaw");
-        if (DataRaw.empty()) {
-            DataRaw = fmt::format("{{ \"data\": {} }}", a.get<std::string>("data"));
-        }
-        StackFlows::pzmq Context(rpc_socket_path);
-        Context.call_rpc_action(set_fun, DataRaw,
-                                [set_fun](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
-                                    if (data->size() > 0) {
-                                        std::cout << data->string() << std::endl;
-                                    } else {
-                                        std::cout << "call " << set_fun << " faile!" << std::endl;
-                                    }
-                                });
-    } else {
-        if (get_fun.length() == 0) {
-            std::cout << "please set data: -d or -D" << std::endl;
-            return;
-        }
-        StackFlows::pzmq Context(rpc_socket_path);
-        Context.call_rpc_action(get_fun, "{}",
-                                [get_fun](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
-                                    if (data->size() > 0) {
-                                        std::cout << data->string() << std::endl;
-                                    } else {
-                                        std::cout << "call " << get_fun << " faile! " << std::endl;
-                                    }
-                                });
-    }
-}
-
-void get_info_fun(cmdline::parser &a, std::string fun)
-{
-    StackFlows::pzmq Context(rpc_socket_path);
-    Context.call_rpc_action(fun, "{\"data\":\"None\"}",
-                            [fun](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
-                                if (data->size() > 0) {
-                                    std::cout << data->string() << std::endl;
-                                } else {
-                                    std::cout << "call " << fun << " faile!" << std::endl;
-                                }
-                            });
-}
-
-void set_info_fun(cmdline::parser &a, std::string fun)
-{
-    if (a.exist("data") || a.exist("DataRaw")) {
-        std::string DataRaw = a.get<std::string>("DataRaw");
-        if (DataRaw.empty()) {
-            DataRaw = fmt::format("{{ \"data\": {} }}", a.get<std::string>("data"));
-        }
-        StackFlows::pzmq Context(rpc_socket_path);
-        Context.call_rpc_action(fun, DataRaw,
-                                [fun](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
-                                    if (data->size() > 0) {
-                                        std::cout << data->string() << std::endl;
-                                    } else {
-                                        std::cout << "call " << fun << "faile!" << std::endl;
-                                    }
-                                });
-    }
-}
-
-void fan_speed_fun(cmdline::parser &a)
-{
-    StackFlows::pzmq Context(rpc_socket_path);
-    Context.call_rpc_action("fan_get_speed", "{}",
-                            [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
-                                if (data->size() > 0) {
-                                    std::cout << data->string() << std::endl;
-                                } else {
-                                    std::cout << "call fan_get_speed faile!" << std::endl;
-                                }
-                            });
-}
-
-void fan_pwm_fun(cmdline::parser &a)
-{
-    if (a.exist("data") || a.exist("DataRaw")) {
-        std::string DataRaw = a.get<std::string>("DataRaw");
-        if (DataRaw.empty()) {
-            DataRaw = fmt::format("{{ \"data\": {} }}", a.get<std::string>("data"));
-        }
-        StackFlows::pzmq Context(rpc_socket_path);
-        Context.call_rpc_action("fan_set_pwm", DataRaw,
-                                [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
-                                    if (data->size() > 0) {
-                                        std::cout << data->string() << std::endl;
-                                    } else {
-                                        std::cout << "call fan_set_pwm faile!" << std::endl;
-                                    }
-                                });
-    } else {
-        StackFlows::pzmq Context(rpc_socket_path);
-        Context.call_rpc_action("fan_get_pwm", "{}",
-                                [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
-                                    if (data->size() > 0) {
-                                        std::cout << data->string() << std::endl;
-                                    } else {
-                                        std::cout << "call fan_get_pwm faile!" << std::endl;
-                                    }
-                                });
-    }
-}
-
-void rgb_fun(cmdline::parser &a)
-{
-    if (a.exist("data") || a.exist("DataRaw")) {
-        std::string DataRaw = a.get<std::string>("DataRaw");
-        if (DataRaw.empty()) {
-            DataRaw = fmt::format("{{ \"data\": {} }}", a.get<std::string>("data"));
-        }
-        StackFlows::pzmq Context(rpc_socket_path);
-        Context.call_rpc_action("rgb_set_mode", DataRaw,
-                                [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
-                                    if (data->size() > 0) {
-                                        std::cout << data->string() << std::endl;
-                                    } else {
-                                        std::cout << "call rgb_set_mode faile!" << std::endl;
-                                    }
-                                });
-    } else {
-        StackFlows::pzmq Context(rpc_socket_path);
-        Context.call_rpc_action("rgb_get_mode", "{}",
-                                [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
-                                    if (data->size() > 0) {
-                                        std::cout << data->string() << std::endl;
-                                    } else {
-                                        std::cout << "call rgb_get_mode faile!" << std::endl;
-                                    }
-                                });
-    }
-}
-
-void exec_fun(cmdline::parser &a)
-{
-    if (a.exist("list")) {
-        StackFlows::pzmq Context(rpc_socket_path);
-        std::cout << "list" << std::endl;
-        Context.call_rpc_action("list_action", "None",
-                                [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
-                                    if (data->size() > 0) {
-                                        try {
-                                            nlohmann::json _data = nlohmann::json::parse(data->string());
-                                            std::cout << _data.dump(4) << std::endl;
-                                        } catch (const nlohmann::json::parse_error &e) {
-                                            std::cerr << "解析失败：" << e.what() << std::endl;
-                                        }
-                                    } else {
-                                        std::cout << "call list_action faile!" << std::endl;
-                                    }
-                                });
-        return;
-    }
-
-    const std::string fun  = a.get<std::string>("fun");
-    const std::string data = a.get<std::string>("data");
-    if (fun.length() == 0) {
-        std::cout << "fun is empty! cli exec -f <fun> -d <data>" << std::endl;
-        return;
-    }
-
-    StackFlows::pzmq Context(rpc_socket_path);
-    Context.call_rpc_action(fun, data, [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
-        if (data->size() > 0) {
-            std::cout << "" << data->string() << std::endl;
-        } else {
-            std::cout << "call ax650_ec_prox faile!" << std::endl;
-        }
-    });
-}
-
-void echo_fun(cmdline::parser &a)
-{
-    if (a.exist("button")) {
-        StackFlows::pzmq Context(pub_socket_path, ZMQ_SUB,
-                                 [](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) {
-                                     std::cout << data->string() << std::endl;
-                                 });
-        while (condition) {
-            usleep(100 * 1000);
-        }
-        exit(0);
-    }
-}
-
-void signalHandler(int signum)
-{
-    condition = 0;
-    std::exit(0);  // 这会调用全局对象的析构函数
-}
-
-struct call_fun {
-    std::string fun;
-    char flage;
-    std::string dec;
-    std::function<void(cmdline::parser &a)> fun_call;
-};
-
-int main(int argc, char *argv[])
-{
-    if (argc < 2) {
-    error_print_exit:
-        std::cout << R"help(
-用法：git [--version] [--help] 
-           <命令> [<参数>]
-
-这些是各种场合常见的 cli 命令：
-
-直接操作 ec （参见：cli help device）
-   device    直接执行一个操作，使用指令传递参数
-
-自定义操作 ec （参见：git help exec）
-   exec       执行一个原始函数调用
-
-订阅 prox 的通道数据（参见：git help echo）
-   echo      显示 prox 的通道数据
-
-命令 'cli help -a' 和 'cli help -g' 显示可用的子命令和一些概念帮助。
-查看 'cli help <命令>' 或 'cli help <概念>' 以获取给定子命令或概念的
-帮助。
-)help" << std::endl;
-        return 0;
-    }
-    signal(SIGTERM, signalHandler);
-    signal(SIGINT, signalHandler);
-    rpc_socket_path = getenv("AX650C_EC_PROXY_RPC_SOCKET");
-    if (rpc_socket_path == NULL) {
-        rpc_socket_path = "ipc:///tmp/rpc.ec_prox";
-    }
-    pub_socket_path = getenv("AX650C_EC_PROXY_PUB_SOCKET");
-    if (pub_socket_path == NULL) {
-        pub_socket_path = "ipc:///tmp/llm/ec_prox.event.socket";
-    }
-    if (std::string(argv[1]) == "device") {
-        cmdline::parser a;
-        // clang-format off
-        std::list<struct call_fun> cmd_list = {
-            {"rgb", 'r', "rgb_mode", std::bind(&info_fun, std::placeholders::_1, "rgb_set_mode", "rgb_get_mode")},
-            {"rgb_size", 0, "rgb_size", std::bind(&info_fun, std::placeholders::_1, "rgb_set_size", "rgb_get_size")},
-            {"rgb_get_color", 0, "rgb_get_color", std::bind(&info_fun, std::placeholders::_1, "rgb_get_color", "")},
-            {"rgb_set_color", 0, "rgb set color ,example: cli --rgb_set_color -d '{\"rgb_index\":0,\"rgb_color\":255}'", std::bind(&info_fun, std::placeholders::_1, "rgb_set_color", "")},
-            {"fan", 'f', "fan_pwm", std::bind(&info_fun, std::placeholders::_1, "fan_set_pwm", "fan_get_pwm")},
-            {"fanspeed", 'F', "fan_speed", std::bind(&info_fun, std::placeholders::_1, "", "fan_get_speed")},
-            {"board", 'B', "board_get_power_info", std::bind(&info_fun, std::placeholders::_1, "", "board_get_power_info")},
-            {"ip", 'i', "ip fun", std::bind(&info_fun, std::placeholders::_1, "ip_set", "ip_get")},
-            {"lcd", 'l', "lcd fun", std::bind(&info_fun, std::placeholders::_1, "lcd_set_mode", "lcd_get_mode")},
-            {"lcd_ram", 0, "lcd_ram", std::bind(&info_fun, std::placeholders::_1, "lcd_set_ram", "")},
-            {"lcd_echo", 'E', "lcd_echo", std::bind(&info_fun, std::placeholders::_1, "lcd_echo", "")},
-            {"vddcpu", 'c', "vddcpu fun", std::bind(&info_fun, std::placeholders::_1, "vddcpu_set", "vddcpu_get")},
-            {"modbus_speed", 0, "modbus_speed", std::bind(&info_fun, std::placeholders::_1, "modbus_set_speed", "modbus_get_speed")},
-            {"ext_power", 'p', "ext_power fun", std::bind(&info_fun, std::placeholders::_1, "ext_power", "")},
-            {"board_power", 'b', "board_power fun", std::bind(&info_fun, std::placeholders::_1, "board_power", "")},
-            {"pcie0", 0, "pcie0_set_switch fun", std::bind(&info_fun, std::placeholders::_1, "pcie0_set_switch", "")},
-            {"pcie1", 0, "pcie1_set_switch fun", std::bind(&info_fun, std::placeholders::_1, "pcie1_set_switch", "")},
-            {"gl3510_reset", 0, "gl3510_reset fun", std::bind(&info_fun, std::placeholders::_1, "gl3510_reset", "")},
-            {"usbds1_big", 0, "usbds1_set_big_power fun", std::bind(&info_fun, std::placeholders::_1, "usbds1_set_big_power", "")},
-            {"usbds2_big", 0, "usbds2_set_big_power fun", std::bind(&info_fun, std::placeholders::_1, "usbds2_set_big_power", "")},
-            {"usbds1", 0, "usbds1_set_switch fun", std::bind(&info_fun, std::placeholders::_1, "usbds1_set_switch", "")},
-            {"usbds2", 0, "usbds2_set_switch fun", std::bind(&info_fun, std::placeholders::_1, "usbds2_set_switch", "")},
-            {"usbds3", 0, "usbds3_set_switch fun", std::bind(&info_fun, std::placeholders::_1, "usbds3_set_switch", "")},
-            {"grove_uart", 0, "grove_uart_set_switch fun", std::bind(&info_fun, std::placeholders::_1, "grove_uart_set_switch", "")},
-            {"grove_iic", 0, "grove_iic_set_switch fun", std::bind(&info_fun, std::placeholders::_1, "grove_iic_set_switch", "")},
-            {"flash_switch", 0, "flash_save_switch fun", std::bind(&info_fun, std::placeholders::_1, "flash_save_switch", "")},
-            {"flash_value", 0, "flash_save_value_config fun", std::bind(&info_fun, std::placeholders::_1, "flash_save_value_config", "")},
-            {"poweroff", 0, "poweroff fun", std::bind(&info_fun, std::placeholders::_1, "poweroff", "")},
-
-        };
-        // clang-format on
-        for (auto &cmd : cmd_list) {
-            a.add(cmd.fun, cmd.flage, cmd.dec);
-        }
-        a.add<std::string>("data", 'd', "call param", false);
-        a.add<std::string>("DataRaw", 'D', "call param raw", false);
-        a.parse_check(argc, argv);
-        for (auto &cmd : cmd_list) {
-            if (a.exist(cmd.fun)) {
-                cmd.fun_call(a);
-            }
-        }
-    } else if (std::string(argv[1]) == "exec") {
-        cmdline::parser a;
-        a.add("list", 'l', "list call function");
-        a.add<std::string>("fun", 'f', "call ax650c_ec_proxy function", false);
-        a.add<std::string>("data", 'd', "call ax650c_ec_proxy function input data", false);
-        a.parse_check(argc, argv);
-        exec_fun(a);
-    } else if (std::string(argv[1]) == "echo") {
-        cmdline::parser a;
-        a.add("button", 'b', "button event");
-        a.parse_check(argc, argv);
-        echo_fun(a);
-        std::cout << a.usage() << std::endl;
-        return 0;
-    } else {
-        goto error_print_exit;
-    }
-
-    return 0;
-}
diff --git a/projects/llm_framework/main_ax650_ec_proxy/src/ec_proxy_main.cpp b/projects/llm_framework/main_ax650_ec_proxy/src/ec_proxy_main.cpp
deleted file mode 100644
index 38c40fdb..00000000
--- a/projects/llm_framework/main_ax650_ec_proxy/src/ec_proxy_main.cpp
+++ /dev/null
@@ -1,963 +0,0 @@
-/*
- * SPDX-FileCopyrightText: 2024 M5Stack Technology CO LTD
- *
- * SPDX-License-Identifier: MIT
- */
-#include "StackFlow.h"
-#include <signal.h>
-#include <sys/stat.h>
-#include <sys/types.h>
-#include <unistd.h>
-#include <base64.h>
-#include <fstream>
-#include <iostream>
-#include <stdexcept>
-#include <modbus.h>
-#include <arpa/inet.h>
-// #include "cmdline.hpp"
-#include <../../static_lib/include/fmt/core.h>
-#include <../../static_lib/include/fmt/format.h>
-#include "../../../../SDK/components/utilities/include/sample_log.h"
-
-int main_exit_flage = 0;
-static void __sigint(int iSigNo)
-{
-    SLOGW("llm_ec_prox will be exit!");
-    main_exit_flage = 1;
-}
-
-#define CONFIG_AUTO_SET(obj, key)             \
-    if (config_body.contains(#key))           \
-        mode_config_.key = config_body[#key]; \
-    else if (obj.contains(#key))              \
-        mode_config_.key = obj[#key];
-
-using namespace StackFlows;
-
-class llm_ec_prox : public StackFlow {
-private:
-    enum { EVENT_LOAD_CONFIG = EVENT_EXPORT + 1, EVENT_QUEUE_PLAY };
-    std::string ec_prox_event_channel = "ipc:///tmp/llm/ec_prox.event.socket";
-    modbus_t *modbus_ctx;
-    std::mutex modbus_mtx_;
-    typedef struct {
-        int reg_index;    // 16位寄存器索引
-        int byte_offset;  // 在寄存器中字节偏移（0-低字节，1-高字节）
-    } RegPosition;
-    std::unique_ptr<pzmq> pub_ctx_;
-    std::string StackFlow_pack_temp =
-        R"format({{ "created":{},"data":{},"object":{},"request_id":{},"work_id":"ax650_ec_proxy" }})format";
-    std::string StackFlow_pack_all_temp =
-        R"format({{ "created":{},"data":{},"error":{{"code":{},"message":{}}},"object":{},"request_id":{},"work_id":"ax650_ec_proxy" }})format";
-    std::string return_success_result(std::string data, std::string object = "\"\"", std::string request_id = "\"\"")
-    {
-        return fmt::format(StackFlow_pack_temp, std::time(nullptr), data, object, request_id);
-    }
-
-    static llm_ec_prox *self;
-
-    void loop()
-    {
-        uint8_t butt = 0;
-        while (!main_exit_flage) {
-            {
-                std::unique_lock<std::mutex> lock(this->modbus_mtx_);
-                uint8_t tab_rp_bits[2];
-                int rc = modbus_read_input_bits(modbus_ctx, 4, 1, tab_rp_bits);
-                if (rc == -1) {
-                    fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
-                } else {
-                    if (tab_rp_bits[0] != butt) {
-                        if (tab_rp_bits[0]) {
-                            pub_ctx_->send_data(return_success_result("1"));
-                        } else {
-                            pub_ctx_->send_data(return_success_result("0"));
-                        }
-                        butt = tab_rp_bits[0];
-                    }
-                }
-                // std::cout << "butt:" << (int)butt << std::endl;
-            }
-            usleep(20 * 1000);
-        }
-    }
-
-    std::string board_get_power_info(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        // mbpoll -m rtu -b 115200 -P none -a 1 -r 1 -t 3 -c 13 -l 10 /dev/ttyS3
-        uint16_t tab_reg[12];
-        int rc = modbus_read_input_registers(modbus_ctx, 0, 12, tab_reg);
-        if (rc == -1) {
-            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
-        } else {
-            return return_success_result(fmt::format(
-                R"format({{ "pcie0_mv": {} , "pcie0_ma": {} , "pcie1_mv": {} , "pcie1_ma": {} , "usb1_mv": {} , "usb1_ma": {} , "usb2_mv": {} , "usb2_ma": {} , "VDD5V_mv": {} , "VDD5V_ma": {} , "EXT5V_mv": {} , "EXT5V_ma": {} }})format",
-                tab_reg[0], tab_reg[1], tab_reg[2], tab_reg[3], tab_reg[4], tab_reg[5], tab_reg[6], tab_reg[7],
-                tab_reg[8], tab_reg[9], tab_reg[10], tab_reg[11]));
-        }
-        return std::string("Error");
-    }
-
-    std::string fan_get_speed(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        // mbpoll -m rtu -b 115200 -P none -a 1 -r 1 -t 3 -c 13 -l 10 /dev/ttyS3
-        uint16_t tab_reg[2];
-        int rc = modbus_read_input_registers(modbus_ctx, 12, 1, tab_reg);
-        if (rc == -1) {
-            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
-        } else {
-            return return_success_result(fmt::format("{}", tab_reg[0]));
-        }
-        return std::string("Error");
-    }
-    std::string fan_get_pwm(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2];
-        int rc = modbus_read_registers(modbus_ctx, 1, 1, tab_reg);
-        if (rc == -1) {
-            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
-        } else {
-            return return_success_result(fmt::format("{}", tab_reg[0]));
-        }
-        return std::string("Error");
-    }
-    std::string fan_set_pwm(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2];
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_register(modbus_ctx, 1, tab_reg[0]);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-        }
-        return return_success_result("\"ok\"");
-    }
-
-    RegPosition get_rgb_reg_pos(int rgb_index)
-    {
-        RegPosition res;
-        int offset      = rgb_index * 3;  // 每个RGB占3字节
-        res.reg_index   = offset / 2;
-        res.byte_offset = offset % 2;
-        return res;
-    }
-    std::string rgb_get_mode(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2];
-        int rc = modbus_read_registers(modbus_ctx, 11, 1, tab_reg);
-        if (rc == -1) {
-            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
-        } else {
-            return return_success_result(fmt::format("{}", tab_reg[0]));
-        }
-        return std::string("Error");
-    }
-    std::string rgb_set_mode(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2];
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_register(modbus_ctx, 11, tab_reg[0]);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-    std::string rgb_get_size(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2];
-        int rc = modbus_read_registers(modbus_ctx, 10, 1, tab_reg);
-        if (rc == -1) {
-            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
-        } else {
-            return return_success_result(fmt::format("{}", tab_reg[0]));
-        }
-        return std::string("Error");
-    }
-    std::string rgb_set_size(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2];
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_register(modbus_ctx, 10, tab_reg[0]);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-    std::string rgb_get_color(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        int rgb_index = 0;
-        int rgb_color = 0;
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            rgb_index               = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        uint16_t tab_reg[2];
-        RegPosition reg_pos = get_rgb_reg_pos(rgb_index);
-        {
-            int rc = modbus_read_registers(modbus_ctx, reg_pos.reg_index + 12, 2, tab_reg);
-            if (rc == -1) {
-                fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
-            } else {
-                for (int i = 0; i < 2; i++) tab_reg[i] = ntohs(tab_reg[i]);
-                uint8_t *rgb_color_reg = (uint8_t *)tab_reg;
-                rgb_color              = rgb_color_reg[reg_pos.byte_offset] << 16;
-                rgb_color |= rgb_color_reg[reg_pos.byte_offset + 1] << 8;
-                rgb_color |= rgb_color_reg[reg_pos.byte_offset + 2];
-                return return_success_result(fmt::format("{}", rgb_color));
-            }
-        }
-        return std::string("Error");
-    }
-    std::string rgb_set_color(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        int rgb_index = 0;
-        int rgb_color = 0;
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            rgb_index               = (int)fan_data["data"]["rgb_index"];
-            rgb_color               = (int)fan_data["data"]["rgb_color"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        uint16_t tab_reg[2];
-        RegPosition reg_pos = get_rgb_reg_pos(rgb_index);
-        {
-            int rc = modbus_read_registers(modbus_ctx, reg_pos.reg_index + 12, 2, tab_reg);
-            if (rc == -1) {
-                fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
-            }
-            for (int i = 0; i < 2; i++) tab_reg[i] = ntohs(tab_reg[i]);
-        }
-        {
-            uint8_t *rgb_color_reg                 = (uint8_t *)tab_reg;
-            rgb_color_reg[reg_pos.byte_offset]     = (rgb_color & 0xff0000) >> 16;
-            rgb_color_reg[reg_pos.byte_offset + 1] = (rgb_color & 0xff00) >> 8;
-            rgb_color_reg[reg_pos.byte_offset + 2] = (rgb_color & 0xff);
-        }
-        for (int i = 0; i < 2; i++) tab_reg[i] = htons(tab_reg[i]);
-        {
-            int rc = modbus_write_registers(modbus_ctx, reg_pos.reg_index + 12, 2, tab_reg);
-            if (rc == -1) {
-                fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-                return std::string("Error");
-            }
-        }
-        return return_success_result("\"ok\"");
-    }
-
-    std::string lcd_get_mode(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2];
-        int rc = modbus_read_registers(modbus_ctx, 86, 1, tab_reg);
-        if (rc == -1) {
-            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
-        } else {
-            return return_success_result(fmt::format("{}", tab_reg[0]));
-        }
-        return std::string("Error");
-    }
-    std::string lcd_set_mode(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2];
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_register(modbus_ctx, 86, tab_reg[0]);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-    std::string lcd_set_ram(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        std::string rgb_ram_base64;
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            rgb_ram_base64          = fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        uint16_t *rgb_ram = (uint16_t *)malloc(((BASE64_DECODE_OUT_SIZE(rgb_ram_base64.length()) / 2) + 1) * 2);
-        int ram_size      = base64_decode(rgb_ram_base64.c_str(), rgb_ram_base64.length(), (uint8_t *)rgb_ram);
-        ram_size          = ram_size > 1870 ? 1870 : ram_size;
-        int rc            = modbus_write_registers(modbus_ctx, 88, ram_size / 2, rgb_ram);
-        free(rgb_ram);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return std::string("Ok");
-    }
-
-    std::string lcd_echo(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        std::string put_strs;
-        uint16_t tab_reg = 0;
-        uint8_t *c_reg   = (uint8_t *)&tab_reg;
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            put_strs                = fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        auto it = put_strs.begin();
-        while (it != put_strs.end()) {
-            c_reg[0] = *it;
-            ++it;
-            if (it != put_strs.end()) {
-                c_reg[1] = *it;
-                ++it;
-            } else {
-                c_reg[1] = 0;
-            }
-            int rc = modbus_write_register(modbus_ctx, 87, tab_reg);
-            if (rc == -1) {
-                fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-                return std::string("Error");
-            }
-        }
-        return return_success_result("\"ok\"");
-    }
-
-    std::string ip_get(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint32_t value      = 0;
-        uint16_t tab_reg[2] = {0};
-
-        int rc = modbus_read_registers(modbus_ctx, 5, 2, tab_reg);
-        if (rc == -1) {
-            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
-        } else {
-            for (int i = 0; i < 2; i++) tab_reg[i] = ntohs(tab_reg[i]);
-            value = ((uint32_t)tab_reg[0] << 16) | tab_reg[1];
-            char ip_str[INET_ADDRSTRLEN];
-            inet_ntop(AF_INET, &value, ip_str, INET_ADDRSTRLEN);
-            return return_success_result(fmt::format("\"{}\"", ip_str));
-        }
-        return std::string("Error");
-    }
-    std::string ip_set(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint32_t value;
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            std::string ip_str      = fan_data["data"];
-            if (inet_pton(AF_INET, ip_str.c_str(), &value) != 1) {
-                return std::string("Error : 无效的 IP 地址\n");
-            }
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        uint16_t tab_reg[2];
-        tab_reg[0] = (value >> 16) & 0xFFFF;
-        tab_reg[1] = value & 0xFFFF;
-        for (int i = 0; i < 2; i++) tab_reg[i] = htons(tab_reg[i]);
-        int rc = modbus_write_registers(modbus_ctx, 5, 2, tab_reg);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-
-    std::string vddcpu_get(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-
-        int rc = modbus_read_registers(modbus_ctx, 2, 1, tab_reg);
-        if (rc == -1) {
-            fprintf(stderr, "读取失败: %s\n", modbus_strerror(errno));
-        } else {
-            return return_success_result(fmt::format("{}", tab_reg[0]));
-        }
-        return std::string("Error");
-    }
-    std::string vddcpu_set(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_registers(modbus_ctx, 2, 1, tab_reg);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-    std::string modbus_set_speed(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        int modbus_speed;
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            modbus_speed            = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        set_modbus_speed(modbus_speed);
-        return return_success_result("\"ok\"");
-    }
-    std::string modbus_get_speed(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        return return_success_result(fmt::format("{}", _modbus_speed));
-    }
-    void set_modbus_speed(int modbus_speed)
-    {
-        uint16_t tab_reg[2];
-        tab_reg[0] = (modbus_speed >> 16) & 0xFFFF;
-        tab_reg[1] = modbus_speed & 0xFFFF;
-        int rc     = modbus_write_registers(modbus_ctx, 3, 2, tab_reg);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            std::exit(-1);
-        }
-        modbus_close(modbus_ctx);
-        modbus_free(modbus_ctx);
-        modbus_ctx = NULL;
-        usleep(100 * 1000);
-        modbus_ctx = modbus_new_rtu("/dev/ttyS3", modbus_speed, 'N', 8, 1);
-        if (modbus_ctx == NULL) {
-            fprintf(stderr, "Unable to create the context\n");
-            std::exit(-1);
-        }
-        modbus_set_slave(modbus_ctx, 1);  // 设置 Modbus 从站地址为1
-        if (modbus_connect(modbus_ctx) == -1) {
-            fprintf(stderr, "Connection failed: %s\n", modbus_strerror(errno));
-            modbus_free(modbus_ctx);
-            std::exit(-1);
-        }
-        _modbus_speed = modbus_speed;
-    }
-    int _modbus_speed = 115200;
-
-    void _init_modbus()
-    {
-        modbus_ctx = modbus_new_rtu("/dev/ttyS3", 115200, 'N', 8, 1);
-        if (modbus_ctx == NULL) {
-            fprintf(stderr, "Unable to create the context\n");
-            std::exit(-1);
-        }
-        modbus_set_slave(modbus_ctx, 1);  // 设置 Modbus 从站地址为1
-        if (modbus_connect(modbus_ctx) == -1) {
-            fprintf(stderr, "Connection failed: %s\n", modbus_strerror(errno));
-            modbus_free(modbus_ctx);
-            std::exit(-1);
-        }
-        uint8_t tab_rp_bits[2];
-        int rc = modbus_read_input_bits(modbus_ctx, 4, 1, tab_rp_bits);
-        if (rc == -1) {
-            _modbus_speed = 1152000;
-            modbus_ctx    = modbus_new_rtu("/dev/ttyS3", _modbus_speed, 'N', 8, 1);
-            if (modbus_ctx == NULL) {
-                fprintf(stderr, "Unable to create the context\n");
-                std::exit(-1);
-            }
-            modbus_set_slave(modbus_ctx, 1);  // 设置 Modbus 从站地址为1
-            if (modbus_connect(modbus_ctx) == -1) {
-                fprintf(stderr, "Connection failed: %s\n", modbus_strerror(errno));
-                modbus_free(modbus_ctx);
-                std::exit(-1);
-            }
-            uint8_t tab_rp_bits[2];
-            int rc = modbus_read_input_bits(modbus_ctx, 4, 1, tab_rp_bits);
-            if (rc == -1) {
-                fprintf(stderr, "modbus 115200/1152000 速度探测失败: %s\n", modbus_strerror(errno));
-                std::exit(-1);
-            }
-        } else {
-            set_modbus_speed(1152000);
-        }
-    }
-
-    std::string ext_poweroff(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_bit(modbus_ctx, 0, 1);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-    std::string board_poweroff(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_bit(modbus_ctx, 1, 1);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-
-    std::string pcie0_set_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_bit(modbus_ctx, 4, 1);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-    std::string pcie1_set_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_bit(modbus_ctx, 5, 1);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-
-    std::string gl3510_reset(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_bit(modbus_ctx, 6, 1);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-    std::string usbds1_set_big_power(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_bit(modbus_ctx, 7, 1);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-    std::string usbds2_set_big_power(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_bit(modbus_ctx, 8, 1);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-    std::string usbds1_set_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_bit(modbus_ctx, 9, 1);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-    std::string usbds2_set_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_bit(modbus_ctx, 10, 1);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-    std::string usbds3_set_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_bit(modbus_ctx, 11, 1);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-
-    std::string grove_uart_set_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_bit(modbus_ctx, 12, 1);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-    std::string grove_iic_set_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_bit(modbus_ctx, 13, 1);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-
-    std::string flash_save_switch(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_bit(modbus_ctx, 14, 1);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-
-    std::string flash_save_value_config(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_bit(modbus_ctx, 15, 1);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-
-    std::string poweroff(StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data)
-    {
-        uint16_t tab_reg[2] = {0};
-        try {
-            nlohmann::json fan_data = nlohmann::json::parse(data->string());
-            tab_reg[0]              = (int)fan_data["data"];
-        } catch (const nlohmann::json::parse_error &e) {
-            std::cerr << "解析失败：" << e.what() << std::endl;
-            return std::string("Error");
-        } catch (...) {
-            std::cout << "data 解析异常" << std::endl;
-            return std::string("Error");
-        }
-        int rc = modbus_write_register(modbus_ctx, 0, tab_reg[0] >= 1000 ? tab_reg[0] : 1000);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        rc = modbus_write_bit(modbus_ctx, 16, 1);
-        if (rc == -1) {
-            fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            return std::string("Error");
-        }
-        return return_success_result("\"ok\"");
-    }
-
-public:
-    llm_ec_prox() : StackFlow("ec_prox")
-    {
-        // setup("", "audio.play", "{\"None\":\"None\"}");
-        self = this;
-        // prob modbus speed
-        _init_modbus();
-
-        pub_ctx_ = std::make_unique<pzmq>(ec_prox_event_channel, ZMQ_PUB);
-
-        // clang-format off
-#define REGISTER_RPC_ACTION(name, func) \
-        rpc_ctx_->register_rpc_action(name, [this](StackFlows::pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data){ \
-            std::unique_lock<std::mutex> lock(this->modbus_mtx_); \
-            return this->func(_pzmq, data); \
-        });
-
-        REGISTER_RPC_ACTION("fan_get_speed",                fan_get_speed);
-        REGISTER_RPC_ACTION("fan_get_pwm",                  fan_get_pwm);
-        REGISTER_RPC_ACTION("fan_set_pwm",                  fan_set_pwm);
-
-        REGISTER_RPC_ACTION("rgb_get_mode",                 rgb_get_mode);
-        REGISTER_RPC_ACTION("rgb_set_mode",                 rgb_set_mode);
-        REGISTER_RPC_ACTION("rgb_get_size",                 rgb_get_size);
-        REGISTER_RPC_ACTION("rgb_set_size",                 rgb_set_size);
-        REGISTER_RPC_ACTION("rgb_get_color",                rgb_get_color);
-        REGISTER_RPC_ACTION("rgb_set_color",                rgb_set_color);
-        
-        REGISTER_RPC_ACTION("board_get_power_info",         board_get_power_info);
-
-        REGISTER_RPC_ACTION("ip_get",                       ip_get);
-        REGISTER_RPC_ACTION("ip_set",                       ip_set);
-
-        REGISTER_RPC_ACTION("lcd_get_mode",                 lcd_get_mode);
-        REGISTER_RPC_ACTION("lcd_set_mode",                 lcd_set_mode);
-        REGISTER_RPC_ACTION("lcd_set_ram",                  lcd_set_ram);
-        REGISTER_RPC_ACTION("lcd_echo",                     lcd_echo);
-        
-
-        REGISTER_RPC_ACTION("vddcpu_get",                   vddcpu_get);
-        REGISTER_RPC_ACTION("vddcpu_set",                   vddcpu_set);
-
-        REGISTER_RPC_ACTION("modbus_get_speed",             modbus_get_speed);
-        REGISTER_RPC_ACTION("modbus_set_speed",             modbus_set_speed);
-
-        REGISTER_RPC_ACTION("ext_poweroff",                   ext_poweroff);
-        REGISTER_RPC_ACTION("board_poweroff",                 board_poweroff);
-        REGISTER_RPC_ACTION("pcie0_set_switch",                 pcie0_set_switch);
-        REGISTER_RPC_ACTION("pcie1_set_switch",                 pcie1_set_switch);
-        REGISTER_RPC_ACTION("gl3510_reset",                 gl3510_reset);
-        REGISTER_RPC_ACTION("usbds1_set_big_power",                 usbds1_set_big_power);
-        REGISTER_RPC_ACTION("usbds2_set_big_power",                 usbds2_set_big_power);
-        REGISTER_RPC_ACTION("usbds1_set_switch",                 usbds1_set_switch);
-        REGISTER_RPC_ACTION("usbds2_set_switch",                 usbds2_set_switch);
-        REGISTER_RPC_ACTION("usbds3_set_switch",                 usbds3_set_switch);
-        REGISTER_RPC_ACTION("grove_uart_set_switch",                 grove_uart_set_switch);
-        REGISTER_RPC_ACTION("grove_iic_set_switch",                 grove_iic_set_switch);
-        REGISTER_RPC_ACTION("flash_save_switch",                 flash_save_switch);
-        REGISTER_RPC_ACTION("flash_save_value_config",                 flash_save_value_config);
-        REGISTER_RPC_ACTION("poweroff",                 poweroff);
-
-
-
-
-#undef REGISTER_RPC_ACTION
-        // clang-format on
-    }
-    int setup(const std::string &work_id, const std::string &object, const std::string &data) override
-    {
-        send(std::string("None"), std::string("None"), std::string(""), unit_name_);
-        return -1;
-    }
-    void ax650_ec_prox_exit()
-    {
-        pub_ctx_.reset();
-        for (int i = 0; i < 3; i++) {
-            uint32_t value = 115200;
-            uint16_t tab_reg[2];
-            tab_reg[0] = (value >> 16) & 0xFFFF;
-            tab_reg[1] = value & 0xFFFF;
-            int rc     = modbus_write_registers(modbus_ctx, 3, 2, tab_reg);
-            if (rc == -1) {
-                fprintf(stderr, "写入失败: %s\n", modbus_strerror(errno));
-            } else {
-                break;
-            }
-        }
-        modbus_close(modbus_ctx);
-        modbus_free(modbus_ctx);
-        std::cout << "all exit!" << std::endl;
-    }
-    ~llm_ec_prox()
-    {
-        ax650_ec_prox_exit();
-    }
-};
-
-llm_ec_prox *llm_ec_prox::self;
-int main(int argc, char *argv[])
-{
-    signal(SIGTERM, __sigint);
-    signal(SIGINT, __sigint);
-    mkdir("/tmp/llm", 0777);
-    llm_ec_prox ec_prox;
-    while (!main_exit_flage) {
-        sleep(1);
-    }
-    ec_prox.llm_firework_exit();
-    return 0;
-}

From e96fcf41dcb689d04cc096275f0caad20baee57d Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Tue, 9 Dec 2025 15:31:53 +0800
Subject: [PATCH 65/79] [update] llm_asr suooported sensevoice, update
 llm_audio supported alsa cap & play, update static lib version

---
 projects/llm_framework/SConstruct             |   2 +-
 projects/llm_framework/main_asr/SConstruct    |  16 +-
 .../mode_sense-voice-small-10s-ax650.json     |  30 ++
 projects/llm_framework/main_asr/src/main.cpp  | 469 +++++++++++++-----
 projects/llm_framework/main_audio/SConstruct  |   6 +-
 .../llm_framework/main_audio/src/alsa_audio.c | 108 +++-
 .../llm_framework/main_audio/src/alsa_audio.h |  20 +
 .../llm_framework/main_audio/src/main.cpp     | 136 ++---
 projects/llm_framework/main_kws/SConstruct    |   6 +-
 .../llm_framework/main_kws_new/SConstruct     |   2 +
 .../llm_framework/main_kws_new/mode_kws.json  |  41 ++
 .../llm_framework/main_kws_new/src/main.cpp   |   3 -
 projects/llm_framework/main_vad/SConstruct    |   3 +-
 13 files changed, 635 insertions(+), 207 deletions(-)
 create mode 100644 projects/llm_framework/main_asr/mode_sense-voice-small-10s-ax650.json
 create mode 100644 projects/llm_framework/main_kws_new/mode_kws.json

diff --git a/projects/llm_framework/SConstruct b/projects/llm_framework/SConstruct
index b6c85dad..731210de 100644
--- a/projects/llm_framework/SConstruct
+++ b/projects/llm_framework/SConstruct
@@ -5,7 +5,7 @@ import shutil
 os.environ['SDK_PATH'] = os.path.normpath(str(Path(os.getcwd())/'..'/'..'/'SDK'))
 os.environ['EXT_COMPONENTS_PATH'] = os.path.normpath(str(Path(os.getcwd())/'..'/'..'/'ext_components'))
 
-version = 'v0.1.2'
+version = 'v0.1.3'
 static_lib = 'static_lib'
 update = False
 
diff --git a/projects/llm_framework/main_asr/SConstruct b/projects/llm_framework/main_asr/SConstruct
index cbc98e40..9bdac46a 100644
--- a/projects/llm_framework/main_asr/SConstruct
+++ b/projects/llm_framework/main_asr/SConstruct
@@ -7,7 +7,7 @@ with open(env['PROJECT_TOOL_S']) as f:
 SRCS = Glob('src/*.c*')
 INCLUDE = [ADir('include'), ADir('.')]
 PRIVATE_INCLUDE = []
-REQUIREMENTS = ['pthread', 'utilities', 'eventpp', 'StackFlow', 'single_header_libs']
+REQUIREMENTS = ['pthread', 'utilities', 'ax_msp', 'eventpp', 'StackFlow', 'single_header_libs']
 STATIC_LIB = []
 DYNAMIC_LIB = []
 DEFINITIONS = []
@@ -17,12 +17,22 @@ LINK_SEARCH_PATH = []
 STATIC_FILES = []
 
 LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
-DEFINITIONS += ['-std=c++17']
+DEFINITIONS += ['-std=c++17', '-fopenmp']
 LINK_SEARCH_PATH += [ADir('../static_lib')]
+REQUIREMENTS += ['ax_engine', 'ax_interpreter', 'ax_sys']
 
 INCLUDE += [ADir('../static_lib/include/sherpa'), ADir('../static_lib/include/sherpa/sherpa-ncnn')]
 LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/ncnn')]
-REQUIREMENTS += ['ncnn', 'sherpa-ncnn-core']
+LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/onnx')]
+REQUIREMENTS += ['ncnn', '', 'onnxruntime']
+LDFLAGS += [
+            '-l:libsherpa-onnx-core.a', '-l:libkaldi-native-fbank-core.a','-l:libkissfft-float.a',
+            '-l:libkaldi-decoder-core.a', '-l:libssentencepiece_core.a', '-l:libsherpa-onnx-fst.a',
+            '-l:libsherpa-onnx-kaldifst-core.a', '-l:libsherpa-onnx-fstfar.a',
+
+            '-l:libsherpa-ncnn-core.a', '-l:libsherpa-ncnn-fst.a', '-l:libsherpa-ncnn-kaldifst-core.a',
+            '-lgomp',
+            ]
 
 STATIC_FILES += Glob('mode_*.json')
 
diff --git a/projects/llm_framework/main_asr/mode_sense-voice-small-10s-ax650.json b/projects/llm_framework/main_asr/mode_sense-voice-small-10s-ax650.json
new file mode 100644
index 00000000..370c8346
--- /dev/null
+++ b/projects/llm_framework/main_asr/mode_sense-voice-small-10s-ax650.json
@@ -0,0 +1,30 @@
+{
+    "mode": "sense-voice-small-10s-ax650",
+    "type": "asr",
+    "homepage": "https://huggingface.co/yunyu1258/qwen2.5-0.5b-ha",
+    "compile_flage": "pulsar2 build --input model-10-seconds.onnx --config config_sensevoice_main_u16.json --output_dir sensevoice-axmodel --output_name model-10-seconds.axmodel --target_hardware AX650 --compiler.check 0",
+    "pulsar_version": "5.0-patch1-fd447d0d",
+    "capabilities": [
+        "Chinese",
+        "English",
+        "Cantonese",
+        "Japanese",
+        "Korean"
+    ],
+    "input_type": [
+        "sys.pcm",
+        "sys.cap.0_0"
+    ],
+    "output_type": [
+        "asr.utf-8"
+    ],
+    "mode_param": {
+        "model_config.sense_voice.model": "model.axmodel",
+        "model_config.tokens": "tokens.txt",
+        "silero_vad.model": "silero_vad.onnx",
+        "model_config.provider": "axera",
+        "silence_timeout": 2000,
+        "awake_delay": 50
+    },
+    "mode_param_bak": {}
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_asr/src/main.cpp b/projects/llm_framework/main_asr/src/main.cpp
index d2dff9ba..19fb3363 100644
--- a/projects/llm_framework/main_asr/src/main.cpp
+++ b/projects/llm_framework/main_asr/src/main.cpp
@@ -5,6 +5,8 @@
  */
 #include "StackFlow.h"
 #include "sherpa-ncnn/csrc/recognizer.h"
+#include "sherpa-onnx/csrc/offline-recognizer.h"
+#include "sherpa-onnx/csrc/voice-activity-detector.h"
 
 #include <iostream>
 #include <signal.h>
@@ -24,6 +26,7 @@
 using namespace StackFlows;
 
 int main_exit_flage = 0;
+
 static void __sigint(int iSigNo)
 {
     SLOGW("llm_asr will be exit!");
@@ -35,17 +38,39 @@ static std::string base_model_config_path_;
 
 typedef std::function<void(const std::string &data, bool finish)> task_callback_t;
 
-#define CONFIG_AUTO_SET(obj, key)             \
+#define NCNN_ASR_CONFIG_AUTO_SET(obj, key)    \
     if (config_body.contains(#key))           \
-        mode_config_.key = config_body[#key]; \
+        ncnn_config_.key = config_body[#key]; \
     else if (obj.contains(#key))              \
-        mode_config_.key = obj[#key];
+        ncnn_config_.key = obj[#key];
+
+#define ONNX_ASR_CONFIG_AUTO_SET(obj, key)        \
+    if (config_body.contains(#key))               \
+        onnx_asr_config_.key = config_body[#key]; \
+    else if (obj.contains(#key))                  \
+        onnx_asr_config_.key = obj[#key];
+
+#define ONNX_VAD_CONFIG_AUTO_SET(obj, key)   \
+    if (config_body.contains(#key))          \
+        vad_config_.key = config_body[#key]; \
+    else if (obj.contains(#key))             \
+        vad_config_.key = obj[#key];
 
 class llm_task {
 private:
-    sherpa_ncnn::RecognizerConfig mode_config_;
-    std::unique_ptr<sherpa_ncnn::Recognizer> recognizer_;
-    std::unique_ptr<sherpa_ncnn::Stream> recognizer_stream_;
+    sherpa_ncnn::RecognizerConfig ncnn_config_;
+    std::unique_ptr<sherpa_ncnn::Recognizer> ncnn_recognizer_;
+    std::unique_ptr<sherpa_ncnn::Stream> ncnn_stream_;
+
+    sherpa_onnx::OfflineRecognizerConfig onnx_asr_config_;
+    sherpa_onnx::VadModelConfig vad_config_;
+    std::unique_ptr<sherpa_onnx::OfflineRecognizer> onnx_recognizer_;
+    std::unique_ptr<sherpa_onnx::VoiceActivityDetector> vad_;
+
+    enum EngineType {
+        ENGINE_NCNN = 0,
+        ENGINE_ONNX = 1,
+    } engine_type_ = ENGINE_NCNN;
 
 public:
     std::string model_;
@@ -57,10 +82,12 @@ class llm_task {
     task_callback_t out_callback_;
     std::atomic_bool audio_flage_;
     std::atomic_bool awake_flage_;
-    int awake_delay_       = 50;
-    int delay_audio_frame_ = 10;
-    buffer_t *pcmdata;
+    int awake_delay_        = 50;
+    int delay_audio_frame_  = 10;
+    float silence_ms_accum_ = 0.0f;
+    float silence_timeout   = 1000.0f;
 
+    buffer_t *pcmdata;
     std::function<void(void)> pause;
 
     bool parse_config(const nlohmann::json &config_body)
@@ -82,18 +109,208 @@ class llm_task {
             SLOGE("setup config_body error");
             return true;
         }
+
+        if (model_.rfind("sherpa-ncnn", 0) == 0) {
+            engine_type_ = ENGINE_NCNN;
+        } else {
+            engine_type_ = ENGINE_ONNX;
+        }
+
         enstream_ = response_format_.find("stream") == std::string::npos ? false : true;
         return false;
     }
 
+    int load_ncnn_model(const nlohmann::json &config_body, const nlohmann::json &file_body)
+    {
+        std::string base_model = base_model_path_ + model_ + "/";
+        SLOGI("base_model (ncnn) %s", base_model.c_str());
+
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.sampling_rate);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.feature_dim);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.encoder_param);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.encoder_bin);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.decoder_param);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.decoder_bin);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.joiner_param);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.joiner_bin);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.tokens);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.encoder_opt.num_threads);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.decoder_opt.num_threads);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.joiner_opt.num_threads);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], decoder_config.method);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], decoder_config.num_active_paths);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule1.must_contain_nonsilence);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule1.min_trailing_silence);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule1.min_utterance_length);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule2.must_contain_nonsilence);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule2.min_trailing_silence);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule2.min_utterance_length);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule3.must_contain_nonsilence);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule3.min_trailing_silence);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule3.min_utterance_length);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], enable_endpoint);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], hotwords_file);
+        NCNN_ASR_CONFIG_AUTO_SET(file_body["mode_param"], hotwords_score);
+
+        if (config_body.contains("awake_delay"))
+            awake_delay_ = config_body["awake_delay"].get<int>();
+        else if (file_body["mode_param"].contains("awake_delay"))
+            awake_delay_ = file_body["mode_param"]["awake_delay"];
+
+        if (config_body.contains("rule1")) {
+            ncnn_config_.endpoint_config.rule1.min_trailing_silence = config_body["rule1"].get<float>();
+            ncnn_config_.endpoint_config.rule1.must_contain_nonsilence =
+                (ncnn_config_.endpoint_config.rule1.min_trailing_silence == 0.0f) ? false : true;
+        }
+        if (config_body.contains("rule2")) {
+            ncnn_config_.endpoint_config.rule2.min_trailing_silence = config_body["rule2"].get<float>();
+            ncnn_config_.endpoint_config.rule2.must_contain_nonsilence =
+                (ncnn_config_.endpoint_config.rule2.min_trailing_silence == 0.0f) ? false : true;
+        }
+        if (config_body.contains("rule3")) {
+            ncnn_config_.endpoint_config.rule3.min_utterance_length = config_body["rule3"].get<float>();
+            ncnn_config_.endpoint_config.rule3.must_contain_nonsilence =
+                (ncnn_config_.endpoint_config.rule3.min_utterance_length == 0.0f) ? false : true;
+        }
+
+        ncnn_config_.model_config.tokens        = base_model + ncnn_config_.model_config.tokens;
+        ncnn_config_.model_config.encoder_param = base_model + ncnn_config_.model_config.encoder_param;
+        ncnn_config_.model_config.encoder_bin   = base_model + ncnn_config_.model_config.encoder_bin;
+        ncnn_config_.model_config.decoder_param = base_model + ncnn_config_.model_config.decoder_param;
+        ncnn_config_.model_config.decoder_bin   = base_model + ncnn_config_.model_config.decoder_bin;
+        ncnn_config_.model_config.joiner_param  = base_model + ncnn_config_.model_config.joiner_param;
+        ncnn_config_.model_config.joiner_bin    = base_model + ncnn_config_.model_config.joiner_bin;
+
+        ncnn_recognizer_ = std::make_unique<sherpa_ncnn::Recognizer>(ncnn_config_);
+        return 0;
+    }
+
+    int load_onnx_model(const nlohmann::json &config_body, const nlohmann::json &file_body)
+    {
+        std::string base_model = base_model_path_ + model_ + "/";
+        SLOGI("base_model (onnx) %s", base_model.c_str());
+
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.sampling_rate);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.feature_dim);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.low_freq);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.dither);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.normalize_samples);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.snip_edges);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.frame_shift_ms);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.frame_length_ms);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.is_librosa);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.remove_dc_offset);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.preemph_coeff);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.window_type);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.nemo_normalize_type);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.num_ceps);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.use_energy);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.is_mfcc);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.is_whisper);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.is_t_one);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.round_to_power_of_two);
+
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.transducer.encoder_filename);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.transducer.decoder_filename);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.transducer.joiner_filename);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.paraformer.model);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.nemo_ctc.model);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.whisper.encoder);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.whisper.decoder);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.whisper.language);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.whisper.task);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.whisper.tail_paddings);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.fire_red_asr.encoder);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.fire_red_asr.decoder);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.tdnn.model);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.zipformer_ctc.model);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.wenet_ctc.model);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.sense_voice.model);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.sense_voice.language);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.sense_voice.use_itn);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.moonshine.preprocessor);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.moonshine.encoder);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.moonshine.uncached_decoder);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.moonshine.cached_decoder);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.dolphin.model);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.canary.encoder);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.canary.decoder);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.canary.src_lang);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.canary.tgt_lang);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.canary.use_pnc);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.omnilingual.model);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.telespeech_ctc);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.tokens);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.num_threads);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.debug);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.provider);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.model_type);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.modeling_unit);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], model_config.bpe_vocab);
+
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], lm_config.model);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], lm_config.scale);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], lm_config.lm_num_threads);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], lm_config.lm_provider);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], lm_config.lodr_fst);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], lm_config.lodr_scale);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], lm_config.lodr_backoff_id);
+
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], ctc_fst_decoder_config.graph);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], ctc_fst_decoder_config.max_active);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], decoding_method);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], max_active_paths);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], hotwords_file);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], hotwords_score);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], blank_penalty);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], rule_fsts);
+        ONNX_ASR_CONFIG_AUTO_SET(file_body["mode_param"], rule_fars);
+
+        ONNX_VAD_CONFIG_AUTO_SET(file_body["mode_param"], silero_vad.model);
+        ONNX_VAD_CONFIG_AUTO_SET(file_body["mode_param"], silero_vad.threshold);
+        ONNX_VAD_CONFIG_AUTO_SET(file_body["mode_param"], silero_vad.min_silence_duration);
+        ONNX_VAD_CONFIG_AUTO_SET(file_body["mode_param"], silero_vad.min_speech_duration);
+        ONNX_VAD_CONFIG_AUTO_SET(file_body["mode_param"], silero_vad.window_size);
+        ONNX_VAD_CONFIG_AUTO_SET(file_body["mode_param"], silero_vad.max_speech_duration);
+        ONNX_VAD_CONFIG_AUTO_SET(file_body["mode_param"], ten_vad.model);
+        ONNX_VAD_CONFIG_AUTO_SET(file_body["mode_param"], ten_vad.threshold);
+        ONNX_VAD_CONFIG_AUTO_SET(file_body["mode_param"], ten_vad.min_silence_duration);
+        ONNX_VAD_CONFIG_AUTO_SET(file_body["mode_param"], ten_vad.min_speech_duration);
+        ONNX_VAD_CONFIG_AUTO_SET(file_body["mode_param"], ten_vad.window_size);
+        ONNX_VAD_CONFIG_AUTO_SET(file_body["mode_param"], ten_vad.max_speech_duration);
+        ONNX_VAD_CONFIG_AUTO_SET(file_body["mode_param"], sample_rate);
+        ONNX_VAD_CONFIG_AUTO_SET(file_body["mode_param"], num_threads);
+        ONNX_VAD_CONFIG_AUTO_SET(file_body["mode_param"], debug);
+
+        if (config_body.contains("awake_delay"))
+            awake_delay_ = config_body["awake_delay"].get<int>();
+        else if (file_body["mode_param"].contains("awake_delay"))
+            awake_delay_ = file_body["mode_param"]["awake_delay"];
+
+        if (config_body.contains("silence_timeout"))
+            silence_timeout = config_body["silence_timeout"].get<int>();
+        else if (file_body["mode_param"].contains("silence_timeout"))
+            silence_timeout = file_body["mode_param"]["silence_timeout"];
+
+        onnx_asr_config_.model_config.sense_voice.model = base_model + onnx_asr_config_.model_config.sense_voice.model;
+        onnx_asr_config_.model_config.tokens            = base_model + onnx_asr_config_.model_config.tokens;
+        vad_config_.silero_vad.model                    = base_model + vad_config_.silero_vad.model;
+
+        onnx_recognizer_ = std::make_unique<sherpa_onnx::OfflineRecognizer>(onnx_asr_config_);
+        vad_             = std::make_unique<sherpa_onnx::VoiceActivityDetector>(vad_config_);
+        return 0;
+    }
+
     int load_model(const nlohmann::json &config_body)
     {
         if (parse_config(config_body)) {
             return -1;
         }
+
         nlohmann::json file_body;
         std::list<std::string> config_file_paths =
             get_config_file_paths(base_model_path_, base_model_config_path_, model_);
+
         try {
             for (auto file_name : config_file_paths) {
                 std::ifstream config_file(file_name);
@@ -110,68 +327,16 @@ class llm_task {
                 SLOGE("all config file miss");
                 return -2;
             }
-            std::string base_model = base_model_path_ + model_ + "/";
-            SLOGI("base_model %s", base_model.c_str());
-
-            CONFIG_AUTO_SET(file_body["mode_param"], feat_config.sampling_rate);
-            CONFIG_AUTO_SET(file_body["mode_param"], feat_config.feature_dim);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.encoder_param);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.encoder_bin);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.decoder_param);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.decoder_bin);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.joiner_param);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.joiner_bin);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.tokens);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.encoder_opt.num_threads);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.decoder_opt.num_threads);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.joiner_opt.num_threads);
-            CONFIG_AUTO_SET(file_body["mode_param"], decoder_config.method);
-            CONFIG_AUTO_SET(file_body["mode_param"], decoder_config.num_active_paths);
-            CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule1.must_contain_nonsilence);
-            CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule1.min_trailing_silence);
-            CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule1.min_utterance_length);
-            CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule2.must_contain_nonsilence);
-            CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule2.min_trailing_silence);
-            CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule2.min_utterance_length);
-            CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule3.must_contain_nonsilence);
-            CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule3.min_trailing_silence);
-            CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule3.min_utterance_length);
-            CONFIG_AUTO_SET(file_body["mode_param"], enable_endpoint);
-            CONFIG_AUTO_SET(file_body["mode_param"], hotwords_file);
-            CONFIG_AUTO_SET(file_body["mode_param"], hotwords_score);
-            if (config_body.contains("awake_delay"))
-                awake_delay_ = config_body["awake_delay"].get<int>();
-            else if (file_body["mode_param"].contains("awake_delay"))
-                awake_delay_ = file_body["mode_param"]["awake_delay"];
-            if (config_body.contains("rule1")) {
-                mode_config_.endpoint_config.rule1.min_trailing_silence = config_body["rule1"].get<float>();
-                mode_config_.endpoint_config.rule1.must_contain_nonsilence =
-                    (mode_config_.endpoint_config.rule1.min_trailing_silence == 0.0f) ? false : true;
-            }
-            if (config_body.contains("rule2")) {
-                mode_config_.endpoint_config.rule2.min_trailing_silence = config_body["rule2"].get<float>();
-                mode_config_.endpoint_config.rule2.must_contain_nonsilence =
-                    (mode_config_.endpoint_config.rule2.min_trailing_silence == 0.0f) ? false : true;
-            }
-            if (config_body.contains("rule3")) {
-                mode_config_.endpoint_config.rule3.min_utterance_length = config_body["rule3"].get<float>();
-                mode_config_.endpoint_config.rule3.must_contain_nonsilence =
-                    (mode_config_.endpoint_config.rule3.min_utterance_length == 0.0f) ? false : true;
-            }
 
-            mode_config_.model_config.tokens        = base_model + mode_config_.model_config.tokens;
-            mode_config_.model_config.encoder_param = base_model + mode_config_.model_config.encoder_param;
-            mode_config_.model_config.encoder_bin   = base_model + mode_config_.model_config.encoder_bin;
-            mode_config_.model_config.decoder_param = base_model + mode_config_.model_config.decoder_param;
-            mode_config_.model_config.decoder_bin   = base_model + mode_config_.model_config.decoder_bin;
-            mode_config_.model_config.joiner_param  = base_model + mode_config_.model_config.joiner_param;
-            mode_config_.model_config.joiner_bin    = base_model + mode_config_.model_config.joiner_bin;
-            recognizer_                             = std::make_unique<sherpa_ncnn::Recognizer>(mode_config_);
+            if (engine_type_ == ENGINE_NCNN) {
+                return load_ncnn_model(config_body, file_body);
+            } else {
+                return load_onnx_model(config_body, file_body);
+            }
         } catch (...) {
             SLOGE("config false");
             return -3;
         }
-        return 0;
     }
 
     void set_output(task_callback_t out_callback)
@@ -179,7 +344,7 @@ class llm_task {
         out_callback_ = out_callback;
     }
 
-    void sys_pcm_on_data(const std::string &raw)
+    void sys_pcm_on_data_ncnn(const std::string &raw)
     {
         static int count = 0;
         if (count < delay_audio_frame_) {
@@ -191,48 +356,110 @@ class llm_task {
         buffer_position_set(pcmdata, 0);
 
         std::vector<float> floatSamples;
-        {
-            int16_t audio_val;
-            while (buffer_read_i16(pcmdata, &audio_val, 1)) {
-                float normalizedSample = static_cast<float>(audio_val) / INT16_MAX;
-                floatSamples.push_back(normalizedSample);
-            }
+        int16_t audio_val;
+        while (buffer_read_i16(pcmdata, &audio_val, 1)) {
+            float normalizedSample = static_cast<float>(audio_val) / INT16_MAX;
+            floatSamples.push_back(normalizedSample);
         }
-
         buffer_resize(pcmdata, 0);
         count = 0;
 
-        if (awake_flage_ && recognizer_stream_) {
-            recognizer_stream_.reset();
+        if (awake_flage_ && ncnn_stream_) {
+            ncnn_stream_.reset();
             awake_flage_ = false;
         }
-        if (!recognizer_stream_) {
-            recognizer_stream_ = recognizer_->CreateStream();
+        if (!ncnn_stream_) {
+            ncnn_stream_ = ncnn_recognizer_->CreateStream();
         }
-        recognizer_stream_->AcceptWaveform(mode_config_.feat_config.sampling_rate, floatSamples.data(),
-                                           floatSamples.size());
-        while (recognizer_->IsReady(recognizer_stream_.get())) {
-            recognizer_->DecodeStream(recognizer_stream_.get());
+
+        ncnn_stream_->AcceptWaveform(ncnn_config_.feat_config.sampling_rate, floatSamples.data(), floatSamples.size());
+
+        while (ncnn_recognizer_->IsReady(ncnn_stream_.get())) {
+            ncnn_recognizer_->DecodeStream(ncnn_stream_.get());
         }
-        std::string text = recognizer_->GetResult(recognizer_stream_.get()).text;
+
+        std::string text = ncnn_recognizer_->GetResult(ncnn_stream_.get()).text;
         std::string lower_text;
         lower_text.resize(text.size());
         std::transform(text.begin(), text.end(), lower_text.begin(), [](const char c) { return std::tolower(c); });
+
         if ((!lower_text.empty()) && out_callback_) out_callback_(lower_text, false);
-        bool is_endpoint = recognizer_->IsEndpoint(recognizer_stream_.get());
+
+        bool is_endpoint = ncnn_recognizer_->IsEndpoint(ncnn_stream_.get());
         if (is_endpoint) {
-            std::cout << "asr have a is_endpoint \n";
-            recognizer_stream_->Finalize();
+            ncnn_stream_->Finalize();
             if ((!lower_text.empty()) && out_callback_) {
                 out_callback_(lower_text, true);
             }
-            recognizer_stream_.reset();
+            ncnn_stream_.reset();
             if (ensleep_) {
                 if (pause) pause();
             }
         }
     }
 
+    void sys_pcm_on_data_onnx(const std::string &raw)
+    {
+        static int count = 0;
+        if (count < delay_audio_frame_) {
+            buffer_write_char(pcmdata, raw.data(), raw.length());
+            count++;
+            return;
+        }
+        buffer_write_char(pcmdata, raw.data(), raw.length());
+        buffer_position_set(pcmdata, 0);
+
+        std::vector<float> floatSamples;
+        int16_t audio_val;
+        while (buffer_read_i16(pcmdata, &audio_val, 1)) {
+            float normalizedSample = static_cast<float>(audio_val) / INT16_MAX;
+            floatSamples.push_back(normalizedSample);
+        }
+        buffer_resize(pcmdata, 0);
+        count = 0;
+
+        vad_->AcceptWaveform(floatSamples.data(), floatSamples.size());
+        while (!vad_->Empty()) {
+            const auto &segment = vad_->Front();
+            auto s              = onnx_recognizer_->CreateStream();
+            s->AcceptWaveform(onnx_asr_config_.feat_config.sampling_rate, segment.samples.data(),
+                              segment.samples.size());
+            onnx_recognizer_->DecodeStream(s.get());
+            const auto &result = s->GetResult();
+            if (!result.text.empty() && out_callback_) {
+                SLOGI("onnx-asr result: %s", result.text.c_str());
+                out_callback_(result.text, true);
+            }
+            vad_->Pop();
+        }
+
+        {
+            bool detected  = vad_->IsSpeechDetected();
+            float chunk_ms = (delay_audio_frame_ + 1) * 10.0f;
+
+            if (detected) {
+                silence_ms_accum_ = 0.0f;
+            } else {
+                silence_ms_accum_ += chunk_ms;
+            }
+            if (silence_ms_accum_ >= silence_timeout) {
+                if (ensleep_) {
+                    if (pause) pause();
+                }
+                silence_ms_accum_ = 0.0f;
+            }
+        }
+    }
+
+    void sys_pcm_on_data(const std::string &raw)
+    {
+        if (engine_type_ == ENGINE_NCNN) {
+            sys_pcm_on_data_ncnn(raw);
+        } else {
+            sys_pcm_on_data_onnx(raw);
+        }
+    }
+
     void kws_awake()
     {
         awake_flage_ = true;
@@ -240,7 +467,8 @@ class llm_task {
 
     bool delete_model()
     {
-        recognizer_.reset();
+        ncnn_recognizer_.reset();
+        onnx_recognizer_.reset();
         return true;
     }
 
@@ -254,7 +482,6 @@ class llm_task {
     void start()
     {
     }
-
     void stop()
     {
     }
@@ -266,8 +493,6 @@ class llm_task {
     }
 };
 
-#undef CONFIG_AUTO_SET
-
 class llm_asr : public StackFlow {
 public:
     enum { EVENT_LOAD_CONFIG = EVENT_EXPORT + 1, EVENT_TASK_PAUSE };
@@ -292,12 +517,14 @@ class llm_asr : public StackFlow {
         if (!(llm_task_obj && llm_channel)) {
             return;
         }
+
         std::string tmp_msg1;
         const std::string *next_data = &data;
         if (finish) {
             tmp_msg1  = data + ".";
             next_data = &tmp_msg1;
         }
+
         if (llm_channel->enstream_) {
             static int count = 0;
             nlohmann::json data_body;
@@ -317,7 +544,7 @@ class llm_asr : public StackFlow {
     {
         int post = 0;
         if (in.length() > 10)
-            for (int i = 0; i < in.length() - 4; i++) {
+            for (int i = 0; i < (int)in.length() - 4; i++) {
                 if ((in[i] == 'd') && (in[i + 1] == 'a') && (in[i + 2] == 't') && (in[i + 3] == 'a')) {
                     post = i + 8;
                     break;
@@ -329,7 +556,6 @@ class llm_asr : public StackFlow {
         } else {
             return 0;
         }
-        return 0;
     }
 
     int decode_mp3(const std::string &in, std::string &out)
@@ -350,15 +576,17 @@ class llm_asr : public StackFlow {
             send("None", "None", error_body, unit_name_);
             return;
         }
+
         std::string tmp_msg1;
         const std::string *next_data = &data;
         int ret;
+
         if (object.find("stream") != std::string::npos) {
             static std::unordered_map<int, std::string> stream_buff;
             try {
                 if (decode_stream(data, tmp_msg1, stream_buff)) {
                     return;
-                };
+                }
             } catch (...) {
                 stream_buff.clear();
                 error_body["code"]    = -25;
@@ -368,6 +596,7 @@ class llm_asr : public StackFlow {
             }
             next_data = &tmp_msg1;
         }
+
         std::string tmp_msg2;
         if (object.find("base64") != std::string::npos) {
             ret = decode_base64((*next_data), tmp_msg2);
@@ -379,6 +608,7 @@ class llm_asr : public StackFlow {
             }
             next_data = &tmp_msg2;
         }
+
         std::string tmp_msg3;
         if (object.find("wav") != std::string::npos) {
             ret = decode_wav((*next_data), tmp_msg3);
@@ -387,6 +617,7 @@ class llm_asr : public StackFlow {
             }
             next_data = &tmp_msg3;
         }
+
         std::string tmp_msg4;
         if (object.find("mp3") != std::string::npos) {
             ret = decode_mp3((*next_data), tmp_msg4);
@@ -395,6 +626,7 @@ class llm_asr : public StackFlow {
             }
             next_data = &tmp_msg4;
         }
+
         llm_task_obj->sys_pcm_on_data((*next_data));
     }
 
@@ -430,7 +662,9 @@ class llm_asr : public StackFlow {
         if ((!audio_url_.empty()) && (llm_task_obj->audio_flage_ == false)) {
             std::weak_ptr<llm_task> _llm_task_obj = llm_task_obj;
             llm_channel->subscriber(audio_url_, [_llm_task_obj](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
-                _llm_task_obj.lock()->sys_pcm_on_data(raw->string());
+                if (auto p = _llm_task_obj.lock()) {
+                    p->sys_pcm_on_data(raw->string());
+                }
             });
             llm_task_obj->audio_flage_ = true;
         }
@@ -452,7 +686,6 @@ class llm_asr : public StackFlow {
     void work(const std::string &work_id, const std::string &object, const std::string &data) override
     {
         SLOGI("llm_asr::work:%s", data.c_str());
-
         nlohmann::json error_body;
         int work_id_num = sample_get_work_id_num(work_id);
         if (llm_task_.find(work_id_num) == llm_task_.end()) {
@@ -467,8 +700,7 @@ class llm_asr : public StackFlow {
 
     void pause(const std::string &work_id, const std::string &object, const std::string &data) override
     {
-        SLOGI("llm_asr::work:%s", data.c_str());
-
+        SLOGI("llm_asr::pause:%s", data.c_str());
         nlohmann::json error_body;
         int work_id_num = sample_get_work_id_num(work_id);
         if (llm_task_.find(work_id_num) == llm_task_.end()) {
@@ -502,9 +734,10 @@ class llm_asr : public StackFlow {
             SLOGE("setup json format error.");
             error_body["code"]    = -2;
             error_body["message"] = "json format error.";
-            send("None", "None", error_body, "kws");
+            send("None", "None", error_body, "asr");
             return -2;
         }
+
         int ret = llm_task_obj->load_model(config_body);
         if (ret == 0) {
             llm_channel->set_output(llm_task_obj->enoutput_);
@@ -514,13 +747,15 @@ class llm_asr : public StackFlow {
                                                std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
                                                std::placeholders::_2));
 
-            for (const auto input : llm_task_obj->inputs_) {
+            for (const auto &input : llm_task_obj->inputs_) {
                 if (input.find("sys") != std::string::npos) {
                     audio_url_                            = unit_call("audio", "cap", input);
                     std::weak_ptr<llm_task> _llm_task_obj = llm_task_obj;
                     llm_channel->subscriber(audio_url_,
                                             [_llm_task_obj](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
-                                                _llm_task_obj.lock()->sys_pcm_on_data(raw->string());
+                                                if (auto p = _llm_task_obj.lock()) {
+                                                    p->sys_pcm_on_data(raw->string());
+                                                }
                                             });
                     llm_task_obj->audio_flage_ = true;
                 } else if (input.find("asr") != std::string::npos) {
@@ -538,12 +773,13 @@ class llm_asr : public StackFlow {
                                          std::placeholders::_2));
                 }
             }
+
             llm_task_[work_id_num] = llm_task_obj;
-            SLOGI("load_mode success");
+            SLOGI("load_model success");
             send("None", "None", LLM_NO_ERROR, work_id);
             return 0;
         } else {
-            SLOGE("load_mode Failed");
+            SLOGE("load_model Failed");
             error_body["code"]    = -5;
             error_body["message"] = "Model loading failed.";
             send("None", "None", error_body, "asr");
@@ -563,13 +799,17 @@ class llm_asr : public StackFlow {
             send("None", "None", error_body, work_id);
             return;
         }
+
         auto llm_channel  = get_channel(work_id);
         auto llm_task_obj = llm_task_[work_id_num];
+
         if (data.find("sys") != std::string::npos) {
             if (audio_url_.empty()) audio_url_ = unit_call("audio", "cap", data);
             std::weak_ptr<llm_task> _llm_task_obj = llm_task_obj;
             llm_channel->subscriber(audio_url_, [_llm_task_obj](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
-                _llm_task_obj.lock()->sys_pcm_on_data(raw->string());
+                if (auto p = _llm_task_obj.lock()) {
+                    p->sys_pcm_on_data(raw->string());
+                }
             });
             llm_task_obj->audio_flage_ = true;
             llm_task_obj->inputs_.push_back(data);
@@ -581,6 +821,7 @@ class llm_asr : public StackFlow {
                                              std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1, std::placeholders::_2));
             llm_task_obj->inputs_.push_back(data);
         }
+
         if (ret) {
             error_body["code"]    = -20;
             error_body["message"] = "link false";
@@ -594,7 +835,6 @@ class llm_asr : public StackFlow {
     void unlink(const std::string &work_id, const std::string &object, const std::string &data) override
     {
         SLOGI("llm_asr::unlink:%s", data.c_str());
-        int ret = 0;
         nlohmann::json error_body;
         int work_id_num = sample_get_work_id_num(work_id);
         if (llm_task_.find(work_id_num) == llm_task_.end()) {
@@ -603,8 +843,10 @@ class llm_asr : public StackFlow {
             send("None", "None", error_body, work_id);
             return;
         }
+
         auto llm_channel = get_channel(work_id);
         llm_channel->stop_subscriber_work_id(data);
+
         auto llm_task_obj = llm_task_[work_id_num];
         for (auto it = llm_task_obj->inputs_.begin(); it != llm_task_obj->inputs_.end();) {
             if (*it == data) {
@@ -619,13 +861,12 @@ class llm_asr : public StackFlow {
     void taskinfo(const std::string &work_id, const std::string &object, const std::string &data) override
     {
         SLOGI("llm_asr::taskinfo:%s", data.c_str());
-
         nlohmann::json req_body;
         int work_id_num = sample_get_work_id_num(work_id);
         if (WORK_ID_NONE == work_id_num) {
             std::vector<std::string> task_list;
             std::transform(llm_task_channel_.begin(), llm_task_channel_.end(), std::back_inserter(task_list),
-                           [](const auto task_channel) { return task_channel.second->work_id_; });
+                           [](const auto &task_channel) { return task_channel.second->work_id_; });
             req_body = task_list;
             send("asr.tasklist", req_body, LLM_NO_ERROR, work_id);
         } else {
@@ -669,17 +910,17 @@ class llm_asr : public StackFlow {
     ~llm_asr()
     {
         while (1) {
-            auto iteam = llm_task_.begin();
-            if (iteam == llm_task_.end()) {
+            auto it = llm_task_.begin();
+            if (it == llm_task_.end()) {
                 break;
             }
-            iteam->second->stop();
-            if (iteam->second->audio_flage_) {
+            it->second->stop();
+            if (it->second->audio_flage_) {
                 unit_call("audio", "cap_stop", "None");
             }
-            get_channel(iteam->first)->stop_subscriber("");
-            iteam->second.reset();
-            llm_task_.erase(iteam->first);
+            get_channel(it->first)->stop_subscriber("");
+            it->second.reset();
+            llm_task_.erase(it->first);
         }
     }
 };
diff --git a/projects/llm_framework/main_audio/SConstruct b/projects/llm_framework/main_audio/SConstruct
index 444248ea..35b70234 100644
--- a/projects/llm_framework/main_audio/SConstruct
+++ b/projects/llm_framework/main_audio/SConstruct
@@ -5,7 +5,11 @@ with open(env['PROJECT_TOOL_S']) as f:
     exec(f.read())
 
 # SRCS = append_srcs_dir(ADir('src'))
-SRCS = Glob('src/*.c*')
+if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
+    SRCS = [AFile('src/sample_audio.c'), AFile('src/main.cpp')]
+else:
+    SRCS = [AFile('src/alsa_audio.c'), AFile('src/main.cpp')]
+
 INCLUDE = [ADir('include'), ADir('.')]
 PRIVATE_INCLUDE = []
 REQUIREMENTS = ['pthread', 'utilities', 'ax_msp', 'eventpp', 'StackFlow', 'single_header_libs']
diff --git a/projects/llm_framework/main_audio/src/alsa_audio.c b/projects/llm_framework/main_audio/src/alsa_audio.c
index 944497a9..fa86a5e4 100644
--- a/projects/llm_framework/main_audio/src/alsa_audio.c
+++ b/projects/llm_framework/main_audio/src/alsa_audio.c
@@ -6,7 +6,11 @@
 #include <errno.h>
 #include <string.h>
 
-static int gcapLoopExit = 0;
+static struct pcm *g_play_pcm = NULL;
+static int gplayLoopExit      = 1;
+static int gcapLoopExit       = 1;
+AlsaConfig cap_config;
+AlsaConfig play_config;
 
 void alsa_cap_start(unsigned int card, unsigned int device, float Volume, int channel, int rate, int bit,
                     AUDIOCallback callback)
@@ -23,7 +27,7 @@ void alsa_cap_start(unsigned int card, unsigned int device, float Volume, int ch
 
     memset(&config, 0, sizeof(config));
     config.channels          = channel;
-    config.rate              = 48000;  // TODO: 部分USB MIC仅支持48k，暂时固定采集为48k
+    config.rate              = 48000;
     config.period_size       = 120;
     config.period_count      = 4;
     config.format            = PCM_FORMAT_S16_LE;
@@ -57,14 +61,16 @@ void alsa_cap_start(unsigned int card, unsigned int device, float Volume, int ch
 
     SRC_STATE *src_state = NULL;
     float *in_float = NULL, *out_float = NULL;
+
     int in_frames  = pcm_get_buffer_size(pcm);
     int out_frames = (int)((float)in_frames * ((float)rate / 48000.0f) + 1);
-    int out_bytes  = out_frames * channel * sizeof(short);
+
+    int src_channels = 1;
 
     if (rate != 48000) {
-        src_state = src_new(SRC_SINC_FASTEST, channel, NULL);
-        in_float  = malloc(in_frames * channel * sizeof(float));
-        out_float = malloc(out_frames * channel * sizeof(float));
+        src_state = src_new(SRC_SINC_FASTEST, src_channels, NULL);
+        in_float  = malloc(in_frames * src_channels * sizeof(float));
+        out_float = malloc(out_frames * src_channels * sizeof(float));
         if (!src_state || !in_float || !out_float) {
             fprintf(stderr, "Unable to allocate resample buffers\n");
             free(buffer);
@@ -82,16 +88,36 @@ void alsa_cap_start(unsigned int card, unsigned int device, float Volume, int ch
             fprintf(stderr, "Error capturing samples - %d (%s)\n", errno, strerror(errno));
             break;
         }
+
         frames_read = ret;
         total_frames_read += frames_read;
 
         if (rate == 48000) {
-            callback(buffer, frames_read * bytes_per_frame);
+            int in_channels = channel;  // 比如 4
+            int16_t *in     = (int16_t *)buffer;
+
+            int16_t *ch0 = malloc(frames_read * sizeof(int16_t));
+            if (!ch0) {
+                fprintf(stderr, "Unable to allocate ch0 buffer\n");
+                break;
+            }
+
+            for (unsigned int i = 0; i < frames_read; ++i) {
+                ch0[i] = in[i * in_channels + 0];
+            }
+
+            callback((const char *)ch0, frames_read * sizeof(int16_t));
+            free(ch0);
+
         } else {
+            int in_channels = channel;  // 比如 4
             short *in_short = (short *)buffer;
-            for (int i = 0; i < frames_read * channel; ++i) {
-                in_float[i] = in_short[i] / 32768.0f;
+
+            for (int i = 0; i < frames_read; ++i) {
+                short s     = in_short[i * in_channels + 0];
+                in_float[i] = s / 32768.0f;
             }
+
             SRC_DATA src_data;
             src_data.data_in       = in_float;
             src_data.input_frames  = frames_read;
@@ -99,20 +125,25 @@ void alsa_cap_start(unsigned int card, unsigned int device, float Volume, int ch
             src_data.output_frames = out_frames;
             src_data.src_ratio     = (double)rate / 48000.0;
             src_data.end_of_input  = 0;
-            int error              = src_process(src_state, &src_data);
+
+            int error = src_process(src_state, &src_data);
             if (error) {
                 fprintf(stderr, "SRC error: %s\n", src_strerror(error));
                 break;
             }
-            // float转short
-            short *out_short = malloc(src_data.output_frames_gen * channel * sizeof(short));
-            for (int i = 0; i < src_data.output_frames_gen * channel; ++i) {
+            int out_samples  = src_data.output_frames_gen;
+            short *out_short = malloc(out_samples * sizeof(short));
+            if (!out_short) {
+                fprintf(stderr, "Unable to allocate out_short buffer\n");
+                break;
+            }
+            for (int i = 0; i < out_samples; ++i) {
                 float sample = out_float[i];
                 if (sample > 1.0f) sample = 1.0f;
                 if (sample < -1.0f) sample = -1.0f;
                 out_short[i] = (short)(sample * 32767.0f);
             }
-            callback((const char *)out_short, src_data.output_frames_gen * channel * sizeof(short));
+            callback((const char *)out_short, out_samples * sizeof(short));
             free(out_short);
         }
     }
@@ -124,7 +155,6 @@ void alsa_cap_start(unsigned int card, unsigned int device, float Volume, int ch
     }
     free(buffer);
     pcm_close(pcm);
-    printf("Total frames captured: %u\n", total_frames_read);
 }
 
 void alsa_close_cap()
@@ -135,4 +165,52 @@ void alsa_close_cap()
 int alsa_cap_status()
 {
     return gcapLoopExit;
+}
+
+void alsa_play(unsigned int card, unsigned int device, float Volume, int channel, int rate, int bit, const void *data,
+               int size)
+{
+    gplayLoopExit = 0;
+
+    struct pcm_config config;
+    memset(&config, 0, sizeof(config));
+    config.channels          = channel;
+    config.rate              = rate;
+    config.period_size       = 1024;
+    config.period_count      = 2;
+    config.format            = PCM_FORMAT_S16_LE;
+    config.silence_threshold = config.period_size * config.period_count;
+    config.stop_threshold    = config.period_size * config.period_count;
+    config.start_threshold   = config.period_size;
+
+    unsigned int pcm_open_flags = PCM_OUT;
+
+    struct pcm *pcm = pcm_open(card, device, pcm_open_flags, &config);
+    if (!pcm || !pcm_is_ready(pcm)) {
+        fprintf(stderr, "Unable to open PCM playback device (%s)\n", pcm ? pcm_get_error(pcm) : "invalid pcm");
+        if (pcm) {
+            pcm_close(pcm);
+        }
+        gplayLoopExit = 2;
+        return;
+    }
+
+    int frames         = pcm_bytes_to_frames(pcm, size);
+    int written_frames = pcm_writei(pcm, data, frames);
+    if (written_frames < 0) {
+        fprintf(stderr, "PCM playback error %s\n", pcm_get_error(pcm));
+    }
+    printf("Played %d frames\n", written_frames);
+    pcm_close(pcm);
+    gplayLoopExit = 2;
+}
+
+void alsa_close_play()
+{
+    gplayLoopExit = 1;
+}
+
+int alsa_play_status()
+{
+    return gplayLoopExit;
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_audio/src/alsa_audio.h b/projects/llm_framework/main_audio/src/alsa_audio.h
index 95c469ab..646af36d 100644
--- a/projects/llm_framework/main_audio/src/alsa_audio.h
+++ b/projects/llm_framework/main_audio/src/alsa_audio.h
@@ -1,5 +1,18 @@
 #pragma once
 
+typedef struct AlsaConfig
+{
+    unsigned int card;
+    unsigned int device;
+    float volume;
+    int channel;
+    int rate;
+    int bit;
+} AlsaConfig;
+
+extern AlsaConfig cap_config;
+extern AlsaConfig play_config;
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -10,6 +23,13 @@ void alsa_cap_start(unsigned int card, unsigned int device, float Volume, int ch
                     AUDIOCallback callback);
 void alsa_close_cap();
 
+void alsa_play(unsigned int card, unsigned int device, float Volume, int channel, int rate, int bit, const void *data,
+             int size);
+
+void alsa_close_play();
+int alsa_cap_status();
+int alsa_play_status();
+
 #ifdef __cplusplus
 }
 #endif
\ No newline at end of file
diff --git a/projects/llm_framework/main_audio/src/main.cpp b/projects/llm_framework/main_audio/src/main.cpp
index de6165fd..3500a447 100644
--- a/projects/llm_framework/main_audio/src/main.cpp
+++ b/projects/llm_framework/main_audio/src/main.cpp
@@ -13,6 +13,7 @@
 #include <iostream>
 #include <stdexcept>
 #include "../../../../SDK/components/utilities/include/sample_log.h"
+#include <global_config.h>
 
 int main_exit_flage = 0;
 static void __sigint(int iSigNo)
@@ -21,8 +22,11 @@ static void __sigint(int iSigNo)
     main_exit_flage = 1;
 }
 
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
 #include "sample_audio.h"
+#else
 #include "alsa_audio.h"
+#endif
 
 #define CONFIG_AUTO_SET(obj, key)             \
     if (config_body.contains(#key))           \
@@ -80,8 +84,13 @@ class llm_audio : public StackFlow {
         }
 
         std::lock_guard<std::mutex> guard(ax_play_mtx);
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
         ax_play(play_config.card, play_config.device, play_config.volume, play_config.channel, play_config.rate,
-                play_config.bit, final_data.c_str(), final_data.length());
+                play_config.bit, audio_data.c_str(), audio_data.length());
+#else
+        alsa_play(play_config.card, play_config.device, play_config.volume, play_config.channel, play_config.rate,
+                  play_config.bit, final_data.c_str(), final_data.length());
+#endif
     }
 
     void hw_play(const std::string &audio_data)
@@ -92,8 +101,13 @@ class llm_audio : public StackFlow {
             final_data = mono_to_stereo_s16le(audio_data);
         }
         std::lock_guard<std::mutex> guard(ax_play_mtx);
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
         ax_play(play_config.card, play_config.device, play_config.volume, play_config.channel, play_config.rate,
                 play_config.bit, audio_data.c_str(), audio_data.length());
+#else
+        alsa_play(play_config.card, play_config.device, play_config.volume, play_config.channel, play_config.rate,
+                  play_config.bit, final_data.c_str(), final_data.length());
+#endif
     }
 
     void hw_cap()
@@ -117,7 +131,11 @@ class llm_audio : public StackFlow {
 
     void _play_stop()
     {
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
         ax_close_play();
+#else
+        alsa_close_play();
+#endif
         if (audio_play_thread_) {
             audio_play_thread_->join();
             audio_play_thread_.reset();
@@ -134,7 +152,7 @@ class llm_audio : public StackFlow {
 
     void _cap_stop()
     {
-#if (defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)) && !defined(CONFIG_AXCL_ENABLED)
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
         ax_close_cap();
 
 #else
@@ -219,8 +237,10 @@ class llm_audio : public StackFlow {
             send("None", "None", error_body, "audio");
             return -2;
         }
-        AX_AUDIO_SAMPLE_CONFIG_t mode_config_;
+
         try {
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
+            AX_AUDIO_SAMPLE_CONFIG_t mode_config_;
             memset(&mode_config_, 0, sizeof(AX_AUDIO_SAMPLE_CONFIG_t));
             if (object == "audio.play") {
                 CONFIG_AUTO_SET(file_body["play_param"], stPoolConfig.MetaSize);
@@ -244,22 +264,7 @@ class llm_audio : public StackFlow {
                 CONFIG_AUTO_SET(file_body["play_param"], stVqeAttr.stAgcCfg.enAgcMode);
                 CONFIG_AUTO_SET(file_body["play_param"], stVqeAttr.stAgcCfg.s16TargetLevel);
                 CONFIG_AUTO_SET(file_body["play_param"], stVqeAttr.stAgcCfg.s16Gain);
-#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
-                CONFIG_AUTO_SET(file_body["play_param"], stHpfAttr.bEnable);
-                CONFIG_AUTO_SET(file_body["play_param"], stHpfAttr.s32GainDb);
-                CONFIG_AUTO_SET(file_body["play_param"], stHpfAttr.s32Freq);
-                CONFIG_AUTO_SET(file_body["play_param"], stLpfAttr.bEnable);
-                CONFIG_AUTO_SET(file_body["play_param"], stLpfAttr.s32GainDb);
-                CONFIG_AUTO_SET(file_body["play_param"], stLpfAttr.s32Samplerate);
-                CONFIG_AUTO_SET(file_body["play_param"], stLpfAttr.s32Freq);
-                CONFIG_AUTO_SET(file_body["play_param"], stEqAttr.bEnable);
-                CONFIG_AUTO_SET(file_body["play_param"], stEqAttr.s32GainDb[0]);
-                CONFIG_AUTO_SET(file_body["play_param"], stEqAttr.s32GainDb[1]);
-                CONFIG_AUTO_SET(file_body["play_param"], stEqAttr.s32GainDb[2]);
-                CONFIG_AUTO_SET(file_body["play_param"], stEqAttr.s32GainDb[3]);
-                CONFIG_AUTO_SET(file_body["play_param"], stEqAttr.s32GainDb[4]);
-                CONFIG_AUTO_SET(file_body["play_param"], stEqAttr.s32Samplerate);
-#endif
+
                 CONFIG_AUTO_SET(file_body["play_param"], gResample);
                 CONFIG_AUTO_SET(file_body["play_param"], enInSampleRate);
                 CONFIG_AUTO_SET(file_body["play_param"], gInstant);
@@ -272,12 +277,12 @@ class llm_audio : public StackFlow {
                 CONFIG_AUTO_SET(file_body["play_param"], bit);
                 if (config_body.contains("stPoolConfig.PartitionName")) {
                     std::string PartitionName = config_body["stPoolConfig.PartitionName"];
-                    for (int i = 0; i < PartitionName.length(); i++) {
+                    for (int i = 0; i < (int)PartitionName.length(); i++) {
                         mode_config_.stPoolConfig.PartitionName[i] = PartitionName[i];
                     }
-                } else if (file_body["cap_param"].contains("stPoolConfig.PartitionName")) {
-                    std::string PartitionName = file_body["cap_param"]["stPoolConfig.PartitionName"];
-                    for (int i = 0; i < PartitionName.length(); i++) {
+                } else if (file_body["play_param"].contains("stPoolConfig.PartitionName")) {
+                    std::string PartitionName = file_body["play_param"]["stPoolConfig.PartitionName"];
+                    for (int i = 0; i < (int)PartitionName.length(); i++) {
                         mode_config_.stPoolConfig.PartitionName[i] = PartitionName[i];
                     }
                 }
@@ -292,7 +297,7 @@ class llm_audio : public StackFlow {
                 CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.enBitwidth);
                 CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.enLinkMode);
                 CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.enSamplerate);
-                // CONFIG_AUTO_SET(file_body["cap_param"], aistAttr.enLayoutMode);
+
                 if (config_body.contains("aistAttr.enLayoutMode"))
                     mode_config_.aistAttr.enLayoutMode = config_body["aistAttr.enLayoutMode"];
                 else if (file_body["cap_param"].contains("aistAttr.enLayoutMode")) {
@@ -316,23 +321,7 @@ class llm_audio : public StackFlow {
                 CONFIG_AUTO_SET(file_body["cap_param"], aistVqeAttr.stAgcCfg.s16TargetLevel);
                 CONFIG_AUTO_SET(file_body["cap_param"], aistVqeAttr.stAgcCfg.s16Gain);
                 CONFIG_AUTO_SET(file_body["cap_param"], aistVqeAttr.stAecCfg.enAecMode);
-#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
-                CONFIG_AUTO_SET(file_body["cap_param"], stHpfAttr.bEnable);
-                CONFIG_AUTO_SET(file_body["cap_param"], stHpfAttr.s32GainDb);
-                CONFIG_AUTO_SET(file_body["cap_param"], stHpfAttr.s32Samplerate);
-                CONFIG_AUTO_SET(file_body["cap_param"], stHpfAttr.s32Freq);
-                CONFIG_AUTO_SET(file_body["cap_param"], stLpfAttr.bEnable);
-                CONFIG_AUTO_SET(file_body["cap_param"], stLpfAttr.s32GainDb);
-                CONFIG_AUTO_SET(file_body["cap_param"], stLpfAttr.s32Samplerate);
-                CONFIG_AUTO_SET(file_body["cap_param"], stLpfAttr.s32Freq);
-                CONFIG_AUTO_SET(file_body["cap_param"], stEqAttr.bEnable);
-                CONFIG_AUTO_SET(file_body["cap_param"], stEqAttr.s32GainDb[0]);
-                CONFIG_AUTO_SET(file_body["cap_param"], stEqAttr.s32GainDb[1]);
-                CONFIG_AUTO_SET(file_body["cap_param"], stEqAttr.s32GainDb[2]);
-                CONFIG_AUTO_SET(file_body["cap_param"], stEqAttr.s32GainDb[3]);
-                CONFIG_AUTO_SET(file_body["cap_param"], stEqAttr.s32GainDb[4]);
-                CONFIG_AUTO_SET(file_body["cap_param"], stEqAttr.s32Samplerate);
-#endif
+
                 CONFIG_AUTO_SET(file_body["cap_param"], gResample);
                 CONFIG_AUTO_SET(file_body["cap_param"], enOutSampleRate);
                 CONFIG_AUTO_SET(file_body["cap_param"], gDbDetection);
@@ -345,12 +334,12 @@ class llm_audio : public StackFlow {
 
                 if (config_body.contains("stPoolConfig.PartitionName")) {
                     std::string PartitionName = config_body["stPoolConfig.PartitionName"];
-                    for (int i = 0; i < PartitionName.length(); i++) {
+                    for (int i = 0; i < (int)PartitionName.length(); i++) {
                         mode_config_.stPoolConfig.PartitionName[i] = PartitionName[i];
                     }
                 } else if (file_body["cap_param"].contains("stPoolConfig.PartitionName")) {
                     std::string PartitionName = file_body["cap_param"]["stPoolConfig.PartitionName"];
-                    for (int i = 0; i < PartitionName.length(); i++) {
+                    for (int i = 0; i < (int)PartitionName.length(); i++) {
                         mode_config_.stPoolConfig.PartitionName[i] = PartitionName[i];
                     }
                 }
@@ -361,6 +350,31 @@ class llm_audio : public StackFlow {
                 }
                 memcpy(&cap_config, &mode_config_, sizeof(AX_AUDIO_SAMPLE_CONFIG_t));
             }
+
+#else
+            AlsaConfig mode_config_;
+            memset(&mode_config_, 0, sizeof(AlsaConfig));
+
+            if (object == "audio.play") {
+                CONFIG_AUTO_SET(file_body["play_param"], card);
+                CONFIG_AUTO_SET(file_body["play_param"], device);
+                CONFIG_AUTO_SET(file_body["play_param"], volume);
+                CONFIG_AUTO_SET(file_body["play_param"], channel);
+                CONFIG_AUTO_SET(file_body["play_param"], rate);
+                CONFIG_AUTO_SET(file_body["play_param"], bit);
+                memcpy(&play_config, &mode_config_, sizeof(AlsaConfig));
+            }
+
+            if (object == "audio.cap") {
+                CONFIG_AUTO_SET(file_body["cap_param"], card);
+                CONFIG_AUTO_SET(file_body["cap_param"], device);
+                CONFIG_AUTO_SET(file_body["cap_param"], volume);
+                CONFIG_AUTO_SET(file_body["cap_param"], channel);
+                CONFIG_AUTO_SET(file_body["cap_param"], rate);
+                CONFIG_AUTO_SET(file_body["cap_param"], bit);
+                memcpy(&cap_config, &mode_config_, sizeof(AlsaConfig));
+            }
+#endif
         } catch (...) {
             error_body["code"]    = -22;
             error_body["message"] = "Parameter format error.";
@@ -471,33 +485,23 @@ class llm_audio : public StackFlow {
     std::string audio_status(pzmq *_pzmq, const std::shared_ptr<pzmq_data> &rawdata)
     {
         std::string _rawdata = rawdata->string();
+
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
+        auto play_state = ax_play_status();
+        auto cap_state  = ax_cap_status();
+#else
+        auto play_state = alsa_play_status();
+        auto cap_state  = alsa_cap_status();
+#endif
+
         if (_rawdata == "play") {
-            if (ax_play_status()) {
-                return std::string("None");
-            } else {
-                return std::string("Runing");
-            }
+            return play_state ? "None" : "Runing";
         } else if (_rawdata == "cap") {
-            if (ax_cap_status()) {
-                return std::string("None");
-            } else {
-                return std::string("Runing");
-            }
+            return cap_state ? "None" : "Runing";
         } else {
             std::ostringstream return_val;
-            return_val << "{\"play\":";
-            if (ax_play_status()) {
-                return_val << "\"None\"";
-            } else {
-                return_val << "\"Runing\"";
-            }
-            return_val << "\"cap\":";
-            if (ax_cap_status()) {
-                return_val << "\"None\"";
-            } else {
-                return_val << "\"Runing\"";
-            }
-            return_val << "\"}";
+            return_val << "{\"play\":" << (play_state ? "\"None\"" : "\"Runing\"")
+                       << ",\"cap\":" << (cap_state ? "\"None\"" : "\"Runing\"") << "}";
             return return_val.str();
         }
     }
diff --git a/projects/llm_framework/main_kws/SConstruct b/projects/llm_framework/main_kws/SConstruct
index b46a1eb1..f746d335 100644
--- a/projects/llm_framework/main_kws/SConstruct
+++ b/projects/llm_framework/main_kws/SConstruct
@@ -29,11 +29,11 @@ INCLUDE += [ADir('../static_lib/include/sherpa'),
             ]
 
 LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/onnx')]
-LDFLAGS += ['-l:libcargs.a',
-            '-l:libsherpa-onnx-core.a', '-l:libkaldi-native-fbank-core.a',
+REQUIREMENTS += ['onnxruntime', 'cargs']
+
+LDFLAGS += ['-l:libsherpa-onnx-core.a', '-l:libkaldi-native-fbank-core.a','-l:libkissfft-float.a',
             '-l:libkaldi-decoder-core.a', '-l:libssentencepiece_core.a']
 
-REQUIREMENTS += ['onnxruntime']
 
 STATIC_FILES += [os.path.join(python_venv, 'sherpa-onnx')]
 STATIC_FILES += Glob('llm-kws_text2token.py')
diff --git a/projects/llm_framework/main_kws_new/SConstruct b/projects/llm_framework/main_kws_new/SConstruct
index 2458e81e..d9e0eca8 100644
--- a/projects/llm_framework/main_kws_new/SConstruct
+++ b/projects/llm_framework/main_kws_new/SConstruct
@@ -57,6 +57,8 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
+STATIC_FILES += Glob('mode_*.json')
+
 env['COMPONENTS'].append({'target':'llm_kws_new-1.9',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
diff --git a/projects/llm_framework/main_kws_new/mode_kws.json b/projects/llm_framework/main_kws_new/mode_kws.json
new file mode 100644
index 00000000..5647a8ab
--- /dev/null
+++ b/projects/llm_framework/main_kws_new/mode_kws.json
@@ -0,0 +1,41 @@
+{
+    "mode": "kws",
+    "type": "kws",
+    "homepage": "",
+    "capabilities": [
+        "Keyword_spotting",
+        "English"
+    ],
+    "input_type": [
+        "sys.pcm",
+        "sys.cap.0_0"
+    ],
+    "output_type": [
+        "kws.bool"
+    ],
+    "mode_param": {
+        "model": "kws.onnx",
+        "wake_wav_file": "/opt/m5stack/data/audio/wakeup_zh_cn.wav",
+        "chunk_size": 32,
+        "threshold": 0.9,
+        "min_continuous_frames": 5,
+        "REFRACTORY_TIME_MS": 2000,
+        "RESAMPLE_RATE": 16000,
+        "FEAT_DIM": 80,
+        "frame_opts.samp_freq": 16000,
+        "frame_opts.frame_length_ms": 25.0,
+        "frame_opts.frame_shift_ms": 10.0,
+        "frame_opts.snip_edges": false,
+        "frame_opts.dither": 0.0,
+        "frame_opts.preemph_coeff": 0.97,
+        "frame_opts.remove_dc_offset": true,
+        "frame_opts.window_type": "povey",
+        "mel_opts.num_bins": 80,
+        "mel_opts.low_freq": 20,
+        "mel_opts.high_freq": 0,
+        "energy_floor": 0.0,
+        "use_energy": false,
+        "raw_energy": true
+    },
+    "mode_param_bak": {}
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_kws_new/src/main.cpp b/projects/llm_framework/main_kws_new/src/main.cpp
index b56f3c60..a395fd97 100644
--- a/projects/llm_framework/main_kws_new/src/main.cpp
+++ b/projects/llm_framework/main_kws_new/src/main.cpp
@@ -191,9 +191,7 @@ class llm_task {
     bool detect_wakeup(const std::vector<float> &scores)
     {
         bool triggered = false;
-        SLOGE("%d", scores.size());
         for (auto score : scores) {
-            printf("%f ", score);
             if (score > threshold) {
                 count_frames++;
                 if (count_frames >= min_continuous_frames) {
@@ -208,7 +206,6 @@ class llm_task {
             }
             frame_index_global++;
         }
-        SLOGE("\n");
         return triggered;
     }
 
diff --git a/projects/llm_framework/main_vad/SConstruct b/projects/llm_framework/main_vad/SConstruct
index 2d802067..f7d02e72 100644
--- a/projects/llm_framework/main_vad/SConstruct
+++ b/projects/llm_framework/main_vad/SConstruct
@@ -23,7 +23,8 @@ LINK_SEARCH_PATH += [ADir('../static_lib')]
 INCLUDE += [ADir('../static_lib/include/sherpa')]
 
 LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/onnx')]
-LDFLAGS += ['-l:libsherpa-onnx-core.a']
+
+LDFLAGS += ['-l:libsherpa-onnx-core.a', '-l:libkaldi-native-fbank-core.a', '-l:libkissfft-float.a']
 
 REQUIREMENTS += ['onnxruntime']
 

From a67466553715e93a63e1306f19f1390331adfac9 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Tue, 9 Dec 2025 18:09:16 +0800
Subject: [PATCH 66/79] [update] kws supported custom 'hi m5' keywords

---
 projects/llm_framework/main_kws/SConstruct    |   3 +-
 .../{main_kws_new => main_kws}/mode_kws.json  |   1 +
 projects/llm_framework/main_kws/src/main.cpp  | 512 +++++++++----
 projects/llm_framework/main_kws_new/Kconfig   |   0
 .../llm_framework/main_kws_new/SConstruct     |  75 --
 .../llm_framework/main_kws_new/src/main.cpp   | 714 ------------------
 6 files changed, 386 insertions(+), 919 deletions(-)
 rename projects/llm_framework/{main_kws_new => main_kws}/mode_kws.json (97%)
 delete mode 100644 projects/llm_framework/main_kws_new/Kconfig
 delete mode 100644 projects/llm_framework/main_kws_new/SConstruct
 delete mode 100644 projects/llm_framework/main_kws_new/src/main.cpp

diff --git a/projects/llm_framework/main_kws/SConstruct b/projects/llm_framework/main_kws/SConstruct
index f746d335..eb00040a 100644
--- a/projects/llm_framework/main_kws/SConstruct
+++ b/projects/llm_framework/main_kws/SConstruct
@@ -23,13 +23,14 @@ LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 
 INCLUDE += [ADir('../static_lib/include/sherpa'),
+            ADir('../static_lib/include/sherpa/fbank'),
             ADir('../static_lib/include/sherpa/sherpa-onnx'),
             ADir('../static_lib/include/sherpa/sherpa-onnx/onnxruntime-src'),
             ADir('../static_lib/include/sherpa/sherpa-onnx/openfst-src')
             ]
 
 LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/onnx')]
-REQUIREMENTS += ['onnxruntime', 'cargs']
+REQUIREMENTS += ['onnxruntime']
 
 LDFLAGS += ['-l:libsherpa-onnx-core.a', '-l:libkaldi-native-fbank-core.a','-l:libkissfft-float.a',
             '-l:libkaldi-decoder-core.a', '-l:libssentencepiece_core.a']
diff --git a/projects/llm_framework/main_kws_new/mode_kws.json b/projects/llm_framework/main_kws/mode_kws.json
similarity index 97%
rename from projects/llm_framework/main_kws_new/mode_kws.json
rename to projects/llm_framework/main_kws/mode_kws.json
index 5647a8ab..c1e11a92 100644
--- a/projects/llm_framework/main_kws_new/mode_kws.json
+++ b/projects/llm_framework/main_kws/mode_kws.json
@@ -15,6 +15,7 @@
     ],
     "mode_param": {
         "model": "kws.onnx",
+        "model_type": "onnx",
         "wake_wav_file": "/opt/m5stack/data/audio/wakeup_zh_cn.wav",
         "chunk_size": 32,
         "threshold": 0.9,
diff --git a/projects/llm_framework/main_kws/src/main.cpp b/projects/llm_framework/main_kws/src/main.cpp
index 9a4f8fc7..51022a83 100644
--- a/projects/llm_framework/main_kws/src/main.cpp
+++ b/projects/llm_framework/main_kws/src/main.cpp
@@ -4,8 +4,6 @@
  * SPDX-License-Identifier: MIT
  */
 #include "StackFlow.h"
-#include "sherpa-onnx/csrc/keyword-spotter.h"
-#include "sherpa-onnx/csrc/parse-options.h"
 
 #include <signal.h>
 #include <sys/stat.h>
@@ -16,7 +14,6 @@
 #include <stdexcept>
 #include <thread_safe_list.h>
 #include "../../../../SDK/components/utilities/include/sample_log.h"
-
 #define BUFFER_IMPLEMENTATION
 #include <stdbool.h>
 #include <stdint.h>
@@ -36,34 +33,78 @@ static std::string base_model_config_path_;
 
 typedef std::function<void(const std::string &data, bool finish)> task_callback_t;
 
-#define CONFIG_AUTO_SET(obj, key)             \
-    if (config_body.contains(#key))           \
-        mode_config_.key = config_body[#key]; \
-    else if (obj.contains(#key))              \
-        mode_config_.key = obj[#key];
+#include "sherpa-onnx/csrc/keyword-spotter.h"
+#include "sherpa-onnx/csrc/parse-options.h"
+
+#include <onnxruntime_cxx_api.h>
+#include "kaldi-native-fbank/csrc/online-feature.h"
+
+typedef struct mode_config_onnx {
+    int chunk_size            = 32;
+    float threshold           = 0.9f;
+    int min_continuous_frames = 5;
+    int REFRACTORY_TIME_MS    = 2000;
+    int RESAMPLE_RATE         = 16000;
+    int FEAT_DIM              = 80;
+    int delay_audio_frame_    = 32;
+} kws_config_onnx;
 
 class llm_task {
 private:
-    sherpa_onnx::KeywordSpotterConfig mode_config_;
-    std::unique_ptr<sherpa_onnx::KeywordSpotter> spotter_;
-    std::unique_ptr<sherpa_onnx::OnlineStream> spotter_stream_;
-
-public:
+    std::string model_type_;
     std::string model_;
     std::string response_format_;
     std::vector<std::string> inputs_;
     std::vector<std::string> kws_;
-    bool enoutput_;
-    bool enoutput_json_;
-    bool enstream_;
-    bool enwake_audio_;
+    bool enoutput_      = true;
+    bool enoutput_json_ = false;
+    bool enstream_      = false;
+    bool enwake_audio_  = true;
     std::atomic_bool audio_flage_;
     task_callback_t out_callback_;
-    int delay_audio_frame_ = 10;
     buffer_t *pcmdata;
     std::string wake_wav_file_;
-
     std::function<void(const std::string &)> play_awake_wav;
+    int delay_audio_frame_ = 10;
+
+    sherpa_onnx::KeywordSpotterConfig sherpa_config_;
+    std::unique_ptr<sherpa_onnx::KeywordSpotter> sherpa_spotter_;
+    std::unique_ptr<sherpa_onnx::OnlineStream> sherpa_stream_;
+
+    kws_config_onnx onnx_config_;
+    std::vector<float> onnx_cache_;
+    std::unique_ptr<Ort::Session> onnx_session_;
+    knf::FbankOptions fbank_opts_;
+    std::unique_ptr<knf::OnlineFbank> fbank_;
+    Ort::Env onnx_env_{ORT_LOGGING_LEVEL_WARNING, "kws"};
+    Ort::SessionOptions session_options_;
+    int count_frames_               = 0;
+    long long last_trigger_time_ms_ = -1e9;
+    long long frame_index_global_   = 0;
+
+public:
+    inline const std::string &model() const
+    {
+        return model_;
+    }
+    inline const std::string &response_format() const
+    {
+        return response_format_;
+    }
+    inline const std::vector<std::string> &inputs() const
+    {
+        return inputs_;
+    }
+    inline bool enoutput() const
+    {
+        return enoutput_;
+    }
+    bool enstream_flag() const
+    {
+        return enstream_;
+    }
+
+    friend class llm_kws;
 
     bool parse_config(const nlohmann::json &config_body)
     {
@@ -71,6 +112,13 @@ class llm_task {
             model_           = config_body.at("model");
             response_format_ = config_body.at("response_format");
             enoutput_        = config_body.at("enoutput");
+
+            if (model_.rfind("sherpa-onnx", 0) == 0) {
+                model_type_ = "sherpa";
+            } else {
+                model_type_ = "onnx";
+            }
+
             if (config_body.contains("enwake_audio")) {
                 enwake_audio_ = config_body["enwake_audio"];
             } else {
@@ -94,21 +142,23 @@ class llm_task {
                     }
                 }
             }
-            enoutput_json_ = response_format_.find("json") == std::string::npos ? false : true;
+            enoutput_json_ = response_format_.find("json") != std::string::npos;
+            enstream_      = response_format_.find("stream") != std::string::npos;
         } catch (...) {
             SLOGE("setup config_body error");
             return true;
         }
-        enstream_ = response_format_.find("stream") == std::string::npos ? false : true;
         return false;
     }
 
-    int load_model(const nlohmann::json &config_body)
-    {
-        if (parse_config(config_body)) {
-            return -1;
-        }
+#define CONFIG_AUTO_SET_SHERPA(obj, key)        \
+    if (config_body.contains(#key))             \
+        sherpa_config_.key = config_body[#key]; \
+    else if (obj.contains(#key))                \
+        sherpa_config_.key = obj[#key];
 
+    int load_model_sherpa(const nlohmann::json &config_body)
+    {
         nlohmann::json file_body;
         std::list<std::string> config_file_paths =
             get_config_file_paths(base_model_path_, base_model_config_path_, model_);
@@ -131,58 +181,77 @@ class llm_task {
             std::string base_model = base_model_path_ + model_ + "/";
             SLOGI("base_model %s", base_model.c_str());
 
-            CONFIG_AUTO_SET(file_body["mode_param"], feat_config.sampling_rate);
-            CONFIG_AUTO_SET(file_body["mode_param"], feat_config.feature_dim);
-            CONFIG_AUTO_SET(file_body["mode_param"], feat_config.low_freq);
-            CONFIG_AUTO_SET(file_body["mode_param"], feat_config.high_freq);
-            CONFIG_AUTO_SET(file_body["mode_param"], feat_config.dither);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.transducer.encoder);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.transducer.decoder);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.transducer.joiner);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.paraformer.encoder);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.paraformer.decoder);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.wenet_ctc.model);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.wenet_ctc.chunk_size);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.wenet_ctc.num_left_chunks);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.zipformer2_ctc.model);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.nemo_ctc.model);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.provider_config.device);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.provider_config.provider);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.provider_config.cuda_config.cudnn_conv_algo_search);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.provider_config.trt_config.trt_max_workspace_size);
-            CONFIG_AUTO_SET(file_body["mode_param"],
-                            model_config.provider_config.trt_config.trt_max_partition_iterations);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.provider_config.trt_config.trt_min_subgraph_size);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.provider_config.trt_config.trt_fp16_enable);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.provider_config.trt_config.trt_detailed_build_log);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.provider_config.trt_config.trt_engine_cache_enable);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.provider_config.trt_config.trt_engine_cache_path);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.provider_config.trt_config.trt_timing_cache_enable);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.provider_config.trt_config.trt_timing_cache_path);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.provider_config.trt_config.trt_dump_subgraphs);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.tokens);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.num_threads);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.warm_up);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.debug);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.model_type);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.modeling_unit);
-            CONFIG_AUTO_SET(file_body["mode_param"], model_config.bpe_vocab);
-            CONFIG_AUTO_SET(file_body["mode_param"], max_active_paths);
-            CONFIG_AUTO_SET(file_body["mode_param"], num_trailing_blanks);
-            CONFIG_AUTO_SET(file_body["mode_param"], keywords_score);
-            CONFIG_AUTO_SET(file_body["mode_param"], keywords_threshold);
-            CONFIG_AUTO_SET(file_body["mode_param"], keywords_file);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], feat_config.sampling_rate);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], feat_config.feature_dim);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], feat_config.low_freq);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], feat_config.high_freq);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], feat_config.dither);
+
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.transducer.encoder);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.transducer.decoder);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.transducer.joiner);
+
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.paraformer.encoder);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.paraformer.decoder);
+
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.wenet_ctc.model);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.wenet_ctc.chunk_size);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.wenet_ctc.num_left_chunks);
+
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.zipformer2_ctc.model);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.nemo_ctc.model);
+
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.provider_config.device);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.provider_config.provider);
+
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"],
+                                   model_config.provider_config.cuda_config.cudnn_conv_algo_search);
+
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"],
+                                   model_config.provider_config.trt_config.trt_max_workspace_size);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"],
+                                   model_config.provider_config.trt_config.trt_max_partition_iterations);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"],
+                                   model_config.provider_config.trt_config.trt_min_subgraph_size);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.provider_config.trt_config.trt_fp16_enable);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"],
+                                   model_config.provider_config.trt_config.trt_detailed_build_log);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"],
+                                   model_config.provider_config.trt_config.trt_engine_cache_enable);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"],
+                                   model_config.provider_config.trt_config.trt_engine_cache_path);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"],
+                                   model_config.provider_config.trt_config.trt_timing_cache_enable);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"],
+                                   model_config.provider_config.trt_config.trt_timing_cache_path);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.provider_config.trt_config.trt_dump_subgraphs);
+
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.tokens);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.num_threads);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.warm_up);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.debug);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.model_type);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.modeling_unit);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], model_config.bpe_vocab);
+
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], max_active_paths);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], num_trailing_blanks);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], keywords_score);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], keywords_threshold);
+            CONFIG_AUTO_SET_SHERPA(file_body["mode_param"], keywords_file);
 
             if (config_body.contains("wake_wav_file"))
                 wake_wav_file_ = config_body["wake_wav_file"];
             else if (file_body["mode_param"].contains("wake_wav_file"))
                 wake_wav_file_ = file_body["mode_param"]["wake_wav_file"];
 
-            mode_config_.model_config.transducer.encoder = base_model + mode_config_.model_config.transducer.encoder;
-            mode_config_.model_config.transducer.decoder = base_model + mode_config_.model_config.transducer.decoder;
-            mode_config_.model_config.transducer.joiner  = base_model + mode_config_.model_config.transducer.joiner;
-            mode_config_.model_config.tokens             = base_model + mode_config_.model_config.tokens;
-            mode_config_.keywords_file                   = base_model + mode_config_.keywords_file;
+            sherpa_config_.model_config.transducer.encoder =
+                base_model + sherpa_config_.model_config.transducer.encoder;
+            sherpa_config_.model_config.transducer.decoder =
+                base_model + sherpa_config_.model_config.transducer.decoder;
+            sherpa_config_.model_config.transducer.joiner = base_model + sherpa_config_.model_config.transducer.joiner;
+            sherpa_config_.model_config.tokens            = base_model + sherpa_config_.model_config.tokens;
+            sherpa_config_.keywords_file                  = base_model + sherpa_config_.keywords_file;
 
             std::ofstream temp_awake_key("/tmp/kws_awake.txt.tmp");
             for (const auto &keyword : kws_) {
@@ -200,7 +269,7 @@ class llm_task {
                 SLOGE("text2token.py or llm-kws_text2token.py not found!");
             }
             awake_key_compile_cmd << "--text /tmp/kws_awake.txt.tmp ";
-            awake_key_compile_cmd << "--tokens " << mode_config_.model_config.tokens << " ";
+            awake_key_compile_cmd << "--tokens " << sherpa_config_.model_config.tokens << " ";
             if (file_body["mode_param"].contains("text2token-tokens-type")) {
                 awake_key_compile_cmd << "--tokens-type "
                                       << file_body["mode_param"]["text2token-tokens-type"].get<std::string>() << " ";
@@ -209,16 +278,189 @@ class llm_task {
                 awake_key_compile_cmd << "--bpe-model " << base_model
                                       << file_body["mode_param"]["text2token-bpe-model"].get<std::string>() << " ";
             }
-            awake_key_compile_cmd << "--output " << mode_config_.keywords_file;
+            awake_key_compile_cmd << "--output " << sherpa_config_.keywords_file;
             system(awake_key_compile_cmd.str().c_str());
-            spotter_        = std::make_unique<sherpa_onnx::KeywordSpotter>(mode_config_);
-            spotter_stream_ = spotter_->CreateStream();
+
+            sherpa_spotter_ = std::make_unique<sherpa_onnx::KeywordSpotter>(sherpa_config_);
+            sherpa_stream_  = sherpa_spotter_->CreateStream();
         } catch (...) {
             SLOGE("config file read false");
             return -3;
         }
+
+        delay_audio_frame_ = 10;
         return 0;
     }
+#undef CONFIG_AUTO_SET_SHERPA
+
+#define CONFIG_AUTO_SET_ONNX(obj, key)        \
+    if (config_body.contains(#key))           \
+        onnx_config_.key = config_body[#key]; \
+    else if (obj.contains(#key))              \
+        onnx_config_.key = obj[#key];
+
+#define OPTS_AUTO_SET(obj, key)              \
+    if (config_body.contains(#key))          \
+        fbank_opts_.key = config_body[#key]; \
+    else if (obj.contains(#key))             \
+        fbank_opts_.key = obj[#key];
+
+    int load_model_onnx(const nlohmann::json &config_body)
+    {
+        nlohmann::json file_body;
+        std::list<std::string> config_file_paths =
+            get_config_file_paths(base_model_path_, base_model_config_path_, model_);
+        try {
+            for (auto file_name : config_file_paths) {
+                std::ifstream config_file(file_name);
+                if (!config_file.is_open()) {
+                    SLOGW("config file :%s miss", file_name.c_str());
+                    continue;
+                }
+                SLOGI("config file :%s read", file_name.c_str());
+                config_file >> file_body;
+                config_file.close();
+                break;
+            }
+            if (file_body.empty()) {
+                SLOGE("all config file miss");
+                return -2;
+            }
+            std::string base_model = base_model_path_ + model_ + "/";
+            SLOGI("base_model %s", base_model.c_str());
+            std::string model_file = base_model + "kws.onnx";
+
+            if (config_body.contains("wake_wav_file"))
+                wake_wav_file_ = config_body["wake_wav_file"];
+            else if (file_body["mode_param"].contains("wake_wav_file"))
+                wake_wav_file_ = file_body["mode_param"]["wake_wav_file"];
+
+            onnx_session_ = std::make_unique<Ort::Session>(onnx_env_, model_file.c_str(), session_options_);
+
+            onnx_cache_.assign(1 * 32 * 88, 0.0f);
+
+            auto &mp = file_body["mode_param"];
+            CONFIG_AUTO_SET_ONNX(mp, chunk_size);
+            CONFIG_AUTO_SET_ONNX(mp, threshold);
+            CONFIG_AUTO_SET_ONNX(mp, min_continuous_frames);
+            CONFIG_AUTO_SET_ONNX(mp, REFRACTORY_TIME_MS);
+            CONFIG_AUTO_SET_ONNX(mp, RESAMPLE_RATE);
+            CONFIG_AUTO_SET_ONNX(mp, FEAT_DIM);
+            CONFIG_AUTO_SET_ONNX(mp, delay_audio_frame_);
+
+            OPTS_AUTO_SET(mp, frame_opts.samp_freq);
+            OPTS_AUTO_SET(mp, frame_opts.frame_length_ms);
+            OPTS_AUTO_SET(mp, frame_opts.frame_shift_ms);
+            OPTS_AUTO_SET(mp, frame_opts.snip_edges);
+            OPTS_AUTO_SET(mp, frame_opts.dither);
+            OPTS_AUTO_SET(mp, frame_opts.preemph_coeff);
+            OPTS_AUTO_SET(mp, frame_opts.remove_dc_offset);
+            OPTS_AUTO_SET(mp, frame_opts.window_type);
+            OPTS_AUTO_SET(mp, mel_opts.num_bins);
+            OPTS_AUTO_SET(mp, mel_opts.low_freq);
+            OPTS_AUTO_SET(mp, mel_opts.high_freq);
+            OPTS_AUTO_SET(mp, energy_floor);
+            OPTS_AUTO_SET(mp, use_energy);
+            OPTS_AUTO_SET(mp, raw_energy);
+
+            fbank_ = std::make_unique<knf::OnlineFbank>(fbank_opts_);
+        } catch (...) {
+            SLOGE("config file read false");
+            return -3;
+        }
+        delay_audio_frame_ = onnx_config_.delay_audio_frame_;
+        return 0;
+    }
+#undef CONFIG_AUTO_SET_ONNX
+#undef OPTS_AUTO_SET
+
+    bool detect_wakeup(const std::vector<float> &scores)
+    {
+        bool triggered = false;
+        for (auto score : scores) {
+            if (score > onnx_config_.threshold) {
+                count_frames_++;
+                if (count_frames_ >= onnx_config_.min_continuous_frames) {
+                    long long trigger_time_ms = (frame_index_global_ - onnx_config_.min_continuous_frames + 1) * 10;
+                    if (trigger_time_ms - last_trigger_time_ms_ >= onnx_config_.REFRACTORY_TIME_MS) {
+                        last_trigger_time_ms_ = trigger_time_ms;
+                        triggered             = true;
+                    }
+                }
+            } else {
+                count_frames_ = 0;
+            }
+            frame_index_global_++;
+        }
+        return triggered;
+    }
+
+    std::vector<std::vector<float>> compute_fbank_kaldi(const std::vector<float> &waveform, int sample_rate,
+                                                        int num_mel_bins)
+    {
+        fbank_.reset();
+        fbank_ = std::make_unique<knf::OnlineFbank>(fbank_opts_);
+        fbank_->AcceptWaveform(sample_rate, waveform.data(), waveform.size());
+        int num_frames = fbank_->NumFramesReady();
+        std::vector<std::vector<float>> features;
+        features.reserve(num_frames);
+        for (int i = 0; i < num_frames; ++i) {
+            const float *frame_data = fbank_->GetFrame(i);
+            std::vector<float> frame(frame_data, frame_data + num_mel_bins);
+            features.push_back(std::move(frame));
+        }
+        return features;
+    }
+
+    std::vector<float> run_inference(const std::vector<float> &audio_chunk_16k)
+    {
+        std::vector<std::vector<float>> fbank_feats;
+        fbank_feats = compute_fbank_kaldi(audio_chunk_16k, onnx_config_.RESAMPLE_RATE, onnx_config_.FEAT_DIM);
+        if (fbank_feats.empty()) {
+            return {};
+        }
+        int T = fbank_feats.size();
+        std::vector<float> mat_flattened;
+        for (const auto &feat : fbank_feats) {
+            mat_flattened.insert(mat_flattened.end(), feat.begin(), feat.end());
+        }
+        std::vector<int64_t> input_shape = {1, static_cast<int64_t>(T), onnx_config_.FEAT_DIM};
+        std::vector<int64_t> cache_shape = {1, 32, 88};
+        Ort::MemoryInfo memory_info      = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
+        Ort::Value input_tensor          = Ort::Value::CreateTensor<float>(
+            memory_info, mat_flattened.data(), mat_flattened.size(), input_shape.data(), input_shape.size());
+        Ort::Value cache_tensor   = Ort::Value::CreateTensor<float>(memory_info, onnx_cache_.data(), onnx_cache_.size(),
+                                                                    cache_shape.data(), cache_shape.size());
+        const char *input_names[] = {"input", "cache"};
+        const char *output_names[] = {"output", "r_cache"};
+        std::vector<Ort::Value> inputs;
+        inputs.push_back(std::move(input_tensor));
+        inputs.push_back(std::move(cache_tensor));
+        auto output_tensors =
+            onnx_session_->Run(Ort::RunOptions{nullptr}, input_names, inputs.data(), 2, output_names, 2);
+        float *out_data                = output_tensors[0].GetTensorMutableData<float>();
+        float *cache_out_data          = output_tensors[1].GetTensorMutableData<float>();
+        std::vector<int64_t> out_shape = output_tensors[0].GetTensorTypeAndShapeInfo().GetShape();
+        size_t out_size                = 1;
+        for (auto dim : out_shape) out_size *= dim;
+        std::vector<float> out_chunk(out_data, out_data + out_size);
+        std::copy(cache_out_data, cache_out_data + onnx_cache_.size(), onnx_cache_.begin());
+        return out_chunk;
+    }
+
+    int load_model(const nlohmann::json &config_body)
+    {
+        if (parse_config(config_body)) {
+            return -1;
+        }
+        if (model_type_ == "onnx") {
+            SLOGE("load onnx kws model");
+            return load_model_onnx(config_body);
+        } else {
+            SLOGE("load sherpa kws model");
+            return load_model_sherpa(config_body);
+        }
+    }
 
     void set_output(task_callback_t out_callback)
     {
@@ -237,32 +479,47 @@ class llm_task {
         buffer_position_set(pcmdata, 0);
 
         std::vector<float> floatSamples;
-        {
-            int16_t audio_val;
-            while (buffer_read_i16(pcmdata, &audio_val, 1)) {
-                float normalizedSample = static_cast<float>(audio_val) / INT16_MAX;
-                floatSamples.push_back(normalizedSample);
+        std::vector<int16_t> int16Samples;
+
+        int16_t audio_val;
+        while (buffer_read_i16(pcmdata, &audio_val, 1)) {
+            int16Samples.push_back(audio_val);
+            if (model_type_ == "onnx") {
+                floatSamples.push_back(static_cast<float>(audio_val) / 1.0f);
+            } else {
+                floatSamples.push_back(static_cast<float>(audio_val) / INT16_MAX);
             }
         }
-
         buffer_resize(pcmdata, 0);
         count = 0;
 
-        spotter_stream_->AcceptWaveform(mode_config_.feat_config.sampling_rate, floatSamples.data(),
-                                        floatSamples.size());
-        while (spotter_->IsReady(spotter_stream_.get())) {
-            spotter_->DecodeStream(spotter_stream_.get());
-        }
-        sherpa_onnx::KeywordResult r = spotter_->GetResult(spotter_stream_.get());
-        if (!r.keyword.empty()) {
-            if (enwake_audio_ && (!wake_wav_file_.empty()) && play_awake_wav) {
-                play_awake_wav(wake_wav_file_);
-            }
-            if (out_callback_) {
-                if (enoutput_json_)
-                    out_callback_(r.AsJsonString(), true);
-                else
+        if (model_type_ == "onnx") {
+            auto scores = run_inference(floatSamples);
+            if (detect_wakeup(scores)) {
+                if (enwake_audio_ && (!wake_wav_file_.empty()) && play_awake_wav) {
+                    play_awake_wav(wake_wav_file_);
+                }
+                if (out_callback_) {
                     out_callback_("", true);
+                }
+            }
+        } else {
+            sherpa_stream_->AcceptWaveform(sherpa_config_.feat_config.sampling_rate, floatSamples.data(),
+                                           floatSamples.size());
+            while (sherpa_spotter_->IsReady(sherpa_stream_.get())) {
+                sherpa_spotter_->DecodeStream(sherpa_stream_.get());
+            }
+            sherpa_onnx::KeywordResult r = sherpa_spotter_->GetResult(sherpa_stream_.get());
+            if (!r.keyword.empty()) {
+                if (enwake_audio_ && (!wake_wav_file_.empty()) && play_awake_wav) {
+                    play_awake_wav(wake_wav_file_);
+                }
+                if (out_callback_) {
+                    if (enoutput_json_)
+                        out_callback_(r.AsJsonString(), true);
+                    else
+                        out_callback_("", true);
+                }
             }
         }
     }
@@ -274,7 +531,10 @@ class llm_task {
 
     bool delete_model()
     {
-        spotter_.reset();
+        sherpa_spotter_.reset();
+        sherpa_stream_.reset();
+        onnx_session_.reset();
+        fbank_.reset();
         return true;
     }
 
@@ -297,7 +557,6 @@ class llm_task {
         buffer_destroy(pcmdata);
     }
 };
-#undef CONFIG_AUTO_SET
 
 class llm_kws : public StackFlow {
 private:
@@ -366,7 +625,7 @@ class llm_kws : public StackFlow {
         fread(wav_data.data(), 1, wav_data.size(), fp);
         fclose(fp);
         int post = 0;
-        for (int i = 0; i < wav_data.size() - 4; i++) {
+        for (int i = 0; i < (int)wav_data.size() - 4; i++) {
             if ((wav_data[i] == 'd') && (wav_data[i + 1] == 'a') && (wav_data[i + 2] == 't') &&
                 (wav_data[i + 3] == 'a')) {
                 post = i + 8;
@@ -403,7 +662,8 @@ class llm_kws : public StackFlow {
         if ((!audio_url_.empty()) && (llm_task_obj->audio_flage_ == false)) {
             std::weak_ptr<llm_task> _llm_task_obj = llm_task_obj;
             llm_channel->subscriber(audio_url_, [_llm_task_obj](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
-                _llm_task_obj.lock()->sys_pcm_on_data(raw->string());
+                auto p = _llm_task_obj.lock();
+                if (p) p->sys_pcm_on_data(raw->string());
             });
             llm_task_obj->audio_flage_ = true;
         }
@@ -411,8 +671,7 @@ class llm_kws : public StackFlow {
 
     void work(const std::string &work_id, const std::string &object, const std::string &data) override
     {
-        SLOGI("llm_asr::work:%s", data.c_str());
-
+        SLOGI("llm_kws::work:%s", data.c_str());
         nlohmann::json error_body;
         int work_id_num = sample_get_work_id_num(work_id);
         if (llm_task_.find(work_id_num) == llm_task_.end()) {
@@ -427,8 +686,7 @@ class llm_kws : public StackFlow {
 
     void pause(const std::string &work_id, const std::string &object, const std::string &data) override
     {
-        SLOGI("llm_asr::work:%s", data.c_str());
-
+        SLOGI("llm_kws::pause:%s", data.c_str());
         nlohmann::json error_body;
         int work_id_num = sample_get_work_id_num(work_id);
         if (llm_task_.find(work_id_num) == llm_task_.end()) {
@@ -495,7 +753,6 @@ class llm_kws : public StackFlow {
             send("None", "None", error_body, "kws");
             return -1;
         }
-
         int work_id_num                             = sample_get_work_id_num(work_id);
         auto llm_channel                            = get_channel(work_id);
         auto llm_task_obj                           = std::make_shared<llm_task>(work_id);
@@ -513,28 +770,26 @@ class llm_kws : public StackFlow {
         }
         int ret = llm_task_obj->load_model(config_body);
         if (ret == 0) {
-            llm_channel->set_output(llm_task_obj->enoutput_);
-            llm_channel->set_stream(llm_task_obj->enstream_);
+            llm_channel->set_output(llm_task_obj->enoutput());
+            llm_channel->set_stream(llm_task_obj->enstream_flag());
             llm_task_obj->play_awake_wav = std::bind(&llm_kws::play_awake_wav, this, std::placeholders::_1);
-            llm_task_obj->set_output(std::bind(&llm_kws::task_output, this, std::weak_ptr<llm_task>(llm_task_obj),
+            llm_task_obj->set_output(std::bind(&llm_kws::task_output, this, _llm_task_obj,
                                                std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
                                                std::placeholders::_2));
-
-            for (const auto input : llm_task_obj->inputs_) {
+            for (const auto &input : llm_task_obj->inputs()) {
                 if (input.find("sys") != std::string::npos) {
                     audio_url_ = unit_call("audio", "cap", "None");
                     llm_channel->subscriber(audio_url_,
                                             [_llm_task_obj](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
-                                                auto llm_task_obj = _llm_task_obj.lock();
-                                                if (llm_task_obj) llm_task_obj->sys_pcm_on_data(raw->string());
+                                                auto p = _llm_task_obj.lock();
+                                                if (p) p->sys_pcm_on_data(raw->string());
                                             });
                     llm_task_obj->audio_flage_ = true;
                 } else if (input.find("kws") != std::string::npos) {
                     llm_task_obj->delay_audio_frame_ = 0;
-                    llm_channel->subscriber_work_id(
-                        "", std::bind(&llm_kws::task_user_data, this, std::weak_ptr<llm_task>(llm_task_obj),
-                                      std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
-                                      std::placeholders::_2));
+                    llm_channel->subscriber_work_id("", std::bind(&llm_kws::task_user_data, this, _llm_task_obj,
+                                                                  std::weak_ptr<llm_channel_obj>(llm_channel),
+                                                                  std::placeholders::_1, std::placeholders::_2));
                 }
             }
             llm_task_[work_id_num] = llm_task_obj;
@@ -558,7 +813,7 @@ class llm_kws : public StackFlow {
         if (WORK_ID_NONE == work_id_num) {
             std::vector<std::string> task_list;
             std::transform(llm_task_channel_.begin(), llm_task_channel_.end(), std::back_inserter(task_list),
-                           [](const auto task_channel) { return task_channel.second->work_id_; });
+                           [](const auto &task_channel) { return task_channel.second->work_id_; });
             req_body = task_list;
             send("kws.tasklist", req_body, LLM_NO_ERROR, work_id);
         } else {
@@ -569,10 +824,10 @@ class llm_kws : public StackFlow {
                 return;
             }
             auto llm_task_obj           = llm_task_[work_id_num];
-            req_body["model"]           = llm_task_obj->model_;
-            req_body["response_format"] = llm_task_obj->response_format_;
-            req_body["enoutput"]        = llm_task_obj->enoutput_;
-            req_body["inputs"]          = llm_task_obj->inputs_;
+            req_body["model"]           = llm_task_obj->model();
+            req_body["response_format"] = llm_task_obj->response_format();
+            req_body["enoutput"]        = llm_task_obj->enoutput();
+            req_body["inputs"]          = llm_task_obj->inputs();
             send("kws.taskinfo", req_body, LLM_NO_ERROR, work_id);
         }
     }
@@ -620,7 +875,6 @@ class llm_kws : public StackFlow {
             _zmq.send_data(out);
             return LLM_NONE;
         }
-
         int work_id_num = sample_get_work_id_num(work_id);
         if (llm_task_.find(work_id_num) == llm_task_.end()) {
             nlohmann::json out_body;
@@ -644,17 +898,17 @@ class llm_kws : public StackFlow {
     ~llm_kws()
     {
         while (1) {
-            auto iteam = llm_task_.begin();
-            if (iteam == llm_task_.end()) {
+            auto it = llm_task_.begin();
+            if (it == llm_task_.end()) {
                 break;
             }
-            iteam->second->stop();
-            if (iteam->second->audio_flage_) {
+            it->second->stop();
+            if (it->second->audio_flage_) {
                 unit_call("audio", "cap_stop", "None");
             }
-            get_channel(iteam->first)->stop_subscriber("");
-            iteam->second.reset();
-            llm_task_.erase(iteam->first);
+            get_channel(it->first)->stop_subscriber("");
+            it->second.reset();
+            llm_task_.erase(it->first);
         }
     }
 };
diff --git a/projects/llm_framework/main_kws_new/Kconfig b/projects/llm_framework/main_kws_new/Kconfig
deleted file mode 100644
index e69de29b..00000000
diff --git a/projects/llm_framework/main_kws_new/SConstruct b/projects/llm_framework/main_kws_new/SConstruct
deleted file mode 100644
index d9e0eca8..00000000
--- a/projects/llm_framework/main_kws_new/SConstruct
+++ /dev/null
@@ -1,75 +0,0 @@
-import os
-
-Import('env')
-with open(env['PROJECT_TOOL_S']) as f:
-    exec(f.read())
-
-SRCS = Glob('src/*.c*')
-INCLUDE = [ADir('include'), ADir('.')]
-PRIVATE_INCLUDE = []
-REQUIREMENTS = ['pthread', 'dl', 'utilities', 'eventpp', 'StackFlow', 'single_header_libs']
-STATIC_LIB = []
-DYNAMIC_LIB = []
-DEFINITIONS = []
-DEFINITIONS_PRIVATE = []
-LDFLAGS = []
-LINK_SEARCH_PATH = []
-STATIC_FILES = []
-
-python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-kws-python-venv_v1.6.tar.gz", 'm5stack_llm-kws-python-venv_v1.6.tar.gz')
-
-DEFINITIONS += ['-std=c++17']
-LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
-LINK_SEARCH_PATH += [ADir('../static_lib')]
-
-INCLUDE += [
-            ADir('../static_lib/include/sherpa/fbank'),
-            ADir('../static_lib/include/sherpa/sherpa-onnx/onnxruntime-src'),
-            ]
-
-LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/fbank')]
-LDFLAGS += [
-            '-l:libkaldi-native-fbank-core.a',
-            '-l:libkissfft-float.a',
-            ]
-
-REQUIREMENTS += ['onnxruntime']
-
-# STATIC_FILES += [os.path.join(python_venv, 'sherpa-onnx')]
-# STATIC_FILES += Glob('llm-kws_text2token.py')
-# STATIC_FILES += Glob('mode_*.json')
-
-IGNORE_FILES = []
-IGNORE_FILES += ['sherpa-onnx']
-
-import json
-if not os.path.exists('../dist'):
-    os.makedirs('../dist')
-ignore = {'ignore':[]}
-try:
-    with open('../dist/fileignore', 'a+') as f:
-        f.seek(0)
-        ignore = json.load(f)
-except:
-    pass
-ignore['ignore'] += IGNORE_FILES
-ignore['ignore'] = list(set(ignore['ignore']))
-with open('../dist/fileignore', 'w') as f:
-    json.dump(ignore, f, indent=4)
-
-STATIC_FILES += Glob('mode_*.json')
-
-env['COMPONENTS'].append({'target':'llm_kws_new-1.9',
-                          'SRCS':SRCS,
-                          'INCLUDE':INCLUDE,
-                          'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
-                          'REQUIREMENTS':REQUIREMENTS,
-                          'STATIC_LIB':STATIC_LIB,
-                          'DYNAMIC_LIB':DYNAMIC_LIB,
-                          'DEFINITIONS':DEFINITIONS,
-                          'DEFINITIONS_PRIVATE':DEFINITIONS_PRIVATE,
-                          'LDFLAGS':LDFLAGS,
-                          'LINK_SEARCH_PATH':LINK_SEARCH_PATH,
-                          'STATIC_FILES':STATIC_FILES,
-                          'REGISTER':'project'
-                          })
diff --git a/projects/llm_framework/main_kws_new/src/main.cpp b/projects/llm_framework/main_kws_new/src/main.cpp
deleted file mode 100644
index a395fd97..00000000
--- a/projects/llm_framework/main_kws_new/src/main.cpp
+++ /dev/null
@@ -1,714 +0,0 @@
-/*
- * SPDX-FileCopyrightText: 2024 M5Stack Technology CO LTD
- *
- * SPDX-License-Identifier: MIT
- */
-#include "StackFlow.h"
-
-#include <signal.h>
-#include <sys/stat.h>
-#include <sys/types.h>
-#include <unistd.h>
-#include <base64.h>
-#include <fstream>
-#include <stdexcept>
-#include <thread_safe_list.h>
-#include "../../../../SDK/components/utilities/include/sample_log.h"
-
-#define BUFFER_IMPLEMENTATION
-#include <stdbool.h>
-#include <stdint.h>
-#include "libs/buffer.h"
-
-#include <onnxruntime_cxx_api.h>
-
-#include "kaldi-native-fbank/csrc/online-feature.h"
-
-using namespace StackFlows;
-
-int main_exit_flage = 0;
-static void __sigint(int iSigNo)
-{
-    SLOGW("llm_kws_new will be exit!");
-    main_exit_flage = 1;
-}
-
-static std::string base_model_path_;
-static std::string base_model_config_path_;
-
-typedef std::function<void(const std::string &data, bool finish)> task_callback_t;
-
-#define CONFIG_AUTO_SET(obj, key)             \
-    if (config_body.contains(#key))           \
-        mode_config_.key = config_body[#key]; \
-    else if (obj.contains(#key))              \
-        mode_config_.key = obj[#key];
-
-class llm_task {
-private:
-    int chunk_size                 = 32;
-    float threshold                = 0.9f;
-    int min_continuous_frames      = 5;
-    int count_frames               = 0;
-    long long last_trigger_time_ms = -1e9;
-    long long frame_index_global   = 0;
-    int REFRACTORY_TIME_MS         = 2000;
-    const int RESAMPLE_RATE        = 16000;
-    const int FEAT_DIM             = 80;
-    std::vector<float> cache;
-    std::unique_ptr<Ort::Session> session;
-    knf::FbankOptions opts_;
-    std::unique_ptr<knf::OnlineFbank> fbank_;
-    Ort::Env env;
-    Ort::SessionOptions session_options;
-
-public:
-    std::string model_;
-    std::string response_format_;
-    std::vector<std::string> inputs_;
-    std::vector<std::string> kws_;
-    bool enoutput_;
-    bool enoutput_json_;
-    bool enstream_;
-    bool enwake_audio_;
-    std::atomic_bool audio_flage_;
-    task_callback_t out_callback_;
-    int delay_audio_frame_ = 32;
-    buffer_t *pcmdata;
-    std::string wake_wav_file_;
-    int file_counter = 0;
-
-    std::function<void(const std::string &)> play_awake_wav;
-
-    bool parse_config(const nlohmann::json &config_body)
-    {
-        try {
-            model_           = config_body.at("model");
-            response_format_ = config_body.at("response_format");
-            enoutput_        = config_body.at("enoutput");
-            if (config_body.contains("enwake_audio")) {
-                enwake_audio_ = config_body["enwake_audio"];
-            } else {
-                enwake_audio_ = true;
-            }
-            if (config_body.contains("input")) {
-                if (config_body["input"].is_string()) {
-                    inputs_.push_back(config_body["input"].get<std::string>());
-                } else if (config_body["input"].is_array()) {
-                    for (auto _in : config_body["input"]) {
-                        inputs_.push_back(_in.get<std::string>());
-                    }
-                }
-            }
-            if (config_body.contains("kws")) {
-                if (config_body["kws"].is_string()) {
-                    kws_.push_back(config_body["kws"].get<std::string>());
-                } else if (config_body["kws"].is_array()) {
-                    for (auto _in : config_body["kws"]) {
-                        kws_.push_back(_in.get<std::string>());
-                    }
-                }
-            }
-            enoutput_json_ = response_format_.find("json") == std::string::npos ? false : true;
-        } catch (...) {
-            SLOGE("setup config_body error");
-            return true;
-        }
-        enstream_ = response_format_.find("stream") == std::string::npos ? false : true;
-        return false;
-    }
-
-    int load_model(const nlohmann::json &config_body)
-    {
-        if (parse_config(config_body)) {
-            return -1;
-        }
-
-        nlohmann::json file_body;
-        std::list<std::string> config_file_paths =
-            get_config_file_paths(base_model_path_, base_model_config_path_, model_);
-        try {
-            for (auto file_name : config_file_paths) {
-                std::ifstream config_file(file_name);
-                if (!config_file.is_open()) {
-                    SLOGW("config file :%s miss", file_name.c_str());
-                    continue;
-                }
-                SLOGI("config file :%s read", file_name.c_str());
-                config_file >> file_body;
-                config_file.close();
-                break;
-            }
-            if (file_body.empty()) {
-                SLOGE("all config file miss");
-                return -2;
-            }
-            std::string base_model = base_model_path_ + model_ + "/";
-            SLOGI("base_model %s", base_model.c_str());
-
-            std::string model_file = base_model + "kws.onnx";
-
-            if (config_body.contains("wake_wav_file"))
-                wake_wav_file_ = config_body["wake_wav_file"];
-            else if (file_body["mode_param"].contains("wake_wav_file"))
-                wake_wav_file_ = file_body["mode_param"]["wake_wav_file"];
-
-            session = std::make_unique<Ort::Session>(env, model_file.c_str(), session_options);
-            cache.assign(1 * 32 * 88, 0.0f);
-
-            opts_.frame_opts.samp_freq        = 16000;
-            opts_.frame_opts.frame_length_ms  = 25.0;
-            opts_.frame_opts.frame_shift_ms   = 10.0;
-            opts_.frame_opts.snip_edges       = false;
-            opts_.frame_opts.dither           = 0.0;
-            opts_.frame_opts.preemph_coeff    = 0.97;
-            opts_.frame_opts.remove_dc_offset = true;
-            opts_.frame_opts.window_type      = "povey";
-
-            opts_.mel_opts.num_bins  = 80;
-            opts_.mel_opts.low_freq  = 20;
-            opts_.mel_opts.high_freq = 0;
-
-            opts_.energy_floor = 0.0;
-            opts_.use_energy   = false;
-
-            opts_.raw_energy = true;
-            // use_log_fbank / use_power 由 knf::OnlineFbank 默认控制，一般一致
-
-            fbank_ = std::make_unique<knf::OnlineFbank>(opts_);
-        } catch (...) {
-            SLOGE("config file read false");
-            return -3;
-        }
-        return 0;
-    }
-
-    void set_output(task_callback_t out_callback)
-    {
-        out_callback_ = out_callback;
-    }
-
-    bool detect_wakeup(const std::vector<float> &scores)
-    {
-        bool triggered = false;
-        for (auto score : scores) {
-            if (score > threshold) {
-                count_frames++;
-                if (count_frames >= min_continuous_frames) {
-                    long long trigger_time_ms = (frame_index_global - min_continuous_frames + 1) * 10;
-                    if (trigger_time_ms - last_trigger_time_ms >= REFRACTORY_TIME_MS) {
-                        last_trigger_time_ms = trigger_time_ms;
-                        triggered            = true;
-                    }
-                }
-            } else {
-                count_frames = 0;
-            }
-            frame_index_global++;
-        }
-        return triggered;
-    }
-
-    std::vector<std::vector<float>> compute_fbank_kaldi(const std::vector<float> &waveform, int sample_rate,
-                                                        int num_mel_bins)
-    {
-        fbank_.reset();
-        fbank_ = std::make_unique<knf::OnlineFbank>(opts_);
-        fbank_->AcceptWaveform(sample_rate, waveform.data(), waveform.size());
-        int num_frames = fbank_->NumFramesReady();
-
-        std::vector<std::vector<float>> features;
-
-        features.reserve(num_frames);
-        for (int i = 0; i < num_frames; ++i) {
-            const float *frame_data = fbank_->GetFrame(i);
-            std::vector<float> frame(frame_data, frame_data + num_mel_bins);
-            features.push_back(std::move(frame));
-        }
-
-        return features;
-    }
-
-    std::vector<float> run_inference(const std::vector<float> &audio_chunk_16k)
-    {
-        std::vector<std::vector<float>> fbank_feats;
-        fbank_feats = compute_fbank_kaldi(audio_chunk_16k, RESAMPLE_RATE, FEAT_DIM);
-
-        if (fbank_feats.empty()) {
-            return {};
-        }
-
-        int T = fbank_feats.size();
-        std::vector<float> mat_flattened;
-        for (const auto &feat : fbank_feats) {
-            mat_flattened.insert(mat_flattened.end(), feat.begin(), feat.end());
-        }
-
-        std::vector<int64_t> input_shape = {1, static_cast<int64_t>(T), FEAT_DIM};
-        std::vector<int64_t> cache_shape = {1, 32, 88};
-
-        Ort::MemoryInfo memory_info = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-        Ort::Value input_tensor     = Ort::Value::CreateTensor<float>(
-            memory_info, mat_flattened.data(), mat_flattened.size(), input_shape.data(), input_shape.size());
-        Ort::Value cache_tensor = Ort::Value::CreateTensor<float>(memory_info, cache.data(), cache.size(),
-                                                                  cache_shape.data(), cache_shape.size());
-
-        const char *input_names[]  = {"input", "cache"};
-        const char *output_names[] = {"output", "r_cache"};
-
-        std::vector<Ort::Value> inputs;
-        inputs.push_back(std::move(input_tensor));
-        inputs.push_back(std::move(cache_tensor));
-
-        auto output_tensors = session->Run(Ort::RunOptions{nullptr}, input_names, inputs.data(), 2, output_names, 2);
-
-        float *out_data       = output_tensors[0].GetTensorMutableData<float>();
-        float *cache_out_data = output_tensors[1].GetTensorMutableData<float>();
-
-        std::vector<int64_t> out_shape = output_tensors[0].GetTensorTypeAndShapeInfo().GetShape();
-        size_t out_size                = 1;
-        for (auto dim : out_shape) out_size *= dim;
-
-        std::vector<float> out_chunk(out_data, out_data + out_size);
-        std::copy(cache_out_data, cache_out_data + cache.size(), cache.begin());
-
-        return out_chunk;
-    }
-
-    void sys_pcm_on_data(const std::string &raw)
-    {
-        static int count = 0;
-        if (count < delay_audio_frame_) {
-            buffer_write_char(pcmdata, raw.data(), raw.length());
-            count++;
-            return;
-        }
-        buffer_write_char(pcmdata, raw.data(), raw.length());
-        buffer_position_set(pcmdata, 0);
-
-        std::vector<float> floatSamples;
-        std::vector<int16_t> int16Samples;
-        {
-            int16_t audio_val;
-            while (buffer_read_i16(pcmdata, &audio_val, 1)) {
-                int16Samples.push_back(audio_val);
-                float normalizedSample = static_cast<float>(audio_val) / 1.0f;
-                float sample           = static_cast<float>(normalizedSample);
-                floatSamples.push_back(sample);
-            }
-        }
-
-        buffer_resize(pcmdata, 0);
-        count = 0;
-
-        auto scores = run_inference(floatSamples);
-        if (detect_wakeup(scores)) {
-            if (enwake_audio_ && (!wake_wav_file_.empty()) && play_awake_wav) {
-                play_awake_wav(wake_wav_file_);
-            }
-            if (out_callback_) out_callback_("", true);
-        }
-    }
-
-    void trigger()
-    {
-        if (out_callback_) out_callback_("", true);
-    }
-
-    bool delete_model()
-    {
-        return true;
-    }
-
-    llm_task(const std::string &workid) : audio_flage_(false)
-    {
-        pcmdata = buffer_create();
-    }
-
-    void start()
-    {
-    }
-
-    void stop()
-    {
-    }
-
-    ~llm_task()
-    {
-        stop();
-        buffer_destroy(pcmdata);
-    }
-};
-#undef CONFIG_AUTO_SET
-
-class llm_kws : public StackFlow {
-private:
-    enum { EVENT_TRIGGER = EVENT_EXPORT + 1 };
-    int task_count_;
-    std::string audio_url_;
-    std::unordered_map<int, std::shared_ptr<llm_task>> llm_task_;
-
-public:
-    llm_kws() : StackFlow("kws_new")
-    {
-        task_count_ = 1;
-        event_queue_.appendListener(EVENT_TRIGGER, std::bind(&llm_kws::trigger, this, std::placeholders::_1));
-        rpc_ctx_->register_rpc_action(
-            "trigger", [this](pzmq *_pzmq, const std::shared_ptr<StackFlows::pzmq_data> &data) -> std::string {
-                this->event_queue_.enqueue(EVENT_TRIGGER,
-                                           std::make_shared<stackflow_data>(data->get_param(0), data->get_param(1)));
-                return LLM_NONE;
-            });
-    }
-
-    void task_output(const std::weak_ptr<llm_task> llm_task_obj_weak,
-                     const std::weak_ptr<llm_channel_obj> llm_channel_weak, const std::string &data, bool finish)
-    {
-        auto llm_task_obj = llm_task_obj_weak.lock();
-        auto llm_channel  = llm_channel_weak.lock();
-        if (!(llm_task_obj && llm_channel)) {
-            return;
-        }
-        std::string tmp_msg1;
-        const std::string *next_data = &data;
-        if (data.empty()) {
-            llm_channel->send(llm_task_obj->response_format_, true, LLM_NO_ERROR);
-            return;
-        }
-        if (finish) {
-            tmp_msg1  = data + ".";
-            next_data = &tmp_msg1;
-        }
-        if (llm_channel->enstream_) {
-            static int count = 0;
-            nlohmann::json data_body;
-            data_body["index"]  = count++;
-            data_body["delta"]  = (*next_data);
-            data_body["finish"] = finish;
-            if (finish) count = 0;
-            SLOGI("send stream:%s", next_data->c_str());
-            llm_channel->send(llm_task_obj->response_format_, data_body, LLM_NO_ERROR);
-        } else if (finish) {
-            SLOGI("send utf-8:%s", next_data->c_str());
-            llm_channel->send(llm_task_obj->response_format_, (*next_data), LLM_NO_ERROR);
-        }
-    }
-
-    void play_awake_wav(const std::string &wav_file)
-    {
-        FILE *fp = fopen(wav_file.c_str(), "rb");
-        if (!fp) {
-            printf("Open %s failed!\n", wav_file.c_str());
-            return;
-        }
-        fseek(fp, 0, SEEK_END);
-        long size = ftell(fp);
-        fseek(fp, 0, SEEK_SET);
-        std::vector<char> wav_data(size);
-        fread(wav_data.data(), 1, wav_data.size(), fp);
-        fclose(fp);
-        int post = 0;
-        for (int i = 0; i < wav_data.size() - 4; i++) {
-            if ((wav_data[i] == 'd') && (wav_data[i + 1] == 'a') && (wav_data[i + 2] == 't') &&
-                (wav_data[i + 3] == 'a')) {
-                post = i + 8;
-                break;
-            }
-        }
-        if (post != 0) {
-            unit_call("audio", "play_raw", std::string((char *)(wav_data.data() + post), size - post));
-        }
-    }
-
-    void task_pause(const std::weak_ptr<llm_task> llm_task_obj_weak,
-                    const std::weak_ptr<llm_channel_obj> llm_channel_weak)
-    {
-        auto llm_task_obj = llm_task_obj_weak.lock();
-        auto llm_channel  = llm_channel_weak.lock();
-        if (!(llm_task_obj && llm_channel)) {
-            return;
-        }
-        if (llm_task_obj->audio_flage_) {
-            if (!audio_url_.empty()) llm_channel->stop_subscriber(audio_url_);
-            llm_task_obj->audio_flage_ = false;
-        }
-    }
-
-    void task_work(const std::weak_ptr<llm_task> llm_task_obj_weak,
-                   const std::weak_ptr<llm_channel_obj> llm_channel_weak)
-    {
-        auto llm_task_obj = llm_task_obj_weak.lock();
-        auto llm_channel  = llm_channel_weak.lock();
-        if (!(llm_task_obj && llm_channel)) {
-            return;
-        }
-        if ((!audio_url_.empty()) && (llm_task_obj->audio_flage_ == false)) {
-            std::weak_ptr<llm_task> _llm_task_obj = llm_task_obj;
-            llm_channel->subscriber(audio_url_, [_llm_task_obj](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
-                _llm_task_obj.lock()->sys_pcm_on_data(raw->string());
-            });
-            llm_task_obj->audio_flage_ = true;
-        }
-    }
-
-    void work(const std::string &work_id, const std::string &object, const std::string &data) override
-    {
-        SLOGI("llm_asr::work:%s", data.c_str());
-
-        nlohmann::json error_body;
-        int work_id_num = sample_get_work_id_num(work_id);
-        if (llm_task_.find(work_id_num) == llm_task_.end()) {
-            error_body["code"]    = -6;
-            error_body["message"] = "Unit Does Not Exist";
-            send("None", "None", error_body, work_id);
-            return;
-        }
-        task_work(llm_task_[work_id_num], get_channel(work_id_num));
-        send("None", "None", LLM_NO_ERROR, work_id);
-    }
-
-    void pause(const std::string &work_id, const std::string &object, const std::string &data) override
-    {
-        SLOGI("llm_asr::work:%s", data.c_str());
-
-        nlohmann::json error_body;
-        int work_id_num = sample_get_work_id_num(work_id);
-        if (llm_task_.find(work_id_num) == llm_task_.end()) {
-            error_body["code"]    = -6;
-            error_body["message"] = "Unit Does Not Exist";
-            send("None", "None", error_body, work_id);
-            return;
-        }
-        task_pause(llm_task_[work_id_num], get_channel(work_id_num));
-        send("None", "None", LLM_NO_ERROR, work_id);
-    }
-
-    void task_user_data(const std::weak_ptr<llm_task> llm_task_obj_weak,
-                        const std::weak_ptr<llm_channel_obj> llm_channel_weak, const std::string &object,
-                        const std::string &data)
-    {
-        nlohmann::json error_body;
-        auto llm_task_obj = llm_task_obj_weak.lock();
-        auto llm_channel  = llm_channel_weak.lock();
-        if (!(llm_task_obj && llm_channel)) {
-            error_body["code"]    = -11;
-            error_body["message"] = "Model run failed.";
-            send("None", "None", error_body, unit_name_);
-            return;
-        }
-        std::string tmp_msg1;
-        const std::string *next_data = &data;
-        int ret;
-        if (object.find("stream") != std::string::npos) {
-            static std::unordered_map<int, std::string> stream_buff;
-            try {
-                if (decode_stream(data, tmp_msg1, stream_buff)) {
-                    return;
-                };
-            } catch (...) {
-                stream_buff.clear();
-                error_body["code"]    = -25;
-                error_body["message"] = "Stream data index error.";
-                send("None", "None", error_body, unit_name_);
-                return;
-            }
-            next_data = &tmp_msg1;
-        }
-        std::string tmp_msg2;
-        if (object.find("base64") != std::string::npos) {
-            ret = decode_base64((*next_data), tmp_msg2);
-            if (ret == -1) {
-                error_body["code"]    = -23;
-                error_body["message"] = "Base64 decoding error.";
-                send("None", "None", error_body, unit_name_);
-                return;
-            }
-            next_data = &tmp_msg2;
-        }
-        llm_task_obj->sys_pcm_on_data((*next_data));
-    }
-
-    int setup(const std::string &work_id, const std::string &object, const std::string &data) override
-    {
-        nlohmann::json error_body;
-        if ((llm_task_channel_.size() - 1) == task_count_) {
-            error_body["code"]    = -21;
-            error_body["message"] = "task full";
-            send("None", "None", error_body, "kws");
-            return -1;
-        }
-
-        int work_id_num                             = sample_get_work_id_num(work_id);
-        auto llm_channel                            = get_channel(work_id);
-        auto llm_task_obj                           = std::make_shared<llm_task>(work_id);
-        std::weak_ptr<llm_task> _llm_task_obj       = llm_task_obj;
-        std::weak_ptr<llm_channel_obj> _llm_channel = llm_channel;
-        nlohmann::json config_body;
-        try {
-            config_body = nlohmann::json::parse(data);
-        } catch (...) {
-            SLOGE("setup json format error.");
-            error_body["code"]    = -2;
-            error_body["message"] = "json format error.";
-            send("None", "None", error_body, "kws");
-            return -2;
-        }
-        int ret = llm_task_obj->load_model(config_body);
-        if (ret == 0) {
-            llm_channel->set_output(llm_task_obj->enoutput_);
-            llm_channel->set_stream(llm_task_obj->enstream_);
-            llm_task_obj->play_awake_wav = std::bind(&llm_kws::play_awake_wav, this, std::placeholders::_1);
-            llm_task_obj->set_output(std::bind(&llm_kws::task_output, this, std::weak_ptr<llm_task>(llm_task_obj),
-                                               std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
-                                               std::placeholders::_2));
-
-            for (const auto input : llm_task_obj->inputs_) {
-                if (input.find("sys") != std::string::npos) {
-                    audio_url_ = unit_call("audio", "cap", "None");
-                    llm_channel->subscriber(audio_url_,
-                                            [_llm_task_obj](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
-                                                auto llm_task_obj = _llm_task_obj.lock();
-                                                if (llm_task_obj) llm_task_obj->sys_pcm_on_data(raw->string());
-                                            });
-                    llm_task_obj->audio_flage_ = true;
-                } else if (input.find("kws") != std::string::npos) {
-                    llm_task_obj->delay_audio_frame_ = 0;
-                    llm_channel->subscriber_work_id(
-                        "", std::bind(&llm_kws::task_user_data, this, std::weak_ptr<llm_task>(llm_task_obj),
-                                      std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1,
-                                      std::placeholders::_2));
-                }
-            }
-            llm_task_[work_id_num] = llm_task_obj;
-            SLOGI("load_mode success");
-            send("None", "None", LLM_NO_ERROR, work_id);
-            return 0;
-        } else {
-            SLOGE("load_mode Failed");
-            error_body["code"]    = -5;
-            error_body["message"] = "Model loading failed.";
-            send("None", "None", error_body, "kws");
-            return -1;
-        }
-    }
-
-    void taskinfo(const std::string &work_id, const std::string &object, const std::string &data) override
-    {
-        SLOGI("llm_kws::taskinfo:%s", data.c_str());
-        nlohmann::json req_body;
-        int work_id_num = sample_get_work_id_num(work_id);
-        if (WORK_ID_NONE == work_id_num) {
-            std::vector<std::string> task_list;
-            std::transform(llm_task_channel_.begin(), llm_task_channel_.end(), std::back_inserter(task_list),
-                           [](const auto task_channel) { return task_channel.second->work_id_; });
-            req_body = task_list;
-            send("kws.tasklist", req_body, LLM_NO_ERROR, work_id);
-        } else {
-            if (llm_task_.find(work_id_num) == llm_task_.end()) {
-                req_body["code"]    = -6;
-                req_body["message"] = "Unit Does Not Exist";
-                send("None", "None", req_body, work_id);
-                return;
-            }
-            auto llm_task_obj           = llm_task_[work_id_num];
-            req_body["model"]           = llm_task_obj->model_;
-            req_body["response_format"] = llm_task_obj->response_format_;
-            req_body["enoutput"]        = llm_task_obj->enoutput_;
-            req_body["inputs"]          = llm_task_obj->inputs_;
-            send("kws.taskinfo", req_body, LLM_NO_ERROR, work_id);
-        }
-    }
-
-    int exit(const std::string &work_id, const std::string &object, const std::string &data) override
-    {
-        SLOGI("llm_kws::exit:%s", data.c_str());
-        nlohmann::json error_body;
-        int work_id_num = sample_get_work_id_num(work_id);
-        if (llm_task_.find(work_id_num) == llm_task_.end()) {
-            error_body["code"]    = -6;
-            error_body["message"] = "Unit Does Not Exist";
-            send("None", "None", error_body, work_id);
-            return -1;
-        }
-        llm_task_[work_id_num]->stop();
-        auto llm_channel = get_channel(work_id_num);
-        llm_channel->stop_subscriber("");
-        if (llm_task_[work_id_num]->audio_flage_) {
-            unit_call("audio", "cap_stop", "None");
-        }
-        llm_task_.erase(work_id_num);
-        send("None", "None", LLM_NO_ERROR, work_id);
-        return 0;
-    }
-
-    std::string trigger(const std::shared_ptr<void> &arg)
-    {
-        std::shared_ptr<stackflow_data> originalPtr = std::static_pointer_cast<stackflow_data>(arg);
-        std::string zmq_url                         = originalPtr->string(0);
-        std::string data                            = originalPtr->string(1);
-        std::string work_id                         = sample_json_str_get(data, "work_id");
-        if (work_id.length() == 0) {
-            nlohmann::json out_body;
-            out_body["request_id"]       = sample_json_str_get(data, "request_id");
-            out_body["work_id"]          = "kws";
-            out_body["created"]          = time(NULL);
-            out_body["object"]           = "";
-            out_body["data"]             = "";
-            out_body["error"]["code"]    = -2;
-            out_body["error"]["message"] = "json format error.";
-            pzmq _zmq(zmq_url, ZMQ_PUSH);
-            std::string out = out_body.dump();
-            out += "\n";
-            _zmq.send_data(out);
-            return LLM_NONE;
-        }
-
-        int work_id_num = sample_get_work_id_num(work_id);
-        if (llm_task_.find(work_id_num) == llm_task_.end()) {
-            nlohmann::json out_body;
-            out_body["request_id"]       = sample_json_str_get(data, "request_id");
-            out_body["work_id"]          = "kws";
-            out_body["created"]          = time(NULL);
-            out_body["object"]           = "";
-            out_body["data"]             = "";
-            out_body["error"]["code"]    = -6;
-            out_body["error"]["message"] = "Unit Does Not Exist";
-            pzmq _zmq(zmq_url, ZMQ_PUSH);
-            std::string out = out_body.dump();
-            out += "\n";
-            _zmq.send_data(out);
-            return LLM_NONE;
-        }
-        llm_task_[work_id_num]->trigger();
-        return LLM_NONE;
-    }
-
-    ~llm_kws()
-    {
-        while (1) {
-            auto iteam = llm_task_.begin();
-            if (iteam == llm_task_.end()) {
-                break;
-            }
-            iteam->second->stop();
-            if (iteam->second->audio_flage_) {
-                unit_call("audio", "cap_stop", "None");
-            }
-            get_channel(iteam->first)->stop_subscriber("");
-            iteam->second.reset();
-            llm_task_.erase(iteam->first);
-        }
-    }
-};
-
-int main(int argc, char *argv[])
-{
-    signal(SIGTERM, __sigint);
-    signal(SIGINT, __sigint);
-    mkdir("/tmp/llm", 0777);
-    llm_kws llm;
-    while (!main_exit_flage) {
-        sleep(1);
-    }
-    return 0;
-}
\ No newline at end of file

From cc9d1bc6bcca23693feb2ed73d317a6492d7d95f Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 18 Dec 2025 10:53:57 +0800
Subject: [PATCH 67/79] [update] perf llm backend & add c tokenizer

---
 .../llm_cosyvoice2.md                         | 223 +++++++
 ext_components/tokenizer/Kconfig              |   5 +
 ext_components/tokenizer/SConstruct           |  52 ++
 .../main_cosy_voice/src/runner/LLM.hpp        |   6 +-
 .../main_cosy_voice/src/runner/Token2wav.hpp  |  18 +-
 .../ax_model_runner/ax_model_runner.hpp       | 205 ++++---
 .../ax_model_runner/ax_model_runner_ax650.cpp | 547 +++++++----------
 .../ax_model_runner/ax_model_runner_ax650.hpp |  17 +-
 .../llm_framework/main_llm/src/runner/LLM.hpp |   8 +-
 .../ax_model_runner/ax_model_runner.hpp       | 205 ++++---
 .../ax_model_runner/ax_model_runner_ax650.cpp | 549 +++++++----------
 .../ax_model_runner/ax_model_runner_ax650.hpp |  17 +-
 .../llm_framework/main_vlm/src/runner/LLM.hpp |  20 +-
 .../ax_model_runner/ax_model_runner.hpp       | 205 ++++---
 .../ax_model_runner/ax_model_runner_ax650.cpp | 552 +++++++-----------
 .../ax_model_runner/ax_model_runner_ax650.hpp |  17 +-
 16 files changed, 1276 insertions(+), 1370 deletions(-)
 create mode 100644 doc/projects_llm_framework_doc/llm_cosyvoice2.md
 create mode 100644 ext_components/tokenizer/Kconfig
 create mode 100644 ext_components/tokenizer/SConstruct

diff --git a/doc/projects_llm_framework_doc/llm_cosyvoice2.md b/doc/projects_llm_framework_doc/llm_cosyvoice2.md
new file mode 100644
index 00000000..7f3a7ab6
--- /dev/null
+++ b/doc/projects_llm_framework_doc/llm_cosyvoice2.md
@@ -0,0 +1,223 @@
+# llm_cosy_voice
+
+使用 npu 加速的文字转语音单元，用于提供文字转语音服务，可使用语音克隆，用于提供多语言转语音服务。
+
+## setup
+
+配置单元工作。
+
+发送 json：
+
+```json
+cosy_voice
+{
+  "request_id": "2",
+  "work_id": "cosy_voice",
+  "action": "setup",
+  "object": "cosy_voice.setup",
+  "data": {
+    "model": "CosyVoice2-0.5B-ax650",
+    "response_format": "file",
+    "input": "tts.utf-8",
+    "enoutput": false
+  }
+}
+```
+
+
+- request_id：参考基本数据解释。
+- work_id：配置单元时，为 `cosy_voice`。
+- action：调用的方法为 `setup`。
+- object：传输的数据类型为 `cosy_voice.setup`。
+- model：使用的模型为 `CosyVoice2-0.5B-ax650` 模型。
+- prompt_files：要克隆的音频信息文件。
+- response_format：返回结果为 `sys.pcm`, 系统音频数据，并直接发送到 llm-audio 模块进行播放。返回结果为 `file`, 生成的音频写 wav 文件，可用 `prompt_data` 指定路径或文件名。
+- input：输入的为 `tts.utf-8`,代表的是从用户输入。
+- enoutput：是否起用用户结果输出。
+
+响应 json：
+
+```json
+{
+    "created": 1761791627,
+    "data": "None",
+    "error": {
+        "code": 0,
+        "message": ""
+    },
+    "object": "None",
+    "request_id": "2",
+    "work_id": "cosy_voice.1000"
+}
+```
+
+- created：消息创建时间，unix 时间。
+- work_id：返回成功创建的 work_id 单元。
+
+## inference
+
+### 流式输入
+
+```json
+{
+    "request_id": "2",
+    "work_id": "cosy_voice.1000",
+    "action": "inference",
+    "object": "cosy_voice.utf-8.stream",
+    "data": {
+        "delta": "今天天气真好！",
+        "index": 0,
+        "finish": true
+    }
+}
+```
+- object：传输的数据类型为 `cosy_voice.utf-8.stream` 代表的是从用户 utf-8 的流式输入
+- delta：流式输入的分段数据
+- index：流式输入的分段索引
+- finish:流式输入是否完成的标志位
+
+### 非流式输入
+
+```json
+{
+    "request_id": "2",
+    "work_id": "cosy_voice.1000",
+    "action": "inference",
+    "object": "cosy_voice.utf-8",
+    "data": "今天天气真好！"
+}
+```
+- object：传输的数据类型为 `cosy_voice.utf-8` 代表的是从用户 utf-8 的非流式输入
+- data：非流式输入的数据
+
+## pause
+
+暂停单元工作。
+
+发送 json：
+
+```json
+{
+  "request_id": "5",
+  "work_id": "cosy_voice.1000",
+  "action": "pause"
+}
+```
+
+响应 json：
+
+```json
+{
+    "created": 1761791706,
+    "data": "None",
+    "error": {
+        "code": 0,
+        "message": ""
+    },
+    "object": "None",
+    "request_id": "5",
+    "work_id": "cosy_voice.1000"
+}
+```
+
+error::code 为 0 表示执行成功。
+
+## exit
+
+单元退出。
+
+发送 json：
+
+```json
+{
+  "request_id": "7",
+  "work_id": "cosy_voice.1000",
+  "action": "exit"
+}
+```
+
+响应 json：
+
+```json
+{
+    "created": 1761791854,
+    "data": "None",
+    "error": {
+        "code": 0,
+        "message": ""
+    },
+    "object": "None",
+    "request_id": "7",
+    "work_id": "cosy_voice.1000"
+}
+```
+
+error::code 为 0 表示执行成功。
+
+## taskinfo
+
+获取任务列表。
+
+发送 json：
+
+```json
+{
+  "request_id": "2",
+  "work_id": "cosy_voice",
+  "action": "taskinfo"
+}
+```
+
+响应 json：
+
+```json
+{
+    "created": 1761791739,
+    "data": [
+        "cosy_voice.1000"
+    ],
+    "error": {
+        "code": 0,
+        "message": ""
+    },
+    "object": "llm.tasklist",
+    "request_id": "2",
+    "work_id": "cosy_voice"
+}
+```
+
+获取任务运行参数。
+
+```json
+{
+  "request_id": "2",
+  "work_id": "cosy_voice.1000",
+  "action": "taskinfo"
+}
+```
+
+响应 json：
+
+```json
+{
+    "created": 1761791761,
+    "data": {
+        "enoutput": false,
+        "inputs": [
+            "tts.utf-8"
+        ],
+        "model": "CosyVoice2-0.5B-ax650",
+        "response_format": "sys.pcm"
+    },
+    "error": {
+        "code": 0,
+        "message": ""
+    },
+    "object": "cosy_voice.taskinfo",
+    "request_id": "2",
+    "work_id": "cosy_voice.1000"
+}
+```
+
+> **注意：work_id 是按照单元的初始化注册顺序增加的，并不是固定的索引值。**  
+> **同类型单元不能配置多个单元同时工作，否则会产生未知错误。例如 tts 和 melo tts 不能同时拍起用工作。**
diff --git a/ext_components/tokenizer/Kconfig b/ext_components/tokenizer/Kconfig
new file mode 100644
index 00000000..1cf11388
--- /dev/null
+++ b/ext_components/tokenizer/Kconfig
@@ -0,0 +1,5 @@
+menuconfig AX_TOKENIZER_ENABLED
+    bool "Enable tokenizer support"
+    default n
+    help
+        enable tokenizer support
\ No newline at end of file
diff --git a/ext_components/tokenizer/SConstruct b/ext_components/tokenizer/SConstruct
new file mode 100644
index 00000000..e393f25f
--- /dev/null
+++ b/ext_components/tokenizer/SConstruct
@@ -0,0 +1,52 @@
+# component2/SConscript
+Import("env")
+import os
+from pathlib import Path
+
+with open(env["PROJECT_TOOL_S"]) as f:
+    exec(f.read())
+
+_SDK_PATH = os.path.normpath(
+    os.environ.get("SDK_PATH", str(Path(os.getcwd()) / ".." / ".."))
+)
+
+env["GIT_REPO_LISTS"]["tokenizer"] = {
+    "url": "https://github.com/ZHEQIUSHUI/tokenizer.git",
+    "commit": "83f41d4b5b9a135c167d44fcdf2a0c56ebacca6d",
+    "path": str(Path(_SDK_PATH) / "github_source" / "tokenizer"),
+}
+
+if "CONFIG_AX_TOKENIZER_ENABLED" in os.environ:
+    check_component("tokenizer")
+    SRCS = []
+    INCLUDE = []
+    PRIVATE_INCLUDE = []
+    REQUIREMENTS = []
+    STATIC_LIB = []
+    DYNAMIC_LIB = []
+    DEFINITIONS = []
+    DEFINITIONS_PRIVATE = []
+    LDFLAGS = []
+    LINK_SEARCH_PATH = []
+
+    INCLUDE += [
+        os.path.join(env["GIT_REPO_LISTS"]["tokenizer"]["path"], "include"),
+    ]
+    print("AX-TOKENIZER INCLUDE:", INCLUDE)
+
+    env["COMPONENTS"].append(
+        {
+            "target": os.path.basename(env["component_dir"]),
+            "SRCS": SRCS,
+            "INCLUDE": INCLUDE,
+            "PRIVATE_INCLUDE": PRIVATE_INCLUDE,
+            "REQUIREMENTS": REQUIREMENTS,
+            "STATIC_LIB": STATIC_LIB,
+            "DYNAMIC_LIB": DYNAMIC_LIB,
+            "DEFINITIONS": DEFINITIONS,
+            "DEFINITIONS_PRIVATE": DEFINITIONS_PRIVATE,
+            "LDFLAGS": LDFLAGS,
+            "LINK_SEARCH_PATH": LINK_SEARCH_PATH,
+            "REGISTER": "static",
+        }
+    )
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp b/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
index 52994ca6..bb187825 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
@@ -246,10 +246,10 @@ class LLM {
     void Deinit()
     {
         for (int i = 0; i < _attr.axmodel_num; i++) {
-            llama_layers[i].layer.release();
+            llama_layers[i].layer.deinit();
         }
-        llama_post.release();
-        llm_decoder.release();
+        llama_post.deinit();
+        llm_decoder.deinit();
         embed_selector.Deinit();
         llm_embed_selector.Deinit();
         speech_embed_selector.Deinit();
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp b/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp
index 0eb0450c..e1daffa5 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/Token2wav.hpp
@@ -145,15 +145,15 @@ class Token2Wav {
 
     void Deinit()
     {
-        flow_encoder_28.release();
-        flow_encoder_53.release();
-        flow_encoder_78.release();
-        flow_encoder_50_final.release();
-        flow_estimator_200.release();
-        flow_estimator_250.release();
-        flow_estimator_300.release();
-        hift_p2_50_first.release();
-        hift_p2_58.release();
+        flow_encoder_28.deinit();
+        flow_encoder_53.deinit();
+        flow_encoder_78.deinit();
+        flow_encoder_50_final.deinit();
+        flow_estimator_200.deinit();
+        flow_estimator_250.deinit();
+        flow_estimator_300.deinit();
+        hift_p2_50_first.deinit();
+        hift_p2_58.deinit();
         flow_embed_selector.Deinit();
     }
 
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner.hpp b/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner.hpp
index c9ed4fa3..c4bcb2ac 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner.hpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner.hpp
@@ -4,31 +4,7 @@
 #include <map>
 #include <stdexcept>
 
-typedef enum _color_space_e
-{
-    axdl_color_space_unknown,
-    axdl_color_space_nv12,
-    axdl_color_space_nv21,
-    axdl_color_space_bgr,
-    axdl_color_space_rgb,
-} ax_color_space_e;
-
-typedef struct _image_t
-{
-    unsigned long long int pPhy;
-    void *pVir;
-    unsigned int nSize;
-    unsigned int nWidth;
-    unsigned int nHeight;
-    ax_color_space_e eDtype;
-    union
-    {
-        int tStride_H, tStride_W, tStride_C;
-    };
-} ax_image_t;
-
-typedef struct
-{
+typedef struct {
     std::string sName;
     unsigned int nIdx;
     std::vector<unsigned int> vShape;
@@ -37,8 +13,7 @@ typedef struct
     void *pVirAddr;
 } ax_runner_tensor_t;
 
-class ax_runner_base
-{
+class ax_runner_base {
 protected:
     std::vector<ax_runner_tensor_t> moutput_tensors;
     std::vector<ax_runner_tensor_t> minput_tensors;
@@ -52,106 +27,124 @@ class ax_runner_base
     std::map<std::string, std::vector<ax_runner_tensor_t>> map_group_output_tensors;
     std::map<std::string, std::vector<ax_runner_tensor_t>> map_group_input_tensors;
 
+    void build_tensor_maps()
+    {
+        map_input_tensors.clear();
+        for (const auto &t : minput_tensors) map_input_tensors[t.sName] = t;
+
+        map_output_tensors.clear();
+        for (const auto &t : moutput_tensors) map_output_tensors[t.sName] = t;
+
+        map_group_input_tensors.clear();
+        for (const auto &grp : mgroup_input_tensors) {
+            for (const auto &t : grp) map_group_input_tensors[t.sName].push_back(t);
+        }
+
+        map_group_output_tensors.clear();
+        for (const auto &grp : mgroup_output_tensors) {
+            for (const auto &t : grp) map_group_output_tensors[t.sName].push_back(t);
+        }
+    }
+
 public:
+    virtual ~ax_runner_base()
+    {
+    }
+
     virtual int init(const char *model_file, bool use_mmap = false) = 0;
-    virtual int init(char *model_buffer, size_t model_size) = 0;
+    virtual int init(char *model_buffer, size_t model_size)         = 0;
+    virtual void deinit()                                           = 0;
 
-    virtual void deinit() = 0;
+    int get_num_inputs()
+    {
+        return minput_tensors.size();
+    };
+    int get_num_outputs()
+    {
+        return moutput_tensors.size();
+    };
+    int get_num_input_groups()
+    {
+        return mgroup_input_tensors.size();
+    };
+    int get_num_output_groups()
+    {
+        return mgroup_output_tensors.size();
+    };
 
-    int get_num_inputs() { return minput_tensors.size(); };
-    int get_num_outputs() { return moutput_tensors.size(); };
+    const ax_runner_tensor_t &get_input(int idx)
+    {
+        return minput_tensors[idx];
+    }
+    const ax_runner_tensor_t *get_inputs_ptr()
+    {
+        return minput_tensors.data();
+    }
 
-    int get_num_input_groups() { return mgroup_input_tensors.size(); };
-    int get_num_output_groups() { return mgroup_output_tensors.size(); };
+    const ax_runner_tensor_t &get_input(const std::string &name)
+    {
+        auto it = map_input_tensors.find(name);
+        if (it == map_input_tensors.end()) throw std::runtime_error("input tensor not found: " + name);
+        return it->second;
+    }
 
-    const ax_runner_tensor_t &get_input(int idx) { return minput_tensors[idx]; }
-    const ax_runner_tensor_t *get_inputs_ptr() { return minput_tensors.data(); }
-    const ax_runner_tensor_t &get_input(std::string name)
+    const ax_runner_tensor_t &get_input(int grpid, int idx)
     {
-        if (map_input_tensors.size() == 0)
-        {
-            for (size_t i = 0; i < minput_tensors.size(); i++)
-            {
-                map_input_tensors[minput_tensors[i].sName] = minput_tensors[i];
-            }
-        }
-        if (map_input_tensors.find(name) == map_input_tensors.end())
-        {
-            throw std::runtime_error("input tensor not found: " + name);
-        }
+        return mgroup_input_tensors[grpid][idx];
+    }
+    const ax_runner_tensor_t *get_inputs_ptr(int grpid)
+    {
+        return mgroup_input_tensors[grpid].data();
+    }
 
-        return map_input_tensors[name];
+    const ax_runner_tensor_t &get_input(int grpid, const std::string &name)
+    {
+        auto it = map_group_input_tensors.find(name);
+        if (it == map_group_input_tensors.end()) throw std::runtime_error("input tensor not found: " + name);
+        if (grpid < 0 || grpid >= (int)it->second.size())
+            throw std::runtime_error("group id out of range for: " + name);
+        return it->second[grpid];
     }
 
-    const ax_runner_tensor_t &get_input(int grpid, int idx) { return mgroup_input_tensors[grpid][idx]; }
-    const ax_runner_tensor_t *get_inputs_ptr(int grpid) { return mgroup_input_tensors[grpid].data(); }
-    const ax_runner_tensor_t &get_input(int grpid, std::string name)
-    {
-        if (map_group_input_tensors.size() == 0)
-        {
-            for (size_t i = 0; i < mgroup_input_tensors.size(); i++)
-            {
-                for (size_t j = 0; j < mgroup_input_tensors[i].size(); j++)
-                {
-                    map_group_input_tensors[mgroup_input_tensors[i][j].sName].push_back(mgroup_input_tensors[i][j]);
-                }
-            }
-        }
-        if (map_group_input_tensors.find(name) == map_group_input_tensors.end())
-        {
-            throw std::runtime_error("input tensor not found: " + name);
-        }
-        return map_group_input_tensors[name][grpid];
-        // return map_input_tensors[name];
+    const ax_runner_tensor_t &get_output(int idx)
+    {
+        return moutput_tensors[idx];
+    }
+    const ax_runner_tensor_t *get_outputs_ptr()
+    {
+        return moutput_tensors.data();
     }
 
-    const ax_runner_tensor_t &get_output(int idx) { return moutput_tensors[idx]; }
-    const ax_runner_tensor_t *get_outputs_ptr() { return moutput_tensors.data(); }
-    const ax_runner_tensor_t &get_output(std::string name)
+    const ax_runner_tensor_t &get_output(const std::string &name)
     {
-        if (map_output_tensors.size() == 0)
-        {
-            for (size_t i = 0; i < moutput_tensors.size(); i++)
-            {
-                map_output_tensors[moutput_tensors[i].sName] = moutput_tensors[i];
-            }
-        }
-        if (map_output_tensors.find(name) == map_output_tensors.end())
-        {
-            throw std::runtime_error("output tensor not found: " + name);
-        }
+        auto it = map_output_tensors.find(name);
+        if (it == map_output_tensors.end()) throw std::runtime_error("output tensor not found: " + name);
+        return it->second;
+    }
 
-        return map_output_tensors[name];
+    const ax_runner_tensor_t &get_output(int grpid, int idx)
+    {
+        return mgroup_output_tensors[grpid][idx];
+    }
+    const ax_runner_tensor_t *get_outputs_ptr(int grpid)
+    {
+        return mgroup_output_tensors[grpid].data();
     }
 
-    const ax_runner_tensor_t &get_output(int grpid, int idx) { return mgroup_output_tensors[grpid][idx]; }
-    const ax_runner_tensor_t *get_outputs_ptr(int grpid) { return mgroup_output_tensors[grpid].data(); }
-    const ax_runner_tensor_t &get_output(int grpid, std::string name)
-    {
-        if (map_group_output_tensors.size() == 0)
-        {
-            for (size_t i = 0; i < mgroup_output_tensors.size(); i++)
-            {
-                for (size_t j = 0; j < mgroup_output_tensors[i].size(); j++)
-                {
-                    map_group_output_tensors[mgroup_output_tensors[i][j].sName].push_back(mgroup_output_tensors[i][j]);
-                }
-            }
-        }
-        if (map_group_output_tensors.find(name) == map_group_output_tensors.end())
-        {
-            throw std::runtime_error("input tensor not found: " + name);
-        }
-        return map_group_output_tensors[name][grpid];
+    const ax_runner_tensor_t &get_output(int grpid, const std::string &name)
+    {
+        auto it = map_group_output_tensors.find(name);
+        if (it == map_group_output_tensors.end()) throw std::runtime_error("output tensor not found: " + name);
+        if (grpid < 0 || grpid >= (int)it->second.size())
+            throw std::runtime_error("group id out of range for: " + name);
+        return it->second[grpid];
     }
 
-    virtual int inference() = 0;
+    virtual int inference()          = 0;
     virtual int inference(int grpid) = 0;
 
     int operator()()
     {
         return inference();
     }
-};
-
-// int ax_cmmcpy(unsigned long long int dst, unsigned long long int src, int size);
+};
\ No newline at end of file
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.cpp b/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.cpp
index a974f0cb..363c3d5c 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.cpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.cpp
@@ -1,8 +1,9 @@
 #include "ax_model_runner_ax650.hpp"
-#include "string.h"
-#include "fstream"
-#include "memory"
-// #include "utilities/file.hpp"
+#include <cstring>
+#include <fstream>
+#include <algorithm>
+#include <memory>
+#include <unordered_set>
 #include <ax_sys_api.h>
 #include <ax_ivps_api.h>
 #include <ax_engine_api.h>
@@ -11,332 +12,211 @@
 #include "sample_log.h"
 
 #define AX_CMM_ALIGN_SIZE 128
-
 const char *AX_CMM_SESSION_NAME = "npu";
 
-typedef enum
-{
+typedef enum {
     AX_ENGINE_ABST_DEFAULT = 0,
-    AX_ENGINE_ABST_CACHED = 1,
+    AX_ENGINE_ABST_CACHED  = 1,
 } AX_ENGINE_ALLOC_BUFFER_STRATEGY_T;
 
-typedef std::pair<AX_ENGINE_ALLOC_BUFFER_STRATEGY_T, AX_ENGINE_ALLOC_BUFFER_STRATEGY_T> INPUT_OUTPUT_ALLOC_STRATEGY;
+struct ax_runner_ax650_handle_t {
+    AX_ENGINE_HANDLE handle     = nullptr;
+    AX_ENGINE_CONTEXT_T context = 0;
+    std::vector<AX_ENGINE_IO_INFO_T *> io_info;
+    std::vector<AX_ENGINE_IO_T> io_data;
+};
 
-static void print_io_info(AX_ENGINE_IO_INFO_T *io_info)
+static int prepare_io_struct_only(AX_ENGINE_IO_INFO_T *info, AX_ENGINE_IO_T *io_data)
 {
-    static std::map<AX_ENGINE_DATA_TYPE_T, const char *> data_type = {
-        {AX_ENGINE_DT_UNKNOWN, "UNKNOWN"},
-        {AX_ENGINE_DT_UINT8, "UINT8"},
-        {AX_ENGINE_DT_UINT16, "UINT16"},
-        {AX_ENGINE_DT_FLOAT32, "FLOAT32"},
-        {AX_ENGINE_DT_SINT16, "SINT16"},
-        {AX_ENGINE_DT_SINT8, "SINT8"},
-        {AX_ENGINE_DT_SINT32, "SINT32"},
-        {AX_ENGINE_DT_UINT32, "UINT32"},
-        {AX_ENGINE_DT_FLOAT64, "FLOAT64"},
-        {AX_ENGINE_DT_UINT10_PACKED, "UINT10_PACKED"},
-        {AX_ENGINE_DT_UINT12_PACKED, "UINT12_PACKED"},
-        {AX_ENGINE_DT_UINT14_PACKED, "UINT14_PACKED"},
-        {AX_ENGINE_DT_UINT16_PACKED, "UINT16_PACKED"},
-    };
-
-    static std::map<AX_ENGINE_COLOR_SPACE_T, const char *> color_type = {
-        {AX_ENGINE_CS_FEATUREMAP, "FEATUREMAP"},
-        {AX_ENGINE_CS_RAW8, "RAW8"},
-        {AX_ENGINE_CS_RAW10, "RAW10"},
-        {AX_ENGINE_CS_RAW12, "RAW12"},
-        {AX_ENGINE_CS_RAW14, "RAW14"},
-        {AX_ENGINE_CS_RAW16, "RAW16"},
-        {AX_ENGINE_CS_NV12, "NV12"},
-        {AX_ENGINE_CS_NV21, "NV21"},
-        {AX_ENGINE_CS_RGB, "RGB"},
-        {AX_ENGINE_CS_BGR, "BGR"},
-        {AX_ENGINE_CS_RGBA, "RGBA"},
-        {AX_ENGINE_CS_GRAY, "GRAY"},
-        {AX_ENGINE_CS_YUV444, "YUV444"},
-    };
-    printf("\ninput size: %d\n", io_info->nInputSize);
-    for (uint32_t i = 0; i < io_info->nInputSize; ++i)
-    {
-        // print shape info,like [batchsize x channel x height x width]
-        auto &info = io_info->pInputs[i];
-        printf("    name: \e[1;32m%8s", info.pName);
-
-        std::string dt = "unknown";
-        if (data_type.find(info.eDataType) != data_type.end())
-        {
-            dt = data_type[info.eDataType];
-            printf(" \e[1;34m[%s] ", dt.c_str());
-        }
-        else
-        {
-            printf(" \e[1;31m[%s] ", dt.c_str());
-        }
-
-        std::string ct = "unknown";
-        if (info.pExtraMeta && color_type.find(info.pExtraMeta->eColorSpace) != color_type.end())
-        {
-            ct = color_type[info.pExtraMeta->eColorSpace];
-            printf("\e[1;34m[%s]", ct.c_str());
-        }
-        else
-        {
-            printf("\e[1;31m[%s]", ct.c_str());
-        }
-        printf(" \n        \e[1;31m");
-
-        for (AX_U8 s = 0; s < info.nShapeSize; s++)
-        {
-            printf("%d", info.pShape[s]);
-            if (s != info.nShapeSize - 1)
-            {
-                printf(" x ");
-            }
-        }
-        printf("\e[0m\n\n");
-    }
+    memset(io_data, 0, sizeof(*io_data));
+    io_data->pInputs    = new AX_ENGINE_IO_BUFFER_T[info->nInputSize];
+    io_data->nInputSize = info->nInputSize;
+    memset(io_data->pInputs, 0, sizeof(AX_ENGINE_IO_BUFFER_T) * info->nInputSize);
 
-    printf("\noutput size: %d\n", io_info->nOutputSize);
-    for (uint32_t i = 0; i < io_info->nOutputSize; ++i)
-    {
-        // print shape info,like [batchsize x channel x height x width]
-        auto &info = io_info->pOutputs[i];
-        printf("    name: \e[1;32m%8s \e[1;34m[%s]\e[0m\n        \e[1;31m", info.pName, data_type[info.eDataType]);
-        for (AX_U8 s = 0; s < info.nShapeSize; s++)
-        {
-            printf("%d", info.pShape[s]);
-            if (s != info.nShapeSize - 1)
-            {
-                printf(" x ");
-            }
-        }
-        printf("\e[0m\n\n");
-    }
-}
+    io_data->pOutputs    = new AX_ENGINE_IO_BUFFER_T[info->nOutputSize];
+    io_data->nOutputSize = info->nOutputSize;
+    memset(io_data->pOutputs, 0, sizeof(AX_ENGINE_IO_BUFFER_T) * info->nOutputSize);
 
-void free_io_index(AX_ENGINE_IO_BUFFER_T *io_buf, int index)
-{
-    for (int i = 0; i < index; ++i)
-    {
-        AX_ENGINE_IO_BUFFER_T *pBuf = io_buf + i;
-        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
-    }
-}
+    for (uint i = 0; i < info->nInputSize; ++i) io_data->pInputs[i].nSize = info->pInputs[i].nSize;
+    for (uint i = 0; i < info->nOutputSize; ++i) io_data->pOutputs[i].nSize = info->pOutputs[i].nSize;
 
-void free_io(AX_ENGINE_IO_T *io)
-{
-    for (size_t j = 0; j < io->nInputSize; ++j)
-    {
-        AX_ENGINE_IO_BUFFER_T *pBuf = io->pInputs + j;
-        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
-    }
-    for (size_t j = 0; j < io->nOutputSize; ++j)
-    {
-        AX_ENGINE_IO_BUFFER_T *pBuf = io->pOutputs + j;
-        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
-    }
-    delete[] io->pInputs;
-    delete[] io->pOutputs;
+    return 0;
 }
 
-static inline int prepare_io(AX_ENGINE_IO_INFO_T *info, AX_ENGINE_IO_T *io_data, INPUT_OUTPUT_ALLOC_STRATEGY strategy)
+static int prepare_io_with_alloc(
+    AX_ENGINE_IO_INFO_T *info, AX_ENGINE_IO_T *io_data,
+    std::pair<AX_ENGINE_ALLOC_BUFFER_STRATEGY_T, AX_ENGINE_ALLOC_BUFFER_STRATEGY_T> strategy,
+    std::vector<std::string> skip_alloc_names = {})
 {
-    memset(io_data, 0, sizeof(*io_data));
-    io_data->pInputs = new AX_ENGINE_IO_BUFFER_T[info->nInputSize];
-    io_data->nInputSize = info->nInputSize;
-
-    auto ret = 0;
-    for (uint i = 0; i < info->nInputSize; ++i)
-    {
-        auto meta = info->pInputs[i];
-        auto buffer = &io_data->pInputs[i];
-        if (strategy.first == AX_ENGINE_ABST_CACHED)
-        {
-            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+    int ret = prepare_io_struct_only(info, io_data);
+    if (ret != 0) return ret;
+
+    for (uint i = 0; i < info->nInputSize; ++i) {
+        auto &buffer = io_data->pInputs[i];
+        if (std::find(skip_alloc_names.begin(), skip_alloc_names.end(), info->pInputs[i].pName) !=
+            skip_alloc_names.end()) {
+            continue;
         }
-        else
-        {
-            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        if (strategy.first == AX_ENGINE_ABST_CACHED) {
+            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer.phyAddr), &buffer.pVirAddr, buffer.nSize, AX_CMM_ALIGN_SIZE,
+                                        (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        } else {
+            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer.phyAddr), &buffer.pVirAddr, buffer.nSize, AX_CMM_ALIGN_SIZE,
+                                  (const AX_S8 *)(AX_CMM_SESSION_NAME));
         }
-
-        if (ret != 0)
-        {
-            free_io_index(io_data->pInputs, i);
-            fprintf(stderr, "Allocate input{%d} { phy: %p, vir: %p, size: %lu Bytes }. fail \n", i, (void *)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+        if (ret != 0) {
+            ALOGE("Alloc input[%d] failed", i);
             return ret;
         }
-        memset(buffer->pVirAddr, 0, meta.nSize);
-        // fprintf(stderr, "Allocate input{%d} { phy: %p, vir: %p, size: %lu Bytes }. \n", i, (void*)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+        memset(buffer.pVirAddr, 0, buffer.nSize);
     }
 
-    io_data->pOutputs = new AX_ENGINE_IO_BUFFER_T[info->nOutputSize];
-    io_data->nOutputSize = info->nOutputSize;
-    for (uint i = 0; i < info->nOutputSize; ++i)
-    {
-        auto meta = info->pOutputs[i];
-        auto buffer = &io_data->pOutputs[i];
-        buffer->nSize = meta.nSize;
-        if (strategy.second == AX_ENGINE_ABST_CACHED)
-        {
-            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+    for (uint i = 0; i < info->nOutputSize; ++i) {
+        auto &buffer = io_data->pOutputs[i];
+        if (std::find(skip_alloc_names.begin(), skip_alloc_names.end(), info->pOutputs[i].pName) !=
+            skip_alloc_names.end()) {
+            continue;
         }
-        else
-        {
-            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        if (strategy.second == AX_ENGINE_ABST_CACHED) {
+            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer.phyAddr), &buffer.pVirAddr, buffer.nSize, AX_CMM_ALIGN_SIZE,
+                                        (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        } else {
+            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer.phyAddr), &buffer.pVirAddr, buffer.nSize, AX_CMM_ALIGN_SIZE,
+                                  (const AX_S8 *)(AX_CMM_SESSION_NAME));
         }
-        if (ret != 0)
-        {
-            fprintf(stderr, "Allocate output{%d} { phy: %p, vir: %p, size: %lu Bytes }. fail \n", i, (void *)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
-            free_io_index(io_data->pInputs, io_data->nInputSize);
-            free_io_index(io_data->pOutputs, i);
+        if (ret != 0) {
+            ALOGE("Alloc output[%d] failed", i);
             return ret;
         }
-        memset(buffer->pVirAddr, 0, meta.nSize);
-        // fprintf(stderr, "Allocate output{%d} { phy: %p, vir: %p, size: %lu Bytes }.\n", i, (void*)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+        memset(buffer.pVirAddr, 0, buffer.nSize);
     }
-
     return 0;
 }
 
-struct ax_joint_runner_ax650_handle_t
-{
-    AX_ENGINE_HANDLE handle;
-    AX_ENGINE_CONTEXT_T context;
-    std::vector<AX_ENGINE_IO_INFO_T *> io_info;
-    std::vector<AX_ENGINE_IO_T> io_data;
-
-    // int algo_width, algo_height;
-    // int algo_colorformat;
-};
-
 int ax_runner_ax650::sub_init()
 {
-    // 4. create context
+    if (!m_handle) return -1;
+
     int ret = AX_ENGINE_CreateContext(m_handle->handle);
-    if (0 != ret)
-    {
-        ALOGE("AX_ENGINE_CreateContext");
-        return ret;
-    }
+    if (ret != 0) return ret;
+
     ret = AX_ENGINE_CreateContextV2(m_handle->handle, &m_handle->context);
-    if (0 != ret)
-    {
-        ALOGE("AX_ENGINE_CreateContextV2");
-        return ret;
-    }
-    // fprintf(stdout, "Engine creating context is done.\n");
+    if (ret != 0) return ret;
 
-    // 5. set io
     AX_U32 io_count = 0;
-    ret = AX_ENGINE_GetGroupIOInfoCount(m_handle->handle, &io_count);
-    if (0 != ret)
-    {
-        ALOGE("AX_ENGINE_GetGroupIOInfoCount");
-        return ret;
-    }
-    // ALOGI("io_count=%d", io_count);
+    ret             = AX_ENGINE_GetGroupIOInfoCount(m_handle->handle, &io_count);
+    if (ret != 0) return ret;
 
     m_handle->io_info.resize(io_count);
     m_handle->io_data.resize(io_count);
     mgroup_input_tensors.resize(io_count);
     mgroup_output_tensors.resize(io_count);
 
-    // fprintf(stdout, "Engine get io info is done. \n");
-
-    // 6. alloc io
-    if (!_parepare_io)
-    {
-        for (size_t grpid = 0; grpid < io_count; grpid++)
-        {
-            AX_ENGINE_IO_INFO_T *io_info = nullptr;
-            ret = AX_ENGINE_GetGroupIOInfo(m_handle->handle, grpid, &io_info);
-            if (0 != ret)
-            {
-                ALOGE("AX_ENGINE_GetIOInfo");
-                return ret;
-            }
-            // print_io_info(io_info);
-
-            m_handle->io_info[grpid] = io_info;
+    std::vector<std::string> skip_alloc_names = {"K_cache", "V_cache"};
+    for (size_t grpid = 0; grpid < io_count; grpid++) {
+        AX_ENGINE_IO_INFO_T *io_info = nullptr;
+        ret                          = AX_ENGINE_GetGroupIOInfo(m_handle->handle, grpid, &io_info);
+        if (ret != 0) return ret;
+        m_handle->io_info[grpid] = io_info;
+
+        if (grpid == 0) {
+            ret = prepare_io_with_alloc(io_info, &m_handle->io_data[grpid],
+                                        {AX_ENGINE_ABST_DEFAULT, AX_ENGINE_ABST_CACHED});
+        } else if (grpid == io_count - 1) {
+            ret = prepare_io_with_alloc(io_info, &m_handle->io_data[grpid],
+                                        {AX_ENGINE_ABST_DEFAULT, AX_ENGINE_ABST_CACHED}, skip_alloc_names);
+        } else {
+            ret = prepare_io_struct_only(io_info, &m_handle->io_data[grpid]);
+        }
+        if (ret != 0) return ret;
+    }
 
-            ret = prepare_io(m_handle->io_info[grpid], &m_handle->io_data[grpid], std::make_pair(AX_ENGINE_ABST_DEFAULT, AX_ENGINE_ABST_CACHED));
-            if (0 != ret)
-            {
-                ALOGE("prepare_io grpid=%d", grpid);
-                return ret;
+    if (io_count > 2) {
+        auto &first_io_data = m_handle->io_data[0];
+        auto &first_io_info = m_handle->io_info[0];
+        auto &last_io_data  = m_handle->io_data[io_count - 1];
+        auto &last_io_info  = m_handle->io_info[io_count - 1];
+        for (uint i = 0; i < last_io_data.nInputSize; ++i) {
+            if (std::find(skip_alloc_names.begin(), skip_alloc_names.end(), last_io_info->pInputs[i].pName) !=
+                skip_alloc_names.end()) {
+                for (uint j = 0; j < first_io_data.nInputSize; ++j) {
+                    if (first_io_info->pInputs[j].pName == last_io_info->pInputs[i].pName) {
+                        last_io_data.pInputs[i].phyAddr  = first_io_data.pInputs[j].phyAddr;
+                        last_io_data.pInputs[i].pVirAddr = first_io_data.pInputs[j].pVirAddr;
+                    }
+                }
             }
         }
 
-        for (size_t grpid = 0; grpid < io_count; grpid++)
-        {
+        for (size_t grpid = 1; grpid < io_count - 1; grpid++) {
             auto &io_info = m_handle->io_info[grpid];
             auto &io_data = m_handle->io_data[grpid];
-            for (size_t i = 0; i < io_info->nOutputSize; i++)
-            {
-                ax_runner_tensor_t tensor;
-                tensor.nIdx = i;
-                tensor.sName = std::string(io_info->pOutputs[i].pName);
-                tensor.nSize = io_info->pOutputs[i].nSize;
-                for (size_t j = 0; j < io_info->pOutputs[i].nShapeSize; j++)
-                {
-                    tensor.vShape.push_back(io_info->pOutputs[i].pShape[j]);
-                }
-                tensor.phyAddr = io_data.pOutputs[i].phyAddr;
-                tensor.pVirAddr = io_data.pOutputs[i].pVirAddr;
-                mgroup_output_tensors[grpid].push_back(tensor);
+
+            size_t min_inputs = std::min(io_info->nInputSize, last_io_data.nInputSize);
+            for (size_t i = 0; i < min_inputs; i++) {
+                io_data.pInputs[i].phyAddr  = last_io_data.pInputs[i].phyAddr;
+                io_data.pInputs[i].pVirAddr = last_io_data.pInputs[i].pVirAddr;
             }
 
-            for (size_t i = 0; i < io_info->nInputSize; i++)
-            {
-                ax_runner_tensor_t tensor;
-                tensor.nIdx = i;
-                tensor.sName = std::string(io_info->pInputs[i].pName);
-                tensor.nSize = io_info->pInputs[i].nSize;
-                for (size_t j = 0; j < io_info->pInputs[i].nShapeSize; j++)
-                {
-                    tensor.vShape.push_back(io_info->pInputs[i].pShape[j]);
-                }
-                tensor.phyAddr = io_data.pInputs[i].phyAddr;
-                tensor.pVirAddr = io_data.pInputs[i].pVirAddr;
-                mgroup_input_tensors[grpid].push_back(tensor);
+            size_t min_outputs = std::min(io_info->nOutputSize, last_io_data.nOutputSize);
+            for (size_t i = 0; i < min_outputs; i++) {
+                io_data.pOutputs[i].phyAddr  = last_io_data.pOutputs[i].phyAddr;
+                io_data.pOutputs[i].pVirAddr = last_io_data.pOutputs[i].pVirAddr;
             }
         }
+    }
 
-        moutput_tensors = mgroup_output_tensors[0];
-        minput_tensors = mgroup_input_tensors[0];
+    for (size_t grpid = 0; grpid < io_count; grpid++) {
+        auto &io_info = m_handle->io_info[grpid];
+        auto &io_data = m_handle->io_data[grpid];
+
+        for (size_t i = 0; i < io_info->nOutputSize; i++) {
+            ax_runner_tensor_t tensor;
+            tensor.nIdx     = i;
+            tensor.sName    = io_info->pOutputs[i].pName ? std::string(io_info->pOutputs[i].pName) : "";
+            tensor.nSize    = io_info->pOutputs[i].nSize;
+            tensor.phyAddr  = io_data.pOutputs[i].phyAddr;
+            tensor.pVirAddr = io_data.pOutputs[i].pVirAddr;
+            for (size_t j = 0; j < io_info->pOutputs[i].nShapeSize; j++) {
+                tensor.vShape.push_back(io_info->pOutputs[i].pShape[j]);
+            }
+            mgroup_output_tensors[grpid].push_back(tensor);
+        }
 
-        _parepare_io = true;
-    }
-    else
-    {
+        for (size_t i = 0; i < io_info->nInputSize; i++) {
+            ax_runner_tensor_t tensor;
+            tensor.nIdx     = i;
+            tensor.sName    = io_info->pInputs[i].pName ? std::string(io_info->pInputs[i].pName) : "";
+            tensor.nSize    = io_info->pInputs[i].nSize;
+            tensor.phyAddr  = io_data.pInputs[i].phyAddr;
+            tensor.pVirAddr = io_data.pInputs[i].pVirAddr;
+            for (size_t j = 0; j < io_info->pInputs[i].nShapeSize; j++) {
+                tensor.vShape.push_back(io_info->pInputs[i].pShape[j]);
+            }
+            mgroup_input_tensors[grpid].push_back(tensor);
+        }
     }
 
-    return ret;
+    if (!mgroup_output_tensors.empty()) moutput_tensors = mgroup_output_tensors[0];
+    if (!mgroup_input_tensors.empty()) minput_tensors = mgroup_input_tensors[0];
+
+    build_tensor_maps();
+
+    return 0;
 }
 
 int ax_runner_ax650::init(const char *model_file, bool use_mmap)
 {
-    if (use_mmap)
-    {
+    if (use_mmap) {
         MMap model_buffer(model_file);
-        if (!model_buffer.data())
-        {
-            ALOGE("mmap");
-            return -1;
-        }
+        if (!model_buffer.data()) return -1;
         auto ret = init((char *)model_buffer.data(), model_buffer.size());
         model_buffer.close_file();
         return ret;
-    }
-    else
-    {
-        char *model_buffer;
-        size_t len;
-        if (!read_file(model_file, &model_buffer, &len))
-        {
-            ALOGE("read_file");
-            return -1;
-        }
+    } else {
+        char *model_buffer = nullptr;
+        size_t len         = 0;
+        if (!read_file(model_file, &model_buffer, &len)) return -1;
         auto ret = init(model_buffer, len);
         delete[] model_buffer;
         return ret;
@@ -345,96 +225,84 @@ int ax_runner_ax650::init(const char *model_file, bool use_mmap)
 
 int ax_runner_ax650::init(char *model_buffer, size_t model_size)
 {
-    if (!m_handle)
-    {
-        m_handle = new ax_joint_runner_ax650_handle_t;
-    }
-
-    static bool b_init = false;
-    if (!b_init)
-    {
-        // 1. init engine
-        AX_ENGINE_NPU_ATTR_T npu_attr;
-        memset(&npu_attr, 0, sizeof(npu_attr));
-        npu_attr.eHardMode = AX_ENGINE_VIRTUAL_NPU_DISABLE;
-        AX_SYS_Init();
-        auto ret = AX_ENGINE_Init(&npu_attr);
-        if (0 != ret)
-        {
-            return ret;
-        }
-        b_init = true;
-    }
-
-    // 3. create handle
-
-    int ret = AX_ENGINE_CreateHandle(&m_handle->handle, model_buffer, model_size);
-    if (0 != ret)
-    {
-        ALOGE("AX_ENGINE_CreateHandle");
+    if (m_handle) deinit();
+    m_handle = new ax_runner_ax650_handle_t;
+    int ret  = AX_ENGINE_CreateHandle(&m_handle->handle, model_buffer, model_size);
+    if (0 != ret) {
+        ALOGE("AX_ENGINE_CreateHandle failed: 0x%x", ret);
+        delete m_handle;
+        m_handle = nullptr;
         return ret;
     }
-    // fprintf(stdout, "Engine creating handle is done.\n");
-
     return sub_init();
 }
 
-void ax_runner_ax650::release()
+void ax_runner_ax650::deinit()
 {
-    if (m_handle && m_handle->handle)
-    {
-        for (size_t i = 0; i < m_handle->io_data.size(); i++)
-        {
-            /* code */
-            free_io(&m_handle->io_data[i]);
+    if (!m_handle) return;
+
+    std::unordered_set<unsigned long> freed_phy_addrs;
+
+    for (size_t g = 0; g < m_handle->io_data.size(); ++g) {
+        auto &io = m_handle->io_data[g];
+
+        if (io.pInputs) {
+            for (size_t j = 0; j < io.nInputSize; ++j) {
+                AX_ENGINE_IO_BUFFER_T *pBuf = io.pInputs + j;
+                if (pBuf->phyAddr != 0) {
+                    if (freed_phy_addrs.find(pBuf->phyAddr) == freed_phy_addrs.end()) {
+                        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+                        freed_phy_addrs.insert(pBuf->phyAddr);
+                    }
+                }
+            }
+            delete[] io.pInputs;
+            io.pInputs = nullptr;
         }
 
-        AX_ENGINE_DestroyHandle(m_handle->handle);
-        m_handle->handle = nullptr;
+        if (io.pOutputs) {
+            for (size_t j = 0; j < io.nOutputSize; ++j) {
+                AX_ENGINE_IO_BUFFER_T *pBuf = io.pOutputs + j;
+                if (pBuf->phyAddr != 0) {
+                    if (freed_phy_addrs.find(pBuf->phyAddr) == freed_phy_addrs.end()) {
+                        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+                        freed_phy_addrs.insert(pBuf->phyAddr);
+                    }
+                }
+            }
+            delete[] io.pOutputs;
+            io.pOutputs = nullptr;
+        }
     }
 
-    if (m_handle)
-    {
-        delete m_handle;
-        m_handle = nullptr;
+    if (m_handle->handle) {
+        AX_ENGINE_DestroyHandle(m_handle->handle);
     }
 
+    delete m_handle;
+    m_handle = nullptr;
+
     moutput_tensors.clear();
     minput_tensors.clear();
     map_input_tensors.clear();
     map_output_tensors.clear();
-
     mgroup_output_tensors.clear();
     mgroup_input_tensors.clear();
     map_group_input_tensors.clear();
     map_group_output_tensors.clear();
-
-    // AX_ENGINE_Deinit();
 }
 
-void ax_runner_ax650::deinit()
+int ax_runner_ax650::inference()
 {
-    if (m_handle && m_handle->handle)
-    {
-        // free_io(&m_handle->io_data);
-        // mtensors.clear();
-        // minput_tensors.clear();
-        // map_input_tensors.clear();
-        // map_tensors.clear();
-        AX_ENGINE_DestroyHandle(m_handle->handle);
-        m_handle->handle = nullptr;
-        // delete m_handle;
-        // m_handle = nullptr;
+    if (!m_handle) return -1;
+    for (size_t i = 0; i < get_num_inputs(); i++) {
+        auto &tensor = get_input(i);
+        AX_SYS_MflushCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
     }
 
-    // AX_ENGINE_Deinit();
-}
-
-int ax_runner_ax650::inference()
-{
     int ret = AX_ENGINE_RunSync(m_handle->handle, &m_handle->io_data[0]);
-    for (size_t i = 0; i < get_num_outputs(); i++)
-    {
+
+    for (size_t i = 0; i < get_num_outputs(); i++) {
         auto &tensor = get_output(i);
         AX_SYS_MinvalidateCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
     }
@@ -443,11 +311,18 @@ int ax_runner_ax650::inference()
 
 int ax_runner_ax650::inference(int grpid)
 {
+    if (!m_handle) return -1;
+    if (grpid < 0 || grpid >= (int)m_handle->io_data.size()) return -1;
+
+    for (size_t i = 0; i < mgroup_input_tensors[grpid].size(); i++) {
+        auto &tensor = mgroup_input_tensors[grpid][i];
+        AX_SYS_MflushCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
+    }
+
     int ret = AX_ENGINE_RunGroupIOSync(m_handle->handle, m_handle->context, grpid, &m_handle->io_data[grpid]);
 
-    for (size_t i = 0; i < get_num_outputs(); i++)
-    {
-        auto &tensor = get_output(grpid, i);
+    for (size_t i = 0; i < mgroup_output_tensors[grpid].size(); i++) {
+        auto &tensor = mgroup_output_tensors[grpid][i];
         AX_SYS_MinvalidateCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
     }
     return ret;
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.hpp b/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.hpp
index 351fb04a..62b65380 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.hpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/ax_model_runner/ax_model_runner_ax650.hpp
@@ -1,20 +1,23 @@
 #pragma once
 #include "ax_model_runner.hpp"
 
-class ax_runner_ax650 : public ax_runner_base
-{
-protected:
-    struct ax_joint_runner_ax650_handle_t *m_handle = nullptr;
-
-    bool _parepare_io = false;
+struct ax_runner_ax650_handle_t;
 
+class ax_runner_ax650 : public ax_runner_base {
+protected:
+    struct ax_runner_ax650_handle_t *m_handle = nullptr;
     int sub_init();
 
 public:
+    ax_runner_ax650() = default;
+    virtual ~ax_runner_ax650()
+    {
+        deinit();
+    }
+
     int init(const char *model_file, bool use_mmap = false) override;
     int init(char *model_buffer, size_t model_size) override;
 
-    void release();
     void deinit() override;
 
     int inference() override;
diff --git a/projects/llm_framework/main_llm/src/runner/LLM.hpp b/projects/llm_framework/main_llm/src/runner/LLM.hpp
index 9de1f1fe..a5fb872f 100644
--- a/projects/llm_framework/main_llm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_llm/src/runner/LLM.hpp
@@ -243,9 +243,9 @@ class LLM {
     void Deinit()
     {
         for (int i = 0; i < _attr.axmodel_num; i++) {
-            llama_layers[i].layer.release();
+            llama_layers[i].layer.deinit();
         }
-        llama_post.release();
+        llama_post.deinit();
         embed_selector.Deinit();
     }
 
@@ -686,9 +686,9 @@ class LLM_CTX {
     void Deinit()
     {
         for (int i = 0; i < _attr.axmodel_num; i++) {
-            llama_layers[i].layer.release();
+            llama_layers[i].layer.deinit();
         }
-        llama_post.release();
+        llama_post.deinit();
         embed_selector.Deinit();
     }
 
diff --git a/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner.hpp b/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner.hpp
index c9ed4fa3..c4bcb2ac 100644
--- a/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner.hpp
+++ b/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner.hpp
@@ -4,31 +4,7 @@
 #include <map>
 #include <stdexcept>
 
-typedef enum _color_space_e
-{
-    axdl_color_space_unknown,
-    axdl_color_space_nv12,
-    axdl_color_space_nv21,
-    axdl_color_space_bgr,
-    axdl_color_space_rgb,
-} ax_color_space_e;
-
-typedef struct _image_t
-{
-    unsigned long long int pPhy;
-    void *pVir;
-    unsigned int nSize;
-    unsigned int nWidth;
-    unsigned int nHeight;
-    ax_color_space_e eDtype;
-    union
-    {
-        int tStride_H, tStride_W, tStride_C;
-    };
-} ax_image_t;
-
-typedef struct
-{
+typedef struct {
     std::string sName;
     unsigned int nIdx;
     std::vector<unsigned int> vShape;
@@ -37,8 +13,7 @@ typedef struct
     void *pVirAddr;
 } ax_runner_tensor_t;
 
-class ax_runner_base
-{
+class ax_runner_base {
 protected:
     std::vector<ax_runner_tensor_t> moutput_tensors;
     std::vector<ax_runner_tensor_t> minput_tensors;
@@ -52,106 +27,124 @@ class ax_runner_base
     std::map<std::string, std::vector<ax_runner_tensor_t>> map_group_output_tensors;
     std::map<std::string, std::vector<ax_runner_tensor_t>> map_group_input_tensors;
 
+    void build_tensor_maps()
+    {
+        map_input_tensors.clear();
+        for (const auto &t : minput_tensors) map_input_tensors[t.sName] = t;
+
+        map_output_tensors.clear();
+        for (const auto &t : moutput_tensors) map_output_tensors[t.sName] = t;
+
+        map_group_input_tensors.clear();
+        for (const auto &grp : mgroup_input_tensors) {
+            for (const auto &t : grp) map_group_input_tensors[t.sName].push_back(t);
+        }
+
+        map_group_output_tensors.clear();
+        for (const auto &grp : mgroup_output_tensors) {
+            for (const auto &t : grp) map_group_output_tensors[t.sName].push_back(t);
+        }
+    }
+
 public:
+    virtual ~ax_runner_base()
+    {
+    }
+
     virtual int init(const char *model_file, bool use_mmap = false) = 0;
-    virtual int init(char *model_buffer, size_t model_size) = 0;
+    virtual int init(char *model_buffer, size_t model_size)         = 0;
+    virtual void deinit()                                           = 0;
 
-    virtual void deinit() = 0;
+    int get_num_inputs()
+    {
+        return minput_tensors.size();
+    };
+    int get_num_outputs()
+    {
+        return moutput_tensors.size();
+    };
+    int get_num_input_groups()
+    {
+        return mgroup_input_tensors.size();
+    };
+    int get_num_output_groups()
+    {
+        return mgroup_output_tensors.size();
+    };
 
-    int get_num_inputs() { return minput_tensors.size(); };
-    int get_num_outputs() { return moutput_tensors.size(); };
+    const ax_runner_tensor_t &get_input(int idx)
+    {
+        return minput_tensors[idx];
+    }
+    const ax_runner_tensor_t *get_inputs_ptr()
+    {
+        return minput_tensors.data();
+    }
 
-    int get_num_input_groups() { return mgroup_input_tensors.size(); };
-    int get_num_output_groups() { return mgroup_output_tensors.size(); };
+    const ax_runner_tensor_t &get_input(const std::string &name)
+    {
+        auto it = map_input_tensors.find(name);
+        if (it == map_input_tensors.end()) throw std::runtime_error("input tensor not found: " + name);
+        return it->second;
+    }
 
-    const ax_runner_tensor_t &get_input(int idx) { return minput_tensors[idx]; }
-    const ax_runner_tensor_t *get_inputs_ptr() { return minput_tensors.data(); }
-    const ax_runner_tensor_t &get_input(std::string name)
+    const ax_runner_tensor_t &get_input(int grpid, int idx)
     {
-        if (map_input_tensors.size() == 0)
-        {
-            for (size_t i = 0; i < minput_tensors.size(); i++)
-            {
-                map_input_tensors[minput_tensors[i].sName] = minput_tensors[i];
-            }
-        }
-        if (map_input_tensors.find(name) == map_input_tensors.end())
-        {
-            throw std::runtime_error("input tensor not found: " + name);
-        }
+        return mgroup_input_tensors[grpid][idx];
+    }
+    const ax_runner_tensor_t *get_inputs_ptr(int grpid)
+    {
+        return mgroup_input_tensors[grpid].data();
+    }
 
-        return map_input_tensors[name];
+    const ax_runner_tensor_t &get_input(int grpid, const std::string &name)
+    {
+        auto it = map_group_input_tensors.find(name);
+        if (it == map_group_input_tensors.end()) throw std::runtime_error("input tensor not found: " + name);
+        if (grpid < 0 || grpid >= (int)it->second.size())
+            throw std::runtime_error("group id out of range for: " + name);
+        return it->second[grpid];
     }
 
-    const ax_runner_tensor_t &get_input(int grpid, int idx) { return mgroup_input_tensors[grpid][idx]; }
-    const ax_runner_tensor_t *get_inputs_ptr(int grpid) { return mgroup_input_tensors[grpid].data(); }
-    const ax_runner_tensor_t &get_input(int grpid, std::string name)
-    {
-        if (map_group_input_tensors.size() == 0)
-        {
-            for (size_t i = 0; i < mgroup_input_tensors.size(); i++)
-            {
-                for (size_t j = 0; j < mgroup_input_tensors[i].size(); j++)
-                {
-                    map_group_input_tensors[mgroup_input_tensors[i][j].sName].push_back(mgroup_input_tensors[i][j]);
-                }
-            }
-        }
-        if (map_group_input_tensors.find(name) == map_group_input_tensors.end())
-        {
-            throw std::runtime_error("input tensor not found: " + name);
-        }
-        return map_group_input_tensors[name][grpid];
-        // return map_input_tensors[name];
+    const ax_runner_tensor_t &get_output(int idx)
+    {
+        return moutput_tensors[idx];
+    }
+    const ax_runner_tensor_t *get_outputs_ptr()
+    {
+        return moutput_tensors.data();
     }
 
-    const ax_runner_tensor_t &get_output(int idx) { return moutput_tensors[idx]; }
-    const ax_runner_tensor_t *get_outputs_ptr() { return moutput_tensors.data(); }
-    const ax_runner_tensor_t &get_output(std::string name)
+    const ax_runner_tensor_t &get_output(const std::string &name)
     {
-        if (map_output_tensors.size() == 0)
-        {
-            for (size_t i = 0; i < moutput_tensors.size(); i++)
-            {
-                map_output_tensors[moutput_tensors[i].sName] = moutput_tensors[i];
-            }
-        }
-        if (map_output_tensors.find(name) == map_output_tensors.end())
-        {
-            throw std::runtime_error("output tensor not found: " + name);
-        }
+        auto it = map_output_tensors.find(name);
+        if (it == map_output_tensors.end()) throw std::runtime_error("output tensor not found: " + name);
+        return it->second;
+    }
 
-        return map_output_tensors[name];
+    const ax_runner_tensor_t &get_output(int grpid, int idx)
+    {
+        return mgroup_output_tensors[grpid][idx];
+    }
+    const ax_runner_tensor_t *get_outputs_ptr(int grpid)
+    {
+        return mgroup_output_tensors[grpid].data();
     }
 
-    const ax_runner_tensor_t &get_output(int grpid, int idx) { return mgroup_output_tensors[grpid][idx]; }
-    const ax_runner_tensor_t *get_outputs_ptr(int grpid) { return mgroup_output_tensors[grpid].data(); }
-    const ax_runner_tensor_t &get_output(int grpid, std::string name)
-    {
-        if (map_group_output_tensors.size() == 0)
-        {
-            for (size_t i = 0; i < mgroup_output_tensors.size(); i++)
-            {
-                for (size_t j = 0; j < mgroup_output_tensors[i].size(); j++)
-                {
-                    map_group_output_tensors[mgroup_output_tensors[i][j].sName].push_back(mgroup_output_tensors[i][j]);
-                }
-            }
-        }
-        if (map_group_output_tensors.find(name) == map_group_output_tensors.end())
-        {
-            throw std::runtime_error("input tensor not found: " + name);
-        }
-        return map_group_output_tensors[name][grpid];
+    const ax_runner_tensor_t &get_output(int grpid, const std::string &name)
+    {
+        auto it = map_group_output_tensors.find(name);
+        if (it == map_group_output_tensors.end()) throw std::runtime_error("output tensor not found: " + name);
+        if (grpid < 0 || grpid >= (int)it->second.size())
+            throw std::runtime_error("group id out of range for: " + name);
+        return it->second[grpid];
     }
 
-    virtual int inference() = 0;
+    virtual int inference()          = 0;
     virtual int inference(int grpid) = 0;
 
     int operator()()
     {
         return inference();
     }
-};
-
-// int ax_cmmcpy(unsigned long long int dst, unsigned long long int src, int size);
+};
\ No newline at end of file
diff --git a/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner_ax650.cpp b/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner_ax650.cpp
index 374261b6..363c3d5c 100644
--- a/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner_ax650.cpp
+++ b/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner_ax650.cpp
@@ -1,8 +1,9 @@
 #include "ax_model_runner_ax650.hpp"
-#include "string.h"
-#include "fstream"
-#include "memory"
-// #include "utilities/file.hpp"
+#include <cstring>
+#include <fstream>
+#include <algorithm>
+#include <memory>
+#include <unordered_set>
 #include <ax_sys_api.h>
 #include <ax_ivps_api.h>
 #include <ax_engine_api.h>
@@ -11,332 +12,211 @@
 #include "sample_log.h"
 
 #define AX_CMM_ALIGN_SIZE 128
-
 const char *AX_CMM_SESSION_NAME = "npu";
 
-typedef enum
-{
+typedef enum {
     AX_ENGINE_ABST_DEFAULT = 0,
-    AX_ENGINE_ABST_CACHED = 1,
+    AX_ENGINE_ABST_CACHED  = 1,
 } AX_ENGINE_ALLOC_BUFFER_STRATEGY_T;
 
-typedef std::pair<AX_ENGINE_ALLOC_BUFFER_STRATEGY_T, AX_ENGINE_ALLOC_BUFFER_STRATEGY_T> INPUT_OUTPUT_ALLOC_STRATEGY;
+struct ax_runner_ax650_handle_t {
+    AX_ENGINE_HANDLE handle     = nullptr;
+    AX_ENGINE_CONTEXT_T context = 0;
+    std::vector<AX_ENGINE_IO_INFO_T *> io_info;
+    std::vector<AX_ENGINE_IO_T> io_data;
+};
 
-static void print_io_info(AX_ENGINE_IO_INFO_T *io_info)
+static int prepare_io_struct_only(AX_ENGINE_IO_INFO_T *info, AX_ENGINE_IO_T *io_data)
 {
-    static std::map<AX_ENGINE_DATA_TYPE_T, const char *> data_type = {
-        {AX_ENGINE_DT_UNKNOWN, "UNKNOWN"},
-        {AX_ENGINE_DT_UINT8, "UINT8"},
-        {AX_ENGINE_DT_UINT16, "UINT16"},
-        {AX_ENGINE_DT_FLOAT32, "FLOAT32"},
-        {AX_ENGINE_DT_SINT16, "SINT16"},
-        {AX_ENGINE_DT_SINT8, "SINT8"},
-        {AX_ENGINE_DT_SINT32, "SINT32"},
-        {AX_ENGINE_DT_UINT32, "UINT32"},
-        {AX_ENGINE_DT_FLOAT64, "FLOAT64"},
-        {AX_ENGINE_DT_UINT10_PACKED, "UINT10_PACKED"},
-        {AX_ENGINE_DT_UINT12_PACKED, "UINT12_PACKED"},
-        {AX_ENGINE_DT_UINT14_PACKED, "UINT14_PACKED"},
-        {AX_ENGINE_DT_UINT16_PACKED, "UINT16_PACKED"},
-    };
-
-    static std::map<AX_ENGINE_COLOR_SPACE_T, const char *> color_type = {
-        {AX_ENGINE_CS_FEATUREMAP, "FEATUREMAP"},
-        {AX_ENGINE_CS_RAW8, "RAW8"},
-        {AX_ENGINE_CS_RAW10, "RAW10"},
-        {AX_ENGINE_CS_RAW12, "RAW12"},
-        {AX_ENGINE_CS_RAW14, "RAW14"},
-        {AX_ENGINE_CS_RAW16, "RAW16"},
-        {AX_ENGINE_CS_NV12, "NV12"},
-        {AX_ENGINE_CS_NV21, "NV21"},
-        {AX_ENGINE_CS_RGB, "RGB"},
-        {AX_ENGINE_CS_BGR, "BGR"},
-        {AX_ENGINE_CS_RGBA, "RGBA"},
-        {AX_ENGINE_CS_GRAY, "GRAY"},
-        {AX_ENGINE_CS_YUV444, "YUV444"},
-    };
-    printf("\ninput size: %d\n", io_info->nInputSize);
-    for (uint32_t i = 0; i < io_info->nInputSize; ++i)
-    {
-        // print shape info,like [batchsize x channel x height x width]
-        auto &info = io_info->pInputs[i];
-        printf("    name: \e[1;32m%8s", info.pName);
-
-        std::string dt = "unknown";
-        if (data_type.find(info.eDataType) != data_type.end())
-        {
-            dt = data_type[info.eDataType];
-            printf(" \e[1;34m[%s] ", dt.c_str());
-        }
-        else
-        {
-            printf(" \e[1;31m[%s] ", dt.c_str());
-        }
-
-        std::string ct = "unknown";
-        if (info.pExtraMeta && color_type.find(info.pExtraMeta->eColorSpace) != color_type.end())
-        {
-            ct = color_type[info.pExtraMeta->eColorSpace];
-            printf("\e[1;34m[%s]", ct.c_str());
-        }
-        else
-        {
-            printf("\e[1;31m[%s]", ct.c_str());
-        }
-        printf(" \n        \e[1;31m");
-
-        for (AX_U8 s = 0; s < info.nShapeSize; s++)
-        {
-            printf("%d", info.pShape[s]);
-            if (s != info.nShapeSize - 1)
-            {
-                printf(" x ");
-            }
-        }
-        printf("\e[0m\n\n");
-    }
+    memset(io_data, 0, sizeof(*io_data));
+    io_data->pInputs    = new AX_ENGINE_IO_BUFFER_T[info->nInputSize];
+    io_data->nInputSize = info->nInputSize;
+    memset(io_data->pInputs, 0, sizeof(AX_ENGINE_IO_BUFFER_T) * info->nInputSize);
 
-    printf("\noutput size: %d\n", io_info->nOutputSize);
-    for (uint32_t i = 0; i < io_info->nOutputSize; ++i)
-    {
-        // print shape info,like [batchsize x channel x height x width]
-        auto &info = io_info->pOutputs[i];
-        printf("    name: \e[1;32m%8s \e[1;34m[%s]\e[0m\n        \e[1;31m", info.pName, data_type[info.eDataType]);
-        for (AX_U8 s = 0; s < info.nShapeSize; s++)
-        {
-            printf("%d", info.pShape[s]);
-            if (s != info.nShapeSize - 1)
-            {
-                printf(" x ");
-            }
-        }
-        printf("\e[0m\n\n");
-    }
-}
+    io_data->pOutputs    = new AX_ENGINE_IO_BUFFER_T[info->nOutputSize];
+    io_data->nOutputSize = info->nOutputSize;
+    memset(io_data->pOutputs, 0, sizeof(AX_ENGINE_IO_BUFFER_T) * info->nOutputSize);
 
-void free_io_index(AX_ENGINE_IO_BUFFER_T *io_buf, int index)
-{
-    for (int i = 0; i < index; ++i)
-    {
-        AX_ENGINE_IO_BUFFER_T *pBuf = io_buf + i;
-        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
-    }
-}
+    for (uint i = 0; i < info->nInputSize; ++i) io_data->pInputs[i].nSize = info->pInputs[i].nSize;
+    for (uint i = 0; i < info->nOutputSize; ++i) io_data->pOutputs[i].nSize = info->pOutputs[i].nSize;
 
-void free_io(AX_ENGINE_IO_T *io)
-{
-    for (size_t j = 0; j < io->nInputSize; ++j)
-    {
-        AX_ENGINE_IO_BUFFER_T *pBuf = io->pInputs + j;
-        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
-    }
-    for (size_t j = 0; j < io->nOutputSize; ++j)
-    {
-        AX_ENGINE_IO_BUFFER_T *pBuf = io->pOutputs + j;
-        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
-    }
-    delete[] io->pInputs;
-    delete[] io->pOutputs;
+    return 0;
 }
 
-static inline int prepare_io(AX_ENGINE_IO_INFO_T *info, AX_ENGINE_IO_T *io_data, INPUT_OUTPUT_ALLOC_STRATEGY strategy)
+static int prepare_io_with_alloc(
+    AX_ENGINE_IO_INFO_T *info, AX_ENGINE_IO_T *io_data,
+    std::pair<AX_ENGINE_ALLOC_BUFFER_STRATEGY_T, AX_ENGINE_ALLOC_BUFFER_STRATEGY_T> strategy,
+    std::vector<std::string> skip_alloc_names = {})
 {
-    memset(io_data, 0, sizeof(*io_data));
-    io_data->pInputs = new AX_ENGINE_IO_BUFFER_T[info->nInputSize];
-    io_data->nInputSize = info->nInputSize;
-
-    auto ret = 0;
-    for (uint i = 0; i < info->nInputSize; ++i)
-    {
-        auto meta = info->pInputs[i];
-        auto buffer = &io_data->pInputs[i];
-        if (strategy.first == AX_ENGINE_ABST_CACHED)
-        {
-            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+    int ret = prepare_io_struct_only(info, io_data);
+    if (ret != 0) return ret;
+
+    for (uint i = 0; i < info->nInputSize; ++i) {
+        auto &buffer = io_data->pInputs[i];
+        if (std::find(skip_alloc_names.begin(), skip_alloc_names.end(), info->pInputs[i].pName) !=
+            skip_alloc_names.end()) {
+            continue;
         }
-        else
-        {
-            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        if (strategy.first == AX_ENGINE_ABST_CACHED) {
+            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer.phyAddr), &buffer.pVirAddr, buffer.nSize, AX_CMM_ALIGN_SIZE,
+                                        (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        } else {
+            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer.phyAddr), &buffer.pVirAddr, buffer.nSize, AX_CMM_ALIGN_SIZE,
+                                  (const AX_S8 *)(AX_CMM_SESSION_NAME));
         }
-
-        if (ret != 0)
-        {
-            free_io_index(io_data->pInputs, i);
-            fprintf(stderr, "Allocate input{%d} { phy: %p, vir: %p, size: %lu Bytes }. fail \n", i, (void *)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+        if (ret != 0) {
+            ALOGE("Alloc input[%d] failed", i);
             return ret;
         }
-        memset(buffer->pVirAddr, 0, meta.nSize);
-        // fprintf(stderr, "Allocate input{%d} { phy: %p, vir: %p, size: %lu Bytes }. \n", i, (void*)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+        memset(buffer.pVirAddr, 0, buffer.nSize);
     }
 
-    io_data->pOutputs = new AX_ENGINE_IO_BUFFER_T[info->nOutputSize];
-    io_data->nOutputSize = info->nOutputSize;
-    for (uint i = 0; i < info->nOutputSize; ++i)
-    {
-        auto meta = info->pOutputs[i];
-        auto buffer = &io_data->pOutputs[i];
-        buffer->nSize = meta.nSize;
-        if (strategy.second == AX_ENGINE_ABST_CACHED)
-        {
-            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+    for (uint i = 0; i < info->nOutputSize; ++i) {
+        auto &buffer = io_data->pOutputs[i];
+        if (std::find(skip_alloc_names.begin(), skip_alloc_names.end(), info->pOutputs[i].pName) !=
+            skip_alloc_names.end()) {
+            continue;
         }
-        else
-        {
-            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        if (strategy.second == AX_ENGINE_ABST_CACHED) {
+            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer.phyAddr), &buffer.pVirAddr, buffer.nSize, AX_CMM_ALIGN_SIZE,
+                                        (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        } else {
+            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer.phyAddr), &buffer.pVirAddr, buffer.nSize, AX_CMM_ALIGN_SIZE,
+                                  (const AX_S8 *)(AX_CMM_SESSION_NAME));
         }
-        if (ret != 0)
-        {
-            fprintf(stderr, "Allocate output{%d} { phy: %p, vir: %p, size: %lu Bytes }. fail \n", i, (void *)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
-            free_io_index(io_data->pInputs, io_data->nInputSize);
-            free_io_index(io_data->pOutputs, i);
+        if (ret != 0) {
+            ALOGE("Alloc output[%d] failed", i);
             return ret;
         }
-        memset(buffer->pVirAddr, 0, meta.nSize);
-        // fprintf(stderr, "Allocate output{%d} { phy: %p, vir: %p, size: %lu Bytes }.\n", i, (void*)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+        memset(buffer.pVirAddr, 0, buffer.nSize);
     }
-
     return 0;
 }
 
-struct ax_joint_runner_ax650_handle_t
-{
-    AX_ENGINE_HANDLE handle;
-    AX_ENGINE_CONTEXT_T context;
-    std::vector<AX_ENGINE_IO_INFO_T *> io_info;
-    std::vector<AX_ENGINE_IO_T> io_data;
-
-    // int algo_width, algo_height;
-    // int algo_colorformat;
-};
-
 int ax_runner_ax650::sub_init()
 {
-    // 4. create context
+    if (!m_handle) return -1;
+
     int ret = AX_ENGINE_CreateContext(m_handle->handle);
-    if (0 != ret)
-    {
-        ALOGE("AX_ENGINE_CreateContext");
-        return ret;
-    }
+    if (ret != 0) return ret;
+
     ret = AX_ENGINE_CreateContextV2(m_handle->handle, &m_handle->context);
-    if (0 != ret)
-    {
-        ALOGE("AX_ENGINE_CreateContextV2");
-        return ret;
-    }
-    // fprintf(stdout, "Engine creating context is done.\n");
+    if (ret != 0) return ret;
 
-    // 5. set io
     AX_U32 io_count = 0;
-    ret = AX_ENGINE_GetGroupIOInfoCount(m_handle->handle, &io_count);
-    if (0 != ret)
-    {
-        ALOGE("AX_ENGINE_GetGroupIOInfoCount");
-        return ret;
-    }
-    // ALOGI("io_count=%d", io_count);
+    ret             = AX_ENGINE_GetGroupIOInfoCount(m_handle->handle, &io_count);
+    if (ret != 0) return ret;
 
     m_handle->io_info.resize(io_count);
     m_handle->io_data.resize(io_count);
     mgroup_input_tensors.resize(io_count);
     mgroup_output_tensors.resize(io_count);
 
-    // fprintf(stdout, "Engine get io info is done. \n");
-
-    // 6. alloc io
-    if (!_parepare_io)
-    {
-        for (size_t grpid = 0; grpid < io_count; grpid++)
-        {
-            AX_ENGINE_IO_INFO_T *io_info = nullptr;
-            ret = AX_ENGINE_GetGroupIOInfo(m_handle->handle, grpid, &io_info);
-            if (0 != ret)
-            {
-                ALOGE("AX_ENGINE_GetIOInfo");
-                return ret;
-            }
-            // print_io_info(io_info);
-
-            m_handle->io_info[grpid] = io_info;
+    std::vector<std::string> skip_alloc_names = {"K_cache", "V_cache"};
+    for (size_t grpid = 0; grpid < io_count; grpid++) {
+        AX_ENGINE_IO_INFO_T *io_info = nullptr;
+        ret                          = AX_ENGINE_GetGroupIOInfo(m_handle->handle, grpid, &io_info);
+        if (ret != 0) return ret;
+        m_handle->io_info[grpid] = io_info;
+
+        if (grpid == 0) {
+            ret = prepare_io_with_alloc(io_info, &m_handle->io_data[grpid],
+                                        {AX_ENGINE_ABST_DEFAULT, AX_ENGINE_ABST_CACHED});
+        } else if (grpid == io_count - 1) {
+            ret = prepare_io_with_alloc(io_info, &m_handle->io_data[grpid],
+                                        {AX_ENGINE_ABST_DEFAULT, AX_ENGINE_ABST_CACHED}, skip_alloc_names);
+        } else {
+            ret = prepare_io_struct_only(io_info, &m_handle->io_data[grpid]);
+        }
+        if (ret != 0) return ret;
+    }
 
-            ret = prepare_io(m_handle->io_info[grpid], &m_handle->io_data[grpid], std::make_pair(AX_ENGINE_ABST_DEFAULT, AX_ENGINE_ABST_CACHED));
-            if (0 != ret)
-            {
-                ALOGE("prepare_io grpid=%d", grpid);
-                return ret;
+    if (io_count > 2) {
+        auto &first_io_data = m_handle->io_data[0];
+        auto &first_io_info = m_handle->io_info[0];
+        auto &last_io_data  = m_handle->io_data[io_count - 1];
+        auto &last_io_info  = m_handle->io_info[io_count - 1];
+        for (uint i = 0; i < last_io_data.nInputSize; ++i) {
+            if (std::find(skip_alloc_names.begin(), skip_alloc_names.end(), last_io_info->pInputs[i].pName) !=
+                skip_alloc_names.end()) {
+                for (uint j = 0; j < first_io_data.nInputSize; ++j) {
+                    if (first_io_info->pInputs[j].pName == last_io_info->pInputs[i].pName) {
+                        last_io_data.pInputs[i].phyAddr  = first_io_data.pInputs[j].phyAddr;
+                        last_io_data.pInputs[i].pVirAddr = first_io_data.pInputs[j].pVirAddr;
+                    }
+                }
             }
         }
 
-        for (size_t grpid = 0; grpid < io_count; grpid++)
-        {
+        for (size_t grpid = 1; grpid < io_count - 1; grpid++) {
             auto &io_info = m_handle->io_info[grpid];
             auto &io_data = m_handle->io_data[grpid];
-            for (size_t i = 0; i < io_info->nOutputSize; i++)
-            {
-                ax_runner_tensor_t tensor;
-                tensor.nIdx = i;
-                tensor.sName = std::string(io_info->pOutputs[i].pName);
-                tensor.nSize = io_info->pOutputs[i].nSize;
-                for (size_t j = 0; j < io_info->pOutputs[i].nShapeSize; j++)
-                {
-                    tensor.vShape.push_back(io_info->pOutputs[i].pShape[j]);
-                }
-                tensor.phyAddr = io_data.pOutputs[i].phyAddr;
-                tensor.pVirAddr = io_data.pOutputs[i].pVirAddr;
-                mgroup_output_tensors[grpid].push_back(tensor);
+
+            size_t min_inputs = std::min(io_info->nInputSize, last_io_data.nInputSize);
+            for (size_t i = 0; i < min_inputs; i++) {
+                io_data.pInputs[i].phyAddr  = last_io_data.pInputs[i].phyAddr;
+                io_data.pInputs[i].pVirAddr = last_io_data.pInputs[i].pVirAddr;
             }
 
-            for (size_t i = 0; i < io_info->nInputSize; i++)
-            {
-                ax_runner_tensor_t tensor;
-                tensor.nIdx = i;
-                tensor.sName = std::string(io_info->pInputs[i].pName);
-                tensor.nSize = io_info->pInputs[i].nSize;
-                for (size_t j = 0; j < io_info->pInputs[i].nShapeSize; j++)
-                {
-                    tensor.vShape.push_back(io_info->pInputs[i].pShape[j]);
-                }
-                tensor.phyAddr = io_data.pInputs[i].phyAddr;
-                tensor.pVirAddr = io_data.pInputs[i].pVirAddr;
-                mgroup_input_tensors[grpid].push_back(tensor);
+            size_t min_outputs = std::min(io_info->nOutputSize, last_io_data.nOutputSize);
+            for (size_t i = 0; i < min_outputs; i++) {
+                io_data.pOutputs[i].phyAddr  = last_io_data.pOutputs[i].phyAddr;
+                io_data.pOutputs[i].pVirAddr = last_io_data.pOutputs[i].pVirAddr;
             }
         }
+    }
 
-        moutput_tensors = mgroup_output_tensors[0];
-        minput_tensors = mgroup_input_tensors[0];
+    for (size_t grpid = 0; grpid < io_count; grpid++) {
+        auto &io_info = m_handle->io_info[grpid];
+        auto &io_data = m_handle->io_data[grpid];
+
+        for (size_t i = 0; i < io_info->nOutputSize; i++) {
+            ax_runner_tensor_t tensor;
+            tensor.nIdx     = i;
+            tensor.sName    = io_info->pOutputs[i].pName ? std::string(io_info->pOutputs[i].pName) : "";
+            tensor.nSize    = io_info->pOutputs[i].nSize;
+            tensor.phyAddr  = io_data.pOutputs[i].phyAddr;
+            tensor.pVirAddr = io_data.pOutputs[i].pVirAddr;
+            for (size_t j = 0; j < io_info->pOutputs[i].nShapeSize; j++) {
+                tensor.vShape.push_back(io_info->pOutputs[i].pShape[j]);
+            }
+            mgroup_output_tensors[grpid].push_back(tensor);
+        }
 
-        _parepare_io = true;
-    }
-    else
-    {
+        for (size_t i = 0; i < io_info->nInputSize; i++) {
+            ax_runner_tensor_t tensor;
+            tensor.nIdx     = i;
+            tensor.sName    = io_info->pInputs[i].pName ? std::string(io_info->pInputs[i].pName) : "";
+            tensor.nSize    = io_info->pInputs[i].nSize;
+            tensor.phyAddr  = io_data.pInputs[i].phyAddr;
+            tensor.pVirAddr = io_data.pInputs[i].pVirAddr;
+            for (size_t j = 0; j < io_info->pInputs[i].nShapeSize; j++) {
+                tensor.vShape.push_back(io_info->pInputs[i].pShape[j]);
+            }
+            mgroup_input_tensors[grpid].push_back(tensor);
+        }
     }
 
-    return ret;
+    if (!mgroup_output_tensors.empty()) moutput_tensors = mgroup_output_tensors[0];
+    if (!mgroup_input_tensors.empty()) minput_tensors = mgroup_input_tensors[0];
+
+    build_tensor_maps();
+
+    return 0;
 }
 
 int ax_runner_ax650::init(const char *model_file, bool use_mmap)
 {
-    if (use_mmap)
-    {
+    if (use_mmap) {
         MMap model_buffer(model_file);
-        if (!model_buffer.data())
-        {
-            ALOGE("mmap");
-            return -1;
-        }
+        if (!model_buffer.data()) return -1;
         auto ret = init((char *)model_buffer.data(), model_buffer.size());
         model_buffer.close_file();
         return ret;
-    }
-    else
-    {
-        char *model_buffer;
-        size_t len;
-        if (!read_file(model_file, &model_buffer, &len))
-        {
-            ALOGE("read_file");
-            return -1;
-        }
+    } else {
+        char *model_buffer = nullptr;
+        size_t len         = 0;
+        if (!read_file(model_file, &model_buffer, &len)) return -1;
         auto ret = init(model_buffer, len);
         delete[] model_buffer;
         return ret;
@@ -345,96 +225,84 @@ int ax_runner_ax650::init(const char *model_file, bool use_mmap)
 
 int ax_runner_ax650::init(char *model_buffer, size_t model_size)
 {
-    if (!m_handle)
-    {
-        m_handle = new ax_joint_runner_ax650_handle_t;
-    }
-
-    static bool b_init = false;
-    if (!b_init)
-    {
-        // 1. init engine
-        AX_ENGINE_NPU_ATTR_T npu_attr;
-        memset(&npu_attr, 0, sizeof(npu_attr));
-        npu_attr.eHardMode = AX_ENGINE_VIRTUAL_NPU_DISABLE;
-        AX_SYS_Init();
-        auto ret = AX_ENGINE_Init(&npu_attr);
-        if (0 != ret)
-        {
-            return ret;
-        }
-        b_init = true;
-    }
-
-    // 3. create handle
-
-    int ret = AX_ENGINE_CreateHandle(&m_handle->handle, model_buffer, model_size);
-    if (0 != ret)
-    {
-        ALOGE("AX_ENGINE_CreateHandle");
+    if (m_handle) deinit();
+    m_handle = new ax_runner_ax650_handle_t;
+    int ret  = AX_ENGINE_CreateHandle(&m_handle->handle, model_buffer, model_size);
+    if (0 != ret) {
+        ALOGE("AX_ENGINE_CreateHandle failed: 0x%x", ret);
+        delete m_handle;
+        m_handle = nullptr;
         return ret;
     }
-    // fprintf(stdout, "Engine creating handle is done.\n");
-
     return sub_init();
 }
 
-void ax_runner_ax650::release()
+void ax_runner_ax650::deinit()
 {
-    if (m_handle && m_handle->handle)
-    {
-        for (size_t i = 0; i < m_handle->io_data.size(); i++)
-        {
-            /* code */
-            free_io(&m_handle->io_data[i]);
+    if (!m_handle) return;
+
+    std::unordered_set<unsigned long> freed_phy_addrs;
+
+    for (size_t g = 0; g < m_handle->io_data.size(); ++g) {
+        auto &io = m_handle->io_data[g];
+
+        if (io.pInputs) {
+            for (size_t j = 0; j < io.nInputSize; ++j) {
+                AX_ENGINE_IO_BUFFER_T *pBuf = io.pInputs + j;
+                if (pBuf->phyAddr != 0) {
+                    if (freed_phy_addrs.find(pBuf->phyAddr) == freed_phy_addrs.end()) {
+                        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+                        freed_phy_addrs.insert(pBuf->phyAddr);
+                    }
+                }
+            }
+            delete[] io.pInputs;
+            io.pInputs = nullptr;
         }
 
-        AX_ENGINE_DestroyHandle(m_handle->handle);
-        m_handle->handle = nullptr;
+        if (io.pOutputs) {
+            for (size_t j = 0; j < io.nOutputSize; ++j) {
+                AX_ENGINE_IO_BUFFER_T *pBuf = io.pOutputs + j;
+                if (pBuf->phyAddr != 0) {
+                    if (freed_phy_addrs.find(pBuf->phyAddr) == freed_phy_addrs.end()) {
+                        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+                        freed_phy_addrs.insert(pBuf->phyAddr);
+                    }
+                }
+            }
+            delete[] io.pOutputs;
+            io.pOutputs = nullptr;
+        }
     }
 
-    if (m_handle)
-    {
-        delete m_handle;
-        m_handle = nullptr;
+    if (m_handle->handle) {
+        AX_ENGINE_DestroyHandle(m_handle->handle);
     }
 
+    delete m_handle;
+    m_handle = nullptr;
+
     moutput_tensors.clear();
     minput_tensors.clear();
     map_input_tensors.clear();
     map_output_tensors.clear();
-
     mgroup_output_tensors.clear();
     mgroup_input_tensors.clear();
     map_group_input_tensors.clear();
     map_group_output_tensors.clear();
-
-    // AX_ENGINE_Deinit();
 }
 
-void ax_runner_ax650::deinit()
+int ax_runner_ax650::inference()
 {
-    if (m_handle && m_handle->handle)
-    {
-        // free_io(&m_handle->io_data);
-        // mtensors.clear();
-        // minput_tensors.clear();
-        // map_input_tensors.clear();
-        // map_tensors.clear();
-        AX_ENGINE_DestroyHandle(m_handle->handle);
-        m_handle->handle = nullptr;
-        // delete m_handle;
-        // m_handle = nullptr;
+    if (!m_handle) return -1;
+    for (size_t i = 0; i < get_num_inputs(); i++) {
+        auto &tensor = get_input(i);
+        AX_SYS_MflushCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
     }
 
-    // AX_ENGINE_Deinit();
-}
-
-int ax_runner_ax650::inference()
-{
     int ret = AX_ENGINE_RunSync(m_handle->handle, &m_handle->io_data[0]);
-    for (size_t i = 0; i < get_num_outputs(); i++)
-    {
+
+    for (size_t i = 0; i < get_num_outputs(); i++) {
         auto &tensor = get_output(i);
         AX_SYS_MinvalidateCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
     }
@@ -443,12 +311,19 @@ int ax_runner_ax650::inference()
 
 int ax_runner_ax650::inference(int grpid)
 {
+    if (!m_handle) return -1;
+    if (grpid < 0 || grpid >= (int)m_handle->io_data.size()) return -1;
+
+    for (size_t i = 0; i < mgroup_input_tensors[grpid].size(); i++) {
+        auto &tensor = mgroup_input_tensors[grpid][i];
+        AX_SYS_MflushCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
+    }
+
     int ret = AX_ENGINE_RunGroupIOSync(m_handle->handle, m_handle->context, grpid, &m_handle->io_data[grpid]);
 
-    for (size_t i = 0; i < get_num_outputs(); i++)
-    {
-        auto &tensor = get_output(grpid, i);
+    for (size_t i = 0; i < mgroup_output_tensors[grpid].size(); i++) {
+        auto &tensor = mgroup_output_tensors[grpid][i];
         AX_SYS_MinvalidateCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
     }
     return ret;
-}
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner_ax650.hpp b/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner_ax650.hpp
index 351fb04a..62b65380 100644
--- a/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner_ax650.hpp
+++ b/projects/llm_framework/main_llm/src/runner/ax_model_runner/ax_model_runner_ax650.hpp
@@ -1,20 +1,23 @@
 #pragma once
 #include "ax_model_runner.hpp"
 
-class ax_runner_ax650 : public ax_runner_base
-{
-protected:
-    struct ax_joint_runner_ax650_handle_t *m_handle = nullptr;
-
-    bool _parepare_io = false;
+struct ax_runner_ax650_handle_t;
 
+class ax_runner_ax650 : public ax_runner_base {
+protected:
+    struct ax_runner_ax650_handle_t *m_handle = nullptr;
     int sub_init();
 
 public:
+    ax_runner_ax650() = default;
+    virtual ~ax_runner_ax650()
+    {
+        deinit();
+    }
+
     int init(const char *model_file, bool use_mmap = false) override;
     int init(char *model_buffer, size_t model_size) override;
 
-    void release();
     void deinit() override;
 
     int inference() override;
diff --git a/projects/llm_framework/main_vlm/src/runner/LLM.hpp b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
index 59dd87e8..3d6ad98c 100644
--- a/projects/llm_framework/main_vlm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
@@ -285,11 +285,11 @@ class LLM {
     void Deinit()
     {
         for (int i = 0; i < _attr.axmodel_num; i++) {
-            llama_layers[i].layer.release();
+            llama_layers[i].layer.deinit();
         }
-        llama_post.release();
-        vpm_encoder.release();
-        vpm_resampler.release();
+        llama_post.deinit();
+        vpm_encoder.deinit();
+        vpm_resampler.deinit();
         embed_selector.Deinit();
     }
 
@@ -856,10 +856,10 @@ class LLM_CTX {
     void Deinit()
     {
         for (int i = 0; i < _attr.axmodel_num; i++) {
-            llama_layers[i].layer.release();
+            llama_layers[i].layer.deinit();
         }
-        llama_post.release();
-        image_encoder.release();
+        llama_post.deinit();
+        image_encoder.deinit();
         embed_selector.Deinit();
     }
 
@@ -1958,10 +1958,10 @@ class LLM_Qwen {
     void Deinit()
     {
         for (int i = 0; i < _attr.axmodel_num; i++) {
-            llama_layers[i].layer.release();
+            llama_layers[i].layer.deinit();
         }
-        llama_post.release();
-        image_encoder.release();
+        llama_post.deinit();
+        image_encoder.deinit();
         embed_selector.Deinit();
     }
 
diff --git a/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner.hpp b/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner.hpp
index d8c78d84..c4bcb2ac 100644
--- a/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner.hpp
@@ -4,31 +4,7 @@
 #include <map>
 #include <stdexcept>
 
-typedef enum _color_space_e
-{
-    axdl_color_space_unknown,
-    axdl_color_space_nv12,
-    axdl_color_space_nv21,
-    axdl_color_space_bgr,
-    axdl_color_space_rgb,
-} ax_color_space_e;
-
-typedef struct _image_t
-{
-    unsigned long long int pPhy;
-    void *pVir;
-    unsigned int nSize;
-    unsigned int nWidth;
-    unsigned int nHeight;
-    ax_color_space_e eDtype;
-    union
-    {
-        int tStride_H, tStride_W, tStride_C;
-    };
-} ax_image_t;
-
-typedef struct
-{
+typedef struct {
     std::string sName;
     unsigned int nIdx;
     std::vector<unsigned int> vShape;
@@ -37,8 +13,7 @@ typedef struct
     void *pVirAddr;
 } ax_runner_tensor_t;
 
-class ax_runner_base
-{
+class ax_runner_base {
 protected:
     std::vector<ax_runner_tensor_t> moutput_tensors;
     std::vector<ax_runner_tensor_t> minput_tensors;
@@ -52,106 +27,124 @@ class ax_runner_base
     std::map<std::string, std::vector<ax_runner_tensor_t>> map_group_output_tensors;
     std::map<std::string, std::vector<ax_runner_tensor_t>> map_group_input_tensors;
 
+    void build_tensor_maps()
+    {
+        map_input_tensors.clear();
+        for (const auto &t : minput_tensors) map_input_tensors[t.sName] = t;
+
+        map_output_tensors.clear();
+        for (const auto &t : moutput_tensors) map_output_tensors[t.sName] = t;
+
+        map_group_input_tensors.clear();
+        for (const auto &grp : mgroup_input_tensors) {
+            for (const auto &t : grp) map_group_input_tensors[t.sName].push_back(t);
+        }
+
+        map_group_output_tensors.clear();
+        for (const auto &grp : mgroup_output_tensors) {
+            for (const auto &t : grp) map_group_output_tensors[t.sName].push_back(t);
+        }
+    }
+
 public:
+    virtual ~ax_runner_base()
+    {
+    }
+
     virtual int init(const char *model_file, bool use_mmap = false) = 0;
-    virtual int init(char *model_buffer, size_t model_size) = 0;
+    virtual int init(char *model_buffer, size_t model_size)         = 0;
+    virtual void deinit()                                           = 0;
 
-    virtual void deinit() = 0;
+    int get_num_inputs()
+    {
+        return minput_tensors.size();
+    };
+    int get_num_outputs()
+    {
+        return moutput_tensors.size();
+    };
+    int get_num_input_groups()
+    {
+        return mgroup_input_tensors.size();
+    };
+    int get_num_output_groups()
+    {
+        return mgroup_output_tensors.size();
+    };
 
-    int get_num_inputs() { return minput_tensors.size(); };
-    int get_num_outputs() { return moutput_tensors.size(); };
+    const ax_runner_tensor_t &get_input(int idx)
+    {
+        return minput_tensors[idx];
+    }
+    const ax_runner_tensor_t *get_inputs_ptr()
+    {
+        return minput_tensors.data();
+    }
 
-    int get_num_input_groups() { return mgroup_input_tensors.size(); };
-    int get_num_output_groups() { return mgroup_output_tensors.size(); };
+    const ax_runner_tensor_t &get_input(const std::string &name)
+    {
+        auto it = map_input_tensors.find(name);
+        if (it == map_input_tensors.end()) throw std::runtime_error("input tensor not found: " + name);
+        return it->second;
+    }
 
-    const ax_runner_tensor_t &get_input(int idx) { return minput_tensors[idx]; }
-    const ax_runner_tensor_t *get_inputs_ptr() { return minput_tensors.data(); }
-    const ax_runner_tensor_t &get_input(std::string name)
+    const ax_runner_tensor_t &get_input(int grpid, int idx)
     {
-        if (map_input_tensors.size() == 0)
-        {
-            for (size_t i = 0; i < minput_tensors.size(); i++)
-            {
-                map_input_tensors[minput_tensors[i].sName] = minput_tensors[i];
-            }
-        }
-        if (map_input_tensors.find(name) == map_input_tensors.end())
-        {
-            throw std::runtime_error("input tensor not found: " + name);
-        }
+        return mgroup_input_tensors[grpid][idx];
+    }
+    const ax_runner_tensor_t *get_inputs_ptr(int grpid)
+    {
+        return mgroup_input_tensors[grpid].data();
+    }
 
-        return map_input_tensors[name];
+    const ax_runner_tensor_t &get_input(int grpid, const std::string &name)
+    {
+        auto it = map_group_input_tensors.find(name);
+        if (it == map_group_input_tensors.end()) throw std::runtime_error("input tensor not found: " + name);
+        if (grpid < 0 || grpid >= (int)it->second.size())
+            throw std::runtime_error("group id out of range for: " + name);
+        return it->second[grpid];
     }
 
-    const ax_runner_tensor_t &get_input(int grpid, int idx) { return mgroup_input_tensors[grpid][idx]; }
-    const ax_runner_tensor_t *get_inputs_ptr(int grpid) { return mgroup_input_tensors[grpid].data(); }
-    const ax_runner_tensor_t &get_input(int grpid, std::string name)
-    {
-        if (map_group_input_tensors.size() == 0)
-        {
-            for (size_t i = 0; i < mgroup_input_tensors.size(); i++)
-            {
-                for (size_t j = 0; j < mgroup_input_tensors[i].size(); j++)
-                {
-                    map_group_input_tensors[mgroup_input_tensors[i][j].sName].push_back(mgroup_input_tensors[i][j]);
-                }
-            }
-        }
-        if (map_group_input_tensors.find(name) == map_group_input_tensors.end())
-        {
-            throw std::runtime_error("input tensor not found: " + name);
-        }
-        return map_group_input_tensors[name][grpid];
-        // return map_input_tensors[name];
+    const ax_runner_tensor_t &get_output(int idx)
+    {
+        return moutput_tensors[idx];
+    }
+    const ax_runner_tensor_t *get_outputs_ptr()
+    {
+        return moutput_tensors.data();
     }
 
-    const ax_runner_tensor_t &get_output(int idx) { return moutput_tensors[idx]; }
-    const ax_runner_tensor_t *get_outputs_ptr() { return moutput_tensors.data(); }
-    const ax_runner_tensor_t &get_output(std::string name)
+    const ax_runner_tensor_t &get_output(const std::string &name)
     {
-        if (map_output_tensors.size() == 0)
-        {
-            for (size_t i = 0; i < moutput_tensors.size(); i++)
-            {
-                map_output_tensors[moutput_tensors[i].sName] = moutput_tensors[i];
-            }
-        }
-        if (map_output_tensors.find(name) == map_output_tensors.end())
-        {
-            throw std::runtime_error("output tensor not found: " + name);
-        }
+        auto it = map_output_tensors.find(name);
+        if (it == map_output_tensors.end()) throw std::runtime_error("output tensor not found: " + name);
+        return it->second;
+    }
 
-        return map_output_tensors[name];
+    const ax_runner_tensor_t &get_output(int grpid, int idx)
+    {
+        return mgroup_output_tensors[grpid][idx];
+    }
+    const ax_runner_tensor_t *get_outputs_ptr(int grpid)
+    {
+        return mgroup_output_tensors[grpid].data();
     }
 
-    const ax_runner_tensor_t &get_output(int grpid, int idx) { return mgroup_output_tensors[grpid][idx]; }
-    const ax_runner_tensor_t *get_outputs_ptr(int grpid) { return mgroup_output_tensors[grpid].data(); }
-    const ax_runner_tensor_t &get_output(int grpid, std::string name)
-    {
-        if (map_group_output_tensors.size() == 0)
-        {
-            for (size_t i = 0; i < mgroup_output_tensors.size(); i++)
-            {
-                for (size_t j = 0; j < mgroup_output_tensors[i].size(); j++)
-                {
-                    map_group_output_tensors[mgroup_output_tensors[i][j].sName].push_back(mgroup_output_tensors[i][j]);
-                }
-            }
-        }
-        if (map_group_output_tensors.find(name) == map_group_output_tensors.end())
-        {
-            throw std::runtime_error("input tensor not found: " + name);
-        }
-        return map_group_output_tensors[name][grpid];
+    const ax_runner_tensor_t &get_output(int grpid, const std::string &name)
+    {
+        auto it = map_group_output_tensors.find(name);
+        if (it == map_group_output_tensors.end()) throw std::runtime_error("output tensor not found: " + name);
+        if (grpid < 0 || grpid >= (int)it->second.size())
+            throw std::runtime_error("group id out of range for: " + name);
+        return it->second[grpid];
     }
 
-    virtual int inference() = 0;
+    virtual int inference()          = 0;
     virtual int inference(int grpid) = 0;
 
     int operator()()
     {
         return inference();
     }
-};
-
-// int ax_cmmcpy(unsigned long long int dst, unsigned long long int src, int size);
\ No newline at end of file
+};
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner_ax650.cpp b/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner_ax650.cpp
index 04542295..363c3d5c 100644
--- a/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner_ax650.cpp
+++ b/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner_ax650.cpp
@@ -1,8 +1,9 @@
 #include "ax_model_runner_ax650.hpp"
-#include "string.h"
-#include "fstream"
-#include "memory"
-// #include "utilities/file.hpp"
+#include <cstring>
+#include <fstream>
+#include <algorithm>
+#include <memory>
+#include <unordered_set>
 #include <ax_sys_api.h>
 #include <ax_ivps_api.h>
 #include <ax_engine_api.h>
@@ -11,332 +12,211 @@
 #include "sample_log.h"
 
 #define AX_CMM_ALIGN_SIZE 128
-
 const char *AX_CMM_SESSION_NAME = "npu";
 
-typedef enum
-{
+typedef enum {
     AX_ENGINE_ABST_DEFAULT = 0,
-    AX_ENGINE_ABST_CACHED = 1,
+    AX_ENGINE_ABST_CACHED  = 1,
 } AX_ENGINE_ALLOC_BUFFER_STRATEGY_T;
 
-typedef std::pair<AX_ENGINE_ALLOC_BUFFER_STRATEGY_T, AX_ENGINE_ALLOC_BUFFER_STRATEGY_T> INPUT_OUTPUT_ALLOC_STRATEGY;
+struct ax_runner_ax650_handle_t {
+    AX_ENGINE_HANDLE handle     = nullptr;
+    AX_ENGINE_CONTEXT_T context = 0;
+    std::vector<AX_ENGINE_IO_INFO_T *> io_info;
+    std::vector<AX_ENGINE_IO_T> io_data;
+};
 
-static void print_io_info(AX_ENGINE_IO_INFO_T *io_info)
+static int prepare_io_struct_only(AX_ENGINE_IO_INFO_T *info, AX_ENGINE_IO_T *io_data)
 {
-    static std::map<AX_ENGINE_DATA_TYPE_T, const char *> data_type = {
-        {AX_ENGINE_DT_UNKNOWN, "UNKNOWN"},
-        {AX_ENGINE_DT_UINT8, "UINT8"},
-        {AX_ENGINE_DT_UINT16, "UINT16"},
-        {AX_ENGINE_DT_FLOAT32, "FLOAT32"},
-        {AX_ENGINE_DT_SINT16, "SINT16"},
-        {AX_ENGINE_DT_SINT8, "SINT8"},
-        {AX_ENGINE_DT_SINT32, "SINT32"},
-        {AX_ENGINE_DT_UINT32, "UINT32"},
-        {AX_ENGINE_DT_FLOAT64, "FLOAT64"},
-        {AX_ENGINE_DT_UINT10_PACKED, "UINT10_PACKED"},
-        {AX_ENGINE_DT_UINT12_PACKED, "UINT12_PACKED"},
-        {AX_ENGINE_DT_UINT14_PACKED, "UINT14_PACKED"},
-        {AX_ENGINE_DT_UINT16_PACKED, "UINT16_PACKED"},
-    };
-
-    static std::map<AX_ENGINE_COLOR_SPACE_T, const char *> color_type = {
-        {AX_ENGINE_CS_FEATUREMAP, "FEATUREMAP"},
-        {AX_ENGINE_CS_RAW8, "RAW8"},
-        {AX_ENGINE_CS_RAW10, "RAW10"},
-        {AX_ENGINE_CS_RAW12, "RAW12"},
-        {AX_ENGINE_CS_RAW14, "RAW14"},
-        {AX_ENGINE_CS_RAW16, "RAW16"},
-        {AX_ENGINE_CS_NV12, "NV12"},
-        {AX_ENGINE_CS_NV21, "NV21"},
-        {AX_ENGINE_CS_RGB, "RGB"},
-        {AX_ENGINE_CS_BGR, "BGR"},
-        {AX_ENGINE_CS_RGBA, "RGBA"},
-        {AX_ENGINE_CS_GRAY, "GRAY"},
-        {AX_ENGINE_CS_YUV444, "YUV444"},
-    };
-    printf("\ninput size: %d\n", io_info->nInputSize);
-    for (uint32_t i = 0; i < io_info->nInputSize; ++i)
-    {
-        // print shape info,like [batchsize x channel x height x width]
-        auto &info = io_info->pInputs[i];
-        printf("    name: \e[1;32m%8s", info.pName);
-
-        std::string dt = "unknown";
-        if (data_type.find(info.eDataType) != data_type.end())
-        {
-            dt = data_type[info.eDataType];
-            printf(" \e[1;34m[%s] ", dt.c_str());
-        }
-        else
-        {
-            printf(" \e[1;31m[%s] ", dt.c_str());
-        }
-
-        std::string ct = "unknown";
-        if (info.pExtraMeta && color_type.find(info.pExtraMeta->eColorSpace) != color_type.end())
-        {
-            ct = color_type[info.pExtraMeta->eColorSpace];
-            printf("\e[1;34m[%s]", ct.c_str());
-        }
-        else
-        {
-            printf("\e[1;31m[%s]", ct.c_str());
-        }
-        printf(" \n        \e[1;31m");
-
-        for (AX_U8 s = 0; s < info.nShapeSize; s++)
-        {
-            printf("%d", info.pShape[s]);
-            if (s != info.nShapeSize - 1)
-            {
-                printf(" x ");
-            }
-        }
-        printf("\e[0m\n\n");
-    }
+    memset(io_data, 0, sizeof(*io_data));
+    io_data->pInputs    = new AX_ENGINE_IO_BUFFER_T[info->nInputSize];
+    io_data->nInputSize = info->nInputSize;
+    memset(io_data->pInputs, 0, sizeof(AX_ENGINE_IO_BUFFER_T) * info->nInputSize);
 
-    printf("\noutput size: %d\n", io_info->nOutputSize);
-    for (uint32_t i = 0; i < io_info->nOutputSize; ++i)
-    {
-        // print shape info,like [batchsize x channel x height x width]
-        auto &info = io_info->pOutputs[i];
-        printf("    name: \e[1;32m%8s \e[1;34m[%s]\e[0m\n        \e[1;31m", info.pName, data_type[info.eDataType]);
-        for (AX_U8 s = 0; s < info.nShapeSize; s++)
-        {
-            printf("%d", info.pShape[s]);
-            if (s != info.nShapeSize - 1)
-            {
-                printf(" x ");
-            }
-        }
-        printf("\e[0m\n\n");
-    }
-}
+    io_data->pOutputs    = new AX_ENGINE_IO_BUFFER_T[info->nOutputSize];
+    io_data->nOutputSize = info->nOutputSize;
+    memset(io_data->pOutputs, 0, sizeof(AX_ENGINE_IO_BUFFER_T) * info->nOutputSize);
 
-void free_io_index(AX_ENGINE_IO_BUFFER_T *io_buf, int index)
-{
-    for (int i = 0; i < index; ++i)
-    {
-        AX_ENGINE_IO_BUFFER_T *pBuf = io_buf + i;
-        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
-    }
-}
+    for (uint i = 0; i < info->nInputSize; ++i) io_data->pInputs[i].nSize = info->pInputs[i].nSize;
+    for (uint i = 0; i < info->nOutputSize; ++i) io_data->pOutputs[i].nSize = info->pOutputs[i].nSize;
 
-void free_io(AX_ENGINE_IO_T *io)
-{
-    for (size_t j = 0; j < io->nInputSize; ++j)
-    {
-        AX_ENGINE_IO_BUFFER_T *pBuf = io->pInputs + j;
-        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
-    }
-    for (size_t j = 0; j < io->nOutputSize; ++j)
-    {
-        AX_ENGINE_IO_BUFFER_T *pBuf = io->pOutputs + j;
-        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
-    }
-    delete[] io->pInputs;
-    delete[] io->pOutputs;
+    return 0;
 }
 
-static inline int prepare_io(AX_ENGINE_IO_INFO_T *info, AX_ENGINE_IO_T *io_data, INPUT_OUTPUT_ALLOC_STRATEGY strategy)
+static int prepare_io_with_alloc(
+    AX_ENGINE_IO_INFO_T *info, AX_ENGINE_IO_T *io_data,
+    std::pair<AX_ENGINE_ALLOC_BUFFER_STRATEGY_T, AX_ENGINE_ALLOC_BUFFER_STRATEGY_T> strategy,
+    std::vector<std::string> skip_alloc_names = {})
 {
-    memset(io_data, 0, sizeof(*io_data));
-    io_data->pInputs = new AX_ENGINE_IO_BUFFER_T[info->nInputSize];
-    io_data->nInputSize = info->nInputSize;
-
-    auto ret = 0;
-    for (uint i = 0; i < info->nInputSize; ++i)
-    {
-        auto meta = info->pInputs[i];
-        auto buffer = &io_data->pInputs[i];
-        if (strategy.first == AX_ENGINE_ABST_CACHED)
-        {
-            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+    int ret = prepare_io_struct_only(info, io_data);
+    if (ret != 0) return ret;
+
+    for (uint i = 0; i < info->nInputSize; ++i) {
+        auto &buffer = io_data->pInputs[i];
+        if (std::find(skip_alloc_names.begin(), skip_alloc_names.end(), info->pInputs[i].pName) !=
+            skip_alloc_names.end()) {
+            continue;
         }
-        else
-        {
-            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        if (strategy.first == AX_ENGINE_ABST_CACHED) {
+            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer.phyAddr), &buffer.pVirAddr, buffer.nSize, AX_CMM_ALIGN_SIZE,
+                                        (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        } else {
+            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer.phyAddr), &buffer.pVirAddr, buffer.nSize, AX_CMM_ALIGN_SIZE,
+                                  (const AX_S8 *)(AX_CMM_SESSION_NAME));
         }
-
-        if (ret != 0)
-        {
-            free_io_index(io_data->pInputs, i);
-            fprintf(stderr, "Allocate input{%d} { phy: %p, vir: %p, size: %lu Bytes }. fail \n", i, (void *)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+        if (ret != 0) {
+            ALOGE("Alloc input[%d] failed", i);
             return ret;
         }
-        memset(buffer->pVirAddr, 0, meta.nSize);
-        // fprintf(stderr, "Allocate input{%d} { phy: %p, vir: %p, size: %lu Bytes }. \n", i, (void*)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+        memset(buffer.pVirAddr, 0, buffer.nSize);
     }
 
-    io_data->pOutputs = new AX_ENGINE_IO_BUFFER_T[info->nOutputSize];
-    io_data->nOutputSize = info->nOutputSize;
-    for (uint i = 0; i < info->nOutputSize; ++i)
-    {
-        auto meta = info->pOutputs[i];
-        auto buffer = &io_data->pOutputs[i];
-        buffer->nSize = meta.nSize;
-        if (strategy.second == AX_ENGINE_ABST_CACHED)
-        {
-            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+    for (uint i = 0; i < info->nOutputSize; ++i) {
+        auto &buffer = io_data->pOutputs[i];
+        if (std::find(skip_alloc_names.begin(), skip_alloc_names.end(), info->pOutputs[i].pName) !=
+            skip_alloc_names.end()) {
+            continue;
         }
-        else
-        {
-            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE, (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        if (strategy.second == AX_ENGINE_ABST_CACHED) {
+            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer.phyAddr), &buffer.pVirAddr, buffer.nSize, AX_CMM_ALIGN_SIZE,
+                                        (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        } else {
+            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer.phyAddr), &buffer.pVirAddr, buffer.nSize, AX_CMM_ALIGN_SIZE,
+                                  (const AX_S8 *)(AX_CMM_SESSION_NAME));
         }
-        if (ret != 0)
-        {
-            fprintf(stderr, "Allocate output{%d} { phy: %p, vir: %p, size: %lu Bytes }. fail \n", i, (void *)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
-            free_io_index(io_data->pInputs, io_data->nInputSize);
-            free_io_index(io_data->pOutputs, i);
+        if (ret != 0) {
+            ALOGE("Alloc output[%d] failed", i);
             return ret;
         }
-        memset(buffer->pVirAddr, 0, meta.nSize);
-        // fprintf(stderr, "Allocate output{%d} { phy: %p, vir: %p, size: %lu Bytes }.\n", i, (void*)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+        memset(buffer.pVirAddr, 0, buffer.nSize);
     }
-
     return 0;
 }
 
-struct ax_joint_runner_ax650_handle_t
-{
-    AX_ENGINE_HANDLE handle;
-    AX_ENGINE_CONTEXT_T context;
-    std::vector<AX_ENGINE_IO_INFO_T *> io_info;
-    std::vector<AX_ENGINE_IO_T> io_data;
-
-    // int algo_width, algo_height;
-    // int algo_colorformat;
-};
-
 int ax_runner_ax650::sub_init()
 {
-    // 4. create context
+    if (!m_handle) return -1;
+
     int ret = AX_ENGINE_CreateContext(m_handle->handle);
-    if (0 != ret)
-    {
-        ALOGE("AX_ENGINE_CreateContext");
-        return ret;
-    }
+    if (ret != 0) return ret;
+
     ret = AX_ENGINE_CreateContextV2(m_handle->handle, &m_handle->context);
-    if (0 != ret)
-    {
-        ALOGE("AX_ENGINE_CreateContextV2");
-        return ret;
-    }
-    // fprintf(stdout, "Engine creating context is done.\n");
+    if (ret != 0) return ret;
 
-    // 5. set io
     AX_U32 io_count = 0;
-    ret = AX_ENGINE_GetGroupIOInfoCount(m_handle->handle, &io_count);
-    if (0 != ret)
-    {
-        ALOGE("AX_ENGINE_GetGroupIOInfoCount");
-        return ret;
-    }
-    // ALOGI("io_count=%d", io_count);
+    ret             = AX_ENGINE_GetGroupIOInfoCount(m_handle->handle, &io_count);
+    if (ret != 0) return ret;
 
     m_handle->io_info.resize(io_count);
     m_handle->io_data.resize(io_count);
     mgroup_input_tensors.resize(io_count);
     mgroup_output_tensors.resize(io_count);
 
-    // fprintf(stdout, "Engine get io info is done. \n");
-
-    // 6. alloc io
-    if (!_parepare_io)
-    {
-        for (size_t grpid = 0; grpid < io_count; grpid++)
-        {
-            AX_ENGINE_IO_INFO_T *io_info = nullptr;
-            ret = AX_ENGINE_GetGroupIOInfo(m_handle->handle, grpid, &io_info);
-            if (0 != ret)
-            {
-                ALOGE("AX_ENGINE_GetIOInfo");
-                return ret;
-            }
-            // print_io_info(io_info);
-
-            m_handle->io_info[grpid] = io_info;
+    std::vector<std::string> skip_alloc_names = {"K_cache", "V_cache"};
+    for (size_t grpid = 0; grpid < io_count; grpid++) {
+        AX_ENGINE_IO_INFO_T *io_info = nullptr;
+        ret                          = AX_ENGINE_GetGroupIOInfo(m_handle->handle, grpid, &io_info);
+        if (ret != 0) return ret;
+        m_handle->io_info[grpid] = io_info;
+
+        if (grpid == 0) {
+            ret = prepare_io_with_alloc(io_info, &m_handle->io_data[grpid],
+                                        {AX_ENGINE_ABST_DEFAULT, AX_ENGINE_ABST_CACHED});
+        } else if (grpid == io_count - 1) {
+            ret = prepare_io_with_alloc(io_info, &m_handle->io_data[grpid],
+                                        {AX_ENGINE_ABST_DEFAULT, AX_ENGINE_ABST_CACHED}, skip_alloc_names);
+        } else {
+            ret = prepare_io_struct_only(io_info, &m_handle->io_data[grpid]);
+        }
+        if (ret != 0) return ret;
+    }
 
-            ret = prepare_io(m_handle->io_info[grpid], &m_handle->io_data[grpid], std::make_pair(AX_ENGINE_ABST_DEFAULT, AX_ENGINE_ABST_CACHED));
-            if (0 != ret)
-            {
-                ALOGE("prepare_io grpid=%d", grpid);
-                return ret;
+    if (io_count > 2) {
+        auto &first_io_data = m_handle->io_data[0];
+        auto &first_io_info = m_handle->io_info[0];
+        auto &last_io_data  = m_handle->io_data[io_count - 1];
+        auto &last_io_info  = m_handle->io_info[io_count - 1];
+        for (uint i = 0; i < last_io_data.nInputSize; ++i) {
+            if (std::find(skip_alloc_names.begin(), skip_alloc_names.end(), last_io_info->pInputs[i].pName) !=
+                skip_alloc_names.end()) {
+                for (uint j = 0; j < first_io_data.nInputSize; ++j) {
+                    if (first_io_info->pInputs[j].pName == last_io_info->pInputs[i].pName) {
+                        last_io_data.pInputs[i].phyAddr  = first_io_data.pInputs[j].phyAddr;
+                        last_io_data.pInputs[i].pVirAddr = first_io_data.pInputs[j].pVirAddr;
+                    }
+                }
             }
         }
 
-        for (size_t grpid = 0; grpid < io_count; grpid++)
-        {
+        for (size_t grpid = 1; grpid < io_count - 1; grpid++) {
             auto &io_info = m_handle->io_info[grpid];
             auto &io_data = m_handle->io_data[grpid];
-            for (size_t i = 0; i < io_info->nOutputSize; i++)
-            {
-                ax_runner_tensor_t tensor;
-                tensor.nIdx = i;
-                tensor.sName = std::string(io_info->pOutputs[i].pName);
-                tensor.nSize = io_info->pOutputs[i].nSize;
-                for (size_t j = 0; j < io_info->pOutputs[i].nShapeSize; j++)
-                {
-                    tensor.vShape.push_back(io_info->pOutputs[i].pShape[j]);
-                }
-                tensor.phyAddr = io_data.pOutputs[i].phyAddr;
-                tensor.pVirAddr = io_data.pOutputs[i].pVirAddr;
-                mgroup_output_tensors[grpid].push_back(tensor);
+
+            size_t min_inputs = std::min(io_info->nInputSize, last_io_data.nInputSize);
+            for (size_t i = 0; i < min_inputs; i++) {
+                io_data.pInputs[i].phyAddr  = last_io_data.pInputs[i].phyAddr;
+                io_data.pInputs[i].pVirAddr = last_io_data.pInputs[i].pVirAddr;
             }
 
-            for (size_t i = 0; i < io_info->nInputSize; i++)
-            {
-                ax_runner_tensor_t tensor;
-                tensor.nIdx = i;
-                tensor.sName = std::string(io_info->pInputs[i].pName);
-                tensor.nSize = io_info->pInputs[i].nSize;
-                for (size_t j = 0; j < io_info->pInputs[i].nShapeSize; j++)
-                {
-                    tensor.vShape.push_back(io_info->pInputs[i].pShape[j]);
-                }
-                tensor.phyAddr = io_data.pInputs[i].phyAddr;
-                tensor.pVirAddr = io_data.pInputs[i].pVirAddr;
-                mgroup_input_tensors[grpid].push_back(tensor);
+            size_t min_outputs = std::min(io_info->nOutputSize, last_io_data.nOutputSize);
+            for (size_t i = 0; i < min_outputs; i++) {
+                io_data.pOutputs[i].phyAddr  = last_io_data.pOutputs[i].phyAddr;
+                io_data.pOutputs[i].pVirAddr = last_io_data.pOutputs[i].pVirAddr;
             }
         }
+    }
 
-        moutput_tensors = mgroup_output_tensors[0];
-        minput_tensors = mgroup_input_tensors[0];
+    for (size_t grpid = 0; grpid < io_count; grpid++) {
+        auto &io_info = m_handle->io_info[grpid];
+        auto &io_data = m_handle->io_data[grpid];
+
+        for (size_t i = 0; i < io_info->nOutputSize; i++) {
+            ax_runner_tensor_t tensor;
+            tensor.nIdx     = i;
+            tensor.sName    = io_info->pOutputs[i].pName ? std::string(io_info->pOutputs[i].pName) : "";
+            tensor.nSize    = io_info->pOutputs[i].nSize;
+            tensor.phyAddr  = io_data.pOutputs[i].phyAddr;
+            tensor.pVirAddr = io_data.pOutputs[i].pVirAddr;
+            for (size_t j = 0; j < io_info->pOutputs[i].nShapeSize; j++) {
+                tensor.vShape.push_back(io_info->pOutputs[i].pShape[j]);
+            }
+            mgroup_output_tensors[grpid].push_back(tensor);
+        }
 
-        _parepare_io = true;
-    }
-    else
-    {
+        for (size_t i = 0; i < io_info->nInputSize; i++) {
+            ax_runner_tensor_t tensor;
+            tensor.nIdx     = i;
+            tensor.sName    = io_info->pInputs[i].pName ? std::string(io_info->pInputs[i].pName) : "";
+            tensor.nSize    = io_info->pInputs[i].nSize;
+            tensor.phyAddr  = io_data.pInputs[i].phyAddr;
+            tensor.pVirAddr = io_data.pInputs[i].pVirAddr;
+            for (size_t j = 0; j < io_info->pInputs[i].nShapeSize; j++) {
+                tensor.vShape.push_back(io_info->pInputs[i].pShape[j]);
+            }
+            mgroup_input_tensors[grpid].push_back(tensor);
+        }
     }
 
-    return ret;
+    if (!mgroup_output_tensors.empty()) moutput_tensors = mgroup_output_tensors[0];
+    if (!mgroup_input_tensors.empty()) minput_tensors = mgroup_input_tensors[0];
+
+    build_tensor_maps();
+
+    return 0;
 }
 
 int ax_runner_ax650::init(const char *model_file, bool use_mmap)
 {
-    if (use_mmap)
-    {
+    if (use_mmap) {
         MMap model_buffer(model_file);
-        if (!model_buffer.data())
-        {
-            ALOGE("mmap");
-            return -1;
-        }
+        if (!model_buffer.data()) return -1;
         auto ret = init((char *)model_buffer.data(), model_buffer.size());
         model_buffer.close_file();
         return ret;
-    }
-    else
-    {
-        char *model_buffer;
-        size_t len;
-        if (!read_file(model_file, &model_buffer, &len))
-        {
-            ALOGE("read_file");
-            return -1;
-        }
+    } else {
+        char *model_buffer = nullptr;
+        size_t len         = 0;
+        if (!read_file(model_file, &model_buffer, &len)) return -1;
         auto ret = init(model_buffer, len);
         delete[] model_buffer;
         return ret;
@@ -345,97 +225,105 @@ int ax_runner_ax650::init(const char *model_file, bool use_mmap)
 
 int ax_runner_ax650::init(char *model_buffer, size_t model_size)
 {
-    if (!m_handle)
-    {
-        m_handle = new ax_joint_runner_ax650_handle_t;
-    }
-
-    static bool b_init = false;
-    if (!b_init)
-    {
-        // 1. init engine
-        AX_ENGINE_NPU_ATTR_T npu_attr;
-        memset(&npu_attr, 0, sizeof(npu_attr));
-        npu_attr.eHardMode = AX_ENGINE_VIRTUAL_NPU_DISABLE;
-        AX_SYS_Init();
-        auto ret = AX_ENGINE_Init(&npu_attr);
-        if (0 != ret)
-        {
-            return ret;
-        }
-        b_init = true;
-    }
-
-    // 3. create handle
-
-    int ret = AX_ENGINE_CreateHandle(&m_handle->handle, model_buffer, model_size);
-    if (0 != ret)
-    {
-        ALOGE("AX_ENGINE_CreateHandle");
+    if (m_handle) deinit();
+    m_handle = new ax_runner_ax650_handle_t;
+    int ret  = AX_ENGINE_CreateHandle(&m_handle->handle, model_buffer, model_size);
+    if (0 != ret) {
+        ALOGE("AX_ENGINE_CreateHandle failed: 0x%x", ret);
+        delete m_handle;
+        m_handle = nullptr;
         return ret;
     }
-    // fprintf(stdout, "Engine creating handle is done.\n");
-
     return sub_init();
 }
 
-void ax_runner_ax650::release()
+void ax_runner_ax650::deinit()
 {
-    if (m_handle && m_handle->handle)
-    {
-        for (size_t i = 0; i < m_handle->io_data.size(); i++)
-        {
-            /* code */
-            free_io(&m_handle->io_data[i]);
+    if (!m_handle) return;
+
+    std::unordered_set<unsigned long> freed_phy_addrs;
+
+    for (size_t g = 0; g < m_handle->io_data.size(); ++g) {
+        auto &io = m_handle->io_data[g];
+
+        if (io.pInputs) {
+            for (size_t j = 0; j < io.nInputSize; ++j) {
+                AX_ENGINE_IO_BUFFER_T *pBuf = io.pInputs + j;
+                if (pBuf->phyAddr != 0) {
+                    if (freed_phy_addrs.find(pBuf->phyAddr) == freed_phy_addrs.end()) {
+                        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+                        freed_phy_addrs.insert(pBuf->phyAddr);
+                    }
+                }
+            }
+            delete[] io.pInputs;
+            io.pInputs = nullptr;
         }
 
-        AX_ENGINE_DestroyHandle(m_handle->handle);
-        m_handle->handle = nullptr;
+        if (io.pOutputs) {
+            for (size_t j = 0; j < io.nOutputSize; ++j) {
+                AX_ENGINE_IO_BUFFER_T *pBuf = io.pOutputs + j;
+                if (pBuf->phyAddr != 0) {
+                    if (freed_phy_addrs.find(pBuf->phyAddr) == freed_phy_addrs.end()) {
+                        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+                        freed_phy_addrs.insert(pBuf->phyAddr);
+                    }
+                }
+            }
+            delete[] io.pOutputs;
+            io.pOutputs = nullptr;
+        }
     }
 
-    if (m_handle)
-    {
-        delete m_handle;
-        m_handle = nullptr;
+    if (m_handle->handle) {
+        AX_ENGINE_DestroyHandle(m_handle->handle);
     }
 
+    delete m_handle;
+    m_handle = nullptr;
+
     moutput_tensors.clear();
     minput_tensors.clear();
     map_input_tensors.clear();
     map_output_tensors.clear();
-
     mgroup_output_tensors.clear();
     mgroup_input_tensors.clear();
     map_group_input_tensors.clear();
     map_group_output_tensors.clear();
-
-    // AX_ENGINE_Deinit();
 }
 
-void ax_runner_ax650::deinit()
+int ax_runner_ax650::inference()
 {
-    if (m_handle && m_handle->handle)
-    {
-        // free_io(&m_handle->io_data);
-        // mtensors.clear();
-        // minput_tensors.clear();
-        // map_input_tensors.clear();
-        // map_tensors.clear();
-        AX_ENGINE_DestroyHandle(m_handle->handle);
-        m_handle->handle = nullptr;
-        // delete m_handle;
-        // m_handle = nullptr;
+    if (!m_handle) return -1;
+    for (size_t i = 0; i < get_num_inputs(); i++) {
+        auto &tensor = get_input(i);
+        AX_SYS_MflushCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
     }
 
-    // AX_ENGINE_Deinit();
-}
+    int ret = AX_ENGINE_RunSync(m_handle->handle, &m_handle->io_data[0]);
 
-int ax_runner_ax650::inference()
-{
-    return AX_ENGINE_RunSync(m_handle->handle, &m_handle->io_data[0]);
+    for (size_t i = 0; i < get_num_outputs(); i++) {
+        auto &tensor = get_output(i);
+        AX_SYS_MinvalidateCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
+    }
+    return ret;
 }
 
 int ax_runner_ax650::inference(int grpid)
 {
-    return AX_ENGINE_RunGroupIOSync(m_handle->handle, m_handle->context, grpid, &m_handle->io_data[grpid]);
+    if (!m_handle) return -1;
+    if (grpid < 0 || grpid >= (int)m_handle->io_data.size()) return -1;
+
+    for (size_t i = 0; i < mgroup_input_tensors[grpid].size(); i++) {
+        auto &tensor = mgroup_input_tensors[grpid][i];
+        AX_SYS_MflushCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
+    }
+
+    int ret = AX_ENGINE_RunGroupIOSync(m_handle->handle, m_handle->context, grpid, &m_handle->io_data[grpid]);
+
+    for (size_t i = 0; i < mgroup_output_tensors[grpid].size(); i++) {
+        auto &tensor = mgroup_output_tensors[grpid][i];
+        AX_SYS_MinvalidateCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
+    }
+    return ret;
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner_ax650.hpp b/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner_ax650.hpp
index 351fb04a..62b65380 100644
--- a/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner_ax650.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/ax_model_runner/ax_model_runner_ax650.hpp
@@ -1,20 +1,23 @@
 #pragma once
 #include "ax_model_runner.hpp"
 
-class ax_runner_ax650 : public ax_runner_base
-{
-protected:
-    struct ax_joint_runner_ax650_handle_t *m_handle = nullptr;
-
-    bool _parepare_io = false;
+struct ax_runner_ax650_handle_t;
 
+class ax_runner_ax650 : public ax_runner_base {
+protected:
+    struct ax_runner_ax650_handle_t *m_handle = nullptr;
     int sub_init();
 
 public:
+    ax_runner_ax650() = default;
+    virtual ~ax_runner_ax650()
+    {
+        deinit();
+    }
+
     int init(const char *model_file, bool use_mmap = false) override;
     int init(char *model_buffer, size_t model_size) override;
 
-    void release();
     void deinit() override;
 
     int inference() override;

From cde5921d3b08040cb10f61e219976040ffeafe7d Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 18 Dec 2025 12:26:44 +0800
Subject: [PATCH 68/79] [update] add legacy llm backend

---
 .../main_cosy_voice/src/main.cpp              |  32 ++
 .../main_cosy_voice/src/runner/LLM.hpp        |  10 +-
 projects/llm_framework/main_llm/src/main.cpp  |  36 +-
 .../llm_framework/main_llm/src/runner/LLM.hpp |  21 +-
 .../legacy/ax_model_runner.hpp                | 177 ++++++++
 .../legacy/ax_model_runner_ax650.cpp          | 411 ++++++++++++++++++
 .../legacy/ax_model_runner_ax650.hpp          |  25 ++
 projects/llm_framework/main_vlm/src/main.cpp  |  32 ++
 .../llm_framework/main_vlm/src/runner/LLM.hpp |  43 +-
 .../legacy/ax_model_runner.hpp                | 177 ++++++++
 .../legacy/ax_model_runner_ax650.cpp          | 411 ++++++++++++++++++
 .../legacy/ax_model_runner_ax650.hpp          |  25 ++
 12 files changed, 1341 insertions(+), 59 deletions(-)
 create mode 100644 projects/llm_framework/main_llm/src/runner/ax_model_runner/legacy/ax_model_runner.hpp
 create mode 100644 projects/llm_framework/main_llm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.cpp
 create mode 100644 projects/llm_framework/main_llm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.hpp
 create mode 100644 projects/llm_framework/main_vlm/src/runner/ax_model_runner/legacy/ax_model_runner.hpp
 create mode 100644 projects/llm_framework/main_vlm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.cpp
 create mode 100644 projects/llm_framework/main_vlm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.hpp

diff --git a/projects/llm_framework/main_cosy_voice/src/main.cpp b/projects/llm_framework/main_cosy_voice/src/main.cpp
index d4adcd3d..a49b11ba 100644
--- a/projects/llm_framework/main_cosy_voice/src/main.cpp
+++ b/projects/llm_framework/main_cosy_voice/src/main.cpp
@@ -92,6 +92,7 @@ class llm_task {
     std::string kvcache_path;
     int precompute_len = 0;
     std::vector<int> _token_ids;
+    static int ax_init_flage_;
     task_callback_t out_callback_;
     bool enoutput_;
     bool enstream_;
@@ -630,9 +631,38 @@ class llm_task {
         return port;
     }
 
+    void _ax_init()
+    {
+        if (!ax_init_flage_) {
+            int ret = AX_SYS_Init();
+            if (0 != ret) {
+                fprintf(stderr, "AX_SYS_Init failed! ret = 0x%x\n", ret);
+            }
+            AX_ENGINE_NPU_ATTR_T npu_attr;
+            memset(&npu_attr, 0, sizeof(npu_attr));
+            ret = AX_ENGINE_Init(&npu_attr);
+            if (0 != ret) {
+                fprintf(stderr, "Init ax-engine failed{0x%8x}.\n", ret);
+            }
+        }
+        ax_init_flage_++;
+    }
+
+    void _ax_deinit()
+    {
+        if (ax_init_flage_ > 0) {
+            --ax_init_flage_;
+            if (!ax_init_flage_) {
+                AX_ENGINE_Deinit();
+                AX_SYS_Deinit();
+            }
+        }
+    }
+
     llm_task(const std::string &workid) : tokenizer_server_flage_(false), port_(getNextPort())
     {
         inference_run_ = std::make_unique<std::thread>(std::bind(&llm_task::run, this));
+        _ax_init();
     }
 
     void start()
@@ -666,10 +696,12 @@ class llm_task {
         if (lToken2Wav_) {
             lToken2Wav_->Deinit();
         }
+        _ax_deinit();
     }
 };
 
 std::atomic<unsigned int> llm_task::next_port_{8070};
+int llm_task::ax_init_flage_ = 0;
 
 #undef CONFIG_AUTO_SET
 
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp b/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
index bb187825..0c170c37 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
@@ -18,6 +18,7 @@
 #include "cqdm.h"
 #include "timer.hpp"
 #include "ax_sys_api.h"
+#include "ax_engine_api.h"
 #include "utils/sampling.hpp"
 #include "utils/utils.hpp"
 
@@ -463,20 +464,11 @@ class LLM {
 
                 layer.layer.inference(_attr.prefill_grpid);
 
-                auto &input_decoder_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
-                auto &input_decoder_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
-
                 auto &output_k_cache = layer.layer.get_output(_attr.prefill_grpid, "K_cache_out");
                 auto &output_v_cache = layer.layer.get_output(_attr.prefill_grpid, "V_cache_out");
 
                 int kv_offset = (_attr.precompute_len + p * _attr.prefill_token_num) * _attr.kv_cache_size;
 
-                memcpy((unsigned short *)input_decoder_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
-                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
-
-                memcpy((unsigned short *)input_decoder_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
-                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
-
                 for (int gid = _attr.prefill_grpid + 1; gid < prefill_split_num + 1; gid++) {
                     auto &input_prefill_k_cache = layer.layer.get_input(gid, "K_cache");
                     memcpy((unsigned short *)input_prefill_k_cache.pVirAddr + kv_offset,
diff --git a/projects/llm_framework/main_llm/src/main.cpp b/projects/llm_framework/main_llm/src/main.cpp
index 8ce63c1f..15128086 100644
--- a/projects/llm_framework/main_llm/src/main.cpp
+++ b/projects/llm_framework/main_llm/src/main.cpp
@@ -67,6 +67,7 @@ class llm_task {
     std::string kvcache_path;
     int precompute_len = 0;
     std::vector<int> _token_ids;
+    static int ax_init_flage_;
     task_callback_t out_callback_;
     bool enoutput_;
     bool enstream_;
@@ -208,7 +209,7 @@ class llm_task {
                     SLOGI("port_=%s model_id=%s content=%s", std::to_string(port_).c_str(),
                           (base_model + std::string("tokenizer")).c_str(), prompt_.c_str());
 
-                    std::this_thread::sleep_for(std::chrono::seconds(15));
+                    std::this_thread::sleep_for(std::chrono::seconds(5));
                 };
 
                 auto process_field = [&](std::string &field, const char *name_for_log) -> bool {
@@ -432,9 +433,38 @@ class llm_task {
         return port;
     }
 
+    void _ax_init()
+    {
+        if (!ax_init_flage_) {
+            int ret = AX_SYS_Init();
+            if (0 != ret) {
+                fprintf(stderr, "AX_SYS_Init failed! ret = 0x%x\n", ret);
+            }
+            AX_ENGINE_NPU_ATTR_T npu_attr;
+            memset(&npu_attr, 0, sizeof(npu_attr));
+            ret = AX_ENGINE_Init(&npu_attr);
+            if (0 != ret) {
+                fprintf(stderr, "Init ax-engine failed{0x%8x}.\n", ret);
+            }
+        }
+        ax_init_flage_++;
+    }
+
+    void _ax_deinit()
+    {
+        if (ax_init_flage_ > 0) {
+            --ax_init_flage_;
+            if (!ax_init_flage_) {
+                AX_ENGINE_Deinit();
+                AX_SYS_Deinit();
+            }
+        }
+    }
+
     llm_task(const std::string &workid) : tokenizer_server_flage_(false), port_(getNextPort())
     {
         inference_run_ = std::make_unique<std::thread>(std::bind(&llm_task::run, this));
+        _ax_init();
     }
 
     void start()
@@ -469,10 +499,12 @@ class llm_task {
         if (lLaMa_ctx_) {
             lLaMa_ctx_->Deinit();
         }
+        _ax_deinit();
     }
 };
 
 std::atomic<unsigned int> llm_task::next_port_{8080};
+int llm_task::ax_init_flage_ = 0;
 
 #undef CONFIG_AUTO_SET
 
@@ -527,7 +559,7 @@ class llm_llm : public StackFlow {
 
     void pause(const std::string &work_id, const std::string &object, const std::string &data) override
     {
-        SLOGI("llm_asr::work:%s", data.c_str());
+        SLOGI("llm_llm::work:%s", data.c_str());
 
         nlohmann::json error_body;
         int work_id_num = sample_get_work_id_num(work_id);
diff --git a/projects/llm_framework/main_llm/src/runner/LLM.hpp b/projects/llm_framework/main_llm/src/runner/LLM.hpp
index a5fb872f..fd89e0cd 100644
--- a/projects/llm_framework/main_llm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_llm/src/runner/LLM.hpp
@@ -7,12 +7,14 @@
 #include "Tokenizer/Tokenizer.hpp"
 #include "LLMEmbedSelector.hpp"
 #include "ax_model_runner/ax_model_runner_ax650.hpp"
+#include "ax_model_runner/legacy/ax_model_runner_ax650.hpp"
 #include "ax_cmm_utils.hpp"
 #include "cqdm.h"
 #include "timer.hpp"
 #include "LLMPostprocess.hpp"
 
-#include <ax_sys_api.h>
+#include "ax_sys_api.h"
+#include "ax_engine_api.h"
 
 #include <arm_neon.h>
 #define ALIGN_DOWN(x, a) ((x) & ~((a) - 1))
@@ -78,14 +80,14 @@ class LLM {
     LLMAttrType _attr;
 
     struct LLMLayer {
-        ax_runner_ax650 layer;
+        ax::legacy::ax_runner_ax650 layer;
         std::string filename;
         MMap layer_buffer;
         std::vector<char> layer_buffer_vec;
     };
 
     std::vector<LLMLayer> llama_layers;
-    ax_runner_ax650 llama_post;
+    ax::legacy::ax_runner_ax650 llama_post;
 
     int prefill_grpid = 1;
     int decode_grpid  = 0;
@@ -243,9 +245,9 @@ class LLM {
     void Deinit()
     {
         for (int i = 0; i < _attr.axmodel_num; i++) {
-            llama_layers[i].layer.deinit();
+            llama_layers[i].layer.release();
         }
-        llama_post.deinit();
+        llama_post.release();
         embed_selector.Deinit();
     }
 
@@ -1245,9 +1247,6 @@ class LLM_CTX {
 
                 layer.layer.inference(_attr.prefill_grpid);
 
-                auto &input_decoder_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
-                auto &input_decoder_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
-
                 auto &input_prefill_k_cache = layer.layer.get_input(_attr.prefill_grpid, "K_cache");
                 auto &input_prefill_v_cache = layer.layer.get_input(_attr.prefill_grpid, "V_cache");
 
@@ -1256,12 +1255,6 @@ class LLM_CTX {
 
                 int kv_offset = (_attr.precompute_len + p * _attr.prefill_token_num) * _attr.kv_cache_size;
 
-                memcpy((unsigned short *)input_decoder_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
-                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
-
-                memcpy((unsigned short *)input_decoder_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
-                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
-
                 memcpy((unsigned short *)input_prefill_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
                        sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
 
diff --git a/projects/llm_framework/main_llm/src/runner/ax_model_runner/legacy/ax_model_runner.hpp b/projects/llm_framework/main_llm/src/runner/ax_model_runner/legacy/ax_model_runner.hpp
new file mode 100644
index 00000000..6b2eccb5
--- /dev/null
+++ b/projects/llm_framework/main_llm/src/runner/ax_model_runner/legacy/ax_model_runner.hpp
@@ -0,0 +1,177 @@
+#pragma once
+#include <vector>
+#include <string>
+#include <map>
+#include <stdexcept>
+
+namespace ax::legacy {
+
+typedef enum _color_space_e {
+    axdl_color_space_unknown,
+    axdl_color_space_nv12,
+    axdl_color_space_nv21,
+    axdl_color_space_bgr,
+    axdl_color_space_rgb,
+} ax_color_space_e;
+
+typedef struct _image_t {
+    unsigned long long int pPhy;
+    void *pVir;
+    unsigned int nSize;
+    unsigned int nWidth;
+    unsigned int nHeight;
+    ax_color_space_e eDtype;
+    union {
+        int tStride_H, tStride_W, tStride_C;
+    };
+} ax_image_t;
+
+typedef struct {
+    std::string sName;
+    unsigned int nIdx;
+    std::vector<unsigned int> vShape;
+    int nSize;
+    unsigned long phyAddr;
+    void *pVirAddr;
+} ax_runner_tensor_t;
+
+class ax_runner_base {
+protected:
+    std::vector<ax_runner_tensor_t> moutput_tensors;
+    std::vector<ax_runner_tensor_t> minput_tensors;
+
+    std::vector<std::vector<ax_runner_tensor_t>> mgroup_output_tensors;
+    std::vector<std::vector<ax_runner_tensor_t>> mgroup_input_tensors;
+
+    std::map<std::string, ax_runner_tensor_t> map_output_tensors;
+    std::map<std::string, ax_runner_tensor_t> map_input_tensors;
+
+    std::map<std::string, std::vector<ax_runner_tensor_t>> map_group_output_tensors;
+    std::map<std::string, std::vector<ax_runner_tensor_t>> map_group_input_tensors;
+
+public:
+    virtual int init(const char *model_file, bool use_mmap = false) = 0;
+    virtual int init(char *model_buffer, size_t model_size)         = 0;
+
+    virtual void deinit() = 0;
+
+    int get_num_inputs()
+    {
+        return minput_tensors.size();
+    };
+    int get_num_outputs()
+    {
+        return moutput_tensors.size();
+    };
+
+    int get_num_input_groups()
+    {
+        return mgroup_input_tensors.size();
+    };
+    int get_num_output_groups()
+    {
+        return mgroup_output_tensors.size();
+    };
+
+    const ax_runner_tensor_t &get_input(int idx)
+    {
+        return minput_tensors[idx];
+    }
+    const ax_runner_tensor_t *get_inputs_ptr()
+    {
+        return minput_tensors.data();
+    }
+    const ax_runner_tensor_t &get_input(std::string name)
+    {
+        if (map_input_tensors.size() == 0) {
+            for (size_t i = 0; i < minput_tensors.size(); i++) {
+                map_input_tensors[minput_tensors[i].sName] = minput_tensors[i];
+            }
+        }
+        if (map_input_tensors.find(name) == map_input_tensors.end()) {
+            throw std::runtime_error("input tensor not found: " + name);
+        }
+
+        return map_input_tensors[name];
+    }
+
+    const ax_runner_tensor_t &get_input(int grpid, int idx)
+    {
+        return mgroup_input_tensors[grpid][idx];
+    }
+    const ax_runner_tensor_t *get_inputs_ptr(int grpid)
+    {
+        return mgroup_input_tensors[grpid].data();
+    }
+    const ax_runner_tensor_t &get_input(int grpid, std::string name)
+    {
+        if (map_group_input_tensors.size() == 0) {
+            for (size_t i = 0; i < mgroup_input_tensors.size(); i++) {
+                for (size_t j = 0; j < mgroup_input_tensors[i].size(); j++) {
+                    map_group_input_tensors[mgroup_input_tensors[i][j].sName].push_back(mgroup_input_tensors[i][j]);
+                }
+            }
+        }
+        if (map_group_input_tensors.find(name) == map_group_input_tensors.end()) {
+            throw std::runtime_error("input tensor not found: " + name);
+        }
+        return map_group_input_tensors[name][grpid];
+        // return map_input_tensors[name];
+    }
+
+    const ax_runner_tensor_t &get_output(int idx)
+    {
+        return moutput_tensors[idx];
+    }
+    const ax_runner_tensor_t *get_outputs_ptr()
+    {
+        return moutput_tensors.data();
+    }
+    const ax_runner_tensor_t &get_output(std::string name)
+    {
+        if (map_output_tensors.size() == 0) {
+            for (size_t i = 0; i < moutput_tensors.size(); i++) {
+                map_output_tensors[moutput_tensors[i].sName] = moutput_tensors[i];
+            }
+        }
+        if (map_output_tensors.find(name) == map_output_tensors.end()) {
+            throw std::runtime_error("output tensor not found: " + name);
+        }
+
+        return map_output_tensors[name];
+    }
+
+    const ax_runner_tensor_t &get_output(int grpid, int idx)
+    {
+        return mgroup_output_tensors[grpid][idx];
+    }
+    const ax_runner_tensor_t *get_outputs_ptr(int grpid)
+    {
+        return mgroup_output_tensors[grpid].data();
+    }
+    const ax_runner_tensor_t &get_output(int grpid, std::string name)
+    {
+        if (map_group_output_tensors.size() == 0) {
+            for (size_t i = 0; i < mgroup_output_tensors.size(); i++) {
+                for (size_t j = 0; j < mgroup_output_tensors[i].size(); j++) {
+                    map_group_output_tensors[mgroup_output_tensors[i][j].sName].push_back(mgroup_output_tensors[i][j]);
+                }
+            }
+        }
+        if (map_group_output_tensors.find(name) == map_group_output_tensors.end()) {
+            throw std::runtime_error("input tensor not found: " + name);
+        }
+        return map_group_output_tensors[name][grpid];
+    }
+
+    virtual int inference()          = 0;
+    virtual int inference(int grpid) = 0;
+
+    int operator()()
+    {
+        return inference();
+    }
+};
+
+// int ax_cmmcpy(unsigned long long int dst, unsigned long long int src, int size);
+}  // namespace ax::legacy
\ No newline at end of file
diff --git a/projects/llm_framework/main_llm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.cpp b/projects/llm_framework/main_llm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.cpp
new file mode 100644
index 00000000..6d6d3704
--- /dev/null
+++ b/projects/llm_framework/main_llm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.cpp
@@ -0,0 +1,411 @@
+#include "ax_model_runner_ax650.hpp"
+#include "string.h"
+#include "fstream"
+#include "memory"
+// #include "utilities/file.hpp"
+#include <ax_sys_api.h>
+#include <ax_ivps_api.h>
+#include <ax_engine_api.h>
+#include <fcntl.h>
+#include "memory_utils.hpp"
+#include "sample_log.h"
+
+#define AX_CMM_ALIGN_SIZE 128
+
+namespace ax::legacy {
+
+const char *AX_CMM_SESSION_NAME = "npu";
+
+typedef enum {
+    AX_ENGINE_ABST_DEFAULT = 0,
+    AX_ENGINE_ABST_CACHED  = 1,
+} AX_ENGINE_ALLOC_BUFFER_STRATEGY_T;
+
+typedef std::pair<AX_ENGINE_ALLOC_BUFFER_STRATEGY_T, AX_ENGINE_ALLOC_BUFFER_STRATEGY_T> INPUT_OUTPUT_ALLOC_STRATEGY;
+
+static void print_io_info(AX_ENGINE_IO_INFO_T *io_info)
+{
+    static std::map<AX_ENGINE_DATA_TYPE_T, const char *> data_type = {
+        {AX_ENGINE_DT_UNKNOWN, "UNKNOWN"},
+        {AX_ENGINE_DT_UINT8, "UINT8"},
+        {AX_ENGINE_DT_UINT16, "UINT16"},
+        {AX_ENGINE_DT_FLOAT32, "FLOAT32"},
+        {AX_ENGINE_DT_SINT16, "SINT16"},
+        {AX_ENGINE_DT_SINT8, "SINT8"},
+        {AX_ENGINE_DT_SINT32, "SINT32"},
+        {AX_ENGINE_DT_UINT32, "UINT32"},
+        {AX_ENGINE_DT_FLOAT64, "FLOAT64"},
+        {AX_ENGINE_DT_UINT10_PACKED, "UINT10_PACKED"},
+        {AX_ENGINE_DT_UINT12_PACKED, "UINT12_PACKED"},
+        {AX_ENGINE_DT_UINT14_PACKED, "UINT14_PACKED"},
+        {AX_ENGINE_DT_UINT16_PACKED, "UINT16_PACKED"},
+    };
+
+    static std::map<AX_ENGINE_COLOR_SPACE_T, const char *> color_type = {
+        {AX_ENGINE_CS_FEATUREMAP, "FEATUREMAP"},
+        {AX_ENGINE_CS_RAW8, "RAW8"},
+        {AX_ENGINE_CS_RAW10, "RAW10"},
+        {AX_ENGINE_CS_RAW12, "RAW12"},
+        {AX_ENGINE_CS_RAW14, "RAW14"},
+        {AX_ENGINE_CS_RAW16, "RAW16"},
+        {AX_ENGINE_CS_NV12, "NV12"},
+        {AX_ENGINE_CS_NV21, "NV21"},
+        {AX_ENGINE_CS_RGB, "RGB"},
+        {AX_ENGINE_CS_BGR, "BGR"},
+        {AX_ENGINE_CS_RGBA, "RGBA"},
+        {AX_ENGINE_CS_GRAY, "GRAY"},
+        {AX_ENGINE_CS_YUV444, "YUV444"},
+    };
+    printf("\ninput size: %d\n", io_info->nInputSize);
+    for (uint32_t i = 0; i < io_info->nInputSize; ++i) {
+        // print shape info,like [batchsize x channel x height x width]
+        auto &info = io_info->pInputs[i];
+        printf("    name: \e[1;32m%8s", info.pName);
+
+        std::string dt = "unknown";
+        if (data_type.find(info.eDataType) != data_type.end()) {
+            dt = data_type[info.eDataType];
+            printf(" \e[1;34m[%s] ", dt.c_str());
+        } else {
+            printf(" \e[1;31m[%s] ", dt.c_str());
+        }
+
+        std::string ct = "unknown";
+        if (info.pExtraMeta && color_type.find(info.pExtraMeta->eColorSpace) != color_type.end()) {
+            ct = color_type[info.pExtraMeta->eColorSpace];
+            printf("\e[1;34m[%s]", ct.c_str());
+        } else {
+            printf("\e[1;31m[%s]", ct.c_str());
+        }
+        printf(" \n        \e[1;31m");
+
+        for (AX_U8 s = 0; s < info.nShapeSize; s++) {
+            printf("%d", info.pShape[s]);
+            if (s != info.nShapeSize - 1) {
+                printf(" x ");
+            }
+        }
+        printf("\e[0m\n\n");
+    }
+
+    printf("\noutput size: %d\n", io_info->nOutputSize);
+    for (uint32_t i = 0; i < io_info->nOutputSize; ++i) {
+        // print shape info,like [batchsize x channel x height x width]
+        auto &info = io_info->pOutputs[i];
+        printf("    name: \e[1;32m%8s \e[1;34m[%s]\e[0m\n        \e[1;31m", info.pName, data_type[info.eDataType]);
+        for (AX_U8 s = 0; s < info.nShapeSize; s++) {
+            printf("%d", info.pShape[s]);
+            if (s != info.nShapeSize - 1) {
+                printf(" x ");
+            }
+        }
+        printf("\e[0m\n\n");
+    }
+}
+
+void free_io_index(AX_ENGINE_IO_BUFFER_T *io_buf, int index)
+{
+    for (int i = 0; i < index; ++i) {
+        AX_ENGINE_IO_BUFFER_T *pBuf = io_buf + i;
+        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+    }
+}
+
+void free_io(AX_ENGINE_IO_T *io)
+{
+    for (size_t j = 0; j < io->nInputSize; ++j) {
+        AX_ENGINE_IO_BUFFER_T *pBuf = io->pInputs + j;
+        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+    }
+    for (size_t j = 0; j < io->nOutputSize; ++j) {
+        AX_ENGINE_IO_BUFFER_T *pBuf = io->pOutputs + j;
+        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+    }
+    delete[] io->pInputs;
+    delete[] io->pOutputs;
+}
+
+static inline int prepare_io(AX_ENGINE_IO_INFO_T *info, AX_ENGINE_IO_T *io_data, INPUT_OUTPUT_ALLOC_STRATEGY strategy)
+{
+    memset(io_data, 0, sizeof(*io_data));
+    io_data->pInputs    = new AX_ENGINE_IO_BUFFER_T[info->nInputSize];
+    io_data->nInputSize = info->nInputSize;
+
+    auto ret = 0;
+    for (uint i = 0; i < info->nInputSize; ++i) {
+        auto meta   = info->pInputs[i];
+        auto buffer = &io_data->pInputs[i];
+        if (strategy.first == AX_ENGINE_ABST_CACHED) {
+            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE,
+                                        (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        } else {
+            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE,
+                                  (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        }
+
+        if (ret != 0) {
+            free_io_index(io_data->pInputs, i);
+            fprintf(stderr, "Allocate input{%d} { phy: %p, vir: %p, size: %lu Bytes }. fail \n", i,
+                    (void *)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+            return ret;
+        }
+        memset(buffer->pVirAddr, 0, meta.nSize);
+        // fprintf(stderr, "Allocate input{%d} { phy: %p, vir: %p, size: %lu Bytes }. \n", i, (void*)buffer->phyAddr,
+        // buffer->pVirAddr, (long)meta.nSize);
+    }
+
+    io_data->pOutputs    = new AX_ENGINE_IO_BUFFER_T[info->nOutputSize];
+    io_data->nOutputSize = info->nOutputSize;
+    for (uint i = 0; i < info->nOutputSize; ++i) {
+        auto meta     = info->pOutputs[i];
+        auto buffer   = &io_data->pOutputs[i];
+        buffer->nSize = meta.nSize;
+        if (strategy.second == AX_ENGINE_ABST_CACHED) {
+            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE,
+                                        (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        } else {
+            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE,
+                                  (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        }
+        if (ret != 0) {
+            fprintf(stderr, "Allocate output{%d} { phy: %p, vir: %p, size: %lu Bytes }. fail \n", i,
+                    (void *)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+            free_io_index(io_data->pInputs, io_data->nInputSize);
+            free_io_index(io_data->pOutputs, i);
+            return ret;
+        }
+        memset(buffer->pVirAddr, 0, meta.nSize);
+        // fprintf(stderr, "Allocate output{%d} { phy: %p, vir: %p, size: %lu Bytes }.\n", i, (void*)buffer->phyAddr,
+        // buffer->pVirAddr, (long)meta.nSize);
+    }
+
+    return 0;
+}
+
+struct ax_joint_runner_ax650_handle_t {
+    AX_ENGINE_HANDLE handle;
+    AX_ENGINE_CONTEXT_T context;
+    std::vector<AX_ENGINE_IO_INFO_T *> io_info;
+    std::vector<AX_ENGINE_IO_T> io_data;
+
+    // int algo_width, algo_height;
+    // int algo_colorformat;
+};
+
+int ax_runner_ax650::sub_init()
+{
+    // 4. create context
+    int ret = AX_ENGINE_CreateContext(m_handle->handle);
+    if (0 != ret) {
+        ALOGE("AX_ENGINE_CreateContext");
+        return ret;
+    }
+    ret = AX_ENGINE_CreateContextV2(m_handle->handle, &m_handle->context);
+    if (0 != ret) {
+        ALOGE("AX_ENGINE_CreateContextV2");
+        return ret;
+    }
+    // fprintf(stdout, "Engine creating context is done.\n");
+
+    // 5. set io
+    AX_U32 io_count = 0;
+    ret             = AX_ENGINE_GetGroupIOInfoCount(m_handle->handle, &io_count);
+    if (0 != ret) {
+        ALOGE("AX_ENGINE_GetGroupIOInfoCount");
+        return ret;
+    }
+    // ALOGI("io_count=%d", io_count);
+
+    m_handle->io_info.resize(io_count);
+    m_handle->io_data.resize(io_count);
+    mgroup_input_tensors.resize(io_count);
+    mgroup_output_tensors.resize(io_count);
+
+    // fprintf(stdout, "Engine get io info is done. \n");
+
+    // 6. alloc io
+    if (!_parepare_io) {
+        for (size_t grpid = 0; grpid < io_count; grpid++) {
+            AX_ENGINE_IO_INFO_T *io_info = nullptr;
+            ret                          = AX_ENGINE_GetGroupIOInfo(m_handle->handle, grpid, &io_info);
+            if (0 != ret) {
+                ALOGE("AX_ENGINE_GetIOInfo");
+                return ret;
+            }
+            // print_io_info(io_info);
+
+            m_handle->io_info[grpid] = io_info;
+
+            ret = prepare_io(m_handle->io_info[grpid], &m_handle->io_data[grpid],
+                             std::make_pair(AX_ENGINE_ABST_DEFAULT, AX_ENGINE_ABST_CACHED));
+            if (0 != ret) {
+                ALOGE("prepare_io grpid=%d", grpid);
+                return ret;
+            }
+        }
+
+        for (size_t grpid = 0; grpid < io_count; grpid++) {
+            auto &io_info = m_handle->io_info[grpid];
+            auto &io_data = m_handle->io_data[grpid];
+            for (size_t i = 0; i < io_info->nOutputSize; i++) {
+                ax_runner_tensor_t tensor;
+                tensor.nIdx  = i;
+                tensor.sName = std::string(io_info->pOutputs[i].pName);
+                tensor.nSize = io_info->pOutputs[i].nSize;
+                for (size_t j = 0; j < io_info->pOutputs[i].nShapeSize; j++) {
+                    tensor.vShape.push_back(io_info->pOutputs[i].pShape[j]);
+                }
+                tensor.phyAddr  = io_data.pOutputs[i].phyAddr;
+                tensor.pVirAddr = io_data.pOutputs[i].pVirAddr;
+                mgroup_output_tensors[grpid].push_back(tensor);
+            }
+
+            for (size_t i = 0; i < io_info->nInputSize; i++) {
+                ax_runner_tensor_t tensor;
+                tensor.nIdx  = i;
+                tensor.sName = std::string(io_info->pInputs[i].pName);
+                tensor.nSize = io_info->pInputs[i].nSize;
+                for (size_t j = 0; j < io_info->pInputs[i].nShapeSize; j++) {
+                    tensor.vShape.push_back(io_info->pInputs[i].pShape[j]);
+                }
+                tensor.phyAddr  = io_data.pInputs[i].phyAddr;
+                tensor.pVirAddr = io_data.pInputs[i].pVirAddr;
+                mgroup_input_tensors[grpid].push_back(tensor);
+            }
+        }
+
+        moutput_tensors = mgroup_output_tensors[0];
+        minput_tensors  = mgroup_input_tensors[0];
+
+        _parepare_io = true;
+    } else {
+    }
+
+    return ret;
+}
+
+int ax_runner_ax650::init(const char *model_file, bool use_mmap)
+{
+    if (use_mmap) {
+        MMap model_buffer(model_file);
+        if (!model_buffer.data()) {
+            ALOGE("mmap");
+            return -1;
+        }
+        auto ret = init((char *)model_buffer.data(), model_buffer.size());
+        model_buffer.close_file();
+        return ret;
+    } else {
+        char *model_buffer;
+        size_t len;
+        if (!read_file(model_file, &model_buffer, &len)) {
+            ALOGE("read_file");
+            return -1;
+        }
+        auto ret = init(model_buffer, len);
+        delete[] model_buffer;
+        return ret;
+    }
+}
+
+int ax_runner_ax650::init(char *model_buffer, size_t model_size)
+{
+    if (!m_handle) {
+        m_handle = new ax_joint_runner_ax650_handle_t;
+    }
+
+    // static bool b_init = false;
+    // if (!b_init) {
+    //     // 1. init engine
+    //     AX_ENGINE_NPU_ATTR_T npu_attr;
+    //     memset(&npu_attr, 0, sizeof(npu_attr));
+    //     npu_attr.eHardMode = AX_ENGINE_VIRTUAL_NPU_DISABLE;
+    //     AX_SYS_Init();
+    //     auto ret = AX_ENGINE_Init(&npu_attr);
+    //     if (0 != ret) {
+    //         return ret;
+    //     }
+    //     b_init = true;
+    // }
+
+    // 3. create handle
+
+    int ret = AX_ENGINE_CreateHandle(&m_handle->handle, model_buffer, model_size);
+    if (0 != ret) {
+        ALOGE("AX_ENGINE_CreateHandle");
+        return ret;
+    }
+    // fprintf(stdout, "Engine creating handle is done.\n");
+
+    return sub_init();
+}
+
+void ax_runner_ax650::release()
+{
+    if (m_handle && m_handle->handle) {
+        for (size_t i = 0; i < m_handle->io_data.size(); i++) {
+            /* code */
+            free_io(&m_handle->io_data[i]);
+        }
+
+        AX_ENGINE_DestroyHandle(m_handle->handle);
+        m_handle->handle = nullptr;
+    }
+
+    if (m_handle) {
+        delete m_handle;
+        m_handle = nullptr;
+    }
+
+    moutput_tensors.clear();
+    minput_tensors.clear();
+    map_input_tensors.clear();
+    map_output_tensors.clear();
+
+    mgroup_output_tensors.clear();
+    mgroup_input_tensors.clear();
+    map_group_input_tensors.clear();
+    map_group_output_tensors.clear();
+
+    // AX_ENGINE_Deinit();
+}
+
+void ax_runner_ax650::deinit()
+{
+    if (m_handle && m_handle->handle) {
+        // free_io(&m_handle->io_data);
+        // mtensors.clear();
+        // minput_tensors.clear();
+        // map_input_tensors.clear();
+        // map_tensors.clear();
+        AX_ENGINE_DestroyHandle(m_handle->handle);
+        m_handle->handle = nullptr;
+        // delete m_handle;
+        // m_handle = nullptr;
+    }
+
+    // AX_ENGINE_Deinit();
+}
+
+int ax_runner_ax650::inference()
+{
+    int ret = AX_ENGINE_RunSync(m_handle->handle, &m_handle->io_data[0]);
+    for (size_t i = 0; i < get_num_outputs(); i++) {
+        auto &tensor = get_output(i);
+        AX_SYS_MinvalidateCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
+    }
+    return ret;
+}
+
+int ax_runner_ax650::inference(int grpid)
+{
+    int ret = AX_ENGINE_RunGroupIOSync(m_handle->handle, m_handle->context, grpid, &m_handle->io_data[grpid]);
+
+    for (size_t i = 0; i < get_num_outputs(); i++) {
+        auto &tensor = get_output(grpid, i);
+        AX_SYS_MinvalidateCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
+    }
+    return ret;
+}
+
+}  // namespace ax::legacy
\ No newline at end of file
diff --git a/projects/llm_framework/main_llm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.hpp b/projects/llm_framework/main_llm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.hpp
new file mode 100644
index 00000000..a21b4319
--- /dev/null
+++ b/projects/llm_framework/main_llm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.hpp
@@ -0,0 +1,25 @@
+#pragma once
+#include "ax_model_runner.hpp"
+
+namespace ax::legacy {
+
+class ax_runner_ax650 : public ax_runner_base {
+protected:
+    struct ax_joint_runner_ax650_handle_t *m_handle = nullptr;
+
+    bool _parepare_io = false;
+
+    int sub_init();
+
+public:
+    int init(const char *model_file, bool use_mmap = false) override;
+    int init(char *model_buffer, size_t model_size) override;
+
+    void release();
+    void deinit() override;
+
+    int inference() override;
+    int inference(int grpid) override;
+};
+
+}  // namespace ax::legacy
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/main.cpp b/projects/llm_framework/main_vlm/src/main.cpp
index 9b552d36..a264a015 100644
--- a/projects/llm_framework/main_vlm/src/main.cpp
+++ b/projects/llm_framework/main_vlm/src/main.cpp
@@ -82,6 +82,7 @@ class llm_task {
     std::string kvcache_path;
     int precompute_len = 0;
     std::vector<int> _token_ids;
+    static int ax_init_flage_;
     task_callback_t out_callback_;
     bool enoutput_;
     bool enstream_;
@@ -572,6 +573,7 @@ class llm_task {
 
     llm_task(const std::string &workid) : tokenizer_server_flage_(false), port_(getNextPort())
     {
+        _ax_init();
     }
 
     void start()
@@ -582,6 +584,34 @@ class llm_task {
     {
     }
 
+    void _ax_init()
+    {
+        if (!ax_init_flage_) {
+            int ret = AX_SYS_Init();
+            if (0 != ret) {
+                fprintf(stderr, "AX_SYS_Init failed! ret = 0x%x\n", ret);
+            }
+            AX_ENGINE_NPU_ATTR_T npu_attr;
+            memset(&npu_attr, 0, sizeof(npu_attr));
+            ret = AX_ENGINE_Init(&npu_attr);
+            if (0 != ret) {
+                fprintf(stderr, "Init ax-engine failed{0x%8x}.\n", ret);
+            }
+        }
+        ax_init_flage_++;
+    }
+
+    void _ax_deinit()
+    {
+        if (ax_init_flage_ > 0) {
+            --ax_init_flage_;
+            if (!ax_init_flage_) {
+                AX_ENGINE_Deinit();
+                AX_SYS_Deinit();
+            }
+        }
+    }
+
     ~llm_task()
     {
         stop();
@@ -598,10 +628,12 @@ class llm_task {
         if (qwen_) {
             qwen_->Deinit();
         }
+        _ax_deinit();
     }
 };
 
 std::atomic<unsigned int> llm_task::next_port_{8090};
+int llm_task::ax_init_flage_ = 0;
 
 #undef CONFIG_AUTO_SET
 
diff --git a/projects/llm_framework/main_vlm/src/runner/LLM.hpp b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
index 3d6ad98c..262bc576 100644
--- a/projects/llm_framework/main_vlm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
@@ -11,11 +11,13 @@
 #include "Tokenizer/Tokenizer.hpp"
 #include "LLMEmbedSelector.hpp"
 #include "ax_model_runner/ax_model_runner_ax650.hpp"
+#include "ax_model_runner/legacy/ax_model_runner_ax650.hpp"
 #include "ax_cmm_utils.hpp"
 #include "cqdm.h"
 #include "timer.hpp"
 #include "opencv2/opencv.hpp"
 #include "ax_sys_api.h"
+#include "ax_engine_api.h"
 #include "LLMPostprocess.hpp"
 
 #define ALIGN_DOWN(x, a) ((x) & ~((a) - 1))
@@ -96,16 +98,16 @@ class LLM {
     LLMAttrType _attr;
 
     struct LLMLayer {
-        ax_runner_ax650 layer;
+        ax::legacy::ax_runner_ax650 layer;
         std::string filename;
         MMap layer_buffer;
         std::vector<char> layer_buffer_vec;
     };
 
     std::vector<LLMLayer> llama_layers;
-    ax_runner_ax650 llama_post;
+    ax::legacy::ax_runner_ax650 llama_post;
 
-    ax_runner_ax650 vpm_encoder, vpm_resampler;
+    ax::legacy::ax_runner_ax650 vpm_encoder, vpm_resampler;
 
     int prefill_grpid = 1;
     int decode_grpid  = 0;
@@ -285,11 +287,11 @@ class LLM {
     void Deinit()
     {
         for (int i = 0; i < _attr.axmodel_num; i++) {
-            llama_layers[i].layer.deinit();
+            llama_layers[i].layer.release();
         }
-        llama_post.deinit();
-        vpm_encoder.deinit();
-        vpm_resampler.deinit();
+        llama_post.release();
+        vpm_encoder.release();
+        vpm_resampler.release();
         embed_selector.Deinit();
     }
 
@@ -978,9 +980,6 @@ class LLM_CTX {
 
                 layer.layer.inference(prefill_grpid);
 
-                auto &input_decoder_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
-                auto &input_decoder_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
-
                 auto &input_prefill_k_cache = layer.layer.get_input(prefill_grpid, "K_cache");
                 auto &input_prefill_v_cache = layer.layer.get_input(prefill_grpid, "V_cache");
 
@@ -989,12 +988,6 @@ class LLM_CTX {
 
                 int kv_offset = (p * _attr.prefill_token_num) * _attr.kv_cache_size;
 
-                memcpy((unsigned short *)input_decoder_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
-                       sizeof(unsigned short) * _attr.prefill_token_num * _attr.kv_cache_size);
-
-                memcpy((unsigned short *)input_decoder_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
-                       sizeof(unsigned short) * _attr.prefill_token_num * _attr.kv_cache_size);
-
                 memcpy((unsigned short *)input_prefill_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
                        sizeof(unsigned short) * _attr.prefill_token_num * _attr.kv_cache_size);
 
@@ -1560,9 +1553,6 @@ class LLM_CTX {
 
                 layer.layer.inference(_attr.prefill_grpid);
 
-                auto &input_decoder_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
-                auto &input_decoder_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
-
                 auto &input_prefill_k_cache = layer.layer.get_input(_attr.prefill_grpid, "K_cache");
                 auto &input_prefill_v_cache = layer.layer.get_input(_attr.prefill_grpid, "V_cache");
 
@@ -1571,12 +1561,6 @@ class LLM_CTX {
 
                 int kv_offset = (_attr.precompute_len + p * _attr.prefill_token_num) * _attr.kv_cache_size;
 
-                memcpy((unsigned short *)input_decoder_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
-                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
-
-                memcpy((unsigned short *)input_decoder_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
-                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
-
                 memcpy((unsigned short *)input_prefill_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
                        sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
 
@@ -2265,20 +2249,11 @@ class LLM_Qwen {
 
                 layer.layer.inference(_attr.prefill_grpid);
 
-                auto &input_decoder_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
-                auto &input_decoder_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
-
                 auto &output_k_cache = layer.layer.get_output(_attr.prefill_grpid, "K_cache_out");
                 auto &output_v_cache = layer.layer.get_output(_attr.prefill_grpid, "V_cache_out");
 
                 int kv_offset = (_attr.precompute_len + p * _attr.prefill_token_num) * _attr.kv_cache_size;
 
-                memcpy((unsigned short *)input_decoder_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
-                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
-
-                memcpy((unsigned short *)input_decoder_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
-                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
-
                 for (int gid = _attr.prefill_grpid + 1; gid < prefill_split_num + 1; gid++) {
                     auto &input_prefill_k_cache = layer.layer.get_input(gid, "K_cache");
 
diff --git a/projects/llm_framework/main_vlm/src/runner/ax_model_runner/legacy/ax_model_runner.hpp b/projects/llm_framework/main_vlm/src/runner/ax_model_runner/legacy/ax_model_runner.hpp
new file mode 100644
index 00000000..6b2eccb5
--- /dev/null
+++ b/projects/llm_framework/main_vlm/src/runner/ax_model_runner/legacy/ax_model_runner.hpp
@@ -0,0 +1,177 @@
+#pragma once
+#include <vector>
+#include <string>
+#include <map>
+#include <stdexcept>
+
+namespace ax::legacy {
+
+typedef enum _color_space_e {
+    axdl_color_space_unknown,
+    axdl_color_space_nv12,
+    axdl_color_space_nv21,
+    axdl_color_space_bgr,
+    axdl_color_space_rgb,
+} ax_color_space_e;
+
+typedef struct _image_t {
+    unsigned long long int pPhy;
+    void *pVir;
+    unsigned int nSize;
+    unsigned int nWidth;
+    unsigned int nHeight;
+    ax_color_space_e eDtype;
+    union {
+        int tStride_H, tStride_W, tStride_C;
+    };
+} ax_image_t;
+
+typedef struct {
+    std::string sName;
+    unsigned int nIdx;
+    std::vector<unsigned int> vShape;
+    int nSize;
+    unsigned long phyAddr;
+    void *pVirAddr;
+} ax_runner_tensor_t;
+
+class ax_runner_base {
+protected:
+    std::vector<ax_runner_tensor_t> moutput_tensors;
+    std::vector<ax_runner_tensor_t> minput_tensors;
+
+    std::vector<std::vector<ax_runner_tensor_t>> mgroup_output_tensors;
+    std::vector<std::vector<ax_runner_tensor_t>> mgroup_input_tensors;
+
+    std::map<std::string, ax_runner_tensor_t> map_output_tensors;
+    std::map<std::string, ax_runner_tensor_t> map_input_tensors;
+
+    std::map<std::string, std::vector<ax_runner_tensor_t>> map_group_output_tensors;
+    std::map<std::string, std::vector<ax_runner_tensor_t>> map_group_input_tensors;
+
+public:
+    virtual int init(const char *model_file, bool use_mmap = false) = 0;
+    virtual int init(char *model_buffer, size_t model_size)         = 0;
+
+    virtual void deinit() = 0;
+
+    int get_num_inputs()
+    {
+        return minput_tensors.size();
+    };
+    int get_num_outputs()
+    {
+        return moutput_tensors.size();
+    };
+
+    int get_num_input_groups()
+    {
+        return mgroup_input_tensors.size();
+    };
+    int get_num_output_groups()
+    {
+        return mgroup_output_tensors.size();
+    };
+
+    const ax_runner_tensor_t &get_input(int idx)
+    {
+        return minput_tensors[idx];
+    }
+    const ax_runner_tensor_t *get_inputs_ptr()
+    {
+        return minput_tensors.data();
+    }
+    const ax_runner_tensor_t &get_input(std::string name)
+    {
+        if (map_input_tensors.size() == 0) {
+            for (size_t i = 0; i < minput_tensors.size(); i++) {
+                map_input_tensors[minput_tensors[i].sName] = minput_tensors[i];
+            }
+        }
+        if (map_input_tensors.find(name) == map_input_tensors.end()) {
+            throw std::runtime_error("input tensor not found: " + name);
+        }
+
+        return map_input_tensors[name];
+    }
+
+    const ax_runner_tensor_t &get_input(int grpid, int idx)
+    {
+        return mgroup_input_tensors[grpid][idx];
+    }
+    const ax_runner_tensor_t *get_inputs_ptr(int grpid)
+    {
+        return mgroup_input_tensors[grpid].data();
+    }
+    const ax_runner_tensor_t &get_input(int grpid, std::string name)
+    {
+        if (map_group_input_tensors.size() == 0) {
+            for (size_t i = 0; i < mgroup_input_tensors.size(); i++) {
+                for (size_t j = 0; j < mgroup_input_tensors[i].size(); j++) {
+                    map_group_input_tensors[mgroup_input_tensors[i][j].sName].push_back(mgroup_input_tensors[i][j]);
+                }
+            }
+        }
+        if (map_group_input_tensors.find(name) == map_group_input_tensors.end()) {
+            throw std::runtime_error("input tensor not found: " + name);
+        }
+        return map_group_input_tensors[name][grpid];
+        // return map_input_tensors[name];
+    }
+
+    const ax_runner_tensor_t &get_output(int idx)
+    {
+        return moutput_tensors[idx];
+    }
+    const ax_runner_tensor_t *get_outputs_ptr()
+    {
+        return moutput_tensors.data();
+    }
+    const ax_runner_tensor_t &get_output(std::string name)
+    {
+        if (map_output_tensors.size() == 0) {
+            for (size_t i = 0; i < moutput_tensors.size(); i++) {
+                map_output_tensors[moutput_tensors[i].sName] = moutput_tensors[i];
+            }
+        }
+        if (map_output_tensors.find(name) == map_output_tensors.end()) {
+            throw std::runtime_error("output tensor not found: " + name);
+        }
+
+        return map_output_tensors[name];
+    }
+
+    const ax_runner_tensor_t &get_output(int grpid, int idx)
+    {
+        return mgroup_output_tensors[grpid][idx];
+    }
+    const ax_runner_tensor_t *get_outputs_ptr(int grpid)
+    {
+        return mgroup_output_tensors[grpid].data();
+    }
+    const ax_runner_tensor_t &get_output(int grpid, std::string name)
+    {
+        if (map_group_output_tensors.size() == 0) {
+            for (size_t i = 0; i < mgroup_output_tensors.size(); i++) {
+                for (size_t j = 0; j < mgroup_output_tensors[i].size(); j++) {
+                    map_group_output_tensors[mgroup_output_tensors[i][j].sName].push_back(mgroup_output_tensors[i][j]);
+                }
+            }
+        }
+        if (map_group_output_tensors.find(name) == map_group_output_tensors.end()) {
+            throw std::runtime_error("input tensor not found: " + name);
+        }
+        return map_group_output_tensors[name][grpid];
+    }
+
+    virtual int inference()          = 0;
+    virtual int inference(int grpid) = 0;
+
+    int operator()()
+    {
+        return inference();
+    }
+};
+
+// int ax_cmmcpy(unsigned long long int dst, unsigned long long int src, int size);
+}  // namespace ax::legacy
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.cpp b/projects/llm_framework/main_vlm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.cpp
new file mode 100644
index 00000000..6d6d3704
--- /dev/null
+++ b/projects/llm_framework/main_vlm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.cpp
@@ -0,0 +1,411 @@
+#include "ax_model_runner_ax650.hpp"
+#include "string.h"
+#include "fstream"
+#include "memory"
+// #include "utilities/file.hpp"
+#include <ax_sys_api.h>
+#include <ax_ivps_api.h>
+#include <ax_engine_api.h>
+#include <fcntl.h>
+#include "memory_utils.hpp"
+#include "sample_log.h"
+
+#define AX_CMM_ALIGN_SIZE 128
+
+namespace ax::legacy {
+
+const char *AX_CMM_SESSION_NAME = "npu";
+
+typedef enum {
+    AX_ENGINE_ABST_DEFAULT = 0,
+    AX_ENGINE_ABST_CACHED  = 1,
+} AX_ENGINE_ALLOC_BUFFER_STRATEGY_T;
+
+typedef std::pair<AX_ENGINE_ALLOC_BUFFER_STRATEGY_T, AX_ENGINE_ALLOC_BUFFER_STRATEGY_T> INPUT_OUTPUT_ALLOC_STRATEGY;
+
+static void print_io_info(AX_ENGINE_IO_INFO_T *io_info)
+{
+    static std::map<AX_ENGINE_DATA_TYPE_T, const char *> data_type = {
+        {AX_ENGINE_DT_UNKNOWN, "UNKNOWN"},
+        {AX_ENGINE_DT_UINT8, "UINT8"},
+        {AX_ENGINE_DT_UINT16, "UINT16"},
+        {AX_ENGINE_DT_FLOAT32, "FLOAT32"},
+        {AX_ENGINE_DT_SINT16, "SINT16"},
+        {AX_ENGINE_DT_SINT8, "SINT8"},
+        {AX_ENGINE_DT_SINT32, "SINT32"},
+        {AX_ENGINE_DT_UINT32, "UINT32"},
+        {AX_ENGINE_DT_FLOAT64, "FLOAT64"},
+        {AX_ENGINE_DT_UINT10_PACKED, "UINT10_PACKED"},
+        {AX_ENGINE_DT_UINT12_PACKED, "UINT12_PACKED"},
+        {AX_ENGINE_DT_UINT14_PACKED, "UINT14_PACKED"},
+        {AX_ENGINE_DT_UINT16_PACKED, "UINT16_PACKED"},
+    };
+
+    static std::map<AX_ENGINE_COLOR_SPACE_T, const char *> color_type = {
+        {AX_ENGINE_CS_FEATUREMAP, "FEATUREMAP"},
+        {AX_ENGINE_CS_RAW8, "RAW8"},
+        {AX_ENGINE_CS_RAW10, "RAW10"},
+        {AX_ENGINE_CS_RAW12, "RAW12"},
+        {AX_ENGINE_CS_RAW14, "RAW14"},
+        {AX_ENGINE_CS_RAW16, "RAW16"},
+        {AX_ENGINE_CS_NV12, "NV12"},
+        {AX_ENGINE_CS_NV21, "NV21"},
+        {AX_ENGINE_CS_RGB, "RGB"},
+        {AX_ENGINE_CS_BGR, "BGR"},
+        {AX_ENGINE_CS_RGBA, "RGBA"},
+        {AX_ENGINE_CS_GRAY, "GRAY"},
+        {AX_ENGINE_CS_YUV444, "YUV444"},
+    };
+    printf("\ninput size: %d\n", io_info->nInputSize);
+    for (uint32_t i = 0; i < io_info->nInputSize; ++i) {
+        // print shape info,like [batchsize x channel x height x width]
+        auto &info = io_info->pInputs[i];
+        printf("    name: \e[1;32m%8s", info.pName);
+
+        std::string dt = "unknown";
+        if (data_type.find(info.eDataType) != data_type.end()) {
+            dt = data_type[info.eDataType];
+            printf(" \e[1;34m[%s] ", dt.c_str());
+        } else {
+            printf(" \e[1;31m[%s] ", dt.c_str());
+        }
+
+        std::string ct = "unknown";
+        if (info.pExtraMeta && color_type.find(info.pExtraMeta->eColorSpace) != color_type.end()) {
+            ct = color_type[info.pExtraMeta->eColorSpace];
+            printf("\e[1;34m[%s]", ct.c_str());
+        } else {
+            printf("\e[1;31m[%s]", ct.c_str());
+        }
+        printf(" \n        \e[1;31m");
+
+        for (AX_U8 s = 0; s < info.nShapeSize; s++) {
+            printf("%d", info.pShape[s]);
+            if (s != info.nShapeSize - 1) {
+                printf(" x ");
+            }
+        }
+        printf("\e[0m\n\n");
+    }
+
+    printf("\noutput size: %d\n", io_info->nOutputSize);
+    for (uint32_t i = 0; i < io_info->nOutputSize; ++i) {
+        // print shape info,like [batchsize x channel x height x width]
+        auto &info = io_info->pOutputs[i];
+        printf("    name: \e[1;32m%8s \e[1;34m[%s]\e[0m\n        \e[1;31m", info.pName, data_type[info.eDataType]);
+        for (AX_U8 s = 0; s < info.nShapeSize; s++) {
+            printf("%d", info.pShape[s]);
+            if (s != info.nShapeSize - 1) {
+                printf(" x ");
+            }
+        }
+        printf("\e[0m\n\n");
+    }
+}
+
+void free_io_index(AX_ENGINE_IO_BUFFER_T *io_buf, int index)
+{
+    for (int i = 0; i < index; ++i) {
+        AX_ENGINE_IO_BUFFER_T *pBuf = io_buf + i;
+        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+    }
+}
+
+void free_io(AX_ENGINE_IO_T *io)
+{
+    for (size_t j = 0; j < io->nInputSize; ++j) {
+        AX_ENGINE_IO_BUFFER_T *pBuf = io->pInputs + j;
+        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+    }
+    for (size_t j = 0; j < io->nOutputSize; ++j) {
+        AX_ENGINE_IO_BUFFER_T *pBuf = io->pOutputs + j;
+        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+    }
+    delete[] io->pInputs;
+    delete[] io->pOutputs;
+}
+
+static inline int prepare_io(AX_ENGINE_IO_INFO_T *info, AX_ENGINE_IO_T *io_data, INPUT_OUTPUT_ALLOC_STRATEGY strategy)
+{
+    memset(io_data, 0, sizeof(*io_data));
+    io_data->pInputs    = new AX_ENGINE_IO_BUFFER_T[info->nInputSize];
+    io_data->nInputSize = info->nInputSize;
+
+    auto ret = 0;
+    for (uint i = 0; i < info->nInputSize; ++i) {
+        auto meta   = info->pInputs[i];
+        auto buffer = &io_data->pInputs[i];
+        if (strategy.first == AX_ENGINE_ABST_CACHED) {
+            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE,
+                                        (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        } else {
+            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE,
+                                  (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        }
+
+        if (ret != 0) {
+            free_io_index(io_data->pInputs, i);
+            fprintf(stderr, "Allocate input{%d} { phy: %p, vir: %p, size: %lu Bytes }. fail \n", i,
+                    (void *)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+            return ret;
+        }
+        memset(buffer->pVirAddr, 0, meta.nSize);
+        // fprintf(stderr, "Allocate input{%d} { phy: %p, vir: %p, size: %lu Bytes }. \n", i, (void*)buffer->phyAddr,
+        // buffer->pVirAddr, (long)meta.nSize);
+    }
+
+    io_data->pOutputs    = new AX_ENGINE_IO_BUFFER_T[info->nOutputSize];
+    io_data->nOutputSize = info->nOutputSize;
+    for (uint i = 0; i < info->nOutputSize; ++i) {
+        auto meta     = info->pOutputs[i];
+        auto buffer   = &io_data->pOutputs[i];
+        buffer->nSize = meta.nSize;
+        if (strategy.second == AX_ENGINE_ABST_CACHED) {
+            ret = AX_SYS_MemAllocCached((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE,
+                                        (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        } else {
+            ret = AX_SYS_MemAlloc((AX_U64 *)(&buffer->phyAddr), &buffer->pVirAddr, meta.nSize, AX_CMM_ALIGN_SIZE,
+                                  (const AX_S8 *)(AX_CMM_SESSION_NAME));
+        }
+        if (ret != 0) {
+            fprintf(stderr, "Allocate output{%d} { phy: %p, vir: %p, size: %lu Bytes }. fail \n", i,
+                    (void *)buffer->phyAddr, buffer->pVirAddr, (long)meta.nSize);
+            free_io_index(io_data->pInputs, io_data->nInputSize);
+            free_io_index(io_data->pOutputs, i);
+            return ret;
+        }
+        memset(buffer->pVirAddr, 0, meta.nSize);
+        // fprintf(stderr, "Allocate output{%d} { phy: %p, vir: %p, size: %lu Bytes }.\n", i, (void*)buffer->phyAddr,
+        // buffer->pVirAddr, (long)meta.nSize);
+    }
+
+    return 0;
+}
+
+struct ax_joint_runner_ax650_handle_t {
+    AX_ENGINE_HANDLE handle;
+    AX_ENGINE_CONTEXT_T context;
+    std::vector<AX_ENGINE_IO_INFO_T *> io_info;
+    std::vector<AX_ENGINE_IO_T> io_data;
+
+    // int algo_width, algo_height;
+    // int algo_colorformat;
+};
+
+int ax_runner_ax650::sub_init()
+{
+    // 4. create context
+    int ret = AX_ENGINE_CreateContext(m_handle->handle);
+    if (0 != ret) {
+        ALOGE("AX_ENGINE_CreateContext");
+        return ret;
+    }
+    ret = AX_ENGINE_CreateContextV2(m_handle->handle, &m_handle->context);
+    if (0 != ret) {
+        ALOGE("AX_ENGINE_CreateContextV2");
+        return ret;
+    }
+    // fprintf(stdout, "Engine creating context is done.\n");
+
+    // 5. set io
+    AX_U32 io_count = 0;
+    ret             = AX_ENGINE_GetGroupIOInfoCount(m_handle->handle, &io_count);
+    if (0 != ret) {
+        ALOGE("AX_ENGINE_GetGroupIOInfoCount");
+        return ret;
+    }
+    // ALOGI("io_count=%d", io_count);
+
+    m_handle->io_info.resize(io_count);
+    m_handle->io_data.resize(io_count);
+    mgroup_input_tensors.resize(io_count);
+    mgroup_output_tensors.resize(io_count);
+
+    // fprintf(stdout, "Engine get io info is done. \n");
+
+    // 6. alloc io
+    if (!_parepare_io) {
+        for (size_t grpid = 0; grpid < io_count; grpid++) {
+            AX_ENGINE_IO_INFO_T *io_info = nullptr;
+            ret                          = AX_ENGINE_GetGroupIOInfo(m_handle->handle, grpid, &io_info);
+            if (0 != ret) {
+                ALOGE("AX_ENGINE_GetIOInfo");
+                return ret;
+            }
+            // print_io_info(io_info);
+
+            m_handle->io_info[grpid] = io_info;
+
+            ret = prepare_io(m_handle->io_info[grpid], &m_handle->io_data[grpid],
+                             std::make_pair(AX_ENGINE_ABST_DEFAULT, AX_ENGINE_ABST_CACHED));
+            if (0 != ret) {
+                ALOGE("prepare_io grpid=%d", grpid);
+                return ret;
+            }
+        }
+
+        for (size_t grpid = 0; grpid < io_count; grpid++) {
+            auto &io_info = m_handle->io_info[grpid];
+            auto &io_data = m_handle->io_data[grpid];
+            for (size_t i = 0; i < io_info->nOutputSize; i++) {
+                ax_runner_tensor_t tensor;
+                tensor.nIdx  = i;
+                tensor.sName = std::string(io_info->pOutputs[i].pName);
+                tensor.nSize = io_info->pOutputs[i].nSize;
+                for (size_t j = 0; j < io_info->pOutputs[i].nShapeSize; j++) {
+                    tensor.vShape.push_back(io_info->pOutputs[i].pShape[j]);
+                }
+                tensor.phyAddr  = io_data.pOutputs[i].phyAddr;
+                tensor.pVirAddr = io_data.pOutputs[i].pVirAddr;
+                mgroup_output_tensors[grpid].push_back(tensor);
+            }
+
+            for (size_t i = 0; i < io_info->nInputSize; i++) {
+                ax_runner_tensor_t tensor;
+                tensor.nIdx  = i;
+                tensor.sName = std::string(io_info->pInputs[i].pName);
+                tensor.nSize = io_info->pInputs[i].nSize;
+                for (size_t j = 0; j < io_info->pInputs[i].nShapeSize; j++) {
+                    tensor.vShape.push_back(io_info->pInputs[i].pShape[j]);
+                }
+                tensor.phyAddr  = io_data.pInputs[i].phyAddr;
+                tensor.pVirAddr = io_data.pInputs[i].pVirAddr;
+                mgroup_input_tensors[grpid].push_back(tensor);
+            }
+        }
+
+        moutput_tensors = mgroup_output_tensors[0];
+        minput_tensors  = mgroup_input_tensors[0];
+
+        _parepare_io = true;
+    } else {
+    }
+
+    return ret;
+}
+
+int ax_runner_ax650::init(const char *model_file, bool use_mmap)
+{
+    if (use_mmap) {
+        MMap model_buffer(model_file);
+        if (!model_buffer.data()) {
+            ALOGE("mmap");
+            return -1;
+        }
+        auto ret = init((char *)model_buffer.data(), model_buffer.size());
+        model_buffer.close_file();
+        return ret;
+    } else {
+        char *model_buffer;
+        size_t len;
+        if (!read_file(model_file, &model_buffer, &len)) {
+            ALOGE("read_file");
+            return -1;
+        }
+        auto ret = init(model_buffer, len);
+        delete[] model_buffer;
+        return ret;
+    }
+}
+
+int ax_runner_ax650::init(char *model_buffer, size_t model_size)
+{
+    if (!m_handle) {
+        m_handle = new ax_joint_runner_ax650_handle_t;
+    }
+
+    // static bool b_init = false;
+    // if (!b_init) {
+    //     // 1. init engine
+    //     AX_ENGINE_NPU_ATTR_T npu_attr;
+    //     memset(&npu_attr, 0, sizeof(npu_attr));
+    //     npu_attr.eHardMode = AX_ENGINE_VIRTUAL_NPU_DISABLE;
+    //     AX_SYS_Init();
+    //     auto ret = AX_ENGINE_Init(&npu_attr);
+    //     if (0 != ret) {
+    //         return ret;
+    //     }
+    //     b_init = true;
+    // }
+
+    // 3. create handle
+
+    int ret = AX_ENGINE_CreateHandle(&m_handle->handle, model_buffer, model_size);
+    if (0 != ret) {
+        ALOGE("AX_ENGINE_CreateHandle");
+        return ret;
+    }
+    // fprintf(stdout, "Engine creating handle is done.\n");
+
+    return sub_init();
+}
+
+void ax_runner_ax650::release()
+{
+    if (m_handle && m_handle->handle) {
+        for (size_t i = 0; i < m_handle->io_data.size(); i++) {
+            /* code */
+            free_io(&m_handle->io_data[i]);
+        }
+
+        AX_ENGINE_DestroyHandle(m_handle->handle);
+        m_handle->handle = nullptr;
+    }
+
+    if (m_handle) {
+        delete m_handle;
+        m_handle = nullptr;
+    }
+
+    moutput_tensors.clear();
+    minput_tensors.clear();
+    map_input_tensors.clear();
+    map_output_tensors.clear();
+
+    mgroup_output_tensors.clear();
+    mgroup_input_tensors.clear();
+    map_group_input_tensors.clear();
+    map_group_output_tensors.clear();
+
+    // AX_ENGINE_Deinit();
+}
+
+void ax_runner_ax650::deinit()
+{
+    if (m_handle && m_handle->handle) {
+        // free_io(&m_handle->io_data);
+        // mtensors.clear();
+        // minput_tensors.clear();
+        // map_input_tensors.clear();
+        // map_tensors.clear();
+        AX_ENGINE_DestroyHandle(m_handle->handle);
+        m_handle->handle = nullptr;
+        // delete m_handle;
+        // m_handle = nullptr;
+    }
+
+    // AX_ENGINE_Deinit();
+}
+
+int ax_runner_ax650::inference()
+{
+    int ret = AX_ENGINE_RunSync(m_handle->handle, &m_handle->io_data[0]);
+    for (size_t i = 0; i < get_num_outputs(); i++) {
+        auto &tensor = get_output(i);
+        AX_SYS_MinvalidateCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
+    }
+    return ret;
+}
+
+int ax_runner_ax650::inference(int grpid)
+{
+    int ret = AX_ENGINE_RunGroupIOSync(m_handle->handle, m_handle->context, grpid, &m_handle->io_data[grpid]);
+
+    for (size_t i = 0; i < get_num_outputs(); i++) {
+        auto &tensor = get_output(grpid, i);
+        AX_SYS_MinvalidateCache(tensor.phyAddr, tensor.pVirAddr, tensor.nSize);
+    }
+    return ret;
+}
+
+}  // namespace ax::legacy
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.hpp b/projects/llm_framework/main_vlm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.hpp
new file mode 100644
index 00000000..a21b4319
--- /dev/null
+++ b/projects/llm_framework/main_vlm/src/runner/ax_model_runner/legacy/ax_model_runner_ax650.hpp
@@ -0,0 +1,25 @@
+#pragma once
+#include "ax_model_runner.hpp"
+
+namespace ax::legacy {
+
+class ax_runner_ax650 : public ax_runner_base {
+protected:
+    struct ax_joint_runner_ax650_handle_t *m_handle = nullptr;
+
+    bool _parepare_io = false;
+
+    int sub_init();
+
+public:
+    int init(const char *model_file, bool use_mmap = false) override;
+    int init(char *model_buffer, size_t model_size) override;
+
+    void release();
+    void deinit() override;
+
+    int inference() override;
+    int inference(int grpid) override;
+};
+
+}  // namespace ax::legacy
\ No newline at end of file

From 50e36093534a05d20697e85003e2741d7a8410ec Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 18 Dec 2025 15:32:37 +0800
Subject: [PATCH 69/79] [update] Reduce model loading time. Optimize model
 loading method

---
 .../main_cosy_voice/src/main.cpp              |  2 +-
 .../main_cosy_voice/src/runner/LLM.hpp        |  6 ++--
 .../llm_framework/main_llm/src/runner/LLM.hpp | 17 +++--------
 projects/llm_framework/main_vlm/src/main.cpp  |  4 +--
 .../llm_framework/main_vlm/src/runner/LLM.hpp | 29 ++++++++++++-------
 5 files changed, 29 insertions(+), 29 deletions(-)

diff --git a/projects/llm_framework/main_cosy_voice/src/main.cpp b/projects/llm_framework/main_cosy_voice/src/main.cpp
index a49b11ba..324e9f38 100644
--- a/projects/llm_framework/main_cosy_voice/src/main.cpp
+++ b/projects/llm_framework/main_cosy_voice/src/main.cpp
@@ -244,7 +244,7 @@ class llm_task {
                     SLOGI("port_=%s model_id=%s content=%s", std::to_string(port_).c_str(),
                           (base_model + std::string("tokenizer")).c_str(), prompt_.c_str());
 
-                    std::this_thread::sleep_for(std::chrono::seconds(15));
+                    std::this_thread::sleep_for(std::chrono::seconds(5));
                 };
 
                 auto process_field = [&](std::string &field, const char *name_for_log) -> bool {
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp b/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
index 0c170c37..987f6816 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/LLM.hpp
@@ -149,7 +149,7 @@ class LLM {
             llama_layers[i].filename = axmodel_path;
 
             if (!attr.b_dynamic_load_axmodel_layer) {
-                int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), false);
+                int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), true);
                 if (ret != 0) {
                     ALOGE("init axmodel(%s) failed", llama_layers[i].filename.c_str());
                     return false;
@@ -172,12 +172,12 @@ class LLM {
             }
         }
 
-        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), false);
+        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), true);
         if (ret != 0) {
             ALOGE("init post axmodel(%s) failed", attr.filename_post_axmodel.c_str());
             return false;
         }
-        ret = llm_decoder.init(attr.filename_decoder_axmodel.c_str(), false);
+        ret = llm_decoder.init(attr.filename_decoder_axmodel.c_str(), true);
         if (ret != 0) {
             ALOGE("init llm decoder axmodel(%s) failed", attr.filename_decoder_axmodel.c_str());
             return false;
diff --git a/projects/llm_framework/main_llm/src/runner/LLM.hpp b/projects/llm_framework/main_llm/src/runner/LLM.hpp
index fd89e0cd..be9dd712 100644
--- a/projects/llm_framework/main_llm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_llm/src/runner/LLM.hpp
@@ -139,7 +139,7 @@ class LLM {
             llama_layers[i].filename = axmodel_path;
 
             if (!attr.b_dynamic_load_axmodel_layer) {
-                int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), false);
+                int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), true);
                 if (ret != 0) {
                     ALOGE("init axmodel(%s) failed", llama_layers[i].filename.c_str());
                     return false;
@@ -162,7 +162,7 @@ class LLM {
             }
         }
 
-        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), false);
+        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), true);
         if (ret != 0) {
             ALOGE("init post axmodel(%s) failed", attr.filename_post_axmodel.c_str());
             return false;
@@ -602,7 +602,7 @@ class LLM_CTX {
             sprintf(axmodel_path, attr.template_filename_axmodel.c_str(), i);
             llama_layers[i].filename = axmodel_path;
 
-            int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), false);
+            int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), true);
             if (ret != 0) {
                 ALOGE("init axmodel(%s) failed", llama_layers[i].filename.c_str());
                 return false;
@@ -612,7 +612,7 @@ class LLM_CTX {
             update_cqdm(&cqdm, i + 2, "count", axmodel_path);
         }
 
-        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), false);
+        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), true);
         if (ret != 0) {
             ALOGE("init post axmodel(%s) failed", attr.filename_post_axmodel.c_str());
             return false;
@@ -810,9 +810,6 @@ class LLM_CTX {
 
                 layer.layer.inference(prefill_grpid);
 
-                auto &input_decoder_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
-                auto &input_decoder_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
-
                 auto &input_prefill_k_cache = layer.layer.get_input(prefill_grpid, "K_cache");
                 auto &input_prefill_v_cache = layer.layer.get_input(prefill_grpid, "V_cache");
 
@@ -821,12 +818,6 @@ class LLM_CTX {
 
                 int kv_offset = (p * _attr.prefill_token_num) * _attr.kv_cache_size;
 
-                memcpy((unsigned short *)input_decoder_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
-                       sizeof(unsigned short) * _attr.prefill_token_num * _attr.kv_cache_size);
-
-                memcpy((unsigned short *)input_decoder_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
-                       sizeof(unsigned short) * _attr.prefill_token_num * _attr.kv_cache_size);
-
                 memcpy((unsigned short *)input_prefill_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
                        sizeof(unsigned short) * _attr.prefill_token_num * _attr.kv_cache_size);
 
diff --git a/projects/llm_framework/main_vlm/src/main.cpp b/projects/llm_framework/main_vlm/src/main.cpp
index a264a015..113236bc 100644
--- a/projects/llm_framework/main_vlm/src/main.cpp
+++ b/projects/llm_framework/main_vlm/src/main.cpp
@@ -263,7 +263,7 @@ class llm_task {
                     SLOGI("port_=%s model_id=%s content=%s", std::to_string(port_).c_str(),
                           (base_model + std::string("tokenizer")).c_str(), prompt_.c_str());
 
-                    std::this_thread::sleep_for(std::chrono::seconds(15));
+                    std::this_thread::sleep_for(std::chrono::seconds(5));
                 };
 
                 auto process_field = [&](std::string &field, const char *name_for_log) -> bool {
@@ -291,7 +291,7 @@ class llm_task {
                     model_type_ = ModelType::Qwen;
                 else if (encoder_name.find("internvl3") != std::string::npos && mode_config_.precompute_len > 0)
                     model_type_ = ModelType::InternVL_CTX;
-                else if (encoder_name.find("internvl3") != std::string::npos)
+                else if ((encoder_name.find("internvl3") != std::string::npos) || (encoder_name.find("vpm") != std::string::npos))
                     model_type_ = ModelType::InternVL;
                 else
                     model_type_ = ModelType::Unknown;
diff --git a/projects/llm_framework/main_vlm/src/runner/LLM.hpp b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
index 262bc576..69b56f33 100644
--- a/projects/llm_framework/main_vlm/src/runner/LLM.hpp
+++ b/projects/llm_framework/main_vlm/src/runner/LLM.hpp
@@ -155,7 +155,7 @@ class LLM {
             llama_layers[i].filename = axmodel_path;
 
             if (!attr.b_dynamic_load_axmodel_layer) {
-                int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), false);
+                int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), true);
                 if (ret != 0) {
                     ALOGE("init axmodel(%s) failed", llama_layers[i].filename.c_str());
                     return false;
@@ -178,7 +178,7 @@ class LLM {
             }
         }
 
-        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), false);
+        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), true);
         if (ret != 0) {
             ALOGE("init post axmodel(%s) failed", attr.filename_post_axmodel.c_str());
             return false;
@@ -188,13 +188,13 @@ class LLM {
         update_cqdm(&cqdm, attr.axmodel_num + 2, "count", axmodel_path);
 
         if (_attr.b_vpm_two_stage) {
-            ret = vpm_encoder.init(attr.filename_vpm_encoder_axmodel.c_str(), false);
+            ret = vpm_encoder.init(attr.filename_vpm_encoder_axmodel.c_str(), true);
             if (ret != 0) {
                 ALOGE("init vpm axmodel(%s) failed", attr.filename_vpm_encoder_axmodel.c_str());
                 return false;
             }
 
-            ret = vpm_resampler.init(attr.filename_vpm_resampler_axmodedl.c_str(), false);
+            ret = vpm_resampler.init(attr.filename_vpm_resampler_axmodedl.c_str(), true);
             if (ret != 0) {
                 ALOGE("init vpm axmodel(%s) failed", attr.filename_vpm_resampler_axmodedl.c_str());
                 return false;
@@ -203,7 +203,7 @@ class LLM {
             _attr.vpm_height = vpm_encoder.get_input(0).vShape[1];
             _attr.vpm_width  = vpm_encoder.get_input(0).vShape[2];
         } else {
-            ret = vpm_resampler.init(attr.filename_vpm_resampler_axmodedl.c_str(), false);
+            ret = vpm_resampler.init(attr.filename_vpm_resampler_axmodedl.c_str(), true);
             if (ret != 0) {
                 ALOGE("init vpm axmodel(%s) failed", attr.filename_vpm_resampler_axmodedl.c_str());
                 return false;
@@ -716,7 +716,7 @@ class LLM_CTX {
             sprintf(axmodel_path, attr.template_filename_axmodel.c_str(), i);
             llama_layers[i].filename = axmodel_path;
 
-            int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), false);
+            int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), true);
             if (ret != 0) {
                 ALOGE("init axmodel(%s) failed", llama_layers[i].filename.c_str());
                 return false;
@@ -726,7 +726,7 @@ class LLM_CTX {
             update_cqdm(&cqdm, i + 2, "count", axmodel_path);
         }
 
-        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), false);
+        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), true);
         if (ret != 0) {
             ALOGE("init post axmodel(%s) failed", attr.filename_post_axmodel.c_str());
             return false;
@@ -1773,7 +1773,7 @@ class LLM_Qwen {
             llama_layers[i].filename = axmodel_path;
 
             if (!attr.b_dynamic_load_axmodel_layer) {
-                int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), false);
+                int ret = llama_layers[i].layer.init(llama_layers[i].filename.c_str(), true);
                 if (ret != 0) {
                     ALOGE("init axmodel(%s) failed", llama_layers[i].filename.c_str());
                     return false;
@@ -1796,7 +1796,7 @@ class LLM_Qwen {
             }
         }
 
-        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), false);
+        int ret = llama_post.init(attr.filename_post_axmodel.c_str(), true);
         if (ret != 0) {
             ALOGE("init post axmodel(%s) failed", attr.filename_post_axmodel.c_str());
             return false;
@@ -1805,7 +1805,7 @@ class LLM_Qwen {
         sprintf(axmodel_path, "init post axmodel ok,remain_cmm(%d MB)", remain_cmm);
         update_cqdm(&cqdm, attr.axmodel_num + 2, "count", axmodel_path);
 
-        ret = image_encoder.init(attr.filename_image_encoder_axmodel.c_str(), false);
+        ret = image_encoder.init(attr.filename_image_encoder_axmodel.c_str(), true);
         if (ret != 0) {
             ALOGE("init image_encoder axmodel(%s) failed", attr.filename_image_encoder_axmodel.c_str());
             return false;
@@ -2249,11 +2249,20 @@ class LLM_Qwen {
 
                 layer.layer.inference(_attr.prefill_grpid);
 
+                auto &input_decoder_k_cache = layer.layer.get_input(decode_grpid, "K_cache");
+                auto &input_decoder_v_cache = layer.layer.get_input(decode_grpid, "V_cache");
+
                 auto &output_k_cache = layer.layer.get_output(_attr.prefill_grpid, "K_cache_out");
                 auto &output_v_cache = layer.layer.get_output(_attr.prefill_grpid, "V_cache_out");
 
                 int kv_offset = (_attr.precompute_len + p * _attr.prefill_token_num) * _attr.kv_cache_size;
 
+                memcpy((unsigned short *)input_decoder_k_cache.pVirAddr + kv_offset, (void *)output_k_cache.pVirAddr,
+                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+
+                memcpy((unsigned short *)input_decoder_v_cache.pVirAddr + kv_offset, (void *)output_v_cache.pVirAddr,
+                       sizeof(unsigned short) * input_num_token * _attr.kv_cache_size);
+
                 for (int gid = _attr.prefill_grpid + 1; gid < prefill_split_num + 1; gid++) {
                     auto &input_prefill_k_cache = layer.layer.get_input(gid, "K_cache");
 

From ea7ddd0254948f6a69f187e2a02a0d0deb5313ab Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 18 Dec 2025 16:03:17 +0800
Subject: [PATCH 70/79] [update] Add CosyVoice tokenizer server  timeout

---
 .../src/runner/Tokenizer/Tokenizer.cpp               | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp
index 5c34cb4c..0829ffaf 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp
@@ -349,11 +349,17 @@ class Tokenizer_Http : public BaseTokenizer {
     bool Init(std::string model_path = "http://localhost:8080", bool b_bos = true, bool b_eos = false) override
     {
         base_url = model_path;
+        if (!test_connect_http(base_url, 20)) {
+            ALOGE("connect %s failed", base_url.c_str());
+            return false;
+        } else {
+            ALOGI("connect %s ok", base_url.c_str());
+        }
         try {
             cli = std::make_shared<httplib::Client>(base_url);
-            cli->set_connection_timeout(1);
-            cli->set_read_timeout(1);
-            cli->set_write_timeout(1);
+            cli->set_connection_timeout(10);
+            cli->set_read_timeout(10);
+            cli->set_write_timeout(10);
             {
                 auto ret = cli->Get("/bos_id");
                 auto rep = ret.value();

From d5685d4419f24eea9034c98bd7c420a07fa9292d Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 18 Dec 2025 17:06:31 +0800
Subject: [PATCH 71/79] [update] kws support axmodel

---
 projects/llm_framework/main_kws/SConstruct    |   8 +-
 projects/llm_framework/main_kws/src/main.cpp  | 129 ++--
 .../main_kws/src/runner/EngineWrapper.cpp     | 357 +++++++++++
 .../main_kws/src/runner/EngineWrapper.hpp     |  53 ++
 .../main_kws/src/runner/utils/checker.h       |  32 +
 .../main_kws/src/runner/utils/io.hpp          | 606 ++++++++++++++++++
 .../main_kws/src/runner/utils/logger.h        |  84 +++
 7 files changed, 1225 insertions(+), 44 deletions(-)
 create mode 100644 projects/llm_framework/main_kws/src/runner/EngineWrapper.cpp
 create mode 100644 projects/llm_framework/main_kws/src/runner/EngineWrapper.hpp
 create mode 100644 projects/llm_framework/main_kws/src/runner/utils/checker.h
 create mode 100644 projects/llm_framework/main_kws/src/runner/utils/io.hpp
 create mode 100644 projects/llm_framework/main_kws/src/runner/utils/logger.h

diff --git a/projects/llm_framework/main_kws/SConstruct b/projects/llm_framework/main_kws/SConstruct
index eb00040a..fc02a377 100644
--- a/projects/llm_framework/main_kws/SConstruct
+++ b/projects/llm_framework/main_kws/SConstruct
@@ -4,10 +4,10 @@ Import('env')
 with open(env['PROJECT_TOOL_S']) as f:
     exec(f.read())
 
-SRCS = Glob('src/*.c*')
+SRCS = append_srcs_dir(ADir('src'))
 INCLUDE = [ADir('include'), ADir('.')]
 PRIVATE_INCLUDE = []
-REQUIREMENTS = ['pthread', 'dl', 'utilities', 'eventpp', 'StackFlow', 'single_header_libs']
+REQUIREMENTS = ['pthread', 'dl', 'utilities', 'ax_msp', 'eventpp', 'StackFlow', 'single_header_libs']
 STATIC_LIB = []
 DYNAMIC_LIB = []
 DEFINITIONS = []
@@ -22,7 +22,8 @@ DEFINITIONS += ['-std=c++17']
 LDFLAGS+=['-Wl,-rpath=/opt/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib', '-Wl,-rpath=/usr/local/m5stack/lib/gcc-10.3', '-Wl,-rpath=/opt/lib', '-Wl,-rpath=/opt/usr/lib', '-Wl,-rpath=./']
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 
-INCLUDE += [ADir('../static_lib/include/sherpa'),
+INCLUDE += [ADir('src/runner'),
+            ADir('../static_lib/include/sherpa'),
             ADir('../static_lib/include/sherpa/fbank'),
             ADir('../static_lib/include/sherpa/sherpa-onnx'),
             ADir('../static_lib/include/sherpa/sherpa-onnx/onnxruntime-src'),
@@ -30,6 +31,7 @@ INCLUDE += [ADir('../static_lib/include/sherpa'),
             ]
 
 LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/onnx')]
+REQUIREMENTS += ['ax_engine', 'ax_interpreter', 'ax_sys']
 REQUIREMENTS += ['onnxruntime']
 
 LDFLAGS += ['-l:libsherpa-onnx-core.a', '-l:libkaldi-native-fbank-core.a','-l:libkissfft-float.a',
diff --git a/projects/llm_framework/main_kws/src/main.cpp b/projects/llm_framework/main_kws/src/main.cpp
index 51022a83..8e70be53 100644
--- a/projects/llm_framework/main_kws/src/main.cpp
+++ b/projects/llm_framework/main_kws/src/main.cpp
@@ -5,6 +5,9 @@
  */
 #include "StackFlow.h"
 
+#include "EngineWrapper.hpp"
+#include "ax_sys_api.h"
+
 #include <signal.h>
 #include <sys/stat.h>
 #include <sys/types.h>
@@ -34,9 +37,6 @@ static std::string base_model_config_path_;
 typedef std::function<void(const std::string &data, bool finish)> task_callback_t;
 
 #include "sherpa-onnx/csrc/keyword-spotter.h"
-#include "sherpa-onnx/csrc/parse-options.h"
-
-#include <onnxruntime_cxx_api.h>
 #include "kaldi-native-fbank/csrc/online-feature.h"
 
 typedef struct mode_config_onnx {
@@ -61,6 +61,7 @@ class llm_task {
     bool enstream_      = false;
     bool enwake_audio_  = true;
     std::atomic_bool audio_flage_;
+    static int ax_init_flage_;
     task_callback_t out_callback_;
     buffer_t *pcmdata;
     std::string wake_wav_file_;
@@ -72,8 +73,8 @@ class llm_task {
     std::unique_ptr<sherpa_onnx::OnlineStream> sherpa_stream_;
 
     kws_config_onnx onnx_config_;
-    std::vector<float> onnx_cache_;
-    std::unique_ptr<Ort::Session> onnx_session_;
+    std::vector<float> axera_cache_;
+    std::unique_ptr<EngineWrapper> axera_session_;
     knf::FbankOptions fbank_opts_;
     std::unique_ptr<knf::OnlineFbank> fbank_;
     Ort::Env onnx_env_{ORT_LOGGING_LEVEL_WARNING, "kws"};
@@ -328,16 +329,20 @@ class llm_task {
             }
             std::string base_model = base_model_path_ + model_ + "/";
             SLOGI("base_model %s", base_model.c_str());
-            std::string model_file = base_model + "kws.onnx";
+            std::string model_file = base_model + "kws.axmodel";
 
             if (config_body.contains("wake_wav_file"))
                 wake_wav_file_ = config_body["wake_wav_file"];
             else if (file_body["mode_param"].contains("wake_wav_file"))
                 wake_wav_file_ = file_body["mode_param"]["wake_wav_file"];
 
-            onnx_session_ = std::make_unique<Ort::Session>(onnx_env_, model_file.c_str(), session_options_);
+            axera_session_ = std::make_unique<EngineWrapper>();
+            if (0 != axera_session_->Init(model_file.c_str())) {
+                SLOGE("Init axera model failed!");
+                return -5;
+            }
 
-            onnx_cache_.assign(1 * 32 * 88, 0.0f);
+            axera_cache_.assign(1 * 32 * 88, 0.0f);
 
             auto &mp = file_body["mode_param"];
             CONFIG_AUTO_SET_ONNX(mp, chunk_size);
@@ -414,37 +419,46 @@ class llm_task {
 
     std::vector<float> run_inference(const std::vector<float> &audio_chunk_16k)
     {
-        std::vector<std::vector<float>> fbank_feats;
-        fbank_feats = compute_fbank_kaldi(audio_chunk_16k, onnx_config_.RESAMPLE_RATE, onnx_config_.FEAT_DIM);
-        if (fbank_feats.empty()) {
+        std::vector<std::vector<float>> fbank_feats =
+            compute_fbank_kaldi(audio_chunk_16k, onnx_config_.RESAMPLE_RATE, onnx_config_.FEAT_DIM);
+        if (fbank_feats.empty()) return {};
+
+        constexpr int FIX_T = 32;
+        const int FEAT_DIM  = onnx_config_.FEAT_DIM;
+
+        std::vector<float> mat_flattened;
+        mat_flattened.resize(FIX_T * FEAT_DIM, 0.0f);
+
+        const int T_in   = static_cast<int>(fbank_feats.size());
+        const int T_copy = std::min(T_in, FIX_T);
+
+        for (int t = 0; t < T_copy; ++t) {
+            if ((int)fbank_feats[t].size() < FEAT_DIM) continue;
+            std::memcpy(mat_flattened.data() + t * FEAT_DIM, fbank_feats[t].data(), sizeof(float) * FEAT_DIM);
+        }
+
+        axera_session_->SetInput(mat_flattened.data(), 0);
+
+        axera_session_->SetInput(axera_cache_.data(), 1);
+
+        int ret = axera_session_->Run();
+        if (ret) {
+            SLOGE("axera_session run failed!");
             return {};
         }
-        int T = fbank_feats.size();
-        std::vector<float> mat_flattened;
-        for (const auto &feat : fbank_feats) {
-            mat_flattened.insert(mat_flattened.end(), feat.begin(), feat.end());
-        }
-        std::vector<int64_t> input_shape = {1, static_cast<int64_t>(T), onnx_config_.FEAT_DIM};
-        std::vector<int64_t> cache_shape = {1, 32, 88};
-        Ort::MemoryInfo memory_info      = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
-        Ort::Value input_tensor          = Ort::Value::CreateTensor<float>(
-            memory_info, mat_flattened.data(), mat_flattened.size(), input_shape.data(), input_shape.size());
-        Ort::Value cache_tensor   = Ort::Value::CreateTensor<float>(memory_info, onnx_cache_.data(), onnx_cache_.size(),
-                                                                    cache_shape.data(), cache_shape.size());
-        const char *input_names[] = {"input", "cache"};
-        const char *output_names[] = {"output", "r_cache"};
-        std::vector<Ort::Value> inputs;
-        inputs.push_back(std::move(input_tensor));
-        inputs.push_back(std::move(cache_tensor));
-        auto output_tensors =
-            onnx_session_->Run(Ort::RunOptions{nullptr}, input_names, inputs.data(), 2, output_names, 2);
-        float *out_data                = output_tensors[0].GetTensorMutableData<float>();
-        float *cache_out_data          = output_tensors[1].GetTensorMutableData<float>();
-        std::vector<int64_t> out_shape = output_tensors[0].GetTensorTypeAndShapeInfo().GetShape();
-        size_t out_size                = 1;
-        for (auto dim : out_shape) out_size *= dim;
-        std::vector<float> out_chunk(out_data, out_data + out_size);
-        std::copy(cache_out_data, cache_out_data + onnx_cache_.size(), onnx_cache_.begin());
+
+        const float *out_ptr = reinterpret_cast<const float *>(axera_session_->GetOutputPtr(0));
+        size_t out_size_f    = axera_session_->GetOutputSize(0) / sizeof(float);
+        std::vector<float> out_chunk(out_ptr, out_ptr + out_size_f);
+
+        const float *cache_ptr = reinterpret_cast<const float *>(axera_session_->GetOutputPtr(1));
+        size_t cache_size_f    = axera_session_->GetOutputSize(1) / sizeof(float);
+        if (cache_size_f != axera_cache_.size()) {
+            SLOGE("cache size mismatch: out=%zu, local=%zu", cache_size_f, axera_cache_.size());
+            return out_chunk;
+        }
+        std::memcpy(axera_cache_.data(), cache_ptr, axera_cache_.size() * sizeof(float));
+
         return out_chunk;
     }
 
@@ -531,16 +545,45 @@ class llm_task {
 
     bool delete_model()
     {
-        sherpa_spotter_.reset();
-        sherpa_stream_.reset();
-        onnx_session_.reset();
-        fbank_.reset();
+        if (sherpa_spotter_) sherpa_spotter_.reset();
+        if (sherpa_stream_) sherpa_stream_.reset();
+        if (axera_session_) axera_session_->Release();
+        if (fbank_) fbank_.reset();
         return true;
     }
 
     llm_task(const std::string &workid) : audio_flage_(false)
     {
         pcmdata = buffer_create();
+        _ax_init();
+    }
+
+    void _ax_init()
+    {
+        if (!ax_init_flage_) {
+            int ret = AX_SYS_Init();
+            if (0 != ret) {
+                fprintf(stderr, "AX_SYS_Init failed! ret = 0x%x\n", ret);
+            }
+            AX_ENGINE_NPU_ATTR_T npu_attr;
+            memset(&npu_attr, 0, sizeof(npu_attr));
+            ret = AX_ENGINE_Init(&npu_attr);
+            if (0 != ret) {
+                fprintf(stderr, "Init ax-engine failed{0x%8x}.\n", ret);
+            }
+        }
+        ax_init_flage_++;
+    }
+
+    void _ax_deinit()
+    {
+        if (ax_init_flage_ > 0) {
+            --ax_init_flage_;
+            if (!ax_init_flage_) {
+                AX_ENGINE_Deinit();
+                AX_SYS_Deinit();
+            }
+        }
     }
 
     void start()
@@ -555,9 +598,13 @@ class llm_task {
     {
         stop();
         buffer_destroy(pcmdata);
+        if (axera_session_) axera_session_->Release();
+        _ax_deinit();
     }
 };
 
+int llm_task::ax_init_flage_ = 0;
+
 class llm_kws : public StackFlow {
 private:
     enum { EVENT_TRIGGER = EVENT_EXPORT + 1 };
diff --git a/projects/llm_framework/main_kws/src/runner/EngineWrapper.cpp b/projects/llm_framework/main_kws/src/runner/EngineWrapper.cpp
new file mode 100644
index 00000000..2ba9ed6d
--- /dev/null
+++ b/projects/llm_framework/main_kws/src/runner/EngineWrapper.cpp
@@ -0,0 +1,357 @@
+/**************************************************************************************************
+ *
+ * Copyright (c) 2019-2023 Axera Semiconductor (Ningbo) Co., Ltd. All Rights Reserved.
+ *
+ * This source file is the property of Axera Semiconductor (Ningbo) Co., Ltd. and
+ * may not be copied or distributed in any isomorphic form without the prior
+ * written consent of Axera Semiconductor (Ningbo) Co., Ltd.
+ *
+ **************************************************************************************************/
+#include "EngineWrapper.hpp"
+#include "utils/io.hpp"
+
+#include <cstdlib>
+#include <global_config.h>
+
+#if defined(CONFIG_AX_650C_MSP_ENABLED)
+static const char *strAlgoModelType[AX_ENGINE_MODEL_TYPE_BUTT] = {"3.6T", "7.2T", "18T"};
+#endif
+
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
+static const char *strAlgoModelType[AX_ENGINE_MODEL_TYPE_BUTT] = {"HalfOCM", "FullOCM"};
+#endif
+
+/// @brief npu type
+typedef enum axNPU_TYPE_E {
+    AX_NPU_DEFAULT = 0,        /* running under default NPU according to system */
+    AX_STD_VNPU_1  = (1 << 0), /* running under STD VNPU1 */
+    AX_STD_VNPU_2  = (1 << 1), /* running under STD VNPU2 */
+    AX_STD_VNPU_3  = (1 << 2), /* running under STD VNPU3 */
+    AX_BL_VNPU_1   = (1 << 3), /* running under BIG-LITTLE VNPU1 */
+    AX_BL_VNPU_2   = (1 << 4)  /* running under BIG-LITTLE VNPU2 */
+} AX_NPU_TYPE_E;
+
+#if defined(CONFIG_AX_650C_MSP_ENABLED)
+static AX_S32 CheckModelVNpu(const std::string &strModel, const AX_ENGINE_MODEL_TYPE_T &eModelType,
+                             const AX_S32 &nNpuType, AX_U32 &nNpuSet)
+{
+    AX_ENGINE_NPU_ATTR_T stNpuAttr;
+    memset(&stNpuAttr, 0x00, sizeof(stNpuAttr));
+
+    auto ret = AX_ENGINE_GetVNPUAttr(&stNpuAttr);
+    if (ret == 0) {
+        // VNPU DISABLE
+        if (stNpuAttr.eHardMode == AX_ENGINE_VIRTUAL_NPU_DISABLE) {
+            nNpuSet = 0x01;  // NON-VNPU (0b111)
+            // printf("%s will run under VNPU-DISABLE [%s]\n", strModel.c_str(), strAlgoModelType[eModelType]);
+        }
+        // STD VNPU
+        else if (stNpuAttr.eHardMode == AX_ENGINE_VIRTUAL_NPU_STD) {
+            // 7.2T & 10.8T no allow
+            if (eModelType == AX_ENGINE_MODEL_TYPE1 || eModelType == AX_ENGINE_MODEL_TYPE2) {
+                // printf("%s model type%d: [%s], no allow run under STD VNPU\n", strModel.c_str(), eModelType,
+                // strAlgoModelType[eModelType]);
+                return -1;
+            }
+
+            // default STD VNPU2
+            if (nNpuType == 0) {
+                nNpuSet = 0x02;  // VNPU2 (0b010)
+                // printf("%s will run under default STD-VNPU2 [%s]\n", strModel.c_str(), strAlgoModelType[eModelType]);
+            } else {
+                if (nNpuType & AX_STD_VNPU_1) {
+                    nNpuSet |= 0x01;  // VNPU1 (0b001)
+                    // printf("%s will run under STD-VNPU1 [%s]\n", strModel.c_str(), strAlgoModelType[eModelType]);
+                }
+                if (nNpuType & AX_STD_VNPU_2) {
+                    nNpuSet |= 0x02;  // VNPU2 (0b010)
+                    // printf("%s will run under STD-VNPU2 [%s]\n", strModel.c_str(), strAlgoModelType[eModelType]);
+                }
+                if (nNpuType & AX_STD_VNPU_3) {
+                    nNpuSet |= 0x04;  // VNPU3 (0b100)
+                    // printf("%s will run under STD-VNPU3 [%s]\n", strModel.c_str(), strAlgoModelType[eModelType]);
+                }
+            }
+        }
+        // BL VNPU
+        else if (stNpuAttr.eHardMode == AX_ENGINE_VIRTUAL_NPU_BIG_LITTLE) {
+            // 10.8T no allow
+            if (eModelType == AX_ENGINE_MODEL_TYPE2) {
+                // printf("%s model type%d: [%s], no allow run under BL VNPU\n", strModel.c_str(), eModelType,
+                // strAlgoModelType[eModelType]);
+                return -1;
+            }
+
+            // default BL VNPU
+            if (nNpuType == 0) {
+                // 7.2T default BL VNPU1
+                if (eModelType == AX_ENGINE_MODEL_TYPE1) {
+                    nNpuSet = 0x01;  // VNPU1 (0b001)
+                    // printf("%s will run under default BL-VNPU1 [%s]\n", strModel.c_str(),
+                    // strAlgoModelType[eModelType]);
+                }
+                // 3.6T default BL VNPU2
+                else {
+                    nNpuSet = 0x02;  // VNPU2 (0b010)
+                    // printf("%s will run under default BL-VNPU2 [%s]\n", strModel.c_str(),
+                    // strAlgoModelType[eModelType]);
+                }
+            } else {
+                // 7.2T
+                if (eModelType == AX_ENGINE_MODEL_TYPE1) {
+                    // no allow set to BL VNPU2
+                    if (nNpuType & AX_BL_VNPU_2) {
+                        // printf("%s model type%d: [%s], no allow run under BL VNPU2\n", strModel.c_str(), eModelType,
+                        // strAlgoModelType[eModelType]);
+                        return -1;
+                    }
+                    if (nNpuType & AX_BL_VNPU_1) {
+                        nNpuSet |= 0x01;  // VNPU1 (0b001)
+                        // printf("%s will run under BL-VNPU1 [%s]\n", strModel.c_str(), strAlgoModelType[eModelType]);
+                    }
+                }
+                // 3.6T
+                else {
+                    if (nNpuType & AX_BL_VNPU_1) {
+                        nNpuSet |= 0x01;  // VNPU1 (0b001)
+                        // printf("%s will run under BL-VNPU1 [%s]\n", strModel.c_str(), strAlgoModelType[eModelType]);
+                    }
+                    if (nNpuType & AX_BL_VNPU_2) {
+                        nNpuSet |= 0x02;  // VNPU2 (0b010)
+                        // printf("%s will run under BL-VNPU2 [%s]\n", strModel.c_str(), strAlgoModelType[eModelType]);
+                    }
+                }
+            }
+        }
+    } else {
+        printf("AX_ENGINE_GetVNPUAttr fail ret = %x\n", ret);
+    }
+
+    return ret;
+}
+#endif
+
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
+static AX_S32 CheckModelVNpu(const std::string &strModel, const AX_ENGINE_MODEL_TYPE_T &eModelType,
+                             const AX_S32 &nNpuType, AX_U32 &nNpuSet)
+{
+    AX_ENGINE_NPU_ATTR_T stNpuAttr;
+    memset(&stNpuAttr, 0x00, sizeof(stNpuAttr));
+
+    auto ret = AX_ENGINE_GetVNPUAttr(&stNpuAttr);
+    if (ret == 0) {
+        // VNPU DISABLE
+        if (stNpuAttr.eHardMode == AX_ENGINE_VIRTUAL_NPU_DISABLE) {
+            nNpuSet = 0x01;  // NON-VNPU (0b111)
+            // ALOGN("%s will run under VNPU-DISABLE [%s]", strModel.c_str(), strAlgoModelType[eModelType]);
+        }
+        // STD VNPU
+        else if (stNpuAttr.eHardMode == AX_ENGINE_VIRTUAL_NPU_ENABLE) {
+            // full ocm model was no allowned
+            if (eModelType == AX_ENGINE_MODEL_TYPE1) {
+                // printf("%s model type%d: [%s], no allow run under STD VNPU", strModel.c_str(), eModelType,
+                // strAlgoModelType[eModelType]);
+                return -1;
+            }
+
+            // default STD VNPU2
+            if (nNpuType == 0) {
+                nNpuSet = 0x02;  // VNPU2 (0b010)
+                // printf("%s will run under default STD-VNPU2 [%s]", strModel.c_str(), strAlgoModelType[eModelType]);
+            } else {
+                if (nNpuType & AX_STD_VNPU_1) {
+                    nNpuSet |= 0x01;  // VNPU1 (0b001)
+                    // printf("%s will run under STD-VNPU1 [%s]", strModel.c_str(), strAlgoModelType[eModelType]);
+                }
+                if (nNpuType & AX_STD_VNPU_2) {
+                    nNpuSet |= 0x02;  // VNPU2 (0b010)
+                    // printf("%s will run under STD-VNPU2 [%s]", strModel.c_str(), strAlgoModelType[eModelType]);
+                }
+            }
+        }
+    } else {
+        printf("AX_ENGINE_GetVNPUAttr fail ret = %x", ret);
+    }
+
+    return ret;
+}
+#endif
+
+int EngineWrapper::Init(const char *strModelPath, uint32_t nNpuType)
+{
+    AX_S32 ret = 0;
+
+    // 1. load model
+    AX_BOOL bLoadModelUseCmm     = AX_TRUE;
+    AX_CHAR *pModelBufferVirAddr = nullptr;
+    AX_U64 u64ModelBufferPhyAddr = 0;
+    AX_U32 nModelBufferSize      = 0;
+
+    std::vector<char> model_buffer;
+
+    if (bLoadModelUseCmm) {
+        if (!utils::read_file(strModelPath, (AX_VOID **)&pModelBufferVirAddr, u64ModelBufferPhyAddr,
+                              nModelBufferSize)) {
+            printf("ALGO read model(%s) fail\n", strModelPath);
+            return -1;
+        }
+    } else {
+        if (!utils::read_file(strModelPath, model_buffer)) {
+            printf("ALGO read model(%s) fail\n", strModelPath);
+            return -1;
+        }
+
+        pModelBufferVirAddr = model_buffer.data();
+        nModelBufferSize    = model_buffer.size();
+    }
+
+    auto freeModelBuffer = [&]() {
+        if (bLoadModelUseCmm) {
+            if (u64ModelBufferPhyAddr != 0) {
+                AX_SYS_MemFree(u64ModelBufferPhyAddr, &pModelBufferVirAddr);
+            }
+        } else {
+            std::vector<char>().swap(model_buffer);
+        }
+        return;
+    };
+
+    // 1.1 Get Model Type
+    AX_ENGINE_MODEL_TYPE_T eModelType = AX_ENGINE_MODEL_TYPE0;
+    ret                               = AX_ENGINE_GetModelType(pModelBufferVirAddr, nModelBufferSize, &eModelType);
+    if (0 != ret || eModelType >= AX_ENGINE_MODEL_TYPE_BUTT) {
+        printf("%s AX_ENGINE_GetModelType fail ret=%x, eModelType=%d\n", strModelPath, eModelType);
+        freeModelBuffer();
+        return -1;
+    }
+
+    // 1.2 Check VNPU
+    AX_ENGINE_NPU_SET_T nNpuSet = 0;
+    ret                         = CheckModelVNpu(strModelPath, eModelType, nNpuType, nNpuSet);
+    if (0 != ret) {
+        printf("ALGO CheckModelVNpu fail\n");
+        freeModelBuffer();
+        return -1;
+    }
+
+    // 2. create handle
+    AX_ENGINE_HANDLE handle = nullptr;
+    ret                     = AX_ENGINE_CreateHandle(&handle, pModelBufferVirAddr, nModelBufferSize);
+    auto deinit_handle      = [&handle]() {
+        if (handle) {
+            AX_ENGINE_DestroyHandle(handle);
+        }
+        return -1;
+    };
+
+    freeModelBuffer();
+
+    if (0 != ret || !handle) {
+        printf("ALGO Create model(%s) handle fail\n", strModelPath);
+
+        return deinit_handle();
+    }
+
+    // 3. create context
+    ret = AX_ENGINE_CreateContext(handle);
+    if (0 != ret) {
+        return deinit_handle();
+    }
+
+    // 4. set io
+    m_io_info = nullptr;
+    ret       = AX_ENGINE_GetIOInfo(handle, &m_io_info);
+    if (0 != ret) {
+        return deinit_handle();
+    }
+    m_input_num  = m_io_info->nInputSize;
+    m_output_num = m_io_info->nOutputSize;
+
+    // 4.1 query io
+    //    AX_IMG_FORMAT_E eDtype;
+    //    ret = utils::query_model_input_size(m_io_info, m_input_size, eDtype);//FIXME.
+    //    if (0 != ret) {
+    //        printf("model(%s) query model input size fail\n", strModelPath.c_str());
+    //        return deinit_handle();
+    //    }
+
+    //    if (!(eDtype == AX_FORMAT_YUV420_SEMIPLANAR ||  eDtype == AX_FORMAT_YUV420_SEMIPLANAR_VU ||
+    //        eDtype == AX_FORMAT_RGB888 || eDtype == AX_FORMAT_BGR888)) {
+    //        printf("model(%s) data type is: 0x%02X, unsupport\n", strModelPath, eDtype);
+    //        return deinit_handle();
+    //    }
+
+    // 4.2 brief io
+#ifdef __DEBUG__
+    printf("brief_io_info\n");
+    utils::brief_io_info(strModelPath, m_io_info);
+#endif
+
+    // 5. Config VNPU
+    //  printf("model(%s) nNpuSet: 0x%08X\n", strModelPath.c_str(), nNpuSet);
+    //  will do nothing for using create handle v2 api
+
+    // 6. prepare io
+    // AX_U32 nIoDepth = (stCtx.vecOutputBufferFlag.size() == 0) ? 1 : stCtx.vecOutputBufferFlag.size();
+    ret = utils::prepare_io(strModelPath, m_io_info, m_io, utils::IO_BUFFER_STRATEGY_DEFAULT);
+    if (0 != ret) {
+        printf("prepare io failed!\n");
+        utils::free_io(m_io);
+        return deinit_handle();
+    }
+
+    m_handle  = handle;
+    m_hasInit = true;
+
+    return 0;
+}
+
+int EngineWrapper::SetInput(void *pInput, int index)
+{
+    return utils::push_io_input(pInput, index, m_io);
+}
+
+int EngineWrapper::Run()
+{
+    if (!m_hasInit) return -1;
+
+    // 7.3 run & benchmark
+    auto ret = AX_ENGINE_RunSync(m_handle, &m_io);
+    if (0 != ret) {
+        printf("AX_ENGINE_RunSync failed. ret=0x%x\n", ret);
+        return ret;
+    }
+
+    return 0;
+}
+
+int EngineWrapper::GetOutput(void *pOutput, int index)
+{
+    return utils::push_io_output(pOutput, index, m_io);
+}
+
+int EngineWrapper::GetInputSize(int index)
+{
+    return m_io.pInputs[index].nSize;
+}
+
+int EngineWrapper::GetOutputSize(int index)
+{
+    return m_io.pOutputs[index].nSize;
+}
+
+void *EngineWrapper::GetOutputPtr(int index)
+{
+    utils::cache_io_flush(&m_io.pOutputs[index]);
+    return m_io.pOutputs[index].pVirAddr;
+}
+
+int EngineWrapper::Release()
+{
+    if (m_handle) {
+        utils::free_io(m_io);
+        AX_ENGINE_DestroyHandle(m_handle);
+        m_handle = nullptr;
+    }
+    return 0;
+}
diff --git a/projects/llm_framework/main_kws/src/runner/EngineWrapper.hpp b/projects/llm_framework/main_kws/src/runner/EngineWrapper.hpp
new file mode 100644
index 00000000..eecc80ea
--- /dev/null
+++ b/projects/llm_framework/main_kws/src/runner/EngineWrapper.hpp
@@ -0,0 +1,53 @@
+/**************************************************************************************************
+ *
+ * Copyright (c) 2019-2023 Axera Semiconductor (Ningbo) Co., Ltd. All Rights Reserved.
+ *
+ * This source file is the property of Axera Semiconductor (Ningbo) Co., Ltd. and
+ * may not be copied or distributed in any isomorphic form without the prior
+ * written consent of Axera Semiconductor (Ningbo) Co., Ltd.
+ *
+ **************************************************************************************************/
+
+#pragma once
+
+#include <string>
+#include <vector>
+#include <cstring>
+#include <array>
+#include <cstdint>
+
+#include "ax_engine_api.h"
+
+class EngineWrapper {
+public:
+    EngineWrapper() : m_hasInit(false), m_handle(nullptr)
+    {
+    }
+
+    ~EngineWrapper()
+    {
+        Release();
+    }
+
+    int Init(const char* strModelPath, uint32_t nNpuType = 0);
+
+    int SetInput(void* pInput, int index);
+
+    int Run();
+
+    int GetOutput(void* pOutput, int index);
+
+    int GetInputSize(int index);
+    int GetOutputSize(int index);
+
+    void* GetOutputPtr(int index);
+
+    int Release();
+
+protected:
+    bool m_hasInit;
+    AX_ENGINE_HANDLE m_handle;
+    AX_ENGINE_IO_INFO_T* m_io_info{};
+    AX_ENGINE_IO_T m_io{};
+    int m_input_num{}, m_output_num{};
+};
\ No newline at end of file
diff --git a/projects/llm_framework/main_kws/src/runner/utils/checker.h b/projects/llm_framework/main_kws/src/runner/utils/checker.h
new file mode 100644
index 00000000..e71a2273
--- /dev/null
+++ b/projects/llm_framework/main_kws/src/runner/utils/checker.h
@@ -0,0 +1,32 @@
+/**************************************************************************************************
+ *
+ * Copyright (c) 2019-2023 Axera Semiconductor (Ningbo) Co., Ltd. All Rights Reserved.
+ *
+ * This source file is the property of Axera Semiconductor (Ningbo) Co., Ltd. and
+ * may not be copied or distributed in any isomorphic form without the prior
+ * written consent of Axera Semiconductor (Ningbo) Co., Ltd.
+ *
+ **************************************************************************************************/
+
+#ifndef AX_CHECKER_H
+#define AX_CHECKER_H
+
+#include "utils/logger.h"
+
+#define CHECK_PTR(p)                       \
+    do {                                   \
+        if (!p) {                          \
+            ALOGE("%s nil pointer\n", #p); \
+            return -1;                     \
+        }                                  \
+    } while (0)
+
+#define CHECK_INITED(p)                     \
+    do {                                    \
+        if (!p->HasInit()) {                \
+            ALOGE("%s has not init\n", #p); \
+            return -1;                      \
+        }                                   \
+    } while (0)
+
+#endif  // AX_CHECKER_H
diff --git a/projects/llm_framework/main_kws/src/runner/utils/io.hpp b/projects/llm_framework/main_kws/src/runner/utils/io.hpp
new file mode 100644
index 00000000..4c79f8fe
--- /dev/null
+++ b/projects/llm_framework/main_kws/src/runner/utils/io.hpp
@@ -0,0 +1,606 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * License); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * AS IS BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+#pragma once
+
+#include <cstdio>
+#include <cstring>
+#include <vector>
+#include <array>
+#include <string>
+#include <fstream>
+#include <cstdint>
+#include <cstdlib>
+
+#include "utils/checker.h"
+#include "ax_sys_api.h"
+#include "ax_engine_type.h"
+
+#define IO_CMM_ALIGN_SIZE 128
+
+namespace utils {
+typedef enum { IO_BUFFER_STRATEGY_DEFAULT, IO_BUFFER_STRATEGY_CACHED } IO_BUFFER_STRATEGY_T;
+
+static inline AX_S32 query_model_input_size(const AX_ENGINE_IO_INFO_T* io_info, std::array<int, 2>& input_size,
+                                            AX_IMG_FORMAT_E& eDtype)
+{
+    int height         = 0;
+    int width          = 0;
+    int size           = 0;
+    int channel        = 0;
+    int data_type_size = 0;
+    auto& input        = io_info->pInputs[0];
+
+    switch (input.eLayout) {
+        case AX_ENGINE_TENSOR_LAYOUT_NHWC:
+            height  = input.pShape[1];
+            width   = input.pShape[2];
+            channel = input.pShape[3];
+            size    = input.nSize;
+            break;
+        case AX_ENGINE_TENSOR_LAYOUT_NCHW:
+            channel = input.pShape[1];
+            height  = input.pShape[2];
+            width   = input.pShape[3];
+            size    = input.nSize;
+            break;
+        default:  // NHWC
+            height  = input.pShape[1];
+            width   = input.pShape[2];
+            channel = input.pShape[3];
+            size    = input.nSize;
+            break;
+    }
+
+    switch (input.eDataType) {
+        case AX_ENGINE_DT_UINT8:
+        case AX_ENGINE_DT_SINT8:
+            data_type_size = 1;
+            break;
+        case AX_ENGINE_DT_UINT16:
+        case AX_ENGINE_DT_SINT16:
+            data_type_size = 2;
+            break;
+        case AX_ENGINE_DT_FLOAT32:
+            data_type_size = 4;
+            break;
+        case AX_ENGINE_DT_SINT32:
+        case AX_ENGINE_DT_UINT32:
+            data_type_size = 4;
+            break;
+        case AX_ENGINE_DT_FLOAT64:
+            data_type_size = 8;
+            break;
+        default:
+            data_type_size = 1;
+            break;
+    }
+
+    if (channel == 0 || height == 0 || width == 0 || size == 0 || data_type_size == 0) {
+        return -1;
+    }
+
+    if (input.pExtraMeta) {
+        switch (input.pExtraMeta->eColorSpace) {
+            case AX_ENGINE_CS_BGR:
+                input_size[0] = height;
+                input_size[1] = width;
+                eDtype        = AX_FORMAT_BGR888;
+                break;
+            case AX_ENGINE_CS_RGB:
+                input_size[0] = height;
+                input_size[1] = width;
+                eDtype        = AX_FORMAT_RGB888;
+                break;
+            case AX_ENGINE_CS_NV12:
+                input_size[0] = height * 2 / 3;
+                input_size[1] = width;
+                eDtype        = AX_FORMAT_YUV420_SEMIPLANAR;
+                break;
+            case AX_ENGINE_CS_NV21:
+                input_size[0] = height * 2 / 3;
+                input_size[1] = width;
+                eDtype        = AX_FORMAT_YUV420_SEMIPLANAR_VU;
+                break;
+            default:  // AX_ENGINE_CS_NV12
+                input_size[0] = height * 2 / 3;
+                input_size[1] = width;
+                eDtype        = AX_FORMAT_YUV420_SEMIPLANAR;
+                break;
+        }
+    } else {
+        input_size[0] = height * 2 / 3;
+        input_size[1] = width;
+        eDtype        = AX_FORMAT_YUV420_SEMIPLANAR;
+    }
+
+    ALOGD("eLayout:%d, eDataType:%d, channel:%d, height:%d, width:%d, size:%d, data_type_size:%d", input.eLayout,
+          input.eDataType, channel, input_size[0], input_size[1], size, data_type_size);
+
+    return 0;
+}
+
+static inline void brief_io_info(std::string strModel, const AX_ENGINE_IO_INFO_T* io_info)
+{
+    auto describe_shape_type = [](AX_ENGINE_TENSOR_LAYOUT_T type) -> const char* {
+        switch (type) {
+            case AX_ENGINE_TENSOR_LAYOUT_NHWC:
+                return "NHWC";
+            case AX_ENGINE_TENSOR_LAYOUT_NCHW:
+                return "NCHW";
+            default:
+                return "unknown";
+        }
+    };
+    auto describe_data_type = [](AX_ENGINE_DATA_TYPE_T type) -> const char* {
+        switch (type) {
+            case AX_ENGINE_DT_UINT8:
+                return "uint8";
+            case AX_ENGINE_DT_UINT16:
+                return "uint16";
+            case AX_ENGINE_DT_FLOAT32:
+                return "float32";
+            case AX_ENGINE_DT_SINT16:
+                return "sint16";
+            case AX_ENGINE_DT_SINT8:
+                return "sint8";
+            case AX_ENGINE_DT_SINT32:
+                return "sint32";
+            case AX_ENGINE_DT_UINT32:
+                return "uint32";
+            case AX_ENGINE_DT_FLOAT64:
+                return "float64";
+            case AX_ENGINE_DT_UINT10_PACKED:
+                return "uint10_packed";
+            case AX_ENGINE_DT_UINT12_PACKED:
+                return "uint12_packed";
+            case AX_ENGINE_DT_UINT14_PACKED:
+                return "uint14_packed";
+            case AX_ENGINE_DT_UINT16_PACKED:
+                return "uint16_packed";
+            default:
+                return "unknown";
+        }
+    };
+    auto describe_memory_type = [](AX_ENGINE_MEMORY_TYPE_T type) -> const char* {
+        switch (type) {
+            case AX_ENGINE_MT_PHYSICAL:
+                return "Physical";
+            case AX_ENGINE_MT_VIRTUAL:
+                return "Virtual";
+            default:
+                return "unknown";
+        }
+    };
+    auto describe_color_space = [](AX_ENGINE_COLOR_SPACE_T cs) -> const char* {
+        switch (cs) {
+            case AX_ENGINE_CS_FEATUREMAP:
+                return "FeatureMap";
+            case AX_ENGINE_CS_BGR:
+                return "BGR";
+            case AX_ENGINE_CS_RGB:
+                return "RGB";
+            case AX_ENGINE_CS_RGBA:
+                return "RGBA";
+            case AX_ENGINE_CS_GRAY:
+                return "GRAY";
+            case AX_ENGINE_CS_NV12:
+                return "NV12";
+            case AX_ENGINE_CS_NV21:
+                return "NV21";
+            case AX_ENGINE_CS_YUV444:
+                return "YUV444";
+            case AX_ENGINE_CS_RAW8:
+                return "RAW8";
+            case AX_ENGINE_CS_RAW10:
+                return "RAW10";
+            case AX_ENGINE_CS_RAW12:
+                return "RAW12";
+            case AX_ENGINE_CS_RAW14:
+                return "RAW14";
+            case AX_ENGINE_CS_RAW16:
+                return "RAW16";
+            default:
+                return "unknown";
+        }
+    };
+    printf("Model Name: %s\n", strModel.c_str());
+    printf("Max Batch Size %d\n", io_info->nMaxBatchSize);
+    printf("Support Dynamic Batch? %s\n", io_info->bDynamicBatchSize == AX_TRUE ? "Yes" : "No");
+
+    for (uint32_t i = 0; i < io_info->nInputSize; ++i) {
+        auto& input = io_info->pInputs[i];
+        printf("Input[%d]: %s\n", i, input.pName);
+        printf("    Shape [");
+        for (uint32_t j = 0; j < input.nShapeSize; ++j) {
+            printf("%d", (int)input.pShape[j]);
+            if (j + 1 < input.nShapeSize) printf(", ");
+        }
+        printf("] %s %s %s %s\n", describe_shape_type(input.eLayout), describe_data_type(input.eDataType),
+               input.pExtraMeta ? describe_color_space(input.pExtraMeta->eColorSpace) : "",
+               input.nQuantizationValue > 0 ? ("Q=" + std::to_string(input.nQuantizationValue)).c_str() : "");
+        printf("    Memory %s\n", describe_memory_type(input.eMemoryType));
+        printf("    Size %u\n", input.nSize);
+    }
+    for (uint32_t i = 0; i < io_info->nOutputSize; ++i) {
+        auto& output = io_info->pOutputs[i];
+        printf("Output[%d]: %s\n", i, output.pName);
+        printf("    Shape [");
+        for (uint32_t j = 0; j < output.nShapeSize; ++j) {
+            printf("%d", (int)output.pShape[j]);
+            if (j + 1 < output.nShapeSize) printf(", ");
+        }
+        printf("] %s %s %s\n", describe_shape_type(output.eLayout), describe_data_type(output.eDataType),
+               output.nQuantizationValue > 0 ? ("Q=" + std::to_string(output.nQuantizationValue)).c_str() : "");
+        printf("    Memory %s\n", describe_memory_type(output.eMemoryType));
+        printf("    Size %u\n", output.nSize);
+    }
+}
+
+static inline AX_S32 alloc_engine_buffer(const std::string& token, const std::string& appendix, size_t index,
+                                         const AX_ENGINE_IOMETA_T* pMeta, AX_ENGINE_IO_BUFFER_T* pBuf,
+                                         IO_BUFFER_STRATEGY_T eStrategy = IO_BUFFER_STRATEGY_DEFAULT)
+{
+    AX_S32 ret = -1;
+    if (eStrategy != IO_BUFFER_STRATEGY_DEFAULT && eStrategy != IO_BUFFER_STRATEGY_CACHED) {
+        fprintf(stderr, "strategy %d not supported\n", (int)eStrategy);
+        return -1;
+    }
+    memset(pBuf, 0, sizeof(AX_ENGINE_IO_BUFFER_T));
+    pBuf->nSize = pMeta->nSize;
+
+    const std::string token_name = "skel_" + token + appendix + std::to_string(index);
+
+    if (eStrategy == IO_BUFFER_STRATEGY_CACHED) {
+        ret = AX_SYS_MemAllocCached((AX_U64*)&pBuf->phyAddr, &pBuf->pVirAddr, pBuf->nSize, IO_CMM_ALIGN_SIZE,
+                                    (const AX_S8*)token_name.c_str());
+    } else {
+        ret = AX_SYS_MemAlloc((AX_U64*)&pBuf->phyAddr, &pBuf->pVirAddr, pBuf->nSize, IO_CMM_ALIGN_SIZE,
+                              (const AX_S8*)token_name.c_str());
+    }
+
+    return ret;
+}
+
+static inline AX_S32 free_engine_buffer(AX_ENGINE_IO_BUFFER_T* pBuf)
+{
+    if (pBuf->phyAddr == 0) {
+        delete[] reinterpret_cast<uint8_t*>(pBuf->pVirAddr);
+    } else {
+        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+    }
+    pBuf->phyAddr  = 0;
+    pBuf->pVirAddr = nullptr;
+
+    return 0;
+}
+
+static inline void free_io_index(AX_ENGINE_IO_BUFFER_T* io_buf, size_t index)
+{
+    AX_ENGINE_IO_BUFFER_T* pBuf = io_buf + index;
+    free_engine_buffer(pBuf);
+}
+
+static inline void free_io(AX_ENGINE_IO_T& io)
+{
+    for (size_t j = 0; j < io.nInputSize; ++j) {
+        AX_ENGINE_IO_BUFFER_T* pBuf = io.pInputs + j;
+        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+    }
+    for (size_t j = 0; j < io.nOutputSize; ++j) {
+        AX_ENGINE_IO_BUFFER_T* pBuf = io.pOutputs + j;
+        AX_SYS_MemFree(pBuf->phyAddr, pBuf->pVirAddr);
+    }
+    delete[] io.pInputs;
+    delete[] io.pOutputs;
+}
+
+static inline void free_io(AX_ENGINE_IO_T& io, std::vector<std::vector<AX_ENGINE_IO_BUFFER_T>>& vecOutputBuffer)
+{
+    if (io.pInputs) {
+        delete[] io.pInputs;
+        io.pInputs = nullptr;
+    }
+
+    if (io.pOutputs) {
+        for (size_t index = 0; index < vecOutputBuffer.size(); ++index) {
+            AX_ENGINE_IO_BUFFER_T* pOutputs = &vecOutputBuffer[index][0];
+            for (size_t j = 0; j < io.nOutputSize; ++j) {
+                free_io_index(pOutputs, j);
+            }
+        }
+
+        delete[] io.pOutputs;
+        io.pOutputs = nullptr;
+    }
+}
+
+static inline int prepare_io(const std::string& token, const AX_ENGINE_IO_INFO_T* info, AX_ENGINE_IO_T& io,
+                             IO_BUFFER_STRATEGY_T strategy)
+{
+    auto ret = 0;
+
+    memset(&io, 0, sizeof(io));
+
+    io.pInputs = new AX_ENGINE_IO_BUFFER_T[info->nInputSize];
+
+    if (!io.pInputs) {
+        goto EXIT;
+    }
+
+    memset(io.pInputs, 0x00, sizeof(AX_ENGINE_IO_BUFFER_T) * info->nInputSize);
+    io.nInputSize = info->nInputSize;
+    for (AX_U32 i = 0; i < info->nInputSize; ++i) {
+        auto meta   = info->pInputs[i];
+        auto buffer = &io.pInputs[i];
+        ret         = alloc_engine_buffer(token, "_input_", i, &meta, buffer, strategy);
+        if (ret != 0) {
+            free_io_index(io.pInputs, i);
+            return ret;
+        }
+    }
+
+    io.pOutputs = new AX_ENGINE_IO_BUFFER_T[info->nOutputSize];
+
+    if (!io.pOutputs) {
+        goto EXIT;
+    }
+
+    memset(io.pOutputs, 0x00, sizeof(AX_ENGINE_IO_BUFFER_T) * info->nOutputSize);
+    io.nOutputSize = info->nOutputSize;
+
+    for (size_t i = 0; i < info->nOutputSize; ++i) {
+        auto meta   = info->pOutputs[i];
+        auto buffer = &io.pOutputs[i];
+        ret         = alloc_engine_buffer(token, "_output_", i, &meta, buffer, strategy);
+        if (ret != 0) {
+            goto EXIT;
+        }
+    }
+
+EXIT:
+    if (ret != 0) {
+        free_io(io);
+        return -1;
+    }
+
+    return 0;
+}
+
+static inline int prepare_io(const std::string& token, const AX_ENGINE_IO_INFO_T* info, AX_ENGINE_IO_T& io,
+                             std::vector<AX_ENGINE_IO_BUFFER_T>& vecOutputBuffer, const IO_BUFFER_STRATEGY_T& strategy)
+{
+    AX_S32 ret = 0;
+    memset(&io, 0, sizeof(io));
+
+    std::vector<AX_ENGINE_IO_BUFFER_T> outputBuffer;
+
+    if (1 != info->nInputSize) {
+        fprintf(stderr, "[ERR]: Only single input was accepted(got %u).\n", info->nInputSize);
+        return -1;
+    }
+
+    io.pInputs = new AX_ENGINE_IO_BUFFER_T[info->nInputSize];
+
+    if (!io.pInputs) {
+        goto EXIT;
+    }
+
+    memset(io.pInputs, 0x00, sizeof(AX_ENGINE_IO_BUFFER_T) * info->nInputSize);
+    io.nInputSize = info->nInputSize;
+
+    for (AX_U32 i = 0; i < info->nInputSize; ++i) {
+        auto meta   = info->pInputs[i];
+        auto buffer = &io.pInputs[i];
+        ret         = alloc_engine_buffer(token, "_input_", i, &meta, buffer, strategy);
+        if (ret != 0) {
+            free_io_index(io.pInputs, i);
+            return ret;
+        }
+    }
+
+    io.pOutputs = new AX_ENGINE_IO_BUFFER_T[info->nOutputSize];
+
+    if (!io.pOutputs) {
+        goto EXIT;
+    }
+
+    for (size_t i = 0; i < info->nOutputSize; ++i) {
+        auto meta   = info->pOutputs[i];
+        auto buffer = &io.pOutputs[i];
+        ret         = alloc_engine_buffer(token, "_output_", i, &meta, buffer, strategy);
+
+        if (ret != 0) {
+            goto EXIT;
+        }
+
+        vecOutputBuffer.push_back(*buffer);
+    }
+
+    memset(io.pOutputs, 0x00, sizeof(AX_ENGINE_IO_BUFFER_T) * info->nOutputSize);
+    io.nOutputSize = info->nOutputSize;
+
+    for (size_t i = 0; i < info->nOutputSize; ++i) {
+        auto buffer = &io.pOutputs[i];
+        *buffer     = vecOutputBuffer[i];
+    }
+
+EXIT:
+    if (ret != 0) {
+        free_io(io);
+        return -1;
+    }
+
+    return 0;
+}
+
+static inline AX_S32 push_io_output(const AX_ENGINE_IO_INFO_T* info, AX_ENGINE_IO_T& io,
+                                    std::vector<AX_ENGINE_IO_BUFFER_T>& outputBuffer)
+{
+    for (size_t i = 0; i < info->nOutputSize; ++i) {
+        auto buffer = &io.pOutputs[i];
+        *buffer     = outputBuffer[i];
+    }
+
+    return 0;
+}
+
+static inline AX_S32 cache_io_flush(const AX_ENGINE_IO_BUFFER_T* io_buf)
+{
+    if (io_buf->phyAddr != 0) {
+        AX_SYS_MflushCache(io_buf->phyAddr, io_buf->pVirAddr, io_buf->nSize);
+    }
+
+    return 0;
+}
+
+static inline AX_S32 push_io_input(void* input, int index, AX_ENGINE_IO_T& io)
+{
+    // img ranks_depth ranks_feat ranks_bev, n_points
+    AX_ENGINE_IO_BUFFER_T* pImg = &io.pInputs[index];
+
+    memcpy(pImg->pVirAddr, input, pImg->nSize);
+    cache_io_flush(pImg);
+    return 0;
+}
+
+static inline AX_S32 push_io_output(void* output, int index, AX_ENGINE_IO_T& io)
+{
+    AX_ENGINE_IO_BUFFER_T* pImg = &io.pOutputs[index];
+    cache_io_flush(pImg);
+    memcpy(output, pImg->pVirAddr, pImg->nSize);
+    return 0;
+}
+
+static inline AX_S32 cpu_copy(AX_U64 nPhyAddrSrc, AX_U64 nPhyAddrDst, AX_U32 nLen)
+{
+    if (nPhyAddrSrc != 0 && nPhyAddrDst != 0 && nLen > 0) {
+        AX_VOID* pSrcVirAddr = AX_SYS_MmapCache(nPhyAddrSrc, nLen);
+        AX_VOID* pDstVirAddr = AX_SYS_MmapCache(nPhyAddrDst, nLen);
+
+        memcpy((AX_VOID*)pDstVirAddr, (AX_VOID*)pSrcVirAddr, nLen);
+
+        AX_SYS_Munmap(pSrcVirAddr, nLen);
+        AX_SYS_Munmap(pDstVirAddr, nLen);
+
+        return 0;
+    }
+
+    return -1;
+}
+
+static inline AX_S32 inc_io_ref_cnt(const AX_VIDEO_FRAME_T& stFrame)
+{
+    if (stFrame.u32BlkId[0] > 0) {
+        AX_POOL_IncreaseRefCnt(stFrame.u32BlkId[0]);
+    }
+    if (stFrame.u32BlkId[1] > 0) {
+        AX_POOL_IncreaseRefCnt(stFrame.u32BlkId[1]);
+    }
+    if (stFrame.u32BlkId[2] > 0) {
+        AX_POOL_IncreaseRefCnt(stFrame.u32BlkId[2]);
+    }
+
+    return 0;
+}
+
+static inline AX_S32 dec_io_ref_cnt(const AX_VIDEO_FRAME_T& stFrame)
+{
+    if (stFrame.u32BlkId[0] > 0) {
+        AX_POOL_DecreaseRefCnt(stFrame.u32BlkId[0]);
+    }
+    if (stFrame.u32BlkId[1] > 0) {
+        AX_POOL_DecreaseRefCnt(stFrame.u32BlkId[1]);
+    }
+    if (stFrame.u32BlkId[2] > 0) {
+        AX_POOL_DecreaseRefCnt(stFrame.u32BlkId[2]);
+    }
+
+    return 0;
+}
+
+static inline bool read_file(const char* path, std::vector<char>& data)
+{
+    std::fstream fs(path, std::ios::in | std::ios::binary);
+
+    if (!fs.is_open()) {
+        return false;
+    }
+
+    fs.seekg(std::ios::end);
+    auto fs_end = fs.tellg();
+    fs.seekg(std::ios::beg);
+    auto fs_beg = fs.tellg();
+
+    auto file_size   = static_cast<size_t>(fs_end - fs_beg);
+    auto vector_size = data.size();
+
+    data.reserve(vector_size + file_size);
+    data.insert(data.end(), std::istreambuf_iterator<char>(fs), std::istreambuf_iterator<char>());
+
+    fs.close();
+
+    return true;
+}
+
+static inline bool read_file(const char* path, AX_VOID** pModelBufferVirAddr, AX_U64& u64ModelBufferPhyAddr,
+                             AX_U32& nModelBufferSize)
+{
+    std::fstream fs(path, std::ios::in | std::ios::binary);
+
+    if (!fs.is_open()) {
+        return false;
+    }
+
+    fs.seekg(0, std::ios::end);
+    int file_size = fs.tellg();
+    fs.seekg(0, std::ios::beg);
+
+    nModelBufferSize = (AX_U32)file_size;
+
+    AX_SYS_MemAlloc(&u64ModelBufferPhyAddr, pModelBufferVirAddr, nModelBufferSize, 0x100, (AX_S8*)"SKEL-CV");
+
+    if (!pModelBufferVirAddr || (u64ModelBufferPhyAddr == 0)) {
+        return false;
+    }
+
+    fs.read((AX_CHAR*)*pModelBufferVirAddr, nModelBufferSize);
+
+    fs.close();
+
+    return true;
+}
+
+static inline void dequant(float** pptrOutput, const AX_ENGINE_IOMETA_T& ptrIoInfo, const AX_ENGINE_IO_BUFFER_T& ioBuf,
+                           float zp, float scale)
+{
+    if (ptrIoInfo.eDataType == AX_ENGINE_DT_FLOAT32) {
+        *pptrOutput = (float*)ioBuf.pVirAddr;
+        return;
+    }
+
+    *pptrOutput    = (float*)malloc(ptrIoInfo.nSize * sizeof(float));
+    uint8_t* pBuf  = (uint8_t*)ioBuf.pVirAddr;
+    float* pOutput = *pptrOutput;
+    // float inv_scale = 1.0f / scale;
+    for (int i = 0; i < ptrIoInfo.nSize; i++) {
+        pOutput[i] = ((float)pBuf[i] - zp) * scale;
+    }
+}
+}  // namespace utils
diff --git a/projects/llm_framework/main_kws/src/runner/utils/logger.h b/projects/llm_framework/main_kws/src/runner/utils/logger.h
new file mode 100644
index 00000000..81969e22
--- /dev/null
+++ b/projects/llm_framework/main_kws/src/runner/utils/logger.h
@@ -0,0 +1,84 @@
+/**************************************************************************************************
+ *
+ * Copyright (c) 2019-2023 Axera Semiconductor (Ningbo) Co., Ltd. All Rights Reserved.
+ *
+ * This source file is the property of Axera Semiconductor (Ningbo) Co., Ltd. and
+ * may not be copied or distributed in any isomorphic form without the prior
+ * written consent of Axera Semiconductor (Ningbo) Co., Ltd.
+ *
+ **************************************************************************************************/
+
+#ifndef SKEL_LOGGER_H
+#define SKEL_LOGGER_H
+
+#include "ax_global_type.h"
+#include "ax_sys_log.h"
+
+#include <stdio.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// #define SKEL_LOG_TAG "SKEL"
+//
+// #define ALOGE(fmt, ...) AX_LOG_ERR_EX(SKEL_LOG_TAG, AX_ID_SKEL, fmt, ##__VA_ARGS__)
+// #define ALOGW(fmt, ...) AX_LOG_WARN_EX(SKEL_LOG_TAG, AX_ID_SKEL, fmt, ##__VA_ARGS__)
+// #define ALOGI(fmt, ...) AX_LOG_INFO_EX(SKEL_LOG_TAG, AX_ID_SKEL, fmt, ##__VA_ARGS__)
+// #define ALOGD(fmt, ...) AX_LOG_DBG_EX(SKEL_LOG_TAG, AX_ID_SKEL, fmt, ##__VA_ARGS__)
+// #define ALOGN(fmt, ...) AX_LOG_NOTICE_EX(SKEL_LOG_TAG, AX_ID_SKEL, fmt, ##__VA_ARGS__)
+
+typedef enum {
+    SKEL_LOG_MIN       = -1,
+    SKEL_LOG_EMERGENCY = 0,
+    SKEL_LOG_ALERT     = 1,
+    SKEL_LOG_CRITICAL  = 2,
+    SKEL_LOG_ERROR     = 3,
+    SKEL_LOG_WARN      = 4,
+    SKEL_LOG_NOTICE    = 5,
+    SKEL_LOG_INFO      = 6,
+    SKEL_LOG_DEBUG     = 7,
+    SKEL_LOG_MAX
+} SKEL_LOG_LEVEL_E;
+
+static SKEL_LOG_LEVEL_E log_level = SKEL_LOG_DEBUG;
+
+#if 1
+#define MACRO_BLACK  "\033[1;30;30m"
+#define MACRO_RED    "\033[1;30;31m"
+#define MACRO_GREEN  "\033[1;30;32m"
+#define MACRO_YELLOW "\033[1;30;33m"
+#define MACRO_BLUE   "\033[1;30;34m"
+#define MACRO_PURPLE "\033[1;30;35m"
+#define MACRO_WHITE  "\033[1;30;37m"
+#define MACRO_END    "\033[0m"
+#else
+#define MACRO_BLACK
+#define MACRO_RED
+#define MACRO_GREEN
+#define MACRO_YELLOW
+#define MACRO_BLUE
+#define MACRO_PURPLE
+#define MACRO_WHITE
+#define MACRO_END
+#endif
+
+#define ALOGE(fmt, ...) printf(MACRO_RED "[E][%32s][%4d]: " fmt MACRO_END "\n", __FUNCTION__, __LINE__, ##__VA_ARGS__)
+#define ALOGW(fmt, ...)             \
+    if (log_level >= SKEL_LOG_WARN) \
+    printf(MACRO_YELLOW "[W][%32s][%4d]: " fmt MACRO_END "\n", __FUNCTION__, __LINE__, ##__VA_ARGS__)
+#define ALOGI(fmt, ...)             \
+    if (log_level >= SKEL_LOG_INFO) \
+    printf(MACRO_GREEN "[I][%32s][%4d]: " fmt MACRO_END "\n", __FUNCTION__, __LINE__, ##__VA_ARGS__)
+#define ALOGD(fmt, ...)              \
+    if (log_level >= SKEL_LOG_DEBUG) \
+    printf(MACRO_WHITE "[D][%32s][%4d]: " fmt MACRO_END "\n", __FUNCTION__, __LINE__, ##__VA_ARGS__)
+#define ALOGN(fmt, ...)               \
+    if (log_level >= SKEL_LOG_NOTICE) \
+    printf(MACRO_PURPLE "[N][%32s][%4d]: " fmt MACRO_END "\n", __FUNCTION__, __LINE__, ##__VA_ARGS__)
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // SKEL_LOGGER_H

From bea45abf1a15e30703308e8651ffd25626365358 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 18 Dec 2025 19:35:02 +0800
Subject: [PATCH 72/79] [update] llm_asr supported zipformer stream model.

---
 projects/llm_framework/main_asr/SConstruct    |   6 +-
 ...erpa-onnx-zipformer-bilingual-zh-en-t.json |  70 +++++++
 projects/llm_framework/main_asr/src/main.cpp  | 182 +++++++++++++++++-
 projects/llm_framework/main_kws/mode_kws.json |   2 +-
 4 files changed, 256 insertions(+), 4 deletions(-)
 create mode 100644 projects/llm_framework/main_asr/mode_sherpa-onnx-zipformer-bilingual-zh-en-t.json

diff --git a/projects/llm_framework/main_asr/SConstruct b/projects/llm_framework/main_asr/SConstruct
index 9bdac46a..217bbc51 100644
--- a/projects/llm_framework/main_asr/SConstruct
+++ b/projects/llm_framework/main_asr/SConstruct
@@ -21,7 +21,11 @@ DEFINITIONS += ['-std=c++17', '-fopenmp']
 LINK_SEARCH_PATH += [ADir('../static_lib')]
 REQUIREMENTS += ['ax_engine', 'ax_interpreter', 'ax_sys']
 
-INCLUDE += [ADir('../static_lib/include/sherpa'), ADir('../static_lib/include/sherpa/sherpa-ncnn')]
+INCLUDE += [ADir('../static_lib/include/sherpa'), 
+            ADir('../static_lib/include/sherpa/sherpa-ncnn'), 
+            ADir('../static_lib/include/sherpa/sherpa-onnx'),
+            ADir('../static_lib/include/sherpa/sherpa-onnx/openfst-src'),
+            ADir('../static_lib/include/sherpa/sherpa-onnx/onnxruntime-src')]
 LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/ncnn')]
 LINK_SEARCH_PATH += [ADir('../static_lib/sherpa/onnx')]
 REQUIREMENTS += ['ncnn', '', 'onnxruntime']
diff --git a/projects/llm_framework/main_asr/mode_sherpa-onnx-zipformer-bilingual-zh-en-t.json b/projects/llm_framework/main_asr/mode_sherpa-onnx-zipformer-bilingual-zh-en-t.json
new file mode 100644
index 00000000..e22da795
--- /dev/null
+++ b/projects/llm_framework/main_asr/mode_sherpa-onnx-zipformer-bilingual-zh-en-t.json
@@ -0,0 +1,70 @@
+{
+    "mode": "sherpa-onnx-zipformer-bilingual-zh-en-t",
+    "type": "asr",
+    "homepage": "",
+    "capabilities": [
+        "Automatic_Speech_Recognition",
+        "Chinese",
+        "English"
+    ],
+    "input_type": [
+        "sys.pcm",
+        "sys.cap.0_0"
+    ],
+    "output_type": [
+        "asr.utf-8",
+        "asr.bool"
+    ],
+    "mode_param": {
+        "model_config.transducer.encoder": "encoder.axmodel",
+        "model_config.transducer.decoder": "decoder.axmodel",
+        "model_config.transducer.joiner": "joiner.axmodel",
+        "model_config.tokens": "tokens.txt",
+        "feat_config.feature_dim": 80,
+        "feat_config.sampling_rate": 16000,
+        "endpoint_config.rule1.min_trailing_silence": 2.4,
+        "endpoint_config.rule2.min_trailing_silence": 1.2,
+        "endpoint_config.rule3.min_utterance_length": 30,
+        "enable_endpoint": true,
+        "awake_delay": 50,
+        "model_config.provider_config.provider": "axera",
+        "model_config.zipformer_meta.encoder_dims": [
+            256,
+            256,
+            256,
+            256,
+            256
+        ],
+        "model_config.zipformer_meta.attention_dims": [
+            192,
+            192,
+            192,
+            192,
+            192
+        ],
+        "model_config.zipformer_meta.num_encoder_layers": [
+            2,
+            2,
+            2,
+            2,
+            2
+        ],
+        "model_config.zipformer_meta.cnn_module_kernels": [
+            31,
+            31,
+            31,
+            31,
+            31
+        ],
+        "model_config.zipformer_meta.left_context_len": [
+            192,
+            96,
+            48,
+            24,
+            96
+        ],
+        "model_config.zipformer_meta.T": 103,
+        "model_config.zipformer_meta.decode_chunk_len": 96,
+        "model_config.zipformer_meta.context_size": 2
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_asr/src/main.cpp b/projects/llm_framework/main_asr/src/main.cpp
index 19fb3363..eb3c80c7 100644
--- a/projects/llm_framework/main_asr/src/main.cpp
+++ b/projects/llm_framework/main_asr/src/main.cpp
@@ -6,6 +6,7 @@
 #include "StackFlow.h"
 #include "sherpa-ncnn/csrc/recognizer.h"
 #include "sherpa-onnx/csrc/offline-recognizer.h"
+#include "sherpa-onnx/csrc/online-recognizer.h"
 #include "sherpa-onnx/csrc/voice-activity-detector.h"
 
 #include <iostream>
@@ -44,6 +45,12 @@ typedef std::function<void(const std::string &data, bool finish)> task_callback_
     else if (obj.contains(#key))              \
         ncnn_config_.key = obj[#key];
 
+#define ONNX_ONLINE_CONFIG_AUTO_SET(obj, key)                \
+    if (config_body.contains(#key))                          \
+        config_body.at(#key).get_to(onnx_online_config.key); \
+    else if ((obj).contains(#key))                           \
+        (obj).at(#key).get_to(onnx_online_config.key);
+
 #define ONNX_ASR_CONFIG_AUTO_SET(obj, key)        \
     if (config_body.contains(#key))               \
         onnx_asr_config_.key = config_body[#key]; \
@@ -63,13 +70,18 @@ class llm_task {
     std::unique_ptr<sherpa_ncnn::Stream> ncnn_stream_;
 
     sherpa_onnx::OfflineRecognizerConfig onnx_asr_config_;
+    sherpa_onnx::OnlineRecognizerConfig onnx_online_config;
+
     sherpa_onnx::VadModelConfig vad_config_;
     std::unique_ptr<sherpa_onnx::OfflineRecognizer> onnx_recognizer_;
+    std::unique_ptr<sherpa_onnx::OnlineRecognizer> onnx_online_recognizer_;
+    std::unique_ptr<sherpa_onnx::OnlineStream> online_stream;
     std::unique_ptr<sherpa_onnx::VoiceActivityDetector> vad_;
 
     enum EngineType {
-        ENGINE_NCNN = 0,
-        ENGINE_ONNX = 1,
+        ENGINE_NCNN   = 0,
+        ENGINE_ONNX   = 1,
+        ENGINE_ONLINE = 3,
     } engine_type_ = ENGINE_NCNN;
 
 public:
@@ -112,6 +124,8 @@ class llm_task {
 
         if (model_.rfind("sherpa-ncnn", 0) == 0) {
             engine_type_ = ENGINE_NCNN;
+        } else if (model_.rfind("sherpa-onnx", 0) == 0) {
+            engine_type_ = ENGINE_ONLINE;
         } else {
             engine_type_ = ENGINE_ONNX;
         }
@@ -301,6 +315,118 @@ class llm_task {
         return 0;
     }
 
+    int load_online_model(const nlohmann::json &config_body, const nlohmann::json &file_body)
+    {
+        std::string base_model = base_model_path_ + model_ + "/";
+        SLOGI("base_model (onnx) %s", base_model.c_str());
+
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.sampling_rate);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.feature_dim);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.low_freq);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.high_freq);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.dither);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.normalize_samples);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.snip_edges);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.frame_shift_ms);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.frame_length_ms);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.is_librosa);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.remove_dc_offset);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.preemph_coeff);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.window_type);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.nemo_normalize_type);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.num_ceps);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.use_energy);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.is_mfcc);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.is_whisper);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.is_t_one);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], feat_config.round_to_power_of_two);
+
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.transducer.encoder);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.transducer.decoder);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.transducer.joiner);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.provider_config.provider);
+
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.paraformer.encoder);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.paraformer.decoder);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.wenet_ctc.model);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.wenet_ctc.chunk_size);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.wenet_ctc.num_left_chunks);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.zipformer2_ctc.model);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.nemo_ctc.model);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.t_one_ctc.model);
+
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.zipformer_meta.encoder_dims);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.zipformer_meta.attention_dims);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.zipformer_meta.num_encoder_layers);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.zipformer_meta.cnn_module_kernels);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.zipformer_meta.left_context_len);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.zipformer_meta.T);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.zipformer_meta.decode_chunk_len);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.zipformer_meta.context_size);
+
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.tokens);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.num_threads);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.warm_up);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.model_type);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.modeling_unit);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.bpe_vocab);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], model_config.tokens_buf);
+
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], lm_config.model);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], lm_config.scale);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], lm_config.lm_num_threads);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], lm_config.lm_provider);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], lm_config.lodr_fst);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], lm_config.lodr_scale);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], lm_config.lodr_backoff_id);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], lm_config.shallow_fusion);
+
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule1.must_contain_nonsilence);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule1.min_trailing_silence);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule1.min_utterance_length);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule2.must_contain_nonsilence);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule2.min_trailing_silence);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule2.min_utterance_length);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule3.must_contain_nonsilence);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule3.min_trailing_silence);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], endpoint_config.rule3.min_utterance_length);
+
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], ctc_fst_decoder_config.graph);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], ctc_fst_decoder_config.max_active);
+
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], enable_endpoint);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], decoding_method);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], max_active_paths);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], hotwords_file);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], hotwords_score);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], blank_penalty);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], temperature_scale);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], rule_fsts);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], rule_fars);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], reset_encoder);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], hr.dict_dir);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], hr.lexicon);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], hr.rule_fsts);
+        ONNX_ONLINE_CONFIG_AUTO_SET(file_body["mode_param"], hotwords_buf);
+
+        if (config_body.contains("awake_delay"))
+            awake_delay_ = config_body["awake_delay"].get<int>();
+        else if (file_body["mode_param"].contains("awake_delay"))
+            awake_delay_ = file_body["mode_param"]["awake_delay"];
+
+        onnx_online_config.model_config.transducer.encoder =
+            base_model + onnx_online_config.model_config.transducer.encoder;
+        onnx_online_config.model_config.transducer.decoder =
+            base_model + onnx_online_config.model_config.transducer.decoder;
+        onnx_online_config.model_config.transducer.joiner =
+            base_model + onnx_online_config.model_config.transducer.joiner;
+        onnx_online_config.model_config.tokens = base_model + onnx_online_config.model_config.tokens;
+
+        onnx_online_recognizer_ = std::make_unique<sherpa_onnx::OnlineRecognizer>(onnx_online_config);
+
+        return 0;
+    }
+
     int load_model(const nlohmann::json &config_body)
     {
         if (parse_config(config_body)) {
@@ -330,6 +456,8 @@ class llm_task {
 
             if (engine_type_ == ENGINE_NCNN) {
                 return load_ncnn_model(config_body, file_body);
+            } else if (engine_type_ == ENGINE_ONLINE) {
+                return load_online_model(config_body, file_body);
             } else {
                 return load_onnx_model(config_body, file_body);
             }
@@ -451,10 +579,60 @@ class llm_task {
         }
     }
 
+    void sys_pcm_on_data_online(const std::string &raw)
+    {
+        static int count = 0;
+        if (count < delay_audio_frame_) {
+            buffer_write_char(pcmdata, raw.data(), raw.length());
+            count++;
+            return;
+        }
+        buffer_write_char(pcmdata, raw.data(), raw.length());
+        buffer_position_set(pcmdata, 0);
+
+        std::vector<float> floatSamples;
+        int16_t audio_val;
+        while (buffer_read_i16(pcmdata, &audio_val, 1)) {
+            float normalizedSample = static_cast<float>(audio_val) / INT16_MAX;
+            floatSamples.push_back(normalizedSample);
+        }
+        buffer_resize(pcmdata, 0);
+        count = 0;
+
+        if (!online_stream) online_stream = onnx_online_recognizer_->CreateStream();
+        online_stream->AcceptWaveform(onnx_online_config.feat_config.sampling_rate, floatSamples.data(),
+                                      floatSamples.size());
+
+        while (onnx_online_recognizer_->IsReady(online_stream.get())) {
+            onnx_online_recognizer_->DecodeStream(online_stream.get());
+        }
+
+        auto text = onnx_online_recognizer_->GetResult(online_stream.get()).text;
+        std::string lower_text;
+        lower_text.resize(text.size());
+        std::transform(text.begin(), text.end(), lower_text.begin(), [](auto c) { return std::tolower(c); });
+
+        if ((!lower_text.empty()) && out_callback_) out_callback_(lower_text, false);
+
+        bool is_endpoint = onnx_online_recognizer_->IsEndpoint(online_stream.get());
+
+        if (is_endpoint) {
+            if ((!lower_text.empty()) && out_callback_) {
+                out_callback_(lower_text, true);
+            }
+            online_stream.reset();
+            if (ensleep_) {
+                if (pause) pause();
+            }
+        }
+    }
+
     void sys_pcm_on_data(const std::string &raw)
     {
         if (engine_type_ == ENGINE_NCNN) {
             sys_pcm_on_data_ncnn(raw);
+        } else if (engine_type_ == ENGINE_ONLINE) {
+            sys_pcm_on_data_online(raw);
         } else {
             sys_pcm_on_data_onnx(raw);
         }
diff --git a/projects/llm_framework/main_kws/mode_kws.json b/projects/llm_framework/main_kws/mode_kws.json
index c1e11a92..8cff9bf2 100644
--- a/projects/llm_framework/main_kws/mode_kws.json
+++ b/projects/llm_framework/main_kws/mode_kws.json
@@ -14,7 +14,7 @@
         "kws.bool"
     ],
     "mode_param": {
-        "model": "kws.onnx",
+        "model": "kws.axmodel",
         "model_type": "onnx",
         "wake_wav_file": "/opt/m5stack/data/audio/wakeup_zh_cn.wav",
         "chunk_size": 32,

From 3f608d2d6089efa569679ba0223b319042a28ef2 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 19 Dec 2025 11:59:50 +0800
Subject: [PATCH 73/79] [update] add asr, kws model config

---
 .../mode_sense-voice-small-10s-ax650.json     |   4 +-
 ...mall-bilingual-zh-en-2023-02-16-ax650.json |  70 +++++++++++
 ...mer-small-bilingual-zh-en-2023-02-16.json} |   4 +-
 projects/llm_framework/main_asr/src/main.cpp  |  13 +-
 projects/llm_framework/main_audio/SConstruct  |   2 +-
 .../main_audio/audio_pyramid.json             | 115 ++++++++++++++++++
 .../llm_framework/main_audio/src/main.cpp     |   2 +-
 .../main_kws/mode_kws-ax650.json              |  41 +++++++
 projects/llm_framework/main_kws/mode_kws.json |   5 +-
 projects/llm_framework/main_kws/src/main.cpp  |  57 +++++----
 projects/llm_framework/tools/llm_pack.py      |  12 ++
 11 files changed, 281 insertions(+), 44 deletions(-)
 create mode 100644 projects/llm_framework/main_asr/mode_sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16-ax650.json
 rename projects/llm_framework/main_asr/{mode_sherpa-onnx-zipformer-bilingual-zh-en-t.json => mode_sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16.json} (91%)
 create mode 100644 projects/llm_framework/main_audio/audio_pyramid.json
 create mode 100644 projects/llm_framework/main_kws/mode_kws-ax650.json

diff --git a/projects/llm_framework/main_asr/mode_sense-voice-small-10s-ax650.json b/projects/llm_framework/main_asr/mode_sense-voice-small-10s-ax650.json
index 370c8346..60185b9c 100644
--- a/projects/llm_framework/main_asr/mode_sense-voice-small-10s-ax650.json
+++ b/projects/llm_framework/main_asr/mode_sense-voice-small-10s-ax650.json
@@ -1,7 +1,7 @@
 {
     "mode": "sense-voice-small-10s-ax650",
     "type": "asr",
-    "homepage": "https://huggingface.co/yunyu1258/qwen2.5-0.5b-ha",
+    "homepage": "https://huggingface.co/M5Stack/SenseVoiceSmall-axmodel",
     "compile_flage": "pulsar2 build --input model-10-seconds.onnx --config config_sensevoice_main_u16.json --output_dir sensevoice-axmodel --output_name model-10-seconds.axmodel --target_hardware AX650 --compiler.check 0",
     "pulsar_version": "5.0-patch1-fd447d0d",
     "capabilities": [
@@ -21,7 +21,7 @@
     "mode_param": {
         "model_config.sense_voice.model": "model.axmodel",
         "model_config.tokens": "tokens.txt",
-        "silero_vad.model": "silero_vad.onnx",
+        "silero_vad.model": "silero_vad.ort",
         "model_config.provider": "axera",
         "silence_timeout": 2000,
         "awake_delay": 50
diff --git a/projects/llm_framework/main_asr/mode_sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16-ax650.json b/projects/llm_framework/main_asr/mode_sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16-ax650.json
new file mode 100644
index 00000000..41348868
--- /dev/null
+++ b/projects/llm_framework/main_asr/mode_sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16-ax650.json
@@ -0,0 +1,70 @@
+{
+    "mode": "sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16-ax650",
+    "type": "asr",
+    "homepage": "https://huggingface.co/csukuangfj/k2fsa-zipformer-bilingual-zh-en-t",
+    "capabilities": [
+        "Automatic_Speech_Recognition",
+        "Chinese",
+        "English"
+    ],
+    "input_type": [
+        "sys.pcm",
+        "sys.cap.0_0"
+    ],
+    "output_type": [
+        "asr.utf-8",
+        "asr.bool"
+    ],
+    "mode_param": {
+        "model_config.transducer.encoder": "encoder.axmodel",
+        "model_config.transducer.decoder": "decoder.axmodel",
+        "model_config.transducer.joiner": "joiner.axmodel",
+        "model_config.tokens": "tokens.txt",
+        "feat_config.feature_dim": 80,
+        "feat_config.sampling_rate": 16000,
+        "endpoint_config.rule1.min_trailing_silence": 2.4,
+        "endpoint_config.rule2.min_trailing_silence": 1.2,
+        "endpoint_config.rule3.min_utterance_length": 30,
+        "enable_endpoint": true,
+        "awake_delay": 50,
+        "model_config.provider_config.provider": "axera",
+        "model_config.zipformer_meta.encoder_dims": [
+            256,
+            256,
+            256,
+            256,
+            256
+        ],
+        "model_config.zipformer_meta.attention_dims": [
+            192,
+            192,
+            192,
+            192,
+            192
+        ],
+        "model_config.zipformer_meta.num_encoder_layers": [
+            2,
+            2,
+            2,
+            2,
+            2
+        ],
+        "model_config.zipformer_meta.cnn_module_kernels": [
+            31,
+            31,
+            31,
+            31,
+            31
+        ],
+        "model_config.zipformer_meta.left_context_len": [
+            192,
+            96,
+            48,
+            24,
+            96
+        ],
+        "model_config.zipformer_meta.T": 103,
+        "model_config.zipformer_meta.decode_chunk_len": 96,
+        "model_config.zipformer_meta.context_size": 2
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_asr/mode_sherpa-onnx-zipformer-bilingual-zh-en-t.json b/projects/llm_framework/main_asr/mode_sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16.json
similarity index 91%
rename from projects/llm_framework/main_asr/mode_sherpa-onnx-zipformer-bilingual-zh-en-t.json
rename to projects/llm_framework/main_asr/mode_sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16.json
index e22da795..368cccb6 100644
--- a/projects/llm_framework/main_asr/mode_sherpa-onnx-zipformer-bilingual-zh-en-t.json
+++ b/projects/llm_framework/main_asr/mode_sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16.json
@@ -1,7 +1,7 @@
 {
-    "mode": "sherpa-onnx-zipformer-bilingual-zh-en-t",
+    "mode": "sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16",
     "type": "asr",
-    "homepage": "",
+    "homepage": "https://huggingface.co/csukuangfj/k2fsa-zipformer-bilingual-zh-en-t",
     "capabilities": [
         "Automatic_Speech_Recognition",
         "Chinese",
diff --git a/projects/llm_framework/main_asr/src/main.cpp b/projects/llm_framework/main_asr/src/main.cpp
index eb3c80c7..59242c97 100644
--- a/projects/llm_framework/main_asr/src/main.cpp
+++ b/projects/llm_framework/main_asr/src/main.cpp
@@ -73,6 +73,7 @@ class llm_task {
     sherpa_onnx::OnlineRecognizerConfig onnx_online_config;
 
     sherpa_onnx::VadModelConfig vad_config_;
+    std::unique_ptr<sherpa_onnx::OfflineStream> offline_stream_;
     std::unique_ptr<sherpa_onnx::OfflineRecognizer> onnx_recognizer_;
     std::unique_ptr<sherpa_onnx::OnlineRecognizer> onnx_online_recognizer_;
     std::unique_ptr<sherpa_onnx::OnlineStream> online_stream;
@@ -549,16 +550,16 @@ class llm_task {
         vad_->AcceptWaveform(floatSamples.data(), floatSamples.size());
         while (!vad_->Empty()) {
             const auto &segment = vad_->Front();
-            auto s              = onnx_recognizer_->CreateStream();
-            s->AcceptWaveform(onnx_asr_config_.feat_config.sampling_rate, segment.samples.data(),
-                              segment.samples.size());
-            onnx_recognizer_->DecodeStream(s.get());
-            const auto &result = s->GetResult();
+            if (!offline_stream_) offline_stream_ = onnx_recognizer_->CreateStream();
+            offline_stream_->AcceptWaveform(onnx_asr_config_.feat_config.sampling_rate, segment.samples.data(),
+                                            segment.samples.size());
+            onnx_recognizer_->DecodeStream(offline_stream_.get());
+            const auto &result = offline_stream_->GetResult();
             if (!result.text.empty() && out_callback_) {
-                SLOGI("onnx-asr result: %s", result.text.c_str());
                 out_callback_(result.text, true);
             }
             vad_->Pop();
+            offline_stream_.reset();
         }
 
         {
diff --git a/projects/llm_framework/main_audio/SConstruct b/projects/llm_framework/main_audio/SConstruct
index 35b70234..2afd08d8 100644
--- a/projects/llm_framework/main_audio/SConstruct
+++ b/projects/llm_framework/main_audio/SConstruct
@@ -29,7 +29,7 @@ REQUIREMENTS += ['ax_sys', 'ax_audio', 'ax_audio_3a', 'tinyalsa', 'fdk-aac', 'op
 if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
     REQUIREMENTS += ['ax_interpreter', 'ax_fdk', 'ax_opus']
 
-STATIC_FILES += [AFile('audio.json'), AFile('audio_kit.json')]
+STATIC_FILES += [AFile('audio.json'), AFile('audio_kit.json'), AFile('audio_pyramid.json')]
 STATIC_FILES += Glob('mode_*.json')
 
 env['COMPONENTS'].append({'target':'llm_audio-1.7',
diff --git a/projects/llm_framework/main_audio/audio_pyramid.json b/projects/llm_framework/main_audio/audio_pyramid.json
new file mode 100644
index 00000000..6a722154
--- /dev/null
+++ b/projects/llm_framework/main_audio/audio_pyramid.json
@@ -0,0 +1,115 @@
+{
+    "mode": "None",
+    "type": "audio",
+    "capabilities": [
+        "play",
+        "cap"
+    ],
+    "input_type": [
+        "rpc.audio.wav.base64",
+        "rpc.audio.pcm.base64"
+    ],
+    "output_type": [
+        "audio.pcm.stream"
+    ],
+    "play_param": {
+        "card": 0,
+        "device": 0,
+        "volume": 0.5,
+        "channel": 2,
+        "rate": 48000,
+        "bit": 16,
+        "stPoolConfig.MetaSize": 8192,
+        "stPoolConfig.BlkSize": 32768,
+        "stPoolConfig.BlkCnt": 37,
+        "stPoolConfig.IsMergeMode": 0,
+        "stPoolConfig.CacheMode": 0,
+        "stPoolConfig.PartitionName": "anonymous",
+        "stAttr.enBitwidth": 1,
+        "stAttr.enSoundmode": 0,
+        "stAttr.u32ChnCnt": 2,
+        "stAttr.enLinkMode": 0,
+        "stAttr.enSamplerate": 16000,
+        "stAttr.U32Depth": 30,
+        "stAttr.u32PeriodSize": 160,
+        "stAttr.u32PeriodCount": 8,
+        "stAttr.bInsertSilence": 0,
+        "stVqeAttr.s32SampleRate": 16000,
+        "stVqeAttr.u32FrameSamples": 160,
+        "stVqeAttr.stNsCfg.bNsEnable": 0,
+        "stVqeAttr.stNsCfg.enAggressivenessLevel": 2,
+        "stVqeAttr.stAgcCfg.bAgcEnable": 0,
+        "stVqeAttr.stAgcCfg.enAgcMode": 2,
+        "stVqeAttr.stAgcCfg.s16TargetLevel": -3,
+        "stVqeAttr.stAgcCfg.s16Gain": 9,
+        "stHpfAttr.bEnable": 0,
+        "stHpfAttr.s32GainDb": -3,
+        "stHpfAttr.s32Samplerate": 16000,
+        "stHpfAttr.s32Freq": 200,
+        "stLpfAttr.bEnable": 0,
+        "stLpfAttr.s32GainDb": 0,
+        "stLpfAttr.s32Samplerate": 16000,
+        "stLpfAttr.s32Freq": 3000,
+        "stEqAttr.bEnable": 0,
+        "stEqAttr.s32GainDb[0]": -10,
+        "stEqAttr.s32GainDb[1]": -3,
+        "stEqAttr.s32GainDb[2]": 3,
+        "stEqAttr.s32GainDb[3]": 5,
+        "stEqAttr.s32GainDb[4]": 10,
+        "stEqAttr.s32Samplerate": 16000,
+        "gResample": 0,
+        "enInSampleRate": 16000,
+        "gInstant": 0,
+        "gInsertSilence": 0
+    },
+    "cap_param": {
+        "sys_pcm_cap_channel": "ipc:///tmp/llm/pcm.cap.socket",
+        "card": 0,
+        "device": 1,
+        "volume": 1.0,
+        "channel": 4,
+        "rate": 16000,
+        "bit": 16,
+        "stPoolConfig.MetaSize": 8192,
+        "stPoolConfig.BlkSize": 7680,
+        "stPoolConfig.BlkCnt": 33,
+        "stPoolConfig.IsMergeMode": 0,
+        "stPoolConfig.CacheMode": 0,
+        "stPoolConfig.PartitionName": "anonymous",
+        "aistAttr.enBitwidth": 1,
+        "aistAttr.enLinkMode": 0,
+        "aistAttr.enSamplerate": 16000,
+        "aistAttr.enLayoutMode": 2,
+        "aistAttr.U32Depth": 30,
+        "aistAttr.u32PeriodSize": 160,
+        "aistAttr.u32PeriodCount": 8,
+        "aistAttr.u32ChnCnt": 2,
+        "aistVqeAttr.s32SampleRate": 16000,
+        "aistVqeAttr.u32FrameSamples": 160,
+        "aistVqeAttr.stNsCfg.bNsEnable": 1,
+        "aistVqeAttr.stNsCfg.enAggressivenessLevel": 2,
+        "aistVqeAttr.stAgcCfg.bAgcEnable": 0,
+        "aistVqeAttr.stAgcCfg.enAgcMode": 2,
+        "aistVqeAttr.stAgcCfg.s16TargetLevel": -3,
+        "aistVqeAttr.stAgcCfg.s16Gain": 9,
+        "aistVqeAttr.stAecCfg.enAecMode": 2,
+        "stHpfAttr.bEnable": 0,
+        "stHpfAttr.s32GainDb": -3,
+        "stHpfAttr.s32Samplerate": 16000,
+        "stHpfAttr.s32Freq": 200,
+        "stLpfAttr.bEnable": 0,
+        "stLpfAttr.s32GainDb": 0,
+        "stLpfAttr.s32Samplerate": 16000,
+        "stLpfAttr.s32Freq": 3000,
+        "stEqAttr.bEnable": 0,
+        "stEqAttr.s32GainDb[0]": -10,
+        "stEqAttr.s32GainDb[1]": -3,
+        "stEqAttr.s32GainDb[2]": 3,
+        "stEqAttr.s32GainDb[3]": 5,
+        "stEqAttr.s32GainDb[4]": 10,
+        "stEqAttr.s32Samplerate": 16000,
+        "gResample": 0,
+        "enOutSampleRate": 16000,
+        "gDbDetection": 0
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_audio/src/main.cpp b/projects/llm_framework/main_audio/src/main.cpp
index 3500a447..216a4d72 100644
--- a/projects/llm_framework/main_audio/src/main.cpp
+++ b/projects/llm_framework/main_audio/src/main.cpp
@@ -209,7 +209,7 @@ class llm_audio : public StackFlow {
         }
 #else
         std::list<std::string> config_file_paths =
-            get_config_file_paths(base_model_path, base_model_config_path, "audio");
+            get_config_file_paths(base_model_path, base_model_config_path, "audio_pyramid");
 #endif
         try {
             config_body = nlohmann::json::parse(data);
diff --git a/projects/llm_framework/main_kws/mode_kws-ax650.json b/projects/llm_framework/main_kws/mode_kws-ax650.json
new file mode 100644
index 00000000..a03f099c
--- /dev/null
+++ b/projects/llm_framework/main_kws/mode_kws-ax650.json
@@ -0,0 +1,41 @@
+{
+    "mode": "kws-ax650",
+    "type": "kws",
+    "homepage": "",
+    "capabilities": [
+        "Keyword_spotting",
+        "English"
+    ],
+    "input_type": [
+        "sys.pcm",
+        "sys.cap.0_0"
+    ],
+    "output_type": [
+        "kws.bool"
+    ],
+    "mode_param": {
+        "model": "kws.axmodel",
+        "model_type": "axera",
+        "wake_wav_file": "/opt/m5stack/data/audio/wakeup_zh_cn.wav",
+        "chunk_size": 32,
+        "threshold": 0.9,
+        "min_continuous_frames": 5,
+        "REFRACTORY_TIME_MS": 2000,
+        "RESAMPLE_RATE": 16000,
+        "FEAT_DIM": 80,
+        "frame_opts.samp_freq": 16000,
+        "frame_opts.frame_length_ms": 25.0,
+        "frame_opts.frame_shift_ms": 10.0,
+        "frame_opts.snip_edges": false,
+        "frame_opts.dither": 0.0,
+        "frame_opts.preemph_coeff": 0.97,
+        "frame_opts.remove_dc_offset": true,
+        "frame_opts.window_type": "povey",
+        "mel_opts.num_bins": 80,
+        "mel_opts.low_freq": 20,
+        "mel_opts.high_freq": 0,
+        "energy_floor": 0.0,
+        "use_energy": false,
+        "raw_energy": true
+    }
+}
\ No newline at end of file
diff --git a/projects/llm_framework/main_kws/mode_kws.json b/projects/llm_framework/main_kws/mode_kws.json
index 8cff9bf2..14f0d4b0 100644
--- a/projects/llm_framework/main_kws/mode_kws.json
+++ b/projects/llm_framework/main_kws/mode_kws.json
@@ -15,7 +15,7 @@
     ],
     "mode_param": {
         "model": "kws.axmodel",
-        "model_type": "onnx",
+        "model_type": "axera",
         "wake_wav_file": "/opt/m5stack/data/audio/wakeup_zh_cn.wav",
         "chunk_size": 32,
         "threshold": 0.9,
@@ -37,6 +37,5 @@
         "energy_floor": 0.0,
         "use_energy": false,
         "raw_energy": true
-    },
-    "mode_param_bak": {}
+    }
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_kws/src/main.cpp b/projects/llm_framework/main_kws/src/main.cpp
index 8e70be53..d14437ff 100644
--- a/projects/llm_framework/main_kws/src/main.cpp
+++ b/projects/llm_framework/main_kws/src/main.cpp
@@ -39,7 +39,7 @@ typedef std::function<void(const std::string &data, bool finish)> task_callback_
 #include "sherpa-onnx/csrc/keyword-spotter.h"
 #include "kaldi-native-fbank/csrc/online-feature.h"
 
-typedef struct mode_config_onnx {
+typedef struct mode_config_axera {
     int chunk_size            = 32;
     float threshold           = 0.9f;
     int min_continuous_frames = 5;
@@ -47,7 +47,7 @@ typedef struct mode_config_onnx {
     int RESAMPLE_RATE         = 16000;
     int FEAT_DIM              = 80;
     int delay_audio_frame_    = 32;
-} kws_config_onnx;
+} kws_config_axera;
 
 class llm_task {
 private:
@@ -72,12 +72,11 @@ class llm_task {
     std::unique_ptr<sherpa_onnx::KeywordSpotter> sherpa_spotter_;
     std::unique_ptr<sherpa_onnx::OnlineStream> sherpa_stream_;
 
-    kws_config_onnx onnx_config_;
+    kws_config_axera axera_config_;
     std::vector<float> axera_cache_;
     std::unique_ptr<EngineWrapper> axera_session_;
     knf::FbankOptions fbank_opts_;
     std::unique_ptr<knf::OnlineFbank> fbank_;
-    Ort::Env onnx_env_{ORT_LOGGING_LEVEL_WARNING, "kws"};
     Ort::SessionOptions session_options_;
     int count_frames_               = 0;
     long long last_trigger_time_ms_ = -1e9;
@@ -117,7 +116,7 @@ class llm_task {
             if (model_.rfind("sherpa-onnx", 0) == 0) {
                 model_type_ = "sherpa";
             } else {
-                model_type_ = "onnx";
+                model_type_ = "axera";
             }
 
             if (config_body.contains("enwake_audio")) {
@@ -294,11 +293,11 @@ class llm_task {
     }
 #undef CONFIG_AUTO_SET_SHERPA
 
-#define CONFIG_AUTO_SET_ONNX(obj, key)        \
+#define CONFIG_AUTO_SET_AXERA(obj, key)        \
     if (config_body.contains(#key))           \
-        onnx_config_.key = config_body[#key]; \
+        axera_config_.key = config_body[#key]; \
     else if (obj.contains(#key))              \
-        onnx_config_.key = obj[#key];
+        axera_config_.key = obj[#key];
 
 #define OPTS_AUTO_SET(obj, key)              \
     if (config_body.contains(#key))          \
@@ -306,7 +305,7 @@ class llm_task {
     else if (obj.contains(#key))             \
         fbank_opts_.key = obj[#key];
 
-    int load_model_onnx(const nlohmann::json &config_body)
+    int load_model_axera(const nlohmann::json &config_body)
     {
         nlohmann::json file_body;
         std::list<std::string> config_file_paths =
@@ -345,13 +344,13 @@ class llm_task {
             axera_cache_.assign(1 * 32 * 88, 0.0f);
 
             auto &mp = file_body["mode_param"];
-            CONFIG_AUTO_SET_ONNX(mp, chunk_size);
-            CONFIG_AUTO_SET_ONNX(mp, threshold);
-            CONFIG_AUTO_SET_ONNX(mp, min_continuous_frames);
-            CONFIG_AUTO_SET_ONNX(mp, REFRACTORY_TIME_MS);
-            CONFIG_AUTO_SET_ONNX(mp, RESAMPLE_RATE);
-            CONFIG_AUTO_SET_ONNX(mp, FEAT_DIM);
-            CONFIG_AUTO_SET_ONNX(mp, delay_audio_frame_);
+            CONFIG_AUTO_SET_AXERA(mp, chunk_size);
+            CONFIG_AUTO_SET_AXERA(mp, threshold);
+            CONFIG_AUTO_SET_AXERA(mp, min_continuous_frames);
+            CONFIG_AUTO_SET_AXERA(mp, REFRACTORY_TIME_MS);
+            CONFIG_AUTO_SET_AXERA(mp, RESAMPLE_RATE);
+            CONFIG_AUTO_SET_AXERA(mp, FEAT_DIM);
+            CONFIG_AUTO_SET_AXERA(mp, delay_audio_frame_);
 
             OPTS_AUTO_SET(mp, frame_opts.samp_freq);
             OPTS_AUTO_SET(mp, frame_opts.frame_length_ms);
@@ -373,7 +372,7 @@ class llm_task {
             SLOGE("config file read false");
             return -3;
         }
-        delay_audio_frame_ = onnx_config_.delay_audio_frame_;
+        delay_audio_frame_ = axera_config_.delay_audio_frame_;
         return 0;
     }
 #undef CONFIG_AUTO_SET_ONNX
@@ -383,11 +382,11 @@ class llm_task {
     {
         bool triggered = false;
         for (auto score : scores) {
-            if (score > onnx_config_.threshold) {
+            if (score > axera_config_.threshold) {
                 count_frames_++;
-                if (count_frames_ >= onnx_config_.min_continuous_frames) {
-                    long long trigger_time_ms = (frame_index_global_ - onnx_config_.min_continuous_frames + 1) * 10;
-                    if (trigger_time_ms - last_trigger_time_ms_ >= onnx_config_.REFRACTORY_TIME_MS) {
+                if (count_frames_ >= axera_config_.min_continuous_frames) {
+                    long long trigger_time_ms = (frame_index_global_ - axera_config_.min_continuous_frames + 1) * 10;
+                    if (trigger_time_ms - last_trigger_time_ms_ >= axera_config_.REFRACTORY_TIME_MS) {
                         last_trigger_time_ms_ = trigger_time_ms;
                         triggered             = true;
                     }
@@ -420,11 +419,11 @@ class llm_task {
     std::vector<float> run_inference(const std::vector<float> &audio_chunk_16k)
     {
         std::vector<std::vector<float>> fbank_feats =
-            compute_fbank_kaldi(audio_chunk_16k, onnx_config_.RESAMPLE_RATE, onnx_config_.FEAT_DIM);
+            compute_fbank_kaldi(audio_chunk_16k, axera_config_.RESAMPLE_RATE, axera_config_.FEAT_DIM);
         if (fbank_feats.empty()) return {};
 
         constexpr int FIX_T = 32;
-        const int FEAT_DIM  = onnx_config_.FEAT_DIM;
+        const int FEAT_DIM  = axera_config_.FEAT_DIM;
 
         std::vector<float> mat_flattened;
         mat_flattened.resize(FIX_T * FEAT_DIM, 0.0f);
@@ -467,11 +466,11 @@ class llm_task {
         if (parse_config(config_body)) {
             return -1;
         }
-        if (model_type_ == "onnx") {
-            SLOGE("load onnx kws model");
-            return load_model_onnx(config_body);
+        if (model_type_ == "axera") {
+            SLOGI("load axera kws model");
+            return load_model_axera(config_body);
         } else {
-            SLOGE("load sherpa kws model");
+            SLOGI("load sherpa kws model");
             return load_model_sherpa(config_body);
         }
     }
@@ -498,7 +497,7 @@ class llm_task {
         int16_t audio_val;
         while (buffer_read_i16(pcmdata, &audio_val, 1)) {
             int16Samples.push_back(audio_val);
-            if (model_type_ == "onnx") {
+            if (model_type_ == "axera") {
                 floatSamples.push_back(static_cast<float>(audio_val) / 1.0f);
             } else {
                 floatSamples.push_back(static_cast<float>(audio_val) / INT16_MAX);
@@ -507,7 +506,7 @@ class llm_task {
         buffer_resize(pcmdata, 0);
         count = 0;
 
-        if (model_type_ == "onnx") {
+        if (model_type_ == "axera") {
             auto scores = run_inference(floatSamples);
             if (detect_wakeup(scores)) {
                 if (enwake_audio_ && (!wake_wav_file_.empty()) && play_awake_wav) {
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 8229e549..80f62e0b 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -397,6 +397,18 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-model-sherpa-ncnn-streaming-zipformer-zh-14M-2023-02-23':[create_data_deb,'llm-model-sherpa-ncnn-streaming-zipformer-zh-14M-2023-02-23', data_version, src_folder, revision],
         'llm-model-sherpa-onnx-kws-zipformer-gigaspeech-3.3M-2024-01-01':[create_data_deb,'llm-model-sherpa-onnx-kws-zipformer-gigaspeech-3.3M-2024-01-01', '0.3', src_folder, revision],
         'llm-model-sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01':[create_data_deb,'llm-model-sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01', '0.3', src_folder, revision],
+        'llm-model-sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16':[create_data_deb,'llm-model-sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16', '0.4', src_folder, revision],
+        'llm-model-sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16-ax650':[create_data_deb,'llm-model-sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16-ax650', '0.4', src_folder, revision],
+        # KWS model
+        ## AX630C
+        'llm-model-kws':[create_data_deb,'llm-model-kws', '0.4', src_folder, revision],
+        ## AX650
+        'llm-model-kws-ax650':[create_data_deb,'llm-model-kws-ax650', '0.4', src_folder, revision],
+        # ASR model
+        ## AX630C
+        'llm-model-sense-voice-small-10s':[create_data_deb,'llm-model-sense-voice-small-10s', '0.4', src_folder, revision],
+        ## AX650
+        'llm-model-sense-voice-small-10s-ax650':[create_data_deb,'llm-model-sense-voice-small-10s-ax650', '0.4', src_folder, revision],
         # TTS model
         'llm-model-single-speaker-english-fast':[create_data_deb,'llm-model-single-speaker-english-fast', '0.3', src_folder, revision],
         'llm-model-single-speaker-fast':[create_data_deb,'llm-model-single-speaker-fast', '0.3', src_folder, revision],

From eff3a479280ca07323f6b0ee039617abd163473d Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 19 Dec 2025 17:31:29 +0800
Subject: [PATCH 74/79] [update] perf llm-asr, kws add  buttons control.

---
 projects/llm_framework/main_asr/src/main.cpp |  74 ++++++++-
 projects/llm_framework/main_kws/src/main.cpp | 162 ++++++++++++++++++-
 projects/llm_framework/tools/llm_pack.py     |   2 +-
 3 files changed, 225 insertions(+), 13 deletions(-)

diff --git a/projects/llm_framework/main_asr/src/main.cpp b/projects/llm_framework/main_asr/src/main.cpp
index 59242c97..78457cfe 100644
--- a/projects/llm_framework/main_asr/src/main.cpp
+++ b/projects/llm_framework/main_asr/src/main.cpp
@@ -85,6 +85,23 @@ class llm_task {
         ENGINE_ONLINE = 3,
     } engine_type_ = ENGINE_NCNN;
 
+    static constexpr int kSampleRate   = 16000;
+    static constexpr int kFrameSamples = 160;
+    int pre_roll_frames_               = 30;
+    std::deque<int16_t> pre_roll_pcm_;
+    bool prev_vad_detected_ = false;
+
+private:
+    void PushPreRollPcm(const int16_t *pcm, size_t n)
+    {
+        pre_roll_pcm_.insert(pre_roll_pcm_.end(), pcm, pcm + n);
+
+        const size_t max_samples = (size_t)pre_roll_frames_ * kFrameSamples;
+        while (pre_roll_pcm_.size() > max_samples) {
+            pre_roll_pcm_.pop_front();
+        }
+    }
+
 public:
     std::string model_;
     std::string response_format_;
@@ -529,48 +546,86 @@ class llm_task {
 
     void sys_pcm_on_data_onnx(const std::string &raw)
     {
+        if (raw.size() >= sizeof(int16_t)) {
+            const int16_t *pcm16 = reinterpret_cast<const int16_t *>(raw.data());
+            size_t n16           = raw.size() / sizeof(int16_t);
+            PushPreRollPcm(pcm16, n16);
+        }
+
         static int count = 0;
         if (count < delay_audio_frame_) {
             buffer_write_char(pcmdata, raw.data(), raw.length());
             count++;
             return;
         }
+
         buffer_write_char(pcmdata, raw.data(), raw.length());
         buffer_position_set(pcmdata, 0);
 
         std::vector<float> floatSamples;
+        floatSamples.reserve((delay_audio_frame_ + 1) * kFrameSamples);
+
         int16_t audio_val;
         while (buffer_read_i16(pcmdata, &audio_val, 1)) {
-            float normalizedSample = static_cast<float>(audio_val) / INT16_MAX;
-            floatSamples.push_back(normalizedSample);
+            floatSamples.push_back(static_cast<float>(audio_val) / 32768.0f);
         }
+
         buffer_resize(pcmdata, 0);
         count = 0;
 
         vad_->AcceptWaveform(floatSamples.data(), floatSamples.size());
+
+        bool detected      = vad_->IsSpeechDetected();
+        bool speech_start  = (!prev_vad_detected_ && detected);
+        prev_vad_detected_ = detected;
+
         while (!vad_->Empty()) {
             const auto &segment = vad_->Front();
-            if (!offline_stream_) offline_stream_ = onnx_recognizer_->CreateStream();
-            offline_stream_->AcceptWaveform(onnx_asr_config_.feat_config.sampling_rate, segment.samples.data(),
-                                            segment.samples.size());
+
+            if (!offline_stream_) {
+                offline_stream_ = onnx_recognizer_->CreateStream();
+            }
+
+            if (speech_start && !pre_roll_pcm_.empty()) {
+                std::vector<float> pre;
+                pre.reserve(pre_roll_pcm_.size());
+                for (int16_t s : pre_roll_pcm_) {
+                    pre.push_back(static_cast<float>(s) / 32768.0f);
+                }
+
+                std::vector<float> merged;
+                merged.reserve(pre.size() + segment.samples.size());
+                merged.insert(merged.end(), pre.begin(), pre.end());
+                merged.insert(merged.end(), segment.samples.begin(), segment.samples.end());
+
+                offline_stream_->AcceptWaveform(kSampleRate, merged.data(), merged.size());
+
+                pre_roll_pcm_.clear();
+                speech_start = false;
+            } else {
+                offline_stream_->AcceptWaveform(kSampleRate, segment.samples.data(), segment.samples.size());
+            }
+
             onnx_recognizer_->DecodeStream(offline_stream_.get());
+
             const auto &result = offline_stream_->GetResult();
             if (!result.text.empty() && out_callback_) {
                 out_callback_(result.text, true);
             }
+
             vad_->Pop();
+
             offline_stream_.reset();
         }
 
         {
-            bool detected  = vad_->IsSpeechDetected();
             float chunk_ms = (delay_audio_frame_ + 1) * 10.0f;
-
             if (detected) {
                 silence_ms_accum_ = 0.0f;
             } else {
                 silence_ms_accum_ += chunk_ms;
             }
+
             if (silence_ms_accum_ >= silence_timeout) {
                 if (ensleep_) {
                     if (pause) pause();
@@ -1034,6 +1089,11 @@ class llm_asr : public StackFlow {
                 ++it;
             }
         }
+
+        if (data.find("sys") != std::string::npos) {
+            llm_task_obj->audio_flage_ = false;
+        }
+
         send("None", "None", LLM_NO_ERROR, work_id);
     }
 
diff --git a/projects/llm_framework/main_kws/src/main.cpp b/projects/llm_framework/main_kws/src/main.cpp
index d14437ff..4883478b 100644
--- a/projects/llm_framework/main_kws/src/main.cpp
+++ b/projects/llm_framework/main_kws/src/main.cpp
@@ -81,6 +81,7 @@ class llm_task {
     int count_frames_               = 0;
     long long last_trigger_time_ms_ = -1e9;
     long long frame_index_global_   = 0;
+    int last_btn_204_state          = -1;
 
 public:
     inline const std::string &model() const
@@ -294,9 +295,9 @@ class llm_task {
 #undef CONFIG_AUTO_SET_SHERPA
 
 #define CONFIG_AUTO_SET_AXERA(obj, key)        \
-    if (config_body.contains(#key))           \
+    if (config_body.contains(#key))            \
         axera_config_.key = config_body[#key]; \
-    else if (obj.contains(#key))              \
+    else if (obj.contains(#key))               \
         axera_config_.key = obj[#key];
 
 #define OPTS_AUTO_SET(obj, key)              \
@@ -537,9 +538,27 @@ class llm_task {
         }
     }
 
-    void trigger()
+    void trigger_wakeup()
     {
-        if (out_callback_) out_callback_("", true);
+        if (enwake_audio_ && (!wake_wav_file_.empty()) && play_awake_wav) {
+            play_awake_wav(wake_wav_file_);
+        }
+        if (out_callback_) {
+            if (enoutput_json_)
+                out_callback_("{\"reason\":\"button_204\"}", true);
+            else
+                out_callback_("", true);
+        }
+    }
+
+    void set_btn_204_state(int state)
+    {
+        last_btn_204_state = state;
+    }
+
+    int get_btn_204_state()
+    {
+        return last_btn_204_state;
     }
 
     bool delete_model()
@@ -790,6 +809,40 @@ class llm_kws : public StackFlow {
         llm_task_obj->sys_pcm_on_data((*next_data));
     }
 
+    void task_buttons_data(const std::weak_ptr<llm_task> llm_task_obj_weak,
+                           const std::weak_ptr<llm_channel_obj> llm_channel_weak, const std::string &object,
+                           const std::string &data)
+    {
+        auto llm_task_obj = llm_task_obj_weak.lock();
+        auto llm_channel  = llm_channel_weak.lock();
+        if (!(llm_task_obj && llm_channel)) {
+            return;
+        }
+        if (data.empty() || (data == "None")) return;
+
+        try {
+            std::string user_msg    = sample_unescapeString(data);
+            nlohmann::json btn_json = nlohmann::json::parse(user_msg);
+
+            if (btn_json.contains("code") && btn_json.contains("vale")) {
+                int current_code = btn_json["code"];
+                int current_vale = btn_json["vale"];
+
+                if (current_vale == 204) {
+                    int last_code = llm_task_obj->get_btn_204_state();
+
+                    if (last_code == 0 && current_code == 1) {
+                        llm_task_obj->trigger_wakeup();
+                    }
+
+                    llm_task_obj->set_btn_204_state(current_code);
+                }
+            }
+        } catch (const std::exception &e) {
+            SLOGE("Button data JSON parse error: %s", e.what());
+        }
+    }
+
     int setup(const std::string &work_id, const std::string &object, const std::string &data) override
     {
         nlohmann::json error_body;
@@ -836,6 +889,17 @@ class llm_kws : public StackFlow {
                     llm_channel->subscriber_work_id("", std::bind(&llm_kws::task_user_data, this, _llm_task_obj,
                                                                   std::weak_ptr<llm_channel_obj>(llm_channel),
                                                                   std::placeholders::_1, std::placeholders::_2));
+                } else if (input.find("buttons_thread") != std::string::npos) {
+                    std::string socket_url = "ipc:///tmp/llm/ec_prox.event.socket";
+                    auto business_logic    = std::bind(
+                        &llm_kws::task_buttons_data, this, std::weak_ptr<llm_task>(llm_task_obj),
+                        std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1, std::placeholders::_2);
+
+                    llm_channel->subscriber(
+                        socket_url, [llm_channel, business_logic](StackFlows::pzmq *p,
+                                                                  const std::shared_ptr<StackFlows::pzmq_data> &d) {
+                            llm_channel->subscriber_event_call(business_logic, p, d);
+                        });
                 }
             }
             llm_task_[work_id_num] = llm_task_obj;
@@ -851,6 +915,94 @@ class llm_kws : public StackFlow {
         }
     }
 
+    void link(const std::string &work_id, const std::string &object, const std::string &data) override
+    {
+        SLOGI("llm_kws::link:%s", data.c_str());
+        int ret = 0;
+        nlohmann::json error_body;
+
+        int work_id_num = sample_get_work_id_num(work_id);
+        if (llm_task_.find(work_id_num) == llm_task_.end()) {
+            error_body["code"]    = -6;
+            error_body["message"] = "Unit Does Not Exist";
+            send("None", "None", error_body, work_id);
+            return;
+        }
+
+        auto llm_channel  = get_channel(work_id);
+        auto llm_task_obj = llm_task_[work_id_num];
+
+        if (data.find("sys") != std::string::npos) {
+            if (audio_url_.empty()) audio_url_ = unit_call("audio", "cap", data);
+
+            std::weak_ptr<llm_task> _llm_task_obj = llm_task_obj;
+            llm_channel->subscriber(audio_url_, [_llm_task_obj](pzmq *_pzmq, const std::shared_ptr<pzmq_data> &raw) {
+                if (auto p = _llm_task_obj.lock()) p->sys_pcm_on_data(raw->string());
+            });
+
+            llm_task_obj->audio_flage_ = true;
+            llm_task_obj->inputs_.push_back(data);
+        } else if (data.find("buttons_thread") != std::string::npos) {
+            std::string socket_url = "ipc:///tmp/llm/ec_prox.event.socket";
+            auto business_logic =
+                std::bind(&llm_kws::task_buttons_data, this, std::weak_ptr<llm_task>(llm_task_obj),
+                          std::weak_ptr<llm_channel_obj>(llm_channel), std::placeholders::_1, std::placeholders::_2);
+
+            llm_channel->subscriber(
+                socket_url,
+                [llm_channel, business_logic](StackFlows::pzmq *p, const std::shared_ptr<StackFlows::pzmq_data> &d) {
+                    llm_channel->subscriber_event_call(business_logic, p, d);
+                });
+
+            llm_task_obj->inputs_.push_back(data);
+        } else {
+            error_body["code"]    = -22;
+            error_body["message"] = "unsupported link target";
+            send("None", "None", error_body, work_id);
+            return;
+        }
+
+        if (ret) {
+            error_body["code"]    = -20;
+            error_body["message"] = "link false";
+            send("None", "None", error_body, work_id);
+            return;
+        }
+        send("None", "None", LLM_NO_ERROR, work_id);
+    }
+
+    void unlink(const std::string &work_id, const std::string &object, const std::string &data) override
+    {
+        SLOGI("llm_kws::unlink:%s", data.c_str());
+        nlohmann::json error_body;
+
+        int work_id_num = sample_get_work_id_num(work_id);
+        if (llm_task_.find(work_id_num) == llm_task_.end()) {
+            error_body["code"]    = -6;
+            error_body["message"] = "Unit Does Not Exist";
+            send("None", "None", error_body, work_id);
+            return;
+        }
+
+        auto llm_channel  = get_channel(work_id);
+        auto llm_task_obj = llm_task_[work_id_num];
+
+        llm_channel->stop_subscriber_work_id(data);
+
+        for (auto it = llm_task_obj->inputs_.begin(); it != llm_task_obj->inputs_.end();) {
+            if (*it == data)
+                it = llm_task_obj->inputs_.erase(it);
+            else
+                ++it;
+        }
+
+        if (data.find("sys") != std::string::npos) {
+            llm_task_obj->audio_flage_ = false;
+        }
+
+        send("None", "None", LLM_NO_ERROR, work_id);
+    }
+
     void taskinfo(const std::string &work_id, const std::string &object, const std::string &data) override
     {
         SLOGI("llm_kws::taskinfo:%s", data.c_str());
@@ -937,7 +1089,7 @@ class llm_kws : public StackFlow {
             _zmq.send_data(out);
             return LLM_NONE;
         }
-        llm_task_[work_id_num]->trigger();
+        llm_task_[work_id_num]->trigger_wakeup();
         return LLM_NONE;
     }
 
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 80f62e0b..36a7f745 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -480,7 +480,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-model-qwen3-vl-2B-Int4-ax630c':[create_data_deb,'llm-model-qwen3-vl-2B-Int4-ax630c', '0.5', src_folder, revision],
         ## AX650
         'llm-model-qwen2.5-0.5B-Int4-ax650':[create_data_deb,'llm-model-qwen2.5-0.5B-Int4-ax650', '0.4', src_folder, revision],
-        'llm-model-qwen2.5-HA-0.5B-ctx-ax650':[create_data_deb,'llm-model-qwen2.5-HA-0.5B-ctx-ax650', '0.5', src_folder, revision],
+        'llm-model-qwen2.5-HA-0.5B-ctx-ax650':[create_data_deb,'llm-model-qwen2.5-HA-0.5B-ctx-ax650', '0.6', src_folder, revision],
         'llm-model-qwen2.5-1.5B-Int4-ax650':[create_data_deb,'llm-model-qwen2.5-1.5B-Int4-ax650', '0.4', src_folder, revision],
         'llm-model-qwen2.5-3B-Int4-ax650':[create_data_deb,'llm-model-qwen2.5-3B-Int4-ax650', '0.4', src_folder, revision],
         'llm-model-qwen2.5-7B-Int4-ax650':[create_data_deb,'llm-model-qwen2.5-7B-Int4-ax650', '0.4', src_folder, revision],

From 7b671f21e03f105bf9062591eb389a54150ad5a6 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Mon, 22 Dec 2025 12:19:30 +0800
Subject: [PATCH 75/79] [update] update melotts play stop cap

---
 .../llm_framework/main_audio/src/main.cpp     |  8 ++---
 .../llm_framework/main_melotts/src/main.cpp   | 36 +++++++++++++++++++
 2 files changed, 40 insertions(+), 4 deletions(-)

diff --git a/projects/llm_framework/main_audio/src/main.cpp b/projects/llm_framework/main_audio/src/main.cpp
index 216a4d72..6a54c7b2 100644
--- a/projects/llm_framework/main_audio/src/main.cpp
+++ b/projects/llm_framework/main_audio/src/main.cpp
@@ -495,13 +495,13 @@ class llm_audio : public StackFlow {
 #endif
 
         if (_rawdata == "play") {
-            return play_state ? "None" : "Runing";
+            return play_state ? "None" : "Running";
         } else if (_rawdata == "cap") {
-            return cap_state ? "None" : "Runing";
+            return cap_state ? "None" : "Running";
         } else {
             std::ostringstream return_val;
-            return_val << "{\"play\":" << (play_state ? "\"None\"" : "\"Runing\"")
-                       << ",\"cap\":" << (cap_state ? "\"None\"" : "\"Runing\"") << "}";
+            return_val << "{\"play\":" << (play_state ? "\"None\"" : "\"Running\"")
+                       << ",\"cap\":" << (cap_state ? "\"None\"" : "\"Running\"") << "}";
             return return_val.str();
         }
     }
diff --git a/projects/llm_framework/main_melotts/src/main.cpp b/projects/llm_framework/main_melotts/src/main.cpp
index c4e2122c..4b6f3511 100644
--- a/projects/llm_framework/main_melotts/src/main.cpp
+++ b/projects/llm_framework/main_melotts/src/main.cpp
@@ -22,6 +22,7 @@
 #include <samplerate.h>
 #include "../../../../SDK/components/utilities/include/sample_log.h"
 #include "subprocess.h"
+#include <global_config.h>
 
 using namespace StackFlows;
 
@@ -87,6 +88,7 @@ class llm_task {
     task_callback_t out_callback_;
     bool enaudio_;
     int awake_delay_ = 1000;
+    bool cap_;
     std::string tts_string_stream_buff;
 
     bool parse_config(const nlohmann::json &config_body)
@@ -255,12 +257,42 @@ class llm_task {
     {
         try {
             std::vector<int16_t> wav_pcm_data;
+#if !defined(CONFIG_AX_620E_MSP_ENABLED) && !defined(CONFIG_AX_620Q_MSP_ENABLED)
+            std::string initial_status = unit_call("audio", "audio_status", "sys");
+            if (!cap_ && initial_status.find("\"cap\":\"Running\"") != std::string::npos) {
+                unit_call("audio", "cap_stop_all", "sys");
+                cap_ = true;
+            }
+#endif
             if (msg_str.empty()) {
                 if (out_callback_) {
                     std::string output = wav_pcm_data.empty() ? std::string()
                                                               : std::string((char *)wav_pcm_data.data(),
                                                                             wav_pcm_data.size() * sizeof(int16_t));
                     out_callback_(output, finish);
+#if !defined(CONFIG_AX_620E_MSP_ENABLED) && !defined(CONFIG_AX_620Q_MSP_ENABLED)
+                    int none_count           = 0;
+                    const int max_iterations = 100;
+
+                    for (int i = 0; i < max_iterations; ++i) {
+                        std::string current_status = unit_call("audio", "audio_status", "sys");
+                        if (current_status.find("\"play\":\"None\"") != std::string::npos) {
+                            none_count++;
+                        } else {
+                            none_count = 0;
+                        }
+                        if (none_count >= 5) {
+                            break;
+                        }
+
+                        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+                    }
+
+                    if (cap_) {
+                        unit_call("audio", "cap", "sys");
+                        cap_ = false;
+                    }
+#endif
                 }
                 return false;
             }
@@ -528,7 +560,11 @@ class llm_tts : public StackFlow {
             llm_channel->send(llm_task_obj->response_format_, base64_data, LLM_NO_ERROR);
         }
         if (llm_task_obj->response_format_.find("sys") != std::string::npos) {
+#if defined(CONFIG_AX_620E_MSP_ENABLED) || defined(CONFIG_AX_620Q_MSP_ENABLED)
             unit_call("audio", "queue_play", data);
+#else
+            unit_call("audio", "play_raw", data);
+#endif
         }
     }
 

From 3ab3d87a19aee4e661a0288e314ba90dd8d71c11 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 26 Dec 2025 10:29:20 +0800
Subject: [PATCH 76/79] [update] update package version

---
 .../llm_cosyvoice2.md                         |  2 +-
 projects/llm_framework/main_asr/SConstruct    |  2 +-
 projects/llm_framework/main_audio/SConstruct  |  2 +-
 .../llm_framework/main_cosy_voice/SConstruct  |  4 +--
 .../src/runner/Tokenizer/Tokenizer.cpp        |  4 +--
 projects/llm_framework/main_kws/SConstruct    |  2 +-
 projects/llm_framework/main_llm/SConstruct    |  2 +-
 .../src/runner/Tokenizer/Tokenizer.cpp        |  4 +--
 .../llm_framework/main_melotts/SConstruct     |  2 +-
 .../models/mode_melotts-en-au-ax650.json      |  2 +-
 .../models/mode_melotts-en-br-ax650.json      |  2 +-
 .../models/mode_melotts-en-default-ax650.json |  2 +-
 .../models/mode_melotts-en-india-ax650.json   |  2 +-
 .../models/mode_melotts-en-us-ax650.json      |  2 +-
 .../models/mode_melotts-es-es-ax650.json      |  2 +-
 .../models/mode_melotts-ja-jp-ax650.json      |  2 +-
 .../models/mode_melotts-zh-cn-ax650.json      |  2 +-
 projects/llm_framework/main_vlm/SConstruct    |  2 +-
 .../src/runner/Tokenizer/Tokenizer.cpp        |  6 ++--
 projects/llm_framework/tools/llm_pack.py      | 32 +++++++++----------
 20 files changed, 40 insertions(+), 40 deletions(-)

diff --git a/doc/projects_llm_framework_doc/llm_cosyvoice2.md b/doc/projects_llm_framework_doc/llm_cosyvoice2.md
index 7f3a7ab6..19a001f7 100644
--- a/doc/projects_llm_framework_doc/llm_cosyvoice2.md
+++ b/doc/projects_llm_framework_doc/llm_cosyvoice2.md
@@ -31,7 +31,7 @@ cosy_voice
 - object：传输的数据类型为 `cosy_voice.setup`。
 - model：使用的模型为 `CosyVoice2-0.5B-ax650` 模型。
 - prompt_files：要克隆的音频信息文件。
-- response_format：返回结果为 `sys.pcm`, 系统音频数据，并直接发送到 llm-audio 模块进行播放。返回结果为 `file`, 生成的音频写 wav 文件，可用 `prompt_data` 指定路径或文件名。
+- response_format：返回结果为 `sys.pcm`, 系统音频数据，并直接发送到 llm-audio 模块进行播放。返回结果为 `file`, 生成的音频写 wav 文件，可用 `prompt_dir` 指定路径或文件名。
 - input：输入的为 `tts.utf-8`,代表的是从用户输入。
 - enoutput：是否起用用户结果输出。
 
diff --git a/projects/llm_framework/main_asr/SConstruct b/projects/llm_framework/main_asr/SConstruct
index 217bbc51..1098aab6 100644
--- a/projects/llm_framework/main_asr/SConstruct
+++ b/projects/llm_framework/main_asr/SConstruct
@@ -40,7 +40,7 @@ LDFLAGS += [
 
 STATIC_FILES += Glob('mode_*.json')
 
-env['COMPONENTS'].append({'target':'llm_asr-1.7',
+env['COMPONENTS'].append({'target':'llm_asr-1.8',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_audio/SConstruct b/projects/llm_framework/main_audio/SConstruct
index 2afd08d8..3e1c8d65 100644
--- a/projects/llm_framework/main_audio/SConstruct
+++ b/projects/llm_framework/main_audio/SConstruct
@@ -32,7 +32,7 @@ if 'CONFIG_AX_620E_MSP_ENABLED' in os.environ:
 STATIC_FILES += [AFile('audio.json'), AFile('audio_kit.json'), AFile('audio_pyramid.json')]
 STATIC_FILES += Glob('mode_*.json')
 
-env['COMPONENTS'].append({'target':'llm_audio-1.7',
+env['COMPONENTS'].append({'target':'llm_audio-1.8',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_cosy_voice/SConstruct b/projects/llm_framework/main_cosy_voice/SConstruct
index 499edf79..3e5fa778 100644
--- a/projects/llm_framework/main_cosy_voice/SConstruct
+++ b/projects/llm_framework/main_cosy_voice/SConstruct
@@ -17,7 +17,7 @@ LDFLAGS = []
 LINK_SEARCH_PATH = []
 STATIC_FILES = []
 
-python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-cosy-voice-python-venv_v1.7.tar.gz", 'm5stack_llm-cosy-voice-python-venv_v1.7.tar.gz')
+python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-cosy-voice-python-venv_v1.8.tar.gz", 'm5stack_llm-cosy-voice-python-venv_v1.8.tar.gz')
 
 # REQUIREMENTS += ['Backward_cpp']
 # DYNAMIC_LIB += [ AFile('../static_lib/libdw.so.1'),
@@ -68,7 +68,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_cosy_voice-1.9',
+env['COMPONENTS'].append({'target':'llm_cosy_voice-1.10',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp
index 0829ffaf..cfea0d83 100644
--- a/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp
+++ b/projects/llm_framework/main_cosy_voice/src/runner/Tokenizer/Tokenizer.cpp
@@ -349,7 +349,7 @@ class Tokenizer_Http : public BaseTokenizer {
     bool Init(std::string model_path = "http://localhost:8080", bool b_bos = true, bool b_eos = false) override
     {
         base_url = model_path;
-        if (!test_connect_http(base_url, 20)) {
+        if (!test_connect_http(base_url, 300)) {
             ALOGE("connect %s failed", base_url.c_str());
             return false;
         } else {
@@ -395,7 +395,7 @@ class Tokenizer_Http : public BaseTokenizer {
     bool Init(std::string model_path = "http://localhost:8080") override
     {
         base_url = model_path;
-        if (!test_connect_http(base_url, 20)) {
+        if (!test_connect_http(base_url, 300)) {
             ALOGE("connect %s failed", base_url.c_str());
             return false;
         } else {
diff --git a/projects/llm_framework/main_kws/SConstruct b/projects/llm_framework/main_kws/SConstruct
index fc02a377..3ad5f920 100644
--- a/projects/llm_framework/main_kws/SConstruct
+++ b/projects/llm_framework/main_kws/SConstruct
@@ -60,7 +60,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_kws-1.9',
+env['COMPONENTS'].append({'target':'llm_kws-1.10',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_llm/SConstruct b/projects/llm_framework/main_llm/SConstruct
index d77a4fb7..305f4254 100644
--- a/projects/llm_framework/main_llm/SConstruct
+++ b/projects/llm_framework/main_llm/SConstruct
@@ -66,7 +66,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_llm-1.11',
+env['COMPONENTS'].append({'target':'llm_llm-1.12',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.cpp b/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.cpp
index 0829ffaf..cfea0d83 100644
--- a/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.cpp
+++ b/projects/llm_framework/main_llm/src/runner/Tokenizer/Tokenizer.cpp
@@ -349,7 +349,7 @@ class Tokenizer_Http : public BaseTokenizer {
     bool Init(std::string model_path = "http://localhost:8080", bool b_bos = true, bool b_eos = false) override
     {
         base_url = model_path;
-        if (!test_connect_http(base_url, 20)) {
+        if (!test_connect_http(base_url, 300)) {
             ALOGE("connect %s failed", base_url.c_str());
             return false;
         } else {
@@ -395,7 +395,7 @@ class Tokenizer_Http : public BaseTokenizer {
     bool Init(std::string model_path = "http://localhost:8080") override
     {
         base_url = model_path;
-        if (!test_connect_http(base_url, 20)) {
+        if (!test_connect_http(base_url, 300)) {
             ALOGE("connect %s failed", base_url.c_str());
             return false;
         } else {
diff --git a/projects/llm_framework/main_melotts/SConstruct b/projects/llm_framework/main_melotts/SConstruct
index 5466d8e1..e0dea067 100644
--- a/projects/llm_framework/main_melotts/SConstruct
+++ b/projects/llm_framework/main_melotts/SConstruct
@@ -32,7 +32,7 @@ REQUIREMENTS += ['onnxruntime']
 
 STATIC_FILES += Glob('models/mode_*.json')
 
-env['COMPONENTS'].append({'target':'llm_melotts-1.9',
+env['COMPONENTS'].append({'target':'llm_melotts-1.10',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-en-au-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-en-au-ax650.json
index 8b94e7ef..539a8ef2 100644
--- a/projects/llm_framework/main_melotts/models/mode_melotts-en-au-ax650.json
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-en-au-ax650.json
@@ -25,7 +25,7 @@
         "verbalizer": "en_tn_verbalizer.fst",
         "spacker_speed": 1.2,
         "mode_rate": 44100,
-        "audio_rate": 16000,
+        "audio_rate": 48000,
         "awake_delay": 1000
     }
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-en-br-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-en-br-ax650.json
index ef43e582..40bf41d2 100644
--- a/projects/llm_framework/main_melotts/models/mode_melotts-en-br-ax650.json
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-en-br-ax650.json
@@ -25,7 +25,7 @@
         "verbalizer": "en_tn_verbalizer.fst",
         "spacker_speed": 1.2,
         "mode_rate": 44100,
-        "audio_rate": 16000,
+        "audio_rate": 48000,
         "awake_delay": 1000
     }
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-en-default-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-en-default-ax650.json
index 3e19b401..04af8b0a 100644
--- a/projects/llm_framework/main_melotts/models/mode_melotts-en-default-ax650.json
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-en-default-ax650.json
@@ -25,7 +25,7 @@
         "verbalizer": "en_tn_verbalizer.fst",
         "spacker_speed": 1.2,
         "mode_rate": 44100,
-        "audio_rate": 16000,
+        "audio_rate": 48000,
         "awake_delay": 1000
     }
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-en-india-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-en-india-ax650.json
index ee0cab64..1eab34c5 100644
--- a/projects/llm_framework/main_melotts/models/mode_melotts-en-india-ax650.json
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-en-india-ax650.json
@@ -25,7 +25,7 @@
         "verbalizer": "en_tn_verbalizer.fst",
         "spacker_speed": 1.2,
         "mode_rate": 44100,
-        "audio_rate": 16000,
+        "audio_rate": 48000,
         "awake_delay": 1000
     }
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-en-us-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-en-us-ax650.json
index 6ca7181e..8490b248 100644
--- a/projects/llm_framework/main_melotts/models/mode_melotts-en-us-ax650.json
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-en-us-ax650.json
@@ -25,7 +25,7 @@
         "verbalizer": "en_tn_verbalizer.fst",
         "spacker_speed": 1.2,
         "mode_rate": 44100,
-        "audio_rate": 16000,
+        "audio_rate": 48000,
         "awake_delay": 1000
     }
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-es-es-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-es-es-ax650.json
index 7012574b..04c0cb09 100644
--- a/projects/llm_framework/main_melotts/models/mode_melotts-es-es-ax650.json
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-es-es-ax650.json
@@ -23,7 +23,7 @@
         "lexicon": "lexicon-es.txt",
         "spacker_speed": 1.2,
         "mode_rate": 44100,
-        "audio_rate": 16000,
+        "audio_rate": 48000,
         "awake_delay": 1000
     }
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-ja-jp-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-ja-jp-ax650.json
index 49f26aa7..58cd3a81 100644
--- a/projects/llm_framework/main_melotts/models/mode_melotts-ja-jp-ax650.json
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-ja-jp-ax650.json
@@ -25,7 +25,7 @@
         "verbalizer": "ja_tn_verbalizer.fst",
         "spacker_speed": 1.1,
         "mode_rate": 44100,
-        "audio_rate": 16000,
+        "audio_rate": 48000,
         "awake_delay": 1000
     }
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_melotts/models/mode_melotts-zh-cn-ax650.json b/projects/llm_framework/main_melotts/models/mode_melotts-zh-cn-ax650.json
index 41060977..cfe7e6f0 100644
--- a/projects/llm_framework/main_melotts/models/mode_melotts-zh-cn-ax650.json
+++ b/projects/llm_framework/main_melotts/models/mode_melotts-zh-cn-ax650.json
@@ -25,7 +25,7 @@
         "verbalizer": "zh_tn_verbalizer.fst",
         "spacker_speed": 1.1,
         "mode_rate": 44100,
-        "audio_rate": 16000,
+        "audio_rate": 48000,
         "awake_delay": 1000
     }
 }
\ No newline at end of file
diff --git a/projects/llm_framework/main_vlm/SConstruct b/projects/llm_framework/main_vlm/SConstruct
index 2d05a3b2..2e78455e 100644
--- a/projects/llm_framework/main_vlm/SConstruct
+++ b/projects/llm_framework/main_vlm/SConstruct
@@ -76,7 +76,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_vlm-1.10',
+env['COMPONENTS'].append({'target':'llm_vlm-1.11',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp b/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp
index 8877a2b9..5d1caf77 100644
--- a/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp
+++ b/projects/llm_framework/main_vlm/src/runner/Tokenizer/Tokenizer.cpp
@@ -25,7 +25,7 @@ class Tokenizer_Http : public BaseTokenizer {
     bool Init(std::string model_path) override
     {
         base_url = model_path;
-        if (!test_connect_http(base_url, 20)) {
+        if (!test_connect_http(base_url, 300)) {
             ALOGE("connect %s failed", base_url.c_str());
             return false;
         } else {
@@ -153,7 +153,7 @@ class Tokenizer_Http : public BaseTokenizer {
     bool Init(std::string model_path = "http://localhost:8080", bool b_bos = true, bool b_eos = false) override
     {
         base_url = model_path;
-        if (!test_connect_http(base_url, 30)) {
+        if (!test_connect_http(base_url, 300)) {
             ALOGE("connect %s failed", base_url.c_str());
             return false;
         } else {
@@ -200,7 +200,7 @@ class Tokenizer_Http : public BaseTokenizer {
     bool Init_new(std::string model_path, bool b_bos, bool b_eos) override
     {
         base_url = model_path;
-        if (!test_connect_http(base_url, 30)) {
+        if (!test_connect_http(base_url, 300)) {
             ALOGE("connect %s failed", base_url.c_str());
             return false;
         } else {
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 36a7f745..f2b58d8d 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -372,23 +372,23 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
     # cpu_count = 50
     Tasks = {
         # Unit
-        'lib-llm':[create_lib_deb,'lib-llm', '1.8', src_folder, revision],
+        'lib-llm':[create_lib_deb,'lib-llm', '1.9', src_folder, revision],
         'llm-sys':[create_bin_deb,'llm-sys', '1.6', src_folder, revision],
-        'llm-audio':[create_bin_deb,'llm-audio', '1.7', src_folder, revision],
-        'llm-kws':[create_bin_deb,'llm-kws', '1.9', src_folder, revision],
-        'llm-asr':[create_bin_deb,'llm-asr', '1.7', src_folder, revision],
-        'llm-llm':[create_bin_deb,'llm-llm', '1.11', src_folder, revision],
+        'llm-audio':[create_bin_deb,'llm-audio', '1.8', src_folder, revision],
+        'llm-kws':[create_bin_deb,'llm-kws', '1.10', src_folder, revision],
+        'llm-asr':[create_bin_deb,'llm-asr', '1.8', src_folder, revision],
+        'llm-llm':[create_bin_deb,'llm-llm', '1.12', src_folder, revision],
         'llm-tts':[create_bin_deb,'llm-tts', '1.6', src_folder, revision],
-        'llm-melotts':[create_bin_deb,'llm-melotts', '1.9', src_folder, revision],
+        'llm-melotts':[create_bin_deb,'llm-melotts', '1.10', src_folder, revision],
         'llm-camera':[create_bin_deb,'llm-camera', '1.9', src_folder, revision, 'lib-llm'],
-        'llm-vlm':[create_bin_deb,'llm-vlm', '1.10', src_folder, revision],
+        'llm-vlm':[create_bin_deb,'llm-vlm', '1.11', src_folder, revision],
         'llm-yolo':[create_bin_deb,'llm-yolo', '1.9', src_folder, revision],
         'llm-skel':[create_bin_deb,'llm-skel', version, src_folder, revision],
         'llm-depth-anything':[create_bin_deb,'llm-depth-anything', '1.7', src_folder, revision],
         'llm-vad':[create_bin_deb,'llm-vad', '1.8', src_folder, revision],
         'llm-whisper':[create_bin_deb,'llm-whisper', '1.8', src_folder, revision],
         'llm-openai-api':[create_bin_deb,'llm-openai-api', '1.9', src_folder, revision],
-        'llm-cosy-voice':[create_bin_deb,'llm-cosy-voice', '1.9', src_folder, revision],
+        'llm-cosy-voice':[create_bin_deb,'llm-cosy-voice', '1.10', src_folder, revision],
         # keyword spotting Audio file
         'llm-model-audio-en-us':[create_data_deb,'llm-model-audio-en-us', data_version, src_folder, revision],
         'llm-model-audio-zh-cn':[create_data_deb,'llm-model-audio-zh-cn', data_version, src_folder, revision],
@@ -425,14 +425,14 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-model-melotts-ja-jp':[create_data_deb,'llm-model-melotts-ja-jp', '0.6', src_folder, revision],
         'llm-model-melotts-es-es':[create_data_deb,'llm-model-melotts-es-es', '0.5', src_folder, revision],
         ## AX650
-        'llm-model-melotts-zh-cn-ax650':[create_data_deb,'llm-model-melotts-zh-cn-ax650', '0.6', src_folder, revision],
-        'llm-model-melotts-en-au-ax650':[create_data_deb,'llm-model-melotts-en-au-ax650', '0.6', src_folder, revision],
-        'llm-model-melotts-en-br-ax650':[create_data_deb,'llm-model-melotts-en-br-ax650', '0.6', src_folder, revision],
-        'llm-model-melotts-en-default-ax650':[create_data_deb,'llm-model-melotts-en-default-ax650', '0.6', src_folder, revision],
-        'llm-model-melotts-en-india-ax650':[create_data_deb,'llm-model-melotts-en-india-ax650', '0.6', src_folder, revision],
-        'llm-model-melotts-en-us-ax650':[create_data_deb,'llm-model-melotts-en-us-ax650', '0.6', src_folder, revision],
-        'llm-model-melotts-ja-jp-ax650':[create_data_deb,'llm-model-melotts-ja-jp-ax650', '0.6', src_folder, revision],
-        'llm-model-melotts-es-es-ax650':[create_data_deb,'llm-model-melotts-es-es-ax650', '0.5', src_folder, revision],
+        'llm-model-melotts-zh-cn-ax650':[create_data_deb,'llm-model-melotts-zh-cn-ax650', '0.7', src_folder, revision],
+        'llm-model-melotts-en-au-ax650':[create_data_deb,'llm-model-melotts-en-au-ax650', '0.7', src_folder, revision],
+        'llm-model-melotts-en-br-ax650':[create_data_deb,'llm-model-melotts-en-br-ax650', '0.7', src_folder, revision],
+        'llm-model-melotts-en-default-ax650':[create_data_deb,'llm-model-melotts-en-default-ax650', '0.7', src_folder, revision],
+        'llm-model-melotts-en-india-ax650':[create_data_deb,'llm-model-melotts-en-india-ax650', '0.7', src_folder, revision],
+        'llm-model-melotts-en-us-ax650':[create_data_deb,'llm-model-melotts-en-us-ax650', '0.7', src_folder, revision],
+        'llm-model-melotts-ja-jp-ax650':[create_data_deb,'llm-model-melotts-ja-jp-ax650', '0.7', src_folder, revision],
+        'llm-model-melotts-es-es-ax650':[create_data_deb,'llm-model-melotts-es-es-ax650', '0.6', src_folder, revision],
         # CosyVoice2 model
         ## AX650
         'llm-model-CosyVoice2-0.5B-ax650':[create_data_deb,'llm-model-CosyVoice2-0.5B-ax650', '0.6', src_folder, revision],

From 9c7ba31bb5d9ae5c79ff322db69c6e7aecedc451 Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 26 Dec 2025 16:53:28 +0800
Subject: [PATCH 77/79] [update] update llm-asr & model config

---
 .../mode_sense-voice-small-10s-ax650.json     |   1 +
 projects/llm_framework/main_asr/src/main.cpp  | 193 ++++++++++++------
 2 files changed, 130 insertions(+), 64 deletions(-)

diff --git a/projects/llm_framework/main_asr/mode_sense-voice-small-10s-ax650.json b/projects/llm_framework/main_asr/mode_sense-voice-small-10s-ax650.json
index 60185b9c..5448a8d8 100644
--- a/projects/llm_framework/main_asr/mode_sense-voice-small-10s-ax650.json
+++ b/projects/llm_framework/main_asr/mode_sense-voice-small-10s-ax650.json
@@ -20,6 +20,7 @@
     ],
     "mode_param": {
         "model_config.sense_voice.model": "model.axmodel",
+        "model_config.sense_voice.use_itn": true,
         "model_config.tokens": "tokens.txt",
         "silero_vad.model": "silero_vad.ort",
         "model_config.provider": "axera",
diff --git a/projects/llm_framework/main_asr/src/main.cpp b/projects/llm_framework/main_asr/src/main.cpp
index 78457cfe..024de85f 100644
--- a/projects/llm_framework/main_asr/src/main.cpp
+++ b/projects/llm_framework/main_asr/src/main.cpp
@@ -120,6 +120,14 @@ class llm_task {
     buffer_t *pcmdata;
     std::function<void(void)> pause;
 
+    bool is_using_itn() const
+    {
+        if (engine_type_ == ENGINE_ONNX) {
+            return onnx_asr_config_.model_config.sense_voice.use_itn;
+        }
+        return false;
+    }
+
     bool parse_config(const nlohmann::json &config_body)
     {
         try {
@@ -546,91 +554,146 @@ class llm_task {
 
     void sys_pcm_on_data_onnx(const std::string &raw)
     {
-        if (raw.size() >= sizeof(int16_t)) {
-            const int16_t *pcm16 = reinterpret_cast<const int16_t *>(raw.data());
-            size_t n16           = raw.size() / sizeof(int16_t);
-            PushPreRollPcm(pcm16, n16);
-        }
-
-        static int count = 0;
-        if (count < delay_audio_frame_) {
+        if (delay_audio_frame_ == 0) {
             buffer_write_char(pcmdata, raw.data(), raw.length());
-            count++;
-            return;
-        }
+            buffer_position_set(pcmdata, 0);
 
-        buffer_write_char(pcmdata, raw.data(), raw.length());
-        buffer_position_set(pcmdata, 0);
+            std::vector<float> floatSamples;
+            int16_t audio_val;
+            while (buffer_read_i16(pcmdata, &audio_val, 1)) {
+                float normalizedSample = static_cast<float>(audio_val) / INT16_MAX;
+                floatSamples.push_back(normalizedSample);
+            }
 
-        std::vector<float> floatSamples;
-        floatSamples.reserve((delay_audio_frame_ + 1) * kFrameSamples);
+            buffer_resize(pcmdata, 0);
+            int32_t window_size = vad_config_.silero_vad.window_size;
+            int32_t i           = 0;
+            std::string final_text;
 
-        int16_t audio_val;
-        while (buffer_read_i16(pcmdata, &audio_val, 1)) {
-            floatSamples.push_back(static_cast<float>(audio_val) / 32768.0f);
-        }
+            while (i < floatSamples.size()) {
+                if (i + window_size <= floatSamples.size()) {
+                    vad_->AcceptWaveform(floatSamples.data() + i, window_size);
+                } else {
+                    vad_->Flush();
+                }
+                i += window_size;
 
-        buffer_resize(pcmdata, 0);
-        count = 0;
+                while (!vad_->Empty()) {
+                    const auto &segment = vad_->Front();
+                    float duration      = segment.samples.size() / 16000.f;
+                    float start_time    = segment.start / 16000.f;
+                    float end_time      = start_time + duration;
 
-        vad_->AcceptWaveform(floatSamples.data(), floatSamples.size());
+                    if (duration < 0.1f) {
+                        vad_->Pop();
+                        continue;
+                    }
 
-        bool detected      = vad_->IsSpeechDetected();
-        bool speech_start  = (!prev_vad_detected_ && detected);
-        prev_vad_detected_ = detected;
+                    if (!offline_stream_) offline_stream_ = onnx_recognizer_->CreateStream();
 
-        while (!vad_->Empty()) {
-            const auto &segment = vad_->Front();
+                    offline_stream_->AcceptWaveform(onnx_asr_config_.feat_config.sampling_rate, segment.samples.data(),
+                                                    segment.samples.size());
 
-            if (!offline_stream_) {
-                offline_stream_ = onnx_recognizer_->CreateStream();
-            }
+                    onnx_recognizer_->DecodeStream(offline_stream_.get());
+                    const auto &result = offline_stream_->GetResult();
 
-            if (speech_start && !pre_roll_pcm_.empty()) {
-                std::vector<float> pre;
-                pre.reserve(pre_roll_pcm_.size());
-                for (int16_t s : pre_roll_pcm_) {
-                    pre.push_back(static_cast<float>(s) / 32768.0f);
-                }
+                    final_text += result.text;
 
-                std::vector<float> merged;
-                merged.reserve(pre.size() + segment.samples.size());
-                merged.insert(merged.end(), pre.begin(), pre.end());
-                merged.insert(merged.end(), segment.samples.begin(), segment.samples.end());
+                    vad_->Pop();
+                    offline_stream_.reset();
+                }
+            }
 
-                offline_stream_->AcceptWaveform(kSampleRate, merged.data(), merged.size());
+            if (out_callback_) {
+                out_callback_(final_text, true);
+            }
+        } else {
+            if (raw.size() >= sizeof(int16_t)) {
+                const int16_t *pcm16 = reinterpret_cast<const int16_t *>(raw.data());
+                size_t n16           = raw.size() / sizeof(int16_t);
+                PushPreRollPcm(pcm16, n16);
+            }
 
-                pre_roll_pcm_.clear();
-                speech_start = false;
-            } else {
-                offline_stream_->AcceptWaveform(kSampleRate, segment.samples.data(), segment.samples.size());
+            static int count = 0;
+            if (count < delay_audio_frame_) {
+                buffer_write_char(pcmdata, raw.data(), raw.length());
+                count++;
+                return;
             }
 
-            onnx_recognizer_->DecodeStream(offline_stream_.get());
+            buffer_write_char(pcmdata, raw.data(), raw.length());
+            buffer_position_set(pcmdata, 0);
+
+            std::vector<float> floatSamples;
+            floatSamples.reserve((delay_audio_frame_ + 1) * kFrameSamples);
 
-            const auto &result = offline_stream_->GetResult();
-            if (!result.text.empty() && out_callback_) {
-                out_callback_(result.text, true);
+            int16_t audio_val;
+            while (buffer_read_i16(pcmdata, &audio_val, 1)) {
+                floatSamples.push_back(static_cast<float>(audio_val) / 32768.0f);
             }
 
-            vad_->Pop();
+            buffer_resize(pcmdata, 0);
+            count = 0;
 
-            offline_stream_.reset();
-        }
+            vad_->AcceptWaveform(floatSamples.data(), floatSamples.size());
 
-        {
-            float chunk_ms = (delay_audio_frame_ + 1) * 10.0f;
-            if (detected) {
-                silence_ms_accum_ = 0.0f;
-            } else {
-                silence_ms_accum_ += chunk_ms;
+            bool detected      = vad_->IsSpeechDetected();
+            bool speech_start  = (!prev_vad_detected_ && detected);
+            prev_vad_detected_ = detected;
+
+            while (!vad_->Empty()) {
+                const auto &segment = vad_->Front();
+
+                if (!offline_stream_) {
+                    offline_stream_ = onnx_recognizer_->CreateStream();
+                }
+
+                if (speech_start && !pre_roll_pcm_.empty()) {
+                    std::vector<float> pre;
+                    pre.reserve(pre_roll_pcm_.size());
+                    for (int16_t s : pre_roll_pcm_) {
+                        pre.push_back(static_cast<float>(s) / 32768.0f);
+                    }
+
+                    std::vector<float> merged;
+                    merged.reserve(pre.size() + segment.samples.size());
+                    merged.insert(merged.end(), pre.begin(), pre.end());
+                    merged.insert(merged.end(), segment.samples.begin(), segment.samples.end());
+
+                    offline_stream_->AcceptWaveform(kSampleRate, merged.data(), merged.size());
+
+                    pre_roll_pcm_.clear();
+                    speech_start = false;
+                } else {
+                    offline_stream_->AcceptWaveform(kSampleRate, segment.samples.data(), segment.samples.size());
+                }
+
+                onnx_recognizer_->DecodeStream(offline_stream_.get());
+
+                const auto &result = offline_stream_->GetResult();
+                if (!result.text.empty() && out_callback_) {
+                    out_callback_(result.text, true);
+                }
+
+                vad_->Pop();
+
+                offline_stream_.reset();
             }
 
-            if (silence_ms_accum_ >= silence_timeout) {
-                if (ensleep_) {
-                    if (pause) pause();
+            {
+                float chunk_ms = (delay_audio_frame_ + 1) * 10.0f;
+                if (detected) {
+                    silence_ms_accum_ = 0.0f;
+                } else {
+                    silence_ms_accum_ += chunk_ms;
+                }
+
+                if (silence_ms_accum_ >= silence_timeout) {
+                    if (ensleep_) {
+                        if (pause) pause();
+                    }
+                    silence_ms_accum_ = 0.0f;
                 }
-                silence_ms_accum_ = 0.0f;
             }
         }
     }
@@ -755,8 +818,10 @@ class llm_asr : public StackFlow {
         std::string tmp_msg1;
         const std::string *next_data = &data;
         if (finish) {
-            tmp_msg1  = data + ".";
-            next_data = &tmp_msg1;
+            if (!llm_task_obj->is_using_itn()) {
+                tmp_msg1  = data + ".";
+                next_data = &tmp_msg1;
+            }
         }
 
         if (llm_channel->enstream_) {

From 312791ed9e85fe0c22afe00c2b838bd06252c20a Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Fri, 26 Dec 2025 17:11:23 +0800
Subject: [PATCH 78/79] [update] update llm-openai-api version

---
 projects/llm_framework/main_openai_api/SConstruct | 4 ++--
 projects/llm_framework/tools/llm_pack.py          | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/projects/llm_framework/main_openai_api/SConstruct b/projects/llm_framework/main_openai_api/SConstruct
index f6e45d1d..d2c3c816 100644
--- a/projects/llm_framework/main_openai_api/SConstruct
+++ b/projects/llm_framework/main_openai_api/SConstruct
@@ -18,7 +18,7 @@ LINK_SEARCH_PATH = []
 STATIC_FILES = []
 
 
-ModuleLLMOpenAIPluginPath = wget_github_commit('https://github.com/m5stack/ModuleLLM-OpenAI-Plugin.git', '50efecaf566a04c1786b7eda2208a0cdefd53dde', True)
+ModuleLLMOpenAIPluginPath = wget_github_commit('https://github.com/m5stack/ModuleLLM-OpenAI-Plugin.git', '52e951b32551f37299afc596c9cd3e8e20354054', True)
 python_venv = check_wget_down("https://m5stack.oss-cn-shenzhen.aliyuncs.com/resource/linux/llm/m5stack_llm-openai-api-python-venv_v1.6.tar.gz", 'm5stack_llm-openai-api-python-venv_v1.6.tar.gz')
 
 
@@ -52,7 +52,7 @@ ignore['ignore'] = list(set(ignore['ignore']))
 with open('../dist/fileignore', 'w') as f:
     json.dump(ignore, f, indent=4)
 
-env['COMPONENTS'].append({'target':'llm_openai_api-1.9',
+env['COMPONENTS'].append({'target':'llm_openai_api-1.10',
                           'SRCS':SRCS,
                           'INCLUDE':INCLUDE,
                           'PRIVATE_INCLUDE':PRIVATE_INCLUDE,
diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index f2b58d8d..766240e9 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -387,7 +387,7 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-depth-anything':[create_bin_deb,'llm-depth-anything', '1.7', src_folder, revision],
         'llm-vad':[create_bin_deb,'llm-vad', '1.8', src_folder, revision],
         'llm-whisper':[create_bin_deb,'llm-whisper', '1.8', src_folder, revision],
-        'llm-openai-api':[create_bin_deb,'llm-openai-api', '1.9', src_folder, revision],
+        'llm-openai-api':[create_bin_deb,'llm-openai-api', '1.10', src_folder, revision],
         'llm-cosy-voice':[create_bin_deb,'llm-cosy-voice', '1.10', src_folder, revision],
         # keyword spotting Audio file
         'llm-model-audio-en-us':[create_data_deb,'llm-model-audio-en-us', data_version, src_folder, revision],

From 9f34887f488f0e97a378a6e309d0f16d172b55af Mon Sep 17 00:00:00 2001
From: LittleMouse <niuchenyang@m5stack.com>
Date: Thu, 8 Jan 2026 14:48:09 +0800
Subject: [PATCH 79/79] [update] llm-model-audio version

---
 projects/llm_framework/tools/llm_pack.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/projects/llm_framework/tools/llm_pack.py b/projects/llm_framework/tools/llm_pack.py
index 766240e9..050e61e5 100755
--- a/projects/llm_framework/tools/llm_pack.py
+++ b/projects/llm_framework/tools/llm_pack.py
@@ -390,8 +390,8 @@ def create_bin_deb(package_name, version, src_folder, revision = 'm5stack1', dep
         'llm-openai-api':[create_bin_deb,'llm-openai-api', '1.10', src_folder, revision],
         'llm-cosy-voice':[create_bin_deb,'llm-cosy-voice', '1.10', src_folder, revision],
         # keyword spotting Audio file
-        'llm-model-audio-en-us':[create_data_deb,'llm-model-audio-en-us', data_version, src_folder, revision],
-        'llm-model-audio-zh-cn':[create_data_deb,'llm-model-audio-zh-cn', data_version, src_folder, revision],
+        'llm-model-audio-en-us':[create_data_deb,'llm-model-audio-en-us', '0.3', src_folder, revision],
+        'llm-model-audio-zh-cn':[create_data_deb,'llm-model-audio-zh-cn', '0.3', src_folder, revision],
         # Kaldi model        
         'llm-model-sherpa-ncnn-streaming-zipformer-20M-2023-02-17':[create_data_deb,'llm-model-sherpa-ncnn-streaming-zipformer-20M-2023-02-17', data_version, src_folder, revision],
         'llm-model-sherpa-ncnn-streaming-zipformer-zh-14M-2023-02-23':[create_data_deb,'llm-model-sherpa-ncnn-streaming-zipformer-zh-14M-2023-02-23', data_version, src_folder, revision],